diff --git a/.gitattributes b/.gitattributes index f5cec68804c6eda86026b596ccaf6baa5ac0a6f9..45bc2e74ea5cf1a1a521678bcfcdac69bbd4b8d0 100644 --- a/.gitattributes +++ b/.gitattributes @@ -2396,3 +2396,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/Temporal[[:space:]]Rate[[:space:]]Reduction[[:space:]]Clustering[[:space:]]for[[:space:]]Human[[:space:]]Motion[[:space:]]Segmentation/f14c6fc6-9149-4e8b-a6dd-a5de071b6ce5_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Temporal[[:space:]]Unlearnable[[:space:]]Examples_[[:space:]]Preventing[[:space:]]Personal[[:space:]]Video[[:space:]]Data[[:space:]]from[[:space:]]Unauthorized[[:space:]]Exploitation[[:space:]]by[[:space:]]Object[[:space:]]Tracking/fe5a296a-27d1-4006-9167-632435e6a7a3_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Temporal-aware[[:space:]]Query[[:space:]]Routing[[:space:]]for[[:space:]]Real-time[[:space:]]Video[[:space:]]Instance[[:space:]]Segmentation/6956edf3-7339-4603-a3e5-6531a6c8894e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Tensor-aggregated[[:space:]]LoRA[[:space:]]in[[:space:]]Federated[[:space:]]Fine-tuning/cfdeda6d-a34d-4c06-a268-6da1c14a07ae_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TerraMind_[[:space:]]Large-Scale[[:space:]]Generative[[:space:]]Multimodality[[:space:]]for[[:space:]]Earth[[:space:]]Observation/c9c92fd3-e6cf-49ea-bfb4-aa3c5b130a2c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Test-Time[[:space:]]Prompt[[:space:]]Tuning[[:space:]]for[[:space:]]Zero-Shot[[:space:]]Depth[[:space:]]Completion/0a2ddfb1-a84e-4b84-bcf1-12ab1957047a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Test-Time[[:space:]]Retrieval-Augmented[[:space:]]Adaptation[[:space:]]for[[:space:]]Vision-Language[[:space:]]Models/abcad282-0784-41dd-b230-4b9b3488f3c1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Test-time[[:space:]]Adaptation[[:space:]]for[[:space:]]Foundation[[:space:]]Medical[[:space:]]Segmentation[[:space:]]Model[[:space:]]Without[[:space:]]Parametric[[:space:]]Updates/8a26d38c-cedb-499b-b41f-c629d5b94654_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Text[[:space:]]Embedding[[:space:]]Knows[[:space:]]How[[:space:]]to[[:space:]]Quantize[[:space:]]Text-Guided[[:space:]]Diffusion[[:space:]]Models/885bbebf-0678-4829-a8d8-7dbca73ecb4a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Text-IRSTD_[[:space:]]Leveraging[[:space:]]Semantic[[:space:]]Text[[:space:]]to[[:space:]]Promote[[:space:]]Infrared[[:space:]]Small[[:space:]]Target[[:space:]]Detection[[:space:]]in[[:space:]]Complex[[:space:]]Scenes/f9aefd65-d52a-4731-ba81-d317b6edbcb4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Text-guided[[:space:]]Visual[[:space:]]Prompt[[:space:]]DINO[[:space:]]for[[:space:]]Generic[[:space:]]Segmentation/0364b0ce-3511-4546-ac4b-3ddb77cbd170_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Text-to-Any-Skeleton[[:space:]]Motion[[:space:]]Generation[[:space:]]Without[[:space:]]Retargeting/a12eb351-f679-4d5e-8b40-f5c8f124976d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Text2Outfit_[[:space:]]Controllable[[:space:]]Outfit[[:space:]]Generation[[:space:]]with[[:space:]]Multimodal[[:space:]]Language[[:space:]]Models/bbd1b25f-9206-484c-b93d-8a139709e547_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Text2VDM_[[:space:]]Text[[:space:]]to[[:space:]]Vector[[:space:]]Displacement[[:space:]]Maps[[:space:]]for[[:space:]]Expressive[[:space:]]and[[:space:]]Interactive[[:space:]]3D[[:space:]]Sculpting/011b91bd-0c46-4896-9ef0-7e4bb90ad7cb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TextMaster_[[:space:]]A[[:space:]]Unified[[:space:]]Framework[[:space:]]for[[:space:]]Realistic[[:space:]]Text[[:space:]]Editing[[:space:]]via[[:space:]]Glyph-Style[[:space:]]Dual-Control/948bb483-aee7-45ff-9c3e-421173343c17_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TextSSR_[[:space:]]Diffusion-based[[:space:]]Data[[:space:]]Synthesis[[:space:]]for[[:space:]]Scene[[:space:]]Text[[:space:]]Recognition/6a7bfaa2-52e3-41ce-8b5a-5f025a33ee77_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Textured[[:space:]]3D[[:space:]]Regenerative[[:space:]]Morphing[[:space:]]with[[:space:]]3D[[:space:]]Diffusion[[:space:]]Prior/0ee49b9a-7099-4408-8e6f-37bd5bc28baf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Best[[:space:]]of[[:space:]]Both[[:space:]]Worlds_[[:space:]]Integrating[[:space:]]Language[[:space:]]Models[[:space:]]and[[:space:]]Diffusion[[:space:]]Models[[:space:]]for[[:space:]]Video[[:space:]]Generation/b1aafd05-162c-4fda-a00d-cb706731d3d7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Curse[[:space:]]of[[:space:]]Conditions_[[:space:]]Analyzing[[:space:]]and[[:space:]]Improving[[:space:]]Optimal[[:space:]]Transport[[:space:]]for[[:space:]]Conditional[[:space:]]Flow-Based[[:space:]]Generation/2ff56f19-5f85-4cca-a5d4-ab26816c514c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Devil[[:space:]]is[[:space:]]in[[:space:]]the[[:space:]]Spurious[[:space:]]Correlations_[[:space:]]Boosting[[:space:]]Moment[[:space:]]Retrieval[[:space:]]with[[:space:]]Dynamic[[:space:]]Learning/10bd97eb-949b-4ebe-ab75-25cc73578707_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Inter-Intra[[:space:]]Modal[[:space:]]Measure_[[:space:]]A[[:space:]]Predictive[[:space:]]Lens[[:space:]]on[[:space:]]Fine-Tuning[[:space:]]Outcomes[[:space:]]in[[:space:]]Vision-Language[[:space:]]Models/f4964a00-88d3-4539-871c-39c0ed82d2c6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Scalability[[:space:]]of[[:space:]]Simplicity_[[:space:]]Empirical[[:space:]]Analysis[[:space:]]of[[:space:]]Vision-Language[[:space:]]Learning[[:space:]]with[[:space:]]a[[:space:]]Single[[:space:]]Transformer/2b8a94f1-28e8-4218-82c7-da4dba2c45e1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Silent[[:space:]]Assistant_[[:space:]]NoiseQuery[[:space:]]as[[:space:]]Implicit[[:space:]]Guidance[[:space:]]for[[:space:]]Goal-Driven[[:space:]]Image[[:space:]]Generation/76b9c942-da58-47e6-8ca6-ac0446eb367e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/The[[:space:]]Source[[:space:]]Image[[:space:]]is[[:space:]]the[[:space:]]Best[[:space:]]Attention[[:space:]]for[[:space:]]Infrared[[:space:]]and[[:space:]]Visible[[:space:]]Image[[:space:]]Fusion/cab6b62f-c3b3-45e2-8877-ae8b2379307d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Thermal[[:space:]]Polarimetric[[:space:]]Multi-view[[:space:]]Stereo/493fd467-1c08-4433-ac25-63db3b58338a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Think[[:space:]]Twice_[[:space:]]Test-Time[[:space:]]Reasoning[[:space:]]for[[:space:]]Robust[[:space:]]CLIP[[:space:]]Zero-Shot[[:space:]]Classification/87c172f1-4134-4138-9793-b4455c2d3c37_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TikZero_[[:space:]]Zero-Shot[[:space:]]Text-Guided[[:space:]]Graphics[[:space:]]Program[[:space:]]Synthesis/5102f011-579e-4213-9591-ec3e5e763b92_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Tile-wise[[:space:]]vs.[[:space:]]Image-wise_[[:space:]]Random-Tile[[:space:]]Loss[[:space:]]and[[:space:]]Training[[:space:]]Paradigm[[:space:]]for[[:space:]]Gaussian[[:space:]]Splatting/e220fb2d-345d-4c10-96de-761db897dc9b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Tiling[[:space:]]artifacts[[:space:]]and[[:space:]]trade-offs[[:space:]]of[[:space:]]feature[[:space:]]normalization[[:space:]]in[[:space:]]the[[:space:]]segmentation[[:space:]]of[[:space:]]large[[:space:]]biological[[:space:]]images/eaeeea1f-eecf-4d2e-8480-0e89bb127b12_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Time-Aware[[:space:]]Auto[[:space:]]White[[:space:]]Balance[[:space:]]in[[:space:]]Mobile[[:space:]]Photography/59296f3d-511f-4955-b6aa-60113b360f05_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TimeBooth_[[:space:]]Disentangled[[:space:]]Facial[[:space:]]Invariant[[:space:]]Representation[[:space:]]for[[:space:]]Diverse[[:space:]]and[[:space:]]Personalized[[:space:]]Face[[:space:]]Aging/b7999e96-8abf-47f1-9b7e-e5c9222126c0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TimeExpert_[[:space:]]An[[:space:]]Expert-Guided[[:space:]]Video[[:space:]]LLM[[:space:]]for[[:space:]]Video[[:space:]]Temporal[[:space:]]Grounding/a4a233d0-73c0-4327-bc43-78ef9be737e9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TimeFormer_[[:space:]]Capturing[[:space:]]Temporal[[:space:]]Relationships[[:space:]]of[[:space:]]Deformable[[:space:]]3D[[:space:]]Gaussians[[:space:]]for[[:space:]]Robust[[:space:]]Reconstruction/8cdc06bd-0ca5-4f6b-b012-50c8329288a4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Timestep-Aware[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Extreme[[:space:]]Image[[:space:]]Rescaling/5b02e9a3-c733-4ab2-9aa6-549d5b3e9de2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TinyViM_[[:space:]]Frequency[[:space:]]Decoupling[[:space:]]for[[:space:]]Tiny[[:space:]]Hybrid[[:space:]]Vision[[:space:]]Mamba/f0ba958e-31c7-4013-97d5-2552433cb070_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/To[[:space:]]Label[[:space:]]or[[:space:]]Not[[:space:]]to[[:space:]]Label_[[:space:]]PALM[[:space:]]-[[:space:]]A[[:space:]]Predictive[[:space:]]Model[[:space:]]for[[:space:]]Evaluating[[:space:]]Sample[[:space:]]Efficiency[[:space:]]in[[:space:]]Active[[:space:]]Learning[[:space:]]Models/8bac4e62-edf8-4833-a2b8-a238e4c4c68b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ToF-Splatting_[[:space:]]Dense[[:space:]]SLAM[[:space:]]using[[:space:]]Sparse[[:space:]]Time-of-Flight[[:space:]]Depth[[:space:]]and[[:space:]]Multi-Frame[[:space:]]Integration/93b627e5-bb27-4dc6-aa8c-e29320f5f5eb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Token[[:space:]]Activation[[:space:]]Map[[:space:]]to[[:space:]]Visually[[:space:]]Explain[[:space:]]Multimodal[[:space:]]LLMs/b852d151-66d5-47a7-8861-5fbe9cff3adc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Token-Efficient[[:space:]]VLM_[[:space:]]High-Resolution[[:space:]]Image[[:space:]]Understanding[[:space:]]via[[:space:]]Dynamic[[:space:]]Region[[:space:]]Proposal/b324faad-4fe9-44a7-a833-a82bac07835f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TokenUnify_[[:space:]]Scaling[[:space:]]Up[[:space:]]Autoregressive[[:space:]]Pretraining[[:space:]]for[[:space:]]Neuron[[:space:]]Segmentation/0bebb7e8-7094-4aab-a136-ac9635434593_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TokensGen_[[:space:]]Harnessing[[:space:]]Condensed[[:space:]]Tokens[[:space:]]for[[:space:]]Long[[:space:]]Video[[:space:]]Generation/439118b2-5f82-4c2e-9d89-cceba716e36e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ToolVQA_[[:space:]]A[[:space:]]Dataset[[:space:]]for[[:space:]]Multi-step[[:space:]]Reasoning[[:space:]]VQA[[:space:]]with[[:space:]]External[[:space:]]Tools/dd0c7eba-b554-4e21-9d01-9fb4080c4a97_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Top2Pano_[[:space:]]Learning[[:space:]]to[[:space:]]Generate[[:space:]]Indoor[[:space:]]Panoramas[[:space:]]from[[:space:]]Top-Down[[:space:]]View/ddd3f9c4-d03d-4cd7-96eb-1ef2931311e2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TopicGeo_[[:space:]]An[[:space:]]Efficient[[:space:]]Unified[[:space:]]Framework[[:space:]]for[[:space:]]Geolocation/43f123eb-2c45-4de4-8a4a-4958a152775f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TopoTTA_[[:space:]]Topology-Enhanced[[:space:]]Test-Time[[:space:]]Adaptation[[:space:]]for[[:space:]]Tubular[[:space:]]Structure[[:space:]]Segmentation/a88fe374-66af-4512-b87e-ddbd06fc2c78_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/TorchAdapt_[[:space:]]Towards[[:space:]]Light-Agnostic[[:space:]]Real-Time[[:space:]]Visual[[:space:]]Perception/60f819c6-a0ae-4fb0-be50-050c7068618b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Toward[[:space:]]Better[[:space:]]Out-painting_[[:space:]]Improving[[:space:]]the[[:space:]]Image[[:space:]]Composition[[:space:]]with[[:space:]]Initialization[[:space:]]Policy[[:space:]]Model/2c55e2cf-c2bf-4623-a4a2-6a1a33a77533_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Toward[[:space:]]Fair[[:space:]]and[[:space:]]Accurate[[:space:]]Cross-Domain[[:space:]]Medical[[:space:]]Image[[:space:]]Segmentation_[[:space:]]A[[:space:]]VLM-Driven[[:space:]]Active[[:space:]]Domain[[:space:]]Adaptation[[:space:]]Paradigm/0cab628c-59eb-4cba-b70b-2c418218511f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Toward[[:space:]]Long-Tailed[[:space:]]Online[[:space:]]Anomaly[[:space:]]Detection[[:space:]]through[[:space:]]Class-Agnostic[[:space:]]Concepts/331fec76-66a3-4c9b-96fc-eeb1ae13b1b6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Toward[[:space:]]Material-Agnostic[[:space:]]System[[:space:]]Identification[[:space:]]from[[:space:]]Videos/58fcfc32-a696-4a6d-a6cd-3c218994c3ba_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Accurate[[:space:]]and[[:space:]]Efficient[[:space:]]3D[[:space:]]Object[[:space:]]Detection[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving_[[:space:]]A[[:space:]]Mixture[[:space:]]of[[:space:]]Experts[[:space:]]Computing[[:space:]]System[[:space:]]on[[:space:]]Edge/8be189e3-e5ec-449b-9eb4-67198582e1c8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Adversarial[[:space:]]Robustness[[:space:]]via[[:space:]]Debiased[[:space:]]High-Confidence[[:space:]]Logit[[:space:]]Alignment/be7e63af-2995-4db4-b164-1814191af90b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Annotation-Free[[:space:]]Evaluation_[[:space:]]KPAScore[[:space:]]for[[:space:]]Human[[:space:]]Keypoint[[:space:]]Detection/c6c6dbbf-396b-41fc-b50d-460da67eb4ab_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Comprehensive[[:space:]]Lecture[[:space:]]Slides[[:space:]]Understanding_[[:space:]]Large-scale[[:space:]]Dataset[[:space:]]and[[:space:]]Effective[[:space:]]Method/58e52446-3fb4-48af-8a99-0aa5fc34cde3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Cross-modal[[:space:]]Backward-compatible[[:space:]]Representation[[:space:]]Learning[[:space:]]for[[:space:]]Vision-Language[[:space:]]Models/5c932632-b5a7-42ba-9baf-b3ca38b774d6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Effective[[:space:]]Foundation[[:space:]]Model[[:space:]]Adaptation[[:space:]]for[[:space:]]Extreme[[:space:]]Cross-Domain[[:space:]]Few-Shot[[:space:]]Learning/2b5d04f4-f92c-4faf-b657-95623ea87551_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Efficient[[:space:]]General[[:space:]]Feature[[:space:]]Prediction[[:space:]]in[[:space:]]Masked[[:space:]]Skeleton[[:space:]]Modeling/23b6091d-1a61-4a72-bde7-0fb203b07465_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Explicit[[:space:]]Exoskeleton[[:space:]]for[[:space:]]the[[:space:]]Reconstruction[[:space:]]of[[:space:]]Complicated[[:space:]]3D[[:space:]]Human[[:space:]]Avatars/83fba3ab-8c7e-4045-994b-8f70e4fc8d14_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Fine-grained[[:space:]]Interactive[[:space:]]Segmentation[[:space:]]in[[:space:]]Images[[:space:]]and[[:space:]]Videos/b24a62fe-1619-470e-918c-9d3158f1e3f4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Foundational[[:space:]]Models[[:space:]]for[[:space:]]Single-Chip[[:space:]]Radar/6536d138-fd2a-4e4b-a53a-8d73d592cae1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Higher[[:space:]]Effective[[:space:]]Rank[[:space:]]in[[:space:]]Parameter-Efficient[[:space:]]Fine-tuning[[:space:]]using[[:space:]]Khatri-Rao[[:space:]]Product/5863579c-c0a2-49e2-85ef-c6b4d99e52c6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Human-like[[:space:]]Virtual[[:space:]]Beings_[[:space:]]Simulating[[:space:]]Human[[:space:]]Behavior[[:space:]]in[[:space:]]3D[[:space:]]Scenes/099d59e1-3cf8-46ca-a7c2-5f8c85602d06_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Immersive[[:space:]]Human-X[[:space:]]Interaction_[[:space:]]A[[:space:]]Real-Time[[:space:]]Framework[[:space:]]for[[:space:]]Physically[[:space:]]Plausible[[:space:]]Motion[[:space:]]Synthesis/4564e033-7cc9-41ba-b940-20704a40b38a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Long-Horizon[[:space:]]Vision-Language-Action[[:space:]]System_[[:space:]]Reasoning,[[:space:]]Acting[[:space:]]and[[:space:]]Memory/d546fd0e-5080-4957-803b-5e64e0ab9d88_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]More[[:space:]]Diverse[[:space:]]and[[:space:]]Challenging[[:space:]]Pre-training[[:space:]]for[[:space:]]Point[[:space:]]Cloud[[:space:]]Learning_[[:space:]]Self-Supervised[[:space:]]Cross[[:space:]]Reconstruction[[:space:]]with[[:space:]]Decoupled[[:space:]]Views/3caf6ed2-ef92-4e47-ab15-82d341d3a49f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Omnimodal[[:space:]]Expressions[[:space:]]and[[:space:]]Reasoning[[:space:]]in[[:space:]]Referring[[:space:]]Audio-Visual[[:space:]]Segmentation/7d9586d8-b2bc-4332-a23e-c79a83d7ffa9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Towards[[:space:]]Open-World[[:space:]]Generation[[:space:]]of[[:space:]]Stereo[[:space:]]Images[[:space:]]and[[:space:]]Unsupervised[[:space:]]Matching/2355f2bc-16e1-4dc5-88a7-53284ff42cd2_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/Tensor-aggregated LoRA in Federated Fine-tuning/cfdeda6d-a34d-4c06-a268-6da1c14a07ae_content_list.json b/2025/Tensor-aggregated LoRA in Federated Fine-tuning/cfdeda6d-a34d-4c06-a268-6da1c14a07ae_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1c400c1c7664d37f9b4fc00c5ffba40a4df04752 --- /dev/null +++ b/2025/Tensor-aggregated LoRA in Federated Fine-tuning/cfdeda6d-a34d-4c06-a268-6da1c14a07ae_content_list.json @@ -0,0 +1,1498 @@ +[ + { + "type": "text", + "text": "Tensor-aggregated LoRA in Federated Fine-tuning", + "text_level": 1, + "bbox": [ + 243, + 130, + 756, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhixuan Li $^{1*}$ , Binqian Xu $^{1*}$ , Xiangbo Shu $^{1\\dagger}$ , Jiachao Zhang $^{2}$ , Yazhou Yao $^{1}$ , Guo-Sen Xie $^{1}$ , Jinhui Tang $^{3}$ $^{1}$ Nanjing University of Science and Technology \n $^{2}$ Nanjing Institute of Technology \n $^{3}$ Nanjing Forestry University", + "bbox": [ + 99, + 179, + 906, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lzxlizhixuan@njust.edu.cn, xubinq11@gmail.com, shuxb@njust.edu.cn, zhangjc07@foxmail.com, yazhou.yao@njust.edu.cn, gsxiehm@gmail.com, tangjh@njfu.edu.cn", + "bbox": [ + 109, + 253, + 893, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The combination of Large Language Models (LLMs) and Federated Learning (FL) to leverage privacy-preserving data has emerged as a promising approach to further enhance the Parameter-Efficient Fine-Tuning (PEFT) capabilities of LLMs. In real-world FL settings with resource heterogeneity, the training process of Low-Rank Adaptation (LoRA), the representative PEFT method, still faces two major challenges: aggregation noise and aggregation misalignment. In this paper, we propose a novel Tensor-aggregated LoRA (Te-LoRA) in Federated Finetuning based on an alternating-freeze training strategy to avoid aggregating noise without additional server-side computational costs, while mitigating aggregation suboptimality caused by parameter misalignment between heterogeneous LoRAs. Especially in addressing the aggregation suboptimality issue, we design the Pre-Aggregation Alignment strategy (PAA-strategy) and Tensor-to-Matrix strategy (T2M-strategy) for aligning heterogeneous LoRAs and aggregating them into an united tensor, which is then decomposed into matrices adapted for client download. Extensive experiments demonstrate the effectiveness and robustness of Te-LoRA in both homogeneous and heterogeneous settings.", + "bbox": [ + 89, + 353, + 483, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 731, + 220, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large Language Models (LLMs) have demonstrated exceptional performance across various task scenarios, such as search engines [20], chatbots [4], healthcare [37], and more. When using pre-trained LLMs to adapt to downstream tasks, significant computational resources are required to fine-tune the model parameters. Therefore, to improve the applicability of LLMs, it is necessary to fine-tune specific", + "bbox": [ + 89, + 757, + 482, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54b59272811ac57b49a56cb3edf41ae5257c02136f2020640ceed37df523cae8.jpg", + "image_caption": [ + "Figure 1. Comparison of different federated fine-tuning LoRAs. Compared to methods (a), (b), and (c), our Te-LoRA performs better in addressing aggregation noise, communication cost, and aggregation misalignment issues." + ], + "image_footnote": [], + "bbox": [ + 516, + 320, + 911, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "datasets [9, 17, 25]. However, collecting such datasets comes with associated costs and privacy concerns [5, 18, 38]. Consequently, researchers have turned to Federated Learning (FL) as a technical approach to fine-tune LLMs using distributed client data without compromising data privacy [2, 30, 33]. In this context, researchers have explored the application of Parameter-Efficient Fine-Tuning (PEFT) [11, 19, 23] methods for federated fine-tuning of LLMs, saving substantial computational resources. Among these, Low-Rank Adaptation (LoRA) [14] has gained attention due to its significant reduction in the number of communication parameters.", + "bbox": [ + 511, + 626, + 906, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, simply applying LoRA in real FL environment presents several challenges, one of the key issues being aggregation noise. In detail, traditional federated fine-tuning of LLMs [36, 46] averages modules A and B of local LoRAs separately, leading to discrepancies between the aggregated and ideal global LoRAs, thereby introducing noise into the", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions", + "bbox": [ + 107, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1058", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "update process. To address this issue, as shown in Fig. 1, the two main solutions are as follows: 1) Merging modules A and B [3, 43], where A and B are first merged by matrix multiplication and then aggregated. 2) Alternating the training of modules A and B [22, 34], where either A or B is kept frozen for consistency at each round, with only the learning B or A being individually aggregated. Compared to the \"Merging\" strategy, the \"Alternating\" strategy reduces communication costs by nearly half, as it transmits only A or B per round.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The existing \"Alternating\" method [22] further reduces communication costs by selecting weights from a global LoRA with a unified initial rank. However, this homogenized assumption contradicts the core feature of resource heterogeneity in FL [16, 35, 40], where real-world scenarios involve clients with local LoRAs of different initial ranks. When LoRA modules A or B with heterogeneous ranks are uploaded and aggregated, an explicit rank dimension misalignment arises. Even when adopting the \"merging\" strategy for weight aggregation through matrix multiplication, semantically driven feature alignment issues may still persist. This multi-level parameter/feature mismatch caused by rank heterogeneity and data heterogeneity can be referred to as the aggregation misalignment problem.", + "bbox": [ + 89, + 243, + 482, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To simultaneously mitigate the issues of aggregation noise and aggregation misalignment in heterogeneous settings, as shown in Fig. 1, we propose a novel Tensor-aggregated LoRA called Te-LoRA. Te-LoRA consists of two main strategies: First, due to the impact of heterogeneity, the ranks of the matrices uploaded by each client to the server are often inconsistent. We present a Pre-Aggregation Alignment strategy (PAA-strategy) to align rank while minimizing information loss or conflict. Second, to preserve the potential relationships and dependencies between different clients, we design the key Tensor-to-Matrix strategy (T2M-strategy) for decomposing the tensor stacked from aligned matrices. T2M-strategy can adapt to the aggregation of heterogeneous LoRAs while effectively preserving the key components from each client. The main contributions of this work are summarized as follows:", + "bbox": [ + 89, + 455, + 482, + 698 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Alternating-freeze training for Tensor version of LoRA. We propose a novel Tensor-aggregated LoRA (Te-LoRA) in Federated Fine-tuning to efficiently aggregate the main components of each client based on an alternating-freeze training approach, avoiding aggregation noise while alleviating aggregation suboptimality among heterogeneous LoRAs.", + "- Alignment of heterogeneous LoRAs for server aggregation. We design a Pre-Aggregation Alignment strategy (PAA-strategy) to align heterogeneous LoRAs, while preserving the unique features of each client and minimizing general information loss.", + "- Decomposition of aggregated LoRAs for client down" + ], + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "load. We devise a Tensor-to-Matrix strategy (T2M-strategy) that decomposes the tensor into matrices adapted for client download, after aggregating LoRAs into an united tensor covering relationships and dependencies among LoRAs.", + "bbox": [ + 526, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 183, + 653, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Parameter-Efficient Fine-Tuning", + "text_level": 1, + "bbox": [ + 513, + 208, + 800, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To effectively fine-tune LLMs while reducing computational resources and storage requirements, Parameter-Efficient Fine-Tuning (PEFT) techniques have emerged. Among these, Prefix-Tuning [27] guides the pre-trained language model's generation process by optimizing a continuous sequence of task-specific vectors (called prefixes) while keeping the model parameters frozen. Soft prompts [23] enable frozen language models to perform specific tasks by propagating signals of labeled examples. Both methods are effective for task-specific adaptation. LoRA [14] reduces memory overhead by representing weight updates with low-rank matrices. AdaLoRA [47] optimizes incremental updates of pre-trained language models by adaptively allocating parameter budgets and using singular value decomposition, effectively reducing the parameter budget and avoiding intensive computation. LongLoRA [6] extends the context length of LLMs at low computational cost through sparse local attention and parameter-efficient fine-tuning. SplitLoRA [28], based on Split Federated Learning (SFL), improves training efficiency by partitioning the model and combining the parallel training of federated learning with the model splitting advantage of split learning. PEFT-based methods for LLMs, especially LoRA and its variants, have advanced fine-tuning techniques, showing excellent performance due to their simplicity and ability to reduce parameters. Thus, we fine-tune LLMs with LoRA in this paper.", + "bbox": [ + 511, + 231, + 903, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Federated Fine-Tuning Homogeneous LoRA", + "text_level": 1, + "bbox": [ + 513, + 636, + 888, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In federated learning, several works have applied LoRA to achieve reduced communication and improved efficiency. For example, Zhang et al. [46] fine-tuned LLMs by aggregating heterogeneous instruction data from multiple clients in a federated learning framework, improving the model's generalization ability on new tasks while ensuring user privacy and data security. Wu et al. [44] introduced the FedBiOT algorithm, which compresses LLMs and splits them into two components, using a dual-layer optimization approach to achieve efficient fine-tuning and reduce computational and communication costs. FLoCoRA [12] also reduces communication overhead by freezing the originally randomized neural network parameters during training, and only training and transmitting the added LoRA adapters. However, it does not fully address the stability issues brought about by extremely low-bit quantization. Sun", + "bbox": [ + 511, + 659, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1059", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/431a37c8c6f64fce36911e3c265e56238e30c9c4afbb8bd6e61d58f9a101a5f3.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed Te-LoRA. It is based on a framework for alternating training of the LoRA modules B and A. When the clients' LoRA have different ranks, the B or A modules are obtained after client training and subsequently uploaded to the server for aggregation. On the server, the heterogeneous B or A modules are first aligned using the PAA-strategy. In the PAA-strategy, $\\mu$ and $\\nu$ are all 1-column vectors of size having $r_{\\mathrm{max}}$ and $r_k$ , respectively. Then, the aligned B or A are aggregated using the T2M-strategy. The detailed operations will be discussed in Sec. 3." + ], + "image_footnote": [], + "bbox": [ + 94, + 93, + 929, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "et al. [39] pointed out that separately integrating the two LoRA matrices does not fully approximate the original, and proposed FFA-LoRA, which fixes the randomly initialized matrices while fine-tuning only the zero-initialized ones. Compared to FFA-LoRA, the proposed Te-LoRA, based on an alternating freezing framework, not only preserves the training of all LoRA modules but also effectively addresses the issue of aggregation noise.", + "bbox": [ + 88, + 406, + 480, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Federated Fine-tuning Heterogeneous LoRA", + "text_level": 1, + "bbox": [ + 89, + 542, + 465, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to non-IID data heterogeneity [26] sampled from the Dirichlet distribution and heterogeneous hardware resources, clients need to adjust the LoRA rank accordingly. Existing work has proposed methods to address this issue. For instance, SLoRA [2] addresses LoRA's limitations in high data heterogeneity scenarios with data-driven initialization, achieving performance similar to full finetuning while reducing training time and communication costs. FLoRA [43] uses a stacked aggregation method to eliminate noise in low-rank adapter (LoRA) aggregation, ensuring noise-free global updates. It supports heterogeneous LoRA configurations across clients, allowing dynamic adjustments based on data complexity and resources. However, this approach leads to linear growth in communication complexity, higher computational and storage demands, and reduced flexibility for supporting multiple task-specific adapters. Cho et al. [8] deploy LoRA modules with different ranks across clients, using zero-padding and truncation for aggregation, combining the benefits of high- and low-rank LoRA. However, the padded zeros dilute useful information from high-quality clients, reducing optimization efficiency and potentially limiting perfor", + "bbox": [ + 91, + 569, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mance. FlexLoRA [3] dynamically adjusts the LoRA rank for clients and uses Singular Value Decomposition (SVD) to reallocate global weights, enhancing resource utilization and model generalization. However, SVD truncation adds computational and storage overhead, and the resulting errors may limit performance.", + "bbox": [ + 511, + 406, + 906, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LoRA-A2 [22] alternates freezing the A and B components of the module and dynamically selects key parameters based on the importance of each client's data, effectively reducing parameter inconsistencies during the aggregation of low-rank adapters and significantly cutting communication costs. However, the method assumes that the global LoRA has a uniform initial rank, whereas in practice, each client's LoRA begins with a different rank. Compared to the related LoRA-A2, we train the initial heterogeneous-rank LoRA directly on the clients. While LoRA-A2 focuses on adaptively selecting client ranks, we emphasize the effective alignment of heterogeneous ranks on the server side. During alignment, unlike FlexLoRA and FLoRA, which use merging strategies, we designed the PAA-strategy and T2M-strategy for adaptive alignment, preserving the effective information within clients and dependencies across clients.", + "bbox": [ + 511, + 500, + 908, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 758, + 604, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 786, + 653, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Low-Rank Adaptation. The application of Low-Rank Adaptation (LoRA) [15] facilitates a reduction in the number of trainable parameters during the fine-tuning of a pretrained model. Specifically, the weight matrix $W \\in \\mathbb{R}^{d \\times l}$ in the fine-tuned model can be represented as the sum of the original weight matrix $W_0 \\in \\mathbb{R}^{d \\times l}$ and an updated", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1060", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "weight matrix $\\triangle W\\in \\mathbb{R}^{d\\times l}$ . $\\triangle W$ is implemented by low-rank decomposition into smaller matrices $B\\in \\mathbb{R}^{d\\times r}$ and $A\\in \\mathbb{R}^{r\\times l}$ :", + "bbox": [ + 89, + 90, + 485, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW = W _ {0} + \\triangle W = W _ {0} + B A, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 145, + 482, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where rank $r$ is much smaller than both dimensions $d$ and $l$ , thereby significantly reducing the number of trainable parameters for the weights.", + "bbox": [ + 89, + 170, + 482, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Aggregation Noise. Currently, FedAvg [30], a widely used FL algorithm, updates the global model by performing a weighted average of $n$ local clients in each communication round. When LoRA is applied to FL, training and transmitting only the updated low-rank matrices B and A significantly reduces both computational and communication costs, compared to directly updating and transmitting $\\triangle W$ . Combining the aforementioned FedAvg and LoRA, FedIT [46] is proposed for federated fine-tuning, where the global LoRA matrices A and B are updated by weighted averaging of the local LoRA matrices $A_{i}$ and $B_{i}$ from $n$ clients, as follows:", + "bbox": [ + 89, + 215, + 483, + 397 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\triangle W = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\triangle W _ {i} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} B _ {i} \\times \\frac {1}{n} \\sum_ {i = 1} ^ {n} A _ {i}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 404, + 482, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, when each weight is averaged and aggregated separately, there is a bias between the aggregated parameters (left) and ideal (right), as shown below:", + "bbox": [ + 89, + 450, + 483, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{n} \\sum_ {i = 1} ^ {n} B _ {i} \\times \\frac {1}{n} \\sum_ {i = 1} ^ {n} A _ {i} \\neq \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(B _ {i} A _ {i}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 503, + 482, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One solution is to first merge $A_{i}$ and $B_{i}$ from each client by matrix multiplication and then perform average aggregation, i.e., $\\triangle W = \\frac{1}{n}\\sum_{i=1}^{n}(B_{i}A_{i})$ . Subsequently, the aggregated $\\triangle W \\in \\mathbb{R}^{d\\times l}$ must be decomposed into the new $A_{i} \\in \\mathbb{R}^{r\\times l}$ and $B_{i} \\in \\mathbb{R}^{d\\times r}$ for each client. This method introduces complex matrix decomposition, making post-decomposition training prone to instability. The details of another simple and effective solution are as follows.", + "bbox": [ + 89, + 550, + 483, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Alternating Freezing Technique. In [22], a simple alternating freezing technique is introduced to mitigate the issue of aggregation noise when applying LoRA in FL. Specifically, in round $t$ , the aggregated $\\bar{A}^{(t)}$ remains frozen, and each client only trains $B_{i}^{(t)}$ , as follows:", + "bbox": [ + 89, + 671, + 483, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{n} \\sum_ {i = 1} ^ {n} B _ {i} ^ {(t)} \\times \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\bar {A} ^ {(t)} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(B _ {i} ^ {(t)} \\bar {A} ^ {(t)}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 758, + 482, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, in round $t + 1$ , the aggregated $\\bar{B}^{(t + 1)}$ is frozen, and only $A_{i}^{(t + 1)}$ is trainable on each client, as shown below:", + "bbox": [ + 89, + 806, + 483, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{n} \\sum_ {i = 1} ^ {n} \\bar {B} ^ {(t + 1)} \\times \\frac {1}{n} \\sum_ {i = 1} ^ {n} A _ {i} ^ {(t + 1)} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\bar {B} ^ {(t + 1)} A _ {i} ^ {(t + 1)}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 848, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, A and B complete iterative training and aggregation through the alternating freezing technique. However, when faced with data heterogeneity and rank heterogeneity at the client, this approach does not achieve good alignment of modules A or B at the server.", + "bbox": [ + 511, + 90, + 906, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 175, + 687, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In real FL environments with resource heterogeneity, the training of Low-Rank Adaptation (LoRA) faces two major challenges: aggregation noise and aggregation misalignment. To address these issues, we propose a novel federated fine-tuning method with Tensor-aggregated LoRA (Te-LoRA) based on an alternating freezing training strategy, as shown in Fig. 2. Te-LoRA consists of three strategies: the alternating freezing training strategy, the Pre-Aggregation Alignment Strategy (PAA-strategy), and the Tensor Decomposition Aggregation Strategy (T2M-strategy). Based on the existing alternating freezing training strategy (Eqs. (4) and (5)), Te-LoRA first effectively avoids aggregation noise (Eq. (3)). Subsequently, the PAA-strategy aims to achieve parameter alignment between different rank modules A and B with minimal information loss. Furthermore, the T2M-strategy adaptively aligns and aggregates the main components among clients to achieve more fine-grained feature alignment. The PAA and T2M strategies complement each other, effectively mitigating the issue of aggregation misalignment.", + "bbox": [ + 511, + 196, + 906, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-Aggregation Alignment strategy (PAA-strategy). The PAA-strategy is presented to achieve the optimal transformation from heterogeneous LoRAs to homogeneous LoRAs with minimal information loss for parameter alignment on the server side. In the implementation of the alternating freezing FL training strategy, the modules transmitted between the client and server are either B or A. Assuming that the module B is uploaded in the current round, module A remains frozen and stays on the client side. After the upload, the server receives a set of $K$ modules B, i.e., $\\{B_k \\in \\mathbb{R}^{d \\times r_k}\\}_{k=1}^K$ . The module with the highest rank is selected as the target module $B_{\\max} = B_{\\arg \\max_{k=1,\\dots,K} r_k}$ , and the remaining modules are aligned towards the target module. Specifically, the distance matrix $\\mathcal{C}$ between $B_k$ and $B_{\\max}$ is calculated as follows:", + "bbox": [ + 511, + 500, + 908, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} (m, n) = \\sum_ {i = 1} ^ {d} \\left(b _ {k} ^ {i m} - b _ {\\max } ^ {i n}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 737, + 906, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{C} \\in \\mathbb{R}^{r_k \\times r_{\\max}}$ , $b_k^{im}$ and $b_{\\max}^{in}$ are the elements of the $i^{\\text{th}}$ row, $m^{\\text{th}}$ column of matrix $B_k$ , and $n^{\\text{th}}$ column of matrix $B_{\\max}$ , respectively. To smooth the matching relationship between the source and target matrices, the matrix $\\mathcal{C}$ is regularized: $\\mathcal{K} = \\exp\\left(-\\frac{\\mathcal{C}}{\\eta}\\right)$ , where $\\eta$ is the regularization parameter. Next, based on the matrix $\\mathcal{K}$ , we first assume that the source distribution $p$ and the target distribution $q$", + "bbox": [ + 511, + 787, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1061", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "follow a uniform distribution, i.e., $p = \\left[\\frac{1}{r_k},\\frac{1}{r_k},\\dots \\frac{1}{r_k}\\right]^\\top$ , $q = \\left[\\frac{1}{r_{\\mathrm{max}}},\\frac{1}{r_{\\mathrm{max}}},\\dots \\frac{1}{r_{\\mathrm{max}}}\\right]^\\top$ . Then, using the Sinkhorn-Knopp algorithm [31], we iteratively update the adjustment factors $\\mu$ and $\\nu$ to make the probability matrix $\\mathcal{G}$ converge to a solution that satisfies the marginal constraints of $p$ and $q$ , as follows:", + "bbox": [ + 89, + 88, + 483, + 200 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu = \\frac {p}{\\mathcal {K} \\cdot \\nu}, \\quad \\nu = \\frac {q}{\\mathcal {K} ^ {\\top} \\cdot \\mu}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 209, + 482, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the adjustment factors $\\mu$ and $\\nu$ are both 1-column vectors of size $r_{\\mathrm{max}}$ and $r_k$ , respectively. By iteratively updating until convergence, the probability matrix $\\mathcal{G}$ , representing the optimal alignment between the source and target features, is shown below:", + "bbox": [ + 89, + 247, + 483, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} = \\mu \\cdot \\mathcal {K} \\cdot \\nu^ {\\top}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 330, + 482, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, the source matrix $B_{k}$ is transformed using the matrix $\\mathcal{G}$ to achieve alignment of different ranks with minimal information loss, as follows:", + "bbox": [ + 89, + 359, + 483, + 405 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {B _ {k}} = B _ {k} \\cdot \\mathcal {G}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 414, + 482, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widetilde{B_k} \\in \\mathbb{R}^{d \\times r_{\\max}}$ is the aligned matrix. The remaining module B and the module A from another round both use the same pre-aggregation process for parameter alignment. Discussions. Although previous research in FL, such as FedOTP [24] and FedAli [10], has explored Optimal Transport (OT) for fine-grained feature alignment, FedOTP [24] uses OT to align local visual features with global and local text features, smoothing the global consensus and local personalization. To balance personalization and generalization, FedAli [10] quantifies the cost of embedding inputs into prototypes using OT, effectively aligning embeddings and reducing client inconsistencies. Compared to these studies, our PAA strategy differs in three ways: 1) The others focus on data heterogeneity in personalized FL, while our work addresses resource heterogeneity in real FL; 2) Previous work aligns personalized with global knowledge via OT, while our work aligns heterogeneous weights into homogeneous ones using OT to address rank mismatches from resource heterogeneity; 3) Previous studies focus on aligning prompts and prototypes, while our work targets alignment methods for heterogeneous LoRAs. To the best of our knowledge, this is the first research to apply OT for heterogeneous LoRA alignment in resource-heterogeneous FL.", + "bbox": [ + 89, + 441, + 483, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tensor-to-Matrix strategy (T2M-strategy). To maintain the relationships and dependencies between different clients, the Tensor-to-Matrix strategy (T2M-strategy) is designed to adaptively aggregate the LoRAs aligned by the PAA-strategy, rather than simply averaging them. Specifically, given the rank-aligned set $\\{\\widetilde{B_k}\\}_{k = 1}^K$ of modules B, $\\widehat{\\mathcal{B}}$ can be stacked from $\\{\\widetilde{B_k}\\}_{k = 1}^K$ to form a new tensor", + "bbox": [ + 89, + 791, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1: Te-LoRA", + "text_level": 1, + "bbox": [ + 529, + 92, + 692, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialization: each client $k$ initializes the LoRA $\\triangle W_{k} = B_{k}A_{k}$ with $B_{k}\\in \\mathbb{R}^{d\\times r_{k}}$ and $A_{k}\\in \\mathbb{R}^{r_{k}\\times d}$ .", + "bbox": [ + 529, + 112, + 906, + 143 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. for $t = 1,2,\\ldots ,T$ do:", + "2. Select K participants for round t", + "3. if $t\\% 2 = 1$ then:", + "4. for $k = 1,2,\\dots ,K$ in parallel do:", + "5. $B_{k}^{(t + 1)} = \\mathrm{LocalTraining}\\left(B^{(t)},k\\right)$", + "6. $\\widetilde{B}_k^{(t + 1)}\\gets \\mathrm{PAA}\\left(B_k^{(t + 1)}\\right)$", + "7. $\\mathbb{B}_{*}^{(t + 1)}\\gets \\mathrm{T2M}\\left(\\widetilde{B}_{1}^{(t + 1)},\\widetilde{B}_{2}^{(t + 1)},\\dots,\\widetilde{B}_{\\mathrm{K}}^{(t + 1)}\\right)$", + "8. $\\overline{A}^{(t + 1)} = A^{(t)}$", + "9. end for", + "10. else:", + "11. for $\\mathrm{k} = 1,2,\\cdots ,\\mathrm{K}$ in parallel do:", + "12. $A_{k}^{(t + 1)} = \\mathrm{LocalTraining}\\left(A^{(t)},k\\right)$", + "13. $\\widetilde{A}_k^{(t + 1)}\\gets \\mathrm{PAA}\\left(A_k^{(t + 1)}\\right)$", + "14. $\\mathbb{A}_{*}^{(t + 1)}\\gets \\mathrm{T2M}\\left(\\widetilde{A}_{1}^{(t + 1)},\\widetilde{A}_{2}^{(t + 1)},\\dots ,\\widetilde{A}_{\\mathrm{K}}^{(t + 1)}\\right)$", + "15. $\\overline{B}^{(t + 1)} = B^{(t)}$", + "16. end for", + "17. end if", + "18. end for" + ], + "bbox": [ + 529, + 147, + 906, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "representing the intermediate state before aggregation, i.e., $\\widehat{\\mathcal{B}} = \\bigcup_{k=1}^{K} \\widetilde{B_k}$ , where $\\widetilde{B_k} \\in \\mathbb{R}^{d \\times r_{\\max}}$ and $\\widehat{\\mathcal{B}} \\in \\mathbb{R}^{d \\times r_{\\max} \\times K}$ . Here, the aggregation process can be viewed as a tensor decomposition process. Inspired by [21], three versions of the T2M-strategy (T2M $_1$ , T2M $_2$ , and T2M $_3$ ) is explored to decompose $\\widehat{\\mathcal{B}}$ into the result $\\mathbb{B}_*$ . The generalized form of the decomposition is as follows:", + "bbox": [ + 511, + 527, + 906, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {B}} = \\mathbb {B} _ {*} \\times_ {1} U \\times_ {2} V \\times_ {3} X, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 645, + 903, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{B}_*$ is a core tensor, $U,V$ , and $X$ serve as the factor matrices, and the symbol $\\times_{i}$ ( $i = 1,2,3$ ) denotes mode- $i$ tensor-matrix multiplication. $\\mathrm{T2M_1}$ is as follows:", + "bbox": [ + 511, + 676, + 906, + 722 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {B}} = \\mathbb {B} _ {*} ^ {(1)} \\times_ {1} U ^ {(1)} \\times_ {2} V ^ {(1)} = \\left[ \\left[ \\mathbb {B} _ {*} ^ {(1)}; U ^ {(1)}, V ^ {(1)}, \\mathbb {I} \\right] \\right], \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 732, + 903, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{B}_{*}^{(1)}\\in \\mathbb{R}^{1\\times 1\\times K}$ , $U^{(1)}\\in \\mathbb{R}^{d\\times 1}$ , $V^{(1)}\\in \\mathbb{R}^{r_{max}\\times 1}$ and $X = \\mathbb{I}$ is the $K\\times K$ identity matrix. [ ] denotes the shorthand. $\\mathbb{B}_{*}^{(1)}$ represents the weight coefficient of different clients, reflecting the importance scores. And the aggregated module is $\\bar{B}$ , where $\\bar{B}_{i,j} = \\sum_{k = 1}^{K}\\hat{\\mathcal{B}}_{i,j,k}\\cdot \\mathbb{B}_{*1,1,k}^{(1)}$ . The formulae of T2M2 is as follows:", + "bbox": [ + 511, + 770, + 906, + 868 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\mathcal {B}} = \\mathbb {B} _ {*} ^ {(2)} \\times_ {3} X ^ {(2)} = \\left[ \\left[ \\mathbb {B} _ {*} ^ {(2)}; \\mathbb {I}, \\mathbb {I}, X ^ {(2)} \\right] \\right], \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 877, + 903, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1062", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{B}_*^{(2)}\\in \\mathbb{R}^{d\\times r_{max}\\times 1},X^{(2)}\\in \\mathbb{R}^{K\\times 1},U = \\mathbb{I}$ and $V = \\mathbb{I}$ are the $d\\times d$ and $r_{max}\\times r_{max}$ identity matrices. $\\mathbb{B}_*^{(2)}$ is the aggregation result capturing the common and representative information, i.e., $\\bar{B} = \\mathbb{B}_*^{(2)}$ . T2M3 is a combined version of T2M1 and T2M2, as shown below:", + "bbox": [ + 89, + 88, + 480, + 167 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {B} _ {*} ^ {(3)} = \\mathrm {T} 2 \\mathrm {M} _ {2} (\\widehat {\\mathcal {B}} \\odot \\mathrm {T} 2 \\mathrm {M} _ {1} (\\widehat {\\mathcal {B}})), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 172, + 480, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbb{B}_{*}^{(1)} = \\mathrm{T2M}_{1}(\\widehat{\\mathcal{B}})$ , $\\odot$ denotes hadamard product used to obtain the new weighted tensor $\\widehat{\\mathcal{B}}^{(1)}\\in \\mathbb{R}^{d\\times r_{\\max}\\times K}$ , i.e., $\\widehat{\\mathcal{B}}_{i,j,k}^{(1)} = \\widehat{\\mathcal{B}}_{i,j,k}\\cdot \\mathbb{B}_{*1,1,k}^{(1)}$ . The final outcome, $\\mathbb{B}_{*}^{(3)}\\in \\mathbb{R}^{d\\times r_{\\max}\\times 1}$ , is the aggregated module, i.e., $\\bar{B} = \\mathbb{B}_{*}^{(3)}$ . After completing aggregation on the server side, the module $\\bar{B}$ is distributed to each client for the next round of training. The T2M-strategy is also applicable to the decomposition of module A when module B is frozen for aggregation. Algorithm 1 provides the pseudo-code for Te-LoRA.", + "bbox": [ + 89, + 196, + 482, + 344 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 356, + 222, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 380, + 264, + 396 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison methods. We compare the proposed Te-LoRA with three state-of-the-art approaches that support federated fine-tuning of heterogeneous LoRA. \"Zeropadding\" [8] scales the lower ranks of heterogeneous LoRAs from clients by padding them with zeros up to the maximum rank on the server side. \"FLoRA\" [43] stacks all heterogeneous LoRAs together and then performs matrix multiplication for aggregation on the server side. \"FlexLoRA\" [3] aggregates heterogeneous LoRAs after matrix multiplication, then decomposes them into modules A and B via SVD. \"Local\" refers to the average test results from all clients independently training on their local data, serving as an important baseline to validate the effectiveness of FL. To ensure a fair comparison, we reproduced the comparison methods based on the existing codebase [45], with the same LLM, data splits, and training configurations, and conducted all the experiments.", + "bbox": [ + 89, + 401, + 482, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and evaluation. We used three datasets to train Te-LoRA: Alpaca-GPT4 [32], an English instruction-following dataset consisting of 52k samples generated by GPT-4 [1] based on Alpaca's self-learning [42]; Databricks-dolly-15k (Dolly) [46], an open-source dataset created by Databricks' staff, which includes tasks such as brainstorming, classification, closed quizzes, open quizzes, and summarization; and Wizard dataset [29] is a quiz dataset including 70k instruction-output pairs. For federated training, datasets were partitioned using a Dirichlet distribution $(\\alpha = 0.5)$ . Furthermore, we separately consider closed [7] and open evaluation benchmarks, using MMLU [13] for closed evaluation and MT-Bench [48] for open evaluation. MMLU is a widely used Q&A dataset for LLM fine-tuning, including 14,024 questions across 57 different topics to assess the logical reasoning ability of LLMs. MT-Bench is", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/676e42c707fbbd2855c66906e5387d1a3971a5c90e8816bef5e5c4a4a1d69079.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMMLUMT-Bench
HomoWizardDollyAlpaca-GPT4Wizard
Local20.8725.1338.662.92
Zero-padding [8]22.0326.5944.883.17
FLoRA [43]22.9127.7345.153.22
FlexLoRA [3]22.9728.2145.343.26
Te-LoRA (Ours)23.3528.4445.863.31
", + "bbox": [ + 513, + 88, + 915, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of Te-LoRA with other methods on MMLU and MT-bench. 'Homo' represents settings with homogeneous LoRA ranks.", + "bbox": [ + 511, + 220, + 903, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "a set of challenging multi-round open-ended questions for assessing chat assistants [48].", + "bbox": [ + 511, + 273, + 903, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. For the model configurations, we use the pre-trained Llama2-7B [41] as the base model and conduct 200 rounds of FL communication. The initial learning rate is set to 5e-5 and decays to 1e-6 after 200 rounds using a cosine scheduler. In each round, two clients are randomly selected, each training for 10 epochs with a batch size of 16. Referring to FLoRA's setup [43] in Homogeneous LoRA, the rank of LoRA is 16. Referring to FLoRA's simulation of a real-world scenario where clients have heterogeneous computational resources, in Heterogeneous LoRA, we assign [64, 32, 16, 16, 8, 8, 4, 4, 4, 4] as different local LoRA ranks applied to 10 clients. All experiments are conducted on a single NVIDIA 3090 GPU.", + "bbox": [ + 511, + 304, + 905, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison Experiment", + "text_level": 1, + "bbox": [ + 511, + 513, + 740, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Homogeneous LoRA. In the homogeneous LoRA setup, we configure all clients to share the same LoRA rank of 16. As shown in Tab. 1, we draw the following conclusions: 1) All four federated fine-tuning LoRA methods outperform \"Local\" with Te-LoRA achieving a 7.2 improvement in the MMLU benchmark compared to \"Local\" under AlpacaGPT4 training. 2) Among the comparison methods, Zeropadding, which is affected by aggregation noise, performs the worst. Notably, in homogeneous LoRA, Zero-padding degrades to a traditional approach where modules B and A are aggregated separately on the server. 3) FlexLoRA outperforms FLoRA. Both aggregate the matrix-multiplied weights from modules B and A, but FlexLoRA uses SVD to decompose the aggregated weights and sends new modules to clients, while FLoRA merges the aggregated weights with the client LLM and reinitialized LoRA training. The reinitialized LoRA may cause instability in client training. 4) Te-LoRA outperforms all methods across tasks by effectively resolving aggregation noise and misalignment.", + "bbox": [ + 511, + 537, + 905, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Heterogeneous LoRA. As described in Sec. 4.1, to simulate the scenario of resource heterogeneity among clients in real-world settings, we apply different local LoRA ranks of [64, 32, 16, 16, 8, 8, 4, 4, 4, 4] to 10 clients. By comparing the results of Heterogeneous LoRA (Heter) in Tab. 2", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1063", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5346bed9902febd3d97e417e549602e86b5418b18b2e998e973ceb78fa5502b5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method HeterWizardMMLU DollyAlpaca-GPT4MT-Bench Wizard
Local20.5325.1038.542.89
Zero-padding [8]21.8626.4844.473.15
FLoRA [43]22.7927.6044.943.20
FlexLoRA [3]23.4928.1845.483.28
Te-LoRA (Ours)23.7128.3746.163.33
", + "bbox": [ + 91, + 88, + 493, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of Te-LoRA with other methods on MMLU and MT-bench. 'Heter' represents settings with heterogeneous LoRA ranks.", + "bbox": [ + 89, + 220, + 485, + 262 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a882a4bbec1565049a862cb21dff494fc6d5d29ebbd130dabdce99984e790f1b.jpg", + "image_caption": [ + "Figure 3. Comparison of the communication cost. Te-LoRA achieves lower overhead compared to the other methods." + ], + "image_footnote": [], + "bbox": [ + 150, + 263, + 433, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with Homogeneous LoRA (Homo) in Tab. 1, we make several interesting observations. First, both Zero-padding and FLoRA show a performance drop in the Heter configuration compared to Homo, with Zero-padding decreasing by up to 0.41 and FLoRA by up to 0.21. Since FLoRA mitigates aggregation noise, it still outperforms Zero-padding. Second, Te-LoRA and FlexLoRA under Heter outperform Homo in most evaluation tasks, demonstrating their ability to effectively handle resource heterogeneity and optimize resource utilization. Because Te-LoRA more effectively addresses the aggregation misalignment issue, it still maintains the best performance under Heter. Finally, Te-LoRA and FlexLoRA show slight performance drops when transitioning from Homo to Heter on the Dolly dataset. This may be due to Dolly's smaller dataset (15k) compared to Alpaca-GPT4 (52k) and Wizard (70k). In the Heter configuration, the 64 and 32 rank LoRA models may overfit on clients with less data, leading to poorer aggregation results.", + "bbox": [ + 91, + 491, + 483, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Communication overhead. Communication cost plays a crucial role in real-world FL. Therefore, we compared the total upload and download communication costs in a single round across different methods. To ensure a fair comparison, we normalized the transmission parameters. As shown in Fig. 3, the proposed Te-LoRA has the lowest communication cost, as it uploads only module A or B per round. In contrast, FlexLoRA and Zero-padding require both modules to be uploaded and downloaded each round, while", + "bbox": [ + 89, + 765, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b59d3aeb35bfebc475f63bdc06fa2eb670132a07db7784e43ed4da720fa6d8db.jpg", + "image_caption": [ + "Figure 4. Impact of different ranks (64, 32, 16, 8, and 4)." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "FLoRA incurs the highest cost due to transmitting the result of multiplying matrices B and A during download. The quantitative comparison highlights the advantage of the alternating freezing technique.", + "bbox": [ + 511, + 263, + 903, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of rank. To further investigate the impact of different rank configurations on homogeneous-rank clients, we conducted experiments on the \"Local\" baseline with five rank settings (64, 32, 16, 8, and 4). Training was performed on the Wizard and Alpaca-GPT4 datasets, and evaluation was conducted using MT-Bench and MMLU, respectively. As illustrated in Fig. 4, setting the rank to 4 for all clients, which aligns with the most resource-constrained ones, leads to a significant performance decline. While higher ranks can improve performance, they are not feasible for resource-limited clients. Therefore, adopting a heterogeneous rank configuration enables better resource utilization and further enhances performance.", + "bbox": [ + 511, + 324, + 906, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Experiment", + "text_level": 1, + "bbox": [ + 511, + 530, + 714, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the PAA-strategy and T2M-strategy in Te-LoRA, we conducted several ablation experiments. Specifically, \"w/o PPA\" refers to replacing the PAA-strategy with zero-padding in the heterogeneous experiment, which combines the alternating freezing method with zero-padding. \"w/o T2M\" indicates replacing the T2M-strategy with the weighted average aggregation of FedAvg in the heterogeneous experiments.", + "bbox": [ + 511, + 551, + 905, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effects of PAA-strategy. To compare with the alignment effect of zero-padding and PAA-strategy, we visualized the cosine similarity between clients after alignment using zero-padding and PAA-strategy, as well as the change curves of the singular value distribution before and after alignment. As shown in Fig. 5, we visualized the cosine similarity between four clients, where most clients exhibit higher cosine similarity after alignment with the PAA strategy compared to zero-padding. This indicates that the PAA-strategy not only aligns the parameters between heterogeneous LoRAs on the server side but also reduces the information loss caused by the alignment. To better illustrate the differences before and after alignment, we compared the normalized singular value distribution curves, as shown in Fig. 6. It can be observed that the distribution of singular", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1064", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ef770263f10047c1d54633ddc3902e1bfcada4c1c54588a4331c4793d97191e.jpg", + "image_caption": [ + "(a) On PAA-strategy." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 272, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/222b1392e865779e914184d839452b0446077c1a3ffe6a26d3fdf2b96175725d.jpg", + "image_caption": [ + "(b) On Zero-padding." + ], + "image_footnote": [], + "bbox": [ + 303, + 88, + 482, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/916fe394f873563fc333a4c7b2bb75581782655c9a2f9d57f14f88297f76a713.jpg", + "image_caption": [ + "(a) In round four." + ], + "image_footnote": [], + "bbox": [ + 91, + 294, + 302, + 435 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2418d0de271994115dcbbffb36daed19cd0e7533d00dd048817d9f410097873.jpg", + "image_caption": [ + "Figure 5. Cosine similarity visualization of the local model processed by PAA-strategy or Zero-padding with the global model.", + "(b) In round seven.", + "Figure 6. Singular value distribution curves. The source LoRA and those transformed by zero-padding and the PAA strategy undergo SVD to plot the singular value distribution curves." + ], + "image_footnote": [], + "bbox": [ + 305, + 292, + 480, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "values after applying the PAA strategy is closer to that of the original LoRA, while the distribution after zero-padding is relatively farther away. This indicates that the PAA strategy retains more of the major components to a certain extent and is closer to the original LoRA compared to zero-padding.", + "bbox": [ + 89, + 517, + 483, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further validate the effect of PPA quantitatively, we conducted an ablation study, as shown in Tab. 3. In all scenarios, \"w/o PAA\" performs worse than $\\mathrm{T2M}_2$ , with a performance drop of at least 1.3 points in each case. This validates that the PAA-strategy, compared to Zero-padding, can more effectively reduce information loss when aligning parameters between different levels of the A and B modules. Meanwhile, compared to the FlexLoRA in Tab. 2, \"w/o T2M\" performs better in both the Dolly and AlpacaGPT4 scenarios. This suggests that combining the PAA-strategy and alternating freezing strategy avoids aggregation noise while achieving parameter alignment between different levels of modules A and B, reducing information loss.", + "bbox": [ + 89, + 595, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effects of T2M-strategy. To evaluate the impact of this strategy, as shown in Tab. 3, we performed an ablation experiment with \"w/o T2M\". First, \"w/o T2M\" performs worse in all three scenarios compared to the $\\mathrm{T2M}_2$ method, especially in the Alpaca-GPT4 dataset, where w/o T2M is about 0.63 points lower than $\\mathrm{T2M}_2$ . This validates that the T2M-strategy, compared to the weighted average", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5932433fc0182a4a9a1ef5aecc629e483174ad12ae695d342c39b8fd124281a7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method HeterWizardMMLU DollyAlpaca-GPT4
w/o PAA22.4027.1044.77
w/o T2M23.3828.2145.53
T2M122.9128.1445.30
T2M223.7128.3746.16
T2M323.6728.3045.99
", + "bbox": [ + 514, + 88, + 903, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Validation of Te-LoRA components. 'Heter' represents settings with heterogeneous LoRA ranks.", + "bbox": [ + 511, + 220, + 903, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "aggregation operation of FedAvg, better preserves the key information from each client, enabling finer feature alignment. In comparison to Zero-padding in Tab. 2, \"w/o PAA\" achieves higher MMLU scores for all three datasets: 0.54, 0.38, and 0.3 points higher on Wizard, Dolly, and Alpaca-GPT4, respectively. This indicates that the T2M-strategy combined with alternating freezing more effectively aligns features while avoiding aggregation noise.", + "bbox": [ + 511, + 256, + 903, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effects of $\\mathrm{T2M_1}$ , $\\mathrm{T2M_2}$ , and $\\mathrm{T2M_3}$ . As observed from Tab. 3, $\\mathrm{T2M_1}$ performs slightly worse than $\\mathrm{T2M_2}$ . In the Wizard, Dolly, and Alpaca-GPT4 datasets, $\\mathrm{T2M_2}$ scores 0.2, 0.23, and 0.86 points higher than $\\mathrm{T2M_1}$ , respectively. While $\\mathrm{T2M_3}$ outperforms $\\mathrm{T2M_1}$ on all three datasets, it still trails $\\mathrm{T2M_2}$ , though the results are nearly identical. For example, in the Wizard, Dolly, and Alpaca-GPT4 datasets, $\\mathrm{T2M_2}$ scores 0.004, 0.007, and 0.005 points higher than $\\mathrm{T2M_3}$ . This suggests that although the tensor $\\mathbb{B}_*^1 \\in \\mathbb{R}^{1 \\times 1 \\times k}$ obtained from the compression of $\\widehat{\\mathcal{B}} \\in \\mathbb{R}^{d \\times r_{\\max} \\times k}$ can preserve the proportion of each client, it is slightly less effective in retaining client information. On the other hand, the tensor $\\mathbb{B}_* \\in \\mathbb{R}^{d \\times r_{\\max} \\times 1}$ obtained from the compression of $\\widehat{\\mathcal{B}}$ better preserves the key information from each client and does not introduce redundancy in information like $\\mathbb{B}_*^3 \\in \\mathbb{R}^{d \\times r_{\\max} \\times 1}$ does.", + "bbox": [ + 511, + 377, + 906, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel Tensor-aggregation LoRA method (Te-LoRA) to address the noise and misalignment issues in the aggregation of low-rank adaptation within federated learning under resource heterogeneity. Building upon an alternating freezing strategy, Te-LoRA incorporates Pre-Aggregation Alignment (PAA) and Tensor Decomposition Aggregation (T2M) strategies. The PAA strategy utilises OT technology to align the weights of heterogeneous LoRA modules A and B from different clients into homogeneous weights, while the T2M-strategy adaptively aggregates key components to preserve interdependencies. Experimental results demonstrate that Te-LoRA effectively reduces communication costs and significantly improves fine-tuning performance, offering a viable solution for efficient federated fine-tuning in resource-constrained environments.", + "bbox": [ + 511, + 659, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1065", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. The work is supported by the National Natural Science Foundation of China (Grant No. 62222207, 62427808, 62472208, 62276134).", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 148, + 187, + 165 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.6", + "[2] Sara Babakniya, Ahmed Roushdy Elkordy, Yahya H Ezzeldin, Qingfeng Liu, Kee-Bong Song, Mostafa El-Khamy, and Salman Avestimehr. Slora: Federated parameter efficient fine-tuning of language models. arXiv preprint arXiv:2308.06522, 2023.1, 3", + "[3] Jiamu Bai, Daoyuan Chen, Bingchen Qian, Liuyi Yao, and Yaliang Li. Federated fine-tuning of large language models under heterogeneous language tasks and client resources. arXiv e-prints, pages arXiv-2402, 2024. 2, 3, 6, 7", + "[4] Desirée Bill and Theodor Eriksson. Fine-tuning a llm using reinforcement learning from human feedback for a therapy chatbot application, 2023. 1", + "[5] Daoyuan Chen, Yilun Huang, Zhijian Ma, Hesen Chen, Xuchen Pan, Ce Ge, Dawei Gao, Yuexiang Xie, Zhaoyang Liu, Jinyang Gao, et al. Data-juicer: A one-stop data processing system for large language models. In Companion of the 2024 International Conference on Management of Data, pages 120-134, 2024. 1", + "[6] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. Longlora: Efficient finetuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 2", + "[7] Yew Ken Chia, Pengfei Hong, Lidong Bing, and Soujanya Poria. Instructeval: Towards holistic evaluation of instruction-tuned large language models. arXiv preprint arXiv:2306.04757, 2023. 6", + "[8] Yae Jee Cho, Luyang Liu, Zheng Xu, Aldi Fahrezi, and Gauri Joshi. Heterogeneous lora for federated fine-tuning of ondevice foundation models. arXiv preprint arXiv:2401.06432, 2024. 3, 6, 7", + "[9] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023. 1", + "[10] Sannara Ek, Kaile Wang, François Portet, Philippe Lalanda, and Jiannong Cao. Fedali: Personalized federated learning with aligned prototypes through optimal transport. arXiv preprint arXiv:2411.10595, 2024. 5", + "[11] Junyao Gao, Yanan Sun, Fei Shen, Xin Jiang, Zhening Xing, Kai Chen, and Cairong Zhao. Faceshot: Bring any character into life. arXiv preprint arXiv:2503.00740, 2025. 1", + "[12] Lucas Grativol, Mathieu Leonardon, Guillaume Muller, Virginie Fresse, and Matthieu Arzel. Flocora: Federated learning compression with low-rank adaptation. In 2024 32nd European Signal Processing Conference (EUSIPCO), pages 1786-1790. IEEE, 2024. 2" + ], + "bbox": [ + 93, + 172, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 6", + "[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2", + "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3", + "[16] Ahmed Imteaj, Urmish Thakker, Shiqiang Wang, Jian Li, and M Hadi Amini. A survey on federated learning for resource-constrained IoT devices. IEEE Internet of Things Journal, 9(1):1-24, 2021. 2", + "[17] Xin Jiang, Hao Tang, Junyao Gao, Xiaoyu Du, Shengfeng He, and Zechao Li. Delving into multimodal prompting for fine-grained visual classification. In Proceedings of the AAAI conference on artificial intelligence, pages 2570-2578, 2024. 1", + "[18] Xin Jiang, Hao Tang, and Zechao Li. Global meets local: Dual activation hashing network for large-scale fine-grained image retrieval. IEEE Transactions on Knowledge and Data Engineering, 2024. 1", + "[19] Xin Jiang, Hao Tang, Rui Yan, Jinhui Tang, and Zechao Li. Dvf: Advancing robust and accurate fine-grained image retrieval with retrieval guidelines. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2379-2388, 2024. 1", + "[20] Dominique Kelly, Yimin Chen, Sarah E Cornwell, Nicole S Delellis, Alex Mayhew, Sodiq Onaolapo, and Victoria L Rubin. Bing chat: The future of search engines? Proceedings of the Association for Information Science and Technology, 60(1):1007-1009, 2023. 1", + "[21] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51(3):455-500, 2009. 5", + "[22] Jabin Koo, Minwoo Jang, and Jungseul Ok. Towards robust and efficient federated low-rank adaptation with heterogeneous clients. arXiv preprint arXiv:2410.22815, 2024. 2, 3, 4", + "[23] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 2", + "[24] Hongxia Li, Wei Huang, Jingya Wang, and Ye Shi. Global and local prompts cooperation via optimal transport for federated learning. In CVPR, pages 12151-12161, 2024. 5", + "[25] Pengpeng Li, Xiangbo Shu, Chun-Mei Feng, Yifei Feng, Wangmeng Zuo, and Jinhui Tang. Surgical video workflow analysis via visual-language learning. npj Health Systems, 2 (1):5, 2025. 1", + "[26] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of fedavg on non-iid data. arXiv preprint arXiv:1907.02189, 2019. 3", + "[27] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1066", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Zheng Lin, Xuanjie Hu, Yuxin Zhang, Zhe Chen, Zihan Fang, Xianhao Chen, Ang Li, Praneeth Vepakomma, and Yue Gao. Splitlora: A split parameter-efficient fin-tuning framework for large language models. arXiv preprint arXiv:2407.00952, 2024. 2", + "[29] Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583, 2023. 6", + "[30] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 4", + "[31] Edouard Pauwels and Samuel Vaiter. The derivatives of sinkhorn-knopp converge. SIAM Journal on Optimization, 33(3):1494-1517, 2023. 5", + "[32] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 6", + "[33] Zhen Qin, Daoyuan Chen, Bingchen Qian, Bolin Ding, Yaliang Li, and Shuiguang Deng. Federated full-parameter tuning of billion-sized language models with communication cost under 18 kilobytes. arXiv preprint arXiv:2312.06353, 2023. 1", + "[34] Riccardo Salami, Pietro Buzzega, Matteo Mosconi, Jacopo Bonato, Luigi Sabetta, and Simone Calderara. Closed-form merging of parameter-efficient modules for federated continual learning. arXiv preprint arXiv:2410.17961, 2024. 2", + "[35] Xiangbo Shu, Jinhui Tang, Guo-Jun Qi, Wei Liu, and Jian Yang. Hierarchical long short-term concurrent memory for human interaction recognition. IEEE transactions on pattern analysis and machine intelligence, 43(3):1110-1118, 2019. 2", + "[36] Xiangbo Shu, Liyan Zhang, Guo-Jun Qi, Wei Liu, and Jinhui Tang. Spatiotemporal co-attention recurrent neural networks for human-skeleton motion prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3300-3315, 2021. 1", + "[37] Karan Singhal, Shekoofeh Azizi, Tao Tu, S Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, et al. Large language models encode clinical knowledge. Nature, 620 (7972):172-180, 2023. 1", + "[38] Tianyu Song, Shumin Fan, Pengpeng Li, Jiyu Jin, Guiyue Jin, and Lei Fan. Learning an effective transformer for remote sensing satellite image dehazing. IEEE Geoscience and Remote Sensing Letters, 20:1-5, 2023. 1", + "[39] Youbang Sun, Zitao Li, Yaliang Li, and Bolin Ding. Improving lora in privacy-preserving federated learning. CoRR, 2024. 3", + "[40] Jinhui Tang, Xiangbo Shu, Rui Yan, and Liyan Zhang. Coherence constrained graph listm for group activity recognition. IEEE transactions on pattern analysis and machine intelligence, 44(2):636-647, 2019. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 6", + "[42] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560, 2022. 6", + "[43] Ziyao Wang, Zheyu Shen, Yexiao He, Guoheng Sun, Hongyi Wang, Lingjuan Lyu, and Ang Li. Flora: Federated fin-tuning large language models with heterogeneous low-rank adaptations. arXiv preprint arXiv:2409.05976, 2024. 2, 3, 6, 7", + "[44] Feijie Wu, Zitao Li, Yaliang Li, Bolin Ding, and Jing Gao. Fedbiot: Llm local fine-tuning in federated learning without full model. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3345-3355, 2024. 2", + "[45] Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. Openfedllm: Training large language models on decentralized private data via federated learning. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 6137-6147, 2024. 6", + "[46] Jianyi Zhang, Saeed Vahidian, Martin Kuo, Chunyuan Li, Ruiyi Zhang, Tong Yu, Guoyin Wang, and Yiran Chen. Towards building the federatedgpt: Federated instruction tuning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6915-6919. IEEE, 2024. 1, 2, 4, 6", + "[47] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adalora: Adaptive budget allocation for parameter-efficient fine-tuning. arXiv preprint arXiv:2303.10512, 2023. 2", + "[48] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6" + ], + "bbox": [ + 514, + 90, + 903, + 667 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1067", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Tensor-aggregated LoRA in Federated Fine-tuning/cfdeda6d-a34d-4c06-a268-6da1c14a07ae_model.json b/2025/Tensor-aggregated LoRA in Federated Fine-tuning/cfdeda6d-a34d-4c06-a268-6da1c14a07ae_model.json new file mode 100644 index 0000000000000000000000000000000000000000..29514319ce1301f1b2750c295d0a66d8ad3efa38 --- /dev/null +++ b/2025/Tensor-aggregated LoRA in Federated Fine-tuning/cfdeda6d-a34d-4c06-a268-6da1c14a07ae_model.json @@ -0,0 +1,2244 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.244, + 0.131, + 0.757, + 0.154 + ], + "angle": 0, + "content": "Tensor-aggregated LoRA in Federated Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.907, + 0.253 + ], + "angle": 0, + "content": "Zhixuan Li\\(^{1*}\\), Binqian Xu\\(^{1*}\\), Xiangbo Shu\\(^{1\\dagger}\\), Jiachao Zhang\\(^{2}\\), Yazhou Yao\\(^{1}\\), Guo-Sen Xie\\(^{1}\\), Jinhui Tang\\(^{3}\\) \n\\(^{1}\\)Nanjing University of Science and Technology \n\\(^{2}\\)Nanjing Institute of Technology \n\\(^{3}\\)Nanjing Forestry University" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.255, + 0.895, + 0.287 + ], + "angle": 0, + "content": "lzxlizhixuan@njust.edu.cn, xubinq11@gmail.com, shuxb@njust.edu.cn, zhangjc07@foxmail.com, yazhou.yao@njust.edu.cn, gsxiehm@gmail.com, tangjh@njfu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.485, + 0.687 + ], + "angle": 0, + "content": "The combination of Large Language Models (LLMs) and Federated Learning (FL) to leverage privacy-preserving data has emerged as a promising approach to further enhance the Parameter-Efficient Fine-Tuning (PEFT) capabilities of LLMs. In real-world FL settings with resource heterogeneity, the training process of Low-Rank Adaptation (LoRA), the representative PEFT method, still faces two major challenges: aggregation noise and aggregation misalignment. In this paper, we propose a novel Tensor-aggregated LoRA (Te-LoRA) in Federated Finetuning based on an alternating-freeze training strategy to avoid aggregating noise without additional server-side computational costs, while mitigating aggregation suboptimality caused by parameter misalignment between heterogeneous LoRAs. Especially in addressing the aggregation suboptimality issue, we design the Pre-Aggregation Alignment strategy (PAA-strategy) and Tensor-to-Matrix strategy (T2M-strategy) for aligning heterogeneous LoRAs and aggregating them into an united tensor, which is then decomposed into matrices adapted for client download. Extensive experiments demonstrate the effectiveness and robustness of Te-LoRA in both homogeneous and heterogeneous settings." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.732, + 0.222, + 0.748 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Large Language Models (LLMs) have demonstrated exceptional performance across various task scenarios, such as search engines [20], chatbots [4], healthcare [37], and more. When using pre-trained LLMs to adapt to downstream tasks, significant computational resources are required to fine-tune the model parameters. Therefore, to improve the applicability of LLMs, it is necessary to fine-tune specific" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.321, + 0.913, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.548, + 0.908, + 0.605 + ], + "angle": 0, + "content": "Figure 1. Comparison of different federated fine-tuning LoRAs. Compared to methods (a), (b), and (c), our Te-LoRA performs better in addressing aggregation noise, communication cost, and aggregation misalignment issues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.808 + ], + "angle": 0, + "content": "datasets [9, 17, 25]. However, collecting such datasets comes with associated costs and privacy concerns [5, 18, 38]. Consequently, researchers have turned to Federated Learning (FL) as a technical approach to fine-tune LLMs using distributed client data without compromising data privacy [2, 30, 33]. In this context, researchers have explored the application of Parameter-Efficient Fine-Tuning (PEFT) [11, 19, 23] methods for federated fine-tuning of LLMs, saving substantial computational resources. Among these, Low-Rank Adaptation (LoRA) [14] has gained attention due to its significant reduction in the number of communication parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "However, simply applying LoRA in real FL environment presents several challenges, one of the key issues being aggregation noise. In detail, traditional federated fine-tuning of LLMs [36, 46] averages modules A and B of local LoRAs separately, leading to discrepancies between the aggregated and ideal global LoRAs, thereby introducing noise into the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "*Equal contributions" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.234, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.234, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1058" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "update process. To address this issue, as shown in Fig. 1, the two main solutions are as follows: 1) Merging modules A and B [3, 43], where A and B are first merged by matrix multiplication and then aggregated. 2) Alternating the training of modules A and B [22, 34], where either A or B is kept frozen for consistency at each round, with only the learning B or A being individually aggregated. Compared to the \"Merging\" strategy, the \"Alternating\" strategy reduces communication costs by nearly half, as it transmits only A or B per round." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.483, + 0.456 + ], + "angle": 0, + "content": "The existing \"Alternating\" method [22] further reduces communication costs by selecting weights from a global LoRA with a unified initial rank. However, this homogenized assumption contradicts the core feature of resource heterogeneity in FL [16, 35, 40], where real-world scenarios involve clients with local LoRAs of different initial ranks. When LoRA modules A or B with heterogeneous ranks are uploaded and aggregated, an explicit rank dimension misalignment arises. Even when adopting the \"merging\" strategy for weight aggregation through matrix multiplication, semantically driven feature alignment issues may still persist. This multi-level parameter/feature mismatch caused by rank heterogeneity and data heterogeneity can be referred to as the aggregation misalignment problem." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.699 + ], + "angle": 0, + "content": "To simultaneously mitigate the issues of aggregation noise and aggregation misalignment in heterogeneous settings, as shown in Fig. 1, we propose a novel Tensor-aggregated LoRA called Te-LoRA. Te-LoRA consists of two main strategies: First, due to the impact of heterogeneity, the ranks of the matrices uploaded by each client to the server are often inconsistent. We present a Pre-Aggregation Alignment strategy (PAA-strategy) to align rank while minimizing information loss or conflict. Second, to preserve the potential relationships and dependencies between different clients, we design the key Tensor-to-Matrix strategy (T2M-strategy) for decomposing the tensor stacked from aligned matrices. T2M-strategy can adapt to the aggregation of heterogeneous LoRAs while effectively preserving the key components from each client. The main contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.81 + ], + "angle": 0, + "content": "- Alternating-freeze training for Tensor version of LoRA. We propose a novel Tensor-aggregated LoRA (Te-LoRA) in Federated Fine-tuning to efficiently aggregate the main components of each client based on an alternating-freeze training approach, avoiding aggregation noise while alleviating aggregation suboptimality among heterogeneous LoRAs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.482, + 0.884 + ], + "angle": 0, + "content": "- Alignment of heterogeneous LoRAs for server aggregation. We design a Pre-Aggregation Alignment strategy (PAA-strategy) to align heterogeneous LoRAs, while preserving the unique features of each client and minimizing general information loss." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- Decomposition of aggregated LoRAs for client down" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "load. We devise a Tensor-to-Matrix strategy (T2M-strategy) that decomposes the tensor into matrices adapted for client download, after aggregating LoRAs into an united tensor covering relationships and dependencies among LoRAs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.184, + 0.655, + 0.199 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.209, + 0.802, + 0.226 + ], + "angle": 0, + "content": "2.1. Parameter-Efficient Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.232, + 0.905, + 0.626 + ], + "angle": 0, + "content": "To effectively fine-tune LLMs while reducing computational resources and storage requirements, Parameter-Efficient Fine-Tuning (PEFT) techniques have emerged. Among these, Prefix-Tuning [27] guides the pre-trained language model's generation process by optimizing a continuous sequence of task-specific vectors (called prefixes) while keeping the model parameters frozen. Soft prompts [23] enable frozen language models to perform specific tasks by propagating signals of labeled examples. Both methods are effective for task-specific adaptation. LoRA [14] reduces memory overhead by representing weight updates with low-rank matrices. AdaLoRA [47] optimizes incremental updates of pre-trained language models by adaptively allocating parameter budgets and using singular value decomposition, effectively reducing the parameter budget and avoiding intensive computation. LongLoRA [6] extends the context length of LLMs at low computational cost through sparse local attention and parameter-efficient fine-tuning. SplitLoRA [28], based on Split Federated Learning (SFL), improves training efficiency by partitioning the model and combining the parallel training of federated learning with the model splitting advantage of split learning. PEFT-based methods for LLMs, especially LoRA and its variants, have advanced fine-tuning techniques, showing excellent performance due to their simplicity and ability to reduce parameters. Thus, we fine-tune LLMs with LoRA in this paper." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.637, + 0.89, + 0.653 + ], + "angle": 0, + "content": "2.2. Federated Fine-Tuning Homogeneous LoRA" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In federated learning, several works have applied LoRA to achieve reduced communication and improved efficiency. For example, Zhang et al. [46] fine-tuned LLMs by aggregating heterogeneous instruction data from multiple clients in a federated learning framework, improving the model's generalization ability on new tasks while ensuring user privacy and data security. Wu et al. [44] introduced the FedBiOT algorithm, which compresses LLMs and splits them into two components, using a dual-layer optimization approach to achieve efficient fine-tuning and reduce computational and communication costs. FLoCoRA [12] also reduces communication overhead by freezing the originally randomized neural network parameters during training, and only training and transmitting the added LoRA adapters. However, it does not fully address the stability issues brought about by extremely low-bit quantization. Sun" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1059" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.094, + 0.93, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.317, + 0.908, + 0.388 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed Te-LoRA. It is based on a framework for alternating training of the LoRA modules B and A. When the clients' LoRA have different ranks, the B or A modules are obtained after client training and subsequently uploaded to the server for aggregation. On the server, the heterogeneous B or A modules are first aligned using the PAA-strategy. In the PAA-strategy, \\(\\mu\\) and \\(\\nu\\) are all 1-column vectors of size having \\(r_{\\mathrm{max}}\\) and \\(r_k\\), respectively. Then, the aligned B or A are aggregated using the T2M-strategy. The detailed operations will be discussed in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.407, + 0.482, + 0.528 + ], + "angle": 0, + "content": "et al. [39] pointed out that separately integrating the two LoRA matrices does not fully approximate the original, and proposed FFA-LoRA, which fixes the randomly initialized matrices while fine-tuning only the zero-initialized ones. Compared to FFA-LoRA, the proposed Te-LoRA, based on an alternating freezing framework, not only preserves the training of all LoRA modules but also effectively addresses the issue of aggregation noise." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.544, + 0.467, + 0.56 + ], + "angle": 0, + "content": "2.3. Federated Fine-tuning Heterogeneous LoRA" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.57, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Due to non-IID data heterogeneity [26] sampled from the Dirichlet distribution and heterogeneous hardware resources, clients need to adjust the LoRA rank accordingly. Existing work has proposed methods to address this issue. For instance, SLoRA [2] addresses LoRA's limitations in high data heterogeneity scenarios with data-driven initialization, achieving performance similar to full finetuning while reducing training time and communication costs. FLoRA [43] uses a stacked aggregation method to eliminate noise in low-rank adapter (LoRA) aggregation, ensuring noise-free global updates. It supports heterogeneous LoRA configurations across clients, allowing dynamic adjustments based on data complexity and resources. However, this approach leads to linear growth in communication complexity, higher computational and storage demands, and reduced flexibility for supporting multiple task-specific adapters. Cho et al. [8] deploy LoRA modules with different ranks across clients, using zero-padding and truncation for aggregation, combining the benefits of high- and low-rank LoRA. However, the padded zeros dilute useful information from high-quality clients, reducing optimization efficiency and potentially limiting perfor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.907, + 0.497 + ], + "angle": 0, + "content": "mance. FlexLoRA [3] dynamically adjusts the LoRA rank for clients and uses Singular Value Decomposition (SVD) to reallocate global weights, enhancing resource utilization and model generalization. However, SVD truncation adds computational and storage overhead, and the resulting errors may limit performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.909, + 0.742 + ], + "angle": 0, + "content": "LoRA-A2 [22] alternates freezing the A and B components of the module and dynamically selects key parameters based on the importance of each client's data, effectively reducing parameter inconsistencies during the aggregation of low-rank adapters and significantly cutting communication costs. However, the method assumes that the global LoRA has a uniform initial rank, whereas in practice, each client's LoRA begins with a different rank. Compared to the related LoRA-A2, we train the initial heterogeneous-rank LoRA directly on the clients. While LoRA-A2 focuses on adaptively selecting client ranks, we emphasize the effective alignment of heterogeneous ranks on the server side. During alignment, unlike FlexLoRA and FLoRA, which use merging strategies, we designed the PAA-strategy and T2M-strategy for adaptive alignment, preserving the effective information within clients and dependencies across clients." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.76, + 0.605, + 0.775 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.655, + 0.801 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Low-Rank Adaptation. The application of Low-Rank Adaptation (LoRA) [15] facilitates a reduction in the number of trainable parameters during the fine-tuning of a pretrained model. Specifically, the weight matrix \\( W \\in \\mathbb{R}^{d \\times l} \\) in the fine-tuned model can be represented as the sum of the original weight matrix \\( W_0 \\in \\mathbb{R}^{d \\times l} \\) and an updated" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1060" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.136 + ], + "angle": 0, + "content": "weight matrix \\(\\triangle W\\in \\mathbb{R}^{d\\times l}\\). \\(\\triangle W\\) is implemented by low-rank decomposition into smaller matrices \\(B\\in \\mathbb{R}^{d\\times r}\\) and \\(A\\in \\mathbb{R}^{r\\times l}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.146, + 0.483, + 0.163 + ], + "angle": 0, + "content": "\\[\nW = W _ {0} + \\triangle W = W _ {0} + B A, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.171, + 0.483, + 0.217 + ], + "angle": 0, + "content": "where rank \\( r \\) is much smaller than both dimensions \\( d \\) and \\( l \\), thereby significantly reducing the number of trainable parameters for the weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.217, + 0.484, + 0.398 + ], + "angle": 0, + "content": "Aggregation Noise. Currently, FedAvg [30], a widely used FL algorithm, updates the global model by performing a weighted average of \\( n \\) local clients in each communication round. When LoRA is applied to FL, training and transmitting only the updated low-rank matrices B and A significantly reduces both computational and communication costs, compared to directly updating and transmitting \\( \\triangle W \\). Combining the aforementioned FedAvg and LoRA, FedIT [46] is proposed for federated fine-tuning, where the global LoRA matrices A and B are updated by weighted averaging of the local LoRA matrices \\( A_{i} \\) and \\( B_{i} \\) from \\( n \\) clients, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.405, + 0.483, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\triangle W = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\triangle W _ {i} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} B _ {i} \\times \\frac {1}{n} \\sum_ {i = 1} ^ {n} A _ {i}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.484, + 0.497 + ], + "angle": 0, + "content": "However, when each weight is averaged and aggregated separately, there is a bias between the aggregated parameters (left) and ideal (right), as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.505, + 0.483, + 0.545 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{n} \\sum_ {i = 1} ^ {n} B _ {i} \\times \\frac {1}{n} \\sum_ {i = 1} ^ {n} A _ {i} \\neq \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(B _ {i} A _ {i}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.672 + ], + "angle": 0, + "content": "One solution is to first merge \\( A_{i} \\) and \\( B_{i} \\) from each client by matrix multiplication and then perform average aggregation, i.e., \\( \\triangle W = \\frac{1}{n}\\sum_{i=1}^{n}(B_{i}A_{i}) \\). Subsequently, the aggregated \\( \\triangle W \\in \\mathbb{R}^{d\\times l} \\) must be decomposed into the new \\( A_{i} \\in \\mathbb{R}^{r\\times l} \\) and \\( B_{i} \\in \\mathbb{R}^{d\\times r} \\) for each client. This method introduces complex matrix decomposition, making post-decomposition training prone to instability. The details of another simple and effective solution are as follows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.484, + 0.751 + ], + "angle": 0, + "content": "Alternating Freezing Technique. In [22], a simple alternating freezing technique is introduced to mitigate the issue of aggregation noise when applying LoRA in FL. Specifically, in round \\( t \\), the aggregated \\( \\bar{A}^{(t)} \\) remains frozen, and each client only trains \\( B_{i}^{(t)} \\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.759, + 0.483, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{n} \\sum_ {i = 1} ^ {n} B _ {i} ^ {(t)} \\times \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\bar {A} ^ {(t)} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(B _ {i} ^ {(t)} \\bar {A} ^ {(t)}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.807, + 0.484, + 0.841 + ], + "angle": 0, + "content": "Then, in round \\( t + 1 \\), the aggregated \\( \\bar{B}^{(t + 1)} \\) is frozen, and only \\( A_{i}^{(t + 1)} \\) is trainable on each client, as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.849, + 0.484, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{n} \\sum_ {i = 1} ^ {n} \\bar {B} ^ {(t + 1)} \\times \\frac {1}{n} \\sum_ {i = 1} ^ {n} A _ {i} ^ {(t + 1)} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\bar {B} ^ {(t + 1)} A _ {i} ^ {(t + 1)}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.168 + ], + "angle": 0, + "content": "Finally, A and B complete iterative training and aggregation through the alternating freezing technique. However, when faced with data heterogeneity and rank heterogeneity at the client, this approach does not achieve good alignment of modules A or B at the server." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.176, + 0.688, + 0.192 + ], + "angle": 0, + "content": "3.2. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.197, + 0.907, + 0.499 + ], + "angle": 0, + "content": "In real FL environments with resource heterogeneity, the training of Low-Rank Adaptation (LoRA) faces two major challenges: aggregation noise and aggregation misalignment. To address these issues, we propose a novel federated fine-tuning method with Tensor-aggregated LoRA (Te-LoRA) based on an alternating freezing training strategy, as shown in Fig. 2. Te-LoRA consists of three strategies: the alternating freezing training strategy, the Pre-Aggregation Alignment Strategy (PAA-strategy), and the Tensor Decomposition Aggregation Strategy (T2M-strategy). Based on the existing alternating freezing training strategy (Eqs. (4) and (5)), Te-LoRA first effectively avoids aggregation noise (Eq. (3)). Subsequently, the PAA-strategy aims to achieve parameter alignment between different rank modules A and B with minimal information loss. Furthermore, the T2M-strategy adaptively aligns and aggregates the main components among clients to achieve more fine-grained feature alignment. The PAA and T2M strategies complement each other, effectively mitigating the issue of aggregation misalignment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.5, + 0.909, + 0.732 + ], + "angle": 0, + "content": "Pre-Aggregation Alignment strategy (PAA-strategy). The PAA-strategy is presented to achieve the optimal transformation from heterogeneous LoRAs to homogeneous LoRAs with minimal information loss for parameter alignment on the server side. In the implementation of the alternating freezing FL training strategy, the modules transmitted between the client and server are either B or A. Assuming that the module B is uploaded in the current round, module A remains frozen and stays on the client side. After the upload, the server receives a set of \\( K \\) modules B, i.e., \\( \\{B_k \\in \\mathbb{R}^{d \\times r_k}\\}_{k=1}^K \\). The module with the highest rank is selected as the target module \\( B_{\\max} = B_{\\arg \\max_{k=1,\\dots,K} r_k} \\), and the remaining modules are aligned towards the target module. Specifically, the distance matrix \\( \\mathcal{C} \\) between \\( B_k \\) and \\( B_{\\max} \\) is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.738, + 0.907, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} (m, n) = \\sum_ {i = 1} ^ {d} \\left(b _ {k} ^ {i m} - b _ {\\max } ^ {i n}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.789, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathcal{C} \\in \\mathbb{R}^{r_k \\times r_{\\max}}\\), \\(b_k^{im}\\) and \\(b_{\\max}^{in}\\) are the elements of the \\(i^{\\text{th}}\\) row, \\(m^{\\text{th}}\\) column of matrix \\(B_k\\), and \\(n^{\\text{th}}\\) column of matrix \\(B_{\\max}\\), respectively. To smooth the matching relationship between the source and target matrices, the matrix \\(\\mathcal{C}\\) is regularized: \\(\\mathcal{K} = \\exp\\left(-\\frac{\\mathcal{C}}{\\eta}\\right)\\), where \\(\\eta\\) is the regularization parameter. Next, based on the matrix \\(\\mathcal{K}\\), we first assume that the source distribution \\(p\\) and the target distribution \\(q\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1061" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.089, + 0.484, + 0.202 + ], + "angle": 0, + "content": "follow a uniform distribution, i.e., \\( p = \\left[\\frac{1}{r_k},\\frac{1}{r_k},\\dots \\frac{1}{r_k}\\right]^\\top \\), \\( q = \\left[\\frac{1}{r_{\\mathrm{max}}},\\frac{1}{r_{\\mathrm{max}}},\\dots \\frac{1}{r_{\\mathrm{max}}}\\right]^\\top \\). Then, using the Sinkhorn-Knopp algorithm [31], we iteratively update the adjustment factors \\( \\mu \\) and \\( \\nu \\) to make the probability matrix \\( \\mathcal{G} \\) converge to a solution that satisfies the marginal constraints of \\( p \\) and \\( q \\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.21, + 0.483, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\mu = \\frac {p}{\\mathcal {K} \\cdot \\nu}, \\quad \\nu = \\frac {q}{\\mathcal {K} ^ {\\top} \\cdot \\mu}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.484, + 0.323 + ], + "angle": 0, + "content": "where the adjustment factors \\(\\mu\\) and \\(\\nu\\) are both 1-column vectors of size \\(r_{\\mathrm{max}}\\) and \\(r_k\\), respectively. By iteratively updating until convergence, the probability matrix \\(\\mathcal{G}\\), representing the optimal alignment between the source and target features, is shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.332, + 0.483, + 0.352 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} = \\mu \\cdot \\mathcal {K} \\cdot \\nu^ {\\top}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.484, + 0.406 + ], + "angle": 0, + "content": "Subsequently, the source matrix \\( B_{k} \\) is transformed using the matrix \\( \\mathcal{G} \\) to achieve alignment of different ranks with minimal information loss, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.415, + 0.483, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\widetilde {B _ {k}} = B _ {k} \\cdot \\mathcal {G}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.484, + 0.79 + ], + "angle": 0, + "content": "where \\(\\widetilde{B_k} \\in \\mathbb{R}^{d \\times r_{\\max}}\\) is the aligned matrix. The remaining module B and the module A from another round both use the same pre-aggregation process for parameter alignment. Discussions. Although previous research in FL, such as FedOTP [24] and FedAli [10], has explored Optimal Transport (OT) for fine-grained feature alignment, FedOTP [24] uses OT to align local visual features with global and local text features, smoothing the global consensus and local personalization. To balance personalization and generalization, FedAli [10] quantifies the cost of embedding inputs into prototypes using OT, effectively aligning embeddings and reducing client inconsistencies. Compared to these studies, our PAA strategy differs in three ways: 1) The others focus on data heterogeneity in personalized FL, while our work addresses resource heterogeneity in real FL; 2) Previous work aligns personalized with global knowledge via OT, while our work aligns heterogeneous weights into homogeneous ones using OT to address rank mismatches from resource heterogeneity; 3) Previous studies focus on aligning prompts and prototypes, while our work targets alignment methods for heterogeneous LoRAs. To the best of our knowledge, this is the first research to apply OT for heterogeneous LoRA alignment in resource-heterogeneous FL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.792, + 0.484, + 0.904 + ], + "angle": 0, + "content": "Tensor-to-Matrix strategy (T2M-strategy). To maintain the relationships and dependencies between different clients, the Tensor-to-Matrix strategy (T2M-strategy) is designed to adaptively aggregate the LoRAs aligned by the PAA-strategy, rather than simply averaging them. Specifically, given the rank-aligned set \\(\\{\\widetilde{B_k}\\}_{k = 1}^K\\) of modules B, \\(\\widehat{\\mathcal{B}}\\) can be stacked from \\(\\{\\widetilde{B_k}\\}_{k = 1}^K\\) to form a new tensor" + }, + { + "type": "title", + "bbox": [ + 0.531, + 0.093, + 0.693, + 0.108 + ], + "angle": 0, + "content": "Algorithm 1: Te-LoRA" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.113, + 0.907, + 0.144 + ], + "angle": 0, + "content": "Initialization: each client \\(k\\) initializes the LoRA \\(\\triangle W_{k} = B_{k}A_{k}\\) with \\(B_{k}\\in \\mathbb{R}^{d\\times r_{k}}\\) and \\(A_{k}\\in \\mathbb{R}^{r_{k}\\times d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.148, + 0.694, + 0.163 + ], + "angle": 0, + "content": "1. for \\( t = 1,2,\\ldots ,T \\) do:" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.167, + 0.775, + 0.183 + ], + "angle": 0, + "content": "2. Select K participants for round t" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.187, + 0.674, + 0.201 + ], + "angle": 0, + "content": "3. if \\( t\\% 2 = 1 \\) then:" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.207, + 0.8, + 0.222 + ], + "angle": 0, + "content": "4. for \\(k = 1,2,\\dots ,K\\) in parallel do:" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.224, + 0.824, + 0.242 + ], + "angle": 0, + "content": "5. \\(B_{k}^{(t + 1)} = \\mathrm{LocalTraining}\\left(B^{(t)},k\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.242, + 0.767, + 0.267 + ], + "angle": 0, + "content": "6. \\(\\widetilde{B}_k^{(t + 1)}\\gets \\mathrm{PAA}\\left(B_k^{(t + 1)}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.266, + 0.893, + 0.29 + ], + "angle": 0, + "content": "7. \\(\\mathbb{B}_{*}^{(t + 1)}\\gets \\mathrm{T2M}\\left(\\widetilde{B}_{1}^{(t + 1)},\\widetilde{B}_{2}^{(t + 1)},\\dots,\\widetilde{B}_{\\mathrm{K}}^{(t + 1)}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.289, + 0.691, + 0.306 + ], + "angle": 0, + "content": "8. \\(\\overline{A}^{(t + 1)} = A^{(t)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.313, + 0.628, + 0.326 + ], + "angle": 0, + "content": "9. end for" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.332, + 0.608, + 0.345 + ], + "angle": 0, + "content": "10. else:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.351, + 0.808, + 0.366 + ], + "angle": 0, + "content": "11. for \\( \\mathrm{k} = 1,2,\\cdots ,\\mathrm{K} \\) in parallel do:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.368, + 0.831, + 0.387 + ], + "angle": 0, + "content": "12. \\(A_{k}^{(t + 1)} = \\mathrm{LocalTraining}\\left(A^{(t)},k\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.386, + 0.774, + 0.41 + ], + "angle": 0, + "content": "13. \\(\\widetilde{A}_k^{(t + 1)}\\gets \\mathrm{PAA}\\left(A_k^{(t + 1)}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.409, + 0.899, + 0.434 + ], + "angle": 0, + "content": "14. \\(\\mathbb{A}_{*}^{(t + 1)}\\gets \\mathrm{T2M}\\left(\\widetilde{A}_{1}^{(t + 1)},\\widetilde{A}_{2}^{(t + 1)},\\dots ,\\widetilde{A}_{\\mathrm{K}}^{(t + 1)}\\right)\\)" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.433, + 0.7, + 0.45 + ], + "angle": 0, + "content": "15. \\(\\overline{B}^{(t + 1)} = B^{(t)}\\)" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.457, + 0.636, + 0.47 + ], + "angle": 0, + "content": "16. end for" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.476, + 0.612, + 0.49 + ], + "angle": 0, + "content": "17. end if" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.496, + 0.609, + 0.509 + ], + "angle": 0, + "content": "18. end for" + }, + { + "type": "list", + "bbox": [ + 0.531, + 0.148, + 0.907, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.907, + 0.636 + ], + "angle": 0, + "content": "representing the intermediate state before aggregation, i.e., \\(\\widehat{\\mathcal{B}} = \\bigcup_{k=1}^{K} \\widetilde{B_k}\\), where \\(\\widetilde{B_k} \\in \\mathbb{R}^{d \\times r_{\\max}}\\) and \\(\\widehat{\\mathcal{B}} \\in \\mathbb{R}^{d \\times r_{\\max} \\times K}\\). Here, the aggregation process can be viewed as a tensor decomposition process. Inspired by [21], three versions of the T2M-strategy (T2M\\(_1\\), T2M\\(_2\\), and T2M\\(_3\\)) is explored to decompose \\(\\widehat{\\mathcal{B}}\\) into the result \\(\\mathbb{B}_*\\). The generalized form of the decomposition is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.646, + 0.905, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {B}} = \\mathbb {B} _ {*} \\times_ {1} U \\times_ {2} V \\times_ {3} X, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.907, + 0.723 + ], + "angle": 0, + "content": "where \\(\\mathbb{B}_*\\) is a core tensor, \\(U,V\\), and \\(X\\) serve as the factor matrices, and the symbol \\(\\times_{i}\\) (\\(i = 1,2,3\\)) denotes mode-\\(i\\) tensor-matrix multiplication. \\(\\mathrm{T2M_1}\\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.733, + 0.905, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {B}} = \\mathbb {B} _ {*} ^ {(1)} \\times_ {1} U ^ {(1)} \\times_ {2} V ^ {(1)} = \\left[ \\left[ \\mathbb {B} _ {*} ^ {(1)}; U ^ {(1)}, V ^ {(1)}, \\mathbb {I} \\right] \\right], \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.771, + 0.907, + 0.869 + ], + "angle": 0, + "content": "where \\(\\mathbb{B}_{*}^{(1)}\\in \\mathbb{R}^{1\\times 1\\times K}\\), \\(U^{(1)}\\in \\mathbb{R}^{d\\times 1}\\), \\(V^{(1)}\\in \\mathbb{R}^{r_{max}\\times 1}\\) and \\(X = \\mathbb{I}\\) is the \\(K\\times K\\) identity matrix. [ ] denotes the shorthand. \\(\\mathbb{B}_{*}^{(1)}\\) represents the weight coefficient of different clients, reflecting the importance scores. And the aggregated module is \\(\\bar{B}\\), where \\(\\bar{B}_{i,j} = \\sum_{k = 1}^{K}\\hat{\\mathcal{B}}_{i,j,k}\\cdot \\mathbb{B}_{*1,1,k}^{(1)}\\). The formulae of T2M2 is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.878, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\mathcal {B}} = \\mathbb {B} _ {*} ^ {(2)} \\times_ {3} X ^ {(2)} = \\left[ \\left[ \\mathbb {B} _ {*} ^ {(2)}; \\mathbb {I}, \\mathbb {I}, X ^ {(2)} \\right] \\right], \\tag {12}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1062" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.089, + 0.482, + 0.169 + ], + "angle": 0, + "content": "where \\(\\mathbb{B}_*^{(2)}\\in \\mathbb{R}^{d\\times r_{max}\\times 1},X^{(2)}\\in \\mathbb{R}^{K\\times 1},U = \\mathbb{I}\\) and \\(V = \\mathbb{I}\\) are the \\(d\\times d\\) and \\(r_{max}\\times r_{max}\\) identity matrices. \\(\\mathbb{B}_*^{(2)}\\) is the aggregation result capturing the common and representative information, i.e., \\(\\bar{B} = \\mathbb{B}_*^{(2)}\\). T2M3 is a combined version of T2M1 and T2M2, as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.173, + 0.482, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\mathbb {B} _ {*} ^ {(3)} = \\mathrm {T} 2 \\mathrm {M} _ {2} (\\widehat {\\mathcal {B}} \\odot \\mathrm {T} 2 \\mathrm {M} _ {1} (\\widehat {\\mathcal {B}})), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.483, + 0.345 + ], + "angle": 0, + "content": "where \\(\\mathbb{B}_{*}^{(1)} = \\mathrm{T2M}_{1}(\\widehat{\\mathcal{B}})\\), \\(\\odot\\) denotes hadamard product used to obtain the new weighted tensor \\(\\widehat{\\mathcal{B}}^{(1)}\\in \\mathbb{R}^{d\\times r_{\\max}\\times K}\\), i.e., \\(\\widehat{\\mathcal{B}}_{i,j,k}^{(1)} = \\widehat{\\mathcal{B}}_{i,j,k}\\cdot \\mathbb{B}_{*1,1,k}^{(1)}\\). The final outcome, \\(\\mathbb{B}_{*}^{(3)}\\in \\mathbb{R}^{d\\times r_{\\max}\\times 1}\\), is the aggregated module, i.e., \\(\\bar{B} = \\mathbb{B}_{*}^{(3)}\\). After completing aggregation on the server side, the module \\(\\bar{B}\\) is distributed to each client for the next round of training. The T2M-strategy is also applicable to the decomposition of module A when module B is frozen for aggregation. Algorithm 1 provides the pseudo-code for Te-LoRA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.357, + 0.223, + 0.373 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.381, + 0.266, + 0.397 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.483, + 0.659 + ], + "angle": 0, + "content": "Comparison methods. We compare the proposed Te-LoRA with three state-of-the-art approaches that support federated fine-tuning of heterogeneous LoRA. \"Zeropadding\" [8] scales the lower ranks of heterogeneous LoRAs from clients by padding them with zeros up to the maximum rank on the server side. \"FLoRA\" [43] stacks all heterogeneous LoRAs together and then performs matrix multiplication for aggregation on the server side. \"FlexLoRA\" [3] aggregates heterogeneous LoRAs after matrix multiplication, then decomposes them into modules A and B via SVD. \"Local\" refers to the average test results from all clients independently training on their local data, serving as an important baseline to validate the effectiveness of FL. To ensure a fair comparison, we reproduced the comparison methods based on the existing codebase [45], with the same LLM, data splits, and training configurations, and conducted all the experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Datasets and evaluation. We used three datasets to train Te-LoRA: Alpaca-GPT4 [32], an English instruction-following dataset consisting of 52k samples generated by GPT-4 [1] based on Alpaca's self-learning [42]; Databricks-dolly-15k (Dolly) [46], an open-source dataset created by Databricks' staff, which includes tasks such as brainstorming, classification, closed quizzes, open quizzes, and summarization; and Wizard dataset [29] is a quiz dataset including 70k instruction-output pairs. For federated training, datasets were partitioned using a Dirichlet distribution \\((\\alpha = 0.5)\\). Furthermore, we separately consider closed [7] and open evaluation benchmarks, using MMLU [13] for closed evaluation and MT-Bench [48] for open evaluation. MMLU is a widely used Q&A dataset for LLM fine-tuning, including 14,024 questions across 57 different topics to assess the logical reasoning ability of LLMs. MT-Bench is" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.089, + 0.916, + 0.212 + ], + "angle": 0, + "content": "
MethodMMLUMT-Bench
HomoWizardDollyAlpaca-GPT4Wizard
Local20.8725.1338.662.92
Zero-padding [8]22.0326.5944.883.17
FLoRA [43]22.9127.7345.153.22
FlexLoRA [3]22.9728.2145.343.26
Te-LoRA (Ours)23.3528.4445.863.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.221, + 0.905, + 0.263 + ], + "angle": 0, + "content": "Table 1. Comparison of Te-LoRA with other methods on MMLU and MT-bench. 'Homo' represents settings with homogeneous LoRA ranks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.905, + 0.304 + ], + "angle": 0, + "content": "a set of challenging multi-round open-ended questions for assessing chat assistants [48]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.906, + 0.501 + ], + "angle": 0, + "content": "Training details. For the model configurations, we use the pre-trained Llama2-7B [41] as the base model and conduct 200 rounds of FL communication. The initial learning rate is set to 5e-5 and decays to 1e-6 after 200 rounds using a cosine scheduler. In each round, two clients are randomly selected, each training for 10 epochs with a batch size of 16. Referring to FLoRA's setup [43] in Homogeneous LoRA, the rank of LoRA is 16. Referring to FLoRA's simulation of a real-world scenario where clients have heterogeneous computational resources, in Heterogeneous LoRA, we assign [64, 32, 16, 16, 8, 8, 4, 4, 4, 4] as different local LoRA ranks applied to 10 clients. All experiments are conducted on a single NVIDIA 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.741, + 0.531 + ], + "angle": 0, + "content": "4.2. Comparison Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Homogeneous LoRA. In the homogeneous LoRA setup, we configure all clients to share the same LoRA rank of 16. As shown in Tab. 1, we draw the following conclusions: 1) All four federated fine-tuning LoRA methods outperform \"Local\" with Te-LoRA achieving a 7.2 improvement in the MMLU benchmark compared to \"Local\" under AlpacaGPT4 training. 2) Among the comparison methods, Zeropadding, which is affected by aggregation noise, performs the worst. Notably, in homogeneous LoRA, Zero-padding degrades to a traditional approach where modules B and A are aggregated separately on the server. 3) FlexLoRA outperforms FLoRA. Both aggregate the matrix-multiplied weights from modules B and A, but FlexLoRA uses SVD to decompose the aggregated weights and sends new modules to clients, while FLoRA merges the aggregated weights with the client LLM and reinitialized LoRA training. The reinitialized LoRA may cause instability in client training. 4) Te-LoRA outperforms all methods across tasks by effectively resolving aggregation noise and misalignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Heterogeneous LoRA. As described in Sec. 4.1, to simulate the scenario of resource heterogeneity among clients in real-world settings, we apply different local LoRA ranks of [64, 32, 16, 16, 8, 8, 4, 4, 4, 4] to 10 clients. By comparing the results of Heterogeneous LoRA (Heter) in Tab. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1063" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.495, + 0.213 + ], + "angle": 0, + "content": "
Method HeterWizardMMLU DollyAlpaca-GPT4MT-Bench Wizard
Local20.5325.1038.542.89
Zero-padding [8]21.8626.4844.473.15
FLoRA [43]22.7927.6044.943.20
FlexLoRA [3]23.4928.1845.483.28
Te-LoRA (Ours)23.7128.3746.163.33
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.221, + 0.486, + 0.263 + ], + "angle": 0, + "content": "Table 2. Comparison of Te-LoRA with other methods on MMLU and MT-bench. 'Heter' represents settings with heterogeneous LoRA ranks." + }, + { + "type": "image", + "bbox": [ + 0.151, + 0.264, + 0.434, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.453, + 0.485, + 0.483 + ], + "angle": 0, + "content": "Figure 3. Comparison of the communication cost. Te-LoRA achieves lower overhead compared to the other methods." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.492, + 0.485, + 0.765 + ], + "angle": 0, + "content": "with Homogeneous LoRA (Homo) in Tab. 1, we make several interesting observations. First, both Zero-padding and FLoRA show a performance drop in the Heter configuration compared to Homo, with Zero-padding decreasing by up to 0.41 and FLoRA by up to 0.21. Since FLoRA mitigates aggregation noise, it still outperforms Zero-padding. Second, Te-LoRA and FlexLoRA under Heter outperform Homo in most evaluation tasks, demonstrating their ability to effectively handle resource heterogeneity and optimize resource utilization. Because Te-LoRA more effectively addresses the aggregation misalignment issue, it still maintains the best performance under Heter. Finally, Te-LoRA and FlexLoRA show slight performance drops when transitioning from Homo to Heter on the Dolly dataset. This may be due to Dolly's smaller dataset (15k) compared to Alpaca-GPT4 (52k) and Wizard (70k). In the Heter configuration, the 64 and 32 rank LoRA models may overfit on clients with less data, leading to poorer aggregation results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Communication overhead. Communication cost plays a crucial role in real-world FL. Therefore, we compared the total upload and download communication costs in a single round across different methods. To ensure a fair comparison, we normalized the transmission parameters. As shown in Fig. 3, the proposed Te-LoRA has the lowest communication cost, as it uploads only module A or B per round. In contrast, FlexLoRA and Zero-padding require both modules to be uploaded and downloaded each round, while" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.241, + 0.88, + 0.256 + ], + "angle": 0, + "content": "Figure 4. Impact of different ranks (64, 32, 16, 8, and 4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.264, + 0.905, + 0.325 + ], + "angle": 0, + "content": "FLoRA incurs the highest cost due to transmitting the result of multiplying matrices B and A during download. The quantitative comparison highlights the advantage of the alternating freezing technique." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Impact of rank. To further investigate the impact of different rank configurations on homogeneous-rank clients, we conducted experiments on the \"Local\" baseline with five rank settings (64, 32, 16, 8, and 4). Training was performed on the Wizard and Alpaca-GPT4 datasets, and evaluation was conducted using MT-Bench and MMLU, respectively. As illustrated in Fig. 4, setting the rank to 4 for all clients, which aligns with the most resource-constrained ones, leads to a significant performance decline. While higher ranks can improve performance, they are not feasible for resource-limited clients. Therefore, adopting a heterogeneous rank configuration enables better resource utilization and further enhances performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.715, + 0.548 + ], + "angle": 0, + "content": "4.3. Ablation Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.674 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the PAA-strategy and T2M-strategy in Te-LoRA, we conducted several ablation experiments. Specifically, \"w/o PPA\" refers to replacing the PAA-strategy with zero-padding in the heterogeneous experiment, which combines the alternating freezing method with zero-padding. \"w/o T2M\" indicates replacing the T2M-strategy with the weighted average aggregation of FedAvg in the heterogeneous experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The effects of PAA-strategy. To compare with the alignment effect of zero-padding and PAA-strategy, we visualized the cosine similarity between clients after alignment using zero-padding and PAA-strategy, as well as the change curves of the singular value distribution before and after alignment. As shown in Fig. 5, we visualized the cosine similarity between four clients, where most clients exhibit higher cosine similarity after alignment with the PAA strategy compared to zero-padding. This indicates that the PAA-strategy not only aligns the parameters between heterogeneous LoRAs on the server side but also reduces the information loss caused by the alignment. To better illustrate the differences before and after alignment, we compared the normalized singular value distribution curves, as shown in Fig. 6. It can be observed that the distribution of singular" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1064" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.273, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.228, + 0.237, + 0.242 + ], + "angle": 0, + "content": "(a) On PAA-strategy." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.089, + 0.483, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.337, + 0.228, + 0.451, + 0.242 + ], + "angle": 0, + "content": "(b) On Zero-padding." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.253, + 0.483, + 0.281 + ], + "angle": 0, + "content": "Figure 5. Cosine similarity visualization of the local model processed by PAA-strategy or Zero-padding with the global model." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.295, + 0.303, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.44, + 0.245, + 0.451 + ], + "angle": 0, + "content": "(a) In round four." + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.294, + 0.481, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.44, + 0.445, + 0.451 + ], + "angle": 0, + "content": "(b) In round seven." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.506 + ], + "angle": 0, + "content": "Figure 6. Singular value distribution curves. The source LoRA and those transformed by zero-padding and the PAA strategy undergo SVD to plot the singular value distribution curves." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.595 + ], + "angle": 0, + "content": "values after applying the PAA strategy is closer to that of the original LoRA, while the distribution after zero-padding is relatively farther away. This indicates that the PAA strategy retains more of the major components to a certain extent and is closer to the original LoRA compared to zero-padding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.793 + ], + "angle": 0, + "content": "To further validate the effect of PPA quantitatively, we conducted an ablation study, as shown in Tab. 3. In all scenarios, \"w/o PAA\" performs worse than \\(\\mathrm{T2M}_2\\), with a performance drop of at least 1.3 points in each case. This validates that the PAA-strategy, compared to Zero-padding, can more effectively reduce information loss when aligning parameters between different levels of the A and B modules. Meanwhile, compared to the FlexLoRA in Tab. 2, \"w/o T2M\" performs better in both the Dolly and AlpacaGPT4 scenarios. This suggests that combining the PAA-strategy and alternating freezing strategy avoids aggregation noise while achieving parameter alignment between different levels of modules A and B, reducing information loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The effects of T2M-strategy. To evaluate the impact of this strategy, as shown in Tab. 3, we performed an ablation experiment with \"w/o T2M\". First, \"w/o T2M\" performs worse in all three scenarios compared to the \\( \\mathrm{T2M}_2 \\) method, especially in the Alpaca-GPT4 dataset, where w/o T2M is about 0.63 points lower than \\( \\mathrm{T2M}_2 \\). This validates that the T2M-strategy, compared to the weighted average" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.905, + 0.212 + ], + "angle": 0, + "content": "
Method HeterWizardMMLU DollyAlpaca-GPT4
w/o PAA22.4027.1044.77
w/o T2M23.3828.2145.53
T2M122.9128.1445.30
T2M223.7128.3746.16
T2M323.6728.3045.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.222, + 0.905, + 0.25 + ], + "angle": 0, + "content": "Table 3. Validation of Te-LoRA components. 'Heter' represents settings with heterogeneous LoRA ranks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.905, + 0.378 + ], + "angle": 0, + "content": "aggregation operation of FedAvg, better preserves the key information from each client, enabling finer feature alignment. In comparison to Zero-padding in Tab. 2, \"w/o PAA\" achieves higher MMLU scores for all three datasets: 0.54, 0.38, and 0.3 points higher on Wizard, Dolly, and Alpaca-GPT4, respectively. This indicates that the T2M-strategy combined with alternating freezing more effectively aligns features while avoiding aggregation noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.907, + 0.621 + ], + "angle": 0, + "content": "The effects of \\(\\mathrm{T2M_1}\\), \\(\\mathrm{T2M_2}\\), and \\(\\mathrm{T2M_3}\\). As observed from Tab. 3, \\(\\mathrm{T2M_1}\\) performs slightly worse than \\(\\mathrm{T2M_2}\\). In the Wizard, Dolly, and Alpaca-GPT4 datasets, \\(\\mathrm{T2M_2}\\) scores 0.2, 0.23, and 0.86 points higher than \\(\\mathrm{T2M_1}\\), respectively. While \\(\\mathrm{T2M_3}\\) outperforms \\(\\mathrm{T2M_1}\\) on all three datasets, it still trails \\(\\mathrm{T2M_2}\\), though the results are nearly identical. For example, in the Wizard, Dolly, and Alpaca-GPT4 datasets, \\(\\mathrm{T2M_2}\\) scores 0.004, 0.007, and 0.005 points higher than \\(\\mathrm{T2M_3}\\). This suggests that although the tensor \\(\\mathbb{B}_*^1 \\in \\mathbb{R}^{1 \\times 1 \\times k}\\) obtained from the compression of \\(\\widehat{\\mathcal{B}} \\in \\mathbb{R}^{d \\times r_{\\max} \\times k}\\) can preserve the proportion of each client, it is slightly less effective in retaining client information. On the other hand, the tensor \\(\\mathbb{B}_* \\in \\mathbb{R}^{d \\times r_{\\max} \\times 1}\\) obtained from the compression of \\(\\widehat{\\mathcal{B}}\\) better preserves the key information from each client and does not introduce redundancy in information like \\(\\mathbb{B}_*^3 \\in \\mathbb{R}^{d \\times r_{\\max} \\times 1}\\) does." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.634, + 0.65 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a novel Tensor-aggregation LoRA method (Te-LoRA) to address the noise and misalignment issues in the aggregation of low-rank adaptation within federated learning under resource heterogeneity. Building upon an alternating freezing strategy, Te-LoRA incorporates Pre-Aggregation Alignment (PAA) and Tensor Decomposition Aggregation (T2M) strategies. The PAA strategy utilises OT technology to align the weights of heterogeneous LoRA modules A and B from different clients into homogeneous weights, while the T2M-strategy adaptively aggregates key components to preserve interdependencies. Experimental results demonstrate that Te-LoRA effectively reduces communication costs and significantly improves fine-tuning performance, offering a viable solution for efficient federated fine-tuning in resource-constrained environments." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1065" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "Acknowledgements. The work is supported by the National Natural Science Foundation of China (Grant No. 62222207, 62427808, 62472208, 62276134)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.149, + 0.188, + 0.166 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.174, + 0.483, + 0.243 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.245, + 0.484, + 0.313 + ], + "angle": 0, + "content": "[2] Sara Babakniya, Ahmed Roushdy Elkordy, Yahya H Ezzeldin, Qingfeng Liu, Kee-Bong Song, Mostafa El-Khamy, and Salman Avestimehr. Slora: Federated parameter efficient fine-tuning of language models. arXiv preprint arXiv:2308.06522, 2023.1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.314, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[3] Jiamu Bai, Daoyuan Chen, Bingchen Qian, Liuyi Yao, and Yaliang Li. Federated fine-tuning of large language models under heterogeneous language tasks and client resources. arXiv e-prints, pages arXiv-2402, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.371, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[4] Desirée Bill and Theodor Eriksson. Fine-tuning a llm using reinforcement learning from human feedback for a therapy chatbot application, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.413, + 0.485, + 0.495 + ], + "angle": 0, + "content": "[5] Daoyuan Chen, Yilun Huang, Zhijian Ma, Hesen Chen, Xuchen Pan, Ce Ge, Dawei Gao, Yuexiang Xie, Zhaoyang Liu, Jinyang Gao, et al. Data-juicer: A one-stop data processing system for large language models. In Companion of the 2024 International Conference on Management of Data, pages 120-134, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[6] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. Longlora: Efficient finetuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[7] Yew Ken Chia, Pengfei Hong, Lidong Bing, and Soujanya Poria. Instructeval: Towards holistic evaluation of instruction-tuned large language models. arXiv preprint arXiv:2306.04757, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[8] Yae Jee Cho, Luyang Liu, Zheng Xu, Aldi Fahrezi, and Gauri Joshi. Heterogeneous lora for federated fine-tuning of ondevice foundation models. arXiv preprint arXiv:2401.06432, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[9] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[10] Sannara Ek, Kaile Wang, François Portet, Philippe Lalanda, and Jiannong Cao. Fedali: Personalized federated learning with aligned prototypes through optimal transport. arXiv preprint arXiv:2411.10595, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[11] Junyao Gao, Yanan Sun, Fei Shen, Xin Jiang, Zhening Xing, Kai Chen, and Cairong Zhao. Faceshot: Bring any character into life. arXiv preprint arXiv:2503.00740, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Lucas Grativol, Mathieu Leonardon, Guillaume Muller, Virginie Fresse, and Matthieu Arzel. Flocora: Federated learning compression with low-rank adaptation. In 2024 32nd European Signal Processing Conference (EUSIPCO), pages 1786-1790. IEEE, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.174, + 0.485, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[16] Ahmed Imteaj, Urmish Thakker, Shiqiang Wang, Jian Li, and M Hadi Amini. A survey on federated learning for resource-constrained IoT devices. IEEE Internet of Things Journal, 9(1):1-24, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[17] Xin Jiang, Hao Tang, Junyao Gao, Xiaoyu Du, Shengfeng He, and Zechao Li. Delving into multimodal prompting for fine-grained visual classification. In Proceedings of the AAAI conference on artificial intelligence, pages 2570-2578, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[18] Xin Jiang, Hao Tang, and Zechao Li. Global meets local: Dual activation hashing network for large-scale fine-grained image retrieval. IEEE Transactions on Knowledge and Data Engineering, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[19] Xin Jiang, Hao Tang, Rui Yan, Jinhui Tang, and Zechao Li. Dvf: Advancing robust and accurate fine-grained image retrieval with retrieval guidelines. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2379-2388, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.517, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[20] Dominique Kelly, Yimin Chen, Sarah E Cornwell, Nicole S Delellis, Alex Mayhew, Sodiq Onaolapo, and Victoria L Rubin. Bing chat: The future of search engines? Proceedings of the Association for Information Science and Technology, 60(1):1007-1009, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.588, + 0.904, + 0.616 + ], + "angle": 0, + "content": "[21] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51(3):455-500, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[22] Jabin Koo, Minwoo Jang, and Jungseul Ok. Towards robust and efficient federated low-rank adaptation with heterogeneous clients. arXiv preprint arXiv:2410.22815, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[23] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[24] Hongxia Li, Wei Huang, Jingya Wang, and Ye Shi. Global and local prompts cooperation via optimal transport for federated learning. In CVPR, pages 12151-12161, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[25] Pengpeng Li, Xiangbo Shu, Chun-Mei Feng, Yifei Feng, Wangmeng Zuo, and Jinhui Tang. Surgical video workflow analysis via visual-language learning. npj Health Systems, 2 (1):5, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[26] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of fedavg on non-iid data. arXiv preprint arXiv:1907.02189, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1066" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[28] Zheng Lin, Xuanjie Hu, Yuxin Zhang, Zhe Chen, Zihan Fang, Xianhao Chen, Ang Li, Praneeth Vepakomma, and Yue Gao. Splitlora: A split parameter-efficient fin-tuning framework for large language models. arXiv preprint arXiv:2407.00952, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[29] Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[30] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[31] Edouard Pauwels and Samuel Vaiter. The derivatives of sinkhorn-knopp converge. SIAM Journal on Optimization, 33(3):1494-1517, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[32] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[33] Zhen Qin, Daoyuan Chen, Bingchen Qian, Bolin Ding, Yaliang Li, and Shuiguang Deng. Federated full-parameter tuning of billion-sized language models with communication cost under 18 kilobytes. arXiv preprint arXiv:2312.06353, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[34] Riccardo Salami, Pietro Buzzega, Matteo Mosconi, Jacopo Bonato, Luigi Sabetta, and Simone Calderara. Closed-form merging of parameter-efficient modules for federated continual learning. arXiv preprint arXiv:2410.17961, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[35] Xiangbo Shu, Jinhui Tang, Guo-Jun Qi, Wei Liu, and Jian Yang. Hierarchical long short-term concurrent memory for human interaction recognition. IEEE transactions on pattern analysis and machine intelligence, 43(3):1110-1118, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[36] Xiangbo Shu, Liyan Zhang, Guo-Jun Qi, Wei Liu, and Jinhui Tang. Spatiotemporal co-attention recurrent neural networks for human-skeleton motion prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3300-3315, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[37] Karan Singhal, Shekoofeh Azizi, Tao Tu, S Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, et al. Large language models encode clinical knowledge. Nature, 620 (7972):172-180, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[38] Tianyu Song, Shumin Fan, Pengpeng Li, Jiyu Jin, Guiyue Jin, and Lei Fan. Learning an effective transformer for remote sensing satellite image dehazing. IEEE Geoscience and Remote Sensing Letters, 20:1-5, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[39] Youbang Sun, Zitao Li, Yaliang Li, and Bolin Ding. Improving lora in privacy-preserving federated learning. CoRR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[40] Jinhui Tang, Xiangbo Shu, Rui Yan, and Liyan Zhang. Coherence constrained graph listm for group activity recognition. IEEE transactions on pattern analysis and machine intelligence, 44(2):636-647, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[42] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.221, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[43] Ziyao Wang, Zheyu Shen, Yexiao He, Guoheng Sun, Hongyi Wang, Lingjuan Lyu, and Ang Li. Flora: Federated fin-tuning large language models with heterogeneous low-rank adaptations. arXiv preprint arXiv:2409.05976, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.291, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[44] Feijie Wu, Zitao Li, Yaliang Li, Bolin Ding, and Jing Gao. Fedbiot: Llm local fine-tuning in federated learning without full model. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3345-3355, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.361, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[45] Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. Openfedllm: Training large language models on decentralized private data via federated learning. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 6137-6147, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.446, + 0.905, + 0.528 + ], + "angle": 0, + "content": "[46] Jianyi Zhang, Saeed Vahidian, Martin Kuo, Chunyuan Li, Ruiyi Zhang, Tong Yu, Guoyin Wang, and Yiran Chen. Towards building the federatedgpt: Federated instruction tuning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6915-6919. IEEE, 2024. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.53, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[47] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adalora: Adaptive budget allocation for parameter-efficient fine-tuning. arXiv preprint arXiv:2303.10512, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.6, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[48] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.669 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1067" + } + ] +] \ No newline at end of file diff --git a/2025/Tensor-aggregated LoRA in Federated Fine-tuning/cfdeda6d-a34d-4c06-a268-6da1c14a07ae_origin.pdf b/2025/Tensor-aggregated LoRA in Federated Fine-tuning/cfdeda6d-a34d-4c06-a268-6da1c14a07ae_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..220991b44f73fdcfff2ed27899b8be0272ab736f --- /dev/null +++ b/2025/Tensor-aggregated LoRA in Federated Fine-tuning/cfdeda6d-a34d-4c06-a268-6da1c14a07ae_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f7c657733b87dc2a3ac5ecd31a3064801c3b538ee04e5da28df8318fa3e4499 +size 3301595 diff --git a/2025/Tensor-aggregated LoRA in Federated Fine-tuning/full.md b/2025/Tensor-aggregated LoRA in Federated Fine-tuning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..44c9ea2feaa293ec476064f261238aa6aab72655 --- /dev/null +++ b/2025/Tensor-aggregated LoRA in Federated Fine-tuning/full.md @@ -0,0 +1,316 @@ +# Tensor-aggregated LoRA in Federated Fine-tuning + +Zhixuan Li $^{1*}$ , Binqian Xu $^{1*}$ , Xiangbo Shu $^{1\dagger}$ , Jiachao Zhang $^{2}$ , Yazhou Yao $^{1}$ , Guo-Sen Xie $^{1}$ , Jinhui Tang $^{3}$ $^{1}$ Nanjing University of Science and Technology + $^{2}$ Nanjing Institute of Technology + $^{3}$ Nanjing Forestry University + +lzxlizhixuan@njust.edu.cn, xubinq11@gmail.com, shuxb@njust.edu.cn, zhangjc07@foxmail.com, yazhou.yao@njust.edu.cn, gsxiehm@gmail.com, tangjh@njfu.edu.cn + +# Abstract + +The combination of Large Language Models (LLMs) and Federated Learning (FL) to leverage privacy-preserving data has emerged as a promising approach to further enhance the Parameter-Efficient Fine-Tuning (PEFT) capabilities of LLMs. In real-world FL settings with resource heterogeneity, the training process of Low-Rank Adaptation (LoRA), the representative PEFT method, still faces two major challenges: aggregation noise and aggregation misalignment. In this paper, we propose a novel Tensor-aggregated LoRA (Te-LoRA) in Federated Finetuning based on an alternating-freeze training strategy to avoid aggregating noise without additional server-side computational costs, while mitigating aggregation suboptimality caused by parameter misalignment between heterogeneous LoRAs. Especially in addressing the aggregation suboptimality issue, we design the Pre-Aggregation Alignment strategy (PAA-strategy) and Tensor-to-Matrix strategy (T2M-strategy) for aligning heterogeneous LoRAs and aggregating them into an united tensor, which is then decomposed into matrices adapted for client download. Extensive experiments demonstrate the effectiveness and robustness of Te-LoRA in both homogeneous and heterogeneous settings. + +# 1. Introduction + +Large Language Models (LLMs) have demonstrated exceptional performance across various task scenarios, such as search engines [20], chatbots [4], healthcare [37], and more. When using pre-trained LLMs to adapt to downstream tasks, significant computational resources are required to fine-tune the model parameters. Therefore, to improve the applicability of LLMs, it is necessary to fine-tune specific + +![](images/54b59272811ac57b49a56cb3edf41ae5257c02136f2020640ceed37df523cae8.jpg) +Figure 1. Comparison of different federated fine-tuning LoRAs. Compared to methods (a), (b), and (c), our Te-LoRA performs better in addressing aggregation noise, communication cost, and aggregation misalignment issues. + +datasets [9, 17, 25]. However, collecting such datasets comes with associated costs and privacy concerns [5, 18, 38]. Consequently, researchers have turned to Federated Learning (FL) as a technical approach to fine-tune LLMs using distributed client data without compromising data privacy [2, 30, 33]. In this context, researchers have explored the application of Parameter-Efficient Fine-Tuning (PEFT) [11, 19, 23] methods for federated fine-tuning of LLMs, saving substantial computational resources. Among these, Low-Rank Adaptation (LoRA) [14] has gained attention due to its significant reduction in the number of communication parameters. + +However, simply applying LoRA in real FL environment presents several challenges, one of the key issues being aggregation noise. In detail, traditional federated fine-tuning of LLMs [36, 46] averages modules A and B of local LoRAs separately, leading to discrepancies between the aggregated and ideal global LoRAs, thereby introducing noise into the + +update process. To address this issue, as shown in Fig. 1, the two main solutions are as follows: 1) Merging modules A and B [3, 43], where A and B are first merged by matrix multiplication and then aggregated. 2) Alternating the training of modules A and B [22, 34], where either A or B is kept frozen for consistency at each round, with only the learning B or A being individually aggregated. Compared to the "Merging" strategy, the "Alternating" strategy reduces communication costs by nearly half, as it transmits only A or B per round. + +The existing "Alternating" method [22] further reduces communication costs by selecting weights from a global LoRA with a unified initial rank. However, this homogenized assumption contradicts the core feature of resource heterogeneity in FL [16, 35, 40], where real-world scenarios involve clients with local LoRAs of different initial ranks. When LoRA modules A or B with heterogeneous ranks are uploaded and aggregated, an explicit rank dimension misalignment arises. Even when adopting the "merging" strategy for weight aggregation through matrix multiplication, semantically driven feature alignment issues may still persist. This multi-level parameter/feature mismatch caused by rank heterogeneity and data heterogeneity can be referred to as the aggregation misalignment problem. + +To simultaneously mitigate the issues of aggregation noise and aggregation misalignment in heterogeneous settings, as shown in Fig. 1, we propose a novel Tensor-aggregated LoRA called Te-LoRA. Te-LoRA consists of two main strategies: First, due to the impact of heterogeneity, the ranks of the matrices uploaded by each client to the server are often inconsistent. We present a Pre-Aggregation Alignment strategy (PAA-strategy) to align rank while minimizing information loss or conflict. Second, to preserve the potential relationships and dependencies between different clients, we design the key Tensor-to-Matrix strategy (T2M-strategy) for decomposing the tensor stacked from aligned matrices. T2M-strategy can adapt to the aggregation of heterogeneous LoRAs while effectively preserving the key components from each client. The main contributions of this work are summarized as follows: + +- Alternating-freeze training for Tensor version of LoRA. We propose a novel Tensor-aggregated LoRA (Te-LoRA) in Federated Fine-tuning to efficiently aggregate the main components of each client based on an alternating-freeze training approach, avoiding aggregation noise while alleviating aggregation suboptimality among heterogeneous LoRAs. +- Alignment of heterogeneous LoRAs for server aggregation. We design a Pre-Aggregation Alignment strategy (PAA-strategy) to align heterogeneous LoRAs, while preserving the unique features of each client and minimizing general information loss. +- Decomposition of aggregated LoRAs for client down + +load. We devise a Tensor-to-Matrix strategy (T2M-strategy) that decomposes the tensor into matrices adapted for client download, after aggregating LoRAs into an united tensor covering relationships and dependencies among LoRAs. + +# 2. Related Work + +# 2.1. Parameter-Efficient Fine-Tuning + +To effectively fine-tune LLMs while reducing computational resources and storage requirements, Parameter-Efficient Fine-Tuning (PEFT) techniques have emerged. Among these, Prefix-Tuning [27] guides the pre-trained language model's generation process by optimizing a continuous sequence of task-specific vectors (called prefixes) while keeping the model parameters frozen. Soft prompts [23] enable frozen language models to perform specific tasks by propagating signals of labeled examples. Both methods are effective for task-specific adaptation. LoRA [14] reduces memory overhead by representing weight updates with low-rank matrices. AdaLoRA [47] optimizes incremental updates of pre-trained language models by adaptively allocating parameter budgets and using singular value decomposition, effectively reducing the parameter budget and avoiding intensive computation. LongLoRA [6] extends the context length of LLMs at low computational cost through sparse local attention and parameter-efficient fine-tuning. SplitLoRA [28], based on Split Federated Learning (SFL), improves training efficiency by partitioning the model and combining the parallel training of federated learning with the model splitting advantage of split learning. PEFT-based methods for LLMs, especially LoRA and its variants, have advanced fine-tuning techniques, showing excellent performance due to their simplicity and ability to reduce parameters. Thus, we fine-tune LLMs with LoRA in this paper. + +# 2.2. Federated Fine-Tuning Homogeneous LoRA + +In federated learning, several works have applied LoRA to achieve reduced communication and improved efficiency. For example, Zhang et al. [46] fine-tuned LLMs by aggregating heterogeneous instruction data from multiple clients in a federated learning framework, improving the model's generalization ability on new tasks while ensuring user privacy and data security. Wu et al. [44] introduced the FedBiOT algorithm, which compresses LLMs and splits them into two components, using a dual-layer optimization approach to achieve efficient fine-tuning and reduce computational and communication costs. FLoCoRA [12] also reduces communication overhead by freezing the originally randomized neural network parameters during training, and only training and transmitting the added LoRA adapters. However, it does not fully address the stability issues brought about by extremely low-bit quantization. Sun + +![](images/431a37c8c6f64fce36911e3c265e56238e30c9c4afbb8bd6e61d58f9a101a5f3.jpg) +Figure 2. An overview of the proposed Te-LoRA. It is based on a framework for alternating training of the LoRA modules B and A. When the clients' LoRA have different ranks, the B or A modules are obtained after client training and subsequently uploaded to the server for aggregation. On the server, the heterogeneous B or A modules are first aligned using the PAA-strategy. In the PAA-strategy, $\mu$ and $\nu$ are all 1-column vectors of size having $r_{\mathrm{max}}$ and $r_k$ , respectively. Then, the aligned B or A are aggregated using the T2M-strategy. The detailed operations will be discussed in Sec. 3. + +et al. [39] pointed out that separately integrating the two LoRA matrices does not fully approximate the original, and proposed FFA-LoRA, which fixes the randomly initialized matrices while fine-tuning only the zero-initialized ones. Compared to FFA-LoRA, the proposed Te-LoRA, based on an alternating freezing framework, not only preserves the training of all LoRA modules but also effectively addresses the issue of aggregation noise. + +# 2.3. Federated Fine-tuning Heterogeneous LoRA + +Due to non-IID data heterogeneity [26] sampled from the Dirichlet distribution and heterogeneous hardware resources, clients need to adjust the LoRA rank accordingly. Existing work has proposed methods to address this issue. For instance, SLoRA [2] addresses LoRA's limitations in high data heterogeneity scenarios with data-driven initialization, achieving performance similar to full finetuning while reducing training time and communication costs. FLoRA [43] uses a stacked aggregation method to eliminate noise in low-rank adapter (LoRA) aggregation, ensuring noise-free global updates. It supports heterogeneous LoRA configurations across clients, allowing dynamic adjustments based on data complexity and resources. However, this approach leads to linear growth in communication complexity, higher computational and storage demands, and reduced flexibility for supporting multiple task-specific adapters. Cho et al. [8] deploy LoRA modules with different ranks across clients, using zero-padding and truncation for aggregation, combining the benefits of high- and low-rank LoRA. However, the padded zeros dilute useful information from high-quality clients, reducing optimization efficiency and potentially limiting perfor + +mance. FlexLoRA [3] dynamically adjusts the LoRA rank for clients and uses Singular Value Decomposition (SVD) to reallocate global weights, enhancing resource utilization and model generalization. However, SVD truncation adds computational and storage overhead, and the resulting errors may limit performance. + +LoRA-A2 [22] alternates freezing the A and B components of the module and dynamically selects key parameters based on the importance of each client's data, effectively reducing parameter inconsistencies during the aggregation of low-rank adapters and significantly cutting communication costs. However, the method assumes that the global LoRA has a uniform initial rank, whereas in practice, each client's LoRA begins with a different rank. Compared to the related LoRA-A2, we train the initial heterogeneous-rank LoRA directly on the clients. While LoRA-A2 focuses on adaptively selecting client ranks, we emphasize the effective alignment of heterogeneous ranks on the server side. During alignment, unlike FlexLoRA and FLoRA, which use merging strategies, we designed the PAA-strategy and T2M-strategy for adaptive alignment, preserving the effective information within clients and dependencies across clients. + +# 3. Method + +# 3.1. Preliminaries + +Low-Rank Adaptation. The application of Low-Rank Adaptation (LoRA) [15] facilitates a reduction in the number of trainable parameters during the fine-tuning of a pretrained model. Specifically, the weight matrix $W \in \mathbb{R}^{d \times l}$ in the fine-tuned model can be represented as the sum of the original weight matrix $W_0 \in \mathbb{R}^{d \times l}$ and an updated + +weight matrix $\triangle W\in \mathbb{R}^{d\times l}$ . $\triangle W$ is implemented by low-rank decomposition into smaller matrices $B\in \mathbb{R}^{d\times r}$ and $A\in \mathbb{R}^{r\times l}$ : + +$$ +W = W _ {0} + \triangle W = W _ {0} + B A, \tag {1} +$$ + +where rank $r$ is much smaller than both dimensions $d$ and $l$ , thereby significantly reducing the number of trainable parameters for the weights. + +Aggregation Noise. Currently, FedAvg [30], a widely used FL algorithm, updates the global model by performing a weighted average of $n$ local clients in each communication round. When LoRA is applied to FL, training and transmitting only the updated low-rank matrices B and A significantly reduces both computational and communication costs, compared to directly updating and transmitting $\triangle W$ . Combining the aforementioned FedAvg and LoRA, FedIT [46] is proposed for federated fine-tuning, where the global LoRA matrices A and B are updated by weighted averaging of the local LoRA matrices $A_{i}$ and $B_{i}$ from $n$ clients, as follows: + +$$ +\triangle W = \frac {1}{n} \sum_ {i = 1} ^ {n} \triangle W _ {i} = \frac {1}{n} \sum_ {i = 1} ^ {n} B _ {i} \times \frac {1}{n} \sum_ {i = 1} ^ {n} A _ {i}. \tag {2} +$$ + +However, when each weight is averaged and aggregated separately, there is a bias between the aggregated parameters (left) and ideal (right), as shown below: + +$$ +\frac {1}{n} \sum_ {i = 1} ^ {n} B _ {i} \times \frac {1}{n} \sum_ {i = 1} ^ {n} A _ {i} \neq \frac {1}{n} \sum_ {i = 1} ^ {n} \left(B _ {i} A _ {i}\right). \tag {3} +$$ + +One solution is to first merge $A_{i}$ and $B_{i}$ from each client by matrix multiplication and then perform average aggregation, i.e., $\triangle W = \frac{1}{n}\sum_{i=1}^{n}(B_{i}A_{i})$ . Subsequently, the aggregated $\triangle W \in \mathbb{R}^{d\times l}$ must be decomposed into the new $A_{i} \in \mathbb{R}^{r\times l}$ and $B_{i} \in \mathbb{R}^{d\times r}$ for each client. This method introduces complex matrix decomposition, making post-decomposition training prone to instability. The details of another simple and effective solution are as follows. + +Alternating Freezing Technique. In [22], a simple alternating freezing technique is introduced to mitigate the issue of aggregation noise when applying LoRA in FL. Specifically, in round $t$ , the aggregated $\bar{A}^{(t)}$ remains frozen, and each client only trains $B_{i}^{(t)}$ , as follows: + +$$ +\frac {1}{n} \sum_ {i = 1} ^ {n} B _ {i} ^ {(t)} \times \frac {1}{n} \sum_ {i = 1} ^ {n} \bar {A} ^ {(t)} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left(B _ {i} ^ {(t)} \bar {A} ^ {(t)}\right). \tag {4} +$$ + +Then, in round $t + 1$ , the aggregated $\bar{B}^{(t + 1)}$ is frozen, and only $A_{i}^{(t + 1)}$ is trainable on each client, as shown below: + +$$ +\frac {1}{n} \sum_ {i = 1} ^ {n} \bar {B} ^ {(t + 1)} \times \frac {1}{n} \sum_ {i = 1} ^ {n} A _ {i} ^ {(t + 1)} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left(\bar {B} ^ {(t + 1)} A _ {i} ^ {(t + 1)}\right). \tag {5} +$$ + +Finally, A and B complete iterative training and aggregation through the alternating freezing technique. However, when faced with data heterogeneity and rank heterogeneity at the client, this approach does not achieve good alignment of modules A or B at the server. + +# 3.2. Proposed Method + +In real FL environments with resource heterogeneity, the training of Low-Rank Adaptation (LoRA) faces two major challenges: aggregation noise and aggregation misalignment. To address these issues, we propose a novel federated fine-tuning method with Tensor-aggregated LoRA (Te-LoRA) based on an alternating freezing training strategy, as shown in Fig. 2. Te-LoRA consists of three strategies: the alternating freezing training strategy, the Pre-Aggregation Alignment Strategy (PAA-strategy), and the Tensor Decomposition Aggregation Strategy (T2M-strategy). Based on the existing alternating freezing training strategy (Eqs. (4) and (5)), Te-LoRA first effectively avoids aggregation noise (Eq. (3)). Subsequently, the PAA-strategy aims to achieve parameter alignment between different rank modules A and B with minimal information loss. Furthermore, the T2M-strategy adaptively aligns and aggregates the main components among clients to achieve more fine-grained feature alignment. The PAA and T2M strategies complement each other, effectively mitigating the issue of aggregation misalignment. + +Pre-Aggregation Alignment strategy (PAA-strategy). The PAA-strategy is presented to achieve the optimal transformation from heterogeneous LoRAs to homogeneous LoRAs with minimal information loss for parameter alignment on the server side. In the implementation of the alternating freezing FL training strategy, the modules transmitted between the client and server are either B or A. Assuming that the module B is uploaded in the current round, module A remains frozen and stays on the client side. After the upload, the server receives a set of $K$ modules B, i.e., $\{B_k \in \mathbb{R}^{d \times r_k}\}_{k=1}^K$ . The module with the highest rank is selected as the target module $B_{\max} = B_{\arg \max_{k=1,\dots,K} r_k}$ , and the remaining modules are aligned towards the target module. Specifically, the distance matrix $\mathcal{C}$ between $B_k$ and $B_{\max}$ is calculated as follows: + +$$ +\mathcal {C} (m, n) = \sum_ {i = 1} ^ {d} \left(b _ {k} ^ {i m} - b _ {\max } ^ {i n}\right), \tag {6} +$$ + +where $\mathcal{C} \in \mathbb{R}^{r_k \times r_{\max}}$ , $b_k^{im}$ and $b_{\max}^{in}$ are the elements of the $i^{\text{th}}$ row, $m^{\text{th}}$ column of matrix $B_k$ , and $n^{\text{th}}$ column of matrix $B_{\max}$ , respectively. To smooth the matching relationship between the source and target matrices, the matrix $\mathcal{C}$ is regularized: $\mathcal{K} = \exp\left(-\frac{\mathcal{C}}{\eta}\right)$ , where $\eta$ is the regularization parameter. Next, based on the matrix $\mathcal{K}$ , we first assume that the source distribution $p$ and the target distribution $q$ + +follow a uniform distribution, i.e., $p = \left[\frac{1}{r_k},\frac{1}{r_k},\dots \frac{1}{r_k}\right]^\top$ , $q = \left[\frac{1}{r_{\mathrm{max}}},\frac{1}{r_{\mathrm{max}}},\dots \frac{1}{r_{\mathrm{max}}}\right]^\top$ . Then, using the Sinkhorn-Knopp algorithm [31], we iteratively update the adjustment factors $\mu$ and $\nu$ to make the probability matrix $\mathcal{G}$ converge to a solution that satisfies the marginal constraints of $p$ and $q$ , as follows: + +$$ +\mu = \frac {p}{\mathcal {K} \cdot \nu}, \quad \nu = \frac {q}{\mathcal {K} ^ {\top} \cdot \mu}, \tag {7} +$$ + +where the adjustment factors $\mu$ and $\nu$ are both 1-column vectors of size $r_{\mathrm{max}}$ and $r_k$ , respectively. By iteratively updating until convergence, the probability matrix $\mathcal{G}$ , representing the optimal alignment between the source and target features, is shown below: + +$$ +\mathcal {G} = \mu \cdot \mathcal {K} \cdot \nu^ {\top}. \tag {8} +$$ + +Subsequently, the source matrix $B_{k}$ is transformed using the matrix $\mathcal{G}$ to achieve alignment of different ranks with minimal information loss, as follows: + +$$ +\widetilde {B _ {k}} = B _ {k} \cdot \mathcal {G}, \tag {9} +$$ + +where $\widetilde{B_k} \in \mathbb{R}^{d \times r_{\max}}$ is the aligned matrix. The remaining module B and the module A from another round both use the same pre-aggregation process for parameter alignment. Discussions. Although previous research in FL, such as FedOTP [24] and FedAli [10], has explored Optimal Transport (OT) for fine-grained feature alignment, FedOTP [24] uses OT to align local visual features with global and local text features, smoothing the global consensus and local personalization. To balance personalization and generalization, FedAli [10] quantifies the cost of embedding inputs into prototypes using OT, effectively aligning embeddings and reducing client inconsistencies. Compared to these studies, our PAA strategy differs in three ways: 1) The others focus on data heterogeneity in personalized FL, while our work addresses resource heterogeneity in real FL; 2) Previous work aligns personalized with global knowledge via OT, while our work aligns heterogeneous weights into homogeneous ones using OT to address rank mismatches from resource heterogeneity; 3) Previous studies focus on aligning prompts and prototypes, while our work targets alignment methods for heterogeneous LoRAs. To the best of our knowledge, this is the first research to apply OT for heterogeneous LoRA alignment in resource-heterogeneous FL. + +Tensor-to-Matrix strategy (T2M-strategy). To maintain the relationships and dependencies between different clients, the Tensor-to-Matrix strategy (T2M-strategy) is designed to adaptively aggregate the LoRAs aligned by the PAA-strategy, rather than simply averaging them. Specifically, given the rank-aligned set $\{\widetilde{B_k}\}_{k = 1}^K$ of modules B, $\widehat{\mathcal{B}}$ can be stacked from $\{\widetilde{B_k}\}_{k = 1}^K$ to form a new tensor + +# Algorithm 1: Te-LoRA + +Initialization: each client $k$ initializes the LoRA $\triangle W_{k} = B_{k}A_{k}$ with $B_{k}\in \mathbb{R}^{d\times r_{k}}$ and $A_{k}\in \mathbb{R}^{r_{k}\times d}$ . + +1. for $t = 1,2,\ldots ,T$ do: +2. Select K participants for round t +3. if $t\% 2 = 1$ then: +4. for $k = 1,2,\dots ,K$ in parallel do: +5. $B_{k}^{(t + 1)} = \mathrm{LocalTraining}\left(B^{(t)},k\right)$ +6. $\widetilde{B}_k^{(t + 1)}\gets \mathrm{PAA}\left(B_k^{(t + 1)}\right)$ +7. $\mathbb{B}_{*}^{(t + 1)}\gets \mathrm{T2M}\left(\widetilde{B}_{1}^{(t + 1)},\widetilde{B}_{2}^{(t + 1)},\dots,\widetilde{B}_{\mathrm{K}}^{(t + 1)}\right)$ +8. $\overline{A}^{(t + 1)} = A^{(t)}$ +9. end for +10. else: +11. for $\mathrm{k} = 1,2,\cdots ,\mathrm{K}$ in parallel do: +12. $A_{k}^{(t + 1)} = \mathrm{LocalTraining}\left(A^{(t)},k\right)$ +13. $\widetilde{A}_k^{(t + 1)}\gets \mathrm{PAA}\left(A_k^{(t + 1)}\right)$ +14. $\mathbb{A}_{*}^{(t + 1)}\gets \mathrm{T2M}\left(\widetilde{A}_{1}^{(t + 1)},\widetilde{A}_{2}^{(t + 1)},\dots ,\widetilde{A}_{\mathrm{K}}^{(t + 1)}\right)$ +15. $\overline{B}^{(t + 1)} = B^{(t)}$ +16. end for +17. end if +18. end for + +representing the intermediate state before aggregation, i.e., $\widehat{\mathcal{B}} = \bigcup_{k=1}^{K} \widetilde{B_k}$ , where $\widetilde{B_k} \in \mathbb{R}^{d \times r_{\max}}$ and $\widehat{\mathcal{B}} \in \mathbb{R}^{d \times r_{\max} \times K}$ . Here, the aggregation process can be viewed as a tensor decomposition process. Inspired by [21], three versions of the T2M-strategy (T2M $_1$ , T2M $_2$ , and T2M $_3$ ) is explored to decompose $\widehat{\mathcal{B}}$ into the result $\mathbb{B}_*$ . The generalized form of the decomposition is as follows: + +$$ +\widehat {\mathcal {B}} = \mathbb {B} _ {*} \times_ {1} U \times_ {2} V \times_ {3} X, \tag {10} +$$ + +where $\mathbb{B}_*$ is a core tensor, $U,V$ , and $X$ serve as the factor matrices, and the symbol $\times_{i}$ ( $i = 1,2,3$ ) denotes mode- $i$ tensor-matrix multiplication. $\mathrm{T2M_1}$ is as follows: + +$$ +\widehat {\mathcal {B}} = \mathbb {B} _ {*} ^ {(1)} \times_ {1} U ^ {(1)} \times_ {2} V ^ {(1)} = \left[ \left[ \mathbb {B} _ {*} ^ {(1)}; U ^ {(1)}, V ^ {(1)}, \mathbb {I} \right] \right], \tag {11} +$$ + +where $\mathbb{B}_{*}^{(1)}\in \mathbb{R}^{1\times 1\times K}$ , $U^{(1)}\in \mathbb{R}^{d\times 1}$ , $V^{(1)}\in \mathbb{R}^{r_{max}\times 1}$ and $X = \mathbb{I}$ is the $K\times K$ identity matrix. [ ] denotes the shorthand. $\mathbb{B}_{*}^{(1)}$ represents the weight coefficient of different clients, reflecting the importance scores. And the aggregated module is $\bar{B}$ , where $\bar{B}_{i,j} = \sum_{k = 1}^{K}\hat{\mathcal{B}}_{i,j,k}\cdot \mathbb{B}_{*1,1,k}^{(1)}$ . The formulae of T2M2 is as follows: + +$$ +\widehat {\mathcal {B}} = \mathbb {B} _ {*} ^ {(2)} \times_ {3} X ^ {(2)} = \left[ \left[ \mathbb {B} _ {*} ^ {(2)}; \mathbb {I}, \mathbb {I}, X ^ {(2)} \right] \right], \tag {12} +$$ + +where $\mathbb{B}_*^{(2)}\in \mathbb{R}^{d\times r_{max}\times 1},X^{(2)}\in \mathbb{R}^{K\times 1},U = \mathbb{I}$ and $V = \mathbb{I}$ are the $d\times d$ and $r_{max}\times r_{max}$ identity matrices. $\mathbb{B}_*^{(2)}$ is the aggregation result capturing the common and representative information, i.e., $\bar{B} = \mathbb{B}_*^{(2)}$ . T2M3 is a combined version of T2M1 and T2M2, as shown below: + +$$ +\mathbb {B} _ {*} ^ {(3)} = \mathrm {T} 2 \mathrm {M} _ {2} (\widehat {\mathcal {B}} \odot \mathrm {T} 2 \mathrm {M} _ {1} (\widehat {\mathcal {B}})), \tag {13} +$$ + +where $\mathbb{B}_{*}^{(1)} = \mathrm{T2M}_{1}(\widehat{\mathcal{B}})$ , $\odot$ denotes hadamard product used to obtain the new weighted tensor $\widehat{\mathcal{B}}^{(1)}\in \mathbb{R}^{d\times r_{\max}\times K}$ , i.e., $\widehat{\mathcal{B}}_{i,j,k}^{(1)} = \widehat{\mathcal{B}}_{i,j,k}\cdot \mathbb{B}_{*1,1,k}^{(1)}$ . The final outcome, $\mathbb{B}_{*}^{(3)}\in \mathbb{R}^{d\times r_{\max}\times 1}$ , is the aggregated module, i.e., $\bar{B} = \mathbb{B}_{*}^{(3)}$ . After completing aggregation on the server side, the module $\bar{B}$ is distributed to each client for the next round of training. The T2M-strategy is also applicable to the decomposition of module A when module B is frozen for aggregation. Algorithm 1 provides the pseudo-code for Te-LoRA. + +# 4. Experiments + +# 4.1. Experiment Setup + +Comparison methods. We compare the proposed Te-LoRA with three state-of-the-art approaches that support federated fine-tuning of heterogeneous LoRA. "Zeropadding" [8] scales the lower ranks of heterogeneous LoRAs from clients by padding them with zeros up to the maximum rank on the server side. "FLoRA" [43] stacks all heterogeneous LoRAs together and then performs matrix multiplication for aggregation on the server side. "FlexLoRA" [3] aggregates heterogeneous LoRAs after matrix multiplication, then decomposes them into modules A and B via SVD. "Local" refers to the average test results from all clients independently training on their local data, serving as an important baseline to validate the effectiveness of FL. To ensure a fair comparison, we reproduced the comparison methods based on the existing codebase [45], with the same LLM, data splits, and training configurations, and conducted all the experiments. + +Datasets and evaluation. We used three datasets to train Te-LoRA: Alpaca-GPT4 [32], an English instruction-following dataset consisting of 52k samples generated by GPT-4 [1] based on Alpaca's self-learning [42]; Databricks-dolly-15k (Dolly) [46], an open-source dataset created by Databricks' staff, which includes tasks such as brainstorming, classification, closed quizzes, open quizzes, and summarization; and Wizard dataset [29] is a quiz dataset including 70k instruction-output pairs. For federated training, datasets were partitioned using a Dirichlet distribution $(\alpha = 0.5)$ . Furthermore, we separately consider closed [7] and open evaluation benchmarks, using MMLU [13] for closed evaluation and MT-Bench [48] for open evaluation. MMLU is a widely used Q&A dataset for LLM fine-tuning, including 14,024 questions across 57 different topics to assess the logical reasoning ability of LLMs. MT-Bench is + +
MethodMMLUMT-Bench
HomoWizardDollyAlpaca-GPT4Wizard
Local20.8725.1338.662.92
Zero-padding [8]22.0326.5944.883.17
FLoRA [43]22.9127.7345.153.22
FlexLoRA [3]22.9728.2145.343.26
Te-LoRA (Ours)23.3528.4445.863.31
+ +Table 1. Comparison of Te-LoRA with other methods on MMLU and MT-bench. 'Homo' represents settings with homogeneous LoRA ranks. + +a set of challenging multi-round open-ended questions for assessing chat assistants [48]. + +Training details. For the model configurations, we use the pre-trained Llama2-7B [41] as the base model and conduct 200 rounds of FL communication. The initial learning rate is set to 5e-5 and decays to 1e-6 after 200 rounds using a cosine scheduler. In each round, two clients are randomly selected, each training for 10 epochs with a batch size of 16. Referring to FLoRA's setup [43] in Homogeneous LoRA, the rank of LoRA is 16. Referring to FLoRA's simulation of a real-world scenario where clients have heterogeneous computational resources, in Heterogeneous LoRA, we assign [64, 32, 16, 16, 8, 8, 4, 4, 4, 4] as different local LoRA ranks applied to 10 clients. All experiments are conducted on a single NVIDIA 3090 GPU. + +# 4.2. Comparison Experiment + +Homogeneous LoRA. In the homogeneous LoRA setup, we configure all clients to share the same LoRA rank of 16. As shown in Tab. 1, we draw the following conclusions: 1) All four federated fine-tuning LoRA methods outperform "Local" with Te-LoRA achieving a 7.2 improvement in the MMLU benchmark compared to "Local" under AlpacaGPT4 training. 2) Among the comparison methods, Zeropadding, which is affected by aggregation noise, performs the worst. Notably, in homogeneous LoRA, Zero-padding degrades to a traditional approach where modules B and A are aggregated separately on the server. 3) FlexLoRA outperforms FLoRA. Both aggregate the matrix-multiplied weights from modules B and A, but FlexLoRA uses SVD to decompose the aggregated weights and sends new modules to clients, while FLoRA merges the aggregated weights with the client LLM and reinitialized LoRA training. The reinitialized LoRA may cause instability in client training. 4) Te-LoRA outperforms all methods across tasks by effectively resolving aggregation noise and misalignment. + +Heterogeneous LoRA. As described in Sec. 4.1, to simulate the scenario of resource heterogeneity among clients in real-world settings, we apply different local LoRA ranks of [64, 32, 16, 16, 8, 8, 4, 4, 4, 4] to 10 clients. By comparing the results of Heterogeneous LoRA (Heter) in Tab. 2 + +
Method HeterWizardMMLU DollyAlpaca-GPT4MT-Bench Wizard
Local20.5325.1038.542.89
Zero-padding [8]21.8626.4844.473.15
FLoRA [43]22.7927.6044.943.20
FlexLoRA [3]23.4928.1845.483.28
Te-LoRA (Ours)23.7128.3746.163.33
+ +Table 2. Comparison of Te-LoRA with other methods on MMLU and MT-bench. 'Heter' represents settings with heterogeneous LoRA ranks. + +![](images/a882a4bbec1565049a862cb21dff494fc6d5d29ebbd130dabdce99984e790f1b.jpg) +Figure 3. Comparison of the communication cost. Te-LoRA achieves lower overhead compared to the other methods. + +with Homogeneous LoRA (Homo) in Tab. 1, we make several interesting observations. First, both Zero-padding and FLoRA show a performance drop in the Heter configuration compared to Homo, with Zero-padding decreasing by up to 0.41 and FLoRA by up to 0.21. Since FLoRA mitigates aggregation noise, it still outperforms Zero-padding. Second, Te-LoRA and FlexLoRA under Heter outperform Homo in most evaluation tasks, demonstrating their ability to effectively handle resource heterogeneity and optimize resource utilization. Because Te-LoRA more effectively addresses the aggregation misalignment issue, it still maintains the best performance under Heter. Finally, Te-LoRA and FlexLoRA show slight performance drops when transitioning from Homo to Heter on the Dolly dataset. This may be due to Dolly's smaller dataset (15k) compared to Alpaca-GPT4 (52k) and Wizard (70k). In the Heter configuration, the 64 and 32 rank LoRA models may overfit on clients with less data, leading to poorer aggregation results. + +Communication overhead. Communication cost plays a crucial role in real-world FL. Therefore, we compared the total upload and download communication costs in a single round across different methods. To ensure a fair comparison, we normalized the transmission parameters. As shown in Fig. 3, the proposed Te-LoRA has the lowest communication cost, as it uploads only module A or B per round. In contrast, FlexLoRA and Zero-padding require both modules to be uploaded and downloaded each round, while + +![](images/b59d3aeb35bfebc475f63bdc06fa2eb670132a07db7784e43ed4da720fa6d8db.jpg) +Figure 4. Impact of different ranks (64, 32, 16, 8, and 4). + +FLoRA incurs the highest cost due to transmitting the result of multiplying matrices B and A during download. The quantitative comparison highlights the advantage of the alternating freezing technique. + +Impact of rank. To further investigate the impact of different rank configurations on homogeneous-rank clients, we conducted experiments on the "Local" baseline with five rank settings (64, 32, 16, 8, and 4). Training was performed on the Wizard and Alpaca-GPT4 datasets, and evaluation was conducted using MT-Bench and MMLU, respectively. As illustrated in Fig. 4, setting the rank to 4 for all clients, which aligns with the most resource-constrained ones, leads to a significant performance decline. While higher ranks can improve performance, they are not feasible for resource-limited clients. Therefore, adopting a heterogeneous rank configuration enables better resource utilization and further enhances performance. + +# 4.3. Ablation Experiment + +To demonstrate the effectiveness of the PAA-strategy and T2M-strategy in Te-LoRA, we conducted several ablation experiments. Specifically, "w/o PPA" refers to replacing the PAA-strategy with zero-padding in the heterogeneous experiment, which combines the alternating freezing method with zero-padding. "w/o T2M" indicates replacing the T2M-strategy with the weighted average aggregation of FedAvg in the heterogeneous experiments. + +The effects of PAA-strategy. To compare with the alignment effect of zero-padding and PAA-strategy, we visualized the cosine similarity between clients after alignment using zero-padding and PAA-strategy, as well as the change curves of the singular value distribution before and after alignment. As shown in Fig. 5, we visualized the cosine similarity between four clients, where most clients exhibit higher cosine similarity after alignment with the PAA strategy compared to zero-padding. This indicates that the PAA-strategy not only aligns the parameters between heterogeneous LoRAs on the server side but also reduces the information loss caused by the alignment. To better illustrate the differences before and after alignment, we compared the normalized singular value distribution curves, as shown in Fig. 6. It can be observed that the distribution of singular + +![](images/7ef770263f10047c1d54633ddc3902e1bfcada4c1c54588a4331c4793d97191e.jpg) +(a) On PAA-strategy. + +![](images/222b1392e865779e914184d839452b0446077c1a3ffe6a26d3fdf2b96175725d.jpg) +(b) On Zero-padding. + +![](images/916fe394f873563fc333a4c7b2bb75581782655c9a2f9d57f14f88297f76a713.jpg) +(a) In round four. + +![](images/c2418d0de271994115dcbbffb36daed19cd0e7533d00dd048817d9f410097873.jpg) +Figure 5. Cosine similarity visualization of the local model processed by PAA-strategy or Zero-padding with the global model. +(b) In round seven. +Figure 6. Singular value distribution curves. The source LoRA and those transformed by zero-padding and the PAA strategy undergo SVD to plot the singular value distribution curves. + +values after applying the PAA strategy is closer to that of the original LoRA, while the distribution after zero-padding is relatively farther away. This indicates that the PAA strategy retains more of the major components to a certain extent and is closer to the original LoRA compared to zero-padding. + +To further validate the effect of PPA quantitatively, we conducted an ablation study, as shown in Tab. 3. In all scenarios, "w/o PAA" performs worse than $\mathrm{T2M}_2$ , with a performance drop of at least 1.3 points in each case. This validates that the PAA-strategy, compared to Zero-padding, can more effectively reduce information loss when aligning parameters between different levels of the A and B modules. Meanwhile, compared to the FlexLoRA in Tab. 2, "w/o T2M" performs better in both the Dolly and AlpacaGPT4 scenarios. This suggests that combining the PAA-strategy and alternating freezing strategy avoids aggregation noise while achieving parameter alignment between different levels of modules A and B, reducing information loss. + +The effects of T2M-strategy. To evaluate the impact of this strategy, as shown in Tab. 3, we performed an ablation experiment with "w/o T2M". First, "w/o T2M" performs worse in all three scenarios compared to the $\mathrm{T2M}_2$ method, especially in the Alpaca-GPT4 dataset, where w/o T2M is about 0.63 points lower than $\mathrm{T2M}_2$ . This validates that the T2M-strategy, compared to the weighted average + +
Method HeterWizardMMLU DollyAlpaca-GPT4
w/o PAA22.4027.1044.77
w/o T2M23.3828.2145.53
T2M122.9128.1445.30
T2M223.7128.3746.16
T2M323.6728.3045.99
+ +Table 3. Validation of Te-LoRA components. 'Heter' represents settings with heterogeneous LoRA ranks. + +aggregation operation of FedAvg, better preserves the key information from each client, enabling finer feature alignment. In comparison to Zero-padding in Tab. 2, "w/o PAA" achieves higher MMLU scores for all three datasets: 0.54, 0.38, and 0.3 points higher on Wizard, Dolly, and Alpaca-GPT4, respectively. This indicates that the T2M-strategy combined with alternating freezing more effectively aligns features while avoiding aggregation noise. + +The effects of $\mathrm{T2M_1}$ , $\mathrm{T2M_2}$ , and $\mathrm{T2M_3}$ . As observed from Tab. 3, $\mathrm{T2M_1}$ performs slightly worse than $\mathrm{T2M_2}$ . In the Wizard, Dolly, and Alpaca-GPT4 datasets, $\mathrm{T2M_2}$ scores 0.2, 0.23, and 0.86 points higher than $\mathrm{T2M_1}$ , respectively. While $\mathrm{T2M_3}$ outperforms $\mathrm{T2M_1}$ on all three datasets, it still trails $\mathrm{T2M_2}$ , though the results are nearly identical. For example, in the Wizard, Dolly, and Alpaca-GPT4 datasets, $\mathrm{T2M_2}$ scores 0.004, 0.007, and 0.005 points higher than $\mathrm{T2M_3}$ . This suggests that although the tensor $\mathbb{B}_*^1 \in \mathbb{R}^{1 \times 1 \times k}$ obtained from the compression of $\widehat{\mathcal{B}} \in \mathbb{R}^{d \times r_{\max} \times k}$ can preserve the proportion of each client, it is slightly less effective in retaining client information. On the other hand, the tensor $\mathbb{B}_* \in \mathbb{R}^{d \times r_{\max} \times 1}$ obtained from the compression of $\widehat{\mathcal{B}}$ better preserves the key information from each client and does not introduce redundancy in information like $\mathbb{B}_*^3 \in \mathbb{R}^{d \times r_{\max} \times 1}$ does. + +# 5. Conclusion + +In this paper, we introduce a novel Tensor-aggregation LoRA method (Te-LoRA) to address the noise and misalignment issues in the aggregation of low-rank adaptation within federated learning under resource heterogeneity. Building upon an alternating freezing strategy, Te-LoRA incorporates Pre-Aggregation Alignment (PAA) and Tensor Decomposition Aggregation (T2M) strategies. The PAA strategy utilises OT technology to align the weights of heterogeneous LoRA modules A and B from different clients into homogeneous weights, while the T2M-strategy adaptively aggregates key components to preserve interdependencies. Experimental results demonstrate that Te-LoRA effectively reduces communication costs and significantly improves fine-tuning performance, offering a viable solution for efficient federated fine-tuning in resource-constrained environments. + +Acknowledgements. The work is supported by the National Natural Science Foundation of China (Grant No. 62222207, 62427808, 62472208, 62276134). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.6 +[2] Sara Babakniya, Ahmed Roushdy Elkordy, Yahya H Ezzeldin, Qingfeng Liu, Kee-Bong Song, Mostafa El-Khamy, and Salman Avestimehr. Slora: Federated parameter efficient fine-tuning of language models. arXiv preprint arXiv:2308.06522, 2023.1, 3 +[3] Jiamu Bai, Daoyuan Chen, Bingchen Qian, Liuyi Yao, and Yaliang Li. Federated fine-tuning of large language models under heterogeneous language tasks and client resources. arXiv e-prints, pages arXiv-2402, 2024. 2, 3, 6, 7 +[4] Desirée Bill and Theodor Eriksson. Fine-tuning a llm using reinforcement learning from human feedback for a therapy chatbot application, 2023. 1 +[5] Daoyuan Chen, Yilun Huang, Zhijian Ma, Hesen Chen, Xuchen Pan, Ce Ge, Dawei Gao, Yuexiang Xie, Zhaoyang Liu, Jinyang Gao, et al. Data-juicer: A one-stop data processing system for large language models. In Companion of the 2024 International Conference on Management of Data, pages 120-134, 2024. 1 +[6] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. Longlora: Efficient finetuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 2 +[7] Yew Ken Chia, Pengfei Hong, Lidong Bing, and Soujanya Poria. Instructeval: Towards holistic evaluation of instruction-tuned large language models. arXiv preprint arXiv:2306.04757, 2023. 6 +[8] Yae Jee Cho, Luyang Liu, Zheng Xu, Aldi Fahrezi, and Gauri Joshi. Heterogeneous lora for federated fine-tuning of ondevice foundation models. arXiv preprint arXiv:2401.06432, 2024. 3, 6, 7 +[9] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023. 1 +[10] Sannara Ek, Kaile Wang, François Portet, Philippe Lalanda, and Jiannong Cao. Fedali: Personalized federated learning with aligned prototypes through optimal transport. arXiv preprint arXiv:2411.10595, 2024. 5 +[11] Junyao Gao, Yanan Sun, Fei Shen, Xin Jiang, Zhening Xing, Kai Chen, and Cairong Zhao. Faceshot: Bring any character into life. arXiv preprint arXiv:2503.00740, 2025. 1 +[12] Lucas Grativol, Mathieu Leonardon, Guillaume Muller, Virginie Fresse, and Matthieu Arzel. Flocora: Federated learning compression with low-rank adaptation. In 2024 32nd European Signal Processing Conference (EUSIPCO), pages 1786-1790. IEEE, 2024. 2 + +[13] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 6 +[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2 +[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3 +[16] Ahmed Imteaj, Urmish Thakker, Shiqiang Wang, Jian Li, and M Hadi Amini. A survey on federated learning for resource-constrained IoT devices. IEEE Internet of Things Journal, 9(1):1-24, 2021. 2 +[17] Xin Jiang, Hao Tang, Junyao Gao, Xiaoyu Du, Shengfeng He, and Zechao Li. Delving into multimodal prompting for fine-grained visual classification. In Proceedings of the AAAI conference on artificial intelligence, pages 2570-2578, 2024. 1 +[18] Xin Jiang, Hao Tang, and Zechao Li. Global meets local: Dual activation hashing network for large-scale fine-grained image retrieval. IEEE Transactions on Knowledge and Data Engineering, 2024. 1 +[19] Xin Jiang, Hao Tang, Rui Yan, Jinhui Tang, and Zechao Li. Dvf: Advancing robust and accurate fine-grained image retrieval with retrieval guidelines. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2379-2388, 2024. 1 +[20] Dominique Kelly, Yimin Chen, Sarah E Cornwell, Nicole S Delellis, Alex Mayhew, Sodiq Onaolapo, and Victoria L Rubin. Bing chat: The future of search engines? Proceedings of the Association for Information Science and Technology, 60(1):1007-1009, 2023. 1 +[21] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51(3):455-500, 2009. 5 +[22] Jabin Koo, Minwoo Jang, and Jungseul Ok. Towards robust and efficient federated low-rank adaptation with heterogeneous clients. arXiv preprint arXiv:2410.22815, 2024. 2, 3, 4 +[23] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 2 +[24] Hongxia Li, Wei Huang, Jingya Wang, and Ye Shi. Global and local prompts cooperation via optimal transport for federated learning. In CVPR, pages 12151-12161, 2024. 5 +[25] Pengpeng Li, Xiangbo Shu, Chun-Mei Feng, Yifei Feng, Wangmeng Zuo, and Jinhui Tang. Surgical video workflow analysis via visual-language learning. npj Health Systems, 2 (1):5, 2025. 1 +[26] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of fedavg on non-iid data. arXiv preprint arXiv:1907.02189, 2019. 3 +[27] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2 + +[28] Zheng Lin, Xuanjie Hu, Yuxin Zhang, Zhe Chen, Zihan Fang, Xianhao Chen, Ang Li, Praneeth Vepakomma, and Yue Gao. Splitlora: A split parameter-efficient fin-tuning framework for large language models. arXiv preprint arXiv:2407.00952, 2024. 2 +[29] Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583, 2023. 6 +[30] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 4 +[31] Edouard Pauwels and Samuel Vaiter. The derivatives of sinkhorn-knopp converge. SIAM Journal on Optimization, 33(3):1494-1517, 2023. 5 +[32] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 6 +[33] Zhen Qin, Daoyuan Chen, Bingchen Qian, Bolin Ding, Yaliang Li, and Shuiguang Deng. Federated full-parameter tuning of billion-sized language models with communication cost under 18 kilobytes. arXiv preprint arXiv:2312.06353, 2023. 1 +[34] Riccardo Salami, Pietro Buzzega, Matteo Mosconi, Jacopo Bonato, Luigi Sabetta, and Simone Calderara. Closed-form merging of parameter-efficient modules for federated continual learning. arXiv preprint arXiv:2410.17961, 2024. 2 +[35] Xiangbo Shu, Jinhui Tang, Guo-Jun Qi, Wei Liu, and Jian Yang. Hierarchical long short-term concurrent memory for human interaction recognition. IEEE transactions on pattern analysis and machine intelligence, 43(3):1110-1118, 2019. 2 +[36] Xiangbo Shu, Liyan Zhang, Guo-Jun Qi, Wei Liu, and Jinhui Tang. Spatiotemporal co-attention recurrent neural networks for human-skeleton motion prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3300-3315, 2021. 1 +[37] Karan Singhal, Shekoofeh Azizi, Tao Tu, S Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, et al. Large language models encode clinical knowledge. Nature, 620 (7972):172-180, 2023. 1 +[38] Tianyu Song, Shumin Fan, Pengpeng Li, Jiyu Jin, Guiyue Jin, and Lei Fan. Learning an effective transformer for remote sensing satellite image dehazing. IEEE Geoscience and Remote Sensing Letters, 20:1-5, 2023. 1 +[39] Youbang Sun, Zitao Li, Yaliang Li, and Bolin Ding. Improving lora in privacy-preserving federated learning. CoRR, 2024. 3 +[40] Jinhui Tang, Xiangbo Shu, Rui Yan, and Liyan Zhang. Coherence constrained graph listm for group activity recognition. IEEE transactions on pattern analysis and machine intelligence, 44(2):636-647, 2019. 2 + +[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 6 +[42] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560, 2022. 6 +[43] Ziyao Wang, Zheyu Shen, Yexiao He, Guoheng Sun, Hongyi Wang, Lingjuan Lyu, and Ang Li. Flora: Federated fin-tuning large language models with heterogeneous low-rank adaptations. arXiv preprint arXiv:2409.05976, 2024. 2, 3, 6, 7 +[44] Feijie Wu, Zitao Li, Yaliang Li, Bolin Ding, and Jing Gao. Fedbiot: Llm local fine-tuning in federated learning without full model. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3345-3355, 2024. 2 +[45] Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. Openfedllm: Training large language models on decentralized private data via federated learning. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 6137-6147, 2024. 6 +[46] Jianyi Zhang, Saeed Vahidian, Martin Kuo, Chunyuan Li, Ruiyi Zhang, Tong Yu, Guoyin Wang, and Yiran Chen. Towards building the federatedgpt: Federated instruction tuning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6915-6919. IEEE, 2024. 1, 2, 4, 6 +[47] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adalora: Adaptive budget allocation for parameter-efficient fine-tuning. arXiv preprint arXiv:2303.10512, 2023. 2 +[48] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6 \ No newline at end of file diff --git a/2025/Tensor-aggregated LoRA in Federated Fine-tuning/images.zip b/2025/Tensor-aggregated LoRA in Federated Fine-tuning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..15b75dc78262b0ad6d6cb119912d3e294f7c5ae5 --- /dev/null +++ b/2025/Tensor-aggregated LoRA in Federated Fine-tuning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80c251214890cc28be44baf1386cc6fc6acc8914f558c7b7de01c96baa6be7ee +size 397441 diff --git a/2025/Tensor-aggregated LoRA in Federated Fine-tuning/layout.json b/2025/Tensor-aggregated LoRA in Federated Fine-tuning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..049b5d8e4ca98719643c4e193f5e1af64bf1bbc7 --- /dev/null +++ b/2025/Tensor-aggregated LoRA in Federated Fine-tuning/layout.json @@ -0,0 +1,9703 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 149, + 103, + 463, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 103, + 463, + 121 + ], + "spans": [ + { + "bbox": [ + 149, + 103, + 463, + 121 + ], + "type": "text", + "content": "Tensor-aggregated LoRA in Federated Fine-tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "text", + "content": "Zhixuan Li" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Binqian Xu" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Xiangbo Shu" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Jiachao Zhang" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Yazhou Yao" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Guo-Sen Xie" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "text", + "content": ", Jinhui Tang" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "text", + "content": "Nanjing University of Science and Technology \n" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "text", + "content": "Nanjing Institute of Technology \n" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 555, + 200 + ], + "type": "text", + "content": "Nanjing Forestry University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 201, + 547, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 201, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 67, + 201, + 547, + 227 + ], + "type": "text", + "content": "lzxlizhixuan@njust.edu.cn, xubinq11@gmail.com, shuxb@njust.edu.cn, zhangjc07@foxmail.com, yazhou.yao@njust.edu.cn, gsxiehm@gmail.com, tangjh@njfu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 280, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 544 + ], + "type": "text", + "content": "The combination of Large Language Models (LLMs) and Federated Learning (FL) to leverage privacy-preserving data has emerged as a promising approach to further enhance the Parameter-Efficient Fine-Tuning (PEFT) capabilities of LLMs. In real-world FL settings with resource heterogeneity, the training process of Low-Rank Adaptation (LoRA), the representative PEFT method, still faces two major challenges: aggregation noise and aggregation misalignment. In this paper, we propose a novel Tensor-aggregated LoRA (Te-LoRA) in Federated Finetuning based on an alternating-freeze training strategy to avoid aggregating noise without additional server-side computational costs, while mitigating aggregation suboptimality caused by parameter misalignment between heterogeneous LoRAs. Especially in addressing the aggregation suboptimality issue, we design the Pre-Aggregation Alignment strategy (PAA-strategy) and Tensor-to-Matrix strategy (T2M-strategy) for aligning heterogeneous LoRAs and aggregating them into an united tensor, which is then decomposed into matrices adapted for client download. Extensive experiments demonstrate the effectiveness and robustness of Te-LoRA in both homogeneous and heterogeneous settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 579, + 135, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 135, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 135, + 592 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "text", + "content": "Large Language Models (LLMs) have demonstrated exceptional performance across various task scenarios, such as search engines [20], chatbots [4], healthcare [37], and more. When using pre-trained LLMs to adapt to downstream tasks, significant computational resources are required to fine-tune the model parameters. Therefore, to improve the applicability of LLMs, it is necessary to fine-tune specific" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 254, + 558, + 423 + ], + "blocks": [ + { + "bbox": [ + 316, + 254, + 558, + 423 + ], + "lines": [ + { + "bbox": [ + 316, + 254, + 558, + 423 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 558, + 423 + ], + "type": "image", + "image_path": "54b59272811ac57b49a56cb3edf41ae5257c02136f2020640ceed37df523cae8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 434, + 555, + 479 + ], + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 479 + ], + "type": "text", + "content": "Figure 1. Comparison of different federated fine-tuning LoRAs. Compared to methods (a), (b), and (c), our Te-LoRA performs better in addressing aggregation noise, communication cost, and aggregation misalignment issues." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 496, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 639 + ], + "type": "text", + "content": "datasets [9, 17, 25]. However, collecting such datasets comes with associated costs and privacy concerns [5, 18, 38]. Consequently, researchers have turned to Federated Learning (FL) as a technical approach to fine-tune LLMs using distributed client data without compromising data privacy [2, 30, 33]. In this context, researchers have explored the application of Parameter-Efficient Fine-Tuning (PEFT) [11, 19, 23] methods for federated fine-tuning of LLMs, saving substantial computational resources. Among these, Low-Rank Adaptation (LoRA) [14] has gained attention due to its significant reduction in the number of communication parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "However, simply applying LoRA in real FL environment presents several challenges, one of the key issues being aggregation noise. In detail, traditional federated fine-tuning of LLMs [36, 46] averages modules A and B of local LoRAs separately, leading to discrepancies between the aggregated and ideal global LoRAs, thereby introducing noise into the" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "text", + "content": "*Equal contributions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1058" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "update process. To address this issue, as shown in Fig. 1, the two main solutions are as follows: 1) Merging modules A and B [3, 43], where A and B are first merged by matrix multiplication and then aggregated. 2) Alternating the training of modules A and B [22, 34], where either A or B is kept frozen for consistency at each round, with only the learning B or A being individually aggregated. Compared to the \"Merging\" strategy, the \"Alternating\" strategy reduces communication costs by nearly half, as it transmits only A or B per round." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 295, + 361 + ], + "type": "text", + "content": "The existing \"Alternating\" method [22] further reduces communication costs by selecting weights from a global LoRA with a unified initial rank. However, this homogenized assumption contradicts the core feature of resource heterogeneity in FL [16, 35, 40], where real-world scenarios involve clients with local LoRAs of different initial ranks. When LoRA modules A or B with heterogeneous ranks are uploaded and aggregated, an explicit rank dimension misalignment arises. Even when adopting the \"merging\" strategy for weight aggregation through matrix multiplication, semantically driven feature alignment issues may still persist. This multi-level parameter/feature mismatch caused by rank heterogeneity and data heterogeneity can be referred to as the aggregation misalignment problem." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 361, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 553 + ], + "type": "text", + "content": "To simultaneously mitigate the issues of aggregation noise and aggregation misalignment in heterogeneous settings, as shown in Fig. 1, we propose a novel Tensor-aggregated LoRA called Te-LoRA. Te-LoRA consists of two main strategies: First, due to the impact of heterogeneity, the ranks of the matrices uploaded by each client to the server are often inconsistent. We present a Pre-Aggregation Alignment strategy (PAA-strategy) to align rank while minimizing information loss or conflict. Second, to preserve the potential relationships and dependencies between different clients, we design the key Tensor-to-Matrix strategy (T2M-strategy) for decomposing the tensor stacked from aligned matrices. T2M-strategy can adapt to the aggregation of heterogeneous LoRAs while effectively preserving the key components from each client. The main contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "text", + "content": "- Alternating-freeze training for Tensor version of LoRA. We propose a novel Tensor-aggregated LoRA (Te-LoRA) in Federated Fine-tuning to efficiently aggregate the main components of each client based on an alternating-freeze training approach, avoiding aggregation noise while alleviating aggregation suboptimality among heterogeneous LoRAs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 294, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 700 + ], + "type": "text", + "content": "- Alignment of heterogeneous LoRAs for server aggregation. We design a Pre-Aggregation Alignment strategy (PAA-strategy) to align heterogeneous LoRAs, while preserving the unique features of each client and minimizing general information loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 294, + 713 + ], + "type": "text", + "content": "- Decomposition of aggregated LoRAs for client down" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 133 + ], + "type": "text", + "content": "load. We devise a Tensor-to-Matrix strategy (T2M-strategy) that decomposes the tensor into matrices adapted for client download, after aggregating LoRAs into an united tensor covering relationships and dependencies among LoRAs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 145, + 400, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 145, + 400, + 157 + ], + "spans": [ + { + "bbox": [ + 314, + 145, + 400, + 157 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 165, + 490, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 165, + 490, + 178 + ], + "spans": [ + { + "bbox": [ + 314, + 165, + 490, + 178 + ], + "type": "text", + "content": "2.1. Parameter-Efficient Fine-Tuning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 183, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 495 + ], + "type": "text", + "content": "To effectively fine-tune LLMs while reducing computational resources and storage requirements, Parameter-Efficient Fine-Tuning (PEFT) techniques have emerged. Among these, Prefix-Tuning [27] guides the pre-trained language model's generation process by optimizing a continuous sequence of task-specific vectors (called prefixes) while keeping the model parameters frozen. Soft prompts [23] enable frozen language models to perform specific tasks by propagating signals of labeled examples. Both methods are effective for task-specific adaptation. LoRA [14] reduces memory overhead by representing weight updates with low-rank matrices. AdaLoRA [47] optimizes incremental updates of pre-trained language models by adaptively allocating parameter budgets and using singular value decomposition, effectively reducing the parameter budget and avoiding intensive computation. LongLoRA [6] extends the context length of LLMs at low computational cost through sparse local attention and parameter-efficient fine-tuning. SplitLoRA [28], based on Split Federated Learning (SFL), improves training efficiency by partitioning the model and combining the parallel training of federated learning with the model splitting advantage of split learning. PEFT-based methods for LLMs, especially LoRA and its variants, have advanced fine-tuning techniques, showing excellent performance due to their simplicity and ability to reduce parameters. Thus, we fine-tune LLMs with LoRA in this paper." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 504, + 544, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 504, + 544, + 517 + ], + "spans": [ + { + "bbox": [ + 314, + 504, + 544, + 517 + ], + "type": "text", + "content": "2.2. Federated Fine-Tuning Homogeneous LoRA" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 713 + ], + "type": "text", + "content": "In federated learning, several works have applied LoRA to achieve reduced communication and improved efficiency. For example, Zhang et al. [46] fine-tuned LLMs by aggregating heterogeneous instruction data from multiple clients in a federated learning framework, improving the model's generalization ability on new tasks while ensuring user privacy and data security. Wu et al. [44] introduced the FedBiOT algorithm, which compresses LLMs and splits them into two components, using a dual-layer optimization approach to achieve efficient fine-tuning and reduce computational and communication costs. FLoCoRA [12] also reduces communication overhead by freezing the originally randomized neural network parameters during training, and only training and transmitting the added LoRA adapters. However, it does not fully address the stability issues brought about by extremely low-bit quantization. Sun" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1059" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 74, + 569, + 242 + ], + "blocks": [ + { + "bbox": [ + 58, + 74, + 569, + 242 + ], + "lines": [ + { + "bbox": [ + 58, + 74, + 569, + 242 + ], + "spans": [ + { + "bbox": [ + 58, + 74, + 569, + 242 + ], + "type": "image", + "image_path": "431a37c8c6f64fce36911e3c265e56238e30c9c4afbb8bd6e61d58f9a101a5f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed Te-LoRA. It is based on a framework for alternating training of the LoRA modules B and A. When the clients' LoRA have different ranks, the B or A modules are obtained after client training and subsequently uploaded to the server for aggregation. On the server, the heterogeneous B or A modules are first aligned using the PAA-strategy. In the PAA-strategy, " + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "text", + "content": " are all 1-column vectors of size having " + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{max}}" + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 54, + 251, + 555, + 307 + ], + "type": "text", + "content": ", respectively. Then, the aligned B or A are aggregated using the T2M-strategy. The detailed operations will be discussed in Sec. 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 322, + 294, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 322, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 322, + 294, + 418 + ], + "type": "text", + "content": "et al. [39] pointed out that separately integrating the two LoRA matrices does not fully approximate the original, and proposed FFA-LoRA, which fixes the randomly initialized matrices while fine-tuning only the zero-initialized ones. Compared to FFA-LoRA, the proposed Te-LoRA, based on an alternating freezing framework, not only preserves the training of all LoRA modules but also effectively addresses the issue of aggregation noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 430, + 285, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 285, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 285, + 443 + ], + "type": "text", + "content": "2.3. Federated Fine-tuning Heterogeneous LoRA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 451, + 296, + 714 + ], + "type": "text", + "content": "Due to non-IID data heterogeneity [26] sampled from the Dirichlet distribution and heterogeneous hardware resources, clients need to adjust the LoRA rank accordingly. Existing work has proposed methods to address this issue. For instance, SLoRA [2] addresses LoRA's limitations in high data heterogeneity scenarios with data-driven initialization, achieving performance similar to full finetuning while reducing training time and communication costs. FLoRA [43] uses a stacked aggregation method to eliminate noise in low-rank adapter (LoRA) aggregation, ensuring noise-free global updates. It supports heterogeneous LoRA configurations across clients, allowing dynamic adjustments based on data complexity and resources. However, this approach leads to linear growth in communication complexity, higher computational and storage demands, and reduced flexibility for supporting multiple task-specific adapters. Cho et al. [8] deploy LoRA modules with different ranks across clients, using zero-padding and truncation for aggregation, combining the benefits of high- and low-rank LoRA. However, the padded zeros dilute useful information from high-quality clients, reducing optimization efficiency and potentially limiting perfor" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 322, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 393 + ], + "type": "text", + "content": "mance. FlexLoRA [3] dynamically adjusts the LoRA rank for clients and uses Singular Value Decomposition (SVD) to reallocate global weights, enhancing resource utilization and model generalization. However, SVD truncation adds computational and storage overhead, and the resulting errors may limit performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 396, + 556, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 587 + ], + "type": "text", + "content": "LoRA-A2 [22] alternates freezing the A and B components of the module and dynamically selects key parameters based on the importance of each client's data, effectively reducing parameter inconsistencies during the aggregation of low-rank adapters and significantly cutting communication costs. However, the method assumes that the global LoRA has a uniform initial rank, whereas in practice, each client's LoRA begins with a different rank. Compared to the related LoRA-A2, we train the initial heterogeneous-rank LoRA directly on the clients. While LoRA-A2 focuses on adaptively selecting client ranks, we emphasize the effective alignment of heterogeneous ranks on the server side. During alignment, unlike FlexLoRA and FLoRA, which use merging strategies, we designed the PAA-strategy and T2M-strategy for adaptive alignment, preserving the effective information within clients and dependencies across clients." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 601, + 370, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 370, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 370, + 613 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 623, + 400, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 400, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 400, + 634 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Low-Rank Adaptation. The application of Low-Rank Adaptation (LoRA) [15] facilitates a reduction in the number of trainable parameters during the fine-tuning of a pretrained model. Specifically, the weight matrix " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{d \\times l}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " in the fine-tuned model can be represented as the sum of the original weight matrix " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "W_0 \\in \\mathbb{R}^{d \\times l}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " and an updated" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1060" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "text", + "content": "weight matrix " + }, + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "inline_equation", + "content": "\\triangle W\\in \\mathbb{R}^{d\\times l}" + }, + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "inline_equation", + "content": "\\triangle W" + }, + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "text", + "content": " is implemented by low-rank decomposition into smaller matrices " + }, + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "inline_equation", + "content": "B\\in \\mathbb{R}^{d\\times r}" + }, + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "inline_equation", + "content": "A\\in \\mathbb{R}^{r\\times l}" + }, + { + "bbox": [ + 55, + 72, + 297, + 107 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 115, + 295, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 115, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 110, + 115, + 295, + 129 + ], + "type": "interline_equation", + "content": "W = W _ {0} + \\triangle W = W _ {0} + B A, \\tag {1}", + "image_path": "98e0eada85276b4618d318105efd736b952cb82de17a309e34d18c1d9b7af932.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 135, + 295, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 135, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 295, + 171 + ], + "type": "text", + "content": "where rank " + }, + { + "bbox": [ + 55, + 135, + 295, + 171 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 135, + 295, + 171 + ], + "type": "text", + "content": " is much smaller than both dimensions " + }, + { + "bbox": [ + 55, + 135, + 295, + 171 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 135, + 295, + 171 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 135, + 295, + 171 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 135, + 295, + 171 + ], + "type": "text", + "content": ", thereby significantly reducing the number of trainable parameters for the weights." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "type": "text", + "content": "Aggregation Noise. Currently, FedAvg [30], a widely used FL algorithm, updates the global model by performing a weighted average of " + }, + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "type": "text", + "content": " local clients in each communication round. When LoRA is applied to FL, training and transmitting only the updated low-rank matrices B and A significantly reduces both computational and communication costs, compared to directly updating and transmitting " + }, + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "type": "inline_equation", + "content": "\\triangle W" + }, + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "type": "text", + "content": ". Combining the aforementioned FedAvg and LoRA, FedIT [46] is proposed for federated fine-tuning, where the global LoRA matrices A and B are updated by weighted averaging of the local LoRA matrices " + }, + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 171, + 296, + 315 + ], + "type": "text", + "content": " clients, as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 320, + 295, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 320, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 82, + 320, + 295, + 352 + ], + "type": "interline_equation", + "content": "\\triangle W = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\triangle W _ {i} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} B _ {i} \\times \\frac {1}{n} \\sum_ {i = 1} ^ {n} A _ {i}. \\tag {2}", + "image_path": "7617f94e68c475b61f1b0a905886293222bc503cbe0bea9a050b2e3ddd303768.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 357, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 393 + ], + "type": "text", + "content": "However, when each weight is averaged and aggregated separately, there is a bias between the aggregated parameters (left) and ideal (right), as shown below:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 399, + 295, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 399, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 94, + 399, + 295, + 431 + ], + "type": "interline_equation", + "content": "\\frac {1}{n} \\sum_ {i = 1} ^ {n} B _ {i} \\times \\frac {1}{n} \\sum_ {i = 1} ^ {n} A _ {i} \\neq \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(B _ {i} A _ {i}\\right). \\tag {3}", + "image_path": "a3139c342a50502cef7aa9f2db18730dc1b6f1835475569c803e39c51c39d00a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "text", + "content": "One solution is to first merge " + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "text", + "content": " from each client by matrix multiplication and then perform average aggregation, i.e., " + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\triangle W = \\frac{1}{n}\\sum_{i=1}^{n}(B_{i}A_{i})" + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "text", + "content": ". Subsequently, the aggregated " + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\triangle W \\in \\mathbb{R}^{d\\times l}" + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "text", + "content": " must be decomposed into the new " + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "inline_equation", + "content": "A_{i} \\in \\mathbb{R}^{r\\times l}" + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "inline_equation", + "content": "B_{i} \\in \\mathbb{R}^{d\\times r}" + }, + { + "bbox": [ + 55, + 436, + 296, + 532 + ], + "type": "text", + "content": " for each client. This method introduces complex matrix decomposition, making post-decomposition training prone to instability. The details of another simple and effective solution are as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 532, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 296, + 594 + ], + "type": "text", + "content": "Alternating Freezing Technique. In [22], a simple alternating freezing technique is introduced to mitigate the issue of aggregation noise when applying LoRA in FL. Specifically, in round " + }, + { + "bbox": [ + 55, + 532, + 296, + 594 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 532, + 296, + 594 + ], + "type": "text", + "content": ", the aggregated " + }, + { + "bbox": [ + 55, + 532, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\bar{A}^{(t)}" + }, + { + "bbox": [ + 55, + 532, + 296, + 594 + ], + "type": "text", + "content": " remains frozen, and each client only trains " + }, + { + "bbox": [ + 55, + 532, + 296, + 594 + ], + "type": "inline_equation", + "content": "B_{i}^{(t)}" + }, + { + "bbox": [ + 55, + 532, + 296, + 594 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 80, + 601, + 295, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 601, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 80, + 601, + 295, + 632 + ], + "type": "interline_equation", + "content": "\\frac {1}{n} \\sum_ {i = 1} ^ {n} B _ {i} ^ {(t)} \\times \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\bar {A} ^ {(t)} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(B _ {i} ^ {(t)} \\bar {A} ^ {(t)}\\right). \\tag {4}", + "image_path": "c9d195936cda3bd0adb9979a0ec3876d479a25e5c79adbf7961906e7643c7dd2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 639, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 666 + ], + "type": "text", + "content": "Then, in round " + }, + { + "bbox": [ + 55, + 639, + 296, + 666 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 639, + 296, + 666 + ], + "type": "text", + "content": ", the aggregated " + }, + { + "bbox": [ + 55, + 639, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\bar{B}^{(t + 1)}" + }, + { + "bbox": [ + 55, + 639, + 296, + 666 + ], + "type": "text", + "content": " is frozen, and only " + }, + { + "bbox": [ + 55, + 639, + 296, + 666 + ], + "type": "inline_equation", + "content": "A_{i}^{(t + 1)}" + }, + { + "bbox": [ + 55, + 639, + 296, + 666 + ], + "type": "text", + "content": " is trainable on each client, as shown below:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 672, + 296, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 672, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 672, + 296, + 714 + ], + "type": "interline_equation", + "content": "\\frac {1}{n} \\sum_ {i = 1} ^ {n} \\bar {B} ^ {(t + 1)} \\times \\frac {1}{n} \\sum_ {i = 1} ^ {n} A _ {i} ^ {(t + 1)} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(\\bar {B} ^ {(t + 1)} A _ {i} ^ {(t + 1)}\\right). \\tag {5}", + "image_path": "35416dbbb45230ee581263c3d6c302346719d65c9d25cebae0b29e94a67cc433.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": "Finally, A and B complete iterative training and aggregation through the alternating freezing technique. However, when faced with data heterogeneity and rank heterogeneity at the client, this approach does not achieve good alignment of modules A or B at the server." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 139, + 421, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 139, + 421, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 421, + 152 + ], + "type": "text", + "content": "3.2. Proposed Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 156, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 395 + ], + "type": "text", + "content": "In real FL environments with resource heterogeneity, the training of Low-Rank Adaptation (LoRA) faces two major challenges: aggregation noise and aggregation misalignment. To address these issues, we propose a novel federated fine-tuning method with Tensor-aggregated LoRA (Te-LoRA) based on an alternating freezing training strategy, as shown in Fig. 2. Te-LoRA consists of three strategies: the alternating freezing training strategy, the Pre-Aggregation Alignment Strategy (PAA-strategy), and the Tensor Decomposition Aggregation Strategy (T2M-strategy). Based on the existing alternating freezing training strategy (Eqs. (4) and (5)), Te-LoRA first effectively avoids aggregation noise (Eq. (3)). Subsequently, the PAA-strategy aims to achieve parameter alignment between different rank modules A and B with minimal information loss. Furthermore, the T2M-strategy adaptively aligns and aggregates the main components among clients to achieve more fine-grained feature alignment. The PAA and T2M strategies complement each other, effectively mitigating the issue of aggregation misalignment." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "text", + "content": "Pre-Aggregation Alignment strategy (PAA-strategy). The PAA-strategy is presented to achieve the optimal transformation from heterogeneous LoRAs to homogeneous LoRAs with minimal information loss for parameter alignment on the server side. In the implementation of the alternating freezing FL training strategy, the modules transmitted between the client and server are either B or A. Assuming that the module B is uploaded in the current round, module A remains frozen and stays on the client side. After the upload, the server receives a set of " + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "text", + "content": " modules B, i.e., " + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\{B_k \\in \\mathbb{R}^{d \\times r_k}\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "text", + "content": ". The module with the highest rank is selected as the target module " + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "inline_equation", + "content": "B_{\\max} = B_{\\arg \\max_{k=1,\\dots,K} r_k}" + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "text", + "content": ", and the remaining modules are aligned towards the target module. Specifically, the distance matrix " + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "inline_equation", + "content": "B_k" + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "inline_equation", + "content": "B_{\\max}" + }, + { + "bbox": [ + 313, + 396, + 556, + 579 + ], + "type": "text", + "content": " is calculated as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 370, + 584, + 555, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 584, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 370, + 584, + 555, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {C} (m, n) = \\sum_ {i = 1} ^ {d} \\left(b _ {k} ^ {i m} - b _ {\\max } ^ {i n}\\right), \\tag {6}", + "image_path": "19a65fa040aee96b39539f72fc0aa48213f35be90b76bcc67f2fedba2149cda2.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\in \\mathbb{R}^{r_k \\times r_{\\max}}" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "b_k^{im}" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "b_{\\max}^{in}" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": " are the elements of the " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": " row, " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "m^{\\text{th}}" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": " column of matrix " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "B_k" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "n^{\\text{th}}" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": " column of matrix " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "B_{\\max}" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": ", respectively. To smooth the matching relationship between the source and target matrices, the matrix " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": " is regularized: " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{K} = \\exp\\left(-\\frac{\\mathcal{C}}{\\eta}\\right)" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": " is the regularization parameter. Next, based on the matrix " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": ", we first assume that the source distribution " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "text", + "content": " and the target distribution " + }, + { + "bbox": [ + 313, + 624, + 556, + 715 + ], + "type": "inline_equation", + "content": "q" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1061" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "text", + "content": "follow a uniform distribution, i.e., " + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "inline_equation", + "content": "p = \\left[\\frac{1}{r_k},\\frac{1}{r_k},\\dots \\frac{1}{r_k}\\right]^\\top" + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "inline_equation", + "content": "q = \\left[\\frac{1}{r_{\\mathrm{max}}},\\frac{1}{r_{\\mathrm{max}}},\\dots \\frac{1}{r_{\\mathrm{max}}}\\right]^\\top" + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "text", + "content": ". Then, using the Sinkhorn-Knopp algorithm [31], we iteratively update the adjustment factors " + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "text", + "content": " to make the probability matrix " + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "text", + "content": " converge to a solution that satisfies the marginal constraints of " + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 70, + 296, + 159 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 119, + 166, + 295, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 166, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 119, + 166, + 295, + 190 + ], + "type": "interline_equation", + "content": "\\mu = \\frac {p}{\\mathcal {K} \\cdot \\nu}, \\quad \\nu = \\frac {q}{\\mathcal {K} ^ {\\top} \\cdot \\mu}, \\tag {7}", + "image_path": "0b2f3d47e9328f8a6f7ea9efaaf14f69c8aaca54ef6af35bb2f4e82288dc2d1a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "type": "text", + "content": "where the adjustment factors " + }, + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "type": "text", + "content": " are both 1-column vectors of size " + }, + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "type": "text", + "content": ", respectively. By iteratively updating until convergence, the probability matrix " + }, + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 196, + 296, + 255 + ], + "type": "text", + "content": ", representing the optimal alignment between the source and target features, is shown below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 262, + 295, + 278 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 262, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 141, + 262, + 295, + 278 + ], + "type": "interline_equation", + "content": "\\mathcal {G} = \\mu \\cdot \\mathcal {K} \\cdot \\nu^ {\\top}. \\tag {8}", + "image_path": "94abb66f923c7af772c34d33cd34ea0ec40b400488248b8a782944bcf778798b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 285, + 296, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 321 + ], + "type": "text", + "content": "Subsequently, the source matrix " + }, + { + "bbox": [ + 55, + 285, + 296, + 321 + ], + "type": "inline_equation", + "content": "B_{k}" + }, + { + "bbox": [ + 55, + 285, + 296, + 321 + ], + "type": "text", + "content": " is transformed using the matrix " + }, + { + "bbox": [ + 55, + 285, + 296, + 321 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 285, + 296, + 321 + ], + "type": "text", + "content": " to achieve alignment of different ranks with minimal information loss, as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 328, + 295, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 328, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 146, + 328, + 295, + 343 + ], + "type": "interline_equation", + "content": "\\widetilde {B _ {k}} = B _ {k} \\cdot \\mathcal {G}, \\tag {9}", + "image_path": "f8f3bc8f8810ba1f88dbd305f9b4e914034e0063297a5403e977fefa55784fb0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 350, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 350, + 296, + 625 + ], + "type": "inline_equation", + "content": "\\widetilde{B_k} \\in \\mathbb{R}^{d \\times r_{\\max}}" + }, + { + "bbox": [ + 55, + 350, + 296, + 625 + ], + "type": "text", + "content": " is the aligned matrix. The remaining module B and the module A from another round both use the same pre-aggregation process for parameter alignment. Discussions. Although previous research in FL, such as FedOTP [24] and FedAli [10], has explored Optimal Transport (OT) for fine-grained feature alignment, FedOTP [24] uses OT to align local visual features with global and local text features, smoothing the global consensus and local personalization. To balance personalization and generalization, FedAli [10] quantifies the cost of embedding inputs into prototypes using OT, effectively aligning embeddings and reducing client inconsistencies. Compared to these studies, our PAA strategy differs in three ways: 1) The others focus on data heterogeneity in personalized FL, while our work addresses resource heterogeneity in real FL; 2) Previous work aligns personalized with global knowledge via OT, while our work aligns heterogeneous weights into homogeneous ones using OT to address rank mismatches from resource heterogeneity; 3) Previous studies focus on aligning prompts and prototypes, while our work targets alignment methods for heterogeneous LoRAs. To the best of our knowledge, this is the first research to apply OT for heterogeneous LoRA alignment in resource-heterogeneous FL." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 627, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 296, + 715 + ], + "type": "text", + "content": "Tensor-to-Matrix strategy (T2M-strategy). To maintain the relationships and dependencies between different clients, the Tensor-to-Matrix strategy (T2M-strategy) is designed to adaptively aggregate the LoRAs aligned by the PAA-strategy, rather than simply averaging them. Specifically, given the rank-aligned set " + }, + { + "bbox": [ + 55, + 627, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{B_k}\\}_{k = 1}^K" + }, + { + "bbox": [ + 55, + 627, + 296, + 715 + ], + "type": "text", + "content": " of modules B, " + }, + { + "bbox": [ + 55, + 627, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 627, + 296, + 715 + ], + "type": "text", + "content": " can be stacked from " + }, + { + "bbox": [ + 55, + 627, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{B_k}\\}_{k = 1}^K" + }, + { + "bbox": [ + 55, + 627, + 296, + 715 + ], + "type": "text", + "content": " to form a new tensor" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 324, + 73, + 424, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 73, + 424, + 85 + ], + "spans": [ + { + "bbox": [ + 324, + 73, + 424, + 85 + ], + "type": "text", + "content": "Algorithm 1: Te-LoRA" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 324, + 89, + 555, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 89, + 555, + 114 + ], + "spans": [ + { + "bbox": [ + 324, + 89, + 555, + 114 + ], + "type": "text", + "content": "Initialization: each client " + }, + { + "bbox": [ + 324, + 89, + 555, + 114 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 324, + 89, + 555, + 114 + ], + "type": "text", + "content": " initializes the LoRA " + }, + { + "bbox": [ + 324, + 89, + 555, + 114 + ], + "type": "inline_equation", + "content": "\\triangle W_{k} = B_{k}A_{k}" + }, + { + "bbox": [ + 324, + 89, + 555, + 114 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 324, + 89, + 555, + 114 + ], + "type": "inline_equation", + "content": "B_{k}\\in \\mathbb{R}^{d\\times r_{k}}" + }, + { + "bbox": [ + 324, + 89, + 555, + 114 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 324, + 89, + 555, + 114 + ], + "type": "inline_equation", + "content": "A_{k}\\in \\mathbb{R}^{r_{k}\\times d}" + }, + { + "bbox": [ + 324, + 89, + 555, + 114 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 117, + 555, + 403 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 325, + 117, + 424, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 117, + 424, + 129 + ], + "spans": [ + { + "bbox": [ + 325, + 117, + 424, + 129 + ], + "type": "text", + "content": "1. for " + }, + { + "bbox": [ + 325, + 117, + 424, + 129 + ], + "type": "inline_equation", + "content": "t = 1,2,\\ldots ,T" + }, + { + "bbox": [ + 325, + 117, + 424, + 129 + ], + "type": "text", + "content": " do:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 132, + 474, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 132, + 474, + 144 + ], + "spans": [ + { + "bbox": [ + 324, + 132, + 474, + 144 + ], + "type": "text", + "content": "2. Select K participants for round t" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 324, + 148, + 412, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 148, + 412, + 159 + ], + "spans": [ + { + "bbox": [ + 324, + 148, + 412, + 159 + ], + "type": "text", + "content": "3. if " + }, + { + "bbox": [ + 324, + 148, + 412, + 159 + ], + "type": "inline_equation", + "content": "t\\% 2 = 1" + }, + { + "bbox": [ + 324, + 148, + 412, + 159 + ], + "type": "text", + "content": " then:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 163, + 489, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 163, + 489, + 175 + ], + "spans": [ + { + "bbox": [ + 324, + 163, + 489, + 175 + ], + "type": "text", + "content": "4. for " + }, + { + "bbox": [ + 324, + 163, + 489, + 175 + ], + "type": "inline_equation", + "content": "k = 1,2,\\dots ,K" + }, + { + "bbox": [ + 324, + 163, + 489, + 175 + ], + "type": "text", + "content": " in parallel do:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 177, + 504, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 177, + 504, + 191 + ], + "spans": [ + { + "bbox": [ + 324, + 177, + 504, + 191 + ], + "type": "text", + "content": "5. " + }, + { + "bbox": [ + 324, + 177, + 504, + 191 + ], + "type": "inline_equation", + "content": "B_{k}^{(t + 1)} = \\mathrm{LocalTraining}\\left(B^{(t)},k\\right)" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 324, + 191, + 469, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 191, + 469, + 211 + ], + "spans": [ + { + "bbox": [ + 324, + 191, + 469, + 211 + ], + "type": "text", + "content": "6. " + }, + { + "bbox": [ + 324, + 191, + 469, + 211 + ], + "type": "inline_equation", + "content": "\\widetilde{B}_k^{(t + 1)}\\gets \\mathrm{PAA}\\left(B_k^{(t + 1)}\\right)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 210, + 546, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 210, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 324, + 210, + 546, + 229 + ], + "type": "text", + "content": "7. " + }, + { + "bbox": [ + 324, + 210, + 546, + 229 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_{*}^{(t + 1)}\\gets \\mathrm{T2M}\\left(\\widetilde{B}_{1}^{(t + 1)},\\widetilde{B}_{2}^{(t + 1)},\\dots,\\widetilde{B}_{\\mathrm{K}}^{(t + 1)}\\right)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 228, + 422, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 228, + 422, + 242 + ], + "spans": [ + { + "bbox": [ + 324, + 228, + 422, + 242 + ], + "type": "text", + "content": "8. " + }, + { + "bbox": [ + 324, + 228, + 422, + 242 + ], + "type": "inline_equation", + "content": "\\overline{A}^{(t + 1)} = A^{(t)}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 247, + 384, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 247, + 384, + 258 + ], + "spans": [ + { + "bbox": [ + 324, + 247, + 384, + 258 + ], + "type": "text", + "content": "9. end for" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 262, + 372, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 262, + 372, + 273 + ], + "spans": [ + { + "bbox": [ + 325, + 262, + 372, + 273 + ], + "type": "text", + "content": "10. else:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 325, + 277, + 494, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 277, + 494, + 289 + ], + "spans": [ + { + "bbox": [ + 325, + 277, + 494, + 289 + ], + "type": "text", + "content": "11. for " + }, + { + "bbox": [ + 325, + 277, + 494, + 289 + ], + "type": "inline_equation", + "content": "\\mathrm{k} = 1,2,\\cdots ,\\mathrm{K}" + }, + { + "bbox": [ + 325, + 277, + 494, + 289 + ], + "type": "text", + "content": " in parallel do:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 325, + 291, + 508, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 291, + 508, + 306 + ], + "spans": [ + { + "bbox": [ + 325, + 291, + 508, + 306 + ], + "type": "text", + "content": "12. " + }, + { + "bbox": [ + 325, + 291, + 508, + 306 + ], + "type": "inline_equation", + "content": "A_{k}^{(t + 1)} = \\mathrm{LocalTraining}\\left(A^{(t)},k\\right)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 325, + 305, + 473, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 305, + 473, + 324 + ], + "spans": [ + { + "bbox": [ + 325, + 305, + 473, + 324 + ], + "type": "text", + "content": "13. " + }, + { + "bbox": [ + 325, + 305, + 473, + 324 + ], + "type": "inline_equation", + "content": "\\widetilde{A}_k^{(t + 1)}\\gets \\mathrm{PAA}\\left(A_k^{(t + 1)}\\right)" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 325, + 323, + 550, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 323, + 550, + 343 + ], + "spans": [ + { + "bbox": [ + 325, + 323, + 550, + 343 + ], + "type": "text", + "content": "14. " + }, + { + "bbox": [ + 325, + 323, + 550, + 343 + ], + "type": "inline_equation", + "content": "\\mathbb{A}_{*}^{(t + 1)}\\gets \\mathrm{T2M}\\left(\\widetilde{A}_{1}^{(t + 1)},\\widetilde{A}_{2}^{(t + 1)},\\dots ,\\widetilde{A}_{\\mathrm{K}}^{(t + 1)}\\right)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 325, + 342, + 428, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 342, + 428, + 356 + ], + "spans": [ + { + "bbox": [ + 325, + 342, + 428, + 356 + ], + "type": "text", + "content": "15. " + }, + { + "bbox": [ + 325, + 342, + 428, + 356 + ], + "type": "inline_equation", + "content": "\\overline{B}^{(t + 1)} = B^{(t)}" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 325, + 361, + 389, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 361, + 389, + 372 + ], + "spans": [ + { + "bbox": [ + 325, + 361, + 389, + 372 + ], + "type": "text", + "content": "16. end for" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 325, + 376, + 374, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 376, + 374, + 388 + ], + "spans": [ + { + "bbox": [ + 325, + 376, + 374, + 388 + ], + "type": "text", + "content": "17. end if" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 325, + 392, + 372, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 392, + 372, + 403 + ], + "spans": [ + { + "bbox": [ + 325, + 392, + 372, + 403 + ], + "type": "text", + "content": "18. end for" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "text", + "content": "representing the intermediate state before aggregation, i.e., " + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{B}} = \\bigcup_{k=1}^{K} \\widetilde{B_k}" + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\widetilde{B_k} \\in \\mathbb{R}^{d \\times r_{\\max}}" + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{B}} \\in \\mathbb{R}^{d \\times r_{\\max} \\times K}" + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "text", + "content": ". Here, the aggregation process can be viewed as a tensor decomposition process. Inspired by [21], three versions of the T2M-strategy (T2M" + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "inline_equation", + "content": "_1" + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "text", + "content": ", T2M" + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "inline_equation", + "content": "_2" + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "text", + "content": ", and T2M" + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "inline_equation", + "content": "_3" + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "text", + "content": ") is explored to decompose " + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "text", + "content": " into the result " + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_*" + }, + { + "bbox": [ + 313, + 418, + 555, + 503 + ], + "type": "text", + "content": ". The generalized form of the decomposition is as follows:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 377, + 511, + 553, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 511, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 377, + 511, + 553, + 526 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {B}} = \\mathbb {B} _ {*} \\times_ {1} U \\times_ {2} V \\times_ {3} X, \\tag {10}", + "image_path": "7b8bee512c86386163bbdb52208d8798dab9654a8f8cda5f84ea7bd96dd4d0de.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_*" + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "text", + "content": " is a core tensor, " + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "inline_equation", + "content": "U,V" + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "text", + "content": " serve as the factor matrices, and the symbol " + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "inline_equation", + "content": "\\times_{i}" + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "inline_equation", + "content": "i = 1,2,3" + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "text", + "content": ") denotes mode-" + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "text", + "content": " tensor-matrix multiplication. " + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_1}" + }, + { + "bbox": [ + 313, + 536, + 555, + 572 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 318, + 580, + 553, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 580, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 318, + 580, + 553, + 601 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {B}} = \\mathbb {B} _ {*} ^ {(1)} \\times_ {1} U ^ {(1)} \\times_ {2} V ^ {(1)} = \\left[ \\left[ \\mathbb {B} _ {*} ^ {(1)}; U ^ {(1)}, V ^ {(1)}, \\mathbb {I} \\right] \\right], \\tag {11}", + "image_path": "e1b0c2bca3728161264defa4faddfbd2400ac6bdcd3b0c29e899a8e908d72825.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_{*}^{(1)}\\in \\mathbb{R}^{1\\times 1\\times K}" + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "inline_equation", + "content": "U^{(1)}\\in \\mathbb{R}^{d\\times 1}" + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "inline_equation", + "content": "V^{(1)}\\in \\mathbb{R}^{r_{max}\\times 1}" + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "inline_equation", + "content": "X = \\mathbb{I}" + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "inline_equation", + "content": "K\\times K" + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "text", + "content": " identity matrix. [ ] denotes the shorthand. " + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_{*}^{(1)}" + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "text", + "content": " represents the weight coefficient of different clients, reflecting the importance scores. And the aggregated module is " + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "inline_equation", + "content": "\\bar{B}" + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "inline_equation", + "content": "\\bar{B}_{i,j} = \\sum_{k = 1}^{K}\\hat{\\mathcal{B}}_{i,j,k}\\cdot \\mathbb{B}_{*1,1,k}^{(1)}" + }, + { + "bbox": [ + 313, + 610, + 555, + 688 + ], + "type": "text", + "content": ". The formulae of T2M2 is as follows:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 351, + 695, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 695, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 351, + 695, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\widehat {\\mathcal {B}} = \\mathbb {B} _ {*} ^ {(2)} \\times_ {3} X ^ {(2)} = \\left[ \\left[ \\mathbb {B} _ {*} ^ {(2)}; \\mathbb {I}, \\mathbb {I}, X ^ {(2)} \\right] \\right], \\tag {12}", + "image_path": "7a8a267796fbb86d877ad2a68b447f778136b5f6f781dfaa0d268c66e1b5a537.jpg" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1062" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_*^{(2)}\\in \\mathbb{R}^{d\\times r_{max}\\times 1},X^{(2)}\\in \\mathbb{R}^{K\\times 1},U = \\mathbb{I}" + }, + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "inline_equation", + "content": "V = \\mathbb{I}" + }, + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "inline_equation", + "content": "d\\times d" + }, + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "inline_equation", + "content": "r_{max}\\times r_{max}" + }, + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "text", + "content": " identity matrices. " + }, + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_*^{(2)}" + }, + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "text", + "content": " is the aggregation result capturing the common and representative information, i.e., " + }, + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\bar{B} = \\mathbb{B}_*^{(2)}" + }, + { + "bbox": [ + 55, + 70, + 294, + 133 + ], + "type": "text", + "content": ". T2M3 is a combined version of T2M1 and T2M2, as shown below:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 137, + 294, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 137, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 111, + 137, + 294, + 152 + ], + "type": "interline_equation", + "content": "\\mathbb {B} _ {*} ^ {(3)} = \\mathrm {T} 2 \\mathrm {M} _ {2} (\\widehat {\\mathcal {B}} \\odot \\mathrm {T} 2 \\mathrm {M} _ {1} (\\widehat {\\mathcal {B}})), \\tag {13}", + "image_path": "0982ca897d5a5516d56c6eefc6750e232df7518f2e0628da676236760a8535d8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_{*}^{(1)} = \\mathrm{T2M}_{1}(\\widehat{\\mathcal{B}})" + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "text", + "content": " denotes hadamard product used to obtain the new weighted tensor " + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{B}}^{(1)}\\in \\mathbb{R}^{d\\times r_{\\max}\\times K}" + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{B}}_{i,j,k}^{(1)} = \\widehat{\\mathcal{B}}_{i,j,k}\\cdot \\mathbb{B}_{*1,1,k}^{(1)}" + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "text", + "content": ". The final outcome, " + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_{*}^{(3)}\\in \\mathbb{R}^{d\\times r_{\\max}\\times 1}" + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "text", + "content": ", is the aggregated module, i.e., " + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "inline_equation", + "content": "\\bar{B} = \\mathbb{B}_{*}^{(3)}" + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "text", + "content": ". After completing aggregation on the server side, the module " + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "inline_equation", + "content": "\\bar{B}" + }, + { + "bbox": [ + 55, + 156, + 295, + 273 + ], + "type": "text", + "content": " is distributed to each client for the next round of training. The T2M-strategy is also applicable to the decomposition of module A when module B is frozen for aggregation. Algorithm 1 provides the pseudo-code for Te-LoRA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 282, + 136, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 136, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 136, + 295 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 301, + 162, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 162, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 162, + 314 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 318, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 295, + 521 + ], + "type": "text", + "content": "Comparison methods. We compare the proposed Te-LoRA with three state-of-the-art approaches that support federated fine-tuning of heterogeneous LoRA. \"Zeropadding\" [8] scales the lower ranks of heterogeneous LoRAs from clients by padding them with zeros up to the maximum rank on the server side. \"FLoRA\" [43] stacks all heterogeneous LoRAs together and then performs matrix multiplication for aggregation on the server side. \"FlexLoRA\" [3] aggregates heterogeneous LoRAs after matrix multiplication, then decomposes them into modules A and B via SVD. \"Local\" refers to the average test results from all clients independently training on their local data, serving as an important baseline to validate the effectiveness of FL. To ensure a fair comparison, we reproduced the comparison methods based on the existing codebase [45], with the same LLM, data splits, and training configurations, and conducted all the experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "Datasets and evaluation. We used three datasets to train Te-LoRA: Alpaca-GPT4 [32], an English instruction-following dataset consisting of 52k samples generated by GPT-4 [1] based on Alpaca's self-learning [42]; Databricks-dolly-15k (Dolly) [46], an open-source dataset created by Databricks' staff, which includes tasks such as brainstorming, classification, closed quizzes, open quizzes, and summarization; and Wizard dataset [29] is a quiz dataset including 70k instruction-output pairs. For federated training, datasets were partitioned using a Dirichlet distribution " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "(\\alpha = 0.5)" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": ". Furthermore, we separately consider closed [7] and open evaluation benchmarks, using MMLU [13] for closed evaluation and MT-Bench [48] for open evaluation. MMLU is a widely used Q&A dataset for LLM fine-tuning, including 14,024 questions across 57 different topics to assess the logical reasoning ability of LLMs. MT-Bench is" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 314, + 70, + 560, + 167 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 560, + 167 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 560, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 560, + 167 + ], + "type": "table", + "html": "
MethodMMLUMT-Bench
HomoWizardDollyAlpaca-GPT4Wizard
Local20.8725.1338.662.92
Zero-padding [8]22.0326.5944.883.17
FLoRA [43]22.9127.7345.153.22
FlexLoRA [3]22.9728.2145.343.26
Te-LoRA (Ours)23.3528.4445.863.31
", + "image_path": "676e42c707fbbd2855c66906e5387d1a3971a5c90e8816bef5e5c4a4a1d69079.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 175, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 208 + ], + "type": "text", + "content": "Table 1. Comparison of Te-LoRA with other methods on MMLU and MT-bench. 'Homo' represents settings with homogeneous LoRA ranks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 240 + ], + "type": "text", + "content": "a set of challenging multi-round open-ended questions for assessing chat assistants [48]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 241, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 554, + 396 + ], + "type": "text", + "content": "Training details. For the model configurations, we use the pre-trained Llama2-7B [41] as the base model and conduct 200 rounds of FL communication. The initial learning rate is set to 5e-5 and decays to 1e-6 after 200 rounds using a cosine scheduler. In each round, two clients are randomly selected, each training for 10 epochs with a batch size of 16. Referring to FLoRA's setup [43] in Homogeneous LoRA, the rank of LoRA is 16. Referring to FLoRA's simulation of a real-world scenario where clients have heterogeneous computational resources, in Heterogeneous LoRA, we assign [64, 32, 16, 16, 8, 8, 4, 4, 4, 4] as different local LoRA ranks applied to 10 clients. All experiments are conducted on a single NVIDIA 3090 GPU." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 407, + 453, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 453, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 453, + 420 + ], + "type": "text", + "content": "4.2. Comparison Experiment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 426, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 653 + ], + "type": "text", + "content": "Homogeneous LoRA. In the homogeneous LoRA setup, we configure all clients to share the same LoRA rank of 16. As shown in Tab. 1, we draw the following conclusions: 1) All four federated fine-tuning LoRA methods outperform \"Local\" with Te-LoRA achieving a 7.2 improvement in the MMLU benchmark compared to \"Local\" under AlpacaGPT4 training. 2) Among the comparison methods, Zeropadding, which is affected by aggregation noise, performs the worst. Notably, in homogeneous LoRA, Zero-padding degrades to a traditional approach where modules B and A are aggregated separately on the server. 3) FlexLoRA outperforms FLoRA. Both aggregate the matrix-multiplied weights from modules B and A, but FlexLoRA uses SVD to decompose the aggregated weights and sends new modules to clients, while FLoRA merges the aggregated weights with the client LLM and reinitialized LoRA training. The reinitialized LoRA may cause instability in client training. 4) Te-LoRA outperforms all methods across tasks by effectively resolving aggregation noise and misalignment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "Heterogeneous LoRA. As described in Sec. 4.1, to simulate the scenario of resource heterogeneity among clients in real-world settings, we apply different local LoRA ranks of [64, 32, 16, 16, 8, 8, 4, 4, 4, 4] to 10 clients. By comparing the results of Heterogeneous LoRA (Heter) in Tab. 2" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1063" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 302, + 168 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 302, + 168 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 302, + 168 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 302, + 168 + ], + "type": "table", + "html": "
Method HeterWizardMMLU DollyAlpaca-GPT4MT-Bench Wizard
Local20.5325.1038.542.89
Zero-padding [8]21.8626.4844.473.15
FLoRA [43]22.7927.6044.943.20
FlexLoRA [3]23.4928.1845.483.28
Te-LoRA (Ours)23.7128.3746.163.33
", + "image_path": "5346bed9902febd3d97e417e549602e86b5418b18b2e998e973ceb78fa5502b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 175, + 297, + 208 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 297, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 297, + 208 + ], + "type": "text", + "content": "Table 2. Comparison of Te-LoRA with other methods on MMLU and MT-bench. 'Heter' represents settings with heterogeneous LoRA ranks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 92, + 209, + 265, + 348 + ], + "blocks": [ + { + "bbox": [ + 92, + 209, + 265, + 348 + ], + "lines": [ + { + "bbox": [ + 92, + 209, + 265, + 348 + ], + "spans": [ + { + "bbox": [ + 92, + 209, + 265, + 348 + ], + "type": "image", + "image_path": "a882a4bbec1565049a862cb21dff494fc6d5d29ebbd130dabdce99984e790f1b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 358, + 296, + 382 + ], + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 382 + ], + "type": "text", + "content": "Figure 3. Comparison of the communication cost. Te-LoRA achieves lower overhead compared to the other methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 389, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 296, + 605 + ], + "type": "text", + "content": "with Homogeneous LoRA (Homo) in Tab. 1, we make several interesting observations. First, both Zero-padding and FLoRA show a performance drop in the Heter configuration compared to Homo, with Zero-padding decreasing by up to 0.41 and FLoRA by up to 0.21. Since FLoRA mitigates aggregation noise, it still outperforms Zero-padding. Second, Te-LoRA and FlexLoRA under Heter outperform Homo in most evaluation tasks, demonstrating their ability to effectively handle resource heterogeneity and optimize resource utilization. Because Te-LoRA more effectively addresses the aggregation misalignment issue, it still maintains the best performance under Heter. Finally, Te-LoRA and FlexLoRA show slight performance drops when transitioning from Homo to Heter on the Dolly dataset. This may be due to Dolly's smaller dataset (15k) compared to Alpaca-GPT4 (52k) and Wizard (70k). In the Heter configuration, the 64 and 32 rank LoRA models may overfit on clients with less data, leading to poorer aggregation results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": "Communication overhead. Communication cost plays a crucial role in real-world FL. Therefore, we compared the total upload and download communication costs in a single round across different methods. To ensure a fair comparison, we normalized the transmission parameters. As shown in Fig. 3, the proposed Te-LoRA has the lowest communication cost, as it uploads only module A or B per round. In contrast, FlexLoRA and Zero-padding require both modules to be uploaded and downloaded each round, while" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 182 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 182 + ], + "type": "image", + "image_path": "b59d3aeb35bfebc475f63bdc06fa2eb670132a07db7784e43ed4da720fa6d8db.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 190, + 538, + 202 + ], + "lines": [ + { + "bbox": [ + 329, + 190, + 538, + 202 + ], + "spans": [ + { + "bbox": [ + 329, + 190, + 538, + 202 + ], + "type": "text", + "content": "Figure 4. Impact of different ranks (64, 32, 16, 8, and 4)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 209, + 553, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 553, + 257 + ], + "type": "text", + "content": "FLoRA incurs the highest cost due to transmitting the result of multiplying matrices B and A during download. The quantitative comparison highlights the advantage of the alternating freezing technique." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 257, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 413 + ], + "type": "text", + "content": "Impact of rank. To further investigate the impact of different rank configurations on homogeneous-rank clients, we conducted experiments on the \"Local\" baseline with five rank settings (64, 32, 16, 8, and 4). Training was performed on the Wizard and Alpaca-GPT4 datasets, and evaluation was conducted using MT-Bench and MMLU, respectively. As illustrated in Fig. 4, setting the rank to 4 for all clients, which aligns with the most resource-constrained ones, leads to a significant performance decline. While higher ranks can improve performance, they are not feasible for resource-limited clients. Therefore, adopting a heterogeneous rank configuration enables better resource utilization and further enhances performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 420, + 437, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 437, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 437, + 434 + ], + "type": "text", + "content": "4.3. Ablation Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 533 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the PAA-strategy and T2M-strategy in Te-LoRA, we conducted several ablation experiments. Specifically, \"w/o PPA\" refers to replacing the PAA-strategy with zero-padding in the heterogeneous experiment, which combines the alternating freezing method with zero-padding. \"w/o T2M\" indicates replacing the T2M-strategy with the weighted average aggregation of FedAvg in the heterogeneous experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "The effects of PAA-strategy. To compare with the alignment effect of zero-padding and PAA-strategy, we visualized the cosine similarity between clients after alignment using zero-padding and PAA-strategy, as well as the change curves of the singular value distribution before and after alignment. As shown in Fig. 5, we visualized the cosine similarity between four clients, where most clients exhibit higher cosine similarity after alignment with the PAA strategy compared to zero-padding. This indicates that the PAA-strategy not only aligns the parameters between heterogeneous LoRAs on the server side but also reduces the information loss caused by the alignment. To better illustrate the differences before and after alignment, we compared the normalized singular value distribution curves, as shown in Fig. 6. It can be observed that the distribution of singular" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1064" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 167, + 178 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 167, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 167, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 167, + 178 + ], + "type": "image", + "image_path": "7ef770263f10047c1d54633ddc3902e1bfcada4c1c54588a4331c4793d97191e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 180, + 145, + 191 + ], + "lines": [ + { + "bbox": [ + 77, + 180, + 145, + 191 + ], + "spans": [ + { + "bbox": [ + 77, + 180, + 145, + 191 + ], + "type": "text", + "content": "(a) On PAA-strategy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 70, + 295, + 178 + ], + "blocks": [ + { + "bbox": [ + 186, + 70, + 295, + 178 + ], + "lines": [ + { + "bbox": [ + 186, + 70, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 186, + 70, + 295, + 178 + ], + "type": "image", + "image_path": "222b1392e865779e914184d839452b0446077c1a3ffe6a26d3fdf2b96175725d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 180, + 276, + 191 + ], + "lines": [ + { + "bbox": [ + 206, + 180, + 276, + 191 + ], + "spans": [ + { + "bbox": [ + 206, + 180, + 276, + 191 + ], + "type": "text", + "content": "(b) On Zero-padding." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 233, + 185, + 345 + ], + "blocks": [ + { + "bbox": [ + 56, + 233, + 185, + 345 + ], + "lines": [ + { + "bbox": [ + 56, + 233, + 185, + 345 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 185, + 345 + ], + "type": "image", + "image_path": "916fe394f873563fc333a4c7b2bb75581782655c9a2f9d57f14f88297f76a713.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 348, + 149, + 357 + ], + "lines": [ + { + "bbox": [ + 94, + 348, + 149, + 357 + ], + "spans": [ + { + "bbox": [ + 94, + 348, + 149, + 357 + ], + "type": "text", + "content": "(a) In round four." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 187, + 232, + 294, + 346 + ], + "blocks": [ + { + "bbox": [ + 55, + 200, + 295, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 200, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 295, + 222 + ], + "type": "text", + "content": "Figure 5. Cosine similarity visualization of the local model processed by PAA-strategy or Zero-padding with the global model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 187, + 232, + 294, + 346 + ], + "lines": [ + { + "bbox": [ + 187, + 232, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 187, + 232, + 294, + 346 + ], + "type": "image", + "image_path": "c2418d0de271994115dcbbffb36daed19cd0e7533d00dd048817d9f410097873.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 348, + 272, + 357 + ], + "lines": [ + { + "bbox": [ + 211, + 348, + 272, + 357 + ], + "spans": [ + { + "bbox": [ + 211, + 348, + 272, + 357 + ], + "type": "text", + "content": "(b) In round seven." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 366, + 296, + 400 + ], + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 400 + ], + "type": "text", + "content": "Figure 6. Singular value distribution curves. The source LoRA and those transformed by zero-padding and the PAA strategy undergo SVD to plot the singular value distribution curves." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 410, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 471 + ], + "type": "text", + "content": "values after applying the PAA strategy is closer to that of the original LoRA, while the distribution after zero-padding is relatively farther away. This indicates that the PAA strategy retains more of the major components to a certain extent and is closer to the original LoRA compared to zero-padding." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 472, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 628 + ], + "type": "text", + "content": "To further validate the effect of PPA quantitatively, we conducted an ablation study, as shown in Tab. 3. In all scenarios, \"w/o PAA\" performs worse than " + }, + { + "bbox": [ + 55, + 472, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M}_2" + }, + { + "bbox": [ + 55, + 472, + 295, + 628 + ], + "type": "text", + "content": ", with a performance drop of at least 1.3 points in each case. This validates that the PAA-strategy, compared to Zero-padding, can more effectively reduce information loss when aligning parameters between different levels of the A and B modules. Meanwhile, compared to the FlexLoRA in Tab. 2, \"w/o T2M\" performs better in both the Dolly and AlpacaGPT4 scenarios. This suggests that combining the PAA-strategy and alternating freezing strategy avoids aggregation noise while achieving parameter alignment between different levels of modules A and B, reducing information loss." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "The effects of T2M-strategy. To evaluate the impact of this strategy, as shown in Tab. 3, we performed an ablation experiment with \"w/o T2M\". First, \"w/o T2M\" performs worse in all three scenarios compared to the " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M}_2" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " method, especially in the Alpaca-GPT4 dataset, where w/o T2M is about 0.63 points lower than " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M}_2" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ". This validates that the T2M-strategy, compared to the weighted average" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 167 + ], + "type": "table", + "html": "
Method HeterWizardMMLU DollyAlpaca-GPT4
w/o PAA22.4027.1044.77
w/o T2M23.3828.2145.53
T2M122.9128.1445.30
T2M223.7128.3746.16
T2M323.6728.3045.99
", + "image_path": "5932433fc0182a4a9a1ef5aecc629e483174ad12ae695d342c39b8fd124281a7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 175, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 198 + ], + "type": "text", + "content": "Table 3. Validation of Te-LoRA components. 'Heter' represents settings with heterogeneous LoRA ranks." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 553, + 299 + ], + "type": "text", + "content": "aggregation operation of FedAvg, better preserves the key information from each client, enabling finer feature alignment. In comparison to Zero-padding in Tab. 2, \"w/o PAA\" achieves higher MMLU scores for all three datasets: 0.54, 0.38, and 0.3 points higher on Wizard, Dolly, and Alpaca-GPT4, respectively. This indicates that the T2M-strategy combined with alternating freezing more effectively aligns features while avoiding aggregation noise." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": "The effects of " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_1}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_2}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_3}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": ". As observed from Tab. 3, " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_1}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": " performs slightly worse than " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_2}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": ". In the Wizard, Dolly, and Alpaca-GPT4 datasets, " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_2}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": " scores 0.2, 0.23, and 0.86 points higher than " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_1}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": ", respectively. While " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_3}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": " outperforms " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_1}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": " on all three datasets, it still trails " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_2}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": ", though the results are nearly identical. For example, in the Wizard, Dolly, and Alpaca-GPT4 datasets, " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_2}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": " scores 0.004, 0.007, and 0.005 points higher than " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{T2M_3}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": ". This suggests that although the tensor " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_*^1 \\in \\mathbb{R}^{1 \\times 1 \\times k}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": " obtained from the compression of " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{B}} \\in \\mathbb{R}^{d \\times r_{\\max} \\times k}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": " can preserve the proportion of each client, it is slightly less effective in retaining client information. On the other hand, the tensor " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_* \\in \\mathbb{R}^{d \\times r_{\\max} \\times 1}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": " obtained from the compression of " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": " better preserves the key information from each client and does not introduce redundancy in information like " + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_*^3 \\in \\mathbb{R}^{d \\times r_{\\max} \\times 1}" + }, + { + "bbox": [ + 313, + 299, + 555, + 491 + ], + "type": "text", + "content": " does." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a novel Tensor-aggregation LoRA method (Te-LoRA) to address the noise and misalignment issues in the aggregation of low-rank adaptation within federated learning under resource heterogeneity. Building upon an alternating freezing strategy, Te-LoRA incorporates Pre-Aggregation Alignment (PAA) and Tensor Decomposition Aggregation (T2M) strategies. The PAA strategy utilises OT technology to align the weights of heterogeneous LoRA modules A and B from different clients into homogeneous weights, while the T2M-strategy adaptively aggregates key components to preserve interdependencies. Experimental results demonstrate that Te-LoRA effectively reduces communication costs and significantly improves fine-tuning performance, offering a viable solution for efficient federated fine-tuning in resource-constrained environments." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1065" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "Acknowledgements. The work is supported by the National Natural Science Foundation of China (Grant No. 62222207, 62427808, 62472208, 62276134)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 115, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 115, + 131 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 115, + 131 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 137, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 137, + 295, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 295, + 192 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023.6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 194, + 296, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 247 + ], + "type": "text", + "content": "[2] Sara Babakniya, Ahmed Roushdy Elkordy, Yahya H Ezzeldin, Qingfeng Liu, Kee-Bong Song, Mostafa El-Khamy, and Salman Avestimehr. Slora: Federated parameter efficient fine-tuning of language models. arXiv preprint arXiv:2308.06522, 2023.1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 248, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 248, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 248, + 295, + 293 + ], + "type": "text", + "content": "[3] Jiamu Bai, Daoyuan Chen, Bingchen Qian, Liuyi Yao, and Yaliang Li. Federated fine-tuning of large language models under heterogeneous language tasks and client resources. arXiv e-prints, pages arXiv-2402, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 293, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 294, + 326 + ], + "type": "text", + "content": "[4] Desirée Bill and Theodor Eriksson. Fine-tuning a llm using reinforcement learning from human feedback for a therapy chatbot application, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 296, + 392 + ], + "type": "text", + "content": "[5] Daoyuan Chen, Yilun Huang, Zhijian Ma, Hesen Chen, Xuchen Pan, Ce Ge, Dawei Gao, Yuexiang Xie, Zhaoyang Liu, Jinyang Gao, et al. Data-juicer: A one-stop data processing system for large language models. In Companion of the 2024 International Conference on Management of Data, pages 120-134, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 295, + 435 + ], + "type": "text", + "content": "[6] Yukang Chen, Shengju Qian, Haotian Tang, Xin Lai, Zhijian Liu, Song Han, and Jiaya Jia. Longlora: Efficient finetuning of long-context large language models. arXiv preprint arXiv:2309.12307, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 437, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 295, + 480 + ], + "type": "text", + "content": "[7] Yew Ken Chia, Pengfei Hong, Lidong Bing, and Soujanya Poria. Instructeval: Towards holistic evaluation of instruction-tuned large language models. arXiv preprint arXiv:2306.04757, 2023. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 524 + ], + "type": "text", + "content": "[8] Yae Jee Cho, Luyang Liu, Zheng Xu, Aldi Fahrezi, and Gauri Joshi. Heterogeneous lora for federated fine-tuning of ondevice foundation models. arXiv preprint arXiv:2401.06432, 2024. 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 525, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 581 + ], + "type": "text", + "content": "[9] Ning Ding, Yujia Qin, Guang Yang, Fuchao Wei, Zonghan Yang, Yusheng Su, Shengding Hu, Yulin Chen, Chi-Min Chan, Weize Chen, et al. Parameter-efficient fine-tuning of large-scale pre-trained language models. Nature Machine Intelligence, 5(3):220-235, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 582, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 295, + 624 + ], + "type": "text", + "content": "[10] Sannara Ek, Kaile Wang, François Portet, Philippe Lalanda, and Jiannong Cao. Fedali: Personalized federated learning with aligned prototypes through optimal transport. arXiv preprint arXiv:2411.10595, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 658 + ], + "type": "text", + "content": "[11] Junyao Gao, Yanan Sun, Fei Shen, Xin Jiang, Zhening Xing, Kai Chen, and Cairong Zhao. Faceshot: Bring any character into life. arXiv preprint arXiv:2503.00740, 2025. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[12] Lucas Grativol, Mathieu Leonardon, Guillaume Muller, Virginie Fresse, and Matthieu Arzel. Flocora: Federated learning compression with low-rank adaptation. In 2024 32nd European Signal Processing Conference (EUSIPCO), pages 1786-1790. IEEE, 2024. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "text", + "content": "[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "text", + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[16] Ahmed Imteaj, Urmish Thakker, Shiqiang Wang, Jian Li, and M Hadi Amini. A survey on federated learning for resource-constrained IoT devices. IEEE Internet of Things Journal, 9(1):1-24, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 253, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 306 + ], + "type": "text", + "content": "[17] Xin Jiang, Hao Tang, Junyao Gao, Xiaoyu Du, Shengfeng He, and Zechao Li. Delving into multimodal prompting for fine-grained visual classification. In Proceedings of the AAAI conference on artificial intelligence, pages 2570-2578, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[18] Xin Jiang, Hao Tang, and Zechao Li. Global meets local: Dual activation hashing network for large-scale fine-grained image retrieval. IEEE Transactions on Knowledge and Data Engineering, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 354, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 407 + ], + "type": "text", + "content": "[19] Xin Jiang, Hao Tang, Rui Yan, Jinhui Tang, and Zechao Li. Dvf: Advancing robust and accurate fine-grained image retrieval with retrieval guidelines. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2379-2388, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 409, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 409, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 317, + 409, + 553, + 463 + ], + "type": "text", + "content": "[20] Dominique Kelly, Yimin Chen, Sarah E Cornwell, Nicole S Delellis, Alex Mayhew, Sodiq Onaolapo, and Victoria L Rubin. Bing chat: The future of search engines? Proceedings of the Association for Information Science and Technology, 60(1):1007-1009, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 465, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 465, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 465, + 553, + 487 + ], + "type": "text", + "content": "[21] Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. SIAM review, 51(3):455-500, 2009. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 488, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 553, + 531 + ], + "type": "text", + "content": "[22] Jabin Koo, Minwoo Jang, and Jungseul Ok. Towards robust and efficient federated low-rank adaptation with heterogeneous clients. arXiv preprint arXiv:2410.22815, 2024. 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "text", + "content": "[23] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "text", + "content": "[24] Hongxia Li, Wei Huang, Jingya Wang, and Ye Shi. Global and local prompts cooperation via optimal transport for federated learning. In CVPR, pages 12151-12161, 2024. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "text", + "content": "[25] Pengpeng Li, Xiangbo Shu, Chun-Mei Feng, Yifei Feng, Wangmeng Zuo, and Jinhui Tang. Surgical video workflow analysis via visual-language learning. npj Health Systems, 2 (1):5, 2025. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[26] Xiang Li, Kaixuan Huang, Wenhao Yang, Shusen Wang, and Zhihua Zhang. On the convergence of fedavg on non-iid data. arXiv preprint arXiv:1907.02189, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[27] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1066" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[28] Zheng Lin, Xuanjie Hu, Yuxin Zhang, Zhe Chen, Zihan Fang, Xianhao Chen, Ang Li, Praneeth Vepakomma, and Yue Gao. Splitlora: A split parameter-efficient fin-tuning framework for large language models. arXiv preprint arXiv:2407.00952, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 194 + ], + "type": "text", + "content": "[29] Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 250 + ], + "type": "text", + "content": "[30] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Aguera y Arcas. Communication-efficient learning of deep networks from decentralized data. In Artificial intelligence and statistics, pages 1273-1282. PMLR, 2017. 1, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 284 + ], + "type": "text", + "content": "[31] Edouard Pauwels and Samuel Vaiter. The derivatives of sinkhorn-knopp converge. SIAM Journal on Optimization, 33(3):1494-1517, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "text", + "content": "[32] Baolin Peng, Chunyuan Li, Pengcheng He, Michel Galley, and Jianfeng Gao. Instruction tuning with gpt-4. arXiv preprint arXiv:2304.03277, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 374 + ], + "type": "text", + "content": "[33] Zhen Qin, Daoyuan Chen, Bingchen Qian, Bolin Ding, Yaliang Li, and Shuiguang Deng. Federated full-parameter tuning of billion-sized language models with communication cost under 18 kilobytes. arXiv preprint arXiv:2312.06353, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 421 + ], + "type": "text", + "content": "[34] Riccardo Salami, Pietro Buzzega, Matteo Mosconi, Jacopo Bonato, Luigi Sabetta, and Simone Calderara. Closed-form merging of parameter-efficient modules for federated continual learning. arXiv preprint arXiv:2410.17961, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 475 + ], + "type": "text", + "content": "[35] Xiangbo Shu, Jinhui Tang, Guo-Jun Qi, Wei Liu, and Jian Yang. Hierarchical long short-term concurrent memory for human interaction recognition. IEEE transactions on pattern analysis and machine intelligence, 43(3):1110-1118, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 532 + ], + "type": "text", + "content": "[36] Xiangbo Shu, Liyan Zhang, Guo-Jun Qi, Wei Liu, and Jinhui Tang. Spatiotemporal co-attention recurrent neural networks for human-skeleton motion prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3300-3315, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "text", + "content": "[37] Karan Singhal, Shekoofeh Azizi, Tao Tu, S Sara Mahdavi, Jason Wei, Hyung Won Chung, Nathan Scales, Ajay Tanwani, Heather Cole-Lewis, Stephen Pfohl, et al. Large language models encode clinical knowledge. Nature, 620 (7972):172-180, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "text", + "content": "[38] Tianyu Song, Shumin Fan, Pengpeng Li, Jiyu Jin, Guiyue Jin, and Lei Fan. Learning an effective transformer for remote sensing satellite image dehazing. IEEE Geoscience and Remote Sensing Letters, 20:1-5, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "text", + "content": "[39] Youbang Sun, Zitao Li, Yaliang Li, and Bolin Ding. Improving lora in privacy-preserving federated learning. CoRR, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[40] Jinhui Tang, Xiangbo Shu, Rui Yan, and Liyan Zhang. Coherence constrained graph listm for group activity recognition. IEEE transactions on pattern analysis and machine intelligence, 44(2):636-647, 2019. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 529 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "type": "text", + "content": "[41] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 553, + 173 + ], + "type": "text", + "content": "[42] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instructions. arXiv preprint arXiv:2212.10560, 2022. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 175, + 553, + 228 + ], + "type": "text", + "content": "[43] Ziyao Wang, Zheyu Shen, Yexiao He, Guoheng Sun, Hongyi Wang, Lingjuan Lyu, and Ang Li. Flora: Federated fin-tuning large language models with heterogeneous low-rank adaptations. arXiv preprint arXiv:2409.05976, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 230, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 230, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 315, + 230, + 553, + 283 + ], + "type": "text", + "content": "[44] Feijie Wu, Zitao Li, Yaliang Li, Bolin Ding, and Jing Gao. Fedbiot: Llm local fine-tuning in federated learning without full model. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 3345-3355, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 285, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 285, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 315, + 285, + 553, + 351 + ], + "type": "text", + "content": "[45] Rui Ye, Wenhao Wang, Jingyi Chai, Dihan Li, Zexi Li, Yinda Xu, Yaxin Du, Yanfeng Wang, and Siheng Chen. Openfedllm: Training large language models on decentralized private data via federated learning. In Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 6137-6147, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 353, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 353, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 315, + 353, + 553, + 418 + ], + "type": "text", + "content": "[46] Jianyi Zhang, Saeed Vahidian, Martin Kuo, Chunyuan Li, Ruiyi Zhang, Tong Yu, Guoyin Wang, and Yiran Chen. Towards building the federatedgpt: Federated instruction tuning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6915-6919. IEEE, 2024. 1, 2, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 419, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 419, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 315, + 419, + 553, + 473 + ], + "type": "text", + "content": "[47] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. Adalora: Adaptive budget allocation for parameter-efficient fine-tuning. arXiv preprint arXiv:2303.10512, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 475, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 475, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 315, + 475, + 553, + 529 + ], + "type": "text", + "content": "[48] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1067" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/c9c92fd3-e6cf-49ea-bfb4-aa3c5b130a2c_content_list.json b/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/c9c92fd3-e6cf-49ea-bfb4-aa3c5b130a2c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2470e37685aa1991774d44992341e84a4d87f0cf --- /dev/null +++ b/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/c9c92fd3-e6cf-49ea-bfb4-aa3c5b130a2c_content_list.json @@ -0,0 +1,1470 @@ +[ + { + "type": "text", + "text": "TerraMind: Large-Scale Generative Multimodality for Earth Observation", + "text_level": 1, + "bbox": [ + 125, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a508ecc15dcc21c6803d3b04d4f20a8dcabf19832964f2fe4ee1a9d82220c68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 178, + 908, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ IBM Research - Europe $^{2}$ ETH Zurich $^{3}$ Forschungszentrum Jülich $^{4}$ European Space Agency $\\Phi$ -Lab $^{5}$ NASA IMPACT $^{6}$ University of Iceland", + "bbox": [ + 199, + 256, + 808, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "johnannes.jakubikl@ibm.com", + "bbox": [ + 393, + 296, + 617, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c2719cca28b3707036b1e21c1c8178c88eade472b909cf3ef0639ccd403e1524.jpg", + "image_caption": [ + "Figure 1. TerraMind represents the first any-to-any generative, and large-scale multimodal model for Earth observation pre-trained on 500 billion tokens from global geospatial data. The model digests multi-scale representations at pixel-level and token-level simultaneously. TerraMindv1 unlocks (i) generation, (ii) zero-shot and finetuning applications, and (iii) \"Thinking-in-Modalities\" finetuning and inference." + ], + "image_footnote": [], + "bbox": [ + 89, + 352, + 651, + 648 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6bb9405779306eee986642cf6d11a4df535fdd5d3976835632bf50bec8e86bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 352, + 905, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f7c9a63048871fd77a2eec6e0ea9896413b1fa97b6cae2bce0e8cf481bc98f2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 465, + 905, + 647 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 715, + 326, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present TerraMind, the first any-to-any generative, multimodal deep learning model for Earth observation (EO). Unlike other approaches, TerraMind is pretrained on dual-scale representations combining both token-level and pixel-level data across modalities. On a token level, TerraMind encodes high-level contextual information to learn cross-modal relationships, while on a pixel level, TerraMind lever", + "bbox": [ + 89, + 750, + 486, + 856 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ages fine-grained representations to capture critical spatial nuances. In this paper, we demonstrate that (i) TerraMind achieves beyond state-of-the-art performance in community-standard benchmarks, (ii) TerraMind can leverage \"thinking in modalities\" (TiM)—the capability of generating additional artificial data during finetuning and inference to improve the model output—and (iii) TerraMind's dual-scale early fusion approach results in well-structured embedding spaces. Models and code have been open-sourced at https://huggingface.co.ibm-esa-geospatialandhttps://github.com.ibm/terramind.", + "bbox": [ + 511, + 718, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 91, + 875, + 204, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Equal supervision", + "bbox": [ + 91, + 888, + 199, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7383", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Earth observation (EO) increasingly benefits from multimodality because of the important integration of complementary information from different data sources. This becomes particularly relevant as EO is spatiotemporally sparse due to low revisiting times or weather phenomena like cloud coverage. Vice versa, for computer vision, EO data is an important playground for the development of new approaches as there is significant publicly available data of very high quality and complexity. The available modalities range from sensors of different satellite missions to relevant complementary information like digital elevation.", + "bbox": [ + 89, + 119, + 485, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce TerraMind as the first any-to-any generative multimodal model for EO. With TerraMind, we introduce a dual-scale pretraining on pixel-level and token-level and demonstrate benefits over training primarily on tokens. TerraMind encodes high-level contextual information in tokens to enable correlation learning and scaling, while, additionally capturing important fine-grained representations using pixel-level inputs. During pretraining, TerraMind predicts masked target tokens so that our pretraining objective boils down to a cross-modal patch classification problem that results in high-quality latent spaces. TerraMind is pretrained on a custom global-scale geospatial dataset named TerraMesh with nine million samples that have been aligned spatiotemporally and across modalities [7]. In addition to radar and optical satellite images of the Copernicus Sentinel-1 (S-1) and Sentinel-2 (S-2) missions, our dataset contains task-specific modalities such as land use/land cover (LULC) and normalized difference vegetation index (NDVI) maps, metadata like digital elevation models (DEM) and geographic coordinates, and natural language in the form of captions. To the best of our knowledge, TerraMind represents the first truly generative, multimodal deep learning model for EO. Additionally, in contrast to other recent models that utilize masked autoencoders like [54], contrastive learning, or diffusion techniques, TerraMind uniquely demonstrates benefits of leveraging token-based pretraining for EO.", + "bbox": [ + 91, + 291, + 483, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We provide an overview of TerraMind's performance in a community-standard benchmark [49] in Figure 2 and highlight the any-to-any generative capabilities of TerraMind in Figure 3. Our key contributions are as follows: (i) We introduce a dual-scale approach for generative multimodal pre-training leveraging data on pixel-level and token-level, which outperforms other fusion approaches and enhances embedding space structures. (ii) We introduce thinking in modalities - similar to chain-of-thought approaches in LLMs - for multi-modal models in EO, demonstrating that infusing generated data during finetuning improves the performance. (iii) We demonstrate that TerraMind outperforms other geospatial foundation models both in unimodal and multimodal settings.", + "bbox": [ + 88, + 689, + 486, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 656, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Computer vision in Earth observation. Computer vision (CV) has significantly advanced EO [76]. Many CV techniques, originally developed for natural image processing, have been adapted to EO [62], often with minimal modifications. A wide range of tasks benefit from these methods, including classification [16], semantic segmentation [72] (e.g., land cover mapping [20, 21]), change detection [59] (e.g., disaster response [19]), object detection [39] (e.g., vessel identification [55]), and regression (e.g., biomass estimation [53]). Deep learning architectures like CNNs [75] and Vision Transformers (ViTs) [17] have demonstrated strong performance, often surpassing traditional remote sensing (RS) methods. However, EO presents unique challenges, including diverse sensor modalities [4] and geospatial heterogeneity [46]. An emerging paradigm in EO is self-supervised learning (SSL) [64] and geospatial foundation models (GFMs) [45], which aim to leverage vast amounts of unlabeled RS data to develop general purpose task models [32]. While off-the-shelf CV models have shown promising results [36], they do not fully exploit the unique characteristics of geospatial data. Many GFMs still rely on generic CV architectures [50], which were not explicitly designed to handle the complexities of EO, such as heterogeneous sensor sources (e.g., optical, radar, DEM) [29], integrated with auxiliary data (e.g., text) [42, 47], and expert knowledge (e.g., prioritizing specific bands or indexes). In this direction, TerraMind better integrates domain-specific properties, developing a customized and expandable multimodal learning strategy.", + "bbox": [ + 511, + 114, + 906, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodality in CV. Multimodal CV is driven by the integration of diverse data streams [69], such as natural images [74], natural language text [10], temporal video data [58], and weather [70], within large foundation models [8].", + "bbox": [ + 511, + 539, + 908, + 602 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/89ab3f4a24a8f0f22c145b98d7aa131473829751dc280f48d7fdfe81a559cdff.jpg", + "image_caption": [ + "Figure 2. TerraMind outperforms other geospatial foundation models on PANGAEA benchmark [49] in finetuning. Performance is measured in mIoU and min-max scaled per dataset." + ], + "image_footnote": [], + "bbox": [ + 516, + 625, + 900, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7384", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/26ad38ec634ed36b3b55e5dbf6b8398dd71b67db16bdc57353ddbf06d8c0c07f.jpg", + "image_caption": [ + "Figure 3. Chained generation example of TerraMindv1-B starting from either optical, radar, or digital elevation data. Left is input, middle is artificially generated data by TerraMind, right represents ground truths and tokenizer reconstructions, respectively." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 908, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Starting from the alignment of images and texts [57], these models moved beyond simple feature extraction, towards nuanced contextual understanding. The ability to combine several modalities allows for unprecedented capabilities in complex tasks [30], evidenced by the rapid advancement of multimodal Large Language Models (MLLMs) [30], that excel in tasks such as scene understanding [12], visual question answering [18], and video analysis [24]. Recent advances in architectures [31] and large scale pre-training [11] have enabled the development of models that learn highly effective cross-modal representations [41], which can then be adapted to a wide variety of downstream tasks [66].", + "bbox": [ + 88, + 333, + 483, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodality in EO. Multimodality in EO originates from data fusion and is typically understood as the integration of SAR and optical data [13, 25, 28, 38] or the combination of optical data with vector data [5]. Some studies have explored alternative combinations of data. In [15], the authors introduce a contrastive framework for comparing RS images and street views. Even different optical sensors can be considered different modalities [48, 61]. Similarly, several multi-view images (i.e. multimodal) datasets [26, 44, 54] are introduced. More recent approaches combined text and images [40], both for discriminative [42] and generative [34] purposes. Lately, different GFMs are trained in a multimodal way [4, 54, 68], still focusing either on a specific set of modalities (e.g., vision [68], [3]) or tasks (e.g., generative [34]). Compared to multi-scale high-quality generation models for optical data, like MetaEarth [71], our approach allows to generate any modality from any other pretraining modality. To the best of our knowledge, no existing model has combined a wide and diverse amount of modalities both for discriminative and generative purposes, as TerraMind does. We provide a comparison in Table 1.", + "bbox": [ + 89, + 515, + 483, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset", + "text_level": 1, + "bbox": [ + 89, + 844, + 181, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the pretraining of TerraMind and its tokenizers, we create a multimodal dataset called TerraMesh [7], which will", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/e99accac7b0552322e0145e2b418e3c374d12d577dbd6d217a88426840df1874.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelModalitiesAny-to-Any GenerationMulti-Scale Features
RemoteCLIPoptical, textXX
CROMAoptical, radarXX
AnySataerial, optical, radar, NAIPXX
DeCURoptical, radarXX
DOFAoptical, radar, hyperspectral, NAIPXX
MetaEarthoptical (unimodal)X✓
Galileooptical, radar, elevation, weather, location, population, ...X✓
TerraMindoptical, radar, land use, elevation, vegetation index, location, text✓✓
", + "bbox": [ + 514, + 329, + 906, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison of TerraMind to other model architectures. TerraMind represents a first-of-its-kind generative, multimodal model.", + "bbox": [ + 511, + 609, + 906, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "be open-sourced to the community. TerraMesh builds on existing datasets, which we expand by adding modalities from external data sources or by applying pseudo-labeling. We provide an overview of the aligned image modalities and a detailed dataset description in the supplementary material.", + "bbox": [ + 511, + 671, + 906, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Base datasets. TerraMesh is based on SSL4EO-S12 [6, 65] and MajorTOM-Core [23], two unlabeled remote sensing datasets containing co-aligned radar and optical imagery from Sentinel-1 and Sentinel-2 satellites. SSL4EO-S12 has lower geographic coverage but is multi-seasonal. MajorTOM-Core covers most of the Earth's land surface at a single timestamp. For MajorTOM-Core, we apply a subsampling scheme based on LULC classes and ecoregions. TerraMesh includes a total of approximately 9 million globally distributed samples from both Sentinel-1 and Sentinel-2,", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7385", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "each measuring $264 \\times 264$ pixels at $10\\mathrm{m}$ resolution.", + "bbox": [ + 89, + 90, + 428, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additional modalities. We obtain co-aligned yearly LULC maps by ESRI with nine land use classes. Additionally, we leverage SEnSeI v2 [22] as a cloud and ice annotation model and update the ESRI LULC classes for better spatiotemporal alignment. NDVI maps are computed using the corresponding spectral bands from Sentinel-2. DEM is extracted from the Copernicus DEM 30m dataset [2], which provides global coverage of the Earth's elevation at a 30m resolution. Captions are generated synthetically by constructing RGB images from Sentinel-2 patches using the corresponding spectral bands and processing them with LLaVANext [37]. A tailored prompt guides the model to describe the content of each image as described in [47]. For geolocations, we round latitude and longitude from the center of each patch to the nearest quarter degree and store the discretized coordinates as strings in a pre-defined format.", + "bbox": [ + 89, + 106, + 485, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 89, + 363, + 191, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TerraMind pretraining is two-staged following [52]. We first pretrain unimodal tokenizer models, tokenize the modalities, and then leverage token-level and pixel-level input to pretrain the TerraMind encoder-decoder architecture. We describe those individual stages in the following.", + "bbox": [ + 89, + 388, + 483, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Tokenization", + "text_level": 1, + "bbox": [ + 89, + 474, + 227, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We develop modality-specific tokenizers to encode each modality into a sequence of discrete tokens for pretraining and decode token sequences back to images. Thus, TerraMind is in principle compatible with any modality, as long as it can be tokenized and aligned with other modalities. For reasons of space, we delegate most experiments related to the tokenizer performances to the supplementary material.", + "bbox": [ + 89, + 497, + 483, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-like modalities. We train autoencoder-based architectures with a quantization step in the bottleneck for image-like modalities such as S-1, S-2, LULC, NDVI, and DEM. Tokenizer encoders process an input image and generate a latent representation for each $16 \\times 16$ patch, which is then discretized with finite-scalar-quantization (FSQ) [51] into one of $N$ codewords. All tokenizers use a vocabulary size of 16K besides the simpler LULC modality for which we use 4K. These codewords are then used by the diffusion decoder to reconstruct the original image. The benefit of leveraging diffusion decoders lies in facilitating cross-modal generation in TerraMind by transforming tokens back into images. By mapping each codeword to a unique integer in $\\{0, 1, \\dots, N - 1\\}$ , we obtain discrete tokens for each image patch. We pretrain the tokenizers in a self-supervised setting. FSQ as quantization method enhances training stability [51] compared to vector quantization [63] by eliminating the need for codebook-related loss terms. Notably, FSQ is", + "bbox": [ + 89, + 604, + 483, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "heavily influenced by ideas of neural compression [27]. For example, on 12-band S-2 images, we achieve compression rates of over $3000\\mathrm{x}$ by applying quantization. We summarize the architecture of our tokenizers in Figure 4. The main objective of the overall tokenizer is to encode image patches consistently into discrete tokens based on semantic similarity to enable cross-modal correlation learning. Therefore, the loss of some details is an expected trade-off, since the focus is on grouping similar patches rather than preserving all fine-grained features. Naturally, more accurate reconstructions facilitate cross-modal generation, however the main focus of the pretraining lies on consistent cross-modal correlation learning. We provided further details on the pretraining of the tokenizers in the supplementary material.", + "bbox": [ + 511, + 90, + 908, + 303 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/71da8b1dfb813f0f4d5e280206571aa38f59c2eb458f021352749f9465058d54.jpg", + "image_caption": [ + "Figure 4. Tokenizer for image-like modalities combining finite-scalar quantization [51] with diffusion decoding." + ], + "image_footnote": [], + "bbox": [ + 519, + 318, + 906, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sequence-like modalities. We treat both captions and geolocations as text and use a single text tokenizer to process both modalities. By discretizing the geographic coordinates and representing them as strings, we introduce special coordinate tokens into the vocabulary. This allows us to encode geolocations as a sequence of discrete tokens, beginning with a latitude token followed by a longitude token. For textual data, we modify the existing WordPiece tokenizer [33].", + "bbox": [ + 511, + 476, + 908, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Pre-training", + "text_level": 1, + "bbox": [ + 511, + 606, + 645, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture. TerraMind uses a symmetric Transformer-based encoder-decoder architecture proposed in [52], which is designed to process sequences of multimodal tokens. In addition to discrete tokens, TerraMind accepts pixel-level inputs, specifically satellite imagery and digital elevation maps. For pixel-level inputs, we apply learnable patch-wise linear projections to generate patch embeddings for each $16 \\times 16$ patch, similar to the approach used in ViT [17].", + "bbox": [ + 511, + 628, + 908, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dual-scale early fusion. In contrast to [52], we not only embed token-level data but additionally leverage pixel-level data across a range of input modalities to introduce a dual-scale feature representation to support the structuring of the embedding space. Both tokens and patches represent a 16x16 pixel area. Tokens represent this area via a single discrete integer value, while the image patches describe the same area with the actual floating point sensor data. Thus, during pretraining, the model not only learns a correlation between modalities (i.e., cross-modal learning) but also between dif", + "bbox": [ + 511, + 750, + 910, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "https://planetarycomputer.microsoft.com/dataset/io-lulc-annual-v02", + "bbox": [ + 91, + 888, + 423, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7386", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ferent levels of abstraction within the same modality. The low-level token information enables cross-modal correlation learning, while adding pixel level input accounts for spatial nuances. Based on dual-scale features the model further learns to better structure pixel-level data in the embedding space via the corresponding information from the discrete token. We illustrate the pretraining paradigm in Figure 5. The model is agnostic to processing tokens or patches in the input space, while the target is generally token-level data. We use six pixel-level modalities and eight token-level modalities.", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e0986ca258c6aa8b5d527085e8ce35c951950eed3bc2982fada04e93f0661922.jpg", + "image_caption": [ + "Figure 5. Illustration of the pre-training task. Given an encoded multimodal sample of random subsets of patches and input tokens, the decoder predicts target tokens for the masked input." + ], + "image_footnote": [], + "bbox": [ + 91, + 256, + 472, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Masking strategy. TerraMind applies a masked modeling approach in the token space following [52]. The model leverages a set of randomly selected target tokens that have to be reconstructed from a randomly selected set of input tokens and pixel-level data. During pre-training, we sample input and target data from a Dirichlet distribution.", + "bbox": [ + 89, + 438, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We opt for masked token reconstruction to familiarize the model with the absence of entire modalities, which is crucial for a high usability of a multimodal model in Earth observation. During pre-training, the model learns an internal representation of unseen modalities which is expected to benefit a range of downstream applications. In addition, sampling input and target tokens improves the computational efficiency of the pre-training, as each token is a compressed representation of a patch with compression factors of between 250x and 3000x depending on the modality. Finally, without tokenized representations of the image-like modalities, it is challenging to learn the correlation to sequence-like modalities. The overall training objective of TerraMind boils down to a cross-modal patch-level classification problem optimized via a cross entropy loss:", + "bbox": [ + 89, + 529, + 483, + 755 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} = - \\sum_ {i = 1} ^ {N} y _ {i} \\log \\left(p _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 763, + 483, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{i}$ is the one-hot encoded true class of token $i$ , $p_{i}$ is the predicted probability for token $i$ , $N$ is the total number of possible tokens. Interestingly, we can infer an upper bound loss for a random model where the cross entropy loss will collapse to the natural logarithm of the vocabulary size $\\mathcal{L}_{\\mathrm{CE,random}} = -\\sum_{i=1}^{N} y_{i} \\log \\left( \\frac{1}{N} \\right) = \\log(N)$ .", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scaling. We trained three versions of TerraMind scaling across model size, compute, and data. In addition, we pretrain different versions of TerraMind with respect to the number of dual-scale features. TerraMindv1-B is pre-trained on 500B tokens for 6 days on 32 NVIDIA A100 GPUs. The model uses dual-scale features from both token-level and pixel-level. During initial experiments, we observed significant improvements from scaling model size when switching from a tiny backbone to a small backbone to a base backbone. Therefore, we pre-trained TerraMindv1-L on a large backbone with 500B tokens on 32 NVIDIA A100 GPUs trained for 10 days. Finally, to better understand the effect of scaling across the dual-scale feature representation, we pre-train TerraMindv1-B-single as a single-scale model on primarily token-level data with optical S-2 L2A data as only pixel-level input (compared to pixel-level S-2 L1C, S-2 RGB, S-1 GRD, S-1 RTC, and DEM in TerraMindv1-B and -L). TerraMindv1-B-single is pretrained on 500B tokens from over one million samples for 6 days on 32 NVIDIA A100 GPUs. We summarize the scaling behavior in model size, compute, and data in Figure 9 of the supplementary material. We additionally provide final validation losses in Table 9 comparing v1-B and v1-L with the theoretical random loss.", + "bbox": [ + 511, + 90, + 906, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Generation", + "text_level": 1, + "bbox": [ + 511, + 450, + 637, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once pretrained, TerraMind can generate tokens for any modality, conditioned on any subset of input modalities. The generative capabilities unlock various zero-shot tasks, such as water body segmentation. For the generation of image-like modalities, the decoder receives mask tokens for the modality to be generated and predicts the corresponding tokens based on the encoded input. For sequence-like modalities, the decoder generates the output autoregressively. After generating tokens from the target modality, the corresponding tokenizer decoder allows to map from token-space to image or text space. TerraMind further supports chained generation which ensures consistency across generated modalities. The chained generation represents a conditional probability distribution where the prior probability distribution is determined by the input modality, and all subsequent modalities are generated conditioned on the input modality and potentially other generated modalities.", + "bbox": [ + 511, + 472, + 908, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Thinking-in-Modalities", + "text_level": 1, + "bbox": [ + 511, + 756, + 728, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thinking in Modalities (TiM) is a recursive fine-tuning and inference technique designed to enhance multimodal learning by leveraging the generative capabilities of the model itself. Given an input $x \\in \\mathcal{X}$ (e.g., an optical satellite image), the model first generates additional synthetic modalities $\\tilde{x} = f_{\\mathrm{gen}}(x)$ on a token-level using a learned generative function $f_{\\mathrm{gen}}$ . These generated tokens are then concatenated with the original input and jointly processed by the downstream", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7387", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "model $f$ (e.g., TerraMind encoder with a segmentation head), yielding the final output $y = f(x, f_{\\mathrm{gen}}(x))$ . This formulation allows the model to reason over both observed and inferred modalities, effectively enriching the input space. TiM can leverage multiple generated modalities which are then generated in a chained approach. For example, for $k$ modalities, the input is augmented with newly generated modalities:", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x} ^ {(k + 1)} = \\tilde {x} ^ {(k)} \\cup f _ {\\text {g e n}} (\\tilde {x} ^ {(k)}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 205, + 483, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the final model output is described by:", + "bbox": [ + 89, + 234, + 372, + 250 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ny = f \\left(\\tilde {x} ^ {(K)}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 258, + 483, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This recursive augmentation mimics a chain-of-thought process, enabling the model to iteratively refine its internal representation, particularly in scenarios with missing modalities.", + "bbox": [ + 89, + 286, + 485, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 348, + 223, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we describe the performance gains resulting from TerraMind and experiment with the unlocked capabilities of any-to-any generation and Thinking-in-Modalities.", + "bbox": [ + 89, + 373, + 483, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Foundational experiments", + "text_level": 1, + "bbox": [ + 89, + 428, + 328, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multimodality vs. unimodality. As a first motivational experiment, we outline the benefit of using multimodal data in Earth observation at the example of water body mapping. Specifically, we leverage the ViT-B encoders from the unimodal tokenizer models for S-1, S-2, and LULC, concatenate their embeddings, and train a segmentation head with four ConvNeXt [43] blocks as a late fusion approach. The results in Table 2 (left) suggest that regardless of which modalities we combine, the combination of two modalities always outperforms each unimodal model. Combining all three modalities achieves the best overall performance.", + "bbox": [ + 89, + 449, + 483, + 614 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4a94b084cb60ebef806fd521fdeb71037ea6eccbcf9986eac6b25c653d4d8bbd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
InputLate fusionToken-level fusion
S-161.0163.94 (2.93pp↑)
S-272.7076.32 (3.62pp↑)
LULC71.7770.96 (0.81pp↓)
S-1 + S-273.8376.74 (2.91pp↑)
S-1 + LULC73.8673.76 (0.10pp↓)
S-2 + LULC75.6577.04 (1.39pp↑)
S-1 + S-2 + LULC76.0076.88 (0.88pp↑)
", + "bbox": [ + 127, + 627, + 446, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Token-level fusion vs. late fusion. In Table 2 (right), we investigate the effects of fusing the inputs on a token level", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "through masked token reconstruction. We observe that token-level fusion outperforms late fusion. The performance gains are particularly high when LULC data is not available. This suggests that early fusion captures an internal representation of the multimodal state—especially pronounced for LULC—that benefits fine-tuning. With those findings in mind, we will explore the effects of using additional multi-modal pixel-level input in a dual-scale pretraining in Section 5.5.", + "bbox": [ + 511, + 90, + 908, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Generation experiments", + "text_level": 1, + "bbox": [ + 511, + 234, + 735, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TerraMind supports any-to-any generation. In the following, we provide examples of the generation performance starting from: (i) an information-rich modality, like optical S-2 L2A data, and (ii) minimal information based on the geolocation. In Figure 3, we observe that TerraMind performs strongly in generating image-like modalities like S-1, LULC, and DEM from optical S-2 L2A data. We provide a quantitative overview on the quality of the generations on unseen validation data in Table 3. Overall, we observe an interesting asymmetry in the generative performance of TerraMind where (a) radar-to-optical generation achieves reasonable quality in terms of SSIM and PSNR – indicating structural and visual fidelity with some perceptual degradation – and (b) optical-to-radar generation yields higher PSNR values but lower SSIM, suggesting visually plausible outputs that lack strong structural alignment. The quality of generated DEM suggests to be structurally very strong, but noisy. The errors for DEM generations suggest that the level of altitude is difficult to infer for the model. We compare these scores with the reconstruction quality of the auto-encoding tokenizers in the supplementary material that can serve as upper bounds. Additionally, we provide experiments on the generation quality using token-level instead of pixel-level inputs. Finally, we demonstrate the quality of generations at kilometer scale in Figures 19 and 20.", + "bbox": [ + 511, + 261, + 908, + 640 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8d44d87fb7279771d4eb4bbf789b50b91b744118f08d738966d3b9c0de1324f7.jpg", + "table_caption": [ + "Table 2. Water body mapping on Sen1Floods11 [9] measured in IoU on water class. Model sizes and architectures are comparable. Left column: Late fusion of tokenizers. The average improvement of full multimodality over the individual unimodal performance is 7.5pp IoU. Right column: Finetuning results of TerraMindv1-B-single as a mid fusion approach based on masked correlation learning. Gains over late fusion in percentage points in parentheses." + ], + "table_footnote": [], + "table_body": "
ModalitiesMAE↓RMSE↓SSIM↑PSNR↑
S-1 GRD → S-2 L2A0.0740.1160.75026.210
S-1 GRD → DEM163.0320.80.87820.694
S-1 GRD → NDVI0.1800.2250.43818.990
S-1 RTC → S-2 L2A0.1130.1940.69524.251
S-1 RTC → DEM298.8799.20.87320.009
S-1 RTC → NDVI0.1720.2110.46519.529
S-2 L2A → S-1 GRD2.9423.8770.53128.678
S-2 L2A → S-1 RTC2.6363.3910.43028.993
S-2 L2A → DEM215.8745.50.94220.616
", + "bbox": [ + 519, + 660, + 901, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Quantitative evaluation of generations on unseen global validation dataset using 10 diffusion steps. MAE and RMSE metrics are in physical units: meter (DEM), reflectance (S-2), and db (S-1).", + "bbox": [ + 511, + 820, + 908, + 863 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7388", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ecd65734e16cd2da1375f189a89942b655bab4f6d4ea51b928ebdd4441055148.jpg", + "image_caption": [ + "(a) Input: S-2 L2A data capturing Singapore in January 2025." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 279, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93cf79200582ee94491ba9ee7aa21762717149e729068fe33c90012c0c793f3d.jpg", + "image_caption": [ + "(b) Generation: S-1 RTC composition generated by TerraMind." + ], + "image_footnote": [], + "bbox": [ + 295, + 90, + 477, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9582e25b72bba7331bcf2355f6ee19a7a7832cdd841268bd8932a766d10c44f.jpg", + "image_caption": [ + "(c) Input: S-2 L2A data capturing Northern Spain in January 2025." + ], + "image_footnote": [], + "bbox": [ + 94, + 208, + 279, + 295 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2339bc758ec8ffc0eb95784cdb7350fbdf51c620f199fe7eeeba6e7f859f7db4.jpg", + "image_caption": [ + "(d) Generation: S-1 GRD composition generated by TerraMind.", + "Figure 6. Generated S-1 imagery using TerraMind. We provide large-scale visualizations in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 295, + 209, + 477, + 294 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Zero-shot experiments", + "text_level": 1, + "bbox": [ + 89, + 388, + 300, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on its generative capabilities, TerraMind unlocks several zero-shot applications, like land-use segmentation, water body mapping, geo-localization, and vegetation mapping. In the following, we focus on water body mapping and geo-localization as image- and sequence-level zero-shot tasks.", + "bbox": [ + 89, + 411, + 483, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Water body mapping. In Table 4, we compare the zero-shot performance of TerraMind with its fine-tuned performance and other finetuned benchmarks for water body mapping. Overall, TerraMindv1-B achieves a zero-shot IoU of $45.4\\%$ compared to SOTA-level fine-tuning performance of $82.2\\%$ of DeCUR. In ablations with TerraMindv1-B-single trained on DynamicWorld LULC data, we boost this to up to $69.8\\%$ suggesting that TerraMind harnesses up to over $80\\%$ of the SOTA performance in zero-shot setting. Additionally, it's notable that none of the benchmarking model can be applied in a zero-shot context, highlighting the relevance of TerraMind's capabilities.", + "bbox": [ + 89, + 487, + 483, + 667 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5e41b8de076f12a79156ced1ee4defed1e0191d11046bc88d737cdc73d98364e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelInputTypeIoUWater
TerraMindv1-BS-2zero-shot45.40
TerraMindv1-B-singleS-2zero-shot69.75
Prithvi 2.0 / DeCUR / ...zero-shotN/A
Baseline [9]S-2finetune31.25
Prithvi 2.0 300MS-2finetune80.97
DeCURS-2finetune82.17
", + "bbox": [ + 130, + 679, + 444, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Geo-localization. TerraMind is able to predict the geolocation of a specific data instance. To better visualize the geolocation capabilities, we prompt the model for the most", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "likely locations of the land use class \"bare land\" (deserts etc.) in a Monte-Carlo-sampling in Figure 7. The probability distribution of the model fits the expectation of where to find bare land, highlighting the Sahara region and middle-east, as well as Mexico and Southern California.", + "bbox": [ + 511, + 90, + 905, + 165 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/442bc09a3bff1182ff8bb275d44d364622dcffd33c27729448ba71c7b3172b1d.jpg", + "image_caption": [ + "Figure 7. Prediction distribution of the land use class \"bare land\" with a sampling temperature of $T = 1.0$ using TerraMindv1-B-single. TerraMind has an accurate internal representation of the geolocation of specific contexts, like land use classes." + ], + "image_footnote": [], + "bbox": [ + 563, + 179, + 854, + 290 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Few-shot experiments", + "text_level": 1, + "bbox": [ + 511, + 388, + 718, + 404 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TerraMind is trained via a cross-modal patch classification objective. Thus, we expect a well-structured latent space that clusters different concepts accurately. To investigate our hypothesis, we apply 1-Nearest-Neighbor (1-NN) classification experiments in the community-standard setting of 1-shot 5-way on two datasets: EuroSAT and METER-ML. In those experiments, there are no weight updates of any kind, so that we can assess the quality of the embedding space structure. In Table 5, we observe that TerraMind outperforms several other benchmarks from both the CV and EO domain on the EuroSAT dataset by at least 10pp in accuracy. Our results further show that for methane source classification on METER-ML, TerraMind outperforms benchmark models and generalizes to high-resolution NAIP data with one order of magnitude higher resolution than the pre-training data. We present additional experiments with other few-shot settings in the supplementary material.", + "bbox": [ + 511, + 411, + 906, + 669 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d53fc5b399a97b4b6fdef0fd3e89d1a0c2c282ff6812aac36344421d2cf4f603.jpg", + "table_caption": [ + "Table 4. Zero-shot results of TerraMind on water body mapping compared to fine-tuned performance of benchmarks." + ], + "table_footnote": [], + "table_body": "
ModelInputEuroSATMETER-ML
CLIP-ViT-B/16S-2 RGB57.0029.15
CLIP-ViT-B/16NAIP-32.01
DeCURS-2 L1C50.5427.87
Prithvi 1.0 100MS-2 L1C60.1126.08
Prithvi 2.0 300MS-2 L1C61.0628.26
TerraMindv1-BS-2 L1C70.8333.90
TerraMindv1-BNAIP-32.23
", + "bbox": [ + 524, + 681, + 893, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. 1-shot 5-way classification results on EuroSAT and METER-ML measured in mean accuracy $\\uparrow$ , averaged over 200 runs. TerraMind outperforms benchmarks from CV and EO domain, suggesting a well-structured latent space.", + "bbox": [ + 511, + 825, + 906, + 881 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7389", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4412d721907d41d67686f46428339dc69d078241be78a050b87e7eaf63439ffb.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Performance evaluation of TerraMind using the PANGAEA evaluation protocol indicates higher mIoU values (↑) and lower rank values (↓). The best model per column is highlighted in bold, the second best is underscored. We indicate unimodal datasets with *. Encoders are frozen for pretrained models, while U-Net and ViT baselines are trained from scratch for each specific task." + ], + "table_body": "
ModelBurnSr*MADOS*PASTISSen1Fl11FBP*DEN*CTM-SSSN7*AI4Farms*Avg. mIoUAvg. Rank
CROMA82.4267.5532.3290.8951.8338.2949.3859.2825.6555.296.61
DOFA80.6359.5830.0289.3743.1839.2951.3361.8427.0753.598.22
GFM-Swin76.9064.7121.2472.6067.1834.0946.9860.8927.1952.4210.00
Prithvi 1.0 100M83.6249.9833.9390.3746.8127.8643.0756.5426.8651.0011.00
RemoteCLIP76.5960.0018.2374.2669.1931.7852.0557.7625.1251.6611.22
SatlasNet79.9655.8617.5190.3050.9736.3146.9761.8825.1351.6510.67
Scale-MAE76.6857.3224.5574.1367.1935.1125.4262.9621.4749.4311.44
SpectralGPT80.4757.9935.4489.0733.4237.8546.9558.8626.7551.8710.11
S.-S12-MoCo81.5851.7634.4989.2653.0235.4448.5857.6425.3853.0210.06
S.-S12-DINO81.7249.3736.1888.6151.1534.8148.6656.4725.6252.5110.89
S.-S12-MAE81.9149.9032.0387.7951.9234.0845.8057.1324.6951.6912.39
S.-S12-Data2Vec81.9144.3634.3288.1548.8235.9054.0358.2324.2352.2210.72
UNet Baseline84.5154.7931.6091.4260.4739.4647.5762.0946.3457.584.89
ViT Baseline81.5848.1938.5387.6659.3236.8344.0852.5738.3754.1310.28
TerraMindv1-B82.4269.5240.5190.6259.7237.8755.8060.6128.1258.353.94
TerraMindv1-L82.9375.5743.1390.7863.3837.8955.0459.9827.4759.573.44
", + "bbox": [ + 91, + 88, + 903, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Fine-tuning experiments", + "text_level": 1, + "bbox": [ + 89, + 404, + 316, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides the novel capabilities that TerraMind introduces, we benchmark the fine-tuning performance of TerraMind in both unimodal and multimodal settings following the community-standard PANGAEA benchmark [49]. We summarize the results in Table 6. Overall, TerraMindv1-B outperforms all other GeoFMs by at least 3pp avg. mIoU. Importantly, we observe that TerraMind is the only foundation model approach in EO that across the PANGAEA benchmark outperforms task-specific U-Net models. Performance increases by approximately 2pp avg. mIoU for TerraMindv1-L, with a peak of 5pp in multimodal datasets. Furthermore, TerraMindv1-L outperforms also specialised ViT baselines by 5pp avg. mIoU. Note that per suggestion of the PANGAEA authors, we exclude the xView2 and BioMassters task as we could not reproduce the reported performances. Finally, we assess the impact of leveraging multimodal data as input to TerraMindv1-B compared to utilizing either optical or radar data as unimodal input to better understand the effect of leveraging multimodal data in finetuning. We observe that across all three multimodal tasks, TerraMindv1-B performs best with access to both optical and radar data.", + "bbox": [ + 88, + 428, + 485, + 747 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/76d0dcb5a3b001a3aee1560a4be356fce95718ddbaf9e018da73acd5d653d386.jpg", + "table_caption": [], + "table_footnote": [ + "Table 7. Benefit of using multimodal input in the PANGAEA benchmark reported in mIoU $(\\%)\\uparrow$" + ], + "table_body": "
PASTISSen1F11CTM-SS
S-120.0480.3924.45
S-240.2089.5750.90
S-1 + S-240.5190.6255.80
", + "bbox": [ + 138, + 762, + 436, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Thinking in modalities", + "text_level": 1, + "bbox": [ + 511, + 404, + 723, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We additionally evaluate the value of TiM tuning on water body mapping. We use S-1 or S-2 to generate artificial LULC data as additional input. Our results in Table 8 indicate a superior performance of TiM tuning compared to leveraging uni-modal data by up to 2pp mIoU. This finding points us in the direction of TerraMind being able to generate data that improve downstream task performance. We provide additional results in the appendix.", + "bbox": [ + 511, + 426, + 906, + 549 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e57d722a55f43bd2375e5d31b0f63658d104c5a8813a596746414f0e63eeafbb.jpg", + "table_caption": [], + "table_footnote": [ + "Table 8. Thinking-in-modalities (TiM) tuning compared with standard full fine-tuning approaches on the Sen1Floods11 dataset." + ], + "table_body": "
Fine-TuningInputIoUWatermIoU
TerraMindv1-BS-168.0081.06
TerraMindv1-BS-282.2689.70
TerraMindv1-B TiMS-1 + gen. LULC72.2583.65
TerraMindv1-B TiMS-2 + gen. LULC84.7591.14
", + "bbox": [ + 514, + 561, + 906, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 723, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "TerraMind's approach of combining token-level and pixel-level data has unlocked a range of new model capabilities in EO. TerraMind demonstrates not only beyond state-of-the-art performance in community-standard benchmarks, it also represents the first fully generative multimodal model in the domain. Because of the ability of integrating heterogeneous data sources, we expect that TerraMind-like models will expand to multi-temporal, multi-resolution, and hyperspectral data to fully leverage the data rich ecosystem available in the Earth Observation domain.", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7390", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] A. Hore and D. Ziou. Image quality metrics: PSNR vs. SSIM. In Proc. 20th International Conference on Pattern Recognition (ICPR), pp. 2366-2369, 2010. 16", + "[2] European Space Agency. Copernicus dem. http://dx.doi.org/10.5270/ESA-c5d3d65, 2022.4", + "[3] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Anysat: An earth observation model for any resolutions, scales, and modalities. arXiv preprint arXiv:2412.14123, 2024. 3", + "[4] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Omnisat: Self-supervised modality fusion for earth observation, 2024. 2, 3", + "[5] Nicolas Audebert, Bertrand Le Saux, and Sébastien Lefèvre. Joint Learning from Earth Observation and OpenStreetMap Data to Get Faster Better Semantic Maps. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1552-1560, 2017. 3", + "[6] Benedikt Blumenstiel, Nassim Ait Ali Braham, Conrad M Albrecht, Stefano Maurogiovanni, and Paolo Fraccaro. SSL4EOS12 v1.1 - A Multimodal, Multiseasonal Dataset for Pretraining. arXiv preprint arXiv:2503.00168, 2025. 3, 13", + "[7] Benedikt Blumenstiel, Paolo Fraccaro, Valerio Marsocci, Johannes Jakubik, Stefano Maurogiovanni, Mikolaj Czerkawski, Rocco Sedona, Gabriele Cavallaro, Thomas Brunschwiler, Juan Bernabe-Moreno, and Nicolas Longépé. Terramesh: A planetary mosaic of multimodal earth observation data. arXiv preprint arXiv:2504.11172, 2025. 2, 3", + "[8] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2", + "[9] Derrick Bonafilia, Beth Tellman, Tyler Anderson, and Erica Issenberg. Sen1floods11: A georeferenced dataset to train and test deep learning flood algorithms for sentinel-1. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020. 6, 7", + "[10] Florian Bordes, Richard Yuanzhe Pang, Anurag Ajay, Alexander C Li, Adrien Bardes, Suzanne Petryk, Oscar Manas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 2", + "[11] Jize Cao, Zhe Gan, Yu Cheng, Licheng Yu, Yen-Chun Chen, and Jingjing Liu. Behind the scene: Revealing the secrets of pre-trained vision-and-language models. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 565-580. Springer, 2020. 3", + "[12] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. Maplm: A real-world large-scale vision-language benchmark for map and traffic scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21819-21830, 2024. 3" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yuxing Chen and Lorenzo Bruzzone. Self-supervised change detection in multi-view remote sensing images. arXiv preprint arXiv:2103.05969, 2021. 3", + "[14] Chenwei Wang, et al. SAR Target Image Generation Method Using Azimuth-Controllable Generative Adversarial Network. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS), Vol. 15, 2022. Online: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9933645&tag=1.16", + "[15] Fabian Deuser, Konrad Habel, and Norbert Oswald. Sample4geo: Hard negative sampling for cross-view geolocation. arXiv preprint arXiv:2303.11851, 2023. 3", + "[16] Ivica Dimitrovski, Ivan Kitanovski, Dragi Kocev, and Nikola Simidjievski. Current trends in deep learning for earth observation: An open-source benchmark arena for image classification. ISPRS Journal of Photogrammetry and Remote Sensing, 197:18-35, 2023. 2", + "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 2, 4", + "[18] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 2023. 3", + "[19] Victor Durnov. xview2 1st place solution. 2", + "[20] Adam Van Etten, Dave Lindenbaum, and Todd M. Bacastow. Spacenet: A remote sensing dataset and challenge series, 2019. 2", + "[21] Casper Fibaek, Luke Camilleri, Andreas Luyts, Nikolaos Dionelis, and Bertrand Le Saux. PhilEO Bench: Evaluating Geo-Spatial Foundation Models, In Proc. Int Geoscience and Remote Sensing Symposium (IGARSS), 2024. 2", + "[22] Alistair Francis. Sensor independent cloud and shadow masking with partial labels and multimodal inputs. IEEE Transactions on Geoscience and Remote Sensing, 2024. 4, 13", + "[23] Alistair Francis and Mikolaj Czerkawski. Major tom: Expandable datasets for earth observation. arXiv preprint arXiv:2402.12095, 2024. 3, 13", + "[24] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 3", + "[25] Anthony Fuller, Korean Millard, and James R. Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders, 2023. 3", + "[26] Anatol Garioud, Nicolas Gonthier, Loic Landrieu, Apolline De Wit, Marion Valette, Marc Poupee, Sebastien Giordano, and Boris Wattrelos. FLAIR: a country-scale land cover semantic segmentation dataset from multi-source optical imagery. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7391", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Carlos Gomes, Isabelle Wittmann, Damien Robert, Johannes Jakubik, Tim Reichelt, Michele Martone, Stefano Maurogiovanni, Rikard Vinge, Jonas Hurst, Erik Scheurer, et al. Lossy neural compression for geospatial analytics: A review. arXiv preprint arXiv:2503.01505, 2025. 4", + "[28] Sebastian Hafner, Yifang Ban, and Andrea Nascetti. Unsupervised domain adaptation for global urban extraction using sentinel-1 sar and sentinel-2 msi data. Remote Sensing of Environment, 280:113192, 2022. 3", + "[29] Boran Han, Shuai Zhang, Xingjian Shi, and Markus Reichstein. Bridging remote sensors with multisensor geospatial foundation models, 2024. 2", + "[30] Soyeon Caren Han, Feiqi Cao, Josiah Poon, and Roberto Navigli. Multimodal large language models and tunings: Vision, language, sensors, audio, and beyond. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11294-11295, 2024. 3", + "[31] Jitesh Jain, Jianwei Yang, and Humphrey Shi. Vcoder: Versatile vision encoders for multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27992-28002, 2024. 3", + "[32] Johannes Jakubik, Sujit Roy, C. E. Phillips, Paolo Fraccaro, Denys Godwin, Bianca Zadrozny, Daniela Szwarcman, Carlos Gomes, Gabby Nyirjesy, Blair Edwards, Daiki Kimura, Naomi Simumba, Linsong Chu, S. Karthik Mikkavilli, Devyani Lambhate, Kamal Das, Ranjini Bangalore, Dario Oliveira, Michal Muszynski, Kumar Ankur, Muthukumaran Ramasubramanian, Iksha Gurung, Sam Khallaghi, Hanxi, Li, Michael Cecil, Maryam Ahmadi, Fatemeh Kordi, Hamed Alemohammad, Manil Maskey, Raghu Ganti, Kommy Weldemariam, and Rahul Ramachandran. Foundation models for generalist geospatial artificial intelligence, 2023. 2", + "[33] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 4", + "[34] Samar Khanna, Patrick Liu, Linqi Zhou, Chenlin Meng, Robin Rombach, Marshall Burke, David Lobell, and Stefano Ermon. Diffusionsat: A generative foundation model for satellite imagery, 2023. 3", + "[35] Kohei Arai, Michihiro Mikamo, and Shunsuke Onishi. Method for Image Quality Evaluation of Satellite-based SAR Data. International Journal of Advanced Computer Science and Applications (IJACSA), Vol. 14, No. 7, 2023. Online: http://thesai.org/Downloads/Volume14No7/Paper_13-Method_for/Image_Quality_Evaluation_of_Satellite_based_SAR_Data.pdf.16", + "[36] Saad Lahrichi, Zion Sheng, Shufan Xia, Kyle Bradbury, and Jordan Malof. Is self-supervised pre-training on satellite imagery better than imagenet? a systematic study with sentinel-2. arXiv preprint arXiv:2502.10669, 2025. 2", + "[37] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llavanext: Stronger llms supercharge multimodal capabilities in the wild, 2024. 4, 13", + "[38] Jiaxin Li, Danfeng Hong, Lianru Gao, Jing Yao, Ke Zheng, Bing Zhang, and Jocelyn Chanussot. Deep learning in mul" + ], + "bbox": [ + 91, + 92, + 483, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "timodal remote sensing data fusion: A comprehensive review. International Journal of Applied Earth Observation and Geoinformation, 112:102926, 2022. 3", + "[39] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 2", + "[40] Xiang Li, Congcong Wen, Yuan Hu, Zhenghang Yuan, and Xiao Xiang Zhu. Vision-language models in remote sensing: Current progress and future trends, 2024. 3", + "[41] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19325-19337, 2023. 3", + "[42] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing, 2024. 2, 3", + "[43] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s, 2022. 6", + "[44] Gabriel Machado, Edemir Ferreira, Keiller Nogueira, Hugo Oliveira, Matheus Brito, Pedro Henrique Targino Gama, and Jefersson Alex dos Santos. Airround and cv-brct: Novel multiview datasets for scene classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:488-503, 2020. 3", + "[45] Gengchen Mai, Chris Cundy, Kristy Choi, Yingjie Hu, Ni Lao, and Stefano Ermon. Towards a foundation model for geospatial artificial intelligence (vision paper). In Proceedings of the 30th International Conference on Advances in Geographic Information Systems, New York, NY, USA, 2022. Association for Computing Machinery. 2", + "[46] Oscar Manas, Alexandre Lacoste, Xavier Giró-i Nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2", + "[47] Clive Tinashe Marimo, Benedikt Blumenstiel, Maximilian Nitsche, Johannes Jakubik, and Thomas Brunschwiler. Beyond the visible: Multispectral vision-language learning for earth observation. arXiv preprint arXiv:2503.15969, 2025. 2, 4, 13", + "[48] Valerio Marsocci and Nicolas Audebert. Cross-sensor self-supervised training and alignment for remote sensing, 2024. 3", + "[49] Valerio Marsocci, Yuru Jia, Georges Le Bellier, David Kerekes, Liang Zeng, Sebastian Hafner, Sebastian Gerard, Eric Brune, Ritu Yadav, Ali Shibli, et al. Pangaea: A global and inclusive benchmark for geospatial foundation models. arXiv preprint arXiv:2412.04204, 2024. 2, 8, 18", + "[50] Matias Mendieta, Boran Han, Xingjian Shi, Yi Zhu, Chen Chen, and Mu Li. Gfm: Building geospatial foundation models via continual pretraining. arXiv preprint arXiv:2302.04476, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7392", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023. 4, 15", + "[52] David Mizrahi, Roman Bachmann, Oğuzhan Fatih Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, and Amir Zamir. 4m: Massively multimodal masked modeling, 2023. 4, 5", + "[53] Andrea Nascetti, RITU YADAV, Kirill Brodt, Qixun Qu, Hongwei Fan, Yuri Shendryk, Isha Shah, and Christine Chung. Biomasssters: A benchmark dataset for forest biomass estimation using multi-modal satellite time-series. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2", + "[54] Vishal Nedungadi, Ankit Kariryaa, Stefan Oehmcke, Serge Belongie, Christian Igel, and Nico Lang. Mmearth: Exploring multi-modal pretext tasks for geospatial representation learning. arXiv preprint arXiv:2405.02771, 2024. 2, 3", + "[55] Fernando Paolo, Tsu ting Tim Lin, Ritwik Gupta, Bryce Goodman, Nirav Patel, Daniel Kuster, David Kroodsma, and Jared Dunnmon. xview3-sar: Detecting dark fishing activity using synthetic aperture radar imagery, 2022. 2", + "[56] Prabhishek Singh and Raj Shree. Analysis and effects of speckle noise in SAR images. In Proc. International Conference on Advances in Computing, Communication, & Automation (ICACCA), 2016. DOI: 10.1109/ICAC-CAF.2016.7748978. Online: http://ieeexplore.ieee.org/document/7748978.16", + "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PmLR, 2021. 3, 17", + "[58] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2", + "[59] Ayesha Shafique, Guo Cao, Zia Khan, Muhammad Asad, and Muhammad Aslam. Deep learning-based change detection in remote sensing images: A review. Remote Sensing, 14(4): 871, 2022. 2", + "[60] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 17", + "[61] Aidan M Swope, Xander H Rudelis, and Kyle T Story. Representation learning for remote sensing: An unsupervised sensor fusion approach. arXiv preprint arXiv:2108.05094, 2021. 3", + "[62] Devis Tuia, Konrad Schindler, Begüm Demir, Gustau Camps-Valls, Xiao Xiang Zhu, Mrinalini Kochupillai, Sašo Džeroski, Jan N. van Rijn, Holger H. Hoos, Fabio Del Frate, Mihai Datcu, Jorge-Arnulfo Quiane-Ruiz, Volker Markl, Bertrand Le Saux, and Rochelle Schneider. Artificial intelligence to advance earth observation: a perspective, 2023. 2", + "[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Yi Wang, Conrad M Albrecht, Nassim Ait Ali Braham, Lichao Mou, and Xiao Xiang Zhu. Self-supervised learning in remote sensing: A review. arXiv preprint arXiv:2206.13188, 2022. 2", + "[65] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eos12: A large-scale multimodal, multitemporal dataset for self-supervised learning in earth observation [software and data sets]. IEEE Geoscience and Remote Sensing Magazine, 11 (3):98-106, 2023. 3", + "[66] Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu, Zhe Chen, Wenhai Wang, Xizhou Zhu, Lewei Lu, Tong Lu, et al. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. Advances in Neural Information Processing Systems, 37:69925-69975, 2025. 3", + "[67] Xinyu Bai and Feng Xu. Accelerating Diffusion for SAR-to-Optical Image Translation via Adversarial Consistency Distillation, 2024. Online: http://arxiv.org/pdf/2407.06095.16", + "[68] Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J. Stewart, Joëlle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. Neural plasticity-inspired foundation model for observing the earth crossing modalities, 2024. 3", + "[69] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 2", + "[70] Qidong Yang, Jonathan Giezendanner, Daniel Salles Civitarese, Johannes Jakubik, Eric Schmitt, Anirban Chandra, Jeremy Vila, Detlef Hohl, Chris Hill, Campbell Watson, et al. Multi-modal graph neural networks for localized off-grid weather forecasting. arXiv preprint arXiv:2410.12938, 2024. 2", + "[71] Zhiping Yu, Chenyang Liu, Liqin Liu, Zhenwei Shi, and Zhengxia Zou. Metaearth: A generative foundation model for global-scale remote sensing image generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3", + "[72] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 2", + "[73] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600-612, 2004. 16", + "[74] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[75] Linying Zhao and Shunping Ji. Cnn, rn, or vit? an evaluation of different deep learning architectures for spatio-temporal representation of sentinel time series. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16:44-56, 2022. 2", + "[76] Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7393", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "learning in remote sensing: A comprehensive review and list of resources. IEEE geoscience and remote sensing magazine, 5(4):8-36, 2017. 2", + "bbox": [ + 122, + 92, + 485, + 133 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "7394", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/c9c92fd3-e6cf-49ea-bfb4-aa3c5b130a2c_model.json b/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/c9c92fd3-e6cf-49ea-bfb4-aa3c5b130a2c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ac9114bbfeea78356d5bf4276f1038e4e547b1b6 --- /dev/null +++ b/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/c9c92fd3-e6cf-49ea-bfb4-aa3c5b130a2c_model.json @@ -0,0 +1,2336 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.131, + 0.872, + 0.154 + ], + "angle": 0, + "content": "TerraMind: Large-Scale Generative Multimodality for Earth Observation" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.179, + 0.909, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.257, + 0.81, + 0.295 + ], + "angle": 0, + "content": "\\(^{1}\\)IBM Research - Europe \\(^{2}\\)ETH Zurich \\(^{3}\\)Forschungszentrum Jülich \\(^{4}\\)European Space Agency \\(\\Phi\\)-Lab \\(^{5}\\)NASA IMPACT \\(^{6}\\)University of Iceland" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.297, + 0.618, + 0.311 + ], + "angle": 0, + "content": "johnannes.jakubikl@ibm.com" + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.353, + 0.653, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.353, + 0.906, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.467, + 0.906, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.661, + 0.908, + 0.707 + ], + "angle": 0, + "content": "Figure 1. TerraMind represents the first any-to-any generative, and large-scale multimodal model for Earth observation pre-trained on 500 billion tokens from global geospatial data. The model digests multi-scale representations at pixel-level and token-level simultaneously. TerraMindv1 unlocks (i) generation, (ii) zero-shot and finetuning applications, and (iii) \"Thinking-in-Modalities\" finetuning and inference." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.716, + 0.327, + 0.733 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.487, + 0.857 + ], + "angle": 0, + "content": "We present TerraMind, the first any-to-any generative, multimodal deep learning model for Earth observation (EO). Unlike other approaches, TerraMind is pretrained on dual-scale representations combining both token-level and pixel-level data across modalities. On a token level, TerraMind encodes high-level contextual information to learn cross-modal relationships, while on a pixel level, TerraMind lever" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.909, + 0.884 + ], + "angle": 0, + "content": "ages fine-grained representations to capture critical spatial nuances. In this paper, we demonstrate that (i) TerraMind achieves beyond state-of-the-art performance in community-standard benchmarks, (ii) TerraMind can leverage \"thinking in modalities\" (TiM)—the capability of generating additional artificial data during finetuning and inference to improve the model output—and (iii) TerraMind's dual-scale early fusion approach results in well-structured embedding spaces. Models and code have been open-sourced at https://huggingface.co.ibm-esa-geospatialandhttps://github.com.ibm/terramind." + }, + { + "type": "page_footnote", + "bbox": [ + 0.092, + 0.876, + 0.205, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.889, + 0.2, + 0.9 + ], + "angle": 0, + "content": "† Equal supervision" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.876, + 0.205, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7383" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.121, + 0.486, + 0.288 + ], + "angle": 0, + "content": "Earth observation (EO) increasingly benefits from multimodality because of the important integration of complementary information from different data sources. This becomes particularly relevant as EO is spatiotemporally sparse due to low revisiting times or weather phenomena like cloud coverage. Vice versa, for computer vision, EO data is an important playground for the development of new approaches as there is significant publicly available data of very high quality and complexity. The available modalities range from sensors of different satellite missions to relevant complementary information like digital elevation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.292, + 0.485, + 0.685 + ], + "angle": 0, + "content": "In this work, we introduce TerraMind as the first any-to-any generative multimodal model for EO. With TerraMind, we introduce a dual-scale pretraining on pixel-level and token-level and demonstrate benefits over training primarily on tokens. TerraMind encodes high-level contextual information in tokens to enable correlation learning and scaling, while, additionally capturing important fine-grained representations using pixel-level inputs. During pretraining, TerraMind predicts masked target tokens so that our pretraining objective boils down to a cross-modal patch classification problem that results in high-quality latent spaces. TerraMind is pretrained on a custom global-scale geospatial dataset named TerraMesh with nine million samples that have been aligned spatiotemporally and across modalities [7]. In addition to radar and optical satellite images of the Copernicus Sentinel-1 (S-1) and Sentinel-2 (S-2) missions, our dataset contains task-specific modalities such as land use/land cover (LULC) and normalized difference vegetation index (NDVI) maps, metadata like digital elevation models (DEM) and geographic coordinates, and natural language in the form of captions. To the best of our knowledge, TerraMind represents the first truly generative, multimodal deep learning model for EO. Additionally, in contrast to other recent models that utilize masked autoencoders like [54], contrastive learning, or diffusion techniques, TerraMind uniquely demonstrates benefits of leveraging token-based pretraining for EO." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.69, + 0.487, + 0.902 + ], + "angle": 0, + "content": "We provide an overview of TerraMind's performance in a community-standard benchmark [49] in Figure 2 and highlight the any-to-any generative capabilities of TerraMind in Figure 3. Our key contributions are as follows: (i) We introduce a dual-scale approach for generative multimodal pre-training leveraging data on pixel-level and token-level, which outperforms other fusion approaches and enhances embedding space structures. (ii) We introduce thinking in modalities - similar to chain-of-thought approaches in LLMs - for multi-modal models in EO, demonstrating that infusing generated data during finetuning improves the performance. (iii) We demonstrate that TerraMind outperforms other geospatial foundation models both in unimodal and multimodal settings." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.657, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.116, + 0.908, + 0.538 + ], + "angle": 0, + "content": "Computer vision in Earth observation. Computer vision (CV) has significantly advanced EO [76]. Many CV techniques, originally developed for natural image processing, have been adapted to EO [62], often with minimal modifications. A wide range of tasks benefit from these methods, including classification [16], semantic segmentation [72] (e.g., land cover mapping [20, 21]), change detection [59] (e.g., disaster response [19]), object detection [39] (e.g., vessel identification [55]), and regression (e.g., biomass estimation [53]). Deep learning architectures like CNNs [75] and Vision Transformers (ViTs) [17] have demonstrated strong performance, often surpassing traditional remote sensing (RS) methods. However, EO presents unique challenges, including diverse sensor modalities [4] and geospatial heterogeneity [46]. An emerging paradigm in EO is self-supervised learning (SSL) [64] and geospatial foundation models (GFMs) [45], which aim to leverage vast amounts of unlabeled RS data to develop general purpose task models [32]. While off-the-shelf CV models have shown promising results [36], they do not fully exploit the unique characteristics of geospatial data. Many GFMs still rely on generic CV architectures [50], which were not explicitly designed to handle the complexities of EO, such as heterogeneous sensor sources (e.g., optical, radar, DEM) [29], integrated with auxiliary data (e.g., text) [42, 47], and expert knowledge (e.g., prioritizing specific bands or indexes). In this direction, TerraMind better integrates domain-specific properties, developing a customized and expandable multimodal learning strategy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.54, + 0.909, + 0.603 + ], + "angle": 0, + "content": "Multimodality in CV. Multimodal CV is driven by the integration of diverse data streams [69], such as natural images [74], natural language text [10], temporal video data [58], and weather [70], within large foundation models [8]." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.625, + 0.901, + 0.844 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.855, + 0.911, + 0.898 + ], + "angle": 0, + "content": "Figure 2. TerraMind outperforms other geospatial foundation models on PANGAEA benchmark [49] in finetuning. Performance is measured in mIoU and min-max scaled per dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7384" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.91, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.279, + 0.908, + 0.308 + ], + "angle": 0, + "content": "Figure 3. Chained generation example of TerraMindv1-B starting from either optical, radar, or digital elevation data. Left is input, middle is artificially generated data by TerraMind, right represents ground truths and tokenizer reconstructions, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.334, + 0.484, + 0.514 + ], + "angle": 0, + "content": "Starting from the alignment of images and texts [57], these models moved beyond simple feature extraction, towards nuanced contextual understanding. The ability to combine several modalities allows for unprecedented capabilities in complex tasks [30], evidenced by the rapid advancement of multimodal Large Language Models (MLLMs) [30], that excel in tasks such as scene understanding [12], visual question answering [18], and video analysis [24]. Recent advances in architectures [31] and large scale pre-training [11] have enabled the development of models that learn highly effective cross-modal representations [41], which can then be adapted to a wide variety of downstream tasks [66]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.833 + ], + "angle": 0, + "content": "Multimodality in EO. Multimodality in EO originates from data fusion and is typically understood as the integration of SAR and optical data [13, 25, 28, 38] or the combination of optical data with vector data [5]. Some studies have explored alternative combinations of data. In [15], the authors introduce a contrastive framework for comparing RS images and street views. Even different optical sensors can be considered different modalities [48, 61]. Similarly, several multi-view images (i.e. multimodal) datasets [26, 44, 54] are introduced. More recent approaches combined text and images [40], both for discriminative [42] and generative [34] purposes. Lately, different GFMs are trained in a multimodal way [4, 54, 68], still focusing either on a specific set of modalities (e.g., vision [68], [3]) or tasks (e.g., generative [34]). Compared to multi-scale high-quality generation models for optical data, like MetaEarth [71], our approach allows to generate any modality from any other pretraining modality. To the best of our knowledge, no existing model has combined a wide and diverse amount of modalities both for discriminative and generative purposes, as TerraMind does. We provide a comparison in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.182, + 0.86 + ], + "angle": 0, + "content": "3. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For the pretraining of TerraMind and its tokenizers, we create a multimodal dataset called TerraMesh [7], which will" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.33, + 0.907, + 0.6 + ], + "angle": 0, + "content": "
ModelModalitiesAny-to-Any GenerationMulti-Scale Features
RemoteCLIPoptical, textXX
CROMAoptical, radarXX
AnySataerial, optical, radar, NAIPXX
DeCURoptical, radarXX
DOFAoptical, radar, hyperspectral, NAIPXX
MetaEarthoptical (unimodal)X✓
Galileooptical, radar, elevation, weather, location, population, ...X✓
TerraMindoptical, radar, land use, elevation, vegetation index, location, text✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.611, + 0.908, + 0.641 + ], + "angle": 0, + "content": "Table 1. Comparison of TerraMind to other model architectures. TerraMind represents a first-of-its-kind generative, multimodal model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.748 + ], + "angle": 0, + "content": "be open-sourced to the community. TerraMesh builds on existing datasets, which we expand by adding modalities from external data sources or by applying pseudo-labeling. We provide an overview of the aligned image modalities and a detailed dataset description in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Base datasets. TerraMesh is based on SSL4EO-S12 [6, 65] and MajorTOM-Core [23], two unlabeled remote sensing datasets containing co-aligned radar and optical imagery from Sentinel-1 and Sentinel-2 satellites. SSL4EO-S12 has lower geographic coverage but is multi-seasonal. MajorTOM-Core covers most of the Earth's land surface at a single timestamp. For MajorTOM-Core, we apply a subsampling scheme based on LULC classes and ecoregions. TerraMesh includes a total of approximately 9 million globally distributed samples from both Sentinel-1 and Sentinel-2," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7385" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.429, + 0.106 + ], + "angle": 0, + "content": "each measuring \\(264 \\times 264\\) pixels at \\(10\\mathrm{m}\\) resolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.486, + 0.351 + ], + "angle": 0, + "content": "Additional modalities. We obtain co-aligned yearly LULC maps by ESRI with nine land use classes. Additionally, we leverage SEnSeI v2 [22] as a cloud and ice annotation model and update the ESRI LULC classes for better spatiotemporal alignment. NDVI maps are computed using the corresponding spectral bands from Sentinel-2. DEM is extracted from the Copernicus DEM 30m dataset [2], which provides global coverage of the Earth's elevation at a 30m resolution. Captions are generated synthetically by constructing RGB images from Sentinel-2 patches using the corresponding spectral bands and processing them with LLaVANext [37]. A tailored prompt guides the model to describe the content of each image as described in [47]. For geolocations, we round latitude and longitude from the center of each patch to the nearest quarter degree and store the discretized coordinates as strings in a pre-defined format." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.364, + 0.192, + 0.38 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.485, + 0.466 + ], + "angle": 0, + "content": "TerraMind pretraining is two-staged following [52]. We first pretrain unimodal tokenizer models, tokenize the modalities, and then leverage token-level and pixel-level input to pretrain the TerraMind encoder-decoder architecture. We describe those individual stages in the following." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.476, + 0.228, + 0.491 + ], + "angle": 0, + "content": "4.1. Tokenization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.484, + 0.604 + ], + "angle": 0, + "content": "We develop modality-specific tokenizers to encode each modality into a sequence of discrete tokens for pretraining and decode token sequences back to images. Thus, TerraMind is in principle compatible with any modality, as long as it can be tokenized and aligned with other modalities. For reasons of space, we delegate most experiments related to the tokenizer performances to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Image-like modalities. We train autoencoder-based architectures with a quantization step in the bottleneck for image-like modalities such as S-1, S-2, LULC, NDVI, and DEM. Tokenizer encoders process an input image and generate a latent representation for each \\(16 \\times 16\\) patch, which is then discretized with finite-scalar-quantization (FSQ) [51] into one of \\(N\\) codewords. All tokenizers use a vocabulary size of 16K besides the simpler LULC modality for which we use 4K. These codewords are then used by the diffusion decoder to reconstruct the original image. The benefit of leveraging diffusion decoders lies in facilitating cross-modal generation in TerraMind by transforming tokens back into images. By mapping each codeword to a unique integer in \\(\\{0, 1, \\dots, N - 1\\}\\), we obtain discrete tokens for each image patch. We pretrain the tokenizers in a self-supervised setting. FSQ as quantization method enhances training stability [51] compared to vector quantization [63] by eliminating the need for codebook-related loss terms. Notably, FSQ is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.304 + ], + "angle": 0, + "content": "heavily influenced by ideas of neural compression [27]. For example, on 12-band S-2 images, we achieve compression rates of over \\(3000\\mathrm{x}\\) by applying quantization. We summarize the architecture of our tokenizers in Figure 4. The main objective of the overall tokenizer is to encode image patches consistently into discrete tokens based on semantic similarity to enable cross-modal correlation learning. Therefore, the loss of some details is an expected trade-off, since the focus is on grouping similar patches rather than preserving all fine-grained features. Naturally, more accurate reconstructions facilitate cross-modal generation, however the main focus of the pretraining lies on consistent cross-modal correlation learning. We provided further details on the pretraining of the tokenizers in the supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.319, + 0.907, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.433, + 0.909, + 0.462 + ], + "angle": 0, + "content": "Figure 4. Tokenizer for image-like modalities combining finite-scalar quantization [51] with diffusion decoding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.909, + 0.6 + ], + "angle": 0, + "content": "Sequence-like modalities. We treat both captions and geolocations as text and use a single text tokenizer to process both modalities. By discretizing the geographic coordinates and representing them as strings, we introduce special coordinate tokens into the vocabulary. This allows us to encode geolocations as a sequence of discrete tokens, beginning with a latitude token followed by a longitude token. For textual data, we modify the existing WordPiece tokenizer [33]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.646, + 0.623 + ], + "angle": 0, + "content": "4.2. Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.75 + ], + "angle": 0, + "content": "Architecture. TerraMind uses a symmetric Transformer-based encoder-decoder architecture proposed in [52], which is designed to process sequences of multimodal tokens. In addition to discrete tokens, TerraMind accepts pixel-level inputs, specifically satellite imagery and digital elevation maps. For pixel-level inputs, we apply learnable patch-wise linear projections to generate patch embeddings for each \\(16 \\times 16\\) patch, similar to the approach used in ViT [17]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Dual-scale early fusion. In contrast to [52], we not only embed token-level data but additionally leverage pixel-level data across a range of input modalities to introduce a dual-scale feature representation to support the structuring of the embedding space. Both tokens and patches represent a 16x16 pixel area. Tokens represent this area via a single discrete integer value, while the image patches describe the same area with the actual floating point sensor data. Thus, during pretraining, the model not only learns a correlation between modalities (i.e., cross-modal learning) but also between dif" + }, + { + "type": "page_footnote", + "bbox": [ + 0.092, + 0.89, + 0.424, + 0.9 + ], + "angle": 0, + "content": "https://planetarycomputer.microsoft.com/dataset/io-lulc-annual-v02" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7386" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.244 + ], + "angle": 0, + "content": "ferent levels of abstraction within the same modality. The low-level token information enables cross-modal correlation learning, while adding pixel level input accounts for spatial nuances. Based on dual-scale features the model further learns to better structure pixel-level data in the embedding space via the corresponding information from the discrete token. We illustrate the pretraining paradigm in Figure 5. The model is agnostic to processing tokens or patches in the input space, while the target is generally token-level data. We use six pixel-level modalities and eight token-level modalities." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.257, + 0.473, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.425 + ], + "angle": 0, + "content": "Figure 5. Illustration of the pre-training task. Given an encoded multimodal sample of random subsets of patches and input tokens, the decoder predicts target tokens for the masked input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.484, + 0.529 + ], + "angle": 0, + "content": "Masking strategy. TerraMind applies a masked modeling approach in the token space following [52]. The model leverages a set of randomly selected target tokens that have to be reconstructed from a randomly selected set of input tokens and pixel-level data. During pre-training, we sample input and target data from a Dirichlet distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.756 + ], + "angle": 0, + "content": "We opt for masked token reconstruction to familiarize the model with the absence of entire modalities, which is crucial for a high usability of a multimodal model in Earth observation. During pre-training, the model learns an internal representation of unseen modalities which is expected to benefit a range of downstream applications. In addition, sampling input and target tokens improves the computational efficiency of the pre-training, as each token is a compressed representation of a patch with compression factors of between 250x and 3000x depending on the modality. Finally, without tokenized representations of the image-like modalities, it is challenging to learn the correlation to sequence-like modalities. The overall training objective of TerraMind boils down to a cross-modal patch-level classification problem optimized via a cross entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.764, + 0.484, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} = - \\sum_ {i = 1} ^ {N} y _ {i} \\log \\left(p _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.904 + ], + "angle": 0, + "content": "where \\( y_{i} \\) is the one-hot encoded true class of token \\( i \\), \\( p_{i} \\) is the predicted probability for token \\( i \\), \\( N \\) is the total number of possible tokens. Interestingly, we can infer an upper bound loss for a random model where the cross entropy loss will collapse to the natural logarithm of the vocabulary size \\( \\mathcal{L}_{\\mathrm{CE,random}} = -\\sum_{i=1}^{N} y_{i} \\log \\left( \\frac{1}{N} \\right) = \\log(N) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.44 + ], + "angle": 0, + "content": "Scaling. We trained three versions of TerraMind scaling across model size, compute, and data. In addition, we pretrain different versions of TerraMind with respect to the number of dual-scale features. TerraMindv1-B is pre-trained on 500B tokens for 6 days on 32 NVIDIA A100 GPUs. The model uses dual-scale features from both token-level and pixel-level. During initial experiments, we observed significant improvements from scaling model size when switching from a tiny backbone to a small backbone to a base backbone. Therefore, we pre-trained TerraMindv1-L on a large backbone with 500B tokens on 32 NVIDIA A100 GPUs trained for 10 days. Finally, to better understand the effect of scaling across the dual-scale feature representation, we pre-train TerraMindv1-B-single as a single-scale model on primarily token-level data with optical S-2 L2A data as only pixel-level input (compared to pixel-level S-2 L1C, S-2 RGB, S-1 GRD, S-1 RTC, and DEM in TerraMindv1-B and -L). TerraMindv1-B-single is pretrained on 500B tokens from over one million samples for 6 days on 32 NVIDIA A100 GPUs. We summarize the scaling behavior in model size, compute, and data in Figure 9 of the supplementary material. We additionally provide final validation losses in Table 9 comparing v1-B and v1-L with the theoretical random loss." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.451, + 0.638, + 0.466 + ], + "angle": 0, + "content": "4.3. Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.909, + 0.731 + ], + "angle": 0, + "content": "Once pretrained, TerraMind can generate tokens for any modality, conditioned on any subset of input modalities. The generative capabilities unlock various zero-shot tasks, such as water body segmentation. For the generation of image-like modalities, the decoder receives mask tokens for the modality to be generated and predicts the corresponding tokens based on the encoded input. For sequence-like modalities, the decoder generates the output autoregressively. After generating tokens from the target modality, the corresponding tokenizer decoder allows to map from token-space to image or text space. TerraMind further supports chained generation which ensures consistency across generated modalities. The chained generation represents a conditional probability distribution where the prior probability distribution is determined by the input modality, and all subsequent modalities are generated conditioned on the input modality and potentially other generated modalities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.729, + 0.773 + ], + "angle": 0, + "content": "4.4. Thinking-in-Modalities" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Thinking in Modalities (TiM) is a recursive fine-tuning and inference technique designed to enhance multimodal learning by leveraging the generative capabilities of the model itself. Given an input \\( x \\in \\mathcal{X} \\) (e.g., an optical satellite image), the model first generates additional synthetic modalities \\( \\tilde{x} = f_{\\mathrm{gen}}(x) \\) on a token-level using a learned generative function \\( f_{\\mathrm{gen}} \\). These generated tokens are then concatenated with the original input and jointly processed by the downstream" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7387" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "model \\( f \\) (e.g., TerraMind encoder with a segmentation head), yielding the final output \\( y = f(x, f_{\\mathrm{gen}}(x)) \\). This formulation allows the model to reason over both observed and inferred modalities, effectively enriching the input space. TiM can leverage multiple generated modalities which are then generated in a chained approach. For example, for \\( k \\) modalities, the input is augmented with newly generated modalities:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.206, + 0.484, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\tilde {x} ^ {(k + 1)} = \\tilde {x} ^ {(k)} \\cup f _ {\\text {g e n}} (\\tilde {x} ^ {(k)}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.235, + 0.374, + 0.25 + ], + "angle": 0, + "content": "and the final model output is described by:" + }, + { + "type": "equation", + "bbox": [ + 0.239, + 0.259, + 0.484, + 0.278 + ], + "angle": 0, + "content": "\\[\ny = f \\left(\\tilde {x} ^ {(K)}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.287, + 0.486, + 0.334 + ], + "angle": 0, + "content": "This recursive augmentation mimics a chain-of-thought process, enabling the model to iteratively refine its internal representation, particularly in scenarios with missing modalities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.349, + 0.224, + 0.366 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.484, + 0.421 + ], + "angle": 0, + "content": "In this section, we describe the performance gains resulting from TerraMind and experiment with the unlocked capabilities of any-to-any generation and Thinking-in-Modalities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.429, + 0.329, + 0.445 + ], + "angle": 0, + "content": "5.1. Foundational experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.485, + 0.616 + ], + "angle": 0, + "content": "Multimodality vs. unimodality. As a first motivational experiment, we outline the benefit of using multimodal data in Earth observation at the example of water body mapping. Specifically, we leverage the ViT-B encoders from the unimodal tokenizer models for S-1, S-2, and LULC, concatenate their embeddings, and train a segmentation head with four ConvNeXt [43] blocks as a late fusion approach. The results in Table 2 (left) suggest that regardless of which modalities we combine, the combination of two modalities always outperforms each unimodal model. Combining all three modalities achieves the best overall performance." + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.628, + 0.447, + 0.747 + ], + "angle": 0, + "content": "
InputLate fusionToken-level fusion
S-161.0163.94 (2.93pp↑)
S-272.7076.32 (3.62pp↑)
LULC71.7770.96 (0.81pp↓)
S-1 + S-273.8376.74 (2.91pp↑)
S-1 + LULC73.8673.76 (0.10pp↓)
S-2 + LULC75.6577.04 (1.39pp↑)
S-1 + S-2 + LULC76.0076.88 (0.88pp↑)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.758, + 0.485, + 0.855 + ], + "angle": 0, + "content": "Table 2. Water body mapping on Sen1Floods11 [9] measured in IoU on water class. Model sizes and architectures are comparable. Left column: Late fusion of tokenizers. The average improvement of full multimodality over the individual unimodal performance is 7.5pp IoU. Right column: Finetuning results of TerraMindv1-B-single as a mid fusion approach based on masked correlation learning. Gains over late fusion in percentage points in parentheses." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Token-level fusion vs. late fusion. In Table 2 (right), we investigate the effects of fusing the inputs on a token level" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.213 + ], + "angle": 0, + "content": "through masked token reconstruction. We observe that token-level fusion outperforms late fusion. The performance gains are particularly high when LULC data is not available. This suggests that early fusion captures an internal representation of the multimodal state—especially pronounced for LULC—that benefits fine-tuning. With those findings in mind, we will explore the effects of using additional multi-modal pixel-level input in a dual-scale pretraining in Section 5.5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.236, + 0.736, + 0.252 + ], + "angle": 0, + "content": "5.2. Generation experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.909, + 0.641 + ], + "angle": 0, + "content": "TerraMind supports any-to-any generation. In the following, we provide examples of the generation performance starting from: (i) an information-rich modality, like optical S-2 L2A data, and (ii) minimal information based on the geolocation. In Figure 3, we observe that TerraMind performs strongly in generating image-like modalities like S-1, LULC, and DEM from optical S-2 L2A data. We provide a quantitative overview on the quality of the generations on unseen validation data in Table 3. Overall, we observe an interesting asymmetry in the generative performance of TerraMind where (a) radar-to-optical generation achieves reasonable quality in terms of SSIM and PSNR – indicating structural and visual fidelity with some perceptual degradation – and (b) optical-to-radar generation yields higher PSNR values but lower SSIM, suggesting visually plausible outputs that lack strong structural alignment. The quality of generated DEM suggests to be structurally very strong, but noisy. The errors for DEM generations suggest that the level of altitude is difficult to infer for the model. We compare these scores with the reconstruction quality of the auto-encoding tokenizers in the supplementary material that can serve as upper bounds. Additionally, we provide experiments on the generation quality using token-level instead of pixel-level inputs. Finally, we demonstrate the quality of generations at kilometer scale in Figures 19 and 20." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.661, + 0.903, + 0.811 + ], + "angle": 0, + "content": "
ModalitiesMAE↓RMSE↓SSIM↑PSNR↑
S-1 GRD → S-2 L2A0.0740.1160.75026.210
S-1 GRD → DEM163.0320.80.87820.694
S-1 GRD → NDVI0.1800.2250.43818.990
S-1 RTC → S-2 L2A0.1130.1940.69524.251
S-1 RTC → DEM298.8799.20.87320.009
S-1 RTC → NDVI0.1720.2110.46519.529
S-2 L2A → S-1 GRD2.9423.8770.53128.678
S-2 L2A → S-1 RTC2.6363.3910.43028.993
S-2 L2A → DEM215.8745.50.94220.616
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.821, + 0.909, + 0.864 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation of generations on unseen global validation dataset using 10 diffusion steps. MAE and RMSE metrics are in physical units: meter (DEM), reflectance (S-2), and db (S-1)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7388" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.091, + 0.28, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.17, + 0.282, + 0.194 + ], + "angle": 0, + "content": "(a) Input: S-2 L2A data capturing Singapore in January 2025." + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.092, + 0.478, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.17, + 0.483, + 0.194 + ], + "angle": 0, + "content": "(b) Generation: S-1 RTC composition generated by TerraMind." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.209, + 0.28, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.3, + 0.282, + 0.324 + ], + "angle": 0, + "content": "(c) Input: S-2 L2A data capturing Northern Spain in January 2025." + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.21, + 0.478, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.3, + 0.483, + 0.324 + ], + "angle": 0, + "content": "(d) Generation: S-1 GRD composition generated by TerraMind." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.336, + 0.483, + 0.365 + ], + "angle": 0, + "content": "Figure 6. Generated S-1 imagery using TerraMind. We provide large-scale visualizations in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.39, + 0.301, + 0.405 + ], + "angle": 0, + "content": "5.3. Zero-shot experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.485, + 0.487 + ], + "angle": 0, + "content": "Based on its generative capabilities, TerraMind unlocks several zero-shot applications, like land-use segmentation, water body mapping, geo-localization, and vegetation mapping. In the following, we focus on water body mapping and geo-localization as image- and sequence-level zero-shot tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.484, + 0.669 + ], + "angle": 0, + "content": "Water body mapping. In Table 4, we compare the zero-shot performance of TerraMind with its fine-tuned performance and other finetuned benchmarks for water body mapping. Overall, TerraMindv1-B achieves a zero-shot IoU of \\(45.4\\%\\) compared to SOTA-level fine-tuning performance of \\(82.2\\%\\) of DeCUR. In ablations with TerraMindv1-B-single trained on DynamicWorld LULC data, we boost this to up to \\(69.8\\%\\) suggesting that TerraMind harnesses up to over \\(80\\%\\) of the SOTA performance in zero-shot setting. Additionally, it's notable that none of the benchmarking model can be applied in a zero-shot context, highlighting the relevance of TerraMind's capabilities." + }, + { + "type": "table", + "bbox": [ + 0.131, + 0.68, + 0.445, + 0.801 + ], + "angle": 0, + "content": "
ModelInputTypeIoUWater
TerraMindv1-BS-2zero-shot45.40
TerraMindv1-B-singleS-2zero-shot69.75
Prithvi 2.0 / DeCUR / ...zero-shotN/A
Baseline [9]S-2finetune31.25
Prithvi 2.0 300MS-2finetune80.97
DeCURS-2finetune82.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Table 4. Zero-shot results of TerraMind on water body mapping compared to fine-tuned performance of benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Geo-localization. TerraMind is able to predict the geolocation of a specific data instance. To better visualize the geolocation capabilities, we prompt the model for the most" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.166 + ], + "angle": 0, + "content": "likely locations of the land use class \"bare land\" (deserts etc.) in a Monte-Carlo-sampling in Figure 7. The probability distribution of the model fits the expectation of where to find bare land, highlighting the Sahara region and middle-east, as well as Mexico and Southern California." + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.18, + 0.856, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.305, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 7. Prediction distribution of the land use class \"bare land\" with a sampling temperature of \\( T = 1.0 \\) using TerraMindv1-B-single. TerraMind has an accurate internal representation of the geolocation of specific contexts, like land use classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.39, + 0.719, + 0.405 + ], + "angle": 0, + "content": "5.4. Few-shot experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.908, + 0.67 + ], + "angle": 0, + "content": "TerraMind is trained via a cross-modal patch classification objective. Thus, we expect a well-structured latent space that clusters different concepts accurately. To investigate our hypothesis, we apply 1-Nearest-Neighbor (1-NN) classification experiments in the community-standard setting of 1-shot 5-way on two datasets: EuroSAT and METER-ML. In those experiments, there are no weight updates of any kind, so that we can assess the quality of the embedding space structure. In Table 5, we observe that TerraMind outperforms several other benchmarks from both the CV and EO domain on the EuroSAT dataset by at least 10pp in accuracy. Our results further show that for methane source classification on METER-ML, TerraMind outperforms benchmark models and generalizes to high-resolution NAIP data with one order of magnitude higher resolution than the pre-training data. We present additional experiments with other few-shot settings in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.682, + 0.895, + 0.815 + ], + "angle": 0, + "content": "
ModelInputEuroSATMETER-ML
CLIP-ViT-B/16S-2 RGB57.0029.15
CLIP-ViT-B/16NAIP-32.01
DeCURS-2 L1C50.5427.87
Prithvi 1.0 100MS-2 L1C60.1126.08
Prithvi 2.0 300MS-2 L1C61.0628.26
TerraMindv1-BS-2 L1C70.8333.90
TerraMindv1-BNAIP-32.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.826, + 0.908, + 0.882 + ], + "angle": 0, + "content": "Table 5. 1-shot 5-way classification results on EuroSAT and METER-ML measured in mean accuracy \\(\\uparrow\\), averaged over 200 runs. TerraMind outperforms benchmarks from CV and EO domain, suggesting a well-structured latent space." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7389" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.328 + ], + "angle": 0, + "content": "
ModelBurnSr*MADOS*PASTISSen1Fl11FBP*DEN*CTM-SSSN7*AI4Farms*Avg. mIoUAvg. Rank
CROMA82.4267.5532.3290.8951.8338.2949.3859.2825.6555.296.61
DOFA80.6359.5830.0289.3743.1839.2951.3361.8427.0753.598.22
GFM-Swin76.9064.7121.2472.6067.1834.0946.9860.8927.1952.4210.00
Prithvi 1.0 100M83.6249.9833.9390.3746.8127.8643.0756.5426.8651.0011.00
RemoteCLIP76.5960.0018.2374.2669.1931.7852.0557.7625.1251.6611.22
SatlasNet79.9655.8617.5190.3050.9736.3146.9761.8825.1351.6510.67
Scale-MAE76.6857.3224.5574.1367.1935.1125.4262.9621.4749.4311.44
SpectralGPT80.4757.9935.4489.0733.4237.8546.9558.8626.7551.8710.11
S.-S12-MoCo81.5851.7634.4989.2653.0235.4448.5857.6425.3853.0210.06
S.-S12-DINO81.7249.3736.1888.6151.1534.8148.6656.4725.6252.5110.89
S.-S12-MAE81.9149.9032.0387.7951.9234.0845.8057.1324.6951.6912.39
S.-S12-Data2Vec81.9144.3634.3288.1548.8235.9054.0358.2324.2352.2210.72
UNet Baseline84.5154.7931.6091.4260.4739.4647.5762.0946.3457.584.89
ViT Baseline81.5848.1938.5387.6659.3236.8344.0852.5738.3754.1310.28
TerraMindv1-B82.4269.5240.5190.6259.7237.8755.8060.6128.1258.353.94
TerraMindv1-L82.9375.5743.1390.7863.3837.8955.0459.9827.4759.573.44
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.337, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Table 6. Performance evaluation of TerraMind using the PANGAEA evaluation protocol indicates higher mIoU values (↑) and lower rank values (↓). The best model per column is highlighted in bold, the second best is underscored. We indicate unimodal datasets with *. Encoders are frozen for pretrained models, while U-Net and ViT baselines are trained from scratch for each specific task." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.405, + 0.317, + 0.422 + ], + "angle": 0, + "content": "5.5. Fine-tuning experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.429, + 0.486, + 0.748 + ], + "angle": 0, + "content": "Besides the novel capabilities that TerraMind introduces, we benchmark the fine-tuning performance of TerraMind in both unimodal and multimodal settings following the community-standard PANGAEA benchmark [49]. We summarize the results in Table 6. Overall, TerraMindv1-B outperforms all other GeoFMs by at least 3pp avg. mIoU. Importantly, we observe that TerraMind is the only foundation model approach in EO that across the PANGAEA benchmark outperforms task-specific U-Net models. Performance increases by approximately 2pp avg. mIoU for TerraMindv1-L, with a peak of 5pp in multimodal datasets. Furthermore, TerraMindv1-L outperforms also specialised ViT baselines by 5pp avg. mIoU. Note that per suggestion of the PANGAEA authors, we exclude the xView2 and BioMassters task as we could not reproduce the reported performances. Finally, we assess the impact of leveraging multimodal data as input to TerraMindv1-B compared to utilizing either optical or radar data as unimodal input to better understand the effect of leveraging multimodal data in finetuning. We observe that across all three multimodal tasks, TerraMindv1-B performs best with access to both optical and radar data." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.763, + 0.437, + 0.842 + ], + "angle": 0, + "content": "
PASTISSen1F11CTM-SS
S-120.0480.3924.45
S-240.2089.5750.90
S-1 + S-240.5190.6255.80
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.852, + 0.485, + 0.88 + ], + "angle": 0, + "content": "Table 7. Benefit of using multimodal input in the PANGAEA benchmark reported in mIoU \\((\\%)\\uparrow\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.405, + 0.725, + 0.422 + ], + "angle": 0, + "content": "5.6. Thinking in modalities" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.55 + ], + "angle": 0, + "content": "We additionally evaluate the value of TiM tuning on water body mapping. We use S-1 or S-2 to generate artificial LULC data as additional input. Our results in Table 8 indicate a superior performance of TiM tuning compared to leveraging uni-modal data by up to 2pp mIoU. This finding points us in the direction of TerraMind being able to generate data that improve downstream task performance. We provide additional results in the appendix." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.562, + 0.907, + 0.655 + ], + "angle": 0, + "content": "
Fine-TuningInputIoUWatermIoU
TerraMindv1-BS-168.0081.06
TerraMindv1-BS-282.2689.70
TerraMindv1-B TiMS-1 + gen. LULC72.2583.65
TerraMindv1-B TiMS-2 + gen. LULC84.7591.14
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.665, + 0.909, + 0.694 + ], + "angle": 0, + "content": "Table 8. Thinking-in-modalities (TiM) tuning compared with standard full fine-tuning approaches on the Sen1Floods11 dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.634, + 0.74 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "TerraMind's approach of combining token-level and pixel-level data has unlocked a range of new model capabilities in EO. TerraMind demonstrates not only beyond state-of-the-art performance in community-standard benchmarks, it also represents the first fully generative multimodal model in the domain. Because of the ability of integrating heterogeneous data sources, we expect that TerraMind-like models will expand to multi-temporal, multi-resolution, and hyperspectral data to fully leverage the data rich ecosystem available in the Earth Observation domain." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7390" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.156 + ], + "angle": 0, + "content": "[1] A. Hore and D. Ziou. Image quality metrics: PSNR vs. SSIM. In Proc. 20th International Conference on Pattern Recognition (ICPR), pp. 2366-2369, 2010. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.485, + 0.184 + ], + "angle": 0, + "content": "[2] European Space Agency. Copernicus dem. http://dx.doi.org/10.5270/ESA-c5d3d65, 2022.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.187, + 0.483, + 0.241 + ], + "angle": 0, + "content": "[3] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Anysat: An earth observation model for any resolutions, scales, and modalities. arXiv preprint arXiv:2412.14123, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.483, + 0.284 + ], + "angle": 0, + "content": "[4] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Omnisat: Self-supervised modality fusion for earth observation, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.484, + 0.367 + ], + "angle": 0, + "content": "[5] Nicolas Audebert, Bertrand Le Saux, and Sébastien Lefèvre. Joint Learning from Earth Observation and OpenStreetMap Data to Get Faster Better Semantic Maps. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1552-1560, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.484, + 0.436 + ], + "angle": 0, + "content": "[6] Benedikt Blumenstiel, Nassim Ait Ali Braham, Conrad M Albrecht, Stefano Maurogiovanni, and Paolo Fraccaro. SSL4EOS12 v1.1 - A Multimodal, Multiseasonal Dataset for Pretraining. arXiv preprint arXiv:2503.00168, 2025. 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[7] Benedikt Blumenstiel, Paolo Fraccaro, Valerio Marsocci, Johannes Jakubik, Stefano Maurogiovanni, Mikolaj Czerkawski, Rocco Sedona, Gabriele Cavallaro, Thomas Brunschwiler, Juan Bernabe-Moreno, and Nicolas Longépé. Terramesh: A planetary mosaic of multimodal earth observation data. arXiv preprint arXiv:2504.11172, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[8] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.484, + 0.663 + ], + "angle": 0, + "content": "[9] Derrick Bonafilia, Beth Tellman, Tyler Anderson, and Erica Issenberg. Sen1floods11: A georeferenced dataset to train and test deep learning flood algorithms for sentinel-1. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.664, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[10] Florian Bordes, Richard Yuanzhe Pang, Anurag Ajay, Alexander C Li, Adrien Bardes, Suzanne Petryk, Oscar Manas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[11] Jize Cao, Zhe Gan, Yu Cheng, Licheng Yu, Yen-Chun Chen, and Jingjing Liu. Behind the scene: Revealing the secrets of pre-trained vision-and-language models. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 565-580. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. Maplm: A real-world large-scale vision-language benchmark for map and traffic scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21819-21830, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[13] Yuxing Chen and Lorenzo Bruzzone. Self-supervised change detection in multi-view remote sensing images. arXiv preprint arXiv:2103.05969, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.908, + 0.22 + ], + "angle": 0, + "content": "[14] Chenwei Wang, et al. SAR Target Image Generation Method Using Azimuth-Controllable Generative Adversarial Network. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS), Vol. 15, 2022. Online: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9933645&tag=1.16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.223, + 0.908, + 0.264 + ], + "angle": 0, + "content": "[15] Fabian Deuser, Konrad Habel, and Norbert Oswald. Sample4geo: Hard negative sampling for cross-view geolocation. arXiv preprint arXiv:2303.11851, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.267, + 0.908, + 0.334 + ], + "angle": 0, + "content": "[16] Ivica Dimitrovski, Ivan Kitanovski, Dragi Kocev, and Nikola Simidjievski. Current trends in deep learning for earth observation: An open-source benchmark arena for image classification. ISPRS Journal of Photogrammetry and Remote Sensing, 197:18-35, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.338, + 0.907, + 0.42 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.424, + 0.906, + 0.479 + ], + "angle": 0, + "content": "[18] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.482, + 0.811, + 0.495 + ], + "angle": 0, + "content": "[19] Victor Durnov. xview2 1st place solution. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.499, + 0.908, + 0.538 + ], + "angle": 0, + "content": "[20] Adam Van Etten, Dave Lindenbaum, and Todd M. Bacastow. Spacenet: A remote sensing dataset and challenge series, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.542, + 0.906, + 0.598 + ], + "angle": 0, + "content": "[21] Casper Fibaek, Luke Camilleri, Andreas Luyts, Nikolaos Dionelis, and Bertrand Le Saux. PhilEO Bench: Evaluating Geo-Spatial Foundation Models, In Proc. Int Geoscience and Remote Sensing Symposium (IGARSS), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.907, + 0.641 + ], + "angle": 0, + "content": "[22] Alistair Francis. Sensor independent cloud and shadow masking with partial labels and multimodal inputs. IEEE Transactions on Geoscience and Remote Sensing, 2024. 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.907, + 0.685 + ], + "angle": 0, + "content": "[23] Alistair Francis and Mikolaj Czerkawski. Major tom: Expandable datasets for earth observation. arXiv preprint arXiv:2402.12095, 2024. 3, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.688, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[24] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[25] Anthony Fuller, Korean Millard, and James R. Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[26] Anatol Garioud, Nicolas Gonthier, Loic Landrieu, Apolline De Wit, Marion Valette, Marc Poupee, Sebastien Giordano, and Boris Wattrelos. FLAIR: a country-scale land cover semantic segmentation dataset from multi-source optical imagery. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7391" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.161 + ], + "angle": 0, + "content": "[27] Carlos Gomes, Isabelle Wittmann, Damien Robert, Johannes Jakubik, Tim Reichelt, Michele Martone, Stefano Maurogiovanni, Rikard Vinge, Jonas Hurst, Erik Scheurer, et al. Lossy neural compression for geospatial analytics: A review. arXiv preprint arXiv:2503.01505, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.484, + 0.217 + ], + "angle": 0, + "content": "[28] Sebastian Hafner, Yifang Ban, and Andrea Nascetti. Unsupervised domain adaptation for global urban extraction using sentinel-1 sar and sentinel-2 msi data. Remote Sensing of Environment, 280:113192, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[29] Boran Han, Shuai Zhang, Xingjian Shi, and Markus Reichstein. Bridging remote sensors with multisensor geospatial foundation models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.484, + 0.329 + ], + "angle": 0, + "content": "[30] Soyeon Caren Han, Feiqi Cao, Josiah Poon, and Roberto Navigli. Multimodal large language models and tunings: Vision, language, sensors, audio, and beyond. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11294-11295, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[31] Jitesh Jain, Jianwei Yang, and Humphrey Shi. Vcoder: Versatile vision encoders for multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27992-28002, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.484, + 0.538 + ], + "angle": 0, + "content": "[32] Johannes Jakubik, Sujit Roy, C. E. Phillips, Paolo Fraccaro, Denys Godwin, Bianca Zadrozny, Daniela Szwarcman, Carlos Gomes, Gabby Nyirjesy, Blair Edwards, Daiki Kimura, Naomi Simumba, Linsong Chu, S. Karthik Mikkavilli, Devyani Lambhate, Kamal Das, Ranjini Bangalore, Dario Oliveira, Michal Muszynski, Kumar Ankur, Muthukumaran Ramasubramanian, Iksha Gurung, Sam Khallaghi, Hanxi, Li, Michael Cecil, Maryam Ahmadi, Fatemeh Kordi, Hamed Alemohammad, Manil Maskey, Raghu Ganti, Kommy Weldemariam, and Rahul Ramachandran. Foundation models for generalist geospatial artificial intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[33] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.65 + ], + "angle": 0, + "content": "[34] Samar Khanna, Patrick Liu, Linqi Zhou, Chenlin Meng, Robin Rombach, Marshall Burke, David Lobell, and Stefano Ermon. Diffusionsat: A generative foundation model for satellite imagery, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[35] Kohei Arai, Michihiro Mikamo, and Shunsuke Onishi. Method for Image Quality Evaluation of Satellite-based SAR Data. International Journal of Advanced Computer Science and Applications (IJACSA), Vol. 14, No. 7, 2023. Online: http://thesai.org/Downloads/Volume14No7/Paper_13-Method_for/Image_Quality_Evaluation_of_Satellite_based_SAR_Data.pdf.16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[36] Saad Lahrichi, Zion Sheng, Shufan Xia, Kyle Bradbury, and Jordan Malof. Is self-supervised pre-training on satellite imagery better than imagenet? a systematic study with sentinel-2. arXiv preprint arXiv:2502.10669, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[37] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llavanext: Stronger llms supercharge multimodal capabilities in the wild, 2024. 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[38] Jiaxin Li, Danfeng Hong, Lianru Gao, Jing Yao, Ke Zheng, Bing Zhang, and Jocelyn Chanussot. Deep learning in mul" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.134 + ], + "angle": 0, + "content": "timodal remote sensing data fusion: A comprehensive review. International Journal of Applied Earth Observation and Geoinformation, 112:102926, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.908, + 0.192 + ], + "angle": 0, + "content": "[39] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.907, + 0.236 + ], + "angle": 0, + "content": "[40] Xiang Li, Congcong Wen, Yuan Hu, Zhenghang Yuan, and Xiao Xiang Zhu. Vision-language models in remote sensing: Current progress and future trends, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.239, + 0.908, + 0.308 + ], + "angle": 0, + "content": "[41] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19325-19337, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.908, + 0.366 + ], + "angle": 0, + "content": "[42] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.369, + 0.907, + 0.41 + ], + "angle": 0, + "content": "[43] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.908, + 0.496 + ], + "angle": 0, + "content": "[44] Gabriel Machado, Edemir Ferreira, Keiller Nogueira, Hugo Oliveira, Matheus Brito, Pedro Henrique Targino Gama, and Jefersson Alex dos Santos. Airround and cv-brct: Novel multiview datasets for scene classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:488-503, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.499, + 0.908, + 0.583 + ], + "angle": 0, + "content": "[45] Gengchen Mai, Chris Cundy, Kristy Choi, Yingjie Hu, Ni Lao, and Stefano Ermon. Towards a foundation model for geospatial artificial intelligence (vision paper). In Proceedings of the 30th International Conference on Advances in Geographic Information Systems, New York, NY, USA, 2022. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.585, + 0.908, + 0.655 + ], + "angle": 0, + "content": "[46] Oscar Manas, Alexandre Lacoste, Xavier Giró-i Nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.908, + 0.726 + ], + "angle": 0, + "content": "[47] Clive Tinashe Marimo, Benedikt Blumenstiel, Maximilian Nitsche, Johannes Jakubik, and Thomas Brunschwiler. Beyond the visible: Multispectral vision-language learning for earth observation. arXiv preprint arXiv:2503.15969, 2025. 2, 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.908, + 0.77 + ], + "angle": 0, + "content": "[48] Valerio Marsocci and Nicolas Audebert. Cross-sensor self-supervised training and alignment for remote sensing, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.908, + 0.843 + ], + "angle": 0, + "content": "[49] Valerio Marsocci, Yuru Jia, Georges Le Bellier, David Kerekes, Liang Zeng, Sebastian Hafner, Sebastian Gerard, Eric Brune, Ritu Yadav, Ali Shibli, et al. Pangaea: A global and inclusive benchmark for geospatial foundation models. arXiv preprint arXiv:2412.04204, 2024. 2, 8, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[50] Matias Mendieta, Boran Han, Xingjian Shi, Yi Zhu, Chen Chen, and Mu Li. Gfm: Building geospatial foundation models via continual pretraining. arXiv preprint arXiv:2302.04476, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7392" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[51] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023. 4, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[52] David Mizrahi, Roman Bachmann, Oğuzhan Fatih Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, and Amir Zamir. 4m: Massively multimodal masked modeling, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[53] Andrea Nascetti, RITU YADAV, Kirill Brodt, Qixun Qu, Hongwei Fan, Yuri Shendryk, Isha Shah, and Christine Chung. Biomasssters: A benchmark dataset for forest biomass estimation using multi-modal satellite time-series. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[54] Vishal Nedungadi, Ankit Kariryaa, Stefan Oehmcke, Serge Belongie, Christian Igel, and Nico Lang. Mmearth: Exploring multi-modal pretext tasks for geospatial representation learning. arXiv preprint arXiv:2405.02771, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.376 + ], + "angle": 0, + "content": "[55] Fernando Paolo, Tsu ting Tim Lin, Ritwik Gupta, Bryce Goodman, Nirav Patel, Daniel Kuster, David Kroodsma, and Jared Dunnmon. xview3-sar: Detecting dark fishing activity using synthetic aperture radar imagery, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[56] Prabhishek Singh and Raj Shree. Analysis and effects of speckle noise in SAR images. In Proc. International Conference on Advances in Computing, Communication, & Automation (ICACCA), 2016. DOI: 10.1109/ICAC-CAF.2016.7748978. Online: http://ieeexplore.ieee.org/document/7748978.16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PmLR, 2021. 3, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[58] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.484, + 0.672 + ], + "angle": 0, + "content": "[59] Ayesha Shafique, Guo Cao, Zia Khan, Muhammad Asad, and Muhammad Aslam. Deep learning-based change detection in remote sensing images: A review. Remote Sensing, 14(4): 871, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[60] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[61] Aidan M Swope, Xander H Rudelis, and Kyle T Story. Representation learning for remote sensing: An unsupervised sensor fusion approach. arXiv preprint arXiv:2108.05094, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[62] Devis Tuia, Konrad Schindler, Begüm Demir, Gustau Camps-Valls, Xiao Xiang Zhu, Mrinalini Kochupillai, Sašo Džeroski, Jan N. van Rijn, Holger H. Hoos, Fabio Del Frate, Mihai Datcu, Jorge-Arnulfo Quiane-Ruiz, Volker Markl, Bertrand Le Saux, and Rochelle Schneider. Artificial intelligence to advance earth observation: a perspective, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[64] Yi Wang, Conrad M Albrecht, Nassim Ait Ali Braham, Lichao Mou, and Xiao Xiang Zhu. Self-supervised learning in remote sensing: A review. arXiv preprint arXiv:2206.13188, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.217 + ], + "angle": 0, + "content": "[65] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eos12: A large-scale multimodal, multitemporal dataset for self-supervised learning in earth observation [software and data sets]. IEEE Geoscience and Remote Sensing Magazine, 11 (3):98-106, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.907, + 0.3 + ], + "angle": 0, + "content": "[66] Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu, Zhe Chen, Wenhai Wang, Xizhou Zhu, Lewei Lu, Tong Lu, et al. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. Advances in Neural Information Processing Systems, 37:69925-69975, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.907, + 0.355 + ], + "angle": 0, + "content": "[67] Xinyu Bai and Feng Xu. Accelerating Diffusion for SAR-to-Optical Image Translation via Adversarial Consistency Distillation, 2024. Online: http://arxiv.org/pdf/2407.06095.16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.906, + 0.426 + ], + "angle": 0, + "content": "[68] Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J. Stewart, Joëlle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. Neural plasticity-inspired foundation model for observing the earth crossing modalities, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.907, + 0.482 + ], + "angle": 0, + "content": "[69] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.907, + 0.564 + ], + "angle": 0, + "content": "[70] Qidong Yang, Jonathan Giezendanner, Daniel Salles Civitarese, Johannes Jakubik, Eric Schmitt, Anirban Chandra, Jeremy Vila, Detlef Hohl, Chris Hill, Campbell Watson, et al. Multi-modal graph neural networks for localized off-grid weather forecasting. arXiv preprint arXiv:2410.12938, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.907, + 0.633 + ], + "angle": 0, + "content": "[71] Zhiping Yu, Chenyang Liu, Liqin Liu, Zhenwei Shi, and Zhengxia Zou. Metaearth: A generative foundation model for global-scale remote sensing image generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.907, + 0.69 + ], + "angle": 0, + "content": "[72] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.907, + 0.747 + ], + "angle": 0, + "content": "[73] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600-612, 2004. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.907, + 0.801 + ], + "angle": 0, + "content": "[74] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.907, + 0.872 + ], + "angle": 0, + "content": "[75] Linying Zhao and Shunping Ji. Cnn, rn, or vit? an evaluation of different deep learning architectures for spatio-temporal representation of sentinel time series. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16:44-56, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[76] Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7393" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.124, + 0.093, + 0.486, + 0.134 + ], + "angle": 0, + "content": "learning in remote sensing: A comprehensive review and list of resources. IEEE geoscience and remote sensing magazine, 5(4):8-36, 2017. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7394" + } + ] +] \ No newline at end of file diff --git a/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/c9c92fd3-e6cf-49ea-bfb4-aa3c5b130a2c_origin.pdf b/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/c9c92fd3-e6cf-49ea-bfb4-aa3c5b130a2c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..24beadff37db576c9b8c081a7dd16c7dd532102b --- /dev/null +++ b/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/c9c92fd3-e6cf-49ea-bfb4-aa3c5b130a2c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:192e86c9def24dc61d8c3ac87c58d189565c4344374a70e6ff18d82f6754b674 +size 26296577 diff --git a/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/full.md b/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c9fe381a9facc9f8908ea88baa481dd72a9fe76f --- /dev/null +++ b/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/full.md @@ -0,0 +1,295 @@ +# TerraMind: Large-Scale Generative Multimodality for Earth Observation + +![](images/6a508ecc15dcc21c6803d3b04d4f20a8dcabf19832964f2fe4ee1a9d82220c68.jpg) + +$^{1}$ IBM Research - Europe $^{2}$ ETH Zurich $^{3}$ Forschungszentrum Jülich $^{4}$ European Space Agency $\Phi$ -Lab $^{5}$ NASA IMPACT $^{6}$ University of Iceland + +johnannes.jakubikl@ibm.com + +![](images/c2719cca28b3707036b1e21c1c8178c88eade472b909cf3ef0639ccd403e1524.jpg) +Figure 1. TerraMind represents the first any-to-any generative, and large-scale multimodal model for Earth observation pre-trained on 500 billion tokens from global geospatial data. The model digests multi-scale representations at pixel-level and token-level simultaneously. TerraMindv1 unlocks (i) generation, (ii) zero-shot and finetuning applications, and (iii) "Thinking-in-Modalities" finetuning and inference. + +![](images/b6bb9405779306eee986642cf6d11a4df535fdd5d3976835632bf50bec8e86bd.jpg) + +![](images/f7c9a63048871fd77a2eec6e0ea9896413b1fa97b6cae2bce0e8cf481bc98f2a.jpg) + +# Abstract + +We present TerraMind, the first any-to-any generative, multimodal deep learning model for Earth observation (EO). Unlike other approaches, TerraMind is pretrained on dual-scale representations combining both token-level and pixel-level data across modalities. On a token level, TerraMind encodes high-level contextual information to learn cross-modal relationships, while on a pixel level, TerraMind lever + +ages fine-grained representations to capture critical spatial nuances. In this paper, we demonstrate that (i) TerraMind achieves beyond state-of-the-art performance in community-standard benchmarks, (ii) TerraMind can leverage "thinking in modalities" (TiM)—the capability of generating additional artificial data during finetuning and inference to improve the model output—and (iii) TerraMind's dual-scale early fusion approach results in well-structured embedding spaces. Models and code have been open-sourced at https://huggingface.co.ibm-esa-geospatialandhttps://github.com.ibm/terramind. + +# 1. Introduction + +Earth observation (EO) increasingly benefits from multimodality because of the important integration of complementary information from different data sources. This becomes particularly relevant as EO is spatiotemporally sparse due to low revisiting times or weather phenomena like cloud coverage. Vice versa, for computer vision, EO data is an important playground for the development of new approaches as there is significant publicly available data of very high quality and complexity. The available modalities range from sensors of different satellite missions to relevant complementary information like digital elevation. + +In this work, we introduce TerraMind as the first any-to-any generative multimodal model for EO. With TerraMind, we introduce a dual-scale pretraining on pixel-level and token-level and demonstrate benefits over training primarily on tokens. TerraMind encodes high-level contextual information in tokens to enable correlation learning and scaling, while, additionally capturing important fine-grained representations using pixel-level inputs. During pretraining, TerraMind predicts masked target tokens so that our pretraining objective boils down to a cross-modal patch classification problem that results in high-quality latent spaces. TerraMind is pretrained on a custom global-scale geospatial dataset named TerraMesh with nine million samples that have been aligned spatiotemporally and across modalities [7]. In addition to radar and optical satellite images of the Copernicus Sentinel-1 (S-1) and Sentinel-2 (S-2) missions, our dataset contains task-specific modalities such as land use/land cover (LULC) and normalized difference vegetation index (NDVI) maps, metadata like digital elevation models (DEM) and geographic coordinates, and natural language in the form of captions. To the best of our knowledge, TerraMind represents the first truly generative, multimodal deep learning model for EO. Additionally, in contrast to other recent models that utilize masked autoencoders like [54], contrastive learning, or diffusion techniques, TerraMind uniquely demonstrates benefits of leveraging token-based pretraining for EO. + +We provide an overview of TerraMind's performance in a community-standard benchmark [49] in Figure 2 and highlight the any-to-any generative capabilities of TerraMind in Figure 3. Our key contributions are as follows: (i) We introduce a dual-scale approach for generative multimodal pre-training leveraging data on pixel-level and token-level, which outperforms other fusion approaches and enhances embedding space structures. (ii) We introduce thinking in modalities - similar to chain-of-thought approaches in LLMs - for multi-modal models in EO, demonstrating that infusing generated data during finetuning improves the performance. (iii) We demonstrate that TerraMind outperforms other geospatial foundation models both in unimodal and multimodal settings. + +# 2. Related Work + +Computer vision in Earth observation. Computer vision (CV) has significantly advanced EO [76]. Many CV techniques, originally developed for natural image processing, have been adapted to EO [62], often with minimal modifications. A wide range of tasks benefit from these methods, including classification [16], semantic segmentation [72] (e.g., land cover mapping [20, 21]), change detection [59] (e.g., disaster response [19]), object detection [39] (e.g., vessel identification [55]), and regression (e.g., biomass estimation [53]). Deep learning architectures like CNNs [75] and Vision Transformers (ViTs) [17] have demonstrated strong performance, often surpassing traditional remote sensing (RS) methods. However, EO presents unique challenges, including diverse sensor modalities [4] and geospatial heterogeneity [46]. An emerging paradigm in EO is self-supervised learning (SSL) [64] and geospatial foundation models (GFMs) [45], which aim to leverage vast amounts of unlabeled RS data to develop general purpose task models [32]. While off-the-shelf CV models have shown promising results [36], they do not fully exploit the unique characteristics of geospatial data. Many GFMs still rely on generic CV architectures [50], which were not explicitly designed to handle the complexities of EO, such as heterogeneous sensor sources (e.g., optical, radar, DEM) [29], integrated with auxiliary data (e.g., text) [42, 47], and expert knowledge (e.g., prioritizing specific bands or indexes). In this direction, TerraMind better integrates domain-specific properties, developing a customized and expandable multimodal learning strategy. + +Multimodality in CV. Multimodal CV is driven by the integration of diverse data streams [69], such as natural images [74], natural language text [10], temporal video data [58], and weather [70], within large foundation models [8]. + +![](images/89ab3f4a24a8f0f22c145b98d7aa131473829751dc280f48d7fdfe81a559cdff.jpg) +Figure 2. TerraMind outperforms other geospatial foundation models on PANGAEA benchmark [49] in finetuning. Performance is measured in mIoU and min-max scaled per dataset. + +![](images/26ad38ec634ed36b3b55e5dbf6b8398dd71b67db16bdc57353ddbf06d8c0c07f.jpg) +Figure 3. Chained generation example of TerraMindv1-B starting from either optical, radar, or digital elevation data. Left is input, middle is artificially generated data by TerraMind, right represents ground truths and tokenizer reconstructions, respectively. + +Starting from the alignment of images and texts [57], these models moved beyond simple feature extraction, towards nuanced contextual understanding. The ability to combine several modalities allows for unprecedented capabilities in complex tasks [30], evidenced by the rapid advancement of multimodal Large Language Models (MLLMs) [30], that excel in tasks such as scene understanding [12], visual question answering [18], and video analysis [24]. Recent advances in architectures [31] and large scale pre-training [11] have enabled the development of models that learn highly effective cross-modal representations [41], which can then be adapted to a wide variety of downstream tasks [66]. + +Multimodality in EO. Multimodality in EO originates from data fusion and is typically understood as the integration of SAR and optical data [13, 25, 28, 38] or the combination of optical data with vector data [5]. Some studies have explored alternative combinations of data. In [15], the authors introduce a contrastive framework for comparing RS images and street views. Even different optical sensors can be considered different modalities [48, 61]. Similarly, several multi-view images (i.e. multimodal) datasets [26, 44, 54] are introduced. More recent approaches combined text and images [40], both for discriminative [42] and generative [34] purposes. Lately, different GFMs are trained in a multimodal way [4, 54, 68], still focusing either on a specific set of modalities (e.g., vision [68], [3]) or tasks (e.g., generative [34]). Compared to multi-scale high-quality generation models for optical data, like MetaEarth [71], our approach allows to generate any modality from any other pretraining modality. To the best of our knowledge, no existing model has combined a wide and diverse amount of modalities both for discriminative and generative purposes, as TerraMind does. We provide a comparison in Table 1. + +# 3. Dataset + +For the pretraining of TerraMind and its tokenizers, we create a multimodal dataset called TerraMesh [7], which will + +
ModelModalitiesAny-to-Any GenerationMulti-Scale Features
RemoteCLIPoptical, textXX
CROMAoptical, radarXX
AnySataerial, optical, radar, NAIPXX
DeCURoptical, radarXX
DOFAoptical, radar, hyperspectral, NAIPXX
MetaEarthoptical (unimodal)X✓
Galileooptical, radar, elevation, weather, location, population, ...X✓
TerraMindoptical, radar, land use, elevation, vegetation index, location, text✓✓
+ +Table 1. Comparison of TerraMind to other model architectures. TerraMind represents a first-of-its-kind generative, multimodal model. + +be open-sourced to the community. TerraMesh builds on existing datasets, which we expand by adding modalities from external data sources or by applying pseudo-labeling. We provide an overview of the aligned image modalities and a detailed dataset description in the supplementary material. + +Base datasets. TerraMesh is based on SSL4EO-S12 [6, 65] and MajorTOM-Core [23], two unlabeled remote sensing datasets containing co-aligned radar and optical imagery from Sentinel-1 and Sentinel-2 satellites. SSL4EO-S12 has lower geographic coverage but is multi-seasonal. MajorTOM-Core covers most of the Earth's land surface at a single timestamp. For MajorTOM-Core, we apply a subsampling scheme based on LULC classes and ecoregions. TerraMesh includes a total of approximately 9 million globally distributed samples from both Sentinel-1 and Sentinel-2, + +each measuring $264 \times 264$ pixels at $10\mathrm{m}$ resolution. + +Additional modalities. We obtain co-aligned yearly LULC maps by ESRI with nine land use classes. Additionally, we leverage SEnSeI v2 [22] as a cloud and ice annotation model and update the ESRI LULC classes for better spatiotemporal alignment. NDVI maps are computed using the corresponding spectral bands from Sentinel-2. DEM is extracted from the Copernicus DEM 30m dataset [2], which provides global coverage of the Earth's elevation at a 30m resolution. Captions are generated synthetically by constructing RGB images from Sentinel-2 patches using the corresponding spectral bands and processing them with LLaVANext [37]. A tailored prompt guides the model to describe the content of each image as described in [47]. For geolocations, we round latitude and longitude from the center of each patch to the nearest quarter degree and store the discretized coordinates as strings in a pre-defined format. + +# 4. Methods + +TerraMind pretraining is two-staged following [52]. We first pretrain unimodal tokenizer models, tokenize the modalities, and then leverage token-level and pixel-level input to pretrain the TerraMind encoder-decoder architecture. We describe those individual stages in the following. + +# 4.1. Tokenization + +We develop modality-specific tokenizers to encode each modality into a sequence of discrete tokens for pretraining and decode token sequences back to images. Thus, TerraMind is in principle compatible with any modality, as long as it can be tokenized and aligned with other modalities. For reasons of space, we delegate most experiments related to the tokenizer performances to the supplementary material. + +Image-like modalities. We train autoencoder-based architectures with a quantization step in the bottleneck for image-like modalities such as S-1, S-2, LULC, NDVI, and DEM. Tokenizer encoders process an input image and generate a latent representation for each $16 \times 16$ patch, which is then discretized with finite-scalar-quantization (FSQ) [51] into one of $N$ codewords. All tokenizers use a vocabulary size of 16K besides the simpler LULC modality for which we use 4K. These codewords are then used by the diffusion decoder to reconstruct the original image. The benefit of leveraging diffusion decoders lies in facilitating cross-modal generation in TerraMind by transforming tokens back into images. By mapping each codeword to a unique integer in $\{0, 1, \dots, N - 1\}$ , we obtain discrete tokens for each image patch. We pretrain the tokenizers in a self-supervised setting. FSQ as quantization method enhances training stability [51] compared to vector quantization [63] by eliminating the need for codebook-related loss terms. Notably, FSQ is + +heavily influenced by ideas of neural compression [27]. For example, on 12-band S-2 images, we achieve compression rates of over $3000\mathrm{x}$ by applying quantization. We summarize the architecture of our tokenizers in Figure 4. The main objective of the overall tokenizer is to encode image patches consistently into discrete tokens based on semantic similarity to enable cross-modal correlation learning. Therefore, the loss of some details is an expected trade-off, since the focus is on grouping similar patches rather than preserving all fine-grained features. Naturally, more accurate reconstructions facilitate cross-modal generation, however the main focus of the pretraining lies on consistent cross-modal correlation learning. We provided further details on the pretraining of the tokenizers in the supplementary material. + +![](images/71da8b1dfb813f0f4d5e280206571aa38f59c2eb458f021352749f9465058d54.jpg) +Figure 4. Tokenizer for image-like modalities combining finite-scalar quantization [51] with diffusion decoding. + +Sequence-like modalities. We treat both captions and geolocations as text and use a single text tokenizer to process both modalities. By discretizing the geographic coordinates and representing them as strings, we introduce special coordinate tokens into the vocabulary. This allows us to encode geolocations as a sequence of discrete tokens, beginning with a latitude token followed by a longitude token. For textual data, we modify the existing WordPiece tokenizer [33]. + +# 4.2. Pre-training + +Architecture. TerraMind uses a symmetric Transformer-based encoder-decoder architecture proposed in [52], which is designed to process sequences of multimodal tokens. In addition to discrete tokens, TerraMind accepts pixel-level inputs, specifically satellite imagery and digital elevation maps. For pixel-level inputs, we apply learnable patch-wise linear projections to generate patch embeddings for each $16 \times 16$ patch, similar to the approach used in ViT [17]. + +Dual-scale early fusion. In contrast to [52], we not only embed token-level data but additionally leverage pixel-level data across a range of input modalities to introduce a dual-scale feature representation to support the structuring of the embedding space. Both tokens and patches represent a 16x16 pixel area. Tokens represent this area via a single discrete integer value, while the image patches describe the same area with the actual floating point sensor data. Thus, during pretraining, the model not only learns a correlation between modalities (i.e., cross-modal learning) but also between dif + +ferent levels of abstraction within the same modality. The low-level token information enables cross-modal correlation learning, while adding pixel level input accounts for spatial nuances. Based on dual-scale features the model further learns to better structure pixel-level data in the embedding space via the corresponding information from the discrete token. We illustrate the pretraining paradigm in Figure 5. The model is agnostic to processing tokens or patches in the input space, while the target is generally token-level data. We use six pixel-level modalities and eight token-level modalities. + +![](images/e0986ca258c6aa8b5d527085e8ce35c951950eed3bc2982fada04e93f0661922.jpg) +Figure 5. Illustration of the pre-training task. Given an encoded multimodal sample of random subsets of patches and input tokens, the decoder predicts target tokens for the masked input. + +Masking strategy. TerraMind applies a masked modeling approach in the token space following [52]. The model leverages a set of randomly selected target tokens that have to be reconstructed from a randomly selected set of input tokens and pixel-level data. During pre-training, we sample input and target data from a Dirichlet distribution. + +We opt for masked token reconstruction to familiarize the model with the absence of entire modalities, which is crucial for a high usability of a multimodal model in Earth observation. During pre-training, the model learns an internal representation of unseen modalities which is expected to benefit a range of downstream applications. In addition, sampling input and target tokens improves the computational efficiency of the pre-training, as each token is a compressed representation of a patch with compression factors of between 250x and 3000x depending on the modality. Finally, without tokenized representations of the image-like modalities, it is challenging to learn the correlation to sequence-like modalities. The overall training objective of TerraMind boils down to a cross-modal patch-level classification problem optimized via a cross entropy loss: + +$$ +\mathcal {L} _ {\mathrm {C E}} = - \sum_ {i = 1} ^ {N} y _ {i} \log \left(p _ {i}\right), \tag {1} +$$ + +where $y_{i}$ is the one-hot encoded true class of token $i$ , $p_{i}$ is the predicted probability for token $i$ , $N$ is the total number of possible tokens. Interestingly, we can infer an upper bound loss for a random model where the cross entropy loss will collapse to the natural logarithm of the vocabulary size $\mathcal{L}_{\mathrm{CE,random}} = -\sum_{i=1}^{N} y_{i} \log \left( \frac{1}{N} \right) = \log(N)$ . + +Scaling. We trained three versions of TerraMind scaling across model size, compute, and data. In addition, we pretrain different versions of TerraMind with respect to the number of dual-scale features. TerraMindv1-B is pre-trained on 500B tokens for 6 days on 32 NVIDIA A100 GPUs. The model uses dual-scale features from both token-level and pixel-level. During initial experiments, we observed significant improvements from scaling model size when switching from a tiny backbone to a small backbone to a base backbone. Therefore, we pre-trained TerraMindv1-L on a large backbone with 500B tokens on 32 NVIDIA A100 GPUs trained for 10 days. Finally, to better understand the effect of scaling across the dual-scale feature representation, we pre-train TerraMindv1-B-single as a single-scale model on primarily token-level data with optical S-2 L2A data as only pixel-level input (compared to pixel-level S-2 L1C, S-2 RGB, S-1 GRD, S-1 RTC, and DEM in TerraMindv1-B and -L). TerraMindv1-B-single is pretrained on 500B tokens from over one million samples for 6 days on 32 NVIDIA A100 GPUs. We summarize the scaling behavior in model size, compute, and data in Figure 9 of the supplementary material. We additionally provide final validation losses in Table 9 comparing v1-B and v1-L with the theoretical random loss. + +# 4.3. Generation + +Once pretrained, TerraMind can generate tokens for any modality, conditioned on any subset of input modalities. The generative capabilities unlock various zero-shot tasks, such as water body segmentation. For the generation of image-like modalities, the decoder receives mask tokens for the modality to be generated and predicts the corresponding tokens based on the encoded input. For sequence-like modalities, the decoder generates the output autoregressively. After generating tokens from the target modality, the corresponding tokenizer decoder allows to map from token-space to image or text space. TerraMind further supports chained generation which ensures consistency across generated modalities. The chained generation represents a conditional probability distribution where the prior probability distribution is determined by the input modality, and all subsequent modalities are generated conditioned on the input modality and potentially other generated modalities. + +# 4.4. Thinking-in-Modalities + +Thinking in Modalities (TiM) is a recursive fine-tuning and inference technique designed to enhance multimodal learning by leveraging the generative capabilities of the model itself. Given an input $x \in \mathcal{X}$ (e.g., an optical satellite image), the model first generates additional synthetic modalities $\tilde{x} = f_{\mathrm{gen}}(x)$ on a token-level using a learned generative function $f_{\mathrm{gen}}$ . These generated tokens are then concatenated with the original input and jointly processed by the downstream + +model $f$ (e.g., TerraMind encoder with a segmentation head), yielding the final output $y = f(x, f_{\mathrm{gen}}(x))$ . This formulation allows the model to reason over both observed and inferred modalities, effectively enriching the input space. TiM can leverage multiple generated modalities which are then generated in a chained approach. For example, for $k$ modalities, the input is augmented with newly generated modalities: + +$$ +\tilde {x} ^ {(k + 1)} = \tilde {x} ^ {(k)} \cup f _ {\text {g e n}} (\tilde {x} ^ {(k)}), \tag {2} +$$ + +and the final model output is described by: + +$$ +y = f \left(\tilde {x} ^ {(K)}\right). \tag {3} +$$ + +This recursive augmentation mimics a chain-of-thought process, enabling the model to iteratively refine its internal representation, particularly in scenarios with missing modalities. + +# 5. Experiments + +In this section, we describe the performance gains resulting from TerraMind and experiment with the unlocked capabilities of any-to-any generation and Thinking-in-Modalities. + +# 5.1. Foundational experiments + +Multimodality vs. unimodality. As a first motivational experiment, we outline the benefit of using multimodal data in Earth observation at the example of water body mapping. Specifically, we leverage the ViT-B encoders from the unimodal tokenizer models for S-1, S-2, and LULC, concatenate their embeddings, and train a segmentation head with four ConvNeXt [43] blocks as a late fusion approach. The results in Table 2 (left) suggest that regardless of which modalities we combine, the combination of two modalities always outperforms each unimodal model. Combining all three modalities achieves the best overall performance. + +
InputLate fusionToken-level fusion
S-161.0163.94 (2.93pp↑)
S-272.7076.32 (3.62pp↑)
LULC71.7770.96 (0.81pp↓)
S-1 + S-273.8376.74 (2.91pp↑)
S-1 + LULC73.8673.76 (0.10pp↓)
S-2 + LULC75.6577.04 (1.39pp↑)
S-1 + S-2 + LULC76.0076.88 (0.88pp↑)
+ +Token-level fusion vs. late fusion. In Table 2 (right), we investigate the effects of fusing the inputs on a token level + +through masked token reconstruction. We observe that token-level fusion outperforms late fusion. The performance gains are particularly high when LULC data is not available. This suggests that early fusion captures an internal representation of the multimodal state—especially pronounced for LULC—that benefits fine-tuning. With those findings in mind, we will explore the effects of using additional multi-modal pixel-level input in a dual-scale pretraining in Section 5.5. + +# 5.2. Generation experiments + +TerraMind supports any-to-any generation. In the following, we provide examples of the generation performance starting from: (i) an information-rich modality, like optical S-2 L2A data, and (ii) minimal information based on the geolocation. In Figure 3, we observe that TerraMind performs strongly in generating image-like modalities like S-1, LULC, and DEM from optical S-2 L2A data. We provide a quantitative overview on the quality of the generations on unseen validation data in Table 3. Overall, we observe an interesting asymmetry in the generative performance of TerraMind where (a) radar-to-optical generation achieves reasonable quality in terms of SSIM and PSNR – indicating structural and visual fidelity with some perceptual degradation – and (b) optical-to-radar generation yields higher PSNR values but lower SSIM, suggesting visually plausible outputs that lack strong structural alignment. The quality of generated DEM suggests to be structurally very strong, but noisy. The errors for DEM generations suggest that the level of altitude is difficult to infer for the model. We compare these scores with the reconstruction quality of the auto-encoding tokenizers in the supplementary material that can serve as upper bounds. Additionally, we provide experiments on the generation quality using token-level instead of pixel-level inputs. Finally, we demonstrate the quality of generations at kilometer scale in Figures 19 and 20. + +Table 2. Water body mapping on Sen1Floods11 [9] measured in IoU on water class. Model sizes and architectures are comparable. Left column: Late fusion of tokenizers. The average improvement of full multimodality over the individual unimodal performance is 7.5pp IoU. Right column: Finetuning results of TerraMindv1-B-single as a mid fusion approach based on masked correlation learning. Gains over late fusion in percentage points in parentheses. + +
ModalitiesMAE↓RMSE↓SSIM↑PSNR↑
S-1 GRD → S-2 L2A0.0740.1160.75026.210
S-1 GRD → DEM163.0320.80.87820.694
S-1 GRD → NDVI0.1800.2250.43818.990
S-1 RTC → S-2 L2A0.1130.1940.69524.251
S-1 RTC → DEM298.8799.20.87320.009
S-1 RTC → NDVI0.1720.2110.46519.529
S-2 L2A → S-1 GRD2.9423.8770.53128.678
S-2 L2A → S-1 RTC2.6363.3910.43028.993
S-2 L2A → DEM215.8745.50.94220.616
+ +Table 3. Quantitative evaluation of generations on unseen global validation dataset using 10 diffusion steps. MAE and RMSE metrics are in physical units: meter (DEM), reflectance (S-2), and db (S-1). + +![](images/ecd65734e16cd2da1375f189a89942b655bab4f6d4ea51b928ebdd4441055148.jpg) +(a) Input: S-2 L2A data capturing Singapore in January 2025. + +![](images/93cf79200582ee94491ba9ee7aa21762717149e729068fe33c90012c0c793f3d.jpg) +(b) Generation: S-1 RTC composition generated by TerraMind. + +![](images/a9582e25b72bba7331bcf2355f6ee19a7a7832cdd841268bd8932a766d10c44f.jpg) +(c) Input: S-2 L2A data capturing Northern Spain in January 2025. + +![](images/2339bc758ec8ffc0eb95784cdb7350fbdf51c620f199fe7eeeba6e7f859f7db4.jpg) +(d) Generation: S-1 GRD composition generated by TerraMind. +Figure 6. Generated S-1 imagery using TerraMind. We provide large-scale visualizations in the supplementary material. + +# 5.3. Zero-shot experiments + +Based on its generative capabilities, TerraMind unlocks several zero-shot applications, like land-use segmentation, water body mapping, geo-localization, and vegetation mapping. In the following, we focus on water body mapping and geo-localization as image- and sequence-level zero-shot tasks. + +Water body mapping. In Table 4, we compare the zero-shot performance of TerraMind with its fine-tuned performance and other finetuned benchmarks for water body mapping. Overall, TerraMindv1-B achieves a zero-shot IoU of $45.4\%$ compared to SOTA-level fine-tuning performance of $82.2\%$ of DeCUR. In ablations with TerraMindv1-B-single trained on DynamicWorld LULC data, we boost this to up to $69.8\%$ suggesting that TerraMind harnesses up to over $80\%$ of the SOTA performance in zero-shot setting. Additionally, it's notable that none of the benchmarking model can be applied in a zero-shot context, highlighting the relevance of TerraMind's capabilities. + +
ModelInputTypeIoUWater
TerraMindv1-BS-2zero-shot45.40
TerraMindv1-B-singleS-2zero-shot69.75
Prithvi 2.0 / DeCUR / ...zero-shotN/A
Baseline [9]S-2finetune31.25
Prithvi 2.0 300MS-2finetune80.97
DeCURS-2finetune82.17
+ +Geo-localization. TerraMind is able to predict the geolocation of a specific data instance. To better visualize the geolocation capabilities, we prompt the model for the most + +likely locations of the land use class "bare land" (deserts etc.) in a Monte-Carlo-sampling in Figure 7. The probability distribution of the model fits the expectation of where to find bare land, highlighting the Sahara region and middle-east, as well as Mexico and Southern California. + +![](images/442bc09a3bff1182ff8bb275d44d364622dcffd33c27729448ba71c7b3172b1d.jpg) +Figure 7. Prediction distribution of the land use class "bare land" with a sampling temperature of $T = 1.0$ using TerraMindv1-B-single. TerraMind has an accurate internal representation of the geolocation of specific contexts, like land use classes. + +# 5.4. Few-shot experiments + +TerraMind is trained via a cross-modal patch classification objective. Thus, we expect a well-structured latent space that clusters different concepts accurately. To investigate our hypothesis, we apply 1-Nearest-Neighbor (1-NN) classification experiments in the community-standard setting of 1-shot 5-way on two datasets: EuroSAT and METER-ML. In those experiments, there are no weight updates of any kind, so that we can assess the quality of the embedding space structure. In Table 5, we observe that TerraMind outperforms several other benchmarks from both the CV and EO domain on the EuroSAT dataset by at least 10pp in accuracy. Our results further show that for methane source classification on METER-ML, TerraMind outperforms benchmark models and generalizes to high-resolution NAIP data with one order of magnitude higher resolution than the pre-training data. We present additional experiments with other few-shot settings in the supplementary material. + +Table 4. Zero-shot results of TerraMind on water body mapping compared to fine-tuned performance of benchmarks. + +
ModelInputEuroSATMETER-ML
CLIP-ViT-B/16S-2 RGB57.0029.15
CLIP-ViT-B/16NAIP-32.01
DeCURS-2 L1C50.5427.87
Prithvi 1.0 100MS-2 L1C60.1126.08
Prithvi 2.0 300MS-2 L1C61.0628.26
TerraMindv1-BS-2 L1C70.8333.90
TerraMindv1-BNAIP-32.23
+ +Table 5. 1-shot 5-way classification results on EuroSAT and METER-ML measured in mean accuracy $\uparrow$ , averaged over 200 runs. TerraMind outperforms benchmarks from CV and EO domain, suggesting a well-structured latent space. + +
ModelBurnSr*MADOS*PASTISSen1Fl11FBP*DEN*CTM-SSSN7*AI4Farms*Avg. mIoUAvg. Rank
CROMA82.4267.5532.3290.8951.8338.2949.3859.2825.6555.296.61
DOFA80.6359.5830.0289.3743.1839.2951.3361.8427.0753.598.22
GFM-Swin76.9064.7121.2472.6067.1834.0946.9860.8927.1952.4210.00
Prithvi 1.0 100M83.6249.9833.9390.3746.8127.8643.0756.5426.8651.0011.00
RemoteCLIP76.5960.0018.2374.2669.1931.7852.0557.7625.1251.6611.22
SatlasNet79.9655.8617.5190.3050.9736.3146.9761.8825.1351.6510.67
Scale-MAE76.6857.3224.5574.1367.1935.1125.4262.9621.4749.4311.44
SpectralGPT80.4757.9935.4489.0733.4237.8546.9558.8626.7551.8710.11
S.-S12-MoCo81.5851.7634.4989.2653.0235.4448.5857.6425.3853.0210.06
S.-S12-DINO81.7249.3736.1888.6151.1534.8148.6656.4725.6252.5110.89
S.-S12-MAE81.9149.9032.0387.7951.9234.0845.8057.1324.6951.6912.39
S.-S12-Data2Vec81.9144.3634.3288.1548.8235.9054.0358.2324.2352.2210.72
UNet Baseline84.5154.7931.6091.4260.4739.4647.5762.0946.3457.584.89
ViT Baseline81.5848.1938.5387.6659.3236.8344.0852.5738.3754.1310.28
TerraMindv1-B82.4269.5240.5190.6259.7237.8755.8060.6128.1258.353.94
TerraMindv1-L82.9375.5743.1390.7863.3837.8955.0459.9827.4759.573.44
+ +Table 6. Performance evaluation of TerraMind using the PANGAEA evaluation protocol indicates higher mIoU values (↑) and lower rank values (↓). The best model per column is highlighted in bold, the second best is underscored. We indicate unimodal datasets with *. Encoders are frozen for pretrained models, while U-Net and ViT baselines are trained from scratch for each specific task. + +# 5.5. Fine-tuning experiments + +Besides the novel capabilities that TerraMind introduces, we benchmark the fine-tuning performance of TerraMind in both unimodal and multimodal settings following the community-standard PANGAEA benchmark [49]. We summarize the results in Table 6. Overall, TerraMindv1-B outperforms all other GeoFMs by at least 3pp avg. mIoU. Importantly, we observe that TerraMind is the only foundation model approach in EO that across the PANGAEA benchmark outperforms task-specific U-Net models. Performance increases by approximately 2pp avg. mIoU for TerraMindv1-L, with a peak of 5pp in multimodal datasets. Furthermore, TerraMindv1-L outperforms also specialised ViT baselines by 5pp avg. mIoU. Note that per suggestion of the PANGAEA authors, we exclude the xView2 and BioMassters task as we could not reproduce the reported performances. Finally, we assess the impact of leveraging multimodal data as input to TerraMindv1-B compared to utilizing either optical or radar data as unimodal input to better understand the effect of leveraging multimodal data in finetuning. We observe that across all three multimodal tasks, TerraMindv1-B performs best with access to both optical and radar data. + +
PASTISSen1F11CTM-SS
S-120.0480.3924.45
S-240.2089.5750.90
S-1 + S-240.5190.6255.80
+ +Table 7. Benefit of using multimodal input in the PANGAEA benchmark reported in mIoU $(\%)\uparrow$ + +# 5.6. Thinking in modalities + +We additionally evaluate the value of TiM tuning on water body mapping. We use S-1 or S-2 to generate artificial LULC data as additional input. Our results in Table 8 indicate a superior performance of TiM tuning compared to leveraging uni-modal data by up to 2pp mIoU. This finding points us in the direction of TerraMind being able to generate data that improve downstream task performance. We provide additional results in the appendix. + +
Fine-TuningInputIoUWatermIoU
TerraMindv1-BS-168.0081.06
TerraMindv1-BS-282.2689.70
TerraMindv1-B TiMS-1 + gen. LULC72.2583.65
TerraMindv1-B TiMS-2 + gen. LULC84.7591.14
+ +Table 8. Thinking-in-modalities (TiM) tuning compared with standard full fine-tuning approaches on the Sen1Floods11 dataset. + +# 6. Conclusion + +TerraMind's approach of combining token-level and pixel-level data has unlocked a range of new model capabilities in EO. TerraMind demonstrates not only beyond state-of-the-art performance in community-standard benchmarks, it also represents the first fully generative multimodal model in the domain. Because of the ability of integrating heterogeneous data sources, we expect that TerraMind-like models will expand to multi-temporal, multi-resolution, and hyperspectral data to fully leverage the data rich ecosystem available in the Earth Observation domain. + +# References + +[1] A. Hore and D. Ziou. Image quality metrics: PSNR vs. SSIM. In Proc. 20th International Conference on Pattern Recognition (ICPR), pp. 2366-2369, 2010. 16 +[2] European Space Agency. Copernicus dem. http://dx.doi.org/10.5270/ESA-c5d3d65, 2022.4 +[3] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Anysat: An earth observation model for any resolutions, scales, and modalities. arXiv preprint arXiv:2412.14123, 2024. 3 +[4] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Omnisat: Self-supervised modality fusion for earth observation, 2024. 2, 3 +[5] Nicolas Audebert, Bertrand Le Saux, and Sébastien Lefèvre. Joint Learning from Earth Observation and OpenStreetMap Data to Get Faster Better Semantic Maps. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1552-1560, 2017. 3 +[6] Benedikt Blumenstiel, Nassim Ait Ali Braham, Conrad M Albrecht, Stefano Maurogiovanni, and Paolo Fraccaro. SSL4EOS12 v1.1 - A Multimodal, Multiseasonal Dataset for Pretraining. arXiv preprint arXiv:2503.00168, 2025. 3, 13 +[7] Benedikt Blumenstiel, Paolo Fraccaro, Valerio Marsocci, Johannes Jakubik, Stefano Maurogiovanni, Mikolaj Czerkawski, Rocco Sedona, Gabriele Cavallaro, Thomas Brunschwiler, Juan Bernabe-Moreno, and Nicolas Longépé. Terramesh: A planetary mosaic of multimodal earth observation data. arXiv preprint arXiv:2504.11172, 2025. 2, 3 +[8] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2 +[9] Derrick Bonafilia, Beth Tellman, Tyler Anderson, and Erica Issenberg. Sen1floods11: A georeferenced dataset to train and test deep learning flood algorithms for sentinel-1. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020. 6, 7 +[10] Florian Bordes, Richard Yuanzhe Pang, Anurag Ajay, Alexander C Li, Adrien Bardes, Suzanne Petryk, Oscar Manas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 2 +[11] Jize Cao, Zhe Gan, Yu Cheng, Licheng Yu, Yen-Chun Chen, and Jingjing Liu. Behind the scene: Revealing the secrets of pre-trained vision-and-language models. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 565-580. Springer, 2020. 3 +[12] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. Maplm: A real-world large-scale vision-language benchmark for map and traffic scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21819-21830, 2024. 3 + +[13] Yuxing Chen and Lorenzo Bruzzone. Self-supervised change detection in multi-view remote sensing images. arXiv preprint arXiv:2103.05969, 2021. 3 +[14] Chenwei Wang, et al. SAR Target Image Generation Method Using Azimuth-Controllable Generative Adversarial Network. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS), Vol. 15, 2022. Online: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9933645&tag=1.16 +[15] Fabian Deuser, Konrad Habel, and Norbert Oswald. Sample4geo: Hard negative sampling for cross-view geolocation. arXiv preprint arXiv:2303.11851, 2023. 3 +[16] Ivica Dimitrovski, Ivan Kitanovski, Dragi Kocev, and Nikola Simidjievski. Current trends in deep learning for earth observation: An open-source benchmark arena for image classification. ISPRS Journal of Photogrammetry and Remote Sensing, 197:18-35, 2023. 2 +[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 2, 4 +[18] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 2023. 3 +[19] Victor Durnov. xview2 1st place solution. 2 +[20] Adam Van Etten, Dave Lindenbaum, and Todd M. Bacastow. Spacenet: A remote sensing dataset and challenge series, 2019. 2 +[21] Casper Fibaek, Luke Camilleri, Andreas Luyts, Nikolaos Dionelis, and Bertrand Le Saux. PhilEO Bench: Evaluating Geo-Spatial Foundation Models, In Proc. Int Geoscience and Remote Sensing Symposium (IGARSS), 2024. 2 +[22] Alistair Francis. Sensor independent cloud and shadow masking with partial labels and multimodal inputs. IEEE Transactions on Geoscience and Remote Sensing, 2024. 4, 13 +[23] Alistair Francis and Mikolaj Czerkawski. Major tom: Expandable datasets for earth observation. arXiv preprint arXiv:2402.12095, 2024. 3, 13 +[24] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 3 +[25] Anthony Fuller, Korean Millard, and James R. Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders, 2023. 3 +[26] Anatol Garioud, Nicolas Gonthier, Loic Landrieu, Apolline De Wit, Marion Valette, Marc Poupee, Sebastien Giordano, and Boris Wattrelos. FLAIR: a country-scale land cover semantic segmentation dataset from multi-source optical imagery. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3 + +[27] Carlos Gomes, Isabelle Wittmann, Damien Robert, Johannes Jakubik, Tim Reichelt, Michele Martone, Stefano Maurogiovanni, Rikard Vinge, Jonas Hurst, Erik Scheurer, et al. Lossy neural compression for geospatial analytics: A review. arXiv preprint arXiv:2503.01505, 2025. 4 +[28] Sebastian Hafner, Yifang Ban, and Andrea Nascetti. Unsupervised domain adaptation for global urban extraction using sentinel-1 sar and sentinel-2 msi data. Remote Sensing of Environment, 280:113192, 2022. 3 +[29] Boran Han, Shuai Zhang, Xingjian Shi, and Markus Reichstein. Bridging remote sensors with multisensor geospatial foundation models, 2024. 2 +[30] Soyeon Caren Han, Feiqi Cao, Josiah Poon, and Roberto Navigli. Multimodal large language models and tunings: Vision, language, sensors, audio, and beyond. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11294-11295, 2024. 3 +[31] Jitesh Jain, Jianwei Yang, and Humphrey Shi. Vcoder: Versatile vision encoders for multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27992-28002, 2024. 3 +[32] Johannes Jakubik, Sujit Roy, C. E. Phillips, Paolo Fraccaro, Denys Godwin, Bianca Zadrozny, Daniela Szwarcman, Carlos Gomes, Gabby Nyirjesy, Blair Edwards, Daiki Kimura, Naomi Simumba, Linsong Chu, S. Karthik Mikkavilli, Devyani Lambhate, Kamal Das, Ranjini Bangalore, Dario Oliveira, Michal Muszynski, Kumar Ankur, Muthukumaran Ramasubramanian, Iksha Gurung, Sam Khallaghi, Hanxi, Li, Michael Cecil, Maryam Ahmadi, Fatemeh Kordi, Hamed Alemohammad, Manil Maskey, Raghu Ganti, Kommy Weldemariam, and Rahul Ramachandran. Foundation models for generalist geospatial artificial intelligence, 2023. 2 +[33] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 4 +[34] Samar Khanna, Patrick Liu, Linqi Zhou, Chenlin Meng, Robin Rombach, Marshall Burke, David Lobell, and Stefano Ermon. Diffusionsat: A generative foundation model for satellite imagery, 2023. 3 +[35] Kohei Arai, Michihiro Mikamo, and Shunsuke Onishi. Method for Image Quality Evaluation of Satellite-based SAR Data. International Journal of Advanced Computer Science and Applications (IJACSA), Vol. 14, No. 7, 2023. Online: http://thesai.org/Downloads/Volume14No7/Paper_13-Method_for/Image_Quality_Evaluation_of_Satellite_based_SAR_Data.pdf.16 +[36] Saad Lahrichi, Zion Sheng, Shufan Xia, Kyle Bradbury, and Jordan Malof. Is self-supervised pre-training on satellite imagery better than imagenet? a systematic study with sentinel-2. arXiv preprint arXiv:2502.10669, 2025. 2 +[37] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llavanext: Stronger llms supercharge multimodal capabilities in the wild, 2024. 4, 13 +[38] Jiaxin Li, Danfeng Hong, Lianru Gao, Jing Yao, Ke Zheng, Bing Zhang, and Jocelyn Chanussot. Deep learning in mul + +timodal remote sensing data fusion: A comprehensive review. International Journal of Applied Earth Observation and Geoinformation, 112:102926, 2022. 3 +[39] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 2 +[40] Xiang Li, Congcong Wen, Yuan Hu, Zhenghang Yuan, and Xiao Xiang Zhu. Vision-language models in remote sensing: Current progress and future trends, 2024. 3 +[41] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19325-19337, 2023. 3 +[42] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing, 2024. 2, 3 +[43] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s, 2022. 6 +[44] Gabriel Machado, Edemir Ferreira, Keiller Nogueira, Hugo Oliveira, Matheus Brito, Pedro Henrique Targino Gama, and Jefersson Alex dos Santos. Airround and cv-brct: Novel multiview datasets for scene classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:488-503, 2020. 3 +[45] Gengchen Mai, Chris Cundy, Kristy Choi, Yingjie Hu, Ni Lao, and Stefano Ermon. Towards a foundation model for geospatial artificial intelligence (vision paper). In Proceedings of the 30th International Conference on Advances in Geographic Information Systems, New York, NY, USA, 2022. Association for Computing Machinery. 2 +[46] Oscar Manas, Alexandre Lacoste, Xavier Giró-i Nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2 +[47] Clive Tinashe Marimo, Benedikt Blumenstiel, Maximilian Nitsche, Johannes Jakubik, and Thomas Brunschwiler. Beyond the visible: Multispectral vision-language learning for earth observation. arXiv preprint arXiv:2503.15969, 2025. 2, 4, 13 +[48] Valerio Marsocci and Nicolas Audebert. Cross-sensor self-supervised training and alignment for remote sensing, 2024. 3 +[49] Valerio Marsocci, Yuru Jia, Georges Le Bellier, David Kerekes, Liang Zeng, Sebastian Hafner, Sebastian Gerard, Eric Brune, Ritu Yadav, Ali Shibli, et al. Pangaea: A global and inclusive benchmark for geospatial foundation models. arXiv preprint arXiv:2412.04204, 2024. 2, 8, 18 +[50] Matias Mendieta, Boran Han, Xingjian Shi, Yi Zhu, Chen Chen, and Mu Li. Gfm: Building geospatial foundation models via continual pretraining. arXiv preprint arXiv:2302.04476, 2023. 2 + +[51] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023. 4, 15 +[52] David Mizrahi, Roman Bachmann, Oğuzhan Fatih Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, and Amir Zamir. 4m: Massively multimodal masked modeling, 2023. 4, 5 +[53] Andrea Nascetti, RITU YADAV, Kirill Brodt, Qixun Qu, Hongwei Fan, Yuri Shendryk, Isha Shah, and Christine Chung. Biomasssters: A benchmark dataset for forest biomass estimation using multi-modal satellite time-series. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2 +[54] Vishal Nedungadi, Ankit Kariryaa, Stefan Oehmcke, Serge Belongie, Christian Igel, and Nico Lang. Mmearth: Exploring multi-modal pretext tasks for geospatial representation learning. arXiv preprint arXiv:2405.02771, 2024. 2, 3 +[55] Fernando Paolo, Tsu ting Tim Lin, Ritwik Gupta, Bryce Goodman, Nirav Patel, Daniel Kuster, David Kroodsma, and Jared Dunnmon. xview3-sar: Detecting dark fishing activity using synthetic aperture radar imagery, 2022. 2 +[56] Prabhishek Singh and Raj Shree. Analysis and effects of speckle noise in SAR images. In Proc. International Conference on Advances in Computing, Communication, & Automation (ICACCA), 2016. DOI: 10.1109/ICAC-CAF.2016.7748978. Online: http://ieeexplore.ieee.org/document/7748978.16 +[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PmLR, 2021. 3, 17 +[58] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2 +[59] Ayesha Shafique, Guo Cao, Zia Khan, Muhammad Asad, and Muhammad Aslam. Deep learning-based change detection in remote sensing images: A review. Remote Sensing, 14(4): 871, 2022. 2 +[60] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 17 +[61] Aidan M Swope, Xander H Rudelis, and Kyle T Story. Representation learning for remote sensing: An unsupervised sensor fusion approach. arXiv preprint arXiv:2108.05094, 2021. 3 +[62] Devis Tuia, Konrad Schindler, Begüm Demir, Gustau Camps-Valls, Xiao Xiang Zhu, Mrinalini Kochupillai, Sašo Džeroski, Jan N. van Rijn, Holger H. Hoos, Fabio Del Frate, Mihai Datcu, Jorge-Arnulfo Quiane-Ruiz, Volker Markl, Bertrand Le Saux, and Rochelle Schneider. Artificial intelligence to advance earth observation: a perspective, 2023. 2 +[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4 + +[64] Yi Wang, Conrad M Albrecht, Nassim Ait Ali Braham, Lichao Mou, and Xiao Xiang Zhu. Self-supervised learning in remote sensing: A review. arXiv preprint arXiv:2206.13188, 2022. 2 +[65] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eos12: A large-scale multimodal, multitemporal dataset for self-supervised learning in earth observation [software and data sets]. IEEE Geoscience and Remote Sensing Magazine, 11 (3):98-106, 2023. 3 +[66] Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu, Zhe Chen, Wenhai Wang, Xizhou Zhu, Lewei Lu, Tong Lu, et al. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. Advances in Neural Information Processing Systems, 37:69925-69975, 2025. 3 +[67] Xinyu Bai and Feng Xu. Accelerating Diffusion for SAR-to-Optical Image Translation via Adversarial Consistency Distillation, 2024. Online: http://arxiv.org/pdf/2407.06095.16 +[68] Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J. Stewart, Joëlle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. Neural plasticity-inspired foundation model for observing the earth crossing modalities, 2024. 3 +[69] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 2 +[70] Qidong Yang, Jonathan Giezendanner, Daniel Salles Civitarese, Johannes Jakubik, Eric Schmitt, Anirban Chandra, Jeremy Vila, Detlef Hohl, Chris Hill, Campbell Watson, et al. Multi-modal graph neural networks for localized off-grid weather forecasting. arXiv preprint arXiv:2410.12938, 2024. 2 +[71] Zhiping Yu, Chenyang Liu, Liqin Liu, Zhenwei Shi, and Zhengxia Zou. Metaearth: A generative foundation model for global-scale remote sensing image generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3 +[72] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 2 +[73] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600-612, 2004. 16 +[74] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[75] Linying Zhao and Shunping Ji. Cnn, rn, or vit? an evaluation of different deep learning architectures for spatio-temporal representation of sentinel time series. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16:44-56, 2022. 2 +[76] Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep + +learning in remote sensing: A comprehensive review and list of resources. IEEE geoscience and remote sensing magazine, 5(4):8-36, 2017. 2 \ No newline at end of file diff --git a/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/images.zip b/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cefbf487d26ccde02ca061dcc374c0291f364ca2 --- /dev/null +++ b/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d833eb1fbb90d136c51c9e2fe0a2d643eb39c2144daaa51c61aea0c913fb12e1 +size 762157 diff --git a/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/layout.json b/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a24d862a457f300e6c0d272775c6af50cb2f23f9 --- /dev/null +++ b/2025/TerraMind_ Large-Scale Generative Multimodality for Earth Observation/layout.json @@ -0,0 +1,7824 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 77, + 103, + 533, + 121 + ], + "type": "text", + "content": "TerraMind: Large-Scale Generative Multimodality for Earth Observation" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 61, + 141, + 556, + 201 + ], + "blocks": [ + { + "bbox": [ + 61, + 141, + 556, + 201 + ], + "lines": [ + { + "bbox": [ + 61, + 141, + 556, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 556, + 201 + ], + "type": "image", + "image_path": "6a508ecc15dcc21c6803d3b04d4f20a8dcabf19832964f2fe4ee1a9d82220c68.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "spans": [ + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "text", + "content": "IBM Research - Europe " + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "text", + "content": "ETH Zurich " + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "text", + "content": "Forschungszentrum Jülich " + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "text", + "content": "European Space Agency " + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "text", + "content": "-Lab " + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "text", + "content": "NASA IMPACT " + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 122, + 203, + 495, + 233 + ], + "type": "text", + "content": "University of Iceland" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 241, + 235, + 378, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 235, + 378, + 246 + ], + "spans": [ + { + "bbox": [ + 241, + 235, + 378, + 246 + ], + "type": "text", + "content": "johnannes.jakubikl@ibm.com" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 55, + 279, + 399, + 514 + ], + "blocks": [ + { + "bbox": [ + 55, + 279, + 399, + 514 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 399, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 399, + 514 + ], + "type": "image", + "image_path": "c2719cca28b3707036b1e21c1c8178c88eade472b909cf3ef0639ccd403e1524.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 523, + 555, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 523, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 555, + 559 + ], + "type": "text", + "content": "Figure 1. TerraMind represents the first any-to-any generative, and large-scale multimodal model for Earth observation pre-trained on 500 billion tokens from global geospatial data. The model digests multi-scale representations at pixel-level and token-level simultaneously. TerraMindv1 unlocks (i) generation, (ii) zero-shot and finetuning applications, and (iii) \"Thinking-in-Modalities\" finetuning and inference." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 409, + 279, + 554, + 365 + ], + "blocks": [ + { + "bbox": [ + 409, + 279, + 554, + 365 + ], + "lines": [ + { + "bbox": [ + 409, + 279, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 409, + 279, + 554, + 365 + ], + "type": "image", + "image_path": "b6bb9405779306eee986642cf6d11a4df535fdd5d3976835632bf50bec8e86bd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 409, + 369, + 554, + 513 + ], + "blocks": [ + { + "bbox": [ + 409, + 369, + 554, + 513 + ], + "lines": [ + { + "bbox": [ + 409, + 369, + 554, + 513 + ], + "spans": [ + { + "bbox": [ + 409, + 369, + 554, + 513 + ], + "type": "image", + "image_path": "f7c9a63048871fd77a2eec6e0ea9896413b1fa97b6cae2bce0e8cf481bc98f2a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 567, + 200, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 567, + 200, + 580 + ], + "spans": [ + { + "bbox": [ + 152, + 567, + 200, + 580 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 298, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 298, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 298, + 678 + ], + "type": "text", + "content": "We present TerraMind, the first any-to-any generative, multimodal deep learning model for Earth observation (EO). Unlike other approaches, TerraMind is pretrained on dual-scale representations combining both token-level and pixel-level data across modalities. On a token level, TerraMind encodes high-level contextual information to learn cross-modal relationships, while on a pixel level, TerraMind lever" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 569, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 556, + 700 + ], + "type": "text", + "content": "ages fine-grained representations to capture critical spatial nuances. In this paper, we demonstrate that (i) TerraMind achieves beyond state-of-the-art performance in community-standard benchmarks, (ii) TerraMind can leverage \"thinking in modalities\" (TiM)—the capability of generating additional artificial data during finetuning and inference to improve the model output—and (iii) TerraMind's dual-scale early fusion approach results in well-structured embedding spaces. Models and code have been open-sourced at https://huggingface.co.ibm-esa-geospatialandhttps://github.com.ibm/terramind." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 693, + 125, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 693, + 125, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 125, + 703 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 704, + 122, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 704, + 122, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 704, + 122, + 712 + ], + "type": "text", + "content": "† Equal supervision" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7383" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 95, + 297, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 95, + 297, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 95, + 297, + 228 + ], + "type": "text", + "content": "Earth observation (EO) increasingly benefits from multimodality because of the important integration of complementary information from different data sources. This becomes particularly relevant as EO is spatiotemporally sparse due to low revisiting times or weather phenomena like cloud coverage. Vice versa, for computer vision, EO data is an important playground for the development of new approaches as there is significant publicly available data of very high quality and complexity. The available modalities range from sensors of different satellite missions to relevant complementary information like digital elevation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 231, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 542 + ], + "type": "text", + "content": "In this work, we introduce TerraMind as the first any-to-any generative multimodal model for EO. With TerraMind, we introduce a dual-scale pretraining on pixel-level and token-level and demonstrate benefits over training primarily on tokens. TerraMind encodes high-level contextual information in tokens to enable correlation learning and scaling, while, additionally capturing important fine-grained representations using pixel-level inputs. During pretraining, TerraMind predicts masked target tokens so that our pretraining objective boils down to a cross-modal patch classification problem that results in high-quality latent spaces. TerraMind is pretrained on a custom global-scale geospatial dataset named TerraMesh with nine million samples that have been aligned spatiotemporally and across modalities [7]. In addition to radar and optical satellite images of the Copernicus Sentinel-1 (S-1) and Sentinel-2 (S-2) missions, our dataset contains task-specific modalities such as land use/land cover (LULC) and normalized difference vegetation index (NDVI) maps, metadata like digital elevation models (DEM) and geographic coordinates, and natural language in the form of captions. To the best of our knowledge, TerraMind represents the first truly generative, multimodal deep learning model for EO. Additionally, in contrast to other recent models that utilize masked autoencoders like [54], contrastive learning, or diffusion techniques, TerraMind uniquely demonstrates benefits of leveraging token-based pretraining for EO." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 546, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 546, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 546, + 298, + 714 + ], + "type": "text", + "content": "We provide an overview of TerraMind's performance in a community-standard benchmark [49] in Figure 2 and highlight the any-to-any generative capabilities of TerraMind in Figure 3. Our key contributions are as follows: (i) We introduce a dual-scale approach for generative multimodal pre-training leveraging data on pixel-level and token-level, which outperforms other fusion approaches and enhances embedding space structures. (ii) We introduce thinking in modalities - similar to chain-of-thought approaches in LLMs - for multi-modal models in EO, demonstrating that infusing generated data during finetuning improves the performance. (iii) We demonstrate that TerraMind outperforms other geospatial foundation models both in unimodal and multimodal settings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 402, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 402, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 402, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 426 + ], + "type": "text", + "content": "Computer vision in Earth observation. Computer vision (CV) has significantly advanced EO [76]. Many CV techniques, originally developed for natural image processing, have been adapted to EO [62], often with minimal modifications. A wide range of tasks benefit from these methods, including classification [16], semantic segmentation [72] (e.g., land cover mapping [20, 21]), change detection [59] (e.g., disaster response [19]), object detection [39] (e.g., vessel identification [55]), and regression (e.g., biomass estimation [53]). Deep learning architectures like CNNs [75] and Vision Transformers (ViTs) [17] have demonstrated strong performance, often surpassing traditional remote sensing (RS) methods. However, EO presents unique challenges, including diverse sensor modalities [4] and geospatial heterogeneity [46]. An emerging paradigm in EO is self-supervised learning (SSL) [64] and geospatial foundation models (GFMs) [45], which aim to leverage vast amounts of unlabeled RS data to develop general purpose task models [32]. While off-the-shelf CV models have shown promising results [36], they do not fully exploit the unique characteristics of geospatial data. Many GFMs still rely on generic CV architectures [50], which were not explicitly designed to handle the complexities of EO, such as heterogeneous sensor sources (e.g., optical, radar, DEM) [29], integrated with auxiliary data (e.g., text) [42, 47], and expert knowledge (e.g., prioritizing specific bands or indexes). In this direction, TerraMind better integrates domain-specific properties, developing a customized and expandable multimodal learning strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 427, + 556, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 556, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 556, + 477 + ], + "type": "text", + "content": "Multimodality in CV. Multimodal CV is driven by the integration of diverse data streams [69], such as natural images [74], natural language text [10], temporal video data [58], and weather [70], within large foundation models [8]." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 495, + 551, + 668 + ], + "blocks": [ + { + "bbox": [ + 316, + 495, + 551, + 668 + ], + "lines": [ + { + "bbox": [ + 316, + 495, + 551, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 495, + 551, + 668 + ], + "type": "image", + "image_path": "89ab3f4a24a8f0f22c145b98d7aa131473829751dc280f48d7fdfe81a559cdff.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 677, + 557, + 711 + ], + "lines": [ + { + "bbox": [ + 313, + 677, + 557, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 557, + 711 + ], + "type": "text", + "content": "Figure 2. TerraMind outperforms other geospatial foundation models on PANGAEA benchmark [49] in finetuning. Performance is measured in mIoU and min-max scaled per dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7384" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 556, + 214 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 556, + 214 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 556, + 214 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 556, + 214 + ], + "type": "image", + "image_path": "26ad38ec634ed36b3b55e5dbf6b8398dd71b67db16bdc57353ddbf06d8c0c07f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "type": "text", + "content": "Figure 3. Chained generation example of TerraMindv1-B starting from either optical, radar, or digital elevation data. Left is input, middle is artificially generated data by TerraMind, right represents ground truths and tokenizer reconstructions, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 264, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 264, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 264, + 296, + 407 + ], + "type": "text", + "content": "Starting from the alignment of images and texts [57], these models moved beyond simple feature extraction, towards nuanced contextual understanding. The ability to combine several modalities allows for unprecedented capabilities in complex tasks [30], evidenced by the rapid advancement of multimodal Large Language Models (MLLMs) [30], that excel in tasks such as scene understanding [12], visual question answering [18], and video analysis [24]. Recent advances in architectures [31] and large scale pre-training [11] have enabled the development of models that learn highly effective cross-modal representations [41], which can then be adapted to a wide variety of downstream tasks [66]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 408, + 296, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 659 + ], + "type": "text", + "content": "Multimodality in EO. Multimodality in EO originates from data fusion and is typically understood as the integration of SAR and optical data [13, 25, 28, 38] or the combination of optical data with vector data [5]. Some studies have explored alternative combinations of data. In [15], the authors introduce a contrastive framework for comparing RS images and street views. Even different optical sensors can be considered different modalities [48, 61]. Similarly, several multi-view images (i.e. multimodal) datasets [26, 44, 54] are introduced. More recent approaches combined text and images [40], both for discriminative [42] and generative [34] purposes. Lately, different GFMs are trained in a multimodal way [4, 54, 68], still focusing either on a specific set of modalities (e.g., vision [68], [3]) or tasks (e.g., generative [34]). Compared to multi-scale high-quality generation models for optical data, like MetaEarth [71], our approach allows to generate any modality from any other pretraining modality. To the best of our knowledge, no existing model has combined a wide and diverse amount of modalities both for discriminative and generative purposes, as TerraMind does. We provide a comparison in Table 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 669, + 111, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 111, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 111, + 681 + ], + "type": "text", + "content": "3. Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "For the pretraining of TerraMind and its tokenizers, we create a multimodal dataset called TerraMesh [7], which will" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 261, + 555, + 475 + ], + "blocks": [ + { + "bbox": [ + 315, + 261, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 315, + 261, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 315, + 261, + 555, + 475 + ], + "type": "table", + "html": "
ModelModalitiesAny-to-Any GenerationMulti-Scale Features
RemoteCLIPoptical, textXX
CROMAoptical, radarXX
AnySataerial, optical, radar, NAIPXX
DeCURoptical, radarXX
DOFAoptical, radar, hyperspectral, NAIPXX
MetaEarthoptical (unimodal)X✓
Galileooptical, radar, elevation, weather, location, population, ...X✓
TerraMindoptical, radar, land use, elevation, vegetation index, location, text✓✓
", + "image_path": "e99accac7b0552322e0145e2b418e3c374d12d577dbd6d217a88426840df1874.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 483, + 555, + 507 + ], + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 507 + ], + "type": "text", + "content": "Table 1. Comparison of TerraMind to other model architectures. TerraMind represents a first-of-its-kind generative, multimodal model." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 532, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 592 + ], + "type": "text", + "content": "be open-sourced to the community. TerraMesh builds on existing datasets, which we expand by adding modalities from external data sources or by applying pseudo-labeling. We provide an overview of the aligned image modalities and a detailed dataset description in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Base datasets. TerraMesh is based on SSL4EO-S12 [6, 65] and MajorTOM-Core [23], two unlabeled remote sensing datasets containing co-aligned radar and optical imagery from Sentinel-1 and Sentinel-2 satellites. SSL4EO-S12 has lower geographic coverage but is multi-seasonal. MajorTOM-Core covers most of the Earth's land surface at a single timestamp. For MajorTOM-Core, we apply a subsampling scheme based on LULC classes and ecoregions. TerraMesh includes a total of approximately 9 million globally distributed samples from both Sentinel-1 and Sentinel-2," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7385" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 262, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 262, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 262, + 83 + ], + "type": "text", + "content": "each measuring " + }, + { + "bbox": [ + 55, + 72, + 262, + 83 + ], + "type": "inline_equation", + "content": "264 \\times 264" + }, + { + "bbox": [ + 55, + 72, + 262, + 83 + ], + "type": "text", + "content": " pixels at " + }, + { + "bbox": [ + 55, + 72, + 262, + 83 + ], + "type": "inline_equation", + "content": "10\\mathrm{m}" + }, + { + "bbox": [ + 55, + 72, + 262, + 83 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 297, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 297, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 297, + 277 + ], + "type": "text", + "content": "Additional modalities. We obtain co-aligned yearly LULC maps by ESRI with nine land use classes. Additionally, we leverage SEnSeI v2 [22] as a cloud and ice annotation model and update the ESRI LULC classes for better spatiotemporal alignment. NDVI maps are computed using the corresponding spectral bands from Sentinel-2. DEM is extracted from the Copernicus DEM 30m dataset [2], which provides global coverage of the Earth's elevation at a 30m resolution. Captions are generated synthetically by constructing RGB images from Sentinel-2 patches using the corresponding spectral bands and processing them with LLaVANext [37]. A tailored prompt guides the model to describe the content of each image as described in [47]. For geolocations, we round latitude and longitude from the center of each patch to the nearest quarter degree and store the discretized coordinates as strings in a pre-defined format." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 288, + 117, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 117, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 117, + 300 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 296, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 369 + ], + "type": "text", + "content": "TerraMind pretraining is two-staged following [52]. We first pretrain unimodal tokenizer models, tokenize the modalities, and then leverage token-level and pixel-level input to pretrain the TerraMind encoder-decoder architecture. We describe those individual stages in the following." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 139, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 139, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 139, + 388 + ], + "type": "text", + "content": "4.1. Tokenization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 394, + 296, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 296, + 478 + ], + "type": "text", + "content": "We develop modality-specific tokenizers to encode each modality into a sequence of discrete tokens for pretraining and decode token sequences back to images. Thus, TerraMind is in principle compatible with any modality, as long as it can be tokenized and aligned with other modalities. For reasons of space, we delegate most experiments related to the tokenizer performances to the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 479, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 694 + ], + "type": "text", + "content": "Image-like modalities. We train autoencoder-based architectures with a quantization step in the bottleneck for image-like modalities such as S-1, S-2, LULC, NDVI, and DEM. Tokenizer encoders process an input image and generate a latent representation for each " + }, + { + "bbox": [ + 55, + 479, + 296, + 694 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 479, + 296, + 694 + ], + "type": "text", + "content": " patch, which is then discretized with finite-scalar-quantization (FSQ) [51] into one of " + }, + { + "bbox": [ + 55, + 479, + 296, + 694 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 479, + 296, + 694 + ], + "type": "text", + "content": " codewords. All tokenizers use a vocabulary size of 16K besides the simpler LULC modality for which we use 4K. These codewords are then used by the diffusion decoder to reconstruct the original image. The benefit of leveraging diffusion decoders lies in facilitating cross-modal generation in TerraMind by transforming tokens back into images. By mapping each codeword to a unique integer in " + }, + { + "bbox": [ + 55, + 479, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\{0, 1, \\dots, N - 1\\}" + }, + { + "bbox": [ + 55, + 479, + 296, + 694 + ], + "type": "text", + "content": ", we obtain discrete tokens for each image patch. We pretrain the tokenizers in a self-supervised setting. FSQ as quantization method enhances training stability [51] compared to vector quantization [63] by eliminating the need for codebook-related loss terms. Notably, FSQ is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "content": "heavily influenced by ideas of neural compression [27]. For example, on 12-band S-2 images, we achieve compression rates of over " + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "inline_equation", + "content": "3000\\mathrm{x}" + }, + { + "bbox": [ + 313, + 72, + 556, + 240 + ], + "type": "text", + "content": " by applying quantization. We summarize the architecture of our tokenizers in Figure 4. The main objective of the overall tokenizer is to encode image patches consistently into discrete tokens based on semantic similarity to enable cross-modal correlation learning. Therefore, the loss of some details is an expected trade-off, since the focus is on grouping similar patches rather than preserving all fine-grained features. Naturally, more accurate reconstructions facilitate cross-modal generation, however the main focus of the pretraining lies on consistent cross-modal correlation learning. We provided further details on the pretraining of the tokenizers in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 252, + 555, + 330 + ], + "blocks": [ + { + "bbox": [ + 318, + 252, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 318, + 252, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 318, + 252, + 555, + 330 + ], + "type": "image", + "image_path": "71da8b1dfb813f0f4d5e280206571aa38f59c2eb458f021352749f9465058d54.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 342, + 556, + 365 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 365 + ], + "type": "text", + "content": "Figure 4. Tokenizer for image-like modalities combining finite-scalar quantization [51] with diffusion decoding." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 377, + 556, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 556, + 475 + ], + "type": "text", + "content": "Sequence-like modalities. We treat both captions and geolocations as text and use a single text tokenizer to process both modalities. By discretizing the geographic coordinates and representing them as strings, we introduce special coordinate tokens into the vocabulary. This allows us to encode geolocations as a sequence of discrete tokens, beginning with a latitude token followed by a longitude token. For textual data, we modify the existing WordPiece tokenizer [33]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 480, + 395, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 395, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 395, + 493 + ], + "type": "text", + "content": "4.2. Pre-training" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "text", + "content": "Architecture. TerraMind uses a symmetric Transformer-based encoder-decoder architecture proposed in [52], which is designed to process sequences of multimodal tokens. In addition to discrete tokens, TerraMind accepts pixel-level inputs, specifically satellite imagery and digital elevation maps. For pixel-level inputs, we apply learnable patch-wise linear projections to generate patch embeddings for each " + }, + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 498, + 556, + 594 + ], + "type": "text", + "content": " patch, similar to the approach used in ViT [17]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 713 + ], + "type": "text", + "content": "Dual-scale early fusion. In contrast to [52], we not only embed token-level data but additionally leverage pixel-level data across a range of input modalities to introduce a dual-scale feature representation to support the structuring of the embedding space. Both tokens and patches represent a 16x16 pixel area. Tokens represent this area via a single discrete integer value, while the image patches describe the same area with the actual floating point sensor data. Thus, during pretraining, the model not only learns a correlation between modalities (i.e., cross-modal learning) but also between dif" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 56, + 704, + 259, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 704, + 259, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 704, + 259, + 712 + ], + "type": "text", + "content": "https://planetarycomputer.microsoft.com/dataset/io-lulc-annual-v02" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7386" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "ferent levels of abstraction within the same modality. The low-level token information enables cross-modal correlation learning, while adding pixel level input accounts for spatial nuances. Based on dual-scale features the model further learns to better structure pixel-level data in the embedding space via the corresponding information from the discrete token. We illustrate the pretraining paradigm in Figure 5. The model is agnostic to processing tokens or patches in the input space, while the target is generally token-level data. We use six pixel-level modalities and eight token-level modalities." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 203, + 289, + 290 + ], + "blocks": [ + { + "bbox": [ + 56, + 203, + 289, + 290 + ], + "lines": [ + { + "bbox": [ + 56, + 203, + 289, + 290 + ], + "spans": [ + { + "bbox": [ + 56, + 203, + 289, + 290 + ], + "type": "image", + "image_path": "e0986ca258c6aa8b5d527085e8ce35c951950eed3bc2982fada04e93f0661922.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 303, + 296, + 336 + ], + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 336 + ], + "type": "text", + "content": "Figure 5. Illustration of the pre-training task. Given an encoded multimodal sample of random subsets of patches and input tokens, the decoder predicts target tokens for the masked input." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 347, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 418 + ], + "type": "text", + "content": "Masking strategy. TerraMind applies a masked modeling approach in the token space following [52]. The model leverages a set of randomly selected target tokens that have to be reconstructed from a randomly selected set of input tokens and pixel-level data. During pre-training, we sample input and target data from a Dirichlet distribution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 419, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 598 + ], + "type": "text", + "content": "We opt for masked token reconstruction to familiarize the model with the absence of entire modalities, which is crucial for a high usability of a multimodal model in Earth observation. During pre-training, the model learns an internal representation of unseen modalities which is expected to benefit a range of downstream applications. In addition, sampling input and target tokens improves the computational efficiency of the pre-training, as each token is a compressed representation of a patch with compression factors of between 250x and 3000x depending on the modality. Finally, without tokenized representations of the image-like modalities, it is challenging to learn the correlation to sequence-like modalities. The overall training objective of TerraMind boils down to a cross-modal patch-level classification problem optimized via a cross entropy loss:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 605, + 296, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 605, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 125, + 605, + 296, + 637 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} = - \\sum_ {i = 1} ^ {N} y _ {i} \\log \\left(p _ {i}\\right), \\tag {1}", + "image_path": "5c964073da44bbe5bd363bfad93b2c657fb846153d402f0c333cc6ceb4c0c0de.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " is the one-hot encoded true class of token " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " is the predicted probability for token " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": " is the total number of possible tokens. Interestingly, we can infer an upper bound loss for a random model where the cross entropy loss will collapse to the natural logarithm of the vocabulary size " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE,random}} = -\\sum_{i=1}^{N} y_{i} \\log \\left( \\frac{1}{N} \\right) = \\log(N)" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 348 + ], + "type": "text", + "content": "Scaling. We trained three versions of TerraMind scaling across model size, compute, and data. In addition, we pretrain different versions of TerraMind with respect to the number of dual-scale features. TerraMindv1-B is pre-trained on 500B tokens for 6 days on 32 NVIDIA A100 GPUs. The model uses dual-scale features from both token-level and pixel-level. During initial experiments, we observed significant improvements from scaling model size when switching from a tiny backbone to a small backbone to a base backbone. Therefore, we pre-trained TerraMindv1-L on a large backbone with 500B tokens on 32 NVIDIA A100 GPUs trained for 10 days. Finally, to better understand the effect of scaling across the dual-scale feature representation, we pre-train TerraMindv1-B-single as a single-scale model on primarily token-level data with optical S-2 L2A data as only pixel-level input (compared to pixel-level S-2 L1C, S-2 RGB, S-1 GRD, S-1 RTC, and DEM in TerraMindv1-B and -L). TerraMindv1-B-single is pretrained on 500B tokens from over one million samples for 6 days on 32 NVIDIA A100 GPUs. We summarize the scaling behavior in model size, compute, and data in Figure 9 of the supplementary material. We additionally provide final validation losses in Table 9 comparing v1-B and v1-L with the theoretical random loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 357, + 390, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 390, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 390, + 369 + ], + "type": "text", + "content": "4.3. Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 374, + 556, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 556, + 578 + ], + "type": "text", + "content": "Once pretrained, TerraMind can generate tokens for any modality, conditioned on any subset of input modalities. The generative capabilities unlock various zero-shot tasks, such as water body segmentation. For the generation of image-like modalities, the decoder receives mask tokens for the modality to be generated and predicts the corresponding tokens based on the encoded input. For sequence-like modalities, the decoder generates the output autoregressively. After generating tokens from the target modality, the corresponding tokenizer decoder allows to map from token-space to image or text space. TerraMind further supports chained generation which ensures consistency across generated modalities. The chained generation represents a conditional probability distribution where the prior probability distribution is determined by the input modality, and all subsequent modalities are generated conditioned on the input modality and potentially other generated modalities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 599, + 446, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 446, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 446, + 612 + ], + "type": "text", + "content": "4.4. Thinking-in-Modalities" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Thinking in Modalities (TiM) is a recursive fine-tuning and inference technique designed to enhance multimodal learning by leveraging the generative capabilities of the model itself. Given an input " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " (e.g., an optical satellite image), the model first generates additional synthetic modalities " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{x} = f_{\\mathrm{gen}}(x)" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " on a token-level using a learned generative function " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{gen}}" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": ". These generated tokens are then concatenated with the original input and jointly processed by the downstream" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7387" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "model " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " (e.g., TerraMind encoder with a segmentation head), yielding the final output " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "y = f(x, f_{\\mathrm{gen}}(x))" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". This formulation allows the model to reason over both observed and inferred modalities, effectively enriching the input space. TiM can leverage multiple generated modalities which are then generated in a chained approach. For example, for " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " modalities, the input is augmented with newly generated modalities:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 163, + 296, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 163, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 118, + 163, + 296, + 178 + ], + "type": "interline_equation", + "content": "\\tilde {x} ^ {(k + 1)} = \\tilde {x} ^ {(k)} \\cup f _ {\\text {g e n}} (\\tilde {x} ^ {(k)}), \\tag {2}", + "image_path": "60b6eb23b3dd22e279e951a2d476553e023a1c300c717e47891d033a65012a8c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 228, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 228, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 228, + 198 + ], + "type": "text", + "content": "and the final model output is described by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 205, + 296, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 205, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 146, + 205, + 296, + 220 + ], + "type": "interline_equation", + "content": "y = f \\left(\\tilde {x} ^ {(K)}\\right). \\tag {3}", + "image_path": "0ccb2554b06baa9d62c6a74f70c7943a8df9804907704db9b3a3219bda8f6be8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 227, + 297, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 297, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 297, + 264 + ], + "type": "text", + "content": "This recursive augmentation mimics a chain-of-thought process, enabling the model to iteratively refine its internal representation, particularly in scenarios with missing modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 276, + 137, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 137, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 137, + 289 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 296, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 296, + 333 + ], + "type": "text", + "content": "In this section, we describe the performance gains resulting from TerraMind and experiment with the unlocked capabilities of any-to-any generation and Thinking-in-Modalities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 339, + 201, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 201, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 201, + 352 + ], + "type": "text", + "content": "5.1. Foundational experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 356, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 487 + ], + "type": "text", + "content": "Multimodality vs. unimodality. As a first motivational experiment, we outline the benefit of using multimodal data in Earth observation at the example of water body mapping. Specifically, we leverage the ViT-B encoders from the unimodal tokenizer models for S-1, S-2, and LULC, concatenate their embeddings, and train a segmentation head with four ConvNeXt [43] blocks as a late fusion approach. The results in Table 2 (left) suggest that regardless of which modalities we combine, the combination of two modalities always outperforms each unimodal model. Combining all three modalities achieves the best overall performance." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 78, + 497, + 273, + 591 + ], + "blocks": [ + { + "bbox": [ + 78, + 497, + 273, + 591 + ], + "lines": [ + { + "bbox": [ + 78, + 497, + 273, + 591 + ], + "spans": [ + { + "bbox": [ + 78, + 497, + 273, + 591 + ], + "type": "table", + "html": "
InputLate fusionToken-level fusion
S-161.0163.94 (2.93pp↑)
S-272.7076.32 (3.62pp↑)
LULC71.7770.96 (0.81pp↓)
S-1 + S-273.8376.74 (2.91pp↑)
S-1 + LULC73.8673.76 (0.10pp↓)
S-2 + LULC75.6577.04 (1.39pp↑)
S-1 + S-2 + LULC76.0076.88 (0.88pp↑)
", + "image_path": "4a94b084cb60ebef806fd521fdeb71037ea6eccbcf9986eac6b25c653d4d8bbd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Token-level fusion vs. late fusion. In Table 2 (right), we investigate the effects of fusing the inputs on a token level" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "content": "through masked token reconstruction. We observe that token-level fusion outperforms late fusion. The performance gains are particularly high when LULC data is not available. This suggests that early fusion captures an internal representation of the multimodal state—especially pronounced for LULC—that benefits fine-tuning. With those findings in mind, we will explore the effects of using additional multi-modal pixel-level input in a dual-scale pretraining in Section 5.5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 186, + 450, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 450, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 450, + 199 + ], + "type": "text", + "content": "5.2. Generation experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 207, + 556, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 556, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 556, + 507 + ], + "type": "text", + "content": "TerraMind supports any-to-any generation. In the following, we provide examples of the generation performance starting from: (i) an information-rich modality, like optical S-2 L2A data, and (ii) minimal information based on the geolocation. In Figure 3, we observe that TerraMind performs strongly in generating image-like modalities like S-1, LULC, and DEM from optical S-2 L2A data. We provide a quantitative overview on the quality of the generations on unseen validation data in Table 3. Overall, we observe an interesting asymmetry in the generative performance of TerraMind where (a) radar-to-optical generation achieves reasonable quality in terms of SSIM and PSNR – indicating structural and visual fidelity with some perceptual degradation – and (b) optical-to-radar generation yields higher PSNR values but lower SSIM, suggesting visually plausible outputs that lack strong structural alignment. The quality of generated DEM suggests to be structurally very strong, but noisy. The errors for DEM generations suggest that the level of altitude is difficult to infer for the model. We compare these scores with the reconstruction quality of the auto-encoding tokenizers in the supplementary material that can serve as upper bounds. Additionally, we provide experiments on the generation quality using token-level instead of pixel-level inputs. Finally, we demonstrate the quality of generations at kilometer scale in Figures 19 and 20." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 318, + 523, + 552, + 642 + ], + "blocks": [ + { + "bbox": [ + 55, + 600, + 296, + 677 + ], + "lines": [ + { + "bbox": [ + 55, + 600, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 296, + 677 + ], + "type": "text", + "content": "Table 2. Water body mapping on Sen1Floods11 [9] measured in IoU on water class. Model sizes and architectures are comparable. Left column: Late fusion of tokenizers. The average improvement of full multimodality over the individual unimodal performance is 7.5pp IoU. Right column: Finetuning results of TerraMindv1-B-single as a mid fusion approach based on masked correlation learning. Gains over late fusion in percentage points in parentheses." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 523, + 552, + 642 + ], + "lines": [ + { + "bbox": [ + 318, + 523, + 552, + 642 + ], + "spans": [ + { + "bbox": [ + 318, + 523, + 552, + 642 + ], + "type": "table", + "html": "
ModalitiesMAE↓RMSE↓SSIM↑PSNR↑
S-1 GRD → S-2 L2A0.0740.1160.75026.210
S-1 GRD → DEM163.0320.80.87820.694
S-1 GRD → NDVI0.1800.2250.43818.990
S-1 RTC → S-2 L2A0.1130.1940.69524.251
S-1 RTC → DEM298.8799.20.87320.009
S-1 RTC → NDVI0.1720.2110.46519.529
S-2 L2A → S-1 GRD2.9423.8770.53128.678
S-2 L2A → S-1 RTC2.6363.3910.43028.993
S-2 L2A → DEM215.8745.50.94220.616
", + "image_path": "8d44d87fb7279771d4eb4bbf789b50b91b744118f08d738966d3b9c0de1324f7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 650, + 556, + 684 + ], + "lines": [ + { + "bbox": [ + 313, + 650, + 556, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 556, + 684 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation of generations on unseen global validation dataset using 10 diffusion steps. MAE and RMSE metrics are in physical units: meter (DEM), reflectance (S-2), and db (S-1)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7388" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 171, + 131 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 171, + 131 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 171, + 131 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 171, + 131 + ], + "type": "image", + "image_path": "ecd65734e16cd2da1375f189a89942b655bab4f6d4ea51b928ebdd4441055148.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 134, + 172, + 153 + ], + "lines": [ + { + "bbox": [ + 56, + 134, + 172, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 134, + 172, + 153 + ], + "type": "text", + "content": "(a) Input: S-2 L2A data capturing Singapore in January 2025." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 72, + 292, + 131 + ], + "blocks": [ + { + "bbox": [ + 181, + 72, + 292, + 131 + ], + "lines": [ + { + "bbox": [ + 181, + 72, + 292, + 131 + ], + "spans": [ + { + "bbox": [ + 181, + 72, + 292, + 131 + ], + "type": "image", + "image_path": "93cf79200582ee94491ba9ee7aa21762717149e729068fe33c90012c0c793f3d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 134, + 295, + 153 + ], + "lines": [ + { + "bbox": [ + 179, + 134, + 295, + 153 + ], + "spans": [ + { + "bbox": [ + 179, + 134, + 295, + 153 + ], + "type": "text", + "content": "(b) Generation: S-1 RTC composition generated by TerraMind." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 165, + 171, + 234 + ], + "blocks": [ + { + "bbox": [ + 58, + 165, + 171, + 234 + ], + "lines": [ + { + "bbox": [ + 58, + 165, + 171, + 234 + ], + "spans": [ + { + "bbox": [ + 58, + 165, + 171, + 234 + ], + "type": "image", + "image_path": "a9582e25b72bba7331bcf2355f6ee19a7a7832cdd841268bd8932a766d10c44f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 237, + 172, + 256 + ], + "lines": [ + { + "bbox": [ + 56, + 237, + 172, + 256 + ], + "spans": [ + { + "bbox": [ + 56, + 237, + 172, + 256 + ], + "type": "text", + "content": "(c) Input: S-2 L2A data capturing Northern Spain in January 2025." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 181, + 166, + 292, + 233 + ], + "blocks": [ + { + "bbox": [ + 181, + 166, + 292, + 233 + ], + "lines": [ + { + "bbox": [ + 181, + 166, + 292, + 233 + ], + "spans": [ + { + "bbox": [ + 181, + 166, + 292, + 233 + ], + "type": "image", + "image_path": "2339bc758ec8ffc0eb95784cdb7350fbdf51c620f199fe7eeeba6e7f859f7db4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 237, + 295, + 256 + ], + "lines": [ + { + "bbox": [ + 179, + 237, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 179, + 237, + 295, + 256 + ], + "type": "text", + "content": "(d) Generation: S-1 GRD composition generated by TerraMind." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 266, + 295, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 289 + ], + "type": "text", + "content": "Figure 6. Generated S-1 imagery using TerraMind. We provide large-scale visualizations in the supplementary material." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 308, + 184, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 184, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 184, + 320 + ], + "type": "text", + "content": "5.3. Zero-shot experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "type": "text", + "content": "Based on its generative capabilities, TerraMind unlocks several zero-shot applications, like land-use segmentation, water body mapping, geo-localization, and vegetation mapping. In the following, we focus on water body mapping and geo-localization as image- and sequence-level zero-shot tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 386, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 296, + 529 + ], + "type": "text", + "content": "Water body mapping. In Table 4, we compare the zero-shot performance of TerraMind with its fine-tuned performance and other finetuned benchmarks for water body mapping. Overall, TerraMindv1-B achieves a zero-shot IoU of " + }, + { + "bbox": [ + 55, + 386, + 296, + 529 + ], + "type": "inline_equation", + "content": "45.4\\%" + }, + { + "bbox": [ + 55, + 386, + 296, + 529 + ], + "type": "text", + "content": " compared to SOTA-level fine-tuning performance of " + }, + { + "bbox": [ + 55, + 386, + 296, + 529 + ], + "type": "inline_equation", + "content": "82.2\\%" + }, + { + "bbox": [ + 55, + 386, + 296, + 529 + ], + "type": "text", + "content": " of DeCUR. In ablations with TerraMindv1-B-single trained on DynamicWorld LULC data, we boost this to up to " + }, + { + "bbox": [ + 55, + 386, + 296, + 529 + ], + "type": "inline_equation", + "content": "69.8\\%" + }, + { + "bbox": [ + 55, + 386, + 296, + 529 + ], + "type": "text", + "content": " suggesting that TerraMind harnesses up to over " + }, + { + "bbox": [ + 55, + 386, + 296, + 529 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 386, + 296, + 529 + ], + "type": "text", + "content": " of the SOTA performance in zero-shot setting. Additionally, it's notable that none of the benchmarking model can be applied in a zero-shot context, highlighting the relevance of TerraMind's capabilities." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 80, + 538, + 272, + 634 + ], + "blocks": [ + { + "bbox": [ + 80, + 538, + 272, + 634 + ], + "lines": [ + { + "bbox": [ + 80, + 538, + 272, + 634 + ], + "spans": [ + { + "bbox": [ + 80, + 538, + 272, + 634 + ], + "type": "table", + "html": "
ModelInputTypeIoUWater
TerraMindv1-BS-2zero-shot45.40
TerraMindv1-B-singleS-2zero-shot69.75
Prithvi 2.0 / DeCUR / ...zero-shotN/A
Baseline [9]S-2finetune31.25
Prithvi 2.0 300MS-2finetune80.97
DeCURS-2finetune82.17
", + "image_path": "5e41b8de076f12a79156ced1ee4defed1e0191d11046bc88d737cdc73d98364e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Geo-localization. TerraMind is able to predict the geolocation of a specific data instance. To better visualize the geolocation capabilities, we prompt the model for the most" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": "likely locations of the land use class \"bare land\" (deserts etc.) in a Monte-Carlo-sampling in Figure 7. The probability distribution of the model fits the expectation of where to find bare land, highlighting the Sahara region and middle-east, as well as Mexico and Southern California." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 345, + 142, + 523, + 230 + ], + "blocks": [ + { + "bbox": [ + 345, + 142, + 523, + 230 + ], + "lines": [ + { + "bbox": [ + 345, + 142, + 523, + 230 + ], + "spans": [ + { + "bbox": [ + 345, + 142, + 523, + 230 + ], + "type": "image", + "image_path": "442bc09a3bff1182ff8bb275d44d364622dcffd33c27729448ba71c7b3172b1d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 241, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 285 + ], + "type": "text", + "content": "Figure 7. Prediction distribution of the land use class \"bare land\" with a sampling temperature of " + }, + { + "bbox": [ + 313, + 241, + 555, + 285 + ], + "type": "inline_equation", + "content": "T = 1.0" + }, + { + "bbox": [ + 313, + 241, + 555, + 285 + ], + "type": "text", + "content": " using TerraMindv1-B-single. TerraMind has an accurate internal representation of the geolocation of specific contexts, like land use classes." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 308, + 440, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 440, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 440, + 320 + ], + "type": "text", + "content": "5.4. Few-shot experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 326, + 555, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 530 + ], + "type": "text", + "content": "TerraMind is trained via a cross-modal patch classification objective. Thus, we expect a well-structured latent space that clusters different concepts accurately. To investigate our hypothesis, we apply 1-Nearest-Neighbor (1-NN) classification experiments in the community-standard setting of 1-shot 5-way on two datasets: EuroSAT and METER-ML. In those experiments, there are no weight updates of any kind, so that we can assess the quality of the embedding space structure. In Table 5, we observe that TerraMind outperforms several other benchmarks from both the CV and EO domain on the EuroSAT dataset by at least 10pp in accuracy. Our results further show that for methane source classification on METER-ML, TerraMind outperforms benchmark models and generalizes to high-resolution NAIP data with one order of magnitude higher resolution than the pre-training data. We present additional experiments with other few-shot settings in the supplementary material." + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 321, + 540, + 547, + 645 + ], + "blocks": [ + { + "bbox": [ + 55, + 642, + 295, + 665 + ], + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 665 + ], + "type": "text", + "content": "Table 4. Zero-shot results of TerraMind on water body mapping compared to fine-tuned performance of benchmarks." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 540, + 547, + 645 + ], + "lines": [ + { + "bbox": [ + 321, + 540, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 321, + 540, + 547, + 645 + ], + "type": "table", + "html": "
ModelInputEuroSATMETER-ML
CLIP-ViT-B/16S-2 RGB57.0029.15
CLIP-ViT-B/16NAIP-32.01
DeCURS-2 L1C50.5427.87
Prithvi 1.0 100MS-2 L1C60.1126.08
Prithvi 2.0 300MS-2 L1C61.0628.26
TerraMindv1-BS-2 L1C70.8333.90
TerraMindv1-BNAIP-32.23
", + "image_path": "d53fc5b399a97b4b6fdef0fd3e89d1a0c2c282ff6812aac36344421d2cf4f603.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 654, + 555, + 698 + ], + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 698 + ], + "type": "text", + "content": "Table 5. 1-shot 5-way classification results on EuroSAT and METER-ML measured in mean accuracy " + }, + { + "bbox": [ + 313, + 654, + 555, + 698 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 313, + 654, + 555, + 698 + ], + "type": "text", + "content": ", averaged over 200 runs. TerraMind outperforms benchmarks from CV and EO domain, suggesting a well-structured latent space." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7389" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 259 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 259 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 259 + ], + "type": "table", + "html": "
ModelBurnSr*MADOS*PASTISSen1Fl11FBP*DEN*CTM-SSSN7*AI4Farms*Avg. mIoUAvg. Rank
CROMA82.4267.5532.3290.8951.8338.2949.3859.2825.6555.296.61
DOFA80.6359.5830.0289.3743.1839.2951.3361.8427.0753.598.22
GFM-Swin76.9064.7121.2472.6067.1834.0946.9860.8927.1952.4210.00
Prithvi 1.0 100M83.6249.9833.9390.3746.8127.8643.0756.5426.8651.0011.00
RemoteCLIP76.5960.0018.2374.2669.1931.7852.0557.7625.1251.6611.22
SatlasNet79.9655.8617.5190.3050.9736.3146.9761.8825.1351.6510.67
Scale-MAE76.6857.3224.5574.1367.1935.1125.4262.9621.4749.4311.44
SpectralGPT80.4757.9935.4489.0733.4237.8546.9558.8626.7551.8710.11
S.-S12-MoCo81.5851.7634.4989.2653.0235.4448.5857.6425.3853.0210.06
S.-S12-DINO81.7249.3736.1888.6151.1534.8148.6656.4725.6252.5110.89
S.-S12-MAE81.9149.9032.0387.7951.9234.0845.8057.1324.6951.6912.39
S.-S12-Data2Vec81.9144.3634.3288.1548.8235.9054.0358.2324.2352.2210.72
UNet Baseline84.5154.7931.6091.4260.4739.4647.5762.0946.3457.584.89
ViT Baseline81.5848.1938.5387.6659.3236.8344.0852.5738.3754.1310.28
TerraMindv1-B82.4269.5240.5190.6259.7237.8755.8060.6128.1258.353.94
TerraMindv1-L82.9375.5743.1390.7863.3837.8955.0459.9827.4759.573.44
", + "image_path": "4412d721907d41d67686f46428339dc69d078241be78a050b87e7eaf63439ffb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 266, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 555, + 301 + ], + "type": "text", + "content": "Table 6. Performance evaluation of TerraMind using the PANGAEA evaluation protocol indicates higher mIoU values (↑) and lower rank values (↓). The best model per column is highlighted in bold, the second best is underscored. We indicate unimodal datasets with *. Encoders are frozen for pretrained models, while U-Net and ViT baselines are trained from scratch for each specific task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 320, + 194, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 194, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 194, + 334 + ], + "type": "text", + "content": "5.5. Fine-tuning experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 339, + 297, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 339, + 297, + 592 + ], + "spans": [ + { + "bbox": [ + 54, + 339, + 297, + 592 + ], + "type": "text", + "content": "Besides the novel capabilities that TerraMind introduces, we benchmark the fine-tuning performance of TerraMind in both unimodal and multimodal settings following the community-standard PANGAEA benchmark [49]. We summarize the results in Table 6. Overall, TerraMindv1-B outperforms all other GeoFMs by at least 3pp avg. mIoU. Importantly, we observe that TerraMind is the only foundation model approach in EO that across the PANGAEA benchmark outperforms task-specific U-Net models. Performance increases by approximately 2pp avg. mIoU for TerraMindv1-L, with a peak of 5pp in multimodal datasets. Furthermore, TerraMindv1-L outperforms also specialised ViT baselines by 5pp avg. mIoU. Note that per suggestion of the PANGAEA authors, we exclude the xView2 and BioMassters task as we could not reproduce the reported performances. Finally, we assess the impact of leveraging multimodal data as input to TerraMindv1-B compared to utilizing either optical or radar data as unimodal input to better understand the effect of leveraging multimodal data in finetuning. We observe that across all three multimodal tasks, TerraMindv1-B performs best with access to both optical and radar data." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 85, + 604, + 267, + 666 + ], + "blocks": [ + { + "bbox": [ + 85, + 604, + 267, + 666 + ], + "lines": [ + { + "bbox": [ + 85, + 604, + 267, + 666 + ], + "spans": [ + { + "bbox": [ + 85, + 604, + 267, + 666 + ], + "type": "table", + "html": "
PASTISSen1F11CTM-SS
S-120.0480.3924.45
S-240.2089.5750.90
S-1 + S-240.5190.6255.80
", + "image_path": "76d0dcb5a3b001a3aee1560a4be356fce95718ddbaf9e018da73acd5d653d386.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 674, + 296, + 696 + ], + "lines": [ + { + "bbox": [ + 55, + 674, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 296, + 696 + ], + "type": "text", + "content": "Table 7. Benefit of using multimodal input in the PANGAEA benchmark reported in mIoU " + }, + { + "bbox": [ + 55, + 674, + 296, + 696 + ], + "type": "inline_equation", + "content": "(\\%)\\uparrow" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 320, + 443, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 443, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 443, + 334 + ], + "type": "text", + "content": "5.6. Thinking in modalities" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 338, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 435 + ], + "type": "text", + "content": "We additionally evaluate the value of TiM tuning on water body mapping. We use S-1 or S-2 to generate artificial LULC data as additional input. Our results in Table 8 indicate a superior performance of TiM tuning compared to leveraging uni-modal data by up to 2pp mIoU. This finding points us in the direction of TerraMind being able to generate data that improve downstream task performance. We provide additional results in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 445, + 555, + 518 + ], + "blocks": [ + { + "bbox": [ + 315, + 445, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 315, + 445, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 315, + 445, + 555, + 518 + ], + "type": "table", + "html": "
Fine-TuningInputIoUWatermIoU
TerraMindv1-BS-168.0081.06
TerraMindv1-BS-282.2689.70
TerraMindv1-B TiMS-1 + gen. LULC72.2583.65
TerraMindv1-B TiMS-2 + gen. LULC84.7591.14
", + "image_path": "e57d722a55f43bd2375e5d31b0f63658d104c5a8813a596746414f0e63eeafbb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 526, + 556, + 549 + ], + "lines": [ + { + "bbox": [ + 313, + 526, + 556, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 556, + 549 + ], + "type": "text", + "content": "Table 8. Thinking-in-modalities (TiM) tuning compared with standard full fine-tuning approaches on the Sen1Floods11 dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "TerraMind's approach of combining token-level and pixel-level data has unlocked a range of new model capabilities in EO. TerraMind demonstrates not only beyond state-of-the-art performance in community-standard benchmarks, it also represents the first fully generative multimodal model in the domain. Because of the ability of integrating heterogeneous data sources, we expect that TerraMind-like models will expand to multi-temporal, multi-resolution, and hyperspectral data to fully leverage the data rich ecosystem available in the Earth Observation domain." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7390" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 123 + ], + "type": "text", + "content": "[1] A. Hore and D. Ziou. Image quality metrics: PSNR vs. SSIM. In Proc. 20th International Conference on Pattern Recognition (ICPR), pp. 2366-2369, 2010. 16" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 145 + ], + "type": "text", + "content": "[2] European Space Agency. Copernicus dem. http://dx.doi.org/10.5270/ESA-c5d3d65, 2022.4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 148, + 295, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 295, + 190 + ], + "type": "text", + "content": "[3] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Anysat: An earth observation model for any resolutions, scales, and modalities. arXiv preprint arXiv:2412.14123, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 192, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 295, + 224 + ], + "type": "text", + "content": "[4] Guillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Omnisat: Self-supervised modality fusion for earth observation, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 225, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 296, + 290 + ], + "type": "text", + "content": "[5] Nicolas Audebert, Bertrand Le Saux, and Sébastien Lefèvre. Joint Learning from Earth Observation and OpenStreetMap Data to Get Faster Better Semantic Maps. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1552-1560, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 292, + 296, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 296, + 345 + ], + "type": "text", + "content": "[6] Benedikt Blumenstiel, Nassim Ait Ali Braham, Conrad M Albrecht, Stefano Maurogiovanni, and Paolo Fraccaro. SSL4EOS12 v1.1 - A Multimodal, Multiseasonal Dataset for Pretraining. arXiv preprint arXiv:2503.00168, 2025. 3, 13" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 296, + 413 + ], + "type": "text", + "content": "[7] Benedikt Blumenstiel, Paolo Fraccaro, Valerio Marsocci, Johannes Jakubik, Stefano Maurogiovanni, Mikolaj Czerkawski, Rocco Sedona, Gabriele Cavallaro, Thomas Brunschwiler, Juan Bernabe-Moreno, and Nicolas Longépé. Terramesh: A planetary mosaic of multimodal earth observation data. arXiv preprint arXiv:2504.11172, 2025. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 468 + ], + "type": "text", + "content": "[8] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 470, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 296, + 525 + ], + "type": "text", + "content": "[9] Derrick Bonafilia, Beth Tellman, Tyler Anderson, and Erica Issenberg. Sen1floods11: A georeferenced dataset to train and test deep learning flood algorithms for sentinel-1. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 579 + ], + "type": "text", + "content": "[10] Florian Bordes, Richard Yuanzhe Pang, Anurag Ajay, Alexander C Li, Adrien Bardes, Suzanne Petryk, Oscar Manas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, et al. An introduction to vision-language modeling. arXiv preprint arXiv:2405.17247, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 581, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 296, + 647 + ], + "type": "text", + "content": "[11] Jize Cao, Zhe Gan, Yu Cheng, Licheng Yu, Yen-Chun Chen, and Jingjing Liu. Behind the scene: Revealing the secrets of pre-trained vision-and-language models. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 565-580. Springer, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "type": "text", + "content": "[12] Xu Cao, Tong Zhou, Yunsheng Ma, Wenqian Ye, Can Cui, Kun Tang, Zhipeng Cao, Kaizhao Liang, Ziran Wang, James M Rehg, et al. Maplm: A real-world large-scale vision-language benchmark for map and traffic scene understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21819-21830, 2024. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[13] Yuxing Chen and Lorenzo Bruzzone. Self-supervised change detection in multi-view remote sensing images. arXiv preprint arXiv:2103.05969, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 108, + 555, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 555, + 174 + ], + "type": "text", + "content": "[14] Chenwei Wang, et al. SAR Target Image Generation Method Using Azimuth-Controllable Generative Adversarial Network. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (JSTARS), Vol. 15, 2022. Online: http://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9933645&tag=1.16" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 176, + 555, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 176, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 176, + 555, + 209 + ], + "type": "text", + "content": "[15] Fabian Deuser, Konrad Habel, and Norbert Oswald. Sample4geo: Hard negative sampling for cross-view geolocation. arXiv preprint arXiv:2303.11851, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 211, + 555, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 211, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 317, + 211, + 555, + 264 + ], + "type": "text", + "content": "[16] Ivica Dimitrovski, Ivan Kitanovski, Dragi Kocev, and Nikola Simidjievski. Current trends in deep learning for earth observation: An open-source benchmark arena for image classification. ISPRS Journal of Photogrammetry and Remote Sensing, 197:18-35, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 267, + 555, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 267, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 267, + 555, + 332 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 335, + 554, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 335, + 554, + 379 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 554, + 379 + ], + "type": "text", + "content": "[18] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, Wenlong Huang, et al. Palm-e: An embodied multimodal language model. 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 381, + 496, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 381, + 496, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 381, + 496, + 392 + ], + "type": "text", + "content": "[19] Victor Durnov. xview2 1st place solution. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 395, + 555, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 395, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 395, + 555, + 426 + ], + "type": "text", + "content": "[20] Adam Van Etten, Dave Lindenbaum, and Todd M. Bacastow. Spacenet: A remote sensing dataset and challenge series, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 429, + 554, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 554, + 473 + ], + "type": "text", + "content": "[21] Casper Fibaek, Luke Camilleri, Andreas Luyts, Nikolaos Dionelis, and Bertrand Le Saux. PhilEO Bench: Evaluating Geo-Spatial Foundation Models, In Proc. Int Geoscience and Remote Sensing Symposium (IGARSS), 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 475, + 555, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 555, + 507 + ], + "type": "text", + "content": "[22] Alistair Francis. Sensor independent cloud and shadow masking with partial labels and multimodal inputs. IEEE Transactions on Geoscience and Remote Sensing, 2024. 4, 13" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 510, + 555, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 555, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 555, + 542 + ], + "type": "text", + "content": "[23] Alistair Francis and Mikolaj Czerkawski. Major tom: Expandable datasets for earth observation. arXiv preprint arXiv:2402.12095, 2024. 3, 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 544, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 544, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 554, + 599 + ], + "type": "text", + "content": "[24] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 601, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 633 + ], + "type": "text", + "content": "[25] Anthony Fuller, Korean Millard, and James R. Green. Croma: Remote sensing representations with contrastive radar-optical masked autoencoders, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 636, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 554, + 712 + ], + "type": "text", + "content": "[26] Anatol Garioud, Nicolas Gonthier, Loic Landrieu, Apolline De Wit, Marion Valette, Marc Poupee, Sebastien Giordano, and Boris Wattrelos. FLAIR: a country-scale land cover semantic segmentation dataset from multi-source optical imagery. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "7391" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "type": "text", + "content": "[27] Carlos Gomes, Isabelle Wittmann, Damien Robert, Johannes Jakubik, Tim Reichelt, Michele Martone, Stefano Maurogiovanni, Rikard Vinge, Jonas Hurst, Erik Scheurer, et al. Lossy neural compression for geospatial analytics: A review. arXiv preprint arXiv:2503.01505, 2025. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 171 + ], + "type": "text", + "content": "[28] Sebastian Hafner, Yifang Ban, and Andrea Nascetti. Unsupervised domain adaptation for global urban extraction using sentinel-1 sar and sentinel-2 msi data. Remote Sensing of Environment, 280:113192, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 205 + ], + "type": "text", + "content": "[29] Boran Han, Shuai Zhang, Xingjian Shi, and Markus Reichstein. Bridging remote sensors with multisensor geospatial foundation models, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 296, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 296, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 296, + 260 + ], + "type": "text", + "content": "[30] Soyeon Caren Han, Feiqi Cao, Josiah Poon, and Roberto Navigli. Multimodal large language models and tunings: Vision, language, sensors, audio, and beyond. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11294-11295, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "text", + "content": "[31] Jitesh Jain, Jianwei Yang, and Humphrey Shi. Vcoder: Versatile vision encoders for multimodal large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27992-28002, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 296, + 426 + ], + "type": "text", + "content": "[32] Johannes Jakubik, Sujit Roy, C. E. Phillips, Paolo Fraccaro, Denys Godwin, Bianca Zadrozny, Daniela Szwarcman, Carlos Gomes, Gabby Nyirjesy, Blair Edwards, Daiki Kimura, Naomi Simumba, Linsong Chu, S. Karthik Mikkavilli, Devyani Lambhate, Kamal Das, Ranjini Bangalore, Dario Oliveira, Michal Muszynski, Kumar Ankur, Muthukumaran Ramasubramanian, Iksha Gurung, Sam Khallaghi, Hanxi, Li, Michael Cecil, Maryam Ahmadi, Fatemeh Kordi, Hamed Alemohammad, Manil Maskey, Raghu Ganti, Kommy Weldemariam, and Rahul Ramachandran. Foundation models for generalist geospatial artificial intelligence, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "type": "text", + "content": "[33] Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "text", + "content": "[34] Samar Khanna, Patrick Liu, Linqi Zhou, Chenlin Meng, Robin Rombach, Marshall Burke, David Lobell, and Stefano Ermon. Diffusionsat: A generative foundation model for satellite imagery, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 514, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 296, + 602 + ], + "type": "text", + "content": "[35] Kohei Arai, Michihiro Mikamo, and Shunsuke Onishi. Method for Image Quality Evaluation of Satellite-based SAR Data. International Journal of Advanced Computer Science and Applications (IJACSA), Vol. 14, No. 7, 2023. Online: http://thesai.org/Downloads/Volume14No7/Paper_13-Method_for/Image_Quality_Evaluation_of_Satellite_based_SAR_Data.pdf.16" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 602, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 296, + 647 + ], + "type": "text", + "content": "[36] Saad Lahrichi, Zion Sheng, Shufan Xia, Kyle Bradbury, and Jordan Malof. Is self-supervised pre-training on satellite imagery better than imagenet? a systematic study with sentinel-2. arXiv preprint arXiv:2502.10669, 2025. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "text", + "content": "[37] Bo Li, Kaichen Zhang, Hao Zhang, Dong Guo, Renrui Zhang, Feng Li, Yuanhan Zhang, Ziwei Liu, and Chunyuan Li. Llavanext: Stronger llms supercharge multimodal capabilities in the wild, 2024. 4, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "type": "text", + "content": "[38] Jiaxin Li, Danfeng Hong, Lianru Gao, Jing Yao, Ke Zheng, Bing Zhang, and Jocelyn Chanussot. Deep learning in mul" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "timodal remote sensing data fusion: A comprehensive review. International Journal of Applied Earth Observation and Geoinformation, 112:102926, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 108, + 555, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 555, + 152 + ], + "type": "text", + "content": "[39] Ke Li, Gang Wan, Gong Cheng, Liqui Meng, and Junwei Han. Object detection in optical remote sensing images: A survey and a new benchmark. ISPRS journal of photogrammetry and remote sensing, 159:296-307, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 153, + 555, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 555, + 186 + ], + "type": "text", + "content": "[40] Xiang Li, Congcong Wen, Yuan Hu, Zhenghang Yuan, and Xiao Xiang Zhu. Vision-language models in remote sensing: Current progress and future trends, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 189, + 555, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 189, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 189, + 555, + 243 + ], + "type": "text", + "content": "[41] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19325-19337, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 246, + 555, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 555, + 289 + ], + "type": "text", + "content": "[42] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing, 2024. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 292, + 555, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 292, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 292, + 555, + 324 + ], + "type": "text", + "content": "[43] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 392 + ], + "type": "text", + "content": "[44] Gabriel Machado, Edemir Ferreira, Keiller Nogueira, Hugo Oliveira, Matheus Brito, Pedro Henrique Targino Gama, and Jefersson Alex dos Santos. Airround and cv-brct: Novel multiview datasets for scene classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:488-503, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 395, + 555, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 395, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 316, + 395, + 555, + 461 + ], + "type": "text", + "content": "[45] Gengchen Mai, Chris Cundy, Kristy Choi, Yingjie Hu, Ni Lao, and Stefano Ermon. Towards a foundation model for geospatial artificial intelligence (vision paper). In Proceedings of the 30th International Conference on Advances in Geographic Information Systems, New York, NY, USA, 2022. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 463, + 555, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 463, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 463, + 555, + 518 + ], + "type": "text", + "content": "[46] Oscar Manas, Alexandre Lacoste, Xavier Giró-i Nieto, David Vazquez, and Pau Rodriguez. Seasonal contrast: Unsupervised pre-training from uncurated remote sensing data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9414-9423, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 520, + 555, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 555, + 574 + ], + "type": "text", + "content": "[47] Clive Tinashe Marimo, Benedikt Blumenstiel, Maximilian Nitsche, Johannes Jakubik, and Thomas Brunschwiler. Beyond the visible: Multispectral vision-language learning for earth observation. arXiv preprint arXiv:2503.15969, 2025. 2, 4, 13" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 577, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 555, + 609 + ], + "type": "text", + "content": "[48] Valerio Marsocci and Nicolas Audebert. Cross-sensor self-supervised training and alignment for remote sensing, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 612, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 555, + 667 + ], + "type": "text", + "content": "[49] Valerio Marsocci, Yuru Jia, Georges Le Bellier, David Kerekes, Liang Zeng, Sebastian Hafner, Sebastian Gerard, Eric Brune, Ritu Yadav, Ali Shibli, et al. Pangaea: A global and inclusive benchmark for geospatial foundation models. arXiv preprint arXiv:2412.04204, 2024. 2, 8, 18" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "type": "text", + "content": "[50] Matias Mendieta, Boran Han, Xingjian Shi, Yi Zhu, Chen Chen, and Mu Li. Gfm: Building geospatial foundation models via continual pretraining. arXiv preprint arXiv:2302.04476, 2023. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7392" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[51] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple. arXiv preprint arXiv:2309.15505, 2023. 4, 15" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "type": "text", + "content": "[52] David Mizrahi, Roman Bachmann, Oğuzhan Fatih Kar, Teresa Yeo, Mingfei Gao, Afshin Dehghan, and Amir Zamir. 4m: Massively multimodal masked modeling, 2023. 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 206 + ], + "type": "text", + "content": "[53] Andrea Nascetti, RITU YADAV, Kirill Brodt, Qixun Qu, Hongwei Fan, Yuri Shendryk, Isha Shah, and Christine Chung. Biomasssters: A benchmark dataset for forest biomass estimation using multi-modal satellite time-series. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[54] Vishal Nedungadi, Ankit Kariryaa, Stefan Oehmcke, Serge Belongie, Christian Igel, and Nico Lang. Mmearth: Exploring multi-modal pretext tasks for geospatial representation learning. arXiv preprint arXiv:2405.02771, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 297 + ], + "type": "text", + "content": "[55] Fernando Paolo, Tsu ting Tim Lin, Ritwik Gupta, Bryce Goodman, Nirav Patel, Daniel Kuster, David Kroodsma, and Jared Dunnmon. xview3-sar: Detecting dark fishing activity using synthetic aperture radar imagery, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 296, + 364 + ], + "type": "text", + "content": "[56] Prabhishek Singh and Raj Shree. Analysis and effects of speckle noise in SAR images. In Proc. International Conference on Advances in Computing, Communication, & Automation (ICACCA), 2016. DOI: 10.1109/ICAC-CAF.2016.7748978. Online: http://ieeexplore.ieee.org/document/7748978.16" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 431 + ], + "type": "text", + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PmLR, 2021. 3, 17" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "type": "text", + "content": "[58] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 296, + 532 + ], + "type": "text", + "content": "[59] Ayesha Shafique, Guo Cao, Zia Khan, Muhammad Asad, and Muhammad Aslam. Deep learning-based change detection in remote sensing images: A review. Remote Sensing, 14(4): 871, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 296, + 567 + ], + "type": "text", + "content": "[60] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. Advances in neural information processing systems, 30, 2017. 17" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 601 + ], + "type": "text", + "content": "[61] Aidan M Swope, Xander H Rudelis, and Kyle T Story. Representation learning for remote sensing: An unsupervised sensor fusion approach. arXiv preprint arXiv:2108.05094, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 296, + 677 + ], + "type": "text", + "content": "[62] Devis Tuia, Konrad Schindler, Begüm Demir, Gustau Camps-Valls, Xiao Xiang Zhu, Mrinalini Kochupillai, Sašo Džeroski, Jan N. van Rijn, Holger H. Hoos, Fabio Del Frate, Mihai Datcu, Jorge-Arnulfo Quiane-Ruiz, Volker Markl, Bertrand Le Saux, and Rochelle Schneider. Artificial intelligence to advance earth observation: a perspective, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "text", + "content": "[63] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 106 + ], + "type": "text", + "content": "[64] Yi Wang, Conrad M Albrecht, Nassim Ait Ali Braham, Lichao Mou, and Xiao Xiang Zhu. Self-supervised learning in remote sensing: A review. arXiv preprint arXiv:2206.13188, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 171 + ], + "type": "text", + "content": "[65] Yi Wang, Nassim Ait Ali Braham, Zhitong Xiong, Chenying Liu, Conrad M Albrecht, and Xiao Xiang Zhu. Ssl4eos12: A large-scale multimodal, multitemporal dataset for self-supervised learning in earth observation [software and data sets]. IEEE Geoscience and Remote Sensing Magazine, 11 (3):98-106, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 555, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 555, + 237 + ], + "type": "text", + "content": "[66] Jiannan Wu, Muyan Zhong, Sen Xing, Zeqiang Lai, Zhaoyang Liu, Zhe Chen, Wenhai Wang, Xizhou Zhu, Lewei Lu, Tong Lu, et al. Visionllm v2: An end-to-end generalist multimodal large language model for hundreds of vision-language tasks. Advances in Neural Information Processing Systems, 37:69925-69975, 2025. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 555, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 555, + 281 + ], + "type": "text", + "content": "[67] Xinyu Bai and Feng Xu. Accelerating Diffusion for SAR-to-Optical Image Translation via Adversarial Consistency Distillation, 2024. Online: http://arxiv.org/pdf/2407.06095.16" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 282, + 554, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 554, + 337 + ], + "type": "text", + "content": "[68] Zhitong Xiong, Yi Wang, Fahong Zhang, Adam J. Stewart, Joëlle Hanna, Damian Borth, Ioannis Papoutsis, Bertrand Le Saux, Gustau Camps-Valls, and Xiao Xiang Zhu. Neural plasticity-inspired foundation model for observing the earth crossing modalities, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 555, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 555, + 381 + ], + "type": "text", + "content": "[69] Lingxiao Yang, Ru-Yuan Zhang, Yanchen Wang, and Xiaohua Xie. Mma: Multi-modal adapter for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23826-23837, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 555, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 555, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 555, + 446 + ], + "type": "text", + "content": "[70] Qidong Yang, Jonathan Giezendanner, Daniel Salles Civitarese, Johannes Jakubik, Eric Schmitt, Anirban Chandra, Jeremy Vila, Detlef Hohl, Chris Hill, Campbell Watson, et al. Multi-modal graph neural networks for localized off-grid weather forecasting. arXiv preprint arXiv:2410.12938, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 449, + 555, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 501 + ], + "type": "text", + "content": "[71] Zhiping Yu, Chenyang Liu, Liqin Liu, Zhenwei Shi, and Zhengxia Zou. Metaearth: A generative foundation model for global-scale remote sensing image generation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 503, + 555, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 555, + 546 + ], + "type": "text", + "content": "[72] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 548, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 555, + 591 + ], + "type": "text", + "content": "[73] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, vol. 13, no. 4, pp. 600-612, 2004. 16" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 593, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 555, + 634 + ], + "type": "text", + "content": "[74] Jingyi Zhang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 636, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 555, + 690 + ], + "type": "text", + "content": "[75] Linying Zhao and Shunping Ji. Cnn, rn, or vit? an evaluation of different deep learning architectures for spatio-temporal representation of sentinel time series. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 16:44-56, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[76] Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7393" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 73, + 297, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 73, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 73, + 297, + 106 + ], + "type": "text", + "content": "learning in remote sensing: A comprehensive review and list of resources. IEEE geoscience and remote sensing magazine, 5(4):8-36, 2017. 2" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "7394" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/0a2ddfb1-a84e-4b84-bcf1-12ab1957047a_content_list.json b/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/0a2ddfb1-a84e-4b84-bcf1-12ab1957047a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a5971b18b5c74f6c362b30e86a317faf64072838 --- /dev/null +++ b/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/0a2ddfb1-a84e-4b84-bcf1-12ab1957047a_content_list.json @@ -0,0 +1,1630 @@ +[ + { + "type": "text", + "text": "Test-Time Prompt Tuning for Zero-Shot Depth Completion", + "text_level": 1, + "bbox": [ + 200, + 130, + 797, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chanhwi Jeong1 Inhwan Bae1 Jin-Hwi Park2* Hae-Gon Jeon3* \n1GIST, 2Chung-Ang University, 3Yonsei University", + "bbox": [ + 238, + 179, + 763, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chanhwij, inhwanbae}@gm.gist.ac.kr, jinhwipark@cau.ac.kr, earboll@yonsei.ac.kr", + "bbox": [ + 150, + 219, + 844, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfa64ab433e1b9170c18b40a0a8f199cf234bd6e3c86a2ea8349236d98564ad5.jpg", + "image_caption": [ + "Figure 1. Comparison of methods for correcting the inaccurate scale of initial depth estimates produced by the depth foundation model using sensor-derived sparse depth with an accurate scale, represented as yellow dots. (a) The depth foundation model generates an initial depth map from a single RGB input. (b) The least squares fitting aligns the initial depth map with the sparse depth [54]. (c) The parameters of the foundation model are fine-tuned under supervision from the sparse depth. (d) Alternatively, the model's parameters are frozen, and a visual prompt is learned to adjust the depth estimates. Please refer to Fig. 3 for depth profiling results and error analysis." + ], + "image_footnote": [], + "bbox": [ + 91, + 243, + 906, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 455, + 326, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zero-shot depth completion using metric scales remains challenging, primarily due to performance limitations such as domain specificity and sensor characteristics. One recent emerging solution is to integrate monocular depth foundation models into depth completion frameworks, yet such efforts still face issues with suboptimal performance and often require further adaptation to the target task. Surprisingly, we find that a simple test-time training, which finetunes monocular depth foundation models on sparse depth measurements from sensors just as it is, yields reasonable results. However, this test-time training obviously incurs high computational costs and introduces biases towards specific conditions, making it impractical for real-world scenarios. In this paper, we introduce a new approach toward parameter-efficient zero-shot depth completion. Our key idea in this work is to leverage visual prompt tuning, achieving sensor-specific depth scale adaptation without forgetting foundational knowledge. Experimental results on diverse datasets demonstrate that our approach outperforms relevant state-of-the-art methods, showing superior generalization and efficiency. Code is publicly available at https://github.com/ch5374/TestPromptDC", + "bbox": [ + 86, + 477, + 485, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 825, + 220, + 842 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth completion, the technique of densifying sparse input measurements from active sensors based on correspond", + "bbox": [ + 89, + 848, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing RGB images, is crucial in various applications including 3D reconstruction [72], autonomous driving [39] and robotics [33]. Beyond traditional methods that rely on handcrafted features [34] and optimization [14], data-driven approaches, which are mainly based on convolutional neural networks [27, 46, 47] and vision transformers [64, 84], have achieved the best performance with public benchmarks [15, 25].", + "bbox": [ + 511, + 455, + 908, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, models trained on a single dataset perform well on the test split of that dataset, but exhibit significant performance degradation in unseen environments. The primary reason lies in the scarcity of data available to use for depth completion. Since depth completion datasets require both dense ground truth depth maps and RGB images obtained from well-calibrated sensor fusion setups, large-scale dataset constructions must collect and label image-depth pairs across environmental variations. Obviously, this makes it difficult to achieve generalizable depth completion, including zero-shot inferences.", + "bbox": [ + 511, + 580, + 908, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One interesting research direction in [40, 52, 54] to mitigate the data scarcity is to integrate foundation models for monocular depth estimation [4, 17, 22, 30, 57, 58, 60, 61, 81-83] trained on huge-scale datasets into depth completion for better generalization. This is quite feasible because foundation models can generate approximate metric or relative depth from a single input image as demonstrated in Fig. 1-(a), actual scene depths can be obtained by fitting sparse depth measurements onto these predictions. Unfortunately, there are still two challenges: (1) they require carefully designed", + "bbox": [ + 511, + 750, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors", + "bbox": [ + 109, + 887, + 238, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9443", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "additional networks, as the existing foundation models struggle to encode sparse depth information; (2) additional training for the depth completion task is also necessary. These two challenges result in either an increase in computational complexity or the introduction of new biases in the training phase. In this paper, we explore one important question: how can we take full advantage of monocular depth foundation models for depth completion? Our work is the first to answer this question by investigating the most effective foundation models, without requiring additional network design or fine-tuning of the pretrained model.", + "bbox": [ + 89, + 90, + 485, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We initiate our investigation by proposing a simple yet highly effective strategy, which is fine-tuning the foundation models in the inference phase, only using sparse depth measurements as supervision. As an example in Fig. 1-(b), compared to the sparse depth fitting onto dense depths [40, 52, 54], fine-tuning the foundation models preserves the original depth while allowing the scale to adapt locally (Fig. 1-(c)). Of course, updating all parameters requires significant computing resources and poses the risk of dataset bias, potentially causing models to lose their own knowledge.", + "bbox": [ + 89, + 258, + 485, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a more efficient alternative, we present a test-time prompt tuning that adapts foundation models at inference time without modifying its parameters (Fig. 1-(d)). We introduce a small set of learnable parameters that guide their prediction while keeping the foundation model frozen. While existing visual prompt [1, 28] uses fixed-size prompts positioned at specific locations (e.g., fixed positions, random placements, or padding) for image classification, we introduce a pixelwise visual prompt specifically designed for depth completion. Compared to previous works that operate at a global or patch level, our pixelwise visual prompt is initialized to zero and added to input images, allowing us to leverage the inherent spatial structures of depth maps. We validate our visual prompt tuning across diverse benchmark datasets and environments, achieving superior zero-shot performance.", + "bbox": [ + 89, + 426, + 485, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 684, + 227, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Monocular Depth Estimation", + "text_level": 1, + "bbox": [ + 89, + 710, + 352, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional monocular depth estimation depends on handcrafted visual cues [20, 67] and scene-specific assumptions [21, 42]. With the advent of deep learning, subsequent methods [16, 35, 36] using data-driven representations have significantly outperformed the traditional ones reducing the reliance on manually designed features and rigid scene-specific assumptions. Recent works have explored techniques such as data mixing [55, 60, 61] and augmentation [25, 66] to achieve zero-shot generality. Moreover, some works in [10, 17, 30, 50, 56, 85] exploit the generative priors of pretrained diffusion models [65]. However, inferring", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "metric-scale depth from single images still remains challenging. To address this, metric-scale depth estimation from monocular images has evolved by incorporating global depth distributions [19], scene-aware conditioning [3] and camera intrinsics [4, 22, 22, 57, 58, 83]. While earlier methods rely on known camera parameters [22, 83], recent approaches estimate focal length [4, 57, 58] directly from learning features, enhancing generalization across diverse environments without any explicit camera metadata. Nevertheless, these methods obviously lag behind approaches that use sensor measurements like LiDAR in terms of accuracy.", + "bbox": [ + 511, + 90, + 908, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Depth Completion", + "text_level": 1, + "bbox": [ + 511, + 273, + 692, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the inherent issues with monocular depth estimation, research on unguided depth completion [12, 23, 73, 75], transforming sparse depth data from sensors into dense depth maps without guidance RGB images, has been studied. However, most mainstream depth completion is to use RGB-depth pairs to directly leverage scene contexts such as edges and textures [26, 46, 47, 68, 70, 79, 80, 84, 87]. Furthermore, constructing pixelwise affinity [6, 7, 41, 43, 45, 51, 53] from corresponding RGB images has proven to be effective as it helps preserve depth discontinuity when propagating sparse depth. Despite these advancements, such models often struggle to show the same performance in unseen domains. To mitigate this, recent approaches have leveraged monocular depth foundation models for depth completion. For instance, works in [29, 54] introduces interesting frameworks that exploit features and predict depth from foundation models to perform adaptive depth refinement with sparse depth inputs. Additionally, diffusion-based depth completion methods [24, 74] have adopted the test-time optimization of latent representations, guided by input sparse depths. However, these approaches often require training on additional data [29, 54] or suffer from extremely slow inference speeds [24, 74].", + "bbox": [ + 511, + 299, + 908, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Prompting for Foundation Models", + "text_level": 1, + "bbox": [ + 511, + 648, + 813, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Foundation models [5, 9, 31, 59, 62, 65] exhibit remarkable zero-shot performance across a wide range of tasks as well as depth perception. In natural language processing, prompting schemes have been introduced to leverage these foundation models effectively for downstream tasks [5, 37, 38, 44, 63, 77]. To apply such prompting schemes to the computer vision domain, works in [28, 86] have introduced visual prompt tuning, which incorporates trainable visual tokens into models for specific downstream tasks. Meanwhile, one interesting work in [1] presents learnable padding prompts appended to input images. More recently, visual prompt tuning has also been extended to test-time adaptation scenarios [13, 48, 49] for better adaptation to unseen domains. Inspired by them [1, 49], we propose a test-time prompt tuning strategy for depth completion.", + "bbox": [ + 511, + 674, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9444", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1a945e7b4c26cb908271025f0d40222334bfb3f4620048d27dbd692a0fc1d723.jpg", + "image_caption": [ + "(a) Supervised training models" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 282, + 165 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ee70203d226eb9cccedaa32f6acbe04e5f707396c58d885efa2bba94661acade.jpg", + "image_caption": [ + "(b) Test-time optimization models" + ], + "image_footnote": [], + "bbox": [ + 297, + 95, + 506, + 159 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/af5aeed83572c04e80106469f99864c72eabe6bcf42e5217be0105d3e10c28e9.jpg", + "image_caption": [ + "(c) Our fine-tuning approach" + ], + "image_footnote": [], + "bbox": [ + 535, + 95, + 710, + 159 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6a181e379d0decf80db53c867c259f57effe87e5fceb59b8822cff99481a2b5b.jpg", + "image_caption": [ + "(d) Our visual prompting approach" + ], + "image_footnote": [], + "bbox": [ + 727, + 95, + 903, + 159 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 266, + 225, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our approach for parameter-efficient zero-shot depth completion through test-time training. In Sec. 3.1, we discuss the role of depth foundation models in depth completion, emphasizing their generalization capabilities while identifying key challenges. Next, in Sec. 3.2, we explore a fine-tuning strategy to adapt the foundation model for sparse depth measurements. This approach achieves impressive performance without relying on additional network design or prompt engineering; however, it introduces computational overhead and poses generalization challenges. To address these limitations, in Sec. 3.3, we propose a novel pixel-wise visual prompting method, enabling efficient adaptation without any parameter modification of the foundation model. Finally, in Sec. 3.4, we describe implementation details, including optimization strategies, dataset preprocessing and experimental settings.", + "bbox": [ + 89, + 292, + 483, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Rationale: Usage of Depth Foundation Model", + "text_level": 1, + "bbox": [ + 89, + 549, + 475, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent works have attempted to leverage monocular depth foundation models for depth completion. These models offer strong generalization capabilities and are considered to be attractive solutions for better zero-shot performance. However, incorporating them into depth completion frameworks faces several technical challenges, as described below:", + "bbox": [ + 89, + 573, + 483, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Incompatibility with sparse depth input handling. Since monocular depth foundation models are exclusively designed to infer depth from single RGB images, they do not have mechanisms to take sparse depth measurements as guidance. As a result, the existing methods require additional components to process sparse depths and to integrate them into the models, increasing computational complexity as shown in Fig. 2-(a).", + "bbox": [ + 89, + 669, + 483, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scale mismatch between prediction and sensor depth. Current depth foundation models are able to estimate approximate metric-scale depths [3, 4, 58] as well as relative-scale depth [17, 30, 60, 61] with high-fidelity depth discontinuities. Unfortunately, the depth foundation models for metric-scale inference exhibit systematic scale mismatches when compared to sensor-derived depths. While least squares scale", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b2a7a78d9ad26d5c41b4e35c857d384557f110a32d60c90daca7770c08ca217d.jpg", + "image_caption": [ + "Figure 2. Comparison of the previous foundation model-based depth completion approaches and our approach. (a) Existing methods [40, 52, 54] design additional networks to integrate sensor-based depth measurements into the foundation model. (b) Recent approaches [24, 74] optimize latent representations in diffusion-based foundation models [17, 30] without dense GT but are computationally expensive. (c) Our fine-tuning method adapts the model with sparse depth supervision, and (d) our visual prompting approach enables efficient adaptation while keeping the foundation model frozen.", + "(a) Least squares fitting", + "Figure 3. Qualitative comparison of different strategies on the depth map from the data used in Fig. 1. The top row shows the predicted depth maps, while the bottom row presents the depth profiles along the indicated horizontal line. The red $\\times$ marks indicate the sparse depth measurements." + ], + "image_footnote": [], + "bbox": [ + 514, + 263, + 643, + 375 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ed49bf74b93ebcd72f25803b4ce9918a24788aa2e185869a03979475c9569076.jpg", + "image_caption": [ + "(b) Fine-tuning" + ], + "image_footnote": [], + "bbox": [ + 645, + 263, + 774, + 375 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c5778b9833e91ee17ac1158e65636d8a2fbc54a1f8148bd23ae65050e4aa65a3.jpg", + "image_caption": [ + "(c) Visual Prompting" + ], + "image_footnote": [], + "bbox": [ + 777, + 263, + 905, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fitting in [54] can globally align depth scales, it fails in regions where depth varies significantly. For instance, as shown in Fig. 3, the least squares fitting locally adjusts the correct depth scale at areas near initial depth seeds (e.g., pixel range 100–150), but fails it in other regions (e.g., pixel range 50–100).", + "bbox": [ + 511, + 483, + 906, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Limitations of existing adaptation strategies. Previous approaches require ground truth (GT) annotations for adaptation, which are costly and labor-intensive (Fig. 2-(a)). For example, works in [40, 54] rely on large-scale datasets, while UniDC [52] adopts a few-shot strategy within the target domain. However, generating new GT labels to cover all scenarios in-the-wild, is impractical, necessitating a more efficient alternative. One promising solution is a test-time optimization, which enables adaptation to new environments without any additional GT annotation (Fig. 2-(b)). Recent contemporary works in [24, 74] have introduced similar strategies by optimizing latent representations in diffusion-based models (Fig. 2-(b)). However, these approaches come with noticeable limitations, including high computational costs and slow inference speeds.", + "bbox": [ + 511, + 580, + 906, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these limitations, we propose fine-tuning depth foundation models in the inference phase, directly using sparse depth measurements as supervision without introducing additional network components (Fig. 2-(c)). This eliminates the dependency on extensive GT data and enables real-time adaptation to new environments.", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9445", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Finetuning monocular depth foundation model", + "text_level": 1, + "bbox": [ + 89, + 90, + 483, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $F_{\\theta}$ denote a depth foundation model, parametrized by $\\theta$ , and $I \\in \\mathbb{R}^{3 \\times H \\times W}$ is an RGB image. Typically, $F_{\\theta}$ takes only the $I$ as input and generates a dense depth map $D \\in \\mathbb{R}^{1 \\times H \\times W}$ as output, which can be expressed as follows:", + "bbox": [ + 89, + 113, + 485, + 175 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\theta} (I) = D. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 180, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since $D$ is derived solely from a single image, it inherently suffers from scale ambiguity. To mitigate this issue, a sparse depth $S$ , obtained from an active sensor with accurate scales, is employed as supervision for $F_{\\theta}$ . This approach allows the model to alleviate the depth scale ambiguity.", + "bbox": [ + 89, + 200, + 483, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, to use only valid sparse depth values as supervision, we define a validity mask $M$ as follows:", + "bbox": [ + 89, + 279, + 483, + 310 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {i, j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} S _ {i, j} \\text {i s v a l i d} \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 314, + 483, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(i,j)$ is the pixel coordinate in the image.", + "bbox": [ + 89, + 359, + 410, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this mask, we initially adjust the scale of the foundation depth by roughly aligning it using the valid sparse depth measurements in the least squares fitting process, as formulated below:", + "bbox": [ + 89, + 376, + 483, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {0}, s _ {1} = \\min _ {\\left(s _ {0}, s _ {1}\\right)} \\| M \\odot \\left(\\left(s _ {0} D + s _ {1}\\right) - S\\right) \\| _ {F}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 441, + 483, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes an element-wise multiplication. $s_0$ and $s_1$ mean scale and shift values, respectively, and $\\| \\cdot \\| _F$ indicates the Frobenius norm. Next, we optimize the model's parameters $\\theta$ using only the valid sparse depth measurements as supervision. The optimal parameters, denoted by $\\theta^{*}$ , are determined by minimizing the following loss:", + "bbox": [ + 89, + 470, + 483, + 563 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {*} = \\arg \\min _ {\\theta} \\sum_ {i, j} M _ {i, j} \\odot | s _ {0} F _ {\\theta} (I) _ {i, j} + s _ {1} - S _ {i, j} | ^ {p}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 566, + 483, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p = 1$ and $p = 2$ correspond to the $L_{1}$ and $L_{2}$ losses, respectively and the final loss is the sum of both. To isolate the effects of fine-tuning itself, we do not incorporate specialized loss functions.", + "bbox": [ + 89, + 604, + 485, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Pixelwise Visual Prompt", + "text_level": 1, + "bbox": [ + 89, + 678, + 313, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the computational overhead and potential knowledge degradation of the fine-tuning of the foundation model, we propose a pixelwise visual prompt, a novel test-time adaptation method designed specifically for depth completion (Fig. 2-(d)). Unlike conventional fine-tuning, our approach enables efficient adaptation while keeping the model's parameters frozen.", + "bbox": [ + 89, + 702, + 485, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a frozen pretrained foundation model $F_{\\theta}$ and a sample consisting of a pair of image and sparse depth measurement $(I,S)$ , we aim to learn a pixelwise visual prompt $V_{\\phi} \\in \\mathbb{R}^{3 \\times H \\times W}$ , parameterized by $\\phi$ . The prompt is added directly to the input image, forming a prompted image $I + V_{\\phi}$ and we optimize it by minimizing the depth prediction error", + "bbox": [ + 89, + 810, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in the training phase as follows:", + "bbox": [ + 511, + 90, + 725, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi^ {*} = \\arg \\min _ {\\phi} \\sum_ {i, j} M _ {i, j} \\odot \\left| F _ {\\theta} \\left(I + V _ {\\phi}\\right) _ {i, j} - S _ {i, j} \\right| ^ {p}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 111, + 906, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, only the prompt parameters $\\phi$ are optimized, while $\\theta$ remains frozen. Once optimized, the visual prompt $V_{\\phi^*}$ is applied to the image as below:", + "bbox": [ + 511, + 148, + 906, + 195 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\theta} \\left(I + V _ {\\phi^ {*}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 199, + 906, + 217 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to achieving superior performance compared to existing methods, this approach also enhances usability, providing an additional advantage. Since the pixelwise visual prompt is initialized to zero and added to the image, it ensures a minimal level of depth prediction performance even without prompt tuning. Additionally, users can dynamically adjust the time allocated for prompt tuning based on their available resources and the required inference speed, allowing for stable and consistent performance improvement.", + "bbox": [ + 511, + 220, + 908, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Difference with existing visual prompt methods. While visual prompt tunings [1, 28] have been previously explored to adapt high-level representations using fixed-size prompts at predefined locations, such approaches lack the spatial granularity needed for precise depth refinement. These methods often focus on global or patch-level modifications, making them less effective for tasks requiring fine-grained adjustments, such as depth completion. In contrast, our pixelwise visual prompt is designed to adapt at the individual pixel level, enabling localized corrections that align depth predictions more accurately with sparse sensor measurements. Our approach not only enhances adaptation flexibility but also maintains the strong generalization capabilities of foundation models.", + "bbox": [ + 511, + 361, + 908, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 582, + 732, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our implementation is based on PyTorch and runs on a single NVIDIA RTX 3090 GPU. We use the AdamW optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ . The learning rate is set to 1e-6 for fine-tuning and 1e-3 for visual prompting. The training process consists of 80 steps for fine-tuning and 50 steps for visual prompting. For all depth foundation models, we use their official pretrained weights and specifically the v2-old version for UniDepthV2 [58], while other models are evaluated in our ablation study. For models that do not support the original image size, we first resize the images while preserving their aspect ratio.", + "bbox": [ + 511, + 604, + 908, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 784, + 638, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we begin by outlining the experimental setup in Sec. 4.1, including the evaluation protocols, and introduce the datasets. We then compare our method with state-of-the-art (SOTA) methods for zero-shot depth completion tasks, in Sec. 4.2. Lastly, in Sec. 4.3, we present an ablation study to analyze the impact of our methods.", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9446", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db46f0ece6530514232395271c50d03d3db527790e9f90a9745040aecef78c4a.jpg", + "image_caption": [ + "Figure 4. Qualitative result and depth error map of various depth completion methods on the IBims-1, VOID 1500, and DDAD datasets." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 908, + 470 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/de1800b1d1c6f94b72bb40abf488c7c2726b4800b38d341f4df0145e0fb10c84.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIBims-1VOIDNYUv2KITTIDDAD
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
NLSPN [51]0.0490.1910.4271.3530.4400.7161.3352.0762.4989.231
CompletionFormer [84]0.0580.2060.6421.6840.1860.3740.9521.9352.5189.471
VPP4DC [2]0.0620.2280.1660.6060.0770.2470.4131.6091.3446.781
BP-Net [71]0.0780.2890.2700.742----2.2708.344
OGNI-DC [87]0.0590.1860.1750.593----1.8676.876
Zero-DC [24]0.0780.2030.1870.6070.0690.1580.3971.4131.6817.122
Marigold-DC [74]0.0450.1660.1520.5510.0480.1240.4341.4652.3646.449
Ours-LS0.1130.2210.1870.5710.1230.2161.2052.8015.67611.97
Ours-FT (Sec. 3.2)0.0490.1520.1380.5140.0480.1070.3831.3291.7755.881
Ours-VP (Sec. 3.3)0.0430.1610.1320.5080.0410.1060.3431.3331.4825.811
", + "bbox": [ + 91, + 494, + 901, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison with SOTA methods on five zero-shot benchmarks. Most methods used values reported in [74], while the results for Zero-DC [24] are evaluated based on the DepthFM model [17]. For the same reason as in [74], we do not include results for BP-Net [71] and OGNI-DC [87] on NYUv2 and KITTI. Bold: Best, Underline: Second best.", + "bbox": [ + 89, + 669, + 906, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 727, + 282, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our test-time training strategy is closely aligned with recent works [24, 74], which focus on optimizing the model at inference time. Following their methodology, we adapt the model using only sparse depth measurements during test-time, ensuring a fully zero-shot approach while effectively refining depth scale estimation. To maintain a fair evaluation, we strictly exclude the dataset (e.g., ScanNet [8]) used for training the depth foundation model [58] from the test. Furthermore, in our fine-tuning experiments, we prevent the model from accumulating dataset-specific biases by fine-tuning on", + "bbox": [ + 89, + 750, + 485, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a single sample and resetting it to its original parameters before proceeding to the next evaluation. This ensures that our approach remains consistent with the principles of test-time optimization and maintains its generalization ability across diverse environments.", + "bbox": [ + 511, + 728, + 906, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our proposed method in a zero-shot setting using five real-world datasets that are not seen during the training of the base model [58]. These datasets encompass both indoor and outdoor environments, varying image resolutions, and different levels of sparse depth density, ensuring a comprehensive assessment of generalization ability.", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9447", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4b2f2f1ba073f013763a8a73e279734a2b233cb21263c9303ab8c3bca59cbfad.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison under varying levels of depth sparsity across NYUv2 and KITTI. Despite significantly reduced depth measurements, our approach recovers high-quality dense depth maps, demonstrating robustness to input sparsity in the depth completion." + ], + "image_footnote": [], + "bbox": [ + 101, + 87, + 877, + 388 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bc10024ed1d80feb86656f803c1face5c090a60d61cfaef5694db2803dbb70ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNYUv2 (200 samples)NYUv2 (100 samples)NYUv2 (50 samples)NYUv2 (5 samples)KITTI (32 channels)KITTI (16 channels)KITTI (8 channels)KITTI (4 channels)
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
Zero-DC [24]0.0960.1970.1240.2340.1610.2820.4470.6230.5401.9160.6942.3231.0082.9041.6874.184
Marigold-DC [74]0.0730.1750.0900.2030.1120.2360.3580.5410.8752.1930.9872.6121.1863.1761.6694.240
Ours-VP0.0530.1330.0650.1520.0800.1800.1380.2470.3961.5400.4751.8510.5672.0930.8122.644
", + "bbox": [ + 91, + 434, + 906, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison on varying sparsity. Bold: Best, Underline: Second best.", + "bbox": [ + 228, + 516, + 764, + 531 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) IBims-1 [32] is an indoor dataset consisting of 100 RGB-D pairs with a resolution of $480 \\times 640$ . We use the entire dataset for testing. Since sparse depth data is not provided, we follow [74] to generate a depth valid mask and construct sparse depth by randomly sampling 1,000 depth points.", + "(2) VOID [78] contains RGB-D frames from 56 sequences, with 48 designated for training and 8 for testing. The official test set comprises 800 frames, and sparse depth maps are obtained using the VIO system. Following [74], we evaluate our method on this test set using 1,500 sparse depth points." + ], + "bbox": [ + 88, + 546, + 482, + 705 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) NYUv2 [69] comprises 464 scenes recorded using a Microsoft Kinect sensor. The official split is divided into 249 training and 215 test scenes. We use the official test set of 654 images for evaluation. Following a commonly used setup [11, 35, 46], we downsample and center-crop the images to $228 \\times 304$ before conducting experiments. Also, sparse depth is obtained by randomly selecting 500 samples from the depth map.", + "(4) KITTIDC [15] consists of over 90K RGB images paired with LiDAR data. Since the depth obtained from LiDAR is inherently sparse, the ground truth depth is generated by aggregating multiple LiDAR frames. For our experiments," + ], + "bbox": [ + 88, + 710, + 486, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we use the 1,000 pairs provided for validation. To reduce noise in the input sparse depth, we apply the filtering method used in [76].", + "bbox": [ + 511, + 546, + 906, + 590 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(5) DDAD [18] is an outdoor RGB-D driving dataset collected using a Luminar-H2 LiDAR sensor with a resolution of $1216 \\times 1936$ . It consists of 12K training samples and 40K validation samples. We use that 40K validation samples for evaluation. Also, sparse depth is generated by randomly sampling $20\\%$ ." + ], + "bbox": [ + 511, + 595, + 908, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For quantitative evaluation, we employ the following widely used metrics, Root Mean Squared Error (RMSE) and Mean Squared Error (MAE), to assess performance.", + "bbox": [ + 511, + 689, + 906, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Zero-shot Depth Completion Results", + "text_level": 1, + "bbox": [ + 511, + 757, + 831, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our methodology not only with conventional approaches [51, 71, 84] but also with models [2, 24, 74, 87] that focus on improving zero-shot depth completion performance. Notably, as shown in Tab. 1, our simple finetuning methodology (Ours-FT) outperforms many models designed for generalization across most datasets. This result highlights the effectiveness of using sparse depth as supervision, demonstrating that the depth foundation model can", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9448", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a3b85b66d3ba3b79576813b85eba2e3b8f9be959e37658d4d22a57b25596994.jpg", + "image_caption": [ + "Figure 6. Visualization of our visual prompt, predicted depth, and corresponding error map at each test-time training iteration. As the visual prompt progressively learns from the sparse depth guidance, the predicted depth scale becomes increasingly accurate, which is reflected by the gradual reduction in the error maps." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 496, + 229 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/26e95e01c45afef44ee2f5e41a8cce746a03f618c37b188d89bbe1f4809bea53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 903, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0ce672cd629eefda500f704ce26381a9db97bca8c3ea13d2952a27addbc5fea8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIBims-1NYUv2
MAE↓RMSE↓MAE↓RMSE↓
UniDepthV2-LS [58]0.1130.2210.1230.216
UniDepthV2-VP0.0430.1610.0410.106
Metric3dV2-LS [22]0.1100.2100.1740.275
Metric3dV2-VP0.0390.1470.0690.114
DepthAnytingV2-LS [82]0.1430.2620.2190.326
DepthAnytingV2-VP0.0400.1490.0550.122
ZoeDepth-LS [3]0.6230.8380.1740.275
ZoeDepth-VP0.1760.3410.0490.114
", + "bbox": [ + 91, + 291, + 480, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "be trained to achieve competitive performance. Furthermore, our proposed pixelwise visual prompting method (Ours-VP) surpasses simple fine-tuning in most cases, illustrating that learning visual prompts can be both more efficient and more effective than fine-tuning the entire model. This finding suggests that visual prompting provides a promising alternative to full fine-tuning, offering a favorable balance between computational efficiency and performance. As shown in Fig. 4, the qualitative results indicate that existing SOTA models, which are trained on synthetic data, tend to capture fine-detailed depth. However, their performance is less favorable when examined through error maps, suggesting that accurate scale estimation is a prerequisite for high-quality depth prediction and takes precedence over fine-detail reconstruction.", + "bbox": [ + 88, + 516, + 485, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to sparsity variations. To further assess our model's adaptability to real-world scenarios, we conduct experiments under varying levels of input sparsity. As shown in Tab. 2, our method consistently outperforms SOTA models even as the number of sparse depth points decreases. Notably, our model exhibits greater robustness to sparse depth bias, which is crucial in real-world applications where the number of available depth measurements can vary significantly. As seen in Fig. 5, comparing methods, scale is often misinterpreted when sparse depth is limited, while our approach maintains accuracy, making it highly suitable for", + "bbox": [ + 89, + 734, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b87f007a24d13c8a95e1574abf23502e29c7fa193e8f749c7a0e2fec62db05c3.jpg", + "table_caption": [ + "Table 3. Performance comparison of our visual prompting method across different depth foundation models. Our approach consistently enhances performance, demonstrating its adaptability to various models." + ], + "table_footnote": [], + "table_body": "
MethodMAE ↓RMSE ↓Learnable Parameters
Fine-tuning (Sec. 3.2)0.0490.152100%
Random Patch [1]0.1110.2210.11%
Fixed patch [1]0.0790.1850.11%
Padding patch [1]0.0500.1670.11%
Pixelwise (Sec. 3.3)0.0430.1610.25%
", + "bbox": [ + 513, + 291, + 903, + 386 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison of different visual prompting designs in terms of accuracy and parameter efficiency on dataset Ibims-1 [32]. Our pixelwise visual prompt achieves fine-tuning-level performance while significantly reducing learnable parameters.", + "bbox": [ + 511, + 387, + 906, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "real-world deployment.", + "bbox": [ + 511, + 459, + 669, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of visual prompt in scale alignment. We next analyze the test-time training process of our pixelwise visual prompt. As shown in Fig. 6, the scale accuracy progressively improves with each test-time training step. When the initial scale is generally well-aligned (left example), the prompt refines predictions gradually, leading to a smooth and consistent improvement. In contrast, when large-scale discrepancies exist due to reflective surfaces like mirrors (right example), the prompt first corrects the dominant errors in these regions before adjusting the overall scale. This adaptive behavior effectively mitigates the scale discrepancy problem discussed in Sec. 3.1. After approximately 10 test-time training iterations, the model stabilizes, achieving a well-calibrated depth scale. This highlights the efficiency of our visual prompt learning process, ensuring rapid convergence with minimal computational overhead.", + "bbox": [ + 511, + 477, + 908, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 727, + 666, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Depth foundation models. We further validate our method by applying it to various depth foundation models beyond UniDepthV2 [58], as shown in Tab. 3. Our visual prompting method consistently improves performance across all foundation models. Interestingly, while UniDepthV2 achieves the best results on NYUv2, Metric3dV2 performs better on IBims-1, suggesting that different foundation models may excel with different datasets. This reinforces the importance of model-agnostic test-time training strategies like visual prompting.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9449", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/23eeb6afd25eea115d5c1470899518b0147fa1448ca34ac0bfff0af98ba8ec24.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRequired Memory (MB)Inference Speed (s)
Zero-DC [24]872720.56
Marigold-DC [74]21748121.04
Ours-FT83722.95
Ours-VP39912.93
", + "bbox": [ + 91, + 88, + 482, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparison of test-time training methods measured on the NYUv2 dataset [69].", + "bbox": [ + 89, + 171, + 482, + 200 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompting design. We compare our pixelwise visual prompt against three conventional visual prompting methods: random patches, fixed patches, and padding-based prompts. As shown in Tab. 4, all prompting techniques reduce computational costs by freezing foundation model parameters, but existing prompting methods as proposed by [1] fail to match the performance of fine-tuning. This suggests a trade-off between parameter efficiency and accuracy in conventional prompting techniques. In contrast, our pixelwise visual prompt bridges this gap, achieving fine-tuning-level performance while maintaining efficiency. This is because our method preserves fine-grained spatial adjustments, ensuring that depth refinements occur at the pixel level rather than through global modifications. These results highlight the effectiveness of our approach in maximizing both performance and efficiency.", + "bbox": [ + 89, + 219, + 485, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational efficiency. We compare our approach with existing test-time optimization methods, including Zero-DC [24] and Marigold-DC [74], to highlight its efficiency in memory usage and inference speed. As shown in Tab. 5, our pixelwise visual prompting (Ours-VP) requires significantly less memory (3991 MB) and achieves a fast inference speed of 2.93s, outperforming Zero-DC (20.56s) and Marigold-DC (121.04s). Even our fine-tuning variant (Ours-FT) remains computationally efficient. These results demonstrate that our method avoids the heavy computational cost of full fine-tuning while adapting effectively, making it highly practical for real-world use.", + "bbox": [ + 89, + 465, + 483, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Learning rate & test-time training step. In Fig. 7, we analyze the learning behavior of fine-tuning versus visual prompting. Fine-tuning exhibits unstable learning dynamics, particularly when using a high learning rate, as abrupt parameter shifts distort depth priors and reduce prediction stability in early iterations. Additionally, fine-tuning requires carefully tuned learning rates to prevent performance degradation. In contrast, our visual prompt method is robust to different learning rates, consistently achieving stable and rapid convergence. As seen in Fig. 7, visual prompting requires significantly fewer iterations to reach a reliable depth estimation, making it more efficient and practical for real-time applications.", + "bbox": [ + 89, + 652, + 483, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of structures of learned visual prompt. Lastly, we analyze the characteristics of learned patterns in the visual prompts. As shown in Fig. 8, regions belonging to differ", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/faa3f2119c3b540808cedc0165dc1e866163e9a9b0f13ad0590736983aaa6271.jpg", + "image_caption": [ + "Figure 7. Comparison of fine-tuning and visual prompting methods. The fine-tuning method shows unstable convergence at high learning rates, while the visual prompting method remains stable across different rates and converges faster. The dotted line indicates the baseline using least squares fitting, while visual prompting achieves superior performance with fewer iterations." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 883, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f00c7d6a531235052dec6e6bc5804e6f4c0e76a6e3f9e29a9d7d55c6c4c18b47.jpg", + "image_caption": [ + "(a) RGB image", + "Figure 8. Visualization of the RGB image, visual prompt, and the combined RGB+Visual Prompt. The colored boxes highlight key regions with structural differences and similarities." + ], + "image_footnote": [], + "bbox": [ + 514, + 277, + 620, + 339 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/14d0e7f05248035dcf778ca1c45328a703582c37b415ab456451cdb7c1689865.jpg", + "image_caption": [ + "(b) Visual Prompt" + ], + "image_footnote": [], + "bbox": [ + 620, + 277, + 759, + 338 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c151595d74f290b6b2cfe49350abd2900fe9e2c797e2f0e80aa1a9dadc0a76d.jpg", + "image_caption": [ + "(c) RGB+Visual Prompt" + ], + "image_footnote": [], + "bbox": [ + 759, + 277, + 800, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ent planes, shown in the yellow and green boxes, show that entirely different visual prompt structures are required to accurately adjust distances between the two walls. Conversely, regions on the same plane, shown in cyan and magenta boxes, have relatively similar structures, introducing the same influence to the input. To represent the metric-aware 3D geometry, these visual prompts are learned to adjust the depth to align with sensor measurements, enabling high-fidelity depth estimation while keeping the depth foundation model frozen.", + "bbox": [ + 511, + 424, + 908, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 588, + 633, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study explores how to effectively utilize foundation models for zero-shot depth completion. We find that simply fine-tuning the depth foundation model with sparse depth measurements outperforms existing methods, demonstrating strong adaptability. However, fine-tuning introduces computational overhead and risks knowledge forgetting. To address this, we propose pixelwise visual prompt tuning, which efficiently refines depth predictions without modifying model parameters. Our method achieves fine-tuning-level performance while maintaining efficiency, making it well-suited for zero-shot adaptation across diverse environments.", + "bbox": [ + 511, + 613, + 908, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2024-00338439) and the MSIT(Ministry of Science and ICT), Korea, under the Graduate School of Metaverse Convergence support program(ITP-2025-RS-2024-00418847) supervised by the IITP(Institute for Information & Communications Technology Planning & Evaluation)", + "bbox": [ + 511, + 784, + 910, + 890 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9450", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 2, 4, 7, 8", + "[2] Luca Bartolomei, Matteo Poggi, Andrea Conti, Fabio Tosi, and Stefano Mattoccia. Revisiting depth completion from a stereo matching perspective for cross-domain generalization. In 2024 International Conference on 3D Vision (3DV), 2024. 5, 6", + "[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 2, 3, 7", + "[4] Alexey Bochkovskiy, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 1, 2, 3", + "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Ma teusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. Proceedings of the Neural Information Processing Systems (NeurIPS), 2020. 2", + "[6] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2018. 2", + "[7] Xinjing Cheng, Peng Wang, Chenye Guan, and Ruigang Yang. Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019. 2", + "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 2", + "[10] Yiqun Duan, Xianda Guo, and Zhengbiao Zhu. Diffusion-depth: Diffusion denoising approach for monocular depth estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2", + "[11] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Proceedings of the Neural Information Processing Systems (NeurIPS), 2014. 6", + "[12] Abdelrahman Eldisokey, Michael Felsberg, Karl Holmquist, and Mikael Persson. Uncertainty-aware cnns for depth com" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pletion: Uncertainty from beginning to end. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[13] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[14] David Ferstl, Christian Reinbacher, René Ranftl, Matthias Rüther, and Horst Bischof. Image guided depth upsampling using anisotropic total generalized variation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2013. 1", + "[15] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 1, 6", + "[16] Clément Godard, Oisin Mac Aodha, and Gabriel J. Brostow. Unsupervised monocular depth estimation with left-right consistency. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2", + "[17] Ming Gui, Johannes S. Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Bjorn Ommer. Depthfm: Fast monocular depth estimation with flow matching. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025. 1, 2, 3, 5", + "[18] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6", + "[19] Vitor Campagholo Guizilini, Pavel Tokmakov, Achal Dave, and Rares Ambrus. Grin: Zero-shot metric depth with pixel-level diffusion. arXiv preprint arXiv:2409.09896, 2024. 2", + "[20] Derek Hoiem, Alexei A. Efros, and Martial Hebert. Automatic photo pop-up. Proceedings of ACM SIGGRAPH, 2005. 2", + "[21] Derek Hoiem, Alexei A. Efros, and Martial Hebert. Recovering surface layout from an image. International Journal on Computer Vision (IJCV), 2007. 2", + "[22] Mu Hu, Wei Yin, China. Xiaoyan Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 1, 2, 7", + "[23] Zixuan Huang, Junming Fan, Shenggan Cheng, Shuai Yi, Xiaogang Wang, and Hongsheng Li. Hms-net: Hierarchical multi-scale sparsity-invariant network for sparse depth completion. IEEE Transactions on Image Processing (TIP), 2018. 2", + "[24] Lee Hyoseok, Kyeong Seon Kim, Kwon Byung-Ki, and TaeHyun Oh. Zero-shot depth completion via test-time alignment with affine-invariant depth prior. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025. 2, 3, 5, 6, 8" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9451", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Dmitry Yu. Ignatov, Andrey D. Ignatov, and Radu Timofte. Virtually enriched nyu depth v2 dataset for monocular depth estimation: Do we need artificial augmentation? Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2024. 1, 2", + "[26] Saif Muhammad Imran, Yunfei Long, Xiaoming Liu, and Daniel Morris. Depth coefficients for depth completion. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[27] Maximilian Jaritz, Raoul de Charette, Émilie Wirbel, Xavier Perrotton, and Fawzi Nashashibi. Sparse and dense data with cnns: Depth completion and semantic segmentation. 2018 International Conference on 3D Vision (3DV), 2018. 1", + "[28] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 4", + "[29] Jinyoung Jun, Jae-Han Lee, and Chang-Su Kim. Masked spatial propagation network for sparsity-adaptive depth refinement. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[30] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3", + "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloé Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólár, and Ross B. Girshick. Segment anything. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[32] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision Workshop (ECCVW), 2018. 6, 7", + "[33] Torsten Kröger. Opening the door to new sensor-based robot applications—the reflexes motion libraries. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2011. 1", + "[34] Jason Ku, Ali Harakeh, and Steven L. Waslander. In defense of classical image processing: Fast depth completion on the cpu. 2018 15th Conference on Computer and Robot Vision (CRV), 2018. 1", + "[35] Iro Laina, C. Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In Proceedings of the International Conference on 3D Vision (3DV), 2016. 2, 6", + "[36] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 2", + "[37] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021. 2" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), 2021. 2", + "[39] Ming Liang, Binh Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 1", + "[40] Haotong Lin, Sida Peng, Jingxiao Chen, Songyou Peng, Ji-aming Sun, Minghuan Liu, Hujun Bao, Jiashi Feng, Xiaowei Zhou, and Bingyi Kang. Prompting depth anything for 4k resolution accurate metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3", + "[41] Yuankai Lin, Hua Yang, Tao Cheng, Wending Zhou, and Zhouping Yin. Dyspn: Learning dynamic affinity for image-guided depth completion. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2024. 2", + "[42] Beyang Liu, Stephen Gould, and Daphne Koller. Single image depth estimation from predicted semantic labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 2", + "[43] Sifei Liu, Shalini De Mello, Jinwei Gu, Guangyu Zhong, Ming-Hsuan Yang, and Jan Kautz. Learning affinity via spatial propagation networks. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2017. 2", + "[44] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association of Computational Linguistics (ACL), 2022. 2", + "[45] Xin Liu, Xiaofei Shao, Boqian Wang, Yali Li, and Shengjin Wang. Graphcspn: Geometry-aware depth completion via dynamic gcns. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[46] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 1, 2, 6", + "[47] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2019. 1, 2", + "[48] Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swaprompt: Test-time prompt adaptation for vision-language models. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2023. 2", + "[49] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2022. 2", + "[50] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Finetuning image-conditional diffusion models is easier than you think. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9452", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 5, 6", + "[52] Jin-Hwi Park and Hae-Gon Jeon. A simple yet universal framework for depth completion. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3", + "[53] Jin-Hwi Park, Jaesung Choe, Inhwan Bae, and Hae-Gon Jeon. Learning affinity with hyperbolic representation for spatial propagation. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 2", + "[54] Jin-Hwi Park, Chanhwi Jeong, Junoh Lee, and Hae-Gon Jeon. Depth prompting for sensor-agnostic depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3", + "[55] Nikolay Patakin, Anna Vorontsova, Mikhail Artemyev, and Anton Konushin. Single-stage 3d geometry-preserving depth estimation model training on dataset mixtures with uncalibrated stereo data. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[56] Suraj Patni, Aradhye Agarwal, and Chetan Arora. Ecodepth: Effective conditioning of diffusion models for monocular depth estimation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[57] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2", + "[58] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc van Gool. Undepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. 1, 2, 3, 4, 5, 7", + "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 2", + "[60] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2019. 1, 2, 3", + "[61] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 3", + "[62] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya K. Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Wu, Ross B. Girshick, Piotr Doll'ar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2", + "[63] Laria Reynolds and Kyle McDonell. Prompt programming for large language models: Beyond the few-shot paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, 2021. 2", + "[64] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1", + "[65] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[66] Kieran Saunders, George Vogiatzis, and Luis J Manso. Self-supervised monocular depth estimation: Let's talk about the weather. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[67] Ashutosh Saxena, Sung H. Chung, and A. Ng. Learning depth from single monocular images. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2005. 2", + "[68] Shreyas S. Shivakumar, Ty Nguyen, Steven W. Chen, and Camillo Jose Taylor. Dfusenet: Deep fusion of rgb and sparse depth information for image guided dense depth completion. IEEE Intelligent Transportation Systems Conference (ITSC), 2019. 2", + "[69] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Proceedings of the European Conference on Computer Vision (ECCV), 2012. 6, 8", + "[70] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing (TIP), 2019. 2", + "[71] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5, 6", + "[72] Yifu Tao, Marija Popović, Yiduo Wang, Sundara Tejaswi Digumarti, Nived Chebrolu, and Maurice F. Fallon. 3d lidar reconstruction with probabilistic depth completion for robotic navigation. In Proceedings of the IEEE International Conference on Intelligent Robots and Systems (IROS), 2022. 1", + "[73] Jonas Uhrig, N. Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. International Conference on 3D Vision (3DV), 2017. 2", + "[74] Massimiliano Viola, Kevin Qu, Nando Metzger, Bingxin Ke, Alexander Becker, Konrad Schindler, and Anton Obukhov. Marigold-dc: Zero-shot monocular depth completion with guided diffusion. arXiv preprint arXiv:2412.13389, 2024. 2, 3, 5, 6, 8", + "[75] Yufei Wang, Yuchao Dai, Qi Liu, Peng Yang, Jiadai Sun, and Bo Li. Cu-net: Lidar depth-only completion with coupled u-net. IEEE Robotics and Automation Letters (RA-L), 2022. 2" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9453", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 6", + "[77] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2023. 2", + "[78] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters (RA-L), 2020. 6", + "[79] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Baobei Xu, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[80] Zhiqiang Yan, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Rignet++: Efficient repetitive image guided network for depth completion. arXiv preprint arXiv:2309.00655, 2023. 2", + "[81] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1", + "[82] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2024. 7", + "[83] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2", + "[84] Youming Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and S. Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 6", + "[85] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[86] Jiawen Zhu, Simiao Lai, Xin Chen, D. Wang, and Hutchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[87] Yiming Zuo and Jia Deng. Ogni-dc: Robust depth completion with optimization-guided neural iterations. In Proceedings of the European Conference on Computer Vision (ECCV), 2024, 2, 5, 6" + ], + "bbox": [ + 91, + 90, + 485, + 838 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "9454", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/0a2ddfb1-a84e-4b84-bcf1-12ab1957047a_model.json b/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/0a2ddfb1-a84e-4b84-bcf1-12ab1957047a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..269fd68988c4c6064d396be7394ed10d5ca90c38 --- /dev/null +++ b/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/0a2ddfb1-a84e-4b84-bcf1-12ab1957047a_model.json @@ -0,0 +1,2677 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.131, + 0.799, + 0.154 + ], + "angle": 0, + "content": "Test-Time Prompt Tuning for Zero-Shot Depth Completion" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.18, + 0.764, + 0.218 + ], + "angle": 0, + "content": "Chanhwi Jeong1 Inhwan Bae1 Jin-Hwi Park2* Hae-Gon Jeon3* \n1GIST, 2Chung-Ang University, 3Yonsei University" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.22, + 0.846, + 0.234 + ], + "angle": 0, + "content": "{chanhwij, inhwanbae}@gm.gist.ac.kr, jinhwipark@cau.ac.kr, earboll@yonsei.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.244, + 0.907, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.364, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 1. Comparison of methods for correcting the inaccurate scale of initial depth estimates produced by the depth foundation model using sensor-derived sparse depth with an accurate scale, represented as yellow dots. (a) The depth foundation model generates an initial depth map from a single RGB input. (b) The least squares fitting aligns the initial depth map with the sparse depth [54]. (c) The parameters of the foundation model are fine-tuned under supervision from the sparse depth. (d) Alternatively, the model's parameters are frozen, and a visual prompt is learned to adjust the depth estimates. Please refer to Fig. 3 for depth profiling results and error analysis." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.456, + 0.327, + 0.471 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.478, + 0.486, + 0.811 + ], + "angle": 0, + "content": "Zero-shot depth completion using metric scales remains challenging, primarily due to performance limitations such as domain specificity and sensor characteristics. One recent emerging solution is to integrate monocular depth foundation models into depth completion frameworks, yet such efforts still face issues with suboptimal performance and often require further adaptation to the target task. Surprisingly, we find that a simple test-time training, which finetunes monocular depth foundation models on sparse depth measurements from sensors just as it is, yields reasonable results. However, this test-time training obviously incurs high computational costs and introduces biases towards specific conditions, making it impractical for real-world scenarios. In this paper, we introduce a new approach toward parameter-efficient zero-shot depth completion. Our key idea in this work is to leverage visual prompt tuning, achieving sensor-specific depth scale adaptation without forgetting foundational knowledge. Experimental results on diverse datasets demonstrate that our approach outperforms relevant state-of-the-art methods, showing superior generalization and efficiency. Code is publicly available at https://github.com/ch5374/TestPromptDC" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.827, + 0.222, + 0.843 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.485, + 0.88 + ], + "angle": 0, + "content": "Depth completion, the technique of densifying sparse input measurements from active sensors based on correspond" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.909, + 0.579 + ], + "angle": 0, + "content": "ing RGB images, is crucial in various applications including 3D reconstruction [72], autonomous driving [39] and robotics [33]. Beyond traditional methods that rely on handcrafted features [34] and optimization [14], data-driven approaches, which are mainly based on convolutional neural networks [27, 46, 47] and vision transformers [64, 84], have achieved the best performance with public benchmarks [15, 25]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.909, + 0.747 + ], + "angle": 0, + "content": "However, models trained on a single dataset perform well on the test split of that dataset, but exhibit significant performance degradation in unseen environments. The primary reason lies in the scarcity of data available to use for depth completion. Since depth completion datasets require both dense ground truth depth maps and RGB images obtained from well-calibrated sensor fusion setups, large-scale dataset constructions must collect and label image-depth pairs across environmental variations. Obviously, this makes it difficult to achieve generalizable depth completion, including zero-shot inferences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.902 + ], + "angle": 0, + "content": "One interesting research direction in [40, 52, 54] to mitigate the data scarcity is to integrate foundation models for monocular depth estimation [4, 17, 22, 30, 57, 58, 60, 61, 81-83] trained on huge-scale datasets into depth completion for better generalization. This is quite feasible because foundation models can generate approximate metric or relative depth from a single input image as demonstrated in Fig. 1-(a), actual scene depths can be obtained by fitting sparse depth measurements onto these predictions. Unfortunately, there are still two challenges: (1) they require carefully designed" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.24, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9443" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.258 + ], + "angle": 0, + "content": "additional networks, as the existing foundation models struggle to encode sparse depth information; (2) additional training for the depth completion task is also necessary. These two challenges result in either an increase in computational complexity or the introduction of new biases in the training phase. In this paper, we explore one important question: how can we take full advantage of monocular depth foundation models for depth completion? Our work is the first to answer this question by investigating the most effective foundation models, without requiring additional network design or fine-tuning of the pretrained model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.486, + 0.426 + ], + "angle": 0, + "content": "We initiate our investigation by proposing a simple yet highly effective strategy, which is fine-tuning the foundation models in the inference phase, only using sparse depth measurements as supervision. As an example in Fig. 1-(b), compared to the sparse depth fitting onto dense depths [40, 52, 54], fine-tuning the foundation models preserves the original depth while allowing the scale to adapt locally (Fig. 1-(c)). Of course, updating all parameters requires significant computing resources and poses the risk of dataset bias, potentially causing models to lose their own knowledge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.486, + 0.668 + ], + "angle": 0, + "content": "As a more efficient alternative, we present a test-time prompt tuning that adapts foundation models at inference time without modifying its parameters (Fig. 1-(d)). We introduce a small set of learnable parameters that guide their prediction while keeping the foundation model frozen. While existing visual prompt [1, 28] uses fixed-size prompts positioned at specific locations (e.g., fixed positions, random placements, or padding) for image classification, we introduce a pixelwise visual prompt specifically designed for depth completion. Compared to previous works that operate at a global or patch level, our pixelwise visual prompt is initialized to zero and added to input images, allowing us to leverage the inherent spatial structures of depth maps. We validate our visual prompt tuning across diverse benchmark datasets and environments, achieving superior zero-shot performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.228, + 0.701 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.353, + 0.728 + ], + "angle": 0, + "content": "2.1. Monocular Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Traditional monocular depth estimation depends on handcrafted visual cues [20, 67] and scene-specific assumptions [21, 42]. With the advent of deep learning, subsequent methods [16, 35, 36] using data-driven representations have significantly outperformed the traditional ones reducing the reliance on manually designed features and rigid scene-specific assumptions. Recent works have explored techniques such as data mixing [55, 60, 61] and augmentation [25, 66] to achieve zero-shot generality. Moreover, some works in [10, 17, 30, 50, 56, 85] exploit the generative priors of pretrained diffusion models [65]. However, inferring" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.259 + ], + "angle": 0, + "content": "metric-scale depth from single images still remains challenging. To address this, metric-scale depth estimation from monocular images has evolved by incorporating global depth distributions [19], scene-aware conditioning [3] and camera intrinsics [4, 22, 22, 57, 58, 83]. While earlier methods rely on known camera parameters [22, 83], recent approaches estimate focal length [4, 57, 58] directly from learning features, enhancing generalization across diverse environments without any explicit camera metadata. Nevertheless, these methods obviously lag behind approaches that use sensor measurements like LiDAR in terms of accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.275, + 0.693, + 0.291 + ], + "angle": 0, + "content": "2.2. Depth Completion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.909, + 0.633 + ], + "angle": 0, + "content": "Due to the inherent issues with monocular depth estimation, research on unguided depth completion [12, 23, 73, 75], transforming sparse depth data from sensors into dense depth maps without guidance RGB images, has been studied. However, most mainstream depth completion is to use RGB-depth pairs to directly leverage scene contexts such as edges and textures [26, 46, 47, 68, 70, 79, 80, 84, 87]. Furthermore, constructing pixelwise affinity [6, 7, 41, 43, 45, 51, 53] from corresponding RGB images has proven to be effective as it helps preserve depth discontinuity when propagating sparse depth. Despite these advancements, such models often struggle to show the same performance in unseen domains. To mitigate this, recent approaches have leveraged monocular depth foundation models for depth completion. For instance, works in [29, 54] introduces interesting frameworks that exploit features and predict depth from foundation models to perform adaptive depth refinement with sparse depth inputs. Additionally, diffusion-based depth completion methods [24, 74] have adopted the test-time optimization of latent representations, guided by input sparse depths. However, these approaches often require training on additional data [29, 54] or suffer from extremely slow inference speeds [24, 74]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.65, + 0.815, + 0.667 + ], + "angle": 0, + "content": "2.3. Prompting for Foundation Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Foundation models [5, 9, 31, 59, 62, 65] exhibit remarkable zero-shot performance across a wide range of tasks as well as depth perception. In natural language processing, prompting schemes have been introduced to leverage these foundation models effectively for downstream tasks [5, 37, 38, 44, 63, 77]. To apply such prompting schemes to the computer vision domain, works in [28, 86] have introduced visual prompt tuning, which incorporates trainable visual tokens into models for specific downstream tasks. Meanwhile, one interesting work in [1] presents learnable padding prompts appended to input images. More recently, visual prompt tuning has also been extended to test-time adaptation scenarios [13, 48, 49] for better adaptation to unseen domains. Inspired by them [1, 49], we propose a test-time prompt tuning strategy for depth completion." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9444" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.283, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.167, + 0.272, + 0.178 + ], + "angle": 0, + "content": "(a) Supervised training models" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.096, + 0.508, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.165, + 0.497, + 0.177 + ], + "angle": 0, + "content": "(b) Test-time optimization models" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.096, + 0.712, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.165, + 0.702, + 0.178 + ], + "angle": 0, + "content": "(c) Our fine-tuning approach" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.096, + 0.905, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.165, + 0.905, + 0.178 + ], + "angle": 0, + "content": "(d) Our visual prompting approach" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.18, + 0.907, + 0.249 + ], + "angle": 0, + "content": "Figure 2. Comparison of the previous foundation model-based depth completion approaches and our approach. (a) Existing methods [40, 52, 54] design additional networks to integrate sensor-based depth measurements into the foundation model. (b) Recent approaches [24, 74] optimize latent representations in diffusion-based foundation models [17, 30] without dense GT but are computationally expensive. (c) Our fine-tuning method adapts the model with sparse depth supervision, and (d) our visual prompting approach enables efficient adaptation while keeping the foundation model frozen." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.267, + 0.226, + 0.284 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.485, + 0.536 + ], + "angle": 0, + "content": "In this section, we present our approach for parameter-efficient zero-shot depth completion through test-time training. In Sec. 3.1, we discuss the role of depth foundation models in depth completion, emphasizing their generalization capabilities while identifying key challenges. Next, in Sec. 3.2, we explore a fine-tuning strategy to adapt the foundation model for sparse depth measurements. This approach achieves impressive performance without relying on additional network design or prompt engineering; however, it introduces computational overhead and poses generalization challenges. To address these limitations, in Sec. 3.3, we propose a novel pixel-wise visual prompting method, enabling efficient adaptation without any parameter modification of the foundation model. Finally, in Sec. 3.4, we describe implementation details, including optimization strategies, dataset preprocessing and experimental settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.55, + 0.476, + 0.566 + ], + "angle": 0, + "content": "3.1. Rationale: Usage of Depth Foundation Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.664 + ], + "angle": 0, + "content": "Recent works have attempted to leverage monocular depth foundation models for depth completion. These models offer strong generalization capabilities and are considered to be attractive solutions for better zero-shot performance. However, incorporating them into depth completion frameworks faces several technical challenges, as described below:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.484, + 0.79 + ], + "angle": 0, + "content": "Incompatibility with sparse depth input handling. Since monocular depth foundation models are exclusively designed to infer depth from single RGB images, they do not have mechanisms to take sparse depth measurements as guidance. As a result, the existing methods require additional components to process sparse depths and to integrate them into the models, increasing computational complexity as shown in Fig. 2-(a)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Scale mismatch between prediction and sensor depth. Current depth foundation models are able to estimate approximate metric-scale depths [3, 4, 58] as well as relative-scale depth [17, 30, 60, 61] with high-fidelity depth discontinuities. Unfortunately, the depth foundation models for metric-scale inference exhibit systematic scale mismatches when compared to sensor-derived depths. While least squares scale" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.265, + 0.645, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.376, + 0.643, + 0.388 + ], + "angle": 0, + "content": "(a) Least squares fitting" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.265, + 0.776, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.376, + 0.756, + 0.388 + ], + "angle": 0, + "content": "(b) Fine-tuning" + }, + { + "type": "image", + "bbox": [ + 0.778, + 0.265, + 0.906, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.376, + 0.899, + 0.388 + ], + "angle": 0, + "content": "(c) Visual Prompting" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.39, + 0.906, + 0.459 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison of different strategies on the depth map from the data used in Fig. 1. The top row shows the predicted depth maps, while the bottom row presents the depth profiles along the indicated horizontal line. The red \\(\\times\\) marks indicate the sparse depth measurements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.907, + 0.576 + ], + "angle": 0, + "content": "fitting in [54] can globally align depth scales, it fails in regions where depth varies significantly. For instance, as shown in Fig. 3, the least squares fitting locally adjusts the correct depth scale at areas near initial depth seeds (e.g., pixel range 100–150), but fails it in other regions (e.g., pixel range 50–100)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.808 + ], + "angle": 0, + "content": "Limitations of existing adaptation strategies. Previous approaches require ground truth (GT) annotations for adaptation, which are costly and labor-intensive (Fig. 2-(a)). For example, works in [40, 54] rely on large-scale datasets, while UniDC [52] adopts a few-shot strategy within the target domain. However, generating new GT labels to cover all scenarios in-the-wild, is impractical, necessitating a more efficient alternative. One promising solution is a test-time optimization, which enables adaptation to new environments without any additional GT annotation (Fig. 2-(b)). Recent contemporary works in [24, 74] have introduced similar strategies by optimizing latent representations in diffusion-based models (Fig. 2-(b)). However, these approaches come with noticeable limitations, including high computational costs and slow inference speeds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To address these limitations, we propose fine-tuning depth foundation models in the inference phase, directly using sparse depth measurements as supervision without introducing additional network components (Fig. 2-(c)). This eliminates the dependency on extensive GT data and enables real-time adaptation to new environments." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9445" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.485, + 0.108 + ], + "angle": 0, + "content": "3.2. Finetuning monocular depth foundation model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.176 + ], + "angle": 0, + "content": "Let \\(F_{\\theta}\\) denote a depth foundation model, parametrized by \\(\\theta\\), and \\(I \\in \\mathbb{R}^{3 \\times H \\times W}\\) is an RGB image. Typically, \\(F_{\\theta}\\) takes only the \\(I\\) as input and generates a dense depth map \\(D \\in \\mathbb{R}^{1 \\times H \\times W}\\) as output, which can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.181, + 0.484, + 0.198 + ], + "angle": 0, + "content": "\\[\nF _ {\\theta} (I) = D. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.202, + 0.485, + 0.279 + ], + "angle": 0, + "content": "Since \\( D \\) is derived solely from a single image, it inherently suffers from scale ambiguity. To mitigate this issue, a sparse depth \\( S \\), obtained from an active sensor with accurate scales, is employed as supervision for \\( F_{\\theta} \\). This approach allows the model to alleviate the depth scale ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.485, + 0.311 + ], + "angle": 0, + "content": "First, to use only valid sparse depth values as supervision, we define a validity mask \\(M\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.315, + 0.484, + 0.356 + ], + "angle": 0, + "content": "\\[\nM _ {i, j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} S _ {i, j} \\text {i s v a l i d} \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.411, + 0.376 + ], + "angle": 0, + "content": "where \\((i,j)\\) is the pixel coordinate in the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.485, + 0.437 + ], + "angle": 0, + "content": "With this mask, we initially adjust the scale of the foundation depth by roughly aligning it using the valid sparse depth measurements in the least squares fitting process, as formulated below:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.443, + 0.484, + 0.468 + ], + "angle": 0, + "content": "\\[\ns _ {0}, s _ {1} = \\min _ {\\left(s _ {0}, s _ {1}\\right)} \\| M \\odot \\left(\\left(s _ {0} D + s _ {1}\\right) - S\\right) \\| _ {F}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.564 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes an element-wise multiplication. \\(s_0\\) and \\(s_1\\) mean scale and shift values, respectively, and \\(\\| \\cdot \\| _F\\) indicates the Frobenius norm. Next, we optimize the model's parameters \\(\\theta\\) using only the valid sparse depth measurements as supervision. The optimal parameters, denoted by \\(\\theta^{*}\\), are determined by minimizing the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.567, + 0.484, + 0.601 + ], + "angle": 0, + "content": "\\[\n\\theta^ {*} = \\arg \\min _ {\\theta} \\sum_ {i, j} M _ {i, j} \\odot | s _ {0} F _ {\\theta} (I) _ {i, j} + s _ {1} - S _ {i, j} | ^ {p}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.486, + 0.665 + ], + "angle": 0, + "content": "where \\( p = 1 \\) and \\( p = 2 \\) correspond to the \\( L_{1} \\) and \\( L_{2} \\) losses, respectively and the final loss is the sum of both. To isolate the effects of fine-tuning itself, we do not incorporate specialized loss functions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.315, + 0.696 + ], + "angle": 0, + "content": "3.3. Pixelwise Visual Prompt" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.486, + 0.809 + ], + "angle": 0, + "content": "To address the computational overhead and potential knowledge degradation of the fine-tuning of the foundation model, we propose a pixelwise visual prompt, a novel test-time adaptation method designed specifically for depth completion (Fig. 2-(d)). Unlike conventional fine-tuning, our approach enables efficient adaptation while keeping the model's parameters frozen." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Given a frozen pretrained foundation model \\(F_{\\theta}\\) and a sample consisting of a pair of image and sparse depth measurement \\((I,S)\\), we aim to learn a pixelwise visual prompt \\(V_{\\phi} \\in \\mathbb{R}^{3 \\times H \\times W}\\), parameterized by \\(\\phi\\). The prompt is added directly to the input image, forming a prompted image \\(I + V_{\\phi}\\) and we optimize it by minimizing the depth prediction error" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.726, + 0.107 + ], + "angle": 0, + "content": "in the training phase as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.112, + 0.907, + 0.145 + ], + "angle": 0, + "content": "\\[\n\\phi^ {*} = \\arg \\min _ {\\phi} \\sum_ {i, j} M _ {i, j} \\odot \\left| F _ {\\theta} \\left(I + V _ {\\phi}\\right) _ {i, j} - S _ {i, j} \\right| ^ {p}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.149, + 0.907, + 0.196 + ], + "angle": 0, + "content": "Here, only the prompt parameters \\(\\phi\\) are optimized, while \\(\\theta\\) remains frozen. Once optimized, the visual prompt \\(V_{\\phi^*}\\) is applied to the image as below:" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.2, + 0.907, + 0.218 + ], + "angle": 0, + "content": "\\[\nF _ {\\theta} \\left(I + V _ {\\phi^ {*}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.222, + 0.909, + 0.359 + ], + "angle": 0, + "content": "In addition to achieving superior performance compared to existing methods, this approach also enhances usability, providing an additional advantage. Since the pixelwise visual prompt is initialized to zero and added to the image, it ensures a minimal level of depth prediction performance even without prompt tuning. Additionally, users can dynamically adjust the time allocated for prompt tuning based on their available resources and the required inference speed, allowing for stable and consistent performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.909, + 0.573 + ], + "angle": 0, + "content": "Difference with existing visual prompt methods. While visual prompt tunings [1, 28] have been previously explored to adapt high-level representations using fixed-size prompts at predefined locations, such approaches lack the spatial granularity needed for precise depth refinement. These methods often focus on global or patch-level modifications, making them less effective for tasks requiring fine-grained adjustments, such as depth completion. In contrast, our pixelwise visual prompt is designed to adapt at the individual pixel level, enabling localized corrections that align depth predictions more accurately with sparse sensor measurements. Our approach not only enhances adaptation flexibility but also maintains the strong generalization capabilities of foundation models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.733, + 0.599 + ], + "angle": 0, + "content": "3.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.909, + 0.773 + ], + "angle": 0, + "content": "Our implementation is based on PyTorch and runs on a single NVIDIA RTX 3090 GPU. We use the AdamW optimizer with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\). The learning rate is set to 1e-6 for fine-tuning and 1e-3 for visual prompting. The training process consists of 80 steps for fine-tuning and 50 steps for visual prompting. For all depth foundation models, we use their official pretrained weights and specifically the v2-old version for UniDepthV2 [58], while other models are evaluated in our ablation study. For models that do not support the original image size, we first resize the images while preserving their aspect ratio." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.64, + 0.802 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this section, we begin by outlining the experimental setup in Sec. 4.1, including the evaluation protocols, and introduce the datasets. We then compare our method with state-of-the-art (SOTA) methods for zero-shot depth completion tasks, in Sec. 4.2. Lastly, in Sec. 4.3, we present an ablation study to analyze the impact of our methods." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "9446" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.909, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.474, + 0.899, + 0.488 + ], + "angle": 0, + "content": "Figure 4. Qualitative result and depth error map of various depth completion methods on the IBims-1, VOID 1500, and DDAD datasets." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.495, + 0.903, + 0.668 + ], + "angle": 0, + "content": "
MethodIBims-1VOIDNYUv2KITTIDDAD
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
NLSPN [51]0.0490.1910.4271.3530.4400.7161.3352.0762.4989.231
CompletionFormer [84]0.0580.2060.6421.6840.1860.3740.9521.9352.5189.471
VPP4DC [2]0.0620.2280.1660.6060.0770.2470.4131.6091.3446.781
BP-Net [71]0.0780.2890.2700.742----2.2708.344
OGNI-DC [87]0.0590.1860.1750.593----1.8676.876
Zero-DC [24]0.0780.2030.1870.6070.0690.1580.3971.4131.6817.122
Marigold-DC [74]0.0450.1660.1520.5510.0480.1240.4341.4652.3646.449
Ours-LS0.1130.2210.1870.5710.1230.2161.2052.8015.67611.97
Ours-FT (Sec. 3.2)0.0490.1520.1380.5140.0480.1070.3831.3291.7755.881
Ours-VP (Sec. 3.3)0.0430.1610.1320.5080.0410.1060.3431.3331.4825.811
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.67, + 0.907, + 0.713 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with SOTA methods on five zero-shot benchmarks. Most methods used values reported in [74], while the results for Zero-DC [24] are evaluated based on the DepthFM model [17]. For the same reason as in [74], we do not include results for BP-Net [71] and OGNI-DC [87] on NYUv2 and KITTI. Bold: Best, Underline: Second best." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.283, + 0.745 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Our test-time training strategy is closely aligned with recent works [24, 74], which focus on optimizing the model at inference time. Following their methodology, we adapt the model using only sparse depth measurements during test-time, ensuring a fully zero-shot approach while effectively refining depth scale estimation. To maintain a fair evaluation, we strictly exclude the dataset (e.g., ScanNet [8]) used for training the depth foundation model [58] from the test. Furthermore, in our fine-tuning experiments, we prevent the model from accumulating dataset-specific biases by fine-tuning on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.729, + 0.907, + 0.803 + ], + "angle": 0, + "content": "a single sample and resetting it to its original parameters before proceeding to the next evaluation. This ensures that our approach remains consistent with the principles of test-time optimization and maintains its generalization ability across diverse environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We evaluate our proposed method in a zero-shot setting using five real-world datasets that are not seen during the training of the base model [58]. These datasets encompass both indoor and outdoor environments, varying image resolutions, and different levels of sparse depth density, ensuring a comprehensive assessment of generalization ability." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9447" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.088, + 0.878, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.393, + 0.908, + 0.424 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison under varying levels of depth sparsity across NYUv2 and KITTI. Despite significantly reduced depth measurements, our approach recovers high-quality dense depth maps, demonstrating robustness to input sparsity in the depth completion." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.435, + 0.907, + 0.515 + ], + "angle": 0, + "content": "
MethodNYUv2 (200 samples)NYUv2 (100 samples)NYUv2 (50 samples)NYUv2 (5 samples)KITTI (32 channels)KITTI (16 channels)KITTI (8 channels)KITTI (4 channels)
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
Zero-DC [24]0.0960.1970.1240.2340.1610.2820.4470.6230.5401.9160.6942.3231.0082.9041.6874.184
Marigold-DC [74]0.0730.1750.0900.2030.1120.2360.3580.5410.8752.1930.9872.6121.1863.1761.6694.240
Ours-VP0.0530.1330.0650.1520.0800.1800.1380.2470.3961.5400.4751.8510.5672.0930.8122.644
" + }, + { + "type": "table_caption", + "bbox": [ + 0.229, + 0.517, + 0.766, + 0.532 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on varying sparsity. Bold: Best, Underline: Second best." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.547, + 0.483, + 0.625 + ], + "angle": 0, + "content": "(1) IBims-1 [32] is an indoor dataset consisting of 100 RGB-D pairs with a resolution of \\(480 \\times 640\\). We use the entire dataset for testing. Since sparse depth data is not provided, we follow [74] to generate a depth valid mask and construct sparse depth by randomly sampling 1,000 depth points." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.63, + 0.483, + 0.707 + ], + "angle": 0, + "content": "(2) VOID [78] contains RGB-D frames from 56 sequences, with 48 designated for training and 8 for testing. The official test set comprises 800 frames, and sparse depth maps are obtained using the VIO system. Following [74], we evaluate our method on this test set using 1,500 sparse depth points." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.547, + 0.483, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.712, + 0.483, + 0.835 + ], + "angle": 0, + "content": "(1) NYUv2 [69] comprises 464 scenes recorded using a Microsoft Kinect sensor. The official split is divided into 249 training and 215 test scenes. We use the official test set of 654 images for evaluation. Following a commonly used setup [11, 35, 46], we downsample and center-crop the images to \\(228 \\times 304\\) before conducting experiments. Also, sparse depth is obtained by randomly selecting 500 samples from the depth map." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.487, + 0.903 + ], + "angle": 0, + "content": "(4) KITTIDC [15] consists of over 90K RGB images paired with LiDAR data. Since the depth obtained from LiDAR is inherently sparse, the ground truth depth is generated by aggregating multiple LiDAR frames. For our experiments," + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.712, + 0.487, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.592 + ], + "angle": 0, + "content": "we use the 1,000 pairs provided for validation. To reduce noise in the input sparse depth, we apply the filtering method used in [76]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.687 + ], + "angle": 0, + "content": "(5) DDAD [18] is an outdoor RGB-D driving dataset collected using a Luminar-H2 LiDAR sensor with a resolution of \\(1216 \\times 1936\\). It consists of 12K training samples and 40K validation samples. We use that 40K validation samples for evaluation. Also, sparse depth is generated by randomly sampling \\(20\\%\\)." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.908, + 0.751 + ], + "angle": 0, + "content": "Evaluation Metrics. For quantitative evaluation, we employ the following widely used metrics, Root Mean Squared Error (RMSE) and Mean Squared Error (MAE), to assess performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.833, + 0.775 + ], + "angle": 0, + "content": "4.2. Zero-shot Depth Completion Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We compare our methodology not only with conventional approaches [51, 71, 84] but also with models [2, 24, 74, 87] that focus on improving zero-shot depth completion performance. Notably, as shown in Tab. 1, our simple finetuning methodology (Ours-FT) outperforms many models designed for generalization across most datasets. This result highlights the effectiveness of using sparse depth as supervision, demonstrating that the depth foundation model can" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9448" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.497, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.089, + 0.905, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.234, + 0.908, + 0.276 + ], + "angle": 0, + "content": "Figure 6. Visualization of our visual prompt, predicted depth, and corresponding error map at each test-time training iteration. As the visual prompt progressively learns from the sparse depth guidance, the predicted depth scale becomes increasingly accurate, which is reflected by the gradual reduction in the error maps." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.437 + ], + "angle": 0, + "content": "
MethodIBims-1NYUv2
MAE↓RMSE↓MAE↓RMSE↓
UniDepthV2-LS [58]0.1130.2210.1230.216
UniDepthV2-VP0.0430.1610.0410.106
Metric3dV2-LS [22]0.1100.2100.1740.275
Metric3dV2-VP0.0390.1470.0690.114
DepthAnytingV2-LS [82]0.1430.2620.2190.326
DepthAnytingV2-VP0.0400.1490.0550.122
ZoeDepth-LS [3]0.6230.8380.1740.275
ZoeDepth-VP0.1760.3410.0490.114
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.439, + 0.485, + 0.493 + ], + "angle": 0, + "content": "Table 3. Performance comparison of our visual prompting method across different depth foundation models. Our approach consistently enhances performance, demonstrating its adaptability to various models." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.517, + 0.486, + 0.729 + ], + "angle": 0, + "content": "be trained to achieve competitive performance. Furthermore, our proposed pixelwise visual prompting method (Ours-VP) surpasses simple fine-tuning in most cases, illustrating that learning visual prompts can be both more efficient and more effective than fine-tuning the entire model. This finding suggests that visual prompting provides a promising alternative to full fine-tuning, offering a favorable balance between computational efficiency and performance. As shown in Fig. 4, the qualitative results indicate that existing SOTA models, which are trained on synthetic data, tend to capture fine-detailed depth. However, their performance is less favorable when examined through error maps, suggesting that accurate scale estimation is a prerequisite for high-quality depth prediction and takes precedence over fine-detail reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Robustness to sparsity variations. To further assess our model's adaptability to real-world scenarios, we conduct experiments under varying levels of input sparsity. As shown in Tab. 2, our method consistently outperforms SOTA models even as the number of sparse depth points decreases. Notably, our model exhibits greater robustness to sparse depth bias, which is crucial in real-world applications where the number of available depth measurements can vary significantly. As seen in Fig. 5, comparing methods, scale is often misinterpreted when sparse depth is limited, while our approach maintains accuracy, making it highly suitable for" + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.292, + 0.905, + 0.387 + ], + "angle": 0, + "content": "
MethodMAE ↓RMSE ↓Learnable Parameters
Fine-tuning (Sec. 3.2)0.0490.152100%
Random Patch [1]0.1110.2210.11%
Fixed patch [1]0.0790.1850.11%
Padding patch [1]0.0500.1670.11%
Pixelwise (Sec. 3.3)0.0430.1610.25%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.445 + ], + "angle": 0, + "content": "Table 4. Comparison of different visual prompting designs in terms of accuracy and parameter efficiency on dataset Ibims-1 [32]. Our pixelwise visual prompt achieves fine-tuning-level performance while significantly reducing learnable parameters." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.46, + 0.671, + 0.474 + ], + "angle": 0, + "content": "real-world deployment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.909, + 0.72 + ], + "angle": 0, + "content": "Effectiveness of visual prompt in scale alignment. We next analyze the test-time training process of our pixelwise visual prompt. As shown in Fig. 6, the scale accuracy progressively improves with each test-time training step. When the initial scale is generally well-aligned (left example), the prompt refines predictions gradually, leading to a smooth and consistent improvement. In contrast, when large-scale discrepancies exist due to reflective surfaces like mirrors (right example), the prompt first corrects the dominant errors in these regions before adjusting the overall scale. This adaptive behavior effectively mitigates the scale discrepancy problem discussed in Sec. 3.1. After approximately 10 test-time training iterations, the model stabilizes, achieving a well-calibrated depth scale. This highlights the efficiency of our visual prompt learning process, ensuring rapid convergence with minimal computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.667, + 0.744 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Depth foundation models. We further validate our method by applying it to various depth foundation models beyond UniDepthV2 [58], as shown in Tab. 3. Our visual prompting method consistently improves performance across all foundation models. Interestingly, while UniDepthV2 achieves the best results on NYUv2, Metric3dV2 performs better on IBims-1, suggesting that different foundation models may excel with different datasets. This reinforces the importance of model-agnostic test-time training strategies like visual prompting." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9449" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.483, + 0.171 + ], + "angle": 0, + "content": "
MethodRequired Memory (MB)Inference Speed (s)
Zero-DC [24]872720.56
Marigold-DC [74]21748121.04
Ours-FT83722.95
Ours-VP39912.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.172, + 0.483, + 0.201 + ], + "angle": 0, + "content": "Table 5. Comparison of test-time training methods measured on the NYUv2 dataset [69]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.22, + 0.486, + 0.462 + ], + "angle": 0, + "content": "Prompting design. We compare our pixelwise visual prompt against three conventional visual prompting methods: random patches, fixed patches, and padding-based prompts. As shown in Tab. 4, all prompting techniques reduce computational costs by freezing foundation model parameters, but existing prompting methods as proposed by [1] fail to match the performance of fine-tuning. This suggests a trade-off between parameter efficiency and accuracy in conventional prompting techniques. In contrast, our pixelwise visual prompt bridges this gap, achieving fine-tuning-level performance while maintaining efficiency. This is because our method preserves fine-grained spatial adjustments, ensuring that depth refinements occur at the pixel level rather than through global modifications. These results highlight the effectiveness of our approach in maximizing both performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.648 + ], + "angle": 0, + "content": "Computational efficiency. We compare our approach with existing test-time optimization methods, including Zero-DC [24] and Marigold-DC [74], to highlight its efficiency in memory usage and inference speed. As shown in Tab. 5, our pixelwise visual prompting (Ours-VP) requires significantly less memory (3991 MB) and achieves a fast inference speed of 2.93s, outperforming Zero-DC (20.56s) and Marigold-DC (121.04s). Even our fine-tuning variant (Ours-FT) remains computationally efficient. These results demonstrate that our method avoids the heavy computational cost of full fine-tuning while adapting effectively, making it highly practical for real-world use." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.484, + 0.851 + ], + "angle": 0, + "content": "Learning rate & test-time training step. In Fig. 7, we analyze the learning behavior of fine-tuning versus visual prompting. Fine-tuning exhibits unstable learning dynamics, particularly when using a high learning rate, as abrupt parameter shifts distort depth priors and reduce prediction stability in early iterations. Additionally, fine-tuning requires carefully tuned learning rates to prevent performance degradation. In contrast, our visual prompt method is robust to different learning rates, consistently achieving stable and rapid convergence. As seen in Fig. 7, visual prompting requires significantly fewer iterations to reach a reliable depth estimation, making it more efficient and practical for real-time applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Analysis of structures of learned visual prompt. Lastly, we analyze the characteristics of learned patterns in the visual prompts. As shown in Fig. 8, regions belonging to differ" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.089, + 0.885, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.193, + 0.907, + 0.275 + ], + "angle": 0, + "content": "Figure 7. Comparison of fine-tuning and visual prompting methods. The fine-tuning method shows unstable convergence at high learning rates, while the visual prompting method remains stable across different rates and converges faster. The dotted line indicates the baseline using least squares fitting, while visual prompting achieves superior performance with fewer iterations." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.279, + 0.621, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.34, + 0.612, + 0.351 + ], + "angle": 0, + "content": "(a) RGB image" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.279, + 0.76, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.339, + 0.758, + 0.35 + ], + "angle": 0, + "content": "(b) Visual Prompt" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.279, + 0.801, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.339, + 0.905, + 0.35 + ], + "angle": 0, + "content": "(c) RGB+Visual Prompt" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.356, + 0.907, + 0.398 + ], + "angle": 0, + "content": "Figure 8. Visualization of the RGB image, visual prompt, and the combined RGB+Visual Prompt. The colored boxes highlight key regions with structural differences and similarities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.909, + 0.575 + ], + "angle": 0, + "content": "ent planes, shown in the yellow and green boxes, show that entirely different visual prompt structures are required to accurately adjust distances between the two walls. Conversely, regions on the same plane, shown in cyan and magenta boxes, have relatively similar structures, introducing the same influence to the input. To represent the metric-aware 3D geometry, these visual prompts are learned to adjust the depth to align with sensor measurements, enabling high-fidelity depth estimation while keeping the depth foundation model frozen." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.634, + 0.605 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.78 + ], + "angle": 0, + "content": "This study explores how to effectively utilize foundation models for zero-shot depth completion. We find that simply fine-tuning the depth foundation model with sparse depth measurements outperforms existing methods, demonstrating strong adaptability. However, fine-tuning introduces computational overhead and risks knowledge forgetting. To address this, we propose pixelwise visual prompt tuning, which efficiently refines depth predictions without modifying model parameters. Our method achieves fine-tuning-level performance while maintaining efficiency, making it well-suited for zero-shot adaptation across diverse environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.785, + 0.911, + 0.891 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2024-00338439) and the MSIT(Ministry of Science and ICT), Korea, under the Graduate School of Metaverse Convergence support program(ITP-2025-RS-2024-00418847) supervised by the IITP(Institute for Information & Communications Technology Planning & Evaluation)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9450" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.169 + ], + "angle": 0, + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Luca Bartolomei, Matteo Poggi, Andrea Conti, Fabio Tosi, and Stefano Mattoccia. Revisiting depth completion from a stereo matching perspective for cross-domain generalization. In 2024 International Conference on 3D Vision (3DV), 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.484, + 0.297 + ], + "angle": 0, + "content": "[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.299, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[4] Alexey Bochkovskiy, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.484, + 0.507 + ], + "angle": 0, + "content": "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Ma teusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. Proceedings of the Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.484, + 0.563 + ], + "angle": 0, + "content": "[6] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[7] Xinjing Cheng, Peng Wang, Chenye Guan, and Ruigang Yang. Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.76 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[10] Yiqun Duan, Xianda Guo, and Zhengbiao Zhu. Diffusion-depth: Diffusion denoising approach for monocular depth estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[11] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Proceedings of the Neural Information Processing Systems (NeurIPS), 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Abdelrahman Eldisokey, Michael Felsberg, Karl Holmquist, and Mikael Persson. Uncertainty-aware cnns for depth com" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "pletion: Uncertainty from beginning to end. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[13] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[14] David Ferstl, Christian Reinbacher, René Ranftl, Matthias Rüther, and Horst Bischof. Image guided depth upsampling using anisotropic total generalized variation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[15] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.907, + 0.39 + ], + "angle": 0, + "content": "[16] Clément Godard, Oisin Mac Aodha, and Gabriel J. Brostow. Unsupervised monocular depth estimation with left-right consistency. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.906, + 0.474 + ], + "angle": 0, + "content": "[17] Ming Gui, Johannes S. Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Bjorn Ommer. Depthfm: Fast monocular depth estimation with flow matching. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.907, + 0.544 + ], + "angle": 0, + "content": "[18] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.907, + 0.588 + ], + "angle": 0, + "content": "[19] Vitor Campagholo Guizilini, Pavel Tokmakov, Achal Dave, and Rares Ambrus. Grin: Zero-shot metric depth with pixel-level diffusion. arXiv preprint arXiv:2409.09896, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.907, + 0.629 + ], + "angle": 0, + "content": "[20] Derek Hoiem, Alexei A. Efros, and Martial Hebert. Automatic photo pop-up. Proceedings of ACM SIGGRAPH, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.907, + 0.674 + ], + "angle": 0, + "content": "[21] Derek Hoiem, Alexei A. Efros, and Martial Hebert. Recovering surface layout from an image. International Journal on Computer Vision (IJCV), 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.907, + 0.759 + ], + "angle": 0, + "content": "[22] Mu Hu, Wei Yin, China. Xiaoyan Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.907, + 0.828 + ], + "angle": 0, + "content": "[23] Zixuan Huang, Junming Fan, Shenggan Cheng, Shuai Yi, Xiaogang Wang, and Hongsheng Li. Hms-net: Hierarchical multi-scale sparsity-invariant network for sparse depth completion. IEEE Transactions on Image Processing (TIP), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.907, + 0.899 + ], + "angle": 0, + "content": "[24] Lee Hyoseok, Kyeong Seon Kim, Kwon Byung-Ki, and TaeHyun Oh. Zero-shot depth completion via test-time alignment with affine-invariant depth prior. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025. 2, 3, 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9451" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.161 + ], + "angle": 0, + "content": "[25] Dmitry Yu. Ignatov, Andrey D. Ignatov, and Radu Timofte. Virtually enriched nyu depth v2 dataset for monocular depth estimation: Do we need artificial augmentation? Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[26] Saif Muhammad Imran, Yunfei Long, Xiaoming Liu, and Daniel Morris. Depth coefficients for depth completion. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[27] Maximilian Jaritz, Raoul de Charette, Émilie Wirbel, Xavier Perrotton, and Fawzi Nashashibi. Sparse and dense data with cnns: Depth completion and semantic segmentation. 2018 International Conference on 3D Vision (3DV), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[28] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.389 + ], + "angle": 0, + "content": "[29] Jinyoung Jun, Jae-Han Lee, and Chang-Su Kim. Masked spatial propagation network for sparsity-adaptive depth refinement. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.484, + 0.473 + ], + "angle": 0, + "content": "[30] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloé Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólár, and Ross B. Girshick. Segment anything. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[32] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision Workshop (ECCVW), 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[33] Torsten Kröger. Opening the door to new sensor-based robot applications—the reflexes motion libraries. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[34] Jason Ku, Ali Harakeh, and Steven L. Waslander. In defense of classical image processing: Fast depth completion on the cpu. 2018 15th Conference on Computer and Robot Vision (CRV), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[35] Iro Laina, C. Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In Proceedings of the International Conference on 3D Vision (3DV), 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[36] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[37] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[38] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.907, + 0.203 + ], + "angle": 0, + "content": "[39] Ming Liang, Binh Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[40] Haotong Lin, Sida Peng, Jingxiao Chen, Songyou Peng, Ji-aming Sun, Minghuan Liu, Hujun Bao, Jiashi Feng, Xiaowei Zhou, and Bingyi Kang. Prompting depth anything for 4k resolution accurate metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[41] Yuankai Lin, Hua Yang, Tao Cheng, Wending Zhou, and Zhouping Yin. Dyspn: Learning dynamic affinity for image-guided depth completion. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.907, + 0.399 + ], + "angle": 0, + "content": "[42] Beyang Liu, Stephen Gould, and Daphne Koller. Single image depth estimation from predicted semantic labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[43] Sifei Liu, Shalini De Mello, Jinwei Gu, Guangyu Zhong, Ming-Hsuan Yang, and Jan Kautz. Learning affinity via spatial propagation networks. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.906, + 0.524 + ], + "angle": 0, + "content": "[44] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association of Computational Linguistics (ACL), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[45] Xin Liu, Xiaofei Shao, Boqian Wang, Yali Li, and Shengjin Wang. Graphcspn: Geometry-aware depth completion via dynamic gcns. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.906, + 0.636 + ], + "angle": 0, + "content": "[46] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.906, + 0.704 + ], + "angle": 0, + "content": "[47] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[48] Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swaprompt: Test-time prompt adaptation for vision-language models. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[49] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.833, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[50] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Finetuning image-conditional diffusion models is easier than you think. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9452" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[51] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.485, + 0.205 + ], + "angle": 0, + "content": "[52] Jin-Hwi Park and Hae-Gon Jeon. A simple yet universal framework for depth completion. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[53] Jin-Hwi Park, Jaesung Choe, Inhwan Bae, and Hae-Gon Jeon. Learning affinity with hyperbolic representation for spatial propagation. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.32 + ], + "angle": 0, + "content": "[54] Jin-Hwi Park, Chanhwi Jeong, Junoh Lee, and Hae-Gon Jeon. Depth prompting for sensor-agnostic depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[55] Nikolay Patakin, Anna Vorontsova, Mikhail Artemyev, and Anton Konushin. Single-stage 3d geometry-preserving depth estimation model training on dataset mixtures with uncalibrated stereo data. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[56] Suraj Patni, Aradhye Agarwal, and Chetan Arora. Ecodepth: Effective conditioning of diffusion models for monocular depth estimation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[57] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[58] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc van Gool. Undepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[60] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[61] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[62] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya K. Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.908, + 0.135 + ], + "angle": 0, + "content": "Wu, Ross B. Girshick, Piotr Doll'ar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.192 + ], + "angle": 0, + "content": "[63] Laria Reynolds and Kyle McDonell. Prompt programming for large language models: Beyond the few-shot paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.908, + 0.248 + ], + "angle": 0, + "content": "[64] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[65] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.908, + 0.376 + ], + "angle": 0, + "content": "[66] Kieran Saunders, George Vogiatzis, and Luis J Manso. Self-supervised monocular depth estimation: Let's talk about the weather. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.419 + ], + "angle": 0, + "content": "[67] Ashutosh Saxena, Sung H. Chung, and A. Ng. Learning depth from single monocular images. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.908, + 0.489 + ], + "angle": 0, + "content": "[68] Shreyas S. Shivakumar, Ty Nguyen, Steven W. Chen, and Camillo Jose Taylor. Dfusenet: Deep fusion of rgb and sparse depth information for image guided dense depth completion. IEEE Intelligent Transportation Systems Conference (ITSC), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.546 + ], + "angle": 0, + "content": "[69] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Proceedings of the European Conference on Computer Vision (ECCV), 2012. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.908, + 0.589 + ], + "angle": 0, + "content": "[70] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing (TIP), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.908, + 0.646 + ], + "angle": 0, + "content": "[71] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.908, + 0.73 + ], + "angle": 0, + "content": "[72] Yifu Tao, Marija Popović, Yiduo Wang, Sundara Tejaswi Digumarti, Nived Chebrolu, and Maurice F. Fallon. 3d lidar reconstruction with probabilistic depth completion for robotic navigation. In Proceedings of the IEEE International Conference on Intelligent Robots and Systems (IROS), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.908, + 0.773 + ], + "angle": 0, + "content": "[73] Jonas Uhrig, N. Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. International Conference on 3D Vision (3DV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.908, + 0.843 + ], + "angle": 0, + "content": "[74] Massimiliano Viola, Kevin Qu, Nando Metzger, Bingxin Ke, Alexander Becker, Konrad Schindler, and Anton Obukhov. Marigold-dc: Zero-shot monocular depth completion with guided diffusion. arXiv preprint arXiv:2412.13389, 2024. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[75] Yufei Wang, Yuchao Dai, Qi Liu, Peng Yang, Jiadai Sun, and Bo Li. Cu-net: Lidar depth-only completion with coupled u-net. IEEE Robotics and Automation Letters (RA-L), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9453" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[76] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[77] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.485, + 0.275 + ], + "angle": 0, + "content": "[78] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters (RA-L), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[79] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Baobei Xu, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.386 + ], + "angle": 0, + "content": "[80] Zhiqiang Yan, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Rignet++: Efficient repetitive image guided network for depth completion. arXiv preprint arXiv:2309.00655, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[81] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.483, + 0.515 + ], + "angle": 0, + "content": "[82] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[83] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[84] Youming Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and S. Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.483, + 0.724 + ], + "angle": 0, + "content": "[85] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.483, + 0.783 + ], + "angle": 0, + "content": "[86] Jiawen Zhu, Simiao Lai, Xin Chen, D. Wang, and Hutchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.786, + 0.483, + 0.839 + ], + "angle": 0, + "content": "[87] Yiming Zuo and Jia Deng. Ogni-dc: Robust depth completion with optimization-guided neural iterations. In Proceedings of the European Conference on Computer Vision (ECCV), 2024, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9454" + } + ] +] \ No newline at end of file diff --git a/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/0a2ddfb1-a84e-4b84-bcf1-12ab1957047a_origin.pdf b/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/0a2ddfb1-a84e-4b84-bcf1-12ab1957047a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5773a616b72aa74e4f1fdda6becc60f4fa226d66 --- /dev/null +++ b/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/0a2ddfb1-a84e-4b84-bcf1-12ab1957047a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c979b5719dde52b3dfc14ad411c835f32377969db55362d447ad3348bd6e4c34 +size 10805094 diff --git a/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/full.md b/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..637831665ffcde4469cbf9897a6e0713d7a43348 --- /dev/null +++ b/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/full.md @@ -0,0 +1,342 @@ +# Test-Time Prompt Tuning for Zero-Shot Depth Completion + +Chanhwi Jeong1 Inhwan Bae1 Jin-Hwi Park2* Hae-Gon Jeon3* +1GIST, 2Chung-Ang University, 3Yonsei University + +{chanhwij, inhwanbae}@gm.gist.ac.kr, jinhwipark@cau.ac.kr, earboll@yonsei.ac.kr + +![](images/cfa64ab433e1b9170c18b40a0a8f199cf234bd6e3c86a2ea8349236d98564ad5.jpg) +Figure 1. Comparison of methods for correcting the inaccurate scale of initial depth estimates produced by the depth foundation model using sensor-derived sparse depth with an accurate scale, represented as yellow dots. (a) The depth foundation model generates an initial depth map from a single RGB input. (b) The least squares fitting aligns the initial depth map with the sparse depth [54]. (c) The parameters of the foundation model are fine-tuned under supervision from the sparse depth. (d) Alternatively, the model's parameters are frozen, and a visual prompt is learned to adjust the depth estimates. Please refer to Fig. 3 for depth profiling results and error analysis. + +# Abstract + +Zero-shot depth completion using metric scales remains challenging, primarily due to performance limitations such as domain specificity and sensor characteristics. One recent emerging solution is to integrate monocular depth foundation models into depth completion frameworks, yet such efforts still face issues with suboptimal performance and often require further adaptation to the target task. Surprisingly, we find that a simple test-time training, which finetunes monocular depth foundation models on sparse depth measurements from sensors just as it is, yields reasonable results. However, this test-time training obviously incurs high computational costs and introduces biases towards specific conditions, making it impractical for real-world scenarios. In this paper, we introduce a new approach toward parameter-efficient zero-shot depth completion. Our key idea in this work is to leverage visual prompt tuning, achieving sensor-specific depth scale adaptation without forgetting foundational knowledge. Experimental results on diverse datasets demonstrate that our approach outperforms relevant state-of-the-art methods, showing superior generalization and efficiency. Code is publicly available at https://github.com/ch5374/TestPromptDC + +# 1. Introduction + +Depth completion, the technique of densifying sparse input measurements from active sensors based on correspond + +ing RGB images, is crucial in various applications including 3D reconstruction [72], autonomous driving [39] and robotics [33]. Beyond traditional methods that rely on handcrafted features [34] and optimization [14], data-driven approaches, which are mainly based on convolutional neural networks [27, 46, 47] and vision transformers [64, 84], have achieved the best performance with public benchmarks [15, 25]. + +However, models trained on a single dataset perform well on the test split of that dataset, but exhibit significant performance degradation in unseen environments. The primary reason lies in the scarcity of data available to use for depth completion. Since depth completion datasets require both dense ground truth depth maps and RGB images obtained from well-calibrated sensor fusion setups, large-scale dataset constructions must collect and label image-depth pairs across environmental variations. Obviously, this makes it difficult to achieve generalizable depth completion, including zero-shot inferences. + +One interesting research direction in [40, 52, 54] to mitigate the data scarcity is to integrate foundation models for monocular depth estimation [4, 17, 22, 30, 57, 58, 60, 61, 81-83] trained on huge-scale datasets into depth completion for better generalization. This is quite feasible because foundation models can generate approximate metric or relative depth from a single input image as demonstrated in Fig. 1-(a), actual scene depths can be obtained by fitting sparse depth measurements onto these predictions. Unfortunately, there are still two challenges: (1) they require carefully designed + +additional networks, as the existing foundation models struggle to encode sparse depth information; (2) additional training for the depth completion task is also necessary. These two challenges result in either an increase in computational complexity or the introduction of new biases in the training phase. In this paper, we explore one important question: how can we take full advantage of monocular depth foundation models for depth completion? Our work is the first to answer this question by investigating the most effective foundation models, without requiring additional network design or fine-tuning of the pretrained model. + +We initiate our investigation by proposing a simple yet highly effective strategy, which is fine-tuning the foundation models in the inference phase, only using sparse depth measurements as supervision. As an example in Fig. 1-(b), compared to the sparse depth fitting onto dense depths [40, 52, 54], fine-tuning the foundation models preserves the original depth while allowing the scale to adapt locally (Fig. 1-(c)). Of course, updating all parameters requires significant computing resources and poses the risk of dataset bias, potentially causing models to lose their own knowledge. + +As a more efficient alternative, we present a test-time prompt tuning that adapts foundation models at inference time without modifying its parameters (Fig. 1-(d)). We introduce a small set of learnable parameters that guide their prediction while keeping the foundation model frozen. While existing visual prompt [1, 28] uses fixed-size prompts positioned at specific locations (e.g., fixed positions, random placements, or padding) for image classification, we introduce a pixelwise visual prompt specifically designed for depth completion. Compared to previous works that operate at a global or patch level, our pixelwise visual prompt is initialized to zero and added to input images, allowing us to leverage the inherent spatial structures of depth maps. We validate our visual prompt tuning across diverse benchmark datasets and environments, achieving superior zero-shot performance. + +# 2. Related work + +# 2.1. Monocular Depth Estimation + +Traditional monocular depth estimation depends on handcrafted visual cues [20, 67] and scene-specific assumptions [21, 42]. With the advent of deep learning, subsequent methods [16, 35, 36] using data-driven representations have significantly outperformed the traditional ones reducing the reliance on manually designed features and rigid scene-specific assumptions. Recent works have explored techniques such as data mixing [55, 60, 61] and augmentation [25, 66] to achieve zero-shot generality. Moreover, some works in [10, 17, 30, 50, 56, 85] exploit the generative priors of pretrained diffusion models [65]. However, inferring + +metric-scale depth from single images still remains challenging. To address this, metric-scale depth estimation from monocular images has evolved by incorporating global depth distributions [19], scene-aware conditioning [3] and camera intrinsics [4, 22, 22, 57, 58, 83]. While earlier methods rely on known camera parameters [22, 83], recent approaches estimate focal length [4, 57, 58] directly from learning features, enhancing generalization across diverse environments without any explicit camera metadata. Nevertheless, these methods obviously lag behind approaches that use sensor measurements like LiDAR in terms of accuracy. + +# 2.2. Depth Completion + +Due to the inherent issues with monocular depth estimation, research on unguided depth completion [12, 23, 73, 75], transforming sparse depth data from sensors into dense depth maps without guidance RGB images, has been studied. However, most mainstream depth completion is to use RGB-depth pairs to directly leverage scene contexts such as edges and textures [26, 46, 47, 68, 70, 79, 80, 84, 87]. Furthermore, constructing pixelwise affinity [6, 7, 41, 43, 45, 51, 53] from corresponding RGB images has proven to be effective as it helps preserve depth discontinuity when propagating sparse depth. Despite these advancements, such models often struggle to show the same performance in unseen domains. To mitigate this, recent approaches have leveraged monocular depth foundation models for depth completion. For instance, works in [29, 54] introduces interesting frameworks that exploit features and predict depth from foundation models to perform adaptive depth refinement with sparse depth inputs. Additionally, diffusion-based depth completion methods [24, 74] have adopted the test-time optimization of latent representations, guided by input sparse depths. However, these approaches often require training on additional data [29, 54] or suffer from extremely slow inference speeds [24, 74]. + +# 2.3. Prompting for Foundation Models + +Foundation models [5, 9, 31, 59, 62, 65] exhibit remarkable zero-shot performance across a wide range of tasks as well as depth perception. In natural language processing, prompting schemes have been introduced to leverage these foundation models effectively for downstream tasks [5, 37, 38, 44, 63, 77]. To apply such prompting schemes to the computer vision domain, works in [28, 86] have introduced visual prompt tuning, which incorporates trainable visual tokens into models for specific downstream tasks. Meanwhile, one interesting work in [1] presents learnable padding prompts appended to input images. More recently, visual prompt tuning has also been extended to test-time adaptation scenarios [13, 48, 49] for better adaptation to unseen domains. Inspired by them [1, 49], we propose a test-time prompt tuning strategy for depth completion. + +![](images/1a945e7b4c26cb908271025f0d40222334bfb3f4620048d27dbd692a0fc1d723.jpg) +(a) Supervised training models + +![](images/ee70203d226eb9cccedaa32f6acbe04e5f707396c58d885efa2bba94661acade.jpg) +(b) Test-time optimization models + +![](images/af5aeed83572c04e80106469f99864c72eabe6bcf42e5217be0105d3e10c28e9.jpg) +(c) Our fine-tuning approach + +![](images/6a181e379d0decf80db53c867c259f57effe87e5fceb59b8822cff99481a2b5b.jpg) +(d) Our visual prompting approach + +# 3. Methodology + +In this section, we present our approach for parameter-efficient zero-shot depth completion through test-time training. In Sec. 3.1, we discuss the role of depth foundation models in depth completion, emphasizing their generalization capabilities while identifying key challenges. Next, in Sec. 3.2, we explore a fine-tuning strategy to adapt the foundation model for sparse depth measurements. This approach achieves impressive performance without relying on additional network design or prompt engineering; however, it introduces computational overhead and poses generalization challenges. To address these limitations, in Sec. 3.3, we propose a novel pixel-wise visual prompting method, enabling efficient adaptation without any parameter modification of the foundation model. Finally, in Sec. 3.4, we describe implementation details, including optimization strategies, dataset preprocessing and experimental settings. + +# 3.1. Rationale: Usage of Depth Foundation Model + +Recent works have attempted to leverage monocular depth foundation models for depth completion. These models offer strong generalization capabilities and are considered to be attractive solutions for better zero-shot performance. However, incorporating them into depth completion frameworks faces several technical challenges, as described below: + +Incompatibility with sparse depth input handling. Since monocular depth foundation models are exclusively designed to infer depth from single RGB images, they do not have mechanisms to take sparse depth measurements as guidance. As a result, the existing methods require additional components to process sparse depths and to integrate them into the models, increasing computational complexity as shown in Fig. 2-(a). + +Scale mismatch between prediction and sensor depth. Current depth foundation models are able to estimate approximate metric-scale depths [3, 4, 58] as well as relative-scale depth [17, 30, 60, 61] with high-fidelity depth discontinuities. Unfortunately, the depth foundation models for metric-scale inference exhibit systematic scale mismatches when compared to sensor-derived depths. While least squares scale + +![](images/b2a7a78d9ad26d5c41b4e35c857d384557f110a32d60c90daca7770c08ca217d.jpg) +Figure 2. Comparison of the previous foundation model-based depth completion approaches and our approach. (a) Existing methods [40, 52, 54] design additional networks to integrate sensor-based depth measurements into the foundation model. (b) Recent approaches [24, 74] optimize latent representations in diffusion-based foundation models [17, 30] without dense GT but are computationally expensive. (c) Our fine-tuning method adapts the model with sparse depth supervision, and (d) our visual prompting approach enables efficient adaptation while keeping the foundation model frozen. +(a) Least squares fitting +Figure 3. Qualitative comparison of different strategies on the depth map from the data used in Fig. 1. The top row shows the predicted depth maps, while the bottom row presents the depth profiles along the indicated horizontal line. The red $\times$ marks indicate the sparse depth measurements. + +![](images/ed49bf74b93ebcd72f25803b4ce9918a24788aa2e185869a03979475c9569076.jpg) +(b) Fine-tuning + +![](images/c5778b9833e91ee17ac1158e65636d8a2fbc54a1f8148bd23ae65050e4aa65a3.jpg) +(c) Visual Prompting + +fitting in [54] can globally align depth scales, it fails in regions where depth varies significantly. For instance, as shown in Fig. 3, the least squares fitting locally adjusts the correct depth scale at areas near initial depth seeds (e.g., pixel range 100–150), but fails it in other regions (e.g., pixel range 50–100). + +Limitations of existing adaptation strategies. Previous approaches require ground truth (GT) annotations for adaptation, which are costly and labor-intensive (Fig. 2-(a)). For example, works in [40, 54] rely on large-scale datasets, while UniDC [52] adopts a few-shot strategy within the target domain. However, generating new GT labels to cover all scenarios in-the-wild, is impractical, necessitating a more efficient alternative. One promising solution is a test-time optimization, which enables adaptation to new environments without any additional GT annotation (Fig. 2-(b)). Recent contemporary works in [24, 74] have introduced similar strategies by optimizing latent representations in diffusion-based models (Fig. 2-(b)). However, these approaches come with noticeable limitations, including high computational costs and slow inference speeds. + +To address these limitations, we propose fine-tuning depth foundation models in the inference phase, directly using sparse depth measurements as supervision without introducing additional network components (Fig. 2-(c)). This eliminates the dependency on extensive GT data and enables real-time adaptation to new environments. + +# 3.2. Finetuning monocular depth foundation model + +Let $F_{\theta}$ denote a depth foundation model, parametrized by $\theta$ , and $I \in \mathbb{R}^{3 \times H \times W}$ is an RGB image. Typically, $F_{\theta}$ takes only the $I$ as input and generates a dense depth map $D \in \mathbb{R}^{1 \times H \times W}$ as output, which can be expressed as follows: + +$$ +F _ {\theta} (I) = D. \tag {1} +$$ + +Since $D$ is derived solely from a single image, it inherently suffers from scale ambiguity. To mitigate this issue, a sparse depth $S$ , obtained from an active sensor with accurate scales, is employed as supervision for $F_{\theta}$ . This approach allows the model to alleviate the depth scale ambiguity. + +First, to use only valid sparse depth values as supervision, we define a validity mask $M$ as follows: + +$$ +M _ {i, j} = \left\{ \begin{array}{l l} 1, & \text {i f} S _ {i, j} \text {i s v a l i d} \\ 0, & \text {o t h e r w i s e ,} \end{array} \right. \tag {2} +$$ + +where $(i,j)$ is the pixel coordinate in the image. + +With this mask, we initially adjust the scale of the foundation depth by roughly aligning it using the valid sparse depth measurements in the least squares fitting process, as formulated below: + +$$ +s _ {0}, s _ {1} = \min _ {\left(s _ {0}, s _ {1}\right)} \| M \odot \left(\left(s _ {0} D + s _ {1}\right) - S\right) \| _ {F}, \tag {3} +$$ + +where $\odot$ denotes an element-wise multiplication. $s_0$ and $s_1$ mean scale and shift values, respectively, and $\| \cdot \| _F$ indicates the Frobenius norm. Next, we optimize the model's parameters $\theta$ using only the valid sparse depth measurements as supervision. The optimal parameters, denoted by $\theta^{*}$ , are determined by minimizing the following loss: + +$$ +\theta^ {*} = \arg \min _ {\theta} \sum_ {i, j} M _ {i, j} \odot | s _ {0} F _ {\theta} (I) _ {i, j} + s _ {1} - S _ {i, j} | ^ {p}, \tag {4} +$$ + +where $p = 1$ and $p = 2$ correspond to the $L_{1}$ and $L_{2}$ losses, respectively and the final loss is the sum of both. To isolate the effects of fine-tuning itself, we do not incorporate specialized loss functions. + +# 3.3. Pixelwise Visual Prompt + +To address the computational overhead and potential knowledge degradation of the fine-tuning of the foundation model, we propose a pixelwise visual prompt, a novel test-time adaptation method designed specifically for depth completion (Fig. 2-(d)). Unlike conventional fine-tuning, our approach enables efficient adaptation while keeping the model's parameters frozen. + +Given a frozen pretrained foundation model $F_{\theta}$ and a sample consisting of a pair of image and sparse depth measurement $(I,S)$ , we aim to learn a pixelwise visual prompt $V_{\phi} \in \mathbb{R}^{3 \times H \times W}$ , parameterized by $\phi$ . The prompt is added directly to the input image, forming a prompted image $I + V_{\phi}$ and we optimize it by minimizing the depth prediction error + +in the training phase as follows: + +$$ +\phi^ {*} = \arg \min _ {\phi} \sum_ {i, j} M _ {i, j} \odot \left| F _ {\theta} \left(I + V _ {\phi}\right) _ {i, j} - S _ {i, j} \right| ^ {p}. \tag {5} +$$ + +Here, only the prompt parameters $\phi$ are optimized, while $\theta$ remains frozen. Once optimized, the visual prompt $V_{\phi^*}$ is applied to the image as below: + +$$ +F _ {\theta} \left(I + V _ {\phi^ {*}}\right). \tag {6} +$$ + +In addition to achieving superior performance compared to existing methods, this approach also enhances usability, providing an additional advantage. Since the pixelwise visual prompt is initialized to zero and added to the image, it ensures a minimal level of depth prediction performance even without prompt tuning. Additionally, users can dynamically adjust the time allocated for prompt tuning based on their available resources and the required inference speed, allowing for stable and consistent performance improvement. + +Difference with existing visual prompt methods. While visual prompt tunings [1, 28] have been previously explored to adapt high-level representations using fixed-size prompts at predefined locations, such approaches lack the spatial granularity needed for precise depth refinement. These methods often focus on global or patch-level modifications, making them less effective for tasks requiring fine-grained adjustments, such as depth completion. In contrast, our pixelwise visual prompt is designed to adapt at the individual pixel level, enabling localized corrections that align depth predictions more accurately with sparse sensor measurements. Our approach not only enhances adaptation flexibility but also maintains the strong generalization capabilities of foundation models. + +# 3.4. Implementation Details + +Our implementation is based on PyTorch and runs on a single NVIDIA RTX 3090 GPU. We use the AdamW optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ . The learning rate is set to 1e-6 for fine-tuning and 1e-3 for visual prompting. The training process consists of 80 steps for fine-tuning and 50 steps for visual prompting. For all depth foundation models, we use their official pretrained weights and specifically the v2-old version for UniDepthV2 [58], while other models are evaluated in our ablation study. For models that do not support the original image size, we first resize the images while preserving their aspect ratio. + +# 4. Experiment + +In this section, we begin by outlining the experimental setup in Sec. 4.1, including the evaluation protocols, and introduce the datasets. We then compare our method with state-of-the-art (SOTA) methods for zero-shot depth completion tasks, in Sec. 4.2. Lastly, in Sec. 4.3, we present an ablation study to analyze the impact of our methods. + +![](images/db46f0ece6530514232395271c50d03d3db527790e9f90a9745040aecef78c4a.jpg) +Figure 4. Qualitative result and depth error map of various depth completion methods on the IBims-1, VOID 1500, and DDAD datasets. + +
MethodIBims-1VOIDNYUv2KITTIDDAD
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
NLSPN [51]0.0490.1910.4271.3530.4400.7161.3352.0762.4989.231
CompletionFormer [84]0.0580.2060.6421.6840.1860.3740.9521.9352.5189.471
VPP4DC [2]0.0620.2280.1660.6060.0770.2470.4131.6091.3446.781
BP-Net [71]0.0780.2890.2700.742----2.2708.344
OGNI-DC [87]0.0590.1860.1750.593----1.8676.876
Zero-DC [24]0.0780.2030.1870.6070.0690.1580.3971.4131.6817.122
Marigold-DC [74]0.0450.1660.1520.5510.0480.1240.4341.4652.3646.449
Ours-LS0.1130.2210.1870.5710.1230.2161.2052.8015.67611.97
Ours-FT (Sec. 3.2)0.0490.1520.1380.5140.0480.1070.3831.3291.7755.881
Ours-VP (Sec. 3.3)0.0430.1610.1320.5080.0410.1060.3431.3331.4825.811
+ +Table 1. Quantitative comparison with SOTA methods on five zero-shot benchmarks. Most methods used values reported in [74], while the results for Zero-DC [24] are evaluated based on the DepthFM model [17]. For the same reason as in [74], we do not include results for BP-Net [71] and OGNI-DC [87] on NYUv2 and KITTI. Bold: Best, Underline: Second best. + +# 4.1. Experimental Setup + +Our test-time training strategy is closely aligned with recent works [24, 74], which focus on optimizing the model at inference time. Following their methodology, we adapt the model using only sparse depth measurements during test-time, ensuring a fully zero-shot approach while effectively refining depth scale estimation. To maintain a fair evaluation, we strictly exclude the dataset (e.g., ScanNet [8]) used for training the depth foundation model [58] from the test. Furthermore, in our fine-tuning experiments, we prevent the model from accumulating dataset-specific biases by fine-tuning on + +a single sample and resetting it to its original parameters before proceeding to the next evaluation. This ensures that our approach remains consistent with the principles of test-time optimization and maintains its generalization ability across diverse environments. + +We evaluate our proposed method in a zero-shot setting using five real-world datasets that are not seen during the training of the base model [58]. These datasets encompass both indoor and outdoor environments, varying image resolutions, and different levels of sparse depth density, ensuring a comprehensive assessment of generalization ability. + +![](images/4b2f2f1ba073f013763a8a73e279734a2b233cb21263c9303ab8c3bca59cbfad.jpg) +Figure 5. Qualitative comparison under varying levels of depth sparsity across NYUv2 and KITTI. Despite significantly reduced depth measurements, our approach recovers high-quality dense depth maps, demonstrating robustness to input sparsity in the depth completion. + +
MethodNYUv2 (200 samples)NYUv2 (100 samples)NYUv2 (50 samples)NYUv2 (5 samples)KITTI (32 channels)KITTI (16 channels)KITTI (8 channels)KITTI (4 channels)
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
Zero-DC [24]0.0960.1970.1240.2340.1610.2820.4470.6230.5401.9160.6942.3231.0082.9041.6874.184
Marigold-DC [74]0.0730.1750.0900.2030.1120.2360.3580.5410.8752.1930.9872.6121.1863.1761.6694.240
Ours-VP0.0530.1330.0650.1520.0800.1800.1380.2470.3961.5400.4751.8510.5672.0930.8122.644
+ +Table 2. Quantitative comparison on varying sparsity. Bold: Best, Underline: Second best. + +(1) IBims-1 [32] is an indoor dataset consisting of 100 RGB-D pairs with a resolution of $480 \times 640$ . We use the entire dataset for testing. Since sparse depth data is not provided, we follow [74] to generate a depth valid mask and construct sparse depth by randomly sampling 1,000 depth points. +(2) VOID [78] contains RGB-D frames from 56 sequences, with 48 designated for training and 8 for testing. The official test set comprises 800 frames, and sparse depth maps are obtained using the VIO system. Following [74], we evaluate our method on this test set using 1,500 sparse depth points. + +(1) NYUv2 [69] comprises 464 scenes recorded using a Microsoft Kinect sensor. The official split is divided into 249 training and 215 test scenes. We use the official test set of 654 images for evaluation. Following a commonly used setup [11, 35, 46], we downsample and center-crop the images to $228 \times 304$ before conducting experiments. Also, sparse depth is obtained by randomly selecting 500 samples from the depth map. +(4) KITTIDC [15] consists of over 90K RGB images paired with LiDAR data. Since the depth obtained from LiDAR is inherently sparse, the ground truth depth is generated by aggregating multiple LiDAR frames. For our experiments, + +we use the 1,000 pairs provided for validation. To reduce noise in the input sparse depth, we apply the filtering method used in [76]. + +(5) DDAD [18] is an outdoor RGB-D driving dataset collected using a Luminar-H2 LiDAR sensor with a resolution of $1216 \times 1936$ . It consists of 12K training samples and 40K validation samples. We use that 40K validation samples for evaluation. Also, sparse depth is generated by randomly sampling $20\%$ . + +Evaluation Metrics. For quantitative evaluation, we employ the following widely used metrics, Root Mean Squared Error (RMSE) and Mean Squared Error (MAE), to assess performance. + +# 4.2. Zero-shot Depth Completion Results + +We compare our methodology not only with conventional approaches [51, 71, 84] but also with models [2, 24, 74, 87] that focus on improving zero-shot depth completion performance. Notably, as shown in Tab. 1, our simple finetuning methodology (Ours-FT) outperforms many models designed for generalization across most datasets. This result highlights the effectiveness of using sparse depth as supervision, demonstrating that the depth foundation model can + +![](images/2a3b85b66d3ba3b79576813b85eba2e3b8f9be959e37658d4d22a57b25596994.jpg) +Figure 6. Visualization of our visual prompt, predicted depth, and corresponding error map at each test-time training iteration. As the visual prompt progressively learns from the sparse depth guidance, the predicted depth scale becomes increasingly accurate, which is reflected by the gradual reduction in the error maps. + +![](images/26e95e01c45afef44ee2f5e41a8cce746a03f618c37b188d89bbe1f4809bea53.jpg) + +
MethodIBims-1NYUv2
MAE↓RMSE↓MAE↓RMSE↓
UniDepthV2-LS [58]0.1130.2210.1230.216
UniDepthV2-VP0.0430.1610.0410.106
Metric3dV2-LS [22]0.1100.2100.1740.275
Metric3dV2-VP0.0390.1470.0690.114
DepthAnytingV2-LS [82]0.1430.2620.2190.326
DepthAnytingV2-VP0.0400.1490.0550.122
ZoeDepth-LS [3]0.6230.8380.1740.275
ZoeDepth-VP0.1760.3410.0490.114
+ +be trained to achieve competitive performance. Furthermore, our proposed pixelwise visual prompting method (Ours-VP) surpasses simple fine-tuning in most cases, illustrating that learning visual prompts can be both more efficient and more effective than fine-tuning the entire model. This finding suggests that visual prompting provides a promising alternative to full fine-tuning, offering a favorable balance between computational efficiency and performance. As shown in Fig. 4, the qualitative results indicate that existing SOTA models, which are trained on synthetic data, tend to capture fine-detailed depth. However, their performance is less favorable when examined through error maps, suggesting that accurate scale estimation is a prerequisite for high-quality depth prediction and takes precedence over fine-detail reconstruction. + +Robustness to sparsity variations. To further assess our model's adaptability to real-world scenarios, we conduct experiments under varying levels of input sparsity. As shown in Tab. 2, our method consistently outperforms SOTA models even as the number of sparse depth points decreases. Notably, our model exhibits greater robustness to sparse depth bias, which is crucial in real-world applications where the number of available depth measurements can vary significantly. As seen in Fig. 5, comparing methods, scale is often misinterpreted when sparse depth is limited, while our approach maintains accuracy, making it highly suitable for + +Table 3. Performance comparison of our visual prompting method across different depth foundation models. Our approach consistently enhances performance, demonstrating its adaptability to various models. + +
MethodMAE ↓RMSE ↓Learnable Parameters
Fine-tuning (Sec. 3.2)0.0490.152100%
Random Patch [1]0.1110.2210.11%
Fixed patch [1]0.0790.1850.11%
Padding patch [1]0.0500.1670.11%
Pixelwise (Sec. 3.3)0.0430.1610.25%
+ +Table 4. Comparison of different visual prompting designs in terms of accuracy and parameter efficiency on dataset Ibims-1 [32]. Our pixelwise visual prompt achieves fine-tuning-level performance while significantly reducing learnable parameters. + +real-world deployment. + +Effectiveness of visual prompt in scale alignment. We next analyze the test-time training process of our pixelwise visual prompt. As shown in Fig. 6, the scale accuracy progressively improves with each test-time training step. When the initial scale is generally well-aligned (left example), the prompt refines predictions gradually, leading to a smooth and consistent improvement. In contrast, when large-scale discrepancies exist due to reflective surfaces like mirrors (right example), the prompt first corrects the dominant errors in these regions before adjusting the overall scale. This adaptive behavior effectively mitigates the scale discrepancy problem discussed in Sec. 3.1. After approximately 10 test-time training iterations, the model stabilizes, achieving a well-calibrated depth scale. This highlights the efficiency of our visual prompt learning process, ensuring rapid convergence with minimal computational overhead. + +# 4.3. Ablation Study + +Depth foundation models. We further validate our method by applying it to various depth foundation models beyond UniDepthV2 [58], as shown in Tab. 3. Our visual prompting method consistently improves performance across all foundation models. Interestingly, while UniDepthV2 achieves the best results on NYUv2, Metric3dV2 performs better on IBims-1, suggesting that different foundation models may excel with different datasets. This reinforces the importance of model-agnostic test-time training strategies like visual prompting. + +
MethodRequired Memory (MB)Inference Speed (s)
Zero-DC [24]872720.56
Marigold-DC [74]21748121.04
Ours-FT83722.95
Ours-VP39912.93
+ +Table 5. Comparison of test-time training methods measured on the NYUv2 dataset [69]. + +Prompting design. We compare our pixelwise visual prompt against three conventional visual prompting methods: random patches, fixed patches, and padding-based prompts. As shown in Tab. 4, all prompting techniques reduce computational costs by freezing foundation model parameters, but existing prompting methods as proposed by [1] fail to match the performance of fine-tuning. This suggests a trade-off between parameter efficiency and accuracy in conventional prompting techniques. In contrast, our pixelwise visual prompt bridges this gap, achieving fine-tuning-level performance while maintaining efficiency. This is because our method preserves fine-grained spatial adjustments, ensuring that depth refinements occur at the pixel level rather than through global modifications. These results highlight the effectiveness of our approach in maximizing both performance and efficiency. + +Computational efficiency. We compare our approach with existing test-time optimization methods, including Zero-DC [24] and Marigold-DC [74], to highlight its efficiency in memory usage and inference speed. As shown in Tab. 5, our pixelwise visual prompting (Ours-VP) requires significantly less memory (3991 MB) and achieves a fast inference speed of 2.93s, outperforming Zero-DC (20.56s) and Marigold-DC (121.04s). Even our fine-tuning variant (Ours-FT) remains computationally efficient. These results demonstrate that our method avoids the heavy computational cost of full fine-tuning while adapting effectively, making it highly practical for real-world use. + +Learning rate & test-time training step. In Fig. 7, we analyze the learning behavior of fine-tuning versus visual prompting. Fine-tuning exhibits unstable learning dynamics, particularly when using a high learning rate, as abrupt parameter shifts distort depth priors and reduce prediction stability in early iterations. Additionally, fine-tuning requires carefully tuned learning rates to prevent performance degradation. In contrast, our visual prompt method is robust to different learning rates, consistently achieving stable and rapid convergence. As seen in Fig. 7, visual prompting requires significantly fewer iterations to reach a reliable depth estimation, making it more efficient and practical for real-time applications. + +Analysis of structures of learned visual prompt. Lastly, we analyze the characteristics of learned patterns in the visual prompts. As shown in Fig. 8, regions belonging to differ + +![](images/faa3f2119c3b540808cedc0165dc1e866163e9a9b0f13ad0590736983aaa6271.jpg) +Figure 7. Comparison of fine-tuning and visual prompting methods. The fine-tuning method shows unstable convergence at high learning rates, while the visual prompting method remains stable across different rates and converges faster. The dotted line indicates the baseline using least squares fitting, while visual prompting achieves superior performance with fewer iterations. + +![](images/f00c7d6a531235052dec6e6bc5804e6f4c0e76a6e3f9e29a9d7d55c6c4c18b47.jpg) +(a) RGB image +Figure 8. Visualization of the RGB image, visual prompt, and the combined RGB+Visual Prompt. The colored boxes highlight key regions with structural differences and similarities. + +![](images/14d0e7f05248035dcf778ca1c45328a703582c37b415ab456451cdb7c1689865.jpg) +(b) Visual Prompt + +![](images/2c151595d74f290b6b2cfe49350abd2900fe9e2c797e2f0e80aa1a9dadc0a76d.jpg) +(c) RGB+Visual Prompt + +ent planes, shown in the yellow and green boxes, show that entirely different visual prompt structures are required to accurately adjust distances between the two walls. Conversely, regions on the same plane, shown in cyan and magenta boxes, have relatively similar structures, introducing the same influence to the input. To represent the metric-aware 3D geometry, these visual prompts are learned to adjust the depth to align with sensor measurements, enabling high-fidelity depth estimation while keeping the depth foundation model frozen. + +# 5. Conclusion + +This study explores how to effectively utilize foundation models for zero-shot depth completion. We find that simply fine-tuning the depth foundation model with sparse depth measurements outperforms existing methods, demonstrating strong adaptability. However, fine-tuning introduces computational overhead and risks knowledge forgetting. To address this, we propose pixelwise visual prompt tuning, which efficiently refines depth predictions without modifying model parameters. Our method achieves fine-tuning-level performance while maintaining efficiency, making it well-suited for zero-shot adaptation across diverse environments. + +Acknowledgement This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2024-00338439) and the MSIT(Ministry of Science and ICT), Korea, under the Graduate School of Metaverse Convergence support program(ITP-2025-RS-2024-00418847) supervised by the IITP(Institute for Information & Communications Technology Planning & Evaluation) + +# References + +[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 2, 4, 7, 8 +[2] Luca Bartolomei, Matteo Poggi, Andrea Conti, Fabio Tosi, and Stefano Mattoccia. Revisiting depth completion from a stereo matching perspective for cross-domain generalization. In 2024 International Conference on 3D Vision (3DV), 2024. 5, 6 +[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 2, 3, 7 +[4] Alexey Bochkovskiy, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 1, 2, 3 +[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Ma teusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. Proceedings of the Neural Information Processing Systems (NeurIPS), 2020. 2 +[6] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2018. 2 +[7] Xinjing Cheng, Peng Wang, Chenye Guan, and Ruigang Yang. Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019. 2 +[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 2 +[10] Yiqun Duan, Xianda Guo, and Zhengbiao Zhu. Diffusion-depth: Diffusion denoising approach for monocular depth estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2 +[11] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Proceedings of the Neural Information Processing Systems (NeurIPS), 2014. 6 +[12] Abdelrahman Eldisokey, Michael Felsberg, Karl Holmquist, and Mikael Persson. Uncertainty-aware cnns for depth com + +pletion: Uncertainty from beginning to end. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[13] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[14] David Ferstl, Christian Reinbacher, René Ranftl, Matthias Rüther, and Horst Bischof. Image guided depth upsampling using anisotropic total generalized variation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2013. 1 +[15] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 1, 6 +[16] Clément Godard, Oisin Mac Aodha, and Gabriel J. Brostow. Unsupervised monocular depth estimation with left-right consistency. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2 +[17] Ming Gui, Johannes S. Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Bjorn Ommer. Depthfm: Fast monocular depth estimation with flow matching. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025. 1, 2, 3, 5 +[18] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6 +[19] Vitor Campagholo Guizilini, Pavel Tokmakov, Achal Dave, and Rares Ambrus. Grin: Zero-shot metric depth with pixel-level diffusion. arXiv preprint arXiv:2409.09896, 2024. 2 +[20] Derek Hoiem, Alexei A. Efros, and Martial Hebert. Automatic photo pop-up. Proceedings of ACM SIGGRAPH, 2005. 2 +[21] Derek Hoiem, Alexei A. Efros, and Martial Hebert. Recovering surface layout from an image. International Journal on Computer Vision (IJCV), 2007. 2 +[22] Mu Hu, Wei Yin, China. Xiaoyan Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 1, 2, 7 +[23] Zixuan Huang, Junming Fan, Shenggan Cheng, Shuai Yi, Xiaogang Wang, and Hongsheng Li. Hms-net: Hierarchical multi-scale sparsity-invariant network for sparse depth completion. IEEE Transactions on Image Processing (TIP), 2018. 2 +[24] Lee Hyoseok, Kyeong Seon Kim, Kwon Byung-Ki, and TaeHyun Oh. Zero-shot depth completion via test-time alignment with affine-invariant depth prior. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025. 2, 3, 5, 6, 8 + +[25] Dmitry Yu. Ignatov, Andrey D. Ignatov, and Radu Timofte. Virtually enriched nyu depth v2 dataset for monocular depth estimation: Do we need artificial augmentation? Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2024. 1, 2 +[26] Saif Muhammad Imran, Yunfei Long, Xiaoming Liu, and Daniel Morris. Depth coefficients for depth completion. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[27] Maximilian Jaritz, Raoul de Charette, Émilie Wirbel, Xavier Perrotton, and Fawzi Nashashibi. Sparse and dense data with cnns: Depth completion and semantic segmentation. 2018 International Conference on 3D Vision (3DV), 2018. 1 +[28] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 4 +[29] Jinyoung Jun, Jae-Han Lee, and Chang-Su Kim. Masked spatial propagation network for sparsity-adaptive depth refinement. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[30] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3 +[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloé Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólár, and Ross B. Girshick. Segment anything. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[32] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision Workshop (ECCVW), 2018. 6, 7 +[33] Torsten Kröger. Opening the door to new sensor-based robot applications—the reflexes motion libraries. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2011. 1 +[34] Jason Ku, Ali Harakeh, and Steven L. Waslander. In defense of classical image processing: Fast depth completion on the cpu. 2018 15th Conference on Computer and Robot Vision (CRV), 2018. 1 +[35] Iro Laina, C. Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In Proceedings of the International Conference on 3D Vision (3DV), 2016. 2, 6 +[36] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 2 +[37] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021. 2 + +[38] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), 2021. 2 +[39] Ming Liang, Binh Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 1 +[40] Haotong Lin, Sida Peng, Jingxiao Chen, Songyou Peng, Ji-aming Sun, Minghuan Liu, Hujun Bao, Jiashi Feng, Xiaowei Zhou, and Bingyi Kang. Prompting depth anything for 4k resolution accurate metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3 +[41] Yuankai Lin, Hua Yang, Tao Cheng, Wending Zhou, and Zhouping Yin. Dyspn: Learning dynamic affinity for image-guided depth completion. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2024. 2 +[42] Beyang Liu, Stephen Gould, and Daphne Koller. Single image depth estimation from predicted semantic labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 2 +[43] Sifei Liu, Shalini De Mello, Jinwei Gu, Guangyu Zhong, Ming-Hsuan Yang, and Jan Kautz. Learning affinity via spatial propagation networks. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2017. 2 +[44] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association of Computational Linguistics (ACL), 2022. 2 +[45] Xin Liu, Xiaofei Shao, Boqian Wang, Yali Li, and Shengjin Wang. Graphcspn: Geometry-aware depth completion via dynamic gcns. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[46] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 1, 2, 6 +[47] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2019. 1, 2 +[48] Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swaprompt: Test-time prompt adaptation for vision-language models. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2023. 2 +[49] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2022. 2 +[50] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Finetuning image-conditional diffusion models is easier than you think. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 2 + +[51] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 5, 6 +[52] Jin-Hwi Park and Hae-Gon Jeon. A simple yet universal framework for depth completion. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3 +[53] Jin-Hwi Park, Jaesung Choe, Inhwan Bae, and Hae-Gon Jeon. Learning affinity with hyperbolic representation for spatial propagation. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 2 +[54] Jin-Hwi Park, Chanhwi Jeong, Junoh Lee, and Hae-Gon Jeon. Depth prompting for sensor-agnostic depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3 +[55] Nikolay Patakin, Anna Vorontsova, Mikhail Artemyev, and Anton Konushin. Single-stage 3d geometry-preserving depth estimation model training on dataset mixtures with uncalibrated stereo data. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[56] Suraj Patni, Aradhye Agarwal, and Chetan Arora. Ecodepth: Effective conditioning of diffusion models for monocular depth estimation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[57] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2 +[58] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc van Gool. Undepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. 1, 2, 3, 4, 5, 7 +[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 2 +[60] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2019. 1, 2, 3 +[61] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 3 +[62] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya K. Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan + +Wu, Ross B. Girshick, Piotr Doll'ar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2 +[63] Laria Reynolds and Kyle McDonell. Prompt programming for large language models: Beyond the few-shot paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, 2021. 2 +[64] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1 +[65] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[66] Kieran Saunders, George Vogiatzis, and Luis J Manso. Self-supervised monocular depth estimation: Let's talk about the weather. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[67] Ashutosh Saxena, Sung H. Chung, and A. Ng. Learning depth from single monocular images. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2005. 2 +[68] Shreyas S. Shivakumar, Ty Nguyen, Steven W. Chen, and Camillo Jose Taylor. Dfusenet: Deep fusion of rgb and sparse depth information for image guided dense depth completion. IEEE Intelligent Transportation Systems Conference (ITSC), 2019. 2 +[69] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Proceedings of the European Conference on Computer Vision (ECCV), 2012. 6, 8 +[70] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing (TIP), 2019. 2 +[71] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5, 6 +[72] Yifu Tao, Marija Popović, Yiduo Wang, Sundara Tejaswi Digumarti, Nived Chebrolu, and Maurice F. Fallon. 3d lidar reconstruction with probabilistic depth completion for robotic navigation. In Proceedings of the IEEE International Conference on Intelligent Robots and Systems (IROS), 2022. 1 +[73] Jonas Uhrig, N. Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. International Conference on 3D Vision (3DV), 2017. 2 +[74] Massimiliano Viola, Kevin Qu, Nando Metzger, Bingxin Ke, Alexander Becker, Konrad Schindler, and Anton Obukhov. Marigold-dc: Zero-shot monocular depth completion with guided diffusion. arXiv preprint arXiv:2412.13389, 2024. 2, 3, 5, 6, 8 +[75] Yufei Wang, Yuchao Dai, Qi Liu, Peng Yang, Jiadai Sun, and Bo Li. Cu-net: Lidar depth-only completion with coupled u-net. IEEE Robotics and Automation Letters (RA-L), 2022. 2 + +[76] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 6 +[77] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2023. 2 +[78] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters (RA-L), 2020. 6 +[79] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Baobei Xu, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[80] Zhiqiang Yan, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Rignet++: Efficient repetitive image guided network for depth completion. arXiv preprint arXiv:2309.00655, 2023. 2 +[81] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1 +[82] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2024. 7 +[83] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2 +[84] Youming Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and S. Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 6 +[85] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[86] Jiawen Zhu, Simiao Lai, Xin Chen, D. Wang, and Hutchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[87] Yiming Zuo and Jia Deng. Ogni-dc: Robust depth completion with optimization-guided neural iterations. In Proceedings of the European Conference on Computer Vision (ECCV), 2024, 2, 5, 6 \ No newline at end of file diff --git a/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/images.zip b/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9916f5893ab35eb240dd8a1528552adbca689275 --- /dev/null +++ b/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:719fba4ef11dff8dd26977bc77442c58bc0c0391be629e5e59a7d3bb12498118 +size 899817 diff --git a/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/layout.json b/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..956524f16a3bb4a35b3ff4a40dcb0dd510846165 --- /dev/null +++ b/2025/Test-Time Prompt Tuning for Zero-Shot Depth Completion/layout.json @@ -0,0 +1,8884 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 488, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 488, + 121 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 488, + 121 + ], + "type": "text", + "content": "Test-Time Prompt Tuning for Zero-Shot Depth Completion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 142, + 467, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 142, + 467, + 172 + ], + "spans": [ + { + "bbox": [ + 146, + 142, + 467, + 172 + ], + "type": "text", + "content": "Chanhwi Jeong1 Inhwan Bae1 Jin-Hwi Park2* Hae-Gon Jeon3* \n1GIST, 2Chung-Ang University, 3Yonsei University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 92, + 174, + 517, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 174, + 517, + 185 + ], + "spans": [ + { + "bbox": [ + 92, + 174, + 517, + 185 + ], + "type": "text", + "content": "{chanhwij, inhwanbae}@gm.gist.ac.kr, jinhwipark@cau.ac.kr, earboll@yonsei.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 193, + 555, + 285 + ], + "blocks": [ + { + "bbox": [ + 56, + 193, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 56, + 193, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 555, + 285 + ], + "type": "image", + "image_path": "cfa64ab433e1b9170c18b40a0a8f199cf234bd6e3c86a2ea8349236d98564ad5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 555, + 344 + ], + "type": "text", + "content": "Figure 1. Comparison of methods for correcting the inaccurate scale of initial depth estimates produced by the depth foundation model using sensor-derived sparse depth with an accurate scale, represented as yellow dots. (a) The depth foundation model generates an initial depth map from a single RGB input. (b) The least squares fitting aligns the initial depth map with the sparse depth [54]. (c) The parameters of the foundation model are fine-tuned under supervision from the sparse depth. (d) Alternatively, the model's parameters are frozen, and a visual prompt is learned to adjust the depth estimates. Please refer to Fig. 3 for depth profiling results and error analysis." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 361, + 200, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 361, + 200, + 373 + ], + "spans": [ + { + "bbox": [ + 151, + 361, + 200, + 373 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 378, + 297, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 378, + 297, + 642 + ], + "spans": [ + { + "bbox": [ + 53, + 378, + 297, + 642 + ], + "type": "text", + "content": "Zero-shot depth completion using metric scales remains challenging, primarily due to performance limitations such as domain specificity and sensor characteristics. One recent emerging solution is to integrate monocular depth foundation models into depth completion frameworks, yet such efforts still face issues with suboptimal performance and often require further adaptation to the target task. Surprisingly, we find that a simple test-time training, which finetunes monocular depth foundation models on sparse depth measurements from sensors just as it is, yields reasonable results. However, this test-time training obviously incurs high computational costs and introduces biases towards specific conditions, making it impractical for real-world scenarios. In this paper, we introduce a new approach toward parameter-efficient zero-shot depth completion. Our key idea in this work is to leverage visual prompt tuning, achieving sensor-specific depth scale adaptation without forgetting foundational knowledge. Experimental results on diverse datasets demonstrate that our approach outperforms relevant state-of-the-art methods, showing superior generalization and efficiency. Code is publicly available at https://github.com/ch5374/TestPromptDC" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 135, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 135, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 135, + 667 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 672, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 296, + 696 + ], + "type": "text", + "content": "Depth completion, the technique of densifying sparse input measurements from active sensors based on correspond" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 361, + 556, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 458 + ], + "type": "text", + "content": "ing RGB images, is crucial in various applications including 3D reconstruction [72], autonomous driving [39] and robotics [33]. Beyond traditional methods that rely on handcrafted features [34] and optimization [14], data-driven approaches, which are mainly based on convolutional neural networks [27, 46, 47] and vision transformers [64, 84], have achieved the best performance with public benchmarks [15, 25]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 460, + 556, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 591 + ], + "type": "text", + "content": "However, models trained on a single dataset perform well on the test split of that dataset, but exhibit significant performance degradation in unseen environments. The primary reason lies in the scarcity of data available to use for depth completion. Since depth completion datasets require both dense ground truth depth maps and RGB images obtained from well-calibrated sensor fusion setups, large-scale dataset constructions must collect and label image-depth pairs across environmental variations. Obviously, this makes it difficult to achieve generalizable depth completion, including zero-shot inferences." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 714 + ], + "type": "text", + "content": "One interesting research direction in [40, 52, 54] to mitigate the data scarcity is to integrate foundation models for monocular depth estimation [4, 17, 22, 30, 57, 58, 60, 61, 81-83] trained on huge-scale datasets into depth completion for better generalization. This is quite feasible because foundation models can generate approximate metric or relative depth from a single input image as demonstrated in Fig. 1-(a), actual scene depths can be obtained by fitting sparse depth measurements onto these predictions. Unfortunately, there are still two challenges: (1) they require carefully designed" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 146, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 713 + ], + "type": "text", + "content": "*Corresponding authors" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9443" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 204 + ], + "type": "text", + "content": "additional networks, as the existing foundation models struggle to encode sparse depth information; (2) additional training for the depth completion task is also necessary. These two challenges result in either an increase in computational complexity or the introduction of new biases in the training phase. In this paper, we explore one important question: how can we take full advantage of monocular depth foundation models for depth completion? Our work is the first to answer this question by investigating the most effective foundation models, without requiring additional network design or fine-tuning of the pretrained model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 297, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 297, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 297, + 337 + ], + "type": "text", + "content": "We initiate our investigation by proposing a simple yet highly effective strategy, which is fine-tuning the foundation models in the inference phase, only using sparse depth measurements as supervision. As an example in Fig. 1-(b), compared to the sparse depth fitting onto dense depths [40, 52, 54], fine-tuning the foundation models preserves the original depth while allowing the scale to adapt locally (Fig. 1-(c)). Of course, updating all parameters requires significant computing resources and poses the risk of dataset bias, potentially causing models to lose their own knowledge." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 338, + 297, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 297, + 529 + ], + "type": "text", + "content": "As a more efficient alternative, we present a test-time prompt tuning that adapts foundation models at inference time without modifying its parameters (Fig. 1-(d)). We introduce a small set of learnable parameters that guide their prediction while keeping the foundation model frozen. While existing visual prompt [1, 28] uses fixed-size prompts positioned at specific locations (e.g., fixed positions, random placements, or padding) for image classification, we introduce a pixelwise visual prompt specifically designed for depth completion. Compared to previous works that operate at a global or patch level, our pixelwise visual prompt is initialized to zero and added to input images, allowing us to leverage the inherent spatial structures of depth maps. We validate our visual prompt tuning across diverse benchmark datasets and environments, achieving superior zero-shot performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 542, + 139, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 139, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 139, + 555 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 563, + 216, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 216, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 216, + 576 + ], + "type": "text", + "content": "2.1. Monocular Depth Estimation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "Traditional monocular depth estimation depends on handcrafted visual cues [20, 67] and scene-specific assumptions [21, 42]. With the advent of deep learning, subsequent methods [16, 35, 36] using data-driven representations have significantly outperformed the traditional ones reducing the reliance on manually designed features and rigid scene-specific assumptions. Recent works have explored techniques such as data mixing [55, 60, 61] and augmentation [25, 66] to achieve zero-shot generality. Moreover, some works in [10, 17, 30, 50, 56, 85] exploit the generative priors of pretrained diffusion models [65]. However, inferring" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 205 + ], + "type": "text", + "content": "metric-scale depth from single images still remains challenging. To address this, metric-scale depth estimation from monocular images has evolved by incorporating global depth distributions [19], scene-aware conditioning [3] and camera intrinsics [4, 22, 22, 57, 58, 83]. While earlier methods rely on known camera parameters [22, 83], recent approaches estimate focal length [4, 57, 58] directly from learning features, enhancing generalization across diverse environments without any explicit camera metadata. Nevertheless, these methods obviously lag behind approaches that use sensor measurements like LiDAR in terms of accuracy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 217, + 424, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 424, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 424, + 230 + ], + "type": "text", + "content": "2.2. Depth Completion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 237, + 556, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 556, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 556, + 501 + ], + "type": "text", + "content": "Due to the inherent issues with monocular depth estimation, research on unguided depth completion [12, 23, 73, 75], transforming sparse depth data from sensors into dense depth maps without guidance RGB images, has been studied. However, most mainstream depth completion is to use RGB-depth pairs to directly leverage scene contexts such as edges and textures [26, 46, 47, 68, 70, 79, 80, 84, 87]. Furthermore, constructing pixelwise affinity [6, 7, 41, 43, 45, 51, 53] from corresponding RGB images has proven to be effective as it helps preserve depth discontinuity when propagating sparse depth. Despite these advancements, such models often struggle to show the same performance in unseen domains. To mitigate this, recent approaches have leveraged monocular depth foundation models for depth completion. For instance, works in [29, 54] introduces interesting frameworks that exploit features and predict depth from foundation models to perform adaptive depth refinement with sparse depth inputs. Additionally, diffusion-based depth completion methods [24, 74] have adopted the test-time optimization of latent representations, guided by input sparse depths. However, these approaches often require training on additional data [29, 54] or suffer from extremely slow inference speeds [24, 74]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 514, + 498, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 498, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 498, + 528 + ], + "type": "text", + "content": "2.3. Prompting for Foundation Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": "Foundation models [5, 9, 31, 59, 62, 65] exhibit remarkable zero-shot performance across a wide range of tasks as well as depth perception. In natural language processing, prompting schemes have been introduced to leverage these foundation models effectively for downstream tasks [5, 37, 38, 44, 63, 77]. To apply such prompting schemes to the computer vision domain, works in [28, 86] have introduced visual prompt tuning, which incorporates trainable visual tokens into models for specific downstream tasks. Meanwhile, one interesting work in [1] presents learnable padding prompts appended to input images. More recently, visual prompt tuning has also been extended to test-time adaptation scenarios [13, 48, 49] for better adaptation to unseen domains. Inspired by them [1, 49], we propose a test-time prompt tuning strategy for depth completion." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9444" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 173, + 131 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 173, + 131 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 173, + 131 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 173, + 131 + ], + "type": "image", + "image_path": "1a945e7b4c26cb908271025f0d40222334bfb3f4620048d27dbd692a0fc1d723.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 132, + 166, + 140 + ], + "lines": [ + { + "bbox": [ + 63, + 132, + 166, + 140 + ], + "spans": [ + { + "bbox": [ + 63, + 132, + 166, + 140 + ], + "type": "text", + "content": "(a) Supervised training models" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 76, + 310, + 126 + ], + "blocks": [ + { + "bbox": [ + 182, + 76, + 310, + 126 + ], + "lines": [ + { + "bbox": [ + 182, + 76, + 310, + 126 + ], + "spans": [ + { + "bbox": [ + 182, + 76, + 310, + 126 + ], + "type": "image", + "image_path": "ee70203d226eb9cccedaa32f6acbe04e5f707396c58d885efa2bba94661acade.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 130, + 304, + 140 + ], + "lines": [ + { + "bbox": [ + 189, + 130, + 304, + 140 + ], + "spans": [ + { + "bbox": [ + 189, + 130, + 304, + 140 + ], + "type": "text", + "content": "(b) Test-time optimization models" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 328, + 76, + 435, + 126 + ], + "blocks": [ + { + "bbox": [ + 328, + 76, + 435, + 126 + ], + "lines": [ + { + "bbox": [ + 328, + 76, + 435, + 126 + ], + "spans": [ + { + "bbox": [ + 328, + 76, + 435, + 126 + ], + "type": "image", + "image_path": "af5aeed83572c04e80106469f99864c72eabe6bcf42e5217be0105d3e10c28e9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 130, + 429, + 140 + ], + "lines": [ + { + "bbox": [ + 332, + 130, + 429, + 140 + ], + "spans": [ + { + "bbox": [ + 332, + 130, + 429, + 140 + ], + "type": "text", + "content": "(c) Our fine-tuning approach" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 445, + 76, + 553, + 126 + ], + "blocks": [ + { + "bbox": [ + 445, + 76, + 553, + 126 + ], + "lines": [ + { + "bbox": [ + 445, + 76, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 445, + 76, + 553, + 126 + ], + "type": "image", + "image_path": "6a181e379d0decf80db53c867c259f57effe87e5fceb59b8822cff99481a2b5b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 130, + 553, + 140 + ], + "lines": [ + { + "bbox": [ + 437, + 130, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 437, + 130, + 553, + 140 + ], + "type": "text", + "content": "(d) Our visual prompting approach" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 211, + 138, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 138, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 138, + 224 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 232, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 424 + ], + "type": "text", + "content": "In this section, we present our approach for parameter-efficient zero-shot depth completion through test-time training. In Sec. 3.1, we discuss the role of depth foundation models in depth completion, emphasizing their generalization capabilities while identifying key challenges. Next, in Sec. 3.2, we explore a fine-tuning strategy to adapt the foundation model for sparse depth measurements. This approach achieves impressive performance without relying on additional network design or prompt engineering; however, it introduces computational overhead and poses generalization challenges. To address these limitations, in Sec. 3.3, we propose a novel pixel-wise visual prompting method, enabling efficient adaptation without any parameter modification of the foundation model. Finally, in Sec. 3.4, we describe implementation details, including optimization strategies, dataset preprocessing and experimental settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 435, + 291, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 291, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 291, + 448 + ], + "type": "text", + "content": "3.1. Rationale: Usage of Depth Foundation Model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 454, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 525 + ], + "type": "text", + "content": "Recent works have attempted to leverage monocular depth foundation models for depth completion. These models offer strong generalization capabilities and are considered to be attractive solutions for better zero-shot performance. However, incorporating them into depth completion frameworks faces several technical challenges, as described below:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 530, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 296, + 625 + ], + "type": "text", + "content": "Incompatibility with sparse depth input handling. Since monocular depth foundation models are exclusively designed to infer depth from single RGB images, they do not have mechanisms to take sparse depth measurements as guidance. As a result, the existing methods require additional components to process sparse depths and to integrate them into the models, increasing computational complexity as shown in Fig. 2-(a)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Scale mismatch between prediction and sensor depth. Current depth foundation models are able to estimate approximate metric-scale depths [3, 4, 58] as well as relative-scale depth [17, 30, 60, 61] with high-fidelity depth discontinuities. Unfortunately, the depth foundation models for metric-scale inference exhibit systematic scale mismatches when compared to sensor-derived depths. While least squares scale" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 315, + 209, + 394, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 142, + 555, + 197 + ], + "lines": [ + { + "bbox": [ + 55, + 142, + 555, + 197 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 555, + 197 + ], + "type": "text", + "content": "Figure 2. Comparison of the previous foundation model-based depth completion approaches and our approach. (a) Existing methods [40, 52, 54] design additional networks to integrate sensor-based depth measurements into the foundation model. (b) Recent approaches [24, 74] optimize latent representations in diffusion-based foundation models [17, 30] without dense GT but are computationally expensive. (c) Our fine-tuning method adapts the model with sparse depth supervision, and (d) our visual prompting approach enables efficient adaptation while keeping the foundation model frozen." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 209, + 394, + 297 + ], + "lines": [ + { + "bbox": [ + 315, + 209, + 394, + 297 + ], + "spans": [ + { + "bbox": [ + 315, + 209, + 394, + 297 + ], + "type": "image", + "image_path": "b2a7a78d9ad26d5c41b4e35c857d384557f110a32d60c90daca7770c08ca217d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 297, + 393, + 307 + ], + "lines": [ + { + "bbox": [ + 318, + 297, + 393, + 307 + ], + "spans": [ + { + "bbox": [ + 318, + 297, + 393, + 307 + ], + "type": "text", + "content": "(a) Least squares fitting" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 308, + 554, + 363 + ], + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 363 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison of different strategies on the depth map from the data used in Fig. 1. The top row shows the predicted depth maps, while the bottom row presents the depth profiles along the indicated horizontal line. The red " + }, + { + "bbox": [ + 313, + 308, + 554, + 363 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 308, + 554, + 363 + ], + "type": "text", + "content": " marks indicate the sparse depth measurements." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 395, + 209, + 474, + 297 + ], + "blocks": [ + { + "bbox": [ + 395, + 209, + 474, + 297 + ], + "lines": [ + { + "bbox": [ + 395, + 209, + 474, + 297 + ], + "spans": [ + { + "bbox": [ + 395, + 209, + 474, + 297 + ], + "type": "image", + "image_path": "ed49bf74b93ebcd72f25803b4ce9918a24788aa2e185869a03979475c9569076.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 297, + 462, + 307 + ], + "lines": [ + { + "bbox": [ + 413, + 297, + 462, + 307 + ], + "spans": [ + { + "bbox": [ + 413, + 297, + 462, + 307 + ], + "type": "text", + "content": "(b) Fine-tuning" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 476, + 209, + 554, + 297 + ], + "blocks": [ + { + "bbox": [ + 476, + 209, + 554, + 297 + ], + "lines": [ + { + "bbox": [ + 476, + 209, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 476, + 209, + 554, + 297 + ], + "type": "image", + "image_path": "c5778b9833e91ee17ac1158e65636d8a2fbc54a1f8148bd23ae65050e4aa65a3.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 297, + 550, + 307 + ], + "lines": [ + { + "bbox": [ + 484, + 297, + 550, + 307 + ], + "spans": [ + { + "bbox": [ + 484, + 297, + 550, + 307 + ], + "type": "text", + "content": "(c) Visual Prompting" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 383, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 456 + ], + "type": "text", + "content": "fitting in [54] can globally align depth scales, it fails in regions where depth varies significantly. For instance, as shown in Fig. 3, the least squares fitting locally adjusts the correct depth scale at areas near initial depth seeds (e.g., pixel range 100–150), but fails it in other regions (e.g., pixel range 50–100)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 460, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 639 + ], + "type": "text", + "content": "Limitations of existing adaptation strategies. Previous approaches require ground truth (GT) annotations for adaptation, which are costly and labor-intensive (Fig. 2-(a)). For example, works in [40, 54] rely on large-scale datasets, while UniDC [52] adopts a few-shot strategy within the target domain. However, generating new GT labels to cover all scenarios in-the-wild, is impractical, necessitating a more efficient alternative. One promising solution is a test-time optimization, which enables adaptation to new environments without any additional GT annotation (Fig. 2-(b)). Recent contemporary works in [24, 74] have introduced similar strategies by optimizing latent representations in diffusion-based models (Fig. 2-(b)). However, these approaches come with noticeable limitations, including high computational costs and slow inference speeds." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "To address these limitations, we propose fine-tuning depth foundation models in the inference phase, directly using sparse depth measurements as supervision without introducing additional network components (Fig. 2-(c)). This eliminates the dependency on extensive GT data and enables real-time adaptation to new environments." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9445" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 85 + ], + "type": "text", + "content": "3.2. Finetuning monocular depth foundation model" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "text", + "content": " denote a depth foundation model, parametrized by " + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "text", + "content": " is an RGB image. Typically, " + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "text", + "content": " takes only the " + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "text", + "content": " as input and generates a dense depth map " + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "inline_equation", + "content": "D \\in \\mathbb{R}^{1 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 90, + 297, + 139 + ], + "type": "text", + "content": " as output, which can be expressed as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 149, + 143, + 296, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 143, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 149, + 143, + 296, + 156 + ], + "type": "interline_equation", + "content": "F _ {\\theta} (I) = D. \\tag {1}", + "image_path": "675f29dc4f36bc0aa75780da692925d0c05abf3040034307e18721db0896c0d0.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 159, + 296, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 220 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 55, + 159, + 296, + 220 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 159, + 296, + 220 + ], + "type": "text", + "content": " is derived solely from a single image, it inherently suffers from scale ambiguity. To mitigate this issue, a sparse depth " + }, + { + "bbox": [ + 55, + 159, + 296, + 220 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 159, + 296, + 220 + ], + "type": "text", + "content": ", obtained from an active sensor with accurate scales, is employed as supervision for " + }, + { + "bbox": [ + 55, + 159, + 296, + 220 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 55, + 159, + 296, + 220 + ], + "type": "text", + "content": ". This approach allows the model to alleviate the depth scale ambiguity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 221, + 296, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 246 + ], + "type": "text", + "content": "First, to use only valid sparse depth values as supervision, we define a validity mask " + }, + { + "bbox": [ + 55, + 221, + 296, + 246 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 221, + 296, + 246 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 249, + 296, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 249, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 116, + 249, + 296, + 281 + ], + "type": "interline_equation", + "content": "M _ {i, j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} S _ {i, j} \\text {i s v a l i d} \\\\ 0, & \\text {o t h e r w i s e ,} \\end{array} \\right. \\tag {2}", + "image_path": "429874ff1afcb183b5827dbdc574e7b0d2535f4a5f44bcd15a74d2eb88fe4344.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 285, + 251, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 251, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 251, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 285, + 251, + 297 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 285, + 251, + 297 + ], + "type": "text", + "content": " is the pixel coordinate in the image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 298, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 346 + ], + "type": "text", + "content": "With this mask, we initially adjust the scale of the foundation depth by roughly aligning it using the valid sparse depth measurements in the least squares fitting process, as formulated below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 350, + 296, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 350, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 85, + 350, + 296, + 370 + ], + "type": "interline_equation", + "content": "s _ {0}, s _ {1} = \\min _ {\\left(s _ {0}, s _ {1}\\right)} \\| M \\odot \\left(\\left(s _ {0} D + s _ {1}\\right) - S\\right) \\| _ {F}, \\tag {3}", + "image_path": "79771963c536ca32a7d237565ba7af19f3367d7ae00c4adf3696635f4e7a61da.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "content": " denotes an element-wise multiplication. " + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "inline_equation", + "content": "s_0" + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "inline_equation", + "content": "s_1" + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "content": " mean scale and shift values, respectively, and " + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _F" + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "content": " indicates the Frobenius norm. Next, we optimize the model's parameters " + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "content": " using only the valid sparse depth measurements as supervision. The optimal parameters, denoted by " + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "inline_equation", + "content": "\\theta^{*}" + }, + { + "bbox": [ + 55, + 373, + 296, + 446 + ], + "type": "text", + "content": ", are determined by minimizing the following loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 449, + 296, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 449, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 63, + 449, + 296, + 475 + ], + "type": "interline_equation", + "content": "\\theta^ {*} = \\arg \\min _ {\\theta} \\sum_ {i, j} M _ {i, j} \\odot | s _ {0} F _ {\\theta} (I) _ {i, j} + s _ {1} - S _ {i, j} | ^ {p}, \\tag {4}", + "image_path": "e5a95cb01055aca66ef6183acfb42c02585834fc07c7916e61a7aa9cdab984c9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 479, + 297, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 297, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 297, + 526 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 479, + 297, + 526 + ], + "type": "inline_equation", + "content": "p = 1" + }, + { + "bbox": [ + 55, + 479, + 297, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 479, + 297, + 526 + ], + "type": "inline_equation", + "content": "p = 2" + }, + { + "bbox": [ + 55, + 479, + 297, + 526 + ], + "type": "text", + "content": " correspond to the " + }, + { + "bbox": [ + 55, + 479, + 297, + 526 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 479, + 297, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 479, + 297, + 526 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 479, + 297, + 526 + ], + "type": "text", + "content": " losses, respectively and the final loss is the sum of both. To isolate the effects of fine-tuning itself, we do not incorporate specialized loss functions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 537, + 192, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 192, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 192, + 551 + ], + "type": "text", + "content": "3.3. Pixelwise Visual Prompt" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 556, + 297, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 297, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 297, + 640 + ], + "type": "text", + "content": "To address the computational overhead and potential knowledge degradation of the fine-tuning of the foundation model, we propose a pixelwise visual prompt, a novel test-time adaptation method designed specifically for depth completion (Fig. 2-(d)). Unlike conventional fine-tuning, our approach enables efficient adaptation while keeping the model's parameters frozen." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "content": "Given a frozen pretrained foundation model " + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "content": " and a sample consisting of a pair of image and sparse depth measurement " + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "inline_equation", + "content": "(I,S)" + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "content": ", we aim to learn a pixelwise visual prompt " + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "inline_equation", + "content": "V_{\\phi} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "content": ". The prompt is added directly to the input image, forming a prompted image " + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "inline_equation", + "content": "I + V_{\\phi}" + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "content": " and we optimize it by minimizing the depth prediction error" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 444, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 444, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 444, + 84 + ], + "type": "text", + "content": "in the training phase as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 88, + 555, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 88, + 555, + 114 + ], + "spans": [ + { + "bbox": [ + 325, + 88, + 555, + 114 + ], + "type": "interline_equation", + "content": "\\phi^ {*} = \\arg \\min _ {\\phi} \\sum_ {i, j} M _ {i, j} \\odot \\left| F _ {\\theta} \\left(I + V _ {\\phi}\\right) _ {i, j} - S _ {i, j} \\right| ^ {p}. \\tag {5}", + "image_path": "698433a41ab63f3f3fb76d352a0a6640cec03206c58d8c99a0c76000eaa553bc.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 118, + 555, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 555, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 555, + 155 + ], + "type": "text", + "content": "Here, only the prompt parameters " + }, + { + "bbox": [ + 313, + 118, + 555, + 155 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 118, + 555, + 155 + ], + "type": "text", + "content": " are optimized, while " + }, + { + "bbox": [ + 313, + 118, + 555, + 155 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 118, + 555, + 155 + ], + "type": "text", + "content": " remains frozen. Once optimized, the visual prompt " + }, + { + "bbox": [ + 313, + 118, + 555, + 155 + ], + "type": "inline_equation", + "content": "V_{\\phi^*}" + }, + { + "bbox": [ + 313, + 118, + 555, + 155 + ], + "type": "text", + "content": " is applied to the image as below:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 405, + 158, + 555, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 158, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 405, + 158, + 555, + 172 + ], + "type": "interline_equation", + "content": "F _ {\\theta} \\left(I + V _ {\\phi^ {*}}\\right). \\tag {6}", + "image_path": "340b774d8de79eee8b9e824118c35ad0cecdba5206cb798eea6f813ed7dba261.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 175, + 556, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 556, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 556, + 284 + ], + "type": "text", + "content": "In addition to achieving superior performance compared to existing methods, this approach also enhances usability, providing an additional advantage. Since the pixelwise visual prompt is initialized to zero and added to the image, it ensures a minimal level of depth prediction performance even without prompt tuning. Additionally, users can dynamically adjust the time allocated for prompt tuning based on their available resources and the required inference speed, allowing for stable and consistent performance improvement." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 286, + 556, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 556, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 556, + 453 + ], + "type": "text", + "content": "Difference with existing visual prompt methods. While visual prompt tunings [1, 28] have been previously explored to adapt high-level representations using fixed-size prompts at predefined locations, such approaches lack the spatial granularity needed for precise depth refinement. These methods often focus on global or patch-level modifications, making them less effective for tasks requiring fine-grained adjustments, such as depth completion. In contrast, our pixelwise visual prompt is designed to adapt at the individual pixel level, enabling localized corrections that align depth predictions more accurately with sparse sensor measurements. Our approach not only enhances adaptation flexibility but also maintains the strong generalization capabilities of foundation models." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 461, + 448, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 448, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 448, + 474 + ], + "type": "text", + "content": "3.4. Implementation Details" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 479, + 556, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 556, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 556, + 612 + ], + "type": "text", + "content": "Our implementation is based on PyTorch and runs on a single NVIDIA RTX 3090 GPU. We use the AdamW optimizer with " + }, + { + "bbox": [ + 313, + 479, + 556, + 612 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 479, + 556, + 612 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 479, + 556, + 612 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 313, + 479, + 556, + 612 + ], + "type": "text", + "content": ". The learning rate is set to 1e-6 for fine-tuning and 1e-3 for visual prompting. The training process consists of 80 steps for fine-tuning and 50 steps for visual prompting. For all depth foundation models, we use their official pretrained weights and specifically the v2-old version for UniDepthV2 [58], while other models are evaluated in our ablation study. For models that do not support the original image size, we first resize the images while preserving their aspect ratio." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 621, + 391, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 391, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 391, + 635 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "In this section, we begin by outlining the experimental setup in Sec. 4.1, including the evaluation protocols, and introduce the datasets. We then compare our method with state-of-the-art (SOTA) methods for zero-shot depth completion tasks, in Sec. 4.2. Lastly, in Sec. 4.3, we present an ablation study to analyze the impact of our methods." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "9446" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 556, + 373 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 556, + 373 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 556, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 556, + 373 + ], + "type": "image", + "image_path": "db46f0ece6530514232395271c50d03d3db527790e9f90a9745040aecef78c4a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 375, + 550, + 386 + ], + "lines": [ + { + "bbox": [ + 59, + 375, + 550, + 386 + ], + "spans": [ + { + "bbox": [ + 59, + 375, + 550, + 386 + ], + "type": "text", + "content": "Figure 4. Qualitative result and depth error map of various depth completion methods on the IBims-1, VOID 1500, and DDAD datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 392, + 552, + 529 + ], + "blocks": [ + { + "bbox": [ + 56, + 392, + 552, + 529 + ], + "lines": [ + { + "bbox": [ + 56, + 392, + 552, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 552, + 529 + ], + "type": "table", + "html": "
MethodIBims-1VOIDNYUv2KITTIDDAD
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
NLSPN [51]0.0490.1910.4271.3530.4400.7161.3352.0762.4989.231
CompletionFormer [84]0.0580.2060.6421.6840.1860.3740.9521.9352.5189.471
VPP4DC [2]0.0620.2280.1660.6060.0770.2470.4131.6091.3446.781
BP-Net [71]0.0780.2890.2700.742----2.2708.344
OGNI-DC [87]0.0590.1860.1750.593----1.8676.876
Zero-DC [24]0.0780.2030.1870.6070.0690.1580.3971.4131.6817.122
Marigold-DC [74]0.0450.1660.1520.5510.0480.1240.4341.4652.3646.449
Ours-LS0.1130.2210.1870.5710.1230.2161.2052.8015.67611.97
Ours-FT (Sec. 3.2)0.0490.1520.1380.5140.0480.1070.3831.3291.7755.881
Ours-VP (Sec. 3.3)0.0430.1610.1320.5080.0410.1060.3431.3331.4825.811
", + "image_path": "de1800b1d1c6f94b72bb40abf488c7c2726b4800b38d341f4df0145e0fb10c84.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 530, + 555, + 564 + ], + "lines": [ + { + "bbox": [ + 55, + 530, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 555, + 564 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with SOTA methods on five zero-shot benchmarks. Most methods used values reported in [74], while the results for Zero-DC [24] are evaluated based on the DepthFM model [17]. For the same reason as in [74], we do not include results for BP-Net [71] and OGNI-DC [87] on NYUv2 and KITTI. Bold: Best, Underline: Second best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 576, + 173, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 173, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 173, + 590 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": "Our test-time training strategy is closely aligned with recent works [24, 74], which focus on optimizing the model at inference time. Following their methodology, we adapt the model using only sparse depth measurements during test-time, ensuring a fully zero-shot approach while effectively refining depth scale estimation. To maintain a fair evaluation, we strictly exclude the dataset (e.g., ScanNet [8]) used for training the depth foundation model [58] from the test. Furthermore, in our fine-tuning experiments, we prevent the model from accumulating dataset-specific biases by fine-tuning on" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 577, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 555, + 635 + ], + "type": "text", + "content": "a single sample and resetting it to its original parameters before proceeding to the next evaluation. This ensures that our approach remains consistent with the principles of test-time optimization and maintains its generalization ability across diverse environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "We evaluate our proposed method in a zero-shot setting using five real-world datasets that are not seen during the training of the base model [58]. These datasets encompass both indoor and outdoor environments, varying image resolutions, and different levels of sparse depth density, ensuring a comprehensive assessment of generalization ability." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9447" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 69, + 537, + 308 + ], + "blocks": [ + { + "bbox": [ + 62, + 69, + 537, + 308 + ], + "lines": [ + { + "bbox": [ + 62, + 69, + 537, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 69, + 537, + 308 + ], + "type": "image", + "image_path": "4b2f2f1ba073f013763a8a73e279734a2b233cb21263c9303ab8c3bca59cbfad.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 311, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 311, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 555, + 335 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison under varying levels of depth sparsity across NYUv2 and KITTI. Despite significantly reduced depth measurements, our approach recovers high-quality dense depth maps, demonstrating robustness to input sparsity in the depth completion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 344, + 555, + 407 + ], + "blocks": [ + { + "bbox": [ + 56, + 344, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 56, + 344, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 555, + 407 + ], + "type": "table", + "html": "
MethodNYUv2 (200 samples)NYUv2 (100 samples)NYUv2 (50 samples)NYUv2 (5 samples)KITTI (32 channels)KITTI (16 channels)KITTI (8 channels)KITTI (4 channels)
MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓MAE↓RMSE↓
Zero-DC [24]0.0960.1970.1240.2340.1610.2820.4470.6230.5401.9160.6942.3231.0082.9041.6874.184
Marigold-DC [74]0.0730.1750.0900.2030.1120.2360.3580.5410.8752.1930.9872.6121.1863.1761.6694.240
Ours-VP0.0530.1330.0650.1520.0800.1800.1380.2470.3961.5400.4751.8510.5672.0930.8122.644
", + "image_path": "bc10024ed1d80feb86656f803c1face5c090a60d61cfaef5694db2803dbb70ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 409, + 468, + 421 + ], + "lines": [ + { + "bbox": [ + 140, + 409, + 468, + 421 + ], + "spans": [ + { + "bbox": [ + 140, + 409, + 468, + 421 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on varying sparsity. Bold: Best, Underline: Second best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 433, + 295, + 559 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 54, + 433, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 433, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 433, + 295, + 495 + ], + "type": "text", + "content": "(1) IBims-1 [32] is an indoor dataset consisting of 100 RGB-D pairs with a resolution of " + }, + { + "bbox": [ + 54, + 433, + 295, + 495 + ], + "type": "inline_equation", + "content": "480 \\times 640" + }, + { + "bbox": [ + 54, + 433, + 295, + 495 + ], + "type": "text", + "content": ". We use the entire dataset for testing. Since sparse depth data is not provided, we follow [74] to generate a depth valid mask and construct sparse depth by randomly sampling 1,000 depth points." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 498, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 559 + ], + "type": "text", + "content": "(2) VOID [78] contains RGB-D frames from 56 sequences, with 48 designated for training and 8 for testing. The official test set comprises 800 frames, and sparse depth maps are obtained using the VIO system. Following [74], we evaluate our method on this test set using 1,500 sparse depth points." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 563, + 298, + 715 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 54, + 563, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 563, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 54, + 563, + 295, + 661 + ], + "type": "text", + "content": "(1) NYUv2 [69] comprises 464 scenes recorded using a Microsoft Kinect sensor. The official split is divided into 249 training and 215 test scenes. We use the official test set of 654 images for evaluation. Following a commonly used setup [11, 35, 46], we downsample and center-crop the images to " + }, + { + "bbox": [ + 54, + 563, + 295, + 661 + ], + "type": "inline_equation", + "content": "228 \\times 304" + }, + { + "bbox": [ + 54, + 563, + 295, + 661 + ], + "type": "text", + "content": " before conducting experiments. Also, sparse depth is obtained by randomly selecting 500 samples from the depth map." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 665, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 298, + 715 + ], + "type": "text", + "content": "(4) KITTIDC [15] consists of over 90K RGB images paired with LiDAR data. Since the depth obtained from LiDAR is inherently sparse, the ground truth depth is generated by aggregating multiple LiDAR frames. For our experiments," + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "type": "text", + "content": "we use the 1,000 pairs provided for validation. To reduce noise in the input sparse depth, we apply the filtering method used in [76]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 556, + 544 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 472, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 544 + ], + "type": "text", + "content": "(5) DDAD [18] is an outdoor RGB-D driving dataset collected using a Luminar-H2 LiDAR sensor with a resolution of " + }, + { + "bbox": [ + 313, + 472, + 556, + 544 + ], + "type": "inline_equation", + "content": "1216 \\times 1936" + }, + { + "bbox": [ + 313, + 472, + 556, + 544 + ], + "type": "text", + "content": ". It consists of 12K training samples and 40K validation samples. We use that 40K validation samples for evaluation. Also, sparse depth is generated by randomly sampling " + }, + { + "bbox": [ + 313, + 472, + 556, + 544 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 472, + 556, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 546, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 594 + ], + "type": "text", + "content": "Evaluation Metrics. For quantitative evaluation, we employ the following widely used metrics, Root Mean Squared Error (RMSE) and Mean Squared Error (MAE), to assess performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 600, + 509, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 509, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 509, + 613 + ], + "type": "text", + "content": "4.2. Zero-shot Depth Completion Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "We compare our methodology not only with conventional approaches [51, 71, 84] but also with models [2, 24, 74, 87] that focus on improving zero-shot depth completion performance. Notably, as shown in Tab. 1, our simple finetuning methodology (Ours-FT) outperforms many models designed for generalization across most datasets. This result highlights the effectiveness of using sparse depth as supervision, demonstrating that the depth foundation model can" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9448" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 304, + 182 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 304, + 182 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 304, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 304, + 182 + ], + "type": "image", + "image_path": "2a3b85b66d3ba3b79576813b85eba2e3b8f9be959e37658d4d22a57b25596994.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 185, + 555, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 555, + 218 + ], + "type": "text", + "content": "Figure 6. Visualization of our visual prompt, predicted depth, and corresponding error map at each test-time training iteration. As the visual prompt progressively learns from the sparse depth guidance, the predicted depth scale becomes increasingly accurate, which is reflected by the gradual reduction in the error maps." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 553, + 182 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 553, + 182 + ], + "type": "image", + "image_path": "26e95e01c45afef44ee2f5e41a8cce746a03f618c37b188d89bbe1f4809bea53.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 231, + 294, + 346 + ], + "blocks": [ + { + "bbox": [ + 56, + 231, + 294, + 346 + ], + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 346 + ], + "type": "table", + "html": "
MethodIBims-1NYUv2
MAE↓RMSE↓MAE↓RMSE↓
UniDepthV2-LS [58]0.1130.2210.1230.216
UniDepthV2-VP0.0430.1610.0410.106
Metric3dV2-LS [22]0.1100.2100.1740.275
Metric3dV2-VP0.0390.1470.0690.114
DepthAnytingV2-LS [82]0.1430.2620.2190.326
DepthAnytingV2-VP0.0400.1490.0550.122
ZoeDepth-LS [3]0.6230.8380.1740.275
ZoeDepth-VP0.1760.3410.0490.114
", + "image_path": "0ce672cd629eefda500f704ce26381a9db97bca8c3ea13d2952a27addbc5fea8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 409, + 297, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 409, + 297, + 577 + ], + "spans": [ + { + "bbox": [ + 54, + 409, + 297, + 577 + ], + "type": "text", + "content": "be trained to achieve competitive performance. Furthermore, our proposed pixelwise visual prompting method (Ours-VP) surpasses simple fine-tuning in most cases, illustrating that learning visual prompts can be both more efficient and more effective than fine-tuning the entire model. This finding suggests that visual prompting provides a promising alternative to full fine-tuning, offering a favorable balance between computational efficiency and performance. As shown in Fig. 4, the qualitative results indicate that existing SOTA models, which are trained on synthetic data, tend to capture fine-detailed depth. However, their performance is less favorable when examined through error maps, suggesting that accurate scale estimation is a prerequisite for high-quality depth prediction and takes precedence over fine-detail reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": "Robustness to sparsity variations. To further assess our model's adaptability to real-world scenarios, we conduct experiments under varying levels of input sparsity. As shown in Tab. 2, our method consistently outperforms SOTA models even as the number of sparse depth points decreases. Notably, our model exhibits greater robustness to sparse depth bias, which is crucial in real-world applications where the number of available depth measurements can vary significantly. As seen in Fig. 5, comparing methods, scale is often misinterpreted when sparse depth is limited, while our approach maintains accuracy, making it highly suitable for" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 314, + 231, + 553, + 306 + ], + "blocks": [ + { + "bbox": [ + 55, + 347, + 296, + 390 + ], + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 390 + ], + "type": "text", + "content": "Table 3. Performance comparison of our visual prompting method across different depth foundation models. Our approach consistently enhances performance, demonstrating its adaptability to various models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 231, + 553, + 306 + ], + "lines": [ + { + "bbox": [ + 314, + 231, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 314, + 231, + 553, + 306 + ], + "type": "table", + "html": "
MethodMAE ↓RMSE ↓Learnable Parameters
Fine-tuning (Sec. 3.2)0.0490.152100%
Random Patch [1]0.1110.2210.11%
Fixed patch [1]0.0790.1850.11%
Padding patch [1]0.0500.1670.11%
Pixelwise (Sec. 3.3)0.0430.1610.25%
", + "image_path": "b87f007a24d13c8a95e1574abf23502e29c7fa193e8f749c7a0e2fec62db05c3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 307, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 352 + ], + "type": "text", + "content": "Table 4. Comparison of different visual prompting designs in terms of accuracy and parameter efficiency on dataset Ibims-1 [32]. Our pixelwise visual prompt achieves fine-tuning-level performance while significantly reducing learnable parameters." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 364, + 410, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 410, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 410, + 375 + ], + "type": "text", + "content": "real-world deployment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 378, + 556, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 556, + 570 + ], + "type": "text", + "content": "Effectiveness of visual prompt in scale alignment. We next analyze the test-time training process of our pixelwise visual prompt. As shown in Fig. 6, the scale accuracy progressively improves with each test-time training step. When the initial scale is generally well-aligned (left example), the prompt refines predictions gradually, leading to a smooth and consistent improvement. In contrast, when large-scale discrepancies exist due to reflective surfaces like mirrors (right example), the prompt first corrects the dominant errors in these regions before adjusting the overall scale. This adaptive behavior effectively mitigates the scale discrepancy problem discussed in Sec. 3.1. After approximately 10 test-time training iterations, the model stabilizes, achieving a well-calibrated depth scale. This highlights the efficiency of our visual prompt learning process, ensuring rapid convergence with minimal computational overhead." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 576, + 408, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 408, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 408, + 589 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Depth foundation models. We further validate our method by applying it to various depth foundation models beyond UniDepthV2 [58], as shown in Tab. 3. Our visual prompting method consistently improves performance across all foundation models. Interestingly, while UniDepthV2 achieves the best results on NYUv2, Metric3dV2 performs better on IBims-1, suggesting that different foundation models may excel with different datasets. This reinforces the importance of model-agnostic test-time training strategies like visual prompting." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9449" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 295, + 135 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 295, + 135 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 295, + 135 + ], + "type": "table", + "html": "
MethodRequired Memory (MB)Inference Speed (s)
Zero-DC [24]872720.56
Marigold-DC [74]21748121.04
Ours-FT83722.95
Ours-VP39912.93
", + "image_path": "23eeb6afd25eea115d5c1470899518b0147fa1448ca34ac0bfff0af98ba8ec24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 136, + 295, + 159 + ], + "lines": [ + { + "bbox": [ + 55, + 136, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 295, + 159 + ], + "type": "text", + "content": "Table 5. Comparison of test-time training methods measured on the NYUv2 dataset [69]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 174, + 297, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 297, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 297, + 365 + ], + "type": "text", + "content": "Prompting design. We compare our pixelwise visual prompt against three conventional visual prompting methods: random patches, fixed patches, and padding-based prompts. As shown in Tab. 4, all prompting techniques reduce computational costs by freezing foundation model parameters, but existing prompting methods as proposed by [1] fail to match the performance of fine-tuning. This suggests a trade-off between parameter efficiency and accuracy in conventional prompting techniques. In contrast, our pixelwise visual prompt bridges this gap, achieving fine-tuning-level performance while maintaining efficiency. This is because our method preserves fine-grained spatial adjustments, ensuring that depth refinements occur at the pixel level rather than through global modifications. These results highlight the effectiveness of our approach in maximizing both performance and efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 513 + ], + "type": "text", + "content": "Computational efficiency. We compare our approach with existing test-time optimization methods, including Zero-DC [24] and Marigold-DC [74], to highlight its efficiency in memory usage and inference speed. As shown in Tab. 5, our pixelwise visual prompting (Ours-VP) requires significantly less memory (3991 MB) and achieves a fast inference speed of 2.93s, outperforming Zero-DC (20.56s) and Marigold-DC (121.04s). Even our fine-tuning variant (Ours-FT) remains computationally efficient. These results demonstrate that our method avoids the heavy computational cost of full fine-tuning while adapting effectively, making it highly practical for real-world use." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 673 + ], + "type": "text", + "content": "Learning rate & test-time training step. In Fig. 7, we analyze the learning behavior of fine-tuning versus visual prompting. Fine-tuning exhibits unstable learning dynamics, particularly when using a high learning rate, as abrupt parameter shifts distort depth priors and reduce prediction stability in early iterations. Additionally, fine-tuning requires carefully tuned learning rates to prevent performance degradation. In contrast, our visual prompt method is robust to different learning rates, consistently achieving stable and rapid convergence. As seen in Fig. 7, visual prompting requires significantly fewer iterations to reach a reliable depth estimation, making it more efficient and practical for real-time applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Analysis of structures of learned visual prompt. Lastly, we analyze the characteristics of learned patterns in the visual prompts. As shown in Fig. 8, regions belonging to differ" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 541, + 152 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 541, + 152 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 541, + 152 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 541, + 152 + ], + "type": "image", + "image_path": "faa3f2119c3b540808cedc0165dc1e866163e9a9b0f13ad0590736983aaa6271.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 152, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 217 + ], + "type": "text", + "content": "Figure 7. Comparison of fine-tuning and visual prompting methods. The fine-tuning method shows unstable convergence at high learning rates, while the visual prompting method remains stable across different rates and converges faster. The dotted line indicates the baseline using least squares fitting, while visual prompting achieves superior performance with fewer iterations." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 220, + 380, + 269 + ], + "blocks": [ + { + "bbox": [ + 315, + 220, + 380, + 269 + ], + "lines": [ + { + "bbox": [ + 315, + 220, + 380, + 269 + ], + "spans": [ + { + "bbox": [ + 315, + 220, + 380, + 269 + ], + "type": "image", + "image_path": "f00c7d6a531235052dec6e6bc5804e6f4c0e76a6e3f9e29a9d7d55c6c4c18b47.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 269, + 374, + 277 + ], + "lines": [ + { + "bbox": [ + 327, + 269, + 374, + 277 + ], + "spans": [ + { + "bbox": [ + 327, + 269, + 374, + 277 + ], + "type": "text", + "content": "(a) RGB image" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 281, + 555, + 315 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 315 + ], + "type": "text", + "content": "Figure 8. Visualization of the RGB image, visual prompt, and the combined RGB+Visual Prompt. The colored boxes highlight key regions with structural differences and similarities." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 380, + 220, + 465, + 268 + ], + "blocks": [ + { + "bbox": [ + 380, + 220, + 465, + 268 + ], + "lines": [ + { + "bbox": [ + 380, + 220, + 465, + 268 + ], + "spans": [ + { + "bbox": [ + 380, + 220, + 465, + 268 + ], + "type": "image", + "image_path": "14d0e7f05248035dcf778ca1c45328a703582c37b415ab456451cdb7c1689865.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 268, + 463, + 277 + ], + "lines": [ + { + "bbox": [ + 409, + 268, + 463, + 277 + ], + "spans": [ + { + "bbox": [ + 409, + 268, + 463, + 277 + ], + "type": "text", + "content": "(b) Visual Prompt" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 465, + 220, + 490, + 268 + ], + "blocks": [ + { + "bbox": [ + 465, + 220, + 490, + 268 + ], + "lines": [ + { + "bbox": [ + 465, + 220, + 490, + 268 + ], + "spans": [ + { + "bbox": [ + 465, + 220, + 490, + 268 + ], + "type": "image", + "image_path": "2c151595d74f290b6b2cfe49350abd2900fe9e2c797e2f0e80aa1a9dadc0a76d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 480, + 268, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 480, + 268, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 480, + 268, + 553, + 277 + ], + "type": "text", + "content": "(c) RGB+Visual Prompt" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 336, + 556, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 556, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 556, + 455 + ], + "type": "text", + "content": "ent planes, shown in the yellow and green boxes, show that entirely different visual prompt structures are required to accurately adjust distances between the two walls. Conversely, regions on the same plane, shown in cyan and magenta boxes, have relatively similar structures, introducing the same influence to the input. To represent the metric-aware 3D geometry, these visual prompts are learned to adjust the depth to align with sensor measurements, enabling high-fidelity depth estimation while keeping the depth foundation model frozen." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 388, + 479 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 486, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 617 + ], + "type": "text", + "content": "This study explores how to effectively utilize foundation models for zero-shot depth completion. We find that simply fine-tuning the depth foundation model with sparse depth measurements outperforms existing methods, demonstrating strong adaptability. However, fine-tuning introduces computational overhead and risks knowledge forgetting. To address this, we propose pixelwise visual prompt tuning, which efficiently refines depth predictions without modifying model parameters. Our method achieves fine-tuning-level performance while maintaining efficiency, making it well-suited for zero-shot adaptation across diverse environments." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 621, + 557, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 557, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 557, + 705 + ], + "type": "text", + "content": "Acknowledgement This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2024-00338439) and the MSIT(Ministry of Science and ICT), Korea, under the Graduate School of Metaverse Convergence support program(ITP-2025-RS-2024-00418847) supervised by the IITP(Institute for Information & Communications Technology Planning & Evaluation)" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9450" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 133 + ], + "type": "text", + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022. 2, 4, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Luca Bartolomei, Matteo Poggi, Andrea Conti, Fabio Tosi, and Stefano Mattoccia. Revisiting depth completion from a stereo matching perspective for cross-domain generalization. In 2024 International Conference on 3D Vision (3DV), 2024. 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 296, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 296, + 235 + ], + "type": "text", + "content": "[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 280 + ], + "type": "text", + "content": "[4] Alexey Bochkovskiy, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R. Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 296, + 401 + ], + "type": "text", + "content": "[5] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Ma teusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. Proceedings of the Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 402, + 296, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 296, + 445 + ], + "type": "text", + "content": "[6] Xinjing Cheng, Peng Wang, and Ruigang Yang. Learning depth with convolutional spatial propagation network. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 447, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 501 + ], + "type": "text", + "content": "[7] Xinjing Cheng, Peng Wang, Chenye Guan, and Ruigang Yang. Cspn++: Learning context and resource aware convolutional spatial propagation networks for depth completion. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 502, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 555 + ], + "type": "text", + "content": "[8] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 557, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 601 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 602, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 645 + ], + "type": "text", + "content": "[10] Yiqun Duan, Xianda Guo, and Zhengbiao Zhu. Diffusion-depth: Diffusion denoising approach for monocular depth estimation. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 690 + ], + "type": "text", + "content": "[11] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Proceedings of the Neural Information Processing Systems (NeurIPS), 2014. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "text", + "content": "[12] Abdelrahman Eldisokey, Michael Felsberg, Karl Holmquist, and Mikael Persson. Uncertainty-aware cnns for depth com" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "pletion: Uncertainty from beginning to end. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "text", + "content": "[13] Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "text", + "content": "[14] David Ferstl, Christian Reinbacher, René Ranftl, Matthias Rüther, and Horst Bischof. Image guided depth upsampling using anisotropic total generalized variation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2013. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 262 + ], + "type": "text", + "content": "[15] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 1, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 555, + 308 + ], + "type": "text", + "content": "[16] Clément Godard, Oisin Mac Aodha, and Gabriel J. Brostow. Unsupervised monocular depth estimation with left-right consistency. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 310, + 554, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 554, + 375 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 554, + 375 + ], + "type": "text", + "content": "[17] Ming Gui, Johannes S. Fischer, Ulrich Prestel, Pingchuan Ma, Dmytro Kotovenko, Olga Grebenkova, Stefan Andreas Baumann, Vincent Tao Hu, and Bjorn Ommer. Depthfm: Fast monocular depth estimation with flow matching. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025. 1, 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 376, + 555, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 555, + 430 + ], + "type": "text", + "content": "[18] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 433, + 555, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 555, + 465 + ], + "type": "text", + "content": "[19] Vitor Campagholo Guizilini, Pavel Tokmakov, Achal Dave, and Rares Ambrus. Grin: Zero-shot metric depth with pixel-level diffusion. arXiv preprint arXiv:2409.09896, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 467, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 555, + 498 + ], + "type": "text", + "content": "[20] Derek Hoiem, Alexei A. Efros, and Martial Hebert. Automatic photo pop-up. Proceedings of ACM SIGGRAPH, 2005. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 501, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 555, + 533 + ], + "type": "text", + "content": "[21] Derek Hoiem, Alexei A. Efros, and Martial Hebert. Recovering surface layout from an image. International Journal on Computer Vision (IJCV), 2007. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 555, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 555, + 601 + ], + "type": "text", + "content": "[22] Mu Hu, Wei Yin, China. Xiaoyan Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2024. 1, 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 602, + 555, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 555, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 555, + 655 + ], + "type": "text", + "content": "[23] Zixuan Huang, Junming Fan, Shenggan Cheng, Shuai Yi, Xiaogang Wang, and Hongsheng Li. Hms-net: Hierarchical multi-scale sparsity-invariant network for sparse depth completion. IEEE Transactions on Image Processing (TIP), 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[24] Lee Hyoseok, Kyeong Seon Kim, Kwon Byung-Ki, and TaeHyun Oh. Zero-shot depth completion via test-time alignment with affine-invariant depth prior. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2025. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9451" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 127 + ], + "type": "text", + "content": "[25] Dmitry Yu. Ignatov, Andrey D. Ignatov, and Radu Timofte. Virtually enriched nyu depth v2 dataset for monocular depth estimation: Do we need artificial augmentation? Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop (CVPRW), 2024. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "type": "text", + "content": "[26] Saif Muhammad Imran, Yunfei Long, Xiaoming Liu, and Daniel Morris. Depth coefficients for depth completion. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "text", + "content": "[27] Maximilian Jaritz, Raoul de Charette, Émilie Wirbel, Xavier Perrotton, and Fawzi Nashashibi. Sparse and dense data with cnns: Depth completion and semantic segmentation. 2018 International Conference on 3D Vision (3DV), 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 262 + ], + "type": "text", + "content": "[28] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 308 + ], + "type": "text", + "content": "[29] Jinyoung Jun, Jae-Han Lee, and Chang-Su Kim. Masked spatial propagation network for sparsity-adaptive depth refinement. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 296, + 374 + ], + "type": "text", + "content": "[30] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 442 + ], + "type": "text", + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloé Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólár, and Ross B. Girshick. Segment anything. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 487 + ], + "type": "text", + "content": "[32] Tobias Koch, Lukas Liebel, Friedrich Fraundorfer, and Marco Korner. Evaluation of cnn-based single-image depth estimation methods. In Proceedings of the European Conference on Computer Vision Workshop (ECCVW), 2018. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 533 + ], + "type": "text", + "content": "[33] Torsten Kröger. Opening the door to new sensor-based robot applications—the reflexes motion libraries. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2011. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "type": "text", + "content": "[34] Jason Ku, Ali Harakeh, and Steven L. Waslander. In defense of classical image processing: Fast depth completion on the cpu. 2018 15th Conference on Computer and Robot Vision (CRV), 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "type": "text", + "content": "[35] Iro Laina, C. Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In Proceedings of the International Conference on 3D Vision (3DV), 2016. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 667 + ], + "type": "text", + "content": "[36] Jin Han Lee, Myung-Kyu Han, Dong Wook Ko, and Il Hong Suh. From big to small: Multi-scale local planar guidance for monocular depth estimation. arXiv preprint arXiv:1907.10326, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[37] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[38] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 555, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 160 + ], + "type": "text", + "content": "[39] Ming Liang, Binh Yang, Shenlong Wang, and Raquel Urtasun. Deep continuous fusion for multi-sensor 3d object detection. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 555, + 227 + ], + "type": "text", + "content": "[40] Haotong Lin, Sida Peng, Jingxiao Chen, Songyou Peng, Ji-aming Sun, Minghuan Liu, Hujun Bao, Jiashi Feng, Xiaowei Zhou, and Bingyi Kang. Prompting depth anything for 4k resolution accurate metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 271 + ], + "type": "text", + "content": "[41] Yuankai Lin, Hua Yang, Tao Cheng, Wending Zhou, and Zhouping Yin. Dyspn: Learning dynamic affinity for image-guided depth completion. IEEE Transactions on Circuits and Systems for Video Technology (TCSVT), 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 273, + 555, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 555, + 316 + ], + "type": "text", + "content": "[42] Beyang Liu, Stephen Gould, and Daphne Koller. Single image depth estimation from predicted semantic labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 317, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 554, + 360 + ], + "type": "text", + "content": "[43] Sifei Liu, Shalini De Mello, Jinwei Gu, Guangyu Zhong, Ming-Hsuan Yang, and Jan Kautz. Learning affinity via spatial propagation networks. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 361, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 554, + 415 + ], + "type": "text", + "content": "[44] Xiao Liu, Kaixuan Ji, Yicheng Fu, Zhengxiao Du, Zhilin Yang, and Jie Tang. P-tuning v2: Prompt tuning can be comparable to fine-tuning universally across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association of Computational Linguistics (ACL), 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 459 + ], + "type": "text", + "content": "[45] Xin Liu, Xiaofei Shao, Boqian Wang, Yali Li, and Shengjin Wang. Graphcspn: Geometry-aware depth completion via dynamic gcns. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 460, + 554, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 503 + ], + "type": "text", + "content": "[46] Fangchang Ma and Sertac Karaman. Sparse-to-dense: Depth prediction from sparse depth samples and a single image. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2018. 1, 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 505, + 554, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 554, + 557 + ], + "type": "text", + "content": "[47] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 554, + 602 + ], + "type": "text", + "content": "[48] Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swaprompt: Test-time prompt adaptation for vision-language models. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 604, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 554, + 658 + ], + "type": "text", + "content": "[49] Shu Manli, Nie Weili, Huang De-An, Yu Zhiding, Goldstein Tom, Anandkumar Anima, and Xiao Chaowei. Test-time prompt tuning for zero-shot generalization in vision-language models. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 659, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 554, + 713 + ], + "type": "text", + "content": "[50] Gonzalo Martin Garcia, Karim Abou Zeid, Christian Schmidt, Daan de Geus, Alexander Hermans, and Bastian Leibe. Finetuning image-conditional diffusion models is easier than you think. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9452" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "text", + "content": "[51] Jinsun Park, Kyungdon Joo, Zhe Hu, Chi Liu, and In So Kweon. Non-local spatial propagation network for depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2020. 2, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "text", + "content": "[52] Jin-Hwi Park and Hae-Gon Jeon. A simple yet universal framework for depth completion. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 207 + ], + "type": "text", + "content": "[53] Jin-Hwi Park, Jaesung Choe, Inhwan Bae, and Hae-Gon Jeon. Learning affinity with hyperbolic representation for spatial propagation. In Proceedings of the International Conference on Machine Learning (ICML), 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 253 + ], + "type": "text", + "content": "[54] Jin-Hwi Park, Chanhwi Jeong, Junoh Lee, and Hae-Gon Jeon. Depth prompting for sensor-agnostic depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 319 + ], + "type": "text", + "content": "[55] Nikolay Patakin, Anna Vorontsova, Mikhail Artemyev, and Anton Konushin. Single-stage 3d geometry-preserving depth estimation model training on dataset mixtures with uncalibrated stereo data. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "type": "text", + "content": "[56] Suraj Patni, Aradhye Agarwal, and Chetan Arora. Ecodepth: Effective conditioning of diffusion models for monocular depth estimation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 296, + 432 + ], + "type": "text", + "content": "[57] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 487 + ], + "type": "text", + "content": "[58] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc van Gool. Undepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 296, + 566 + ], + "type": "text", + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning (ICML), 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 296, + 623 + ], + "type": "text", + "content": "[60] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2019. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "type": "text", + "content": "[61] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "text", + "content": "[62] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya K. Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 106 + ], + "type": "text", + "content": "Wu, Ross B. Girshick, Piotr Doll'ar, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 555, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 152 + ], + "type": "text", + "content": "[63] Laria Reynolds and Kyle McDonell. Prompt programming for large language models: Beyond the few-shot paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "type": "text", + "content": "[64] Kyeongha Rho, Jinsung Ha, and Youngjung Kim. Guideformer: Transformers for image guided depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 252 + ], + "type": "text", + "content": "[65] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 555, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 297 + ], + "type": "text", + "content": "[66] Kieran Saunders, George Vogiatzis, and Luis J Manso. Self-supervised monocular depth estimation: Let's talk about the weather. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 331 + ], + "type": "text", + "content": "[67] Ashutosh Saxena, Sung H. Chung, and A. Ng. Learning depth from single monocular images. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2005. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 555, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 387 + ], + "type": "text", + "content": "[68] Shreyas S. Shivakumar, Ty Nguyen, Steven W. Chen, and Camillo Jose Taylor. Dfusenet: Deep fusion of rgb and sparse depth information for image guided dense depth completion. IEEE Intelligent Transportation Systems Conference (ITSC), 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "text", + "content": "[69] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Proceedings of the European Conference on Computer Vision (ECCV), 2012. 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 555, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 466 + ], + "type": "text", + "content": "[70] Jie Tang, Fei-Peng Tian, Wei Feng, Jian Li, and Ping Tan. Learning guided convolutional network for depth completion. IEEE Transactions on Image Processing (TIP), 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 468, + 555, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 555, + 511 + ], + "type": "text", + "content": "[71] Jie Tang, Fei-Peng Tian, Boshi An, Jian Li, and Ping Tan. Bilateral propagation network for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 512, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 555, + 578 + ], + "type": "text", + "content": "[72] Yifu Tao, Marija Popović, Yiduo Wang, Sundara Tejaswi Digumarti, Nived Chebrolu, and Maurice F. Fallon. 3d lidar reconstruction with probabilistic depth completion for robotic navigation. In Proceedings of the IEEE International Conference on Intelligent Robots and Systems (IROS), 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 555, + 612 + ], + "type": "text", + "content": "[73] Jonas Uhrig, N. Schneider, Lukas Schneider, Uwe Franke, Thomas Brox, and Andreas Geiger. Sparsity invariant cnns. International Conference on 3D Vision (3DV), 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 667 + ], + "type": "text", + "content": "[74] Massimiliano Viola, Kevin Qu, Nando Metzger, Bingxin Ke, Alexander Becker, Konrad Schindler, and Anton Obukhov. Marigold-dc: Zero-shot monocular depth completion with guided diffusion. arXiv preprint arXiv:2412.13389, 2024. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "type": "text", + "content": "[75] Yufei Wang, Yuchao Dai, Qi Liu, Peng Yang, Jiadai Sun, and Bo Li. Cu-net: Lidar depth-only completion with coupled u-net. IEEE Robotics and Automation Letters (RA-L), 2022. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9453" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 664 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[76] Yufei Wang, Bo Li, Ge Zhang, Qi Liu, Tao Gao, and Yuchao Dai. Lrru: Long-short range recurrent updating networks for depth completion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 184 + ], + "type": "text", + "content": "[77] Yuxin Wen, Neel Jain, John Kirchenbauer, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Hard prompts made easy: Gradient-based discrete optimization for prompt tuning and discovery. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 217 + ], + "type": "text", + "content": "[78] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters (RA-L), 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 295, + 262 + ], + "type": "text", + "content": "[79] Zhiqiang Yan, Kun Wang, Xiang Li, Zhenyu Zhang, Baobei Xu, Jun Li, and Jian Yang. Rignet: Repetitive image guided network for depth completion. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 305 + ], + "type": "text", + "content": "[80] Zhiqiang Yan, Xiang Li, Zhenyu Zhang, Jun Li, and Jian Yang. Rignet++: Efficient repetitive image guided network for depth completion. arXiv preprint arXiv:2309.00655, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "type": "text", + "content": "[81] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 407 + ], + "type": "text", + "content": "[82] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In Proceedings of the Neural Information Processing Systems (NeurIPS), 2024. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 295, + 463 + ], + "type": "text", + "content": "[83] Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 465, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 295, + 520 + ], + "type": "text", + "content": "[84] Youming Zhang, Xianda Guo, Matteo Poggi, Zheng Zhu, Guan Huang, and S. Mattoccia. Completionformer: Depth completion with convolutions and vision transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 521, + 295, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 295, + 573 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 295, + 573 + ], + "type": "text", + "content": "[85] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 576, + 295, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 295, + 620 + ], + "type": "text", + "content": "[86] Jiawen Zhu, Simiao Lai, Xin Chen, D. Wang, and Hutchuan Lu. Visual prompt multi-modal tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 622, + 295, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 622, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 622, + 295, + 664 + ], + "type": "text", + "content": "[87] Yiming Zuo and Jia Deng. Ogni-dc: Robust depth completion with optimization-guided neural iterations. In Proceedings of the European Conference on Computer Vision (ECCV), 2024, 2, 5, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9454" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/abcad282-0784-41dd-b230-4b9b3488f3c1_content_list.json b/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/abcad282-0784-41dd-b230-4b9b3488f3c1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..18624b2ed95aa5d328948b625c3feca3b747e239 --- /dev/null +++ b/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/abcad282-0784-41dd-b230-4b9b3488f3c1_content_list.json @@ -0,0 +1,1609 @@ +[ + { + "type": "text", + "text": "Test-Time Retrieval-Augmented Adaptation for Vision-Language Models", + "text_level": 1, + "bbox": [ + 130, + 130, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinqi Fan $^{1}$ , Xueli Chen $^{2}$ , Luoxiao Yang $^{3}$ , Chuin Hong Yap $^{1}$ , Rizwan Qureshi $^{4}$ , Qi Dou $^{5}$ , Moi Hoon Yap $^{1}$ , Mubarak Shah $^{4}$", + "bbox": [ + 184, + 179, + 810, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Manchester Metropolitan University $^{2}$ Hong Kong Metropolitan University $^{3}$ Xi'an University of Technology $^{4}$ University of Central Florida $^{5}$ Chinese University of Hong Kong", + "bbox": [ + 107, + 215, + 888, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language models (VLMs) have shown promise in test-time adaptation tasks due to their remarkable capabilities in understanding and reasoning about visual content through natural language descriptions. However, training VLMs typically demands substantial computational resources, and they often struggle to adapt efficiently to new domains or tasks. Additionally, dynamically estimating the test distribution from streaming data at test time remains a significant challenge. In this work, we propose a novel test-time retrieval-augmented adaptation (TT-RAA) method that enables VLMs to maintain high performance across diverse visual recognition tasks without the need for task-specific training or large computational overhead. During inference, TT-RAA employs a streaming mixture of Gaussian database (SMGD) to continuously estimate test distributions, requiring minimal storage. Then, TT-RAA retrieves the most relevant information from the SMGD, enhancing the original VLM outputs. A key limitation of CLIP-based VLMs is their inter-modal vision-language optimization, which does not optimize vision-space similarity, leading to larger intra-modal variance. To address this, we propose a multimodal retrieval augmentation module that transforms the SMGD into a unified multimodal space, enabling retrieval that aligns both vision and language modalities. Extensive experiments across both cross-domain and out-of-distribution benchmarks comprising fourteen datasets demonstrate TT-RAA's superior performance compared to state-of-the-art methods. Ablation studies and hyperparameter analyses further validate the effectiveness of the proposed modules. The source code of our work is available at https://github.com/xinqi-fan/TT-RAA.", + "bbox": [ + 89, + 316, + 483, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 799, + 220, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language models (VLMs) have significantly advanced visual understanding by leveraging natural language supervision with contrastive language-image pre-training (CLIP) notably pioneering this field [34]. The zero-shot capability enables this model to be deployed in various do", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3deb82f03189eccad0fff94c65752ec7775e644dc2e4c0a9a72c340c33e33994.jpg", + "image_caption": [ + "Figure 1. Illustration of our proposed test time retrieval augmented adaptation (TT-RAA). Without TT-RAA, the CLIP prediction will be the final prediction. With TT-RAA, the image feature will be used to retrieve the most similar database centers which will be used to improve the final prediction in a training-free manner." + ], + "image_footnote": [], + "bbox": [ + 526, + 289, + 895, + 454 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mains. This foundational work has catalyzed a wave of architectural innovations, including domain-specific variants and architectural enhancements that build upon CLIP's core principles [52]. Some variants have focused on scaling up capacities, learning from noisy text supervision, and fine-grained alignment [16, 50]. However, these models consistently encounter challenges when deployed in specialized domains, where the distribution of images often differs substantially from their web-scale training data [11, 33]. This challenge has become even more prominent with the recent introduction of more sophisticated large VLMs such as GPT-4V [31], LLaVA [25], and Llama 3.2 Vision [1]. While these advanced models excel at integrating visual and textual data, the high computational cost of fine-tuning them for domain-specific tasks poses a major challenge, especially in specialized visual domains that differ significantly from general web images.", + "bbox": [ + 511, + 550, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Test-time adaptation (TTA) has emerged as a promising solution that enables models to adapt to new domains without accessing source domain training data [23]. Traditional TTA approaches include self-training-based methods that leverage pseudo-labels [22, 41, 49], self-supervised learning-based methods that incorporate aux", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8810", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "iliary tasks [42], batch normalization [28, 37], and entropy minimization-based methods [30, 46]. For VLMs, recent works have explored prompt-based adaptation methods [10, 38] to bridge the domain gap between web-scale training data and specific downstream tasks. However, these methods often incur substantial computational overhead, highlighting the need for more efficient approaches [10, 17]. Advancements in training-free methods have explored the use of cache mechanisms [17, 55] and distribution-based methods [51, 58] to improve the efficiency. Despite these advances, dynamically estimating the test distribution from streaming data at test time remains a significant challenge, due to the test sample arriving one at a time.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce test-time retrieval-augmented adaptation (TT-RAA), a novel training-free method that allows VLMs to sustain strong performance across various visual recognition tasks without requiring task-specific training or high computational costs (Fig. 1). TT-RAA is a training-free framework, which uses a streaming mixture of Gaussian database (SMGD) to dynamically capture test distribution characteristics with minimal storage requirements at inference time. TT-RAA retrieves relevant information from the SMGD to enhance the performance of the original VLM. Furthermore, an inherent limitation of CLIP-based VLMs is their lack of optimization for intramodal similarity, which does not optimize similarity in the vision space, resulting in greater intramodal variance [24]. To mitigate this, we introduce a multimodal retrieval augmentation (MRA) module that projects the SMGD into a shared multimodal embedding space, facilitating retrieval that harmonizes vision and language modalities. Extensive testing across ten cross-domain datasets and four out-of-distribution datasets confirms that TT-RAA delivers superior performance relative to state-of-the-art methods, with additional ablation studies and hyperparameter analyses supporting the effectiveness of the proposed modules. The key contributions of this work include:", + "bbox": [ + 91, + 291, + 483, + 652 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce test-time retrieval-augmented adaptation (TT-RAA), an efficient, training-free framework that significantly improves domain adaptability without additional fine-tuning.", + "- We propose streaming mixture of Gaussian database (SMGD) for dynamic, efficient, and robust estimation of test domain statistics during inference.", + "- We propose multimodal retrieval augmentation (MRA), a novel mechanism addressing CLIP's inherent intra-modal similarity limitation, aligning both vision and language modalities effectively." + ], + "bbox": [ + 89, + 656, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The rest of this paper is organized as follows: Section 2 provides some related works, Section 3 describes the proposed method. Experimental results and performance evaluation are presented in Section 4. Finally, Section 5 concludes the article and outlines future research directions.", + "bbox": [ + 89, + 824, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Vision Language Model", + "text_level": 1, + "bbox": [ + 513, + 114, + 732, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The development of vision language models (VLMs) can be traced back to earlier works that explored vision-language pre-training for specific tasks, where VisualBERT [21] and UNITER [5] pioneered joint representation learning between visual and textual modalities for tasks like visual question answering. The field witnessed a paradigm shift with CLIP [34], which demonstrated impressive zero-shot capabilities by learning transferable visual representations through contrastive learning between images and texts at web scale. ALIGN [16] further scaled up this approach with 1.8 billion noisy image-text pairs and introduced noise-robust training strategies. FILIP [50] enhanced CLIP by introducing fine-grained region-word alignment, enabling more detailed vision-language correspondences. FLAVA [39] introduced a single foundation model that can tackle vision, language, and multimodal tasks. More recently, the emergence of multimodal large language models has pushed the boundaries of vision-language understanding, where GPT-4V [31] extends GPT-4's language capabilities to handle visual inputs, LLaVA [25] achieves strong visual conversation abilities by aligning vision encoders with language models through minimal task-specific training, and Llama 3.2 [1] demonstrates remarkable vision and language capabilities with open models. Other notable VLMs include VinVL [53], which enhanced visual understanding by incorporating object detection and segmentation, and VisualGPT [4], which demonstrated zero-shot transfer to various vision-language tasks. These models have continued to push the boundaries of multimodal representation learning, enabling a wide range of applications from intelligent assistants to creative tools. While these newer models demonstrate impressive capabilities in combining visual and textual understanding, they face similar domain adaptation challenges but at an even larger scale. The computational demands for fine-tuning these models on domain-specific tasks present a significant barrier, particularly for specialized applications where the visual domain differs markedly from general web imagery.", + "bbox": [ + 511, + 138, + 906, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Test Time Adaptation", + "text_level": 1, + "bbox": [ + 513, + 726, + 718, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Normal domain adaptation techniques require access to both source and target domain data, but test-time adaptation (TTA) adapts models to the target domain without source data, making it particularly useful for scenarios with privacy or legal restrictions [23]. To adapt a pre-trained model to a target domain, methods like self-training refine model predictions using pseudo-labels. Approaches such as centroid-based [22], neighbor-based [49], and optimization-based [41] strategies enhance the quality of these pseudolabels. Self-supervised methods introduce auxiliary tasks to", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8811", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "align features, exemplified by test-time training (TTT) [42] with a self-supervised rotation prediction task. Batch normalization [28, 37] techniques and entropy minimization strategies [30, 46] have also been adopted in TTA tasks.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent advancements focus on VLMs like CLIP, using domain-specific prompts from test data. Test-time prompt tuning (TPT) [38] fine-tunes a learnable prompt with each testing sample, while DiffTPT [10] uses pre-trained diffusion models to enhance test data diversity. Although these methods are promising, they still require a large amount of computational resources for their training-based adaptation, making it hard to adapt large VLMs. To reduce costs, a training-free dynamic adapter (TDA) [17] has been introduced, using a lightweight dynamic key-value cache without backpropagation. However, its cache only captures the streaming test data with a limited cache capacity without the ability to estimate the full spectrum of the test data.", + "bbox": [ + 89, + 152, + 482, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Retrieval Augmented Generation", + "text_level": 1, + "bbox": [ + 89, + 357, + 380, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Retrieval-augmented generation (RAG) has emerged as a promising solution to enhance large language models (LLMs) by incorporating external knowledge during inference, effectively addressing challenges like hallucination and outdated knowledge [20]. In its naive form, RAG follows a three-stage process: indexing, where documents are chunked and encoded into vectors; retrieval, where relevant documents are retrieved based on semantic similarity with the input query; and generation, where retrieved documents along with the query are fed to LLMs for response generation [12]. Recent efforts have extended RAG to VLMs, including methods such as I2I/T2I adaptors [27], RA-TTA [19], and neural priming [45]. These methods rely on additional databases constructed from large-scale datasets like LAION-2B/5B, which often incur high computational and storage costs during both construction and inference. In contrast, our work focuses on constructing a database dynamically from incoming test samples, without relying on additional large-scale databases. In addition, we employed a multimodal retrieval augmentation module that transforms our database into a unified multimodal space, enabling retrieval that aligns both vision and language modalities.", + "bbox": [ + 89, + 378, + 482, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 739, + 181, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we introduce test-time retrieval-augmented Adaptation (TT-RAA) to adapt vision-language models (VLMs) to new domains (Fig. 2). TT-RAA leverages a streaming mixture of Gaussian database (SMGD) to model test data statistics in real time. As test data streams in, we dynamically update the SMGD with Gaussian centers representing the distribution of vision features. Each test sample's vision features query the SMGD to retrieve relevant Gaussian centers, providing contextual information.", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since CLIP's joint vision-text optimization does not explicitly optimize vision-space similarity, we transform the Gaussian centers into a multimodal space using text embeddings. This allows retrieval in a space that is aligned with both vision and language modalities. Final predictions combine CLIP outputs with retrieval results from both vision and multimodal spaces, enhancing accuracy and robustness. TT-RAA thus provides a powerful framework for adapting VLMs to new domains by leveraging both vision and language information.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Contrastive Language-Image Pretraining", + "text_level": 1, + "bbox": [ + 511, + 250, + 867, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive language-image pre-training (CLIP) [34] is a vision-language model that has two separate encoders for processing images and text, respectively. These encoders are trained to project their respective inputs into a shared embedding space where related images and texts are positioned close to each other. Consider a $K$ -class classification problem where the CLIP's objective in the zero-shot setting is to match images with their most relevant textual descriptions using an image encoder $E_I$ and a text encoder $E_T$ . To obtain the textual descriptions $T$ , $K$ -class names are concatenated with hand-crafted prompts and then mapped into the $D$ -dimensional text embeddings $Z$ using the text encoder $E_T$ as", + "bbox": [ + 511, + 272, + 903, + 467 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nZ = E _ {T} (T) \\in \\mathbb {R} ^ {D \\times K}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 467, + 903, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the $t$ -th test samples $x^t$ , CLIP vision encoder produces image features", + "bbox": [ + 511, + 489, + 903, + 518 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {t} = E _ {I} \\left(x ^ {t}\\right) \\in \\mathbb {R} ^ {D}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 527, + 903, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP was trained using contrastive loss to minimize the distance between corresponding text and image embeddings. When making predictions, the CLIP output can be obtained as", + "bbox": [ + 511, + 554, + 903, + 612 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathrm {C L I P}} ^ {t} = Z ^ {T} f ^ {t} \\in \\mathbb {R} ^ {K}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 613, + 903, + 631 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, we can obtain the final prediction for test sample $x^{t}$ as $\\hat{y}^t = \\operatorname*{argmax}_k P_{\\mathrm{CLIP}}^t (k)$ , where $k \\in \\{1,2,\\dots,K\\}$ .", + "bbox": [ + 511, + 636, + 903, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Streaming Mixture of Gaussian Database", + "text_level": 1, + "bbox": [ + 511, + 681, + 867, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work addresses a challenging training-free test-time adaptation problem where models must adapt to streaming test data under two constraints: (1) Both input and output distributions of the source and target domains may be subject to distribution shifts. (2) Test samples arrive sequentially, one at a time, rather than in batches, reflecting real-world deployment. This setting is demanding because it requires training-free, zero-shot adaptation without access to batch statistics. Therefore, it is important to estimate the statistics of the new domain by capturing its full spectrum, and based on this we can retrieve the most relevant information to enhance the final prediction. In this work, we propose a Streaming Mixture of Gaussian Database (SMGD)", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8812", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/48eaa854a7bc8dfedcc7f1a059c0ccea08966a7c72bb9e3f052827cc67cf1ee8.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed test-time retrieval-augmented adaptation (TT-RAA) method. TT-RAA leverages a streaming mixture of gaussian database (SMGD) to dynamically model the distribution of streaming test data in vision space $\\mathcal{V}$ . The multimodal retrieval augmentation (MRA) module projects these Gaussian centers into a unified multimodal space $\\mathcal{M}$ , enabling effective retrieval aligned across both vision and language modalities. Combining retrieval results from vision and multimodal spaces with CLIP's original predictions, TT-RAA significantly enhances VLM performance without additional training." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 906, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "module by using a Gaussian mixture model to estimate the dynamic statistics from the test time streaming data.", + "bbox": [ + 89, + 404, + 482, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We assume that feature embedding $f^t$ of the $t$ -th sample $x_{t}$ is drawn from a mixture of Gaussian distributions", + "bbox": [ + 89, + 435, + 482, + 464 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {t} \\sim \\sum_ {k} w _ {k} \\mathcal {N} \\left(\\mu_ {k}, \\Sigma_ {k}\\right), \\forall k = 1, 2, \\dots , K, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 474, + 482, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_k \\in \\mathbb{R}^D$ and $\\Sigma_k \\in \\mathbb{R}^{D \\times D}$ are the mean and covariance of the $k$ -th Gaussian component $\\mathcal{N}$ respectively. The mixture weights $w_k \\geq 0$ , $\\sum_k w_k = 1$ , and $K$ is the number of classes. For simplicity, we assume the weights $w_k = p(y = k) = 1 / K$ . Here, SMGD is a key-value database, where the keys are Gaussian mean vectors $\\{\\mu_k\\}_{k=1}^K$ and the values are the corresponding pseudo labels of these Gaussian mean vectors. The covariance matrices $\\{\\Sigma_k\\}_{k=1}^K$ are also recorded which will involve in the retrieval during test time adaptation. Given the test time setting, all samples come in a stream. Our goal is to estimate the optimal mean $\\mu_k$ and covariance $\\Sigma_k$ for the class $k$ .", + "bbox": [ + 88, + 515, + 482, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To estimate the optimal value of the mean $\\mu_{k}$ and covariance $\\Sigma_{k}$ , we draw inspiration from streaming data statistics estimation [35]. Given a test sample $x^{t}$ at time $t$ , its pseudo label $\\hat{l}_t$ is determined by its CLIP prediction as $\\hat{l}_t = \\operatorname*{argmax}_{k} P_{\\mathrm{CLIP}}^t (k)$ . Initially, we set the mean vector $\\mu_{k}$ to the first test sample with its pseudo label as $k$ from the streaming test data. We set the initial covariance $\\Sigma_{k} = \\sigma^{2}\\mathbf{I}$ with $\\sigma = 0.1$ . We also record the corresponding entropy $h_k$ as the entropy of the first test sample's CLIP logit. Once the mean vector $\\mu_{k}$ and covariance matrix $\\Sigma_{k}$ are initialized, we progressively update them using selected test samples with high-quality pseudo labels, as measured by entropy. If the test sample $x^{t}$ at time $t$ has a pseudo label $\\hat{l}^t = k$ and an", + "bbox": [ + 89, + 698, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "entropy smaller than $h_k^{t-1}$ , the mean $\\mu_k^t$ and the covariance $\\Sigma_k^t$ at time $t$ are updated as", + "bbox": [ + 511, + 404, + 906, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {k} ^ {t} = (1 - \\eta) \\mu_ {k} ^ {t - 1} + \\eta f ^ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 446, + 903, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma_ {k} ^ {t} = (1 - \\eta) \\Sigma_ {k} ^ {t - 1} + \\eta \\left(f ^ {t} - \\mu_ {k} ^ {t}\\right) \\left(f ^ {t} - \\mu_ {k} ^ {t}\\right) ^ {T}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 468, + 903, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta$ is the update coefficient. Otherwise, $\\mu_k^t = \\mu_k^{t-1}$ and $\\Sigma_k^t = \\Sigma_k^{t-1}$ for $k \\in \\{1, 2, \\dots, K\\}$ , meaning the mean vectors and the covariance matrices remain unchanged if there is no test sample with pseudo label $k$ or if the test sample does not have a high-quality pseudo label. A high-quality pseudo label is identified when its entropy is lower than that of the Gaussian mean vector. Once the mean and covariance are updated according to Eqs. (5) and (6), the corresponding entropy is updated accordingly as", + "bbox": [ + 511, + 500, + 906, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {k} ^ {t} = (1 - \\eta) h _ {k} ^ {t - 1} + \\eta H \\left(P _ {\\mathrm {C L I P}} ^ {t}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 648, + 903, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{H}(P_{\\mathrm{CLIP}}^t) = -\\sum_{k=1}^{K} P_{\\mathrm{CLIP}}^t(k) \\log(P_{\\mathrm{CLIP}}^t(k))$ is the entropy of the CLIP logits $P_{\\mathrm{CLIP}}^t$ .", + "bbox": [ + 511, + 681, + 903, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, we keep updating a high-quality database with high confidence as measured by the entropy of test-time streaming data's pseudo labels. The SMGD $S_{G}$ is composed of the following:", + "bbox": [ + 511, + 714, + 903, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {G} = \\{G, L \\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 787, + 903, + 804 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG = \\left[ \\mu_ {0} ^ {t}, \\mu_ {1} ^ {t}, \\dots , \\mu_ {K} ^ {t} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 808, + 903, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $G \\in \\mathbb{R}^{D \\times K}$ represents the Gaussian centers of the SMGD, and each column of $L \\in \\mathbb{R}^{K \\times K}$ denotes a $K$ -dimensional one-hot pseudo-label vector. Along with SMGD, we have the covariances $\\{\\Sigma_0^t, \\Sigma_1^t, \\dots, \\Sigma_K^t\\}$ .", + "bbox": [ + 511, + 838, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8813", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/121c712bc59063ad88e26279103aff13c2399aba643331d2b1c7b9a2486b98c2.jpg", + "image_caption": [ + "Figure 3. Motivation of multimodal retrieval augmentation. Cosine similarity distributions of matched and unmatched image-text pairs (inter-modal) exhibit less overlap than those of matched and unmatched image-image pairs (intra-modal). This suggests that samples are more easily distinguishable in the multimodal space than in the vision space." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 460, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Multimodal Retrieval Augmentation", + "text_level": 1, + "bbox": [ + 89, + 316, + 406, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can directly retrieve the most similar data center from the SMGD to improve the final prediction. However, the CLIP model was optimized to reduce the inter-modal (vision-text) similarities rather than the intra-modal (vision-vision or text-text) similarities [24]. This optimization process leaves a disadvantage that similar images in the vision (image) feature space are not well clustered [43]. As shown in Fig. 3, the cosine similarity distributions of matched and unmatched image-text pairs (inter-modal) exhibit less overlap than those of matched and unmatched image-image pairs (intra-modal). This indicates that matched and unmatched pairs are more easily distinguishable in the multimodal space than in the vision-only space, a distinction shaped by CLIP's image-text contrastive learning. Therefore, in addition to the vision space retrieval augmentation, we also added a multimodal space retrieval augmentation, which compares the similarities of the samples in the CLIP text-image space rather than the vision space.", + "bbox": [ + 89, + 338, + 483, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vision Space Retrieval Augmentation. To fully leverage the available information from the vision space, we include two terms: similarity retrieval and discriminant analysis.", + "bbox": [ + 89, + 609, + 483, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarity Retrieval. This term treats the test sample as a query and retrieves the most similar class center from the SMGD based on their similarities within the vision space. Given the SMGD $S_{G} = \\{G, L\\}$ and test image features $f^{t}$ generated from the CLIP's image encoder, the prediction can be calculated as follows", + "bbox": [ + 89, + 656, + 483, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {s i m} \\left(f ^ {t}\\right) = L A \\left(G ^ {T} f ^ {t}\\right) \\in \\mathbb {R} ^ {K}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 753, + 482, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A(z) = \\exp (-\\beta (1 - z))$ is an adaptation function with a sharpness ratio $\\beta$ [54].", + "bbox": [ + 89, + 779, + 482, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discriminant Analysis. This term leverages probabilistic information derived from the Gaussian means and covariances of the SMGD by using a closed-form expression from Gaussian discriminant analysis [2]. Under the assumption that all Gaussian components share an identical covariance $\\Sigma^t$ , the discriminant score can be computed as", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Omega_ {d i s c} (f ^ {t}) = G ^ {T} \\Sigma^ {t - 1} f ^ {t} - \\frac {1}{2} \\operatorname {d i a g} \\left(G ^ {T} \\Sigma^ {t - 1} G\\right) \\tag {11} \\\\ + \\log \\frac {1}{K} \\mathbf {1} _ {K} \\in \\mathbb {R} ^ {K}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 88, + 906, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Sigma^t$ is the mean of $K$ covariances $\\{\\Sigma_k^t\\}_{k=1}^K$ . $\\Omega_{disc}(f^t)$ contains scores that reflect how well $f^t$ matches different classes, considering both the distance from the class mean and the spread of the class distribution. Therefore, the discriminant analysis-based term can be given as", + "bbox": [ + 511, + 157, + 906, + 234 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {d i s c} \\left(f ^ {t}\\right) = L \\Omega_ {d i s c} \\left(f ^ {t}\\right) \\in \\mathbb {R} ^ {K}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 246, + 906, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After that, the prediction from the vision space $\\mathcal{V}$ retrieval augmentation can then be calculated as", + "bbox": [ + 511, + 276, + 905, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathrm {R}} ^ {\\mathcal {V}} \\left(f ^ {t}\\right) = P _ {s i m} \\left(f ^ {t}\\right) + P _ {d i s c} \\left(f ^ {t}\\right) \\in \\mathbb {R} ^ {K}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 316, + 906, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal Space Retrieval Augmentation. During test-time adaptation, multimodal space retrieval augmentation treats the test sample as a query and retrieves the most similar Gaussian center from the SMGD by evaluating their distances within CLIP's multimodal space. This process involves measuring the Kullback-Leibler (KL) divergence between the image features of the test sample or Gaussian centers and the textual features of the categories in CLIP's embedding space.", + "bbox": [ + 511, + 345, + 905, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the SMGD $S_{G} = \\{G,L\\}$ and the text embeddings $Z\\in \\mathbb{R}^{D\\times K}$ generated from CLIP's text encoder, we transform Gaussian centers $G$ of the SMGD from the vision space $\\mathcal{V}$ to the multimodal space $\\mathcal{M}$ as $\\Psi$ :", + "bbox": [ + 511, + 482, + 905, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi = \\sigma \\left(Z ^ {T} G\\right) \\in \\mathbb {R} ^ {K \\times K}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 551, + 906, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ is a sigmoid function. We also project the test sample $f^{t}$ to the multimodal space $\\mathcal{M}$ as", + "bbox": [ + 511, + 580, + 905, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\psi = \\sigma \\left(Z ^ {T} f ^ {t}\\right) \\in \\mathbb {R} ^ {K}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 622, + 903, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use Kullback-Leibler (KL) divergence to compare the similarities of the two distributions in the multimodal space as", + "bbox": [ + 511, + 651, + 905, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {k} = K L (\\psi | | \\Psi_ {k}), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 696, + 903, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Psi_{k}$ is the $k$ -th column vector of $\\Psi$ , and $KL(P|Q) = \\sum_{x\\in \\mathcal{X}}P(x)\\log \\left(\\frac{P(x)}{Q(x)}\\right)$ . Therefore, $\\Phi = [\\Phi_1,\\Phi_2,\\dots ,\\Phi_K]\\in \\mathbb{R}^K$ consists of KL-divergence values. Since the most similar samples have low KL-divergence values (close to 0), they receive small weights. To address this, we negate the values in $\\Phi$ . The prediction of the multimodal space $\\mathcal{M}$ retrieval augmentation will be given as", + "bbox": [ + 511, + 719, + 905, + 833 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {R} ^ {\\mathcal {M}} = - L \\Phi \\in \\mathbb {R} ^ {K}. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 845, + 903, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we can obtain the output of the multimodal retrieval augmentation (MRA) module $P_{\\mathrm{R}}$ as a combination of the", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8814", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dd5972123e5ec6fb949d290b5328f1504905225cda1301239fa7b97ff0a87c4f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAircraftCaltech101CarsDTDEuroSATFlower102Food101PetsSUN397UCF101Average
CLIP-ViT-B/1623.2293.5566.1145.0450.4266.9982.8686.9265.6365.1664.59
CoOp18.4793.7064.5141.9246.3968.7185.3089.1464.1566.5563.88
CoCoOp22.2993.7964.9045.4539.2370.8583.9790.4666.8968.4464.63
TPT24.7894.1666.8747.7542.4468.9884.6787.7965.5068.0465.10
DiffTPT25.6092.4967.0147.0043.1370.1087.2388.2265.7462.6765.47
MTA25.3294.1368.0545.5938.7168.2684.9588.2264.9868.1164.63
DN24.3093.6064.0045.7053.3068.0086.0087.7066.5068.4065.75
ZERO25.2193.6668.0446.1234.3367.6886.5387.7565.0367.7767.72
DMN24.8494.1265.6444.3947.7771.3884.4889.0766.2866.7565.47
TDA23.9194.2467.2847.4058.0071.4286.1488.6367.6270.6667.53
TT-RAA (ours)25.3894.0866.4247.9966.1272.6886.0989.8367.6971.2968.76
", + "bbox": [ + 94, + 88, + 903, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Results on the cross-domain benchmark. Our TT-RAA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, four training-based adaptation methods (CoOp, CoCoOp, TPT, and DiffTPT), and five training-free adaptation methods (MTA, DN, DMN, ZERO, and TDA).", + "bbox": [ + 89, + 276, + 906, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "vision space retrieval argumentation result $P_{\\mathrm{R}}^{\\mathcal{V}}$ and multimodal space augmentation result $P_{\\mathrm{R}}^{\\mathcal{M}}$ as", + "bbox": [ + 89, + 327, + 482, + 359 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathrm {R}} = P _ {\\mathrm {R}} ^ {\\mathcal {V}} + P _ {\\mathrm {R}} ^ {\\mathcal {M}}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 366, + 482, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After that, we compute our final TT-RAA logits by combining the original CLIP logits $P_{\\mathrm{CLIP}}$ with the MRA prediction $P_{\\mathrm{R}}$ using a weighting factor $\\alpha$ as", + "bbox": [ + 89, + 395, + 483, + 441 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathrm {T T} - \\mathrm {R A A}} = P _ {\\mathrm {C L I P}} + \\alpha P _ {\\mathrm {R}}. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 450, + 482, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the end, we can obtain the final prediction of TT-RAA as", + "bbox": [ + 89, + 478, + 483, + 493 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} = \\underset {k} {\\operatorname {a r g m a x}} P _ {\\mathrm {T T} - \\mathrm {R A A}} (k), \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 508, + 482, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $k\\in \\{1,2,\\dots,K\\}$", + "bbox": [ + 89, + 537, + 256, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 566, + 223, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 89, + 590, + 187, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the method's test time adaptation capability across different domains and variations within a familiar domain, we conducted experiments on two benchmarks: a cross-domain (CD) benchmark comprising 10 datasets and an out-of-distribution (OOD) benchmark comprising 4 datasets. The CD benchmark primarily assesses transferability by testing on domains that diverge significantly from the original training distribution. Specifically, it includes Aircraft [26], Caltech101 [9], Cars [18], DTD [6], EuroSAT [13], Flower102 [29], Food101 [3], Pets [32], SUN397 [48], and UCF101 [40]. In contrast, the OOD benchmark focuses on robustness to distribution shifts within the same general domain. Specifically, it includes ImageNet-A [15], ImageNet-V2 [36], ImageNet-R [14], and ImageNet-S [47].", + "bbox": [ + 89, + 613, + 483, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 848, + 308, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement our method based on the pre-trained CLIP architecture [34], which comprises a vision transformer", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(ViT) based image encoder [7] and a transformer-based text encoder [44]. Our experimental setup follows the challenging single-image test-time adaptation scenario, processing samples sequentially with a batch size of 1. To maintain computational efficiency and eliminate the need for backpropagation during inference, we utilize the hand-crafted prompts as proposed in [34] rather than learnable prompts. The experiments are conducted using the PyTorch framework, with evaluations performed on a single NVIDIA RTX A5000 GPU. Following standard practice in similar tasks, we employ top-1 accuracy as our evaluation metric.", + "bbox": [ + 511, + 327, + 906, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparisons with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 511, + 501, + 883, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive experiments to evaluate TT-RAA against state-of-the-art approaches on the CD benchmark in Table 1 and OOD benchmark in Table 2. The competing methods include the CLIP baseline [34], training-based adaptation methods, and training-free adaptation methods. Discussions on Competing Methods. In the field of efficient domain adaptation, several methods require training, either in the training stage or at the test time, to align the model more closely with the target distribution. Context optimization (CoOp) [57], conditional context optimization (CoCoOp) [56] both involve learning prompts during a training phase. CoOp focuses on optimizing prompt tokens that are fixed for each domain, while CoCoOp extends this to generate a conditional prompt token for each image. Test-time prompt tuning (TPT) [38] and diffusion model for TPT (DiffTPT) [10], on the other hand, adapt the model at test time by tuning prompt embeddings. TPT performs direct gradient-based prompt tuning using test data, while DiffTPT uses new data generated by diffusion models.", + "bbox": [ + 511, + 523, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to training-based methods, we also compare with several training-free adaptation methods, including two distribution-based methods, one entropy-based method, and two cache-based methods. Distribution-based methods evaluated are MeanShift for Test-time Augmentation (MTA) [51] and Distribution Normalization (DN) [58].", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8815", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a54d7495ac1af9a8efe4e9d87b22c6a6a3d51bd405b7491bff715157fa59936b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodImageNet-AImageNet-V2ImageNet-RImageNet-SAverage
CLIP-ViT-B/1649.8961.8877.6548.2459.42
CoOp49.7164.2075.2147.9959.28
CoCoOp50.6364.0776.1848.7559.91
TPT54.7763.4577.0647.9460.81
DiffTPT55.6865.1075.0046.8060.52
MTA57.4163.6176.9248.5861.63
DN58.7162.8980.2048.9462.69
ZERO59.6164.1677.2248.4062.35
DMN58.2865.1778.5553.2063.80
TDA60.1164.6780.2450.5463.89
TT-RAA (Ours)60.5964.6980.5849.9863.96
", + "bbox": [ + 217, + 88, + 779, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Results on the out-of-distribution benchmark. Our TT-RAA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, four training-based adaptation methods (CoOp, CoCoOp, TPT, and DiffTPT), and five training-free adaptation methods (MTA, ZERO, DN, DMN, and TDA).", + "bbox": [ + 89, + 286, + 906, + 329 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "They improve the test data distribution estimation via MeanShift algorithm and distribution normalization, respectively. Cache-based methods include the Dual Memory Network (DMN) [55] and TDA [17]. DMN comprises a dynamic and a static cache, while TDA employs a positive and a negative cache. For a fair comparison, we reproduced DMN's zero-shot results using our standardized prompts instead of DMN's original LLM-generated prompts, and we reduced the batch size to 1. We also compare against ZERO [8], an entropy-based method. ZERO sets the temperature of most confident predictions as zero to approximate marginal entropy minimization.", + "bbox": [ + 88, + 335, + 480, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on the Cross-domain Benchmark. Despite having the advantage of task-specific training, training-based adaptation methods, such as CoOp, CoCoOp, TPT, and DiffTPT, show limited generalization capabilities. TT-RAA outperforms these methods by significant margins (about $3\\% - 4\\%$ ). On fine-grained recognition tasks such as Flowers102, TT-RAA achieves $72.68\\%$ accuracy, surpassing CoCoOp $(70.85\\%)$ by $1.83\\%$ . This gap widens further on specialized datasets like SUN397, where TT-RAA $(67.69\\%)$ demonstrates superior feature adaptation compared to CoCoOp $(64.15\\%)$ . These results suggest that static prompt learning, even with input-conditional mechanisms, may not be sufficient enough for handling significant domain shifts.", + "bbox": [ + 89, + 520, + 482, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In comparison to existing training-free adaptation approaches, TT-RAA demonstrates consistent superiority across various domain shifts. On remote sensing data (EuroSAT), TT-RAA $(66.12\\%)$ significantly outperforms TDA $(58.00\\%)$ by $8.12\\%$ . This substantial gain can be attributed to our SMGD effectively capturing the unique characteristics of satellite imagery. Similarly, for action recognition (UCF101), TT-RAA $(71.29\\%)$ surpasses TDA $(70.66\\%)$ by $0.63\\%$ , demonstrating robust feature adaptation. The performance advantages extend to other specialized domains as well. This consistent improvement across diverse domains validates the effectiveness of our method in bridging", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "domain gaps in test-time settings.", + "bbox": [ + 511, + 337, + 736, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on the Out-of-distribution Benchmark. The experimental results in Table 2 demonstrate that our proposed TT-RAA method achieves state-of-the-art performance on the OOD benchmark with an average accuracy of $63.96\\%$ across all ImageNet variants, outperforming both training-based adaptation methods (CoOp, CoCoOp, TPT, DiffTPT) and training-free approaches (MTA, DN, ZERO, DMN, TDA). Notably, TT-RAA exhibits good performance on ImageNet-R $(80.58\\%)$ and ImageNet-A $(60.59\\%)$ , highlighting its robust generalization capabilities under diverse distribution shifts. When examining performance across different distribution shifts, we observe that most methods achieve their highest accuracy on ImageNet-R and ImageNet-V2. ImageNet-S with sketches poses the greatest challenge for all methods, indicating that semantic variations remain difficult to address even with advanced adaptation techniques.", + "bbox": [ + 511, + 353, + 903, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 619, + 679, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies on the CD benchmark to systematically evaluate the effectiveness of each proposed component in TT-RAA. Figure 4 presents the quantitative results of our analysis with both ResNet-50 and ViT-B/16 backbones. We establish our baseline using the TDA method with one-shot capacity, which achieves $59.82\\%$ (ResNet-50) and $67.04\\%$ (ViT-B/16) accuracy.", + "bbox": [ + 511, + 642, + 903, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Streaming Mixture of Gaussian Database Module. When integrating the SMGD module with the baseline, we observe a performance improvement of about $1.4\\%$ with ViT-B/16 and $1.2\\%$ with ResNet-50. This enhancement demonstrates that our SMGD effectively captures the dynamic distribution of test samples while maintaining computational efficiency. Its adaptive nature in updating class statistics online is useful for handling distribution shifts.", + "bbox": [ + 511, + 750, + 903, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multimodal Retrieval Augmentation Module. Incorporating the MRA module also improved the performance by", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8816", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b56bbdf756d4d985f5e316e87d7d3fff514621606d1f8f38adb12e338164e796.jpg", + "image_caption": [ + "Figure 4. Ablation studies of TT-RAA. The SMGD and MRA modules independently improve performance over the baseline. Their combination yields further performance gains, demonstrating complementary benefits." + ], + "image_footnote": [], + "bbox": [ + 106, + 90, + 472, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "about $0.4\\%$ and $0.2\\%$ with ViT-B/16 and ResNet-50, respectively. This gain supports our hypothesis regarding vision space misalignment in CLIP, which stems from its vision-text joint optimization objective. The performance improvement demonstrates that projecting both the SMGD database and test samples from vision space to multimodal space effectively addresses this misalignment.", + "bbox": [ + 89, + 324, + 483, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Complementary Benefits of Proposed Modules. When combining both SMGD and MRA modules (denoted as TT-RAA), we achieve the best performance of $68.76\\%$ with ViT-B/16 and $61.25\\%$ with ResNet-50, representing a substantial improvement of $1.72\\%$ and $1.43\\%$ over the baseline. These gains suggest that SMGD and MRA offer complementary and synergistic benefits.", + "bbox": [ + 89, + 430, + 483, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Feature Backbone Impact. The effectiveness of our approach is further validated across two vision ecoders: ViT-B/16 and ResNet-50. These consistent improvements across different backbones suggest that our adaptation strategy effectively addresses fundamental domain shift challenges rather than exploiting architecture-specific characteristics.", + "bbox": [ + 89, + 536, + 483, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Hyperparameter Analysis", + "text_level": 1, + "bbox": [ + 89, + 652, + 326, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The hyperparameter analysis of the update coefficient $\\eta$ on the UCF101 is shown in Fig. 5a. When $\\eta = 0$ , the model maintains a static Gaussian center initialized at the start, resulting in poor adaptation. From 0 to 0.55, accuracy increases sharply, highlighting the importance of incorporating new information. However, the performance deterioration beyond 0.65 reveals a critical threshold where excessive emphasis on new samples compromises the model's stability. This observation aligns with theoretical expectations, as larger $\\eta$ values reduce the influence of historical information, potentially leading to over-adaptation to new samples. The empirical optimal value of $\\eta = 0.65$ suggests that maintaining approximately $35\\%$ of historical information while incorporating $65\\%$ of new information achieves the best trade-off.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97b338899f73e273bce87e9f8c68fbabc888cebfba23eb72dbaa4d7fb3760afa.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 521, + 89, + 705, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b695c877f870a06d95745d4684e2c0d472f39cd20e7a595ee0db7bf07168f3ae.jpg", + "image_caption": [ + "(b)", + "Figure 5. (a) Hyperparameter analysis of the update coefficient $\\eta$ . It shows that the best coefficient is 0.65. (b) Experiment with additional data. The method can accept additional data, which improve the performance significantly." + ], + "image_footnote": [], + "bbox": [ + 709, + 93, + 897, + 199 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Experiments with Access to Additional Data", + "text_level": 1, + "bbox": [ + 511, + 310, + 888, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the standard test-time adaptation setting, access to the target domain's training set is typically restricted. However, in many cases, the training set data might be available. Finetuning large-scale vision-language models in such scenarios remains computationally expensive. This raises an important question: how can performance be improved using the training data without computationally expensive finetuning? With a slight modification to TT-RAA, we propose leveraging the target domain training set (if accessible prior to test time) to construct a database and calculate class centers using a Gaussian mixture model. By doing so, we perform the same MRA process as in TT-RAA. This modified approach, referred to as TT-RAA-A, demonstrates significant performance improvements while maintaining a low computational requirement. The corresponding results are presented in Fig. 5b, where TT-RAA-A outperforms TT-RAA by a large margin, highlighting the benefits of incorporating target training data in a training-free manner, which open a new direction on adaptation without training.", + "bbox": [ + 511, + 332, + 906, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced TT-RAA, a novel test-time retrieval augmented adaptation method that effectively addresses the computationally expensive domain adaptation challenges faced by vision-language models. Experimental results validated the effectiveness of SMGD by continuously estimating the test distribution and compressing all useful information into a single representation for each class without discarding old samples. By transforming Gaussian centers from vision space into a shared multimodal space, MRA enables more effective retrieval that leverages both visual and textual features. The success of TT-RAA opens up new possibilities for deploying vision-language models in real-world applications where computational resources are limited and rapid adaptation to new domains is crucial. Further exploration on adapting our method to larger multimodal foundation and large language models is promising.", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8817", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Meta AI. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. Meta AI Report, 2024. 1, 2", + "[2] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 5", + "[3] Lukas Bossard, Matthieu Guillaumin, et al. Food-101: Mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 6", + "[4] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. VisualGPT: Data-efficient adaptation of pretrained language models for image captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18030-18040, 2022. 2", + "[5] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. European Conference on Computer Vision, pages 104-120, 2020. 2", + "[6] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2014. 6", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6", + "[8] Matteo Farina, Gianni Franchi, Giovanni Iacca, Massimiliano Mancini, and Elisa Ricci. Frustratingly easy test-time adaptation of vision-language models. Advances in Neural Information Processing Systems, 2024. 7", + "[9] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2004. 6", + "[10] Chun-Mei Feng, Kai Yu, Yong Liu, et al. Diverse data augmentation with diffusions for effective test-time prompt tuning. In IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 2, 3, 6", + "[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, et al. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. 1", + "[12] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, et al. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2024. 3", + "[13] Patrick Helber, Benjamin Bischke, et al. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 6", + "[14] Dan Hendrycks, Steven Basart, Norman Mu, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6", + "[15] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. International Conference on Machine Learning, pages 4904-4916, 2021. 1, 2", + "[17] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 7", + "[18] Jonathan Krause, Michael Stark, et al. 3D object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshops, 2013. 6", + "[19] Youngjun Lee, Doyoung Kim, Junhyeok Kang, Jihwan Bang, Hwanjun Song, et al. RA-TTA: Retrieval-augmented test-time adaptation for vision-language models. In International Conference on Learning Representations, 2025. 3", + "[20] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, and Vladimir others Karpukhin. Retrievalaugmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems, pages 9459-9474, 2020. 3", + "[21] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2", + "[22] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 1, 2", + "[23] Jian Liang, Ran He, et al. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, pages 1-34, 2024. 1, 2", + "[24] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 2, 5", + "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 1, 2", + "[26] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, et al. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6", + "[27] Yifei Ming and Yixuan Li. Understanding retrieval-augmented task adaptation for vision-language models. In International Conference on Machine Learning, pages 35719-35743, 2024. 3", + "[28] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, et al. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 2, 3", + "[29] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 6", + "[30] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8818", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 2, 3", + "[31] OpenAI. Gpt-4v(ison) system card. In OpenAI Technical Report, 2023. 1, 2", + "[32] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 6", + "[33] Rizwan Qureshi, Ranjan Sapkota, Abbas Shah, et al. Thinking beyond tokens: From brain-inspired intelligence to cognitive foundations for artificial general intelligence and its societal impact. arXiv preprint arXiv:2507.00951, 2025. 1", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 1, 2, 3, 6", + "[35] Aditi Raghunathan, Prateek Jain, et al. Learning mixture of gaussians with streaming data. Advances in Neural Information Processing Systems, 30, 2017. 4", + "[36] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, et al. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, 2019. 6", + "[37] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, et al. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 2020. 2, 3", + "[38] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 2022. 2, 3, 6", + "[39] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2", + "[40] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 6", + "[41] Tao Sun, Cheng Lu, and Haibin Ling. Prior knowledge guided unsupervised domain adaptation. In European Conference on Computer Vision. Springer, 2022. 1, 2", + "[42] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2, 3", + "[43] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. SuS-X: Training-free name-only transfer of vision-language models. In IEEE/CVF International Conference on Computer Vision, pages 2725-2736, 2023. 5", + "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6", + "[45] Matthew Wallingford, Vivek Ramanujan, Alex Fang, Aditya Kusupati, Roozbeh Mottaghi, Aniruddha Kembhavi, et al." + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Neural priming for sample-efficient adaptation. Advances in Neural Information Processing Systems, 2023. 3", + "[46] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 2, 3", + "[47] Haohan Wang, Songwei Ge, et al. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, 2019. 6", + "[48] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In IEEE Conference on Computer Vision and Pattern Recognition, 2010. 6", + "[49] Shiqi Yang, Joost Van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. Advances in Neural Information Processing Systems, 34:29393-29405, 2021. 1, 2", + "[50] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. International Conference on Learning Representations, 2022. 1, 2", + "[51] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23783-23793, 2024. 2, 6", + "[52] Jingyi Zhang, Jiaxing Huang, et al. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[53] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. VinVL: Revisiting visual representations in vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 2", + "[54] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision, pages 493-510. Springer, 2022. 5", + "[55] Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28718-28728, 2024. 2, 7", + "[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 6", + "[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 6", + "[58] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser Nam Lim. Test-time distribution normalization for contrastively learned visual-language models. Advances in Neural Information Processing Systems, pages 47105-47123, 2023. 2, 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8819", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/abcad282-0784-41dd-b230-4b9b3488f3c1_model.json b/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/abcad282-0784-41dd-b230-4b9b3488f3c1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5bc94fcf94ac93458474726771800d08b524b85a --- /dev/null +++ b/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/abcad282-0784-41dd-b230-4b9b3488f3c1_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.132, + 0.131, + 0.868, + 0.154 + ], + "angle": 0, + "content": "Test-Time Retrieval-Augmented Adaptation for Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.18, + 0.812, + 0.217 + ], + "angle": 0, + "content": "Xinqi Fan\\(^{1}\\), Xueli Chen\\(^{2}\\), Luoxiao Yang\\(^{3}\\), Chuin Hong Yap\\(^{1}\\), Rizwan Qureshi\\(^{4}\\), Qi Dou\\(^{5}\\), Moi Hoon Yap\\(^{1}\\), Mubarak Shah\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.217, + 0.89, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)Manchester Metropolitan University \\(^{2}\\)Hong Kong Metropolitan University \\(^{3}\\)Xi'an University of Technology \\(^{4}\\)University of Central Florida \\(^{5}\\)Chinese University of Hong Kong" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.485, + 0.789 + ], + "angle": 0, + "content": "Vision-language models (VLMs) have shown promise in test-time adaptation tasks due to their remarkable capabilities in understanding and reasoning about visual content through natural language descriptions. However, training VLMs typically demands substantial computational resources, and they often struggle to adapt efficiently to new domains or tasks. Additionally, dynamically estimating the test distribution from streaming data at test time remains a significant challenge. In this work, we propose a novel test-time retrieval-augmented adaptation (TT-RAA) method that enables VLMs to maintain high performance across diverse visual recognition tasks without the need for task-specific training or large computational overhead. During inference, TT-RAA employs a streaming mixture of Gaussian database (SMGD) to continuously estimate test distributions, requiring minimal storage. Then, TT-RAA retrieves the most relevant information from the SMGD, enhancing the original VLM outputs. A key limitation of CLIP-based VLMs is their inter-modal vision-language optimization, which does not optimize vision-space similarity, leading to larger intra-modal variance. To address this, we propose a multimodal retrieval augmentation module that transforms the SMGD into a unified multimodal space, enabling retrieval that aligns both vision and language modalities. Extensive experiments across both cross-domain and out-of-distribution benchmarks comprising fourteen datasets demonstrate TT-RAA's superior performance compared to state-of-the-art methods. Ablation studies and hyperparameter analyses further validate the effectiveness of the proposed modules. The source code of our work is available at https://github.com/xinqi-fan/TT-RAA." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.8, + 0.222, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Vision-language models (VLMs) have significantly advanced visual understanding by leveraging natural language supervision with contrastive language-image pre-training (CLIP) notably pioneering this field [34]. The zero-shot capability enables this model to be deployed in various do" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.29, + 0.896, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.468, + 0.908, + 0.538 + ], + "angle": 0, + "content": "Figure 1. Illustration of our proposed test time retrieval augmented adaptation (TT-RAA). Without TT-RAA, the CLIP prediction will be the final prediction. With TT-RAA, the image feature will be used to retrieve the most similar database centers which will be used to improve the final prediction in a training-free manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.809 + ], + "angle": 0, + "content": "mains. This foundational work has catalyzed a wave of architectural innovations, including domain-specific variants and architectural enhancements that build upon CLIP's core principles [52]. Some variants have focused on scaling up capacities, learning from noisy text supervision, and fine-grained alignment [16, 50]. However, these models consistently encounter challenges when deployed in specialized domains, where the distribution of images often differs substantially from their web-scale training data [11, 33]. This challenge has become even more prominent with the recent introduction of more sophisticated large VLMs such as GPT-4V [31], LLaVA [25], and Llama 3.2 Vision [1]. While these advanced models excel at integrating visual and textual data, the high computational cost of fine-tuning them for domain-specific tasks poses a major challenge, especially in specialized visual domains that differ significantly from general web images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Test-time adaptation (TTA) has emerged as a promising solution that enables models to adapt to new domains without accessing source domain training data [23]. Traditional TTA approaches include self-training-based methods that leverage pseudo-labels [22, 41, 49], self-supervised learning-based methods that incorporate aux" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8810" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "iliary tasks [42], batch normalization [28, 37], and entropy minimization-based methods [30, 46]. For VLMs, recent works have explored prompt-based adaptation methods [10, 38] to bridge the domain gap between web-scale training data and specific downstream tasks. However, these methods often incur substantial computational overhead, highlighting the need for more efficient approaches [10, 17]. Advancements in training-free methods have explored the use of cache mechanisms [17, 55] and distribution-based methods [51, 58] to improve the efficiency. Despite these advances, dynamically estimating the test distribution from streaming data at test time remains a significant challenge, due to the test sample arriving one at a time." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.292, + 0.485, + 0.653 + ], + "angle": 0, + "content": "In this work, we introduce test-time retrieval-augmented adaptation (TT-RAA), a novel training-free method that allows VLMs to sustain strong performance across various visual recognition tasks without requiring task-specific training or high computational costs (Fig. 1). TT-RAA is a training-free framework, which uses a streaming mixture of Gaussian database (SMGD) to dynamically capture test distribution characteristics with minimal storage requirements at inference time. TT-RAA retrieves relevant information from the SMGD to enhance the performance of the original VLM. Furthermore, an inherent limitation of CLIP-based VLMs is their lack of optimization for intramodal similarity, which does not optimize similarity in the vision space, resulting in greater intramodal variance [24]. To mitigate this, we introduce a multimodal retrieval augmentation (MRA) module that projects the SMGD into a shared multimodal embedding space, facilitating retrieval that harmonizes vision and language modalities. Extensive testing across ten cross-domain datasets and four out-of-distribution datasets confirms that TT-RAA delivers superior performance relative to state-of-the-art methods, with additional ablation studies and hyperparameter analyses supporting the effectiveness of the proposed modules. The key contributions of this work include:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.657, + 0.483, + 0.716 + ], + "angle": 0, + "content": "- We introduce test-time retrieval-augmented adaptation (TT-RAA), an efficient, training-free framework that significantly improves domain adaptability without additional fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.718, + 0.483, + 0.761 + ], + "angle": 0, + "content": "- We propose streaming mixture of Gaussian database (SMGD) for dynamic, efficient, and robust estimation of test domain statistics during inference." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.763, + 0.483, + 0.822 + ], + "angle": 0, + "content": "- We propose multimodal retrieval augmentation (MRA), a novel mechanism addressing CLIP's inherent intra-modal similarity limitation, aligning both vision and language modalities effectively." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.657, + 0.483, + 0.822 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.9 + ], + "angle": 0, + "content": "The rest of this paper is organized as follows: Section 2 provides some related works, Section 3 describes the proposed method. Experimental results and performance evaluation are presented in Section 4. Finally, Section 5 concludes the article and outlines future research directions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.116, + 0.733, + 0.132 + ], + "angle": 0, + "content": "2.1. Vision Language Model" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.139, + 0.908, + 0.715 + ], + "angle": 0, + "content": "The development of vision language models (VLMs) can be traced back to earlier works that explored vision-language pre-training for specific tasks, where VisualBERT [21] and UNITER [5] pioneered joint representation learning between visual and textual modalities for tasks like visual question answering. The field witnessed a paradigm shift with CLIP [34], which demonstrated impressive zero-shot capabilities by learning transferable visual representations through contrastive learning between images and texts at web scale. ALIGN [16] further scaled up this approach with 1.8 billion noisy image-text pairs and introduced noise-robust training strategies. FILIP [50] enhanced CLIP by introducing fine-grained region-word alignment, enabling more detailed vision-language correspondences. FLAVA [39] introduced a single foundation model that can tackle vision, language, and multimodal tasks. More recently, the emergence of multimodal large language models has pushed the boundaries of vision-language understanding, where GPT-4V [31] extends GPT-4's language capabilities to handle visual inputs, LLaVA [25] achieves strong visual conversation abilities by aligning vision encoders with language models through minimal task-specific training, and Llama 3.2 [1] demonstrates remarkable vision and language capabilities with open models. Other notable VLMs include VinVL [53], which enhanced visual understanding by incorporating object detection and segmentation, and VisualGPT [4], which demonstrated zero-shot transfer to various vision-language tasks. These models have continued to push the boundaries of multimodal representation learning, enabling a wide range of applications from intelligent assistants to creative tools. While these newer models demonstrate impressive capabilities in combining visual and textual understanding, they face similar domain adaptation challenges but at an even larger scale. The computational demands for fine-tuning these models on domain-specific tasks present a significant barrier, particularly for specialized applications where the visual domain differs markedly from general web imagery." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.727, + 0.719, + 0.743 + ], + "angle": 0, + "content": "2.2. Test Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Normal domain adaptation techniques require access to both source and target domain data, but test-time adaptation (TTA) adapts models to the target domain without source data, making it particularly useful for scenarios with privacy or legal restrictions [23]. To adapt a pre-trained model to a target domain, methods like self-training refine model predictions using pseudo-labels. Approaches such as centroid-based [22], neighbor-based [49], and optimization-based [41] strategies enhance the quality of these pseudolabels. Self-supervised methods introduce auxiliary tasks to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "8811" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "align features, exemplified by test-time training (TTT) [42] with a self-supervised rotation prediction task. Batch normalization [28, 37] techniques and entropy minimization strategies [30, 46] have also been adopted in TTA tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.349 + ], + "angle": 0, + "content": "Recent advancements focus on VLMs like CLIP, using domain-specific prompts from test data. Test-time prompt tuning (TPT) [38] fine-tunes a learnable prompt with each testing sample, while DiffTPT [10] uses pre-trained diffusion models to enhance test data diversity. Although these methods are promising, they still require a large amount of computational resources for their training-based adaptation, making it hard to adapt large VLMs. To reduce costs, a training-free dynamic adapter (TDA) [17] has been introduced, using a lightweight dynamic key-value cache without backpropagation. However, its cache only captures the streaming test data with a limited cache capacity without the ability to estimate the full spectrum of the test data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.358, + 0.382, + 0.373 + ], + "angle": 0, + "content": "2.3. Retrieval Augmented Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.483, + 0.726 + ], + "angle": 0, + "content": "Retrieval-augmented generation (RAG) has emerged as a promising solution to enhance large language models (LLMs) by incorporating external knowledge during inference, effectively addressing challenges like hallucination and outdated knowledge [20]. In its naive form, RAG follows a three-stage process: indexing, where documents are chunked and encoded into vectors; retrieval, where relevant documents are retrieved based on semantic similarity with the input query; and generation, where retrieved documents along with the query are fed to LLMs for response generation [12]. Recent efforts have extended RAG to VLMs, including methods such as I2I/T2I adaptors [27], RA-TTA [19], and neural priming [45]. These methods rely on additional databases constructed from large-scale datasets like LAION-2B/5B, which often incur high computational and storage costs during both construction and inference. In contrast, our work focuses on constructing a database dynamically from incoming test samples, without relying on additional large-scale databases. In addition, we employed a multimodal retrieval augmentation module that transforms our database into a unified multimodal space, enabling retrieval that aligns both vision and language modalities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.182, + 0.755 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce test-time retrieval-augmented Adaptation (TT-RAA) to adapt vision-language models (VLMs) to new domains (Fig. 2). TT-RAA leverages a streaming mixture of Gaussian database (SMGD) to model test data statistics in real time. As test data streams in, we dynamically update the SMGD with Gaussian centers representing the distribution of vision features. Each test sample's vision features query the SMGD to retrieve relevant Gaussian centers, providing contextual information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "Since CLIP's joint vision-text optimization does not explicitly optimize vision-space similarity, we transform the Gaussian centers into a multimodal space using text embeddings. This allows retrieval in a space that is aligned with both vision and language modalities. Final predictions combine CLIP outputs with retrieval results from both vision and multimodal spaces, enhancing accuracy and robustness. TT-RAA thus provides a powerful framework for adapting VLMs to new domains by leveraging both vision and language information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.251, + 0.868, + 0.268 + ], + "angle": 0, + "content": "3.1. Contrastive Language-Image Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.905, + 0.468 + ], + "angle": 0, + "content": "Contrastive language-image pre-training (CLIP) [34] is a vision-language model that has two separate encoders for processing images and text, respectively. These encoders are trained to project their respective inputs into a shared embedding space where related images and texts are positioned close to each other. Consider a \\(K\\)-class classification problem where the CLIP's objective in the zero-shot setting is to match images with their most relevant textual descriptions using an image encoder \\(E_I\\) and a text encoder \\(E_T\\). To obtain the textual descriptions \\(T\\), \\(K\\)-class names are concatenated with hand-crafted prompts and then mapped into the \\(D\\)-dimensional text embeddings \\(Z\\) using the text encoder \\(E_T\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.468, + 0.905, + 0.485 + ], + "angle": 0, + "content": "\\[\nZ = E _ {T} (T) \\in \\mathbb {R} ^ {D \\times K}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.49, + 0.905, + 0.52 + ], + "angle": 0, + "content": "Given the \\(t\\)-th test samples \\(x^t\\), CLIP vision encoder produces image features" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.529, + 0.905, + 0.547 + ], + "angle": 0, + "content": "\\[\nf ^ {t} = E _ {I} \\left(x ^ {t}\\right) \\in \\mathbb {R} ^ {D}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.555, + 0.905, + 0.613 + ], + "angle": 0, + "content": "CLIP was trained using contrastive loss to minimize the distance between corresponding text and image embeddings. When making predictions, the CLIP output can be obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.614, + 0.905, + 0.632 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathrm {C L I P}} ^ {t} = Z ^ {T} f ^ {t} \\in \\mathbb {R} ^ {K}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.637, + 0.905, + 0.675 + ], + "angle": 0, + "content": "Then, we can obtain the final prediction for test sample \\( x^{t} \\) as \\( \\hat{y}^t = \\operatorname*{argmax}_k P_{\\mathrm{CLIP}}^t (k) \\), where \\( k \\in \\{1,2,\\dots,K\\} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.868, + 0.699 + ], + "angle": 0, + "content": "3.2. Streaming Mixture of Gaussian Database" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our work addresses a challenging training-free test-time adaptation problem where models must adapt to streaming test data under two constraints: (1) Both input and output distributions of the source and target domains may be subject to distribution shifts. (2) Test samples arrive sequentially, one at a time, rather than in batches, reflecting real-world deployment. This setting is demanding because it requires training-free, zero-shot adaptation without access to batch statistics. Therefore, it is important to estimate the statistics of the new domain by capturing its full spectrum, and based on this we can retrieve the most relevant information to enhance the final prediction. In this work, we propose a Streaming Mixture of Gaussian Database (SMGD)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8812" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.907, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.327, + 0.907, + 0.398 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed test-time retrieval-augmented adaptation (TT-RAA) method. TT-RAA leverages a streaming mixture of gaussian database (SMGD) to dynamically model the distribution of streaming test data in vision space \\(\\mathcal{V}\\). The multimodal retrieval augmentation (MRA) module projects these Gaussian centers into a unified multimodal space \\(\\mathcal{M}\\), enabling effective retrieval aligned across both vision and language modalities. Combining retrieval results from vision and multimodal spaces with CLIP's original predictions, TT-RAA significantly enhances VLM performance without additional training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.405, + 0.483, + 0.435 + ], + "angle": 0, + "content": "module by using a Gaussian mixture model to estimate the dynamic statistics from the test time streaming data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.483, + 0.465 + ], + "angle": 0, + "content": "We assume that feature embedding \\( f^t \\) of the \\( t \\)-th sample \\( x_{t} \\) is drawn from a mixture of Gaussian distributions" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.475, + 0.483, + 0.506 + ], + "angle": 0, + "content": "\\[\nf ^ {t} \\sim \\sum_ {k} w _ {k} \\mathcal {N} \\left(\\mu_ {k}, \\Sigma_ {k}\\right), \\forall k = 1, 2, \\dots , K, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.516, + 0.483, + 0.698 + ], + "angle": 0, + "content": "where \\(\\mu_k \\in \\mathbb{R}^D\\) and \\(\\Sigma_k \\in \\mathbb{R}^{D \\times D}\\) are the mean and covariance of the \\(k\\)-th Gaussian component \\(\\mathcal{N}\\) respectively. The mixture weights \\(w_k \\geq 0\\), \\(\\sum_k w_k = 1\\), and \\(K\\) is the number of classes. For simplicity, we assume the weights \\(w_k = p(y = k) = 1 / K\\). Here, SMGD is a key-value database, where the keys are Gaussian mean vectors \\(\\{\\mu_k\\}_{k=1}^K\\) and the values are the corresponding pseudo labels of these Gaussian mean vectors. The covariance matrices \\(\\{\\Sigma_k\\}_{k=1}^K\\) are also recorded which will involve in the retrieval during test time adaptation. Given the test time setting, all samples come in a stream. Our goal is to estimate the optimal mean \\(\\mu_k\\) and covariance \\(\\Sigma_k\\) for the class \\(k\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To estimate the optimal value of the mean \\(\\mu_{k}\\) and covariance \\(\\Sigma_{k}\\), we draw inspiration from streaming data statistics estimation [35]. Given a test sample \\(x^{t}\\) at time \\(t\\), its pseudo label \\(\\hat{l}_t\\) is determined by its CLIP prediction as \\(\\hat{l}_t = \\operatorname*{argmax}_{k} P_{\\mathrm{CLIP}}^t (k)\\). Initially, we set the mean vector \\(\\mu_{k}\\) to the first test sample with its pseudo label as \\(k\\) from the streaming test data. We set the initial covariance \\(\\Sigma_{k} = \\sigma^{2}\\mathbf{I}\\) with \\(\\sigma = 0.1\\). We also record the corresponding entropy \\(h_k\\) as the entropy of the first test sample's CLIP logit. Once the mean vector \\(\\mu_{k}\\) and covariance matrix \\(\\Sigma_{k}\\) are initialized, we progressively update them using selected test samples with high-quality pseudo labels, as measured by entropy. If the test sample \\(x^{t}\\) at time \\(t\\) has a pseudo label \\(\\hat{l}^t = k\\) and an" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.405, + 0.907, + 0.438 + ], + "angle": 0, + "content": "entropy smaller than \\( h_k^{t-1} \\), the mean \\( \\mu_k^t \\) and the covariance \\( \\Sigma_k^t \\) at time \\( t \\) are updated as" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.448, + 0.905, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mu_ {k} ^ {t} = (1 - \\eta) \\mu_ {k} ^ {t - 1} + \\eta f ^ {t}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.469, + 0.905, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\Sigma_ {k} ^ {t} = (1 - \\eta) \\Sigma_ {k} ^ {t - 1} + \\eta \\left(f ^ {t} - \\mu_ {k} ^ {t}\\right) \\left(f ^ {t} - \\mu_ {k} ^ {t}\\right) ^ {T}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.907, + 0.639 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the update coefficient. Otherwise, \\(\\mu_k^t = \\mu_k^{t-1}\\) and \\(\\Sigma_k^t = \\Sigma_k^{t-1}\\) for \\(k \\in \\{1, 2, \\dots, K\\}\\), meaning the mean vectors and the covariance matrices remain unchanged if there is no test sample with pseudo label \\(k\\) or if the test sample does not have a high-quality pseudo label. A high-quality pseudo label is identified when its entropy is lower than that of the Gaussian mean vector. Once the mean and covariance are updated according to Eqs. (5) and (6), the corresponding entropy is updated accordingly as" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.65, + 0.905, + 0.669 + ], + "angle": 0, + "content": "\\[\nh _ {k} ^ {t} = (1 - \\eta) h _ {k} ^ {t - 1} + \\eta H \\left(P _ {\\mathrm {C L I P}} ^ {t}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.905, + 0.715 + ], + "angle": 0, + "content": "where \\(\\mathrm{H}(P_{\\mathrm{CLIP}}^t) = -\\sum_{k=1}^{K} P_{\\mathrm{CLIP}}^t(k) \\log(P_{\\mathrm{CLIP}}^t(k))\\) is the entropy of the CLIP logits \\(P_{\\mathrm{CLIP}}^t\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.905, + 0.775 + ], + "angle": 0, + "content": "In this way, we keep updating a high-quality database with high confidence as measured by the entropy of test-time streaming data's pseudo labels. The SMGD \\( S_{G} \\) is composed of the following:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.789, + 0.905, + 0.805 + ], + "angle": 0, + "content": "\\[\nS _ {G} = \\{G, L \\}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.809, + 0.905, + 0.827 + ], + "angle": 0, + "content": "\\[\nG = \\left[ \\mu_ {0} ^ {t}, \\mu_ {1} ^ {t}, \\dots , \\mu_ {K} ^ {t} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.839, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(G \\in \\mathbb{R}^{D \\times K}\\) represents the Gaussian centers of the SMGD, and each column of \\(L \\in \\mathbb{R}^{K \\times K}\\) denotes a \\(K\\)-dimensional one-hot pseudo-label vector. Along with SMGD, we have the covariances \\(\\{\\Sigma_0^t, \\Sigma_1^t, \\dots, \\Sigma_K^t\\}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8813" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.092, + 0.462, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.228, + 0.485, + 0.313 + ], + "angle": 0, + "content": "Figure 3. Motivation of multimodal retrieval augmentation. Cosine similarity distributions of matched and unmatched image-text pairs (inter-modal) exhibit less overlap than those of matched and unmatched image-image pairs (intra-modal). This suggests that samples are more easily distinguishable in the multimodal space than in the vision space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.318, + 0.408, + 0.334 + ], + "angle": 0, + "content": "3.3. Multimodal Retrieval Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.484, + 0.61 + ], + "angle": 0, + "content": "We can directly retrieve the most similar data center from the SMGD to improve the final prediction. However, the CLIP model was optimized to reduce the inter-modal (vision-text) similarities rather than the intra-modal (vision-vision or text-text) similarities [24]. This optimization process leaves a disadvantage that similar images in the vision (image) feature space are not well clustered [43]. As shown in Fig. 3, the cosine similarity distributions of matched and unmatched image-text pairs (inter-modal) exhibit less overlap than those of matched and unmatched image-image pairs (intra-modal). This indicates that matched and unmatched pairs are more easily distinguishable in the multimodal space than in the vision-only space, a distinction shaped by CLIP's image-text contrastive learning. Therefore, in addition to the vision space retrieval augmentation, we also added a multimodal space retrieval augmentation, which compares the similarities of the samples in the CLIP text-image space rather than the vision space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Vision Space Retrieval Augmentation. To fully leverage the available information from the vision space, we include two terms: similarity retrieval and discriminant analysis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.484, + 0.746 + ], + "angle": 0, + "content": "Similarity Retrieval. This term treats the test sample as a query and retrieves the most similar class center from the SMGD based on their similarities within the vision space. Given the SMGD \\( S_{G} = \\{G, L\\} \\) and test image features \\( f^{t} \\) generated from the CLIP's image encoder, the prediction can be calculated as follows" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.754, + 0.483, + 0.773 + ], + "angle": 0, + "content": "\\[\nP _ {s i m} \\left(f ^ {t}\\right) = L A \\left(G ^ {T} f ^ {t}\\right) \\in \\mathbb {R} ^ {K}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.81 + ], + "angle": 0, + "content": "where \\( A(z) = \\exp (-\\beta (1 - z)) \\) is an adaptation function with a sharpness ratio \\( \\beta \\) [54]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Discriminant Analysis. This term leverages probabilistic information derived from the Gaussian means and covariances of the SMGD by using a closed-form expression from Gaussian discriminant analysis [2]. Under the assumption that all Gaussian components share an identical covariance \\(\\Sigma^t\\), the discriminant score can be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.089, + 0.907, + 0.148 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Omega_ {d i s c} (f ^ {t}) = G ^ {T} \\Sigma^ {t - 1} f ^ {t} - \\frac {1}{2} \\operatorname {d i a g} \\left(G ^ {T} \\Sigma^ {t - 1} G\\right) \\tag {11} \\\\ + \\log \\frac {1}{K} \\mathbf {1} _ {K} \\in \\mathbb {R} ^ {K}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.158, + 0.907, + 0.235 + ], + "angle": 0, + "content": "where \\(\\Sigma^t\\) is the mean of \\(K\\) covariances \\(\\{\\Sigma_k^t\\}_{k=1}^K\\). \\(\\Omega_{disc}(f^t)\\) contains scores that reflect how well \\(f^t\\) matches different classes, considering both the distance from the class mean and the spread of the class distribution. Therefore, the discriminant analysis-based term can be given as" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.247, + 0.907, + 0.267 + ], + "angle": 0, + "content": "\\[\nP _ {d i s c} \\left(f ^ {t}\\right) = L \\Omega_ {d i s c} \\left(f ^ {t}\\right) \\in \\mathbb {R} ^ {K}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.277, + 0.906, + 0.307 + ], + "angle": 0, + "content": "After that, the prediction from the vision space \\(\\mathcal{V}\\) retrieval augmentation can then be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.317, + 0.907, + 0.337 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathrm {R}} ^ {\\mathcal {V}} \\left(f ^ {t}\\right) = P _ {s i m} \\left(f ^ {t}\\right) + P _ {d i s c} \\left(f ^ {t}\\right) \\in \\mathbb {R} ^ {K}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.906, + 0.482 + ], + "angle": 0, + "content": "Multimodal Space Retrieval Augmentation. During test-time adaptation, multimodal space retrieval augmentation treats the test sample as a query and retrieves the most similar Gaussian center from the SMGD by evaluating their distances within CLIP's multimodal space. This process involves measuring the Kullback-Leibler (KL) divergence between the image features of the test sample or Gaussian centers and the textual features of the categories in CLIP's embedding space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.906, + 0.543 + ], + "angle": 0, + "content": "Given the SMGD \\( S_{G} = \\{G,L\\} \\) and the text embeddings \\( Z\\in \\mathbb{R}^{D\\times K} \\) generated from CLIP's text encoder, we transform Gaussian centers \\( G \\) of the SMGD from the vision space \\( \\mathcal{V} \\) to the multimodal space \\( \\mathcal{M} \\) as \\( \\Psi \\):" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.553, + 0.907, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\Psi = \\sigma \\left(Z ^ {T} G\\right) \\in \\mathbb {R} ^ {K \\times K}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.582, + 0.906, + 0.613 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) is a sigmoid function. We also project the test sample \\(f^{t}\\) to the multimodal space \\(\\mathcal{M}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.623, + 0.905, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\psi = \\sigma \\left(Z ^ {T} f ^ {t}\\right) \\in \\mathbb {R} ^ {K}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.652, + 0.906, + 0.695 + ], + "angle": 0, + "content": "We use Kullback-Leibler (KL) divergence to compare the similarities of the two distributions in the multimodal space as" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.698, + 0.905, + 0.714 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {k} = K L (\\psi | | \\Psi_ {k}), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.906, + 0.834 + ], + "angle": 0, + "content": "where \\(\\Psi_{k}\\) is the \\(k\\)-th column vector of \\(\\Psi\\), and \\(KL(P|Q) = \\sum_{x\\in \\mathcal{X}}P(x)\\log \\left(\\frac{P(x)}{Q(x)}\\right)\\). Therefore, \\(\\Phi = [\\Phi_1,\\Phi_2,\\dots ,\\Phi_K]\\in \\mathbb{R}^K\\) consists of KL-divergence values. Since the most similar samples have low KL-divergence values (close to 0), they receive small weights. To address this, we negate the values in \\(\\Phi\\). The prediction of the multimodal space \\(\\mathcal{M}\\) retrieval augmentation will be given as" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.846, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\nP _ {R} ^ {\\mathcal {M}} = - L \\Phi \\in \\mathbb {R} ^ {K}. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Then, we can obtain the output of the multimodal retrieval augmentation (MRA) module \\(P_{\\mathrm{R}}\\) as a combination of the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8814" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.267 + ], + "angle": 0, + "content": "
MethodAircraftCaltech101CarsDTDEuroSATFlower102Food101PetsSUN397UCF101Average
CLIP-ViT-B/1623.2293.5566.1145.0450.4266.9982.8686.9265.6365.1664.59
CoOp18.4793.7064.5141.9246.3968.7185.3089.1464.1566.5563.88
CoCoOp22.2993.7964.9045.4539.2370.8583.9790.4666.8968.4464.63
TPT24.7894.1666.8747.7542.4468.9884.6787.7965.5068.0465.10
DiffTPT25.6092.4967.0147.0043.1370.1087.2388.2265.7462.6765.47
MTA25.3294.1368.0545.5938.7168.2684.9588.2264.9868.1164.63
DN24.3093.6064.0045.7053.3068.0086.0087.7066.5068.4065.75
ZERO25.2193.6668.0446.1234.3367.6886.5387.7565.0367.7767.72
DMN24.8494.1265.6444.3947.7771.3884.4889.0766.2866.7565.47
TDA23.9194.2467.2847.4058.0071.4286.1488.6367.6270.6667.53
TT-RAA (ours)25.3894.0866.4247.9966.1272.6886.0989.8367.6971.2968.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.277, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Table 1. Results on the cross-domain benchmark. Our TT-RAA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, four training-based adaptation methods (CoOp, CoCoOp, TPT, and DiffTPT), and five training-free adaptation methods (MTA, DN, DMN, ZERO, and TDA)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.483, + 0.36 + ], + "angle": 0, + "content": "vision space retrieval argumentation result \\(P_{\\mathrm{R}}^{\\mathcal{V}}\\) and multimodal space augmentation result \\(P_{\\mathrm{R}}^{\\mathcal{M}}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.367, + 0.483, + 0.387 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathrm {R}} = P _ {\\mathrm {R}} ^ {\\mathcal {V}} + P _ {\\mathrm {R}} ^ {\\mathcal {M}}. \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.396, + 0.484, + 0.442 + ], + "angle": 0, + "content": "After that, we compute our final TT-RAA logits by combining the original CLIP logits \\( P_{\\mathrm{CLIP}} \\) with the MRA prediction \\( P_{\\mathrm{R}} \\) using a weighting factor \\( \\alpha \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.452, + 0.483, + 0.469 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathrm {T T} - \\mathrm {R A A}} = P _ {\\mathrm {C L I P}} + \\alpha P _ {\\mathrm {R}}. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.484, + 0.494 + ], + "angle": 0, + "content": "In the end, we can obtain the final prediction of TT-RAA as" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.509, + 0.483, + 0.533 + ], + "angle": 0, + "content": "\\[\n\\hat {y} = \\underset {k} {\\operatorname {a r g m a x}} P _ {\\mathrm {T T} - \\mathrm {R A A}} (k), \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.258, + 0.555 + ], + "angle": 0, + "content": "where \\(k\\in \\{1,2,\\dots,K\\}\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.567, + 0.224, + 0.584 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.188, + 0.606 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.84 + ], + "angle": 0, + "content": "To validate the method's test time adaptation capability across different domains and variations within a familiar domain, we conducted experiments on two benchmarks: a cross-domain (CD) benchmark comprising 10 datasets and an out-of-distribution (OOD) benchmark comprising 4 datasets. The CD benchmark primarily assesses transferability by testing on domains that diverge significantly from the original training distribution. Specifically, it includes Aircraft [26], Caltech101 [9], Cars [18], DTD [6], EuroSAT [13], Flower102 [29], Food101 [3], Pets [32], SUN397 [48], and UCF101 [40]. In contrast, the OOD benchmark focuses on robustness to distribution shifts within the same general domain. Specifically, it includes ImageNet-A [15], ImageNet-V2 [36], ImageNet-R [14], and ImageNet-S [47]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.31, + 0.866 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We implement our method based on the pre-trained CLIP architecture [34], which comprises a vision transformer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.907, + 0.495 + ], + "angle": 0, + "content": "(ViT) based image encoder [7] and a transformer-based text encoder [44]. Our experimental setup follows the challenging single-image test-time adaptation scenario, processing samples sequentially with a batch size of 1. To maintain computational efficiency and eliminate the need for backpropagation during inference, we utilize the hand-crafted prompts as proposed in [34] rather than learnable prompts. The experiments are conducted using the PyTorch framework, with evaluations performed on a single NVIDIA RTX A5000 GPU. Following standard practice in similar tasks, we employ top-1 accuracy as our evaluation metric." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.502, + 0.885, + 0.518 + ], + "angle": 0, + "content": "4.3. Comparisons with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.81 + ], + "angle": 0, + "content": "We conduct extensive experiments to evaluate TT-RAA against state-of-the-art approaches on the CD benchmark in Table 1 and OOD benchmark in Table 2. The competing methods include the CLIP baseline [34], training-based adaptation methods, and training-free adaptation methods. Discussions on Competing Methods. In the field of efficient domain adaptation, several methods require training, either in the training stage or at the test time, to align the model more closely with the target distribution. Context optimization (CoOp) [57], conditional context optimization (CoCoOp) [56] both involve learning prompts during a training phase. CoOp focuses on optimizing prompt tokens that are fixed for each domain, while CoCoOp extends this to generate a conditional prompt token for each image. Test-time prompt tuning (TPT) [38] and diffusion model for TPT (DiffTPT) [10], on the other hand, adapt the model at test time by tuning prompt embeddings. TPT performs direct gradient-based prompt tuning using test data, while DiffTPT uses new data generated by diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In addition to training-based methods, we also compare with several training-free adaptation methods, including two distribution-based methods, one entropy-based method, and two cache-based methods. Distribution-based methods evaluated are MeanShift for Test-time Augmentation (MTA) [51] and Distribution Normalization (DN) [58]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8815" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.218, + 0.089, + 0.781, + 0.276 + ], + "angle": 0, + "content": "
MethodImageNet-AImageNet-V2ImageNet-RImageNet-SAverage
CLIP-ViT-B/1649.8961.8877.6548.2459.42
CoOp49.7164.2075.2147.9959.28
CoCoOp50.6364.0776.1848.7559.91
TPT54.7763.4577.0647.9460.81
DiffTPT55.6865.1075.0046.8060.52
MTA57.4163.6176.9248.5861.63
DN58.7162.8980.2048.9462.69
ZERO59.6164.1677.2248.4062.35
DMN58.2865.1778.5553.2063.80
TDA60.1164.6780.2450.5463.89
TT-RAA (Ours)60.5964.6980.5849.9863.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.287, + 0.907, + 0.33 + ], + "angle": 0, + "content": "Table 2. Results on the out-of-distribution benchmark. Our TT-RAA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, four training-based adaptation methods (CoOp, CoCoOp, TPT, and DiffTPT), and five training-free adaptation methods (MTA, ZERO, DN, DMN, and TDA)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.482, + 0.519 + ], + "angle": 0, + "content": "They improve the test data distribution estimation via MeanShift algorithm and distribution normalization, respectively. Cache-based methods include the Dual Memory Network (DMN) [55] and TDA [17]. DMN comprises a dynamic and a static cache, while TDA employs a positive and a negative cache. For a fair comparison, we reproduced DMN's zero-shot results using our standardized prompts instead of DMN's original LLM-generated prompts, and we reduced the batch size to 1. We also compare against ZERO [8], an entropy-based method. ZERO sets the temperature of most confident predictions as zero to approximate marginal entropy minimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Results on the Cross-domain Benchmark. Despite having the advantage of task-specific training, training-based adaptation methods, such as CoOp, CoCoOp, TPT, and DiffTPT, show limited generalization capabilities. TT-RAA outperforms these methods by significant margins (about \\(3\\% - 4\\%\\)). On fine-grained recognition tasks such as Flowers102, TT-RAA achieves \\(72.68\\%\\) accuracy, surpassing CoCoOp \\((70.85\\%)\\) by \\(1.83\\%\\). This gap widens further on specialized datasets like SUN397, where TT-RAA \\((67.69\\%)\\) demonstrates superior feature adaptation compared to CoCoOp \\((64.15\\%)\\). These results suggest that static prompt learning, even with input-conditional mechanisms, may not be sufficient enough for handling significant domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In comparison to existing training-free adaptation approaches, TT-RAA demonstrates consistent superiority across various domain shifts. On remote sensing data (EuroSAT), TT-RAA \\((66.12\\%)\\) significantly outperforms TDA \\((58.00\\%)\\) by \\(8.12\\%\\). This substantial gain can be attributed to our SMGD effectively capturing the unique characteristics of satellite imagery. Similarly, for action recognition (UCF101), TT-RAA \\((71.29\\%)\\) surpasses TDA \\((70.66\\%)\\) by \\(0.63\\%\\), demonstrating robust feature adaptation. The performance advantages extend to other specialized domains as well. This consistent improvement across diverse domains validates the effectiveness of our method in bridging" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.338, + 0.738, + 0.352 + ], + "angle": 0, + "content": "domain gaps in test-time settings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.354, + 0.905, + 0.61 + ], + "angle": 0, + "content": "Results on the Out-of-distribution Benchmark. The experimental results in Table 2 demonstrate that our proposed TT-RAA method achieves state-of-the-art performance on the OOD benchmark with an average accuracy of \\(63.96\\%\\) across all ImageNet variants, outperforming both training-based adaptation methods (CoOp, CoCoOp, TPT, DiffTPT) and training-free approaches (MTA, DN, ZERO, DMN, TDA). Notably, TT-RAA exhibits good performance on ImageNet-R \\((80.58\\%)\\) and ImageNet-A \\((60.59\\%)\\), highlighting its robust generalization capabilities under diverse distribution shifts. When examining performance across different distribution shifts, we observe that most methods achieve their highest accuracy on ImageNet-R and ImageNet-V2. ImageNet-S with sketches poses the greatest challenge for all methods, indicating that semantic variations remain difficult to address even with advanced adaptation techniques." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.68, + 0.636 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.905, + 0.748 + ], + "angle": 0, + "content": "We conduct ablation studies on the CD benchmark to systematically evaluate the effectiveness of each proposed component in TT-RAA. Figure 4 presents the quantitative results of our analysis with both ResNet-50 and ViT-B/16 backbones. We establish our baseline using the TDA method with one-shot capacity, which achieves \\(59.82\\%\\) (ResNet-50) and \\(67.04\\%\\) (ViT-B/16) accuracy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.75, + 0.905, + 0.87 + ], + "angle": 0, + "content": "Streaming Mixture of Gaussian Database Module. When integrating the SMGD module with the baseline, we observe a performance improvement of about \\(1.4\\%\\) with ViT-B/16 and \\(1.2\\%\\) with ResNet-50. This enhancement demonstrates that our SMGD effectively captures the dynamic distribution of test samples while maintaining computational efficiency. Its adaptive nature in updating class statistics online is useful for handling distribution shifts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Multimodal Retrieval Augmentation Module. Incorporating the MRA module also improved the performance by" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "8816" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.091, + 0.473, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.242, + 0.485, + 0.3 + ], + "angle": 0, + "content": "Figure 4. Ablation studies of TT-RAA. The SMGD and MRA modules independently improve performance over the baseline. Their combination yields further performance gains, demonstrating complementary benefits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.431 + ], + "angle": 0, + "content": "about \\(0.4\\%\\) and \\(0.2\\%\\) with ViT-B/16 and ResNet-50, respectively. This gain supports our hypothesis regarding vision space misalignment in CLIP, which stems from its vision-text joint optimization objective. The performance improvement demonstrates that projecting both the SMGD database and test samples from vision space to multimodal space effectively addresses this misalignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.536 + ], + "angle": 0, + "content": "Complementary Benefits of Proposed Modules. When combining both SMGD and MRA modules (denoted as TT-RAA), we achieve the best performance of \\(68.76\\%\\) with ViT-B/16 and \\(61.25\\%\\) with ResNet-50, representing a substantial improvement of \\(1.72\\%\\) and \\(1.43\\%\\) over the baseline. These gains suggest that SMGD and MRA offer complementary and synergistic benefits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Analysis of Feature Backbone Impact. The effectiveness of our approach is further validated across two vision ecoders: ViT-B/16 and ResNet-50. These consistent improvements across different backbones suggest that our adaptation strategy effectively addresses fundamental domain shift challenges rather than exploiting architecture-specific characteristics." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.653, + 0.327, + 0.669 + ], + "angle": 0, + "content": "4.5. Hyperparameter Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The hyperparameter analysis of the update coefficient \\(\\eta\\) on the UCF101 is shown in Fig. 5a. When \\(\\eta = 0\\), the model maintains a static Gaussian center initialized at the start, resulting in poor adaptation. From 0 to 0.55, accuracy increases sharply, highlighting the importance of incorporating new information. However, the performance deterioration beyond 0.65 reveals a critical threshold where excessive emphasis on new samples compromises the model's stability. This observation aligns with theoretical expectations, as larger \\(\\eta\\) values reduce the influence of historical information, potentially leading to over-adaptation to new samples. The empirical optimal value of \\(\\eta = 0.65\\) suggests that maintaining approximately \\(35\\%\\) of historical information while incorporating \\(65\\%\\) of new information achieves the best trade-off." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.09, + 0.707, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.205, + 0.623, + 0.217 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.094, + 0.898, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.205, + 0.815, + 0.216 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.229, + 0.908, + 0.287 + ], + "angle": 0, + "content": "Figure 5. (a) Hyperparameter analysis of the update coefficient \\(\\eta\\). It shows that the best coefficient is 0.65. (b) Experiment with additional data. The method can accept additional data, which improve the performance significantly." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.311, + 0.89, + 0.327 + ], + "angle": 0, + "content": "4.6. Experiments with Access to Additional Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.907, + 0.621 + ], + "angle": 0, + "content": "In the standard test-time adaptation setting, access to the target domain's training set is typically restricted. However, in many cases, the training set data might be available. Finetuning large-scale vision-language models in such scenarios remains computationally expensive. This raises an important question: how can performance be improved using the training data without computationally expensive finetuning? With a slight modification to TT-RAA, we propose leveraging the target domain training set (if accessible prior to test time) to construct a database and calculate class centers using a Gaussian mixture model. By doing so, we perform the same MRA process as in TT-RAA. This modified approach, referred to as TT-RAA-A, demonstrates significant performance improvements while maintaining a low computational requirement. The corresponding results are presented in Fig. 5b, where TT-RAA-A outperforms TT-RAA by a large margin, highlighting the benefits of incorporating target training data in a training-free manner, which open a new direction on adaptation without training." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.634, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced TT-RAA, a novel test-time retrieval augmented adaptation method that effectively addresses the computationally expensive domain adaptation challenges faced by vision-language models. Experimental results validated the effectiveness of SMGD by continuously estimating the test distribution and compressing all useful information into a single representation for each class without discarding old samples. By transforming Gaussian centers from vision space into a shared multimodal space, MRA enables more effective retrieval that leverages both visual and textual features. The success of TT-RAA opens up new possibilities for deploying vision-language models in real-world applications where computational resources are limited and rapid adaptation to new domains is crucial. Further exploration on adapting our method to larger multimodal foundation and large language models is promising." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8817" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.143 + ], + "angle": 0, + "content": "[1] Meta AI. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. Meta AI Report, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.145, + 0.483, + 0.171 + ], + "angle": 0, + "content": "[2] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.173, + 0.482, + 0.213 + ], + "angle": 0, + "content": "[3] Lukas Bossard, Matthieu Guillaumin, et al. Food-101: Mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.482, + 0.282 + ], + "angle": 0, + "content": "[4] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. VisualGPT: Data-efficient adaptation of pretrained language models for image captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18030-18040, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.482, + 0.34 + ], + "angle": 0, + "content": "[5] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. European Conference on Computer Vision, pages 104-120, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.342, + 0.482, + 0.382 + ], + "angle": 0, + "content": "[6] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.384, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.482, + 0.494 + ], + "angle": 0, + "content": "[8] Matteo Farina, Gianni Franchi, Giovanni Iacca, Massimiliano Mancini, and Elisa Ricci. Frustratingly easy test-time adaptation of vision-language models. Advances in Neural Information Processing Systems, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[9] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.566, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[10] Chun-Mei Feng, Kai Yu, Yong Liu, et al. Diverse data augmentation with diffusions for effective test-time prompt tuning. In IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.622, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, et al. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.678, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[12] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, et al. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[13] Patrick Helber, Benjamin Bischke, et al. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[14] Dan Hendrycks, Steven Basart, Norman Mu, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[15] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. International Conference on Machine Learning, pages 4904-4916, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[17] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[18] Jonathan Krause, Michael Stark, et al. 3D object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshops, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[19] Youngjun Lee, Doyoung Kim, Junhyeok Kang, Jihwan Bang, Hwanjun Song, et al. RA-TTA: Retrieval-augmented test-time adaptation for vision-language models. In International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[20] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, and Vladimir others Karpukhin. Retrievalaugmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems, pages 9459-9474, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[21] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[22] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[23] Jian Liang, Ran He, et al. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, pages 1-34, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.547, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[24] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.904, + 0.702 + ], + "angle": 0, + "content": "[26] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, et al. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[27] Yifei Ming and Yixuan Li. Understanding retrieval-augmented task adaptation for vision-language models. In International Conference on Machine Learning, pages 35719-35743, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[28] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, et al. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[29] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8818" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.149 + ], + "angle": 0, + "content": "[31] OpenAI. Gpt-4v(ison) system card. In OpenAI Technical Report, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[32] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[33] Rizwan Qureshi, Ranjan Sapkota, Abbas Shah, et al. Thinking beyond tokens: From brain-inspired intelligence to cognitive foundations for artificial general intelligence and its societal impact. arXiv preprint arXiv:2507.00951, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.307, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[35] Aditi Raghunathan, Prateek Jain, et al. Learning mixture of gaussians with streaming data. Advances in Neural Information Processing Systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.391 + ], + "angle": 0, + "content": "[36] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, et al. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.393, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[37] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, et al. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[38] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[39] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[40] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[41] Tao Sun, Cheng Lu, and Haibin Ling. Prior knowledge guided unsupervised domain adaptation. In European Conference on Computer Vision. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[42] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[43] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. SuS-X: Training-free name-only transfer of vision-language models. In IEEE/CVF International Conference on Computer Vision, pages 2725-2736, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[45] Matthew Wallingford, Vivek Ramanujan, Alex Fang, Aditya Kusupati, Roozbeh Mottaghi, Aniruddha Kembhavi, et al." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Neural priming for sample-efficient adaptation. Advances in Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[46] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[47] Haohan Wang, Songwei Ge, et al. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[48] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In IEEE Conference on Computer Vision and Pattern Recognition, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[49] Shiqi Yang, Joost Van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. Advances in Neural Information Processing Systems, 34:29393-29405, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[50] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. International Conference on Learning Representations, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[51] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23783-23793, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[52] Jingyi Zhang, Jiaxing Huang, et al. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[53] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. VinVL: Revisiting visual representations in vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[54] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision, pages 493-510. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[55] Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28718-28728, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[58] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser Nam Lim. Test-time distribution normalization for contrastively learned visual-language models. Advances in Neural Information Processing Systems, pages 47105-47123, 2023. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8819" + } + ] +] \ No newline at end of file diff --git a/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/abcad282-0784-41dd-b230-4b9b3488f3c1_origin.pdf b/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/abcad282-0784-41dd-b230-4b9b3488f3c1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..60931c780ee9719feb16ca69277b53cbe72192f3 --- /dev/null +++ b/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/abcad282-0784-41dd-b230-4b9b3488f3c1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f5f0f1762cb78c392742793858abf157dd9646b749193251c351fb17e914f19a +size 482601 diff --git a/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/full.md b/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..17753d836a6a5fbd3feec2cf9a4845df8c7d18d0 --- /dev/null +++ b/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/full.md @@ -0,0 +1,346 @@ +# Test-Time Retrieval-Augmented Adaptation for Vision-Language Models + +Xinqi Fan $^{1}$ , Xueli Chen $^{2}$ , Luoxiao Yang $^{3}$ , Chuin Hong Yap $^{1}$ , Rizwan Qureshi $^{4}$ , Qi Dou $^{5}$ , Moi Hoon Yap $^{1}$ , Mubarak Shah $^{4}$ + +$^{1}$ Manchester Metropolitan University $^{2}$ Hong Kong Metropolitan University $^{3}$ Xi'an University of Technology $^{4}$ University of Central Florida $^{5}$ Chinese University of Hong Kong + +# Abstract + +Vision-language models (VLMs) have shown promise in test-time adaptation tasks due to their remarkable capabilities in understanding and reasoning about visual content through natural language descriptions. However, training VLMs typically demands substantial computational resources, and they often struggle to adapt efficiently to new domains or tasks. Additionally, dynamically estimating the test distribution from streaming data at test time remains a significant challenge. In this work, we propose a novel test-time retrieval-augmented adaptation (TT-RAA) method that enables VLMs to maintain high performance across diverse visual recognition tasks without the need for task-specific training or large computational overhead. During inference, TT-RAA employs a streaming mixture of Gaussian database (SMGD) to continuously estimate test distributions, requiring minimal storage. Then, TT-RAA retrieves the most relevant information from the SMGD, enhancing the original VLM outputs. A key limitation of CLIP-based VLMs is their inter-modal vision-language optimization, which does not optimize vision-space similarity, leading to larger intra-modal variance. To address this, we propose a multimodal retrieval augmentation module that transforms the SMGD into a unified multimodal space, enabling retrieval that aligns both vision and language modalities. Extensive experiments across both cross-domain and out-of-distribution benchmarks comprising fourteen datasets demonstrate TT-RAA's superior performance compared to state-of-the-art methods. Ablation studies and hyperparameter analyses further validate the effectiveness of the proposed modules. The source code of our work is available at https://github.com/xinqi-fan/TT-RAA. + +# 1. Introduction + +Vision-language models (VLMs) have significantly advanced visual understanding by leveraging natural language supervision with contrastive language-image pre-training (CLIP) notably pioneering this field [34]. The zero-shot capability enables this model to be deployed in various do + +![](images/3deb82f03189eccad0fff94c65752ec7775e644dc2e4c0a9a72c340c33e33994.jpg) +Figure 1. Illustration of our proposed test time retrieval augmented adaptation (TT-RAA). Without TT-RAA, the CLIP prediction will be the final prediction. With TT-RAA, the image feature will be used to retrieve the most similar database centers which will be used to improve the final prediction in a training-free manner. + +mains. This foundational work has catalyzed a wave of architectural innovations, including domain-specific variants and architectural enhancements that build upon CLIP's core principles [52]. Some variants have focused on scaling up capacities, learning from noisy text supervision, and fine-grained alignment [16, 50]. However, these models consistently encounter challenges when deployed in specialized domains, where the distribution of images often differs substantially from their web-scale training data [11, 33]. This challenge has become even more prominent with the recent introduction of more sophisticated large VLMs such as GPT-4V [31], LLaVA [25], and Llama 3.2 Vision [1]. While these advanced models excel at integrating visual and textual data, the high computational cost of fine-tuning them for domain-specific tasks poses a major challenge, especially in specialized visual domains that differ significantly from general web images. + +Test-time adaptation (TTA) has emerged as a promising solution that enables models to adapt to new domains without accessing source domain training data [23]. Traditional TTA approaches include self-training-based methods that leverage pseudo-labels [22, 41, 49], self-supervised learning-based methods that incorporate aux + +iliary tasks [42], batch normalization [28, 37], and entropy minimization-based methods [30, 46]. For VLMs, recent works have explored prompt-based adaptation methods [10, 38] to bridge the domain gap between web-scale training data and specific downstream tasks. However, these methods often incur substantial computational overhead, highlighting the need for more efficient approaches [10, 17]. Advancements in training-free methods have explored the use of cache mechanisms [17, 55] and distribution-based methods [51, 58] to improve the efficiency. Despite these advances, dynamically estimating the test distribution from streaming data at test time remains a significant challenge, due to the test sample arriving one at a time. + +In this work, we introduce test-time retrieval-augmented adaptation (TT-RAA), a novel training-free method that allows VLMs to sustain strong performance across various visual recognition tasks without requiring task-specific training or high computational costs (Fig. 1). TT-RAA is a training-free framework, which uses a streaming mixture of Gaussian database (SMGD) to dynamically capture test distribution characteristics with minimal storage requirements at inference time. TT-RAA retrieves relevant information from the SMGD to enhance the performance of the original VLM. Furthermore, an inherent limitation of CLIP-based VLMs is their lack of optimization for intramodal similarity, which does not optimize similarity in the vision space, resulting in greater intramodal variance [24]. To mitigate this, we introduce a multimodal retrieval augmentation (MRA) module that projects the SMGD into a shared multimodal embedding space, facilitating retrieval that harmonizes vision and language modalities. Extensive testing across ten cross-domain datasets and four out-of-distribution datasets confirms that TT-RAA delivers superior performance relative to state-of-the-art methods, with additional ablation studies and hyperparameter analyses supporting the effectiveness of the proposed modules. The key contributions of this work include: + +- We introduce test-time retrieval-augmented adaptation (TT-RAA), an efficient, training-free framework that significantly improves domain adaptability without additional fine-tuning. +- We propose streaming mixture of Gaussian database (SMGD) for dynamic, efficient, and robust estimation of test domain statistics during inference. +- We propose multimodal retrieval augmentation (MRA), a novel mechanism addressing CLIP's inherent intra-modal similarity limitation, aligning both vision and language modalities effectively. + +The rest of this paper is organized as follows: Section 2 provides some related works, Section 3 describes the proposed method. Experimental results and performance evaluation are presented in Section 4. Finally, Section 5 concludes the article and outlines future research directions. + +# 2. Related Work + +# 2.1. Vision Language Model + +The development of vision language models (VLMs) can be traced back to earlier works that explored vision-language pre-training for specific tasks, where VisualBERT [21] and UNITER [5] pioneered joint representation learning between visual and textual modalities for tasks like visual question answering. The field witnessed a paradigm shift with CLIP [34], which demonstrated impressive zero-shot capabilities by learning transferable visual representations through contrastive learning between images and texts at web scale. ALIGN [16] further scaled up this approach with 1.8 billion noisy image-text pairs and introduced noise-robust training strategies. FILIP [50] enhanced CLIP by introducing fine-grained region-word alignment, enabling more detailed vision-language correspondences. FLAVA [39] introduced a single foundation model that can tackle vision, language, and multimodal tasks. More recently, the emergence of multimodal large language models has pushed the boundaries of vision-language understanding, where GPT-4V [31] extends GPT-4's language capabilities to handle visual inputs, LLaVA [25] achieves strong visual conversation abilities by aligning vision encoders with language models through minimal task-specific training, and Llama 3.2 [1] demonstrates remarkable vision and language capabilities with open models. Other notable VLMs include VinVL [53], which enhanced visual understanding by incorporating object detection and segmentation, and VisualGPT [4], which demonstrated zero-shot transfer to various vision-language tasks. These models have continued to push the boundaries of multimodal representation learning, enabling a wide range of applications from intelligent assistants to creative tools. While these newer models demonstrate impressive capabilities in combining visual and textual understanding, they face similar domain adaptation challenges but at an even larger scale. The computational demands for fine-tuning these models on domain-specific tasks present a significant barrier, particularly for specialized applications where the visual domain differs markedly from general web imagery. + +# 2.2. Test Time Adaptation + +Normal domain adaptation techniques require access to both source and target domain data, but test-time adaptation (TTA) adapts models to the target domain without source data, making it particularly useful for scenarios with privacy or legal restrictions [23]. To adapt a pre-trained model to a target domain, methods like self-training refine model predictions using pseudo-labels. Approaches such as centroid-based [22], neighbor-based [49], and optimization-based [41] strategies enhance the quality of these pseudolabels. Self-supervised methods introduce auxiliary tasks to + +align features, exemplified by test-time training (TTT) [42] with a self-supervised rotation prediction task. Batch normalization [28, 37] techniques and entropy minimization strategies [30, 46] have also been adopted in TTA tasks. + +Recent advancements focus on VLMs like CLIP, using domain-specific prompts from test data. Test-time prompt tuning (TPT) [38] fine-tunes a learnable prompt with each testing sample, while DiffTPT [10] uses pre-trained diffusion models to enhance test data diversity. Although these methods are promising, they still require a large amount of computational resources for their training-based adaptation, making it hard to adapt large VLMs. To reduce costs, a training-free dynamic adapter (TDA) [17] has been introduced, using a lightweight dynamic key-value cache without backpropagation. However, its cache only captures the streaming test data with a limited cache capacity without the ability to estimate the full spectrum of the test data. + +# 2.3. Retrieval Augmented Generation + +Retrieval-augmented generation (RAG) has emerged as a promising solution to enhance large language models (LLMs) by incorporating external knowledge during inference, effectively addressing challenges like hallucination and outdated knowledge [20]. In its naive form, RAG follows a three-stage process: indexing, where documents are chunked and encoded into vectors; retrieval, where relevant documents are retrieved based on semantic similarity with the input query; and generation, where retrieved documents along with the query are fed to LLMs for response generation [12]. Recent efforts have extended RAG to VLMs, including methods such as I2I/T2I adaptors [27], RA-TTA [19], and neural priming [45]. These methods rely on additional databases constructed from large-scale datasets like LAION-2B/5B, which often incur high computational and storage costs during both construction and inference. In contrast, our work focuses on constructing a database dynamically from incoming test samples, without relying on additional large-scale databases. In addition, we employed a multimodal retrieval augmentation module that transforms our database into a unified multimodal space, enabling retrieval that aligns both vision and language modalities. + +# 3. Method + +In this work, we introduce test-time retrieval-augmented Adaptation (TT-RAA) to adapt vision-language models (VLMs) to new domains (Fig. 2). TT-RAA leverages a streaming mixture of Gaussian database (SMGD) to model test data statistics in real time. As test data streams in, we dynamically update the SMGD with Gaussian centers representing the distribution of vision features. Each test sample's vision features query the SMGD to retrieve relevant Gaussian centers, providing contextual information. + +Since CLIP's joint vision-text optimization does not explicitly optimize vision-space similarity, we transform the Gaussian centers into a multimodal space using text embeddings. This allows retrieval in a space that is aligned with both vision and language modalities. Final predictions combine CLIP outputs with retrieval results from both vision and multimodal spaces, enhancing accuracy and robustness. TT-RAA thus provides a powerful framework for adapting VLMs to new domains by leveraging both vision and language information. + +# 3.1. Contrastive Language-Image Pretraining + +Contrastive language-image pre-training (CLIP) [34] is a vision-language model that has two separate encoders for processing images and text, respectively. These encoders are trained to project their respective inputs into a shared embedding space where related images and texts are positioned close to each other. Consider a $K$ -class classification problem where the CLIP's objective in the zero-shot setting is to match images with their most relevant textual descriptions using an image encoder $E_I$ and a text encoder $E_T$ . To obtain the textual descriptions $T$ , $K$ -class names are concatenated with hand-crafted prompts and then mapped into the $D$ -dimensional text embeddings $Z$ using the text encoder $E_T$ as + +$$ +Z = E _ {T} (T) \in \mathbb {R} ^ {D \times K}. \tag {1} +$$ + +Given the $t$ -th test samples $x^t$ , CLIP vision encoder produces image features + +$$ +f ^ {t} = E _ {I} \left(x ^ {t}\right) \in \mathbb {R} ^ {D}. \tag {2} +$$ + +CLIP was trained using contrastive loss to minimize the distance between corresponding text and image embeddings. When making predictions, the CLIP output can be obtained as + +$$ +P _ {\mathrm {C L I P}} ^ {t} = Z ^ {T} f ^ {t} \in \mathbb {R} ^ {K}. \tag {3} +$$ + +Then, we can obtain the final prediction for test sample $x^{t}$ as $\hat{y}^t = \operatorname*{argmax}_k P_{\mathrm{CLIP}}^t (k)$ , where $k \in \{1,2,\dots,K\}$ . + +# 3.2. Streaming Mixture of Gaussian Database + +Our work addresses a challenging training-free test-time adaptation problem where models must adapt to streaming test data under two constraints: (1) Both input and output distributions of the source and target domains may be subject to distribution shifts. (2) Test samples arrive sequentially, one at a time, rather than in batches, reflecting real-world deployment. This setting is demanding because it requires training-free, zero-shot adaptation without access to batch statistics. Therefore, it is important to estimate the statistics of the new domain by capturing its full spectrum, and based on this we can retrieve the most relevant information to enhance the final prediction. In this work, we propose a Streaming Mixture of Gaussian Database (SMGD) + +![](images/48eaa854a7bc8dfedcc7f1a059c0ccea08966a7c72bb9e3f052827cc67cf1ee8.jpg) +Figure 2. Overview of the proposed test-time retrieval-augmented adaptation (TT-RAA) method. TT-RAA leverages a streaming mixture of gaussian database (SMGD) to dynamically model the distribution of streaming test data in vision space $\mathcal{V}$ . The multimodal retrieval augmentation (MRA) module projects these Gaussian centers into a unified multimodal space $\mathcal{M}$ , enabling effective retrieval aligned across both vision and language modalities. Combining retrieval results from vision and multimodal spaces with CLIP's original predictions, TT-RAA significantly enhances VLM performance without additional training. + +module by using a Gaussian mixture model to estimate the dynamic statistics from the test time streaming data. + +We assume that feature embedding $f^t$ of the $t$ -th sample $x_{t}$ is drawn from a mixture of Gaussian distributions + +$$ +f ^ {t} \sim \sum_ {k} w _ {k} \mathcal {N} \left(\mu_ {k}, \Sigma_ {k}\right), \forall k = 1, 2, \dots , K, \tag {4} +$$ + +where $\mu_k \in \mathbb{R}^D$ and $\Sigma_k \in \mathbb{R}^{D \times D}$ are the mean and covariance of the $k$ -th Gaussian component $\mathcal{N}$ respectively. The mixture weights $w_k \geq 0$ , $\sum_k w_k = 1$ , and $K$ is the number of classes. For simplicity, we assume the weights $w_k = p(y = k) = 1 / K$ . Here, SMGD is a key-value database, where the keys are Gaussian mean vectors $\{\mu_k\}_{k=1}^K$ and the values are the corresponding pseudo labels of these Gaussian mean vectors. The covariance matrices $\{\Sigma_k\}_{k=1}^K$ are also recorded which will involve in the retrieval during test time adaptation. Given the test time setting, all samples come in a stream. Our goal is to estimate the optimal mean $\mu_k$ and covariance $\Sigma_k$ for the class $k$ . + +To estimate the optimal value of the mean $\mu_{k}$ and covariance $\Sigma_{k}$ , we draw inspiration from streaming data statistics estimation [35]. Given a test sample $x^{t}$ at time $t$ , its pseudo label $\hat{l}_t$ is determined by its CLIP prediction as $\hat{l}_t = \operatorname*{argmax}_{k} P_{\mathrm{CLIP}}^t (k)$ . Initially, we set the mean vector $\mu_{k}$ to the first test sample with its pseudo label as $k$ from the streaming test data. We set the initial covariance $\Sigma_{k} = \sigma^{2}\mathbf{I}$ with $\sigma = 0.1$ . We also record the corresponding entropy $h_k$ as the entropy of the first test sample's CLIP logit. Once the mean vector $\mu_{k}$ and covariance matrix $\Sigma_{k}$ are initialized, we progressively update them using selected test samples with high-quality pseudo labels, as measured by entropy. If the test sample $x^{t}$ at time $t$ has a pseudo label $\hat{l}^t = k$ and an + +entropy smaller than $h_k^{t-1}$ , the mean $\mu_k^t$ and the covariance $\Sigma_k^t$ at time $t$ are updated as + +$$ +\mu_ {k} ^ {t} = (1 - \eta) \mu_ {k} ^ {t - 1} + \eta f ^ {t}, \tag {5} +$$ + +$$ +\Sigma_ {k} ^ {t} = (1 - \eta) \Sigma_ {k} ^ {t - 1} + \eta \left(f ^ {t} - \mu_ {k} ^ {t}\right) \left(f ^ {t} - \mu_ {k} ^ {t}\right) ^ {T}, \tag {6} +$$ + +where $\eta$ is the update coefficient. Otherwise, $\mu_k^t = \mu_k^{t-1}$ and $\Sigma_k^t = \Sigma_k^{t-1}$ for $k \in \{1, 2, \dots, K\}$ , meaning the mean vectors and the covariance matrices remain unchanged if there is no test sample with pseudo label $k$ or if the test sample does not have a high-quality pseudo label. A high-quality pseudo label is identified when its entropy is lower than that of the Gaussian mean vector. Once the mean and covariance are updated according to Eqs. (5) and (6), the corresponding entropy is updated accordingly as + +$$ +h _ {k} ^ {t} = (1 - \eta) h _ {k} ^ {t - 1} + \eta H \left(P _ {\mathrm {C L I P}} ^ {t}\right), \tag {7} +$$ + +where $\mathrm{H}(P_{\mathrm{CLIP}}^t) = -\sum_{k=1}^{K} P_{\mathrm{CLIP}}^t(k) \log(P_{\mathrm{CLIP}}^t(k))$ is the entropy of the CLIP logits $P_{\mathrm{CLIP}}^t$ . + +In this way, we keep updating a high-quality database with high confidence as measured by the entropy of test-time streaming data's pseudo labels. The SMGD $S_{G}$ is composed of the following: + +$$ +S _ {G} = \{G, L \}, \tag {8} +$$ + +$$ +G = \left[ \mu_ {0} ^ {t}, \mu_ {1} ^ {t}, \dots , \mu_ {K} ^ {t} \right], \tag {9} +$$ + +where $G \in \mathbb{R}^{D \times K}$ represents the Gaussian centers of the SMGD, and each column of $L \in \mathbb{R}^{K \times K}$ denotes a $K$ -dimensional one-hot pseudo-label vector. Along with SMGD, we have the covariances $\{\Sigma_0^t, \Sigma_1^t, \dots, \Sigma_K^t\}$ . + +![](images/121c712bc59063ad88e26279103aff13c2399aba643331d2b1c7b9a2486b98c2.jpg) +Figure 3. Motivation of multimodal retrieval augmentation. Cosine similarity distributions of matched and unmatched image-text pairs (inter-modal) exhibit less overlap than those of matched and unmatched image-image pairs (intra-modal). This suggests that samples are more easily distinguishable in the multimodal space than in the vision space. + +# 3.3. Multimodal Retrieval Augmentation + +We can directly retrieve the most similar data center from the SMGD to improve the final prediction. However, the CLIP model was optimized to reduce the inter-modal (vision-text) similarities rather than the intra-modal (vision-vision or text-text) similarities [24]. This optimization process leaves a disadvantage that similar images in the vision (image) feature space are not well clustered [43]. As shown in Fig. 3, the cosine similarity distributions of matched and unmatched image-text pairs (inter-modal) exhibit less overlap than those of matched and unmatched image-image pairs (intra-modal). This indicates that matched and unmatched pairs are more easily distinguishable in the multimodal space than in the vision-only space, a distinction shaped by CLIP's image-text contrastive learning. Therefore, in addition to the vision space retrieval augmentation, we also added a multimodal space retrieval augmentation, which compares the similarities of the samples in the CLIP text-image space rather than the vision space. + +Vision Space Retrieval Augmentation. To fully leverage the available information from the vision space, we include two terms: similarity retrieval and discriminant analysis. + +Similarity Retrieval. This term treats the test sample as a query and retrieves the most similar class center from the SMGD based on their similarities within the vision space. Given the SMGD $S_{G} = \{G, L\}$ and test image features $f^{t}$ generated from the CLIP's image encoder, the prediction can be calculated as follows + +$$ +P _ {s i m} \left(f ^ {t}\right) = L A \left(G ^ {T} f ^ {t}\right) \in \mathbb {R} ^ {K}, \tag {10} +$$ + +where $A(z) = \exp (-\beta (1 - z))$ is an adaptation function with a sharpness ratio $\beta$ [54]. + +Discriminant Analysis. This term leverages probabilistic information derived from the Gaussian means and covariances of the SMGD by using a closed-form expression from Gaussian discriminant analysis [2]. Under the assumption that all Gaussian components share an identical covariance $\Sigma^t$ , the discriminant score can be computed as + +$$ +\begin{array}{l} \Omega_ {d i s c} (f ^ {t}) = G ^ {T} \Sigma^ {t - 1} f ^ {t} - \frac {1}{2} \operatorname {d i a g} \left(G ^ {T} \Sigma^ {t - 1} G\right) \tag {11} \\ + \log \frac {1}{K} \mathbf {1} _ {K} \in \mathbb {R} ^ {K}, \\ \end{array} +$$ + +where $\Sigma^t$ is the mean of $K$ covariances $\{\Sigma_k^t\}_{k=1}^K$ . $\Omega_{disc}(f^t)$ contains scores that reflect how well $f^t$ matches different classes, considering both the distance from the class mean and the spread of the class distribution. Therefore, the discriminant analysis-based term can be given as + +$$ +P _ {d i s c} \left(f ^ {t}\right) = L \Omega_ {d i s c} \left(f ^ {t}\right) \in \mathbb {R} ^ {K}. \tag {12} +$$ + +After that, the prediction from the vision space $\mathcal{V}$ retrieval augmentation can then be calculated as + +$$ +P _ {\mathrm {R}} ^ {\mathcal {V}} \left(f ^ {t}\right) = P _ {s i m} \left(f ^ {t}\right) + P _ {d i s c} \left(f ^ {t}\right) \in \mathbb {R} ^ {K}, \tag {13} +$$ + +Multimodal Space Retrieval Augmentation. During test-time adaptation, multimodal space retrieval augmentation treats the test sample as a query and retrieves the most similar Gaussian center from the SMGD by evaluating their distances within CLIP's multimodal space. This process involves measuring the Kullback-Leibler (KL) divergence between the image features of the test sample or Gaussian centers and the textual features of the categories in CLIP's embedding space. + +Given the SMGD $S_{G} = \{G,L\}$ and the text embeddings $Z\in \mathbb{R}^{D\times K}$ generated from CLIP's text encoder, we transform Gaussian centers $G$ of the SMGD from the vision space $\mathcal{V}$ to the multimodal space $\mathcal{M}$ as $\Psi$ : + +$$ +\Psi = \sigma \left(Z ^ {T} G\right) \in \mathbb {R} ^ {K \times K}, \tag {14} +$$ + +where $\sigma (\cdot)$ is a sigmoid function. We also project the test sample $f^{t}$ to the multimodal space $\mathcal{M}$ as + +$$ +\psi = \sigma \left(Z ^ {T} f ^ {t}\right) \in \mathbb {R} ^ {K}. \tag {15} +$$ + +We use Kullback-Leibler (KL) divergence to compare the similarities of the two distributions in the multimodal space as + +$$ +\Phi_ {k} = K L (\psi | | \Psi_ {k}), \tag {16} +$$ + +where $\Psi_{k}$ is the $k$ -th column vector of $\Psi$ , and $KL(P|Q) = \sum_{x\in \mathcal{X}}P(x)\log \left(\frac{P(x)}{Q(x)}\right)$ . Therefore, $\Phi = [\Phi_1,\Phi_2,\dots ,\Phi_K]\in \mathbb{R}^K$ consists of KL-divergence values. Since the most similar samples have low KL-divergence values (close to 0), they receive small weights. To address this, we negate the values in $\Phi$ . The prediction of the multimodal space $\mathcal{M}$ retrieval augmentation will be given as + +$$ +P _ {R} ^ {\mathcal {M}} = - L \Phi \in \mathbb {R} ^ {K}. \tag {17} +$$ + +Then, we can obtain the output of the multimodal retrieval augmentation (MRA) module $P_{\mathrm{R}}$ as a combination of the + +
MethodAircraftCaltech101CarsDTDEuroSATFlower102Food101PetsSUN397UCF101Average
CLIP-ViT-B/1623.2293.5566.1145.0450.4266.9982.8686.9265.6365.1664.59
CoOp18.4793.7064.5141.9246.3968.7185.3089.1464.1566.5563.88
CoCoOp22.2993.7964.9045.4539.2370.8583.9790.4666.8968.4464.63
TPT24.7894.1666.8747.7542.4468.9884.6787.7965.5068.0465.10
DiffTPT25.6092.4967.0147.0043.1370.1087.2388.2265.7462.6765.47
MTA25.3294.1368.0545.5938.7168.2684.9588.2264.9868.1164.63
DN24.3093.6064.0045.7053.3068.0086.0087.7066.5068.4065.75
ZERO25.2193.6668.0446.1234.3367.6886.5387.7565.0367.7767.72
DMN24.8494.1265.6444.3947.7771.3884.4889.0766.2866.7565.47
TDA23.9194.2467.2847.4058.0071.4286.1488.6367.6270.6667.53
TT-RAA (ours)25.3894.0866.4247.9966.1272.6886.0989.8367.6971.2968.76
+ +Table 1. Results on the cross-domain benchmark. Our TT-RAA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, four training-based adaptation methods (CoOp, CoCoOp, TPT, and DiffTPT), and five training-free adaptation methods (MTA, DN, DMN, ZERO, and TDA). + +vision space retrieval argumentation result $P_{\mathrm{R}}^{\mathcal{V}}$ and multimodal space augmentation result $P_{\mathrm{R}}^{\mathcal{M}}$ as + +$$ +P _ {\mathrm {R}} = P _ {\mathrm {R}} ^ {\mathcal {V}} + P _ {\mathrm {R}} ^ {\mathcal {M}}. \tag {18} +$$ + +After that, we compute our final TT-RAA logits by combining the original CLIP logits $P_{\mathrm{CLIP}}$ with the MRA prediction $P_{\mathrm{R}}$ using a weighting factor $\alpha$ as + +$$ +P _ {\mathrm {T T} - \mathrm {R A A}} = P _ {\mathrm {C L I P}} + \alpha P _ {\mathrm {R}}. \tag {19} +$$ + +In the end, we can obtain the final prediction of TT-RAA as + +$$ +\hat {y} = \underset {k} {\operatorname {a r g m a x}} P _ {\mathrm {T T} - \mathrm {R A A}} (k), \tag {20} +$$ + +where $k\in \{1,2,\dots,K\}$ + +# 4. Experiments + +# 4.1. Dataset + +To validate the method's test time adaptation capability across different domains and variations within a familiar domain, we conducted experiments on two benchmarks: a cross-domain (CD) benchmark comprising 10 datasets and an out-of-distribution (OOD) benchmark comprising 4 datasets. The CD benchmark primarily assesses transferability by testing on domains that diverge significantly from the original training distribution. Specifically, it includes Aircraft [26], Caltech101 [9], Cars [18], DTD [6], EuroSAT [13], Flower102 [29], Food101 [3], Pets [32], SUN397 [48], and UCF101 [40]. In contrast, the OOD benchmark focuses on robustness to distribution shifts within the same general domain. Specifically, it includes ImageNet-A [15], ImageNet-V2 [36], ImageNet-R [14], and ImageNet-S [47]. + +# 4.2. Implementation Details + +We implement our method based on the pre-trained CLIP architecture [34], which comprises a vision transformer + +(ViT) based image encoder [7] and a transformer-based text encoder [44]. Our experimental setup follows the challenging single-image test-time adaptation scenario, processing samples sequentially with a batch size of 1. To maintain computational efficiency and eliminate the need for backpropagation during inference, we utilize the hand-crafted prompts as proposed in [34] rather than learnable prompts. The experiments are conducted using the PyTorch framework, with evaluations performed on a single NVIDIA RTX A5000 GPU. Following standard practice in similar tasks, we employ top-1 accuracy as our evaluation metric. + +# 4.3. Comparisons with State-of-the-art Methods + +We conduct extensive experiments to evaluate TT-RAA against state-of-the-art approaches on the CD benchmark in Table 1 and OOD benchmark in Table 2. The competing methods include the CLIP baseline [34], training-based adaptation methods, and training-free adaptation methods. Discussions on Competing Methods. In the field of efficient domain adaptation, several methods require training, either in the training stage or at the test time, to align the model more closely with the target distribution. Context optimization (CoOp) [57], conditional context optimization (CoCoOp) [56] both involve learning prompts during a training phase. CoOp focuses on optimizing prompt tokens that are fixed for each domain, while CoCoOp extends this to generate a conditional prompt token for each image. Test-time prompt tuning (TPT) [38] and diffusion model for TPT (DiffTPT) [10], on the other hand, adapt the model at test time by tuning prompt embeddings. TPT performs direct gradient-based prompt tuning using test data, while DiffTPT uses new data generated by diffusion models. + +In addition to training-based methods, we also compare with several training-free adaptation methods, including two distribution-based methods, one entropy-based method, and two cache-based methods. Distribution-based methods evaluated are MeanShift for Test-time Augmentation (MTA) [51] and Distribution Normalization (DN) [58]. + +
MethodImageNet-AImageNet-V2ImageNet-RImageNet-SAverage
CLIP-ViT-B/1649.8961.8877.6548.2459.42
CoOp49.7164.2075.2147.9959.28
CoCoOp50.6364.0776.1848.7559.91
TPT54.7763.4577.0647.9460.81
DiffTPT55.6865.1075.0046.8060.52
MTA57.4163.6176.9248.5861.63
DN58.7162.8980.2048.9462.69
ZERO59.6164.1677.2248.4062.35
DMN58.2865.1778.5553.2063.80
TDA60.1164.6780.2450.5463.89
TT-RAA (Ours)60.5964.6980.5849.9863.96
+ +Table 2. Results on the out-of-distribution benchmark. Our TT-RAA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, four training-based adaptation methods (CoOp, CoCoOp, TPT, and DiffTPT), and five training-free adaptation methods (MTA, ZERO, DN, DMN, and TDA). + +They improve the test data distribution estimation via MeanShift algorithm and distribution normalization, respectively. Cache-based methods include the Dual Memory Network (DMN) [55] and TDA [17]. DMN comprises a dynamic and a static cache, while TDA employs a positive and a negative cache. For a fair comparison, we reproduced DMN's zero-shot results using our standardized prompts instead of DMN's original LLM-generated prompts, and we reduced the batch size to 1. We also compare against ZERO [8], an entropy-based method. ZERO sets the temperature of most confident predictions as zero to approximate marginal entropy minimization. + +Results on the Cross-domain Benchmark. Despite having the advantage of task-specific training, training-based adaptation methods, such as CoOp, CoCoOp, TPT, and DiffTPT, show limited generalization capabilities. TT-RAA outperforms these methods by significant margins (about $3\% - 4\%$ ). On fine-grained recognition tasks such as Flowers102, TT-RAA achieves $72.68\%$ accuracy, surpassing CoCoOp $(70.85\%)$ by $1.83\%$ . This gap widens further on specialized datasets like SUN397, where TT-RAA $(67.69\%)$ demonstrates superior feature adaptation compared to CoCoOp $(64.15\%)$ . These results suggest that static prompt learning, even with input-conditional mechanisms, may not be sufficient enough for handling significant domain shifts. + +In comparison to existing training-free adaptation approaches, TT-RAA demonstrates consistent superiority across various domain shifts. On remote sensing data (EuroSAT), TT-RAA $(66.12\%)$ significantly outperforms TDA $(58.00\%)$ by $8.12\%$ . This substantial gain can be attributed to our SMGD effectively capturing the unique characteristics of satellite imagery. Similarly, for action recognition (UCF101), TT-RAA $(71.29\%)$ surpasses TDA $(70.66\%)$ by $0.63\%$ , demonstrating robust feature adaptation. The performance advantages extend to other specialized domains as well. This consistent improvement across diverse domains validates the effectiveness of our method in bridging + +domain gaps in test-time settings. + +Results on the Out-of-distribution Benchmark. The experimental results in Table 2 demonstrate that our proposed TT-RAA method achieves state-of-the-art performance on the OOD benchmark with an average accuracy of $63.96\%$ across all ImageNet variants, outperforming both training-based adaptation methods (CoOp, CoCoOp, TPT, DiffTPT) and training-free approaches (MTA, DN, ZERO, DMN, TDA). Notably, TT-RAA exhibits good performance on ImageNet-R $(80.58\%)$ and ImageNet-A $(60.59\%)$ , highlighting its robust generalization capabilities under diverse distribution shifts. When examining performance across different distribution shifts, we observe that most methods achieve their highest accuracy on ImageNet-R and ImageNet-V2. ImageNet-S with sketches poses the greatest challenge for all methods, indicating that semantic variations remain difficult to address even with advanced adaptation techniques. + +# 4.4. Ablation Studies + +We conduct ablation studies on the CD benchmark to systematically evaluate the effectiveness of each proposed component in TT-RAA. Figure 4 presents the quantitative results of our analysis with both ResNet-50 and ViT-B/16 backbones. We establish our baseline using the TDA method with one-shot capacity, which achieves $59.82\%$ (ResNet-50) and $67.04\%$ (ViT-B/16) accuracy. + +Streaming Mixture of Gaussian Database Module. When integrating the SMGD module with the baseline, we observe a performance improvement of about $1.4\%$ with ViT-B/16 and $1.2\%$ with ResNet-50. This enhancement demonstrates that our SMGD effectively captures the dynamic distribution of test samples while maintaining computational efficiency. Its adaptive nature in updating class statistics online is useful for handling distribution shifts. + +Multimodal Retrieval Augmentation Module. Incorporating the MRA module also improved the performance by + +![](images/b56bbdf756d4d985f5e316e87d7d3fff514621606d1f8f38adb12e338164e796.jpg) +Figure 4. Ablation studies of TT-RAA. The SMGD and MRA modules independently improve performance over the baseline. Their combination yields further performance gains, demonstrating complementary benefits. + +about $0.4\%$ and $0.2\%$ with ViT-B/16 and ResNet-50, respectively. This gain supports our hypothesis regarding vision space misalignment in CLIP, which stems from its vision-text joint optimization objective. The performance improvement demonstrates that projecting both the SMGD database and test samples from vision space to multimodal space effectively addresses this misalignment. + +Complementary Benefits of Proposed Modules. When combining both SMGD and MRA modules (denoted as TT-RAA), we achieve the best performance of $68.76\%$ with ViT-B/16 and $61.25\%$ with ResNet-50, representing a substantial improvement of $1.72\%$ and $1.43\%$ over the baseline. These gains suggest that SMGD and MRA offer complementary and synergistic benefits. + +Analysis of Feature Backbone Impact. The effectiveness of our approach is further validated across two vision ecoders: ViT-B/16 and ResNet-50. These consistent improvements across different backbones suggest that our adaptation strategy effectively addresses fundamental domain shift challenges rather than exploiting architecture-specific characteristics. + +# 4.5. Hyperparameter Analysis + +The hyperparameter analysis of the update coefficient $\eta$ on the UCF101 is shown in Fig. 5a. When $\eta = 0$ , the model maintains a static Gaussian center initialized at the start, resulting in poor adaptation. From 0 to 0.55, accuracy increases sharply, highlighting the importance of incorporating new information. However, the performance deterioration beyond 0.65 reveals a critical threshold where excessive emphasis on new samples compromises the model's stability. This observation aligns with theoretical expectations, as larger $\eta$ values reduce the influence of historical information, potentially leading to over-adaptation to new samples. The empirical optimal value of $\eta = 0.65$ suggests that maintaining approximately $35\%$ of historical information while incorporating $65\%$ of new information achieves the best trade-off. + +![](images/97b338899f73e273bce87e9f8c68fbabc888cebfba23eb72dbaa4d7fb3760afa.jpg) +(a) + +![](images/b695c877f870a06d95745d4684e2c0d472f39cd20e7a595ee0db7bf07168f3ae.jpg) +(b) +Figure 5. (a) Hyperparameter analysis of the update coefficient $\eta$ . It shows that the best coefficient is 0.65. (b) Experiment with additional data. The method can accept additional data, which improve the performance significantly. + +# 4.6. Experiments with Access to Additional Data + +In the standard test-time adaptation setting, access to the target domain's training set is typically restricted. However, in many cases, the training set data might be available. Finetuning large-scale vision-language models in such scenarios remains computationally expensive. This raises an important question: how can performance be improved using the training data without computationally expensive finetuning? With a slight modification to TT-RAA, we propose leveraging the target domain training set (if accessible prior to test time) to construct a database and calculate class centers using a Gaussian mixture model. By doing so, we perform the same MRA process as in TT-RAA. This modified approach, referred to as TT-RAA-A, demonstrates significant performance improvements while maintaining a low computational requirement. The corresponding results are presented in Fig. 5b, where TT-RAA-A outperforms TT-RAA by a large margin, highlighting the benefits of incorporating target training data in a training-free manner, which open a new direction on adaptation without training. + +# 5. Conclusion + +In this paper, we introduced TT-RAA, a novel test-time retrieval augmented adaptation method that effectively addresses the computationally expensive domain adaptation challenges faced by vision-language models. Experimental results validated the effectiveness of SMGD by continuously estimating the test distribution and compressing all useful information into a single representation for each class without discarding old samples. By transforming Gaussian centers from vision space into a shared multimodal space, MRA enables more effective retrieval that leverages both visual and textual features. The success of TT-RAA opens up new possibilities for deploying vision-language models in real-world applications where computational resources are limited and rapid adaptation to new domains is crucial. Further exploration on adapting our method to larger multimodal foundation and large language models is promising. + +# References + +[1] Meta AI. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. Meta AI Report, 2024. 1, 2 +[2] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 5 +[3] Lukas Bossard, Matthieu Guillaumin, et al. Food-101: Mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 6 +[4] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. VisualGPT: Data-efficient adaptation of pretrained language models for image captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18030-18040, 2022. 2 +[5] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. European Conference on Computer Vision, pages 104-120, 2020. 2 +[6] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2014. 6 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6 +[8] Matteo Farina, Gianni Franchi, Giovanni Iacca, Massimiliano Mancini, and Elisa Ricci. Frustratingly easy test-time adaptation of vision-language models. Advances in Neural Information Processing Systems, 2024. 7 +[9] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2004. 6 +[10] Chun-Mei Feng, Kai Yu, Yong Liu, et al. Diverse data augmentation with diffusions for effective test-time prompt tuning. In IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 2, 3, 6 +[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, et al. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. 1 +[12] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, et al. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2024. 3 +[13] Patrick Helber, Benjamin Bischke, et al. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 6 +[14] Dan Hendrycks, Steven Basart, Norman Mu, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6 +[15] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6 + +[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. International Conference on Machine Learning, pages 4904-4916, 2021. 1, 2 +[17] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 7 +[18] Jonathan Krause, Michael Stark, et al. 3D object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshops, 2013. 6 +[19] Youngjun Lee, Doyoung Kim, Junhyeok Kang, Jihwan Bang, Hwanjun Song, et al. RA-TTA: Retrieval-augmented test-time adaptation for vision-language models. In International Conference on Learning Representations, 2025. 3 +[20] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, and Vladimir others Karpukhin. Retrievalaugmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems, pages 9459-9474, 2020. 3 +[21] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2 +[22] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 1, 2 +[23] Jian Liang, Ran He, et al. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, pages 1-34, 2024. 1, 2 +[24] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 2, 5 +[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 1, 2 +[26] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, et al. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6 +[27] Yifei Ming and Yixuan Li. Understanding retrieval-augmented task adaptation for vision-language models. In International Conference on Machine Learning, pages 35719-35743, 2024. 3 +[28] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, et al. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 2, 3 +[29] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 6 +[30] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable + +test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 2, 3 +[31] OpenAI. Gpt-4v(ison) system card. In OpenAI Technical Report, 2023. 1, 2 +[32] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 6 +[33] Rizwan Qureshi, Ranjan Sapkota, Abbas Shah, et al. Thinking beyond tokens: From brain-inspired intelligence to cognitive foundations for artificial general intelligence and its societal impact. arXiv preprint arXiv:2507.00951, 2025. 1 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 1, 2, 3, 6 +[35] Aditi Raghunathan, Prateek Jain, et al. Learning mixture of gaussians with streaming data. Advances in Neural Information Processing Systems, 30, 2017. 4 +[36] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, et al. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, 2019. 6 +[37] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, et al. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 2020. 2, 3 +[38] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 2022. 2, 3, 6 +[39] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2 +[40] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 6 +[41] Tao Sun, Cheng Lu, and Haibin Ling. Prior knowledge guided unsupervised domain adaptation. In European Conference on Computer Vision. Springer, 2022. 1, 2 +[42] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2, 3 +[43] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. SuS-X: Training-free name-only transfer of vision-language models. In IEEE/CVF International Conference on Computer Vision, pages 2725-2736, 2023. 5 +[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6 +[45] Matthew Wallingford, Vivek Ramanujan, Alex Fang, Aditya Kusupati, Roozbeh Mottaghi, Aniruddha Kembhavi, et al. + +Neural priming for sample-efficient adaptation. Advances in Neural Information Processing Systems, 2023. 3 +[46] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 2, 3 +[47] Haohan Wang, Songwei Ge, et al. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, 2019. 6 +[48] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In IEEE Conference on Computer Vision and Pattern Recognition, 2010. 6 +[49] Shiqi Yang, Joost Van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. Advances in Neural Information Processing Systems, 34:29393-29405, 2021. 1, 2 +[50] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. International Conference on Learning Representations, 2022. 1, 2 +[51] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23783-23793, 2024. 2, 6 +[52] Jingyi Zhang, Jiaxing Huang, et al. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[53] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. VinVL: Revisiting visual representations in vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 2 +[54] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision, pages 493-510. Springer, 2022. 5 +[55] Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28718-28728, 2024. 2, 7 +[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 6 +[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 6 +[58] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser Nam Lim. Test-time distribution normalization for contrastively learned visual-language models. Advances in Neural Information Processing Systems, pages 47105-47123, 2023. 2, 6 \ No newline at end of file diff --git a/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/images.zip b/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3a89ad61df4d0a92d9dd6248020daf7d45a90452 --- /dev/null +++ b/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9538ffbde1e81697bc96e5e8ba4c378575dce273dbcc59c9611a5f91d9eb266d +size 392389 diff --git a/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/layout.json b/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8aae76168bf712a4a85a7588f350e7beab102e27 --- /dev/null +++ b/2025/Test-Time Retrieval-Augmented Adaptation for Vision-Language Models/layout.json @@ -0,0 +1,9648 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 103, + 531, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 103, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 80, + 103, + 531, + 121 + ], + "type": "text", + "content": "Test-Time Retrieval-Augmented Adaptation for Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "spans": [ + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "text", + "content": "Xinqi Fan" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "text", + "content": ", Xueli Chen" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "text", + "content": ", Luoxiao Yang" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "text", + "content": ", Chuin Hong Yap" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "text", + "content": ", Rizwan Qureshi" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "text", + "content": ", Qi Dou" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "text", + "content": ", Moi Hoon Yap" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "text", + "content": ", Mubarak Shah" + }, + { + "bbox": [ + 113, + 142, + 496, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 171, + 544, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 171, + 544, + 200 + ], + "spans": [ + { + "bbox": [ + 66, + 171, + 544, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 171, + 544, + 200 + ], + "type": "text", + "content": "Manchester Metropolitan University " + }, + { + "bbox": [ + 66, + 171, + 544, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 171, + 544, + 200 + ], + "type": "text", + "content": "Hong Kong Metropolitan University " + }, + { + "bbox": [ + 66, + 171, + 544, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 171, + 544, + 200 + ], + "type": "text", + "content": "Xi'an University of Technology " + }, + { + "bbox": [ + 66, + 171, + 544, + 200 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 66, + 171, + 544, + 200 + ], + "type": "text", + "content": "University of Central Florida " + }, + { + "bbox": [ + 66, + 171, + 544, + 200 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 66, + 171, + 544, + 200 + ], + "type": "text", + "content": "Chinese University of Hong Kong" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 251, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 624 + ], + "type": "text", + "content": "Vision-language models (VLMs) have shown promise in test-time adaptation tasks due to their remarkable capabilities in understanding and reasoning about visual content through natural language descriptions. However, training VLMs typically demands substantial computational resources, and they often struggle to adapt efficiently to new domains or tasks. Additionally, dynamically estimating the test distribution from streaming data at test time remains a significant challenge. In this work, we propose a novel test-time retrieval-augmented adaptation (TT-RAA) method that enables VLMs to maintain high performance across diverse visual recognition tasks without the need for task-specific training or large computational overhead. During inference, TT-RAA employs a streaming mixture of Gaussian database (SMGD) to continuously estimate test distributions, requiring minimal storage. Then, TT-RAA retrieves the most relevant information from the SMGD, enhancing the original VLM outputs. A key limitation of CLIP-based VLMs is their inter-modal vision-language optimization, which does not optimize vision-space similarity, leading to larger intra-modal variance. To address this, we propose a multimodal retrieval augmentation module that transforms the SMGD into a unified multimodal space, enabling retrieval that aligns both vision and language modalities. Extensive experiments across both cross-domain and out-of-distribution benchmarks comprising fourteen datasets demonstrate TT-RAA's superior performance compared to state-of-the-art methods. Ablation studies and hyperparameter analyses further validate the effectiveness of the proposed modules. The source code of our work is available at https://github.com/xinqi-fan/TT-RAA." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 633, + 135, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 633, + 135, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 633, + 135, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Vision-language models (VLMs) have significantly advanced visual understanding by leveraging natural language supervision with contrastive language-image pre-training (CLIP) notably pioneering this field [34]. The zero-shot capability enables this model to be deployed in various do" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 229, + 548, + 360 + ], + "blocks": [ + { + "bbox": [ + 322, + 229, + 548, + 360 + ], + "lines": [ + { + "bbox": [ + 322, + 229, + 548, + 360 + ], + "spans": [ + { + "bbox": [ + 322, + 229, + 548, + 360 + ], + "type": "image", + "image_path": "3deb82f03189eccad0fff94c65752ec7775e644dc2e4c0a9a72c340c33e33994.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 370, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 426 + ], + "type": "text", + "content": "Figure 1. Illustration of our proposed test time retrieval augmented adaptation (TT-RAA). Without TT-RAA, the CLIP prediction will be the final prediction. With TT-RAA, the image feature will be used to retrieve the most similar database centers which will be used to improve the final prediction in a training-free manner." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 436, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 640 + ], + "type": "text", + "content": "mains. This foundational work has catalyzed a wave of architectural innovations, including domain-specific variants and architectural enhancements that build upon CLIP's core principles [52]. Some variants have focused on scaling up capacities, learning from noisy text supervision, and fine-grained alignment [16, 50]. However, these models consistently encounter challenges when deployed in specialized domains, where the distribution of images often differs substantially from their web-scale training data [11, 33]. This challenge has become even more prominent with the recent introduction of more sophisticated large VLMs such as GPT-4V [31], LLaVA [25], and Llama 3.2 Vision [1]. While these advanced models excel at integrating visual and textual data, the high computational cost of fine-tuning them for domain-specific tasks poses a major challenge, especially in specialized visual domains that differ significantly from general web images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Test-time adaptation (TTA) has emerged as a promising solution that enables models to adapt to new domains without accessing source domain training data [23]. Traditional TTA approaches include self-training-based methods that leverage pseudo-labels [22, 41, 49], self-supervised learning-based methods that incorporate aux" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8810" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "iliary tasks [42], batch normalization [28, 37], and entropy minimization-based methods [30, 46]. For VLMs, recent works have explored prompt-based adaptation methods [10, 38] to bridge the domain gap between web-scale training data and specific downstream tasks. However, these methods often incur substantial computational overhead, highlighting the need for more efficient approaches [10, 17]. Advancements in training-free methods have explored the use of cache mechanisms [17, 55] and distribution-based methods [51, 58] to improve the efficiency. Despite these advances, dynamically estimating the test distribution from streaming data at test time remains a significant challenge, due to the test sample arriving one at a time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 231, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 517 + ], + "type": "text", + "content": "In this work, we introduce test-time retrieval-augmented adaptation (TT-RAA), a novel training-free method that allows VLMs to sustain strong performance across various visual recognition tasks without requiring task-specific training or high computational costs (Fig. 1). TT-RAA is a training-free framework, which uses a streaming mixture of Gaussian database (SMGD) to dynamically capture test distribution characteristics with minimal storage requirements at inference time. TT-RAA retrieves relevant information from the SMGD to enhance the performance of the original VLM. Furthermore, an inherent limitation of CLIP-based VLMs is their lack of optimization for intramodal similarity, which does not optimize similarity in the vision space, resulting in greater intramodal variance [24]. To mitigate this, we introduce a multimodal retrieval augmentation (MRA) module that projects the SMGD into a shared multimodal embedding space, facilitating retrieval that harmonizes vision and language modalities. Extensive testing across ten cross-domain datasets and four out-of-distribution datasets confirms that TT-RAA delivers superior performance relative to state-of-the-art methods, with additional ablation studies and hyperparameter analyses supporting the effectiveness of the proposed modules. The key contributions of this work include:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 520, + 295, + 651 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 520, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 567 + ], + "type": "text", + "content": "- We introduce test-time retrieval-augmented adaptation (TT-RAA), an efficient, training-free framework that significantly improves domain adaptability without additional fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 568, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 602 + ], + "type": "text", + "content": "- We propose streaming mixture of Gaussian database (SMGD) for dynamic, efficient, and robust estimation of test domain statistics during inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 604, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 295, + 651 + ], + "type": "text", + "content": "- We propose multimodal retrieval augmentation (MRA), a novel mechanism addressing CLIP's inherent intra-modal similarity limitation, aligning both vision and language modalities effectively." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 653, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 712 + ], + "type": "text", + "content": "The rest of this paper is organized as follows: Section 2 provides some related works, Section 3 describes the proposed method. Experimental results and performance evaluation are presented in Section 4. Finally, Section 5 concludes the article and outlines future research directions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 91, + 448, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 448, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 448, + 104 + ], + "type": "text", + "content": "2.1. Vision Language Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 110, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 566 + ], + "type": "text", + "content": "The development of vision language models (VLMs) can be traced back to earlier works that explored vision-language pre-training for specific tasks, where VisualBERT [21] and UNITER [5] pioneered joint representation learning between visual and textual modalities for tasks like visual question answering. The field witnessed a paradigm shift with CLIP [34], which demonstrated impressive zero-shot capabilities by learning transferable visual representations through contrastive learning between images and texts at web scale. ALIGN [16] further scaled up this approach with 1.8 billion noisy image-text pairs and introduced noise-robust training strategies. FILIP [50] enhanced CLIP by introducing fine-grained region-word alignment, enabling more detailed vision-language correspondences. FLAVA [39] introduced a single foundation model that can tackle vision, language, and multimodal tasks. More recently, the emergence of multimodal large language models has pushed the boundaries of vision-language understanding, where GPT-4V [31] extends GPT-4's language capabilities to handle visual inputs, LLaVA [25] achieves strong visual conversation abilities by aligning vision encoders with language models through minimal task-specific training, and Llama 3.2 [1] demonstrates remarkable vision and language capabilities with open models. Other notable VLMs include VinVL [53], which enhanced visual understanding by incorporating object detection and segmentation, and VisualGPT [4], which demonstrated zero-shot transfer to various vision-language tasks. These models have continued to push the boundaries of multimodal representation learning, enabling a wide range of applications from intelligent assistants to creative tools. While these newer models demonstrate impressive capabilities in combining visual and textual understanding, they face similar domain adaptation challenges but at an even larger scale. The computational demands for fine-tuning these models on domain-specific tasks present a significant barrier, particularly for specialized applications where the visual domain differs markedly from general web imagery." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 575, + 440, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 575, + 440, + 588 + ], + "spans": [ + { + "bbox": [ + 314, + 575, + 440, + 588 + ], + "type": "text", + "content": "2.2. Test Time Adaptation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Normal domain adaptation techniques require access to both source and target domain data, but test-time adaptation (TTA) adapts models to the target domain without source data, making it particularly useful for scenarios with privacy or legal restrictions [23]. To adapt a pre-trained model to a target domain, methods like self-training refine model predictions using pseudo-labels. Approaches such as centroid-based [22], neighbor-based [49], and optimization-based [41] strategies enhance the quality of these pseudolabels. Self-supervised methods introduce auxiliary tasks to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8811" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "align features, exemplified by test-time training (TTT) [42] with a self-supervised rotation prediction task. Batch normalization [28, 37] techniques and entropy minimization strategies [30, 46] have also been adopted in TTA tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 276 + ], + "type": "text", + "content": "Recent advancements focus on VLMs like CLIP, using domain-specific prompts from test data. Test-time prompt tuning (TPT) [38] fine-tunes a learnable prompt with each testing sample, while DiffTPT [10] uses pre-trained diffusion models to enhance test data diversity. Although these methods are promising, they still require a large amount of computational resources for their training-based adaptation, making it hard to adapt large VLMs. To reduce costs, a training-free dynamic adapter (TDA) [17] has been introduced, using a lightweight dynamic key-value cache without backpropagation. However, its cache only captures the streaming test data with a limited cache capacity without the ability to estimate the full spectrum of the test data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 283, + 233, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 233, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 233, + 295 + ], + "type": "text", + "content": "2.3. Retrieval Augmented Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 574 + ], + "type": "text", + "content": "Retrieval-augmented generation (RAG) has emerged as a promising solution to enhance large language models (LLMs) by incorporating external knowledge during inference, effectively addressing challenges like hallucination and outdated knowledge [20]. In its naive form, RAG follows a three-stage process: indexing, where documents are chunked and encoded into vectors; retrieval, where relevant documents are retrieved based on semantic similarity with the input query; and generation, where retrieved documents along with the query are fed to LLMs for response generation [12]. Recent efforts have extended RAG to VLMs, including methods such as I2I/T2I adaptors [27], RA-TTA [19], and neural priming [45]. These methods rely on additional databases constructed from large-scale datasets like LAION-2B/5B, which often incur high computational and storage costs during both construction and inference. In contrast, our work focuses on constructing a database dynamically from incoming test samples, without relying on additional large-scale databases. In addition, we employed a multimodal retrieval augmentation module that transforms our database into a unified multimodal space, enabling retrieval that aligns both vision and language modalities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 586, + 111, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 111, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 111, + 597 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "In this work, we introduce test-time retrieval-augmented Adaptation (TT-RAA) to adapt vision-language models (VLMs) to new domains (Fig. 2). TT-RAA leverages a streaming mixture of Gaussian database (SMGD) to model test data statistics in real time. As test data streams in, we dynamically update the SMGD with Gaussian centers representing the distribution of vision features. Each test sample's vision features query the SMGD to retrieve relevant Gaussian centers, providing contextual information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "Since CLIP's joint vision-text optimization does not explicitly optimize vision-space similarity, we transform the Gaussian centers into a multimodal space using text embeddings. This allows retrieval in a space that is aligned with both vision and language modalities. Final predictions combine CLIP outputs with retrieval results from both vision and multimodal spaces, enhancing accuracy and robustness. TT-RAA thus provides a powerful framework for adapting VLMs to new domains by leveraging both vision and language information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 198, + 531, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 531, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 531, + 212 + ], + "type": "text", + "content": "3.1. Contrastive Language-Image Pretraining" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "text", + "content": "Contrastive language-image pre-training (CLIP) [34] is a vision-language model that has two separate encoders for processing images and text, respectively. These encoders are trained to project their respective inputs into a shared embedding space where related images and texts are positioned close to each other. Consider a " + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "text", + "content": "-class classification problem where the CLIP's objective in the zero-shot setting is to match images with their most relevant textual descriptions using an image encoder " + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "inline_equation", + "content": "E_I" + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "inline_equation", + "content": "E_T" + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "text", + "content": ". To obtain the textual descriptions " + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "text", + "content": "-class names are concatenated with hand-crafted prompts and then mapped into the " + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "text", + "content": "-dimensional text embeddings " + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "text", + "content": " using the text encoder " + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "inline_equation", + "content": "E_T" + }, + { + "bbox": [ + 313, + 216, + 553, + 370 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 386, + 370, + 553, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 370, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 386, + 370, + 553, + 384 + ], + "type": "interline_equation", + "content": "Z = E _ {T} (T) \\in \\mathbb {R} ^ {D \\times K}. \\tag {1}", + "image_path": "ab077b0597bc586f697193e26856080891973a0a8aeb44cad159a6beb184dc15.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 388, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 553, + 411 + ], + "type": "text", + "content": "Given the " + }, + { + "bbox": [ + 313, + 388, + 553, + 411 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 388, + 553, + 411 + ], + "type": "text", + "content": "-th test samples " + }, + { + "bbox": [ + 313, + 388, + 553, + 411 + ], + "type": "inline_equation", + "content": "x^t" + }, + { + "bbox": [ + 313, + 388, + 553, + 411 + ], + "type": "text", + "content": ", CLIP vision encoder produces image features" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 392, + 418, + 553, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 418, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 392, + 418, + 553, + 433 + ], + "type": "interline_equation", + "content": "f ^ {t} = E _ {I} \\left(x ^ {t}\\right) \\in \\mathbb {R} ^ {D}. \\tag {2}", + "image_path": "4f81329d786ae6f595e675c8b3a5f2adb9f0cdcc25feee6080c68b6e3e0baa1d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 439, + 553, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 553, + 485 + ], + "type": "text", + "content": "CLIP was trained using contrastive loss to minimize the distance between corresponding text and image embeddings. When making predictions, the CLIP output can be obtained as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 386, + 486, + 553, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 486, + 553, + 500 + ], + "spans": [ + { + "bbox": [ + 386, + 486, + 553, + 500 + ], + "type": "interline_equation", + "content": "P _ {\\mathrm {C L I P}} ^ {t} = Z ^ {T} f ^ {t} \\in \\mathbb {R} ^ {K}. \\tag {3}", + "image_path": "89d9e8597a9f22b4afd0995990f554c99c6ff7aefcd4f89d0b69d2e5cfe62fa1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 504, + 553, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 553, + 534 + ], + "type": "text", + "content": "Then, we can obtain the final prediction for test sample " + }, + { + "bbox": [ + 313, + 504, + 553, + 534 + ], + "type": "inline_equation", + "content": "x^{t}" + }, + { + "bbox": [ + 313, + 504, + 553, + 534 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 504, + 553, + 534 + ], + "type": "inline_equation", + "content": "\\hat{y}^t = \\operatorname*{argmax}_k P_{\\mathrm{CLIP}}^t (k)" + }, + { + "bbox": [ + 313, + 504, + 553, + 534 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 504, + 553, + 534 + ], + "type": "inline_equation", + "content": "k \\in \\{1,2,\\dots,K\\}" + }, + { + "bbox": [ + 313, + 504, + 553, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 540, + 531, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 531, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 531, + 553 + ], + "type": "text", + "content": "3.2. Streaming Mixture of Gaussian Database" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "Our work addresses a challenging training-free test-time adaptation problem where models must adapt to streaming test data under two constraints: (1) Both input and output distributions of the source and target domains may be subject to distribution shifts. (2) Test samples arrive sequentially, one at a time, rather than in batches, reflecting real-world deployment. This setting is demanding because it requires training-free, zero-shot adaptation without access to batch statistics. Therefore, it is important to estimate the statistics of the new domain by capturing its full spectrum, and based on this we can retrieve the most relevant information to enhance the final prediction. In this work, we propose a Streaming Mixture of Gaussian Database (SMGD)" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8812" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 555, + 250 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 555, + 250 + ], + "type": "image", + "image_path": "48eaa854a7bc8dfedcc7f1a059c0ccea08966a7c72bb9e3f052827cc67cf1ee8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 258, + 555, + 315 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 555, + 315 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed test-time retrieval-augmented adaptation (TT-RAA) method. TT-RAA leverages a streaming mixture of gaussian database (SMGD) to dynamically model the distribution of streaming test data in vision space " + }, + { + "bbox": [ + 55, + 258, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 258, + 555, + 315 + ], + "type": "text", + "content": ". The multimodal retrieval augmentation (MRA) module projects these Gaussian centers into a unified multimodal space " + }, + { + "bbox": [ + 55, + 258, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 258, + 555, + 315 + ], + "type": "text", + "content": ", enabling effective retrieval aligned across both vision and language modalities. Combining retrieval results from vision and multimodal spaces with CLIP's original predictions, TT-RAA significantly enhances VLM performance without additional training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 320, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 320, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 320, + 295, + 344 + ], + "type": "text", + "content": "module by using a Gaussian mixture model to estimate the dynamic statistics from the test time streaming data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 345, + 295, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 295, + 368 + ], + "type": "text", + "content": "We assume that feature embedding " + }, + { + "bbox": [ + 55, + 345, + 295, + 368 + ], + "type": "inline_equation", + "content": "f^t" + }, + { + "bbox": [ + 55, + 345, + 295, + 368 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 55, + 345, + 295, + 368 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 345, + 295, + 368 + ], + "type": "text", + "content": "-th sample " + }, + { + "bbox": [ + 55, + 345, + 295, + 368 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 345, + 295, + 368 + ], + "type": "text", + "content": " is drawn from a mixture of Gaussian distributions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 376, + 295, + 400 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 376, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 88, + 376, + 295, + 400 + ], + "type": "interline_equation", + "content": "f ^ {t} \\sim \\sum_ {k} w _ {k} \\mathcal {N} \\left(\\mu_ {k}, \\Sigma_ {k}\\right), \\forall k = 1, 2, \\dots , K, \\tag {4}", + "image_path": "8a7131515c1f0a85bd84a426cd9038c7377124086fd70e7b8efc071e7809b085.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\mu_k \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\Sigma_k \\in \\mathbb{R}^{D \\times D}" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": " are the mean and covariance of the " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": "-th Gaussian component " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": " respectively. The mixture weights " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "w_k \\geq 0" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\sum_k w_k = 1" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": " is the number of classes. For simplicity, we assume the weights " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "w_k = p(y = k) = 1 / K" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": ". Here, SMGD is a key-value database, where the keys are Gaussian mean vectors " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\{\\mu_k\\}_{k=1}^K" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": " and the values are the corresponding pseudo labels of these Gaussian mean vectors. The covariance matrices " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\{\\Sigma_k\\}_{k=1}^K" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": " are also recorded which will involve in the retrieval during test time adaptation. Given the test time setting, all samples come in a stream. Our goal is to estimate the optimal mean " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\mu_k" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": " and covariance " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "\\Sigma_k" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": " for the class " + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 408, + 295, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": "To estimate the optimal value of the mean " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mu_{k}" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": " and covariance " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma_{k}" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": ", we draw inspiration from streaming data statistics estimation [35]. Given a test sample " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "x^{t}" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": ", its pseudo label " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{l}_t" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": " is determined by its CLIP prediction as " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{l}_t = \\operatorname*{argmax}_{k} P_{\\mathrm{CLIP}}^t (k)" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": ". Initially, we set the mean vector " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mu_{k}" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": " to the first test sample with its pseudo label as " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": " from the streaming test data. We set the initial covariance " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma_{k} = \\sigma^{2}\\mathbf{I}" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\sigma = 0.1" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": ". We also record the corresponding entropy " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "h_k" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": " as the entropy of the first test sample's CLIP logit. Once the mean vector " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mu_{k}" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": " and covariance matrix " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\Sigma_{k}" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": " are initialized, we progressively update them using selected test samples with high-quality pseudo labels, as measured by entropy. If the test sample " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "x^{t}" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": " has a pseudo label " + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{l}^t = k" + }, + { + "bbox": [ + 55, + 553, + 295, + 713 + ], + "type": "text", + "content": " and an" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 320, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 346 + ], + "type": "text", + "content": "entropy smaller than " + }, + { + "bbox": [ + 313, + 320, + 555, + 346 + ], + "type": "inline_equation", + "content": "h_k^{t-1}" + }, + { + "bbox": [ + 313, + 320, + 555, + 346 + ], + "type": "text", + "content": ", the mean " + }, + { + "bbox": [ + 313, + 320, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mu_k^t" + }, + { + "bbox": [ + 313, + 320, + 555, + 346 + ], + "type": "text", + "content": " and the covariance " + }, + { + "bbox": [ + 313, + 320, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\Sigma_k^t" + }, + { + "bbox": [ + 313, + 320, + 555, + 346 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 320, + 555, + 346 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 320, + 555, + 346 + ], + "type": "text", + "content": " are updated as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 341, + 354, + 553, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 354, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 341, + 354, + 553, + 369 + ], + "type": "interline_equation", + "content": "\\mu_ {k} ^ {t} = (1 - \\eta) \\mu_ {k} ^ {t - 1} + \\eta f ^ {t}, \\tag {5}", + "image_path": "2e70575f38d926e71879df0eec6cc864db9714bc6c193c74b6f3096816d26c24.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 341, + 371, + 553, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 371, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 341, + 371, + 553, + 385 + ], + "type": "interline_equation", + "content": "\\Sigma_ {k} ^ {t} = (1 - \\eta) \\Sigma_ {k} ^ {t - 1} + \\eta \\left(f ^ {t} - \\mu_ {k} ^ {t}\\right) \\left(f ^ {t} - \\mu_ {k} ^ {t}\\right) ^ {T}, \\tag {6}", + "image_path": "bbcfc8ebaf4365c56e0269ca01bf0c75bc2864fc257b4a1cf520331055f91d04.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "type": "text", + "content": " is the update coefficient. Otherwise, " + }, + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\mu_k^t = \\mu_k^{t-1}" + }, + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\Sigma_k^t = \\Sigma_k^{t-1}" + }, + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "type": "inline_equation", + "content": "k \\in \\{1, 2, \\dots, K\\}" + }, + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "type": "text", + "content": ", meaning the mean vectors and the covariance matrices remain unchanged if there is no test sample with pseudo label " + }, + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 396, + 555, + 506 + ], + "type": "text", + "content": " or if the test sample does not have a high-quality pseudo label. A high-quality pseudo label is identified when its entropy is lower than that of the Gaussian mean vector. Once the mean and covariance are updated according to Eqs. (5) and (6), the corresponding entropy is updated accordingly as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 364, + 514, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 514, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 364, + 514, + 553, + 529 + ], + "type": "interline_equation", + "content": "h _ {k} ^ {t} = (1 - \\eta) h _ {k} ^ {t - 1} + \\eta H \\left(P _ {\\mathrm {C L I P}} ^ {t}\\right), \\tag {7}", + "image_path": "81be83e8684a5b81bc319254d10f1da0f2db4407607df7628bfaea02693694f3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 540, + 553, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 553, + 566 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 540, + 553, + 566 + ], + "type": "inline_equation", + "content": "\\mathrm{H}(P_{\\mathrm{CLIP}}^t) = -\\sum_{k=1}^{K} P_{\\mathrm{CLIP}}^t(k) \\log(P_{\\mathrm{CLIP}}^t(k))" + }, + { + "bbox": [ + 313, + 540, + 553, + 566 + ], + "type": "text", + "content": " is the entropy of the CLIP logits " + }, + { + "bbox": [ + 313, + 540, + 553, + 566 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{CLIP}}^t" + }, + { + "bbox": [ + 313, + 540, + 553, + 566 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 566, + 553, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 553, + 613 + ], + "type": "text", + "content": "In this way, we keep updating a high-quality database with high confidence as measured by the entropy of test-time streaming data's pseudo labels. The SMGD " + }, + { + "bbox": [ + 313, + 566, + 553, + 613 + ], + "type": "inline_equation", + "content": "S_{G}" + }, + { + "bbox": [ + 313, + 566, + 553, + 613 + ], + "type": "text", + "content": " is composed of the following:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 384, + 624, + 553, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 624, + 553, + 637 + ], + "spans": [ + { + "bbox": [ + 384, + 624, + 553, + 637 + ], + "type": "interline_equation", + "content": "S _ {G} = \\{G, L \\}, \\tag {8}", + "image_path": "4d666dcd878b1010fe0ef753cf5a680c27e2c733340ba7a4d148b417f35a1a5f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 390, + 640, + 553, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 640, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 390, + 640, + 553, + 654 + ], + "type": "interline_equation", + "content": "G = \\left[ \\mu_ {0} ^ {t}, \\mu_ {1} ^ {t}, \\dots , \\mu_ {K} ^ {t} \\right], \\tag {9}", + "image_path": "f1c0e42e686a94906d3f110d43d706ac9b804aa0c2a9c5326a4c7629dfc3723e.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "inline_equation", + "content": "G \\in \\mathbb{R}^{D \\times K}" + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "text", + "content": " represents the Gaussian centers of the SMGD, and each column of " + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "inline_equation", + "content": "L \\in \\mathbb{R}^{K \\times K}" + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "text", + "content": " denotes a " + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "text", + "content": "-dimensional one-hot pseudo-label vector. Along with SMGD, we have the covariances " + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\Sigma_0^t, \\Sigma_1^t, \\dots, \\Sigma_K^t\\}" + }, + { + "bbox": [ + 313, + 664, + 555, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8813" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 282, + 177 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 282, + 177 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 282, + 177 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 282, + 177 + ], + "type": "image", + "image_path": "121c712bc59063ad88e26279103aff13c2399aba643331d2b1c7b9a2486b98c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 180, + 296, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 247 + ], + "type": "text", + "content": "Figure 3. Motivation of multimodal retrieval augmentation. Cosine similarity distributions of matched and unmatched image-text pairs (inter-modal) exhibit less overlap than those of matched and unmatched image-image pairs (intra-modal). This suggests that samples are more easily distinguishable in the multimodal space than in the vision space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 249, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 249, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 249, + 264 + ], + "type": "text", + "content": "3.3. Multimodal Retrieval Augmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 268, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 483 + ], + "type": "text", + "content": "We can directly retrieve the most similar data center from the SMGD to improve the final prediction. However, the CLIP model was optimized to reduce the inter-modal (vision-text) similarities rather than the intra-modal (vision-vision or text-text) similarities [24]. This optimization process leaves a disadvantage that similar images in the vision (image) feature space are not well clustered [43]. As shown in Fig. 3, the cosine similarity distributions of matched and unmatched image-text pairs (inter-modal) exhibit less overlap than those of matched and unmatched image-image pairs (intra-modal). This indicates that matched and unmatched pairs are more easily distinguishable in the multimodal space than in the vision-only space, a distinction shaped by CLIP's image-text contrastive learning. Therefore, in addition to the vision space retrieval augmentation, we also added a multimodal space retrieval augmentation, which compares the similarities of the samples in the CLIP text-image space rather than the vision space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 519 + ], + "type": "text", + "content": "Vision Space Retrieval Augmentation. To fully leverage the available information from the vision space, we include two terms: similarity retrieval and discriminant analysis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 520, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 590 + ], + "type": "text", + "content": "Similarity Retrieval. This term treats the test sample as a query and retrieves the most similar class center from the SMGD based on their similarities within the vision space. Given the SMGD " + }, + { + "bbox": [ + 55, + 520, + 296, + 590 + ], + "type": "inline_equation", + "content": "S_{G} = \\{G, L\\}" + }, + { + "bbox": [ + 55, + 520, + 296, + 590 + ], + "type": "text", + "content": " and test image features " + }, + { + "bbox": [ + 55, + 520, + 296, + 590 + ], + "type": "inline_equation", + "content": "f^{t}" + }, + { + "bbox": [ + 55, + 520, + 296, + 590 + ], + "type": "text", + "content": " generated from the CLIP's image encoder, the prediction can be calculated as follows" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 110, + 597, + 295, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 597, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 110, + 597, + 295, + 612 + ], + "type": "interline_equation", + "content": "P _ {s i m} \\left(f ^ {t}\\right) = L A \\left(G ^ {T} f ^ {t}\\right) \\in \\mathbb {R} ^ {K}, \\tag {10}", + "image_path": "803cad7445cdeb287fe8250247582f32075629299d2cba96a167f0b91750f345.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 617, + 295, + 641 + ], + "type": "inline_equation", + "content": "A(z) = \\exp (-\\beta (1 - z))" + }, + { + "bbox": [ + 55, + 617, + 295, + 641 + ], + "type": "text", + "content": " is an adaptation function with a sharpness ratio " + }, + { + "bbox": [ + 55, + 617, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 617, + 295, + 641 + ], + "type": "text", + "content": " [54]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Discriminant Analysis. This term leverages probabilistic information derived from the Gaussian means and covariances of the SMGD by using a closed-form expression from Gaussian discriminant analysis [2]. Under the assumption that all Gaussian components share an identical covariance " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma^t" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": ", the discriminant score can be computed as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 331, + 70, + 555, + 117 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 70, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 555, + 117 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Omega_ {d i s c} (f ^ {t}) = G ^ {T} \\Sigma^ {t - 1} f ^ {t} - \\frac {1}{2} \\operatorname {d i a g} \\left(G ^ {T} \\Sigma^ {t - 1} G\\right) \\tag {11} \\\\ + \\log \\frac {1}{K} \\mathbf {1} _ {K} \\in \\mathbb {R} ^ {K}, \\\\ \\end{array}", + "image_path": "7dc8e1e1f2ee18351fb6ae5374c0e7005b2d3c86a839877774392c0ea56cad5e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "type": "inline_equation", + "content": "\\Sigma^t" + }, + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "type": "text", + "content": " is the mean of " + }, + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "type": "text", + "content": " covariances " + }, + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "type": "inline_equation", + "content": "\\{\\Sigma_k^t\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "type": "inline_equation", + "content": "\\Omega_{disc}(f^t)" + }, + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "type": "text", + "content": " contains scores that reflect how well " + }, + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "type": "inline_equation", + "content": "f^t" + }, + { + "bbox": [ + 313, + 125, + 555, + 186 + ], + "type": "text", + "content": " matches different classes, considering both the distance from the class mean and the spread of the class distribution. Therefore, the discriminant analysis-based term can be given as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 369, + 195, + 555, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 195, + 555, + 211 + ], + "spans": [ + { + "bbox": [ + 369, + 195, + 555, + 211 + ], + "type": "interline_equation", + "content": "P _ {d i s c} \\left(f ^ {t}\\right) = L \\Omega_ {d i s c} \\left(f ^ {t}\\right) \\in \\mathbb {R} ^ {K}. \\tag {12}", + "image_path": "bbfa0d1d4c76a62d231e0e13460fd2ded793cd925b59f128a93324b9a00769f6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 219, + 554, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 554, + 243 + ], + "type": "text", + "content": "After that, the prediction from the vision space " + }, + { + "bbox": [ + 313, + 219, + 554, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 219, + 554, + 243 + ], + "type": "text", + "content": " retrieval augmentation can then be calculated as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 251, + 555, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 251, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 351, + 251, + 555, + 266 + ], + "type": "interline_equation", + "content": "P _ {\\mathrm {R}} ^ {\\mathcal {V}} \\left(f ^ {t}\\right) = P _ {s i m} \\left(f ^ {t}\\right) + P _ {d i s c} \\left(f ^ {t}\\right) \\in \\mathbb {R} ^ {K}, \\tag {13}", + "image_path": "5f93d058053552310da1bb0525d09a866c32fdb1c39ed445d789905d70c2fdf7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 274, + 554, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 554, + 381 + ], + "type": "text", + "content": "Multimodal Space Retrieval Augmentation. During test-time adaptation, multimodal space retrieval augmentation treats the test sample as a query and retrieves the most similar Gaussian center from the SMGD by evaluating their distances within CLIP's multimodal space. This process involves measuring the Kullback-Leibler (KL) divergence between the image features of the test sample or Gaussian centers and the textual features of the categories in CLIP's embedding space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "text", + "content": "Given the SMGD " + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "inline_equation", + "content": "S_{G} = \\{G,L\\}" + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "text", + "content": " and the text embeddings " + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "inline_equation", + "content": "Z\\in \\mathbb{R}^{D\\times K}" + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "text", + "content": " generated from CLIP's text encoder, we transform Gaussian centers " + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "text", + "content": " of the SMGD from the vision space " + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "text", + "content": " to the multimodal space " + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 382, + 554, + 430 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 382, + 437, + 555, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 437, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 382, + 437, + 555, + 452 + ], + "type": "interline_equation", + "content": "\\Psi = \\sigma \\left(Z ^ {T} G\\right) \\in \\mathbb {R} ^ {K \\times K}, \\tag {14}", + "image_path": "54dcc6d429fdc977059ec4a709ed2c781a8805756bc1b319e28f07838a8fbb69.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 460, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 460, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 313, + 460, + 554, + 485 + ], + "type": "text", + "content": " is a sigmoid function. We also project the test sample " + }, + { + "bbox": [ + 313, + 460, + 554, + 485 + ], + "type": "inline_equation", + "content": "f^{t}" + }, + { + "bbox": [ + 313, + 460, + 554, + 485 + ], + "type": "text", + "content": " to the multimodal space " + }, + { + "bbox": [ + 313, + 460, + 554, + 485 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 460, + 554, + 485 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 389, + 493, + 553, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 493, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 389, + 493, + 553, + 507 + ], + "type": "interline_equation", + "content": "\\psi = \\sigma \\left(Z ^ {T} f ^ {t}\\right) \\in \\mathbb {R} ^ {K}. \\tag {15}", + "image_path": "59dc4f2aef69381e27f3341971ae3391833dc3c397243cb11e23a4e739ce5411.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 516, + 554, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 554, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 554, + 550 + ], + "type": "text", + "content": "We use Kullback-Leibler (KL) divergence to compare the similarities of the two distributions in the multimodal space as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 393, + 552, + 553, + 565 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 552, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 393, + 552, + 553, + 565 + ], + "type": "interline_equation", + "content": "\\Phi_ {k} = K L (\\psi | | \\Psi_ {k}), \\tag {16}", + "image_path": "27102fa228559875b34c11c1f857ef6041d7114f122f54c1e17b16ef56fc40b5.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\Psi_{k}" + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "text", + "content": "-th column vector of " + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "inline_equation", + "content": "KL(P|Q) = \\sum_{x\\in \\mathcal{X}}P(x)\\log \\left(\\frac{P(x)}{Q(x)}\\right)" + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "text", + "content": ". Therefore, " + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\Phi = [\\Phi_1,\\Phi_2,\\dots ,\\Phi_K]\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "text", + "content": " consists of KL-divergence values. Since the most similar samples have low KL-divergence values (close to 0), they receive small weights. To address this, we negate the values in " + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "text", + "content": ". The prediction of the multimodal space " + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 570, + 554, + 660 + ], + "type": "text", + "content": " retrieval augmentation will be given as" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 391, + 670, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 670, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 391, + 670, + 553, + 685 + ], + "type": "interline_equation", + "content": "P _ {R} ^ {\\mathcal {M}} = - L \\Phi \\in \\mathbb {R} ^ {K}. \\tag {17}", + "image_path": "a7d5154865c5793a0ee8825356c374fbddea2f0da97f7141e187340f5302fb28.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Then, we can obtain the output of the multimodal retrieval augmentation (MRA) module " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{R}}" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " as a combination of the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8814" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 211 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 211 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 211 + ], + "type": "table", + "html": "
MethodAircraftCaltech101CarsDTDEuroSATFlower102Food101PetsSUN397UCF101Average
CLIP-ViT-B/1623.2293.5566.1145.0450.4266.9982.8686.9265.6365.1664.59
CoOp18.4793.7064.5141.9246.3968.7185.3089.1464.1566.5563.88
CoCoOp22.2993.7964.9045.4539.2370.8583.9790.4666.8968.4464.63
TPT24.7894.1666.8747.7542.4468.9884.6787.7965.5068.0465.10
DiffTPT25.6092.4967.0147.0043.1370.1087.2388.2265.7462.6765.47
MTA25.3294.1368.0545.5938.7168.2684.9588.2264.9868.1164.63
DN24.3093.6064.0045.7053.3068.0086.0087.7066.5068.4065.75
ZERO25.2193.6668.0446.1234.3367.6886.5387.7565.0367.7767.72
DMN24.8494.1265.6444.3947.7771.3884.4889.0766.2866.7565.47
TDA23.9194.2467.2847.4058.0071.4286.1488.6367.6270.6667.53
TT-RAA (ours)25.3894.0866.4247.9966.1272.6886.0989.8367.6971.2968.76
", + "image_path": "dd5972123e5ec6fb949d290b5328f1504905225cda1301239fa7b97ff0a87c4f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 219, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 219, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 555, + 253 + ], + "type": "text", + "content": "Table 1. Results on the cross-domain benchmark. Our TT-RAA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, four training-based adaptation methods (CoOp, CoCoOp, TPT, and DiffTPT), and five training-free adaptation methods (MTA, DN, DMN, ZERO, and TDA)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 259, + 295, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 295, + 285 + ], + "type": "text", + "content": "vision space retrieval argumentation result " + }, + { + "bbox": [ + 55, + 259, + 295, + 285 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{R}}^{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 259, + 295, + 285 + ], + "type": "text", + "content": " and multimodal space augmentation result " + }, + { + "bbox": [ + 55, + 259, + 295, + 285 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{R}}^{\\mathcal{M}}" + }, + { + "bbox": [ + 55, + 259, + 295, + 285 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 290, + 295, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 290, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 137, + 290, + 295, + 306 + ], + "type": "interline_equation", + "content": "P _ {\\mathrm {R}} = P _ {\\mathrm {R}} ^ {\\mathcal {V}} + P _ {\\mathrm {R}} ^ {\\mathcal {M}}. \\tag {18}", + "image_path": "090a7000481fc6abac61ecb179e0b3464a4ffdf6c87442be546c270e18f7e08e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 313, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 296, + 350 + ], + "type": "text", + "content": "After that, we compute our final TT-RAA logits by combining the original CLIP logits " + }, + { + "bbox": [ + 55, + 313, + 296, + 350 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 313, + 296, + 350 + ], + "type": "text", + "content": " with the MRA prediction " + }, + { + "bbox": [ + 55, + 313, + 296, + 350 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{R}}" + }, + { + "bbox": [ + 55, + 313, + 296, + 350 + ], + "type": "text", + "content": " using a weighting factor " + }, + { + "bbox": [ + 55, + 313, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 313, + 296, + 350 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 357, + 295, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 357, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 116, + 357, + 295, + 371 + ], + "type": "interline_equation", + "content": "P _ {\\mathrm {T T} - \\mathrm {R A A}} = P _ {\\mathrm {C L I P}} + \\alpha P _ {\\mathrm {R}}. \\tag {19}", + "image_path": "921d8c64c027d846c7a5065be5f8bc567fe00de4af6d172da8f8dd7cac4cc95e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 379, + 296, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 296, + 391 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 296, + 391 + ], + "type": "text", + "content": "In the end, we can obtain the final prediction of TT-RAA as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 403, + 295, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 403, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 118, + 403, + 295, + 422 + ], + "type": "interline_equation", + "content": "\\hat {y} = \\underset {k} {\\operatorname {a r g m a x}} P _ {\\mathrm {T T} - \\mathrm {R A A}} (k), \\tag {20}", + "image_path": "1be5e023fcdf5bd0c23d3967a99b38460e24326ba0725373f006cbc0c66f2aff.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 157, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 157, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 157, + 439 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 426, + 157, + 439 + ], + "type": "inline_equation", + "content": "k\\in \\{1,2,\\dots,K\\}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 449, + 137, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 137, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 137, + 462 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 468, + 115, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 115, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 115, + 479 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 486, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 665 + ], + "type": "text", + "content": "To validate the method's test time adaptation capability across different domains and variations within a familiar domain, we conducted experiments on two benchmarks: a cross-domain (CD) benchmark comprising 10 datasets and an out-of-distribution (OOD) benchmark comprising 4 datasets. The CD benchmark primarily assesses transferability by testing on domains that diverge significantly from the original training distribution. Specifically, it includes Aircraft [26], Caltech101 [9], Cars [18], DTD [6], EuroSAT [13], Flower102 [29], Food101 [3], Pets [32], SUN397 [48], and UCF101 [40]. In contrast, the OOD benchmark focuses on robustness to distribution shifts within the same general domain. Specifically, it includes ImageNet-A [15], ImageNet-V2 [36], ImageNet-R [14], and ImageNet-S [47]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 672, + 189, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 189, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 189, + 685 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "We implement our method based on the pre-trained CLIP architecture [34], which comprises a vision transformer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 259, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 555, + 392 + ], + "type": "text", + "content": "(ViT) based image encoder [7] and a transformer-based text encoder [44]. Our experimental setup follows the challenging single-image test-time adaptation scenario, processing samples sequentially with a batch size of 1. To maintain computational efficiency and eliminate the need for backpropagation during inference, we utilize the hand-crafted prompts as proposed in [34] rather than learnable prompts. The experiments are conducted using the PyTorch framework, with evaluations performed on a single NVIDIA RTX A5000 GPU. Following standard practice in similar tasks, we employ top-1 accuracy as our evaluation metric." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 397, + 541, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 541, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 541, + 410 + ], + "type": "text", + "content": "4.3. Comparisons with State-of-the-art Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 415, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 641 + ], + "type": "text", + "content": "We conduct extensive experiments to evaluate TT-RAA against state-of-the-art approaches on the CD benchmark in Table 1 and OOD benchmark in Table 2. The competing methods include the CLIP baseline [34], training-based adaptation methods, and training-free adaptation methods. Discussions on Competing Methods. In the field of efficient domain adaptation, several methods require training, either in the training stage or at the test time, to align the model more closely with the target distribution. Context optimization (CoOp) [57], conditional context optimization (CoCoOp) [56] both involve learning prompts during a training phase. CoOp focuses on optimizing prompt tokens that are fixed for each domain, while CoCoOp extends this to generate a conditional prompt token for each image. Test-time prompt tuning (TPT) [38] and diffusion model for TPT (DiffTPT) [10], on the other hand, adapt the model at test time by tuning prompt embeddings. TPT performs direct gradient-based prompt tuning using test data, while DiffTPT uses new data generated by diffusion models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "In addition to training-based methods, we also compare with several training-free adaptation methods, including two distribution-based methods, one entropy-based method, and two cache-based methods. Distribution-based methods evaluated are MeanShift for Test-time Augmentation (MTA) [51] and Distribution Normalization (DN) [58]." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8815" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 70, + 477, + 218 + ], + "blocks": [ + { + "bbox": [ + 133, + 70, + 477, + 218 + ], + "lines": [ + { + "bbox": [ + 133, + 70, + 477, + 218 + ], + "spans": [ + { + "bbox": [ + 133, + 70, + 477, + 218 + ], + "type": "table", + "html": "
MethodImageNet-AImageNet-V2ImageNet-RImageNet-SAverage
CLIP-ViT-B/1649.8961.8877.6548.2459.42
CoOp49.7164.2075.2147.9959.28
CoCoOp50.6364.0776.1848.7559.91
TPT54.7763.4577.0647.9460.81
DiffTPT55.6865.1075.0046.8060.52
MTA57.4163.6176.9248.5861.63
DN58.7162.8980.2048.9462.69
ZERO59.6164.1677.2248.4062.35
DMN58.2865.1778.5553.2063.80
TDA60.1164.6780.2450.5463.89
TT-RAA (Ours)60.5964.6980.5849.9863.96
", + "image_path": "a54d7495ac1af9a8efe4e9d87b22c6a6a3d51bd405b7491bff715157fa59936b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 227, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 227, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 555, + 261 + ], + "type": "text", + "content": "Table 2. Results on the out-of-distribution benchmark. Our TT-RAA is compared with several state-of-the-art methods designed for vision-language models: the baseline method CLIP, four training-based adaptation methods (CoOp, CoCoOp, TPT, and DiffTPT), and five training-free adaptation methods (MTA, ZERO, DN, DMN, and TDA)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 266, + 294, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 294, + 411 + ], + "type": "text", + "content": "They improve the test data distribution estimation via MeanShift algorithm and distribution normalization, respectively. Cache-based methods include the Dual Memory Network (DMN) [55] and TDA [17]. DMN comprises a dynamic and a static cache, while TDA employs a positive and a negative cache. For a fair comparison, we reproduced DMN's zero-shot results using our standardized prompts instead of DMN's original LLM-generated prompts, and we reduced the batch size to 1. We also compare against ZERO [8], an entropy-based method. ZERO sets the temperature of most confident predictions as zero to approximate marginal entropy minimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "text", + "content": "Results on the Cross-domain Benchmark. Despite having the advantage of task-specific training, training-based adaptation methods, such as CoOp, CoCoOp, TPT, and DiffTPT, show limited generalization capabilities. TT-RAA outperforms these methods by significant margins (about " + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "inline_equation", + "content": "3\\% - 4\\%" + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "text", + "content": "). On fine-grained recognition tasks such as Flowers102, TT-RAA achieves " + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "inline_equation", + "content": "72.68\\%" + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "text", + "content": " accuracy, surpassing CoCoOp " + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "inline_equation", + "content": "(70.85\\%)" + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "inline_equation", + "content": "1.83\\%" + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "text", + "content": ". This gap widens further on specialized datasets like SUN397, where TT-RAA " + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "inline_equation", + "content": "(67.69\\%)" + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "text", + "content": " demonstrates superior feature adaptation compared to CoCoOp " + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "inline_equation", + "content": "(64.15\\%)" + }, + { + "bbox": [ + 55, + 412, + 295, + 568 + ], + "type": "text", + "content": ". These results suggest that static prompt learning, even with input-conditional mechanisms, may not be sufficient enough for handling significant domain shifts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "In comparison to existing training-free adaptation approaches, TT-RAA demonstrates consistent superiority across various domain shifts. On remote sensing data (EuroSAT), TT-RAA " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "(66.12\\%)" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " significantly outperforms TDA " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "(58.00\\%)" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "8.12\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ". This substantial gain can be attributed to our SMGD effectively capturing the unique characteristics of satellite imagery. Similarly, for action recognition (UCF101), TT-RAA " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "(71.29\\%)" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " surpasses TDA " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "(70.66\\%)" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "0.63\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ", demonstrating robust feature adaptation. The performance advantages extend to other specialized domains as well. This consistent improvement across diverse domains validates the effectiveness of our method in bridging" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 267, + 451, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 451, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 451, + 278 + ], + "type": "text", + "content": "domain gaps in test-time settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 280, + 553, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 553, + 483 + ], + "type": "text", + "content": "Results on the Out-of-distribution Benchmark. The experimental results in Table 2 demonstrate that our proposed TT-RAA method achieves state-of-the-art performance on the OOD benchmark with an average accuracy of " + }, + { + "bbox": [ + 313, + 280, + 553, + 483 + ], + "type": "inline_equation", + "content": "63.96\\%" + }, + { + "bbox": [ + 313, + 280, + 553, + 483 + ], + "type": "text", + "content": " across all ImageNet variants, outperforming both training-based adaptation methods (CoOp, CoCoOp, TPT, DiffTPT) and training-free approaches (MTA, DN, ZERO, DMN, TDA). Notably, TT-RAA exhibits good performance on ImageNet-R " + }, + { + "bbox": [ + 313, + 280, + 553, + 483 + ], + "type": "inline_equation", + "content": "(80.58\\%)" + }, + { + "bbox": [ + 313, + 280, + 553, + 483 + ], + "type": "text", + "content": " and ImageNet-A " + }, + { + "bbox": [ + 313, + 280, + 553, + 483 + ], + "type": "inline_equation", + "content": "(60.59\\%)" + }, + { + "bbox": [ + 313, + 280, + 553, + 483 + ], + "type": "text", + "content": ", highlighting its robust generalization capabilities under diverse distribution shifts. When examining performance across different distribution shifts, we observe that most methods achieve their highest accuracy on ImageNet-R and ImageNet-V2. ImageNet-S with sketches poses the greatest challenge for all methods, indicating that semantic variations remain difficult to address even with advanced adaptation techniques." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 491, + 416, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 416, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 416, + 503 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 509, + 553, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 553, + 592 + ], + "type": "text", + "content": "We conduct ablation studies on the CD benchmark to systematically evaluate the effectiveness of each proposed component in TT-RAA. Figure 4 presents the quantitative results of our analysis with both ResNet-50 and ViT-B/16 backbones. We establish our baseline using the TDA method with one-shot capacity, which achieves " + }, + { + "bbox": [ + 313, + 509, + 553, + 592 + ], + "type": "inline_equation", + "content": "59.82\\%" + }, + { + "bbox": [ + 313, + 509, + 553, + 592 + ], + "type": "text", + "content": " (ResNet-50) and " + }, + { + "bbox": [ + 313, + 509, + 553, + 592 + ], + "type": "inline_equation", + "content": "67.04\\%" + }, + { + "bbox": [ + 313, + 509, + 553, + 592 + ], + "type": "text", + "content": " (ViT-B/16) accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 594, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 689 + ], + "type": "text", + "content": "Streaming Mixture of Gaussian Database Module. When integrating the SMGD module with the baseline, we observe a performance improvement of about " + }, + { + "bbox": [ + 313, + 594, + 553, + 689 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 313, + 594, + 553, + 689 + ], + "type": "text", + "content": " with ViT-B/16 and " + }, + { + "bbox": [ + 313, + 594, + 553, + 689 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 313, + 594, + 553, + 689 + ], + "type": "text", + "content": " with ResNet-50. This enhancement demonstrates that our SMGD effectively captures the dynamic distribution of test samples while maintaining computational efficiency. Its adaptive nature in updating class statistics online is useful for handling distribution shifts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Multimodal Retrieval Augmentation Module. Incorporating the MRA module also improved the performance by" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8816" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 72, + 289, + 180 + ], + "blocks": [ + { + "bbox": [ + 65, + 72, + 289, + 180 + ], + "lines": [ + { + "bbox": [ + 65, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 289, + 180 + ], + "type": "image", + "image_path": "b56bbdf756d4d985f5e316e87d7d3fff514621606d1f8f38adb12e338164e796.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 191, + 296, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 296, + 237 + ], + "type": "text", + "content": "Figure 4. Ablation studies of TT-RAA. The SMGD and MRA modules independently improve performance over the baseline. Their combination yields further performance gains, demonstrating complementary benefits." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 257, + 296, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 341 + ], + "type": "text", + "content": "about " + }, + { + "bbox": [ + 55, + 257, + 296, + 341 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 55, + 257, + 296, + 341 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 257, + 296, + 341 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 55, + 257, + 296, + 341 + ], + "type": "text", + "content": " with ViT-B/16 and ResNet-50, respectively. This gain supports our hypothesis regarding vision space misalignment in CLIP, which stems from its vision-text joint optimization objective. The performance improvement demonstrates that projecting both the SMGD database and test samples from vision space to multimodal space effectively addresses this misalignment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "type": "text", + "content": "Complementary Benefits of Proposed Modules. When combining both SMGD and MRA modules (denoted as TT-RAA), we achieve the best performance of " + }, + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "type": "inline_equation", + "content": "68.76\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "type": "text", + "content": " with ViT-B/16 and " + }, + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "type": "inline_equation", + "content": "61.25\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "type": "text", + "content": " with ResNet-50, representing a substantial improvement of " + }, + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "type": "inline_equation", + "content": "1.72\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "type": "inline_equation", + "content": "1.43\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 424 + ], + "type": "text", + "content": " over the baseline. These gains suggest that SMGD and MRA offer complementary and synergistic benefits." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 296, + 509 + ], + "type": "text", + "content": "Analysis of Feature Backbone Impact. The effectiveness of our approach is further validated across two vision ecoders: ViT-B/16 and ResNet-50. These consistent improvements across different backbones suggest that our adaptation strategy effectively addresses fundamental domain shift challenges rather than exploiting architecture-specific characteristics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 517, + 200, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 200, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 200, + 529 + ], + "type": "text", + "content": "4.5. Hyperparameter Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "The hyperparameter analysis of the update coefficient " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " on the UCF101 is shown in Fig. 5a. When " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\eta = 0" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": ", the model maintains a static Gaussian center initialized at the start, resulting in poor adaptation. From 0 to 0.55, accuracy increases sharply, highlighting the importance of incorporating new information. However, the performance deterioration beyond 0.65 reveals a critical threshold where excessive emphasis on new samples compromises the model's stability. This observation aligns with theoretical expectations, as larger " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " values reduce the influence of historical information, potentially leading to over-adaptation to new samples. The empirical optimal value of " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\eta = 0.65" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " suggests that maintaining approximately " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " of historical information while incorporating " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " of new information achieves the best trade-off." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 71, + 432, + 159 + ], + "blocks": [ + { + "bbox": [ + 319, + 71, + 432, + 159 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 432, + 159 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 432, + 159 + ], + "type": "image", + "image_path": "97b338899f73e273bce87e9f8c68fbabc888cebfba23eb72dbaa4d7fb3760afa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 162, + 381, + 171 + ], + "lines": [ + { + "bbox": [ + 370, + 162, + 381, + 171 + ], + "spans": [ + { + "bbox": [ + 370, + 162, + 381, + 171 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 434, + 74, + 549, + 158 + ], + "blocks": [ + { + "bbox": [ + 434, + 74, + 549, + 158 + ], + "lines": [ + { + "bbox": [ + 434, + 74, + 549, + 158 + ], + "spans": [ + { + "bbox": [ + 434, + 74, + 549, + 158 + ], + "type": "image", + "image_path": "b695c877f870a06d95745d4684e2c0d472f39cd20e7a595ee0db7bf07168f3ae.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 162, + 498, + 171 + ], + "lines": [ + { + "bbox": [ + 487, + 162, + 498, + 171 + ], + "spans": [ + { + "bbox": [ + 487, + 162, + 498, + 171 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 181, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 227 + ], + "type": "text", + "content": "Figure 5. (a) Hyperparameter analysis of the update coefficient " + }, + { + "bbox": [ + 313, + 181, + 555, + 227 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 181, + 555, + 227 + ], + "type": "text", + "content": ". It shows that the best coefficient is 0.65. (b) Experiment with additional data. The method can accept additional data, which improve the performance significantly." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 246, + 544, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 544, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 544, + 258 + ], + "type": "text", + "content": "4.6. Experiments with Access to Additional Data" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 263, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 491 + ], + "type": "text", + "content": "In the standard test-time adaptation setting, access to the target domain's training set is typically restricted. However, in many cases, the training set data might be available. Finetuning large-scale vision-language models in such scenarios remains computationally expensive. This raises an important question: how can performance be improved using the training data without computationally expensive finetuning? With a slight modification to TT-RAA, we propose leveraging the target domain training set (if accessible prior to test time) to construct a database and calculate class centers using a Gaussian mixture model. By doing so, we perform the same MRA process as in TT-RAA. This modified approach, referred to as TT-RAA-A, demonstrates significant performance improvements while maintaining a low computational requirement. The corresponding results are presented in Fig. 5b, where TT-RAA-A outperforms TT-RAA by a large margin, highlighting the benefits of incorporating target training data in a training-free manner, which open a new direction on adaptation without training." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduced TT-RAA, a novel test-time retrieval augmented adaptation method that effectively addresses the computationally expensive domain adaptation challenges faced by vision-language models. Experimental results validated the effectiveness of SMGD by continuously estimating the test distribution and compressing all useful information into a single representation for each class without discarding old samples. By transforming Gaussian centers from vision space into a shared multimodal space, MRA enables more effective retrieval that leverages both visual and textual features. The success of TT-RAA opens up new possibilities for deploying vision-language models in real-world applications where computational resources are limited and rapid adaptation to new domains is crucial. Further exploration on adapting our method to larger multimodal foundation and large language models is promising." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8817" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 113 + ], + "type": "text", + "content": "[1] Meta AI. Llama 3.2: Revolutionizing edge ai and vision with open, customizable models. Meta AI Report, 2024. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 114, + 295, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 114, + 295, + 135 + ], + "spans": [ + { + "bbox": [ + 62, + 114, + 295, + 135 + ], + "type": "text", + "content": "[2] Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning. Springer, 2006. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 137, + 294, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 137, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 62, + 137, + 294, + 168 + ], + "type": "text", + "content": "[3] Lukas Bossard, Matthieu Guillaumin, et al. Food-101: Mining discriminative components with random forests. In European Conference on Computer Vision, 2014. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 170, + 294, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 294, + 223 + ], + "type": "text", + "content": "[4] Jun Chen, Han Guo, Kai Yi, Boyang Li, and Mohamed Elhoseiny. VisualGPT: Data-efficient adaptation of pretrained language models for image captioning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18030-18040, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 224, + 294, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 269 + ], + "type": "text", + "content": "[5] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. European Conference on Computer Vision, pages 104-120, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 270, + 294, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 294, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 294, + 302 + ], + "type": "text", + "content": "[6] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, , and A. Vedaldi. Describing textures in the wild. In IEEE Conference on Computer Vision and Pattern Recognition, 2014. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 304, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 294, + 346 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 347, + 294, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 294, + 391 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 294, + 391 + ], + "type": "text", + "content": "[8] Matteo Farina, Gianni Franchi, Giovanni Iacca, Massimiliano Mancini, and Elisa Ricci. Frustratingly easy test-time adaptation of vision-language models. Advances in Neural Information Processing Systems, 2024. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 392, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 446 + ], + "type": "text", + "content": "[9] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2004. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 448, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 448, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 448, + 295, + 491 + ], + "type": "text", + "content": "[10] Chun-Mei Feng, Kai Yu, Yong Liu, et al. Diverse data augmentation with diffusions for effective test-time prompt tuning. In IEEE/CVF International Conference on Computer Vision, pages 2704-2714, 2023. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 492, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 492, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 57, + 492, + 295, + 535 + ], + "type": "text", + "content": "[11] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, et al. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "type": "text", + "content": "[12] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, et al. Retrieval-augmented generation for large language models: A survey. arXiv preprint arXiv:2312.10997, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 581, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 624 + ], + "type": "text", + "content": "[13] Patrick Helber, Benjamin Bischke, et al. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "text", + "content": "[14] Dan Hendrycks, Steven Basart, Norman Mu, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[15] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15262-15271, 2021. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. International Conference on Machine Learning, pages 4904-4916, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "text", + "content": "[17] Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "type": "text", + "content": "[18] Jonathan Krause, Michael Stark, et al. 3D object representations for fine-grained categorization. In IEEE International Conference on Computer Vision Workshops, 2013. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 251 + ], + "type": "text", + "content": "[19] Youngjun Lee, Doyoung Kim, Junhyeok Kang, Jihwan Bang, Hwanjun Song, et al. RA-TTA: Retrieval-augmented test-time adaptation for vision-language models. In International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 254, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 307 + ], + "type": "text", + "content": "[20] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, and Vladimir others Karpukhin. Retrievalaugmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems, pages 9459-9474, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[21] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. VisualBERT: A simple and perform-. mant baseline for vision and language. arXiv preprint arXiv:1908.03557, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 398 + ], + "type": "text", + "content": "[22] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In International conference on machine learning, pages 6028-6039. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 432 + ], + "type": "text", + "content": "[23] Jian Liang, Ran He, et al. A comprehensive survey on test-time adaptation under distribution shifts. International Journal of Computer Vision, pages 1-34, 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 433, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 433, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 433, + 553, + 487 + ], + "type": "text", + "content": "[24] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. Advances in Neural Information Processing Systems, 35:17612-17625, 2022. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 488, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 553, + 521 + ], + "type": "text", + "content": "[25] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 555 + ], + "type": "text", + "content": "[26] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, et al. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 557, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 599 + ], + "type": "text", + "content": "[27] Yifei Ming and Yixuan Li. Understanding retrieval-augmented task adaptation for vision-language models. In International Conference on Machine Learning, pages 35719-35743, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "text", + "content": "[28] Zachary Nado, Shreyas Padhy, D Sculley, Alexander D'Amour, Balaji Lakshminarayanan, et al. Evaluating prediction-time batch normalization for robustness under covariate shift. arXiv preprint arXiv:2006.10963, 2020. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[29] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In Indian Conference on Computer Vision, Graphics and Image Processing, 2008. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[30] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8818" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "text", + "content": "test-time adaptation in dynamic wild world. In International Conference on Learning Representations, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "type": "text", + "content": "[31] OpenAI. Gpt-4v(ison) system card. In OpenAI Technical Report, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "type": "text", + "content": "[32] Omkar M. Parkhi, Andrea Vedaldi, Andrew Zisserman, and C. V. Jawahar. Cats and dogs. In IEEE Conference on Computer Vision and Pattern Recognition, 2012. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "type": "text", + "content": "[33] Rizwan Qureshi, Ranjan Sapkota, Abbas Shah, et al. Thinking beyond tokens: From brain-inspired intelligence to cognitive foundations for artificial general intelligence and its societal impact. arXiv preprint arXiv:2507.00951, 2025. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 198, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 241 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 243, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 243, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 57, + 243, + 295, + 275 + ], + "type": "text", + "content": "[35] Aditi Raghunathan, Prateek Jain, et al. Learning mixture of gaussians with streaming data. Advances in Neural Information Processing Systems, 30, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 277, + 294, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 309 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 309 + ], + "type": "text", + "content": "[36] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, et al. Do imagenet classifiers generalize to imagenet? In International Conference on Machine Learning, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 311, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 311, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 311, + 294, + 354 + ], + "type": "text", + "content": "[37] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, et al. Improving robustness against common corruptions by covariate shift adaptation. Advances in Neural Information Processing Systems, 2020. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 355, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 409 + ], + "type": "text", + "content": "[38] Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems, 2022. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 476 + ], + "type": "text", + "content": "[39] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15638-15650, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 478, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 510 + ], + "type": "text", + "content": "[40] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. CoRR, abs/1212.0402, 2012. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "text", + "content": "[41] Tao Sun, Cheng Lu, and Haibin Ling. Prior knowledge guided unsupervised domain adaptation. In European Conference on Computer Vision. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 545, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 599 + ], + "type": "text", + "content": "[42] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International Conference on Machine Learning, pages 9229-9248. PMLR, 2020. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "text", + "content": "[43] Vishaal Udandarao, Ankush Gupta, and Samuel Albanie. SuS-X: Training-free name-only transfer of vision-language models. In IEEE/CVF International Conference on Computer Vision, pages 2725-2736, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[44] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[45] Matthew Wallingford, Vivek Ramanujan, Alex Fang, Aditya Kusupati, Roozbeh Mottaghi, Aniruddha Kembhavi, et al." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Neural priming for sample-efficient adaptation. Advances in Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[46] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations, 2021. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "text", + "content": "[47] Haohan Wang, Songwei Ge, et al. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[48] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In IEEE Conference on Computer Vision and Pattern Recognition, 2010. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[49] Shiqi Yang, Joost Van de Weijer, Luis Herranz, Shangling Jui, et al. Exploiting the intrinsic neighborhood structure for source-free domain adaptation. Advances in Neural Information Processing Systems, 34:29393-29405, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 262, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 315 + ], + "type": "text", + "content": "[50] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. International Conference on Learning Representations, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 317, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 371 + ], + "type": "text", + "content": "[51] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23783-23793, 2024. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "text", + "content": "[52] Jingyi Zhang, Jiaxing Huang, et al. Vision-language models for vision tasks: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 405, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 459 + ], + "type": "text", + "content": "[53] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. VinVL: Revisiting visual representations in vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5579-5588, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 514 + ], + "type": "text", + "content": "[54] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision, pages 493-510. Springer, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 515, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 569 + ], + "type": "text", + "content": "[55] Yabin Zhang, Wenjie Zhu, Hui Tang, Zhiyuan Ma, Kaiyang Zhou, and Lei Zhang. Dual memory networks: A versatile adaptation approach for vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28718-28728, 2024. 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "text", + "content": "[56] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16816-16825, 2022. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 615, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 656 + ], + "type": "text", + "content": "[57] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[58] Yifei Zhou, Juntao Ren, Fengyu Li, Ramin Zabih, and Ser Nam Lim. Test-time distribution normalization for contrastively learned visual-language models. Advances in Neural Information Processing Systems, pages 47105-47123, 2023. 2, 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8819" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/8a26d38c-cedb-499b-b41f-c629d5b94654_content_list.json b/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/8a26d38c-cedb-499b-b41f-c629d5b94654_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..211665cf1358ea2128a60178ac1539809a82bf2b --- /dev/null +++ b/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/8a26d38c-cedb-499b-b41f-c629d5b94654_content_list.json @@ -0,0 +1,1618 @@ +[ + { + "type": "text", + "text": "Test-time Adaptation for Foundation Medical Segmentation Model without Parametric Updates", + "text_level": 1, + "bbox": [ + 122, + 128, + 875, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kecheng Chen $^{1}$ , Xinyu Luo $^{1}$ , Tiexin Qin $^{1}$ , Jie Liu $^{1}$ , Hui Liu $^{1}$", + "bbox": [ + 250, + 202, + 750, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Victor Ho Fun Lee², Hong Yan¹, Haoliang Li¹*", + "bbox": [ + 305, + 220, + 696, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ City University of Hong Kong $^{2}$ HKU {ck.ee@my.cityu.edu.hk} $\\star$ : Corresponding author", + "bbox": [ + 127, + 238, + 872, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 259, + 297, + 336, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundation medical segmentation models, with MedSAM being the most popular, have achieved promising performance across organs and lesions. However, MedSAM still suffers from compromised performance on specific lesions with intricate structures and appearance, as well as bounding box prompt-induced perturbations. Although current test-time adaptation (TTA) methods for medical image segmentation may tackle this issue, partial (e.g., batch normalization) or whole parametric updates restrict their effectiveness due to limited update signals or catastrophic forgetting in large models. Meanwhile, these approaches ignore the computational complexity during adaptation, which is particularly significant for modern foundation models. To this end, our theoretical analyses reveal that directly refining image embeddings is feasible to approach the same goal as parametric updates under the MedSAM architecture, which enables us to realize high computational efficiency and segmentation performance without the risk of catastrophic forgetting. Under this framework, we propose to encourage maximizing factorized conditional probabilities of the posterior prediction probability using a proposed distribution-approximated latent conditional random field loss combined with an entropy minimization loss. Experiments show that we achieve about $3\\%$ Dice score improvements across three datasets while reducing computational complexity by over 7 times. Codes are available at here.", + "bbox": [ + 109, + 329, + 485, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 112, + 780, + 243, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, foundation segmentation models for medical images have attracted massive attention from the medical imaging community [31], as medical image segmentation plays an indispensable role in many downstream tasks (e.g., surgical navigation and treatment planning). Medical Segment Anything Model [15] (MedSAM) is the most popular foundation segmentation model, which", + "bbox": [ + 109, + 806, + 483, + 912 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/347ac5eb387cc0f60aa3964b10faa3ddae8b15b1101414eca83eff5e22822baa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Update TypeMethodDatasetsGFLOPs↓
OCODSCGM
Direct Inference87.7780.5756.67-
NormalizationTENT [25] (ICLR'21)86.2478.0854.34743.5
InTEnt [6] (CVPR'24)86.9582.4049.81744.8
GraTa [32] (AAAI'25)88.5780.4155.142975.6
PASS [30] (TMI'24)84.0571.8154.80745.8
FullMEMO [33] (NeurIPS'22)88.1280.0457.485646.5
CRF-SOD [24] (CVPR'23)88.9579.7956.811115.2
LatentOurs90.8484.1959.4997.3
", + "bbox": [ + 519, + 297, + 890, + 401 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Table 1. Average Dice coefficient $(\\uparrow)$ of different TTA methods on three datasets. Normalization-based methods are customized to BN layers, exhibiting marginal (even degraded) performance for LN-based foundation medical segmentation model. We report the computational complexity using GFLOPs (Giga Floating Point Operations), including forward and backward propagations. We leverage latent updates to achieve $\\sim 7\\times$ GFLOPs reduction compared with parametric update-based counterparts with performance gains.", + "bbox": [ + 511, + 411, + 883, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "exhibits impressive universality across organs and lesions. Since MedSAM is pre-trained on large-scale medical image datasets, it typically enables the high-quality segmentation mask of the region of interest (RoI) in a zero-shot manner, which has contributed to extensive areas, e.g., semi/weak-supervised medical image segmentation [29] or anomaly detection [16] using provided segmentation masks as pseudo labels or RoIs.", + "bbox": [ + 509, + 547, + 883, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, due to the extreme modality imbalance of medical images (where CT, MRI, and endoscopy images are dominant), MedSAM has compromised performance on less-represented modalities (e.g., eye fundus images) [15]. It is also difficult for MedSAM to maneuver specific lesions with intricate structures in the segmentation area [15], e.g., optic cup segmentation suffers from many vessel-like branching structures. Although supervised fine-tuning on specific modalities and lesions with annotated masks seems effective in tackling these issues [28], the overwhelming computational consumption of fine-tuning and inherent data scarcity (or privacy) concerns regarding annotated medical images may impede this process. Moreover, some test-time perturbations are still encountered, even though a well-calibrated and fine-tuned model has been prepared. For instance,", + "bbox": [ + 509, + 670, + 883, + 912 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20075", + "bbox": [ + 478, + 950, + 517, + 963 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mild looseness or shrinkage perturbations for bounding box prompts are sufficient to cause performance degradation for MedSAM (see Tables 2 and 3), which is clinically unfavorable as delineating a perfect prompt is usually labor-intensive and time-consuming for physicians.", + "bbox": [ + 109, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above-mentioned issues, test-time adaptation (TTA) for foundation medical segmentation models is a feasible solution that enables performance gains on a specific unlabeled target (test) domain using the pre-trained model only. Although recent TTA approaches for medical image segmentation achieve promising results, we argue that scaling them to foundation segmentation models is difficult. First, most methods utilize various self-supervised objectives (e.g., prediction consistency over different augmentations [2, 6, 30]) to conduct batch normalization (BN)-centric fin-tuning, which may endure limited update signals with marginal adaptation as modern foundation models employ layer normalization (LN) without source-trained statistics like BN. Second, some approaches [27] update the whole model to explore more gains, but even minimal parameter adjustments in large models can precipitate catastrophic forgetting [10]. Third, most methods ignore computational complexity during the adaptation, which is particularly significant for foundation medical segmentation models1, since continuous forward and backward propagations on a large model are computationally intensive, especially on resource-constrained edge devices. Overall, existing methods struggle with maintaining a trade-off between sufficient update signals and knowledge preservation, and ignore computational consumption of the adaptation for large models.", + "bbox": [ + 109, + 167, + 482, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose a novel TTA framework for foundation medical segmentation models like MedSAM. Compared to existing methods, one of the most pivotal differences is that our proposed method updates latent representations rather than any model parameters², which can not only endow maximal update flexibility under a forgetting-free context but also circumvent heavy forward and backward computations on sophisticated architectures (e.g., image encoder). Specifically, by incorporating an off-the-shelf conditionally independent assumption between the input image X and the prediction Y in MedSAM, we provide a theoretical analysis to uncover that directly refining the image embedding Z (a.k.a. latent refinement) can encourage maximizing factorized conditional probabilities (including latent fidelity $P(\\mathbf{Z}|\\mathbf{X})$ and segmentation likelihood $P(\\mathbf{Y}|\\mathbf{Z}))$ of posterior prediction probability $P(\\mathbf{Y}|\\mathbf{X})$ , in a similar objective like parametric updates. Under this latent", + "bbox": [ + 109, + 577, + 482, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "refinement framework, we devise a novel distributionapproximated latent conditional random field (CRF) loss to leverage useful cues in the input space for explicitly maximizing the posterior distribution of the latent fidelity. Meanwhile, we utilize entropy minimization to enforce the usability of the final predictions for implicitly maximizing segmentation likelihood. Finally, the mask decoder can directly impose the refined latent without inferring the image embedding again. Our contributions include three-fold:", + "bbox": [ + 511, + 90, + 883, + 241 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We provide theoretical analyses to uncover that direct latent refinement is feasible to approach the same goal as parametric updates under the MedSAM architecture, which enables us to realize high computational efficiency and segmentation performance (refers to Table 1) without the risk of catastrophic forgetting.", + "- A distribution-approximated latent CRF loss based on a Maximum Mean Discrepancy-aware co-occurrent measure is proposed to introduce potential useful cues from input space, leading to sufficient statistics of image embedding over the input image.", + "- The latent refinement framework is validated on three multi-center medical image segmentation datasets with strong appearance variations and simulated prompt shifts. Extensive experiments demonstrate that our method can not only achieve average $\\sim 3\\%$ \n- Dice score improvements among three datasets but also enjoy around $7\\times$ computational complexity reduction compared with existing methods." + ], + "bbox": [ + 513, + 243, + 883, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 546, + 658, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Medical Image Segmentation with TTA", + "text_level": 1, + "bbox": [ + 511, + 571, + 849, + 587 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various test-time adaptation (TTA) methodologies have been developed to address distribution shifts during inference [3, 17, 25]. Recent advancements have extended TTA frameworks to more complex tasks, including super-resolution [4] and blind image quality assessment [19]. In the field of medical image segmentation, TTA tackles the common multi-center issue [13], where different hospitals may have specific imaging parameters with strong appearance variations [26]. To this end, some methods conducted the optimization of the BN layer, manipulating either the affine-transformation parameters [30] or source-trained BN statistics [6], where many self-supervised objectives and sophisticated alignment of source-trained statistics in the BN layer are leveraged. Meanwhile, some approaches [27] fine-tune the overall model to explore more gains using elaborated objectives and knowledge accumulation strategies. We argue that such parametric updates suffer from restricted effectiveness due to limited BN-centric update signals and catastrophic forgetting for a large pretrained model, respectively. The computational com", + "bbox": [ + 511, + 595, + 883, + 912 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1MedSAM obtains image embedding using an intricate ViT-based encoder (accounts for $98.5\\%$ computations) and a lightweight decoder. 2We refer to partial or whole model updates as parametric updates hereafter.", + "bbox": [ + 112, + 861, + 482, + 911 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20076", + "bbox": [ + 478, + 950, + 519, + 962 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "plexity is also ignored during adaptation. There are limited works specific to the TTA of foundation medical segmentation models. For example, Huang et al. [9] proposed an auxiliary online learning (AuxOL) and adaptive fusion for the foundation medical image segmentation model. However, AuxOL requires clinicians' annotations for model training, leading to potential unfeasibility in many scenarios. Beyond the medical image, Schön et al. [20] leveraged the clicked point prompts as cues for adaptation, which is inapplicable for MedSAM due to its bounding box-based prompts.", + "bbox": [ + 109, + 90, + 480, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, there is a lack of a customized approach for the TTA problem of foundation medical segmentation models when considering both effectiveness and efficiency.", + "bbox": [ + 111, + 257, + 480, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Latent Refinement for TTA", + "text_level": 1, + "bbox": [ + 112, + 313, + 357, + 328 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Typically, latent refinement is applied in the neural image compression community [1, 5, 14, 21, 23] to reduce the bit consumption of latent representation at compression time. Since the image compression task has a strong self-supervised signal that corresponds to reconstructing an original raw image by an encoder-decoder architecture [1, 5, 14, 21], it is more straightforward to leverage the latent refinement technique. Instead, there is no direct-supervised signal for medical image segmentation during inference, leaving unexplored space. To the best of our knowledge, we are the first to introduce the concept of latent refinement to test-time adaptation of medical image segmentation. More importantly, we provide theoretical analyses to uncover that directly refining the latent representation of the input image is feasible to approach the same goal as parametric update-based counterparts, which is not developed by previous literature.", + "bbox": [ + 109, + 335, + 483, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 112, + 607, + 246, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary and Theoretical Analysis", + "text_level": 1, + "bbox": [ + 112, + 632, + 434, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper mainly discusses the common MedSAM architecture in [15] with the bounding box prompt. During the test-time adaptation, the bounding box prompt is provided and fixed but is omitted in the following derivations for notation simplification.", + "bbox": [ + 109, + 655, + 482, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 1. (Architecture of MedSAM) Unlike U-Net and its variants [18], MedSAM has no skip connections between an image encoder $E_{\\theta}$ parameterized by $\\theta$ and a mask decoder $D_{\\varphi}$ parameterized by $\\varphi$ , which enables multiple interactions by reusing the image embedding. In light of this, the following proposition can be derived:", + "bbox": [ + 109, + 731, + 482, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proposition 1. Let $\\mathbf{X} \\in \\mathcal{X}$ , $\\mathbf{Z} \\in \\mathcal{Z}$ , and $\\mathbf{Y} \\in \\mathcal{Y}$ be an input image, its latent embedding, and its corresponding prediction. For a MedSAM model, $\\mathbf{Y}$ is strictly conditionally independent of $\\mathbf{X}$ given $\\mathbf{Z}$ , i.e., $\\mathbf{Y} \\perp \\mathbf{X} \\mid \\mathbf{Z}$ .", + "bbox": [ + 109, + 821, + 482, + 881 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof: Once the (deepest) latent embedding $\\mathbf{Z}$ is calculated by $\\mathbf{Z} = E_{\\theta}(\\mathbf{X})$ , the upsampling function $\\mathbf{Y} =$", + "bbox": [ + 111, + 882, + 482, + 912 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$D_{\\varphi}(\\mathbf{Z})$ is not directly related to the input image $\\mathbf{X}$ or its any latent variants, leading to $\\mathbf{Y}\\perp \\mathbf{X}\\mid \\mathbf{Z}$", + "bbox": [ + 511, + 90, + 880, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1. (General objective) Under maximum a posteriori (MAP) estimation, the objective of TTA for image segmentation is computing the optimal prediction $\\mathbf{Y}^*$ by maximizing the posterior probability conditioned on $\\mathbf{X}$ :", + "bbox": [ + 511, + 122, + 880, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} ^ {*} = D _ {\\varphi^ {*}} \\left(E _ {\\theta^ {*}} (\\mathbf {X})\\right), \\varphi^ {*}, \\theta^ {*} = \\arg \\max _ {\\theta , \\varphi} [ P (\\mathbf {Y} | \\mathbf {X}; \\theta , \\varphi) ] ^ {3}.\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 205, + 893, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Various TTA strategies are regarded as implicit posterior maximization by self-supervised loss-based partial or overall parameter updates. Here, we incorporate the architecture of the MedSAM as a factorized precondition to yield an extension of Theorem 1 as follows.", + "bbox": [ + 511, + 238, + 880, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Corollary 1. (Factorized objective of MedSAM) The log-posterior $\\log P(\\mathbf{Y}|\\mathbf{X})$ during test-time adaptation can be factorized into the sum of the segmentation likelihood $\\log P(\\mathbf{Y}|\\mathbf{Z})$ and the latent fidelity term $\\log P(\\mathbf{Z}|\\mathbf{X})$ based on $\\mathbf{Y}\\perp \\mathbf{X}\\mid \\mathbf{Z}$ , i.e.,", + "bbox": [ + 511, + 314, + 880, + 388 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} ^ {*} = D _ {\\varphi^ {*}} \\left(\\mathbf {Z} ^ {*}\\right), \\mathbf {Z} ^ {*} = E _ {\\theta^ {*}} (\\mathbf {X})\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 398, + 803, + 417 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi^ {*}, \\theta^ {*} = \\arg \\max _ {\\varphi , \\theta} [ \\log P (\\mathbf {Y} | \\mathbf {Z}; \\varphi) + \\log P (\\mathbf {Z} | \\mathbf {X}; \\theta) ],\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 417, + 872, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Proof: By recalling Prop. 1, $P(\\mathbf{Y}|\\mathbf{X}) \\propto P(\\mathbf{Y}|\\mathbf{Z}) \\cdot P(\\mathbf{Z}|\\mathbf{X})$ holds. Substituting this into the MAP objective in Theorem 1 yields the factorized objective above. Please refer to the Appendix for more details.", + "bbox": [ + 511, + 450, + 880, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Latent Refinement in MedSAM framework", + "text_level": 1, + "bbox": [ + 511, + 523, + 880, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem. For Corollary 1, one of the commonly adopted solutions for achieving optimal prediction $\\mathbf{Y}^*$ is directly to jointly optimize partial or whole parameters in $\\{\\varphi ,\\theta \\}$ , but two challenges are encountered as follows.", + "bbox": [ + 511, + 545, + 880, + 606 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Partial model updates. Most existing approaches are BN-centric strategies, which are inapplicable for LN-based MedSAM. Although updating the affine transformation parameters in LN layers is feasible, the performance may be marginal due to restricted model modifications especially in strong domain shifts.", + "- Whole model updates. It is difficult to maintain a trade-off between TTA performance and catastrophic forgetting. Meanwhile, optimizing the overall model is computationally expensive due to the considerable number of parameters in modern large models." + ], + "bbox": [ + 513, + 607, + 880, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Solution. In light of this, we are interested in directly optimizing the latent embedding rather than the model parameters. Assume the encoder $E_{\\theta}$ and the decoder $D_{\\varphi}$ are fixed during TTA, this scheme can not only circumvent the above-mentioned challenges (i.e., it is likely to update more \"parameters\" in the forgetting-free context) but also be equivalent to theoretically achieving the", + "bbox": [ + 511, + 772, + 880, + 877 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3In practice, one would optimize the expectation of the probability. To be simplified, we omit it in later formulations.", + "bbox": [ + 511, + 886, + 880, + 910 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20077", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1f61326352df930d53a2bb4300a8ce0ed762f2b9e4a917310b3e0e59c6740a5b.jpg", + "image_caption": [ + "Figure 1. The framework of the proposed method, including initial embedding inference, TTA via latent refinement, and final inference phases. Distribution-approximated latent CRF loss is visualized, where the loss between $\\mathbf{z}_i$ and $\\mathbf{z}_j$ is computed." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "same factorized objective of MedSAM as Corollary 1,", + "bbox": [ + 112, + 342, + 482, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} ^ {*} = D _ {\\varphi} (\\mathbf {Z} ^ {*}),\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 378, + 351, + 396 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} ^ {*} = \\arg \\max _ {\\mathbf {Z}} \\left[ \\log P (\\mathbf {Y} | \\mathbf {Z}; \\varphi) + \\log P (\\mathbf {Z} | \\mathbf {X}; \\theta) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 398, + 480, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Z} = E_{\\theta}(\\mathbf{X})$ can be initialized and refined iteratively. Typically, the total parameter number of $\\mathbf{Z}$ is overly larger than that of the normalization layer, resulting in more update flexibility. Under frozen $\\theta$ and $\\varphi$ , there is no knowledge forgetting during optimization. These two perspectives are more likely to activate better TTA performance collaboratively than existing methods. Effectiveness of Latent Refinement during TTA. First, the $\\log P(\\mathbf{Z}|\\mathbf{X})$ term encourages $\\mathbf{Z}$ to maintain sufficient statistics over the input $\\mathbf{X}$ , which is highly significant for the MedSAM framework during the TTA because 1) domain-specific variations at test time may degrade salient characteristics (e.g., segmentation cues like obvious edges or contours) of latent embedding, leading to suboptimal segmentation performance; 2) sufficient statistics relative to the input $\\mathbf{X}$ can encourage the conditional independence assumption of MedSAM, i.e., $\\mathbf{Y} \\perp \\mathbf{X} \\mid \\mathbf{Z}$ , to hold. Second, $\\log P(\\mathbf{Y}|\\mathbf{Z})$ guarantees the usability of latent embedding for final predictions, preventing any collapsed solutions.", + "bbox": [ + 111, + 426, + 483, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Practical implementation. For Eq. (1), one would update the initial latent embedding $\\mathbf{Z} = \\mathbf{Z}^0 = E_\\theta (\\mathbf{X})$ with $m\\geq 1$ steps, for each step $\\mathbf{Z}^{m} =$", + "bbox": [ + 112, + 729, + 482, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} ^ {m - 1} - \\epsilon \\cdot \\frac {\\partial [ - \\log P (\\mathbf {Y} ^ {m - 1} | \\mathbf {Z} ^ {m - 1} ; \\varphi) - \\log P (\\mathbf {Z} ^ {m - 1} | \\mathbf {X} ; \\theta ]}{\\partial \\mathbf {Z} ^ {m - 1}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 777, + 509, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon$ is the learning rate. The final prediction can be represented as $\\mathbf{Y}^m = D_{\\varphi}(\\mathbf{Z}^m)$ .", + "bbox": [ + 111, + 821, + 482, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Distribution-Approximated Latent CRF", + "text_level": 1, + "bbox": [ + 112, + 859, + 457, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To Eq. (2), one of the most important components is maximizing the log-probability $P(\\mathbf{Z}^{m - 1}|\\mathbf{X};\\theta)$ , i.e.,", + "bbox": [ + 111, + 881, + 482, + 912 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the latent fidelity term. For MedSAM, the volume of the image encoder parameter $\\theta$ is considerable, resulting in a high computation during the forward and backward propagations in the TTA phase. Benefiting from our latent refinement framework, we propose circumventing the image encoder $\\theta$ to directly enforce sufficient statistics over the input image $\\mathbf{X}$ , i.e., maximizing $\\log P(\\mathbf{Z}^{m-1}|\\mathbf{X})$ . By doing so, we can leverage potential useful cues in the input space, which is typically ignored by existing TTA methods that focus more on the label space (e.g., prediction consistency [2]).", + "bbox": [ + 511, + 342, + 883, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we introduce the concept of conditional random field (CRF) [22], which is usually employed to refine the final predictions by enforcing the co-occurrence between the label space $\\mathcal{V}$ and the input space $\\mathcal{X}$ . For our purpose, we extend the CRF to the co-occurrence between the latent space $\\mathcal{Z}$ and the input space $\\mathcal{X}$ . Theoretically, minimizing the CRF energy is equivalent to maximizing the posterior probability of the random field [11], which coincides with our optimization objective in Eq. (2) via a computationally efficient manner. Formally, given a random field $\\mathbf{Z} = E_{\\theta}(\\mathbf{X})$ over a set of vectors $\\{\\mathbf{z}_1,\\mathbf{z}_2,\\dots ,\\mathbf{z}_{N'}\\}$ , where $\\mathbf{Z}\\in \\mathbb{R}^{W^{\\prime}\\times H^{\\prime}\\times C^{\\prime}}$ is the latent embedding with a spatial resolution of $N^{\\prime} = W^{\\prime}\\times H^{\\prime}$ and a channel of $C^\\prime$ , i.e., $\\mathbf{z}\\in \\mathbb{R}^{1\\times C^{\\prime}}$ . Consider another random field $\\mathbf{X}$ over a set of vectors $\\{\\mathbf{x}_1,\\mathbf{x}_2,\\dots ,\\mathbf{x}_N\\}$ , where $\\mathbf{X}\\in \\mathbb{R}^{W\\times H\\times C}$ is the input intensities with a spatial resolution of $N = W\\times H$ and a channel of $C$ , i.e., $\\mathbf{x}\\in \\mathbb{R}^{1\\times C}$ .", + "bbox": [ + 511, + 513, + 883, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Typically, $\\mathbf{x}_i$ and $\\mathbf{z}_i$ are related in a one-to-one manner, i.e., $\\mathbf{x}_i$ represents a color vector at pixel $i$ in the input space and $\\mathbf{z}_i$ is its corresponding predicted label in the label space. However, the input and latent spaces are dimensionally heterogeneous, i.e., $N \\neq N'$ , which makes constructing a random field in this scenario challenging. Here, a novel distribution-approximated latent CRF (DAL-CRF) is proposed to eliminate the dimen", + "bbox": [ + 511, + 791, + 883, + 912 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20078", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sionally heterogeneous problem between the input and latent spaces. Specifically, by leveraging the fact that each latent vector $\\mathbf{z}_i$ corresponds to a receptive field in the input space with a set of intensity vectors, it is feasible to identify related intensity vectors for each latent vector, i.e.,", + "bbox": [ + 111, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {i} \\rightarrow \\mathbf {X} _ {i} = \\left\\{\\mathbf {x} _ {i} ^ {l} \\right\\} _ {l = 1} ^ {L}, i = 1, 2, \\dots , N ^ {\\prime}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 190, + 482, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L$ denotes the number of intensity vectors in the receptive field, where $L = 256$ corresponding to a $16\\times 16$ receptive field in MedSAM. Then, by following Veksler [24], the pairwise potential is only considered to be a regularization loss to encourage the minimization of CRF energy, which can be formulated as $\\mathcal{L}_{\\mathrm{DAL - CRF}} =$", + "bbox": [ + 111, + 218, + 483, + 309 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {1}{N ^ {\\prime}} \\sum_ {i, j \\in \\mathcal {G}} \\left(\\frac {1}{2} \\operatorname {M M D} \\left(\\mathbb {P} \\left(\\mathbf {X} _ {i}\\right), \\mathbb {P} \\left(\\mathbf {X} _ {j}\\right)\\right)\\right) \\cdot \\exp \\left(- \\| \\mathbf {z} _ {i} - \\mathbf {z} _ {j} \\| _ {F}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 316, + 482, + 366 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike existing CRF methods [24], which utilize the Gaussian kernel to evaluate the correlation between two color vectors, we introduce the Maximum Mean Discrepancy (MMD) to measure the distance between two vector sets. Specifically, each vector set is first approximated as an empirical distribution, i.e., $\\mathbf{x}_i^l\\sim \\mathbb{P}(\\mathbf{X}_i)$ . Then, each empirical distribution is represented by a kernel mean embedding in a reproducing kernel Hilbert space (RKHS) $\\mathcal{H}$ , i.e., $\\mu_{\\mathbb{P}} = \\mathbb{E}_{\\mathbf{x}\\sim \\mathbb{P}}[\\phi (\\mathbf{x})] = \\int_{\\mathcal{X}}k(\\mathbf{x},\\cdot)d\\mathbb{P} = \\mathbb{E}_{\\mathbf{x}\\sim \\mathbb{P}}[k(\\mathbf{x},\\cdot)]$ , where $k:\\mathcal{X}\\times \\mathcal{X}\\to \\mathbb{R}$ is a reproducing kernel (Gaussian RBF kernel here) and corresponding feature map $\\phi :\\mathcal{X}\\rightarrow \\mathcal{H}$ . The discrepancy can be measured between mean embeddings, i.e.,", + "bbox": [ + 111, + 367, + 483, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M M D} \\left(\\mathbb {P} \\left(\\mathbf {X} _ {i}\\right), \\mathbb {P} \\left(\\mathbf {X} _ {j}\\right)\\right) = \\| \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\phi \\left(\\mathbf {x} _ {i} ^ {l}\\right) - \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\phi \\left(\\mathbf {x} _ {j} ^ {l}\\right) \\| _ {\\mathcal {H}} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 585, + 496, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MMD enables all-order correlation measurement using Gaussian kernel via unbiased empirical estimation. This may be useful for better discrepancy measurement, as the distribution of the intensity space may be highly complex and non-linear. For the second term in Eq. (4), we utilize the Frobenius norm $\\| \\cdot \\| _F$ to calculate the feature compatibility in a co-occurrent probability using exponential zero-to-one normalization.", + "bbox": [ + 111, + 640, + 482, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Integrating Bounding Box Prompt. The graph $\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}$ in Eq. (4) is usually either a complete (dense-connected) graph [11] or a sparse position-related neighboring graph [24] on $\\mathbf{Z}$ . Instead, we propose to consider the bounding box prompt as the prior to construct $K$ -random pairs inside the given bounding box, which can not only exclude those unimportant latent features far away from the bounding box prompt to enhance the computational efficiency but also focus more on the interested region with better performance.", + "bbox": [ + 111, + 761, + 483, + 912 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intuitive Understanding. For Eq. (4), one actually encourages being more discriminative in latent representations between $\\mathbf{z}_i$ and $\\mathbf{z}_j$ . Since if two empirical distributions have a large distribution discrepancy (i.e., the MMD distance is large) in the input space, there are more penalties assigned to the pair who has larger similarity (i.e., co-occurrent probability is larger)) in the latent space. By minimizing $\\mathcal{L}_{\\mathrm{DAL - CRF}}$ , one can enforce the latent representations to equip with sufficient statistics as in the input space.", + "bbox": [ + 511, + 90, + 883, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall Objective at Test-time Adaptation", + "text_level": 1, + "bbox": [ + 511, + 251, + 870, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Without accessing the ground-truth label, it is infeasible to maximize the segmentation likelihood $\\log P(\\mathbf{Y}^m|\\mathbf{Z}^m;\\varphi)$ directly. Therefore, we adopt a common predictive entropy minimization to implicitly achieve this goal [8]. Formally,", + "bbox": [ + 511, + 273, + 883, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E M}} = H (P (\\mathbf {Y} | \\mathbf {Z})) = - \\sum_ {h, v \\in C} P (y ^ {h, v} | \\mathbf {Z}) \\log P (y ^ {h, v} | \\mathbf {Z}),\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 359, + 901, + 391 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\quad C = \\left\\{\\left(h, v\\right) \\mid P \\left(y ^ {h, v} \\mid \\mathbf {Z}\\right) \\geq \\alpha \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 395, + 883, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y^{h,v}$ denotes a single prediction in the position $(h,v)$ of $\\mathbf{Y}$ . Eq. (6) is similar to foreground-background balanced entropy loss in Dong et al. [6]. However, we only utilize the foreground entropy and set a high confidence threshold (e.g., $\\alpha = 0.95$ ) due to the following reasons. First, since the CRF-based loss usually requires a volumetric regularization to avoid the trivial solutions [24] (e.g., all labels are assigned to background), the entropy regularization over a high confidence foreground region can be interpreted as an implicit volumetric regularization, i.e., ensuring these foreground pixels confident enough for preventing trivial solutions. Second, we argue that background entropy is unnecessary for a foundation segmentation model, as these models have learned a good foreground-background balance.", + "bbox": [ + 511, + 425, + 883, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By considering the proposed distributionapproximated latent CRF loss and the customized entropy loss, the practical latent refinement scheme is reformulated as follows.", + "bbox": [ + 511, + 652, + 883, + 712 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} ^ {m} = \\mathbf {Z} ^ {m - 1} - \\epsilon \\cdot \\frac {\\partial \\left[ \\mathcal {L} _ {\\mathrm {D A L - C R F}} + \\lambda \\mathcal {L} _ {\\mathrm {E M}} \\right]}{\\partial \\mathbf {Z} ^ {m - 1}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 720, + 882, + 751 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a balance coefficient.", + "bbox": [ + 511, + 760, + 728, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 784, + 728, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall framework is illustrated in Figure 1. Superficially, the initial image embedding $\\mathbf{Z}^0$ is inferred only once. Once the refined latent representation $\\mathbf{Z}^*$ is obtained, the upsampling can be conducted directly without requiring image embedding again. This mechanism is well compatible with the interactive scheme of the MedSAM. We implement the proposed method based on", + "bbox": [ + 511, + 806, + 883, + 912 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20079", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/46724d48282b0d2b92c3026b1a7e1cf4060d70b11092ae966452024d939f9fd5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Update TypeMethodDice Coefficient ↑Average Surface Distance ↓
D1D2D3D4Avg.D1D2D3D4Avg.
Looseness promptsDirect Inference90.9987.1687.0684.7287.489.487.518.747.628.33
Normalization updateTENT (ICLR'21)91.2785.4085.0681.8485.899.098.7010.058.889.18
InTEnt (CVPR'24)91.3687.4387.4786.1387.608.807.518.267.668.06
GraTa (AAAI'25)90.8987.0086.9584.6487.379.487.548.807.658.37
PASS (TMI'24)80.8679.6382.0384.0281.6320.4012.2211.697.9713.07
Full updatesMEMO (NeurIPS'22)87.2685.4886.4284.4785.9112.458.279.157.769.40
CRF-SOD (CVPR'23)89.5784.0184.8082.5885.2410.548.9910.188.649.58
Latent updatesOurs89.3588.7389.8388.1889.0211.406.696.926.077.77
Shrinkage promptsDirect Inference81.0567.6072.1068.5772.3318.5117.6117.7614.9217.20
Normalization updateTENT (ICLR'21)80.2265.1769.2562.1069.1818.8519.2619.3917.5818.77
InTEnt (CVPR'24)82.7270.0674.5073.4975.1916.8615.2816.1313.8715.53
GraTa (AAAI'25)80.8767.3271.9468.4672.1517.2716.5416.5314.5216.21
PASS (TMI'24)66.8158.6663.3156.7861.3932.3922.8922.7420.1924.55
Full updatesMEMO (NeurIPS'22)81.0267.5472.1068.8472.3718.5217.7617.7714.7817.20
CRF-SOD (CVPR'23)81.8469.2674.7372.0874.4819.2718.8318.6314.7317.86
Latent updatesOurs83.0773.7278.5475.7777.7716.7414.3813.7211.2614.02
Perfectness promptsDirect Inference87.5880.2581.3278.4681.9012.7611.2112.2710.2411.62
Normalization updateTENT (ICLR'21)87.6077.9878.1872.9979.1812.8312.2113.9212.6912.91
InTEnt (CVPR'24)86.5484.2585.1781.6584.4010.989.239.858.959.75
GraTa (AAAI'25)87.3979.9781.1478.3681.7112.9611.3812.3810.2911.75
PASS (TMI'24)74.5969.8673.8871.3472.4125.5317.3216.5813.8518.32
Full updatesMEMO (NeurIPS'22)87.3980.1881.3278.4981.8412.9711.2412.2710.2311.67
CRF-SOD (CVPR'23)86.0777.2279.1076.2479.6513.9612.5913.6311.212.84
Latent updatesOurs86.8585.1186.7284.4985.7913.738.638.797.489.65
", + "bbox": [ + 124, + 89, + 877, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison of TTA results on fundus image OC segmentation between different methods. The best and second-best results are shown in the bold and the underline, respectively. For loose prompts, we report the average of two looseness ratios. For shrinkage prompts, we report the average of two shrinkage ratios. Full results can be found in the Appendix.", + "bbox": [ + 111, + 430, + 883, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "official MedSAM checkpoints (ViT-B pre-trained models) and codes using a one-step optimization strategy $(m = 1)$ by following previous TTA settings. Due to the extremely various sizes of RoIs in medical images, we leverage the initial segmentation result $\\mathbf{Y}^0$ and the given bounding box prompt to explore a dynamic adjustment strategy for important hyperparameters. Specifically, the learning rate $\\epsilon$ with the Adam optimizer is linearly interpolated from $5\\times 10^{-2}$ to $1\\times 10^{-1}$ according to the ratio of the initial foreground to the entire image $f_{e}$ (i.e., the rough size of the RoI), where $f_{e}\\leq 0.05$ and $f_{e}\\geq 0.2$ are fixedly set to $5\\times 10^{-2}$ and $1\\times 10^{-1}$ , respectively. Note that we use commonly-used definitions over the $f_{e}$ as adopted by Gao et al. [7], i.e., $f_{e}\\leq 0.05$ , $0.05 < f_{e} < 0.2$ , and $f_{e}\\geq 0.2$ as small, medium, and large RoIs. Similarly, the $\\lambda$ is also RoI-specific adjustment, where the small RoI $(f_{e}\\leq 0.05)$ requires a similar contribution between the $\\mathcal{L}_{\\mathrm{DAL - CRF}}$ and the $\\mathcal{L}_{\\mathrm{EM}}$ , thus the $\\lambda$ is set to $1\\times 10^{3}$ . The $K$ is set to 4 by balancing computational consumption and performance (refers to Table 5(a)). $\\alpha$ in entropy loss is set to 0.95 empirically.", + "bbox": [ + 111, + 484, + 483, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 112, + 823, + 243, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics We evaluate our method on two public multi-center medical image segmentation datasets with different data modalities. (1) Fundus [26]: Fundus datasets are retinal fundus images", + "bbox": [ + 111, + 851, + 483, + 912 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "collected from 4 healthcare centers (D1 to D4 for short). For each retinal fundus image, there are two segmentation objectives, including the optic cup (OC) (mainly small RoIs) and optic disc (OD) (mainly medium or large RoIs). (2) Spinal Cord Gray Matter (SCGM) [12]: SCGM datasets are collected from 4 institutions (D1 to D4 for short) with different imaging manufacturers for spinal MRI image gray matter segmentation. We use the Dice Similarity Coefficient (DSC) and Average Surface Distance (ASD) to measure the accuracy.", + "bbox": [ + 511, + 484, + 883, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment Settings. Appearance Perturbations: Note that our adopted multi-center datasets have strong domain shifts with appearance variations, we thus ignore test-time appearance perturbations (e.g., changes in contrast, noise, and blur). Prompt Perturbations: We mainly consider three kinds of bounding box prompts to evaluate TTA performance. (i) Perfectness prompt $(PP)$ : The perfect prompt can be obtained by groundtruth (GT) mask, where the top-left and lower-right coordinates of the PP correspond to the minimal and maximal horizontal and vertical coordinates of the GT mask, respectively. A unique PP is utilized for TTA. (ii) Looseness prompts $(LP)$ : Clinically, it is labor-intensive and time-consuming to obtain the PP. Instead, a moderately loose prompt can be efficient but may trigger the degradation of the model with the need for TTA. Therefore, we define the looseness rate $\\sigma$ , which is the ratio of the coordinate shift relative to the width $w$ and", + "bbox": [ + 511, + 640, + 883, + 911 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20080", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0007e14b6606369f862dca70c689fb83c814d3ca09c4d3e0be7cb7cd54b1e975.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Update TypeMethodDice Coefficient ↑Average Surface Distance ↓
D1D2D3D4Avg.D1D2D3D4Avg.
Looseness promptsDirect Inference83.0578.8681.8275.5479.8133.2032.8730.2931.9032.06
Normalization updateTENT (ICLR'21)81.6176.9680.0074.8578.3537.5836.5433.5334.3535.50
InTEnt (CVPR'24)83.6176.7180.0071.8378.0332.4637.0132.5739.1435.29
GraTa (AAAI'25)84.6480.0484.1082.8282.9030.0230.9325.9128.9128.94
PASS (TMI'24)84.5781.7782.2977.3481.4918.0226.2021.0924.6822.49
Full model updatesMEMO (NeurIPS'22)86.2680.0183.9478.1882.1026.7330.9326.2728.5428.12
CRF-SOD (CVPR'23)87.0382.8683.3379.5283.1825.1926.2523.7226.5325.42
Latent updatesOurs90.8583.5587.7782.2486.1017.5923.5418.5021.7520.34
Shrinkage promptsDirect Inference92.7086.1790.9088.6789.6112.6017.6612.7712.1013.78
Normalization updateTENT (ICLR'21)90.6385.6389.6686.8888.2015.9018.4414.2712.4815.27
InTEnt (CVPR'24)90.9187.2790.8888.7089.4415.4916.4712.6812.0514.17
GraTa (AAAI'25)92.6185.6091.0088.7189.4812.7018.3212.6112.1113.93
PASS (TMI'24)91.4678.8585.9275.7082.9813.7225.1417.5423.0519.86
Full model updatesMEMO (NeurIPS'22)92.1283.8290.8788.7788.8913.5220.1812.7512.0414.62
CRF-SOD (CVPR'23)94.4284.0792.3491.4390.569.3619.8310.709.3012.30
Latent updatesOurs94.2486.7793.2392.4291.669.8016.259.458.2110.93
Perfectness promptsDirect Inference96.4992.8793.8992.3693.906.099.488.648.368.14
Normalization updateTENT (ICLR'21)94.2090.7893.3990.3892.1810.1212.019.3210.3910.46
InTEnt (CVPR'24)94.1291.5794.4293.4793.3910.1211.018.017.039.04
GraTa (AAAI'25)94.4692.6693.9192.3193.336.149.738.618.428.22
PASS (TMI'24)94.9085.3584.3686.1487.688.5817.5217.6213.9814.42
Full model updatesMEMO (NeurIPS'22)95.7791.4893.8992.3193.367.8011.288.618.419.02
CRF-SOD (CVPR'23)96.6691.1593.7693.4193.755.7811.428.737.268.29
Latent updatesOurs96.6892.6795.1594.5794.765.699.516.855.987.00
", + "bbox": [ + 124, + 89, + 877, + 417 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dcd1c88bd483377793d6fb8aebc07cb7c8fe25e5aea6e45a3f2ac6a39ec039c4.jpg", + "table_caption": [ + "Table 3. Quantitative comparison of TTA results on fundus image OD segmentation between different methods. The best and second-best results are shown in the bold and the underline, respectively. For loose prompts, we report the average of two looseness ratios. For shrinkage prompts, we report the average of two shrinkage ratios. Full results can be found in the Appendix." + ], + "table_footnote": [], + "table_body": "
Update TypeMethodDice Coefficient ↑Average Surface Distance ↓
D1D2D3D4Avg.D1D2D3D4Avg.
Direct Inference55.3554.2061.5456.6756.9410.2211.3915.2715.9513.20
Normalization updateTENT (ICLR'21)49.6652.9659.3955.3854.349.9410.4315.8516.1213.08
InTEnt (CVPR'24)50.9152.1245.8550.3849.819.3510.9922.5719.0115.48
GraTa (AAAI'25)56.3555.1463.6658.1258.319.2610.4414.3615.1512.30
PASS (TMI'24)55.2454.8061.9452.8856.2110.5811.0115.0220.1413.99
Full updatesMEMO (NeurIPS'22)56.2055.2061.2057.3457.489.2710.4514.8915.3912.50
CRF-SOD (CVPR'23)54.2054.0861.8957.0756.8110.3710.8315.3615.5813.03
Latent updatesOurs57.9255.4266.0058.6259.498.3510.5814.1414.1511.75
", + "bbox": [ + 124, + 474, + 877, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Quantitative comparison of TTA results on MRI image SCGM segmentation between different methods. SCGM segmentation only reported PP-based results due to its very challenging intricate structures and appearance perturbations.", + "bbox": [ + 111, + 617, + 883, + 646 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "height $h$ of the PP, to simulate real-world loose scenarios as follows: $x(y)_{tl}^{LP}: x(y)_{tl} - \\sigma \\cdot w(h), x(y)_{lr}^{LP}: x(y)_{lr} + \\sigma \\cdot w(h)$ , where $x(y)_{tl}$ and $x(y)_{lr}$ denote the vertical (horizontal) coordinates of the top-left and lower-right points of the PP. Each loose prompt is represented as $(x_{tl}^{LP}, y_{tl}^{LP}, x_{lr}^{LP}, y_{lr}^{LP})$ . Here, $\\sigma$ is set to 0.1 and 0.2, as we observe 1) too small $\\sigma (< 0.1)$ is close to the PP without obvious changes and 2) too large $\\sigma (>0.2)$ will usually result in significantly overlaps with other RoIs with unacceptable initial results. We report the average of different loose rates for comprehensive evaluation. (iii) Shrinkage prompts (SP): Similarly, we define the shrinkage ratio $\\gamma$ to simulate the scenario that the bounding box prompt shifts to the interior of the PP as follows: $x(y)_{tl}^{SP}: x(y)_{tl} + \\gamma \\cdot w(h), x(y)_{lr}^{SP}: x(y)_{lr} - \\gamma \\cdot w(h)$ . Each shrinkage prompt is represented", + "bbox": [ + 111, + 661, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as $(x_{tl}^{SP},y_{tl}^{SP},x_{lr}^{SP},y_{lr}^{SP})$ . Similarly, $\\gamma$ is set to 0.1 and 0.2. We report the average of different shrinkage rates for comprehensive evaluation.", + "bbox": [ + 511, + 660, + 883, + 707 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. In this study, two kinds of TTA approaches for medical image segmentation are utilized for comparison. (1) Normalization-based TTA: Most baseline methods conduct the update of affine transformation parameters and manipulation of source-trained statistics in BN layers, which the MedSAM lacks due to the usage of LN layers. We, therefore, discard source-trained statistics-related components and only optimize the affine-transformation parameters for baseline methods, including TENT (entropy minimization), InTEnt (foreground-background-balanced entropy weighting), GraTa (augmentation-based gradient alignment), and PASS (input decora", + "bbox": [ + 509, + 715, + 883, + 912 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20081", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8335dfc2ddad3f2cf9acaacdc3c84db4ed78b5f2dbab02b8e0637480f57b225b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 87, + 215, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b8854cfa7bdad07e0cbc26f9997a562d607cbcd164d539630d8eb048a3d29057.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 217, + 89, + 307, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48cc3745ba6e2555f858576859fe914657a51126e49bd64a3eb8271db1be7070.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 90, + 401, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aac781c46819eefe84498bfe43ad6bc4575f42090109af93f3e13f29b658ea43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 90, + 496, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f2b83411b3b336fc2918554b16c1b2164441e88e422fdc837eafe03cb34d31ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 90, + 594, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2056b3dc42a43ba70e316950ad6da654ecf0490c1dcf1c7e1995abd38ad1a7c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 152, + 217, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b3fa64e6721c9ecdfe38623fb6d45b7c6f2a0c3497ab809793946259341c1a6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 152, + 310, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a329bc565350d7c16c811867415bc5f3c649dfb2f25ea812c4e3cfb40418fab6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 152, + 403, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/454e1cd0268541b851d7ef1cc044798c0cc936cefcadc7e2fa52e6e587ec0b18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 152, + 496, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/177140834aa9cfb310e60a7642480a327e71a0422358f2adbc12ef36337b3ebf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 152, + 606, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/410503a618ac85d9ce271c839342f438d16d31253fff6c29ee0b59dda936fc30.jpg", + "image_caption": [ + "Figure 2. Visualized process of latent refinement with iterations. The first, second, and third rows denote the segmentation result, corresponding visualization of channel-average refined latent embedding (which is from the region in bounding box prompt and has been fused with the prompt for a better view), and corresponding DAL-CRF loss for each spatial position. You may zoom in for a better view." + ], + "image_footnote": [], + "bbox": [ + 116, + 214, + 218, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5eb53ed1ccbaf88dc2a92044519d2d6cb845941a4cae5de354b426b84914977c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 220, + 214, + 310, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c99c923fc7a2dbda4ce6d619541d9cd9be0e5a1497ef1f5d31adc1ef218b523.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 214, + 401, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5913bd22dec9afd75d38b5010f23e210b39c4b79b82d02a82f9a01f979e6a487.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 403, + 214, + 496, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9315e7c096c20f395efbab305ec890adf41a33bd637ba0792d755cc1bfde0acf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 214, + 606, + 279 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/897899b585be90332e5d615c9f606212631fc5d15f3ea3f2bc5c8dbcf2514810.jpg", + "table_caption": [ + "(a)" + ], + "table_footnote": [], + "table_body": "
KD1D2D3D4Avg.
293.2487.0692.5186.5189.75
494.1687.1691.8387.5790.14
694.6787.0892.3487.9290.37
894.8686.5490.8188.0190.03
", + "bbox": [ + 616, + 90, + 880, + 155 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/319d69781d59d710475f75ecbf83b1bd8eb56d807467904e59386a0c1823c8ec.jpg", + "table_caption": [ + "(b)" + ], + "table_footnote": [], + "table_body": "
LossD1D2D3D4Avg.
DI89.1984.3187.9282.3985.95
\\( \\mathcal{L}_{\\text{InTEnt}} \\)79.3484.8988.8386.0884.78
\\( \\mathcal{L}_{\\text{TENT}} \\)85.0984.9989.4685.5986.28
\\( \\mathcal{L}_{\\text{CRF-SOD}} \\)88.9386.1788.4287.3387.71
\\( \\mathcal{L}_{\\text{DAL-CRF}} \\)94.1686.9690.8587.5789.88
", + "bbox": [ + 616, + 171, + 880, + 252 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/79b39651b916e9ac2b8f1068b8c75122345b4f615646509e9c94e64dbd123660.jpg", + "table_caption": [ + "Table 5. (a) Hyperparameter Analyses in terms of $K$ . (b) Integration of the latent refinement framework and other TTA loss functions $\\nu$ s. Our proposed DAL-CRF, where DI denotes direct inference without adaptation on OD segmentation." + ], + "table_footnote": [], + "table_body": "
MethodD1D2D3D4Avg.
Direct Inference89.1984.3187.9282.3985.95
+ LEM85.8285.3188.1083.2186.01
+ LDAL-CRF94.1686.9690.8587.5789.88
+ LEM + LDAL-CRF94.1687.1691.8387.4290.14
DenseCRF [11]-Postprocessing82.7079.3980.9073.7679.18
", + "bbox": [ + 116, + 361, + 478, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation Study on Different Components of the OD segmentation with loose prompts, in terms of Dice score.", + "bbox": [ + 111, + 440, + 480, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tor and alternating momentum updating strategy). (2) Full update-based TTA: The whole model parameters are updated based on customized losses, including MEMO (augmentation-based marginal entropy minimization) and CRF-SOD (sparse CRF regularization loss). We leverage their open-source codes to apply into the architecture of MedSAM. To be fair, all baseline methods adopt a one-step optimization strategy.", + "bbox": [ + 111, + 481, + 482, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Analysis. From Tables 2, 3, and 4, some conclusions can be summarized. (1) Our proposed method without parametric updates surpasses all baselines in most cases with the best or second-best performance, showcasing the effectiveness of latent refinement. (2) It seems that full parameter updates are more suitable to large size RoIs than normalization-based updates, e.g., the OD. (3) For very challenging OC and SCGM tasks, most parametric update-based approaches have compromised performance compared with direct inference, but our proposed method performs well.", + "bbox": [ + 111, + 606, + 482, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Analysis. From Figure 2, our proposed method can achieve gradually better TTA segmentation with iterations (Note: we reduce the learning rate with multistep optimization rather than a one-step manner for understanding the optimization process better). Meanwhile, the latent embedding is gradually unsmoothed, which may be reasonable as the DAL-CRF loss enforces more statistics over the input image, leading to more discriminative information. The gradual decrease of DAL-", + "bbox": [ + 111, + 776, + 482, + 911 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CRF loss in each spatial point of the latent embedding is also observed, showcasing its effectiveness.", + "bbox": [ + 511, + 364, + 882, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computation Complexity. As illustrated in Table 1, we achieve around $7 \\times$ GFLOPs reduction compared with common usages. This is reasonable as 1) sophisticated augmentations (e.g., MEMO) and multiple gradient backward processes (e.g., GraTa) are computationally intensive. 2) For normalization-based updates, the gradient computation over normalization layers still passes through the entire model during backpropagation, which is unfavorable as MedSAM's image encoder is sophisticated with considerable consumption. 3) We circumvent any augmentation and the forward/backward process to the image encoder in a more efficient manner.", + "bbox": [ + 511, + 396, + 883, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study & Other TTA losses under the latent refinement framework. As shown in Table 6, $\\mathcal{L}_{\\mathrm{DAL - CRF}}$ significantly contributes more compared to entropy minimization loss and simply post-processing predictions fail to achieve acceptable performance on challenging medical images with strong appearance perturbations. Meanwhile, $\\mathcal{L}_{\\mathrm{DAL - CRF}}$ still performs better, even though other TTA losses are integrated into the latent refinement framework as shown in 5(b). Compared to the CRF-SOD (using input and output spaces), our DAL-CRF exhibits superiority using the latent-based co-occurrence regularization for foundation models.", + "bbox": [ + 511, + 578, + 883, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 763, + 632, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel TTA framework for foundation medical segmentation without parametric updates. Our theoretical analysis shows that direct latent refinement is feasible within the MedSAM architecture. By maximizing factorized conditional probabilities by our DAL-CRF loss combined with entropy minimization, we achieve better performance for effectiveness and efficiency.", + "bbox": [ + 511, + 791, + 883, + 912 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20082", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 112, + 90, + 272, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the Hong Kong Innovation and Technology Commission (ITC) (InnoHK Project CIMDA), in part by the Institute of Digital Medicine of City University of Hong Kong (CityU) under Project 9229503, in part by Chow Sang Sang Donation and Matching Fund (Project 9229161), and in part by the CityU under Project 7006087.", + "bbox": [ + 111, + 114, + 483, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 114, + 243, + 209, + 258 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kecheng Chen, Pingping Zhang, Tiexin Qin, Shiqi Wang, Hong Yan, and Haoliang Li. Test-time adaptation for image compression with distribution regularization. In The Thirteenth International Conference on Learning Representations, 2025. 3", + "[2] Ziyang Chen, Yiwen Ye, Yongsheng Pan, and Yong Xia. Gradient alignment improves test-time adaptation for medical image segmentation. The 39th Annual AAAI Conference on Artificial Intelligence, 2025. 2, 4", + "[3] Mohammad Zalbagi Darestani, Jiayu Liu, and Reinhard Heckel. Test-time training can close the natural distribution shift performance gap in deep learning based compressed sensing. In International Conference on Machine Learning, pages 4754-4776. PMLR, 2022. 2", + "[4] Zeshuai Deng, Zhuokun Chen, Shuaicheng Niu, Thomas Li, Bohan Zhuang, and Mingkui Tan. Efficient test-time adaptation for super-resolution with second-order degradation and reconstruction. Advances in Neural Information Processing Systems, 36:74671-74701, 2023. 2", + "[5] JCMSA Djelouah and Christopher Schroers. Content adaptive optimization for neural image compression. In Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit, pages 1-5, 2019. 3", + "[6] Haoyu Dong, Nicholas Konz, Hanxue Gu, and Maciej A Mazurowski. Medical image segmentation with intent: Integrated entropy weighting for single image test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5046-5055, 2024. 1, 2, 5", + "[7] Yunhe Gao, Rui Huang, Yiwei Yang, Jie Zhang, Kainan Shao, Changjuan Tao, Yuanyuan Chen, Dimitris N. Metaxas, Hongsheng Li, and Ming Chen. Focusnetv2: Imbalanced large and small organ segmentation with adversarial shape constraint for head and neck ct images. Medical Image Analysis, 67:101831, 2021. 6", + "[8] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 5", + "[9] Tianyu Huang, Tao Zhou, Weidi Xie, Shuo Wang, Qi Dou, and Yizhe Zhang. Improving segment anything on the fly: Auxiliary online learning and adaptive fusion for medical image segmentation. arXiv preprint arXiv:2406.00956, 2024.3", + "[10] Yiming Ju, Ziyi Ni, Xingrun Xing, Zhixiong Zeng, Hanyu Zhao, Siqi Fan, and Zheng Zhang. Mitigating" + ], + "bbox": [ + 114, + 267, + 483, + 912 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "training imbalance in LLM fine-tuning via selective parameter merging. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15952-15959, Miami, Florida, USA, 2024. Association for Computational Linguistics. 2", + "[11] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 4, 5, 8", + "[12] Haoliang Li, YuFei Wang, Renjie Wan, Shiqi Wang, TieQiang Li, and Alex Kot. Domain generalization for medical imaging classification with linear-dependency regularization. Advances in Neural Information Processing Systems, 33:3118-3129, 2020. 6", + "[13] Zhenbing Liu, Fengfeng Wu, Yumeng Wang, Mengyu Yang, and Xipeng Pan. Fedcl: Federated contrastive learning for multi-center medical image classification. Pattern Recognition, 143:109739, 2023. 2", + "[14] Yue Lv, Jinxi Xiang, Jun Zhang, Wenming Yang, Xiao Han, and Wei Yang. Dynamic low-rank instance adaptation for universal neural image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 632-642, 2023. 3", + "[15] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 3", + "[16] Yuchen Mao, Hongwei Bran Li, LAI Yinyi, Giorgos Papanastasiou, Peng Qi, Yunjie Yang, and Chengjia Wang. Semi-supervised medical image segmentation via knowledge mining from large models. 1", + "[17] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 2", + "[18] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3", + "[19] Subhadeep Roy, Shankhanil Mitra, Soma Biswas, and Rajiv Soundararajan. Test time adaptation for blind image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16742-16751, 2023. 2", + "[20] Robin Schon, Julian Lorenz, Katja Ludwig, and Rainer Lienhart. Adapting the segment anything model during usage in novel situations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3616-3626, 2024. 3", + "[21] Sheng Shen, Huanjing Yue, and Jingyu Yang. Decap dapter: Exploring efficient decoder-side adapter for bridging screen content and natural image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12887-12896, 2023. 3", + "[22] Charles Sutton, Andrew McCallum, et al. An introduction to conditional random fields. Foundations and Trends® in Machine Learning, 4(4):267-373, 2012. 4" + ], + "bbox": [ + 516, + 92, + 883, + 912 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20083", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Koki Tsubota, Hiroaki Akutsu, and Kiyoharu Aizawa. Universal deep image compression via content-adaptive optimization with adapters. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2529-2538, 2023. 3", + "[24] Olga Veksler. Test time adaptation with regularized loss for weakly supervised salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7360-7369, 2023. 1, 5", + "[25] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 1, 2", + "[26] Shujun Wang, Lequan Yu, Kang Li, Xin Yang, Chi-Wing Fu, and Pheng-Ann Heng. Dofe: Domain-oriented feature embedding for generalizable fundus image segmentation on unseen datasets. IEEE Transactions on Medical Imaging, 39(12):4237-4248, 2020. 2, 6", + "[27] Ruxue Wen, Hangjie Yuan, Dong Ni, Wenbo Xiao, and Yaoyao Wu. From denoising training to test-time adaptation: Enhancing domain generalization for medical image segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 464-474, 2024. 2", + "[28] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 1", + "[29] Guoping Xu, Xiaoxue Qian, Hua Chieh Shao, Jax Luo, Weiguo Lu, and You Zhang. A sam-guided and match-based semi-supervised segmentation framework for medical imaging. arXiv preprint arXiv:2411.16949, 2024. 1", + "[30] Chuyan Zhang, Hao Zheng, Xin You, Yefeng Zheng, and Yun Gu. Pass: test-time prompting to adapt styles and semantic shapes in medical image segmentation. IEEE Transactions on Medical Imaging, 2024. 1, 2", + "[31] Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Yejia Zhang, and Danny Z Chen. Input augmentation with sam: Boosting medical image segmentation with segmentation foundation model. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2023. 1", + "[32] Chen, Ziyang and Ye, Yiwen and Pan, Yongsheng and Xia, Yong. Gradient Alignment Improves Test-Time Adaptation for Medical Image Segmentation. In AAAI, 2025. 1", + "[33] Zhang, Marvin and Levine, Sergey and Finn, Chelsea. Memo: Test time robustness via adaptation and augmentation Segmentation. In Advances in neural information processing systems, pages 38629-38642, 2022. 1" + ], + "bbox": [ + 114, + 90, + 480, + 824 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20084", + "bbox": [ + 478, + 950, + 517, + 962 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/8a26d38c-cedb-499b-b41f-c629d5b94654_model.json b/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/8a26d38c-cedb-499b-b41f-c629d5b94654_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b82aa34ce0ad74f789cef5c1cb4b2728792c3186 --- /dev/null +++ b/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/8a26d38c-cedb-499b-b41f-c629d5b94654_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.13, + 0.877, + 0.176 + ], + "angle": 0, + "content": "Test-time Adaptation for Foundation Medical Segmentation Model without Parametric Updates" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.203, + 0.751, + 0.221 + ], + "angle": 0, + "content": "Kecheng Chen\\(^{1}\\), Xinyu Luo\\(^{1}\\), Tiexin Qin\\(^{1}\\), Jie Liu\\(^{1}\\), Hui Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.221, + 0.697, + 0.239 + ], + "angle": 0, + "content": "Victor Ho Fun Lee², Hong Yan¹, Haoliang Li¹*" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.239, + 0.874, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)City University of Hong Kong \\(^{2}\\)HKU {ck.ee@my.cityu.edu.hk} \\(\\star\\): Corresponding author" + }, + { + "type": "title", + "bbox": [ + 0.26, + 0.299, + 0.337, + 0.315 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.33, + 0.486, + 0.754 + ], + "angle": 0, + "content": "Foundation medical segmentation models, with MedSAM being the most popular, have achieved promising performance across organs and lesions. However, MedSAM still suffers from compromised performance on specific lesions with intricate structures and appearance, as well as bounding box prompt-induced perturbations. Although current test-time adaptation (TTA) methods for medical image segmentation may tackle this issue, partial (e.g., batch normalization) or whole parametric updates restrict their effectiveness due to limited update signals or catastrophic forgetting in large models. Meanwhile, these approaches ignore the computational complexity during adaptation, which is particularly significant for modern foundation models. To this end, our theoretical analyses reveal that directly refining image embeddings is feasible to approach the same goal as parametric updates under the MedSAM architecture, which enables us to realize high computational efficiency and segmentation performance without the risk of catastrophic forgetting. Under this framework, we propose to encourage maximizing factorized conditional probabilities of the posterior prediction probability using a proposed distribution-approximated latent conditional random field loss combined with an entropy minimization loss. Experiments show that we achieve about \\(3\\%\\) Dice score improvements across three datasets while reducing computational complexity by over 7 times. Codes are available at here." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.781, + 0.245, + 0.797 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.807, + 0.485, + 0.913 + ], + "angle": 0, + "content": "Recently, foundation segmentation models for medical images have attracted massive attention from the medical imaging community [31], as medical image segmentation plays an indispensable role in many downstream tasks (e.g., surgical navigation and treatment planning). Medical Segment Anything Model [15] (MedSAM) is the most popular foundation segmentation model, which" + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.298, + 0.892, + 0.402 + ], + "angle": 0, + "content": "
Update TypeMethodDatasetsGFLOPs↓
OCODSCGM
Direct Inference87.7780.5756.67-
NormalizationTENT [25] (ICLR'21)86.2478.0854.34743.5
InTEnt [6] (CVPR'24)86.9582.4049.81744.8
GraTa [32] (AAAI'25)88.5780.4155.142975.6
PASS [30] (TMI'24)84.0571.8154.80745.8
FullMEMO [33] (NeurIPS'22)88.1280.0457.485646.5
CRF-SOD [24] (CVPR'23)88.9579.7956.811115.2
LatentOurs90.8484.1959.4997.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.412, + 0.885, + 0.538 + ], + "angle": 0, + "content": "Table 1. Average Dice coefficient \\((\\uparrow)\\) of different TTA methods on three datasets. Normalization-based methods are customized to BN layers, exhibiting marginal (even degraded) performance for LN-based foundation medical segmentation model. We report the computational complexity using GFLOPs (Giga Floating Point Operations), including forward and backward propagations. We leverage latent updates to achieve \\(\\sim 7\\times\\) GFLOPs reduction compared with parametric update-based counterparts with performance gains." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.548, + 0.884, + 0.67 + ], + "angle": 0, + "content": "exhibits impressive universality across organs and lesions. Since MedSAM is pre-trained on large-scale medical image datasets, it typically enables the high-quality segmentation mask of the region of interest (RoI) in a zero-shot manner, which has contributed to extensive areas, e.g., semi/weak-supervised medical image segmentation [29] or anomaly detection [16] using provided segmentation masks as pseudo labels or RoIs." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.671, + 0.884, + 0.913 + ], + "angle": 0, + "content": "However, due to the extreme modality imbalance of medical images (where CT, MRI, and endoscopy images are dominant), MedSAM has compromised performance on less-represented modalities (e.g., eye fundus images) [15]. It is also difficult for MedSAM to maneuver specific lesions with intricate structures in the segmentation area [15], e.g., optic cup segmentation suffers from many vessel-like branching structures. Although supervised fine-tuning on specific modalities and lesions with annotated masks seems effective in tackling these issues [28], the overwhelming computational consumption of fine-tuning and inherent data scarcity (or privacy) concerns regarding annotated medical images may impede this process. Moreover, some test-time perturbations are still encountered, even though a well-calibrated and fine-tuned model has been prepared. For instance," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.964 + ], + "angle": 0, + "content": "20075" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "mild looseness or shrinkage perturbations for bounding box prompts are sufficient to cause performance degradation for MedSAM (see Tables 2 and 3), which is clinically unfavorable as delineating a perfect prompt is usually labor-intensive and time-consuming for physicians." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.169, + 0.483, + 0.576 + ], + "angle": 0, + "content": "To address the above-mentioned issues, test-time adaptation (TTA) for foundation medical segmentation models is a feasible solution that enables performance gains on a specific unlabeled target (test) domain using the pre-trained model only. Although recent TTA approaches for medical image segmentation achieve promising results, we argue that scaling them to foundation segmentation models is difficult. First, most methods utilize various self-supervised objectives (e.g., prediction consistency over different augmentations [2, 6, 30]) to conduct batch normalization (BN)-centric fin-tuning, which may endure limited update signals with marginal adaptation as modern foundation models employ layer normalization (LN) without source-trained statistics like BN. Second, some approaches [27] update the whole model to explore more gains, but even minimal parameter adjustments in large models can precipitate catastrophic forgetting [10]. Third, most methods ignore computational complexity during the adaptation, which is particularly significant for foundation medical segmentation models1, since continuous forward and backward propagations on a large model are computationally intensive, especially on resource-constrained edge devices. Overall, existing methods struggle with maintaining a trade-off between sufficient update signals and knowledge preservation, and ignore computational consumption of the adaptation for large models." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.578, + 0.483, + 0.849 + ], + "angle": 0, + "content": "To this end, we propose a novel TTA framework for foundation medical segmentation models like MedSAM. Compared to existing methods, one of the most pivotal differences is that our proposed method updates latent representations rather than any model parameters², which can not only endow maximal update flexibility under a forgetting-free context but also circumvent heavy forward and backward computations on sophisticated architectures (e.g., image encoder). Specifically, by incorporating an off-the-shelf conditionally independent assumption between the input image X and the prediction Y in MedSAM, we provide a theoretical analysis to uncover that directly refining the image embedding Z (a.k.a. latent refinement) can encourage maximizing factorized conditional probabilities (including latent fidelity \\(P(\\mathbf{Z}|\\mathbf{X})\\) and segmentation likelihood \\(P(\\mathbf{Y}|\\mathbf{Z}))\\) of posterior prediction probability \\(P(\\mathbf{Y}|\\mathbf{X})\\), in a similar objective like parametric updates. Under this latent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.885, + 0.242 + ], + "angle": 0, + "content": "refinement framework, we devise a novel distributionapproximated latent conditional random field (CRF) loss to leverage useful cues in the input space for explicitly maximizing the posterior distribution of the latent fidelity. Meanwhile, we utilize entropy minimization to enforce the usability of the final predictions for implicitly maximizing segmentation likelihood. Finally, the mask decoder can directly impose the refined latent without inferring the image embedding again. Our contributions include three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.244, + 0.884, + 0.334 + ], + "angle": 0, + "content": "- We provide theoretical analyses to uncover that direct latent refinement is feasible to approach the same goal as parametric updates under the MedSAM architecture, which enables us to realize high computational efficiency and segmentation performance (refers to Table 1) without the risk of catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.334, + 0.882, + 0.409 + ], + "angle": 0, + "content": "- A distribution-approximated latent CRF loss based on a Maximum Mean Discrepancy-aware co-occurrent measure is proposed to introduce potential useful cues from input space, leading to sufficient statistics of image embedding over the input image." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.41, + 0.882, + 0.53 + ], + "angle": 0, + "content": "- The latent refinement framework is validated on three multi-center medical image segmentation datasets with strong appearance variations and simulated prompt shifts. Extensive experiments demonstrate that our method can not only achieve average \\(\\sim 3\\%\\)\n- Dice score improvements among three datasets but also enjoy around \\(7\\times\\) computational complexity reduction compared with existing methods." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.244, + 0.884, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.547, + 0.66, + 0.563 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.851, + 0.588 + ], + "angle": 0, + "content": "2.1. Medical Image Segmentation with TTA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.885, + 0.913 + ], + "angle": 0, + "content": "Various test-time adaptation (TTA) methodologies have been developed to address distribution shifts during inference [3, 17, 25]. Recent advancements have extended TTA frameworks to more complex tasks, including super-resolution [4] and blind image quality assessment [19]. In the field of medical image segmentation, TTA tackles the common multi-center issue [13], where different hospitals may have specific imaging parameters with strong appearance variations [26]. To this end, some methods conducted the optimization of the BN layer, manipulating either the affine-transformation parameters [30] or source-trained BN statistics [6], where many self-supervised objectives and sophisticated alignment of source-trained statistics in the BN layer are leveraged. Meanwhile, some approaches [27] fine-tune the overall model to explore more gains using elaborated objectives and knowledge accumulation strategies. We argue that such parametric updates suffer from restricted effectiveness due to limited BN-centric update signals and catastrophic forgetting for a large pretrained model, respectively. The computational com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.862, + 0.483, + 0.912 + ], + "angle": 0, + "content": "1MedSAM obtains image embedding using an intricate ViT-based encoder (accounts for \\(98.5\\%\\) computations) and a lightweight decoder. 2We refer to partial or whole model updates as parametric updates hereafter." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.52, + 0.963 + ], + "angle": 0, + "content": "20076" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.111, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "plexity is also ignored during adaptation. There are limited works specific to the TTA of foundation medical segmentation models. For example, Huang et al. [9] proposed an auxiliary online learning (AuxOL) and adaptive fusion for the foundation medical image segmentation model. However, AuxOL requires clinicians' annotations for model training, leading to potential unfeasibility in many scenarios. Beyond the medical image, Schön et al. [20] leveraged the clicked point prompts as cues for adaptation, which is inapplicable for MedSAM due to its bounding box-based prompts." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.258, + 0.482, + 0.304 + ], + "angle": 0, + "content": "Thus, there is a lack of a customized approach for the TTA problem of foundation medical segmentation models when considering both effectiveness and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.314, + 0.358, + 0.329 + ], + "angle": 0, + "content": "2.2. Latent Refinement for TTA" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.337, + 0.485, + 0.594 + ], + "angle": 0, + "content": "Typically, latent refinement is applied in the neural image compression community [1, 5, 14, 21, 23] to reduce the bit consumption of latent representation at compression time. Since the image compression task has a strong self-supervised signal that corresponds to reconstructing an original raw image by an encoder-decoder architecture [1, 5, 14, 21], it is more straightforward to leverage the latent refinement technique. Instead, there is no direct-supervised signal for medical image segmentation during inference, leaving unexplored space. To the best of our knowledge, we are the first to introduce the concept of latent refinement to test-time adaptation of medical image segmentation. More importantly, we provide theoretical analyses to uncover that directly refining the latent representation of the input image is feasible to approach the same goal as parametric update-based counterparts, which is not developed by previous literature." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.608, + 0.248, + 0.626 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.633, + 0.436, + 0.649 + ], + "angle": 0, + "content": "3.1. Preliminary and Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.656, + 0.483, + 0.731 + ], + "angle": 0, + "content": "This paper mainly discusses the common MedSAM architecture in [15] with the bounding box prompt. During the test-time adaptation, the bounding box prompt is provided and fixed but is omitted in the following derivations for notation simplification." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.732, + 0.483, + 0.821 + ], + "angle": 0, + "content": "Definition 1. (Architecture of MedSAM) Unlike U-Net and its variants [18], MedSAM has no skip connections between an image encoder \\( E_{\\theta} \\) parameterized by \\( \\theta \\) and a mask decoder \\( D_{\\varphi} \\) parameterized by \\( \\varphi \\), which enables multiple interactions by reusing the image embedding. In light of this, the following proposition can be derived:" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.822, + 0.483, + 0.882 + ], + "angle": 0, + "content": "Proposition 1. Let \\(\\mathbf{X} \\in \\mathcal{X}\\), \\(\\mathbf{Z} \\in \\mathcal{Z}\\), and \\(\\mathbf{Y} \\in \\mathcal{Y}\\) be an input image, its latent embedding, and its corresponding prediction. For a MedSAM model, \\(\\mathbf{Y}\\) is strictly conditionally independent of \\(\\mathbf{X}\\) given \\(\\mathbf{Z}\\), i.e., \\(\\mathbf{Y} \\perp \\mathbf{X} \\mid \\mathbf{Z}\\)." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.883, + 0.483, + 0.913 + ], + "angle": 0, + "content": "Proof: Once the (deepest) latent embedding \\(\\mathbf{Z}\\) is calculated by \\(\\mathbf{Z} = E_{\\theta}(\\mathbf{X})\\), the upsampling function \\(\\mathbf{Y} =\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.882, + 0.122 + ], + "angle": 0, + "content": "\\(D_{\\varphi}(\\mathbf{Z})\\) is not directly related to the input image \\(\\mathbf{X}\\) or its any latent variants, leading to \\(\\mathbf{Y}\\perp \\mathbf{X}\\mid \\mathbf{Z}\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.882, + 0.182 + ], + "angle": 0, + "content": "Theorem 1. (General objective) Under maximum a posteriori (MAP) estimation, the objective of TTA for image segmentation is computing the optimal prediction \\(\\mathbf{Y}^*\\) by maximizing the posterior probability conditioned on \\(\\mathbf{X}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.207, + 0.894, + 0.231 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} ^ {*} = D _ {\\varphi^ {*}} \\left(E _ {\\theta^ {*}} (\\mathbf {X})\\right), \\varphi^ {*}, \\theta^ {*} = \\arg \\max _ {\\theta , \\varphi} [ P (\\mathbf {Y} | \\mathbf {X}; \\theta , \\varphi) ] ^ {3}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.239, + 0.882, + 0.314 + ], + "angle": 0, + "content": "Various TTA strategies are regarded as implicit posterior maximization by self-supervised loss-based partial or overall parameter updates. Here, we incorporate the architecture of the MedSAM as a factorized precondition to yield an extension of Theorem 1 as follows." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.315, + 0.882, + 0.39 + ], + "angle": 0, + "content": "Corollary 1. (Factorized objective of MedSAM) The log-posterior \\(\\log P(\\mathbf{Y}|\\mathbf{X})\\) during test-time adaptation can be factorized into the sum of the segmentation likelihood \\(\\log P(\\mathbf{Y}|\\mathbf{Z})\\) and the latent fidelity term \\(\\log P(\\mathbf{Z}|\\mathbf{X})\\) based on \\(\\mathbf{Y}\\perp \\mathbf{X}\\mid \\mathbf{Z}\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.4, + 0.804, + 0.418 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} ^ {*} = D _ {\\varphi^ {*}} \\left(\\mathbf {Z} ^ {*}\\right), \\mathbf {Z} ^ {*} = E _ {\\theta^ {*}} (\\mathbf {X})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.419, + 0.873, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\varphi^ {*}, \\theta^ {*} = \\arg \\max _ {\\varphi , \\theta} [ \\log P (\\mathbf {Y} | \\mathbf {Z}; \\varphi) + \\log P (\\mathbf {Z} | \\mathbf {X}; \\theta) ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.452, + 0.882, + 0.513 + ], + "angle": 0, + "content": "Proof: By recalling Prop. 1, \\( P(\\mathbf{Y}|\\mathbf{X}) \\propto P(\\mathbf{Y}|\\mathbf{Z}) \\cdot P(\\mathbf{Z}|\\mathbf{X}) \\) holds. Substituting this into the MAP objective in Theorem 1 yields the factorized objective above. Please refer to the Appendix for more details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.524, + 0.882, + 0.539 + ], + "angle": 0, + "content": "3.2. Latent Refinement in MedSAM framework" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.546, + 0.882, + 0.607 + ], + "angle": 0, + "content": "Problem. For Corollary 1, one of the commonly adopted solutions for achieving optimal prediction \\(\\mathbf{Y}^*\\) is directly to jointly optimize partial or whole parameters in \\(\\{\\varphi ,\\theta \\}\\), but two challenges are encountered as follows." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.608, + 0.882, + 0.697 + ], + "angle": 0, + "content": "- Partial model updates. Most existing approaches are BN-centric strategies, which are inapplicable for LN-based MedSAM. Although updating the affine transformation parameters in LN layers is feasible, the performance may be marginal due to restricted model modifications especially in strong domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.698, + 0.882, + 0.772 + ], + "angle": 0, + "content": "- Whole model updates. It is difficult to maintain a trade-off between TTA performance and catastrophic forgetting. Meanwhile, optimizing the overall model is computationally expensive due to the considerable number of parameters in modern large models." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.608, + 0.882, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.773, + 0.882, + 0.878 + ], + "angle": 0, + "content": "Solution. In light of this, we are interested in directly optimizing the latent embedding rather than the model parameters. Assume the encoder \\( E_{\\theta} \\) and the decoder \\( D_{\\varphi} \\) are fixed during TTA, this scheme can not only circumvent the above-mentioned challenges (i.e., it is likely to update more \"parameters\" in the forgetting-free context) but also be equivalent to theoretically achieving the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.887, + 0.882, + 0.911 + ], + "angle": 0, + "content": "3In practice, one would optimize the expectation of the probability. To be simplified, we omit it in later formulations." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.518, + 0.963 + ], + "angle": 0, + "content": "20077" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.289, + 0.884, + 0.318 + ], + "angle": 0, + "content": "Figure 1. The framework of the proposed method, including initial embedding inference, TTA via latent refinement, and final inference phases. Distribution-approximated latent CRF loss is visualized, where the loss between \\(\\mathbf{z}_i\\) and \\(\\mathbf{z}_j\\) is computed." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.343, + 0.483, + 0.361 + ], + "angle": 0, + "content": "same factorized objective of MedSAM as Corollary 1," + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.38, + 0.352, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} ^ {*} = D _ {\\varphi} (\\mathbf {Z} ^ {*}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.4, + 0.482, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} ^ {*} = \\arg \\max _ {\\mathbf {Z}} \\left[ \\log P (\\mathbf {Y} | \\mathbf {Z}; \\varphi) + \\log P (\\mathbf {Z} | \\mathbf {X}; \\theta) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.427, + 0.485, + 0.729 + ], + "angle": 0, + "content": "where \\(\\mathbf{Z} = E_{\\theta}(\\mathbf{X})\\) can be initialized and refined iteratively. Typically, the total parameter number of \\(\\mathbf{Z}\\) is overly larger than that of the normalization layer, resulting in more update flexibility. Under frozen \\(\\theta\\) and \\(\\varphi\\), there is no knowledge forgetting during optimization. These two perspectives are more likely to activate better TTA performance collaboratively than existing methods. Effectiveness of Latent Refinement during TTA. First, the \\(\\log P(\\mathbf{Z}|\\mathbf{X})\\) term encourages \\(\\mathbf{Z}\\) to maintain sufficient statistics over the input \\(\\mathbf{X}\\), which is highly significant for the MedSAM framework during the TTA because 1) domain-specific variations at test time may degrade salient characteristics (e.g., segmentation cues like obvious edges or contours) of latent embedding, leading to suboptimal segmentation performance; 2) sufficient statistics relative to the input \\(\\mathbf{X}\\) can encourage the conditional independence assumption of MedSAM, i.e., \\(\\mathbf{Y} \\perp \\mathbf{X} \\mid \\mathbf{Z}\\), to hold. Second, \\(\\log P(\\mathbf{Y}|\\mathbf{Z})\\) guarantees the usability of latent embedding for final predictions, preventing any collapsed solutions." + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.73, + 0.483, + 0.774 + ], + "angle": 0, + "content": "Practical implementation. For Eq. (1), one would update the initial latent embedding \\(\\mathbf{Z} = \\mathbf{Z}^0 = E_\\theta (\\mathbf{X})\\) with \\(m\\geq 1\\) steps, for each step \\(\\mathbf{Z}^{m} =\\)" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.779, + 0.511, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} ^ {m - 1} - \\epsilon \\cdot \\frac {\\partial [ - \\log P (\\mathbf {Y} ^ {m - 1} | \\mathbf {Z} ^ {m - 1} ; \\varphi) - \\log P (\\mathbf {Z} ^ {m - 1} | \\mathbf {X} ; \\theta ]}{\\partial \\mathbf {Z} ^ {m - 1}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.823, + 0.483, + 0.854 + ], + "angle": 0, + "content": "where \\(\\epsilon\\) is the learning rate. The final prediction can be represented as \\(\\mathbf{Y}^m = D_{\\varphi}(\\mathbf{Z}^m)\\)." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.86, + 0.458, + 0.877 + ], + "angle": 0, + "content": "3.3. Distribution-Approximated Latent CRF" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.882, + 0.483, + 0.914 + ], + "angle": 0, + "content": "To Eq. (2), one of the most important components is maximizing the log-probability \\( P(\\mathbf{Z}^{m - 1}|\\mathbf{X};\\theta) \\), i.e.," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.884, + 0.51 + ], + "angle": 0, + "content": "the latent fidelity term. For MedSAM, the volume of the image encoder parameter \\(\\theta\\) is considerable, resulting in a high computation during the forward and backward propagations in the TTA phase. Benefiting from our latent refinement framework, we propose circumventing the image encoder \\(\\theta\\) to directly enforce sufficient statistics over the input image \\(\\mathbf{X}\\), i.e., maximizing \\(\\log P(\\mathbf{Z}^{m-1}|\\mathbf{X})\\). By doing so, we can leverage potential useful cues in the input space, which is typically ignored by existing TTA methods that focus more on the label space (e.g., prediction consistency [2])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.884, + 0.786 + ], + "angle": 0, + "content": "Therefore, we introduce the concept of conditional random field (CRF) [22], which is usually employed to refine the final predictions by enforcing the co-occurrence between the label space \\(\\mathcal{V}\\) and the input space \\(\\mathcal{X}\\). For our purpose, we extend the CRF to the co-occurrence between the latent space \\(\\mathcal{Z}\\) and the input space \\(\\mathcal{X}\\). Theoretically, minimizing the CRF energy is equivalent to maximizing the posterior probability of the random field [11], which coincides with our optimization objective in Eq. (2) via a computationally efficient manner. Formally, given a random field \\(\\mathbf{Z} = E_{\\theta}(\\mathbf{X})\\) over a set of vectors \\(\\{\\mathbf{z}_1,\\mathbf{z}_2,\\dots ,\\mathbf{z}_{N'}\\}\\), where \\(\\mathbf{Z}\\in \\mathbb{R}^{W^{\\prime}\\times H^{\\prime}\\times C^{\\prime}}\\) is the latent embedding with a spatial resolution of \\(N^{\\prime} = W^{\\prime}\\times H^{\\prime}\\) and a channel of \\(C^\\prime\\), i.e., \\(\\mathbf{z}\\in \\mathbb{R}^{1\\times C^{\\prime}}\\). Consider another random field \\(\\mathbf{X}\\) over a set of vectors \\(\\{\\mathbf{x}_1,\\mathbf{x}_2,\\dots ,\\mathbf{x}_N\\}\\), where \\(\\mathbf{X}\\in \\mathbb{R}^{W\\times H\\times C}\\) is the input intensities with a spatial resolution of \\(N = W\\times H\\) and a channel of \\(C\\), i.e., \\(\\mathbf{x}\\in \\mathbb{R}^{1\\times C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.885, + 0.913 + ], + "angle": 0, + "content": "Typically, \\(\\mathbf{x}_i\\) and \\(\\mathbf{z}_i\\) are related in a one-to-one manner, i.e., \\(\\mathbf{x}_i\\) represents a color vector at pixel \\(i\\) in the input space and \\(\\mathbf{z}_i\\) is its corresponding predicted label in the label space. However, the input and latent spaces are dimensionally heterogeneous, i.e., \\(N \\neq N'\\), which makes constructing a random field in this scenario challenging. Here, a novel distribution-approximated latent CRF (DAL-CRF) is proposed to eliminate the dimen" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "20078" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.112, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "sionally heterogeneous problem between the input and latent spaces. Specifically, by leveraging the fact that each latent vector \\(\\mathbf{z}_i\\) corresponds to a receptive field in the input space with a set of intensity vectors, it is feasible to identify related intensity vectors for each latent vector, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.191, + 0.483, + 0.21 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {i} \\rightarrow \\mathbf {X} _ {i} = \\left\\{\\mathbf {x} _ {i} ^ {l} \\right\\} _ {l = 1} ^ {L}, i = 1, 2, \\dots , N ^ {\\prime}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.219, + 0.484, + 0.31 + ], + "angle": 0, + "content": "where \\(L\\) denotes the number of intensity vectors in the receptive field, where \\(L = 256\\) corresponding to a \\(16\\times 16\\) receptive field in MedSAM. Then, by following Veksler [24], the pairwise potential is only considered to be a regularization loss to encourage the minimization of CRF energy, which can be formulated as \\(\\mathcal{L}_{\\mathrm{DAL - CRF}} =\\)" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.317, + 0.483, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\frac {1}{N ^ {\\prime}} \\sum_ {i, j \\in \\mathcal {G}} \\left(\\frac {1}{2} \\operatorname {M M D} \\left(\\mathbb {P} \\left(\\mathbf {X} _ {i}\\right), \\mathbb {P} \\left(\\mathbf {X} _ {j}\\right)\\right)\\right) \\cdot \\exp \\left(- \\| \\mathbf {z} _ {i} - \\mathbf {z} _ {j} \\| _ {F}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.368, + 0.485, + 0.566 + ], + "angle": 0, + "content": "Unlike existing CRF methods [24], which utilize the Gaussian kernel to evaluate the correlation between two color vectors, we introduce the Maximum Mean Discrepancy (MMD) to measure the distance between two vector sets. Specifically, each vector set is first approximated as an empirical distribution, i.e., \\(\\mathbf{x}_i^l\\sim \\mathbb{P}(\\mathbf{X}_i)\\). Then, each empirical distribution is represented by a kernel mean embedding in a reproducing kernel Hilbert space (RKHS) \\(\\mathcal{H}\\), i.e., \\(\\mu_{\\mathbb{P}} = \\mathbb{E}_{\\mathbf{x}\\sim \\mathbb{P}}[\\phi (\\mathbf{x})] = \\int_{\\mathcal{X}}k(\\mathbf{x},\\cdot)d\\mathbb{P} = \\mathbb{E}_{\\mathbf{x}\\sim \\mathbb{P}}[k(\\mathbf{x},\\cdot)]\\), where \\(k:\\mathcal{X}\\times \\mathcal{X}\\to \\mathbb{R}\\) is a reproducing kernel (Gaussian RBF kernel here) and corresponding feature map \\(\\phi :\\mathcal{X}\\rightarrow \\mathcal{H}\\). The discrepancy can be measured between mean embeddings, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.586, + 0.498, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M M D} \\left(\\mathbb {P} \\left(\\mathbf {X} _ {i}\\right), \\mathbb {P} \\left(\\mathbf {X} _ {j}\\right)\\right) = \\| \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\phi \\left(\\mathbf {x} _ {i} ^ {l}\\right) - \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\phi \\left(\\mathbf {x} _ {j} ^ {l}\\right) \\| _ {\\mathcal {H}} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.641, + 0.483, + 0.761 + ], + "angle": 0, + "content": "MMD enables all-order correlation measurement using Gaussian kernel via unbiased empirical estimation. This may be useful for better discrepancy measurement, as the distribution of the intensity space may be highly complex and non-linear. For the second term in Eq. (4), we utilize the Frobenius norm \\(\\| \\cdot \\| _F\\) to calculate the feature compatibility in a co-occurrent probability using exponential zero-to-one normalization." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.762, + 0.484, + 0.913 + ], + "angle": 0, + "content": "Integrating Bounding Box Prompt. The graph \\(\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}\\) in Eq. (4) is usually either a complete (dense-connected) graph [11] or a sparse position-related neighboring graph [24] on \\(\\mathbf{Z}\\). Instead, we propose to consider the bounding box prompt as the prior to construct \\(K\\)-random pairs inside the given bounding box, which can not only exclude those unimportant latent features far away from the bounding box prompt to enhance the computational efficiency but also focus more on the interested region with better performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.885, + 0.244 + ], + "angle": 0, + "content": "Intuitive Understanding. For Eq. (4), one actually encourages being more discriminative in latent representations between \\(\\mathbf{z}_i\\) and \\(\\mathbf{z}_j\\). Since if two empirical distributions have a large distribution discrepancy (i.e., the MMD distance is large) in the input space, there are more penalties assigned to the pair who has larger similarity (i.e., co-occurrent probability is larger)) in the latent space. By minimizing \\(\\mathcal{L}_{\\mathrm{DAL - CRF}}\\), one can enforce the latent representations to equip with sufficient statistics as in the input space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.252, + 0.872, + 0.268 + ], + "angle": 0, + "content": "3.4. Overall Objective at Test-time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.884, + 0.35 + ], + "angle": 0, + "content": "Without accessing the ground-truth label, it is infeasible to maximize the segmentation likelihood \\(\\log P(\\mathbf{Y}^m|\\mathbf{Z}^m;\\varphi)\\) directly. Therefore, we adopt a common predictive entropy minimization to implicitly achieve this goal [8]. Formally," + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.36, + 0.903, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E M}} = H (P (\\mathbf {Y} | \\mathbf {Z})) = - \\sum_ {h, v \\in C} P (y ^ {h, v} | \\mathbf {Z}) \\log P (y ^ {h, v} | \\mathbf {Z}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.396, + 0.884, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\quad C = \\left\\{\\left(h, v\\right) \\mid P \\left(y ^ {h, v} \\mid \\mathbf {Z}\\right) \\geq \\alpha \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.884, + 0.652 + ], + "angle": 0, + "content": "where \\( y^{h,v} \\) denotes a single prediction in the position \\( (h,v) \\) of \\( \\mathbf{Y} \\). Eq. (6) is similar to foreground-background balanced entropy loss in Dong et al. [6]. However, we only utilize the foreground entropy and set a high confidence threshold (e.g., \\( \\alpha = 0.95 \\)) due to the following reasons. First, since the CRF-based loss usually requires a volumetric regularization to avoid the trivial solutions [24] (e.g., all labels are assigned to background), the entropy regularization over a high confidence foreground region can be interpreted as an implicit volumetric regularization, i.e., ensuring these foreground pixels confident enough for preventing trivial solutions. Second, we argue that background entropy is unnecessary for a foundation segmentation model, as these models have learned a good foreground-background balance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.884, + 0.713 + ], + "angle": 0, + "content": "By considering the proposed distributionapproximated latent CRF loss and the customized entropy loss, the practical latent refinement scheme is reformulated as follows." + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.721, + 0.883, + 0.752 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} ^ {m} = \\mathbf {Z} ^ {m - 1} - \\epsilon \\cdot \\frac {\\partial \\left[ \\mathcal {L} _ {\\mathrm {D A L - C R F}} + \\lambda \\mathcal {L} _ {\\mathrm {E M}} \\right]}{\\partial \\mathbf {Z} ^ {m - 1}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.761, + 0.729, + 0.774 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a balance coefficient." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.73, + 0.8 + ], + "angle": 0, + "content": "3.5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.884, + 0.913 + ], + "angle": 0, + "content": "The overall framework is illustrated in Figure 1. Superficially, the initial image embedding \\(\\mathbf{Z}^0\\) is inferred only once. Once the refined latent representation \\(\\mathbf{Z}^*\\) is obtained, the upsampling can be conducted directly without requiring image embedding again. This mechanism is well compatible with the interactive scheme of the MedSAM. We implement the proposed method based on" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "20079" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.125, + 0.09, + 0.878, + 0.42 + ], + "angle": 0, + "content": "
Update TypeMethodDice Coefficient ↑Average Surface Distance ↓
D1D2D3D4Avg.D1D2D3D4Avg.
Looseness promptsDirect Inference90.9987.1687.0684.7287.489.487.518.747.628.33
Normalization updateTENT (ICLR'21)91.2785.4085.0681.8485.899.098.7010.058.889.18
InTEnt (CVPR'24)91.3687.4387.4786.1387.608.807.518.267.668.06
GraTa (AAAI'25)90.8987.0086.9584.6487.379.487.548.807.658.37
PASS (TMI'24)80.8679.6382.0384.0281.6320.4012.2211.697.9713.07
Full updatesMEMO (NeurIPS'22)87.2685.4886.4284.4785.9112.458.279.157.769.40
CRF-SOD (CVPR'23)89.5784.0184.8082.5885.2410.548.9910.188.649.58
Latent updatesOurs89.3588.7389.8388.1889.0211.406.696.926.077.77
Shrinkage promptsDirect Inference81.0567.6072.1068.5772.3318.5117.6117.7614.9217.20
Normalization updateTENT (ICLR'21)80.2265.1769.2562.1069.1818.8519.2619.3917.5818.77
InTEnt (CVPR'24)82.7270.0674.5073.4975.1916.8615.2816.1313.8715.53
GraTa (AAAI'25)80.8767.3271.9468.4672.1517.2716.5416.5314.5216.21
PASS (TMI'24)66.8158.6663.3156.7861.3932.3922.8922.7420.1924.55
Full updatesMEMO (NeurIPS'22)81.0267.5472.1068.8472.3718.5217.7617.7714.7817.20
CRF-SOD (CVPR'23)81.8469.2674.7372.0874.4819.2718.8318.6314.7317.86
Latent updatesOurs83.0773.7278.5475.7777.7716.7414.3813.7211.2614.02
Perfectness promptsDirect Inference87.5880.2581.3278.4681.9012.7611.2112.2710.2411.62
Normalization updateTENT (ICLR'21)87.6077.9878.1872.9979.1812.8312.2113.9212.6912.91
InTEnt (CVPR'24)86.5484.2585.1781.6584.4010.989.239.858.959.75
GraTa (AAAI'25)87.3979.9781.1478.3681.7112.9611.3812.3810.2911.75
PASS (TMI'24)74.5969.8673.8871.3472.4125.5317.3216.5813.8518.32
Full updatesMEMO (NeurIPS'22)87.3980.1881.3278.4981.8412.9711.2412.2710.2311.67
CRF-SOD (CVPR'23)86.0777.2279.1076.2479.6513.9612.5913.6311.212.84
Latent updatesOurs86.8585.1186.7284.4985.7913.738.638.797.489.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.431, + 0.884, + 0.474 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of TTA results on fundus image OC segmentation between different methods. The best and second-best results are shown in the bold and the underline, respectively. For loose prompts, we report the average of two looseness ratios. For shrinkage prompts, we report the average of two shrinkage ratios. Full results can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.485, + 0.484, + 0.804 + ], + "angle": 0, + "content": "official MedSAM checkpoints (ViT-B pre-trained models) and codes using a one-step optimization strategy \\((m = 1)\\) by following previous TTA settings. Due to the extremely various sizes of RoIs in medical images, we leverage the initial segmentation result \\(\\mathbf{Y}^0\\) and the given bounding box prompt to explore a dynamic adjustment strategy for important hyperparameters. Specifically, the learning rate \\(\\epsilon\\) with the Adam optimizer is linearly interpolated from \\(5\\times 10^{-2}\\) to \\(1\\times 10^{-1}\\) according to the ratio of the initial foreground to the entire image \\(f_{e}\\) (i.e., the rough size of the RoI), where \\(f_{e}\\leq 0.05\\) and \\(f_{e}\\geq 0.2\\) are fixedly set to \\(5\\times 10^{-2}\\) and \\(1\\times 10^{-1}\\), respectively. Note that we use commonly-used definitions over the \\(f_{e}\\) as adopted by Gao et al. [7], i.e., \\(f_{e}\\leq 0.05\\), \\(0.05 < f_{e} < 0.2\\), and \\(f_{e}\\geq 0.2\\) as small, medium, and large RoIs. Similarly, the \\(\\lambda\\) is also RoI-specific adjustment, where the small RoI \\((f_{e}\\leq 0.05)\\) requires a similar contribution between the \\(\\mathcal{L}_{\\mathrm{DAL - CRF}}\\) and the \\(\\mathcal{L}_{\\mathrm{EM}}\\), thus the \\(\\lambda\\) is set to \\(1\\times 10^{3}\\). The \\(K\\) is set to 4 by balancing computational consumption and performance (refers to Table 5(a)). \\(\\alpha\\) in entropy loss is set to 0.95 empirically." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.824, + 0.245, + 0.841 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.852, + 0.484, + 0.914 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics We evaluate our method on two public multi-center medical image segmentation datasets with different data modalities. (1) Fundus [26]: Fundus datasets are retinal fundus images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.884, + 0.637 + ], + "angle": 0, + "content": "collected from 4 healthcare centers (D1 to D4 for short). For each retinal fundus image, there are two segmentation objectives, including the optic cup (OC) (mainly small RoIs) and optic disc (OD) (mainly medium or large RoIs). (2) Spinal Cord Gray Matter (SCGM) [12]: SCGM datasets are collected from 4 institutions (D1 to D4 for short) with different imaging manufacturers for spinal MRI image gray matter segmentation. We use the Dice Similarity Coefficient (DSC) and Average Surface Distance (ASD) to measure the accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.885, + 0.912 + ], + "angle": 0, + "content": "Experiment Settings. Appearance Perturbations: Note that our adopted multi-center datasets have strong domain shifts with appearance variations, we thus ignore test-time appearance perturbations (e.g., changes in contrast, noise, and blur). Prompt Perturbations: We mainly consider three kinds of bounding box prompts to evaluate TTA performance. (i) Perfectness prompt \\((PP)\\): The perfect prompt can be obtained by groundtruth (GT) mask, where the top-left and lower-right coordinates of the PP correspond to the minimal and maximal horizontal and vertical coordinates of the GT mask, respectively. A unique PP is utilized for TTA. (ii) Looseness prompts \\((LP)\\): Clinically, it is labor-intensive and time-consuming to obtain the PP. Instead, a moderately loose prompt can be efficient but may trigger the degradation of the model with the need for TTA. Therefore, we define the looseness rate \\(\\sigma\\), which is the ratio of the coordinate shift relative to the width \\(w\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "20080" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.125, + 0.09, + 0.878, + 0.419 + ], + "angle": 0, + "content": "
Update TypeMethodDice Coefficient ↑Average Surface Distance ↓
D1D2D3D4Avg.D1D2D3D4Avg.
Looseness promptsDirect Inference83.0578.8681.8275.5479.8133.2032.8730.2931.9032.06
Normalization updateTENT (ICLR'21)81.6176.9680.0074.8578.3537.5836.5433.5334.3535.50
InTEnt (CVPR'24)83.6176.7180.0071.8378.0332.4637.0132.5739.1435.29
GraTa (AAAI'25)84.6480.0484.1082.8282.9030.0230.9325.9128.9128.94
PASS (TMI'24)84.5781.7782.2977.3481.4918.0226.2021.0924.6822.49
Full model updatesMEMO (NeurIPS'22)86.2680.0183.9478.1882.1026.7330.9326.2728.5428.12
CRF-SOD (CVPR'23)87.0382.8683.3379.5283.1825.1926.2523.7226.5325.42
Latent updatesOurs90.8583.5587.7782.2486.1017.5923.5418.5021.7520.34
Shrinkage promptsDirect Inference92.7086.1790.9088.6789.6112.6017.6612.7712.1013.78
Normalization updateTENT (ICLR'21)90.6385.6389.6686.8888.2015.9018.4414.2712.4815.27
InTEnt (CVPR'24)90.9187.2790.8888.7089.4415.4916.4712.6812.0514.17
GraTa (AAAI'25)92.6185.6091.0088.7189.4812.7018.3212.6112.1113.93
PASS (TMI'24)91.4678.8585.9275.7082.9813.7225.1417.5423.0519.86
Full model updatesMEMO (NeurIPS'22)92.1283.8290.8788.7788.8913.5220.1812.7512.0414.62
CRF-SOD (CVPR'23)94.4284.0792.3491.4390.569.3619.8310.709.3012.30
Latent updatesOurs94.2486.7793.2392.4291.669.8016.259.458.2110.93
Perfectness promptsDirect Inference96.4992.8793.8992.3693.906.099.488.648.368.14
Normalization updateTENT (ICLR'21)94.2090.7893.3990.3892.1810.1212.019.3210.3910.46
InTEnt (CVPR'24)94.1291.5794.4293.4793.3910.1211.018.017.039.04
GraTa (AAAI'25)94.4692.6693.9192.3193.336.149.738.618.428.22
PASS (TMI'24)94.9085.3584.3686.1487.688.5817.5217.6213.9814.42
Full model updatesMEMO (NeurIPS'22)95.7791.4893.8992.3193.367.8011.288.618.419.02
CRF-SOD (CVPR'23)96.6691.1593.7693.4193.755.7811.428.737.268.29
Latent updatesOurs96.6892.6795.1594.5794.765.699.516.855.987.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.43, + 0.884, + 0.472 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of TTA results on fundus image OD segmentation between different methods. The best and second-best results are shown in the bold and the underline, respectively. For loose prompts, we report the average of two looseness ratios. For shrinkage prompts, we report the average of two shrinkage ratios. Full results can be found in the Appendix." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.476, + 0.878, + 0.607 + ], + "angle": 0, + "content": "
Update TypeMethodDice Coefficient ↑Average Surface Distance ↓
D1D2D3D4Avg.D1D2D3D4Avg.
Direct Inference55.3554.2061.5456.6756.9410.2211.3915.2715.9513.20
Normalization updateTENT (ICLR'21)49.6652.9659.3955.3854.349.9410.4315.8516.1213.08
InTEnt (CVPR'24)50.9152.1245.8550.3849.819.3510.9922.5719.0115.48
GraTa (AAAI'25)56.3555.1463.6658.1258.319.2610.4414.3615.1512.30
PASS (TMI'24)55.2454.8061.9452.8856.2110.5811.0115.0220.1413.99
Full updatesMEMO (NeurIPS'22)56.2055.2061.2057.3457.489.2710.4514.8915.3912.50
CRF-SOD (CVPR'23)54.2054.0861.8957.0756.8110.3710.8315.3615.5813.03
Latent updatesOurs57.9255.4266.0058.6259.498.3510.5814.1414.1511.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.618, + 0.884, + 0.647 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of TTA results on MRI image SCGM segmentation between different methods. SCGM segmentation only reported PP-based results due to its very challenging intricate structures and appearance perturbations." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.662, + 0.484, + 0.904 + ], + "angle": 0, + "content": "height \\(h\\) of the PP, to simulate real-world loose scenarios as follows: \\(x(y)_{tl}^{LP}: x(y)_{tl} - \\sigma \\cdot w(h), x(y)_{lr}^{LP}: x(y)_{lr} + \\sigma \\cdot w(h)\\), where \\(x(y)_{tl}\\) and \\(x(y)_{lr}\\) denote the vertical (horizontal) coordinates of the top-left and lower-right points of the PP. Each loose prompt is represented as \\((x_{tl}^{LP}, y_{tl}^{LP}, x_{lr}^{LP}, y_{lr}^{LP})\\). Here, \\(\\sigma\\) is set to 0.1 and 0.2, as we observe 1) too small \\(\\sigma (< 0.1)\\) is close to the PP without obvious changes and 2) too large \\(\\sigma (>0.2)\\) will usually result in significantly overlaps with other RoIs with unacceptable initial results. We report the average of different loose rates for comprehensive evaluation. (iii) Shrinkage prompts (SP): Similarly, we define the shrinkage ratio \\(\\gamma\\) to simulate the scenario that the bounding box prompt shifts to the interior of the PP as follows: \\(x(y)_{tl}^{SP}: x(y)_{tl} + \\gamma \\cdot w(h), x(y)_{lr}^{SP}: x(y)_{lr} - \\gamma \\cdot w(h)\\). Each shrinkage prompt is represented" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.884, + 0.708 + ], + "angle": 0, + "content": "as \\((x_{tl}^{SP},y_{tl}^{SP},x_{lr}^{SP},y_{lr}^{SP})\\). Similarly, \\(\\gamma\\) is set to 0.1 and 0.2. We report the average of different shrinkage rates for comprehensive evaluation." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.716, + 0.884, + 0.913 + ], + "angle": 0, + "content": "Baselines. In this study, two kinds of TTA approaches for medical image segmentation are utilized for comparison. (1) Normalization-based TTA: Most baseline methods conduct the update of affine transformation parameters and manipulation of source-trained statistics in BN layers, which the MedSAM lacks due to the usage of LN layers. We, therefore, discard source-trained statistics-related components and only optimize the affine-transformation parameters for baseline methods, including TENT (entropy minimization), InTEnt (foreground-background-balanced entropy weighting), GraTa (augmentation-based gradient alignment), and PASS (input decora" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.518, + 0.963 + ], + "angle": 0, + "content": "20081" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.088, + 0.216, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.09, + 0.308, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.091, + 0.402, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.091, + 0.498, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.091, + 0.596, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.154, + 0.218, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.154, + 0.311, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.154, + 0.404, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.405, + 0.154, + 0.498, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.154, + 0.607, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.215, + 0.219, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.221, + 0.215, + 0.311, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.215, + 0.403, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.215, + 0.498, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.215, + 0.607, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.282, + 0.609, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Visualized process of latent refinement with iterations. The first, second, and third rows denote the segmentation result, corresponding visualization of channel-average refined latent embedding (which is from the region in bounding box prompt and has been fused with the prompt for a better view), and corresponding DAL-CRF loss for each spatial position. You may zoom in for a better view." + }, + { + "type": "table_caption", + "bbox": [ + 0.743, + 0.083, + 0.758, + 0.09 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "table", + "bbox": [ + 0.617, + 0.092, + 0.882, + 0.156 + ], + "angle": 0, + "content": "
KD1D2D3D4Avg.
293.2487.0692.5186.5189.75
494.1687.1691.8387.5790.14
694.6787.0892.3487.9290.37
894.8686.5490.8188.0190.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.742, + 0.161, + 0.757, + 0.17 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "table", + "bbox": [ + 0.617, + 0.172, + 0.882, + 0.253 + ], + "angle": 0, + "content": "
LossD1D2D3D4Avg.
DI89.1984.3187.9282.3985.95
\\( \\mathcal{L}_{\\text{InTEnt}} \\)79.3484.8988.8386.0884.78
\\( \\mathcal{L}_{\\text{TENT}} \\)85.0984.9989.4685.5986.28
\\( \\mathcal{L}_{\\text{CRF-SOD}} \\)88.9386.1788.4287.3387.71
\\( \\mathcal{L}_{\\text{DAL-CRF}} \\)94.1686.9690.8587.5789.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.608, + 0.267, + 0.889, + 0.349 + ], + "angle": 0, + "content": "Table 5. (a) Hyperparameter Analyses in terms of \\( K \\). (b) Integration of the latent refinement framework and other TTA loss functions \\( \\nu \\) s. Our proposed DAL-CRF, where DI denotes direct inference without adaptation on OD segmentation." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.362, + 0.48, + 0.43 + ], + "angle": 0, + "content": "
MethodD1D2D3D4Avg.
Direct Inference89.1984.3187.9282.3985.95
+ LEM85.8285.3188.1083.2186.01
+ LDAL-CRF94.1686.9690.8587.5789.88
+ LEM + LDAL-CRF94.1687.1691.8387.4290.14
DenseCRF [11]-Postprocessing82.7079.3980.9073.7679.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.441, + 0.482, + 0.469 + ], + "angle": 0, + "content": "Table 6. Ablation Study on Different Components of the OD segmentation with loose prompts, in terms of Dice score." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.482, + 0.483, + 0.603 + ], + "angle": 0, + "content": "tor and alternating momentum updating strategy). (2) Full update-based TTA: The whole model parameters are updated based on customized losses, including MEMO (augmentation-based marginal entropy minimization) and CRF-SOD (sparse CRF regularization loss). We leverage their open-source codes to apply into the architecture of MedSAM. To be fair, all baseline methods adopt a one-step optimization strategy." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.607, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Quantitative Analysis. From Tables 2, 3, and 4, some conclusions can be summarized. (1) Our proposed method without parametric updates surpasses all baselines in most cases with the best or second-best performance, showcasing the effectiveness of latent refinement. (2) It seems that full parameter updates are more suitable to large size RoIs than normalization-based updates, e.g., the OD. (3) For very challenging OC and SCGM tasks, most parametric update-based approaches have compromised performance compared with direct inference, but our proposed method performs well." + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.777, + 0.483, + 0.912 + ], + "angle": 0, + "content": "Qualitative Analysis. From Figure 2, our proposed method can achieve gradually better TTA segmentation with iterations (Note: we reduce the learning rate with multistep optimization rather than a one-step manner for understanding the optimization process better). Meanwhile, the latent embedding is gradually unsmoothed, which may be reasonable as the DAL-CRF loss enforces more statistics over the input image, leading to more discriminative information. The gradual decrease of DAL-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.366, + 0.883, + 0.395 + ], + "angle": 0, + "content": "CRF loss in each spatial point of the latent embedding is also observed, showcasing its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.884, + 0.577 + ], + "angle": 0, + "content": "Computation Complexity. As illustrated in Table 1, we achieve around \\(7 \\times\\) GFLOPs reduction compared with common usages. This is reasonable as 1) sophisticated augmentations (e.g., MEMO) and multiple gradient backward processes (e.g., GraTa) are computationally intensive. 2) For normalization-based updates, the gradient computation over normalization layers still passes through the entire model during backpropagation, which is unfavorable as MedSAM's image encoder is sophisticated with considerable consumption. 3) We circumvent any augmentation and the forward/backward process to the image encoder in a more efficient manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.884, + 0.76 + ], + "angle": 0, + "content": "Ablation Study & Other TTA losses under the latent refinement framework. As shown in Table 6, \\(\\mathcal{L}_{\\mathrm{DAL - CRF}}\\) significantly contributes more compared to entropy minimization loss and simply post-processing predictions fail to achieve acceptable performance on challenging medical images with strong appearance perturbations. Meanwhile, \\(\\mathcal{L}_{\\mathrm{DAL - CRF}}\\) still performs better, even though other TTA losses are integrated into the latent refinement framework as shown in 5(b). Compared to the CRF-SOD (using input and output spaces), our DAL-CRF exhibits superiority using the latent-based co-occurrence regularization for foundation models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.765, + 0.633, + 0.781 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.884, + 0.913 + ], + "angle": 0, + "content": "We propose a novel TTA framework for foundation medical segmentation without parametric updates. Our theoretical analysis shows that direct latent refinement is feasible within the MedSAM architecture. By maximizing factorized conditional probabilities by our DAL-CRF loss combined with entropy minimization, we achieve better performance for effectiveness and efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "20082" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.114, + 0.091, + 0.274, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.115, + 0.484, + 0.226 + ], + "angle": 0, + "content": "This work was supported in part by the Hong Kong Innovation and Technology Commission (ITC) (InnoHK Project CIMDA), in part by the Institute of Digital Medicine of City University of Hong Kong (CityU) under Project 9229503, in part by Chow Sang Sang Donation and Matching Fund (Project 9229161), and in part by the CityU under Project 7006087." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.244, + 0.21, + 0.259 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.268, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[1] Kecheng Chen, Pingping Zhang, Tiexin Qin, Shiqi Wang, Hong Yan, and Haoliang Li. Test-time adaptation for image compression with distribution regularization. In The Thirteenth International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.121, + 0.34, + 0.484, + 0.396 + ], + "angle": 0, + "content": "[2] Ziyang Chen, Yiwen Ye, Yongsheng Pan, and Yong Xia. Gradient alignment improves test-time adaptation for medical image segmentation. The 39th Annual AAAI Conference on Artificial Intelligence, 2025. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.399, + 0.484, + 0.467 + ], + "angle": 0, + "content": "[3] Mohammad Zalbagi Darestani, Jiayu Liu, and Reinhard Heckel. Test-time training can close the natural distribution shift performance gap in deep learning based compressed sensing. In International Conference on Machine Learning, pages 4754-4776. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.47, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[4] Zeshuai Deng, Zhuokun Chen, Shuaicheng Niu, Thomas Li, Bohan Zhuang, and Mingkui Tan. Efficient test-time adaptation for super-resolution with second-order degradation and reconstruction. Advances in Neural Information Processing Systems, 36:74671-74701, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.541, + 0.482, + 0.597 + ], + "angle": 0, + "content": "[5] JCMSA Djelouah and Christopher Schroers. Content adaptive optimization for neural image compression. In Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit, pages 1-5, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.599, + 0.482, + 0.681 + ], + "angle": 0, + "content": "[6] Haoyu Dong, Nicholas Konz, Hanxue Gu, and Maciej A Mazurowski. Medical image segmentation with intent: Integrated entropy weighting for single image test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5046-5055, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.684, + 0.482, + 0.767 + ], + "angle": 0, + "content": "[7] Yunhe Gao, Rui Huang, Yiwei Yang, Jie Zhang, Kainan Shao, Changjuan Tao, Yuanyuan Chen, Dimitris N. Metaxas, Hongsheng Li, and Ming Chen. Focusnetv2: Imbalanced large and small organ segmentation with adversarial shape constraint for head and neck ct images. Medical Image Analysis, 67:101831, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.769, + 0.482, + 0.811 + ], + "angle": 0, + "content": "[8] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.122, + 0.813, + 0.482, + 0.881 + ], + "angle": 0, + "content": "[9] Tianyu Huang, Tao Zhou, Weidi Xie, Shuo Wang, Qi Dou, and Yizhe Zhang. Improving segment anything on the fly: Auxiliary online learning and adaptive fusion for medical image segmentation. arXiv preprint arXiv:2406.00956, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.885, + 0.482, + 0.913 + ], + "angle": 0, + "content": "[10] Yiming Ju, Ziyi Ni, Xingrun Xing, Zhixiong Zeng, Hanyu Zhao, Siqi Fan, and Zheng Zhang. Mitigating" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.268, + 0.484, + 0.913 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.884, + 0.162 + ], + "angle": 0, + "content": "training imbalance in LLM fine-tuning via selective parameter merging. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15952-15959, Miami, Florida, USA, 2024. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.883, + 0.218 + ], + "angle": 0, + "content": "[11] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.883, + 0.289 + ], + "angle": 0, + "content": "[12] Haoliang Li, YuFei Wang, Renjie Wan, Shiqi Wang, TieQiang Li, and Alex Kot. Domain generalization for medical imaging classification with linear-dependency regularization. Advances in Neural Information Processing Systems, 33:3118-3129, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.883, + 0.346 + ], + "angle": 0, + "content": "[13] Zhenbing Liu, Fengfeng Wu, Yumeng Wang, Mengyu Yang, and Xipeng Pan. Fedcl: Federated contrastive learning for multi-center medical image classification. Pattern Recognition, 143:109739, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.883, + 0.417 + ], + "angle": 0, + "content": "[14] Yue Lv, Jinxi Xiang, Jun Zhang, Wenming Yang, Xiao Han, and Wei Yang. Dynamic low-rank instance adaptation for universal neural image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 632-642, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.883, + 0.46 + ], + "angle": 0, + "content": "[15] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.883, + 0.517 + ], + "angle": 0, + "content": "[16] Yuchen Mao, Hongwei Bran Li, LAI Yinyi, Giorgos Papanastasiou, Peng Qi, Yunjie Yang, and Chengjia Wang. Semi-supervised medical image segmentation via knowledge mining from large models. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.883, + 0.586 + ], + "angle": 0, + "content": "[17] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.883, + 0.658 + ], + "angle": 0, + "content": "[18] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.883, + 0.728 + ], + "angle": 0, + "content": "[19] Subhadeep Roy, Shankhanil Mitra, Soma Biswas, and Rajiv Soundararajan. Test time adaptation for blind image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16742-16751, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.883, + 0.799 + ], + "angle": 0, + "content": "[20] Robin Schon, Julian Lorenz, Katja Ludwig, and Rainer Lienhart. Adapting the segment anything model during usage in novel situations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3616-3626, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.801, + 0.883, + 0.87 + ], + "angle": 0, + "content": "[21] Sheng Shen, Huanjing Yue, and Jingyu Yang. Decap dapter: Exploring efficient decoder-side adapter for bridging screen content and natural image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12887-12896, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.871, + 0.883, + 0.913 + ], + "angle": 0, + "content": "[22] Charles Sutton, Andrew McCallum, et al. An introduction to conditional random fields. Foundations and Trends® in Machine Learning, 4(4):267-373, 2012. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.884, + 0.913 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "20083" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.115, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[23] Koki Tsubota, Hiroaki Akutsu, and Kiyoharu Aizawa. Universal deep image compression via content-adaptive optimization with adapters. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2529-2538, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[24] Olga Veksler. Test time adaptation with regularized loss for weakly supervised salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7360-7369, 2023. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.234, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[25] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.291, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[26] Shujun Wang, Lequan Yu, Kang Li, Xin Yang, Chi-Wing Fu, and Pheng-Ann Heng. Dofe: Domain-oriented feature embedding for generalizable fundus image segmentation on unseen datasets. IEEE Transactions on Medical Imaging, 39(12):4237-4248, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.361, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[27] Ruxue Wen, Hangjie Yuan, Dong Ni, Wenbo Xiao, and Yaoyao Wu. From denoising training to test-time adaptation: Enhancing domain generalization for medical image segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 464-474, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.446, + 0.482, + 0.512 + ], + "angle": 0, + "content": "[28] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.516, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[29] Guoping Xu, Xiaoxue Qian, Hua Chieh Shao, Jax Luo, Weiguo Lu, and You Zhang. A sam-guided and match-based semi-supervised segmentation framework for medical imaging. arXiv preprint arXiv:2411.16949, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.573, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[30] Chuyan Zhang, Hao Zheng, Xin You, Yefeng Zheng, and Yun Gu. Pass: test-time prompting to adapt styles and semantic shapes in medical image segmentation. IEEE Transactions on Medical Imaging, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.63, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[31] Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Yejia Zhang, and Danny Z Chen. Input augmentation with sam: Boosting medical image segmentation with segmentation foundation model. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.714, + 0.482, + 0.767 + ], + "angle": 0, + "content": "[32] Chen, Ziyang and Ye, Yiwen and Pan, Yongsheng and Xia, Yong. Gradient Alignment Improves Test-Time Adaptation for Medical Image Segmentation. In AAAI, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.117, + 0.77, + 0.482, + 0.825 + ], + "angle": 0, + "content": "[33] Zhang, Marvin and Levine, Sergey and Finn, Chelsea. Memo: Test time robustness via adaptation and augmentation Segmentation. In Advances in neural information processing systems, pages 38629-38642, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.092, + 0.482, + 0.825 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.952, + 0.519, + 0.963 + ], + "angle": 0, + "content": "20084" + } + ] +] \ No newline at end of file diff --git a/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/8a26d38c-cedb-499b-b41f-c629d5b94654_origin.pdf b/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/8a26d38c-cedb-499b-b41f-c629d5b94654_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6604bceab3d4dbb09e9cc61572cfa7e649bd9847 --- /dev/null +++ b/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/8a26d38c-cedb-499b-b41f-c629d5b94654_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b50f206d2d321ec7bb33727925f7e427c3382f4057171dd9ff61e08293b69cf +size 531694 diff --git a/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/full.md b/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/full.md new file mode 100644 index 0000000000000000000000000000000000000000..28b80a5fa809fe2c6d30dff194bbb9aa8b285cd0 --- /dev/null +++ b/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/full.md @@ -0,0 +1,308 @@ +# Test-time Adaptation for Foundation Medical Segmentation Model without Parametric Updates + +Kecheng Chen $^{1}$ , Xinyu Luo $^{1}$ , Tiexin Qin $^{1}$ , Jie Liu $^{1}$ , Hui Liu $^{1}$ + +Victor Ho Fun Lee², Hong Yan¹, Haoliang Li¹* + +$^{1}$ City University of Hong Kong $^{2}$ HKU {ck.ee@my.cityu.edu.hk} $\star$ : Corresponding author + +# Abstract + +Foundation medical segmentation models, with MedSAM being the most popular, have achieved promising performance across organs and lesions. However, MedSAM still suffers from compromised performance on specific lesions with intricate structures and appearance, as well as bounding box prompt-induced perturbations. Although current test-time adaptation (TTA) methods for medical image segmentation may tackle this issue, partial (e.g., batch normalization) or whole parametric updates restrict their effectiveness due to limited update signals or catastrophic forgetting in large models. Meanwhile, these approaches ignore the computational complexity during adaptation, which is particularly significant for modern foundation models. To this end, our theoretical analyses reveal that directly refining image embeddings is feasible to approach the same goal as parametric updates under the MedSAM architecture, which enables us to realize high computational efficiency and segmentation performance without the risk of catastrophic forgetting. Under this framework, we propose to encourage maximizing factorized conditional probabilities of the posterior prediction probability using a proposed distribution-approximated latent conditional random field loss combined with an entropy minimization loss. Experiments show that we achieve about $3\%$ Dice score improvements across three datasets while reducing computational complexity by over 7 times. Codes are available at here. + +# 1. Introduction + +Recently, foundation segmentation models for medical images have attracted massive attention from the medical imaging community [31], as medical image segmentation plays an indispensable role in many downstream tasks (e.g., surgical navigation and treatment planning). Medical Segment Anything Model [15] (MedSAM) is the most popular foundation segmentation model, which + +
Update TypeMethodDatasetsGFLOPs↓
OCODSCGM
Direct Inference87.7780.5756.67-
NormalizationTENT [25] (ICLR'21)86.2478.0854.34743.5
InTEnt [6] (CVPR'24)86.9582.4049.81744.8
GraTa [32] (AAAI'25)88.5780.4155.142975.6
PASS [30] (TMI'24)84.0571.8154.80745.8
FullMEMO [33] (NeurIPS'22)88.1280.0457.485646.5
CRF-SOD [24] (CVPR'23)88.9579.7956.811115.2
LatentOurs90.8484.1959.4997.3
+ +Table 1. Average Dice coefficient $(\uparrow)$ of different TTA methods on three datasets. Normalization-based methods are customized to BN layers, exhibiting marginal (even degraded) performance for LN-based foundation medical segmentation model. We report the computational complexity using GFLOPs (Giga Floating Point Operations), including forward and backward propagations. We leverage latent updates to achieve $\sim 7\times$ GFLOPs reduction compared with parametric update-based counterparts with performance gains. + +exhibits impressive universality across organs and lesions. Since MedSAM is pre-trained on large-scale medical image datasets, it typically enables the high-quality segmentation mask of the region of interest (RoI) in a zero-shot manner, which has contributed to extensive areas, e.g., semi/weak-supervised medical image segmentation [29] or anomaly detection [16] using provided segmentation masks as pseudo labels or RoIs. + +However, due to the extreme modality imbalance of medical images (where CT, MRI, and endoscopy images are dominant), MedSAM has compromised performance on less-represented modalities (e.g., eye fundus images) [15]. It is also difficult for MedSAM to maneuver specific lesions with intricate structures in the segmentation area [15], e.g., optic cup segmentation suffers from many vessel-like branching structures. Although supervised fine-tuning on specific modalities and lesions with annotated masks seems effective in tackling these issues [28], the overwhelming computational consumption of fine-tuning and inherent data scarcity (or privacy) concerns regarding annotated medical images may impede this process. Moreover, some test-time perturbations are still encountered, even though a well-calibrated and fine-tuned model has been prepared. For instance, + +mild looseness or shrinkage perturbations for bounding box prompts are sufficient to cause performance degradation for MedSAM (see Tables 2 and 3), which is clinically unfavorable as delineating a perfect prompt is usually labor-intensive and time-consuming for physicians. + +To address the above-mentioned issues, test-time adaptation (TTA) for foundation medical segmentation models is a feasible solution that enables performance gains on a specific unlabeled target (test) domain using the pre-trained model only. Although recent TTA approaches for medical image segmentation achieve promising results, we argue that scaling them to foundation segmentation models is difficult. First, most methods utilize various self-supervised objectives (e.g., prediction consistency over different augmentations [2, 6, 30]) to conduct batch normalization (BN)-centric fin-tuning, which may endure limited update signals with marginal adaptation as modern foundation models employ layer normalization (LN) without source-trained statistics like BN. Second, some approaches [27] update the whole model to explore more gains, but even minimal parameter adjustments in large models can precipitate catastrophic forgetting [10]. Third, most methods ignore computational complexity during the adaptation, which is particularly significant for foundation medical segmentation models1, since continuous forward and backward propagations on a large model are computationally intensive, especially on resource-constrained edge devices. Overall, existing methods struggle with maintaining a trade-off between sufficient update signals and knowledge preservation, and ignore computational consumption of the adaptation for large models. + +To this end, we propose a novel TTA framework for foundation medical segmentation models like MedSAM. Compared to existing methods, one of the most pivotal differences is that our proposed method updates latent representations rather than any model parameters², which can not only endow maximal update flexibility under a forgetting-free context but also circumvent heavy forward and backward computations on sophisticated architectures (e.g., image encoder). Specifically, by incorporating an off-the-shelf conditionally independent assumption between the input image X and the prediction Y in MedSAM, we provide a theoretical analysis to uncover that directly refining the image embedding Z (a.k.a. latent refinement) can encourage maximizing factorized conditional probabilities (including latent fidelity $P(\mathbf{Z}|\mathbf{X})$ and segmentation likelihood $P(\mathbf{Y}|\mathbf{Z}))$ of posterior prediction probability $P(\mathbf{Y}|\mathbf{X})$ , in a similar objective like parametric updates. Under this latent + +refinement framework, we devise a novel distributionapproximated latent conditional random field (CRF) loss to leverage useful cues in the input space for explicitly maximizing the posterior distribution of the latent fidelity. Meanwhile, we utilize entropy minimization to enforce the usability of the final predictions for implicitly maximizing segmentation likelihood. Finally, the mask decoder can directly impose the refined latent without inferring the image embedding again. Our contributions include three-fold: + +- We provide theoretical analyses to uncover that direct latent refinement is feasible to approach the same goal as parametric updates under the MedSAM architecture, which enables us to realize high computational efficiency and segmentation performance (refers to Table 1) without the risk of catastrophic forgetting. +- A distribution-approximated latent CRF loss based on a Maximum Mean Discrepancy-aware co-occurrent measure is proposed to introduce potential useful cues from input space, leading to sufficient statistics of image embedding over the input image. +- The latent refinement framework is validated on three multi-center medical image segmentation datasets with strong appearance variations and simulated prompt shifts. Extensive experiments demonstrate that our method can not only achieve average $\sim 3\%$ +- Dice score improvements among three datasets but also enjoy around $7\times$ computational complexity reduction compared with existing methods. + +# 2. Related works + +# 2.1. Medical Image Segmentation with TTA + +Various test-time adaptation (TTA) methodologies have been developed to address distribution shifts during inference [3, 17, 25]. Recent advancements have extended TTA frameworks to more complex tasks, including super-resolution [4] and blind image quality assessment [19]. In the field of medical image segmentation, TTA tackles the common multi-center issue [13], where different hospitals may have specific imaging parameters with strong appearance variations [26]. To this end, some methods conducted the optimization of the BN layer, manipulating either the affine-transformation parameters [30] or source-trained BN statistics [6], where many self-supervised objectives and sophisticated alignment of source-trained statistics in the BN layer are leveraged. Meanwhile, some approaches [27] fine-tune the overall model to explore more gains using elaborated objectives and knowledge accumulation strategies. We argue that such parametric updates suffer from restricted effectiveness due to limited BN-centric update signals and catastrophic forgetting for a large pretrained model, respectively. The computational com + +plexity is also ignored during adaptation. There are limited works specific to the TTA of foundation medical segmentation models. For example, Huang et al. [9] proposed an auxiliary online learning (AuxOL) and adaptive fusion for the foundation medical image segmentation model. However, AuxOL requires clinicians' annotations for model training, leading to potential unfeasibility in many scenarios. Beyond the medical image, Schön et al. [20] leveraged the clicked point prompts as cues for adaptation, which is inapplicable for MedSAM due to its bounding box-based prompts. + +Thus, there is a lack of a customized approach for the TTA problem of foundation medical segmentation models when considering both effectiveness and efficiency. + +# 2.2. Latent Refinement for TTA + +Typically, latent refinement is applied in the neural image compression community [1, 5, 14, 21, 23] to reduce the bit consumption of latent representation at compression time. Since the image compression task has a strong self-supervised signal that corresponds to reconstructing an original raw image by an encoder-decoder architecture [1, 5, 14, 21], it is more straightforward to leverage the latent refinement technique. Instead, there is no direct-supervised signal for medical image segmentation during inference, leaving unexplored space. To the best of our knowledge, we are the first to introduce the concept of latent refinement to test-time adaptation of medical image segmentation. More importantly, we provide theoretical analyses to uncover that directly refining the latent representation of the input image is feasible to approach the same goal as parametric update-based counterparts, which is not developed by previous literature. + +# 3. Methodology + +# 3.1. Preliminary and Theoretical Analysis + +This paper mainly discusses the common MedSAM architecture in [15] with the bounding box prompt. During the test-time adaptation, the bounding box prompt is provided and fixed but is omitted in the following derivations for notation simplification. + +Definition 1. (Architecture of MedSAM) Unlike U-Net and its variants [18], MedSAM has no skip connections between an image encoder $E_{\theta}$ parameterized by $\theta$ and a mask decoder $D_{\varphi}$ parameterized by $\varphi$ , which enables multiple interactions by reusing the image embedding. In light of this, the following proposition can be derived: + +Proposition 1. Let $\mathbf{X} \in \mathcal{X}$ , $\mathbf{Z} \in \mathcal{Z}$ , and $\mathbf{Y} \in \mathcal{Y}$ be an input image, its latent embedding, and its corresponding prediction. For a MedSAM model, $\mathbf{Y}$ is strictly conditionally independent of $\mathbf{X}$ given $\mathbf{Z}$ , i.e., $\mathbf{Y} \perp \mathbf{X} \mid \mathbf{Z}$ . + +Proof: Once the (deepest) latent embedding $\mathbf{Z}$ is calculated by $\mathbf{Z} = E_{\theta}(\mathbf{X})$ , the upsampling function $\mathbf{Y} =$ + +$D_{\varphi}(\mathbf{Z})$ is not directly related to the input image $\mathbf{X}$ or its any latent variants, leading to $\mathbf{Y}\perp \mathbf{X}\mid \mathbf{Z}$ + +Theorem 1. (General objective) Under maximum a posteriori (MAP) estimation, the objective of TTA for image segmentation is computing the optimal prediction $\mathbf{Y}^*$ by maximizing the posterior probability conditioned on $\mathbf{X}$ : + +$$ +\mathbf {Y} ^ {*} = D _ {\varphi^ {*}} \left(E _ {\theta^ {*}} (\mathbf {X})\right), \varphi^ {*}, \theta^ {*} = \arg \max _ {\theta , \varphi} [ P (\mathbf {Y} | \mathbf {X}; \theta , \varphi) ] ^ {3}. +$$ + +Various TTA strategies are regarded as implicit posterior maximization by self-supervised loss-based partial or overall parameter updates. Here, we incorporate the architecture of the MedSAM as a factorized precondition to yield an extension of Theorem 1 as follows. + +Corollary 1. (Factorized objective of MedSAM) The log-posterior $\log P(\mathbf{Y}|\mathbf{X})$ during test-time adaptation can be factorized into the sum of the segmentation likelihood $\log P(\mathbf{Y}|\mathbf{Z})$ and the latent fidelity term $\log P(\mathbf{Z}|\mathbf{X})$ based on $\mathbf{Y}\perp \mathbf{X}\mid \mathbf{Z}$ , i.e., + +$$ +\mathbf {Y} ^ {*} = D _ {\varphi^ {*}} \left(\mathbf {Z} ^ {*}\right), \mathbf {Z} ^ {*} = E _ {\theta^ {*}} (\mathbf {X}) +$$ + +$$ +\varphi^ {*}, \theta^ {*} = \arg \max _ {\varphi , \theta} [ \log P (\mathbf {Y} | \mathbf {Z}; \varphi) + \log P (\mathbf {Z} | \mathbf {X}; \theta) ], +$$ + +Proof: By recalling Prop. 1, $P(\mathbf{Y}|\mathbf{X}) \propto P(\mathbf{Y}|\mathbf{Z}) \cdot P(\mathbf{Z}|\mathbf{X})$ holds. Substituting this into the MAP objective in Theorem 1 yields the factorized objective above. Please refer to the Appendix for more details. + +# 3.2. Latent Refinement in MedSAM framework + +Problem. For Corollary 1, one of the commonly adopted solutions for achieving optimal prediction $\mathbf{Y}^*$ is directly to jointly optimize partial or whole parameters in $\{\varphi ,\theta \}$ , but two challenges are encountered as follows. + +- Partial model updates. Most existing approaches are BN-centric strategies, which are inapplicable for LN-based MedSAM. Although updating the affine transformation parameters in LN layers is feasible, the performance may be marginal due to restricted model modifications especially in strong domain shifts. +- Whole model updates. It is difficult to maintain a trade-off between TTA performance and catastrophic forgetting. Meanwhile, optimizing the overall model is computationally expensive due to the considerable number of parameters in modern large models. + +Solution. In light of this, we are interested in directly optimizing the latent embedding rather than the model parameters. Assume the encoder $E_{\theta}$ and the decoder $D_{\varphi}$ are fixed during TTA, this scheme can not only circumvent the above-mentioned challenges (i.e., it is likely to update more "parameters" in the forgetting-free context) but also be equivalent to theoretically achieving the + +![](images/1f61326352df930d53a2bb4300a8ce0ed762f2b9e4a917310b3e0e59c6740a5b.jpg) +Figure 1. The framework of the proposed method, including initial embedding inference, TTA via latent refinement, and final inference phases. Distribution-approximated latent CRF loss is visualized, where the loss between $\mathbf{z}_i$ and $\mathbf{z}_j$ is computed. + +same factorized objective of MedSAM as Corollary 1, + +$$ +\mathbf {Y} ^ {*} = D _ {\varphi} (\mathbf {Z} ^ {*}), +$$ + +$$ +\mathbf {Z} ^ {*} = \arg \max _ {\mathbf {Z}} \left[ \log P (\mathbf {Y} | \mathbf {Z}; \varphi) + \log P (\mathbf {Z} | \mathbf {X}; \theta) \right], \tag {1} +$$ + +where $\mathbf{Z} = E_{\theta}(\mathbf{X})$ can be initialized and refined iteratively. Typically, the total parameter number of $\mathbf{Z}$ is overly larger than that of the normalization layer, resulting in more update flexibility. Under frozen $\theta$ and $\varphi$ , there is no knowledge forgetting during optimization. These two perspectives are more likely to activate better TTA performance collaboratively than existing methods. Effectiveness of Latent Refinement during TTA. First, the $\log P(\mathbf{Z}|\mathbf{X})$ term encourages $\mathbf{Z}$ to maintain sufficient statistics over the input $\mathbf{X}$ , which is highly significant for the MedSAM framework during the TTA because 1) domain-specific variations at test time may degrade salient characteristics (e.g., segmentation cues like obvious edges or contours) of latent embedding, leading to suboptimal segmentation performance; 2) sufficient statistics relative to the input $\mathbf{X}$ can encourage the conditional independence assumption of MedSAM, i.e., $\mathbf{Y} \perp \mathbf{X} \mid \mathbf{Z}$ , to hold. Second, $\log P(\mathbf{Y}|\mathbf{Z})$ guarantees the usability of latent embedding for final predictions, preventing any collapsed solutions. + +Practical implementation. For Eq. (1), one would update the initial latent embedding $\mathbf{Z} = \mathbf{Z}^0 = E_\theta (\mathbf{X})$ with $m\geq 1$ steps, for each step $\mathbf{Z}^{m} =$ + +$$ +\mathbf {Z} ^ {m - 1} - \epsilon \cdot \frac {\partial [ - \log P (\mathbf {Y} ^ {m - 1} | \mathbf {Z} ^ {m - 1} ; \varphi) - \log P (\mathbf {Z} ^ {m - 1} | \mathbf {X} ; \theta ]}{\partial \mathbf {Z} ^ {m - 1}} \tag {2} +$$ + +where $\epsilon$ is the learning rate. The final prediction can be represented as $\mathbf{Y}^m = D_{\varphi}(\mathbf{Z}^m)$ . + +# 3.3. Distribution-Approximated Latent CRF + +To Eq. (2), one of the most important components is maximizing the log-probability $P(\mathbf{Z}^{m - 1}|\mathbf{X};\theta)$ , i.e., + +the latent fidelity term. For MedSAM, the volume of the image encoder parameter $\theta$ is considerable, resulting in a high computation during the forward and backward propagations in the TTA phase. Benefiting from our latent refinement framework, we propose circumventing the image encoder $\theta$ to directly enforce sufficient statistics over the input image $\mathbf{X}$ , i.e., maximizing $\log P(\mathbf{Z}^{m-1}|\mathbf{X})$ . By doing so, we can leverage potential useful cues in the input space, which is typically ignored by existing TTA methods that focus more on the label space (e.g., prediction consistency [2]). + +Therefore, we introduce the concept of conditional random field (CRF) [22], which is usually employed to refine the final predictions by enforcing the co-occurrence between the label space $\mathcal{V}$ and the input space $\mathcal{X}$ . For our purpose, we extend the CRF to the co-occurrence between the latent space $\mathcal{Z}$ and the input space $\mathcal{X}$ . Theoretically, minimizing the CRF energy is equivalent to maximizing the posterior probability of the random field [11], which coincides with our optimization objective in Eq. (2) via a computationally efficient manner. Formally, given a random field $\mathbf{Z} = E_{\theta}(\mathbf{X})$ over a set of vectors $\{\mathbf{z}_1,\mathbf{z}_2,\dots ,\mathbf{z}_{N'}\}$ , where $\mathbf{Z}\in \mathbb{R}^{W^{\prime}\times H^{\prime}\times C^{\prime}}$ is the latent embedding with a spatial resolution of $N^{\prime} = W^{\prime}\times H^{\prime}$ and a channel of $C^\prime$ , i.e., $\mathbf{z}\in \mathbb{R}^{1\times C^{\prime}}$ . Consider another random field $\mathbf{X}$ over a set of vectors $\{\mathbf{x}_1,\mathbf{x}_2,\dots ,\mathbf{x}_N\}$ , where $\mathbf{X}\in \mathbb{R}^{W\times H\times C}$ is the input intensities with a spatial resolution of $N = W\times H$ and a channel of $C$ , i.e., $\mathbf{x}\in \mathbb{R}^{1\times C}$ . + +Typically, $\mathbf{x}_i$ and $\mathbf{z}_i$ are related in a one-to-one manner, i.e., $\mathbf{x}_i$ represents a color vector at pixel $i$ in the input space and $\mathbf{z}_i$ is its corresponding predicted label in the label space. However, the input and latent spaces are dimensionally heterogeneous, i.e., $N \neq N'$ , which makes constructing a random field in this scenario challenging. Here, a novel distribution-approximated latent CRF (DAL-CRF) is proposed to eliminate the dimen + +sionally heterogeneous problem between the input and latent spaces. Specifically, by leveraging the fact that each latent vector $\mathbf{z}_i$ corresponds to a receptive field in the input space with a set of intensity vectors, it is feasible to identify related intensity vectors for each latent vector, i.e., + +$$ +\mathbf {z} _ {i} \rightarrow \mathbf {X} _ {i} = \left\{\mathbf {x} _ {i} ^ {l} \right\} _ {l = 1} ^ {L}, i = 1, 2, \dots , N ^ {\prime}, \tag {3} +$$ + +where $L$ denotes the number of intensity vectors in the receptive field, where $L = 256$ corresponding to a $16\times 16$ receptive field in MedSAM. Then, by following Veksler [24], the pairwise potential is only considered to be a regularization loss to encourage the minimization of CRF energy, which can be formulated as $\mathcal{L}_{\mathrm{DAL - CRF}} =$ + +$$ +\frac {1}{N ^ {\prime}} \sum_ {i, j \in \mathcal {G}} \left(\frac {1}{2} \operatorname {M M D} \left(\mathbb {P} \left(\mathbf {X} _ {i}\right), \mathbb {P} \left(\mathbf {X} _ {j}\right)\right)\right) \cdot \exp \left(- \| \mathbf {z} _ {i} - \mathbf {z} _ {j} \| _ {F}\right). \tag {4} +$$ + +Unlike existing CRF methods [24], which utilize the Gaussian kernel to evaluate the correlation between two color vectors, we introduce the Maximum Mean Discrepancy (MMD) to measure the distance between two vector sets. Specifically, each vector set is first approximated as an empirical distribution, i.e., $\mathbf{x}_i^l\sim \mathbb{P}(\mathbf{X}_i)$ . Then, each empirical distribution is represented by a kernel mean embedding in a reproducing kernel Hilbert space (RKHS) $\mathcal{H}$ , i.e., $\mu_{\mathbb{P}} = \mathbb{E}_{\mathbf{x}\sim \mathbb{P}}[\phi (\mathbf{x})] = \int_{\mathcal{X}}k(\mathbf{x},\cdot)d\mathbb{P} = \mathbb{E}_{\mathbf{x}\sim \mathbb{P}}[k(\mathbf{x},\cdot)]$ , where $k:\mathcal{X}\times \mathcal{X}\to \mathbb{R}$ is a reproducing kernel (Gaussian RBF kernel here) and corresponding feature map $\phi :\mathcal{X}\rightarrow \mathcal{H}$ . The discrepancy can be measured between mean embeddings, i.e., + +$$ +\operatorname {M M D} \left(\mathbb {P} \left(\mathbf {X} _ {i}\right), \mathbb {P} \left(\mathbf {X} _ {j}\right)\right) = \| \frac {1}{L} \sum_ {l = 1} ^ {L} \phi \left(\mathbf {x} _ {i} ^ {l}\right) - \frac {1}{L} \sum_ {l = 1} ^ {L} \phi \left(\mathbf {x} _ {j} ^ {l}\right) \| _ {\mathcal {H}} ^ {2}. \tag {5} +$$ + +MMD enables all-order correlation measurement using Gaussian kernel via unbiased empirical estimation. This may be useful for better discrepancy measurement, as the distribution of the intensity space may be highly complex and non-linear. For the second term in Eq. (4), we utilize the Frobenius norm $\| \cdot \| _F$ to calculate the feature compatibility in a co-occurrent probability using exponential zero-to-one normalization. + +Integrating Bounding Box Prompt. The graph $\mathcal{G} = \{\mathcal{V},\mathcal{E}\}$ in Eq. (4) is usually either a complete (dense-connected) graph [11] or a sparse position-related neighboring graph [24] on $\mathbf{Z}$ . Instead, we propose to consider the bounding box prompt as the prior to construct $K$ -random pairs inside the given bounding box, which can not only exclude those unimportant latent features far away from the bounding box prompt to enhance the computational efficiency but also focus more on the interested region with better performance. + +Intuitive Understanding. For Eq. (4), one actually encourages being more discriminative in latent representations between $\mathbf{z}_i$ and $\mathbf{z}_j$ . Since if two empirical distributions have a large distribution discrepancy (i.e., the MMD distance is large) in the input space, there are more penalties assigned to the pair who has larger similarity (i.e., co-occurrent probability is larger)) in the latent space. By minimizing $\mathcal{L}_{\mathrm{DAL - CRF}}$ , one can enforce the latent representations to equip with sufficient statistics as in the input space. + +# 3.4. Overall Objective at Test-time Adaptation + +Without accessing the ground-truth label, it is infeasible to maximize the segmentation likelihood $\log P(\mathbf{Y}^m|\mathbf{Z}^m;\varphi)$ directly. Therefore, we adopt a common predictive entropy minimization to implicitly achieve this goal [8]. Formally, + +$$ +\mathcal {L} _ {\mathrm {E M}} = H (P (\mathbf {Y} | \mathbf {Z})) = - \sum_ {h, v \in C} P (y ^ {h, v} | \mathbf {Z}) \log P (y ^ {h, v} | \mathbf {Z}), +$$ + +$$ +\text {w h e r e} \quad C = \left\{\left(h, v\right) \mid P \left(y ^ {h, v} \mid \mathbf {Z}\right) \geq \alpha \right\}, \tag {6} +$$ + +where $y^{h,v}$ denotes a single prediction in the position $(h,v)$ of $\mathbf{Y}$ . Eq. (6) is similar to foreground-background balanced entropy loss in Dong et al. [6]. However, we only utilize the foreground entropy and set a high confidence threshold (e.g., $\alpha = 0.95$ ) due to the following reasons. First, since the CRF-based loss usually requires a volumetric regularization to avoid the trivial solutions [24] (e.g., all labels are assigned to background), the entropy regularization over a high confidence foreground region can be interpreted as an implicit volumetric regularization, i.e., ensuring these foreground pixels confident enough for preventing trivial solutions. Second, we argue that background entropy is unnecessary for a foundation segmentation model, as these models have learned a good foreground-background balance. + +By considering the proposed distributionapproximated latent CRF loss and the customized entropy loss, the practical latent refinement scheme is reformulated as follows. + +$$ +\mathbf {Z} ^ {m} = \mathbf {Z} ^ {m - 1} - \epsilon \cdot \frac {\partial \left[ \mathcal {L} _ {\mathrm {D A L - C R F}} + \lambda \mathcal {L} _ {\mathrm {E M}} \right]}{\partial \mathbf {Z} ^ {m - 1}}, \tag {7} +$$ + +where $\lambda$ is a balance coefficient. + +# 3.5. Implementation Details + +The overall framework is illustrated in Figure 1. Superficially, the initial image embedding $\mathbf{Z}^0$ is inferred only once. Once the refined latent representation $\mathbf{Z}^*$ is obtained, the upsampling can be conducted directly without requiring image embedding again. This mechanism is well compatible with the interactive scheme of the MedSAM. We implement the proposed method based on + +
Update TypeMethodDice Coefficient ↑Average Surface Distance ↓
D1D2D3D4Avg.D1D2D3D4Avg.
Looseness promptsDirect Inference90.9987.1687.0684.7287.489.487.518.747.628.33
Normalization updateTENT (ICLR'21)91.2785.4085.0681.8485.899.098.7010.058.889.18
InTEnt (CVPR'24)91.3687.4387.4786.1387.608.807.518.267.668.06
GraTa (AAAI'25)90.8987.0086.9584.6487.379.487.548.807.658.37
PASS (TMI'24)80.8679.6382.0384.0281.6320.4012.2211.697.9713.07
Full updatesMEMO (NeurIPS'22)87.2685.4886.4284.4785.9112.458.279.157.769.40
CRF-SOD (CVPR'23)89.5784.0184.8082.5885.2410.548.9910.188.649.58
Latent updatesOurs89.3588.7389.8388.1889.0211.406.696.926.077.77
Shrinkage promptsDirect Inference81.0567.6072.1068.5772.3318.5117.6117.7614.9217.20
Normalization updateTENT (ICLR'21)80.2265.1769.2562.1069.1818.8519.2619.3917.5818.77
InTEnt (CVPR'24)82.7270.0674.5073.4975.1916.8615.2816.1313.8715.53
GraTa (AAAI'25)80.8767.3271.9468.4672.1517.2716.5416.5314.5216.21
PASS (TMI'24)66.8158.6663.3156.7861.3932.3922.8922.7420.1924.55
Full updatesMEMO (NeurIPS'22)81.0267.5472.1068.8472.3718.5217.7617.7714.7817.20
CRF-SOD (CVPR'23)81.8469.2674.7372.0874.4819.2718.8318.6314.7317.86
Latent updatesOurs83.0773.7278.5475.7777.7716.7414.3813.7211.2614.02
Perfectness promptsDirect Inference87.5880.2581.3278.4681.9012.7611.2112.2710.2411.62
Normalization updateTENT (ICLR'21)87.6077.9878.1872.9979.1812.8312.2113.9212.6912.91
InTEnt (CVPR'24)86.5484.2585.1781.6584.4010.989.239.858.959.75
GraTa (AAAI'25)87.3979.9781.1478.3681.7112.9611.3812.3810.2911.75
PASS (TMI'24)74.5969.8673.8871.3472.4125.5317.3216.5813.8518.32
Full updatesMEMO (NeurIPS'22)87.3980.1881.3278.4981.8412.9711.2412.2710.2311.67
CRF-SOD (CVPR'23)86.0777.2279.1076.2479.6513.9612.5913.6311.212.84
Latent updatesOurs86.8585.1186.7284.4985.7913.738.638.797.489.65
+ +Table 2. Quantitative comparison of TTA results on fundus image OC segmentation between different methods. The best and second-best results are shown in the bold and the underline, respectively. For loose prompts, we report the average of two looseness ratios. For shrinkage prompts, we report the average of two shrinkage ratios. Full results can be found in the Appendix. + +official MedSAM checkpoints (ViT-B pre-trained models) and codes using a one-step optimization strategy $(m = 1)$ by following previous TTA settings. Due to the extremely various sizes of RoIs in medical images, we leverage the initial segmentation result $\mathbf{Y}^0$ and the given bounding box prompt to explore a dynamic adjustment strategy for important hyperparameters. Specifically, the learning rate $\epsilon$ with the Adam optimizer is linearly interpolated from $5\times 10^{-2}$ to $1\times 10^{-1}$ according to the ratio of the initial foreground to the entire image $f_{e}$ (i.e., the rough size of the RoI), where $f_{e}\leq 0.05$ and $f_{e}\geq 0.2$ are fixedly set to $5\times 10^{-2}$ and $1\times 10^{-1}$ , respectively. Note that we use commonly-used definitions over the $f_{e}$ as adopted by Gao et al. [7], i.e., $f_{e}\leq 0.05$ , $0.05 < f_{e} < 0.2$ , and $f_{e}\geq 0.2$ as small, medium, and large RoIs. Similarly, the $\lambda$ is also RoI-specific adjustment, where the small RoI $(f_{e}\leq 0.05)$ requires a similar contribution between the $\mathcal{L}_{\mathrm{DAL - CRF}}$ and the $\mathcal{L}_{\mathrm{EM}}$ , thus the $\lambda$ is set to $1\times 10^{3}$ . The $K$ is set to 4 by balancing computational consumption and performance (refers to Table 5(a)). $\alpha$ in entropy loss is set to 0.95 empirically. + +# 4. Experiments + +Datasets and Evaluation Metrics We evaluate our method on two public multi-center medical image segmentation datasets with different data modalities. (1) Fundus [26]: Fundus datasets are retinal fundus images + +collected from 4 healthcare centers (D1 to D4 for short). For each retinal fundus image, there are two segmentation objectives, including the optic cup (OC) (mainly small RoIs) and optic disc (OD) (mainly medium or large RoIs). (2) Spinal Cord Gray Matter (SCGM) [12]: SCGM datasets are collected from 4 institutions (D1 to D4 for short) with different imaging manufacturers for spinal MRI image gray matter segmentation. We use the Dice Similarity Coefficient (DSC) and Average Surface Distance (ASD) to measure the accuracy. + +Experiment Settings. Appearance Perturbations: Note that our adopted multi-center datasets have strong domain shifts with appearance variations, we thus ignore test-time appearance perturbations (e.g., changes in contrast, noise, and blur). Prompt Perturbations: We mainly consider three kinds of bounding box prompts to evaluate TTA performance. (i) Perfectness prompt $(PP)$ : The perfect prompt can be obtained by groundtruth (GT) mask, where the top-left and lower-right coordinates of the PP correspond to the minimal and maximal horizontal and vertical coordinates of the GT mask, respectively. A unique PP is utilized for TTA. (ii) Looseness prompts $(LP)$ : Clinically, it is labor-intensive and time-consuming to obtain the PP. Instead, a moderately loose prompt can be efficient but may trigger the degradation of the model with the need for TTA. Therefore, we define the looseness rate $\sigma$ , which is the ratio of the coordinate shift relative to the width $w$ and + +
Update TypeMethodDice Coefficient ↑Average Surface Distance ↓
D1D2D3D4Avg.D1D2D3D4Avg.
Looseness promptsDirect Inference83.0578.8681.8275.5479.8133.2032.8730.2931.9032.06
Normalization updateTENT (ICLR'21)81.6176.9680.0074.8578.3537.5836.5433.5334.3535.50
InTEnt (CVPR'24)83.6176.7180.0071.8378.0332.4637.0132.5739.1435.29
GraTa (AAAI'25)84.6480.0484.1082.8282.9030.0230.9325.9128.9128.94
PASS (TMI'24)84.5781.7782.2977.3481.4918.0226.2021.0924.6822.49
Full model updatesMEMO (NeurIPS'22)86.2680.0183.9478.1882.1026.7330.9326.2728.5428.12
CRF-SOD (CVPR'23)87.0382.8683.3379.5283.1825.1926.2523.7226.5325.42
Latent updatesOurs90.8583.5587.7782.2486.1017.5923.5418.5021.7520.34
Shrinkage promptsDirect Inference92.7086.1790.9088.6789.6112.6017.6612.7712.1013.78
Normalization updateTENT (ICLR'21)90.6385.6389.6686.8888.2015.9018.4414.2712.4815.27
InTEnt (CVPR'24)90.9187.2790.8888.7089.4415.4916.4712.6812.0514.17
GraTa (AAAI'25)92.6185.6091.0088.7189.4812.7018.3212.6112.1113.93
PASS (TMI'24)91.4678.8585.9275.7082.9813.7225.1417.5423.0519.86
Full model updatesMEMO (NeurIPS'22)92.1283.8290.8788.7788.8913.5220.1812.7512.0414.62
CRF-SOD (CVPR'23)94.4284.0792.3491.4390.569.3619.8310.709.3012.30
Latent updatesOurs94.2486.7793.2392.4291.669.8016.259.458.2110.93
Perfectness promptsDirect Inference96.4992.8793.8992.3693.906.099.488.648.368.14
Normalization updateTENT (ICLR'21)94.2090.7893.3990.3892.1810.1212.019.3210.3910.46
InTEnt (CVPR'24)94.1291.5794.4293.4793.3910.1211.018.017.039.04
GraTa (AAAI'25)94.4692.6693.9192.3193.336.149.738.618.428.22
PASS (TMI'24)94.9085.3584.3686.1487.688.5817.5217.6213.9814.42
Full model updatesMEMO (NeurIPS'22)95.7791.4893.8992.3193.367.8011.288.618.419.02
CRF-SOD (CVPR'23)96.6691.1593.7693.4193.755.7811.428.737.268.29
Latent updatesOurs96.6892.6795.1594.5794.765.699.516.855.987.00
+ +Table 3. Quantitative comparison of TTA results on fundus image OD segmentation between different methods. The best and second-best results are shown in the bold and the underline, respectively. For loose prompts, we report the average of two looseness ratios. For shrinkage prompts, we report the average of two shrinkage ratios. Full results can be found in the Appendix. + +
Update TypeMethodDice Coefficient ↑Average Surface Distance ↓
D1D2D3D4Avg.D1D2D3D4Avg.
Direct Inference55.3554.2061.5456.6756.9410.2211.3915.2715.9513.20
Normalization updateTENT (ICLR'21)49.6652.9659.3955.3854.349.9410.4315.8516.1213.08
InTEnt (CVPR'24)50.9152.1245.8550.3849.819.3510.9922.5719.0115.48
GraTa (AAAI'25)56.3555.1463.6658.1258.319.2610.4414.3615.1512.30
PASS (TMI'24)55.2454.8061.9452.8856.2110.5811.0115.0220.1413.99
Full updatesMEMO (NeurIPS'22)56.2055.2061.2057.3457.489.2710.4514.8915.3912.50
CRF-SOD (CVPR'23)54.2054.0861.8957.0756.8110.3710.8315.3615.5813.03
Latent updatesOurs57.9255.4266.0058.6259.498.3510.5814.1414.1511.75
+ +Table 4. Quantitative comparison of TTA results on MRI image SCGM segmentation between different methods. SCGM segmentation only reported PP-based results due to its very challenging intricate structures and appearance perturbations. + +height $h$ of the PP, to simulate real-world loose scenarios as follows: $x(y)_{tl}^{LP}: x(y)_{tl} - \sigma \cdot w(h), x(y)_{lr}^{LP}: x(y)_{lr} + \sigma \cdot w(h)$ , where $x(y)_{tl}$ and $x(y)_{lr}$ denote the vertical (horizontal) coordinates of the top-left and lower-right points of the PP. Each loose prompt is represented as $(x_{tl}^{LP}, y_{tl}^{LP}, x_{lr}^{LP}, y_{lr}^{LP})$ . Here, $\sigma$ is set to 0.1 and 0.2, as we observe 1) too small $\sigma (< 0.1)$ is close to the PP without obvious changes and 2) too large $\sigma (>0.2)$ will usually result in significantly overlaps with other RoIs with unacceptable initial results. We report the average of different loose rates for comprehensive evaluation. (iii) Shrinkage prompts (SP): Similarly, we define the shrinkage ratio $\gamma$ to simulate the scenario that the bounding box prompt shifts to the interior of the PP as follows: $x(y)_{tl}^{SP}: x(y)_{tl} + \gamma \cdot w(h), x(y)_{lr}^{SP}: x(y)_{lr} - \gamma \cdot w(h)$ . Each shrinkage prompt is represented + +as $(x_{tl}^{SP},y_{tl}^{SP},x_{lr}^{SP},y_{lr}^{SP})$ . Similarly, $\gamma$ is set to 0.1 and 0.2. We report the average of different shrinkage rates for comprehensive evaluation. + +Baselines. In this study, two kinds of TTA approaches for medical image segmentation are utilized for comparison. (1) Normalization-based TTA: Most baseline methods conduct the update of affine transformation parameters and manipulation of source-trained statistics in BN layers, which the MedSAM lacks due to the usage of LN layers. We, therefore, discard source-trained statistics-related components and only optimize the affine-transformation parameters for baseline methods, including TENT (entropy minimization), InTEnt (foreground-background-balanced entropy weighting), GraTa (augmentation-based gradient alignment), and PASS (input decora + +![](images/8335dfc2ddad3f2cf9acaacdc3c84db4ed78b5f2dbab02b8e0637480f57b225b.jpg) + +![](images/b8854cfa7bdad07e0cbc26f9997a562d607cbcd164d539630d8eb048a3d29057.jpg) + +![](images/48cc3745ba6e2555f858576859fe914657a51126e49bd64a3eb8271db1be7070.jpg) + +![](images/aac781c46819eefe84498bfe43ad6bc4575f42090109af93f3e13f29b658ea43.jpg) + +![](images/f2b83411b3b336fc2918554b16c1b2164441e88e422fdc837eafe03cb34d31ad.jpg) + +![](images/2056b3dc42a43ba70e316950ad6da654ecf0490c1dcf1c7e1995abd38ad1a7c9.jpg) + +![](images/b3fa64e6721c9ecdfe38623fb6d45b7c6f2a0c3497ab809793946259341c1a6d.jpg) + +![](images/a329bc565350d7c16c811867415bc5f3c649dfb2f25ea812c4e3cfb40418fab6.jpg) + +![](images/454e1cd0268541b851d7ef1cc044798c0cc936cefcadc7e2fa52e6e587ec0b18.jpg) + +![](images/177140834aa9cfb310e60a7642480a327e71a0422358f2adbc12ef36337b3ebf.jpg) + +![](images/410503a618ac85d9ce271c839342f438d16d31253fff6c29ee0b59dda936fc30.jpg) +Figure 2. Visualized process of latent refinement with iterations. The first, second, and third rows denote the segmentation result, corresponding visualization of channel-average refined latent embedding (which is from the region in bounding box prompt and has been fused with the prompt for a better view), and corresponding DAL-CRF loss for each spatial position. You may zoom in for a better view. + +![](images/5eb53ed1ccbaf88dc2a92044519d2d6cb845941a4cae5de354b426b84914977c.jpg) + +![](images/9c99c923fc7a2dbda4ce6d619541d9cd9be0e5a1497ef1f5d31adc1ef218b523.jpg) + +![](images/5913bd22dec9afd75d38b5010f23e210b39c4b79b82d02a82f9a01f979e6a487.jpg) + +![](images/9315e7c096c20f395efbab305ec890adf41a33bd637ba0792d755cc1bfde0acf.jpg) + +(a) + +
KD1D2D3D4Avg.
293.2487.0692.5186.5189.75
494.1687.1691.8387.5790.14
694.6787.0892.3487.9290.37
894.8686.5490.8188.0190.03
+ +(b) + +
LossD1D2D3D4Avg.
DI89.1984.3187.9282.3985.95
\( \mathcal{L}_{\text{InTEnt}} \)79.3484.8988.8386.0884.78
\( \mathcal{L}_{\text{TENT}} \)85.0984.9989.4685.5986.28
\( \mathcal{L}_{\text{CRF-SOD}} \)88.9386.1788.4287.3387.71
\( \mathcal{L}_{\text{DAL-CRF}} \)94.1686.9690.8587.5789.88
+ +Table 5. (a) Hyperparameter Analyses in terms of $K$ . (b) Integration of the latent refinement framework and other TTA loss functions $\nu$ s. Our proposed DAL-CRF, where DI denotes direct inference without adaptation on OD segmentation. + +
MethodD1D2D3D4Avg.
Direct Inference89.1984.3187.9282.3985.95
+ LEM85.8285.3188.1083.2186.01
+ LDAL-CRF94.1686.9690.8587.5789.88
+ LEM + LDAL-CRF94.1687.1691.8387.4290.14
DenseCRF [11]-Postprocessing82.7079.3980.9073.7679.18
+ +Table 6. Ablation Study on Different Components of the OD segmentation with loose prompts, in terms of Dice score. + +tor and alternating momentum updating strategy). (2) Full update-based TTA: The whole model parameters are updated based on customized losses, including MEMO (augmentation-based marginal entropy minimization) and CRF-SOD (sparse CRF regularization loss). We leverage their open-source codes to apply into the architecture of MedSAM. To be fair, all baseline methods adopt a one-step optimization strategy. + +Quantitative Analysis. From Tables 2, 3, and 4, some conclusions can be summarized. (1) Our proposed method without parametric updates surpasses all baselines in most cases with the best or second-best performance, showcasing the effectiveness of latent refinement. (2) It seems that full parameter updates are more suitable to large size RoIs than normalization-based updates, e.g., the OD. (3) For very challenging OC and SCGM tasks, most parametric update-based approaches have compromised performance compared with direct inference, but our proposed method performs well. + +Qualitative Analysis. From Figure 2, our proposed method can achieve gradually better TTA segmentation with iterations (Note: we reduce the learning rate with multistep optimization rather than a one-step manner for understanding the optimization process better). Meanwhile, the latent embedding is gradually unsmoothed, which may be reasonable as the DAL-CRF loss enforces more statistics over the input image, leading to more discriminative information. The gradual decrease of DAL- + +CRF loss in each spatial point of the latent embedding is also observed, showcasing its effectiveness. + +Computation Complexity. As illustrated in Table 1, we achieve around $7 \times$ GFLOPs reduction compared with common usages. This is reasonable as 1) sophisticated augmentations (e.g., MEMO) and multiple gradient backward processes (e.g., GraTa) are computationally intensive. 2) For normalization-based updates, the gradient computation over normalization layers still passes through the entire model during backpropagation, which is unfavorable as MedSAM's image encoder is sophisticated with considerable consumption. 3) We circumvent any augmentation and the forward/backward process to the image encoder in a more efficient manner. + +Ablation Study & Other TTA losses under the latent refinement framework. As shown in Table 6, $\mathcal{L}_{\mathrm{DAL - CRF}}$ significantly contributes more compared to entropy minimization loss and simply post-processing predictions fail to achieve acceptable performance on challenging medical images with strong appearance perturbations. Meanwhile, $\mathcal{L}_{\mathrm{DAL - CRF}}$ still performs better, even though other TTA losses are integrated into the latent refinement framework as shown in 5(b). Compared to the CRF-SOD (using input and output spaces), our DAL-CRF exhibits superiority using the latent-based co-occurrence regularization for foundation models. + +# 5. Conclusion + +We propose a novel TTA framework for foundation medical segmentation without parametric updates. Our theoretical analysis shows that direct latent refinement is feasible within the MedSAM architecture. By maximizing factorized conditional probabilities by our DAL-CRF loss combined with entropy minimization, we achieve better performance for effectiveness and efficiency. + +# Acknowledgement + +This work was supported in part by the Hong Kong Innovation and Technology Commission (ITC) (InnoHK Project CIMDA), in part by the Institute of Digital Medicine of City University of Hong Kong (CityU) under Project 9229503, in part by Chow Sang Sang Donation and Matching Fund (Project 9229161), and in part by the CityU under Project 7006087. + +# References + +[1] Kecheng Chen, Pingping Zhang, Tiexin Qin, Shiqi Wang, Hong Yan, and Haoliang Li. Test-time adaptation for image compression with distribution regularization. In The Thirteenth International Conference on Learning Representations, 2025. 3 +[2] Ziyang Chen, Yiwen Ye, Yongsheng Pan, and Yong Xia. Gradient alignment improves test-time adaptation for medical image segmentation. The 39th Annual AAAI Conference on Artificial Intelligence, 2025. 2, 4 +[3] Mohammad Zalbagi Darestani, Jiayu Liu, and Reinhard Heckel. Test-time training can close the natural distribution shift performance gap in deep learning based compressed sensing. In International Conference on Machine Learning, pages 4754-4776. PMLR, 2022. 2 +[4] Zeshuai Deng, Zhuokun Chen, Shuaicheng Niu, Thomas Li, Bohan Zhuang, and Mingkui Tan. Efficient test-time adaptation for super-resolution with second-order degradation and reconstruction. Advances in Neural Information Processing Systems, 36:74671-74701, 2023. 2 +[5] JCMSA Djelouah and Christopher Schroers. Content adaptive optimization for neural image compression. In Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit, pages 1-5, 2019. 3 +[6] Haoyu Dong, Nicholas Konz, Hanxue Gu, and Maciej A Mazurowski. Medical image segmentation with intent: Integrated entropy weighting for single image test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5046-5055, 2024. 1, 2, 5 +[7] Yunhe Gao, Rui Huang, Yiwei Yang, Jie Zhang, Kainan Shao, Changjuan Tao, Yuanyuan Chen, Dimitris N. Metaxas, Hongsheng Li, and Ming Chen. Focusnetv2: Imbalanced large and small organ segmentation with adversarial shape constraint for head and neck ct images. Medical Image Analysis, 67:101831, 2021. 6 +[8] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 5 +[9] Tianyu Huang, Tao Zhou, Weidi Xie, Shuo Wang, Qi Dou, and Yizhe Zhang. Improving segment anything on the fly: Auxiliary online learning and adaptive fusion for medical image segmentation. arXiv preprint arXiv:2406.00956, 2024.3 +[10] Yiming Ju, Ziyi Ni, Xingrun Xing, Zhixiong Zeng, Hanyu Zhao, Siqi Fan, and Zheng Zhang. Mitigating + +training imbalance in LLM fine-tuning via selective parameter merging. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15952-15959, Miami, Florida, USA, 2024. Association for Computational Linguistics. 2 +[11] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 4, 5, 8 +[12] Haoliang Li, YuFei Wang, Renjie Wan, Shiqi Wang, TieQiang Li, and Alex Kot. Domain generalization for medical imaging classification with linear-dependency regularization. Advances in Neural Information Processing Systems, 33:3118-3129, 2020. 6 +[13] Zhenbing Liu, Fengfeng Wu, Yumeng Wang, Mengyu Yang, and Xipeng Pan. Fedcl: Federated contrastive learning for multi-center medical image classification. Pattern Recognition, 143:109739, 2023. 2 +[14] Yue Lv, Jinxi Xiang, Jun Zhang, Wenming Yang, Xiao Han, and Wei Yang. Dynamic low-rank instance adaptation for universal neural image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 632-642, 2023. 3 +[15] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 3 +[16] Yuchen Mao, Hongwei Bran Li, LAI Yinyi, Giorgos Papanastasiou, Peng Qi, Yunjie Yang, and Chengjia Wang. Semi-supervised medical image segmentation via knowledge mining from large models. 1 +[17] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 2 +[18] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3 +[19] Subhadeep Roy, Shankhanil Mitra, Soma Biswas, and Rajiv Soundararajan. Test time adaptation for blind image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16742-16751, 2023. 2 +[20] Robin Schon, Julian Lorenz, Katja Ludwig, and Rainer Lienhart. Adapting the segment anything model during usage in novel situations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3616-3626, 2024. 3 +[21] Sheng Shen, Huanjing Yue, and Jingyu Yang. Decap dapter: Exploring efficient decoder-side adapter for bridging screen content and natural image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12887-12896, 2023. 3 +[22] Charles Sutton, Andrew McCallum, et al. An introduction to conditional random fields. Foundations and Trends® in Machine Learning, 4(4):267-373, 2012. 4 + +[23] Koki Tsubota, Hiroaki Akutsu, and Kiyoharu Aizawa. Universal deep image compression via content-adaptive optimization with adapters. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2529-2538, 2023. 3 +[24] Olga Veksler. Test time adaptation with regularized loss for weakly supervised salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7360-7369, 2023. 1, 5 +[25] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 1, 2 +[26] Shujun Wang, Lequan Yu, Kang Li, Xin Yang, Chi-Wing Fu, and Pheng-Ann Heng. Dofe: Domain-oriented feature embedding for generalizable fundus image segmentation on unseen datasets. IEEE Transactions on Medical Imaging, 39(12):4237-4248, 2020. 2, 6 +[27] Ruxue Wen, Hangjie Yuan, Dong Ni, Wenbo Xiao, and Yaoyao Wu. From denoising training to test-time adaptation: Enhancing domain generalization for medical image segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 464-474, 2024. 2 +[28] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 1 +[29] Guoping Xu, Xiaoxue Qian, Hua Chieh Shao, Jax Luo, Weiguo Lu, and You Zhang. A sam-guided and match-based semi-supervised segmentation framework for medical imaging. arXiv preprint arXiv:2411.16949, 2024. 1 +[30] Chuyan Zhang, Hao Zheng, Xin You, Yefeng Zheng, and Yun Gu. Pass: test-time prompting to adapt styles and semantic shapes in medical image segmentation. IEEE Transactions on Medical Imaging, 2024. 1, 2 +[31] Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Yejia Zhang, and Danny Z Chen. Input augmentation with sam: Boosting medical image segmentation with segmentation foundation model. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2023. 1 +[32] Chen, Ziyang and Ye, Yiwen and Pan, Yongsheng and Xia, Yong. Gradient Alignment Improves Test-Time Adaptation for Medical Image Segmentation. In AAAI, 2025. 1 +[33] Zhang, Marvin and Levine, Sergey and Finn, Chelsea. Memo: Test time robustness via adaptation and augmentation Segmentation. In Advances in neural information processing systems, pages 38629-38642, 2022. 1 \ No newline at end of file diff --git a/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/images.zip b/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b0a9c853f7790c325492e76e1abda9cfa714c887 --- /dev/null +++ b/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1485a3354619850e0cc6164035e2dd94c9b43985c2d9ba57285dc54a7953a18e +size 756549 diff --git a/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/layout.json b/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..41d7317c99848f303220e9459f9f65b410d4684e --- /dev/null +++ b/2025/Test-time Adaptation for Foundation Medical Segmentation Model Without Parametric Updates/layout.json @@ -0,0 +1,9293 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 536, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 536, + 139 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 536, + 139 + ], + "type": "text", + "content": "Test-time Adaptation for Foundation Medical Segmentation Model without Parametric Updates" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 160, + 459, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 160, + 459, + 175 + ], + "spans": [ + { + "bbox": [ + 153, + 160, + 459, + 175 + ], + "type": "text", + "content": "Kecheng Chen" + }, + { + "bbox": [ + 153, + 160, + 459, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 459, + 175 + ], + "type": "text", + "content": ", Xinyu Luo" + }, + { + "bbox": [ + 153, + 160, + 459, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 459, + 175 + ], + "type": "text", + "content": ", Tiexin Qin" + }, + { + "bbox": [ + 153, + 160, + 459, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 459, + 175 + ], + "type": "text", + "content": ", Jie Liu" + }, + { + "bbox": [ + 153, + 160, + 459, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 153, + 160, + 459, + 175 + ], + "type": "text", + "content": ", Hui Liu" + }, + { + "bbox": [ + 153, + 160, + 459, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 175, + 426, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 175, + 426, + 189 + ], + "spans": [ + { + "bbox": [ + 187, + 175, + 426, + 189 + ], + "type": "text", + "content": "Victor Ho Fun Lee², Hong Yan¹, Haoliang Li¹*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 78, + 189, + 534, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 189, + 534, + 203 + ], + "spans": [ + { + "bbox": [ + 78, + 189, + 534, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 189, + 534, + 203 + ], + "type": "text", + "content": "City University of Hong Kong " + }, + { + "bbox": [ + 78, + 189, + 534, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 189, + 534, + 203 + ], + "type": "text", + "content": "HKU {ck.ee@my.cityu.edu.hk} " + }, + { + "bbox": [ + 78, + 189, + 534, + 203 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 78, + 189, + 534, + 203 + ], + "type": "text", + "content": ": Corresponding author" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 159, + 236, + 206, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 236, + 206, + 249 + ], + "spans": [ + { + "bbox": [ + 159, + 236, + 206, + 249 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 261, + 297, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 261, + 297, + 597 + ], + "spans": [ + { + "bbox": [ + 67, + 261, + 297, + 597 + ], + "type": "text", + "content": "Foundation medical segmentation models, with MedSAM being the most popular, have achieved promising performance across organs and lesions. However, MedSAM still suffers from compromised performance on specific lesions with intricate structures and appearance, as well as bounding box prompt-induced perturbations. Although current test-time adaptation (TTA) methods for medical image segmentation may tackle this issue, partial (e.g., batch normalization) or whole parametric updates restrict their effectiveness due to limited update signals or catastrophic forgetting in large models. Meanwhile, these approaches ignore the computational complexity during adaptation, which is particularly significant for modern foundation models. To this end, our theoretical analyses reveal that directly refining image embeddings is feasible to approach the same goal as parametric updates under the MedSAM architecture, which enables us to realize high computational efficiency and segmentation performance without the risk of catastrophic forgetting. Under this framework, we propose to encourage maximizing factorized conditional probabilities of the posterior prediction probability using a proposed distribution-approximated latent conditional random field loss combined with an entropy minimization loss. Experiments show that we achieve about " + }, + { + "bbox": [ + 67, + 261, + 297, + 597 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 67, + 261, + 297, + 597 + ], + "type": "text", + "content": " Dice score improvements across three datasets while reducing computational complexity by over 7 times. Codes are available at here." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 69, + 618, + 149, + 631 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 618, + 149, + 631 + ], + "spans": [ + { + "bbox": [ + 69, + 618, + 149, + 631 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 639, + 296, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 639, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 67, + 639, + 296, + 723 + ], + "type": "text", + "content": "Recently, foundation segmentation models for medical images have attracted massive attention from the medical imaging community [31], as medical image segmentation plays an indispensable role in many downstream tasks (e.g., surgical navigation and treatment planning). Medical Segment Anything Model [15] (MedSAM) is the most popular foundation segmentation model, which" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 318, + 236, + 545, + 318 + ], + "blocks": [ + { + "bbox": [ + 318, + 236, + 545, + 318 + ], + "lines": [ + { + "bbox": [ + 318, + 236, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 318, + 236, + 545, + 318 + ], + "type": "table", + "html": "
Update TypeMethodDatasetsGFLOPs↓
OCODSCGM
Direct Inference87.7780.5756.67-
NormalizationTENT [25] (ICLR'21)86.2478.0854.34743.5
InTEnt [6] (CVPR'24)86.9582.4049.81744.8
GraTa [32] (AAAI'25)88.5780.4155.142975.6
PASS [30] (TMI'24)84.0571.8154.80745.8
FullMEMO [33] (NeurIPS'22)88.1280.0457.485646.5
CRF-SOD [24] (CVPR'23)88.9579.7956.811115.2
LatentOurs90.8484.1959.4997.3
", + "image_path": "347ac5eb387cc0f60aa3964b10faa3ddae8b15b1101414eca83eff5e22822baa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 326, + 541, + 426 + ], + "lines": [ + { + "bbox": [ + 313, + 326, + 541, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 541, + 426 + ], + "type": "text", + "content": "Table 1. Average Dice coefficient " + }, + { + "bbox": [ + 313, + 326, + 541, + 426 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 313, + 326, + 541, + 426 + ], + "type": "text", + "content": " of different TTA methods on three datasets. Normalization-based methods are customized to BN layers, exhibiting marginal (even degraded) performance for LN-based foundation medical segmentation model. We report the computational complexity using GFLOPs (Giga Floating Point Operations), including forward and backward propagations. We leverage latent updates to achieve " + }, + { + "bbox": [ + 313, + 326, + 541, + 426 + ], + "type": "inline_equation", + "content": "\\sim 7\\times" + }, + { + "bbox": [ + 313, + 326, + 541, + 426 + ], + "type": "text", + "content": " GFLOPs reduction compared with parametric update-based counterparts with performance gains." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 312, + 434, + 541, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 434, + 541, + 530 + ], + "spans": [ + { + "bbox": [ + 312, + 434, + 541, + 530 + ], + "type": "text", + "content": "exhibits impressive universality across organs and lesions. Since MedSAM is pre-trained on large-scale medical image datasets, it typically enables the high-quality segmentation mask of the region of interest (RoI) in a zero-shot manner, which has contributed to extensive areas, e.g., semi/weak-supervised medical image segmentation [29] or anomaly detection [16] using provided segmentation masks as pseudo labels or RoIs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 531, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 531, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 312, + 531, + 541, + 723 + ], + "type": "text", + "content": "However, due to the extreme modality imbalance of medical images (where CT, MRI, and endoscopy images are dominant), MedSAM has compromised performance on less-represented modalities (e.g., eye fundus images) [15]. It is also difficult for MedSAM to maneuver specific lesions with intricate structures in the segmentation area [15], e.g., optic cup segmentation suffers from many vessel-like branching structures. Although supervised fine-tuning on specific modalities and lesions with annotated masks seems effective in tackling these issues [28], the overwhelming computational consumption of fine-tuning and inherent data scarcity (or privacy) concerns regarding annotated medical images may impede this process. Moreover, some test-time perturbations are still encountered, even though a well-calibrated and fine-tuned model has been prepared. For instance," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 763 + ], + "type": "text", + "content": "20075" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 294, + 133 + ], + "type": "text", + "content": "mild looseness or shrinkage perturbations for bounding box prompts are sufficient to cause performance degradation for MedSAM (see Tables 2 and 3), which is clinically unfavorable as delineating a perfect prompt is usually labor-intensive and time-consuming for physicians." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 133, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 133, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 67, + 133, + 295, + 456 + ], + "type": "text", + "content": "To address the above-mentioned issues, test-time adaptation (TTA) for foundation medical segmentation models is a feasible solution that enables performance gains on a specific unlabeled target (test) domain using the pre-trained model only. Although recent TTA approaches for medical image segmentation achieve promising results, we argue that scaling them to foundation segmentation models is difficult. First, most methods utilize various self-supervised objectives (e.g., prediction consistency over different augmentations [2, 6, 30]) to conduct batch normalization (BN)-centric fin-tuning, which may endure limited update signals with marginal adaptation as modern foundation models employ layer normalization (LN) without source-trained statistics like BN. Second, some approaches [27] update the whole model to explore more gains, but even minimal parameter adjustments in large models can precipitate catastrophic forgetting [10]. Third, most methods ignore computational complexity during the adaptation, which is particularly significant for foundation medical segmentation models1, since continuous forward and backward propagations on a large model are computationally intensive, especially on resource-constrained edge devices. Overall, existing methods struggle with maintaining a trade-off between sufficient update signals and knowledge preservation, and ignore computational consumption of the adaptation for large models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 457, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 457, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 67, + 457, + 295, + 672 + ], + "type": "text", + "content": "To this end, we propose a novel TTA framework for foundation medical segmentation models like MedSAM. Compared to existing methods, one of the most pivotal differences is that our proposed method updates latent representations rather than any model parameters², which can not only endow maximal update flexibility under a forgetting-free context but also circumvent heavy forward and backward computations on sophisticated architectures (e.g., image encoder). Specifically, by incorporating an off-the-shelf conditionally independent assumption between the input image X and the prediction Y in MedSAM, we provide a theoretical analysis to uncover that directly refining the image embedding Z (a.k.a. latent refinement) can encourage maximizing factorized conditional probabilities (including latent fidelity " + }, + { + "bbox": [ + 67, + 457, + 295, + 672 + ], + "type": "inline_equation", + "content": "P(\\mathbf{Z}|\\mathbf{X})" + }, + { + "bbox": [ + 67, + 457, + 295, + 672 + ], + "type": "text", + "content": " and segmentation likelihood " + }, + { + "bbox": [ + 67, + 457, + 295, + 672 + ], + "type": "inline_equation", + "content": "P(\\mathbf{Y}|\\mathbf{Z}))" + }, + { + "bbox": [ + 67, + 457, + 295, + 672 + ], + "type": "text", + "content": " of posterior prediction probability " + }, + { + "bbox": [ + 67, + 457, + 295, + 672 + ], + "type": "inline_equation", + "content": "P(\\mathbf{Y}|\\mathbf{X})" + }, + { + "bbox": [ + 67, + 457, + 295, + 672 + ], + "type": "text", + "content": ", in a similar objective like parametric updates. Under this latent" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 541, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 541, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 541, + 191 + ], + "type": "text", + "content": "refinement framework, we devise a novel distributionapproximated latent conditional random field (CRF) loss to leverage useful cues in the input space for explicitly maximizing the posterior distribution of the latent fidelity. Meanwhile, we utilize entropy minimization to enforce the usability of the final predictions for implicitly maximizing segmentation likelihood. Finally, the mask decoder can directly impose the refined latent without inferring the image embedding again. Our contributions include three-fold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 193, + 541, + 419 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 193, + 541, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 193, + 541, + 264 + ], + "spans": [ + { + "bbox": [ + 314, + 193, + 541, + 264 + ], + "type": "text", + "content": "- We provide theoretical analyses to uncover that direct latent refinement is feasible to approach the same goal as parametric updates under the MedSAM architecture, which enables us to realize high computational efficiency and segmentation performance (refers to Table 1) without the risk of catastrophic forgetting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 264, + 539, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 264, + 539, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 264, + 539, + 323 + ], + "type": "text", + "content": "- A distribution-approximated latent CRF loss based on a Maximum Mean Discrepancy-aware co-occurrent measure is proposed to introduce potential useful cues from input space, leading to sufficient statistics of image embedding over the input image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 324, + 539, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 324, + 539, + 419 + ], + "spans": [ + { + "bbox": [ + 314, + 324, + 539, + 419 + ], + "type": "text", + "content": "- The latent refinement framework is validated on three multi-center medical image segmentation datasets with strong appearance variations and simulated prompt shifts. Extensive experiments demonstrate that our method can not only achieve average " + }, + { + "bbox": [ + 314, + 324, + 539, + 419 + ], + "type": "inline_equation", + "content": "\\sim 3\\%" + }, + { + "bbox": [ + 314, + 324, + 539, + 419 + ], + "type": "text", + "content": "\n- Dice score improvements among three datasets but also enjoy around " + }, + { + "bbox": [ + 314, + 324, + 539, + 419 + ], + "type": "inline_equation", + "content": "7\\times" + }, + { + "bbox": [ + 314, + 324, + 539, + 419 + ], + "type": "text", + "content": " computational complexity reduction compared with existing methods." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 433, + 403, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 433, + 403, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 433, + 403, + 445 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 453, + 520, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 520, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 520, + 465 + ], + "type": "text", + "content": "2.1. Medical Image Segmentation with TTA" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 472, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 541, + 723 + ], + "type": "text", + "content": "Various test-time adaptation (TTA) methodologies have been developed to address distribution shifts during inference [3, 17, 25]. Recent advancements have extended TTA frameworks to more complex tasks, including super-resolution [4] and blind image quality assessment [19]. In the field of medical image segmentation, TTA tackles the common multi-center issue [13], where different hospitals may have specific imaging parameters with strong appearance variations [26]. To this end, some methods conducted the optimization of the BN layer, manipulating either the affine-transformation parameters [30] or source-trained BN statistics [6], where many self-supervised objectives and sophisticated alignment of source-trained statistics in the BN layer are leveraged. Meanwhile, some approaches [27] fine-tune the overall model to explore more gains using elaborated objectives and knowledge accumulation strategies. We argue that such parametric updates suffer from restricted effectiveness due to limited BN-centric update signals and catastrophic forgetting for a large pretrained model, respectively. The computational com" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 69, + 682, + 295, + 722 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 682, + 295, + 722 + ], + "spans": [ + { + "bbox": [ + 69, + 682, + 295, + 722 + ], + "type": "text", + "content": "1MedSAM obtains image embedding using an intricate ViT-based encoder (accounts for " + }, + { + "bbox": [ + 69, + 682, + 295, + 722 + ], + "type": "inline_equation", + "content": "98.5\\%" + }, + { + "bbox": [ + 69, + 682, + 295, + 722 + ], + "type": "text", + "content": " computations) and a lightweight decoder. 2We refer to partial or whole model updates as parametric updates hereafter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 753, + 318, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 318, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 318, + 762 + ], + "type": "text", + "content": "20076" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 294, + 203 + ], + "type": "text", + "content": "plexity is also ignored during adaptation. There are limited works specific to the TTA of foundation medical segmentation models. For example, Huang et al. [9] proposed an auxiliary online learning (AuxOL) and adaptive fusion for the foundation medical image segmentation model. However, AuxOL requires clinicians' annotations for model training, leading to potential unfeasibility in many scenarios. Beyond the medical image, Schön et al. [20] leveraged the clicked point prompts as cues for adaptation, which is inapplicable for MedSAM due to its bounding box-based prompts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 204, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 204, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 68, + 204, + 294, + 240 + ], + "type": "text", + "content": "Thus, there is a lack of a customized approach for the TTA problem of foundation medical segmentation models when considering both effectiveness and efficiency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 248, + 219, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 248, + 219, + 260 + ], + "spans": [ + { + "bbox": [ + 69, + 248, + 219, + 260 + ], + "type": "text", + "content": "2.2. Latent Refinement for TTA" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 266, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 266, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 67, + 266, + 296, + 470 + ], + "type": "text", + "content": "Typically, latent refinement is applied in the neural image compression community [1, 5, 14, 21, 23] to reduce the bit consumption of latent representation at compression time. Since the image compression task has a strong self-supervised signal that corresponds to reconstructing an original raw image by an encoder-decoder architecture [1, 5, 14, 21], it is more straightforward to leverage the latent refinement technique. Instead, there is no direct-supervised signal for medical image segmentation during inference, leaving unexplored space. To the best of our knowledge, we are the first to introduce the concept of latent refinement to test-time adaptation of medical image segmentation. More importantly, we provide theoretical analyses to uncover that directly refining the latent representation of the input image is feasible to approach the same goal as parametric update-based counterparts, which is not developed by previous literature." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 481, + 151, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 481, + 151, + 495 + ], + "spans": [ + { + "bbox": [ + 69, + 481, + 151, + 495 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 501, + 266, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 501, + 266, + 514 + ], + "spans": [ + { + "bbox": [ + 69, + 501, + 266, + 514 + ], + "type": "text", + "content": "3.1. Preliminary and Theoretical Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 519, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 519, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 67, + 519, + 295, + 578 + ], + "type": "text", + "content": "This paper mainly discusses the common MedSAM architecture in [15] with the bounding box prompt. During the test-time adaptation, the bounding box prompt is provided and fixed but is omitted in the following derivations for notation simplification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 579, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 579, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 67, + 579, + 295, + 650 + ], + "type": "text", + "content": "Definition 1. (Architecture of MedSAM) Unlike U-Net and its variants [18], MedSAM has no skip connections between an image encoder " + }, + { + "bbox": [ + 67, + 579, + 295, + 650 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 67, + 579, + 295, + 650 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 67, + 579, + 295, + 650 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 67, + 579, + 295, + 650 + ], + "type": "text", + "content": " and a mask decoder " + }, + { + "bbox": [ + 67, + 579, + 295, + 650 + ], + "type": "inline_equation", + "content": "D_{\\varphi}" + }, + { + "bbox": [ + 67, + 579, + 295, + 650 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 67, + 579, + 295, + 650 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 67, + 579, + 295, + 650 + ], + "type": "text", + "content": ", which enables multiple interactions by reusing the image embedding. In light of this, the following proposition can be derived:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "text", + "content": "Proposition 1. Let " + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathcal{X}" + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathcal{Z}" + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\in \\mathcal{Y}" + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "text", + "content": " be an input image, its latent embedding, and its corresponding prediction. For a MedSAM model, " + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "text", + "content": " is strictly conditionally independent of " + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\perp \\mathbf{X} \\mid \\mathbf{Z}" + }, + { + "bbox": [ + 67, + 651, + 295, + 698 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 699, + 295, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 699, + 295, + 723 + ], + "spans": [ + { + "bbox": [ + 68, + 699, + 295, + 723 + ], + "type": "text", + "content": "Proof: Once the (deepest) latent embedding " + }, + { + "bbox": [ + 68, + 699, + 295, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 68, + 699, + 295, + 723 + ], + "type": "text", + "content": " is calculated by " + }, + { + "bbox": [ + 68, + 699, + 295, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = E_{\\theta}(\\mathbf{X})" + }, + { + "bbox": [ + 68, + 699, + 295, + 723 + ], + "type": "text", + "content": ", the upsampling function " + }, + { + "bbox": [ + 68, + 699, + 295, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} =" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 539, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 539, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 539, + 96 + ], + "type": "inline_equation", + "content": "D_{\\varphi}(\\mathbf{Z})" + }, + { + "bbox": [ + 313, + 72, + 539, + 96 + ], + "type": "text", + "content": " is not directly related to the input image " + }, + { + "bbox": [ + 313, + 72, + 539, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 72, + 539, + 96 + ], + "type": "text", + "content": " or its any latent variants, leading to " + }, + { + "bbox": [ + 313, + 72, + 539, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}\\perp \\mathbf{X}\\mid \\mathbf{Z}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 97, + 539, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 539, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 539, + 144 + ], + "type": "text", + "content": "Theorem 1. (General objective) Under maximum a posteriori (MAP) estimation, the objective of TTA for image segmentation is computing the optimal prediction " + }, + { + "bbox": [ + 313, + 97, + 539, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^*" + }, + { + "bbox": [ + 313, + 97, + 539, + 144 + ], + "type": "text", + "content": " by maximizing the posterior probability conditioned on " + }, + { + "bbox": [ + 313, + 97, + 539, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 97, + 539, + 144 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 163, + 547, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 163, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 314, + 163, + 547, + 182 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} ^ {*} = D _ {\\varphi^ {*}} \\left(E _ {\\theta^ {*}} (\\mathbf {X})\\right), \\varphi^ {*}, \\theta^ {*} = \\arg \\max _ {\\theta , \\varphi} [ P (\\mathbf {Y} | \\mathbf {X}; \\theta , \\varphi) ] ^ {3}.", + "image_path": "62e89afd040a03620b3a17ce742a919f564c0c39eee81d4ca2b543f77109e13e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 189, + 539, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 539, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 539, + 248 + ], + "type": "text", + "content": "Various TTA strategies are regarded as implicit posterior maximization by self-supervised loss-based partial or overall parameter updates. Here, we incorporate the architecture of the MedSAM as a factorized precondition to yield an extension of Theorem 1 as follows." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 249, + 539, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 539, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 539, + 308 + ], + "type": "text", + "content": "Corollary 1. (Factorized objective of MedSAM) The log-posterior " + }, + { + "bbox": [ + 313, + 249, + 539, + 308 + ], + "type": "inline_equation", + "content": "\\log P(\\mathbf{Y}|\\mathbf{X})" + }, + { + "bbox": [ + 313, + 249, + 539, + 308 + ], + "type": "text", + "content": " during test-time adaptation can be factorized into the sum of the segmentation likelihood " + }, + { + "bbox": [ + 313, + 249, + 539, + 308 + ], + "type": "inline_equation", + "content": "\\log P(\\mathbf{Y}|\\mathbf{Z})" + }, + { + "bbox": [ + 313, + 249, + 539, + 308 + ], + "type": "text", + "content": " and the latent fidelity term " + }, + { + "bbox": [ + 313, + 249, + 539, + 308 + ], + "type": "inline_equation", + "content": "\\log P(\\mathbf{Z}|\\mathbf{X})" + }, + { + "bbox": [ + 313, + 249, + 539, + 308 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 313, + 249, + 539, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}\\perp \\mathbf{X}\\mid \\mathbf{Z}" + }, + { + "bbox": [ + 313, + 249, + 539, + 308 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 363, + 316, + 492, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 316, + 492, + 331 + ], + "spans": [ + { + "bbox": [ + 363, + 316, + 492, + 331 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} ^ {*} = D _ {\\varphi^ {*}} \\left(\\mathbf {Z} ^ {*}\\right), \\mathbf {Z} ^ {*} = E _ {\\theta^ {*}} (\\mathbf {X})", + "image_path": "31268354f641f524e8d98b15b9e978cb54987dac8234148ca6eefdae246a1509.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 331, + 534, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 331, + 534, + 350 + ], + "spans": [ + { + "bbox": [ + 321, + 331, + 534, + 350 + ], + "type": "interline_equation", + "content": "\\varphi^ {*}, \\theta^ {*} = \\arg \\max _ {\\varphi , \\theta} [ \\log P (\\mathbf {Y} | \\mathbf {Z}; \\varphi) + \\log P (\\mathbf {Z} | \\mathbf {X}; \\theta) ],", + "image_path": "6ec5658eac46876213190a438289c5ff5abba3a6abe7211ce3068cf157a7a8d4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 357, + 539, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 539, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 539, + 406 + ], + "type": "text", + "content": "Proof: By recalling Prop. 1, " + }, + { + "bbox": [ + 313, + 357, + 539, + 406 + ], + "type": "inline_equation", + "content": "P(\\mathbf{Y}|\\mathbf{X}) \\propto P(\\mathbf{Y}|\\mathbf{Z}) \\cdot P(\\mathbf{Z}|\\mathbf{X})" + }, + { + "bbox": [ + 313, + 357, + 539, + 406 + ], + "type": "text", + "content": " holds. Substituting this into the MAP objective in Theorem 1 yields the factorized objective above. Please refer to the Appendix for more details." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 415, + 539, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 539, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 539, + 426 + ], + "type": "text", + "content": "3.2. Latent Refinement in MedSAM framework" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 432, + 539, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 539, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 539, + 480 + ], + "type": "text", + "content": "Problem. For Corollary 1, one of the commonly adopted solutions for achieving optimal prediction " + }, + { + "bbox": [ + 313, + 432, + 539, + 480 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^*" + }, + { + "bbox": [ + 313, + 432, + 539, + 480 + ], + "type": "text", + "content": " is directly to jointly optimize partial or whole parameters in " + }, + { + "bbox": [ + 313, + 432, + 539, + 480 + ], + "type": "inline_equation", + "content": "\\{\\varphi ,\\theta \\}" + }, + { + "bbox": [ + 313, + 432, + 539, + 480 + ], + "type": "text", + "content": ", but two challenges are encountered as follows." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 481, + 539, + 611 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 314, + 481, + 539, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 481, + 539, + 552 + ], + "spans": [ + { + "bbox": [ + 314, + 481, + 539, + 552 + ], + "type": "text", + "content": "- Partial model updates. Most existing approaches are BN-centric strategies, which are inapplicable for LN-based MedSAM. Although updating the affine transformation parameters in LN layers is feasible, the performance may be marginal due to restricted model modifications especially in strong domain shifts." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 552, + 539, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 539, + 611 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 539, + 611 + ], + "type": "text", + "content": "- Whole model updates. It is difficult to maintain a trade-off between TTA performance and catastrophic forgetting. Meanwhile, optimizing the overall model is computationally expensive due to the considerable number of parameters in modern large models." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 612, + 539, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 539, + 695 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 539, + 695 + ], + "type": "text", + "content": "Solution. In light of this, we are interested in directly optimizing the latent embedding rather than the model parameters. Assume the encoder " + }, + { + "bbox": [ + 313, + 612, + 539, + 695 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 313, + 612, + 539, + 695 + ], + "type": "text", + "content": " and the decoder " + }, + { + "bbox": [ + 313, + 612, + 539, + 695 + ], + "type": "inline_equation", + "content": "D_{\\varphi}" + }, + { + "bbox": [ + 313, + 612, + 539, + 695 + ], + "type": "text", + "content": " are fixed during TTA, this scheme can not only circumvent the above-mentioned challenges (i.e., it is likely to update more \"parameters\" in the forgetting-free context) but also be equivalent to theoretically achieving the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 702, + 539, + 721 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 702, + 539, + 721 + ], + "spans": [ + { + "bbox": [ + 313, + 702, + 539, + 721 + ], + "type": "text", + "content": "3In practice, one would optimize the expectation of the probability. To be simplified, we omit it in later formulations." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "20077" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 222 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 222 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 222 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 222 + ], + "type": "image", + "image_path": "1f61326352df930d53a2bb4300a8ce0ed762f2b9e4a917310b3e0e59c6740a5b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 228, + 541, + 251 + ], + "lines": [ + { + "bbox": [ + 68, + 228, + 541, + 251 + ], + "spans": [ + { + "bbox": [ + 68, + 228, + 541, + 251 + ], + "type": "text", + "content": "Figure 1. The framework of the proposed method, including initial embedding inference, TTA via latent refinement, and final inference phases. Distribution-approximated latent CRF loss is visualized, where the loss between " + }, + { + "bbox": [ + 68, + 228, + 541, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 68, + 228, + 541, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 228, + 541, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j" + }, + { + "bbox": [ + 68, + 228, + 541, + 251 + ], + "type": "text", + "content": " is computed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 69, + 271, + 295, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 271, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 69, + 271, + 295, + 285 + ], + "type": "text", + "content": "same factorized objective of MedSAM as Corollary 1," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 300, + 215, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 300, + 215, + 314 + ], + "spans": [ + { + "bbox": [ + 148, + 300, + 215, + 314 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} ^ {*} = D _ {\\varphi} (\\mathbf {Z} ^ {*}),", + "image_path": "1c60a57377772b70c14157f9730e60043059422fdccceeb7ba81c24d8cc79fb7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 76, + 316, + 294, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 316, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 76, + 316, + 294, + 334 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} ^ {*} = \\arg \\max _ {\\mathbf {Z}} \\left[ \\log P (\\mathbf {Y} | \\mathbf {Z}; \\varphi) + \\log P (\\mathbf {Z} | \\mathbf {X}; \\theta) \\right], \\tag {1}", + "image_path": "8fcf0cebe4fc7b4b7a15c815a0fb227af3318cb34fdb19ef1a78de7d865e9f83.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = E_{\\theta}(\\mathbf{X})" + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "text", + "content": " can be initialized and refined iteratively. Typically, the total parameter number of " + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "text", + "content": " is overly larger than that of the normalization layer, resulting in more update flexibility. Under frozen " + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "text", + "content": ", there is no knowledge forgetting during optimization. These two perspectives are more likely to activate better TTA performance collaboratively than existing methods. Effectiveness of Latent Refinement during TTA. First, the " + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\log P(\\mathbf{Z}|\\mathbf{X})" + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "text", + "content": " term encourages " + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "text", + "content": " to maintain sufficient statistics over the input " + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "text", + "content": ", which is highly significant for the MedSAM framework during the TTA because 1) domain-specific variations at test time may degrade salient characteristics (e.g., segmentation cues like obvious edges or contours) of latent embedding, leading to suboptimal segmentation performance; 2) sufficient statistics relative to the input " + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "text", + "content": " can encourage the conditional independence assumption of MedSAM, i.e., " + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\mathbf{Y} \\perp \\mathbf{X} \\mid \\mathbf{Z}" + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "text", + "content": ", to hold. Second, " + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\log P(\\mathbf{Y}|\\mathbf{Z})" + }, + { + "bbox": [ + 68, + 338, + 296, + 577 + ], + "type": "text", + "content": " guarantees the usability of latent embedding for final predictions, preventing any collapsed solutions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 578, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 578, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 69, + 578, + 295, + 613 + ], + "type": "text", + "content": "Practical implementation. For Eq. (1), one would update the initial latent embedding " + }, + { + "bbox": [ + 69, + 578, + 295, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = \\mathbf{Z}^0 = E_\\theta (\\mathbf{X})" + }, + { + "bbox": [ + 69, + 578, + 295, + 613 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 69, + 578, + 295, + 613 + ], + "type": "inline_equation", + "content": "m\\geq 1" + }, + { + "bbox": [ + 69, + 578, + 295, + 613 + ], + "type": "text", + "content": " steps, for each step " + }, + { + "bbox": [ + 69, + 578, + 295, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^{m} =" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 616, + 312, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 616, + 312, + 651 + ], + "spans": [ + { + "bbox": [ + 69, + 616, + 312, + 651 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} ^ {m - 1} - \\epsilon \\cdot \\frac {\\partial [ - \\log P (\\mathbf {Y} ^ {m - 1} | \\mathbf {Z} ^ {m - 1} ; \\varphi) - \\log P (\\mathbf {Z} ^ {m - 1} | \\mathbf {X} ; \\theta ]}{\\partial \\mathbf {Z} ^ {m - 1}} \\tag {2}", + "image_path": "7caeed8fecd0a563f8b801c1ac3d570183aa1065801bfeee46bb3775ae5a5b23.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 651, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 651, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 68, + 651, + 295, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 651, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 68, + 651, + 295, + 676 + ], + "type": "text", + "content": " is the learning rate. The final prediction can be represented as " + }, + { + "bbox": [ + 68, + 651, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^m = D_{\\varphi}(\\mathbf{Z}^m)" + }, + { + "bbox": [ + 68, + 651, + 295, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 681, + 280, + 694 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 681, + 280, + 694 + ], + "spans": [ + { + "bbox": [ + 69, + 681, + 280, + 694 + ], + "type": "text", + "content": "3.3. Distribution-Approximated Latent CRF" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 698, + 295, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 698, + 295, + 723 + ], + "spans": [ + { + "bbox": [ + 68, + 698, + 295, + 723 + ], + "type": "text", + "content": "To Eq. (2), one of the most important components is maximizing the log-probability " + }, + { + "bbox": [ + 68, + 698, + 295, + 723 + ], + "type": "inline_equation", + "content": "P(\\mathbf{Z}^{m - 1}|\\mathbf{X};\\theta)" + }, + { + "bbox": [ + 68, + 698, + 295, + 723 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 271, + 541, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 541, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 541, + 403 + ], + "type": "text", + "content": "the latent fidelity term. For MedSAM, the volume of the image encoder parameter " + }, + { + "bbox": [ + 313, + 271, + 541, + 403 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 271, + 541, + 403 + ], + "type": "text", + "content": " is considerable, resulting in a high computation during the forward and backward propagations in the TTA phase. Benefiting from our latent refinement framework, we propose circumventing the image encoder " + }, + { + "bbox": [ + 313, + 271, + 541, + 403 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 271, + 541, + 403 + ], + "type": "text", + "content": " to directly enforce sufficient statistics over the input image " + }, + { + "bbox": [ + 313, + 271, + 541, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 271, + 541, + 403 + ], + "type": "text", + "content": ", i.e., maximizing " + }, + { + "bbox": [ + 313, + 271, + 541, + 403 + ], + "type": "inline_equation", + "content": "\\log P(\\mathbf{Z}^{m-1}|\\mathbf{X})" + }, + { + "bbox": [ + 313, + 271, + 541, + 403 + ], + "type": "text", + "content": ". By doing so, we can leverage potential useful cues in the input space, which is typically ignored by existing TTA methods that focus more on the label space (e.g., prediction consistency [2])." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": "Therefore, we introduce the concept of conditional random field (CRF) [22], which is usually employed to refine the final predictions by enforcing the co-occurrence between the label space " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": " and the input space " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": ". For our purpose, we extend the CRF to the co-occurrence between the latent space " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": " and the input space " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": ". Theoretically, minimizing the CRF energy is equivalent to maximizing the posterior probability of the random field [11], which coincides with our optimization objective in Eq. (2) via a computationally efficient manner. Formally, given a random field " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = E_{\\theta}(\\mathbf{X})" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": " over a set of vectors " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{z}_1,\\mathbf{z}_2,\\dots ,\\mathbf{z}_{N'}\\}" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}\\in \\mathbb{R}^{W^{\\prime}\\times H^{\\prime}\\times C^{\\prime}}" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": " is the latent embedding with a spatial resolution of " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "N^{\\prime} = W^{\\prime}\\times H^{\\prime}" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": " and a channel of " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "C^\\prime" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{z}\\in \\mathbb{R}^{1\\times C^{\\prime}}" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": ". Consider another random field " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": " over a set of vectors " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_1,\\mathbf{x}_2,\\dots ,\\mathbf{x}_N\\}" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\in \\mathbb{R}^{W\\times H\\times C}" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": " is the input intensities with a spatial resolution of " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "N = W\\times H" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": " and a channel of " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^{1\\times C}" + }, + { + "bbox": [ + 313, + 407, + 541, + 622 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "content": "Typically, " + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "content": " are related in a one-to-one manner, i.e., " + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "content": " represents a color vector at pixel " + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "content": " in the input space and " + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "content": " is its corresponding predicted label in the label space. However, the input and latent spaces are dimensionally heterogeneous, i.e., " + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "inline_equation", + "content": "N \\neq N'" + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "content": ", which makes constructing a random field in this scenario challenging. Here, a novel distribution-approximated latent CRF (DAL-CRF) is proposed to eliminate the dimen" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "20078" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 68, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 294, + 144 + ], + "type": "text", + "content": "sionally heterogeneous problem between the input and latent spaces. Specifically, by leveraging the fact that each latent vector " + }, + { + "bbox": [ + 68, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 68, + 72, + 294, + 144 + ], + "type": "text", + "content": " corresponds to a receptive field in the input space with a set of intensity vectors, it is feasible to identify related intensity vectors for each latent vector, i.e.," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 151, + 295, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 151, + 295, + 166 + ], + "spans": [ + { + "bbox": [ + 102, + 151, + 295, + 166 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {i} \\rightarrow \\mathbf {X} _ {i} = \\left\\{\\mathbf {x} _ {i} ^ {l} \\right\\} _ {l = 1} ^ {L}, i = 1, 2, \\dots , N ^ {\\prime}, \\tag {3}", + "image_path": "445b0ebcdd5c2a7e974087c9f47de07830bc6d08b4dcfe496d60e6ee64b523c1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 68, + 173, + 296, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 173, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 68, + 173, + 296, + 245 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 68, + 173, + 296, + 245 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 68, + 173, + 296, + 245 + ], + "type": "text", + "content": " denotes the number of intensity vectors in the receptive field, where " + }, + { + "bbox": [ + 68, + 173, + 296, + 245 + ], + "type": "inline_equation", + "content": "L = 256" + }, + { + "bbox": [ + 68, + 173, + 296, + 245 + ], + "type": "text", + "content": " corresponding to a " + }, + { + "bbox": [ + 68, + 173, + 296, + 245 + ], + "type": "inline_equation", + "content": "16\\times 16" + }, + { + "bbox": [ + 68, + 173, + 296, + 245 + ], + "type": "text", + "content": " receptive field in MedSAM. Then, by following Veksler [24], the pairwise potential is only considered to be a regularization loss to encourage the minimization of CRF energy, which can be formulated as " + }, + { + "bbox": [ + 68, + 173, + 296, + 245 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DAL - CRF}} =" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 251, + 295, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 251, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 69, + 251, + 295, + 290 + ], + "type": "interline_equation", + "content": "\\frac {1}{N ^ {\\prime}} \\sum_ {i, j \\in \\mathcal {G}} \\left(\\frac {1}{2} \\operatorname {M M D} \\left(\\mathbb {P} \\left(\\mathbf {X} _ {i}\\right), \\mathbb {P} \\left(\\mathbf {X} _ {j}\\right)\\right)\\right) \\cdot \\exp \\left(- \\| \\mathbf {z} _ {i} - \\mathbf {z} _ {j} \\| _ {F}\\right). \\tag {4}", + "image_path": "f3831c36673ebe12d15af30a70237b13ae458c5f356670d6a169397f8ff8bc55.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "type": "text", + "content": "Unlike existing CRF methods [24], which utilize the Gaussian kernel to evaluate the correlation between two color vectors, we introduce the Maximum Mean Discrepancy (MMD) to measure the distance between two vector sets. Specifically, each vector set is first approximated as an empirical distribution, i.e., " + }, + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i^l\\sim \\mathbb{P}(\\mathbf{X}_i)" + }, + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "type": "text", + "content": ". Then, each empirical distribution is represented by a kernel mean embedding in a reproducing kernel Hilbert space (RKHS) " + }, + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathbb{P}} = \\mathbb{E}_{\\mathbf{x}\\sim \\mathbb{P}}[\\phi (\\mathbf{x})] = \\int_{\\mathcal{X}}k(\\mathbf{x},\\cdot)d\\mathbb{P} = \\mathbb{E}_{\\mathbf{x}\\sim \\mathbb{P}}[k(\\mathbf{x},\\cdot)]" + }, + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "type": "inline_equation", + "content": "k:\\mathcal{X}\\times \\mathcal{X}\\to \\mathbb{R}" + }, + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "type": "text", + "content": " is a reproducing kernel (Gaussian RBF kernel here) and corresponding feature map " + }, + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\phi :\\mathcal{X}\\rightarrow \\mathcal{H}" + }, + { + "bbox": [ + 68, + 291, + 296, + 448 + ], + "type": "text", + "content": ". The discrepancy can be measured between mean embeddings, i.e.," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 464, + 304, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 464, + 304, + 506 + ], + "spans": [ + { + "bbox": [ + 69, + 464, + 304, + 506 + ], + "type": "interline_equation", + "content": "\\operatorname {M M D} \\left(\\mathbb {P} \\left(\\mathbf {X} _ {i}\\right), \\mathbb {P} \\left(\\mathbf {X} _ {j}\\right)\\right) = \\| \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\phi \\left(\\mathbf {x} _ {i} ^ {l}\\right) - \\frac {1}{L} \\sum_ {l = 1} ^ {L} \\phi \\left(\\mathbf {x} _ {j} ^ {l}\\right) \\| _ {\\mathcal {H}} ^ {2}. \\tag {5}", + "image_path": "d90c85aee52d37636d23a29c7558b3cfb12888db24ea919358e5b4c50e9d307f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 507, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 507, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 68, + 507, + 295, + 602 + ], + "type": "text", + "content": "MMD enables all-order correlation measurement using Gaussian kernel via unbiased empirical estimation. This may be useful for better discrepancy measurement, as the distribution of the intensity space may be highly complex and non-linear. For the second term in Eq. (4), we utilize the Frobenius norm " + }, + { + "bbox": [ + 68, + 507, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _F" + }, + { + "bbox": [ + 68, + 507, + 295, + 602 + ], + "type": "text", + "content": " to calculate the feature compatibility in a co-occurrent probability using exponential zero-to-one normalization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 603, + 296, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 603, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 68, + 603, + 296, + 723 + ], + "type": "text", + "content": "Integrating Bounding Box Prompt. The graph " + }, + { + "bbox": [ + 68, + 603, + 296, + 723 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}" + }, + { + "bbox": [ + 68, + 603, + 296, + 723 + ], + "type": "text", + "content": " in Eq. (4) is usually either a complete (dense-connected) graph [11] or a sparse position-related neighboring graph [24] on " + }, + { + "bbox": [ + 68, + 603, + 296, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 68, + 603, + 296, + 723 + ], + "type": "text", + "content": ". Instead, we propose to consider the bounding box prompt as the prior to construct " + }, + { + "bbox": [ + 68, + 603, + 296, + 723 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 68, + 603, + 296, + 723 + ], + "type": "text", + "content": "-random pairs inside the given bounding box, which can not only exclude those unimportant latent features far away from the bounding box prompt to enhance the computational efficiency but also focus more on the interested region with better performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 541, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 541, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 541, + 193 + ], + "type": "text", + "content": "Intuitive Understanding. For Eq. (4), one actually encourages being more discriminative in latent representations between " + }, + { + "bbox": [ + 313, + 72, + 541, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_i" + }, + { + "bbox": [ + 313, + 72, + 541, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 541, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_j" + }, + { + "bbox": [ + 313, + 72, + 541, + 193 + ], + "type": "text", + "content": ". Since if two empirical distributions have a large distribution discrepancy (i.e., the MMD distance is large) in the input space, there are more penalties assigned to the pair who has larger similarity (i.e., co-occurrent probability is larger)) in the latent space. By minimizing " + }, + { + "bbox": [ + 313, + 72, + 541, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DAL - CRF}}" + }, + { + "bbox": [ + 313, + 72, + 541, + 193 + ], + "type": "text", + "content": ", one can enforce the latent representations to equip with sufficient statistics as in the input space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 199, + 533, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 533, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 533, + 212 + ], + "type": "text", + "content": "3.4. Overall Objective at Test-time Adaptation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 217, + 541, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 541, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 541, + 277 + ], + "type": "text", + "content": "Without accessing the ground-truth label, it is infeasible to maximize the segmentation likelihood " + }, + { + "bbox": [ + 313, + 217, + 541, + 277 + ], + "type": "inline_equation", + "content": "\\log P(\\mathbf{Y}^m|\\mathbf{Z}^m;\\varphi)" + }, + { + "bbox": [ + 313, + 217, + 541, + 277 + ], + "type": "text", + "content": " directly. Therefore, we adopt a common predictive entropy minimization to implicitly achieve this goal [8]. Formally," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 285, + 552, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 552, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 552, + 310 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E M}} = H (P (\\mathbf {Y} | \\mathbf {Z})) = - \\sum_ {h, v \\in C} P (y ^ {h, v} | \\mathbf {Z}) \\log P (y ^ {h, v} | \\mathbf {Z}),", + "image_path": "e53349144539c695bea743801228e9d4e75f3d766327e5df576c0cff8483554d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 313, + 541, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 313, + 541, + 328 + ], + "spans": [ + { + "bbox": [ + 347, + 313, + 541, + 328 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\quad C = \\left\\{\\left(h, v\\right) \\mid P \\left(y ^ {h, v} \\mid \\mathbf {Z}\\right) \\geq \\alpha \\right\\}, \\tag {6}", + "image_path": "b44b8701368f80208e7fdd00e40c4c3b7de5bd8201bcfdf937c01e96bcad9882.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 337, + 541, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 541, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 541, + 516 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 337, + 541, + 516 + ], + "type": "inline_equation", + "content": "y^{h,v}" + }, + { + "bbox": [ + 313, + 337, + 541, + 516 + ], + "type": "text", + "content": " denotes a single prediction in the position " + }, + { + "bbox": [ + 313, + 337, + 541, + 516 + ], + "type": "inline_equation", + "content": "(h,v)" + }, + { + "bbox": [ + 313, + 337, + 541, + 516 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 337, + 541, + 516 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 313, + 337, + 541, + 516 + ], + "type": "text", + "content": ". Eq. (6) is similar to foreground-background balanced entropy loss in Dong et al. [6]. However, we only utilize the foreground entropy and set a high confidence threshold (e.g., " + }, + { + "bbox": [ + 313, + 337, + 541, + 516 + ], + "type": "inline_equation", + "content": "\\alpha = 0.95" + }, + { + "bbox": [ + 313, + 337, + 541, + 516 + ], + "type": "text", + "content": ") due to the following reasons. First, since the CRF-based loss usually requires a volumetric regularization to avoid the trivial solutions [24] (e.g., all labels are assigned to background), the entropy regularization over a high confidence foreground region can be interpreted as an implicit volumetric regularization, i.e., ensuring these foreground pixels confident enough for preventing trivial solutions. Second, we argue that background entropy is unnecessary for a foundation segmentation model, as these models have learned a good foreground-background balance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 517, + 541, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 541, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 541, + 564 + ], + "type": "text", + "content": "By considering the proposed distributionapproximated latent CRF loss and the customized entropy loss, the practical latent refinement scheme is reformulated as follows." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 339, + 571, + 540, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 571, + 540, + 595 + ], + "spans": [ + { + "bbox": [ + 339, + 571, + 540, + 595 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} ^ {m} = \\mathbf {Z} ^ {m - 1} - \\epsilon \\cdot \\frac {\\partial \\left[ \\mathcal {L} _ {\\mathrm {D A L - C R F}} + \\lambda \\mathcal {L} _ {\\mathrm {E M}} \\right]}{\\partial \\mathbf {Z} ^ {m - 1}}, \\tag {7}", + "image_path": "3c6069ce7b1143b49376055734ccd831353a16b1aa2daecfc4b110f81cd8f28a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 602, + 446, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 446, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 446, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 602, + 446, + 613 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 602, + 446, + 613 + ], + "type": "text", + "content": " is a balance coefficient." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 621, + 446, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 446, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 446, + 633 + ], + "type": "text", + "content": "3.5. Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 639, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 541, + 723 + ], + "type": "text", + "content": "The overall framework is illustrated in Figure 1. Superficially, the initial image embedding " + }, + { + "bbox": [ + 313, + 639, + 541, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^0" + }, + { + "bbox": [ + 313, + 639, + 541, + 723 + ], + "type": "text", + "content": " is inferred only once. Once the refined latent representation " + }, + { + "bbox": [ + 313, + 639, + 541, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^*" + }, + { + "bbox": [ + 313, + 639, + 541, + 723 + ], + "type": "text", + "content": " is obtained, the upsampling can be conducted directly without requiring image embedding again. This mechanism is well compatible with the interactive scheme of the MedSAM. We implement the proposed method based on" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "20079" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 76, + 71, + 537, + 332 + ], + "blocks": [ + { + "bbox": [ + 76, + 71, + 537, + 332 + ], + "lines": [ + { + "bbox": [ + 76, + 71, + 537, + 332 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 537, + 332 + ], + "type": "table", + "html": "
Update TypeMethodDice Coefficient ↑Average Surface Distance ↓
D1D2D3D4Avg.D1D2D3D4Avg.
Looseness promptsDirect Inference90.9987.1687.0684.7287.489.487.518.747.628.33
Normalization updateTENT (ICLR'21)91.2785.4085.0681.8485.899.098.7010.058.889.18
InTEnt (CVPR'24)91.3687.4387.4786.1387.608.807.518.267.668.06
GraTa (AAAI'25)90.8987.0086.9584.6487.379.487.548.807.658.37
PASS (TMI'24)80.8679.6382.0384.0281.6320.4012.2211.697.9713.07
Full updatesMEMO (NeurIPS'22)87.2685.4886.4284.4785.9112.458.279.157.769.40
CRF-SOD (CVPR'23)89.5784.0184.8082.5885.2410.548.9910.188.649.58
Latent updatesOurs89.3588.7389.8388.1889.0211.406.696.926.077.77
Shrinkage promptsDirect Inference81.0567.6072.1068.5772.3318.5117.6117.7614.9217.20
Normalization updateTENT (ICLR'21)80.2265.1769.2562.1069.1818.8519.2619.3917.5818.77
InTEnt (CVPR'24)82.7270.0674.5073.4975.1916.8615.2816.1313.8715.53
GraTa (AAAI'25)80.8767.3271.9468.4672.1517.2716.5416.5314.5216.21
PASS (TMI'24)66.8158.6663.3156.7861.3932.3922.8922.7420.1924.55
Full updatesMEMO (NeurIPS'22)81.0267.5472.1068.8472.3718.5217.7617.7714.7817.20
CRF-SOD (CVPR'23)81.8469.2674.7372.0874.4819.2718.8318.6314.7317.86
Latent updatesOurs83.0773.7278.5475.7777.7716.7414.3813.7211.2614.02
Perfectness promptsDirect Inference87.5880.2581.3278.4681.9012.7611.2112.2710.2411.62
Normalization updateTENT (ICLR'21)87.6077.9878.1872.9979.1812.8312.2113.9212.6912.91
InTEnt (CVPR'24)86.5484.2585.1781.6584.4010.989.239.858.959.75
GraTa (AAAI'25)87.3979.9781.1478.3681.7112.9611.3812.3810.2911.75
PASS (TMI'24)74.5969.8673.8871.3472.4125.5317.3216.5813.8518.32
Full updatesMEMO (NeurIPS'22)87.3980.1881.3278.4981.8412.9711.2412.2710.2311.67
CRF-SOD (CVPR'23)86.0777.2279.1076.2479.6513.9612.5913.6311.212.84
Latent updatesOurs86.8585.1186.7284.4985.7913.738.638.797.489.65
", + "image_path": "46724d48282b0d2b92c3026b1a7e1cf4060d70b11092ae966452024d939f9fd5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 341, + 541, + 375 + ], + "lines": [ + { + "bbox": [ + 68, + 341, + 541, + 375 + ], + "spans": [ + { + "bbox": [ + 68, + 341, + 541, + 375 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of TTA results on fundus image OC segmentation between different methods. The best and second-best results are shown in the bold and the underline, respectively. For loose prompts, we report the average of two looseness ratios. For shrinkage prompts, we report the average of two shrinkage ratios. Full results can be found in the Appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": "official MedSAM checkpoints (ViT-B pre-trained models) and codes using a one-step optimization strategy " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "(m = 1)" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " by following previous TTA settings. Due to the extremely various sizes of RoIs in medical images, we leverage the initial segmentation result " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}^0" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " and the given bounding box prompt to explore a dynamic adjustment strategy for important hyperparameters. Specifically, the learning rate " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " with the Adam optimizer is linearly interpolated from " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "5\\times 10^{-2}" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "1\\times 10^{-1}" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " according to the ratio of the initial foreground to the entire image " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " (i.e., the rough size of the RoI), where " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "f_{e}\\leq 0.05" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "f_{e}\\geq 0.2" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " are fixedly set to " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "5\\times 10^{-2}" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "1\\times 10^{-1}" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": ", respectively. Note that we use commonly-used definitions over the " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " as adopted by Gao et al. [7], i.e., " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "f_{e}\\leq 0.05" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "0.05 < f_{e} < 0.2" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "f_{e}\\geq 0.2" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " as small, medium, and large RoIs. Similarly, the " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " is also RoI-specific adjustment, where the small RoI " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "(f_{e}\\leq 0.05)" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " requires a similar contribution between the " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DAL - CRF}}" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{EM}}" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": ", thus the " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "1\\times 10^{3}" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " is set to 4 by balancing computational consumption and performance (refers to Table 5(a)). " + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 68, + 384, + 296, + 636 + ], + "type": "text", + "content": " in entropy loss is set to 0.95 empirically." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 652, + 149, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 652, + 149, + 666 + ], + "spans": [ + { + "bbox": [ + 69, + 652, + 149, + 666 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 674, + 296, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 674, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 68, + 674, + 296, + 723 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics We evaluate our method on two public multi-center medical image segmentation datasets with different data modalities. (1) Fundus [26]: Fundus datasets are retinal fundus images" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 384, + 541, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 541, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 541, + 504 + ], + "type": "text", + "content": "collected from 4 healthcare centers (D1 to D4 for short). For each retinal fundus image, there are two segmentation objectives, including the optic cup (OC) (mainly small RoIs) and optic disc (OD) (mainly medium or large RoIs). (2) Spinal Cord Gray Matter (SCGM) [12]: SCGM datasets are collected from 4 institutions (D1 to D4 for short) with different imaging manufacturers for spinal MRI image gray matter segmentation. We use the Dice Similarity Coefficient (DSC) and Average Surface Distance (ASD) to measure the accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 507, + 541, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 541, + 722 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 541, + 722 + ], + "type": "text", + "content": "Experiment Settings. Appearance Perturbations: Note that our adopted multi-center datasets have strong domain shifts with appearance variations, we thus ignore test-time appearance perturbations (e.g., changes in contrast, noise, and blur). Prompt Perturbations: We mainly consider three kinds of bounding box prompts to evaluate TTA performance. (i) Perfectness prompt " + }, + { + "bbox": [ + 313, + 507, + 541, + 722 + ], + "type": "inline_equation", + "content": "(PP)" + }, + { + "bbox": [ + 313, + 507, + 541, + 722 + ], + "type": "text", + "content": ": The perfect prompt can be obtained by groundtruth (GT) mask, where the top-left and lower-right coordinates of the PP correspond to the minimal and maximal horizontal and vertical coordinates of the GT mask, respectively. A unique PP is utilized for TTA. (ii) Looseness prompts " + }, + { + "bbox": [ + 313, + 507, + 541, + 722 + ], + "type": "inline_equation", + "content": "(LP)" + }, + { + "bbox": [ + 313, + 507, + 541, + 722 + ], + "type": "text", + "content": ": Clinically, it is labor-intensive and time-consuming to obtain the PP. Instead, a moderately loose prompt can be efficient but may trigger the degradation of the model with the need for TTA. Therefore, we define the looseness rate " + }, + { + "bbox": [ + 313, + 507, + 541, + 722 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 507, + 541, + 722 + ], + "type": "text", + "content": ", which is the ratio of the coordinate shift relative to the width " + }, + { + "bbox": [ + 313, + 507, + 541, + 722 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 507, + 541, + 722 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "20080" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 76, + 71, + 537, + 331 + ], + "blocks": [ + { + "bbox": [ + 76, + 71, + 537, + 331 + ], + "lines": [ + { + "bbox": [ + 76, + 71, + 537, + 331 + ], + "spans": [ + { + "bbox": [ + 76, + 71, + 537, + 331 + ], + "type": "table", + "html": "
Update TypeMethodDice Coefficient ↑Average Surface Distance ↓
D1D2D3D4Avg.D1D2D3D4Avg.
Looseness promptsDirect Inference83.0578.8681.8275.5479.8133.2032.8730.2931.9032.06
Normalization updateTENT (ICLR'21)81.6176.9680.0074.8578.3537.5836.5433.5334.3535.50
InTEnt (CVPR'24)83.6176.7180.0071.8378.0332.4637.0132.5739.1435.29
GraTa (AAAI'25)84.6480.0484.1082.8282.9030.0230.9325.9128.9128.94
PASS (TMI'24)84.5781.7782.2977.3481.4918.0226.2021.0924.6822.49
Full model updatesMEMO (NeurIPS'22)86.2680.0183.9478.1882.1026.7330.9326.2728.5428.12
CRF-SOD (CVPR'23)87.0382.8683.3379.5283.1825.1926.2523.7226.5325.42
Latent updatesOurs90.8583.5587.7782.2486.1017.5923.5418.5021.7520.34
Shrinkage promptsDirect Inference92.7086.1790.9088.6789.6112.6017.6612.7712.1013.78
Normalization updateTENT (ICLR'21)90.6385.6389.6686.8888.2015.9018.4414.2712.4815.27
InTEnt (CVPR'24)90.9187.2790.8888.7089.4415.4916.4712.6812.0514.17
GraTa (AAAI'25)92.6185.6091.0088.7189.4812.7018.3212.6112.1113.93
PASS (TMI'24)91.4678.8585.9275.7082.9813.7225.1417.5423.0519.86
Full model updatesMEMO (NeurIPS'22)92.1283.8290.8788.7788.8913.5220.1812.7512.0414.62
CRF-SOD (CVPR'23)94.4284.0792.3491.4390.569.3619.8310.709.3012.30
Latent updatesOurs94.2486.7793.2392.4291.669.8016.259.458.2110.93
Perfectness promptsDirect Inference96.4992.8793.8992.3693.906.099.488.648.368.14
Normalization updateTENT (ICLR'21)94.2090.7893.3990.3892.1810.1212.019.3210.3910.46
InTEnt (CVPR'24)94.1291.5794.4293.4793.3910.1211.018.017.039.04
GraTa (AAAI'25)94.4692.6693.9192.3193.336.149.738.618.428.22
PASS (TMI'24)94.9085.3584.3686.1487.688.5817.5217.6213.9814.42
Full model updatesMEMO (NeurIPS'22)95.7791.4893.8992.3193.367.8011.288.618.419.02
CRF-SOD (CVPR'23)96.6691.1593.7693.4193.755.7811.428.737.268.29
Latent updatesOurs96.6892.6795.1594.5794.765.699.516.855.987.00
", + "image_path": "0007e14b6606369f862dca70c689fb83c814d3ca09c4d3e0be7cb7cd54b1e975.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 76, + 376, + 537, + 480 + ], + "blocks": [ + { + "bbox": [ + 68, + 340, + 541, + 373 + ], + "lines": [ + { + "bbox": [ + 68, + 340, + 541, + 373 + ], + "spans": [ + { + "bbox": [ + 68, + 340, + 541, + 373 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of TTA results on fundus image OD segmentation between different methods. The best and second-best results are shown in the bold and the underline, respectively. For loose prompts, we report the average of two looseness ratios. For shrinkage prompts, we report the average of two shrinkage ratios. Full results can be found in the Appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 376, + 537, + 480 + ], + "lines": [ + { + "bbox": [ + 76, + 376, + 537, + 480 + ], + "spans": [ + { + "bbox": [ + 76, + 376, + 537, + 480 + ], + "type": "table", + "html": "
Update TypeMethodDice Coefficient ↑Average Surface Distance ↓
D1D2D3D4Avg.D1D2D3D4Avg.
Direct Inference55.3554.2061.5456.6756.9410.2211.3915.2715.9513.20
Normalization updateTENT (ICLR'21)49.6652.9659.3955.3854.349.9410.4315.8516.1213.08
InTEnt (CVPR'24)50.9152.1245.8550.3849.819.3510.9922.5719.0115.48
GraTa (AAAI'25)56.3555.1463.6658.1258.319.2610.4414.3615.1512.30
PASS (TMI'24)55.2454.8061.9452.8856.2110.5811.0115.0220.1413.99
Full updatesMEMO (NeurIPS'22)56.2055.2061.2057.3457.489.2710.4514.8915.3912.50
CRF-SOD (CVPR'23)54.2054.0861.8957.0756.8110.3710.8315.3615.5813.03
Latent updatesOurs57.9255.4266.0058.6259.498.3510.5814.1414.1511.75
", + "image_path": "dcd1c88bd483377793d6fb8aebc07cb7c8fe25e5aea6e45a3f2ac6a39ec039c4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 489, + 541, + 512 + ], + "lines": [ + { + "bbox": [ + 68, + 489, + 541, + 512 + ], + "spans": [ + { + "bbox": [ + 68, + 489, + 541, + 512 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of TTA results on MRI image SCGM segmentation between different methods. SCGM segmentation only reported PP-based results due to its very challenging intricate structures and appearance perturbations." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "text", + "content": "height " + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "text", + "content": " of the PP, to simulate real-world loose scenarios as follows: " + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "inline_equation", + "content": "x(y)_{tl}^{LP}: x(y)_{tl} - \\sigma \\cdot w(h), x(y)_{lr}^{LP}: x(y)_{lr} + \\sigma \\cdot w(h)" + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "inline_equation", + "content": "x(y)_{tl}" + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "inline_equation", + "content": "x(y)_{lr}" + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "text", + "content": " denote the vertical (horizontal) coordinates of the top-left and lower-right points of the PP. Each loose prompt is represented as " + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "inline_equation", + "content": "(x_{tl}^{LP}, y_{tl}^{LP}, x_{lr}^{LP}, y_{lr}^{LP})" + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "text", + "content": " is set to 0.1 and 0.2, as we observe 1) too small " + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\sigma (< 0.1)" + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "text", + "content": " is close to the PP without obvious changes and 2) too large " + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\sigma (>0.2)" + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "text", + "content": " will usually result in significantly overlaps with other RoIs with unacceptable initial results. We report the average of different loose rates for comprehensive evaluation. (iii) Shrinkage prompts (SP): Similarly, we define the shrinkage ratio " + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "text", + "content": " to simulate the scenario that the bounding box prompt shifts to the interior of the PP as follows: " + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "inline_equation", + "content": "x(y)_{tl}^{SP}: x(y)_{tl} + \\gamma \\cdot w(h), x(y)_{lr}^{SP}: x(y)_{lr} - \\gamma \\cdot w(h)" + }, + { + "bbox": [ + 68, + 524, + 296, + 715 + ], + "type": "text", + "content": ". Each shrinkage prompt is represented" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 523, + 541, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 541, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 541, + 560 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 313, + 523, + 541, + 560 + ], + "type": "inline_equation", + "content": "(x_{tl}^{SP},y_{tl}^{SP},x_{lr}^{SP},y_{lr}^{SP})" + }, + { + "bbox": [ + 313, + 523, + 541, + 560 + ], + "type": "text", + "content": ". Similarly, " + }, + { + "bbox": [ + 313, + 523, + 541, + 560 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 523, + 541, + 560 + ], + "type": "text", + "content": " is set to 0.1 and 0.2. We report the average of different shrinkage rates for comprehensive evaluation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 567, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 567, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 312, + 567, + 541, + 723 + ], + "type": "text", + "content": "Baselines. In this study, two kinds of TTA approaches for medical image segmentation are utilized for comparison. (1) Normalization-based TTA: Most baseline methods conduct the update of affine transformation parameters and manipulation of source-trained statistics in BN layers, which the MedSAM lacks due to the usage of LN layers. We, therefore, discard source-trained statistics-related components and only optimize the affine-transformation parameters for baseline methods, including TENT (entropy minimization), InTEnt (foreground-background-balanced entropy weighting), GraTa (augmentation-based gradient alignment), and PASS (input decora" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "20081" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 132, + 121 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 132, + 121 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 132, + 121 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 132, + 121 + ], + "type": "image", + "image_path": "8335dfc2ddad3f2cf9acaacdc3c84db4ed78b5f2dbab02b8e0637480f57b225b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 133, + 71, + 188, + 121 + ], + "blocks": [ + { + "bbox": [ + 133, + 71, + 188, + 121 + ], + "lines": [ + { + "bbox": [ + 133, + 71, + 188, + 121 + ], + "spans": [ + { + "bbox": [ + 133, + 71, + 188, + 121 + ], + "type": "image", + "image_path": "b8854cfa7bdad07e0cbc26f9997a562d607cbcd164d539630d8eb048a3d29057.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 190, + 72, + 246, + 121 + ], + "blocks": [ + { + "bbox": [ + 190, + 72, + 246, + 121 + ], + "lines": [ + { + "bbox": [ + 190, + 72, + 246, + 121 + ], + "spans": [ + { + "bbox": [ + 190, + 72, + 246, + 121 + ], + "type": "image", + "image_path": "48cc3745ba6e2555f858576859fe914657a51126e49bd64a3eb8271db1be7070.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 247, + 72, + 304, + 121 + ], + "blocks": [ + { + "bbox": [ + 247, + 72, + 304, + 121 + ], + "lines": [ + { + "bbox": [ + 247, + 72, + 304, + 121 + ], + "spans": [ + { + "bbox": [ + 247, + 72, + 304, + 121 + ], + "type": "image", + "image_path": "aac781c46819eefe84498bfe43ad6bc4575f42090109af93f3e13f29b658ea43.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 364, + 121 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 364, + 121 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 364, + 121 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 364, + 121 + ], + "type": "image", + "image_path": "f2b83411b3b336fc2918554b16c1b2164441e88e422fdc837eafe03cb34d31ad.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 71, + 121, + 133, + 169 + ], + "blocks": [ + { + "bbox": [ + 71, + 121, + 133, + 169 + ], + "lines": [ + { + "bbox": [ + 71, + 121, + 133, + 169 + ], + "spans": [ + { + "bbox": [ + 71, + 121, + 133, + 169 + ], + "type": "image", + "image_path": "2056b3dc42a43ba70e316950ad6da654ecf0490c1dcf1c7e1995abd38ad1a7c9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 134, + 121, + 190, + 169 + ], + "blocks": [ + { + "bbox": [ + 134, + 121, + 190, + 169 + ], + "lines": [ + { + "bbox": [ + 134, + 121, + 190, + 169 + ], + "spans": [ + { + "bbox": [ + 134, + 121, + 190, + 169 + ], + "type": "image", + "image_path": "b3fa64e6721c9ecdfe38623fb6d45b7c6f2a0c3497ab809793946259341c1a6d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 190, + 121, + 247, + 169 + ], + "blocks": [ + { + "bbox": [ + 190, + 121, + 247, + 169 + ], + "lines": [ + { + "bbox": [ + 190, + 121, + 247, + 169 + ], + "spans": [ + { + "bbox": [ + 190, + 121, + 247, + 169 + ], + "type": "image", + "image_path": "a329bc565350d7c16c811867415bc5f3c649dfb2f25ea812c4e3cfb40418fab6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 247, + 121, + 304, + 169 + ], + "blocks": [ + { + "bbox": [ + 247, + 121, + 304, + 169 + ], + "lines": [ + { + "bbox": [ + 247, + 121, + 304, + 169 + ], + "spans": [ + { + "bbox": [ + 247, + 121, + 304, + 169 + ], + "type": "image", + "image_path": "454e1cd0268541b851d7ef1cc044798c0cc936cefcadc7e2fa52e6e587ec0b18.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 121, + 371, + 169 + ], + "blocks": [ + { + "bbox": [ + 306, + 121, + 371, + 169 + ], + "lines": [ + { + "bbox": [ + 306, + 121, + 371, + 169 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 371, + 169 + ], + "type": "image", + "image_path": "177140834aa9cfb310e60a7642480a327e71a0422358f2adbc12ef36337b3ebf.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 71, + 170, + 134, + 221 + ], + "blocks": [ + { + "bbox": [ + 71, + 170, + 134, + 221 + ], + "lines": [ + { + "bbox": [ + 71, + 170, + 134, + 221 + ], + "spans": [ + { + "bbox": [ + 71, + 170, + 134, + 221 + ], + "type": "image", + "image_path": "410503a618ac85d9ce271c839342f438d16d31253fff6c29ee0b59dda936fc30.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 223, + 372, + 277 + ], + "lines": [ + { + "bbox": [ + 68, + 223, + 372, + 277 + ], + "spans": [ + { + "bbox": [ + 68, + 223, + 372, + 277 + ], + "type": "text", + "content": "Figure 2. Visualized process of latent refinement with iterations. The first, second, and third rows denote the segmentation result, corresponding visualization of channel-average refined latent embedding (which is from the region in bounding box prompt and has been fused with the prompt for a better view), and corresponding DAL-CRF loss for each spatial position. You may zoom in for a better view." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 135, + 170, + 190, + 221 + ], + "blocks": [ + { + "bbox": [ + 135, + 170, + 190, + 221 + ], + "lines": [ + { + "bbox": [ + 135, + 170, + 190, + 221 + ], + "spans": [ + { + "bbox": [ + 135, + 170, + 190, + 221 + ], + "type": "image", + "image_path": "5eb53ed1ccbaf88dc2a92044519d2d6cb845941a4cae5de354b426b84914977c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 190, + 170, + 246, + 221 + ], + "blocks": [ + { + "bbox": [ + 190, + 170, + 246, + 221 + ], + "lines": [ + { + "bbox": [ + 190, + 170, + 246, + 221 + ], + "spans": [ + { + "bbox": [ + 190, + 170, + 246, + 221 + ], + "type": "image", + "image_path": "9c99c923fc7a2dbda4ce6d619541d9cd9be0e5a1497ef1f5d31adc1ef218b523.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 247, + 170, + 304, + 221 + ], + "blocks": [ + { + "bbox": [ + 247, + 170, + 304, + 221 + ], + "lines": [ + { + "bbox": [ + 247, + 170, + 304, + 221 + ], + "spans": [ + { + "bbox": [ + 247, + 170, + 304, + 221 + ], + "type": "image", + "image_path": "5913bd22dec9afd75d38b5010f23e210b39c4b79b82d02a82f9a01f979e6a487.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 170, + 371, + 221 + ], + "blocks": [ + { + "bbox": [ + 306, + 170, + 371, + 221 + ], + "lines": [ + { + "bbox": [ + 306, + 170, + 371, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 170, + 371, + 221 + ], + "type": "image", + "image_path": "9315e7c096c20f395efbab305ec890adf41a33bd637ba0792d755cc1bfde0acf.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 377, + 72, + 539, + 123 + ], + "blocks": [ + { + "bbox": [ + 454, + 65, + 463, + 71 + ], + "lines": [ + { + "bbox": [ + 454, + 65, + 463, + 71 + ], + "spans": [ + { + "bbox": [ + 454, + 65, + 463, + 71 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 377, + 72, + 539, + 123 + ], + "lines": [ + { + "bbox": [ + 377, + 72, + 539, + 123 + ], + "spans": [ + { + "bbox": [ + 377, + 72, + 539, + 123 + ], + "type": "table", + "html": "
KD1D2D3D4Avg.
293.2487.0692.5186.5189.75
494.1687.1691.8387.5790.14
694.6787.0892.3487.9290.37
894.8686.5490.8188.0190.03
", + "image_path": "897899b585be90332e5d615c9f606212631fc5d15f3ea3f2bc5c8dbcf2514810.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 377, + 136, + 539, + 200 + ], + "blocks": [ + { + "bbox": [ + 454, + 127, + 463, + 134 + ], + "lines": [ + { + "bbox": [ + 454, + 127, + 463, + 134 + ], + "spans": [ + { + "bbox": [ + 454, + 127, + 463, + 134 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 377, + 136, + 539, + 200 + ], + "lines": [ + { + "bbox": [ + 377, + 136, + 539, + 200 + ], + "spans": [ + { + "bbox": [ + 377, + 136, + 539, + 200 + ], + "type": "table", + "html": "
LossD1D2D3D4Avg.
DI89.1984.3187.9282.3985.95
\\( \\mathcal{L}_{\\text{InTEnt}} \\)79.3484.8988.8386.0884.78
\\( \\mathcal{L}_{\\text{TENT}} \\)85.0984.9989.4685.5986.28
\\( \\mathcal{L}_{\\text{CRF-SOD}} \\)88.9386.1788.4287.3387.71
\\( \\mathcal{L}_{\\text{DAL-CRF}} \\)94.1686.9690.8587.5789.88
", + "image_path": "319d69781d59d710475f75ecbf83b1bd8eb56d807467904e59386a0c1823c8ec.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 71, + 286, + 293, + 340 + ], + "blocks": [ + { + "bbox": [ + 372, + 211, + 544, + 276 + ], + "lines": [ + { + "bbox": [ + 372, + 211, + 544, + 276 + ], + "spans": [ + { + "bbox": [ + 372, + 211, + 544, + 276 + ], + "type": "text", + "content": "Table 5. (a) Hyperparameter Analyses in terms of " + }, + { + "bbox": [ + 372, + 211, + 544, + 276 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 372, + 211, + 544, + 276 + ], + "type": "text", + "content": ". (b) Integration of the latent refinement framework and other TTA loss functions " + }, + { + "bbox": [ + 372, + 211, + 544, + 276 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 372, + 211, + 544, + 276 + ], + "type": "text", + "content": " s. Our proposed DAL-CRF, where DI denotes direct inference without adaptation on OD segmentation." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 286, + 293, + 340 + ], + "lines": [ + { + "bbox": [ + 71, + 286, + 293, + 340 + ], + "spans": [ + { + "bbox": [ + 71, + 286, + 293, + 340 + ], + "type": "table", + "html": "
MethodD1D2D3D4Avg.
Direct Inference89.1984.3187.9282.3985.95
+ LEM85.8285.3188.1083.2186.01
+ LDAL-CRF94.1686.9690.8587.5789.88
+ LEM + LDAL-CRF94.1687.1691.8387.4290.14
DenseCRF [11]-Postprocessing82.7079.3980.9073.7679.18
", + "image_path": "79b39651b916e9ac2b8f1068b8c75122345b4f615646509e9c94e64dbd123660.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 68, + 349, + 294, + 371 + ], + "lines": [ + { + "bbox": [ + 68, + 349, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 68, + 349, + 294, + 371 + ], + "type": "text", + "content": "Table 6. Ablation Study on Different Components of the OD segmentation with loose prompts, in terms of Dice score." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 68, + 381, + 295, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 381, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 68, + 381, + 295, + 477 + ], + "type": "text", + "content": "tor and alternating momentum updating strategy). (2) Full update-based TTA: The whole model parameters are updated based on customized losses, including MEMO (augmentation-based marginal entropy minimization) and CRF-SOD (sparse CRF regularization loss). We leverage their open-source codes to apply into the architecture of MedSAM. To be fair, all baseline methods adopt a one-step optimization strategy." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 68, + 480, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 480, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 68, + 480, + 295, + 612 + ], + "type": "text", + "content": "Quantitative Analysis. From Tables 2, 3, and 4, some conclusions can be summarized. (1) Our proposed method without parametric updates surpasses all baselines in most cases with the best or second-best performance, showcasing the effectiveness of latent refinement. (2) It seems that full parameter updates are more suitable to large size RoIs than normalization-based updates, e.g., the OD. (3) For very challenging OC and SCGM tasks, most parametric update-based approaches have compromised performance compared with direct inference, but our proposed method performs well." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 68, + 615, + 295, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 615, + 295, + 722 + ], + "spans": [ + { + "bbox": [ + 68, + 615, + 295, + 722 + ], + "type": "text", + "content": "Qualitative Analysis. From Figure 2, our proposed method can achieve gradually better TTA segmentation with iterations (Note: we reduce the learning rate with multistep optimization rather than a one-step manner for understanding the optimization process better). Meanwhile, the latent embedding is gradually unsmoothed, which may be reasonable as the DAL-CRF loss enforces more statistics over the input image, leading to more discriminative information. The gradual decrease of DAL-" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 289, + 540, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 540, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 540, + 312 + ], + "type": "text", + "content": "CRF loss in each spatial point of the latent embedding is also observed, showcasing its effectiveness." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 314, + 541, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 541, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 541, + 456 + ], + "type": "text", + "content": "Computation Complexity. As illustrated in Table 1, we achieve around " + }, + { + "bbox": [ + 313, + 314, + 541, + 456 + ], + "type": "inline_equation", + "content": "7 \\times" + }, + { + "bbox": [ + 313, + 314, + 541, + 456 + ], + "type": "text", + "content": " GFLOPs reduction compared with common usages. This is reasonable as 1) sophisticated augmentations (e.g., MEMO) and multiple gradient backward processes (e.g., GraTa) are computationally intensive. 2) For normalization-based updates, the gradient computation over normalization layers still passes through the entire model during backpropagation, which is unfavorable as MedSAM's image encoder is sophisticated with considerable consumption. 3) We circumvent any augmentation and the forward/backward process to the image encoder in a more efficient manner." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 458, + 541, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 541, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 541, + 601 + ], + "type": "text", + "content": "Ablation Study & Other TTA losses under the latent refinement framework. As shown in Table 6, " + }, + { + "bbox": [ + 313, + 458, + 541, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DAL - CRF}}" + }, + { + "bbox": [ + 313, + 458, + 541, + 601 + ], + "type": "text", + "content": " significantly contributes more compared to entropy minimization loss and simply post-processing predictions fail to achieve acceptable performance on challenging medical images with strong appearance perturbations. Meanwhile, " + }, + { + "bbox": [ + 313, + 458, + 541, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{DAL - CRF}}" + }, + { + "bbox": [ + 313, + 458, + 541, + 601 + ], + "type": "text", + "content": " still performs better, even though other TTA losses are integrated into the latent refinement framework as shown in 5(b). Compared to the CRF-SOD (using input and output spaces), our DAL-CRF exhibits superiority using the latent-based co-occurrence regularization for foundation models." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 605, + 387, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 387, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 387, + 618 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 541, + 723 + ], + "type": "text", + "content": "We propose a novel TTA framework for foundation medical segmentation without parametric updates. Our theoretical analysis shows that direct latent refinement is feasible within the MedSAM architecture. By maximizing factorized conditional probabilities by our DAL-CRF loss combined with entropy minimization, we achieve better performance for effectiveness and efficiency." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "20082" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 69, + 72, + 167, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 72, + 167, + 85 + ], + "spans": [ + { + "bbox": [ + 69, + 72, + 167, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 91, + 296, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 91, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 68, + 91, + 296, + 178 + ], + "type": "text", + "content": "This work was supported in part by the Hong Kong Innovation and Technology Commission (ITC) (InnoHK Project CIMDA), in part by the Institute of Digital Medicine of City University of Hong Kong (CityU) under Project 9229503, in part by Chow Sang Sang Donation and Matching Fund (Project 9229161), and in part by the CityU under Project 7006087." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 193, + 128, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 193, + 128, + 205 + ], + "spans": [ + { + "bbox": [ + 70, + 193, + 128, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 212, + 296, + 723 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 74, + 212, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 212, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 74, + 212, + 296, + 267 + ], + "type": "text", + "content": "[1] Kecheng Chen, Pingping Zhang, Tiexin Qin, Shiqi Wang, Hong Yan, and Haoliang Li. Test-time adaptation for image compression with distribution regularization. In The Thirteenth International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 269, + 296, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 269, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 74, + 269, + 296, + 313 + ], + "type": "text", + "content": "[2] Ziyang Chen, Yiwen Ye, Yongsheng Pan, and Yong Xia. Gradient alignment improves test-time adaptation for medical image segmentation. The 39th Annual AAAI Conference on Artificial Intelligence, 2025. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 316, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 316, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 74, + 316, + 296, + 369 + ], + "type": "text", + "content": "[3] Mohammad Zalbagi Darestani, Jiayu Liu, and Reinhard Heckel. Test-time training can close the natural distribution shift performance gap in deep learning based compressed sensing. In International Conference on Machine Learning, pages 4754-4776. PMLR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 372, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 372, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 74, + 372, + 294, + 426 + ], + "type": "text", + "content": "[4] Zeshuai Deng, Zhuokun Chen, Shuaicheng Niu, Thomas Li, Bohan Zhuang, and Mingkui Tan. Efficient test-time adaptation for super-resolution with second-order degradation and reconstruction. Advances in Neural Information Processing Systems, 36:74671-74701, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 428, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 428, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 74, + 428, + 294, + 472 + ], + "type": "text", + "content": "[5] JCMSA Djelouah and Christopher Schroers. Content adaptive optimization for neural image compression. In Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit, pages 1-5, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 74, + 474, + 294, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 474, + 294, + 539 + ], + "spans": [ + { + "bbox": [ + 74, + 474, + 294, + 539 + ], + "type": "text", + "content": "[6] Haoyu Dong, Nicholas Konz, Hanxue Gu, and Maciej A Mazurowski. Medical image segmentation with intent: Integrated entropy weighting for single image test-time adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5046-5055, 2024. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 541, + 294, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 541, + 294, + 607 + ], + "spans": [ + { + "bbox": [ + 74, + 541, + 294, + 607 + ], + "type": "text", + "content": "[7] Yunhe Gao, Rui Huang, Yiwei Yang, Jie Zhang, Kainan Shao, Changjuan Tao, Yuanyuan Chen, Dimitris N. Metaxas, Hongsheng Li, and Ming Chen. Focusnetv2: Imbalanced large and small organ segmentation with adversarial shape constraint for head and neck ct images. Medical Image Analysis, 67:101831, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 74, + 609, + 294, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 609, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 74, + 609, + 294, + 642 + ], + "type": "text", + "content": "[8] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 74, + 643, + 294, + 697 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 643, + 294, + 697 + ], + "spans": [ + { + "bbox": [ + 74, + 643, + 294, + 697 + ], + "type": "text", + "content": "[9] Tianyu Huang, Tao Zhou, Weidi Xie, Shuo Wang, Qi Dou, and Yizhe Zhang. Improving segment anything on the fly: Auxiliary online learning and adaptive fusion for medical image segmentation. arXiv preprint arXiv:2406.00956, 2024.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 70, + 700, + 294, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 700, + 294, + 723 + ], + "spans": [ + { + "bbox": [ + 70, + 700, + 294, + 723 + ], + "type": "text", + "content": "[10] Yiming Ju, Ziyi Ni, Xingrun Xing, Zhixiong Zeng, Hanyu Zhao, Siqi Fan, and Zheng Zhang. Mitigating" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 541, + 723 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 541, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 541, + 128 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 541, + 128 + ], + "type": "text", + "content": "training imbalance in LLM fine-tuning via selective parameter merging. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 15952-15959, Miami, Florida, USA, 2024. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 540, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 540, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 540, + 172 + ], + "type": "text", + "content": "[11] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 4, 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 540, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 540, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 540, + 228 + ], + "type": "text", + "content": "[12] Haoliang Li, YuFei Wang, Renjie Wan, Shiqi Wang, TieQiang Li, and Alex Kot. Domain generalization for medical imaging classification with linear-dependency regularization. Advances in Neural Information Processing Systems, 33:3118-3129, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 540, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 540, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 540, + 274 + ], + "type": "text", + "content": "[13] Zhenbing Liu, Fengfeng Wu, Yumeng Wang, Mengyu Yang, and Xipeng Pan. Fedcl: Federated contrastive learning for multi-center medical image classification. Pattern Recognition, 143:109739, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 275, + 540, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 540, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 540, + 330 + ], + "type": "text", + "content": "[14] Yue Lv, Jinxi Xiang, Jun Zhang, Wenming Yang, Xiao Han, and Wei Yang. Dynamic low-rank instance adaptation for universal neural image compression. In Proceedings of the 31st ACM International Conference on Multimedia, pages 632-642, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 540, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 540, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 540, + 364 + ], + "type": "text", + "content": "[15] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 540, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 540, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 540, + 409 + ], + "type": "text", + "content": "[16] Yuchen Mao, Hongwei Bran Li, LAI Yinyi, Giorgos Papanastasiou, Peng Qi, Yunjie Yang, and Chengjia Wang. Semi-supervised medical image segmentation via knowledge mining from large models. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 540, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 540, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 540, + 464 + ], + "type": "text", + "content": "[17] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 466, + 540, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 540, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 540, + 521 + ], + "type": "text", + "content": "[18] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 522, + 540, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 540, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 540, + 576 + ], + "type": "text", + "content": "[19] Subhadeep Roy, Shankhanil Mitra, Soma Biswas, and Rajiv Soundararajan. Test time adaptation for blind image quality assessment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16742-16751, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 578, + 540, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 540, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 540, + 632 + ], + "type": "text", + "content": "[20] Robin Schon, Julian Lorenz, Katja Ludwig, and Rainer Lienhart. Adapting the segment anything model during usage in novel situations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3616-3626, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 634, + 540, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 634, + 540, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 634, + 540, + 689 + ], + "type": "text", + "content": "[21] Sheng Shen, Huanjing Yue, and Jingyu Yang. Decap dapter: Exploring efficient decoder-side adapter for bridging screen content and natural image compression. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12887-12896, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 689, + 540, + 723 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 689, + 540, + 723 + ], + "spans": [ + { + "bbox": [ + 316, + 689, + 540, + 723 + ], + "type": "text", + "content": "[22] Charles Sutton, Andrew McCallum, et al. An introduction to conditional random fields. Foundations and Trends® in Machine Learning, 4(4):267-373, 2012. 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "20083" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 70, + 72, + 294, + 653 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 70, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 294, + 127 + ], + "type": "text", + "content": "[23] Koki Tsubota, Hiroaki Akutsu, and Kiyoharu Aizawa. Universal deep image compression via content-adaptive optimization with adapters. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2529-2538, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 71, + 129, + 294, + 183 + ], + "type": "text", + "content": "[24] Olga Veksler. Test time adaptation with regularized loss for weakly supervised salient object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7360-7369, 2023. 1, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 71, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 71, + 185, + 294, + 228 + ], + "type": "text", + "content": "[25] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 230, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 230, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 71, + 230, + 294, + 284 + ], + "type": "text", + "content": "[26] Shujun Wang, Lequan Yu, Kang Li, Xin Yang, Chi-Wing Fu, and Pheng-Ann Heng. Dofe: Domain-oriented feature embedding for generalizable fundus image segmentation on unseen datasets. IEEE Transactions on Medical Imaging, 39(12):4237-4248, 2020. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 285, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 285, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 71, + 285, + 294, + 350 + ], + "type": "text", + "content": "[27] Ruxue Wen, Hangjie Yuan, Dong Ni, Wenbo Xiao, and Yaoyao Wu. From denoising training to test-time adaptation: Enhancing domain generalization for medical image segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 464-474, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 353, + 294, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 353, + 294, + 405 + ], + "spans": [ + { + "bbox": [ + 71, + 353, + 294, + 405 + ], + "type": "text", + "content": "[28] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 408, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 408, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 71, + 408, + 294, + 452 + ], + "type": "text", + "content": "[29] Guoping Xu, Xiaoxue Qian, Hua Chieh Shao, Jax Luo, Weiguo Lu, and You Zhang. A sam-guided and match-based semi-supervised segmentation framework for medical imaging. arXiv preprint arXiv:2411.16949, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 71, + 453, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 453, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 71, + 453, + 294, + 496 + ], + "type": "text", + "content": "[30] Chuyan Zhang, Hao Zheng, Xin You, Yefeng Zheng, and Yun Gu. Pass: test-time prompting to adapt styles and semantic shapes in medical image segmentation. IEEE Transactions on Medical Imaging, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 71, + 498, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 498, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 71, + 498, + 294, + 563 + ], + "type": "text", + "content": "[31] Yizhe Zhang, Tao Zhou, Shuo Wang, Peixian Liang, Yejia Zhang, and Danny Z Chen. Input augmentation with sam: Boosting medical image segmentation with segmentation foundation model. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 129-139. Springer, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 71, + 565, + 294, + 607 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 565, + 294, + 607 + ], + "spans": [ + { + "bbox": [ + 71, + 565, + 294, + 607 + ], + "type": "text", + "content": "[32] Chen, Ziyang and Ye, Yiwen and Pan, Yongsheng and Xia, Yong. Gradient Alignment Improves Test-Time Adaptation for Medical Image Segmentation. In AAAI, 2025. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 71, + 609, + 294, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 609, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 71, + 609, + 294, + 653 + ], + "type": "text", + "content": "[33] Zhang, Marvin and Levine, Sergey and Finn, Chelsea. Memo: Test time robustness via adaptation and augmentation Segmentation. In Advances in neural information processing systems, pages 38629-38642, 2022. 1" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "spans": [ + { + "bbox": [ + 293, + 753, + 317, + 762 + ], + "type": "text", + "content": "20084" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/885bbebf-0678-4829-a8d8-7dbca73ecb4a_content_list.json b/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/885bbebf-0678-4829-a8d8-7dbca73ecb4a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..305291568b651c6dae88a56cae6d712dd2d5b827 --- /dev/null +++ b/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/885bbebf-0678-4829-a8d8-7dbca73ecb4a_content_list.json @@ -0,0 +1,1686 @@ +[ + { + "type": "text", + "text": "Text Embedding Knows How to Quantize Text-Guided Diffusion Models", + "text_level": 1, + "bbox": [ + 135, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongjae Lee Myungjun Son Dongjea Kang Seung-Won Jung\\* Korea University", + "bbox": [ + 212, + 180, + 789, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jimmy9704,sonbill,kedjk11,swjung83}@korea.ac.kr", + "bbox": [ + 272, + 218, + 723, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 328, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the success of diffusion models in image generation tasks such as text-to-image, the enormous computational complexity of diffusion models limits their use in resource-constrained environments. To address this, network quantization has emerged as a promising solution for designing efficient diffusion models. However, existing diffusion model quantization methods do not consider input conditions, such as text prompts, as an essential source of information for quantization. In this paper, we propose a novel quantization method dubbed Quantization of Language-to-Image diffusion models using text Prompts (QLIP). QLIP leverages text prompts to guide the selection of bit precision for every layer at each time step. In addition, QLIP can be seamlessly integrated into existing quantization methods to enhance quantization efficiency. Our extensive experiments demonstrate the effectiveness of QLIP in reducing computational complexity and improving the quality of the generated images across various datasets. The source code is available at our project page https://github.com/jimmy9704/QLIP.", + "bbox": [ + 88, + 301, + 485, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 636, + 223, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have achieved remarkable success in various generative tasks, particularly in text-to-image generation [39, 40, 57]. However, the huge computational complexity of diffusion models restricts their practical applicability, especially in resource-constrained environments. Specifically, diffusion models typically require hundreds of denoising steps to produce high-quality samples, making them considerably slower than generative adversarial networks (GANs) [10]. To this end, many studies introduced advanced training-free samplers to reduce the number of denoising iterations [2, 29, 31, 48]. However, the decrease in iterations alone is insufficient since the denoising models used in each iteration possess billions of parameters [4, 37, 42], leading to increased computational com", + "bbox": [ + 89, + 662, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plexity and high memory consumption.", + "bbox": [ + 511, + 270, + 774, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Besides other approaches toward light-weight network designs, such as network architecture search [8, 28], network pruning [12, 50], and knowledge distillation [23, 56], network quantization has received particular attention due to its simplicity, the increasing demand for low-precision operations, and related hardware support [9, 55, 58]. Although various network quantization methods have been proposed for convolutional neural networks (CNNs), the direct use of such methods for diffusion models has shown significant performance degradation [45]. As tailored solutions for diffusion models, several methods have been proposed considering the time step of diffusion models for quantization. For example, PTQ4DM [45] performed 8-bit quantization by constructing a calibration dataset based on the time step. Q-Diffusion [24] introduced an advanced method based on BRECQ [25] and performed the performance analysis of quantization methods for text-guided diffusion models. In addition, TDQ [47] performed quantization with varying activation scaling based on the diffusion time step. However, these methods do not use the text prompt provided to the diffusion model as valuable information for quantization, leaving room for further improvement.", + "bbox": [ + 511, + 287, + 906, + 635 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meanwhile, recent studies highlight the advantages of dynamic bit precision quantization considering input conditions [30, 52]. For example, in the case of image superresolution, CADyQ [18] and AdaBM [17] performed dynamic quantization according to the complexity of the images and the importance of layers. RefQSR [22] clustered the image patches and assigned high-bit and low-bit precisions to the reference and query patches, respectively. These dynamic quantization methods allow for the cost-effective assignment of bit precision, enhancing quantization efficiency. However, such input-adaptive quantization methods have not yet been explored in diffusion models.", + "bbox": [ + 511, + 638, + 908, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a novel quantization method, dubbed Quantization of Language-to-Image diffusion models using text Prompts (QLIP), which determines the bit precision of diffusion models according to the text prompt. Figure 1 shows our observation: The quality of the gener", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* corresponding author", + "bbox": [ + 107, + 887, + 232, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15426", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4eb91b1bf14e292483921a31f8f9a8f863e441e39dd39c885c7b363aaf81194f.jpg", + "image_caption": [ + "Figure 1. Examples of the images generated by Stable Diffusion v1.4 using the text prompts given below images. When the text prompt details the shape, texture, and properties for the target object, the generated images have rich details and textures, as shown on the right; whereas the generated images are not very different from the images generated using only the keyword when the text prompt is less specific to the target object, as shown on the left." + ], + "image_footnote": [ + "Pikachu, the symbol and eternal mascot of Pokémon, is the most famous of over 1,000 Pokémon and is a Pokémon with an electric attribute. It appeared in various video games, TV shows, movies, and merchandise since its introduction in the late 1990s." + ], + "bbox": [ + 119, + 89, + 235, + 179 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ec0073a3ac985fc020281a1356d7493a41b8f94db770caf202fbbe6f261ae37a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 89, + 359, + 179 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bb842b45fe9d55c3ed0f3fbb581de141ca28649f5f0ab8b643855dd452716366.jpg", + "image_caption": [ + "Pikachu" + ], + "image_footnote": [], + "bbox": [ + 379, + 89, + 496, + 179 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/13febf055745aa57c7ff2933467dfa31a53dd9f95741acbbdc98aed283bc802d.jpg", + "image_caption": [], + "image_footnote": [ + "Pikachu is a small, yellow, mouse-like creature with large, black-tipped ears that resemble lightning bolts. It has rosy cheeks, a small mouth, and large, expressive eyes. Pikachu's body is covered in soft, yellow fur, and it has a long, lightning bolt-shaped tail." + ], + "bbox": [ + 504, + 90, + 620, + 178 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/84c03b2ff81d7d3d6496d316cfdd9546f634ee4c1abc10364779983f9d7bca2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 90, + 754, + 178 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8e769f7a176ca997ed6c3bcd792144e82df40fa92ec0790e889b314586e7ef37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 90, + 880, + 178 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ated images increases as the text prompt has more specific information about the image to be generated. This suggests that high-bit precision operations are necessary to generate images with vivid details and textures corresponding to detailed text prompts; whereas, low-bit precision operations can be sufficient to generate images corresponding to text prompts that lack specific guidance on details and textures.", + "bbox": [ + 88, + 334, + 480, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by this observation, we quantized the diffusion model [39] using Q-diffusion [24] with high-bit and low-bit settings and compared the quality of the generated images with the full-precision model. As shown in Figure 2, the performance drop of the low-bit model is noticeable when the quality of the generated images is high for the full-precision model. In contrast, when the quality is low for the full-precision model, the high-bit and low-bit models perform similarly to the full-precision model. These results suggest that diffusion models can be efficiently quantized without sacrificing performance if the quality of the generated images can be first predicted and then used to determine bit precision. Accordingly, our proposed QLIP predicts the image quality from the text prompt and determines the quantization bit precision for diffusion models.", + "bbox": [ + 88, + 441, + 482, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, previous studies [13, 24, 45, 47] revealed that the activation distributions of diffusion models vary across layers over time steps. Inspired by these studies, QLIP uses learnable parameters to determine the bit precision for each layer and for each timestep. Our contributions can be summarized as follows:", + "bbox": [ + 89, + 669, + 482, + 758 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, QLIP is the first approach that explicitly utilizes text prompts for dynamic quantization in diffusion models.", + "- We propose leveraging text prompts to predict the image quality, which is used to guide the selection of bit precision for every layer in different denoising time steps.", + "- Our proposed QLIP can be integrated into existing diffusion quantization methods to enhance quantization efficiency." + ], + "bbox": [ + 89, + 761, + 482, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 332, + 650, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Computationally efficient diffusion models", + "text_level": 1, + "bbox": [ + 511, + 358, + 875, + 375 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models generate high-quality images by iteratively removing noise from random noise. Despite their capability to generate high-quality images, diffusion models encounter challenges in real-world applications due to the time-consuming iterative denoising process. Thus, many efforts have been made to make the diffusion process computationally efficient, including reducing the number of iterative diffusion processes via a non-Markovian process [48], adjusting the variance scheduling [33], and employing knowledge distillation [32, 43].", + "bbox": [ + 509, + 381, + 903, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, since the computational complexity of diffusion model architectures also limits their practical applicability, especially in resource-constrained environments, research efforts have been made to lighten denoising models. For example, SnapFusion [26] introduced an efficient U-Net architecture based on Stable Diffusion [39], while BKSDM [19] further lightened the architecture through knowledge distillation.", + "bbox": [ + 511, + 532, + 903, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Quantization of diffusion models", + "text_level": 1, + "bbox": [ + 511, + 666, + 802, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Quantization involves mapping 32-bit floating-point values of weights and activations to lower-bit values. Numerous studies have explored quantization to compress and accelerate neural networks [5, 18], which can be classified into two approaches: quantization-aware training (QAT) and post-training quantization (PTQ). Recent studies [7, 35, 41, 51] on quantizing diffusion models have focused mainly on PTQ due to the inability to use training data for privacy and commercial reasons, as well as the high computational cost of QAT. Specifically, PTQ4DM [45] and Q-diffusion [24] performed quantization of diffusion models by generating a time step-aware calibration set that consists of data sampled across various time steps using the full-precision denoising model. PTQD [13] separated the", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15427", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6ff6701630ba33acf936ede1f816cfe9163784d4bf9677452a8cbfa01de6b369.jpg", + "image_caption": [ + "Figure 2. Analysis of the effect of quantization and the quality of generated images, where the image quality is measured using the GIQA score [11]. On the left side, the x-axis represents the quality of images generated by the full-precision model, and the y-axis represents the quality difference between the generated images using the full-precision and quantized models. The images are generated by Stable Diffusion v1.4 [39] using 5000 captions from the COCO 2017 dataset [27] with a size of $512 \\times 512$ . The high-bit (W4A10) and low-bit (W4A6) quantized models are obtained by applying Q-diffusion [24]. On the right side, examples of images generated using models that are unquantized or quantized w/wo QLIP are shown for two different image quality levels." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 539, + 277 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/80c76b4ebb90825bf8842e67d5e322bd47ed58ffab25c4c264224097e2bc7cd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 90, + 684, + 170 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2f3acb12f551e68b939b3cb4065830534c18ed65279ab62d59565e8ad6c8b99b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 90, + 785, + 169 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9fe47efa25c24fb108e881a9e4640205b270f3ed8d3eaad9928b690b7a828026.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 90, + 888, + 169 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c1d25611b96184d123cbae5f1e8557c49dd2aa2096161f03caad283c5414bc72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 172, + 684, + 267 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a36df503c5555d2cebe04756dd2aec3253cc66cec91e1bb83b8c9178f52becc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 172, + 785, + 266 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4cb782038d7fdd814494d66a7f43d49c5a862caf53c38483fc0545de08a3bc84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 787, + 172, + 888, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "quantization noise from the diffusion noise to reduce quantization errors. Additionally, PTQD introduced a time step-based mixed-precision method but did not employ varying bit precision for each layer. Furthermore, TDQ [47] performed quantization by adaptively scaling the activations for different diffusion time steps and layers, but assigned the same bit precision for all time steps. Although these diffusion model quantization methods have shown progress in reducing the complexity of noise estimation models, they do not adjust quantization parameters based on input conditions such as text prompts.", + "bbox": [ + 88, + 398, + 483, + 565 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Input-adaptive dynamic quantization", + "text_level": 1, + "bbox": [ + 89, + 575, + 413, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent advances in hardware have enabled mixed-precision operations, allowing various layers to use different bit precisions to improve performance and computational efficiency [9, 55]. Many methods have been proposed to determine the bit precision of each layer of CNNs [9, 30, 55]. For example, several input-adaptive dynamic quantization methods [17, 18, 22, 52] have been introduced for super-resolution, which adjust bit precisions based on the complexity of image patches and the sensitivity to quantization. However, diffusion models have not yet explored such input-adaptive quantization methods. Therefore, this paper proposes a novel quantization method for text-guided diffusion models that applies adaptive quantization using input conditions.", + "bbox": [ + 89, + 597, + 483, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 821, + 181, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 847, + 230, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion model. The reverse process of a diffusion model gradually removes noise from random data $\\mathbf{x}_t$ to", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "produce a high-quality image $\\mathbf{x}_0$ , which can be expressed as $p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t) = \\mathcal{N}(\\mathbf{x}_{t - 1};\\mu_\\theta (\\mathbf{x}_t,t),\\sigma_t^2\\mathbf{I})$ . The denoising process at time step $t$ using the noise estimation model $\\epsilon_{\\theta}$ can then be represented as follows:", + "bbox": [ + 511, + 398, + 906, + 460 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {\\theta} (\\mathbf {x} _ {t}, t) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t)\\right), \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 465, + 906, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t},\\bar{\\alpha}_{t}$ , and $\\beta_{t}$ are the parameters that control the strength of the Gaussian noise in each step, and $\\sigma_{t}$ is a variance schedule [16].", + "bbox": [ + 511, + 503, + 905, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion model quantization. Diffusion model quantization reduces the complexity of a noise estimation model that requires hundreds of iterative sampling processes. The denoising process at time step $t$ using the quantized noise estimation model $\\hat{\\epsilon}_{\\theta}$ and the quantized data $\\dot{\\mathbf{x}}_t$ can be expressed as follows:", + "bbox": [ + 511, + 549, + 905, + 638 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\mu} _ {\\theta} \\left(\\hat {\\mathbf {x}} _ {t}, t\\right) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\hat {\\mathbf {x}} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\hat {\\epsilon} _ {\\theta} \\left(\\hat {\\mathbf {x}} _ {t}, t\\right)\\right), \\tag {2} \\\\ w h e r e \\hat {\\epsilon} _ {\\theta} = Q _ {\\mathbf {b} ^ {t}} (\\epsilon_ {\\theta}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 657, + 903, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\mathbf{b}^t\\in \\mathbb{R}^K$ represents a set of quantization bit precisions for $K$ layers at time step $t$ , and $Q_{\\mathbf{b}^t}$ is a function that performs layer-wise and time step-wise quantization using $\\mathbf{b}^t$ . To our knowledge, all previous diffusion quantization methods set the same bit precision to all layers, i.e., $\\mathbf{b}^t (1) = \\mathbf{b}^t (2) = \\dots = \\mathbf{b}^t (K)$ , and only a few methods [13, 51] performed time step-wise dynamic quantization, i.e., $\\mathbf{b}^{t_i}\\neq \\mathbf{b}^{t_j}$ for some time steps $t_i$ and $t_j$ , $t_i\\neq t_j$ . However, such methods do not adapt the bit precision for each layer. Our proposed QLIP determines $\\mathbf{b}^t$ by training so that the quantization error can be effectively reduced without sacrificing the performance of diffusion models.", + "bbox": [ + 511, + 718, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15428", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b4b5aca18b19d9b948d6eda78d52618574b5054bc506997ebc73f8da3e0518e8.jpg", + "image_caption": [ + "Figure 3. Overview of QLIP that consists of two key modules: T2Q module that predicts image quality from the input text prompt and Q2B module that determines bit precisions for all denoising layers and time steps. Only the Q2B module is trained during QLIP training. See Sections 3.3 and 3.4 for the details about the pre-training of the T2Q module and diffusion models." + ], + "image_footnote": [], + "bbox": [ + 122, + 89, + 875, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Framework overview", + "text_level": 1, + "bbox": [ + 89, + 316, + 290, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our observations shown in Figure 2 indicate that bit precisions for quantization can be determined cost-effectively given the quality of the generated image. However, this image quality remains unknown during the inference stage, motivating us to predict it using an available input, i.e., text prompts. Specifically, if text prompts have rich and detailed information, the expected quality of the generated image is high, and vice versa. To this end, we introduce a text-to-quality (T2Q) module $\\phi$ that predicts image quality $q$ from the embedding vector $\\mathbf{z}$ , i.e., $q = \\phi(\\mathbf{z})$ , where $\\mathbf{z}$ is obtained by CLIP-Text [38] using the input text prompt.", + "bbox": [ + 89, + 338, + 483, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, as we discussed, high-bit quantization is required to generate images with vivid details and textures corresponding to detailed text prompts; while low-bit quantization is sufficient to generate images using less specific text prompts. Therefore, we design a quality-to-bit (Q2B) module $\\psi$ that maps image quality $q$ to appropriate bit precision, i.e., $\\mathbf{b}^t = \\psi(q)$ . Note that the Q2B module determines bit precisions for each time step for different layers since the activation distributions of diffusion models vary across layers over time steps.", + "bbox": [ + 89, + 505, + 483, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, the proposed QLIP obtains a quantized diffusion model $\\hat{\\epsilon}_{\\theta} = Q_{\\psi (\\phi (\\mathbf{z}))}(\\epsilon_{\\theta})$ by learning the parameters of $\\psi$ and $\\phi$ . The overview of the proposed QLIP is shown in Figure 3. The following subsections detail the T2Q and Q2B modules and their training procedure and loss functions.", + "bbox": [ + 89, + 656, + 483, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Text-to-Quality estimation", + "text_level": 1, + "bbox": [ + 89, + 756, + 331, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The T2Q module, denoted as $\\phi$ , aims to predict the image quality from the input text prompt. $\\phi$ consists of three linear layers and outputs an image quality score $q$ as $q = \\phi(\\mathbf{z})$ , where $\\mathbf{z} \\in \\mathbb{R}^{C_{clip}}$ represents the CLIP text embedding, and $C_{clip}$ denotes its channel dimension. For the training of the T2Q module, we collect a dataset $\\mathcal{T}_{T2Q} = \\{\\mathbf{z}^i, \\mathbf{x}_0^i\\}_{i=1}^{N_{T2Q}}$ , where $\\mathbf{z}^i$ and $\\mathbf{x}_0^i$ are the $i$ -th sample of the CLIP text embedding and its corresponding generated", + "bbox": [ + 89, + 777, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "image obtained using the full-precision diffusion model, respectively, and $N_{T2Q}$ is the number of images in $\\mathcal{T}_{T2Q}$ . Let $\\bar{q}^i$ represent the (pseudo) ground-truth image quality of $\\mathbf{x}_0^i$ , where we used GIQA [11] as the default quality measure, with an additional normalization to make the range of quality values [0, 1]. See Section 4.4 for the performance analysis with different quality measures. The T2Q module is trained using the loss $L_{t2q}$ :", + "bbox": [ + 511, + 316, + 906, + 439 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t 2 q} = \\frac {1}{N _ {T 2 Q}} \\sum_ {i = 1} ^ {N _ {T 2 Q}} \\left(\\bar {q} ^ {i} - \\phi \\left(\\mathbf {z} ^ {i}\\right)\\right) ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 452, + 906, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once the T2Q module is trained, it is frozen during the training of the Q2B module and served as a quality predictor to help determine quantization bit precisions.", + "bbox": [ + 511, + 506, + 905, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Quality-to-Bit estimation", + "text_level": 1, + "bbox": [ + 511, + 563, + 741, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In text-guided diffusion models, the richer and more detailed the information in the text prompt is, the higher the image quality tends to be, requiring the use of higher bits. Therefore, the image quality $q$ predicted by the T2Q module can serve as a direct hint for bit selection. To this end, the proposed Q2B module obtains a probability vector $\\mathbf{p}_q \\in \\mathbb{R}^K$ for bit selection as follows:", + "bbox": [ + 511, + 585, + 905, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {q} = \\sigma \\left((q - 0. 5) \\mathbf {s} + \\mathbf {o}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 705, + 906, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{s} \\in \\mathbb{R}^K$ and $\\mathbf{o} \\in \\mathbb{R}^K$ represent sets of learnable parameters, and $\\sigma$ denotes the sigmoid function. Furthermore, to adjust the bit precision for each time step, additional learnable parameters $\\mathbf{u}_m^t \\in \\mathbb{R}^K$ and $\\mathbf{u}_h^t \\in \\mathbb{R}^K$ are introduced to obtain probability vectors $\\mathbf{p}_m^t \\in \\mathbb{R}^K$ and $\\mathbf{p}_h^t \\in \\mathbb{R}^K$ as", + "bbox": [ + 511, + 736, + 905, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {j} ^ {t} = \\sigma \\left(\\mathbf {u} _ {j} ^ {t}\\right), \\quad j \\in \\{m, h \\}, t = 1, 2, \\dots , T, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 828, + 903, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ is the total number of denoising time steps.", + "bbox": [ + 511, + 854, + 857, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These probability vectors, $\\mathbf{p}_q$ , $\\mathbf{p}_m^t$ , and $\\mathbf{p}_h^t$ , are used to determine the bit precision for each layer and each time", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15429", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "step. Specifically, we define a set of supported bit precis \nsions as $\\mathcal{B} = \\{b_{low},b_{med},b_{high}\\}$ . Then, the probabilities of \nselecting $b_{low},b_{med}$ , and $b_{high}$ are obtained as", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {b _ {l o w}} ^ {t} = \\left(1 - \\mathbf {p} _ {q}\\right) \\odot \\left(1 - \\mathbf {p} _ {m} ^ {t}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 141, + 356, + 160 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {b _ {m e d}} ^ {t} = \\left(1 - \\mathbf {p} _ {q}\\right) \\odot \\mathbf {p} _ {m} ^ {t} + \\mathbf {p} _ {q} \\odot \\left(1 - \\mathbf {p} _ {h} ^ {t}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 162, + 482, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {b _ {h i g h}} ^ {t} = \\mathbf {p} _ {q} \\odot \\mathbf {p} _ {h} ^ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 183, + 269, + 203 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ represents element-wise multiplication, $\\mathbf{p}_{b_{low}}^{t} \\in \\mathbb{R}^{K}$ , $\\mathbf{p}_{b_{med}}^{t} \\in \\mathbb{R}^{K}$ , and $\\mathbf{p}_{b_{high}}^{t} \\in \\mathbb{R}^{K}$ are the probability vectors for bit selection. Finally, the bit precision for each layer at time step $t$ is determined as follows:", + "bbox": [ + 89, + 207, + 483, + 270 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} ^ {t} (k) = \\operatorname * {a r g m a x} _ {i \\in \\{b _ {\\text {l o w}}, b _ {\\text {m e d}}, b _ {\\text {h i g h}} \\}} \\mathbf {p} _ {i} ^ {t} (k), k = 1, 2, \\dots , K. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 273, + 483, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the training of $\\mathbf{u}_m^t$ and $\\mathbf{u}_h^t$ for all $t$ poses a challenge as $T$ increases. In addition, the activation distributions of the diffusion models are known to be similar to each other for adjacent time steps [13, 47]. Consequently, we learn the parameters for every $M$ time step, i.e., $\\mathbf{u}_j^{Mt}$ and use the same parameters for the intermediate time steps, i.e., $\\mathbf{u}_j^{M(t + 1) - 1} = \\mathbf{u}_j^{M(t + 1) - 2} = \\dots = \\mathbf{u}_j^{Mt}$ , $j\\in \\{m,h\\}$ . Next, previous studies [1, 6] revealed that the initial stage of the reverse process involves generating the rough context of an image corresponding to the input text prompt; thus, the use of low bits in the initial time steps could decrease the correlation between the text prompt and the generated image. Therefore, we set all values in $\\mathbf{p}_q$ to 1 for the first $m$ time steps in Equation (6) to enforce the selection of high bits in the initial stage of the reverse process. In summary, the Q2B module has $2K$ ( $\\mathbf{s}$ and $\\mathbf{o}$ ) and $2KT / M$ ( $\\mathbf{u}_m^t$ and $\\mathbf{u}_h^t$ ) trainable parameters.", + "bbox": [ + 89, + 308, + 483, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The bit selection in Equation (7) is non-differentiable, making it unable to backpropagate. Thus, we adopt a straight-through estimator [3] to render this process differentiable:", + "bbox": [ + 89, + 566, + 483, + 626 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {a}} ^ {t} (k) =\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 630, + 163, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} Q _ {\\mathbf {b} ^ {t} (k)} \\left(\\mathbf {a} ^ {t} (k)\\right) & \\text {f o r w a r d}, \\\\ \\sum_ {i \\in \\left\\{b _ {l o w}, b _ {m e d}, b _ {h i g h} \\right\\}} \\mathbf {p} _ {i} ^ {t} (k) \\cdot Q _ {i} \\left(\\mathbf {a} ^ {t} (k)\\right) & \\text {b a c k w a r d}. \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 651, + 483, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbf{a}^t (k)$ and $\\hat{\\mathbf{a}}^t (k)$ represent the activations of the full-precision and quantized models at time step $t$ of the $k$ -th layer. In other words, during backpropagation, the activations quantized by all bit precisions in $\\mathcal{B}$ are weighted by their corresponding estimated probabilities. Since diffusion models are less sensitive to the quantization of the weights [13, 24], the weights are quantized using a fixed precision, which is empirically chosen as 4 bits for all layers and time steps.", + "bbox": [ + 89, + 704, + 483, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the full precision diffusion model, we obtain its quantized weights by applying the baseline diffusion quantization methods [13, 24]. The quantization parameters, including the maximum and minimum values for clipping, are", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "also predetermined for all bit precisions in $\\mathcal{B}$ . Then, while freezing all the other modules in Figure 3, only the Q2B module is trained using a small calibration set. The Q2B module is trained at randomly sampled time steps $t$ using the loss function $L_{QLIP}$ :", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {Q L I P} = \\left(\\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) - \\hat {\\epsilon} _ {\\theta} (\\hat {\\mathbf {x}} _ {t}, t)\\right) ^ {2} \\\\ + \\lambda_ {b i t} \\left(b _ {h i g h} \\cdot \\sum_ {k = 1} ^ {K} \\mathbf {p} _ {b _ {h i g h}} ^ {t} (k) + b _ {m e d} \\cdot \\sum_ {k = 1} ^ {K} \\mathbf {p} _ {b _ {m e d}} ^ {t} (k)\\right). \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 175, + 905, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, the first term measures the difference between the results of the full-precision and quantized models, and the second term measures the bit lengths for the high- and medium-precision models. $\\lambda_{bit}$ is a weighting parameter that balances between these two terms.", + "bbox": [ + 511, + 252, + 905, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 340, + 645, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 511, + 364, + 700, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and baseline quantization methods. To evaluate the effectiveness of QLIP across different diffusion models, we conducted experiments using two diffusion networks: BK-SDM-Tiny-2M [19] and Stable Diffusion v1.4 [39]. Our experiments involve two datasets, COCO2017 [27] and Conceptual Captions [46]. To train the T2Q module, 10k images were generated by Stable Diffusion v1.4 [39] using 10k text annotations in the COCO2017 training dataset as input text prompts. We generated another 1024 images using unused text annotations in the COCO2017 training dataset to train the Q2B module. We evaluated the effectiveness of our proposed QLIP by applying it to the baseline diffusion quantization methods, Q-diffusion [24] and PTQD [13]. In the implementation of PTQD, we set the noise correction parameter to $b_{med}$ , while noise correction was not applied if the estimated image quality by the T2Q module is below 0.3 or above 0.7. We trained the T2Q module for 3 epochs using the Adam optimizer [20] with a learning rate of 0.001 and the Q2B module for 5000 iterations using BRECQ [25], as adopted in Q-diffusion and PTQD, using RTX 3090. For performance evaluation, following the baseline quantization method [24], we generated 10k images using 10k text annotations in the COCO2017 validation dataset. Moreover, to evaluate the generalization ability of the quantized models, we used Conceptual Captions to generate additional 10k images.", + "bbox": [ + 511, + 387, + 906, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We report widely adopted metrics such as FID [15] and sFID [44] to evaluate image quality. In addition, we use CLIP Score [14] to assess the degree of matching between the generated images and their corresponding text prompts. To quantify computational efficiency, we measure BitOPs [22, 53], which account for the required operations in both the denoising model and addi", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15430", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ca2bb2ebb84b59dc44eadd91a3fcdfcfb45924e662480e934f03e41c8e6b09b7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
COCO2017
MethodBitoptionsFAB↓BitOPs (T)↓FID↓sFID↓CLIP Score↑
BK-SDM [19]W32A3232.0010.4623.7966.190.3069
Q-diffusion [24]W4A1616.001.0330.0273.250.3068
+QLIPW4A{8,16,32}12.140.8830.0173.240.3063
PTQD [13]W4A1616.001.0330.2777.180.3069
+QLIPW4A{8,16,32}12.140.8830.0273.260.3063
", + "bbox": [ + 91, + 89, + 906, + 205 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/004a962e7c72cb89b5b418b838a66c76a8ada03a5f368b33ec81e8cf060cb041.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Conceptual Captions
MethodBitoptionsFAB↓BitOPs (T)↓FID↓sFID↓CLIP Score↑
BK-SDM [19]W32A3232.0010.4620.0047.800.2983
Q-diffusion [24]W4A1616.001.0324.6862.700.2965
+QLIPW4A{8,16,32}10.580.8224.7259.540.2964
PTQD [13]W4A1616.001.0324.6662.600.2965
+QLIPW4A{8,16,32}10.580.8224.7159.250.2966
", + "bbox": [ + 91, + 205, + 906, + 318 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/30cfcdc1a6a3afa28ff14724ca66374030c63e9ecdffaffae304df6cda1fca67.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons of various quantization methods at a resolution of ${512} \\times {512}$ using BK-SDM-Tiny-2M." + ], + "table_footnote": [], + "table_body": "
COCO2017
MethodFAB↓FID↓sFID↓CLIP Score↑
Stable Diffusion [39]32.0022.2365.110.3174
Q-diffusion [24]8.0023.4066.570.3126
+QLIP7.8621.6164.320.3120
PTQD [13]8.0022.7568.630.3126
+QLIP7.8621.3565.810.3120
Conceptual Captions
MethodFAB↓FID↓sFID↓CLIP Score↑
Stable Diffusion [39]32.0015.6745.610.3078
Q-diffusion [24]8.0020.7448.330.2984
+QLIP7.7719.8947.630.2979
PTQD [13]8.0025.4055.010.2972
+QLIP7.7720.4248.830.2975
", + "bbox": [ + 93, + 364, + 480, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tional modules for QLIP implementation, and feature average bit-width (FAB) [17, 18] for the denoising model. In particular, BitOPs are measured as $MACs \\cdot \\frac{b_w}{32} \\cdot \\frac{b_a}{32}$ , where $MACs$ denotes multiply-accumulate computations, and $b_w$ and $b_a$ represent the bitwidths of weights and activations, respectively.", + "bbox": [ + 89, + 688, + 482, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyperparameter settings. We conducted experiments on various hyperparameter settings for the COCO2017 validation dataset using Q-diffusion as the baseline in Stable Diffusion v1.4. First, Figure 4(a) presents the results obtained with various values of the weight term $\\lambda_{bit}$ in Equation (9), where $\\lambda_{bit} = 1$ showed an effective balance between FID and FAB. Next, Figure 4(b) shows the results for different $M$ values in $\\mathbf{u}_j^{Mt}, j\\in \\{m,h\\}$ when the maximum timestep", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7d7ab1170578e7defac034e9ae1cc329b60e5430439d381860a725b32f21f9d9.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons of various quantization methods at a resolution of $512\\times 512$ using Stable Diffusion v1.4. For the bit precision options, W4A8 is used for Q-diffusion and PTQD, and W4A{6,8,10} are used for QLIP." + ], + "table_footnote": [], + "table_body": "
MethodBitoptionsRuntime (s)↓FAB↓FID↓
BK-SDM [19]W32A326.5032.0020.00
Q-diffusion [24]W4A84.538.0028.32
Q-diffusion [24]W4A165.5816.0024.68
+QLIPW4A{8,16,32}4.8510.5824.72
", + "bbox": [ + 514, + 364, + 908, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Runtime comparisons, where W4A{8,16,32} is used for QLIP. Measured on the Conceptual Captions dataset.", + "bbox": [ + 511, + 455, + 906, + 486 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/66a7115134cd2c554b04b9f63bf7ba14c45fb8d1169aae7c49db9a7befd6be79.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 514, + 498, + 705, + 612 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/92b289e7a2dc3dab7c5a4ddf27e1f7efff5167b6e9fa91450f34cedcdfc7c06d.jpg", + "image_caption": [ + "(b)", + "Figure 4. Results on Q-diffusion using different hyperparameter settings of QLIP (W4A{6,8,10}): (a) $\\lambda_{bit}$ and (b) $M$ ." + ], + "image_footnote": [], + "bbox": [ + 707, + 500, + 903, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$T$ was 1000, where we obtained the best scores with $M = 200$ .", + "bbox": [ + 511, + 686, + 906, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative results", + "text_level": 1, + "bbox": [ + 511, + 727, + 702, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with existing quantization methods. To demonstrate the effectiveness and applicability of QLIP, we tested QLIP on two diffusion model quantization methods: Q-diffusion [24] and PTQD [13]. Tables 1 and 2 provide comparison results for the COCO2017 [27] and Conceptual Captions [46] datasets using lightweight diffusion model BK-SDM-Tiny-2M [19] and Stable Diffusion v1.4 [39], respectively.", + "bbox": [ + 511, + 750, + 906, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, for the COCO2017 dataset, QLIP effectively reduced computational complexity of the base", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15431", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ba0280fd6a05472baa38fa2c4ace5ffcc10e002c2a9b7dcf1f64e80c2e8de31.jpg", + "image_caption": [ + "Figure 5. Examples of variations in FAB by QLIP for the texts with different levels of richness and detail, along with the generated images." + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 859, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9452066d6a2a8f82a7aed4389a807bdf6f4ff3ffd6c433e5465dc93f57dcbc78.jpg", + "image_caption": [ + "Figure 6. Examples of the bit selection results and generated images. When the text prompt is specifically describing the image to be generated, the predicted image quality is high, and high and medium bits are assigned for most layers and time steps, as shown on the top left. On the other hand, when the text prompt has less detailed information, the predicted image quality is low, and low bits are assigned for many layers and time steps, as shown on the bottom left. On the right side, the generated images using Q-diffusion with and without applying QLIP are shown, demonstrating the effectiveness of QLIP in saving bits while improving the quality of generated images." + ], + "image_footnote": [], + "bbox": [ + 135, + 325, + 859, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "line diffusion quantization methods. For example, Q-diffusion+QLIP (W4A{8,16,32}) achieved a reduction in FAB by 3.86 while maintaining similar FID and sFID, compared to Q-diffusion (W4A16). Furthermore, the use of QLIP consistently led to improvements when applied to PTQD, as demonstrated in Table 1. For example, PTQD+QLIP (W4A{8,16,32}) enhanced FID and sFID by 0.25 and 3.92, respectively, while also reducing FAB and BitOPs. The results on Conceptual Captions followed trends similar to those observed in COCO2017. For example, Q-diffusion+QLIP (W4A{8,16,32}) demonstrated reduced computational complexity compared to Q-diffusion on Conceptual Captions.", + "bbox": [ + 88, + 655, + 480, + 851 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 presents the results of applying lower-bit precision combinations to Stable Diffusion v1.4. Here, PTQD+QLIP (W4A{6,8,10}) achieved a 4.98 improvement", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in FID compared to PTQD (W4A8), while maintaining similar FAB using bit combinations efficiently. Although our proposed QLIP drives the quantization methods to reduce computational complexity and improve image quality, the CLIP Scores remained similar to those of the baseline quantization methods, indicating that the fitness to the text prompts is not significantly influenced by QLIP. Additional results on larger diffusion models can be found in the supplementary material.", + "bbox": [ + 511, + 655, + 906, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Runtime measurement. Runtime was measured with Cutlass [34], including the T2Q and Q2B modules, as shown in Table 3. The results show that runtime was reduced to a level similar to W4A8 while maintaining an FID score comparable to W4A16. Measurements were conducted on an RTX 3090. Additional analysis on memory requirements can be found in the supplementary material.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15432", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative results", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5 visualizes the changes in FAB by the richness and detail level of the text, along with the generated images. Here, ChatGPT-4o was used to create texts of varying complexity that describe a given keyword. Specifically, the prompt used was \"Create four sentences that describe how 'keyword' looks like, gradually getting richer and longer.\" As can be seen, QLIP assigned higher bits as the text describes the keyword more specifically and in greater detail, demonstrating that it utilizes information from the text to determine bit precisions for effective synthesis of vivid details and textures.", + "bbox": [ + 89, + 112, + 483, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6 shows several results comparing Q-diffusion with and without QLIP when applied to Stable Diffusion v1.4. As can be seen, our QLIP generates more detailed and natural images compared to PTQD by using various bit combinations. Furthermore, Figure 6 visualizes the bit precision maps determined by QLIP. When the T2Q module predicts high image quality from input text prompts, QLIP allocates more bits to enable rendering of vivid and accurate details from rich text prompts. Otherwise, QLIP employs lower bits for layers that are less sensitive to quantization, based on the text prompt and the time step. For example, when given less specific text prompts, the correlation between the generated image and the text prompt is relatively less critical, allowing for low-bit quantization in the cross-attention blocks. These results highlight the effectiveness and significance of dynamically allocating bit precisions according to text prompts.", + "bbox": [ + 89, + 279, + 483, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 545, + 251, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies on the T2Q module. We report the results on the COCO2017 validation dataset using Stable Diffusion v1.4 with Q-diffusion as the baseline quantization method. Table 4 presents the results of the T2Q modules trained using various perceptual image quality metrics, including Realism score [21], CLIP-IQA [54], and GIQA [11]. The performance of the T2Q module was evaluated using Spearman's rank-order correlation coefficient (SROCC) [49] and Pearson's linear correlation coefficient (PLCC) [36] between the actual quality measures and the predictions of the T2Q module. QLIP with GIQA showed the most substantial improvements in FAB and FID along with the highest SROCC and PLCC scores.", + "bbox": [ + 89, + 568, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies on the Q2B module. We conducted experiments on several variants of the Q2B module using Q-diffusion and Stable Diffusion v1.4 on the COCO2017 validation dataset. As shown in Table 5, when using either $\\mathbf{p}_q$ alone or in combination with $\\mathbf{p}_h^t$ , effective reduction in FAB was achieved at the expense of increase in FID. When using $\\mathbf{p}_q$ and $\\mathbf{p}_m^t$ , FID decreased while FAB increased. Finally, when $\\mathbf{p}_q$ , $\\mathbf{p}_h^t$ , and $\\mathbf{p}_m^t$ were all involved, balanced results were obtained in two metrics. Implementation details of", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0cd2d20b96951a579726cc2e44876115471f56ef637a10687a1c82b00a242583.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSROCC↑PLCC↑FAB↓FID↓
Q-diffusion [24]--8.0023.40
+QLIP
w/ Realism score0.51320.50228.1022.18
w/ CLIP-IQA0.71270.70828.5421.81
w/ GIQA0.80470.81087.8621.61
", + "bbox": [ + 517, + 88, + 903, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4ead45ab3e361af5c5db5cb5122037a125bcd30b4fcea725c9d0ffb4c57ed0cc.jpg", + "table_caption": [ + "Table 4. Ablation studies on the T2Q module: For the bit precision options, W4A8 in Q-diffusion and +QLIP W4A{6,8,10} were used." + ], + "table_footnote": [], + "table_body": "
MethodBitoptionsFAB↓FID↓
Q-diffusion [24]W4A88.0023.40
+ pqW4A{6,10}7.5726.91
+ pq + ptW4A{6,8,10}6.7329.37
+ pq + pmtW4A{6,8,10}8.6021.96
+ pq + pmt + pth (QLIP)W4A{6,8,10}7.8621.61
", + "bbox": [ + 517, + 255, + 901, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation studies on the Q2B module.", + "bbox": [ + 571, + 364, + 846, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the ablation studies on the Q2B module can be found in the supplementary material.", + "bbox": [ + 511, + 410, + 903, + 440 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 460, + 633, + 476 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed quantization method for diffusion models, called QLIP, is the first method that uses text prompts for quantization in diffusion models. From our observation that the bit precision for quantization can be effectively determined based on the quality of the generated images, we designed the T2Q module that predicts image quality from input text prompts. We then introduced the Q2B module that allocates appropriate bit precision for each time step for different layers. By integrating the T2Q and Q2B modules into existing diffusion model quantization methods, we demonstrated the effectiveness of the proposed QLIP. QLIP presents a promising solution for quantizing diffusion models due to its ability to improve the quality of generated images using text prompts while reducing computational costs.", + "bbox": [ + 509, + 488, + 906, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future directions. Similar to other generative models, QLIP also has the potential to generate counterfeit images for malicious purposes, potentially undermining trust in visual media. In a future study, we will extend QLIP to consider other input conditions, such as images or segmentation maps, in addition to text prompts.", + "bbox": [ + 511, + 715, + 905, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 828, + 671, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2022-NR070077).", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15433", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. eDiff-I: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 5", + "[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: An analytic estimate of the optimal reverse variance in diffusion probabilistic models. In Proceedings of the International Conference on Learning Representations, 2022. 1", + "[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.5", + "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2023.1", + "[5] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. PACT: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 2", + "[6] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5", + "[7] Huanpeng Chu, Wei Wu, Chengjie Zang, and Kun Yuan. QNCD: Quantization noise correction for diffusion models. arXiv preprint arXiv:2403.19140, 2024. 2", + "[8] Xiangxiang Chu, Shun Lu, Xudong Li, and Bo Zhang. Mix-Path: A unified approach for one-shot neural architecture search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5972–5981, 2023. 1", + "[9] Zhen Dong, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. HAWQ: Hessian aware quantization of neural networks with mixed-precision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 293–302, 2019. 1, 3", + "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proceedings of the Advances in Neural Information Processing Systems, pages 2672–2680, 2014. 1", + "[11] Shuyang Gu, Jianmin Bao, Dong Chen, and Fang Wen. GIQA: Generated image quality assessment. In Proceedings of the European Conference on Computer Vision, pages 369–385, 2020. 3, 4, 8", + "[12] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Proceedings of the Advances in Neural Information Processing Systems, pages 1135-1143, 2015. 1", + "[13] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. PTQD: Accurate post-training quantization" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for diffusion models. In Proceedings of the Advances in Neural Information Processing Systems, pages 13237-13249, 2023. 2, 3, 5, 6", + "[14] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 5", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the Advances in Neural Information Processing Systems, pages 6629-6640, 2017. 5", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 3", + "[17] Cheeun Hong and Kyoung Mu Lee. AdaBM: On-the-fly adaptive bit mapping for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2641-2650, 2024. 1, 3, 6", + "[18] Cheeun Hong, Sungyong Baik, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. CADyQ: Content-aware dynamic quantization for image super-resolution. In Proceedings of the European Conference on Computer Vision, pages 367-383, 2022. 1, 2, 3, 6", + "[19] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. BK-SDM: A lightweight, fast, and cheap version of stable diffusion. In Proceedings of the European Conference on Computer Vision, 2024. 2, 5, 6", + "[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations, 2015. 5", + "[21] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Proceedings of the Advances in Neural Information Processing Systems, pages 3927-3936, 2019. 8", + "[22] Hongjae Lee, Jun-Sang Yoo, and Seung-Won Jung. RefQSR: Reference-based quantization for image super-resolution networks. IEEE Transactions on Image Processing, 33: 2823-2834, 2024. 1, 3, 5", + "[23] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17413-17424, 2023. 1", + "[24] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 1, 2, 3, 5, 6, 8", + "[25] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. BRECQ: Pushing the limit of post-training quantization by block reconstruction. In Proceedings of the International Conference on Learning Representations, 2021. 1, 5" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15434", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. SnapFusion: Text-to-image diffusion model on mobile devices within two seconds. In Proceedings of the Advances in Neural Information Processing Systems, pages 20662-20678, 2024. 2", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755, 2014. 3, 5, 6", + "[28] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In Proceedings of the European Conference on Computer Vision, pages 19–34, 2018. 1", + "[29] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In Proceedings of the International Conference on Learning Representations, 2022. 1", + "[30] Zhenhua Liu, Yunhe Wang, Kai Han, Siwei Ma, and Wen Gao. Instance-aware dynamic neural network quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12434-12443, 2022. 1, 3", + "[31] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Proceedings of the Advances in Neural Information Processing Systems, pages 5775-5787, 2022. 1", + "[32] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 2", + "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Proceedings of the International Conference on Machine Learning, pages 8162-8171, 2021. 2", + "[34] NVIDIA. CUTLASS, 2025. https://github.com/NVIDIA/cutlass.7", + "[35] Nilesh Prasad Pandey, Marios Fournarakis, Chirag Patel, and Markus Nagel. Softmax bias correction for quantized generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshop, pages 1453-1458, 2023. 2", + "[36] Karl Pearson. Notes on regression and inheritance in the case of two parents. Proceedings of the Royal Society of London, 58:240-242, 1895. 8", + "[37] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *SDXL: Improving latent diffusion models for high-resolution image synthesis*. In *Proceedings of the International Conference on Learning Representations*, 2024. 1", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763, 2021. 4", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10674-10685, 2022. 1, 2, 3, 5, 6", + "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1", + "[41] Hyogon Ryu, Seohyun Lim, and Hyunjung Shim. Memory-efficient personalization using quantized diffusion model. In Proceedings of the European Conference on Computer Vision, 2024. 2", + "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In Proceedings of the Advances in Neural Information Processing Systems, pages 36479–36494, 2022. 1", + "[43] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Proceedings of the International Conference on Learning Representations, 2022. 2", + "[44] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In Proceedings of the Advances in Neural Information Processing Systems, pages 2234-2242, 2016. 5", + "[45] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1972-1981, 2023. 1, 2", + "[46] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 2556-2565, 2018. 5, 6", + "[47] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. In Proceedings of the Advances in Neural Information Processing Systems, pages 48686-48698, 2023. 1, 2, 3, 5", + "[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Proceedings of the International Conference on Learning Representations, 2021. 1, 2", + "[49] Charles Spearman. The proof and measurement of association between two things. The American Journal of Psychology, 15(1):72--101, 1961. 8" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15435", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. CHIP: Channel independence-based pruning for compact neural networks. In Proceedings of the Advances in Neural Information Processing Systems, pages 24604-24616, 2021. 1", + "[51] Siao Tang, Xin Wang, Hong Chen, Chaoyu Guan, Zewen Wu, Yansong Tang, and Wenwu Zhu. Post-training quantization with progressive calibration and activation relaxing for text-to-image diffusion models. In Proceedings of the European Conference on Computer Vision, 2024. 2, 3", + "[52] Senmao Tian, Ming Lu, Jiaming Liu, Yandong Guo, Yurong Chen, and Shunli Zhang. CABM: Content-aware bit mapping for single image super-resolution network with large input. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1756-1765, 2023. 1, 3", + "[53] Mart Van Baalen, Christos Louizos, Markus Nagel, Rana Ali Amjad, Ying Wang, Tijmen Blankevoort, and Max Welling. Bayesian bits: Unifying quantization and pruning. In Proceedings of the Advances in Neural Information Processing Systems, pages 5741-5752, 2020. 5", + "[54] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 8", + "[55] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. HAQ: Hardware-aware automated quantization with mixed precision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8612-8620, 2019. 1, 3", + "[56] Chuanguang Yang, Zhulin An, Helong Zhou, Fuzhen Zhuang, Yongjun Xu, and Qian Zhang. Online knowledge distillation via mutual contrastive learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10212-10227, 2023. 1", + "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1", + "[58] Xuechen Zhang, Zheng Li, Samet Oymak, and Jiasi Chen. Text-to-3d generative ai on mobile devices: Measurements and optimizations. In Proceedings of the Workshop on Emerging Multimedia Systems, pages 8–14, 2023. 1" + ], + "bbox": [ + 91, + 90, + 480, + 696 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15436", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/885bbebf-0678-4829-a8d8-7dbca73ecb4a_model.json b/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/885bbebf-0678-4829-a8d8-7dbca73ecb4a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d73a3da45307f2cb9cf23375e7660ca06eb056e0 --- /dev/null +++ b/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/885bbebf-0678-4829-a8d8-7dbca73ecb4a_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.865, + 0.153 + ], + "angle": 0, + "content": "Text Embedding Knows How to Quantize Text-Guided Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.181, + 0.79, + 0.218 + ], + "angle": 0, + "content": "Hongjae Lee Myungjun Son Dongjea Kang Seung-Won Jung\\* Korea University" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.219, + 0.725, + 0.235 + ], + "angle": 0, + "content": "{jimmy9704,sonbill,kedjk11,swjung83}@korea.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.329, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.303, + 0.486, + 0.606 + ], + "angle": 0, + "content": "Despite the success of diffusion models in image generation tasks such as text-to-image, the enormous computational complexity of diffusion models limits their use in resource-constrained environments. To address this, network quantization has emerged as a promising solution for designing efficient diffusion models. However, existing diffusion model quantization methods do not consider input conditions, such as text prompts, as an essential source of information for quantization. In this paper, we propose a novel quantization method dubbed Quantization of Language-to-Image diffusion models using text Prompts (QLIP). QLIP leverages text prompts to guide the selection of bit precision for every layer at each time step. In addition, QLIP can be seamlessly integrated into existing quantization methods to enhance quantization efficiency. Our extensive experiments demonstrate the effectiveness of QLIP in reducing computational complexity and improving the quality of the generated images across various datasets. The source code is available at our project page https://github.com/jimmy9704/QLIP." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.637, + 0.225, + 0.654 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Diffusion models have achieved remarkable success in various generative tasks, particularly in text-to-image generation [39, 40, 57]. However, the huge computational complexity of diffusion models restricts their practical applicability, especially in resource-constrained environments. Specifically, diffusion models typically require hundreds of denoising steps to produce high-quality samples, making them considerably slower than generative adversarial networks (GANs) [10]. To this end, many studies introduced advanced training-free samplers to reduce the number of denoising iterations [2, 29, 31, 48]. However, the decrease in iterations alone is insufficient since the denoising models used in each iteration possess billions of parameters [4, 37, 42], leading to increased computational com" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.271, + 0.776, + 0.286 + ], + "angle": 0, + "content": "plexity and high memory consumption." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.908, + 0.636 + ], + "angle": 0, + "content": "Besides other approaches toward light-weight network designs, such as network architecture search [8, 28], network pruning [12, 50], and knowledge distillation [23, 56], network quantization has received particular attention due to its simplicity, the increasing demand for low-precision operations, and related hardware support [9, 55, 58]. Although various network quantization methods have been proposed for convolutional neural networks (CNNs), the direct use of such methods for diffusion models has shown significant performance degradation [45]. As tailored solutions for diffusion models, several methods have been proposed considering the time step of diffusion models for quantization. For example, PTQ4DM [45] performed 8-bit quantization by constructing a calibration dataset based on the time step. Q-Diffusion [24] introduced an advanced method based on BRECQ [25] and performed the performance analysis of quantization methods for text-guided diffusion models. In addition, TDQ [47] performed quantization with varying activation scaling based on the diffusion time step. However, these methods do not use the text prompt provided to the diffusion model as valuable information for quantization, leaving room for further improvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.909, + 0.822 + ], + "angle": 0, + "content": "Meanwhile, recent studies highlight the advantages of dynamic bit precision quantization considering input conditions [30, 52]. For example, in the case of image superresolution, CADyQ [18] and AdaBM [17] performed dynamic quantization according to the complexity of the images and the importance of layers. RefQSR [22] clustered the image patches and assigned high-bit and low-bit precisions to the reference and query patches, respectively. These dynamic quantization methods allow for the cost-effective assignment of bit precision, enhancing quantization efficiency. However, such input-adaptive quantization methods have not yet been explored in diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose a novel quantization method, dubbed Quantization of Language-to-Image diffusion models using text Prompts (QLIP), which determines the bit precision of diffusion models according to the text prompt. Figure 1 shows our observation: The quality of the gener" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.233, + 0.901 + ], + "angle": 0, + "content": "* corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "15426" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.09, + 0.236, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.108, + 0.186, + 0.374, + 0.242 + ], + "angle": 0, + "content": "Pikachu, the symbol and eternal mascot of Pokémon, is the most famous of over 1,000 Pokémon and is a Pokémon with an electric attribute. It appeared in various video games, TV shows, movies, and merchandise since its introduction in the late 1990s." + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.09, + 0.361, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.09, + 0.497, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.481, + 0.208, + 0.52, + 0.219 + ], + "angle": 0, + "content": "Pikachu" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.621, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.091, + 0.756, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.091, + 0.882, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.628, + 0.181, + 0.892, + 0.236 + ], + "angle": 0, + "content": "Pikachu is a small, yellow, mouse-like creature with large, black-tipped ears that resemble lightning bolts. It has rosy cheeks, a small mouth, and large, expressive eyes. Pikachu's body is covered in soft, yellow fur, and it has a long, lightning bolt-shaped tail." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.253, + 0.908, + 0.308 + ], + "angle": 0, + "content": "Figure 1. Examples of the images generated by Stable Diffusion v1.4 using the text prompts given below images. When the text prompt details the shape, texture, and properties for the target object, the generated images have rich details and textures, as shown on the right; whereas the generated images are not very different from the images generated using only the keyword when the text prompt is less specific to the target object, as shown on the left." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.335, + 0.482, + 0.44 + ], + "angle": 0, + "content": "ated images increases as the text prompt has more specific information about the image to be generated. This suggests that high-bit precision operations are necessary to generate images with vivid details and textures corresponding to detailed text prompts; whereas, low-bit precision operations can be sufficient to generate images corresponding to text prompts that lack specific guidance on details and textures." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.442, + 0.483, + 0.669 + ], + "angle": 0, + "content": "Motivated by this observation, we quantized the diffusion model [39] using Q-diffusion [24] with high-bit and low-bit settings and compared the quality of the generated images with the full-precision model. As shown in Figure 2, the performance drop of the low-bit model is noticeable when the quality of the generated images is high for the full-precision model. In contrast, when the quality is low for the full-precision model, the high-bit and low-bit models perform similarly to the full-precision model. These results suggest that diffusion models can be efficiently quantized without sacrificing performance if the quality of the generated images can be first predicted and then used to determine bit precision. Accordingly, our proposed QLIP predicts the image quality from the text prompt and determines the quantization bit precision for diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Furthermore, previous studies [13, 24, 45, 47] revealed that the activation distributions of diffusion models vary across layers over time steps. Inspired by these studies, QLIP uses learnable parameters to determine the bit precision for each layer and for each timestep. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.762, + 0.483, + 0.806 + ], + "angle": 0, + "content": "- To the best of our knowledge, QLIP is the first approach that explicitly utilizes text prompts for dynamic quantization in diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.808, + 0.482, + 0.852 + ], + "angle": 0, + "content": "- We propose leveraging text prompts to predict the image quality, which is used to guide the selection of bit precision for every layer in different denoising time steps." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.853, + 0.482, + 0.898 + ], + "angle": 0, + "content": "- Our proposed QLIP can be integrated into existing diffusion quantization methods to enhance quantization efficiency." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.762, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.333, + 0.651, + 0.349 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.359, + 0.876, + 0.375 + ], + "angle": 0, + "content": "2.1. Computationally efficient diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.382, + 0.905, + 0.533 + ], + "angle": 0, + "content": "Diffusion models generate high-quality images by iteratively removing noise from random noise. Despite their capability to generate high-quality images, diffusion models encounter challenges in real-world applications due to the time-consuming iterative denoising process. Thus, many efforts have been made to make the diffusion process computationally efficient, including reducing the number of iterative diffusion processes via a non-Markovian process [48], adjusting the variance scheduling [33], and employing knowledge distillation [32, 43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.905, + 0.655 + ], + "angle": 0, + "content": "Meanwhile, since the computational complexity of diffusion model architectures also limits their practical applicability, especially in resource-constrained environments, research efforts have been made to lighten denoising models. For example, SnapFusion [26] introduced an efficient U-Net architecture based on Stable Diffusion [39], while BKSDM [19] further lightened the architecture through knowledge distillation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.667, + 0.803, + 0.683 + ], + "angle": 0, + "content": "2.2. Quantization of diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Quantization involves mapping 32-bit floating-point values of weights and activations to lower-bit values. Numerous studies have explored quantization to compress and accelerate neural networks [5, 18], which can be classified into two approaches: quantization-aware training (QAT) and post-training quantization (PTQ). Recent studies [7, 35, 41, 51] on quantizing diffusion models have focused mainly on PTQ due to the inability to use training data for privacy and commercial reasons, as well as the high computational cost of QAT. Specifically, PTQ4DM [45] and Q-diffusion [24] performed quantization of diffusion models by generating a time step-aware calibration set that consists of data sampled across various time steps using the full-precision denoising model. PTQD [13] separated the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15427" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.54, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.091, + 0.685, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.092, + 0.786, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.092, + 0.889, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.174, + 0.685, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.174, + 0.787, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.174, + 0.889, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.289, + 0.907, + 0.374 + ], + "angle": 0, + "content": "Figure 2. Analysis of the effect of quantization and the quality of generated images, where the image quality is measured using the GIQA score [11]. On the left side, the x-axis represents the quality of images generated by the full-precision model, and the y-axis represents the quality difference between the generated images using the full-precision and quantized models. The images are generated by Stable Diffusion v1.4 [39] using 5000 captions from the COCO 2017 dataset [27] with a size of \\(512 \\times 512\\). The high-bit (W4A10) and low-bit (W4A6) quantized models are obtained by applying Q-diffusion [24]. On the right side, examples of images generated using models that are unquantized or quantized w/wo QLIP are shown for two different image quality levels." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.399, + 0.484, + 0.566 + ], + "angle": 0, + "content": "quantization noise from the diffusion noise to reduce quantization errors. Additionally, PTQD introduced a time step-based mixed-precision method but did not employ varying bit precision for each layer. Furthermore, TDQ [47] performed quantization by adaptively scaling the activations for different diffusion time steps and layers, but assigned the same bit precision for all time steps. Although these diffusion model quantization methods have shown progress in reducing the complexity of noise estimation models, they do not adjust quantization parameters based on input conditions such as text prompts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.576, + 0.414, + 0.592 + ], + "angle": 0, + "content": "2.3. Input-adaptive dynamic quantization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.81 + ], + "angle": 0, + "content": "Recent advances in hardware have enabled mixed-precision operations, allowing various layers to use different bit precisions to improve performance and computational efficiency [9, 55]. Many methods have been proposed to determine the bit precision of each layer of CNNs [9, 30, 55]. For example, several input-adaptive dynamic quantization methods [17, 18, 22, 52] have been introduced for super-resolution, which adjust bit precisions based on the complexity of image patches and the sensitivity to quantization. However, diffusion models have not yet explored such input-adaptive quantization methods. Therefore, this paper proposes a novel quantization method for text-guided diffusion models that applies adaptive quantization using input conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.823, + 0.182, + 0.839 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.231, + 0.863 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Diffusion model. The reverse process of a diffusion model gradually removes noise from random data \\(\\mathbf{x}_t\\) to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.461 + ], + "angle": 0, + "content": "produce a high-quality image \\(\\mathbf{x}_0\\), which can be expressed as \\(p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t) = \\mathcal{N}(\\mathbf{x}_{t - 1};\\mu_\\theta (\\mathbf{x}_t,t),\\sigma_t^2\\mathbf{I})\\). The denoising process at time step \\(t\\) using the noise estimation model \\(\\epsilon_{\\theta}\\) can then be represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.466, + 0.907, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\mu_ {\\theta} (\\mathbf {x} _ {t}, t) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t)\\right), \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.906, + 0.55 + ], + "angle": 0, + "content": "where \\(\\alpha_{t},\\bar{\\alpha}_{t}\\), and \\(\\beta_{t}\\) are the parameters that control the strength of the Gaussian noise in each step, and \\(\\sigma_{t}\\) is a variance schedule [16]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.906, + 0.64 + ], + "angle": 0, + "content": "Diffusion model quantization. Diffusion model quantization reduces the complexity of a noise estimation model that requires hundreds of iterative sampling processes. The denoising process at time step \\( t \\) using the quantized noise estimation model \\( \\hat{\\epsilon}_{\\theta} \\) and the quantized data \\( \\dot{\\mathbf{x}}_t \\) can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.659, + 0.905, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\mu} _ {\\theta} \\left(\\hat {\\mathbf {x}} _ {t}, t\\right) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\hat {\\mathbf {x}} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\hat {\\epsilon} _ {\\theta} \\left(\\hat {\\mathbf {x}} _ {t}, t\\right)\\right), \\tag {2} \\\\ w h e r e \\hat {\\epsilon} _ {\\theta} = Q _ {\\mathbf {b} ^ {t}} (\\epsilon_ {\\theta}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{b}^t\\in \\mathbb{R}^K\\) represents a set of quantization bit precisions for \\(K\\) layers at time step \\(t\\), and \\(Q_{\\mathbf{b}^t}\\) is a function that performs layer-wise and time step-wise quantization using \\(\\mathbf{b}^t\\). To our knowledge, all previous diffusion quantization methods set the same bit precision to all layers, i.e., \\(\\mathbf{b}^t (1) = \\mathbf{b}^t (2) = \\dots = \\mathbf{b}^t (K)\\), and only a few methods [13, 51] performed time step-wise dynamic quantization, i.e., \\(\\mathbf{b}^{t_i}\\neq \\mathbf{b}^{t_j}\\) for some time steps \\(t_i\\) and \\(t_j\\), \\(t_i\\neq t_j\\). However, such methods do not adapt the bit precision for each layer. Our proposed QLIP determines \\(\\mathbf{b}^t\\) by training so that the quantization error can be effectively reduced without sacrificing the performance of diffusion models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15428" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.124, + 0.09, + 0.877, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.249, + 0.908, + 0.293 + ], + "angle": 0, + "content": "Figure 3. Overview of QLIP that consists of two key modules: T2Q module that predicts image quality from the input text prompt and Q2B module that determines bit precisions for all denoising layers and time steps. Only the Q2B module is trained during QLIP training. See Sections 3.3 and 3.4 for the details about the pre-training of the T2Q module and diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.317, + 0.292, + 0.332 + ], + "angle": 0, + "content": "3.2. Framework overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.484, + 0.506 + ], + "angle": 0, + "content": "Our observations shown in Figure 2 indicate that bit precisions for quantization can be determined cost-effectively given the quality of the generated image. However, this image quality remains unknown during the inference stage, motivating us to predict it using an available input, i.e., text prompts. Specifically, if text prompts have rich and detailed information, the expected quality of the generated image is high, and vice versa. To this end, we introduce a text-to-quality (T2Q) module \\(\\phi\\) that predicts image quality \\(q\\) from the embedding vector \\(\\mathbf{z}\\), i.e., \\(q = \\phi(\\mathbf{z})\\), where \\(\\mathbf{z}\\) is obtained by CLIP-Text [38] using the input text prompt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.656 + ], + "angle": 0, + "content": "Next, as we discussed, high-bit quantization is required to generate images with vivid details and textures corresponding to detailed text prompts; while low-bit quantization is sufficient to generate images using less specific text prompts. Therefore, we design a quality-to-bit (Q2B) module \\(\\psi\\) that maps image quality \\(q\\) to appropriate bit precision, i.e., \\(\\mathbf{b}^t = \\psi(q)\\). Note that the Q2B module determines bit precisions for each time step for different layers since the activation distributions of diffusion models vary across layers over time steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.484, + 0.747 + ], + "angle": 0, + "content": "In summary, the proposed QLIP obtains a quantized diffusion model \\(\\hat{\\epsilon}_{\\theta} = Q_{\\psi (\\phi (\\mathbf{z}))}(\\epsilon_{\\theta})\\) by learning the parameters of \\(\\psi\\) and \\(\\phi\\). The overview of the proposed QLIP is shown in Figure 3. The following subsections detail the T2Q and Q2B modules and their training procedure and loss functions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.332, + 0.773 + ], + "angle": 0, + "content": "3.3. Text-to-Quality estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The T2Q module, denoted as \\(\\phi\\), aims to predict the image quality from the input text prompt. \\(\\phi\\) consists of three linear layers and outputs an image quality score \\(q\\) as \\(q = \\phi(\\mathbf{z})\\), where \\(\\mathbf{z} \\in \\mathbb{R}^{C_{clip}}\\) represents the CLIP text embedding, and \\(C_{clip}\\) denotes its channel dimension. For the training of the T2Q module, we collect a dataset \\(\\mathcal{T}_{T2Q} = \\{\\mathbf{z}^i, \\mathbf{x}_0^i\\}_{i=1}^{N_{T2Q}}\\), where \\(\\mathbf{z}^i\\) and \\(\\mathbf{x}_0^i\\) are the \\(i\\)-th sample of the CLIP text embedding and its corresponding generated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.44 + ], + "angle": 0, + "content": "image obtained using the full-precision diffusion model, respectively, and \\(N_{T2Q}\\) is the number of images in \\(\\mathcal{T}_{T2Q}\\). Let \\(\\bar{q}^i\\) represent the (pseudo) ground-truth image quality of \\(\\mathbf{x}_0^i\\), where we used GIQA [11] as the default quality measure, with an additional normalization to make the range of quality values [0, 1]. See Section 4.4 for the performance analysis with different quality measures. The T2Q module is trained using the loss \\(L_{t2q}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.453, + 0.907, + 0.497 + ], + "angle": 0, + "content": "\\[\nL _ {t 2 q} = \\frac {1}{N _ {T 2 Q}} \\sum_ {i = 1} ^ {N _ {T 2 Q}} \\left(\\bar {q} ^ {i} - \\phi \\left(\\mathbf {z} ^ {i}\\right)\\right) ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.553 + ], + "angle": 0, + "content": "Once the T2Q module is trained, it is frozen during the training of the Q2B module and served as a quality predictor to help determine quantization bit precisions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.564, + 0.743, + 0.58 + ], + "angle": 0, + "content": "3.4. Quality-to-Bit estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.906, + 0.694 + ], + "angle": 0, + "content": "In text-guided diffusion models, the richer and more detailed the information in the text prompt is, the higher the image quality tends to be, requiring the use of higher bits. Therefore, the image quality \\( q \\) predicted by the T2Q module can serve as a direct hint for bit selection. To this end, the proposed Q2B module obtains a probability vector \\( \\mathbf{p}_q \\in \\mathbb{R}^K \\) for bit selection as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.706, + 0.907, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {q} = \\sigma \\left((q - 0. 5) \\mathbf {s} + \\mathbf {o}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.737, + 0.906, + 0.827 + ], + "angle": 0, + "content": "where \\(\\mathbf{s} \\in \\mathbb{R}^K\\) and \\(\\mathbf{o} \\in \\mathbb{R}^K\\) represent sets of learnable parameters, and \\(\\sigma\\) denotes the sigmoid function. Furthermore, to adjust the bit precision for each time step, additional learnable parameters \\(\\mathbf{u}_m^t \\in \\mathbb{R}^K\\) and \\(\\mathbf{u}_h^t \\in \\mathbb{R}^K\\) are introduced to obtain probability vectors \\(\\mathbf{p}_m^t \\in \\mathbb{R}^K\\) and \\(\\mathbf{p}_h^t \\in \\mathbb{R}^K\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.829, + 0.905, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {j} ^ {t} = \\sigma \\left(\\mathbf {u} _ {j} ^ {t}\\right), \\quad j \\in \\{m, h \\}, t = 1, 2, \\dots , T, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.855, + 0.859, + 0.87 + ], + "angle": 0, + "content": "where \\(T\\) is the total number of denoising time steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "These probability vectors, \\(\\mathbf{p}_q\\), \\(\\mathbf{p}_m^t\\), and \\(\\mathbf{p}_h^t\\), are used to determine the bit precision for each layer and each time" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15429" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "step. Specifically, we define a set of supported bit precis \nsions as \\(\\mathcal{B} = \\{b_{low},b_{med},b_{high}\\}\\) . Then, the probabilities of \nselecting \\(b_{low},b_{med}\\) , and \\(b_{high}\\) are obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.142, + 0.357, + 0.161 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {b _ {l o w}} ^ {t} = \\left(1 - \\mathbf {p} _ {q}\\right) \\odot \\left(1 - \\mathbf {p} _ {m} ^ {t}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.163, + 0.483, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {b _ {m e d}} ^ {t} = \\left(1 - \\mathbf {p} _ {q}\\right) \\odot \\mathbf {p} _ {m} ^ {t} + \\mathbf {p} _ {q} \\odot \\left(1 - \\mathbf {p} _ {h} ^ {t}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.184, + 0.27, + 0.204 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {b _ {h i g h}} ^ {t} = \\mathbf {p} _ {q} \\odot \\mathbf {p} _ {h} ^ {t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.208, + 0.485, + 0.271 + ], + "angle": 0, + "content": "where \\(\\odot\\) represents element-wise multiplication, \\(\\mathbf{p}_{b_{low}}^{t} \\in \\mathbb{R}^{K}\\), \\(\\mathbf{p}_{b_{med}}^{t} \\in \\mathbb{R}^{K}\\), and \\(\\mathbf{p}_{b_{high}}^{t} \\in \\mathbb{R}^{K}\\) are the probability vectors for bit selection. Finally, the bit precision for each layer at time step \\(t\\) is determined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.275, + 0.484, + 0.304 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} ^ {t} (k) = \\operatorname * {a r g m a x} _ {i \\in \\{b _ {\\text {l o w}}, b _ {\\text {m e d}}, b _ {\\text {h i g h}} \\}} \\mathbf {p} _ {i} ^ {t} (k), k = 1, 2, \\dots , K. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.484, + 0.568 + ], + "angle": 0, + "content": "However, the training of \\(\\mathbf{u}_m^t\\) and \\(\\mathbf{u}_h^t\\) for all \\(t\\) poses a challenge as \\(T\\) increases. In addition, the activation distributions of the diffusion models are known to be similar to each other for adjacent time steps [13, 47]. Consequently, we learn the parameters for every \\(M\\) time step, i.e., \\(\\mathbf{u}_j^{Mt}\\) and use the same parameters for the intermediate time steps, i.e., \\(\\mathbf{u}_j^{M(t + 1) - 1} = \\mathbf{u}_j^{M(t + 1) - 2} = \\dots = \\mathbf{u}_j^{Mt}\\), \\(j\\in \\{m,h\\}\\). Next, previous studies [1, 6] revealed that the initial stage of the reverse process involves generating the rough context of an image corresponding to the input text prompt; thus, the use of low bits in the initial time steps could decrease the correlation between the text prompt and the generated image. Therefore, we set all values in \\(\\mathbf{p}_q\\) to 1 for the first \\(m\\) time steps in Equation (6) to enforce the selection of high bits in the initial stage of the reverse process. In summary, the Q2B module has \\(2K\\) (\\(\\mathbf{s}\\) and \\(\\mathbf{o}\\)) and \\(2KT / M\\) (\\(\\mathbf{u}_m^t\\) and \\(\\mathbf{u}_h^t\\)) trainable parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.627 + ], + "angle": 0, + "content": "The bit selection in Equation (7) is non-differentiable, making it unable to backpropagate. Thus, we adopt a straight-through estimator [3] to render this process differentiable:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.631, + 0.165, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {a}} ^ {t} (k) =\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.652, + 0.484, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} Q _ {\\mathbf {b} ^ {t} (k)} \\left(\\mathbf {a} ^ {t} (k)\\right) & \\text {f o r w a r d}, \\\\ \\sum_ {i \\in \\left\\{b _ {l o w}, b _ {m e d}, b _ {h i g h} \\right\\}} \\mathbf {p} _ {i} ^ {t} (k) \\cdot Q _ {i} \\left(\\mathbf {a} ^ {t} (k)\\right) & \\text {b a c k w a r d}. \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Here, \\(\\mathbf{a}^t (k)\\) and \\(\\hat{\\mathbf{a}}^t (k)\\) represent the activations of the full-precision and quantized models at time step \\(t\\) of the \\(k\\)-th layer. In other words, during backpropagation, the activations quantized by all bit precisions in \\(\\mathcal{B}\\) are weighted by their corresponding estimated probabilities. Since diffusion models are less sensitive to the quantization of the weights [13, 24], the weights are quantized using a fixed precision, which is empirically chosen as 4 bits for all layers and time steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Given the full precision diffusion model, we obtain its quantized weights by applying the baseline diffusion quantization methods [13, 24]. The quantization parameters, including the maximum and minimum values for clipping, are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.169 + ], + "angle": 0, + "content": "also predetermined for all bit precisions in \\(\\mathcal{B}\\). Then, while freezing all the other modules in Figure 3, only the Q2B module is trained using a small calibration set. The Q2B module is trained at randomly sampled time steps \\(t\\) using the loss function \\(L_{QLIP}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.176, + 0.906, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {Q L I P} = \\left(\\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) - \\hat {\\epsilon} _ {\\theta} (\\hat {\\mathbf {x}} _ {t}, t)\\right) ^ {2} \\\\ + \\lambda_ {b i t} \\left(b _ {h i g h} \\cdot \\sum_ {k = 1} ^ {K} \\mathbf {p} _ {b _ {h i g h}} ^ {t} (k) + b _ {m e d} \\cdot \\sum_ {k = 1} ^ {K} \\mathbf {p} _ {b _ {m e d}} ^ {t} (k)\\right). \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.253, + 0.906, + 0.328 + ], + "angle": 0, + "content": "Here, the first term measures the difference between the results of the full-precision and quantized models, and the second term measures the bit lengths for the high- and medium-precision models. \\(\\lambda_{bit}\\) is a weighting parameter that balances between these two terms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.341, + 0.646, + 0.358 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.366, + 0.702, + 0.383 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Datasets and baseline quantization methods. To evaluate the effectiveness of QLIP across different diffusion models, we conducted experiments using two diffusion networks: BK-SDM-Tiny-2M [19] and Stable Diffusion v1.4 [39]. Our experiments involve two datasets, COCO2017 [27] and Conceptual Captions [46]. To train the T2Q module, 10k images were generated by Stable Diffusion v1.4 [39] using 10k text annotations in the COCO2017 training dataset as input text prompts. We generated another 1024 images using unused text annotations in the COCO2017 training dataset to train the Q2B module. We evaluated the effectiveness of our proposed QLIP by applying it to the baseline diffusion quantization methods, Q-diffusion [24] and PTQD [13]. In the implementation of PTQD, we set the noise correction parameter to \\( b_{med} \\), while noise correction was not applied if the estimated image quality by the T2Q module is below 0.3 or above 0.7. We trained the T2Q module for 3 epochs using the Adam optimizer [20] with a learning rate of 0.001 and the Q2B module for 5000 iterations using BRECQ [25], as adopted in Q-diffusion and PTQD, using RTX 3090. For performance evaluation, following the baseline quantization method [24], we generated 10k images using 10k text annotations in the COCO2017 validation dataset. Moreover, to evaluate the generalization ability of the quantized models, we used Conceptual Captions to generate additional 10k images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics. We report widely adopted metrics such as FID [15] and sFID [44] to evaluate image quality. In addition, we use CLIP Score [14] to assess the degree of matching between the generated images and their corresponding text prompts. To quantify computational efficiency, we measure BitOPs [22, 53], which account for the required operations in both the denoising model and addi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15430" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.09, + 0.907, + 0.206 + ], + "angle": 0, + "content": "
COCO2017
MethodBitoptionsFAB↓BitOPs (T)↓FID↓sFID↓CLIP Score↑
BK-SDM [19]W32A3232.0010.4623.7966.190.3069
Q-diffusion [24]W4A1616.001.0330.0273.250.3068
+QLIPW4A{8,16,32}12.140.8830.0173.240.3063
PTQD [13]W4A1616.001.0330.2777.180.3069
+QLIPW4A{8,16,32}12.140.8830.0273.260.3063
" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.207, + 0.907, + 0.319 + ], + "angle": 0, + "content": "
Conceptual Captions
MethodBitoptionsFAB↓BitOPs (T)↓FID↓sFID↓CLIP Score↑
BK-SDM [19]W32A3232.0010.4620.0047.800.2983
Q-diffusion [24]W4A1616.001.0324.6862.700.2965
+QLIPW4A{8,16,32}10.580.8224.7259.540.2964
PTQD [13]W4A1616.001.0324.6662.600.2965
+QLIPW4A{8,16,32}10.580.8224.7159.250.2966
" + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.327, + 0.856, + 0.342 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons of various quantization methods at a resolution of \\( {512} \\times {512} \\) using BK-SDM-Tiny-2M." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.366, + 0.482, + 0.596 + ], + "angle": 0, + "content": "
COCO2017
MethodFAB↓FID↓sFID↓CLIP Score↑
Stable Diffusion [39]32.0022.2365.110.3174
Q-diffusion [24]8.0023.4066.570.3126
+QLIP7.8621.6164.320.3120
PTQD [13]8.0022.7568.630.3126
+QLIP7.8621.3565.810.3120
Conceptual Captions
MethodFAB↓FID↓sFID↓CLIP Score↑
Stable Diffusion [39]32.0015.6745.610.3078
Q-diffusion [24]8.0020.7448.330.2984
+QLIP7.7719.8947.630.2979
PTQD [13]8.0025.4055.010.2972
+QLIP7.7720.4248.830.2975
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.605, + 0.483, + 0.661 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons of various quantization methods at a resolution of \\(512\\times 512\\) using Stable Diffusion v1.4. For the bit precision options, W4A8 is used for Q-diffusion and PTQD, and W4A{6,8,10} are used for QLIP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.78 + ], + "angle": 0, + "content": "tional modules for QLIP implementation, and feature average bit-width (FAB) [17, 18] for the denoising model. In particular, BitOPs are measured as \\( MACs \\cdot \\frac{b_w}{32} \\cdot \\frac{b_a}{32} \\), where \\( MACs \\) denotes multiply-accumulate computations, and \\( b_w \\) and \\( b_a \\) represent the bitwidths of weights and activations, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.904 + ], + "angle": 0, + "content": "Hyperparameter settings. We conducted experiments on various hyperparameter settings for the COCO2017 validation dataset using Q-diffusion as the baseline in Stable Diffusion v1.4. First, Figure 4(a) presents the results obtained with various values of the weight term \\(\\lambda_{bit}\\) in Equation (9), where \\(\\lambda_{bit} = 1\\) showed an effective balance between FID and FAB. Next, Figure 4(b) shows the results for different \\(M\\) values in \\(\\mathbf{u}_j^{Mt}, j\\in \\{m,h\\}\\) when the maximum timestep" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.365, + 0.91, + 0.448 + ], + "angle": 0, + "content": "
MethodBitoptionsRuntime (s)↓FAB↓FID↓
BK-SDM [19]W32A326.5032.0020.00
Q-diffusion [24]W4A84.538.0028.32
Q-diffusion [24]W4A165.5816.0024.68
+QLIPW4A{8,16,32}4.8510.5824.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.457, + 0.907, + 0.487 + ], + "angle": 0, + "content": "Table 3. Runtime comparisons, where W4A{8,16,32} is used for QLIP. Measured on the Conceptual Captions dataset." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.499, + 0.707, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.613, + 0.637, + 0.623 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.5, + 0.905, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.613, + 0.834, + 0.623 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.635, + 0.907, + 0.663 + ], + "angle": 0, + "content": "Figure 4. Results on Q-diffusion using different hyperparameter settings of QLIP (W4A{6,8,10}): (a) \\(\\lambda_{bit}\\) and (b) \\(M\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.907, + 0.718 + ], + "angle": 0, + "content": "\\(T\\) was 1000, where we obtained the best scores with \\(M = 200\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.704, + 0.743 + ], + "angle": 0, + "content": "4.2. Quantitative results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.871 + ], + "angle": 0, + "content": "Comparison with existing quantization methods. To demonstrate the effectiveness and applicability of QLIP, we tested QLIP on two diffusion model quantization methods: Q-diffusion [24] and PTQD [13]. Tables 1 and 2 provide comparison results for the COCO2017 [27] and Conceptual Captions [46] datasets using lightweight diffusion model BK-SDM-Tiny-2M [19] and Stable Diffusion v1.4 [39], respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, for the COCO2017 dataset, QLIP effectively reduced computational complexity of the base" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15431" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.089, + 0.86, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.298, + 0.907, + 0.314 + ], + "angle": 0, + "content": "Figure 5. Examples of variations in FAB by QLIP for the texts with different levels of richness and detail, along with the generated images." + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.326, + 0.861, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.56, + 0.908, + 0.631 + ], + "angle": 0, + "content": "Figure 6. Examples of the bit selection results and generated images. When the text prompt is specifically describing the image to be generated, the predicted image quality is high, and high and medium bits are assigned for most layers and time steps, as shown on the top left. On the other hand, when the text prompt has less detailed information, the predicted image quality is low, and low bits are assigned for many layers and time steps, as shown on the bottom left. On the right side, the generated images using Q-diffusion with and without applying QLIP are shown, demonstrating the effectiveness of QLIP in saving bits while improving the quality of generated images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.656, + 0.482, + 0.852 + ], + "angle": 0, + "content": "line diffusion quantization methods. For example, Q-diffusion+QLIP (W4A{8,16,32}) achieved a reduction in FAB by 3.86 while maintaining similar FID and sFID, compared to Q-diffusion (W4A16). Furthermore, the use of QLIP consistently led to improvements when applied to PTQD, as demonstrated in Table 1. For example, PTQD+QLIP (W4A{8,16,32}) enhanced FID and sFID by 0.25 and 3.92, respectively, while also reducing FAB and BitOPs. The results on Conceptual Captions followed trends similar to those observed in COCO2017. For example, Q-diffusion+QLIP (W4A{8,16,32}) demonstrated reduced computational complexity compared to Q-diffusion on Conceptual Captions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Table 2 presents the results of applying lower-bit precision combinations to Stable Diffusion v1.4. Here, PTQD+QLIP (W4A{6,8,10}) achieved a 4.98 improvement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.793 + ], + "angle": 0, + "content": "in FID compared to PTQD (W4A8), while maintaining similar FAB using bit combinations efficiently. Although our proposed QLIP drives the quantization methods to reduce computational complexity and improve image quality, the CLIP Scores remained similar to those of the baseline quantization methods, indicating that the fitness to the text prompts is not significantly influenced by QLIP. Additional results on larger diffusion models can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Runtime measurement. Runtime was measured with Cutlass [34], including the T2Q and Q2B modules, as shown in Table 3. The results show that runtime was reduced to a level similar to W4A8 while maintaining an FID score comparable to W4A16. Measurements were conducted on an RTX 3090. Additional analysis on memory requirements can be found in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15432" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.107 + ], + "angle": 0, + "content": "4.3. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.278 + ], + "angle": 0, + "content": "Figure 5 visualizes the changes in FAB by the richness and detail level of the text, along with the generated images. Here, ChatGPT-4o was used to create texts of varying complexity that describe a given keyword. Specifically, the prompt used was \"Create four sentences that describe how 'keyword' looks like, gradually getting richer and longer.\" As can be seen, QLIP assigned higher bits as the text describes the keyword more specifically and in greater detail, demonstrating that it utilizes information from the text to determine bit precisions for effective synthesis of vivid details and textures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.484, + 0.537 + ], + "angle": 0, + "content": "Figure 6 shows several results comparing Q-diffusion with and without QLIP when applied to Stable Diffusion v1.4. As can be seen, our QLIP generates more detailed and natural images compared to PTQD by using various bit combinations. Furthermore, Figure 6 visualizes the bit precision maps determined by QLIP. When the T2Q module predicts high image quality from input text prompts, QLIP allocates more bits to enable rendering of vivid and accurate details from rich text prompts. Otherwise, QLIP employs lower bits for layers that are less sensitive to quantization, based on the text prompt and the time step. For example, when given less specific text prompts, the correlation between the generated image and the text prompt is relatively less critical, allowing for low-bit quantization in the cross-attention blocks. These results highlight the effectiveness and significance of dynamically allocating bit precisions according to text prompts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.546, + 0.253, + 0.561 + ], + "angle": 0, + "content": "4.4. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Ablation studies on the T2Q module. We report the results on the COCO2017 validation dataset using Stable Diffusion v1.4 with Q-diffusion as the baseline quantization method. Table 4 presents the results of the T2Q modules trained using various perceptual image quality metrics, including Realism score [21], CLIP-IQA [54], and GIQA [11]. The performance of the T2Q module was evaluated using Spearman's rank-order correlation coefficient (SROCC) [49] and Pearson's linear correlation coefficient (PLCC) [36] between the actual quality measures and the predictions of the T2Q module. QLIP with GIQA showed the most substantial improvements in FAB and FID along with the highest SROCC and PLCC scores." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Ablation studies on the Q2B module. We conducted experiments on several variants of the Q2B module using Q-diffusion and Stable Diffusion v1.4 on the COCO2017 validation dataset. As shown in Table 5, when using either \\(\\mathbf{p}_q\\) alone or in combination with \\(\\mathbf{p}_h^t\\), effective reduction in FAB was achieved at the expense of increase in FID. When using \\(\\mathbf{p}_q\\) and \\(\\mathbf{p}_m^t\\), FID decreased while FAB increased. Finally, when \\(\\mathbf{p}_q\\), \\(\\mathbf{p}_h^t\\), and \\(\\mathbf{p}_m^t\\) were all involved, balanced results were obtained in two metrics. Implementation details of" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.187 + ], + "angle": 0, + "content": "
MethodSROCC↑PLCC↑FAB↓FID↓
Q-diffusion [24]--8.0023.40
+QLIP
w/ Realism score0.51320.50228.1022.18
w/ CLIP-IQA0.71270.70828.5421.81
w/ GIQA0.80470.81087.8621.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.196, + 0.905, + 0.237 + ], + "angle": 0, + "content": "Table 4. Ablation studies on the T2Q module: For the bit precision options, W4A8 in Q-diffusion and +QLIP W4A{6,8,10} were used." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.256, + 0.903, + 0.356 + ], + "angle": 0, + "content": "
MethodBitoptionsFAB↓FID↓
Q-diffusion [24]W4A88.0023.40
+ pqW4A{6,10}7.5726.91
+ pq + ptW4A{6,8,10}6.7329.37
+ pq + pmtW4A{6,8,10}8.6021.96
+ pq + pmt + pth (QLIP)W4A{6,8,10}7.8621.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.572, + 0.365, + 0.847, + 0.379 + ], + "angle": 0, + "content": "Table 5. Ablation studies on the Q2B module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.905, + 0.441 + ], + "angle": 0, + "content": "the ablation studies on the Q2B module can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.461, + 0.634, + 0.477 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.489, + 0.907, + 0.715 + ], + "angle": 0, + "content": "The proposed quantization method for diffusion models, called QLIP, is the first method that uses text prompts for quantization in diffusion models. From our observation that the bit precision for quantization can be effectively determined based on the quality of the generated images, we designed the T2Q module that predicts image quality from input text prompts. We then introduced the Q2B module that allocates appropriate bit precision for each time step for different layers. By integrating the T2Q and Q2B modules into existing diffusion model quantization methods, we demonstrated the effectiveness of the proposed QLIP. QLIP presents a promising solution for quantizing diffusion models due to its ability to improve the quality of generated images using text prompts while reducing computational costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Limitations and future directions. Similar to other generative models, QLIP also has the potential to generate counterfeit images for malicious purposes, potentially undermining trust in visual media. In a future study, we will extend QLIP to consider other input conditions, such as images or segmentation maps, in addition to text prompts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.829, + 0.673, + 0.845 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2022-NR070077)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15433" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. eDiff-I: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.253 + ], + "angle": 0, + "content": "[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: An analytic estimate of the optimal reverse variance in diffusion probabilistic models. In Proceedings of the International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.45 + ], + "angle": 0, + "content": "[5] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. PACT: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[6] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.563 + ], + "angle": 0, + "content": "[7] Huanpeng Chu, Wei Wu, Chengjie Zang, and Kun Yuan. QNCD: Quantization noise correction for diffusion models. arXiv preprint arXiv:2403.19140, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[8] Xiangxiang Chu, Shun Lu, Xudong Li, and Bo Zhang. Mix-Path: A unified approach for one-shot neural architecture search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5972–5981, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[9] Zhen Dong, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. HAWQ: Hessian aware quantization of neural networks with mixed-precision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 293–302, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proceedings of the Advances in Neural Information Processing Systems, pages 2672–2680, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[11] Shuyang Gu, Jianmin Bao, Dong Chen, and Fang Wen. GIQA: Generated image quality assessment. In Proceedings of the European Conference on Computer Vision, pages 369–385, 2020. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[12] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Proceedings of the Advances in Neural Information Processing Systems, pages 1135-1143, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[13] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. PTQD: Accurate post-training quantization" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "for diffusion models. In Proceedings of the Advances in Neural Information Processing Systems, pages 13237-13249, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[14] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.209, + 0.905, + 0.278 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the Advances in Neural Information Processing Systems, pages 6629-6640, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.905, + 0.335 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.905, + 0.395 + ], + "angle": 0, + "content": "[17] Cheeun Hong and Kyoung Mu Lee. AdaBM: On-the-fly adaptive bit mapping for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2641-2650, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.398, + 0.907, + 0.466 + ], + "angle": 0, + "content": "[18] Cheeun Hong, Sungyong Baik, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. CADyQ: Content-aware dynamic quantization for image super-resolution. In Proceedings of the European Conference on Computer Vision, pages 367-383, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[19] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. BK-SDM: A lightweight, fast, and cheap version of stable diffusion. In Proceedings of the European Conference on Computer Vision, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.905, + 0.569 + ], + "angle": 0, + "content": "[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.572, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[21] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Proceedings of the Advances in Neural Information Processing Systems, pages 3927-3936, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.643, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[22] Hongjae Lee, Jun-Sang Yoo, and Seung-Won Jung. RefQSR: Reference-based quantization for image super-resolution networks. IEEE Transactions on Image Processing, 33: 2823-2834, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[23] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17413-17424, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[24] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. BRECQ: Pushing the limit of post-training quantization by block reconstruction. In Proceedings of the International Conference on Learning Representations, 2021. 1, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15434" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[26] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. SnapFusion: Text-to-image diffusion model on mobile devices within two seconds. In Proceedings of the Advances in Neural Information Processing Systems, pages 20662-20678, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755, 2014. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[28] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In Proceedings of the European Conference on Computer Vision, pages 19–34, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[29] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In Proceedings of the International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[30] Zhenhua Liu, Yunhe Wang, Kai Han, Siwei Ma, and Wen Gao. Instance-aware dynamic neural network quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12434-12443, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[31] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Proceedings of the Advances in Neural Information Processing Systems, pages 5775-5787, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[32] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Proceedings of the International Conference on Machine Learning, pages 8162-8171, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[34] NVIDIA. CUTLASS, 2025. https://github.com/NVIDIA/cutlass.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[35] Nilesh Prasad Pandey, Marios Fournarakis, Chirag Patel, and Markus Nagel. Softmax bias correction for quantized generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshop, pages 1453-1458, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[36] Karl Pearson. Notes on regression and inheritance in the case of two parents. Proceedings of the Royal Society of London, 58:240-242, 1895. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[37] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *SDXL: Improving latent diffusion models for high-resolution image synthesis*. In *Proceedings of the International Conference on Learning Representations*, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10674-10685, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[41] Hyogon Ryu, Seohyun Lim, and Hyunjung Shim. Memory-efficient personalization using quantized diffusion model. In Proceedings of the European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In Proceedings of the Advances in Neural Information Processing Systems, pages 36479–36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[43] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Proceedings of the International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[44] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In Proceedings of the Advances in Neural Information Processing Systems, pages 2234-2242, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[45] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1972-1981, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[46] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 2556-2565, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[47] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. In Proceedings of the Advances in Neural Information Processing Systems, pages 48686-48698, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Proceedings of the International Conference on Learning Representations, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Charles Spearman. The proof and measurement of association between two things. The American Journal of Psychology, 15(1):72--101, 1961. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15435" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[50] Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. CHIP: Channel independence-based pruning for compact neural networks. In Proceedings of the Advances in Neural Information Processing Systems, pages 24604-24616, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[51] Siao Tang, Xin Wang, Hong Chen, Chaoyu Guan, Zewen Wu, Yansong Tang, and Wenwu Zhu. Post-training quantization with progressive calibration and activation relaxing for text-to-image diffusion models. In Proceedings of the European Conference on Computer Vision, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[52] Senmao Tian, Ming Lu, Jiaming Liu, Yandong Guo, Yurong Chen, and Shunli Zhang. CABM: Content-aware bit mapping for single image super-resolution network with large input. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1756-1765, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.318, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[53] Mart Van Baalen, Christos Louizos, Markus Nagel, Rana Ali Amjad, Ying Wang, Tijmen Blankevoort, and Max Welling. Bayesian bits: Unifying quantization and pruning. In Proceedings of the Advances in Neural Information Processing Systems, pages 5741-5752, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.389, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[54] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[55] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. HAQ: Hardware-aware automated quantization with mixed precision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8612-8620, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.516, + 0.482, + 0.584 + ], + "angle": 0, + "content": "[56] Chuanguang Yang, Zhulin An, Helong Zhou, Fuzhen Zhuang, Yongjun Xu, and Qian Zhang. Online knowledge distillation via mutual contrastive learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10212-10227, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.586, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.482, + 0.698 + ], + "angle": 0, + "content": "[58] Xuechen Zhang, Zheng Li, Samet Oymak, and Jiasi Chen. Text-to-3d generative ai on mobile devices: Measurements and optimizations. In Proceedings of the Workshop on Emerging Multimedia Systems, pages 8–14, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15436" + } + ] +] \ No newline at end of file diff --git a/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/885bbebf-0678-4829-a8d8-7dbca73ecb4a_origin.pdf b/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/885bbebf-0678-4829-a8d8-7dbca73ecb4a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..483c7e80e6012ea7c37fc832f021a5c6c938f992 --- /dev/null +++ b/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/885bbebf-0678-4829-a8d8-7dbca73ecb4a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b9db8c01fa856a135a861a570b8a0530db6edf74dbf1bc0829638e87fbeb0dae +size 17549482 diff --git a/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/full.md b/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..015177fdac1835bdfa1e55d1e705fa57cb799f6a --- /dev/null +++ b/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/full.md @@ -0,0 +1,343 @@ +# Text Embedding Knows How to Quantize Text-Guided Diffusion Models + +Hongjae Lee Myungjun Son Dongjea Kang Seung-Won Jung\* Korea University + +{jimmy9704,sonbill,kedjk11,swjung83}@korea.ac.kr + +# Abstract + +Despite the success of diffusion models in image generation tasks such as text-to-image, the enormous computational complexity of diffusion models limits their use in resource-constrained environments. To address this, network quantization has emerged as a promising solution for designing efficient diffusion models. However, existing diffusion model quantization methods do not consider input conditions, such as text prompts, as an essential source of information for quantization. In this paper, we propose a novel quantization method dubbed Quantization of Language-to-Image diffusion models using text Prompts (QLIP). QLIP leverages text prompts to guide the selection of bit precision for every layer at each time step. In addition, QLIP can be seamlessly integrated into existing quantization methods to enhance quantization efficiency. Our extensive experiments demonstrate the effectiveness of QLIP in reducing computational complexity and improving the quality of the generated images across various datasets. The source code is available at our project page https://github.com/jimmy9704/QLIP. + +# 1. Introduction + +Diffusion models have achieved remarkable success in various generative tasks, particularly in text-to-image generation [39, 40, 57]. However, the huge computational complexity of diffusion models restricts their practical applicability, especially in resource-constrained environments. Specifically, diffusion models typically require hundreds of denoising steps to produce high-quality samples, making them considerably slower than generative adversarial networks (GANs) [10]. To this end, many studies introduced advanced training-free samplers to reduce the number of denoising iterations [2, 29, 31, 48]. However, the decrease in iterations alone is insufficient since the denoising models used in each iteration possess billions of parameters [4, 37, 42], leading to increased computational com + +plexity and high memory consumption. + +Besides other approaches toward light-weight network designs, such as network architecture search [8, 28], network pruning [12, 50], and knowledge distillation [23, 56], network quantization has received particular attention due to its simplicity, the increasing demand for low-precision operations, and related hardware support [9, 55, 58]. Although various network quantization methods have been proposed for convolutional neural networks (CNNs), the direct use of such methods for diffusion models has shown significant performance degradation [45]. As tailored solutions for diffusion models, several methods have been proposed considering the time step of diffusion models for quantization. For example, PTQ4DM [45] performed 8-bit quantization by constructing a calibration dataset based on the time step. Q-Diffusion [24] introduced an advanced method based on BRECQ [25] and performed the performance analysis of quantization methods for text-guided diffusion models. In addition, TDQ [47] performed quantization with varying activation scaling based on the diffusion time step. However, these methods do not use the text prompt provided to the diffusion model as valuable information for quantization, leaving room for further improvement. + +Meanwhile, recent studies highlight the advantages of dynamic bit precision quantization considering input conditions [30, 52]. For example, in the case of image superresolution, CADyQ [18] and AdaBM [17] performed dynamic quantization according to the complexity of the images and the importance of layers. RefQSR [22] clustered the image patches and assigned high-bit and low-bit precisions to the reference and query patches, respectively. These dynamic quantization methods allow for the cost-effective assignment of bit precision, enhancing quantization efficiency. However, such input-adaptive quantization methods have not yet been explored in diffusion models. + +In this paper, we propose a novel quantization method, dubbed Quantization of Language-to-Image diffusion models using text Prompts (QLIP), which determines the bit precision of diffusion models according to the text prompt. Figure 1 shows our observation: The quality of the gener + +Figure 1. Examples of the images generated by Stable Diffusion v1.4 using the text prompts given below images. When the text prompt details the shape, texture, and properties for the target object, the generated images have rich details and textures, as shown on the right; whereas the generated images are not very different from the images generated using only the keyword when the text prompt is less specific to the target object, as shown on the left. +![](images/4eb91b1bf14e292483921a31f8f9a8f863e441e39dd39c885c7b363aaf81194f.jpg) +Pikachu, the symbol and eternal mascot of Pokémon, is the most famous of over 1,000 Pokémon and is a Pokémon with an electric attribute. It appeared in various video games, TV shows, movies, and merchandise since its introduction in the late 1990s. + +![](images/ec0073a3ac985fc020281a1356d7493a41b8f94db770caf202fbbe6f261ae37a.jpg) + +![](images/bb842b45fe9d55c3ed0f3fbb581de141ca28649f5f0ab8b643855dd452716366.jpg) +Pikachu + +![](images/13febf055745aa57c7ff2933467dfa31a53dd9f95741acbbdc98aed283bc802d.jpg) +Pikachu is a small, yellow, mouse-like creature with large, black-tipped ears that resemble lightning bolts. It has rosy cheeks, a small mouth, and large, expressive eyes. Pikachu's body is covered in soft, yellow fur, and it has a long, lightning bolt-shaped tail. + +![](images/84c03b2ff81d7d3d6496d316cfdd9546f634ee4c1abc10364779983f9d7bca2f.jpg) + +![](images/8e769f7a176ca997ed6c3bcd792144e82df40fa92ec0790e889b314586e7ef37.jpg) + +ated images increases as the text prompt has more specific information about the image to be generated. This suggests that high-bit precision operations are necessary to generate images with vivid details and textures corresponding to detailed text prompts; whereas, low-bit precision operations can be sufficient to generate images corresponding to text prompts that lack specific guidance on details and textures. + +Motivated by this observation, we quantized the diffusion model [39] using Q-diffusion [24] with high-bit and low-bit settings and compared the quality of the generated images with the full-precision model. As shown in Figure 2, the performance drop of the low-bit model is noticeable when the quality of the generated images is high for the full-precision model. In contrast, when the quality is low for the full-precision model, the high-bit and low-bit models perform similarly to the full-precision model. These results suggest that diffusion models can be efficiently quantized without sacrificing performance if the quality of the generated images can be first predicted and then used to determine bit precision. Accordingly, our proposed QLIP predicts the image quality from the text prompt and determines the quantization bit precision for diffusion models. + +Furthermore, previous studies [13, 24, 45, 47] revealed that the activation distributions of diffusion models vary across layers over time steps. Inspired by these studies, QLIP uses learnable parameters to determine the bit precision for each layer and for each timestep. Our contributions can be summarized as follows: + +- To the best of our knowledge, QLIP is the first approach that explicitly utilizes text prompts for dynamic quantization in diffusion models. +- We propose leveraging text prompts to predict the image quality, which is used to guide the selection of bit precision for every layer in different denoising time steps. +- Our proposed QLIP can be integrated into existing diffusion quantization methods to enhance quantization efficiency. + +# 2. Related work + +# 2.1. Computationally efficient diffusion models + +Diffusion models generate high-quality images by iteratively removing noise from random noise. Despite their capability to generate high-quality images, diffusion models encounter challenges in real-world applications due to the time-consuming iterative denoising process. Thus, many efforts have been made to make the diffusion process computationally efficient, including reducing the number of iterative diffusion processes via a non-Markovian process [48], adjusting the variance scheduling [33], and employing knowledge distillation [32, 43]. + +Meanwhile, since the computational complexity of diffusion model architectures also limits their practical applicability, especially in resource-constrained environments, research efforts have been made to lighten denoising models. For example, SnapFusion [26] introduced an efficient U-Net architecture based on Stable Diffusion [39], while BKSDM [19] further lightened the architecture through knowledge distillation. + +# 2.2. Quantization of diffusion models + +Quantization involves mapping 32-bit floating-point values of weights and activations to lower-bit values. Numerous studies have explored quantization to compress and accelerate neural networks [5, 18], which can be classified into two approaches: quantization-aware training (QAT) and post-training quantization (PTQ). Recent studies [7, 35, 41, 51] on quantizing diffusion models have focused mainly on PTQ due to the inability to use training data for privacy and commercial reasons, as well as the high computational cost of QAT. Specifically, PTQ4DM [45] and Q-diffusion [24] performed quantization of diffusion models by generating a time step-aware calibration set that consists of data sampled across various time steps using the full-precision denoising model. PTQD [13] separated the + +![](images/6ff6701630ba33acf936ede1f816cfe9163784d4bf9677452a8cbfa01de6b369.jpg) +Figure 2. Analysis of the effect of quantization and the quality of generated images, where the image quality is measured using the GIQA score [11]. On the left side, the x-axis represents the quality of images generated by the full-precision model, and the y-axis represents the quality difference between the generated images using the full-precision and quantized models. The images are generated by Stable Diffusion v1.4 [39] using 5000 captions from the COCO 2017 dataset [27] with a size of $512 \times 512$ . The high-bit (W4A10) and low-bit (W4A6) quantized models are obtained by applying Q-diffusion [24]. On the right side, examples of images generated using models that are unquantized or quantized w/wo QLIP are shown for two different image quality levels. + +![](images/80c76b4ebb90825bf8842e67d5e322bd47ed58ffab25c4c264224097e2bc7cd7.jpg) + +![](images/2f3acb12f551e68b939b3cb4065830534c18ed65279ab62d59565e8ad6c8b99b.jpg) + +![](images/9fe47efa25c24fb108e881a9e4640205b270f3ed8d3eaad9928b690b7a828026.jpg) + +![](images/c1d25611b96184d123cbae5f1e8557c49dd2aa2096161f03caad283c5414bc72.jpg) + +![](images/a36df503c5555d2cebe04756dd2aec3253cc66cec91e1bb83b8c9178f52becc7.jpg) + +![](images/4cb782038d7fdd814494d66a7f43d49c5a862caf53c38483fc0545de08a3bc84.jpg) + +quantization noise from the diffusion noise to reduce quantization errors. Additionally, PTQD introduced a time step-based mixed-precision method but did not employ varying bit precision for each layer. Furthermore, TDQ [47] performed quantization by adaptively scaling the activations for different diffusion time steps and layers, but assigned the same bit precision for all time steps. Although these diffusion model quantization methods have shown progress in reducing the complexity of noise estimation models, they do not adjust quantization parameters based on input conditions such as text prompts. + +# 2.3. Input-adaptive dynamic quantization + +Recent advances in hardware have enabled mixed-precision operations, allowing various layers to use different bit precisions to improve performance and computational efficiency [9, 55]. Many methods have been proposed to determine the bit precision of each layer of CNNs [9, 30, 55]. For example, several input-adaptive dynamic quantization methods [17, 18, 22, 52] have been introduced for super-resolution, which adjust bit precisions based on the complexity of image patches and the sensitivity to quantization. However, diffusion models have not yet explored such input-adaptive quantization methods. Therefore, this paper proposes a novel quantization method for text-guided diffusion models that applies adaptive quantization using input conditions. + +# 3. Method + +# 3.1. Preliminaries + +Diffusion model. The reverse process of a diffusion model gradually removes noise from random data $\mathbf{x}_t$ to + +produce a high-quality image $\mathbf{x}_0$ , which can be expressed as $p_{\theta}(\mathbf{x}_{t - 1}|\mathbf{x}_t) = \mathcal{N}(\mathbf{x}_{t - 1};\mu_\theta (\mathbf{x}_t,t),\sigma_t^2\mathbf{I})$ . The denoising process at time step $t$ using the noise estimation model $\epsilon_{\theta}$ can then be represented as follows: + +$$ +\mu_ {\theta} (\mathbf {x} _ {t}, t) = \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathbf {x} _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\theta} (\mathbf {x} _ {t}, t)\right), \quad (1) +$$ + +where $\alpha_{t},\bar{\alpha}_{t}$ , and $\beta_{t}$ are the parameters that control the strength of the Gaussian noise in each step, and $\sigma_{t}$ is a variance schedule [16]. + +Diffusion model quantization. Diffusion model quantization reduces the complexity of a noise estimation model that requires hundreds of iterative sampling processes. The denoising process at time step $t$ using the quantized noise estimation model $\hat{\epsilon}_{\theta}$ and the quantized data $\dot{\mathbf{x}}_t$ can be expressed as follows: + +$$ +\begin{array}{l} \hat {\mu} _ {\theta} \left(\hat {\mathbf {x}} _ {t}, t\right) = \frac {1}{\sqrt {\alpha_ {t}}} \left(\hat {\mathbf {x}} _ {t} - \frac {\beta_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \hat {\epsilon} _ {\theta} \left(\hat {\mathbf {x}} _ {t}, t\right)\right), \tag {2} \\ w h e r e \hat {\epsilon} _ {\theta} = Q _ {\mathbf {b} ^ {t}} (\epsilon_ {\theta}). \\ \end{array} +$$ + +Here, $\mathbf{b}^t\in \mathbb{R}^K$ represents a set of quantization bit precisions for $K$ layers at time step $t$ , and $Q_{\mathbf{b}^t}$ is a function that performs layer-wise and time step-wise quantization using $\mathbf{b}^t$ . To our knowledge, all previous diffusion quantization methods set the same bit precision to all layers, i.e., $\mathbf{b}^t (1) = \mathbf{b}^t (2) = \dots = \mathbf{b}^t (K)$ , and only a few methods [13, 51] performed time step-wise dynamic quantization, i.e., $\mathbf{b}^{t_i}\neq \mathbf{b}^{t_j}$ for some time steps $t_i$ and $t_j$ , $t_i\neq t_j$ . However, such methods do not adapt the bit precision for each layer. Our proposed QLIP determines $\mathbf{b}^t$ by training so that the quantization error can be effectively reduced without sacrificing the performance of diffusion models. + +![](images/b4b5aca18b19d9b948d6eda78d52618574b5054bc506997ebc73f8da3e0518e8.jpg) +Figure 3. Overview of QLIP that consists of two key modules: T2Q module that predicts image quality from the input text prompt and Q2B module that determines bit precisions for all denoising layers and time steps. Only the Q2B module is trained during QLIP training. See Sections 3.3 and 3.4 for the details about the pre-training of the T2Q module and diffusion models. + +# 3.2. Framework overview + +Our observations shown in Figure 2 indicate that bit precisions for quantization can be determined cost-effectively given the quality of the generated image. However, this image quality remains unknown during the inference stage, motivating us to predict it using an available input, i.e., text prompts. Specifically, if text prompts have rich and detailed information, the expected quality of the generated image is high, and vice versa. To this end, we introduce a text-to-quality (T2Q) module $\phi$ that predicts image quality $q$ from the embedding vector $\mathbf{z}$ , i.e., $q = \phi(\mathbf{z})$ , where $\mathbf{z}$ is obtained by CLIP-Text [38] using the input text prompt. + +Next, as we discussed, high-bit quantization is required to generate images with vivid details and textures corresponding to detailed text prompts; while low-bit quantization is sufficient to generate images using less specific text prompts. Therefore, we design a quality-to-bit (Q2B) module $\psi$ that maps image quality $q$ to appropriate bit precision, i.e., $\mathbf{b}^t = \psi(q)$ . Note that the Q2B module determines bit precisions for each time step for different layers since the activation distributions of diffusion models vary across layers over time steps. + +In summary, the proposed QLIP obtains a quantized diffusion model $\hat{\epsilon}_{\theta} = Q_{\psi (\phi (\mathbf{z}))}(\epsilon_{\theta})$ by learning the parameters of $\psi$ and $\phi$ . The overview of the proposed QLIP is shown in Figure 3. The following subsections detail the T2Q and Q2B modules and their training procedure and loss functions. + +# 3.3. Text-to-Quality estimation + +The T2Q module, denoted as $\phi$ , aims to predict the image quality from the input text prompt. $\phi$ consists of three linear layers and outputs an image quality score $q$ as $q = \phi(\mathbf{z})$ , where $\mathbf{z} \in \mathbb{R}^{C_{clip}}$ represents the CLIP text embedding, and $C_{clip}$ denotes its channel dimension. For the training of the T2Q module, we collect a dataset $\mathcal{T}_{T2Q} = \{\mathbf{z}^i, \mathbf{x}_0^i\}_{i=1}^{N_{T2Q}}$ , where $\mathbf{z}^i$ and $\mathbf{x}_0^i$ are the $i$ -th sample of the CLIP text embedding and its corresponding generated + +image obtained using the full-precision diffusion model, respectively, and $N_{T2Q}$ is the number of images in $\mathcal{T}_{T2Q}$ . Let $\bar{q}^i$ represent the (pseudo) ground-truth image quality of $\mathbf{x}_0^i$ , where we used GIQA [11] as the default quality measure, with an additional normalization to make the range of quality values [0, 1]. See Section 4.4 for the performance analysis with different quality measures. The T2Q module is trained using the loss $L_{t2q}$ : + +$$ +L _ {t 2 q} = \frac {1}{N _ {T 2 Q}} \sum_ {i = 1} ^ {N _ {T 2 Q}} \left(\bar {q} ^ {i} - \phi \left(\mathbf {z} ^ {i}\right)\right) ^ {2}. \tag {3} +$$ + +Once the T2Q module is trained, it is frozen during the training of the Q2B module and served as a quality predictor to help determine quantization bit precisions. + +# 3.4. Quality-to-Bit estimation + +In text-guided diffusion models, the richer and more detailed the information in the text prompt is, the higher the image quality tends to be, requiring the use of higher bits. Therefore, the image quality $q$ predicted by the T2Q module can serve as a direct hint for bit selection. To this end, the proposed Q2B module obtains a probability vector $\mathbf{p}_q \in \mathbb{R}^K$ for bit selection as follows: + +$$ +\mathbf {p} _ {q} = \sigma \left((q - 0. 5) \mathbf {s} + \mathbf {o}\right), \tag {4} +$$ + +where $\mathbf{s} \in \mathbb{R}^K$ and $\mathbf{o} \in \mathbb{R}^K$ represent sets of learnable parameters, and $\sigma$ denotes the sigmoid function. Furthermore, to adjust the bit precision for each time step, additional learnable parameters $\mathbf{u}_m^t \in \mathbb{R}^K$ and $\mathbf{u}_h^t \in \mathbb{R}^K$ are introduced to obtain probability vectors $\mathbf{p}_m^t \in \mathbb{R}^K$ and $\mathbf{p}_h^t \in \mathbb{R}^K$ as + +$$ +\mathbf {p} _ {j} ^ {t} = \sigma \left(\mathbf {u} _ {j} ^ {t}\right), \quad j \in \{m, h \}, t = 1, 2, \dots , T, \tag {5} +$$ + +where $T$ is the total number of denoising time steps. + +These probability vectors, $\mathbf{p}_q$ , $\mathbf{p}_m^t$ , and $\mathbf{p}_h^t$ , are used to determine the bit precision for each layer and each time + +step. Specifically, we define a set of supported bit precis +sions as $\mathcal{B} = \{b_{low},b_{med},b_{high}\}$ . Then, the probabilities of +selecting $b_{low},b_{med}$ , and $b_{high}$ are obtained as + +$$ +\mathbf {p} _ {b _ {l o w}} ^ {t} = \left(1 - \mathbf {p} _ {q}\right) \odot \left(1 - \mathbf {p} _ {m} ^ {t}\right), +$$ + +$$ +\mathbf {p} _ {b _ {m e d}} ^ {t} = \left(1 - \mathbf {p} _ {q}\right) \odot \mathbf {p} _ {m} ^ {t} + \mathbf {p} _ {q} \odot \left(1 - \mathbf {p} _ {h} ^ {t}\right), \tag {6} +$$ + +$$ +\mathbf {p} _ {b _ {h i g h}} ^ {t} = \mathbf {p} _ {q} \odot \mathbf {p} _ {h} ^ {t}, +$$ + +where $\odot$ represents element-wise multiplication, $\mathbf{p}_{b_{low}}^{t} \in \mathbb{R}^{K}$ , $\mathbf{p}_{b_{med}}^{t} \in \mathbb{R}^{K}$ , and $\mathbf{p}_{b_{high}}^{t} \in \mathbb{R}^{K}$ are the probability vectors for bit selection. Finally, the bit precision for each layer at time step $t$ is determined as follows: + +$$ +\mathbf {b} ^ {t} (k) = \operatorname * {a r g m a x} _ {i \in \{b _ {\text {l o w}}, b _ {\text {m e d}}, b _ {\text {h i g h}} \}} \mathbf {p} _ {i} ^ {t} (k), k = 1, 2, \dots , K. \tag {7} +$$ + +However, the training of $\mathbf{u}_m^t$ and $\mathbf{u}_h^t$ for all $t$ poses a challenge as $T$ increases. In addition, the activation distributions of the diffusion models are known to be similar to each other for adjacent time steps [13, 47]. Consequently, we learn the parameters for every $M$ time step, i.e., $\mathbf{u}_j^{Mt}$ and use the same parameters for the intermediate time steps, i.e., $\mathbf{u}_j^{M(t + 1) - 1} = \mathbf{u}_j^{M(t + 1) - 2} = \dots = \mathbf{u}_j^{Mt}$ , $j\in \{m,h\}$ . Next, previous studies [1, 6] revealed that the initial stage of the reverse process involves generating the rough context of an image corresponding to the input text prompt; thus, the use of low bits in the initial time steps could decrease the correlation between the text prompt and the generated image. Therefore, we set all values in $\mathbf{p}_q$ to 1 for the first $m$ time steps in Equation (6) to enforce the selection of high bits in the initial stage of the reverse process. In summary, the Q2B module has $2K$ ( $\mathbf{s}$ and $\mathbf{o}$ ) and $2KT / M$ ( $\mathbf{u}_m^t$ and $\mathbf{u}_h^t$ ) trainable parameters. + +The bit selection in Equation (7) is non-differentiable, making it unable to backpropagate. Thus, we adopt a straight-through estimator [3] to render this process differentiable: + +$$ +\hat {\mathbf {a}} ^ {t} (k) = +$$ + +$$ +\left\{ \begin{array}{l l} Q _ {\mathbf {b} ^ {t} (k)} \left(\mathbf {a} ^ {t} (k)\right) & \text {f o r w a r d}, \\ \sum_ {i \in \left\{b _ {l o w}, b _ {m e d}, b _ {h i g h} \right\}} \mathbf {p} _ {i} ^ {t} (k) \cdot Q _ {i} \left(\mathbf {a} ^ {t} (k)\right) & \text {b a c k w a r d}. \end{array} \right. \tag {8} +$$ + +Here, $\mathbf{a}^t (k)$ and $\hat{\mathbf{a}}^t (k)$ represent the activations of the full-precision and quantized models at time step $t$ of the $k$ -th layer. In other words, during backpropagation, the activations quantized by all bit precisions in $\mathcal{B}$ are weighted by their corresponding estimated probabilities. Since diffusion models are less sensitive to the quantization of the weights [13, 24], the weights are quantized using a fixed precision, which is empirically chosen as 4 bits for all layers and time steps. + +Given the full precision diffusion model, we obtain its quantized weights by applying the baseline diffusion quantization methods [13, 24]. The quantization parameters, including the maximum and minimum values for clipping, are + +also predetermined for all bit precisions in $\mathcal{B}$ . Then, while freezing all the other modules in Figure 3, only the Q2B module is trained using a small calibration set. The Q2B module is trained at randomly sampled time steps $t$ using the loss function $L_{QLIP}$ : + +$$ +\begin{array}{l} L _ {Q L I P} = \left(\epsilon_ {\theta} (\mathbf {x} _ {t}, t) - \hat {\epsilon} _ {\theta} (\hat {\mathbf {x}} _ {t}, t)\right) ^ {2} \\ + \lambda_ {b i t} \left(b _ {h i g h} \cdot \sum_ {k = 1} ^ {K} \mathbf {p} _ {b _ {h i g h}} ^ {t} (k) + b _ {m e d} \cdot \sum_ {k = 1} ^ {K} \mathbf {p} _ {b _ {m e d}} ^ {t} (k)\right). \tag {9} \\ \end{array} +$$ + +Here, the first term measures the difference between the results of the full-precision and quantized models, and the second term measures the bit lengths for the high- and medium-precision models. $\lambda_{bit}$ is a weighting parameter that balances between these two terms. + +# 4. Experiments + +# 4.1. Experimental setup + +Datasets and baseline quantization methods. To evaluate the effectiveness of QLIP across different diffusion models, we conducted experiments using two diffusion networks: BK-SDM-Tiny-2M [19] and Stable Diffusion v1.4 [39]. Our experiments involve two datasets, COCO2017 [27] and Conceptual Captions [46]. To train the T2Q module, 10k images were generated by Stable Diffusion v1.4 [39] using 10k text annotations in the COCO2017 training dataset as input text prompts. We generated another 1024 images using unused text annotations in the COCO2017 training dataset to train the Q2B module. We evaluated the effectiveness of our proposed QLIP by applying it to the baseline diffusion quantization methods, Q-diffusion [24] and PTQD [13]. In the implementation of PTQD, we set the noise correction parameter to $b_{med}$ , while noise correction was not applied if the estimated image quality by the T2Q module is below 0.3 or above 0.7. We trained the T2Q module for 3 epochs using the Adam optimizer [20] with a learning rate of 0.001 and the Q2B module for 5000 iterations using BRECQ [25], as adopted in Q-diffusion and PTQD, using RTX 3090. For performance evaluation, following the baseline quantization method [24], we generated 10k images using 10k text annotations in the COCO2017 validation dataset. Moreover, to evaluate the generalization ability of the quantized models, we used Conceptual Captions to generate additional 10k images. + +Evaluation metrics. We report widely adopted metrics such as FID [15] and sFID [44] to evaluate image quality. In addition, we use CLIP Score [14] to assess the degree of matching between the generated images and their corresponding text prompts. To quantify computational efficiency, we measure BitOPs [22, 53], which account for the required operations in both the denoising model and addi + +
COCO2017
MethodBitoptionsFAB↓BitOPs (T)↓FID↓sFID↓CLIP Score↑
BK-SDM [19]W32A3232.0010.4623.7966.190.3069
Q-diffusion [24]W4A1616.001.0330.0273.250.3068
+QLIPW4A{8,16,32}12.140.8830.0173.240.3063
PTQD [13]W4A1616.001.0330.2777.180.3069
+QLIPW4A{8,16,32}12.140.8830.0273.260.3063
+ +
Conceptual Captions
MethodBitoptionsFAB↓BitOPs (T)↓FID↓sFID↓CLIP Score↑
BK-SDM [19]W32A3232.0010.4620.0047.800.2983
Q-diffusion [24]W4A1616.001.0324.6862.700.2965
+QLIPW4A{8,16,32}10.580.8224.7259.540.2964
PTQD [13]W4A1616.001.0324.6662.600.2965
+QLIPW4A{8,16,32}10.580.8224.7159.250.2966
+ +Table 1. Quantitative comparisons of various quantization methods at a resolution of ${512} \times {512}$ using BK-SDM-Tiny-2M. + +
COCO2017
MethodFAB↓FID↓sFID↓CLIP Score↑
Stable Diffusion [39]32.0022.2365.110.3174
Q-diffusion [24]8.0023.4066.570.3126
+QLIP7.8621.6164.320.3120
PTQD [13]8.0022.7568.630.3126
+QLIP7.8621.3565.810.3120
Conceptual Captions
MethodFAB↓FID↓sFID↓CLIP Score↑
Stable Diffusion [39]32.0015.6745.610.3078
Q-diffusion [24]8.0020.7448.330.2984
+QLIP7.7719.8947.630.2979
PTQD [13]8.0025.4055.010.2972
+QLIP7.7720.4248.830.2975
+ +tional modules for QLIP implementation, and feature average bit-width (FAB) [17, 18] for the denoising model. In particular, BitOPs are measured as $MACs \cdot \frac{b_w}{32} \cdot \frac{b_a}{32}$ , where $MACs$ denotes multiply-accumulate computations, and $b_w$ and $b_a$ represent the bitwidths of weights and activations, respectively. + +Hyperparameter settings. We conducted experiments on various hyperparameter settings for the COCO2017 validation dataset using Q-diffusion as the baseline in Stable Diffusion v1.4. First, Figure 4(a) presents the results obtained with various values of the weight term $\lambda_{bit}$ in Equation (9), where $\lambda_{bit} = 1$ showed an effective balance between FID and FAB. Next, Figure 4(b) shows the results for different $M$ values in $\mathbf{u}_j^{Mt}, j\in \{m,h\}$ when the maximum timestep + +Table 2. Quantitative comparisons of various quantization methods at a resolution of $512\times 512$ using Stable Diffusion v1.4. For the bit precision options, W4A8 is used for Q-diffusion and PTQD, and W4A{6,8,10} are used for QLIP. + +
MethodBitoptionsRuntime (s)↓FAB↓FID↓
BK-SDM [19]W32A326.5032.0020.00
Q-diffusion [24]W4A84.538.0028.32
Q-diffusion [24]W4A165.5816.0024.68
+QLIPW4A{8,16,32}4.8510.5824.72
+ +Table 3. Runtime comparisons, where W4A{8,16,32} is used for QLIP. Measured on the Conceptual Captions dataset. + +![](images/66a7115134cd2c554b04b9f63bf7ba14c45fb8d1169aae7c49db9a7befd6be79.jpg) +(a) + +![](images/92b289e7a2dc3dab7c5a4ddf27e1f7efff5167b6e9fa91450f34cedcdfc7c06d.jpg) +(b) +Figure 4. Results on Q-diffusion using different hyperparameter settings of QLIP (W4A{6,8,10}): (a) $\lambda_{bit}$ and (b) $M$ . + +$T$ was 1000, where we obtained the best scores with $M = 200$ . + +# 4.2. Quantitative results + +Comparison with existing quantization methods. To demonstrate the effectiveness and applicability of QLIP, we tested QLIP on two diffusion model quantization methods: Q-diffusion [24] and PTQD [13]. Tables 1 and 2 provide comparison results for the COCO2017 [27] and Conceptual Captions [46] datasets using lightweight diffusion model BK-SDM-Tiny-2M [19] and Stable Diffusion v1.4 [39], respectively. + +As shown in Table 1, for the COCO2017 dataset, QLIP effectively reduced computational complexity of the base + +![](images/3ba0280fd6a05472baa38fa2c4ace5ffcc10e002c2a9b7dcf1f64e80c2e8de31.jpg) +Figure 5. Examples of variations in FAB by QLIP for the texts with different levels of richness and detail, along with the generated images. + +![](images/9452066d6a2a8f82a7aed4389a807bdf6f4ff3ffd6c433e5465dc93f57dcbc78.jpg) +Figure 6. Examples of the bit selection results and generated images. When the text prompt is specifically describing the image to be generated, the predicted image quality is high, and high and medium bits are assigned for most layers and time steps, as shown on the top left. On the other hand, when the text prompt has less detailed information, the predicted image quality is low, and low bits are assigned for many layers and time steps, as shown on the bottom left. On the right side, the generated images using Q-diffusion with and without applying QLIP are shown, demonstrating the effectiveness of QLIP in saving bits while improving the quality of generated images. + +line diffusion quantization methods. For example, Q-diffusion+QLIP (W4A{8,16,32}) achieved a reduction in FAB by 3.86 while maintaining similar FID and sFID, compared to Q-diffusion (W4A16). Furthermore, the use of QLIP consistently led to improvements when applied to PTQD, as demonstrated in Table 1. For example, PTQD+QLIP (W4A{8,16,32}) enhanced FID and sFID by 0.25 and 3.92, respectively, while also reducing FAB and BitOPs. The results on Conceptual Captions followed trends similar to those observed in COCO2017. For example, Q-diffusion+QLIP (W4A{8,16,32}) demonstrated reduced computational complexity compared to Q-diffusion on Conceptual Captions. + +Table 2 presents the results of applying lower-bit precision combinations to Stable Diffusion v1.4. Here, PTQD+QLIP (W4A{6,8,10}) achieved a 4.98 improvement + +in FID compared to PTQD (W4A8), while maintaining similar FAB using bit combinations efficiently. Although our proposed QLIP drives the quantization methods to reduce computational complexity and improve image quality, the CLIP Scores remained similar to those of the baseline quantization methods, indicating that the fitness to the text prompts is not significantly influenced by QLIP. Additional results on larger diffusion models can be found in the supplementary material. + +Runtime measurement. Runtime was measured with Cutlass [34], including the T2Q and Q2B modules, as shown in Table 3. The results show that runtime was reduced to a level similar to W4A8 while maintaining an FID score comparable to W4A16. Measurements were conducted on an RTX 3090. Additional analysis on memory requirements can be found in the supplementary material. + +# 4.3. Qualitative results + +Figure 5 visualizes the changes in FAB by the richness and detail level of the text, along with the generated images. Here, ChatGPT-4o was used to create texts of varying complexity that describe a given keyword. Specifically, the prompt used was "Create four sentences that describe how 'keyword' looks like, gradually getting richer and longer." As can be seen, QLIP assigned higher bits as the text describes the keyword more specifically and in greater detail, demonstrating that it utilizes information from the text to determine bit precisions for effective synthesis of vivid details and textures. + +Figure 6 shows several results comparing Q-diffusion with and without QLIP when applied to Stable Diffusion v1.4. As can be seen, our QLIP generates more detailed and natural images compared to PTQD by using various bit combinations. Furthermore, Figure 6 visualizes the bit precision maps determined by QLIP. When the T2Q module predicts high image quality from input text prompts, QLIP allocates more bits to enable rendering of vivid and accurate details from rich text prompts. Otherwise, QLIP employs lower bits for layers that are less sensitive to quantization, based on the text prompt and the time step. For example, when given less specific text prompts, the correlation between the generated image and the text prompt is relatively less critical, allowing for low-bit quantization in the cross-attention blocks. These results highlight the effectiveness and significance of dynamically allocating bit precisions according to text prompts. + +# 4.4. Ablation studies + +Ablation studies on the T2Q module. We report the results on the COCO2017 validation dataset using Stable Diffusion v1.4 with Q-diffusion as the baseline quantization method. Table 4 presents the results of the T2Q modules trained using various perceptual image quality metrics, including Realism score [21], CLIP-IQA [54], and GIQA [11]. The performance of the T2Q module was evaluated using Spearman's rank-order correlation coefficient (SROCC) [49] and Pearson's linear correlation coefficient (PLCC) [36] between the actual quality measures and the predictions of the T2Q module. QLIP with GIQA showed the most substantial improvements in FAB and FID along with the highest SROCC and PLCC scores. + +Ablation studies on the Q2B module. We conducted experiments on several variants of the Q2B module using Q-diffusion and Stable Diffusion v1.4 on the COCO2017 validation dataset. As shown in Table 5, when using either $\mathbf{p}_q$ alone or in combination with $\mathbf{p}_h^t$ , effective reduction in FAB was achieved at the expense of increase in FID. When using $\mathbf{p}_q$ and $\mathbf{p}_m^t$ , FID decreased while FAB increased. Finally, when $\mathbf{p}_q$ , $\mathbf{p}_h^t$ , and $\mathbf{p}_m^t$ were all involved, balanced results were obtained in two metrics. Implementation details of + +
MethodSROCC↑PLCC↑FAB↓FID↓
Q-diffusion [24]--8.0023.40
+QLIP
w/ Realism score0.51320.50228.1022.18
w/ CLIP-IQA0.71270.70828.5421.81
w/ GIQA0.80470.81087.8621.61
+ +Table 4. Ablation studies on the T2Q module: For the bit precision options, W4A8 in Q-diffusion and +QLIP W4A{6,8,10} were used. + +
MethodBitoptionsFAB↓FID↓
Q-diffusion [24]W4A88.0023.40
+ pqW4A{6,10}7.5726.91
+ pq + ptW4A{6,8,10}6.7329.37
+ pq + pmtW4A{6,8,10}8.6021.96
+ pq + pmt + pth (QLIP)W4A{6,8,10}7.8621.61
+ +Table 5. Ablation studies on the Q2B module. + +the ablation studies on the Q2B module can be found in the supplementary material. + +# 5. Conclusion + +The proposed quantization method for diffusion models, called QLIP, is the first method that uses text prompts for quantization in diffusion models. From our observation that the bit precision for quantization can be effectively determined based on the quality of the generated images, we designed the T2Q module that predicts image quality from input text prompts. We then introduced the Q2B module that allocates appropriate bit precision for each time step for different layers. By integrating the T2Q and Q2B modules into existing diffusion model quantization methods, we demonstrated the effectiveness of the proposed QLIP. QLIP presents a promising solution for quantizing diffusion models due to its ability to improve the quality of generated images using text prompts while reducing computational costs. + +Limitations and future directions. Similar to other generative models, QLIP also has the potential to generate counterfeit images for malicious purposes, potentially undermining trust in visual media. In a future study, we will extend QLIP to consider other input conditions, such as images or segmentation maps, in addition to text prompts. + +# Acknowledgement + +This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2022-NR070077). + +# References + +[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. eDiff-I: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 5 +[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: An analytic estimate of the optimal reverse variance in diffusion probabilistic models. In Proceedings of the International Conference on Learning Representations, 2022. 1 +[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.5 +[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2023.1 +[5] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. PACT: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 2 +[6] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5 +[7] Huanpeng Chu, Wei Wu, Chengjie Zang, and Kun Yuan. QNCD: Quantization noise correction for diffusion models. arXiv preprint arXiv:2403.19140, 2024. 2 +[8] Xiangxiang Chu, Shun Lu, Xudong Li, and Bo Zhang. Mix-Path: A unified approach for one-shot neural architecture search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5972–5981, 2023. 1 +[9] Zhen Dong, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. HAWQ: Hessian aware quantization of neural networks with mixed-precision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 293–302, 2019. 1, 3 +[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proceedings of the Advances in Neural Information Processing Systems, pages 2672–2680, 2014. 1 +[11] Shuyang Gu, Jianmin Bao, Dong Chen, and Fang Wen. GIQA: Generated image quality assessment. In Proceedings of the European Conference on Computer Vision, pages 369–385, 2020. 3, 4, 8 +[12] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Proceedings of the Advances in Neural Information Processing Systems, pages 1135-1143, 2015. 1 +[13] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. PTQD: Accurate post-training quantization + +for diffusion models. In Proceedings of the Advances in Neural Information Processing Systems, pages 13237-13249, 2023. 2, 3, 5, 6 +[14] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 5 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the Advances in Neural Information Processing Systems, pages 6629-6640, 2017. 5 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 3 +[17] Cheeun Hong and Kyoung Mu Lee. AdaBM: On-the-fly adaptive bit mapping for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2641-2650, 2024. 1, 3, 6 +[18] Cheeun Hong, Sungyong Baik, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. CADyQ: Content-aware dynamic quantization for image super-resolution. In Proceedings of the European Conference on Computer Vision, pages 367-383, 2022. 1, 2, 3, 6 +[19] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. BK-SDM: A lightweight, fast, and cheap version of stable diffusion. In Proceedings of the European Conference on Computer Vision, 2024. 2, 5, 6 +[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations, 2015. 5 +[21] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Proceedings of the Advances in Neural Information Processing Systems, pages 3927-3936, 2019. 8 +[22] Hongjae Lee, Jun-Sang Yoo, and Seung-Won Jung. RefQSR: Reference-based quantization for image super-resolution networks. IEEE Transactions on Image Processing, 33: 2823-2834, 2024. 1, 3, 5 +[23] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17413-17424, 2023. 1 +[24] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 1, 2, 3, 5, 6, 8 +[25] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. BRECQ: Pushing the limit of post-training quantization by block reconstruction. In Proceedings of the International Conference on Learning Representations, 2021. 1, 5 + +[26] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. SnapFusion: Text-to-image diffusion model on mobile devices within two seconds. In Proceedings of the Advances in Neural Information Processing Systems, pages 20662-20678, 2024. 2 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755, 2014. 3, 5, 6 +[28] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In Proceedings of the European Conference on Computer Vision, pages 19–34, 2018. 1 +[29] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In Proceedings of the International Conference on Learning Representations, 2022. 1 +[30] Zhenhua Liu, Yunhe Wang, Kai Han, Siwei Ma, and Wen Gao. Instance-aware dynamic neural network quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12434-12443, 2022. 1, 3 +[31] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Proceedings of the Advances in Neural Information Processing Systems, pages 5775-5787, 2022. 1 +[32] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 2 +[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Proceedings of the International Conference on Machine Learning, pages 8162-8171, 2021. 2 +[34] NVIDIA. CUTLASS, 2025. https://github.com/NVIDIA/cutlass.7 +[35] Nilesh Prasad Pandey, Marios Fournarakis, Chirag Patel, and Markus Nagel. Softmax bias correction for quantized generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshop, pages 1453-1458, 2023. 2 +[36] Karl Pearson. Notes on regression and inheritance in the case of two parents. Proceedings of the Royal Society of London, 58:240-242, 1895. 8 +[37] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *SDXL: Improving latent diffusion models for high-resolution image synthesis*. In *Proceedings of the International Conference on Learning Representations*, 2024. 1 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763, 2021. 4 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10674-10685, 2022. 1, 2, 3, 5, 6 +[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1 +[41] Hyogon Ryu, Seohyun Lim, and Hyunjung Shim. Memory-efficient personalization using quantized diffusion model. In Proceedings of the European Conference on Computer Vision, 2024. 2 +[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In Proceedings of the Advances in Neural Information Processing Systems, pages 36479–36494, 2022. 1 +[43] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Proceedings of the International Conference on Learning Representations, 2022. 2 +[44] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In Proceedings of the Advances in Neural Information Processing Systems, pages 2234-2242, 2016. 5 +[45] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1972-1981, 2023. 1, 2 +[46] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 2556-2565, 2018. 5, 6 +[47] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. In Proceedings of the Advances in Neural Information Processing Systems, pages 48686-48698, 2023. 1, 2, 3, 5 +[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Proceedings of the International Conference on Learning Representations, 2021. 1, 2 +[49] Charles Spearman. The proof and measurement of association between two things. The American Journal of Psychology, 15(1):72--101, 1961. 8 + +[50] Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. CHIP: Channel independence-based pruning for compact neural networks. In Proceedings of the Advances in Neural Information Processing Systems, pages 24604-24616, 2021. 1 +[51] Siao Tang, Xin Wang, Hong Chen, Chaoyu Guan, Zewen Wu, Yansong Tang, and Wenwu Zhu. Post-training quantization with progressive calibration and activation relaxing for text-to-image diffusion models. In Proceedings of the European Conference on Computer Vision, 2024. 2, 3 +[52] Senmao Tian, Ming Lu, Jiaming Liu, Yandong Guo, Yurong Chen, and Shunli Zhang. CABM: Content-aware bit mapping for single image super-resolution network with large input. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1756-1765, 2023. 1, 3 +[53] Mart Van Baalen, Christos Louizos, Markus Nagel, Rana Ali Amjad, Ying Wang, Tijmen Blankevoort, and Max Welling. Bayesian bits: Unifying quantization and pruning. In Proceedings of the Advances in Neural Information Processing Systems, pages 5741-5752, 2020. 5 +[54] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 8 +[55] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. HAQ: Hardware-aware automated quantization with mixed precision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8612-8620, 2019. 1, 3 +[56] Chuanguang Yang, Zhulin An, Helong Zhou, Fuzhen Zhuang, Yongjun Xu, and Qian Zhang. Online knowledge distillation via mutual contrastive learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10212-10227, 2023. 1 +[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1 +[58] Xuechen Zhang, Zheng Li, Samet Oymak, and Jiasi Chen. Text-to-3d generative ai on mobile devices: Measurements and optimizations. In Proceedings of the Workshop on Emerging Multimedia Systems, pages 8–14, 2023. 1 \ No newline at end of file diff --git a/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/images.zip b/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..51d8430bee879688106c5456728def0ef96a5e7b --- /dev/null +++ b/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c19192dad00e12412572257afc16c4f21306dd94a99034a97b08c5f60c5e99c0 +size 675289 diff --git a/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/layout.json b/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1432b1d8e6e1e49dee7f1ecf184751ba143c1993 --- /dev/null +++ b/2025/Text Embedding Knows How to Quantize Text-Guided Diffusion Models/layout.json @@ -0,0 +1,9678 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "type": "text", + "content": "Text Embedding Knows How to Quantize Text-Guided Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 143, + 483, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 143, + 483, + 172 + ], + "spans": [ + { + "bbox": [ + 130, + 143, + 483, + 172 + ], + "type": "text", + "content": "Hongjae Lee Myungjun Son Dongjea Kang Seung-Won Jung\\* Korea University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 173, + 443, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 173, + 443, + 186 + ], + "spans": [ + { + "bbox": [ + 167, + 173, + 443, + 186 + ], + "type": "text", + "content": "{jimmy9704,sonbill,kedjk11,swjung83}@korea.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 201, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 201, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 201, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 239, + 297, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 479 + ], + "type": "text", + "content": "Despite the success of diffusion models in image generation tasks such as text-to-image, the enormous computational complexity of diffusion models limits their use in resource-constrained environments. To address this, network quantization has emerged as a promising solution for designing efficient diffusion models. However, existing diffusion model quantization methods do not consider input conditions, such as text prompts, as an essential source of information for quantization. In this paper, we propose a novel quantization method dubbed Quantization of Language-to-Image diffusion models using text Prompts (QLIP). QLIP leverages text prompts to guide the selection of bit precision for every layer at each time step. In addition, QLIP can be seamlessly integrated into existing quantization methods to enhance quantization efficiency. Our extensive experiments demonstrate the effectiveness of QLIP in reducing computational complexity and improving the quality of the generated images across various datasets. The source code is available at our project page https://github.com/jimmy9704/QLIP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 504, + 137, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 137, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 137, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 525, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 693 + ], + "type": "text", + "content": "Diffusion models have achieved remarkable success in various generative tasks, particularly in text-to-image generation [39, 40, 57]. However, the huge computational complexity of diffusion models restricts their practical applicability, especially in resource-constrained environments. Specifically, diffusion models typically require hundreds of denoising steps to produce high-quality samples, making them considerably slower than generative adversarial networks (GANs) [10]. To this end, many studies introduced advanced training-free samplers to reduce the number of denoising iterations [2, 29, 31, 48]. However, the decrease in iterations alone is insufficient since the denoising models used in each iteration possess billions of parameters [4, 37, 42], leading to increased computational com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 214, + 474, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 474, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 474, + 226 + ], + "type": "text", + "content": "plexity and high memory consumption." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 228, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 503 + ], + "type": "text", + "content": "Besides other approaches toward light-weight network designs, such as network architecture search [8, 28], network pruning [12, 50], and knowledge distillation [23, 56], network quantization has received particular attention due to its simplicity, the increasing demand for low-precision operations, and related hardware support [9, 55, 58]. Although various network quantization methods have been proposed for convolutional neural networks (CNNs), the direct use of such methods for diffusion models has shown significant performance degradation [45]. As tailored solutions for diffusion models, several methods have been proposed considering the time step of diffusion models for quantization. For example, PTQ4DM [45] performed 8-bit quantization by constructing a calibration dataset based on the time step. Q-Diffusion [24] introduced an advanced method based on BRECQ [25] and performed the performance analysis of quantization methods for text-guided diffusion models. In addition, TDQ [47] performed quantization with varying activation scaling based on the diffusion time step. However, these methods do not use the text prompt provided to the diffusion model as valuable information for quantization, leaving room for further improvement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 506, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 651 + ], + "type": "text", + "content": "Meanwhile, recent studies highlight the advantages of dynamic bit precision quantization considering input conditions [30, 52]. For example, in the case of image superresolution, CADyQ [18] and AdaBM [17] performed dynamic quantization according to the complexity of the images and the importance of layers. RefQSR [22] clustered the image patches and assigned high-bit and low-bit precisions to the reference and query patches, respectively. These dynamic quantization methods allow for the cost-effective assignment of bit precision, enhancing quantization efficiency. However, such input-adaptive quantization methods have not yet been explored in diffusion models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we propose a novel quantization method, dubbed Quantization of Language-to-Image diffusion models using text Prompts (QLIP), which determines the bit precision of diffusion models according to the text prompt. Figure 1 shows our observation: The quality of the gener" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 142, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 142, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 142, + 713 + ], + "type": "text", + "content": "* corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15426" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 71, + 144, + 142 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 144, + 142 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 144, + 142 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 144, + 142 + ], + "type": "image", + "image_path": "4eb91b1bf14e292483921a31f8f9a8f863e441e39dd39c885c7b363aaf81194f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 147, + 228, + 191 + ], + "lines": [ + { + "bbox": [ + 66, + 147, + 228, + 191 + ], + "spans": [ + { + "bbox": [ + 66, + 147, + 228, + 191 + ], + "type": "text", + "content": "Pikachu, the symbol and eternal mascot of Pokémon, is the most famous of over 1,000 Pokémon and is a Pokémon with an electric attribute. It appeared in various video games, TV shows, movies, and merchandise since its introduction in the late 1990s." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 54, + 200, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 54, + 200, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 54, + 200, + 555, + 243 + ], + "type": "text", + "content": "Figure 1. Examples of the images generated by Stable Diffusion v1.4 using the text prompts given below images. When the text prompt details the shape, texture, and properties for the target object, the generated images have rich details and textures, as shown on the right; whereas the generated images are not very different from the images generated using only the keyword when the text prompt is less specific to the target object, as shown on the left." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 149, + 71, + 220, + 142 + ], + "blocks": [ + { + "bbox": [ + 149, + 71, + 220, + 142 + ], + "lines": [ + { + "bbox": [ + 149, + 71, + 220, + 142 + ], + "spans": [ + { + "bbox": [ + 149, + 71, + 220, + 142 + ], + "type": "image", + "image_path": "ec0073a3ac985fc020281a1356d7493a41b8f94db770caf202fbbe6f261ae37a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 232, + 71, + 304, + 142 + ], + "blocks": [ + { + "bbox": [ + 232, + 71, + 304, + 142 + ], + "lines": [ + { + "bbox": [ + 232, + 71, + 304, + 142 + ], + "spans": [ + { + "bbox": [ + 232, + 71, + 304, + 142 + ], + "type": "image", + "image_path": "bb842b45fe9d55c3ed0f3fbb581de141ca28649f5f0ab8b643855dd452716366.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 294, + 164, + 318, + 173 + ], + "lines": [ + { + "bbox": [ + 294, + 164, + 318, + 173 + ], + "spans": [ + { + "bbox": [ + 294, + 164, + 318, + 173 + ], + "type": "text", + "content": "Pikachu" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 380, + 141 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 380, + 141 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 380, + 141 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 380, + 141 + ], + "type": "image", + "image_path": "13febf055745aa57c7ff2933467dfa31a53dd9f95741acbbdc98aed283bc802d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 143, + 545, + 186 + ], + "lines": [ + { + "bbox": [ + 384, + 143, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 384, + 143, + 545, + 186 + ], + "type": "text", + "content": "Pikachu is a small, yellow, mouse-like creature with large, black-tipped ears that resemble lightning bolts. It has rosy cheeks, a small mouth, and large, expressive eyes. Pikachu's body is covered in soft, yellow fur, and it has a long, lightning bolt-shaped tail." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 392, + 72, + 462, + 141 + ], + "blocks": [ + { + "bbox": [ + 392, + 72, + 462, + 141 + ], + "lines": [ + { + "bbox": [ + 392, + 72, + 462, + 141 + ], + "spans": [ + { + "bbox": [ + 392, + 72, + 462, + 141 + ], + "type": "image", + "image_path": "84c03b2ff81d7d3d6496d316cfdd9546f634ee4c1abc10364779983f9d7bca2f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 468, + 72, + 539, + 141 + ], + "blocks": [ + { + "bbox": [ + 468, + 72, + 539, + 141 + ], + "lines": [ + { + "bbox": [ + 468, + 72, + 539, + 141 + ], + "spans": [ + { + "bbox": [ + 468, + 72, + 539, + 141 + ], + "type": "image", + "image_path": "8e769f7a176ca997ed6c3bcd792144e82df40fa92ec0790e889b314586e7ef37.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 265, + 294, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 265, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 54, + 265, + 294, + 348 + ], + "type": "text", + "content": "ated images increases as the text prompt has more specific information about the image to be generated. This suggests that high-bit precision operations are necessary to generate images with vivid details and textures corresponding to detailed text prompts; whereas, low-bit precision operations can be sufficient to generate images corresponding to text prompts that lack specific guidance on details and textures." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 350, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 295, + 529 + ], + "type": "text", + "content": "Motivated by this observation, we quantized the diffusion model [39] using Q-diffusion [24] with high-bit and low-bit settings and compared the quality of the generated images with the full-precision model. As shown in Figure 2, the performance drop of the low-bit model is noticeable when the quality of the generated images is high for the full-precision model. In contrast, when the quality is low for the full-precision model, the high-bit and low-bit models perform similarly to the full-precision model. These results suggest that diffusion models can be efficiently quantized without sacrificing performance if the quality of the generated images can be first predicted and then used to determine bit precision. Accordingly, our proposed QLIP predicts the image quality from the text prompt and determines the quantization bit precision for diffusion models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 530, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 295, + 601 + ], + "type": "text", + "content": "Furthermore, previous studies [13, 24, 45, 47] revealed that the activation distributions of diffusion models vary across layers over time steps. Inspired by these studies, QLIP uses learnable parameters to determine the bit precision for each layer and for each timestep. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 603, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 638 + ], + "type": "text", + "content": "- To the best of our knowledge, QLIP is the first approach that explicitly utilizes text prompts for dynamic quantization in diffusion models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 639, + 294, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 294, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 294, + 674 + ], + "type": "text", + "content": "- We propose leveraging text prompts to predict the image quality, which is used to guide the selection of bit precision for every layer in different denoising time steps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 675, + 294, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 294, + 711 + ], + "type": "text", + "content": "- Our proposed QLIP can be integrated into existing diffusion quantization methods to enhance quantization efficiency." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 263, + 398, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 263, + 398, + 276 + ], + "spans": [ + { + "bbox": [ + 314, + 263, + 398, + 276 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 284, + 536, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 536, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 536, + 297 + ], + "type": "text", + "content": "2.1. Computationally efficient diffusion models" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 302, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 302, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 312, + 302, + 553, + 422 + ], + "type": "text", + "content": "Diffusion models generate high-quality images by iteratively removing noise from random noise. Despite their capability to generate high-quality images, diffusion models encounter challenges in real-world applications due to the time-consuming iterative denoising process. Thus, many efforts have been made to make the diffusion process computationally efficient, including reducing the number of iterative diffusion processes via a non-Markovian process [48], adjusting the variance scheduling [33], and employing knowledge distillation [32, 43]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 422, + 553, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 553, + 518 + ], + "type": "text", + "content": "Meanwhile, since the computational complexity of diffusion model architectures also limits their practical applicability, especially in resource-constrained environments, research efforts have been made to lighten denoising models. For example, SnapFusion [26] introduced an efficient U-Net architecture based on Stable Diffusion [39], while BKSDM [19] further lightened the architecture through knowledge distillation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 528, + 491, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 491, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 491, + 540 + ], + "type": "text", + "content": "2.2. Quantization of diffusion models" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Quantization involves mapping 32-bit floating-point values of weights and activations to lower-bit values. Numerous studies have explored quantization to compress and accelerate neural networks [5, 18], which can be classified into two approaches: quantization-aware training (QAT) and post-training quantization (PTQ). Recent studies [7, 35, 41, 51] on quantizing diffusion models have focused mainly on PTQ due to the inability to use training data for privacy and commercial reasons, as well as the high computational cost of QAT. Specifically, PTQ4DM [45] and Q-diffusion [24] performed quantization of diffusion models by generating a time step-aware calibration set that consists of data sampled across various time steps using the full-precision denoising model. PTQD [13] separated the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15427" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 330, + 220 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 330, + 220 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 330, + 220 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 330, + 220 + ], + "type": "image", + "image_path": "6ff6701630ba33acf936ede1f816cfe9163784d4bf9677452a8cbfa01de6b369.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 228, + 555, + 296 + ], + "lines": [ + { + "bbox": [ + 54, + 228, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 555, + 296 + ], + "type": "text", + "content": "Figure 2. Analysis of the effect of quantization and the quality of generated images, where the image quality is measured using the GIQA score [11]. On the left side, the x-axis represents the quality of images generated by the full-precision model, and the y-axis represents the quality difference between the generated images using the full-precision and quantized models. The images are generated by Stable Diffusion v1.4 [39] using 5000 captions from the COCO 2017 dataset [27] with a size of " + }, + { + "bbox": [ + 54, + 228, + 555, + 296 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 54, + 228, + 555, + 296 + ], + "type": "text", + "content": ". The high-bit (W4A10) and low-bit (W4A6) quantized models are obtained by applying Q-diffusion [24]. On the right side, examples of images generated using models that are unquantized or quantized w/wo QLIP are shown for two different image quality levels." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 342, + 72, + 419, + 135 + ], + "blocks": [ + { + "bbox": [ + 342, + 72, + 419, + 135 + ], + "lines": [ + { + "bbox": [ + 342, + 72, + 419, + 135 + ], + "spans": [ + { + "bbox": [ + 342, + 72, + 419, + 135 + ], + "type": "image", + "image_path": "80c76b4ebb90825bf8842e67d5e322bd47ed58ffab25c4c264224097e2bc7cd7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 419, + 72, + 481, + 134 + ], + "blocks": [ + { + "bbox": [ + 419, + 72, + 481, + 134 + ], + "lines": [ + { + "bbox": [ + 419, + 72, + 481, + 134 + ], + "spans": [ + { + "bbox": [ + 419, + 72, + 481, + 134 + ], + "type": "image", + "image_path": "2f3acb12f551e68b939b3cb4065830534c18ed65279ab62d59565e8ad6c8b99b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 482, + 72, + 544, + 134 + ], + "blocks": [ + { + "bbox": [ + 482, + 72, + 544, + 134 + ], + "lines": [ + { + "bbox": [ + 482, + 72, + 544, + 134 + ], + "spans": [ + { + "bbox": [ + 482, + 72, + 544, + 134 + ], + "type": "image", + "image_path": "9fe47efa25c24fb108e881a9e4640205b270f3ed8d3eaad9928b690b7a828026.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 345, + 137, + 419, + 212 + ], + "blocks": [ + { + "bbox": [ + 345, + 137, + 419, + 212 + ], + "lines": [ + { + "bbox": [ + 345, + 137, + 419, + 212 + ], + "spans": [ + { + "bbox": [ + 345, + 137, + 419, + 212 + ], + "type": "image", + "image_path": "c1d25611b96184d123cbae5f1e8557c49dd2aa2096161f03caad283c5414bc72.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 419, + 137, + 481, + 211 + ], + "blocks": [ + { + "bbox": [ + 419, + 137, + 481, + 211 + ], + "lines": [ + { + "bbox": [ + 419, + 137, + 481, + 211 + ], + "spans": [ + { + "bbox": [ + 419, + 137, + 481, + 211 + ], + "type": "image", + "image_path": "a36df503c5555d2cebe04756dd2aec3253cc66cec91e1bb83b8c9178f52becc7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 482, + 137, + 544, + 211 + ], + "blocks": [ + { + "bbox": [ + 482, + 137, + 544, + 211 + ], + "lines": [ + { + "bbox": [ + 482, + 137, + 544, + 211 + ], + "spans": [ + { + "bbox": [ + 482, + 137, + 544, + 211 + ], + "type": "image", + "image_path": "4cb782038d7fdd814494d66a7f43d49c5a862caf53c38483fc0545de08a3bc84.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 316, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 296, + 448 + ], + "type": "text", + "content": "quantization noise from the diffusion noise to reduce quantization errors. Additionally, PTQD introduced a time step-based mixed-precision method but did not employ varying bit precision for each layer. Furthermore, TDQ [47] performed quantization by adaptively scaling the activations for different diffusion time steps and layers, but assigned the same bit precision for all time steps. Although these diffusion model quantization methods have shown progress in reducing the complexity of noise estimation models, they do not adjust quantization parameters based on input conditions such as text prompts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 456, + 253, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 253, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 253, + 468 + ], + "type": "text", + "content": "2.3. Input-adaptive dynamic quantization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 473, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 641 + ], + "type": "text", + "content": "Recent advances in hardware have enabled mixed-precision operations, allowing various layers to use different bit precisions to improve performance and computational efficiency [9, 55]. Many methods have been proposed to determine the bit precision of each layer of CNNs [9, 30, 55]. For example, several input-adaptive dynamic quantization methods [17, 18, 22, 52] have been introduced for super-resolution, which adjust bit precisions based on the complexity of image patches and the sensitivity to quantization. However, diffusion models have not yet explored such input-adaptive quantization methods. Therefore, this paper proposes a novel quantization method for text-guided diffusion models that applies adaptive quantization using input conditions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 651, + 111, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 111, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 111, + 664 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 671, + 141, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 141, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 141, + 683 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Diffusion model. The reverse process of a diffusion model gradually removes noise from random data " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 316, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 365 + ], + "type": "text", + "content": "produce a high-quality image " + }, + { + "bbox": [ + 313, + 316, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 316, + 555, + 365 + ], + "type": "text", + "content": ", which can be expressed as " + }, + { + "bbox": [ + 313, + 316, + 555, + 365 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\mathbf{x}_{t - 1}|\\mathbf{x}_t) = \\mathcal{N}(\\mathbf{x}_{t - 1};\\mu_\\theta (\\mathbf{x}_t,t),\\sigma_t^2\\mathbf{I})" + }, + { + "bbox": [ + 313, + 316, + 555, + 365 + ], + "type": "text", + "content": ". The denoising process at time step " + }, + { + "bbox": [ + 313, + 316, + 555, + 365 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 316, + 555, + 365 + ], + "type": "text", + "content": " using the noise estimation model " + }, + { + "bbox": [ + 313, + 316, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 316, + 555, + 365 + ], + "type": "text", + "content": " can then be represented as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 340, + 369, + 555, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 369, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 340, + 369, + 555, + 396 + ], + "type": "interline_equation", + "content": "\\mu_ {\\theta} (\\mathbf {x} _ {t}, t) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {x} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t)\\right), \\quad (1)", + "image_path": "3c90fb4c893235de63b22214178057f721b1e703357a2173303bc4567501f8e0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 399, + 554, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 435 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 399, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\alpha_{t},\\bar{\\alpha}_{t}" + }, + { + "bbox": [ + 313, + 399, + 554, + 435 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 399, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 399, + 554, + 435 + ], + "type": "text", + "content": " are the parameters that control the strength of the Gaussian noise in each step, and " + }, + { + "bbox": [ + 313, + 399, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 313, + 399, + 554, + 435 + ], + "type": "text", + "content": " is a variance schedule [16]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 435, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 554, + 506 + ], + "type": "text", + "content": "Diffusion model quantization. Diffusion model quantization reduces the complexity of a noise estimation model that requires hundreds of iterative sampling processes. The denoising process at time step " + }, + { + "bbox": [ + 313, + 435, + 554, + 506 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 435, + 554, + 506 + ], + "type": "text", + "content": " using the quantized noise estimation model " + }, + { + "bbox": [ + 313, + 435, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta}" + }, + { + "bbox": [ + 313, + 435, + 554, + 506 + ], + "type": "text", + "content": " and the quantized data " + }, + { + "bbox": [ + 313, + 435, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\dot{\\mathbf{x}}_t" + }, + { + "bbox": [ + 313, + 435, + 554, + 506 + ], + "type": "text", + "content": " can be expressed as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 340, + 521, + 553, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 521, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 340, + 521, + 553, + 563 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\mu} _ {\\theta} \\left(\\hat {\\mathbf {x}} _ {t}, t\\right) = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\hat {\\mathbf {x}} _ {t} - \\frac {\\beta_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\hat {\\epsilon} _ {\\theta} \\left(\\hat {\\mathbf {x}} _ {t}, t\\right)\\right), \\tag {2} \\\\ w h e r e \\hat {\\epsilon} _ {\\theta} = Q _ {\\mathbf {b} ^ {t}} (\\epsilon_ {\\theta}). \\\\ \\end{array}", + "image_path": "26c6ae294f78db0dae3599374651d385b878e942ef107e8898b85cd71395e0ea.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^t\\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "content": " represents a set of quantization bit precisions for " + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "content": " layers at time step " + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "inline_equation", + "content": "Q_{\\mathbf{b}^t}" + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "content": " is a function that performs layer-wise and time step-wise quantization using " + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^t" + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "content": ". To our knowledge, all previous diffusion quantization methods set the same bit precision to all layers, i.e., " + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^t (1) = \\mathbf{b}^t (2) = \\dots = \\mathbf{b}^t (K)" + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "content": ", and only a few methods [13, 51] performed time step-wise dynamic quantization, i.e., " + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^{t_i}\\neq \\mathbf{b}^{t_j}" + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "content": " for some time steps " + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "inline_equation", + "content": "t_i\\neq t_j" + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "content": ". However, such methods do not adapt the bit precision for each layer. Our proposed QLIP determines " + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^t" + }, + { + "bbox": [ + 313, + 569, + 555, + 713 + ], + "type": "text", + "content": " by training so that the quantization error can be effectively reduced without sacrificing the performance of diffusion models." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15428" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 71, + 536, + 189 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 536, + 189 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 536, + 189 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 536, + 189 + ], + "type": "image", + "image_path": "b4b5aca18b19d9b948d6eda78d52618574b5054bc506997ebc73f8da3e0518e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 197, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 54, + 197, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 54, + 197, + 555, + 232 + ], + "type": "text", + "content": "Figure 3. Overview of QLIP that consists of two key modules: T2Q module that predicts image quality from the input text prompt and Q2B module that determines bit precisions for all denoising layers and time steps. Only the Q2B module is trained during QLIP training. See Sections 3.3 and 3.4 for the details about the pre-training of the T2Q module and diffusion models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 178, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 178, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 178, + 262 + ], + "type": "text", + "content": "3.2. Framework overview" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "type": "text", + "content": "Our observations shown in Figure 2 indicate that bit precisions for quantization can be determined cost-effectively given the quality of the generated image. However, this image quality remains unknown during the inference stage, motivating us to predict it using an available input, i.e., text prompts. Specifically, if text prompts have rich and detailed information, the expected quality of the generated image is high, and vice versa. To this end, we introduce a text-to-quality (T2Q) module " + }, + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "type": "text", + "content": " that predicts image quality " + }, + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "type": "text", + "content": " from the embedding vector " + }, + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "type": "inline_equation", + "content": "q = \\phi(\\mathbf{z})" + }, + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 268, + 296, + 400 + ], + "type": "text", + "content": " is obtained by CLIP-Text [38] using the input text prompt." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 400, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 519 + ], + "type": "text", + "content": "Next, as we discussed, high-bit quantization is required to generate images with vivid details and textures corresponding to detailed text prompts; while low-bit quantization is sufficient to generate images using less specific text prompts. Therefore, we design a quality-to-bit (Q2B) module " + }, + { + "bbox": [ + 55, + 400, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 400, + 296, + 519 + ], + "type": "text", + "content": " that maps image quality " + }, + { + "bbox": [ + 55, + 400, + 296, + 519 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 400, + 296, + 519 + ], + "type": "text", + "content": " to appropriate bit precision, i.e., " + }, + { + "bbox": [ + 55, + 400, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^t = \\psi(q)" + }, + { + "bbox": [ + 55, + 400, + 296, + 519 + ], + "type": "text", + "content": ". Note that the Q2B module determines bit precisions for each time step for different layers since the activation distributions of diffusion models vary across layers over time steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 520, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 591 + ], + "type": "text", + "content": "In summary, the proposed QLIP obtains a quantized diffusion model " + }, + { + "bbox": [ + 55, + 520, + 296, + 591 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon}_{\\theta} = Q_{\\psi (\\phi (\\mathbf{z}))}(\\epsilon_{\\theta})" + }, + { + "bbox": [ + 55, + 520, + 296, + 591 + ], + "type": "text", + "content": " by learning the parameters of " + }, + { + "bbox": [ + 55, + 520, + 296, + 591 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 520, + 296, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 520, + 296, + 591 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 520, + 296, + 591 + ], + "type": "text", + "content": ". The overview of the proposed QLIP is shown in Figure 3. The following subsections detail the T2Q and Q2B modules and their training procedure and loss functions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 599, + 203, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 203, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 203, + 612 + ], + "type": "text", + "content": "3.3. Text-to-Quality estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": "The T2Q module, denoted as " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", aims to predict the image quality from the input text prompt. " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " consists of three linear layers and outputs an image quality score " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "q = \\phi(\\mathbf{z})" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^{C_{clip}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " represents the CLIP text embedding, and " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "C_{clip}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " denotes its channel dimension. For the training of the T2Q module, we collect a dataset " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{T2Q} = \\{\\mathbf{z}^i, \\mathbf{x}_0^i\\}_{i=1}^{N_{T2Q}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^i" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^i" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": "-th sample of the CLIP text embedding and its corresponding generated" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "text", + "content": "image obtained using the full-precision diffusion model, respectively, and " + }, + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "inline_equation", + "content": "N_{T2Q}" + }, + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "text", + "content": " is the number of images in " + }, + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{T2Q}" + }, + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\bar{q}^i" + }, + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "text", + "content": " represent the (pseudo) ground-truth image quality of " + }, + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0^i" + }, + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "text", + "content": ", where we used GIQA [11] as the default quality measure, with an additional normalization to make the range of quality values [0, 1]. See Section 4.4 for the performance analysis with different quality measures. The T2Q module is trained using the loss " + }, + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "inline_equation", + "content": "L_{t2q}" + }, + { + "bbox": [ + 313, + 251, + 555, + 348 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 358, + 555, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 358, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 359, + 358, + 555, + 393 + ], + "type": "interline_equation", + "content": "L _ {t 2 q} = \\frac {1}{N _ {T 2 Q}} \\sum_ {i = 1} ^ {N _ {T 2 Q}} \\left(\\bar {q} ^ {i} - \\phi \\left(\\mathbf {z} ^ {i}\\right)\\right) ^ {2}. \\tag {3}", + "image_path": "f95abbaaee16455618e3d6ef1c54d7bba8eaba15dc2dfcd2c0cf57d6360d5d1e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 401, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 437 + ], + "type": "text", + "content": "Once the T2Q module is trained, it is frozen during the training of the Q2B module and served as a quality predictor to help determine quantization bit precisions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 446, + 454, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 454, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 454, + 459 + ], + "type": "text", + "content": "3.4. Quality-to-Bit estimation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 464, + 554, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 549 + ], + "type": "text", + "content": "In text-guided diffusion models, the richer and more detailed the information in the text prompt is, the higher the image quality tends to be, requiring the use of higher bits. Therefore, the image quality " + }, + { + "bbox": [ + 313, + 464, + 554, + 549 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 464, + 554, + 549 + ], + "type": "text", + "content": " predicted by the T2Q module can serve as a direct hint for bit selection. To this end, the proposed Q2B module obtains a probability vector " + }, + { + "bbox": [ + 313, + 464, + 554, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_q \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 464, + 554, + 549 + ], + "type": "text", + "content": " for bit selection as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 381, + 559, + 555, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 559, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 381, + 559, + 555, + 574 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {q} = \\sigma \\left((q - 0. 5) \\mathbf {s} + \\mathbf {o}\\right), \\tag {4}", + "image_path": "6b6a22a95b088d6bd4c711aa6ffaedfb56bf9a2c803f660178a511ed7eedcf1a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{o} \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "text", + "content": " represent sets of learnable parameters, and " + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "text", + "content": " denotes the sigmoid function. Furthermore, to adjust the bit precision for each time step, additional learnable parameters " + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_m^t \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_h^t \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "text", + "content": " are introduced to obtain probability vectors " + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_m^t \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_h^t \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 583, + 554, + 654 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 346, + 656, + 553, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 656, + 553, + 671 + ], + "spans": [ + { + "bbox": [ + 346, + 656, + 553, + 671 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {j} ^ {t} = \\sigma \\left(\\mathbf {u} _ {j} ^ {t}\\right), \\quad j \\in \\{m, h \\}, t = 1, 2, \\dots , T, \\tag {5}", + "image_path": "a031b6e3b186d1fa75059ace2fa608dff8f82f4117f6465a200e772f9f50f060.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 525, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 525, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 525, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 525, + 689 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 677, + 525, + 689 + ], + "type": "text", + "content": " is the total number of denoising time steps." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "These probability vectors, " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_q" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_m^t" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_h^t" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": ", are used to determine the bit precision for each layer and each time" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15429" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "step. Specifically, we define a set of supported bit precis \nsions as " + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{b_{low},b_{med},b_{high}\\}" + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": " . Then, the probabilities of \nselecting " + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "inline_equation", + "content": "b_{low},b_{med}" + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "inline_equation", + "content": "b_{high}" + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": " are obtained as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 112, + 218, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 112, + 218, + 127 + ], + "spans": [ + { + "bbox": [ + 91, + 112, + 218, + 127 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {b _ {l o w}} ^ {t} = \\left(1 - \\mathbf {p} _ {q}\\right) \\odot \\left(1 - \\mathbf {p} _ {m} ^ {t}\\right),", + "image_path": "0a8d72bad006adee4658a427c9fda916502c6dffc1623c86a7cbb6f753b7847a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 90, + 129, + 295, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 129, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 90, + 129, + 295, + 144 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {b _ {m e d}} ^ {t} = \\left(1 - \\mathbf {p} _ {q}\\right) \\odot \\mathbf {p} _ {m} ^ {t} + \\mathbf {p} _ {q} \\odot \\left(1 - \\mathbf {p} _ {h} ^ {t}\\right), \\tag {6}", + "image_path": "09588836cd754ec0b30c374faefda548356df29f1680f8be7f76fdeca3c6b9aa.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 145, + 165, + 161 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 145, + 165, + 161 + ], + "spans": [ + { + "bbox": [ + 89, + 145, + 165, + 161 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {b _ {h i g h}} ^ {t} = \\mathbf {p} _ {q} \\odot \\mathbf {p} _ {h} ^ {t},", + "image_path": "3e2587f67380f0b9ddaa51756ed871af3fd61ccf256e6f8eef2d528186ef8eb8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "type": "text", + "content": " represents element-wise multiplication, " + }, + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{b_{low}}^{t} \\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{b_{med}}^{t} \\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{b_{high}}^{t} \\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "type": "text", + "content": " are the probability vectors for bit selection. Finally, the bit precision for each layer at time step " + }, + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 164, + 296, + 214 + ], + "type": "text", + "content": " is determined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 217, + 296, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 240 + ], + "type": "interline_equation", + "content": "\\mathbf {b} ^ {t} (k) = \\operatorname * {a r g m a x} _ {i \\in \\{b _ {\\text {l o w}}, b _ {\\text {m e d}}, b _ {\\text {h i g h}} \\}} \\mathbf {p} _ {i} ^ {t} (k), k = 1, 2, \\dots , K. \\tag {7}", + "image_path": "f1389e42cc1926cdc714b9edcd3cc236c10fafac47b88119faf4ea7e2825260c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": "However, the training of " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_m^t" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_h^t" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": " for all " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": " poses a challenge as " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": " increases. In addition, the activation distributions of the diffusion models are known to be similar to each other for adjacent time steps [13, 47]. Consequently, we learn the parameters for every " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": " time step, i.e., " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_j^{Mt}" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": " and use the same parameters for the intermediate time steps, i.e., " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_j^{M(t + 1) - 1} = \\mathbf{u}_j^{M(t + 1) - 2} = \\dots = \\mathbf{u}_j^{Mt}" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "j\\in \\{m,h\\}" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": ". Next, previous studies [1, 6] revealed that the initial stage of the reverse process involves generating the rough context of an image corresponding to the input text prompt; thus, the use of low bits in the initial time steps could decrease the correlation between the text prompt and the generated image. Therefore, we set all values in " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_q" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": " to 1 for the first " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": " time steps in Equation (6) to enforce the selection of high bits in the initial stage of the reverse process. In summary, the Q2B module has " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "2K" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "2KT / M" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_m^t" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_h^t" + }, + { + "bbox": [ + 55, + 244, + 296, + 449 + ], + "type": "text", + "content": ") trainable parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 449, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 496 + ], + "type": "text", + "content": "The bit selection in Equation (7) is non-differentiable, making it unable to backpropagate. Thus, we adopt a straight-through estimator [3] to render this process differentiable:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 499, + 100, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 499, + 100, + 513 + ], + "spans": [ + { + "bbox": [ + 65, + 499, + 100, + 513 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {a}} ^ {t} (k) =", + "image_path": "a8f603ac29540f4685d473d17440a2425b348bc599425c50fee50dde839d121f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 516, + 296, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 516, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 67, + 516, + 296, + 557 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} Q _ {\\mathbf {b} ^ {t} (k)} \\left(\\mathbf {a} ^ {t} (k)\\right) & \\text {f o r w a r d}, \\\\ \\sum_ {i \\in \\left\\{b _ {l o w}, b _ {m e d}, b _ {h i g h} \\right\\}} \\mathbf {p} _ {i} ^ {t} (k) \\cdot Q _ {i} \\left(\\mathbf {a} ^ {t} (k)\\right) & \\text {b a c k w a r d}. \\end{array} \\right. \\tag {8}", + "image_path": "61c8eb81c0e2392d05292835303456065bda41689f7bdbdd21dd8dbb38c0a254.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^t (k)" + }, + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{a}}^t (k)" + }, + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "text", + "content": " represent the activations of the full-precision and quantized models at time step " + }, + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "text", + "content": "-th layer. In other words, during backpropagation, the activations quantized by all bit precisions in " + }, + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 558, + 296, + 665 + ], + "type": "text", + "content": " are weighted by their corresponding estimated probabilities. Since diffusion models are less sensitive to the quantization of the weights [13, 24], the weights are quantized using a fixed precision, which is empirically chosen as 4 bits for all layers and time steps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Given the full precision diffusion model, we obtain its quantized weights by applying the baseline diffusion quantization methods [13, 24]. The quantization parameters, including the maximum and minimum values for clipping, are" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "also predetermined for all bit precisions in " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ". Then, while freezing all the other modules in Figure 3, only the Q2B module is trained using a small calibration set. The Q2B module is trained at randomly sampled time steps " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " using the loss function " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "L_{QLIP}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 139, + 554, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 139, + 554, + 199 + ], + "spans": [ + { + "bbox": [ + 320, + 139, + 554, + 199 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {Q L I P} = \\left(\\epsilon_ {\\theta} (\\mathbf {x} _ {t}, t) - \\hat {\\epsilon} _ {\\theta} (\\hat {\\mathbf {x}} _ {t}, t)\\right) ^ {2} \\\\ + \\lambda_ {b i t} \\left(b _ {h i g h} \\cdot \\sum_ {k = 1} ^ {K} \\mathbf {p} _ {b _ {h i g h}} ^ {t} (k) + b _ {m e d} \\cdot \\sum_ {k = 1} ^ {K} \\mathbf {p} _ {b _ {m e d}} ^ {t} (k)\\right). \\tag {9} \\\\ \\end{array}", + "image_path": "7699f83e50ee27a74dd77af40162f5d64c73a903ccb3833c595e7b94add61e23.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 200, + 554, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 554, + 259 + ], + "type": "text", + "content": "Here, the first term measures the difference between the results of the full-precision and quantized models, and the second term measures the bit lengths for the high- and medium-precision models. " + }, + { + "bbox": [ + 313, + 200, + 554, + 259 + ], + "type": "inline_equation", + "content": "\\lambda_{bit}" + }, + { + "bbox": [ + 313, + 200, + 554, + 259 + ], + "type": "text", + "content": " is a weighting parameter that balances between these two terms." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 270, + 395, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 395, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 395, + 283 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 289, + 429, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 429, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 429, + 303 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 307, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 629 + ], + "type": "text", + "content": "Datasets and baseline quantization methods. To evaluate the effectiveness of QLIP across different diffusion models, we conducted experiments using two diffusion networks: BK-SDM-Tiny-2M [19] and Stable Diffusion v1.4 [39]. Our experiments involve two datasets, COCO2017 [27] and Conceptual Captions [46]. To train the T2Q module, 10k images were generated by Stable Diffusion v1.4 [39] using 10k text annotations in the COCO2017 training dataset as input text prompts. We generated another 1024 images using unused text annotations in the COCO2017 training dataset to train the Q2B module. We evaluated the effectiveness of our proposed QLIP by applying it to the baseline diffusion quantization methods, Q-diffusion [24] and PTQD [13]. In the implementation of PTQD, we set the noise correction parameter to " + }, + { + "bbox": [ + 313, + 307, + 555, + 629 + ], + "type": "inline_equation", + "content": "b_{med}" + }, + { + "bbox": [ + 313, + 307, + 555, + 629 + ], + "type": "text", + "content": ", while noise correction was not applied if the estimated image quality by the T2Q module is below 0.3 or above 0.7. We trained the T2Q module for 3 epochs using the Adam optimizer [20] with a learning rate of 0.001 and the Q2B module for 5000 iterations using BRECQ [25], as adopted in Q-diffusion and PTQD, using RTX 3090. For performance evaluation, following the baseline quantization method [24], we generated 10k images using 10k text annotations in the COCO2017 validation dataset. Moreover, to evaluate the generalization ability of the quantized models, we used Conceptual Captions to generate additional 10k images." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Evaluation metrics. We report widely adopted metrics such as FID [15] and sFID [44] to evaluate image quality. In addition, we use CLIP Score [14] to assess the degree of matching between the generated images and their corresponding text prompts. To quantify computational efficiency, we measure BitOPs [22, 53], which account for the required operations in both the denoising model and addi" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15430" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 555, + 163 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 555, + 163 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 555, + 163 + ], + "type": "table", + "html": "
COCO2017
MethodBitoptionsFAB↓BitOPs (T)↓FID↓sFID↓CLIP Score↑
BK-SDM [19]W32A3232.0010.4623.7966.190.3069
Q-diffusion [24]W4A1616.001.0330.0273.250.3068
+QLIPW4A{8,16,32}12.140.8830.0173.240.3063
PTQD [13]W4A1616.001.0330.2777.180.3069
+QLIPW4A{8,16,32}12.140.8830.0273.260.3063
", + "image_path": "ca2bb2ebb84b59dc44eadd91a3fcdfcfb45924e662480e934f03e41c8e6b09b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 163, + 555, + 252 + ], + "blocks": [ + { + "bbox": [ + 56, + 163, + 555, + 252 + ], + "lines": [ + { + "bbox": [ + 56, + 163, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 555, + 252 + ], + "type": "table", + "html": "
Conceptual Captions
MethodBitoptionsFAB↓BitOPs (T)↓FID↓sFID↓CLIP Score↑
BK-SDM [19]W32A3232.0010.4620.0047.800.2983
Q-diffusion [24]W4A1616.001.0324.6862.700.2965
+QLIPW4A{8,16,32}10.580.8224.7259.540.2964
PTQD [13]W4A1616.001.0324.6662.600.2965
+QLIPW4A{8,16,32}10.580.8224.7159.250.2966
", + "image_path": "004a962e7c72cb89b5b418b838a66c76a8ada03a5f368b33ec81e8cf060cb041.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 289, + 294, + 472 + ], + "blocks": [ + { + "bbox": [ + 85, + 258, + 523, + 270 + ], + "lines": [ + { + "bbox": [ + 85, + 258, + 523, + 270 + ], + "spans": [ + { + "bbox": [ + 85, + 258, + 523, + 270 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons of various quantization methods at a resolution of " + }, + { + "bbox": [ + 85, + 258, + 523, + 270 + ], + "type": "inline_equation", + "content": "{512} \\times {512}" + }, + { + "bbox": [ + 85, + 258, + 523, + 270 + ], + "type": "text", + "content": " using BK-SDM-Tiny-2M." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 289, + 294, + 472 + ], + "lines": [ + { + "bbox": [ + 57, + 289, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 57, + 289, + 294, + 472 + ], + "type": "table", + "html": "
COCO2017
MethodFAB↓FID↓sFID↓CLIP Score↑
Stable Diffusion [39]32.0022.2365.110.3174
Q-diffusion [24]8.0023.4066.570.3126
+QLIP7.8621.6164.320.3120
PTQD [13]8.0022.7568.630.3126
+QLIP7.8621.3565.810.3120
Conceptual Captions
MethodFAB↓FID↓sFID↓CLIP Score↑
Stable Diffusion [39]32.0015.6745.610.3078
Q-diffusion [24]8.0020.7448.330.2984
+QLIP7.7719.8947.630.2979
PTQD [13]8.0025.4055.010.2972
+QLIP7.7720.4248.830.2975
", + "image_path": "30cfcdc1a6a3afa28ff14724ca66374030c63e9ecdffaffae304df6cda1fca67.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "type": "text", + "content": "tional modules for QLIP implementation, and feature average bit-width (FAB) [17, 18] for the denoising model. In particular, BitOPs are measured as " + }, + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "type": "inline_equation", + "content": "MACs \\cdot \\frac{b_w}{32} \\cdot \\frac{b_a}{32}" + }, + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "type": "inline_equation", + "content": "MACs" + }, + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "type": "text", + "content": " denotes multiply-accumulate computations, and " + }, + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "type": "inline_equation", + "content": "b_w" + }, + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "type": "inline_equation", + "content": "b_a" + }, + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "type": "text", + "content": " represent the bitwidths of weights and activations, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "Hyperparameter settings. We conducted experiments on various hyperparameter settings for the COCO2017 validation dataset using Q-diffusion as the baseline in Stable Diffusion v1.4. First, Figure 4(a) presents the results obtained with various values of the weight term " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{bit}" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " in Equation (9), where " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\lambda_{bit} = 1" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " showed an effective balance between FID and FAB. Next, Figure 4(b) shows the results for different " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " values in " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_j^{Mt}, j\\in \\{m,h\\}" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " when the maximum timestep" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 289, + 556, + 354 + ], + "blocks": [ + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons of various quantization methods at a resolution of " + }, + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 55, + 479, + 295, + 523 + ], + "type": "text", + "content": " using Stable Diffusion v1.4. For the bit precision options, W4A8 is used for Q-diffusion and PTQD, and W4A{6,8,10} are used for QLIP." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 289, + 556, + 354 + ], + "lines": [ + { + "bbox": [ + 315, + 289, + 556, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 289, + 556, + 354 + ], + "type": "table", + "html": "
MethodBitoptionsRuntime (s)↓FAB↓FID↓
BK-SDM [19]W32A326.5032.0020.00
Q-diffusion [24]W4A84.538.0028.32
Q-diffusion [24]W4A165.5816.0024.68
+QLIPW4A{8,16,32}4.8510.5824.72
", + "image_path": "7d7ab1170578e7defac034e9ae1cc329b60e5430439d381860a725b32f21f9d9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 361, + 555, + 385 + ], + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 385 + ], + "type": "text", + "content": "Table 3. Runtime comparisons, where W4A{8,16,32} is used for QLIP. Measured on the Conceptual Captions dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 395, + 432, + 485 + ], + "blocks": [ + { + "bbox": [ + 315, + 395, + 432, + 485 + ], + "lines": [ + { + "bbox": [ + 315, + 395, + 432, + 485 + ], + "spans": [ + { + "bbox": [ + 315, + 395, + 432, + 485 + ], + "type": "image", + "image_path": "66a7115134cd2c554b04b9f63bf7ba14c45fb8d1169aae7c49db9a7befd6be79.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 485, + 389, + 493 + ], + "lines": [ + { + "bbox": [ + 381, + 485, + 389, + 493 + ], + "spans": [ + { + "bbox": [ + 381, + 485, + 389, + 493 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 433, + 396, + 553, + 484 + ], + "blocks": [ + { + "bbox": [ + 433, + 396, + 553, + 484 + ], + "lines": [ + { + "bbox": [ + 433, + 396, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 433, + 396, + 553, + 484 + ], + "type": "image", + "image_path": "92b289e7a2dc3dab7c5a4ddf27e1f7efff5167b6e9fa91450f34cedcdfc7c06d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 485, + 510, + 493 + ], + "lines": [ + { + "bbox": [ + 501, + 485, + 510, + 493 + ], + "spans": [ + { + "bbox": [ + 501, + 485, + 510, + 493 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 502, + 555, + 525 + ], + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 525 + ], + "type": "text", + "content": "Figure 4. Results on Q-diffusion using different hyperparameter settings of QLIP (W4A{6,8,10}): (a) " + }, + { + "bbox": [ + 313, + 502, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\lambda_{bit}" + }, + { + "bbox": [ + 313, + 502, + 555, + 525 + ], + "type": "text", + "content": " and (b) " + }, + { + "bbox": [ + 313, + 502, + 555, + 525 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 502, + 555, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 544, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 568 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 544, + 555, + 568 + ], + "type": "text", + "content": " was 1000, where we obtained the best scores with " + }, + { + "bbox": [ + 313, + 544, + 555, + 568 + ], + "type": "inline_equation", + "content": "M = 200" + }, + { + "bbox": [ + 313, + 544, + 555, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 576, + 430, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 430, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 430, + 588 + ], + "type": "text", + "content": "4.2. Quantitative results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 689 + ], + "type": "text", + "content": "Comparison with existing quantization methods. To demonstrate the effectiveness and applicability of QLIP, we tested QLIP on two diffusion model quantization methods: Q-diffusion [24] and PTQD [13]. Tables 1 and 2 provide comparison results for the COCO2017 [27] and Conceptual Captions [46] datasets using lightweight diffusion model BK-SDM-Tiny-2M [19] and Stable Diffusion v1.4 [39], respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "As shown in Table 1, for the COCO2017 dataset, QLIP effectively reduced computational complexity of the base" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15431" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 526, + 228 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 526, + 228 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 526, + 228 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 526, + 228 + ], + "type": "image", + "image_path": "3ba0280fd6a05472baa38fa2c4ace5ffcc10e002c2a9b7dcf1f64e80c2e8de31.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 555, + 248 + ], + "type": "text", + "content": "Figure 5. Examples of variations in FAB by QLIP for the texts with different levels of richness and detail, along with the generated images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 83, + 258, + 526, + 434 + ], + "blocks": [ + { + "bbox": [ + 83, + 258, + 526, + 434 + ], + "lines": [ + { + "bbox": [ + 83, + 258, + 526, + 434 + ], + "spans": [ + { + "bbox": [ + 83, + 258, + 526, + 434 + ], + "type": "image", + "image_path": "9452066d6a2a8f82a7aed4389a807bdf6f4ff3ffd6c433e5465dc93f57dcbc78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 443, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 54, + 443, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 443, + 555, + 499 + ], + "type": "text", + "content": "Figure 6. Examples of the bit selection results and generated images. When the text prompt is specifically describing the image to be generated, the predicted image quality is high, and high and medium bits are assigned for most layers and time steps, as shown on the top left. On the other hand, when the text prompt has less detailed information, the predicted image quality is low, and low bits are assigned for many layers and time steps, as shown on the bottom left. On the right side, the generated images using Q-diffusion with and without applying QLIP are shown, demonstrating the effectiveness of QLIP in saving bits while improving the quality of generated images." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 519, + 294, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 519, + 294, + 674 + ], + "spans": [ + { + "bbox": [ + 54, + 519, + 294, + 674 + ], + "type": "text", + "content": "line diffusion quantization methods. For example, Q-diffusion+QLIP (W4A{8,16,32}) achieved a reduction in FAB by 3.86 while maintaining similar FID and sFID, compared to Q-diffusion (W4A16). Furthermore, the use of QLIP consistently led to improvements when applied to PTQD, as demonstrated in Table 1. For example, PTQD+QLIP (W4A{8,16,32}) enhanced FID and sFID by 0.25 and 3.92, respectively, while also reducing FAB and BitOPs. The results on Conceptual Captions followed trends similar to those observed in COCO2017. For example, Q-diffusion+QLIP (W4A{8,16,32}) demonstrated reduced computational complexity compared to Q-diffusion on Conceptual Captions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "Table 2 presents the results of applying lower-bit precision combinations to Stable Diffusion v1.4. Here, PTQD+QLIP (W4A{6,8,10}) achieved a 4.98 improvement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 519, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 628 + ], + "type": "text", + "content": "in FID compared to PTQD (W4A8), while maintaining similar FAB using bit combinations efficiently. Although our proposed QLIP drives the quantization methods to reduce computational complexity and improve image quality, the CLIP Scores remained similar to those of the baseline quantization methods, indicating that the fitness to the text prompts is not significantly influenced by QLIP. Additional results on larger diffusion models can be found in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Runtime measurement. Runtime was measured with Cutlass [34], including the T2Q and Q2B modules, as shown in Table 3. The results show that runtime was reduced to a level similar to W4A8 while maintaining an FID score comparable to W4A16. Measurements were conducted on an RTX 3090. Additional analysis on memory requirements can be found in the supplementary material." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15432" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 84 + ], + "type": "text", + "content": "4.3. Qualitative results" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 220 + ], + "type": "text", + "content": "Figure 5 visualizes the changes in FAB by the richness and detail level of the text, along with the generated images. Here, ChatGPT-4o was used to create texts of varying complexity that describe a given keyword. Specifically, the prompt used was \"Create four sentences that describe how 'keyword' looks like, gradually getting richer and longer.\" As can be seen, QLIP assigned higher bits as the text describes the keyword more specifically and in greater detail, demonstrating that it utilizes information from the text to determine bit precisions for effective synthesis of vivid details and textures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 221, + 296, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 425 + ], + "type": "text", + "content": "Figure 6 shows several results comparing Q-diffusion with and without QLIP when applied to Stable Diffusion v1.4. As can be seen, our QLIP generates more detailed and natural images compared to PTQD by using various bit combinations. Furthermore, Figure 6 visualizes the bit precision maps determined by QLIP. When the T2Q module predicts high image quality from input text prompts, QLIP allocates more bits to enable rendering of vivid and accurate details from rich text prompts. Otherwise, QLIP employs lower bits for layers that are less sensitive to quantization, based on the text prompt and the time step. For example, when given less specific text prompts, the correlation between the generated image and the text prompt is relatively less critical, allowing for low-bit quantization in the cross-attention blocks. These results highlight the effectiveness and significance of dynamically allocating bit precisions according to text prompts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 432, + 154, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 154, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 154, + 444 + ], + "type": "text", + "content": "4.4. Ablation studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 605 + ], + "type": "text", + "content": "Ablation studies on the T2Q module. We report the results on the COCO2017 validation dataset using Stable Diffusion v1.4 with Q-diffusion as the baseline quantization method. Table 4 presents the results of the T2Q modules trained using various perceptual image quality metrics, including Realism score [21], CLIP-IQA [54], and GIQA [11]. The performance of the T2Q module was evaluated using Spearman's rank-order correlation coefficient (SROCC) [49] and Pearson's linear correlation coefficient (PLCC) [36] between the actual quality measures and the predictions of the T2Q module. QLIP with GIQA showed the most substantial improvements in FAB and FID along with the highest SROCC and PLCC scores." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Ablation studies on the Q2B module. We conducted experiments on several variants of the Q2B module using Q-diffusion and Stable Diffusion v1.4 on the COCO2017 validation dataset. As shown in Table 5, when using either " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_q" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " alone or in combination with " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_h^t" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ", effective reduction in FAB was achieved at the expense of increase in FID. When using " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_q" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_m^t" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ", FID decreased while FAB increased. Finally, when " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_q" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_h^t" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_m^t" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " were all involved, balanced results were obtained in two metrics. Implementation details of" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 148 + ], + "type": "table", + "html": "
MethodSROCC↑PLCC↑FAB↓FID↓
Q-diffusion [24]--8.0023.40
+QLIP
w/ Realism score0.51320.50228.1022.18
w/ CLIP-IQA0.71270.70828.5421.81
w/ GIQA0.80470.81087.8621.61
", + "image_path": "0cd2d20b96951a579726cc2e44876115471f56ef637a10687a1c82b00a242583.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 202, + 552, + 281 + ], + "blocks": [ + { + "bbox": [ + 313, + 155, + 553, + 187 + ], + "lines": [ + { + "bbox": [ + 313, + 155, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 553, + 187 + ], + "type": "text", + "content": "Table 4. Ablation studies on the T2Q module: For the bit precision options, W4A8 in Q-diffusion and +QLIP W4A{6,8,10} were used." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 202, + 552, + 281 + ], + "lines": [ + { + "bbox": [ + 317, + 202, + 552, + 281 + ], + "spans": [ + { + "bbox": [ + 317, + 202, + 552, + 281 + ], + "type": "table", + "html": "
MethodBitoptionsFAB↓FID↓
Q-diffusion [24]W4A88.0023.40
+ pqW4A{6,10}7.5726.91
+ pq + ptW4A{6,8,10}6.7329.37
+ pq + pmtW4A{6,8,10}8.6021.96
+ pq + pmt + pth (QLIP)W4A{6,8,10}7.8621.61
", + "image_path": "4ead45ab3e361af5c5db5cb5122037a125bcd30b4fcea725c9d0ffb4c57ed0cc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 350, + 289, + 518, + 300 + ], + "lines": [ + { + "bbox": [ + 350, + 289, + 518, + 300 + ], + "spans": [ + { + "bbox": [ + 350, + 289, + 518, + 300 + ], + "type": "text", + "content": "Table 5. Ablation studies on the Q2B module." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 325, + 553, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 553, + 349 + ], + "type": "text", + "content": "the ablation studies on the Q2B module can be found in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 365, + 388, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 388, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 388, + 377 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 387, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 387, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 312, + 387, + 555, + 566 + ], + "type": "text", + "content": "The proposed quantization method for diffusion models, called QLIP, is the first method that uses text prompts for quantization in diffusion models. From our observation that the bit precision for quantization can be effectively determined based on the quality of the generated images, we designed the T2Q module that predicts image quality from input text prompts. We then introduced the Q2B module that allocates appropriate bit precision for each time step for different layers. By integrating the T2Q and Q2B modules into existing diffusion model quantization methods, we demonstrated the effectiveness of the proposed QLIP. QLIP presents a promising solution for quantizing diffusion models due to its ability to improve the quality of generated images using text prompts while reducing computational costs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 567, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 640 + ], + "type": "text", + "content": "Limitations and future directions. Similar to other generative models, QLIP also has the potential to generate counterfeit images for malicious purposes, potentially undermining trust in visual media. In a future study, we will extend QLIP to consider other input conditions, such as images or segmentation maps, in addition to text prompts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 656, + 411, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 656, + 411, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 411, + 669 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. RS-2022-NR070077)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15433" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. eDiff-I: Text-to-image diffusion models with ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 200 + ], + "type": "text", + "content": "[2] Fan Bao, Chongxuan Li, Jun Zhu, and Bo Zhang. Analytic-DPM: An analytic estimate of the optimal reverse variance in diffusion probabilistic models. In Proceedings of the International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 245 + ], + "type": "text", + "content": "[3] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 294, + 301 + ], + "type": "text", + "content": "[4] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2023.1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 295, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 356 + ], + "type": "text", + "content": "[5] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. PACT: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 413 + ], + "type": "text", + "content": "[6] Jooyoung Choi, Jungbeom Lee, Chaehun Shin, Sungwon Kim, Hyunwoo Kim, and Sungroh Yoon. Perception prioritized training of diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11472-11481, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 414, + 294, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 445 + ], + "type": "text", + "content": "[7] Huanpeng Chu, Wei Wu, Chengjie Zang, and Kun Yuan. QNCD: Quantization noise correction for diffusion models. arXiv preprint arXiv:2403.19140, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 491 + ], + "type": "text", + "content": "[8] Xiangxiang Chu, Shun Lu, Xudong Li, and Bo Zhang. Mix-Path: A unified approach for one-shot neural architecture search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5972–5981, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 294, + 546 + ], + "type": "text", + "content": "[9] Zhen Dong, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. HAWQ: Hessian aware quantization of neural networks with mixed-precision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 293–302, 2019. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 294, + 601 + ], + "type": "text", + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proceedings of the Advances in Neural Information Processing Systems, pages 2672–2680, 2014. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "type": "text", + "content": "[11] Shuyang Gu, Jianmin Bao, Dong Chen, and Fang Wen. GIQA: Generated image quality assessment. In Proceedings of the European Conference on Computer Vision, pages 369–385, 2020. 3, 4, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "text", + "content": "[12] Song Han, Jeff Pool, John Tran, and William Dally. Learning both weights and connections for efficient neural network. In Proceedings of the Advances in Neural Information Processing Systems, pages 1135-1143, 2015. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "text", + "content": "[13] Yefei He, Luping Liu, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. PTQD: Accurate post-training quantization" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "for diffusion models. In Proceedings of the Advances in Neural Information Processing Systems, pages 13237-13249, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 108, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 163 + ], + "type": "text", + "content": "[14] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 165, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 165, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 165, + 553, + 220 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the Advances in Neural Information Processing Systems, pages 6629-6640, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 222, + 553, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 553, + 265 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Proceedings of the Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 268, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 553, + 312 + ], + "type": "text", + "content": "[17] Cheeun Hong and Kyoung Mu Lee. AdaBM: On-the-fly adaptive bit mapping for image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2641-2650, 2024. 1, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 315, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 315, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 315, + 555, + 369 + ], + "type": "text", + "content": "[18] Cheeun Hong, Sungyong Baik, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. CADyQ: Content-aware dynamic quantization for image super-resolution. In Proceedings of the European Conference on Computer Vision, pages 367-383, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "text", + "content": "[19] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. BK-SDM: A lightweight, fast, and cheap version of stable diffusion. In Proceedings of the European Conference on Computer Vision, 2024. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 417, + 553, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 553, + 450 + ], + "type": "text", + "content": "[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proceedings of the International Conference on Learning Representations, 2015. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 453, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 453, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 453, + 553, + 507 + ], + "type": "text", + "content": "[21] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In Proceedings of the Advances in Neural Information Processing Systems, pages 3927-3936, 2019. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 509, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 509, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 317, + 509, + 553, + 553 + ], + "type": "text", + "content": "[22] Hongjae Lee, Jun-Sang Yoo, and Seung-Won Jung. RefQSR: Reference-based quantization for image super-resolution networks. IEEE Transactions on Image Processing, 33: 2823-2834, 2024. 1, 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 555, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 599 + ], + "type": "text", + "content": "[23] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17413-17424, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 601, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 656 + ], + "type": "text", + "content": "[24] Xiuyu Li, Yijiang Liu, Long Lian, Huanrui Yang, Zhen Dong, Daniel Kang, Shanghang Zhang, and Kurt Keutzer. Q-diffusion: Quantizing diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17535-17545, 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[25] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. BRECQ: Pushing the limit of post-training quantization by block reconstruction. In Proceedings of the International Conference on Learning Representations, 2021. 1, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15434" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "text", + "content": "[26] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. SnapFusion: Text-to-image diffusion model on mobile devices within two seconds. In Proceedings of the Advances in Neural Information Processing Systems, pages 20662-20678, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In Proceedings of the European Conference on Computer Vision, pages 740-755, 2014. 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[28] Chenxi Liu, Barret Zoph, Maxim Neumann, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. In Proceedings of the European Conference on Computer Vision, pages 19–34, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 296 + ], + "type": "text", + "content": "[29] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In Proceedings of the International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "text", + "content": "[30] Zhenhua Liu, Yunhe Wang, Kai Han, Siwei Ma, and Wen Gao. Instance-aware dynamic neural network quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12434-12443, 2022. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "text", + "content": "[31] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps. In Proceedings of the Advances in Neural Information Processing Systems, pages 5775-5787, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 410, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 464 + ], + "type": "text", + "content": "[32] Chenlin Meng, Robin Rombach, Ruiqi Gao, Diederik Kingma, Stefano Ermon, Jonathan Ho, and Tim Salimans. On distillation of guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14297-14306, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "type": "text", + "content": "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In Proceedings of the International Conference on Machine Learning, pages 8162-8171, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 511, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 532 + ], + "type": "text", + "content": "[34] NVIDIA. CUTLASS, 2025. https://github.com/NVIDIA/cutlass.7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "text", + "content": "[35] Nilesh Prasad Pandey, Marios Fournarakis, Chirag Patel, and Markus Nagel. Softmax bias correction for quantized generative models. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshop, pages 1453-1458, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "text", + "content": "[36] Karl Pearson. Notes on regression and inheritance in the case of two parents. Proceedings of the Royal Society of London, 58:240-242, 1895. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[37] Dustin Podell, *Zion English*, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. *SDXL: Improving latent diffusion models for high-resolution image synthesis*. In *Proceedings of the International Conference on Learning Representations*, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proceedings of the International Conference on Machine Learning, pages 8748-8763, 2021. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10674-10685, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 239 + ], + "type": "text", + "content": "[40] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22500-22510, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "type": "text", + "content": "[41] Hyogon Ryu, Seohyun Lim, and Hyunjung Shim. Memory-efficient personalization using quantized diffusion model. In Proceedings of the European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 363 + ], + "type": "text", + "content": "[42] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding. In Proceedings of the Advances in Neural Information Processing Systems, pages 36479–36494, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 407 + ], + "type": "text", + "content": "[43] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Proceedings of the International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[44] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. Improved techniques for training gans. In Proceedings of the Advances in Neural Information Processing Systems, pages 2234-2242, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "type": "text", + "content": "[45] Yuzhang Shang, Zhihang Yuan, Bin Xie, Bingzhe Wu, and Yan Yan. Post-training quantization on diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1972-1981, 2023. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 522, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 578 + ], + "type": "text", + "content": "[46] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pages 2556-2565, 2018. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "text", + "content": "[47] Junhyuk So, Jungwon Lee, Daehyun Ahn, Hyungjun Kim, and Eunhyeok Park. Temporal dynamic quantization for diffusion models. In Proceedings of the Advances in Neural Information Processing Systems, pages 48686-48698, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "text", + "content": "[48] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Proceedings of the International Conference on Learning Representations, 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[49] Charles Spearman. The proof and measurement of association between two things. The American Journal of Psychology, 15(1):72--101, 1961. 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15435" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 552 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[50] Yang Sui, Miao Yin, Yi Xie, Huy Phan, Saman Aliari Zonouz, and Bo Yuan. CHIP: Channel independence-based pruning for compact neural networks. In Proceedings of the Advances in Neural Information Processing Systems, pages 24604-24616, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[51] Siao Tang, Xin Wang, Hong Chen, Chaoyu Guan, Zewen Wu, Yansong Tang, and Wenwu Zhu. Post-training quantization with progressive calibration and activation relaxing for text-to-image diffusion models. In Proceedings of the European Conference on Computer Vision, 2024. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 250 + ], + "type": "text", + "content": "[52] Senmao Tian, Ming Lu, Jiaming Liu, Yandong Guo, Yurong Chen, and Shunli Zhang. CABM: Content-aware bit mapping for single image super-resolution network with large input. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1756-1765, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 306 + ], + "type": "text", + "content": "[53] Mart Van Baalen, Christos Louizos, Markus Nagel, Rana Ali Amjad, Ying Wang, Tijmen Blankevoort, and Max Welling. Bayesian bits: Unifying quantization and pruning. In Proceedings of the Advances in Neural Information Processing Systems, pages 5741-5752, 2020. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "text", + "content": "[54] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 407 + ], + "type": "text", + "content": "[55] Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. HAQ: Hardware-aware automated quantization with mixed precision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8612-8620, 2019. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 408, + 294, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 408, + 294, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 294, + 462 + ], + "type": "text", + "content": "[56] Chuanguang Yang, Zhulin An, Helong Zhou, Fuzhen Zhuang, Yongjun Xu, and Qian Zhang. Online knowledge distillation via mutual contrastive learning for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10212-10227, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 464, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 508 + ], + "type": "text", + "content": "[57] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 294, + 552 + ], + "type": "text", + "content": "[58] Xuechen Zhang, Zheng Li, Samet Oymak, and Jiasi Chen. Text-to-3d generative ai on mobile devices: Measurements and optimizations. In Proceedings of the Workshop on Emerging Multimedia Systems, pages 8–14, 2023. 1" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15436" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/f9aefd65-d52a-4731-ba81-d317b6edbcb4_content_list.json b/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/f9aefd65-d52a-4731-ba81-d317b6edbcb4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a11cb77fa285ba0c1ce61fe9a38ebb8fb3f0be0a --- /dev/null +++ b/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/f9aefd65-d52a-4731-ba81-d317b6edbcb4_content_list.json @@ -0,0 +1,1523 @@ +[ + { + "type": "text", + "text": "Text-IRSTD: Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes", + "text_level": 1, + "bbox": [ + 127, + 128, + 870, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feng Huang*, Shuyuan Zheng*, Zhaobing Qiu*,†, Huanxian Liu, Huanxin Bai, Liqiong Chen† Fuzhou University, Fuzhou, China", + "bbox": [ + 127, + 202, + 867, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "huangf@fzu.edu.cn, zhengshuyuan200147@gmail.com, qiuzhaobing@fzu.edu.cn, liuhuanxian10@gmail.com, mw136296@gmail.com, liqionchen@fzu.edu.cn.", + "bbox": [ + 173, + 241, + 810, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 309, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Infrared small target detection is currently a hot and challenging task in computer vision. Existing methods usually focus on mining visual features of targets, which struggles to cope with complex and diverse detection scenarios. The main reason is that infrared small targets have limited image information on their own, thus relying only on visual features fails to discriminate targets and interferences, leading to lower detection performance. To address this issue, we introduce a novel approach leveraging semantic text to guide infrared small target detection, called Text-IRSTD. It innovatively expands classical IRSTD to text-guided IRSTD, providing a new research idea. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. On the other hand, we propose a progressive cross-modal semantic interaction decoder (PCSID) to facilitate information fusion between texts and images. In addition, we construct a new benchmark consisting of 2,755 infrared images of different scenarios with fuzzy semantic textual annotations, called FZDT. Extensive experimental results demonstrate that our method achieves better detection performance and target contour recovery than the state-of-the-art methods. Moreover, proposed Text-IRSTD shows strong generalization and wide application prospects in unseen detection scenarios. The dataset and code are available at https://github.com/Zhengsy0407/Text-IRSTD.", + "bbox": [ + 86, + 340, + 483, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Infrared small target detection (IRSTD) has become an important and challenging issue in computer vision, which is widely used in both civil and military applications [12, 13, 19]. However, due to the long range detection needs, imaged targets usually occupy only a few pixels and lack tex", + "bbox": [ + 89, + 787, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ture details [20, 27, 32]. Therefore, how to accurately detect infrared (IR) small targets from complex backgrounds remains an open issue.", + "bbox": [ + 511, + 310, + 906, + 356 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ebbd5bdf2e26354bb118083726691eb37d4388e61f010ef790af37280950c44f.jpg", + "image_caption": [ + "Figure 1. Visual comparison between the SOTA model and proposed Text-IRSTD. (a) Detection results of the SOTA model [29], which relies only on visual features. (b) Detection results of proposed Text-IRSTD, which uses both textual and visual features." + ], + "image_footnote": [], + "bbox": [ + 514, + 369, + 924, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the early days, many traditional methods emerged to improve detection accuracy. They mainly consist of filter-based methods [11, 23], local saliency-based methods [1, 24, 34] and sparse representation-based methods [8, 14, 25]. However, reliance on artificial feature representations prevents them from further mining the deep semantic information of real targets. In contrast, deep learning (DL)-based methods autonomously learn multi-layered features in a data-driven manner and achieve significant performance improvements. For example, Dai et al. [2] propose asymmetric contextual modulation (ACM) module to better interact with high-level semantics and low-level detail features. For higher feature utilization, Zhang et al. [33] propose attention-guided pyramid context network (AGPC-Net), which mines the correlations between feature map", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10635", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "patches. In addition, researchers have also attempted to fuse prior information for higher detection accuracy [7, 28].", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, limited by scarce image features, existing methods fail to perform satisfactory in complex scenes. As shown in Fig. 1, the SOTA model still suffers from many false detections due to limited ability to discriminate between targets and interference. Recently, with the development of large-scale visual language models [17], semantic text-image-based models [10, 22] have made impressive progress in generic target detection. However, applying them directly to infrared small target detection works poorly. The main reason is that ambiguous target categories in infrared images invalidate existing data annotation and feature fusion methods.", + "bbox": [ + 88, + 121, + 480, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address above challenges, this paper proposes a novel approach leveraging semantic text to guide infrared small target detection, called Text-IRSTD. It innovatively expands classical IRSTD to text-guided IRSTD, providing a new idea for IRSTD. As shown in Fig. 1, by utilizing both text and image information, proposed Text-IRSTD can further suppress clutter interference and improve detection performance. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. This prompt provides more text-image semantic associations that contribute to clutter suppression and target contour recovery. On the other hand, we propose a progressive cross-modal semantic interaction decoder (PCSID) that fuses both text and image information to further improve the detection performance. PCSID consists of two core components: text-guided feature aggregation (TGFA) block and text-guided semantic interaction (TGSI) block. Specifically, TGFA first utilizes text information to guide cross-layer visual feature aggregation for key feature extraction. Where key features (e.g., contrast [24]) refer to those features that are more important than others (e.g., intensity) for IRSTD. Then, TGSI block simultaneously performs feature modulation from two different perspectives, thus realizing a more refined text-image information interaction. In addition, we construct a new benchmark consisting of 2,755 infrared images of different scenarios with fuzzy semantic textual annotations, called FZDT. Extensive experiments on multiple public datasets demonstrate that our approach achieves better detection performance and generalization than representative state-of-the-art algorithms.", + "bbox": [ + 91, + 303, + 483, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 768, + 232, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we first briefly review existing methods for IRSTD. Then, an introduction to text-image models in other domains is presented.", + "bbox": [ + 89, + 792, + 482, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. IRSTD Methods", + "text_level": 1, + "bbox": [ + 89, + 848, + 253, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the early days, many traditional methods emerged to improve detection accuracy. They mainly consist of filter", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "based methods [11, 23], local saliency-based methods [1, 24, 34] and sparse representation-based methods [8, 14, 25]. However, reliance on artificial feature representations prevents them from further mining the deep semantic information of real targets. In contrast, deep learning (DL)-based methods autonomously learn multi-layered features in a data-driven manner and have made some impressive progresses recently. To mitigate feature loss of IR small targets, several researchers focus on improving the cross-layer feature fusion approach. For example, Dai et al. [2] propose an asymmetric contextual modulation (ACM) module to better interact with high-level semantics and low-level detail features. For higher feature utilization, Zhang et al. [33] propose the attention-guided pyramid context network (AGPCNet), which mines the correlations between feature map patches. In order to interact cross-layer features incrementally, Li et al. [9] propose dense nested attention network (DNA-Net). Several researchers also attempted to improve the model structure. For example, Wu et al. [21] embed a smaller U-Net into the U-Net backbone to learn target features from multiple levels and scales. In addition, Hou et al. [7] try to introduce prior information and design a robust infrared small target detection network (RISTDnet) combining artificial features and neural networks.", + "bbox": [ + 511, + 90, + 903, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods have made many impressive progress in IRSTD. However, the visual features of IR small targets are quite limited, which renders their performance unsatisfactory in complex and diverse detection scenarios. In contrast, we explore a new research idea to introduce text semantic features into IRSTD and devise a novel text-image fusion detection method. Utilizing added textual information, proposed Text-IRSTD can further suppress background interference and improve detection performance.", + "bbox": [ + 511, + 455, + 906, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Text-Image Models", + "text_level": 1, + "bbox": [ + 511, + 604, + 697, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the advancement of deep learning techniques and the support of large datasets, text-image models has achieved success. In particular, CLIP [17] utilizes 400 million matched pairs datasets and contrast learning techniques to demonstrate powerful performance and wide application prospects in multiple domains and tasks. CLIP significantly improves text-image alignment and avoids the reliance on explicit labeling. With the support of CLIP models, text-driven generic target detection has also developed rapidly.", + "bbox": [ + 511, + 627, + 903, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For better detection performance, RegionCLIP [35] leverages CLIP to match image regions with template captions, achieving fine-grained alignment between image and text features. ViLD [5] designs a training method based on knowledge distillation, which distills the knowledge from a pretrained open-vocabulary image classification model into a two-stage detector. CORA [22] mitigates the distribution gap between overall image features and regional features when using CLIP directly for detector training by introduc", + "bbox": [ + 511, + 763, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10636", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing learnable prompts.", + "bbox": [ + 89, + 90, + 243, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although existing text-image models for generic target detection are illuminating, they are difficult to apply to IRSTD tasks. The reason is that it is difficult to establish appropriate text-image semantic associations and interactions for IR small targets with ambiguous categories. Therefore, it is important and challenging to investigate text-image detection models dedicated to IRSTD tasks.", + "bbox": [ + 89, + 109, + 483, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Fuzzy Semantic Text Prompt For IRSTD", + "text_level": 1, + "bbox": [ + 89, + 234, + 457, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic text prompts aim to establish a proper matching relationship between text and images, which is crucial for cross-modal target detection. In generic target detection [5, 35], researchers usually directly describe specific categories of targets, such as \"A photo of a girl\" and \"A photo of a soccer\" in Fig. 2(a). In this way, it is possible to focus not only on the visual features of a single target, but also to further explore semantic relations between multiple targets, e.g. a small round target near a girl is more likely to be a soccer. However, IR small targets typically occupy only a few pixels and lack visual features such as texture. As shown in Fig. 2(b), real targets are usually presented as bright spots that cannot be further identified as a specific category such as an aircraft or a bird. Therefore, using existing text prompts to describe ambiguous IR small target categories will lead to incorrect text-image matching relationships, impairing detection accuracy and model generalizability.", + "bbox": [ + 89, + 263, + 483, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this issue, we propose a novel fuzzy semantic text prompt for IRSTD, which includes two templates: scenario description and task description. Specifically, for known scenes with semantic information, we use the scenario description and describe one IR image as “A photo of a [Interested Region] target in the [Scene] background”, where [Interested Region] denotes the region where targets need to be detected, e.g., sky target and ground target, and [Scene] denotes the detection scene, e.g., sky-ground background, ocean-sky background. For example, we describe the IR image Fig. 2(b) as: “A photo of a sky target in the sky and ground background”, which not only establishes correct and robust matching relationships, but also provides more semantic associations. In this way, the model can focus on more shared visual features of sky targets, such as shape contours, which helps target shape recovery and detection. Moreover, as shown in Fig. 1, more semantic associations are established and utilized to further suppress complex ground interference and improve detection accuracy. For unseen scenes without semantic information, we use the task description, i.e., the default text prompt is “A photo of a small target in the background”.", + "bbox": [ + 89, + 537, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3c93771244c902d6f7ce0c8add431b5339216eedb7f9ecd02a04f93bbdb5de0a.jpg", + "image_caption": [ + "(a) Typical prompt template for object detection" + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 924, + 191 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/baaf457a050d1fa96472c3237b42a17d6217b0aa5abb75d08af7403f2b6bd15a.jpg", + "image_caption": [ + "(b) Proposed fuzzy semantic prompt for IRSTD", + "Figure 2. (a) Typical semantic text prompt for generic target detection. (b) Proposed fuzzy semantic text prompt for IRSTD." + ], + "image_footnote": [], + "bbox": [ + 516, + 209, + 921, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 404, + 604, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first present the overall architecture of proposed Text-IRSTD. Then, Sections 4.2 and 4.3 introduce two core components: TGFA block for aggregating key visual features and TGSI block for facilitating text-image information interaction, respectively.", + "bbox": [ + 511, + 430, + 906, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 511, + 513, + 709, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 3 illustrates the overall structure of Text-IRSTD. First, one infrared image $I_{img}$ and the corresponding text prompt $I_{text}$ are fed into the image encoder $\\mathcal{F}_i$ and text encoder $\\mathcal{F}_t$ , respectively, as follows:", + "bbox": [ + 511, + 537, + 906, + 597 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E _ {i m g} = \\mathcal {F} _ {i} (I _ {i m g}), \\\\ \\Gamma_ {i m g} = \\mathcal {F} _ {i} (I _ {i m g}). \\end{array} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 606, + 903, + 630 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE _ {t e x t} = \\mathcal {F} _ {t} (I _ {t e x t}),\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 627, + 772, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $E_{img}$ and $E_{text}$ denote the image and text encoder outputs respectively. Note that $\\mathcal{F}_i$ consists of four groups of ResNeSt [30] blocks, while $\\mathcal{F}_t$ denote the pre-trained language model CLIP with frozen weights.", + "bbox": [ + 511, + 651, + 906, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, $E_{img}$ and $E_{text}$ are fed into PCSID for text-image feature interaction and enhancement. The result of cross-modal interaction $D_{cm}$ can be expressed as:", + "bbox": [ + 511, + 710, + 905, + 756 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {c m} = \\operatorname {P C S I D} \\left(E _ {i m g}, E _ {t e x t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 767, + 903, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where PCSID consists of three groups of TGFA blocks and TGSI blocks in series. TGFA and TGSI will be introduced in Sections 4.2 and 4.3, respectively. Note that the fourth layer output $E_{\\mathrm{img}}^{(4)}$ of $\\mathcal{F}_i$ is enhanced by detail perception module (DPM) and then fed to PCSID.", + "bbox": [ + 511, + 792, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we apply a $1 \\times 1$ convolution $f_{c}^{1}$ and sigmoid function to the decoder outputs $D_{cm}$ . Thus, the cross-modal", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10637", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2558a3d70209199205f0131ab5cf4506674ec914a6dc7c4a69d479ada9ea695b.jpg", + "image_caption": [ + "Figure 3. Overview of the proposed Text-IRSTD. It consists of three main components: text encoder, image encoder, and cross-modal decoder PCSID with their outputs represented as $E_{text}$ , $E_{img}^{(i)}$ , and $D_{cm}$ , respectively. Where DPM denotes detail perception module, TGFA Block denotes text-guided feature aggregation block and TGSI Block denotes text-guided semantic interaction block." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 911, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "detection result $O_{cm}$ can be obtained as follows:", + "bbox": [ + 89, + 395, + 413, + 411 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nO _ {c m} = \\operatorname {S i g m o i d} \\left(f _ {c} ^ {1} \\left(D _ {c m}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 422, + 482, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Text-guided Feature Aggregation Block", + "text_level": 1, + "bbox": [ + 89, + 454, + 429, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Effective aggregation of visual features is crucial for detection models. Limited by scarce visual information, existing methods struggle to capture key features from high-level and low-level feature maps. To this end, we propose the novel text-guided cross-layer feature aggregation block TGFA (shown in Fig. 4). TGFA transforms text features $E_{text}$ into modulation information through MLP, guiding the model to capture key features that are easily neglected in complex scenes. The details are as follows:", + "bbox": [ + 89, + 477, + 483, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {t, 1}, F _ {t, 2} = \\operatorname {S p i l t} \\left(\\operatorname {M L P} \\left(E _ {\\text {t e x t}}\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 627, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {f} ^ {l} = F _ {l} \\otimes F _ {t, 1}, F _ {f} ^ {h} = F _ {h} \\otimes F _ {t, 2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 657, + 482, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{l}$ and $F_{h}$ denote low-level feature and high-level feature, $F_{f}^{l}$ and $F_{f}^{h}$ denote the fusion features after text modulation, $\\mathrm{Spit}(\\cdot)$ denotes splitting the feature into two equal groups along the channel dimension, $\\otimes$ denotes element-wise multiplication.", + "bbox": [ + 89, + 686, + 483, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Low-level features contain rich target locations and details, but are easily lost at deep layers because small IR targets occupy only a few pixels. Therefore, we devise the detail perception module DPM as shown in Fig. 4. Specifically, for a given low-level fusion feature $F_{f}^{l}$ , a $1 \\times 1$ convolution is first applied to expand the channel dimension by a factor of $\\mu$ , and then divide the resulting feature map into two groups $F_{l,1} \\in \\mathbb{R}^{\\frac{\\mu C}{2} \\times H \\times W}$ and $F_{l,2} \\in \\mathbb{R}^{\\frac{\\mu C}{2} \\times H \\times W}$ along the channel dimension.", + "bbox": [ + 89, + 762, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/70253ee2c60fa7c4fa46f28876b041049a1cf2953ca66ed2da13be1d445dffb2.jpg", + "image_caption": [ + "Figure 4. Structure of TGFA block, where DPM denotes detail perception module and CA denotes channel attention." + ], + "image_footnote": [], + "bbox": [ + 516, + 397, + 908, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, we use $3 \\times 3$ and $5 \\times 5$ depth-wise convolutions to capture important local features such as contrast at different scales and concatenate these two groups of features along the channel dimension. The multi-scale detailed feature mapping $F_{ms}$ is obtained from the following equation:", + "bbox": [ + 511, + 609, + 906, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {l, 1}, F _ {l, 2} = \\operatorname {S p l i t} \\left(f _ {c} ^ {1} \\left(F _ {f} ^ {l}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 696, + 903, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {m s} = f _ {c} ^ {1} \\left(\\operatorname {C o n t a c t} \\left(f _ {d w c} ^ {3} \\left(F _ {l, 1}\\right), f _ {d w c} ^ {5} \\left(F _ {l, 2}\\right)\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 726, + 903, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{Contact}(\\cdot)$ denotes concatenating the two groups of features along the channel dimension, $f_{c}^{1}$ denotes $1 \\times 1$ convolution, $f_{dwc}^{3}$ and $f_{dwc}^{5}$ denote $3 \\times 3$ and $5 \\times 5$ depth-wise convolutions.", + "bbox": [ + 511, + 751, + 906, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the multi-scale detailed feature mapping $F_{ms}$ , we use point-wise convolution to aggregate the contextual features at each spatial location. For simplicity, it can be described as:", + "bbox": [ + 511, + 811, + 906, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {D P} = \\sigma \\left(\\operatorname {P W C o n v} _ {2} \\left(\\gamma \\left(\\operatorname {P W C o n v} _ {1} \\left(F _ {m s}\\right)\\right)\\right) \\otimes F _ {m s}, \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 883, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10638", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{PWConv}_1$ and $\\mathrm{PWConv}_2$ denote the point-wise convolution with kernel sizes of $\\frac{C}{4} \\times C \\times 1 \\times 1$ and $C \\times \\frac{C}{4} \\times 1 \\times 1$ , $\\sigma$ and $\\gamma$ represent the sigmoid activation function and rectified linear unit.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, high-level features can provide more accurate semantic information, while adjusting the weights of different channels can adaptively enhance the representation of key information and suppress redundant features. Therefore, for a given high-level fusion features $F_{f}^{h}$ , we first use global average pooling on each channel to obtain the global feature representation $z_{c} \\in \\mathbb{R}^{C \\times 1 \\times 1}$ , which can be described as:", + "bbox": [ + 89, + 152, + 482, + 271 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {c} = \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} F _ {f} ^ {h} [:, i, j ]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 282, + 482, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, a nonlinear mapping of this global information is performed through two fully connected layers to establish dependencies between the channels.", + "bbox": [ + 89, + 337, + 482, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the channel feature maps are normalized using the sigmoid function, and an element-wise multiplication is performed with the $F_{f}^{h}$ , the output of CA is defined as:", + "bbox": [ + 89, + 383, + 483, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {C A} = \\sigma \\left(F c _ {2} \\left(F c _ {1} \\left(z _ {c}\\right)\\right)\\right) \\otimes F _ {f} ^ {h}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 443, + 482, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Fc_{1} \\in \\mathbb{R}^{\\frac{C}{r} \\times C}$ and $Fc_{2} \\in \\mathbb{R}^{C \\times \\frac{C}{r}}$ represent the two fully connected operations, $r$ denotes the channel reduction rate.", + "bbox": [ + 89, + 474, + 482, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Text-guided Semantic Interaction Block", + "text_level": 1, + "bbox": [ + 89, + 534, + 431, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the limited visual information, existing models struggle to effectively suppress complex interference and recover target contours. Therefore, we introduce semantic text prompts and propose TGSI block, which achieves better detection performance through text-image information interaction. Figure 3 shows the detailed structure of TGSI block. First, we take the image feature $M$ and the text feature $E_{text}$ as inputs, where $E_{text}$ is obtained by feeding the text prompt $T_{text}$ into the CLIP. Notably, we freeze the weights of the text encoder to preserve the original capabilities of the pre-trained model.", + "bbox": [ + 89, + 556, + 482, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, we design a Feed-forward Network (FFN) to further explore the deep semantic information embedded in the text features. This helps the model capture more fine-grained semantic associations between the text, the target, and the regions of interest. Specifically, the text feature $E_{text}$ is mapped into two learnable semantic parameters $\\alpha_{t}$ and $\\beta_{t}$ after going through the FFN, which can be defined as:", + "bbox": [ + 89, + 723, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {t}, \\beta_ {t} = \\operatorname {F F N} \\left(E _ {\\text {t e x t}}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 845, + 482, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, the semantic parameters $\\alpha_{t}$ and $\\beta_{t}$ modulate the image feature $M$ from the perspective of gain control and", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "offset adjustment, respectively. Specifically, the gain parameter $\\alpha_{t}$ directs the network to pay more attention to the regions of interest of the text prompt and reduces false alarms in other regions. The offset parameter $\\beta_{t}$ provides finer tuning and helps in the recovery of target contour details. The fused semantic features $\\hat{M}$ are obtained by:", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} = \\left(1 + \\alpha_ {t}\\right) \\otimes M + \\beta_ {t}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 194, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 224, + 645, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 251, + 799, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: All experiments are conducted on proposed FZDT dataset, which contains three public image datasets NUAA-SIRST [2], NUDT-SIRST [9], IRSTD-1k [32], and corresponding fuzzy semantic text prompts. Following existing works [26], each dataset is divided into training set, validation set and test set in the ratio of $6:2:2$ .", + "bbox": [ + 511, + 273, + 903, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics: For comprehensive algorithmic comparison, the evaluation metrics include intersection over union (IoU), false alarm rate $(\\mathrm{F}_a)$ , F-measure $(\\mathrm{F}_1)$ , and probability of detection $(\\mathrm{P}_d)$ [29].", + "bbox": [ + 511, + 364, + 903, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 436, + 730, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the efficacy of Text-IRSTD, we use the common CLIP-ViT-B/32 [17] as the text encoder. Our network is implemented with Pytorch on a single NVIDIA GeForce RTX 3090 GPU. We adopt the AdamW optimizer with a learning rate of 0.0001 and a weight decay of 0.05. Every images are normalized and randomly cropped to $256 \\times 256$ . The batch size and epoch are set as 8 and 500, respectively.", + "bbox": [ + 511, + 458, + 903, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The comparison algorithms for experiments include eight state-of-the-art DL-based methods: ACMNet [2], ALCNet [3], AGPCNet [33], DNANet [9], UIUNet [21], RDIAN [18], DMFNet [6] and SCTrans [29], and five recent traditional methods: IPI [4], PSTNN [31], FKRW [15], GSWLCM [16] and RUMFR [1]. To further illustrate the advancement of our method, Text-IRSTD without text prompts is also selected as the baseline method.", + "bbox": [ + 511, + 565, + 905, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3.Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 696, + 705, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The quantitative results of different methods are presented in Tab. 1. Overall, our Text-IRSTD achieves the best performance across all evaluation metrics. For example, the $\\mathrm{F}_a$ of our method achieves as low as $1.032 \\times 10^{-6}$ on NUDT-SIRST, while also reaching the highest $\\mathrm{P}_d$ of $99.73\\%$ , demonstrating our superior capability in suppressing complex interference. Additionally, we obtain the best IoU of $95.84\\%$ , which is $1.01\\%$ higher than the second best, indicating better target contour recovery capability. Traditional methods have lower metrics and exhibit poor performance. DL-based methods can further mine deep semantic features in a data-driven manner, and thus have significant", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10639", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2a2e6c604f3190f9337dc6336cc9180b29cf734b2a0840cc3f683ef913262059.jpg", + "table_caption": [ + "Table 1. Quantitative results of different methods. Results for the metrics of IoU(%, Pd(%) F, $(10^{-6})$ and $\\mathrm{F_1}(\\%)$ are presented. The best values are highlighted with bold, the second best values are underlined." + ], + "table_footnote": [], + "table_body": "
MethodNUDT-SIRSTIRSTD-1k
IoUPdFaF1IoUPdFaF1
IPI [4]38.1590.78369.961.5230.9578.91180.455.15
PSTNN [31]17.2154.4781.6229.4525.5862.9678.9141.35
FKRW [15]25.3860.26107.538.5415.8954.5426.7527.08
GSWLCM [16]4.90965.0015.719.3013.06859.2526.946.045
RUMFR [1]26.0867.8996.4239.2510.2653.8725.2918.59
ACMNet [2]68.1496.5717.3883.4161.1487.8841.7974.17
ALCNet [3]72.5696.057.57286.7858.2789.9044.6972.92
AGPCNet [33]85.9797.637.22893.5162.5191.5820.9777.68
DNANet [9]94.7098.683.21297.3164.5190.2318.2778.77
UIUNet [21]89.1597.115.50694.2364.3290.2428.3379.02
RDIAN [18]86.0898.156.71192.3661.8688.5641.4476.27
DMFNet [6]87.6798.422.29493.3064.4590.1418.2478.52
SCTrans [29]94.8398.422.06597.1962.9592.2423.4078.04
Text-IRSTD (w/o text)95.2598.941.66397.3765.5091.5616.1378.73
Text-IRSTD (Ours)95.8499.731.03297.9569.5792.5914.9779.24
", + "bbox": [ + 91, + 121, + 906, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "improvements in all metrics. However, the limited visual information of the target makes it difficult to further improve their detection accuracy in complex scenes. For instance, these SOTA DL-based methods achieve a maximum IoU of only $64.51\\%$ on the IRSTD-1k, which is difficult to satisfy higher detection requirements. By incorporating text information and designing cross-modal information interaction methods, the proposed Text-IRSTD achieves up to $69.57\\%$ performance improvement in IoU.", + "bbox": [ + 88, + 393, + 480, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It is worth noting that our method is able to produce satisfactory results even when using the default text input ('w/o text') in the testing stage. This phenomenon indicates that our method does not rely completely on detailed text prompts, and the inherent advantages of the model design contribute to the performance of IRSTD.", + "bbox": [ + 89, + 531, + 482, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 633, + 272, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We visualize the detection results of six representative algorithms, as shown in Fig. 5 and Fig. 6. It can be seen that traditional algorithms (e.g., PSTNN) have a large number of missed detections and false alarms due to their limited artificial feature representation capabilities. DL-based methods have made significant progress in background suppression by mining high-level semantic features. However, they all showed false alarms in the third row of images in Fig. 5. The reason is that the limited visual information of infrared images makes it difficult to further suppress interference and improve detection performance.", + "bbox": [ + 89, + 657, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast, proposed Text-IRSTD accurately detects targets in every scene and suppresses almost all background interference. By introducing semantic text to expand the limited visual information, Text-IRSTD can further improve the detection accuracy in complex scenes. As shown", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in Fig. 5, thanks to the PCSID, we realize high-performance interactions across modal information, which can focus more on the region of interest to suppress complex interferences, while perceive and enhance the detail information to guide the accurate detection and contour recovery of targets.", + "bbox": [ + 511, + 393, + 906, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.5. Performance in Unseen Scenarios", + "text_level": 1, + "bbox": [ + 511, + 500, + 805, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In practical applications, models often need to face unseen detection scenarios. The detection performance in unseen scenarios can evaluate the practicality and generalization of models to some extent. To further validate the advancement of proposed methods, we compare the detection capabilities of different methods in unseen scenarios.", + "bbox": [ + 511, + 523, + 905, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we train different models using the NUDT-SIRST dataset and then test their detection performance using the unseen NUAA-SIRST dataset. Tab. 2 presents the detection results of different models on the unseen dataset. It can be seen that our method outperforms the state-of-the-art models in terms of IoU, $\\mathrm{P}_d$ and $\\mathrm{F}_a$ , implying better detection accuracy and scene generalization.", + "bbox": [ + 511, + 614, + 905, + 722 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0420314caad6bea81d9dbe2a48203e725bb648f4be90fb143b2e0e94a0a2e94e.jpg", + "table_caption": [ + "Table 2. Detection performance of different methods for the unseen dataset." + ], + "table_footnote": [], + "table_body": "
MethodNUDT-SIRST (Train)
NUAA-SIRST (Test)
IoU(%)Pd(%)Fa(10-6)
DNANet [9]35.8076.14141.3
UIUNet [21]53.3280.7335.15
RDIAN [18]39.3577.9843.68
DMFNet [6]43.7079.8160.07
SCTrans [29]50.9783.4837.71
Text-IRSTD (Ours)54.5187.1529.86
", + "bbox": [ + 519, + 771, + 901, + 897 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10640", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c7f75ae26b84ebd43b3c8a5af085f4cc2386cb5095160628f0e017f5d7e532e6.jpg", + "image_caption": [ + "Figure 5. Visual results of different IRSTD methods. The enlarged view of the target is shown in image corners. The red, blue, and yellow boxes represent correctly detected targets, missed targets, and false detections, respectively." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.6. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 554, + 243, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Text Prompt: In order to verify the effectiveness of fuzzy text prompts, we designed three textual prompt variants for experiments and the results are shown in Tab. 3. It can be seen that the use of [Interested Region] target as a prompt to describe the IR small target helps to avoid the wrong text-image matching relationship caused by the use of explicit categories of prompts. Meanwhile, the description of [Scene] background further improves the detection performance in complex scenes by establishing more appropriate semantic associations. More importantly, experiments demonstrate the possibility of improving IRSTD performance with text and provide an effective technical path for further research.", + "bbox": [ + 89, + 579, + 482, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of PCSID: To validate the effectiveness of each component in PCSID, we perform several ablation studies on the IRSTD-1k, and the results are shown in Tab. 4. The results of the experiment indicate that both TGSI block and TGFA block improve the performance of the baseline model, and when use together, the model achieve the best results. This further confirms the importance of their syn", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ergistic effect in enhancing the performance of small target detection.", + "bbox": [ + 511, + 554, + 906, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of TGFA: To demonstrate the effectiveness of TGFA block, we introduce three network variants for comparison, as shown in Tab. 5. It can be seen that proposed TGFA block achieves the best performance. The reason is that text features as modulation information can optimize the complementarity and interaction between cross-layer features. Moreover, DPM and CA can better perceive and enhance important visual features, such as contrast, which facilitate visual feature fusion and subsequent cross-modal information interaction.", + "bbox": [ + 511, + 590, + 908, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of TGSI Block: We use four different text-image feature interaction methods to train the Text-IRSTD. As shown in Tab. 6, simply concatenating text and image features along the channel dimension ignores semantic inconsistencies across modalities, leading to inadequate semantic alignment. Without gain parameter $\\alpha_{t}$ , the model cannot dynamically focus on the regions of interest indicated by the text prompt. Similarly, without offset parameter $\\beta_{t}$ , the model cannot make fine adjustments to the image features using the text features, thus limiting the effective", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10641", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bf04e86d4cd1b77d4130bdad29c70bbfcf1e731f615c910984f5314bbf04bf1.jpg", + "image_caption": [ + "Figure 6. 3D visualization results corresponding to those in Fig. 5. Different colors represent distinct grayscale values, gradually increasing from blue to yellow." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 906, + 400 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1e92defea8b0ad4d3553e26ec0b3c7f37dce7e9fd4f552350f461d7b25eb536f.jpg", + "table_caption": [ + "Table 3. Ablation study of different prompts." + ], + "table_footnote": [], + "table_body": "
TypePromptIoU(%)Pd(%)Fa(10-6)
Conventional promptsA photo of a [class]62.2290.5830.36
Variants of conventionA photo of a [class] in the [Scene] background66.6589.8921.69
Variants of OursA [Interested Region] target in the [Scene] background69.0191.6818.86
OursA photo of a [Interested Region] target in the [Scene] background69.5792.5914.97
", + "bbox": [ + 94, + 467, + 906, + 550 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4a57e05157528cdff16b095dd2342dab25b33c1ec6de555a16418520f824475e.jpg", + "table_caption": [ + "Table 4. Ablation study of the TGFA block and TGSI block." + ], + "table_footnote": [], + "table_body": "
BaselineTGFATGSIIoU(%)Pd(%)Fa(10-6)
✓61.6789.2257.97
✓✓67.7591.8719.83
✓✓66.1991.9226.39
✓✓✓69.5792.5914.97
", + "bbox": [ + 91, + 575, + 482, + 656 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ff17e5dd11c04c6728edd2866bf21800e0c6dcb6cb3d8eac8f18597e076d950c.jpg", + "table_caption": [ + "Table 5. Ablation study of the TGFA block." + ], + "table_footnote": [], + "table_body": "
ModelIoU(%)Pd(%)Fa(10-6)
TGFA w/o text guide68.7592.0223.68
TGFA w/o CA68.0191.2426.24
TGFA w/o DPM67.1689.5617.95
TGFA (Ours)69.5792.5914.97
", + "bbox": [ + 91, + 680, + 488, + 762 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6615fb772fd786c8da72dbdf6494bdc45c5f894a8380b58eb08a36455cf71ece.jpg", + "table_caption": [ + "Table 6. Ablation study of the TGSI block." + ], + "table_footnote": [], + "table_body": "
ModelIoU(%)Pd(%)Fa(10-6)
Concatenation67.7390.5729.49
TGSI w/o αt66.6590.1335.85
TGSI w/o βt68.0591.2116.22
TGSI (Ours)69.5792.5914.97
", + "bbox": [ + 91, + 785, + 483, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ness of target shape reconstruction. By combining both, the", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "proposed TGSI Block achieves the best performance.", + "bbox": [ + 511, + 559, + 864, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 599, + 633, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we extend the IRSTD task and propose a text-guided IRSTD framework to address the issue that current models struggle to further improve detection accuracy by relying only on limited visual information. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. This prompt provides more text-image semantic associations that contribute to clutter suppression and target contour recovery. On the other hand, we propose PCSID to facilitate the information interaction between text and images. Moreover, we have constructed the first text-image bimodal dataset, called FZDT, to provide the basis for subsequent text-guided IRSTD research. Extensive experimental results demonstrate that our method achieves better detection performance and target contour recovery than the state-of-the-art methods. More surprisingly, the proposed Text-IRSTD demonstrates strong generalization and wide application prospects in unseen detection scenarios.", + "bbox": [ + 509, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10642", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62405060, in part by the Nature Science Foundation of Fujian Province under Grant 2022J05113 and 2024J01245, and in part by the Educational Research Program for Young and Middle-aged Teachers of Fujian Province under Grant JAT210035.", + "bbox": [ + 89, + 113, + 483, + 199 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 224, + 187, + 242 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Liqiong Chen, Tong Wu, Shuyuan Zheng, Zhaobing Qiu, and Feng Huang. Robust unsupervised multifeature representation for infrared small target detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:10306-10323, 2024. 1, 2, 5, 6", + "[2] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Asymmetric contextual modulation for infrared small target detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 950-959, 2021. 1, 2, 5, 6", + "[3] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Attentional local contrast networks for infrared small target detection. IEEE transactions on geoscience and remote sensing, 59(11):9813-9824, 2021. 5, 6", + "[4] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE transactions on image processing, 22(12):4996-5009, 2013. 5, 6", + "[5] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2, 3", + "[6] Tan Guo, Baojiang Zhou, Fulin Luo, Lei Zhang, and Xinbo Gao. Dmfnet: Dual-encoder multi-stage feature fusion network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-14, 2024. 5, 6", + "[7] Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Ye Zhao, Hao-liang Zheng, and Wei Zhang. Ristdnet: Robust infrared small target detection network. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2021. 2", + "[8] Xuan Kong, Chunping Yang, Siying Cao, Chaohai Li, and Zhenming Peng. Infrared small target detection via nonconvex tensor fibered rank approximation. IEEE Transactions on Geoscience and Remote Sensing, 60:1-21, 2022. 1, 2", + "[9] Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo. Dense nested attention network for infrared small target detection. IEEE Transactions on Image Processing, 32:1745-1758, 2022. 2, 5, 6", + "[10] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2" + ], + "bbox": [ + 93, + 251, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Yongsong Li, Zhengzhou Li, Yu Shen, and Jie Li. Infrared small target detection based on 1-d difference of guided filtering. IEEE Geoscience and Remote Sensing Letters, 20: 1-5, 2023. 1, 2", + "[12] Fanzhao Lin, Kexin Bao, Yong Li, Dan Zeng, and Shiming Ge. Learning contrast-enhanced shape-biased representations for infrared small target detection. IEEE Transactions on Image Processing, 33:3047-3058, 2024. 1", + "[13] Qiankun Liu, Rui Liu, Bolun Zheng, Hongkui Wang, and Ying Fu. Infrared small target detection with scale and location sensitivity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17490-17499, 2024. 1", + "[14] Yuan Luo, Xiaorun Li, Shuhan Chen, Chaoqun Xia, and Liaoying Zhao. Imnn-lwec: A novel infrared small target detection based on spatial-temporal tensor model. IEEE Transactions on Geoscience and Remote Sensing, 60:1-22, 2022. 1, 2", + "[15] Yao Qin, Lorenzo Bruzzone, Chengqiang Gao, and Biao Li. Infrared small target detection based on facet kernel and random walker. IEEE Transactions on Geoscience and Remote Sensing, 57(9):7104-7118, 2019. 5, 6", + "[16] Zhaobing Qiu, Yong Ma, Fan Fan, Jun Huang, and Lang Wu. Global sparsity-weighted local contrast measure for infrared small target detection. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 5, 6", + "[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5", + "[18] Heng Sun, Junxiang Bai, Fan Yang, and Xiangzhi Bai. Receptive-field and direction induced attention network for infrared dim small target detection with a large-scale dataset irdst. IEEE Transactions on Geoscience and Remote Sensing, 61:1-13, 2023. 5, 6", + "[19] Xiaozhong Tong, Zhen Zuo, Shaojing Su, Junyu Wei, Xiaoyong Sun, Peng Wu, and Zongqing Zhao. St-trans: Spatial-temporal transformer for infrared small target detection in sequential images. IEEE Transactions on Geoscience and Remote Sensing, 62:1-19, 2024. 1", + "[20] Keyan Wang, Xueyan Wu, Peicheng Zhou, Zuntian Chen, Rui Zhang, Liyun Yang, and Yunsong Li. Afe-net: Attention-guided feature enhancement network for infrared small target detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:4208-4221, 2024. 1", + "[21] Xin Wu, Danfeng Hong, and Jocelyn Chanussot. Uiu-net: U-net in u-net for infrared small object detection. IEEE Transactions on Image Processing, 32:364-376, 2022. 2, 5, 6", + "[22] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7031-7040, 2023. 2", + "[23] Yunhong Xin, Jiao Zhou, and Yishuan Chen. Dual multiscale filter with SSS and GW for infrared small target detec" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10643", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. Infrared Physics & Technology, 81:97-108, 2017. 1, 2", + "[24] Yunkai Xu, Minjie Wan, Xiaojie Zhang, Jian Wu, Yili Chen, Qian Chen, and Guohua Gu. Infrared small target detection based on local contrast-weighted multidirectional derivative. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-16, 2023. 1, 2", + "[25]Fuju Yan,Guili Xu,Junpu Wang,Quan Wu,and Zhengsheng Wang. Infrared small target detection via schatten capped pnorm-based non-convex tensor low-rank approximation.IEEE Geoscience and Remote Sensing Letters,20: 1-5,2023.1,2", + "[26] Bo Yang, Xinyu Zhang, Jian Zhang, Jun Luo, Mingliang Zhou, and Yangjun Pi. Eflnet: Enhancing feature learning network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-11, 2024. 5", + "[27] Xinyi Ying, Li Liu, Yingqian Wang, Ruojing Li, Nuo Chen, Zaiping Lin, Weidong Sheng, and Shilin Zhou. Mapping degeneration meets label evolution: Learning infrared small target detection with single point supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15528-15538, 2023. 1", + "[28] Chuang Yu, Yunpeng Liu, Shuhang Wu, Xin Xia, Zhuhua Hu, Deyan Lan, and Xin Liu. Pay attention to local contrast learning networks for infrared small target detection. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 2", + "[29] Shuai Yuan, Hanlin Qin, Xiang Yan, Naveed Akhtar, and Ajmal Mian. Sctransnet: Spatial-channel cross transformer network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-15, 2024. 1, 5, 6", + "[30] Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R Manmatha, et al. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2736–2746, 2022. 3", + "[31] Landan Zhang and Zhenming Peng. Infrared small target detection based on partial sum of the tensor nuclear norm. Remote Sensing, 11(4):382, 2019. 5, 6", + "[32] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 877-886, 2022. 1, 5", + "[33] Tianfang Zhang, Lei Li, Siying Cao, Tian Pu, and Zhenming Peng. Attention-guided pyramid context networks for detecting infrared small target under complex background. IEEE Transactions on Aerospace and Electronic Systems, 59(4): 4250-4261, 2023. 1, 2, 5, 6", + "[34] Xiangyue Zhang, Jingyu Ru, and Chengdong Wu. Infrared small target detection based on gradient correlation filtering and contrast measurement. IEEE Transactions on Geoscience and Remote Sensing, 61:1-12, 2023. 1, 2", + "[35] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16793-16803, 2022. 2, 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10644", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/f9aefd65-d52a-4731-ba81-d317b6edbcb4_model.json b/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/f9aefd65-d52a-4731-ba81-d317b6edbcb4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fa08bcf3f537de9931150341f3bbc4b00df3e1fc --- /dev/null +++ b/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/f9aefd65-d52a-4731-ba81-d317b6edbcb4_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.13, + 0.872, + 0.177 + ], + "angle": 0, + "content": "Text-IRSTD: Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.203, + 0.868, + 0.239 + ], + "angle": 0, + "content": "Feng Huang*, Shuyuan Zheng*, Zhaobing Qiu*,†, Huanxian Liu, Huanxin Bai, Liqiong Chen† Fuzhou University, Fuzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.242, + 0.812, + 0.275 + ], + "angle": 0, + "content": "huangf@fzu.edu.cn, zhengshuyuan200147@gmail.com, qiuzhaobing@fzu.edu.cn, liuhuanxian10@gmail.com, mw136296@gmail.com, liqionchen@fzu.edu.cn." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.31, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.341, + 0.485, + 0.735 + ], + "angle": 0, + "content": "Infrared small target detection is currently a hot and challenging task in computer vision. Existing methods usually focus on mining visual features of targets, which struggles to cope with complex and diverse detection scenarios. The main reason is that infrared small targets have limited image information on their own, thus relying only on visual features fails to discriminate targets and interferences, leading to lower detection performance. To address this issue, we introduce a novel approach leveraging semantic text to guide infrared small target detection, called Text-IRSTD. It innovatively expands classical IRSTD to text-guided IRSTD, providing a new research idea. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. On the other hand, we propose a progressive cross-modal semantic interaction decoder (PCSID) to facilitate information fusion between texts and images. In addition, we construct a new benchmark consisting of 2,755 infrared images of different scenarios with fuzzy semantic textual annotations, called FZDT. Extensive experimental results demonstrate that our method achieves better detection performance and target contour recovery than the state-of-the-art methods. Moreover, proposed Text-IRSTD shows strong generalization and wide application prospects in unseen detection scenarios. The dataset and code are available at https://github.com/Zhengsy0407/Text-IRSTD." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Infrared small target detection (IRSTD) has become an important and challenging issue in computer vision, which is widely used in both civil and military applications [12, 13, 19]. However, due to the long range detection needs, imaged targets usually occupy only a few pixels and lack tex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.907, + 0.357 + ], + "angle": 0, + "content": "ture details [20, 27, 32]. Therefore, how to accurately detect infrared (IR) small targets from complex backgrounds remains an open issue." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.371, + 0.925, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.597, + 0.909, + 0.655 + ], + "angle": 0, + "content": "Figure 1. Visual comparison between the SOTA model and proposed Text-IRSTD. (a) Detection results of the SOTA model [29], which relies only on visual features. (b) Detection results of proposed Text-IRSTD, which uses both textual and visual features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In the early days, many traditional methods emerged to improve detection accuracy. They mainly consist of filter-based methods [11, 23], local saliency-based methods [1, 24, 34] and sparse representation-based methods [8, 14, 25]. However, reliance on artificial feature representations prevents them from further mining the deep semantic information of real targets. In contrast, deep learning (DL)-based methods autonomously learn multi-layered features in a data-driven manner and achieve significant performance improvements. For example, Dai et al. [2] propose asymmetric contextual modulation (ACM) module to better interact with high-level semantics and low-level detail features. For higher feature utilization, Zhang et al. [33] propose attention-guided pyramid context network (AGPC-Net), which mines the correlations between feature map" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10635" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "patches. In addition, researchers have also attempted to fuse prior information for higher detection accuracy [7, 28]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.122, + 0.482, + 0.302 + ], + "angle": 0, + "content": "However, limited by scarce image features, existing methods fail to perform satisfactory in complex scenes. As shown in Fig. 1, the SOTA model still suffers from many false detections due to limited ability to discriminate between targets and interference. Recently, with the development of large-scale visual language models [17], semantic text-image-based models [10, 22] have made impressive progress in generic target detection. However, applying them directly to infrared small target detection works poorly. The main reason is that ambiguous target categories in infrared images invalidate existing data annotation and feature fusion methods." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.304, + 0.485, + 0.758 + ], + "angle": 0, + "content": "To address above challenges, this paper proposes a novel approach leveraging semantic text to guide infrared small target detection, called Text-IRSTD. It innovatively expands classical IRSTD to text-guided IRSTD, providing a new idea for IRSTD. As shown in Fig. 1, by utilizing both text and image information, proposed Text-IRSTD can further suppress clutter interference and improve detection performance. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. This prompt provides more text-image semantic associations that contribute to clutter suppression and target contour recovery. On the other hand, we propose a progressive cross-modal semantic interaction decoder (PCSID) that fuses both text and image information to further improve the detection performance. PCSID consists of two core components: text-guided feature aggregation (TGFA) block and text-guided semantic interaction (TGSI) block. Specifically, TGFA first utilizes text information to guide cross-layer visual feature aggregation for key feature extraction. Where key features (e.g., contrast [24]) refer to those features that are more important than others (e.g., intensity) for IRSTD. Then, TGSI block simultaneously performs feature modulation from two different perspectives, thus realizing a more refined text-image information interaction. In addition, we construct a new benchmark consisting of 2,755 infrared images of different scenarios with fuzzy semantic textual annotations, called FZDT. Extensive experiments on multiple public datasets demonstrate that our approach achieves better detection performance and generalization than representative state-of-the-art algorithms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.769, + 0.233, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.483, + 0.84 + ], + "angle": 0, + "content": "In this section, we first briefly review existing methods for IRSTD. Then, an introduction to text-image models in other domains is presented." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.254, + 0.863 + ], + "angle": 0, + "content": "2.1. IRSTD Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In the early days, many traditional methods emerged to improve detection accuracy. They mainly consist of filter" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.453 + ], + "angle": 0, + "content": "based methods [11, 23], local saliency-based methods [1, 24, 34] and sparse representation-based methods [8, 14, 25]. However, reliance on artificial feature representations prevents them from further mining the deep semantic information of real targets. In contrast, deep learning (DL)-based methods autonomously learn multi-layered features in a data-driven manner and have made some impressive progresses recently. To mitigate feature loss of IR small targets, several researchers focus on improving the cross-layer feature fusion approach. For example, Dai et al. [2] propose an asymmetric contextual modulation (ACM) module to better interact with high-level semantics and low-level detail features. For higher feature utilization, Zhang et al. [33] propose the attention-guided pyramid context network (AGPCNet), which mines the correlations between feature map patches. In order to interact cross-layer features incrementally, Li et al. [9] propose dense nested attention network (DNA-Net). Several researchers also attempted to improve the model structure. For example, Wu et al. [21] embed a smaller U-Net into the U-Net backbone to learn target features from multiple levels and scales. In addition, Hou et al. [7] try to introduce prior information and design a robust infrared small target detection network (RISTDnet) combining artificial features and neural networks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.592 + ], + "angle": 0, + "content": "Existing methods have made many impressive progress in IRSTD. However, the visual features of IR small targets are quite limited, which renders their performance unsatisfactory in complex and diverse detection scenarios. In contrast, we explore a new research idea to introduce text semantic features into IRSTD and devise a novel text-image fusion detection method. Utilizing added textual information, proposed Text-IRSTD can further suppress background interference and improve detection performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.605, + 0.698, + 0.621 + ], + "angle": 0, + "content": "2.2. Text-Image Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.905, + 0.764 + ], + "angle": 0, + "content": "With the advancement of deep learning techniques and the support of large datasets, text-image models has achieved success. In particular, CLIP [17] utilizes 400 million matched pairs datasets and contrast learning techniques to demonstrate powerful performance and wide application prospects in multiple domains and tasks. CLIP significantly improves text-image alignment and avoids the reliance on explicit labeling. With the support of CLIP models, text-driven generic target detection has also developed rapidly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.902 + ], + "angle": 0, + "content": "For better detection performance, RegionCLIP [35] leverages CLIP to match image regions with template captions, achieving fine-grained alignment between image and text features. ViLD [5] designs a training method based on knowledge distillation, which distills the knowledge from a pretrained open-vocabulary image classification model into a two-stage detector. CORA [22] mitigates the distribution gap between overall image features and regional features when using CLIP directly for detector training by introduc" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10636" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.244, + 0.107 + ], + "angle": 0, + "content": "ing learnable prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.11, + 0.485, + 0.215 + ], + "angle": 0, + "content": "Although existing text-image models for generic target detection are illuminating, they are difficult to apply to IRSTD tasks. The reason is that it is difficult to establish appropriate text-image semantic associations and interactions for IR small targets with ambiguous categories. Therefore, it is important and challenging to investigate text-image detection models dedicated to IRSTD tasks." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.236, + 0.458, + 0.254 + ], + "angle": 0, + "content": "3. Fuzzy Semantic Text Prompt For IRSTD" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.265, + 0.485, + 0.536 + ], + "angle": 0, + "content": "Semantic text prompts aim to establish a proper matching relationship between text and images, which is crucial for cross-modal target detection. In generic target detection [5, 35], researchers usually directly describe specific categories of targets, such as \"A photo of a girl\" and \"A photo of a soccer\" in Fig. 2(a). In this way, it is possible to focus not only on the visual features of a single target, but also to further explore semantic relations between multiple targets, e.g. a small round target near a girl is more likely to be a soccer. However, IR small targets typically occupy only a few pixels and lack visual features such as texture. As shown in Fig. 2(b), real targets are usually presented as bright spots that cannot be further identified as a specific category such as an aircraft or a bird. Therefore, using existing text prompts to describe ambiguous IR small target categories will lead to incorrect text-image matching relationships, impairing detection accuracy and model generalizability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To address this issue, we propose a novel fuzzy semantic text prompt for IRSTD, which includes two templates: scenario description and task description. Specifically, for known scenes with semantic information, we use the scenario description and describe one IR image as “A photo of a [Interested Region] target in the [Scene] background”, where [Interested Region] denotes the region where targets need to be detected, e.g., sky target and ground target, and [Scene] denotes the detection scene, e.g., sky-ground background, ocean-sky background. For example, we describe the IR image Fig. 2(b) as: “A photo of a sky target in the sky and ground background”, which not only establishes correct and robust matching relationships, but also provides more semantic associations. In this way, the model can focus on more shared visual features of sky targets, such as shape contours, which helps target shape recovery and detection. Moreover, as shown in Fig. 1, more semantic associations are established and utilized to further suppress complex ground interference and improve detection accuracy. For unseen scenes without semantic information, we use the task description, i.e., the default text prompt is “A photo of a small target in the background”." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.091, + 0.925, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.195, + 0.836, + 0.207 + ], + "angle": 0, + "content": "(a) Typical prompt template for object detection" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.21, + 0.922, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.341, + 0.835, + 0.353 + ], + "angle": 0, + "content": "(b) Proposed fuzzy semantic prompt for IRSTD" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.368, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Figure 2. (a) Typical semantic text prompt for generic target detection. (b) Proposed fuzzy semantic text prompt for IRSTD." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.405, + 0.605, + 0.421 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.507 + ], + "angle": 0, + "content": "In this section, we first present the overall architecture of proposed Text-IRSTD. Then, Sections 4.2 and 4.3 introduce two core components: TGFA block for aggregating key visual features and TGSI block for facilitating text-image information interaction, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.71, + 0.53 + ], + "angle": 0, + "content": "4.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Fig. 3 illustrates the overall structure of Text-IRSTD. First, one infrared image \\(I_{img}\\) and the corresponding text prompt \\(I_{text}\\) are fed into the image encoder \\(\\mathcal{F}_i\\) and text encoder \\(\\mathcal{F}_t\\), respectively, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.607, + 0.905, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E _ {i m g} = \\mathcal {F} _ {i} (I _ {i m g}), \\\\ \\Gamma_ {i m g} = \\mathcal {F} _ {i} (I _ {i m g}). \\end{array} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.628, + 0.773, + 0.642 + ], + "angle": 0, + "content": "\\[\nE _ {t e x t} = \\mathcal {F} _ {t} (I _ {t e x t}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.907, + 0.711 + ], + "angle": 0, + "content": "where \\( E_{img} \\) and \\( E_{text} \\) denote the image and text encoder outputs respectively. Note that \\( \\mathcal{F}_i \\) consists of four groups of ResNeSt [30] blocks, while \\( \\mathcal{F}_t \\) denote the pre-trained language model CLIP with frozen weights." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.906, + 0.757 + ], + "angle": 0, + "content": "Then, \\( E_{img} \\) and \\( E_{text} \\) are fed into PCSID for text-image feature interaction and enhancement. The result of cross-modal interaction \\( D_{cm} \\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.768, + 0.905, + 0.784 + ], + "angle": 0, + "content": "\\[\nD _ {c m} = \\operatorname {P C S I D} \\left(E _ {i m g}, E _ {t e x t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.793, + 0.906, + 0.87 + ], + "angle": 0, + "content": "where PCSID consists of three groups of TGFA blocks and TGSI blocks in series. TGFA and TGSI will be introduced in Sections 4.2 and 4.3, respectively. Note that the fourth layer output \\( E_{\\mathrm{img}}^{(4)} \\) of \\( \\mathcal{F}_i \\) is enhanced by detail perception module (DPM) and then fed to PCSID." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Finally, we apply a \\(1 \\times 1\\) convolution \\(f_{c}^{1}\\) and sigmoid function to the decoder outputs \\(D_{cm}\\). Thus, the cross-modal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10637" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.912, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.336, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 3. Overview of the proposed Text-IRSTD. It consists of three main components: text encoder, image encoder, and cross-modal decoder PCSID with their outputs represented as \\( E_{text} \\), \\( E_{img}^{(i)} \\), and \\( D_{cm} \\), respectively. Where DPM denotes detail perception module, TGFA Block denotes text-guided feature aggregation block and TGSI Block denotes text-guided semantic interaction block." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.414, + 0.412 + ], + "angle": 0, + "content": "detection result \\(O_{cm}\\) can be obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.424, + 0.483, + 0.442 + ], + "angle": 0, + "content": "\\[\nO _ {c m} = \\operatorname {S i g m o i d} \\left(f _ {c} ^ {1} \\left(D _ {c m}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.455, + 0.431, + 0.472 + ], + "angle": 0, + "content": "4.2. Text-guided Feature Aggregation Block" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Effective aggregation of visual features is crucial for detection models. Limited by scarce visual information, existing methods struggle to capture key features from high-level and low-level feature maps. To this end, we propose the novel text-guided cross-layer feature aggregation block TGFA (shown in Fig. 4). TGFA transforms text features \\( E_{text} \\) into modulation information through MLP, guiding the model to capture key features that are easily neglected in complex scenes. The details are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.628, + 0.483, + 0.645 + ], + "angle": 0, + "content": "\\[\nF _ {t, 1}, F _ {t, 2} = \\operatorname {S p i l t} \\left(\\operatorname {M L P} \\left(E _ {\\text {t e x t}}\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.659, + 0.483, + 0.68 + ], + "angle": 0, + "content": "\\[\nF _ {f} ^ {l} = F _ {l} \\otimes F _ {t, 1}, F _ {f} ^ {h} = F _ {h} \\otimes F _ {t, 2}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.484, + 0.763 + ], + "angle": 0, + "content": "where \\(F_{l}\\) and \\(F_{h}\\) denote low-level feature and high-level feature, \\(F_{f}^{l}\\) and \\(F_{f}^{h}\\) denote the fusion features after text modulation, \\(\\mathrm{Spit}(\\cdot)\\) denotes splitting the feature into two equal groups along the channel dimension, \\(\\otimes\\) denotes element-wise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Low-level features contain rich target locations and details, but are easily lost at deep layers because small IR targets occupy only a few pixels. Therefore, we devise the detail perception module DPM as shown in Fig. 4. Specifically, for a given low-level fusion feature \\( F_{f}^{l} \\), a \\( 1 \\times 1 \\) convolution is first applied to expand the channel dimension by a factor of \\( \\mu \\), and then divide the resulting feature map into two groups \\( F_{l,1} \\in \\mathbb{R}^{\\frac{\\mu C}{2} \\times H \\times W} \\) and \\( F_{l,2} \\in \\mathbb{R}^{\\frac{\\mu C}{2} \\times H \\times W} \\) along the channel dimension." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.398, + 0.909, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.575, + 0.907, + 0.604 + ], + "angle": 0, + "content": "Figure 4. Structure of TGFA block, where DPM denotes detail perception module and CA denotes channel attention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.687 + ], + "angle": 0, + "content": "Then, we use \\(3 \\times 3\\) and \\(5 \\times 5\\) depth-wise convolutions to capture important local features such as contrast at different scales and concatenate these two groups of features along the channel dimension. The multi-scale detailed feature mapping \\(F_{ms}\\) is obtained from the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.697, + 0.905, + 0.716 + ], + "angle": 0, + "content": "\\[\nF _ {l, 1}, F _ {l, 2} = \\operatorname {S p l i t} \\left(f _ {c} ^ {1} \\left(F _ {f} ^ {l}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.727, + 0.905, + 0.746 + ], + "angle": 0, + "content": "\\[\nF _ {m s} = f _ {c} ^ {1} \\left(\\operatorname {C o n t a c t} \\left(f _ {d w c} ^ {3} \\left(F _ {l, 1}\\right), f _ {d w c} ^ {5} \\left(F _ {l, 2}\\right)\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.752, + 0.907, + 0.811 + ], + "angle": 0, + "content": "where \\(\\operatorname{Contact}(\\cdot)\\) denotes concatenating the two groups of features along the channel dimension, \\(f_{c}^{1}\\) denotes \\(1 \\times 1\\) convolution, \\(f_{dwc}^{3}\\) and \\(f_{dwc}^{5}\\) denote \\(3 \\times 3\\) and \\(5 \\times 5\\) depth-wise convolutions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.813, + 0.908, + 0.872 + ], + "angle": 0, + "content": "After obtaining the multi-scale detailed feature mapping \\( F_{ms} \\), we use point-wise convolution to aggregate the contextual features at each spatial location. For simplicity, it can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.885, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\nF _ {D P} = \\sigma \\left(\\operatorname {P W C o n v} _ {2} \\left(\\gamma \\left(\\operatorname {P W C o n v} _ {1} \\left(F _ {m s}\\right)\\right)\\right) \\otimes F _ {m s}, \\right. \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10638" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "where \\(\\mathrm{PWConv}_1\\) and \\(\\mathrm{PWConv}_2\\) denote the point-wise convolution with kernel sizes of \\(\\frac{C}{4} \\times C \\times 1 \\times 1\\) and \\(C \\times \\frac{C}{4} \\times 1 \\times 1\\), \\(\\sigma\\) and \\(\\gamma\\) represent the sigmoid activation function and rectified linear unit." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.272 + ], + "angle": 0, + "content": "Moreover, high-level features can provide more accurate semantic information, while adjusting the weights of different channels can adaptively enhance the representation of key information and suppress redundant features. Therefore, for a given high-level fusion features \\( F_{f}^{h} \\), we first use global average pooling on each channel to obtain the global feature representation \\( z_{c} \\in \\mathbb{R}^{C \\times 1 \\times 1} \\), which can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.283, + 0.483, + 0.327 + ], + "angle": 0, + "content": "\\[\nz _ {c} = \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} F _ {f} ^ {h} [:, i, j ]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.483, + 0.383 + ], + "angle": 0, + "content": "Then, a nonlinear mapping of this global information is performed through two fully connected layers to establish dependencies between the channels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.484, + 0.432 + ], + "angle": 0, + "content": "Finally, the channel feature maps are normalized using the sigmoid function, and an element-wise multiplication is performed with the \\( F_{f}^{h} \\), the output of CA is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.444, + 0.483, + 0.466 + ], + "angle": 0, + "content": "\\[\nF _ {C A} = \\sigma \\left(F c _ {2} \\left(F c _ {1} \\left(z _ {c}\\right)\\right)\\right) \\otimes F _ {f} ^ {h}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.522 + ], + "angle": 0, + "content": "where \\( Fc_{1} \\in \\mathbb{R}^{\\frac{C}{r} \\times C} \\) and \\( Fc_{2} \\in \\mathbb{R}^{C \\times \\frac{C}{r}} \\) represent the two fully connected operations, \\( r \\) denotes the channel reduction rate." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.535, + 0.433, + 0.55 + ], + "angle": 0, + "content": "4.3. Text-guided Semantic Interaction Block" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.483, + 0.722 + ], + "angle": 0, + "content": "Due to the limited visual information, existing models struggle to effectively suppress complex interference and recover target contours. Therefore, we introduce semantic text prompts and propose TGSI block, which achieves better detection performance through text-image information interaction. Figure 3 shows the detailed structure of TGSI block. First, we take the image feature \\( M \\) and the text feature \\( E_{text} \\) as inputs, where \\( E_{text} \\) is obtained by feeding the text prompt \\( T_{text} \\) into the CLIP. Notably, we freeze the weights of the text encoder to preserve the original capabilities of the pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.724, + 0.483, + 0.842 + ], + "angle": 0, + "content": "Furthermore, we design a Feed-forward Network (FFN) to further explore the deep semantic information embedded in the text features. This helps the model capture more fine-grained semantic associations between the text, the target, and the regions of interest. Specifically, the text feature \\( E_{text} \\) is mapped into two learnable semantic parameters \\( \\alpha_{t} \\) and \\( \\beta_{t} \\) after going through the FFN, which can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.846, + 0.483, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {t}, \\beta_ {t} = \\operatorname {F F N} \\left(E _ {\\text {t e x t}}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Finally, the semantic parameters \\(\\alpha_{t}\\) and \\(\\beta_{t}\\) modulate the image feature \\(M\\) from the perspective of gain control and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "offset adjustment, respectively. Specifically, the gain parameter \\(\\alpha_{t}\\) directs the network to pay more attention to the regions of interest of the text prompt and reduces false alarms in other regions. The offset parameter \\(\\beta_{t}\\) provides finer tuning and helps in the recovery of target contour details. The fused semantic features \\(\\hat{M}\\) are obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.195, + 0.905, + 0.212 + ], + "angle": 0, + "content": "\\[\n\\hat {M} = \\left(1 + \\alpha_ {t}\\right) \\otimes M + \\beta_ {t}. \\tag {12}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.226, + 0.646, + 0.243 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.252, + 0.8, + 0.266 + ], + "angle": 0, + "content": "5.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Datasets: All experiments are conducted on proposed FZDT dataset, which contains three public image datasets NUAA-SIRST [2], NUDT-SIRST [9], IRSTD-1k [32], and corresponding fuzzy semantic text prompts. Following existing works [26], each dataset is divided into training set, validation set and test set in the ratio of \\( 6:2:2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.365, + 0.905, + 0.426 + ], + "angle": 0, + "content": "Evaluation Metrics: For comprehensive algorithmic comparison, the evaluation metrics include intersection over union (IoU), false alarm rate \\((\\mathrm{F}_a)\\), F-measure \\((\\mathrm{F}_1)\\), and probability of detection \\((\\mathrm{P}_d)\\) [29]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.437, + 0.731, + 0.452 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.905, + 0.565 + ], + "angle": 0, + "content": "To evaluate the efficacy of Text-IRSTD, we use the common CLIP-ViT-B/32 [17] as the text encoder. Our network is implemented with Pytorch on a single NVIDIA GeForce RTX 3090 GPU. We adopt the AdamW optimizer with a learning rate of 0.0001 and a weight decay of 0.05. Every images are normalized and randomly cropped to \\(256 \\times 256\\). The batch size and epoch are set as 8 and 500, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.906, + 0.686 + ], + "angle": 0, + "content": "The comparison algorithms for experiments include eight state-of-the-art DL-based methods: ACMNet [2], ALCNet [3], AGPCNet [33], DNANet [9], UIUNet [21], RDIAN [18], DMFNet [6] and SCTrans [29], and five recent traditional methods: IPI [4], PSTNN [31], FKRW [15], GSWLCM [16] and RUMFR [1]. To further illustrate the advancement of our method, Text-IRSTD without text prompts is also selected as the baseline method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.707, + 0.713 + ], + "angle": 0, + "content": "5.3.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The quantitative results of different methods are presented in Tab. 1. Overall, our Text-IRSTD achieves the best performance across all evaluation metrics. For example, the \\( \\mathrm{F}_a \\) of our method achieves as low as \\( 1.032 \\times 10^{-6} \\) on NUDT-SIRST, while also reaching the highest \\( \\mathrm{P}_d \\) of \\( 99.73\\% \\), demonstrating our superior capability in suppressing complex interference. Additionally, we obtain the best IoU of \\( 95.84\\% \\), which is \\( 1.01\\% \\) higher than the second best, indicating better target contour recovery capability. Traditional methods have lower metrics and exhibit poor performance. DL-based methods can further mine deep semantic features in a data-driven manner, and thus have significant" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10639" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative results of different methods. Results for the metrics of IoU(%, Pd(%) F, \\( (10^{-6}) \\) and \\( \\mathrm{F_1}(\\%) \\) are presented. The best values are highlighted with bold, the second best values are underlined." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.122, + 0.907, + 0.384 + ], + "angle": 0, + "content": "
MethodNUDT-SIRSTIRSTD-1k
IoUPdFaF1IoUPdFaF1
IPI [4]38.1590.78369.961.5230.9578.91180.455.15
PSTNN [31]17.2154.4781.6229.4525.5862.9678.9141.35
FKRW [15]25.3860.26107.538.5415.8954.5426.7527.08
GSWLCM [16]4.90965.0015.719.3013.06859.2526.946.045
RUMFR [1]26.0867.8996.4239.2510.2653.8725.2918.59
ACMNet [2]68.1496.5717.3883.4161.1487.8841.7974.17
ALCNet [3]72.5696.057.57286.7858.2789.9044.6972.92
AGPCNet [33]85.9797.637.22893.5162.5191.5820.9777.68
DNANet [9]94.7098.683.21297.3164.5190.2318.2778.77
UIUNet [21]89.1597.115.50694.2364.3290.2428.3379.02
RDIAN [18]86.0898.156.71192.3661.8688.5641.4476.27
DMFNet [6]87.6798.422.29493.3064.4590.1418.2478.52
SCTrans [29]94.8398.422.06597.1962.9592.2423.4078.04
Text-IRSTD (w/o text)95.2598.941.66397.3765.5091.5616.1378.73
Text-IRSTD (Ours)95.8499.731.03297.9569.5792.5914.9779.24
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.395, + 0.482, + 0.531 + ], + "angle": 0, + "content": "improvements in all metrics. However, the limited visual information of the target makes it difficult to further improve their detection accuracy in complex scenes. For instance, these SOTA DL-based methods achieve a maximum IoU of only \\(64.51\\%\\) on the IRSTD-1k, which is difficult to satisfy higher detection requirements. By incorporating text information and designing cross-modal information interaction methods, the proposed Text-IRSTD achieves up to \\(69.57\\%\\) performance improvement in IoU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.483, + 0.624 + ], + "angle": 0, + "content": "It is worth noting that our method is able to produce satisfactory results even when using the default text input ('w/o text') in the testing stage. This phenomenon indicates that our method does not rely completely on detailed text prompts, and the inherent advantages of the model design contribute to the performance of IRSTD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.635, + 0.274, + 0.65 + ], + "angle": 0, + "content": "5.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.825 + ], + "angle": 0, + "content": "We visualize the detection results of six representative algorithms, as shown in Fig. 5 and Fig. 6. It can be seen that traditional algorithms (e.g., PSTNN) have a large number of missed detections and false alarms due to their limited artificial feature representation capabilities. DL-based methods have made significant progress in background suppression by mining high-level semantic features. However, they all showed false alarms in the third row of images in Fig. 5. The reason is that the limited visual information of infrared images makes it difficult to further suppress interference and improve detection performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In contrast, proposed Text-IRSTD accurately detects targets in every scene and suppresses almost all background interference. By introducing semantic text to expand the limited visual information, Text-IRSTD can further improve the detection accuracy in complex scenes. As shown" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.487 + ], + "angle": 0, + "content": "in Fig. 5, thanks to the PCSID, we realize high-performance interactions across modal information, which can focus more on the region of interest to suppress complex interferences, while perceive and enhance the detail information to guide the accurate detection and contour recovery of targets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.501, + 0.807, + 0.515 + ], + "angle": 0, + "content": "5.5. Performance in Unseen Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.614 + ], + "angle": 0, + "content": "In practical applications, models often need to face unseen detection scenarios. The detection performance in unseen scenarios can evaluate the practicality and generalization of models to some extent. To further validate the advancement of proposed methods, we compare the detection capabilities of different methods in unseen scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.906, + 0.723 + ], + "angle": 0, + "content": "Specifically, we train different models using the NUDT-SIRST dataset and then test their detection performance using the unseen NUAA-SIRST dataset. Tab. 2 presents the detection results of different models on the unseen dataset. It can be seen that our method outperforms the state-of-the-art models in terms of IoU, \\( \\mathrm{P}_d \\) and \\( \\mathrm{F}_a \\), implying better detection accuracy and scene generalization." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.74, + 0.905, + 0.768 + ], + "angle": 0, + "content": "Table 2. Detection performance of different methods for the unseen dataset." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.772, + 0.902, + 0.898 + ], + "angle": 0, + "content": "
MethodNUDT-SIRST (Train)
NUAA-SIRST (Test)
IoU(%)Pd(%)Fa(10-6)
DNANet [9]35.8076.14141.3
UIUNet [21]53.3280.7335.15
RDIAN [18]39.3577.9843.68
DMFNet [6]43.7079.8160.07
SCTrans [29]50.9783.4837.71
Text-IRSTD (Ours)54.5187.1529.86
" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10640" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.907, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.515, + 0.907, + 0.545 + ], + "angle": 0, + "content": "Figure 5. Visual results of different IRSTD methods. The enlarged view of the target is shown in image corners. The red, blue, and yellow boxes represent correctly detected targets, missed targets, and false detections, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.555, + 0.245, + 0.571 + ], + "angle": 0, + "content": "5.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Impact of Text Prompt: In order to verify the effectiveness of fuzzy text prompts, we designed three textual prompt variants for experiments and the results are shown in Tab. 3. It can be seen that the use of [Interested Region] target as a prompt to describe the IR small target helps to avoid the wrong text-image matching relationship caused by the use of explicit categories of prompts. Meanwhile, the description of [Scene] background further improves the detection performance in complex scenes by establishing more appropriate semantic associations. More importantly, experiments demonstrate the possibility of improving IRSTD performance with text and provide an effective technical path for further research." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Impact of PCSID: To validate the effectiveness of each component in PCSID, we perform several ablation studies on the IRSTD-1k, and the results are shown in Tab. 4. The results of the experiment indicate that both TGSI block and TGFA block improve the performance of the baseline model, and when use together, the model achieve the best results. This further confirms the importance of their syn" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.555, + 0.907, + 0.584 + ], + "angle": 0, + "content": "ergistic effect in enhancing the performance of small target detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.909, + 0.743 + ], + "angle": 0, + "content": "Impact of TGFA: To demonstrate the effectiveness of TGFA block, we introduce three network variants for comparison, as shown in Tab. 5. It can be seen that proposed TGFA block achieves the best performance. The reason is that text features as modulation information can optimize the complementarity and interaction between cross-layer features. Moreover, DPM and CA can better perceive and enhance important visual features, such as contrast, which facilitate visual feature fusion and subsequent cross-modal information interaction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Impact of TGSI Block: We use four different text-image feature interaction methods to train the Text-IRSTD. As shown in Tab. 6, simply concatenating text and image features along the channel dimension ignores semantic inconsistencies across modalities, leading to inadequate semantic alignment. Without gain parameter \\(\\alpha_{t}\\), the model cannot dynamically focus on the regions of interest indicated by the text prompt. Similarly, without offset parameter \\(\\beta_{t}\\), the model cannot make fine adjustments to the image features using the text features, thus limiting the effective" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10641" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.086, + 0.907, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.408, + 0.908, + 0.437 + ], + "angle": 0, + "content": "Figure 6. 3D visualization results corresponding to those in Fig. 5. Different colors represent distinct grayscale values, gradually increasing from blue to yellow." + }, + { + "type": "table_caption", + "bbox": [ + 0.364, + 0.451, + 0.632, + 0.466 + ], + "angle": 0, + "content": "Table 3. Ablation study of different prompts." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.468, + 0.907, + 0.551 + ], + "angle": 0, + "content": "
TypePromptIoU(%)Pd(%)Fa(10-6)
Conventional promptsA photo of a [class]62.2290.5830.36
Variants of conventionA photo of a [class] in the [Scene] background66.6589.8921.69
Variants of OursA [Interested Region] target in the [Scene] background69.0191.6818.86
OursA photo of a [Interested Region] target in the [Scene] background69.5792.5914.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.558, + 0.466, + 0.572 + ], + "angle": 0, + "content": "Table 4. Ablation study of the TGFA block and TGSI block." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.576, + 0.483, + 0.657 + ], + "angle": 0, + "content": "
BaselineTGFATGSIIoU(%)Pd(%)Fa(10-6)
✓61.6789.2257.97
✓✓67.7591.8719.83
✓✓66.1991.9226.39
✓✓✓69.5792.5914.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.157, + 0.663, + 0.416, + 0.677 + ], + "angle": 0, + "content": "Table 5. Ablation study of the TGFA block." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.681, + 0.489, + 0.763 + ], + "angle": 0, + "content": "
ModelIoU(%)Pd(%)Fa(10-6)
TGFA w/o text guide68.7592.0223.68
TGFA w/o CA68.0191.2426.24
TGFA w/o DPM67.1689.5617.95
TGFA (Ours)69.5792.5914.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.158, + 0.769, + 0.414, + 0.782 + ], + "angle": 0, + "content": "Table 6. Ablation study of the TGSI block." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.786, + 0.484, + 0.868 + ], + "angle": 0, + "content": "
ModelIoU(%)Pd(%)Fa(10-6)
Concatenation67.7390.5729.49
TGSI w/o αt66.6590.1335.85
TGSI w/o βt68.0591.2116.22
TGSI (Ours)69.5792.5914.97
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "ness of target shape reconstruction. By combining both, the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.56, + 0.866, + 0.576 + ], + "angle": 0, + "content": "proposed TGSI Block achieves the best performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.6, + 0.634, + 0.615 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we extend the IRSTD task and propose a text-guided IRSTD framework to address the issue that current models struggle to further improve detection accuracy by relying only on limited visual information. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. This prompt provides more text-image semantic associations that contribute to clutter suppression and target contour recovery. On the other hand, we propose PCSID to facilitate the information interaction between text and images. Moreover, we have constructed the first text-image bimodal dataset, called FZDT, to provide the basis for subsequent text-guided IRSTD research. Extensive experimental results demonstrate that our method achieves better detection performance and target contour recovery than the state-of-the-art methods. More surprisingly, the proposed Text-IRSTD demonstrates strong generalization and wide application prospects in unseen detection scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10642" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.2 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62405060, in part by the Nature Science Foundation of Fujian Province under Grant 2022J05113 and 2024J01245, and in part by the Educational Research Program for Young and Middle-aged Teachers of Fujian Province under Grant JAT210035." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.226, + 0.188, + 0.243 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.252, + 0.484, + 0.32 + ], + "angle": 0, + "content": "[1] Liqiong Chen, Tong Wu, Shuyuan Zheng, Zhaobing Qiu, and Feng Huang. Robust unsupervised multifeature representation for infrared small target detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:10306-10323, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.323, + 0.484, + 0.391 + ], + "angle": 0, + "content": "[2] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Asymmetric contextual modulation for infrared small target detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 950-959, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.393, + 0.484, + 0.448 + ], + "angle": 0, + "content": "[3] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Attentional local contrast networks for infrared small target detection. IEEE transactions on geoscience and remote sensing, 59(11):9813-9824, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.45, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[4] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE transactions on image processing, 22(12):4996-5009, 2013. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.52, + 0.484, + 0.575 + ], + "angle": 0, + "content": "[5] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.484, + 0.632 + ], + "angle": 0, + "content": "[6] Tan Guo, Baojiang Zhou, Fulin Luo, Lei Zhang, and Xinbo Gao. Dmfnet: Dual-encoder multi-stage feature fusion network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-14, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.484, + 0.689 + ], + "angle": 0, + "content": "[7] Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Ye Zhao, Hao-liang Zheng, and Wei Zhang. Ristdnet: Robust infrared small target detection network. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[8] Xuan Kong, Chunping Yang, Siying Cao, Chaohai Li, and Zhenming Peng. Infrared small target detection via nonconvex tensor fibered rank approximation. IEEE Transactions on Geoscience and Remote Sensing, 60:1-21, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[9] Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo. Dense nested attention network for infrared small target detection. IEEE Transactions on Image Processing, 32:1745-1758, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.252, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[11] Yongsong Li, Zhengzhou Li, Yu Shen, and Jie Li. Infrared small target detection based on 1-d difference of guided filtering. IEEE Geoscience and Remote Sensing Letters, 20: 1-5, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[12] Fanzhao Lin, Kexin Bao, Yong Li, Dan Zeng, and Shiming Ge. Learning contrast-enhanced shape-biased representations for infrared small target detection. IEEE Transactions on Image Processing, 33:3047-3058, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.906, + 0.272 + ], + "angle": 0, + "content": "[13] Qiankun Liu, Rui Liu, Bolun Zheng, Hongkui Wang, and Ying Fu. Infrared small target detection with scale and location sensitivity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17490-17499, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.342 + ], + "angle": 0, + "content": "[14] Yuan Luo, Xiaorun Li, Shuhan Chen, Chaoqun Xia, and Liaoying Zhao. Imnn-lwec: A novel infrared small target detection based on spatial-temporal tensor model. IEEE Transactions on Geoscience and Remote Sensing, 60:1-22, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[15] Yao Qin, Lorenzo Bruzzone, Chengqiang Gao, and Biao Li. Infrared small target detection based on facet kernel and random walker. IEEE Transactions on Geoscience and Remote Sensing, 57(9):7104-7118, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[16] Zhaobing Qiu, Yong Ma, Fan Fan, Jun Huang, and Lang Wu. Global sparsity-weighted local contrast measure for infrared small target detection. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[18] Heng Sun, Junxiang Bai, Fan Yang, and Xiangzhi Bai. Receptive-field and direction induced attention network for infrared dim small target detection with a large-scale dataset irdst. IEEE Transactions on Geoscience and Remote Sensing, 61:1-13, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.906, + 0.677 + ], + "angle": 0, + "content": "[19] Xiaozhong Tong, Zhen Zuo, Shaojing Su, Junyu Wei, Xiaoyong Sun, Peng Wu, and Zongqing Zhao. St-trans: Spatial-temporal transformer for infrared small target detection in sequential images. IEEE Transactions on Geoscience and Remote Sensing, 62:1-19, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[20] Keyan Wang, Xueyan Wu, Peicheng Zhou, Zuntian Chen, Rui Zhang, Liyun Yang, and Yunsong Li. Afe-net: Attention-guided feature enhancement network for infrared small target detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:4208-4221, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.803 + ], + "angle": 0, + "content": "[21] Xin Wu, Danfeng Hong, and Jocelyn Chanussot. Uiu-net: U-net in u-net for infrared small object detection. IEEE Transactions on Image Processing, 32:364-376, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.908, + 0.873 + ], + "angle": 0, + "content": "[22] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7031-7040, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Yunhong Xin, Jiao Zhou, and Yishuan Chen. Dual multiscale filter with SSS and GW for infrared small target detec" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10643" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "tion. Infrared Physics & Technology, 81:97-108, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[24] Yunkai Xu, Minjie Wan, Xiaojie Zhang, Jian Wu, Yili Chen, Qian Chen, and Guohua Gu. Infrared small target detection based on local contrast-weighted multidirectional derivative. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-16, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[25]Fuju Yan,Guili Xu,Junpu Wang,Quan Wu,and Zhengsheng Wang. Infrared small target detection via schatten capped pnorm-based non-convex tensor low-rank approximation.IEEE Geoscience and Remote Sensing Letters,20: 1-5,2023.1,2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[26] Bo Yang, Xinyu Zhang, Jian Zhang, Jun Luo, Mingliang Zhou, and Yangjun Pi. Eflnet: Enhancing feature learning network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-11, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[27] Xinyi Ying, Li Liu, Yingqian Wang, Ruojing Li, Nuo Chen, Zaiping Lin, Weidong Sheng, and Shilin Zhou. Mapping degeneration meets label evolution: Learning infrared small target detection with single point supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15528-15538, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[28] Chuang Yu, Yunpeng Liu, Shuhang Wu, Xin Xia, Zhuhua Hu, Deyan Lan, and Xin Liu. Pay attention to local contrast learning networks for infrared small target detection. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[29] Shuai Yuan, Hanlin Qin, Xiang Yan, Naveed Akhtar, and Ajmal Mian. Sctransnet: Spatial-channel cross transformer network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-15, 2024. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[30] Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R Manmatha, et al. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2736–2746, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[31] Landan Zhang and Zhenming Peng. Infrared small target detection based on partial sum of the tensor nuclear norm. Remote Sensing, 11(4):382, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.624, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[32] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 877-886, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[33] Tianfang Zhang, Lei Li, Siying Cao, Tian Pu, and Zhenming Peng. Attention-guided pyramid context networks for detecting infrared small target under complex background. IEEE Transactions on Aerospace and Electronic Systems, 59(4): 4250-4261, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[34] Xiangyue Zhang, Jingyu Ru, and Chengdong Wu. Infrared small target detection based on gradient correlation filtering and contrast measurement. IEEE Transactions on Geoscience and Remote Sensing, 61:1-12, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[35] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16793-16803, 2022. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10644" + } + ] +] \ No newline at end of file diff --git a/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/f9aefd65-d52a-4731-ba81-d317b6edbcb4_origin.pdf b/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/f9aefd65-d52a-4731-ba81-d317b6edbcb4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9c6016d0d39936e1518b417d257d5d3b080af533 --- /dev/null +++ b/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/f9aefd65-d52a-4731-ba81-d317b6edbcb4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1ffddc6a00459b945df908042ab57f1ec052765ee81385eab5f115f7930256f8 +size 14010157 diff --git a/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/full.md b/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c1bf64ec7bb4b7977c233a77b30843c5d2f86836 --- /dev/null +++ b/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/full.md @@ -0,0 +1,302 @@ +# Text-IRSTD: Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes + +Feng Huang*, Shuyuan Zheng*, Zhaobing Qiu*,†, Huanxian Liu, Huanxin Bai, Liqiong Chen† Fuzhou University, Fuzhou, China + +huangf@fzu.edu.cn, zhengshuyuan200147@gmail.com, qiuzhaobing@fzu.edu.cn, liuhuanxian10@gmail.com, mw136296@gmail.com, liqionchen@fzu.edu.cn. + +# Abstract + +Infrared small target detection is currently a hot and challenging task in computer vision. Existing methods usually focus on mining visual features of targets, which struggles to cope with complex and diverse detection scenarios. The main reason is that infrared small targets have limited image information on their own, thus relying only on visual features fails to discriminate targets and interferences, leading to lower detection performance. To address this issue, we introduce a novel approach leveraging semantic text to guide infrared small target detection, called Text-IRSTD. It innovatively expands classical IRSTD to text-guided IRSTD, providing a new research idea. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. On the other hand, we propose a progressive cross-modal semantic interaction decoder (PCSID) to facilitate information fusion between texts and images. In addition, we construct a new benchmark consisting of 2,755 infrared images of different scenarios with fuzzy semantic textual annotations, called FZDT. Extensive experimental results demonstrate that our method achieves better detection performance and target contour recovery than the state-of-the-art methods. Moreover, proposed Text-IRSTD shows strong generalization and wide application prospects in unseen detection scenarios. The dataset and code are available at https://github.com/Zhengsy0407/Text-IRSTD. + +# 1. Introduction + +Infrared small target detection (IRSTD) has become an important and challenging issue in computer vision, which is widely used in both civil and military applications [12, 13, 19]. However, due to the long range detection needs, imaged targets usually occupy only a few pixels and lack tex + +ture details [20, 27, 32]. Therefore, how to accurately detect infrared (IR) small targets from complex backgrounds remains an open issue. + +![](images/ebbd5bdf2e26354bb118083726691eb37d4388e61f010ef790af37280950c44f.jpg) +Figure 1. Visual comparison between the SOTA model and proposed Text-IRSTD. (a) Detection results of the SOTA model [29], which relies only on visual features. (b) Detection results of proposed Text-IRSTD, which uses both textual and visual features. + +In the early days, many traditional methods emerged to improve detection accuracy. They mainly consist of filter-based methods [11, 23], local saliency-based methods [1, 24, 34] and sparse representation-based methods [8, 14, 25]. However, reliance on artificial feature representations prevents them from further mining the deep semantic information of real targets. In contrast, deep learning (DL)-based methods autonomously learn multi-layered features in a data-driven manner and achieve significant performance improvements. For example, Dai et al. [2] propose asymmetric contextual modulation (ACM) module to better interact with high-level semantics and low-level detail features. For higher feature utilization, Zhang et al. [33] propose attention-guided pyramid context network (AGPC-Net), which mines the correlations between feature map + +patches. In addition, researchers have also attempted to fuse prior information for higher detection accuracy [7, 28]. + +However, limited by scarce image features, existing methods fail to perform satisfactory in complex scenes. As shown in Fig. 1, the SOTA model still suffers from many false detections due to limited ability to discriminate between targets and interference. Recently, with the development of large-scale visual language models [17], semantic text-image-based models [10, 22] have made impressive progress in generic target detection. However, applying them directly to infrared small target detection works poorly. The main reason is that ambiguous target categories in infrared images invalidate existing data annotation and feature fusion methods. + +To address above challenges, this paper proposes a novel approach leveraging semantic text to guide infrared small target detection, called Text-IRSTD. It innovatively expands classical IRSTD to text-guided IRSTD, providing a new idea for IRSTD. As shown in Fig. 1, by utilizing both text and image information, proposed Text-IRSTD can further suppress clutter interference and improve detection performance. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. This prompt provides more text-image semantic associations that contribute to clutter suppression and target contour recovery. On the other hand, we propose a progressive cross-modal semantic interaction decoder (PCSID) that fuses both text and image information to further improve the detection performance. PCSID consists of two core components: text-guided feature aggregation (TGFA) block and text-guided semantic interaction (TGSI) block. Specifically, TGFA first utilizes text information to guide cross-layer visual feature aggregation for key feature extraction. Where key features (e.g., contrast [24]) refer to those features that are more important than others (e.g., intensity) for IRSTD. Then, TGSI block simultaneously performs feature modulation from two different perspectives, thus realizing a more refined text-image information interaction. In addition, we construct a new benchmark consisting of 2,755 infrared images of different scenarios with fuzzy semantic textual annotations, called FZDT. Extensive experiments on multiple public datasets demonstrate that our approach achieves better detection performance and generalization than representative state-of-the-art algorithms. + +# 2. Related Work + +In this section, we first briefly review existing methods for IRSTD. Then, an introduction to text-image models in other domains is presented. + +# 2.1. IRSTD Methods + +In the early days, many traditional methods emerged to improve detection accuracy. They mainly consist of filter + +based methods [11, 23], local saliency-based methods [1, 24, 34] and sparse representation-based methods [8, 14, 25]. However, reliance on artificial feature representations prevents them from further mining the deep semantic information of real targets. In contrast, deep learning (DL)-based methods autonomously learn multi-layered features in a data-driven manner and have made some impressive progresses recently. To mitigate feature loss of IR small targets, several researchers focus on improving the cross-layer feature fusion approach. For example, Dai et al. [2] propose an asymmetric contextual modulation (ACM) module to better interact with high-level semantics and low-level detail features. For higher feature utilization, Zhang et al. [33] propose the attention-guided pyramid context network (AGPCNet), which mines the correlations between feature map patches. In order to interact cross-layer features incrementally, Li et al. [9] propose dense nested attention network (DNA-Net). Several researchers also attempted to improve the model structure. For example, Wu et al. [21] embed a smaller U-Net into the U-Net backbone to learn target features from multiple levels and scales. In addition, Hou et al. [7] try to introduce prior information and design a robust infrared small target detection network (RISTDnet) combining artificial features and neural networks. + +Existing methods have made many impressive progress in IRSTD. However, the visual features of IR small targets are quite limited, which renders their performance unsatisfactory in complex and diverse detection scenarios. In contrast, we explore a new research idea to introduce text semantic features into IRSTD and devise a novel text-image fusion detection method. Utilizing added textual information, proposed Text-IRSTD can further suppress background interference and improve detection performance. + +# 2.2. Text-Image Models + +With the advancement of deep learning techniques and the support of large datasets, text-image models has achieved success. In particular, CLIP [17] utilizes 400 million matched pairs datasets and contrast learning techniques to demonstrate powerful performance and wide application prospects in multiple domains and tasks. CLIP significantly improves text-image alignment and avoids the reliance on explicit labeling. With the support of CLIP models, text-driven generic target detection has also developed rapidly. + +For better detection performance, RegionCLIP [35] leverages CLIP to match image regions with template captions, achieving fine-grained alignment between image and text features. ViLD [5] designs a training method based on knowledge distillation, which distills the knowledge from a pretrained open-vocabulary image classification model into a two-stage detector. CORA [22] mitigates the distribution gap between overall image features and regional features when using CLIP directly for detector training by introduc + +ing learnable prompts. + +Although existing text-image models for generic target detection are illuminating, they are difficult to apply to IRSTD tasks. The reason is that it is difficult to establish appropriate text-image semantic associations and interactions for IR small targets with ambiguous categories. Therefore, it is important and challenging to investigate text-image detection models dedicated to IRSTD tasks. + +# 3. Fuzzy Semantic Text Prompt For IRSTD + +Semantic text prompts aim to establish a proper matching relationship between text and images, which is crucial for cross-modal target detection. In generic target detection [5, 35], researchers usually directly describe specific categories of targets, such as "A photo of a girl" and "A photo of a soccer" in Fig. 2(a). In this way, it is possible to focus not only on the visual features of a single target, but also to further explore semantic relations between multiple targets, e.g. a small round target near a girl is more likely to be a soccer. However, IR small targets typically occupy only a few pixels and lack visual features such as texture. As shown in Fig. 2(b), real targets are usually presented as bright spots that cannot be further identified as a specific category such as an aircraft or a bird. Therefore, using existing text prompts to describe ambiguous IR small target categories will lead to incorrect text-image matching relationships, impairing detection accuracy and model generalizability. + +To address this issue, we propose a novel fuzzy semantic text prompt for IRSTD, which includes two templates: scenario description and task description. Specifically, for known scenes with semantic information, we use the scenario description and describe one IR image as “A photo of a [Interested Region] target in the [Scene] background”, where [Interested Region] denotes the region where targets need to be detected, e.g., sky target and ground target, and [Scene] denotes the detection scene, e.g., sky-ground background, ocean-sky background. For example, we describe the IR image Fig. 2(b) as: “A photo of a sky target in the sky and ground background”, which not only establishes correct and robust matching relationships, but also provides more semantic associations. In this way, the model can focus on more shared visual features of sky targets, such as shape contours, which helps target shape recovery and detection. Moreover, as shown in Fig. 1, more semantic associations are established and utilized to further suppress complex ground interference and improve detection accuracy. For unseen scenes without semantic information, we use the task description, i.e., the default text prompt is “A photo of a small target in the background”. + +![](images/3c93771244c902d6f7ce0c8add431b5339216eedb7f9ecd02a04f93bbdb5de0a.jpg) +(a) Typical prompt template for object detection + +![](images/baaf457a050d1fa96472c3237b42a17d6217b0aa5abb75d08af7403f2b6bd15a.jpg) +(b) Proposed fuzzy semantic prompt for IRSTD +Figure 2. (a) Typical semantic text prompt for generic target detection. (b) Proposed fuzzy semantic text prompt for IRSTD. + +# 4. Method + +In this section, we first present the overall architecture of proposed Text-IRSTD. Then, Sections 4.2 and 4.3 introduce two core components: TGFA block for aggregating key visual features and TGSI block for facilitating text-image information interaction, respectively. + +# 4.1. Overall Architecture + +Fig. 3 illustrates the overall structure of Text-IRSTD. First, one infrared image $I_{img}$ and the corresponding text prompt $I_{text}$ are fed into the image encoder $\mathcal{F}_i$ and text encoder $\mathcal{F}_t$ , respectively, as follows: + +$$ +\begin{array}{l} E _ {i m g} = \mathcal {F} _ {i} (I _ {i m g}), \\ \Gamma_ {i m g} = \mathcal {F} _ {i} (I _ {i m g}). \end{array} \tag {1} +$$ + +$$ +E _ {t e x t} = \mathcal {F} _ {t} (I _ {t e x t}), +$$ + +where $E_{img}$ and $E_{text}$ denote the image and text encoder outputs respectively. Note that $\mathcal{F}_i$ consists of four groups of ResNeSt [30] blocks, while $\mathcal{F}_t$ denote the pre-trained language model CLIP with frozen weights. + +Then, $E_{img}$ and $E_{text}$ are fed into PCSID for text-image feature interaction and enhancement. The result of cross-modal interaction $D_{cm}$ can be expressed as: + +$$ +D _ {c m} = \operatorname {P C S I D} \left(E _ {i m g}, E _ {t e x t}\right), \tag {2} +$$ + +where PCSID consists of three groups of TGFA blocks and TGSI blocks in series. TGFA and TGSI will be introduced in Sections 4.2 and 4.3, respectively. Note that the fourth layer output $E_{\mathrm{img}}^{(4)}$ of $\mathcal{F}_i$ is enhanced by detail perception module (DPM) and then fed to PCSID. + +Finally, we apply a $1 \times 1$ convolution $f_{c}^{1}$ and sigmoid function to the decoder outputs $D_{cm}$ . Thus, the cross-modal + +![](images/2558a3d70209199205f0131ab5cf4506674ec914a6dc7c4a69d479ada9ea695b.jpg) +Figure 3. Overview of the proposed Text-IRSTD. It consists of three main components: text encoder, image encoder, and cross-modal decoder PCSID with their outputs represented as $E_{text}$ , $E_{img}^{(i)}$ , and $D_{cm}$ , respectively. Where DPM denotes detail perception module, TGFA Block denotes text-guided feature aggregation block and TGSI Block denotes text-guided semantic interaction block. + +detection result $O_{cm}$ can be obtained as follows: + +$$ +O _ {c m} = \operatorname {S i g m o i d} \left(f _ {c} ^ {1} \left(D _ {c m}\right)\right). \tag {3} +$$ + +# 4.2. Text-guided Feature Aggregation Block + +Effective aggregation of visual features is crucial for detection models. Limited by scarce visual information, existing methods struggle to capture key features from high-level and low-level feature maps. To this end, we propose the novel text-guided cross-layer feature aggregation block TGFA (shown in Fig. 4). TGFA transforms text features $E_{text}$ into modulation information through MLP, guiding the model to capture key features that are easily neglected in complex scenes. The details are as follows: + +$$ +F _ {t, 1}, F _ {t, 2} = \operatorname {S p i l t} \left(\operatorname {M L P} \left(E _ {\text {t e x t}}\right)\right), \tag {4} +$$ + +$$ +F _ {f} ^ {l} = F _ {l} \otimes F _ {t, 1}, F _ {f} ^ {h} = F _ {h} \otimes F _ {t, 2}, \tag {5} +$$ + +where $F_{l}$ and $F_{h}$ denote low-level feature and high-level feature, $F_{f}^{l}$ and $F_{f}^{h}$ denote the fusion features after text modulation, $\mathrm{Spit}(\cdot)$ denotes splitting the feature into two equal groups along the channel dimension, $\otimes$ denotes element-wise multiplication. + +Low-level features contain rich target locations and details, but are easily lost at deep layers because small IR targets occupy only a few pixels. Therefore, we devise the detail perception module DPM as shown in Fig. 4. Specifically, for a given low-level fusion feature $F_{f}^{l}$ , a $1 \times 1$ convolution is first applied to expand the channel dimension by a factor of $\mu$ , and then divide the resulting feature map into two groups $F_{l,1} \in \mathbb{R}^{\frac{\mu C}{2} \times H \times W}$ and $F_{l,2} \in \mathbb{R}^{\frac{\mu C}{2} \times H \times W}$ along the channel dimension. + +![](images/70253ee2c60fa7c4fa46f28876b041049a1cf2953ca66ed2da13be1d445dffb2.jpg) +Figure 4. Structure of TGFA block, where DPM denotes detail perception module and CA denotes channel attention. + +Then, we use $3 \times 3$ and $5 \times 5$ depth-wise convolutions to capture important local features such as contrast at different scales and concatenate these two groups of features along the channel dimension. The multi-scale detailed feature mapping $F_{ms}$ is obtained from the following equation: + +$$ +F _ {l, 1}, F _ {l, 2} = \operatorname {S p l i t} \left(f _ {c} ^ {1} \left(F _ {f} ^ {l}\right)\right), \tag {6} +$$ + +$$ +F _ {m s} = f _ {c} ^ {1} \left(\operatorname {C o n t a c t} \left(f _ {d w c} ^ {3} \left(F _ {l, 1}\right), f _ {d w c} ^ {5} \left(F _ {l, 2}\right)\right)\right), \tag {7} +$$ + +where $\operatorname{Contact}(\cdot)$ denotes concatenating the two groups of features along the channel dimension, $f_{c}^{1}$ denotes $1 \times 1$ convolution, $f_{dwc}^{3}$ and $f_{dwc}^{5}$ denote $3 \times 3$ and $5 \times 5$ depth-wise convolutions. + +After obtaining the multi-scale detailed feature mapping $F_{ms}$ , we use point-wise convolution to aggregate the contextual features at each spatial location. For simplicity, it can be described as: + +$$ +F _ {D P} = \sigma \left(\operatorname {P W C o n v} _ {2} \left(\gamma \left(\operatorname {P W C o n v} _ {1} \left(F _ {m s}\right)\right)\right) \otimes F _ {m s}, \right. \tag {8} +$$ + +where $\mathrm{PWConv}_1$ and $\mathrm{PWConv}_2$ denote the point-wise convolution with kernel sizes of $\frac{C}{4} \times C \times 1 \times 1$ and $C \times \frac{C}{4} \times 1 \times 1$ , $\sigma$ and $\gamma$ represent the sigmoid activation function and rectified linear unit. + +Moreover, high-level features can provide more accurate semantic information, while adjusting the weights of different channels can adaptively enhance the representation of key information and suppress redundant features. Therefore, for a given high-level fusion features $F_{f}^{h}$ , we first use global average pooling on each channel to obtain the global feature representation $z_{c} \in \mathbb{R}^{C \times 1 \times 1}$ , which can be described as: + +$$ +z _ {c} = \frac {1}{H \times W} \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} F _ {f} ^ {h} [:, i, j ]. \tag {9} +$$ + +Then, a nonlinear mapping of this global information is performed through two fully connected layers to establish dependencies between the channels. + +Finally, the channel feature maps are normalized using the sigmoid function, and an element-wise multiplication is performed with the $F_{f}^{h}$ , the output of CA is defined as: + +$$ +F _ {C A} = \sigma \left(F c _ {2} \left(F c _ {1} \left(z _ {c}\right)\right)\right) \otimes F _ {f} ^ {h}, \tag {10} +$$ + +where $Fc_{1} \in \mathbb{R}^{\frac{C}{r} \times C}$ and $Fc_{2} \in \mathbb{R}^{C \times \frac{C}{r}}$ represent the two fully connected operations, $r$ denotes the channel reduction rate. + +# 4.3. Text-guided Semantic Interaction Block + +Due to the limited visual information, existing models struggle to effectively suppress complex interference and recover target contours. Therefore, we introduce semantic text prompts and propose TGSI block, which achieves better detection performance through text-image information interaction. Figure 3 shows the detailed structure of TGSI block. First, we take the image feature $M$ and the text feature $E_{text}$ as inputs, where $E_{text}$ is obtained by feeding the text prompt $T_{text}$ into the CLIP. Notably, we freeze the weights of the text encoder to preserve the original capabilities of the pre-trained model. + +Furthermore, we design a Feed-forward Network (FFN) to further explore the deep semantic information embedded in the text features. This helps the model capture more fine-grained semantic associations between the text, the target, and the regions of interest. Specifically, the text feature $E_{text}$ is mapped into two learnable semantic parameters $\alpha_{t}$ and $\beta_{t}$ after going through the FFN, which can be defined as: + +$$ +\alpha_ {t}, \beta_ {t} = \operatorname {F F N} \left(E _ {\text {t e x t}}\right). \tag {11} +$$ + +Finally, the semantic parameters $\alpha_{t}$ and $\beta_{t}$ modulate the image feature $M$ from the perspective of gain control and + +offset adjustment, respectively. Specifically, the gain parameter $\alpha_{t}$ directs the network to pay more attention to the regions of interest of the text prompt and reduces false alarms in other regions. The offset parameter $\beta_{t}$ provides finer tuning and helps in the recovery of target contour details. The fused semantic features $\hat{M}$ are obtained by: + +$$ +\hat {M} = \left(1 + \alpha_ {t}\right) \otimes M + \beta_ {t}. \tag {12} +$$ + +# 5. Experiments + +# 5.1. Datasets and Evaluation Metrics + +Datasets: All experiments are conducted on proposed FZDT dataset, which contains three public image datasets NUAA-SIRST [2], NUDT-SIRST [9], IRSTD-1k [32], and corresponding fuzzy semantic text prompts. Following existing works [26], each dataset is divided into training set, validation set and test set in the ratio of $6:2:2$ . + +Evaluation Metrics: For comprehensive algorithmic comparison, the evaluation metrics include intersection over union (IoU), false alarm rate $(\mathrm{F}_a)$ , F-measure $(\mathrm{F}_1)$ , and probability of detection $(\mathrm{P}_d)$ [29]. + +# 5.2. Implementation Details + +To evaluate the efficacy of Text-IRSTD, we use the common CLIP-ViT-B/32 [17] as the text encoder. Our network is implemented with Pytorch on a single NVIDIA GeForce RTX 3090 GPU. We adopt the AdamW optimizer with a learning rate of 0.0001 and a weight decay of 0.05. Every images are normalized and randomly cropped to $256 \times 256$ . The batch size and epoch are set as 8 and 500, respectively. + +The comparison algorithms for experiments include eight state-of-the-art DL-based methods: ACMNet [2], ALCNet [3], AGPCNet [33], DNANet [9], UIUNet [21], RDIAN [18], DMFNet [6] and SCTrans [29], and five recent traditional methods: IPI [4], PSTNN [31], FKRW [15], GSWLCM [16] and RUMFR [1]. To further illustrate the advancement of our method, Text-IRSTD without text prompts is also selected as the baseline method. + +# 5.3.Quantitative Results + +The quantitative results of different methods are presented in Tab. 1. Overall, our Text-IRSTD achieves the best performance across all evaluation metrics. For example, the $\mathrm{F}_a$ of our method achieves as low as $1.032 \times 10^{-6}$ on NUDT-SIRST, while also reaching the highest $\mathrm{P}_d$ of $99.73\%$ , demonstrating our superior capability in suppressing complex interference. Additionally, we obtain the best IoU of $95.84\%$ , which is $1.01\%$ higher than the second best, indicating better target contour recovery capability. Traditional methods have lower metrics and exhibit poor performance. DL-based methods can further mine deep semantic features in a data-driven manner, and thus have significant + +Table 1. Quantitative results of different methods. Results for the metrics of IoU(%, Pd(%) F, $(10^{-6})$ and $\mathrm{F_1}(\%)$ are presented. The best values are highlighted with bold, the second best values are underlined. + +
MethodNUDT-SIRSTIRSTD-1k
IoUPdFaF1IoUPdFaF1
IPI [4]38.1590.78369.961.5230.9578.91180.455.15
PSTNN [31]17.2154.4781.6229.4525.5862.9678.9141.35
FKRW [15]25.3860.26107.538.5415.8954.5426.7527.08
GSWLCM [16]4.90965.0015.719.3013.06859.2526.946.045
RUMFR [1]26.0867.8996.4239.2510.2653.8725.2918.59
ACMNet [2]68.1496.5717.3883.4161.1487.8841.7974.17
ALCNet [3]72.5696.057.57286.7858.2789.9044.6972.92
AGPCNet [33]85.9797.637.22893.5162.5191.5820.9777.68
DNANet [9]94.7098.683.21297.3164.5190.2318.2778.77
UIUNet [21]89.1597.115.50694.2364.3290.2428.3379.02
RDIAN [18]86.0898.156.71192.3661.8688.5641.4476.27
DMFNet [6]87.6798.422.29493.3064.4590.1418.2478.52
SCTrans [29]94.8398.422.06597.1962.9592.2423.4078.04
Text-IRSTD (w/o text)95.2598.941.66397.3765.5091.5616.1378.73
Text-IRSTD (Ours)95.8499.731.03297.9569.5792.5914.9779.24
+ +improvements in all metrics. However, the limited visual information of the target makes it difficult to further improve their detection accuracy in complex scenes. For instance, these SOTA DL-based methods achieve a maximum IoU of only $64.51\%$ on the IRSTD-1k, which is difficult to satisfy higher detection requirements. By incorporating text information and designing cross-modal information interaction methods, the proposed Text-IRSTD achieves up to $69.57\%$ performance improvement in IoU. + +It is worth noting that our method is able to produce satisfactory results even when using the default text input ('w/o text') in the testing stage. This phenomenon indicates that our method does not rely completely on detailed text prompts, and the inherent advantages of the model design contribute to the performance of IRSTD. + +# 5.4. Qualitative Results + +We visualize the detection results of six representative algorithms, as shown in Fig. 5 and Fig. 6. It can be seen that traditional algorithms (e.g., PSTNN) have a large number of missed detections and false alarms due to their limited artificial feature representation capabilities. DL-based methods have made significant progress in background suppression by mining high-level semantic features. However, they all showed false alarms in the third row of images in Fig. 5. The reason is that the limited visual information of infrared images makes it difficult to further suppress interference and improve detection performance. + +In contrast, proposed Text-IRSTD accurately detects targets in every scene and suppresses almost all background interference. By introducing semantic text to expand the limited visual information, Text-IRSTD can further improve the detection accuracy in complex scenes. As shown + +in Fig. 5, thanks to the PCSID, we realize high-performance interactions across modal information, which can focus more on the region of interest to suppress complex interferences, while perceive and enhance the detail information to guide the accurate detection and contour recovery of targets. + +# 5.5. Performance in Unseen Scenarios + +In practical applications, models often need to face unseen detection scenarios. The detection performance in unseen scenarios can evaluate the practicality and generalization of models to some extent. To further validate the advancement of proposed methods, we compare the detection capabilities of different methods in unseen scenarios. + +Specifically, we train different models using the NUDT-SIRST dataset and then test their detection performance using the unseen NUAA-SIRST dataset. Tab. 2 presents the detection results of different models on the unseen dataset. It can be seen that our method outperforms the state-of-the-art models in terms of IoU, $\mathrm{P}_d$ and $\mathrm{F}_a$ , implying better detection accuracy and scene generalization. + +Table 2. Detection performance of different methods for the unseen dataset. + +
MethodNUDT-SIRST (Train)
NUAA-SIRST (Test)
IoU(%)Pd(%)Fa(10-6)
DNANet [9]35.8076.14141.3
UIUNet [21]53.3280.7335.15
RDIAN [18]39.3577.9843.68
DMFNet [6]43.7079.8160.07
SCTrans [29]50.9783.4837.71
Text-IRSTD (Ours)54.5187.1529.86
+ +![](images/c7f75ae26b84ebd43b3c8a5af085f4cc2386cb5095160628f0e017f5d7e532e6.jpg) +Figure 5. Visual results of different IRSTD methods. The enlarged view of the target is shown in image corners. The red, blue, and yellow boxes represent correctly detected targets, missed targets, and false detections, respectively. + +# 5.6. Ablation Study + +Impact of Text Prompt: In order to verify the effectiveness of fuzzy text prompts, we designed three textual prompt variants for experiments and the results are shown in Tab. 3. It can be seen that the use of [Interested Region] target as a prompt to describe the IR small target helps to avoid the wrong text-image matching relationship caused by the use of explicit categories of prompts. Meanwhile, the description of [Scene] background further improves the detection performance in complex scenes by establishing more appropriate semantic associations. More importantly, experiments demonstrate the possibility of improving IRSTD performance with text and provide an effective technical path for further research. + +Impact of PCSID: To validate the effectiveness of each component in PCSID, we perform several ablation studies on the IRSTD-1k, and the results are shown in Tab. 4. The results of the experiment indicate that both TGSI block and TGFA block improve the performance of the baseline model, and when use together, the model achieve the best results. This further confirms the importance of their syn + +ergistic effect in enhancing the performance of small target detection. + +Impact of TGFA: To demonstrate the effectiveness of TGFA block, we introduce three network variants for comparison, as shown in Tab. 5. It can be seen that proposed TGFA block achieves the best performance. The reason is that text features as modulation information can optimize the complementarity and interaction between cross-layer features. Moreover, DPM and CA can better perceive and enhance important visual features, such as contrast, which facilitate visual feature fusion and subsequent cross-modal information interaction. + +Impact of TGSI Block: We use four different text-image feature interaction methods to train the Text-IRSTD. As shown in Tab. 6, simply concatenating text and image features along the channel dimension ignores semantic inconsistencies across modalities, leading to inadequate semantic alignment. Without gain parameter $\alpha_{t}$ , the model cannot dynamically focus on the regions of interest indicated by the text prompt. Similarly, without offset parameter $\beta_{t}$ , the model cannot make fine adjustments to the image features using the text features, thus limiting the effective + +![](images/3bf04e86d4cd1b77d4130bdad29c70bbfcf1e731f615c910984f5314bbf04bf1.jpg) +Figure 6. 3D visualization results corresponding to those in Fig. 5. Different colors represent distinct grayscale values, gradually increasing from blue to yellow. + +Table 3. Ablation study of different prompts. + +
TypePromptIoU(%)Pd(%)Fa(10-6)
Conventional promptsA photo of a [class]62.2290.5830.36
Variants of conventionA photo of a [class] in the [Scene] background66.6589.8921.69
Variants of OursA [Interested Region] target in the [Scene] background69.0191.6818.86
OursA photo of a [Interested Region] target in the [Scene] background69.5792.5914.97
+ +Table 4. Ablation study of the TGFA block and TGSI block. + +
BaselineTGFATGSIIoU(%)Pd(%)Fa(10-6)
✓61.6789.2257.97
✓✓67.7591.8719.83
✓✓66.1991.9226.39
✓✓✓69.5792.5914.97
+ +Table 5. Ablation study of the TGFA block. + +
ModelIoU(%)Pd(%)Fa(10-6)
TGFA w/o text guide68.7592.0223.68
TGFA w/o CA68.0191.2426.24
TGFA w/o DPM67.1689.5617.95
TGFA (Ours)69.5792.5914.97
+ +Table 6. Ablation study of the TGSI block. + +
ModelIoU(%)Pd(%)Fa(10-6)
Concatenation67.7390.5729.49
TGSI w/o αt66.6590.1335.85
TGSI w/o βt68.0591.2116.22
TGSI (Ours)69.5792.5914.97
+ +ness of target shape reconstruction. By combining both, the + +proposed TGSI Block achieves the best performance. + +# 6. Conclusion + +In this paper, we extend the IRSTD task and propose a text-guided IRSTD framework to address the issue that current models struggle to further improve detection accuracy by relying only on limited visual information. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. This prompt provides more text-image semantic associations that contribute to clutter suppression and target contour recovery. On the other hand, we propose PCSID to facilitate the information interaction between text and images. Moreover, we have constructed the first text-image bimodal dataset, called FZDT, to provide the basis for subsequent text-guided IRSTD research. Extensive experimental results demonstrate that our method achieves better detection performance and target contour recovery than the state-of-the-art methods. More surprisingly, the proposed Text-IRSTD demonstrates strong generalization and wide application prospects in unseen detection scenarios. + +# Acknowledgments + +This work was supported in part by the National Natural Science Foundation of China under Grant 62405060, in part by the Nature Science Foundation of Fujian Province under Grant 2022J05113 and 2024J01245, and in part by the Educational Research Program for Young and Middle-aged Teachers of Fujian Province under Grant JAT210035. + +# References + +[1] Liqiong Chen, Tong Wu, Shuyuan Zheng, Zhaobing Qiu, and Feng Huang. Robust unsupervised multifeature representation for infrared small target detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:10306-10323, 2024. 1, 2, 5, 6 +[2] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Asymmetric contextual modulation for infrared small target detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 950-959, 2021. 1, 2, 5, 6 +[3] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Attentional local contrast networks for infrared small target detection. IEEE transactions on geoscience and remote sensing, 59(11):9813-9824, 2021. 5, 6 +[4] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE transactions on image processing, 22(12):4996-5009, 2013. 5, 6 +[5] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2, 3 +[6] Tan Guo, Baojiang Zhou, Fulin Luo, Lei Zhang, and Xinbo Gao. Dmfnet: Dual-encoder multi-stage feature fusion network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-14, 2024. 5, 6 +[7] Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Ye Zhao, Hao-liang Zheng, and Wei Zhang. Ristdnet: Robust infrared small target detection network. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2021. 2 +[8] Xuan Kong, Chunping Yang, Siying Cao, Chaohai Li, and Zhenming Peng. Infrared small target detection via nonconvex tensor fibered rank approximation. IEEE Transactions on Geoscience and Remote Sensing, 60:1-21, 2022. 1, 2 +[9] Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo. Dense nested attention network for infrared small target detection. IEEE Transactions on Image Processing, 32:1745-1758, 2022. 2, 5, 6 +[10] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2 + +[11] Yongsong Li, Zhengzhou Li, Yu Shen, and Jie Li. Infrared small target detection based on 1-d difference of guided filtering. IEEE Geoscience and Remote Sensing Letters, 20: 1-5, 2023. 1, 2 +[12] Fanzhao Lin, Kexin Bao, Yong Li, Dan Zeng, and Shiming Ge. Learning contrast-enhanced shape-biased representations for infrared small target detection. IEEE Transactions on Image Processing, 33:3047-3058, 2024. 1 +[13] Qiankun Liu, Rui Liu, Bolun Zheng, Hongkui Wang, and Ying Fu. Infrared small target detection with scale and location sensitivity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17490-17499, 2024. 1 +[14] Yuan Luo, Xiaorun Li, Shuhan Chen, Chaoqun Xia, and Liaoying Zhao. Imnn-lwec: A novel infrared small target detection based on spatial-temporal tensor model. IEEE Transactions on Geoscience and Remote Sensing, 60:1-22, 2022. 1, 2 +[15] Yao Qin, Lorenzo Bruzzone, Chengqiang Gao, and Biao Li. Infrared small target detection based on facet kernel and random walker. IEEE Transactions on Geoscience and Remote Sensing, 57(9):7104-7118, 2019. 5, 6 +[16] Zhaobing Qiu, Yong Ma, Fan Fan, Jun Huang, and Lang Wu. Global sparsity-weighted local contrast measure for infrared small target detection. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 5, 6 +[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5 +[18] Heng Sun, Junxiang Bai, Fan Yang, and Xiangzhi Bai. Receptive-field and direction induced attention network for infrared dim small target detection with a large-scale dataset irdst. IEEE Transactions on Geoscience and Remote Sensing, 61:1-13, 2023. 5, 6 +[19] Xiaozhong Tong, Zhen Zuo, Shaojing Su, Junyu Wei, Xiaoyong Sun, Peng Wu, and Zongqing Zhao. St-trans: Spatial-temporal transformer for infrared small target detection in sequential images. IEEE Transactions on Geoscience and Remote Sensing, 62:1-19, 2024. 1 +[20] Keyan Wang, Xueyan Wu, Peicheng Zhou, Zuntian Chen, Rui Zhang, Liyun Yang, and Yunsong Li. Afe-net: Attention-guided feature enhancement network for infrared small target detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:4208-4221, 2024. 1 +[21] Xin Wu, Danfeng Hong, and Jocelyn Chanussot. Uiu-net: U-net in u-net for infrared small object detection. IEEE Transactions on Image Processing, 32:364-376, 2022. 2, 5, 6 +[22] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7031-7040, 2023. 2 +[23] Yunhong Xin, Jiao Zhou, and Yishuan Chen. Dual multiscale filter with SSS and GW for infrared small target detec + +tion. Infrared Physics & Technology, 81:97-108, 2017. 1, 2 +[24] Yunkai Xu, Minjie Wan, Xiaojie Zhang, Jian Wu, Yili Chen, Qian Chen, and Guohua Gu. Infrared small target detection based on local contrast-weighted multidirectional derivative. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-16, 2023. 1, 2 +[25]Fuju Yan,Guili Xu,Junpu Wang,Quan Wu,and Zhengsheng Wang. Infrared small target detection via schatten capped pnorm-based non-convex tensor low-rank approximation.IEEE Geoscience and Remote Sensing Letters,20: 1-5,2023.1,2 +[26] Bo Yang, Xinyu Zhang, Jian Zhang, Jun Luo, Mingliang Zhou, and Yangjun Pi. Eflnet: Enhancing feature learning network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-11, 2024. 5 +[27] Xinyi Ying, Li Liu, Yingqian Wang, Ruojing Li, Nuo Chen, Zaiping Lin, Weidong Sheng, and Shilin Zhou. Mapping degeneration meets label evolution: Learning infrared small target detection with single point supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15528-15538, 2023. 1 +[28] Chuang Yu, Yunpeng Liu, Shuhang Wu, Xin Xia, Zhuhua Hu, Deyan Lan, and Xin Liu. Pay attention to local contrast learning networks for infrared small target detection. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 2 +[29] Shuai Yuan, Hanlin Qin, Xiang Yan, Naveed Akhtar, and Ajmal Mian. Sctransnet: Spatial-channel cross transformer network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-15, 2024. 1, 5, 6 +[30] Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R Manmatha, et al. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2736–2746, 2022. 3 +[31] Landan Zhang and Zhenming Peng. Infrared small target detection based on partial sum of the tensor nuclear norm. Remote Sensing, 11(4):382, 2019. 5, 6 +[32] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 877-886, 2022. 1, 5 +[33] Tianfang Zhang, Lei Li, Siying Cao, Tian Pu, and Zhenming Peng. Attention-guided pyramid context networks for detecting infrared small target under complex background. IEEE Transactions on Aerospace and Electronic Systems, 59(4): 4250-4261, 2023. 1, 2, 5, 6 +[34] Xiangyue Zhang, Jingyu Ru, and Chengdong Wu. Infrared small target detection based on gradient correlation filtering and contrast measurement. IEEE Transactions on Geoscience and Remote Sensing, 61:1-12, 2023. 1, 2 +[35] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16793-16803, 2022. 2, 3 \ No newline at end of file diff --git a/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/images.zip b/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b7be8c464afdb29c23493d4889c6e0c2f48b291a --- /dev/null +++ b/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c66110dc6a5028b173060f4b7ede1f82b36f7ae13da53593d5024b2b430052fc +size 688365 diff --git a/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/layout.json b/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c7cd6d991d360483d5061b8c17d2e43aac039fa1 --- /dev/null +++ b/2025/Text-IRSTD_ Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes/layout.json @@ -0,0 +1,7782 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 78, + 102, + 533, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 102, + 533, + 140 + ], + "spans": [ + { + "bbox": [ + 78, + 102, + 533, + 140 + ], + "type": "text", + "content": "Text-IRSTD: Leveraging Semantic Text to Promote Infrared Small Target Detection in Complex Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 160, + 531, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 160, + 531, + 189 + ], + "spans": [ + { + "bbox": [ + 78, + 160, + 531, + 189 + ], + "type": "text", + "content": "Feng Huang*, Shuyuan Zheng*, Zhaobing Qiu*,†, Huanxian Liu, Huanxin Bai, Liqiong Chen† Fuzhou University, Fuzhou, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 191, + 496, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 191, + 496, + 217 + ], + "spans": [ + { + "bbox": [ + 106, + 191, + 496, + 217 + ], + "type": "text", + "content": "huangf@fzu.edu.cn, zhengshuyuan200147@gmail.com, qiuzhaobing@fzu.edu.cn, liuhuanxian10@gmail.com, mw136296@gmail.com, liqionchen@fzu.edu.cn." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 245, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 245, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 245, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 296, + 582 + ], + "type": "text", + "content": "Infrared small target detection is currently a hot and challenging task in computer vision. Existing methods usually focus on mining visual features of targets, which struggles to cope with complex and diverse detection scenarios. The main reason is that infrared small targets have limited image information on their own, thus relying only on visual features fails to discriminate targets and interferences, leading to lower detection performance. To address this issue, we introduce a novel approach leveraging semantic text to guide infrared small target detection, called Text-IRSTD. It innovatively expands classical IRSTD to text-guided IRSTD, providing a new research idea. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. On the other hand, we propose a progressive cross-modal semantic interaction decoder (PCSID) to facilitate information fusion between texts and images. In addition, we construct a new benchmark consisting of 2,755 infrared images of different scenarios with fuzzy semantic textual annotations, called FZDT. Extensive experimental results demonstrate that our method achieves better detection performance and target contour recovery than the state-of-the-art methods. Moreover, proposed Text-IRSTD shows strong generalization and wide application prospects in unseen detection scenarios. The dataset and code are available at https://github.com/Zhengsy0407/Text-IRSTD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 624, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 685 + ], + "type": "text", + "content": "Infrared small target detection (IRSTD) has become an important and challenging issue in computer vision, which is widely used in both civil and military applications [12, 13, 19]. However, due to the long range detection needs, imaged targets usually occupy only a few pixels and lack tex" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 246, + 555, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 282 + ], + "type": "text", + "content": "ture details [20, 27, 32]. Therefore, how to accurately detect infrared (IR) small targets from complex backgrounds remains an open issue." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 293, + 566, + 460 + ], + "blocks": [ + { + "bbox": [ + 315, + 293, + 566, + 460 + ], + "lines": [ + { + "bbox": [ + 315, + 293, + 566, + 460 + ], + "spans": [ + { + "bbox": [ + 315, + 293, + 566, + 460 + ], + "type": "image", + "image_path": "ebbd5bdf2e26354bb118083726691eb37d4388e61f010ef790af37280950c44f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 472, + 556, + 518 + ], + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 518 + ], + "type": "text", + "content": "Figure 1. Visual comparison between the SOTA model and proposed Text-IRSTD. (a) Detection results of the SOTA model [29], which relies only on visual features. (b) Detection results of proposed Text-IRSTD, which uses both textual and visual features." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "In the early days, many traditional methods emerged to improve detection accuracy. They mainly consist of filter-based methods [11, 23], local saliency-based methods [1, 24, 34] and sparse representation-based methods [8, 14, 25]. However, reliance on artificial feature representations prevents them from further mining the deep semantic information of real targets. In contrast, deep learning (DL)-based methods autonomously learn multi-layered features in a data-driven manner and achieve significant performance improvements. For example, Dai et al. [2] propose asymmetric contextual modulation (ACM) module to better interact with high-level semantics and low-level detail features. For higher feature utilization, Zhang et al. [33] propose attention-guided pyramid context network (AGPC-Net), which mines the correlations between feature map" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10635" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "patches. In addition, researchers have also attempted to fuse prior information for higher detection accuracy [7, 28]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 96, + 294, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 96, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 54, + 96, + 294, + 239 + ], + "type": "text", + "content": "However, limited by scarce image features, existing methods fail to perform satisfactory in complex scenes. As shown in Fig. 1, the SOTA model still suffers from many false detections due to limited ability to discriminate between targets and interference. Recently, with the development of large-scale visual language models [17], semantic text-image-based models [10, 22] have made impressive progress in generic target detection. However, applying them directly to infrared small target detection works poorly. The main reason is that ambiguous target categories in infrared images invalidate existing data annotation and feature fusion methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 240, + 296, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 296, + 600 + ], + "type": "text", + "content": "To address above challenges, this paper proposes a novel approach leveraging semantic text to guide infrared small target detection, called Text-IRSTD. It innovatively expands classical IRSTD to text-guided IRSTD, providing a new idea for IRSTD. As shown in Fig. 1, by utilizing both text and image information, proposed Text-IRSTD can further suppress clutter interference and improve detection performance. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. This prompt provides more text-image semantic associations that contribute to clutter suppression and target contour recovery. On the other hand, we propose a progressive cross-modal semantic interaction decoder (PCSID) that fuses both text and image information to further improve the detection performance. PCSID consists of two core components: text-guided feature aggregation (TGFA) block and text-guided semantic interaction (TGSI) block. Specifically, TGFA first utilizes text information to guide cross-layer visual feature aggregation for key feature extraction. Where key features (e.g., contrast [24]) refer to those features that are more important than others (e.g., intensity) for IRSTD. Then, TGSI block simultaneously performs feature modulation from two different perspectives, thus realizing a more refined text-image information interaction. In addition, we construct a new benchmark consisting of 2,755 infrared images of different scenarios with fuzzy semantic textual annotations, called FZDT. Extensive experiments on multiple public datasets demonstrate that our approach achieves better detection performance and generalization than representative state-of-the-art algorithms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 609, + 142, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 142, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 142, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 628, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 665 + ], + "type": "text", + "content": "In this section, we first briefly review existing methods for IRSTD. Then, an introduction to text-image models in other domains is presented." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 155, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 155, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 155, + 683 + ], + "type": "text", + "content": "2.1. IRSTD Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "In the early days, many traditional methods emerged to improve detection accuracy. They mainly consist of filter" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 358 + ], + "type": "text", + "content": "based methods [11, 23], local saliency-based methods [1, 24, 34] and sparse representation-based methods [8, 14, 25]. However, reliance on artificial feature representations prevents them from further mining the deep semantic information of real targets. In contrast, deep learning (DL)-based methods autonomously learn multi-layered features in a data-driven manner and have made some impressive progresses recently. To mitigate feature loss of IR small targets, several researchers focus on improving the cross-layer feature fusion approach. For example, Dai et al. [2] propose an asymmetric contextual modulation (ACM) module to better interact with high-level semantics and low-level detail features. For higher feature utilization, Zhang et al. [33] propose the attention-guided pyramid context network (AGPCNet), which mines the correlations between feature map patches. In order to interact cross-layer features incrementally, Li et al. [9] propose dense nested attention network (DNA-Net). Several researchers also attempted to improve the model structure. For example, Wu et al. [21] embed a smaller U-Net into the U-Net backbone to learn target features from multiple levels and scales. In addition, Hou et al. [7] try to introduce prior information and design a robust infrared small target detection network (RISTDnet) combining artificial features and neural networks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 361, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 468 + ], + "type": "text", + "content": "Existing methods have made many impressive progress in IRSTD. However, the visual features of IR small targets are quite limited, which renders their performance unsatisfactory in complex and diverse detection scenarios. In contrast, we explore a new research idea to introduce text semantic features into IRSTD and devise a novel text-image fusion detection method. Utilizing added textual information, proposed Text-IRSTD can further suppress background interference and improve detection performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 479, + 427, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 427, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 427, + 491 + ], + "type": "text", + "content": "2.2. Text-Image Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 553, + 605 + ], + "type": "text", + "content": "With the advancement of deep learning techniques and the support of large datasets, text-image models has achieved success. In particular, CLIP [17] utilizes 400 million matched pairs datasets and contrast learning techniques to demonstrate powerful performance and wide application prospects in multiple domains and tasks. CLIP significantly improves text-image alignment and avoids the reliance on explicit labeling. With the support of CLIP models, text-driven generic target detection has also developed rapidly." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": "For better detection performance, RegionCLIP [35] leverages CLIP to match image regions with template captions, achieving fine-grained alignment between image and text features. ViLD [5] designs a training method based on knowledge distillation, which distills the knowledge from a pretrained open-vocabulary image classification model into a two-stage detector. CORA [22] mitigates the distribution gap between overall image features and regional features when using CLIP directly for detector training by introduc" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10636" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 149, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 149, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 149, + 84 + ], + "type": "text", + "content": "ing learnable prompts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 296, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 296, + 170 + ], + "type": "text", + "content": "Although existing text-image models for generic target detection are illuminating, they are difficult to apply to IRSTD tasks. The reason is that it is difficult to establish appropriate text-image semantic associations and interactions for IR small targets with ambiguous categories. Therefore, it is important and challenging to investigate text-image detection models dedicated to IRSTD tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 280, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 280, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 280, + 201 + ], + "type": "text", + "content": "3. Fuzzy Semantic Text Prompt For IRSTD" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 209, + 296, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 424 + ], + "type": "text", + "content": "Semantic text prompts aim to establish a proper matching relationship between text and images, which is crucial for cross-modal target detection. In generic target detection [5, 35], researchers usually directly describe specific categories of targets, such as \"A photo of a girl\" and \"A photo of a soccer\" in Fig. 2(a). In this way, it is possible to focus not only on the visual features of a single target, but also to further explore semantic relations between multiple targets, e.g. a small round target near a girl is more likely to be a soccer. However, IR small targets typically occupy only a few pixels and lack visual features such as texture. As shown in Fig. 2(b), real targets are usually presented as bright spots that cannot be further identified as a specific category such as an aircraft or a bird. Therefore, using existing text prompts to describe ambiguous IR small target categories will lead to incorrect text-image matching relationships, impairing detection accuracy and model generalizability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 714 + ], + "type": "text", + "content": "To address this issue, we propose a novel fuzzy semantic text prompt for IRSTD, which includes two templates: scenario description and task description. Specifically, for known scenes with semantic information, we use the scenario description and describe one IR image as “A photo of a [Interested Region] target in the [Scene] background”, where [Interested Region] denotes the region where targets need to be detected, e.g., sky target and ground target, and [Scene] denotes the detection scene, e.g., sky-ground background, ocean-sky background. For example, we describe the IR image Fig. 2(b) as: “A photo of a sky target in the sky and ground background”, which not only establishes correct and robust matching relationships, but also provides more semantic associations. In this way, the model can focus on more shared visual features of sky targets, such as shape contours, which helps target shape recovery and detection. Moreover, as shown in Fig. 1, more semantic associations are established and utilized to further suppress complex ground interference and improve detection accuracy. For unseen scenes without semantic information, we use the task description, i.e., the default text prompt is “A photo of a small target in the background”." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 566, + 152 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 566, + 152 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 566, + 152 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 566, + 152 + ], + "type": "image", + "image_path": "3c93771244c902d6f7ce0c8add431b5339216eedb7f9ecd02a04f93bbdb5de0a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 154, + 511, + 163 + ], + "lines": [ + { + "bbox": [ + 369, + 154, + 511, + 163 + ], + "spans": [ + { + "bbox": [ + 369, + 154, + 511, + 163 + ], + "type": "text", + "content": "(a) Typical prompt template for object detection" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 166, + 564, + 268 + ], + "blocks": [ + { + "bbox": [ + 316, + 166, + 564, + 268 + ], + "lines": [ + { + "bbox": [ + 316, + 166, + 564, + 268 + ], + "spans": [ + { + "bbox": [ + 316, + 166, + 564, + 268 + ], + "type": "image", + "image_path": "baaf457a050d1fa96472c3237b42a17d6217b0aa5abb75d08af7403f2b6bd15a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 270, + 511, + 279 + ], + "lines": [ + { + "bbox": [ + 369, + 270, + 511, + 279 + ], + "spans": [ + { + "bbox": [ + 369, + 270, + 511, + 279 + ], + "type": "text", + "content": "(b) Proposed fuzzy semantic prompt for IRSTD" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 291, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 314 + ], + "type": "text", + "content": "Figure 2. (a) Typical semantic text prompt for generic target detection. (b) Proposed fuzzy semantic text prompt for IRSTD." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 320, + 370, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 370, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 370, + 333 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 341, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 401 + ], + "type": "text", + "content": "In this section, we first present the overall architecture of proposed Text-IRSTD. Then, Sections 4.2 and 4.3 introduce two core components: TGFA block for aggregating key visual features and TGSI block for facilitating text-image information interaction, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 407, + 434, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 434, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 434, + 419 + ], + "type": "text", + "content": "4.1. Overall Architecture" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "type": "text", + "content": "Fig. 3 illustrates the overall structure of Text-IRSTD. First, one infrared image " + }, + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "type": "inline_equation", + "content": "I_{img}" + }, + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "type": "text", + "content": " and the corresponding text prompt " + }, + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "type": "inline_equation", + "content": "I_{text}" + }, + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "type": "text", + "content": " are fed into the image encoder " + }, + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "type": "text", + "content": " and text encoder " + }, + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 313, + 426, + 555, + 473 + ], + "type": "text", + "content": ", respectively, as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 394, + 480, + 553, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 480, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 394, + 480, + 553, + 499 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E _ {i m g} = \\mathcal {F} _ {i} (I _ {i m g}), \\\\ \\Gamma_ {i m g} = \\mathcal {F} _ {i} (I _ {i m g}). \\end{array} \\tag {1}", + "image_path": "34262690c224fc59e859f7c7ae0aca8fe91f90dd5460849a615046098e20d41f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 397, + 497, + 473, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 497, + 473, + 508 + ], + "spans": [ + { + "bbox": [ + 397, + 497, + 473, + 508 + ], + "type": "interline_equation", + "content": "E _ {t e x t} = \\mathcal {F} _ {t} (I _ {t e x t}),", + "image_path": "0da05a58162ac2fa2158b6fd36c11c412f017d131277a548f90cbbd5c58d993b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 516, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 555, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 516, + 555, + 563 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 313, + 516, + 555, + 563 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 516, + 555, + 563 + ], + "type": "inline_equation", + "content": "E_{text}" + }, + { + "bbox": [ + 313, + 516, + 555, + 563 + ], + "type": "text", + "content": " denote the image and text encoder outputs respectively. Note that " + }, + { + "bbox": [ + 313, + 516, + 555, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 313, + 516, + 555, + 563 + ], + "type": "text", + "content": " consists of four groups of ResNeSt [30] blocks, while " + }, + { + "bbox": [ + 313, + 516, + 555, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t" + }, + { + "bbox": [ + 313, + 516, + 555, + 563 + ], + "type": "text", + "content": " denote the pre-trained language model CLIP with frozen weights." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 563, + 554, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 599 + ], + "type": "text", + "content": "Then, " + }, + { + "bbox": [ + 313, + 563, + 554, + 599 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 313, + 563, + 554, + 599 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 563, + 554, + 599 + ], + "type": "inline_equation", + "content": "E_{text}" + }, + { + "bbox": [ + 313, + 563, + 554, + 599 + ], + "type": "text", + "content": " are fed into PCSID for text-image feature interaction and enhancement. The result of cross-modal interaction " + }, + { + "bbox": [ + 313, + 563, + 554, + 599 + ], + "type": "inline_equation", + "content": "D_{cm}" + }, + { + "bbox": [ + 313, + 563, + 554, + 599 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 372, + 608, + 553, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 608, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 372, + 608, + 553, + 620 + ], + "type": "interline_equation", + "content": "D _ {c m} = \\operatorname {P C S I D} \\left(E _ {i m g}, E _ {t e x t}\\right), \\tag {2}", + "image_path": "1ef6123df263c5a24c93389b8b0d7f8fa9b88b182912d8e113bc1193f32fe03b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 628, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 554, + 689 + ], + "type": "text", + "content": "where PCSID consists of three groups of TGFA blocks and TGSI blocks in series. TGFA and TGSI will be introduced in Sections 4.2 and 4.3, respectively. Note that the fourth layer output " + }, + { + "bbox": [ + 313, + 628, + 554, + 689 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{img}}^{(4)}" + }, + { + "bbox": [ + 313, + 628, + 554, + 689 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 628, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 313, + 628, + 554, + 689 + ], + "type": "text", + "content": " is enhanced by detail perception module (DPM) and then fed to PCSID." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Finally, we apply a " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " convolution " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{c}^{1}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " and sigmoid function to the decoder outputs " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "D_{cm}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": ". Thus, the cross-modal" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10637" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 558, + 259 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 558, + 259 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 558, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 558, + 259 + ], + "type": "image", + "image_path": "2558a3d70209199205f0131ab5cf4506674ec914a6dc7c4a69d479ada9ea695b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 266, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 301 + ], + "type": "text", + "content": "Figure 3. Overview of the proposed Text-IRSTD. It consists of three main components: text encoder, image encoder, and cross-modal decoder PCSID with their outputs represented as " + }, + { + "bbox": [ + 54, + 266, + 555, + 301 + ], + "type": "inline_equation", + "content": "E_{text}" + }, + { + "bbox": [ + 54, + 266, + 555, + 301 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 266, + 555, + 301 + ], + "type": "inline_equation", + "content": "E_{img}^{(i)}" + }, + { + "bbox": [ + 54, + 266, + 555, + 301 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 266, + 555, + 301 + ], + "type": "inline_equation", + "content": "D_{cm}" + }, + { + "bbox": [ + 54, + 266, + 555, + 301 + ], + "type": "text", + "content": ", respectively. Where DPM denotes detail perception module, TGFA Block denotes text-guided feature aggregation block and TGSI Block denotes text-guided semantic interaction block." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 313, + 253, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 253, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 253, + 326 + ], + "type": "text", + "content": "detection result " + }, + { + "bbox": [ + 55, + 313, + 253, + 326 + ], + "type": "inline_equation", + "content": "O_{cm}" + }, + { + "bbox": [ + 55, + 313, + 253, + 326 + ], + "type": "text", + "content": " can be obtained as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 335, + 295, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 335, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 116, + 335, + 295, + 350 + ], + "type": "interline_equation", + "content": "O _ {c m} = \\operatorname {S i g m o i d} \\left(f _ {c} ^ {1} \\left(D _ {c m}\\right)\\right). \\tag {3}", + "image_path": "155c4796718dd06ba63ccf4dcb88fa4cde401fcb54dfba39fd5b5468b6fc5258.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 263, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 263, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 263, + 373 + ], + "type": "text", + "content": "4.2. Text-guided Feature Aggregation Block" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 378, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 486 + ], + "type": "text", + "content": "Effective aggregation of visual features is crucial for detection models. Limited by scarce visual information, existing methods struggle to capture key features from high-level and low-level feature maps. To this end, we propose the novel text-guided cross-layer feature aggregation block TGFA (shown in Fig. 4). TGFA transforms text features " + }, + { + "bbox": [ + 55, + 378, + 296, + 486 + ], + "type": "inline_equation", + "content": "E_{text}" + }, + { + "bbox": [ + 55, + 378, + 296, + 486 + ], + "type": "text", + "content": " into modulation information through MLP, guiding the model to capture key features that are easily neglected in complex scenes. The details are as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 108, + 497, + 295, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 497, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 108, + 497, + 295, + 510 + ], + "type": "interline_equation", + "content": "F _ {t, 1}, F _ {t, 2} = \\operatorname {S p i l t} \\left(\\operatorname {M L P} \\left(E _ {\\text {t e x t}}\\right)\\right), \\tag {4}", + "image_path": "d4caef792eda45ceb8d32df83c900fa1d6ee7327da0d87c84d998e36939dc317.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 521, + 295, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 521, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 105, + 521, + 295, + 538 + ], + "type": "interline_equation", + "content": "F _ {f} ^ {l} = F _ {l} \\otimes F _ {t, 1}, F _ {f} ^ {h} = F _ {h} \\otimes F _ {t, 2}, \\tag {5}", + "image_path": "d609fe3ab9f4e59af7cc94c2f751271de0c638e6e84b3bba59e4a4656f4301d2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "inline_equation", + "content": "F_{h}" + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "content": " denote low-level feature and high-level feature, " + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "inline_equation", + "content": "F_{f}^{l}" + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "inline_equation", + "content": "F_{f}^{h}" + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "content": " denote the fusion features after text modulation, " + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{Spit}(\\cdot)" + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "content": " denotes splitting the feature into two equal groups along the channel dimension, " + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 55, + 544, + 296, + 604 + ], + "type": "text", + "content": " denotes element-wise multiplication." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": "Low-level features contain rich target locations and details, but are easily lost at deep layers because small IR targets occupy only a few pixels. Therefore, we devise the detail perception module DPM as shown in Fig. 4. Specifically, for a given low-level fusion feature " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{f}^{l}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": ", a " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " convolution is first applied to expand the channel dimension by a factor of " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": ", and then divide the resulting feature map into two groups " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{l,1} \\in \\mathbb{R}^{\\frac{\\mu C}{2} \\times H \\times W}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{l,2} \\in \\mathbb{R}^{\\frac{\\mu C}{2} \\times H \\times W}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " along the channel dimension." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 315, + 556, + 449 + ], + "blocks": [ + { + "bbox": [ + 316, + 315, + 556, + 449 + ], + "lines": [ + { + "bbox": [ + 316, + 315, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 556, + 449 + ], + "type": "image", + "image_path": "70253ee2c60fa7c4fa46f28876b041049a1cf2953ca66ed2da13be1d445dffb2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 455, + 555, + 478 + ], + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 478 + ], + "type": "text", + "content": "Figure 4. Structure of TGFA block, where DPM denotes detail perception module and CA denotes channel attention." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "text", + "content": "Then, we use " + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "text", + "content": " depth-wise convolutions to capture important local features such as contrast at different scales and concatenate these two groups of features along the channel dimension. The multi-scale detailed feature mapping " + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "inline_equation", + "content": "F_{ms}" + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "text", + "content": " is obtained from the following equation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 378, + 552, + 553, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 552, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 378, + 552, + 553, + 567 + ], + "type": "interline_equation", + "content": "F _ {l, 1}, F _ {l, 2} = \\operatorname {S p l i t} \\left(f _ {c} ^ {1} \\left(F _ {f} ^ {l}\\right)\\right), \\tag {6}", + "image_path": "a0a69a430d9a6549d0a9a1b05b7bfe0a985fddde9b8605d42901c84ef11a6b81.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 575, + 553, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 575, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 339, + 575, + 553, + 590 + ], + "type": "interline_equation", + "content": "F _ {m s} = f _ {c} ^ {1} \\left(\\operatorname {C o n t a c t} \\left(f _ {d w c} ^ {3} \\left(F _ {l, 1}\\right), f _ {d w c} ^ {5} \\left(F _ {l, 2}\\right)\\right)\\right), \\tag {7}", + "image_path": "228f1ecb2339fe731cd00fe3d517c03041dec401571cfdf5e8abc36d875b4203.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\operatorname{Contact}(\\cdot)" + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "text", + "content": " denotes concatenating the two groups of features along the channel dimension, " + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "inline_equation", + "content": "f_{c}^{1}" + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "text", + "content": " convolution, " + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "inline_equation", + "content": "f_{dwc}^{3}" + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "inline_equation", + "content": "f_{dwc}^{5}" + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 313, + 595, + 555, + 642 + ], + "type": "text", + "content": " depth-wise convolutions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 643, + 555, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 555, + 690 + ], + "type": "text", + "content": "After obtaining the multi-scale detailed feature mapping " + }, + { + "bbox": [ + 313, + 643, + 555, + 690 + ], + "type": "inline_equation", + "content": "F_{ms}" + }, + { + "bbox": [ + 313, + 643, + 555, + 690 + ], + "type": "text", + "content": ", we use point-wise convolution to aggregate the contextual features at each spatial location. For simplicity, it can be described as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 320, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "F _ {D P} = \\sigma \\left(\\operatorname {P W C o n v} _ {2} \\left(\\gamma \\left(\\operatorname {P W C o n v} _ {1} \\left(F _ {m s}\\right)\\right)\\right) \\otimes F _ {m s}, \\right. \\tag {8}", + "image_path": "f2a047309327898fc7844759277feaf0ac81dc6e82bf615dd393057f8de5ea93.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10638" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "\\mathrm{PWConv}_1" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "\\mathrm{PWConv}_2" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": " denote the point-wise convolution with kernel sizes of " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "\\frac{C}{4} \\times C \\times 1 \\times 1" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "C \\times \\frac{C}{4} \\times 1 \\times 1" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": " represent the sigmoid activation function and rectified linear unit." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "text", + "content": "Moreover, high-level features can provide more accurate semantic information, while adjusting the weights of different channels can adaptively enhance the representation of key information and suppress redundant features. Therefore, for a given high-level fusion features " + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "inline_equation", + "content": "F_{f}^{h}" + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "text", + "content": ", we first use global average pooling on each channel to obtain the global feature representation " + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "inline_equation", + "content": "z_{c} \\in \\mathbb{R}^{C \\times 1 \\times 1}" + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "text", + "content": ", which can be described as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 108, + 224, + 295, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 224, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 108, + 224, + 295, + 258 + ], + "type": "interline_equation", + "content": "z _ {c} = \\frac {1}{H \\times W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} F _ {f} ^ {h} [:, i, j ]. \\tag {9}", + "image_path": "6c8874aead843f83d2ebaef8898c0ed3fc988a6b801ce68eb54fb653a3e451f3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 267, + 295, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 295, + 303 + ], + "type": "text", + "content": "Then, a nonlinear mapping of this global information is performed through two fully connected layers to establish dependencies between the channels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 304, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 342 + ], + "type": "text", + "content": "Finally, the channel feature maps are normalized using the sigmoid function, and an element-wise multiplication is performed with the " + }, + { + "bbox": [ + 55, + 304, + 296, + 342 + ], + "type": "inline_equation", + "content": "F_{f}^{h}" + }, + { + "bbox": [ + 55, + 304, + 296, + 342 + ], + "type": "text", + "content": ", the output of CA is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 351, + 295, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 351, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 295, + 369 + ], + "type": "interline_equation", + "content": "F _ {C A} = \\sigma \\left(F c _ {2} \\left(F c _ {1} \\left(z _ {c}\\right)\\right)\\right) \\otimes F _ {f} ^ {h}, \\tag {10}", + "image_path": "d5cfa979ec211a7d1801500c7755ca69364601496b8134cb71e9eeba1b027c57.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 376, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 376, + 295, + 413 + ], + "type": "inline_equation", + "content": "Fc_{1} \\in \\mathbb{R}^{\\frac{C}{r} \\times C}" + }, + { + "bbox": [ + 55, + 376, + 295, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 376, + 295, + 413 + ], + "type": "inline_equation", + "content": "Fc_{2} \\in \\mathbb{R}^{C \\times \\frac{C}{r}}" + }, + { + "bbox": [ + 55, + 376, + 295, + 413 + ], + "type": "text", + "content": " represent the two fully connected operations, " + }, + { + "bbox": [ + 55, + 376, + 295, + 413 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 376, + 295, + 413 + ], + "type": "text", + "content": " denotes the channel reduction rate." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 423, + 264, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 264, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 264, + 435 + ], + "type": "text", + "content": "4.3. Text-guided Semantic Interaction Block" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "text", + "content": "Due to the limited visual information, existing models struggle to effectively suppress complex interference and recover target contours. Therefore, we introduce semantic text prompts and propose TGSI block, which achieves better detection performance through text-image information interaction. Figure 3 shows the detailed structure of TGSI block. First, we take the image feature " + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "text", + "content": " and the text feature " + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "inline_equation", + "content": "E_{text}" + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "text", + "content": " as inputs, where " + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "inline_equation", + "content": "E_{text}" + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "text", + "content": " is obtained by feeding the text prompt " + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "inline_equation", + "content": "T_{text}" + }, + { + "bbox": [ + 55, + 441, + 295, + 571 + ], + "type": "text", + "content": " into the CLIP. Notably, we freeze the weights of the text encoder to preserve the original capabilities of the pre-trained model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 573, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 295, + 666 + ], + "type": "text", + "content": "Furthermore, we design a Feed-forward Network (FFN) to further explore the deep semantic information embedded in the text features. This helps the model capture more fine-grained semantic associations between the text, the target, and the regions of interest. Specifically, the text feature " + }, + { + "bbox": [ + 55, + 573, + 295, + 666 + ], + "type": "inline_equation", + "content": "E_{text}" + }, + { + "bbox": [ + 55, + 573, + 295, + 666 + ], + "type": "text", + "content": " is mapped into two learnable semantic parameters " + }, + { + "bbox": [ + 55, + 573, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 573, + 295, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 573, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 55, + 573, + 295, + 666 + ], + "type": "text", + "content": " after going through the FFN, which can be defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 128, + 670, + 295, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 670, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 128, + 670, + 295, + 683 + ], + "type": "interline_equation", + "content": "\\alpha_ {t}, \\beta_ {t} = \\operatorname {F F N} \\left(E _ {\\text {t e x t}}\\right). \\tag {11}", + "image_path": "e057c88f720b52ffba73ae3ab0b3cb5b357b3c9e208be906a7caaf84e2239165.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Finally, the semantic parameters " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": " modulate the image feature " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": " from the perspective of gain control and" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "offset adjustment, respectively. Specifically, the gain parameter " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " directs the network to pay more attention to the regions of interest of the text prompt and reduces false alarms in other regions. The offset parameter " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " provides finer tuning and helps in the recovery of target contour details. The fused semantic features " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " are obtained by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 380, + 154, + 553, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 154, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 380, + 154, + 553, + 167 + ], + "type": "interline_equation", + "content": "\\hat {M} = \\left(1 + \\alpha_ {t}\\right) \\otimes M + \\beta_ {t}. \\tag {12}", + "image_path": "2e512c400774c4e2ffc5a659573c2a26beb2039a1ae1b119d86d17c2a75cc419.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 178, + 395, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 395, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 395, + 192 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 199, + 489, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 489, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 489, + 210 + ], + "type": "text", + "content": "5.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 217, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 288 + ], + "type": "text", + "content": "Datasets: All experiments are conducted on proposed FZDT dataset, which contains three public image datasets NUAA-SIRST [2], NUDT-SIRST [9], IRSTD-1k [32], and corresponding fuzzy semantic text prompts. Following existing works [26], each dataset is divided into training set, validation set and test set in the ratio of " + }, + { + "bbox": [ + 313, + 217, + 553, + 288 + ], + "type": "inline_equation", + "content": "6:2:2" + }, + { + "bbox": [ + 313, + 217, + 553, + 288 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 289, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 337 + ], + "type": "text", + "content": "Evaluation Metrics: For comprehensive algorithmic comparison, the evaluation metrics include intersection over union (IoU), false alarm rate " + }, + { + "bbox": [ + 313, + 289, + 553, + 337 + ], + "type": "inline_equation", + "content": "(\\mathrm{F}_a)" + }, + { + "bbox": [ + 313, + 289, + 553, + 337 + ], + "type": "text", + "content": ", F-measure " + }, + { + "bbox": [ + 313, + 289, + 553, + 337 + ], + "type": "inline_equation", + "content": "(\\mathrm{F}_1)" + }, + { + "bbox": [ + 313, + 289, + 553, + 337 + ], + "type": "text", + "content": ", and probability of detection " + }, + { + "bbox": [ + 313, + 289, + 553, + 337 + ], + "type": "inline_equation", + "content": "(\\mathrm{P}_d)" + }, + { + "bbox": [ + 313, + 289, + 553, + 337 + ], + "type": "text", + "content": " [29]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 346, + 447, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 447, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 447, + 357 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 363, + 553, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 553, + 447 + ], + "type": "text", + "content": "To evaluate the efficacy of Text-IRSTD, we use the common CLIP-ViT-B/32 [17] as the text encoder. Our network is implemented with Pytorch on a single NVIDIA GeForce RTX 3090 GPU. We adopt the AdamW optimizer with a learning rate of 0.0001 and a weight decay of 0.05. Every images are normalized and randomly cropped to " + }, + { + "bbox": [ + 313, + 363, + 553, + 447 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 363, + 553, + 447 + ], + "type": "text", + "content": ". The batch size and epoch are set as 8 and 500, respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 448, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 554, + 543 + ], + "type": "text", + "content": "The comparison algorithms for experiments include eight state-of-the-art DL-based methods: ACMNet [2], ALCNet [3], AGPCNet [33], DNANet [9], UIUNet [21], RDIAN [18], DMFNet [6] and SCTrans [29], and five recent traditional methods: IPI [4], PSTNN [31], FKRW [15], GSWLCM [16] and RUMFR [1]. To further illustrate the advancement of our method, Text-IRSTD without text prompts is also selected as the baseline method." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 552, + 432, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 432, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 432, + 564 + ], + "type": "text", + "content": "5.3.Quantitative Results" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "The quantitative results of different methods are presented in Tab. 1. Overall, our Text-IRSTD achieves the best performance across all evaluation metrics. For example, the " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_a" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " of our method achieves as low as " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "1.032 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " on NUDT-SIRST, while also reaching the highest " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_d" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "99.73\\%" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": ", demonstrating our superior capability in suppressing complex interference. Additionally, we obtain the best IoU of " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "95.84\\%" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": ", which is " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "1.01\\%" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " higher than the second best, indicating better target contour recovery capability. Traditional methods have lower metrics and exhibit poor performance. DL-based methods can further mine deep semantic features in a data-driven manner, and thus have significant" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10639" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 96, + 555, + 304 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative results of different methods. Results for the metrics of IoU(%, Pd(%) F, " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "(10^{-6})" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\mathrm{F_1}(\\%)" + }, + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": " are presented. The best values are highlighted with bold, the second best values are underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 96, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 56, + 96, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 555, + 304 + ], + "type": "table", + "html": "
MethodNUDT-SIRSTIRSTD-1k
IoUPdFaF1IoUPdFaF1
IPI [4]38.1590.78369.961.5230.9578.91180.455.15
PSTNN [31]17.2154.4781.6229.4525.5862.9678.9141.35
FKRW [15]25.3860.26107.538.5415.8954.5426.7527.08
GSWLCM [16]4.90965.0015.719.3013.06859.2526.946.045
RUMFR [1]26.0867.8996.4239.2510.2653.8725.2918.59
ACMNet [2]68.1496.5717.3883.4161.1487.8841.7974.17
ALCNet [3]72.5696.057.57286.7858.2789.9044.6972.92
AGPCNet [33]85.9797.637.22893.5162.5191.5820.9777.68
DNANet [9]94.7098.683.21297.3164.5190.2318.2778.77
UIUNet [21]89.1597.115.50694.2364.3290.2428.3379.02
RDIAN [18]86.0898.156.71192.3661.8688.5641.4476.27
DMFNet [6]87.6798.422.29493.3064.4590.1418.2478.52
SCTrans [29]94.8398.422.06597.1962.9592.2423.4078.04
Text-IRSTD (w/o text)95.2598.941.66397.3765.5091.5616.1378.73
Text-IRSTD (Ours)95.8499.731.03297.9569.5792.5914.9779.24
", + "image_path": "2a2e6c604f3190f9337dc6336cc9180b29cf734b2a0840cc3f683ef913262059.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 312, + 294, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 294, + 420 + ], + "type": "text", + "content": "improvements in all metrics. However, the limited visual information of the target makes it difficult to further improve their detection accuracy in complex scenes. For instance, these SOTA DL-based methods achieve a maximum IoU of only " + }, + { + "bbox": [ + 54, + 312, + 294, + 420 + ], + "type": "inline_equation", + "content": "64.51\\%" + }, + { + "bbox": [ + 54, + 312, + 294, + 420 + ], + "type": "text", + "content": " on the IRSTD-1k, which is difficult to satisfy higher detection requirements. By incorporating text information and designing cross-modal information interaction methods, the proposed Text-IRSTD achieves up to " + }, + { + "bbox": [ + 54, + 312, + 294, + 420 + ], + "type": "inline_equation", + "content": "69.57\\%" + }, + { + "bbox": [ + 54, + 312, + 294, + 420 + ], + "type": "text", + "content": " performance improvement in IoU." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 421, + 295, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 295, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 295, + 494 + ], + "type": "text", + "content": "It is worth noting that our method is able to produce satisfactory results even when using the default text input ('w/o text') in the testing stage. This phenomenon indicates that our method does not rely completely on detailed text prompts, and the inherent advantages of the model design contribute to the performance of IRSTD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 502, + 167, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 167, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 167, + 514 + ], + "type": "text", + "content": "5.4. Qualitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "type": "text", + "content": "We visualize the detection results of six representative algorithms, as shown in Fig. 5 and Fig. 6. It can be seen that traditional algorithms (e.g., PSTNN) have a large number of missed detections and false alarms due to their limited artificial feature representation capabilities. DL-based methods have made significant progress in background suppression by mining high-level semantic features. However, they all showed false alarms in the third row of images in Fig. 5. The reason is that the limited visual information of infrared images makes it difficult to further suppress interference and improve detection performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "In contrast, proposed Text-IRSTD accurately detects targets in every scene and suppresses almost all background interference. By introducing semantic text to expand the limited visual information, Text-IRSTD can further improve the detection accuracy in complex scenes. As shown" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 312, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 385 + ], + "type": "text", + "content": "in Fig. 5, thanks to the PCSID, we realize high-performance interactions across modal information, which can focus more on the region of interest to suppress complex interferences, while perceive and enhance the detail information to guide the accurate detection and contour recovery of targets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 396, + 493, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 493, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 493, + 407 + ], + "type": "text", + "content": "5.5. Performance in Unseen Scenarios" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 415, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 486 + ], + "type": "text", + "content": "In practical applications, models often need to face unseen detection scenarios. The detection performance in unseen scenarios can evaluate the practicality and generalization of models to some extent. To further validate the advancement of proposed methods, we compare the detection capabilities of different methods in unseen scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 487, + 554, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 572 + ], + "type": "text", + "content": "Specifically, we train different models using the NUDT-SIRST dataset and then test their detection performance using the unseen NUAA-SIRST dataset. Tab. 2 presents the detection results of different models on the unseen dataset. It can be seen that our method outperforms the state-of-the-art models in terms of IoU, " + }, + { + "bbox": [ + 313, + 487, + 554, + 572 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_d" + }, + { + "bbox": [ + 313, + 487, + 554, + 572 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 487, + 554, + 572 + ], + "type": "inline_equation", + "content": "\\mathrm{F}_a" + }, + { + "bbox": [ + 313, + 487, + 554, + 572 + ], + "type": "text", + "content": ", implying better detection accuracy and scene generalization." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 611, + 552, + 711 + ], + "blocks": [ + { + "bbox": [ + 313, + 586, + 553, + 608 + ], + "lines": [ + { + "bbox": [ + 313, + 586, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 553, + 608 + ], + "type": "text", + "content": "Table 2. Detection performance of different methods for the unseen dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 611, + 552, + 711 + ], + "lines": [ + { + "bbox": [ + 318, + 611, + 552, + 711 + ], + "spans": [ + { + "bbox": [ + 318, + 611, + 552, + 711 + ], + "type": "table", + "html": "
MethodNUDT-SIRST (Train)
NUAA-SIRST (Test)
IoU(%)Pd(%)Fa(10-6)
DNANet [9]35.8076.14141.3
UIUNet [21]53.3280.7335.15
RDIAN [18]39.3577.9843.68
DMFNet [6]43.7079.8160.07
SCTrans [29]50.9783.4837.71
Text-IRSTD (Ours)54.5187.1529.86
", + "image_path": "0420314caad6bea81d9dbe2a48203e725bb648f4be90fb143b2e0e94a0a2e94e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10640" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 401 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 401 + ], + "type": "image", + "image_path": "c7f75ae26b84ebd43b3c8a5af085f4cc2386cb5095160628f0e017f5d7e532e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 407, + 555, + 431 + ], + "lines": [ + { + "bbox": [ + 55, + 407, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 555, + 431 + ], + "type": "text", + "content": "Figure 5. Visual results of different IRSTD methods. The enlarged view of the target is shown in image corners. The red, blue, and yellow boxes represent correctly detected targets, missed targets, and false detections, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 439, + 149, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 149, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 149, + 452 + ], + "type": "text", + "content": "5.6. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 459, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 626 + ], + "type": "text", + "content": "Impact of Text Prompt: In order to verify the effectiveness of fuzzy text prompts, we designed three textual prompt variants for experiments and the results are shown in Tab. 3. It can be seen that the use of [Interested Region] target as a prompt to describe the IR small target helps to avoid the wrong text-image matching relationship caused by the use of explicit categories of prompts. Meanwhile, the description of [Scene] background further improves the detection performance in complex scenes by establishing more appropriate semantic associations. More importantly, experiments demonstrate the possibility of improving IRSTD performance with text and provide an effective technical path for further research." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Impact of PCSID: To validate the effectiveness of each component in PCSID, we perform several ablation studies on the IRSTD-1k, and the results are shown in Tab. 4. The results of the experiment indicate that both TGSI block and TGFA block improve the performance of the baseline model, and when use together, the model achieve the best results. This further confirms the importance of their syn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 439, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 462 + ], + "type": "text", + "content": "ergistic effect in enhancing the performance of small target detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 468, + 556, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 588 + ], + "type": "text", + "content": "Impact of TGFA: To demonstrate the effectiveness of TGFA block, we introduce three network variants for comparison, as shown in Tab. 5. It can be seen that proposed TGFA block achieves the best performance. The reason is that text features as modulation information can optimize the complementarity and interaction between cross-layer features. Moreover, DPM and CA can better perceive and enhance important visual features, such as contrast, which facilitate visual feature fusion and subsequent cross-modal information interaction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Impact of TGSI Block: We use four different text-image feature interaction methods to train the Text-IRSTD. As shown in Tab. 6, simply concatenating text and image features along the channel dimension ignores semantic inconsistencies across modalities, leading to inadequate semantic alignment. Without gain parameter " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", the model cannot dynamically focus on the regions of interest indicated by the text prompt. Similarly, without offset parameter " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", the model cannot make fine adjustments to the image features using the text features, thus limiting the effective" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10641" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 555, + 317 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 555, + 317 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 555, + 317 + ], + "type": "image", + "image_path": "3bf04e86d4cd1b77d4130bdad29c70bbfcf1e731f615c910984f5314bbf04bf1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 323, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 555, + 346 + ], + "type": "text", + "content": "Figure 6. 3D visualization results corresponding to those in Fig. 5. Different colors represent distinct grayscale values, gradually increasing from blue to yellow." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 370, + 555, + 436 + ], + "blocks": [ + { + "bbox": [ + 222, + 357, + 386, + 369 + ], + "lines": [ + { + "bbox": [ + 222, + 357, + 386, + 369 + ], + "spans": [ + { + "bbox": [ + 222, + 357, + 386, + 369 + ], + "type": "text", + "content": "Table 3. Ablation study of different prompts." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 370, + 555, + 436 + ], + "lines": [ + { + "bbox": [ + 58, + 370, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 58, + 370, + 555, + 436 + ], + "type": "table", + "html": "
TypePromptIoU(%)Pd(%)Fa(10-6)
Conventional promptsA photo of a [class]62.2290.5830.36
Variants of conventionA photo of a [class] in the [Scene] background66.6589.8921.69
Variants of OursA [Interested Region] target in the [Scene] background69.0191.6818.86
OursA photo of a [Interested Region] target in the [Scene] background69.5792.5914.97
", + "image_path": "1e92defea8b0ad4d3553e26ec0b3c7f37dce7e9fd4f552350f461d7b25eb536f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 456, + 295, + 520 + ], + "blocks": [ + { + "bbox": [ + 65, + 441, + 285, + 453 + ], + "lines": [ + { + "bbox": [ + 65, + 441, + 285, + 453 + ], + "spans": [ + { + "bbox": [ + 65, + 441, + 285, + 453 + ], + "type": "text", + "content": "Table 4. Ablation study of the TGFA block and TGSI block." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 456, + 295, + 520 + ], + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 520 + ], + "type": "table", + "html": "
BaselineTGFATGSIIoU(%)Pd(%)Fa(10-6)
✓61.6789.2257.97
✓✓67.7591.8719.83
✓✓66.1991.9226.39
✓✓✓69.5792.5914.97
", + "image_path": "4a57e05157528cdff16b095dd2342dab25b33c1ec6de555a16418520f824475e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 539, + 299, + 604 + ], + "blocks": [ + { + "bbox": [ + 96, + 525, + 254, + 536 + ], + "lines": [ + { + "bbox": [ + 96, + 525, + 254, + 536 + ], + "spans": [ + { + "bbox": [ + 96, + 525, + 254, + 536 + ], + "type": "text", + "content": "Table 5. Ablation study of the TGFA block." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 539, + 299, + 604 + ], + "lines": [ + { + "bbox": [ + 56, + 539, + 299, + 604 + ], + "spans": [ + { + "bbox": [ + 56, + 539, + 299, + 604 + ], + "type": "table", + "html": "
ModelIoU(%)Pd(%)Fa(10-6)
TGFA w/o text guide68.7592.0223.68
TGFA w/o CA68.0191.2426.24
TGFA w/o DPM67.1689.5617.95
TGFA (Ours)69.5792.5914.97
", + "image_path": "ff17e5dd11c04c6728edd2866bf21800e0c6dcb6cb3d8eac8f18597e076d950c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 56, + 622, + 296, + 687 + ], + "blocks": [ + { + "bbox": [ + 96, + 609, + 253, + 619 + ], + "lines": [ + { + "bbox": [ + 96, + 609, + 253, + 619 + ], + "spans": [ + { + "bbox": [ + 96, + 609, + 253, + 619 + ], + "type": "text", + "content": "Table 6. Ablation study of the TGSI block." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 622, + 296, + 687 + ], + "lines": [ + { + "bbox": [ + 56, + 622, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 56, + 622, + 296, + 687 + ], + "type": "table", + "html": "
ModelIoU(%)Pd(%)Fa(10-6)
Concatenation67.7390.5729.49
TGSI w/o αt66.6590.1335.85
TGSI w/o βt68.0591.2116.22
TGSI (Ours)69.5792.5914.97
", + "image_path": "6615fb772fd786c8da72dbdf6494bdc45c5f894a8380b58eb08a36455cf71ece.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "ness of target shape reconstruction. By combining both, the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 443, + 529, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 529, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 529, + 456 + ], + "type": "text", + "content": "proposed TGSI Block achieves the best performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 388, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 388, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 388, + 487 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we extend the IRSTD task and propose a text-guided IRSTD framework to address the issue that current models struggle to further improve detection accuracy by relying only on limited visual information. On the one hand, we devise a novel fuzzy semantic text prompt to accommodate ambiguous target categories. This prompt provides more text-image semantic associations that contribute to clutter suppression and target contour recovery. On the other hand, we propose PCSID to facilitate the information interaction between text and images. Moreover, we have constructed the first text-image bimodal dataset, called FZDT, to provide the basis for subsequent text-guided IRSTD research. Extensive experimental results demonstrate that our method achieves better detection performance and target contour recovery than the state-of-the-art methods. More surprisingly, the proposed Text-IRSTD demonstrates strong generalization and wide application prospects in unseen detection scenarios." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10642" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 158 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62405060, in part by the Nature Science Foundation of Fujian Province under Grant 2022J05113 and 2024J01245, and in part by the Educational Research Program for Young and Middle-aged Teachers of Fujian Province under Grant JAT210035." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 199, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 199, + 296, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 199, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 61, + 199, + 296, + 253 + ], + "type": "text", + "content": "[1] Liqiong Chen, Tong Wu, Shuyuan Zheng, Zhaobing Qiu, and Feng Huang. Robust unsupervised multifeature representation for infrared small target detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:10306-10323, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 255, + 296, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 255, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 61, + 255, + 296, + 309 + ], + "type": "text", + "content": "[2] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Asymmetric contextual modulation for infrared small target detection. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 950-959, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 311, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 311, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 311, + 296, + 354 + ], + "type": "text", + "content": "[3] Yimian Dai, Yiquan Wu, Fei Zhou, and Kobus Barnard. Attentional local contrast networks for infrared small target detection. IEEE transactions on geoscience and remote sensing, 59(11):9813-9824, 2021. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 356, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 356, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 356, + 296, + 409 + ], + "type": "text", + "content": "[4] Chenqiang Gao, Deyu Meng, Yi Yang, Yongtao Wang, Xiaofang Zhou, and Alexander G Hauptmann. Infrared patch-image model for small target detection in a single image. IEEE transactions on image processing, 22(12):4996-5009, 2013. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 411, + 296, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 296, + 455 + ], + "type": "text", + "content": "[5] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 457, + 296, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 296, + 500 + ], + "type": "text", + "content": "[6] Tan Guo, Baojiang Zhou, Fulin Luo, Lei Zhang, and Xinbo Gao. Dmfnet: Dual-encoder multi-stage feature fusion network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-14, 2024. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 296, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 545 + ], + "type": "text", + "content": "[7] Qingyu Hou, Zhipeng Wang, Fanjiao Tan, Ye Zhao, Hao-liang Zheng, and Wei Zhang. Ristdnet: Robust infrared small target detection network. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 547, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 296, + 590 + ], + "type": "text", + "content": "[8] Xuan Kong, Chunping Yang, Siying Cao, Chaohai Li, and Zhenming Peng. Infrared small target detection via nonconvex tensor fibered rank approximation. IEEE Transactions on Geoscience and Remote Sensing, 60:1-21, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 592, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 296, + 645 + ], + "type": "text", + "content": "[9] Boyang Li, Chao Xiao, Longguang Wang, Yingqian Wang, Zaiping Lin, Miao Li, Wei An, and Yulan Guo. Dense nested attention network for infrared small target detection. IEEE Transactions on Image Processing, 32:1745-1758, 2022. 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "type": "text", + "content": "[10] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10965-10975, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[11] Yongsong Li, Zhengzhou Li, Yu Shen, and Jie Li. Infrared small target detection based on 1-d difference of guided filtering. IEEE Geoscience and Remote Sensing Letters, 20: 1-5, 2023. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "text", + "content": "[12] Fanzhao Lin, Kexin Bao, Yong Li, Dan Zeng, and Shiming Ge. Learning contrast-enhanced shape-biased representations for infrared small target detection. IEEE Transactions on Image Processing, 33:3047-3058, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 162, + 554, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 554, + 215 + ], + "type": "text", + "content": "[13] Qiankun Liu, Rui Liu, Bolun Zheng, Hongkui Wang, and Ying Fu. Infrared small target detection with scale and location sensitivity. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17490-17499, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 554, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 270 + ], + "type": "text", + "content": "[14] Yuan Luo, Xiaorun Li, Shuhan Chen, Chaoqun Xia, and Liaoying Zhao. Imnn-lwec: A novel infrared small target detection based on spatial-temporal tensor model. IEEE Transactions on Geoscience and Remote Sensing, 60:1-22, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "text", + "content": "[15] Yao Qin, Lorenzo Bruzzone, Chengqiang Gao, and Biao Li. Infrared small target detection based on facet kernel and random walker. IEEE Transactions on Geoscience and Remote Sensing, 57(9):7104-7118, 2019. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 316, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 554, + 360 + ], + "type": "text", + "content": "[16] Zhaobing Qiu, Yong Ma, Fan Fan, Jun Huang, and Lang Wu. Global sparsity-weighted local contrast measure for infrared small target detection. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 361, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 554, + 425 + ], + "type": "text", + "content": "[17] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 426, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 554, + 480 + ], + "type": "text", + "content": "[18] Heng Sun, Junxiang Bai, Fan Yang, and Xiangzhi Bai. Receptive-field and direction induced attention network for infrared dim small target detection with a large-scale dataset irdst. IEEE Transactions on Geoscience and Remote Sensing, 61:1-13, 2023. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 482, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 554, + 536 + ], + "type": "text", + "content": "[19] Xiaozhong Tong, Zhen Zuo, Shaojing Su, Junyu Wei, Xiaoyong Sun, Peng Wu, and Zongqing Zhao. St-trans: Spatial-temporal transformer for infrared small target detection in sequential images. IEEE Transactions on Geoscience and Remote Sensing, 62:1-19, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 537, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 602 + ], + "type": "text", + "content": "[20] Keyan Wang, Xueyan Wu, Peicheng Zhou, Zuntian Chen, Rui Zhang, Liyun Yang, and Yunsong Li. Afe-net: Attention-guided feature enhancement network for infrared small target detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 17:4208-4221, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "type": "text", + "content": "[21] Xin Wu, Danfeng Hong, and Jocelyn Chanussot. Uiu-net: U-net in u-net for infrared small object detection. IEEE Transactions on Image Processing, 32:364-376, 2022. 2, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 636, + 555, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 555, + 691 + ], + "type": "text", + "content": "[22] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7031-7040, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[23] Yunhong Xin, Jiao Zhou, and Yishuan Chen. Dual multiscale filter with SSS and GW for infrared small target detec" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10643" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "tion. Infrared Physics & Technology, 81:97-108, 2017. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[24] Yunkai Xu, Minjie Wan, Xiaojie Zhang, Jian Wu, Yili Chen, Qian Chen, and Guohua Gu. Infrared small target detection based on local contrast-weighted multidirectional derivative. IEEE Transactions on Geoscience and Remote Sensing, 61: 1-16, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 204 + ], + "type": "text", + "content": "[25]Fuju Yan,Guili Xu,Junpu Wang,Quan Wu,and Zhengsheng Wang. Infrared small target detection via schatten capped pnorm-based non-convex tensor low-rank approximation.IEEE Geoscience and Remote Sensing Letters,20: 1-5,2023.1,2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "type": "text", + "content": "[26] Bo Yang, Xinyu Zhang, Jian Zhang, Jun Luo, Mingliang Zhou, and Yangjun Pi. Eflnet: Enhancing feature learning network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-11, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 316 + ], + "type": "text", + "content": "[27] Xinyi Ying, Li Liu, Yingqian Wang, Ruojing Li, Nuo Chen, Zaiping Lin, Weidong Sheng, and Shilin Zhou. Mapping degeneration meets label evolution: Learning infrared small target detection with single point supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15528-15538, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 317, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 360 + ], + "type": "text", + "content": "[28] Chuang Yu, Yunpeng Liu, Shuhang Wu, Xin Xia, Zhuhua Hu, Deyan Lan, and Xin Liu. Pay attention to local contrast learning networks for infrared small target detection. IEEE Geoscience and Remote Sensing Letters, 19:1-5, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 404 + ], + "type": "text", + "content": "[29] Shuai Yuan, Hanlin Qin, Xiang Yan, Naveed Akhtar, and Ajmal Mian. Sctransnet: Spatial-channel cross transformer network for infrared small target detection. IEEE Transactions on Geoscience and Remote Sensing, 62:1-15, 2024. 1, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 405, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 294, + 459 + ], + "type": "text", + "content": "[30] Hang Zhang, Chongruo Wu, Zhongyue Zhang, Yi Zhu, Haibin Lin, Zhi Zhang, Yue Sun, Tong He, Jonas Mueller, R Manmatha, et al. Resnest: Split-attention networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2736–2746, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "text", + "content": "[31] Landan Zhang and Zhenming Peng. Infrared small target detection based on partial sum of the tensor nuclear norm. Remote Sensing, 11(4):382, 2019. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 494, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 494, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 494, + 294, + 547 + ], + "type": "text", + "content": "[32] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 877-886, 2022. 1, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[33] Tianfang Zhang, Lei Li, Siying Cao, Tian Pu, and Zhenming Peng. Attention-guided pyramid context networks for detecting infrared small target under complex background. IEEE Transactions on Aerospace and Electronic Systems, 59(4): 4250-4261, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 604, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 647 + ], + "type": "text", + "content": "[34] Xiangyue Zhang, Jingyu Ru, and Chengdong Wu. Infrared small target detection based on gradient correlation filtering and contrast measurement. IEEE Transactions on Geoscience and Remote Sensing, 61:1-12, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[35] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16793-16803, 2022. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10644" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Text-guided Visual Prompt DINO for Generic Segmentation/0364b0ce-3511-4546-ac4b-3ddb77cbd170_content_list.json b/2025/Text-guided Visual Prompt DINO for Generic Segmentation/0364b0ce-3511-4546-ac4b-3ddb77cbd170_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d1d3d759bbaed7274fced2452801df9205fa261c --- /dev/null +++ b/2025/Text-guided Visual Prompt DINO for Generic Segmentation/0364b0ce-3511-4546-ac4b-3ddb77cbd170_content_list.json @@ -0,0 +1,1873 @@ +[ + { + "type": "text", + "text": "Text-guided Visual Prompt DINO for Generic Segmentation", + "text_level": 1, + "bbox": [ + 194, + 130, + 802, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuchen Guan $^{1*‡}$ , Chong Sun $^{2*†}$ , Canmiao Fu $^{2*}$ , Zhipeng Huang $^{2}$ , Chun Yuan $^{1†}$ , Chen Li $^{2}$ $^{1}$ Tsinghua Shenzhen International Graduate School, Tsinghua University \n $^{2}$ WeChat AI, Tencent Inc.", + "bbox": [ + 148, + 179, + 846, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zero-Shot Segmentation via Visual Prompt", + "bbox": [ + 102, + 247, + 269, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e45deb8b198c74b7630209764108ebb61d60c61a538f52efb612fb59470c6d18.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 255, + 300, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eaebf852eba30e883eff7ec2a75fc9b18025760ac819f6d6d74b26e735242734.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 255, + 491, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Cross-Image Segmentation", + "bbox": [ + 192, + 358, + 401, + 371 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/344751fe6bbf4ec16e7cdd984b13d713b4dd4fd5c0d63fdbf6bf0d19219ed9ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 255, + 699, + 353 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3b56b0d126228125b9af241c58353b01639eb23a8ff83f6a18c340dc88d47ec6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 255, + 895, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual In-Image Segmentation", + "bbox": [ + 611, + 358, + 797, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zero-Shot Segmentation via Text Prompt", + "bbox": [ + 102, + 373, + 261, + 382 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4863ca13a43dacc00841f33624d25d6f860cef1b950829c7a95604e336ad325d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 382, + 267, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf18a0dee20706dfe0821158eae9379cb1666290a01c69d002e96a6611306641.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 382, + 433, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text In-Image Segmentation", + "bbox": [ + 179, + 484, + 352, + 498 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/53ef319c81eab998c5d0685eeb3e283914d3ad8a93348245339ee6840ad3b30d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 382, + 669, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e19643d3f04e3b5794d8bf8a5ef2f7e301c14245e7ed5b44a49a1645b98f870b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 671, + 382, + 895, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text Panoptic Segmentation", + "bbox": [ + 580, + 484, + 751, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zero-Shot part Segmentation via Text/Visual Prompt", + "bbox": [ + 102, + 500, + 305, + 510 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9df8bf5121a3e1cc6b08953b1cef5d3a0ee690e40cba00a72f329948c7dfb30d.jpg", + "image_caption": [ + "Figure 1. Detection/Segmentation samples. Our Prompt-DINO is a text-guided visual Prompt DINO. Prompt-DINO supports single or multiple visual or text prompts for in-image or cross-image generic segmentation. Based on the granularity of the provided prompts, Prompt-DINO is capable of performing both panoptic and part segmentation." + ], + "image_footnote": [], + "bbox": [ + 101, + 510, + 287, + 608 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80779919295d9afce6d868cafd836ea76fadca1c0670bf94789f9fe64e0aa97b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 510, + 473, + 609 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/83f5ab435ed5f02ce9ec3bd1697698c2c228a088f649f3a452b3fb39c2899b1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 473, + 510, + 676, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual Part Segmentation", + "bbox": [ + 310, + 613, + 467, + 626 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6703fb5ff0fc67b914d10a2def8853e5c821b28d48ca721d3605ac56d0293707.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 510, + 771, + 608 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1920781344e8b6fb191b66b01cf79d33013a837af902051c8f1306b2608e2bb3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 510, + 893, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text Part Segmentation", + "bbox": [ + 718, + 614, + 864, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 689, + 326, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in multimodal vision models have highlighted limitations in late-stage feature fusion and suboptimal query selection for hybrid prompts open-world segmentation, alongside constraints from caption-derived vocabularies. To address these challenges, we propose Prompt-DINO, a text-guided visual Prompt DINO framework featuring three key innovations. First, we introduce an early fusion mechanism that unifies text/visual prompts", + "bbox": [ + 89, + 720, + 485, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and backbone features at the initial encoding stage, enabling deeper cross-modal interactions to resolve semantic ambiguities. Second, we design order-aligned query selection for DETR-based architectures, explicitly optimizing the structural alignment between text and visual queries during decoding to enhance semantic-spatial consistency. Third, we develop a generative data engine powered by the Recognize Anything via Prompt (RAP) model, which synthesizes 0.5B diverse training instances through a dualpath cross-verification pipeline, reducing label noise by $80.5\\%$ compared to conventional approaches. Extensive experiments demonstrate that Prompt-DINO achieves state-of-the-art performance on open-world detection bench", + "bbox": [ + 511, + 691, + 906, + 888 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution.", + "bbox": [ + 112, + 849, + 233, + 862 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Authors. Emails: waynecsun@tencent.com, yuanc@sz.tsinghua.edu.cn", + "bbox": [ + 116, + 862, + 437, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Work done as interns at WeChat.", + "bbox": [ + 116, + 887, + 299, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21288", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "marks while significantly expanding semantic coverage beyond fixed-vocabulary constraints. Our work establishes a new paradigm for scalable multimodal detection and data generation in open-world scenarios. Data&Code are available at https://github.com/WeChatCV/WeVisionOne.", + "bbox": [ + 89, + 90, + 485, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 195, + 222, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in vision models [5, 43-46, 58] have focused on unifying diverse prompting strategies—such as text, visual, or hybrid inputs—to enable versatile object understanding and segmentation. Pioneering works like [16, 21, 22, 60] have demonstrated the potential of multimodal prompting for tasks ranging from open-set detection to segmentation. Despite their emerging capabilities, openworld detection and segmentation methods leveraging multimodal prompts demonstrate inherent technical deficiencies.1) Late-stage fusion of multimodal features. [22, 60] process text, visual or other prompts, and backbone features in separate subspaces, merging them only in later encoding stages. [16] similarly adopts contrastive learning to align modalities without joint optimization at the encoding stage. This delayed interaction restricts cross-modal alignment, particularly in ambiguous scenarios where early fusion of text and visual cues could disambiguate object semantics. 2) Insufficient exploration of query selection mechanisms. Most of these methods employ generic crossattention for text-guided query selection, or handle queries corresponding to text and visual prompts independently, none of which explicitly optimizes the structural alignment between queries corresponding to text and visual cues respectively. Like [16] processes text and visual prompts independently. This results in the selection of sub-optimal queries that do not promote optimization between text and images.", + "bbox": [ + 91, + 220, + 483, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond these architectural limitations, a broader challenge persists across the field: the reliance on caption-derived vocabularies or fixed-vocabulary, which inherently constrain generalization to novel or composite objects. [16, 22, 60] derive their object vocabularies from entity words extracted from image captions or use fixed-vocabulary detector (like TAP [32]). While this approach avoids manually predefined categories, it remains bounded by the entities explicitly mentioned in the training captions. For instance, TRex-2 [16] employ TAP [32] to annotate SA-1B [19] with a category name from a dictionary of 2560 classes. Nevertheless, the reality that SA-1B [19] is associated with a vocabulary significantly exceeding 2560 substantially constrains the dataset's potential.", + "bbox": [ + 89, + 628, + 483, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose Prompt-DINO, a novel text-guided visual Prompt DINO, that addresses these challenges through two architectural innovations and a paradigm-shifting data engine. We pioneer an early fusion", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mechanism that deeply integrates text, visual, and backbone features at the initial encoding stage. Unlike [16, 22, 60], our model unifies multimodal signals into a cohesive representation from the outset. This enables richer cross-modal interactions, reducing ambiguity in open-set scenarios. And we introduce order-aligned query selection, a novel paradigm for DETR-based architectures that explicitly optimizes the alignment between text queries and visual queries during decoding. By enhancing semantic correlation and spatial consistency within query priorities, our method facilitates the optimization of text and image query selection in a unified direction, thereby surpassing the heuristic cross-attention mechanisms employed in prior research. Also, we overcome the fixed vocabulary bottleneck through a generative data engine with dual-path cross-verification pipeline. We proposal an novel Recognize Anything via Prompt (RAP) model, a 0.5B-parameter multimodal large language model(MLLM) that generates semantic labels for input images and their corresponding masks. Our engine leverages RAP to synthesize 0.5B diverse training instances with enhanced semantic relevance. To ensure label fidelity, we further implement a novel dual-path cross-verification pipeline that substantially reduces label noise by $80.5\\%$ compared to conventional single-path approaches. This integrated data engine pioneers scalable open-world dataset generation, bridging limited annotation vocabularies with real-world detection needs.", + "bbox": [ + 511, + 90, + 906, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are threefold:", + "bbox": [ + 532, + 500, + 841, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose an early fusion mechanism that integrates text, visual, and backbone features during initial encoding stage, unifying multimodal signals into a cohesive representation.", + "- We introduce order-aligned query selection, a new paradigm for DETR-based architectures that explicitly optimizes the alignment between text queries and visual queries during decoding.", + "- We develop a generative data engine powered by our RAP, incorporating a dual-path cross-verification pipeline. This approach synthesizes diverse and semantically relevant training data, and establishes a new paradigm for scalable open-world data generation." + ], + "bbox": [ + 513, + 515, + 903, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 720, + 653, + 736 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. General Recognition via Prompts", + "text_level": 1, + "bbox": [ + 513, + 741, + 805, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in general recognition models (e.g., detection/segmentation [10, 11, 18, 24, 27, 30, 49, 53, 56]) have demonstrated that multimodal alignment between vision and language modalities serves as the cornerstone for open-world understanding. These frameworks typically employ joint text-visual prompting mechanisms to achieve object localization through aligned cross-modal feature spaces.", + "bbox": [ + 511, + 763, + 906, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text Prompt Paradigm. Existing methods predominantly focus on text-based cross-modal alignment. GLIP", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21289", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/919e3f7873156ad81dc5661fa82214035904da1c6d36d4c1bf6d2b1d3bc2f8d2.jpg", + "image_caption": [ + "Figure 2. The Overall Framework of Prompt-DINO. Prompt-DINO is constructed based on the DINO architecture. Images are processed by ViT to extract multi-scale features, while visual and text prompts are processed by deformable attention transformer layers and CLIP's text encoder, respectively, to derive embeddings. Through early fusion layers in the backbone encoder and perception-aware query selection, Prompt-DINO perform deep fusion and alignment of visual and text in two steps, with the final predictions generated by the Decoder." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 911, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[24] pioneers the adaptation of image-text pretraining to downstream detection tasks through deep feature fusion, establishing language-aware visual representations. GroundingDINO [27] further extends this paradigm by demonstrating that multi-stage cross-attention fusion significantly enhances performance in open-vocabulary detection. While these approaches achieve remarkable progress, their heavy reliance on textual semantics limits their capacity to handle visual concepts with ambiguous linguistic descriptions.", + "bbox": [ + 88, + 489, + 482, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Prompt Paradigm. Complementary to text prompts, visual prompting (points/boxes/masks) provides geometric guidance for specific object localization. SAM [19] establishes a milestone in interactive segmentation through decoupled visual prompt encoding, though lacking semantic awareness. Subsequent works like SEEM [60] and DINOv [22] enhance this paradigm by integrating semantic context and exemplar-based reasoning. A notable hybrid approach is T-Rex2 [16], which combines both modalities but still treats visual prompts as auxiliary geometric cues. Our key insight lies in that existing methods inadequately exploit the semantic synergy between visual prompts and language guidance, leaving visual prompting's full potential underdeveloped for language-aware feature learning.", + "bbox": [ + 89, + 626, + 482, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Data Engine for Open-World Learning", + "text_level": 1, + "bbox": [ + 89, + 848, + 426, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scalable data curation has become pivotal for training robust open-world detector/segmentation models. SAM2 [36]", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "introduces a cyclic model-annotator collaboration framework, iteratively refining the SA-V dataset through human-machine interactions. Alternative strategies employ cross-model verification: YOLO-World [6] curates CC3M-Lite by relabeling CC3M [39] with GLIP and CLIP, while T-Rex2 [16] propagates pseudo-labels across SA-1B and LAION400M [34]. Current methods face three key limitations: (1) Over-reliance on noun phrase extraction from captions misses more detailed concepts and introduces too many redundant nouns; (2) Single-path verification propagates annotation errors; (3) Semantic ambiguity in freeform text introduces label noise. Our data engine addresses these issues and enables the automatic generation of high-quality training data with reduced noise.", + "bbox": [ + 511, + 489, + 906, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 707, + 604, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by the limitations observed in existing hybrid prompting methods for detection and segmentation tasks, we introduce Prompt-DINO, an end-to-end generic segmentation model built upon the DINO architecture [52]. Our model natively integrates multimodal prompting capabilities for both textual and visual inputs while maintaining architectural simplicity. In this section, we first formalize the promptable segmentation paradigm and outline the core design prompt encoders of Prompt-DINO (§3.1). We then systematically address three critical challenges: (1) early fusion with visual prompts through cross-modal feature align-", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21290", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ment(§3.2), (2) order-aware query selection for maintaining semantic consistency(§3.3) and (3) construction of a scalable data engine for open-world generalization(§3.4). Finally, we introduce the training objective(§3.5).", + "bbox": [ + 89, + 90, + 483, + 154 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Promptable Segmentation Task", + "text_level": 1, + "bbox": [ + 89, + 160, + 369, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Prompt-DINO processes an input image $\\mathcal{I}$ , with accompanying visual prompts $\\mathcal{P}^v = \\{\\mathcal{P}_1^v, \\dots, \\mathcal{P}_K^v\\}$ , and textual prompts $\\mathcal{P}^t = \\{\\mathcal{P}_1^t, \\dots, \\mathcal{P}_K^t\\}$ , where $K$ is the number of interactive prompts per image. Following established architectures in DETR [57] and DINO [52], our system first extracts hierarchical visual features $F = \\{f_1, \\dots, f_L\\}$ through a vision backbone network (e.g., ViT [9] or Swin Transformer [28]), where $L$ represents the number of feature pyramid levels.", + "bbox": [ + 89, + 181, + 483, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Prompt Encoder. Specifically, our visual encoding module employs $L$ transformer blocks with deformable attention mechanisms [22]. For each target object, we initialize a learnable query embedding $q_{i}$ and progressively refine it through multi-scale feature aggregation. The transformation at each layer $l$ can be formulated as:", + "bbox": [ + 89, + 318, + 483, + 409 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} ^ {l} = M S D A \\left(v _ {i} ^ {l - 1}, f _ {l}\\right), l \\in [ 1, \\dots , L ], \\quad \\text {w i t h} \\quad v _ {i} ^ {0} = q _ {i} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 417, + 482, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $MSDA(\\cdot)$ denotes multi-scale deformable attention operation. The final visual prompt representations $V = \\{v_{1},\\dots,v_{K}\\}$ are obtained as $v_{i} = v_{i}^{L}$ .", + "bbox": [ + 89, + 446, + 483, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Text Prompt Encoder. Building upon insights from recent open-world detection studies, we employ a CLIP [34] text encoder rather than conventional BERT [8] approaches used in earlier models like [24, 27]. This design choice enables better romantic correlation modeling for input words, phrases, and short sentences, generating text prompts $T = \\{t_1,\\dots,t_K\\}$ that effectively guide cross-modal feature alignment.", + "bbox": [ + 89, + 492, + 483, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The integrated representation (text(T), visual(V), multiscale features(F)) feeds into the backbone encoder for multimodal fusion, enabling dynamic adaptation to prompts while preserving cross-modal alignment.", + "bbox": [ + 89, + 613, + 483, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Early Fusion with Visual Prompts", + "text_level": 1, + "bbox": [ + 89, + 681, + 387, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We are the first to propose the early fusion framework that integrates text prompts, visual prompts, and input image features through multi-modal interaction. While this fusion mechanism effectively reduces the modality gap between textual and visual features, enhances prompt-image correlation, and improves target recall, We found unintended consequences of severe hallucinations in our experiments. This phenomenon stems from the inherent limitation of conventional cross-attention mechanisms: they compulsively reconstruct Key features most similar to the Query features, even when all Key features are substantially dissimilar to the current Query. To address this non-target feature reconstruction issue, we introduce a novel gated cross", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ae748a4d13be1c914e0c223da049e14d11c49083d6f7c9c661587aa6ed6cb778.jpg", + "image_caption": [ + "Figure 3. Structure of Early Fusion Layer. Notably, the contrastive alignment of visual and text prompts is completed before early fusion. By employing Gated Cross-Attention, text prompts are reconstructed with image features, while visual prompts and image features mutually reconstruct each other." + ], + "image_footnote": [], + "bbox": [ + 516, + 85, + 908, + 334 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "attention mechanism incorporating a learnable background token $(B)$ . Reconstruction occurs exclusively when the Query-Key similarity exceeds an adaptive threshold; otherwise, the background token automatically activates to suppress hallucination. As illustrated in Fig. 3, our architecture comprises three core components: 1) a self-attention layer for intra-modal refinement, 2) the proposed gated cross-attention layer for inter-modal interaction, and 3) a feedforward neural network (FFN) for feature transformation. The cross-attention module implements three interaction pathways: feature-to-visual prompt, visual prompt-to-feature, and feature-to-text prompt. For visualization clarity, Fig. 3 presents only stage-2 features mapped to text space through our two-phase training scheme (detailed in §3.5). The enhanced cross-modal fusion operates through the following formulation:", + "bbox": [ + 511, + 409, + 908, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i} ^ {\\prime} = \\operatorname {G a t e d A t t n} (f _ {i}, t _ {i}, t _ {i})\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 656, + 803, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} ^ {\\prime} = \\operatorname {G a t e d A t t n} \\left(f _ {i}, v _ {i}, v _ {i}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 676, + 903, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} ^ {\\prime} = \\operatorname {G a t e d A t t n} (v _ {i}, f _ {i}, f _ {i})\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 695, + 800, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the gated attention mechanism is defined as:", + "bbox": [ + 513, + 720, + 852, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {G a t e d A t t n} (Q, K, V) = \\operatorname {s o f t m a x} (\\frac {Q \\cdot [ K , B ]}{\\sqrt {d _ {k}}}) [ V, B ] \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 739, + 906, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $B$ denotes our learnable background token. Following TRex-2 [16] methodology, we employ region-level contrastive learning to align visual and textual modalities, utilizing symmetric cross-entropy loss [34] for embedding alignment:", + "bbox": [ + 511, + 770, + 906, + 844 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = - \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\log \\frac {v _ {i} \\cdot t _ {i}}{\\sum_ {j = 1} ^ {K} v _ {i} \\cdot t _ {j}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 833, + 903, + 873 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our experience, symmetric cross-entropy losses have better convergence than InfoNCE loss [31].", + "bbox": [ + 511, + 871, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21291", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the context of training detection and segmentation tasks, it is common practice to construct negative samples. For negative prompt construction, while text prompts naturally leverage category vocabularies for negative sampling, visual prompts require explicit negative construction. We address this by aggregating visual prompt features across all categories within each batch and computing inter-category averages to generate negative visual prompts. This approach ensures sufficient diversity in negative samples while maintaining semantic relevance.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate annotation granularity discrepancies across datasets (e.g., \"bird\" in Dataset A vs. \"pigeon\" in Dataset B), we implement a dataset-aware sampling strategy. Our novel data sampler ensures each training batch contains only intra-dataset samples, preventing semantically similar but categorically distinct concepts from becoming conflicting negative samples. This is particularly critical in visual-prompt open-world detection/segmentation scenarios where label hierarchies may differ substantially across source datasets.", + "bbox": [ + 89, + 242, + 480, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Order-aware query selection", + "text_level": 1, + "bbox": [ + 89, + 401, + 346, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prompt-DINO adopts a DETR-based architecture that selects top-K encoder features through similarity computation between encoder features and prompt embeddings, following DINO [52] and MaskDINO [21]. However, our approach fundamentally diverges from existing methods by simultaneously considering both textual-visual query similarity and inter-query correlations with visual prompts. To further strengthen the alignment between visual and textual prompts, we introduce order-aware alignment during query selection, which establishes the second-level alignment between visual and textual prompts during training. Specifically, we employ Kendall's $\\tau$ coefficient to quantify the ordinal consistency [12] between queries selected by visual prompts and text prompts, the Kendall's $\\tau$ coefficient can be expressed as: $P_{\\mathrm{c}} - P_{\\mathrm{d}}$", + "bbox": [ + 89, + 422, + 482, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau = \\frac {P _ {c} - P _ {d}}{\\frac {1}{2} N (N - 1)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 637, + 480, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P_{c}$ represents concordant pairs and $P_{d}$ represents discordant pairs, $N$ represents the total number of queries. Given the similarity score between query embeddings and prompt embeddings $S^{v} = s_{1}^{v},\\ldots ,s_{N}^{v}$ and $S^t = s_1^t,\\dots,s_N^t$ we calculate $P_{c}$ and $P_{d}$ as follows:", + "bbox": [ + 89, + 670, + 482, + 743 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tau = \\frac {\\sum_ {i} \\sum_ {j < i} \\operatorname {s g n} \\left(s _ {i} ^ {t} - s _ {j} ^ {t}\\right) \\cdot \\operatorname {s g n} \\left(s _ {i} ^ {v} - s _ {j} ^ {v}\\right)}{\\frac {1}{2} N (N - 1)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 743, + 480, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the sgn is not differentiable, we approximate the sign function with the tanh function and utilize the opposite of $\\tau$ as the order alignment loss:", + "bbox": [ + 89, + 777, + 482, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {o r d e r}} = - \\frac {\\sum_ {i} \\sum_ {j < i} \\tanh \\left(s _ {i} ^ {t} - s _ {j} ^ {t}\\right) \\cdot \\tanh \\left(s _ {i} ^ {v} - s _ {j} ^ {v}\\right)}{\\frac {1}{2} N (N - 1)} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 821, + 480, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed alignment mechanism operates at two complementary levels: The primary alignment directly optimizes the learning objective through $\\mathcal{L}_{\\text{align}}$ (as given in", + "bbox": [ + 89, + 856, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eq. 4), while our novel order-aware alignment provides indirect assurance through two critical aspects: (1) Semantic proximity between corresponding text and visual prompts for candidate queries naturally reduces overall training complexity by minimizing feature space discrepancies; (2) Training-inference consistency preservation ensures equivalent substitutability between text and visual prompts during inference through maintained representational parity.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 5 demonstrates the effectiveness of our proposed method. Prior to applying the order-aware loss optimization, we observe significant disorder in both textual and visual prompt queries, with no discernible alignment between corresponding semantic representations. The optimization process induces a remarkable transformation - post-optimization analysis reveals substantially enhanced consistency between semantically equivalent textual and visual prompt queries.", + "bbox": [ + 511, + 213, + 903, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Data Engine", + "text_level": 1, + "bbox": [ + 513, + 361, + 645, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effective visual concept learning from rich image-text pairs is fundamental to enhancing the generalization capabilities of open-vocabulary object detectors. Existing methods predominantly rely on captioning models([3, 23, 26] etc.), coupled with lexical analysis tools like SpaCy [14] to extract noun phrases. These noun phrases are subsequently employed to generate pseudo-labels via open-set detection models ([24, 27]). Nonetheless, these techniques have significant drawbacks. They frequently omit valid labels, introduce erroneous annotations, and are susceptible to irrelevant lexical noise. All of these issues detrimentally affect detection performance.", + "bbox": [ + 511, + 383, + 905, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address these challenges, we propose a novel Recognize Anything via Prompting (RAP) model, a 0.5B MLLM, which significantly enhances annotation accuracy. Our framework further integrates an innovative dual-path cross-validation pipeline that substantially reduces label noise. This comprehensive approach enables the fully automated generation of high-quality open-world detection datasets while preserving the fidelity of annotations.", + "bbox": [ + 511, + 566, + 905, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recognize Anything via Prompting. Despite the remarkable capabilities of the Segment Anything Model (SAM) [19] in detecting a wide array of objects within images, and the substantial progress achieved in region-level understanding through multimodal models [1, 2, 33, 35, 51], a critical question persists: Can we effectively assign semantic labels to arbitrary mask prompts? To address this challenge, we introduce RAP, a novel framework that bridges visual segmentation with semantic recognition. As illustrated in Fig. 4, our RAP model processes raw image inputs alongside their corresponding object masks to generate precise semantic annotations for the specified regions. The proposed RAP comprises three key components:", + "bbox": [ + 511, + 688, + 905, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1. Feature extraction. We perform masked pooling over", + "bbox": [ + 534, + 885, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21292", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4b51941a84aacf5fc7d55a29f37b0b50e268f690b5918da5e650c453af95325f.jpg", + "image_caption": [ + "Figure 4. Data Engine and RAP Frameworks. Left: The data engine with dual-path cross-verification pipeline. Right: The proposed Recognize Anything via Prompting (RAP), a 0.5B-parameter MLLM model." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 632, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/525074c2087d8413737cf9a47c9ee6667778f38f68f406135b0da0974b7c1eea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 655, + 88, + 905, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "multi-layer image features to extract region-specific features corresponding to input masks, and concatenate the mask token embeddings obtained via a projection layer with the original image tokens to form a unified representation. This preserves both global contextual information from the complete image and localized object-specific features within masked regions, effectively integrating scene-level understanding with instance-level visual cues.", + "bbox": [ + 88, + 270, + 482, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2. Ground Data Cleaning. To standardize the format of our ground truth data, we leveraged the Qwen model [48] to eliminate redundant descriptive phrases through comprehensive text normalization. Specifically, we transformed verbose expressions into their canonical semantic equivalents (e.g., converting \"any part of the sky\" to \"sky\"). This preprocessing strategy serves dual purposes: (1) it reduces the RAP's learning complexity by minimizing lexical variability, and (2) it mitigates the risk of hallucinations by ensuring precise semantic alignment between textual annotations and visual content.", + "bbox": [ + 88, + 393, + 482, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3. Training. The proposed methodology employs a two-stage training strategy. During the first stage, we conduct pre-training using 10M samples, specifically optimizing only the projector while keeping other model components frozen. The second stage implements instruction-tuning, where we unfreeze and jointly optimize the entire model to enhance task-specific adaptation capabilities.", + "bbox": [ + 88, + 561, + 482, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dual-path cross-verification pipeline. We propose a dual-pipeline annotation framework combining bottom-up and top-down approaches with cross-verification filtering, consisting of three key components:", + "bbox": [ + 88, + 669, + 482, + 731 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Top-Down Pipeline. Our top-down pipeline leverages the Qwen-VL model [42] to extract semantic tags from images. These tags are subsequently fed into an open-set detection model to generate corresponding bounding boxes. This pipeline demonstrates strong capability in detecting large-scale salient objects (e.g., trees, roads, sky).", + "2. Bottom-Up Pipeline. To address the inherent limitation of top-down pipeline in capturing fine-grained details, we implement a complementary bottom-up strategy using the RAP integrated with SAM [19]. This combination enables effective detection of smaller objects and detailed" + ], + "bbox": [ + 89, + 732, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "components (e.g., eyes, noses, windows) that are frequently missed by conventional top-down pipeline.", + "bbox": [ + 511, + 270, + 903, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3. Cross-Verification. The detection outputs from both pipelines undergo rigorous validation through a three-stage process. First, We employ the Hungarian matching algorithm to establish correspondences between bounding boxes from different pipelines. Next, we utilize the BGE model [47] to compute similarity scores between the tags associated with the matched bounding boxes. Finally, we filter out the candidate pairs whose similarity scores were below the threshold of empirical determination to further clean up the annotated results.", + "bbox": [ + 511, + 300, + 905, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Training Objective", + "text_level": 1, + "bbox": [ + 511, + 460, + 694, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building upon previous DINO-like models [52], we employ a combination of L1 loss and GIoU loss [37] for bounding box regression, while for mask segmentation, we adopt cross-entropy loss and dice loss [29]. Following the methodology established in GroudingDINO [27], we utilize contrastive loss between predicted objects and text prompts for classification tasks. To accelerate model convergence, we incorporate denoising training as implemented in DINO [52]. The complete objective function for Prompt DINO can be expressed as:", + "bbox": [ + 511, + 482, + 905, + 633 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {b b o x} + \\mathcal {L} _ {m a s k} + \\mathcal {L} _ {D N} + \\mathcal {L} _ {a l i g n} + \\mathcal {L} _ {o r d e r} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 646, + 903, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our training methodology adopts a progressive multi-stage strategy: we first optimize the text prompt while keeping the visual components fixed, followed by joint optimization of both textual and visual prompt representations.", + "bbox": [ + 511, + 674, + 906, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 748, + 645, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 511, + 772, + 611, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Prompt-DINO, we employ ViT-L [9] as the backbone and fine-tune CLIP-L [34] as our text prompt encoder. The visual prompt encoder consists of 6 layers of deformable cross-attention transformer layers, with a hidden dimension of 256. We adopt the \"two-stage\" structure proposed by DINO [52], selecting 900 queries for decoding after the backbone encoder.", + "bbox": [ + 511, + 794, + 905, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21293", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/447feed9f38ad163ba0efef88a0e8682476bebbb75ac07b62a548321e5729834.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackbonePrompt TypeCOCO (in-domain)LVIS (in-domain)ADE20K (out-domain)
mask APbox APmask APbox APPQmask APbox AP
Mask2Former [4]Swin-LClose-Set48.652.1-----
OneFormer [15]Swin-LClose-Set48.9------
MaskDINO* [21]Swin-LClose-Set50.656.2-----
kMaX-DeepLab [50]Swin-LClose-Set58.1------
GLIPv2 [53]Swin-HText Prompt48.9------
X-Decoder [59]DaViT-LText Prompt46.7---21.813.138.1
OpenSeeD [54]Swin-LText Prompt53.258.221.023.019.715.017.7
T-Rex2 [16]Swin-LText Prompt-52.2*-45.8*---
APE [40]ViT-LText Prompt49.358.353.059.627.224.429.6
DINOv [22]Swin-LVisual Prompt50.454.2--23.215.114.3
T-Rex2 [16]Swin-LVisual Prompt-46.5*-45.3*---
Prompt-DINOViT-LText Prompt51.158.656.560.129.624.628.6
Prompt-DINOViT-LVisual Prompt51.058.055.659.935.928.032.7
", + "bbox": [ + 94, + 68, + 903, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Generic Segmentation results of Prompt-DINO on multiple datasets. \"-\" denotes the model does not have performance reported or does not have ability for the specific task. \"+\" denotes the performance reported under different domain settings. MaskDINO uses the results reported in DINOv [22]", + "bbox": [ + 89, + 294, + 906, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets. During the training process, we utilized three categories of datasets: open-source detection datasets. These include COCO [25] with 118K images and LVIS [13] with 160K images, both of which have mask annotations. For datasets without mask annotations, such as Object365 [38], OID [20], and V3Det [41], we first apply SAM to generate mask annotations. Due to the presence of low-quality data and long-tail distribution in Object365 and OID, which hinder general object recognition, we resampled them, selecting 700K and 1M data points, respectively. grounding datasets. This includes the GoldG [17] dataset with 100K images. self-built datasets. We re-announce and filter CC3M [39] and SA-1B [19] using the RAP and dual-path cross-validation data engine introduced in §3.4, yielding high-quality datasets of 3M and 7M images, respectively, which constitute our primary training data. For testing, we utilize COCO and LVIS for in-domain evaluation and ADE20K [55], SeginW [59], and Cityscapes [7] for out-domain evaluation.", + "bbox": [ + 88, + 342, + 480, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. We employ Average Precision (AP) as the evaluation metric, utilizing mask AP for instance segmentation, and box AP for object detection. For ADE20K and CtiyScapes, we employ Panoptic Quality(PQ) and mean intersection over union (mIoU) for panoptic segmentation as well. For SegInW, we adopt AP-Average and AP-Median as the evaluation metric.", + "bbox": [ + 89, + 628, + 482, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Generic Segmentation", + "text_level": 1, + "bbox": [ + 89, + 742, + 297, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness and generalization capability of our Prompt-DINO, we conduct evaluations on both indomain and out-domain datasets, with the results presented in Tab. 1.", + "bbox": [ + 89, + 763, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In-Domain Segmentation. We conduct in-domain experiments on COCO and LVIS. Compared to methods based on visual prompts, our Prompt-DINO achieved significantly better results. Compared to the previous work DINOv [22], we improve mask AP by 0.6 and box AP by 3.8", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on COCO. Compared to methods based on text prompts, Prompt-DINO also demonstrates superior results when using text prompts. For example, compared to APE [40], we lead by 3.5 in mask AP and 0.5 in box AP on LVIS.", + "bbox": [ + 511, + 342, + 903, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Out-Domain Segmentation. We use ADE-20K and Cityscapes as an out-domain dataset and conduct zero-shot testing on it, demonstrating the strong generalization ability of our Prompt-DINO. Under the out-domain setting (where data marked with * indicates performance reported under different domain settings), our method achieves the best performance on ADE20K across all methods for both prompt types, whether using visual prompts or text prompts. Compared to the state-of-the-art method APE on ADE20K, we improve PQ by 8.7, mask AP by 3.6, and box AP by 18.4, representing increases of $32.0\\%$ , $14.8\\%$ , and $10.5\\%$ , respectively. Among methods based on visual prompts, compared to the previous method DINOv, we improve PQ by 12.7, mask AP by 12.9, and box AP by 18.4, representing increases of $54.7\\%$ , $85.4\\%$ , and $128.7\\%$ , respectively. Compared to APE on Cityscapes, we also improve PQ by 1.0 and mIoU by 8.0.", + "bbox": [ + 511, + 402, + 905, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Segmentation in the Wild. To further evaluate the generalization capabilities of our model, we perform zero-shot testing on the SeginW benchmark, which includes 25 datasets. Compared to the previous work DINOv, when using visual prompts, we improve the average AP by 16.2 and the median AP by 14.5. Among text-prompt-based methods, we also surpass APE by 6.8 in average AP and 6.9 in median AP. This further validates the strong generalization ability of our Prompt-DINO.", + "bbox": [ + 511, + 660, + 905, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Effectiveness Analysis", + "text_level": 1, + "bbox": [ + 511, + 803, + 718, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of our proposed components, we conduct ablation studies on Text Prompt Alignment, Early Fusion, Order Alignment, and Data Engine. As our baseline, we implement the naive MaskDINO method with the FC head replaced with a user-prompted visual embed", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21294", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c060a39977021d7a6cd317d419f6976afa0eea310ebce380062ebff384d9a7f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTypeCityScapeSegInW
PQmIoUAP-avgAP-med
kMaX-DeepLab-L [50]in-domain68.483.5--
X-Decoder-L [59]out-domain38.152.022.332.3
OpenSeeD-L [54]out-domain41.447.836.138.7
APE-L [40]out-domain33.344.249.652.2
DINOv [22]out-domain--40.644.6
Prompt-DINO [40]out-domain34.352.256.459.1
", + "bbox": [ + 94, + 68, + 478, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d153fc5c4170300da08ca9fb90898aac05b643b1c6a3e36c210bcf415d9c5de2.jpg", + "table_caption": [ + "Table 2. Zero-shot Segmentation results of Prompt-DINO on the CityScape [7] and SegInW [59] dataset." + ], + "table_footnote": [], + "table_body": "
BaselineMethodADE20K (out-domain)
Text AlignEarly FusionOrder AlignPQmAPbAP
✓×××25.020.423.5
✓✓××32.725.530.5
✓✓✓×34.726.731.7
✓✓✓✓35.928.032.7
", + "bbox": [ + 93, + 210, + 480, + 285 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ding. The ablation results are presented in Tab. 3.", + "bbox": [ + 89, + 306, + 418, + 320 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Early Fusion. By comparing rows 1 and 2 in Tab. 3, it can be observed that incorporating text align in the backbone encoder (stage one of DINO) increases the mask AP and box AP on ADE20K by 5.1 and 7.0, respectively, and PQ by 7.7, compared to the baseline. This indicates that naive alignment of text and visual can enhance the model's general recognition capability. Comparing rows 2 and 3 in Tab. 3 shows that early fusion of visual prompts, guided by text alignment, significantly improves the model's performance, with mask AP and box AP on ADE20K increasing by 1.2 and 1.2, respectively, and PQ by 2.0. This demonstrates that visual prompts effectively complement features that are difficult to describe with text prompts and, under the guidance of text prompts, achieve more generalizable expressive power.", + "bbox": [ + 89, + 321, + 483, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f54bb0f07caa5d1ec7f50ff2be7491c0a9c7f05cfd7227fc660e53e168957e68.jpg", + "image_caption": [ + "Figure 5. Order Alignment Comparison. Left: Visual and text should demonstrate strong correlation, but the queries selected by the two prompts exhibit significant disorder. Right: Order alignment significantly improves the consistency of visual-text query selection and reinforces the alignment between corresponding semantics." + ], + "image_footnote": [], + "bbox": [ + 94, + 553, + 287, + 674 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b80a80b6a055e66a2f9122a8603a864e4e08f3ca850e750c8f8edac2873955b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 553, + 477, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Text-guided Visual Prompts. Rows 1, 2, and 3 in Tab. 3 show that merely adding text alignment provides limited improvement to the model. However, after early fusion to eliminate semantic ambiguity, the two types of prompts can serve as complementary information, providing stronger recognition capabilities. Next, we introduced order alignment during the query selection stage to further guide visual prompts with text. Rows 3 and 4 in Tab. 3 demonstrate that our two-stage guidance deeply", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/01ab29b901c447db640264a901dde471c25a001ec72a6dc947e097922d6e528a.jpg", + "image_caption": [ + "Figure 6. Annotation quality and quantity in the data engine." + ], + "image_footnote": [], + "bbox": [ + 517, + 63, + 700, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/244c7a54e5839ef223edd1855bfdd42f7d4c3832ace980e9a3522012903f9ddb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 63, + 903, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "aligns visual and text prompts, leading to further performance improvements. Fig. 5 also demonstrates that our order alignment method further enhances the alignment between visual and text, resulting in a more consistent selection order for queries.", + "bbox": [ + 511, + 213, + 903, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Data Engine. Within the data engine, we employ a cross-validation module to filter data annotated by the top-down and bottom-up pipelines. As illustrated in Fig. 6, the cross-validation module filters approximately $80.5\\%$ of instance annotations and enhances the annotation clip similarity from 0.16 to 0.23, highlighting the effectiveness of the cross-validation module in generating high-quality data. As shown in Tab. 4, incorporating data generated by our data engine into the training set improves Prompt-DINO's box AP by $12\\%$ on LVIS and $5.9\\%$ on ADE20K. To further demonstrate the effectiveness of RAP, we process a subset of SA-1B using different data engines in Tab. 5, revealing that RAP enhances model performance by $1.6\\%$ mask AP and $2.6\\%$ box AP on ADE20K. This indicates that the proposed data engine effectively enhances data quality, while RAP generates more diverse data, significantly improving the model's generalization capability.", + "bbox": [ + 511, + 287, + 906, + 546 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c5e19d07751aaf12a15780da3f75f7e20e8d1a8ca53ab3565ce5c6f20245deb1.jpg", + "table_caption": [ + "Table 3. Ablation of proposed components." + ], + "table_footnote": [], + "table_body": "
MethodLVIS (in-domain)ADE20K (out-domain)
mask APbox APPQmask APbox AP
w/o DataEngine46.047.930.823.426.8
w Data Engine55.659.935.928.032.7
", + "bbox": [ + 517, + 550, + 901, + 609 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aa96fb539b914b20f7034adf588dccb37a4be3087cfff6a1398c14e892cafe78.jpg", + "table_caption": [ + "Table 4. Ablation of Data Engine (Visual Prompt Results)." + ], + "table_footnote": [], + "table_body": "
DatasetCOCO (in-domain)ADE20K (out-domain)
mask APbox APPQmask APbox AP
COCO+GoldG49.855.718.211.011.6
COCO+SA-1B (QwenVL)50.156.722.317.820.4
COCO+SA-1B (QwenVL + RAP)50.858.122.519.423.0
", + "bbox": [ + 517, + 626, + 901, + 685 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation of Training Dataset (Text Prompt Results). For efficient training, we use a 2M subset of SA-1B.", + "bbox": [ + 511, + 686, + 905, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 715, + 633, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present Prompt-DINO, a text-guided visual Prompt DINO for generic segmentation via three innovations: 1) Early fusion of text, visual, and backbone features during initial encoding to resolve cross-modal ambiguity; 2) Order-aligned query selection for semantic-spatial consistency between text and visual cues in DETR decoding; 3) A generative data engine (RAP + dual-path verification) synthesizing 0.5B high-fidelity training instances, reducing label noise by $80.5\\%$ . Prompt-DINO sets new state-of-the-art results while enabling scalable adaptation to real-world object diversity.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21295", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant(KJZD20230923115106012, KJZD20230923 114916032, GJHZ20240218113604008), National Natural Science Foundation of China (No. 62106149).", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 219 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 5", + "[2] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 5", + "[3] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 5", + "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 7", + "[5] Runxi Cheng, Feng Xiong, Yongxian Wei, Wanyun Zhu, and Chun Yuan. Whoever started the interference should end it: Guiding data-free model merging via task vectors. arXiv preprint arXiv:2503.08099, 2025. 2", + "[6] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 3", + "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 7, 8", + "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 4", + "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 6", + "[10] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level" + ], + "bbox": [ + 93, + 227, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "labels. In European conference on computer vision, pages 540-557. Springer, 2022. 2", + "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2", + "[12] Yuchen Guan, Runxi Cheng, Kang Liu, and Chun Yuan. Enhancing logits distillation with plug&play kendall's $\\tau$ ranking loss, 2025. 5", + "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 7", + "[14] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, Adriane Boyd, et al. spacy: Industrial-strength natural language processing in python. 2020. 5", + "[15] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2989-2998, 2023. 7", + "[16] Qing Jiang, Feng Li, Zhaoyang Zeng, Tianhe Ren, Shilong Liu, and Lei Zhang. T-rex2: Towards generic object detection via text-visual prompt synergy. In European Conference on Computer Vision, pages 38-57. Springer, 2024. 2, 3, 4, 7", + "[17] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 7", + "[18] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 2", + "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2, 3, 5, 6, 7", + "[20] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 7", + "[21] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3041-3050, 2023. 2, 5, 7", + "[22] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Jianwei Yang, Chunyuan Li, et al. Visual in-context prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21296", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Pattern Recognition, pages 12861-12871, 2024. 2, 3, 4, 7, 8", + "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 5", + "[24] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 2, 3, 4, 5", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 7", + "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 5", + "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 2, 3, 4, 5, 6", + "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4", + "[29] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6", + "[30] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European conference on computer vision, pages 728-755. Springer, 2022. 2", + "[31] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4", + "[32] Ting Pan, Lulu Tang, Xinlong Wang, and Shiguang Shan. Tokenize anything via prompting. In European Conference on Computer Vision, pages 330-348. Springer, 2024. 2", + "[33] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 5", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3, 4, 6", + "[35] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 5", + "[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3", + "[37] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6", + "[38] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 7", + "[39] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 3, 7", + "[40] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13193–13203, 2024. 7, 8", + "[41] Jiaqi Wang, Pan Zhang, Tao Chu, Yuhang Cao, Yujie Zhou, Tong Wu, Bin Wang, Conghui He, and Dahua Lin. V3det: Vast vocabulary visual detection dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19844-19854, 2023. 7", + "[42] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6", + "[43] Yuanrui Wang, Cong Han, YafeiLi, Zhipeng Jin, Xiawei Li, SiNan Du, Wen Tao, Yi Yang, shuanglong li, Chun Yuan, and Liu Lin. Uniglyph: Unified segmentation-conditioned diffusion for precise visual text synthesis, 2025. 2", + "[44] Hongyang Wei, Shuaizheng Liu, Chun Yuan, and Lei Zhang. Perceive, understand and restore: Real-world image superresolution with autoregressive multimodal generative models, 2025." + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21297", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Yongxian Wei, Runxi Cheng, Weike Jin, Enneng Yang, Li Shen, Lu Hou, Sinan Du, Chun Yuan, Xiaochun Cao, and Dacheng Tao. Unifying multimodal large language model capabilities and modalities via model merging. arXiv preprint arXiv:2505.19892, 2025.", + "[46] Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Chun Yuan, and Dacheng Tao. Open-vocabulary customization from clip via data-free knowledge distillation. In The Thirteenth International Conference on Learning Representations, 2025. 2", + "[47] Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muenighoff. C-pack: Packaged resources to advance general chinese embedding, 2023. 6", + "[48] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 6", + "[49] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 2", + "[50] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In European Conference on Computer Vision, pages 288-307. Springer, 2022. 7, 8", + "[51] Yuqian Yuan, Wentong Li, Jian Liu, Dongqi Tang, Xinjie Luo, Chi Qin, Lei Zhang, and Jianke Zhu. Osprey: Pixel understanding with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28202-28211, 2024. 5", + "[52] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 4, 5, 6", + "[53] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems, 35:36067-36080, 2022. 2, 7", + "[54] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 7, 8" + ], + "bbox": [ + 89, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7", + "[56] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European conference on computer vision, pages 350-368. Springer, 2022. 2", + "[57] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4", + "[58] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting, 2023. 2", + "[59] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15116-15127, 2023. 7, 8", + "[60] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in neural information processing systems, 36:19769-19782, 2023. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 470 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21298", + "bbox": [ + 478, + 943, + 517, + 954 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Text-guided Visual Prompt DINO for Generic Segmentation/0364b0ce-3511-4546-ac4b-3ddb77cbd170_model.json b/2025/Text-guided Visual Prompt DINO for Generic Segmentation/0364b0ce-3511-4546-ac4b-3ddb77cbd170_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4796d038745986a931851fe07a5c5fdbec808a42 --- /dev/null +++ b/2025/Text-guided Visual Prompt DINO for Generic Segmentation/0364b0ce-3511-4546-ac4b-3ddb77cbd170_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.131, + 0.803, + 0.154 + ], + "angle": 0, + "content": "Text-guided Visual Prompt DINO for Generic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.18, + 0.848, + 0.234 + ], + "angle": 0, + "content": "Yuchen Guan\\(^{1*‡}\\), Chong Sun\\(^{2*†}\\), Canmiao Fu\\(^{2*}\\), Zhipeng Huang\\(^{2}\\), Chun Yuan\\(^{1†}\\), Chen Li\\(^{2}\\) \n\\(^{1}\\)Tsinghua Shenzhen International Graduate School, Tsinghua University \n\\(^{2}\\)WeChat AI, Tencent Inc." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.248, + 0.27, + 0.257 + ], + "angle": 0, + "content": "Zero-Shot Segmentation via Visual Prompt" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.256, + 0.302, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.256, + 0.493, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.359, + 0.402, + 0.372 + ], + "angle": 0, + "content": "Visual Cross-Image Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.256, + 0.7, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.256, + 0.896, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.612, + 0.359, + 0.798, + 0.373 + ], + "angle": 0, + "content": "Visual In-Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.374, + 0.262, + 0.383 + ], + "angle": 0, + "content": "Zero-Shot Segmentation via Text Prompt" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.383, + 0.268, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.383, + 0.434, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.486, + 0.354, + 0.499 + ], + "angle": 0, + "content": "Text In-Image Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.44, + 0.383, + 0.67, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.673, + 0.383, + 0.896, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.581, + 0.486, + 0.753, + 0.499 + ], + "angle": 0, + "content": "Text Panoptic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.501, + 0.307, + 0.511 + ], + "angle": 0, + "content": "Zero-Shot part Segmentation via Text/Visual Prompt" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.511, + 0.288, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.511, + 0.474, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.474, + 0.511, + 0.678, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.311, + 0.614, + 0.468, + 0.627 + ], + "angle": 0, + "content": "Visual Part Segmentation" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.511, + 0.772, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.511, + 0.895, + 0.609 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.72, + 0.615, + 0.865, + 0.628 + ], + "angle": 0, + "content": "Text Part Segmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.635, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Figure 1. Detection/Segmentation samples. Our Prompt-DINO is a text-guided visual Prompt DINO. Prompt-DINO supports single or multiple visual or text prompts for in-image or cross-image generic segmentation. Based on the granularity of the provided prompts, Prompt-DINO is capable of performing both panoptic and part segmentation." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.69, + 0.327, + 0.705 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.722, + 0.486, + 0.844 + ], + "angle": 0, + "content": "Recent advancements in multimodal vision models have highlighted limitations in late-stage feature fusion and suboptimal query selection for hybrid prompts open-world segmentation, alongside constraints from caption-derived vocabularies. To address these challenges, we propose Prompt-DINO, a text-guided visual Prompt DINO framework featuring three key innovations. First, we introduce an early fusion mechanism that unifies text/visual prompts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.907, + 0.889 + ], + "angle": 0, + "content": "and backbone features at the initial encoding stage, enabling deeper cross-modal interactions to resolve semantic ambiguities. Second, we design order-aligned query selection for DETR-based architectures, explicitly optimizing the structural alignment between text and visual queries during decoding to enhance semantic-spatial consistency. Third, we develop a generative data engine powered by the Recognize Anything via Prompt (RAP) model, which synthesizes 0.5B diverse training instances through a dualpath cross-verification pipeline, reducing label noise by \\(80.5\\%\\) compared to conventional approaches. Extensive experiments demonstrate that Prompt-DINO achieves state-of-the-art performance on open-world detection bench" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.851, + 0.234, + 0.863 + ], + "angle": 0, + "content": "* Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.863, + 0.439, + 0.888 + ], + "angle": 0, + "content": "† Corresponding Authors. Emails: waynecsun@tencent.com, yuanc@sz.tsinghua.edu.cn" + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.888, + 0.3, + 0.899 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Work done as interns at WeChat." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.851, + 0.439, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "21288" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.182 + ], + "angle": 0, + "content": "marks while significantly expanding semantic coverage beyond fixed-vocabulary constraints. Our work establishes a new paradigm for scalable multimodal detection and data generation in open-world scenarios. Data&Code are available at https://github.com/WeChatCV/WeVisionOne." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.196, + 0.223, + 0.212 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.221, + 0.485, + 0.629 + ], + "angle": 0, + "content": "Recent advancements in vision models [5, 43-46, 58] have focused on unifying diverse prompting strategies—such as text, visual, or hybrid inputs—to enable versatile object understanding and segmentation. Pioneering works like [16, 21, 22, 60] have demonstrated the potential of multimodal prompting for tasks ranging from open-set detection to segmentation. Despite their emerging capabilities, openworld detection and segmentation methods leveraging multimodal prompts demonstrate inherent technical deficiencies.1) Late-stage fusion of multimodal features. [22, 60] process text, visual or other prompts, and backbone features in separate subspaces, merging them only in later encoding stages. [16] similarly adopts contrastive learning to align modalities without joint optimization at the encoding stage. This delayed interaction restricts cross-modal alignment, particularly in ambiguous scenarios where early fusion of text and visual cues could disambiguate object semantics. 2) Insufficient exploration of query selection mechanisms. Most of these methods employ generic crossattention for text-guided query selection, or handle queries corresponding to text and visual prompts independently, none of which explicitly optimizes the structural alignment between queries corresponding to text and visual cues respectively. Like [16] processes text and visual prompts independently. This results in the selection of sub-optimal queries that do not promote optimization between text and images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Beyond these architectural limitations, a broader challenge persists across the field: the reliance on caption-derived vocabularies or fixed-vocabulary, which inherently constrain generalization to novel or composite objects. [16, 22, 60] derive their object vocabularies from entity words extracted from image captions or use fixed-vocabulary detector (like TAP [32]). While this approach avoids manually predefined categories, it remains bounded by the entities explicitly mentioned in the training captions. For instance, TRex-2 [16] employ TAP [32] to annotate SA-1B [19] with a category name from a dictionary of 2560 classes. Nevertheless, the reality that SA-1B [19] is associated with a vocabulary significantly exceeding 2560 substantially constrains the dataset's potential." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this work, we propose Prompt-DINO, a novel text-guided visual Prompt DINO, that addresses these challenges through two architectural innovations and a paradigm-shifting data engine. We pioneer an early fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.499 + ], + "angle": 0, + "content": "mechanism that deeply integrates text, visual, and backbone features at the initial encoding stage. Unlike [16, 22, 60], our model unifies multimodal signals into a cohesive representation from the outset. This enables richer cross-modal interactions, reducing ambiguity in open-set scenarios. And we introduce order-aligned query selection, a novel paradigm for DETR-based architectures that explicitly optimizes the alignment between text queries and visual queries during decoding. By enhancing semantic correlation and spatial consistency within query priorities, our method facilitates the optimization of text and image query selection in a unified direction, thereby surpassing the heuristic cross-attention mechanisms employed in prior research. Also, we overcome the fixed vocabulary bottleneck through a generative data engine with dual-path cross-verification pipeline. We proposal an novel Recognize Anything via Prompt (RAP) model, a 0.5B-parameter multimodal large language model(MLLM) that generates semantic labels for input images and their corresponding masks. Our engine leverages RAP to synthesize 0.5B diverse training instances with enhanced semantic relevance. To ensure label fidelity, we further implement a novel dual-path cross-verification pipeline that substantially reduces label noise by \\(80.5\\%\\) compared to conventional single-path approaches. This integrated data engine pioneers scalable open-world dataset generation, bridging limited annotation vocabularies with real-world detection needs." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.501, + 0.842, + 0.514 + ], + "angle": 0, + "content": "To summarize, our contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.516, + 0.905, + 0.574 + ], + "angle": 0, + "content": "- We propose an early fusion mechanism that integrates text, visual, and backbone features during initial encoding stage, unifying multimodal signals into a cohesive representation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.577, + 0.905, + 0.636 + ], + "angle": 0, + "content": "- We introduce order-aligned query selection, a new paradigm for DETR-based architectures that explicitly optimizes the alignment between text queries and visual queries during decoding." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.637, + 0.905, + 0.713 + ], + "angle": 0, + "content": "- We develop a generative data engine powered by our RAP, incorporating a dual-path cross-verification pipeline. This approach synthesizes diverse and semantically relevant training data, and establishes a new paradigm for scalable open-world data generation." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.516, + 0.905, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.722, + 0.655, + 0.737 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.742, + 0.806, + 0.758 + ], + "angle": 0, + "content": "2.1. General Recognition via Prompts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Recent advances in general recognition models (e.g., detection/segmentation [10, 11, 18, 24, 27, 30, 49, 53, 56]) have demonstrated that multimodal alignment between vision and language modalities serves as the cornerstone for open-world understanding. These frameworks typically employ joint text-visual prompting mechanisms to achieve object localization through aligned cross-modal feature spaces." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Text Prompt Paradigm. Existing methods predominantly focus on text-based cross-modal alignment. GLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "21289" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.912, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.412, + 0.908, + 0.482 + ], + "angle": 0, + "content": "Figure 2. The Overall Framework of Prompt-DINO. Prompt-DINO is constructed based on the DINO architecture. Images are processed by ViT to extract multi-scale features, while visual and text prompts are processed by deformable attention transformer layers and CLIP's text encoder, respectively, to derive embeddings. Through early fusion layers in the backbone encoder and perception-aware query selection, Prompt-DINO perform deep fusion and alignment of visual and text in two steps, with the final predictions generated by the Decoder." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.49, + 0.483, + 0.626 + ], + "angle": 0, + "content": "[24] pioneers the adaptation of image-text pretraining to downstream detection tasks through deep feature fusion, establishing language-aware visual representations. GroundingDINO [27] further extends this paradigm by demonstrating that multi-stage cross-attention fusion significantly enhances performance in open-vocabulary detection. While these approaches achieve remarkable progress, their heavy reliance on textual semantics limits their capacity to handle visual concepts with ambiguous linguistic descriptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Visual Prompt Paradigm. Complementary to text prompts, visual prompting (points/boxes/masks) provides geometric guidance for specific object localization. SAM [19] establishes a milestone in interactive segmentation through decoupled visual prompt encoding, though lacking semantic awareness. Subsequent works like SEEM [60] and DINOv [22] enhance this paradigm by integrating semantic context and exemplar-based reasoning. A notable hybrid approach is T-Rex2 [16], which combines both modalities but still treats visual prompts as auxiliary geometric cues. Our key insight lies in that existing methods inadequately exploit the semantic synergy between visual prompts and language guidance, leaving visual prompting's full potential underdeveloped for language-aware feature learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.427, + 0.866 + ], + "angle": 0, + "content": "2.2. Data Engine for Open-World Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Scalable data curation has become pivotal for training robust open-world detector/segmentation models. SAM2 [36]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.907, + 0.703 + ], + "angle": 0, + "content": "introduces a cyclic model-annotator collaboration framework, iteratively refining the SA-V dataset through human-machine interactions. Alternative strategies employ cross-model verification: YOLO-World [6] curates CC3M-Lite by relabeling CC3M [39] with GLIP and CLIP, while T-Rex2 [16] propagates pseudo-labels across SA-1B and LAION400M [34]. Current methods face three key limitations: (1) Over-reliance on noun phrase extraction from captions misses more detailed concepts and introduces too many redundant nouns; (2) Single-path verification propagates annotation errors; (3) Semantic ambiguity in freeform text introduces label noise. Our data engine addresses these issues and enables the automatic generation of high-quality training data with reduced noise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.708, + 0.605, + 0.723 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Motivated by the limitations observed in existing hybrid prompting methods for detection and segmentation tasks, we introduce Prompt-DINO, an end-to-end generic segmentation model built upon the DINO architecture [52]. Our model natively integrates multimodal prompting capabilities for both textual and visual inputs while maintaining architectural simplicity. In this section, we first formalize the promptable segmentation paradigm and outline the core design prompt encoders of Prompt-DINO (§3.1). We then systematically address three critical challenges: (1) early fusion with visual prompts through cross-modal feature align-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "21290" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.155 + ], + "angle": 0, + "content": "ment(§3.2), (2) order-aware query selection for maintaining semantic consistency(§3.3) and (3) construction of a scalable data engine for open-world generalization(§3.4). Finally, we introduce the training objective(§3.5)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.161, + 0.37, + 0.178 + ], + "angle": 0, + "content": "3.1. Promptable Segmentation Task" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.485, + 0.318 + ], + "angle": 0, + "content": "The Prompt-DINO processes an input image \\(\\mathcal{I}\\), with accompanying visual prompts \\(\\mathcal{P}^v = \\{\\mathcal{P}_1^v, \\dots, \\mathcal{P}_K^v\\}\\), and textual prompts \\(\\mathcal{P}^t = \\{\\mathcal{P}_1^t, \\dots, \\mathcal{P}_K^t\\}\\), where \\(K\\) is the number of interactive prompts per image. Following established architectures in DETR [57] and DINO [52], our system first extracts hierarchical visual features \\(F = \\{f_1, \\dots, f_L\\}\\) through a vision backbone network (e.g., ViT [9] or Swin Transformer [28]), where \\(L\\) represents the number of feature pyramid levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.41 + ], + "angle": 0, + "content": "Visual Prompt Encoder. Specifically, our visual encoding module employs \\( L \\) transformer blocks with deformable attention mechanisms [22]. For each target object, we initialize a learnable query embedding \\( q_{i} \\) and progressively refine it through multi-scale feature aggregation. The transformation at each layer \\( l \\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.419, + 0.483, + 0.439 + ], + "angle": 0, + "content": "\\[\nv _ {i} ^ {l} = M S D A \\left(v _ {i} ^ {l - 1}, f _ {l}\\right), l \\in [ 1, \\dots , L ], \\quad \\text {w i t h} \\quad v _ {i} ^ {0} = q _ {i} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.447, + 0.484, + 0.493 + ], + "angle": 0, + "content": "where \\(MSDA(\\cdot)\\) denotes multi-scale deformable attention operation. The final visual prompt representations \\(V = \\{v_{1},\\dots,v_{K}\\}\\) are obtained as \\(v_{i} = v_{i}^{L}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.613 + ], + "angle": 0, + "content": "Text Prompt Encoder. Building upon insights from recent open-world detection studies, we employ a CLIP [34] text encoder rather than conventional BERT [8] approaches used in earlier models like [24, 27]. This design choice enables better romantic correlation modeling for input words, phrases, and short sentences, generating text prompts \\( T = \\{t_1,\\dots,t_K\\} \\) that effectively guide cross-modal feature alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.675 + ], + "angle": 0, + "content": "The integrated representation (text(T), visual(V), multiscale features(F)) feeds into the backbone encoder for multimodal fusion, enabling dynamic adaptation to prompts while preserving cross-modal alignment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.388, + 0.699 + ], + "angle": 0, + "content": "3.2. Early Fusion with Visual Prompts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We are the first to propose the early fusion framework that integrates text prompts, visual prompts, and input image features through multi-modal interaction. While this fusion mechanism effectively reduces the modality gap between textual and visual features, enhances prompt-image correlation, and improves target recall, We found unintended consequences of severe hallucinations in our experiments. This phenomenon stems from the inherent limitation of conventional cross-attention mechanisms: they compulsively reconstruct Key features most similar to the Query features, even when all Key features are substantially dissimilar to the current Query. To address this non-target feature reconstruction issue, we introduce a novel gated cross" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.087, + 0.91, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Figure 3. Structure of Early Fusion Layer. Notably, the contrastive alignment of visual and text prompts is completed before early fusion. By employing Gated Cross-Attention, text prompts are reconstructed with image features, while visual prompts and image features mutually reconstruct each other." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.909, + 0.651 + ], + "angle": 0, + "content": "attention mechanism incorporating a learnable background token \\((B)\\). Reconstruction occurs exclusively when the Query-Key similarity exceeds an adaptive threshold; otherwise, the background token automatically activates to suppress hallucination. As illustrated in Fig. 3, our architecture comprises three core components: 1) a self-attention layer for intra-modal refinement, 2) the proposed gated cross-attention layer for inter-modal interaction, and 3) a feedforward neural network (FFN) for feature transformation. The cross-attention module implements three interaction pathways: feature-to-visual prompt, visual prompt-to-feature, and feature-to-text prompt. For visualization clarity, Fig. 3 presents only stage-2 features mapped to text space through our two-phase training scheme (detailed in §3.5). The enhanced cross-modal fusion operates through the following formulation:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.657, + 0.804, + 0.675 + ], + "angle": 0, + "content": "\\[\nt _ {i} ^ {\\prime} = \\operatorname {G a t e d A t t n} (f _ {i}, t _ {i}, t _ {i})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.677, + 0.905, + 0.694 + ], + "angle": 0, + "content": "\\[\nv _ {i} ^ {\\prime} = \\operatorname {G a t e d A t t n} \\left(f _ {i}, v _ {i}, v _ {i}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.696, + 0.802, + 0.713 + ], + "angle": 0, + "content": "\\[\nf _ {i} ^ {\\prime} = \\operatorname {G a t e d A t t n} (v _ {i}, f _ {i}, f _ {i})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.721, + 0.853, + 0.736 + ], + "angle": 0, + "content": "where the gated attention mechanism is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.741, + 0.907, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\operatorname {G a t e d A t t n} (Q, K, V) = \\operatorname {s o f t m a x} (\\frac {Q \\cdot [ K , B ]}{\\sqrt {d _ {k}}}) [ V, B ] \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.771, + 0.907, + 0.845 + ], + "angle": 0, + "content": "Here, \\(B\\) denotes our learnable background token. Following TRex-2 [16] methodology, we employ region-level contrastive learning to align visual and textual modalities, utilizing symmetric cross-entropy loss [34] for embedding alignment:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.834, + 0.905, + 0.874 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = - \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\log \\frac {v _ {i} \\cdot t _ {i}}{\\sum_ {j = 1} ^ {K} v _ {i} \\cdot t _ {j}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.872, + 0.905, + 0.902 + ], + "angle": 0, + "content": "In our experience, symmetric cross-entropy losses have better convergence than InfoNCE loss [31]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.518, + 0.956 + ], + "angle": 0, + "content": "21291" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "In the context of training detection and segmentation tasks, it is common practice to construct negative samples. For negative prompt construction, while text prompts naturally leverage category vocabularies for negative sampling, visual prompts require explicit negative construction. We address this by aggregating visual prompt features across all categories within each batch and computing inter-category averages to generate negative visual prompts. This approach ensures sufficient diversity in negative samples while maintaining semantic relevance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.482, + 0.392 + ], + "angle": 0, + "content": "To mitigate annotation granularity discrepancies across datasets (e.g., \"bird\" in Dataset A vs. \"pigeon\" in Dataset B), we implement a dataset-aware sampling strategy. Our novel data sampler ensures each training batch contains only intra-dataset samples, preventing semantically similar but categorically distinct concepts from becoming conflicting negative samples. This is particularly critical in visual-prompt open-world detection/segmentation scenarios where label hierarchies may differ substantially across source datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.402, + 0.347, + 0.418 + ], + "angle": 0, + "content": "3.3. Order-aware query selection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.483, + 0.649 + ], + "angle": 0, + "content": "Prompt-DINO adopts a DETR-based architecture that selects top-K encoder features through similarity computation between encoder features and prompt embeddings, following DINO [52] and MaskDINO [21]. However, our approach fundamentally diverges from existing methods by simultaneously considering both textual-visual query similarity and inter-query correlations with visual prompts. To further strengthen the alignment between visual and textual prompts, we introduce order-aware alignment during query selection, which establishes the second-level alignment between visual and textual prompts during training. Specifically, we employ Kendall's \\(\\tau\\) coefficient to quantify the ordinal consistency [12] between queries selected by visual prompts and text prompts, the Kendall's \\(\\tau\\) coefficient can be expressed as: \\(P_{\\mathrm{c}} - P_{\\mathrm{d}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.638, + 0.482, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\tau = \\frac {P _ {c} - P _ {d}}{\\frac {1}{2} N (N - 1)} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.744 + ], + "angle": 0, + "content": "where \\(P_{c}\\) represents concordant pairs and \\(P_{d}\\) represents discordant pairs, \\(N\\) represents the total number of queries. Given the similarity score between query embeddings and prompt embeddings \\(S^{v} = s_{1}^{v},\\ldots ,s_{N}^{v}\\) and \\(S^t = s_1^t,\\dots,s_N^t\\) we calculate \\(P_{c}\\) and \\(P_{d}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.744, + 0.482, + 0.779 + ], + "angle": 0, + "content": "\\[\n\\tau = \\frac {\\sum_ {i} \\sum_ {j < i} \\operatorname {s g n} \\left(s _ {i} ^ {t} - s _ {j} ^ {t}\\right) \\cdot \\operatorname {s g n} \\left(s _ {i} ^ {v} - s _ {j} ^ {v}\\right)}{\\frac {1}{2} N (N - 1)} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Since the sgn is not differentiable, we approximate the sign function with the tanh function and utilize the opposite of \\(\\tau\\) as the order alignment loss:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.822, + 0.482, + 0.856 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {o r d e r}} = - \\frac {\\sum_ {i} \\sum_ {j < i} \\tanh \\left(s _ {i} ^ {t} - s _ {j} ^ {t}\\right) \\cdot \\tanh \\left(s _ {i} ^ {v} - s _ {j} ^ {v}\\right)}{\\frac {1}{2} N (N - 1)} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.857, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our proposed alignment mechanism operates at two complementary levels: The primary alignment directly optimizes the learning objective through \\(\\mathcal{L}_{\\text{align}}\\) (as given in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "Eq. 4), while our novel order-aware alignment provides indirect assurance through two critical aspects: (1) Semantic proximity between corresponding text and visual prompts for candidate queries naturally reduces overall training complexity by minimizing feature space discrepancies; (2) Training-inference consistency preservation ensures equivalent substitutability between text and visual prompts during inference through maintained representational parity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Fig. 5 demonstrates the effectiveness of our proposed method. Prior to applying the order-aware loss optimization, we observe significant disorder in both textual and visual prompt queries, with no discernible alignment between corresponding semantic representations. The optimization process induces a remarkable transformation - post-optimization analysis reveals substantially enhanced consistency between semantically equivalent textual and visual prompt queries." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.646, + 0.377 + ], + "angle": 0, + "content": "3.4. Data Engine" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.906, + 0.566 + ], + "angle": 0, + "content": "Effective visual concept learning from rich image-text pairs is fundamental to enhancing the generalization capabilities of open-vocabulary object detectors. Existing methods predominantly rely on captioning models([3, 23, 26] etc.), coupled with lexical analysis tools like SpaCy [14] to extract noun phrases. These noun phrases are subsequently employed to generate pseudo-labels via open-set detection models ([24, 27]). Nonetheless, these techniques have significant drawbacks. They frequently omit valid labels, introduce erroneous annotations, and are susceptible to irrelevant lexical noise. All of these issues detrimentally affect detection performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.906, + 0.688 + ], + "angle": 0, + "content": "To address these challenges, we propose a novel Recognize Anything via Prompting (RAP) model, a 0.5B MLLM, which significantly enhances annotation accuracy. Our framework further integrates an innovative dual-path cross-validation pipeline that substantially reduces label noise. This comprehensive approach enables the fully automated generation of high-quality open-world detection datasets while preserving the fidelity of annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.885 + ], + "angle": 0, + "content": "Recognize Anything via Prompting. Despite the remarkable capabilities of the Segment Anything Model (SAM) [19] in detecting a wide array of objects within images, and the substantial progress achieved in region-level understanding through multimodal models [1, 2, 33, 35, 51], a critical question persists: Can we effectively assign semantic labels to arbitrary mask prompts? To address this challenge, we introduce RAP, a novel framework that bridges visual segmentation with semantic recognition. As illustrated in Fig. 4, our RAP model processes raw image inputs alongside their corresponding object masks to generate precise semantic annotations for the specified regions. The proposed RAP comprises three key components:" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1. Feature extraction. We perform masked pooling over" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21292" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.09, + 0.633, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.089, + 0.906, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.235, + 0.908, + 0.263 + ], + "angle": 0, + "content": "Figure 4. Data Engine and RAP Frameworks. Left: The data engine with dual-path cross-verification pipeline. Right: The proposed Recognize Anything via Prompting (RAP), a 0.5B-parameter MLLM model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.271, + 0.483, + 0.391 + ], + "angle": 0, + "content": "multi-layer image features to extract region-specific features corresponding to input masks, and concatenate the mask token embeddings obtained via a projection layer with the original image tokens to form a unified representation. This preserves both global contextual information from the complete image and localized object-specific features within masked regions, effectively integrating scene-level understanding with instance-level visual cues." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.394, + 0.483, + 0.56 + ], + "angle": 0, + "content": "2. Ground Data Cleaning. To standardize the format of our ground truth data, we leveraged the Qwen model [48] to eliminate redundant descriptive phrases through comprehensive text normalization. Specifically, we transformed verbose expressions into their canonical semantic equivalents (e.g., converting \"any part of the sky\" to \"sky\"). This preprocessing strategy serves dual purposes: (1) it reduces the RAP's learning complexity by minimizing lexical variability, and (2) it mitigates the risk of hallucinations by ensuring precise semantic alignment between textual annotations and visual content." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.562, + 0.483, + 0.669 + ], + "angle": 0, + "content": "3. Training. The proposed methodology employs a two-stage training strategy. During the first stage, we conduct pre-training using 10M samples, specifically optimizing only the projector while keeping other model components frozen. The second stage implements instruction-tuning, where we unfreeze and jointly optimize the entire model to enhance task-specific adaptation capabilities." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.67, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Dual-path cross-verification pipeline. We propose a dual-pipeline annotation framework combining bottom-up and top-down approaches with cross-verification filtering, consisting of three key components:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.824 + ], + "angle": 0, + "content": "1. Top-Down Pipeline. Our top-down pipeline leverages the Qwen-VL model [42] to extract semantic tags from images. These tags are subsequently fed into an open-set detection model to generate corresponding bounding boxes. This pipeline demonstrates strong capability in detecting large-scale salient objects (e.g., trees, roads, sky)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2. Bottom-Up Pipeline. To address the inherent limitation of top-down pipeline in capturing fine-grained details, we implement a complementary bottom-up strategy using the RAP integrated with SAM [19]. This combination enables effective detection of smaller objects and detailed" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.905, + 0.3 + ], + "angle": 0, + "content": "components (e.g., eyes, noses, windows) that are frequently missed by conventional top-down pipeline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.906, + 0.452 + ], + "angle": 0, + "content": "3. Cross-Verification. The detection outputs from both pipelines undergo rigorous validation through a three-stage process. First, We employ the Hungarian matching algorithm to establish correspondences between bounding boxes from different pipelines. Next, we utilize the BGE model [47] to compute similarity scores between the tags associated with the matched bounding boxes. Finally, we filter out the candidate pairs whose similarity scores were below the threshold of empirical determination to further clean up the annotated results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.461, + 0.695, + 0.478 + ], + "angle": 0, + "content": "3.5. Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.906, + 0.635 + ], + "angle": 0, + "content": "Building upon previous DINO-like models [52], we employ a combination of L1 loss and GIoU loss [37] for bounding box regression, while for mask segmentation, we adopt cross-entropy loss and dice loss [29]. Following the methodology established in GroudingDINO [27], we utilize contrastive loss between predicted objects and text prompts for classification tasks. To accelerate model convergence, we incorporate denoising training as implemented in DINO [52]. The complete objective function for Prompt DINO can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.647, + 0.905, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {b b o x} + \\mathcal {L} _ {m a s k} + \\mathcal {L} _ {D N} + \\mathcal {L} _ {a l i g n} + \\mathcal {L} _ {o r d e r} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.908, + 0.737 + ], + "angle": 0, + "content": "Our training methodology adopts a progressive multi-stage strategy: we first optimize the text prompt while keeping the visual components fixed, followed by joint optimization of both textual and visual prompt representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.749, + 0.646, + 0.765 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.612, + 0.79 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.9 + ], + "angle": 0, + "content": "In Prompt-DINO, we employ ViT-L [9] as the backbone and fine-tune CLIP-L [34] as our text prompt encoder. The visual prompt encoder consists of 6 layers of deformable cross-attention transformer layers, with a hidden dimension of 256. We adopt the \"two-stage\" structure proposed by DINO [52], selecting 900 queries for decoding after the backbone encoder." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21293" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.069, + 0.904, + 0.293 + ], + "angle": 0, + "content": "
MethodBackbonePrompt TypeCOCO (in-domain)LVIS (in-domain)ADE20K (out-domain)
mask APbox APmask APbox APPQmask APbox AP
Mask2Former [4]Swin-LClose-Set48.652.1-----
OneFormer [15]Swin-LClose-Set48.9------
MaskDINO* [21]Swin-LClose-Set50.656.2-----
kMaX-DeepLab [50]Swin-LClose-Set58.1------
GLIPv2 [53]Swin-HText Prompt48.9------
X-Decoder [59]DaViT-LText Prompt46.7---21.813.138.1
OpenSeeD [54]Swin-LText Prompt53.258.221.023.019.715.017.7
T-Rex2 [16]Swin-LText Prompt-52.2*-45.8*---
APE [40]ViT-LText Prompt49.358.353.059.627.224.429.6
DINOv [22]Swin-LVisual Prompt50.454.2--23.215.114.3
T-Rex2 [16]Swin-LVisual Prompt-46.5*-45.3*---
Prompt-DINOViT-LText Prompt51.158.656.560.129.624.628.6
Prompt-DINOViT-LVisual Prompt51.058.055.659.935.928.032.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.295, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Table 1. Generic Segmentation results of Prompt-DINO on multiple datasets. \"-\" denotes the model does not have performance reported or does not have ability for the specific task. \"+\" denotes the performance reported under different domain settings. MaskDINO uses the results reported in DINOv [22]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.482, + 0.629 + ], + "angle": 0, + "content": "Datasets. During the training process, we utilized three categories of datasets: open-source detection datasets. These include COCO [25] with 118K images and LVIS [13] with 160K images, both of which have mask annotations. For datasets without mask annotations, such as Object365 [38], OID [20], and V3Det [41], we first apply SAM to generate mask annotations. Due to the presence of low-quality data and long-tail distribution in Object365 and OID, which hinder general object recognition, we resampled them, selecting 700K and 1M data points, respectively. grounding datasets. This includes the GoldG [17] dataset with 100K images. self-built datasets. We re-announce and filter CC3M [39] and SA-1B [19] using the RAP and dual-path cross-validation data engine introduced in §3.4, yielding high-quality datasets of 3M and 7M images, respectively, which constitute our primary training data. For testing, we utilize COCO and LVIS for in-domain evaluation and ADE20K [55], SeginW [59], and Cityscapes [7] for out-domain evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Metrics. We employ Average Precision (AP) as the evaluation metric, utilizing mask AP for instance segmentation, and box AP for object detection. For ADE20K and CtiyScapes, we employ Panoptic Quality(PQ) and mean intersection over union (mIoU) for panoptic segmentation as well. For SegInW, we adopt AP-Average and AP-Median as the evaluation metric." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.298, + 0.759 + ], + "angle": 0, + "content": "4.2. Generic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.825 + ], + "angle": 0, + "content": "To evaluate the effectiveness and generalization capability of our Prompt-DINO, we conduct evaluations on both indomain and out-domain datasets, with the results presented in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In-Domain Segmentation. We conduct in-domain experiments on COCO and LVIS. Compared to methods based on visual prompts, our Prompt-DINO achieved significantly better results. Compared to the previous work DINOv [22], we improve mask AP by 0.6 and box AP by 3.8" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.905, + 0.403 + ], + "angle": 0, + "content": "on COCO. Compared to methods based on text prompts, Prompt-DINO also demonstrates superior results when using text prompts. For example, compared to APE [40], we lead by 3.5 in mask AP and 0.5 in box AP on LVIS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.906, + 0.659 + ], + "angle": 0, + "content": "Out-Domain Segmentation. We use ADE-20K and Cityscapes as an out-domain dataset and conduct zero-shot testing on it, demonstrating the strong generalization ability of our Prompt-DINO. Under the out-domain setting (where data marked with * indicates performance reported under different domain settings), our method achieves the best performance on ADE20K across all methods for both prompt types, whether using visual prompts or text prompts. Compared to the state-of-the-art method APE on ADE20K, we improve PQ by 8.7, mask AP by 3.6, and box AP by 18.4, representing increases of \\(32.0\\%\\), \\(14.8\\%\\), and \\(10.5\\%\\), respectively. Among methods based on visual prompts, compared to the previous method DINOv, we improve PQ by 12.7, mask AP by 12.9, and box AP by 18.4, representing increases of \\(54.7\\%\\), \\(85.4\\%\\), and \\(128.7\\%\\), respectively. Compared to APE on Cityscapes, we also improve PQ by 1.0 and mIoU by 8.0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.906, + 0.797 + ], + "angle": 0, + "content": "Segmentation in the Wild. To further evaluate the generalization capabilities of our model, we perform zero-shot testing on the SeginW benchmark, which includes 25 datasets. Compared to the previous work DINOv, when using visual prompts, we improve the average AP by 16.2 and the median AP by 14.5. Among text-prompt-based methods, we also surpass APE by 6.8 in average AP and 6.9 in median AP. This further validates the strong generalization ability of our Prompt-DINO." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.72, + 0.82 + ], + "angle": 0, + "content": "4.3. Effectiveness Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To validate the effectiveness of our proposed components, we conduct ablation studies on Text Prompt Alignment, Early Fusion, Order Alignment, and Data Engine. As our baseline, we implement the naive MaskDINO method with the FC head replaced with a user-prompted visual embed" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "21294" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.069, + 0.48, + 0.178 + ], + "angle": 0, + "content": "
MethodTypeCityScapeSegInW
PQmIoUAP-avgAP-med
kMaX-DeepLab-L [50]in-domain68.483.5--
X-Decoder-L [59]out-domain38.152.022.332.3
OpenSeeD-L [54]out-domain41.447.836.138.7
APE-L [40]out-domain33.344.249.652.2
DINOv [22]out-domain--40.644.6
Prompt-DINO [40]out-domain34.352.256.459.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.18, + 0.482, + 0.208 + ], + "angle": 0, + "content": "Table 2. Zero-shot Segmentation results of Prompt-DINO on the CityScape [7] and SegInW [59] dataset." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.211, + 0.481, + 0.286 + ], + "angle": 0, + "content": "
BaselineMethodADE20K (out-domain)
Text AlignEarly FusionOrder AlignPQmAPbAP
✓×××25.020.423.5
✓✓××32.725.530.5
✓✓✓×34.726.731.7
✓✓✓✓35.928.032.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.155, + 0.289, + 0.419, + 0.303 + ], + "angle": 0, + "content": "Table 3. Ablation of proposed components." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.307, + 0.419, + 0.321 + ], + "angle": 0, + "content": "ding. The ablation results are presented in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.549 + ], + "angle": 0, + "content": "Effectiveness of Early Fusion. By comparing rows 1 and 2 in Tab. 3, it can be observed that incorporating text align in the backbone encoder (stage one of DINO) increases the mask AP and box AP on ADE20K by 5.1 and 7.0, respectively, and PQ by 7.7, compared to the baseline. This indicates that naive alignment of text and visual can enhance the model's general recognition capability. Comparing rows 2 and 3 in Tab. 3 shows that early fusion of visual prompts, guided by text alignment, significantly improves the model's performance, with mask AP and box AP on ADE20K increasing by 1.2 and 1.2, respectively, and PQ by 2.0. This demonstrates that visual prompts effectively complement features that are difficult to describe with text prompts and, under the guidance of text prompts, achieve more generalizable expressive power." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.554, + 0.288, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.554, + 0.478, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.678, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Figure 5. Order Alignment Comparison. Left: Visual and text should demonstrate strong correlation, but the queries selected by the two prompts exhibit significant disorder. Right: Order alignment significantly improves the consistency of visual-text query selection and reinforces the alignment between corresponding semantics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Text-guided Visual Prompts. Rows 1, 2, and 3 in Tab. 3 show that merely adding text alignment provides limited improvement to the model. However, after early fusion to eliminate semantic ambiguity, the two types of prompts can serve as complementary information, providing stronger recognition capabilities. Next, we introduced order alignment during the query selection stage to further guide visual prompts with text. Rows 3 and 4 in Tab. 3 demonstrate that our two-stage guidance deeply" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.064, + 0.701, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.064, + 0.904, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.187, + 0.89, + 0.202 + ], + "angle": 0, + "content": "Figure 6. Annotation quality and quantity in the data engine." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.905, + 0.289 + ], + "angle": 0, + "content": "aligns visual and text prompts, leading to further performance improvements. Fig. 5 also demonstrates that our order alignment method further enhances the alignment between visual and text, resulting in a more consistent selection order for queries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.907, + 0.547 + ], + "angle": 0, + "content": "Effectiveness of Data Engine. Within the data engine, we employ a cross-validation module to filter data annotated by the top-down and bottom-up pipelines. As illustrated in Fig. 6, the cross-validation module filters approximately \\(80.5\\%\\) of instance annotations and enhances the annotation clip similarity from 0.16 to 0.23, highlighting the effectiveness of the cross-validation module in generating high-quality data. As shown in Tab. 4, incorporating data generated by our data engine into the training set improves Prompt-DINO's box AP by \\(12\\%\\) on LVIS and \\(5.9\\%\\) on ADE20K. To further demonstrate the effectiveness of RAP, we process a subset of SA-1B using different data engines in Tab. 5, revealing that RAP enhances model performance by \\(1.6\\%\\) mask AP and \\(2.6\\%\\) box AP on ADE20K. This indicates that the proposed data engine effectively enhances data quality, while RAP generates more diverse data, significantly improving the model's generalization capability." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.551, + 0.903, + 0.611 + ], + "angle": 0, + "content": "
MethodLVIS (in-domain)ADE20K (out-domain)
mask APbox APPQmask APbox AP
w/o DataEngine46.047.930.823.426.8
w Data Engine55.659.935.928.032.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.534, + 0.613, + 0.885, + 0.627 + ], + "angle": 0, + "content": "Table 4. Ablation of Data Engine (Visual Prompt Results)." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.627, + 0.903, + 0.686 + ], + "angle": 0, + "content": "
DatasetCOCO (in-domain)ADE20K (out-domain)
mask APbox APPQmask APbox AP
COCO+GoldG49.855.718.211.011.6
COCO+SA-1B (QwenVL)50.156.722.317.820.4
COCO+SA-1B (QwenVL + RAP)50.858.122.519.423.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.687, + 0.906, + 0.715 + ], + "angle": 0, + "content": "Table 5. Ablation of Training Dataset (Text Prompt Results). For efficient training, we use a 2M subset of SA-1B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.716, + 0.634, + 0.731 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We present Prompt-DINO, a text-guided visual Prompt DINO for generic segmentation via three innovations: 1) Early fusion of text, visual, and backbone features during initial encoding to resolve cross-modal ambiguity; 2) Order-aligned query selection for semantic-spatial consistency between text and visual cues in DETR decoding; 3) A generative data engine (RAP + dual-path verification) synthesizing 0.5B high-fidelity training instances, reducing label noise by \\(80.5\\%\\). Prompt-DINO sets new state-of-the-art results while enabling scalable adaptation to real-world object diversity." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21295" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This work was partially supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant(KJZD20230923115106012, KJZD20230923 114916032, GJHZ20240218113604008), National Natural Science Foundation of China (No. 62106149)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.22 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.285 + ], + "angle": 0, + "content": "[1] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.286, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[2] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.342, + 0.484, + 0.41 + ], + "angle": 0, + "content": "[3] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.484, + 0.481 + ], + "angle": 0, + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.482, + 0.484, + 0.537 + ], + "angle": 0, + "content": "[5] Runxi Cheng, Feng Xiong, Yongxian Wei, Wanyun Zhu, and Chun Yuan. Whoever started the interference should end it: Guiding data-free model merging via task vectors. arXiv preprint arXiv:2503.08099, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[6] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.484, + 0.691 + ], + "angle": 0, + "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.692, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "labels. In European conference on computer vision, pages 540-557. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[12] Yuchen Guan, Runxi Cheng, Kang Liu, and Chun Yuan. Enhancing logits distillation with plug&play kendall's \\(\\tau\\) ranking loss, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.904, + 0.316 + ], + "angle": 0, + "content": "[14] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, Adriane Boyd, et al. spacy: Industrial-strength natural language processing in python. 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.907, + 0.386 + ], + "angle": 0, + "content": "[15] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2989-2998, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[16] Qing Jiang, Feng Li, Zhaoyang Zeng, Tianhe Ren, Shilong Liu, and Lei Zhang. T-rex2: Towards generic object detection via text-visual prompt synergy. In European Conference on Computer Vision, pages 38-57. Springer, 2024. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.444, + 0.905, + 0.512 + ], + "angle": 0, + "content": "[17] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[18] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[20] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[21] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3041-3050, 2023. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[22] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Jianwei Yang, Chunyuan Li, et al. Visual in-context prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "21296" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "and Pattern Recognition, pages 12861-12871, 2024. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[24] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[29] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[30] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European conference on computer vision, pages 728-755. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[31] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[32] Ting Pan, Lulu Tang, Xinlong Wang, and Shiguang Shan. Tokenize anything via prompting. In European Conference on Computer Vision, pages 330-348. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[33] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[35] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[37] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[38] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[39] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[40] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13193–13203, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[41] Jiaqi Wang, Pan Zhang, Tao Chu, Yuhang Cao, Yujie Zhou, Tong Wu, Bin Wang, Conghui He, and Dahua Lin. V3det: Vast vocabulary visual detection dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19844-19854, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[42] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[43] Yuanrui Wang, Cong Han, YafeiLi, Zhipeng Jin, Xiawei Li, SiNan Du, Wen Tao, Yi Yang, shuanglong li, Chun Yuan, and Liu Lin. Uniglyph: Unified segmentation-conditioned diffusion for precise visual text synthesis, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[44] Hongyang Wei, Shuaizheng Liu, Chun Yuan, and Lei Zhang. Perceive, understand and restore: Real-world image superresolution with autoregressive multimodal generative models, 2025." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "21297" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[45] Yongxian Wei, Runxi Cheng, Weike Jin, Enneng Yang, Li Shen, Lu Hou, Sinan Du, Chun Yuan, Xiaochun Cao, and Dacheng Tao. Unifying multimodal large language model capabilities and modalities via model merging. arXiv preprint arXiv:2505.19892, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.164, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[46] Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Chun Yuan, and Dacheng Tao. Open-vocabulary customization from clip via data-free knowledge distillation. In The Thirteenth International Conference on Learning Representations, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.235, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[47] Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muenighoff. C-pack: Packaged resources to advance general chinese embedding, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.279, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[48] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.489, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[49] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.561, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[50] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In European Conference on Computer Vision, pages 288-307. Springer, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.618, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[51] Yuqian Yuan, Wentong Li, Jian Liu, Dongqi Tang, Xinjie Luo, Chi Qin, Lei Zhang, and Jianke Zhu. Osprey: Pixel understanding with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28202-28211, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[52] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[53] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems, 35:36067-36080, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[54] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[56] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European conference on computer vision, pages 350-368. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[57] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.904, + 0.319 + ], + "angle": 0, + "content": "[58] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[59] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15116-15127, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.471 + ], + "angle": 0, + "content": "[60] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in neural information processing systems, 36:19769-19782, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.519, + 0.955 + ], + "angle": 0, + "content": "21298" + } + ] +] \ No newline at end of file diff --git a/2025/Text-guided Visual Prompt DINO for Generic Segmentation/0364b0ce-3511-4546-ac4b-3ddb77cbd170_origin.pdf b/2025/Text-guided Visual Prompt DINO for Generic Segmentation/0364b0ce-3511-4546-ac4b-3ddb77cbd170_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dfd4eae4e4a846e25c56ae8a5e1743bb023b7c6a --- /dev/null +++ b/2025/Text-guided Visual Prompt DINO for Generic Segmentation/0364b0ce-3511-4546-ac4b-3ddb77cbd170_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc57e9df335dab9d6c864efddeaeda6c8a7d39f5edc78c7c345b163590492bd8 +size 3423199 diff --git a/2025/Text-guided Visual Prompt DINO for Generic Segmentation/full.md b/2025/Text-guided Visual Prompt DINO for Generic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b1429823b6f29ae4aad2e129a6368e7d2492a1ea --- /dev/null +++ b/2025/Text-guided Visual Prompt DINO for Generic Segmentation/full.md @@ -0,0 +1,366 @@ +# Text-guided Visual Prompt DINO for Generic Segmentation + +Yuchen Guan $^{1*‡}$ , Chong Sun $^{2*†}$ , Canmiao Fu $^{2*}$ , Zhipeng Huang $^{2}$ , Chun Yuan $^{1†}$ , Chen Li $^{2}$ $^{1}$ Tsinghua Shenzhen International Graduate School, Tsinghua University + $^{2}$ WeChat AI, Tencent Inc. + +Zero-Shot Segmentation via Visual Prompt + +![](images/e45deb8b198c74b7630209764108ebb61d60c61a538f52efb612fb59470c6d18.jpg) + +![](images/eaebf852eba30e883eff7ec2a75fc9b18025760ac819f6d6d74b26e735242734.jpg) + +Visual Cross-Image Segmentation + +![](images/344751fe6bbf4ec16e7cdd984b13d713b4dd4fd5c0d63fdbf6bf0d19219ed9ae.jpg) + +![](images/3b56b0d126228125b9af241c58353b01639eb23a8ff83f6a18c340dc88d47ec6.jpg) + +Visual In-Image Segmentation + +Zero-Shot Segmentation via Text Prompt + +![](images/4863ca13a43dacc00841f33624d25d6f860cef1b950829c7a95604e336ad325d.jpg) + +![](images/bf18a0dee20706dfe0821158eae9379cb1666290a01c69d002e96a6611306641.jpg) + +Text In-Image Segmentation + +![](images/53ef319c81eab998c5d0685eeb3e283914d3ad8a93348245339ee6840ad3b30d.jpg) + +![](images/e19643d3f04e3b5794d8bf8a5ef2f7e301c14245e7ed5b44a49a1645b98f870b.jpg) + +Text Panoptic Segmentation + +Zero-Shot part Segmentation via Text/Visual Prompt + +![](images/9df8bf5121a3e1cc6b08953b1cef5d3a0ee690e40cba00a72f329948c7dfb30d.jpg) +Figure 1. Detection/Segmentation samples. Our Prompt-DINO is a text-guided visual Prompt DINO. Prompt-DINO supports single or multiple visual or text prompts for in-image or cross-image generic segmentation. Based on the granularity of the provided prompts, Prompt-DINO is capable of performing both panoptic and part segmentation. + +![](images/80779919295d9afce6d868cafd836ea76fadca1c0670bf94789f9fe64e0aa97b.jpg) + +![](images/83f5ab435ed5f02ce9ec3bd1697698c2c228a088f649f3a452b3fb39c2899b1d.jpg) + +Visual Part Segmentation + +![](images/6703fb5ff0fc67b914d10a2def8853e5c821b28d48ca721d3605ac56d0293707.jpg) + +![](images/1920781344e8b6fb191b66b01cf79d33013a837af902051c8f1306b2608e2bb3.jpg) + +Text Part Segmentation + +# Abstract + +Recent advancements in multimodal vision models have highlighted limitations in late-stage feature fusion and suboptimal query selection for hybrid prompts open-world segmentation, alongside constraints from caption-derived vocabularies. To address these challenges, we propose Prompt-DINO, a text-guided visual Prompt DINO framework featuring three key innovations. First, we introduce an early fusion mechanism that unifies text/visual prompts + +and backbone features at the initial encoding stage, enabling deeper cross-modal interactions to resolve semantic ambiguities. Second, we design order-aligned query selection for DETR-based architectures, explicitly optimizing the structural alignment between text and visual queries during decoding to enhance semantic-spatial consistency. Third, we develop a generative data engine powered by the Recognize Anything via Prompt (RAP) model, which synthesizes 0.5B diverse training instances through a dualpath cross-verification pipeline, reducing label noise by $80.5\%$ compared to conventional approaches. Extensive experiments demonstrate that Prompt-DINO achieves state-of-the-art performance on open-world detection bench + +marks while significantly expanding semantic coverage beyond fixed-vocabulary constraints. Our work establishes a new paradigm for scalable multimodal detection and data generation in open-world scenarios. Data&Code are available at https://github.com/WeChatCV/WeVisionOne. + +# 1. Introduction + +Recent advancements in vision models [5, 43-46, 58] have focused on unifying diverse prompting strategies—such as text, visual, or hybrid inputs—to enable versatile object understanding and segmentation. Pioneering works like [16, 21, 22, 60] have demonstrated the potential of multimodal prompting for tasks ranging from open-set detection to segmentation. Despite their emerging capabilities, openworld detection and segmentation methods leveraging multimodal prompts demonstrate inherent technical deficiencies.1) Late-stage fusion of multimodal features. [22, 60] process text, visual or other prompts, and backbone features in separate subspaces, merging them only in later encoding stages. [16] similarly adopts contrastive learning to align modalities without joint optimization at the encoding stage. This delayed interaction restricts cross-modal alignment, particularly in ambiguous scenarios where early fusion of text and visual cues could disambiguate object semantics. 2) Insufficient exploration of query selection mechanisms. Most of these methods employ generic crossattention for text-guided query selection, or handle queries corresponding to text and visual prompts independently, none of which explicitly optimizes the structural alignment between queries corresponding to text and visual cues respectively. Like [16] processes text and visual prompts independently. This results in the selection of sub-optimal queries that do not promote optimization between text and images. + +Beyond these architectural limitations, a broader challenge persists across the field: the reliance on caption-derived vocabularies or fixed-vocabulary, which inherently constrain generalization to novel or composite objects. [16, 22, 60] derive their object vocabularies from entity words extracted from image captions or use fixed-vocabulary detector (like TAP [32]). While this approach avoids manually predefined categories, it remains bounded by the entities explicitly mentioned in the training captions. For instance, TRex-2 [16] employ TAP [32] to annotate SA-1B [19] with a category name from a dictionary of 2560 classes. Nevertheless, the reality that SA-1B [19] is associated with a vocabulary significantly exceeding 2560 substantially constrains the dataset's potential. + +In this work, we propose Prompt-DINO, a novel text-guided visual Prompt DINO, that addresses these challenges through two architectural innovations and a paradigm-shifting data engine. We pioneer an early fusion + +mechanism that deeply integrates text, visual, and backbone features at the initial encoding stage. Unlike [16, 22, 60], our model unifies multimodal signals into a cohesive representation from the outset. This enables richer cross-modal interactions, reducing ambiguity in open-set scenarios. And we introduce order-aligned query selection, a novel paradigm for DETR-based architectures that explicitly optimizes the alignment between text queries and visual queries during decoding. By enhancing semantic correlation and spatial consistency within query priorities, our method facilitates the optimization of text and image query selection in a unified direction, thereby surpassing the heuristic cross-attention mechanisms employed in prior research. Also, we overcome the fixed vocabulary bottleneck through a generative data engine with dual-path cross-verification pipeline. We proposal an novel Recognize Anything via Prompt (RAP) model, a 0.5B-parameter multimodal large language model(MLLM) that generates semantic labels for input images and their corresponding masks. Our engine leverages RAP to synthesize 0.5B diverse training instances with enhanced semantic relevance. To ensure label fidelity, we further implement a novel dual-path cross-verification pipeline that substantially reduces label noise by $80.5\%$ compared to conventional single-path approaches. This integrated data engine pioneers scalable open-world dataset generation, bridging limited annotation vocabularies with real-world detection needs. + +To summarize, our contributions are threefold: + +- We propose an early fusion mechanism that integrates text, visual, and backbone features during initial encoding stage, unifying multimodal signals into a cohesive representation. +- We introduce order-aligned query selection, a new paradigm for DETR-based architectures that explicitly optimizes the alignment between text queries and visual queries during decoding. +- We develop a generative data engine powered by our RAP, incorporating a dual-path cross-verification pipeline. This approach synthesizes diverse and semantically relevant training data, and establishes a new paradigm for scalable open-world data generation. + +# 2. Related Work + +# 2.1. General Recognition via Prompts + +Recent advances in general recognition models (e.g., detection/segmentation [10, 11, 18, 24, 27, 30, 49, 53, 56]) have demonstrated that multimodal alignment between vision and language modalities serves as the cornerstone for open-world understanding. These frameworks typically employ joint text-visual prompting mechanisms to achieve object localization through aligned cross-modal feature spaces. + +Text Prompt Paradigm. Existing methods predominantly focus on text-based cross-modal alignment. GLIP + +![](images/919e3f7873156ad81dc5661fa82214035904da1c6d36d4c1bf6d2b1d3bc2f8d2.jpg) +Figure 2. The Overall Framework of Prompt-DINO. Prompt-DINO is constructed based on the DINO architecture. Images are processed by ViT to extract multi-scale features, while visual and text prompts are processed by deformable attention transformer layers and CLIP's text encoder, respectively, to derive embeddings. Through early fusion layers in the backbone encoder and perception-aware query selection, Prompt-DINO perform deep fusion and alignment of visual and text in two steps, with the final predictions generated by the Decoder. + +[24] pioneers the adaptation of image-text pretraining to downstream detection tasks through deep feature fusion, establishing language-aware visual representations. GroundingDINO [27] further extends this paradigm by demonstrating that multi-stage cross-attention fusion significantly enhances performance in open-vocabulary detection. While these approaches achieve remarkable progress, their heavy reliance on textual semantics limits their capacity to handle visual concepts with ambiguous linguistic descriptions. + +Visual Prompt Paradigm. Complementary to text prompts, visual prompting (points/boxes/masks) provides geometric guidance for specific object localization. SAM [19] establishes a milestone in interactive segmentation through decoupled visual prompt encoding, though lacking semantic awareness. Subsequent works like SEEM [60] and DINOv [22] enhance this paradigm by integrating semantic context and exemplar-based reasoning. A notable hybrid approach is T-Rex2 [16], which combines both modalities but still treats visual prompts as auxiliary geometric cues. Our key insight lies in that existing methods inadequately exploit the semantic synergy between visual prompts and language guidance, leaving visual prompting's full potential underdeveloped for language-aware feature learning. + +# 2.2. Data Engine for Open-World Learning + +Scalable data curation has become pivotal for training robust open-world detector/segmentation models. SAM2 [36] + +introduces a cyclic model-annotator collaboration framework, iteratively refining the SA-V dataset through human-machine interactions. Alternative strategies employ cross-model verification: YOLO-World [6] curates CC3M-Lite by relabeling CC3M [39] with GLIP and CLIP, while T-Rex2 [16] propagates pseudo-labels across SA-1B and LAION400M [34]. Current methods face three key limitations: (1) Over-reliance on noun phrase extraction from captions misses more detailed concepts and introduces too many redundant nouns; (2) Single-path verification propagates annotation errors; (3) Semantic ambiguity in freeform text introduces label noise. Our data engine addresses these issues and enables the automatic generation of high-quality training data with reduced noise. + +# 3. Method + +Motivated by the limitations observed in existing hybrid prompting methods for detection and segmentation tasks, we introduce Prompt-DINO, an end-to-end generic segmentation model built upon the DINO architecture [52]. Our model natively integrates multimodal prompting capabilities for both textual and visual inputs while maintaining architectural simplicity. In this section, we first formalize the promptable segmentation paradigm and outline the core design prompt encoders of Prompt-DINO (§3.1). We then systematically address three critical challenges: (1) early fusion with visual prompts through cross-modal feature align- + +ment(§3.2), (2) order-aware query selection for maintaining semantic consistency(§3.3) and (3) construction of a scalable data engine for open-world generalization(§3.4). Finally, we introduce the training objective(§3.5). + +# 3.1. Promptable Segmentation Task + +The Prompt-DINO processes an input image $\mathcal{I}$ , with accompanying visual prompts $\mathcal{P}^v = \{\mathcal{P}_1^v, \dots, \mathcal{P}_K^v\}$ , and textual prompts $\mathcal{P}^t = \{\mathcal{P}_1^t, \dots, \mathcal{P}_K^t\}$ , where $K$ is the number of interactive prompts per image. Following established architectures in DETR [57] and DINO [52], our system first extracts hierarchical visual features $F = \{f_1, \dots, f_L\}$ through a vision backbone network (e.g., ViT [9] or Swin Transformer [28]), where $L$ represents the number of feature pyramid levels. + +Visual Prompt Encoder. Specifically, our visual encoding module employs $L$ transformer blocks with deformable attention mechanisms [22]. For each target object, we initialize a learnable query embedding $q_{i}$ and progressively refine it through multi-scale feature aggregation. The transformation at each layer $l$ can be formulated as: + +$$ +v _ {i} ^ {l} = M S D A \left(v _ {i} ^ {l - 1}, f _ {l}\right), l \in [ 1, \dots , L ], \quad \text {w i t h} \quad v _ {i} ^ {0} = q _ {i} \tag {1} +$$ + +where $MSDA(\cdot)$ denotes multi-scale deformable attention operation. The final visual prompt representations $V = \{v_{1},\dots,v_{K}\}$ are obtained as $v_{i} = v_{i}^{L}$ . + +Text Prompt Encoder. Building upon insights from recent open-world detection studies, we employ a CLIP [34] text encoder rather than conventional BERT [8] approaches used in earlier models like [24, 27]. This design choice enables better romantic correlation modeling for input words, phrases, and short sentences, generating text prompts $T = \{t_1,\dots,t_K\}$ that effectively guide cross-modal feature alignment. + +The integrated representation (text(T), visual(V), multiscale features(F)) feeds into the backbone encoder for multimodal fusion, enabling dynamic adaptation to prompts while preserving cross-modal alignment. + +# 3.2. Early Fusion with Visual Prompts + +We are the first to propose the early fusion framework that integrates text prompts, visual prompts, and input image features through multi-modal interaction. While this fusion mechanism effectively reduces the modality gap between textual and visual features, enhances prompt-image correlation, and improves target recall, We found unintended consequences of severe hallucinations in our experiments. This phenomenon stems from the inherent limitation of conventional cross-attention mechanisms: they compulsively reconstruct Key features most similar to the Query features, even when all Key features are substantially dissimilar to the current Query. To address this non-target feature reconstruction issue, we introduce a novel gated cross + +![](images/ae748a4d13be1c914e0c223da049e14d11c49083d6f7c9c661587aa6ed6cb778.jpg) +Figure 3. Structure of Early Fusion Layer. Notably, the contrastive alignment of visual and text prompts is completed before early fusion. By employing Gated Cross-Attention, text prompts are reconstructed with image features, while visual prompts and image features mutually reconstruct each other. + +attention mechanism incorporating a learnable background token $(B)$ . Reconstruction occurs exclusively when the Query-Key similarity exceeds an adaptive threshold; otherwise, the background token automatically activates to suppress hallucination. As illustrated in Fig. 3, our architecture comprises three core components: 1) a self-attention layer for intra-modal refinement, 2) the proposed gated cross-attention layer for inter-modal interaction, and 3) a feedforward neural network (FFN) for feature transformation. The cross-attention module implements three interaction pathways: feature-to-visual prompt, visual prompt-to-feature, and feature-to-text prompt. For visualization clarity, Fig. 3 presents only stage-2 features mapped to text space through our two-phase training scheme (detailed in §3.5). The enhanced cross-modal fusion operates through the following formulation: + +$$ +t _ {i} ^ {\prime} = \operatorname {G a t e d A t t n} (f _ {i}, t _ {i}, t _ {i}) +$$ + +$$ +v _ {i} ^ {\prime} = \operatorname {G a t e d A t t n} \left(f _ {i}, v _ {i}, v _ {i}\right) \tag {2} +$$ + +$$ +f _ {i} ^ {\prime} = \operatorname {G a t e d A t t n} (v _ {i}, f _ {i}, f _ {i}) +$$ + +where the gated attention mechanism is defined as: + +$$ +\operatorname {G a t e d A t t n} (Q, K, V) = \operatorname {s o f t m a x} (\frac {Q \cdot [ K , B ]}{\sqrt {d _ {k}}}) [ V, B ] \quad (3) +$$ + +Here, $B$ denotes our learnable background token. Following TRex-2 [16] methodology, we employ region-level contrastive learning to align visual and textual modalities, utilizing symmetric cross-entropy loss [34] for embedding alignment: + +$$ +\mathcal {L} _ {\text {a l i g n}} = - \frac {1}{K} \sum_ {i = 1} ^ {K} \log \frac {v _ {i} \cdot t _ {i}}{\sum_ {j = 1} ^ {K} v _ {i} \cdot t _ {j}} \tag {4} +$$ + +In our experience, symmetric cross-entropy losses have better convergence than InfoNCE loss [31]. + +In the context of training detection and segmentation tasks, it is common practice to construct negative samples. For negative prompt construction, while text prompts naturally leverage category vocabularies for negative sampling, visual prompts require explicit negative construction. We address this by aggregating visual prompt features across all categories within each batch and computing inter-category averages to generate negative visual prompts. This approach ensures sufficient diversity in negative samples while maintaining semantic relevance. + +To mitigate annotation granularity discrepancies across datasets (e.g., "bird" in Dataset A vs. "pigeon" in Dataset B), we implement a dataset-aware sampling strategy. Our novel data sampler ensures each training batch contains only intra-dataset samples, preventing semantically similar but categorically distinct concepts from becoming conflicting negative samples. This is particularly critical in visual-prompt open-world detection/segmentation scenarios where label hierarchies may differ substantially across source datasets. + +# 3.3. Order-aware query selection + +Prompt-DINO adopts a DETR-based architecture that selects top-K encoder features through similarity computation between encoder features and prompt embeddings, following DINO [52] and MaskDINO [21]. However, our approach fundamentally diverges from existing methods by simultaneously considering both textual-visual query similarity and inter-query correlations with visual prompts. To further strengthen the alignment between visual and textual prompts, we introduce order-aware alignment during query selection, which establishes the second-level alignment between visual and textual prompts during training. Specifically, we employ Kendall's $\tau$ coefficient to quantify the ordinal consistency [12] between queries selected by visual prompts and text prompts, the Kendall's $\tau$ coefficient can be expressed as: $P_{\mathrm{c}} - P_{\mathrm{d}}$ + +$$ +\tau = \frac {P _ {c} - P _ {d}}{\frac {1}{2} N (N - 1)} \tag {5} +$$ + +where $P_{c}$ represents concordant pairs and $P_{d}$ represents discordant pairs, $N$ represents the total number of queries. Given the similarity score between query embeddings and prompt embeddings $S^{v} = s_{1}^{v},\ldots ,s_{N}^{v}$ and $S^t = s_1^t,\dots,s_N^t$ we calculate $P_{c}$ and $P_{d}$ as follows: + +$$ +\tau = \frac {\sum_ {i} \sum_ {j < i} \operatorname {s g n} \left(s _ {i} ^ {t} - s _ {j} ^ {t}\right) \cdot \operatorname {s g n} \left(s _ {i} ^ {v} - s _ {j} ^ {v}\right)}{\frac {1}{2} N (N - 1)} \tag {6} +$$ + +Since the sgn is not differentiable, we approximate the sign function with the tanh function and utilize the opposite of $\tau$ as the order alignment loss: + +$$ +\mathcal {L} _ {\text {o r d e r}} = - \frac {\sum_ {i} \sum_ {j < i} \tanh \left(s _ {i} ^ {t} - s _ {j} ^ {t}\right) \cdot \tanh \left(s _ {i} ^ {v} - s _ {j} ^ {v}\right)}{\frac {1}{2} N (N - 1)} \tag {7} +$$ + +Our proposed alignment mechanism operates at two complementary levels: The primary alignment directly optimizes the learning objective through $\mathcal{L}_{\text{align}}$ (as given in + +Eq. 4), while our novel order-aware alignment provides indirect assurance through two critical aspects: (1) Semantic proximity between corresponding text and visual prompts for candidate queries naturally reduces overall training complexity by minimizing feature space discrepancies; (2) Training-inference consistency preservation ensures equivalent substitutability between text and visual prompts during inference through maintained representational parity. + +Fig. 5 demonstrates the effectiveness of our proposed method. Prior to applying the order-aware loss optimization, we observe significant disorder in both textual and visual prompt queries, with no discernible alignment between corresponding semantic representations. The optimization process induces a remarkable transformation - post-optimization analysis reveals substantially enhanced consistency between semantically equivalent textual and visual prompt queries. + +# 3.4. Data Engine + +Effective visual concept learning from rich image-text pairs is fundamental to enhancing the generalization capabilities of open-vocabulary object detectors. Existing methods predominantly rely on captioning models([3, 23, 26] etc.), coupled with lexical analysis tools like SpaCy [14] to extract noun phrases. These noun phrases are subsequently employed to generate pseudo-labels via open-set detection models ([24, 27]). Nonetheless, these techniques have significant drawbacks. They frequently omit valid labels, introduce erroneous annotations, and are susceptible to irrelevant lexical noise. All of these issues detrimentally affect detection performance. + +To address these challenges, we propose a novel Recognize Anything via Prompting (RAP) model, a 0.5B MLLM, which significantly enhances annotation accuracy. Our framework further integrates an innovative dual-path cross-validation pipeline that substantially reduces label noise. This comprehensive approach enables the fully automated generation of high-quality open-world detection datasets while preserving the fidelity of annotations. + +Recognize Anything via Prompting. Despite the remarkable capabilities of the Segment Anything Model (SAM) [19] in detecting a wide array of objects within images, and the substantial progress achieved in region-level understanding through multimodal models [1, 2, 33, 35, 51], a critical question persists: Can we effectively assign semantic labels to arbitrary mask prompts? To address this challenge, we introduce RAP, a novel framework that bridges visual segmentation with semantic recognition. As illustrated in Fig. 4, our RAP model processes raw image inputs alongside their corresponding object masks to generate precise semantic annotations for the specified regions. The proposed RAP comprises three key components: + +1. Feature extraction. We perform masked pooling over + +![](images/4b51941a84aacf5fc7d55a29f37b0b50e268f690b5918da5e650c453af95325f.jpg) +Figure 4. Data Engine and RAP Frameworks. Left: The data engine with dual-path cross-verification pipeline. Right: The proposed Recognize Anything via Prompting (RAP), a 0.5B-parameter MLLM model. + +![](images/525074c2087d8413737cf9a47c9ee6667778f38f68f406135b0da0974b7c1eea.jpg) + +multi-layer image features to extract region-specific features corresponding to input masks, and concatenate the mask token embeddings obtained via a projection layer with the original image tokens to form a unified representation. This preserves both global contextual information from the complete image and localized object-specific features within masked regions, effectively integrating scene-level understanding with instance-level visual cues. + +2. Ground Data Cleaning. To standardize the format of our ground truth data, we leveraged the Qwen model [48] to eliminate redundant descriptive phrases through comprehensive text normalization. Specifically, we transformed verbose expressions into their canonical semantic equivalents (e.g., converting "any part of the sky" to "sky"). This preprocessing strategy serves dual purposes: (1) it reduces the RAP's learning complexity by minimizing lexical variability, and (2) it mitigates the risk of hallucinations by ensuring precise semantic alignment between textual annotations and visual content. + +3. Training. The proposed methodology employs a two-stage training strategy. During the first stage, we conduct pre-training using 10M samples, specifically optimizing only the projector while keeping other model components frozen. The second stage implements instruction-tuning, where we unfreeze and jointly optimize the entire model to enhance task-specific adaptation capabilities. + +Dual-path cross-verification pipeline. We propose a dual-pipeline annotation framework combining bottom-up and top-down approaches with cross-verification filtering, consisting of three key components: + +1. Top-Down Pipeline. Our top-down pipeline leverages the Qwen-VL model [42] to extract semantic tags from images. These tags are subsequently fed into an open-set detection model to generate corresponding bounding boxes. This pipeline demonstrates strong capability in detecting large-scale salient objects (e.g., trees, roads, sky). +2. Bottom-Up Pipeline. To address the inherent limitation of top-down pipeline in capturing fine-grained details, we implement a complementary bottom-up strategy using the RAP integrated with SAM [19]. This combination enables effective detection of smaller objects and detailed + +components (e.g., eyes, noses, windows) that are frequently missed by conventional top-down pipeline. + +3. Cross-Verification. The detection outputs from both pipelines undergo rigorous validation through a three-stage process. First, We employ the Hungarian matching algorithm to establish correspondences between bounding boxes from different pipelines. Next, we utilize the BGE model [47] to compute similarity scores between the tags associated with the matched bounding boxes. Finally, we filter out the candidate pairs whose similarity scores were below the threshold of empirical determination to further clean up the annotated results. + +# 3.5. Training Objective + +Building upon previous DINO-like models [52], we employ a combination of L1 loss and GIoU loss [37] for bounding box regression, while for mask segmentation, we adopt cross-entropy loss and dice loss [29]. Following the methodology established in GroudingDINO [27], we utilize contrastive loss between predicted objects and text prompts for classification tasks. To accelerate model convergence, we incorporate denoising training as implemented in DINO [52]. The complete objective function for Prompt DINO can be expressed as: + +$$ +\mathcal {L} = \mathcal {L} _ {c l s} + \mathcal {L} _ {b b o x} + \mathcal {L} _ {m a s k} + \mathcal {L} _ {D N} + \mathcal {L} _ {a l i g n} + \mathcal {L} _ {o r d e r} \tag {8} +$$ + +Our training methodology adopts a progressive multi-stage strategy: we first optimize the text prompt while keeping the visual components fixed, followed by joint optimization of both textual and visual prompt representations. + +# 4. Experiments + +# 4.1. Settings + +In Prompt-DINO, we employ ViT-L [9] as the backbone and fine-tune CLIP-L [34] as our text prompt encoder. The visual prompt encoder consists of 6 layers of deformable cross-attention transformer layers, with a hidden dimension of 256. We adopt the "two-stage" structure proposed by DINO [52], selecting 900 queries for decoding after the backbone encoder. + +
MethodBackbonePrompt TypeCOCO (in-domain)LVIS (in-domain)ADE20K (out-domain)
mask APbox APmask APbox APPQmask APbox AP
Mask2Former [4]Swin-LClose-Set48.652.1-----
OneFormer [15]Swin-LClose-Set48.9------
MaskDINO* [21]Swin-LClose-Set50.656.2-----
kMaX-DeepLab [50]Swin-LClose-Set58.1------
GLIPv2 [53]Swin-HText Prompt48.9------
X-Decoder [59]DaViT-LText Prompt46.7---21.813.138.1
OpenSeeD [54]Swin-LText Prompt53.258.221.023.019.715.017.7
T-Rex2 [16]Swin-LText Prompt-52.2*-45.8*---
APE [40]ViT-LText Prompt49.358.353.059.627.224.429.6
DINOv [22]Swin-LVisual Prompt50.454.2--23.215.114.3
T-Rex2 [16]Swin-LVisual Prompt-46.5*-45.3*---
Prompt-DINOViT-LText Prompt51.158.656.560.129.624.628.6
Prompt-DINOViT-LVisual Prompt51.058.055.659.935.928.032.7
+ +Table 1. Generic Segmentation results of Prompt-DINO on multiple datasets. "-" denotes the model does not have performance reported or does not have ability for the specific task. "+" denotes the performance reported under different domain settings. MaskDINO uses the results reported in DINOv [22] + +Datasets. During the training process, we utilized three categories of datasets: open-source detection datasets. These include COCO [25] with 118K images and LVIS [13] with 160K images, both of which have mask annotations. For datasets without mask annotations, such as Object365 [38], OID [20], and V3Det [41], we first apply SAM to generate mask annotations. Due to the presence of low-quality data and long-tail distribution in Object365 and OID, which hinder general object recognition, we resampled them, selecting 700K and 1M data points, respectively. grounding datasets. This includes the GoldG [17] dataset with 100K images. self-built datasets. We re-announce and filter CC3M [39] and SA-1B [19] using the RAP and dual-path cross-validation data engine introduced in §3.4, yielding high-quality datasets of 3M and 7M images, respectively, which constitute our primary training data. For testing, we utilize COCO and LVIS for in-domain evaluation and ADE20K [55], SeginW [59], and Cityscapes [7] for out-domain evaluation. + +Metrics. We employ Average Precision (AP) as the evaluation metric, utilizing mask AP for instance segmentation, and box AP for object detection. For ADE20K and CtiyScapes, we employ Panoptic Quality(PQ) and mean intersection over union (mIoU) for panoptic segmentation as well. For SegInW, we adopt AP-Average and AP-Median as the evaluation metric. + +# 4.2. Generic Segmentation + +To evaluate the effectiveness and generalization capability of our Prompt-DINO, we conduct evaluations on both indomain and out-domain datasets, with the results presented in Tab. 1. + +In-Domain Segmentation. We conduct in-domain experiments on COCO and LVIS. Compared to methods based on visual prompts, our Prompt-DINO achieved significantly better results. Compared to the previous work DINOv [22], we improve mask AP by 0.6 and box AP by 3.8 + +on COCO. Compared to methods based on text prompts, Prompt-DINO also demonstrates superior results when using text prompts. For example, compared to APE [40], we lead by 3.5 in mask AP and 0.5 in box AP on LVIS. + +Out-Domain Segmentation. We use ADE-20K and Cityscapes as an out-domain dataset and conduct zero-shot testing on it, demonstrating the strong generalization ability of our Prompt-DINO. Under the out-domain setting (where data marked with * indicates performance reported under different domain settings), our method achieves the best performance on ADE20K across all methods for both prompt types, whether using visual prompts or text prompts. Compared to the state-of-the-art method APE on ADE20K, we improve PQ by 8.7, mask AP by 3.6, and box AP by 18.4, representing increases of $32.0\%$ , $14.8\%$ , and $10.5\%$ , respectively. Among methods based on visual prompts, compared to the previous method DINOv, we improve PQ by 12.7, mask AP by 12.9, and box AP by 18.4, representing increases of $54.7\%$ , $85.4\%$ , and $128.7\%$ , respectively. Compared to APE on Cityscapes, we also improve PQ by 1.0 and mIoU by 8.0. + +Segmentation in the Wild. To further evaluate the generalization capabilities of our model, we perform zero-shot testing on the SeginW benchmark, which includes 25 datasets. Compared to the previous work DINOv, when using visual prompts, we improve the average AP by 16.2 and the median AP by 14.5. Among text-prompt-based methods, we also surpass APE by 6.8 in average AP and 6.9 in median AP. This further validates the strong generalization ability of our Prompt-DINO. + +# 4.3. Effectiveness Analysis + +To validate the effectiveness of our proposed components, we conduct ablation studies on Text Prompt Alignment, Early Fusion, Order Alignment, and Data Engine. As our baseline, we implement the naive MaskDINO method with the FC head replaced with a user-prompted visual embed + +
MethodTypeCityScapeSegInW
PQmIoUAP-avgAP-med
kMaX-DeepLab-L [50]in-domain68.483.5--
X-Decoder-L [59]out-domain38.152.022.332.3
OpenSeeD-L [54]out-domain41.447.836.138.7
APE-L [40]out-domain33.344.249.652.2
DINOv [22]out-domain--40.644.6
Prompt-DINO [40]out-domain34.352.256.459.1
+ +Table 2. Zero-shot Segmentation results of Prompt-DINO on the CityScape [7] and SegInW [59] dataset. + +
BaselineMethodADE20K (out-domain)
Text AlignEarly FusionOrder AlignPQmAPbAP
✓×××25.020.423.5
✓✓××32.725.530.5
✓✓✓×34.726.731.7
✓✓✓✓35.928.032.7
+ +ding. The ablation results are presented in Tab. 3. + +Effectiveness of Early Fusion. By comparing rows 1 and 2 in Tab. 3, it can be observed that incorporating text align in the backbone encoder (stage one of DINO) increases the mask AP and box AP on ADE20K by 5.1 and 7.0, respectively, and PQ by 7.7, compared to the baseline. This indicates that naive alignment of text and visual can enhance the model's general recognition capability. Comparing rows 2 and 3 in Tab. 3 shows that early fusion of visual prompts, guided by text alignment, significantly improves the model's performance, with mask AP and box AP on ADE20K increasing by 1.2 and 1.2, respectively, and PQ by 2.0. This demonstrates that visual prompts effectively complement features that are difficult to describe with text prompts and, under the guidance of text prompts, achieve more generalizable expressive power. + +![](images/f54bb0f07caa5d1ec7f50ff2be7491c0a9c7f05cfd7227fc660e53e168957e68.jpg) +Figure 5. Order Alignment Comparison. Left: Visual and text should demonstrate strong correlation, but the queries selected by the two prompts exhibit significant disorder. Right: Order alignment significantly improves the consistency of visual-text query selection and reinforces the alignment between corresponding semantics. + +![](images/b80a80b6a055e66a2f9122a8603a864e4e08f3ca850e750c8f8edac2873955b4.jpg) + +Effectiveness of Text-guided Visual Prompts. Rows 1, 2, and 3 in Tab. 3 show that merely adding text alignment provides limited improvement to the model. However, after early fusion to eliminate semantic ambiguity, the two types of prompts can serve as complementary information, providing stronger recognition capabilities. Next, we introduced order alignment during the query selection stage to further guide visual prompts with text. Rows 3 and 4 in Tab. 3 demonstrate that our two-stage guidance deeply + +![](images/01ab29b901c447db640264a901dde471c25a001ec72a6dc947e097922d6e528a.jpg) +Figure 6. Annotation quality and quantity in the data engine. + +![](images/244c7a54e5839ef223edd1855bfdd42f7d4c3832ace980e9a3522012903f9ddb.jpg) + +aligns visual and text prompts, leading to further performance improvements. Fig. 5 also demonstrates that our order alignment method further enhances the alignment between visual and text, resulting in a more consistent selection order for queries. + +Effectiveness of Data Engine. Within the data engine, we employ a cross-validation module to filter data annotated by the top-down and bottom-up pipelines. As illustrated in Fig. 6, the cross-validation module filters approximately $80.5\%$ of instance annotations and enhances the annotation clip similarity from 0.16 to 0.23, highlighting the effectiveness of the cross-validation module in generating high-quality data. As shown in Tab. 4, incorporating data generated by our data engine into the training set improves Prompt-DINO's box AP by $12\%$ on LVIS and $5.9\%$ on ADE20K. To further demonstrate the effectiveness of RAP, we process a subset of SA-1B using different data engines in Tab. 5, revealing that RAP enhances model performance by $1.6\%$ mask AP and $2.6\%$ box AP on ADE20K. This indicates that the proposed data engine effectively enhances data quality, while RAP generates more diverse data, significantly improving the model's generalization capability. + +Table 3. Ablation of proposed components. + +
MethodLVIS (in-domain)ADE20K (out-domain)
mask APbox APPQmask APbox AP
w/o DataEngine46.047.930.823.426.8
w Data Engine55.659.935.928.032.7
+ +Table 4. Ablation of Data Engine (Visual Prompt Results). + +
DatasetCOCO (in-domain)ADE20K (out-domain)
mask APbox APPQmask APbox AP
COCO+GoldG49.855.718.211.011.6
COCO+SA-1B (QwenVL)50.156.722.317.820.4
COCO+SA-1B (QwenVL + RAP)50.858.122.519.423.0
+ +Table 5. Ablation of Training Dataset (Text Prompt Results). For efficient training, we use a 2M subset of SA-1B. + +# 5. Conclusion + +We present Prompt-DINO, a text-guided visual Prompt DINO for generic segmentation via three innovations: 1) Early fusion of text, visual, and backbone features during initial encoding to resolve cross-modal ambiguity; 2) Order-aligned query selection for semantic-spatial consistency between text and visual cues in DETR decoding; 3) A generative data engine (RAP + dual-path verification) synthesizing 0.5B high-fidelity training instances, reducing label noise by $80.5\%$ . Prompt-DINO sets new state-of-the-art results while enabling scalable adaptation to real-world object diversity. + +# Acknowledgements + +This work was partially supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant(KJZD20230923115106012, KJZD20230923 114916032, GJHZ20240218113604008), National Natural Science Foundation of China (No. 62106149). + +# References + +[1] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 5 +[2] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 5 +[3] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 5 +[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 7 +[5] Runxi Cheng, Feng Xiong, Yongxian Wei, Wanyun Zhu, and Chun Yuan. Whoever started the interference should end it: Guiding data-free model merging via task vectors. arXiv preprint arXiv:2503.08099, 2025. 2 +[6] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 3 +[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 7, 8 +[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 4 +[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 6 +[10] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level + +labels. In European conference on computer vision, pages 540-557. Springer, 2022. 2 +[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2 +[12] Yuchen Guan, Runxi Cheng, Kang Liu, and Chun Yuan. Enhancing logits distillation with plug&play kendall's $\tau$ ranking loss, 2025. 5 +[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 7 +[14] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, Adriane Boyd, et al. spacy: Industrial-strength natural language processing in python. 2020. 5 +[15] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2989-2998, 2023. 7 +[16] Qing Jiang, Feng Li, Zhaoyang Zeng, Tianhe Ren, Shilong Liu, and Lei Zhang. T-rex2: Towards generic object detection via text-visual prompt synergy. In European Conference on Computer Vision, pages 38-57. Springer, 2024. 2, 3, 4, 7 +[17] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 7 +[18] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 2 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2, 3, 5, 6, 7 +[20] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 7 +[21] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3041-3050, 2023. 2, 5, 7 +[22] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Jianwei Yang, Chunyuan Li, et al. Visual in-context prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision + +and Pattern Recognition, pages 12861-12871, 2024. 2, 3, 4, 7, 8 +[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 5 +[24] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 2, 3, 4, 5 +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 7 +[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 5 +[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 2, 3, 4, 5, 6 +[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4 +[29] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6 +[30] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European conference on computer vision, pages 728-755. Springer, 2022. 2 +[31] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4 +[32] Ting Pan, Lulu Tang, Xinlong Wang, and Shiguang Shan. Tokenize anything via prompting. In European Conference on Computer Vision, pages 330-348. Springer, 2024. 2 +[33] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 5 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning + +transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3, 4, 6 +[35] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 5 +[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3 +[37] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6 +[38] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 7 +[39] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 3, 7 +[40] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13193–13203, 2024. 7, 8 +[41] Jiaqi Wang, Pan Zhang, Tao Chu, Yuhang Cao, Yujie Zhou, Tong Wu, Bin Wang, Conghui He, and Dahua Lin. V3det: Vast vocabulary visual detection dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19844-19854, 2023. 7 +[42] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6 +[43] Yuanrui Wang, Cong Han, YafeiLi, Zhipeng Jin, Xiawei Li, SiNan Du, Wen Tao, Yi Yang, shuanglong li, Chun Yuan, and Liu Lin. Uniglyph: Unified segmentation-conditioned diffusion for precise visual text synthesis, 2025. 2 +[44] Hongyang Wei, Shuaizheng Liu, Chun Yuan, and Lei Zhang. Perceive, understand and restore: Real-world image superresolution with autoregressive multimodal generative models, 2025. + +[45] Yongxian Wei, Runxi Cheng, Weike Jin, Enneng Yang, Li Shen, Lu Hou, Sinan Du, Chun Yuan, Xiaochun Cao, and Dacheng Tao. Unifying multimodal large language model capabilities and modalities via model merging. arXiv preprint arXiv:2505.19892, 2025. +[46] Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Chun Yuan, and Dacheng Tao. Open-vocabulary customization from clip via data-free knowledge distillation. In The Thirteenth International Conference on Learning Representations, 2025. 2 +[47] Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muenighoff. C-pack: Packaged resources to advance general chinese embedding, 2023. 6 +[48] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 6 +[49] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 2 +[50] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In European Conference on Computer Vision, pages 288-307. Springer, 2022. 7, 8 +[51] Yuqian Yuan, Wentong Li, Jian Liu, Dongqi Tang, Xinjie Luo, Chi Qin, Lei Zhang, and Jianke Zhu. Osprey: Pixel understanding with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28202-28211, 2024. 5 +[52] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 4, 5, 6 +[53] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems, 35:36067-36080, 2022. 2, 7 +[54] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 7, 8 + +[55] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7 +[56] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European conference on computer vision, pages 350-368. Springer, 2022. 2 +[57] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4 +[58] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting, 2023. 2 +[59] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15116-15127, 2023. 7, 8 +[60] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in neural information processing systems, 36:19769-19782, 2023. 2, 3 \ No newline at end of file diff --git a/2025/Text-guided Visual Prompt DINO for Generic Segmentation/images.zip b/2025/Text-guided Visual Prompt DINO for Generic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4c52ccf29e6b0ff150a6af298b6017c29fcb6dd0 --- /dev/null +++ b/2025/Text-guided Visual Prompt DINO for Generic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81735341b378badf79ab4e129de117484e0732554dee7eae0607b7855db70920 +size 651713 diff --git a/2025/Text-guided Visual Prompt DINO for Generic Segmentation/layout.json b/2025/Text-guided Visual Prompt DINO for Generic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e83a3180403c0747d808ec13c3732a18da60aa61 --- /dev/null +++ b/2025/Text-guided Visual Prompt DINO for Generic Segmentation/layout.json @@ -0,0 +1,8882 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 491, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 491, + 121 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 491, + 121 + ], + "type": "text", + "content": "Text-guided Visual Prompt DINO for Generic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "spans": [ + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "text", + "content": "Yuchen Guan" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{1*‡}" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "text", + "content": ", Chong Sun" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{2*†}" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "text", + "content": ", Canmiao Fu" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "text", + "content": ", Zhipeng Huang" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "text", + "content": ", Chun Yuan" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{1†}" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "text", + "content": ", Chen Li" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "text", + "content": "Tsinghua Shenzhen International Graduate School, Tsinghua University \n" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 142, + 518, + 185 + ], + "type": "text", + "content": "WeChat AI, Tencent Inc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 196, + 165, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 196, + 165, + 203 + ], + "spans": [ + { + "bbox": [ + 63, + 196, + 165, + 203 + ], + "type": "text", + "content": "Zero-Shot Segmentation via Visual Prompt" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 202, + 184, + 281 + ], + "blocks": [ + { + "bbox": [ + 61, + 202, + 184, + 281 + ], + "lines": [ + { + "bbox": [ + 61, + 202, + 184, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 202, + 184, + 281 + ], + "type": "image", + "image_path": "e45deb8b198c74b7630209764108ebb61d60c61a538f52efb612fb59470c6d18.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 184, + 202, + 301, + 281 + ], + "blocks": [ + { + "bbox": [ + 184, + 202, + 301, + 281 + ], + "lines": [ + { + "bbox": [ + 184, + 202, + 301, + 281 + ], + "spans": [ + { + "bbox": [ + 184, + 202, + 301, + 281 + ], + "type": "image", + "image_path": "eaebf852eba30e883eff7ec2a75fc9b18025760ac819f6d6d74b26e735242734.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 118, + 284, + 246, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 284, + 246, + 294 + ], + "spans": [ + { + "bbox": [ + 118, + 284, + 246, + 294 + ], + "type": "text", + "content": "Visual Cross-Image Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 202, + 428, + 280 + ], + "blocks": [ + { + "bbox": [ + 313, + 202, + 428, + 280 + ], + "lines": [ + { + "bbox": [ + 313, + 202, + 428, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 428, + 280 + ], + "type": "image", + "image_path": "344751fe6bbf4ec16e7cdd984b13d713b4dd4fd5c0d63fdbf6bf0d19219ed9ae.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 436, + 202, + 548, + 280 + ], + "blocks": [ + { + "bbox": [ + 436, + 202, + 548, + 280 + ], + "lines": [ + { + "bbox": [ + 436, + 202, + 548, + 280 + ], + "spans": [ + { + "bbox": [ + 436, + 202, + 548, + 280 + ], + "type": "image", + "image_path": "3b56b0d126228125b9af241c58353b01639eb23a8ff83f6a18c340dc88d47ec6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 374, + 284, + 488, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 284, + 488, + 295 + ], + "spans": [ + { + "bbox": [ + 374, + 284, + 488, + 295 + ], + "type": "text", + "content": "Visual In-Image Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 296, + 160, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 296, + 160, + 303 + ], + "spans": [ + { + "bbox": [ + 63, + 296, + 160, + 303 + ], + "type": "text", + "content": "Zero-Shot Segmentation via Text Prompt" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 62, + 303, + 164, + 380 + ], + "blocks": [ + { + "bbox": [ + 62, + 303, + 164, + 380 + ], + "lines": [ + { + "bbox": [ + 62, + 303, + 164, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 164, + 380 + ], + "type": "image", + "image_path": "4863ca13a43dacc00841f33624d25d6f860cef1b950829c7a95604e336ad325d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 164, + 303, + 265, + 380 + ], + "blocks": [ + { + "bbox": [ + 164, + 303, + 265, + 380 + ], + "lines": [ + { + "bbox": [ + 164, + 303, + 265, + 380 + ], + "spans": [ + { + "bbox": [ + 164, + 303, + 265, + 380 + ], + "type": "image", + "image_path": "bf18a0dee20706dfe0821158eae9379cb1666290a01c69d002e96a6611306641.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 110, + 384, + 216, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 384, + 216, + 395 + ], + "spans": [ + { + "bbox": [ + 110, + 384, + 216, + 395 + ], + "type": "text", + "content": "Text In-Image Segmentation" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 269, + 303, + 410, + 380 + ], + "blocks": [ + { + "bbox": [ + 269, + 303, + 410, + 380 + ], + "lines": [ + { + "bbox": [ + 269, + 303, + 410, + 380 + ], + "spans": [ + { + "bbox": [ + 269, + 303, + 410, + 380 + ], + "type": "image", + "image_path": "53ef319c81eab998c5d0685eeb3e283914d3ad8a93348245339ee6840ad3b30d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 411, + 303, + 548, + 380 + ], + "blocks": [ + { + "bbox": [ + 411, + 303, + 548, + 380 + ], + "lines": [ + { + "bbox": [ + 411, + 303, + 548, + 380 + ], + "spans": [ + { + "bbox": [ + 411, + 303, + 548, + 380 + ], + "type": "image", + "image_path": "e19643d3f04e3b5794d8bf8a5ef2f7e301c14245e7ed5b44a49a1645b98f870b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 355, + 384, + 460, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 384, + 460, + 395 + ], + "spans": [ + { + "bbox": [ + 355, + 384, + 460, + 395 + ], + "type": "text", + "content": "Text Panoptic Segmentation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 63, + 396, + 187, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 396, + 187, + 404 + ], + "spans": [ + { + "bbox": [ + 63, + 396, + 187, + 404 + ], + "type": "text", + "content": "Zero-Shot part Segmentation via Text/Visual Prompt" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 62, + 404, + 176, + 482 + ], + "blocks": [ + { + "bbox": [ + 62, + 404, + 176, + 482 + ], + "lines": [ + { + "bbox": [ + 62, + 404, + 176, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 176, + 482 + ], + "type": "image", + "image_path": "9df8bf5121a3e1cc6b08953b1cef5d3a0ee690e40cba00a72f329948c7dfb30d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 502, + 555, + 537 + ], + "lines": [ + { + "bbox": [ + 55, + 502, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 555, + 537 + ], + "type": "text", + "content": "Figure 1. Detection/Segmentation samples. Our Prompt-DINO is a text-guided visual Prompt DINO. Prompt-DINO supports single or multiple visual or text prompts for in-image or cross-image generic segmentation. Based on the granularity of the provided prompts, Prompt-DINO is capable of performing both panoptic and part segmentation." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 176, + 404, + 290, + 483 + ], + "blocks": [ + { + "bbox": [ + 176, + 404, + 290, + 483 + ], + "lines": [ + { + "bbox": [ + 176, + 404, + 290, + 483 + ], + "spans": [ + { + "bbox": [ + 176, + 404, + 290, + 483 + ], + "type": "image", + "image_path": "80779919295d9afce6d868cafd836ea76fadca1c0670bf94789f9fe64e0aa97b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 290, + 404, + 414, + 483 + ], + "blocks": [ + { + "bbox": [ + 290, + 404, + 414, + 483 + ], + "lines": [ + { + "bbox": [ + 290, + 404, + 414, + 483 + ], + "spans": [ + { + "bbox": [ + 290, + 404, + 414, + 483 + ], + "type": "image", + "image_path": "83f5ab435ed5f02ce9ec3bd1697698c2c228a088f649f3a452b3fb39c2899b1d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 190, + 486, + 286, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 486, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 190, + 486, + 286, + 496 + ], + "type": "text", + "content": "Visual Part Segmentation" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 421, + 404, + 472, + 482 + ], + "blocks": [ + { + "bbox": [ + 421, + 404, + 472, + 482 + ], + "lines": [ + { + "bbox": [ + 421, + 404, + 472, + 482 + ], + "spans": [ + { + "bbox": [ + 421, + 404, + 472, + 482 + ], + "type": "image", + "image_path": "6703fb5ff0fc67b914d10a2def8853e5c821b28d48ca721d3605ac56d0293707.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 472, + 404, + 547, + 482 + ], + "blocks": [ + { + "bbox": [ + 472, + 404, + 547, + 482 + ], + "lines": [ + { + "bbox": [ + 472, + 404, + 547, + 482 + ], + "spans": [ + { + "bbox": [ + 472, + 404, + 547, + 482 + ], + "type": "image", + "image_path": "1920781344e8b6fb191b66b01cf79d33013a837af902051c8f1306b2608e2bb3.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 440, + 487, + 529, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 440, + 487, + 529, + 497 + ], + "spans": [ + { + "bbox": [ + 440, + 487, + 529, + 497 + ], + "type": "text", + "content": "Text Part Segmentation" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 152, + 546, + 200, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 546, + 200, + 558 + ], + "spans": [ + { + "bbox": [ + 152, + 546, + 200, + 558 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 571, + 297, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 297, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 297, + 668 + ], + "type": "text", + "content": "Recent advancements in multimodal vision models have highlighted limitations in late-stage feature fusion and suboptimal query selection for hybrid prompts open-world segmentation, alongside constraints from caption-derived vocabularies. To address these challenges, we propose Prompt-DINO, a text-guided visual Prompt DINO framework featuring three key innovations. First, we introduce an early fusion mechanism that unifies text/visual prompts" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 548, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 704 + ], + "type": "text", + "content": "and backbone features at the initial encoding stage, enabling deeper cross-modal interactions to resolve semantic ambiguities. Second, we design order-aligned query selection for DETR-based architectures, explicitly optimizing the structural alignment between text and visual queries during decoding to enhance semantic-spatial consistency. Third, we develop a generative data engine powered by the Recognize Anything via Prompt (RAP) model, which synthesizes 0.5B diverse training instances through a dualpath cross-verification pipeline, reducing label noise by " + }, + { + "bbox": [ + 313, + 548, + 555, + 704 + ], + "type": "inline_equation", + "content": "80.5\\%" + }, + { + "bbox": [ + 313, + 548, + 555, + 704 + ], + "type": "text", + "content": " compared to conventional approaches. Extensive experiments demonstrate that Prompt-DINO achieves state-of-the-art performance on open-world detection bench" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 673, + 143, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 673, + 143, + 683 + ], + "spans": [ + { + "bbox": [ + 69, + 673, + 143, + 683 + ], + "type": "text", + "content": "* Equal Contribution." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 71, + 683, + 268, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 683, + 268, + 703 + ], + "spans": [ + { + "bbox": [ + 71, + 683, + 268, + 703 + ], + "type": "text", + "content": "† Corresponding Authors. Emails: waynecsun@tencent.com, yuanc@sz.tsinghua.edu.cn" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 71, + 703, + 183, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 703, + 183, + 712 + ], + "spans": [ + { + "bbox": [ + 71, + 703, + 183, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 71, + 703, + 183, + 712 + ], + "type": "text", + "content": " Work done as interns at WeChat." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "21288" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 144 + ], + "type": "text", + "content": "marks while significantly expanding semantic coverage beyond fixed-vocabulary constraints. Our work establishes a new paradigm for scalable multimodal detection and data generation in open-world scenarios. Data&Code are available at https://github.com/WeChatCV/WeVisionOne." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 155, + 136, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 136, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 136, + 167 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 498 + ], + "type": "text", + "content": "Recent advancements in vision models [5, 43-46, 58] have focused on unifying diverse prompting strategies—such as text, visual, or hybrid inputs—to enable versatile object understanding and segmentation. Pioneering works like [16, 21, 22, 60] have demonstrated the potential of multimodal prompting for tasks ranging from open-set detection to segmentation. Despite their emerging capabilities, openworld detection and segmentation methods leveraging multimodal prompts demonstrate inherent technical deficiencies.1) Late-stage fusion of multimodal features. [22, 60] process text, visual or other prompts, and backbone features in separate subspaces, merging them only in later encoding stages. [16] similarly adopts contrastive learning to align modalities without joint optimization at the encoding stage. This delayed interaction restricts cross-modal alignment, particularly in ambiguous scenarios where early fusion of text and visual cues could disambiguate object semantics. 2) Insufficient exploration of query selection mechanisms. Most of these methods employ generic crossattention for text-guided query selection, or handle queries corresponding to text and visual prompts independently, none of which explicitly optimizes the structural alignment between queries corresponding to text and visual cues respectively. Like [16] processes text and visual prompts independently. This results in the selection of sub-optimal queries that do not promote optimization between text and images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": "Beyond these architectural limitations, a broader challenge persists across the field: the reliance on caption-derived vocabularies or fixed-vocabulary, which inherently constrain generalization to novel or composite objects. [16, 22, 60] derive their object vocabularies from entity words extracted from image captions or use fixed-vocabulary detector (like TAP [32]). While this approach avoids manually predefined categories, it remains bounded by the entities explicitly mentioned in the training captions. For instance, TRex-2 [16] employ TAP [32] to annotate SA-1B [19] with a category name from a dictionary of 2560 classes. Nevertheless, the reality that SA-1B [19] is associated with a vocabulary significantly exceeding 2560 substantially constrains the dataset's potential." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "In this work, we propose Prompt-DINO, a novel text-guided visual Prompt DINO, that addresses these challenges through two architectural innovations and a paradigm-shifting data engine. We pioneer an early fusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 395 + ], + "type": "text", + "content": "mechanism that deeply integrates text, visual, and backbone features at the initial encoding stage. Unlike [16, 22, 60], our model unifies multimodal signals into a cohesive representation from the outset. This enables richer cross-modal interactions, reducing ambiguity in open-set scenarios. And we introduce order-aligned query selection, a novel paradigm for DETR-based architectures that explicitly optimizes the alignment between text queries and visual queries during decoding. By enhancing semantic correlation and spatial consistency within query priorities, our method facilitates the optimization of text and image query selection in a unified direction, thereby surpassing the heuristic cross-attention mechanisms employed in prior research. Also, we overcome the fixed vocabulary bottleneck through a generative data engine with dual-path cross-verification pipeline. We proposal an novel Recognize Anything via Prompt (RAP) model, a 0.5B-parameter multimodal large language model(MLLM) that generates semantic labels for input images and their corresponding masks. Our engine leverages RAP to synthesize 0.5B diverse training instances with enhanced semantic relevance. To ensure label fidelity, we further implement a novel dual-path cross-verification pipeline that substantially reduces label noise by " + }, + { + "bbox": [ + 313, + 72, + 555, + 395 + ], + "type": "inline_equation", + "content": "80.5\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 395 + ], + "type": "text", + "content": " compared to conventional single-path approaches. This integrated data engine pioneers scalable open-world dataset generation, bridging limited annotation vocabularies with real-world detection needs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 396, + 515, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 396, + 515, + 407 + ], + "spans": [ + { + "bbox": [ + 326, + 396, + 515, + 407 + ], + "type": "text", + "content": "To summarize, our contributions are threefold:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 408, + 553, + 564 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 408, + 553, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 408, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 314, + 408, + 553, + 454 + ], + "type": "text", + "content": "- We propose an early fusion mechanism that integrates text, visual, and backbone features during initial encoding stage, unifying multimodal signals into a cohesive representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 456, + 553, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 456, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 314, + 456, + 553, + 503 + ], + "type": "text", + "content": "- We introduce order-aligned query selection, a new paradigm for DETR-based architectures that explicitly optimizes the alignment between text queries and visual queries during decoding." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 504, + 553, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 504, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 314, + 504, + 553, + 564 + ], + "type": "text", + "content": "- We develop a generative data engine powered by our RAP, incorporating a dual-path cross-verification pipeline. This approach synthesizes diverse and semantically relevant training data, and establishes a new paradigm for scalable open-world data generation." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 571, + 400, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 571, + 400, + 583 + ], + "spans": [ + { + "bbox": [ + 314, + 571, + 400, + 583 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 587, + 493, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 587, + 493, + 600 + ], + "spans": [ + { + "bbox": [ + 314, + 587, + 493, + 600 + ], + "type": "text", + "content": "2.1. General Recognition via Prompts" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 689 + ], + "type": "text", + "content": "Recent advances in general recognition models (e.g., detection/segmentation [10, 11, 18, 24, 27, 30, 49, 53, 56]) have demonstrated that multimodal alignment between vision and language modalities serves as the cornerstone for open-world understanding. These frameworks typically employ joint text-visual prompting mechanisms to achieve object localization through aligned cross-modal feature spaces." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Text Prompt Paradigm. Existing methods predominantly focus on text-based cross-modal alignment. GLIP" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "21289" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 558, + 322 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 558, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 558, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 558, + 322 + ], + "type": "image", + "image_path": "919e3f7873156ad81dc5661fa82214035904da1c6d36d4c1bf6d2b1d3bc2f8d2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 326, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 54, + 326, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 555, + 381 + ], + "type": "text", + "content": "Figure 2. The Overall Framework of Prompt-DINO. Prompt-DINO is constructed based on the DINO architecture. Images are processed by ViT to extract multi-scale features, while visual and text prompts are processed by deformable attention transformer layers and CLIP's text encoder, respectively, to derive embeddings. Through early fusion layers in the backbone encoder and perception-aware query selection, Prompt-DINO perform deep fusion and alignment of visual and text in two steps, with the final predictions generated by the Decoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 388, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 495 + ], + "type": "text", + "content": "[24] pioneers the adaptation of image-text pretraining to downstream detection tasks through deep feature fusion, establishing language-aware visual representations. GroundingDINO [27] further extends this paradigm by demonstrating that multi-stage cross-attention fusion significantly enhances performance in open-vocabulary detection. While these approaches achieve remarkable progress, their heavy reliance on textual semantics limits their capacity to handle visual concepts with ambiguous linguistic descriptions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 664 + ], + "type": "text", + "content": "Visual Prompt Paradigm. Complementary to text prompts, visual prompting (points/boxes/masks) provides geometric guidance for specific object localization. SAM [19] establishes a milestone in interactive segmentation through decoupled visual prompt encoding, though lacking semantic awareness. Subsequent works like SEEM [60] and DINOv [22] enhance this paradigm by integrating semantic context and exemplar-based reasoning. A notable hybrid approach is T-Rex2 [16], which combines both modalities but still treats visual prompts as auxiliary geometric cues. Our key insight lies in that existing methods inadequately exploit the semantic synergy between visual prompts and language guidance, leaving visual prompting's full potential underdeveloped for language-aware feature learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 672, + 261, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 261, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 261, + 685 + ], + "type": "text", + "content": "2.2. Data Engine for Open-World Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Scalable data curation has become pivotal for training robust open-world detector/segmentation models. SAM2 [36]" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "text", + "content": "introduces a cyclic model-annotator collaboration framework, iteratively refining the SA-V dataset through human-machine interactions. Alternative strategies employ cross-model verification: YOLO-World [6] curates CC3M-Lite by relabeling CC3M [39] with GLIP and CLIP, while T-Rex2 [16] propagates pseudo-labels across SA-1B and LAION400M [34]. Current methods face three key limitations: (1) Over-reliance on noun phrase extraction from captions misses more detailed concepts and introduces too many redundant nouns; (2) Single-path verification propagates annotation errors; (3) Semantic ambiguity in freeform text introduces label noise. Our data engine addresses these issues and enables the automatic generation of high-quality training data with reduced noise." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 560, + 370, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 370, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 370, + 572 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Motivated by the limitations observed in existing hybrid prompting methods for detection and segmentation tasks, we introduce Prompt-DINO, an end-to-end generic segmentation model built upon the DINO architecture [52]. Our model natively integrates multimodal prompting capabilities for both textual and visual inputs while maintaining architectural simplicity. In this section, we first formalize the promptable segmentation paradigm and outline the core design prompt encoders of Prompt-DINO (§3.1). We then systematically address three critical challenges: (1) early fusion with visual prompts through cross-modal feature align-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "21290" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 122 + ], + "type": "text", + "content": "ment(§3.2), (2) order-aware query selection for maintaining semantic consistency(§3.3) and (3) construction of a scalable data engine for open-world generalization(§3.4). Finally, we introduce the training objective(§3.5)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 127, + 226, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 127, + 226, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 226, + 140 + ], + "type": "text", + "content": "3.1. Promptable Segmentation Task" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": "The Prompt-DINO processes an input image " + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": ", with accompanying visual prompts " + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^v = \\{\\mathcal{P}_1^v, \\dots, \\mathcal{P}_K^v\\}" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": ", and textual prompts " + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^t = \\{\\mathcal{P}_1^t, \\dots, \\mathcal{P}_K^t\\}" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": " is the number of interactive prompts per image. Following established architectures in DETR [57] and DINO [52], our system first extracts hierarchical visual features " + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "F = \\{f_1, \\dots, f_L\\}" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": " through a vision backbone network (e.g., ViT [9] or Swin Transformer [28]), where " + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 144, + 296, + 251 + ], + "type": "text", + "content": " represents the number of feature pyramid levels." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 252, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 324 + ], + "type": "text", + "content": "Visual Prompt Encoder. Specifically, our visual encoding module employs " + }, + { + "bbox": [ + 55, + 252, + 296, + 324 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 252, + 296, + 324 + ], + "type": "text", + "content": " transformer blocks with deformable attention mechanisms [22]. For each target object, we initialize a learnable query embedding " + }, + { + "bbox": [ + 55, + 252, + 296, + 324 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 55, + 252, + 296, + 324 + ], + "type": "text", + "content": " and progressively refine it through multi-scale feature aggregation. The transformation at each layer " + }, + { + "bbox": [ + 55, + 252, + 296, + 324 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 252, + 296, + 324 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 331, + 295, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 331, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 61, + 331, + 295, + 347 + ], + "type": "interline_equation", + "content": "v _ {i} ^ {l} = M S D A \\left(v _ {i} ^ {l - 1}, f _ {l}\\right), l \\in [ 1, \\dots , L ], \\quad \\text {w i t h} \\quad v _ {i} ^ {0} = q _ {i} \\tag {1}", + "image_path": "ea7c40c290073ae678f23f6f821239f8a18d05e1c3cbd14d5354bff84cc199d0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "inline_equation", + "content": "MSDA(\\cdot)" + }, + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "text", + "content": " denotes multi-scale deformable attention operation. The final visual prompt representations " + }, + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "inline_equation", + "content": "V = \\{v_{1},\\dots,v_{K}\\}" + }, + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "text", + "content": " are obtained as " + }, + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "inline_equation", + "content": "v_{i} = v_{i}^{L}" + }, + { + "bbox": [ + 55, + 354, + 296, + 390 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 390, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 485 + ], + "type": "text", + "content": "Text Prompt Encoder. Building upon insights from recent open-world detection studies, we employ a CLIP [34] text encoder rather than conventional BERT [8] approaches used in earlier models like [24, 27]. This design choice enables better romantic correlation modeling for input words, phrases, and short sentences, generating text prompts " + }, + { + "bbox": [ + 55, + 390, + 296, + 485 + ], + "type": "inline_equation", + "content": "T = \\{t_1,\\dots,t_K\\}" + }, + { + "bbox": [ + 55, + 390, + 296, + 485 + ], + "type": "text", + "content": " that effectively guide cross-modal feature alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 534 + ], + "type": "text", + "content": "The integrated representation (text(T), visual(V), multiscale features(F)) feeds into the backbone encoder for multimodal fusion, enabling dynamic adaptation to prompts while preserving cross-modal alignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 237, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 237, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 237, + 553 + ], + "type": "text", + "content": "3.2. Early Fusion with Visual Prompts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "We are the first to propose the early fusion framework that integrates text prompts, visual prompts, and input image features through multi-modal interaction. While this fusion mechanism effectively reduces the modality gap between textual and visual features, enhances prompt-image correlation, and improves target recall, We found unintended consequences of severe hallucinations in our experiments. This phenomenon stems from the inherent limitation of conventional cross-attention mechanisms: they compulsively reconstruct Key features most similar to the Query features, even when all Key features are substantially dissimilar to the current Query. To address this non-target feature reconstruction issue, we introduce a novel gated cross" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 68, + 556, + 265 + ], + "blocks": [ + { + "bbox": [ + 316, + 68, + 556, + 265 + ], + "lines": [ + { + "bbox": [ + 316, + 68, + 556, + 265 + ], + "spans": [ + { + "bbox": [ + 316, + 68, + 556, + 265 + ], + "type": "image", + "image_path": "ae748a4d13be1c914e0c223da049e14d11c49083d6f7c9c661587aa6ed6cb778.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 265, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 321 + ], + "type": "text", + "content": "Figure 3. Structure of Early Fusion Layer. Notably, the contrastive alignment of visual and text prompts is completed before early fusion. By employing Gated Cross-Attention, text prompts are reconstructed with image features, while visual prompts and image features mutually reconstruct each other." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 324, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 556, + 515 + ], + "type": "text", + "content": "attention mechanism incorporating a learnable background token " + }, + { + "bbox": [ + 313, + 324, + 556, + 515 + ], + "type": "inline_equation", + "content": "(B)" + }, + { + "bbox": [ + 313, + 324, + 556, + 515 + ], + "type": "text", + "content": ". Reconstruction occurs exclusively when the Query-Key similarity exceeds an adaptive threshold; otherwise, the background token automatically activates to suppress hallucination. As illustrated in Fig. 3, our architecture comprises three core components: 1) a self-attention layer for intra-modal refinement, 2) the proposed gated cross-attention layer for inter-modal interaction, and 3) a feedforward neural network (FFN) for feature transformation. The cross-attention module implements three interaction pathways: feature-to-visual prompt, visual prompt-to-feature, and feature-to-text prompt. For visualization clarity, Fig. 3 presents only stage-2 features mapped to text space through our two-phase training scheme (detailed in §3.5). The enhanced cross-modal fusion operates through the following formulation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 380, + 520, + 492, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 520, + 492, + 534 + ], + "spans": [ + { + "bbox": [ + 380, + 520, + 492, + 534 + ], + "type": "interline_equation", + "content": "t _ {i} ^ {\\prime} = \\operatorname {G a t e d A t t n} (f _ {i}, t _ {i}, t _ {i})", + "image_path": "a7f47ffdb0e766c0259dcf5988f8a51c188dd361e0cdcfa06fd619b9cb9f3b34.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 378, + 536, + 553, + 549 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 536, + 553, + 549 + ], + "spans": [ + { + "bbox": [ + 378, + 536, + 553, + 549 + ], + "type": "interline_equation", + "content": "v _ {i} ^ {\\prime} = \\operatorname {G a t e d A t t n} \\left(f _ {i}, v _ {i}, v _ {i}\\right) \\tag {2}", + "image_path": "66429f10b4900ca9cb2f9c38a332ce03d10278397b2ea0dfc176ae1ac43c0ef2.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 378, + 551, + 490, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 551, + 490, + 564 + ], + "spans": [ + { + "bbox": [ + 378, + 551, + 490, + 564 + ], + "type": "interline_equation", + "content": "f _ {i} ^ {\\prime} = \\operatorname {G a t e d A t t n} (v _ {i}, f _ {i}, f _ {i})", + "image_path": "27687196fb3c6ec29d9a2b95ba020a203573b075d92454b41f41c16454749ff4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 571, + 522, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 571, + 522, + 582 + ], + "spans": [ + { + "bbox": [ + 314, + 571, + 522, + 582 + ], + "type": "text", + "content": "where the gated attention mechanism is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 327, + 586, + 555, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 586, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 327, + 586, + 555, + 606 + ], + "type": "interline_equation", + "content": "\\operatorname {G a t e d A t t n} (Q, K, V) = \\operatorname {s o f t m a x} (\\frac {Q \\cdot [ K , B ]}{\\sqrt {d _ {k}}}) [ V, B ] \\quad (3)", + "image_path": "8e447136a8d94005e6bc6c9b472bbe06c4d3478013f5e58226f7dad43a75ad45.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 610, + 555, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 555, + 669 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 610, + 555, + 669 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 610, + 555, + 669 + ], + "type": "text", + "content": " denotes our learnable background token. Following TRex-2 [16] methodology, we employ region-level contrastive learning to align visual and textual modalities, utilizing symmetric cross-entropy loss [34] for embedding alignment:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 361, + 660, + 553, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 660, + 553, + 692 + ], + "spans": [ + { + "bbox": [ + 361, + 660, + 553, + 692 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = - \\frac {1}{K} \\sum_ {i = 1} ^ {K} \\log \\frac {v _ {i} \\cdot t _ {i}}{\\sum_ {j = 1} ^ {K} v _ {i} \\cdot t _ {j}} \\tag {4}", + "image_path": "f069664e3619ffd05c758e1de342cc8312c76e18aca504e09eb717c1a56954c2.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 690, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 690, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 690, + 553, + 714 + ], + "type": "text", + "content": "In our experience, symmetric cross-entropy losses have better convergence than InfoNCE loss [31]." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "21291" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "In the context of training detection and segmentation tasks, it is common practice to construct negative samples. For negative prompt construction, while text prompts naturally leverage category vocabularies for negative sampling, visual prompts require explicit negative construction. We address this by aggregating visual prompt features across all categories within each batch and computing inter-category averages to generate negative visual prompts. This approach ensures sufficient diversity in negative samples while maintaining semantic relevance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 294, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 294, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 294, + 310 + ], + "type": "text", + "content": "To mitigate annotation granularity discrepancies across datasets (e.g., \"bird\" in Dataset A vs. \"pigeon\" in Dataset B), we implement a dataset-aware sampling strategy. Our novel data sampler ensures each training batch contains only intra-dataset samples, preventing semantically similar but categorically distinct concepts from becoming conflicting negative samples. This is particularly critical in visual-prompt open-world detection/segmentation scenarios where label hierarchies may differ substantially across source datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 318, + 212, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 212, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 212, + 331 + ], + "type": "text", + "content": "3.3. Order-aware query selection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 514 + ], + "type": "text", + "content": "Prompt-DINO adopts a DETR-based architecture that selects top-K encoder features through similarity computation between encoder features and prompt embeddings, following DINO [52] and MaskDINO [21]. However, our approach fundamentally diverges from existing methods by simultaneously considering both textual-visual query similarity and inter-query correlations with visual prompts. To further strengthen the alignment between visual and textual prompts, we introduce order-aware alignment during query selection, which establishes the second-level alignment between visual and textual prompts during training. Specifically, we employ Kendall's " + }, + { + "bbox": [ + 55, + 335, + 295, + 514 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 335, + 295, + 514 + ], + "type": "text", + "content": " coefficient to quantify the ordinal consistency [12] between queries selected by visual prompts and text prompts, the Kendall's " + }, + { + "bbox": [ + 55, + 335, + 295, + 514 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 335, + 295, + 514 + ], + "type": "text", + "content": " coefficient can be expressed as: " + }, + { + "bbox": [ + 55, + 335, + 295, + 514 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{c}} - P_{\\mathrm{d}}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 505, + 294, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 505, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 138, + 505, + 294, + 531 + ], + "type": "interline_equation", + "content": "\\tau = \\frac {P _ {c} - P _ {d}}{\\frac {1}{2} N (N - 1)} \\tag {5}", + "image_path": "954ea3f90222a553fe90b6bb93e694939ef2f205725123ca7b4a51d7c2e90dd9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "text", + "content": " represents concordant pairs and " + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "text", + "content": " represents discordant pairs, " + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "text", + "content": " represents the total number of queries. Given the similarity score between query embeddings and prompt embeddings " + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "inline_equation", + "content": "S^{v} = s_{1}^{v},\\ldots ,s_{N}^{v}" + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "inline_equation", + "content": "S^t = s_1^t,\\dots,s_N^t" + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "text", + "content": " we calculate " + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "inline_equation", + "content": "P_{c}" + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 55, + 531, + 295, + 589 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 589, + 294, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 589, + 294, + 616 + ], + "spans": [ + { + "bbox": [ + 87, + 589, + 294, + 616 + ], + "type": "interline_equation", + "content": "\\tau = \\frac {\\sum_ {i} \\sum_ {j < i} \\operatorname {s g n} \\left(s _ {i} ^ {t} - s _ {j} ^ {t}\\right) \\cdot \\operatorname {s g n} \\left(s _ {i} ^ {v} - s _ {j} ^ {v}\\right)}{\\frac {1}{2} N (N - 1)} \\tag {6}", + "image_path": "b78bcafc801c740cb4f7044b1c10db9f4f0900e68af71e8282757c5a1f1a26c1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 616, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 651 + ], + "type": "text", + "content": "Since the sgn is not differentiable, we approximate the sign function with the tanh function and utilize the opposite of " + }, + { + "bbox": [ + 55, + 616, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 616, + 295, + 651 + ], + "type": "text", + "content": " as the order alignment loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 651, + 294, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 651, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 61, + 651, + 294, + 677 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {o r d e r}} = - \\frac {\\sum_ {i} \\sum_ {j < i} \\tanh \\left(s _ {i} ^ {t} - s _ {j} ^ {t}\\right) \\cdot \\tanh \\left(s _ {i} ^ {v} - s _ {j} ^ {v}\\right)}{\\frac {1}{2} N (N - 1)} \\tag {7}", + "image_path": "720dcf1f0e6ec739ed0d271beeb7451ae8fd40bc763a7dca467d20ff7747f2cc.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 678, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 295, + 714 + ], + "type": "text", + "content": "Our proposed alignment mechanism operates at two complementary levels: The primary alignment directly optimizes the learning objective through " + }, + { + "bbox": [ + 55, + 678, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{align}}" + }, + { + "bbox": [ + 55, + 678, + 295, + 714 + ], + "type": "text", + "content": " (as given in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "Eq. 4), while our novel order-aware alignment provides indirect assurance through two critical aspects: (1) Semantic proximity between corresponding text and visual prompts for candidate queries naturally reduces overall training complexity by minimizing feature space discrepancies; (2) Training-inference consistency preservation ensures equivalent substitutability between text and visual prompts during inference through maintained representational parity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 169, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 553, + 276 + ], + "type": "text", + "content": "Fig. 5 demonstrates the effectiveness of our proposed method. Prior to applying the order-aware loss optimization, we observe significant disorder in both textual and visual prompt queries, with no discernible alignment between corresponding semantic representations. The optimization process induces a remarkable transformation - post-optimization analysis reveals substantially enhanced consistency between semantically equivalent textual and visual prompt queries." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 286, + 395, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 395, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 395, + 298 + ], + "type": "text", + "content": "3.4. Data Engine" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 304, + 554, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 448 + ], + "type": "text", + "content": "Effective visual concept learning from rich image-text pairs is fundamental to enhancing the generalization capabilities of open-vocabulary object detectors. Existing methods predominantly rely on captioning models([3, 23, 26] etc.), coupled with lexical analysis tools like SpaCy [14] to extract noun phrases. These noun phrases are subsequently employed to generate pseudo-labels via open-set detection models ([24, 27]). Nonetheless, these techniques have significant drawbacks. They frequently omit valid labels, introduce erroneous annotations, and are susceptible to irrelevant lexical noise. All of these issues detrimentally affect detection performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 544 + ], + "type": "text", + "content": "To address these challenges, we propose a novel Recognize Anything via Prompting (RAP) model, a 0.5B MLLM, which significantly enhances annotation accuracy. Our framework further integrates an innovative dual-path cross-validation pipeline that substantially reduces label noise. This comprehensive approach enables the fully automated generation of high-quality open-world detection datasets while preserving the fidelity of annotations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 700 + ], + "type": "text", + "content": "Recognize Anything via Prompting. Despite the remarkable capabilities of the Segment Anything Model (SAM) [19] in detecting a wide array of objects within images, and the substantial progress achieved in region-level understanding through multimodal models [1, 2, 33, 35, 51], a critical question persists: Can we effectively assign semantic labels to arbitrary mask prompts? To address this challenge, we introduce RAP, a novel framework that bridges visual segmentation with semantic recognition. As illustrated in Fig. 4, our RAP model processes raw image inputs alongside their corresponding object masks to generate precise semantic annotations for the specified regions. The proposed RAP comprises three key components:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 327, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 327, + 701, + 553, + 713 + ], + "type": "text", + "content": "1. Feature extraction. We perform masked pooling over" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "21292" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 387, + 178 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 387, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 387, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 387, + 178 + ], + "type": "image", + "image_path": "4b51941a84aacf5fc7d55a29f37b0b50e268f690b5918da5e650c453af95325f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 555, + 208 + ], + "type": "text", + "content": "Figure 4. Data Engine and RAP Frameworks. Left: The data engine with dual-path cross-verification pipeline. Right: The proposed Recognize Anything via Prompting (RAP), a 0.5B-parameter MLLM model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 401, + 70, + 554, + 178 + ], + "blocks": [ + { + "bbox": [ + 401, + 70, + 554, + 178 + ], + "lines": [ + { + "bbox": [ + 401, + 70, + 554, + 178 + ], + "spans": [ + { + "bbox": [ + 401, + 70, + 554, + 178 + ], + "type": "image", + "image_path": "525074c2087d8413737cf9a47c9ee6667778f38f68f406135b0da0974b7c1eea.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 214, + 295, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 214, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 54, + 214, + 295, + 309 + ], + "type": "text", + "content": "multi-layer image features to extract region-specific features corresponding to input masks, and concatenate the mask token embeddings obtained via a projection layer with the original image tokens to form a unified representation. This preserves both global contextual information from the complete image and localized object-specific features within masked regions, effectively integrating scene-level understanding with instance-level visual cues." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 312, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 295, + 443 + ], + "type": "text", + "content": "2. Ground Data Cleaning. To standardize the format of our ground truth data, we leveraged the Qwen model [48] to eliminate redundant descriptive phrases through comprehensive text normalization. Specifically, we transformed verbose expressions into their canonical semantic equivalents (e.g., converting \"any part of the sky\" to \"sky\"). This preprocessing strategy serves dual purposes: (1) it reduces the RAP's learning complexity by minimizing lexical variability, and (2) it mitigates the risk of hallucinations by ensuring precise semantic alignment between textual annotations and visual content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 445, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 445, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 295, + 529 + ], + "type": "text", + "content": "3. Training. The proposed methodology employs a two-stage training strategy. During the first stage, we conduct pre-training using 10M samples, specifically optimizing only the projector while keeping other model components frozen. The second stage implements instruction-tuning, where we unfreeze and jointly optimize the entire model to enhance task-specific adaptation capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 530, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 530, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 54, + 530, + 295, + 579 + ], + "type": "text", + "content": "Dual-path cross-verification pipeline. We propose a dual-pipeline annotation framework combining bottom-up and top-down approaches with cross-verification filtering, consisting of three key components:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 580, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 652 + ], + "type": "text", + "content": "1. Top-Down Pipeline. Our top-down pipeline leverages the Qwen-VL model [42] to extract semantic tags from images. These tags are subsequently fed into an open-set detection model to generate corresponding bounding boxes. This pipeline demonstrates strong capability in detecting large-scale salient objects (e.g., trees, roads, sky)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "2. Bottom-Up Pipeline. To address the inherent limitation of top-down pipeline in capturing fine-grained details, we implement a complementary bottom-up strategy using the RAP integrated with SAM [19]. This combination enables effective detection of smaller objects and detailed" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 214, + 553, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 553, + 237 + ], + "type": "text", + "content": "components (e.g., eyes, noses, windows) that are frequently missed by conventional top-down pipeline." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 238, + 554, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 554, + 357 + ], + "type": "text", + "content": "3. Cross-Verification. The detection outputs from both pipelines undergo rigorous validation through a three-stage process. First, We employ the Hungarian matching algorithm to establish correspondences between bounding boxes from different pipelines. Next, we utilize the BGE model [47] to compute similarity scores between the tags associated with the matched bounding boxes. Finally, we filter out the candidate pairs whose similarity scores were below the threshold of empirical determination to further clean up the annotated results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 425, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 425, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 425, + 378 + ], + "type": "text", + "content": "3.5. Training Objective" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 382, + 554, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 554, + 502 + ], + "type": "text", + "content": "Building upon previous DINO-like models [52], we employ a combination of L1 loss and GIoU loss [37] for bounding box regression, while for mask segmentation, we adopt cross-entropy loss and dice loss [29]. Following the methodology established in GroudingDINO [27], we utilize contrastive loss between predicted objects and text prompts for classification tasks. To accelerate model convergence, we incorporate denoising training as implemented in DINO [52]. The complete objective function for Prompt DINO can be expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 512, + 553, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 512, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 323, + 512, + 553, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c l s} + \\mathcal {L} _ {b b o x} + \\mathcal {L} _ {m a s k} + \\mathcal {L} _ {D N} + \\mathcal {L} _ {a l i g n} + \\mathcal {L} _ {o r d e r} \\tag {8}", + "image_path": "d1710fadc4f1f347f7a5ae1b71d13924e280e91eda0c1512917aef97129eec95.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 555, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 583 + ], + "type": "text", + "content": "Our training methodology adopts a progressive multi-stage strategy: we first optimize the text prompt while keeping the visual components fixed, followed by joint optimization of both textual and visual prompt representations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 593, + 395, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 395, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 395, + 605 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 612, + 374, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 374, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 374, + 625 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 712 + ], + "type": "text", + "content": "In Prompt-DINO, we employ ViT-L [9] as the backbone and fine-tune CLIP-L [34] as our text prompt encoder. The visual prompt encoder consists of 6 layers of deformable cross-attention transformer layers, with a hidden dimension of 256. We adopt the \"two-stage\" structure proposed by DINO [52], selecting 900 queries for decoding after the backbone encoder." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "21293" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 54, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 58, + 54, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 54, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 54, + 553, + 232 + ], + "type": "table", + "html": "
MethodBackbonePrompt TypeCOCO (in-domain)LVIS (in-domain)ADE20K (out-domain)
mask APbox APmask APbox APPQmask APbox AP
Mask2Former [4]Swin-LClose-Set48.652.1-----
OneFormer [15]Swin-LClose-Set48.9------
MaskDINO* [21]Swin-LClose-Set50.656.2-----
kMaX-DeepLab [50]Swin-LClose-Set58.1------
GLIPv2 [53]Swin-HText Prompt48.9------
X-Decoder [59]DaViT-LText Prompt46.7---21.813.138.1
OpenSeeD [54]Swin-LText Prompt53.258.221.023.019.715.017.7
T-Rex2 [16]Swin-LText Prompt-52.2*-45.8*---
APE [40]ViT-LText Prompt49.358.353.059.627.224.429.6
DINOv [22]Swin-LVisual Prompt50.454.2--23.215.114.3
T-Rex2 [16]Swin-LVisual Prompt-46.5*-45.3*---
Prompt-DINOViT-LText Prompt51.158.656.560.129.624.628.6
Prompt-DINOViT-LVisual Prompt51.058.055.659.935.928.032.7
", + "image_path": "447feed9f38ad163ba0efef88a0e8682476bebbb75ac07b62a548321e5729834.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 233, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 555, + 266 + ], + "type": "text", + "content": "Table 1. Generic Segmentation results of Prompt-DINO on multiple datasets. \"-\" denotes the model does not have performance reported or does not have ability for the specific task. \"+\" denotes the performance reported under different domain settings. MaskDINO uses the results reported in DINOv [22]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 271, + 294, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 294, + 498 + ], + "type": "text", + "content": "Datasets. During the training process, we utilized three categories of datasets: open-source detection datasets. These include COCO [25] with 118K images and LVIS [13] with 160K images, both of which have mask annotations. For datasets without mask annotations, such as Object365 [38], OID [20], and V3Det [41], we first apply SAM to generate mask annotations. Due to the presence of low-quality data and long-tail distribution in Object365 and OID, which hinder general object recognition, we resampled them, selecting 700K and 1M data points, respectively. grounding datasets. This includes the GoldG [17] dataset with 100K images. self-built datasets. We re-announce and filter CC3M [39] and SA-1B [19] using the RAP and dual-path cross-validation data engine introduced in §3.4, yielding high-quality datasets of 3M and 7M images, respectively, which constitute our primary training data. For testing, we utilize COCO and LVIS for in-domain evaluation and ADE20K [55], SeginW [59], and Cityscapes [7] for out-domain evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 582 + ], + "type": "text", + "content": "Metrics. We employ Average Precision (AP) as the evaluation metric, utilizing mask AP for instance segmentation, and box AP for object detection. For ADE20K and CtiyScapes, we employ Panoptic Quality(PQ) and mean intersection over union (mIoU) for panoptic segmentation as well. For SegInW, we adopt AP-Average and AP-Median as the evaluation metric." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 588, + 182, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 182, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 182, + 601 + ], + "type": "text", + "content": "4.2. Generic Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "content": "To evaluate the effectiveness and generalization capability of our Prompt-DINO, we conduct evaluations on both indomain and out-domain datasets, with the results presented in Tab. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "In-Domain Segmentation. We conduct in-domain experiments on COCO and LVIS. Compared to methods based on visual prompts, our Prompt-DINO achieved significantly better results. Compared to the previous work DINOv [22], we improve mask AP by 0.6 and box AP by 3.8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 271, + 553, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 553, + 319 + ], + "type": "text", + "content": "on COCO. Compared to methods based on text prompts, Prompt-DINO also demonstrates superior results when using text prompts. For example, compared to APE [40], we lead by 3.5 in mask AP and 0.5 in box AP on LVIS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "text", + "content": "Out-Domain Segmentation. We use ADE-20K and Cityscapes as an out-domain dataset and conduct zero-shot testing on it, demonstrating the strong generalization ability of our Prompt-DINO. Under the out-domain setting (where data marked with * indicates performance reported under different domain settings), our method achieves the best performance on ADE20K across all methods for both prompt types, whether using visual prompts or text prompts. Compared to the state-of-the-art method APE on ADE20K, we improve PQ by 8.7, mask AP by 3.6, and box AP by 18.4, representing increases of " + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "inline_equation", + "content": "32.0\\%" + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "inline_equation", + "content": "14.8\\%" + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "inline_equation", + "content": "10.5\\%" + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "text", + "content": ", respectively. Among methods based on visual prompts, compared to the previous method DINOv, we improve PQ by 12.7, mask AP by 12.9, and box AP by 18.4, representing increases of " + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "inline_equation", + "content": "54.7\\%" + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "inline_equation", + "content": "85.4\\%" + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "inline_equation", + "content": "128.7\\%" + }, + { + "bbox": [ + 313, + 319, + 554, + 521 + ], + "type": "text", + "content": ", respectively. Compared to APE on Cityscapes, we also improve PQ by 1.0 and mIoU by 8.0." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 523, + 554, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 554, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 554, + 631 + ], + "type": "text", + "content": "Segmentation in the Wild. To further evaluate the generalization capabilities of our model, we perform zero-shot testing on the SeginW benchmark, which includes 25 datasets. Compared to the previous work DINOv, when using visual prompts, we improve the average AP by 16.2 and the median AP by 14.5. Among text-prompt-based methods, we also surpass APE by 6.8 in average AP and 6.9 in median AP. This further validates the strong generalization ability of our Prompt-DINO." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 636, + 440, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 440, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 440, + 649 + ], + "type": "text", + "content": "4.3. Effectiveness Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "To validate the effectiveness of our proposed components, we conduct ablation studies on Text Prompt Alignment, Early Fusion, Order Alignment, and Data Engine. As our baseline, we implement the naive MaskDINO method with the FC head replaced with a user-prompted visual embed" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "21294" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 54, + 293, + 140 + ], + "blocks": [ + { + "bbox": [ + 58, + 54, + 293, + 140 + ], + "lines": [ + { + "bbox": [ + 58, + 54, + 293, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 54, + 293, + 140 + ], + "type": "table", + "html": "
MethodTypeCityScapeSegInW
PQmIoUAP-avgAP-med
kMaX-DeepLab-L [50]in-domain68.483.5--
X-Decoder-L [59]out-domain38.152.022.332.3
OpenSeeD-L [54]out-domain41.447.836.138.7
APE-L [40]out-domain33.344.249.652.2
DINOv [22]out-domain--40.644.6
Prompt-DINO [40]out-domain34.352.256.459.1
", + "image_path": "c060a39977021d7a6cd317d419f6976afa0eea310ebce380062ebff384d9a7f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 167, + 294, + 226 + ], + "blocks": [ + { + "bbox": [ + 56, + 142, + 294, + 164 + ], + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 164 + ], + "type": "text", + "content": "Table 2. Zero-shot Segmentation results of Prompt-DINO on the CityScape [7] and SegInW [59] dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 167, + 294, + 226 + ], + "lines": [ + { + "bbox": [ + 57, + 167, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 57, + 167, + 294, + 226 + ], + "type": "table", + "html": "
BaselineMethodADE20K (out-domain)
Text AlignEarly FusionOrder AlignPQmAPbAP
✓×××25.020.423.5
✓✓××32.725.530.5
✓✓✓×34.726.731.7
✓✓✓✓35.928.032.7
", + "image_path": "d153fc5c4170300da08ca9fb90898aac05b643b1c6a3e36c210bcf415d9c5de2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 243, + 256, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 256, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 256, + 254 + ], + "type": "text", + "content": "ding. The ablation results are presented in Tab. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 255, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 434 + ], + "type": "text", + "content": "Effectiveness of Early Fusion. By comparing rows 1 and 2 in Tab. 3, it can be observed that incorporating text align in the backbone encoder (stage one of DINO) increases the mask AP and box AP on ADE20K by 5.1 and 7.0, respectively, and PQ by 7.7, compared to the baseline. This indicates that naive alignment of text and visual can enhance the model's general recognition capability. Comparing rows 2 and 3 in Tab. 3 shows that early fusion of visual prompts, guided by text alignment, significantly improves the model's performance, with mask AP and box AP on ADE20K increasing by 1.2 and 1.2, respectively, and PQ by 2.0. This demonstrates that visual prompts effectively complement features that are difficult to describe with text prompts and, under the guidance of text prompts, achieve more generalizable expressive power." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 438, + 176, + 534 + ], + "blocks": [ + { + "bbox": [ + 58, + 438, + 176, + 534 + ], + "lines": [ + { + "bbox": [ + 58, + 438, + 176, + 534 + ], + "spans": [ + { + "bbox": [ + 58, + 438, + 176, + 534 + ], + "type": "image", + "image_path": "f54bb0f07caa5d1ec7f50ff2be7491c0a9c7f05cfd7227fc660e53e168957e68.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 602 + ], + "type": "text", + "content": "Figure 5. Order Alignment Comparison. Left: Visual and text should demonstrate strong correlation, but the queries selected by the two prompts exhibit significant disorder. Right: Order alignment significantly improves the consistency of visual-text query selection and reinforces the alignment between corresponding semantics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 175, + 438, + 292, + 534 + ], + "blocks": [ + { + "bbox": [ + 175, + 438, + 292, + 534 + ], + "lines": [ + { + "bbox": [ + 175, + 438, + 292, + 534 + ], + "spans": [ + { + "bbox": [ + 175, + 438, + 292, + 534 + ], + "type": "image", + "image_path": "b80a80b6a055e66a2f9122a8603a864e4e08f3ca850e750c8f8edac2873955b4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Effectiveness of Text-guided Visual Prompts. Rows 1, 2, and 3 in Tab. 3 show that merely adding text alignment provides limited improvement to the model. However, after early fusion to eliminate semantic ambiguity, the two types of prompts can serve as complementary information, providing stronger recognition capabilities. Next, we introduced order alignment during the query selection stage to further guide visual prompts with text. Rows 3 and 4 in Tab. 3 demonstrate that our two-stage guidance deeply" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 50, + 429, + 144 + ], + "blocks": [ + { + "bbox": [ + 317, + 50, + 429, + 144 + ], + "lines": [ + { + "bbox": [ + 317, + 50, + 429, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 50, + 429, + 144 + ], + "type": "image", + "image_path": "01ab29b901c447db640264a901dde471c25a001ec72a6dc947e097922d6e528a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 148, + 544, + 159 + ], + "lines": [ + { + "bbox": [ + 323, + 148, + 544, + 159 + ], + "spans": [ + { + "bbox": [ + 323, + 148, + 544, + 159 + ], + "type": "text", + "content": "Figure 6. Annotation quality and quantity in the data engine." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 441, + 50, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 441, + 50, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 441, + 50, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 441, + 50, + 553, + 144 + ], + "type": "image", + "image_path": "244c7a54e5839ef223edd1855bfdd42f7d4c3832ace980e9a3522012903f9ddb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 169, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 553, + 228 + ], + "type": "text", + "content": "aligns visual and text prompts, leading to further performance improvements. Fig. 5 also demonstrates that our order alignment method further enhances the alignment between visual and text, resulting in a more consistent selection order for queries." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "type": "text", + "content": "Effectiveness of Data Engine. Within the data engine, we employ a cross-validation module to filter data annotated by the top-down and bottom-up pipelines. As illustrated in Fig. 6, the cross-validation module filters approximately " + }, + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "type": "inline_equation", + "content": "80.5\\%" + }, + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "type": "text", + "content": " of instance annotations and enhances the annotation clip similarity from 0.16 to 0.23, highlighting the effectiveness of the cross-validation module in generating high-quality data. As shown in Tab. 4, incorporating data generated by our data engine into the training set improves Prompt-DINO's box AP by " + }, + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "type": "text", + "content": " on LVIS and " + }, + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "type": "text", + "content": " on ADE20K. To further demonstrate the effectiveness of RAP, we process a subset of SA-1B using different data engines in Tab. 5, revealing that RAP enhances model performance by " + }, + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "type": "text", + "content": " mask AP and " + }, + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 313, + 228, + 555, + 433 + ], + "type": "text", + "content": " box AP on ADE20K. This indicates that the proposed data engine effectively enhances data quality, while RAP generates more diverse data, significantly improving the model's generalization capability." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 317, + 436, + 552, + 483 + ], + "blocks": [ + { + "bbox": [ + 94, + 228, + 256, + 239 + ], + "lines": [ + { + "bbox": [ + 94, + 228, + 256, + 239 + ], + "spans": [ + { + "bbox": [ + 94, + 228, + 256, + 239 + ], + "type": "text", + "content": "Table 3. Ablation of proposed components." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 436, + 552, + 483 + ], + "lines": [ + { + "bbox": [ + 317, + 436, + 552, + 483 + ], + "spans": [ + { + "bbox": [ + 317, + 436, + 552, + 483 + ], + "type": "table", + "html": "
MethodLVIS (in-domain)ADE20K (out-domain)
mask APbox APPQmask APbox AP
w/o DataEngine46.047.930.823.426.8
w Data Engine55.659.935.928.032.7
", + "image_path": "c5e19d07751aaf12a15780da3f75f7e20e8d1a8ca53ab3565ce5c6f20245deb1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 317, + 496, + 552, + 543 + ], + "blocks": [ + { + "bbox": [ + 326, + 485, + 541, + 496 + ], + "lines": [ + { + "bbox": [ + 326, + 485, + 541, + 496 + ], + "spans": [ + { + "bbox": [ + 326, + 485, + 541, + 496 + ], + "type": "text", + "content": "Table 4. Ablation of Data Engine (Visual Prompt Results)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 496, + 552, + 543 + ], + "lines": [ + { + "bbox": [ + 317, + 496, + 552, + 543 + ], + "spans": [ + { + "bbox": [ + 317, + 496, + 552, + 543 + ], + "type": "table", + "html": "
DatasetCOCO (in-domain)ADE20K (out-domain)
mask APbox APPQmask APbox AP
COCO+GoldG49.855.718.211.011.6
COCO+SA-1B (QwenVL)50.156.722.317.820.4
COCO+SA-1B (QwenVL + RAP)50.858.122.519.423.0
", + "image_path": "aa96fb539b914b20f7034adf588dccb37a4be3087cfff6a1398c14e892cafe78.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 544, + 554, + 566 + ], + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 566 + ], + "type": "text", + "content": "Table 5. Ablation of Training Dataset (Text Prompt Results). For efficient training, we use a 2M subset of SA-1B." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 567, + 388, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 388, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 388, + 578 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "We present Prompt-DINO, a text-guided visual Prompt DINO for generic segmentation via three innovations: 1) Early fusion of text, visual, and backbone features during initial encoding to resolve cross-modal ambiguity; 2) Order-aligned query selection for semantic-spatial consistency between text and visual cues in DETR decoding; 3) A generative data engine (RAP + dual-path verification) synthesizing 0.5B high-fidelity training instances, reducing label noise by " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "80.5\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". Prompt-DINO sets new state-of-the-art results while enabling scalable adaptation to real-world object diversity." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "21295" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work was partially supported by the National Key R&D Program of China (2022YFB4701400/4701402), SSTIC Grant(KJZD20230923115106012, KJZD20230923 114916032, GJHZ20240218113604008), National Natural Science Foundation of China (No. 62106149)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 174 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 180, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 296, + 225 + ], + "type": "text", + "content": "[1] Chi Chen, Ruoyu Qin, Fuwen Luo, Xiaoyue Mi, Peng Li, Maosong Sun, and Yang Liu. Position-enhanced visual instruction tuning for multimodal large language models. arXiv preprint arXiv:2308.13437, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 226, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 226, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 226, + 296, + 269 + ], + "type": "text", + "content": "[2] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 270, + 296, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 296, + 324 + ], + "type": "text", + "content": "[3] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370-387. Springer, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 326, + 296, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 296, + 380 + ], + "type": "text", + "content": "[4] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 381, + 296, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 381, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 381, + 296, + 425 + ], + "type": "text", + "content": "[5] Runxi Cheng, Feng Xiong, Yongxian Wei, Wanyun Zhu, and Chun Yuan. Whoever started the interference should end it: Guiding data-free model merging via task vectors. arXiv preprint arXiv:2503.08099, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 426, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 296, + 480 + ], + "type": "text", + "content": "[6] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 296, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 296, + 547 + ], + "type": "text", + "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 548, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 296, + 624 + ], + "type": "text", + "content": "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 625, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 296, + 689 + ], + "type": "text", + "content": "[9] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[10] Golnaz Ghiasi, Xiuye Gu, Yin Cui, and Tsung-Yi Lin. Scaling open-vocabulary image segmentation with image-level" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "labels. In European conference on computer vision, pages 540-557. Springer, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "text", + "content": "[12] Yuchen Guan, Runxi Cheng, Kang Liu, and Chun Yuan. Enhancing logits distillation with plug&play kendall's " + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "text", + "content": " ranking loss, 2025. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[13] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "type": "text", + "content": "[14] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, Adriane Boyd, et al. spacy: Industrial-strength natural language processing in python. 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 555, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 305 + ], + "type": "text", + "content": "[15] Jitesh Jain, Jiachen Li, Mang Tik Chiu, Ali Hassani, Nikita Orlov, and Humphrey Shi. Oneformer: One transformer to rule universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2989-2998, 2023. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 306, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 350 + ], + "type": "text", + "content": "[16] Qing Jiang, Feng Li, Zhaoyang Zeng, Tianhe Ren, Shilong Liu, and Lei Zhang. T-rex2: Towards generic object detection via text-visual prompt synergy. In European Conference on Computer Vision, pages 38-57. Springer, 2024. 2, 3, 4, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 351, + 553, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 351, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 351, + 553, + 405 + ], + "type": "text", + "content": "[17] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 406, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 553, + 459 + ], + "type": "text", + "content": "[18] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 526, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 601 + ], + "type": "text", + "content": "[20] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 668 + ], + "type": "text", + "content": "[21] Feng Li, Hao Zhang, Huaizhe Xu, Shilong Liu, Lei Zhang, Lionel M Ni, and Heung-Yeung Shum. Mask dino: Towards a unified transformer-based framework for object detection and segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3041-3050, 2023. 2, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 714 + ], + "type": "text", + "content": "[22] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Jianwei Yang, Chunyuan Li, et al. Visual in-context prompting. In Proceedings of the IEEE/CVF Conference on Computer Vision" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "21296" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "and Pattern Recognition, pages 12861-12871, 2024. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "text", + "content": "[24] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10965-10975, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13, pages 740-755. Springer, 2014. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 319 + ], + "type": "text", + "content": "[26] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 386 + ], + "type": "text", + "content": "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 442 + ], + "type": "text", + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 497 + ], + "type": "text", + "content": "[29] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international conference on 3D vision (3DV), pages 565-571. IEEE, 2016. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 565 + ], + "type": "text", + "content": "[30] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, et al. Simple open-vocabulary object detection. In European conference on computer vision, pages 728-755. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "text", + "content": "[31] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 634 + ], + "type": "text", + "content": "[32] Ting Pan, Lulu Tang, Xinlong Wang, and Shiguang Shan. Tokenize anything via prompting. In European Conference on Computer Vision, pages 330-348. Springer, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[33] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3, 4, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 173 + ], + "type": "text", + "content": "[35] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[36] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 232, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 296 + ], + "type": "text", + "content": "[37] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 353 + ], + "type": "text", + "content": "[38] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8430-8439, 2019. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 355, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 419 + ], + "type": "text", + "content": "[39] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 422, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 487 + ], + "type": "text", + "content": "[40] Yunhang Shen, Chaoyou Fu, Peixian Chen, Mengdan Zhang, Ke Li, Xing Sun, Yunsheng Wu, Shaohui Lin, and Rongrong Ji. Aligning and prompting everything all at once for universal visual perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13193–13203, 2024. 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "type": "text", + "content": "[41] Jiaqi Wang, Pan Zhang, Tao Chu, Yuhang Cao, Yujie Zhou, Tong Wu, Bin Wang, Conghui He, and Dahua Lin. V3det: Vast vocabulary visual detection dataset. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19844-19854, 2023. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 545, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 621 + ], + "type": "text", + "content": "[42] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "text", + "content": "[43] Yuanrui Wang, Cong Han, YafeiLi, Zhipeng Jin, Xiawei Li, SiNan Du, Wen Tao, Yi Yang, shuanglong li, Chun Yuan, and Liu Lin. Uniglyph: Unified segmentation-conditioned diffusion for precise visual text synthesis, 2025. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[44] Hongyang Wei, Shuaizheng Liu, Chun Yuan, and Lei Zhang. Perceive, understand and restore: Real-world image superresolution with autoregressive multimodal generative models, 2025." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "21297" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[45] Yongxian Wei, Runxi Cheng, Weike Jin, Enneng Yang, Li Shen, Lu Hou, Sinan Du, Chun Yuan, Xiaochun Cao, and Dacheng Tao. Unifying multimodal large language model capabilities and modalities via model merging. arXiv preprint arXiv:2505.19892, 2025." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 295, + 184 + ], + "type": "text", + "content": "[46] Yongxian Wei, Zixuan Hu, Li Shen, Zhenyi Wang, Chun Yuan, and Dacheng Tao. Open-vocabulary customization from clip via data-free knowledge distillation. In The Thirteenth International Conference on Learning Representations, 2025. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 186, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 186, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 57, + 186, + 295, + 218 + ], + "type": "text", + "content": "[47] Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muenighoff. C-pack: Packaged resources to advance general chinese embedding, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 220, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 294, + 384 + ], + "type": "text", + "content": "[48] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhang Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zhihao Fan. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 387, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 387, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 57, + 387, + 294, + 441 + ], + "type": "text", + "content": "[49] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. Detclip: Dictionary-enriched visual-concept paralleled pretraining for open-world detection. Advances in Neural Information Processing Systems, 35:9125-9138, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 444, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 444, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 444, + 294, + 487 + ], + "type": "text", + "content": "[50] Qihang Yu, Huiyu Wang, Siyuan Qiao, Maxwell Collins, Yukun Zhu, Hartwig Adam, Alan Yuille, and Liang-Chieh Chen. k-means mask transformer. In European Conference on Computer Vision, pages 288-307. Springer, 2022. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 489, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 489, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 57, + 489, + 295, + 544 + ], + "type": "text", + "content": "[51] Yuqian Yuan, Wentong Li, Jian Liu, Dongqi Tang, Xinjie Luo, Chi Qin, Lei Zhang, and Jianke Zhu. Osprey: Pixel understanding with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28202-28211, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 545, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 294, + 589 + ], + "type": "text", + "content": "[52] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 3, 4, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 591, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 294, + 656 + ], + "type": "text", + "content": "[53] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. Advances in Neural Information Processing Systems, 35:36067-36080, 2022. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "text", + "content": "[54] Hao Zhang, Feng Li, Xueyan Zou, Shilong Liu, Chunyuan Li, Jianwei Yang, and Lei Zhang. A simple framework for open-vocabulary segmentation and detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1020-1031, 2023. 7, 8" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 373 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "text", + "content": "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "text", + "content": "[56] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European conference on computer vision, pages 350-368. Springer, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "text", + "content": "[57] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 252 + ], + "type": "text", + "content": "[58] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 253, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 317 + ], + "type": "text", + "content": "[59] Xueyan Zou, Zi-Yi Dou, Jianwei Yang, Zhe Gan, Linjie Li, Chunyuan Li, Xiyang Dai, Harkirat Behl, Jianfeng Wang, Lu Yuan, et al. Generalized decoding for pixel, image, and language. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15116-15127, 2023. 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "text", + "content": "[60] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in neural information processing systems, 36:19769-19782, 2023. 2, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 756 + ], + "type": "text", + "content": "21298" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/a12eb351-f679-4d5e-8b40-f5c8f124976d_content_list.json b/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/a12eb351-f679-4d5e-8b40-f5c8f124976d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2026e1a7431e2d49c8434959536aa7435c73a2b7 --- /dev/null +++ b/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/a12eb351-f679-4d5e-8b40-f5c8f124976d_content_list.json @@ -0,0 +1,1767 @@ +[ + { + "type": "text", + "text": "Text-to-Any-Skeleton Motion Generation Without Retargeting", + "text_level": 1, + "bbox": [ + 183, + 130, + 815, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qingyuan Liu $^{1}$ Ke Lu $^{1,2}$ Kun Dong $^{1}$ Jian Xue $^{1*}$ Zehai Niu $^{1}$ Jinbao Wang $^{3}$ $^{1}$ University of Chinese Academy of Sciences $^{2}$ Peng Cheng Laboratory $^{3}$ Shenzhen University \n{liuqingyuan23, dongkun22, niuzehai18}@mails.ucas.ac.cn, {luk, xuejian}@ucas.ac.cn \nwangjb@szu.edu.cn", + "bbox": [ + 127, + 179, + 869, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0cd5482add302f1fb2103c2541798d95e98ddc2b6570060363de3a01e9babddb.jpg", + "image_caption": [ + "Figure 1. Introducing OmniSkel: A text-driven motion generation framework that directly generates motions for any skeletons without retargeting. Unlike traditional two-stage pipelines (top right) that require complex retargeting procedures, our single-stage approach (bottom right) only needs text input and target bone lengths (which are easily obtainable) to generate desired motions. As shown on the left, OmniSkel creates natural movements like \"pushed but did not fall\" and \"walks up stairs with right hand on handrail\" for characters with different body proportions while maintaining action fidelity." + ], + "image_footnote": [], + "bbox": [ + 98, + 287, + 544, + 473 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/035bd83486ffd72d587dabd95b58552474e2a4a2f0673be8c20efcde3b6495f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 287, + 890, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 571, + 326, + 587 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in text-driven motion generation have shown notable progress. However, these methods are typically limited to standardized skeletons and rely on a cumbersome targeting process to adapt to varying skeletal configurations of diverse characters. In this paper, we present OmniSkel, a novel framework that directly generates high-quality human motions for any user-defined skeleton without retargeting. Specifically, we introduce a skeleton-aware RVQ-VAE, which utilizes Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into motion encoding and reconstruction. Moreover, we propose a simple yet effective training-free approach, Motion Restoration Optimizer (MRO), to ensure zero bone length error while preserving motion smoothness. To support this research, we construct SkeleMotion3D, a large-scale text-skeleton-motion dataset based on HumanML3D. Extensive experiments demonstrate the excel", + "bbox": [ + 88, + 607, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lent robustness and generalization of our method.", + "bbox": [ + 513, + 573, + 841, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 617, + 643, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed remarkable progress in text-to-motion (T2M) generation, showing promising potential in applications such as gaming, metaverse, and virtual/augmented reality [10, 16, 32, 41, 50, 51]. This technology enables the intuitive creation of 3D character animations through natural language descriptions, significantly simplifying content creation workflows in animation [20], virtual reality [16], video games [30, 47], and robotics [5, 22].", + "bbox": [ + 509, + 643, + 906, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these advancements, a fundamental limitation restricts the practical application of current text-to-motion generation methods. As illustrated in Figure 1, existing approaches primarily focus on enhancing the motion quality and text-motion alignment, neglecting variations in skeletons and only generating human motions for standardized skeletons. However, real-world production environments routinely require varying skeletal configurations for diverse", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12926", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "characters. To this end, practitioners must resort to skeleton retargeting - a technically demanding process that necessitates significant expertise and frequently fails to produce usable results, inevitably leading to laborious tuning and production delays. Therefore, there is a pressing demand for straightforward motion generation given arbitrary skeletons.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we present OmniSkel, a novel framework that can directly generate high-quality human motion sequences for arbitrary character skeletons. As shown in Figure 1, our approach eliminates the complex retargeting stage required by traditional pipelines, offering a single-stage solution that directly produces character-specific motion without specialized expertise. Specifically, we propose a Skeleton-aware RVQ-VAE (SR-VAE) architecture, which utilizes Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into the motion encoding and reconstruction. Unlike previous approaches, our design effectively separates skeleton-invariant features, which is crucial for arbitrary skeleton adaptation. To ensure zero bone length error while preserving motion smoothness, we further propose a training-free approach, Motion Restoration Optimizer (MRO). With the above designs, our OmniSkel can deliver high-quality motion generation precisely in line with textual descriptions and character skeletons, providing robust support for practical text-to-motion applications.", + "bbox": [ + 91, + 199, + 482, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To facilitate our research, we construct SkeleMotion-3D, a large-scale dataset based on HumanML3D, pairing textual descriptions with motion sequences across diverse skeletal configurations. This dataset establishes a strong foundation for motion generation with arbitrary skeletons.", + "bbox": [ + 89, + 503, + 482, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Before delving into the details, we summarize our core contributions in this work as follows:", + "bbox": [ + 89, + 580, + 482, + 611 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we present the first text-to-motion generation framework capable of directly producing motions for arbitrary character skeletons without retargeting, significantly simplifying creation workflows in real-world production environments.", + "- We propose a novel Skeleton-aware RVQ-VAE (SR-VAE) architecture, which employs Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into the motion encoding and reconstruction. Moreover, we further introduce Motion Reconstruction Optimizer (MRO) during inference, ensuring zero bone length error while preserving motion smoothness.", + "- Based on HumanML3D, we construct SkeleMotion-3D, a large-scale text-skeleton-motion dataset. This establishes a foundation for arbitrary skeleton motion generation, facilitating future research in this field.", + "- To verify the effectiveness of our method, we conduct extensive experiments on challenging datasets. Results demonstrate the superior robustness and generalization of" + ], + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "our method. Further analysis reveals the contribution of each component to the performance improvement.", + "bbox": [ + 526, + 90, + 906, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 136, + 653, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-Motion Generation", + "text_level": 1, + "bbox": [ + 513, + 162, + 756, + 176 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-motion generation has evolved significantly as a vital research area in computer animation and human-computer interaction.", + "bbox": [ + 511, + 184, + 903, + 228 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early Approaches. Initial methods established text-to-motion mappings through alignment techniques [3, 4, 9, 26, 36, 46], but produced motions that lacked naturalism and diversity due to their deterministic nature.", + "bbox": [ + 511, + 229, + 903, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Probabilistic Frameworks. VAE-based approaches [6, 16, 28, 31, 32] model distributions of possible motions, while vision-language alignment methods [27, 33, 37, 40] leveraged pre-trained models for better semantic consistency.", + "bbox": [ + 511, + 291, + 903, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent Advances. Contemporary approaches include diffusion-based models [7, 10, 12, 21, 38, 41, 51, 52] generating high-fidelity motions through progressive denoising. Vector Quantized VAE (VQ-VAE) approaches treat motions as discrete token sequences, including autoregressive models [19, 45, 50, 54, 55], bidirectional methods [18, 35], and hybrid approaches [34]. Part-based generation methods [13, 54, 55] provide control over individual body components.", + "bbox": [ + 511, + 352, + 903, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Limitations. Existing approaches remain constrained to standard skeletal structures, creating a gap between research and industry needs where adaptability to various character rigs is essential.", + "bbox": [ + 511, + 488, + 903, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work addresses this limitation by developing a skeleton-agnostic framework that generates semantically consistent motions for arbitrary skeletal structures.", + "bbox": [ + 511, + 549, + 903, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Motion Retargeting", + "text_level": 1, + "bbox": [ + 513, + 604, + 700, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motion retargeting adapts existing motion data between characters with different proportions and structures. Historically, Gleicher et al. [14] pioneered this field through spacetime optimization with kinematic constraints. Following this work, Lee and Shin [23] applied inverse kinematics with B-spline curve fitting, while Choi and Ko [11] simultaneously developed an online method preserving high-frequency details. Tak and Ko [39] later enhanced these approaches by utilizing dynamics constraints for physical plausibility.", + "bbox": [ + 511, + 628, + 906, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the advent of deep learning, new approaches emerged, including Villegas et al. [44] with unsupervised retargeting via recurrent networks, Lim et al. [25] disentangling pose and movement, and Aberman et al. [1] designing skeleton-aware networks for cross-structural retargeting. More recently, research has shifted towards preserving motion semantics. For instance, Zhang et al. [49] developed a residual retargeting network with skeleton-aware", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12927", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bcf229b8eb283236c8777696ce593b163ed5de0c0c167ae0b375369dde5c54a6.jpg", + "image_caption": [ + "(a) Skeleton-aware Residual VQ-VAE", + "(b) Masked Transformer", + "(c) Residual Transformer", + "Figure 2. Method overview. (a) Skeleton-aware Residual VQ-VAE: our framework processes source motion and skeleton through K-GCA (Kinematic Graph Cross Attention), which divides the input into Q, K, and V branches, creating a hierarchical representation with motion-invariant features. The residual quantization approach progressively captures motion details across multiple layers, enabling high-fidelity reconstruction for arbitrary target skeletons. (b) Masked Transformer: base-layer motion tokens $t^0$ are randomly masked during training and the text-conditioned transformer learns to predict these tokens simultaneously, capturing fundamental motion patterns. (c) Residual Transformer: builds upon the base layer by progressively predicting higher-level residual tokens $t^v$ conditioned on both text and previous layer tokens, enhancing motion fidelity and skeleton-specific details." + ], + "image_footnote": [], + "bbox": [ + 98, + 107, + 883, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and shape-aware modules. Similarly, Lee et al. [24] proposed Skeleton-Agnostic Motion Embedding (SAME) using graph convolutional networks to separate skeleton information from motion while preserving semantics. Building upon these semantic approaches, Zhang et al. [48] introduced Semantics-aware Motion reTargeting (SMT), leveraging vision-language models to extract and maintain motion semantics through a two-stage pipeline with skeleton-aware pre-training and semantic constraints.", + "bbox": [ + 88, + 484, + 482, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, despite this significant progress, many approaches still struggle with maintaining semantic consistency across dramatically different skeletal structures.", + "bbox": [ + 89, + 619, + 483, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 89, + 679, + 199, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to generate motion sequences $\\mathbf{M}_{1:T}$ of length $T$ conditioned on both a text description $c$ and a target skeletal structure $\\mathbf{S}$ , where $\\mathbf{M}_t\\in \\mathbb{R}^{J\\times D_{\\mathrm{m}}}$ with $J$ denoting the number of joints and $D_{\\mathrm{m}}$ representing the dimension of joint features. The skeletal structure $\\mathbf{S}\\in \\mathbb{R}^{J\\times D_{\\mathrm{s}}}$ characterizes the static joint properties of the target character. As illustrated in Figure 2, our framework consists of two main components: a skeleton-aware VAE module that enables motion generation for arbitrary skeletal structures (Section 3.2), and a cooperative transformer architecture for motion token prediction (Section 3.3). We first introduce our newly proposed dataset and its representation (Section 3.1) to facilitate better understanding of our method. Then, we detail", + "bbox": [ + 89, + 703, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the complete training pipeline (Section 3.3) and inference process (Section 3.4).", + "bbox": [ + 511, + 484, + 906, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. SkeleMotion-3D Dataset", + "text_level": 1, + "bbox": [ + 511, + 522, + 736, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our dataset comprises three key components: textual descriptions, motion sequences, and corresponding skeletal structures. Building upon the HumanML3D dataset [16], which combines motion data from AMASS [29] and HumanAct12 [15], we significantly expand the motion-skeleton pairs through a skeleton randomization process inspired by [24].", + "bbox": [ + 511, + 544, + 906, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature Representation. We represent motion sequences as $\\mathbf{M} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}$ , where the motion feature $F_{\\mathrm{motion}}$ with dimension $D_{\\mathrm{m}}$ consists of:", + "bbox": [ + 511, + 651, + 905, + 695 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {m o t i o n}} = \\left[ F _ {\\text {t o p o}}, F _ {\\text {r o o t}}, F _ {\\text {x y z}}, F _ {\\text {r o t}}, F _ {\\text {v e l}} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 704, + 903, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "comprising skeletal topology features, root joint features, relative spatial positions, 6D rotation features, and velocity features.", + "bbox": [ + 511, + 729, + 905, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The skeletal structures are encoded as $\\mathbf{S} \\in \\mathbb{R}^{J \\times D_{\\mathrm{s}}}$ , where the skeletal feature $F_{\\mathrm{skel}}$ with dimension $D_{\\mathrm{s}}$ consists of:", + "bbox": [ + 511, + 773, + 905, + 816 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\mathrm {s k e l}} = \\left[ F _ {\\mathrm {t o p o}}, F _ {\\mathrm {s t a t i c}} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 819, + 903, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{\\mathrm{topo}}$ represents topology features and $F_{\\mathrm{static}}$ denotes static skeletal properties.", + "bbox": [ + 511, + 840, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset Construction. Following the preprocessing pipeline established by HumanML3D, we standardize the", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12928", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "raw motion sequences through temporal normalization (20 FPS), sequence cropping (maximum 10 seconds), skeletal retargeting, and orientation alignment $(\\mathrm{Z}+$ direction). The text annotations from HumanML3D, collected via Amazon Mechanical Turk with quality control measures, are preserved in our dataset.", + "bbox": [ + 89, + 90, + 480, + 180 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset Statistics. The SkeleMotion-3D dataset contains 14,616 distinct motion patterns, 29,233 different skeletal configurations, 43,848 unique text-motion-skeleton triplets, and 44,970 textual descriptions with 5,371 unique words. This represents a substantial increase compared to the training dataset (949 motion patterns and 4,745 motion-skeleton pairs) used in [24]'s motion retargeting network. The significant expansion enhances the robustness and generalization capability of our skeleton-aware motion generation model.", + "bbox": [ + 89, + 181, + 482, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Skeleton-aware Residual VQ-VAE", + "text_level": 1, + "bbox": [ + 89, + 327, + 390, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose Skeleton-aware Residual VQ-VAE (SR-VAE), a novel architecture that extends the conventional Residual VQ-VAE to enable skeleton-conditioned motion generation. Conventional motion VQ-VAEs [17, 19, 50, 53] and recent RVQ-VAEs [18, 45] focus on reconstructing source motions with identical skeletal structures. In contrast, as illustrated in Figure 2, SR-VAE adopts an asymmetric VAE architecture, consisting of an encoder, a decoder, a residual quantization module, and a Kinetic Graph Cross Attention (K-GCA) module, enabling motion reconstruction for arbitrary target skeletons.", + "bbox": [ + 89, + 348, + 482, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture Design. Our SR-VAE processes motion generation through three main stages. First, the input motion sequence $\\mathbf{M}_{\\mathrm{in}} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}$ and its corresponding skeleton structure $\\mathbf{S}_{\\mathrm{in}} \\in \\mathbb{R}^{J \\times D_{\\mathrm{s}}}$ are fused via K-GCA before being encoded into latent features. Following MoMask [18], these features undergo residual quantization to produce $V + 1$ ordered code sequences. Finally, the decoder incorporates target skeleton features $\\mathbf{S}_{\\mathrm{tgt}} \\in \\mathbb{R}^{J \\times D_{\\mathrm{s}}}$ through multiple K-GCA layers at different decoding stages to generate the target motion $\\mathbf{M}_{\\mathrm{tgt}} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}$ .", + "bbox": [ + 89, + 515, + 482, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\hat{\\mathbf{M}}_{\\mathrm{tgt}} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}$ denote the predicted motion sequence and $\\mathbf{M}_{\\mathrm{tgt}}$ denote the ground truth target motion. For residual quantization, let $\\mathbf{r}^v \\in \\mathbb{R}^{T \\times d}$ represent the residual features at layer $v \\in \\{0,1,\\dots,V\\}$ and $\\mathbf{b}^v \\in \\mathbb{R}^{T \\times d}$ denote the corresponding quantized codes, where $d$ is the dimension of the latent representation. Additionally, let $\\mathbf{D}_{\\mathrm{xyz}} \\in \\mathbb{R}^{T \\times J \\times 3}$ and $\\hat{\\mathbf{D}}_{\\mathrm{xyz}} \\in \\mathbb{R}^{T \\times J \\times 3}$ represent the ground truth and predicted joint positions in 3D space, respectively. The overall training objective consists of three terms:", + "bbox": [ + 89, + 667, + 482, + 803 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\left\\| \\mathbf {M} _ {\\mathrm {t g t}} - \\hat {\\mathbf {M}} _ {\\mathrm {t g t}} \\right\\| _ {1} + \\beta \\sum_ {v = 0} ^ {V} \\left\\| \\mathbf {r} ^ {v} - \\operatorname {s g} \\left[ \\mathbf {b} ^ {v} \\right] \\right\\| _ {2} ^ {2} + \\lambda \\left\\| \\mathbf {D} _ {\\mathrm {x y z}} - \\hat {\\mathbf {D}} _ {\\mathrm {x y z}} \\right\\| _ {1} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 809, + 480, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the first term is motion reconstruction loss, the second term is commitment loss for residual quantization, and the third term specifically focuses on joint positions accu", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/548a422267deed8ed1c2be045bca439657900ab4efadc69eb8ac7262e605e982.jpg", + "image_caption": [ + "Figure 3. Architecture of our Kinetic Graph Cross Attention." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "racy. Here, $\\mathrm{sg}[\\cdot ]$ denotes the stop-gradient operation, $\\beta$ and $\\lambda$ are weighting factors for the commitment loss and joint position loss, respectively. The loss function is optimized via a straight-through gradient estimator [42]. Following [18, 35, 50], we employ exponential moving average for codebook updates and reset.", + "bbox": [ + 511, + 234, + 903, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Kinetic Graph Cross Attention. To effectively fuse motion and skeletal features, we propose Kinetic Graph Cross Attention (K-GCA), as illustrated in Figure 3, extending the traditional Graph Attention Networks (GAT) [8, 43] for motion-skeleton feature fusion. Unlike standard GAT that performs self-attention within a single feature type, K-GCA enables cross-modal attention between motion and skeletal features while preserving the skeletal graph structure.", + "bbox": [ + 511, + 325, + 905, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given motion features $\\mathbf{X} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}$ and skeleton features $\\mathbf{S} \\in \\mathbb{R}^{J \\times D_{\\mathrm{s}}}$ , K-GCA first projects them into a shared space using learnable projection matrices $\\mathbf{W}_{\\mathrm{m}} \\in \\mathbb{R}^{D_{\\mathrm{m}} \\times d_k}$ and $\\mathbf{W}_{\\mathrm{s}} \\in \\mathbb{R}^{D_{\\mathrm{s}} \\times d_k}$ , where $d_k$ is the dimension of the projected space:", + "bbox": [ + 511, + 445, + 903, + 521 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {W} _ {\\mathrm {m}} \\mathbf {X}, \\quad \\mathbf {K} = \\mathbf {V} = \\mathbf {W} _ {\\mathrm {s}} \\mathbf {S}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 532, + 903, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The attention scores between joints $j$ and $k$ are computed using learnable vectors $\\mathbf{a}^{\\mathrm{src}}\\in \\mathbb{R}^{d_k}$ and $\\mathbf{a}^{\\mathrm{dst}}\\in \\mathbb{R}^{d_k}$ :", + "bbox": [ + 511, + 561, + 903, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {j k} = \\text {L e a k y R e L U} \\left(\\mathbf {a} ^ {\\mathrm {s r c}} \\cdot \\mathbf {Q} _ {j} + \\mathbf {a} ^ {\\mathrm {d s t}} \\cdot \\mathbf {K} _ {k}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 603, + 903, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To preserve the skeletal structure, we mask the attention scores using the skeleton adjacency matrix $\\mathbf{A} \\in \\{0,1\\}^{J \\times J}$ , where $\\mathbf{A}_{jk} = 1$ if joints $j$ and $k$ are connected in the skeleton graph, and 0 otherwise. We set attention scores for disconnected joints to negative infinity before applying softmax:", + "bbox": [ + 511, + 633, + 905, + 720 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {e} _ {j k} = \\left\\{ \\begin{array}{l l} e _ {j k}, & \\text {i f} \\mathbf {A} _ {j k} = 1 \\\\ - \\infty , & \\text {i f} \\mathbf {A} _ {j k} = 0 \\end{array} \\right., \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 720, + 903, + 760 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {j k} = \\frac {\\exp \\left(\\tilde {e} _ {j k}\\right)}{\\sum_ {l = 1} ^ {J} \\exp \\left(\\tilde {e} _ {j l}\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 768, + 903, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final output $\\hat{\\mathbf{X}}\\in \\mathbb{R}^{T\\times J\\times D_{\\mathrm{m}}}$ is computed by applying the attention weights to value matrix $\\mathbf{V}$ in each head, concatenating results from all heads, and adding back to the input through residual connection:", + "bbox": [ + 511, + 810, + 903, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {X}} = \\operatorname {L a y e r N o r m} (\\mathbf {X} + \\operatorname {M u l t i H e a d} (\\alpha , \\mathbf {V})). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 883, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12929", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where MultiHead $(\\alpha, \\mathbf{V})$ represents the multi-head weighted combination of values using the computed attention weights $\\alpha$ . This design allows K-GCA to capture dynamic relationships between motion and skeletal features while maintaining the anatomical constraints imposed by the skeletal structure. Compared to standard GAT, our approach provides two key advantages: (1) the ability to perform cross-modal attention between motion and skeletal features, and (2) the incorporation of skeletal structure constraints through attention masking.", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Pipeline", + "text_level": 1, + "bbox": [ + 89, + 252, + 261, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our training pipeline follows a sequential approach for skeleton-agnostic motion generation from text, as shown in Figure 2.", + "bbox": [ + 89, + 275, + 483, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SR-VAE Training. First, we train the Skeleton-aware Residual VQ-VAE (SR-VAE) to establish skeleton-agnostic motion representation. This model encodes motion sequences into hierarchical tokens and decodes them back to motion features, independent of skeletal structure. After training, both encoder $\\mathcal{E}$ and decoder $\\mathcal{D}$ are frozen.", + "bbox": [ + 89, + 321, + 483, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tokenization of Training Data. Using the frozen SR-VAE encoder, we convert all training motion sequences into hierarchical token representations. For each motion $\\mathbf{M}_{\\mathrm{in}}$ with corresponding skeleton $\\mathbf{S}_{\\mathrm{in}}$ , we extract multi-level token sequences $\\mathbf{t}^{0:V}$ as targets for our text-to-token models.", + "bbox": [ + 89, + 412, + 483, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Masked Transformer Training. Inspired by MoMask [18], we implement a bidirectional Masked Transformer to generate base-layer tokens $(\\mathbf{t}^0)$ . This design offers parallel generation capability and better global motion coherence compared to autoregressive approaches. During training, the model learns to predict original tokens at masked positions from text prompts.", + "bbox": [ + 89, + 488, + 483, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Residual Transformer Training. Finally, we train the Residual Transformer to predict higher-layer tokens $(\\mathbf{t}^{1:V})$ encoding motion refinements. Following MoMask [18], this transformer takes both text descriptions and tokens from previous layers as input, focusing on finer motion details while maintaining consistency with the base motion.", + "bbox": [ + 89, + 594, + 483, + 684 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Inference Process", + "text_level": 1, + "bbox": [ + 89, + 695, + 261, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, OmniSkel generates skeleton-aware motions through the following steps:", + "bbox": [ + 89, + 718, + 482, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Base Token Generation: Given a text prompt, the Masked Transformer generates the base-layer tokens $\\hat{\\mathbf{t}}^0$ that capture primary motion patterns.", + "bbox": [ + 89, + 748, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hierarchical Refinement: The Residual Transformer progressively generates higher-layer tokens $\\hat{\\mathbf{t}}^{1:V}$ , starting with layer 1 and proceeding sequentially. Each layer's prediction utilizes both the text prompt and all previously generated token layers.", + "bbox": [ + 89, + 794, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion Decoding: Once we have the complete token hierarchy $\\hat{\\mathbf{t}}^{0:V}$ , we feed these tokens along with the target skele", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ton structure $\\mathbf{S}_{\\mathrm{tgt}}$ into the frozen SR-VAE decoder to obtain motion features.", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1: Motion Restoration Optimizer" + ], + "code_body": "Input: Measurement data: $F_{\\mathrm{topo}}$ , $F_{\\mathrm{root}}$ , $F_{\\mathrm{xyz}}$ ; Bone lengths: $S$ \nOutput: Smooth motion sequence $M_{\\mathrm{smooth}}$ with guaranteed skeletal consistency", + "guess_lang": "txt", + "bbox": [ + 526, + 167, + 866, + 229 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "1 // Step 1: Restore initial motion using position-based approach;" + ], + "code_body": "2 $M_{\\mathrm{ric}} \\gets \\mathrm{ric}(F_{\\mathrm{topo}}, F_{\\mathrm{root}}, F_{\\mathrm{xyz}})$ ;", + "guess_lang": "txt", + "bbox": [ + 519, + 265, + 728, + 280 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "3 $M_{\\mathrm{smooth}} \\gets \\emptyset$", + "guess_lang": "latex", + "bbox": [ + 519, + 287, + 629, + 300 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "4 foreach frame $f\\in M_{\\mathrm{ric}}$ do" + ], + "code_body": "// Step 2: Extract unit direction vectors for each bone (except the root); \nInitialize set $\\mathcal{U} \\gets \\emptyset$ ; \nfor $i \\gets 1$ to $J - 1$ do \nLet $\\mathbf{p}_i$ be the position of joint $i$ in frame $f$ ; \nLet $\\mathbf{p}_{\\text{parent}(i)}$ be the position of joint $i$ 's parent; \nCompute displacement: $\\mathbf{v}_i \\gets \\mathbf{p}_i - \\mathbf{p}_{\\text{parent}(i)}$ ; \nCompute unit direction: $\\mathbf{u}_i \\gets \\frac{\\mathbf{v}_i}{|\\mathbf{v}_i|}$ ; \nAdd $\\mathbf{u}_i$ to $\\mathcal{U}$ ; \nend \n// Step 3: Reconstruct the current frame using extracted directions, bone lengths, and topology; \n $f_{\\text{rec}} \\gets \\text{ReconstructFrame}(\\mathcal{U}, S, F_{\\text{topo}}, F_{\\text{root}})$ ; \nAdd $f_{\\text{rec}}$ to $M_{\\text{smooth}}$ ; \nend \nreturn $M_{\\text{smooth}}$ ;", + "guess_lang": "txt", + "bbox": [ + 517, + 321, + 880, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature-to-Motion Conversion: Following the generation of motion features by our SR-VAE decoder, we face the challenge of transforming these abstract representations into production-ready motion data. Contemporary text-to-motion frameworks typically rely on either position-based restoration (offering fluidity but with bone length inconsistencies) or rotation-based restoration (maintaining skeletal integrity but introducing jitter and penetration artifacts).", + "bbox": [ + 511, + 667, + 906, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To overcome these limitations, we employ our Motion Restoration Optimizer (MRO, detailed in Algorithm 1), which synergistically combines the strengths of both approaches. MRO guarantees two critical properties simultaneously: absolute preservation of skeletal proportions (zero bone length error) and maintenance of smooth trajectory characteristics, producing physically plausible output $\\hat{\\mathbf{M}}_{\\mathrm{tgt}}$ .", + "bbox": [ + 511, + 795, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12930", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4a7cc8c474e8b6a9e599bcabee8587733a8087ec8699fa1f08004712cb143857.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID↓R Precision↑MPJPE↓Skeleton Error↓Jitter Error↓Skating Ratio ↓
Top 1Top 2Top 3
SR-VAE (full model)0.0270.5080.7000.7930.03513.7781.8240.074
w/ Single-head K-GCA0.0400.4970.6850.7890.03412.2942.9350.061
w/o K-GCA0.0650.4950.6890.7850.04117.7269.5940.092
w/o RVQ0.1650.4670.6390.7750.06119.7263.5940.092
w/ Symmetric Reconstruction0.8850.4500.6410.7450.09789.7479.8670.067
w/ Rotation-based Restoration0.0270.5080.7000.7930.0630.0030.7240.193
w/ MRO0.0270.5080.7000.7930.0350.002.0840.090
", + "bbox": [ + 94, + 88, + 919, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Ablation study of SR-VAE components tested on the SkeleMotion-3D test set. Our SR-VAE is trained on the SkeleMotion-3D dataset. To evaluate SR-VAE's reconstruction performance for arbitrary skeletal motions, we input random skeleton motions to the encoder and use the standard skeleton from the HumanML3D dataset as the target skeleton.", + "bbox": [ + 89, + 255, + 906, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 323, + 225, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive experiments to evaluate our approach across multiple dimensions, focusing on both semantic accuracy and physical plausibility.", + "bbox": [ + 89, + 349, + 482, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We assess our model using established semantic metrics from prior work [16]: (1) Frechet Inception Distance (FID) for overall motion quality; (2) $R$ -Precision and multimodal distance for text-motion alignment.", + "bbox": [ + 89, + 396, + 483, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To comprehensively evaluate physical plausibility—a critical aspect for animation applications—we introduce several specialized metrics:", + "bbox": [ + 89, + 473, + 483, + 518 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Skating Ratio: Measures foot sliding artifacts where feet unnaturally glide across the ground", + "- Skeleton Error: Quantifies bone length variations throughout motion sequences", + "- MPJPE (Mean Per-Joint Position Error): Assesses positional accuracy of joints", + "- Jitter Error: Evaluates motion continuity and natural flow. Detailed formulations of these metrics appear in the supplementary materials." + ], + "bbox": [ + 89, + 518, + 482, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our PyTorch implementation features a Skeleton-aware Residual VQ-VAE (SR-VAE) with residual blocks following T2M-GPT [50] and MoMask [18] architectures. We use a downscale factor of 4, with a residual quantization comprising 6 layers (each containing 512 codes of 512 dimensions) and quantization dropout ratio of 0.2. For the SR-VAE training objective, we set commitment loss weight $\\beta$ to 0.02 and joint position loss weight $\\lambda$ to 2.0. The K-GCA module employs 2-head attention.", + "bbox": [ + 89, + 657, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Both transformer models consist of 6 layers with 6 attention heads and 384-dimensional latent representations, trained on our SkeleMotion-3D dataset. We use a learning rate of 2e-4 with linear warm-up over 2000 iterations, batch sizes of 512 for SR-VAE and 64 for transformers. During inference, we apply classifier-free guidance (scale 4 for", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "masked transformer, 5 for refinement transformer) with 10 iterative decoding steps. All experiments run on NVIDIA RTX 4090 GPUs.", + "bbox": [ + 511, + 324, + 906, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Skeleton-Aware Disentanglement Analysis", + "text_level": 1, + "bbox": [ + 511, + 378, + 872, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A cornerstone of our approach is SR-VAE's ability to disentangle motion patterns from skeletal structures. We conduct a targeted experiment to validate this skeleton-agnostic capability.", + "bbox": [ + 511, + 401, + 906, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 4 presents our analysis of four distinct motion patterns, each performed by three different skeletal structures with varying proportions. We process these motion-skeleton pairs through our SR-VAE encoder $\\mathcal{E}$ and visualize the resulting base layer tokens $\\mathbf{t}^0$ on aligned line charts.", + "bbox": [ + 511, + 462, + 905, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results reveal that motions with identical semantic patterns but executed by anatomically diverse skeletons produce remarkably similar token sequences. This consistency demonstrates that our encoder $\\mathcal{E}$ successfully extracts skeleton-invariant motion features, preserving essential semantic content regardless of the executing skeletal structure.", + "bbox": [ + 511, + 537, + 905, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This confirms that SR-VAE effectively decomposes motion into two orthogonal components: (1) skeleton-invariant pattern information encoded in tokens, and (2) skeleton-specific structural information provided separately to the decoder. This disentanglement enables our model to generate physically plausible animations for arbitrary skeletal structures while preserving intended motion semantics.", + "bbox": [ + 511, + 643, + 905, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 757, + 666, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate SR-VAE's skeleton-aware capabilities, we conduct experiments using random input skeletons while decoding to standard HumanML3D [16] skeletons. This evaluates both semantic alignment and cross-skeleton reconstruction capabilities.", + "bbox": [ + 511, + 779, + 905, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents our ablation results on the SkeleMotion-3D test set. The complete SR-VAE model achieves superior performance with an FID of 0.026 and strong R", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12931", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b812883ec6b3350c3f600d7743fa6720c62b8839b3af77d3b9c2eddcbeb9abd9.jpg", + "image_caption": [ + "Figure 4. Motion token analysis across skeletal structures. This figure demonstrates token sequences generated by the SR-VAE encoder for four distinct motion patterns, each performed by three skeletal structures with different proportions. The consistent token patterns across each motion type (columns) despite skeletal variations demonstrates our encoder's ability to extract skeleton-invariant motion features. Top row shows skeletal visualizations while bottom row displays the corresponding quantized token values, revealing clear pattern preservation across different anatomical configurations." + ], + "image_footnote": [], + "bbox": [ + 94, + 98, + 903, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a793afe065b0353315d59d1958cec69c529528111794b872f152209c4e887f07.jpg", + "image_caption": [ + "Figure 5. Text-to-motion generation across different skeletal structures. Our OmniSkel framework generates diverse motions responding to different text prompts (columns) while adapting to different skeletal anatomies (rows). The human skeleton on the left is randomly selected from Mixamo [2], demonstrating OmniSkel's ability to maintain semantic consistency across dramatically different skeletal structures without retargeting." + ], + "image_footnote": [], + "bbox": [ + 107, + 335, + 846, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Precision scores (0.508/0.700/0.793 for Top-1/2/3), demonstrating excellent semantic consistency.", + "bbox": [ + 89, + 690, + 482, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Architecture Components. Using symmetric reconstruction (same skeleton for encoder and decoder) drastically increases FID to 0.885 and skeleton error to 89.747, confirming that asymmetric processing is essential for effective cross-skeleton learning. Removing K-GCA degrades FID to 0.065 with a significant increase in jitter error (9.594 vs. 1.824), demonstrating its importance in capturing cross-skeleton geometric correspondences.", + "bbox": [ + 89, + 727, + 482, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without RVQ, FID increases substantially to 0.165 and MPJPE to 0.061, highlighting RVQ's crucial role in maintaining semantic consistency across different skeletal struc", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tures. The single-head K-GCA variant shows moderate performance decline (FID: 0.040), indicating that multi-head attention more effectively captures the rich geometric relationships between skeletons.", + "bbox": [ + 511, + 690, + 906, + 751 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motion Restoration. For motion restoration approaches, rotation-based methods achieve perfect skeleton reconstruction (skeleton error: 0.00) but significantly increase jitter error (30.724) and skating performance (0.193). In contrast, our MRO technique also achieves perfect skeleton reconstruction (skeleton error: 0.00) while maintaining low jitter error (2.084) and better skating ratio (0.090). This demonstrates MRO's effectiveness in preserving physical plausibility while maintaining motion quality across differ-", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12932", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ent skeletal structures.", + "bbox": [ + 89, + 90, + 240, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our ablation study confirms each component's critical contribution to the overall system: K-GCA and RVQ enhance cross-skeleton reconstruction and semantic alignment, while MRO ensures physical consistency in the generated motions. These findings validate our architectural design choices for skeleton-aware motion generation and demonstrate SR-VAE's ability to effectively handle arbitrary skeleton motion retargeting with high fidelity.", + "bbox": [ + 89, + 107, + 483, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Qualitative and Quantitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 239, + 416, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Evaluation. We evaluate our methods against state-of-the-art text-to-motion approaches as shown in Table 2. The fundamental distinction of our approach is the ability to directly generate motions for arbitrary skeleton structures—a paradigm shift from existing methods that are restricted to fixed, pre-defined skeletons.", + "bbox": [ + 89, + 263, + 483, + 354 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/25075906b47c32c24ee07742f8ee8efde32afd1c8e4dd9b30a65c9f3c1c2d7d4.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Comparison with text-to-motion methods on standard skeleton. Our approach introduces a novel pipeline trained on the SkeleMotion-3D dataset, which fundamentally differs from traditional T2M methods. For fair comparison with existing methods trained on HumanML3D, we configured our models (SR-VAE and OmniSkel) to target the standard HumanML3D skeleton during inference. Despite being designed for arbitrary skeleton generation, our methods achieve competitive performance on this fixed skeleton setting, demonstrating strong text-to-motion capabilities even when constrained to a single skeleton structure." + ], + "table_body": "
DatasetsMethodsFID↓R Precision↑
Top 1Top 2Top 3
Human-ML3DTM2T [17]1.501±.0170.424±.0030.618±.0030.729±.002
T2M [16]1.087±.0210.455±.0030.636±.0030.736±.002
MDM [41]0.544±.0440.320±.0050.498±.0040.611±.007
MLD [10]0.473±.0130.481±.0030.673±.0030.772±.002
MotionDiffuse [51]0.630±.0010.491±.0010.681±.0010.782±.001
T2M-GPT [50]0.141±.0050.492±.0030.679±.0020.775±.002
ReMoDiffuse [52]0.103±.0040.510±.0050.698±.0060.795±.004
AttT2M [54]0.112±.0060.499±.0030.690±.0020.786±.002
Momask [18]0.045±.0020.521±.0020.713±.0020.807±.002
Skele-Motion-3D OmniSkel (ours)0.027±.0000.508±.0020.700±.0020.793±.002
0.084±.0040.507±.0020.699±.0030.797±.002
", + "bbox": [ + 93, + 369, + 480, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This fundamental difference presents a challenge for direct comparison, as our methods operate on the newly introduced SkeleMotion-3D dataset with diverse skeleton structures, while previous approaches are confined to the HumanML3D dataset with a single skeleton configuration. Notably, existing methods cannot be trained or evaluated on our dataset due to their inherent architectural limitations in handling skeleton variations.", + "bbox": [ + 89, + 703, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To provide a meaningful comparison that highlights our text-to-motion capabilities, we configure our models at inference time to target the standard HumanML3D skeleton structure. Despite being designed for the more complex task of arbitrary skeleton generation, both SR-VAE", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and OmniSkel achieve competitive performance on the HumanML3D test set, with comparable or superior metrics in FID and R-Precision compared to specialized methods.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These results demonstrate that our methods successfully balance adaptability with generation quality, enabling cross-skeleton applications without compromising performance on standard benchmarks.", + "bbox": [ + 511, + 137, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. Figure 5 illustrates OmniSkel's capability to generate motions across different skeletal structures. The visualization demonstrates: (1) diverse motions for different textual descriptions, and (2) consistent semantic motion across distinct skeletal anatomies.", + "bbox": [ + 511, + 196, + 903, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Each column represents a different textual prompt, from complex sequences to simpler actions. The model successfully captures the nuanced details of these descriptions, including locomotion styles and sequential actions.", + "bbox": [ + 511, + 273, + 903, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This comparison demonstrates OmniSkel's precise perception and effective integration of different semantics and arbitrary skeletons. More visualization results are provided in the supplementary materials.", + "bbox": [ + 511, + 334, + 903, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Performance Comparison", + "text_level": 1, + "bbox": [ + 511, + 405, + 748, + 422 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c332ea08bd070095fbe679d531b3ca84211318d9a04017df1acbddcb30173e87.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParametersInference TimeGPU Memory
MoMask [18]44.85M0.075s1434M
Ours38.73M0.099s1402M
", + "bbox": [ + 517, + 440, + 901, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Computational cost comparison on RTX 4090.", + "bbox": [ + 544, + 494, + 872, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method achieves comparable performance to state-of-the-art approaches while maintaining slight GPU memory advantages as shown in Table 3 and uniquely enabling end-to-end motion generation for arbitrary skeletal specifications. This efficiency demonstrates the practical viability of our skeleton-agnostic framework for real-world applications requiring diverse character animations.", + "bbox": [ + 511, + 527, + 903, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 648, + 761, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, we present OmniSkel, a text-driven motion generation framework that enables motion synthesis for arbitrary human skeletons through our novel K-GCA and SR-VAE architectures. Our Motion Restoration Optimizer (MRO) addresses the cross-skeleton motion reconstruction challenge, ensuring both motion smoothness and skeletal consistency. Leveraging our SkeleMotion-3D dataset, which pairs textual descriptions with diverse skeletal configurations, our approach bridges the gap between research and practical applications. Experiments demonstrate that OmniSkel achieves competitive performance while maintaining semantic alignment with textual descriptions. This work establishes a foundation for more inclusive motion generation systems that can serve diverse applications without requiring conventional retargeting approaches.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12933", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grant 62320106007 and Grant 62032022; and in part by the Guangdong Provincial Key Laboratory under Grant 2023B1212060076.", + "bbox": [ + 89, + 114, + 483, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 203 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2", + "[2] Adobe. Mixamo. Online Animation Library, 2023. Available at: https://www MIXamo.com/.7", + "[3] Hyemin Ahn, Timothy Ha, Yunho Choi, Hwiyeon Yoo, and Songhwai Oh. Text2action: Generative adversarial synthesis from language to action. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5915-5920. IEEE, 2018. 2", + "[4] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719–728. IEEE, 2019. 2", + "[5] André Antakli, Erik Hermann, Ingo Zinnikus, Han Du, and Klaus Fischer. Intelligent distributed human motion simulation in human-robot collaboration environments. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 319–324, 2018. 1", + "[6] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 2", + "[7] Gustav Bredell, Kyriakos Flouris, Krishna Chaitanya, Ertunc Erdil, and Ender Konukoglu. Explicitly minimizing the blur error of variational autoencoders. arXiv preprint arXiv:2304.05939, 2023. 2", + "[8] Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks? In International Conference on Learning Representations, 2022. 4", + "[9] Haoye Cai, Chunyan Bai, Yu-Wing Tai, and Chi-Keung Tang. Deep video generation, prediction and completion of human action sequences. In Proceedings of the European Conference on Computer Vision (ECCV), pages 366–382, 2018. 2", + "[10] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 1, 2, 8", + "[11] Kwang-Jin Choi and Hyeong-Seok Ko. Online motion retargeting. The Journal of Visualization and Computer Animation, 11(5):223-235, 2000. 2", + "[12] Xuehao Gao, Yang Yang, Zhenyu Xie, Shaoyi Du, Zhongqian Sun, and Yang Wu. Guess: Gradually enriching synthesis for text-driven human motion generation. IEEE Transactions on Visualization and Computer Graphics, 2024. 2" + ], + "bbox": [ + 93, + 213, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1396-1406, 2021. 2", + "[14] Michael Gleicher. Retargetting motion to new characters. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, pages 33-42, 1998. 2", + "[15] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 3", + "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 1, 2, 3, 6, 8", + "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 4, 8", + "[18] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 2, 4, 5, 6, 8", + "[19] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2, 4", + "[20] Moritz Kappel, Vladislav Golyanik, Mohamed Elgharib, Jann-Ole Henningson, Hans-Peter Seidel, Susana Castillo, Christian Theobalt, and Marcus Magnor. High-fidelity neural human motion transfer from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1541-1550, 2021. 1", + "[21] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 2", + "[22] Hema Koppula and Ashutosh Saxena. Learning spatiotemporal structure from rgb-d videos for human activity detection and anticipation. In International conference on machine learning, pages 792-800. PMLR, 2013. 1", + "[23] Jehee Lee and Sung Yong Shin. A hierarchical approach to interactive motion editing for human-like figures. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 39-48, 1999. 2", + "[24] Sunmin Lee, Taeho Kang, Jungnam Park, Jehee Lee, and Jungdam Won. Same: Skeleton-agnostic motion embedding for character animation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3, 4", + "[25] Jongin Lim, Hyung Jin Chang, and Jin Young Choi. PMnet: learning of disentangled pose and movement for unsupervised motion retargeting. In Proc. BMVC, 2019. 2", + "[26] Angela S Lin, Lemeng Wu, Rodolfo Corona, Kevin Tai, Qixing Huang, and Raymond J Mooney. Generating animated" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12934", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "videos of human activities from natural language descriptions. Learning, 2018(1), 2018. 2", + "[27] Junfan Lin, Jianlong Chang, Lingbo Liu, Guanbin Li, Liang Lin, Qi Tian, and Chang-wen Chen. Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23222-23231, 2023. 2", + "[28] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In International Conference on Machine Learning, pages 32939-32977. PMLR, 2024. 2", + "[29] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5442-5451, 2019. 3", + "[30] Dennis Majoe, Lars Widmer, and Juerg Gutknecht. Enhanced motion interaction for multimedia applications. In Proceedings of the 7th International Conference on Advances in Mobile Computing and Multimedia, pages 13-19, 2009. 1", + "[31] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2", + "[32] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 1, 2", + "[33] Mathis Petrovich, Michael J Black, and Gül Varol. Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9488-9497, 2023. 2", + "[34] Ekkasit Pinyoanuntapong, Muhammad Usama Saleem, Pu Wang, Minwoo Lee, Srijan Das, and Chen Chen. Bamm: Bidirectional autoregressive motion model. arXiv preprint arXiv:2403.19435, 2024. 2", + "[35] Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 4", + "[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems, 109:13-26, 2018. 2", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[38] Haowen Sun, Ruikun Zheng, Haibin Huang, Chongyang Ma, Hui Huang, and Ruizhen Hu. Lgtm: Local-to-global text-driven human motion diffusion model. In ACM SIGGRAPH 2024 Conference Papers, pages 1-9, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Seyoon Tak and Hyeong-Seok Ko. A physically-based motion retargeting filter. ACM Trans. Graph., 24(1):98-117, 2005. 2", + "[40] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2", + "[41] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 1, 2, 8", + "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems, 30, 2017. 4", + "[43] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph Attention Networks. International Conference on Learning Representations, 2018. accepted as poster. 4", + "[44] Ruben Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic networks for unsupervised motion retargeting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8639-8648, 2018. 2", + "[45] Congyi Wang. T2m-hifigpt: generating high quality human motion from textual descriptions with residual discrete representations. arXiv preprint arXiv:2312.10628, 2023. 2, 4", + "[46] Zhenyi Wang, Ping Yu, Yang Zhao, Ruiyi Zhang, Yufan Zhou, Junsong Yuan, and Changyou Chen. Learning diverse stochastic human-action generators by learning smooth latent transitions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12281-12288, 2020. 2", + "[47] Mohammed Yeasin, Ediz Polat, and Rajeev Sharma. A multiobject tracking framework for interactive multimedia applications. IEEE transactions on multimedia, 6(3):398-405, 2004. 1", + "[48] Haodong Zhang, ZhiKe Chen, Haocheng Xu, Lei Hao, Xiaofei Wu, Songcen Xu, Zhensong Zhang, Yue Wang, and Rong Xiong. Semantics-aware motion retargeting with vision-language models. 2024. 3", + "[49] Jiaxu Zhang, Junwu Weng, Di Kang, Fang Zhao, Shaoli Huang, Xuefei Zhe, Linchao Bao, Ying Shan, Jue Wang, and Zhigang Tu. Skinned motion retargeting with residual perception of motion semantics & geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13864-13872, 2023. 2", + "[50] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 1, 2, 4, 6, 8", + "[51] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 1, 2, 8", + "[52] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116, 2023. 2, 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12935", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motionpt: Finetuned llms are general-purpose motion generators. arXiv preprint arXiv:2306.10900, 2023. 4", + "[54] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 509-519, 2023. 2, 8", + "[55] Qiran Zou, Shangyuan Yuan, Shian Du, Yu Wang, Chang Liu, Yi Xu, Jie Chen, and Xiangyang Ji. Parco: Part-coordinating text-to-motion synthesis. arXiv preprint arXiv:2403.18512, 2024. 2" + ], + "bbox": [ + 91, + 92, + 480, + 273 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12936", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/a12eb351-f679-4d5e-8b40-f5c8f124976d_model.json b/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/a12eb351-f679-4d5e-8b40-f5c8f124976d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a86adbf7952ef6ca60b234634146c3ef67f03c8c --- /dev/null +++ b/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/a12eb351-f679-4d5e-8b40-f5c8f124976d_model.json @@ -0,0 +1,2466 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.184, + 0.131, + 0.816, + 0.154 + ], + "angle": 0, + "content": "Text-to-Any-Skeleton Motion Generation Without Retargeting" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.18, + 0.87, + 0.252 + ], + "angle": 0, + "content": "Qingyuan Liu\\(^{1}\\) Ke Lu\\(^{1,2}\\) Kun Dong\\(^{1}\\) Jian Xue\\(^{1*}\\) Zehai Niu\\(^{1}\\) Jinbao Wang\\(^{3}\\) \n\\(^{1}\\)University of Chinese Academy of Sciences \\(^{2}\\)Peng Cheng Laboratory \\(^{3}\\)Shenzhen University \n{liuqingyuan23, dongkun22, niuzehai18}@mails.ucas.ac.cn, {luk, xuejian}@ucas.ac.cn \nwangjb@szu.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.288, + 0.545, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.289, + 0.891, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.49, + 0.907, + 0.56 + ], + "angle": 0, + "content": "Figure 1. Introducing OmniSkel: A text-driven motion generation framework that directly generates motions for any skeletons without retargeting. Unlike traditional two-stage pipelines (top right) that require complex retargeting procedures, our single-stage approach (bottom right) only needs text input and target bone lengths (which are easily obtainable) to generate desired motions. As shown on the left, OmniSkel creates natural movements like \"pushed but did not fall\" and \"walks up stairs with right hand on handrail\" for characters with different body proportions while maintaining action fidelity." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.572, + 0.327, + 0.588 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.608, + 0.486, + 0.865 + ], + "angle": 0, + "content": "Recent advances in text-driven motion generation have shown notable progress. However, these methods are typically limited to standardized skeletons and rely on a cumbersome targeting process to adapt to varying skeletal configurations of diverse characters. In this paper, we present OmniSkel, a novel framework that directly generates high-quality human motions for any user-defined skeleton without retargeting. Specifically, we introduce a skeleton-aware RVQ-VAE, which utilizes Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into motion encoding and reconstruction. Moreover, we propose a simple yet effective training-free approach, Motion Restoration Optimizer (MRO), to ensure zero bone length error while preserving motion smoothness. To support this research, we construct SkeleMotion3D, a large-scale text-skeleton-motion dataset based on HumanML3D. Extensive experiments demonstrate the excel" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.574, + 0.842, + 0.589 + ], + "angle": 0, + "content": "lent robustness and generalization of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.618, + 0.645, + 0.634 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Recent years have witnessed remarkable progress in text-to-motion (T2M) generation, showing promising potential in applications such as gaming, metaverse, and virtual/augmented reality [10, 16, 32, 41, 50, 51]. This technology enables the intuitive creation of 3D character animations through natural language descriptions, significantly simplifying content creation workflows in animation [20], virtual reality [16], video games [30, 47], and robotics [5, 22]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Despite these advancements, a fundamental limitation restricts the practical application of current text-to-motion generation methods. As illustrated in Figure 1, existing approaches primarily focus on enhancing the motion quality and text-motion alignment, neglecting variations in skeletons and only generating human motions for standardized skeletons. However, real-world production environments routinely require varying skeletal configurations for diverse" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12926" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "characters. To this end, practitioners must resort to skeleton retargeting - a technically demanding process that necessitates significant expertise and frequently fails to produce usable results, inevitably leading to laborious tuning and production delays. Therefore, there is a pressing demand for straightforward motion generation given arbitrary skeletons." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.2, + 0.483, + 0.501 + ], + "angle": 0, + "content": "To address this issue, we present OmniSkel, a novel framework that can directly generate high-quality human motion sequences for arbitrary character skeletons. As shown in Figure 1, our approach eliminates the complex retargeting stage required by traditional pipelines, offering a single-stage solution that directly produces character-specific motion without specialized expertise. Specifically, we propose a Skeleton-aware RVQ-VAE (SR-VAE) architecture, which utilizes Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into the motion encoding and reconstruction. Unlike previous approaches, our design effectively separates skeleton-invariant features, which is crucial for arbitrary skeleton adaptation. To ensure zero bone length error while preserving motion smoothness, we further propose a training-free approach, Motion Restoration Optimizer (MRO). With the above designs, our OmniSkel can deliver high-quality motion generation precisely in line with textual descriptions and character skeletons, providing robust support for practical text-to-motion applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.483, + 0.581 + ], + "angle": 0, + "content": "To facilitate our research, we construct SkeleMotion-3D, a large-scale dataset based on HumanML3D, pairing textual descriptions with motion sequences across diverse skeletal configurations. This dataset establishes a strong foundation for motion generation with arbitrary skeletons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.612 + ], + "angle": 0, + "content": "Before delving into the details, we summarize our core contributions in this work as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.614, + 0.483, + 0.689 + ], + "angle": 0, + "content": "- To the best of our knowledge, we present the first text-to-motion generation framework capable of directly producing motions for arbitrary character skeletons without retargeting, significantly simplifying creation workflows in real-world production environments." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.69, + 0.483, + 0.796 + ], + "angle": 0, + "content": "- We propose a novel Skeleton-aware RVQ-VAE (SR-VAE) architecture, which employs Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into the motion encoding and reconstruction. Moreover, we further introduce Motion Reconstruction Optimizer (MRO) during inference, ensuring zero bone length error while preserving motion smoothness." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.855 + ], + "angle": 0, + "content": "- Based on HumanML3D, we construct SkeleMotion-3D, a large-scale text-skeleton-motion dataset. This establishes a foundation for arbitrary skeleton motion generation, facilitating future research in this field." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- To verify the effectiveness of our method, we conduct extensive experiments on challenging datasets. Results demonstrate the superior robustness and generalization of" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "our method. Further analysis reveals the contribution of each component to the performance improvement." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.137, + 0.655, + 0.153 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.163, + 0.757, + 0.178 + ], + "angle": 0, + "content": "2.1. Text-to-Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.185, + 0.905, + 0.229 + ], + "angle": 0, + "content": "Text-to-motion generation has evolved significantly as a vital research area in computer animation and human-computer interaction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.231, + 0.905, + 0.291 + ], + "angle": 0, + "content": "Early Approaches. Initial methods established text-to-motion mappings through alignment techniques [3, 4, 9, 26, 36, 46], but produced motions that lacked naturalism and diversity due to their deterministic nature." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.905, + 0.352 + ], + "angle": 0, + "content": "Probabilistic Frameworks. VAE-based approaches [6, 16, 28, 31, 32] model distributions of possible motions, while vision-language alignment methods [27, 33, 37, 40] leveraged pre-trained models for better semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.353, + 0.905, + 0.488 + ], + "angle": 0, + "content": "Recent Advances. Contemporary approaches include diffusion-based models [7, 10, 12, 21, 38, 41, 51, 52] generating high-fidelity motions through progressive denoising. Vector Quantized VAE (VQ-VAE) approaches treat motions as discrete token sequences, including autoregressive models [19, 45, 50, 54, 55], bidirectional methods [18, 35], and hybrid approaches [34]. Part-based generation methods [13, 54, 55] provide control over individual body components." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.489, + 0.905, + 0.548 + ], + "angle": 0, + "content": "Limitations. Existing approaches remain constrained to standard skeletal structures, creating a gap between research and industry needs where adaptability to various character rigs is essential." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.55, + 0.905, + 0.595 + ], + "angle": 0, + "content": "Our work addresses this limitation by developing a skeleton-agnostic framework that generates semantically consistent motions for arbitrary skeletal structures." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.606, + 0.702, + 0.623 + ], + "angle": 0, + "content": "2.2. Motion Retargeting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Motion retargeting adapts existing motion data between characters with different proportions and structures. Historically, Gleicher et al. [14] pioneered this field through spacetime optimization with kinematic constraints. Following this work, Lee and Shin [23] applied inverse kinematics with B-spline curve fitting, while Choi and Ko [11] simultaneously developed an online method preserving high-frequency details. Tak and Ko [39] later enhanced these approaches by utilizing dynamics constraints for physical plausibility." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "With the advent of deep learning, new approaches emerged, including Villegas et al. [44] with unsupervised retargeting via recurrent networks, Lim et al. [25] disentangling pose and movement, and Aberman et al. [1] designing skeleton-aware networks for cross-structural retargeting. More recently, research has shifted towards preserving motion semantics. For instance, Zhang et al. [49] developed a residual retargeting network with skeleton-aware" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12927" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.09, + 0.616, + 0.104 + ], + "angle": 0, + "content": "(a) Skeleton-aware Residual VQ-VAE" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.108, + 0.885, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.335, + 0.388, + 0.348 + ], + "angle": 0, + "content": "(b) Masked Transformer" + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.336, + 0.78, + 0.349 + ], + "angle": 0, + "content": "(c) Residual Transformer" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.36, + 0.908, + 0.459 + ], + "angle": 0, + "content": "Figure 2. Method overview. (a) Skeleton-aware Residual VQ-VAE: our framework processes source motion and skeleton through K-GCA (Kinematic Graph Cross Attention), which divides the input into Q, K, and V branches, creating a hierarchical representation with motion-invariant features. The residual quantization approach progressively captures motion details across multiple layers, enabling high-fidelity reconstruction for arbitrary target skeletons. (b) Masked Transformer: base-layer motion tokens \\( t^0 \\) are randomly masked during training and the text-conditioned transformer learns to predict these tokens simultaneously, capturing fundamental motion patterns. (c) Residual Transformer: builds upon the base layer by progressively predicting higher-level residual tokens \\( t^v \\) conditioned on both text and previous layer tokens, enhancing motion fidelity and skeleton-specific details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.485, + 0.483, + 0.62 + ], + "angle": 0, + "content": "and shape-aware modules. Similarly, Lee et al. [24] proposed Skeleton-Agnostic Motion Embedding (SAME) using graph convolutional networks to separate skeleton information from motion while preserving semantics. Building upon these semantic approaches, Zhang et al. [48] introduced Semantics-aware Motion reTargeting (SMT), leveraging vision-language models to extract and maintain motion semantics through a two-stage pipeline with skeleton-aware pre-training and semantic constraints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.484, + 0.667 + ], + "angle": 0, + "content": "However, despite this significant progress, many approaches still struggle with maintaining semantic consistency across dramatically different skeletal structures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.2, + 0.697 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our goal is to generate motion sequences \\(\\mathbf{M}_{1:T}\\) of length \\(T\\) conditioned on both a text description \\(c\\) and a target skeletal structure \\(\\mathbf{S}\\), where \\(\\mathbf{M}_t\\in \\mathbb{R}^{J\\times D_{\\mathrm{m}}}\\) with \\(J\\) denoting the number of joints and \\(D_{\\mathrm{m}}\\) representing the dimension of joint features. The skeletal structure \\(\\mathbf{S}\\in \\mathbb{R}^{J\\times D_{\\mathrm{s}}}\\) characterizes the static joint properties of the target character. As illustrated in Figure 2, our framework consists of two main components: a skeleton-aware VAE module that enables motion generation for arbitrary skeletal structures (Section 3.2), and a cooperative transformer architecture for motion token prediction (Section 3.3). We first introduce our newly proposed dataset and its representation (Section 3.1) to facilitate better understanding of our method. Then, we detail" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.515 + ], + "angle": 0, + "content": "the complete training pipeline (Section 3.3) and inference process (Section 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.523, + 0.738, + 0.537 + ], + "angle": 0, + "content": "3.1. SkeleMotion-3D Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.651 + ], + "angle": 0, + "content": "Our dataset comprises three key components: textual descriptions, motion sequences, and corresponding skeletal structures. Building upon the HumanML3D dataset [16], which combines motion data from AMASS [29] and HumanAct12 [15], we significantly expand the motion-skeleton pairs through a skeleton randomization process inspired by [24]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.652, + 0.906, + 0.696 + ], + "angle": 0, + "content": "Feature Representation. We represent motion sequences as \\(\\mathbf{M} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}\\), where the motion feature \\(F_{\\mathrm{motion}}\\) with dimension \\(D_{\\mathrm{m}}\\) consists of:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.705, + 0.905, + 0.722 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {m o t i o n}} = \\left[ F _ {\\text {t o p o}}, F _ {\\text {r o o t}}, F _ {\\text {x y z}}, F _ {\\text {r o t}}, F _ {\\text {v e l}} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.906, + 0.773 + ], + "angle": 0, + "content": "comprising skeletal topology features, root joint features, relative spatial positions, 6D rotation features, and velocity features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.774, + 0.906, + 0.817 + ], + "angle": 0, + "content": "The skeletal structures are encoded as \\(\\mathbf{S} \\in \\mathbb{R}^{J \\times D_{\\mathrm{s}}}\\), where the skeletal feature \\(F_{\\mathrm{skel}}\\) with dimension \\(D_{\\mathrm{s}}\\) consists of:" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.82, + 0.905, + 0.836 + ], + "angle": 0, + "content": "\\[\nF _ {\\mathrm {s k e l}} = \\left[ F _ {\\mathrm {t o p o}}, F _ {\\mathrm {s t a t i c}} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.871 + ], + "angle": 0, + "content": "where \\(F_{\\mathrm{topo}}\\) represents topology features and \\(F_{\\mathrm{static}}\\) denotes static skeletal properties." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Dataset Construction. Following the preprocessing pipeline established by HumanML3D, we standardize the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12928" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.181 + ], + "angle": 0, + "content": "raw motion sequences through temporal normalization (20 FPS), sequence cropping (maximum 10 seconds), skeletal retargeting, and orientation alignment \\((\\mathrm{Z}+\\) direction). The text annotations from HumanML3D, collected via Amazon Mechanical Turk with quality control measures, are preserved in our dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.319 + ], + "angle": 0, + "content": "Dataset Statistics. The SkeleMotion-3D dataset contains 14,616 distinct motion patterns, 29,233 different skeletal configurations, 43,848 unique text-motion-skeleton triplets, and 44,970 textual descriptions with 5,371 unique words. This represents a substantial increase compared to the training dataset (949 motion patterns and 4,745 motion-skeleton pairs) used in [24]'s motion retargeting network. The significant expansion enhances the robustness and generalization capability of our skeleton-aware motion generation model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.328, + 0.391, + 0.343 + ], + "angle": 0, + "content": "3.2. Skeleton-aware Residual VQ-VAE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.483, + 0.515 + ], + "angle": 0, + "content": "We propose Skeleton-aware Residual VQ-VAE (SR-VAE), a novel architecture that extends the conventional Residual VQ-VAE to enable skeleton-conditioned motion generation. Conventional motion VQ-VAEs [17, 19, 50, 53] and recent RVQ-VAEs [18, 45] focus on reconstructing source motions with identical skeletal structures. In contrast, as illustrated in Figure 2, SR-VAE adopts an asymmetric VAE architecture, consisting of an encoder, a decoder, a residual quantization module, and a Kinetic Graph Cross Attention (K-GCA) module, enabling motion reconstruction for arbitrary target skeletons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.483, + 0.667 + ], + "angle": 0, + "content": "Architecture Design. Our SR-VAE processes motion generation through three main stages. First, the input motion sequence \\(\\mathbf{M}_{\\mathrm{in}} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}\\) and its corresponding skeleton structure \\(\\mathbf{S}_{\\mathrm{in}} \\in \\mathbb{R}^{J \\times D_{\\mathrm{s}}}\\) are fused via K-GCA before being encoded into latent features. Following MoMask [18], these features undergo residual quantization to produce \\(V + 1\\) ordered code sequences. Finally, the decoder incorporates target skeleton features \\(\\mathbf{S}_{\\mathrm{tgt}} \\in \\mathbb{R}^{J \\times D_{\\mathrm{s}}}\\) through multiple K-GCA layers at different decoding stages to generate the target motion \\(\\mathbf{M}_{\\mathrm{tgt}} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.483, + 0.804 + ], + "angle": 0, + "content": "Let \\(\\hat{\\mathbf{M}}_{\\mathrm{tgt}} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}\\) denote the predicted motion sequence and \\(\\mathbf{M}_{\\mathrm{tgt}}\\) denote the ground truth target motion. For residual quantization, let \\(\\mathbf{r}^v \\in \\mathbb{R}^{T \\times d}\\) represent the residual features at layer \\(v \\in \\{0,1,\\dots,V\\}\\) and \\(\\mathbf{b}^v \\in \\mathbb{R}^{T \\times d}\\) denote the corresponding quantized codes, where \\(d\\) is the dimension of the latent representation. Additionally, let \\(\\mathbf{D}_{\\mathrm{xyz}} \\in \\mathbb{R}^{T \\times J \\times 3}\\) and \\(\\hat{\\mathbf{D}}_{\\mathrm{xyz}} \\in \\mathbb{R}^{T \\times J \\times 3}\\) represent the ground truth and predicted joint positions in 3D space, respectively. The overall training objective consists of three terms:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.81, + 0.482, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\left\\| \\mathbf {M} _ {\\mathrm {t g t}} - \\hat {\\mathbf {M}} _ {\\mathrm {t g t}} \\right\\| _ {1} + \\beta \\sum_ {v = 0} ^ {V} \\left\\| \\mathbf {r} ^ {v} - \\operatorname {s g} \\left[ \\mathbf {b} ^ {v} \\right] \\right\\| _ {2} ^ {2} + \\lambda \\left\\| \\mathbf {D} _ {\\mathrm {x y z}} - \\hat {\\mathbf {D}} _ {\\mathrm {x y z}} \\right\\| _ {1} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where the first term is motion reconstruction loss, the second term is commitment loss for residual quantization, and the third term specifically focuses on joint positions accu" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.194, + 0.89, + 0.208 + ], + "angle": 0, + "content": "Figure 3. Architecture of our Kinetic Graph Cross Attention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.235, + 0.905, + 0.325 + ], + "angle": 0, + "content": "racy. Here, \\(\\mathrm{sg}[\\cdot ]\\) denotes the stop-gradient operation, \\(\\beta\\) and \\(\\lambda\\) are weighting factors for the commitment loss and joint position loss, respectively. The loss function is optimized via a straight-through gradient estimator [42]. Following [18, 35, 50], we employ exponential moving average for codebook updates and reset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.906, + 0.446 + ], + "angle": 0, + "content": "Kinetic Graph Cross Attention. To effectively fuse motion and skeletal features, we propose Kinetic Graph Cross Attention (K-GCA), as illustrated in Figure 3, extending the traditional Graph Attention Networks (GAT) [8, 43] for motion-skeleton feature fusion. Unlike standard GAT that performs self-attention within a single feature type, K-GCA enables cross-modal attention between motion and skeletal features while preserving the skeletal graph structure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.905, + 0.522 + ], + "angle": 0, + "content": "Given motion features \\(\\mathbf{X} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}\\) and skeleton features \\(\\mathbf{S} \\in \\mathbb{R}^{J \\times D_{\\mathrm{s}}}\\), K-GCA first projects them into a shared space using learnable projection matrices \\(\\mathbf{W}_{\\mathrm{m}} \\in \\mathbb{R}^{D_{\\mathrm{m}} \\times d_k}\\) and \\(\\mathbf{W}_{\\mathrm{s}} \\in \\mathbb{R}^{D_{\\mathrm{s}} \\times d_k}\\), where \\(d_k\\) is the dimension of the projected space:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.534, + 0.905, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {W} _ {\\mathrm {m}} \\mathbf {X}, \\quad \\mathbf {K} = \\mathbf {V} = \\mathbf {W} _ {\\mathrm {s}} \\mathbf {S}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.563, + 0.905, + 0.593 + ], + "angle": 0, + "content": "The attention scores between joints \\(j\\) and \\(k\\) are computed using learnable vectors \\(\\mathbf{a}^{\\mathrm{src}}\\in \\mathbb{R}^{d_k}\\) and \\(\\mathbf{a}^{\\mathrm{dst}}\\in \\mathbb{R}^{d_k}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.604, + 0.905, + 0.622 + ], + "angle": 0, + "content": "\\[\ne _ {j k} = \\text {L e a k y R e L U} \\left(\\mathbf {a} ^ {\\mathrm {s r c}} \\cdot \\mathbf {Q} _ {j} + \\mathbf {a} ^ {\\mathrm {d s t}} \\cdot \\mathbf {K} _ {k}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.906, + 0.721 + ], + "angle": 0, + "content": "To preserve the skeletal structure, we mask the attention scores using the skeleton adjacency matrix \\(\\mathbf{A} \\in \\{0,1\\}^{J \\times J}\\), where \\(\\mathbf{A}_{jk} = 1\\) if joints \\(j\\) and \\(k\\) are connected in the skeleton graph, and 0 otherwise. We set attention scores for disconnected joints to negative infinity before applying softmax:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.721, + 0.905, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\tilde {e} _ {j k} = \\left\\{ \\begin{array}{l l} e _ {j k}, & \\text {i f} \\mathbf {A} _ {j k} = 1 \\\\ - \\infty , & \\text {i f} \\mathbf {A} _ {j k} = 0 \\end{array} \\right., \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.769, + 0.905, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {j k} = \\frac {\\exp \\left(\\tilde {e} _ {j k}\\right)}{\\sum_ {l = 1} ^ {J} \\exp \\left(\\tilde {e} _ {j l}\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.873 + ], + "angle": 0, + "content": "The final output \\(\\hat{\\mathbf{X}}\\in \\mathbb{R}^{T\\times J\\times D_{\\mathrm{m}}}\\) is computed by applying the attention weights to value matrix \\(\\mathbf{V}\\) in each head, concatenating results from all heads, and adding back to the input through residual connection:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.884, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {X}} = \\operatorname {L a y e r N o r m} (\\mathbf {X} + \\operatorname {M u l t i H e a d} (\\alpha , \\mathbf {V})). \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12929" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.244 + ], + "angle": 0, + "content": "where MultiHead \\((\\alpha, \\mathbf{V})\\) represents the multi-head weighted combination of values using the computed attention weights \\(\\alpha\\). This design allows K-GCA to capture dynamic relationships between motion and skeletal features while maintaining the anatomical constraints imposed by the skeletal structure. Compared to standard GAT, our approach provides two key advantages: (1) the ability to perform cross-modal attention between motion and skeletal features, and (2) the incorporation of skeletal structure constraints through attention masking." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.253, + 0.262, + 0.27 + ], + "angle": 0, + "content": "3.3. Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.484, + 0.32 + ], + "angle": 0, + "content": "Our training pipeline follows a sequential approach for skeleton-agnostic motion generation from text, as shown in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.412 + ], + "angle": 0, + "content": "SR-VAE Training. First, we train the Skeleton-aware Residual VQ-VAE (SR-VAE) to establish skeleton-agnostic motion representation. This model encodes motion sequences into hierarchical tokens and decodes them back to motion features, independent of skeletal structure. After training, both encoder \\(\\mathcal{E}\\) and decoder \\(\\mathcal{D}\\) are frozen." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.488 + ], + "angle": 0, + "content": "Tokenization of Training Data. Using the frozen SR-VAE encoder, we convert all training motion sequences into hierarchical token representations. For each motion \\(\\mathbf{M}_{\\mathrm{in}}\\) with corresponding skeleton \\(\\mathbf{S}_{\\mathrm{in}}\\), we extract multi-level token sequences \\(\\mathbf{t}^{0:V}\\) as targets for our text-to-token models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.594 + ], + "angle": 0, + "content": "Masked Transformer Training. Inspired by MoMask [18], we implement a bidirectional Masked Transformer to generate base-layer tokens \\((\\mathbf{t}^0)\\). This design offers parallel generation capability and better global motion coherence compared to autoregressive approaches. During training, the model learns to predict original tokens at masked positions from text prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.685 + ], + "angle": 0, + "content": "Residual Transformer Training. Finally, we train the Residual Transformer to predict higher-layer tokens \\((\\mathbf{t}^{1:V})\\) encoding motion refinements. Following MoMask [18], this transformer takes both text descriptions and tokens from previous layers as input, focusing on finer motion details while maintaining consistency with the base motion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.262, + 0.71 + ], + "angle": 0, + "content": "3.4. Inference Process" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.748 + ], + "angle": 0, + "content": "During inference, OmniSkel generates skeleton-aware motions through the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Base Token Generation: Given a text prompt, the Masked Transformer generates the base-layer tokens \\(\\hat{\\mathbf{t}}^0\\) that capture primary motion patterns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Hierarchical Refinement: The Residual Transformer progressively generates higher-layer tokens \\(\\hat{\\mathbf{t}}^{1:V}\\), starting with layer 1 and proceeding sequentially. Each layer's prediction utilizes both the text prompt and all previously generated token layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Motion Decoding: Once we have the complete token hierarchy \\(\\hat{\\mathbf{t}}^{0:V}\\), we feed these tokens along with the target skele" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "ton structure \\(\\mathbf{S}_{\\mathrm{tgt}}\\) into the frozen SR-VAE decoder to obtain motion features." + }, + { + "type": "code_caption", + "bbox": [ + 0.526, + 0.153, + 0.82, + 0.167 + ], + "angle": 0, + "content": "Algorithm 1: Motion Restoration Optimizer" + }, + { + "type": "code", + "bbox": [ + 0.527, + 0.169, + 0.867, + 0.23 + ], + "angle": 0, + "content": "Input: Measurement data: \\(F_{\\mathrm{topo}}\\), \\(F_{\\mathrm{root}}\\), \\(F_{\\mathrm{xyz}}\\); Bone lengths: \\(S\\) \nOutput: Smooth motion sequence \\(M_{\\mathrm{smooth}}\\) with guaranteed skeletal consistency" + }, + { + "type": "code_caption", + "bbox": [ + 0.521, + 0.236, + 0.797, + 0.265 + ], + "angle": 0, + "content": "1 // Step 1: Restore initial motion using position-based approach;" + }, + { + "type": "code", + "bbox": [ + 0.521, + 0.266, + 0.73, + 0.281 + ], + "angle": 0, + "content": "2 \\( M_{\\mathrm{ric}} \\gets \\mathrm{ric}(F_{\\mathrm{topo}}, F_{\\mathrm{root}}, F_{\\mathrm{xyz}}) \\);" + }, + { + "type": "code", + "bbox": [ + 0.521, + 0.288, + 0.63, + 0.301 + ], + "angle": 0, + "content": "3 \\(M_{\\mathrm{smooth}} \\gets \\emptyset\\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.521, + 0.308, + 0.715, + 0.322 + ], + "angle": 0, + "content": "4 foreach frame \\(f\\in M_{\\mathrm{ric}}\\) do" + }, + { + "type": "code", + "bbox": [ + 0.519, + 0.323, + 0.882, + 0.628 + ], + "angle": 0, + "content": "// Step 2: Extract unit direction vectors for each bone (except the root); \nInitialize set \\(\\mathcal{U} \\gets \\emptyset\\); \nfor \\(i \\gets 1\\) to \\(J - 1\\) do \nLet \\(\\mathbf{p}_i\\) be the position of joint \\(i\\) in frame \\(f\\); \nLet \\(\\mathbf{p}_{\\text{parent}(i)}\\) be the position of joint \\(i\\)'s parent; \nCompute displacement: \\(\\mathbf{v}_i \\gets \\mathbf{p}_i - \\mathbf{p}_{\\text{parent}(i)}\\); \nCompute unit direction: \\(\\mathbf{u}_i \\gets \\frac{\\mathbf{v}_i}{|\\mathbf{v}_i|}\\); \nAdd \\(\\mathbf{u}_i\\) to \\(\\mathcal{U}\\); \nend \n// Step 3: Reconstruct the current frame using extracted directions, bone lengths, and topology; \n\\(f_{\\text{rec}} \\gets \\text{ReconstructFrame}(\\mathcal{U}, S, F_{\\text{topo}}, F_{\\text{root}})\\); \nAdd \\(f_{\\text{rec}}\\) to \\(M_{\\text{smooth}}\\); \nend \nreturn \\(M_{\\text{smooth}}\\);" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.907, + 0.789 + ], + "angle": 0, + "content": "Feature-to-Motion Conversion: Following the generation of motion features by our SR-VAE decoder, we face the challenge of transforming these abstract representations into production-ready motion data. Contemporary text-to-motion frameworks typically rely on either position-based restoration (offering fluidity but with bone length inconsistencies) or rotation-based restoration (maintaining skeletal integrity but introducing jitter and penetration artifacts)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To overcome these limitations, we employ our Motion Restoration Optimizer (MRO, detailed in Algorithm 1), which synergistically combines the strengths of both approaches. MRO guarantees two critical properties simultaneously: absolute preservation of skeletal proportions (zero bone length error) and maintenance of smooth trajectory characteristics, producing physically plausible output \\(\\hat{\\mathbf{M}}_{\\mathrm{tgt}}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12930" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.921, + 0.246 + ], + "angle": 0, + "content": "
MethodFID↓R Precision↑MPJPE↓Skeleton Error↓Jitter Error↓Skating Ratio ↓
Top 1Top 2Top 3
SR-VAE (full model)0.0270.5080.7000.7930.03513.7781.8240.074
w/ Single-head K-GCA0.0400.4970.6850.7890.03412.2942.9350.061
w/o K-GCA0.0650.4950.6890.7850.04117.7269.5940.092
w/o RVQ0.1650.4670.6390.7750.06119.7263.5940.092
w/ Symmetric Reconstruction0.8850.4500.6410.7450.09789.7479.8670.067
w/ Rotation-based Restoration0.0270.5080.7000.7930.0630.0030.7240.193
w/ MRO0.0270.5080.7000.7930.0350.002.0840.090
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.256, + 0.908, + 0.3 + ], + "angle": 0, + "content": "Table 1. Ablation study of SR-VAE components tested on the SkeleMotion-3D test set. Our SR-VAE is trained on the SkeleMotion-3D dataset. To evaluate SR-VAE's reconstruction performance for arbitrary skeletal motions, we input random skeleton motions to the encoder and use the standard skeleton from the HumanML3D dataset as the target skeleton." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.324, + 0.226, + 0.343 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.396 + ], + "angle": 0, + "content": "We conduct extensive experiments to evaluate our approach across multiple dimensions, focusing on both semantic accuracy and physical plausibility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.484, + 0.472 + ], + "angle": 0, + "content": "Evaluation Metrics. We assess our model using established semantic metrics from prior work [16]: (1) Frechet Inception Distance (FID) for overall motion quality; (2) \\(R\\)-Precision and multimodal distance for text-motion alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.519 + ], + "angle": 0, + "content": "To comprehensively evaluate physical plausibility—a critical aspect for animation applications—we introduce several specialized metrics:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.52, + 0.483, + 0.55 + ], + "angle": 0, + "content": "- Skating Ratio: Measures foot sliding artifacts where feet unnaturally glide across the ground" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.55, + 0.482, + 0.581 + ], + "angle": 0, + "content": "- Skeleton Error: Quantifies bone length variations throughout motion sequences" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.581, + 0.482, + 0.611 + ], + "angle": 0, + "content": "- MPJPE (Mean Per-Joint Position Error): Assesses positional accuracy of joints" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.482, + 0.657 + ], + "angle": 0, + "content": "- Jitter Error: Evaluates motion continuity and natural flow. Detailed formulations of these metrics appear in the supplementary materials." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Implementation Details. Our PyTorch implementation features a Skeleton-aware Residual VQ-VAE (SR-VAE) with residual blocks following T2M-GPT [50] and MoMask [18] architectures. We use a downscale factor of 4, with a residual quantization comprising 6 layers (each containing 512 codes of 512 dimensions) and quantization dropout ratio of 0.2. For the SR-VAE training objective, we set commitment loss weight \\(\\beta\\) to 0.02 and joint position loss weight \\(\\lambda\\) to 2.0. The K-GCA module employs 2-head attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Both transformer models consist of 6 layers with 6 attention heads and 384-dimensional latent representations, trained on our SkeleMotion-3D dataset. We use a learning rate of 2e-4 with linear warm-up over 2000 iterations, batch sizes of 512 for SR-VAE and 64 for transformers. During inference, we apply classifier-free guidance (scale 4 for" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.325, + 0.907, + 0.37 + ], + "angle": 0, + "content": "masked transformer, 5 for refinement transformer) with 10 iterative decoding steps. All experiments run on NVIDIA RTX 4090 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.38, + 0.874, + 0.396 + ], + "angle": 0, + "content": "4.1. Skeleton-Aware Disentanglement Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.463 + ], + "angle": 0, + "content": "A cornerstone of our approach is SR-VAE's ability to disentangle motion patterns from skeletal structures. We conduct a targeted experiment to validate this skeleton-agnostic capability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.538 + ], + "angle": 0, + "content": "Figure 4 presents our analysis of four distinct motion patterns, each performed by three different skeletal structures with varying proportions. We process these motion-skeleton pairs through our SR-VAE encoder \\(\\mathcal{E}\\) and visualize the resulting base layer tokens \\(\\mathbf{t}^0\\) on aligned line charts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.643 + ], + "angle": 0, + "content": "The results reveal that motions with identical semantic patterns but executed by anatomically diverse skeletons produce remarkably similar token sequences. This consistency demonstrates that our encoder \\(\\mathcal{E}\\) successfully extracts skeleton-invariant motion features, preserving essential semantic content regardless of the executing skeletal structure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.75 + ], + "angle": 0, + "content": "This confirms that SR-VAE effectively decomposes motion into two orthogonal components: (1) skeleton-invariant pattern information encoded in tokens, and (2) skeleton-specific structural information provided separately to the decoder. This disentanglement enables our model to generate physically plausible animations for arbitrary skeletal structures while preserving intended motion semantics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.667, + 0.774 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.856 + ], + "angle": 0, + "content": "To validate SR-VAE's skeleton-aware capabilities, we conduct experiments using random input skeletons while decoding to standard HumanML3D [16] skeletons. This evaluates both semantic alignment and cross-skeleton reconstruction capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Table 1 presents our ablation results on the SkeleMotion-3D test set. The complete SR-VAE model achieves superior performance with an FID of 0.026 and strong R" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12931" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.099, + 0.904, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.249, + 0.908, + 0.32 + ], + "angle": 0, + "content": "Figure 4. Motion token analysis across skeletal structures. This figure demonstrates token sequences generated by the SR-VAE encoder for four distinct motion patterns, each performed by three skeletal structures with different proportions. The consistent token patterns across each motion type (columns) despite skeletal variations demonstrates our encoder's ability to extract skeleton-invariant motion features. Top row shows skeletal visualizations while bottom row displays the corresponding quantized token values, revealing clear pattern preservation across different anatomical configurations." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.336, + 0.847, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.609, + 0.908, + 0.665 + ], + "angle": 0, + "content": "Figure 5. Text-to-motion generation across different skeletal structures. Our OmniSkel framework generates diverse motions responding to different text prompts (columns) while adapting to different skeletal anatomies (rows). The human skeleton on the left is randomly selected from Mixamo [2], demonstrating OmniSkel's ability to maintain semantic consistency across dramatically different skeletal structures without retargeting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.691, + 0.483, + 0.722 + ], + "angle": 0, + "content": "Precision scores (0.508/0.700/0.793 for Top-1/2/3), demonstrating excellent semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.85 + ], + "angle": 0, + "content": "Architecture Components. Using symmetric reconstruction (same skeleton for encoder and decoder) drastically increases FID to 0.885 and skeleton error to 89.747, confirming that asymmetric processing is essential for effective cross-skeleton learning. Removing K-GCA degrades FID to 0.065 with a significant increase in jitter error (9.594 vs. 1.824), demonstrating its importance in capturing cross-skeleton geometric correspondences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Without RVQ, FID increases substantially to 0.165 and MPJPE to 0.061, highlighting RVQ's crucial role in maintaining semantic consistency across different skeletal struc" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.907, + 0.752 + ], + "angle": 0, + "content": "tures. The single-head K-GCA variant shows moderate performance decline (FID: 0.040), indicating that multi-head attention more effectively captures the rich geometric relationships between skeletons." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Motion Restoration. For motion restoration approaches, rotation-based methods achieve perfect skeleton reconstruction (skeleton error: 0.00) but significantly increase jitter error (30.724) and skating performance (0.193). In contrast, our MRO technique also achieves perfect skeleton reconstruction (skeleton error: 0.00) while maintaining low jitter error (2.084) and better skating ratio (0.090). This demonstrates MRO's effectiveness in preserving physical plausibility while maintaining motion quality across differ-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12932" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.241, + 0.105 + ], + "angle": 0, + "content": "ent skeletal structures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.484, + 0.231 + ], + "angle": 0, + "content": "Our ablation study confirms each component's critical contribution to the overall system: K-GCA and RVQ enhance cross-skeleton reconstruction and semantic alignment, while MRO ensures physical consistency in the generated motions. These findings validate our architectural design choices for skeleton-aware motion generation and demonstrate SR-VAE's ability to effectively handle arbitrary skeleton motion retargeting with high fidelity." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.241, + 0.418, + 0.258 + ], + "angle": 0, + "content": "4.3. Qualitative and Quantitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.265, + 0.484, + 0.356 + ], + "angle": 0, + "content": "Quantitative Evaluation. We evaluate our methods against state-of-the-art text-to-motion approaches as shown in Table 2. The fundamental distinction of our approach is the ability to directly generate motions for arbitrary skeleton structures—a paradigm shift from existing methods that are restricted to fixed, pre-defined skeletons." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.37, + 0.482, + 0.535 + ], + "angle": 0, + "content": "
DatasetsMethodsFID↓R Precision↑
Top 1Top 2Top 3
Human-ML3DTM2T [17]1.501±.0170.424±.0030.618±.0030.729±.002
T2M [16]1.087±.0210.455±.0030.636±.0030.736±.002
MDM [41]0.544±.0440.320±.0050.498±.0040.611±.007
MLD [10]0.473±.0130.481±.0030.673±.0030.772±.002
MotionDiffuse [51]0.630±.0010.491±.0010.681±.0010.782±.001
T2M-GPT [50]0.141±.0050.492±.0030.679±.0020.775±.002
ReMoDiffuse [52]0.103±.0040.510±.0050.698±.0060.795±.004
AttT2M [54]0.112±.0060.499±.0030.690±.0020.786±.002
Momask [18]0.045±.0020.521±.0020.713±.0020.807±.002
Skele-Motion-3D OmniSkel (ours)0.027±.0000.508±.0020.700±.0020.793±.002
0.084±.0040.507±.0020.699±.0030.797±.002
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.686 + ], + "angle": 0, + "content": "Table 2. Comparison with text-to-motion methods on standard skeleton. Our approach introduces a novel pipeline trained on the SkeleMotion-3D dataset, which fundamentally differs from traditional T2M methods. For fair comparison with existing methods trained on HumanML3D, we configured our models (SR-VAE and OmniSkel) to target the standard HumanML3D skeleton during inference. Despite being designed for arbitrary skeleton generation, our methods achieve competitive performance on this fixed skeleton setting, demonstrating strong text-to-motion capabilities even when constrained to a single skeleton structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.825 + ], + "angle": 0, + "content": "This fundamental difference presents a challenge for direct comparison, as our methods operate on the newly introduced SkeleMotion-3D dataset with diverse skeleton structures, while previous approaches are confined to the HumanML3D dataset with a single skeleton configuration. Notably, existing methods cannot be trained or evaluated on our dataset due to their inherent architectural limitations in handling skeleton variations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To provide a meaningful comparison that highlights our text-to-motion capabilities, we configure our models at inference time to target the standard HumanML3D skeleton structure. Despite being designed for the more complex task of arbitrary skeleton generation, both SR-VAE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "and OmniSkel achieve competitive performance on the HumanML3D test set, with comparable or superior metrics in FID and R-Precision compared to specialized methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.905, + 0.197 + ], + "angle": 0, + "content": "These results demonstrate that our methods successfully balance adaptability with generation quality, enabling cross-skeleton applications without compromising performance on standard benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.905, + 0.273 + ], + "angle": 0, + "content": "Qualitative Results. Figure 5 illustrates OmniSkel's capability to generate motions across different skeletal structures. The visualization demonstrates: (1) diverse motions for different textual descriptions, and (2) consistent semantic motion across distinct skeletal anatomies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.905, + 0.334 + ], + "angle": 0, + "content": "Each column represents a different textual prompt, from complex sequences to simpler actions. The model successfully captures the nuanced details of these descriptions, including locomotion styles and sequential actions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.905, + 0.396 + ], + "angle": 0, + "content": "This comparison demonstrates OmniSkel's precise perception and effective integration of different semantics and arbitrary skeletons. More visualization results are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.406, + 0.749, + 0.423 + ], + "angle": 0, + "content": "4.4. Performance Comparison" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.441, + 0.903, + 0.486 + ], + "angle": 0, + "content": "
MethodParametersInference TimeGPU Memory
MoMask [18]44.85M0.075s1434M
Ours38.73M0.099s1402M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.545, + 0.496, + 0.874, + 0.51 + ], + "angle": 0, + "content": "Table 3. Computational cost comparison on RTX 4090." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.905, + 0.634 + ], + "angle": 0, + "content": "Our method achieves comparable performance to state-of-the-art approaches while maintaining slight GPU memory advantages as shown in Table 3 and uniquely enabling end-to-end motion generation for arbitrary skeletal specifications. This efficiency demonstrates the practical viability of our skeleton-agnostic framework for real-world applications requiring diverse character animations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.762, + 0.665 + ], + "angle": 0, + "content": "5. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In conclusion, we present OmniSkel, a text-driven motion generation framework that enables motion synthesis for arbitrary human skeletons through our novel K-GCA and SR-VAE architectures. Our Motion Restoration Optimizer (MRO) addresses the cross-skeleton motion reconstruction challenge, ensuring both motion smoothness and skeletal consistency. Leveraging our SkeleMotion-3D dataset, which pairs textual descriptions with diverse skeletal configurations, our approach bridges the gap between research and practical applications. Experiments demonstrate that OmniSkel achieves competitive performance while maintaining semantic alignment with textual descriptions. This work establishes a foundation for more inclusive motion generation systems that can serve diverse applications without requiring conventional retargeting approaches." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12933" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.177 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62320106007 and Grant 62032022; and in part by the Guangdong Provincial Key Laboratory under Grant 2023B1212060076." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.188, + 0.188, + 0.204 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.214, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.271, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[2] Adobe. Mixamo. Online Animation Library, 2023. Available at: https://www MIXamo.com/.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.299, + 0.483, + 0.366 + ], + "angle": 0, + "content": "[3] Hyemin Ahn, Timothy Ha, Yunho Choi, Hwiyeon Yoo, and Songhwai Oh. Text2action: Generative adversarial synthesis from language to action. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5915-5920. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[4] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719–728. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[5] André Antakli, Erik Hermann, Ingo Zinnikus, Han Du, and Klaus Fischer. Intelligent distributed human motion simulation in human-robot collaboration environments. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 319–324, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.482, + 0.549 + ], + "angle": 0, + "content": "[6] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.605 + ], + "angle": 0, + "content": "[7] Gustav Bredell, Kyriakos Flouris, Krishna Chaitanya, Ertunc Erdil, and Ender Konukoglu. Explicitly minimizing the blur error of variational autoencoders. arXiv preprint arXiv:2304.05939, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[8] Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks? In International Conference on Learning Representations, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[9] Haoye Cai, Chunyan Bai, Yu-Wing Tai, and Chi-Keung Tang. Deep video generation, prediction and completion of human action sequences. In Proceedings of the European Conference on Computer Vision (ECCV), pages 366–382, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[10] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[11] Kwang-Jin Choi and Hyeong-Seok Ko. Online motion retargeting. The Journal of Visualization and Computer Animation, 11(5):223-235, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[12] Xuehao Gao, Yang Yang, Zhenyu Xie, Shaoyi Du, Zhongqian Sun, and Yang Wu. Guess: Gradually enriching synthesis for text-driven human motion generation. IEEE Transactions on Visualization and Computer Graphics, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.214, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[13] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1396-1406, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Michael Gleicher. Retargetting motion to new characters. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, pages 33-42, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[15] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[18] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.904, + 0.516 + ], + "angle": 0, + "content": "[19] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[20] Moritz Kappel, Vladislav Golyanik, Mohamed Elgharib, Jann-Ole Henningson, Hans-Peter Seidel, Susana Castillo, Christian Theobalt, and Marcus Magnor. High-fidelity neural human motion transfer from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1541-1550, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[21] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[22] Hema Koppula and Ashutosh Saxena. Learning spatiotemporal structure from rgb-d videos for human activity detection and anticipation. In International conference on machine learning, pages 792-800. PMLR, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.717, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[23] Jehee Lee and Sung Yong Shin. A hierarchical approach to interactive motion editing for human-like figures. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 39-48, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[24] Sunmin Lee, Taeho Kang, Jungnam Park, Jehee Lee, and Jungdam Won. Same: Skeleton-agnostic motion embedding for character animation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[25] Jongin Lim, Hyung Jin Chang, and Jin Young Choi. PMnet: learning of disentangled pose and movement for unsupervised motion retargeting. In Proc. BMVC, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Angela S Lin, Lemeng Wu, Rodolfo Corona, Kevin Tai, Qixing Huang, and Raymond J Mooney. Generating animated" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12934" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "videos of human activities from natural language descriptions. Learning, 2018(1), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[27] Junfan Lin, Jianlong Chang, Lingbo Liu, Guanbin Li, Liang Lin, Qi Tian, and Chang-wen Chen. Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23222-23231, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[28] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In International Conference on Machine Learning, pages 32939-32977. PMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[29] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5442-5451, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.411 + ], + "angle": 0, + "content": "[30] Dennis Majoe, Lars Widmer, and Juerg Gutknecht. Enhanced motion interaction for multimedia applications. In Proceedings of the 7th International Conference on Advances in Mobile Computing and Multimedia, pages 13-19, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[31] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[32] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[33] Mathis Petrovich, Michael J Black, and Gül Varol. Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9488-9497, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[34] Ekkasit Pinyoanuntapong, Muhammad Usama Saleem, Pu Wang, Minwoo Lee, Srijan Das, and Chen Chen. Bamm: Bidirectional autoregressive motion model. arXiv preprint arXiv:2403.19435, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[35] Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems, 109:13-26, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Haowen Sun, Ruikun Zheng, Haibin Huang, Chongyang Ma, Hui Huang, and Ruizhen Hu. Lgtm: Local-to-global text-driven human motion diffusion model. In ACM SIGGRAPH 2024 Conference Papers, pages 1-9, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[39] Seyoon Tak and Hyeong-Seok Ko. A physically-based motion retargeting filter. ACM Trans. Graph., 24(1):98-117, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.134, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[40] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[41] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[43] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph Attention Networks. International Conference on Learning Representations, 2018. accepted as poster. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[44] Ruben Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic networks for unsupervised motion retargeting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8639-8648, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[45] Congyi Wang. T2m-hifigpt: generating high quality human motion from textual descriptions with residual discrete representations. arXiv preprint arXiv:2312.10628, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[46] Zhenyi Wang, Ping Yu, Yang Zhao, Ruiyi Zhang, Yufan Zhou, Junsong Yuan, and Changyou Chen. Learning diverse stochastic human-action generators by learning smooth latent transitions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12281-12288, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[47] Mohammed Yeasin, Ediz Polat, and Rajeev Sharma. A multiobject tracking framework for interactive multimedia applications. IEEE transactions on multimedia, 6(3):398-405, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[48] Haodong Zhang, ZhiKe Chen, Haocheng Xu, Lei Hao, Xiaofei Wu, Songcen Xu, Zhensong Zhang, Yue Wang, and Rong Xiong. Semantics-aware motion retargeting with vision-language models. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[49] Jiaxu Zhang, Junwu Weng, Di Kang, Fang Zhao, Shaoli Huang, Xuefei Zhe, Linchao Bao, Ying Shan, Jue Wang, and Zhigang Tu. Skinned motion retargeting with residual perception of motion semantics & geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13864-13872, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[50] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 1, 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[51] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116, 2023. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12935" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[53] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motionpt: Finetuned llms are general-purpose motion generators. arXiv preprint arXiv:2306.10900, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[54] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 509-519, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[55] Qiran Zou, Shangyuan Yuan, Shian Du, Yu Wang, Chang Liu, Yi Xu, Jie Chen, and Xiangyang Ji. Parco: Part-coordinating text-to-motion synthesis. arXiv preprint arXiv:2403.18512, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "12936" + } + ] +] \ No newline at end of file diff --git a/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/a12eb351-f679-4d5e-8b40-f5c8f124976d_origin.pdf b/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/a12eb351-f679-4d5e-8b40-f5c8f124976d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..84e0764867d7821fb0c042fdd6ae7f7dfc5531f8 --- /dev/null +++ b/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/a12eb351-f679-4d5e-8b40-f5c8f124976d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b5e6c80f102ba77f3dabc74a936a0ae813479de4e0538599a45dbac2a480444 +size 10597118 diff --git a/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/full.md b/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d4fb868a093c488a109f4bc32fc84eda8a4f2ce2 --- /dev/null +++ b/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/full.md @@ -0,0 +1,373 @@ +# Text-to-Any-Skeleton Motion Generation Without Retargeting + +Qingyuan Liu $^{1}$ Ke Lu $^{1,2}$ Kun Dong $^{1}$ Jian Xue $^{1*}$ Zehai Niu $^{1}$ Jinbao Wang $^{3}$ $^{1}$ University of Chinese Academy of Sciences $^{2}$ Peng Cheng Laboratory $^{3}$ Shenzhen University +{liuqingyuan23, dongkun22, niuzehai18}@mails.ucas.ac.cn, {luk, xuejian}@ucas.ac.cn +wangjb@szu.edu.cn + +![](images/0cd5482add302f1fb2103c2541798d95e98ddc2b6570060363de3a01e9babddb.jpg) +Figure 1. Introducing OmniSkel: A text-driven motion generation framework that directly generates motions for any skeletons without retargeting. Unlike traditional two-stage pipelines (top right) that require complex retargeting procedures, our single-stage approach (bottom right) only needs text input and target bone lengths (which are easily obtainable) to generate desired motions. As shown on the left, OmniSkel creates natural movements like "pushed but did not fall" and "walks up stairs with right hand on handrail" for characters with different body proportions while maintaining action fidelity. + +![](images/035bd83486ffd72d587dabd95b58552474e2a4a2f0673be8c20efcde3b6495f2.jpg) + +# Abstract + +Recent advances in text-driven motion generation have shown notable progress. However, these methods are typically limited to standardized skeletons and rely on a cumbersome targeting process to adapt to varying skeletal configurations of diverse characters. In this paper, we present OmniSkel, a novel framework that directly generates high-quality human motions for any user-defined skeleton without retargeting. Specifically, we introduce a skeleton-aware RVQ-VAE, which utilizes Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into motion encoding and reconstruction. Moreover, we propose a simple yet effective training-free approach, Motion Restoration Optimizer (MRO), to ensure zero bone length error while preserving motion smoothness. To support this research, we construct SkeleMotion3D, a large-scale text-skeleton-motion dataset based on HumanML3D. Extensive experiments demonstrate the excel + +lent robustness and generalization of our method. + +# 1. Introduction + +Recent years have witnessed remarkable progress in text-to-motion (T2M) generation, showing promising potential in applications such as gaming, metaverse, and virtual/augmented reality [10, 16, 32, 41, 50, 51]. This technology enables the intuitive creation of 3D character animations through natural language descriptions, significantly simplifying content creation workflows in animation [20], virtual reality [16], video games [30, 47], and robotics [5, 22]. + +Despite these advancements, a fundamental limitation restricts the practical application of current text-to-motion generation methods. As illustrated in Figure 1, existing approaches primarily focus on enhancing the motion quality and text-motion alignment, neglecting variations in skeletons and only generating human motions for standardized skeletons. However, real-world production environments routinely require varying skeletal configurations for diverse + +characters. To this end, practitioners must resort to skeleton retargeting - a technically demanding process that necessitates significant expertise and frequently fails to produce usable results, inevitably leading to laborious tuning and production delays. Therefore, there is a pressing demand for straightforward motion generation given arbitrary skeletons. + +To address this issue, we present OmniSkel, a novel framework that can directly generate high-quality human motion sequences for arbitrary character skeletons. As shown in Figure 1, our approach eliminates the complex retargeting stage required by traditional pipelines, offering a single-stage solution that directly produces character-specific motion without specialized expertise. Specifically, we propose a Skeleton-aware RVQ-VAE (SR-VAE) architecture, which utilizes Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into the motion encoding and reconstruction. Unlike previous approaches, our design effectively separates skeleton-invariant features, which is crucial for arbitrary skeleton adaptation. To ensure zero bone length error while preserving motion smoothness, we further propose a training-free approach, Motion Restoration Optimizer (MRO). With the above designs, our OmniSkel can deliver high-quality motion generation precisely in line with textual descriptions and character skeletons, providing robust support for practical text-to-motion applications. + +To facilitate our research, we construct SkeleMotion-3D, a large-scale dataset based on HumanML3D, pairing textual descriptions with motion sequences across diverse skeletal configurations. This dataset establishes a strong foundation for motion generation with arbitrary skeletons. + +Before delving into the details, we summarize our core contributions in this work as follows: + +- To the best of our knowledge, we present the first text-to-motion generation framework capable of directly producing motions for arbitrary character skeletons without retargeting, significantly simplifying creation workflows in real-world production environments. +- We propose a novel Skeleton-aware RVQ-VAE (SR-VAE) architecture, which employs Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into the motion encoding and reconstruction. Moreover, we further introduce Motion Reconstruction Optimizer (MRO) during inference, ensuring zero bone length error while preserving motion smoothness. +- Based on HumanML3D, we construct SkeleMotion-3D, a large-scale text-skeleton-motion dataset. This establishes a foundation for arbitrary skeleton motion generation, facilitating future research in this field. +- To verify the effectiveness of our method, we conduct extensive experiments on challenging datasets. Results demonstrate the superior robustness and generalization of + +our method. Further analysis reveals the contribution of each component to the performance improvement. + +# 2. Related Work + +# 2.1. Text-to-Motion Generation + +Text-to-motion generation has evolved significantly as a vital research area in computer animation and human-computer interaction. + +Early Approaches. Initial methods established text-to-motion mappings through alignment techniques [3, 4, 9, 26, 36, 46], but produced motions that lacked naturalism and diversity due to their deterministic nature. + +Probabilistic Frameworks. VAE-based approaches [6, 16, 28, 31, 32] model distributions of possible motions, while vision-language alignment methods [27, 33, 37, 40] leveraged pre-trained models for better semantic consistency. + +Recent Advances. Contemporary approaches include diffusion-based models [7, 10, 12, 21, 38, 41, 51, 52] generating high-fidelity motions through progressive denoising. Vector Quantized VAE (VQ-VAE) approaches treat motions as discrete token sequences, including autoregressive models [19, 45, 50, 54, 55], bidirectional methods [18, 35], and hybrid approaches [34]. Part-based generation methods [13, 54, 55] provide control over individual body components. + +Limitations. Existing approaches remain constrained to standard skeletal structures, creating a gap between research and industry needs where adaptability to various character rigs is essential. + +Our work addresses this limitation by developing a skeleton-agnostic framework that generates semantically consistent motions for arbitrary skeletal structures. + +# 2.2. Motion Retargeting + +Motion retargeting adapts existing motion data between characters with different proportions and structures. Historically, Gleicher et al. [14] pioneered this field through spacetime optimization with kinematic constraints. Following this work, Lee and Shin [23] applied inverse kinematics with B-spline curve fitting, while Choi and Ko [11] simultaneously developed an online method preserving high-frequency details. Tak and Ko [39] later enhanced these approaches by utilizing dynamics constraints for physical plausibility. + +With the advent of deep learning, new approaches emerged, including Villegas et al. [44] with unsupervised retargeting via recurrent networks, Lim et al. [25] disentangling pose and movement, and Aberman et al. [1] designing skeleton-aware networks for cross-structural retargeting. More recently, research has shifted towards preserving motion semantics. For instance, Zhang et al. [49] developed a residual retargeting network with skeleton-aware + +![](images/bcf229b8eb283236c8777696ce593b163ed5de0c0c167ae0b375369dde5c54a6.jpg) +(a) Skeleton-aware Residual VQ-VAE +(b) Masked Transformer +(c) Residual Transformer +Figure 2. Method overview. (a) Skeleton-aware Residual VQ-VAE: our framework processes source motion and skeleton through K-GCA (Kinematic Graph Cross Attention), which divides the input into Q, K, and V branches, creating a hierarchical representation with motion-invariant features. The residual quantization approach progressively captures motion details across multiple layers, enabling high-fidelity reconstruction for arbitrary target skeletons. (b) Masked Transformer: base-layer motion tokens $t^0$ are randomly masked during training and the text-conditioned transformer learns to predict these tokens simultaneously, capturing fundamental motion patterns. (c) Residual Transformer: builds upon the base layer by progressively predicting higher-level residual tokens $t^v$ conditioned on both text and previous layer tokens, enhancing motion fidelity and skeleton-specific details. + +and shape-aware modules. Similarly, Lee et al. [24] proposed Skeleton-Agnostic Motion Embedding (SAME) using graph convolutional networks to separate skeleton information from motion while preserving semantics. Building upon these semantic approaches, Zhang et al. [48] introduced Semantics-aware Motion reTargeting (SMT), leveraging vision-language models to extract and maintain motion semantics through a two-stage pipeline with skeleton-aware pre-training and semantic constraints. + +However, despite this significant progress, many approaches still struggle with maintaining semantic consistency across dramatically different skeletal structures. + +# 3. Approach + +Our goal is to generate motion sequences $\mathbf{M}_{1:T}$ of length $T$ conditioned on both a text description $c$ and a target skeletal structure $\mathbf{S}$ , where $\mathbf{M}_t\in \mathbb{R}^{J\times D_{\mathrm{m}}}$ with $J$ denoting the number of joints and $D_{\mathrm{m}}$ representing the dimension of joint features. The skeletal structure $\mathbf{S}\in \mathbb{R}^{J\times D_{\mathrm{s}}}$ characterizes the static joint properties of the target character. As illustrated in Figure 2, our framework consists of two main components: a skeleton-aware VAE module that enables motion generation for arbitrary skeletal structures (Section 3.2), and a cooperative transformer architecture for motion token prediction (Section 3.3). We first introduce our newly proposed dataset and its representation (Section 3.1) to facilitate better understanding of our method. Then, we detail + +the complete training pipeline (Section 3.3) and inference process (Section 3.4). + +# 3.1. SkeleMotion-3D Dataset + +Our dataset comprises three key components: textual descriptions, motion sequences, and corresponding skeletal structures. Building upon the HumanML3D dataset [16], which combines motion data from AMASS [29] and HumanAct12 [15], we significantly expand the motion-skeleton pairs through a skeleton randomization process inspired by [24]. + +Feature Representation. We represent motion sequences as $\mathbf{M} \in \mathbb{R}^{T \times J \times D_{\mathrm{m}}}$ , where the motion feature $F_{\mathrm{motion}}$ with dimension $D_{\mathrm{m}}$ consists of: + +$$ +F _ {\text {m o t i o n}} = \left[ F _ {\text {t o p o}}, F _ {\text {r o o t}}, F _ {\text {x y z}}, F _ {\text {r o t}}, F _ {\text {v e l}} \right], \tag {1} +$$ + +comprising skeletal topology features, root joint features, relative spatial positions, 6D rotation features, and velocity features. + +The skeletal structures are encoded as $\mathbf{S} \in \mathbb{R}^{J \times D_{\mathrm{s}}}$ , where the skeletal feature $F_{\mathrm{skel}}$ with dimension $D_{\mathrm{s}}$ consists of: + +$$ +F _ {\mathrm {s k e l}} = \left[ F _ {\mathrm {t o p o}}, F _ {\mathrm {s t a t i c}} \right], \tag {2} +$$ + +where $F_{\mathrm{topo}}$ represents topology features and $F_{\mathrm{static}}$ denotes static skeletal properties. + +Dataset Construction. Following the preprocessing pipeline established by HumanML3D, we standardize the + +raw motion sequences through temporal normalization (20 FPS), sequence cropping (maximum 10 seconds), skeletal retargeting, and orientation alignment $(\mathrm{Z}+$ direction). The text annotations from HumanML3D, collected via Amazon Mechanical Turk with quality control measures, are preserved in our dataset. + +Dataset Statistics. The SkeleMotion-3D dataset contains 14,616 distinct motion patterns, 29,233 different skeletal configurations, 43,848 unique text-motion-skeleton triplets, and 44,970 textual descriptions with 5,371 unique words. This represents a substantial increase compared to the training dataset (949 motion patterns and 4,745 motion-skeleton pairs) used in [24]'s motion retargeting network. The significant expansion enhances the robustness and generalization capability of our skeleton-aware motion generation model. + +# 3.2. Skeleton-aware Residual VQ-VAE + +We propose Skeleton-aware Residual VQ-VAE (SR-VAE), a novel architecture that extends the conventional Residual VQ-VAE to enable skeleton-conditioned motion generation. Conventional motion VQ-VAEs [17, 19, 50, 53] and recent RVQ-VAEs [18, 45] focus on reconstructing source motions with identical skeletal structures. In contrast, as illustrated in Figure 2, SR-VAE adopts an asymmetric VAE architecture, consisting of an encoder, a decoder, a residual quantization module, and a Kinetic Graph Cross Attention (K-GCA) module, enabling motion reconstruction for arbitrary target skeletons. + +Architecture Design. Our SR-VAE processes motion generation through three main stages. First, the input motion sequence $\mathbf{M}_{\mathrm{in}} \in \mathbb{R}^{T \times J \times D_{\mathrm{m}}}$ and its corresponding skeleton structure $\mathbf{S}_{\mathrm{in}} \in \mathbb{R}^{J \times D_{\mathrm{s}}}$ are fused via K-GCA before being encoded into latent features. Following MoMask [18], these features undergo residual quantization to produce $V + 1$ ordered code sequences. Finally, the decoder incorporates target skeleton features $\mathbf{S}_{\mathrm{tgt}} \in \mathbb{R}^{J \times D_{\mathrm{s}}}$ through multiple K-GCA layers at different decoding stages to generate the target motion $\mathbf{M}_{\mathrm{tgt}} \in \mathbb{R}^{T \times J \times D_{\mathrm{m}}}$ . + +Let $\hat{\mathbf{M}}_{\mathrm{tgt}} \in \mathbb{R}^{T \times J \times D_{\mathrm{m}}}$ denote the predicted motion sequence and $\mathbf{M}_{\mathrm{tgt}}$ denote the ground truth target motion. For residual quantization, let $\mathbf{r}^v \in \mathbb{R}^{T \times d}$ represent the residual features at layer $v \in \{0,1,\dots,V\}$ and $\mathbf{b}^v \in \mathbb{R}^{T \times d}$ denote the corresponding quantized codes, where $d$ is the dimension of the latent representation. Additionally, let $\mathbf{D}_{\mathrm{xyz}} \in \mathbb{R}^{T \times J \times 3}$ and $\hat{\mathbf{D}}_{\mathrm{xyz}} \in \mathbb{R}^{T \times J \times 3}$ represent the ground truth and predicted joint positions in 3D space, respectively. The overall training objective consists of three terms: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \left\| \mathbf {M} _ {\mathrm {t g t}} - \hat {\mathbf {M}} _ {\mathrm {t g t}} \right\| _ {1} + \beta \sum_ {v = 0} ^ {V} \left\| \mathbf {r} ^ {v} - \operatorname {s g} \left[ \mathbf {b} ^ {v} \right] \right\| _ {2} ^ {2} + \lambda \left\| \mathbf {D} _ {\mathrm {x y z}} - \hat {\mathbf {D}} _ {\mathrm {x y z}} \right\| _ {1} \tag {3} +$$ + +where the first term is motion reconstruction loss, the second term is commitment loss for residual quantization, and the third term specifically focuses on joint positions accu + +![](images/548a422267deed8ed1c2be045bca439657900ab4efadc69eb8ac7262e605e982.jpg) +Figure 3. Architecture of our Kinetic Graph Cross Attention. + +racy. Here, $\mathrm{sg}[\cdot ]$ denotes the stop-gradient operation, $\beta$ and $\lambda$ are weighting factors for the commitment loss and joint position loss, respectively. The loss function is optimized via a straight-through gradient estimator [42]. Following [18, 35, 50], we employ exponential moving average for codebook updates and reset. + +Kinetic Graph Cross Attention. To effectively fuse motion and skeletal features, we propose Kinetic Graph Cross Attention (K-GCA), as illustrated in Figure 3, extending the traditional Graph Attention Networks (GAT) [8, 43] for motion-skeleton feature fusion. Unlike standard GAT that performs self-attention within a single feature type, K-GCA enables cross-modal attention between motion and skeletal features while preserving the skeletal graph structure. + +Given motion features $\mathbf{X} \in \mathbb{R}^{T \times J \times D_{\mathrm{m}}}$ and skeleton features $\mathbf{S} \in \mathbb{R}^{J \times D_{\mathrm{s}}}$ , K-GCA first projects them into a shared space using learnable projection matrices $\mathbf{W}_{\mathrm{m}} \in \mathbb{R}^{D_{\mathrm{m}} \times d_k}$ and $\mathbf{W}_{\mathrm{s}} \in \mathbb{R}^{D_{\mathrm{s}} \times d_k}$ , where $d_k$ is the dimension of the projected space: + +$$ +\mathbf {Q} = \mathbf {W} _ {\mathrm {m}} \mathbf {X}, \quad \mathbf {K} = \mathbf {V} = \mathbf {W} _ {\mathrm {s}} \mathbf {S}. \tag {4} +$$ + +The attention scores between joints $j$ and $k$ are computed using learnable vectors $\mathbf{a}^{\mathrm{src}}\in \mathbb{R}^{d_k}$ and $\mathbf{a}^{\mathrm{dst}}\in \mathbb{R}^{d_k}$ : + +$$ +e _ {j k} = \text {L e a k y R e L U} \left(\mathbf {a} ^ {\mathrm {s r c}} \cdot \mathbf {Q} _ {j} + \mathbf {a} ^ {\mathrm {d s t}} \cdot \mathbf {K} _ {k}\right), \tag {5} +$$ + +To preserve the skeletal structure, we mask the attention scores using the skeleton adjacency matrix $\mathbf{A} \in \{0,1\}^{J \times J}$ , where $\mathbf{A}_{jk} = 1$ if joints $j$ and $k$ are connected in the skeleton graph, and 0 otherwise. We set attention scores for disconnected joints to negative infinity before applying softmax: + +$$ +\tilde {e} _ {j k} = \left\{ \begin{array}{l l} e _ {j k}, & \text {i f} \mathbf {A} _ {j k} = 1 \\ - \infty , & \text {i f} \mathbf {A} _ {j k} = 0 \end{array} \right., \tag {6} +$$ + +$$ +\alpha_ {j k} = \frac {\exp \left(\tilde {e} _ {j k}\right)}{\sum_ {l = 1} ^ {J} \exp \left(\tilde {e} _ {j l}\right)}, \tag {7} +$$ + +The final output $\hat{\mathbf{X}}\in \mathbb{R}^{T\times J\times D_{\mathrm{m}}}$ is computed by applying the attention weights to value matrix $\mathbf{V}$ in each head, concatenating results from all heads, and adding back to the input through residual connection: + +$$ +\hat {\mathbf {X}} = \operatorname {L a y e r N o r m} (\mathbf {X} + \operatorname {M u l t i H e a d} (\alpha , \mathbf {V})). \tag {8} +$$ + +where MultiHead $(\alpha, \mathbf{V})$ represents the multi-head weighted combination of values using the computed attention weights $\alpha$ . This design allows K-GCA to capture dynamic relationships between motion and skeletal features while maintaining the anatomical constraints imposed by the skeletal structure. Compared to standard GAT, our approach provides two key advantages: (1) the ability to perform cross-modal attention between motion and skeletal features, and (2) the incorporation of skeletal structure constraints through attention masking. + +# 3.3. Training Pipeline + +Our training pipeline follows a sequential approach for skeleton-agnostic motion generation from text, as shown in Figure 2. + +SR-VAE Training. First, we train the Skeleton-aware Residual VQ-VAE (SR-VAE) to establish skeleton-agnostic motion representation. This model encodes motion sequences into hierarchical tokens and decodes them back to motion features, independent of skeletal structure. After training, both encoder $\mathcal{E}$ and decoder $\mathcal{D}$ are frozen. + +Tokenization of Training Data. Using the frozen SR-VAE encoder, we convert all training motion sequences into hierarchical token representations. For each motion $\mathbf{M}_{\mathrm{in}}$ with corresponding skeleton $\mathbf{S}_{\mathrm{in}}$ , we extract multi-level token sequences $\mathbf{t}^{0:V}$ as targets for our text-to-token models. + +Masked Transformer Training. Inspired by MoMask [18], we implement a bidirectional Masked Transformer to generate base-layer tokens $(\mathbf{t}^0)$ . This design offers parallel generation capability and better global motion coherence compared to autoregressive approaches. During training, the model learns to predict original tokens at masked positions from text prompts. + +Residual Transformer Training. Finally, we train the Residual Transformer to predict higher-layer tokens $(\mathbf{t}^{1:V})$ encoding motion refinements. Following MoMask [18], this transformer takes both text descriptions and tokens from previous layers as input, focusing on finer motion details while maintaining consistency with the base motion. + +# 3.4. Inference Process + +During inference, OmniSkel generates skeleton-aware motions through the following steps: + +Base Token Generation: Given a text prompt, the Masked Transformer generates the base-layer tokens $\hat{\mathbf{t}}^0$ that capture primary motion patterns. + +Hierarchical Refinement: The Residual Transformer progressively generates higher-layer tokens $\hat{\mathbf{t}}^{1:V}$ , starting with layer 1 and proceeding sequentially. Each layer's prediction utilizes both the text prompt and all previously generated token layers. + +Motion Decoding: Once we have the complete token hierarchy $\hat{\mathbf{t}}^{0:V}$ , we feed these tokens along with the target skele + +ton structure $\mathbf{S}_{\mathrm{tgt}}$ into the frozen SR-VAE decoder to obtain motion features. + +Algorithm 1: Motion Restoration Optimizer +```txt +Input: Measurement data: $F_{\mathrm{topo}}$ , $F_{\mathrm{root}}$ , $F_{\mathrm{xyz}}$ ; Bone lengths: $S$ +Output: Smooth motion sequence $M_{\mathrm{smooth}}$ with guaranteed skeletal consistency +``` + +1 // Step 1: Restore initial motion using position-based approach; +```txt +2 $M_{\mathrm{ric}} \gets \mathrm{ric}(F_{\mathrm{topo}}, F_{\mathrm{root}}, F_{\mathrm{xyz}})$ ; +``` + +```latex +3 $M_{\mathrm{smooth}} \gets \emptyset$ +``` + +4 foreach frame $f\in M_{\mathrm{ric}}$ do +```txt +// Step 2: Extract unit direction vectors for each bone (except the root); +Initialize set $\mathcal{U} \gets \emptyset$ ; +for $i \gets 1$ to $J - 1$ do +Let $\mathbf{p}_i$ be the position of joint $i$ in frame $f$ ; +Let $\mathbf{p}_{\text{parent}(i)}$ be the position of joint $i$ 's parent; +Compute displacement: $\mathbf{v}_i \gets \mathbf{p}_i - \mathbf{p}_{\text{parent}(i)}$ ; +Compute unit direction: $\mathbf{u}_i \gets \frac{\mathbf{v}_i}{|\mathbf{v}_i|}$ ; +Add $\mathbf{u}_i$ to $\mathcal{U}$ ; +end +// Step 3: Reconstruct the current frame using extracted directions, bone lengths, and topology; + $f_{\text{rec}} \gets \text{ReconstructFrame}(\mathcal{U}, S, F_{\text{topo}}, F_{\text{root}})$ ; +Add $f_{\text{rec}}$ to $M_{\text{smooth}}$ ; +end +return $M_{\text{smooth}}$ ; +``` + +Feature-to-Motion Conversion: Following the generation of motion features by our SR-VAE decoder, we face the challenge of transforming these abstract representations into production-ready motion data. Contemporary text-to-motion frameworks typically rely on either position-based restoration (offering fluidity but with bone length inconsistencies) or rotation-based restoration (maintaining skeletal integrity but introducing jitter and penetration artifacts). + +To overcome these limitations, we employ our Motion Restoration Optimizer (MRO, detailed in Algorithm 1), which synergistically combines the strengths of both approaches. MRO guarantees two critical properties simultaneously: absolute preservation of skeletal proportions (zero bone length error) and maintenance of smooth trajectory characteristics, producing physically plausible output $\hat{\mathbf{M}}_{\mathrm{tgt}}$ . + +
MethodFID↓R Precision↑MPJPE↓Skeleton Error↓Jitter Error↓Skating Ratio ↓
Top 1Top 2Top 3
SR-VAE (full model)0.0270.5080.7000.7930.03513.7781.8240.074
w/ Single-head K-GCA0.0400.4970.6850.7890.03412.2942.9350.061
w/o K-GCA0.0650.4950.6890.7850.04117.7269.5940.092
w/o RVQ0.1650.4670.6390.7750.06119.7263.5940.092
w/ Symmetric Reconstruction0.8850.4500.6410.7450.09789.7479.8670.067
w/ Rotation-based Restoration0.0270.5080.7000.7930.0630.0030.7240.193
w/ MRO0.0270.5080.7000.7930.0350.002.0840.090
+ +Table 1. Ablation study of SR-VAE components tested on the SkeleMotion-3D test set. Our SR-VAE is trained on the SkeleMotion-3D dataset. To evaluate SR-VAE's reconstruction performance for arbitrary skeletal motions, we input random skeleton motions to the encoder and use the standard skeleton from the HumanML3D dataset as the target skeleton. + +# 4. Experiments + +We conduct extensive experiments to evaluate our approach across multiple dimensions, focusing on both semantic accuracy and physical plausibility. + +Evaluation Metrics. We assess our model using established semantic metrics from prior work [16]: (1) Frechet Inception Distance (FID) for overall motion quality; (2) $R$ -Precision and multimodal distance for text-motion alignment. + +To comprehensively evaluate physical plausibility—a critical aspect for animation applications—we introduce several specialized metrics: + +- Skating Ratio: Measures foot sliding artifacts where feet unnaturally glide across the ground +- Skeleton Error: Quantifies bone length variations throughout motion sequences +- MPJPE (Mean Per-Joint Position Error): Assesses positional accuracy of joints +- Jitter Error: Evaluates motion continuity and natural flow. Detailed formulations of these metrics appear in the supplementary materials. + +Implementation Details. Our PyTorch implementation features a Skeleton-aware Residual VQ-VAE (SR-VAE) with residual blocks following T2M-GPT [50] and MoMask [18] architectures. We use a downscale factor of 4, with a residual quantization comprising 6 layers (each containing 512 codes of 512 dimensions) and quantization dropout ratio of 0.2. For the SR-VAE training objective, we set commitment loss weight $\beta$ to 0.02 and joint position loss weight $\lambda$ to 2.0. The K-GCA module employs 2-head attention. + +Both transformer models consist of 6 layers with 6 attention heads and 384-dimensional latent representations, trained on our SkeleMotion-3D dataset. We use a learning rate of 2e-4 with linear warm-up over 2000 iterations, batch sizes of 512 for SR-VAE and 64 for transformers. During inference, we apply classifier-free guidance (scale 4 for + +masked transformer, 5 for refinement transformer) with 10 iterative decoding steps. All experiments run on NVIDIA RTX 4090 GPUs. + +# 4.1. Skeleton-Aware Disentanglement Analysis + +A cornerstone of our approach is SR-VAE's ability to disentangle motion patterns from skeletal structures. We conduct a targeted experiment to validate this skeleton-agnostic capability. + +Figure 4 presents our analysis of four distinct motion patterns, each performed by three different skeletal structures with varying proportions. We process these motion-skeleton pairs through our SR-VAE encoder $\mathcal{E}$ and visualize the resulting base layer tokens $\mathbf{t}^0$ on aligned line charts. + +The results reveal that motions with identical semantic patterns but executed by anatomically diverse skeletons produce remarkably similar token sequences. This consistency demonstrates that our encoder $\mathcal{E}$ successfully extracts skeleton-invariant motion features, preserving essential semantic content regardless of the executing skeletal structure. + +This confirms that SR-VAE effectively decomposes motion into two orthogonal components: (1) skeleton-invariant pattern information encoded in tokens, and (2) skeleton-specific structural information provided separately to the decoder. This disentanglement enables our model to generate physically plausible animations for arbitrary skeletal structures while preserving intended motion semantics. + +# 4.2. Ablation Study + +To validate SR-VAE's skeleton-aware capabilities, we conduct experiments using random input skeletons while decoding to standard HumanML3D [16] skeletons. This evaluates both semantic alignment and cross-skeleton reconstruction capabilities. + +Table 1 presents our ablation results on the SkeleMotion-3D test set. The complete SR-VAE model achieves superior performance with an FID of 0.026 and strong R + +![](images/b812883ec6b3350c3f600d7743fa6720c62b8839b3af77d3b9c2eddcbeb9abd9.jpg) +Figure 4. Motion token analysis across skeletal structures. This figure demonstrates token sequences generated by the SR-VAE encoder for four distinct motion patterns, each performed by three skeletal structures with different proportions. The consistent token patterns across each motion type (columns) despite skeletal variations demonstrates our encoder's ability to extract skeleton-invariant motion features. Top row shows skeletal visualizations while bottom row displays the corresponding quantized token values, revealing clear pattern preservation across different anatomical configurations. + +![](images/a793afe065b0353315d59d1958cec69c529528111794b872f152209c4e887f07.jpg) +Figure 5. Text-to-motion generation across different skeletal structures. Our OmniSkel framework generates diverse motions responding to different text prompts (columns) while adapting to different skeletal anatomies (rows). The human skeleton on the left is randomly selected from Mixamo [2], demonstrating OmniSkel's ability to maintain semantic consistency across dramatically different skeletal structures without retargeting. + +Precision scores (0.508/0.700/0.793 for Top-1/2/3), demonstrating excellent semantic consistency. + +Architecture Components. Using symmetric reconstruction (same skeleton for encoder and decoder) drastically increases FID to 0.885 and skeleton error to 89.747, confirming that asymmetric processing is essential for effective cross-skeleton learning. Removing K-GCA degrades FID to 0.065 with a significant increase in jitter error (9.594 vs. 1.824), demonstrating its importance in capturing cross-skeleton geometric correspondences. + +Without RVQ, FID increases substantially to 0.165 and MPJPE to 0.061, highlighting RVQ's crucial role in maintaining semantic consistency across different skeletal struc + +tures. The single-head K-GCA variant shows moderate performance decline (FID: 0.040), indicating that multi-head attention more effectively captures the rich geometric relationships between skeletons. + +Motion Restoration. For motion restoration approaches, rotation-based methods achieve perfect skeleton reconstruction (skeleton error: 0.00) but significantly increase jitter error (30.724) and skating performance (0.193). In contrast, our MRO technique also achieves perfect skeleton reconstruction (skeleton error: 0.00) while maintaining low jitter error (2.084) and better skating ratio (0.090). This demonstrates MRO's effectiveness in preserving physical plausibility while maintaining motion quality across differ- + +ent skeletal structures. + +Our ablation study confirms each component's critical contribution to the overall system: K-GCA and RVQ enhance cross-skeleton reconstruction and semantic alignment, while MRO ensures physical consistency in the generated motions. These findings validate our architectural design choices for skeleton-aware motion generation and demonstrate SR-VAE's ability to effectively handle arbitrary skeleton motion retargeting with high fidelity. + +# 4.3. Qualitative and Quantitative Analysis + +Quantitative Evaluation. We evaluate our methods against state-of-the-art text-to-motion approaches as shown in Table 2. The fundamental distinction of our approach is the ability to directly generate motions for arbitrary skeleton structures—a paradigm shift from existing methods that are restricted to fixed, pre-defined skeletons. + +
DatasetsMethodsFID↓R Precision↑
Top 1Top 2Top 3
Human-ML3DTM2T [17]1.501±.0170.424±.0030.618±.0030.729±.002
T2M [16]1.087±.0210.455±.0030.636±.0030.736±.002
MDM [41]0.544±.0440.320±.0050.498±.0040.611±.007
MLD [10]0.473±.0130.481±.0030.673±.0030.772±.002
MotionDiffuse [51]0.630±.0010.491±.0010.681±.0010.782±.001
T2M-GPT [50]0.141±.0050.492±.0030.679±.0020.775±.002
ReMoDiffuse [52]0.103±.0040.510±.0050.698±.0060.795±.004
AttT2M [54]0.112±.0060.499±.0030.690±.0020.786±.002
Momask [18]0.045±.0020.521±.0020.713±.0020.807±.002
Skele-Motion-3D OmniSkel (ours)0.027±.0000.508±.0020.700±.0020.793±.002
0.084±.0040.507±.0020.699±.0030.797±.002
+ +Table 2. Comparison with text-to-motion methods on standard skeleton. Our approach introduces a novel pipeline trained on the SkeleMotion-3D dataset, which fundamentally differs from traditional T2M methods. For fair comparison with existing methods trained on HumanML3D, we configured our models (SR-VAE and OmniSkel) to target the standard HumanML3D skeleton during inference. Despite being designed for arbitrary skeleton generation, our methods achieve competitive performance on this fixed skeleton setting, demonstrating strong text-to-motion capabilities even when constrained to a single skeleton structure. + +This fundamental difference presents a challenge for direct comparison, as our methods operate on the newly introduced SkeleMotion-3D dataset with diverse skeleton structures, while previous approaches are confined to the HumanML3D dataset with a single skeleton configuration. Notably, existing methods cannot be trained or evaluated on our dataset due to their inherent architectural limitations in handling skeleton variations. + +To provide a meaningful comparison that highlights our text-to-motion capabilities, we configure our models at inference time to target the standard HumanML3D skeleton structure. Despite being designed for the more complex task of arbitrary skeleton generation, both SR-VAE + +and OmniSkel achieve competitive performance on the HumanML3D test set, with comparable or superior metrics in FID and R-Precision compared to specialized methods. + +These results demonstrate that our methods successfully balance adaptability with generation quality, enabling cross-skeleton applications without compromising performance on standard benchmarks. + +Qualitative Results. Figure 5 illustrates OmniSkel's capability to generate motions across different skeletal structures. The visualization demonstrates: (1) diverse motions for different textual descriptions, and (2) consistent semantic motion across distinct skeletal anatomies. + +Each column represents a different textual prompt, from complex sequences to simpler actions. The model successfully captures the nuanced details of these descriptions, including locomotion styles and sequential actions. + +This comparison demonstrates OmniSkel's precise perception and effective integration of different semantics and arbitrary skeletons. More visualization results are provided in the supplementary materials. + +# 4.4. Performance Comparison + +
MethodParametersInference TimeGPU Memory
MoMask [18]44.85M0.075s1434M
Ours38.73M0.099s1402M
+ +Table 3. Computational cost comparison on RTX 4090. + +Our method achieves comparable performance to state-of-the-art approaches while maintaining slight GPU memory advantages as shown in Table 3 and uniquely enabling end-to-end motion generation for arbitrary skeletal specifications. This efficiency demonstrates the practical viability of our skeleton-agnostic framework for real-world applications requiring diverse character animations. + +# 5. Discussion and Conclusion + +In conclusion, we present OmniSkel, a text-driven motion generation framework that enables motion synthesis for arbitrary human skeletons through our novel K-GCA and SR-VAE architectures. Our Motion Restoration Optimizer (MRO) addresses the cross-skeleton motion reconstruction challenge, ensuring both motion smoothness and skeletal consistency. Leveraging our SkeleMotion-3D dataset, which pairs textual descriptions with diverse skeletal configurations, our approach bridges the gap between research and practical applications. Experiments demonstrate that OmniSkel achieves competitive performance while maintaining semantic alignment with textual descriptions. This work establishes a foundation for more inclusive motion generation systems that can serve diverse applications without requiring conventional retargeting approaches. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China under Grant 62320106007 and Grant 62032022; and in part by the Guangdong Provincial Key Laboratory under Grant 2023B1212060076. + +# References + +[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2 +[2] Adobe. Mixamo. Online Animation Library, 2023. Available at: https://www MIXamo.com/.7 +[3] Hyemin Ahn, Timothy Ha, Yunho Choi, Hwiyeon Yoo, and Songhwai Oh. Text2action: Generative adversarial synthesis from language to action. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5915-5920. IEEE, 2018. 2 +[4] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719–728. IEEE, 2019. 2 +[5] André Antakli, Erik Hermann, Ingo Zinnikus, Han Du, and Klaus Fischer. Intelligent distributed human motion simulation in human-robot collaboration environments. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 319–324, 2018. 1 +[6] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 2 +[7] Gustav Bredell, Kyriakos Flouris, Krishna Chaitanya, Ertunc Erdil, and Ender Konukoglu. Explicitly minimizing the blur error of variational autoencoders. arXiv preprint arXiv:2304.05939, 2023. 2 +[8] Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks? In International Conference on Learning Representations, 2022. 4 +[9] Haoye Cai, Chunyan Bai, Yu-Wing Tai, and Chi-Keung Tang. Deep video generation, prediction and completion of human action sequences. In Proceedings of the European Conference on Computer Vision (ECCV), pages 366–382, 2018. 2 +[10] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 1, 2, 8 +[11] Kwang-Jin Choi and Hyeong-Seok Ko. Online motion retargeting. The Journal of Visualization and Computer Animation, 11(5):223-235, 2000. 2 +[12] Xuehao Gao, Yang Yang, Zhenyu Xie, Shaoyi Du, Zhongqian Sun, and Yang Wu. Guess: Gradually enriching synthesis for text-driven human motion generation. IEEE Transactions on Visualization and Computer Graphics, 2024. 2 + +[13] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1396-1406, 2021. 2 +[14] Michael Gleicher. Retargetting motion to new characters. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, pages 33-42, 1998. 2 +[15] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 3 +[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 1, 2, 3, 6, 8 +[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 4, 8 +[18] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 2, 4, 5, 6, 8 +[19] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2, 4 +[20] Moritz Kappel, Vladislav Golyanik, Mohamed Elgharib, Jann-Ole Henningson, Hans-Peter Seidel, Susana Castillo, Christian Theobalt, and Marcus Magnor. High-fidelity neural human motion transfer from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1541-1550, 2021. 1 +[21] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 2 +[22] Hema Koppula and Ashutosh Saxena. Learning spatiotemporal structure from rgb-d videos for human activity detection and anticipation. In International conference on machine learning, pages 792-800. PMLR, 2013. 1 +[23] Jehee Lee and Sung Yong Shin. A hierarchical approach to interactive motion editing for human-like figures. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 39-48, 1999. 2 +[24] Sunmin Lee, Taeho Kang, Jungnam Park, Jehee Lee, and Jungdam Won. Same: Skeleton-agnostic motion embedding for character animation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3, 4 +[25] Jongin Lim, Hyung Jin Chang, and Jin Young Choi. PMnet: learning of disentangled pose and movement for unsupervised motion retargeting. In Proc. BMVC, 2019. 2 +[26] Angela S Lin, Lemeng Wu, Rodolfo Corona, Kevin Tai, Qixing Huang, and Raymond J Mooney. Generating animated + +videos of human activities from natural language descriptions. Learning, 2018(1), 2018. 2 +[27] Junfan Lin, Jianlong Chang, Lingbo Liu, Guanbin Li, Liang Lin, Qi Tian, and Chang-wen Chen. Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23222-23231, 2023. 2 +[28] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In International Conference on Machine Learning, pages 32939-32977. PMLR, 2024. 2 +[29] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5442-5451, 2019. 3 +[30] Dennis Majoe, Lars Widmer, and Juerg Gutknecht. Enhanced motion interaction for multimedia applications. In Proceedings of the 7th International Conference on Advances in Mobile Computing and Multimedia, pages 13-19, 2009. 1 +[31] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2 +[32] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 1, 2 +[33] Mathis Petrovich, Michael J Black, and Gül Varol. Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9488-9497, 2023. 2 +[34] Ekkasit Pinyoanuntapong, Muhammad Usama Saleem, Pu Wang, Minwoo Lee, Srijan Das, and Chen Chen. Bamm: Bidirectional autoregressive motion model. arXiv preprint arXiv:2403.19435, 2024. 2 +[35] Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 4 +[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems, 109:13-26, 2018. 2 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[38] Haowen Sun, Ruikun Zheng, Haibin Huang, Chongyang Ma, Hui Huang, and Ruizhen Hu. Lgtm: Local-to-global text-driven human motion diffusion model. In ACM SIGGRAPH 2024 Conference Papers, pages 1-9, 2024. 2 + +[39] Seyoon Tak and Hyeong-Seok Ko. A physically-based motion retargeting filter. ACM Trans. Graph., 24(1):98-117, 2005. 2 +[40] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2 +[41] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 1, 2, 8 +[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems, 30, 2017. 4 +[43] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph Attention Networks. International Conference on Learning Representations, 2018. accepted as poster. 4 +[44] Ruben Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic networks for unsupervised motion retargeting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8639-8648, 2018. 2 +[45] Congyi Wang. T2m-hifigpt: generating high quality human motion from textual descriptions with residual discrete representations. arXiv preprint arXiv:2312.10628, 2023. 2, 4 +[46] Zhenyi Wang, Ping Yu, Yang Zhao, Ruiyi Zhang, Yufan Zhou, Junsong Yuan, and Changyou Chen. Learning diverse stochastic human-action generators by learning smooth latent transitions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12281-12288, 2020. 2 +[47] Mohammed Yeasin, Ediz Polat, and Rajeev Sharma. A multiobject tracking framework for interactive multimedia applications. IEEE transactions on multimedia, 6(3):398-405, 2004. 1 +[48] Haodong Zhang, ZhiKe Chen, Haocheng Xu, Lei Hao, Xiaofei Wu, Songcen Xu, Zhensong Zhang, Yue Wang, and Rong Xiong. Semantics-aware motion retargeting with vision-language models. 2024. 3 +[49] Jiaxu Zhang, Junwu Weng, Di Kang, Fang Zhao, Shaoli Huang, Xuefei Zhe, Linchao Bao, Ying Shan, Jue Wang, and Zhigang Tu. Skinned motion retargeting with residual perception of motion semantics & geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13864-13872, 2023. 2 +[50] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 1, 2, 4, 6, 8 +[51] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 1, 2, 8 +[52] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116, 2023. 2, 8 + +[53] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motionpt: Finetuned llms are general-purpose motion generators. arXiv preprint arXiv:2306.10900, 2023. 4 +[54] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 509-519, 2023. 2, 8 +[55] Qiran Zou, Shangyuan Yuan, Shian Du, Yu Wang, Chang Liu, Yi Xu, Jie Chen, and Xiangyang Ji. Parco: Part-coordinating text-to-motion synthesis. arXiv preprint arXiv:2403.18512, 2024. 2 \ No newline at end of file diff --git a/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/images.zip b/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..030ee3e170a985cd25aa5ec90bcab29bd3f74d52 --- /dev/null +++ b/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea448779066cc2fb667b02720ccbff9ea1d4f5c1e42f1ca6419e4e88599712e5 +size 459248 diff --git a/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/layout.json b/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bb4266903331a9ac34f7679df21373bb80f96738 --- /dev/null +++ b/2025/Text-to-Any-Skeleton Motion Generation Without Retargeting/layout.json @@ -0,0 +1,9535 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 112, + 103, + 499, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 103, + 499, + 121 + ], + "spans": [ + { + "bbox": [ + 112, + 103, + 499, + 121 + ], + "type": "text", + "content": "Text-to-Any-Skeleton Motion Generation Without Retargeting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "spans": [ + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "text", + "content": "Qingyuan Liu" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "text", + "content": " Ke Lu" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "text", + "content": " Kun Dong" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "text", + "content": " Jian Xue" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "text", + "content": " Zehai Niu" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "text", + "content": " Jinbao Wang" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences " + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "text", + "content": "Peng Cheng Laboratory " + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 142, + 532, + 199 + ], + "type": "text", + "content": "Shenzhen University \n{liuqingyuan23, dongkun22, niuzehai18}@mails.ucas.ac.cn, {luk, xuejian}@ucas.ac.cn \nwangjb@szu.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 228, + 333, + 375 + ], + "blocks": [ + { + "bbox": [ + 60, + 228, + 333, + 375 + ], + "lines": [ + { + "bbox": [ + 60, + 228, + 333, + 375 + ], + "spans": [ + { + "bbox": [ + 60, + 228, + 333, + 375 + ], + "type": "image", + "image_path": "0cd5482add302f1fb2103c2541798d95e98ddc2b6570060363de3a01e9babddb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 388, + 555, + 443 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 555, + 443 + ], + "type": "text", + "content": "Figure 1. Introducing OmniSkel: A text-driven motion generation framework that directly generates motions for any skeletons without retargeting. Unlike traditional two-stage pipelines (top right) that require complex retargeting procedures, our single-stage approach (bottom right) only needs text input and target bone lengths (which are easily obtainable) to generate desired motions. As shown on the left, OmniSkel creates natural movements like \"pushed but did not fall\" and \"walks up stairs with right hand on handrail\" for characters with different body proportions while maintaining action fidelity." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 336, + 228, + 545, + 375 + ], + "blocks": [ + { + "bbox": [ + 336, + 228, + 545, + 375 + ], + "lines": [ + { + "bbox": [ + 336, + 228, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 336, + 228, + 545, + 375 + ], + "type": "image", + "image_path": "035bd83486ffd72d587dabd95b58552474e2a4a2f0673be8c20efcde3b6495f2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 453, + 200, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 453, + 200, + 465 + ], + "spans": [ + { + "bbox": [ + 151, + 453, + 200, + 465 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 481, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 481, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 481, + 297, + 685 + ], + "type": "text", + "content": "Recent advances in text-driven motion generation have shown notable progress. However, these methods are typically limited to standardized skeletons and rely on a cumbersome targeting process to adapt to varying skeletal configurations of diverse characters. In this paper, we present OmniSkel, a novel framework that directly generates high-quality human motions for any user-defined skeleton without retargeting. Specifically, we introduce a skeleton-aware RVQ-VAE, which utilizes Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into motion encoding and reconstruction. Moreover, we propose a simple yet effective training-free approach, Motion Restoration Optimizer (MRO), to ensure zero bone length error while preserving motion smoothness. To support this research, we construct SkeleMotion3D, a large-scale text-skeleton-motion dataset based on HumanML3D. Extensive experiments demonstrate the excel" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 454, + 515, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 454, + 515, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 454, + 515, + 466 + ], + "type": "text", + "content": "lent robustness and generalization of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 489, + 394, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 489, + 394, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 489, + 394, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 510, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 616 + ], + "type": "text", + "content": "Recent years have witnessed remarkable progress in text-to-motion (T2M) generation, showing promising potential in applications such as gaming, metaverse, and virtual/augmented reality [10, 16, 32, 41, 50, 51]. This technology enables the intuitive creation of 3D character animations through natural language descriptions, significantly simplifying content creation workflows in animation [20], virtual reality [16], video games [30, 47], and robotics [5, 22]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Despite these advancements, a fundamental limitation restricts the practical application of current text-to-motion generation methods. As illustrated in Figure 1, existing approaches primarily focus on enhancing the motion quality and text-motion alignment, neglecting variations in skeletons and only generating human motions for standardized skeletons. However, real-world production environments routinely require varying skeletal configurations for diverse" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12926" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "characters. To this end, practitioners must resort to skeleton retargeting - a technically demanding process that necessitates significant expertise and frequently fails to produce usable results, inevitably leading to laborious tuning and production delays. Therefore, there is a pressing demand for straightforward motion generation given arbitrary skeletons." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 158, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 158, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 295, + 396 + ], + "type": "text", + "content": "To address this issue, we present OmniSkel, a novel framework that can directly generate high-quality human motion sequences for arbitrary character skeletons. As shown in Figure 1, our approach eliminates the complex retargeting stage required by traditional pipelines, offering a single-stage solution that directly produces character-specific motion without specialized expertise. Specifically, we propose a Skeleton-aware RVQ-VAE (SR-VAE) architecture, which utilizes Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into the motion encoding and reconstruction. Unlike previous approaches, our design effectively separates skeleton-invariant features, which is crucial for arbitrary skeleton adaptation. To ensure zero bone length error while preserving motion smoothness, we further propose a training-free approach, Motion Restoration Optimizer (MRO). With the above designs, our OmniSkel can deliver high-quality motion generation precisely in line with textual descriptions and character skeletons, providing robust support for practical text-to-motion applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 399, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 460 + ], + "type": "text", + "content": "To facilitate our research, we construct SkeleMotion-3D, a large-scale dataset based on HumanML3D, pairing textual descriptions with motion sequences across diverse skeletal configurations. This dataset establishes a strong foundation for motion generation with arbitrary skeletons." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 484 + ], + "type": "text", + "content": "Before delving into the details, we summarize our core contributions in this work as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "content": "- To the best of our knowledge, we present the first text-to-motion generation framework capable of directly producing motions for arbitrary character skeletons without retargeting, significantly simplifying creation workflows in real-world production environments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 546, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 630 + ], + "type": "text", + "content": "- We propose a novel Skeleton-aware RVQ-VAE (SR-VAE) architecture, which employs Kinematic Graph Cross Attention (K-GCA) to effectively integrate skeletal information into the motion encoding and reconstruction. Moreover, we further introduce Motion Reconstruction Optimizer (MRO) during inference, ensuring zero bone length error while preserving motion smoothness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "content": "- Based on HumanML3D, we construct SkeleMotion-3D, a large-scale text-skeleton-motion dataset. This establishes a foundation for arbitrary skeleton motion generation, facilitating future research in this field." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "- To verify the effectiveness of our method, we conduct extensive experiments on challenging datasets. Results demonstrate the superior robustness and generalization of" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 555, + 96 + ], + "type": "text", + "content": "our method. Further analysis reveals the contribution of each component to the performance improvement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 108, + 400, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 400, + 121 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 400, + 121 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 129, + 463, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 129, + 463, + 140 + ], + "spans": [ + { + "bbox": [ + 314, + 129, + 463, + 140 + ], + "type": "text", + "content": "2.1. Text-to-Motion Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 146, + 553, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 553, + 181 + ], + "type": "text", + "content": "Text-to-motion generation has evolved significantly as a vital research area in computer animation and human-computer interaction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 182, + 553, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 230 + ], + "type": "text", + "content": "Early Approaches. Initial methods established text-to-motion mappings through alignment techniques [3, 4, 9, 26, 36, 46], but produced motions that lacked naturalism and diversity due to their deterministic nature." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 231, + 553, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 553, + 278 + ], + "type": "text", + "content": "Probabilistic Frameworks. VAE-based approaches [6, 16, 28, 31, 32] model distributions of possible motions, while vision-language alignment methods [27, 33, 37, 40] leveraged pre-trained models for better semantic consistency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 279, + 553, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 553, + 386 + ], + "type": "text", + "content": "Recent Advances. Contemporary approaches include diffusion-based models [7, 10, 12, 21, 38, 41, 51, 52] generating high-fidelity motions through progressive denoising. Vector Quantized VAE (VQ-VAE) approaches treat motions as discrete token sequences, including autoregressive models [19, 45, 50, 54, 55], bidirectional methods [18, 35], and hybrid approaches [34]. Part-based generation methods [13, 54, 55] provide control over individual body components." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 387, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 553, + 434 + ], + "type": "text", + "content": "Limitations. Existing approaches remain constrained to standard skeletal structures, creating a gap between research and industry needs where adaptability to various character rigs is essential." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 553, + 471 + ], + "type": "text", + "content": "Our work addresses this limitation by developing a skeleton-agnostic framework that generates semantically consistent motions for arbitrary skeletal structures." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 479, + 429, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 479, + 429, + 493 + ], + "spans": [ + { + "bbox": [ + 314, + 479, + 429, + 493 + ], + "type": "text", + "content": "2.2. Motion Retargeting" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 617 + ], + "type": "text", + "content": "Motion retargeting adapts existing motion data between characters with different proportions and structures. Historically, Gleicher et al. [14] pioneered this field through spacetime optimization with kinematic constraints. Following this work, Lee and Shin [23] applied inverse kinematics with B-spline curve fitting, while Choi and Ko [11] simultaneously developed an online method preserving high-frequency details. Tak and Ko [39] later enhanced these approaches by utilizing dynamics constraints for physical plausibility." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "With the advent of deep learning, new approaches emerged, including Villegas et al. [44] with unsupervised retargeting via recurrent networks, Lim et al. [25] disentangling pose and movement, and Aberman et al. [1] designing skeleton-aware networks for cross-structural retargeting. More recently, research has shifted towards preserving motion semantics. For instance, Zhang et al. [49] developed a residual retargeting network with skeleton-aware" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12927" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 85, + 541, + 261 + ], + "blocks": [ + { + "bbox": [ + 241, + 71, + 376, + 82 + ], + "lines": [ + { + "bbox": [ + 241, + 71, + 376, + 82 + ], + "spans": [ + { + "bbox": [ + 241, + 71, + 376, + 82 + ], + "type": "text", + "content": "(a) Skeleton-aware Residual VQ-VAE" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 85, + 541, + 261 + ], + "lines": [ + { + "bbox": [ + 60, + 85, + 541, + 261 + ], + "spans": [ + { + "bbox": [ + 60, + 85, + 541, + 261 + ], + "type": "image", + "image_path": "bcf229b8eb283236c8777696ce593b163ed5de0c0c167ae0b375369dde5c54a6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 265, + 237, + 275 + ], + "lines": [ + { + "bbox": [ + 149, + 265, + 237, + 275 + ], + "spans": [ + { + "bbox": [ + 149, + 265, + 237, + 275 + ], + "type": "text", + "content": "(b) Masked Transformer" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 389, + 266, + 477, + 276 + ], + "lines": [ + { + "bbox": [ + 389, + 266, + 477, + 276 + ], + "spans": [ + { + "bbox": [ + 389, + 266, + 477, + 276 + ], + "type": "text", + "content": "(c) Residual Transformer" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 285, + 555, + 363 + ], + "lines": [ + { + "bbox": [ + 54, + 285, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 555, + 363 + ], + "type": "text", + "content": "Figure 2. Method overview. (a) Skeleton-aware Residual VQ-VAE: our framework processes source motion and skeleton through K-GCA (Kinematic Graph Cross Attention), which divides the input into Q, K, and V branches, creating a hierarchical representation with motion-invariant features. The residual quantization approach progressively captures motion details across multiple layers, enabling high-fidelity reconstruction for arbitrary target skeletons. (b) Masked Transformer: base-layer motion tokens " + }, + { + "bbox": [ + 54, + 285, + 555, + 363 + ], + "type": "inline_equation", + "content": "t^0" + }, + { + "bbox": [ + 54, + 285, + 555, + 363 + ], + "type": "text", + "content": " are randomly masked during training and the text-conditioned transformer learns to predict these tokens simultaneously, capturing fundamental motion patterns. (c) Residual Transformer: builds upon the base layer by progressively predicting higher-level residual tokens " + }, + { + "bbox": [ + 54, + 285, + 555, + 363 + ], + "type": "inline_equation", + "content": "t^v" + }, + { + "bbox": [ + 54, + 285, + 555, + 363 + ], + "type": "text", + "content": " conditioned on both text and previous layer tokens, enhancing motion fidelity and skeleton-specific details." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 384, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 295, + 491 + ], + "type": "text", + "content": "and shape-aware modules. Similarly, Lee et al. [24] proposed Skeleton-Agnostic Motion Embedding (SAME) using graph convolutional networks to separate skeleton information from motion while preserving semantics. Building upon these semantic approaches, Zhang et al. [48] introduced Semantics-aware Motion reTargeting (SMT), leveraging vision-language models to extract and maintain motion semantics through a two-stage pipeline with skeleton-aware pre-training and semantic constraints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 491, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 528 + ], + "type": "text", + "content": "However, despite this significant progress, many approaches still struggle with maintaining semantic consistency across dramatically different skeletal structures." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 538, + 122, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 122, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 122, + 552 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": "Our goal is to generate motion sequences " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{1:T}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " conditioned on both a text description " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " and a target skeletal structure " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_t\\in \\mathbb{R}^{J\\times D_{\\mathrm{m}}}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " denoting the number of joints and " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{m}}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " representing the dimension of joint features. The skeletal structure " + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S}\\in \\mathbb{R}^{J\\times D_{\\mathrm{s}}}" + }, + { + "bbox": [ + 55, + 557, + 296, + 713 + ], + "type": "text", + "content": " characterizes the static joint properties of the target character. As illustrated in Figure 2, our framework consists of two main components: a skeleton-aware VAE module that enables motion generation for arbitrary skeletal structures (Section 3.2), and a cooperative transformer architecture for motion token prediction (Section 3.3). We first introduce our newly proposed dataset and its representation (Section 3.1) to facilitate better understanding of our method. Then, we detail" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 384, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 407 + ], + "type": "text", + "content": "the complete training pipeline (Section 3.3) and inference process (Section 3.4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 414, + 451, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 451, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 451, + 425 + ], + "type": "text", + "content": "3.1. SkeleMotion-3D Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 431, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 515 + ], + "type": "text", + "content": "Our dataset comprises three key components: textual descriptions, motion sequences, and corresponding skeletal structures. Building upon the HumanML3D dataset [16], which combines motion data from AMASS [29] and HumanAct12 [15], we significantly expand the motion-skeleton pairs through a skeleton randomization process inspired by [24]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 516, + 554, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 554, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 554, + 551 + ], + "type": "text", + "content": "Feature Representation. We represent motion sequences as " + }, + { + "bbox": [ + 313, + 516, + 554, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}" + }, + { + "bbox": [ + 313, + 516, + 554, + 551 + ], + "type": "text", + "content": ", where the motion feature " + }, + { + "bbox": [ + 313, + 516, + 554, + 551 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{motion}}" + }, + { + "bbox": [ + 313, + 516, + 554, + 551 + ], + "type": "text", + "content": " with dimension " + }, + { + "bbox": [ + 313, + 516, + 554, + 551 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{m}}" + }, + { + "bbox": [ + 313, + 516, + 554, + 551 + ], + "type": "text", + "content": " consists of:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 357, + 558, + 553, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 558, + 553, + 571 + ], + "spans": [ + { + "bbox": [ + 357, + 558, + 553, + 571 + ], + "type": "interline_equation", + "content": "F _ {\\text {m o t i o n}} = \\left[ F _ {\\text {t o p o}}, F _ {\\text {r o o t}}, F _ {\\text {x y z}}, F _ {\\text {r o t}}, F _ {\\text {v e l}} \\right], \\tag {1}", + "image_path": "f666e5a65797a2fc077270af4af6e1ba843878b4ff9f8b2486996e2caa877f84.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 578, + 554, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 554, + 612 + ], + "type": "text", + "content": "comprising skeletal topology features, root joint features, relative spatial positions, 6D rotation features, and velocity features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 613, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 554, + 647 + ], + "type": "text", + "content": "The skeletal structures are encoded as " + }, + { + "bbox": [ + 313, + 613, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{J \\times D_{\\mathrm{s}}}" + }, + { + "bbox": [ + 313, + 613, + 554, + 647 + ], + "type": "text", + "content": ", where the skeletal feature " + }, + { + "bbox": [ + 313, + 613, + 554, + 647 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{skel}}" + }, + { + "bbox": [ + 313, + 613, + 554, + 647 + ], + "type": "text", + "content": " with dimension " + }, + { + "bbox": [ + 313, + 613, + 554, + 647 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 613, + 554, + 647 + ], + "type": "text", + "content": " consists of:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 390, + 649, + 553, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 649, + 553, + 662 + ], + "spans": [ + { + "bbox": [ + 390, + 649, + 553, + 662 + ], + "type": "interline_equation", + "content": "F _ {\\mathrm {s k e l}} = \\left[ F _ {\\mathrm {t o p o}}, F _ {\\mathrm {s t a t i c}} \\right], \\tag {2}", + "image_path": "b0f7566a3a2c94cd56a855418d1c05459b1317c11a203eb4b0a4e503218ec37b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 666, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 666, + 554, + 689 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{topo}}" + }, + { + "bbox": [ + 313, + 666, + 554, + 689 + ], + "type": "text", + "content": " represents topology features and " + }, + { + "bbox": [ + 313, + 666, + 554, + 689 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{static}}" + }, + { + "bbox": [ + 313, + 666, + 554, + 689 + ], + "type": "text", + "content": " denotes static skeletal properties." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Dataset Construction. Following the preprocessing pipeline established by HumanML3D, we standardize the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12928" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "content": "raw motion sequences through temporal normalization (20 FPS), sequence cropping (maximum 10 seconds), skeletal retargeting, and orientation alignment " + }, + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "inline_equation", + "content": "(\\mathrm{Z}+" + }, + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "content": " direction). The text annotations from HumanML3D, collected via Amazon Mechanical Turk with quality control measures, are preserved in our dataset." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 252 + ], + "type": "text", + "content": "Dataset Statistics. The SkeleMotion-3D dataset contains 14,616 distinct motion patterns, 29,233 different skeletal configurations, 43,848 unique text-motion-skeleton triplets, and 44,970 textual descriptions with 5,371 unique words. This represents a substantial increase compared to the training dataset (949 motion patterns and 4,745 motion-skeleton pairs) used in [24]'s motion retargeting network. The significant expansion enhances the robustness and generalization capability of our skeleton-aware motion generation model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 259, + 239, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 239, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 239, + 271 + ], + "type": "text", + "content": "3.2. Skeleton-aware Residual VQ-VAE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 276, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 295, + 407 + ], + "type": "text", + "content": "We propose Skeleton-aware Residual VQ-VAE (SR-VAE), a novel architecture that extends the conventional Residual VQ-VAE to enable skeleton-conditioned motion generation. Conventional motion VQ-VAEs [17, 19, 50, 53] and recent RVQ-VAEs [18, 45] focus on reconstructing source motions with identical skeletal structures. In contrast, as illustrated in Figure 2, SR-VAE adopts an asymmetric VAE architecture, consisting of an encoder, a decoder, a residual quantization module, and a Kinetic Graph Cross Attention (K-GCA) module, enabling motion reconstruction for arbitrary target skeletons." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "text", + "content": "Architecture Design. Our SR-VAE processes motion generation through three main stages. First, the input motion sequence " + }, + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{in}} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}" + }, + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "text", + "content": " and its corresponding skeleton structure " + }, + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{in}} \\in \\mathbb{R}^{J \\times D_{\\mathrm{s}}}" + }, + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "text", + "content": " are fused via K-GCA before being encoded into latent features. Following MoMask [18], these features undergo residual quantization to produce " + }, + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "inline_equation", + "content": "V + 1" + }, + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "text", + "content": " ordered code sequences. Finally, the decoder incorporates target skeleton features " + }, + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{tgt}} \\in \\mathbb{R}^{J \\times D_{\\mathrm{s}}}" + }, + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "text", + "content": " through multiple K-GCA layers at different decoding stages to generate the target motion " + }, + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{tgt}} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}" + }, + { + "bbox": [ + 55, + 408, + 295, + 528 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_{\\mathrm{tgt}} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}" + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "text", + "content": " denote the predicted motion sequence and " + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{tgt}}" + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "text", + "content": " denote the ground truth target motion. For residual quantization, let " + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{r}^v \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "text", + "content": " represent the residual features at layer " + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "inline_equation", + "content": "v \\in \\{0,1,\\dots,V\\}" + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{b}^v \\in \\mathbb{R}^{T \\times d}" + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "text", + "content": " denote the corresponding quantized codes, where " + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "text", + "content": " is the dimension of the latent representation. Additionally, let " + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{D}_{\\mathrm{xyz}} \\in \\mathbb{R}^{T \\times J \\times 3}" + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{D}}_{\\mathrm{xyz}} \\in \\mathbb{R}^{T \\times J \\times 3}" + }, + { + "bbox": [ + 55, + 529, + 295, + 636 + ], + "type": "text", + "content": " represent the ground truth and predicted joint positions in 3D space, respectively. The overall training objective consists of three terms:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 294, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 294, + 677 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\left\\| \\mathbf {M} _ {\\mathrm {t g t}} - \\hat {\\mathbf {M}} _ {\\mathrm {t g t}} \\right\\| _ {1} + \\beta \\sum_ {v = 0} ^ {V} \\left\\| \\mathbf {r} ^ {v} - \\operatorname {s g} \\left[ \\mathbf {b} ^ {v} \\right] \\right\\| _ {2} ^ {2} + \\lambda \\left\\| \\mathbf {D} _ {\\mathrm {x y z}} - \\hat {\\mathbf {D}} _ {\\mathrm {x y z}} \\right\\| _ {1} \\tag {3}", + "image_path": "3b6f238e1f91be03c7334ccb23a2037390128d591247314dafdecc2f92d9f638.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "where the first term is motion reconstruction loss, the second term is commitment loss for residual quantization, and the third term specifically focuses on joint positions accu" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 145 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 145 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 145 + ], + "type": "image", + "image_path": "548a422267deed8ed1c2be045bca439657900ab4efadc69eb8ac7262e605e982.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 153, + 544, + 164 + ], + "lines": [ + { + "bbox": [ + 323, + 153, + 544, + 164 + ], + "spans": [ + { + "bbox": [ + 323, + 153, + 544, + 164 + ], + "type": "text", + "content": "Figure 3. Architecture of our Kinetic Graph Cross Attention." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 186, + 553, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 257 + ], + "type": "text", + "content": "racy. Here, " + }, + { + "bbox": [ + 313, + 186, + 553, + 257 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}[\\cdot ]" + }, + { + "bbox": [ + 313, + 186, + 553, + 257 + ], + "type": "text", + "content": " denotes the stop-gradient operation, " + }, + { + "bbox": [ + 313, + 186, + 553, + 257 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 186, + 553, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 186, + 553, + 257 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 186, + 553, + 257 + ], + "type": "text", + "content": " are weighting factors for the commitment loss and joint position loss, respectively. The loss function is optimized via a straight-through gradient estimator [42]. Following [18, 35, 50], we employ exponential moving average for codebook updates and reset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 258, + 554, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 554, + 353 + ], + "type": "text", + "content": "Kinetic Graph Cross Attention. To effectively fuse motion and skeletal features, we propose Kinetic Graph Cross Attention (K-GCA), as illustrated in Figure 3, extending the traditional Graph Attention Networks (GAT) [8, 43] for motion-skeleton feature fusion. Unlike standard GAT that performs self-attention within a single feature type, K-GCA enables cross-modal attention between motion and skeletal features while preserving the skeletal graph structure." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "type": "text", + "content": "Given motion features " + }, + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{T \\times J \\times D_{\\mathrm{m}}}" + }, + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "type": "text", + "content": " and skeleton features " + }, + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^{J \\times D_{\\mathrm{s}}}" + }, + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "type": "text", + "content": ", K-GCA first projects them into a shared space using learnable projection matrices " + }, + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{m}} \\in \\mathbb{R}^{D_{\\mathrm{m}} \\times d_k}" + }, + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathrm{s}} \\in \\mathbb{R}^{D_{\\mathrm{s}} \\times d_k}" + }, + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "type": "inline_equation", + "content": "d_k" + }, + { + "bbox": [ + 313, + 353, + 553, + 413 + ], + "type": "text", + "content": " is the dimension of the projected space:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 366, + 422, + 553, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 422, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 366, + 422, + 553, + 436 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {W} _ {\\mathrm {m}} \\mathbf {X}, \\quad \\mathbf {K} = \\mathbf {V} = \\mathbf {W} _ {\\mathrm {s}} \\mathbf {S}. \\tag {4}", + "image_path": "aa7beaceabc73fac89671c50a1e556076b566989e4935810961a114e53e9315c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 445, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 553, + 469 + ], + "type": "text", + "content": "The attention scores between joints " + }, + { + "bbox": [ + 313, + 445, + 553, + 469 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 445, + 553, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 445, + 553, + 469 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 445, + 553, + 469 + ], + "type": "text", + "content": " are computed using learnable vectors " + }, + { + "bbox": [ + 313, + 445, + 553, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^{\\mathrm{src}}\\in \\mathbb{R}^{d_k}" + }, + { + "bbox": [ + 313, + 445, + 553, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 445, + 553, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{a}^{\\mathrm{dst}}\\in \\mathbb{R}^{d_k}" + }, + { + "bbox": [ + 313, + 445, + 553, + 469 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 348, + 478, + 553, + 492 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 478, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 348, + 478, + 553, + 492 + ], + "type": "interline_equation", + "content": "e _ {j k} = \\text {L e a k y R e L U} \\left(\\mathbf {a} ^ {\\mathrm {s r c}} \\cdot \\mathbf {Q} _ {j} + \\mathbf {a} ^ {\\mathrm {d s t}} \\cdot \\mathbf {K} _ {k}\\right), \\tag {5}", + "image_path": "624e6f744ddfbf3a68ed5084b0108d57dfae2856b581cefd9ad65c0d709b4105.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 502, + 554, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 571 + ], + "type": "text", + "content": "To preserve the skeletal structure, we mask the attention scores using the skeleton adjacency matrix " + }, + { + "bbox": [ + 313, + 502, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\{0,1\\}^{J \\times J}" + }, + { + "bbox": [ + 313, + 502, + 554, + 571 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 502, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{jk} = 1" + }, + { + "bbox": [ + 313, + 502, + 554, + 571 + ], + "type": "text", + "content": " if joints " + }, + { + "bbox": [ + 313, + 502, + 554, + 571 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 502, + 554, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 502, + 554, + 571 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 502, + 554, + 571 + ], + "type": "text", + "content": " are connected in the skeleton graph, and 0 otherwise. We set attention scores for disconnected joints to negative infinity before applying softmax:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 375, + 571, + 553, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 571, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 375, + 571, + 553, + 602 + ], + "type": "interline_equation", + "content": "\\tilde {e} _ {j k} = \\left\\{ \\begin{array}{l l} e _ {j k}, & \\text {i f} \\mathbf {A} _ {j k} = 1 \\\\ - \\infty , & \\text {i f} \\mathbf {A} _ {j k} = 0 \\end{array} \\right., \\tag {6}", + "image_path": "28c73c14743b80aaa5f1e5a09640dbd64fb6cdc56443b68e6a9b144c67e55ca8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 386, + 609, + 553, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 609, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 386, + 609, + 553, + 638 + ], + "type": "interline_equation", + "content": "\\alpha_ {j k} = \\frac {\\exp \\left(\\tilde {e} _ {j k}\\right)}{\\sum_ {l = 1} ^ {J} \\exp \\left(\\tilde {e} _ {j l}\\right)}, \\tag {7}", + "image_path": "f64fee98a112d018d10559e9569bcdcebbfd2e19d5d9a484834d496336bbf371.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 553, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 691 + ], + "type": "text", + "content": "The final output " + }, + { + "bbox": [ + 313, + 642, + 553, + 691 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}\\in \\mathbb{R}^{T\\times J\\times D_{\\mathrm{m}}}" + }, + { + "bbox": [ + 313, + 642, + 553, + 691 + ], + "type": "text", + "content": " is computed by applying the attention weights to value matrix " + }, + { + "bbox": [ + 313, + 642, + 553, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 642, + 553, + 691 + ], + "type": "text", + "content": " in each head, concatenating results from all heads, and adding back to the input through residual connection:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 347, + 700, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 347, + 700, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {X}} = \\operatorname {L a y e r N o r m} (\\mathbf {X} + \\operatorname {M u l t i H e a d} (\\alpha , \\mathbf {V})). \\tag {8}", + "image_path": "013bbcc2963b84748ecbfbd750357e6dcf34e9bc4ecc6357ff49766c96bb3c5b.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12929" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "where MultiHead " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "(\\alpha, \\mathbf{V})" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " represents the multi-head weighted combination of values using the computed attention weights " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ". This design allows K-GCA to capture dynamic relationships between motion and skeletal features while maintaining the anatomical constraints imposed by the skeletal structure. Compared to standard GAT, our approach provides two key advantages: (1) the ability to perform cross-modal attention between motion and skeletal features, and (2) the incorporation of skeletal structure constraints through attention masking." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 200, + 160, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 200, + 160, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 160, + 213 + ], + "type": "text", + "content": "3.3. Training Pipeline" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 218, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 296, + 253 + ], + "type": "text", + "content": "Our training pipeline follows a sequential approach for skeleton-agnostic motion generation from text, as shown in Figure 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": "SR-VAE Training. First, we train the Skeleton-aware Residual VQ-VAE (SR-VAE) to establish skeleton-agnostic motion representation. This model encodes motion sequences into hierarchical tokens and decodes them back to motion features, independent of skeletal structure. After training, both encoder " + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": " are frozen." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 327, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 386 + ], + "type": "text", + "content": "Tokenization of Training Data. Using the frozen SR-VAE encoder, we convert all training motion sequences into hierarchical token representations. For each motion " + }, + { + "bbox": [ + 55, + 327, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{in}}" + }, + { + "bbox": [ + 55, + 327, + 296, + 386 + ], + "type": "text", + "content": " with corresponding skeleton " + }, + { + "bbox": [ + 55, + 327, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{in}}" + }, + { + "bbox": [ + 55, + 327, + 296, + 386 + ], + "type": "text", + "content": ", we extract multi-level token sequences " + }, + { + "bbox": [ + 55, + 327, + 296, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^{0:V}" + }, + { + "bbox": [ + 55, + 327, + 296, + 386 + ], + "type": "text", + "content": " as targets for our text-to-token models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 387, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 470 + ], + "type": "text", + "content": "Masked Transformer Training. Inspired by MoMask [18], we implement a bidirectional Masked Transformer to generate base-layer tokens " + }, + { + "bbox": [ + 55, + 387, + 296, + 470 + ], + "type": "inline_equation", + "content": "(\\mathbf{t}^0)" + }, + { + "bbox": [ + 55, + 387, + 296, + 470 + ], + "type": "text", + "content": ". This design offers parallel generation capability and better global motion coherence compared to autoregressive approaches. During training, the model learns to predict original tokens at masked positions from text prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "text", + "content": "Residual Transformer Training. Finally, we train the Residual Transformer to predict higher-layer tokens " + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "inline_equation", + "content": "(\\mathbf{t}^{1:V})" + }, + { + "bbox": [ + 55, + 471, + 296, + 542 + ], + "type": "text", + "content": " encoding motion refinements. Following MoMask [18], this transformer takes both text descriptions and tokens from previous layers as input, focusing on finer motion details while maintaining consistency with the base motion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 551, + 160, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 160, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 160, + 562 + ], + "type": "text", + "content": "3.4. Inference Process" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "type": "text", + "content": "During inference, OmniSkel generates skeleton-aware motions through the following steps:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 593, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 628 + ], + "type": "text", + "content": "Base Token Generation: Given a text prompt, the Masked Transformer generates the base-layer tokens " + }, + { + "bbox": [ + 55, + 593, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}^0" + }, + { + "bbox": [ + 55, + 593, + 295, + 628 + ], + "type": "text", + "content": " that capture primary motion patterns." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": "Hierarchical Refinement: The Residual Transformer progressively generates higher-layer tokens " + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}^{1:V}" + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": ", starting with layer 1 and proceeding sequentially. Each layer's prediction utilizes both the text prompt and all previously generated token layers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Motion Decoding: Once we have the complete token hierarchy " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}^{0:V}" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": ", we feed these tokens along with the target skele" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "ton structure " + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_{\\mathrm{tgt}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": " into the frozen SR-VAE decoder to obtain motion features." + } + ] + } + ], + "index": 12 + }, + { + "type": "code", + "bbox": [ + 322, + 133, + 530, + 182 + ], + "blocks": [ + { + "bbox": [ + 321, + 121, + 501, + 132 + ], + "lines": [ + { + "bbox": [ + 321, + 121, + 501, + 132 + ], + "spans": [ + { + "bbox": [ + 321, + 121, + 501, + 132 + ], + "type": "text", + "content": "Algorithm 1: Motion Restoration Optimizer" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "lines": [ + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "spans": [ + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "type": "text", + "content": "Input: Measurement data: " + }, + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{topo}}" + }, + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{root}}" + }, + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{xyz}}" + }, + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "type": "text", + "content": "; Bone lengths: " + }, + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "type": "text", + "content": " \nOutput: Smooth motion sequence " + }, + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{smooth}}" + }, + { + "bbox": [ + 322, + 133, + 530, + 182 + ], + "type": "text", + "content": " with guaranteed skeletal consistency" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_body" + } + ], + "index": 14, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "type": "code", + "bbox": [ + 318, + 210, + 446, + 222 + ], + "blocks": [ + { + "bbox": [ + 318, + 186, + 487, + 209 + ], + "lines": [ + { + "bbox": [ + 318, + 186, + 487, + 209 + ], + "spans": [ + { + "bbox": [ + 318, + 186, + 487, + 209 + ], + "type": "text", + "content": "1 // Step 1: Restore initial motion using position-based approach;" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 318, + 210, + 446, + 222 + ], + "lines": [ + { + "bbox": [ + 318, + 210, + 446, + 222 + ], + "spans": [ + { + "bbox": [ + 318, + 210, + 446, + 222 + ], + "type": "text", + "content": "2 " + }, + { + "bbox": [ + 318, + 210, + 446, + 222 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{ric}} \\gets \\mathrm{ric}(F_{\\mathrm{topo}}, F_{\\mathrm{root}}, F_{\\mathrm{xyz}})" + }, + { + "bbox": [ + 318, + 210, + 446, + 222 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "code_body" + } + ], + "index": 16, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "type": "code", + "bbox": [ + 318, + 228, + 385, + 238 + ], + "blocks": [ + { + "bbox": [ + 318, + 228, + 385, + 238 + ], + "lines": [ + { + "bbox": [ + 318, + 228, + 385, + 238 + ], + "spans": [ + { + "bbox": [ + 318, + 228, + 385, + 238 + ], + "type": "text", + "content": "3 " + }, + { + "bbox": [ + 318, + 228, + 385, + 238 + ], + "type": "inline_equation", + "content": "M_{\\mathrm{smooth}} \\gets \\emptyset" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "code_body" + } + ], + "index": 17, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 317, + 255, + 539, + 497 + ], + "blocks": [ + { + "bbox": [ + 318, + 243, + 437, + 255 + ], + "lines": [ + { + "bbox": [ + 318, + 243, + 437, + 255 + ], + "spans": [ + { + "bbox": [ + 318, + 243, + 437, + 255 + ], + "type": "text", + "content": "4 foreach frame " + }, + { + "bbox": [ + 318, + 243, + 437, + 255 + ], + "type": "inline_equation", + "content": "f\\in M_{\\mathrm{ric}}" + }, + { + "bbox": [ + 318, + 243, + 437, + 255 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "lines": [ + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": "// Step 2: Extract unit direction vectors for each bone (except the root); \nInitialize set " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{U} \\gets \\emptyset" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": "; \nfor " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "i \\gets 1" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "J - 1" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": " do \nLet " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": " be the position of joint " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": "; \nLet " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\text{parent}(i)}" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": " be the position of joint " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": "'s parent; \nCompute displacement: " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i \\gets \\mathbf{p}_i - \\mathbf{p}_{\\text{parent}(i)}" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": "; \nCompute unit direction: " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i \\gets \\frac{\\mathbf{v}_i}{|\\mathbf{v}_i|}" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": "; \nAdd " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": "; \nend \n// Step 3: Reconstruct the current frame using extracted directions, bone lengths, and topology; \n" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "f_{\\text{rec}} \\gets \\text{ReconstructFrame}(\\mathcal{U}, S, F_{\\text{topo}}, F_{\\text{root}})" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": "; \nAdd " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "f_{\\text{rec}}" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "M_{\\text{smooth}}" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": "; \nend \nreturn " + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "inline_equation", + "content": "M_{\\text{smooth}}" + }, + { + "bbox": [ + 317, + 255, + 539, + 497 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "code_body" + } + ], + "index": 19, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 313, + 529, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 624 + ], + "type": "text", + "content": "Feature-to-Motion Conversion: Following the generation of motion features by our SR-VAE decoder, we face the challenge of transforming these abstract representations into production-ready motion data. Contemporary text-to-motion frameworks typically rely on either position-based restoration (offering fluidity but with bone length inconsistencies) or rotation-based restoration (maintaining skeletal integrity but introducing jitter and penetration artifacts)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "To overcome these limitations, we employ our Motion Restoration Optimizer (MRO, detailed in Algorithm 1), which synergistically combines the strengths of both approaches. MRO guarantees two critical properties simultaneously: absolute preservation of skeletal proportions (zero bone length error) and maintenance of smooth trajectory characteristics, producing physically plausible output " + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_{\\mathrm{tgt}}" + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12930" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 563, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 563, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 563, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 563, + 194 + ], + "type": "table", + "html": "
MethodFID↓R Precision↑MPJPE↓Skeleton Error↓Jitter Error↓Skating Ratio ↓
Top 1Top 2Top 3
SR-VAE (full model)0.0270.5080.7000.7930.03513.7781.8240.074
w/ Single-head K-GCA0.0400.4970.6850.7890.03412.2942.9350.061
w/o K-GCA0.0650.4950.6890.7850.04117.7269.5940.092
w/o RVQ0.1650.4670.6390.7750.06119.7263.5940.092
w/ Symmetric Reconstruction0.8850.4500.6410.7450.09789.7479.8670.067
w/ Rotation-based Restoration0.0270.5080.7000.7930.0630.0030.7240.193
w/ MRO0.0270.5080.7000.7930.0350.002.0840.090
", + "image_path": "4a7cc8c474e8b6a9e599bcabee8587733a8087ec8699fa1f08004712cb143857.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 202, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 555, + 237 + ], + "type": "text", + "content": "Table 1. Ablation study of SR-VAE components tested on the SkeleMotion-3D test set. Our SR-VAE is trained on the SkeleMotion-3D dataset. To evaluate SR-VAE's reconstruction performance for arbitrary skeletal motions, we input random skeleton motions to the encoder and use the standard skeleton from the HumanML3D dataset as the target skeleton." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 256, + 138, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 138, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 138, + 271 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 295, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 313 + ], + "type": "text", + "content": "We conduct extensive experiments to evaluate our approach across multiple dimensions, focusing on both semantic accuracy and physical plausibility." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 314, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 373 + ], + "type": "text", + "content": "Evaluation Metrics. We assess our model using established semantic metrics from prior work [16]: (1) Frechet Inception Distance (FID) for overall motion quality; (2) " + }, + { + "bbox": [ + 55, + 314, + 296, + 373 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 314, + 296, + 373 + ], + "type": "text", + "content": "-Precision and multimodal distance for text-motion alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 375, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 411 + ], + "type": "text", + "content": "To comprehensively evaluate physical plausibility—a critical aspect for animation applications—we introduce several specialized metrics:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 411, + 295, + 520 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 411, + 295, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 435 + ], + "type": "text", + "content": "- Skating Ratio: Measures foot sliding artifacts where feet unnaturally glide across the ground" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 435, + 294, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 294, + 460 + ], + "type": "text", + "content": "- Skeleton Error: Quantifies bone length variations throughout motion sequences" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 460, + 294, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 294, + 483 + ], + "type": "text", + "content": "- MPJPE (Mean Per-Joint Position Error): Assesses positional accuracy of joints" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 483, + 294, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 294, + 520 + ], + "type": "text", + "content": "- Jitter Error: Evaluates motion continuity and natural flow. Detailed formulations of these metrics appear in the supplementary materials." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": "Implementation Details. Our PyTorch implementation features a Skeleton-aware Residual VQ-VAE (SR-VAE) with residual blocks following T2M-GPT [50] and MoMask [18] architectures. We use a downscale factor of 4, with a residual quantization comprising 6 layers (each containing 512 codes of 512 dimensions) and quantization dropout ratio of 0.2. For the SR-VAE training objective, we set commitment loss weight " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " to 0.02 and joint position loss weight " + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 521, + 295, + 640 + ], + "type": "text", + "content": " to 2.0. The K-GCA module employs 2-head attention." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Both transformer models consist of 6 layers with 6 attention heads and 384-dimensional latent representations, trained on our SkeleMotion-3D dataset. We use a learning rate of 2e-4 with linear warm-up over 2000 iterations, batch sizes of 512 for SR-VAE and 64 for transformers. During inference, we apply classifier-free guidance (scale 4 for" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 257, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 293 + ], + "type": "text", + "content": "masked transformer, 5 for refinement transformer) with 10 iterative decoding steps. All experiments run on NVIDIA RTX 4090 GPUs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 300, + 534, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 534, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 534, + 313 + ], + "type": "text", + "content": "4.1. Skeleton-Aware Disentanglement Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 366 + ], + "type": "text", + "content": "A cornerstone of our approach is SR-VAE's ability to disentangle motion patterns from skeletal structures. We conduct a targeted experiment to validate this skeleton-agnostic capability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 366, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 426 + ], + "type": "text", + "content": "Figure 4 presents our analysis of four distinct motion patterns, each performed by three different skeletal structures with varying proportions. We process these motion-skeleton pairs through our SR-VAE encoder " + }, + { + "bbox": [ + 313, + 366, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 366, + 554, + 426 + ], + "type": "text", + "content": " and visualize the resulting base layer tokens " + }, + { + "bbox": [ + 313, + 366, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^0" + }, + { + "bbox": [ + 313, + 366, + 554, + 426 + ], + "type": "text", + "content": " on aligned line charts." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 426, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 509 + ], + "type": "text", + "content": "The results reveal that motions with identical semantic patterns but executed by anatomically diverse skeletons produce remarkably similar token sequences. This consistency demonstrates that our encoder " + }, + { + "bbox": [ + 313, + 426, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 426, + 554, + 509 + ], + "type": "text", + "content": " successfully extracts skeleton-invariant motion features, preserving essential semantic content regardless of the executing skeletal structure." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 510, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 594 + ], + "type": "text", + "content": "This confirms that SR-VAE effectively decomposes motion into two orthogonal components: (1) skeleton-invariant pattern information encoded in tokens, and (2) skeleton-specific structural information provided separately to the decoder. This disentanglement enables our model to generate physically plausible animations for arbitrary skeletal structures while preserving intended motion semantics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 600, + 408, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 408, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 408, + 613 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 677 + ], + "type": "text", + "content": "To validate SR-VAE's skeleton-aware capabilities, we conduct experiments using random input skeletons while decoding to standard HumanML3D [16] skeletons. This evaluates both semantic alignment and cross-skeleton reconstruction capabilities." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Table 1 presents our ablation results on the SkeleMotion-3D test set. The complete SR-VAE model achieves superior performance with an FID of 0.026 and strong R" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12931" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 78, + 553, + 189 + ], + "blocks": [ + { + "bbox": [ + 58, + 78, + 553, + 189 + ], + "lines": [ + { + "bbox": [ + 58, + 78, + 553, + 189 + ], + "spans": [ + { + "bbox": [ + 58, + 78, + 553, + 189 + ], + "type": "image", + "image_path": "b812883ec6b3350c3f600d7743fa6720c62b8839b3af77d3b9c2eddcbeb9abd9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 197, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 197, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 555, + 253 + ], + "type": "text", + "content": "Figure 4. Motion token analysis across skeletal structures. This figure demonstrates token sequences generated by the SR-VAE encoder for four distinct motion patterns, each performed by three skeletal structures with different proportions. The consistent token patterns across each motion type (columns) despite skeletal variations demonstrates our encoder's ability to extract skeleton-invariant motion features. Top row shows skeletal visualizations while bottom row displays the corresponding quantized token values, revealing clear pattern preservation across different anatomical configurations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 66, + 266, + 518, + 473 + ], + "blocks": [ + { + "bbox": [ + 66, + 266, + 518, + 473 + ], + "lines": [ + { + "bbox": [ + 66, + 266, + 518, + 473 + ], + "spans": [ + { + "bbox": [ + 66, + 266, + 518, + 473 + ], + "type": "image", + "image_path": "a793afe065b0353315d59d1958cec69c529528111794b872f152209c4e887f07.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 482, + 555, + 526 + ], + "lines": [ + { + "bbox": [ + 55, + 482, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 555, + 526 + ], + "type": "text", + "content": "Figure 5. Text-to-motion generation across different skeletal structures. Our OmniSkel framework generates diverse motions responding to different text prompts (columns) while adapting to different skeletal anatomies (rows). The human skeleton on the left is randomly selected from Mixamo [2], demonstrating OmniSkel's ability to maintain semantic consistency across dramatically different skeletal structures without retargeting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 547, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 295, + 571 + ], + "type": "text", + "content": "Precision scores (0.508/0.700/0.793 for Top-1/2/3), demonstrating excellent semantic consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 673 + ], + "type": "text", + "content": "Architecture Components. Using symmetric reconstruction (same skeleton for encoder and decoder) drastically increases FID to 0.885 and skeleton error to 89.747, confirming that asymmetric processing is essential for effective cross-skeleton learning. Removing K-GCA degrades FID to 0.065 with a significant increase in jitter error (9.594 vs. 1.824), demonstrating its importance in capturing cross-skeleton geometric correspondences." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Without RVQ, FID increases substantially to 0.165 and MPJPE to 0.061, highlighting RVQ's crucial role in maintaining semantic consistency across different skeletal struc" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "type": "text", + "content": "tures. The single-head K-GCA variant shows moderate performance decline (FID: 0.040), indicating that multi-head attention more effectively captures the rich geometric relationships between skeletons." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Motion Restoration. For motion restoration approaches, rotation-based methods achieve perfect skeleton reconstruction (skeleton error: 0.00) but significantly increase jitter error (30.724) and skating performance (0.193). In contrast, our MRO technique also achieves perfect skeleton reconstruction (skeleton error: 0.00) while maintaining low jitter error (2.084) and better skating ratio (0.090). This demonstrates MRO's effectiveness in preserving physical plausibility while maintaining motion quality across differ-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12932" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 147, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 147, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 147, + 83 + ], + "type": "text", + "content": "ent skeletal structures." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 296, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 296, + 182 + ], + "type": "text", + "content": "Our ablation study confirms each component's critical contribution to the overall system: K-GCA and RVQ enhance cross-skeleton reconstruction and semantic alignment, while MRO ensures physical consistency in the generated motions. These findings validate our architectural design choices for skeleton-aware motion generation and demonstrate SR-VAE's ability to effectively handle arbitrary skeleton motion retargeting with high fidelity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 190, + 255, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 255, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 255, + 204 + ], + "type": "text", + "content": "4.3. Qualitative and Quantitative Analysis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 209, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 281 + ], + "type": "text", + "content": "Quantitative Evaluation. We evaluate our methods against state-of-the-art text-to-motion approaches as shown in Table 2. The fundamental distinction of our approach is the ability to directly generate motions for arbitrary skeleton structures—a paradigm shift from existing methods that are restricted to fixed, pre-defined skeletons." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 293, + 294, + 423 + ], + "blocks": [ + { + "bbox": [ + 57, + 293, + 294, + 423 + ], + "lines": [ + { + "bbox": [ + 57, + 293, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 57, + 293, + 294, + 423 + ], + "type": "table", + "html": "
DatasetsMethodsFID↓R Precision↑
Top 1Top 2Top 3
Human-ML3DTM2T [17]1.501±.0170.424±.0030.618±.0030.729±.002
T2M [16]1.087±.0210.455±.0030.636±.0030.736±.002
MDM [41]0.544±.0440.320±.0050.498±.0040.611±.007
MLD [10]0.473±.0130.481±.0030.673±.0030.772±.002
MotionDiffuse [51]0.630±.0010.491±.0010.681±.0010.782±.001
T2M-GPT [50]0.141±.0050.492±.0030.679±.0020.775±.002
ReMoDiffuse [52]0.103±.0040.510±.0050.698±.0060.795±.004
AttT2M [54]0.112±.0060.499±.0030.690±.0020.786±.002
Momask [18]0.045±.0020.521±.0020.713±.0020.807±.002
Skele-Motion-3D OmniSkel (ours)0.027±.0000.508±.0020.700±.0020.793±.002
0.084±.0040.507±.0020.699±.0030.797±.002
", + "image_path": "25075906b47c32c24ee07742f8ee8efde32afd1c8e4dd9b30a65c9f3c1c2d7d4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 433, + 296, + 543 + ], + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 543 + ], + "type": "text", + "content": "Table 2. Comparison with text-to-motion methods on standard skeleton. Our approach introduces a novel pipeline trained on the SkeleMotion-3D dataset, which fundamentally differs from traditional T2M methods. For fair comparison with existing methods trained on HumanML3D, we configured our models (SR-VAE and OmniSkel) to target the standard HumanML3D skeleton during inference. Despite being designed for arbitrary skeleton generation, our methods achieve competitive performance on this fixed skeleton setting, demonstrating strong text-to-motion capabilities even when constrained to a single skeleton structure." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 653 + ], + "type": "text", + "content": "This fundamental difference presents a challenge for direct comparison, as our methods operate on the newly introduced SkeleMotion-3D dataset with diverse skeleton structures, while previous approaches are confined to the HumanML3D dataset with a single skeleton configuration. Notably, existing methods cannot be trained or evaluated on our dataset due to their inherent architectural limitations in handling skeleton variations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "To provide a meaningful comparison that highlights our text-to-motion capabilities, we configure our models at inference time to target the standard HumanML3D skeleton structure. Despite being designed for the more complex task of arbitrary skeleton generation, both SR-VAE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "and OmniSkel achieve competitive performance on the HumanML3D test set, with comparable or superior metrics in FID and R-Precision compared to specialized methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 109, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 553, + 156 + ], + "type": "text", + "content": "These results demonstrate that our methods successfully balance adaptability with generation quality, enabling cross-skeleton applications without compromising performance on standard benchmarks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 156, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 216 + ], + "type": "text", + "content": "Qualitative Results. Figure 5 illustrates OmniSkel's capability to generate motions across different skeletal structures. The visualization demonstrates: (1) diverse motions for different textual descriptions, and (2) consistent semantic motion across distinct skeletal anatomies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 217, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 264 + ], + "type": "text", + "content": "Each column represents a different textual prompt, from complex sequences to simpler actions. The model successfully captures the nuanced details of these descriptions, including locomotion styles and sequential actions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 265, + 553, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 553, + 313 + ], + "type": "text", + "content": "This comparison demonstrates OmniSkel's precise perception and effective integration of different semantics and arbitrary skeletons. More visualization results are provided in the supplementary materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 321, + 458, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 458, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 458, + 335 + ], + "type": "text", + "content": "4.4. Performance Comparison" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 317, + 349, + 552, + 384 + ], + "blocks": [ + { + "bbox": [ + 317, + 349, + 552, + 384 + ], + "lines": [ + { + "bbox": [ + 317, + 349, + 552, + 384 + ], + "spans": [ + { + "bbox": [ + 317, + 349, + 552, + 384 + ], + "type": "table", + "html": "
MethodParametersInference TimeGPU Memory
MoMask [18]44.85M0.075s1434M
Ours38.73M0.099s1402M
", + "image_path": "c332ea08bd070095fbe679d531b3ca84211318d9a04017df1acbddcb30173e87.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 333, + 392, + 534, + 403 + ], + "lines": [ + { + "bbox": [ + 333, + 392, + 534, + 403 + ], + "spans": [ + { + "bbox": [ + 333, + 392, + 534, + 403 + ], + "type": "text", + "content": "Table 3. Computational cost comparison on RTX 4090." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 418, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 553, + 502 + ], + "type": "text", + "content": "Our method achieves comparable performance to state-of-the-art approaches while maintaining slight GPU memory advantages as shown in Table 3 and uniquely enabling end-to-end motion generation for arbitrary skeletal specifications. This efficiency demonstrates the practical viability of our skeleton-agnostic framework for real-world applications requiring diverse character animations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 514, + 466, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 466, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 466, + 526 + ], + "type": "text", + "content": "5. Discussion and Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "In conclusion, we present OmniSkel, a text-driven motion generation framework that enables motion synthesis for arbitrary human skeletons through our novel K-GCA and SR-VAE architectures. Our Motion Restoration Optimizer (MRO) addresses the cross-skeleton motion reconstruction challenge, ensuring both motion smoothness and skeletal consistency. Leveraging our SkeleMotion-3D dataset, which pairs textual descriptions with diverse skeletal configurations, our approach bridges the gap between research and practical applications. Experiments demonstrate that OmniSkel achieves competitive performance while maintaining semantic alignment with textual descriptions. This work establishes a foundation for more inclusive motion generation systems that can serve diverse applications without requiring conventional retargeting approaches." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12933" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 140 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62320106007 and Grant 62032022; and in part by the Guangdong Provincial Key Laboratory under Grant 2023B1212060076." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 169, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 169, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 169, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 169, + 296, + 212 + ], + "type": "text", + "content": "[1] Kfir Aberman, Peizhuo Li, Dani Lischinski, Olga Sorkine-Hornung, Daniel Cohen-Or, and Baoquan Chen. Skeleton-aware networks for deep motion retargeting. ACM Transactions on Graphics (TOG), 39(4):62-1, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 214, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 214, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 61, + 214, + 295, + 235 + ], + "type": "text", + "content": "[2] Adobe. Mixamo. Online Animation Library, 2023. Available at: https://www MIXamo.com/.7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 236, + 295, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 236, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 61, + 236, + 295, + 289 + ], + "type": "text", + "content": "[3] Hyemin Ahn, Timothy Ha, Yunho Choi, Hwiyeon Yoo, and Songhwai Oh. Text2action: Generative adversarial synthesis from language to action. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5915-5920. IEEE, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 292, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 294, + 334 + ], + "type": "text", + "content": "[4] Chaitanya Ahuja and Louis-Philippe Morency. Language2pose: Natural language grounded pose forecasting. In 2019 International Conference on 3D Vision (3DV), pages 719–728. IEEE, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 336, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 295, + 390 + ], + "type": "text", + "content": "[5] André Antakli, Erik Hermann, Ingo Zinnikus, Han Du, and Klaus Fischer. Intelligent distributed human motion simulation in human-robot collaboration environments. In Proceedings of the 18th International Conference on Intelligent Virtual Agents, pages 319–324, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 434 + ], + "type": "text", + "content": "[6] Nikos Athanasiou, Mathis Petrovich, Michael J Black, and Gül Varol. Teach: Temporal action composition for 3d humans. In 2022 International Conference on 3D Vision (3DV), pages 414-423. IEEE, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "type": "text", + "content": "[7] Gustav Bredell, Kyriakos Flouris, Krishna Chaitanya, Ertunc Erdil, and Ender Konukoglu. Explicitly minimizing the blur error of variational autoencoders. arXiv preprint arXiv:2304.05939, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 481, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 294, + 513 + ], + "type": "text", + "content": "[8] Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks? In International Conference on Learning Representations, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 567 + ], + "type": "text", + "content": "[9] Haoye Cai, Chunyan Bai, Yu-Wing Tai, and Chi-Keung Tang. Deep video generation, prediction and completion of human action sequences. In Proceedings of the European Conference on Computer Vision (ECCV), pages 366–382, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 570, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 295, + 624 + ], + "type": "text", + "content": "[10] Xin Chen, Biao Jiang, Wen Liu, Zilong Huang, Bin Fu, Tao Chen, and Gang Yu. Executing your commands via motion diffusion in latent space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18000-18010, 2023. 1, 2, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 657 + ], + "type": "text", + "content": "[11] Kwang-Jin Choi and Hyeong-Seok Ko. Online motion retargeting. The Journal of Visualization and Computer Animation, 11(5):223-235, 2000. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[12] Xuehao Gao, Yang Yang, Zhenyu Xie, Shaoyi Du, Zhongqian Sun, and Yang Wu. Guess: Gradually enriching synthesis for text-driven human motion generation. IEEE Transactions on Visualization and Computer Graphics, 2024. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[13] Anindita Ghosh, Noshaba Cheema, Cennet Oguz, Christian Theobalt, and Philipp Slusallek. Synthesis of compositional animations from textual descriptions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1396-1406, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 162 + ], + "type": "text", + "content": "[14] Michael Gleicher. Retargetting motion to new characters. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, pages 33-42, 1998. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "text", + "content": "[15] Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia, pages 2021-2029, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "text", + "content": "[16] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5152-5161, 2022. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[17] Chuan Guo, Xinxin Zuo, Sen Wang, and Li Cheng. Tm2t: Stochastic and tokenized modeling for the reciprocal generation of 3d human motions and texts. In European Conference on Computer Vision, pages 580-597. Springer, 2022. 4, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 321, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 374 + ], + "type": "text", + "content": "[18] Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1900-1910, 2024. 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 376, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 408 + ], + "type": "text", + "content": "[19] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. arXiv preprint arXiv:2306.14795, 2023. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 411, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 475 + ], + "type": "text", + "content": "[20] Moritz Kappel, Vladislav Golyanik, Mohamed Elgharib, Jann-Ole Henningson, Hans-Peter Seidel, Susana Castillo, Christian Theobalt, and Marcus Magnor. High-fidelity neural human motion transfer from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1541-1550, 2021. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 478, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 520 + ], + "type": "text", + "content": "[21] Jihoon Kim, Jiseob Kim, and Sungjoon Choi. Flame: Freeform language-based motion synthesis & editing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8255-8263, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "text", + "content": "[22] Hema Koppula and Ashutosh Saxena. Learning spatiotemporal structure from rgb-d videos for human activity detection and anticipation. In International conference on machine learning, pages 792-800. PMLR, 2013. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "type": "text", + "content": "[23] Jehee Lee and Sung Yong Shin. A hierarchical approach to interactive motion editing for human-like figures. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pages 39-48, 1999. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "type": "text", + "content": "[24] Sunmin Lee, Taeho Kang, Jungnam Park, Jehee Lee, and Jungdam Won. Same: Skeleton-agnostic motion embedding for character animation. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[25] Jongin Lim, Hyung Jin Chang, and Jin Young Choi. PMnet: learning of disentangled pose and movement for unsupervised motion retargeting. In Proc. BMVC, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[26] Angela S Lin, Lemeng Wu, Rodolfo Corona, Kevin Tai, Qixing Huang, and Raymond J Mooney. Generating animated" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12934" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "videos of human activities from natural language descriptions. Learning, 2018(1), 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "text", + "content": "[27] Junfan Lin, Jianlong Chang, Lingbo Liu, Guanbin Li, Liang Lin, Qi Tian, and Chang-wen Chen. Being comes from not-being: Open-vocabulary text-to-motion generation with wordless training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23222-23231, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 295, + 215 + ], + "type": "text", + "content": "[28] Shunlin Lu, Ling-Hao Chen, Ailing Zeng, Jing Lin, Ruimao Zhang, Lei Zhang, and Heung-Yeung Shum. Humantomato: Text-aligned whole-body motion generation. In International Conference on Machine Learning, pages 32939-32977. PMLR, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "text", + "content": "[29] Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5442-5451, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 325 + ], + "type": "text", + "content": "[30] Dennis Majoe, Lars Widmer, and Juerg Gutknecht. Enhanced motion interaction for multimedia applications. In Proceedings of the 7th International Conference on Advances in Mobile Computing and Multimedia, pages 13-19, 2009. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 371 + ], + "type": "text", + "content": "[31] Mathis Petrovich, Michael J Black, and Gül Varol. Action-conditioned 3d human motion synthesis with transformer vae. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10985-10995, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "text", + "content": "[32] Mathis Petrovich, Michael J Black, and Gül Varol. Temos: Generating diverse human motions from textual descriptions. In European Conference on Computer Vision, pages 480-497. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "text", + "content": "[33] Mathis Petrovich, Michael J Black, and Gül Varol. Tmr: Text-to-motion retrieval using contrastive 3d human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9488-9497, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[34] Ekkasit Pinyoanuntapong, Muhammad Usama Saleem, Pu Wang, Minwoo Lee, Srijan Das, and Chen Chen. Bamm: Bidirectional autoregressive motion model. arXiv preprint arXiv:2403.19435, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 548 + ], + "type": "text", + "content": "[35] Ekkasit Pinyoanuntapong, Pu Wang, Minwoo Lee, and Chen Chen. Mmm: Generative masked motion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 601 + ], + "type": "text", + "content": "[36] Matthias Plappert, Christian Mandery, and Tamim Asfour. Learning a bidirectional mapping between human whole-body motion and natural language using deep recurrent neural networks. Robotics and Autonomous Systems, 109:13-26, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 668 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[38] Haowen Sun, Ruikun Zheng, Haibin Huang, Chongyang Ma, Hui Huang, and Ruizhen Hu. Lgtm: Local-to-global text-driven human motion diffusion model. In ACM SIGGRAPH 2024 Conference Papers, pages 1-9, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[39] Seyoon Tak and Hyeong-Seok Ko. A physically-based motion retargeting filter. ACM Trans. Graph., 24(1):98-117, 2005. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 150 + ], + "type": "text", + "content": "[40] Guy Tevet, Brian Gordon, Amir Hertz, Amit H Bermano, and Daniel Cohen-Or. Motionclip: Exposing human motion generation to clip space. In European Conference on Computer Vision, pages 358–374. Springer, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[41] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 1, 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 194, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 227 + ], + "type": "text", + "content": "[42] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in Neural Information Processing Systems, 30, 2017. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[43] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph Attention Networks. International Conference on Learning Representations, 2018. accepted as poster. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 326 + ], + "type": "text", + "content": "[44] Ruben Villegas, Jimei Yang, Duygu Ceylan, and Honglak Lee. Neural kinematic networks for unsupervised motion retargeting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8639-8648, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "text", + "content": "[45] Congyi Wang. T2m-hifigpt: generating high quality human motion from textual descriptions with residual discrete representations. arXiv preprint arXiv:2312.10628, 2023. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "text", + "content": "[46] Zhenyi Wang, Ping Yu, Yang Zhao, Ruiyi Zhang, Yufan Zhou, Junsong Yuan, and Changyou Chen. Learning diverse stochastic human-action generators by learning smooth latent transitions. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12281-12288, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 458 + ], + "type": "text", + "content": "[47] Mohammed Yeasin, Ediz Polat, and Rajeev Sharma. A multiobject tracking framework for interactive multimedia applications. IEEE transactions on multimedia, 6(3):398-405, 2004. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 503 + ], + "type": "text", + "content": "[48] Haodong Zhang, ZhiKe Chen, Haocheng Xu, Lei Hao, Xiaofei Wu, Songcen Xu, Zhensong Zhang, Yue Wang, and Rong Xiong. Semantics-aware motion retargeting with vision-language models. 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 569 + ], + "type": "text", + "content": "[49] Jiaxu Zhang, Junwu Weng, Di Kang, Fang Zhao, Shaoli Huang, Xuefei Zhe, Linchao Bao, Ying Shan, Jue Wang, and Zhigang Tu. Skinned motion retargeting with residual perception of motion semantics & geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13864-13872, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "text", + "content": "[50] Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. T2m-gpt: Generating human motion from textual descriptions with discrete representations. arXiv preprint arXiv:2301.06052, 2023. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[51] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022. 1, 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[52] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. arXiv preprint arXiv:2304.01116, 2023. 2, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12935" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 217 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[53] Yaqi Zhang, Di Huang, Bin Liu, Shixiang Tang, Yan Lu, Lu Chen, Lei Bai, Qi Chu, Nenghai Yu, and Wanli Ouyang. Motionpt: Finetuned llms are general-purpose motion generators. arXiv preprint arXiv:2306.10900, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[54] Chongyang Zhong, Lei Hu, Zihao Zhang, and Shihong Xia. Attt2m: Text-driven human motion generation with multiperspective attention mechanism. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 509-519, 2023. 2, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 217 + ], + "type": "text", + "content": "[55] Qiran Zou, Shangyuan Yuan, Shian Du, Yu Wang, Chang Liu, Yi Xu, Jie Chen, and Xiangyang Ji. Parco: Part-coordinating text-to-motion synthesis. arXiv preprint arXiv:2403.18512, 2024. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12936" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/bbd1b25f-9206-484c-b93d-8a139709e547_content_list.json b/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/bbd1b25f-9206-484c-b93d-8a139709e547_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f38ae2250406679745337d095e981cda59460274 --- /dev/null +++ b/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/bbd1b25f-9206-484c-b93d-8a139709e547_content_list.json @@ -0,0 +1,1721 @@ +[ + { + "type": "text", + "text": "Text2Outfit: Controllable Outfit Generation with Multimodal Language Models", + "text_level": 1, + "bbox": [ + 96, + 130, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuanhao Zhai $^{1*}$ Yen-Liang Lin $^{2}$ Minxu Peng $^{2}$ Larry S. Davis $^{2}$ Ashwin Chandramouli $^{2}$ Junsong Yuan $^{1}$ David Doermann $^{1}$ $^{1}$ State University of New York at Buffalo $^{2}$ Amazon", + "bbox": [ + 183, + 178, + 813, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 282, + 326, + 300 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing outfit recommendation frameworks focus on outfit compatibility prediction and complementary item retrieval. We present a text-driven outfit generation framework, Text2Outfit, which generates outfits controlled by text prompts. Our framework supports two forms of outfit recommendation: 1) Text-to-outfit generation, where the prompt includes the specification for each outfit item (e.g., product features), and the model retrieves items that match the prompt and are stylistically compatible. 2) Seed-to-outfit generation, where the prompt includes the specification for a seed item, and the model both predicts which product types the outfit should include (referred to as composition generation) and retrieves the remaining items to build an outfit. We develop a large language model (LLM) framework that learns the cross-modal mapping between text and image set, and predicts a set of embeddings and compositions to retrieve outfit items. We devise an attention masking mechanism in LLM to handle the alignment between text descriptions and image tokens from different categories. We conduct experiments on the Polyvore dataset and evaluate the quality of the generated outfits from two perspectives: 1) feature matching for outfit items, and 2) outfit visual compatibility. The results demonstrate that our approach significantly outperforms the baseline methods in text to outfit generation.", + "bbox": [ + 88, + 316, + 485, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 726, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fashion outfit recommendation system constructs outfits from items that go well together (e.g., an outfit that includes a top, a bottom, shoes, bags, etc). It is an important recommendation problem for online retailers, as customers frequently shop for items that can be worn together. Existing outfit recommendation methods mainly focus on compatibility prediction (determining whether a set of fashion items in an outfit go well together) [7, 9, 21, 29, 30] or outfit", + "bbox": [ + 89, + 752, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/551a976129217ba657b18b4a9617d29aa5fba0114455200d98cd4ea9c06172cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 522, + 286, + 900, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8a2a6a981607b614c210f361a4ae13fa33bdbd53f4c9747a507fcb40bdd143b.jpg", + "image_caption": [ + "Figure 1. A text to outfit generation framework that generates outfits driven by the text prompt. Our framework supports two forms of outfit recommendation. For text-to-outfit generation, the prompt includes outfit-level and item-level specifications. The model retrieves outfit items that match the texts in the prompt and are stylistically compatible. For seed-to-outfit generation, the prompt includes the specification for a seed item (a seed image or text description). The model predicts the composition (outfit categories) and retrieves the items to build outfits." + ], + "image_footnote": [], + "bbox": [ + 522, + 380, + 898, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "complementary item retrieval (completing a partial outfit by finding a compatible item) [20, 25]. The outfit items are predicted by a pre-trained model and are not controlled by the text prompt. For example, generating outfits with desired attributes, e.g., wide-leg sweatpants.", + "bbox": [ + 511, + 652, + 906, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent work ([14, 15]) for text to image retrieval leverages a pre-trained LLM model to learn image embeddings. They perform better retrieval than non-LLM based approaches (e.g., CLIP [24]) especially for longer text descriptions. However, these approaches focus on retrieving single images so they are not designed for retrieving image sets like outfits.", + "bbox": [ + 511, + 731, + 908, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present an outfit generation framework (Text2Outfit) that generates outfits driven by a text prompt (cf. Figure 1). Our framework learns the cross-modal mapping (text to image set) and predict the embeddings to retrieve outfits", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during an internship at Amazon.", + "bbox": [ + 107, + 887, + 395, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16165", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "that match the product features specified in the text prompt and are visually compatible. We instantiate our framework in two outfit recommendation scenarios using text prompts. In the first scenario (text-to-outfit generation), users specify the text prompt that describes both outfit-level features (e.g., an outfit for spring and causal) and individual items (e.g., a blouse in white with an off-shoulder design). Our model retrieves the complementary outfit items that match the prompt to build outfits. In the second scenario (seed-to-outfit generation), users specify the requirements for a few items (e.g., a seed item), using either images or texts, e.g., a white knitted sweater, and the model both completes the composition and retrieves complementary items to complete the outfit. We consider the case of just one seed item, but the framework can be extended to include multiple seed items. To build complete outfits, we predict the outfit composition (outfit product categories) for the remaining items and retrieve items in the predicted categories to build outfits.", + "bbox": [ + 89, + 90, + 485, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use learnable image tokens to generate embeddings for image retrieval similar to ([14, 15]). In contrast to prior work, we use a set of image tokens to retrieve outfits (image set) beyond a single image. Specifically, we incorporate a set of image tokens for different product categories (e.g., top, bottom, shoes). For each image token, we use the MLP layers on the top of the hidden state vector from the LLM to generate the embeddings. We train the framework in an end-to-end manner (fine-tune both LLM model and MLP layers) using LoRA [10] with a loss function that considers both product feature matching and outfit compatibility.", + "bbox": [ + 89, + 364, + 483, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The prompt includes the texts describing multiple items, which introduces the challenge of aligning the outfit texts and image tokens. It is important to avoid the case - where the image token from one category wrongly attends to the text descriptions from other categories. We present a masking mechanism in the self-attention layers to facilitate the needed attention learning (cf. Section 3.1 and Figure 3).", + "bbox": [ + 89, + 534, + 483, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For seed-to-outfit generation, since we only have one seed item as input, we must select the composition (outfit categories) for the remaining items. For example, for specific occasions/seasons, some categories should not be included. So, for summer outfits, the coat category should not be included in the final outfit. We train the LLM model to predict outfit compositions (outfit category combinations) and retrieve the items in the predicted categories.", + "bbox": [ + 89, + 642, + 483, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate our framework on the public Polyvore Outfit dataset [29] and measure the quality of generated outfits from two perspectives: 1) feature matching for outfit items and 2) outfit compatibility. Since the Polyvore dataset does not have text product attributes, we use Claude-3 [3] to extract the product features for each outfit item (we verify the precision of Claude-3 with human annotations in Section 4.1). The experiment section shows that our approach outperforms the baseline approaches in the product feature matching and", + "bbox": [ + 89, + 763, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "outfit compatibility.", + "bbox": [ + 513, + 92, + 648, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, the main technical contributions of our work are:", + "bbox": [ + 511, + 107, + 906, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a text-driven outfit generation framework that generates outfits based on text prompts or images, facilitating outfit recommendation in two forms.", + "- We develop a large language model framework to predict the embeddings (and compositions) to retrieve the outfit items where they match the input prompt and are visually compatible.", + "- We devise an attention masking mechanism to handle the alignment between the text descriptions in the prompt and the image tokens from different outfit categories." + ], + "bbox": [ + 511, + 138, + 906, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 305, + 653, + 321 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Outfit Recommendation", + "text_level": 1, + "bbox": [ + 513, + 330, + 735, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Outfit Compatibility Prediction: Outfit compatibility prediction evaluates how well a set of fashion items are compatible with each other in an outfit. Early works [21, 30] focused on pairwise item-level compatibility by leveraging item co-view or co-purchase information. However, compatibility is determined only by the pairwise items, not the entire outfits. Later works such as [7, 9, 27, 29] addressed the compatibility problem at the outfit level. They use different types of architectures to model the relationship between outfit items (e.g., bidirectional LSTM [9], subspace embeddings [27, 29] and graph convolution networks (GCN) [7]). These methods can be used for the fill-in-the-blank (FITB) problem - given a subset of items in an outfit and a set of candidate items, select the most compatible candidate. [4] extends FITB to build an entire outfit from a seed item - it starts from a seed item and iteratively finds the complementary items to build an outfit. However, the composition (outfit category combination) is pre-determined from the existing outfits. It is unclear how to compute the compositions for the new seed items.", + "bbox": [ + 511, + 354, + 908, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These approaches are mainly designed for compatibility prediction and FITB. In contrast, our framework generates outfits based on the input text prompt and predicts the outfit composition given a seed item.", + "bbox": [ + 511, + 657, + 906, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Outfit Complementary Item Retrieval: Complementary item retrieval is to find a compatible item(s) to complete a partial outfit from a large database. Lin et al. [20] designed a category-based attention mechanism that enables scalable indexing and searching for complementary items. Sarka et al. [25] generated the embeddings using a transformer-based architecture that encodes the compatibility for the partial outfit. These approaches address the FITB problem in a scalable manner (support large-scale indexing and complementary item retrieval), which differs from our problem settings.", + "bbox": [ + 511, + 719, + 908, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Outfit Retrieval: CLIP4Outfit [11] trained transformer encoders to learn cross-modal similarity (the similarity be", + "bbox": [ + 511, + 869, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16166", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tween text descriptions and outfit images) and used outfit-level embeddings to retrieve the most relevant outfit given the text descriptions. However, this approach considers the whole outfit as a candidate for retrieval. To be able to match a variety of text descriptions, the approach requires constructing a large-scale outfit dataset that covers different types of outfits, which is not feasible in practice. Moreover, their framework does not support outfit generation based on a given seed item. In contrast, our model is an outfit-building approach, retrieving individual items and assembling them into complete outfits.", + "bbox": [ + 89, + 90, + 485, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Text-to-Image Retrieval", + "text_level": 1, + "bbox": [ + 89, + 266, + 312, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most related work focuses on text-to-single image retrieval. CLIP [12] learned joint representations of text and images through large-scale contrastive learning. These representations can be fine-tuned on the domain-specific datasets to achieve higher performance, such as Fashion-CLIP [6] for the fashion domain. Later works (e.g., Flamingo [1], BLIP [16] and BLIP-2 [17]) learned a unified vision-language representation space and translated visual features into LLM to generate captions for images. The unified vision-language representation (e.g., Q-Former [17]) can be used for text-to-image retrieval. Some recent work leverages the LLM model to predict embeddings for retrieval [14, 15], which performs better than CLIP embeddings for longer captions. FROMAGe [15] takes the interleaved image and text as input and produce image embeddings and text as output. GILL [14] extends FROMAGe to support image retrieval and image synthesis. Similar to [14, 15], our framework also uses learnable image tokens for retrieval. In contrast to prior work on single image retrieval, our approach focuses on retrieving multiple items to build the outfits (image set retrieval).", + "bbox": [ + 91, + 287, + 485, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Text-to-Image/Outfit Synthesis", + "text_level": 1, + "bbox": [ + 89, + 613, + 362, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most work focuses on generating a single image given the text prompt [5, 8, 23, 26]. Some recent work [13, 18] generates a full-body human image wearing garments that reflect the provided textual descriptions. However, the synthesized images may not fully preserve the attribute details specified in the prompt as mentioned by [18]. In addition, to obtain shoppable content, these approaches require using a retrieval system to retrieve similar images from the product catalog (the generated images may not be able to find similar catalog images). In contrast, our framework is designed for outfit generation and retrieves outfit items that preserve the attributes specified in the prompt.", + "bbox": [ + 89, + 636, + 485, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 829, + 181, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 illustrates the system overview of our framework. Our framework supports both forms: 1) text-to-outfit generation and 2) seed-to-outfit generation. For text-to-outfit", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "generation, the task is to find the outfit items that match the text descriptions described in the input prompt. We use a set of image tokens and learn the image embeddings for different categories. We fine-tune the LLM to learn the embeddings using the outfit retrieval and compatibility losses, which encourage embeddings to retrieve the items that match the product features and are visually compatible. We devise a masking attention approach to handle the alignment between the text descriptions in the prompt and the image tokens. Details are described in Section 3.1.", + "bbox": [ + 511, + 90, + 906, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For seed-to-outfit generation, the task is to find outfit items to complete an outfit given a seed item and a prompt that describes the global outfit features. We train the LLM model to predict the outfit compositions (outfit category combination) and the embeddings to search the relevant outfit items to generate outfits. Details are described in Section 3.2", + "bbox": [ + 511, + 243, + 908, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Text-to-Outfit Generation", + "text_level": 1, + "bbox": [ + 511, + 359, + 748, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We are given an input text prompt: $T = \\{T_{1}, T_{2}, \\dots, T_{k}\\}$ , where each text description $T_{i}$ describes the outfit-level specification (e.g., outfit for spring and casual) or the item-level specification (e.g., a blouse in white with off-shoulder design). The order of the text descriptions for each category is arbitrary. The goal is to find the outfit items that match the text descriptions in the prompt (cf. Figure 2 (a)).", + "bbox": [ + 511, + 383, + 908, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To produce embeddings, we add a set of image tokens to LLM: $\\mathrm{[IMG_{cat_1},\\dots,IMG_{cat_n}]}$ ; each image token corresponds to a high-level category (we use 11 image tokens for Polyvore dataset; the category list is described in Section 4.3). The hidden state vectors of these image tokens from LLM are denoted as $\\mathrm{h_{cat_1}},\\ldots,\\mathrm{h_{cat_n}}$ . We learn a set of MLP layers: $\\mathrm{W}_{\\mathrm{ret}} = \\{\\mathrm{W}_{\\mathrm{cat}_i}^{\\mathrm{ret}},\\ldots,\\mathrm{W}_{\\mathrm{cat}_n}^{\\mathrm{ret}}\\}$ to transform the hidden state vectors to the embeddings: $\\mathbf{e} = \\{\\mathbf{e}_{\\mathrm{cat}_1},\\dots,\\mathbf{e}_{\\mathrm{cat}_n}\\}$ , where embedding $\\mathbf{e}_{\\mathrm{cat}_i}$ is used to retrieve the outfit images for the category $i$ .", + "bbox": [ + 511, + 489, + 908, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An outfit item is represented by an image I and its text descriptions T. We use a visual encoder $E_{\\mathrm{img}}$ to extract the visual embeddings from the image and use a text encoder $E_{\\mathrm{txt}}$ to extract text embeddings from the text descriptions. The multi-modal feature of an outfit item is denoted by: $\\mathrm{f} = \\mathrm{W}_{\\mathrm{item}} \\cdot (E_{\\mathrm{img}}(\\mathrm{I}) \\| E_{\\mathrm{txt}}(\\mathrm{T}))$ , where $\\|$ is the concatenation and $\\mathrm{W}_{\\mathrm{item}}$ is a learnable MLP model that is fine-tuned with the outfit retrieval loss.", + "bbox": [ + 511, + 642, + 908, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We train our framework end-to-end using the following losses: outfit retrieval loss, attention mask loss, and outfit compatibility loss. The outfit retrieval loss retrieves items that match the product features mentioned in the prompt. The attention mask loss handles the alignment between the outfit text descriptions and the image tokens. The outfit compatibility loss improves the visual compatibility of the retrieved outfit items.", + "bbox": [ + 511, + 763, + 908, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Outfit retrieval loss. A training batch includes a set of posi", + "bbox": [ + 511, + 885, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16167", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/74a93b5998007ea98ffb514748fe9107e8c3c7078b81c18045d95b64a778c15a.jpg", + "image_caption": [ + "(a) Text-to-outfit generation (Sec. 3.1)", + "(b) Seed-to-outfit generation (Sec 3.2)" + ], + "image_footnote": [], + "bbox": [ + 98, + 114, + 467, + 297 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f482bf3b625cb9b4a66ee812d99b8218998b9b35812c83a7a1860fe6d2e4cbef.jpg", + "image_caption": [ + "Figure 2. System overview of the text to outfit generation framework. For text-to-outfit generation (a), the LLM framework predicts a set of embeddings for different outfit categories, which are used to retrieve the outfit items. We devise an attention masking approach to handle the alignment between the text descriptions in the prompt and the image tokens from different categories (details are in Section 3.1). For seed-to-outfit generation (b), the LLM framework takes a seed item as input, predicts the outfit compositions (outfit categories), and retrieves items in the predicted categories to build the outfits (details are in Section 3.2)." + ], + "image_footnote": [], + "bbox": [ + 472, + 117, + 903, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tive outfits. We randomly sample a set of negative outfits for each positive outfit, where each item in the negative outfits has the same high-level category as the positive outfit. Since there are no ground truth negative outfits, random negatives are typically used in outfit training (e.g., [20, 25, 29]). We extract the occasion, season, color and product features of the outfit items using Claude-3 (details are in Section 4.1). To simulate the prompt for training, we randomly select the product features of each outfit item and use Claude-3 to generate the free-form prompt. The goal is to learn the embeddings so that the similarity between the predicted embeddings and the positive outfits is greater than the negative ones.", + "bbox": [ + 88, + 411, + 483, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, a batch includes a set of triples: $\\Upsilon = \\{T, p, N\\}$ , where $T$ is the prompt, $p$ is the positive outfit and $N$ is a set of negative outfits. The outfit retrieval loss is defined as:", + "bbox": [ + 89, + 606, + 483, + 666 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e t}} (p, N; \\mathrm {e}) = \\sum_ {i = 1} ^ {n} \\mathcal {L} _ {\\text {c a t} _ {i}} \\left(\\mathrm {f} _ {\\text {c a t} _ {i}} ^ {p}, \\mathrm {f} _ {\\text {c a t} _ {i}} ^ {N}; \\mathrm {e} _ {\\text {c a t} _ {i}}\\right) \\xi (\\text {c a t} _ {i}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 674, + 483, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We combine the loss for each category to obtain the outfit retrieval loss. $\\xi$ is the indicator function denoting if the category $i$ is presented in the outfit. The loss for each category is defined as:", + "bbox": [ + 89, + 722, + 483, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c a t} _ {i}} \\left(\\mathrm {f} _ {\\mathrm {c a t} _ {i}} ^ {p}, \\mathrm {f} _ {\\mathrm {c a t} _ {i}} ^ {N}; \\mathrm {e} _ {\\mathrm {c a t} _ {i}}\\right) = - \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathrm {f} _ {\\mathrm {c a t} _ {i}} ^ {p} , \\mathrm {e} _ {\\mathrm {c a t} _ {i}}\\right) / \\tau\\right)}{\\sum_ {n = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(\\mathrm {f} _ {\\mathrm {c a t} _ {i}} ^ {n} , \\mathrm {e} _ {\\mathrm {c a t} _ {i}}\\right) / \\tau\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 789, + 491, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We adopt the InfoNCE loss ([28]). $\\mathrm{f}_{\\mathrm{cat}_i}^p$ is the feature for positive outfit item and $\\mathrm{f}_{\\mathrm{cat}_i}^n$ is the feature for negative outfit item. $\\mathrm{e}_{\\mathrm{cat}_i}$ is the embeddings. $\\tau$ is the temperature. We use the cosine function for the sim function.", + "bbox": [ + 89, + 839, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attention mask loss. The prompt includes the text descriptions for multiple categories. We observed the alignment problem where the image token from a category may attend to the text descriptions from other categories. For example, the top image token attends to the shoe text descriptions. An example prompt: \"An outfit for spring and formal event. Black high heels. A black dress with a rose print. A red bracelet with gemstones.\" The text describes the features for fine-grained categories like high-heel, dress, etc, but the image tokens correspond to high-level categories (e.g., tops, bottoms, shoes). These pose the challenge for the LLM model to find the relations between the outfit text descriptions and multiple image tokens, which affects the product feature matching performance (cf. Table 3)).", + "bbox": [ + 511, + 411, + 906, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, we design a masking approach in the self-attention layers of the LLM decoder. Figure 3 illustrates the mask attention for the top image token. We incorporate the mask loss to guide attention learning, where the mask loss is applied to all image tokens. We compute the binary cross-entropy loss (BCE) for each image token between the ground truth mask and the predicted attention values. The mask loss is defined as:", + "bbox": [ + 511, + 623, + 908, + 743 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {k} \\mathrm {B C E} _ {i} \\left(\\alpha_ {j}, m _ {j} ^ {\\mathrm {G T}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 753, + 906, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $n$ is the number of image tokens. $k$ is the number of tokens from the text descriptions. $\\alpha_{j}$ is the predicted attention value and $m_j^{\\mathrm{GT}}$ is the ground truth mask.", + "bbox": [ + 511, + 809, + 906, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compatibility loss (CP. loss). We include a compatibility loss to facilitate the embedding learning that improves the compatibility of the generated outfit items. We trained an", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16168", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/548c25207238ed1786a2f06c74d360ef7aa061ce00aee8d4f68a34c2ff0cd21b.jpg", + "image_caption": [ + "Figure 3. The figure illustrates the masked attention for the top image token, where the mask facilitates the top image token to attend to the right text descriptions (outfit text and top text sections, highlighted in yellow) in the prompt. Details are in Section 3.1." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 482, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "outfit compatibility model (details are in Section 4.2), which predicts a compatibility score given a set of outfit items. In training, we freeze the model weights of the compatibility model (denoted as $\\mathcal{C}$ ) and fine-tune the embeddings (e) such that the compatibility score is maximized. The compatibility loss is defined as:", + "bbox": [ + 89, + 402, + 483, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c p}} (e) = - \\mathcal {C} \\left(\\mathrm {W} _ {\\mathrm {c p}} \\cdot \\mathrm {e}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 503, + 483, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training and Inference. We train the framework end-to-end using LoRA [2] by minimizing the following loss function:", + "bbox": [ + 89, + 532, + 482, + 563 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta , \\mathrm {W} _ {\\text {r e t}}, \\mathrm {W} _ {\\text {i t e m}}, \\mathrm {W} _ {\\mathrm {c p}}} \\mathcal {L} \\left(\\theta , \\mathrm {W} _ {\\text {r e t}}, \\mathrm {W} _ {\\text {i t e m}}, \\mathrm {W} _ {\\mathrm {c p}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 571, + 482, + 601 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {1} \\cdot \\mathcal {L} _ {\\mathrm {r e t}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\mathrm {m a s k}} + \\lambda_ {3} \\cdot \\mathcal {L} _ {\\mathrm {c p}},\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 601, + 424, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta$ is LLM model weights. $W_{\\mathrm{ret}}$ , $W_{\\mathrm{item}}$ , and $W_{\\mathrm{cp}}$ are the MLP models for outfit retrieval, outfit item feature extraction, and compatibility, respectively. $\\mathcal{L}_{\\mathrm{ret}}$ is the outfit retrieval loss, $\\mathcal{L}_{\\mathrm{mask}}$ is the mask loss and $\\mathcal{L}_{\\mathrm{cp}}$ is the compatibility loss. $\\lambda_1, \\lambda_2$ , and $\\lambda_3$ are empirically chosen to make the magnitudes of the losses similar in scale. We report the results of using different losses in Table 3.", + "bbox": [ + 89, + 627, + 483, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the inference, we feed the text prompt to the LLM model to predict the embeddings and use them to retrieve the outfit items.", + "bbox": [ + 89, + 733, + 483, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Seed-to-Outfit Generation", + "text_level": 1, + "bbox": [ + 89, + 787, + 326, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The task is to generate a complete outfit given an input seed item and a text prompt describing the global outfit features. Unlike the text-to-outfit generation task, where users specify the texts for each outfit item, here, users only specify the information for the seed item. Therefore, we need to predict the composition of the remaining items. One can retrieve", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the outfit items from all the categories to build an outfit. However, some categories do not go well with certain types of occasions, seasons, or the seed types. For example, coats should not be included for summer outfits, or an all-body dress should not be selected if the seed is a top.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Composition loss. To address this problem, we train the LLM model to predict the outfit compositions (a set of outfit categories), where the compositions are learned from outfit datasets (cf. Figure 2 (b)). Specifically, the input includes a prompt and a seed item (users provide an image or a text description for the seed item. For the image, we use a visual encoder to extract the embeddings and feed them to the LLM). To predict the compositions, we add two MLP models (one for predicting the high-level category and another for predicting the fine-grained category) on the top of the hidden state vectors of the high-level category to predict the composition. The composition loss is defined as:", + "bbox": [ + 511, + 167, + 906, + 348 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o m p}} = \\sum_ {i = 1} ^ {n} \\mathrm {C E} _ {i} \\left(y _ {\\mathrm {p r e d}} ^ {\\mathrm {h c}}, y _ {\\mathrm {g t}} ^ {\\mathrm {h c}}\\right) + \\mathrm {C E} _ {i} \\left(y _ {\\mathrm {p r e d}} ^ {\\mathrm {f g}}, y _ {\\mathrm {g t}} ^ {\\mathrm {f g}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 372, + 906, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the hc denotes the high-level category and fg denotes the fine-grained category. CE is the cross-entropy loss. We train the framework with LoRA end-to-end with the following loss:", + "bbox": [ + 511, + 421, + 906, + 482 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {4} \\cdot \\mathcal {L} _ {\\text {r e t}} + \\lambda_ {5} \\cdot \\mathcal {L} _ {\\mathrm {c p}} + \\lambda_ {6} \\cdot \\mathcal {L} _ {\\text {c o m p}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 494, + 906, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss function includes the retrieval, compatibility, and composition prediction losses. We train the model to predict the composition of the seed item belonging to the top category. The performance of other categories for the seed item is in Appendix Table 2. In inference, we search for outfit items in the predicted high-level and fine-grained categories to compose an outfit.", + "bbox": [ + 511, + 522, + 905, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 642, + 645, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 511, + 667, + 609, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the model performance on the Polyvore dataset [29]. The dataset includes two different sets: a disjoint set and a non-disjoint set. The non-disjoint set contains 53,306 training outfits and 10,000 testing outfits, while the disjoint set contains 16,995 training outfits and 15,145 testing outfits. We train our framework on the non-disjoint training set as it contains a larger number of outfits and report the performance on both non-disjoint and dis-joint testing sets (results are in the Appendix Section 1).", + "bbox": [ + 511, + 688, + 906, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Polyvore dataset does not have annotated product attributes. Motivated by recent works that leverage LLM for annotations (e.g., [19, 31]), we leverage Claude-3 [3] to extract outfit-level attributes (occasion and season) and item-level attributes (color and product features). We verify", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16169", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the agreement between the predicted attributes from Claude-3 and humans on a subset of testing samples (500 outfits and 2629 items). The results show the Claude-3 annotated attributes achieve $94.7\\%$ overall precision $(97.3\\%)$ for color, $95.5\\%$ for product feature, $92.2\\%$ for occasion, and $93.6\\%$ for season), verifying the feasibility of using Claude-3 for attribute annotations.", + "bbox": [ + 89, + 90, + 485, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 208, + 210, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the quality of the retrieved outfits in two ways: 1) feature matching for outfit items and 2) outfit compatibility. To evaluate feature matching, we compute the number of items in the top-10 retrieval results that match the product features specified in the prompt. We evaluate the feature matching for color, occasion, season, and product feature attributes for the outfit items, which are critical attributes for outfits. To handle non-exact matches, where the product features are different but their meanings are similar, we use ModernBERT embedding [22] to evaluate the affinity of two features (we select the thresholds of different attributes using a validation set). We report the results using precision@top-10 (denoted as P@10).", + "bbox": [ + 89, + 229, + 485, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate outfit compatibility, we train an outfit compatibility model using a SOTA architecture (OutfitTransformer [25]), where the model takes a set of outfit items as input, and predicts a compatibility score. We replace the ResNet features in the original architecture with the FashionCLIP embeddings [6] and use the item image and the text description as input. We achieve better outfit compatibility prediction than the original architecture. We achieve $94.93\\%$ (93% for OutfitTransformer) and $93.88\\%$ AUC (88% for OutfitTransformer) for the non-disjoint and disjoint sets respectively. We report the compatibility (CP) score for the top 10 retrieved outfits.", + "bbox": [ + 89, + 428, + 483, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There is no ground-truth outfit composition given a seed and prompt. To obtain a quantitative quality score, we use Claude-3 (we also evaluate the performance with Pixtral Large from Mistral AI, see Appendix Section 2) to evaluate the quality of the predicted compositions. Recent work (e.g., [33]) shows that strong LLMs achieve high agreement with humans. We evaluate the quality in two ways: 1) We use Claude-3 to predict a quality score for each predicted composition and report an average quality score (denoted as Comp. score). 2) Given two compositions generated by the baseline and our approach, we ask Claude-3 to pick the better composition and report the winning rate of our approach (denoted as Comp. win rate).", + "bbox": [ + 89, + 609, + 485, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Baselines and Model Settings", + "text_level": 1, + "bbox": [ + 89, + 816, + 351, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There is no existing framework for text-to-outfit retrieval tasks ([11] is designed to retrieve entire outfits, which is different from our settings as mentioned in Section 2). We convert a set of baseline approaches to the outfit retrieval", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tasks and report the performance. In Table 1, we categorize these baselines into the text-to-single-item retrieval and text-to-outfit retrieval approaches. Existing methods for text-to-image retrieval [6, 14, 24] mainly focus on single-image retrieval. To make these methods applicable to outfit retrieval problem, we make the following changes: 1) We convert the outfit-level specification (occasion/season) into item-level specification. 2) We use the text descriptions of each category to retrieve the outfit items (assuming that the text descriptions to category correspondence are given).", + "bbox": [ + 511, + 90, + 906, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For text-to-outfit retrieval, we converted the GILL [14] framework to support multiple-category retrieval. For text-to-outfit retrieval methods, the text descriptions to category correspondence information are not given, which is directly predicted by the models (except the fixed mask approach in Table 1, which uses the text-to-category correspondence to generate the fixed mask for attention). The details of the baseline approaches are described below.", + "bbox": [ + 511, + 243, + 908, + 363 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- CLIP [24]: We convert the outfit-level specification into item-level specifications and perform retrieval for each category. We use ViT-L-14 version.", + "- Fashion-CLIP 2.0 [6]: We use the same settings as CLIP and use Fashion-CLIP to retrieve items for each category. We use version 2.0, which achieves better performance than version 1.0.", + "- BLIP-2 [17]: We use the pre-trained image encoder and Q-Former from BLIP-2. Like CLIP, we perform image retrieval for each category using the corresponding text description. We use the ViT-g version.", + "- GILL [14] + Multiple categories: We add the learnable image tokens for each high-level category for multiple image retrieval. We fine-tune the MLP layers to learn embeddings while keeping the LLM model fixed following the original training approach. We use the same prompt format as our model. Note that we compare our method with an improved version of GiLL, in which we replace the original CLIP vision encoder with Fashion-CLIP" + ], + "bbox": [ + 511, + 364, + 906, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Settings: For our framework, we use the pretrained OPT 6.7B model [32] (same as GILL) as the LLM backbone. We fine-tune the LLM with LoRA (we set the rank to 64 for Q, K, and V projection matrices). We use Fashion-CLIP 2.0 [6] as the visual encoder $E_{\\mathrm{img}}$ and the text encoder $E_{\\mathrm{text}}$ . Fashion-CLIP 2.0 achieves better performance than CLIP embeddings. In model training, the visual and text encoder weights are fixed. We use the bi-directional masks for image tokens. We use 11 image tokens for the Polyvore dataset, corresponding to the following high-level categories: tops, bottoms, all-body, outerwear, shoes, bags, jewelry, accessories, sunglasses, hats, and scarves.", + "bbox": [ + 511, + 652, + 908, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Text-to-Outfit Performance Evaluation", + "text_level": 1, + "bbox": [ + 511, + 847, + 846, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 summarizes the text-to-outfit retrieval results. We use the testing set of the Polyvore dataset to generate the", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16170", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a4ea5b61ccf13e886c2ca23428576f7fd81720cf696b79894a712663172add94.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodModalityColor (P@10)Product feature (P@10)Season (P@10)Occasion (P@10)Avg (P@10)CP. score
Text-to-single-item retrieval
CLIP [24]Text only96.0670.9190.2276.6283.4663.49
Fashion-CLIP [6]Text only96.8280.4289.6881.2787.0564.07
BLIP-2 [17]Text only93.2966.9485.4684.5182.5565.09
CLIP [24]Text to image83.5961.3285.4967.0174.3549.77
Fashion-CLIP [6]Text to image87.5972.6785.9168.8878.7653.03
BLIP-2 [17]Text to image86.3966.6985.6267.7576.6151.05
Text-to-outfit retrieval
GILL [14] + multiple categoryText to image48.1948.1586.8769.6963.2374.26
Text2Outfit + learnable mask (ours)Text to image78.4276.5486.6070.4578.0066.28
Text2Outfit + learnable mask (ours)Text to image/text86.0584.8195.6993.6690.0584.05
Text2Outfit + fixed mask (ours)Text to image/text96.6186.2794.8792.4892.5682.51
", + "bbox": [ + 91, + 87, + 906, + 238 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bd5f349ad2db606f4ac8c9e01ec3de7460a6f6e3a05d59df12187f6f8c2c8192.jpg", + "table_caption": [ + "Table 1. Text-to-outfit retrieval performance comparison. P@10 denotes precision at top-10 and CP. score denotes compatibility score using [25] for top-10 retrieved outfits. Note that for single-item retrieval approaches, the text descriptions and its corresponding category are given while the outfit retrieval approaches are not given such information. The bold text denotes the best performance and the underlines denote the second best performance. Details are in Section 4.2." + ], + "table_footnote": [], + "table_body": "
MethodCompositionSeason (P@10)Occasion (P@10)Comp. score (c3)Comp win rate (c3)CP. scoreCP. score (c3)
Seed text + prompt (baseline)All categories97.3296.202.596%80.822.78
Seed text + prompt (ours)Predicted97.5696.014.5794%86.183.85
Seed image + prompt (baseline)All categories95.0087.502.858%72.882.66
Seed image + prompt (ours)Predicted96.4391.804.3292%83.363.83
", + "bbox": [ + 91, + 316, + 903, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Seed-to-outfit retrieval performance comparison. The scores with (c3) are evaluated by Claude-3. Details are in Section 4.5.", + "bbox": [ + 104, + 388, + 888, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "prompt. Given a test outfit including multiple outfit items, we randomly select outfit attributes (occasion and season) and item attributes (color and product features) and use Claude-3 to generate the free-form prompt. We use the generated prompt to evaluate the outfit retrieval performance.", + "bbox": [ + 89, + 430, + 482, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our approaches achieve $90.05\\%$ and $92.56\\%$ average $\\mathrm{P@10}$ using the learnable mask and fixed mask respectively, showing better feature-matching precision than the baseline approaches. The GILL + multi-category approach has an alignment issue (the alignment between the outfit texts and the images tokens) while our approach uses the learnable mask to alleviate the alignment issue. We fine-tune the model end-to-end with LoRA instead of using the frozen LLM model as GILL, improving the performance (see Table 3 for detailed analysis). Our approach with a learnable mask also performs better than the single-item retrieval approaches, where they are given the text description to category correspondence information, while our approach does not.", + "bbox": [ + 88, + 506, + 482, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, our approach shows better outfit compatibility than the baseline approaches. We achieve the compatibility score 84.05 using the learnable mask setting (we compute the average compatibility scores for the ground truth positive (86.95) and negative (22.99) outfits on the Polyvore testing set using OutfitTransformer. The generated outfits from our method achieve similar compatibility as the positive outfits). Single-item retrieval approaches perform worse on outfit compatibility because they find items that match the text descriptions but do not model the item relationships in an outfit. Our framework is fine-tuned on the outfit data, and incorporates the outfit retrieval and compatibility losses,", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which improve overall compatibility. Figure 4 shows example results of our framework.", + "bbox": [ + 511, + 430, + 906, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Seed-to-Outfit Performance Evaluation", + "text_level": 1, + "bbox": [ + 511, + 468, + 849, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 summarizes the results for seed-to-outfit retrieval. We used the Polyvore test set to construct the input (seed and prompt) for evaluation. Given a test outfit including multiple outfit items, we randomly select the outfit attributes (occasion and season) and use Claude-3 to generate the free-form prompt.", + "bbox": [ + 511, + 489, + 908, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our approach with a baseline approach that uses all 11 high-level categories for the composition. We feed the compositions (e.g., category 1/fine-grained category 1, category 2/fine-grained category 2, ...) of the baseline and our approach to Claude-3 for evaluation (for the baseline approach, the fine-grained category is obtained from the top-1 retrieved outfit). Our approach shows better composition scores (the score is between 1 (poor) to 5 (excellent)) and winning rates than the baseline (e.g., $94\\%$ and $92\\%$ winning rate of using seed text and seed image, respectively). We verify the agreement between Claude-3 and humans on a subset of samples (500 outfitsits). The Spearman's rank correlation coefficient is 0.73, which indicates they are highly correlated. We report the visual compatibility using both OutfitTransformer [25] (CP. score) and Claude-3 (CP. score (c3)). Our approach shows better visual compatibility than the baseline approach.", + "bbox": [ + 511, + 582, + 908, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 848, + 666, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multimodal feature: We compare the performance of single-modality (image) vs. multimodality (image + text)", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16171", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/825440f05325dd9143ac5306f2fee0c8c3a808764019b2227d9e82421a698dcf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModalityLoRAMask LossCP. lossColor (P@10)Product feature (P@10)Season (P@10)Occasion (P@10)Avg (P@10)CP. score
1Image---48.3153.0385.1068.6263.7653.43
2Image + text---44.6047.1888.9481.0765.4578.62
3Image + text✓--74.7475.0195.6093.9084.8286.95
4Image + text✓Global CE-79.2882.2396.0294.4387.9985.63
5Image + text✓Local BCE-85.5884.8795.8993.4789.9583.43
6Image + text✓Local BCE✓86.0584.8195.6993.6690.0584.05
", + "bbox": [ + 91, + 87, + 906, + 174 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study of different components in our framework. Details are described in Section 4.6.", + "bbox": [ + 189, + 184, + 803, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(a) Text-to-outfit generation", + "text_level": 1, + "bbox": [ + 107, + 218, + 313, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt: Spring outfit for casual. A canvas upper flats in a pink shade. A jeans in black with high-waisted accents. A sleeveless top in a white shade. A denim jacket in a blue shade.", + "bbox": [ + 106, + 244, + 387, + 289 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4efe064e518d69a18e7ff1e1e3a626f988c797752138e13c95524d7bfbe7e1e.jpg", + "image_caption": [ + "GILL + Multiple Category" + ], + "image_footnote": [], + "bbox": [ + 107, + 316, + 173, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9be4457d91776668bdd16f14a1962ac1ec65ed9777cf919efe78084094afba9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 316, + 248, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18efad0ff3ea480b301f1711bda6c09bac818333885e997ff6ca54342dda0da2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 316, + 323, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce5b83a67e98e063fbfe395c9395b168cdd7586bd7974ca14a54dfedb23e1344.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 316, + 400, + 367 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1317f57e31be65ee5a819c8d98d7c40e2bead98444790a87697636326383654.jpg", + "image_caption": [ + "Text2Outfit + learnable mask (ours)" + ], + "image_footnote": [], + "bbox": [ + 106, + 397, + 173, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/030bdd17657e64778cf578361ddbb6a322b541c65e28e1fce63eb8a89d9d08fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 181, + 397, + 248, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ba38cb2439c77f1e4cb73a6098cc169320807725a6cefd3cfc5a9a5720569a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 397, + 321, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71edc74bafd512ef522181b76693381aaf120b531651abe27973106c59e5f6a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 397, + 398, + 448 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b816c838c4695ed9cd965bfe7e7a0976de041081e3a02d1eb6872a094d8b300.jpg", + "image_caption": [ + "Text2Outfit + fixed mask (ours)" + ], + "image_footnote": [], + "bbox": [ + 106, + 483, + 173, + 532 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d168ea803e55d552c8bf367a9ea55d9770ad0f256f9cfe9b0f420f2d34ef8b4a.jpg", + "image_caption": [ + "Figure 4. Qualitative results. We compare the results for text-to-outfit generation with the extended GILL [14] approach that supports multiple image retrieval and our framework (learnable mask and fixed mask). The non-matched items are marked with boxes. We show the predicted compositions and retrieval results for seed-to-outfit generation with a seed image or text. More visual results are in the Appendix Section 4." + ], + "image_footnote": [], + "bbox": [ + 181, + 483, + 248, + 532 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f4c32543f29a9a9b14363f4dc103a17f4c8f0bc3f5e0ff9318190bd215d2144.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 484, + 321, + 532 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b6177f2df1e6f18c2d3d414e179c2abef279d14d3e9cd1bc187de643f8c8c17c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 333, + 484, + 398, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b) Seed-to-outfit generation", + "text_level": 1, + "bbox": [ + 419, + 218, + 632, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt: Outfit for casual in spring.", + "bbox": [ + 419, + 243, + 591, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/29714022b8032df6c9d2e00402574ad4af879ccb5a9fb37ebe02796f802a0f77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 256, + 895, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt: Formal events outfit for spring weather.", + "bbox": [ + 419, + 400, + 658, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78d6996f0a84d87d05b1fb3de30f9b519db6b6178dedb92d37c151e3ba37b3d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 412, + 895, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(row-1 and row-2 in Table 3). Using multimodality improves the average P@10 from $63.76\\%$ to $65.45\\%$ for feature matching.", + "bbox": [ + 88, + 648, + 485, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LoRA: We compare the performance of fine-tuned LLM (with LoRA) vs. frozen LLM (row-2 and row-3 in Table 3). Using LoRA significantly improves the average P@10 from $65.45\\%$ to $84.82\\%$ and improves the compatibility score from $78.62\\%$ to $86.95\\%$ .", + "bbox": [ + 89, + 699, + 483, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mask loss: We investigate different loss functions for the attention masking: Global CE and Local BCE (row-4 and row-5 in Table 3). For Global CE, we consider the attention values from all the text tokens to each image token as a distribution and optimize a global cross-entropy loss. We optimize each attention value for Local BCE with a binary cross-entropy loss. Local BCE performs better, as it optimizes the attention directly with the ground truth mask,", + "bbox": [ + 89, + 779, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "while global CE optimizes the overall distribution of the attention.", + "bbox": [ + 511, + 648, + 906, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compatibility loss: We analyze the impact of including a compatibility loss in the training loss (cf. Equation 5). It further improves compatibility (row-6 in Table 3).", + "bbox": [ + 511, + 680, + 906, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 736, + 640, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a framework for text-based outfit generation. We explore two different forms of outfit generation: 1) text-to-outfit generation and 2) seed-to-outfit generation tasks, which enable the controllable outfit generation. In addition, we introduce a masking mechanism which better handles the alignment between outfit texts and the image tokens, and achieves comparable performance when using the ground-truth masks. The experimental results demonstrate that our model outperforms the baseline approaches for text to outfit retrieval tasks.", + "bbox": [ + 511, + 761, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16172", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 118, + 187, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 3", + "[2] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, et al. Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2, pages 929-947, 2024. 5", + "[3] Anthropic. Introducing the next generation of claude. https://www.anthropic.com/news/claude-3-family, 2024.2,5", + "[4] Federico Becattini, Federico Maria Teotini, and Alberto Del Bimbo. Transformer-based graph neural networks for outfit generation, 2023. 2", + "[5] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3", + "[6] Patrick John Chia, Giuseppe Attanasio, Federico Bianchi, Silvia Terragni, Ana Rita Magalhães, Diogo Goncalves, Ciro Greco, and Jacopo Tagliabue. Contrastive language and vision learning of general fashion concepts. Scientific Reports, 12 (1):18958, 2022. 3, 6, 7", + "[7] Guillem Cucurull, Perouz Taslakian, and David Vazquez. Context-aware visual compatibility prediction. In CVPR, pages 12617-12626, 2019. 1, 2", + "[8] Runpei Dong, Chunrui Han, Yang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. DreamLLM: Synergistic multimodal comprehension and creation. In ICLR, 2024. 3", + "[9] Xintong Han, Zuxuan Wu, Yu-Gang Jiang, and Larry S Davis. Learning fashion compatibility with bidirectional lstms. In ACM MM, pages 1078-1086, 2017. 1, 2", + "[10] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 2", + "[11] Junkyu Jang, Eugene Hwang, and Sung-Hyuk Park. Lost your style? navigating with semantic-level approach for text-to-outfit retrieval. In WACV, pages 8066-8075, 2024. 2, 6", + "[12] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 3", + "[13] Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven" + ], + "bbox": [ + 93, + 143, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3", + "[14] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. NeurIPS, 36, 2023. 1, 2, 3, 6, 7, 8", + "[15] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In ICML, pages 17283-17300. PMLR, 2023. 1, 2, 3", + "[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3", + "[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 3, 6, 7", + "[18] Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, Kwan-Yee Lin, and Wayne Wu. Cosmicman: A text-to-image foundation model for humans. In CVPR, pages 6955-6965, 2024. 3", + "[19] Wei Li, Hehe Fan, Yongkang Wong, Yi Yang, and Mohan Kankanhalli. Improving context understanding in multimodal large language models via multimodal composition learning. In *Forty-first International Conference on Machine Learning*, 2024. 5", + "[20] Yen-Liang Lin, Son Tran, and Larry S Davis. Fashion outfit complementary item retrieval. In CVPR, pages 3311-3319, 2020. 1, 2, 4", + "[21] Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, pages 43-52, 2015. 1, 2", + "[22] Zach Nussbaum, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. Nomic embed: Training a reproducible long context text embedder, 2024. 6", + "[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 3", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 6, 7", + "[25] Rohan Sarkar, Navaneeth Bodla, Mariya I Vasileva, Yen-Liang Lin, Anurag Beniwal, Alan Lu, and Gerard Medioni. Outfittransformer: Learning outfit representations for fashion recommendation. In WACV, pages 3601-3609, 2023. 1, 2, 4, 6, 7", + "[26] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 3", + "[27] Reuben Tan, Mariya I Vasileva, Kate Saenko, and Bryan A Plummer. Learning similarity conditions without explicit supervision. In ICCV, pages 10373-10382, 2019. 2" + ], + "bbox": [ + 516, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16173", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 4", + "[29] Mariya I Vasileva, Bryan A Plummer, Krishna Dusad, Shreya Rajpal, Ranjitha Kumar, and David Forsyth. Learning type-aware embeddings for fashion compatibility. In ECCV, pages 390-405, 2018. 1, 2, 4, 5", + "[30] Andreas Veit, Balazs Kovacs, Sean Bell, Julian McAuley, Kavita Bala, and Serge Belongie. Learning visual clothing style with heterogeneous dyadic co-occurrences. In ICCV, pages 4642–4650, 2015. 1, 2", + "[31] Shuohang Wang, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. Want to reduce labeling cost? gpt-3 can help, 2021. 5", + "[32] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 6", + "[33] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6" + ], + "bbox": [ + 91, + 92, + 485, + 457 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16174", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/bbd1b25f-9206-484c-b93d-8a139709e547_model.json b/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/bbd1b25f-9206-484c-b93d-8a139709e547_model.json new file mode 100644 index 0000000000000000000000000000000000000000..56951ca5349826ac1fde6f976900c47a80f5c2be --- /dev/null +++ b/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/bbd1b25f-9206-484c-b93d-8a139709e547_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.097, + 0.131, + 0.9, + 0.154 + ], + "angle": 0, + "content": "Text2Outfit: Controllable Outfit Generation with Multimodal Language Models" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.179, + 0.815, + 0.252 + ], + "angle": 0, + "content": "Yuanhao Zhai\\(^{1*}\\) Yen-Liang Lin\\(^{2}\\) Minxu Peng\\(^{2}\\) Larry S. Davis\\(^{2}\\) Ashwin Chandramouli\\(^{2}\\) Junsong Yuan\\(^{1}\\) David Doermann\\(^{1}\\) \\(^{1}\\)State University of New York at Buffalo \\(^{2}\\)Amazon" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.284, + 0.327, + 0.301 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.317, + 0.486, + 0.681 + ], + "angle": 0, + "content": "Existing outfit recommendation frameworks focus on outfit compatibility prediction and complementary item retrieval. We present a text-driven outfit generation framework, Text2Outfit, which generates outfits controlled by text prompts. Our framework supports two forms of outfit recommendation: 1) Text-to-outfit generation, where the prompt includes the specification for each outfit item (e.g., product features), and the model retrieves items that match the prompt and are stylistically compatible. 2) Seed-to-outfit generation, where the prompt includes the specification for a seed item, and the model both predicts which product types the outfit should include (referred to as composition generation) and retrieves the remaining items to build an outfit. We develop a large language model (LLM) framework that learns the cross-modal mapping between text and image set, and predicts a set of embeddings and compositions to retrieve outfit items. We devise an attention masking mechanism in LLM to handle the alignment between text descriptions and image tokens from different categories. We conduct experiments on the Polyvore dataset and evaluate the quality of the generated outfits from two perspectives: 1) feature matching for outfit items, and 2) outfit visual compatibility. The results demonstrate that our approach significantly outperforms the baseline methods in text to outfit generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Fashion outfit recommendation system constructs outfits from items that go well together (e.g., an outfit that includes a top, a bottom, shoes, bags, etc). It is an important recommendation problem for online retailers, as customers frequently shop for items that can be worn together. Existing outfit recommendation methods mainly focus on compatibility prediction (determining whether a set of fashion items in an outfit go well together) [7, 9, 21, 29, 30] or outfit" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.287, + 0.901, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.381, + 0.9, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.495, + 0.907, + 0.62 + ], + "angle": 0, + "content": "Figure 1. A text to outfit generation framework that generates outfits driven by the text prompt. Our framework supports two forms of outfit recommendation. For text-to-outfit generation, the prompt includes outfit-level and item-level specifications. The model retrieves outfit items that match the texts in the prompt and are stylistically compatible. For seed-to-outfit generation, the prompt includes the specification for a seed item (a seed image or text description). The model predicts the composition (outfit categories) and retrieves the items to build outfits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.907, + 0.73 + ], + "angle": 0, + "content": "complementary item retrieval (completing a partial outfit by finding a compatible item) [20, 25]. The outfit items are predicted by a pre-trained model and are not controlled by the text prompt. For example, generating outfits with desired attributes, e.g., wide-leg sweatpants." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.909, + 0.837 + ], + "angle": 0, + "content": "Recent work ([14, 15]) for text to image retrieval leverages a pre-trained LLM model to learn image embeddings. They perform better retrieval than non-LLM based approaches (e.g., CLIP [24]) especially for longer text descriptions. However, these approaches focus on retrieving single images so they are not designed for retrieving image sets like outfits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We present an outfit generation framework (Text2Outfit) that generates outfits driven by a text prompt (cf. Figure 1). Our framework learns the cross-modal mapping (text to image set) and predict the embeddings to retrieve outfits" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.396, + 0.901 + ], + "angle": 0, + "content": "*This work was done during an internship at Amazon." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16165" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.364 + ], + "angle": 0, + "content": "that match the product features specified in the text prompt and are visually compatible. We instantiate our framework in two outfit recommendation scenarios using text prompts. In the first scenario (text-to-outfit generation), users specify the text prompt that describes both outfit-level features (e.g., an outfit for spring and causal) and individual items (e.g., a blouse in white with an off-shoulder design). Our model retrieves the complementary outfit items that match the prompt to build outfits. In the second scenario (seed-to-outfit generation), users specify the requirements for a few items (e.g., a seed item), using either images or texts, e.g., a white knitted sweater, and the model both completes the composition and retrieves complementary items to complete the outfit. We consider the case of just one seed item, but the framework can be extended to include multiple seed items. To build complete outfits, we predict the outfit composition (outfit product categories) for the remaining items and retrieve items in the predicted categories to build outfits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.485, + 0.532 + ], + "angle": 0, + "content": "We use learnable image tokens to generate embeddings for image retrieval similar to ([14, 15]). In contrast to prior work, we use a set of image tokens to retrieve outfits (image set) beyond a single image. Specifically, we incorporate a set of image tokens for different product categories (e.g., top, bottom, shoes). For each image token, we use the MLP layers on the top of the hidden state vector from the LLM to generate the embeddings. We train the framework in an end-to-end manner (fine-tune both LLM model and MLP layers) using LoRA [10] with a loss function that considers both product feature matching and outfit compatibility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.485, + 0.64 + ], + "angle": 0, + "content": "The prompt includes the texts describing multiple items, which introduces the challenge of aligning the outfit texts and image tokens. It is important to avoid the case - where the image token from one category wrongly attends to the text descriptions from other categories. We present a masking mechanism in the self-attention layers to facilitate the needed attention learning (cf. Section 3.1 and Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.485, + 0.763 + ], + "angle": 0, + "content": "For seed-to-outfit generation, since we only have one seed item as input, we must select the composition (outfit categories) for the remaining items. For example, for specific occasions/seasons, some categories should not be included. So, for summer outfits, the coat category should not be included in the final outfit. We train the LLM model to predict outfit compositions (outfit category combinations) and retrieve the items in the predicted categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We evaluate our framework on the public Polyvore Outfit dataset [29] and measure the quality of generated outfits from two perspectives: 1) feature matching for outfit items and 2) outfit compatibility. Since the Polyvore dataset does not have text product attributes, we use Claude-3 [3] to extract the product features for each outfit item (we verify the precision of Claude-3 with human annotations in Section 4.1). The experiment section shows that our approach outperforms the baseline approaches in the product feature matching and" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.649, + 0.107 + ], + "angle": 0, + "content": "outfit compatibility." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.108, + 0.907, + 0.137 + ], + "angle": 0, + "content": "To summarize, the main technical contributions of our work are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.139, + 0.907, + 0.183 + ], + "angle": 0, + "content": "- We propose a text-driven outfit generation framework that generates outfits based on text prompts or images, facilitating outfit recommendation in two forms." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.184, + 0.907, + 0.243 + ], + "angle": 0, + "content": "- We develop a large language model framework to predict the embeddings (and compositions) to retrieve the outfit items where they match the input prompt and are visually compatible." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.245, + 0.905, + 0.289 + ], + "angle": 0, + "content": "- We devise an attention masking mechanism to handle the alignment between the text descriptions in the prompt and the image tokens from different outfit categories." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.139, + 0.907, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.306, + 0.655, + 0.322 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.332, + 0.736, + 0.346 + ], + "angle": 0, + "content": "2.1. Outfit Recommendation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.355, + 0.909, + 0.656 + ], + "angle": 0, + "content": "Outfit Compatibility Prediction: Outfit compatibility prediction evaluates how well a set of fashion items are compatible with each other in an outfit. Early works [21, 30] focused on pairwise item-level compatibility by leveraging item co-view or co-purchase information. However, compatibility is determined only by the pairwise items, not the entire outfits. Later works such as [7, 9, 27, 29] addressed the compatibility problem at the outfit level. They use different types of architectures to model the relationship between outfit items (e.g., bidirectional LSTM [9], subspace embeddings [27, 29] and graph convolution networks (GCN) [7]). These methods can be used for the fill-in-the-blank (FITB) problem - given a subset of items in an outfit and a set of candidate items, select the most compatible candidate. [4] extends FITB to build an entire outfit from a seed item - it starts from a seed item and iteratively finds the complementary items to build an outfit. However, the composition (outfit category combination) is pre-determined from the existing outfits. It is unclear how to compute the compositions for the new seed items." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.658, + 0.907, + 0.718 + ], + "angle": 0, + "content": "These approaches are mainly designed for compatibility prediction and FITB. In contrast, our framework generates outfits based on the input text prompt and predicts the outfit composition given a seed item." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Outfit Complementary Item Retrieval: Complementary item retrieval is to find a compatible item(s) to complete a partial outfit from a large database. Lin et al. [20] designed a category-based attention mechanism that enables scalable indexing and searching for complementary items. Sarka et al. [25] generated the embeddings using a transformer-based architecture that encodes the compatibility for the partial outfit. These approaches address the FITB problem in a scalable manner (support large-scale indexing and complementary item retrieval), which differs from our problem settings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Outfit Retrieval: CLIP4Outfit [11] trained transformer encoders to learn cross-modal similarity (the similarity be" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "16166" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.258 + ], + "angle": 0, + "content": "tween text descriptions and outfit images) and used outfit-level embeddings to retrieve the most relevant outfit given the text descriptions. However, this approach considers the whole outfit as a candidate for retrieval. To be able to match a variety of text descriptions, the approach requires constructing a large-scale outfit dataset that covers different types of outfits, which is not feasible in practice. Moreover, their framework does not support outfit generation based on a given seed item. In contrast, our model is an outfit-building approach, retrieving individual items and assembling them into complete outfits." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.267, + 0.313, + 0.283 + ], + "angle": 0, + "content": "2.2. Text-to-Image Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.289, + 0.486, + 0.606 + ], + "angle": 0, + "content": "Most related work focuses on text-to-single image retrieval. CLIP [12] learned joint representations of text and images through large-scale contrastive learning. These representations can be fine-tuned on the domain-specific datasets to achieve higher performance, such as Fashion-CLIP [6] for the fashion domain. Later works (e.g., Flamingo [1], BLIP [16] and BLIP-2 [17]) learned a unified vision-language representation space and translated visual features into LLM to generate captions for images. The unified vision-language representation (e.g., Q-Former [17]) can be used for text-to-image retrieval. Some recent work leverages the LLM model to predict embeddings for retrieval [14, 15], which performs better than CLIP embeddings for longer captions. FROMAGe [15] takes the interleaved image and text as input and produce image embeddings and text as output. GILL [14] extends FROMAGe to support image retrieval and image synthesis. Similar to [14, 15], our framework also uses learnable image tokens for retrieval. In contrast to prior work on single image retrieval, our approach focuses on retrieving multiple items to build the outfits (image set retrieval)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.614, + 0.364, + 0.631 + ], + "angle": 0, + "content": "2.3. Text-to-Image/Outfit Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.486, + 0.818 + ], + "angle": 0, + "content": "Most work focuses on generating a single image given the text prompt [5, 8, 23, 26]. Some recent work [13, 18] generates a full-body human image wearing garments that reflect the provided textual descriptions. However, the synthesized images may not fully preserve the attribute details specified in the prompt as mentioned by [18]. In addition, to obtain shoppable content, these approaches require using a retrieval system to retrieve similar images from the product catalog (the generated images may not be able to find similar catalog images). In contrast, our framework is designed for outfit generation and retrieves outfit items that preserve the attributes specified in the prompt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.183, + 0.847 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Figure 2 illustrates the system overview of our framework. Our framework supports both forms: 1) text-to-outfit generation and 2) seed-to-outfit generation. For text-to-outfit" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.242 + ], + "angle": 0, + "content": "generation, the task is to find the outfit items that match the text descriptions described in the input prompt. We use a set of image tokens and learn the image embeddings for different categories. We fine-tune the LLM to learn the embeddings using the outfit retrieval and compatibility losses, which encourage embeddings to retrieve the items that match the product features and are visually compatible. We devise a masking attention approach to handle the alignment between the text descriptions in the prompt and the image tokens. Details are described in Section 3.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.909, + 0.349 + ], + "angle": 0, + "content": "For seed-to-outfit generation, the task is to find outfit items to complete an outfit given a seed item and a prompt that describes the global outfit features. We train the LLM model to predict the outfit compositions (outfit category combination) and the embeddings to search the relevant outfit items to generate outfits. Details are described in Section 3.2" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.361, + 0.749, + 0.377 + ], + "angle": 0, + "content": "3.1. Text-to-Outfit Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.909, + 0.49 + ], + "angle": 0, + "content": "We are given an input text prompt: \\( T = \\{T_{1}, T_{2}, \\dots, T_{k}\\} \\), where each text description \\( T_{i} \\) describes the outfit-level specification (e.g., outfit for spring and casual) or the item-level specification (e.g., a blouse in white with off-shoulder design). The order of the text descriptions for each category is arbitrary. The goal is to find the outfit items that match the text descriptions in the prompt (cf. Figure 2 (a))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.909, + 0.642 + ], + "angle": 0, + "content": "To produce embeddings, we add a set of image tokens to LLM: \\(\\mathrm{[IMG_{cat_1},\\dots,IMG_{cat_n}]}\\); each image token corresponds to a high-level category (we use 11 image tokens for Polyvore dataset; the category list is described in Section 4.3). The hidden state vectors of these image tokens from LLM are denoted as \\(\\mathrm{h_{cat_1}},\\ldots,\\mathrm{h_{cat_n}}\\). We learn a set of MLP layers: \\(\\mathrm{W}_{\\mathrm{ret}} = \\{\\mathrm{W}_{\\mathrm{cat}_i}^{\\mathrm{ret}},\\ldots,\\mathrm{W}_{\\mathrm{cat}_n}^{\\mathrm{ret}}\\}\\) to transform the hidden state vectors to the embeddings: \\(\\mathbf{e} = \\{\\mathbf{e}_{\\mathrm{cat}_1},\\dots,\\mathbf{e}_{\\mathrm{cat}_n}\\}\\), where embedding \\(\\mathbf{e}_{\\mathrm{cat}_i}\\) is used to retrieve the outfit images for the category \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.909, + 0.763 + ], + "angle": 0, + "content": "An outfit item is represented by an image I and its text descriptions T. We use a visual encoder \\( E_{\\mathrm{img}} \\) to extract the visual embeddings from the image and use a text encoder \\( E_{\\mathrm{txt}} \\) to extract text embeddings from the text descriptions. The multi-modal feature of an outfit item is denoted by: \\( \\mathrm{f} = \\mathrm{W}_{\\mathrm{item}} \\cdot (E_{\\mathrm{img}}(\\mathrm{I}) \\| E_{\\mathrm{txt}}(\\mathrm{T})) \\), where \\( \\| \\) is the concatenation and \\( \\mathrm{W}_{\\mathrm{item}} \\) is a learnable MLP model that is fine-tuned with the outfit retrieval loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.909, + 0.884 + ], + "angle": 0, + "content": "We train our framework end-to-end using the following losses: outfit retrieval loss, attention mask loss, and outfit compatibility loss. The outfit retrieval loss retrieves items that match the product features mentioned in the prompt. The attention mask loss handles the alignment between the outfit text descriptions and the image tokens. The outfit compatibility loss improves the visual compatibility of the retrieved outfit items." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Outfit retrieval loss. A training batch includes a set of posi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16167" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.093, + 0.317, + 0.106 + ], + "angle": 0, + "content": "(a) Text-to-outfit generation (Sec. 3.1)" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.115, + 0.468, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.476, + 0.093, + 0.689, + 0.106 + ], + "angle": 0, + "content": "(b) Seed-to-outfit generation (Sec 3.2)" + }, + { + "type": "image", + "bbox": [ + 0.473, + 0.118, + 0.904, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.315, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Figure 2. System overview of the text to outfit generation framework. For text-to-outfit generation (a), the LLM framework predicts a set of embeddings for different outfit categories, which are used to retrieve the outfit items. We devise an attention masking approach to handle the alignment between the text descriptions in the prompt and the image tokens from different categories (details are in Section 3.1). For seed-to-outfit generation (b), the LLM framework takes a seed item as input, predicts the outfit compositions (outfit categories), and retrieves items in the predicted categories to build the outfits (details are in Section 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.412, + 0.484, + 0.605 + ], + "angle": 0, + "content": "tive outfits. We randomly sample a set of negative outfits for each positive outfit, where each item in the negative outfits has the same high-level category as the positive outfit. Since there are no ground truth negative outfits, random negatives are typically used in outfit training (e.g., [20, 25, 29]). We extract the occasion, season, color and product features of the outfit items using Claude-3 (details are in Section 4.1). To simulate the prompt for training, we randomly select the product features of each outfit item and use Claude-3 to generate the free-form prompt. The goal is to learn the embeddings so that the similarity between the predicted embeddings and the positive outfits is greater than the negative ones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.667 + ], + "angle": 0, + "content": "Specifically, a batch includes a set of triples: \\(\\Upsilon = \\{T, p, N\\}\\), where \\(T\\) is the prompt, \\(p\\) is the positive outfit and \\(N\\) is a set of negative outfits. The outfit retrieval loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.675, + 0.484, + 0.715 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e t}} (p, N; \\mathrm {e}) = \\sum_ {i = 1} ^ {n} \\mathcal {L} _ {\\text {c a t} _ {i}} \\left(\\mathrm {f} _ {\\text {c a t} _ {i}} ^ {p}, \\mathrm {f} _ {\\text {c a t} _ {i}} ^ {N}; \\mathrm {e} _ {\\text {c a t} _ {i}}\\right) \\xi (\\text {c a t} _ {i}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.485, + 0.784 + ], + "angle": 0, + "content": "We combine the loss for each category to obtain the outfit retrieval loss. \\(\\xi\\) is the indicator function denoting if the category \\(i\\) is presented in the outfit. The loss for each category is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.79, + 0.493, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c a t} _ {i}} \\left(\\mathrm {f} _ {\\mathrm {c a t} _ {i}} ^ {p}, \\mathrm {f} _ {\\mathrm {c a t} _ {i}} ^ {N}; \\mathrm {e} _ {\\mathrm {c a t} _ {i}}\\right) = - \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathrm {f} _ {\\mathrm {c a t} _ {i}} ^ {p} , \\mathrm {e} _ {\\mathrm {c a t} _ {i}}\\right) / \\tau\\right)}{\\sum_ {n = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(\\mathrm {f} _ {\\mathrm {c a t} _ {i}} ^ {n} , \\mathrm {e} _ {\\mathrm {c a t} _ {i}}\\right) / \\tau\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.9 + ], + "angle": 0, + "content": "We adopt the InfoNCE loss ([28]). \\( \\mathrm{f}_{\\mathrm{cat}_i}^p \\) is the feature for positive outfit item and \\( \\mathrm{f}_{\\mathrm{cat}_i}^n \\) is the feature for negative outfit item. \\( \\mathrm{e}_{\\mathrm{cat}_i} \\) is the embeddings. \\( \\tau \\) is the temperature. We use the cosine function for the sim function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.908, + 0.622 + ], + "angle": 0, + "content": "Attention mask loss. The prompt includes the text descriptions for multiple categories. We observed the alignment problem where the image token from a category may attend to the text descriptions from other categories. For example, the top image token attends to the shoe text descriptions. An example prompt: \"An outfit for spring and formal event. Black high heels. A black dress with a rose print. A red bracelet with gemstones.\" The text describes the features for fine-grained categories like high-heel, dress, etc, but the image tokens correspond to high-level categories (e.g., tops, bottoms, shoes). These pose the challenge for the LLM model to find the relations between the outfit text descriptions and multiple image tokens, which affects the product feature matching performance (cf. Table 3))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.909, + 0.744 + ], + "angle": 0, + "content": "To address this issue, we design a masking approach in the self-attention layers of the LLM decoder. Figure 3 illustrates the mask attention for the top image token. We incorporate the mask loss to guide attention learning, where the mask loss is applied to all image tokens. We compute the binary cross-entropy loss (BCE) for each image token between the ground truth mask and the predicted attention values. The mask loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.755, + 0.907, + 0.798 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {k} \\mathrm {B C E} _ {i} \\left(\\alpha_ {j}, m _ {j} ^ {\\mathrm {G T}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.908, + 0.855 + ], + "angle": 0, + "content": "where \\( n \\) is the number of image tokens. \\( k \\) is the number of tokens from the text descriptions. \\( \\alpha_{j} \\) is the predicted attention value and \\( m_j^{\\mathrm{GT}} \\) is the ground truth mask." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Compatibility loss (CP. loss). We include a compatibility loss to facilitate the embedding learning that improves the compatibility of the generated outfit items. We trained an" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16168" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.483, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.321, + 0.485, + 0.378 + ], + "angle": 0, + "content": "Figure 3. The figure illustrates the masked attention for the top image token, where the mask facilitates the top image token to attend to the right text descriptions (outfit text and top text sections, highlighted in yellow) in the prompt. Details are in Section 3.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.493 + ], + "angle": 0, + "content": "outfit compatibility model (details are in Section 4.2), which predicts a compatibility score given a set of outfit items. In training, we freeze the model weights of the compatibility model (denoted as \\(\\mathcal{C}\\)) and fine-tune the embeddings (e) such that the compatibility score is maximized. The compatibility loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.505, + 0.484, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c p}} (e) = - \\mathcal {C} \\left(\\mathrm {W} _ {\\mathrm {c p}} \\cdot \\mathrm {e}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.483, + 0.564 + ], + "angle": 0, + "content": "Training and Inference. We train the framework end-to-end using LoRA [2] by minimizing the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.573, + 0.483, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta , \\mathrm {W} _ {\\text {r e t}}, \\mathrm {W} _ {\\text {i t e m}}, \\mathrm {W} _ {\\mathrm {c p}}} \\mathcal {L} \\left(\\theta , \\mathrm {W} _ {\\text {r e t}}, \\mathrm {W} _ {\\text {i t e m}}, \\mathrm {W} _ {\\mathrm {c p}}\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.602, + 0.426, + 0.619 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {1} \\cdot \\mathcal {L} _ {\\mathrm {r e t}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\mathrm {m a s k}} + \\lambda_ {3} \\cdot \\mathcal {L} _ {\\mathrm {c p}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.734 + ], + "angle": 0, + "content": "where \\(\\theta\\) is LLM model weights. \\(W_{\\mathrm{ret}}\\), \\(W_{\\mathrm{item}}\\), and \\(W_{\\mathrm{cp}}\\) are the MLP models for outfit retrieval, outfit item feature extraction, and compatibility, respectively. \\(\\mathcal{L}_{\\mathrm{ret}}\\) is the outfit retrieval loss, \\(\\mathcal{L}_{\\mathrm{mask}}\\) is the mask loss and \\(\\mathcal{L}_{\\mathrm{cp}}\\) is the compatibility loss. \\(\\lambda_1, \\lambda_2\\), and \\(\\lambda_3\\) are empirically chosen to make the magnitudes of the losses similar in scale. We report the results of using different losses in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.484, + 0.779 + ], + "angle": 0, + "content": "In the inference, we feed the text prompt to the LLM model to predict the embeddings and use them to retrieve the outfit items." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.328, + 0.803 + ], + "angle": 0, + "content": "3.2. Seed-to-Outfit Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The task is to generate a complete outfit given an input seed item and a text prompt describing the global outfit features. Unlike the text-to-outfit generation task, where users specify the texts for each outfit item, here, users only specify the information for the seed item. Therefore, we need to predict the composition of the remaining items. One can retrieve" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "the outfit items from all the categories to build an outfit. However, some categories do not go well with certain types of occasions, seasons, or the seed types. For example, coats should not be included for summer outfits, or an all-body dress should not be selected if the seed is a top." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.349 + ], + "angle": 0, + "content": "Composition loss. To address this problem, we train the LLM model to predict the outfit compositions (a set of outfit categories), where the compositions are learned from outfit datasets (cf. Figure 2 (b)). Specifically, the input includes a prompt and a seed item (users provide an image or a text description for the seed item. For the image, we use a visual encoder to extract the embeddings and feed them to the LLM). To predict the compositions, we add two MLP models (one for predicting the high-level category and another for predicting the fine-grained category) on the top of the hidden state vectors of the high-level category to predict the composition. The composition loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.373, + 0.907, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o m p}} = \\sum_ {i = 1} ^ {n} \\mathrm {C E} _ {i} \\left(y _ {\\mathrm {p r e d}} ^ {\\mathrm {h c}}, y _ {\\mathrm {g t}} ^ {\\mathrm {h c}}\\right) + \\mathrm {C E} _ {i} \\left(y _ {\\mathrm {p r e d}} ^ {\\mathrm {f g}}, y _ {\\mathrm {g t}} ^ {\\mathrm {f g}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.908, + 0.483 + ], + "angle": 0, + "content": "where the hc denotes the high-level category and fg denotes the fine-grained category. CE is the cross-entropy loss. We train the framework with LoRA end-to-end with the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.495, + 0.907, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {4} \\cdot \\mathcal {L} _ {\\text {r e t}} + \\lambda_ {5} \\cdot \\mathcal {L} _ {\\mathrm {c p}} + \\lambda_ {6} \\cdot \\mathcal {L} _ {\\text {c o m p}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.629 + ], + "angle": 0, + "content": "The loss function includes the retrieval, compatibility, and composition prediction losses. We train the model to predict the composition of the seed item belonging to the top category. The performance of other categories for the seed item is in Appendix Table 2. In inference, we search for outfit items in the predicted high-level and fine-grained categories to compose an outfit." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.643, + 0.646, + 0.66 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.61, + 0.683 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.825 + ], + "angle": 0, + "content": "We evaluate the model performance on the Polyvore dataset [29]. The dataset includes two different sets: a disjoint set and a non-disjoint set. The non-disjoint set contains 53,306 training outfits and 10,000 testing outfits, while the disjoint set contains 16,995 training outfits and 15,145 testing outfits. We train our framework on the non-disjoint training set as it contains a larger number of outfits and report the performance on both non-disjoint and dis-joint testing sets (results are in the Appendix Section 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The Polyvore dataset does not have annotated product attributes. Motivated by recent works that leverage LLM for annotations (e.g., [19, 31]), we leverage Claude-3 [3] to extract outfit-level attributes (occasion and season) and item-level attributes (color and product features). We verify" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16169" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.198 + ], + "angle": 0, + "content": "the agreement between the predicted attributes from Claude-3 and humans on a subset of testing samples (500 outfits and 2629 items). The results show the Claude-3 annotated attributes achieve \\(94.7\\%\\) overall precision \\((97.3\\%)\\) for color, \\(95.5\\%\\) for product feature, \\(92.2\\%\\) for occasion, and \\(93.6\\%\\) for season), verifying the feasibility of using Claude-3 for attribute annotations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.209, + 0.212, + 0.224 + ], + "angle": 0, + "content": "4.2. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.231, + 0.486, + 0.426 + ], + "angle": 0, + "content": "We evaluate the quality of the retrieved outfits in two ways: 1) feature matching for outfit items and 2) outfit compatibility. To evaluate feature matching, we compute the number of items in the top-10 retrieval results that match the product features specified in the prompt. We evaluate the feature matching for color, occasion, season, and product feature attributes for the outfit items, which are critical attributes for outfits. To handle non-exact matches, where the product features are different but their meanings are similar, we use ModernBERT embedding [22] to evaluate the affinity of two features (we select the thresholds of different attributes using a validation set). We report the results using precision@top-10 (denoted as P@10)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.485, + 0.609 + ], + "angle": 0, + "content": "To evaluate outfit compatibility, we train an outfit compatibility model using a SOTA architecture (OutfitTransformer [25]), where the model takes a set of outfit items as input, and predicts a compatibility score. We replace the ResNet features in the original architecture with the FashionCLIP embeddings [6] and use the item image and the text description as input. We achieve better outfit compatibility prediction than the original architecture. We achieve \\(94.93\\%\\) (93% for OutfitTransformer) and \\(93.88\\%\\) AUC (88% for OutfitTransformer) for the non-disjoint and disjoint sets respectively. We report the compatibility (CP) score for the top 10 retrieved outfits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.486, + 0.807 + ], + "angle": 0, + "content": "There is no ground-truth outfit composition given a seed and prompt. To obtain a quantitative quality score, we use Claude-3 (we also evaluate the performance with Pixtral Large from Mistral AI, see Appendix Section 2) to evaluate the quality of the predicted compositions. Recent work (e.g., [33]) shows that strong LLMs achieve high agreement with humans. We evaluate the quality in two ways: 1) We use Claude-3 to predict a quality score for each predicted composition and report an average quality score (denoted as Comp. score). 2) Given two compositions generated by the baseline and our approach, we ask Claude-3 to pick the better composition and report the winning rate of our approach (denoted as Comp. win rate)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.352, + 0.835 + ], + "angle": 0, + "content": "4.3. Baselines and Model Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.901 + ], + "angle": 0, + "content": "There is no existing framework for text-to-outfit retrieval tasks ([11] is designed to retrieve entire outfits, which is different from our settings as mentioned in Section 2). We convert a set of baseline approaches to the outfit retrieval" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.242 + ], + "angle": 0, + "content": "tasks and report the performance. In Table 1, we categorize these baselines into the text-to-single-item retrieval and text-to-outfit retrieval approaches. Existing methods for text-to-image retrieval [6, 14, 24] mainly focus on single-image retrieval. To make these methods applicable to outfit retrieval problem, we make the following changes: 1) We convert the outfit-level specification (occasion/season) into item-level specification. 2) We use the text descriptions of each category to retrieve the outfit items (assuming that the text descriptions to category correspondence are given)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.909, + 0.364 + ], + "angle": 0, + "content": "For text-to-outfit retrieval, we converted the GILL [14] framework to support multiple-category retrieval. For text-to-outfit retrieval methods, the text descriptions to category correspondence information are not given, which is directly predicted by the models (except the fixed mask approach in Table 1, which uses the text-to-category correspondence to generate the fixed mask for attention). The details of the baseline approaches are described below." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.366, + 0.905, + 0.41 + ], + "angle": 0, + "content": "- CLIP [24]: We convert the outfit-level specification into item-level specifications and perform retrieval for each category. We use ViT-L-14 version." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.411, + 0.907, + 0.469 + ], + "angle": 0, + "content": "- Fashion-CLIP 2.0 [6]: We use the same settings as CLIP and use Fashion-CLIP to retrieve items for each category. We use version 2.0, which achieves better performance than version 1.0." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.471, + 0.905, + 0.531 + ], + "angle": 0, + "content": "- BLIP-2 [17]: We use the pre-trained image encoder and Q-Former from BLIP-2. Like CLIP, we perform image retrieval for each category using the corresponding text description. We use the ViT-g version." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.532, + 0.905, + 0.652 + ], + "angle": 0, + "content": "- GILL [14] + Multiple categories: We add the learnable image tokens for each high-level category for multiple image retrieval. We fine-tune the MLP layers to learn embeddings while keeping the LLM model fixed following the original training approach. We use the same prompt format as our model. Note that we compare our method with an improved version of GiLL, in which we replace the original CLIP vision encoder with Fashion-CLIP" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.366, + 0.907, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.909, + 0.835 + ], + "angle": 0, + "content": "Model Settings: For our framework, we use the pretrained OPT 6.7B model [32] (same as GILL) as the LLM backbone. We fine-tune the LLM with LoRA (we set the rank to 64 for Q, K, and V projection matrices). We use Fashion-CLIP 2.0 [6] as the visual encoder \\( E_{\\mathrm{img}} \\) and the text encoder \\( E_{\\mathrm{text}} \\). Fashion-CLIP 2.0 achieves better performance than CLIP embeddings. In model training, the visual and text encoder weights are fixed. We use the bi-directional masks for image tokens. We use 11 image tokens for the Polyvore dataset, corresponding to the following high-level categories: tops, bottoms, all-body, outerwear, shoes, bags, jewelry, accessories, sunglasses, hats, and scarves." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.848, + 0.862 + ], + "angle": 0, + "content": "4.4. Text-to-Outfit Performance Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Table 1 summarizes the text-to-outfit retrieval results. We use the testing set of the Polyvore dataset to generate the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16170" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.088, + 0.907, + 0.239 + ], + "angle": 0, + "content": "
MethodModalityColor (P@10)Product feature (P@10)Season (P@10)Occasion (P@10)Avg (P@10)CP. score
Text-to-single-item retrieval
CLIP [24]Text only96.0670.9190.2276.6283.4663.49
Fashion-CLIP [6]Text only96.8280.4289.6881.2787.0564.07
BLIP-2 [17]Text only93.2966.9485.4684.5182.5565.09
CLIP [24]Text to image83.5961.3285.4967.0174.3549.77
Fashion-CLIP [6]Text to image87.5972.6785.9168.8878.7653.03
BLIP-2 [17]Text to image86.3966.6985.6267.7576.6151.05
Text-to-outfit retrieval
GILL [14] + multiple categoryText to image48.1948.1586.8769.6963.2374.26
Text2Outfit + learnable mask (ours)Text to image78.4276.5486.6070.4578.0066.28
Text2Outfit + learnable mask (ours)Text to image/text86.0584.8195.6993.6690.0584.05
Text2Outfit + fixed mask (ours)Text to image/text96.6186.2794.8792.4892.5682.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.249, + 0.907, + 0.306 + ], + "angle": 0, + "content": "Table 1. Text-to-outfit retrieval performance comparison. P@10 denotes precision at top-10 and CP. score denotes compatibility score using [25] for top-10 retrieved outfits. Note that for single-item retrieval approaches, the text descriptions and its corresponding category are given while the outfit retrieval approaches are not given such information. The bold text denotes the best performance and the underlines denote the second best performance. Details are in Section 4.2." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.318, + 0.905, + 0.381 + ], + "angle": 0, + "content": "
MethodCompositionSeason (P@10)Occasion (P@10)Comp. score (c3)Comp win rate (c3)CP. scoreCP. score (c3)
Seed text + prompt (baseline)All categories97.3296.202.596%80.822.78
Seed text + prompt (ours)Predicted97.5696.014.5794%86.183.85
Seed image + prompt (baseline)All categories95.0087.502.858%72.882.66
Seed image + prompt (ours)Predicted96.4391.804.3292%83.363.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.39, + 0.89, + 0.404 + ], + "angle": 0, + "content": "Table 2. Seed-to-outfit retrieval performance comparison. The scores with (c3) are evaluated by Claude-3. Details are in Section 4.5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.483, + 0.506 + ], + "angle": 0, + "content": "prompt. Given a test outfit including multiple outfit items, we randomly select outfit attributes (occasion and season) and item attributes (color and product features) and use Claude-3 to generate the free-form prompt. We use the generated prompt to evaluate the outfit retrieval performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.507, + 0.483, + 0.717 + ], + "angle": 0, + "content": "Our approaches achieve \\(90.05\\%\\) and \\(92.56\\%\\) average \\(\\mathrm{P@10}\\) using the learnable mask and fixed mask respectively, showing better feature-matching precision than the baseline approaches. The GILL + multi-category approach has an alignment issue (the alignment between the outfit texts and the images tokens) while our approach uses the learnable mask to alleviate the alignment issue. We fine-tune the model end-to-end with LoRA instead of using the frozen LLM model as GILL, improving the performance (see Table 3 for detailed analysis). Our approach with a learnable mask also performs better than the single-item retrieval approaches, where they are given the text description to category correspondence information, while our approach does not." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "In addition, our approach shows better outfit compatibility than the baseline approaches. We achieve the compatibility score 84.05 using the learnable mask setting (we compute the average compatibility scores for the ground truth positive (86.95) and negative (22.99) outfits on the Polyvore testing set using OutfitTransformer. The generated outfits from our method achieve similar compatibility as the positive outfits). Single-item retrieval approaches perform worse on outfit compatibility because they find items that match the text descriptions but do not model the item relationships in an outfit. Our framework is fine-tuned on the outfit data, and incorporates the outfit retrieval and compatibility losses," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.461 + ], + "angle": 0, + "content": "which improve overall compatibility. Figure 4 shows example results of our framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.851, + 0.485 + ], + "angle": 0, + "content": "4.5. Seed-to-Outfit Performance Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.909, + 0.581 + ], + "angle": 0, + "content": "Table 2 summarizes the results for seed-to-outfit retrieval. We used the Polyvore test set to construct the input (seed and prompt) for evaluation. Given a test outfit including multiple outfit items, we randomly select the outfit attributes (occasion and season) and use Claude-3 to generate the free-form prompt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.909, + 0.84 + ], + "angle": 0, + "content": "We compare our approach with a baseline approach that uses all 11 high-level categories for the composition. We feed the compositions (e.g., category 1/fine-grained category 1, category 2/fine-grained category 2, ...) of the baseline and our approach to Claude-3 for evaluation (for the baseline approach, the fine-grained category is obtained from the top-1 retrieved outfit). Our approach shows better composition scores (the score is between 1 (poor) to 5 (excellent)) and winning rates than the baseline (e.g., \\(94\\%\\) and \\(92\\%\\) winning rate of using seed text and seed image, respectively). We verify the agreement between Claude-3 and humans on a subset of samples (500 outfitsits). The Spearman's rank correlation coefficient is 0.73, which indicates they are highly correlated. We report the visual compatibility using both OutfitTransformer [25] (CP. score) and Claude-3 (CP. score (c3)). Our approach shows better visual compatibility than the baseline approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.667, + 0.866 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Multimodal feature: We compare the performance of single-modality (image) vs. multimodality (image + text)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16171" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.088, + 0.907, + 0.175 + ], + "angle": 0, + "content": "
ModalityLoRAMask LossCP. lossColor (P@10)Product feature (P@10)Season (P@10)Occasion (P@10)Avg (P@10)CP. score
1Image---48.3153.0385.1068.6263.7653.43
2Image + text---44.6047.1888.9481.0765.4578.62
3Image + text✓--74.7475.0195.6093.9084.8286.95
4Image + text✓Global CE-79.2882.2396.0294.4387.9985.63
5Image + text✓Local BCE-85.5884.8795.8993.4789.9583.43
6Image + text✓Local BCE✓86.0584.8195.6993.6690.0584.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.191, + 0.185, + 0.804, + 0.199 + ], + "angle": 0, + "content": "Table 3. Ablation study of different components in our framework. Details are described in Section 4.6." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.219, + 0.314, + 0.236 + ], + "angle": 0, + "content": "(a) Text-to-outfit generation" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.245, + 0.388, + 0.29 + ], + "angle": 0, + "content": "Prompt: Spring outfit for casual. A canvas upper flats in a pink shade. A jeans in black with high-waisted accents. A sleeveless top in a white shade. A denim jacket in a blue shade." + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.305, + 0.239, + 0.315 + ], + "angle": 0, + "content": "GILL + Multiple Category" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.317, + 0.174, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.317, + 0.249, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.317, + 0.324, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.317, + 0.401, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.386, + 0.291, + 0.396 + ], + "angle": 0, + "content": "Text2Outfit + learnable mask (ours)" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.398, + 0.174, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.398, + 0.249, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.398, + 0.323, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.398, + 0.4, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.471, + 0.268, + 0.482 + ], + "angle": 0, + "content": "Text2Outfit + fixed mask (ours)" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.484, + 0.174, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.484, + 0.249, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.261, + 0.485, + 0.323, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.334, + 0.485, + 0.4, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.42, + 0.219, + 0.633, + 0.235 + ], + "angle": 0, + "content": "(b) Seed-to-outfit generation" + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.244, + 0.593, + 0.256 + ], + "angle": 0, + "content": "Prompt: Outfit for casual in spring." + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.257, + 0.896, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.42, + 0.401, + 0.66, + 0.413 + ], + "angle": 0, + "content": "Prompt: Formal events outfit for spring weather." + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.414, + 0.896, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.568, + 0.907, + 0.623 + ], + "angle": 0, + "content": "Figure 4. Qualitative results. We compare the results for text-to-outfit generation with the extended GILL [14] approach that supports multiple image retrieval and our framework (learnable mask and fixed mask). The non-matched items are marked with boxes. We show the predicted compositions and retrieval results for seed-to-outfit generation with a seed image or text. More visual results are in the Appendix Section 4." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.65, + 0.486, + 0.697 + ], + "angle": 0, + "content": "(row-1 and row-2 in Table 3). Using multimodality improves the average P@10 from \\(63.76\\%\\) to \\(65.45\\%\\) for feature matching." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.484, + 0.776 + ], + "angle": 0, + "content": "LoRA: We compare the performance of fine-tuned LLM (with LoRA) vs. frozen LLM (row-2 and row-3 in Table 3). Using LoRA significantly improves the average P@10 from \\(65.45\\%\\) to \\(84.82\\%\\) and improves the compatibility score from \\(78.62\\%\\) to \\(86.95\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Mask loss: We investigate different loss functions for the attention masking: Global CE and Local BCE (row-4 and row-5 in Table 3). For Global CE, we consider the attention values from all the text tokens to each image token as a distribution and optimize a global cross-entropy loss. We optimize each attention value for Local BCE with a binary cross-entropy loss. Local BCE performs better, as it optimizes the attention directly with the ground truth mask," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.679 + ], + "angle": 0, + "content": "while global CE optimizes the overall distribution of the attention." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.907, + 0.727 + ], + "angle": 0, + "content": "Compatibility loss: We analyze the impact of including a compatibility loss in the training loss (cf. Equation 5). It further improves compatibility (row-6 in Table 3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.737, + 0.641, + 0.753 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We present a framework for text-based outfit generation. We explore two different forms of outfit generation: 1) text-to-outfit generation and 2) seed-to-outfit generation tasks, which enable the controllable outfit generation. In addition, we introduce a masking mechanism which better handles the alignment between outfit texts and the image tokens, and achieves comparable performance when using the ground-truth masks. The experimental results demonstrate that our model outperforms the baseline approaches for text to outfit retrieval tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16172" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.119, + 0.188, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.145, + 0.486, + 0.213 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.215, + 0.485, + 0.325 + ], + "angle": 0, + "content": "[2] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, et al. Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2, pages 929-947, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.485, + 0.367 + ], + "angle": 0, + "content": "[3] Anthropic. Introducing the next generation of claude. https://www.anthropic.com/news/claude-3-family, 2024.2,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.483, + 0.409 + ], + "angle": 0, + "content": "[4] Federico Becattini, Federico Maria Teotini, and Alberto Del Bimbo. Transformer-based graph neural networks for outfit generation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.48 + ], + "angle": 0, + "content": "[5] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.549 + ], + "angle": 0, + "content": "[6] Patrick John Chia, Giuseppe Attanasio, Federico Bianchi, Silvia Terragni, Ana Rita Magalhães, Diogo Goncalves, Ciro Greco, and Jacopo Tagliabue. Contrastive language and vision learning of general fashion concepts. Scientific Reports, 12 (1):18958, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.592 + ], + "angle": 0, + "content": "[7] Guillem Cucurull, Perouz Taslakian, and David Vazquez. Context-aware visual compatibility prediction. In CVPR, pages 12617-12626, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[8] Runpei Dong, Chunrui Han, Yang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. DreamLLM: Synergistic multimodal comprehension and creation. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[9] Xintong Han, Zuxuan Wu, Yu-Gang Jiang, and Larry S Davis. Learning fashion compatibility with bidirectional lstms. In ACM MM, pages 1078-1086, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.485, + 0.746 + ], + "angle": 0, + "content": "[10] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[11] Junkyu Jang, Eugene Hwang, and Sung-Hyuk Park. Lost your style? navigating with semantic-level approach for text-to-outfit retrieval. In WACV, pages 8066-8075, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[12] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.145, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.094, + 0.906, + 0.12 + ], + "angle": 0, + "content": "controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[14] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. NeurIPS, 36, 2023. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.908, + 0.219 + ], + "angle": 0, + "content": "[15] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In ICML, pages 17283-17300. PMLR, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.907, + 0.276 + ], + "angle": 0, + "content": "[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.906, + 0.376 + ], + "angle": 0, + "content": "[18] Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, Kwan-Yee Lin, and Wayne Wu. Cosmicman: A text-to-image foundation model for humans. In CVPR, pages 6955-6965, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.379, + 0.907, + 0.446 + ], + "angle": 0, + "content": "[19] Wei Li, Hehe Fan, Yongkang Wong, Yi Yang, and Mohan Kankanhalli. Improving context understanding in multimodal large language models via multimodal composition learning. In *Forty-first International Conference on Machine Learning*, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.907, + 0.489 + ], + "angle": 0, + "content": "[20] Yen-Liang Lin, Son Tran, and Larry S Davis. Fashion outfit complementary item retrieval. In CVPR, pages 3311-3319, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.907, + 0.56 + ], + "angle": 0, + "content": "[21] Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, pages 43-52, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[22] Zach Nussbaum, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. Nomic embed: Training a reproducible long context text embedder, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.907, + 0.659 + ], + "angle": 0, + "content": "[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.662, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.908, + 0.8 + ], + "angle": 0, + "content": "[25] Rohan Sarkar, Navaneeth Bodla, Mariya I Vasileva, Yen-Liang Lin, Anurag Beniwal, Alan Lu, and Gerard Medioni. Outfittransformer: Learning outfit representations for fashion recommendation. In WACV, pages 3601-3609, 2023. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[26] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[27] Reuben Tan, Mariya I Vasileva, Kate Saenko, and Bryan A Plummer. Learning similarity conditions without explicit supervision. In ICCV, pages 10373-10382, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16173" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.133 + ], + "angle": 0, + "content": "[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.485, + 0.19 + ], + "angle": 0, + "content": "[29] Mariya I Vasileva, Bryan A Plummer, Krishna Dusad, Shreya Rajpal, Ranjitha Kumar, and David Forsyth. Learning type-aware embeddings for fashion compatibility. In ECCV, pages 390-405, 2018. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[30] Andreas Veit, Balazs Kovacs, Sean Bell, Julian McAuley, Kavita Bala, and Serge Belongie. Learning visual clothing style with heterogeneous dyadic co-occurrences. In ICCV, pages 4642–4650, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[31] Shuohang Wang, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. Want to reduce labeling cost? gpt-3 can help, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.484, + 0.359 + ], + "angle": 0, + "content": "[32] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.484, + 0.458 + ], + "angle": 0, + "content": "[33] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16174" + } + ] +] \ No newline at end of file diff --git a/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/bbd1b25f-9206-484c-b93d-8a139709e547_origin.pdf b/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/bbd1b25f-9206-484c-b93d-8a139709e547_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..16168f8332df05d79ea469c0f7448adc786894da --- /dev/null +++ b/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/bbd1b25f-9206-484c-b93d-8a139709e547_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e7d18ceaf890b277ef03cf38b0ba5f9d6a7fb93a583203bf08871fe17539c864 +size 5750507 diff --git a/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/full.md b/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0952e4bffa94ac9c2f70869bca4354bc43f26a30 --- /dev/null +++ b/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/full.md @@ -0,0 +1,322 @@ +# Text2Outfit: Controllable Outfit Generation with Multimodal Language Models + +Yuanhao Zhai $^{1*}$ Yen-Liang Lin $^{2}$ Minxu Peng $^{2}$ Larry S. Davis $^{2}$ Ashwin Chandramouli $^{2}$ Junsong Yuan $^{1}$ David Doermann $^{1}$ $^{1}$ State University of New York at Buffalo $^{2}$ Amazon + +# Abstract + +Existing outfit recommendation frameworks focus on outfit compatibility prediction and complementary item retrieval. We present a text-driven outfit generation framework, Text2Outfit, which generates outfits controlled by text prompts. Our framework supports two forms of outfit recommendation: 1) Text-to-outfit generation, where the prompt includes the specification for each outfit item (e.g., product features), and the model retrieves items that match the prompt and are stylistically compatible. 2) Seed-to-outfit generation, where the prompt includes the specification for a seed item, and the model both predicts which product types the outfit should include (referred to as composition generation) and retrieves the remaining items to build an outfit. We develop a large language model (LLM) framework that learns the cross-modal mapping between text and image set, and predicts a set of embeddings and compositions to retrieve outfit items. We devise an attention masking mechanism in LLM to handle the alignment between text descriptions and image tokens from different categories. We conduct experiments on the Polyvore dataset and evaluate the quality of the generated outfits from two perspectives: 1) feature matching for outfit items, and 2) outfit visual compatibility. The results demonstrate that our approach significantly outperforms the baseline methods in text to outfit generation. + +# 1. Introduction + +Fashion outfit recommendation system constructs outfits from items that go well together (e.g., an outfit that includes a top, a bottom, shoes, bags, etc). It is an important recommendation problem for online retailers, as customers frequently shop for items that can be worn together. Existing outfit recommendation methods mainly focus on compatibility prediction (determining whether a set of fashion items in an outfit go well together) [7, 9, 21, 29, 30] or outfit + +![](images/551a976129217ba657b18b4a9617d29aa5fba0114455200d98cd4ea9c06172cc.jpg) + +![](images/b8a2a6a981607b614c210f361a4ae13fa33bdbd53f4c9747a507fcb40bdd143b.jpg) +Figure 1. A text to outfit generation framework that generates outfits driven by the text prompt. Our framework supports two forms of outfit recommendation. For text-to-outfit generation, the prompt includes outfit-level and item-level specifications. The model retrieves outfit items that match the texts in the prompt and are stylistically compatible. For seed-to-outfit generation, the prompt includes the specification for a seed item (a seed image or text description). The model predicts the composition (outfit categories) and retrieves the items to build outfits. + +complementary item retrieval (completing a partial outfit by finding a compatible item) [20, 25]. The outfit items are predicted by a pre-trained model and are not controlled by the text prompt. For example, generating outfits with desired attributes, e.g., wide-leg sweatpants. + +Recent work ([14, 15]) for text to image retrieval leverages a pre-trained LLM model to learn image embeddings. They perform better retrieval than non-LLM based approaches (e.g., CLIP [24]) especially for longer text descriptions. However, these approaches focus on retrieving single images so they are not designed for retrieving image sets like outfits. + +We present an outfit generation framework (Text2Outfit) that generates outfits driven by a text prompt (cf. Figure 1). Our framework learns the cross-modal mapping (text to image set) and predict the embeddings to retrieve outfits + +that match the product features specified in the text prompt and are visually compatible. We instantiate our framework in two outfit recommendation scenarios using text prompts. In the first scenario (text-to-outfit generation), users specify the text prompt that describes both outfit-level features (e.g., an outfit for spring and causal) and individual items (e.g., a blouse in white with an off-shoulder design). Our model retrieves the complementary outfit items that match the prompt to build outfits. In the second scenario (seed-to-outfit generation), users specify the requirements for a few items (e.g., a seed item), using either images or texts, e.g., a white knitted sweater, and the model both completes the composition and retrieves complementary items to complete the outfit. We consider the case of just one seed item, but the framework can be extended to include multiple seed items. To build complete outfits, we predict the outfit composition (outfit product categories) for the remaining items and retrieve items in the predicted categories to build outfits. + +We use learnable image tokens to generate embeddings for image retrieval similar to ([14, 15]). In contrast to prior work, we use a set of image tokens to retrieve outfits (image set) beyond a single image. Specifically, we incorporate a set of image tokens for different product categories (e.g., top, bottom, shoes). For each image token, we use the MLP layers on the top of the hidden state vector from the LLM to generate the embeddings. We train the framework in an end-to-end manner (fine-tune both LLM model and MLP layers) using LoRA [10] with a loss function that considers both product feature matching and outfit compatibility. + +The prompt includes the texts describing multiple items, which introduces the challenge of aligning the outfit texts and image tokens. It is important to avoid the case - where the image token from one category wrongly attends to the text descriptions from other categories. We present a masking mechanism in the self-attention layers to facilitate the needed attention learning (cf. Section 3.1 and Figure 3). + +For seed-to-outfit generation, since we only have one seed item as input, we must select the composition (outfit categories) for the remaining items. For example, for specific occasions/seasons, some categories should not be included. So, for summer outfits, the coat category should not be included in the final outfit. We train the LLM model to predict outfit compositions (outfit category combinations) and retrieve the items in the predicted categories. + +We evaluate our framework on the public Polyvore Outfit dataset [29] and measure the quality of generated outfits from two perspectives: 1) feature matching for outfit items and 2) outfit compatibility. Since the Polyvore dataset does not have text product attributes, we use Claude-3 [3] to extract the product features for each outfit item (we verify the precision of Claude-3 with human annotations in Section 4.1). The experiment section shows that our approach outperforms the baseline approaches in the product feature matching and + +outfit compatibility. + +To summarize, the main technical contributions of our work are: + +- We propose a text-driven outfit generation framework that generates outfits based on text prompts or images, facilitating outfit recommendation in two forms. +- We develop a large language model framework to predict the embeddings (and compositions) to retrieve the outfit items where they match the input prompt and are visually compatible. +- We devise an attention masking mechanism to handle the alignment between the text descriptions in the prompt and the image tokens from different outfit categories. + +# 2. Related Work + +# 2.1. Outfit Recommendation + +Outfit Compatibility Prediction: Outfit compatibility prediction evaluates how well a set of fashion items are compatible with each other in an outfit. Early works [21, 30] focused on pairwise item-level compatibility by leveraging item co-view or co-purchase information. However, compatibility is determined only by the pairwise items, not the entire outfits. Later works such as [7, 9, 27, 29] addressed the compatibility problem at the outfit level. They use different types of architectures to model the relationship between outfit items (e.g., bidirectional LSTM [9], subspace embeddings [27, 29] and graph convolution networks (GCN) [7]). These methods can be used for the fill-in-the-blank (FITB) problem - given a subset of items in an outfit and a set of candidate items, select the most compatible candidate. [4] extends FITB to build an entire outfit from a seed item - it starts from a seed item and iteratively finds the complementary items to build an outfit. However, the composition (outfit category combination) is pre-determined from the existing outfits. It is unclear how to compute the compositions for the new seed items. + +These approaches are mainly designed for compatibility prediction and FITB. In contrast, our framework generates outfits based on the input text prompt and predicts the outfit composition given a seed item. + +Outfit Complementary Item Retrieval: Complementary item retrieval is to find a compatible item(s) to complete a partial outfit from a large database. Lin et al. [20] designed a category-based attention mechanism that enables scalable indexing and searching for complementary items. Sarka et al. [25] generated the embeddings using a transformer-based architecture that encodes the compatibility for the partial outfit. These approaches address the FITB problem in a scalable manner (support large-scale indexing and complementary item retrieval), which differs from our problem settings. + +Outfit Retrieval: CLIP4Outfit [11] trained transformer encoders to learn cross-modal similarity (the similarity be + +tween text descriptions and outfit images) and used outfit-level embeddings to retrieve the most relevant outfit given the text descriptions. However, this approach considers the whole outfit as a candidate for retrieval. To be able to match a variety of text descriptions, the approach requires constructing a large-scale outfit dataset that covers different types of outfits, which is not feasible in practice. Moreover, their framework does not support outfit generation based on a given seed item. In contrast, our model is an outfit-building approach, retrieving individual items and assembling them into complete outfits. + +# 2.2. Text-to-Image Retrieval + +Most related work focuses on text-to-single image retrieval. CLIP [12] learned joint representations of text and images through large-scale contrastive learning. These representations can be fine-tuned on the domain-specific datasets to achieve higher performance, such as Fashion-CLIP [6] for the fashion domain. Later works (e.g., Flamingo [1], BLIP [16] and BLIP-2 [17]) learned a unified vision-language representation space and translated visual features into LLM to generate captions for images. The unified vision-language representation (e.g., Q-Former [17]) can be used for text-to-image retrieval. Some recent work leverages the LLM model to predict embeddings for retrieval [14, 15], which performs better than CLIP embeddings for longer captions. FROMAGe [15] takes the interleaved image and text as input and produce image embeddings and text as output. GILL [14] extends FROMAGe to support image retrieval and image synthesis. Similar to [14, 15], our framework also uses learnable image tokens for retrieval. In contrast to prior work on single image retrieval, our approach focuses on retrieving multiple items to build the outfits (image set retrieval). + +# 2.3. Text-to-Image/Outfit Synthesis + +Most work focuses on generating a single image given the text prompt [5, 8, 23, 26]. Some recent work [13, 18] generates a full-body human image wearing garments that reflect the provided textual descriptions. However, the synthesized images may not fully preserve the attribute details specified in the prompt as mentioned by [18]. In addition, to obtain shoppable content, these approaches require using a retrieval system to retrieve similar images from the product catalog (the generated images may not be able to find similar catalog images). In contrast, our framework is designed for outfit generation and retrieves outfit items that preserve the attributes specified in the prompt. + +# 3. Method + +Figure 2 illustrates the system overview of our framework. Our framework supports both forms: 1) text-to-outfit generation and 2) seed-to-outfit generation. For text-to-outfit + +generation, the task is to find the outfit items that match the text descriptions described in the input prompt. We use a set of image tokens and learn the image embeddings for different categories. We fine-tune the LLM to learn the embeddings using the outfit retrieval and compatibility losses, which encourage embeddings to retrieve the items that match the product features and are visually compatible. We devise a masking attention approach to handle the alignment between the text descriptions in the prompt and the image tokens. Details are described in Section 3.1. + +For seed-to-outfit generation, the task is to find outfit items to complete an outfit given a seed item and a prompt that describes the global outfit features. We train the LLM model to predict the outfit compositions (outfit category combination) and the embeddings to search the relevant outfit items to generate outfits. Details are described in Section 3.2 + +# 3.1. Text-to-Outfit Generation + +We are given an input text prompt: $T = \{T_{1}, T_{2}, \dots, T_{k}\}$ , where each text description $T_{i}$ describes the outfit-level specification (e.g., outfit for spring and casual) or the item-level specification (e.g., a blouse in white with off-shoulder design). The order of the text descriptions for each category is arbitrary. The goal is to find the outfit items that match the text descriptions in the prompt (cf. Figure 2 (a)). + +To produce embeddings, we add a set of image tokens to LLM: $\mathrm{[IMG_{cat_1},\dots,IMG_{cat_n}]}$ ; each image token corresponds to a high-level category (we use 11 image tokens for Polyvore dataset; the category list is described in Section 4.3). The hidden state vectors of these image tokens from LLM are denoted as $\mathrm{h_{cat_1}},\ldots,\mathrm{h_{cat_n}}$ . We learn a set of MLP layers: $\mathrm{W}_{\mathrm{ret}} = \{\mathrm{W}_{\mathrm{cat}_i}^{\mathrm{ret}},\ldots,\mathrm{W}_{\mathrm{cat}_n}^{\mathrm{ret}}\}$ to transform the hidden state vectors to the embeddings: $\mathbf{e} = \{\mathbf{e}_{\mathrm{cat}_1},\dots,\mathbf{e}_{\mathrm{cat}_n}\}$ , where embedding $\mathbf{e}_{\mathrm{cat}_i}$ is used to retrieve the outfit images for the category $i$ . + +An outfit item is represented by an image I and its text descriptions T. We use a visual encoder $E_{\mathrm{img}}$ to extract the visual embeddings from the image and use a text encoder $E_{\mathrm{txt}}$ to extract text embeddings from the text descriptions. The multi-modal feature of an outfit item is denoted by: $\mathrm{f} = \mathrm{W}_{\mathrm{item}} \cdot (E_{\mathrm{img}}(\mathrm{I}) \| E_{\mathrm{txt}}(\mathrm{T}))$ , where $\|$ is the concatenation and $\mathrm{W}_{\mathrm{item}}$ is a learnable MLP model that is fine-tuned with the outfit retrieval loss. + +We train our framework end-to-end using the following losses: outfit retrieval loss, attention mask loss, and outfit compatibility loss. The outfit retrieval loss retrieves items that match the product features mentioned in the prompt. The attention mask loss handles the alignment between the outfit text descriptions and the image tokens. The outfit compatibility loss improves the visual compatibility of the retrieved outfit items. + +Outfit retrieval loss. A training batch includes a set of posi + +![](images/74a93b5998007ea98ffb514748fe9107e8c3c7078b81c18045d95b64a778c15a.jpg) +(a) Text-to-outfit generation (Sec. 3.1) +(b) Seed-to-outfit generation (Sec 3.2) + +![](images/f482bf3b625cb9b4a66ee812d99b8218998b9b35812c83a7a1860fe6d2e4cbef.jpg) +Figure 2. System overview of the text to outfit generation framework. For text-to-outfit generation (a), the LLM framework predicts a set of embeddings for different outfit categories, which are used to retrieve the outfit items. We devise an attention masking approach to handle the alignment between the text descriptions in the prompt and the image tokens from different categories (details are in Section 3.1). For seed-to-outfit generation (b), the LLM framework takes a seed item as input, predicts the outfit compositions (outfit categories), and retrieves items in the predicted categories to build the outfits (details are in Section 3.2). + +tive outfits. We randomly sample a set of negative outfits for each positive outfit, where each item in the negative outfits has the same high-level category as the positive outfit. Since there are no ground truth negative outfits, random negatives are typically used in outfit training (e.g., [20, 25, 29]). We extract the occasion, season, color and product features of the outfit items using Claude-3 (details are in Section 4.1). To simulate the prompt for training, we randomly select the product features of each outfit item and use Claude-3 to generate the free-form prompt. The goal is to learn the embeddings so that the similarity between the predicted embeddings and the positive outfits is greater than the negative ones. + +Specifically, a batch includes a set of triples: $\Upsilon = \{T, p, N\}$ , where $T$ is the prompt, $p$ is the positive outfit and $N$ is a set of negative outfits. The outfit retrieval loss is defined as: + +$$ +\mathcal {L} _ {\text {r e t}} (p, N; \mathrm {e}) = \sum_ {i = 1} ^ {n} \mathcal {L} _ {\text {c a t} _ {i}} \left(\mathrm {f} _ {\text {c a t} _ {i}} ^ {p}, \mathrm {f} _ {\text {c a t} _ {i}} ^ {N}; \mathrm {e} _ {\text {c a t} _ {i}}\right) \xi (\text {c a t} _ {i}) \tag {1} +$$ + +We combine the loss for each category to obtain the outfit retrieval loss. $\xi$ is the indicator function denoting if the category $i$ is presented in the outfit. The loss for each category is defined as: + +$$ +\mathcal {L} _ {\mathrm {c a t} _ {i}} \left(\mathrm {f} _ {\mathrm {c a t} _ {i}} ^ {p}, \mathrm {f} _ {\mathrm {c a t} _ {i}} ^ {N}; \mathrm {e} _ {\mathrm {c a t} _ {i}}\right) = - \frac {\exp \left(\operatorname {s i m} \left(\mathrm {f} _ {\mathrm {c a t} _ {i}} ^ {p} , \mathrm {e} _ {\mathrm {c a t} _ {i}}\right) / \tau\right)}{\sum_ {n = 1} ^ {N} \exp \left(\operatorname {s i m} \left(\mathrm {f} _ {\mathrm {c a t} _ {i}} ^ {n} , \mathrm {e} _ {\mathrm {c a t} _ {i}}\right) / \tau\right)}, \tag {2} +$$ + +We adopt the InfoNCE loss ([28]). $\mathrm{f}_{\mathrm{cat}_i}^p$ is the feature for positive outfit item and $\mathrm{f}_{\mathrm{cat}_i}^n$ is the feature for negative outfit item. $\mathrm{e}_{\mathrm{cat}_i}$ is the embeddings. $\tau$ is the temperature. We use the cosine function for the sim function. + +Attention mask loss. The prompt includes the text descriptions for multiple categories. We observed the alignment problem where the image token from a category may attend to the text descriptions from other categories. For example, the top image token attends to the shoe text descriptions. An example prompt: "An outfit for spring and formal event. Black high heels. A black dress with a rose print. A red bracelet with gemstones." The text describes the features for fine-grained categories like high-heel, dress, etc, but the image tokens correspond to high-level categories (e.g., tops, bottoms, shoes). These pose the challenge for the LLM model to find the relations between the outfit text descriptions and multiple image tokens, which affects the product feature matching performance (cf. Table 3)). + +To address this issue, we design a masking approach in the self-attention layers of the LLM decoder. Figure 3 illustrates the mask attention for the top image token. We incorporate the mask loss to guide attention learning, where the mask loss is applied to all image tokens. We compute the binary cross-entropy loss (BCE) for each image token between the ground truth mask and the predicted attention values. The mask loss is defined as: + +$$ +\mathcal {L} _ {\text {m a s k}} = \sum_ {i = 1} ^ {n} \sum_ {j = 1} ^ {k} \mathrm {B C E} _ {i} \left(\alpha_ {j}, m _ {j} ^ {\mathrm {G T}}\right), \tag {3} +$$ + +where $n$ is the number of image tokens. $k$ is the number of tokens from the text descriptions. $\alpha_{j}$ is the predicted attention value and $m_j^{\mathrm{GT}}$ is the ground truth mask. + +Compatibility loss (CP. loss). We include a compatibility loss to facilitate the embedding learning that improves the compatibility of the generated outfit items. We trained an + +![](images/548c25207238ed1786a2f06c74d360ef7aa061ce00aee8d4f68a34c2ff0cd21b.jpg) +Figure 3. The figure illustrates the masked attention for the top image token, where the mask facilitates the top image token to attend to the right text descriptions (outfit text and top text sections, highlighted in yellow) in the prompt. Details are in Section 3.1. + +outfit compatibility model (details are in Section 4.2), which predicts a compatibility score given a set of outfit items. In training, we freeze the model weights of the compatibility model (denoted as $\mathcal{C}$ ) and fine-tune the embeddings (e) such that the compatibility score is maximized. The compatibility loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {c p}} (e) = - \mathcal {C} \left(\mathrm {W} _ {\mathrm {c p}} \cdot \mathrm {e}\right) \tag {4} +$$ + +Training and Inference. We train the framework end-to-end using LoRA [2] by minimizing the following loss function: + +$$ +\min _ {\theta , \mathrm {W} _ {\text {r e t}}, \mathrm {W} _ {\text {i t e m}}, \mathrm {W} _ {\mathrm {c p}}} \mathcal {L} \left(\theta , \mathrm {W} _ {\text {r e t}}, \mathrm {W} _ {\text {i t e m}}, \mathrm {W} _ {\mathrm {c p}}\right) \tag {5} +$$ + +$$ +\mathcal {L} = \lambda_ {1} \cdot \mathcal {L} _ {\mathrm {r e t}} + \lambda_ {2} \cdot \mathcal {L} _ {\mathrm {m a s k}} + \lambda_ {3} \cdot \mathcal {L} _ {\mathrm {c p}}, +$$ + +where $\theta$ is LLM model weights. $W_{\mathrm{ret}}$ , $W_{\mathrm{item}}$ , and $W_{\mathrm{cp}}$ are the MLP models for outfit retrieval, outfit item feature extraction, and compatibility, respectively. $\mathcal{L}_{\mathrm{ret}}$ is the outfit retrieval loss, $\mathcal{L}_{\mathrm{mask}}$ is the mask loss and $\mathcal{L}_{\mathrm{cp}}$ is the compatibility loss. $\lambda_1, \lambda_2$ , and $\lambda_3$ are empirically chosen to make the magnitudes of the losses similar in scale. We report the results of using different losses in Table 3. + +In the inference, we feed the text prompt to the LLM model to predict the embeddings and use them to retrieve the outfit items. + +# 3.2. Seed-to-Outfit Generation + +The task is to generate a complete outfit given an input seed item and a text prompt describing the global outfit features. Unlike the text-to-outfit generation task, where users specify the texts for each outfit item, here, users only specify the information for the seed item. Therefore, we need to predict the composition of the remaining items. One can retrieve + +the outfit items from all the categories to build an outfit. However, some categories do not go well with certain types of occasions, seasons, or the seed types. For example, coats should not be included for summer outfits, or an all-body dress should not be selected if the seed is a top. + +Composition loss. To address this problem, we train the LLM model to predict the outfit compositions (a set of outfit categories), where the compositions are learned from outfit datasets (cf. Figure 2 (b)). Specifically, the input includes a prompt and a seed item (users provide an image or a text description for the seed item. For the image, we use a visual encoder to extract the embeddings and feed them to the LLM). To predict the compositions, we add two MLP models (one for predicting the high-level category and another for predicting the fine-grained category) on the top of the hidden state vectors of the high-level category to predict the composition. The composition loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {c o m p}} = \sum_ {i = 1} ^ {n} \mathrm {C E} _ {i} \left(y _ {\mathrm {p r e d}} ^ {\mathrm {h c}}, y _ {\mathrm {g t}} ^ {\mathrm {h c}}\right) + \mathrm {C E} _ {i} \left(y _ {\mathrm {p r e d}} ^ {\mathrm {f g}}, y _ {\mathrm {g t}} ^ {\mathrm {f g}}\right), \tag {6} +$$ + +where the hc denotes the high-level category and fg denotes the fine-grained category. CE is the cross-entropy loss. We train the framework with LoRA end-to-end with the following loss: + +$$ +\mathcal {L} = \lambda_ {4} \cdot \mathcal {L} _ {\text {r e t}} + \lambda_ {5} \cdot \mathcal {L} _ {\mathrm {c p}} + \lambda_ {6} \cdot \mathcal {L} _ {\text {c o m p}} \tag {7} +$$ + +The loss function includes the retrieval, compatibility, and composition prediction losses. We train the model to predict the composition of the seed item belonging to the top category. The performance of other categories for the seed item is in Appendix Table 2. In inference, we search for outfit items in the predicted high-level and fine-grained categories to compose an outfit. + +# 4. Experiments + +# 4.1. Dataset + +We evaluate the model performance on the Polyvore dataset [29]. The dataset includes two different sets: a disjoint set and a non-disjoint set. The non-disjoint set contains 53,306 training outfits and 10,000 testing outfits, while the disjoint set contains 16,995 training outfits and 15,145 testing outfits. We train our framework on the non-disjoint training set as it contains a larger number of outfits and report the performance on both non-disjoint and dis-joint testing sets (results are in the Appendix Section 1). + +The Polyvore dataset does not have annotated product attributes. Motivated by recent works that leverage LLM for annotations (e.g., [19, 31]), we leverage Claude-3 [3] to extract outfit-level attributes (occasion and season) and item-level attributes (color and product features). We verify + +the agreement between the predicted attributes from Claude-3 and humans on a subset of testing samples (500 outfits and 2629 items). The results show the Claude-3 annotated attributes achieve $94.7\%$ overall precision $(97.3\%)$ for color, $95.5\%$ for product feature, $92.2\%$ for occasion, and $93.6\%$ for season), verifying the feasibility of using Claude-3 for attribute annotations. + +# 4.2. Evaluation + +We evaluate the quality of the retrieved outfits in two ways: 1) feature matching for outfit items and 2) outfit compatibility. To evaluate feature matching, we compute the number of items in the top-10 retrieval results that match the product features specified in the prompt. We evaluate the feature matching for color, occasion, season, and product feature attributes for the outfit items, which are critical attributes for outfits. To handle non-exact matches, where the product features are different but their meanings are similar, we use ModernBERT embedding [22] to evaluate the affinity of two features (we select the thresholds of different attributes using a validation set). We report the results using precision@top-10 (denoted as P@10). + +To evaluate outfit compatibility, we train an outfit compatibility model using a SOTA architecture (OutfitTransformer [25]), where the model takes a set of outfit items as input, and predicts a compatibility score. We replace the ResNet features in the original architecture with the FashionCLIP embeddings [6] and use the item image and the text description as input. We achieve better outfit compatibility prediction than the original architecture. We achieve $94.93\%$ (93% for OutfitTransformer) and $93.88\%$ AUC (88% for OutfitTransformer) for the non-disjoint and disjoint sets respectively. We report the compatibility (CP) score for the top 10 retrieved outfits. + +There is no ground-truth outfit composition given a seed and prompt. To obtain a quantitative quality score, we use Claude-3 (we also evaluate the performance with Pixtral Large from Mistral AI, see Appendix Section 2) to evaluate the quality of the predicted compositions. Recent work (e.g., [33]) shows that strong LLMs achieve high agreement with humans. We evaluate the quality in two ways: 1) We use Claude-3 to predict a quality score for each predicted composition and report an average quality score (denoted as Comp. score). 2) Given two compositions generated by the baseline and our approach, we ask Claude-3 to pick the better composition and report the winning rate of our approach (denoted as Comp. win rate). + +# 4.3. Baselines and Model Settings + +There is no existing framework for text-to-outfit retrieval tasks ([11] is designed to retrieve entire outfits, which is different from our settings as mentioned in Section 2). We convert a set of baseline approaches to the outfit retrieval + +tasks and report the performance. In Table 1, we categorize these baselines into the text-to-single-item retrieval and text-to-outfit retrieval approaches. Existing methods for text-to-image retrieval [6, 14, 24] mainly focus on single-image retrieval. To make these methods applicable to outfit retrieval problem, we make the following changes: 1) We convert the outfit-level specification (occasion/season) into item-level specification. 2) We use the text descriptions of each category to retrieve the outfit items (assuming that the text descriptions to category correspondence are given). + +For text-to-outfit retrieval, we converted the GILL [14] framework to support multiple-category retrieval. For text-to-outfit retrieval methods, the text descriptions to category correspondence information are not given, which is directly predicted by the models (except the fixed mask approach in Table 1, which uses the text-to-category correspondence to generate the fixed mask for attention). The details of the baseline approaches are described below. + +- CLIP [24]: We convert the outfit-level specification into item-level specifications and perform retrieval for each category. We use ViT-L-14 version. +- Fashion-CLIP 2.0 [6]: We use the same settings as CLIP and use Fashion-CLIP to retrieve items for each category. We use version 2.0, which achieves better performance than version 1.0. +- BLIP-2 [17]: We use the pre-trained image encoder and Q-Former from BLIP-2. Like CLIP, we perform image retrieval for each category using the corresponding text description. We use the ViT-g version. +- GILL [14] + Multiple categories: We add the learnable image tokens for each high-level category for multiple image retrieval. We fine-tune the MLP layers to learn embeddings while keeping the LLM model fixed following the original training approach. We use the same prompt format as our model. Note that we compare our method with an improved version of GiLL, in which we replace the original CLIP vision encoder with Fashion-CLIP + +Model Settings: For our framework, we use the pretrained OPT 6.7B model [32] (same as GILL) as the LLM backbone. We fine-tune the LLM with LoRA (we set the rank to 64 for Q, K, and V projection matrices). We use Fashion-CLIP 2.0 [6] as the visual encoder $E_{\mathrm{img}}$ and the text encoder $E_{\mathrm{text}}$ . Fashion-CLIP 2.0 achieves better performance than CLIP embeddings. In model training, the visual and text encoder weights are fixed. We use the bi-directional masks for image tokens. We use 11 image tokens for the Polyvore dataset, corresponding to the following high-level categories: tops, bottoms, all-body, outerwear, shoes, bags, jewelry, accessories, sunglasses, hats, and scarves. + +# 4.4. Text-to-Outfit Performance Evaluation + +Table 1 summarizes the text-to-outfit retrieval results. We use the testing set of the Polyvore dataset to generate the + +
MethodModalityColor (P@10)Product feature (P@10)Season (P@10)Occasion (P@10)Avg (P@10)CP. score
Text-to-single-item retrieval
CLIP [24]Text only96.0670.9190.2276.6283.4663.49
Fashion-CLIP [6]Text only96.8280.4289.6881.2787.0564.07
BLIP-2 [17]Text only93.2966.9485.4684.5182.5565.09
CLIP [24]Text to image83.5961.3285.4967.0174.3549.77
Fashion-CLIP [6]Text to image87.5972.6785.9168.8878.7653.03
BLIP-2 [17]Text to image86.3966.6985.6267.7576.6151.05
Text-to-outfit retrieval
GILL [14] + multiple categoryText to image48.1948.1586.8769.6963.2374.26
Text2Outfit + learnable mask (ours)Text to image78.4276.5486.6070.4578.0066.28
Text2Outfit + learnable mask (ours)Text to image/text86.0584.8195.6993.6690.0584.05
Text2Outfit + fixed mask (ours)Text to image/text96.6186.2794.8792.4892.5682.51
+ +Table 1. Text-to-outfit retrieval performance comparison. P@10 denotes precision at top-10 and CP. score denotes compatibility score using [25] for top-10 retrieved outfits. Note that for single-item retrieval approaches, the text descriptions and its corresponding category are given while the outfit retrieval approaches are not given such information. The bold text denotes the best performance and the underlines denote the second best performance. Details are in Section 4.2. + +
MethodCompositionSeason (P@10)Occasion (P@10)Comp. score (c3)Comp win rate (c3)CP. scoreCP. score (c3)
Seed text + prompt (baseline)All categories97.3296.202.596%80.822.78
Seed text + prompt (ours)Predicted97.5696.014.5794%86.183.85
Seed image + prompt (baseline)All categories95.0087.502.858%72.882.66
Seed image + prompt (ours)Predicted96.4391.804.3292%83.363.83
+ +Table 2. Seed-to-outfit retrieval performance comparison. The scores with (c3) are evaluated by Claude-3. Details are in Section 4.5. + +prompt. Given a test outfit including multiple outfit items, we randomly select outfit attributes (occasion and season) and item attributes (color and product features) and use Claude-3 to generate the free-form prompt. We use the generated prompt to evaluate the outfit retrieval performance. + +Our approaches achieve $90.05\%$ and $92.56\%$ average $\mathrm{P@10}$ using the learnable mask and fixed mask respectively, showing better feature-matching precision than the baseline approaches. The GILL + multi-category approach has an alignment issue (the alignment between the outfit texts and the images tokens) while our approach uses the learnable mask to alleviate the alignment issue. We fine-tune the model end-to-end with LoRA instead of using the frozen LLM model as GILL, improving the performance (see Table 3 for detailed analysis). Our approach with a learnable mask also performs better than the single-item retrieval approaches, where they are given the text description to category correspondence information, while our approach does not. + +In addition, our approach shows better outfit compatibility than the baseline approaches. We achieve the compatibility score 84.05 using the learnable mask setting (we compute the average compatibility scores for the ground truth positive (86.95) and negative (22.99) outfits on the Polyvore testing set using OutfitTransformer. The generated outfits from our method achieve similar compatibility as the positive outfits). Single-item retrieval approaches perform worse on outfit compatibility because they find items that match the text descriptions but do not model the item relationships in an outfit. Our framework is fine-tuned on the outfit data, and incorporates the outfit retrieval and compatibility losses, + +which improve overall compatibility. Figure 4 shows example results of our framework. + +# 4.5. Seed-to-Outfit Performance Evaluation + +Table 2 summarizes the results for seed-to-outfit retrieval. We used the Polyvore test set to construct the input (seed and prompt) for evaluation. Given a test outfit including multiple outfit items, we randomly select the outfit attributes (occasion and season) and use Claude-3 to generate the free-form prompt. + +We compare our approach with a baseline approach that uses all 11 high-level categories for the composition. We feed the compositions (e.g., category 1/fine-grained category 1, category 2/fine-grained category 2, ...) of the baseline and our approach to Claude-3 for evaluation (for the baseline approach, the fine-grained category is obtained from the top-1 retrieved outfit). Our approach shows better composition scores (the score is between 1 (poor) to 5 (excellent)) and winning rates than the baseline (e.g., $94\%$ and $92\%$ winning rate of using seed text and seed image, respectively). We verify the agreement between Claude-3 and humans on a subset of samples (500 outfitsits). The Spearman's rank correlation coefficient is 0.73, which indicates they are highly correlated. We report the visual compatibility using both OutfitTransformer [25] (CP. score) and Claude-3 (CP. score (c3)). Our approach shows better visual compatibility than the baseline approach. + +# 4.6. Ablation Study + +Multimodal feature: We compare the performance of single-modality (image) vs. multimodality (image + text) + +
ModalityLoRAMask LossCP. lossColor (P@10)Product feature (P@10)Season (P@10)Occasion (P@10)Avg (P@10)CP. score
1Image---48.3153.0385.1068.6263.7653.43
2Image + text---44.6047.1888.9481.0765.4578.62
3Image + text✓--74.7475.0195.6093.9084.8286.95
4Image + text✓Global CE-79.2882.2396.0294.4387.9985.63
5Image + text✓Local BCE-85.5884.8795.8993.4789.9583.43
6Image + text✓Local BCE✓86.0584.8195.6993.6690.0584.05
+ +Table 3. Ablation study of different components in our framework. Details are described in Section 4.6. + +# (a) Text-to-outfit generation + +Prompt: Spring outfit for casual. A canvas upper flats in a pink shade. A jeans in black with high-waisted accents. A sleeveless top in a white shade. A denim jacket in a blue shade. + +![](images/d4efe064e518d69a18e7ff1e1e3a626f988c797752138e13c95524d7bfbe7e1e.jpg) +GILL + Multiple Category + +![](images/9be4457d91776668bdd16f14a1962ac1ec65ed9777cf919efe78084094afba9e.jpg) + +![](images/18efad0ff3ea480b301f1711bda6c09bac818333885e997ff6ca54342dda0da2.jpg) + +![](images/ce5b83a67e98e063fbfe395c9395b168cdd7586bd7974ca14a54dfedb23e1344.jpg) + +![](images/a1317f57e31be65ee5a819c8d98d7c40e2bead98444790a87697636326383654.jpg) +Text2Outfit + learnable mask (ours) + +![](images/030bdd17657e64778cf578361ddbb6a322b541c65e28e1fce63eb8a89d9d08fd.jpg) + +![](images/6ba38cb2439c77f1e4cb73a6098cc169320807725a6cefd3cfc5a9a5720569a7.jpg) + +![](images/71edc74bafd512ef522181b76693381aaf120b531651abe27973106c59e5f6a9.jpg) + +![](images/1b816c838c4695ed9cd965bfe7e7a0976de041081e3a02d1eb6872a094d8b300.jpg) +Text2Outfit + fixed mask (ours) + +![](images/d168ea803e55d552c8bf367a9ea55d9770ad0f256f9cfe9b0f420f2d34ef8b4a.jpg) +Figure 4. Qualitative results. We compare the results for text-to-outfit generation with the extended GILL [14] approach that supports multiple image retrieval and our framework (learnable mask and fixed mask). The non-matched items are marked with boxes. We show the predicted compositions and retrieval results for seed-to-outfit generation with a seed image or text. More visual results are in the Appendix Section 4. + +![](images/5f4c32543f29a9a9b14363f4dc103a17f4c8f0bc3f5e0ff9318190bd215d2144.jpg) + +![](images/b6177f2df1e6f18c2d3d414e179c2abef279d14d3e9cd1bc187de643f8c8c17c.jpg) + +# (b) Seed-to-outfit generation + +Prompt: Outfit for casual in spring. + +![](images/29714022b8032df6c9d2e00402574ad4af879ccb5a9fb37ebe02796f802a0f77.jpg) + +Prompt: Formal events outfit for spring weather. + +![](images/78d6996f0a84d87d05b1fb3de30f9b519db6b6178dedb92d37c151e3ba37b3d1.jpg) + +(row-1 and row-2 in Table 3). Using multimodality improves the average P@10 from $63.76\%$ to $65.45\%$ for feature matching. + +LoRA: We compare the performance of fine-tuned LLM (with LoRA) vs. frozen LLM (row-2 and row-3 in Table 3). Using LoRA significantly improves the average P@10 from $65.45\%$ to $84.82\%$ and improves the compatibility score from $78.62\%$ to $86.95\%$ . + +Mask loss: We investigate different loss functions for the attention masking: Global CE and Local BCE (row-4 and row-5 in Table 3). For Global CE, we consider the attention values from all the text tokens to each image token as a distribution and optimize a global cross-entropy loss. We optimize each attention value for Local BCE with a binary cross-entropy loss. Local BCE performs better, as it optimizes the attention directly with the ground truth mask, + +while global CE optimizes the overall distribution of the attention. + +Compatibility loss: We analyze the impact of including a compatibility loss in the training loss (cf. Equation 5). It further improves compatibility (row-6 in Table 3). + +# 5. Conclusions + +We present a framework for text-based outfit generation. We explore two different forms of outfit generation: 1) text-to-outfit generation and 2) seed-to-outfit generation tasks, which enable the controllable outfit generation. In addition, we introduce a masking mechanism which better handles the alignment between outfit texts and the image tokens, and achieves comparable performance when using the ground-truth masks. The experimental results demonstrate that our model outperforms the baseline approaches for text to outfit retrieval tasks. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 3 +[2] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, et al. Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2, pages 929-947, 2024. 5 +[3] Anthropic. Introducing the next generation of claude. https://www.anthropic.com/news/claude-3-family, 2024.2,5 +[4] Federico Becattini, Federico Maria Teotini, and Alberto Del Bimbo. Transformer-based graph neural networks for outfit generation, 2023. 2 +[5] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3 +[6] Patrick John Chia, Giuseppe Attanasio, Federico Bianchi, Silvia Terragni, Ana Rita Magalhães, Diogo Goncalves, Ciro Greco, and Jacopo Tagliabue. Contrastive language and vision learning of general fashion concepts. Scientific Reports, 12 (1):18958, 2022. 3, 6, 7 +[7] Guillem Cucurull, Perouz Taslakian, and David Vazquez. Context-aware visual compatibility prediction. In CVPR, pages 12617-12626, 2019. 1, 2 +[8] Runpei Dong, Chunrui Han, Yang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. DreamLLM: Synergistic multimodal comprehension and creation. In ICLR, 2024. 3 +[9] Xintong Han, Zuxuan Wu, Yu-Gang Jiang, and Larry S Davis. Learning fashion compatibility with bidirectional lstms. In ACM MM, pages 1078-1086, 2017. 1, 2 +[10] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 2 +[11] Junkyu Jang, Eugene Hwang, and Sung-Hyuk Park. Lost your style? navigating with semantic-level approach for text-to-outfit retrieval. In WACV, pages 8066-8075, 2024. 2, 6 +[12] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 3 +[13] Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven + +controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3 +[14] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. NeurIPS, 36, 2023. 1, 2, 3, 6, 7, 8 +[15] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In ICML, pages 17283-17300. PMLR, 2023. 1, 2, 3 +[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3 +[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 3, 6, 7 +[18] Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, Kwan-Yee Lin, and Wayne Wu. Cosmicman: A text-to-image foundation model for humans. In CVPR, pages 6955-6965, 2024. 3 +[19] Wei Li, Hehe Fan, Yongkang Wong, Yi Yang, and Mohan Kankanhalli. Improving context understanding in multimodal large language models via multimodal composition learning. In *Forty-first International Conference on Machine Learning*, 2024. 5 +[20] Yen-Liang Lin, Son Tran, and Larry S Davis. Fashion outfit complementary item retrieval. In CVPR, pages 3311-3319, 2020. 1, 2, 4 +[21] Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, pages 43-52, 2015. 1, 2 +[22] Zach Nussbaum, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. Nomic embed: Training a reproducible long context text embedder, 2024. 6 +[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 3 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 6, 7 +[25] Rohan Sarkar, Navaneeth Bodla, Mariya I Vasileva, Yen-Liang Lin, Anurag Beniwal, Alan Lu, and Gerard Medioni. Outfittransformer: Learning outfit representations for fashion recommendation. In WACV, pages 3601-3609, 2023. 1, 2, 4, 6, 7 +[26] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 3 +[27] Reuben Tan, Mariya I Vasileva, Kate Saenko, and Bryan A Plummer. Learning similarity conditions without explicit supervision. In ICCV, pages 10373-10382, 2019. 2 + +[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 4 +[29] Mariya I Vasileva, Bryan A Plummer, Krishna Dusad, Shreya Rajpal, Ranjitha Kumar, and David Forsyth. Learning type-aware embeddings for fashion compatibility. In ECCV, pages 390-405, 2018. 1, 2, 4, 5 +[30] Andreas Veit, Balazs Kovacs, Sean Bell, Julian McAuley, Kavita Bala, and Serge Belongie. Learning visual clothing style with heterogeneous dyadic co-occurrences. In ICCV, pages 4642–4650, 2015. 1, 2 +[31] Shuohang Wang, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. Want to reduce labeling cost? gpt-3 can help, 2021. 5 +[32] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 6 +[33] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6 \ No newline at end of file diff --git a/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/images.zip b/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..acbea4d0ffd7ae6a3be2fce7683fcb2e9a76c250 --- /dev/null +++ b/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3e6c32460658f8d95c13585bebf021a5590f822955966ef820fd7568ee02683 +size 409970 diff --git a/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/layout.json b/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5ef9343b3f3adcf368a1a156bb4d15a8e1d9941d --- /dev/null +++ b/2025/Text2Outfit_ Controllable Outfit Generation with Multimodal Language Models/layout.json @@ -0,0 +1,8008 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "text", + "content": "Text2Outfit: Controllable Outfit Generation with Multimodal Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "spans": [ + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "text", + "content": "Yuanhao Zhai" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "text", + "content": " Yen-Liang Lin" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "text", + "content": " Minxu Peng" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "text", + "content": " Larry S. Davis" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "text", + "content": " Ashwin Chandramouli" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "text", + "content": " Junsong Yuan" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "text", + "content": " David Doermann" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "text", + "content": "State University of New York at Buffalo " + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 112, + 141, + 498, + 199 + ], + "type": "text", + "content": "Amazon" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 224, + 200, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 224, + 200, + 238 + ], + "spans": [ + { + "bbox": [ + 151, + 224, + 200, + 238 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 251, + 297, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 539 + ], + "type": "text", + "content": "Existing outfit recommendation frameworks focus on outfit compatibility prediction and complementary item retrieval. We present a text-driven outfit generation framework, Text2Outfit, which generates outfits controlled by text prompts. Our framework supports two forms of outfit recommendation: 1) Text-to-outfit generation, where the prompt includes the specification for each outfit item (e.g., product features), and the model retrieves items that match the prompt and are stylistically compatible. 2) Seed-to-outfit generation, where the prompt includes the specification for a seed item, and the model both predicts which product types the outfit should include (referred to as composition generation) and retrieves the remaining items to build an outfit. We develop a large language model (LLM) framework that learns the cross-modal mapping between text and image set, and predicts a set of embeddings and compositions to retrieve outfit items. We devise an attention masking mechanism in LLM to handle the alignment between text descriptions and image tokens from different categories. We conduct experiments on the Polyvore dataset and evaluate the quality of the generated outfits from two perspectives: 1) feature matching for outfit items, and 2) outfit visual compatibility. The results demonstrate that our approach significantly outperforms the baseline methods in text to outfit generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "content": "Fashion outfit recommendation system constructs outfits from items that go well together (e.g., an outfit that includes a top, a bottom, shoes, bags, etc). It is an important recommendation problem for online retailers, as customers frequently shop for items that can be worn together. Existing outfit recommendation methods mainly focus on compatibility prediction (determining whether a set of fashion items in an outfit go well together) [7, 9, 21, 29, 30] or outfit" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 227, + 551, + 296 + ], + "blocks": [ + { + "bbox": [ + 320, + 227, + 551, + 296 + ], + "lines": [ + { + "bbox": [ + 320, + 227, + 551, + 296 + ], + "spans": [ + { + "bbox": [ + 320, + 227, + 551, + 296 + ], + "type": "image", + "image_path": "551a976129217ba657b18b4a9617d29aa5fba0114455200d98cd4ea9c06172cc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 301, + 550, + 380 + ], + "blocks": [ + { + "bbox": [ + 320, + 301, + 550, + 380 + ], + "lines": [ + { + "bbox": [ + 320, + 301, + 550, + 380 + ], + "spans": [ + { + "bbox": [ + 320, + 301, + 550, + 380 + ], + "type": "image", + "image_path": "b8a2a6a981607b614c210f361a4ae13fa33bdbd53f4c9747a507fcb40bdd143b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 392, + 555, + 491 + ], + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 491 + ], + "type": "text", + "content": "Figure 1. A text to outfit generation framework that generates outfits driven by the text prompt. Our framework supports two forms of outfit recommendation. For text-to-outfit generation, the prompt includes outfit-level and item-level specifications. The model retrieves outfit items that match the texts in the prompt and are stylistically compatible. For seed-to-outfit generation, the prompt includes the specification for a seed item (a seed image or text description). The model predicts the composition (outfit categories) and retrieves the items to build outfits." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 517, + 555, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 578 + ], + "type": "text", + "content": "complementary item retrieval (completing a partial outfit by finding a compatible item) [20, 25]. The outfit items are predicted by a pre-trained model and are not controlled by the text prompt. For example, generating outfits with desired attributes, e.g., wide-leg sweatpants." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 579, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 556, + 662 + ], + "type": "text", + "content": "Recent work ([14, 15]) for text to image retrieval leverages a pre-trained LLM model to learn image embeddings. They perform better retrieval than non-LLM based approaches (e.g., CLIP [24]) especially for longer text descriptions. However, these approaches focus on retrieving single images so they are not designed for retrieving image sets like outfits." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "We present an outfit generation framework (Text2Outfit) that generates outfits driven by a text prompt (cf. Figure 1). Our framework learns the cross-modal mapping (text to image set) and predict the embeddings to retrieve outfits" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 242, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 242, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 242, + 713 + ], + "type": "text", + "content": "*This work was done during an internship at Amazon." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16165" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 288 + ], + "type": "text", + "content": "that match the product features specified in the text prompt and are visually compatible. We instantiate our framework in two outfit recommendation scenarios using text prompts. In the first scenario (text-to-outfit generation), users specify the text prompt that describes both outfit-level features (e.g., an outfit for spring and causal) and individual items (e.g., a blouse in white with an off-shoulder design). Our model retrieves the complementary outfit items that match the prompt to build outfits. In the second scenario (seed-to-outfit generation), users specify the requirements for a few items (e.g., a seed item), using either images or texts, e.g., a white knitted sweater, and the model both completes the composition and retrieves complementary items to complete the outfit. We consider the case of just one seed item, but the framework can be extended to include multiple seed items. To build complete outfits, we predict the outfit composition (outfit product categories) for the remaining items and retrieve items in the predicted categories to build outfits." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 421 + ], + "type": "text", + "content": "We use learnable image tokens to generate embeddings for image retrieval similar to ([14, 15]). In contrast to prior work, we use a set of image tokens to retrieve outfits (image set) beyond a single image. Specifically, we incorporate a set of image tokens for different product categories (e.g., top, bottom, shoes). For each image token, we use the MLP layers on the top of the hidden state vector from the LLM to generate the embeddings. We train the framework in an end-to-end manner (fine-tune both LLM model and MLP layers) using LoRA [10] with a loss function that considers both product feature matching and outfit compatibility." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 423, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 506 + ], + "type": "text", + "content": "The prompt includes the texts describing multiple items, which introduces the challenge of aligning the outfit texts and image tokens. It is important to avoid the case - where the image token from one category wrongly attends to the text descriptions from other categories. We present a masking mechanism in the self-attention layers to facilitate the needed attention learning (cf. Section 3.1 and Figure 3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 509, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 604 + ], + "type": "text", + "content": "For seed-to-outfit generation, since we only have one seed item as input, we must select the composition (outfit categories) for the remaining items. For example, for specific occasions/seasons, some categories should not be included. So, for summer outfits, the coat category should not be included in the final outfit. We train the LLM model to predict outfit compositions (outfit category combinations) and retrieve the items in the predicted categories." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 297, + 715 + ], + "type": "text", + "content": "We evaluate our framework on the public Polyvore Outfit dataset [29] and measure the quality of generated outfits from two perspectives: 1) feature matching for outfit items and 2) outfit compatibility. Since the Polyvore dataset does not have text product attributes, we use Claude-3 [3] to extract the product features for each outfit item (we verify the precision of Claude-3 with human annotations in Section 4.1). The experiment section shows that our approach outperforms the baseline approaches in the product feature matching and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 73, + 397, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 397, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 397, + 84 + ], + "type": "text", + "content": "outfit compatibility." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 85, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 108 + ], + "type": "text", + "content": "To summarize, the main technical contributions of our work are:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 110, + 555, + 228 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 110, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 110, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 110, + 555, + 144 + ], + "type": "text", + "content": "- We propose a text-driven outfit generation framework that generates outfits based on text prompts or images, facilitating outfit recommendation in two forms." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 145, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 555, + 192 + ], + "type": "text", + "content": "- We develop a large language model framework to predict the embeddings (and compositions) to retrieve the outfit items where they match the input prompt and are visually compatible." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 194, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 228 + ], + "type": "text", + "content": "- We devise an attention masking mechanism to handle the alignment between the text descriptions in the prompt and the image tokens from different outfit categories." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 242, + 400, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 242, + 400, + 255 + ], + "spans": [ + { + "bbox": [ + 314, + 242, + 400, + 255 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 262, + 450, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 262, + 450, + 274 + ], + "spans": [ + { + "bbox": [ + 314, + 262, + 450, + 274 + ], + "type": "text", + "content": "2.1. Outfit Recommendation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 281, + 556, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 556, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 556, + 519 + ], + "type": "text", + "content": "Outfit Compatibility Prediction: Outfit compatibility prediction evaluates how well a set of fashion items are compatible with each other in an outfit. Early works [21, 30] focused on pairwise item-level compatibility by leveraging item co-view or co-purchase information. However, compatibility is determined only by the pairwise items, not the entire outfits. Later works such as [7, 9, 27, 29] addressed the compatibility problem at the outfit level. They use different types of architectures to model the relationship between outfit items (e.g., bidirectional LSTM [9], subspace embeddings [27, 29] and graph convolution networks (GCN) [7]). These methods can be used for the fill-in-the-blank (FITB) problem - given a subset of items in an outfit and a set of candidate items, select the most compatible candidate. [4] extends FITB to build an entire outfit from a seed item - it starts from a seed item and iteratively finds the complementary items to build an outfit. However, the composition (outfit category combination) is pre-determined from the existing outfits. It is unclear how to compute the compositions for the new seed items." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 568 + ], + "type": "text", + "content": "These approaches are mainly designed for compatibility prediction and FITB. In contrast, our framework generates outfits based on the input text prompt and predicts the outfit composition given a seed item." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 689 + ], + "type": "text", + "content": "Outfit Complementary Item Retrieval: Complementary item retrieval is to find a compatible item(s) to complete a partial outfit from a large database. Lin et al. [20] designed a category-based attention mechanism that enables scalable indexing and searching for complementary items. Sarka et al. [25] generated the embeddings using a transformer-based architecture that encodes the compatibility for the partial outfit. These approaches address the FITB problem in a scalable manner (support large-scale indexing and complementary item retrieval), which differs from our problem settings." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "content": "Outfit Retrieval: CLIP4Outfit [11] trained transformer encoders to learn cross-modal similarity (the similarity be" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16166" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 204 + ], + "type": "text", + "content": "tween text descriptions and outfit images) and used outfit-level embeddings to retrieve the most relevant outfit given the text descriptions. However, this approach considers the whole outfit as a candidate for retrieval. To be able to match a variety of text descriptions, the approach requires constructing a large-scale outfit dataset that covers different types of outfits, which is not feasible in practice. Moreover, their framework does not support outfit generation based on a given seed item. In contrast, our model is an outfit-building approach, retrieving individual items and assembling them into complete outfits." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 211, + 191, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 191, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 191, + 224 + ], + "type": "text", + "content": "2.2. Text-to-Image Retrieval" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 228, + 297, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 297, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 297, + 479 + ], + "type": "text", + "content": "Most related work focuses on text-to-single image retrieval. CLIP [12] learned joint representations of text and images through large-scale contrastive learning. These representations can be fine-tuned on the domain-specific datasets to achieve higher performance, such as Fashion-CLIP [6] for the fashion domain. Later works (e.g., Flamingo [1], BLIP [16] and BLIP-2 [17]) learned a unified vision-language representation space and translated visual features into LLM to generate captions for images. The unified vision-language representation (e.g., Q-Former [17]) can be used for text-to-image retrieval. Some recent work leverages the LLM model to predict embeddings for retrieval [14, 15], which performs better than CLIP embeddings for longer captions. FROMAGe [15] takes the interleaved image and text as input and produce image embeddings and text as output. GILL [14] extends FROMAGe to support image retrieval and image synthesis. Similar to [14, 15], our framework also uses learnable image tokens for retrieval. In contrast to prior work on single image retrieval, our approach focuses on retrieving multiple items to build the outfits (image set retrieval)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 222, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 222, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 222, + 499 + ], + "type": "text", + "content": "2.3. Text-to-Image/Outfit Synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 504, + 297, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 297, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 297, + 647 + ], + "type": "text", + "content": "Most work focuses on generating a single image given the text prompt [5, 8, 23, 26]. Some recent work [13, 18] generates a full-body human image wearing garments that reflect the provided textual descriptions. However, the synthesized images may not fully preserve the attribute details specified in the prompt as mentioned by [18]. In addition, to obtain shoppable content, these approaches require using a retrieval system to retrieve similar images from the product catalog (the generated images may not be able to find similar catalog images). In contrast, our framework is designed for outfit generation and retrieves outfit items that preserve the attributes specified in the prompt." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 657, + 111, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 111, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 111, + 670 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Figure 2 illustrates the system overview of our framework. Our framework supports both forms: 1) text-to-outfit generation and 2) seed-to-outfit generation. For text-to-outfit" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 191 + ], + "type": "text", + "content": "generation, the task is to find the outfit items that match the text descriptions described in the input prompt. We use a set of image tokens and learn the image embeddings for different categories. We fine-tune the LLM to learn the embeddings using the outfit retrieval and compatibility losses, which encourage embeddings to retrieve the items that match the product features and are visually compatible. We devise a masking attention approach to handle the alignment between the text descriptions in the prompt and the image tokens. Details are described in Section 3.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 193, + 556, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 556, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 556, + 276 + ], + "type": "text", + "content": "For seed-to-outfit generation, the task is to find outfit items to complete an outfit given a seed item and a prompt that describes the global outfit features. We train the LLM model to predict the outfit compositions (outfit category combination) and the embeddings to search the relevant outfit items to generate outfits. Details are described in Section 3.2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 285, + 458, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 458, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 458, + 298 + ], + "type": "text", + "content": "3.1. Text-to-Outfit Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 304, + 556, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 556, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 556, + 388 + ], + "type": "text", + "content": "We are given an input text prompt: " + }, + { + "bbox": [ + 313, + 304, + 556, + 388 + ], + "type": "inline_equation", + "content": "T = \\{T_{1}, T_{2}, \\dots, T_{k}\\}" + }, + { + "bbox": [ + 313, + 304, + 556, + 388 + ], + "type": "text", + "content": ", where each text description " + }, + { + "bbox": [ + 313, + 304, + 556, + 388 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 313, + 304, + 556, + 388 + ], + "type": "text", + "content": " describes the outfit-level specification (e.g., outfit for spring and casual) or the item-level specification (e.g., a blouse in white with off-shoulder design). The order of the text descriptions for each category is arbitrary. The goal is to find the outfit items that match the text descriptions in the prompt (cf. Figure 2 (a))." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "text", + "content": "To produce embeddings, we add a set of image tokens to LLM: " + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "inline_equation", + "content": "\\mathrm{[IMG_{cat_1},\\dots,IMG_{cat_n}]}" + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "text", + "content": "; each image token corresponds to a high-level category (we use 11 image tokens for Polyvore dataset; the category list is described in Section 4.3). The hidden state vectors of these image tokens from LLM are denoted as " + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "inline_equation", + "content": "\\mathrm{h_{cat_1}},\\ldots,\\mathrm{h_{cat_n}}" + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "text", + "content": ". We learn a set of MLP layers: " + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "inline_equation", + "content": "\\mathrm{W}_{\\mathrm{ret}} = \\{\\mathrm{W}_{\\mathrm{cat}_i}^{\\mathrm{ret}},\\ldots,\\mathrm{W}_{\\mathrm{cat}_n}^{\\mathrm{ret}}\\}" + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "text", + "content": " to transform the hidden state vectors to the embeddings: " + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{e} = \\{\\mathbf{e}_{\\mathrm{cat}_1},\\dots,\\mathbf{e}_{\\mathrm{cat}_n}\\}" + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "text", + "content": ", where embedding " + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{\\mathrm{cat}_i}" + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "text", + "content": " is used to retrieve the outfit images for the category " + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 388, + 556, + 508 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "type": "text", + "content": "An outfit item is represented by an image I and its text descriptions T. We use a visual encoder " + }, + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "type": "text", + "content": " to extract the visual embeddings from the image and use a text encoder " + }, + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{txt}}" + }, + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "type": "text", + "content": " to extract text embeddings from the text descriptions. The multi-modal feature of an outfit item is denoted by: " + }, + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{f} = \\mathrm{W}_{\\mathrm{item}} \\cdot (E_{\\mathrm{img}}(\\mathrm{I}) \\| E_{\\mathrm{txt}}(\\mathrm{T}))" + }, + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\|" + }, + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "type": "text", + "content": " is the concatenation and " + }, + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{W}_{\\mathrm{item}}" + }, + { + "bbox": [ + 313, + 509, + 556, + 604 + ], + "type": "text", + "content": " is a learnable MLP model that is fine-tuned with the outfit retrieval loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 700 + ], + "type": "text", + "content": "We train our framework end-to-end using the following losses: outfit retrieval loss, attention mask loss, and outfit compatibility loss. The outfit retrieval loss retrieves items that match the product features mentioned in the prompt. The attention mask loss handles the alignment between the outfit text descriptions and the image tokens. The outfit compatibility loss improves the visual compatibility of the retrieved outfit items." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 701, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 556, + 714 + ], + "type": "text", + "content": "Outfit retrieval loss. A training batch includes a set of posi" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16167" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 91, + 286, + 236 + ], + "blocks": [ + { + "bbox": [ + 64, + 73, + 194, + 83 + ], + "lines": [ + { + "bbox": [ + 64, + 73, + 194, + 83 + ], + "spans": [ + { + "bbox": [ + 64, + 73, + 194, + 83 + ], + "type": "text", + "content": "(a) Text-to-outfit generation (Sec. 3.1)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 91, + 286, + 236 + ], + "lines": [ + { + "bbox": [ + 60, + 91, + 286, + 236 + ], + "spans": [ + { + "bbox": [ + 60, + 91, + 286, + 236 + ], + "type": "image", + "image_path": "74a93b5998007ea98ffb514748fe9107e8c3c7078b81c18045d95b64a778c15a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 73, + 421, + 83 + ], + "lines": [ + { + "bbox": [ + 291, + 73, + 421, + 83 + ], + "spans": [ + { + "bbox": [ + 291, + 73, + 421, + 83 + ], + "type": "text", + "content": "(b) Seed-to-outfit generation (Sec 3.2)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 289, + 93, + 553, + 227 + ], + "blocks": [ + { + "bbox": [ + 289, + 93, + 553, + 227 + ], + "lines": [ + { + "bbox": [ + 289, + 93, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 289, + 93, + 553, + 227 + ], + "type": "image", + "image_path": "f482bf3b625cb9b4a66ee812d99b8218998b9b35812c83a7a1860fe6d2e4cbef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 249, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 54, + 249, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 54, + 249, + 555, + 304 + ], + "type": "text", + "content": "Figure 2. System overview of the text to outfit generation framework. For text-to-outfit generation (a), the LLM framework predicts a set of embeddings for different outfit categories, which are used to retrieve the outfit items. We devise an attention masking approach to handle the alignment between the text descriptions in the prompt and the image tokens from different categories (details are in Section 3.1). For seed-to-outfit generation (b), the LLM framework takes a seed item as input, predicts the outfit compositions (outfit categories), and retrieves items in the predicted categories to build the outfits (details are in Section 3.2)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 326, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 326, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 296, + 479 + ], + "type": "text", + "content": "tive outfits. We randomly sample a set of negative outfits for each positive outfit, where each item in the negative outfits has the same high-level category as the positive outfit. Since there are no ground truth negative outfits, random negatives are typically used in outfit training (e.g., [20, 25, 29]). We extract the occasion, season, color and product features of the outfit items using Claude-3 (details are in Section 4.1). To simulate the prompt for training, we randomly select the product features of each outfit item and use Claude-3 to generate the free-form prompt. The goal is to learn the embeddings so that the similarity between the predicted embeddings and the positive outfits is greater than the negative ones." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "type": "text", + "content": "Specifically, a batch includes a set of triples: " + }, + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\Upsilon = \\{T, p, N\\}" + }, + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "type": "text", + "content": " is the prompt, " + }, + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "type": "text", + "content": " is the positive outfit and " + }, + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 480, + 296, + 528 + ], + "type": "text", + "content": " is a set of negative outfits. The outfit retrieval loss is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 534, + 296, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 534, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 64, + 534, + 296, + 566 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e t}} (p, N; \\mathrm {e}) = \\sum_ {i = 1} ^ {n} \\mathcal {L} _ {\\text {c a t} _ {i}} \\left(\\mathrm {f} _ {\\text {c a t} _ {i}} ^ {p}, \\mathrm {f} _ {\\text {c a t} _ {i}} ^ {N}; \\mathrm {e} _ {\\text {c a t} _ {i}}\\right) \\xi (\\text {c a t} _ {i}) \\tag {1}", + "image_path": "669335d7595da75c89461f20b82b83546bdae53c63d0d22d27030a2c81777276.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 572, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 296, + 620 + ], + "type": "text", + "content": "We combine the loss for each category to obtain the outfit retrieval loss. " + }, + { + "bbox": [ + 55, + 572, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 55, + 572, + 296, + 620 + ], + "type": "text", + "content": " is the indicator function denoting if the category " + }, + { + "bbox": [ + 55, + 572, + 296, + 620 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 572, + 296, + 620 + ], + "type": "text", + "content": " is presented in the outfit. The loss for each category is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 625, + 301, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 301, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 301, + 665 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c a t} _ {i}} \\left(\\mathrm {f} _ {\\mathrm {c a t} _ {i}} ^ {p}, \\mathrm {f} _ {\\mathrm {c a t} _ {i}} ^ {N}; \\mathrm {e} _ {\\mathrm {c a t} _ {i}}\\right) = - \\frac {\\exp \\left(\\operatorname {s i m} \\left(\\mathrm {f} _ {\\mathrm {c a t} _ {i}} ^ {p} , \\mathrm {e} _ {\\mathrm {c a t} _ {i}}\\right) / \\tau\\right)}{\\sum_ {n = 1} ^ {N} \\exp \\left(\\operatorname {s i m} \\left(\\mathrm {f} _ {\\mathrm {c a t} _ {i}} ^ {n} , \\mathrm {e} _ {\\mathrm {c a t} _ {i}}\\right) / \\tau\\right)}, \\tag {2}", + "image_path": "72be7d9ae05fd47ea034af68ccf294387728346f45a38d8d823fe0995a83fc61.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": "We adopt the InfoNCE loss ([28]). " + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{f}_{\\mathrm{cat}_i}^p" + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": " is the feature for positive outfit item and " + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{f}_{\\mathrm{cat}_i}^n" + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": " is the feature for negative outfit item. " + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{e}_{\\mathrm{cat}_i}" + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": " is the embeddings. " + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": " is the temperature. We use the cosine function for the sim function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 326, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 492 + ], + "type": "text", + "content": "Attention mask loss. The prompt includes the text descriptions for multiple categories. We observed the alignment problem where the image token from a category may attend to the text descriptions from other categories. For example, the top image token attends to the shoe text descriptions. An example prompt: \"An outfit for spring and formal event. Black high heels. A black dress with a rose print. A red bracelet with gemstones.\" The text describes the features for fine-grained categories like high-heel, dress, etc, but the image tokens correspond to high-level categories (e.g., tops, bottoms, shoes). These pose the challenge for the LLM model to find the relations between the outfit text descriptions and multiple image tokens, which affects the product feature matching performance (cf. Table 3))." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 494, + 556, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 556, + 589 + ], + "type": "text", + "content": "To address this issue, we design a masking approach in the self-attention layers of the LLM decoder. Figure 3 illustrates the mask attention for the top image token. We incorporate the mask loss to guide attention learning, where the mask loss is applied to all image tokens. We compute the binary cross-entropy loss (BCE) for each image token between the ground truth mask and the predicted attention values. The mask loss is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 597, + 555, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 597, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 361, + 597, + 555, + 632 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\sum_ {i = 1} ^ {n} \\sum_ {j = 1} ^ {k} \\mathrm {B C E} _ {i} \\left(\\alpha_ {j}, m _ {j} ^ {\\mathrm {G T}}\\right), \\tag {3}", + "image_path": "5631efd4b754cb7597669ea770c6e6c04fc6b759663d20f859c7abc8e8b4a54e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 555, + 677 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 641, + 555, + 677 + ], + "type": "text", + "content": " is the number of image tokens. " + }, + { + "bbox": [ + 313, + 641, + 555, + 677 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 641, + 555, + 677 + ], + "type": "text", + "content": " is the number of tokens from the text descriptions. " + }, + { + "bbox": [ + 313, + 641, + 555, + 677 + ], + "type": "inline_equation", + "content": "\\alpha_{j}" + }, + { + "bbox": [ + 313, + 641, + 555, + 677 + ], + "type": "text", + "content": " is the predicted attention value and " + }, + { + "bbox": [ + 313, + 641, + 555, + 677 + ], + "type": "inline_equation", + "content": "m_j^{\\mathrm{GT}}" + }, + { + "bbox": [ + 313, + 641, + 555, + 677 + ], + "type": "text", + "content": " is the ground truth mask." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Compatibility loss (CP. loss). We include a compatibility loss to facilitate the embedding learning that improves the compatibility of the generated outfit items. We trained an" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16168" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 295, + 243 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 295, + 243 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 295, + 243 + ], + "type": "image", + "image_path": "548c25207238ed1786a2f06c74d360ef7aa061ce00aee8d4f68a34c2ff0cd21b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 254, + 296, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 299 + ], + "type": "text", + "content": "Figure 3. The figure illustrates the masked attention for the top image token, where the mask facilitates the top image token to attend to the right text descriptions (outfit text and top text sections, highlighted in yellow) in the prompt. Details are in Section 3.1." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 319, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 390 + ], + "type": "text", + "content": "outfit compatibility model (details are in Section 4.2), which predicts a compatibility score given a set of outfit items. In training, we freeze the model weights of the compatibility model (denoted as " + }, + { + "bbox": [ + 55, + 319, + 296, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 319, + 296, + 390 + ], + "type": "text", + "content": ") and fine-tune the embeddings (e) such that the compatibility score is maximized. The compatibility loss is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 399, + 296, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 399, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 124, + 399, + 296, + 414 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c p}} (e) = - \\mathcal {C} \\left(\\mathrm {W} _ {\\mathrm {c p}} \\cdot \\mathrm {e}\\right) \\tag {4}", + "image_path": "18fbe3d08e3c359da4aaf9d24dc85524048a8802486c4fb7e780f446f18ee686.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 422, + 295, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 446 + ], + "type": "text", + "content": "Training and Inference. We train the framework end-to-end using LoRA [2] by minimizing the following loss function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 453, + 295, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 453, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 90, + 453, + 295, + 476 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta , \\mathrm {W} _ {\\text {r e t}}, \\mathrm {W} _ {\\text {i t e m}}, \\mathrm {W} _ {\\mathrm {c p}}} \\mathcal {L} \\left(\\theta , \\mathrm {W} _ {\\text {r e t}}, \\mathrm {W} _ {\\text {i t e m}}, \\mathrm {W} _ {\\mathrm {c p}}\\right) \\tag {5}", + "image_path": "50b6d06f241c85c809cd4359bdbe946ebcc265242cd8b034ba5a900351dee64d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 476, + 260, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 476, + 260, + 490 + ], + "spans": [ + { + "bbox": [ + 100, + 476, + 260, + 490 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {1} \\cdot \\mathcal {L} _ {\\mathrm {r e t}} + \\lambda_ {2} \\cdot \\mathcal {L} _ {\\mathrm {m a s k}} + \\lambda_ {3} \\cdot \\mathcal {L} _ {\\mathrm {c p}},", + "image_path": "cae31069d88a069f909e5458c691ea8b6c7a84a17de187e6667610340f54f5fb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": " is LLM model weights. " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{ret}}" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{item}}" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{cp}}" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": " are the MLP models for outfit retrieval, outfit item feature extraction, and compatibility, respectively. " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ret}}" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": " is the outfit retrieval loss, " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": " is the mask loss and " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cp}}" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": " is the compatibility loss. " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": " are empirically chosen to make the magnitudes of the losses similar in scale. We report the results of using different losses in Table 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 581, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 296, + 616 + ], + "type": "text", + "content": "In the inference, we feed the text prompt to the LLM model to predict the embeddings and use them to retrieve the outfit items." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 624, + 200, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 200, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 200, + 635 + ], + "type": "text", + "content": "3.2. Seed-to-Outfit Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "The task is to generate a complete outfit given an input seed item and a text prompt describing the global outfit features. Unlike the text-to-outfit generation task, where users specify the texts for each outfit item, here, users only specify the information for the seed item. Therefore, we need to predict the composition of the remaining items. One can retrieve" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "the outfit items from all the categories to build an outfit. However, some categories do not go well with certain types of occasions, seasons, or the seed types. For example, coats should not be included for summer outfits, or an all-body dress should not be selected if the seed is a top." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 276 + ], + "type": "text", + "content": "Composition loss. To address this problem, we train the LLM model to predict the outfit compositions (a set of outfit categories), where the compositions are learned from outfit datasets (cf. Figure 2 (b)). Specifically, the input includes a prompt and a seed item (users provide an image or a text description for the seed item. For the image, we use a visual encoder to extract the embeddings and feed them to the LLM). To predict the compositions, we add two MLP models (one for predicting the high-level category and another for predicting the fine-grained category) on the top of the hidden state vectors of the high-level category to predict the composition. The composition loss is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 295, + 555, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 295, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 328, + 295, + 555, + 326 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o m p}} = \\sum_ {i = 1} ^ {n} \\mathrm {C E} _ {i} \\left(y _ {\\mathrm {p r e d}} ^ {\\mathrm {h c}}, y _ {\\mathrm {g t}} ^ {\\mathrm {h c}}\\right) + \\mathrm {C E} _ {i} \\left(y _ {\\mathrm {p r e d}} ^ {\\mathrm {f g}}, y _ {\\mathrm {g t}} ^ {\\mathrm {f g}}\\right), \\tag {6}", + "image_path": "1f5fedee883d3f585ab03739645d339580b1db5e89d2c671074ac4c3efbeaacf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 334, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 382 + ], + "type": "text", + "content": "where the hc denotes the high-level category and fg denotes the fine-grained category. CE is the cross-entropy loss. We train the framework with LoRA end-to-end with the following loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 352, + 392, + 555, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 392, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 352, + 392, + 555, + 406 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {4} \\cdot \\mathcal {L} _ {\\text {r e t}} + \\lambda_ {5} \\cdot \\mathcal {L} _ {\\mathrm {c p}} + \\lambda_ {6} \\cdot \\mathcal {L} _ {\\text {c o m p}} \\tag {7}", + "image_path": "a8f3c3bc2ca44fba497d996733655839de33ca74b3f79774e4c5b15afc822030.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 498 + ], + "type": "text", + "content": "The loss function includes the retrieval, compatibility, and composition prediction losses. We train the model to predict the composition of the seed item belonging to the top category. The performance of other categories for the seed item is in Appendix Table 2. In inference, we search for outfit items in the predicted high-level and fine-grained categories to compose an outfit." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 509, + 395, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 395, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 395, + 522 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 529, + 373, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 373, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 373, + 540 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 545, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 653 + ], + "type": "text", + "content": "We evaluate the model performance on the Polyvore dataset [29]. The dataset includes two different sets: a disjoint set and a non-disjoint set. The non-disjoint set contains 53,306 training outfits and 10,000 testing outfits, while the disjoint set contains 16,995 training outfits and 15,145 testing outfits. We train our framework on the non-disjoint training set as it contains a larger number of outfits and report the performance on both non-disjoint and dis-joint testing sets (results are in the Appendix Section 1)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "The Polyvore dataset does not have annotated product attributes. Motivated by recent works that leverage LLM for annotations (e.g., [19, 31]), we leverage Claude-3 [3] to extract outfit-level attributes (occasion and season) and item-level attributes (color and product features). We verify" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16169" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": "the agreement between the predicted attributes from Claude-3 and humans on a subset of testing samples (500 outfits and 2629 items). The results show the Claude-3 annotated attributes achieve " + }, + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "inline_equation", + "content": "94.7\\%" + }, + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": " overall precision " + }, + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "inline_equation", + "content": "(97.3\\%)" + }, + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": " for color, " + }, + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "inline_equation", + "content": "95.5\\%" + }, + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": " for product feature, " + }, + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "inline_equation", + "content": "92.2\\%" + }, + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": " for occasion, and " + }, + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "inline_equation", + "content": "93.6\\%" + }, + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": " for season), verifying the feasibility of using Claude-3 for attribute annotations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 165, + 129, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 165, + 129, + 177 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 129, + 177 + ], + "type": "text", + "content": "4.2. Evaluation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 297, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 297, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 297, + 337 + ], + "type": "text", + "content": "We evaluate the quality of the retrieved outfits in two ways: 1) feature matching for outfit items and 2) outfit compatibility. To evaluate feature matching, we compute the number of items in the top-10 retrieval results that match the product features specified in the prompt. We evaluate the feature matching for color, occasion, season, and product feature attributes for the outfit items, which are critical attributes for outfits. To handle non-exact matches, where the product features are different but their meanings are similar, we use ModernBERT embedding [22] to evaluate the affinity of two features (we select the thresholds of different attributes using a validation set). We report the results using precision@top-10 (denoted as P@10)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 339, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 482 + ], + "type": "text", + "content": "To evaluate outfit compatibility, we train an outfit compatibility model using a SOTA architecture (OutfitTransformer [25]), where the model takes a set of outfit items as input, and predicts a compatibility score. We replace the ResNet features in the original architecture with the FashionCLIP embeddings [6] and use the item image and the text description as input. We achieve better outfit compatibility prediction than the original architecture. We achieve " + }, + { + "bbox": [ + 55, + 339, + 296, + 482 + ], + "type": "inline_equation", + "content": "94.93\\%" + }, + { + "bbox": [ + 55, + 339, + 296, + 482 + ], + "type": "text", + "content": " (93% for OutfitTransformer) and " + }, + { + "bbox": [ + 55, + 339, + 296, + 482 + ], + "type": "inline_equation", + "content": "93.88\\%" + }, + { + "bbox": [ + 55, + 339, + 296, + 482 + ], + "type": "text", + "content": " AUC (88% for OutfitTransformer) for the non-disjoint and disjoint sets respectively. We report the compatibility (CP) score for the top 10 retrieved outfits." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 297, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 297, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 297, + 639 + ], + "type": "text", + "content": "There is no ground-truth outfit composition given a seed and prompt. To obtain a quantitative quality score, we use Claude-3 (we also evaluate the performance with Pixtral Large from Mistral AI, see Appendix Section 2) to evaluate the quality of the predicted compositions. Recent work (e.g., [33]) shows that strong LLMs achieve high agreement with humans. We evaluate the quality in two ways: 1) We use Claude-3 to predict a quality score for each predicted composition and report an average quality score (denoted as Comp. score). 2) Given two compositions generated by the baseline and our approach, we ask Claude-3 to pick the better composition and report the winning rate of our approach (denoted as Comp. win rate)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 215, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 215, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 215, + 661 + ], + "type": "text", + "content": "4.3. Baselines and Model Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "There is no existing framework for text-to-outfit retrieval tasks ([11] is designed to retrieve entire outfits, which is different from our settings as mentioned in Section 2). We convert a set of baseline approaches to the outfit retrieval" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 191 + ], + "type": "text", + "content": "tasks and report the performance. In Table 1, we categorize these baselines into the text-to-single-item retrieval and text-to-outfit retrieval approaches. Existing methods for text-to-image retrieval [6, 14, 24] mainly focus on single-image retrieval. To make these methods applicable to outfit retrieval problem, we make the following changes: 1) We convert the outfit-level specification (occasion/season) into item-level specification. 2) We use the text descriptions of each category to retrieve the outfit items (assuming that the text descriptions to category correspondence are given)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 193, + 556, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 556, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 556, + 288 + ], + "type": "text", + "content": "For text-to-outfit retrieval, we converted the GILL [14] framework to support multiple-category retrieval. For text-to-outfit retrieval methods, the text descriptions to category correspondence information are not given, which is directly predicted by the models (except the fixed mask approach in Table 1, which uses the text-to-category correspondence to generate the fixed mask for attention). The details of the baseline approaches are described below." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 289, + 555, + 516 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 313, + 289, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 324 + ], + "type": "text", + "content": "- CLIP [24]: We convert the outfit-level specification into item-level specifications and perform retrieval for each category. We use ViT-L-14 version." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 325, + 555, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 371 + ], + "type": "text", + "content": "- Fashion-CLIP 2.0 [6]: We use the same settings as CLIP and use Fashion-CLIP to retrieve items for each category. We use version 2.0, which achieves better performance than version 1.0." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 373, + 553, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 553, + 420 + ], + "type": "text", + "content": "- BLIP-2 [17]: We use the pre-trained image encoder and Q-Former from BLIP-2. Like CLIP, we perform image retrieval for each category using the corresponding text description. We use the ViT-g version." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 421, + 553, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 553, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 553, + 516 + ], + "type": "text", + "content": "- GILL [14] + Multiple categories: We add the learnable image tokens for each high-level category for multiple image retrieval. We fine-tune the MLP layers to learn embeddings while keeping the LLM model fixed following the original training approach. We use the same prompt format as our model. Note that we compare our method with an improved version of GiLL, in which we replace the original CLIP vision encoder with Fashion-CLIP" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 517, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 556, + 661 + ], + "type": "text", + "content": "Model Settings: For our framework, we use the pretrained OPT 6.7B model [32] (same as GILL) as the LLM backbone. We fine-tune the LLM with LoRA (we set the rank to 64 for Q, K, and V projection matrices). We use Fashion-CLIP 2.0 [6] as the visual encoder " + }, + { + "bbox": [ + 313, + 517, + 556, + 661 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{img}}" + }, + { + "bbox": [ + 313, + 517, + 556, + 661 + ], + "type": "text", + "content": " and the text encoder " + }, + { + "bbox": [ + 313, + 517, + 556, + 661 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{text}}" + }, + { + "bbox": [ + 313, + 517, + 556, + 661 + ], + "type": "text", + "content": ". Fashion-CLIP 2.0 achieves better performance than CLIP embeddings. In model training, the visual and text encoder weights are fixed. We use the bi-directional masks for image tokens. We use 11 image tokens for the Polyvore dataset, corresponding to the following high-level categories: tops, bottoms, all-body, outerwear, shoes, bags, jewelry, accessories, sunglasses, hats, and scarves." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 671, + 518, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 518, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 518, + 682 + ], + "type": "text", + "content": "4.4. Text-to-Outfit Performance Evaluation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Table 1 summarizes the text-to-outfit retrieval results. We use the testing set of the Polyvore dataset to generate the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16170" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 555, + 189 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 555, + 189 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 555, + 189 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 555, + 189 + ], + "type": "table", + "html": "
MethodModalityColor (P@10)Product feature (P@10)Season (P@10)Occasion (P@10)Avg (P@10)CP. score
Text-to-single-item retrieval
CLIP [24]Text only96.0670.9190.2276.6283.4663.49
Fashion-CLIP [6]Text only96.8280.4289.6881.2787.0564.07
BLIP-2 [17]Text only93.2966.9485.4684.5182.5565.09
CLIP [24]Text to image83.5961.3285.4967.0174.3549.77
Fashion-CLIP [6]Text to image87.5972.6785.9168.8878.7653.03
BLIP-2 [17]Text to image86.3966.6985.6267.7576.6151.05
Text-to-outfit retrieval
GILL [14] + multiple categoryText to image48.1948.1586.8769.6963.2374.26
Text2Outfit + learnable mask (ours)Text to image78.4276.5486.6070.4578.0066.28
Text2Outfit + learnable mask (ours)Text to image/text86.0584.8195.6993.6690.0584.05
Text2Outfit + fixed mask (ours)Text to image/text96.6186.2794.8792.4892.5682.51
", + "image_path": "a4ea5b61ccf13e886c2ca23428576f7fd81720cf696b79894a712663172add94.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 251, + 553, + 301 + ], + "blocks": [ + { + "bbox": [ + 55, + 197, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 197, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 555, + 242 + ], + "type": "text", + "content": "Table 1. Text-to-outfit retrieval performance comparison. P@10 denotes precision at top-10 and CP. score denotes compatibility score using [25] for top-10 retrieved outfits. Note that for single-item retrieval approaches, the text descriptions and its corresponding category are given while the outfit retrieval approaches are not given such information. The bold text denotes the best performance and the underlines denote the second best performance. Details are in Section 4.2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 251, + 553, + 301 + ], + "lines": [ + { + "bbox": [ + 56, + 251, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 553, + 301 + ], + "type": "table", + "html": "
MethodCompositionSeason (P@10)Occasion (P@10)Comp. score (c3)Comp win rate (c3)CP. scoreCP. score (c3)
Seed text + prompt (baseline)All categories97.3296.202.596%80.822.78
Seed text + prompt (ours)Predicted97.5696.014.5794%86.183.85
Seed image + prompt (baseline)All categories95.0087.502.858%72.882.66
Seed image + prompt (ours)Predicted96.4391.804.3292%83.363.83
", + "image_path": "bd5f349ad2db606f4ac8c9e01ec3de7460a6f6e3a05d59df12187f6f8c2c8192.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 308, + 544, + 319 + ], + "lines": [ + { + "bbox": [ + 64, + 308, + 544, + 319 + ], + "spans": [ + { + "bbox": [ + 64, + 308, + 544, + 319 + ], + "type": "text", + "content": "Table 2. Seed-to-outfit retrieval performance comparison. The scores with (c3) are evaluated by Claude-3. Details are in Section 4.5." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 341, + 295, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 295, + 400 + ], + "type": "text", + "content": "prompt. Given a test outfit including multiple outfit items, we randomly select outfit attributes (occasion and season) and item attributes (color and product features) and use Claude-3 to generate the free-form prompt. We use the generated prompt to evaluate the outfit retrieval performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 401, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 401, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 401, + 295, + 567 + ], + "type": "text", + "content": "Our approaches achieve " + }, + { + "bbox": [ + 54, + 401, + 295, + 567 + ], + "type": "inline_equation", + "content": "90.05\\%" + }, + { + "bbox": [ + 54, + 401, + 295, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 401, + 295, + 567 + ], + "type": "inline_equation", + "content": "92.56\\%" + }, + { + "bbox": [ + 54, + 401, + 295, + 567 + ], + "type": "text", + "content": " average " + }, + { + "bbox": [ + 54, + 401, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathrm{P@10}" + }, + { + "bbox": [ + 54, + 401, + 295, + 567 + ], + "type": "text", + "content": " using the learnable mask and fixed mask respectively, showing better feature-matching precision than the baseline approaches. The GILL + multi-category approach has an alignment issue (the alignment between the outfit texts and the images tokens) while our approach uses the learnable mask to alleviate the alignment issue. We fine-tune the model end-to-end with LoRA instead of using the frozen LLM model as GILL, improving the performance (see Table 3 for detailed analysis). Our approach with a learnable mask also performs better than the single-item retrieval approaches, where they are given the text description to category correspondence information, while our approach does not." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "In addition, our approach shows better outfit compatibility than the baseline approaches. We achieve the compatibility score 84.05 using the learnable mask setting (we compute the average compatibility scores for the ground truth positive (86.95) and negative (22.99) outfits on the Polyvore testing set using OutfitTransformer. The generated outfits from our method achieve similar compatibility as the positive outfits). Single-item retrieval approaches perform worse on outfit compatibility because they find items that match the text descriptions but do not model the item relationships in an outfit. Our framework is fine-tuned on the outfit data, and incorporates the outfit retrieval and compatibility losses," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 341, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 365 + ], + "type": "text", + "content": "which improve overall compatibility. Figure 4 shows example results of our framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 371, + 520, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 520, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 520, + 384 + ], + "type": "text", + "content": "4.5. Seed-to-Outfit Performance Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 388, + 556, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 460 + ], + "type": "text", + "content": "Table 2 summarizes the results for seed-to-outfit retrieval. We used the Polyvore test set to construct the input (seed and prompt) for evaluation. Given a test outfit including multiple outfit items, we randomly select the outfit attributes (occasion and season) and use Claude-3 to generate the free-form prompt." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 461, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 665 + ], + "type": "text", + "content": "We compare our approach with a baseline approach that uses all 11 high-level categories for the composition. We feed the compositions (e.g., category 1/fine-grained category 1, category 2/fine-grained category 2, ...) of the baseline and our approach to Claude-3 for evaluation (for the baseline approach, the fine-grained category is obtained from the top-1 retrieved outfit). Our approach shows better composition scores (the score is between 1 (poor) to 5 (excellent)) and winning rates than the baseline (e.g., " + }, + { + "bbox": [ + 313, + 461, + 556, + 665 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 313, + 461, + 556, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 461, + 556, + 665 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 313, + 461, + 556, + 665 + ], + "type": "text", + "content": " winning rate of using seed text and seed image, respectively). We verify the agreement between Claude-3 and humans on a subset of samples (500 outfitsits). The Spearman's rank correlation coefficient is 0.73, which indicates they are highly correlated. We report the visual compatibility using both OutfitTransformer [25] (CP. score) and Claude-3 (CP. score (c3)). Our approach shows better visual compatibility than the baseline approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 672, + 408, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 408, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 408, + 685 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Multimodal feature: We compare the performance of single-modality (image) vs. multimodality (image + text)" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16171" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 555, + 138 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 555, + 138 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 555, + 138 + ], + "type": "table", + "html": "
ModalityLoRAMask LossCP. lossColor (P@10)Product feature (P@10)Season (P@10)Occasion (P@10)Avg (P@10)CP. score
1Image---48.3153.0385.1068.6263.7653.43
2Image + text---44.6047.1888.9481.0765.4578.62
3Image + text✓--74.7475.0195.6093.9084.8286.95
4Image + text✓Global CE-79.2882.2396.0294.4387.9985.63
5Image + text✓Local BCE-85.5884.8795.8993.4789.9583.43
6Image + text✓Local BCE✓86.0584.8195.6993.6690.0584.05
", + "image_path": "825440f05325dd9143ac5306f2fee0c8c3a808764019b2227d9e82421a698dcf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 116, + 146, + 492, + 157 + ], + "lines": [ + { + "bbox": [ + 116, + 146, + 492, + 157 + ], + "spans": [ + { + "bbox": [ + 116, + 146, + 492, + 157 + ], + "type": "text", + "content": "Table 3. Ablation study of different components in our framework. Details are described in Section 4.6." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 66, + 173, + 192, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 173, + 192, + 186 + ], + "spans": [ + { + "bbox": [ + 66, + 173, + 192, + 186 + ], + "type": "text", + "content": "(a) Text-to-outfit generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 194, + 237, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 194, + 237, + 229 + ], + "spans": [ + { + "bbox": [ + 65, + 194, + 237, + 229 + ], + "type": "text", + "content": "Prompt: Spring outfit for casual. A canvas upper flats in a pink shade. A jeans in black with high-waisted accents. A sleeveless top in a white shade. A denim jacket in a blue shade." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 66, + 251, + 106, + 291 + ], + "blocks": [ + { + "bbox": [ + 69, + 241, + 146, + 249 + ], + "lines": [ + { + "bbox": [ + 69, + 241, + 146, + 249 + ], + "spans": [ + { + "bbox": [ + 69, + 241, + 146, + 249 + ], + "type": "text", + "content": "GILL + Multiple Category" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 251, + 106, + 291 + ], + "lines": [ + { + "bbox": [ + 66, + 251, + 106, + 291 + ], + "spans": [ + { + "bbox": [ + 66, + 251, + 106, + 291 + ], + "type": "image", + "image_path": "d4efe064e518d69a18e7ff1e1e3a626f988c797752138e13c95524d7bfbe7e1e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 111, + 251, + 152, + 291 + ], + "blocks": [ + { + "bbox": [ + 111, + 251, + 152, + 291 + ], + "lines": [ + { + "bbox": [ + 111, + 251, + 152, + 291 + ], + "spans": [ + { + "bbox": [ + 111, + 251, + 152, + 291 + ], + "type": "image", + "image_path": "9be4457d91776668bdd16f14a1962ac1ec65ed9777cf919efe78084094afba9e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 157, + 251, + 198, + 291 + ], + "blocks": [ + { + "bbox": [ + 157, + 251, + 198, + 291 + ], + "lines": [ + { + "bbox": [ + 157, + 251, + 198, + 291 + ], + "spans": [ + { + "bbox": [ + 157, + 251, + 198, + 291 + ], + "type": "image", + "image_path": "18efad0ff3ea480b301f1711bda6c09bac818333885e997ff6ca54342dda0da2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 204, + 251, + 245, + 291 + ], + "blocks": [ + { + "bbox": [ + 204, + 251, + 245, + 291 + ], + "lines": [ + { + "bbox": [ + 204, + 251, + 245, + 291 + ], + "spans": [ + { + "bbox": [ + 204, + 251, + 245, + 291 + ], + "type": "image", + "image_path": "ce5b83a67e98e063fbfe395c9395b168cdd7586bd7974ca14a54dfedb23e1344.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 65, + 315, + 106, + 354 + ], + "blocks": [ + { + "bbox": [ + 69, + 305, + 178, + 313 + ], + "lines": [ + { + "bbox": [ + 69, + 305, + 178, + 313 + ], + "spans": [ + { + "bbox": [ + 69, + 305, + 178, + 313 + ], + "type": "text", + "content": "Text2Outfit + learnable mask (ours)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 315, + 106, + 354 + ], + "lines": [ + { + "bbox": [ + 65, + 315, + 106, + 354 + ], + "spans": [ + { + "bbox": [ + 65, + 315, + 106, + 354 + ], + "type": "image", + "image_path": "a1317f57e31be65ee5a819c8d98d7c40e2bead98444790a87697636326383654.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 111, + 315, + 152, + 354 + ], + "blocks": [ + { + "bbox": [ + 111, + 315, + 152, + 354 + ], + "lines": [ + { + "bbox": [ + 111, + 315, + 152, + 354 + ], + "spans": [ + { + "bbox": [ + 111, + 315, + 152, + 354 + ], + "type": "image", + "image_path": "030bdd17657e64778cf578361ddbb6a322b541c65e28e1fce63eb8a89d9d08fd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 159, + 315, + 197, + 354 + ], + "blocks": [ + { + "bbox": [ + 159, + 315, + 197, + 354 + ], + "lines": [ + { + "bbox": [ + 159, + 315, + 197, + 354 + ], + "spans": [ + { + "bbox": [ + 159, + 315, + 197, + 354 + ], + "type": "image", + "image_path": "6ba38cb2439c77f1e4cb73a6098cc169320807725a6cefd3cfc5a9a5720569a7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 204, + 315, + 244, + 355 + ], + "blocks": [ + { + "bbox": [ + 204, + 315, + 244, + 355 + ], + "lines": [ + { + "bbox": [ + 204, + 315, + 244, + 355 + ], + "spans": [ + { + "bbox": [ + 204, + 315, + 244, + 355 + ], + "type": "image", + "image_path": "71edc74bafd512ef522181b76693381aaf120b531651abe27973106c59e5f6a9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 65, + 383, + 106, + 422 + ], + "blocks": [ + { + "bbox": [ + 69, + 373, + 164, + 381 + ], + "lines": [ + { + "bbox": [ + 69, + 373, + 164, + 381 + ], + "spans": [ + { + "bbox": [ + 69, + 373, + 164, + 381 + ], + "type": "text", + "content": "Text2Outfit + fixed mask (ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 383, + 106, + 422 + ], + "lines": [ + { + "bbox": [ + 65, + 383, + 106, + 422 + ], + "spans": [ + { + "bbox": [ + 65, + 383, + 106, + 422 + ], + "type": "image", + "image_path": "1b816c838c4695ed9cd965bfe7e7a0976de041081e3a02d1eb6872a094d8b300.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 111, + 383, + 152, + 422 + ], + "blocks": [ + { + "bbox": [ + 111, + 383, + 152, + 422 + ], + "lines": [ + { + "bbox": [ + 111, + 383, + 152, + 422 + ], + "spans": [ + { + "bbox": [ + 111, + 383, + 152, + 422 + ], + "type": "image", + "image_path": "d168ea803e55d552c8bf367a9ea55d9770ad0f256f9cfe9b0f420f2d34ef8b4a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 449, + 555, + 493 + ], + "lines": [ + { + "bbox": [ + 55, + 449, + 555, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 555, + 493 + ], + "type": "text", + "content": "Figure 4. Qualitative results. We compare the results for text-to-outfit generation with the extended GILL [14] approach that supports multiple image retrieval and our framework (learnable mask and fixed mask). The non-matched items are marked with boxes. We show the predicted compositions and retrieval results for seed-to-outfit generation with a seed image or text. More visual results are in the Appendix Section 4." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 159, + 384, + 197, + 422 + ], + "blocks": [ + { + "bbox": [ + 159, + 384, + 197, + 422 + ], + "lines": [ + { + "bbox": [ + 159, + 384, + 197, + 422 + ], + "spans": [ + { + "bbox": [ + 159, + 384, + 197, + 422 + ], + "type": "image", + "image_path": "5f4c32543f29a9a9b14363f4dc103a17f4c8f0bc3f5e0ff9318190bd215d2144.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 204, + 384, + 244, + 422 + ], + "blocks": [ + { + "bbox": [ + 204, + 384, + 244, + 422 + ], + "lines": [ + { + "bbox": [ + 204, + 384, + 244, + 422 + ], + "spans": [ + { + "bbox": [ + 204, + 384, + 244, + 422 + ], + "type": "image", + "image_path": "b6177f2df1e6f18c2d3d414e179c2abef279d14d3e9cd1bc187de643f8c8c17c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 257, + 173, + 387, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 173, + 387, + 186 + ], + "spans": [ + { + "bbox": [ + 257, + 173, + 387, + 186 + ], + "type": "text", + "content": "(b) Seed-to-outfit generation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 257, + 193, + 362, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 193, + 362, + 202 + ], + "spans": [ + { + "bbox": [ + 257, + 193, + 362, + 202 + ], + "type": "text", + "content": "Prompt: Outfit for casual in spring." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 258, + 203, + 548, + 316 + ], + "blocks": [ + { + "bbox": [ + 258, + 203, + 548, + 316 + ], + "lines": [ + { + "bbox": [ + 258, + 203, + 548, + 316 + ], + "spans": [ + { + "bbox": [ + 258, + 203, + 548, + 316 + ], + "type": "image", + "image_path": "29714022b8032df6c9d2e00402574ad4af879ccb5a9fb37ebe02796f802a0f77.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 257, + 317, + 403, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 317, + 403, + 327 + ], + "spans": [ + { + "bbox": [ + 257, + 317, + 403, + 327 + ], + "type": "text", + "content": "Prompt: Formal events outfit for spring weather." + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 258, + 327, + 548, + 439 + ], + "blocks": [ + { + "bbox": [ + 258, + 327, + 548, + 439 + ], + "lines": [ + { + "bbox": [ + 258, + 327, + 548, + 439 + ], + "spans": [ + { + "bbox": [ + 258, + 327, + 548, + 439 + ], + "type": "image", + "image_path": "78d6996f0a84d87d05b1fb3de30f9b519db6b6178dedb92d37c151e3ba37b3d1.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 54, + 514, + 297, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 514, + 297, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 514, + 297, + 552 + ], + "type": "text", + "content": "(row-1 and row-2 in Table 3). Using multimodality improves the average P@10 from " + }, + { + "bbox": [ + 54, + 514, + 297, + 552 + ], + "type": "inline_equation", + "content": "63.76\\%" + }, + { + "bbox": [ + 54, + 514, + 297, + 552 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 514, + 297, + 552 + ], + "type": "inline_equation", + "content": "65.45\\%" + }, + { + "bbox": [ + 54, + 514, + 297, + 552 + ], + "type": "text", + "content": " for feature matching." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 554, + 296, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 614 + ], + "type": "text", + "content": "LoRA: We compare the performance of fine-tuned LLM (with LoRA) vs. frozen LLM (row-2 and row-3 in Table 3). Using LoRA significantly improves the average P@10 from " + }, + { + "bbox": [ + 55, + 554, + 296, + 614 + ], + "type": "inline_equation", + "content": "65.45\\%" + }, + { + "bbox": [ + 55, + 554, + 296, + 614 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 554, + 296, + 614 + ], + "type": "inline_equation", + "content": "84.82\\%" + }, + { + "bbox": [ + 55, + 554, + 296, + 614 + ], + "type": "text", + "content": " and improves the compatibility score from " + }, + { + "bbox": [ + 55, + 554, + 296, + 614 + ], + "type": "inline_equation", + "content": "78.62\\%" + }, + { + "bbox": [ + 55, + 554, + 296, + 614 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 554, + 296, + 614 + ], + "type": "inline_equation", + "content": "86.95\\%" + }, + { + "bbox": [ + 55, + 554, + 296, + 614 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 617, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 297, + 714 + ], + "type": "text", + "content": "Mask loss: We investigate different loss functions for the attention masking: Global CE and Local BCE (row-4 and row-5 in Table 3). For Global CE, we consider the attention values from all the text tokens to each image token as a distribution and optimize a global cross-entropy loss. We optimize each attention value for Local BCE with a binary cross-entropy loss. Local BCE performs better, as it optimizes the attention directly with the ground truth mask," + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 514, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 537 + ], + "type": "text", + "content": "while global CE optimizes the overall distribution of the attention." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 539, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 575 + ], + "type": "text", + "content": "Compatibility loss: We analyze the impact of including a compatibility loss in the training loss (cf. Equation 5). It further improves compatibility (row-6 in Table 3)." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 583, + 392, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 392, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 392, + 596 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 603, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 556, + 713 + ], + "type": "text", + "content": "We present a framework for text-based outfit generation. We explore two different forms of outfit generation: 1) text-to-outfit generation and 2) seed-to-outfit generation tasks, which enable the controllable outfit generation. In addition, we introduce a masking mechanism which better handles the alignment between outfit texts and the image tokens, and achieves comparable performance when using the ground-truth masks. The experimental results demonstrate that our model outperforms the baseline approaches for text to outfit retrieval tasks." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16172" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 114, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 114, + 297, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 297, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 297, + 168 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. NeurIPS, 35: 23716-23736, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 170, + 296, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 296, + 257 + ], + "type": "text", + "content": "[2] Jason Ansel, Edward Yang, Horace He, Natalia Gimelshein, Animesh Jain, Michael Voznesensky, Bin Bao, Peter Bell, David Berard, Evgeni Burovski, et al. Pytorch 2: Faster machine learning through dynamic python bytecode transformation and graph compilation. In Proceedings of the 29th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 2, pages 929-947, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 258, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 296, + 290 + ], + "type": "text", + "content": "[3] Anthropic. Introducing the next generation of claude. https://www.anthropic.com/news/claude-3-family, 2024.2,5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 292, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 295, + 323 + ], + "type": "text", + "content": "[4] Federico Becattini, Federico Maria Teotini, and Alberto Del Bimbo. Transformer-based graph neural networks for outfit generation, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 325, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 380 + ], + "type": "text", + "content": "[5] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "type": "text", + "content": "[6] Patrick John Chia, Giuseppe Attanasio, Federico Bianchi, Silvia Terragni, Ana Rita Magalhães, Diogo Goncalves, Ciro Greco, and Jacopo Tagliabue. Contrastive language and vision learning of general fashion concepts. Scientific Reports, 12 (1):18958, 2022. 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 436, + 295, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 468 + ], + "type": "text", + "content": "[7] Guillem Cucurull, Perouz Taslakian, and David Vazquez. Context-aware visual compatibility prediction. In CVPR, pages 12617-12626, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 295, + 523 + ], + "type": "text", + "content": "[8] Runpei Dong, Chunrui Han, Yang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, Xiangwen Kong, Xiangyu Zhang, Kaisheng Ma, and Li Yi. DreamLLM: Synergistic multimodal comprehension and creation. In ICLR, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 557 + ], + "type": "text", + "content": "[9] Xintong Han, Zuxuan Wu, Yu-Gang Jiang, and Larry S Davis. Learning fashion compatibility with bidirectional lstms. In ACM MM, pages 1078-1086, 2017. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 296, + 590 + ], + "type": "text", + "content": "[10] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 296, + 635 + ], + "type": "text", + "content": "[11] Junkyu Jang, Eugene Hwang, and Sung-Hyuk Park. Lost your style? navigating with semantic-level approach for text-to-outfit retrieval. In WACV, pages 8066-8075, 2024. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 689 + ], + "type": "text", + "content": "[12] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916. PMLR, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[13] Yuming Jiang, Shuai Yang, Haonan Qiu, Wayne Wu, Chen Change Loy, and Ziwei Liu. Text2human: Text-driven" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "type": "text", + "content": "controllable human image generation. ACM Transactions on Graphics (TOG), 41(4):1-11, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 128 + ], + "type": "text", + "content": "[14] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. NeurIPS, 36, 2023. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 130, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 555, + 173 + ], + "type": "text", + "content": "[15] Jing Yu Koh, Ruslan Salakhutdinov, and Daniel Fried. Grounding language models to images for multimodal inputs and outputs. In ICML, pages 17283-17300. PMLR, 2023. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 218 + ], + "type": "text", + "content": "[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 262 + ], + "type": "text", + "content": "[17] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742. PMLR, 2023. 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 265, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 554, + 297 + ], + "type": "text", + "content": "[18] Shikai Li, Jianglin Fu, Kaiyuan Liu, Wentao Wang, Kwan-Yee Lin, and Wayne Wu. Cosmicman: A text-to-image foundation model for humans. In CVPR, pages 6955-6965, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 300, + 555, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 555, + 353 + ], + "type": "text", + "content": "[19] Wei Li, Hehe Fan, Yongkang Wong, Yi Yang, and Mohan Kankanhalli. Improving context understanding in multimodal large language models via multimodal composition learning. In *Forty-first International Conference on Machine Learning*, 2024. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 355, + 555, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 555, + 387 + ], + "type": "text", + "content": "[20] Yen-Liang Lin, Son Tran, and Larry S Davis. Fashion outfit complementary item retrieval. In CVPR, pages 3311-3319, 2020. 1, 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 555, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 555, + 443 + ], + "type": "text", + "content": "[21] Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval, pages 43-52, 2015. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 445, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 554, + 477 + ], + "type": "text", + "content": "[22] Zach Nussbaum, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. Nomic embed: Training a reproducible long context text embedder, 2024. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 479, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 555, + 521 + ], + "type": "text", + "content": "[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 524, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 524, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 524, + 555, + 578 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 580, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 555, + 633 + ], + "type": "text", + "content": "[25] Rohan Sarkar, Navaneeth Bodla, Mariya I Vasileva, Yen-Liang Lin, Anurag Beniwal, Alan Lu, and Gerard Medioni. Outfittransformer: Learning outfit representations for fashion recommendation. In WACV, pages 3601-3609, 2023. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 679 + ], + "type": "text", + "content": "[26] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, pages 14398-14409, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[27] Reuben Tan, Mariya I Vasileva, Kate Saenko, and Bryan A Plummer. Learning similarity conditions without explicit supervision. In ICCV, pages 10373-10382, 2019. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16173" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 362 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 105 + ], + "type": "text", + "content": "[28] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2019. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[29] Mariya I Vasileva, Bryan A Plummer, Krishna Dusad, Shreya Rajpal, Ranjitha Kumar, and David Forsyth. Learning type-aware embeddings for fashion compatibility. In ECCV, pages 390-405, 2018. 1, 2, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "text", + "content": "[30] Andreas Veit, Balazs Kovacs, Sean Bell, Julian McAuley, Kavita Bala, and Serge Belongie. Learning visual clothing style with heterogeneous dyadic co-occurrences. In ICCV, pages 4642–4650, 2015. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 296, + 229 + ], + "type": "text", + "content": "[31] Shuohang Wang, Yang Liu, Yichong Xu, Chenguang Zhu, and Michael Zeng. Want to reduce labeling cost? gpt-3 can help, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 296, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 284 + ], + "type": "text", + "content": "[32] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 296, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 296, + 362 + ], + "type": "text", + "content": "[33] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, Hao Zhang, Joseph E. Gonzalez, and Ion Stoica. Judging LLM-as-a-judge with MT-bench and chatbot arena. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16174" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/011b91bd-0c46-4896-9ef0-7e4bb90ad7cb_content_list.json b/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/011b91bd-0c46-4896-9ef0-7e4bb90ad7cb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..32e72f04ae3396fe61c41095d40e31d1f8b6dadd --- /dev/null +++ b/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/011b91bd-0c46-4896-9ef0-7e4bb90ad7cb_content_list.json @@ -0,0 +1,1614 @@ +[ + { + "type": "text", + "text": "Text2VDM: Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting", + "text_level": 1, + "bbox": [ + 261, + 130, + 736, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hengyu Meng $^{1}$ Duotun Wang $^{1}$ Zhijing Shao $^{1}$ Ligang Liu $^{2}$ Zeyu Wang $^{1,3*}$ $^{1}$ The Hong Kong University of Science and Technology (Guangzhou) \n $^{2}$ University of Science and Technology of China \n $^{3}$ The Hong Kong University of Science and Technology", + "bbox": [ + 130, + 202, + 867, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8808fca1b5050677e7463aca4d4c257791061307f1993715e82713dfa58cbf92.jpg", + "image_caption": [ + "Figure 1. Example VDM brushes generated by Text2VDM and sculpted models in Blender. Text2VDM can produce high-quality brushes for surface details (top row) and geometric structures (bottom row) from text input. Users can rapidly create an expressive model from a plain shape by directly applying these brushes in Blender. Yellow underlined text highlights semantics enhanced by our framework." + ], + "image_footnote": [], + "bbox": [ + 106, + 309, + 895, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 606, + 326, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Professional 3D asset creation often requires diverse sculpting brushes to add surface details and geometric structures. Despite recent progress in 3D generation, producing reusable sculpting brushes compatible with artists' workflows remains an open and challenging problem. These sculpting brushes are typically represented as vector displacement maps (VDMs), which existing models cannot easily generate compared to natural images. This paper presents Text2VDM, a novel framework for text-to-VDM brush generation through the deformation of a dense planar mesh guided by score distillation sampling (SDS). The original SDS loss is designed for generating full objects and struggles with generating desirable sub-object structures from scratch in brush generation. We refer to this issue as semantic coupling, which we address by introducing weighted blending of prompt tokens to SDS, resulting in a", + "bbox": [ + 88, + 637, + 485, + 880 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "more accurate target distribution and semantic guidance. Experiments demonstrate that Text2VDM can generate diverse, high-quality VDM brushes for sculpting surface details and geometric structures. Our generated brushes can be seamlessly integrated into mainstream modeling software, enabling various applications such as mesh stylization and real-time interactive modeling.", + "bbox": [ + 511, + 608, + 908, + 713 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 751, + 645, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sculpting brushes are essential tools in 3D asset creation, as artists often require a variety of brushes to create surface details and geometric structures. In modeling software, 3D sculpting brushes are typically defined as vector displacement maps (VDMs). A VDM is a 2D image where each pixel stores a 3D displacement vector. Through these vectors, VDM brushes can create complex surface details, such as cracks and wood grain, or generate geometric structures", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 114, + 887, + 241, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16882", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "like ears and horns. This allows artists to apply the same geometric pattern iteratively while sculpting.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite significant advances in text-to-image (T2I) [36, 41] and text-to-3D generation [23, 30, 37, 49, 50], existing methods are unsuitable for creating VDM brushes. We summarize the challenges as follows: 1) Since VDMs are not natural images (Figure 2), it is difficult for existing T2I models to generate them directly. 2) From a 3D perspective, a VDM represents mesh deformation through per-vertex displacement vectors from a dense planar mesh. Mapping any generated mesh to a dense planar mesh to create a VDM is non-trivial. 3) Sculpting brushes often involve sub-object structures, whereas most 3D generation methods can only generate full objects. Enabling users to accurately control the generation of sculpting brushes through text prompts in a semantically focused manner remains challenging.", + "bbox": [ + 88, + 123, + 482, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenges of brush generation, we propose Text2VDM, a novel optimization-based framework that generates diverse and controllable VDM brushes from text input. Our approach does not generate VDMs directly from a T2I model. Instead, we address VDM brush generation from a 3D perspective by applying score distillation with a pre-trained T2I model to guide mesh deformation. Our framework supports three ways to initialize a base mesh through a zero-valued, spike-pattern, or user-specified VDM for custom shape control. For mesh deformation, we formulate a Sobolev preconditioned optimization [35] to maintain mesh quality with intrinsic smoothness. We also provide optional region control using a mask of activated mesh deformation, helping users obtain the intended brush effects. The normal maps of the mesh are then rasterized by a differentiable renderer for brush optimization.", + "bbox": [ + 91, + 337, + 482, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We observed that the standard score distillation sampling (SDS) [37] can lead to semantic coupling when supervising the generation of sub-object level structures due to the associated semantics caused by the noisy gradients from the full object. For example, a generated deer's horn should not be a full deer's head, or a generated beard should not include a nose. A straightforward solution is to use negative prompts [18, 59] to exclude undesired semantics, but our experiments show that this semantics suppression approach is ineffective in decoupling semantics and leads to an unstable optimization process. Instead, we propose to enhance the semantics of part-related words by applying weighted blending to the tokens in the prompt. This results in semantically focused text embedding, directing toward a more precise target distribution while reducing noisy gradients during optimization.", + "bbox": [ + 89, + 580, + 482, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our experiments demonstrate that Text2VDM produces high-quality and diverse VDM brushes that can be directly integrated into mainstream modeling software, such as Blender [5] and ZBrush [12]. Compared to existing methods that directly generate full 3D models, our approach ad", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9db4fae690336e980920165d4dd6898a39d5f74a7765dd310c2606c9b3ce2c64.jpg", + "image_caption": [ + "(a) Vector Displacement", + "Figure 2. Difference between vector and standard displacement. The VDM enables full 3D vector displacement, while the height map only allows unidirectional standard displacement." + ], + "image_footnote": [], + "bbox": [ + 522, + 87, + 709, + 213 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/351cde89c79c3a88dab240689ff39d098aef6945e286f373d4bb3c14011f8a9a.jpg", + "image_caption": [ + "(b) Standard Displacement" + ], + "image_footnote": [], + "bbox": [ + 709, + 88, + 897, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dresses a different use case where brush-based user sculpting is desirable. This enables artists to interactively use a variety of brushes to sculpt diverse and expressive models from a plain shape.", + "bbox": [ + 511, + 303, + 903, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper makes the following contributions:", + "bbox": [ + 532, + 364, + 838, + 378 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We first introduce the task of text-to-VDM brush generation, which is challenging to tackle directly using current text-to-image and text-to-3D methods.", + "- We propose Text2VDM, a novel framework for text-to-VDM brush generation that is readily compatible with artists' workflow of 3D asset creation.", + "- We design a novel Semantic Enhancement SDS loss, which uses weighted blending to mitigate semantic coupling for sub-object structure generation." + ], + "bbox": [ + 513, + 378, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 527, + 653, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text to Local 3D Generation and Editing. With recent advances in diffusion models [41] and differentiable 3D representations [1, 33, 35, 44, 46], many methods for text-guided full 3D model generation have emerged [8, 11, 23, 25, 31, 38, 42]. Since 3D content creation is an iterative process that often requires user interaction, more attention has been directed toward localized 3D generation and editing. For example, 3D Highlighter [9] and 3D Paintbrush [10] use text as input, leveraging pre-trained CLIP models [40] or diffusion models [37] to supervise the optimization of neural networks for segmenting the regions of a 3D model that match the text description. Based on the information from these segmented regions, further editing of texture and geometry can be applied to the 3D model. Furthermore, SKED [32] and SketchDream [27] introduce sketches as an additional modality to assist in localized editing. To enable more precise control, FocalDreamer [24], MagicClay [4], and Tip-Editor [60] allow users to specify the editing location directly within the 3D space. These works rely on optimization-based methods to edit specific objects, often resulting in non-reusable editing outcomes. Additionally, each edit requires a lengthy optimization process, making interactivity difficult to achieve.", + "bbox": [ + 511, + 553, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16883", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f6e725640433d0a1fbaf19f0cdc17efc044134368f3d359e2927ed02a61c846d.jpg", + "image_caption": [ + "Figure 3. Overview of Text2VDM. Starting with a dense planar mesh constructed from a zero-valued VDM, users can initialize the mesh volume through a default spike-pattern VDM or a user-specified VDM. Given the text prompt and region mask, we propose the Semantic Enhancement SDS loss $\\mathcal{L}_{\\mathrm{SE}}$ to guide mesh deformations through a mesh Laplacian $L$ iteratively, achieving semantically focused generation of surface details or geometric structures. After generation, vertex displacements are baked into the final VDM." + ], + "image_footnote": [], + "bbox": [ + 109, + 93, + 890, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Priors for 3D Generation. Score distillation sampling (SDS) [37, 52] provides pixel-level guidance by seeking specific modes in a diffusion model, inspiring further research to improve optimization-based 3D generation [3, 26, 53, 54, 56]. Some studies focus on mitigating the \"Janus\" problem [2, 15], while others fine-tune diffusion models with multiview datasets to enhance 3D consistency [28, 45]. Recent research focuses on refining the design of SDS loss to achieve more precise guidance. For instance, Make-it-3D [50] introduces two-stage optimizations to improve textured appearance, while Fantasia3D [8] dynamically modifies the time-dependent weighting function within SDS computations. Additionally, several methods [18, 59] incorporate negative prompts as the conditional term to further refine the optimizations. Although diffusion priors have achieved promising results, their application in generating sub-object structures without global context as a reference is still challenging.", + "bbox": [ + 88, + 460, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Appearance and Geometric Brush Synthesis. The concept of brushes is very common in the creative process of digital artists, serving as a reusable local decorative unit. Appearance brushes focus on color representation and drawing styles in 2D space. With the development of generative models [13, 41], many works have explored the synthesis of procedural material [21, 22] for 3D object texturing and appearance brushes for interactive painting [16, 47], realistic artworks generation [29, 34, 61], and applying stylization [17, 19]. Unlike appearance brushes, geometric brushes focus on modifying geometry by moving the ver", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tices of a mesh in 3D space. VDM brushes, as an extension of standard geometric brushes, provide more complex geometric effects by utilizing VDMs. To the best of our knowledge, only a few techniques adopted the concepts of VDM for generation [43, 55] and geometric texture transfer [14]. Recently, concurrent work [57] explored using a single image as input and leveraged a diffusion model for generating multiview normal maps to guide VDM reconstruction. In comparison, our method is fundamentally different in tackling the semantic coupling problem arising from text guidance. This demonstrates that generating geometric brushes is a highly promising research direction.", + "bbox": [ + 511, + 460, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 647, + 650, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To generate VDM brushes compatible with mainstream modeling software, we begin by constructing a dense mesh from an initial VDM, as shown in Figure 3. We then apply score distillation with Stable Diffusion to guide high-quality mesh deformation formulated as a form of Sobolev preconditioned optimization [35]. To produce the intended sub-object level structure described in the text, we design a Semantic Enhancement SDS loss by applying weighted blending to the tokens in the prompt, effectively handling the issue of semantic coupling in SDS.", + "bbox": [ + 511, + 672, + 908, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Brush Initialization", + "text_level": 1, + "bbox": [ + 511, + 832, + 702, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We provide three methods to initialize a base mesh for brush generation via a zero-valued VDM, a spike-pattern VDM, or a user-specified VDM. A VDM is represented as a $512 \\times$", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16884", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "512 three-channel image, in which each channel stores the displacement in the X, Y, or Z direction, respectively. We first construct a planar grid mesh by creating two triangles for every $2 \\times 2$ pixels and then apply the displacement stored in the VDM to mesh vertices. The values in these three initial VDMs range from 0 to 1, in which 0 represents no displacement, and 1 corresponds to half of the mesh's edge length in the positive axis direction. Since users can apply sculpting brushes symmetrically, our initial VDM does not need to store any negative values.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our three methods for brush initialization facilitate the generation of diverse sculpting brush styles. The zero-valued VDM results in a planar mesh, which is our default setup when no control is provided. The spike-pattern VDM is suitable for generating protruding geometric structures, as it can effectively adjust the Laplacian term in Equation (2) to steer the gradient direction for mesh deformation. For better control of the brush's volume and direction, we also provide an interface for users to create custom VDMs, so the user-specified brush initialization can effectively guide mesh deformation toward the target structure.", + "bbox": [ + 89, + 242, + 482, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Brush Generation via Mesh Deformation", + "text_level": 1, + "bbox": [ + 89, + 417, + 441, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the vertices $v$ on the initialized base mesh, our method aims to learn a mesh deformation to the target brush shape. The vertex positions $\\hat{v}$ after mesh deformation can be expressed by:", + "bbox": [ + 89, + 441, + 483, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {v} = \\underset {v} {\\arg \\min } \\mathcal {L} _ {\\mathrm {S E}} \\left(\\mathcal {D} _ {c} (v), y\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 515, + 482, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ represents the camera setup in a differentiable renderer $\\mathcal{D}$ [20]. The loss function $\\mathcal{L}_{\\mathrm{SE}}$ receives the rendered normal image $\\mathcal{D}_c(v)$ and text input $y$ to evaluate the semantic guidance, which is detailed in Section 3.3. To accompany the external forces from $\\mathcal{L}_{\\mathrm{SE}}$ that drive the mesh deformation with intrinsic smoothness energies, we follow the framework of a Sobolev preconditioned gradient descent [35], where the base mesh is reparameterized by the mesh Laplacian $L$ :", + "bbox": [ + 89, + 550, + 483, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv ^ {*} = (I + \\lambda L) v. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 699, + 482, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This preconditioning involves solving a sparse linear system at every iteration, modifying the gradient descent update for each mesh deformation step to:", + "bbox": [ + 89, + 728, + 483, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv \\leftarrow v - \\eta (I + \\lambda L) ^ {- 1} \\frac {\\partial \\mathcal {L} _ {\\mathrm {S E}}}{\\partial v}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 784, + 482, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta$ is the learning rate, $I$ is the identity matrix, and $\\lambda$ is a hyperparameter to control the extent of gradient diffusion over the entire domain. We set $\\lambda = 15$ throughout our experiments to balance the global structure and fine details during mesh deformation.", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to directly adding a Laplacian regularization term to Equation (1), the preconditioning framework [35] is critical in achieving large mesh deformation while maintaining proper topology with reduced triangle flips (Figure 4). Several works [11, 51] adopt the strategy by Aignerman et al. [1], parameterizing deformation through Jacobian fields that capture the scaling and rotation of each triangle. Although this method effectively smooths vertex displacements, the local deformation represented in Jacobians accumulates, leading to global drifting for open-boundary meshes, making it challenging to bake the mesh as a brush.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, we provide an optional region mask to restrict mesh deformation to the user-defined region during optimization. It helps maintain zero values in unused VDM areas when generating geometric structures. For producing surface details, the region mask ensures that the brush effect meets user-customized requirements (Figure 9).", + "bbox": [ + 511, + 257, + 903, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By framing the VDM generation task as mesh deformation through preconditioned optimization, our framework preserves the favorable initial mesh topology and ensures control throughout the procedure. The resulting mesh preserves the original structure while incorporating rich local deformations, making it well-suited for baking as a brush.", + "bbox": [ + 511, + 348, + 903, + 439 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/33eac003bceaa84327a88f3d78016b0acc15ccd2d09d32395fc717d33840340c.jpg", + "image_caption": [ + "Figure 4. Effect of $\\lambda$ . (a) Low-quality results w/o mesh smoothness, (b) low-quality results w/ direct Laplacian regularization, (c) our adopted preconditioning scheme with $\\lambda = 15$ , and (d) results with over-smoothness caused by a larger $\\lambda$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 450, + 903, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semantic Enhancement Score Distillation", + "text_level": 1, + "bbox": [ + 511, + 685, + 867, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Current text-to-3D generation methods like DreamFusion [37] often optimize a 3D representation parameterized by $\\theta$ so that rendered images $\\mathbf{x} = g(\\theta)$ resemble 2D samples produced by a pre-trained T2I diffusion model for a given text prompt $y$ . $g$ functions as a differentiable renderer. The T2I diffusion model $\\phi$ predicts the sampled noise $\\epsilon_{\\phi}(\\mathbf{x}_t; y, t)$ of a rendered image $\\mathbf{x}_t$ at a noise level $t$ for the text input $y$ . To make rendered images follow the text-conditioned distribution in Stable Diffusion, the SDS loss updates $\\theta$ by estimating the gradient:", + "bbox": [ + 511, + 708, + 906, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon , c} \\left[ \\omega (t) \\left(\\epsilon_ {\\phi} (\\mathbf {x} _ {t}; y, t) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 869, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16885", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\omega (t)$ is a time-dependent weighting function.", + "bbox": [ + 89, + 90, + 434, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the SDS loss cannot effectively supervise sub-object structure generation due to the issue of semantic coupling in full objects. For example, when using the SDS loss to generate a tortoise shell, it also usually generates the tortoise's tail and head, causing semantic coupling (Figure 7). We believe that the issue of semantic coupling in SDS stems from the training data of Stable Diffusion, in which most images contain full objects rather than separate parts. Therefore, the semantics of full objects often appear in the target distribution conditioned on text only describing sub-object structures.", + "bbox": [ + 89, + 106, + 483, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A straightforward approach is using negative distributions via Classifier Score Distillation (CSD) [59] or Variational Score Distillation (VSD) [54] to suppress coupled semantics. CSD employs predefined negative prompts, resulting in more accurate negative distributions than those adaptively learned by VSD [59]:", + "bbox": [ + 89, + 272, + 483, + 363 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {C S D}} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon , c} [ (\\omega_ {\\mathrm {p o s}} \\cdot \\epsilon_ {\\phi} (\\mathbf {x} _ {t}; y, t) \\\\ \\left. - \\omega_ {\\text {n e g}} \\cdot \\epsilon_ {\\phi} \\left(\\mathbf {x} _ {t}; y _ {\\text {n e g}}, t\\right)\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} ], \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 369, + 482, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega_{\\mathrm{pos}}$ and $\\omega_{\\mathrm{neg}}$ denote different weights for positive and negative prompts. It requires two separate inferences with positive and negative prompts to obtain two distributions, which are then subtracted to suppress coupled semantics. However, our experiments show that CSD is ineffective in decoupling semantics because the negative prompt cannot accurately represent the undesirable coupled semantics in the positive prompt (Figure 7). This results in noisy guidance, making CSD less effective in decoupling semantics.", + "bbox": [ + 89, + 424, + 483, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike semantic suppression in CSD, we propose a semantic enhancement method to mitigate semantic coupling by enhancing the semantics of part-related words. This can lead to a more accurate and stable target distribution, as shown in Section 4.3. Our key design is to apply weighted blending to the tokens in the original prompt to obtain a semantically focused text embedding, which serves as stable guidance for the optimization process. We define the Semantic Enhancement SDS loss as:", + "bbox": [ + 89, + 561, + 483, + 695 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S E}} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon , c} \\left[ \\omega (t) \\left(\\epsilon_ {\\phi} ^ {*} (\\mathbf {x} _ {t}; y, t) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 698, + 482, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\phi}^{*}(\\cdot)$ uses a text embedding weighted by Compel [48]. Specifically, we assign a weight $s$ to each word in the prompt and compute the weighted embedding $e_{\\mathrm{w}}$ for each word by blending the original word embedding $e$ and the empty text embedding $e_0$ : $e_{\\mathrm{w}} = e_0 + s\\cdot (e - e_0)$ . By concatenating the weighted embedding of each word in sequence, we obtain the final semantically focused text embedding. In our experiments, we found that using $s = 1.1^2$ for words that require semantic enhancement can achieve stable optimization and effectively alleviate the issue of semantic coupling.", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/32ae9705644a6d017e563c1cd9860a71631b6ca5aae86fbd97317c96a65f39c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 90, + 586, + 135 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f27ac06eeb4621f121cd081556ec2d3db9348260927b25caf3219c37d711eb7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 88, + 661, + 138 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/60e68b185ce5dc93b63e20d6455ebdab6a35f9b0df83f89b8c11c0448c971353.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 138, + 666, + 199 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4282a32cf52a03466831d93baf340c1fc528ac46fea74ed26e4343bec8841c2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 88, + 730, + 138 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bb4a929c0f9e1a93864eb7e74df1588028f307f858761994e32d8d911d56f8ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 143, + 730, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/148f5918c3afdbff8627c7473e8b8896fb5cdc7a43e9740c06c29abdb8700778.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 199, + 591, + 248 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/339bd868ba412fd34b52a72ccc06e9d0577d2f292ff77a135a2fcbae511b9550.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 199, + 661, + 250 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c45a80201127c8dad8be377ad86c137d1bd5ee47d5fa3b0f19a5802e6dab096c.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 594, + 252, + 661, + 304 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3397ba1251c1e967e5002ef250713d758bb8c310c6bce1245e19255ad3fde7d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 199, + 733, + 250 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0e0228dd3321b74f046295a20df5aac87d0b2d6311f97b4313c69627cd38b667.jpg", + "image_caption": [ + "Text2Mesh" + ], + "image_footnote": [], + "bbox": [ + 669, + 252, + 733, + 303 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f8cd0eba056fd7575174ebb4f8c19fc7d066445320ce7a82121e4da9731a19db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 88, + 808, + 138 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9d3fb6c910fab193f57055bfdc9e63c46cc09062e59eb82b9c253272845535f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 741, + 143, + 805, + 194 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/19856accc9267105194cd89ab83bafbe8e6905fab5aba1034dac29276dee11c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 199, + 807, + 250 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9caa1c0362d06a5b26b6e71cb1fabf4b7de490d3442f9d114c3f9945b0a6cce0.jpg", + "image_caption": [ + "Paint-it" + ], + "image_footnote": [], + "bbox": [ + 746, + 253, + 807, + 303 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a4f63aedf8f1c54357004c942e8d84120957007958978fcf502a615f8724a107.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 88, + 877, + 138 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/527e8c7a94b256f4a08d6340dd3a708b83072e779451892ce944bd581d9655e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 143, + 880, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1e19b63b09904c996cbb857f4e270e37afc7fe4b55a0ec45819238b6082dc7d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 200, + 882, + 253 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/46835acade887753b98a215374ec719ec1b5b9f6650e1d078abbf4ba2e5a7a3e.jpg", + "image_caption": [ + "TextDeformer", + "Figure 5. Qualitative comparisons of generated brushes for surface details. Our method captures geometry details guided by text, effectively preserving the surface structure and avoiding mesh distortion." + ], + "image_footnote": [], + "bbox": [ + 818, + 257, + 888, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 404, + 645, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conducted experiments to evaluate the various capabilities of Text2VDM both quantitatively and qualitatively for text-to-VDM brush generation. We then present an ablation study that validates the significance of our key insight into Semantic Enhancement SDS, as well as the effect of the region mask and VDM initialization.", + "bbox": [ + 511, + 430, + 905, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 530, + 723, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To the best of our knowledge, Text2VDM is the first framework to generate VDM brushes from text. We adapted three existing methods for comparison and classified them into two categories. The first category includes Text2Mesh [31] and TextDeformer [11], which generate a brush mesh through text-guided mesh deformation on a planar mesh, following a process similar to ours. For the second category, we opt to directly generate VDM via Paint-it [58]. Notably, this method originally uses SDS to optimize a UNet for generating PBR textures. We reframed it to suit our VDM brush generation task, modifying it to generate VDM through SDS optimization of the UNet. In geometric structures generation experiment (Figure 6), all methods are compared fairly, with the same non-zero VDM initialization and mask applied to each prompt (see Appendix B.2). For surface details (Figure 5), all methods start with a zero-valued VDM and no masks to ensure a fair comparison.", + "bbox": [ + 511, + 551, + 906, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared to other methods, Text2VDM can generate better-quality VDM brushes. Text2Mesh applies displacement to each vertex along normal directions, resulting in limited mesh deformation. TextDeformer indicates the accumulation of local deformations in the Jacobians, which results in global mesh drift, making it challenging to bake", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16886", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2724a6d7f4554fc538ecefda3d302064971f545c7a888bed05ee89898092d0f5.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons of generated brushes for geometric structures. Our method accurately presents key geometric features described by text, facilitating downstream applications in modeling software." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 890, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "these meshes into VDM. Reframed Paint-it VDM generation is equivalent to optimizing the three-axis displacement of each vertex on the mesh with SDS. Although the UNet reduces noise from the SDS [58], smooth regularization is still required to ensure mesh quality, which makes achieving high-quality mesh generation quite challenging.", + "bbox": [ + 89, + 550, + 482, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 89, + 650, + 313, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We quantitatively evaluated our framework regarding generation consistency with text input and mesh quality. We used 40 distinctive text prompts for VDM generation.", + "bbox": [ + 89, + 674, + 482, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generation Consistency with Text. We initially assessed the relevance of the generated results to the text descriptions [40]. 12 different views were rendered for average scores respectively, as presented in Table 1. Our approach achieves the highest scores compared to baseline methods.", + "bbox": [ + 89, + 719, + 482, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mesh Quality. We evaluated mesh quality by examining self-intersection. Paint-it and Text2Mesh, which utilize direct vertex displacement, often converge to a local minimum and disregard the mesh triangulation. While TextDeformer exhibits the lowest self-intersection, its tendency to produce over-smoothed results frequently results in losing object features described in text prompts.", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f065ac1c7781d4bd8f609223644e1fd5d591b0050a087058e65fb3f2d7758953.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation of state-of-the-art methods. The geometry CLIP score is calculated on shaded images with uniform albedo colors [39], and self-intersection is quantified as the ratio of self-intersected mesh faces to the total number of faces." + ], + "table_footnote": [], + "table_body": "
Geometry CLIP Score ↑Mesh Self-Intersection ↓
Paint-it0.237519.42%
Text2Mesh0.24977.18%
TextDeformer0.24770.04%
Ours0.25560.77%
", + "bbox": [ + 514, + 614, + 906, + 680 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f4516ac7e4878e636a83ec549597fad33587ebebb560246d63805b505ce1283b.jpg", + "table_caption": [ + "Table 2. User evaluation of generated VDMs." + ], + "table_footnote": [], + "table_body": "
User Preference ↑Geometry QualityConsistency with Text
Paint-it3.1%1.7%
Text2Mesh18.3%27.3%
TextDeformer3.3%3.4%
Ours75.3%67.6%
", + "bbox": [ + 524, + 719, + 893, + 784 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "User Study. We further conducted a user study to evaluate the effectiveness and expressiveness of our method. A Google Form was utilized to assess 1) geometry quality and 2) consistency with text. We recruited 32 participants, of whom 14 are graduate students majoring in media arts, and 18 are company employees specializing in AI content gen", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16887", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3b7ed331c8cb6a0a13ab499cf1eb6de03182366edcc4db96098b0f16658fb03.jpg", + "image_caption": [ + "Figure 7. Effect of Semantic Enhancement SDS. Our method effectively mitigates semantic coupling issues in SDS, such as generating the tortoise's tail and head or the deer's ear and mouth, by providing more focused semantic guidance. In contrast, the semantic suppression method used in CSD and the other semantic enhancement approach proposed by Attend-and-Excite both lead to an unstable optimization." + ], + "image_footnote": [], + "bbox": [ + 112, + 85, + 883, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "eration. The participants were instructed to choose the preferred renderings of VDM from different methods in randomized order, as shown in Table 2. The results show participants preferred our method by a significant margin.", + "bbox": [ + 88, + 395, + 480, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 467, + 243, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Semantic Enhancement SDS. Figure 7 compares the results generated by the original SDS [37], our Semantic Enhancement SDS with Compel, SDS with another semantic enhancement method in Attend-and-Excite (AaE) [7], and CSD [59] with two different annealed weights for the negative prompts: \"tortoise tail, tortoise head\" and \"deer's ear, deer's mouth.\" We also qualitatively compare these methods by visualizing their performance on semantic decoupling using the same Stable Diffusion for 2D image generation (Figure 8). Compared to semantic suppression using negative prompts and semantic enhancement by AaE, our design of introducing a semantically focused text embedding via Compel to SDS is most effective in resolving the issue of semantic coupling.", + "bbox": [ + 89, + 491, + 483, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key Insight. As discussed in Section 3.3, SDS can result in semantic coupling when generating sub-object structures, leading to artifacts like the tortoise's tail and head or the deer's ear and deer's mouth. We observed that the semantics represented by negative prompts are also coupled. Therefore, meaningless semantics can emerge when applying semantic suppression in SDS, which leads to unstable optimization. Increasing the weight of negative prompts further reduces the overall quality of generated results. AaE enhances the cross-attention map of specific tokens by continuously updating the latent code. However, AaE is not suitable for the SDS framework because the latent code is affected by different camera poses and random noise at each", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "iteration. This temporal randomness undermines the continuity of the latent code, resulting in unstable optimization and subpar results. In contrast, our method uses Compel to enhance semantics and achieves more effective semantic decoupling. Moreover, the semantically focused text embedding produced by Compel is independent of temporal variation. These properties help mitigate semantic coupling during SDS optimization, leading to high-quality sub-object geometric structures. Meanwhile, our method can also enhance the semantics in the text prompt for surface detail generation (see Appendix A.2).", + "bbox": [ + 511, + 395, + 906, + 561 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c51a0f93af5061df6dec77d828d0d07063b2bb03f89e2f86c3d9e4accc8885f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 574, + 906, + 652 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9790439975bac725358d87ac5566b40000efe57edb8e33224a382539f041ce60.jpg", + "image_caption": [ + "Figure 8. Visualization of semantic decoupling performance. The images generated by the T2I model qualitatively visualize sub-object semantics. (a) Coupled semantics in positive and negative prompts lead to meaningless distributions when subtracted. (b) Enhanced semantics by Compel achieves stable decoupling. (c) AaE can only ensure that the enhanced semantics are preserved, but cannot decouple negative ones effectively." + ], + "image_footnote": [], + "bbox": [ + 517, + 654, + 903, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Region Mask. Figure 9 demonstrates two region masks and their control over surface detail generation given a text prompt. Without a region mask, the results can still", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16888", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dac4f56f276654eda9dc7b45b186243d881d04b8ac6b1ba0811cd831b44ad150.jpg", + "image_caption": [ + "Figure 9. Effect of region mask. Region masks can effectively control the pattern of surface details based on text input." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 486, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "match the text prompt but lack a specific pattern. By using a region mask and maintaining an activation ratio of $1/2$ throughout the total iterations as a warm-up stage, we achieve a reasonable tradeoff between mask-result alignment and generation diversity (see Appendix A.3).", + "bbox": [ + 89, + 251, + 483, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of VDM Initialization. Our method demonstrates that user-specified VDMs can effectively control the volume and direction of generated geometric structures. Figure 10 shows that the results are high-quality and match the text descriptions well, such as the elf ear and pauldron. As this initializes the Laplacian term and steers the gradient flow in geometric structure generation, users can easily specify an initial VDM or choose a VDM template provided in our framework to generate expressive results.", + "bbox": [ + 89, + 327, + 483, + 463 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0296df2d3b99aa17fa1d0ba8d6f95e97acdf9cb17cfb7bfd10b1935845e1a98e.jpg", + "image_caption": [ + "Figure 10. Effect of VDM initialization. User-specified VDMs can help achieve the intended final effect of geometric structures by initializing the brush's volume and direction." + ], + "image_footnote": [], + "bbox": [ + 93, + 473, + 486, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Applications", + "text_level": 1, + "bbox": [ + 89, + 756, + 223, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Once various VDMs are generated, users can use these brushes to meet diverse creative needs in modeling software. For example, they can apply VDM brushes for mesh stylization and a real-time iterative modeling process.", + "bbox": [ + 89, + 780, + 482, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Local-to-Global Mesh Stylization. Although mesh stylization is a complex task even for professional artists, combining different surface details allows users to achieve stylization quickly (see Appendix C.3).", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/208b43d4fe5f555dc761423972359ec7717b635dd0351bd595818662e78c59c8.jpg", + "image_caption": [ + "Figure 11. Coarse to fine interactive modeling. By combining geometric structures brushes and surface details brushes for iterative sculpting in modeling software, users can rapidly create an expressive model from a plain shape (top left)." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 908, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Coarse-to-Fine Interactive Modeling. Unlike previous methods [4, 60] that require a lengthy optimization process for each edit and result in non-reusable outcomes, our generated VDM brushes can be directly used in modeling software. This enables users to apply the generated brushes easily and interactively (Figure 11).", + "bbox": [ + 511, + 347, + 906, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 468, + 633, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented Text2VDM, a novel framework for VDM brush generation from text. A VDM is a non-natural 2D image where each pixel stores a 3D displacement vector, making it challenging for existing T2I models to generate. Therefore, we treat VDM generation as diffusion-guided mesh deformation formulated as a form of Sobolev preconditioned optimization. To mitigate semantic coupling issues in SDS, we propose using weighted blending for prompt tokens, achieving high-quality brush generation. Moreover, we introduce two control methods, i.e., region and shape control, to meet customized requirements. The generated VDMs are directly compatible with mainstream modeling software, enabling various applications such as mesh stylization and interactive modeling.", + "bbox": [ + 511, + 501, + 906, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. While our framework can generate high-quality VDM brushes, they may encounter multiview inconsistencies, a common issue introduced by SDS. View-consistent diffusion models like MV2MV [6] may be helpful to further handle this. Additionally, errors of the T2I models may guide wrong 3D generation, such as the wrong number of fingers in the generated skeleton hand. VDMs demonstrate that complex 3D models can be efficiently created using diverse reusable sculpting brushes. Future exploration of 3D generation through the assembly of modular components with similar design principles holds promising research value.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16889", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the Guangdong Provincial Key Lab of Integrated Communication, Sensing and Computation for Ubiquitous Internet of Things #2023B1212010007.", + "bbox": [ + 89, + 114, + 483, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 203 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Noam Aigerman, Kunal Gupta, Vladimir G. Kim, Siddhartha Chaudhuri, Jun Saito, and Thibault Groueix. Neural Jacobian Fields: Learning Intrinsic Mappings of Arbitrary Meshes. ACM Trans. Graph., 41(4), 2022. 2, 4", + "[2] Thiemo Alldieck, Nikos Kolotouros, and Cristian Sminchisescu. Score Distillation Sampling with Learned Manifold Corrective. European Conference on Computer Vision, 2024. 3", + "[3] Mohammadreza Armandpour, Ali Sadeghian, Huangjie Zheng, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the Negative Prompt Algorithm: Transform 2D Diffusion into 3D, alleviate Janus problem and Beyond. arXiv preprint 2304.04968, 2023. 3", + "[4] Amir Barda, Vladimir G. Kim, Noam Aigerman, Amit H. Bermano, and Thibault Groueix. MagicClay: Sculpting Meshes with Generative Neural Fields. SIGGRAPH Asia (Conference track), 2024. 2, 8", + "[5] Blender Foundation. Blender. https://www.blender.org, 2025. Accessed Mar 5, 2025. 2", + "[6] Youcheng Cai, Runshi Li, and Ligang Liu. MV2MV: Multi-View Image Translation via View-Consistent Diffusion Models. ACM Trans. Graph., 2024. 8", + "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-Ecite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models. 42 (4), 2023. 7", + "[8] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3D: Disentangling Geometry and Appearance for High-quality Text-to-3D Content Creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 2, 3", + "[9] Dale Decatur, Itai Lang, and Rana Hanocka. 3D Highlighter: Localizing Regions on 3D Shapes via Text Descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20930-20939, 2023. 2", + "[10] Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3D Paintbrush: Local Stylization of 3D Shapes with Cascaded Score Distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4473-4483, 2024. 2", + "[11] William Gao, Noam Aigerman, Thibault Groueix, Vova Kim, and Rana Hanocka. TextDeformer: Geometry Manipulation Using Text Guidance. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2, 4, 5", + "[12] Maxon Computer GMBH. ZBrush. https://www.maxon.net/zbrush, 1999. Accessed Oct 18, 2024. 2", + "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and" + ], + "bbox": [ + 93, + 212, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yoshua Bengio. Generative Adversarial Nets. Advances in Neural Information Processing Systems, 27, 2014. 3", + "[14] Amir Hertz, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Deep Geometric Texture Synthesis. ACM Trans. Graph., 39(4), 2020. 3", + "[15] Susung Hong, Donghoon Ahn, and Seungryong Kim. Debiasing Scores and Prompts of 2D Diffusion for View-consistent Text-to-3D Generation. In Neural Information Processing Systems, 2023. 3", + "[16] Anita Hu, Nishkrit Desai, Hassan Abu Alhaija, Seung Wook Kim, and Maria Shugrina. Diffusion Texture Painting. New York, NY, USA, 2024. Association for Computing Machinery. 3", + "[17] Teng Hu, Ran Yi, Haokun Zhu, Liang Liu, Jinlong Peng, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Stroke-based Neural Painting and Stylization with Dynamically Predicted Painting Region. New York, NY, USA, 2023. Association for Computing Machinery. 3", + "[18] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free Score Distillation. In The Twelfth International Conference on Learning Representations, 2024. 2, 3", + "[19] Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Björn Ommer. Rethinking Style Transfer: From Pixels to Parameterized Brushstrokes. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12191-12200, 2021. 3", + "[20] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular Primitives for High-Performance Differentiable Rendering. ACM Transactions on Graphics, 39(6), 2020. 4", + "[21] Beichen Li, Liang Shi, and Wojciech Matusik. End-to-end procedural material capture with proxy-free mixed-integer optimization. ACM Transactions on Graphics (TOG), 42(4): 1-15, 2023. 3", + "[22] Beichen Li, Yiwei Hu, Paul Guerrero, Miloš Hasan, Liang Shi, Valentin Deschaintre, and Wojciech Matusik. Procedural material generation with reinforcement learning. ACM Transactions on Graphics (TOG), 43(6):1-14, 2024. 3", + "[23] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast Text-to-3D with Sparse-View Generation and Large Reconstruction Model. arXiv preprint 2311.06214, 2023. 2", + "[24] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. FocalDreamer: Text-Driven 3D Editing via Focal-Fusion Assembly. Proceedings of the AAAI Conference on Artificial Intelligence, 38:3279–3287, 2024. 2", + "[25] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. LucidDreamer: Towards High-Fidelity Text-to-3D Generation via Interval Score Matching. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6517-6526, 2023. 2", + "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-Resolution Text-to-3D Content Creation. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16890", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300-309, 2023. 3", + "[27] Feng-Lin Liu, Hongbo Fu, Yu-Kun Lai, and Lin Gao. SketchDream: Sketch-based Text-to-3D Generation and Editing. ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH 2024), 43(4), 2024. 2", + "[28] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot One Image to 3D Object. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9264–9275, 2023. 3", + "[29] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Ruifeng Deng, Xin Li, Errui Ding, and Hao Wang. Paint Transformer: Feed Forward Neural Painting with Stroke Prediction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6578-6587, 2021. 3", + "[30] Xiaoxiao Long, Yuanchen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3D: Single Image to 3D Using Cross-Domain Diffusion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9970-9980, 2023. 2", + "[31] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2Mesh: Text-Driven Neural Stylization for Meshes. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13482–13492, 2022. 2, 5", + "[32] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2", + "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 2", + "[34] Haoran Mo, Edgar Simo-Serra, Chengying Gao, Changqing Zou, and Ruomei Wang. General virtual sketching framework for vector line art. ACM Transactions on Graphics (TOG), 40:1-14, 2021. 3", + "[35] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large Steps in Inverse Rendering of Geometry. ACM Trans. Graph., 40(6), 2021. 2, 3, 4", + "[36] Openai. DALLE. https://openai.com/index/dall-e-3/, 2024. Accessed Oct 22, 2024. 2", + "[37] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. In The Eleventh International Conference on Learning Representations, 2022. 2, 3, 4, 7", + "[38] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. RichDreamer: A Generalizable Normal-Depth Diffusion Model for Detail Richness in Text-to-3D. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914–9925, 2023. 2", + "[39] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bo, and Xiaoguang Han. RichDreamer: A Generalizable Normal-Depth Diffusion Model for Detail Richness in Text-to-3D. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914–9925, 2024. 6", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In International Conference on Machine Learning, 2021. 2, 6", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022. 2, 3", + "[42] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards Zero-shot Text-to-shape Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18603-18613, 2022. 2", + "[43] Artem Sevastopolsky, Philip-William Grassal, Simon Giebenhain, ShahRukh Athar, Luisa Verdoliva, and Matthias Niessner. Headcraft: Modeling high-detail shape variations for animated 3dmms, 2023. 3", + "[44] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep Marching Tetrahedra: a Hybrid Representation for High-resolution 3D Shape Synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2", + "[45] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view Diffusion for 3D Generation. In The Twelfth International Conference on Learning Representations, 2024. 3", + "[46] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3D Neural Field Generation Using Triplane Diffusion. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2023. 2", + "[47] Maria Shugrina, Chin-Ying Li, and Sanja Fidler. Neural Brushstroke Engine: Learning a Latent Style Space of Interactive Drawing Tools. ACM Trans. Graph., 41(6), 2022. 3", + "[48] Damian Stewart. Compel: A Text Prompt Weighting and Blending Library. https://github.com/damian0815/compel, 2023. Accessed Oct 20, 2024. 5", + "[49] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative Gaussian Splitting for Efficient 3D Content Creation, 2023. 2", + "[50] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-It-3D: High-fidelity 3D Creation from A Single Image with Diffusion Prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22819–22829, 2023. 2, 3", + "[51] Duotun Wang, Hengyu Meng, Zeyu Cai, Zhijing Shao, Qianxi Liu, Lin Wang, Mingming Fan, Xiaohang Zhan, and" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16891", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zeyu Wang. HeadEvolver: Text to Head Avatars via Expressive and Attribute-Preserving Mesh Deformation. International Conference on 3D Vision, 2025. 4", + "[52] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12619-12629, 2023. 3", + "[53] Peihao Wang, Dejia Xu, Zhiwen Fan, Dilin Wang, Sreyas Mohan, Forrest N. Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu, Zhangyang Wang, and Vikas Chandra. Taming Mode Collapse in Score Distillation for Text-to-3D Generation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[54] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and Diverse Text-to-3D Generation with Variational Score Distillation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 3, 5", + "[55] Xingguang Yan, Han-Hung Lee, Ziyu Wan, and Angel X. Chang. An Object is Worth 64x64 Pixels: Generating 3D Object via Image Diffusion, 2024. 3", + "[56] Xiaofeng Yang, Yiwen Chen, Cheng Chen, Chi Zhang, Yi Xu, Xulei Yang, Fayao Liu, and Guosheng Lin. Learn to Optimize Denoising Scores for 3D Generation: A Unified and Improved Diffusion Prior on NeRF and 3D Gaussian Splatting. arXiv preprint 2312.04820, 2023. 3", + "[57] Yuezhi Yang, Qimin Chen, Vladimir Kim, Siddhartha Chaudhuri, Qixing Huang, and Zhiqin Chen. GenVDM: Generating Vector Displacement Maps From a Single Image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3", + "[58] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-Texture Synthesis via Deep Convolutional Texture Map Optimization and Physically-Based Rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5, 6", + "[59] Xin Yu, Yuanchen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and Xiaojuan Qi. Text-to-3D with Classifier Score Distillation. arXiv preprint 2310.19415, 2023. 2, 3, 5, 7", + "[60] Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. TIP-Editor: An Accurate 3D Editor Following Both Text-Prompts and Image-Prompts. ACM Trans. Graph., 43(4), 2024. 2, 8", + "[61] Zhengxia Zou, Tianyang Shi, Shuang Qiu, Yi Yuan, and Zhenwei Shi. Stylized Neural Painting. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15684-15693, 2021. 3" + ], + "bbox": [ + 91, + 90, + 480, + 810 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16892", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/011b91bd-0c46-4896-9ef0-7e4bb90ad7cb_model.json b/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/011b91bd-0c46-4896-9ef0-7e4bb90ad7cb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c3282470447eaf472370407e41b4d3d80ca979b7 --- /dev/null +++ b/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/011b91bd-0c46-4896-9ef0-7e4bb90ad7cb_model.json @@ -0,0 +1,2389 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.263, + 0.131, + 0.738, + 0.177 + ], + "angle": 0, + "content": "Text2VDM: Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.203, + 0.868, + 0.277 + ], + "angle": 0, + "content": "Hengyu Meng\\(^{1}\\) Duotun Wang\\(^{1}\\) Zhijing Shao\\(^{1}\\) Ligang Liu\\(^{2}\\) Zeyu Wang\\(^{1,3*}\\) \n\\(^{1}\\)The Hong Kong University of Science and Technology (Guangzhou) \n\\(^{2}\\)University of Science and Technology of China \n\\(^{3}\\)The Hong Kong University of Science and Technology" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.31, + 0.896, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.551, + 0.908, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Example VDM brushes generated by Text2VDM and sculpted models in Blender. Text2VDM can produce high-quality brushes for surface details (top row) and geometric structures (bottom row) from text input. Users can rapidly create an expressive model from a plain shape by directly applying these brushes in Blender. Yellow underlined text highlights semantics enhanced by our framework." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.607, + 0.328, + 0.623 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.638, + 0.486, + 0.881 + ], + "angle": 0, + "content": "Professional 3D asset creation often requires diverse sculpting brushes to add surface details and geometric structures. Despite recent progress in 3D generation, producing reusable sculpting brushes compatible with artists' workflows remains an open and challenging problem. These sculpting brushes are typically represented as vector displacement maps (VDMs), which existing models cannot easily generate compared to natural images. This paper presents Text2VDM, a novel framework for text-to-VDM brush generation through the deformation of a dense planar mesh guided by score distillation sampling (SDS). The original SDS loss is designed for generating full objects and struggles with generating desirable sub-object structures from scratch in brush generation. We refer to this issue as semantic coupling, which we address by introducing weighted blending of prompt tokens to SDS, resulting in a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.909, + 0.714 + ], + "angle": 0, + "content": "more accurate target distribution and semantic guidance. Experiments demonstrate that Text2VDM can generate diverse, high-quality VDM brushes for sculpting surface details and geometric structures. Our generated brushes can be seamlessly integrated into mainstream modeling software, enabling various applications such as mesh stylization and real-time interactive modeling." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.752, + 0.646, + 0.767 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Sculpting brushes are essential tools in 3D asset creation, as artists often require a variety of brushes to create surface details and geometric structures. In modeling software, 3D sculpting brushes are typically defined as vector displacement maps (VDMs). A VDM is a 2D image where each pixel stores a 3D displacement vector. Through these vectors, VDM brushes can create complex surface details, such as cracks and wood grain, or generate geometric structures" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.242, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16882" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "like ears and horns. This allows artists to apply the same geometric pattern iteratively while sculpting." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.124, + 0.483, + 0.336 + ], + "angle": 0, + "content": "Despite significant advances in text-to-image (T2I) [36, 41] and text-to-3D generation [23, 30, 37, 49, 50], existing methods are unsuitable for creating VDM brushes. We summarize the challenges as follows: 1) Since VDMs are not natural images (Figure 2), it is difficult for existing T2I models to generate them directly. 2) From a 3D perspective, a VDM represents mesh deformation through per-vertex displacement vectors from a dense planar mesh. Mapping any generated mesh to a dense planar mesh to create a VDM is non-trivial. 3) Sculpting brushes often involve sub-object structures, whereas most 3D generation methods can only generate full objects. Enabling users to accurately control the generation of sculpting brushes through text prompts in a semantically focused manner remains challenging." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.338, + 0.483, + 0.579 + ], + "angle": 0, + "content": "To address the challenges of brush generation, we propose Text2VDM, a novel optimization-based framework that generates diverse and controllable VDM brushes from text input. Our approach does not generate VDMs directly from a T2I model. Instead, we address VDM brush generation from a 3D perspective by applying score distillation with a pre-trained T2I model to guide mesh deformation. Our framework supports three ways to initialize a base mesh through a zero-valued, spike-pattern, or user-specified VDM for custom shape control. For mesh deformation, we formulate a Sobolev preconditioned optimization [35] to maintain mesh quality with intrinsic smoothness. We also provide optional region control using a mask of activated mesh deformation, helping users obtain the intended brush effects. The normal maps of the mesh are then rasterized by a differentiable renderer for brush optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.824 + ], + "angle": 0, + "content": "We observed that the standard score distillation sampling (SDS) [37] can lead to semantic coupling when supervising the generation of sub-object level structures due to the associated semantics caused by the noisy gradients from the full object. For example, a generated deer's horn should not be a full deer's head, or a generated beard should not include a nose. A straightforward solution is to use negative prompts [18, 59] to exclude undesired semantics, but our experiments show that this semantics suppression approach is ineffective in decoupling semantics and leads to an unstable optimization process. Instead, we propose to enhance the semantics of part-related words by applying weighted blending to the tokens in the prompt. This results in semantically focused text embedding, directing toward a more precise target distribution while reducing noisy gradients during optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our experiments demonstrate that Text2VDM produces high-quality and diverse VDM brushes that can be directly integrated into mainstream modeling software, such as Blender [5] and ZBrush [12]. Compared to existing methods that directly generate full 3D models, our approach ad" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.088, + 0.71, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.214, + 0.688, + 0.225 + ], + "angle": 0, + "content": "(a) Vector Displacement" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.898, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.753, + 0.214, + 0.889, + 0.225 + ], + "angle": 0, + "content": "(b) Standard Displacement" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.236, + 0.907, + 0.278 + ], + "angle": 0, + "content": "Figure 2. Difference between vector and standard displacement. The VDM enables full 3D vector displacement, while the height map only allows unidirectional standard displacement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.905, + 0.364 + ], + "angle": 0, + "content": "dresses a different use case where brush-based user sculpting is desirable. This enables artists to interactively use a variety of brushes to sculpt diverse and expressive models from a plain shape." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.365, + 0.839, + 0.379 + ], + "angle": 0, + "content": "This paper makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.38, + 0.905, + 0.424 + ], + "angle": 0, + "content": "- We first introduce the task of text-to-VDM brush generation, which is challenging to tackle directly using current text-to-image and text-to-3D methods." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.425, + 0.905, + 0.469 + ], + "angle": 0, + "content": "- We propose Text2VDM, a novel framework for text-to-VDM brush generation that is readily compatible with artists' workflow of 3D asset creation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.47, + 0.905, + 0.515 + ], + "angle": 0, + "content": "- We design a novel Semantic Enhancement SDS loss, which uses weighted blending to mitigate semantic coupling for sub-object structure generation." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.38, + 0.905, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.528, + 0.655, + 0.544 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Text to Local 3D Generation and Editing. With recent advances in diffusion models [41] and differentiable 3D representations [1, 33, 35, 44, 46], many methods for text-guided full 3D model generation have emerged [8, 11, 23, 25, 31, 38, 42]. Since 3D content creation is an iterative process that often requires user interaction, more attention has been directed toward localized 3D generation and editing. For example, 3D Highlighter [9] and 3D Paintbrush [10] use text as input, leveraging pre-trained CLIP models [40] or diffusion models [37] to supervise the optimization of neural networks for segmenting the regions of a 3D model that match the text description. Based on the information from these segmented regions, further editing of texture and geometry can be applied to the 3D model. Furthermore, SKED [32] and SketchDream [27] introduce sketches as an additional modality to assist in localized editing. To enable more precise control, FocalDreamer [24], MagicClay [4], and Tip-Editor [60] allow users to specify the editing location directly within the 3D space. These works rely on optimization-based methods to edit specific objects, often resulting in non-reusable editing outcomes. Additionally, each edit requires a lengthy optimization process, making interactivity difficult to achieve." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16883" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.094, + 0.891, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.378, + 0.908, + 0.437 + ], + "angle": 0, + "content": "Figure 3. Overview of Text2VDM. Starting with a dense planar mesh constructed from a zero-valued VDM, users can initialize the mesh volume through a default spike-pattern VDM or a user-specified VDM. Given the text prompt and region mask, we propose the Semantic Enhancement SDS loss \\(\\mathcal{L}_{\\mathrm{SE}}\\) to guide mesh deformations through a mesh Laplacian \\(L\\) iteratively, achieving semantically focused generation of surface details or geometric structures. After generation, vertex displacements are baked into the final VDM." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Diffusion Priors for 3D Generation. Score distillation sampling (SDS) [37, 52] provides pixel-level guidance by seeking specific modes in a diffusion model, inspiring further research to improve optimization-based 3D generation [3, 26, 53, 54, 56]. Some studies focus on mitigating the \"Janus\" problem [2, 15], while others fine-tune diffusion models with multiview datasets to enhance 3D consistency [28, 45]. Recent research focuses on refining the design of SDS loss to achieve more precise guidance. For instance, Make-it-3D [50] introduces two-stage optimizations to improve textured appearance, while Fantasia3D [8] dynamically modifies the time-dependent weighting function within SDS computations. Additionally, several methods [18, 59] incorporate negative prompts as the conditional term to further refine the optimizations. Although diffusion priors have achieved promising results, their application in generating sub-object structures without global context as a reference is still challenging." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Appearance and Geometric Brush Synthesis. The concept of brushes is very common in the creative process of digital artists, serving as a reusable local decorative unit. Appearance brushes focus on color representation and drawing styles in 2D space. With the development of generative models [13, 41], many works have explored the synthesis of procedural material [21, 22] for 3D object texturing and appearance brushes for interactive painting [16, 47], realistic artworks generation [29, 34, 61], and applying stylization [17, 19]. Unlike appearance brushes, geometric brushes focus on modifying geometry by moving the ver" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.644 + ], + "angle": 0, + "content": "tices of a mesh in 3D space. VDM brushes, as an extension of standard geometric brushes, provide more complex geometric effects by utilizing VDMs. To the best of our knowledge, only a few techniques adopted the concepts of VDM for generation [43, 55] and geometric texture transfer [14]. Recently, concurrent work [57] explored using a single image as input and leveraged a diffusion model for generating multiview normal maps to guide VDM reconstruction. In comparison, our method is fundamentally different in tackling the semantic coupling problem arising from text guidance. This demonstrates that generating geometric brushes is a highly promising research direction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.651, + 0.667 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.825 + ], + "angle": 0, + "content": "To generate VDM brushes compatible with mainstream modeling software, we begin by constructing a dense mesh from an initial VDM, as shown in Figure 3. We then apply score distillation with Stable Diffusion to guide high-quality mesh deformation formulated as a form of Sobolev preconditioned optimization [35]. To produce the intended sub-object level structure described in the text, we design a Semantic Enhancement SDS loss by applying weighted blending to the tokens in the prompt, effectively handling the issue of semantic coupling in SDS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.704, + 0.849 + ], + "angle": 0, + "content": "3.1. Brush Initialization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We provide three methods to initialize a base mesh for brush generation via a zero-valued VDM, a spike-pattern VDM, or a user-specified VDM. A VDM is represented as a \\(512 \\times\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16884" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "512 three-channel image, in which each channel stores the displacement in the X, Y, or Z direction, respectively. We first construct a planar grid mesh by creating two triangles for every \\( 2 \\times 2 \\) pixels and then apply the displacement stored in the VDM to mesh vertices. The values in these three initial VDMs range from 0 to 1, in which 0 represents no displacement, and 1 corresponds to half of the mesh's edge length in the positive axis direction. Since users can apply sculpting brushes symmetrically, our initial VDM does not need to store any negative values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.41 + ], + "angle": 0, + "content": "Our three methods for brush initialization facilitate the generation of diverse sculpting brush styles. The zero-valued VDM results in a planar mesh, which is our default setup when no control is provided. The spike-pattern VDM is suitable for generating protruding geometric structures, as it can effectively adjust the Laplacian term in Equation (2) to steer the gradient direction for mesh deformation. For better control of the brush's volume and direction, we also provide an interface for users to create custom VDMs, so the user-specified brush initialization can effectively guide mesh deformation toward the target structure." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.419, + 0.442, + 0.435 + ], + "angle": 0, + "content": "3.2. Brush Generation via Mesh Deformation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.484, + 0.504 + ], + "angle": 0, + "content": "Given the vertices \\( v \\) on the initialized base mesh, our method aims to learn a mesh deformation to the target brush shape. The vertex positions \\( \\hat{v} \\) after mesh deformation can be expressed by:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.516, + 0.483, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\hat {v} = \\underset {v} {\\arg \\min } \\mathcal {L} _ {\\mathrm {S E}} \\left(\\mathcal {D} _ {c} (v), y\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.484, + 0.687 + ], + "angle": 0, + "content": "where \\(c\\) represents the camera setup in a differentiable renderer \\(\\mathcal{D}\\) [20]. The loss function \\(\\mathcal{L}_{\\mathrm{SE}}\\) receives the rendered normal image \\(\\mathcal{D}_c(v)\\) and text input \\(y\\) to evaluate the semantic guidance, which is detailed in Section 3.3. To accompany the external forces from \\(\\mathcal{L}_{\\mathrm{SE}}\\) that drive the mesh deformation with intrinsic smoothness energies, we follow the framework of a Sobolev preconditioned gradient descent [35], where the base mesh is reparameterized by the mesh Laplacian \\(L\\):" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.7, + 0.483, + 0.718 + ], + "angle": 0, + "content": "\\[\nv ^ {*} = (I + \\lambda L) v. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.485, + 0.775 + ], + "angle": 0, + "content": "This preconditioning involves solving a sparse linear system at every iteration, modifying the gradient descent update for each mesh deformation step to:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.785, + 0.483, + 0.817 + ], + "angle": 0, + "content": "\\[\nv \\leftarrow v - \\eta (I + \\lambda L) ^ {- 1} \\frac {\\partial \\mathcal {L} _ {\\mathrm {S E}}}{\\partial v}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the learning rate, \\(I\\) is the identity matrix, and \\(\\lambda\\) is a hyperparameter to control the extent of gradient diffusion over the entire domain. We set \\(\\lambda = 15\\) throughout our experiments to balance the global structure and fine details during mesh deformation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "Compared to directly adding a Laplacian regularization term to Equation (1), the preconditioning framework [35] is critical in achieving large mesh deformation while maintaining proper topology with reduced triangle flips (Figure 4). Several works [11, 51] adopt the strategy by Aignerman et al. [1], parameterizing deformation through Jacobian fields that capture the scaling and rotation of each triangle. Although this method effectively smooths vertex displacements, the local deformation represented in Jacobians accumulates, leading to global drifting for open-boundary meshes, making it challenging to bake the mesh as a brush." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Additionally, we provide an optional region mask to restrict mesh deformation to the user-defined region during optimization. It helps maintain zero values in unused VDM areas when generating geometric structures. For producing surface details, the region mask ensures that the brush effect meets user-customized requirements (Figure 9)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.905, + 0.44 + ], + "angle": 0, + "content": "By framing the VDM generation task as mesh deformation through preconditioned optimization, our framework preserves the favorable initial mesh topology and ensures control throughout the procedure. The resulting mesh preserves the original structure while incorporating rich local deformations, making it well-suited for baking as a brush." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.451, + 0.905, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.603, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Figure 4. Effect of \\(\\lambda\\). (a) Low-quality results w/o mesh smoothness, (b) low-quality results w/ direct Laplacian regularization, (c) our adopted preconditioning scheme with \\(\\lambda = 15\\), and (d) results with over-smoothness caused by a larger \\(\\lambda\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.686, + 0.868, + 0.7 + ], + "angle": 0, + "content": "3.3. Semantic Enhancement Score Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.709, + 0.907, + 0.86 + ], + "angle": 0, + "content": "Current text-to-3D generation methods like DreamFusion [37] often optimize a 3D representation parameterized by \\(\\theta\\) so that rendered images \\(\\mathbf{x} = g(\\theta)\\) resemble 2D samples produced by a pre-trained T2I diffusion model for a given text prompt \\(y\\). \\(g\\) functions as a differentiable renderer. The T2I diffusion model \\(\\phi\\) predicts the sampled noise \\(\\epsilon_{\\phi}(\\mathbf{x}_t; y, t)\\) of a rendered image \\(\\mathbf{x}_t\\) at a noise level \\(t\\) for the text input \\(y\\). To make rendered images follow the text-conditioned distribution in Stable Diffusion, the SDS loss updates \\(\\theta\\) by estimating the gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.871, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon , c} \\left[ \\omega (t) \\left(\\epsilon_ {\\phi} (\\mathbf {x} _ {t}; y, t) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16885" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.435, + 0.107 + ], + "angle": 0, + "content": "where \\(\\omega (t)\\) is a time-dependent weighting function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.272 + ], + "angle": 0, + "content": "However, the SDS loss cannot effectively supervise sub-object structure generation due to the issue of semantic coupling in full objects. For example, when using the SDS loss to generate a tortoise shell, it also usually generates the tortoise's tail and head, causing semantic coupling (Figure 7). We believe that the issue of semantic coupling in SDS stems from the training data of Stable Diffusion, in which most images contain full objects rather than separate parts. Therefore, the semantics of full objects often appear in the target distribution conditioned on text only describing sub-object structures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.484, + 0.364 + ], + "angle": 0, + "content": "A straightforward approach is using negative distributions via Classifier Score Distillation (CSD) [59] or Variational Score Distillation (VSD) [54] to suppress coupled semantics. CSD employs predefined negative prompts, resulting in more accurate negative distributions than those adaptively learned by VSD [59]:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.37, + 0.483, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {C S D}} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon , c} [ (\\omega_ {\\mathrm {p o s}} \\cdot \\epsilon_ {\\phi} (\\mathbf {x} _ {t}; y, t) \\\\ \\left. - \\omega_ {\\text {n e g}} \\cdot \\epsilon_ {\\phi} \\left(\\mathbf {x} _ {t}; y _ {\\text {n e g}}, t\\right)\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} ], \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.56 + ], + "angle": 0, + "content": "where \\(\\omega_{\\mathrm{pos}}\\) and \\(\\omega_{\\mathrm{neg}}\\) denote different weights for positive and negative prompts. It requires two separate inferences with positive and negative prompts to obtain two distributions, which are then subtracted to suppress coupled semantics. However, our experiments show that CSD is ineffective in decoupling semantics because the negative prompt cannot accurately represent the undesirable coupled semantics in the positive prompt (Figure 7). This results in noisy guidance, making CSD less effective in decoupling semantics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.484, + 0.696 + ], + "angle": 0, + "content": "Unlike semantic suppression in CSD, we propose a semantic enhancement method to mitigate semantic coupling by enhancing the semantics of part-related words. This can lead to a more accurate and stable target distribution, as shown in Section 4.3. Our key design is to apply weighted blending to the tokens in the original prompt to obtain a semantically focused text embedding, which serves as stable guidance for the optimization process. We define the Semantic Enhancement SDS loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.699, + 0.483, + 0.731 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S E}} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon , c} \\left[ \\omega (t) \\left(\\epsilon_ {\\phi} ^ {*} (\\mathbf {x} _ {t}; y, t) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\phi}^{*}(\\cdot)\\) uses a text embedding weighted by Compel [48]. Specifically, we assign a weight \\(s\\) to each word in the prompt and compute the weighted embedding \\(e_{\\mathrm{w}}\\) for each word by blending the original word embedding \\(e\\) and the empty text embedding \\(e_0\\): \\(e_{\\mathrm{w}} = e_0 + s\\cdot (e - e_0)\\). By concatenating the weighted embedding of each word in sequence, we obtain the final semantically focused text embedding. In our experiments, we found that using \\(s = 1.1^2\\) for words that require semantic enhancement can achieve stable optimization and effectively alleviate the issue of semantic coupling." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.092, + 0.588, + 0.136 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.089, + 0.663, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.14, + 0.667, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.089, + 0.732, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.145, + 0.732, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.2, + 0.592, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.2, + 0.662, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.595, + 0.253, + 0.663, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.614, + 0.306, + 0.641, + 0.314 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.2, + 0.735, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.253, + 0.735, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.305, + 0.732, + 0.313 + ], + "angle": 0, + "content": "Text2Mesh" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.089, + 0.809, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.144, + 0.806, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.2, + 0.808, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.255, + 0.808, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.305, + 0.798, + 0.314 + ], + "angle": 0, + "content": "Paint-it" + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.089, + 0.879, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.824, + 0.145, + 0.881, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.202, + 0.883, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.258, + 0.89, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.305, + 0.887, + 0.314 + ], + "angle": 0, + "content": "TextDeformer" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.325, + 0.907, + 0.38 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons of generated brushes for surface details. Our method captures geometry details guided by text, effectively preserving the surface structure and avoiding mesh distortion." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.405, + 0.646, + 0.423 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.522 + ], + "angle": 0, + "content": "We conducted experiments to evaluate the various capabilities of Text2VDM both quantitatively and qualitatively for text-to-VDM brush generation. We then present an ablation study that validates the significance of our key insight into Semantic Enhancement SDS, as well as the effect of the region mask and VDM initialization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.724, + 0.547 + ], + "angle": 0, + "content": "4.1. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.811 + ], + "angle": 0, + "content": "To the best of our knowledge, Text2VDM is the first framework to generate VDM brushes from text. We adapted three existing methods for comparison and classified them into two categories. The first category includes Text2Mesh [31] and TextDeformer [11], which generate a brush mesh through text-guided mesh deformation on a planar mesh, following a process similar to ours. For the second category, we opt to directly generate VDM via Paint-it [58]. Notably, this method originally uses SDS to optimize a UNet for generating PBR textures. We reframed it to suit our VDM brush generation task, modifying it to generate VDM through SDS optimization of the UNet. In geometric structures generation experiment (Figure 6), all methods are compared fairly, with the same non-zero VDM initialization and mask applied to each prompt (see Appendix B.2). For surface details (Figure 5), all methods start with a zero-valued VDM and no masks to ensure a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Compared to other methods, Text2VDM can generate better-quality VDM brushes. Text2Mesh applies displacement to each vertex along normal directions, resulting in limited mesh deformation. TextDeformer indicates the accumulation of local deformations in the Jacobians, which results in global mesh drift, making it challenging to bake" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16886" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.086, + 0.891, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.496, + 0.908, + 0.525 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons of generated brushes for geometric structures. Our method accurately presents key geometric features described by text, facilitating downstream applications in modeling software." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.643 + ], + "angle": 0, + "content": "these meshes into VDM. Reframed Paint-it VDM generation is equivalent to optimizing the three-axis displacement of each vertex on the mesh with SDS. Although the UNet reduces noise from the SDS [58], smooth regularization is still required to ensure mesh quality, which makes achieving high-quality mesh generation quite challenging." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.651, + 0.315, + 0.667 + ], + "angle": 0, + "content": "4.2. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.719 + ], + "angle": 0, + "content": "We quantitatively evaluated our framework regarding generation consistency with text input and mesh quality. We used 40 distinctive text prompts for VDM generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Generation Consistency with Text. We initially assessed the relevance of the generated results to the text descriptions [40]. 12 different views were rendered for average scores respectively, as presented in Table 1. Our approach achieves the highest scores compared to baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Mesh Quality. We evaluated mesh quality by examining self-intersection. Paint-it and Text2Mesh, which utilize direct vertex displacement, often converge to a local minimum and disregard the mesh triangulation. While TextDeformer exhibits the lowest self-intersection, its tendency to produce over-smoothed results frequently results in losing object features described in text prompts." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.55, + 0.907, + 0.604 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation of state-of-the-art methods. The geometry CLIP score is calculated on shaded images with uniform albedo colors [39], and self-intersection is quantified as the ratio of self-intersected mesh faces to the total number of faces." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.616, + 0.908, + 0.681 + ], + "angle": 0, + "content": "
Geometry CLIP Score ↑Mesh Self-Intersection ↓
Paint-it0.237519.42%
Text2Mesh0.24977.18%
TextDeformer0.24770.04%
Ours0.25560.77%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.574, + 0.695, + 0.845, + 0.709 + ], + "angle": 0, + "content": "Table 2. User evaluation of generated VDMs." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.72, + 0.894, + 0.785 + ], + "angle": 0, + "content": "
User Preference ↑Geometry QualityConsistency with Text
Paint-it3.1%1.7%
Text2Mesh18.3%27.3%
TextDeformer3.3%3.4%
Ours75.3%67.6%
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "User Study. We further conducted a user study to evaluate the effectiveness and expressiveness of our method. A Google Form was utilized to assess 1) geometry quality and 2) consistency with text. We recruited 32 participants, of whom 14 are graduate students majoring in media arts, and 18 are company employees specializing in AI content gen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16887" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.086, + 0.885, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.327, + 0.908, + 0.37 + ], + "angle": 0, + "content": "Figure 7. Effect of Semantic Enhancement SDS. Our method effectively mitigates semantic coupling issues in SDS, such as generating the tortoise's tail and head or the deer's ear and mouth, by providing more focused semantic guidance. In contrast, the semantic suppression method used in CSD and the other semantic enhancement approach proposed by Attend-and-Excite both lead to an unstable optimization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.396, + 0.482, + 0.456 + ], + "angle": 0, + "content": "eration. The participants were instructed to choose the preferred renderings of VDM from different methods in randomized order, as shown in Table 2. The results show participants preferred our method by a significant margin." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.245, + 0.484 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.703 + ], + "angle": 0, + "content": "Effect of Semantic Enhancement SDS. Figure 7 compares the results generated by the original SDS [37], our Semantic Enhancement SDS with Compel, SDS with another semantic enhancement method in Attend-and-Excite (AaE) [7], and CSD [59] with two different annealed weights for the negative prompts: \"tortoise tail, tortoise head\" and \"deer's ear, deer's mouth.\" We also qualitatively compare these methods by visualizing their performance on semantic decoupling using the same Stable Diffusion for 2D image generation (Figure 8). Compared to semantic suppression using negative prompts and semantic enhancement by AaE, our design of introducing a semantically focused text embedding via Compel to SDS is most effective in resolving the issue of semantic coupling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Key Insight. As discussed in Section 3.3, SDS can result in semantic coupling when generating sub-object structures, leading to artifacts like the tortoise's tail and head or the deer's ear and deer's mouth. We observed that the semantics represented by negative prompts are also coupled. Therefore, meaningless semantics can emerge when applying semantic suppression in SDS, which leads to unstable optimization. Increasing the weight of negative prompts further reduces the overall quality of generated results. AaE enhances the cross-attention map of specific tokens by continuously updating the latent code. However, AaE is not suitable for the SDS framework because the latent code is affected by different camera poses and random noise at each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.907, + 0.562 + ], + "angle": 0, + "content": "iteration. This temporal randomness undermines the continuity of the latent code, resulting in unstable optimization and subpar results. In contrast, our method uses Compel to enhance semantics and achieves more effective semantic decoupling. Moreover, the semantically focused text embedding produced by Compel is independent of temporal variation. These properties help mitigate semantic coupling during SDS optimization, leading to high-quality sub-object geometric structures. Meanwhile, our method can also enhance the semantics in the text prompt for surface detail generation (see Appendix A.2)." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.575, + 0.907, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.655, + 0.905, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.74, + 0.907, + 0.837 + ], + "angle": 0, + "content": "Figure 8. Visualization of semantic decoupling performance. The images generated by the T2I model qualitatively visualize sub-object semantics. (a) Coupled semantics in positive and negative prompts lead to meaningless distributions when subtracted. (b) Enhanced semantics by Compel achieves stable decoupling. (c) AaE can only ensure that the enhanced semantics are preserved, but cannot decouple negative ones effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effect of Region Mask. Figure 9 demonstrates two region masks and their control over surface detail generation given a text prompt. Without a region mask, the results can still" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16888" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.487, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.198, + 0.483, + 0.227 + ], + "angle": 0, + "content": "Figure 9. Effect of region mask. Region masks can effectively control the pattern of surface details based on text input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.252, + 0.484, + 0.327 + ], + "angle": 0, + "content": "match the text prompt but lack a specific pattern. By using a region mask and maintaining an activation ratio of \\(1/2\\) throughout the total iterations as a warm-up stage, we achieve a reasonable tradeoff between mask-result alignment and generation diversity (see Appendix A.3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.328, + 0.484, + 0.464 + ], + "angle": 0, + "content": "Effect of VDM Initialization. Our method demonstrates that user-specified VDMs can effectively control the volume and direction of generated geometric structures. Figure 10 shows that the results are high-quality and match the text descriptions well, such as the elf ear and pauldron. As this initializes the Laplacian term and steers the gradient flow in geometric structure generation, users can easily specify an initial VDM or choose a VDM template provided in our framework to generate expressive results." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.474, + 0.487, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Figure 10. Effect of VDM initialization. User-specified VDMs can help achieve the intended final effect of geometric structures by initializing the brush's volume and direction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.225, + 0.773 + ], + "angle": 0, + "content": "4.4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Once various VDMs are generated, users can use these brushes to meet diverse creative needs in modeling software. For example, they can apply VDM brushes for mesh stylization and a real-time iterative modeling process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Local-to-Global Mesh Stylization. Although mesh stylization is a complex task even for professional artists, combining different surface details allows users to achieve stylization quickly (see Appendix C.3)." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.909, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.252, + 0.907, + 0.308 + ], + "angle": 0, + "content": "Figure 11. Coarse to fine interactive modeling. By combining geometric structures brushes and surface details brushes for iterative sculpting in modeling software, users can rapidly create an expressive model from a plain shape (top left)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Coarse-to-Fine Interactive Modeling. Unlike previous methods [4, 60] that require a lengthy optimization process for each edit and result in non-reusable outcomes, our generated VDM brushes can be directly used in modeling software. This enables users to apply the generated brushes easily and interactively (Figure 11)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.469, + 0.634, + 0.485 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.713 + ], + "angle": 0, + "content": "We have presented Text2VDM, a novel framework for VDM brush generation from text. A VDM is a non-natural 2D image where each pixel stores a 3D displacement vector, making it challenging for existing T2I models to generate. Therefore, we treat VDM generation as diffusion-guided mesh deformation formulated as a form of Sobolev preconditioned optimization. To mitigate semantic coupling issues in SDS, we propose using weighted blending for prompt tokens, achieving high-quality brush generation. Moreover, we introduce two control methods, i.e., region and shape control, to meet customized requirements. The generated VDMs are directly compatible with mainstream modeling software, enabling various applications such as mesh stylization and interactive modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Limitations and Future Work. While our framework can generate high-quality VDM brushes, they may encounter multiview inconsistencies, a common issue introduced by SDS. View-consistent diffusion models like MV2MV [6] may be helpful to further handle this. Additionally, errors of the T2I models may guide wrong 3D generation, such as the wrong number of fingers in the generated skeleton hand. VDMs demonstrate that complex 3D models can be efficiently created using diverse reusable sculpting brushes. Future exploration of 3D generation through the assembly of modular components with similar design principles holds promising research value." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16889" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.176 + ], + "angle": 0, + "content": "This work was partially supported by the Guangdong Provincial Key Lab of Integrated Communication, Sensing and Computation for Ubiquitous Internet of Things #2023B1212010007." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.188, + 0.188, + 0.204 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.213, + 0.484, + 0.27 + ], + "angle": 0, + "content": "[1] Noam Aigerman, Kunal Gupta, Vladimir G. Kim, Siddhartha Chaudhuri, Jun Saito, and Thibault Groueix. Neural Jacobian Fields: Learning Intrinsic Mappings of Arbitrary Meshes. ACM Trans. Graph., 41(4), 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.27, + 0.484, + 0.325 + ], + "angle": 0, + "content": "[2] Thiemo Alldieck, Nikos Kolotouros, and Cristian Sminchisescu. Score Distillation Sampling with Learned Manifold Corrective. European Conference on Computer Vision, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.326, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[3] Mohammadreza Armandpour, Ali Sadeghian, Huangjie Zheng, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the Negative Prompt Algorithm: Transform 2D Diffusion into 3D, alleviate Janus problem and Beyond. arXiv preprint 2304.04968, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.396, + 0.484, + 0.452 + ], + "angle": 0, + "content": "[4] Amir Barda, Vladimir G. Kim, Noam Aigerman, Amit H. Bermano, and Thibault Groueix. MagicClay: Sculpting Meshes with Generative Neural Fields. SIGGRAPH Asia (Conference track), 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.48, + 0.479 + ], + "angle": 0, + "content": "[5] Blender Foundation. Blender. https://www.blender.org, 2025. Accessed Mar 5, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.522 + ], + "angle": 0, + "content": "[6] Youcheng Cai, Runshi Li, and Ligang Liu. MV2MV: Multi-View Image Translation via View-Consistent Diffusion Models. ACM Trans. Graph., 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.484, + 0.578 + ], + "angle": 0, + "content": "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-Ecite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models. 42 (4), 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.485, + 0.648 + ], + "angle": 0, + "content": "[8] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3D: Disentangling Geometry and Appearance for High-quality Text-to-3D Content Creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[9] Dale Decatur, Itai Lang, and Rana Hanocka. 3D Highlighter: Localizing Regions on 3D Shapes via Text Descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20930-20939, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[10] Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3D Paintbrush: Local Stylization of 3D Shapes with Cascaded Score Distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4473-4483, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[11] William Gao, Noam Aigerman, Thibault Groueix, Vova Kim, and Rana Hanocka. TextDeformer: Geometry Manipulation Using Text Guidance. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.48, + 0.872 + ], + "angle": 0, + "content": "[12] Maxon Computer GMBH. ZBrush. https://www.maxon.net/zbrush, 1999. Accessed Oct 18, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.213, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "Yoshua Bengio. Generative Adversarial Nets. Advances in Neural Information Processing Systems, 27, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[14] Amir Hertz, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Deep Geometric Texture Synthesis. ACM Trans. Graph., 39(4), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[15] Susung Hong, Donghoon Ahn, and Seungryong Kim. Debiasing Scores and Prompts of 2D Diffusion for View-consistent Text-to-3D Generation. In Neural Information Processing Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[16] Anita Hu, Nishkrit Desai, Hassan Abu Alhaija, Seung Wook Kim, and Maria Shugrina. Diffusion Texture Painting. New York, NY, USA, 2024. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.904, + 0.344 + ], + "angle": 0, + "content": "[17] Teng Hu, Ran Yi, Haokun Zhu, Liang Liu, Jinlong Peng, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Stroke-based Neural Painting and Stylization with Dynamically Predicted Painting Region. New York, NY, USA, 2023. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[18] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free Score Distillation. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[19] Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Björn Ommer. Rethinking Style Transfer: From Pixels to Parameterized Brushstrokes. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12191-12200, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[20] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular Primitives for High-Performance Differentiable Rendering. ACM Transactions on Graphics, 39(6), 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[21] Beichen Li, Liang Shi, and Wojciech Matusik. End-to-end procedural material capture with proxy-free mixed-integer optimization. ACM Transactions on Graphics (TOG), 42(4): 1-15, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[22] Beichen Li, Yiwei Hu, Paul Guerrero, Miloš Hasan, Liang Shi, Valentin Deschaintre, and Wojciech Matusik. Procedural material generation with reinforcement learning. ACM Transactions on Graphics (TOG), 43(6):1-14, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[23] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast Text-to-3D with Sparse-View Generation and Large Reconstruction Model. arXiv preprint 2311.06214, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[24] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. FocalDreamer: Text-Driven 3D Editing via Focal-Fusion Assembly. Proceedings of the AAAI Conference on Artificial Intelligence, 38:3279–3287, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[25] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. LucidDreamer: Towards High-Fidelity Text-to-3D Generation via Interval Score Matching. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6517-6526, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-Resolution Text-to-3D Content Creation. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16890" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300-309, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[27] Feng-Lin Liu, Hongbo Fu, Yu-Kun Lai, and Lin Gao. SketchDream: Sketch-based Text-to-3D Generation and Editing. ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH 2024), 43(4), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[28] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot One Image to 3D Object. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9264–9275, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[29] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Ruifeng Deng, Xin Li, Errui Ding, and Hao Wang. Paint Transformer: Feed Forward Neural Painting with Stroke Prediction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6578-6587, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[30] Xiaoxiao Long, Yuanchen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3D: Single Image to 3D Using Cross-Domain Diffusion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9970-9980, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[31] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2Mesh: Text-Driven Neural Stylization for Meshes. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13482–13492, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[32] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[34] Haoran Mo, Edgar Simo-Serra, Chengying Gao, Changqing Zou, and Ruomei Wang. General virtual sketching framework for vector line art. ACM Transactions on Graphics (TOG), 40:1-14, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[35] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large Steps in Inverse Rendering of Geometry. ACM Trans. Graph., 40(6), 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[36] Openai. DALLE. https://openai.com/index/dall-e-3/, 2024. Accessed Oct 22, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[37] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. In The Eleventh International Conference on Learning Representations, 2022. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[38] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. RichDreamer: A Generalizable Normal-Depth Diffusion Model for Detail Richness in Text-to-3D. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914–9925, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[39] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "Bo, and Xiaoguang Han. RichDreamer: A Generalizable Normal-Depth Diffusion Model for Detail Richness in Text-to-3D. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914–9925, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In International Conference on Machine Learning, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.317 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[42] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards Zero-shot Text-to-shape Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18603-18613, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[43] Artem Sevastopolsky, Philip-William Grassal, Simon Giebenhain, ShahRukh Athar, Luisa Verdoliva, and Matthias Niessner. Headcraft: Modeling high-detail shape variations for animated 3dmms, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[44] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep Marching Tetrahedra: a Hybrid Representation for High-resolution 3D Shape Synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[45] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view Diffusion for 3D Generation. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[46] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3D Neural Field Generation Using Triplane Diffusion. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[47] Maria Shugrina, Chin-Ying Li, and Sanja Fidler. Neural Brushstroke Engine: Learning a Latent Style Space of Interactive Drawing Tools. ACM Trans. Graph., 41(6), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[48] Damian Stewart. Compel: A Text Prompt Weighting and Blending Library. https://github.com/damian0815/compel, 2023. Accessed Oct 20, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[49] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative Gaussian Splitting for Efficient 3D Content Creation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[50] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-It-3D: High-fidelity 3D Creation from A Single Image with Diffusion Prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22819–22829, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Duotun Wang, Hengyu Meng, Zeyu Cai, Zhijing Shao, Qianxi Liu, Lin Wang, Mingming Fan, Xiaohang Zhan, and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16891" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Zeyu Wang. HeadEvolver: Text to Head Avatars via Expressive and Attribute-Preserving Mesh Deformation. International Conference on 3D Vision, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[52] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12619-12629, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[53] Peihao Wang, Dejia Xu, Zhiwen Fan, Dilin Wang, Sreyas Mohan, Forrest N. Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu, Zhangyang Wang, and Vikas Chandra. Taming Mode Collapse in Score Distillation for Text-to-3D Generation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[54] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and Diverse Text-to-3D Generation with Variational Score Distillation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.389, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[55] Xingguang Yan, Han-Hung Lee, Ziyu Wan, and Angel X. Chang. An Object is Worth 64x64 Pixels: Generating 3D Object via Image Diffusion, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[56] Xiaofeng Yang, Yiwen Chen, Cheng Chen, Chi Zhang, Yi Xu, Xulei Yang, Fayao Liu, and Guosheng Lin. Learn to Optimize Denoising Scores for 3D Generation: A Unified and Improved Diffusion Prior on NeRF and 3D Gaussian Splatting. arXiv preprint 2312.04820, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.502, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[57] Yuezhi Yang, Qimin Chen, Vladimir Kim, Siddhartha Chaudhuri, Qixing Huang, and Zhiqin Chen. GenVDM: Generating Vector Displacement Maps From a Single Image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.573, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[58] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-Texture Synthesis via Deep Convolutional Texture Map Optimization and Physically-Based Rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.482, + 0.696 + ], + "angle": 0, + "content": "[59] Xin Yu, Yuanchen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and Xiaojuan Qi. Text-to-3D with Classifier Score Distillation. arXiv preprint 2310.19415, 2023. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.7, + 0.482, + 0.754 + ], + "angle": 0, + "content": "[60] Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. TIP-Editor: An Accurate 3D Editor Following Both Text-Prompts and Image-Prompts. ACM Trans. Graph., 43(4), 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.756, + 0.482, + 0.811 + ], + "angle": 0, + "content": "[61] Zhengxia Zou, Tianyang Shi, Shuang Qiu, Yi Yuan, and Zhenwei Shi. Stylized Neural Painting. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15684-15693, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16892" + } + ] +] \ No newline at end of file diff --git a/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/011b91bd-0c46-4896-9ef0-7e4bb90ad7cb_origin.pdf b/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/011b91bd-0c46-4896-9ef0-7e4bb90ad7cb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1cc7458b584474fe1c2804d508206c604637fd13 --- /dev/null +++ b/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/011b91bd-0c46-4896-9ef0-7e4bb90ad7cb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:22102e35405c65a668c546b733d4dba017d6ed33730efd594633b871c8c7e347 +size 2449054 diff --git a/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/full.md b/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..05fb15b38a540a241ad289a4b3b8c78e8e9974e1 --- /dev/null +++ b/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/full.md @@ -0,0 +1,324 @@ +# Text2VDM: Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting + +Hengyu Meng $^{1}$ Duotun Wang $^{1}$ Zhijing Shao $^{1}$ Ligang Liu $^{2}$ Zeyu Wang $^{1,3*}$ $^{1}$ The Hong Kong University of Science and Technology (Guangzhou) + $^{2}$ University of Science and Technology of China + $^{3}$ The Hong Kong University of Science and Technology + +![](images/8808fca1b5050677e7463aca4d4c257791061307f1993715e82713dfa58cbf92.jpg) +Figure 1. Example VDM brushes generated by Text2VDM and sculpted models in Blender. Text2VDM can produce high-quality brushes for surface details (top row) and geometric structures (bottom row) from text input. Users can rapidly create an expressive model from a plain shape by directly applying these brushes in Blender. Yellow underlined text highlights semantics enhanced by our framework. + +# Abstract + +Professional 3D asset creation often requires diverse sculpting brushes to add surface details and geometric structures. Despite recent progress in 3D generation, producing reusable sculpting brushes compatible with artists' workflows remains an open and challenging problem. These sculpting brushes are typically represented as vector displacement maps (VDMs), which existing models cannot easily generate compared to natural images. This paper presents Text2VDM, a novel framework for text-to-VDM brush generation through the deformation of a dense planar mesh guided by score distillation sampling (SDS). The original SDS loss is designed for generating full objects and struggles with generating desirable sub-object structures from scratch in brush generation. We refer to this issue as semantic coupling, which we address by introducing weighted blending of prompt tokens to SDS, resulting in a + +more accurate target distribution and semantic guidance. Experiments demonstrate that Text2VDM can generate diverse, high-quality VDM brushes for sculpting surface details and geometric structures. Our generated brushes can be seamlessly integrated into mainstream modeling software, enabling various applications such as mesh stylization and real-time interactive modeling. + +# 1. Introduction + +Sculpting brushes are essential tools in 3D asset creation, as artists often require a variety of brushes to create surface details and geometric structures. In modeling software, 3D sculpting brushes are typically defined as vector displacement maps (VDMs). A VDM is a 2D image where each pixel stores a 3D displacement vector. Through these vectors, VDM brushes can create complex surface details, such as cracks and wood grain, or generate geometric structures + +like ears and horns. This allows artists to apply the same geometric pattern iteratively while sculpting. + +Despite significant advances in text-to-image (T2I) [36, 41] and text-to-3D generation [23, 30, 37, 49, 50], existing methods are unsuitable for creating VDM brushes. We summarize the challenges as follows: 1) Since VDMs are not natural images (Figure 2), it is difficult for existing T2I models to generate them directly. 2) From a 3D perspective, a VDM represents mesh deformation through per-vertex displacement vectors from a dense planar mesh. Mapping any generated mesh to a dense planar mesh to create a VDM is non-trivial. 3) Sculpting brushes often involve sub-object structures, whereas most 3D generation methods can only generate full objects. Enabling users to accurately control the generation of sculpting brushes through text prompts in a semantically focused manner remains challenging. + +To address the challenges of brush generation, we propose Text2VDM, a novel optimization-based framework that generates diverse and controllable VDM brushes from text input. Our approach does not generate VDMs directly from a T2I model. Instead, we address VDM brush generation from a 3D perspective by applying score distillation with a pre-trained T2I model to guide mesh deformation. Our framework supports three ways to initialize a base mesh through a zero-valued, spike-pattern, or user-specified VDM for custom shape control. For mesh deformation, we formulate a Sobolev preconditioned optimization [35] to maintain mesh quality with intrinsic smoothness. We also provide optional region control using a mask of activated mesh deformation, helping users obtain the intended brush effects. The normal maps of the mesh are then rasterized by a differentiable renderer for brush optimization. + +We observed that the standard score distillation sampling (SDS) [37] can lead to semantic coupling when supervising the generation of sub-object level structures due to the associated semantics caused by the noisy gradients from the full object. For example, a generated deer's horn should not be a full deer's head, or a generated beard should not include a nose. A straightforward solution is to use negative prompts [18, 59] to exclude undesired semantics, but our experiments show that this semantics suppression approach is ineffective in decoupling semantics and leads to an unstable optimization process. Instead, we propose to enhance the semantics of part-related words by applying weighted blending to the tokens in the prompt. This results in semantically focused text embedding, directing toward a more precise target distribution while reducing noisy gradients during optimization. + +Our experiments demonstrate that Text2VDM produces high-quality and diverse VDM brushes that can be directly integrated into mainstream modeling software, such as Blender [5] and ZBrush [12]. Compared to existing methods that directly generate full 3D models, our approach ad + +![](images/9db4fae690336e980920165d4dd6898a39d5f74a7765dd310c2606c9b3ce2c64.jpg) +(a) Vector Displacement +Figure 2. Difference between vector and standard displacement. The VDM enables full 3D vector displacement, while the height map only allows unidirectional standard displacement. + +![](images/351cde89c79c3a88dab240689ff39d098aef6945e286f373d4bb3c14011f8a9a.jpg) +(b) Standard Displacement + +dresses a different use case where brush-based user sculpting is desirable. This enables artists to interactively use a variety of brushes to sculpt diverse and expressive models from a plain shape. + +This paper makes the following contributions: + +- We first introduce the task of text-to-VDM brush generation, which is challenging to tackle directly using current text-to-image and text-to-3D methods. +- We propose Text2VDM, a novel framework for text-to-VDM brush generation that is readily compatible with artists' workflow of 3D asset creation. +- We design a novel Semantic Enhancement SDS loss, which uses weighted blending to mitigate semantic coupling for sub-object structure generation. + +# 2. Related Work + +Text to Local 3D Generation and Editing. With recent advances in diffusion models [41] and differentiable 3D representations [1, 33, 35, 44, 46], many methods for text-guided full 3D model generation have emerged [8, 11, 23, 25, 31, 38, 42]. Since 3D content creation is an iterative process that often requires user interaction, more attention has been directed toward localized 3D generation and editing. For example, 3D Highlighter [9] and 3D Paintbrush [10] use text as input, leveraging pre-trained CLIP models [40] or diffusion models [37] to supervise the optimization of neural networks for segmenting the regions of a 3D model that match the text description. Based on the information from these segmented regions, further editing of texture and geometry can be applied to the 3D model. Furthermore, SKED [32] and SketchDream [27] introduce sketches as an additional modality to assist in localized editing. To enable more precise control, FocalDreamer [24], MagicClay [4], and Tip-Editor [60] allow users to specify the editing location directly within the 3D space. These works rely on optimization-based methods to edit specific objects, often resulting in non-reusable editing outcomes. Additionally, each edit requires a lengthy optimization process, making interactivity difficult to achieve. + +![](images/f6e725640433d0a1fbaf19f0cdc17efc044134368f3d359e2927ed02a61c846d.jpg) +Figure 3. Overview of Text2VDM. Starting with a dense planar mesh constructed from a zero-valued VDM, users can initialize the mesh volume through a default spike-pattern VDM or a user-specified VDM. Given the text prompt and region mask, we propose the Semantic Enhancement SDS loss $\mathcal{L}_{\mathrm{SE}}$ to guide mesh deformations through a mesh Laplacian $L$ iteratively, achieving semantically focused generation of surface details or geometric structures. After generation, vertex displacements are baked into the final VDM. + +Diffusion Priors for 3D Generation. Score distillation sampling (SDS) [37, 52] provides pixel-level guidance by seeking specific modes in a diffusion model, inspiring further research to improve optimization-based 3D generation [3, 26, 53, 54, 56]. Some studies focus on mitigating the "Janus" problem [2, 15], while others fine-tune diffusion models with multiview datasets to enhance 3D consistency [28, 45]. Recent research focuses on refining the design of SDS loss to achieve more precise guidance. For instance, Make-it-3D [50] introduces two-stage optimizations to improve textured appearance, while Fantasia3D [8] dynamically modifies the time-dependent weighting function within SDS computations. Additionally, several methods [18, 59] incorporate negative prompts as the conditional term to further refine the optimizations. Although diffusion priors have achieved promising results, their application in generating sub-object structures without global context as a reference is still challenging. + +Appearance and Geometric Brush Synthesis. The concept of brushes is very common in the creative process of digital artists, serving as a reusable local decorative unit. Appearance brushes focus on color representation and drawing styles in 2D space. With the development of generative models [13, 41], many works have explored the synthesis of procedural material [21, 22] for 3D object texturing and appearance brushes for interactive painting [16, 47], realistic artworks generation [29, 34, 61], and applying stylization [17, 19]. Unlike appearance brushes, geometric brushes focus on modifying geometry by moving the ver + +tices of a mesh in 3D space. VDM brushes, as an extension of standard geometric brushes, provide more complex geometric effects by utilizing VDMs. To the best of our knowledge, only a few techniques adopted the concepts of VDM for generation [43, 55] and geometric texture transfer [14]. Recently, concurrent work [57] explored using a single image as input and leveraged a diffusion model for generating multiview normal maps to guide VDM reconstruction. In comparison, our method is fundamentally different in tackling the semantic coupling problem arising from text guidance. This demonstrates that generating geometric brushes is a highly promising research direction. + +# 3. Methodology + +To generate VDM brushes compatible with mainstream modeling software, we begin by constructing a dense mesh from an initial VDM, as shown in Figure 3. We then apply score distillation with Stable Diffusion to guide high-quality mesh deformation formulated as a form of Sobolev preconditioned optimization [35]. To produce the intended sub-object level structure described in the text, we design a Semantic Enhancement SDS loss by applying weighted blending to the tokens in the prompt, effectively handling the issue of semantic coupling in SDS. + +# 3.1. Brush Initialization + +We provide three methods to initialize a base mesh for brush generation via a zero-valued VDM, a spike-pattern VDM, or a user-specified VDM. A VDM is represented as a $512 \times$ + +512 three-channel image, in which each channel stores the displacement in the X, Y, or Z direction, respectively. We first construct a planar grid mesh by creating two triangles for every $2 \times 2$ pixels and then apply the displacement stored in the VDM to mesh vertices. The values in these three initial VDMs range from 0 to 1, in which 0 represents no displacement, and 1 corresponds to half of the mesh's edge length in the positive axis direction. Since users can apply sculpting brushes symmetrically, our initial VDM does not need to store any negative values. + +Our three methods for brush initialization facilitate the generation of diverse sculpting brush styles. The zero-valued VDM results in a planar mesh, which is our default setup when no control is provided. The spike-pattern VDM is suitable for generating protruding geometric structures, as it can effectively adjust the Laplacian term in Equation (2) to steer the gradient direction for mesh deformation. For better control of the brush's volume and direction, we also provide an interface for users to create custom VDMs, so the user-specified brush initialization can effectively guide mesh deformation toward the target structure. + +# 3.2. Brush Generation via Mesh Deformation + +Given the vertices $v$ on the initialized base mesh, our method aims to learn a mesh deformation to the target brush shape. The vertex positions $\hat{v}$ after mesh deformation can be expressed by: + +$$ +\hat {v} = \underset {v} {\arg \min } \mathcal {L} _ {\mathrm {S E}} \left(\mathcal {D} _ {c} (v), y\right), \tag {1} +$$ + +where $c$ represents the camera setup in a differentiable renderer $\mathcal{D}$ [20]. The loss function $\mathcal{L}_{\mathrm{SE}}$ receives the rendered normal image $\mathcal{D}_c(v)$ and text input $y$ to evaluate the semantic guidance, which is detailed in Section 3.3. To accompany the external forces from $\mathcal{L}_{\mathrm{SE}}$ that drive the mesh deformation with intrinsic smoothness energies, we follow the framework of a Sobolev preconditioned gradient descent [35], where the base mesh is reparameterized by the mesh Laplacian $L$ : + +$$ +v ^ {*} = (I + \lambda L) v. \tag {2} +$$ + +This preconditioning involves solving a sparse linear system at every iteration, modifying the gradient descent update for each mesh deformation step to: + +$$ +v \leftarrow v - \eta (I + \lambda L) ^ {- 1} \frac {\partial \mathcal {L} _ {\mathrm {S E}}}{\partial v}, \tag {3} +$$ + +where $\eta$ is the learning rate, $I$ is the identity matrix, and $\lambda$ is a hyperparameter to control the extent of gradient diffusion over the entire domain. We set $\lambda = 15$ throughout our experiments to balance the global structure and fine details during mesh deformation. + +Compared to directly adding a Laplacian regularization term to Equation (1), the preconditioning framework [35] is critical in achieving large mesh deformation while maintaining proper topology with reduced triangle flips (Figure 4). Several works [11, 51] adopt the strategy by Aignerman et al. [1], parameterizing deformation through Jacobian fields that capture the scaling and rotation of each triangle. Although this method effectively smooths vertex displacements, the local deformation represented in Jacobians accumulates, leading to global drifting for open-boundary meshes, making it challenging to bake the mesh as a brush. + +Additionally, we provide an optional region mask to restrict mesh deformation to the user-defined region during optimization. It helps maintain zero values in unused VDM areas when generating geometric structures. For producing surface details, the region mask ensures that the brush effect meets user-customized requirements (Figure 9). + +By framing the VDM generation task as mesh deformation through preconditioned optimization, our framework preserves the favorable initial mesh topology and ensures control throughout the procedure. The resulting mesh preserves the original structure while incorporating rich local deformations, making it well-suited for baking as a brush. + +![](images/33eac003bceaa84327a88f3d78016b0acc15ccd2d09d32395fc717d33840340c.jpg) +Figure 4. Effect of $\lambda$ . (a) Low-quality results w/o mesh smoothness, (b) low-quality results w/ direct Laplacian regularization, (c) our adopted preconditioning scheme with $\lambda = 15$ , and (d) results with over-smoothness caused by a larger $\lambda$ . + +# 3.3. Semantic Enhancement Score Distillation + +Current text-to-3D generation methods like DreamFusion [37] often optimize a 3D representation parameterized by $\theta$ so that rendered images $\mathbf{x} = g(\theta)$ resemble 2D samples produced by a pre-trained T2I diffusion model for a given text prompt $y$ . $g$ functions as a differentiable renderer. The T2I diffusion model $\phi$ predicts the sampled noise $\epsilon_{\phi}(\mathbf{x}_t; y, t)$ of a rendered image $\mathbf{x}_t$ at a noise level $t$ for the text input $y$ . To make rendered images follow the text-conditioned distribution in Stable Diffusion, the SDS loss updates $\theta$ by estimating the gradient: + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {S D S}} (\phi , \mathbf {x}) = \mathbb {E} _ {t, \epsilon , c} \left[ \omega (t) \left(\epsilon_ {\phi} (\mathbf {x} _ {t}; y, t) - \epsilon\right) \frac {\partial \mathbf {x}}{\partial \theta} \right], \tag {4} +$$ + +where $\omega (t)$ is a time-dependent weighting function. + +However, the SDS loss cannot effectively supervise sub-object structure generation due to the issue of semantic coupling in full objects. For example, when using the SDS loss to generate a tortoise shell, it also usually generates the tortoise's tail and head, causing semantic coupling (Figure 7). We believe that the issue of semantic coupling in SDS stems from the training data of Stable Diffusion, in which most images contain full objects rather than separate parts. Therefore, the semantics of full objects often appear in the target distribution conditioned on text only describing sub-object structures. + +A straightforward approach is using negative distributions via Classifier Score Distillation (CSD) [59] or Variational Score Distillation (VSD) [54] to suppress coupled semantics. CSD employs predefined negative prompts, resulting in more accurate negative distributions than those adaptively learned by VSD [59]: + +$$ +\begin{array}{l} \nabla_ {\theta} \mathcal {L} _ {\mathrm {C S D}} (\phi , \mathbf {x}) = \mathbb {E} _ {t, \epsilon , c} [ (\omega_ {\mathrm {p o s}} \cdot \epsilon_ {\phi} (\mathbf {x} _ {t}; y, t) \\ \left. - \omega_ {\text {n e g}} \cdot \epsilon_ {\phi} \left(\mathbf {x} _ {t}; y _ {\text {n e g}}, t\right)\right) \frac {\partial \mathbf {x}}{\partial \theta} ], \tag {5} \\ \end{array} +$$ + +where $\omega_{\mathrm{pos}}$ and $\omega_{\mathrm{neg}}$ denote different weights for positive and negative prompts. It requires two separate inferences with positive and negative prompts to obtain two distributions, which are then subtracted to suppress coupled semantics. However, our experiments show that CSD is ineffective in decoupling semantics because the negative prompt cannot accurately represent the undesirable coupled semantics in the positive prompt (Figure 7). This results in noisy guidance, making CSD less effective in decoupling semantics. + +Unlike semantic suppression in CSD, we propose a semantic enhancement method to mitigate semantic coupling by enhancing the semantics of part-related words. This can lead to a more accurate and stable target distribution, as shown in Section 4.3. Our key design is to apply weighted blending to the tokens in the original prompt to obtain a semantically focused text embedding, which serves as stable guidance for the optimization process. We define the Semantic Enhancement SDS loss as: + +$$ +\nabla_ {\theta} \mathcal {L} _ {\mathrm {S E}} (\phi , \mathbf {x}) = \mathbb {E} _ {t, \epsilon , c} \left[ \omega (t) \left(\epsilon_ {\phi} ^ {*} (\mathbf {x} _ {t}; y, t) - \epsilon\right) \frac {\partial \mathbf {x}}{\partial \theta} \right], \tag {6} +$$ + +where $\epsilon_{\phi}^{*}(\cdot)$ uses a text embedding weighted by Compel [48]. Specifically, we assign a weight $s$ to each word in the prompt and compute the weighted embedding $e_{\mathrm{w}}$ for each word by blending the original word embedding $e$ and the empty text embedding $e_0$ : $e_{\mathrm{w}} = e_0 + s\cdot (e - e_0)$ . By concatenating the weighted embedding of each word in sequence, we obtain the final semantically focused text embedding. In our experiments, we found that using $s = 1.1^2$ for words that require semantic enhancement can achieve stable optimization and effectively alleviate the issue of semantic coupling. + +![](images/32ae9705644a6d017e563c1cd9860a71631b6ca5aae86fbd97317c96a65f39c3.jpg) + +![](images/f27ac06eeb4621f121cd081556ec2d3db9348260927b25caf3219c37d711eb7e.jpg) + +![](images/60e68b185ce5dc93b63e20d6455ebdab6a35f9b0df83f89b8c11c0448c971353.jpg) + +![](images/4282a32cf52a03466831d93baf340c1fc528ac46fea74ed26e4343bec8841c2d.jpg) + +![](images/bb4a929c0f9e1a93864eb7e74df1588028f307f858761994e32d8d911d56f8ca.jpg) + +![](images/148f5918c3afdbff8627c7473e8b8896fb5cdc7a43e9740c06c29abdb8700778.jpg) + +![](images/339bd868ba412fd34b52a72ccc06e9d0577d2f292ff77a135a2fcbae511b9550.jpg) + +![](images/c45a80201127c8dad8be377ad86c137d1bd5ee47d5fa3b0f19a5802e6dab096c.jpg) +Ours + +![](images/3397ba1251c1e967e5002ef250713d758bb8c310c6bce1245e19255ad3fde7d5.jpg) + +![](images/0e0228dd3321b74f046295a20df5aac87d0b2d6311f97b4313c69627cd38b667.jpg) +Text2Mesh + +![](images/f8cd0eba056fd7575174ebb4f8c19fc7d066445320ce7a82121e4da9731a19db.jpg) + +![](images/9d3fb6c910fab193f57055bfdc9e63c46cc09062e59eb82b9c253272845535f3.jpg) + +![](images/19856accc9267105194cd89ab83bafbe8e6905fab5aba1034dac29276dee11c5.jpg) + +![](images/9caa1c0362d06a5b26b6e71cb1fabf4b7de490d3442f9d114c3f9945b0a6cce0.jpg) +Paint-it + +![](images/a4f63aedf8f1c54357004c942e8d84120957007958978fcf502a615f8724a107.jpg) + +![](images/527e8c7a94b256f4a08d6340dd3a708b83072e779451892ce944bd581d9655e7.jpg) + +![](images/1e19b63b09904c996cbb857f4e270e37afc7fe4b55a0ec45819238b6082dc7d0.jpg) + +![](images/46835acade887753b98a215374ec719ec1b5b9f6650e1d078abbf4ba2e5a7a3e.jpg) +TextDeformer +Figure 5. Qualitative comparisons of generated brushes for surface details. Our method captures geometry details guided by text, effectively preserving the surface structure and avoiding mesh distortion. + +# 4. Experiments + +We conducted experiments to evaluate the various capabilities of Text2VDM both quantitatively and qualitatively for text-to-VDM brush generation. We then present an ablation study that validates the significance of our key insight into Semantic Enhancement SDS, as well as the effect of the region mask and VDM initialization. + +# 4.1. Qualitative Evaluation + +To the best of our knowledge, Text2VDM is the first framework to generate VDM brushes from text. We adapted three existing methods for comparison and classified them into two categories. The first category includes Text2Mesh [31] and TextDeformer [11], which generate a brush mesh through text-guided mesh deformation on a planar mesh, following a process similar to ours. For the second category, we opt to directly generate VDM via Paint-it [58]. Notably, this method originally uses SDS to optimize a UNet for generating PBR textures. We reframed it to suit our VDM brush generation task, modifying it to generate VDM through SDS optimization of the UNet. In geometric structures generation experiment (Figure 6), all methods are compared fairly, with the same non-zero VDM initialization and mask applied to each prompt (see Appendix B.2). For surface details (Figure 5), all methods start with a zero-valued VDM and no masks to ensure a fair comparison. + +Compared to other methods, Text2VDM can generate better-quality VDM brushes. Text2Mesh applies displacement to each vertex along normal directions, resulting in limited mesh deformation. TextDeformer indicates the accumulation of local deformations in the Jacobians, which results in global mesh drift, making it challenging to bake + +![](images/2724a6d7f4554fc538ecefda3d302064971f545c7a888bed05ee89898092d0f5.jpg) +Figure 6. Qualitative comparisons of generated brushes for geometric structures. Our method accurately presents key geometric features described by text, facilitating downstream applications in modeling software. + +these meshes into VDM. Reframed Paint-it VDM generation is equivalent to optimizing the three-axis displacement of each vertex on the mesh with SDS. Although the UNet reduces noise from the SDS [58], smooth regularization is still required to ensure mesh quality, which makes achieving high-quality mesh generation quite challenging. + +# 4.2. Quantitative Evaluation + +We quantitatively evaluated our framework regarding generation consistency with text input and mesh quality. We used 40 distinctive text prompts for VDM generation. + +Generation Consistency with Text. We initially assessed the relevance of the generated results to the text descriptions [40]. 12 different views were rendered for average scores respectively, as presented in Table 1. Our approach achieves the highest scores compared to baseline methods. + +Mesh Quality. We evaluated mesh quality by examining self-intersection. Paint-it and Text2Mesh, which utilize direct vertex displacement, often converge to a local minimum and disregard the mesh triangulation. While TextDeformer exhibits the lowest self-intersection, its tendency to produce over-smoothed results frequently results in losing object features described in text prompts. + +Table 1. Quantitative evaluation of state-of-the-art methods. The geometry CLIP score is calculated on shaded images with uniform albedo colors [39], and self-intersection is quantified as the ratio of self-intersected mesh faces to the total number of faces. + +
Geometry CLIP Score ↑Mesh Self-Intersection ↓
Paint-it0.237519.42%
Text2Mesh0.24977.18%
TextDeformer0.24770.04%
Ours0.25560.77%
+ +Table 2. User evaluation of generated VDMs. + +
User Preference ↑Geometry QualityConsistency with Text
Paint-it3.1%1.7%
Text2Mesh18.3%27.3%
TextDeformer3.3%3.4%
Ours75.3%67.6%
+ +User Study. We further conducted a user study to evaluate the effectiveness and expressiveness of our method. A Google Form was utilized to assess 1) geometry quality and 2) consistency with text. We recruited 32 participants, of whom 14 are graduate students majoring in media arts, and 18 are company employees specializing in AI content gen + +![](images/a3b7ed331c8cb6a0a13ab499cf1eb6de03182366edcc4db96098b0f16658fb03.jpg) +Figure 7. Effect of Semantic Enhancement SDS. Our method effectively mitigates semantic coupling issues in SDS, such as generating the tortoise's tail and head or the deer's ear and mouth, by providing more focused semantic guidance. In contrast, the semantic suppression method used in CSD and the other semantic enhancement approach proposed by Attend-and-Excite both lead to an unstable optimization. + +eration. The participants were instructed to choose the preferred renderings of VDM from different methods in randomized order, as shown in Table 2. The results show participants preferred our method by a significant margin. + +# 4.3. Ablation Study + +Effect of Semantic Enhancement SDS. Figure 7 compares the results generated by the original SDS [37], our Semantic Enhancement SDS with Compel, SDS with another semantic enhancement method in Attend-and-Excite (AaE) [7], and CSD [59] with two different annealed weights for the negative prompts: "tortoise tail, tortoise head" and "deer's ear, deer's mouth." We also qualitatively compare these methods by visualizing their performance on semantic decoupling using the same Stable Diffusion for 2D image generation (Figure 8). Compared to semantic suppression using negative prompts and semantic enhancement by AaE, our design of introducing a semantically focused text embedding via Compel to SDS is most effective in resolving the issue of semantic coupling. + +Key Insight. As discussed in Section 3.3, SDS can result in semantic coupling when generating sub-object structures, leading to artifacts like the tortoise's tail and head or the deer's ear and deer's mouth. We observed that the semantics represented by negative prompts are also coupled. Therefore, meaningless semantics can emerge when applying semantic suppression in SDS, which leads to unstable optimization. Increasing the weight of negative prompts further reduces the overall quality of generated results. AaE enhances the cross-attention map of specific tokens by continuously updating the latent code. However, AaE is not suitable for the SDS framework because the latent code is affected by different camera poses and random noise at each + +iteration. This temporal randomness undermines the continuity of the latent code, resulting in unstable optimization and subpar results. In contrast, our method uses Compel to enhance semantics and achieves more effective semantic decoupling. Moreover, the semantically focused text embedding produced by Compel is independent of temporal variation. These properties help mitigate semantic coupling during SDS optimization, leading to high-quality sub-object geometric structures. Meanwhile, our method can also enhance the semantics in the text prompt for surface detail generation (see Appendix A.2). + +![](images/3c51a0f93af5061df6dec77d828d0d07063b2bb03f89e2f86c3d9e4accc8885f.jpg) + +![](images/9790439975bac725358d87ac5566b40000efe57edb8e33224a382539f041ce60.jpg) +Figure 8. Visualization of semantic decoupling performance. The images generated by the T2I model qualitatively visualize sub-object semantics. (a) Coupled semantics in positive and negative prompts lead to meaningless distributions when subtracted. (b) Enhanced semantics by Compel achieves stable decoupling. (c) AaE can only ensure that the enhanced semantics are preserved, but cannot decouple negative ones effectively. + +Effect of Region Mask. Figure 9 demonstrates two region masks and their control over surface detail generation given a text prompt. Without a region mask, the results can still + +![](images/dac4f56f276654eda9dc7b45b186243d881d04b8ac6b1ba0811cd831b44ad150.jpg) +Figure 9. Effect of region mask. Region masks can effectively control the pattern of surface details based on text input. + +match the text prompt but lack a specific pattern. By using a region mask and maintaining an activation ratio of $1/2$ throughout the total iterations as a warm-up stage, we achieve a reasonable tradeoff between mask-result alignment and generation diversity (see Appendix A.3). + +Effect of VDM Initialization. Our method demonstrates that user-specified VDMs can effectively control the volume and direction of generated geometric structures. Figure 10 shows that the results are high-quality and match the text descriptions well, such as the elf ear and pauldron. As this initializes the Laplacian term and steers the gradient flow in geometric structure generation, users can easily specify an initial VDM or choose a VDM template provided in our framework to generate expressive results. + +![](images/0296df2d3b99aa17fa1d0ba8d6f95e97acdf9cb17cfb7bfd10b1935845e1a98e.jpg) +Figure 10. Effect of VDM initialization. User-specified VDMs can help achieve the intended final effect of geometric structures by initializing the brush's volume and direction. + +# 4.4. Applications + +Once various VDMs are generated, users can use these brushes to meet diverse creative needs in modeling software. For example, they can apply VDM brushes for mesh stylization and a real-time iterative modeling process. + +Local-to-Global Mesh Stylization. Although mesh stylization is a complex task even for professional artists, combining different surface details allows users to achieve stylization quickly (see Appendix C.3). + +![](images/208b43d4fe5f555dc761423972359ec7717b635dd0351bd595818662e78c59c8.jpg) +Figure 11. Coarse to fine interactive modeling. By combining geometric structures brushes and surface details brushes for iterative sculpting in modeling software, users can rapidly create an expressive model from a plain shape (top left). + +Coarse-to-Fine Interactive Modeling. Unlike previous methods [4, 60] that require a lengthy optimization process for each edit and result in non-reusable outcomes, our generated VDM brushes can be directly used in modeling software. This enables users to apply the generated brushes easily and interactively (Figure 11). + +# 5. Conclusion + +We have presented Text2VDM, a novel framework for VDM brush generation from text. A VDM is a non-natural 2D image where each pixel stores a 3D displacement vector, making it challenging for existing T2I models to generate. Therefore, we treat VDM generation as diffusion-guided mesh deformation formulated as a form of Sobolev preconditioned optimization. To mitigate semantic coupling issues in SDS, we propose using weighted blending for prompt tokens, achieving high-quality brush generation. Moreover, we introduce two control methods, i.e., region and shape control, to meet customized requirements. The generated VDMs are directly compatible with mainstream modeling software, enabling various applications such as mesh stylization and interactive modeling. + +Limitations and Future Work. While our framework can generate high-quality VDM brushes, they may encounter multiview inconsistencies, a common issue introduced by SDS. View-consistent diffusion models like MV2MV [6] may be helpful to further handle this. Additionally, errors of the T2I models may guide wrong 3D generation, such as the wrong number of fingers in the generated skeleton hand. VDMs demonstrate that complex 3D models can be efficiently created using diverse reusable sculpting brushes. Future exploration of 3D generation through the assembly of modular components with similar design principles holds promising research value. + +# Acknowledgments + +This work was partially supported by the Guangdong Provincial Key Lab of Integrated Communication, Sensing and Computation for Ubiquitous Internet of Things #2023B1212010007. + +# References + +[1] Noam Aigerman, Kunal Gupta, Vladimir G. Kim, Siddhartha Chaudhuri, Jun Saito, and Thibault Groueix. Neural Jacobian Fields: Learning Intrinsic Mappings of Arbitrary Meshes. ACM Trans. Graph., 41(4), 2022. 2, 4 +[2] Thiemo Alldieck, Nikos Kolotouros, and Cristian Sminchisescu. Score Distillation Sampling with Learned Manifold Corrective. European Conference on Computer Vision, 2024. 3 +[3] Mohammadreza Armandpour, Ali Sadeghian, Huangjie Zheng, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the Negative Prompt Algorithm: Transform 2D Diffusion into 3D, alleviate Janus problem and Beyond. arXiv preprint 2304.04968, 2023. 3 +[4] Amir Barda, Vladimir G. Kim, Noam Aigerman, Amit H. Bermano, and Thibault Groueix. MagicClay: Sculpting Meshes with Generative Neural Fields. SIGGRAPH Asia (Conference track), 2024. 2, 8 +[5] Blender Foundation. Blender. https://www.blender.org, 2025. Accessed Mar 5, 2025. 2 +[6] Youcheng Cai, Runshi Li, and Ligang Liu. MV2MV: Multi-View Image Translation via View-Consistent Diffusion Models. ACM Trans. Graph., 2024. 8 +[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-Ecite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models. 42 (4), 2023. 7 +[8] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3D: Disentangling Geometry and Appearance for High-quality Text-to-3D Content Creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 2, 3 +[9] Dale Decatur, Itai Lang, and Rana Hanocka. 3D Highlighter: Localizing Regions on 3D Shapes via Text Descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20930-20939, 2023. 2 +[10] Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3D Paintbrush: Local Stylization of 3D Shapes with Cascaded Score Distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4473-4483, 2024. 2 +[11] William Gao, Noam Aigerman, Thibault Groueix, Vova Kim, and Rana Hanocka. TextDeformer: Geometry Manipulation Using Text Guidance. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2, 4, 5 +[12] Maxon Computer GMBH. ZBrush. https://www.maxon.net/zbrush, 1999. Accessed Oct 18, 2024. 2 +[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and + +Yoshua Bengio. Generative Adversarial Nets. Advances in Neural Information Processing Systems, 27, 2014. 3 +[14] Amir Hertz, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Deep Geometric Texture Synthesis. ACM Trans. Graph., 39(4), 2020. 3 +[15] Susung Hong, Donghoon Ahn, and Seungryong Kim. Debiasing Scores and Prompts of 2D Diffusion for View-consistent Text-to-3D Generation. In Neural Information Processing Systems, 2023. 3 +[16] Anita Hu, Nishkrit Desai, Hassan Abu Alhaija, Seung Wook Kim, and Maria Shugrina. Diffusion Texture Painting. New York, NY, USA, 2024. Association for Computing Machinery. 3 +[17] Teng Hu, Ran Yi, Haokun Zhu, Liang Liu, Jinlong Peng, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Stroke-based Neural Painting and Stylization with Dynamically Predicted Painting Region. New York, NY, USA, 2023. Association for Computing Machinery. 3 +[18] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free Score Distillation. In The Twelfth International Conference on Learning Representations, 2024. 2, 3 +[19] Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Björn Ommer. Rethinking Style Transfer: From Pixels to Parameterized Brushstrokes. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12191-12200, 2021. 3 +[20] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular Primitives for High-Performance Differentiable Rendering. ACM Transactions on Graphics, 39(6), 2020. 4 +[21] Beichen Li, Liang Shi, and Wojciech Matusik. End-to-end procedural material capture with proxy-free mixed-integer optimization. ACM Transactions on Graphics (TOG), 42(4): 1-15, 2023. 3 +[22] Beichen Li, Yiwei Hu, Paul Guerrero, Miloš Hasan, Liang Shi, Valentin Deschaintre, and Wojciech Matusik. Procedural material generation with reinforcement learning. ACM Transactions on Graphics (TOG), 43(6):1-14, 2024. 3 +[23] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast Text-to-3D with Sparse-View Generation and Large Reconstruction Model. arXiv preprint 2311.06214, 2023. 2 +[24] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. FocalDreamer: Text-Driven 3D Editing via Focal-Fusion Assembly. Proceedings of the AAAI Conference on Artificial Intelligence, 38:3279–3287, 2024. 2 +[25] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. LucidDreamer: Towards High-Fidelity Text-to-3D Generation via Interval Score Matching. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6517-6526, 2023. 2 +[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-Resolution Text-to-3D Content Creation. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300-309, 2023. 3 +[27] Feng-Lin Liu, Hongbo Fu, Yu-Kun Lai, and Lin Gao. SketchDream: Sketch-based Text-to-3D Generation and Editing. ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH 2024), 43(4), 2024. 2 +[28] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot One Image to 3D Object. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9264–9275, 2023. 3 +[29] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Ruifeng Deng, Xin Li, Errui Ding, and Hao Wang. Paint Transformer: Feed Forward Neural Painting with Stroke Prediction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6578-6587, 2021. 3 +[30] Xiaoxiao Long, Yuanchen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3D: Single Image to 3D Using Cross-Domain Diffusion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9970-9980, 2023. 2 +[31] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2Mesh: Text-Driven Neural Stylization for Meshes. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13482–13492, 2022. 2, 5 +[32] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2 +[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 2 +[34] Haoran Mo, Edgar Simo-Serra, Chengying Gao, Changqing Zou, and Ruomei Wang. General virtual sketching framework for vector line art. ACM Transactions on Graphics (TOG), 40:1-14, 2021. 3 +[35] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large Steps in Inverse Rendering of Geometry. ACM Trans. Graph., 40(6), 2021. 2, 3, 4 +[36] Openai. DALLE. https://openai.com/index/dall-e-3/, 2024. Accessed Oct 22, 2024. 2 +[37] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. In The Eleventh International Conference on Learning Representations, 2022. 2, 3, 4, 7 +[38] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. RichDreamer: A Generalizable Normal-Depth Diffusion Model for Detail Richness in Text-to-3D. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914–9925, 2023. 2 +[39] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng + +Bo, and Xiaoguang Han. RichDreamer: A Generalizable Normal-Depth Diffusion Model for Detail Richness in Text-to-3D. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914–9925, 2024. 6 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In International Conference on Machine Learning, 2021. 2, 6 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022. 2, 3 +[42] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards Zero-shot Text-to-shape Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18603-18613, 2022. 2 +[43] Artem Sevastopolsky, Philip-William Grassal, Simon Giebenhain, ShahRukh Athar, Luisa Verdoliva, and Matthias Niessner. Headcraft: Modeling high-detail shape variations for animated 3dmms, 2023. 3 +[44] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep Marching Tetrahedra: a Hybrid Representation for High-resolution 3D Shape Synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2 +[45] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view Diffusion for 3D Generation. In The Twelfth International Conference on Learning Representations, 2024. 3 +[46] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3D Neural Field Generation Using Triplane Diffusion. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2023. 2 +[47] Maria Shugrina, Chin-Ying Li, and Sanja Fidler. Neural Brushstroke Engine: Learning a Latent Style Space of Interactive Drawing Tools. ACM Trans. Graph., 41(6), 2022. 3 +[48] Damian Stewart. Compel: A Text Prompt Weighting and Blending Library. https://github.com/damian0815/compel, 2023. Accessed Oct 20, 2024. 5 +[49] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative Gaussian Splitting for Efficient 3D Content Creation, 2023. 2 +[50] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-It-3D: High-fidelity 3D Creation from A Single Image with Diffusion Prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22819–22829, 2023. 2, 3 +[51] Duotun Wang, Hengyu Meng, Zeyu Cai, Zhijing Shao, Qianxi Liu, Lin Wang, Mingming Fan, Xiaohang Zhan, and + +Zeyu Wang. HeadEvolver: Text to Head Avatars via Expressive and Attribute-Preserving Mesh Deformation. International Conference on 3D Vision, 2025. 4 +[52] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12619-12629, 2023. 3 +[53] Peihao Wang, Dejia Xu, Zhiwen Fan, Dilin Wang, Sreyas Mohan, Forrest N. Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu, Zhangyang Wang, and Vikas Chandra. Taming Mode Collapse in Score Distillation for Text-to-3D Generation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[54] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and Diverse Text-to-3D Generation with Variational Score Distillation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 3, 5 +[55] Xingguang Yan, Han-Hung Lee, Ziyu Wan, and Angel X. Chang. An Object is Worth 64x64 Pixels: Generating 3D Object via Image Diffusion, 2024. 3 +[56] Xiaofeng Yang, Yiwen Chen, Cheng Chen, Chi Zhang, Yi Xu, Xulei Yang, Fayao Liu, and Guosheng Lin. Learn to Optimize Denoising Scores for 3D Generation: A Unified and Improved Diffusion Prior on NeRF and 3D Gaussian Splatting. arXiv preprint 2312.04820, 2023. 3 +[57] Yuezhi Yang, Qimin Chen, Vladimir Kim, Siddhartha Chaudhuri, Qixing Huang, and Zhiqin Chen. GenVDM: Generating Vector Displacement Maps From a Single Image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3 +[58] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-Texture Synthesis via Deep Convolutional Texture Map Optimization and Physically-Based Rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5, 6 +[59] Xin Yu, Yuanchen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and Xiaojuan Qi. Text-to-3D with Classifier Score Distillation. arXiv preprint 2310.19415, 2023. 2, 3, 5, 7 +[60] Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. TIP-Editor: An Accurate 3D Editor Following Both Text-Prompts and Image-Prompts. ACM Trans. Graph., 43(4), 2024. 2, 8 +[61] Zhengxia Zou, Tianyang Shi, Shuang Qiu, Yi Yuan, and Zhenwei Shi. Stylized Neural Painting. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15684-15693, 2021. 3 \ No newline at end of file diff --git a/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/images.zip b/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4716d170a5731cd3df2fb89c9dc1d4daa7f4479d --- /dev/null +++ b/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d4538eb5305b3d82fa2827202719bd034c18fa53e57100afbb56ab52b32815d +size 537123 diff --git a/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/layout.json b/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b07683db063b6bc48b9f34f417b7ad1edc6a7e14 --- /dev/null +++ b/2025/Text2VDM_ Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting/layout.json @@ -0,0 +1,8460 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 160, + 103, + 451, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 103, + 451, + 140 + ], + "spans": [ + { + "bbox": [ + 160, + 103, + 451, + 140 + ], + "type": "text", + "content": "Text2VDM: Text to Vector Displacement Maps for Expressive and Interactive 3D Sculpting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "spans": [ + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "text", + "content": "Hengyu Meng" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "text", + "content": " Duotun Wang" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "text", + "content": " Zhijing Shao" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "text", + "content": " Ligang Liu" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "text", + "content": " Zeyu Wang" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology (Guangzhou) \n" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "text", + "content": "University of Science and Technology of China \n" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 80, + 160, + 531, + 219 + ], + "type": "text", + "content": "The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 65, + 245, + 548, + 430 + ], + "blocks": [ + { + "bbox": [ + 65, + 245, + 548, + 430 + ], + "lines": [ + { + "bbox": [ + 65, + 245, + 548, + 430 + ], + "spans": [ + { + "bbox": [ + 65, + 245, + 548, + 430 + ], + "type": "image", + "image_path": "8808fca1b5050677e7463aca4d4c257791061307f1993715e82713dfa58cbf92.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 436, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 54, + 436, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. Example VDM brushes generated by Text2VDM and sculpted models in Blender. Text2VDM can produce high-quality brushes for surface details (top row) and geometric structures (bottom row) from text input. Users can rapidly create an expressive model from a plain shape by directly applying these brushes in Blender. Yellow underlined text highlights semantics enhanced by our framework." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 480, + 200, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 480, + 200, + 493 + ], + "spans": [ + { + "bbox": [ + 151, + 480, + 200, + 493 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 505, + 297, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 505, + 297, + 697 + ], + "spans": [ + { + "bbox": [ + 54, + 505, + 297, + 697 + ], + "type": "text", + "content": "Professional 3D asset creation often requires diverse sculpting brushes to add surface details and geometric structures. Despite recent progress in 3D generation, producing reusable sculpting brushes compatible with artists' workflows remains an open and challenging problem. These sculpting brushes are typically represented as vector displacement maps (VDMs), which existing models cannot easily generate compared to natural images. This paper presents Text2VDM, a novel framework for text-to-VDM brush generation through the deformation of a dense planar mesh guided by score distillation sampling (SDS). The original SDS loss is designed for generating full objects and struggles with generating desirable sub-object structures from scratch in brush generation. We refer to this issue as semantic coupling, which we address by introducing weighted blending of prompt tokens to SDS, resulting in a" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 482, + 556, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 556, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 556, + 565 + ], + "type": "text", + "content": "more accurate target distribution and semantic guidance. Experiments demonstrate that Text2VDM can generate diverse, high-quality VDM brushes for sculpting surface details and geometric structures. Our generated brushes can be seamlessly integrated into mainstream modeling software, enabling various applications such as mesh stylization and real-time interactive modeling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 595, + 395, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 595, + 395, + 607 + ], + "spans": [ + { + "bbox": [ + 314, + 595, + 395, + 607 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Sculpting brushes are essential tools in 3D asset creation, as artists often require a variety of brushes to create surface details and geometric structures. In modeling software, 3D sculpting brushes are typically defined as vector displacement maps (VDMs). A VDM is a 2D image where each pixel stores a 3D displacement vector. Through these vectors, VDM brushes can create complex surface details, such as cracks and wood grain, or generate geometric structures" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 148, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 148, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 148, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16882" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "like ears and horns. This allows artists to apply the same geometric pattern iteratively while sculpting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 98, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 98, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 98, + 295, + 266 + ], + "type": "text", + "content": "Despite significant advances in text-to-image (T2I) [36, 41] and text-to-3D generation [23, 30, 37, 49, 50], existing methods are unsuitable for creating VDM brushes. We summarize the challenges as follows: 1) Since VDMs are not natural images (Figure 2), it is difficult for existing T2I models to generate them directly. 2) From a 3D perspective, a VDM represents mesh deformation through per-vertex displacement vectors from a dense planar mesh. Mapping any generated mesh to a dense planar mesh to create a VDM is non-trivial. 3) Sculpting brushes often involve sub-object structures, whereas most 3D generation methods can only generate full objects. Enabling users to accurately control the generation of sculpting brushes through text prompts in a semantically focused manner remains challenging." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 267, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 267, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 267, + 295, + 458 + ], + "type": "text", + "content": "To address the challenges of brush generation, we propose Text2VDM, a novel optimization-based framework that generates diverse and controllable VDM brushes from text input. Our approach does not generate VDMs directly from a T2I model. Instead, we address VDM brush generation from a 3D perspective by applying score distillation with a pre-trained T2I model to guide mesh deformation. Our framework supports three ways to initialize a base mesh through a zero-valued, spike-pattern, or user-specified VDM for custom shape control. For mesh deformation, we formulate a Sobolev preconditioned optimization [35] to maintain mesh quality with intrinsic smoothness. We also provide optional region control using a mask of activated mesh deformation, helping users obtain the intended brush effects. The normal maps of the mesh are then rasterized by a differentiable renderer for brush optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 652 + ], + "type": "text", + "content": "We observed that the standard score distillation sampling (SDS) [37] can lead to semantic coupling when supervising the generation of sub-object level structures due to the associated semantics caused by the noisy gradients from the full object. For example, a generated deer's horn should not be a full deer's head, or a generated beard should not include a nose. A straightforward solution is to use negative prompts [18, 59] to exclude undesired semantics, but our experiments show that this semantics suppression approach is ineffective in decoupling semantics and leads to an unstable optimization process. Instead, we propose to enhance the semantics of part-related words by applying weighted blending to the tokens in the prompt. This results in semantically focused text embedding, directing toward a more precise target distribution while reducing noisy gradients during optimization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Our experiments demonstrate that Text2VDM produces high-quality and diverse VDM brushes that can be directly integrated into mainstream modeling software, such as Blender [5] and ZBrush [12]. Compared to existing methods that directly generate full 3D models, our approach ad" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 320, + 69, + 434, + 169 + ], + "blocks": [ + { + "bbox": [ + 320, + 69, + 434, + 169 + ], + "lines": [ + { + "bbox": [ + 320, + 69, + 434, + 169 + ], + "spans": [ + { + "bbox": [ + 320, + 69, + 434, + 169 + ], + "type": "image", + "image_path": "9db4fae690336e980920165d4dd6898a39d5f74a7765dd310c2606c9b3ce2c64.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 169, + 421, + 178 + ], + "lines": [ + { + "bbox": [ + 345, + 169, + 421, + 178 + ], + "spans": [ + { + "bbox": [ + 345, + 169, + 421, + 178 + ], + "type": "text", + "content": "(a) Vector Displacement" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 186, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 220 + ], + "type": "text", + "content": "Figure 2. Difference between vector and standard displacement. The VDM enables full 3D vector displacement, while the height map only allows unidirectional standard displacement." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 549, + 168 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 549, + 168 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 549, + 168 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 549, + 168 + ], + "type": "image", + "image_path": "351cde89c79c3a88dab240689ff39d098aef6945e286f373d4bb3c14011f8a9a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 169, + 544, + 178 + ], + "lines": [ + { + "bbox": [ + 460, + 169, + 544, + 178 + ], + "spans": [ + { + "bbox": [ + 460, + 169, + 544, + 178 + ], + "type": "text", + "content": "(b) Standard Displacement" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 240, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 288 + ], + "type": "text", + "content": "dresses a different use case where brush-based user sculpting is desirable. This enables artists to interactively use a variety of brushes to sculpt diverse and expressive models from a plain shape." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 289, + 513, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 289, + 513, + 300 + ], + "spans": [ + { + "bbox": [ + 326, + 289, + 513, + 300 + ], + "type": "text", + "content": "This paper makes the following contributions:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 300, + 553, + 407 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 300, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 300, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 300, + 553, + 335 + ], + "type": "text", + "content": "- We first introduce the task of text-to-VDM brush generation, which is challenging to tackle directly using current text-to-image and text-to-3D methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 336, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 336, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 314, + 336, + 553, + 371 + ], + "type": "text", + "content": "- We propose Text2VDM, a novel framework for text-to-VDM brush generation that is readily compatible with artists' workflow of 3D asset creation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 372, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 372, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 314, + 372, + 553, + 407 + ], + "type": "text", + "content": "- We design a novel Semantic Enhancement SDS loss, which uses weighted blending to mitigate semantic coupling for sub-object structure generation." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 418, + 400, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 418, + 400, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 400, + 430 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 438, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 714 + ], + "type": "text", + "content": "Text to Local 3D Generation and Editing. With recent advances in diffusion models [41] and differentiable 3D representations [1, 33, 35, 44, 46], many methods for text-guided full 3D model generation have emerged [8, 11, 23, 25, 31, 38, 42]. Since 3D content creation is an iterative process that often requires user interaction, more attention has been directed toward localized 3D generation and editing. For example, 3D Highlighter [9] and 3D Paintbrush [10] use text as input, leveraging pre-trained CLIP models [40] or diffusion models [37] to supervise the optimization of neural networks for segmenting the regions of a 3D model that match the text description. Based on the information from these segmented regions, further editing of texture and geometry can be applied to the 3D model. Furthermore, SKED [32] and SketchDream [27] introduce sketches as an additional modality to assist in localized editing. To enable more precise control, FocalDreamer [24], MagicClay [4], and Tip-Editor [60] allow users to specify the editing location directly within the 3D space. These works rely on optimization-based methods to edit specific objects, often resulting in non-reusable editing outcomes. Additionally, each edit requires a lengthy optimization process, making interactivity difficult to achieve." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16883" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 74, + 545, + 289 + ], + "blocks": [ + { + "bbox": [ + 67, + 74, + 545, + 289 + ], + "lines": [ + { + "bbox": [ + 67, + 74, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 67, + 74, + 545, + 289 + ], + "type": "image", + "image_path": "f6e725640433d0a1fbaf19f0cdc17efc044134368f3d359e2927ed02a61c846d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 299, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 54, + 299, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 54, + 299, + 555, + 346 + ], + "type": "text", + "content": "Figure 3. Overview of Text2VDM. Starting with a dense planar mesh constructed from a zero-valued VDM, users can initialize the mesh volume through a default spike-pattern VDM or a user-specified VDM. Given the text prompt and region mask, we propose the Semantic Enhancement SDS loss " + }, + { + "bbox": [ + 54, + 299, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SE}}" + }, + { + "bbox": [ + 54, + 299, + 555, + 346 + ], + "type": "text", + "content": " to guide mesh deformations through a mesh Laplacian " + }, + { + "bbox": [ + 54, + 299, + 555, + 346 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 54, + 299, + 555, + 346 + ], + "type": "text", + "content": " iteratively, achieving semantically focused generation of surface details or geometric structures. After generation, vertex displacements are baked into the final VDM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 365, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 581 + ], + "type": "text", + "content": "Diffusion Priors for 3D Generation. Score distillation sampling (SDS) [37, 52] provides pixel-level guidance by seeking specific modes in a diffusion model, inspiring further research to improve optimization-based 3D generation [3, 26, 53, 54, 56]. Some studies focus on mitigating the \"Janus\" problem [2, 15], while others fine-tune diffusion models with multiview datasets to enhance 3D consistency [28, 45]. Recent research focuses on refining the design of SDS loss to achieve more precise guidance. For instance, Make-it-3D [50] introduces two-stage optimizations to improve textured appearance, while Fantasia3D [8] dynamically modifies the time-dependent weighting function within SDS computations. Additionally, several methods [18, 59] incorporate negative prompts as the conditional term to further refine the optimizations. Although diffusion priors have achieved promising results, their application in generating sub-object structures without global context as a reference is still challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Appearance and Geometric Brush Synthesis. The concept of brushes is very common in the creative process of digital artists, serving as a reusable local decorative unit. Appearance brushes focus on color representation and drawing styles in 2D space. With the development of generative models [13, 41], many works have explored the synthesis of procedural material [21, 22] for 3D object texturing and appearance brushes for interactive painting [16, 47], realistic artworks generation [29, 34, 61], and applying stylization [17, 19]. Unlike appearance brushes, geometric brushes focus on modifying geometry by moving the ver" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 365, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 510 + ], + "type": "text", + "content": "tices of a mesh in 3D space. VDM brushes, as an extension of standard geometric brushes, provide more complex geometric effects by utilizing VDMs. To the best of our knowledge, only a few techniques adopted the concepts of VDM for generation [43, 55] and geometric texture transfer [14]. Recently, concurrent work [57] explored using a single image as input and leveraged a diffusion model for generating multiview normal maps to guide VDM reconstruction. In comparison, our method is fundamentally different in tackling the semantic coupling problem arising from text guidance. This demonstrates that generating geometric brushes is a highly promising research direction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 513, + 398, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 398, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 398, + 528 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": "To generate VDM brushes compatible with mainstream modeling software, we begin by constructing a dense mesh from an initial VDM, as shown in Figure 3. We then apply score distillation with Stable Diffusion to guide high-quality mesh deformation formulated as a form of Sobolev preconditioned optimization [35]. To produce the intended sub-object level structure described in the text, we design a Semantic Enhancement SDS loss by applying weighted blending to the tokens in the prompt, effectively handling the issue of semantic coupling in SDS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 659, + 430, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 430, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 430, + 672 + ], + "type": "text", + "content": "3.1. Brush Initialization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "We provide three methods to initialize a base mesh for brush generation via a zero-valued VDM, a spike-pattern VDM, or a user-specified VDM. A VDM is represented as a " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "512 \\times" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16884" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "512 three-channel image, in which each channel stores the displacement in the X, Y, or Z direction, respectively. We first construct a planar grid mesh by creating two triangles for every " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " pixels and then apply the displacement stored in the VDM to mesh vertices. The values in these three initial VDMs range from 0 to 1, in which 0 represents no displacement, and 1 corresponds to half of the mesh's edge length in the positive axis direction. Since users can apply sculpting brushes symmetrically, our initial VDM does not need to store any negative values." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 324 + ], + "type": "text", + "content": "Our three methods for brush initialization facilitate the generation of diverse sculpting brush styles. The zero-valued VDM results in a planar mesh, which is our default setup when no control is provided. The spike-pattern VDM is suitable for generating protruding geometric structures, as it can effectively adjust the Laplacian term in Equation (2) to steer the gradient direction for mesh deformation. For better control of the brush's volume and direction, we also provide an interface for users to create custom VDMs, so the user-specified brush initialization can effectively guide mesh deformation toward the target structure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 331, + 270, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 270, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 270, + 344 + ], + "type": "text", + "content": "3.2. Brush Generation via Mesh Deformation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "text", + "content": "Given the vertices " + }, + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "text", + "content": " on the initialized base mesh, our method aims to learn a mesh deformation to the target brush shape. The vertex positions " + }, + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\hat{v}" + }, + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "text", + "content": " after mesh deformation can be expressed by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 408, + 295, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 408, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 117, + 408, + 295, + 428 + ], + "type": "interline_equation", + "content": "\\hat {v} = \\underset {v} {\\arg \\min } \\mathcal {L} _ {\\mathrm {S E}} \\left(\\mathcal {D} _ {c} (v), y\\right), \\tag {1}", + "image_path": "778a6549c572b349638ee86572e085deb6d857464466d5f076f4ce1e6058b07e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": " represents the camera setup in a differentiable renderer " + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": " [20]. The loss function " + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SE}}" + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": " receives the rendered normal image " + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c(v)" + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": " and text input " + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": " to evaluate the semantic guidance, which is detailed in Section 3.3. To accompany the external forces from " + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SE}}" + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": " that drive the mesh deformation with intrinsic smoothness energies, we follow the framework of a Sobolev preconditioned gradient descent [35], where the base mesh is reparameterized by the mesh Laplacian " + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 436, + 296, + 544 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 554, + 295, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 554, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 139, + 554, + 295, + 568 + ], + "type": "interline_equation", + "content": "v ^ {*} = (I + \\lambda L) v. \\tag {2}", + "image_path": "b54e1431390b94b4bbac6480cbcf48ebd06122d33705e7ba2ae0709b49d8f74f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 577, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 296, + 613 + ], + "type": "text", + "content": "This preconditioning involves solving a sparse linear system at every iteration, modifying the gradient descent update for each mesh deformation step to:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 114, + 621, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 621, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 114, + 621, + 295, + 647 + ], + "type": "interline_equation", + "content": "v \\leftarrow v - \\eta (I + \\lambda L) ^ {- 1} \\frac {\\partial \\mathcal {L} _ {\\mathrm {S E}}}{\\partial v}, \\tag {3}", + "image_path": "dbe3ab66a867cee105470bde7959d2d34e99c3e5988106116993c0f984e52bc3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " is the learning rate, " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " is the identity matrix, and " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " is a hyperparameter to control the extent of gradient diffusion over the entire domain. We set " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda = 15" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " throughout our experiments to balance the global structure and fine details during mesh deformation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "Compared to directly adding a Laplacian regularization term to Equation (1), the preconditioning framework [35] is critical in achieving large mesh deformation while maintaining proper topology with reduced triangle flips (Figure 4). Several works [11, 51] adopt the strategy by Aignerman et al. [1], parameterizing deformation through Jacobian fields that capture the scaling and rotation of each triangle. Although this method effectively smooths vertex displacements, the local deformation represented in Jacobians accumulates, leading to global drifting for open-boundary meshes, making it challenging to bake the mesh as a brush." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "text", + "content": "Additionally, we provide an optional region mask to restrict mesh deformation to the user-defined region during optimization. It helps maintain zero values in unused VDM areas when generating geometric structures. For producing surface details, the region mask ensures that the brush effect meets user-customized requirements (Figure 9)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 276, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 348 + ], + "type": "text", + "content": "By framing the VDM generation task as mesh deformation through preconditioned optimization, our framework preserves the favorable initial mesh topology and ensures control throughout the procedure. The resulting mesh preserves the original structure while incorporating rich local deformations, making it well-suited for baking as a brush." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 357, + 553, + 470 + ], + "blocks": [ + { + "bbox": [ + 317, + 357, + 553, + 470 + ], + "lines": [ + { + "bbox": [ + 317, + 357, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 357, + 553, + 470 + ], + "type": "image", + "image_path": "33eac003bceaa84327a88f3d78016b0acc15ccd2d09d32395fc717d33840340c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 477, + 555, + 521 + ], + "lines": [ + { + "bbox": [ + 313, + 477, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 555, + 521 + ], + "type": "text", + "content": "Figure 4. Effect of " + }, + { + "bbox": [ + 313, + 477, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 477, + 555, + 521 + ], + "type": "text", + "content": ". (a) Low-quality results w/o mesh smoothness, (b) low-quality results w/ direct Laplacian regularization, (c) our adopted preconditioning scheme with " + }, + { + "bbox": [ + 313, + 477, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\lambda = 15" + }, + { + "bbox": [ + 313, + 477, + 555, + 521 + ], + "type": "text", + "content": ", and (d) results with over-smoothness caused by a larger " + }, + { + "bbox": [ + 313, + 477, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 477, + 555, + 521 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 543, + 531, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 531, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 531, + 554 + ], + "type": "text", + "content": "3.3. Semantic Enhancement Score Distillation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "text", + "content": "Current text-to-3D generation methods like DreamFusion [37] often optimize a 3D representation parameterized by " + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "text", + "content": " so that rendered images " + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = g(\\theta)" + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "text", + "content": " resemble 2D samples produced by a pre-trained T2I diffusion model for a given text prompt " + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "text", + "content": " functions as a differentiable renderer. The T2I diffusion model " + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "text", + "content": " predicts the sampled noise " + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}(\\mathbf{x}_t; y, t)" + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "text", + "content": " of a rendered image " + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "text", + "content": " at a noise level " + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "text", + "content": " for the text input " + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "text", + "content": ". To make rendered images follow the text-conditioned distribution in Stable Diffusion, the SDS loss updates " + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 561, + 555, + 681 + ], + "type": "text", + "content": " by estimating the gradient:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 689, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 689, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 320, + 689, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S D S}} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon , c} \\left[ \\omega (t) \\left(\\epsilon_ {\\phi} (\\mathbf {x} _ {t}; y, t) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {4}", + "image_path": "0dcf839eec53f920a50424fe713dafcd64514a33932b2d4af5a5edf98c684ce6.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16885" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 266, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 266, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 266, + 84 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 266, + 84 + ], + "type": "inline_equation", + "content": "\\omega (t)" + }, + { + "bbox": [ + 55, + 72, + 266, + 84 + ], + "type": "text", + "content": " is a time-dependent weighting function." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 215 + ], + "type": "text", + "content": "However, the SDS loss cannot effectively supervise sub-object structure generation due to the issue of semantic coupling in full objects. For example, when using the SDS loss to generate a tortoise shell, it also usually generates the tortoise's tail and head, causing semantic coupling (Figure 7). We believe that the issue of semantic coupling in SDS stems from the training data of Stable Diffusion, in which most images contain full objects rather than separate parts. Therefore, the semantics of full objects often appear in the target distribution conditioned on text only describing sub-object structures." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 288 + ], + "type": "text", + "content": "A straightforward approach is using negative distributions via Classifier Score Distillation (CSD) [59] or Variational Score Distillation (VSD) [54] to suppress coupled semantics. CSD employs predefined negative prompts, resulting in more accurate negative distributions than those adaptively learned by VSD [59]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 293, + 295, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 293, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 87, + 293, + 295, + 332 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {C S D}} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon , c} [ (\\omega_ {\\mathrm {p o s}} \\cdot \\epsilon_ {\\phi} (\\mathbf {x} _ {t}; y, t) \\\\ \\left. - \\omega_ {\\text {n e g}} \\cdot \\epsilon_ {\\phi} \\left(\\mathbf {x} _ {t}; y _ {\\text {n e g}}, t\\right)\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} ], \\tag {5} \\\\ \\end{array}", + "image_path": "49fe96a4343de8e18d378f73fb18c3f5ad1c92eafb295e201340fb6d194a34d2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{pos}}" + }, + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{neg}}" + }, + { + "bbox": [ + 55, + 336, + 296, + 443 + ], + "type": "text", + "content": " denote different weights for positive and negative prompts. It requires two separate inferences with positive and negative prompts to obtain two distributions, which are then subtracted to suppress coupled semantics. However, our experiments show that CSD is ineffective in decoupling semantics because the negative prompt cannot accurately represent the undesirable coupled semantics in the positive prompt (Figure 7). This results in noisy guidance, making CSD less effective in decoupling semantics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 445, + 296, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 551 + ], + "type": "text", + "content": "Unlike semantic suppression in CSD, we propose a semantic enhancement method to mitigate semantic coupling by enhancing the semantics of part-related words. This can lead to a more accurate and stable target distribution, as shown in Section 4.3. Our key design is to apply weighted blending to the tokens in the original prompt to obtain a semantically focused text embedding, which serves as stable guidance for the optimization process. We define the Semantic Enhancement SDS loss as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 553, + 295, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 553, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 70, + 553, + 295, + 578 + ], + "type": "interline_equation", + "content": "\\nabla_ {\\theta} \\mathcal {L} _ {\\mathrm {S E}} (\\phi , \\mathbf {x}) = \\mathbb {E} _ {t, \\epsilon , c} \\left[ \\omega (t) \\left(\\epsilon_ {\\phi} ^ {*} (\\mathbf {x} _ {t}; y, t) - \\epsilon\\right) \\frac {\\partial \\mathbf {x}}{\\partial \\theta} \\right], \\tag {6}", + "image_path": "c828dbb3109b3bd8979189e7294cc2bd237c706d258d476a5354c6dde50150dc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\phi}^{*}(\\cdot)" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " uses a text embedding weighted by Compel [48]. Specifically, we assign a weight " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " to each word in the prompt and compute the weighted embedding " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{w}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " for each word by blending the original word embedding " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " and the empty text embedding " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "e_0" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{w}} = e_0 + s\\cdot (e - e_0)" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ". By concatenating the weighted embedding of each word in sequence, we obtain the final semantically focused text embedding. In our experiments, we found that using " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "s = 1.1^2" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " for words that require semantic enhancement can achieve stable optimization and effectively alleviate the issue of semantic coupling." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 359, + 107 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 359, + 107 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 359, + 107 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 359, + 107 + ], + "type": "image", + "image_path": "32ae9705644a6d017e563c1cd9860a71631b6ca5aae86fbd97317c96a65f39c3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 364, + 70, + 405, + 110 + ], + "blocks": [ + { + "bbox": [ + 364, + 70, + 405, + 110 + ], + "lines": [ + { + "bbox": [ + 364, + 70, + 405, + 110 + ], + "spans": [ + { + "bbox": [ + 364, + 70, + 405, + 110 + ], + "type": "image", + "image_path": "f27ac06eeb4621f121cd081556ec2d3db9348260927b25caf3219c37d711eb7e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 364, + 110, + 408, + 158 + ], + "blocks": [ + { + "bbox": [ + 364, + 110, + 408, + 158 + ], + "lines": [ + { + "bbox": [ + 364, + 110, + 408, + 158 + ], + "spans": [ + { + "bbox": [ + 364, + 110, + 408, + 158 + ], + "type": "image", + "image_path": "60e68b185ce5dc93b63e20d6455ebdab6a35f9b0df83f89b8c11c0448c971353.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 408, + 70, + 447, + 110 + ], + "blocks": [ + { + "bbox": [ + 408, + 70, + 447, + 110 + ], + "lines": [ + { + "bbox": [ + 408, + 70, + 447, + 110 + ], + "spans": [ + { + "bbox": [ + 408, + 70, + 447, + 110 + ], + "type": "image", + "image_path": "4282a32cf52a03466831d93baf340c1fc528ac46fea74ed26e4343bec8841c2d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 408, + 114, + 447, + 156 + ], + "blocks": [ + { + "bbox": [ + 408, + 114, + 447, + 156 + ], + "lines": [ + { + "bbox": [ + 408, + 114, + 447, + 156 + ], + "spans": [ + { + "bbox": [ + 408, + 114, + 447, + 156 + ], + "type": "image", + "image_path": "bb4a929c0f9e1a93864eb7e74df1588028f307f858761994e32d8d911d56f8ca.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 321, + 158, + 362, + 197 + ], + "blocks": [ + { + "bbox": [ + 321, + 158, + 362, + 197 + ], + "lines": [ + { + "bbox": [ + 321, + 158, + 362, + 197 + ], + "spans": [ + { + "bbox": [ + 321, + 158, + 362, + 197 + ], + "type": "image", + "image_path": "148f5918c3afdbff8627c7473e8b8896fb5cdc7a43e9740c06c29abdb8700778.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 364, + 158, + 405, + 198 + ], + "blocks": [ + { + "bbox": [ + 364, + 158, + 405, + 198 + ], + "lines": [ + { + "bbox": [ + 364, + 158, + 405, + 198 + ], + "spans": [ + { + "bbox": [ + 364, + 158, + 405, + 198 + ], + "type": "image", + "image_path": "339bd868ba412fd34b52a72ccc06e9d0577d2f292ff77a135a2fcbae511b9550.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 364, + 200, + 405, + 241 + ], + "blocks": [ + { + "bbox": [ + 364, + 200, + 405, + 241 + ], + "lines": [ + { + "bbox": [ + 364, + 200, + 405, + 241 + ], + "spans": [ + { + "bbox": [ + 364, + 200, + 405, + 241 + ], + "type": "image", + "image_path": "c45a80201127c8dad8be377ad86c137d1bd5ee47d5fa3b0f19a5802e6dab096c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 242, + 392, + 248 + ], + "lines": [ + { + "bbox": [ + 375, + 242, + 392, + 248 + ], + "spans": [ + { + "bbox": [ + 375, + 242, + 392, + 248 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 409, + 158, + 449, + 198 + ], + "blocks": [ + { + "bbox": [ + 409, + 158, + 449, + 198 + ], + "lines": [ + { + "bbox": [ + 409, + 158, + 449, + 198 + ], + "spans": [ + { + "bbox": [ + 409, + 158, + 449, + 198 + ], + "type": "image", + "image_path": "3397ba1251c1e967e5002ef250713d758bb8c310c6bce1245e19255ad3fde7d5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 410, + 200, + 449, + 240 + ], + "blocks": [ + { + "bbox": [ + 410, + 200, + 449, + 240 + ], + "lines": [ + { + "bbox": [ + 410, + 200, + 449, + 240 + ], + "spans": [ + { + "bbox": [ + 410, + 200, + 449, + 240 + ], + "type": "image", + "image_path": "0e0228dd3321b74f046295a20df5aac87d0b2d6311f97b4313c69627cd38b667.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 241, + 447, + 247 + ], + "lines": [ + { + "bbox": [ + 414, + 241, + 447, + 247 + ], + "spans": [ + { + "bbox": [ + 414, + 241, + 447, + 247 + ], + "type": "text", + "content": "Text2Mesh" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 454, + 70, + 495, + 110 + ], + "blocks": [ + { + "bbox": [ + 454, + 70, + 495, + 110 + ], + "lines": [ + { + "bbox": [ + 454, + 70, + 495, + 110 + ], + "spans": [ + { + "bbox": [ + 454, + 70, + 495, + 110 + ], + "type": "image", + "image_path": "f8cd0eba056fd7575174ebb4f8c19fc7d066445320ce7a82121e4da9731a19db.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 454, + 114, + 493, + 154 + ], + "blocks": [ + { + "bbox": [ + 454, + 114, + 493, + 154 + ], + "lines": [ + { + "bbox": [ + 454, + 114, + 493, + 154 + ], + "spans": [ + { + "bbox": [ + 454, + 114, + 493, + 154 + ], + "type": "image", + "image_path": "9d3fb6c910fab193f57055bfdc9e63c46cc09062e59eb82b9c253272845535f3.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 455, + 158, + 494, + 198 + ], + "blocks": [ + { + "bbox": [ + 455, + 158, + 494, + 198 + ], + "lines": [ + { + "bbox": [ + 455, + 158, + 494, + 198 + ], + "spans": [ + { + "bbox": [ + 455, + 158, + 494, + 198 + ], + "type": "image", + "image_path": "19856accc9267105194cd89ab83bafbe8e6905fab5aba1034dac29276dee11c5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 457, + 201, + 494, + 240 + ], + "blocks": [ + { + "bbox": [ + 457, + 201, + 494, + 240 + ], + "lines": [ + { + "bbox": [ + 457, + 201, + 494, + 240 + ], + "spans": [ + { + "bbox": [ + 457, + 201, + 494, + 240 + ], + "type": "image", + "image_path": "9caa1c0362d06a5b26b6e71cb1fabf4b7de490d3442f9d114c3f9945b0a6cce0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 241, + 488, + 248 + ], + "lines": [ + { + "bbox": [ + 465, + 241, + 488, + 248 + ], + "spans": [ + { + "bbox": [ + 465, + 241, + 488, + 248 + ], + "type": "text", + "content": "Paint-it" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 501, + 70, + 537, + 110 + ], + "blocks": [ + { + "bbox": [ + 501, + 70, + 537, + 110 + ], + "lines": [ + { + "bbox": [ + 501, + 70, + 537, + 110 + ], + "spans": [ + { + "bbox": [ + 501, + 70, + 537, + 110 + ], + "type": "image", + "image_path": "a4f63aedf8f1c54357004c942e8d84120957007958978fcf502a615f8724a107.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 504, + 114, + 539, + 155 + ], + "blocks": [ + { + "bbox": [ + 504, + 114, + 539, + 155 + ], + "lines": [ + { + "bbox": [ + 504, + 114, + 539, + 155 + ], + "spans": [ + { + "bbox": [ + 504, + 114, + 539, + 155 + ], + "type": "image", + "image_path": "527e8c7a94b256f4a08d6340dd3a708b83072e779451892ce944bd581d9655e7.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 501, + 159, + 540, + 201 + ], + "blocks": [ + { + "bbox": [ + 501, + 159, + 540, + 201 + ], + "lines": [ + { + "bbox": [ + 501, + 159, + 540, + 201 + ], + "spans": [ + { + "bbox": [ + 501, + 159, + 540, + 201 + ], + "type": "image", + "image_path": "1e19b63b09904c996cbb857f4e270e37afc7fe4b55a0ec45819238b6082dc7d0.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 501, + 204, + 544, + 241 + ], + "blocks": [ + { + "bbox": [ + 501, + 204, + 544, + 241 + ], + "lines": [ + { + "bbox": [ + 501, + 204, + 544, + 241 + ], + "spans": [ + { + "bbox": [ + 501, + 204, + 544, + 241 + ], + "type": "image", + "image_path": "46835acade887753b98a215374ec719ec1b5b9f6650e1d078abbf4ba2e5a7a3e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 241, + 542, + 248 + ], + "lines": [ + { + "bbox": [ + 501, + 241, + 542, + 248 + ], + "spans": [ + { + "bbox": [ + 501, + 241, + 542, + 248 + ], + "type": "text", + "content": "TextDeformer" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 257, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 300 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons of generated brushes for surface details. Our method captures geometry details guided by text, effectively preserving the surface structure and avoiding mesh distortion." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 320, + 395, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 395, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 395, + 335 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 341, + 554, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 413 + ], + "type": "text", + "content": "We conducted experiments to evaluate the various capabilities of Text2VDM both quantitatively and qualitatively for text-to-VDM brush generation. We then present an ablation study that validates the significance of our key insight into Semantic Enhancement SDS, as well as the effect of the region mask and VDM initialization." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 420, + 443, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 443, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 443, + 433 + ], + "type": "text", + "content": "4.1. Qualitative Evaluation" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 437, + 555, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 642 + ], + "type": "text", + "content": "To the best of our knowledge, Text2VDM is the first framework to generate VDM brushes from text. We adapted three existing methods for comparison and classified them into two categories. The first category includes Text2Mesh [31] and TextDeformer [11], which generate a brush mesh through text-guided mesh deformation on a planar mesh, following a process similar to ours. For the second category, we opt to directly generate VDM via Paint-it [58]. Notably, this method originally uses SDS to optimize a UNet for generating PBR textures. We reframed it to suit our VDM brush generation task, modifying it to generate VDM through SDS optimization of the UNet. In geometric structures generation experiment (Figure 6), all methods are compared fairly, with the same non-zero VDM initialization and mask applied to each prompt (see Appendix B.2). For surface details (Figure 5), all methods start with a zero-valued VDM and no masks to ensure a fair comparison." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Compared to other methods, Text2VDM can generate better-quality VDM brushes. Text2Mesh applies displacement to each vertex along normal directions, resulting in limited mesh deformation. TextDeformer indicates the accumulation of local deformations in the Jacobians, which results in global mesh drift, making it challenging to bake" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16886" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 545, + 386 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 545, + 386 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 545, + 386 + ], + "type": "image", + "image_path": "2724a6d7f4554fc538ecefda3d302064971f545c7a888bed05ee89898092d0f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 392, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 555, + 415 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons of generated brushes for geometric structures. Our method accurately presents key geometric features described by text, facilitating downstream applications in modeling software." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 436, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 509 + ], + "type": "text", + "content": "these meshes into VDM. Reframed Paint-it VDM generation is equivalent to optimizing the three-axis displacement of each vertex on the mesh with SDS. Although the UNet reduces noise from the SDS [58], smooth regularization is still required to ensure mesh quality, which makes achieving high-quality mesh generation quite challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 515, + 192, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 192, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 192, + 528 + ], + "type": "text", + "content": "4.2. Quantitative Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 534, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 569 + ], + "type": "text", + "content": "We quantitatively evaluated our framework regarding generation consistency with text input and mesh quality. We used 40 distinctive text prompts for VDM generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": "Generation Consistency with Text. We initially assessed the relevance of the generated results to the text descriptions [40]. 12 different views were rendered for average scores respectively, as presented in Table 1. Our approach achieves the highest scores compared to baseline methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Mesh Quality. We evaluated mesh quality by examining self-intersection. Paint-it and Text2Mesh, which utilize direct vertex displacement, often converge to a local minimum and disregard the mesh triangulation. While TextDeformer exhibits the lowest self-intersection, its tendency to produce over-smoothed results frequently results in losing object features described in text prompts." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 487, + 555, + 539 + ], + "blocks": [ + { + "bbox": [ + 313, + 435, + 555, + 478 + ], + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 478 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation of state-of-the-art methods. The geometry CLIP score is calculated on shaded images with uniform albedo colors [39], and self-intersection is quantified as the ratio of self-intersected mesh faces to the total number of faces." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 487, + 555, + 539 + ], + "lines": [ + { + "bbox": [ + 315, + 487, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 315, + 487, + 555, + 539 + ], + "type": "table", + "html": "
Geometry CLIP Score ↑Mesh Self-Intersection ↓
Paint-it0.237519.42%
Text2Mesh0.24977.18%
TextDeformer0.24770.04%
Ours0.25560.77%
", + "image_path": "f065ac1c7781d4bd8f609223644e1fd5d591b0050a087058e65fb3f2d7758953.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 321, + 570, + 547, + 621 + ], + "blocks": [ + { + "bbox": [ + 351, + 550, + 517, + 561 + ], + "lines": [ + { + "bbox": [ + 351, + 550, + 517, + 561 + ], + "spans": [ + { + "bbox": [ + 351, + 550, + 517, + 561 + ], + "type": "text", + "content": "Table 2. User evaluation of generated VDMs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 570, + 547, + 621 + ], + "lines": [ + { + "bbox": [ + 321, + 570, + 547, + 621 + ], + "spans": [ + { + "bbox": [ + 321, + 570, + 547, + 621 + ], + "type": "table", + "html": "
User Preference ↑Geometry QualityConsistency with Text
Paint-it3.1%1.7%
Text2Mesh18.3%27.3%
TextDeformer3.3%3.4%
Ours75.3%67.6%
", + "image_path": "f4516ac7e4878e636a83ec549597fad33587ebebb560246d63805b505ce1283b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "User Study. We further conducted a user study to evaluate the effectiveness and expressiveness of our method. A Google Form was utilized to assess 1) geometry quality and 2) consistency with text. We recruited 32 participants, of whom 14 are graduate students majoring in media arts, and 18 are company employees specializing in AI content gen" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16887" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 541, + 250 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 541, + 250 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 541, + 250 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 541, + 250 + ], + "type": "image", + "image_path": "a3b7ed331c8cb6a0a13ab499cf1eb6de03182366edcc4db96098b0f16658fb03.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 258, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 54, + 258, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 555, + 293 + ], + "type": "text", + "content": "Figure 7. Effect of Semantic Enhancement SDS. Our method effectively mitigates semantic coupling issues in SDS, such as generating the tortoise's tail and head or the deer's ear and mouth, by providing more focused semantic guidance. In contrast, the semantic suppression method used in CSD and the other semantic enhancement approach proposed by Attend-and-Excite both lead to an unstable optimization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 313, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 313, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 313, + 294, + 361 + ], + "type": "text", + "content": "eration. The participants were instructed to choose the preferred renderings of VDM from different methods in randomized order, as shown in Table 2. The results show participants preferred our method by a significant margin." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 370, + 149, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 149, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 149, + 383 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 389, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 556 + ], + "type": "text", + "content": "Effect of Semantic Enhancement SDS. Figure 7 compares the results generated by the original SDS [37], our Semantic Enhancement SDS with Compel, SDS with another semantic enhancement method in Attend-and-Excite (AaE) [7], and CSD [59] with two different annealed weights for the negative prompts: \"tortoise tail, tortoise head\" and \"deer's ear, deer's mouth.\" We also qualitatively compare these methods by visualizing their performance on semantic decoupling using the same Stable Diffusion for 2D image generation (Figure 8). Compared to semantic suppression using negative prompts and semantic enhancement by AaE, our design of introducing a semantically focused text embedding via Compel to SDS is most effective in resolving the issue of semantic coupling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Key Insight. As discussed in Section 3.3, SDS can result in semantic coupling when generating sub-object structures, leading to artifacts like the tortoise's tail and head or the deer's ear and deer's mouth. We observed that the semantics represented by negative prompts are also coupled. Therefore, meaningless semantics can emerge when applying semantic suppression in SDS, which leads to unstable optimization. Increasing the weight of negative prompts further reduces the overall quality of generated results. AaE enhances the cross-attention map of specific tokens by continuously updating the latent code. However, AaE is not suitable for the SDS framework because the latent code is affected by different camera poses and random noise at each" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 313, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 445 + ], + "type": "text", + "content": "iteration. This temporal randomness undermines the continuity of the latent code, resulting in unstable optimization and subpar results. In contrast, our method uses Compel to enhance semantics and achieves more effective semantic decoupling. Moreover, the semantically focused text embedding produced by Compel is independent of temporal variation. These properties help mitigate semantic coupling during SDS optimization, leading to high-quality sub-object geometric structures. Meanwhile, our method can also enhance the semantics in the text prompt for surface detail generation (see Appendix A.2)." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 455, + 555, + 517 + ], + "blocks": [ + { + "bbox": [ + 316, + 455, + 555, + 517 + ], + "lines": [ + { + "bbox": [ + 316, + 455, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 555, + 517 + ], + "type": "image", + "image_path": "3c51a0f93af5061df6dec77d828d0d07063b2bb03f89e2f86c3d9e4accc8885f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 518, + 553, + 578 + ], + "blocks": [ + { + "bbox": [ + 317, + 518, + 553, + 578 + ], + "lines": [ + { + "bbox": [ + 317, + 518, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 518, + 553, + 578 + ], + "type": "image", + "image_path": "9790439975bac725358d87ac5566b40000efe57edb8e33224a382539f041ce60.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 586, + 555, + 662 + ], + "lines": [ + { + "bbox": [ + 313, + 586, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 555, + 662 + ], + "type": "text", + "content": "Figure 8. Visualization of semantic decoupling performance. The images generated by the T2I model qualitatively visualize sub-object semantics. (a) Coupled semantics in positive and negative prompts lead to meaningless distributions when subtracted. (b) Enhanced semantics by Compel achieves stable decoupling. (c) AaE can only ensure that the enhanced semantics are preserved, but cannot decouple negative ones effectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Effect of Region Mask. Figure 9 demonstrates two region masks and their control over surface detail generation given a text prompt. Without a region mask, the results can still" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16888" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 298, + 149 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 298, + 149 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 298, + 149 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 298, + 149 + ], + "type": "image", + "image_path": "dac4f56f276654eda9dc7b45b186243d881d04b8ac6b1ba0811cd831b44ad150.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 156, + 295, + 179 + ], + "lines": [ + { + "bbox": [ + 55, + 156, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 295, + 179 + ], + "type": "text", + "content": "Figure 9. Effect of region mask. Region masks can effectively control the pattern of surface details based on text input." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 199, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 296, + 258 + ], + "type": "text", + "content": "match the text prompt but lack a specific pattern. By using a region mask and maintaining an activation ratio of " + }, + { + "bbox": [ + 55, + 199, + 296, + 258 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 55, + 199, + 296, + 258 + ], + "type": "text", + "content": " throughout the total iterations as a warm-up stage, we achieve a reasonable tradeoff between mask-result alignment and generation diversity (see Appendix A.3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 259, + 296, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 367 + ], + "type": "text", + "content": "Effect of VDM Initialization. Our method demonstrates that user-specified VDMs can effectively control the volume and direction of generated geometric structures. Figure 10 shows that the results are high-quality and match the text descriptions well, such as the elf ear and pauldron. As this initializes the Laplacian term and steers the gradient flow in geometric structure generation, users can easily specify an initial VDM or choose a VDM template provided in our framework to generate expressive results." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 375, + 298, + 540 + ], + "blocks": [ + { + "bbox": [ + 57, + 375, + 298, + 540 + ], + "lines": [ + { + "bbox": [ + 57, + 375, + 298, + 540 + ], + "spans": [ + { + "bbox": [ + 57, + 375, + 298, + 540 + ], + "type": "image", + "image_path": "0296df2d3b99aa17fa1d0ba8d6f95e97acdf9cb17cfb7bfd10b1935845e1a98e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 546, + 296, + 581 + ], + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 581 + ], + "type": "text", + "content": "Figure 10. Effect of VDM initialization. User-specified VDMs can help achieve the intended final effect of geometric structures by initializing the brush's volume and direction." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 137, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 137, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 137, + 612 + ], + "type": "text", + "content": "4.4. Applications" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 666 + ], + "type": "text", + "content": "Once various VDMs are generated, users can use these brushes to meet diverse creative needs in modeling software. For example, they can apply VDM brushes for mesh stylization and a real-time iterative modeling process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Local-to-Global Mesh Stylization. Although mesh stylization is a complex task even for professional artists, combining different surface details allows users to achieve stylization quickly (see Appendix C.3)." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 556, + 191 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 556, + 191 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 556, + 191 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 556, + 191 + ], + "type": "image", + "image_path": "208b43d4fe5f555dc761423972359ec7717b635dd0351bd595818662e78c59c8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 199, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 199, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 555, + 243 + ], + "type": "text", + "content": "Figure 11. Coarse to fine interactive modeling. By combining geometric structures brushes and surface details brushes for iterative sculpting in modeling software, users can rapidly create an expressive model from a plain shape (top left)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": "Coarse-to-Fine Interactive Modeling. Unlike previous methods [4, 60] that require a lengthy optimization process for each edit and result in non-reusable outcomes, our generated VDM brushes can be directly used in modeling software. This enables users to apply the generated brushes easily and interactively (Figure 11)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 371, + 388, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 371, + 388, + 384 + ], + "spans": [ + { + "bbox": [ + 314, + 371, + 388, + 384 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 397, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 564 + ], + "type": "text", + "content": "We have presented Text2VDM, a novel framework for VDM brush generation from text. A VDM is a non-natural 2D image where each pixel stores a 3D displacement vector, making it challenging for existing T2I models to generate. Therefore, we treat VDM generation as diffusion-guided mesh deformation formulated as a form of Sobolev preconditioned optimization. To mitigate semantic coupling issues in SDS, we propose using weighted blending for prompt tokens, achieving high-quality brush generation. Moreover, we introduce two control methods, i.e., region and shape control, to meet customized requirements. The generated VDMs are directly compatible with mainstream modeling software, enabling various applications such as mesh stylization and interactive modeling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Limitations and Future Work. While our framework can generate high-quality VDM brushes, they may encounter multiview inconsistencies, a common issue introduced by SDS. View-consistent diffusion models like MV2MV [6] may be helpful to further handle this. Additionally, errors of the T2I models may guide wrong 3D generation, such as the wrong number of fingers in the generated skeleton hand. VDMs demonstrate that complex 3D models can be efficiently created using diverse reusable sculpting brushes. Future exploration of 3D generation through the assembly of modular components with similar design principles holds promising research value." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16889" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 139 + ], + "type": "text", + "content": "This work was partially supported by the Guangdong Provincial Key Lab of Integrated Communication, Sensing and Computation for Ubiquitous Internet of Things #2023B1212010007." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 161 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 168, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 168, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 168, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 168, + 296, + 213 + ], + "type": "text", + "content": "[1] Noam Aigerman, Kunal Gupta, Vladimir G. Kim, Siddhartha Chaudhuri, Jun Saito, and Thibault Groueix. Neural Jacobian Fields: Learning Intrinsic Mappings of Arbitrary Meshes. ACM Trans. Graph., 41(4), 2022. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 213, + 296, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 213, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 61, + 213, + 296, + 257 + ], + "type": "text", + "content": "[2] Thiemo Alldieck, Nikos Kolotouros, and Cristian Sminchisescu. Score Distillation Sampling with Learned Manifold Corrective. European Conference on Computer Vision, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 258, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 258, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 61, + 258, + 296, + 312 + ], + "type": "text", + "content": "[3] Mohammadreza Armandpour, Ali Sadeghian, Huangjie Zheng, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the Negative Prompt Algorithm: Transform 2D Diffusion into 3D, alleviate Janus problem and Beyond. arXiv preprint 2304.04968, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 313, + 296, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 313, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 61, + 313, + 296, + 357 + ], + "type": "text", + "content": "[4] Amir Barda, Vladimir G. Kim, Noam Aigerman, Amit H. Bermano, and Thibault Groueix. MagicClay: Sculpting Meshes with Generative Neural Fields. SIGGRAPH Asia (Conference track), 2024. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 293, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 293, + 379 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 293, + 379 + ], + "type": "text", + "content": "[5] Blender Foundation. Blender. https://www.blender.org, 2025. Accessed Mar 5, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 413 + ], + "type": "text", + "content": "[6] Youcheng Cai, Runshi Li, and Ligang Liu. MV2MV: Multi-View Image Translation via View-Consistent Diffusion Models. ACM Trans. Graph., 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 414, + 296, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 296, + 457 + ], + "type": "text", + "content": "[7] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-Ecite: Attention-Based Semantic Guidance for Text-to-Image Diffusion Models. 42 (4), 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 459, + 296, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 296, + 513 + ], + "type": "text", + "content": "[8] Rui Chen, Yongwei Chen, Ningxin Jiao, and Kui Jia. Fantasia3D: Disentangling Geometry and Appearance for High-quality Text-to-3D Content Creation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 22246–22256, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "text", + "content": "[9] Dale Decatur, Itai Lang, and Rana Hanocka. 3D Highlighter: Localizing Regions on 3D Shapes via Text Descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20930-20939, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 558, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 296, + 612 + ], + "type": "text", + "content": "[10] Dale Decatur, Itai Lang, Kfir Aberman, and Rana Hanocka. 3D Paintbrush: Local Stylization of 3D Shapes with Cascaded Score Distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4473-4483, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 669 + ], + "type": "text", + "content": "[11] William Gao, Noam Aigerman, Thibault Groueix, Vova Kim, and Rana Hanocka. TextDeformer: Geometry Manipulation Using Text Guidance. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2, 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 293, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 293, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 293, + 690 + ], + "type": "text", + "content": "[12] Maxon Computer GMBH. ZBrush. https://www.maxon.net/zbrush, 1999. Accessed Oct 18, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "Yoshua Bengio. Generative Adversarial Nets. Advances in Neural Information Processing Systems, 27, 2014. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "type": "text", + "content": "[14] Amir Hertz, Rana Hanocka, Raja Giryes, and Daniel Cohen-Or. Deep Geometric Texture Synthesis. ACM Trans. Graph., 39(4), 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[15] Susung Hong, Donghoon Ahn, and Seungryong Kim. Debiasing Scores and Prompts of 2D Diffusion for View-consistent Text-to-3D Generation. In Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 173, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 217 + ], + "type": "text", + "content": "[16] Anita Hu, Nishkrit Desai, Hassan Abu Alhaija, Seung Wook Kim, and Maria Shugrina. Diffusion Texture Painting. New York, NY, USA, 2024. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 217, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 272 + ], + "type": "text", + "content": "[17] Teng Hu, Ran Yi, Haokun Zhu, Liang Liu, Jinlong Peng, Yabiao Wang, Chengjie Wang, and Lizhuang Ma. Stroke-based Neural Painting and Stylization with Dynamically Predicted Painting Region. New York, NY, USA, 2023. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 273, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 315 + ], + "type": "text", + "content": "[18] Oren Katzir, Or Patashnik, Daniel Cohen-Or, and Dani Lischinski. Noise-free Score Distillation. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "text", + "content": "[19] Dmytro Kotovenko, Matthias Wright, Arthur Heimbrecht, and Björn Ommer. Rethinking Style Transfer: From Pixels to Parameterized Brushstrokes. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12191-12200, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[20] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular Primitives for High-Performance Differentiable Rendering. ACM Transactions on Graphics, 39(6), 2020. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 459 + ], + "type": "text", + "content": "[21] Beichen Li, Liang Shi, and Wojciech Matusik. End-to-end procedural material capture with proxy-free mixed-integer optimization. ACM Transactions on Graphics (TOG), 42(4): 1-15, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 460, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 504 + ], + "type": "text", + "content": "[22] Beichen Li, Yiwei Hu, Paul Guerrero, Miloš Hasan, Liang Shi, Valentin Deschaintre, and Wojciech Matusik. Procedural material generation with reinforcement learning. ACM Transactions on Graphics (TOG), 43(6):1-14, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 504, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 558 + ], + "type": "text", + "content": "[23] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3D: Fast Text-to-3D with Sparse-View Generation and Large Reconstruction Model. arXiv preprint 2311.06214, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "text", + "content": "[24] Yuhan Li, Yishun Dou, Yue Shi, Yu Lei, Xuanhong Chen, Yi Zhang, Peng Zhou, and Bingbing Ni. FocalDreamer: Text-Driven 3D Editing via Focal-Fusion Assembly. Proceedings of the AAAI Conference on Artificial Intelligence, 38:3279–3287, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 614, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 669 + ], + "type": "text", + "content": "[25] Yixun Liang, Xin Yang, Jiantao Lin, Haodong Li, Xiaogang Xu, and Yingcong Chen. LucidDreamer: Towards High-Fidelity Text-to-3D Generation via Interval Score Matching. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6517-6526, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[26] Chen-Hsuan Lin, Jun Gao, Luming Tang, Towaki Takikawa, Xiaohui Zeng, Xun Huang, Karsten Kreis, Sanja Fidler, Ming-Yu Liu, and Tsung-Yi Lin. Magic3D: High-Resolution Text-to-3D Content Creation. In Proceedings of" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16890" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 300-309, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 296, + 140 + ], + "type": "text", + "content": "[27] Feng-Lin Liu, Hongbo Fu, Yu-Kun Lai, and Lin Gao. SketchDream: Sketch-based Text-to-3D Generation and Editing. ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH 2024), 43(4), 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 195 + ], + "type": "text", + "content": "[28] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot One Image to 3D Object. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9264–9275, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 252 + ], + "type": "text", + "content": "[29] Songhua Liu, Tianwei Lin, Dongliang He, Fu Li, Ruifeng Deng, Xin Li, Errui Ding, and Hao Wang. Paint Transformer: Feed Forward Neural Painting with Stroke Prediction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6578-6587, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 319 + ], + "type": "text", + "content": "[30] Xiaoxiao Long, Yuanchen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, and Wenping Wang. Wonder3D: Single Image to 3D Using Cross-Domain Diffusion. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9970-9980, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 295, + 374 + ], + "type": "text", + "content": "[31] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2Mesh: Text-Driven Neural Stylization for Meshes. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13482–13492, 2022. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 421 + ], + "type": "text", + "content": "[32] Aryan Mikaeili, Or Perel, Mehdi Safaee, Daniel Cohen-Or, and Ali Mahdavi-Amiri. Sked: Sketch-guided text-based 3d editing. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 476 + ], + "type": "text", + "content": "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 521 + ], + "type": "text", + "content": "[34] Haoran Mo, Edgar Simo-Serra, Chengying Gao, Changqing Zou, and Ruomei Wang. General virtual sketching framework for vector line art. ACM Transactions on Graphics (TOG), 40:1-14, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 522, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 295, + 555 + ], + "type": "text", + "content": "[35] Baptiste Nicolet, Alec Jacobson, and Wenzel Jakob. Large Steps in Inverse Rendering of Geometry. ACM Trans. Graph., 40(6), 2021. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 578 + ], + "type": "text", + "content": "[36] Openai. DALLE. https://openai.com/index/dall-e-3/, 2024. Accessed Oct 22, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 623 + ], + "type": "text", + "content": "[37] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. DreamFusion: Text-to-3D using 2D Diffusion. In The Eleventh International Conference on Learning Representations, 2022. 2, 3, 4, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 690 + ], + "type": "text", + "content": "[38] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng Bo, and Xiaoguang Han. RichDreamer: A Generalizable Normal-Depth Diffusion Model for Detail Richness in Text-to-3D. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914–9925, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[39] Lingteng Qiu, Guanying Chen, Xiaodong Gu, Qi Zuo, Mutian Xu, Yushuang Wu, Weihao Yuan, Zilong Dong, Liefeng" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 128 + ], + "type": "text", + "content": "Bo, and Xiaoguang Han. RichDreamer: A Generalizable Normal-Depth Diffusion Model for Detail Richness in Text-to-3D. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9914–9925, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 195 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In International Conference on Machine Learning, 2021. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 251 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 252, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 317 + ], + "type": "text", + "content": "[42] Aditya Sanghi, Hang Chu, Joseph G Lambourne, Ye Wang, Chin-Yi Cheng, Marco Fumero, and Kamal Rahimi Malekshan. Clip-forge: Towards Zero-shot Text-to-shape Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18603-18613, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 363 + ], + "type": "text", + "content": "[43] Artem Sevastopolsky, Philip-William Grassal, Simon Giebenhain, ShahRukh Athar, Luisa Verdoliva, and Matthias Niessner. Headcraft: Modeling high-detail shape variations for animated 3dmms, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 418 + ], + "type": "text", + "content": "[44] Tianchang Shen, Jun Gao, Kangxue Yin, Ming-Yu Liu, and Sanja Fidler. Deep Marching Tetrahedra: a Hybrid Representation for High-resolution 3D Shape Synthesis. Advances in Neural Information Processing Systems, 34:6087-6101, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 464 + ], + "type": "text", + "content": "[45] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view Diffusion for 3D Generation. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 466, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 521 + ], + "type": "text", + "content": "[46] J. Ryan Shue, Eric Ryan Chan, Ryan Po, Zachary Ankner, Jiajun Wu, and Gordon Wetzstein. 3D Neural Field Generation Using Triplane Diffusion. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20875-20886, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 522, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 564 + ], + "type": "text", + "content": "[47] Maria Shugrina, Chin-Ying Li, and Sanja Fidler. Neural Brushstroke Engine: Learning a Latent Style Space of Interactive Drawing Tools. ACM Trans. Graph., 41(6), 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "text", + "content": "[48] Damian Stewart. Compel: A Text Prompt Weighting and Blending Library. https://github.com/damian0815/compel, 2023. Accessed Oct 20, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "text", + "content": "[49] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. DreamGaussian: Generative Gaussian Splitting for Efficient 3D Content Creation, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "text", + "content": "[50] Junshu Tang, Tengfei Wang, Bo Zhang, Ting Zhang, Ran Yi, Lizhuang Ma, and Dong Chen. Make-It-3D: High-fidelity 3D Creation from A Single Image with Diffusion Prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22819–22829, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[51] Duotun Wang, Hengyu Meng, Zeyu Cai, Zhijing Shao, Qianxi Liu, Lin Wang, Mingming Fan, Xiaohang Zhan, and" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16891" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 642 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Zeyu Wang. HeadEvolver: Text to Head Avatars via Expressive and Attribute-Preserving Mesh Deformation. International Conference on 3D Vision, 2025. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 172 + ], + "type": "text", + "content": "[52] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A Yeh, and Greg Shakhnarovich. Score Jacobian Chaining: Lifting Pretrained 2D Diffusion Models for 3D Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12619-12629, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 239 + ], + "type": "text", + "content": "[53] Peihao Wang, Dejia Xu, Zhiwen Fan, Dilin Wang, Sreyas Mohan, Forrest N. Iandola, Rakesh Ranjan, Yilei Li, Qiang Liu, Zhangyang Wang, and Vikas Chandra. Taming Mode Collapse in Score Distillation for Text-to-3D Generation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 306 + ], + "type": "text", + "content": "[54] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. ProlificDreamer: High-fidelity and Diverse Text-to-3D Generation with Variational Score Distillation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 340 + ], + "type": "text", + "content": "[55] Xingguang Yan, Han-Hung Lee, Ziyu Wan, and Angel X. Chang. An Object is Worth 64x64 Pixels: Generating 3D Object via Image Diffusion, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "text", + "content": "[56] Xiaofeng Yang, Yiwen Chen, Cheng Chen, Chi Zhang, Yi Xu, Xulei Yang, Fayao Liu, and Guosheng Lin. Learn to Optimize Denoising Scores for 3D Generation: A Unified and Improved Diffusion Prior on NeRF and 3D Gaussian Splatting. arXiv preprint 2312.04820, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 397, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 397, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 397, + 294, + 452 + ], + "type": "text", + "content": "[57] Yuezhi Yang, Qimin Chen, Vladimir Kim, Siddhartha Chaudhuri, Qixing Huang, and Zhiqin Chen. GenVDM: Generating Vector Displacement Maps From a Single Image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 453, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 507 + ], + "type": "text", + "content": "[58] Kim Youwang, Tae-Hyun Oh, and Gerard Pons-Moll. Paint-it: Text-to-Texture Synthesis via Deep Convolutional Texture Map Optimization and Physically-Based Rendering. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 509, + 294, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 294, + 551 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 294, + 551 + ], + "type": "text", + "content": "[59] Xin Yu, Yuanchen Guo, Yangguang Li, Ding Liang, Song-Hai Zhang, and Xiaojuan Qi. Text-to-3D with Classifier Score Distillation. arXiv preprint 2310.19415, 2023. 2, 3, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 554, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 294, + 597 + ], + "type": "text", + "content": "[60] Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. TIP-Editor: An Accurate 3D Editor Following Both Text-Prompts and Image-Prompts. ACM Trans. Graph., 43(4), 2024. 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 598, + 294, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 598, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 598, + 294, + 642 + ], + "type": "text", + "content": "[61] Zhengxia Zou, Tianyang Shi, Shuang Qiu, Yi Yuan, and Zhenwei Shi. Stylized Neural Painting. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15684-15693, 2021. 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16892" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/948bb483-aee7-45ff-9c3e-421173343c17_content_list.json b/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/948bb483-aee7-45ff-9c3e-421173343c17_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4255d7ec1a8554806a3be7e18f11ea01e08e2519 --- /dev/null +++ b/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/948bb483-aee7-45ff-9c3e-421173343c17_content_list.json @@ -0,0 +1,1515 @@ +[ + { + "type": "text", + "text": "TextMaster: A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control", + "text_level": 1, + "bbox": [ + 194, + 130, + 805, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenyu Yan $^{1*}$ , Jian Wang $^{1*‡}$ , Aoqiang Wang $^{*†}$ , Yuhan Li $^{2†}$ , Wenxiang Shang $^{1}$ , Zhu Hangcheng $^{1}$ $^{1}$ Taobao & Tmall Group of Alibaba $\\quad^{2}$ Shanghai Jiao Tong University", + "bbox": [ + 119, + 202, + 875, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yeeshai.yzy, alex.wj}@taobao.com, wangaq1999@gmail.com, melodious@sjtu.edu.cn", + "bbox": [ + 151, + 243, + 839, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 292, + 326, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In image editing tasks, high-quality text editing capabilities can significantly reduce both human and material resource costs. Existing methods, however, face significant limitations in terms of stroke accuracy for complex text and controllability of generated text styles. To address these challenges, we propose TextMaster, a solution capable of accurately editing text across various scenarios and image regions, while ensuring proper layout and controllable text style. Our method enhances the accuracy and fidelity of text rendering by incorporating high-resolution standard glyph information and applying perceptual loss within the text editing region. Additionally, we leverage an attention mechanism to compute intermediate layer bounding box regression loss for each character, enabling the model to learn text layout across varying contexts. Furthermore, we propose a novel style injection technique that enables controllable style transfer for the injected text. Through comprehensive experiments, we demonstrate the state-of-the-art performance of our method.", + "bbox": [ + 88, + 325, + 485, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 659, + 220, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in generative models have revolutionized image synthesis, with diffusion models surpassing Generative Adversarial Networks (GANs) in quality, diversity, and controllability [11, 27, 30]. While GAN-based models, such as the original Generative Adversarial Nets [9], have demonstrated impressive realism, they often suffer from training instability and mode collapse. The advent of diffusion models, such as Denoising Diffusion Probabilistic Models (DDPM) [11] and Latent Diffusion Models (LDM) [30], have led to superior image fidelity and text-image alignment, making them the dominant paradigm in", + "bbox": [ + 89, + 685, + 483, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f521f04b12ced89c7d4c20bdaaa0ab70c45795465961721afc5126725215c03d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 297, + 897, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1a40186cd70b72b90d97f626b0fe13d1305df208a6912142f1d40a424713c8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 377, + 895, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e6a607f42e273fcb7aee7c234db8344fc2a5e345eb55bbc0bd435f6932551bcf.jpg", + "image_caption": [ + "Figure 1. The image illustrates the diverse capabilities of our TextMaster, encompassing precise typesetting and layout, consistent style retention, and external style injection." + ], + "image_footnote": [], + "bbox": [ + 531, + 462, + 895, + 563 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "text-to-image (T2I) generation.", + "bbox": [ + 511, + 655, + 720, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To enhance the controllability of T2I generation, researchers have explored methods to inject conditional signals. ControlNet [48] enables spatial constraints by conditioning diffusion models on structural priors, while T2I-Adapter [24] further enhances control by learning task-specific adapters. Similarly,Composer [13] introduces a framework for flexible, composable conditions, allowing intricate constraints in text-guided image synthesis.", + "bbox": [ + 509, + 671, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Another critical direction involves capturing specific styles and characteristics from reference images. Techniques such as Textual Inversion [8] and DreamBooth [31] fine-tune models on personalized concepts, improving fidelity to target attributes. IP-Adapter [45] extends this by efficiently integrating image prompts with textual guidance, while Reference-Efficient [12] and InstantID [41] enable", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. \n†Work done during an internship at Alibaba Group. \n‡Corresponding author.", + "bbox": [ + 89, + 862, + 364, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16112", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c0ac696c6ddcba5add81456471b5bdb237679799fd5bc28ccae4aee133215bc1.jpg", + "image_caption": [ + "Figure 2. The current approach exhibits several deficiencies, including a lack of alignment between the generated text and the intended target text, suboptimal layout organization, and inconsistent in style preservation." + ], + "image_footnote": [], + "bbox": [ + 173, + 92, + 823, + 429 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "identity preservation in personalized generation. Further advancements, such as PhotoMaker [19], refine identity embeddings to achieve photorealistic results.", + "bbox": [ + 89, + 501, + 482, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main challenge of text editing lies in generating accurate textual content while minimizing any noticeable traces of editing. To Achieve this requires, the model needs to reconstruct multiple densely packed characters within localized regions, which places substantial demands on its fine-grained modeling abilities. Recent methods [22, 43] have explored glyph control for text editing. Yet, these approaches often suffer from reduced editing accuracy due to the lack of character-level textual guidance. On the other hand, effectively disentangling content from style presents another significant hurdle. Due to the highly structured nature of character glyphs, independently extracting style features from a given text image remains particularly difficult. Diffusion-based methods [2, 4, 39] attempt to restore text by leveraging the image context to preserve style. Since style information is primarily derived from unmasked regions, this may lead to inconsistencies in style representation during reconstruction. Other approaches [15, 35] seek to disentangle text style from glyph structure using manually constructed datasets, where different texts are rendered with various fonts to create style-content pairs for learning. Despite being designed to simulate real-world scenarios, such datasets still fall short in capturing the authenticity", + "bbox": [ + 89, + 553, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and diversity found in real text images. Furthermore, some methods opt to inject reference styles through text-style image inversion [33, 47], but this comes at the cost of significantly increased inference overhead.", + "bbox": [ + 511, + 501, + 906, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to these challenges, our method introduces further improvements in glyph control by combining character-level text embedding and IoU loss enabling precise regulation of both glyph shape and positioning. For style disentanglement, unlike previous approaches that rely on synthesized data, whose results are often affected by the glyphs in the reference image, we explicitly separate the glyph structure from the style appearance, which is achieved by integrating a style adapter with in-context learning. These innovations enable our model to reconstruct text across deiverse and fully authentic datasets, ultimately delivering robust style disentanglement and superior generalization performance.", + "bbox": [ + 511, + 565, + 908, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By integrating these advanced techniques, the proposed TextMaster framework addresses the critical challenges in text editing, offering unparalleled performance in maintaining style consistency, content accuracy, proper layout, and high-quality visual output. These advancements position TextMaster as a state-of-the-art tool for complex text editing tasks in images [5, 11, 32]. Meanwhile, Our approach is not limited to some specific languages, it can be easily extended to other languages as well.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16113", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 232, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent advancements in generative models, such as FLUX [1], DALL-E 3 [25], and SD3 [7], have significantly improved image synthesis quality, achieving high-resolution fidelity and enhanced semantic coherence. However, challenges remain in accurate text rendering within images. Specifically, ensuring precise glyph structure, spatial consistency, and stylistic control remains a complex task.", + "bbox": [ + 89, + 122, + 480, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Accurate Glyph and Spatial Control. Several studies have addressed the challenges of generating visually coherent and spatially accurate text. GlyphDraw [22] proposes a word-level synthesis approach to ensure coherent Chinese character rendering. GlyphControl [43] introduces glyph-conditioned diffusion to enhance character structure control, while TextDiffuser [4] treats text rendering as a diffusion-based painting task. Recent advancements, such as AnyText [39], extend this to multilingual text generation. Further improvements, such as TextDiffuser-2 [3], leverage large language models for refined text placement. Additionally, models like Glyph-ByT5 [20] and its successor Glyph-ByT5-v2 [21] introduce customized text encoders for multilingual and aesthetic text rendering, improving both accuracy and flexibility. Moreover, GlyphDraw2 [23] combines diffusion models with large-scale language modeling to generate complex text-based posters.", + "bbox": [ + 89, + 234, + 482, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllable Text Style and Appearance. Separating textual content from stylistic attributes is a longstanding challenge. DiffSTE [15] improves text editing by leveraging dual encoders trained on extensive font datasets, allowing font-specific style preservation during synthesis. However, inference still requires explicit font name input. AnyText2 [38] enhances controllability by embedding RGB-based color attributes, but complex styles such as gradients, borders, and shadows remain difficult to represent effectively.", + "bbox": [ + 89, + 500, + 482, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 686, + 181, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text editing tasks can be deconstructed into two aspects: glyphs and styles. As shown in the figure 3, to achieve precise editing control, we propose Glyph-Style Dual-Control. Glyph control involves maintaining structural integrity of characters through glyph constraints, and achieving cohesive text flow via adaptive typography layouts. To maintain a consistent style with the original text, we introduce style control, which includes adjustments for typeface and color, allowing for accurate referencing and injection of text styles. This dual mechanism ensures precise character rendering, harmonized layout composition, and versatile style referencing capabilities.", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 513, + 90, + 643, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed method builds upon Stable Diffusion [29] Inpainting Models. Given an image $x_0$ , a binary mask $m$ specifying the editable region and a textual prompt $P$ describing the target text to generate, the system first encodes both the original image $x_0$ and its masked image $x_0 \\odot (1 - m)$ into latent space using a Variational Autoencoder (VAE) [17], where $\\odot$ denotes element-wise multiplication. Then a conditional latent diffusion model performs sequential denoising with a UNet in the latent space. UNet takes as inputs the noise latent variable $\\hat{z}_t$ , downsampled mask $m'$ , masked image latent $z_{Im}$ , timestep $t$ , and embedding $e_t$ of the textual prompt $P$ from a pre-trained CLIP [28] model, and predicts the noise $\\epsilon_{\\theta}(\\hat{z}_t, z_{Im}, m', t, e_t)$ . The training of diffusion models resolves around optimizing the denoiser network $\\epsilon_{\\theta}$ to minimize the noise prediction loss:", + "bbox": [ + 511, + 112, + 906, + 339 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o i s e}} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\hat {z} _ {t}, z _ {I m}, m ^ {\\prime}, t, e _ {t}\\right) \\right\\| _ {2} ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 348, + 906, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Glyph control", + "text_level": 1, + "bbox": [ + 511, + 376, + 658, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Glyph Constraint", + "text_level": 1, + "bbox": [ + 511, + 398, + 683, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The accurate rendering of glyphs is critical in text editing tasks. To ensure proper formation of character strokes, we have implemented multi-faceted constraints including reference standard character images, precise textual encoding representations, and perceptual loss mechanisms.", + "bbox": [ + 511, + 417, + 905, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Glyph Guidance. Previous work [22, 39, 44] has demonstrated the effective guidance provided by standard font reference images for text content generation. However, these methods design glyph images with per-char alignment positions to accommodate the generation of arbitrary shape masks. Unfortunately, the shape of these masks may unintentionally reveal textual details, such as character count and layout (see Fig.2 problem 1). Therefore, we calculate the outer rectangle of the edit area and inject glyph information by field alignment. The standard quadrilateral mask serves as a better foundation for glyph representation, as a well-coupled relationship between glyphs and layout fosters harmonious content generation.", + "bbox": [ + 511, + 493, + 905, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In detail, we first obtain the bounding box of the text area to be injected and generate the corresponding glyph in this area with standard font in appropriate size. Next, we convert the generated glyph image $\\mathcal{I}_g$ into latent space variables $l_{g}$ using the VAE [17] model. Finally, we concatenate these latent space variables with the latent representation of the image along the channel dimension to ensure that the generated image visually aligns with the original input image while clearly retaining and conveying the glyph information.", + "bbox": [ + 511, + 689, + 905, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Token-separated Text Encoder. The default text encoder of SDXL [27] is not sufficient for representing Chinese semantics. To address this limitations, we employ the ChatGLM [6] text encoder to support both Chinese and English", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16114", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6d330d78d85d480740cb3b849c637df82b4a7eb8e5c53681b849ed6af5e12a27.jpg", + "image_caption": [ + "Figure 3. The framework of TextMaster consists of the following components: the glyph control module, comprising the glyph constraint, update text encoder, and perception module; the layout control is achieved via position-aware attention; and the style control module encompasses the style-adapter and in-context reference." + ], + "image_footnote": [], + "bbox": [ + 174, + 95, + 826, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "characters. ChatGLM employs a dual-stream self-attention architecture: one stream processes positional information to capture text sequence and structure, while the other handles content information to understand semantic relationships.", + "bbox": [ + 88, + 477, + 482, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In text editing tasks, semantic relationships can adversely affect generation outcomes, as character embeddings may differ based on context. To reduce this interference, we adopt a specialized encoding method. First, we tokenize the prompt $P = \\{c_{1}, c_{2}, \\dots, c_{K}\\}$ by treating each character $c_{k} \\in P$ as an individual token, preventing cross-semantic influence. We then sort and concatenate these tokens according to their original positions, using zero padding to ensure consistent sequence lengths. Finally, we use a text encoder $\\mathcal{E}_{t}$ to encode the characters token by token, incorporating position information $p_{k}$ , and generate the corresponding mask encoding $m_{k}$ for each token $c_{k}$ . Each character is encoded as follows:", + "bbox": [ + 88, + 537, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {k} = \\mathcal {E} _ {t} \\left(c _ {k} + \\operatorname {p o s} _ {k} + m _ {k}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 744, + 482, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k\\in \\{1,2,\\dots ,K\\}$ . We obtain the final text conditional features by concatenating the character encodings:", + "bbox": [ + 89, + 772, + 482, + 803 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {t} = \\left[ e _ {1}; e _ {2}; \\dots ; e _ {K} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 816, + 482, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This approach maintains uniform character embeddings across contexts while preserving positional information, ultimately improving the generation quality in text editing tasks.", + "bbox": [ + 89, + 839, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Perception Module. We propose a method based on perceptual loss to maintain the consistency of glyph structures. By calculating the $L_{2}$ loss of the Optical Character Recognition (OCR) features in the edited region between the reconstructed image $\\hat{x}_0$ and the original image $x_0$ , we impose constraints on the generation of glyph structures.", + "bbox": [ + 511, + 477, + 903, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, by denoising the noise prediction $\\epsilon_{t}$ from the denoising network's output, we can derive an estimate $\\hat{z}_0$ from the noise vector $z_{t}$ as follows: $\\hat{z}_{0} = \\frac{1}{\\sqrt{\\bar{\\alpha}_{t}}} z_{t} + \\frac{\\sqrt{1 - \\bar{\\alpha}_{t}}}{\\sqrt{\\bar{\\alpha}_{t}}}\\epsilon_{t}$ , where $\\epsilon_{t} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , and $\\alpha_{t}$ denotes the variance scheduling of the Gaussian noise. Furthermore, we utilize a VAE decoder to obtain an approximate reconstruction estimate $\\hat{x}_0$ . Based on the mask $m$ of the edited region, we crop the original image $x_0$ and the reconstructed image $\\hat{x}_0$ , resulting in the corresponding images of the edited area $x_0^m$ and $\\hat{x}_0^m$ , respectively. We employ the PPOCR [18] model as the glyph encoder $\\phi (\\cdot)$ and extract features $f^{m} = \\phi (x_{0}^{m})$ and $\\hat{f}^{m} = \\phi (\\hat{x}_{0}^{m})$ from the fully connected layer to compute the L2 loss of the glyph features between the original and predicted images.", + "bbox": [ + 511, + 568, + 906, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, we found it necessary to ensure a natural transition between the text and the surrounding background. Therefore, we further introduce pixel-based Mean Squared Error (MSE) loss to ensure the accuracy of the text generation area. The resulting perceptual loss is as follows:", + "bbox": [ + 511, + 789, + 908, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\left\\| f ^ {m} - \\hat {f} ^ {m} \\left\\| _ {2} ^ {2} + \\left\\| x _ {0} ^ {m} - \\hat {x} _ {0} ^ {m} \\right\\| _ {2}. \\right. \\right. \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 877, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16115", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Perceptual loss injects a stronger supervisory signal by constraining the reconstruction of the editing region, ensuring the correct presentation of the text.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/62d7a394b1e974819a58e14e07b3cb5077d36506a8cb893199672fef6d19d767.jpg", + "image_caption": [ + "Figure 4. The attention response map processing flow: Atten represent the top chosen attention response maps for each token." + ], + "image_footnote": [], + "bbox": [ + 106, + 157, + 460, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Adaptive Typography Layout", + "text_level": 1, + "bbox": [ + 89, + 430, + 344, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Position-aware Attention. We expect the model to exhibit accurate positional responses for each character within the masked regions, thereby generating text sequences with proper typographic arrangement. As revealed in prior work [40], different cross-attention blocks demonstrate distinct attention orientations, with certain layers exhibiting stronger responses to layout patterns. As illustrated in Fig. 4, our analysis of text generation tasks shows that specific attention layers demonstrate pronounced perception of character positioning. To optimize layout learning, we calculate the IoU overlap ratio between binarized attention maps and ground-truth character bounding boxes, subsequently selecting the top three layers exhibiting the strongest responses for layout adaptation. This design facilitates improved intra-paragraph layout learning, as illustrated in Fig. 2 problem 2. Specifically, we perform the following steps:", + "bbox": [ + 88, + 449, + 482, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compute attention scores $\\mathbf{A}$ using the query matrix $\\mathbf{Q}$ and key matrix $\\mathbf{K}$ :", + "bbox": [ + 89, + 705, + 482, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} = \\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {d _ {k}}}, \\quad \\mathbf {W} = \\operatorname {s o f t m a x} (\\mathbf {A}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 746, + 482, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thresholding: Generate a binary image $\\mathbf{T}$ by applying a threshold $T = 125$ to the averaged attention map:", + "bbox": [ + 89, + 792, + 483, + 823 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} (i, j) = \\left\\{ \\begin{array}{l l} 2 5 5 & \\text {i f} \\bar {\\mathbf {A}} (i, j) > T \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 834, + 482, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian Blurring: Apply Gaussian blurring to the", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "thresholded image $\\mathbf{T}$ using a Gaussian kernel $\\mathbf{W}(k,l)$ :", + "bbox": [ + 511, + 90, + 877, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} (i, j) = \\sum_ {k = - K} ^ {K} \\sum_ {l = - L} ^ {L} \\mathbf {W} (k, l) \\cdot \\mathbf {T} (i + k, j + l). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 117, + 906, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Complete Intersection over Union (CIOU) Loss: Calculate the CIOU loss between the derived bounding box $\\mathbf{B}_d$ and the ground truth bounding box $\\mathbf{B}_g$ :", + "bbox": [ + 511, + 169, + 906, + 215 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {i o u}} = 1 - \\left(\\operatorname {I o U} - \\frac {\\rho^ {2} \\left(\\mathbf {B} _ {d} , \\mathbf {B} _ {g}\\right)}{c ^ {2}} - \\alpha v\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 226, + 906, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where:", + "text_level": 1, + "bbox": [ + 513, + 268, + 566, + 282 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- IoU: Intersection over Union of the two bounding boxes.", + "- $\\rho(\\mathbf{B}_d, \\mathbf{B}_g)$ : Euclidean distance between the centers.", + "- $c$ : Diagonal length of the smallest enclosing box covering both bounding boxes.", + "- $\\alpha$ : Trade-off parameter.", + "- $v$ : Measure of aspect ratio consistency." + ], + "bbox": [ + 513, + 284, + 903, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive Crop and Mask Strategy. Our experiments revealed that VAE-encoded latent representations often degrade glyph structures in extremely small text. To improve multi-scale generalization, we developed an Adaptive Crop strategy that dynamically adjusts text regions. In our experimental setup, we configured the threshold value at 9 characters. For text segments shorter than this threshold, we performed center-aligned cropping on the images and scaled them to the proportional size corresponding to a 9-character width. When the text length exceeded the threshold, we expanded the cropped region at a 1/10 scaling ratio. As demonstrated in Fig. 5, this threshold-triggered spatial adaptation significantly improves background restoration and visual coherence.", + "bbox": [ + 511, + 375, + 906, + 585 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/577344334d06b8b6a665ee658f4784dd39f653ea950c27affc2ade2c675e9d4b.jpg", + "image_caption": [ + "Prompt:\"健身房\"(Gym)" + ], + "image_footnote": [], + "bbox": [ + 532, + 603, + 643, + 693 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0e838a26693f587f0bc2e7645d88070c43ae949eec390075d2db917dcd44b3c1.jpg", + "image_caption": [ + "Figure 5. The text modification area depicted in the image is merely 30 pixels in height; TextMaster is still capable of executing high-quality edits within this constrained space." + ], + "image_footnote": [], + "bbox": [ + 663, + 603, + 771, + 693 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/442ef971a276f67010e4c7cd5fc45428db240cf307504b5b8c56409586ee87e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 603, + 883, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, to address uncontrollable text generation caused by mask size leakage (Fig. 2 problem 1) and mitigate layout issues from over-reliance on precise bounding dimensions, we propose an adaptive mask augmentation technique. By randomly expanding masked regions around text, this approach enhances robustness to bounding variations while improving the model's capability to generate visually coherent text layouts.", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16116", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Style Injection", + "text_level": 1, + "bbox": [ + 91, + 90, + 238, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Current methods prioritize text accuracy and layout over style consistency. Preserving original text styles enhances visual harmony with image contexts, while integrating external references broadens editing versatility. Although recent works [15, 38] employ style control via datasets artificially constructed from font libraries (conditioned on basic attributes like color and typeface), such synthetic datasets often mismatch real-world distributions and provide coarse-grained control, failing to reproduce complex styles like gradient colors, outlined text, or 3D effects.(Fig. 2 problem 3) To overcome these limitations, we propose a Style-Adapter module with In-Context Learning for precise style manipulation.", + "bbox": [ + 89, + 113, + 483, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.1. Style-Adapter", + "text_level": 1, + "bbox": [ + 89, + 320, + 235, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In style transfer tasks, decoupling style from content is a critical challenge. Text inherently carries strong glyph information, and the segmentation boundary information carried by the style often disrupts the generation results. InstanceStyle [40] decouples image objects and their semantic content. Inspired by this, we aim to force the model to decouple font style from glyph content.", + "bbox": [ + 89, + 339, + 483, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, to enable the model to extract more general style information rather than strictly aligning with the segmentation boundaries of the generated text, we randomly select $n, n \\in \\{1, 2, \\dots, K\\}$ words from the edit region of the original image $x_0$ and the glyph $\\mathcal{G}$ , resulting in cropped images $x_0^n$ and $\\mathcal{G}^n$ . We further apply a text segmentation model $S$ to separate the font style from the background in $x_0^n$ , yielding a segmented font style image $x_S = S(x_0^n)$ .", + "bbox": [ + 89, + 446, + 483, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By employing Dinov2 [26] as an image feature extractor $f$ , we obtain the features containing glyph and style information $x_{f} = f(x_{S})$ and the content features $\\mathcal{G}_f = f(\\mathcal{G}^n)$ . We then subtract the extracted feature vectors to obtain content-decoupled style information $\\mathcal{F}_{\\mathcal{S}} = x_{f} - \\mathcal{G}_{f}$ , which is injected into the cross-attention layers of the UNet using IP-Adapter [45]. By jointly injecting image style features and textual prompts as conditional information into the UNet attention block, we achieve interactive fusion through cross-attention mechanisms with the edit image. Furthermore, since image style injection only requires modifications to masked regions, we restrict the style's response regions to text editing areas by computing the intersection between IPA attention maps and the mask region $m$ .", + "bbox": [ + 89, + 568, + 482, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During text style injection, treating content bias as style feature resolves the coupling issue between glyph and typographic styles. This approach fully unleashes the text-image style transfer capabilities in two key aspects: (1) Enabling preservation of original typographic styles when editing full text segments without contextual references, through injecting source image characteristics into modified content. (2) Providing external style injection functionality that al", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "lows users to edit images using preferred textual exemplars through our style transfer framework.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3.2. In-Context Reference", + "text_level": 1, + "bbox": [ + 511, + 127, + 710, + 140 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our proposed Style-Adapter follows this approach to inject textual attributes from reference characters, achieving robust font style reconstruction. However, sometimes chromatic deviations persist due to the model's tendency to derive pixel color information from the contextual surroundings of the original image.", + "bbox": [ + 511, + 143, + 906, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The model's ability to extract color information from original image contexts led us to implement self-attention mechanisms [14, 46, 49] to manage detailed stylistic features. Based on this, TextMaster introduces an efficient approach to enhance the model's understanding of stylistic nuances in reference text. As illustrated in Figure 3, our framework processes the reference character $c_{s}$ using a segmentation model $S$ . The resized and positioned $c_{s}$ is placed at the bottom of the input image, creating modified context image $x_{c}$ , and latent features $z_{t}^{c}$ . This setup allows the model to learn spatial-semantic relationships between masked regions and the added contextual reference through self-attention.", + "bbox": [ + 511, + 237, + 908, + 431 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {\\text {s e l f}} = l _ {\\mathbf {Q}} \\left(\\varphi_ {\\text {s e l f}} \\left(z _ {t} ^ {c}\\right)\\right), \\mathbf {K} _ {\\text {s e l f}} = l _ {\\mathbf {K}} \\left(\\varphi_ {\\text {s e l f}} \\left(z _ {t} ^ {c}\\right)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 454, + 906, + 470 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} _ {\\text {s e l f}} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {\\text {s e l f}} \\mathbf {K} _ {\\text {s e l f}} ^ {\\top}}{\\sqrt {d _ {\\text {s e l f}}}}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 487, + 905, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, $l_{\\mathbf{Q}}$ and $l_{\\mathbf{K}}$ denotes the linear projection operation, $\\varphi_{\\mathrm{self}}$ represents layer normalization, and $\\mathbf{A}_{\\mathrm{self}}$ corresponds to the attention maps derived from self-attention computation. We term this approach In-Context Reference.", + "bbox": [ + 511, + 523, + 905, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The denoising loss is then formulated as follows:", + "bbox": [ + 532, + 584, + 857, + 598 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o i s e}} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, m ^ {\\prime}, t, e _ {t}\\right) \\right\\| _ {2} ^ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 604, + 903, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our overall loss consists of denoising loss $\\mathcal{L}_{\\mathrm{noise}}$ , perceptual loss $\\mathcal{L}_{\\mathrm{p}}$ and CIOU loss $\\mathcal{L}_{\\mathrm{iou}}$ , This can be expressed as:", + "bbox": [ + 511, + 628, + 905, + 661 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {n o i s e}} + \\beta \\mathcal {L} _ {\\mathrm {p}} + \\gamma \\mathcal {L} _ {\\mathrm {i o u}} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 667, + 903, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $\\beta$ and $\\gamma$ are coefficients that adjust the relative contributions of these loss terms.", + "bbox": [ + 511, + 689, + 905, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During training iterations, we observed that directly using the target reconstruction text as context reference led to an over-reliance on copying contextual glyphs to masked regions during inference, resulting in character deformation. To mitigate this, we randomly select a character as the contextual reference and completely erase it using the erasure model [36], enabling the model to learn reconstruction of remaining characters while decoupling the generated textual content from the referenced typographic style. This strategy enhances the model's capability for accurate reconstruction of color-detail stylistic features and ensures effective content-style disentanglement characteristics.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16117", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/428d9f31253c97456b495e8d66f2e6124cad92e0355bc344bb72d4e3411e3a1d.jpg", + "image_caption": [ + "Figure 6. Comparison of the visualized results between our method and anytext[39], anytext2[38], flux[1], and textctrl[47]." + ], + "image_footnote": [], + "bbox": [ + 137, + 90, + 859, + 369 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7163fc2a7415149ffcf61f2792e25af2da4e7279b3e440eab13fc7e19ce8ca80.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Evaluation BenchmarkModelChineseEnglish
Acc↑NED↑FID↓LPIPS↓Acc↑NED↑FID↓LPIPS↓
AnyText-BenchmarkSDXL----0.08570.146437.020.0557
FLUX----0.15940.318525.670.0471
TextDiffuser2----0.54750.795131.540.1457
AnyText-v1.10.53140.643754.810.11150.51670.587743.970.1303
AnyText20.61330.712030.560.05360.62350.665728.660.0662
TextCtrl----0.80960.82316.010.0234
TextMaster0.92570.948311.930.02470.93580.953410.350.0238
ICDAR13-BenchmarkSDXL----0.55790.102440.520.0735
FLUX----0.15940.137321.640.0581
TextDiffuser2----0.59210.542540.270.1537
AnyText-v1.1----0.55790.614234.560.1536
AnyText2----0.67650.683217.650.0754
TextCtrl----0.80550.811610.770.0278
TextMaster----0.91000.936610.820.0256
TextMaster-BenchmarkSDXL----0.12320.167545.340.0894
FLUX----0.17640.283727.320.0546
TextDiffuser2----0.34270.453719.640.0846
AnyText-v1.10.54670.768127.960.09340.45380.587718.980.1303
AnyText20.57840.643719.610.06380.54990.632714.370.0794
TextCtrl----0.82230.835410.410.0293
TextMaster0.91800.95259.860.02450.91800.93459.730.0237
", + "bbox": [ + 138, + 412, + 864, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Evaluation Results of random text editing on five benchmarks. Accuracy (ACC, NED) is higher the better $(\\uparrow)$ , while FID and LPIPS are lower the better $(\\downarrow)$ .", + "bbox": [ + 89, + 779, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 818, + 223, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 89, + 845, + 186, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our model was trained on the AnyWord-3M dataset containing Chinese/English texts, combining LAION [34]", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(p1/p2) and Wukong [10] (1-2/5) subsets. Given the limitations of existing open-source datasets in both image resolution and diversity, we publicly crawled 100,000 of artistic style posters from the web to enhance the model's style infusion capability during training. Specially, we", + "bbox": [ + 511, + 820, + 906, + 897 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16118", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bdbe5c00d47a426ae4a36bb9c30b149843e293fe93496337effbc6c769da555c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingAcc↑NED↑FID↓LPIPS↓
Base0.50980.587316.260.0454
+Token-separated0.68320.707314.580.0417
+Lp0.74510.763812.950.0369
+Liou0.78160.849513.020.0351
+in-context0.81620.874111.650.0338
+style adapter0.88600.902510.870.0286
", + "bbox": [ + 282, + 88, + 715, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Experimental results of different methods.", + "bbox": [ + 344, + 220, + 650, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "employed an OCR character detection model to generate single-character bounding boxes for precise CIOU loss calculation in character positioning and context reference in style learning. In addition, we used ICDAR13[16], Any-text benchmark[39], and the textmaster dataset we built for evaluation.", + "bbox": [ + 89, + 248, + 483, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Experimental Details", + "text_level": 1, + "bbox": [ + 89, + 345, + 290, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our model was initialized using pre-trained weights from the SDXL[27] model and pre-trained IPA weights from colors[37]. The training process was conducted in two phases. First step, we train five epochs on the AnyWord-3M dataset using 4 H20 GPUs. Set the learning rate to 1e-4, the weight of perception loss $\\beta$ to 0.1, the weight of iou loss $\\gamma$ to 0.05, and the noise-offset of 0.02. In second stage, we continue training Unet with in-context reference on style poster dataset, with learning rate of 1e-5. Then, we train style-adapter module with freezing Unet parameters first, subsequently jointly optimizing IPA-Unet attention interactions, and ultimately enabling full parameter fine-tuning. During training, the resolution is set to 1024.", + "bbox": [ + 89, + 368, + 483, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Comparison Results", + "text_level": 1, + "bbox": [ + 89, + 573, + 284, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.1. Quantitative and Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 594, + 390, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Following the methodology established in prior research [1, 27, 38, 39, 42, 47], we conducted a comprehensive evaluation of our model across three distinct datasets using a standardized approach. We divided the metrics into those that assess content accuracy, including OCR content accuracy and text edit distance, and for evaluating style retention, including FID and LPIPS. We ensured a fair comparison among all methods by evaluating accuracy metrics using random text, while assessing style similarity metrics with the original text. For t2i methods [1, 27, 38, 39], we supplemented them with image description prompts. As shown in Table 1, our method outperforms the current state-of-the-art methods across all metrics. We observe that TextCtrl tends to retain the original text during the editing process, which leads to reduced accuracy in content editing. However, this characteristic gives it an advantage in style similarity metrics when reconstructing from the original text.", + "bbox": [ + 89, + 613, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure 6, we have tested random editing and can see the great advantages of our method for editing", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "arbitrary random content. Also, a large number of qualitative results (see supplementary materials) demonstrate that our model significantly outperforms existing models in both Chinese and English editing effects, particularly in complex Chinese text editing. Moreover, our general editing capability not only maintains accurate content editing but also enhances style retention ability.", + "bbox": [ + 511, + 248, + 905, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 372, + 666, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 2, we conducted ablation evaluations on 500 carefully selected images that represent various aspects of text. The results show that incorporating glyph information significantly improves the model's performance across all metrics. The perceptual losses and token-separated text encoder contribute to a substantial increase in text generation accuracy and overall image quality. Furthermore, considering the attention IoU loss further enhances the model's ability to accurately locate text in images.", + "bbox": [ + 511, + 396, + 905, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Meanwhile, the style injection module not only significantly improves the metric of style consistency but also synergistically enhances the accuracy of text editing. In our experiments, we observe that style injection is essential for text generation: in the absence of style reference, the generated arbitrary-style text tends to produce less accurate glyphs, and the resulting text is more difficult to harmoniously blend with its surroundings.", + "bbox": [ + 511, + 535, + 905, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce TextMaster, a groundbreaking method that, for the first time, supports high-accuracy, consistent style text editing with rational layout capabilities. We ensured the generation of glyph structures through the design of glyph images, a token-separating encoder, and a perceptual loss method. We utilized attention IoU loss to promote the model's learning of adaptive layouts. Additionally, we introduce a style injection method that infuses the style of the modified font into the target font, decoupling semantic entanglement and preserving style information. Extensive qualitative and quantitative experimental results demonstrate that TextMaster outperforms the current state-of-the-art methods.", + "bbox": [ + 511, + 705, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16119", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] BlackForestLab. Flux.1. https://blackforestlabs.ai/announcing-black-forest-labs/,2024.3,7,8", + "[2] Haoxing Chen, Zhuoer Xu, Zhangxuan Gu, Jun Lan, Xing Zheng, Yaohui Li, Changhua Meng, Huijia Zhu, and Weiqiang Wang. Diffuse: Universal text editing diffusion model, 2023. 2", + "[3] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser-2: Unleashing the power of language models for text rendering. arXiv preprint arXiv:2311.16465, 2023. 3", + "[4] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. arXiv preprint, abs/2305.10855, 2023. 2, 3", + "[5] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794, 2021. 2", + "[6] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. ACL, 2022. 3", + "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first international conference on machine learning*, 2024. 3", + "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 1", + "[9] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1", + "[10] Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Minzhe Niu, Xiaodan Liang, Lewei Yao, Runhui Huang, Wei Zhang, Xin Jiang, Chunjing Xu, and Hang Xu. Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark, 2022. 7", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2", + "[12] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. CoRR, 2023. 1", + "[13] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint, abs/2302.09778, 2023. 1", + "[14] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers, 2024. 6" + ], + "bbox": [ + 91, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 2, 3, 6", + "[16] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazán Almazán, and Lluis Pere de las Heras. Icdar 2013 robust reading competition. In 2013 12th International Conference on Document Analysis and Recognition, pages 1484–1493, 2013. 8", + "[17] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3", + "[18] Chenxia Li, Weiwei Liu, Ruoyu Guo, Xiaoting Yin, Kaitao Jiang, Yongkun Du, Yuning Du, Lingfeng Zhu, Baohua Lai, Xiaoguang Hu, Dianhai Yu, and Yanjun Ma. Pp-ocrv3: More attempts for the improvement of ultra lightweightOCR system, 2022. 4", + "[19] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[20] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. arXiv preprint arXiv:2403.09622, 2024. 3", + "[21] Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. Glyph-byt5-v2: A strong aesthetic baseline for accurate multilingual visual text rendering. arXiv preprint arXiv:2406.10208, 2024. 3", + "[22] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently. arXiv preprint, abs/2303.17870, 2023. 2, 3", + "[23] Jian Ma, Yonglin Deng, Chen Chen, Haonan Lu, and Zhenyu Yang. Glyphdraw2: Automatic generation of complex glyph posters with diffusion models and large language models. CoRR, 2024. 3", + "[24] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint, abs/2302.08453, 2023. 1", + "[25] OpenAI. Dall-e3. https://openai.com/index/dall-e-3/, 2023.3", + "[26] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 6", + "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16120", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 8", + "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3", + "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1", + "[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint, abs/2208.12242, 2022. 1", + "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Raphael Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2", + "[33] Joshua Santoso, Christian Simon, and Williem. On manipulating scene text in the wild with diffusion models, 2023. 2", + "[34] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs, 2021. 7", + "[35] Wenda Shi, Yiren Song, Dengming Zhang, Jiaming Liu, and Xingxing Zou. Fonts: Text rendering with typography and style controls, 2025. 2", + "[36] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 6", + "[37] Kolars Team. Kolars: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 8", + "[38] Yuxiang Tuo, Yifeng Geng, and Liefeng Bo. Anytext2: Visual text generation and editing with customizable attributes. arXiv preprint arXiv:2411.15245, 2024. 3, 6, 7, 8", + "[39] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. *ICLR*, 2024. 2, 3, 7, 8", + "[40] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation, 2024. 5, 6", + "[41] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Yibin Wang, Weizhong Zhang, Honghui Xu, and Cheng Jin. Dreamtext: High fidelity scene text synthesis, 2025. 8", + "[43] Yukang Yang, Dongnan Gui, Yuhui Yuan, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. arXiv preprint, abs/2305.18259, 2023. 2, 3", + "[44] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint, 2023. 1, 6", + "[46] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and Anan Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model, 2024. 6", + "[47] Weichao Zeng, Yan Shu, Zhenhang Li, Dongbao Yang, and Yu Zhou. Textctrl: Diffusion-based scene text editing with prior guidance control, 2024. 2, 7, 8", + "[48] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint, abs/2302.05543, 2023. 1", + "[49] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation, 2024. 6" + ], + "bbox": [ + 514, + 92, + 903, + 458 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16121", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/948bb483-aee7-45ff-9c3e-421173343c17_model.json b/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/948bb483-aee7-45ff-9c3e-421173343c17_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0ef07d6f2f111ab941f8a94ab4c4fa89a430209d --- /dev/null +++ b/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/948bb483-aee7-45ff-9c3e-421173343c17_model.json @@ -0,0 +1,2079 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.131, + 0.806, + 0.177 + ], + "angle": 0, + "content": "TextMaster: A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.203, + 0.877, + 0.242 + ], + "angle": 0, + "content": "Zhenyu Yan\\(^{1*}\\), Jian Wang\\(^{1*‡}\\), Aoqiang Wang\\(^{*†}\\), Yuhan Li\\(^{2†}\\), Wenxiang Shang\\(^{1}\\), Zhu Hangcheng\\(^{1}\\) \n\\(^{1}\\)Taobao & Tmall Group of Alibaba \\(\\quad^{2}\\)Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.244, + 0.841, + 0.26 + ], + "angle": 0, + "content": "{yeeshai.yzy, alex.wj}@taobao.com, wangaq1999@gmail.com, melodious@sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.294, + 0.327, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.327, + 0.486, + 0.614 + ], + "angle": 0, + "content": "In image editing tasks, high-quality text editing capabilities can significantly reduce both human and material resource costs. Existing methods, however, face significant limitations in terms of stroke accuracy for complex text and controllability of generated text styles. To address these challenges, we propose TextMaster, a solution capable of accurately editing text across various scenarios and image regions, while ensuring proper layout and controllable text style. Our method enhances the accuracy and fidelity of text rendering by incorporating high-resolution standard glyph information and applying perceptual loss within the text editing region. Additionally, we leverage an attention mechanism to compute intermediate layer bounding box regression loss for each character, enabling the model to learn text layout across varying contexts. Furthermore, we propose a novel style injection technique that enables controllable style transfer for the injected text. Through comprehensive experiments, we demonstrate the state-of-the-art performance of our method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.66, + 0.222, + 0.675 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Recent advances in generative models have revolutionized image synthesis, with diffusion models surpassing Generative Adversarial Networks (GANs) in quality, diversity, and controllability [11, 27, 30]. While GAN-based models, such as the original Generative Adversarial Nets [9], have demonstrated impressive realism, they often suffer from training instability and mode collapse. The advent of diffusion models, such as Denoising Diffusion Probabilistic Models (DDPM) [11] and Latent Diffusion Models (LDM) [30], have led to superior image fidelity and text-image alignment, making them the dominant paradigm in" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.298, + 0.898, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.378, + 0.897, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.463, + 0.897, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.582, + 0.908, + 0.625 + ], + "angle": 0, + "content": "Figure 1. The image illustrates the diverse capabilities of our TextMaster, encompassing precise typesetting and layout, consistent style retention, and external style injection." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.656, + 0.722, + 0.671 + ], + "angle": 0, + "content": "text-to-image (T2I) generation." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.672, + 0.907, + 0.794 + ], + "angle": 0, + "content": "To enhance the controllability of T2I generation, researchers have explored methods to inject conditional signals. ControlNet [48] enables spatial constraints by conditioning diffusion models on structural priors, while T2I-Adapter [24] further enhances control by learning task-specific adapters. Similarly,Composer [13] introduces a framework for flexible, composable conditions, allowing intricate constraints in text-guided image synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Another critical direction involves capturing specific styles and characteristics from reference images. Techniques such as Textual Inversion [8] and DreamBooth [31] fine-tune models on personalized concepts, improving fidelity to target attributes. IP-Adapter [45] extends this by efficiently integrating image prompts with textual guidance, while Reference-Efficient [12] and InstantID [41] enable" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.863, + 0.366, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. \n†Work done during an internship at Alibaba Group. \n‡Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16112" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.174, + 0.093, + 0.825, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.447, + 0.907, + 0.476 + ], + "angle": 0, + "content": "Figure 2. The current approach exhibits several deficiencies, including a lack of alignment between the generated text and the intended target text, suboptimal layout organization, and inconsistent in style preservation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.483, + 0.548 + ], + "angle": 0, + "content": "identity preservation in personalized generation. Further advancements, such as PhotoMaker [19], refine identity embeddings to achieve photorealistic results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The main challenge of text editing lies in generating accurate textual content while minimizing any noticeable traces of editing. To Achieve this requires, the model needs to reconstruct multiple densely packed characters within localized regions, which places substantial demands on its fine-grained modeling abilities. Recent methods [22, 43] have explored glyph control for text editing. Yet, these approaches often suffer from reduced editing accuracy due to the lack of character-level textual guidance. On the other hand, effectively disentangling content from style presents another significant hurdle. Due to the highly structured nature of character glyphs, independently extracting style features from a given text image remains particularly difficult. Diffusion-based methods [2, 4, 39] attempt to restore text by leveraging the image context to preserve style. Since style information is primarily derived from unmasked regions, this may lead to inconsistencies in style representation during reconstruction. Other approaches [15, 35] seek to disentangle text style from glyph structure using manually constructed datasets, where different texts are rendered with various fonts to create style-content pairs for learning. Despite being designed to simulate real-world scenarios, such datasets still fall short in capturing the authenticity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.562 + ], + "angle": 0, + "content": "and diversity found in real text images. Furthermore, some methods opt to inject reference styles through text-style image inversion [33, 47], but this comes at the cost of significantly increased inference overhead." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.909, + 0.763 + ], + "angle": 0, + "content": "In response to these challenges, our method introduces further improvements in glyph control by combining character-level text embedding and IoU loss enabling precise regulation of both glyph shape and positioning. For style disentanglement, unlike previous approaches that rely on synthesized data, whose results are often affected by the glyphs in the reference image, we explicitly separate the glyph structure from the style appearance, which is achieved by integrating a style adapter with in-context learning. These innovations enable our model to reconstruct text across deiverse and fully authentic datasets, ultimately delivering robust style disentanglement and superior generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "By integrating these advanced techniques, the proposed TextMaster framework addresses the critical challenges in text editing, offering unparalleled performance in maintaining style consistency, content accuracy, proper layout, and high-quality visual output. These advancements position TextMaster as a state-of-the-art tool for complex text editing tasks in images [5, 11, 32]. Meanwhile, Our approach is not limited to some specific languages, it can be easily extended to other languages as well." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16113" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.233, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.123, + 0.482, + 0.228 + ], + "angle": 0, + "content": "Recent advancements in generative models, such as FLUX [1], DALL-E 3 [25], and SD3 [7], have significantly improved image synthesis quality, achieving high-resolution fidelity and enhanced semantic coherence. However, challenges remain in accurate text rendering within images. Specifically, ensuring precise glyph structure, spatial consistency, and stylistic control remains a complex task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.236, + 0.483, + 0.493 + ], + "angle": 0, + "content": "Accurate Glyph and Spatial Control. Several studies have addressed the challenges of generating visually coherent and spatially accurate text. GlyphDraw [22] proposes a word-level synthesis approach to ensure coherent Chinese character rendering. GlyphControl [43] introduces glyph-conditioned diffusion to enhance character structure control, while TextDiffuser [4] treats text rendering as a diffusion-based painting task. Recent advancements, such as AnyText [39], extend this to multilingual text generation. Further improvements, such as TextDiffuser-2 [3], leverage large language models for refined text placement. Additionally, models like Glyph-ByT5 [20] and its successor Glyph-ByT5-v2 [21] introduce customized text encoders for multilingual and aesthetic text rendering, improving both accuracy and flexibility. Moreover, GlyphDraw2 [23] combines diffusion models with large-scale language modeling to generate complex text-based posters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.651 + ], + "angle": 0, + "content": "Controllable Text Style and Appearance. Separating textual content from stylistic attributes is a longstanding challenge. DiffSTE [15] improves text editing by leveraging dual encoders trained on extensive font datasets, allowing font-specific style preservation during synthesis. However, inference still requires explicit font name input. AnyText2 [38] enhances controllability by embedding RGB-based color attributes, but complex styles such as gradients, borders, and shadows remain difficult to represent effectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.687, + 0.182, + 0.703 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Text editing tasks can be deconstructed into two aspects: glyphs and styles. As shown in the figure 3, to achieve precise editing control, we propose Glyph-Style Dual-Control. Glyph control involves maintaining structural integrity of characters through glyph constraints, and achieving cohesive text flow via adaptive typography layouts. To maintain a consistent style with the original text, we introduce style control, which includes adjustments for typeface and color, allowing for accurate referencing and injection of text styles. This dual mechanism ensures precise character rendering, harmonized layout composition, and versatile style referencing capabilities." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.644, + 0.108 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Our proposed method builds upon Stable Diffusion [29] Inpainting Models. Given an image \\( x_0 \\), a binary mask \\( m \\) specifying the editable region and a textual prompt \\( P \\) describing the target text to generate, the system first encodes both the original image \\( x_0 \\) and its masked image \\( x_0 \\odot (1 - m) \\) into latent space using a Variational Autoencoder (VAE) [17], where \\( \\odot \\) denotes element-wise multiplication. Then a conditional latent diffusion model performs sequential denoising with a UNet in the latent space. UNet takes as inputs the noise latent variable \\( \\hat{z}_t \\), downsampled mask \\( m' \\), masked image latent \\( z_{Im} \\), timestep \\( t \\), and embedding \\( e_t \\) of the textual prompt \\( P \\) from a pre-trained CLIP [28] model, and predicts the noise \\( \\epsilon_{\\theta}(\\hat{z}_t, z_{Im}, m', t, e_t) \\). The training of diffusion models resolves around optimizing the denoiser network \\( \\epsilon_{\\theta} \\) to minimize the noise prediction loss:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.349, + 0.907, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o i s e}} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\hat {z} _ {t}, z _ {I m}, m ^ {\\prime}, t, e _ {t}\\right) \\right\\| _ {2} ^ {2} \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.377, + 0.659, + 0.394 + ], + "angle": 0, + "content": "3.2. Glyph control" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.399, + 0.684, + 0.415 + ], + "angle": 0, + "content": "3.2.1. Glyph Constraint" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.906, + 0.493 + ], + "angle": 0, + "content": "The accurate rendering of glyphs is critical in text editing tasks. To ensure proper formation of character strokes, we have implemented multi-faceted constraints including reference standard character images, precise textual encoding representations, and perceptual loss mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Glyph Guidance. Previous work [22, 39, 44] has demonstrated the effective guidance provided by standard font reference images for text content generation. However, these methods design glyph images with per-char alignment positions to accommodate the generation of arbitrary shape masks. Unfortunately, the shape of these masks may unintentionally reveal textual details, such as character count and layout (see Fig.2 problem 1). Therefore, we calculate the outer rectangle of the edit area and inject glyph information by field alignment. The standard quadrilateral mask serves as a better foundation for glyph representation, as a well-coupled relationship between glyphs and layout fosters harmonious content generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.839 + ], + "angle": 0, + "content": "In detail, we first obtain the bounding box of the text area to be injected and generate the corresponding glyph in this area with standard font in appropriate size. Next, we convert the generated glyph image \\(\\mathcal{I}_g\\) into latent space variables \\(l_{g}\\) using the VAE [17] model. Finally, we concatenate these latent space variables with the latent representation of the image along the channel dimension to ensure that the generated image visually aligns with the original input image while clearly retaining and conveying the glyph information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Token-separated Text Encoder. The default text encoder of SDXL [27] is not sufficient for representing Chinese semantics. To address this limitations, we employ the ChatGLM [6] text encoder to support both Chinese and English" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16114" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.176, + 0.096, + 0.827, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.409, + 0.908, + 0.455 + ], + "angle": 0, + "content": "Figure 3. The framework of TextMaster consists of the following components: the glyph control module, comprising the glyph constraint, update text encoder, and perception module; the layout control is achieved via position-aware attention; and the style control module encompasses the style-adapter and in-context reference." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.483, + 0.538 + ], + "angle": 0, + "content": "characters. ChatGLM employs a dual-stream self-attention architecture: one stream processes positional information to capture text sequence and structure, while the other handles content information to understand semantic relationships." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.539, + 0.483, + 0.734 + ], + "angle": 0, + "content": "In text editing tasks, semantic relationships can adversely affect generation outcomes, as character embeddings may differ based on context. To reduce this interference, we adopt a specialized encoding method. First, we tokenize the prompt \\( P = \\{c_{1}, c_{2}, \\dots, c_{K}\\} \\) by treating each character \\( c_{k} \\in P \\) as an individual token, preventing cross-semantic influence. We then sort and concatenate these tokens according to their original positions, using zero padding to ensure consistent sequence lengths. Finally, we use a text encoder \\( \\mathcal{E}_{t} \\) to encode the characters token by token, incorporating position information \\( p_{k} \\), and generate the corresponding mask encoding \\( m_{k} \\) for each token \\( c_{k} \\). Each character is encoded as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.746, + 0.483, + 0.763 + ], + "angle": 0, + "content": "\\[\ne _ {k} = \\mathcal {E} _ {t} \\left(c _ {k} + \\operatorname {p o s} _ {k} + m _ {k}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.483, + 0.804 + ], + "angle": 0, + "content": "where \\(k\\in \\{1,2,\\dots ,K\\}\\) . We obtain the final text conditional features by concatenating the character encodings:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.818, + 0.483, + 0.836 + ], + "angle": 0, + "content": "\\[\ne _ {t} = \\left[ e _ {1}; e _ {2}; \\dots ; e _ {K} \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.9 + ], + "angle": 0, + "content": "This approach maintains uniform character embeddings across contexts while preserving positional information, ultimately improving the generation quality in text editing tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.905, + 0.568 + ], + "angle": 0, + "content": "Perception Module. We propose a method based on perceptual loss to maintain the consistency of glyph structures. By calculating the \\( L_{2} \\) loss of the Optical Character Recognition (OCR) features in the edited region between the reconstructed image \\( \\hat{x}_0 \\) and the original image \\( x_0 \\), we impose constraints on the generation of glyph structures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.789 + ], + "angle": 0, + "content": "Specifically, by denoising the noise prediction \\(\\epsilon_{t}\\) from the denoising network's output, we can derive an estimate \\(\\hat{z}_0\\) from the noise vector \\(z_{t}\\) as follows: \\(\\hat{z}_{0} = \\frac{1}{\\sqrt{\\bar{\\alpha}_{t}}} z_{t} + \\frac{\\sqrt{1 - \\bar{\\alpha}_{t}}}{\\sqrt{\\bar{\\alpha}_{t}}}\\epsilon_{t}\\), where \\(\\epsilon_{t} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), and \\(\\alpha_{t}\\) denotes the variance scheduling of the Gaussian noise. Furthermore, we utilize a VAE decoder to obtain an approximate reconstruction estimate \\(\\hat{x}_0\\). Based on the mask \\(m\\) of the edited region, we crop the original image \\(x_0\\) and the reconstructed image \\(\\hat{x}_0\\), resulting in the corresponding images of the edited area \\(x_0^m\\) and \\(\\hat{x}_0^m\\), respectively. We employ the PPOCR [18] model as the glyph encoder \\(\\phi (\\cdot)\\) and extract features \\(f^{m} = \\phi (x_{0}^{m})\\) and \\(\\hat{f}^{m} = \\phi (\\hat{x}_{0}^{m})\\) from the fully connected layer to compute the L2 loss of the glyph features between the original and predicted images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.79, + 0.909, + 0.866 + ], + "angle": 0, + "content": "Moreover, we found it necessary to ensure a natural transition between the text and the surrounding background. Therefore, we further introduce pixel-based Mean Squared Error (MSE) loss to ensure the accuracy of the text generation area. The resulting perceptual loss is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.878, + 0.908, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\left\\| f ^ {m} - \\hat {f} ^ {m} \\left\\| _ {2} ^ {2} + \\left\\| x _ {0} ^ {m} - \\hat {x} _ {0} ^ {m} \\right\\| _ {2}. \\right. \\right. \\qquad (4)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16115" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "Perceptual loss injects a stronger supervisory signal by constraining the reconstruction of the editing region, ensuring the correct presentation of the text." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.158, + 0.462, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.378, + 0.483, + 0.407 + ], + "angle": 0, + "content": "Figure 4. The attention response map processing flow: Atten represent the top chosen attention response maps for each token." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.431, + 0.345, + 0.447 + ], + "angle": 0, + "content": "3.2.2. Adaptive Typography Layout" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.45, + 0.483, + 0.706 + ], + "angle": 0, + "content": "Position-aware Attention. We expect the model to exhibit accurate positional responses for each character within the masked regions, thereby generating text sequences with proper typographic arrangement. As revealed in prior work [40], different cross-attention blocks demonstrate distinct attention orientations, with certain layers exhibiting stronger responses to layout patterns. As illustrated in Fig. 4, our analysis of text generation tasks shows that specific attention layers demonstrate pronounced perception of character positioning. To optimize layout learning, we calculate the IoU overlap ratio between binarized attention maps and ground-truth character bounding boxes, subsequently selecting the top three layers exhibiting the strongest responses for layout adaptation. This design facilitates improved intra-paragraph layout learning, as illustrated in Fig. 2 problem 2. Specifically, we perform the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.483, + 0.737 + ], + "angle": 0, + "content": "Compute attention scores \\(\\mathbf{A}\\) using the query matrix \\(\\mathbf{Q}\\) and key matrix \\(\\mathbf{K}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.747, + 0.483, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} = \\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {d _ {k}}}, \\quad \\mathbf {W} = \\operatorname {s o f t m a x} (\\mathbf {A}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.485, + 0.824 + ], + "angle": 0, + "content": "Thresholding: Generate a binary image \\(\\mathbf{T}\\) by applying a threshold \\(T = 125\\) to the averaged attention map:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.835, + 0.483, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} (i, j) = \\left\\{ \\begin{array}{l l} 2 5 5 & \\text {i f} \\bar {\\mathbf {A}} (i, j) > T \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Gaussian Blurring: Apply Gaussian blurring to the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.879, + 0.108 + ], + "angle": 0, + "content": "thresholded image \\(\\mathbf{T}\\) using a Gaussian kernel \\(\\mathbf{W}(k,l)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.118, + 0.907, + 0.161 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} (i, j) = \\sum_ {k = - K} ^ {K} \\sum_ {l = - L} ^ {L} \\mathbf {W} (k, l) \\cdot \\mathbf {T} (i + k, j + l). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.17, + 0.907, + 0.217 + ], + "angle": 0, + "content": "Complete Intersection over Union (CIOU) Loss: Calculate the CIOU loss between the derived bounding box \\(\\mathbf{B}_d\\) and the ground truth bounding box \\(\\mathbf{B}_g\\):" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.227, + 0.907, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {i o u}} = 1 - \\left(\\operatorname {I o U} - \\frac {\\rho^ {2} \\left(\\mathbf {B} _ {d} , \\mathbf {B} _ {g}\\right)}{c ^ {2}} - \\alpha v\\right). \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.27, + 0.567, + 0.283 + ], + "angle": 0, + "content": "Where:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.285, + 0.903, + 0.3 + ], + "angle": 0, + "content": "- IoU: Intersection over Union of the two bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.3, + 0.872, + 0.315 + ], + "angle": 0, + "content": "- \\(\\rho(\\mathbf{B}_d, \\mathbf{B}_g)\\): Euclidean distance between the centers." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.315, + 0.904, + 0.345 + ], + "angle": 0, + "content": "- \\(c\\): Diagonal length of the smallest enclosing box covering both bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.346, + 0.688, + 0.361 + ], + "angle": 0, + "content": "- \\(\\alpha\\): Trade-off parameter." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.361, + 0.786, + 0.375 + ], + "angle": 0, + "content": "- \\( v \\) : Measure of aspect ratio consistency." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.285, + 0.904, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Adaptive Crop and Mask Strategy. Our experiments revealed that VAE-encoded latent representations often degrade glyph structures in extremely small text. To improve multi-scale generalization, we developed an Adaptive Crop strategy that dynamically adjusts text regions. In our experimental setup, we configured the threshold value at 9 characters. For text segments shorter than this threshold, we performed center-aligned cropping on the images and scaled them to the proportional size corresponding to a 9-character width. When the text length exceeded the threshold, we expanded the cropped region at a 1/10 scaling ratio. As demonstrated in Fig. 5, this threshold-triggered spatial adaptation significantly improves background restoration and visual coherence." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.604, + 0.645, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.695, + 0.625, + 0.706 + ], + "angle": 0, + "content": "Prompt:\"健身房\"(Gym)" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.604, + 0.772, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.604, + 0.885, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.762 + ], + "angle": 0, + "content": "Figure 5. The text modification area depicted in the image is merely 30 pixels in height; TextMaster is still capable of executing high-quality edits within this constrained space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Additionally, to address uncontrollable text generation caused by mask size leakage (Fig. 2 problem 1) and mitigate layout issues from over-reliance on precise bounding dimensions, we propose an adaptive mask augmentation technique. By randomly expanding masked regions around text, this approach enhances robustness to bounding variations while improving the model's capability to generate visually coherent text layouts." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16116" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.24, + 0.108 + ], + "angle": 0, + "content": "3.3. Style Injection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.31 + ], + "angle": 0, + "content": "Current methods prioritize text accuracy and layout over style consistency. Preserving original text styles enhances visual harmony with image contexts, while integrating external references broadens editing versatility. Although recent works [15, 38] employ style control via datasets artificially constructed from font libraries (conditioned on basic attributes like color and typeface), such synthetic datasets often mismatch real-world distributions and provide coarse-grained control, failing to reproduce complex styles like gradient colors, outlined text, or 3D effects.(Fig. 2 problem 3) To overcome these limitations, we propose a Style-Adapter module with In-Context Learning for precise style manipulation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.321, + 0.236, + 0.336 + ], + "angle": 0, + "content": "3.3.1. Style-Adapter" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.446 + ], + "angle": 0, + "content": "In style transfer tasks, decoupling style from content is a critical challenge. Text inherently carries strong glyph information, and the segmentation boundary information carried by the style often disrupts the generation results. InstanceStyle [40] decouples image objects and their semantic content. Inspired by this, we aim to force the model to decouple font style from glyph content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.568 + ], + "angle": 0, + "content": "First, to enable the model to extract more general style information rather than strictly aligning with the segmentation boundaries of the generated text, we randomly select \\( n, n \\in \\{1, 2, \\dots, K\\} \\) words from the edit region of the original image \\( x_0 \\) and the glyph \\( \\mathcal{G} \\), resulting in cropped images \\( x_0^n \\) and \\( \\mathcal{G}^n \\). We further apply a text segmentation model \\( S \\) to separate the font style from the background in \\( x_0^n \\), yielding a segmented font style image \\( x_S = S(x_0^n) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.78 + ], + "angle": 0, + "content": "By employing Dinov2 [26] as an image feature extractor \\( f \\), we obtain the features containing glyph and style information \\( x_{f} = f(x_{S}) \\) and the content features \\( \\mathcal{G}_f = f(\\mathcal{G}^n) \\). We then subtract the extracted feature vectors to obtain content-decoupled style information \\( \\mathcal{F}_{\\mathcal{S}} = x_{f} - \\mathcal{G}_{f} \\), which is injected into the cross-attention layers of the UNet using IP-Adapter [45]. By jointly injecting image style features and textual prompts as conditional information into the UNet attention block, we achieve interactive fusion through cross-attention mechanisms with the edit image. Furthermore, since image style injection only requires modifications to masked regions, we restrict the style's response regions to text editing areas by computing the intersection between IPA attention maps and the mask region \\( m \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.902 + ], + "angle": 0, + "content": "During text style injection, treating content bias as style feature resolves the coupling issue between glyph and typographic styles. This approach fully unleashes the text-image style transfer capabilities in two key aspects: (1) Enabling preservation of original typographic styles when editing full text segments without contextual references, through injecting source image characteristics into modified content. (2) Providing external style injection functionality that al" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "lows users to edit images using preferred textual exemplars through our style transfer framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.128, + 0.712, + 0.141 + ], + "angle": 0, + "content": "3.3.2. In-Context Reference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.145, + 0.907, + 0.236 + ], + "angle": 0, + "content": "Our proposed Style-Adapter follows this approach to inject textual attributes from reference characters, achieving robust font style reconstruction. However, sometimes chromatic deviations persist due to the model's tendency to derive pixel color information from the contextual surroundings of the original image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.909, + 0.432 + ], + "angle": 0, + "content": "The model's ability to extract color information from original image contexts led us to implement self-attention mechanisms [14, 46, 49] to manage detailed stylistic features. Based on this, TextMaster introduces an efficient approach to enhance the model's understanding of stylistic nuances in reference text. As illustrated in Figure 3, our framework processes the reference character \\( c_{s} \\) using a segmentation model \\( S \\). The resized and positioned \\( c_{s} \\) is placed at the bottom of the input image, creating modified context image \\( x_{c} \\), and latent features \\( z_{t}^{c} \\). This setup allows the model to learn spatial-semantic relationships between masked regions and the added contextual reference through self-attention." + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.455, + 0.907, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {\\text {s e l f}} = l _ {\\mathbf {Q}} \\left(\\varphi_ {\\text {s e l f}} \\left(z _ {t} ^ {c}\\right)\\right), \\mathbf {K} _ {\\text {s e l f}} = l _ {\\mathbf {K}} \\left(\\varphi_ {\\text {s e l f}} \\left(z _ {t} ^ {c}\\right)\\right). \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.488, + 0.906, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} _ {\\text {s e l f}} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {\\text {s e l f}} \\mathbf {K} _ {\\text {s e l f}} ^ {\\top}}{\\sqrt {d _ {\\text {s e l f}}}}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.906, + 0.584 + ], + "angle": 0, + "content": "Here, \\(l_{\\mathbf{Q}}\\) and \\(l_{\\mathbf{K}}\\) denotes the linear projection operation, \\(\\varphi_{\\mathrm{self}}\\) represents layer normalization, and \\(\\mathbf{A}_{\\mathrm{self}}\\) corresponds to the attention maps derived from self-attention computation. We term this approach In-Context Reference." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.585, + 0.859, + 0.599 + ], + "angle": 0, + "content": "The denoising loss is then formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.605, + 0.905, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o i s e}} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, m ^ {\\prime}, t, e _ {t}\\right) \\right\\| _ {2} ^ {2} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.63, + 0.906, + 0.662 + ], + "angle": 0, + "content": "Our overall loss consists of denoising loss \\(\\mathcal{L}_{\\mathrm{noise}}\\), perceptual loss \\(\\mathcal{L}_{\\mathrm{p}}\\) and CIOU loss \\(\\mathcal{L}_{\\mathrm{iou}}\\), This can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.668, + 0.905, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {n o i s e}} + \\beta \\mathcal {L} _ {\\mathrm {p}} + \\gamma \\mathcal {L} _ {\\mathrm {i o u}} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.906, + 0.718 + ], + "angle": 0, + "content": "Where \\(\\beta\\) and \\(\\gamma\\) are coefficients that adjust the relative contributions of these loss terms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "During training iterations, we observed that directly using the target reconstruction text as context reference led to an over-reliance on copying contextual glyphs to masked regions during inference, resulting in character deformation. To mitigate this, we randomly select a character as the contextual reference and completely erase it using the erasure model [36], enabling the model to learn reconstruction of remaining characters while decoupling the generated textual content from the referenced typographic style. This strategy enhances the model's capability for accurate reconstruction of color-detail stylistic features and ensures effective content-style disentanglement characteristics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16117" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.138, + 0.091, + 0.861, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.386, + 0.861, + 0.401 + ], + "angle": 0, + "content": "Figure 6. Comparison of the visualized results between our method and anytext[39], anytext2[38], flux[1], and textctrl[47]." + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.414, + 0.866, + 0.77 + ], + "angle": 0, + "content": "
Evaluation BenchmarkModelChineseEnglish
Acc↑NED↑FID↓LPIPS↓Acc↑NED↑FID↓LPIPS↓
AnyText-BenchmarkSDXL----0.08570.146437.020.0557
FLUX----0.15940.318525.670.0471
TextDiffuser2----0.54750.795131.540.1457
AnyText-v1.10.53140.643754.810.11150.51670.587743.970.1303
AnyText20.61330.712030.560.05360.62350.665728.660.0662
TextCtrl----0.80960.82316.010.0234
TextMaster0.92570.948311.930.02470.93580.953410.350.0238
ICDAR13-BenchmarkSDXL----0.55790.102440.520.0735
FLUX----0.15940.137321.640.0581
TextDiffuser2----0.59210.542540.270.1537
AnyText-v1.1----0.55790.614234.560.1536
AnyText2----0.67650.683217.650.0754
TextCtrl----0.80550.811610.770.0278
TextMaster----0.91000.936610.820.0256
TextMaster-BenchmarkSDXL----0.12320.167545.340.0894
FLUX----0.17640.283727.320.0546
TextDiffuser2----0.34270.453719.640.0846
AnyText-v1.10.54670.768127.960.09340.45380.587718.980.1303
AnyText20.57840.643719.610.06380.54990.632714.370.0794
TextCtrl----0.82230.835410.410.0293
TextMaster0.91800.95259.860.02450.91800.93459.730.0237
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.78, + 0.908, + 0.809 + ], + "angle": 0, + "content": "Table 1. Evaluation Results of random text editing on five benchmarks. Accuracy (ACC, NED) is higher the better \\((\\uparrow)\\), while FID and LPIPS are lower the better \\((\\downarrow)\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.224, + 0.837 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.846, + 0.187, + 0.861 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Our model was trained on the AnyWord-3M dataset containing Chinese/English texts, combining LAION [34]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.821, + 0.907, + 0.898 + ], + "angle": 0, + "content": "(p1/p2) and Wukong [10] (1-2/5) subsets. Given the limitations of existing open-source datasets in both image resolution and diversity, we publicly crawled 100,000 of artistic style posters from the web to enhance the model's style infusion capability during training. Specially, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16118" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.283, + 0.089, + 0.716, + 0.212 + ], + "angle": 0, + "content": "
SettingAcc↑NED↑FID↓LPIPS↓
Base0.50980.587316.260.0454
+Token-separated0.68320.707314.580.0417
+Lp0.74510.763812.950.0369
+Liou0.78160.849513.020.0351
+in-context0.81620.874111.650.0338
+style adapter0.88600.902510.870.0286
" + }, + { + "type": "table_caption", + "bbox": [ + 0.346, + 0.222, + 0.651, + 0.236 + ], + "angle": 0, + "content": "Table 2. Experimental results of different methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.249, + 0.484, + 0.339 + ], + "angle": 0, + "content": "employed an OCR character detection model to generate single-character bounding boxes for precise CIOU loss calculation in character positioning and context reference in style learning. In addition, we used ICDAR13[16], Any-text benchmark[39], and the textmaster dataset we built for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.347, + 0.292, + 0.364 + ], + "angle": 0, + "content": "4.2. Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.484, + 0.566 + ], + "angle": 0, + "content": "Our model was initialized using pre-trained weights from the SDXL[27] model and pre-trained IPA weights from colors[37]. The training process was conducted in two phases. First step, we train five epochs on the AnyWord-3M dataset using 4 H20 GPUs. Set the learning rate to 1e-4, the weight of perception loss \\(\\beta\\) to 0.1, the weight of iou loss \\(\\gamma\\) to 0.05, and the noise-offset of 0.02. In second stage, we continue training Unet with in-context reference on style poster dataset, with learning rate of 1e-5. Then, we train style-adapter module with freezing Unet parameters first, subsequently jointly optimizing IPA-Unet attention interactions, and ultimately enabling full parameter fine-tuning. During training, the resolution is set to 1024." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.285, + 0.589 + ], + "angle": 0, + "content": "4.3. Comparison Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.595, + 0.392, + 0.611 + ], + "angle": 0, + "content": "4.3.1. Quantitative and Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Following the methodology established in prior research [1, 27, 38, 39, 42, 47], we conducted a comprehensive evaluation of our model across three distinct datasets using a standardized approach. We divided the metrics into those that assess content accuracy, including OCR content accuracy and text edit distance, and for evaluating style retention, including FID and LPIPS. We ensured a fair comparison among all methods by evaluating accuracy metrics using random text, while assessing style similarity metrics with the original text. For t2i methods [1, 27, 38, 39], we supplemented them with image description prompts. As shown in Table 1, our method outperforms the current state-of-the-art methods across all metrics. We observe that TextCtrl tends to retain the original text during the editing process, which leads to reduced accuracy in content editing. However, this characteristic gives it an advantage in style similarity metrics when reconstructing from the original text." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "As shown in Figure 6, we have tested random editing and can see the great advantages of our method for editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.249, + 0.906, + 0.355 + ], + "angle": 0, + "content": "arbitrary random content. Also, a large number of qualitative results (see supplementary materials) demonstrate that our model significantly outperforms existing models in both Chinese and English editing effects, particularly in complex Chinese text editing. Moreover, our general editing capability not only maintains accurate content editing but also enhances style retention ability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.373, + 0.667, + 0.389 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.906, + 0.534 + ], + "angle": 0, + "content": "As shown in Table 2, we conducted ablation evaluations on 500 carefully selected images that represent various aspects of text. The results show that incorporating glyph information significantly improves the model's performance across all metrics. The perceptual losses and token-separated text encoder contribute to a substantial increase in text generation accuracy and overall image quality. Furthermore, considering the attention IoU loss further enhances the model's ability to accurately locate text in images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.906, + 0.657 + ], + "angle": 0, + "content": "Meanwhile, the style injection module not only significantly improves the metric of style consistency but also synergistically enhances the accuracy of text editing. In our experiments, we observe that style injection is essential for text generation: in the absence of style reference, the generated arbitrary-style text tends to produce less accurate glyphs, and the resulting text is more difficult to harmoniously blend with its surroundings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.906, + 0.9 + ], + "angle": 0, + "content": "In this paper, we introduce TextMaster, a groundbreaking method that, for the first time, supports high-accuracy, consistent style text editing with rational layout capabilities. We ensured the generation of glyph structures through the design of glyph images, a token-separating encoder, and a perceptual loss method. We utilized attention IoU loss to promote the model's learning of adaptive layouts. Additionally, we introduce a style injection method that infuses the style of the modified font into the target font, decoupling semantic entanglement and preserving style information. Extensive qualitative and quantitative experimental results demonstrate that TextMaster outperforms the current state-of-the-art methods." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16119" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] BlackForestLab. Flux.1. https://blackforestlabs.ai/announcing-black-forest-labs/,2024.3,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Haoxing Chen, Zhuoer Xu, Zhangxuan Gu, Jun Lan, Xing Zheng, Yaohui Li, Changhua Meng, Huijia Zhu, and Weiqiang Wang. Diffuse: Universal text editing diffusion model, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.216, + 0.484, + 0.271 + ], + "angle": 0, + "content": "[3] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser-2: Unleashing the power of language models for text rendering. arXiv preprint arXiv:2311.16465, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.273, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[4] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. arXiv preprint, abs/2305.10855, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.317, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[5] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.36, + 0.484, + 0.414 + ], + "angle": 0, + "content": "[6] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. ACL, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.417, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first international conference on machine learning*, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.502, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.56, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[9] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.616, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[10] Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Minzhe Niu, Xiaodan Liang, Lewei Yao, Runhui Huang, Wei Zhang, Xin Jiang, Chunjing Xu, and Hang Xu. Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.484, + 0.785 + ], + "angle": 0, + "content": "[12] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. CoRR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[13] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint, abs/2302.09778, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[14] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[16] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazán Almazán, and Lluis Pere de las Heras. Icdar 2013 robust reading competition. In 2013 12th International Conference on Document Analysis and Recognition, pages 1484–1493, 2013. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[17] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[18] Chenxia Li, Weiwei Liu, Ruoyu Guo, Xiaoting Yin, Kaitao Jiang, Yongkun Du, Yuning Du, Lingfeng Zhu, Baohua Lai, Xiaoguang Hu, Dianhai Yu, and Yanjun Ma. Pp-ocrv3: More attempts for the improvement of ultra lightweightOCR system, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[19] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[20] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. arXiv preprint arXiv:2403.09622, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[21] Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. Glyph-byt5-v2: A strong aesthetic baseline for accurate multilingual visual text rendering. arXiv preprint arXiv:2406.10208, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[22] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently. arXiv preprint, abs/2303.17870, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[23] Jian Ma, Yonglin Deng, Chen Chen, Haonan Lu, and Zhenyu Yang. Glyphdraw2: Automatic generation of complex glyph posters with diffusion models and large language models. CoRR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[24] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint, abs/2302.08453, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[25] OpenAI. Dall-e3. https://openai.com/index/dall-e-3/, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[26] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16120" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.119 + ], + "angle": 0, + "content": "els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint, abs/2208.12242, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Raphael Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.48 + ], + "angle": 0, + "content": "[33] Joshua Santoso, Christian Simon, and Williem. On manipulating scene text in the wild with diffusion models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[34] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[35] Wenda Shi, Yiren Song, Dengming Zhang, Jiaming Liu, and Xingxing Zou. Fonts: Text rendering with typography and style controls, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[36] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[37] Kolars Team. Kolars: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[38] Yuxiang Tuo, Yifeng Geng, and Liefeng Bo. Anytext2: Visual text generation and editing with customizable attributes. arXiv preprint arXiv:2411.15245, 2024. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[39] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. *ICLR*, 2024. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[40] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[41] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[42] Yibin Wang, Weizhong Zhang, Honghui Xu, and Cheng Jin. Dreamtext: High fidelity scene text synthesis, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[43] Yukang Yang, Dongnan Gui, Yuhui Yuan, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. arXiv preprint, abs/2305.18259, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[44] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[46] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and Anan Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[47] Weichao Zeng, Yan Shu, Zhenhang Li, Dongbao Yang, and Yu Zhou. Textctrl: Diffusion-based scene text editing with prior guidance control, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[48] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint, abs/2302.05543, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[49] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.093, + 0.905, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16121" + } + ] +] \ No newline at end of file diff --git a/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/948bb483-aee7-45ff-9c3e-421173343c17_origin.pdf b/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/948bb483-aee7-45ff-9c3e-421173343c17_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..374f51a36ea41f1663e4d83aec604398c5c93de9 --- /dev/null +++ b/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/948bb483-aee7-45ff-9c3e-421173343c17_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a28baddb088e6f99fa264efe6752c556fc48bd10189f9d769c5bacf07ace41db +size 13001792 diff --git a/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/full.md b/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b47d3a2a19a24fb57a4c6adbb6f4f40f5e1353b3 --- /dev/null +++ b/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/full.md @@ -0,0 +1,304 @@ +# TextMaster: A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control + +Zhenyu Yan $^{1*}$ , Jian Wang $^{1*‡}$ , Aoqiang Wang $^{*†}$ , Yuhan Li $^{2†}$ , Wenxiang Shang $^{1}$ , Zhu Hangcheng $^{1}$ $^{1}$ Taobao & Tmall Group of Alibaba $\quad^{2}$ Shanghai Jiao Tong University + +{yeeshai.yzy, alex.wj}@taobao.com, wangaq1999@gmail.com, melodious@sjtu.edu.cn + +# Abstract + +In image editing tasks, high-quality text editing capabilities can significantly reduce both human and material resource costs. Existing methods, however, face significant limitations in terms of stroke accuracy for complex text and controllability of generated text styles. To address these challenges, we propose TextMaster, a solution capable of accurately editing text across various scenarios and image regions, while ensuring proper layout and controllable text style. Our method enhances the accuracy and fidelity of text rendering by incorporating high-resolution standard glyph information and applying perceptual loss within the text editing region. Additionally, we leverage an attention mechanism to compute intermediate layer bounding box regression loss for each character, enabling the model to learn text layout across varying contexts. Furthermore, we propose a novel style injection technique that enables controllable style transfer for the injected text. Through comprehensive experiments, we demonstrate the state-of-the-art performance of our method. + +# 1. Introduction + +Recent advances in generative models have revolutionized image synthesis, with diffusion models surpassing Generative Adversarial Networks (GANs) in quality, diversity, and controllability [11, 27, 30]. While GAN-based models, such as the original Generative Adversarial Nets [9], have demonstrated impressive realism, they often suffer from training instability and mode collapse. The advent of diffusion models, such as Denoising Diffusion Probabilistic Models (DDPM) [11] and Latent Diffusion Models (LDM) [30], have led to superior image fidelity and text-image alignment, making them the dominant paradigm in + +![](images/f521f04b12ced89c7d4c20bdaaa0ab70c45795465961721afc5126725215c03d.jpg) + +![](images/b1a40186cd70b72b90d97f626b0fe13d1305df208a6912142f1d40a424713c8f.jpg) + +![](images/e6a607f42e273fcb7aee7c234db8344fc2a5e345eb55bbc0bd435f6932551bcf.jpg) +Figure 1. The image illustrates the diverse capabilities of our TextMaster, encompassing precise typesetting and layout, consistent style retention, and external style injection. + +text-to-image (T2I) generation. + +To enhance the controllability of T2I generation, researchers have explored methods to inject conditional signals. ControlNet [48] enables spatial constraints by conditioning diffusion models on structural priors, while T2I-Adapter [24] further enhances control by learning task-specific adapters. Similarly,Composer [13] introduces a framework for flexible, composable conditions, allowing intricate constraints in text-guided image synthesis. + +Another critical direction involves capturing specific styles and characteristics from reference images. Techniques such as Textual Inversion [8] and DreamBooth [31] fine-tune models on personalized concepts, improving fidelity to target attributes. IP-Adapter [45] extends this by efficiently integrating image prompts with textual guidance, while Reference-Efficient [12] and InstantID [41] enable + +![](images/c0ac696c6ddcba5add81456471b5bdb237679799fd5bc28ccae4aee133215bc1.jpg) +Figure 2. The current approach exhibits several deficiencies, including a lack of alignment between the generated text and the intended target text, suboptimal layout organization, and inconsistent in style preservation. + +identity preservation in personalized generation. Further advancements, such as PhotoMaker [19], refine identity embeddings to achieve photorealistic results. + +The main challenge of text editing lies in generating accurate textual content while minimizing any noticeable traces of editing. To Achieve this requires, the model needs to reconstruct multiple densely packed characters within localized regions, which places substantial demands on its fine-grained modeling abilities. Recent methods [22, 43] have explored glyph control for text editing. Yet, these approaches often suffer from reduced editing accuracy due to the lack of character-level textual guidance. On the other hand, effectively disentangling content from style presents another significant hurdle. Due to the highly structured nature of character glyphs, independently extracting style features from a given text image remains particularly difficult. Diffusion-based methods [2, 4, 39] attempt to restore text by leveraging the image context to preserve style. Since style information is primarily derived from unmasked regions, this may lead to inconsistencies in style representation during reconstruction. Other approaches [15, 35] seek to disentangle text style from glyph structure using manually constructed datasets, where different texts are rendered with various fonts to create style-content pairs for learning. Despite being designed to simulate real-world scenarios, such datasets still fall short in capturing the authenticity + +and diversity found in real text images. Furthermore, some methods opt to inject reference styles through text-style image inversion [33, 47], but this comes at the cost of significantly increased inference overhead. + +In response to these challenges, our method introduces further improvements in glyph control by combining character-level text embedding and IoU loss enabling precise regulation of both glyph shape and positioning. For style disentanglement, unlike previous approaches that rely on synthesized data, whose results are often affected by the glyphs in the reference image, we explicitly separate the glyph structure from the style appearance, which is achieved by integrating a style adapter with in-context learning. These innovations enable our model to reconstruct text across deiverse and fully authentic datasets, ultimately delivering robust style disentanglement and superior generalization performance. + +By integrating these advanced techniques, the proposed TextMaster framework addresses the critical challenges in text editing, offering unparalleled performance in maintaining style consistency, content accuracy, proper layout, and high-quality visual output. These advancements position TextMaster as a state-of-the-art tool for complex text editing tasks in images [5, 11, 32]. Meanwhile, Our approach is not limited to some specific languages, it can be easily extended to other languages as well. + +# 2. Related Work + +Recent advancements in generative models, such as FLUX [1], DALL-E 3 [25], and SD3 [7], have significantly improved image synthesis quality, achieving high-resolution fidelity and enhanced semantic coherence. However, challenges remain in accurate text rendering within images. Specifically, ensuring precise glyph structure, spatial consistency, and stylistic control remains a complex task. + +Accurate Glyph and Spatial Control. Several studies have addressed the challenges of generating visually coherent and spatially accurate text. GlyphDraw [22] proposes a word-level synthesis approach to ensure coherent Chinese character rendering. GlyphControl [43] introduces glyph-conditioned diffusion to enhance character structure control, while TextDiffuser [4] treats text rendering as a diffusion-based painting task. Recent advancements, such as AnyText [39], extend this to multilingual text generation. Further improvements, such as TextDiffuser-2 [3], leverage large language models for refined text placement. Additionally, models like Glyph-ByT5 [20] and its successor Glyph-ByT5-v2 [21] introduce customized text encoders for multilingual and aesthetic text rendering, improving both accuracy and flexibility. Moreover, GlyphDraw2 [23] combines diffusion models with large-scale language modeling to generate complex text-based posters. + +Controllable Text Style and Appearance. Separating textual content from stylistic attributes is a longstanding challenge. DiffSTE [15] improves text editing by leveraging dual encoders trained on extensive font datasets, allowing font-specific style preservation during synthesis. However, inference still requires explicit font name input. AnyText2 [38] enhances controllability by embedding RGB-based color attributes, but complex styles such as gradients, borders, and shadows remain difficult to represent effectively. + +# 3. Method + +Text editing tasks can be deconstructed into two aspects: glyphs and styles. As shown in the figure 3, to achieve precise editing control, we propose Glyph-Style Dual-Control. Glyph control involves maintaining structural integrity of characters through glyph constraints, and achieving cohesive text flow via adaptive typography layouts. To maintain a consistent style with the original text, we introduce style control, which includes adjustments for typeface and color, allowing for accurate referencing and injection of text styles. This dual mechanism ensures precise character rendering, harmonized layout composition, and versatile style referencing capabilities. + +# 3.1. Preliminary + +Our proposed method builds upon Stable Diffusion [29] Inpainting Models. Given an image $x_0$ , a binary mask $m$ specifying the editable region and a textual prompt $P$ describing the target text to generate, the system first encodes both the original image $x_0$ and its masked image $x_0 \odot (1 - m)$ into latent space using a Variational Autoencoder (VAE) [17], where $\odot$ denotes element-wise multiplication. Then a conditional latent diffusion model performs sequential denoising with a UNet in the latent space. UNet takes as inputs the noise latent variable $\hat{z}_t$ , downsampled mask $m'$ , masked image latent $z_{Im}$ , timestep $t$ , and embedding $e_t$ of the textual prompt $P$ from a pre-trained CLIP [28] model, and predicts the noise $\epsilon_{\theta}(\hat{z}_t, z_{Im}, m', t, e_t)$ . The training of diffusion models resolves around optimizing the denoiser network $\epsilon_{\theta}$ to minimize the noise prediction loss: + +$$ +\mathcal {L} _ {\text {n o i s e}} = \left\| \epsilon - \epsilon_ {\theta} \left(\hat {z} _ {t}, z _ {I m}, m ^ {\prime}, t, e _ {t}\right) \right\| _ {2} ^ {2} \tag {1} +$$ + +# 3.2. Glyph control + +# 3.2.1. Glyph Constraint + +The accurate rendering of glyphs is critical in text editing tasks. To ensure proper formation of character strokes, we have implemented multi-faceted constraints including reference standard character images, precise textual encoding representations, and perceptual loss mechanisms. + +Glyph Guidance. Previous work [22, 39, 44] has demonstrated the effective guidance provided by standard font reference images for text content generation. However, these methods design glyph images with per-char alignment positions to accommodate the generation of arbitrary shape masks. Unfortunately, the shape of these masks may unintentionally reveal textual details, such as character count and layout (see Fig.2 problem 1). Therefore, we calculate the outer rectangle of the edit area and inject glyph information by field alignment. The standard quadrilateral mask serves as a better foundation for glyph representation, as a well-coupled relationship between glyphs and layout fosters harmonious content generation. + +In detail, we first obtain the bounding box of the text area to be injected and generate the corresponding glyph in this area with standard font in appropriate size. Next, we convert the generated glyph image $\mathcal{I}_g$ into latent space variables $l_{g}$ using the VAE [17] model. Finally, we concatenate these latent space variables with the latent representation of the image along the channel dimension to ensure that the generated image visually aligns with the original input image while clearly retaining and conveying the glyph information. + +Token-separated Text Encoder. The default text encoder of SDXL [27] is not sufficient for representing Chinese semantics. To address this limitations, we employ the ChatGLM [6] text encoder to support both Chinese and English + +![](images/6d330d78d85d480740cb3b849c637df82b4a7eb8e5c53681b849ed6af5e12a27.jpg) +Figure 3. The framework of TextMaster consists of the following components: the glyph control module, comprising the glyph constraint, update text encoder, and perception module; the layout control is achieved via position-aware attention; and the style control module encompasses the style-adapter and in-context reference. + +characters. ChatGLM employs a dual-stream self-attention architecture: one stream processes positional information to capture text sequence and structure, while the other handles content information to understand semantic relationships. + +In text editing tasks, semantic relationships can adversely affect generation outcomes, as character embeddings may differ based on context. To reduce this interference, we adopt a specialized encoding method. First, we tokenize the prompt $P = \{c_{1}, c_{2}, \dots, c_{K}\}$ by treating each character $c_{k} \in P$ as an individual token, preventing cross-semantic influence. We then sort and concatenate these tokens according to their original positions, using zero padding to ensure consistent sequence lengths. Finally, we use a text encoder $\mathcal{E}_{t}$ to encode the characters token by token, incorporating position information $p_{k}$ , and generate the corresponding mask encoding $m_{k}$ for each token $c_{k}$ . Each character is encoded as follows: + +$$ +e _ {k} = \mathcal {E} _ {t} \left(c _ {k} + \operatorname {p o s} _ {k} + m _ {k}\right), \tag {2} +$$ + +where $k\in \{1,2,\dots ,K\}$ . We obtain the final text conditional features by concatenating the character encodings: + +$$ +e _ {t} = \left[ e _ {1}; e _ {2}; \dots ; e _ {K} \right]. \tag {3} +$$ + +This approach maintains uniform character embeddings across contexts while preserving positional information, ultimately improving the generation quality in text editing tasks. + +Perception Module. We propose a method based on perceptual loss to maintain the consistency of glyph structures. By calculating the $L_{2}$ loss of the Optical Character Recognition (OCR) features in the edited region between the reconstructed image $\hat{x}_0$ and the original image $x_0$ , we impose constraints on the generation of glyph structures. + +Specifically, by denoising the noise prediction $\epsilon_{t}$ from the denoising network's output, we can derive an estimate $\hat{z}_0$ from the noise vector $z_{t}$ as follows: $\hat{z}_{0} = \frac{1}{\sqrt{\bar{\alpha}_{t}}} z_{t} + \frac{\sqrt{1 - \bar{\alpha}_{t}}}{\sqrt{\bar{\alpha}_{t}}}\epsilon_{t}$ , where $\epsilon_{t} \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , and $\alpha_{t}$ denotes the variance scheduling of the Gaussian noise. Furthermore, we utilize a VAE decoder to obtain an approximate reconstruction estimate $\hat{x}_0$ . Based on the mask $m$ of the edited region, we crop the original image $x_0$ and the reconstructed image $\hat{x}_0$ , resulting in the corresponding images of the edited area $x_0^m$ and $\hat{x}_0^m$ , respectively. We employ the PPOCR [18] model as the glyph encoder $\phi (\cdot)$ and extract features $f^{m} = \phi (x_{0}^{m})$ and $\hat{f}^{m} = \phi (\hat{x}_{0}^{m})$ from the fully connected layer to compute the L2 loss of the glyph features between the original and predicted images. + +Moreover, we found it necessary to ensure a natural transition between the text and the surrounding background. Therefore, we further introduce pixel-based Mean Squared Error (MSE) loss to ensure the accuracy of the text generation area. The resulting perceptual loss is as follows: + +$$ +\mathcal {L} _ {p} = \left\| f ^ {m} - \hat {f} ^ {m} \left\| _ {2} ^ {2} + \left\| x _ {0} ^ {m} - \hat {x} _ {0} ^ {m} \right\| _ {2}. \right. \right. \qquad (4) +$$ + +Perceptual loss injects a stronger supervisory signal by constraining the reconstruction of the editing region, ensuring the correct presentation of the text. + +![](images/62d7a394b1e974819a58e14e07b3cb5077d36506a8cb893199672fef6d19d767.jpg) +Figure 4. The attention response map processing flow: Atten represent the top chosen attention response maps for each token. + +# 3.2.2. Adaptive Typography Layout + +Position-aware Attention. We expect the model to exhibit accurate positional responses for each character within the masked regions, thereby generating text sequences with proper typographic arrangement. As revealed in prior work [40], different cross-attention blocks demonstrate distinct attention orientations, with certain layers exhibiting stronger responses to layout patterns. As illustrated in Fig. 4, our analysis of text generation tasks shows that specific attention layers demonstrate pronounced perception of character positioning. To optimize layout learning, we calculate the IoU overlap ratio between binarized attention maps and ground-truth character bounding boxes, subsequently selecting the top three layers exhibiting the strongest responses for layout adaptation. This design facilitates improved intra-paragraph layout learning, as illustrated in Fig. 2 problem 2. Specifically, we perform the following steps: + +Compute attention scores $\mathbf{A}$ using the query matrix $\mathbf{Q}$ and key matrix $\mathbf{K}$ : + +$$ +\mathbf {A} = \frac {\mathbf {Q K} ^ {\top}}{\sqrt {d _ {k}}}, \quad \mathbf {W} = \operatorname {s o f t m a x} (\mathbf {A}). \tag {5} +$$ + +Thresholding: Generate a binary image $\mathbf{T}$ by applying a threshold $T = 125$ to the averaged attention map: + +$$ +\mathbf {T} (i, j) = \left\{ \begin{array}{l l} 2 5 5 & \text {i f} \bar {\mathbf {A}} (i, j) > T \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {6} +$$ + +Gaussian Blurring: Apply Gaussian blurring to the + +thresholded image $\mathbf{T}$ using a Gaussian kernel $\mathbf{W}(k,l)$ : + +$$ +\mathbf {G} (i, j) = \sum_ {k = - K} ^ {K} \sum_ {l = - L} ^ {L} \mathbf {W} (k, l) \cdot \mathbf {T} (i + k, j + l). \tag {7} +$$ + +Complete Intersection over Union (CIOU) Loss: Calculate the CIOU loss between the derived bounding box $\mathbf{B}_d$ and the ground truth bounding box $\mathbf{B}_g$ : + +$$ +\mathcal {L} _ {\mathrm {i o u}} = 1 - \left(\operatorname {I o U} - \frac {\rho^ {2} \left(\mathbf {B} _ {d} , \mathbf {B} _ {g}\right)}{c ^ {2}} - \alpha v\right). \tag {8} +$$ + +# Where: + +- IoU: Intersection over Union of the two bounding boxes. +- $\rho(\mathbf{B}_d, \mathbf{B}_g)$ : Euclidean distance between the centers. +- $c$ : Diagonal length of the smallest enclosing box covering both bounding boxes. +- $\alpha$ : Trade-off parameter. +- $v$ : Measure of aspect ratio consistency. + +Adaptive Crop and Mask Strategy. Our experiments revealed that VAE-encoded latent representations often degrade glyph structures in extremely small text. To improve multi-scale generalization, we developed an Adaptive Crop strategy that dynamically adjusts text regions. In our experimental setup, we configured the threshold value at 9 characters. For text segments shorter than this threshold, we performed center-aligned cropping on the images and scaled them to the proportional size corresponding to a 9-character width. When the text length exceeded the threshold, we expanded the cropped region at a 1/10 scaling ratio. As demonstrated in Fig. 5, this threshold-triggered spatial adaptation significantly improves background restoration and visual coherence. + +![](images/577344334d06b8b6a665ee658f4784dd39f653ea950c27affc2ade2c675e9d4b.jpg) +Prompt:"健身房"(Gym) + +![](images/0e838a26693f587f0bc2e7645d88070c43ae949eec390075d2db917dcd44b3c1.jpg) +Figure 5. The text modification area depicted in the image is merely 30 pixels in height; TextMaster is still capable of executing high-quality edits within this constrained space. + +![](images/442ef971a276f67010e4c7cd5fc45428db240cf307504b5b8c56409586ee87e7.jpg) + +Additionally, to address uncontrollable text generation caused by mask size leakage (Fig. 2 problem 1) and mitigate layout issues from over-reliance on precise bounding dimensions, we propose an adaptive mask augmentation technique. By randomly expanding masked regions around text, this approach enhances robustness to bounding variations while improving the model's capability to generate visually coherent text layouts. + +# 3.3. Style Injection + +Current methods prioritize text accuracy and layout over style consistency. Preserving original text styles enhances visual harmony with image contexts, while integrating external references broadens editing versatility. Although recent works [15, 38] employ style control via datasets artificially constructed from font libraries (conditioned on basic attributes like color and typeface), such synthetic datasets often mismatch real-world distributions and provide coarse-grained control, failing to reproduce complex styles like gradient colors, outlined text, or 3D effects.(Fig. 2 problem 3) To overcome these limitations, we propose a Style-Adapter module with In-Context Learning for precise style manipulation. + +# 3.3.1. Style-Adapter + +In style transfer tasks, decoupling style from content is a critical challenge. Text inherently carries strong glyph information, and the segmentation boundary information carried by the style often disrupts the generation results. InstanceStyle [40] decouples image objects and their semantic content. Inspired by this, we aim to force the model to decouple font style from glyph content. + +First, to enable the model to extract more general style information rather than strictly aligning with the segmentation boundaries of the generated text, we randomly select $n, n \in \{1, 2, \dots, K\}$ words from the edit region of the original image $x_0$ and the glyph $\mathcal{G}$ , resulting in cropped images $x_0^n$ and $\mathcal{G}^n$ . We further apply a text segmentation model $S$ to separate the font style from the background in $x_0^n$ , yielding a segmented font style image $x_S = S(x_0^n)$ . + +By employing Dinov2 [26] as an image feature extractor $f$ , we obtain the features containing glyph and style information $x_{f} = f(x_{S})$ and the content features $\mathcal{G}_f = f(\mathcal{G}^n)$ . We then subtract the extracted feature vectors to obtain content-decoupled style information $\mathcal{F}_{\mathcal{S}} = x_{f} - \mathcal{G}_{f}$ , which is injected into the cross-attention layers of the UNet using IP-Adapter [45]. By jointly injecting image style features and textual prompts as conditional information into the UNet attention block, we achieve interactive fusion through cross-attention mechanisms with the edit image. Furthermore, since image style injection only requires modifications to masked regions, we restrict the style's response regions to text editing areas by computing the intersection between IPA attention maps and the mask region $m$ . + +During text style injection, treating content bias as style feature resolves the coupling issue between glyph and typographic styles. This approach fully unleashes the text-image style transfer capabilities in two key aspects: (1) Enabling preservation of original typographic styles when editing full text segments without contextual references, through injecting source image characteristics into modified content. (2) Providing external style injection functionality that al + +lows users to edit images using preferred textual exemplars through our style transfer framework. + +# 3.3.2. In-Context Reference + +Our proposed Style-Adapter follows this approach to inject textual attributes from reference characters, achieving robust font style reconstruction. However, sometimes chromatic deviations persist due to the model's tendency to derive pixel color information from the contextual surroundings of the original image. + +The model's ability to extract color information from original image contexts led us to implement self-attention mechanisms [14, 46, 49] to manage detailed stylistic features. Based on this, TextMaster introduces an efficient approach to enhance the model's understanding of stylistic nuances in reference text. As illustrated in Figure 3, our framework processes the reference character $c_{s}$ using a segmentation model $S$ . The resized and positioned $c_{s}$ is placed at the bottom of the input image, creating modified context image $x_{c}$ , and latent features $z_{t}^{c}$ . This setup allows the model to learn spatial-semantic relationships between masked regions and the added contextual reference through self-attention. + +$$ +\mathbf {Q} _ {\text {s e l f}} = l _ {\mathbf {Q}} \left(\varphi_ {\text {s e l f}} \left(z _ {t} ^ {c}\right)\right), \mathbf {K} _ {\text {s e l f}} = l _ {\mathbf {K}} \left(\varphi_ {\text {s e l f}} \left(z _ {t} ^ {c}\right)\right). \tag {9} +$$ + +$$ +\mathbf {A} _ {\text {s e l f}} = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} _ {\text {s e l f}} \mathbf {K} _ {\text {s e l f}} ^ {\top}}{\sqrt {d _ {\text {s e l f}}}}\right). \tag {10} +$$ + +Here, $l_{\mathbf{Q}}$ and $l_{\mathbf{K}}$ denotes the linear projection operation, $\varphi_{\mathrm{self}}$ represents layer normalization, and $\mathbf{A}_{\mathrm{self}}$ corresponds to the attention maps derived from self-attention computation. We term this approach In-Context Reference. + +The denoising loss is then formulated as follows: + +$$ +\mathcal {L} _ {\text {n o i s e}} = \left\| \epsilon - \epsilon_ {\theta} \left(z _ {t}, m ^ {\prime}, t, e _ {t}\right) \right\| _ {2} ^ {2} \tag {11} +$$ + +Our overall loss consists of denoising loss $\mathcal{L}_{\mathrm{noise}}$ , perceptual loss $\mathcal{L}_{\mathrm{p}}$ and CIOU loss $\mathcal{L}_{\mathrm{iou}}$ , This can be expressed as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {n o i s e}} + \beta \mathcal {L} _ {\mathrm {p}} + \gamma \mathcal {L} _ {\mathrm {i o u}} \tag {12} +$$ + +Where $\beta$ and $\gamma$ are coefficients that adjust the relative contributions of these loss terms. + +During training iterations, we observed that directly using the target reconstruction text as context reference led to an over-reliance on copying contextual glyphs to masked regions during inference, resulting in character deformation. To mitigate this, we randomly select a character as the contextual reference and completely erase it using the erasure model [36], enabling the model to learn reconstruction of remaining characters while decoupling the generated textual content from the referenced typographic style. This strategy enhances the model's capability for accurate reconstruction of color-detail stylistic features and ensures effective content-style disentanglement characteristics. + +![](images/428d9f31253c97456b495e8d66f2e6124cad92e0355bc344bb72d4e3411e3a1d.jpg) +Figure 6. Comparison of the visualized results between our method and anytext[39], anytext2[38], flux[1], and textctrl[47]. + +
Evaluation BenchmarkModelChineseEnglish
Acc↑NED↑FID↓LPIPS↓Acc↑NED↑FID↓LPIPS↓
AnyText-BenchmarkSDXL----0.08570.146437.020.0557
FLUX----0.15940.318525.670.0471
TextDiffuser2----0.54750.795131.540.1457
AnyText-v1.10.53140.643754.810.11150.51670.587743.970.1303
AnyText20.61330.712030.560.05360.62350.665728.660.0662
TextCtrl----0.80960.82316.010.0234
TextMaster0.92570.948311.930.02470.93580.953410.350.0238
ICDAR13-BenchmarkSDXL----0.55790.102440.520.0735
FLUX----0.15940.137321.640.0581
TextDiffuser2----0.59210.542540.270.1537
AnyText-v1.1----0.55790.614234.560.1536
AnyText2----0.67650.683217.650.0754
TextCtrl----0.80550.811610.770.0278
TextMaster----0.91000.936610.820.0256
TextMaster-BenchmarkSDXL----0.12320.167545.340.0894
FLUX----0.17640.283727.320.0546
TextDiffuser2----0.34270.453719.640.0846
AnyText-v1.10.54670.768127.960.09340.45380.587718.980.1303
AnyText20.57840.643719.610.06380.54990.632714.370.0794
TextCtrl----0.82230.835410.410.0293
TextMaster0.91800.95259.860.02450.91800.93459.730.0237
+ +Table 1. Evaluation Results of random text editing on five benchmarks. Accuracy (ACC, NED) is higher the better $(\uparrow)$ , while FID and LPIPS are lower the better $(\downarrow)$ . + +# 4. Experiments + +# 4.1. Dataset + +Our model was trained on the AnyWord-3M dataset containing Chinese/English texts, combining LAION [34] + +(p1/p2) and Wukong [10] (1-2/5) subsets. Given the limitations of existing open-source datasets in both image resolution and diversity, we publicly crawled 100,000 of artistic style posters from the web to enhance the model's style infusion capability during training. Specially, we + +
SettingAcc↑NED↑FID↓LPIPS↓
Base0.50980.587316.260.0454
+Token-separated0.68320.707314.580.0417
+Lp0.74510.763812.950.0369
+Liou0.78160.849513.020.0351
+in-context0.81620.874111.650.0338
+style adapter0.88600.902510.870.0286
+ +Table 2. Experimental results of different methods. + +employed an OCR character detection model to generate single-character bounding boxes for precise CIOU loss calculation in character positioning and context reference in style learning. In addition, we used ICDAR13[16], Any-text benchmark[39], and the textmaster dataset we built for evaluation. + +# 4.2. Experimental Details + +Our model was initialized using pre-trained weights from the SDXL[27] model and pre-trained IPA weights from colors[37]. The training process was conducted in two phases. First step, we train five epochs on the AnyWord-3M dataset using 4 H20 GPUs. Set the learning rate to 1e-4, the weight of perception loss $\beta$ to 0.1, the weight of iou loss $\gamma$ to 0.05, and the noise-offset of 0.02. In second stage, we continue training Unet with in-context reference on style poster dataset, with learning rate of 1e-5. Then, we train style-adapter module with freezing Unet parameters first, subsequently jointly optimizing IPA-Unet attention interactions, and ultimately enabling full parameter fine-tuning. During training, the resolution is set to 1024. + +# 4.3. Comparison Results + +# 4.3.1. Quantitative and Qualitative Results + +Following the methodology established in prior research [1, 27, 38, 39, 42, 47], we conducted a comprehensive evaluation of our model across three distinct datasets using a standardized approach. We divided the metrics into those that assess content accuracy, including OCR content accuracy and text edit distance, and for evaluating style retention, including FID and LPIPS. We ensured a fair comparison among all methods by evaluating accuracy metrics using random text, while assessing style similarity metrics with the original text. For t2i methods [1, 27, 38, 39], we supplemented them with image description prompts. As shown in Table 1, our method outperforms the current state-of-the-art methods across all metrics. We observe that TextCtrl tends to retain the original text during the editing process, which leads to reduced accuracy in content editing. However, this characteristic gives it an advantage in style similarity metrics when reconstructing from the original text. + +As shown in Figure 6, we have tested random editing and can see the great advantages of our method for editing + +arbitrary random content. Also, a large number of qualitative results (see supplementary materials) demonstrate that our model significantly outperforms existing models in both Chinese and English editing effects, particularly in complex Chinese text editing. Moreover, our general editing capability not only maintains accurate content editing but also enhances style retention ability. + +# 4.4. Ablation Study + +As shown in Table 2, we conducted ablation evaluations on 500 carefully selected images that represent various aspects of text. The results show that incorporating glyph information significantly improves the model's performance across all metrics. The perceptual losses and token-separated text encoder contribute to a substantial increase in text generation accuracy and overall image quality. Furthermore, considering the attention IoU loss further enhances the model's ability to accurately locate text in images. + +Meanwhile, the style injection module not only significantly improves the metric of style consistency but also synergistically enhances the accuracy of text editing. In our experiments, we observe that style injection is essential for text generation: in the absence of style reference, the generated arbitrary-style text tends to produce less accurate glyphs, and the resulting text is more difficult to harmoniously blend with its surroundings. + +# 5. Conclusion + +In this paper, we introduce TextMaster, a groundbreaking method that, for the first time, supports high-accuracy, consistent style text editing with rational layout capabilities. We ensured the generation of glyph structures through the design of glyph images, a token-separating encoder, and a perceptual loss method. We utilized attention IoU loss to promote the model's learning of adaptive layouts. Additionally, we introduce a style injection method that infuses the style of the modified font into the target font, decoupling semantic entanglement and preserving style information. Extensive qualitative and quantitative experimental results demonstrate that TextMaster outperforms the current state-of-the-art methods. + +# References + +[1] BlackForestLab. Flux.1. https://blackforestlabs.ai/announcing-black-forest-labs/,2024.3,7,8 +[2] Haoxing Chen, Zhuoer Xu, Zhangxuan Gu, Jun Lan, Xing Zheng, Yaohui Li, Changhua Meng, Huijia Zhu, and Weiqiang Wang. Diffuse: Universal text editing diffusion model, 2023. 2 +[3] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser-2: Unleashing the power of language models for text rendering. arXiv preprint arXiv:2311.16465, 2023. 3 +[4] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. arXiv preprint, abs/2305.10855, 2023. 2, 3 +[5] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794, 2021. 2 +[6] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. ACL, 2022. 3 +[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first international conference on machine learning*, 2024. 3 +[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 1 +[9] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1 +[10] Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Minzhe Niu, Xiaodan Liang, Lewei Yao, Runhui Huang, Wei Zhang, Xin Jiang, Chunjing Xu, and Hang Xu. Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark, 2022. 7 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2 +[12] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. CoRR, 2023. 1 +[13] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint, abs/2302.09778, 2023. 1 +[14] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers, 2024. 6 + +[15] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 2, 3, 6 +[16] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazán Almazán, and Lluis Pere de las Heras. Icdar 2013 robust reading competition. In 2013 12th International Conference on Document Analysis and Recognition, pages 1484–1493, 2013. 8 +[17] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3 +[18] Chenxia Li, Weiwei Liu, Ruoyu Guo, Xiaoting Yin, Kaitao Jiang, Yongkun Du, Yuning Du, Lingfeng Zhu, Baohua Lai, Xiaoguang Hu, Dianhai Yu, and Yanjun Ma. Pp-ocrv3: More attempts for the improvement of ultra lightweightOCR system, 2022. 4 +[19] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[20] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. arXiv preprint arXiv:2403.09622, 2024. 3 +[21] Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. Glyph-byt5-v2: A strong aesthetic baseline for accurate multilingual visual text rendering. arXiv preprint arXiv:2406.10208, 2024. 3 +[22] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently. arXiv preprint, abs/2303.17870, 2023. 2, 3 +[23] Jian Ma, Yonglin Deng, Chen Chen, Haonan Lu, and Zhenyu Yang. Glyphdraw2: Automatic generation of complex glyph posters with diffusion models and large language models. CoRR, 2024. 3 +[24] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint, abs/2302.08453, 2023. 1 +[25] OpenAI. Dall-e3. https://openai.com/index/dall-e-3/, 2023.3 +[26] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 6 +[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod + +els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 8 +[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3 +[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint, abs/2208.12242, 2022. 1 +[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Raphael Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2 +[33] Joshua Santoso, Christian Simon, and Williem. On manipulating scene text in the wild with diffusion models, 2023. 2 +[34] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs, 2021. 7 +[35] Wenda Shi, Yiren Song, Dengming Zhang, Jiaming Liu, and Xingxing Zou. Fonts: Text rendering with typography and style controls, 2025. 2 +[36] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 6 +[37] Kolars Team. Kolars: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 8 +[38] Yuxiang Tuo, Yifeng Geng, and Liefeng Bo. Anytext2: Visual text generation and editing with customizable attributes. arXiv preprint arXiv:2411.15245, 2024. 3, 6, 7, 8 +[39] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. *ICLR*, 2024. 2, 3, 7, 8 +[40] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation, 2024. 5, 6 +[41] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1 + +[42] Yibin Wang, Weizhong Zhang, Honghui Xu, and Cheng Jin. Dreamtext: High fidelity scene text synthesis, 2025. 8 +[43] Yukang Yang, Dongnan Gui, Yuhui Yuan, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. arXiv preprint, abs/2305.18259, 2023. 2, 3 +[44] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. Advances in Neural Information Processing Systems, 36, 2024. 3 +[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint, 2023. 1, 6 +[46] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and Anan Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model, 2024. 6 +[47] Weichao Zeng, Yan Shu, Zhenhang Li, Dongbao Yang, and Yu Zhou. Textctrl: Diffusion-based scene text editing with prior guidance control, 2024. 2, 7, 8 +[48] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint, abs/2302.05543, 2023. 1 +[49] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation, 2024. 6 \ No newline at end of file diff --git a/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/images.zip b/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ab01bd26b09105e9bb4f41961f38d3a2d5ef8234 --- /dev/null +++ b/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1002a57b205c3b82979d75e5f9faab842209dec793f996c305b9c4f863b99a69 +size 780330 diff --git a/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/layout.json b/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..47e4bfd82d18ef4ece930256720acd4a6661da82 --- /dev/null +++ b/2025/TextMaster_ A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control/layout.json @@ -0,0 +1,8200 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 493, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 493, + 140 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 493, + 140 + ], + "type": "text", + "content": "TextMaster: A Unified Framework for Realistic Text Editing via Glyph-Style Dual-Control" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "spans": [ + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "text", + "content": "Zhenyu Yan" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "text", + "content": ", Jian Wang" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "inline_equation", + "content": "^{1*‡}" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "text", + "content": ", Aoqiang Wang" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "inline_equation", + "content": "^{*†}" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "text", + "content": ", Yuhan Li" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "inline_equation", + "content": "^{2†}" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "text", + "content": ", Wenxiang Shang" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "text", + "content": ", Zhu Hangcheng" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "text", + "content": "Taobao & Tmall Group of Alibaba " + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "inline_equation", + "content": "\\quad^{2}" + }, + { + "bbox": [ + 73, + 160, + 536, + 191 + ], + "type": "text", + "content": "Shanghai Jiao Tong University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 193, + 514, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 193, + 514, + 205 + ], + "spans": [ + { + "bbox": [ + 93, + 193, + 514, + 205 + ], + "type": "text", + "content": "{yeeshai.yzy, alex.wj}@taobao.com, wangaq1999@gmail.com, melodious@sjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "spans": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 258, + 297, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 486 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 486 + ], + "type": "text", + "content": "In image editing tasks, high-quality text editing capabilities can significantly reduce both human and material resource costs. Existing methods, however, face significant limitations in terms of stroke accuracy for complex text and controllability of generated text styles. To address these challenges, we propose TextMaster, a solution capable of accurately editing text across various scenarios and image regions, while ensuring proper layout and controllable text style. Our method enhances the accuracy and fidelity of text rendering by incorporating high-resolution standard glyph information and applying perceptual loss within the text editing region. Additionally, we leverage an attention mechanism to compute intermediate layer bounding box regression loss for each character, enabling the model to learn text layout across varying contexts. Furthermore, we propose a novel style injection technique that enables controllable style transfer for the injected text. Through comprehensive experiments, we demonstrate the state-of-the-art performance of our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 135, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 135, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 135, + 534 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 543, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 296, + 676 + ], + "type": "text", + "content": "Recent advances in generative models have revolutionized image synthesis, with diffusion models surpassing Generative Adversarial Networks (GANs) in quality, diversity, and controllability [11, 27, 30]. While GAN-based models, such as the original Generative Adversarial Nets [9], have demonstrated impressive realism, they often suffer from training instability and mode collapse. The advent of diffusion models, such as Denoising Diffusion Probabilistic Models (DDPM) [11] and Latent Diffusion Models (LDM) [30], have led to superior image fidelity and text-image alignment, making them the dominant paradigm in" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 236, + 549, + 299 + ], + "blocks": [ + { + "bbox": [ + 324, + 236, + 549, + 299 + ], + "lines": [ + { + "bbox": [ + 324, + 236, + 549, + 299 + ], + "spans": [ + { + "bbox": [ + 324, + 236, + 549, + 299 + ], + "type": "image", + "image_path": "f521f04b12ced89c7d4c20bdaaa0ab70c45795465961721afc5126725215c03d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 324, + 299, + 548, + 365 + ], + "blocks": [ + { + "bbox": [ + 324, + 299, + 548, + 365 + ], + "lines": [ + { + "bbox": [ + 324, + 299, + 548, + 365 + ], + "spans": [ + { + "bbox": [ + 324, + 299, + 548, + 365 + ], + "type": "image", + "image_path": "b1a40186cd70b72b90d97f626b0fe13d1305df208a6912142f1d40a424713c8f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 325, + 366, + 548, + 446 + ], + "blocks": [ + { + "bbox": [ + 325, + 366, + 548, + 446 + ], + "lines": [ + { + "bbox": [ + 325, + 366, + 548, + 446 + ], + "spans": [ + { + "bbox": [ + 325, + 366, + 548, + 446 + ], + "type": "image", + "image_path": "e6a607f42e273fcb7aee7c234db8344fc2a5e345eb55bbc0bd435f6932551bcf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 460, + 555, + 495 + ], + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 495 + ], + "type": "text", + "content": "Figure 1. The image illustrates the diverse capabilities of our TextMaster, encompassing precise typesetting and layout, consistent style retention, and external style injection." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 519, + 441, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 441, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 441, + 531 + ], + "type": "text", + "content": "text-to-image (T2I) generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 532, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 532, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 532, + 555, + 628 + ], + "type": "text", + "content": "To enhance the controllability of T2I generation, researchers have explored methods to inject conditional signals. ControlNet [48] enables spatial constraints by conditioning diffusion models on structural priors, while T2I-Adapter [24] further enhances control by learning task-specific adapters. Similarly,Composer [13] introduces a framework for flexible, composable conditions, allowing intricate constraints in text-guided image synthesis." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Another critical direction involves capturing specific styles and characteristics from reference images. Techniques such as Textual Inversion [8] and DreamBooth [31] fine-tune models on personalized concepts, improving fidelity to target attributes. IP-Adapter [45] extends this by efficiently integrating image prompts with textual guidance, while Reference-Efficient [12] and InstantID [41] enable" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 223, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 223, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 223, + 713 + ], + "type": "text", + "content": "*Equal contribution. \n†Work done during an internship at Alibaba Group. \n‡Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16112" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 106, + 73, + 504, + 340 + ], + "blocks": [ + { + "bbox": [ + 106, + 73, + 504, + 340 + ], + "lines": [ + { + "bbox": [ + 106, + 73, + 504, + 340 + ], + "spans": [ + { + "bbox": [ + 106, + 73, + 504, + 340 + ], + "type": "image", + "image_path": "c0ac696c6ddcba5add81456471b5bdb237679799fd5bc28ccae4aee133215bc1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 354, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 55, + 354, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 555, + 376 + ], + "type": "text", + "content": "Figure 2. The current approach exhibits several deficiencies, including a lack of alignment between the generated text and the intended target text, suboptimal layout organization, and inconsistent in style preservation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 397, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 295, + 434 + ], + "type": "text", + "content": "identity preservation in personalized generation. Further advancements, such as PhotoMaker [19], refine identity embeddings to achieve photorealistic results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 438, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 715 + ], + "type": "text", + "content": "The main challenge of text editing lies in generating accurate textual content while minimizing any noticeable traces of editing. To Achieve this requires, the model needs to reconstruct multiple densely packed characters within localized regions, which places substantial demands on its fine-grained modeling abilities. Recent methods [22, 43] have explored glyph control for text editing. Yet, these approaches often suffer from reduced editing accuracy due to the lack of character-level textual guidance. On the other hand, effectively disentangling content from style presents another significant hurdle. Due to the highly structured nature of character glyphs, independently extracting style features from a given text image remains particularly difficult. Diffusion-based methods [2, 4, 39] attempt to restore text by leveraging the image context to preserve style. Since style information is primarily derived from unmasked regions, this may lead to inconsistencies in style representation during reconstruction. Other approaches [15, 35] seek to disentangle text style from glyph structure using manually constructed datasets, where different texts are rendered with various fonts to create style-content pairs for learning. Despite being designed to simulate real-world scenarios, such datasets still fall short in capturing the authenticity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 397, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 445 + ], + "type": "text", + "content": "and diversity found in real text images. Furthermore, some methods opt to inject reference styles through text-style image inversion [33, 47], but this comes at the cost of significantly increased inference overhead." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 448, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 604 + ], + "type": "text", + "content": "In response to these challenges, our method introduces further improvements in glyph control by combining character-level text embedding and IoU loss enabling precise regulation of both glyph shape and positioning. For style disentanglement, unlike previous approaches that rely on synthesized data, whose results are often affected by the glyphs in the reference image, we explicitly separate the glyph structure from the style appearance, which is achieved by integrating a style adapter with in-context learning. These innovations enable our model to reconstruct text across deiverse and fully authentic datasets, ultimately delivering robust style disentanglement and superior generalization performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "By integrating these advanced techniques, the proposed TextMaster framework addresses the critical challenges in text editing, offering unparalleled performance in maintaining style consistency, content accuracy, proper layout, and high-quality visual output. These advancements position TextMaster as a state-of-the-art tool for complex text editing tasks in images [5, 11, 32]. Meanwhile, Our approach is not limited to some specific languages, it can be easily extended to other languages as well." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16113" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 294, + 180 + ], + "type": "text", + "content": "Recent advancements in generative models, such as FLUX [1], DALL-E 3 [25], and SD3 [7], have significantly improved image synthesis quality, achieving high-resolution fidelity and enhanced semantic coherence. However, challenges remain in accurate text rendering within images. Specifically, ensuring precise glyph structure, spatial consistency, and stylistic control remains a complex task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 295, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 295, + 390 + ], + "type": "text", + "content": "Accurate Glyph and Spatial Control. Several studies have addressed the challenges of generating visually coherent and spatially accurate text. GlyphDraw [22] proposes a word-level synthesis approach to ensure coherent Chinese character rendering. GlyphControl [43] introduces glyph-conditioned diffusion to enhance character structure control, while TextDiffuser [4] treats text rendering as a diffusion-based painting task. Recent advancements, such as AnyText [39], extend this to multilingual text generation. Further improvements, such as TextDiffuser-2 [3], leverage large language models for refined text placement. Additionally, models like Glyph-ByT5 [20] and its successor Glyph-ByT5-v2 [21] introduce customized text encoders for multilingual and aesthetic text rendering, improving both accuracy and flexibility. Moreover, GlyphDraw2 [23] combines diffusion models with large-scale language modeling to generate complex text-based posters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 515 + ], + "type": "text", + "content": "Controllable Text Style and Appearance. Separating textual content from stylistic attributes is a longstanding challenge. DiffSTE [15] improves text editing by leveraging dual encoders trained on extensive font datasets, allowing font-specific style preservation during synthesis. However, inference still requires explicit font name input. AnyText2 [38] enhances controllability by embedding RGB-based color attributes, but complex styles such as gradients, borders, and shadows remain difficult to represent effectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 544, + 111, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 111, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 111, + 556 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Text editing tasks can be deconstructed into two aspects: glyphs and styles. As shown in the figure 3, to achieve precise editing control, we propose Glyph-Style Dual-Control. Glyph control involves maintaining structural integrity of characters through glyph constraints, and achieving cohesive text flow via adaptive typography layouts. To maintain a consistent style with the original text, we introduce style control, which includes adjustments for typeface and color, allowing for accurate referencing and injection of text styles. This dual mechanism ensures precise character rendering, harmonized layout composition, and versatile style referencing capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 394, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 394, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 394, + 85 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": "Our proposed method builds upon Stable Diffusion [29] Inpainting Models. Given an image " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": ", a binary mask " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": " specifying the editable region and a textual prompt " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": " describing the target text to generate, the system first encodes both the original image " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": " and its masked image " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "x_0 \\odot (1 - m)" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": " into latent space using a Variational Autoencoder (VAE) [17], where " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": " denotes element-wise multiplication. Then a conditional latent diffusion model performs sequential denoising with a UNet in the latent space. UNet takes as inputs the noise latent variable " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": ", downsampled mask " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "m'" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": ", masked image latent " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "z_{Im}" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": ", timestep " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": ", and embedding " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "e_t" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": " of the textual prompt " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": " from a pre-trained CLIP [28] model, and predicts the noise " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\hat{z}_t, z_{Im}, m', t, e_t)" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": ". The training of diffusion models resolves around optimizing the denoiser network " + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 89, + 555, + 269 + ], + "type": "text", + "content": " to minimize the noise prediction loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 358, + 276, + 555, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 276, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 358, + 276, + 555, + 291 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o i s e}} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\hat {z} _ {t}, z _ {I m}, m ^ {\\prime}, t, e _ {t}\\right) \\right\\| _ {2} ^ {2} \\tag {1}", + "image_path": "2b47cb0545929b41fd120f4f1a09a9c6c74a3a15449e23f60685fc9fa5436841.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 298, + 403, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 403, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 403, + 312 + ], + "type": "text", + "content": "3.2. Glyph control" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 316, + 418, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 418, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 418, + 328 + ], + "type": "text", + "content": "3.2.1. Glyph Constraint" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 331, + 554, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 390 + ], + "type": "text", + "content": "The accurate rendering of glyphs is critical in text editing tasks. To ensure proper formation of character strokes, we have implemented multi-faceted constraints including reference standard character images, precise textual encoding representations, and perceptual loss mechanisms." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 554, + 545 + ], + "type": "text", + "content": "Glyph Guidance. Previous work [22, 39, 44] has demonstrated the effective guidance provided by standard font reference images for text content generation. However, these methods design glyph images with per-char alignment positions to accommodate the generation of arbitrary shape masks. Unfortunately, the shape of these masks may unintentionally reveal textual details, such as character count and layout (see Fig.2 problem 1). Therefore, we calculate the outer rectangle of the edit area and inject glyph information by field alignment. The standard quadrilateral mask serves as a better foundation for glyph representation, as a well-coupled relationship between glyphs and layout fosters harmonious content generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 546, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 664 + ], + "type": "text", + "content": "In detail, we first obtain the bounding box of the text area to be injected and generate the corresponding glyph in this area with standard font in appropriate size. Next, we convert the generated glyph image " + }, + { + "bbox": [ + 313, + 546, + 554, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_g" + }, + { + "bbox": [ + 313, + 546, + 554, + 664 + ], + "type": "text", + "content": " into latent space variables " + }, + { + "bbox": [ + 313, + 546, + 554, + 664 + ], + "type": "inline_equation", + "content": "l_{g}" + }, + { + "bbox": [ + 313, + 546, + 554, + 664 + ], + "type": "text", + "content": " using the VAE [17] model. Finally, we concatenate these latent space variables with the latent representation of the image along the channel dimension to ensure that the generated image visually aligns with the original input image while clearly retaining and conveying the glyph information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Token-separated Text Encoder. The default text encoder of SDXL [27] is not sufficient for representing Chinese semantics. To address this limitations, we employ the ChatGLM [6] text encoder to support both Chinese and English" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16114" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 76, + 506, + 311 + ], + "blocks": [ + { + "bbox": [ + 107, + 76, + 506, + 311 + ], + "lines": [ + { + "bbox": [ + 107, + 76, + 506, + 311 + ], + "spans": [ + { + "bbox": [ + 107, + 76, + 506, + 311 + ], + "type": "image", + "image_path": "6d330d78d85d480740cb3b849c637df82b4a7eb8e5c53681b849ed6af5e12a27.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 323, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 54, + 323, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 555, + 360 + ], + "type": "text", + "content": "Figure 3. The framework of TextMaster consists of the following components: the glyph control module, comprising the glyph constraint, update text encoder, and perception module; the layout control is achieved via position-aware attention; and the style control module encompasses the style-adapter and in-context reference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 378, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 295, + 426 + ], + "type": "text", + "content": "characters. ChatGLM employs a dual-stream self-attention architecture: one stream processes positional information to capture text sequence and structure, while the other handles content information to understand semantic relationships." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "text", + "content": "In text editing tasks, semantic relationships can adversely affect generation outcomes, as character embeddings may differ based on context. To reduce this interference, we adopt a specialized encoding method. First, we tokenize the prompt " + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "inline_equation", + "content": "P = \\{c_{1}, c_{2}, \\dots, c_{K}\\}" + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "text", + "content": " by treating each character " + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "inline_equation", + "content": "c_{k} \\in P" + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "text", + "content": " as an individual token, preventing cross-semantic influence. We then sort and concatenate these tokens according to their original positions, using zero padding to ensure consistent sequence lengths. Finally, we use a text encoder " + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{t}" + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "text", + "content": " to encode the characters token by token, incorporating position information " + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "inline_equation", + "content": "p_{k}" + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "text", + "content": ", and generate the corresponding mask encoding " + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "inline_equation", + "content": "m_{k}" + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "text", + "content": " for each token " + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "inline_equation", + "content": "c_{k}" + }, + { + "bbox": [ + 54, + 426, + 295, + 581 + ], + "type": "text", + "content": ". Each character is encoded as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 119, + 590, + 295, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 590, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 119, + 590, + 295, + 604 + ], + "type": "interline_equation", + "content": "e _ {k} = \\mathcal {E} _ {t} \\left(c _ {k} + \\operatorname {p o s} _ {k} + m _ {k}\\right), \\tag {2}", + "image_path": "3c1ef9f0bedf65c485243fe48258118278c8cf321ac826d1deb3580c7b8aaa78.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "type": "inline_equation", + "content": "k\\in \\{1,2,\\dots ,K\\}" + }, + { + "bbox": [ + 55, + 612, + 295, + 636 + ], + "type": "text", + "content": " . We obtain the final text conditional features by concatenating the character encodings:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 647, + 295, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 647, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 129, + 647, + 295, + 662 + ], + "type": "interline_equation", + "content": "e _ {t} = \\left[ e _ {1}; e _ {2}; \\dots ; e _ {K} \\right]. \\tag {3}", + "image_path": "f67f15e08426733a55d983af8ada63fa0ffc28d834e84d7635628be490cf3e90.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": "This approach maintains uniform character embeddings across contexts while preserving positional information, ultimately improving the generation quality in text editing tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 378, + 553, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 553, + 449 + ], + "type": "text", + "content": "Perception Module. We propose a method based on perceptual loss to maintain the consistency of glyph structures. By calculating the " + }, + { + "bbox": [ + 313, + 378, + 553, + 449 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 378, + 553, + 449 + ], + "type": "text", + "content": " loss of the Optical Character Recognition (OCR) features in the edited region between the reconstructed image " + }, + { + "bbox": [ + 313, + 378, + 553, + 449 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 313, + 378, + 553, + 449 + ], + "type": "text", + "content": " and the original image " + }, + { + "bbox": [ + 313, + 378, + 553, + 449 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 378, + 553, + 449 + ], + "type": "text", + "content": ", we impose constraints on the generation of glyph structures." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": "Specifically, by denoising the noise prediction " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\epsilon_{t}" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": " from the denoising network's output, we can derive an estimate " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": " from the noise vector " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": " as follows: " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0} = \\frac{1}{\\sqrt{\\bar{\\alpha}_{t}}} z_{t} + \\frac{\\sqrt{1 - \\bar{\\alpha}_{t}}}{\\sqrt{\\bar{\\alpha}_{t}}}\\epsilon_{t}" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\epsilon_{t} \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": " denotes the variance scheduling of the Gaussian noise. Furthermore, we utilize a VAE decoder to obtain an approximate reconstruction estimate " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": ". Based on the mask " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": " of the edited region, we crop the original image " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": " and the reconstructed image " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": ", resulting in the corresponding images of the edited area " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "x_0^m" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\hat{x}_0^m" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": ", respectively. We employ the PPOCR [18] model as the glyph encoder " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": " and extract features " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "f^{m} = \\phi (x_{0}^{m})" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "inline_equation", + "content": "\\hat{f}^{m} = \\phi (\\hat{x}_{0}^{m})" + }, + { + "bbox": [ + 313, + 450, + 555, + 624 + ], + "type": "text", + "content": " from the fully connected layer to compute the L2 loss of the glyph features between the original and predicted images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 625, + 556, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 556, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 556, + 685 + ], + "type": "text", + "content": "Moreover, we found it necessary to ensure a natural transition between the text and the surrounding background. Therefore, we further introduce pixel-based Mean Squared Error (MSE) loss to ensure the accuracy of the text generation area. The resulting perceptual loss is as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 695, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 695, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 354, + 695, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\left\\| f ^ {m} - \\hat {f} ^ {m} \\left\\| _ {2} ^ {2} + \\left\\| x _ {0} ^ {m} - \\hat {x} _ {0} ^ {m} \\right\\| _ {2}. \\right. \\right. \\qquad (4)", + "image_path": "803848fe299927efb6a84bf0fac80bc16513d70b65c008e6492a83ab20798cc5.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16115" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "Perceptual loss injects a stronger supervisory signal by constraining the reconstruction of the editing region, ensuring the correct presentation of the text." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 125, + 282, + 281 + ], + "blocks": [ + { + "bbox": [ + 65, + 125, + 282, + 281 + ], + "lines": [ + { + "bbox": [ + 65, + 125, + 282, + 281 + ], + "spans": [ + { + "bbox": [ + 65, + 125, + 282, + 281 + ], + "type": "image", + "image_path": "62d7a394b1e974819a58e14e07b3cb5077d36506a8cb893199672fef6d19d767.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 299, + 295, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 299, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 295, + 322 + ], + "type": "text", + "content": "Figure 4. The attention response map processing flow: Atten represent the top chosen attention response maps for each token." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 341, + 211, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 211, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 211, + 354 + ], + "type": "text", + "content": "3.2.2. Adaptive Typography Layout" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 356, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 295, + 559 + ], + "type": "text", + "content": "Position-aware Attention. We expect the model to exhibit accurate positional responses for each character within the masked regions, thereby generating text sequences with proper typographic arrangement. As revealed in prior work [40], different cross-attention blocks demonstrate distinct attention orientations, with certain layers exhibiting stronger responses to layout patterns. As illustrated in Fig. 4, our analysis of text generation tasks shows that specific attention layers demonstrate pronounced perception of character positioning. To optimize layout learning, we calculate the IoU overlap ratio between binarized attention maps and ground-truth character bounding boxes, subsequently selecting the top three layers exhibiting the strongest responses for layout adaptation. This design facilitates improved intra-paragraph layout learning, as illustrated in Fig. 2 problem 2. Specifically, we perform the following steps:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 559, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 583 + ], + "type": "text", + "content": "Compute attention scores " + }, + { + "bbox": [ + 55, + 559, + 295, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 559, + 295, + 583 + ], + "type": "text", + "content": " using the query matrix " + }, + { + "bbox": [ + 55, + 559, + 295, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 559, + 295, + 583 + ], + "type": "text", + "content": " and key matrix " + }, + { + "bbox": [ + 55, + 559, + 295, + 583 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 559, + 295, + 583 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 104, + 591, + 295, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 591, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 104, + 591, + 295, + 620 + ], + "type": "interline_equation", + "content": "\\mathbf {A} = \\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {d _ {k}}}, \\quad \\mathbf {W} = \\operatorname {s o f t m a x} (\\mathbf {A}). \\tag {5}", + "image_path": "01580cfb6284a3abc016a85bd1f6a630f9f3fe60522383463c5a35fdf0fd91b1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "type": "text", + "content": "Thresholding: Generate a binary image " + }, + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "type": "text", + "content": " by applying a threshold " + }, + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "type": "inline_equation", + "content": "T = 125" + }, + { + "bbox": [ + 55, + 628, + 296, + 652 + ], + "type": "text", + "content": " to the averaged attention map:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 106, + 661, + 295, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 661, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 106, + 661, + 295, + 693 + ], + "type": "interline_equation", + "content": "\\mathbf {T} (i, j) = \\left\\{ \\begin{array}{l l} 2 5 5 & \\text {i f} \\bar {\\mathbf {A}} (i, j) > T \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}", + "image_path": "b6d7f0127ee1c5d16e590b0d853408d2c06ea0f66f6dcf4eac56fd95761d72eb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "Gaussian Blurring: Apply Gaussian blurring to the" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "type": "text", + "content": "thresholded image " + }, + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "type": "text", + "content": " using a Gaussian kernel " + }, + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "type": "inline_equation", + "content": "\\mathbf{W}(k,l)" + }, + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 93, + 555, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 93, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 329, + 93, + 555, + 127 + ], + "type": "interline_equation", + "content": "\\mathbf {G} (i, j) = \\sum_ {k = - K} ^ {K} \\sum_ {l = - L} ^ {L} \\mathbf {W} (k, l) \\cdot \\mathbf {T} (i + k, j + l). \\tag {7}", + "image_path": "ede21bd5903349865c7bbf0ff824c20d745962bc5fd8906fd555f254522a108c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 134, + 555, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 134, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 555, + 171 + ], + "type": "text", + "content": "Complete Intersection over Union (CIOU) Loss: Calculate the CIOU loss between the derived bounding box " + }, + { + "bbox": [ + 313, + 134, + 555, + 171 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_d" + }, + { + "bbox": [ + 313, + 134, + 555, + 171 + ], + "type": "text", + "content": " and the ground truth bounding box " + }, + { + "bbox": [ + 313, + 134, + 555, + 171 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_g" + }, + { + "bbox": [ + 313, + 134, + 555, + 171 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 179, + 555, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 179, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 348, + 179, + 555, + 207 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {i o u}} = 1 - \\left(\\operatorname {I o U} - \\frac {\\rho^ {2} \\left(\\mathbf {B} _ {d} , \\mathbf {B} _ {g}\\right)}{c ^ {2}} - \\alpha v\\right). \\tag {8}", + "image_path": "186cc6a2ed02e2d39b548c57ea0373edebff29e619d166ca5155ff14c7ece84c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 213, + 347, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 213, + 347, + 224 + ], + "spans": [ + { + "bbox": [ + 314, + 213, + 347, + 224 + ], + "type": "text", + "content": "Where:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 225, + 553, + 297 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 225, + 552, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 225, + 552, + 237 + ], + "spans": [ + { + "bbox": [ + 314, + 225, + 552, + 237 + ], + "type": "text", + "content": "- IoU: Intersection over Union of the two bounding boxes." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 237, + 533, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 237, + 533, + 249 + ], + "spans": [ + { + "bbox": [ + 314, + 237, + 533, + 249 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 237, + 533, + 249 + ], + "type": "inline_equation", + "content": "\\rho(\\mathbf{B}_d, \\mathbf{B}_g)" + }, + { + "bbox": [ + 314, + 237, + 533, + 249 + ], + "type": "text", + "content": ": Euclidean distance between the centers." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 249, + 553, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 249, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 314, + 249, + 553, + 273 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 249, + 553, + 273 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 314, + 249, + 553, + 273 + ], + "type": "text", + "content": ": Diagonal length of the smallest enclosing box covering both bounding boxes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 274, + 421, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 421, + 285 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 421, + 285 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 274, + 421, + 285 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 314, + 274, + 421, + 285 + ], + "type": "text", + "content": ": Trade-off parameter." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 285, + 481, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 285, + 481, + 297 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 481, + 297 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 314, + 285, + 481, + 297 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 314, + 285, + 481, + 297 + ], + "type": "text", + "content": " : Measure of aspect ratio consistency." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 297, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 464 + ], + "type": "text", + "content": "Adaptive Crop and Mask Strategy. Our experiments revealed that VAE-encoded latent representations often degrade glyph structures in extremely small text. To improve multi-scale generalization, we developed an Adaptive Crop strategy that dynamically adjusts text regions. In our experimental setup, we configured the threshold value at 9 characters. For text segments shorter than this threshold, we performed center-aligned cropping on the images and scaled them to the proportional size corresponding to a 9-character width. When the text length exceeded the threshold, we expanded the cropped region at a 1/10 scaling ratio. As demonstrated in Fig. 5, this threshold-triggered spatial adaptation significantly improves background restoration and visual coherence." + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 326, + 478, + 394, + 549 + ], + "blocks": [ + { + "bbox": [ + 326, + 478, + 394, + 549 + ], + "lines": [ + { + "bbox": [ + 326, + 478, + 394, + 549 + ], + "spans": [ + { + "bbox": [ + 326, + 478, + 394, + 549 + ], + "type": "image", + "image_path": "577344334d06b8b6a665ee658f4784dd39f653ea950c27affc2ade2c675e9d4b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 550, + 382, + 559 + ], + "lines": [ + { + "bbox": [ + 327, + 550, + 382, + 559 + ], + "spans": [ + { + "bbox": [ + 327, + 550, + 382, + 559 + ], + "type": "text", + "content": "Prompt:\"健身房\"(Gym)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 406, + 478, + 472, + 549 + ], + "blocks": [ + { + "bbox": [ + 406, + 478, + 472, + 549 + ], + "lines": [ + { + "bbox": [ + 406, + 478, + 472, + 549 + ], + "spans": [ + { + "bbox": [ + 406, + 478, + 472, + 549 + ], + "type": "image", + "image_path": "0e838a26693f587f0bc2e7645d88070c43ae949eec390075d2db917dcd44b3c1.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 570, + 555, + 603 + ], + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 603 + ], + "type": "text", + "content": "Figure 5. The text modification area depicted in the image is merely 30 pixels in height; TextMaster is still capable of executing high-quality edits within this constrained space." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 475, + 478, + 541, + 549 + ], + "blocks": [ + { + "bbox": [ + 475, + 478, + 541, + 549 + ], + "lines": [ + { + "bbox": [ + 475, + 478, + 541, + 549 + ], + "spans": [ + { + "bbox": [ + 475, + 478, + 541, + 549 + ], + "type": "image", + "image_path": "442ef971a276f67010e4c7cd5fc45428db240cf307504b5b8c56409586ee87e7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "Additionally, to address uncontrollable text generation caused by mask size leakage (Fig. 2 problem 1) and mitigate layout issues from over-reliance on precise bounding dimensions, we propose an adaptive mask augmentation technique. By randomly expanding masked regions around text, this approach enhances robustness to bounding variations while improving the model's capability to generate visually coherent text layouts." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16116" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 146, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 146, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 146, + 85 + ], + "type": "text", + "content": "3.3. Style Injection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 245 + ], + "type": "text", + "content": "Current methods prioritize text accuracy and layout over style consistency. Preserving original text styles enhances visual harmony with image contexts, while integrating external references broadens editing versatility. Although recent works [15, 38] employ style control via datasets artificially constructed from font libraries (conditioned on basic attributes like color and typeface), such synthetic datasets often mismatch real-world distributions and provide coarse-grained control, failing to reproduce complex styles like gradient colors, outlined text, or 3D effects.(Fig. 2 problem 3) To overcome these limitations, we propose a Style-Adapter module with In-Context Learning for precise style manipulation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 254, + 144, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 144, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 144, + 266 + ], + "type": "text", + "content": "3.3.1. Style-Adapter" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 353 + ], + "type": "text", + "content": "In style transfer tasks, decoupling style from content is a critical challenge. Text inherently carries strong glyph information, and the segmentation boundary information carried by the style often disrupts the generation results. InstanceStyle [40] decouples image objects and their semantic content. Inspired by this, we aim to force the model to decouple font style from glyph content." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": "First, to enable the model to extract more general style information rather than strictly aligning with the segmentation boundaries of the generated text, we randomly select " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "n, n \\in \\{1, 2, \\dots, K\\}" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": " words from the edit region of the original image " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": " and the glyph " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": ", resulting in cropped images " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "x_0^n" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^n" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": ". We further apply a text segmentation model " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": " to separate the font style from the background in " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "x_0^n" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": ", yielding a segmented font style image " + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "inline_equation", + "content": "x_S = S(x_0^n)" + }, + { + "bbox": [ + 55, + 354, + 296, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": "By employing Dinov2 [26] as an image feature extractor " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": ", we obtain the features containing glyph and style information " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "x_{f} = f(x_{S})" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": " and the content features " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_f = f(\\mathcal{G}^n)" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": ". We then subtract the extracted feature vectors to obtain content-decoupled style information " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathcal{S}} = x_{f} - \\mathcal{G}_{f}" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": ", which is injected into the cross-attention layers of the UNet using IP-Adapter [45]. By jointly injecting image style features and textual prompts as conditional information into the UNet attention block, we achieve interactive fusion through cross-attention mechanisms with the edit image. Furthermore, since image style injection only requires modifications to masked regions, we restrict the style's response regions to text editing areas by computing the intersection between IPA attention maps and the mask region " + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 450, + 295, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "During text style injection, treating content bias as style feature resolves the coupling issue between glyph and typographic styles. This approach fully unleashes the text-image style transfer capabilities in two key aspects: (1) Enabling preservation of original typographic styles when editing full text segments without contextual references, through injecting source image characteristics into modified content. (2) Providing external style injection functionality that al" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "lows users to edit images using preferred textual exemplars through our style transfer framework." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 101, + 435, + 111 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 101, + 435, + 111 + ], + "spans": [ + { + "bbox": [ + 313, + 101, + 435, + 111 + ], + "type": "text", + "content": "3.3.2. In-Context Reference" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 114, + 555, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 555, + 186 + ], + "type": "text", + "content": "Our proposed Style-Adapter follows this approach to inject textual attributes from reference characters, achieving robust font style reconstruction. However, sometimes chromatic deviations persist due to the model's tendency to derive pixel color information from the contextual surroundings of the original image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "type": "text", + "content": "The model's ability to extract color information from original image contexts led us to implement self-attention mechanisms [14, 46, 49] to manage detailed stylistic features. Based on this, TextMaster introduces an efficient approach to enhance the model's understanding of stylistic nuances in reference text. As illustrated in Figure 3, our framework processes the reference character " + }, + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "type": "inline_equation", + "content": "c_{s}" + }, + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "type": "text", + "content": " using a segmentation model " + }, + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "type": "text", + "content": ". The resized and positioned " + }, + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "type": "inline_equation", + "content": "c_{s}" + }, + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "type": "text", + "content": " is placed at the bottom of the input image, creating modified context image " + }, + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "type": "inline_equation", + "content": "x_{c}" + }, + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "type": "text", + "content": ", and latent features " + }, + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "type": "inline_equation", + "content": "z_{t}^{c}" + }, + { + "bbox": [ + 313, + 188, + 556, + 342 + ], + "type": "text", + "content": ". This setup allows the model to learn spatial-semantic relationships between masked regions and the added contextual reference through self-attention." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 344, + 360, + 555, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 360, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 344, + 360, + 555, + 373 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {\\text {s e l f}} = l _ {\\mathbf {Q}} \\left(\\varphi_ {\\text {s e l f}} \\left(z _ {t} ^ {c}\\right)\\right), \\mathbf {K} _ {\\text {s e l f}} = l _ {\\mathbf {K}} \\left(\\varphi_ {\\text {s e l f}} \\left(z _ {t} ^ {c}\\right)\\right). \\tag {9}", + "image_path": "6029659a85587d28e9899a87342f3b7470070efbe38a9ee5c6c4a84befaec409.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 373, + 386, + 554, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 386, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 373, + 386, + 554, + 413 + ], + "type": "interline_equation", + "content": "\\mathbf {A} _ {\\text {s e l f}} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {\\text {s e l f}} \\mathbf {K} _ {\\text {s e l f}} ^ {\\top}}{\\sqrt {d _ {\\text {s e l f}}}}\\right). \\tag {10}", + "image_path": "939a5856975d4811289801c4a9e76d45f4fb574b6c9b3ea82ede59c7e61ff97f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "inline_equation", + "content": "l_{\\mathbf{Q}}" + }, + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "inline_equation", + "content": "l_{\\mathbf{K}}" + }, + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "text", + "content": " denotes the linear projection operation, " + }, + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "inline_equation", + "content": "\\varphi_{\\mathrm{self}}" + }, + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "text", + "content": " represents layer normalization, and " + }, + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_{\\mathrm{self}}" + }, + { + "bbox": [ + 313, + 415, + 554, + 462 + ], + "type": "text", + "content": " corresponds to the attention maps derived from self-attention computation. We term this approach In-Context Reference." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 463, + 525, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 463, + 525, + 474 + ], + "spans": [ + { + "bbox": [ + 326, + 463, + 525, + 474 + ], + "type": "text", + "content": "The denoising loss is then formulated as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 369, + 479, + 553, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 479, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 369, + 479, + 553, + 495 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o i s e}} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, m ^ {\\prime}, t, e _ {t}\\right) \\right\\| _ {2} ^ {2} \\tag {11}", + "image_path": "820b9f9795827a993f99a2595a11119c2d7be2ef1a80650bd7cad7d29d221a7f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 498, + 554, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 524 + ], + "type": "text", + "content": "Our overall loss consists of denoising loss " + }, + { + "bbox": [ + 313, + 498, + 554, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{noise}}" + }, + { + "bbox": [ + 313, + 498, + 554, + 524 + ], + "type": "text", + "content": ", perceptual loss " + }, + { + "bbox": [ + 313, + 498, + 554, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{p}}" + }, + { + "bbox": [ + 313, + 498, + 554, + 524 + ], + "type": "text", + "content": " and CIOU loss " + }, + { + "bbox": [ + 313, + 498, + 554, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{iou}}" + }, + { + "bbox": [ + 313, + 498, + 554, + 524 + ], + "type": "text", + "content": ", This can be expressed as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 380, + 529, + 553, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 529, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 380, + 529, + 553, + 541 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {n o i s e}} + \\beta \\mathcal {L} _ {\\mathrm {p}} + \\gamma \\mathcal {L} _ {\\mathrm {i o u}} \\tag {12}", + "image_path": "398294953d1f8bc6223e6c18950f57b64d4aeaf8469449eba651b908c871b38a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 546, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 568 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 546, + 554, + 568 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 546, + 554, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 546, + 554, + 568 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 546, + 554, + 568 + ], + "type": "text", + "content": " are coefficients that adjust the relative contributions of these loss terms." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "During training iterations, we observed that directly using the target reconstruction text as context reference led to an over-reliance on copying contextual glyphs to masked regions during inference, resulting in character deformation. To mitigate this, we randomly select a character as the contextual reference and completely erase it using the erasure model [36], enabling the model to learn reconstruction of remaining characters while decoupling the generated textual content from the referenced typographic style. This strategy enhances the model's capability for accurate reconstruction of color-detail stylistic features and ensures effective content-style disentanglement characteristics." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16117" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 84, + 72, + 526, + 293 + ], + "blocks": [ + { + "bbox": [ + 84, + 72, + 526, + 293 + ], + "lines": [ + { + "bbox": [ + 84, + 72, + 526, + 293 + ], + "spans": [ + { + "bbox": [ + 84, + 72, + 526, + 293 + ], + "type": "image", + "image_path": "428d9f31253c97456b495e8d66f2e6124cad92e0355bc344bb72d4e3411e3a1d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 305, + 526, + 317 + ], + "lines": [ + { + "bbox": [ + 82, + 305, + 526, + 317 + ], + "spans": [ + { + "bbox": [ + 82, + 305, + 526, + 317 + ], + "type": "text", + "content": "Figure 6. Comparison of the visualized results between our method and anytext[39], anytext2[38], flux[1], and textctrl[47]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 85, + 327, + 529, + 609 + ], + "blocks": [ + { + "bbox": [ + 85, + 327, + 529, + 609 + ], + "lines": [ + { + "bbox": [ + 85, + 327, + 529, + 609 + ], + "spans": [ + { + "bbox": [ + 85, + 327, + 529, + 609 + ], + "type": "table", + "html": "
Evaluation BenchmarkModelChineseEnglish
Acc↑NED↑FID↓LPIPS↓Acc↑NED↑FID↓LPIPS↓
AnyText-BenchmarkSDXL----0.08570.146437.020.0557
FLUX----0.15940.318525.670.0471
TextDiffuser2----0.54750.795131.540.1457
AnyText-v1.10.53140.643754.810.11150.51670.587743.970.1303
AnyText20.61330.712030.560.05360.62350.665728.660.0662
TextCtrl----0.80960.82316.010.0234
TextMaster0.92570.948311.930.02470.93580.953410.350.0238
ICDAR13-BenchmarkSDXL----0.55790.102440.520.0735
FLUX----0.15940.137321.640.0581
TextDiffuser2----0.59210.542540.270.1537
AnyText-v1.1----0.55790.614234.560.1536
AnyText2----0.67650.683217.650.0754
TextCtrl----0.80550.811610.770.0278
TextMaster----0.91000.936610.820.0256
TextMaster-BenchmarkSDXL----0.12320.167545.340.0894
FLUX----0.17640.283727.320.0546
TextDiffuser2----0.34270.453719.640.0846
AnyText-v1.10.54670.768127.960.09340.45380.587718.980.1303
AnyText20.57840.643719.610.06380.54990.632714.370.0794
TextCtrl----0.82230.835410.410.0293
TextMaster0.91800.95259.860.02450.91800.93459.730.0237
", + "image_path": "7163fc2a7415149ffcf61f2792e25af2da4e7279b3e440eab13fc7e19ce8ca80.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 617, + 555, + 640 + ], + "lines": [ + { + "bbox": [ + 55, + 617, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 555, + 640 + ], + "type": "text", + "content": "Table 1. Evaluation Results of random text editing on five benchmarks. Accuracy (ACC, NED) is higher the better " + }, + { + "bbox": [ + 55, + 617, + 555, + 640 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 55, + 617, + 555, + 640 + ], + "type": "text", + "content": ", while FID and LPIPS are lower the better " + }, + { + "bbox": [ + 55, + 617, + 555, + 640 + ], + "type": "inline_equation", + "content": "(\\downarrow)" + }, + { + "bbox": [ + 55, + 617, + 555, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 648, + 137, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 137, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 137, + 662 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 670, + 114, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 114, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 114, + 681 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Our model was trained on the AnyWord-3M dataset containing Chinese/English texts, combining LAION [34]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 650, + 555, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 555, + 711 + ], + "type": "text", + "content": "(p1/p2) and Wukong [10] (1-2/5) subsets. Given the limitations of existing open-source datasets in both image resolution and diversity, we publicly crawled 100,000 of artistic style posters from the web to enhance the model's style infusion capability during training. Specially, we" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16118" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 173, + 70, + 438, + 167 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 438, + 167 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 438, + 167 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 438, + 167 + ], + "type": "table", + "html": "
SettingAcc↑NED↑FID↓LPIPS↓
Base0.50980.587316.260.0454
+Token-separated0.68320.707314.580.0417
+Lp0.74510.763812.950.0369
+Liou0.78160.849513.020.0351
+in-context0.81620.874111.650.0338
+style adapter0.88600.902510.870.0286
", + "image_path": "bdbe5c00d47a426ae4a36bb9c30b149843e293fe93496337effbc6c769da555c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 211, + 175, + 398, + 186 + ], + "lines": [ + { + "bbox": [ + 211, + 175, + 398, + 186 + ], + "spans": [ + { + "bbox": [ + 211, + 175, + 398, + 186 + ], + "type": "text", + "content": "Table 2. Experimental results of different methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 296, + 268 + ], + "type": "text", + "content": "employed an OCR character detection model to generate single-character bounding boxes for precise CIOU loss calculation in character positioning and context reference in style learning. In addition, we used ICDAR13[16], Any-text benchmark[39], and the textmaster dataset we built for evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 178, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 178, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 178, + 288 + ], + "type": "text", + "content": "4.2. Experimental Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 292, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 448 + ], + "type": "text", + "content": "Our model was initialized using pre-trained weights from the SDXL[27] model and pre-trained IPA weights from colors[37]. The training process was conducted in two phases. First step, we train five epochs on the AnyWord-3M dataset using 4 H20 GPUs. Set the learning rate to 1e-4, the weight of perception loss " + }, + { + "bbox": [ + 55, + 292, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 292, + 296, + 448 + ], + "type": "text", + "content": " to 0.1, the weight of iou loss " + }, + { + "bbox": [ + 55, + 292, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 292, + 296, + 448 + ], + "type": "text", + "content": " to 0.05, and the noise-offset of 0.02. In second stage, we continue training Unet with in-context reference on style poster dataset, with learning rate of 1e-5. Then, we train style-adapter module with freezing Unet parameters first, subsequently jointly optimizing IPA-Unet attention interactions, and ultimately enabling full parameter fine-tuning. During training, the resolution is set to 1024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 454, + 174, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 174, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 174, + 466 + ], + "type": "text", + "content": "4.3. Comparison Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 471, + 239, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 239, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 239, + 483 + ], + "type": "text", + "content": "4.3.1. Quantitative and Qualitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 689 + ], + "type": "text", + "content": "Following the methodology established in prior research [1, 27, 38, 39, 42, 47], we conducted a comprehensive evaluation of our model across three distinct datasets using a standardized approach. We divided the metrics into those that assess content accuracy, including OCR content accuracy and text edit distance, and for evaluating style retention, including FID and LPIPS. We ensured a fair comparison among all methods by evaluating accuracy metrics using random text, while assessing style similarity metrics with the original text. For t2i methods [1, 27, 38, 39], we supplemented them with image description prompts. As shown in Table 1, our method outperforms the current state-of-the-art methods across all metrics. We observe that TextCtrl tends to retain the original text during the editing process, which leads to reduced accuracy in content editing. However, this characteristic gives it an advantage in style similarity metrics when reconstructing from the original text." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "As shown in Figure 6, we have tested random editing and can see the great advantages of our method for editing" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 197, + 554, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 197, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 197, + 554, + 281 + ], + "type": "text", + "content": "arbitrary random content. Also, a large number of qualitative results (see supplementary materials) demonstrate that our model significantly outperforms existing models in both Chinese and English editing effects, particularly in complex Chinese text editing. Moreover, our general editing capability not only maintains accurate content editing but also enhances style retention ability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 295, + 408, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 408, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 408, + 308 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 314, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 422 + ], + "type": "text", + "content": "As shown in Table 2, we conducted ablation evaluations on 500 carefully selected images that represent various aspects of text. The results show that incorporating glyph information significantly improves the model's performance across all metrics. The perceptual losses and token-separated text encoder contribute to a substantial increase in text generation accuracy and overall image quality. Furthermore, considering the attention IoU loss further enhances the model's ability to accurately locate text in images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 424, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 554, + 520 + ], + "type": "text", + "content": "Meanwhile, the style injection module not only significantly improves the metric of style consistency but also synergistically enhances the accuracy of text editing. In our experiments, we observe that style injection is essential for text generation: in the absence of style reference, the generated arbitrary-style text tends to produce less accurate glyphs, and the resulting text is more difficult to harmoniously blend with its surroundings." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 559, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 554, + 712 + ], + "type": "text", + "content": "In this paper, we introduce TextMaster, a groundbreaking method that, for the first time, supports high-accuracy, consistent style text editing with rational layout capabilities. We ensured the generation of glyph structures through the design of glyph images, a token-separating encoder, and a perceptual loss method. We utilized attention IoU loss to promote the model's learning of adaptive layouts. Additionally, we introduce a style injection method that infuses the style of the modified font into the target font, decoupling semantic entanglement and preserving style information. Extensive qualitative and quantitative experimental results demonstrate that TextMaster outperforms the current state-of-the-art methods." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16119" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] BlackForestLab. Flux.1. https://blackforestlabs.ai/announcing-black-forest-labs/,2024.3,7,8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 168 + ], + "type": "text", + "content": "[2] Haoxing Chen, Zhuoer Xu, Zhangxuan Gu, Jun Lan, Xing Zheng, Yaohui Li, Changhua Meng, Huijia Zhu, and Weiqiang Wang. Diffuse: Universal text editing diffusion model, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 171, + 296, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 214 + ], + "type": "text", + "content": "[3] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser-2: Unleashing the power of language models for text rendering. arXiv preprint arXiv:2311.16465, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 216, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 216, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 296, + 249 + ], + "type": "text", + "content": "[4] Jingye Chen, Yupan Huang, Tengchao Lv, Lei Cui, Qifeng Chen, and Furu Wei. Textdiffuser: Diffusion models as text painters. arXiv preprint, abs/2305.10855, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 251, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 283 + ], + "type": "text", + "content": "[5] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, pages 8780-8794, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 285, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 285, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 296, + 327 + ], + "type": "text", + "content": "[6] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. ACL, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 330, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 330, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 61, + 330, + 296, + 396 + ], + "type": "text", + "content": "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first international conference on machine learning*, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 397, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 397, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 61, + 397, + 296, + 441 + ], + "type": "text", + "content": "[8] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 443, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 443, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 61, + 443, + 296, + 485 + ], + "type": "text", + "content": "[9] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 487, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 296, + 542 + ], + "type": "text", + "content": "[10] Jiaxi Gu, Xiaojun Meng, Guansong Lu, Lu Hou, Minzhe Niu, Xiaodan Liang, Lewei Yao, Runhui Huang, Wei Zhang, Xin Jiang, Chunjing Xu, and Hang Xu. Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 296, + 577 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 578, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 296, + 621 + ], + "type": "text", + "content": "[12] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. CoRR, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "type": "text", + "content": "[13] Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou.Composer: Creative and controllable image synthesis with composable conditions. arXiv preprint, abs/2302.09778, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 669, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 712 + ], + "type": "text", + "content": "[14] Lianghua Huang, Wei Wang, Zhi-Fan Wu, Yupeng Shi, Huanzhang Dou, Chen Liang, Yutong Feng, Yu Liu, and Jingren Zhou. In-context lora for diffusion transformers, 2024. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] Jiabao Ji, Guanhua Zhang, Zhaowen Wang, Bairu Hou, Zhifei Zhang, Brian Price, and Shiyu Chang. Improving diffusion models for scene text editing with dual encoders. arXiv preprint arXiv:2304.05568, 2023. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 194 + ], + "type": "text", + "content": "[16] Dimosthenis Karatzas, Faisal Shafait, Seiichi Uchida, Masakazu Iwamura, Lluis Gomez i Bigorda, Sergi Robles Mestre, Joan Mas, David Fernandez Mota, Jon Almazán Almazán, and Lluis Pere de las Heras. Icdar 2013 robust reading competition. In 2013 12th International Conference on Document Analysis and Recognition, pages 1484–1493, 2013. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 217 + ], + "type": "text", + "content": "[17] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 273 + ], + "type": "text", + "content": "[18] Chenxia Li, Weiwei Liu, Ruoyu Guo, Xiaoting Yin, Kaitao Jiang, Yongkun Du, Yuning Du, Lingfeng Zhu, Baohua Lai, Xiaoguang Hu, Dianhai Yu, and Yanjun Ma. Pp-ocrv3: More attempts for the improvement of ultra lightweightOCR system, 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 275, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 330 + ], + "type": "text", + "content": "[19] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "text", + "content": "[20] Zeyu Liu, Weicong Liang, Zhanhao Liang, Chong Luo, Ji Li, Gao Huang, and Yuhui Yuan. Glyph-byt5: A customized text encoder for accurate visual text rendering. arXiv preprint arXiv:2403.09622, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 421 + ], + "type": "text", + "content": "[21] Zeyu Liu, Weicong Liang, Yiming Zhao, Bohan Chen, Ji Li, and Yuhui Yuan. Glyph-byt5-v2: A strong aesthetic baseline for accurate multilingual visual text rendering. arXiv preprint arXiv:2406.10208, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "text", + "content": "[22] Jian Ma, Mingjun Zhao, Chen Chen, Ruichen Wang, Di Niu, Haonan Lu, and Xiaodong Lin. Glyphdraw: Learning to draw Chinese characters in image synthesis models coherently. arXiv preprint, abs/2303.17870, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "type": "text", + "content": "[23] Jian Ma, Yonglin Deng, Chen Chen, Haonan Lu, and Zhenyu Yang. Glyphdraw2: Automatic generation of complex glyph posters with diffusion models and large language models. CoRR, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 555 + ], + "type": "text", + "content": "[24] Chong Mou, Xintao Wang, Liangbin Xie, Jian Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint, abs/2302.08453, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 557, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 578 + ], + "type": "text", + "content": "[25] OpenAI. Dall-e3. https://openai.com/index/dall-e-3/, 2023.3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 580, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 678 + ], + "type": "text", + "content": "[26] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2024. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[27] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion mod" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16120" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 94 + ], + "type": "text", + "content": "els for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "type": "text", + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 183 + ], + "type": "text", + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[31] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint, abs/2208.12242, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 348 + ], + "type": "text", + "content": "[32] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Raphael Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 380 + ], + "type": "text", + "content": "[33] Joshua Santoso, Christian Simon, and Williem. On manipulating scene text in the wild with diffusion models, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 435 + ], + "type": "text", + "content": "[34] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 470 + ], + "type": "text", + "content": "[35] Wenda Shi, Yiren Song, Dengming Zhang, Jiaming Liu, and Xingxing Zou. Fonts: Text rendering with typography and style controls, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 535 + ], + "type": "text", + "content": "[36] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. arXiv preprint arXiv:2109.07161, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 536, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 568 + ], + "type": "text", + "content": "[37] Kolars Team. Kolars: Effective training of diffusion model for photorealistic text-to-image synthesis. arXiv preprint, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "text", + "content": "[38] Yuxiang Tuo, Yifeng Geng, and Liefeng Bo. Anytext2: Visual text generation and editing with customizable attributes. arXiv preprint arXiv:2411.15245, 2024. 3, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 635 + ], + "type": "text", + "content": "[39] Yuxiang Tuo, Wangmeng Xiang, Jun-Yan He, Yifeng Geng, and Xuansong Xie. Anytext: Multilingual visual text generation and editing. *ICLR*, 2024. 2, 3, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 669 + ], + "type": "text", + "content": "[40] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation, 2024. 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[41] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, and Anthony Chen. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 363 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 95 + ], + "type": "text", + "content": "[42] Yibin Wang, Weizhong Zhang, Honghui Xu, and Cheng Jin. Dreamtext: High fidelity scene text synthesis, 2025. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[43] Yukang Yang, Dongnan Gui, Yuhui Yuan, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. arXiv preprint, abs/2305.18259, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[44] Yukang Yang, Dongnan Gui, Yuhui Yuan, Weicong Liang, Haisong Ding, Han Hu, and Kai Chen. Glyphcontrol: Glyph conditional control for visual text generation. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 218 + ], + "type": "text", + "content": "[45] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint, 2023. 1, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 252 + ], + "type": "text", + "content": "[46] Jianhao Zeng, Dan Song, Weizhi Nie, Hongshuo Tian, Tongtong Wang, and Anan Liu. Cat-dm: Controllable accelerated virtual try-on with diffusion model, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[47] Weichao Zeng, Yan Shu, Zhenhang Li, Dongbao Yang, and Yu Zhou. Textctrl: Diffusion-based scene text editing with prior guidance control, 2024. 2, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[48] Lvmin Zhang and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint, abs/2302.05543, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 321, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 363 + ], + "type": "text", + "content": "[49] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation, 2024. 6" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16121" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/6a7bfaa2-52e3-41ce-8b5a-5f025a33ee77_content_list.json b/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/6a7bfaa2-52e3-41ce-8b5a-5f025a33ee77_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3351caea1d781c0e5c2fb8728fc378a107460447 --- /dev/null +++ b/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/6a7bfaa2-52e3-41ce-8b5a-5f025a33ee77_content_list.json @@ -0,0 +1,1868 @@ +[ + { + "type": "text", + "text": "TextSSR: Diffusion-based Data Synthesis for Scene Text Recognition", + "text_level": 1, + "bbox": [ + 153, + 130, + 843, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xingsong Ye $^{1}$ , Yongkun Du $^{1}$ , Yunbo Tao $^{1}$ , Zhineng Chen $^{2*}$", + "bbox": [ + 264, + 179, + 736, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ College of Computer Science and Artificial Intelligence, Fudan University, China $^{2}$ Institute of Trustworthy Embodied AI, Fudan University, China", + "bbox": [ + 174, + 198, + 821, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xsye20，zhinchen}@fudan.edu.cn，{ykdu23，ybtao24} $\\text{已}$ m.fudan.edu.cn", + "bbox": [ + 205, + 234, + 777, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene text recognition (STR) suffers from challenges of either less realistic synthetic training data or the difficulty of collecting sufficient high-quality real-world data, limiting the effectiveness of trained models. Meanwhile, despite producing holistically appealing text images, diffusion-based visual text generation methods struggle to synthesize accurate and realistic instance-level text at scale. To tackle this, we introduce TextSSR: a novel pipeline for Synthesizing Scene Text Recognition training data. TextSSR targets three key synthesizing characteristics: accuracy, realism, and scalability. It achieves accuracy through a proposed region-centric text generation with position-glyph enhancement, ensuring proper character placement. It maintains realism by guiding style and appearance generation using contextual hints from surrounding text or background. This character-aware diffusion architecture enjoys precise character-level control and semantic coherence preservation, without relying on natural language prompts. Therefore, TextSSR supports large-scale generation through combinatorial text permutations. Based on these, we present TextSSR-F, a dataset of 3.55 million quality-screened text instances. Extensive experiments show that STR models trained on TextSSR-F outperform those trained on existing synthetic datasets by clear margins on common benchmarks, and further improvements are observed when mixed with real-world training data. Code is available at https://github.com/YesianRohn/TextSSR.", + "bbox": [ + 88, + 319, + 485, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 758, + 220, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene text images are a unique type of image captured in the wild, they differ from general images in the presence of text, which appears with diverse sizes, fonts, and often with varying degrees of distortion and occlusion. However, these texts usually contain high-level semantic information closely related to the captured scene, making STR a crucial", + "bbox": [ + 89, + 784, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ddccf2735e725e1dcbc87411ec22ecc338ae80305ad8baaa25a144930e5067c.jpg", + "image_caption": [ + "Figure 1. Top: Text instances detected by an end-to-end STR model [23], and synthesized by our TextSSR based on the detected examples. Bottom: Accuracy of MAERec [20] across common benchmarks. Peach indicates training on ST [15] and TextOCR [39], while blue denotes that TextSSR-F is additionally included as training data." + ], + "image_footnote": [], + "bbox": [ + 516, + 286, + 897, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "task for complex scene understanding, image search, etc.", + "bbox": [ + 511, + 566, + 888, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent STR practices [9, 10, 20] have shown that besides advanced modeling techniques, high-quality training data also plays a critical role. Training STR models also follows the scaling law from the data perspective [31] and current data size is far from saturated even for English. Enlarging the size of real-world datasets is a feasible way to further enhance recognition performance. However, it is challenging to collect real-world scene text images on an even larger scale. On one hand, diversified data collection is costly and time-consuming. On the other hand, real-world scenes predominantly contain high-frequency, semantically meaningful words, while low-frequency words are difficult to collect. Synthesizing high-quality text images seems to be an effective alternative. However, research on this direction is somewhat laggard. The two most popular synthetic datasets, MJ [18] and ST [15], are constructed based on traditional rendering techniques nearly a decade ago. Although they can be easily scaled up in quantity, recent studies [9, 11] indicate that there are significant accuracy gaps between models trained on MJ&SJ and Union14M-L [20], mainly because MJ&SJ fail to comprehensively describing", + "bbox": [ + 509, + 583, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17464", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the challenge of real-world text instances.", + "bbox": [ + 89, + 90, + 367, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, diffusion models have shown impressive performance in generating text-rich images [4, 5, 41, 50, 57]. However, most of them focus on generating aesthetic text images holistically, e.g., posters. Using data generated by these methods for STR model training encounters several issues. First, these methods are designed for the full image level generation, when switching to synthesizing text in instance-level regions, it would be less effective and ensuring the generation accuracy becomes challenging. Second, current generation methods mainly rely on well-refined natural language as prompts. They produce overly homogeneous and unrealistic text instances with some probability, which is not suitable as training data. Moreover, constructing a diverse set of prompts is inherently difficult, which hinders the ability to generate data at scale.", + "bbox": [ + 89, + 107, + 483, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, given the exceptional performance of diffusion models across various tasks, it is anticipated that there are diffusion-based techniques for producing satisfactory training data for STR. Before proceeding, we first outline three key characteristics that an improved data synthesis method should exhibit:", + "bbox": [ + 89, + 334, + 483, + 422 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Accurate: The synthesized text content should match the given text, ensuring that the synthesized words are readable and free from character misplacement or duplication errors. This is fundamental for maintaining the correctness of the trained STR models.", + "- Realistic: The generated text instances should visually resemble real-world text, simulating scene conditions in the wild. This includes being harmoniously blended into the background, ensuring consistent generation at different sizes and capturing the diversity of text presentation, especially in complex scenes. Achieving realism is essential for enhancing the performance of text recognizers.", + "- Scalable: The method should support extensive training data production. In other words, abandon complex control inputs and design processes, and enable large-scale data generation by leveraging easily accessible and processable resources only." + ], + "bbox": [ + 89, + 428, + 482, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To satisfy these characteristics, we propose TextSSR, a novel diffusion-driven pipeline focusing on Synthesizing training data for Scene Text Recognition. Our approach introduces the following innovations to overcome persistent limitations in text synthesis. First, we implement a region-centric text generation. It employs end-to-end OCR model to acquire text location and size. Moreover, we devise a position-glyph enhancement that further resolves character duplication and positional disorder, ensuring precise text placement and arrangement. Second, we leverage the detected text or only background near the designated region as contextual hints for font style and appearance generation, guiding the generated text to match real-world scenarios better. Third, we design a character-aware diffusion", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4a54ed79f77e42a1a50dc46f1b833ebbbbc2fb8662c7056c6883d837d83a4661.jpg", + "image_caption": [ + "Figure 2. Showcases of TextSSR's capacity to synthesize accurate, realistic and scalable text instances." + ], + "image_footnote": [], + "bbox": [ + 535, + 89, + 885, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "architecture that employs glyph priors rather than natural language prompts as injected conditions. It enables precise character-level control while maintaining full-word semantic coherence. To expand the generation quantity, we use detected text to generate a series of anagrams by rearranging characters within the original bounding boxes, thus supporting large-scale generation. As a result, TextSSR achieves accurate, realistic, and scalable text instance synthesize, as illustrated in Fig. 2.", + "bbox": [ + 511, + 321, + 906, + 457 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through these attempts, we construct the TextSSR-F dataset with 3.55 million quality-screened samples, which is automatically validated by checking their recognitions and pseudo labels. Extensive experiments demonstrate the superiority of TextSSR-F over existing synthesis data, as well as its complementarity with real-world training data.", + "bbox": [ + 511, + 458, + 906, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 531, + 547, + 834, + 561 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose TextSSR, a diffusion-based text synthesis method centered on seamlessly blending text instances into real scenes. It develops a region-centric processing that enables accurate instance-level generation.", + "- TextSSR fully exploits positional and glyph priors of text, achieving accurate, diverse and realistic text synthesis by using only the detected text as input. Moreover, it can be easily scaled up and we construct the TextSSR-F dataset with 3.55 million text instances.", + "- Extensive evaluations validate that STR models trained on TextSSR-F surpass their counterparts trained on existing synthetic data by $3.1\\%$ on common benchmarks, and even better STR models can be obtained when further combined with real-world training data, highlighting the utility of the proposed TextSSR." + ], + "bbox": [ + 513, + 566, + 903, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 808, + 653, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Scene Text Recognition", + "text_level": 1, + "bbox": [ + 513, + 833, + 728, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "STR researches can be categorized from both the model and the data perspectives. From a model point of view, early approaches employ the CTC-based pipeline [36, 52].", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17465", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "They used CNN for visual feature extraction, followed by RNN for sequence modeling, then with CTC loss applied as a constraint. Subsequently, attention mechanisms gradually gained attention in the STR field [24, 26, 35, 38]. Inspired by developments in the NLP field, many methods [2, 12, 13, 44, 46, 48, 58] further refined recognizers from a language modeling perspective. While some approaches [8, 11, 56] continued to explore solutions based on a single vision model and CTC decoding, due to their fast inference nature. Accuracy was steadily improved following these efforts.", + "bbox": [ + 89, + 90, + 480, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From a data perspective, the limited availability of well-annotated real-world training data led to a practical STR training and evaluation protocol: training solely on synthetic data, or a combination of synthetic and real data, followed by evaluation on real-world benchmarks. Besides the six common benchmarks [21, 22, 27, 29, 32, 43]. Union14M-L [20] aggregated publicly available real-world datasets (including TextOCR [39]) and created a real-world training dataset with nearly 4 million instances. However, recent studies on the OCR scaling law [31] suggested that even this substantial volume had not reached the saturation point. This observation implies that STR models can still benefit from additional training data, and high-quality synthetic data could be a feasible supplement, which motivates our study.", + "bbox": [ + 89, + 258, + 482, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Scene Text Synthesis", + "text_level": 1, + "bbox": [ + 89, + 500, + 285, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Early synthesis methods [15, 18, 25] employed rule-based methods to overlay text onto background with transformations simulating real-world appearance. To be more realistic, later research shifted toward deep learning-based generation methods. Among these, GAN-based scene text editing [30, 47, 49] methods aimed to expand data by editing text directly within specified regions, but they were limited by the lack of real paired training data. Recent advances had leveraged diffusion models for both editing [17, 19, 51] and synthesis, with fine-grained control over glyphs [3, 50], as well as positioning at word [41] and character [4, 45, 57] levels. The latest methods [5, 60] incorporate large language or multi-modal models to enhance performance. However, these approaches primarily targeted visually appealing generation for aesthetic purposes, rather than generating data for better STR models. While some work has explored diffusion-based text image synthesis. For example, CTIG-DM [59] employed character sequence as the condition for generation. However, it is less controllable and faced with the dilemma of diversity-realism tradeoff, limiting quality of the generated text. SceneVTG [60] developed a first-erasing-then-synthesizing pipeline. It requires a pre-erased model to remove existing text completely, which is difficult to guarantee in large-scale synthesizing scenarios.", + "bbox": [ + 89, + 523, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve accurate, realistic, and scalable text instance generation, TextSSR proposes a four-step pipeline (detailed in Fig. 3(b)). To handle text instances with varying sizes and locations, we first focus on text region processing, which will be discussed in Sec. 3.1. To enhance accuracy through fine-grained extraction and decoupling of text glyphs and positions, Sec. 3.2 and Sec. 3.3 will detail the specific module design and training strategy, respectively. Sec. 3.4 introduces our data extension method, achieving scalability in data synthesis at scale.", + "bbox": [ + 511, + 114, + 903, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Text Region-centric Processing", + "text_level": 1, + "bbox": [ + 511, + 273, + 785, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instance-level text generation presents significant challenges, as we must handle varying sizes, diverse potential positions, and determine how to select appropriate surrounding information for effective prompting. Synthesis methods that edit the entire image are not suitable for this task. Such approaches create an imbalance, and small text regions lack sufficient pixels for accurate reconstruction. Furthermore, information from the entire scene is often redundant for text region synthesis. For example, in Fig. 3(a), the text \"kills\" to be synthesized shares a similar style and context with the nearby upper text and background, while the region farther away is inconsistent in appearance.", + "bbox": [ + 511, + 296, + 903, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, we propose a generative scheme based on a locality assumption—text styles are more likely to correlate with adjacent text and background information. This method focuses generation at the region level, synthesizing the text using only surrounding local context rather than the entire scene. Specifically, we use the annotated text box to extract the smallest enclosing square around the text region and resize it to a fixed size (e.g., 256). For practical considerations, we uniformly employ an end-to-end OCR tool [23] to obtain text instance size, position, and textual labels. It well reflects real-world scenarios where annotations are often unavailable.", + "bbox": [ + 511, + 479, + 903, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Fine-grained Condition Mechanisms", + "text_level": 1, + "bbox": [ + 511, + 667, + 830, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual text rendering needs to consider two key fine-grained features: glyph structure and positional arrangement, both at word (holistic) and character (individual) levels. Word-level features concern entire word appearance and region placement, while the character-level address individual characters and their relative ordering. To guide the generative process, we implement the following approach:", + "bbox": [ + 511, + 688, + 903, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Word-level Position and Glyph Guidance. For a text $t$ with position $P_{t}$ in the original image, we extract a square region $I_{s}$ where $P_{t}$ becomes $P_{t}^{\\prime}$ (Fig. 3(a), red dashed rectangle). This allows us to establish our desired word-level position guidance. Then we create a binary mask $I_{M}$ that separates text (0) from non-text (1) regions. The size of $I_{M}$ is identical to $I_{s}$ , indicating where the text is present. For", + "bbox": [ + 511, + 795, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17466", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/25fcfe9f298a6c509b7f84f45e39131128ff186b82647f258cc5ba781efdb548.jpg", + "image_caption": [ + "(a) Generative Model" + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 521, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/03062fe75ce00b17cd5082df12f1839f15822ac1dee81edd643d2fe03f880ab9.jpg", + "image_caption": [ + "(b) Synthesis Pipeline", + "Figure 3. Overall architecture of TextSSR's (a) generative model and (b) synthesis pipeline. (a) Given a scene text instance (marked in red) and its content (e.g., \"kills\"), it is preprocessed into the smallest local image containing the text. A text-size mask is then applied to identify non-text regions, while the text content is rendered as priors at both word and character levels. These different conditions are subsequently injected into a diffusion model as training guidance. (b) An end-to-end OCR model extracts the text location and content within the image (Step 1). All possible text words are generated through anagram (Step 2). The generative model then synthesizes text instances in batches (Step 3), and finally, a SOTA STR model filters the samples, retaining only the correctly recognized instances for use (Step 4)." + ], + "image_footnote": [], + "bbox": [ + 522, + 88, + 883, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "glyph guidance, we render the text in a standard font, applying an affine transformation to align it with the position and orientation of $P_{t}^{\\prime}$ (resulting in $I_{G}$ ). As noted in previous work [3, 41, 50], when we close the door on the original styled text, we must at least provide a window with a basic visual style of the text at that location.", + "bbox": [ + 89, + 383, + 482, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Character-level Position and Glyph Guidance. Only word-level information limits the model's understanding to the words, as we observe issues such as repeated or incorrect characters when using only word-level features. We address this by encoding character-level ordering through pixel values, enabling the model to learn spatial relationships between characters. Specifically, we define a maximum character length $L$ . For characters from position 1 to $l$ ( $l \\leq L$ , with truncation if $l > L$ and zero-padding in channels $l + 1$ to $L$ if $l < L$ ), the $i$ -th character is rendered into the $i$ -th channel of an $L$ -channel image $I_g$ . Each character glyph is rendered with a pixel intensity of $i \\times \\left\\lfloor \\frac{P_M}{L} \\right\\rfloor$ , where $P_M$ denotes the maximum number of pixels allocated to distinguish a character from the background. We represent the glyph image tensor as $I_g \\in \\mathbb{R}^{L \\times S \\times S}$ , where $L$ is the number of characters and $S$ is the image size for each character. This creates position-prior pixel values that serve as positional encodings to help the model distinguish character information at each specific position.", + "bbox": [ + 89, + 474, + 482, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision Transformer (ViT) [7] is employed as the Glyph Encoder $G_{\\phi}$ to accommodate input $I_{g}$ and output features compatible with diffusion model's condition inputs:", + "bbox": [ + 89, + 763, + 482, + 810 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {g} = G _ {\\phi} \\left(I _ {g}\\right) \\in \\mathbb {R} ^ {(N + 1) \\times D} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 821, + 482, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $N$ represents the number of patches, the additional one corresponds to a class token, and dimension $D$ is set to match the input requirements of the diffusion process.", + "bbox": [ + 89, + 854, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adapting to Stable Diffusion", + "text_level": 1, + "bbox": [ + 511, + 383, + 769, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We leverage a pretrained model and fine-tune its Variational Auto Encoder (VAE) and retrain the Conditional Diffusion Model (CDM) components for our specific task.", + "bbox": [ + 511, + 405, + 906, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VAE Fine-tuning. The standard VAE, when pretrained on general datasets, may not be sensitive enough to capture the fine-grained spatial and glyph features of scaled text instances. As highlighted by DiffUTE [3] and also confirmed by our ablation experiments, fine-tuning the VAE is essential when working with text images. Therefore, we fine-tune it to enhance the capacity for accurate representation of text regions in a fair, local manner. The VAE consists of an encoder $E_{\\theta}$ and a decoder $D_{\\theta}$ , and we let $V_{\\theta}$ represent the reconstruction after the encoding and decoding process. Following the pre-defined $I_{s}$ , we minimize the following loss function:", + "bbox": [ + 511, + 450, + 906, + 632 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {T e x t S S R - V A E}} = \\left\\| V _ {\\theta} \\left(I _ {s}\\right) - I _ {s} \\right\\| _ {2} ^ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 643, + 906, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CDM Retraining. To mask the actual text instance region while obtaining surrounding contextual information, we apply $I_{M}$ to $I_{s}$ , resulting in $I_{m}$ , through the operation $I_{m} = I_{M} \\cdot I_{s}$ . $I_{m}$ , $I_{G}$ , and $I_{s}$ are fed into the frozen VAE, generating latent representations $Z_{m}$ , $Z_{G}$ , and $Z_{s}$ . Additionally, $Z_{s}$ is further processed with added noise to produce $Z_{T}$ . Moreover, we downsample $I_{M}$ to match their dimensions (e.g., in SD 2.1 [33], height = width = 16), transforming it into $Z_{M}$ .", + "bbox": [ + 511, + 672, + 906, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These latent features, $Z_{M}, Z_{T}, Z_{m}$ , and $Z_{G}$ , are concatenated and then passed through a Conv2D layer to match the channel dimension required by the U-Net (e.g., in SD 2.1, channel = 4). Meanwhile, the Glyph Encoder (see Fig. 3(a)) that produces $F_{g}$ is also integrated into the CDM to extract features from $I_{g}$ . Finally, the CDM processes", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17467", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/93767311cc114282c739edbee2381dd79b84dc47c7be62602fff452f54daf97b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIC13 [21]IC15 [22]ShopSign [54]
SeqAcc(%)↑NED ↑FID-R↓SeqAcc(%)↑NED ↑FID-R↓SeqAcc(%)↑NED ↑FID-R↓
TextDiffuser [4]25.080.4779103.740.820.0551120.660.640.0230136.56
GlyphControl [50]0.550.0472142.9100.0217152.3700.0067140.91
UDiffText [57]37.510.589267.822.550.1206103.43———
AnyText [41]55.070.715375.465.970.218676.0333.580.460289.25
TextDiffuser-2 [5]26.390.534268.310.870.083882.930.820.0411101.46
TextSSR Pre-training75.350.900751.7550.460.736134.6930.280.507554.24
TextSSR Fine-tuning85.170.940253.2862.640.808736.3052.240.705352.84
Real96.510.9895084.880.9486074.840.86990
", + "bbox": [ + 122, + 88, + 874, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparing TextSSR (differences between the pre-training and fine-tuning versions are detailed in Appendix 6.3.1. We use the fine-tuned version, abbreviated as TextSSR, in the rest part of the paper.) and existing synthesis methods on regular text (IC13), irregular text (IC15), and multi-language text (ShopSign) datasets, alongside results from real datasets (Real). SeqAcc and NED denote the recognition accuracy, editing distance, respectively. FID-R score means FID on the text region.", + "bbox": [ + 88, + 253, + 906, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "these inputs and produces the denoised latent representation $Z_{T-1}$ , progressively refining the noisy input to recover the clean latent space. Assuming $\\epsilon$ represents the original added noise and $\\epsilon_{\\theta}$ denotes the entire denoising process that outputs the predicted noise, the modified CDM is trained by minimizing the following objective:", + "bbox": [ + 89, + 335, + 480, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {T e x t S S R - C D M}} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, Z _ {M}, Z _ {m}, Z _ {G}, I _ {g}\\right) \\right\\| _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 433, + 480, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Anagram-based Data Synthesis", + "text_level": 1, + "bbox": [ + 89, + 455, + 369, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned above, our pipeline uses text region and content detected by an OCR engine as inputs. The recognized content is regarded as pseudo-label of the text. Our generation produces a synthesized text instance on the detected region for each pseudo-label rather than the groundtruth, which is typically unknown. Therefore, errors in OCR would not affect the generation. To expand the generation in quantity, we propose an anagram-based data expansion strategy. Unlike previous approaches that arbitrarily edit text [4, 5, 41] or substitute equal-length strings [45, 57], we manipulate internal character order to maintain better region-content coherence. Meanwhile, the diffusion process adaptively handles the varying occupation requirement of different characters, e.g., 'o' typically is wider than 'l' in width. Therefore, our TextSSR leads to a more flexible generation. By changing the internal character order only, the region remain unchanged while still generating many samples with different character arrangements.", + "bbox": [ + 88, + 478, + 482, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model's character-level rendering capability enables correct generation of text from this permutation-based operation. We define the total number of possible permutations for a word of length $l$ as:", + "bbox": [ + 88, + 750, + 482, + 810 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP (l) = l! \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 253, + 816, + 482, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that some words may repeat due to character duplication; they still represent different samples, as the diffusion model is inherently random and there are also variations in cropped regions.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 334, + 645, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. TextSSR Data Evaluation", + "text_level": 1, + "bbox": [ + 511, + 359, + 746, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first evaluate both accuracy and realism of the generation text, as well as the scaling capabilities of TextSSR data in small-scale. Then, we construct large-scale training data and assess its effectiveness at larger scales. Fig. 4 visualizes the examples of TextSSR compared to other methods.", + "bbox": [ + 511, + 383, + 905, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Accuracy Evaluation. Evaluating the accuracy of generated scene text is challenging, as manually verifying alignment with target text is costly. Previous benchmarks typically employ an STR model to automatically assess text consistency. We also follow this scheme. To mitigate potential OCR recognition errors and more effectively reflect the accuracy and difficulty level of our synthetic data, we choose SVTRv2 [11], a SOTA STR model upgraded from SVTR [8], as the recognition model.", + "bbox": [ + 511, + 460, + 905, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tab. 1 gives the accuracy assessment results across different datasets, where IC13 denotes all the methods using text instances from the IC13 dataset [21] to initialize their generation, while others are defined similarly. As can be seen, a large portion of text instances generated by TextSSR are correctly recognized. For more irregular and difficult IC15 [22], and less linguistically ShopSign [54] datasets, despite accuracy declines observed, TextSSR still outperforms the others by significant margins. The results provide strong evidence of TextSSR's correctness.", + "bbox": [ + 511, + 598, + 905, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparing with AnyText [41], the current SOTA, TextSSR outperforms it by $30.1\\%$ and $56.7\\%$ in accuracy on IC13 and IC15, respectively. This is mainly because TextSSR operates generation at the instance level, while AnyText performs on the full image level. Note that methods based on full-image synthesis struggle significantly in complex, challenging scenarios, e.g., IC15. This dataset includes many text instances affected by blur, low-resolution, perspective distortion, and curvature—challenges that previous methods do not address ef", + "bbox": [ + 511, + 750, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17468", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9121d23c28265570b1251692e433996e94cf69452c05f49b2667e846a80332c3.jpg", + "image_caption": [ + "GroundTruth" + ], + "image_footnote": [], + "bbox": [ + 174, + 108, + 261, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7974814931c525ae5a2445b1c8ee63bbca8dbc5d799b35bed606e7c7c31557d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 160, + 261, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/69a9d689acda4cc4721086811c89418badbf764ebf6733cc5ad5a2d5121ba20d.jpg", + "image_caption": [ + "TextDiffuser" + ], + "image_footnote": [], + "bbox": [ + 269, + 108, + 354, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d497c04336417c192bd0cdd89a6796db7ce379891128d5d4e7c663747b8906ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 159, + 354, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e7b4110a01d99aa29900f4def875d7263a84c6c79de7ad05fb8f6b9aeabfda2d.jpg", + "image_caption": [ + "GlyphControl" + ], + "image_footnote": [], + "bbox": [ + 366, + 108, + 447, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/87268eb510134ba42b6dcaa7e96227b170abb248b68d56107c20e92e48bd2e78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 160, + 447, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1d1e1e7e1d79e72a195f7daee2868271683afcaf59d08919e5ec588bfa1d0ded.jpg", + "image_caption": [ + "UDiffText" + ], + "image_footnote": [], + "bbox": [ + 455, + 108, + 542, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77d976b5bab074c34b5386516de6ea7d6b297b579223c2b55c4f056540b02189.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 160, + 540, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ae130331065af71b444b44b67bba559e4fc540bce5f600dd73531fb6c3e74d9.jpg", + "image_caption": [ + "AnyText" + ], + "image_footnote": [], + "bbox": [ + 550, + 108, + 633, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55be7c6ec5685408af1e65fbcbfb30b22a430af3b4faea505b091caacaac5dfe.jpg", + "image_caption": [ + "TextDiffuser-2" + ], + "image_footnote": [], + "bbox": [ + 643, + 107, + 728, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/979797ea2384fdc03db4db3f77e3c74885eb8fa32176bea947736c94dcac860c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 159, + 727, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1fc3ac4e4cf38c09fd51caae7558cc4fba1cd0acd14a0a17c4352a9e5fa31b3a.jpg", + "image_caption": [ + "TextSSR" + ], + "image_footnote": [], + "bbox": [ + 746, + 107, + 823, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/580c1b1f679a309b1ece3cf8109fa10ea0e85923768b88ccb94b0f00a76604c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 160, + 823, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/91f58b865faadcd8c362ebb9a92ea8be5fa2fa37e9b8a311e3fd4f79044be5db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 191, + 261, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1616cd0213d11274b0dc7f59b3f340e4da0fdc54389fb6c6aa74d827dd7a12ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 242, + 261, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/144d84cd9cb27db04d8f1d316b0e72a820473508a589fa5e81d389f1222a24d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 191, + 354, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4c766147a574570cd9a4321677ab085472fc40ee5cc8489c046782d0ab63dcfc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 271, + 242, + 354, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7f477be425f529b864d86c35a7594d04f3d408d8fe66e71ce68c212e2411a2a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 191, + 447, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/34c14200ca0bc58b012ab030e2ea3c0c0356c0d8bf1e844e4c637623356db553.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 242, + 447, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/946b835d48180427168e31c1f5114a2902a62049255d1c9460055f268acbc9b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 191, + 542, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bd50607e466fbfe520083c565dceca8ea45d00241460014852a3e8c0a9c69956.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 242, + 540, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0b1e3369e7d86773fb42e5d4986f67074bb0d18c4c62a82ee69091e436173878.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 191, + 633, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f58f7bd09506221ca649410e84569d3f3ab4b12591e70d016894e1d469df770.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 242, + 633, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f7bc5573365aa391984850322c4e22212ec6fa21a67a4e1a6709acac54648007.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 191, + 728, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ad719daf98b1b7a772cbff0b3a082c10d2e6c1b24dd0f7b094756e37e557db7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 242, + 728, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b49b7ec54917f988f627b93d7e5b13ecb6036b869bd3b7375ff9dfbb9e2b5af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 191, + 810, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b4dc2ee993be44a877141e4a331dbc0ec0394abf6ba70b5e121ce952f4f3c2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 242, + 823, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81bdd850e6f2dcd8d663805deb6aa990c15a7b54876f5363ab0a9db5fa9f46e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 174, + 273, + 261, + 323 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7216de48a3193cad2f787e266d0a040e66403d62fe7fa42851da536da85c990f.jpg", + "image_caption": [ + "Figure 4. Examples synthesized by TextSSR and other existing open-source methods (UDiffText only supports English text for its limited text encoder). TextSSR consistently produces accurate and realistic text instances. Please refer to Appendix 7.1 for more analysis." + ], + "image_footnote": [], + "bbox": [ + 176, + 325, + 259, + 358 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/618549cc2d5d54d058c21ba95453feebce23ca4cfad0b7b43d62f9786b4373f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 273, + 354, + 323 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/248400ce8d95bfcbb656f163cbd3d56f9c618b7acad871450e22dad576339579.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 269, + 325, + 354, + 358 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f0285cf616479a8cbc6999d5b6bd0a035b31100552df8dac25436a3a9898699.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 273, + 447, + 323 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c6b9bdf86c7677436dd02108319aa3bfb9bcef267e0dc2d02a1d7f75a96e13d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 325, + 447, + 358 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/293c00de0c3af1731ef69f3b389c1e49094861a6423249135df63d4e038f62d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 273, + 542, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4a83c1a140a826fa802a317aee4a31086c43c12b6e4642feb6c8fa9aaaafb585.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 273, + 633, + 323 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f539f2a66f6215d63a38f836a25a5b5a33abecb45e9363b6df9c7c63dd164dbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 549, + 325, + 633, + 358 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e8b971d361f01c594e4b038c4ab45d51393422c14216a3a428bc6c2205e6a962.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 273, + 728, + 323 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16ae66567977c09ffca176166f5fa7f89317e43299b35224b8dd41ac60acb7fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 325, + 727, + 358 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ffe4026785d57d371d7826e14228e9968885ea3ca54d62390f401cd2b7fd8899.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 745, + 273, + 810, + 323 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f51e3698a500ccc29c39d9e1a89aef5fb65b83e2dfc5b68c7c7bb889667f8a37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 325, + 821, + 358 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/05a11de2377eb5be31595c28eebfaf9fac4bd58b1ee1c9a2bf8c931e6f306dde.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TypeMethodSeqAcc(%)↑
RegularIrregularAvg
RenderingSynthText [15]48.5526.0839.74
VISD [53]38.0226.8833.65
UnrealText [25]32.0619.8527.28
DiffusionTextDiffuser [4]30.088.7921.74
GlyphControl [50]40.8013.2129.99
AnyText [41]49.8019.6237.97
TextDiffuser-2 [5]40.2610.4828.59
SceneVTG [60]54.9735.5047.34
TextSSR (Ours)59.7635.7150.33
RealReal-L [1]57.9737.6650.01
TextOCR [39]57.8341.5654.05
", + "bbox": [ + 99, + 421, + 475, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of different data synthetic methods, where CRNN [36] trained on different synthetic data and real data of ${30}\\mathrm{k}$ is employed to obtain the results. \"Regular\" denotes the average on IIIT [27], SVT [43], and IC13 [21] datasets, \"Irregular\" represents the average on IC15 [22], SVTP [29], and CUTE [32] datasets, while \"Avg\" is the average of the six datasets.", + "bbox": [ + 89, + 623, + 482, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fectively. In contrast, TextSSR benefits from a region-centered generation pipeline. It excels at handling challenging text instances and maintains reasonable synthesis accuracy. Moreover, the condition for TextSSR is glyph shapes, which can be extended to other languages like Chinese. The results on ShopSign prove TextSSR's multilingual capability. It outperforms the purportedly multilingual-focused AnyText by $18.7\\%$ . More details on accuracy evaluation are provided in Appendix 6.3.2.", + "bbox": [ + 89, + 734, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Realism Evaluation. Following the validation presented in SceneVTG [60], we generate a fixed amount (30k) of data", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to train an identical model under fair settings and then test it on common STR benchmarks [21, 22, 27, 29, 32, 43]. This approach avoids the need for separate verification of text correctness or synthesis quality. If the generated text is inaccurate or unrealistic, this will be reflected in poor recognition performance during testing.", + "bbox": [ + 511, + 425, + 906, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 2, for the compared methods, their average SeqAcc is below $40\\%$ except for SceneVTG [60]. The result is closely linked to their low generation accuracy. Note that rendering-based methods generally perform better than traditional diffusion-based methods in irregular benchmarks, highlighting the challenge of generating highly controllable text images. SceneVTG benefits from its two-stage pipeline of text erasure and then rendering, it reaches a SeqAcc of $47.3\\%$ . Nevertheless, TextSSR, featured by its glyph-based condition injection and anagram-based generation, surpasses SceneVTG by nearly $3\\%$ . Moreover, the gap between models trained on 30k TextSSR and real-world TextOCR is only $3.7\\%$ , which comes from irregular text. These results convincingly demonstrate the realism of TextSSR data, especially in regular text synthesizing. However, further efforts are still needed to improve the synthesis of irregular text.", + "bbox": [ + 511, + 518, + 908, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scalability Evaluation. TextSSR data can be easily scaled up to a large quantity. It is necessary to assess whether the increase in data volume can translate into performance improvements. Therefore, we use the model generated by the 30k data in the realism evaluation as baseline (TextSSR $\\times 1$ ), and re-train CRNN under the same experimental setup but with data doubled every time. Results of this scalability evaluation are presented in Tab. 4. When scaled up", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17469", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1842e14b3f4fb53be5d0c9068e92d8f2ca147ff5d671772f3b8e602cb52c899c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDatasetVolumeSeqAcc(%)↑
IIIT5k [27]SVT [43]IC13 [21]IC15 [22]SVTP [29]CUTE80 [32]AvgContextless [20]
CRNN [36]ST4m90.1382.8490.9072.4572.0982.9981.9033.50
ST6.98m90.8083.0091.8372.3473.8081.6082.0632.61
ST + SynthAdd4m + 1.05m90.4383.9391.1373.5572.7180.2181.9934.53
ST + TextSSR4m + 1.05m92.1387.0293.3577.4277.6784.3785.3355.97
ST + TextSSR4m + 3.55m93.5789.1894.2878.5878.4582.9986.1757.89
ST + TextOCR4m + 1.05m93.1388.1092.5379.6879.2283.6886.0644.93
ST + TextOCR + TextSSR5.05m + 3.55m94.5789.3494.4080.2379.2284.7287.0858.15
+1.44+1.24+1.87+0.55+0.00+1.04+1.02+13.22
MAERec [20]ST4m96.0091.9695.8083.8885.7489.5890.4959.43
ST6.98m96.9092.8995.8084.7086.8292.3691.5861.62
ST + SynthAdd4m + 1.05m95.9793.2095.8085.0486.9890.2891.2160.46
ST + TextSSR4m + 1.05m97.5393.8296.8586.7589.4694.1093.0876.77
ST + TextSSR4m + 3.55m98.1394.7497.3287.8589.6194.1093.6378.17
ST + TextOCR4m + 1.05m98.1095.5297.2089.3490.3996.8794.5774.33
ST + TextOCR + TextSSR5.05m + 3.55m98.2096.7597.2089.9592.7197.9295.4682.03
+0.10+1.23+0.00+0.61+2.32+1.05+0.89+7.70
", + "bbox": [ + 117, + 88, + 879, + 323 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1174999513ea79e912630ec76825a8086defe5fdb5271bb82d7a9d85daaf3115.jpg", + "table_caption": [ + "Table 3. Performance of models trained on different data combinations, where CRNN and MAERec are selected as model representatives. Bold and underline indicate the highest and second highest values, respectively." + ], + "table_footnote": [], + "table_body": "
MethodVolumeSeqAcc(%)↑
RegularIrregularAvg
TextSSR ×130k59.7635.7150.33
TextSSR EL30k24.5810.2318.95
TextSSR ×260k60.4934.4950.29
TextSSR ×4120k65.1739.3455.03
TextSSR ×8240k73.9649.6964.44
TextOCR30k57.8341.5654.05
", + "bbox": [ + 125, + 385, + 447, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. TextSSR scalability evaluation. $\\times n$ refers to perform the anagram-based rendering $n$ times. EL denotes equal length edit.", + "bbox": [ + 89, + 542, + 482, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to $\\times 4$ and $\\times 8$ , accuracy improvements of $4.7\\%$ and $14.1\\%$ are observed, respectively. The result indicates that increasing the data volume indeed contributes to better STR models. Note that models in the two scenarios also surpass the model trained on 30k real-world data by $1.0\\%$ and $10.4\\%$ , respectively. It is observed that TextSSR $\\times 8$ also presents a much better result than real data model in irregular text, highlighting that the challenge of irregularity can be overcome to some extent by increasing the data volume. Meanwhile, we also conducted a comparison experiment similar to [57], i.e., substituting text instances with equal-length strings (TextSSR EL), where the trained model reported quite worse results. This again demonstrates the effectiveness of our anagram-based generation pipeline.", + "bbox": [ + 88, + 598, + 482, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Usability Assessment. We further assess the effectiveness of our synthesis method in supporting the large-scale training of STR models. To ensure the quality of data, we apply a quality-screened step to every sample generated by TextSSR. We still use SVTRv2 [11] to recognize the sample and compare the recognized content with its genera", + "bbox": [ + 88, + 810, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion label. The sample is reserved if they are consistent, otherwise we discard it. The results on different training data combinations are shown in Tab. 3. The observations can be summarized as follows: First, the rendering-based ST dataset reaches a saturation point in terms of accuracy, where $6.98\\mathrm{m}$ samples only show marginal improvement $(0.16\\%)$ on CRNN over the $4\\mathrm{m}$ dataset on the six common benchmarks. This suggests that we have to move towards more realistic data synthesis. Second, when incorporating $1.05\\mathrm{m}$ our synthetic dataset, the two STR models (CRNN and MAERec) gain improvements of $3.4\\%$ and $2.6\\%$ over the $4\\mathrm{m}$ ST dataset on the six common benchmarks, respectively. They are only $0.7\\%$ and $1.5\\%$ lower than ST combined with equivalent size real-world TextOCR. This suggests that TextSSR data is approaching real-world data in quality. Third, when further increasing TextSSR data to $3.55\\mathrm{m}$ , it is seen that the purely synthetic training data $(\\mathrm{ST} + 3.55\\mathrm{m}$ TextSSR) gain performance more approaching the synthetic-real mixed training data $(\\mathrm{ST} + 1.05\\mathrm{m}$ TextOCR). Moreover, when combining the $3.55\\mathrm{m}$ synthetic data with this mixed data, the trained model can report further performance improvements, obtaining $1.02\\%$ on CRNN and $0.89\\%$ on MAERec. This implies that TextSSR data is a valuable supplement to existing training data. Using it individually or collectively both can benefit STR model training. As a result, we mark this $3.55\\mathrm{m}$ dataset as TextSSR-F, an accurate and realistic large-scale synthetic dataset constructed by our TextSSR, and make it publicly available.", + "bbox": [ + 511, + 388, + 906, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our anagram-based synthesis generates a large number of contextless words, and contextless word generation is also discussed in SynthAdd [24], we conduct a comparison between SynthAdd and our TextSSR. As shown", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17470", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fad2a90b44ba7c97c046f6faeb563aefb478236fbdd7c3c01aa790ebba78ffaa.jpg", + "table_caption": [ + "in Tab. 3, there are clear performance margins (3.3% on CRNN and 1.9% on MAERec) between the model trained on traditional SynthAdd and TextSSR. This also indicates the superiority of TextSSR as a STR training data synthetic method." + ], + "table_footnote": [], + "table_body": "
region-centricchar-glyphchar-positionfine-tune VAESeqAcc (%) ↑NED ↑FID-R ↓
20.280.3444102.89
✓73.500.876445.89
✓✓74.150.885246.65
✓✓✓80.260.913345.83
✓✓✓77.210.904350.34
✓✓76.230.899151.94
✓✓✓47.550.621162.52
✓✓✓✓82.990.930850.76
", + "bbox": [ + 106, + 176, + 468, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study of TextSSR components using the TextSSR pre-training model in Tab. 1 on the IC13 dataset.", + "bbox": [ + 89, + 321, + 482, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 377, + 243, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess the proposed components, we start with the TextSSR pre-training model in Tab. 1 (to simplify, trained for only 5k steps), where all components are considered. Then, we stepwisely remove one or some of the components to assess their respective necessity and effectiveness. From the results on Tab. 5, we can see that the removal of each one brings a clear decline in performance, while the absence of \"region-centric\" processing can lead to a catastrophic drop of $35.4\\%$ in accuracy. Additionally, fine-tuning the VAE offers a trade-off, slightly sacrificing visual quality for an improvement in accuracy. This is because real-world scene text also confronts poor visual quality challenge, simulating this to some extent can provide valuable samples for model training.", + "bbox": [ + 88, + 398, + 483, + 609 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9c36a1114f5b860501ab83aac2dbfc51a2844da35800dc8b4bd1174c36a386d3.jpg", + "image_caption": [ + "Figure 5. Performance of TextSSR fine-tuning model with and without region-centric processing across text regions of different sizes (in pixels)." + ], + "image_footnote": [], + "bbox": [ + 135, + 623, + 444, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform another experiment to verify the effectiveness of the region-centric processing. The results in Fig. 5", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "show that without this processing, TextSSR has to synthesize text instance on the whole image. It exhibits a performance trend that first rises and then falls as the size increases. When the region is small, the entire image is scaled severely, resulting in more blurred text regions and insufficient pixels for text rendering. As the region size increases, larger regions and more pixels are allowed, therefore performance increase. When a too large region given, it causes the side effect that constraints the shape of the generated text, and performance decline. In contrast, incorporating region-centric processing allows TextSSR to achieve superior and consistent performance across all sizes.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 6, we also illustratively ablate the character-level position and glyph components. The results indicate that omitting either component leads to issues such as character deformation, incorrect characters, and duplication errors.", + "bbox": [ + 511, + 272, + 903, + 333 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ecc09037c1d9dea1144cce0c4e95e7a29a009db736e71773833247a99d693bd8.jpg", + "image_caption": [ + "Figure 6. Visualizations of TextSSR with and without character-level prior. More examples are provided in Appendix 7.3" + ], + "image_footnote": [], + "bbox": [ + 537, + 344, + 883, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 573, + 633, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed TextSSR to provide STR model training with high-quality synthetic data. It employs the glyph of existing text and surrounding context as the prompt of the diffusion-based synthesizing model, and combines character-level constraints and permutations. TextSSR successfully synthesizes accurate and realistic text instances at a large scale. The TextSSR-F dataset with 3.55 million diverse and realistic instances are constructed. We have conducted extensive experiments to assess TextSSR-F. STR models trained solely on TextSSR-F, and on the combination of TextSSR-F and real-world data both show performance improvements. While the effectiveness of TextSSR has been basically confirmed, for irregular text, there is still an accuracy gap between models trained on TextSSR-F and real-world data. Thus, future work includes the exploration of more controllable ways to further improve the synthesizing quality, especially for challenging instances such as curved and multi-oriented text. We are also interested in extending TextSSR to generate large-scale data in different languages like Chinese.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17471", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by the National Natural Science Foundation of China (Nos. 32341012, 62172103).", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 150, + 187, + 166 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] J. Baek, Y. Matsui, and K. Aizawa. What if we only use real datasets for scene text recognition? toward scene text recognition with fewer labels. In CVPR, pages 3113-3122, 2021. 6, 2", + "[2] D. Bautista and R. Atienza. Scene text recognition with permuted autoregressive sequence models. In ECCV, pages 178-196, 2022. 3", + "[3] H. Chen, Z. Xu, Z. Gu, J. Lan, X. Zheng, Y. Li, C. Meng, H. Zhu, and W. Wang. DiffUTE: Universal text editing diffusion model. In NeurIPS, pages 63062-63074, 2023. 3, 4", + "[4] J. Chen, Y. Huang, T. Lv, L. Cui, Q. Chen, and F. Wei. TextDiffuser: Diffusion models as text painters. In NeurIPS, pages 9353-9387, 2023. 2, 3, 5, 6", + "[5] J. Chen, Y. Huang, T. Lv, L. Cui, Q. Chen, and F. Wei. TextDiffuser-2: Unleashing the power of language models for text rendering. In ECCV, pages 386–402, 2024. 2, 3, 5, 6", + "[6] C.K. Chng, Y. Liu, Y. Sun, C.C. Ng, C. Luo, Z. Ni, C.M. Fang, S. Zhang, J. Han, E. Ding, J. Liu, D. Karatzas, C.S. Chan, and L. Jin. ICDAR2019 robust reading challenge on arbitrary-shaped text-rc-art. In ICDAR, pages 1571–1576, 2019. 1", + "[7] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4", + "[8] Y. Du, Z. Chen, C. Jia, X. Yin, T. Zheng, C. Li, Y. Du, and Y. Jiang. SVTR: Scene text recognition with a single visual model. In IJCAI, pages 884-890, 2022. 3, 5", + "[9] Y. Du, Z. Chen, C. Jia, X. Yin, C. Li, Y. Du, and Y.-G. Jiang. Context perception parallel decoder for scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell., 47(6):4668-4683, 2025. 1", + "[10] Y. Du, Z. Chen, Y. Su, C. Jia, and Y.-G. Jiang. Instruction-guided scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell., 47(4):2723-2738, 2025. 1", + "[11] Y. Du, Z. Chen, H. Xie, C. Jia, and Y. Jiang. SVTRv2: CTC beats encoder-decoder models in scene text recognition. In ICCV, 2025. 1, 3, 5, 7, 2", + "[12] S. Fang, H. Xie, Y. Wang, Z. Mao, and Y. Zhang. Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition. In CVPR, pages 7098-7107, 2021. 3", + "[13] S. Fang, Z. Mao, H. Xie, Y. Wang, C. Yan, and Y. Zhang. Abinet++: Autonomous, bidirectional and iterative language modeling for scene text spotting. IEEE Trans. Pattern Anal. Mach. Intell., 45(6):7123-7141, 2022. 3", + "[14] J. Gu, X. Meng, G. Lu, L. Hou, N. Minzhe, X. Liang, L. Yao, R. Huang, W. Zhang, X. Jiang, C. Xu, and H. Xu. Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. In NeurIPS, pages 26418-26431, 2022. 1" + ], + "bbox": [ + 93, + 175, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] A. Gupta, A. Vedaldi, and A. Zisserman. Synthetic data for text localisation in natural images. In CVPR, pages 2315-2324, 2016. 1, 3, 6", + "[16] M. He, Y. Liu, Z. Yang, S. Zhang, C. Luo, F. Gao, Q. Zheng, Y. Wang, X. Zhang, and L. Jin. ICPR2018 contest on robust reading for multi-type web images. In ICPR, pages 7-12, 2018. 1", + "[17] Santoso J, Simon C, and Williem. On manipulating scene text in the wild with diffusion models. In WACV, pages 5202-5211, 2024. 3", + "[18] M. Jaderberg, K. Simonyan, A. Vedaldi, and A. Zisserman. Synthetic data and artificial neural networks for natural scene text recognition. In Workshop on Deep Learning, NeurIPS, 2014. 1, 3", + "[19] J. Ji, G. Zhang, Z. Wang, B. Hou, Z. Zhang, B. Price, and S. Chang. Improving diffusion models for scene text editing with dual encoders. TMLR, 2024. 3", + "[20] Q. Jiang, J. Wang, D. Peng, C. Liu, and L. Jin. Revisiting scene text recognition: A data perspective. In ICCV, pages 20543-20554, 2023. 1, 3, 7, 2", + "[21] D. Karatzas, F. Shafait, S. Uchida, M. Iwamura, L. Gomez i Bigorda, S. Robles Mestre, J. Mas, D. Fernandez Mota, J. Almazán, and L.P. de las Heras. ICDAR 2013 robust reading competition. In ICDAR, pages 1484-1493, 2013. 3, 5, 6, 7, 1", + "[22] D. Karatzas, L. Gomez-Bigorda, A. Nicolaou, S. Ghosh, A. Bagdanov, M. Iwamura, J. Matas, L. Neumann, V.R. Chandrasekhar, S. Lu, F. Shafait, S. Uchida, and E. Valveny. IC-DAR 2015 competition on robust reading. In ICDAR, pages 1156-1160, 2015. 3, 5, 6, 7, 1", + "[23] C. Li, W. Liu, R. Guo, X. Yin, K. Jiang, Y. Du, Y. Du, L. Zhu, B. Lai, X. Hu, D. Yu, and Y. Ma. PP-OCRv3: More attempts for the improvement of ultra lightweightOCR system. arXiv preprint arXiv:2206.03001, 2022. 1, 3, 2", + "[24] H. Li, P. Wang, C. Shen, and G. Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In AAAI, pages 8610-8617, 2019. 3, 7", + "[25] S. Long and C. Yao. Unrealtext: Synthesizing realistic scene text images from the unreal world. In CVPR, 2020. 3, 6", + "[26] C. Luo, L. Jin, and Z. Sun. Moran: A multi-object rectified attention network for scene text recognition. Pattern Recognit., 90:109-118, 2019. 3", + "[27] A. Mishra, K. Alahari, and CV Jawahar. Scene text recognition using higher order language priors. In BMVC, 2012. 3, 6, 7", + "[28] N. Nayef, C. Liu, J. Ogier, Y. Patel, M. Busta, P.N. Chowdhury, D. Karatzas, W. Khlif, J. Matas, U. Pal, and J. Burie. ICDAR2019 robust reading challenge on multi-lingual scene text detection and recognition-rrc-plt-2019. In ICDAR, pages 1582-1587, 2019. 1", + "[29] T. Q. Phan, P. Shivakumara, S. Tian, and C. L. Tan. Recognizing text with perspective distortion in natural scenes. In ICCV, pages 569-576, 2013. 3, 6, 7", + "[30] Y. Qu, Q. Tan, H. Xie, J. Xu, Y. Wang, and Y. Zhang. Exploring stroke-level modifications for scene text editing. In AAAI, pages 2119-2127, 2023. 3" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17472", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] M. Rang, Z. Bi, C. Liu, Y. Wang, and K. Han. An empirical study of scaling law for scene text recognition. In CVPR, pages 15619-15629, 2024. 1, 3", + "[32] A. Risnumawan, P. Shivakumara, C. S. Chan, and C. L. Tan. A robust arbitrary text detection system for natural scene images. Expert Syst. Appl., 41(18):8027-8048, 2014. 3, 6, 7", + "[33] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 4, 2", + "[34] C. Schuhmann, R. Vencu, R. Beaumont, R. Kaczmarczyk, C. Mullis, A. Katta, T. Coombes, J. Jitsev, and A. Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1", + "[35] F. Sheng, Z. Chen, and B. Xu. NRTR: A no-recurrence sequence-to-sequence model for scene text recognition. In ICDAR, pages 781-786, 2019. 3", + "[36] B. Shi, X. Bai, and C. Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell., 39(11):2298-2304, 2016. 2, 6, 7", + "[37] B. Shi, C. Yao, M. Liao, M. Yang, P. Xu, L. Cui, S. Belongie, S. Lu, and X. Bai. ICDAR2017 competition on reading Chinese text in the wild (rctw-17). In ICDAR, pages 1429–1434, 2017. 1", + "[38] B. Shi, M. Yang, X. Wang, P. Lyu, C. Yao, and X. Bai. ASTER: An attentional scene text recognizer with flexible rectification. IEEE Trans. Pattern Anal. Mach. Intell., 41(9): 2035-2048, 2018. 3", + "[39] A. Singh, G. Pang, M. Toh, J. Huang, W. Galuba, and T. Hassner. TextOCR: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. In CVPR, pages 8802-8812, 2021. 1, 3, 6, 2", + "[40] Y. Sun, Z. Ni, C.K. Chng, Y. Liu, C. Luo, C.C. Ng, J. Han, E. Ding, J. Liu, D. Karatzas, C.S. Chan, and L. Jin. ICDAR 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In ICDAR, pages 1557-1562, 2019. 1", + "[41] Y. Tuo, W. Xiang, J. Y. He, Y. Geng, and X. Xie. AnyText: Multilingual visual text generation and editing. In ICLR, 2024. 2, 3, 4, 5, 6, 1", + "[42] A. Veit, T. Matera, L. Neumann, J. Matas, and S. Belongie. COCO-Text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 1", + "[43] K. Wang, B. Babenko, and S. Belongie. End-to-end scene text recognition. In ICCV, pages 1457-1464, 2011. 3, 6, 7", + "[44] Y. Wang, H. Xie, S. Fang, M. Xing, J. Wang, S. Zhu, and Y. Zhang. Petr: Rethinking the capability of transformer-based language model in scene text recognition. IEEE Trans. Image Process., 31:5585-5598, 2022. 3", + "[45] Y. Wang, W. Zhang, H. Xu, and C. Jin. Dreamtext: High fidelity scene text synthesis. In CVPR, pages 28555-28563, 2025. 3, 5", + "[46] J. Wei, H. Zhan, Y. Lu, X. Tu, B. Yin, C. Liu, and U. Pal. Image as a language: Revisiting scene text recognition via balanced, unified and synchronized vision-language reasoning network. In AAAI, pages 5885-5893, 2024. 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] L. Wu, C. Zhang, J. Liu, J. Han, J. Liu, E. Ding, and X. Bai. Editing text in the wild. In ACM MM, pages 1500-1508, 2019. 3", + "[48] J. Xu, Y. Wang, H. Xie, and Y. Zhang. OTE: Exploring accurate scene text recognition using one token. In CVPR, pages 28327-28336, 2024. 3", + "[49] Q. Yang, J. Huang, and W. Lin. SwapText: Image based texts transfer in scenes. In CVPR, pages 14700-14709, 2020. 3", + "[50] Y. Yang, D. Gui, Y. Yuan, W. Liang, H. Ding, H. Hu, and K. Chen. GlyphControl: Glyph conditional control for visual text generation. In NeurIPS, pages 44050-44066, 2023. 2, 3, 4, 5, 6", + "[51] W. Zeng, Y. Shu, Z. Li, D. Yang, and Y. Zhou. TextCtrl: Diffusion-based scene text editing with prior guidance control. In NeurlPS, pages 138569–138594, 2024. 3", + "[52] C. Zhai, Z. Chen, J. Li, and B. Xu. Chinese image text recognition with blastm-ctc: a segmentation-free method. In CCPR, pages 525–536, 2016. 2", + "[53] F. Zhan, S. Lu, and C. Xue. Verisimilar image synthesis for accurate detection and recognition of texts in scenes. In ECCV, pages 249-266, 2018. 6", + "[54] C. Zhang, W. Ding, G. Peng, F. Fu, and W. Wang. Street view text recognition with deep learning for urban scene understanding in intelligent transportation systems. IEEE Trans. Intell. Transp., 22(7):4727-4743, 2021. 5, 1", + "[55] R. Zhang, Y. Zhou, Q. Jiang, Q. Song, N. Li, K. Zhou, L. Wang, D. Wang, M. Liao, M. Yang, X. Bai, B. Shi, D. Karatzas, S. Lu, and C.V. Jawahar. ICDAR 2019 robust reading challenge on reading chinese text on signboard. In IC-DAR, pages 1577-1581, 2019. 1", + "[56] Z. Zhang, N. Lu, M. Liao, Y. Huang, C. Li, M. Wang, and W. Peng. Self-distillation regularized connectionist temporal classification loss for text recognition: A simple yet effective approach. In AAAI, pages 7441-7449, 2024. 3", + "[57] Y. Zhao and Z. Lian. UDiffText: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In ECCV, pages 217-233, 2024. 2, 3, 5, 7", + "[58] T. Zheng, Z. Chen, S. Fang, H. Xie, and Y.-G. Jiang. CDistNet: Perceiving multi-domain character distance for robust text recognition. Int. J. Comput. Vis., 132(2):300-318, 2024. 3", + "[59] Y. Zhu, Z. Li, T. Wang, M. He, and C. Yao. Conditional text image generation with diffusion models. In CVPR, pages 14235-14245, 2023. 3", + "[60] Y. Zhu, J. Liu, F. Gao, W. Liu, X. Wang, P. Wang, F. Huang, C. Yao, and Z. Yang. Visual text generation in the wild. In ECCV, pages 89-106, 2024. 3, 6" + ], + "bbox": [ + 514, + 90, + 903, + 772 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17473", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/6a7bfaa2-52e3-41ce-8b5a-5f025a33ee77_model.json b/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/6a7bfaa2-52e3-41ce-8b5a-5f025a33ee77_model.json new file mode 100644 index 0000000000000000000000000000000000000000..eb2a965e4d7de2b5a95fd8bda0d2bc9cc9ee1508 --- /dev/null +++ b/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/6a7bfaa2-52e3-41ce-8b5a-5f025a33ee77_model.json @@ -0,0 +1,2563 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.155, + 0.131, + 0.844, + 0.154 + ], + "angle": 0, + "content": "TextSSR: Diffusion-based Data Synthesis for Scene Text Recognition" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.18, + 0.737, + 0.199 + ], + "angle": 0, + "content": "Xingsong Ye\\(^{1}\\), Yongkun Du\\(^{1}\\), Yunbo Tao\\(^{1}\\), Zhineng Chen\\(^{2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.199, + 0.823, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)College of Computer Science and Artificial Intelligence, Fudan University, China \\(^{2}\\)Institute of Trustworthy Embodied AI, Fudan University, China" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.236, + 0.778, + 0.252 + ], + "angle": 0, + "content": "{xsye20，zhinchen}@fudan.edu.cn，{ykdu23，ybtao24} \\(\\text{已}\\) m.fudan.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.32, + 0.486, + 0.729 + ], + "angle": 0, + "content": "Scene text recognition (STR) suffers from challenges of either less realistic synthetic training data or the difficulty of collecting sufficient high-quality real-world data, limiting the effectiveness of trained models. Meanwhile, despite producing holistically appealing text images, diffusion-based visual text generation methods struggle to synthesize accurate and realistic instance-level text at scale. To tackle this, we introduce TextSSR: a novel pipeline for Synthesizing Scene Text Recognition training data. TextSSR targets three key synthesizing characteristics: accuracy, realism, and scalability. It achieves accuracy through a proposed region-centric text generation with position-glyph enhancement, ensuring proper character placement. It maintains realism by guiding style and appearance generation using contextual hints from surrounding text or background. This character-aware diffusion architecture enjoys precise character-level control and semantic coherence preservation, without relying on natural language prompts. Therefore, TextSSR supports large-scale generation through combinatorial text permutations. Based on these, we present TextSSR-F, a dataset of 3.55 million quality-screened text instances. Extensive experiments show that STR models trained on TextSSR-F outperform those trained on existing synthetic datasets by clear margins on common benchmarks, and further improvements are observed when mixed with real-world training data. Code is available at https://github.com/YesianRohn/TextSSR." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.759, + 0.222, + 0.774 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.785, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Scene text images are a unique type of image captured in the wild, they differ from general images in the presence of text, which appears with diverse sizes, fonts, and often with varying degrees of distortion and occlusion. However, these texts usually contain high-level semantic information closely related to the captured scene, making STR a crucial" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.287, + 0.898, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.454, + 0.908, + 0.538 + ], + "angle": 0, + "content": "Figure 1. Top: Text instances detected by an end-to-end STR model [23], and synthesized by our TextSSR based on the detected examples. Bottom: Accuracy of MAERec [20] across common benchmarks. Peach indicates training on ST [15] and TextOCR [39], while blue denotes that TextSSR-F is additionally included as training data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.889, + 0.583 + ], + "angle": 0, + "content": "task for complex scene understanding, image search, etc." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.584, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Recent STR practices [9, 10, 20] have shown that besides advanced modeling techniques, high-quality training data also plays a critical role. Training STR models also follows the scaling law from the data perspective [31] and current data size is far from saturated even for English. Enlarging the size of real-world datasets is a feasible way to further enhance recognition performance. However, it is challenging to collect real-world scene text images on an even larger scale. On one hand, diversified data collection is costly and time-consuming. On the other hand, real-world scenes predominantly contain high-frequency, semantically meaningful words, while low-frequency words are difficult to collect. Synthesizing high-quality text images seems to be an effective alternative. However, research on this direction is somewhat laggard. The two most popular synthetic datasets, MJ [18] and ST [15], are constructed based on traditional rendering techniques nearly a decade ago. Although they can be easily scaled up in quantity, recent studies [9, 11] indicate that there are significant accuracy gaps between models trained on MJ&SJ and Union14M-L [20], mainly because MJ&SJ fail to comprehensively describing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17464" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.368, + 0.106 + ], + "angle": 0, + "content": "the challenge of real-world text instances." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.484, + 0.333 + ], + "angle": 0, + "content": "Recently, diffusion models have shown impressive performance in generating text-rich images [4, 5, 41, 50, 57]. However, most of them focus on generating aesthetic text images holistically, e.g., posters. Using data generated by these methods for STR model training encounters several issues. First, these methods are designed for the full image level generation, when switching to synthesizing text in instance-level regions, it would be less effective and ensuring the generation accuracy becomes challenging. Second, current generation methods mainly rely on well-refined natural language as prompts. They produce overly homogeneous and unrealistic text instances with some probability, which is not suitable as training data. Moreover, constructing a diverse set of prompts is inherently difficult, which hinders the ability to generate data at scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.424 + ], + "angle": 0, + "content": "Nevertheless, given the exceptional performance of diffusion models across various tasks, it is anticipated that there are diffusion-based techniques for producing satisfactory training data for STR. Before proceeding, we first outline three key characteristics that an improved data synthesis method should exhibit:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.429, + 0.483, + 0.503 + ], + "angle": 0, + "content": "- Accurate: The synthesized text content should match the given text, ensuring that the synthesized words are readable and free from character misplacement or duplication errors. This is fundamental for maintaining the correctness of the trained STR models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.505, + 0.483, + 0.61 + ], + "angle": 0, + "content": "- Realistic: The generated text instances should visually resemble real-world text, simulating scene conditions in the wild. This includes being harmoniously blended into the background, ensuring consistent generation at different sizes and capturing the diversity of text presentation, especially in complex scenes. Achieving realism is essential for enhancing the performance of text recognizers." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.611, + 0.483, + 0.686 + ], + "angle": 0, + "content": "- Scalable: The method should support extensive training data production. In other words, abandon complex control inputs and design processes, and enable large-scale data generation by leveraging easily accessible and processable resources only." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.429, + 0.483, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To satisfy these characteristics, we propose TextSSR, a novel diffusion-driven pipeline focusing on Synthesizing training data for Scene Text Recognition. Our approach introduces the following innovations to overcome persistent limitations in text synthesis. First, we implement a region-centric text generation. It employs end-to-end OCR model to acquire text location and size. Moreover, we devise a position-glyph enhancement that further resolves character duplication and positional disorder, ensuring precise text placement and arrangement. Second, we leverage the detected text or only background near the designated region as contextual hints for font style and appearance generation, guiding the generated text to match real-world scenarios better. Third, we design a character-aware diffusion" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.09, + 0.887, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.267, + 0.905, + 0.295 + ], + "angle": 0, + "content": "Figure 2. Showcases of TextSSR's capacity to synthesize accurate, realistic and scalable text instances." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.458 + ], + "angle": 0, + "content": "architecture that employs glyph priors rather than natural language prompts as injected conditions. It enables precise character-level control while maintaining full-word semantic coherence. To expand the generation quantity, we use detected text to generate a series of anagrams by rearranging characters within the original bounding boxes, thus supporting large-scale generation. As a result, TextSSR achieves accurate, realistic, and scalable text instance synthesize, as illustrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.459, + 0.907, + 0.549 + ], + "angle": 0, + "content": "Through these attempts, we construct the TextSSR-F dataset with 3.55 million quality-screened samples, which is automatically validated by checking their recognitions and pseudo labels. Extensive experiments demonstrate the superiority of TextSSR-F over existing synthesis data, as well as its complementarity with real-world training data." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.549, + 0.836, + 0.563 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.567, + 0.905, + 0.627 + ], + "angle": 0, + "content": "- We propose TextSSR, a diffusion-based text synthesis method centered on seamlessly blending text instances into real scenes. It develops a region-centric processing that enables accurate instance-level generation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.628, + 0.905, + 0.701 + ], + "angle": 0, + "content": "- TextSSR fully exploits positional and glyph priors of text, achieving accurate, diverse and realistic text synthesis by using only the detected text as input. Moreover, it can be easily scaled up and we construct the TextSSR-F dataset with 3.55 million text instances." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.702, + 0.905, + 0.793 + ], + "angle": 0, + "content": "- Extensive evaluations validate that STR models trained on TextSSR-F surpass their counterparts trained on existing synthetic data by \\(3.1\\%\\) on common benchmarks, and even better STR models can be obtained when further combined with real-world training data, highlighting the utility of the proposed TextSSR." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.567, + 0.905, + 0.793 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.809, + 0.655, + 0.825 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.834, + 0.729, + 0.85 + ], + "angle": 0, + "content": "2.1. Scene Text Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "STR researches can be categorized from both the model and the data perspectives. From a model point of view, early approaches employ the CTC-based pipeline [36, 52]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17465" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.257 + ], + "angle": 0, + "content": "They used CNN for visual feature extraction, followed by RNN for sequence modeling, then with CTC loss applied as a constraint. Subsequently, attention mechanisms gradually gained attention in the STR field [24, 26, 35, 38]. Inspired by developments in the NLP field, many methods [2, 12, 13, 44, 46, 48, 58] further refined recognizers from a language modeling perspective. While some approaches [8, 11, 56] continued to explore solutions based on a single vision model and CTC decoding, due to their fast inference nature. Accuracy was steadily improved following these efforts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.486 + ], + "angle": 0, + "content": "From a data perspective, the limited availability of well-annotated real-world training data led to a practical STR training and evaluation protocol: training solely on synthetic data, or a combination of synthetic and real data, followed by evaluation on real-world benchmarks. Besides the six common benchmarks [21, 22, 27, 29, 32, 43]. Union14M-L [20] aggregated publicly available real-world datasets (including TextOCR [39]) and created a real-world training dataset with nearly 4 million instances. However, recent studies on the OCR scaling law [31] suggested that even this substantial volume had not reached the saturation point. This observation implies that STR models can still benefit from additional training data, and high-quality synthetic data could be a feasible supplement, which motivates our study." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.5, + 0.286, + 0.515 + ], + "angle": 0, + "content": "2.2. Scene Text Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Early synthesis methods [15, 18, 25] employed rule-based methods to overlay text onto background with transformations simulating real-world appearance. To be more realistic, later research shifted toward deep learning-based generation methods. Among these, GAN-based scene text editing [30, 47, 49] methods aimed to expand data by editing text directly within specified regions, but they were limited by the lack of real paired training data. Recent advances had leveraged diffusion models for both editing [17, 19, 51] and synthesis, with fine-grained control over glyphs [3, 50], as well as positioning at word [41] and character [4, 45, 57] levels. The latest methods [5, 60] incorporate large language or multi-modal models to enhance performance. However, these approaches primarily targeted visually appealing generation for aesthetic purposes, rather than generating data for better STR models. While some work has explored diffusion-based text image synthesis. For example, CTIG-DM [59] employed character sequence as the condition for generation. However, it is less controllable and faced with the dilemma of diversity-realism tradeoff, limiting quality of the generated text. SceneVTG [60] developed a first-erasing-then-synthesizing pipeline. It requires a pre-erased model to remove existing text completely, which is difficult to guarantee in large-scale synthesizing scenarios." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.605, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.905, + 0.266 + ], + "angle": 0, + "content": "To achieve accurate, realistic, and scalable text instance generation, TextSSR proposes a four-step pipeline (detailed in Fig. 3(b)). To handle text instances with varying sizes and locations, we first focus on text region processing, which will be discussed in Sec. 3.1. To enhance accuracy through fine-grained extraction and decoupling of text glyphs and positions, Sec. 3.2 and Sec. 3.3 will detail the specific module design and training strategy, respectively. Sec. 3.4 introduces our data extension method, achieving scalability in data synthesis at scale." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.275, + 0.787, + 0.291 + ], + "angle": 0, + "content": "3.1. Text Region-centric Processing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.905, + 0.478 + ], + "angle": 0, + "content": "Instance-level text generation presents significant challenges, as we must handle varying sizes, diverse potential positions, and determine how to select appropriate surrounding information for effective prompting. Synthesis methods that edit the entire image are not suitable for this task. Such approaches create an imbalance, and small text regions lack sufficient pixels for accurate reconstruction. Furthermore, information from the entire scene is often redundant for text region synthesis. For example, in Fig. 3(a), the text \"kills\" to be synthesized shares a similar style and context with the nearby upper text and background, while the region farther away is inconsistent in appearance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.905, + 0.658 + ], + "angle": 0, + "content": "Therefore, we propose a generative scheme based on a locality assumption—text styles are more likely to correlate with adjacent text and background information. This method focuses generation at the region level, synthesizing the text using only surrounding local context rather than the entire scene. Specifically, we use the annotated text box to extract the smallest enclosing square around the text region and resize it to a fixed size (e.g., 256). For practical considerations, we uniformly employ an end-to-end OCR tool [23] to obtain text instance size, position, and textual labels. It well reflects real-world scenarios where annotations are often unavailable." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.831, + 0.684 + ], + "angle": 0, + "content": "3.2. Fine-grained Condition Mechanisms" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.795 + ], + "angle": 0, + "content": "Visual text rendering needs to consider two key fine-grained features: glyph structure and positional arrangement, both at word (holistic) and character (individual) levels. Word-level features concern entire word appearance and region placement, while the character-level address individual characters and their relative ordering. To guide the generative process, we implement the following approach:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Word-level Position and Glyph Guidance. For a text \\( t \\) with position \\( P_{t} \\) in the original image, we extract a square region \\( I_{s} \\) where \\( P_{t} \\) becomes \\( P_{t}^{\\prime} \\) (Fig. 3(a), red dashed rectangle). This allows us to establish our desired word-level position guidance. Then we create a binary mask \\( I_{M} \\) that separates text (0) from non-text (1) regions. The size of \\( I_{M} \\) is identical to \\( I_{s} \\), indicating where the text is present. For" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17466" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.522, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.246, + 0.365, + 0.257 + ], + "angle": 0, + "content": "(a) Generative Model" + }, + { + "type": "image", + "bbox": [ + 0.524, + 0.089, + 0.885, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.246, + 0.753, + 0.258 + ], + "angle": 0, + "content": "(b) Synthesis Pipeline" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.274, + 0.907, + 0.359 + ], + "angle": 0, + "content": "Figure 3. Overall architecture of TextSSR's (a) generative model and (b) synthesis pipeline. (a) Given a scene text instance (marked in red) and its content (e.g., \"kills\"), it is preprocessed into the smallest local image containing the text. A text-size mask is then applied to identify non-text regions, while the text content is rendered as priors at both word and character levels. These different conditions are subsequently injected into a diffusion model as training guidance. (b) An end-to-end OCR model extracts the text location and content within the image (Step 1). All possible text words are generated through anagram (Step 2). The generative model then synthesizes text instances in batches (Step 3), and finally, a SOTA STR model filters the samples, retaining only the correctly recognized instances for use (Step 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.475 + ], + "angle": 0, + "content": "glyph guidance, we render the text in a standard font, applying an affine transformation to align it with the position and orientation of \\( P_{t}^{\\prime} \\) (resulting in \\( I_{G} \\)). As noted in previous work [3, 41, 50], when we close the door on the original styled text, we must at least provide a window with a basic visual style of the text at that location." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Character-level Position and Glyph Guidance. Only word-level information limits the model's understanding to the words, as we observe issues such as repeated or incorrect characters when using only word-level features. We address this by encoding character-level ordering through pixel values, enabling the model to learn spatial relationships between characters. Specifically, we define a maximum character length \\( L \\). For characters from position 1 to \\( l \\) (\\( l \\leq L \\), with truncation if \\( l > L \\) and zero-padding in channels \\( l + 1 \\) to \\( L \\) if \\( l < L \\)), the \\( i \\)-th character is rendered into the \\( i \\)-th channel of an \\( L \\)-channel image \\( I_g \\). Each character glyph is rendered with a pixel intensity of \\( i \\times \\left\\lfloor \\frac{P_M}{L} \\right\\rfloor \\), where \\( P_M \\) denotes the maximum number of pixels allocated to distinguish a character from the background. We represent the glyph image tensor as \\( I_g \\in \\mathbb{R}^{L \\times S \\times S} \\), where \\( L \\) is the number of characters and \\( S \\) is the image size for each character. This creates position-prior pixel values that serve as positional encodings to help the model distinguish character information at each specific position." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.811 + ], + "angle": 0, + "content": "Vision Transformer (ViT) [7] is employed as the Glyph Encoder \\( G_{\\phi} \\) to accommodate input \\( I_{g} \\) and output features compatible with diffusion model's condition inputs:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.822, + 0.483, + 0.843 + ], + "angle": 0, + "content": "\\[\nF _ {g} = G _ {\\phi} \\left(I _ {g}\\right) \\in \\mathbb {R} ^ {(N + 1) \\times D} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Here, \\( N \\) represents the number of patches, the additional one corresponds to a class token, and dimension \\( D \\) is set to match the input requirements of the diffusion process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.384, + 0.771, + 0.4 + ], + "angle": 0, + "content": "3.3. Adapting to Stable Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.907, + 0.451 + ], + "angle": 0, + "content": "We leverage a pretrained model and fine-tune its Variational Auto Encoder (VAE) and retrain the Conditional Diffusion Model (CDM) components for our specific task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.908, + 0.633 + ], + "angle": 0, + "content": "VAE Fine-tuning. The standard VAE, when pretrained on general datasets, may not be sensitive enough to capture the fine-grained spatial and glyph features of scaled text instances. As highlighted by DiffUTE [3] and also confirmed by our ablation experiments, fine-tuning the VAE is essential when working with text images. Therefore, we fine-tune it to enhance the capacity for accurate representation of text regions in a fair, local manner. The VAE consists of an encoder \\( E_{\\theta} \\) and a decoder \\( D_{\\theta} \\), and we let \\( V_{\\theta} \\) represent the reconstruction after the encoding and decoding process. Following the pre-defined \\( I_{s} \\), we minimize the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.644, + 0.907, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {T e x t S S R - V A E}} = \\left\\| V _ {\\theta} \\left(I _ {s}\\right) - I _ {s} \\right\\| _ {2} ^ {2} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.674, + 0.907, + 0.809 + ], + "angle": 0, + "content": "CDM Retraining. To mask the actual text instance region while obtaining surrounding contextual information, we apply \\( I_{M} \\) to \\( I_{s} \\), resulting in \\( I_{m} \\), through the operation \\( I_{m} = I_{M} \\cdot I_{s} \\). \\( I_{m} \\), \\( I_{G} \\), and \\( I_{s} \\) are fed into the frozen VAE, generating latent representations \\( Z_{m} \\), \\( Z_{G} \\), and \\( Z_{s} \\). Additionally, \\( Z_{s} \\) is further processed with added noise to produce \\( Z_{T} \\). Moreover, we downsample \\( I_{M} \\) to match their dimensions (e.g., in SD 2.1 [33], height = width = 16), transforming it into \\( Z_{M} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "These latent features, \\( Z_{M}, Z_{T}, Z_{m} \\), and \\( Z_{G} \\), are concatenated and then passed through a Conv2D layer to match the channel dimension required by the U-Net (e.g., in SD 2.1, channel = 4). Meanwhile, the Glyph Encoder (see Fig. 3(a)) that produces \\( F_{g} \\) is also integrated into the CDM to extract features from \\( I_{g} \\). Finally, the CDM processes" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17467" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.875, + 0.245 + ], + "angle": 0, + "content": "
MethodIC13 [21]IC15 [22]ShopSign [54]
SeqAcc(%)↑NED ↑FID-R↓SeqAcc(%)↑NED ↑FID-R↓SeqAcc(%)↑NED ↑FID-R↓
TextDiffuser [4]25.080.4779103.740.820.0551120.660.640.0230136.56
GlyphControl [50]0.550.0472142.9100.0217152.3700.0067140.91
UDiffText [57]37.510.589267.822.550.1206103.43———
AnyText [41]55.070.715375.465.970.218676.0333.580.460289.25
TextDiffuser-2 [5]26.390.534268.310.870.083882.930.820.0411101.46
TextSSR Pre-training75.350.900751.7550.460.736134.6930.280.507554.24
TextSSR Fine-tuning85.170.940253.2862.640.808736.3052.240.705352.84
Real96.510.9895084.880.9486074.840.86990
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.254, + 0.908, + 0.311 + ], + "angle": 0, + "content": "Table 1. Comparing TextSSR (differences between the pre-training and fine-tuning versions are detailed in Appendix 6.3.1. We use the fine-tuned version, abbreviated as TextSSR, in the rest part of the paper.) and existing synthesis methods on regular text (IC13), irregular text (IC15), and multi-language text (ShopSign) datasets, alongside results from real datasets (Real). SeqAcc and NED denote the recognition accuracy, editing distance, respectively. FID-R score means FID on the text region." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.482, + 0.429 + ], + "angle": 0, + "content": "these inputs and produces the denoised latent representation \\( Z_{T-1} \\), progressively refining the noisy input to recover the clean latent space. Assuming \\( \\epsilon \\) represents the original added noise and \\( \\epsilon_{\\theta} \\) denotes the entire denoising process that outputs the predicted noise, the modified CDM is trained by minimizing the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.434, + 0.482, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {T e x t S S R - C D M}} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, Z _ {M}, Z _ {m}, Z _ {G}, I _ {g}\\right) \\right\\| _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.457, + 0.37, + 0.473 + ], + "angle": 0, + "content": "3.4. Anagram-based Data Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.479, + 0.483, + 0.75 + ], + "angle": 0, + "content": "As mentioned above, our pipeline uses text region and content detected by an OCR engine as inputs. The recognized content is regarded as pseudo-label of the text. Our generation produces a synthesized text instance on the detected region for each pseudo-label rather than the groundtruth, which is typically unknown. Therefore, errors in OCR would not affect the generation. To expand the generation in quantity, we propose an anagram-based data expansion strategy. Unlike previous approaches that arbitrarily edit text [4, 5, 41] or substitute equal-length strings [45, 57], we manipulate internal character order to maintain better region-content coherence. Meanwhile, the diffusion process adaptively handles the varying occupation requirement of different characters, e.g., 'o' typically is wider than 'l' in width. Therefore, our TextSSR leads to a more flexible generation. By changing the internal character order only, the region remain unchanged while still generating many samples with different character arrangements." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.751, + 0.483, + 0.811 + ], + "angle": 0, + "content": "Our model's character-level rendering capability enables correct generation of text from this permutation-based operation. We define the total number of possible permutations for a word of length \\( l \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.254, + 0.818, + 0.483, + 0.835 + ], + "angle": 0, + "content": "\\[\nP (l) = l! \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Note that some words may repeat due to character duplication; they still represent different samples, as the diffusion model is inherently random and there are also variations in cropped regions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.335, + 0.646, + 0.353 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.361, + 0.747, + 0.377 + ], + "angle": 0, + "content": "4.1. TextSSR Data Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.906, + 0.46 + ], + "angle": 0, + "content": "We first evaluate both accuracy and realism of the generation text, as well as the scaling capabilities of TextSSR data in small-scale. Then, we construct large-scale training data and assess its effectiveness at larger scales. Fig. 4 visualizes the examples of TextSSR compared to other methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.906, + 0.597 + ], + "angle": 0, + "content": "Accuracy Evaluation. Evaluating the accuracy of generated scene text is challenging, as manually verifying alignment with target text is costly. Previous benchmarks typically employ an STR model to automatically assess text consistency. We also follow this scheme. To mitigate potential OCR recognition errors and more effectively reflect the accuracy and difficulty level of our synthetic data, we choose SVTRv2 [11], a SOTA STR model upgraded from SVTR [8], as the recognition model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.748 + ], + "angle": 0, + "content": "Tab. 1 gives the accuracy assessment results across different datasets, where IC13 denotes all the methods using text instances from the IC13 dataset [21] to initialize their generation, while others are defined similarly. As can be seen, a large portion of text instances generated by TextSSR are correctly recognized. For more irregular and difficult IC15 [22], and less linguistically ShopSign [54] datasets, despite accuracy declines observed, TextSSR still outperforms the others by significant margins. The results provide strong evidence of TextSSR's correctness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Comparing with AnyText [41], the current SOTA, TextSSR outperforms it by \\(30.1\\%\\) and \\(56.7\\%\\) in accuracy on IC13 and IC15, respectively. This is mainly because TextSSR operates generation at the instance level, while AnyText performs on the full image level. Note that methods based on full-image synthesis struggle significantly in complex, challenging scenarios, e.g., IC15. This dataset includes many text instances affected by blur, low-resolution, perspective distortion, and curvature—challenges that previous methods do not address ef" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17468" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.177, + 0.093, + 0.256, + 0.105 + ], + "angle": 0, + "content": "GroundTruth" + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.109, + 0.262, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.161, + 0.263, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.274, + 0.094, + 0.349, + 0.105 + ], + "angle": 0, + "content": "TextDiffuser" + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.109, + 0.355, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.16, + 0.356, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.094, + 0.444, + 0.106 + ], + "angle": 0, + "content": "GlyphControl" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.109, + 0.448, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.161, + 0.449, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.094, + 0.526, + 0.105 + ], + "angle": 0, + "content": "UDiffText" + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.109, + 0.543, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.161, + 0.542, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.094, + 0.616, + 0.105 + ], + "angle": 0, + "content": "AnyText" + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.109, + 0.635, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.094, + 0.726, + 0.105 + ], + "angle": 0, + "content": "TextDiffuser-2" + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.108, + 0.73, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.16, + 0.728, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.094, + 0.802, + 0.104 + ], + "angle": 0, + "content": "TextSSR" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.108, + 0.824, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.161, + 0.824, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.177, + 0.192, + 0.262, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.243, + 0.262, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.192, + 0.355, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.243, + 0.356, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.192, + 0.448, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.243, + 0.449, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.192, + 0.543, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.243, + 0.542, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.192, + 0.635, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.243, + 0.635, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.192, + 0.73, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.243, + 0.729, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.192, + 0.811, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.243, + 0.824, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.176, + 0.274, + 0.262, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.327, + 0.261, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.274, + 0.355, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.326, + 0.355, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.274, + 0.448, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.363, + 0.326, + 0.448, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.275, + 0.543, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.275, + 0.635, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.326, + 0.635, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.275, + 0.729, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.326, + 0.728, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.275, + 0.811, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.327, + 0.823, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.371, + 0.907, + 0.4 + ], + "angle": 0, + "content": "Figure 4. Examples synthesized by TextSSR and other existing open-source methods (UDiffText only supports English text for its limited text encoder). TextSSR consistently produces accurate and realistic text instances. Please refer to Appendix 7.1 for more analysis." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.422, + 0.477, + 0.615 + ], + "angle": 0, + "content": "
TypeMethodSeqAcc(%)↑
RegularIrregularAvg
RenderingSynthText [15]48.5526.0839.74
VISD [53]38.0226.8833.65
UnrealText [25]32.0619.8527.28
DiffusionTextDiffuser [4]30.088.7921.74
GlyphControl [50]40.8013.2129.99
AnyText [41]49.8019.6237.97
TextDiffuser-2 [5]40.2610.4828.59
SceneVTG [60]54.9735.5047.34
TextSSR (Ours)59.7635.7150.33
RealReal-L [1]57.9737.6650.01
TextOCR [39]57.8341.5654.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.707 + ], + "angle": 0, + "content": "Table 2. Comparison of different data synthetic methods, where CRNN [36] trained on different synthetic data and real data of \\( {30}\\mathrm{k} \\) is employed to obtain the results. \"Regular\" denotes the average on IIIT [27], SVT [43], and IC13 [21] datasets, \"Irregular\" represents the average on IC15 [22], SVTP [29], and CUTE [32] datasets, while \"Avg\" is the average of the six datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.87 + ], + "angle": 0, + "content": "fectively. In contrast, TextSSR benefits from a region-centered generation pipeline. It excels at handling challenging text instances and maintains reasonable synthesis accuracy. Moreover, the condition for TextSSR is glyph shapes, which can be extended to other languages like Chinese. The results on ShopSign prove TextSSR's multilingual capability. It outperforms the purportedly multilingual-focused AnyText by \\(18.7\\%\\). More details on accuracy evaluation are provided in Appendix 6.3.2." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Realism Evaluation. Following the validation presented in SceneVTG [60], we generate a fixed amount (30k) of data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.907, + 0.517 + ], + "angle": 0, + "content": "to train an identical model under fair settings and then test it on common STR benchmarks [21, 22, 27, 29, 32, 43]. This approach avoids the need for separate verification of text correctness or synthesis quality. If the generated text is inaccurate or unrealistic, this will be reflected in poor recognition performance during testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.909, + 0.777 + ], + "angle": 0, + "content": "As shown in Tab. 2, for the compared methods, their average SeqAcc is below \\(40\\%\\) except for SceneVTG [60]. The result is closely linked to their low generation accuracy. Note that rendering-based methods generally perform better than traditional diffusion-based methods in irregular benchmarks, highlighting the challenge of generating highly controllable text images. SceneVTG benefits from its two-stage pipeline of text erasure and then rendering, it reaches a SeqAcc of \\(47.3\\%\\). Nevertheless, TextSSR, featured by its glyph-based condition injection and anagram-based generation, surpasses SceneVTG by nearly \\(3\\%\\). Moreover, the gap between models trained on 30k TextSSR and real-world TextOCR is only \\(3.7\\%\\), which comes from irregular text. These results convincingly demonstrate the realism of TextSSR data, especially in regular text synthesizing. However, further efforts are still needed to improve the synthesis of irregular text." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Scalability Evaluation. TextSSR data can be easily scaled up to a large quantity. It is necessary to assess whether the increase in data volume can translate into performance improvements. Therefore, we use the model generated by the 30k data in the realism evaluation as baseline (TextSSR \\(\\times 1\\)), and re-train CRNN under the same experimental setup but with data doubled every time. Results of this scalability evaluation are presented in Tab. 4. When scaled up" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17469" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.089, + 0.88, + 0.324 + ], + "angle": 0, + "content": "
ModelDatasetVolumeSeqAcc(%)↑
IIIT5k [27]SVT [43]IC13 [21]IC15 [22]SVTP [29]CUTE80 [32]AvgContextless [20]
CRNN [36]ST4m90.1382.8490.9072.4572.0982.9981.9033.50
ST6.98m90.8083.0091.8372.3473.8081.6082.0632.61
ST + SynthAdd4m + 1.05m90.4383.9391.1373.5572.7180.2181.9934.53
ST + TextSSR4m + 1.05m92.1387.0293.3577.4277.6784.3785.3355.97
ST + TextSSR4m + 3.55m93.5789.1894.2878.5878.4582.9986.1757.89
ST + TextOCR4m + 1.05m93.1388.1092.5379.6879.2283.6886.0644.93
ST + TextOCR + TextSSR5.05m + 3.55m94.5789.3494.4080.2379.2284.7287.0858.15
+1.44+1.24+1.87+0.55+0.00+1.04+1.02+13.22
MAERec [20]ST4m96.0091.9695.8083.8885.7489.5890.4959.43
ST6.98m96.9092.8995.8084.7086.8292.3691.5861.62
ST + SynthAdd4m + 1.05m95.9793.2095.8085.0486.9890.2891.2160.46
ST + TextSSR4m + 1.05m97.5393.8296.8586.7589.4694.1093.0876.77
ST + TextSSR4m + 3.55m98.1394.7497.3287.8589.6194.1093.6378.17
ST + TextOCR4m + 1.05m98.1095.5297.2089.3490.3996.8794.5774.33
ST + TextOCR + TextSSR5.05m + 3.55m98.2096.7597.2089.9592.7197.9295.4682.03
+0.10+1.23+0.00+0.61+2.32+1.05+0.89+7.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.334, + 0.905, + 0.364 + ], + "angle": 0, + "content": "Table 3. Performance of models trained on different data combinations, where CRNN and MAERec are selected as model representatives. Bold and underline indicate the highest and second highest values, respectively." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.386, + 0.449, + 0.533 + ], + "angle": 0, + "content": "
MethodVolumeSeqAcc(%)↑
RegularIrregularAvg
TextSSR ×130k59.7635.7150.33
TextSSR EL30k24.5810.2318.95
TextSSR ×260k60.4934.4950.29
TextSSR ×4120k65.1739.3455.03
TextSSR ×8240k73.9649.6964.44
TextOCR30k57.8341.5654.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.543, + 0.483, + 0.573 + ], + "angle": 0, + "content": "Table 4. TextSSR scalability evaluation. \\( \\times n \\) refers to perform the anagram-based rendering \\( n \\) times. EL denotes equal length edit." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.599, + 0.483, + 0.81 + ], + "angle": 0, + "content": "to \\(\\times 4\\) and \\(\\times 8\\), accuracy improvements of \\(4.7\\%\\) and \\(14.1\\%\\) are observed, respectively. The result indicates that increasing the data volume indeed contributes to better STR models. Note that models in the two scenarios also surpass the model trained on 30k real-world data by \\(1.0\\%\\) and \\(10.4\\%\\), respectively. It is observed that TextSSR \\(\\times 8\\) also presents a much better result than real data model in irregular text, highlighting that the challenge of irregularity can be overcome to some extent by increasing the data volume. Meanwhile, we also conducted a comparison experiment similar to [57], i.e., substituting text instances with equal-length strings (TextSSR EL), where the trained model reported quite worse results. This again demonstrates the effectiveness of our anagram-based generation pipeline." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Usability Assessment. We further assess the effectiveness of our synthesis method in supporting the large-scale training of STR models. To ensure the quality of data, we apply a quality-screened step to every sample generated by TextSSR. We still use SVTRv2 [11] to recognize the sample and compare the recognized content with its genera" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.908, + 0.829 + ], + "angle": 0, + "content": "tion label. The sample is reserved if they are consistent, otherwise we discard it. The results on different training data combinations are shown in Tab. 3. The observations can be summarized as follows: First, the rendering-based ST dataset reaches a saturation point in terms of accuracy, where \\(6.98\\mathrm{m}\\) samples only show marginal improvement \\((0.16\\%)\\) on CRNN over the \\(4\\mathrm{m}\\) dataset on the six common benchmarks. This suggests that we have to move towards more realistic data synthesis. Second, when incorporating \\(1.05\\mathrm{m}\\) our synthetic dataset, the two STR models (CRNN and MAERec) gain improvements of \\(3.4\\%\\) and \\(2.6\\%\\) over the \\(4\\mathrm{m}\\) ST dataset on the six common benchmarks, respectively. They are only \\(0.7\\%\\) and \\(1.5\\%\\) lower than ST combined with equivalent size real-world TextOCR. This suggests that TextSSR data is approaching real-world data in quality. Third, when further increasing TextSSR data to \\(3.55\\mathrm{m}\\), it is seen that the purely synthetic training data \\((\\mathrm{ST} + 3.55\\mathrm{m}\\) TextSSR) gain performance more approaching the synthetic-real mixed training data \\((\\mathrm{ST} + 1.05\\mathrm{m}\\) TextOCR). Moreover, when combining the \\(3.55\\mathrm{m}\\) synthetic data with this mixed data, the trained model can report further performance improvements, obtaining \\(1.02\\%\\) on CRNN and \\(0.89\\%\\) on MAERec. This implies that TextSSR data is a valuable supplement to existing training data. Using it individually or collectively both can benefit STR model training. As a result, we mark this \\(3.55\\mathrm{m}\\) dataset as TextSSR-F, an accurate and realistic large-scale synthetic dataset constructed by our TextSSR, and make it publicly available." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Since our anagram-based synthesis generates a large number of contextless words, and contextless word generation is also discussed in SynthAdd [24], we conduct a comparison between SynthAdd and our TextSSR. As shown" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17470" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.166 + ], + "angle": 0, + "content": "in Tab. 3, there are clear performance margins (3.3% on CRNN and 1.9% on MAERec) between the model trained on traditional SynthAdd and TextSSR. This also indicates the superiority of TextSSR as a STR training data synthetic method." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.177, + 0.47, + 0.314 + ], + "angle": 0, + "content": "
region-centricchar-glyphchar-positionfine-tune VAESeqAcc (%) ↑NED ↑FID-R ↓
20.280.3444102.89
✓73.500.876445.89
✓✓74.150.885246.65
✓✓✓80.260.913345.83
✓✓✓77.210.904350.34
✓✓76.230.899151.94
✓✓✓47.550.621162.52
✓✓✓✓82.990.930850.76
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.323, + 0.483, + 0.351 + ], + "angle": 0, + "content": "Table 5. Ablation study of TextSSR components using the TextSSR pre-training model in Tab. 1 on the IC13 dataset." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.378, + 0.245, + 0.394 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.484, + 0.611 + ], + "angle": 0, + "content": "To assess the proposed components, we start with the TextSSR pre-training model in Tab. 1 (to simplify, trained for only 5k steps), where all components are considered. Then, we stepwisely remove one or some of the components to assess their respective necessity and effectiveness. From the results on Tab. 5, we can see that the removal of each one brings a clear decline in performance, while the absence of \"region-centric\" processing can lead to a catastrophic drop of \\(35.4\\%\\) in accuracy. Additionally, fine-tuning the VAE offers a trade-off, slightly sacrificing visual quality for an improvement in accuracy. This is because real-world scene text also confronts poor visual quality challenge, simulating this to some extent can provide valuable samples for model training." + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.624, + 0.445, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.813, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Figure 5. Performance of TextSSR fine-tuning model with and without region-centric processing across text regions of different sizes (in pixels)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We perform another experiment to verify the effectiveness of the region-centric processing. The results in Fig. 5" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "show that without this processing, TextSSR has to synthesize text instance on the whole image. It exhibits a performance trend that first rises and then falls as the size increases. When the region is small, the entire image is scaled severely, resulting in more blurred text regions and insufficient pixels for text rendering. As the region size increases, larger regions and more pixels are allowed, therefore performance increase. When a too large region given, it causes the side effect that constraints the shape of the generated text, and performance decline. In contrast, incorporating region-centric processing allows TextSSR to achieve superior and consistent performance across all sizes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.905, + 0.334 + ], + "angle": 0, + "content": "In Fig. 6, we also illustratively ablate the character-level position and glyph components. The results indicate that omitting either component leads to issues such as character deformation, incorrect characters, and duplication errors." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.345, + 0.885, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.52, + 0.905, + 0.549 + ], + "angle": 0, + "content": "Figure 6. Visualizations of TextSSR with and without character-level prior. More examples are provided in Appendix 7.3" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.634, + 0.589 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We have proposed TextSSR to provide STR model training with high-quality synthetic data. It employs the glyph of existing text and surrounding context as the prompt of the diffusion-based synthesizing model, and combines character-level constraints and permutations. TextSSR successfully synthesizes accurate and realistic text instances at a large scale. The TextSSR-F dataset with 3.55 million diverse and realistic instances are constructed. We have conducted extensive experiments to assess TextSSR-F. STR models trained solely on TextSSR-F, and on the combination of TextSSR-F and real-world data both show performance improvements. While the effectiveness of TextSSR has been basically confirmed, for irregular text, there is still an accuracy gap between models trained on TextSSR-F and real-world data. Thus, future work includes the exploration of more controllable ways to further improve the synthesizing quality, especially for challenging instances such as curved and multi-oriented text. We are also interested in extending TextSSR to generate large-scale data in different languages like Chinese." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17471" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by the National Natural Science Foundation of China (Nos. 32341012, 62172103)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.151, + 0.188, + 0.167 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.176, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[1] J. Baek, Y. Matsui, and K. Aizawa. What if we only use real datasets for scene text recognition? toward scene text recognition with fewer labels. In CVPR, pages 3113-3122, 2021. 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.234, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[2] D. Bautista and R. Atienza. Scene text recognition with permuted autoregressive sequence models. In ECCV, pages 178-196, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.276, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[3] H. Chen, Z. Xu, Z. Gu, J. Lan, X. Zheng, Y. Li, C. Meng, H. Zhu, and W. Wang. DiffUTE: Universal text editing diffusion model. In NeurIPS, pages 63062-63074, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.319, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[4] J. Chen, Y. Huang, T. Lv, L. Cui, Q. Chen, and F. Wei. TextDiffuser: Diffusion models as text painters. In NeurIPS, pages 9353-9387, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[5] J. Chen, Y. Huang, T. Lv, L. Cui, Q. Chen, and F. Wei. TextDiffuser-2: Unleashing the power of language models for text rendering. In ECCV, pages 386–402, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.406, + 0.484, + 0.473 + ], + "angle": 0, + "content": "[6] C.K. Chng, Y. Liu, Y. Sun, C.C. Ng, C. Luo, Z. Ni, C.M. Fang, S. Zhang, J. Han, E. Ding, J. Liu, D. Karatzas, C.S. Chan, and L. Jin. ICDAR2019 robust reading challenge on arbitrary-shaped text-rc-art. In ICDAR, pages 1571–1576, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[7] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.547, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[8] Y. Du, Z. Chen, C. Jia, X. Yin, T. Zheng, C. Li, Y. Du, and Y. Jiang. SVTR: Scene text recognition with a single visual model. In IJCAI, pages 884-890, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[9] Y. Du, Z. Chen, C. Jia, X. Yin, C. Li, Y. Du, and Y.-G. Jiang. Context perception parallel decoder for scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell., 47(6):4668-4683, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.647, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[10] Y. Du, Z. Chen, Y. Su, C. Jia, and Y.-G. Jiang. Instruction-guided scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell., 47(4):2723-2738, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.689, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[11] Y. Du, Z. Chen, H. Xie, C. Jia, and Y. Jiang. SVTRv2: CTC beats encoder-decoder models in scene text recognition. In ICCV, 2025. 1, 3, 5, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[12] S. Fang, H. Xie, Y. Wang, Z. Mao, and Y. Zhang. Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition. In CVPR, pages 7098-7107, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[13] S. Fang, Z. Mao, H. Xie, Y. Wang, C. Yan, and Y. Zhang. Abinet++: Autonomous, bidirectional and iterative language modeling for scene text spotting. IEEE Trans. Pattern Anal. Mach. Intell., 45(6):7123-7141, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[14] J. Gu, X. Meng, G. Lu, L. Hou, N. Minzhe, X. Liang, L. Yao, R. Huang, W. Zhang, X. Jiang, C. Xu, and H. Xu. Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. In NeurIPS, pages 26418-26431, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.176, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[15] A. Gupta, A. Vedaldi, and A. Zisserman. Synthetic data for text localisation in natural images. In CVPR, pages 2315-2324, 2016. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[16] M. He, Y. Liu, Z. Yang, S. Zhang, C. Luo, F. Gao, Q. Zheng, Y. Wang, X. Zhang, and L. Jin. ICPR2018 contest on robust reading for multi-type web images. In ICPR, pages 7-12, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.906, + 0.235 + ], + "angle": 0, + "content": "[17] Santoso J, Simon C, and Williem. On manipulating scene text in the wild with diffusion models. In WACV, pages 5202-5211, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.238, + 0.906, + 0.292 + ], + "angle": 0, + "content": "[18] M. Jaderberg, K. Simonyan, A. Vedaldi, and A. Zisserman. Synthetic data and artificial neural networks for natural scene text recognition. In Workshop on Deep Learning, NeurIPS, 2014. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.296, + 0.906, + 0.337 + ], + "angle": 0, + "content": "[19] J. Ji, G. Zhang, Z. Wang, B. Hou, Z. Zhang, B. Price, and S. Chang. Improving diffusion models for scene text editing with dual encoders. TMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.906, + 0.38 + ], + "angle": 0, + "content": "[20] Q. Jiang, J. Wang, D. Peng, C. Liu, and L. Jin. Revisiting scene text recognition: A data perspective. In ICCV, pages 20543-20554, 2023. 1, 3, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.383, + 0.906, + 0.45 + ], + "angle": 0, + "content": "[21] D. Karatzas, F. Shafait, S. Uchida, M. Iwamura, L. Gomez i Bigorda, S. Robles Mestre, J. Mas, D. Fernandez Mota, J. Almazán, and L.P. de las Heras. ICDAR 2013 robust reading competition. In ICDAR, pages 1484-1493, 2013. 3, 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.454, + 0.906, + 0.523 + ], + "angle": 0, + "content": "[22] D. Karatzas, L. Gomez-Bigorda, A. Nicolaou, S. Ghosh, A. Bagdanov, M. Iwamura, J. Matas, L. Neumann, V.R. Chandrasekhar, S. Lu, F. Shafait, S. Uchida, and E. Valveny. IC-DAR 2015 competition on robust reading. In ICDAR, pages 1156-1160, 2015. 3, 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.906, + 0.581 + ], + "angle": 0, + "content": "[23] C. Li, W. Liu, R. Guo, X. Yin, K. Jiang, Y. Du, Y. Du, L. Zhu, B. Lai, X. Hu, D. Yu, and Y. Ma. PP-OCRv3: More attempts for the improvement of ultra lightweightOCR system. arXiv preprint arXiv:2206.03001, 2022. 1, 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.583, + 0.906, + 0.625 + ], + "angle": 0, + "content": "[24] H. Li, P. Wang, C. Shen, and G. Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In AAAI, pages 8610-8617, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.627, + 0.906, + 0.654 + ], + "angle": 0, + "content": "[25] S. Long and C. Yao. Unrealtext: Synthesizing realistic scene text images from the unreal world. In CVPR, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.656, + 0.906, + 0.697 + ], + "angle": 0, + "content": "[26] C. Luo, L. Jin, and Z. Sun. Moran: A multi-object rectified attention network for scene text recognition. Pattern Recognit., 90:109-118, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.906, + 0.741 + ], + "angle": 0, + "content": "[27] A. Mishra, K. Alahari, and CV Jawahar. Scene text recognition using higher order language priors. In BMVC, 2012. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.744, + 0.906, + 0.813 + ], + "angle": 0, + "content": "[28] N. Nayef, C. Liu, J. Ogier, Y. Patel, M. Busta, P.N. Chowdhury, D. Karatzas, W. Khlif, J. Matas, U. Pal, and J. Burie. ICDAR2019 robust reading challenge on multi-lingual scene text detection and recognition-rrc-plt-2019. In ICDAR, pages 1582-1587, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[29] T. Q. Phan, P. Shivakumara, S. Tian, and C. L. Tan. Recognizing text with perspective distortion in natural scenes. In ICCV, pages 569-576, 2013. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[30] Y. Qu, Q. Tan, H. Xie, J. Xu, Y. Wang, and Y. Zhang. Exploring stroke-level modifications for scene text editing. In AAAI, pages 2119-2127, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17472" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[31] M. Rang, Z. Bi, C. Liu, Y. Wang, and K. Han. An empirical study of scaling law for scene text recognition. In CVPR, pages 15619-15629, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[32] A. Risnumawan, P. Shivakumara, C. S. Chan, and C. L. Tan. A robust arbitrary text detection system for natural scene images. Expert Syst. Appl., 41(18):8027-8048, 2014. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[33] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[34] C. Schuhmann, R. Vencu, R. Beaumont, R. Kaczmarczyk, C. Mullis, A. Katta, T. Coombes, J. Jitsev, and A. Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[35] F. Sheng, Z. Chen, and B. Xu. NRTR: A no-recurrence sequence-to-sequence model for scene text recognition. In ICDAR, pages 781-786, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[36] B. Shi, X. Bai, and C. Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell., 39(11):2298-2304, 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[37] B. Shi, C. Yao, M. Liao, M. Yang, P. Xu, L. Cui, S. Belongie, S. Lu, and X. Bai. ICDAR2017 competition on reading Chinese text in the wild (rctw-17). In ICDAR, pages 1429–1434, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[38] B. Shi, M. Yang, X. Wang, P. Lyu, C. Yao, and X. Bai. ASTER: An attentional scene text recognizer with flexible rectification. IEEE Trans. Pattern Anal. Mach. Intell., 41(9): 2035-2048, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[39] A. Singh, G. Pang, M. Toh, J. Huang, W. Galuba, and T. Hassner. TextOCR: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. In CVPR, pages 8802-8812, 2021. 1, 3, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[40] Y. Sun, Z. Ni, C.K. Chng, Y. Liu, C. Luo, C.C. Ng, J. Han, E. Ding, J. Liu, D. Karatzas, C.S. Chan, and L. Jin. ICDAR 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In ICDAR, pages 1557-1562, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[41] Y. Tuo, W. Xiang, J. Y. He, Y. Geng, and X. Xie. AnyText: Multilingual visual text generation and editing. In ICLR, 2024. 2, 3, 4, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[42] A. Veit, T. Matera, L. Neumann, J. Matas, and S. Belongie. COCO-Text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[43] K. Wang, B. Babenko, and S. Belongie. End-to-end scene text recognition. In ICCV, pages 1457-1464, 2011. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[44] Y. Wang, H. Xie, S. Fang, M. Xing, J. Wang, S. Zhu, and Y. Zhang. Petr: Rethinking the capability of transformer-based language model in scene text recognition. IEEE Trans. Image Process., 31:5585-5598, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[45] Y. Wang, W. Zhang, H. Xu, and C. Jin. Dreamtext: High fidelity scene text synthesis. In CVPR, pages 28555-28563, 2025. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[46] J. Wei, H. Zhan, Y. Lu, X. Tu, B. Yin, C. Liu, and U. Pal. Image as a language: Revisiting scene text recognition via balanced, unified and synchronized vision-language reasoning network. In AAAI, pages 5885-5893, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[47] L. Wu, C. Zhang, J. Liu, J. Han, J. Liu, E. Ding, and X. Bai. Editing text in the wild. In ACM MM, pages 1500-1508, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[48] J. Xu, Y. Wang, H. Xie, and Y. Zhang. OTE: Exploring accurate scene text recognition using one token. In CVPR, pages 28327-28336, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.179, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[49] Q. Yang, J. Huang, and W. Lin. SwapText: Image based texts transfer in scenes. In CVPR, pages 14700-14709, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.208, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[50] Y. Yang, D. Gui, Y. Yuan, W. Liang, H. Ding, H. Hu, and K. Chen. GlyphControl: Glyph conditional control for visual text generation. In NeurIPS, pages 44050-44066, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.264, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[51] W. Zeng, Y. Shu, Z. Li, D. Yang, and Y. Zhou. TextCtrl: Diffusion-based scene text editing with prior guidance control. In NeurlPS, pages 138569–138594, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.307, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[52] C. Zhai, Z. Chen, J. Li, and B. Xu. Chinese image text recognition with blastm-ctc: a segmentation-free method. In CCPR, pages 525–536, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.35, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[53] F. Zhan, S. Lu, and C. Xue. Verisimilar image synthesis for accurate detection and recognition of texts in scenes. In ECCV, pages 249-266, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[54] C. Zhang, W. Ding, G. Peng, F. Fu, and W. Wang. Street view text recognition with deep learning for urban scene understanding in intelligent transportation systems. IEEE Trans. Intell. Transp., 22(7):4727-4743, 2021. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.449, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[55] R. Zhang, Y. Zhou, Q. Jiang, Q. Song, N. Li, K. Zhou, L. Wang, D. Wang, M. Liao, M. Yang, X. Bai, B. Shi, D. Karatzas, S. Lu, and C.V. Jawahar. ICDAR 2019 robust reading challenge on reading chinese text on signboard. In IC-DAR, pages 1577-1581, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.52, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[56] Z. Zhang, N. Lu, M. Liao, Y. Huang, C. Li, M. Wang, and W. Peng. Self-distillation regularized connectionist temporal classification loss for text recognition: A simple yet effective approach. In AAAI, pages 7441-7449, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.576, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[57] Y. Zhao and Z. Lian. UDiffText: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In ECCV, pages 217-233, 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.633, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[58] T. Zheng, Z. Chen, S. Fang, H. Xie, and Y.-G. Jiang. CDistNet: Perceiving multi-domain character distance for robust text recognition. Int. J. Comput. Vis., 132(2):300-318, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.688, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[59] Y. Zhu, Z. Li, T. Wang, M. He, and C. Yao. Conditional text image generation with diffusion models. In CVPR, pages 14235-14245, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.732, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[60] Y. Zhu, J. Liu, F. Gao, W. Liu, X. Wang, P. Wang, F. Huang, C. Yao, and Z. Yang. Visual text generation in the wild. In ECCV, pages 89-106, 2024. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17473" + } + ] +] \ No newline at end of file diff --git a/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/6a7bfaa2-52e3-41ce-8b5a-5f025a33ee77_origin.pdf b/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/6a7bfaa2-52e3-41ce-8b5a-5f025a33ee77_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7004fbbffeb710b11b35de4c3ff9ceed991b5d0e --- /dev/null +++ b/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/6a7bfaa2-52e3-41ce-8b5a-5f025a33ee77_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f2642e871fd0fa32895186adae539cd3bdbb1aad83fc1d9631426c2d89a30b45 +size 5009004 diff --git a/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/full.md b/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..08d0572b9326bf0abe117326d5dbf986d2207ea9 --- /dev/null +++ b/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/full.md @@ -0,0 +1,356 @@ +# TextSSR: Diffusion-based Data Synthesis for Scene Text Recognition + +Xingsong Ye $^{1}$ , Yongkun Du $^{1}$ , Yunbo Tao $^{1}$ , Zhineng Chen $^{2*}$ + +$^{1}$ College of Computer Science and Artificial Intelligence, Fudan University, China $^{2}$ Institute of Trustworthy Embodied AI, Fudan University, China + +{xsye20，zhinchen}@fudan.edu.cn，{ykdu23，ybtao24} $\text{已}$ m.fudan.edu.cn + +# Abstract + +Scene text recognition (STR) suffers from challenges of either less realistic synthetic training data or the difficulty of collecting sufficient high-quality real-world data, limiting the effectiveness of trained models. Meanwhile, despite producing holistically appealing text images, diffusion-based visual text generation methods struggle to synthesize accurate and realistic instance-level text at scale. To tackle this, we introduce TextSSR: a novel pipeline for Synthesizing Scene Text Recognition training data. TextSSR targets three key synthesizing characteristics: accuracy, realism, and scalability. It achieves accuracy through a proposed region-centric text generation with position-glyph enhancement, ensuring proper character placement. It maintains realism by guiding style and appearance generation using contextual hints from surrounding text or background. This character-aware diffusion architecture enjoys precise character-level control and semantic coherence preservation, without relying on natural language prompts. Therefore, TextSSR supports large-scale generation through combinatorial text permutations. Based on these, we present TextSSR-F, a dataset of 3.55 million quality-screened text instances. Extensive experiments show that STR models trained on TextSSR-F outperform those trained on existing synthetic datasets by clear margins on common benchmarks, and further improvements are observed when mixed with real-world training data. Code is available at https://github.com/YesianRohn/TextSSR. + +# 1. Introduction + +Scene text images are a unique type of image captured in the wild, they differ from general images in the presence of text, which appears with diverse sizes, fonts, and often with varying degrees of distortion and occlusion. However, these texts usually contain high-level semantic information closely related to the captured scene, making STR a crucial + +![](images/9ddccf2735e725e1dcbc87411ec22ecc338ae80305ad8baaa25a144930e5067c.jpg) +Figure 1. Top: Text instances detected by an end-to-end STR model [23], and synthesized by our TextSSR based on the detected examples. Bottom: Accuracy of MAERec [20] across common benchmarks. Peach indicates training on ST [15] and TextOCR [39], while blue denotes that TextSSR-F is additionally included as training data. + +task for complex scene understanding, image search, etc. + +Recent STR practices [9, 10, 20] have shown that besides advanced modeling techniques, high-quality training data also plays a critical role. Training STR models also follows the scaling law from the data perspective [31] and current data size is far from saturated even for English. Enlarging the size of real-world datasets is a feasible way to further enhance recognition performance. However, it is challenging to collect real-world scene text images on an even larger scale. On one hand, diversified data collection is costly and time-consuming. On the other hand, real-world scenes predominantly contain high-frequency, semantically meaningful words, while low-frequency words are difficult to collect. Synthesizing high-quality text images seems to be an effective alternative. However, research on this direction is somewhat laggard. The two most popular synthetic datasets, MJ [18] and ST [15], are constructed based on traditional rendering techniques nearly a decade ago. Although they can be easily scaled up in quantity, recent studies [9, 11] indicate that there are significant accuracy gaps between models trained on MJ&SJ and Union14M-L [20], mainly because MJ&SJ fail to comprehensively describing + +the challenge of real-world text instances. + +Recently, diffusion models have shown impressive performance in generating text-rich images [4, 5, 41, 50, 57]. However, most of them focus on generating aesthetic text images holistically, e.g., posters. Using data generated by these methods for STR model training encounters several issues. First, these methods are designed for the full image level generation, when switching to synthesizing text in instance-level regions, it would be less effective and ensuring the generation accuracy becomes challenging. Second, current generation methods mainly rely on well-refined natural language as prompts. They produce overly homogeneous and unrealistic text instances with some probability, which is not suitable as training data. Moreover, constructing a diverse set of prompts is inherently difficult, which hinders the ability to generate data at scale. + +Nevertheless, given the exceptional performance of diffusion models across various tasks, it is anticipated that there are diffusion-based techniques for producing satisfactory training data for STR. Before proceeding, we first outline three key characteristics that an improved data synthesis method should exhibit: + +- Accurate: The synthesized text content should match the given text, ensuring that the synthesized words are readable and free from character misplacement or duplication errors. This is fundamental for maintaining the correctness of the trained STR models. +- Realistic: The generated text instances should visually resemble real-world text, simulating scene conditions in the wild. This includes being harmoniously blended into the background, ensuring consistent generation at different sizes and capturing the diversity of text presentation, especially in complex scenes. Achieving realism is essential for enhancing the performance of text recognizers. +- Scalable: The method should support extensive training data production. In other words, abandon complex control inputs and design processes, and enable large-scale data generation by leveraging easily accessible and processable resources only. + +To satisfy these characteristics, we propose TextSSR, a novel diffusion-driven pipeline focusing on Synthesizing training data for Scene Text Recognition. Our approach introduces the following innovations to overcome persistent limitations in text synthesis. First, we implement a region-centric text generation. It employs end-to-end OCR model to acquire text location and size. Moreover, we devise a position-glyph enhancement that further resolves character duplication and positional disorder, ensuring precise text placement and arrangement. Second, we leverage the detected text or only background near the designated region as contextual hints for font style and appearance generation, guiding the generated text to match real-world scenarios better. Third, we design a character-aware diffusion + +![](images/4a54ed79f77e42a1a50dc46f1b833ebbbbc2fb8662c7056c6883d837d83a4661.jpg) +Figure 2. Showcases of TextSSR's capacity to synthesize accurate, realistic and scalable text instances. + +architecture that employs glyph priors rather than natural language prompts as injected conditions. It enables precise character-level control while maintaining full-word semantic coherence. To expand the generation quantity, we use detected text to generate a series of anagrams by rearranging characters within the original bounding boxes, thus supporting large-scale generation. As a result, TextSSR achieves accurate, realistic, and scalable text instance synthesize, as illustrated in Fig. 2. + +Through these attempts, we construct the TextSSR-F dataset with 3.55 million quality-screened samples, which is automatically validated by checking their recognitions and pseudo labels. Extensive experiments demonstrate the superiority of TextSSR-F over existing synthesis data, as well as its complementarity with real-world training data. + +Our contributions are summarized as follows: + +- We propose TextSSR, a diffusion-based text synthesis method centered on seamlessly blending text instances into real scenes. It develops a region-centric processing that enables accurate instance-level generation. +- TextSSR fully exploits positional and glyph priors of text, achieving accurate, diverse and realistic text synthesis by using only the detected text as input. Moreover, it can be easily scaled up and we construct the TextSSR-F dataset with 3.55 million text instances. +- Extensive evaluations validate that STR models trained on TextSSR-F surpass their counterparts trained on existing synthetic data by $3.1\%$ on common benchmarks, and even better STR models can be obtained when further combined with real-world training data, highlighting the utility of the proposed TextSSR. + +# 2. Related Work + +# 2.1. Scene Text Recognition + +STR researches can be categorized from both the model and the data perspectives. From a model point of view, early approaches employ the CTC-based pipeline [36, 52]. + +They used CNN for visual feature extraction, followed by RNN for sequence modeling, then with CTC loss applied as a constraint. Subsequently, attention mechanisms gradually gained attention in the STR field [24, 26, 35, 38]. Inspired by developments in the NLP field, many methods [2, 12, 13, 44, 46, 48, 58] further refined recognizers from a language modeling perspective. While some approaches [8, 11, 56] continued to explore solutions based on a single vision model and CTC decoding, due to their fast inference nature. Accuracy was steadily improved following these efforts. + +From a data perspective, the limited availability of well-annotated real-world training data led to a practical STR training and evaluation protocol: training solely on synthetic data, or a combination of synthetic and real data, followed by evaluation on real-world benchmarks. Besides the six common benchmarks [21, 22, 27, 29, 32, 43]. Union14M-L [20] aggregated publicly available real-world datasets (including TextOCR [39]) and created a real-world training dataset with nearly 4 million instances. However, recent studies on the OCR scaling law [31] suggested that even this substantial volume had not reached the saturation point. This observation implies that STR models can still benefit from additional training data, and high-quality synthetic data could be a feasible supplement, which motivates our study. + +# 2.2. Scene Text Synthesis + +Early synthesis methods [15, 18, 25] employed rule-based methods to overlay text onto background with transformations simulating real-world appearance. To be more realistic, later research shifted toward deep learning-based generation methods. Among these, GAN-based scene text editing [30, 47, 49] methods aimed to expand data by editing text directly within specified regions, but they were limited by the lack of real paired training data. Recent advances had leveraged diffusion models for both editing [17, 19, 51] and synthesis, with fine-grained control over glyphs [3, 50], as well as positioning at word [41] and character [4, 45, 57] levels. The latest methods [5, 60] incorporate large language or multi-modal models to enhance performance. However, these approaches primarily targeted visually appealing generation for aesthetic purposes, rather than generating data for better STR models. While some work has explored diffusion-based text image synthesis. For example, CTIG-DM [59] employed character sequence as the condition for generation. However, it is less controllable and faced with the dilemma of diversity-realism tradeoff, limiting quality of the generated text. SceneVTG [60] developed a first-erasing-then-synthesizing pipeline. It requires a pre-erased model to remove existing text completely, which is difficult to guarantee in large-scale synthesizing scenarios. + +# 3. Method + +To achieve accurate, realistic, and scalable text instance generation, TextSSR proposes a four-step pipeline (detailed in Fig. 3(b)). To handle text instances with varying sizes and locations, we first focus on text region processing, which will be discussed in Sec. 3.1. To enhance accuracy through fine-grained extraction and decoupling of text glyphs and positions, Sec. 3.2 and Sec. 3.3 will detail the specific module design and training strategy, respectively. Sec. 3.4 introduces our data extension method, achieving scalability in data synthesis at scale. + +# 3.1. Text Region-centric Processing + +Instance-level text generation presents significant challenges, as we must handle varying sizes, diverse potential positions, and determine how to select appropriate surrounding information for effective prompting. Synthesis methods that edit the entire image are not suitable for this task. Such approaches create an imbalance, and small text regions lack sufficient pixels for accurate reconstruction. Furthermore, information from the entire scene is often redundant for text region synthesis. For example, in Fig. 3(a), the text "kills" to be synthesized shares a similar style and context with the nearby upper text and background, while the region farther away is inconsistent in appearance. + +Therefore, we propose a generative scheme based on a locality assumption—text styles are more likely to correlate with adjacent text and background information. This method focuses generation at the region level, synthesizing the text using only surrounding local context rather than the entire scene. Specifically, we use the annotated text box to extract the smallest enclosing square around the text region and resize it to a fixed size (e.g., 256). For practical considerations, we uniformly employ an end-to-end OCR tool [23] to obtain text instance size, position, and textual labels. It well reflects real-world scenarios where annotations are often unavailable. + +# 3.2. Fine-grained Condition Mechanisms + +Visual text rendering needs to consider two key fine-grained features: glyph structure and positional arrangement, both at word (holistic) and character (individual) levels. Word-level features concern entire word appearance and region placement, while the character-level address individual characters and their relative ordering. To guide the generative process, we implement the following approach: + +Word-level Position and Glyph Guidance. For a text $t$ with position $P_{t}$ in the original image, we extract a square region $I_{s}$ where $P_{t}$ becomes $P_{t}^{\prime}$ (Fig. 3(a), red dashed rectangle). This allows us to establish our desired word-level position guidance. Then we create a binary mask $I_{M}$ that separates text (0) from non-text (1) regions. The size of $I_{M}$ is identical to $I_{s}$ , indicating where the text is present. For + +![](images/25fcfe9f298a6c509b7f84f45e39131128ff186b82647f258cc5ba781efdb548.jpg) +(a) Generative Model + +![](images/03062fe75ce00b17cd5082df12f1839f15822ac1dee81edd643d2fe03f880ab9.jpg) +(b) Synthesis Pipeline +Figure 3. Overall architecture of TextSSR's (a) generative model and (b) synthesis pipeline. (a) Given a scene text instance (marked in red) and its content (e.g., "kills"), it is preprocessed into the smallest local image containing the text. A text-size mask is then applied to identify non-text regions, while the text content is rendered as priors at both word and character levels. These different conditions are subsequently injected into a diffusion model as training guidance. (b) An end-to-end OCR model extracts the text location and content within the image (Step 1). All possible text words are generated through anagram (Step 2). The generative model then synthesizes text instances in batches (Step 3), and finally, a SOTA STR model filters the samples, retaining only the correctly recognized instances for use (Step 4). + +glyph guidance, we render the text in a standard font, applying an affine transformation to align it with the position and orientation of $P_{t}^{\prime}$ (resulting in $I_{G}$ ). As noted in previous work [3, 41, 50], when we close the door on the original styled text, we must at least provide a window with a basic visual style of the text at that location. + +Character-level Position and Glyph Guidance. Only word-level information limits the model's understanding to the words, as we observe issues such as repeated or incorrect characters when using only word-level features. We address this by encoding character-level ordering through pixel values, enabling the model to learn spatial relationships between characters. Specifically, we define a maximum character length $L$ . For characters from position 1 to $l$ ( $l \leq L$ , with truncation if $l > L$ and zero-padding in channels $l + 1$ to $L$ if $l < L$ ), the $i$ -th character is rendered into the $i$ -th channel of an $L$ -channel image $I_g$ . Each character glyph is rendered with a pixel intensity of $i \times \left\lfloor \frac{P_M}{L} \right\rfloor$ , where $P_M$ denotes the maximum number of pixels allocated to distinguish a character from the background. We represent the glyph image tensor as $I_g \in \mathbb{R}^{L \times S \times S}$ , where $L$ is the number of characters and $S$ is the image size for each character. This creates position-prior pixel values that serve as positional encodings to help the model distinguish character information at each specific position. + +Vision Transformer (ViT) [7] is employed as the Glyph Encoder $G_{\phi}$ to accommodate input $I_{g}$ and output features compatible with diffusion model's condition inputs: + +$$ +F _ {g} = G _ {\phi} \left(I _ {g}\right) \in \mathbb {R} ^ {(N + 1) \times D} \tag {1} +$$ + +Here, $N$ represents the number of patches, the additional one corresponds to a class token, and dimension $D$ is set to match the input requirements of the diffusion process. + +# 3.3. Adapting to Stable Diffusion + +We leverage a pretrained model and fine-tune its Variational Auto Encoder (VAE) and retrain the Conditional Diffusion Model (CDM) components for our specific task. + +VAE Fine-tuning. The standard VAE, when pretrained on general datasets, may not be sensitive enough to capture the fine-grained spatial and glyph features of scaled text instances. As highlighted by DiffUTE [3] and also confirmed by our ablation experiments, fine-tuning the VAE is essential when working with text images. Therefore, we fine-tune it to enhance the capacity for accurate representation of text regions in a fair, local manner. The VAE consists of an encoder $E_{\theta}$ and a decoder $D_{\theta}$ , and we let $V_{\theta}$ represent the reconstruction after the encoding and decoding process. Following the pre-defined $I_{s}$ , we minimize the following loss function: + +$$ +\mathcal {L} _ {\text {T e x t S S R - V A E}} = \left\| V _ {\theta} \left(I _ {s}\right) - I _ {s} \right\| _ {2} ^ {2} \tag {2} +$$ + +CDM Retraining. To mask the actual text instance region while obtaining surrounding contextual information, we apply $I_{M}$ to $I_{s}$ , resulting in $I_{m}$ , through the operation $I_{m} = I_{M} \cdot I_{s}$ . $I_{m}$ , $I_{G}$ , and $I_{s}$ are fed into the frozen VAE, generating latent representations $Z_{m}$ , $Z_{G}$ , and $Z_{s}$ . Additionally, $Z_{s}$ is further processed with added noise to produce $Z_{T}$ . Moreover, we downsample $I_{M}$ to match their dimensions (e.g., in SD 2.1 [33], height = width = 16), transforming it into $Z_{M}$ . + +These latent features, $Z_{M}, Z_{T}, Z_{m}$ , and $Z_{G}$ , are concatenated and then passed through a Conv2D layer to match the channel dimension required by the U-Net (e.g., in SD 2.1, channel = 4). Meanwhile, the Glyph Encoder (see Fig. 3(a)) that produces $F_{g}$ is also integrated into the CDM to extract features from $I_{g}$ . Finally, the CDM processes + +
MethodIC13 [21]IC15 [22]ShopSign [54]
SeqAcc(%)↑NED ↑FID-R↓SeqAcc(%)↑NED ↑FID-R↓SeqAcc(%)↑NED ↑FID-R↓
TextDiffuser [4]25.080.4779103.740.820.0551120.660.640.0230136.56
GlyphControl [50]0.550.0472142.9100.0217152.3700.0067140.91
UDiffText [57]37.510.589267.822.550.1206103.43———
AnyText [41]55.070.715375.465.970.218676.0333.580.460289.25
TextDiffuser-2 [5]26.390.534268.310.870.083882.930.820.0411101.46
TextSSR Pre-training75.350.900751.7550.460.736134.6930.280.507554.24
TextSSR Fine-tuning85.170.940253.2862.640.808736.3052.240.705352.84
Real96.510.9895084.880.9486074.840.86990
+ +Table 1. Comparing TextSSR (differences between the pre-training and fine-tuning versions are detailed in Appendix 6.3.1. We use the fine-tuned version, abbreviated as TextSSR, in the rest part of the paper.) and existing synthesis methods on regular text (IC13), irregular text (IC15), and multi-language text (ShopSign) datasets, alongside results from real datasets (Real). SeqAcc and NED denote the recognition accuracy, editing distance, respectively. FID-R score means FID on the text region. + +these inputs and produces the denoised latent representation $Z_{T-1}$ , progressively refining the noisy input to recover the clean latent space. Assuming $\epsilon$ represents the original added noise and $\epsilon_{\theta}$ denotes the entire denoising process that outputs the predicted noise, the modified CDM is trained by minimizing the following objective: + +$$ +\mathcal {L} _ {\text {T e x t S S R - C D M}} = \left\| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, Z _ {M}, Z _ {m}, Z _ {G}, I _ {g}\right) \right\| _ {2} ^ {2} \tag {3} +$$ + +# 3.4. Anagram-based Data Synthesis + +As mentioned above, our pipeline uses text region and content detected by an OCR engine as inputs. The recognized content is regarded as pseudo-label of the text. Our generation produces a synthesized text instance on the detected region for each pseudo-label rather than the groundtruth, which is typically unknown. Therefore, errors in OCR would not affect the generation. To expand the generation in quantity, we propose an anagram-based data expansion strategy. Unlike previous approaches that arbitrarily edit text [4, 5, 41] or substitute equal-length strings [45, 57], we manipulate internal character order to maintain better region-content coherence. Meanwhile, the diffusion process adaptively handles the varying occupation requirement of different characters, e.g., 'o' typically is wider than 'l' in width. Therefore, our TextSSR leads to a more flexible generation. By changing the internal character order only, the region remain unchanged while still generating many samples with different character arrangements. + +Our model's character-level rendering capability enables correct generation of text from this permutation-based operation. We define the total number of possible permutations for a word of length $l$ as: + +$$ +P (l) = l! \tag {4} +$$ + +Note that some words may repeat due to character duplication; they still represent different samples, as the diffusion model is inherently random and there are also variations in cropped regions. + +# 4. Experiments + +# 4.1. TextSSR Data Evaluation + +We first evaluate both accuracy and realism of the generation text, as well as the scaling capabilities of TextSSR data in small-scale. Then, we construct large-scale training data and assess its effectiveness at larger scales. Fig. 4 visualizes the examples of TextSSR compared to other methods. + +Accuracy Evaluation. Evaluating the accuracy of generated scene text is challenging, as manually verifying alignment with target text is costly. Previous benchmarks typically employ an STR model to automatically assess text consistency. We also follow this scheme. To mitigate potential OCR recognition errors and more effectively reflect the accuracy and difficulty level of our synthetic data, we choose SVTRv2 [11], a SOTA STR model upgraded from SVTR [8], as the recognition model. + +Tab. 1 gives the accuracy assessment results across different datasets, where IC13 denotes all the methods using text instances from the IC13 dataset [21] to initialize their generation, while others are defined similarly. As can be seen, a large portion of text instances generated by TextSSR are correctly recognized. For more irregular and difficult IC15 [22], and less linguistically ShopSign [54] datasets, despite accuracy declines observed, TextSSR still outperforms the others by significant margins. The results provide strong evidence of TextSSR's correctness. + +Comparing with AnyText [41], the current SOTA, TextSSR outperforms it by $30.1\%$ and $56.7\%$ in accuracy on IC13 and IC15, respectively. This is mainly because TextSSR operates generation at the instance level, while AnyText performs on the full image level. Note that methods based on full-image synthesis struggle significantly in complex, challenging scenarios, e.g., IC15. This dataset includes many text instances affected by blur, low-resolution, perspective distortion, and curvature—challenges that previous methods do not address ef + +![](images/9121d23c28265570b1251692e433996e94cf69452c05f49b2667e846a80332c3.jpg) +GroundTruth + +![](images/7974814931c525ae5a2445b1c8ee63bbca8dbc5d799b35bed606e7c7c31557d0.jpg) + +![](images/69a9d689acda4cc4721086811c89418badbf764ebf6733cc5ad5a2d5121ba20d.jpg) +TextDiffuser + +![](images/d497c04336417c192bd0cdd89a6796db7ce379891128d5d4e7c663747b8906ca.jpg) + +![](images/e7b4110a01d99aa29900f4def875d7263a84c6c79de7ad05fb8f6b9aeabfda2d.jpg) +GlyphControl + +![](images/87268eb510134ba42b6dcaa7e96227b170abb248b68d56107c20e92e48bd2e78.jpg) + +![](images/1d1e1e7e1d79e72a195f7daee2868271683afcaf59d08919e5ec588bfa1d0ded.jpg) +UDiffText + +![](images/77d976b5bab074c34b5386516de6ea7d6b297b579223c2b55c4f056540b02189.jpg) + +![](images/9ae130331065af71b444b44b67bba559e4fc540bce5f600dd73531fb6c3e74d9.jpg) +AnyText + +![](images/55be7c6ec5685408af1e65fbcbfb30b22a430af3b4faea505b091caacaac5dfe.jpg) +TextDiffuser-2 + +![](images/979797ea2384fdc03db4db3f77e3c74885eb8fa32176bea947736c94dcac860c.jpg) + +![](images/1fc3ac4e4cf38c09fd51caae7558cc4fba1cd0acd14a0a17c4352a9e5fa31b3a.jpg) +TextSSR + +![](images/580c1b1f679a309b1ece3cf8109fa10ea0e85923768b88ccb94b0f00a76604c9.jpg) + +![](images/91f58b865faadcd8c362ebb9a92ea8be5fa2fa37e9b8a311e3fd4f79044be5db.jpg) + +![](images/1616cd0213d11274b0dc7f59b3f340e4da0fdc54389fb6c6aa74d827dd7a12ef.jpg) + +![](images/144d84cd9cb27db04d8f1d316b0e72a820473508a589fa5e81d389f1222a24d3.jpg) + +![](images/4c766147a574570cd9a4321677ab085472fc40ee5cc8489c046782d0ab63dcfc.jpg) + +![](images/7f477be425f529b864d86c35a7594d04f3d408d8fe66e71ce68c212e2411a2a1.jpg) + +![](images/34c14200ca0bc58b012ab030e2ea3c0c0356c0d8bf1e844e4c637623356db553.jpg) + +![](images/946b835d48180427168e31c1f5114a2902a62049255d1c9460055f268acbc9b8.jpg) + +![](images/bd50607e466fbfe520083c565dceca8ea45d00241460014852a3e8c0a9c69956.jpg) + +![](images/0b1e3369e7d86773fb42e5d4986f67074bb0d18c4c62a82ee69091e436173878.jpg) + +![](images/8f58f7bd09506221ca649410e84569d3f3ab4b12591e70d016894e1d469df770.jpg) + +![](images/f7bc5573365aa391984850322c4e22212ec6fa21a67a4e1a6709acac54648007.jpg) + +![](images/3ad719daf98b1b7a772cbff0b3a082c10d2e6c1b24dd0f7b094756e37e557db7.jpg) + +![](images/4b49b7ec54917f988f627b93d7e5b13ecb6036b869bd3b7375ff9dfbb9e2b5af.jpg) + +![](images/4b4dc2ee993be44a877141e4a331dbc0ec0394abf6ba70b5e121ce952f4f3c2e.jpg) + +![](images/81bdd850e6f2dcd8d663805deb6aa990c15a7b54876f5363ab0a9db5fa9f46e2.jpg) + +![](images/7216de48a3193cad2f787e266d0a040e66403d62fe7fa42851da536da85c990f.jpg) +Figure 4. Examples synthesized by TextSSR and other existing open-source methods (UDiffText only supports English text for its limited text encoder). TextSSR consistently produces accurate and realistic text instances. Please refer to Appendix 7.1 for more analysis. + +![](images/618549cc2d5d54d058c21ba95453feebce23ca4cfad0b7b43d62f9786b4373f2.jpg) + +![](images/248400ce8d95bfcbb656f163cbd3d56f9c618b7acad871450e22dad576339579.jpg) + +![](images/8f0285cf616479a8cbc6999d5b6bd0a035b31100552df8dac25436a3a9898699.jpg) + +![](images/c6b9bdf86c7677436dd02108319aa3bfb9bcef267e0dc2d02a1d7f75a96e13d0.jpg) + +![](images/293c00de0c3af1731ef69f3b389c1e49094861a6423249135df63d4e038f62d0.jpg) + +![](images/4a83c1a140a826fa802a317aee4a31086c43c12b6e4642feb6c8fa9aaaafb585.jpg) + +![](images/f539f2a66f6215d63a38f836a25a5b5a33abecb45e9363b6df9c7c63dd164dbf.jpg) + +![](images/e8b971d361f01c594e4b038c4ab45d51393422c14216a3a428bc6c2205e6a962.jpg) + +![](images/16ae66567977c09ffca176166f5fa7f89317e43299b35224b8dd41ac60acb7fe.jpg) + +![](images/ffe4026785d57d371d7826e14228e9968885ea3ca54d62390f401cd2b7fd8899.jpg) + +![](images/f51e3698a500ccc29c39d9e1a89aef5fb65b83e2dfc5b68c7c7bb889667f8a37.jpg) + +
TypeMethodSeqAcc(%)↑
RegularIrregularAvg
RenderingSynthText [15]48.5526.0839.74
VISD [53]38.0226.8833.65
UnrealText [25]32.0619.8527.28
DiffusionTextDiffuser [4]30.088.7921.74
GlyphControl [50]40.8013.2129.99
AnyText [41]49.8019.6237.97
TextDiffuser-2 [5]40.2610.4828.59
SceneVTG [60]54.9735.5047.34
TextSSR (Ours)59.7635.7150.33
RealReal-L [1]57.9737.6650.01
TextOCR [39]57.8341.5654.05
+ +Table 2. Comparison of different data synthetic methods, where CRNN [36] trained on different synthetic data and real data of ${30}\mathrm{k}$ is employed to obtain the results. "Regular" denotes the average on IIIT [27], SVT [43], and IC13 [21] datasets, "Irregular" represents the average on IC15 [22], SVTP [29], and CUTE [32] datasets, while "Avg" is the average of the six datasets. + +fectively. In contrast, TextSSR benefits from a region-centered generation pipeline. It excels at handling challenging text instances and maintains reasonable synthesis accuracy. Moreover, the condition for TextSSR is glyph shapes, which can be extended to other languages like Chinese. The results on ShopSign prove TextSSR's multilingual capability. It outperforms the purportedly multilingual-focused AnyText by $18.7\%$ . More details on accuracy evaluation are provided in Appendix 6.3.2. + +Realism Evaluation. Following the validation presented in SceneVTG [60], we generate a fixed amount (30k) of data + +to train an identical model under fair settings and then test it on common STR benchmarks [21, 22, 27, 29, 32, 43]. This approach avoids the need for separate verification of text correctness or synthesis quality. If the generated text is inaccurate or unrealistic, this will be reflected in poor recognition performance during testing. + +As shown in Tab. 2, for the compared methods, their average SeqAcc is below $40\%$ except for SceneVTG [60]. The result is closely linked to their low generation accuracy. Note that rendering-based methods generally perform better than traditional diffusion-based methods in irregular benchmarks, highlighting the challenge of generating highly controllable text images. SceneVTG benefits from its two-stage pipeline of text erasure and then rendering, it reaches a SeqAcc of $47.3\%$ . Nevertheless, TextSSR, featured by its glyph-based condition injection and anagram-based generation, surpasses SceneVTG by nearly $3\%$ . Moreover, the gap between models trained on 30k TextSSR and real-world TextOCR is only $3.7\%$ , which comes from irregular text. These results convincingly demonstrate the realism of TextSSR data, especially in regular text synthesizing. However, further efforts are still needed to improve the synthesis of irregular text. + +Scalability Evaluation. TextSSR data can be easily scaled up to a large quantity. It is necessary to assess whether the increase in data volume can translate into performance improvements. Therefore, we use the model generated by the 30k data in the realism evaluation as baseline (TextSSR $\times 1$ ), and re-train CRNN under the same experimental setup but with data doubled every time. Results of this scalability evaluation are presented in Tab. 4. When scaled up + +
ModelDatasetVolumeSeqAcc(%)↑
IIIT5k [27]SVT [43]IC13 [21]IC15 [22]SVTP [29]CUTE80 [32]AvgContextless [20]
CRNN [36]ST4m90.1382.8490.9072.4572.0982.9981.9033.50
ST6.98m90.8083.0091.8372.3473.8081.6082.0632.61
ST + SynthAdd4m + 1.05m90.4383.9391.1373.5572.7180.2181.9934.53
ST + TextSSR4m + 1.05m92.1387.0293.3577.4277.6784.3785.3355.97
ST + TextSSR4m + 3.55m93.5789.1894.2878.5878.4582.9986.1757.89
ST + TextOCR4m + 1.05m93.1388.1092.5379.6879.2283.6886.0644.93
ST + TextOCR + TextSSR5.05m + 3.55m94.5789.3494.4080.2379.2284.7287.0858.15
+1.44+1.24+1.87+0.55+0.00+1.04+1.02+13.22
MAERec [20]ST4m96.0091.9695.8083.8885.7489.5890.4959.43
ST6.98m96.9092.8995.8084.7086.8292.3691.5861.62
ST + SynthAdd4m + 1.05m95.9793.2095.8085.0486.9890.2891.2160.46
ST + TextSSR4m + 1.05m97.5393.8296.8586.7589.4694.1093.0876.77
ST + TextSSR4m + 3.55m98.1394.7497.3287.8589.6194.1093.6378.17
ST + TextOCR4m + 1.05m98.1095.5297.2089.3490.3996.8794.5774.33
ST + TextOCR + TextSSR5.05m + 3.55m98.2096.7597.2089.9592.7197.9295.4682.03
+0.10+1.23+0.00+0.61+2.32+1.05+0.89+7.70
+ +Table 3. Performance of models trained on different data combinations, where CRNN and MAERec are selected as model representatives. Bold and underline indicate the highest and second highest values, respectively. + +
MethodVolumeSeqAcc(%)↑
RegularIrregularAvg
TextSSR ×130k59.7635.7150.33
TextSSR EL30k24.5810.2318.95
TextSSR ×260k60.4934.4950.29
TextSSR ×4120k65.1739.3455.03
TextSSR ×8240k73.9649.6964.44
TextOCR30k57.8341.5654.05
+ +Table 4. TextSSR scalability evaluation. $\times n$ refers to perform the anagram-based rendering $n$ times. EL denotes equal length edit. + +to $\times 4$ and $\times 8$ , accuracy improvements of $4.7\%$ and $14.1\%$ are observed, respectively. The result indicates that increasing the data volume indeed contributes to better STR models. Note that models in the two scenarios also surpass the model trained on 30k real-world data by $1.0\%$ and $10.4\%$ , respectively. It is observed that TextSSR $\times 8$ also presents a much better result than real data model in irregular text, highlighting that the challenge of irregularity can be overcome to some extent by increasing the data volume. Meanwhile, we also conducted a comparison experiment similar to [57], i.e., substituting text instances with equal-length strings (TextSSR EL), where the trained model reported quite worse results. This again demonstrates the effectiveness of our anagram-based generation pipeline. + +Usability Assessment. We further assess the effectiveness of our synthesis method in supporting the large-scale training of STR models. To ensure the quality of data, we apply a quality-screened step to every sample generated by TextSSR. We still use SVTRv2 [11] to recognize the sample and compare the recognized content with its genera + +tion label. The sample is reserved if they are consistent, otherwise we discard it. The results on different training data combinations are shown in Tab. 3. The observations can be summarized as follows: First, the rendering-based ST dataset reaches a saturation point in terms of accuracy, where $6.98\mathrm{m}$ samples only show marginal improvement $(0.16\%)$ on CRNN over the $4\mathrm{m}$ dataset on the six common benchmarks. This suggests that we have to move towards more realistic data synthesis. Second, when incorporating $1.05\mathrm{m}$ our synthetic dataset, the two STR models (CRNN and MAERec) gain improvements of $3.4\%$ and $2.6\%$ over the $4\mathrm{m}$ ST dataset on the six common benchmarks, respectively. They are only $0.7\%$ and $1.5\%$ lower than ST combined with equivalent size real-world TextOCR. This suggests that TextSSR data is approaching real-world data in quality. Third, when further increasing TextSSR data to $3.55\mathrm{m}$ , it is seen that the purely synthetic training data $(\mathrm{ST} + 3.55\mathrm{m}$ TextSSR) gain performance more approaching the synthetic-real mixed training data $(\mathrm{ST} + 1.05\mathrm{m}$ TextOCR). Moreover, when combining the $3.55\mathrm{m}$ synthetic data with this mixed data, the trained model can report further performance improvements, obtaining $1.02\%$ on CRNN and $0.89\%$ on MAERec. This implies that TextSSR data is a valuable supplement to existing training data. Using it individually or collectively both can benefit STR model training. As a result, we mark this $3.55\mathrm{m}$ dataset as TextSSR-F, an accurate and realistic large-scale synthetic dataset constructed by our TextSSR, and make it publicly available. + +Since our anagram-based synthesis generates a large number of contextless words, and contextless word generation is also discussed in SynthAdd [24], we conduct a comparison between SynthAdd and our TextSSR. As shown + +in Tab. 3, there are clear performance margins (3.3% on CRNN and 1.9% on MAERec) between the model trained on traditional SynthAdd and TextSSR. This also indicates the superiority of TextSSR as a STR training data synthetic method. + +
region-centricchar-glyphchar-positionfine-tune VAESeqAcc (%) ↑NED ↑FID-R ↓
20.280.3444102.89
✓73.500.876445.89
✓✓74.150.885246.65
✓✓✓80.260.913345.83
✓✓✓77.210.904350.34
✓✓76.230.899151.94
✓✓✓47.550.621162.52
✓✓✓✓82.990.930850.76
+ +Table 5. Ablation study of TextSSR components using the TextSSR pre-training model in Tab. 1 on the IC13 dataset. + +# 4.2. Ablation Study + +To assess the proposed components, we start with the TextSSR pre-training model in Tab. 1 (to simplify, trained for only 5k steps), where all components are considered. Then, we stepwisely remove one or some of the components to assess their respective necessity and effectiveness. From the results on Tab. 5, we can see that the removal of each one brings a clear decline in performance, while the absence of "region-centric" processing can lead to a catastrophic drop of $35.4\%$ in accuracy. Additionally, fine-tuning the VAE offers a trade-off, slightly sacrificing visual quality for an improvement in accuracy. This is because real-world scene text also confronts poor visual quality challenge, simulating this to some extent can provide valuable samples for model training. + +![](images/9c36a1114f5b860501ab83aac2dbfc51a2844da35800dc8b4bd1174c36a386d3.jpg) +Figure 5. Performance of TextSSR fine-tuning model with and without region-centric processing across text regions of different sizes (in pixels). + +We perform another experiment to verify the effectiveness of the region-centric processing. The results in Fig. 5 + +show that without this processing, TextSSR has to synthesize text instance on the whole image. It exhibits a performance trend that first rises and then falls as the size increases. When the region is small, the entire image is scaled severely, resulting in more blurred text regions and insufficient pixels for text rendering. As the region size increases, larger regions and more pixels are allowed, therefore performance increase. When a too large region given, it causes the side effect that constraints the shape of the generated text, and performance decline. In contrast, incorporating region-centric processing allows TextSSR to achieve superior and consistent performance across all sizes. + +In Fig. 6, we also illustratively ablate the character-level position and glyph components. The results indicate that omitting either component leads to issues such as character deformation, incorrect characters, and duplication errors. + +![](images/ecc09037c1d9dea1144cce0c4e95e7a29a009db736e71773833247a99d693bd8.jpg) +Figure 6. Visualizations of TextSSR with and without character-level prior. More examples are provided in Appendix 7.3 + +# 5. Conclusion + +We have proposed TextSSR to provide STR model training with high-quality synthetic data. It employs the glyph of existing text and surrounding context as the prompt of the diffusion-based synthesizing model, and combines character-level constraints and permutations. TextSSR successfully synthesizes accurate and realistic text instances at a large scale. The TextSSR-F dataset with 3.55 million diverse and realistic instances are constructed. We have conducted extensive experiments to assess TextSSR-F. STR models trained solely on TextSSR-F, and on the combination of TextSSR-F and real-world data both show performance improvements. While the effectiveness of TextSSR has been basically confirmed, for irregular text, there is still an accuracy gap between models trained on TextSSR-F and real-world data. Thus, future work includes the exploration of more controllable ways to further improve the synthesizing quality, especially for challenging instances such as curved and multi-oriented text. We are also interested in extending TextSSR to generate large-scale data in different languages like Chinese. + +Acknowledgement This work was supported by the National Natural Science Foundation of China (Nos. 32341012, 62172103). + +# References + +[1] J. Baek, Y. Matsui, and K. Aizawa. What if we only use real datasets for scene text recognition? toward scene text recognition with fewer labels. In CVPR, pages 3113-3122, 2021. 6, 2 +[2] D. Bautista and R. Atienza. Scene text recognition with permuted autoregressive sequence models. In ECCV, pages 178-196, 2022. 3 +[3] H. Chen, Z. Xu, Z. Gu, J. Lan, X. Zheng, Y. Li, C. Meng, H. Zhu, and W. Wang. DiffUTE: Universal text editing diffusion model. In NeurIPS, pages 63062-63074, 2023. 3, 4 +[4] J. Chen, Y. Huang, T. Lv, L. Cui, Q. Chen, and F. Wei. TextDiffuser: Diffusion models as text painters. In NeurIPS, pages 9353-9387, 2023. 2, 3, 5, 6 +[5] J. Chen, Y. Huang, T. Lv, L. Cui, Q. Chen, and F. Wei. TextDiffuser-2: Unleashing the power of language models for text rendering. In ECCV, pages 386–402, 2024. 2, 3, 5, 6 +[6] C.K. Chng, Y. Liu, Y. Sun, C.C. Ng, C. Luo, Z. Ni, C.M. Fang, S. Zhang, J. Han, E. Ding, J. Liu, D. Karatzas, C.S. Chan, and L. Jin. ICDAR2019 robust reading challenge on arbitrary-shaped text-rc-art. In ICDAR, pages 1571–1576, 2019. 1 +[7] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4 +[8] Y. Du, Z. Chen, C. Jia, X. Yin, T. Zheng, C. Li, Y. Du, and Y. Jiang. SVTR: Scene text recognition with a single visual model. In IJCAI, pages 884-890, 2022. 3, 5 +[9] Y. Du, Z. Chen, C. Jia, X. Yin, C. Li, Y. Du, and Y.-G. Jiang. Context perception parallel decoder for scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell., 47(6):4668-4683, 2025. 1 +[10] Y. Du, Z. Chen, Y. Su, C. Jia, and Y.-G. Jiang. Instruction-guided scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell., 47(4):2723-2738, 2025. 1 +[11] Y. Du, Z. Chen, H. Xie, C. Jia, and Y. Jiang. SVTRv2: CTC beats encoder-decoder models in scene text recognition. In ICCV, 2025. 1, 3, 5, 7, 2 +[12] S. Fang, H. Xie, Y. Wang, Z. Mao, and Y. Zhang. Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition. In CVPR, pages 7098-7107, 2021. 3 +[13] S. Fang, Z. Mao, H. Xie, Y. Wang, C. Yan, and Y. Zhang. Abinet++: Autonomous, bidirectional and iterative language modeling for scene text spotting. IEEE Trans. Pattern Anal. Mach. Intell., 45(6):7123-7141, 2022. 3 +[14] J. Gu, X. Meng, G. Lu, L. Hou, N. Minzhe, X. Liang, L. Yao, R. Huang, W. Zhang, X. Jiang, C. Xu, and H. Xu. Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. In NeurIPS, pages 26418-26431, 2022. 1 + +[15] A. Gupta, A. Vedaldi, and A. Zisserman. Synthetic data for text localisation in natural images. In CVPR, pages 2315-2324, 2016. 1, 3, 6 +[16] M. He, Y. Liu, Z. Yang, S. Zhang, C. Luo, F. Gao, Q. Zheng, Y. Wang, X. Zhang, and L. Jin. ICPR2018 contest on robust reading for multi-type web images. In ICPR, pages 7-12, 2018. 1 +[17] Santoso J, Simon C, and Williem. On manipulating scene text in the wild with diffusion models. In WACV, pages 5202-5211, 2024. 3 +[18] M. Jaderberg, K. Simonyan, A. Vedaldi, and A. Zisserman. Synthetic data and artificial neural networks for natural scene text recognition. In Workshop on Deep Learning, NeurIPS, 2014. 1, 3 +[19] J. Ji, G. Zhang, Z. Wang, B. Hou, Z. Zhang, B. Price, and S. Chang. Improving diffusion models for scene text editing with dual encoders. TMLR, 2024. 3 +[20] Q. Jiang, J. Wang, D. Peng, C. Liu, and L. Jin. Revisiting scene text recognition: A data perspective. In ICCV, pages 20543-20554, 2023. 1, 3, 7, 2 +[21] D. Karatzas, F. Shafait, S. Uchida, M. Iwamura, L. Gomez i Bigorda, S. Robles Mestre, J. Mas, D. Fernandez Mota, J. Almazán, and L.P. de las Heras. ICDAR 2013 robust reading competition. In ICDAR, pages 1484-1493, 2013. 3, 5, 6, 7, 1 +[22] D. Karatzas, L. Gomez-Bigorda, A. Nicolaou, S. Ghosh, A. Bagdanov, M. Iwamura, J. Matas, L. Neumann, V.R. Chandrasekhar, S. Lu, F. Shafait, S. Uchida, and E. Valveny. IC-DAR 2015 competition on robust reading. In ICDAR, pages 1156-1160, 2015. 3, 5, 6, 7, 1 +[23] C. Li, W. Liu, R. Guo, X. Yin, K. Jiang, Y. Du, Y. Du, L. Zhu, B. Lai, X. Hu, D. Yu, and Y. Ma. PP-OCRv3: More attempts for the improvement of ultra lightweightOCR system. arXiv preprint arXiv:2206.03001, 2022. 1, 3, 2 +[24] H. Li, P. Wang, C. Shen, and G. Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In AAAI, pages 8610-8617, 2019. 3, 7 +[25] S. Long and C. Yao. Unrealtext: Synthesizing realistic scene text images from the unreal world. In CVPR, 2020. 3, 6 +[26] C. Luo, L. Jin, and Z. Sun. Moran: A multi-object rectified attention network for scene text recognition. Pattern Recognit., 90:109-118, 2019. 3 +[27] A. Mishra, K. Alahari, and CV Jawahar. Scene text recognition using higher order language priors. In BMVC, 2012. 3, 6, 7 +[28] N. Nayef, C. Liu, J. Ogier, Y. Patel, M. Busta, P.N. Chowdhury, D. Karatzas, W. Khlif, J. Matas, U. Pal, and J. Burie. ICDAR2019 robust reading challenge on multi-lingual scene text detection and recognition-rrc-plt-2019. In ICDAR, pages 1582-1587, 2019. 1 +[29] T. Q. Phan, P. Shivakumara, S. Tian, and C. L. Tan. Recognizing text with perspective distortion in natural scenes. In ICCV, pages 569-576, 2013. 3, 6, 7 +[30] Y. Qu, Q. Tan, H. Xie, J. Xu, Y. Wang, and Y. Zhang. Exploring stroke-level modifications for scene text editing. In AAAI, pages 2119-2127, 2023. 3 + +[31] M. Rang, Z. Bi, C. Liu, Y. Wang, and K. Han. An empirical study of scaling law for scene text recognition. In CVPR, pages 15619-15629, 2024. 1, 3 +[32] A. Risnumawan, P. Shivakumara, C. S. Chan, and C. L. Tan. A robust arbitrary text detection system for natural scene images. Expert Syst. Appl., 41(18):8027-8048, 2014. 3, 6, 7 +[33] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 4, 2 +[34] C. Schuhmann, R. Vencu, R. Beaumont, R. Kaczmarczyk, C. Mullis, A. Katta, T. Coombes, J. Jitsev, and A. Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1 +[35] F. Sheng, Z. Chen, and B. Xu. NRTR: A no-recurrence sequence-to-sequence model for scene text recognition. In ICDAR, pages 781-786, 2019. 3 +[36] B. Shi, X. Bai, and C. Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell., 39(11):2298-2304, 2016. 2, 6, 7 +[37] B. Shi, C. Yao, M. Liao, M. Yang, P. Xu, L. Cui, S. Belongie, S. Lu, and X. Bai. ICDAR2017 competition on reading Chinese text in the wild (rctw-17). In ICDAR, pages 1429–1434, 2017. 1 +[38] B. Shi, M. Yang, X. Wang, P. Lyu, C. Yao, and X. Bai. ASTER: An attentional scene text recognizer with flexible rectification. IEEE Trans. Pattern Anal. Mach. Intell., 41(9): 2035-2048, 2018. 3 +[39] A. Singh, G. Pang, M. Toh, J. Huang, W. Galuba, and T. Hassner. TextOCR: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. In CVPR, pages 8802-8812, 2021. 1, 3, 6, 2 +[40] Y. Sun, Z. Ni, C.K. Chng, Y. Liu, C. Luo, C.C. Ng, J. Han, E. Ding, J. Liu, D. Karatzas, C.S. Chan, and L. Jin. ICDAR 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In ICDAR, pages 1557-1562, 2019. 1 +[41] Y. Tuo, W. Xiang, J. Y. He, Y. Geng, and X. Xie. AnyText: Multilingual visual text generation and editing. In ICLR, 2024. 2, 3, 4, 5, 6, 1 +[42] A. Veit, T. Matera, L. Neumann, J. Matas, and S. Belongie. COCO-Text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 1 +[43] K. Wang, B. Babenko, and S. Belongie. End-to-end scene text recognition. In ICCV, pages 1457-1464, 2011. 3, 6, 7 +[44] Y. Wang, H. Xie, S. Fang, M. Xing, J. Wang, S. Zhu, and Y. Zhang. Petr: Rethinking the capability of transformer-based language model in scene text recognition. IEEE Trans. Image Process., 31:5585-5598, 2022. 3 +[45] Y. Wang, W. Zhang, H. Xu, and C. Jin. Dreamtext: High fidelity scene text synthesis. In CVPR, pages 28555-28563, 2025. 3, 5 +[46] J. Wei, H. Zhan, Y. Lu, X. Tu, B. Yin, C. Liu, and U. Pal. Image as a language: Revisiting scene text recognition via balanced, unified and synchronized vision-language reasoning network. In AAAI, pages 5885-5893, 2024. 3 + +[47] L. Wu, C. Zhang, J. Liu, J. Han, J. Liu, E. Ding, and X. Bai. Editing text in the wild. In ACM MM, pages 1500-1508, 2019. 3 +[48] J. Xu, Y. Wang, H. Xie, and Y. Zhang. OTE: Exploring accurate scene text recognition using one token. In CVPR, pages 28327-28336, 2024. 3 +[49] Q. Yang, J. Huang, and W. Lin. SwapText: Image based texts transfer in scenes. In CVPR, pages 14700-14709, 2020. 3 +[50] Y. Yang, D. Gui, Y. Yuan, W. Liang, H. Ding, H. Hu, and K. Chen. GlyphControl: Glyph conditional control for visual text generation. In NeurIPS, pages 44050-44066, 2023. 2, 3, 4, 5, 6 +[51] W. Zeng, Y. Shu, Z. Li, D. Yang, and Y. Zhou. TextCtrl: Diffusion-based scene text editing with prior guidance control. In NeurlPS, pages 138569–138594, 2024. 3 +[52] C. Zhai, Z. Chen, J. Li, and B. Xu. Chinese image text recognition with blastm-ctc: a segmentation-free method. In CCPR, pages 525–536, 2016. 2 +[53] F. Zhan, S. Lu, and C. Xue. Verisimilar image synthesis for accurate detection and recognition of texts in scenes. In ECCV, pages 249-266, 2018. 6 +[54] C. Zhang, W. Ding, G. Peng, F. Fu, and W. Wang. Street view text recognition with deep learning for urban scene understanding in intelligent transportation systems. IEEE Trans. Intell. Transp., 22(7):4727-4743, 2021. 5, 1 +[55] R. Zhang, Y. Zhou, Q. Jiang, Q. Song, N. Li, K. Zhou, L. Wang, D. Wang, M. Liao, M. Yang, X. Bai, B. Shi, D. Karatzas, S. Lu, and C.V. Jawahar. ICDAR 2019 robust reading challenge on reading chinese text on signboard. In IC-DAR, pages 1577-1581, 2019. 1 +[56] Z. Zhang, N. Lu, M. Liao, Y. Huang, C. Li, M. Wang, and W. Peng. Self-distillation regularized connectionist temporal classification loss for text recognition: A simple yet effective approach. In AAAI, pages 7441-7449, 2024. 3 +[57] Y. Zhao and Z. Lian. UDiffText: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In ECCV, pages 217-233, 2024. 2, 3, 5, 7 +[58] T. Zheng, Z. Chen, S. Fang, H. Xie, and Y.-G. Jiang. CDistNet: Perceiving multi-domain character distance for robust text recognition. Int. J. Comput. Vis., 132(2):300-318, 2024. 3 +[59] Y. Zhu, Z. Li, T. Wang, M. He, and C. Yao. Conditional text image generation with diffusion models. In CVPR, pages 14235-14245, 2023. 3 +[60] Y. Zhu, J. Liu, F. Gao, W. Liu, X. Wang, P. Wang, F. Huang, C. Yao, and Z. Yang. Visual text generation in the wild. In ECCV, pages 89-106, 2024. 3, 6 \ No newline at end of file diff --git a/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/images.zip b/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a8f70882f90afa58ebb5d69a875fff18a2c35639 --- /dev/null +++ b/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae42c5bd75fca4acdf21ca0ebe9f71d713b4cc2d5a630f54a5cfe80ba19a0d70 +size 653000 diff --git a/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/layout.json b/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7e64f9304aebaa32ced81161dc42de594e6cb577 --- /dev/null +++ b/2025/TextSSR_ Diffusion-based Data Synthesis for Scene Text Recognition/layout.json @@ -0,0 +1,10206 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 516, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 516, + 121 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 516, + 121 + ], + "type": "text", + "content": "TextSSR: Diffusion-based Data Synthesis for Scene Text Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 142, + 451, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 142, + 451, + 157 + ], + "spans": [ + { + "bbox": [ + 162, + 142, + 451, + 157 + ], + "type": "text", + "content": "Xingsong Ye" + }, + { + "bbox": [ + 162, + 142, + 451, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 142, + 451, + 157 + ], + "type": "text", + "content": ", Yongkun Du" + }, + { + "bbox": [ + 162, + 142, + 451, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 142, + 451, + 157 + ], + "type": "text", + "content": ", Yunbo Tao" + }, + { + "bbox": [ + 162, + 142, + 451, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 142, + 451, + 157 + ], + "type": "text", + "content": ", Zhineng Chen" + }, + { + "bbox": [ + 162, + 142, + 451, + 157 + ], + "type": "inline_equation", + "content": "^{2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 157, + 503, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 157, + 503, + 185 + ], + "spans": [ + { + "bbox": [ + 107, + 157, + 503, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 157, + 503, + 185 + ], + "type": "text", + "content": "College of Computer Science and Artificial Intelligence, Fudan University, China " + }, + { + "bbox": [ + 107, + 157, + 503, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 157, + 503, + 185 + ], + "type": "text", + "content": "Institute of Trustworthy Embodied AI, Fudan University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 186, + 476, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 186, + 476, + 199 + ], + "spans": [ + { + "bbox": [ + 126, + 186, + 476, + 199 + ], + "type": "text", + "content": "{xsye20，zhinchen}@fudan.edu.cn，{ykdu23，ybtao24} " + }, + { + "bbox": [ + 126, + 186, + 476, + 199 + ], + "type": "inline_equation", + "content": "\\text{已}" + }, + { + "bbox": [ + 126, + 186, + 476, + 199 + ], + "type": "text", + "content": " m.fudan.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 253, + 297, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 253, + 297, + 577 + ], + "spans": [ + { + "bbox": [ + 54, + 253, + 297, + 577 + ], + "type": "text", + "content": "Scene text recognition (STR) suffers from challenges of either less realistic synthetic training data or the difficulty of collecting sufficient high-quality real-world data, limiting the effectiveness of trained models. Meanwhile, despite producing holistically appealing text images, diffusion-based visual text generation methods struggle to synthesize accurate and realistic instance-level text at scale. To tackle this, we introduce TextSSR: a novel pipeline for Synthesizing Scene Text Recognition training data. TextSSR targets three key synthesizing characteristics: accuracy, realism, and scalability. It achieves accuracy through a proposed region-centric text generation with position-glyph enhancement, ensuring proper character placement. It maintains realism by guiding style and appearance generation using contextual hints from surrounding text or background. This character-aware diffusion architecture enjoys precise character-level control and semantic coherence preservation, without relying on natural language prompts. Therefore, TextSSR supports large-scale generation through combinatorial text permutations. Based on these, we present TextSSR-F, a dataset of 3.55 million quality-screened text instances. Extensive experiments show that STR models trained on TextSSR-F outperform those trained on existing synthetic datasets by clear margins on common benchmarks, and further improvements are observed when mixed with real-world training data. Code is available at https://github.com/YesianRohn/TextSSR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 135, + 613 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 621, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 296, + 693 + ], + "type": "text", + "content": "Scene text images are a unique type of image captured in the wild, they differ from general images in the presence of text, which appears with diverse sizes, fonts, and often with varying degrees of distortion and occlusion. However, these texts usually contain high-level semantic information closely related to the captured scene, making STR a crucial" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 227, + 549, + 350 + ], + "blocks": [ + { + "bbox": [ + 316, + 227, + 549, + 350 + ], + "lines": [ + { + "bbox": [ + 316, + 227, + 549, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 549, + 350 + ], + "type": "image", + "image_path": "9ddccf2735e725e1dcbc87411ec22ecc338ae80305ad8baaa25a144930e5067c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 359, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 426 + ], + "type": "text", + "content": "Figure 1. Top: Text instances detected by an end-to-end STR model [23], and synthesized by our TextSSR based on the detected examples. Bottom: Accuracy of MAERec [20] across common benchmarks. Peach indicates training on ST [15] and TextOCR [39], while blue denotes that TextSSR-F is additionally included as training data." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 544, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 544, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 544, + 461 + ], + "type": "text", + "content": "task for complex scene understanding, image search, etc." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 462, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 462, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 462, + 555, + 715 + ], + "type": "text", + "content": "Recent STR practices [9, 10, 20] have shown that besides advanced modeling techniques, high-quality training data also plays a critical role. Training STR models also follows the scaling law from the data perspective [31] and current data size is far from saturated even for English. Enlarging the size of real-world datasets is a feasible way to further enhance recognition performance. However, it is challenging to collect real-world scene text images on an even larger scale. On one hand, diversified data collection is costly and time-consuming. On the other hand, real-world scenes predominantly contain high-frequency, semantically meaningful words, while low-frequency words are difficult to collect. Synthesizing high-quality text images seems to be an effective alternative. However, research on this direction is somewhat laggard. The two most popular synthetic datasets, MJ [18] and ST [15], are constructed based on traditional rendering techniques nearly a decade ago. Although they can be easily scaled up in quantity, recent studies [9, 11] indicate that there are significant accuracy gaps between models trained on MJ&SJ and Union14M-L [20], mainly because MJ&SJ fail to comprehensively describing" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17464" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 225, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 225, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 225, + 83 + ], + "type": "text", + "content": "the challenge of real-world text instances." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 296, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 296, + 263 + ], + "type": "text", + "content": "Recently, diffusion models have shown impressive performance in generating text-rich images [4, 5, 41, 50, 57]. However, most of them focus on generating aesthetic text images holistically, e.g., posters. Using data generated by these methods for STR model training encounters several issues. First, these methods are designed for the full image level generation, when switching to synthesizing text in instance-level regions, it would be less effective and ensuring the generation accuracy becomes challenging. Second, current generation methods mainly rely on well-refined natural language as prompts. They produce overly homogeneous and unrealistic text instances with some probability, which is not suitable as training data. Moreover, constructing a diverse set of prompts is inherently difficult, which hinders the ability to generate data at scale." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 265, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 335 + ], + "type": "text", + "content": "Nevertheless, given the exceptional performance of diffusion models across various tasks, it is anticipated that there are diffusion-based techniques for producing satisfactory training data for STR. Before proceeding, we first outline three key characteristics that an improved data synthesis method should exhibit:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 339, + 295, + 543 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 339, + 295, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 295, + 398 + ], + "type": "text", + "content": "- Accurate: The synthesized text content should match the given text, ensuring that the synthesized words are readable and free from character misplacement or duplication errors. This is fundamental for maintaining the correctness of the trained STR models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 483 + ], + "type": "text", + "content": "- Realistic: The generated text instances should visually resemble real-world text, simulating scene conditions in the wild. This includes being harmoniously blended into the background, ensuring consistent generation at different sizes and capturing the diversity of text presentation, especially in complex scenes. Achieving realism is essential for enhancing the performance of text recognizers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 543 + ], + "type": "text", + "content": "- Scalable: The method should support extensive training data production. In other words, abandon complex control inputs and design processes, and enable large-scale data generation by leveraging easily accessible and processable resources only." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "To satisfy these characteristics, we propose TextSSR, a novel diffusion-driven pipeline focusing on Synthesizing training data for Scene Text Recognition. Our approach introduces the following innovations to overcome persistent limitations in text synthesis. First, we implement a region-centric text generation. It employs end-to-end OCR model to acquire text location and size. Moreover, we devise a position-glyph enhancement that further resolves character duplication and positional disorder, ensuring precise text placement and arrangement. Second, we leverage the detected text or only background near the designated region as contextual hints for font style and appearance generation, guiding the generated text to match real-world scenarios better. Third, we design a character-aware diffusion" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 71, + 542, + 202 + ], + "blocks": [ + { + "bbox": [ + 328, + 71, + 542, + 202 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 542, + 202 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 542, + 202 + ], + "type": "image", + "image_path": "4a54ed79f77e42a1a50dc46f1b833ebbbbc2fb8662c7056c6883d837d83a4661.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 211, + 553, + 233 + ], + "lines": [ + { + "bbox": [ + 313, + 211, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 553, + 233 + ], + "type": "text", + "content": "Figure 2. Showcases of TextSSR's capacity to synthesize accurate, realistic and scalable text instances." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 255, + 555, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 362 + ], + "type": "text", + "content": "architecture that employs glyph priors rather than natural language prompts as injected conditions. It enables precise character-level control while maintaining full-word semantic coherence. To expand the generation quantity, we use detected text to generate a series of anagrams by rearranging characters within the original bounding boxes, thus supporting large-scale generation. As a result, TextSSR achieves accurate, realistic, and scalable text instance synthesize, as illustrated in Fig. 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 363, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 434 + ], + "type": "text", + "content": "Through these attempts, we construct the TextSSR-F dataset with 3.55 million quality-screened samples, which is automatically validated by checking their recognitions and pseudo labels. Extensive experiments demonstrate the superiority of TextSSR-F over existing synthesis data, as well as its complementarity with real-world training data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 325, + 434, + 511, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 434, + 511, + 445 + ], + "spans": [ + { + "bbox": [ + 325, + 434, + 511, + 445 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 449, + 553, + 628 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 449, + 553, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 449, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 314, + 449, + 553, + 496 + ], + "type": "text", + "content": "- We propose TextSSR, a diffusion-based text synthesis method centered on seamlessly blending text instances into real scenes. It develops a region-centric processing that enables accurate instance-level generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 497, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 497, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 314, + 497, + 553, + 555 + ], + "type": "text", + "content": "- TextSSR fully exploits positional and glyph priors of text, achieving accurate, diverse and realistic text synthesis by using only the detected text as input. Moreover, it can be easily scaled up and we construct the TextSSR-F dataset with 3.55 million text instances." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 555, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 555, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 314, + 555, + 553, + 628 + ], + "type": "text", + "content": "- Extensive evaluations validate that STR models trained on TextSSR-F surpass their counterparts trained on existing synthetic data by " + }, + { + "bbox": [ + 314, + 555, + 553, + 628 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 314, + 555, + 553, + 628 + ], + "type": "text", + "content": " on common benchmarks, and even better STR models can be obtained when further combined with real-world training data, highlighting the utility of the proposed TextSSR." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 640, + 400, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 640, + 400, + 653 + ], + "spans": [ + { + "bbox": [ + 314, + 640, + 400, + 653 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 660, + 446, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 660, + 446, + 673 + ], + "spans": [ + { + "bbox": [ + 314, + 660, + 446, + 673 + ], + "type": "text", + "content": "2.1. Scene Text Recognition" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "STR researches can be categorized from both the model and the data perspectives. From a model point of view, early approaches employ the CTC-based pipeline [36, 52]." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17465" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 203 + ], + "type": "text", + "content": "They used CNN for visual feature extraction, followed by RNN for sequence modeling, then with CTC loss applied as a constraint. Subsequently, attention mechanisms gradually gained attention in the STR field [24, 26, 35, 38]. Inspired by developments in the NLP field, many methods [2, 12, 13, 44, 46, 48, 58] further refined recognizers from a language modeling perspective. While some approaches [8, 11, 56] continued to explore solutions based on a single vision model and CTC decoding, due to their fast inference nature. Accuracy was steadily improved following these efforts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 384 + ], + "type": "text", + "content": "From a data perspective, the limited availability of well-annotated real-world training data led to a practical STR training and evaluation protocol: training solely on synthetic data, or a combination of synthetic and real data, followed by evaluation on real-world benchmarks. Besides the six common benchmarks [21, 22, 27, 29, 32, 43]. Union14M-L [20] aggregated publicly available real-world datasets (including TextOCR [39]) and created a real-world training dataset with nearly 4 million instances. However, recent studies on the OCR scaling law [31] suggested that even this substantial volume had not reached the saturation point. This observation implies that STR models can still benefit from additional training data, and high-quality synthetic data could be a feasible supplement, which motivates our study." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 396, + 175, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 175, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 175, + 407 + ], + "type": "text", + "content": "2.2. Scene Text Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 713 + ], + "type": "text", + "content": "Early synthesis methods [15, 18, 25] employed rule-based methods to overlay text onto background with transformations simulating real-world appearance. To be more realistic, later research shifted toward deep learning-based generation methods. Among these, GAN-based scene text editing [30, 47, 49] methods aimed to expand data by editing text directly within specified regions, but they were limited by the lack of real paired training data. Recent advances had leveraged diffusion models for both editing [17, 19, 51] and synthesis, with fine-grained control over glyphs [3, 50], as well as positioning at word [41] and character [4, 45, 57] levels. The latest methods [5, 60] incorporate large language or multi-modal models to enhance performance. However, these approaches primarily targeted visually appealing generation for aesthetic purposes, rather than generating data for better STR models. While some work has explored diffusion-based text image synthesis. For example, CTIG-DM [59] employed character sequence as the condition for generation. However, it is less controllable and faced with the dilemma of diversity-realism tradeoff, limiting quality of the generated text. SceneVTG [60] developed a first-erasing-then-synthesizing pipeline. It requires a pre-erased model to remove existing text completely, which is difficult to guarantee in large-scale synthesizing scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 553, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 210 + ], + "type": "text", + "content": "To achieve accurate, realistic, and scalable text instance generation, TextSSR proposes a four-step pipeline (detailed in Fig. 3(b)). To handle text instances with varying sizes and locations, we first focus on text region processing, which will be discussed in Sec. 3.1. To enhance accuracy through fine-grained extraction and decoupling of text glyphs and positions, Sec. 3.2 and Sec. 3.3 will detail the specific module design and training strategy, respectively. Sec. 3.4 introduces our data extension method, achieving scalability in data synthesis at scale." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 217, + 481, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 481, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 481, + 230 + ], + "type": "text", + "content": "3.1. Text Region-centric Processing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 235, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 553, + 378 + ], + "type": "text", + "content": "Instance-level text generation presents significant challenges, as we must handle varying sizes, diverse potential positions, and determine how to select appropriate surrounding information for effective prompting. Synthesis methods that edit the entire image are not suitable for this task. Such approaches create an imbalance, and small text regions lack sufficient pixels for accurate reconstruction. Furthermore, information from the entire scene is often redundant for text region synthesis. For example, in Fig. 3(a), the text \"kills\" to be synthesized shares a similar style and context with the nearby upper text and background, while the region farther away is inconsistent in appearance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 380, + 553, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 553, + 521 + ], + "type": "text", + "content": "Therefore, we propose a generative scheme based on a locality assumption—text styles are more likely to correlate with adjacent text and background information. This method focuses generation at the region level, synthesizing the text using only surrounding local context rather than the entire scene. Specifically, we use the annotated text box to extract the smallest enclosing square around the text region and resize it to a fixed size (e.g., 256). For practical considerations, we uniformly employ an end-to-end OCR tool [23] to obtain text instance size, position, and textual labels. It well reflects real-world scenarios where annotations are often unavailable." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 529, + 508, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 508, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 508, + 541 + ], + "type": "text", + "content": "3.2. Fine-grained Condition Mechanisms" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 629 + ], + "type": "text", + "content": "Visual text rendering needs to consider two key fine-grained features: glyph structure and positional arrangement, both at word (holistic) and character (individual) levels. Word-level features concern entire word appearance and region placement, while the character-level address individual characters and their relative ordering. To guide the generative process, we implement the following approach:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": "Word-level Position and Glyph Guidance. For a text " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " with position " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "P_{t}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " in the original image, we extract a square region " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "P_{t}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " becomes " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "P_{t}^{\\prime}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " (Fig. 3(a), red dashed rectangle). This allows us to establish our desired word-level position guidance. Then we create a binary mask " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "I_{M}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " that separates text (0) from non-text (1) regions. The size of " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "I_{M}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": " is identical to " + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 313, + 630, + 553, + 713 + ], + "type": "text", + "content": ", indicating where the text is present. For" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17466" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 319, + 192 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 319, + 192 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 319, + 192 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 319, + 192 + ], + "type": "image", + "image_path": "25fcfe9f298a6c509b7f84f45e39131128ff186b82647f258cc5ba781efdb548.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 194, + 223, + 203 + ], + "lines": [ + { + "bbox": [ + 163, + 194, + 223, + 203 + ], + "spans": [ + { + "bbox": [ + 163, + 194, + 223, + 203 + ], + "type": "text", + "content": "(a) Generative Model" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 320, + 70, + 541, + 192 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 541, + 192 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 541, + 192 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 541, + 192 + ], + "type": "image", + "image_path": "03062fe75ce00b17cd5082df12f1839f15822ac1dee81edd643d2fe03f880ab9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 194, + 460, + 204 + ], + "lines": [ + { + "bbox": [ + 399, + 194, + 460, + 204 + ], + "spans": [ + { + "bbox": [ + 399, + 194, + 460, + 204 + ], + "type": "text", + "content": "(b) Synthesis Pipeline" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 217, + 555, + 284 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 555, + 284 + ], + "type": "text", + "content": "Figure 3. Overall architecture of TextSSR's (a) generative model and (b) synthesis pipeline. (a) Given a scene text instance (marked in red) and its content (e.g., \"kills\"), it is preprocessed into the smallest local image containing the text. A text-size mask is then applied to identify non-text regions, while the text content is rendered as priors at both word and character levels. These different conditions are subsequently injected into a diffusion model as training guidance. (b) An end-to-end OCR model extracts the text location and content within the image (Step 1). All possible text words are generated through anagram (Step 2). The generative model then synthesizes text instances in batches (Step 3), and finally, a SOTA STR model filters the samples, retaining only the correctly recognized instances for use (Step 4)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 304, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 376 + ], + "type": "text", + "content": "glyph guidance, we render the text in a standard font, applying an affine transformation to align it with the position and orientation of " + }, + { + "bbox": [ + 55, + 304, + 295, + 376 + ], + "type": "inline_equation", + "content": "P_{t}^{\\prime}" + }, + { + "bbox": [ + 55, + 304, + 295, + 376 + ], + "type": "text", + "content": " (resulting in " + }, + { + "bbox": [ + 55, + 304, + 295, + 376 + ], + "type": "inline_equation", + "content": "I_{G}" + }, + { + "bbox": [ + 55, + 304, + 295, + 376 + ], + "type": "text", + "content": "). As noted in previous work [3, 41, 50], when we close the door on the original styled text, we must at least provide a window with a basic visual style of the text at that location." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": "Character-level Position and Glyph Guidance. Only word-level information limits the model's understanding to the words, as we observe issues such as repeated or incorrect characters when using only word-level features. We address this by encoding character-level ordering through pixel values, enabling the model to learn spatial relationships between characters. Specifically, we define a maximum character length " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": ". For characters from position 1 to " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "l \\leq L" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": ", with truncation if " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "l > L" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": " and zero-padding in channels " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "l + 1" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "l < L" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": "), the " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": "-th character is rendered into the " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": "-th channel of an " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": "-channel image " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "I_g" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": ". Each character glyph is rendered with a pixel intensity of " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "i \\times \\left\\lfloor \\frac{P_M}{L} \\right\\rfloor" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "P_M" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": " denotes the maximum number of pixels allocated to distinguish a character from the background. We represent the glyph image tensor as " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "I_g \\in \\mathbb{R}^{L \\times S \\times S}" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": " is the number of characters and " + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 376, + 295, + 604 + ], + "type": "text", + "content": " is the image size for each character. This creates position-prior pixel values that serve as positional encodings to help the model distinguish character information at each specific position." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "text", + "content": "Vision Transformer (ViT) [7] is employed as the Glyph Encoder " + }, + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "inline_equation", + "content": "G_{\\phi}" + }, + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "text", + "content": " to accommodate input " + }, + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "inline_equation", + "content": "I_{g}" + }, + { + "bbox": [ + 55, + 605, + 295, + 642 + ], + "type": "text", + "content": " and output features compatible with diffusion model's condition inputs:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 118, + 651, + 295, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 651, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 118, + 651, + 295, + 667 + ], + "type": "interline_equation", + "content": "F _ {g} = G _ {\\phi} \\left(I _ {g}\\right) \\in \\mathbb {R} ^ {(N + 1) \\times D} \\tag {1}", + "image_path": "046b0c1f3643ecf51159da33c7f19845653c07425daddcf1dd86071e7d63da0c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": " represents the number of patches, the additional one corresponds to a class token, and dimension " + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 677, + 295, + 714 + ], + "type": "text", + "content": " is set to match the input requirements of the diffusion process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 304, + 471, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 471, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 471, + 316 + ], + "type": "text", + "content": "3.3. Adapting to Stable Diffusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 357 + ], + "type": "text", + "content": "We leverage a pretrained model and fine-tune its Variational Auto Encoder (VAE) and retrain the Conditional Diffusion Model (CDM) components for our specific task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": "VAE Fine-tuning. The standard VAE, when pretrained on general datasets, may not be sensitive enough to capture the fine-grained spatial and glyph features of scaled text instances. As highlighted by DiffUTE [3] and also confirmed by our ablation experiments, fine-tuning the VAE is essential when working with text images. Therefore, we fine-tune it to enhance the capacity for accurate representation of text regions in a fair, local manner. The VAE consists of an encoder " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": ", and we let " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "V_{\\theta}" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": " represent the reconstruction after the encoding and decoding process. Following the pre-defined " + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 313, + 357, + 555, + 501 + ], + "type": "text", + "content": ", we minimize the following loss function:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 370, + 510, + 555, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 510, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 370, + 510, + 555, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {T e x t S S R - V A E}} = \\left\\| V _ {\\theta} \\left(I _ {s}\\right) - I _ {s} \\right\\| _ {2} ^ {2} \\tag {2}", + "image_path": "a1255682fd9bf99fec173b7f9ced9b219f07fe603242474c4a2d9b6de3c4fe53.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": "CDM Retraining. To mask the actual text instance region while obtaining surrounding contextual information, we apply " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "I_{M}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": ", through the operation " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "I_{m} = I_{M} \\cdot I_{s}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "I_{m}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "I_{G}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "I_{s}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": " are fed into the frozen VAE, generating latent representations " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "Z_{m}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "Z_{G}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "Z_{s}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": ". Additionally, " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "Z_{s}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": " is further processed with added noise to produce " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "Z_{T}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": ". Moreover, we downsample " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "I_{M}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": " to match their dimensions (e.g., in SD 2.1 [33], height = width = 16), transforming it into " + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "inline_equation", + "content": "Z_{M}" + }, + { + "bbox": [ + 313, + 533, + 555, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "These latent features, " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "Z_{M}, Z_{T}, Z_{m}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "Z_{G}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", are concatenated and then passed through a Conv2D layer to match the channel dimension required by the U-Net (e.g., in SD 2.1, channel = 4). Meanwhile, the Glyph Encoder (see Fig. 3(a)) that produces " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "F_{g}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " is also integrated into the CDM to extract features from " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "I_{g}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ". Finally, the CDM processes" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17467" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 535, + 194 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 535, + 194 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 535, + 194 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 535, + 194 + ], + "type": "table", + "html": "
MethodIC13 [21]IC15 [22]ShopSign [54]
SeqAcc(%)↑NED ↑FID-R↓SeqAcc(%)↑NED ↑FID-R↓SeqAcc(%)↑NED ↑FID-R↓
TextDiffuser [4]25.080.4779103.740.820.0551120.660.640.0230136.56
GlyphControl [50]0.550.0472142.9100.0217152.3700.0067140.91
UDiffText [57]37.510.589267.822.550.1206103.43———
AnyText [41]55.070.715375.465.970.218676.0333.580.460289.25
TextDiffuser-2 [5]26.390.534268.310.870.083882.930.820.0411101.46
TextSSR Pre-training75.350.900751.7550.460.736134.6930.280.507554.24
TextSSR Fine-tuning85.170.940253.2862.640.808736.3052.240.705352.84
Real96.510.9895084.880.9486074.840.86990
", + "image_path": "93767311cc114282c739edbee2381dd79b84dc47c7be62602fff452f54daf97b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 201, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 54, + 201, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 54, + 201, + 555, + 246 + ], + "type": "text", + "content": "Table 1. Comparing TextSSR (differences between the pre-training and fine-tuning versions are detailed in Appendix 6.3.1. We use the fine-tuned version, abbreviated as TextSSR, in the rest part of the paper.) and existing synthesis methods on regular text (IC13), irregular text (IC15), and multi-language text (ShopSign) datasets, alongside results from real datasets (Real). SeqAcc and NED denote the recognition accuracy, editing distance, respectively. FID-R score means FID on the text region." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 266, + 294, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 294, + 339 + ], + "type": "text", + "content": "these inputs and produces the denoised latent representation " + }, + { + "bbox": [ + 55, + 266, + 294, + 339 + ], + "type": "inline_equation", + "content": "Z_{T-1}" + }, + { + "bbox": [ + 55, + 266, + 294, + 339 + ], + "type": "text", + "content": ", progressively refining the noisy input to recover the clean latent space. Assuming " + }, + { + "bbox": [ + 55, + 266, + 294, + 339 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 266, + 294, + 339 + ], + "type": "text", + "content": " represents the original added noise and " + }, + { + "bbox": [ + 55, + 266, + 294, + 339 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 266, + 294, + 339 + ], + "type": "text", + "content": " denotes the entire denoising process that outputs the predicted noise, the modified CDM is trained by minimizing the following objective:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 343, + 294, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 343, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 69, + 343, + 294, + 357 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {T e x t S S R - C D M}} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, Z _ {M}, Z _ {m}, Z _ {G}, I _ {g}\\right) \\right\\| _ {2} ^ {2} \\tag {3}", + "image_path": "1b3da7dbafb48e40846677b22ab4b55c98363bcd0751387253f34d4f043f00f3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 226, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 226, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 226, + 374 + ], + "type": "text", + "content": "3.4. Anagram-based Data Synthesis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 379, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 379, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 379, + 295, + 594 + ], + "type": "text", + "content": "As mentioned above, our pipeline uses text region and content detected by an OCR engine as inputs. The recognized content is regarded as pseudo-label of the text. Our generation produces a synthesized text instance on the detected region for each pseudo-label rather than the groundtruth, which is typically unknown. Therefore, errors in OCR would not affect the generation. To expand the generation in quantity, we propose an anagram-based data expansion strategy. Unlike previous approaches that arbitrarily edit text [4, 5, 41] or substitute equal-length strings [45, 57], we manipulate internal character order to maintain better region-content coherence. Meanwhile, the diffusion process adaptively handles the varying occupation requirement of different characters, e.g., 'o' typically is wider than 'l' in width. Therefore, our TextSSR leads to a more flexible generation. By changing the internal character order only, the region remain unchanged while still generating many samples with different character arrangements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 594, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 642 + ], + "type": "text", + "content": "Our model's character-level rendering capability enables correct generation of text from this permutation-based operation. We define the total number of possible permutations for a word of length " + }, + { + "bbox": [ + 54, + 594, + 295, + 642 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 54, + 594, + 295, + 642 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 155, + 647, + 295, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 647, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 155, + 647, + 295, + 661 + ], + "type": "interline_equation", + "content": "P (l) = l! \\tag {4}", + "image_path": "67d234624e6fe799029a5b9526859b7cf37336bb7594d1f85f5f7b1733b5405b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Note that some words may repeat due to character duplication; they still represent different samples, as the diffusion model is inherently random and there are also variations in cropped regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 265, + 395, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 265, + 395, + 279 + ], + "spans": [ + { + "bbox": [ + 314, + 265, + 395, + 279 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 285, + 457, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 457, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 457, + 298 + ], + "type": "text", + "content": "4.1. TextSSR Data Evaluation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 304, + 554, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 364 + ], + "type": "text", + "content": "We first evaluate both accuracy and realism of the generation text, as well as the scaling capabilities of TextSSR data in small-scale. Then, we construct large-scale training data and assess its effectiveness at larger scales. Fig. 4 visualizes the examples of TextSSR compared to other methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 472 + ], + "type": "text", + "content": "Accuracy Evaluation. Evaluating the accuracy of generated scene text is challenging, as manually verifying alignment with target text is costly. Previous benchmarks typically employ an STR model to automatically assess text consistency. We also follow this scheme. To mitigate potential OCR recognition errors and more effectively reflect the accuracy and difficulty level of our synthetic data, we choose SVTRv2 [11], a SOTA STR model upgraded from SVTR [8], as the recognition model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 474, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 592 + ], + "type": "text", + "content": "Tab. 1 gives the accuracy assessment results across different datasets, where IC13 denotes all the methods using text instances from the IC13 dataset [21] to initialize their generation, while others are defined similarly. As can be seen, a large portion of text instances generated by TextSSR are correctly recognized. For more irregular and difficult IC15 [22], and less linguistically ShopSign [54] datasets, despite accuracy declines observed, TextSSR still outperforms the others by significant margins. The results provide strong evidence of TextSSR's correctness." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": "Comparing with AnyText [41], the current SOTA, TextSSR outperforms it by " + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "inline_equation", + "content": "30.1\\%" + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "inline_equation", + "content": "56.7\\%" + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": " in accuracy on IC13 and IC15, respectively. This is mainly because TextSSR operates generation at the instance level, while AnyText performs on the full image level. Note that methods based on full-image synthesis struggle significantly in complex, challenging scenarios, e.g., IC15. This dataset includes many text instances affected by blur, low-resolution, perspective distortion, and curvature—challenges that previous methods do not address ef" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17468" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 86, + 160, + 125 + ], + "blocks": [ + { + "bbox": [ + 108, + 73, + 156, + 83 + ], + "lines": [ + { + "bbox": [ + 108, + 73, + 156, + 83 + ], + "spans": [ + { + "bbox": [ + 108, + 73, + 156, + 83 + ], + "type": "text", + "content": "GroundTruth" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 107, + 86, + 160, + 125 + ], + "lines": [ + { + "bbox": [ + 107, + 86, + 160, + 125 + ], + "spans": [ + { + "bbox": [ + 107, + 86, + 160, + 125 + ], + "type": "image", + "image_path": "9121d23c28265570b1251692e433996e94cf69452c05f49b2667e846a80332c3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 108, + 127, + 160, + 148 + ], + "blocks": [ + { + "bbox": [ + 108, + 127, + 160, + 148 + ], + "lines": [ + { + "bbox": [ + 108, + 127, + 160, + 148 + ], + "spans": [ + { + "bbox": [ + 108, + 127, + 160, + 148 + ], + "type": "image", + "image_path": "7974814931c525ae5a2445b1c8ee63bbca8dbc5d799b35bed606e7c7c31557d0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 165, + 86, + 217, + 125 + ], + "blocks": [ + { + "bbox": [ + 167, + 74, + 213, + 83 + ], + "lines": [ + { + "bbox": [ + 167, + 74, + 213, + 83 + ], + "spans": [ + { + "bbox": [ + 167, + 74, + 213, + 83 + ], + "type": "text", + "content": "TextDiffuser" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 165, + 86, + 217, + 125 + ], + "lines": [ + { + "bbox": [ + 165, + 86, + 217, + 125 + ], + "spans": [ + { + "bbox": [ + 165, + 86, + 217, + 125 + ], + "type": "image", + "image_path": "69a9d689acda4cc4721086811c89418badbf764ebf6733cc5ad5a2d5121ba20d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 165, + 126, + 217, + 148 + ], + "blocks": [ + { + "bbox": [ + 165, + 126, + 217, + 148 + ], + "lines": [ + { + "bbox": [ + 165, + 126, + 217, + 148 + ], + "spans": [ + { + "bbox": [ + 165, + 126, + 217, + 148 + ], + "type": "image", + "image_path": "d497c04336417c192bd0cdd89a6796db7ce379891128d5d4e7c663747b8906ca.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 224, + 86, + 274, + 125 + ], + "blocks": [ + { + "bbox": [ + 225, + 74, + 271, + 83 + ], + "lines": [ + { + "bbox": [ + 225, + 74, + 271, + 83 + ], + "spans": [ + { + "bbox": [ + 225, + 74, + 271, + 83 + ], + "type": "text", + "content": "GlyphControl" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 224, + 86, + 274, + 125 + ], + "lines": [ + { + "bbox": [ + 224, + 86, + 274, + 125 + ], + "spans": [ + { + "bbox": [ + 224, + 86, + 274, + 125 + ], + "type": "image", + "image_path": "e7b4110a01d99aa29900f4def875d7263a84c6c79de7ad05fb8f6b9aeabfda2d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 224, + 127, + 274, + 148 + ], + "blocks": [ + { + "bbox": [ + 224, + 127, + 274, + 148 + ], + "lines": [ + { + "bbox": [ + 224, + 127, + 274, + 148 + ], + "spans": [ + { + "bbox": [ + 224, + 127, + 274, + 148 + ], + "type": "image", + "image_path": "87268eb510134ba42b6dcaa7e96227b170abb248b68d56107c20e92e48bd2e78.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 279, + 86, + 332, + 125 + ], + "blocks": [ + { + "bbox": [ + 288, + 74, + 321, + 83 + ], + "lines": [ + { + "bbox": [ + 288, + 74, + 321, + 83 + ], + "spans": [ + { + "bbox": [ + 288, + 74, + 321, + 83 + ], + "type": "text", + "content": "UDiffText" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 279, + 86, + 332, + 125 + ], + "lines": [ + { + "bbox": [ + 279, + 86, + 332, + 125 + ], + "spans": [ + { + "bbox": [ + 279, + 86, + 332, + 125 + ], + "type": "image", + "image_path": "1d1e1e7e1d79e72a195f7daee2868271683afcaf59d08919e5ec588bfa1d0ded.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 281, + 127, + 331, + 148 + ], + "blocks": [ + { + "bbox": [ + 281, + 127, + 331, + 148 + ], + "lines": [ + { + "bbox": [ + 281, + 127, + 331, + 148 + ], + "spans": [ + { + "bbox": [ + 281, + 127, + 331, + 148 + ], + "type": "image", + "image_path": "77d976b5bab074c34b5386516de6ea7d6b297b579223c2b55c4f056540b02189.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 337, + 86, + 388, + 148 + ], + "blocks": [ + { + "bbox": [ + 348, + 74, + 376, + 83 + ], + "lines": [ + { + "bbox": [ + 348, + 74, + 376, + 83 + ], + "spans": [ + { + "bbox": [ + 348, + 74, + 376, + 83 + ], + "type": "text", + "content": "AnyText" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 86, + 388, + 148 + ], + "lines": [ + { + "bbox": [ + 337, + 86, + 388, + 148 + ], + "spans": [ + { + "bbox": [ + 337, + 86, + 388, + 148 + ], + "type": "image", + "image_path": "9ae130331065af71b444b44b67bba559e4fc540bce5f600dd73531fb6c3e74d9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 394, + 85, + 446, + 125 + ], + "blocks": [ + { + "bbox": [ + 395, + 74, + 444, + 83 + ], + "lines": [ + { + "bbox": [ + 395, + 74, + 444, + 83 + ], + "spans": [ + { + "bbox": [ + 395, + 74, + 444, + 83 + ], + "type": "text", + "content": "TextDiffuser-2" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 394, + 85, + 446, + 125 + ], + "lines": [ + { + "bbox": [ + 394, + 85, + 446, + 125 + ], + "spans": [ + { + "bbox": [ + 394, + 85, + 446, + 125 + ], + "type": "image", + "image_path": "55be7c6ec5685408af1e65fbcbfb30b22a430af3b4faea505b091caacaac5dfe.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 394, + 126, + 445, + 148 + ], + "blocks": [ + { + "bbox": [ + 394, + 126, + 445, + 148 + ], + "lines": [ + { + "bbox": [ + 394, + 126, + 445, + 148 + ], + "spans": [ + { + "bbox": [ + 394, + 126, + 445, + 148 + ], + "type": "image", + "image_path": "979797ea2384fdc03db4db3f77e3c74885eb8fa32176bea947736c94dcac860c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 457, + 85, + 504, + 125 + ], + "blocks": [ + { + "bbox": [ + 463, + 74, + 490, + 82 + ], + "lines": [ + { + "bbox": [ + 463, + 74, + 490, + 82 + ], + "spans": [ + { + "bbox": [ + 463, + 74, + 490, + 82 + ], + "type": "text", + "content": "TextSSR" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 85, + 504, + 125 + ], + "lines": [ + { + "bbox": [ + 457, + 85, + 504, + 125 + ], + "spans": [ + { + "bbox": [ + 457, + 85, + 504, + 125 + ], + "type": "image", + "image_path": "1fc3ac4e4cf38c09fd51caae7558cc4fba1cd0acd14a0a17c4352a9e5fa31b3a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 452, + 127, + 504, + 148 + ], + "blocks": [ + { + "bbox": [ + 452, + 127, + 504, + 148 + ], + "lines": [ + { + "bbox": [ + 452, + 127, + 504, + 148 + ], + "spans": [ + { + "bbox": [ + 452, + 127, + 504, + 148 + ], + "type": "image", + "image_path": "580c1b1f679a309b1ece3cf8109fa10ea0e85923768b88ccb94b0f00a76604c9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 108, + 152, + 160, + 190 + ], + "blocks": [ + { + "bbox": [ + 108, + 152, + 160, + 190 + ], + "lines": [ + { + "bbox": [ + 108, + 152, + 160, + 190 + ], + "spans": [ + { + "bbox": [ + 108, + 152, + 160, + 190 + ], + "type": "image", + "image_path": "91f58b865faadcd8c362ebb9a92ea8be5fa2fa37e9b8a311e3fd4f79044be5db.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 108, + 192, + 160, + 213 + ], + "blocks": [ + { + "bbox": [ + 108, + 192, + 160, + 213 + ], + "lines": [ + { + "bbox": [ + 108, + 192, + 160, + 213 + ], + "spans": [ + { + "bbox": [ + 108, + 192, + 160, + 213 + ], + "type": "image", + "image_path": "1616cd0213d11274b0dc7f59b3f340e4da0fdc54389fb6c6aa74d827dd7a12ef.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 165, + 152, + 217, + 191 + ], + "blocks": [ + { + "bbox": [ + 165, + 152, + 217, + 191 + ], + "lines": [ + { + "bbox": [ + 165, + 152, + 217, + 191 + ], + "spans": [ + { + "bbox": [ + 165, + 152, + 217, + 191 + ], + "type": "image", + "image_path": "144d84cd9cb27db04d8f1d316b0e72a820473508a589fa5e81d389f1222a24d3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 166, + 192, + 217, + 213 + ], + "blocks": [ + { + "bbox": [ + 166, + 192, + 217, + 213 + ], + "lines": [ + { + "bbox": [ + 166, + 192, + 217, + 213 + ], + "spans": [ + { + "bbox": [ + 166, + 192, + 217, + 213 + ], + "type": "image", + "image_path": "4c766147a574570cd9a4321677ab085472fc40ee5cc8489c046782d0ab63dcfc.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 223, + 152, + 274, + 191 + ], + "blocks": [ + { + "bbox": [ + 223, + 152, + 274, + 191 + ], + "lines": [ + { + "bbox": [ + 223, + 152, + 274, + 191 + ], + "spans": [ + { + "bbox": [ + 223, + 152, + 274, + 191 + ], + "type": "image", + "image_path": "7f477be425f529b864d86c35a7594d04f3d408d8fe66e71ce68c212e2411a2a1.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 223, + 192, + 274, + 213 + ], + "blocks": [ + { + "bbox": [ + 223, + 192, + 274, + 213 + ], + "lines": [ + { + "bbox": [ + 223, + 192, + 274, + 213 + ], + "spans": [ + { + "bbox": [ + 223, + 192, + 274, + 213 + ], + "type": "image", + "image_path": "34c14200ca0bc58b012ab030e2ea3c0c0356c0d8bf1e844e4c637623356db553.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 278, + 152, + 332, + 191 + ], + "blocks": [ + { + "bbox": [ + 278, + 152, + 332, + 191 + ], + "lines": [ + { + "bbox": [ + 278, + 152, + 332, + 191 + ], + "spans": [ + { + "bbox": [ + 278, + 152, + 332, + 191 + ], + "type": "image", + "image_path": "946b835d48180427168e31c1f5114a2902a62049255d1c9460055f268acbc9b8.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 280, + 192, + 331, + 213 + ], + "blocks": [ + { + "bbox": [ + 280, + 192, + 331, + 213 + ], + "lines": [ + { + "bbox": [ + 280, + 192, + 331, + 213 + ], + "spans": [ + { + "bbox": [ + 280, + 192, + 331, + 213 + ], + "type": "image", + "image_path": "bd50607e466fbfe520083c565dceca8ea45d00241460014852a3e8c0a9c69956.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 337, + 152, + 388, + 191 + ], + "blocks": [ + { + "bbox": [ + 337, + 152, + 388, + 191 + ], + "lines": [ + { + "bbox": [ + 337, + 152, + 388, + 191 + ], + "spans": [ + { + "bbox": [ + 337, + 152, + 388, + 191 + ], + "type": "image", + "image_path": "0b1e3369e7d86773fb42e5d4986f67074bb0d18c4c62a82ee69091e436173878.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 337, + 192, + 388, + 213 + ], + "blocks": [ + { + "bbox": [ + 337, + 192, + 388, + 213 + ], + "lines": [ + { + "bbox": [ + 337, + 192, + 388, + 213 + ], + "spans": [ + { + "bbox": [ + 337, + 192, + 388, + 213 + ], + "type": "image", + "image_path": "8f58f7bd09506221ca649410e84569d3f3ab4b12591e70d016894e1d469df770.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 394, + 152, + 446, + 191 + ], + "blocks": [ + { + "bbox": [ + 394, + 152, + 446, + 191 + ], + "lines": [ + { + "bbox": [ + 394, + 152, + 446, + 191 + ], + "spans": [ + { + "bbox": [ + 394, + 152, + 446, + 191 + ], + "type": "image", + "image_path": "f7bc5573365aa391984850322c4e22212ec6fa21a67a4e1a6709acac54648007.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 394, + 192, + 446, + 213 + ], + "blocks": [ + { + "bbox": [ + 394, + 192, + 446, + 213 + ], + "lines": [ + { + "bbox": [ + 394, + 192, + 446, + 213 + ], + "spans": [ + { + "bbox": [ + 394, + 192, + 446, + 213 + ], + "type": "image", + "image_path": "3ad719daf98b1b7a772cbff0b3a082c10d2e6c1b24dd0f7b094756e37e557db7.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 457, + 152, + 496, + 191 + ], + "blocks": [ + { + "bbox": [ + 457, + 152, + 496, + 191 + ], + "lines": [ + { + "bbox": [ + 457, + 152, + 496, + 191 + ], + "spans": [ + { + "bbox": [ + 457, + 152, + 496, + 191 + ], + "type": "image", + "image_path": "4b49b7ec54917f988f627b93d7e5b13ecb6036b869bd3b7375ff9dfbb9e2b5af.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 452, + 192, + 504, + 213 + ], + "blocks": [ + { + "bbox": [ + 452, + 192, + 504, + 213 + ], + "lines": [ + { + "bbox": [ + 452, + 192, + 504, + 213 + ], + "spans": [ + { + "bbox": [ + 452, + 192, + 504, + 213 + ], + "type": "image", + "image_path": "4b4dc2ee993be44a877141e4a331dbc0ec0394abf6ba70b5e121ce952f4f3c2e.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 107, + 217, + 160, + 256 + ], + "blocks": [ + { + "bbox": [ + 107, + 217, + 160, + 256 + ], + "lines": [ + { + "bbox": [ + 107, + 217, + 160, + 256 + ], + "spans": [ + { + "bbox": [ + 107, + 217, + 160, + 256 + ], + "type": "image", + "image_path": "81bdd850e6f2dcd8d663805deb6aa990c15a7b54876f5363ab0a9db5fa9f46e2.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 108, + 258, + 159, + 284 + ], + "blocks": [ + { + "bbox": [ + 108, + 258, + 159, + 284 + ], + "lines": [ + { + "bbox": [ + 108, + 258, + 159, + 284 + ], + "spans": [ + { + "bbox": [ + 108, + 258, + 159, + 284 + ], + "type": "image", + "image_path": "7216de48a3193cad2f787e266d0a040e66403d62fe7fa42851da536da85c990f.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "type": "text", + "content": "Figure 4. Examples synthesized by TextSSR and other existing open-source methods (UDiffText only supports English text for its limited text encoder). TextSSR consistently produces accurate and realistic text instances. Please refer to Appendix 7.1 for more analysis." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 164, + 217, + 217, + 256 + ], + "blocks": [ + { + "bbox": [ + 164, + 217, + 217, + 256 + ], + "lines": [ + { + "bbox": [ + 164, + 217, + 217, + 256 + ], + "spans": [ + { + "bbox": [ + 164, + 217, + 217, + 256 + ], + "type": "image", + "image_path": "618549cc2d5d54d058c21ba95453feebce23ca4cfad0b7b43d62f9786b4373f2.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 165, + 258, + 217, + 284 + ], + "blocks": [ + { + "bbox": [ + 165, + 258, + 217, + 284 + ], + "lines": [ + { + "bbox": [ + 165, + 258, + 217, + 284 + ], + "spans": [ + { + "bbox": [ + 165, + 258, + 217, + 284 + ], + "type": "image", + "image_path": "248400ce8d95bfcbb656f163cbd3d56f9c618b7acad871450e22dad576339579.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 222, + 217, + 274, + 256 + ], + "blocks": [ + { + "bbox": [ + 222, + 217, + 274, + 256 + ], + "lines": [ + { + "bbox": [ + 222, + 217, + 274, + 256 + ], + "spans": [ + { + "bbox": [ + 222, + 217, + 274, + 256 + ], + "type": "image", + "image_path": "8f0285cf616479a8cbc6999d5b6bd0a035b31100552df8dac25436a3a9898699.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 222, + 258, + 274, + 284 + ], + "blocks": [ + { + "bbox": [ + 222, + 258, + 274, + 284 + ], + "lines": [ + { + "bbox": [ + 222, + 258, + 274, + 284 + ], + "spans": [ + { + "bbox": [ + 222, + 258, + 274, + 284 + ], + "type": "image", + "image_path": "c6b9bdf86c7677436dd02108319aa3bfb9bcef267e0dc2d02a1d7f75a96e13d0.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 280, + 217, + 332, + 285 + ], + "blocks": [ + { + "bbox": [ + 280, + 217, + 332, + 285 + ], + "lines": [ + { + "bbox": [ + 280, + 217, + 332, + 285 + ], + "spans": [ + { + "bbox": [ + 280, + 217, + 332, + 285 + ], + "type": "image", + "image_path": "293c00de0c3af1731ef69f3b389c1e49094861a6423249135df63d4e038f62d0.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 337, + 217, + 388, + 256 + ], + "blocks": [ + { + "bbox": [ + 337, + 217, + 388, + 256 + ], + "lines": [ + { + "bbox": [ + 337, + 217, + 388, + 256 + ], + "spans": [ + { + "bbox": [ + 337, + 217, + 388, + 256 + ], + "type": "image", + "image_path": "4a83c1a140a826fa802a317aee4a31086c43c12b6e4642feb6c8fa9aaaafb585.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 336, + 258, + 388, + 284 + ], + "blocks": [ + { + "bbox": [ + 336, + 258, + 388, + 284 + ], + "lines": [ + { + "bbox": [ + 336, + 258, + 388, + 284 + ], + "spans": [ + { + "bbox": [ + 336, + 258, + 388, + 284 + ], + "type": "image", + "image_path": "f539f2a66f6215d63a38f836a25a5b5a33abecb45e9363b6df9c7c63dd164dbf.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 394, + 217, + 446, + 256 + ], + "blocks": [ + { + "bbox": [ + 394, + 217, + 446, + 256 + ], + "lines": [ + { + "bbox": [ + 394, + 217, + 446, + 256 + ], + "spans": [ + { + "bbox": [ + 394, + 217, + 446, + 256 + ], + "type": "image", + "image_path": "e8b971d361f01c594e4b038c4ab45d51393422c14216a3a428bc6c2205e6a962.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 394, + 258, + 445, + 284 + ], + "blocks": [ + { + "bbox": [ + 394, + 258, + 445, + 284 + ], + "lines": [ + { + "bbox": [ + 394, + 258, + 445, + 284 + ], + "spans": [ + { + "bbox": [ + 394, + 258, + 445, + 284 + ], + "type": "image", + "image_path": "16ae66567977c09ffca176166f5fa7f89317e43299b35224b8dd41ac60acb7fe.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 456, + 217, + 496, + 256 + ], + "blocks": [ + { + "bbox": [ + 456, + 217, + 496, + 256 + ], + "lines": [ + { + "bbox": [ + 456, + 217, + 496, + 256 + ], + "spans": [ + { + "bbox": [ + 456, + 217, + 496, + 256 + ], + "type": "image", + "image_path": "ffe4026785d57d371d7826e14228e9968885ea3ca54d62390f401cd2b7fd8899.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 452, + 258, + 503, + 284 + ], + "blocks": [ + { + "bbox": [ + 452, + 258, + 503, + 284 + ], + "lines": [ + { + "bbox": [ + 452, + 258, + 503, + 284 + ], + "spans": [ + { + "bbox": [ + 452, + 258, + 503, + 284 + ], + "type": "image", + "image_path": "f51e3698a500ccc29c39d9e1a89aef5fb65b83e2dfc5b68c7c7bb889667f8a37.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "table", + "bbox": [ + 61, + 334, + 291, + 487 + ], + "blocks": [ + { + "bbox": [ + 61, + 334, + 291, + 487 + ], + "lines": [ + { + "bbox": [ + 61, + 334, + 291, + 487 + ], + "spans": [ + { + "bbox": [ + 61, + 334, + 291, + 487 + ], + "type": "table", + "html": "
TypeMethodSeqAcc(%)↑
RegularIrregularAvg
RenderingSynthText [15]48.5526.0839.74
VISD [53]38.0226.8833.65
UnrealText [25]32.0619.8527.28
DiffusionTextDiffuser [4]30.088.7921.74
GlyphControl [50]40.8013.2129.99
AnyText [41]49.8019.6237.97
TextDiffuser-2 [5]40.2610.4828.59
SceneVTG [60]54.9735.5047.34
TextSSR (Ours)59.7635.7150.33
RealReal-L [1]57.9737.6650.01
TextOCR [39]57.8341.5654.05
", + "image_path": "05a11de2377eb5be31595c28eebfaf9fac4bd58b1ee1c9a2bf8c931e6f306dde.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "table_body" + } + ], + "index": 48 + }, + { + "bbox": [ + 55, + 494, + 295, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 559 + ], + "type": "text", + "content": "Table 2. Comparison of different data synthetic methods, where CRNN [36] trained on different synthetic data and real data of " + }, + { + "bbox": [ + 55, + 494, + 295, + 559 + ], + "type": "inline_equation", + "content": "{30}\\mathrm{k}" + }, + { + "bbox": [ + 55, + 494, + 295, + 559 + ], + "type": "text", + "content": " is employed to obtain the results. \"Regular\" denotes the average on IIIT [27], SVT [43], and IC13 [21] datasets, \"Irregular\" represents the average on IC15 [22], SVTP [29], and CUTE [32] datasets, while \"Avg\" is the average of the six datasets." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": "fectively. In contrast, TextSSR benefits from a region-centered generation pipeline. It excels at handling challenging text instances and maintains reasonable synthesis accuracy. Moreover, the condition for TextSSR is glyph shapes, which can be extended to other languages like Chinese. The results on ShopSign prove TextSSR's multilingual capability. It outperforms the purportedly multilingual-focused AnyText by " + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "inline_equation", + "content": "18.7\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": ". More details on accuracy evaluation are provided in Appendix 6.3.2." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Realism Evaluation. Following the validation presented in SceneVTG [60], we generate a fixed amount (30k) of data" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 337, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 409 + ], + "type": "text", + "content": "to train an identical model under fair settings and then test it on common STR benchmarks [21, 22, 27, 29, 32, 43]. This approach avoids the need for separate verification of text correctness or synthesis quality. If the generated text is inaccurate or unrealistic, this will be reflected in poor recognition performance during testing." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 313, + 411, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 615 + ], + "type": "text", + "content": "As shown in Tab. 2, for the compared methods, their average SeqAcc is below " + }, + { + "bbox": [ + 313, + 411, + 556, + 615 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 411, + 556, + 615 + ], + "type": "text", + "content": " except for SceneVTG [60]. The result is closely linked to their low generation accuracy. Note that rendering-based methods generally perform better than traditional diffusion-based methods in irregular benchmarks, highlighting the challenge of generating highly controllable text images. SceneVTG benefits from its two-stage pipeline of text erasure and then rendering, it reaches a SeqAcc of " + }, + { + "bbox": [ + 313, + 411, + 556, + 615 + ], + "type": "inline_equation", + "content": "47.3\\%" + }, + { + "bbox": [ + 313, + 411, + 556, + 615 + ], + "type": "text", + "content": ". Nevertheless, TextSSR, featured by its glyph-based condition injection and anagram-based generation, surpasses SceneVTG by nearly " + }, + { + "bbox": [ + 313, + 411, + 556, + 615 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 313, + 411, + 556, + 615 + ], + "type": "text", + "content": ". Moreover, the gap between models trained on 30k TextSSR and real-world TextOCR is only " + }, + { + "bbox": [ + 313, + 411, + 556, + 615 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 313, + 411, + 556, + 615 + ], + "type": "text", + "content": ", which comes from irregular text. These results convincingly demonstrate the realism of TextSSR data, especially in regular text synthesizing. However, further efforts are still needed to improve the synthesis of irregular text." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Scalability Evaluation. TextSSR data can be easily scaled up to a large quantity. It is necessary to assess whether the increase in data volume can translate into performance improvements. Therefore, we use the model generated by the 30k data in the realism evaluation as baseline (TextSSR " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\times 1" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "), and re-train CRNN under the same experimental setup but with data doubled every time. Results of this scalability evaluation are presented in Tab. 4. When scaled up" + } + ] + } + ], + "index": 54 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17469" + } + ] + } + ], + "index": 55 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 538, + 256 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 538, + 256 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 538, + 256 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 538, + 256 + ], + "type": "table", + "html": "
ModelDatasetVolumeSeqAcc(%)↑
IIIT5k [27]SVT [43]IC13 [21]IC15 [22]SVTP [29]CUTE80 [32]AvgContextless [20]
CRNN [36]ST4m90.1382.8490.9072.4572.0982.9981.9033.50
ST6.98m90.8083.0091.8372.3473.8081.6082.0632.61
ST + SynthAdd4m + 1.05m90.4383.9391.1373.5572.7180.2181.9934.53
ST + TextSSR4m + 1.05m92.1387.0293.3577.4277.6784.3785.3355.97
ST + TextSSR4m + 3.55m93.5789.1894.2878.5878.4582.9986.1757.89
ST + TextOCR4m + 1.05m93.1388.1092.5379.6879.2283.6886.0644.93
ST + TextOCR + TextSSR5.05m + 3.55m94.5789.3494.4080.2379.2284.7287.0858.15
+1.44+1.24+1.87+0.55+0.00+1.04+1.02+13.22
MAERec [20]ST4m96.0091.9695.8083.8885.7489.5890.4959.43
ST6.98m96.9092.8995.8084.7086.8292.3691.5861.62
ST + SynthAdd4m + 1.05m95.9793.2095.8085.0486.9890.2891.2160.46
ST + TextSSR4m + 1.05m97.5393.8296.8586.7589.4694.1093.0876.77
ST + TextSSR4m + 3.55m98.1394.7497.3287.8589.6194.1093.6378.17
ST + TextOCR4m + 1.05m98.1095.5297.2089.3490.3996.8794.5774.33
ST + TextOCR + TextSSR5.05m + 3.55m98.2096.7597.2089.9592.7197.9295.4682.03
+0.10+1.23+0.00+0.61+2.32+1.05+0.89+7.70
", + "image_path": "1842e14b3f4fb53be5d0c9068e92d8f2ca147ff5d671772f3b8e602cb52c899c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 77, + 305, + 274, + 422 + ], + "blocks": [ + { + "bbox": [ + 55, + 264, + 553, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 553, + 288 + ], + "type": "text", + "content": "Table 3. Performance of models trained on different data combinations, where CRNN and MAERec are selected as model representatives. Bold and underline indicate the highest and second highest values, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 305, + 274, + 422 + ], + "lines": [ + { + "bbox": [ + 77, + 305, + 274, + 422 + ], + "spans": [ + { + "bbox": [ + 77, + 305, + 274, + 422 + ], + "type": "table", + "html": "
MethodVolumeSeqAcc(%)↑
RegularIrregularAvg
TextSSR ×130k59.7635.7150.33
TextSSR EL30k24.5810.2318.95
TextSSR ×260k60.4934.4950.29
TextSSR ×4120k65.1739.3455.03
TextSSR ×8240k73.9649.6964.44
TextOCR30k57.8341.5654.05
", + "image_path": "1174999513ea79e912630ec76825a8086defe5fdb5271bb82d7a9d85daaf3115.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "text", + "content": "Table 4. TextSSR scalability evaluation. " + }, + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\times n" + }, + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "text", + "content": " refers to perform the anagram-based rendering " + }, + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "text", + "content": " times. EL denotes equal length edit." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "text", + "content": ", accuracy improvements of " + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "inline_equation", + "content": "14.1\\%" + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "text", + "content": " are observed, respectively. The result indicates that increasing the data volume indeed contributes to better STR models. Note that models in the two scenarios also surpass the model trained on 30k real-world data by " + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "inline_equation", + "content": "10.4\\%" + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "text", + "content": ", respectively. It is observed that TextSSR " + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 54, + 474, + 295, + 641 + ], + "type": "text", + "content": " also presents a much better result than real data model in irregular text, highlighting that the challenge of irregularity can be overcome to some extent by increasing the data volume. Meanwhile, we also conducted a comparison experiment similar to [57], i.e., substituting text instances with equal-length strings (TextSSR EL), where the trained model reported quite worse results. This again demonstrates the effectiveness of our anagram-based generation pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 295, + 714 + ], + "type": "text", + "content": "Usability Assessment. We further assess the effectiveness of our synthesis method in supporting the large-scale training of STR models. To ensure the quality of data, we apply a quality-screened step to every sample generated by TextSSR. We still use SVTRv2 [11] to recognize the sample and compare the recognized content with its genera" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": "tion label. The sample is reserved if they are consistent, otherwise we discard it. The results on different training data combinations are shown in Tab. 3. The observations can be summarized as follows: First, the rendering-based ST dataset reaches a saturation point in terms of accuracy, where " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "6.98\\mathrm{m}" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " samples only show marginal improvement " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "(0.16\\%)" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " on CRNN over the " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "4\\mathrm{m}" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " dataset on the six common benchmarks. This suggests that we have to move towards more realistic data synthesis. Second, when incorporating " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "1.05\\mathrm{m}" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " our synthetic dataset, the two STR models (CRNN and MAERec) gain improvements of " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " over the " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "4\\mathrm{m}" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " ST dataset on the six common benchmarks, respectively. They are only " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " lower than ST combined with equivalent size real-world TextOCR. This suggests that TextSSR data is approaching real-world data in quality. Third, when further increasing TextSSR data to " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "3.55\\mathrm{m}" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": ", it is seen that the purely synthetic training data " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "(\\mathrm{ST} + 3.55\\mathrm{m}" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " TextSSR) gain performance more approaching the synthetic-real mixed training data " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "(\\mathrm{ST} + 1.05\\mathrm{m}" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " TextOCR). Moreover, when combining the " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "3.55\\mathrm{m}" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " synthetic data with this mixed data, the trained model can report further performance improvements, obtaining " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "1.02\\%" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " on CRNN and " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "0.89\\%" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " on MAERec. This implies that TextSSR data is a valuable supplement to existing training data. Using it individually or collectively both can benefit STR model training. As a result, we mark this " + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "inline_equation", + "content": "3.55\\mathrm{m}" + }, + { + "bbox": [ + 313, + 308, + 555, + 656 + ], + "type": "text", + "content": " dataset as TextSSR-F, an accurate and realistic large-scale synthetic dataset constructed by our TextSSR, and make it publicly available." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Since our anagram-based synthesis generates a large number of contextless words, and contextless word generation is also discussed in SynthAdd [24], we conduct a comparison between SynthAdd and our TextSSR. As shown" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17470" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 140, + 287, + 248 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 131 + ], + "type": "text", + "content": "in Tab. 3, there are clear performance margins (3.3% on CRNN and 1.9% on MAERec) between the model trained on traditional SynthAdd and TextSSR. This also indicates the superiority of TextSSR as a STR training data synthetic method." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 140, + 287, + 248 + ], + "lines": [ + { + "bbox": [ + 65, + 140, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 65, + 140, + 287, + 248 + ], + "type": "table", + "html": "
region-centricchar-glyphchar-positionfine-tune VAESeqAcc (%) ↑NED ↑FID-R ↓
20.280.3444102.89
✓73.500.876445.89
✓✓74.150.885246.65
✓✓✓80.260.913345.83
✓✓✓77.210.904350.34
✓✓76.230.899151.94
✓✓✓47.550.621162.52
✓✓✓✓82.990.930850.76
", + "image_path": "fad2a90b44ba7c97c046f6faeb563aefb478236fbdd7c3c01aa790ebba78ffaa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 295, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 277 + ], + "type": "text", + "content": "Table 5. Ablation study of TextSSR components using the TextSSR pre-training model in Tab. 1 on the IC13 dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 299, + 149, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 149, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 149, + 312 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 316, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 296, + 483 + ], + "type": "text", + "content": "To assess the proposed components, we start with the TextSSR pre-training model in Tab. 1 (to simplify, trained for only 5k steps), where all components are considered. Then, we stepwisely remove one or some of the components to assess their respective necessity and effectiveness. From the results on Tab. 5, we can see that the removal of each one brings a clear decline in performance, while the absence of \"region-centric\" processing can lead to a catastrophic drop of " + }, + { + "bbox": [ + 54, + 316, + 296, + 483 + ], + "type": "inline_equation", + "content": "35.4\\%" + }, + { + "bbox": [ + 54, + 316, + 296, + 483 + ], + "type": "text", + "content": " in accuracy. Additionally, fine-tuning the VAE offers a trade-off, slightly sacrificing visual quality for an improvement in accuracy. This is because real-world scene text also confronts poor visual quality challenge, simulating this to some extent can provide valuable samples for model training." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 83, + 494, + 272, + 633 + ], + "blocks": [ + { + "bbox": [ + 83, + 494, + 272, + 633 + ], + "lines": [ + { + "bbox": [ + 83, + 494, + 272, + 633 + ], + "spans": [ + { + "bbox": [ + 83, + 494, + 272, + 633 + ], + "type": "image", + "image_path": "9c36a1114f5b860501ab83aac2dbfc51a2844da35800dc8b4bd1174c36a386d3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 643, + 295, + 677 + ], + "lines": [ + { + "bbox": [ + 55, + 643, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 295, + 677 + ], + "type": "text", + "content": "Figure 5. Performance of TextSSR fine-tuning model with and without region-centric processing across text regions of different sizes (in pixels)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "We perform another experiment to verify the effectiveness of the region-centric processing. The results in Fig. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "show that without this processing, TextSSR has to synthesize text instance on the whole image. It exhibits a performance trend that first rises and then falls as the size increases. When the region is small, the entire image is scaled severely, resulting in more blurred text regions and insufficient pixels for text rendering. As the region size increases, larger regions and more pixels are allowed, therefore performance increase. When a too large region given, it causes the side effect that constraints the shape of the generated text, and performance decline. In contrast, incorporating region-centric processing allows TextSSR to achieve superior and consistent performance across all sizes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 216, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 264 + ], + "type": "text", + "content": "In Fig. 6, we also illustratively ablate the character-level position and glyph components. The results indicate that omitting either component leads to issues such as character deformation, incorrect characters, and duplication errors." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 329, + 273, + 541, + 401 + ], + "blocks": [ + { + "bbox": [ + 329, + 273, + 541, + 401 + ], + "lines": [ + { + "bbox": [ + 329, + 273, + 541, + 401 + ], + "spans": [ + { + "bbox": [ + 329, + 273, + 541, + 401 + ], + "type": "image", + "image_path": "ecc09037c1d9dea1144cce0c4e95e7a29a009db736e71773833247a99d693bd8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 411, + 553, + 434 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 553, + 434 + ], + "type": "text", + "content": "Figure 6. Visualizations of TextSSR with and without character-level prior. More examples are provided in Appendix 7.3" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 454, + 388, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 388, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 388, + 466 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "We have proposed TextSSR to provide STR model training with high-quality synthetic data. It employs the glyph of existing text and surrounding context as the prompt of the diffusion-based synthesizing model, and combines character-level constraints and permutations. TextSSR successfully synthesizes accurate and realistic text instances at a large scale. The TextSSR-F dataset with 3.55 million diverse and realistic instances are constructed. We have conducted extensive experiments to assess TextSSR-F. STR models trained solely on TextSSR-F, and on the combination of TextSSR-F and real-world data both show performance improvements. While the effectiveness of TextSSR has been basically confirmed, for irregular text, there is still an accuracy gap between models trained on TextSSR-F and real-world data. Thus, future work includes the exploration of more controllable ways to further improve the synthesizing quality, especially for challenging instances such as curved and multi-oriented text. We are also interested in extending TextSSR to generate large-scale data in different languages like Chinese." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17471" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "Acknowledgement This work was supported by the National Natural Science Foundation of China (Nos. 32341012, 62172103)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 115, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 115, + 132 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 115, + 132 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 139, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 139, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 139, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 139, + 296, + 183 + ], + "type": "text", + "content": "[1] J. Baek, Y. Matsui, and K. Aizawa. What if we only use real datasets for scene text recognition? toward scene text recognition with fewer labels. In CVPR, pages 3113-3122, 2021. 6, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 185, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 185, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 185, + 296, + 217 + ], + "type": "text", + "content": "[2] D. Bautista and R. Atienza. Scene text recognition with permuted autoregressive sequence models. In ECCV, pages 178-196, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 218, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 218, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 218, + 296, + 251 + ], + "type": "text", + "content": "[3] H. Chen, Z. Xu, Z. Gu, J. Lan, X. Zheng, Y. Li, C. Meng, H. Zhu, and W. Wang. DiffUTE: Universal text editing diffusion model. In NeurIPS, pages 63062-63074, 2023. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 252, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 252, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 252, + 296, + 285 + ], + "type": "text", + "content": "[4] J. Chen, Y. Huang, T. Lv, L. Cui, Q. Chen, and F. Wei. TextDiffuser: Diffusion models as text painters. In NeurIPS, pages 9353-9387, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 286, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 296, + 319 + ], + "type": "text", + "content": "[5] J. Chen, Y. Huang, T. Lv, L. Cui, Q. Chen, and F. Wei. TextDiffuser-2: Unleashing the power of language models for text rendering. In ECCV, pages 386–402, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 321, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 321, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 296, + 374 + ], + "type": "text", + "content": "[6] C.K. Chng, Y. Liu, Y. Sun, C.C. Ng, C. Luo, Z. Ni, C.M. Fang, S. Zhang, J. Han, E. Ding, J. Liu, D. Karatzas, C.S. Chan, and L. Jin. ICDAR2019 robust reading challenge on arbitrary-shaped text-rc-art. In ICDAR, pages 1571–1576, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 376, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 296, + 431 + ], + "type": "text", + "content": "[7] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 433, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 296, + 465 + ], + "type": "text", + "content": "[8] Y. Du, Z. Chen, C. Jia, X. Yin, T. Zheng, C. Li, Y. Du, and Y. Jiang. SVTR: Scene text recognition with a single visual model. In IJCAI, pages 884-890, 2022. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 467, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 296, + 510 + ], + "type": "text", + "content": "[9] Y. Du, Z. Chen, C. Jia, X. Yin, C. Li, Y. Du, and Y.-G. Jiang. Context perception parallel decoder for scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell., 47(6):4668-4683, 2025. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 512, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 512, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 57, + 512, + 295, + 544 + ], + "type": "text", + "content": "[10] Y. Du, Z. Chen, Y. Su, C. Jia, and Y.-G. Jiang. Instruction-guided scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell., 47(4):2723-2738, 2025. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 545, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 545, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 545, + 296, + 578 + ], + "type": "text", + "content": "[11] Y. Du, Z. Chen, H. Xie, C. Jia, and Y. Jiang. SVTRv2: CTC beats encoder-decoder models in scene text recognition. In ICCV, 2025. 1, 3, 5, 7, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 580, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 296, + 622 + ], + "type": "text", + "content": "[12] S. Fang, H. Xie, Y. Wang, Z. Mao, and Y. Zhang. Read like humans: Autonomous, bidirectional and iterative language modeling for scene text recognition. In CVPR, pages 7098-7107, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 667 + ], + "type": "text", + "content": "[13] S. Fang, Z. Mao, H. Xie, Y. Wang, C. Yan, and Y. Zhang. Abinet++: Autonomous, bidirectional and iterative language modeling for scene text spotting. IEEE Trans. Pattern Anal. Mach. Intell., 45(6):7123-7141, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[14] J. Gu, X. Meng, G. Lu, L. Hou, N. Minzhe, X. Liang, L. Yao, R. Huang, W. Zhang, X. Jiang, C. Xu, and H. Xu. Wukong: A 100 million large-scale chinese cross-modal pre-training benchmark. In NeurIPS, pages 26418-26431, 2022. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[15] A. Gupta, A. Vedaldi, and A. Zisserman. Synthetic data for text localisation in natural images. In CVPR, pages 2315-2324, 2016. 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 554, + 150 + ], + "type": "text", + "content": "[16] M. He, Y. Liu, Z. Yang, S. Zhang, C. Luo, F. Gao, Q. Zheng, Y. Wang, X. Zhang, and L. Jin. ICPR2018 contest on robust reading for multi-type web images. In ICPR, pages 7-12, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 153, + 554, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 554, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 554, + 186 + ], + "type": "text", + "content": "[17] Santoso J, Simon C, and Williem. On manipulating scene text in the wild with diffusion models. In WACV, pages 5202-5211, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 188, + 554, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 554, + 231 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 554, + 231 + ], + "type": "text", + "content": "[18] M. Jaderberg, K. Simonyan, A. Vedaldi, and A. Zisserman. Synthetic data and artificial neural networks for natural scene text recognition. In Workshop on Deep Learning, NeurIPS, 2014. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 234, + 554, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 234, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 554, + 266 + ], + "type": "text", + "content": "[19] J. Ji, G. Zhang, Z. Wang, B. Hou, Z. Zhang, B. Price, and S. Chang. Improving diffusion models for scene text editing with dual encoders. TMLR, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 269, + 554, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 554, + 300 + ], + "type": "text", + "content": "[20] Q. Jiang, J. Wang, D. Peng, C. Liu, and L. Jin. Revisiting scene text recognition: A data perspective. In ICCV, pages 20543-20554, 2023. 1, 3, 7, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 303, + 554, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 303, + 554, + 356 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 554, + 356 + ], + "type": "text", + "content": "[21] D. Karatzas, F. Shafait, S. Uchida, M. Iwamura, L. Gomez i Bigorda, S. Robles Mestre, J. Mas, D. Fernandez Mota, J. Almazán, and L.P. de las Heras. ICDAR 2013 robust reading competition. In ICDAR, pages 1484-1493, 2013. 3, 5, 6, 7, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 359, + 554, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 359, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 359, + 554, + 414 + ], + "type": "text", + "content": "[22] D. Karatzas, L. Gomez-Bigorda, A. Nicolaou, S. Ghosh, A. Bagdanov, M. Iwamura, J. Matas, L. Neumann, V.R. Chandrasekhar, S. Lu, F. Shafait, S. Uchida, and E. Valveny. IC-DAR 2015 competition on robust reading. In ICDAR, pages 1156-1160, 2015. 3, 5, 6, 7, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 416, + 554, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 554, + 460 + ], + "type": "text", + "content": "[23] C. Li, W. Liu, R. Guo, X. Yin, K. Jiang, Y. Du, Y. Du, L. Zhu, B. Lai, X. Hu, D. Yu, and Y. Ma. PP-OCRv3: More attempts for the improvement of ultra lightweightOCR system. arXiv preprint arXiv:2206.03001, 2022. 1, 3, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 461, + 554, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 461, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 461, + 554, + 495 + ], + "type": "text", + "content": "[24] H. Li, P. Wang, C. Shen, and G. Zhang. Show, attend and read: A simple and strong baseline for irregular text recognition. In AAAI, pages 8610-8617, 2019. 3, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 496, + 554, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 496, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 496, + 554, + 517 + ], + "type": "text", + "content": "[25] S. Long and C. Yao. Unrealtext: Synthesizing realistic scene text images from the unreal world. In CVPR, 2020. 3, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 519, + 554, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 519, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 519, + 554, + 552 + ], + "type": "text", + "content": "[26] C. Luo, L. Jin, and Z. Sun. Moran: A multi-object rectified attention network for scene text recognition. Pattern Recognit., 90:109-118, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 554, + 554, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 554, + 586 + ], + "type": "text", + "content": "[27] A. Mishra, K. Alahari, and CV Jawahar. Scene text recognition using higher order language priors. In BMVC, 2012. 3, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 589, + 554, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 589, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 317, + 589, + 554, + 643 + ], + "type": "text", + "content": "[28] N. Nayef, C. Liu, J. Ogier, Y. Patel, M. Busta, P.N. Chowdhury, D. Karatzas, W. Khlif, J. Matas, U. Pal, and J. Burie. ICDAR2019 robust reading challenge on multi-lingual scene text detection and recognition-rrc-plt-2019. In ICDAR, pages 1582-1587, 2019. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 646, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 554, + 678 + ], + "type": "text", + "content": "[29] T. Q. Phan, P. Shivakumara, S. Tian, and C. L. Tan. Recognizing text with perspective distortion in natural scenes. In ICCV, pages 569-576, 2013. 3, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[30] Y. Qu, Q. Tan, H. Xie, J. Xu, Y. Wang, and Y. Zhang. Exploring stroke-level modifications for scene text editing. In AAAI, pages 2119-2127, 2023. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17472" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[31] M. Rang, Z. Bi, C. Liu, Y. Wang, and K. Han. An empirical study of scaling law for scene text recognition. In CVPR, pages 15619-15629, 2024. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[32] A. Risnumawan, P. Shivakumara, C. S. Chan, and C. L. Tan. A robust arbitrary text detection system for natural scene images. Expert Syst. Appl., 41(18):8027-8048, 2014. 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[33] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 4, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[34] C. Schuhmann, R. Vencu, R. Beaumont, R. Kaczmarczyk, C. Mullis, A. Katta, T. Coombes, J. Jitsev, and A. Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "text", + "content": "[35] F. Sheng, Z. Chen, and B. Xu. NRTR: A no-recurrence sequence-to-sequence model for scene text recognition. In ICDAR, pages 781-786, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "type": "text", + "content": "[36] B. Shi, X. Bai, and C. Yao. An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell., 39(11):2298-2304, 2016. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "text", + "content": "[37] B. Shi, C. Yao, M. Liao, M. Yang, P. Xu, L. Cui, S. Belongie, S. Lu, and X. Bai. ICDAR2017 competition on reading Chinese text in the wild (rctw-17). In ICDAR, pages 1429–1434, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 397 + ], + "type": "text", + "content": "[38] B. Shi, M. Yang, X. Wang, P. Lyu, C. Yao, and X. Bai. ASTER: An attentional scene text recognizer with flexible rectification. IEEE Trans. Pattern Anal. Mach. Intell., 41(9): 2035-2048, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "text", + "content": "[39] A. Singh, G. Pang, M. Toh, J. Huang, W. Galuba, and T. Hassner. TextOCR: Towards large-scale end-to-end reasoning for arbitrary-shaped scene text. In CVPR, pages 8802-8812, 2021. 1, 3, 6, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[40] Y. Sun, Z. Ni, C.K. Chng, Y. Liu, C. Luo, C.C. Ng, J. Han, E. Ding, J. Liu, D. Karatzas, C.S. Chan, and L. Jin. ICDAR 2019 competition on large-scale street view text with partial labeling-rrc-lsvt. In ICDAR, pages 1557-1562, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 520 + ], + "type": "text", + "content": "[41] Y. Tuo, W. Xiang, J. Y. He, Y. Geng, and X. Xie. AnyText: Multilingual visual text generation and editing. In ICLR, 2024. 2, 3, 4, 5, 6, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "type": "text", + "content": "[42] A. Veit, T. Matera, L. Neumann, J. Matas, and S. Belongie. COCO-Text: Dataset and benchmark for text detection and recognition in natural images. arXiv preprint arXiv:1601.07140, 2016. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 567, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 590 + ], + "type": "text", + "content": "[43] K. Wang, B. Babenko, and S. Belongie. End-to-end scene text recognition. In ICCV, pages 1457-1464, 2011. 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 591, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 633 + ], + "type": "text", + "content": "[44] Y. Wang, H. Xie, S. Fang, M. Xing, J. Wang, S. Zhu, and Y. Zhang. Petr: Rethinking the capability of transformer-based language model in scene text recognition. IEEE Trans. Image Process., 31:5585-5598, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 667 + ], + "type": "text", + "content": "[45] Y. Wang, W. Zhang, H. Xu, and C. Jin. Dreamtext: High fidelity scene text synthesis. In CVPR, pages 28555-28563, 2025. 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[46] J. Wei, H. Zhan, Y. Lu, X. Tu, B. Yin, C. Liu, and U. Pal. Image as a language: Revisiting scene text recognition via balanced, unified and synchronized vision-language reasoning network. In AAAI, pages 5885-5893, 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 612 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 105 + ], + "type": "text", + "content": "[47] L. Wu, C. Zhang, J. Liu, J. Han, J. Liu, E. Ding, and X. Bai. Editing text in the wild. In ACM MM, pages 1500-1508, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 315, + 107, + 553, + 139 + ], + "type": "text", + "content": "[48] J. Xu, Y. Wang, H. Xie, and Y. Zhang. OTE: Exploring accurate scene text recognition using one token. In CVPR, pages 28327-28336, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 141, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 141, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 315, + 141, + 553, + 163 + ], + "type": "text", + "content": "[49] Q. Yang, J. Huang, and W. Lin. SwapText: Image based texts transfer in scenes. In CVPR, pages 14700-14709, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 164, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 164, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 315, + 164, + 553, + 206 + ], + "type": "text", + "content": "[50] Y. Yang, D. Gui, Y. Yuan, W. Liang, H. Ding, H. Hu, and K. Chen. GlyphControl: Glyph conditional control for visual text generation. In NeurIPS, pages 44050-44066, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 209, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 209, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 315, + 209, + 553, + 241 + ], + "type": "text", + "content": "[51] W. Zeng, Y. Shu, Z. Li, D. Yang, and Y. Zhou. TextCtrl: Diffusion-based scene text editing with prior guidance control. In NeurlPS, pages 138569–138594, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 243, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 243, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 315, + 243, + 553, + 275 + ], + "type": "text", + "content": "[52] C. Zhai, Z. Chen, J. Li, and B. Xu. Chinese image text recognition with blastm-ctc: a segmentation-free method. In CCPR, pages 525–536, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 277, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 277, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 277, + 553, + 308 + ], + "type": "text", + "content": "[53] F. Zhan, S. Lu, and C. Xue. Verisimilar image synthesis for accurate detection and recognition of texts in scenes. In ECCV, pages 249-266, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 310, + 553, + 354 + ], + "type": "text", + "content": "[54] C. Zhang, W. Ding, G. Peng, F. Fu, and W. Wang. Street view text recognition with deep learning for urban scene understanding in intelligent transportation systems. IEEE Trans. Intell. Transp., 22(7):4727-4743, 2021. 5, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 355, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 355, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 315, + 355, + 553, + 410 + ], + "type": "text", + "content": "[55] R. Zhang, Y. Zhou, Q. Jiang, Q. Song, N. Li, K. Zhou, L. Wang, D. Wang, M. Liao, M. Yang, X. Bai, B. Shi, D. Karatzas, S. Lu, and C.V. Jawahar. ICDAR 2019 robust reading challenge on reading chinese text on signboard. In IC-DAR, pages 1577-1581, 2019. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 315, + 411, + 553, + 454 + ], + "type": "text", + "content": "[56] Z. Zhang, N. Lu, M. Liao, Y. Huang, C. Li, M. Wang, and W. Peng. Self-distillation regularized connectionist temporal classification loss for text recognition: A simple yet effective approach. In AAAI, pages 7441-7449, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 456, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 456, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 315, + 456, + 553, + 498 + ], + "type": "text", + "content": "[57] Y. Zhao and Z. Lian. UDiffText: A unified framework for high-quality text synthesis in arbitrary images via character-aware diffusion models. In ECCV, pages 217-233, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 501, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 501, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 315, + 501, + 553, + 542 + ], + "type": "text", + "content": "[58] T. Zheng, Z. Chen, S. Fang, H. Xie, and Y.-G. Jiang. CDistNet: Perceiving multi-domain character distance for robust text recognition. Int. J. Comput. Vis., 132(2):300-318, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 544, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 544, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 315, + 544, + 553, + 578 + ], + "type": "text", + "content": "[59] Y. Zhu, Z. Li, T. Wang, M. He, and C. Yao. Conditional text image generation with diffusion models. In CVPR, pages 14235-14245, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 579, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 579, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 315, + 579, + 553, + 612 + ], + "type": "text", + "content": "[60] Y. Zhu, J. Liu, F. Gao, W. Liu, X. Wang, P. Wang, F. Huang, C. Yao, and Z. Yang. Visual text generation in the wild. In ECCV, pages 89-106, 2024. 3, 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17473" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/0ee49b9a-7099-4408-8e6f-37bd5bc28baf_content_list.json b/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/0ee49b9a-7099-4408-8e6f-37bd5bc28baf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..60ef66a7c6e5ab4a59212e7750161270b7bf40f9 --- /dev/null +++ b/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/0ee49b9a-7099-4408-8e6f-37bd5bc28baf_content_list.json @@ -0,0 +1,1782 @@ +[ + { + "type": "text", + "text": "Textured 3D Regenerative Morphing with 3D Diffusion Prior", + "text_level": 1, + "bbox": [ + 191, + 130, + 807, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Songlin Yang, Yushi Lan, Honghua Chen, Xingang Pan \nS-Lab, Nanyang Technological University", + "bbox": [ + 277, + 180, + 717, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sl.yang888@outlook.com, {yushi001, honghua.chen, xingang.pan}@ntu.edu.sg", + "bbox": [ + 178, + 219, + 813, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae9c7ebaef496fc140b4e44516f54aaefcc49c649abbb99a12be1a5faed571bd.jpg", + "image_caption": [ + "Figure 1. The textured 3D morphing sequences regenerated using our method with 3D diffusion prior. Our method requires no category-specific alignment data for model training and no labor-intensive preprocessing for explicit correspondence, enabling smooth and plausible morphing across diverse cross-category 3D object pairs. More video results can be found here." + ], + "image_footnote": [], + "bbox": [ + 96, + 247, + 906, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 670, + 326, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Textured 3D morphing creates smooth and plausible interpolation sequences between two 3D objects, focusing on transitions in both shape and texture. This is important for creative applications like visual effects in filmmaking. Previous methods rely on establishing point-to-point correspondences and determining smooth deformation trajectories, which inherently restrict them to shape-only morphing on untextured, topologically aligned datasets. This restriction leads to labor-intensive preprocessing and poor generalization. To overcome these challenges, we propose a method for 3D regenerative morphing using a 3D diffusion prior. Unlike previous methods that depend on explicit correspondences and deformations, our method eliminates", + "bbox": [ + 88, + 704, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the additional need for obtaining correspondence and uses the 3D diffusion prior to generate morphing. Specifically, we first introduce a 3D diffusion model and interpolate the source and target information at three levels: initial noise, model parameters, and condition features. We then explore an Attention Fusion strategy to generate smoother morphing sequences. To further improve the plausibility of semantic interpolation and the generated 3D surfaces, we propose two strategies: (a) Token Reordering, where we match approximate tokens based on semantic analysis to guide implicit correspondences in the denoising process of the diffusion model, and (b) Low-Frequency Enhancement, where we enhance low-frequency signals in the tokens to improve the quality of generated surfaces. Experimental results show that our method achieves superior smoothness and", + "bbox": [ + 511, + 672, + 908, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15159", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plausibility in 3D morphing across diverse cross-category object pairs, offering a novel regenerative method for 3D morphing with textured representations.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 165, + 222, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Morphing [35] generates an interpolation sequence between a source and a target, requiring smooth and plausible transitions. This fundamental technique is crucial for creative applications like visual effects in film and media. Depending on the data type, morphing is categorized into image morphing [4, 7, 18, 23, 34, 54, 80] and 3D morphing [21, 29, 64]. Compared to image morphing, 3D morphing aligns more naturally with visual effects production, where objects and transformations inherently operate in 3D space. However, 3D morphing is more challenging, requiring the interpolation of 3D objects holistically (i.e., image morphing can be viewed as a special case of 3D morphing from a specific viewpoint). Our work addresses the task of textured 3D morphing [64], which generates a sequence for two textured 3D representations, aiming for smooth and plausible transitions in shape and texture, as shown in Fig. 1.", + "bbox": [ + 89, + 191, + 483, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous 3D morphing methods mainly focused on morphing shapes, which can be summarized in two steps: first, establishing correspondence [12, 13, 31, 37] between the 3D representations of the source and target objects, and second, determining smooth and plausible deformation trajectories between corresponding 3D points [17, 61, 78]. Following these steps, previous methods blend the two 3D representations with respective weights to obtain a sequence of interpolated 3D representations.", + "bbox": [ + 89, + 431, + 483, + 568 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the scarcity of topologically aligned and textured 3D datasets, previous 3D morphing methods [17, 61, 78] have mainly focused on in-domain untextured datasets, such as FAUST [9] (human shapes) and Shrec'20 [14] (quadruped animals). As a result, these methods are limited to shape-only morphing with two key generalization challenges: (a) Labor-Intensive Preprocessing: Morphing new in-domain 3D data with these methods [17, 78] requires domain-specific alignment with the training datasets through tedious registration [61] and matching [83] steps. (b) Limited Morphing Capacity: The methods [5, 17, 66, 78] suffer from limited object diversity and small datasets, leading to ambiguous and implausible interpolations.", + "bbox": [ + 89, + 568, + 483, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The limitations of previous methods inspire us to consider two critical questions: (a) Is explicit point-to-point correspondence truly necessary? (b) Can we enhance the generalization capability of textured 3D morphing via a generic generative prior?", + "bbox": [ + 89, + 765, + 483, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For correspondence, obtaining dense correspondences between textured 3D representations across categories remains underexplored [83] despite being foundational to previous methods, while explicit constraints can instead limit", + "bbox": [ + 89, + 840, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the creativity of morphing. Therefore, a promising method is using labor-saving implicit correspondences [31, 74] to guide the morphing. For instance, morphing could be formulated as an optimization problem, allowing correspondences to emerge automatically [64]. Alternatively, attention mechanisms in generative models could enable automatic alignment during object blending [23, 55, 80].", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For generative priors, recent advancements in diffusion-based generation models offer two strategies for 3D morphing: using 3D diffusion models directly [10, 33, 70], or enhancing 2D models with 3D priors. However, those hybrid 2D-3D methods [22, 29, 47, 51] face significant challenges: 2D models lack a holistic 3D knowledge, and optimizing the 2D-3D mapping cannot guarantee that the morphing process will maintain 3D consistency across viewpoints. Therefore, directly using a 3D diffusion model to regenerate the interpolated 3D representations (i.e., regenerative morphing [54, 64]) enables authentic 3D morphing and has the potential to be scaled up by using a more state-of-the-art 3D generation model.", + "bbox": [ + 511, + 199, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the above analysis, we adopt a generic 3D diffusion prior that leverages its implicit correspondence and 3D generation capabilities to blend source and target information, enabling the regeneration of interpolated textured 3D representations.", + "bbox": [ + 511, + 397, + 903, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we first introduce a 3D diffusion model [33] and interpolate the information from the source and target at three levels: initial noises, model parameters, and condition features. The 3D representation for each interpolation is then regenerated using the 3D generation model. To improve the 3D diffusion model's ability to generate smooth morphing sequences, we explore an extended Attention Fusion strategy [23, 80]. However, fusing different information weakened the model's denoising capability, and aggressively applying Attention Fusion to all denoising steps for smoother morphing resulted in implausible outcomes. Therefore, to improve the plausibility of semantic interpolation and the generated 3D surfaces, we propose two strategies: (a) Token Reordering: After semantic analysis, we identify semantic correspondences in diffusion tokens and propose matching approximate tokens before attention computation to better guide implicit correspondences in the diffusion space. (b) Low-Frequency Enhancement: Frequency-domain analysis [57] reveals that boosting low-frequency signals improves the surface quality of the 3D diffusion model. Thus, we enhance low-frequency signals at key time steps to preserve the model's ability to generate 3D surfaces.", + "bbox": [ + 511, + 474, + 903, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows: (a) We are the first to use a generic 3D diffusion prior for morphing textured 3D representations, enabling 3D morphing without explicit correspondences. (b) We analyze the merging of source and target information during morphing with", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15160", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a 3D diffusion prior from semantic and frequency perspectives, proposing Token Reordering and Low-Frequency Enhancement to improve plausibility. (c) Extensive experimental results demonstrate that our method achieves superior smoothness and plausibility in performing 3D morphing across diverse cross-category object pairs.", + "bbox": [ + 89, + 90, + 480, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 196, + 232, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Image Morphing", + "text_level": 1, + "bbox": [ + 89, + 223, + 256, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image morphing is a long-standing challenge in computer vision and graphics [3, 67, 84]. Traditional methods [6, 7, 11, 34, 54] use feature-based warping and blending to create smooth transitions but struggle to generate new content, often leading to artifacts. Data-driven methods [4, 18] leverage large single-class datasets to achieve smoother results but are limited in cross-domain or personalized applications due to their reliance on specific data. The methods like DiffMorpher [80] and AID [23] address this by utilizing pre-trained diffusion models on diverse datasets, enabling flexible morphing across a wide range of object categories. Inputting multi-view images to image morphing methods enables 3D-aware morphing.", + "bbox": [ + 89, + 246, + 482, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.3D Morphing", + "text_level": 1, + "bbox": [ + 89, + 454, + 232, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous 3D morphing methods focus on shape-only correspondence [62], and realize 3D morphing through interpolation or deformation between corresponding 3D primitives (e.g., points, vertices, and faces). They can be divided into (a) Axiomatic Methods: They tend to rely on sparse landmarks [15, 30] or use functional maps [44] to address under-constrained mapping spaces, such as MapTree [50] and SmoothShells [16]. Energy-minimizing functions [59, 60] and skinning methods [19, 28] enable deformation control, while optimal transportation [58, 64] approximates shape correspondence. (b) Learning-Based Methods: This category of methods can be categorized into introducing other generative priors and using aligned data. SATR [1] introduces the semantic labels from multi-view images. NSSM [41] uses DINO [43] for sparse landmarks and SRIF [61] introduces large vision models [80] for semantic shape registration and morphing. A recent class of works leverages text prompts as user inputs for driving a deformation towards an arbitrary textual prompt [20, 39, 40], but CLIP [49] objective lacks a full understanding of object details. For data prior, category-specific training [76] on a topology-aligned dataset such as NeuroMorph [17] is also prevailing in learning-based 3D shape analysis.", + "bbox": [ + 89, + 477, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These methods focus on shape-only morphing or rely on rigging annotations, requiring extra effort due to 2D-3D mismatches. In contrast, our work enables textured 3D morphing without learning correspondence or deformation, using a 3D diffusion model to regenerate interpolated rep", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resentations, addressing previous challenges and offering a new direction for 3D morphing research.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. 3D Diffusion Model", + "text_level": 1, + "bbox": [ + 511, + 136, + 700, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Generative 3D priors fall into two types: native 3D diffusion models and 3D-aware diffusion models. Due to the scarcity of 3D data, methods leveraging 2D priors to generate 3D or multi-view content have been proposed. For instance, Score Distillation Sampling [47] distills 3D information from a 2D diffusion model. 3D-aware generation can be divided into two steps: multi-view image generation [56] followed by feed-forward 3D reconstruction [71]. However, these methods inherently lack a 3D latent space. To address this, native 3D diffusion models [32, 77, 79, 81] that encode and learn 3D representations have been introduced. These models typically consist of two steps: training a VAE to encode 3D data and training a diffusion model based on corresponding latent codes. We use Gaussian Anything [33] as the 3D diffusion prior, encoding 3D information in a point cloud latent space.", + "bbox": [ + 511, + 160, + 906, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 421, + 604, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The framework for 3D regenerative morphing based on the 3D diffusion prior, as shown in Fig. 2, consists of three main steps: Basic Interpolation, where essential information is interpolated (Sec. 3.2); Smoothness Improvement, achieved through an Attention Fusion mechanism (Sec. 3.3); and Plausibility Improvement, which involves two strategies, Token Reordering (Sec. 3.4) and Low-Frequency Enhancement (Sec. 3.5).", + "bbox": [ + 511, + 448, + 906, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 584, + 643, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. 3D Diffusion Model", + "text_level": 1, + "bbox": [ + 511, + 607, + 696, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We select Gaussian Anything [33] as our 3D diffusion prior, a two-stage native 3D diffusion model with a structured latent representation (i.e., point cloud) and the DiT [46] structure. It consists of a geometry generation model $\\epsilon_{G}$ and a texture generation model $\\epsilon_{T}$ . In the first stage, the model $\\epsilon_{G}$ takes a Gaussian initial noise $\\mathbf{z}_{G}$ and textual conditioning information $\\mathbf{c}$ as inputs, to generate a structured point cloud representation $\\mathbf{x}_{\\text {point-cloud }} = \\epsilon_{G}(\\mathbf{z}_{G}, \\mathbf{c})$ . In the second stage, $\\mathbf{x}_{\\text {point-cloud }}$ is added to a Gaussian initial noise $\\mathbf{z}_{T}$ , and the resulting variable is denoised by $\\epsilon_{T}$ with condition $\\mathbf{c}$ to obtain the final texture feature $\\mathbf{x}_{\\text {feature }} = \\epsilon_{T}(\\mathbf{x}_{\\text {point-cloud }} + \\mathbf{z}_{T}, \\mathbf{c})$ . Finally, $\\mathbf{x}_{\\text {point-cloud }}$ and $\\mathbf{x}_{\\text {feature }}$ are fed into a pre-trained decoder $\\mathcal{D}$ to produce the final 3D Gaussian representation $\\mathbf{x}_{3D} = \\mathcal{D}(\\mathbf{x}_{\\text {point-cloud}}, \\mathbf{x}_{\\text {feature}})$ , which can be rendered as multi-view images. Notably, we denote the inherently unordered point-cloud-based token sequences between DiT blocks as $\\{h_{j}\\}_{j=1}^{M}$ , where $M$ is the length of token sequence.", + "bbox": [ + 511, + 628, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15161", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cf650e840532d63ca4a1a001466ed896c27bf33835f6e0f00260cf20ea6343b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 88, + 903, + 238 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/58e45fb395ccbe570fb1b0057a38d046a9cc323574d17958fcf3ac9b85d00759.jpg", + "image_caption": [ + "Figure 2. The framework of our method. Beyond basic interpolation, Attention Fusion is explored to improve smoothness, while Token Reordering and Low-Frequency Enhancement are proposed to improve plausibility." + ], + "image_footnote": [], + "bbox": [ + 98, + 241, + 344, + 366 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ad4b57356c64e20c0552fe9d21127835754d300b14fee184931e2e5150eb4cf9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 348, + 239, + 671, + 366 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4c801e05a74c58d48307c7ebab74d66fdcce3bd3b3d2d8208717aabdd1a1e80b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 241, + 898, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2. Attention", + "text_level": 1, + "bbox": [ + 89, + 409, + 202, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The attention [65] mechanism is an important component for the current text-driven diffusion models [10, 46, 51, 52, 70], especially cross-attention and self-attention. Given a latent variable $z \\in \\mathbb{R}^{d_z}$ , a text condition $c \\in \\mathbb{R}^{d_c}$ , and the attention layer with matrices $W_{Q} \\in \\mathbb{R}^{d_{z} \\times d_{q}}$ , $W_{K} \\in \\mathbb{R}^{d_{c} \\times d_{k}}$ , and $W_{V} \\in \\mathbb{R}^{d_{c} \\times d_{v}}$ , the cross-attention is computed as:", + "bbox": [ + 89, + 426, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA (z, c) = A t t n (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d _ {k}}}\\right) V, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 513, + 482, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q = W_{Q}^{\\top}z, K = W_{K}^{\\top}c, V = W_{V}^{\\top}c$ . Self-attention is a specific form of cross-attention, computed as $A(z,z)$ .", + "bbox": [ + 89, + 546, + 483, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Basic Interpolation", + "text_level": 1, + "bbox": [ + 89, + 585, + 274, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Implementation", + "text_level": 1, + "bbox": [ + 89, + 608, + 248, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The basic interpolation has three levels, with source and target weights given by $(1 - \\alpha)$ and $\\alpha$ , respectively, where $\\alpha = 0$ generates the source $\\mathbf{x}_{3D}^{src}$ , and $\\alpha = 1$ generates the target $\\mathbf{x}_{3D}^{tgt}$ .", + "bbox": [ + 89, + 627, + 483, + 686 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a) Initial Noises: The textured 3D representations $\\mathbf{x}_{3D}^{src}$ and $\\mathbf{x}_{3D}^{tgt}$ , are inverted via diffusion inversion [36] to obtain their respective input noises, $[\\mathbf{z}_T^{src},\\mathbf{z}_G^{src}]$ and $[\\mathbf{z}_T^{tgt},\\mathbf{z}_G^{tgt}]$ . To ensure that Gaussian noise properties are preserved, spherical linear interpolation [53] is applied to these noises to generate the interpolated noises, $[\\mathbf{z}_T^\\alpha ,\\mathbf{z}_G^\\alpha ]$", + "(b) Model Parameters: Given $\\mathbf{x}_{3D}^{src}$ and $\\mathbf{x}_{3D}^{tgt}$ , we fine-tune the model using LoRA (Low-Rank Adaptation) [26] to obtain two sets of LoRA parameters. These parameters are then linearly interpolated and fused to obtain the morphing models $\\epsilon_G^\\alpha$ and $\\epsilon_T^\\alpha$ .", + "(c) Condition Features: We adopt text conditions as they provide a more consistent semantic space and are layout-independent compared to DINO-based [43] image condi" + ], + "bbox": [ + 89, + 686, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tions. The text prompts of the source and target 3D representations are encoded via a CLIP [49] encoder into $\\mathbf{c}^{src}$ and $\\mathbf{c}^{tgt}$ , which are linearly interpolated to produce $\\mathbf{c}^{\\alpha}$ .", + "bbox": [ + 511, + 409, + 905, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Problems", + "text_level": 1, + "bbox": [ + 511, + 465, + 625, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Basic interpolation realizes the blending of basic information, but there are still the following problems:", + "bbox": [ + 511, + 484, + 905, + 515 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a) Abrupt Changes (Smoothness): Nonlinear multi-step denoising in diffusion models introduces variability in noise-to-data mapping, while the CLIP encoder's space lacks guaranteed semantic smoothness, leading to abrupt changes (yellow-highlighted part in Fig. 8).", + "(b) Artifacts (Plausibility): Misalignment between conditioning and diffusion spaces disrupts learned mappings, causing artifacts like structure collapse or degraded surface quality (red and yellowhighlighted parts in Fig. 8)." + ], + "bbox": [ + 511, + 516, + 905, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we investigate Attention Fusion (Sec. 3.3) for smoothness and Token Reordering (Sec. 3.4) with Low-Frequency Enhancement (Sec. 3.5) for plausibility.", + "bbox": [ + 511, + 654, + 905, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Attention Fusion", + "text_level": 1, + "bbox": [ + 511, + 710, + 679, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The fusion of attention has been proven effective in improving morphing smoothness in 2D diffusion-based image morphing. However, DiffMorpher [80] only explored self-attention interpolation and did not consider the smoothness of semantic features in conditioning. AID [23] did not address the differences between the original and LoRA models, neglecting conditioning feature interpolation with accurate inversion. Our method combines self-attention and cross-attention fusion, using unified attention features from fine-tuned models to enhance smoothness while encouraging plausible generation.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15162", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, as shown in Fig. 2, we first feed $z^{src}$ , $z^{tgt}$ and $z^{\\alpha}$ into blocks of the morphing model $\\epsilon^{\\alpha}$ to obtain three sets of $(Q^{src}, K^{src}, V^{src})$ , $(Q^{tgt}, K^{tgt}, V^{tgt})$ , and $(Q^{\\alpha}, K^{\\alpha}, V^{\\alpha})$ . Then, based on Eq. (1), denoting concatenation as $[\\cdot, \\cdot]$ , we obtain the fused attention by:", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF u s e d - A t t n (Q ^ {\\alpha}, K ^ {\\alpha}, V ^ {\\alpha}) = A t t n (Q ^ {\\alpha},\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 172, + 361, + 188 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ (1 - \\alpha) K ^ {\\mathrm {s r c}} + \\alpha K ^ {\\mathrm {t g t}}, K ^ {\\alpha} \\right], \\left[ (1 - \\alpha) V ^ {\\mathrm {s r c}} + \\alpha V ^ {\\mathrm {t g t}}, V ^ {\\alpha} \\right]). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 191, + 482, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the interpolated model for all attention calculations reduces quality deterioration caused by Attention Fusion. However, excessive fusion across time steps harms plausibility, causing structural collapse and surface issues (see Fig. 8).", + "bbox": [ + 89, + 218, + 482, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Token Reordering", + "text_level": 1, + "bbox": [ + 89, + 303, + 266, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1. Motivation", + "text_level": 1, + "bbox": [ + 89, + 324, + 212, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The point clouds of 3D objects are tokenized into sequences $\\{h_j\\}_{j=1}^M$ , with each token $h_j$ representing a 3D point. However, point clouds are inherently unordered representations [77], and relying solely on the attention mechanism to match the points in different 3D objects may lead the model to make semantically implausible connections (e.g., the connected tokens may represent \"chair legs\" and \"donut frosting\", respectively). This vanilla use of the diffusion prior does not fully exploit its potential, as works like DIFT [63] show that 2D diffusion features or tokens represent semantics [24, 75, 82]. We infer that 3D diffusion features also capture semantic correspondences. Therefore, why not first pair points with similar semantics and then interpolate within this semantically plausible space?", + "bbox": [ + 88, + 344, + 482, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2. Experimental Analysis", + "text_level": 1, + "bbox": [ + 89, + 564, + 294, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To test our hypothesis, we first present the problem of position-identical tokens in capturing correspondence (i.e., a problem of vanilla Attention Fusion). We then reveal the emergent correspondence between position-different tokens, supporting the validity of Token Reordering.", + "bbox": [ + 88, + 583, + 482, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(a) Problem of Vanilla Attention Fusion. The DiT models tend to have stronger semantic correspondence in the intermediate time steps [75], and shorter distances between two tokens indicate their higher semantic similarity. To test this in 3D diffusion, we first obtain fully aligned 3D objects by scaling to test our model's correspondence. For instance, as shown in Fig. 3, the right chair results from scaling the left chair's x-axis by 0.6 and z-axis by 1.2. This ensures that tokens at the identical sequence positions represent the originally same point-cloud points (i.e., the points shift positions after scaling while retaining the same color and semantics). Ideally, in geometry diffusion, token distance at the same position should first decrease (semantic learning) until step 200, then increase (point location learning). In texture diffusion, it should gradually decrease, ensuring the color information at the identical position consistent. But", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9c96dc898f1a2d206a32ebb1cbb9811a84095920a79f7f6600360799df986be4.jpg", + "image_caption": [ + "Figure 3. Mean squared error of token distances at identical positions. The solid line represents real data, while the red dashed line shows the ideal trend (i.e., if position-identical tokens can capture the correspondence). Light blue blocks highlight their conflicts." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 751, + 171 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/072599d6a2516dedcb71ec73e934ac11bb0b31909dd8b1447ab9a82777a7a7d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 90, + 903, + 171 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/986c9ee0e976479b4eb1c2c85dd4c169515a3a38aabd73e763b897003bf71453.jpg", + "image_caption": [ + "Figure 4. Mean squared error of point distances between token-distance-closest points. The red dashed line indicates the time step range where semantic correspondence emerges." + ], + "image_footnote": [], + "bbox": [ + 517, + 243, + 746, + 323 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7281a51f332585ab661e4035e74d0fe13156ce8b9f9498a2cdcdf3587e1176aa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 243, + 898, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the Fig. 3 indicates that position-identical semantic alignment is decreasing during denoising, especially the texture diffusion. The vanilla Attention Fusion links these positional tokens, which causes artifacts (see Fig. 8, 3rd row).", + "bbox": [ + 511, + 378, + 903, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(b) Emergence of Semantic Correspondence. By extracting the corresponding points in the final point cloud for two tokens and computing their distance, we can assess the semantic matching of the tokens. We obtain token sequences for the same 3D point cloud through different random seeds, meaning tokens at the identical sequence positions do not correspond to the same point-cloud points. For example, as shown in Fig. 4, the 3D point at the same position may correspond to the $1st$ token (the left chair) and the $M-th$ token (the right chair) respectively. As shown in Fig. 4, the distance between points based on geometry token distance decreases with increasing time steps, indicating emergent semantic correspondence. Conversely, the distance based on texture token distance first decreases and then increases, suggesting that the denoising process initially promotes semantic alignment, but later shifts towards learning texture details. This aligns with findings [75] in 2D diffusion.", + "bbox": [ + 511, + 440, + 905, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, by considering both geometry and texture denoising characteristics, we reorder tokens in the intermediate stages to better guide plausible morphing.", + "bbox": [ + 511, + 696, + 903, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.3. Implementation", + "text_level": 1, + "bbox": [ + 511, + 751, + 669, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on these observations, we reorder the token sequences before passing the output of the $i$ -th block to the $(i + 1)$ -th block, where $i \\in \\{1,2,\\dots,N - 1\\}$ and $N$ is the total number of blocks in the diffusion model. As shown in Fig. 2 and Alg. 1, we reorder the source and target token sequences $\\{h_j^{\\mathrm{src}}\\}_{j = 1}^M$ , $\\{h_j^{\\mathrm{tgt}}\\}_{j = 1}^M$ such that tokens with similar distances align at the same index, as follows:", + "bbox": [ + 511, + 768, + 905, + 867 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {j = 1} {\\text {m i n i m i z e}} \\sum_ {j = 1} ^ {M} \\left\\| h _ {j} ^ {\\mathrm {s r c}} - h _ {\\sigma (j)} ^ {\\mathrm {t g t}} \\right\\|, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 866, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15163", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Token Reordering", + "text_level": 1, + "bbox": [ + 91, + 90, + 302, + 106 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Input: Source token sequence $\\{h_j^{\\mathrm{src}}\\}_{j = 1}^M$ target token sequence $\\{h_j^{\\mathrm{tgt}}\\}_{j = 1}^M$", + "2: Output: Reordered sequences $\\{\\tilde{h}_j^{\\mathrm{src}}\\}_{j = 1}^M$ $\\{\\tilde{h}_j^{\\mathrm{tgt}}\\}_{j = 1}^M$", + "3: for $i = 1$ to $N - 1$ do $\\triangleright$ Efficient computation by PyTorch [45] operations", + "4: Compute $D_{j,k} = \\| h_j^{\\mathrm{src}} - h_k^{\\mathrm{tgt}}\\|$ and $\\sigma (j) = \\arg \\min_{k}D_{j,k},\\forall j$", + "5: if $\\alpha < 0.5$ then", + "6: $\\{\\tilde{h}_j^{\\mathrm{tgt}}\\}_{j = 1}^M\\gets \\{h_{\\sigma (j)}^{\\mathrm{tgt}}\\} ,\\{\\tilde{h}_j^{\\mathrm{src}}\\}_{j = 1}^M\\gets \\{h_j^{\\mathrm{src}}\\}_{j = 1}^M$", + "7:els", + "8: $\\{\\tilde{h}_j^{\\mathrm{src}}\\}_{j = 1}^M\\leftarrow \\{h_{\\sigma -1(j)}^{\\mathrm{src}}\\} ,\\{\\tilde{h}_j^{\\mathrm{tgt}}\\}_{j = 1}^M\\leftarrow \\{h_j^{\\mathrm{tgt}}\\}_{j = 1}^M$", + "9: end if", + "10: end for", + "11: Return $\\{\\tilde{h}_j^{\\mathrm{src}}\\}_{j = 1}^M$ $\\{\\tilde{h}_j^{\\mathrm{tgt}}\\}_{j = 1}^M$" + ], + "bbox": [ + 101, + 109, + 480, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $M$ is the number of tokens, and $\\sigma(j)$ denotes the index of the element in the target sequence that best corresponds to the $j$ -th element in the source sequence. To further refine the generation, we set different reordering strategies depending on $\\alpha$ : For $\\alpha \\in [0, 0.5)$ , the target token sequence is reordered based on the source token sequence. For $\\alpha \\in [0.5, 1]$ , the source token sequence is reordered based on the target token sequence.", + "bbox": [ + 89, + 252, + 483, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Low-Frequency Enhancement", + "text_level": 1, + "bbox": [ + 89, + 381, + 357, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5.1. Motivation", + "text_level": 1, + "bbox": [ + 89, + 402, + 212, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to the additional Attention Fusion operations, the misalignment between the condition space and the diffusion space is exacerbated. Excessive Attention Fusion at later time steps significantly degrades the model's denoising capability, with the deterioration becoming more pronounced as the fusion time step increases (see Fig. 8, from the $4th$ to the $5th$ row). This raises the question: what part of the $3D$ diffusion model is influenced by the morphing operations, leading to this decline in performance?", + "bbox": [ + 89, + 421, + 483, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5.2. Experimental Analysis", + "text_level": 1, + "bbox": [ + 89, + 559, + 294, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The deterioration problem is analyzed in the frequency domain to address the significant visual differences between 3D objects, as shown in Fig. 5. In 3D generation, low-frequency noise controls the overall layout, while high-frequency noise governs surface details. Excessive amplification of high-frequency components during denoising can interfere with the low-frequency components, degrading overall quality. Therefore, enhancing low-frequency signals during denoising is crucial to prevent the model from overemphasizing high-frequency noise, thus improving the quality of 3D surface generation. Similar patterns appear in 2D diffusion studies [48, 57, 69], with low frequencies tied to image layout and high frequencies to details.", + "bbox": [ + 89, + 578, + 483, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5.3. Implementation", + "text_level": 1, + "bbox": [ + 89, + 780, + 248, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address this, as shown in Fig. 2, we propose to enhance the low-frequency signal when generating the 3D interpolations. Specifically, the process is defined as:", + "bbox": [ + 89, + 797, + 482, + 840 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF (h) = \\operatorname {F F T} (h), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 845, + 482, + 861 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\omega < \\omega_ {0}} ^ {\\prime} (h) = F _ {\\omega < \\omega_ {0}} (h) \\odot \\text {s c a l e}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 864, + 480, + 878 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nh ^ {\\prime} = \\operatorname {I F F T} \\left(\\left[ F _ {\\omega < \\omega_ {0}} ^ {\\prime} (h), F _ {\\omega \\geq \\omega_ {0}} (h) \\right]\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 881, + 480, + 897 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/53ffca7b0d30eaded2579e00f4f29007da582c821af681bca5592e63b2dcc9a8.jpg", + "image_caption": [ + "Figure 5. The changes of high and low frequency signals in the diffusion model during the denoising process. Visualizing signal amplitudes at different time steps reveals that low-frequency noise varies less than high-frequency noise, with smaller gaps across denoising time steps." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 710, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b4f7cf3696832a2b5bf1f270d9a9f4014f3a46650eebb584120a1fc69cda84b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 89, + 903, + 189 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b952db4e63fc36a45322068f82a9c2568c1fd11e30fc130e4a5df44347460e2a.jpg", + "image_caption": [ + "Figure 6. The effects of scale for improving surface quality." + ], + "image_footnote": [], + "bbox": [ + 521, + 270, + 903, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $h$ represents the tokens, $F(h)$ their Fourier features, and $h'$ the enhanced tokens, $\\omega$ and $\\omega_0$ denote Fourier frequencies and the threshold, with $\\omega < \\omega_0$ indicating low-frequency components. The scale is the enhancement coefficient, and $\\mathrm{FFT}(\\cdot)$ and $\\mathrm{IFFT}(\\cdot)$ are the Fourier transform and inverse Fourier transform. As shown in Fig. 6, increasing the scale continuously enhances the surface generation capability. However, beyond a certain value, the surface does not increase further with larger scale values.", + "bbox": [ + 511, + 363, + 905, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 516, + 643, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 541, + 728, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.1. Morphing Configuration", + "text_level": 1, + "bbox": [ + 511, + 564, + 732, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All experiments use 250 denoising time steps. Finetuning is performed using LoRA models implemented with PEFT [38] using 500 training steps (rank = 16, alpha = 20, and list = ['to_k', 'to_q', 'to_v', 'qkv']) The initial value of $\\alpha$ is sampled from a Beta distribution, $Beta(5,5)$ , over the $[0,1]$ interval, with 20 points selected as interpolation weights. If multiple morphing is allowed, $\\alpha$ can be concentrated in ranges with more variation by adjusting the Beta distribution or using the reschedule strategy from DiffMorpher [80]. For Attention Fusion, we begin at step 1, with geometry diffusion ending at step 160 and texture diffusion at step 5. Token Reordering starts at step 80 and ends at step 200. Low-Frequency Enhancement occurs between steps 200 and 230, with scale = 5 and $\\omega_0 = 0.1\\pi$ . Note that all our results share the same hyperparameters.", + "bbox": [ + 511, + 584, + 906, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1.2. Baselines", + "text_level": 1, + "bbox": [ + 511, + 820, + 624, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(a) Task (Textured 3D Morphing): MorphFlow [64] directly implements morphing on textured 3D volumetric representations, which has the closest setting to ours; (b) Morphing Tricks: DiffMorpher [80] and AID [23] propose fusion", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15164", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2415ebf24f9023a93bceb563a6c4a28292f4446788cfd42507882e57a9ca4d8.jpg", + "image_caption": [ + "Figure 7. Qualitative comparisons of methods in tasks, morphing tricks, and generative priors. More video results can be found here." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 897, + 483 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9ce5ac50ebd45b4ac82109fc648baeb2402e7537029fb4f29b81e09e153ee161.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons of different methods." + ], + "table_footnote": [], + "table_body": "
Quantitative MetricsUser Study
FID↓KID↓STP-G↑SEP-G↑PPL↓PDV↓STP-U↑SEP-U↑
DiffMorpher218.070.3670.230.135.230.05350.4350.300
AID115.720.1240.670.704.680.01180.3800.505
MV-Adapter120.930.1290.630.577.290.01520.2250.350
Luma95.490.0880.830.777.370.00070.4150.330
MorphFlow147.700.1500.870.903.100.00010.5550.505
Ours6.360.0011.001.003.020.00010.9150.950
", + "bbox": [ + 93, + 532, + 483, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "strategies for attention to enhance smoothness; (c) Generative Priors: We compare the performance of using 2D diffusion (DiffMorpher, AID), multi-view diffusion (MV-Adapter [27]), video generation (Luma [2]), and 3D diffusion as prior in 3D (multi-view image) morphing task.", + "bbox": [ + 89, + 645, + 483, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1.3. Metrics", + "text_level": 1, + "bbox": [ + 89, + 729, + 191, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(a) FID [25] and KID [8]: They are calculated by comparing 3,000 images rendered from original and interpolated 3D representations. (b) Structural Plausibility-GPT (STP-G) and Semantic Plausibility-GPT (SEP-G): GPT-4o [42, 68] evaluates structural and semantic plausibility based on testing results, providing scores and explanations. (c) Perceptual Path Length (PPL) and Perceptual Distance Variance (PDV) [80]: PPL sums perceptual losses over 20-frame sequences, reflecting smoothness and consistency, while PDV measures the variance, indicating transition ho", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mogeneity. (d) Structural Plausibility-User (STP-U) and Semantic Plausibility-User (SEP-U): Volunteers score morphing results based on structural and semantic plausibility.", + "bbox": [ + 511, + 517, + 906, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation", + "text_level": 1, + "bbox": [ + 511, + 571, + 635, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1. Textured 3D Morphing", + "text_level": 1, + "bbox": [ + 511, + 594, + 720, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method adopts a 3D generation prior for 3D morphing, offering two main benefits: direct morphing of textured 3D representations, and maintaining structural and semantic consistency in interpolated representations, as shown in Fig. 1. The closest baseline to our method is MorphFlow [64], which morphs between two 3D volumetric representations using optimal transport optimization. However, as shown in Fig. 7, MorphFlow has two main drawbacks. First, the generated quality is low (see Tab. 1). The advantage of volumetric representations for photorealism is diminished due to the dense interpolation process, which requires morphing even points with no color. Second, it lacks an effective generative prior to the intermediate process, limiting its ability to understand the semantic meaning of intermediate stages. As a result, it often introduces artifacts in morphing scenarios, such as generating six legs instead of four when morphing a bear's and a stool's legs (see the \"Bear-Stool\" case in Fig. 7).", + "bbox": [ + 511, + 613, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Based on the quantitative metrics and user study in the", + "bbox": [ + 532, + 885, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15165", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 1, ours not only significantly outperforms MorphFlow in quality but also demonstrates an impressive \"understanding\" capabilities: (a) the effective mapping of semantically similar parts, as shown when morphing a boot into a red teddy bear (see the \"Boot-Teddy Bear\" case in Fig. 1), where our model smoothly splits the boot into legs and transforms them into the bear's facial features, and (b) minimal disconnected artifacts. The regenerative nature of our method ensures the fusion of source and target while considering the distribution of the entire latent space, minimizing issues like 3D structure collapse or disconnections.", + "bbox": [ + 89, + 90, + 482, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.2. 3D-Aware (Multi-View Image) Morphing", + "text_level": 1, + "bbox": [ + 89, + 263, + 421, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate alternatives to our method for textured 3D morphing from two perspectives. First, 3D morphing can be viewed as multi-view image morphing, where multiview source and target images are fed into image morphing methods for pseudo-3D morphing. Second, many image and video generation models, beyond 3D priors, can perform interpolation tasks. Notably, 2D generation models [27, 56, 72-74], trained on larger datasets often produce 3D-consistent images with superior semantic, structural, and textural understanding compared to 3D models. Therefore, we explore three types of generative priors: 2D diffusion, multi-view diffusion, and video generation.", + "bbox": [ + 89, + 281, + 482, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 7 and Tab. 1, compared to state-of-the-art 2D image morphing methods, such as DiffMorpher [80] and AID [23], 2D diffusion models often suffer from mode collapse when influenced by non-object image regions (e.g., DiffMorpher is sensitive to white backgrounds). Their lack of 3D consistency leads to inconsistent morphing results for the same $\\alpha$ across different viewpoints. When comparing with multi-view diffusion models, we observe that their image-based multi-view generation is limited by pixel-aligned morphing. This limitation becomes apparent when matching pixels across large spatial distances—such as aligning a point on the lower edge of a pumpkin's surface with a point on the stem of a mushroom in the image's center, resulting in interpolation errors (see the \"Pumpkpin-Mushroom\" case in Fig. 7). Lastly, video generation models, while demonstrating strong spatial understanding and achieving morphing by specifying the source and target images as the first and last frames, suffer from limited controllability during generation, often producing incomplete or out-of-frame content. Additionally, the structural consistency of intermediate frames remains suboptimal.", + "bbox": [ + 89, + 463, + 482, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 787, + 243, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We observe that texture diffusion requires minimal Attention Fusion (AF), while geometry diffusion offers a larger range, emphasizing shape understanding in 3D morphing. Thus, we incorporate Attention Fusion in texture diffusion from the first to the fifth time step, while progressively extending the final time step for geometry diffusion. As shown", + "bbox": [ + 89, + 809, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8eb74bb7f786c6684884505d273177cfbb261ea66435feb73bacba3fc182c7fa.jpg", + "image_caption": [ + "Figure 8. Ablation experiments of our proposed strategies." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 915, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0ec6c892d0074bb40b78da4251dad5337163e16cd81cadff1ce9cc05a6ea9860.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation of ablation study." + ], + "table_footnote": [], + "table_body": "
Geometry Diffusion Time Step80 Steps120 Steps160 Steps
FID↓PPL↓PDV↓FID↓PPL↓PDV↓FID↓PPL↓PDV↓
Basic6.386.010.05726.386.010.05726.386.010.0572
Basic+AF9.975.650.054496.324.360.0121105.624.440.0110
Basic+AF+TR8.284.060.00709.703.300.00389.413.370.0024
Basic+AF+LE8.093.760.00497.293.240.00227.083.150.0009
Basic+AF+TR+LE6.403.540.00356.373.210.00196.363.020.0001
", + "bbox": [ + 514, + 311, + 900, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Fig. 8, adding Attention Fusion on top of basic interpolation improves smoothness, but extending the final time step too far causes 3D structural collapse. Token Reordering (TR) helps mitigate this issue, though pushing the time step further reduces generative quality. Ultimately, by applying the Low-Frequency Enhancement (LE) strategy, we balance the smoothness and structural plausibility and ensure all frequencies are maintained effectively. We quantitatively evaluate five different pairs under various settings, as shown in Tab. 2. The ablation study demonstrates that achieving a smooth and plausible morphing sequence requires increasing Attention Fusion steps, as well as applying Token Reordering and Low-Frequency Enhancement.", + "bbox": [ + 511, + 429, + 906, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 643, + 638, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a method to achieve smooth and plausible morphing sequences across diverse cross-category 3D object pairs, incorporating Attention Fusion, Token Reordering, and Low-Frequency Enhancement. This introduces a new paradigm for textured 3D morphing, extending beyond the limitations of previous research confined to shape-only morphing on topologically aligned datasets.", + "bbox": [ + 511, + 672, + 905, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and Future Work. Limited by the performance of existing open-source Text-to-3D models, complex 3D morphing remains challenging. In the future, we will enhance fidelity with more advanced 3D diffusion models [70, 81], enable few-shot motion interpolation [55] with temporal consistency, and explore 4D transitions [21]. Please check our supplementary materials for more implementation details and discussions.", + "bbox": [ + 511, + 779, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15166", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 89, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported by the National Research Foundation, Singapore, under its NRF Fellowship Award (NRF-NRFF16-2024-0003). This research is also supported by MOE AcRF Tier 1 (RG97/23) and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAFICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 232, + 187, + 248 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ahmed Abdelreheem, Ivan Skorokhodov, Maks Ovsjanikov, and Peter Wonka. Satr: Zero-shot semantic segmentation of 3d shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15166-15179, 2023. 3", + "[2] Luma Labs AI. Luma dream machine: Ai-powered video content creation. https://dream-machine.lumalabs.ai/, 2025. Accessed: 2025-01-15. 7", + "[3] Alyaa Qusay Aloraibi. Image morphing techniques: A review. 2023. 3", + "[4] Hadar Averbuch-Elor, Daniel Cohen-Or, and Johannes Kopf. Smooth image sequences for data-driven morphing. In computer graphics forum, pages 203–213. Wiley Online Library, 2016. 2, 3", + "[5] Melike Aydinlar and Yusuf Sahillioğlu. Part-based data-driven 3d shape interpolation. Computer-Aided Design, 136: 103027, 2021. 2", + "[6] Thaddeus Beier and Shawn Neely. Feature-based image metamorphosis. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 529-536. 2023. 3", + "[7] Bhumika G Bhatt. Comparative study of triangulation based and feature based image morphing. Signal & Image Processing, 2(4):235, 2011. 2, 3", + "[8] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 7", + "[9] Federica Bogo, Javier Romero, Matthew Loper, and Michael J Black. Faust: Dataset and evaluation for 3d mesh registration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3794-3801, 2014. 2", + "[10] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024. 2, 4", + "[11] Soheil Darabi, Eli Shechtman, Connelly Barnes, Dan B Goldman, and Pradeep Sen. Image melding: Combining inconsistent images using patch-based synthesis. ACM Transactions on graphics (TOG), 31(4):1-10, 2012. 3", + "[12] Jiacheng Deng, Chuxin Wang, Jiahao Lu, Jianfeng He, Tianzhu Zhang, Jiyang Yu, and Zhe Zhang. Se-ornet: Self-ensembling orientation-aware network for unsupervised point cloud shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364–5373, 2023. 2" + ], + "bbox": [ + 93, + 258, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4494-4504, 2024. 2", + "[14] Roberto M Dyke, Yu-Kun Lai, Paul L Rosin, Stefano Zappala, Seana Dykes, Daoliang Guo, Kun Li, Riccardo Marin, Simone Melzi, and Jingyu Yang. Shrec'20: Shape correspondence with non-isometric deformations. Computers & Graphics, 92:28-43, 2020. 2", + "[15] Michal Edelstein, Danielle Ezuz, and Mirela Ben-Chen. Enigma: Evolutionary non-isometric geometry matching. arXiv preprint arXiv:1905.10763, 2019. 3", + "[16] Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi-scale shape registration with functional maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12265-12274, 2020. 3", + "[17] Marvin Eisenberger, David Novotny, Gael Kerchenbaum, Patrick Labatut, Natalia Neverova, Daniel Cremers, and Andrea Vedaldi. Neuromorph: Unsupervised shape interpolation and correspondence in one go. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7473-7483, 2021. 2, 3", + "[18] Noa Fish, Richard Zhang, Lilach Perry, Daniel Cohen-Or, Eli Shechtman, and Connelly Barnes. Image morphing with perceptual constraints and stn alignment. In Computer Graphics Forum, pages 303–313. Wiley Online Library, 2020. 2, 3", + "[19] Lawson Fulton, Vismay Modi, David Duvenaud, David IW Levin, and Alec Jacobson. Latent-space dynamics for reduced deformable simulation. In Computer graphics forum, pages 379-391. Wiley Online Library, 2019. 3", + "[20] William Gao, Noam Aigerman, Thibault Groueix, Vova Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3", + "[21] Chen Geng, Yunzhi Zhang, Shangzhe Wu, and Jiajun Wu. Birth and death of a rose, 2024. 2, 8", + "[22] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19740–19750, 2023. 2", + "[23] Qiyuan He, Jinghao Wang, Ziwei Liu, and Angela Yao. Aid: Attention interpolation of text-to-image diffusion. NeurIPS, 2024. 2, 3, 4, 6, 8", + "[24] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 36, 2024. 5", + "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[26] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15167", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4", + "[27] Zehuan Huang, Yuanchen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024. 7, 8", + "[28] Alec Jacobson, Zhigang Deng, Ladislav Kavan, and John P Lewis. Skinning: Real-time shape deformation (full text not available). In ACM SIGGRAPH 2014 Courses, pages 1-1. 2014. 3", + "[29] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G Kim, and Rana Hanocka. Meshup: Multi-target mesh deformation via blended score distillation. arXiv preprint arXiv:2408.14899, 2024. 2", + "[30] Vladimir G Kim, Yaron Lipman, and Thomas Funkhouser. Blended intrinsic maps. ACM transactions on graphics (TOG), 30(4):1-12, 2011. 3", + "[31] Yushi Lan, Chen Change Loy, and Bo Dai. DDF: Correspondence distillation from nerf-based gan. IJCV, 2022. 2", + "[32] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2025. 3", + "[33] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. In ICLR, 2025. 2, 3", + "[34] Jing Liao, Rodolfo S Lima, Diego Nehab, Hugues Hoppe, Pedro V Sander, and Jinhui Yu. Automating image morphing using structural similarity on a halfway domain. ACM Transactions on Graphics (TOG), 33(5):1-12, 2014. 2, 3", + "[35] Jianchu Lin, Yinxi Gu, Guangxiao Du, Guoqiang Qu, Xiaobing Chen, Yudong Zhang, Shangbing Gao, Zhen Liu, and Nallappan Gunasekaran. 2d/3d image morphing technology from traditional to modern: A survey. Information Fusion, page 102913, 2024. 2", + "[36] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 4", + "[37] Robin Magnet and Maks Ovsjanikov. Scalable and efficient functional map computations on dense meshes. In Computer Graphics Forum, pages 89-101. Wiley Online Library, 2023. 2", + "[38] Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, and Benjamin Bossan. Peft: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft, 2022.6", + "[39] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3", + "[40] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Luca Morreale, Noam Aigerman, Vladimir G Kim, and Niloy J Mitra. Neural semantic surface maps. In Computer Graphics Forum, page e15005. Wiley Online Library, 2024. 3", + "[42] OpenAI. Gpt-4: Openai's fourth-generation language model, 2023. 7", + "[43] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 4", + "[44] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. ACM Transactions on Graphics (ToG), 31(4):1-11, 2012. 3", + "[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3, 4", + "[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3", + "[48] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. *ICLR*, 2024. 6", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4", + "[50] Jing Ren, Simone Melzi, Maks Ovsjanikov, and Peter Wonka. Maptree: Recovering multiple solutions in the space of maps. ACM Transactions on Graphics, 39(6):1-17, 2020. 3", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4", + "[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 4", + "[53] Dvir Samuel, Rami Ben-Ari, Nir Darshan, Haggai Maron, and Gal Chechik. Norm-guided latent space exploration for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 4" + ], + "bbox": [ + 514, + 90, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15168", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Eli Shechtman, Alex Rav-Acha, Michal Irani, and Steve Seitz. Regenerative morphing. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 615-622. IEEE, 2010. 2, 3", + "[55] Liao Shen, Tianqi Liu, Huiqiang Sun, Xinyi Ye, Baopu Li, Jianming Zhang, and Zhiguo Cao. Dreammover: Leveraging the prior of diffusion models for image interpolation with large motion. arXiv preprint arXiv:2409.09605, 2, 2024. 2, 8", + "[56] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3, 8", + "[57] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2024. 2, 6", + "[58] Justin Solomon, Fernando De Goes, Gabriel Peyre, Marco Cuturi, Adrian Butscher, Andy Nguyen, Tao Du, and Leonidas Guibas. Convolutional wasserstein distances: Efficient optimal transportation on geometric domains. ACM Transactions on Graphics (ToG), 34(4):1-11, 2015. 3", + "[59] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, pages 109–116. CiteSeer, 2007. 3", + "[60] Olga Sorkine, Daniel Cohen-Or, Yaron Lipman, Marc Alexa, Christian Rössl, and H-P Seidel. Laplacian surface editing. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH symposium on Geometry processing, pages 175–184, 2004. 3", + "[61] Mingze Sun, Chen Guo, Puhua Jiang, Shiwei Mao, Yurun Chen, and Ruqi Huang. Srif: Semantic shape registration empowered by diffusion-based image morphing and flow estimation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 2, 3", + "[62] Gary KL Tam, Zhi-Quan Cheng, Yu-Kun Lai, Frank C Langbein, Yonghuai Liu, David Marshall, Ralph R Martin, XianFang Sun, and Paul L Rosin. Registration of 3d point clouds and meshes: A survey from rigid to nonrigid. IEEE transactions on visualization and computer graphics, 19(7):1199-1217, 2012. 3", + "[63] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 5", + "[64] Chih-Jung Tsai, Cheng Sun, and Hwann-Tzong Chen. Multiview regenerative morphing with dual flows. In European Conference on Computer Vision, pages 492-509. Springer, 2022. 2, 3, 6, 7", + "[65] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4", + "[66] Shantanu Vyas, Ting-Ju Chen, Ronak R Mohanty, Peng Jiang, and Vinayak R Krishnamurthy. Latent embedded graphs for image and shape interpolation. Computer-Aided Design, 140:103091, 2021. 2", + "[67] George Wolberg. Image morphing: a survey. The visual computer, 14(8-9):360-372, 1998. 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v(ision) is a human-aligned evaluator for text-to-3d generation. In CVPR, 2024. 7", + "[69] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In European Conference on Computer Vision, pages 378-394. Springer, 2025. 6", + "[70] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2, 4, 8", + "[71] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.3", + "[72] Songlin Yang, Wei Wang, Jun Ling, Bo Peng, Xu Tan, and Jing Dong. Context-aware talking-head video editing. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7718-7727, 2023. 8", + "[73] Songlin Yang, Wei Wang, Bo Peng, and Jing Dong. Designing a 3d-aware stylenerf encoder for face editing. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023.", + "[74] Songlin Yang, Wei Wang, Yushi Lan, Xiangyu Fan, Bo Peng, Lei Yang, and Jing Dong. Learning dense correspondence for nef-based face reenactment. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6522-6530, 2024. 2, 8", + "[75] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 5", + "[76] Mehmet Ersin Yumer, Siddhartha Chaudhuri, Jessica K Hodgins, and Levent Burak Kara. Semantic shape editing using deformation handles. ACM Transactions on Graphics (TOG), 34(4):1-12, 2015. 3", + "[77] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3, 5", + "[78] Xiao Zhan, Rao Fu, and Daniel Ritchie. Charactermixer: Rig-aware interpolation of 3d characters. In Computer Graphics Forum, page e15047. Wiley Online Library, 2024. 2", + "[79] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 3", + "[80] Kaiwen Zhang, Yifan Zhou, Xudong Xu, Bo Dai, and Xinggang Pan. Diffmorpher: Unleashing the capability of diffusion models for image morphing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7912-7921, 2024. 2, 3, 4, 6, 7, 8" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15169", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[81] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 3, 8", + "[82] Zhengbo Zhang, Li Xu, Duo Peng, Hossein Rahmani, and Jun Liu. Diff-tracker: text-to-image diffusion models are unsupervised trackers. In European Conference on Computer Vision, pages 319-337. Springer, 2025. 5", + "[83] Junzhe Zhu, Yuanchen Ju, Junyi Zhang, Muhan Wang, Zhecheng Yuan, Kaizhe Hu, and Huazhe Xu. Dense-matcher: Learning 3d semantic correspondence for category-level manipulation from a single demo. arXiv preprint arXiv:2412.05268, 2024. 2", + "[84] Bhushan Zope and Soniya B Zope. A survey of morphing techniques. International Journal of Advanced Engineering, Management and Science, 3(2):239773, 2017. 3" + ], + "bbox": [ + 91, + 92, + 480, + 330 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "15170", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/0ee49b9a-7099-4408-8e6f-37bd5bc28baf_model.json b/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/0ee49b9a-7099-4408-8e6f-37bd5bc28baf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9c4d245acd67a2759180520b0f69fba5f0e488e0 --- /dev/null +++ b/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/0ee49b9a-7099-4408-8e6f-37bd5bc28baf_model.json @@ -0,0 +1,2809 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.192, + 0.131, + 0.808, + 0.154 + ], + "angle": 0, + "content": "Textured 3D Regenerative Morphing with 3D Diffusion Prior" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.181, + 0.718, + 0.218 + ], + "angle": 0, + "content": "Songlin Yang, Yushi Lan, Honghua Chen, Xingang Pan \nS-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.22, + 0.815, + 0.234 + ], + "angle": 0, + "content": "sl.yang888@outlook.com, {yushi001, honghua.chen, xingang.pan}@ntu.edu.sg" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.248, + 0.907, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.613, + 0.908, + 0.657 + ], + "angle": 0, + "content": "Figure 1. The textured 3D morphing sequences regenerated using our method with 3D diffusion prior. Our method requires no category-specific alignment data for model training and no labor-intensive preprocessing for explicit correspondence, enabling smooth and plausible morphing across diverse cross-category 3D object pairs. More video results can be found here." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.671, + 0.327, + 0.687 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Textured 3D morphing creates smooth and plausible interpolation sequences between two 3D objects, focusing on transitions in both shape and texture. This is important for creative applications like visual effects in filmmaking. Previous methods rely on establishing point-to-point correspondences and determining smooth deformation trajectories, which inherently restrict them to shape-only morphing on untextured, topologically aligned datasets. This restriction leads to labor-intensive preprocessing and poor generalization. To overcome these challenges, we propose a method for 3D regenerative morphing using a 3D diffusion prior. Unlike previous methods that depend on explicit correspondences and deformations, our method eliminates" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.9 + ], + "angle": 0, + "content": "the additional need for obtaining correspondence and uses the 3D diffusion prior to generate morphing. Specifically, we first introduce a 3D diffusion model and interpolate the source and target information at three levels: initial noise, model parameters, and condition features. We then explore an Attention Fusion strategy to generate smoother morphing sequences. To further improve the plausibility of semantic interpolation and the generated 3D surfaces, we propose two strategies: (a) Token Reordering, where we match approximate tokens based on semantic analysis to guide implicit correspondences in the denoising process of the diffusion model, and (b) Low-Frequency Enhancement, where we enhance low-frequency signals in the tokens to improve the quality of generated surfaces. Experimental results show that our method achieves superior smoothness and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15159" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "plausibility in 3D morphing across diverse cross-category object pairs, offering a novel regenerative method for 3D morphing with textured representations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.166, + 0.223, + 0.182 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.192, + 0.485, + 0.432 + ], + "angle": 0, + "content": "Morphing [35] generates an interpolation sequence between a source and a target, requiring smooth and plausible transitions. This fundamental technique is crucial for creative applications like visual effects in film and media. Depending on the data type, morphing is categorized into image morphing [4, 7, 18, 23, 34, 54, 80] and 3D morphing [21, 29, 64]. Compared to image morphing, 3D morphing aligns more naturally with visual effects production, where objects and transformations inherently operate in 3D space. However, 3D morphing is more challenging, requiring the interpolation of 3D objects holistically (i.e., image morphing can be viewed as a special case of 3D morphing from a specific viewpoint). Our work addresses the task of textured 3D morphing [64], which generates a sequence for two textured 3D representations, aiming for smooth and plausible transitions in shape and texture, as shown in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.569 + ], + "angle": 0, + "content": "Previous 3D morphing methods mainly focused on morphing shapes, which can be summarized in two steps: first, establishing correspondence [12, 13, 31, 37] between the 3D representations of the source and target objects, and second, determining smooth and plausible deformation trajectories between corresponding 3D points [17, 61, 78]. Following these steps, previous methods blend the two 3D representations with respective weights to obtain a sequence of interpolated 3D representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.485, + 0.765 + ], + "angle": 0, + "content": "Due to the scarcity of topologically aligned and textured 3D datasets, previous 3D morphing methods [17, 61, 78] have mainly focused on in-domain untextured datasets, such as FAUST [9] (human shapes) and Shrec'20 [14] (quadruped animals). As a result, these methods are limited to shape-only morphing with two key generalization challenges: (a) Labor-Intensive Preprocessing: Morphing new in-domain 3D data with these methods [17, 78] requires domain-specific alignment with the training datasets through tedious registration [61] and matching [83] steps. (b) Limited Morphing Capacity: The methods [5, 17, 66, 78] suffer from limited object diversity and small datasets, leading to ambiguous and implausible interpolations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.484, + 0.841 + ], + "angle": 0, + "content": "The limitations of previous methods inspire us to consider two critical questions: (a) Is explicit point-to-point correspondence truly necessary? (b) Can we enhance the generalization capability of textured 3D morphing via a generic generative prior?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For correspondence, obtaining dense correspondences between textured 3D representations across categories remains underexplored [83] despite being foundational to previous methods, while explicit constraints can instead limit" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "the creativity of morphing. Therefore, a promising method is using labor-saving implicit correspondences [31, 74] to guide the morphing. For instance, morphing could be formulated as an optimization problem, allowing correspondences to emerge automatically [64]. Alternatively, attention mechanisms in generative models could enable automatic alignment during object blending [23, 55, 80]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.905, + 0.395 + ], + "angle": 0, + "content": "For generative priors, recent advancements in diffusion-based generation models offer two strategies for 3D morphing: using 3D diffusion models directly [10, 33, 70], or enhancing 2D models with 3D priors. However, those hybrid 2D-3D methods [22, 29, 47, 51] face significant challenges: 2D models lack a holistic 3D knowledge, and optimizing the 2D-3D mapping cannot guarantee that the morphing process will maintain 3D consistency across viewpoints. Therefore, directly using a 3D diffusion model to regenerate the interpolated 3D representations (i.e., regenerative morphing [54, 64]) enables authentic 3D morphing and has the potential to be scaled up by using a more state-of-the-art 3D generation model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.905, + 0.473 + ], + "angle": 0, + "content": "Building on the above analysis, we adopt a generic 3D diffusion prior that leverages its implicit correspondence and 3D generation capabilities to blend source and target information, enabling the regeneration of interpolated textured 3D representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.905, + 0.823 + ], + "angle": 0, + "content": "Specifically, we first introduce a 3D diffusion model [33] and interpolate the information from the source and target at three levels: initial noises, model parameters, and condition features. The 3D representation for each interpolation is then regenerated using the 3D generation model. To improve the 3D diffusion model's ability to generate smooth morphing sequences, we explore an extended Attention Fusion strategy [23, 80]. However, fusing different information weakened the model's denoising capability, and aggressively applying Attention Fusion to all denoising steps for smoother morphing resulted in implausible outcomes. Therefore, to improve the plausibility of semantic interpolation and the generated 3D surfaces, we propose two strategies: (a) Token Reordering: After semantic analysis, we identify semantic correspondences in diffusion tokens and propose matching approximate tokens before attention computation to better guide implicit correspondences in the diffusion space. (b) Low-Frequency Enhancement: Frequency-domain analysis [57] reveals that boosting low-frequency signals improves the surface quality of the 3D diffusion model. Thus, we enhance low-frequency signals at key time steps to preserve the model's ability to generate 3D surfaces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows: (a) We are the first to use a generic 3D diffusion prior for morphing textured 3D representations, enabling 3D morphing without explicit correspondences. (b) We analyze the merging of source and target information during morphing with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15160" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.184 + ], + "angle": 0, + "content": "a 3D diffusion prior from semantic and frequency perspectives, proposing Token Reordering and Low-Frequency Enhancement to improve plausibility. (c) Extensive experimental results demonstrate that our method achieves superior smoothness and plausibility in performing 3D morphing across diverse cross-category object pairs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.198, + 0.233, + 0.213 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.224, + 0.258, + 0.24 + ], + "angle": 0, + "content": "2.1. Image Morphing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.483, + 0.444 + ], + "angle": 0, + "content": "Image morphing is a long-standing challenge in computer vision and graphics [3, 67, 84]. Traditional methods [6, 7, 11, 34, 54] use feature-based warping and blending to create smooth transitions but struggle to generate new content, often leading to artifacts. Data-driven methods [4, 18] leverage large single-class datasets to achieve smoother results but are limited in cross-domain or personalized applications due to their reliance on specific data. The methods like DiffMorpher [80] and AID [23] address this by utilizing pre-trained diffusion models on diverse datasets, enabling flexible morphing across a wide range of object categories. Inputting multi-view images to image morphing methods enables 3D-aware morphing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.455, + 0.233, + 0.47 + ], + "angle": 0, + "content": "2.2.3D Morphing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Previous 3D morphing methods focus on shape-only correspondence [62], and realize 3D morphing through interpolation or deformation between corresponding 3D primitives (e.g., points, vertices, and faces). They can be divided into (a) Axiomatic Methods: They tend to rely on sparse landmarks [15, 30] or use functional maps [44] to address under-constrained mapping spaces, such as MapTree [50] and SmoothShells [16]. Energy-minimizing functions [59, 60] and skinning methods [19, 28] enable deformation control, while optimal transportation [58, 64] approximates shape correspondence. (b) Learning-Based Methods: This category of methods can be categorized into introducing other generative priors and using aligned data. SATR [1] introduces the semantic labels from multi-view images. NSSM [41] uses DINO [43] for sparse landmarks and SRIF [61] introduces large vision models [80] for semantic shape registration and morphing. A recent class of works leverages text prompts as user inputs for driving a deformation towards an arbitrary textual prompt [20, 39, 40], but CLIP [49] objective lacks a full understanding of object details. For data prior, category-specific training [76] on a topology-aligned dataset such as NeuroMorph [17] is also prevailing in learning-based 3D shape analysis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "These methods focus on shape-only morphing or rely on rigging annotations, requiring extra effort due to 2D-3D mismatches. In contrast, our work enables textured 3D morphing without learning correspondence or deformation, using a 3D diffusion model to regenerate interpolated rep" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "resentations, addressing previous challenges and offering a new direction for 3D morphing research." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.137, + 0.702, + 0.151 + ], + "angle": 0, + "content": "2.3. 3D Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.161, + 0.907, + 0.403 + ], + "angle": 0, + "content": "Generative 3D priors fall into two types: native 3D diffusion models and 3D-aware diffusion models. Due to the scarcity of 3D data, methods leveraging 2D priors to generate 3D or multi-view content have been proposed. For instance, Score Distillation Sampling [47] distills 3D information from a 2D diffusion model. 3D-aware generation can be divided into two steps: multi-view image generation [56] followed by feed-forward 3D reconstruction [71]. However, these methods inherently lack a 3D latent space. To address this, native 3D diffusion models [32, 77, 79, 81] that encode and learn 3D representations have been introduced. These models typically consist of two steps: training a VAE to encode 3D data and training a diffusion model based on corresponding latent codes. We use Gaussian Anything [33] as the 3D diffusion prior, encoding 3D information in a point cloud latent space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.422, + 0.605, + 0.437 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.907, + 0.57 + ], + "angle": 0, + "content": "The framework for 3D regenerative morphing based on the 3D diffusion prior, as shown in Fig. 2, consists of three main steps: Basic Interpolation, where essential information is interpolated (Sec. 3.2); Smoothness Improvement, achieved through an Attention Fusion mechanism (Sec. 3.3); and Plausibility Improvement, which involves two strategies, Token Reordering (Sec. 3.4) and Low-Frequency Enhancement (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.585, + 0.644, + 0.601 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.608, + 0.697, + 0.623 + ], + "angle": 0, + "content": "3.1.1. 3D Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We select Gaussian Anything [33] as our 3D diffusion prior, a two-stage native 3D diffusion model with a structured latent representation (i.e., point cloud) and the DiT [46] structure. It consists of a geometry generation model \\(\\epsilon_{G}\\) and a texture generation model \\(\\epsilon_{T}\\). In the first stage, the model \\(\\epsilon_{G}\\) takes a Gaussian initial noise \\(\\mathbf{z}_{G}\\) and textual conditioning information \\(\\mathbf{c}\\) as inputs, to generate a structured point cloud representation \\(\\mathbf{x}_{\\text {point-cloud }} = \\epsilon_{G}(\\mathbf{z}_{G}, \\mathbf{c})\\). In the second stage, \\(\\mathbf{x}_{\\text {point-cloud }}\\) is added to a Gaussian initial noise \\(\\mathbf{z}_{T}\\), and the resulting variable is denoised by \\(\\epsilon_{T}\\) with condition \\(\\mathbf{c}\\) to obtain the final texture feature \\(\\mathbf{x}_{\\text {feature }} = \\epsilon_{T}(\\mathbf{x}_{\\text {point-cloud }} + \\mathbf{z}_{T}, \\mathbf{c})\\). Finally, \\(\\mathbf{x}_{\\text {point-cloud }}\\) and \\(\\mathbf{x}_{\\text {feature }}\\) are fed into a pre-trained decoder \\(\\mathcal{D}\\) to produce the final 3D Gaussian representation \\(\\mathbf{x}_{3D} = \\mathcal{D}(\\mathbf{x}_{\\text {point-cloud}}, \\mathbf{x}_{\\text {feature}})\\), which can be rendered as multi-view images. Notably, we denote the inherently unordered point-cloud-based token sequences between DiT blocks as \\(\\{h_{j}\\}_{j=1}^{M}\\), where \\(M\\) is the length of token sequence." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15161" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.905, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.242, + 0.345, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.241, + 0.672, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.242, + 0.9, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.373, + 0.908, + 0.403 + ], + "angle": 0, + "content": "Figure 2. The framework of our method. Beyond basic interpolation, Attention Fusion is explored to improve smoothness, while Token Reordering and Low-Frequency Enhancement are proposed to improve plausibility." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.41, + 0.204, + 0.423 + ], + "angle": 0, + "content": "3.1.2. Attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.428, + 0.484, + 0.514 + ], + "angle": 0, + "content": "The attention [65] mechanism is an important component for the current text-driven diffusion models [10, 46, 51, 52, 70], especially cross-attention and self-attention. Given a latent variable \\(z \\in \\mathbb{R}^{d_z}\\), a text condition \\(c \\in \\mathbb{R}^{d_c}\\), and the attention layer with matrices \\(W_{Q} \\in \\mathbb{R}^{d_{z} \\times d_{q}}\\), \\(W_{K} \\in \\mathbb{R}^{d_{c} \\times d_{k}}\\), and \\(W_{V} \\in \\mathbb{R}^{d_{c} \\times d_{v}}\\), the cross-attention is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.515, + 0.483, + 0.547 + ], + "angle": 0, + "content": "\\[\nA (z, c) = A t t n (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d _ {k}}}\\right) V, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.579 + ], + "angle": 0, + "content": "where \\(Q = W_{Q}^{\\top}z, K = W_{K}^{\\top}c, V = W_{V}^{\\top}c\\). Self-attention is a specific form of cross-attention, computed as \\(A(z,z)\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.275, + 0.603 + ], + "angle": 0, + "content": "3.2. Basic Interpolation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.609, + 0.249, + 0.624 + ], + "angle": 0, + "content": "3.2.1. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.688 + ], + "angle": 0, + "content": "The basic interpolation has three levels, with source and target weights given by \\((1 - \\alpha)\\) and \\(\\alpha\\), respectively, where \\(\\alpha = 0\\) generates the source \\(\\mathbf{x}_{3D}^{src}\\), and \\(\\alpha = 1\\) generates the target \\(\\mathbf{x}_{3D}^{tgt}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.781 + ], + "angle": 0, + "content": "(a) Initial Noises: The textured 3D representations \\(\\mathbf{x}_{3D}^{src}\\) and \\(\\mathbf{x}_{3D}^{tgt}\\), are inverted via diffusion inversion [36] to obtain their respective input noises, \\([\\mathbf{z}_T^{src},\\mathbf{z}_G^{src}]\\) and \\([\\mathbf{z}_T^{tgt},\\mathbf{z}_G^{tgt}]\\). To ensure that Gaussian noise properties are preserved, spherical linear interpolation [53] is applied to these noises to generate the interpolated noises, \\([\\mathbf{z}_T^\\alpha ,\\mathbf{z}_G^\\alpha ]\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.856 + ], + "angle": 0, + "content": "(b) Model Parameters: Given \\(\\mathbf{x}_{3D}^{src}\\) and \\(\\mathbf{x}_{3D}^{tgt}\\), we fine-tune the model using LoRA (Low-Rank Adaptation) [26] to obtain two sets of LoRA parameters. These parameters are then linearly interpolated and fused to obtain the morphing models \\(\\epsilon_G^\\alpha\\) and \\(\\epsilon_T^\\alpha\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "(c) Condition Features: We adopt text conditions as they provide a more consistent semantic space and are layout-independent compared to DINO-based [43] image condi" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.688, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.906, + 0.456 + ], + "angle": 0, + "content": "tions. The text prompts of the source and target 3D representations are encoded via a CLIP [49] encoder into \\(\\mathbf{c}^{src}\\) and \\(\\mathbf{c}^{tgt}\\), which are linearly interpolated to produce \\(\\mathbf{c}^{\\alpha}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.466, + 0.627, + 0.48 + ], + "angle": 0, + "content": "3.2.2. Problems" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.906, + 0.516 + ], + "angle": 0, + "content": "Basic interpolation realizes the blending of basic information, but there are still the following problems:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.906, + 0.592 + ], + "angle": 0, + "content": "(a) Abrupt Changes (Smoothness): Nonlinear multi-step denoising in diffusion models introduces variability in noise-to-data mapping, while the CLIP encoder's space lacks guaranteed semantic smoothness, leading to abrupt changes (yellow-highlighted part in Fig. 8)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.906, + 0.655 + ], + "angle": 0, + "content": "(b) Artifacts (Plausibility): Misalignment between conditioning and diffusion spaces disrupts learned mappings, causing artifacts like structure collapse or degraded surface quality (red and yellowhighlighted parts in Fig. 8)." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.517, + 0.906, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.906, + 0.701 + ], + "angle": 0, + "content": "Therefore, we investigate Attention Fusion (Sec. 3.3) for smoothness and Token Reordering (Sec. 3.4) with Low-Frequency Enhancement (Sec. 3.5) for plausibility." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.681, + 0.727 + ], + "angle": 0, + "content": "3.3. Attention Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The fusion of attention has been proven effective in improving morphing smoothness in 2D diffusion-based image morphing. However, DiffMorpher [80] only explored self-attention interpolation and did not consider the smoothness of semantic features in conditioning. AID [23] did not address the differences between the original and LoRA models, neglecting conditioning feature interpolation with accurate inversion. Our method combines self-attention and cross-attention fusion, using unified attention features from fine-tuned models to enhance smoothness while encouraging plausible generation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15162" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.091, + 0.483, + 0.168 + ], + "angle": 0, + "content": "Specifically, as shown in Fig. 2, we first feed \\(z^{src}\\), \\(z^{tgt}\\) and \\(z^{\\alpha}\\) into blocks of the morphing model \\(\\epsilon^{\\alpha}\\) to obtain three sets of \\((Q^{src}, K^{src}, V^{src})\\), \\((Q^{tgt}, K^{tgt}, V^{tgt})\\), and \\((Q^{\\alpha}, K^{\\alpha}, V^{\\alpha})\\). Then, based on Eq. (1), denoting concatenation as \\([\\cdot, \\cdot]\\), we obtain the fused attention by:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.174, + 0.362, + 0.189 + ], + "angle": 0, + "content": "\\[\nF u s e d - A t t n (Q ^ {\\alpha}, K ^ {\\alpha}, V ^ {\\alpha}) = A t t n (Q ^ {\\alpha},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.192, + 0.483, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\left[ (1 - \\alpha) K ^ {\\mathrm {s r c}} + \\alpha K ^ {\\mathrm {t g t}}, K ^ {\\alpha} \\right], \\left[ (1 - \\alpha) V ^ {\\mathrm {s r c}} + \\alpha V ^ {\\mathrm {t g t}}, V ^ {\\alpha} \\right]). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.219, + 0.483, + 0.294 + ], + "angle": 0, + "content": "Using the interpolated model for all attention calculations reduces quality deterioration caused by Attention Fusion. However, excessive fusion across time steps harms plausibility, causing structural collapse and surface issues (see Fig. 8)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.304, + 0.267, + 0.319 + ], + "angle": 0, + "content": "3.4. Token Reordering" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.325, + 0.214, + 0.339 + ], + "angle": 0, + "content": "3.4.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.345, + 0.483, + 0.556 + ], + "angle": 0, + "content": "The point clouds of 3D objects are tokenized into sequences \\(\\{h_j\\}_{j=1}^M\\), with each token \\(h_j\\) representing a 3D point. However, point clouds are inherently unordered representations [77], and relying solely on the attention mechanism to match the points in different 3D objects may lead the model to make semantically implausible connections (e.g., the connected tokens may represent \"chair legs\" and \"donut frosting\", respectively). This vanilla use of the diffusion prior does not fully exploit its potential, as works like DIFT [63] show that 2D diffusion features or tokens represent semantics [24, 75, 82]. We infer that 3D diffusion features also capture semantic correspondences. Therefore, why not first pair points with similar semantics and then interpolate within this semantically plausible space?" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.565, + 0.295, + 0.58 + ], + "angle": 0, + "content": "3.4.2. Experimental Analysis" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.483, + 0.659 + ], + "angle": 0, + "content": "To test our hypothesis, we first present the problem of position-identical tokens in capturing correspondence (i.e., a problem of vanilla Attention Fusion). We then reveal the emergent correspondence between position-different tokens, supporting the validity of Token Reordering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "(a) Problem of Vanilla Attention Fusion. The DiT models tend to have stronger semantic correspondence in the intermediate time steps [75], and shorter distances between two tokens indicate their higher semantic similarity. To test this in 3D diffusion, we first obtain fully aligned 3D objects by scaling to test our model's correspondence. For instance, as shown in Fig. 3, the right chair results from scaling the left chair's x-axis by 0.6 and z-axis by 1.2. This ensures that tokens at the identical sequence positions represent the originally same point-cloud points (i.e., the points shift positions after scaling while retaining the same color and semantics). Ideally, in geometry diffusion, token distance at the same position should first decrease (semantic learning) until step 200, then increase (point location learning). In texture diffusion, it should gradually decrease, ensuring the color information at the identical position consistent. But" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.752, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.092, + 0.905, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.179, + 0.905, + 0.235 + ], + "angle": 0, + "content": "Figure 3. Mean squared error of token distances at identical positions. The solid line represents real data, while the red dashed line shows the ideal trend (i.e., if position-identical tokens can capture the correspondence). Light blue blocks highlight their conflicts." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.244, + 0.748, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.244, + 0.9, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.331, + 0.905, + 0.372 + ], + "angle": 0, + "content": "Figure 4. Mean squared error of point distances between token-distance-closest points. The red dashed line indicates the time step range where semantic correspondence emerges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.905, + 0.44 + ], + "angle": 0, + "content": "the Fig. 3 indicates that position-identical semantic alignment is decreasing during denoising, especially the texture diffusion. The vanilla Attention Fusion links these positional tokens, which causes artifacts (see Fig. 8, 3rd row)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.906, + 0.698 + ], + "angle": 0, + "content": "(b) Emergence of Semantic Correspondence. By extracting the corresponding points in the final point cloud for two tokens and computing their distance, we can assess the semantic matching of the tokens. We obtain token sequences for the same 3D point cloud through different random seeds, meaning tokens at the identical sequence positions do not correspond to the same point-cloud points. For example, as shown in Fig. 4, the 3D point at the same position may correspond to the \\(1st\\) token (the left chair) and the \\(M-th\\) token (the right chair) respectively. As shown in Fig. 4, the distance between points based on geometry token distance decreases with increasing time steps, indicating emergent semantic correspondence. Conversely, the distance based on texture token distance first decreases and then increases, suggesting that the denoising process initially promotes semantic alignment, but later shifts towards learning texture details. This aligns with findings [75] in 2D diffusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.905, + 0.744 + ], + "angle": 0, + "content": "Therefore, by considering both geometry and texture denoising characteristics, we reorder tokens in the intermediate stages to better guide plausible morphing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.752, + 0.671, + 0.767 + ], + "angle": 0, + "content": "3.4.3. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.77, + 0.906, + 0.868 + ], + "angle": 0, + "content": "Based on these observations, we reorder the token sequences before passing the output of the \\(i\\)-th block to the \\((i + 1)\\)-th block, where \\(i \\in \\{1,2,\\dots,N - 1\\}\\) and \\(N\\) is the total number of blocks in the diffusion model. As shown in Fig. 2 and Alg. 1, we reorder the source and target token sequences \\(\\{h_j^{\\mathrm{src}}\\}_{j = 1}^M\\), \\(\\{h_j^{\\mathrm{tgt}}\\}_{j = 1}^M\\) such that tokens with similar distances align at the same index, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.867, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\underset {j = 1} {\\text {m i n i m i z e}} \\sum_ {j = 1} ^ {M} \\left\\| h _ {j} ^ {\\mathrm {s r c}} - h _ {\\sigma (j)} ^ {\\mathrm {t g t}} \\right\\|, \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15163" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.092, + 0.303, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Token Reordering" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.111, + 0.473, + 0.126 + ], + "angle": 0, + "content": "1: Input: Source token sequence \\(\\{h_j^{\\mathrm{src}}\\}_{j = 1}^M\\) target token sequence \\(\\{h_j^{\\mathrm{tgt}}\\}_{j = 1}^M\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.126, + 0.373, + 0.139 + ], + "angle": 0, + "content": "2: Output: Reordered sequences \\(\\{\\tilde{h}_j^{\\mathrm{src}}\\}_{j = 1}^M\\) \\(\\{\\tilde{h}_j^{\\mathrm{tgt}}\\}_{j = 1}^M\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.139, + 0.482, + 0.15 + ], + "angle": 0, + "content": "3: for \\( i = 1 \\) to \\( N - 1 \\) do \\( \\triangleright \\) Efficient computation by PyTorch [45] operations" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.15, + 0.456, + 0.161 + ], + "angle": 0, + "content": "4: Compute \\( D_{j,k} = \\| h_j^{\\mathrm{src}} - h_k^{\\mathrm{tgt}}\\| \\) and \\( \\sigma (j) = \\arg \\min_{k}D_{j,k},\\forall j \\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.162, + 0.218, + 0.171 + ], + "angle": 0, + "content": "5: if \\(\\alpha < 0.5\\) then" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.171, + 0.398, + 0.185 + ], + "angle": 0, + "content": "6: \\(\\{\\tilde{h}_j^{\\mathrm{tgt}}\\}_{j = 1}^M\\gets \\{h_{\\sigma (j)}^{\\mathrm{tgt}}\\} ,\\{\\tilde{h}_j^{\\mathrm{src}}\\}_{j = 1}^M\\gets \\{h_j^{\\mathrm{src}}\\}_{j = 1}^M\\)" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.185, + 0.151, + 0.193 + ], + "angle": 0, + "content": "7:els" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.193, + 0.412, + 0.208 + ], + "angle": 0, + "content": "8: \\(\\{\\tilde{h}_j^{\\mathrm{src}}\\}_{j = 1}^M\\leftarrow \\{h_{\\sigma -1(j)}^{\\mathrm{src}}\\} ,\\{\\tilde{h}_j^{\\mathrm{tgt}}\\}_{j = 1}^M\\leftarrow \\{h_j^{\\mathrm{tgt}}\\}_{j = 1}^M\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.209, + 0.169, + 0.217 + ], + "angle": 0, + "content": "9: end if" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.218, + 0.158, + 0.226 + ], + "angle": 0, + "content": "10: end for" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.227, + 0.266, + 0.242 + ], + "angle": 0, + "content": "11: Return \\(\\{\\tilde{h}_j^{\\mathrm{src}}\\}_{j = 1}^M\\) \\(\\{\\tilde{h}_j^{\\mathrm{tgt}}\\}_{j = 1}^M\\)" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.111, + 0.482, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.253, + 0.485, + 0.375 + ], + "angle": 0, + "content": "where \\(M\\) is the number of tokens, and \\(\\sigma(j)\\) denotes the index of the element in the target sequence that best corresponds to the \\(j\\)-th element in the source sequence. To further refine the generation, we set different reordering strategies depending on \\(\\alpha\\): For \\(\\alpha \\in [0, 0.5)\\), the target token sequence is reordered based on the source token sequence. For \\(\\alpha \\in [0.5, 1]\\), the source token sequence is reordered based on the target token sequence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.382, + 0.359, + 0.398 + ], + "angle": 0, + "content": "3.5. Low-Frequency Enhancement" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.404, + 0.214, + 0.417 + ], + "angle": 0, + "content": "3.5.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.558 + ], + "angle": 0, + "content": "Due to the additional Attention Fusion operations, the misalignment between the condition space and the diffusion space is exacerbated. Excessive Attention Fusion at later time steps significantly degrades the model's denoising capability, with the deterioration becoming more pronounced as the fusion time step increases (see Fig. 8, from the \\(4th\\) to the \\(5th\\) row). This raises the question: what part of the \\(3D\\) diffusion model is influenced by the morphing operations, leading to this decline in performance?" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.295, + 0.575 + ], + "angle": 0, + "content": "3.5.2. Experimental Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.484, + 0.776 + ], + "angle": 0, + "content": "The deterioration problem is analyzed in the frequency domain to address the significant visual differences between 3D objects, as shown in Fig. 5. In 3D generation, low-frequency noise controls the overall layout, while high-frequency noise governs surface details. Excessive amplification of high-frequency components during denoising can interfere with the low-frequency components, degrading overall quality. Therefore, enhancing low-frequency signals during denoising is crucial to prevent the model from overemphasizing high-frequency noise, thus improving the quality of 3D surface generation. Similar patterns appear in 2D diffusion studies [48, 57, 69], with low frequencies tied to image layout and high frequencies to details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.781, + 0.249, + 0.795 + ], + "angle": 0, + "content": "3.5.3. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.798, + 0.483, + 0.841 + ], + "angle": 0, + "content": "To address this, as shown in Fig. 2, we propose to enhance the low-frequency signal when generating the 3D interpolations. Specifically, the process is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.847, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\nF (h) = \\operatorname {F F T} (h), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.865, + 0.482, + 0.88 + ], + "angle": 0, + "content": "\\[\nF _ {\\omega < \\omega_ {0}} ^ {\\prime} (h) = F _ {\\omega < \\omega_ {0}} (h) \\odot \\text {s c a l e}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.882, + 0.482, + 0.898 + ], + "angle": 0, + "content": "\\[\nh ^ {\\prime} = \\operatorname {I F F T} \\left(\\left[ F _ {\\omega < \\omega_ {0}} ^ {\\prime} (h), F _ {\\omega \\geq \\omega_ {0}} (h) \\right]\\right), \\tag {6}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.712, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.09, + 0.905, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.195, + 0.906, + 0.265 + ], + "angle": 0, + "content": "Figure 5. The changes of high and low frequency signals in the diffusion model during the denoising process. Visualizing signal amplitudes at different time steps reveals that low-frequency noise varies less than high-frequency noise, with smaller gaps across denoising time steps." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.271, + 0.905, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.343, + 0.888, + 0.357 + ], + "angle": 0, + "content": "Figure 6. The effects of scale for improving surface quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.501 + ], + "angle": 0, + "content": "where \\( h \\) represents the tokens, \\( F(h) \\) their Fourier features, and \\( h' \\) the enhanced tokens, \\( \\omega \\) and \\( \\omega_0 \\) denote Fourier frequencies and the threshold, with \\( \\omega < \\omega_0 \\) indicating low-frequency components. The scale is the enhancement coefficient, and \\( \\mathrm{FFT}(\\cdot) \\) and \\( \\mathrm{IFFT}(\\cdot) \\) are the Fourier transform and inverse Fourier transform. As shown in Fig. 6, increasing the scale continuously enhances the surface generation capability. However, beyond a certain value, the surface does not increase further with larger scale values." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.645, + 0.533 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.73, + 0.558 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.565, + 0.733, + 0.58 + ], + "angle": 0, + "content": "4.1.1. Morphing Configuration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.812 + ], + "angle": 0, + "content": "All experiments use 250 denoising time steps. Finetuning is performed using LoRA models implemented with PEFT [38] using 500 training steps (rank = 16, alpha = 20, and list = ['to_k', 'to_q', 'to_v', 'qkv']) The initial value of \\(\\alpha\\) is sampled from a Beta distribution, \\(Beta(5,5)\\), over the \\([0,1]\\) interval, with 20 points selected as interpolation weights. If multiple morphing is allowed, \\(\\alpha\\) can be concentrated in ranges with more variation by adjusting the Beta distribution or using the reschedule strategy from DiffMorpher [80]. For Attention Fusion, we begin at step 1, with geometry diffusion ending at step 160 and texture diffusion at step 5. Token Reordering starts at step 80 and ends at step 200. Low-Frequency Enhancement occurs between steps 200 and 230, with scale = 5 and \\(\\omega_0 = 0.1\\pi\\). Note that all our results share the same hyperparameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.821, + 0.625, + 0.834 + ], + "angle": 0, + "content": "4.1.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "(a) Task (Textured 3D Morphing): MorphFlow [64] directly implements morphing on textured 3D volumetric representations, which has the closest setting to ours; (b) Morphing Tricks: DiffMorpher [80] and AID [23] propose fusion" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15164" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.898, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.491, + 0.893, + 0.507 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparisons of methods in tasks, morphing tricks, and generative priors. More video results can be found here." + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.516, + 0.454, + 0.53 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons of different methods." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.534, + 0.484, + 0.638 + ], + "angle": 0, + "content": "
Quantitative MetricsUser Study
FID↓KID↓STP-G↑SEP-G↑PPL↓PDV↓STP-U↑SEP-U↑
DiffMorpher218.070.3670.230.135.230.05350.4350.300
AID115.720.1240.670.704.680.01180.3800.505
MV-Adapter120.930.1290.630.577.290.01520.2250.350
Luma95.490.0880.830.777.370.00070.4150.330
MorphFlow147.700.1500.870.903.100.00010.5550.505
Ours6.360.0011.001.003.020.00010.9150.950
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.722 + ], + "angle": 0, + "content": "strategies for attention to enhance smoothness; (c) Generative Priors: We compare the performance of using 2D diffusion (DiffMorpher, AID), multi-view diffusion (MV-Adapter [27]), video generation (Luma [2]), and 3D diffusion as prior in 3D (multi-view image) morphing task." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.731, + 0.192, + 0.744 + ], + "angle": 0, + "content": "4.1.3. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "(a) FID [25] and KID [8]: They are calculated by comparing 3,000 images rendered from original and interpolated 3D representations. (b) Structural Plausibility-GPT (STP-G) and Semantic Plausibility-GPT (SEP-G): GPT-4o [42, 68] evaluates structural and semantic plausibility based on testing results, providing scores and explanations. (c) Perceptual Path Length (PPL) and Perceptual Distance Variance (PDV) [80]: PPL sums perceptual losses over 20-frame sequences, reflecting smoothness and consistency, while PDV measures the variance, indicating transition ho" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.565 + ], + "angle": 0, + "content": "mogeneity. (d) Structural Plausibility-User (STP-U) and Semantic Plausibility-User (SEP-U): Volunteers score morphing results based on structural and semantic plausibility." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.573, + 0.637, + 0.587 + ], + "angle": 0, + "content": "4.2. Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.595, + 0.722, + 0.611 + ], + "angle": 0, + "content": "4.2.1. Textured 3D Morphing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Our method adopts a 3D generation prior for 3D morphing, offering two main benefits: direct morphing of textured 3D representations, and maintaining structural and semantic consistency in interpolated representations, as shown in Fig. 1. The closest baseline to our method is MorphFlow [64], which morphs between two 3D volumetric representations using optimal transport optimization. However, as shown in Fig. 7, MorphFlow has two main drawbacks. First, the generated quality is low (see Tab. 1). The advantage of volumetric representations for photorealism is diminished due to the dense interpolation process, which requires morphing even points with no color. Second, it lacks an effective generative prior to the intermediate process, limiting its ability to understand the semantic meaning of intermediate stages. As a result, it often introduces artifacts in morphing scenarios, such as generating six legs instead of four when morphing a bear's and a stool's legs (see the \"Bear-Stool\" case in Fig. 7)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Based on the quantitative metrics and user study in the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15165" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.258 + ], + "angle": 0, + "content": "Tab. 1, ours not only significantly outperforms MorphFlow in quality but also demonstrates an impressive \"understanding\" capabilities: (a) the effective mapping of semantically similar parts, as shown when morphing a boot into a red teddy bear (see the \"Boot-Teddy Bear\" case in Fig. 1), where our model smoothly splits the boot into legs and transforms them into the bear's facial features, and (b) minimal disconnected artifacts. The regenerative nature of our method ensures the fusion of source and target while considering the distribution of the entire latent space, minimizing issues like 3D structure collapse or disconnections." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.264, + 0.423, + 0.28 + ], + "angle": 0, + "content": "4.2.2. 3D-Aware (Multi-View Image) Morphing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.463 + ], + "angle": 0, + "content": "We evaluate alternatives to our method for textured 3D morphing from two perspectives. First, 3D morphing can be viewed as multi-view image morphing, where multiview source and target images are fed into image morphing methods for pseudo-3D morphing. Second, many image and video generation models, beyond 3D priors, can perform interpolation tasks. Notably, 2D generation models [27, 56, 72-74], trained on larger datasets often produce 3D-consistent images with superior semantic, structural, and textural understanding compared to 3D models. Therefore, we explore three types of generative priors: 2D diffusion, multi-view diffusion, and video generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.781 + ], + "angle": 0, + "content": "As shown in Fig. 7 and Tab. 1, compared to state-of-the-art 2D image morphing methods, such as DiffMorpher [80] and AID [23], 2D diffusion models often suffer from mode collapse when influenced by non-object image regions (e.g., DiffMorpher is sensitive to white backgrounds). Their lack of 3D consistency leads to inconsistent morphing results for the same \\(\\alpha\\) across different viewpoints. When comparing with multi-view diffusion models, we observe that their image-based multi-view generation is limited by pixel-aligned morphing. This limitation becomes apparent when matching pixels across large spatial distances—such as aligning a point on the lower edge of a pumpkin's surface with a point on the stem of a mushroom in the image's center, resulting in interpolation errors (see the \"Pumpkpin-Mushroom\" case in Fig. 7). Lastly, video generation models, while demonstrating strong spatial understanding and achieving morphing by specifying the source and target images as the first and last frames, suffer from limited controllability during generation, often producing incomplete or out-of-frame content. Additionally, the structural consistency of intermediate frames remains suboptimal." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.245, + 0.805 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We observe that texture diffusion requires minimal Attention Fusion (AF), while geometry diffusion offers a larger range, emphasizing shape understanding in 3D morphing. Thus, we incorporate Attention Fusion in texture diffusion from the first to the fifth time step, while progressively extending the final time step for geometry diffusion. As shown" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.916, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.272, + 0.882, + 0.286 + ], + "angle": 0, + "content": "Figure 8. Ablation experiments of our proposed strategies." + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.294, + 0.858, + 0.307 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation of ablation study." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.312, + 0.901, + 0.406 + ], + "angle": 0, + "content": "
Geometry Diffusion Time Step80 Steps120 Steps160 Steps
FID↓PPL↓PDV↓FID↓PPL↓PDV↓FID↓PPL↓PDV↓
Basic6.386.010.05726.386.010.05726.386.010.0572
Basic+AF9.975.650.054496.324.360.0121105.624.440.0110
Basic+AF+TR8.284.060.00709.703.300.00389.413.370.0024
Basic+AF+LE8.093.760.00497.293.240.00227.083.150.0009
Basic+AF+TR+LE6.403.540.00356.373.210.00196.363.020.0001
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.626 + ], + "angle": 0, + "content": "in Fig. 8, adding Attention Fusion on top of basic interpolation improves smoothness, but extending the final time step too far causes 3D structural collapse. Token Reordering (TR) helps mitigate this issue, though pushing the time step further reduces generative quality. Ultimately, by applying the Low-Frequency Enhancement (LE) strategy, we balance the smoothness and structural plausibility and ensure all frequencies are maintained effectively. We quantitatively evaluate five different pairs under various settings, as shown in Tab. 2. The ablation study demonstrates that achieving a smooth and plausible morphing sequence requires increasing Attention Fusion steps, as well as applying Token Reordering and Low-Frequency Enhancement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.645, + 0.64, + 0.661 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.906, + 0.779 + ], + "angle": 0, + "content": "We propose a method to achieve smooth and plausible morphing sequences across diverse cross-category 3D object pairs, incorporating Attention Fusion, Token Reordering, and Low-Frequency Enhancement. This introduces a new paradigm for textured 3D morphing, extending beyond the limitations of previous research confined to shape-only morphing on topologically aligned datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Limitation and Future Work. Limited by the performance of existing open-source Text-to-3D models, complex 3D morphing remains challenging. In the future, we will enhance fidelity with more advanced 3D diffusion models [70, 81], enable few-shot motion interpolation [55] with temporal consistency, and explore 4D transitions [21]. Please check our supplementary materials for more implementation details and discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15166" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This research is supported by the National Research Foundation, Singapore, under its NRF Fellowship Award (NRF-NRFF16-2024-0003). This research is also supported by MOE AcRF Tier 1 (RG97/23) and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAFICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.233, + 0.188, + 0.249 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[1] Ahmed Abdelreheem, Ivan Skorokhodov, Maks Ovsjanikov, and Peter Wonka. Satr: Zero-shot semantic segmentation of 3d shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15166-15179, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.329, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[2] Luma Labs AI. Luma dream machine: Ai-powered video content creation. https://dream-machine.lumalabs.ai/, 2025. Accessed: 2025-01-15. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.371, + 0.484, + 0.397 + ], + "angle": 0, + "content": "[3] Alyaa Qusay Aloraibi. Image morphing techniques: A review. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.399, + 0.484, + 0.453 + ], + "angle": 0, + "content": "[4] Hadar Averbuch-Elor, Daniel Cohen-Or, and Johannes Kopf. Smooth image sequences for data-driven morphing. In computer graphics forum, pages 203–213. Wiley Online Library, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[5] Melike Aydinlar and Yusuf Sahillioğlu. Part-based data-driven 3d shape interpolation. Computer-Aided Design, 136: 103027, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.484, + 0.538 + ], + "angle": 0, + "content": "[6] Thaddeus Beier and Shawn Neely. Feature-based image metamorphosis. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 529-536. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.539, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[7] Bhumika G Bhatt. Comparative study of triangulation based and feature based image morphing. Signal & Image Processing, 2(4):235, 2011. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.484, + 0.622 + ], + "angle": 0, + "content": "[8] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.484, + 0.69 + ], + "angle": 0, + "content": "[9] Federica Bogo, Javier Romero, Matthew Loper, and Michael J Black. Faust: Dataset and evaluation for 3d mesh registration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3794-3801, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[10] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[11] Soheil Darabi, Eli Shechtman, Connelly Barnes, Dan B Goldman, and Pradeep Sen. Image melding: Combining inconsistent images using patch-based synthesis. ACM Transactions on graphics (TOG), 31(4):1-10, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Jiacheng Deng, Chuxin Wang, Jiahao Lu, Jianfeng He, Tianzhu Zhang, Jiyang Yu, and Zhe Zhang. Se-ornet: Self-ensembling orientation-aware network for unsupervised point cloud shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364–5373, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.259, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[13] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4494-4504, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[14] Roberto M Dyke, Yu-Kun Lai, Paul L Rosin, Stefano Zappala, Seana Dykes, Daoliang Guo, Kun Li, Riccardo Marin, Simone Melzi, and Jingyu Yang. Shrec'20: Shape correspondence with non-isometric deformations. Computers & Graphics, 92:28-43, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[15] Michal Edelstein, Danielle Ezuz, and Mirela Ben-Chen. Enigma: Evolutionary non-isometric geometry matching. arXiv preprint arXiv:1905.10763, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.275, + 0.907, + 0.342 + ], + "angle": 0, + "content": "[16] Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi-scale shape registration with functional maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12265-12274, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.343, + 0.908, + 0.427 + ], + "angle": 0, + "content": "[17] Marvin Eisenberger, David Novotny, Gael Kerchenbaum, Patrick Labatut, Natalia Neverova, Daniel Cremers, and Andrea Vedaldi. Neuromorph: Unsupervised shape interpolation and correspondence in one go. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7473-7483, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.428, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[18] Noa Fish, Richard Zhang, Lilach Perry, Daniel Cohen-Or, Eli Shechtman, and Connelly Barnes. Image morphing with perceptual constraints and stn alignment. In Computer Graphics Forum, pages 303–313. Wiley Online Library, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[19] Lawson Fulton, Vismay Modi, David Duvenaud, David IW Levin, and Alec Jacobson. Latent-space dynamics for reduced deformable simulation. In Computer graphics forum, pages 379-391. Wiley Online Library, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[20] William Gao, Noam Aigerman, Thibault Groueix, Vova Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[21] Chen Geng, Yunzhi Zhang, Shangzhe Wu, and Jiajun Wu. Birth and death of a rose, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.622, + 0.906, + 0.692 + ], + "angle": 0, + "content": "[22] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19740–19750, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.906, + 0.733 + ], + "angle": 0, + "content": "[23] Qiyuan He, Jinghao Wang, Ziwei Liu, and Angela Yao. Aid: Attention interpolation of text-to-image diffusion. NeurIPS, 2024. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.734, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[24] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15167" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[27] Zehuan Huang, Yuanchen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[28] Alec Jacobson, Zhigang Deng, Ladislav Kavan, and John P Lewis. Skinning: Real-time shape deformation (full text not available). In ACM SIGGRAPH 2014 Courses, pages 1-1. 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[29] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G Kim, and Rana Hanocka. Meshup: Multi-target mesh deformation via blended score distillation. arXiv preprint arXiv:2408.14899, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[30] Vladimir G Kim, Yaron Lipman, and Thomas Funkhouser. Blended intrinsic maps. ACM transactions on graphics (TOG), 30(4):1-12, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[31] Yushi Lan, Chen Change Loy, and Bo Dai. DDF: Correspondence distillation from nerf-based gan. IJCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[32] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[33] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. In ICLR, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[34] Jing Liao, Rodolfo S Lima, Diego Nehab, Hugues Hoppe, Pedro V Sander, and Jinhui Yu. Automating image morphing using structural similarity on a halfway domain. ACM Transactions on Graphics (TOG), 33(5):1-12, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[35] Jianchu Lin, Yinxi Gu, Guangxiao Du, Guoqiang Qu, Xiaobing Chen, Yudong Zhang, Shangbing Gao, Zhen Liu, and Nallappan Gunasekaran. 2d/3d image morphing technology from traditional to modern: A survey. Information Fusion, page 102913, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[36] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[37] Robin Magnet and Maks Ovsjanikov. Scalable and efficient functional map computations on dense meshes. In Computer Graphics Forum, pages 89-101. Wiley Online Library, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[38] Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, and Benjamin Bossan. Peft: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[39] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.146 + ], + "angle": 0, + "content": "[41] Luca Morreale, Noam Aigerman, Vladimir G Kim, and Niloy J Mitra. Neural semantic surface maps. In Computer Graphics Forum, page e15005. Wiley Online Library, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[42] OpenAI. Gpt-4: Openai's fourth-generation language model, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[43] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[44] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. ACM Transactions on Graphics (ToG), 31(4):1-11, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[48] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. *ICLR*, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[50] Jing Ren, Simone Melzi, Maks Ovsjanikov, and Peter Wonka. Maptree: Recovering multiple solutions in the space of maps. ACM Transactions on Graphics, 39(6):1-17, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.906, + 0.759 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Dvir Samuel, Rami Ben-Ari, Nir Darshan, Haggai Maron, and Gal Chechik. Norm-guided latent space exploration for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15168" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[54] Eli Shechtman, Alex Rav-Acha, Michal Irani, and Steve Seitz. Regenerative morphing. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 615-622. IEEE, 2010. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[55] Liao Shen, Tianqi Liu, Huiqiang Sun, Xinyi Ye, Baopu Li, Jianming Zhang, and Zhiguo Cao. Dreammover: Leveraging the prior of diffusion models for image interpolation with large motion. arXiv preprint arXiv:2409.09605, 2, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.264 + ], + "angle": 0, + "content": "[56] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.267, + 0.483, + 0.322 + ], + "angle": 0, + "content": "[57] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.324, + 0.482, + 0.393 + ], + "angle": 0, + "content": "[58] Justin Solomon, Fernando De Goes, Gabriel Peyre, Marco Cuturi, Adrian Butscher, Andy Nguyen, Tao Du, and Leonidas Guibas. Convolutional wasserstein distances: Efficient optimal transportation on geometric domains. ACM Transactions on Graphics (ToG), 34(4):1-11, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.396, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[59] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, pages 109–116. CiteSeer, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.44, + 0.482, + 0.507 + ], + "angle": 0, + "content": "[60] Olga Sorkine, Daniel Cohen-Or, Yaron Lipman, Marc Alexa, Christian Rössl, and H-P Seidel. Laplacian surface editing. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH symposium on Geometry processing, pages 175–184, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.482, + 0.581 + ], + "angle": 0, + "content": "[61] Mingze Sun, Chen Guo, Puhua Jiang, Shiwei Mao, Yurun Chen, and Ruqi Huang. Srif: Semantic shape registration empowered by diffusion-based image morphing and flow estimation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.584, + 0.482, + 0.666 + ], + "angle": 0, + "content": "[62] Gary KL Tam, Zhi-Quan Cheng, Yu-Kun Lai, Frank C Langbein, Yonghuai Liu, David Marshall, Ralph R Martin, XianFang Sun, and Paul L Rosin. Registration of 3d point clouds and meshes: A survey from rigid to nonrigid. IEEE transactions on visualization and computer graphics, 19(7):1199-1217, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.669, + 0.482, + 0.724 + ], + "angle": 0, + "content": "[63] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.727, + 0.482, + 0.782 + ], + "angle": 0, + "content": "[64] Chih-Jung Tsai, Cheng Sun, and Hwann-Tzong Chen. Multiview regenerative morphing with dual flows. In European Conference on Computer Vision, pages 492-509. Springer, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.785, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[65] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[66] Shantanu Vyas, Ting-Ju Chen, Ronak R Mohanty, Peng Jiang, and Vinayak R Krishnamurthy. Latent embedded graphs for image and shape interpolation. Computer-Aided Design, 140:103091, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[67] George Wolberg. Image morphing: a survey. The visual computer, 14(8-9):360-372, 1998. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[68] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v(ision) is a human-aligned evaluator for text-to-3d generation. In CVPR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[69] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In European Conference on Computer Vision, pages 378-394. Springer, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[70] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[71] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.329, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[72] Songlin Yang, Wei Wang, Jun Ling, Bo Peng, Xu Tan, and Jing Dong. Context-aware talking-head video editing. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7718-7727, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[73] Songlin Yang, Wei Wang, Bo Peng, and Jing Dong. Designing a 3d-aware stylenerf encoder for face editing. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.907, + 0.523 + ], + "angle": 0, + "content": "[74] Songlin Yang, Wei Wang, Yushi Lan, Xiangyu Fan, Bo Peng, Lei Yang, and Jing Dong. Learning dense correspondence for nef-based face reenactment. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6522-6530, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[75] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[76] Mehmet Ersin Yumer, Siddhartha Chaudhuri, Jessica K Hodgins, and Levent Burak Kara. Semantic shape editing using deformation handles. ACM Transactions on Graphics (TOG), 34(4):1-12, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[77] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[78] Xiao Zhan, Rao Fu, and Daniel Ritchie. Charactermixer: Rig-aware interpolation of 3d characters. In Computer Graphics Forum, page e15047. Wiley Online Library, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[79] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[80] Kaiwen Zhang, Yifan Zhou, Xudong Xu, Bo Dai, and Xinggang Pan. Diffmorpher: Unleashing the capability of diffusion models for image morphing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7912-7921, 2024. 2, 3, 4, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15169" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[81] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[82] Zhengbo Zhang, Li Xu, Duo Peng, Hossein Rahmani, and Jun Liu. Diff-tracker: text-to-image diffusion models are unsupervised trackers. In European Conference on Computer Vision, pages 319-337. Springer, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[83] Junzhe Zhu, Yuanchen Ju, Junyi Zhang, Muhan Wang, Zhecheng Yuan, Kaizhe Hu, and Huazhe Xu. Dense-matcher: Learning 3d semantic correspondence for category-level manipulation from a single demo. arXiv preprint arXiv:2412.05268, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[84] Bhushan Zope and Soniya B Zope. A survey of morphing techniques. International Journal of Advanced Engineering, Management and Science, 3(2):239773, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15170" + } + ] +] \ No newline at end of file diff --git a/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/0ee49b9a-7099-4408-8e6f-37bd5bc28baf_origin.pdf b/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/0ee49b9a-7099-4408-8e6f-37bd5bc28baf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..38b9cadb9af4121fd4c6c9bb1de94e9dc2cf09ac --- /dev/null +++ b/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/0ee49b9a-7099-4408-8e6f-37bd5bc28baf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:85d313318afdac9e0271ce244c321ff703c4615da78deb7e0d16545e18dcdcee +size 2560394 diff --git a/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/full.md b/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..87128e2308dc49f0332c6a92d99ab3328610f7ec --- /dev/null +++ b/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/full.md @@ -0,0 +1,373 @@ +# Textured 3D Regenerative Morphing with 3D Diffusion Prior + +Songlin Yang, Yushi Lan, Honghua Chen, Xingang Pan +S-Lab, Nanyang Technological University + +sl.yang888@outlook.com, {yushi001, honghua.chen, xingang.pan}@ntu.edu.sg + +![](images/ae9c7ebaef496fc140b4e44516f54aaefcc49c649abbb99a12be1a5faed571bd.jpg) +Figure 1. The textured 3D morphing sequences regenerated using our method with 3D diffusion prior. Our method requires no category-specific alignment data for model training and no labor-intensive preprocessing for explicit correspondence, enabling smooth and plausible morphing across diverse cross-category 3D object pairs. More video results can be found here. + +# Abstract + +Textured 3D morphing creates smooth and plausible interpolation sequences between two 3D objects, focusing on transitions in both shape and texture. This is important for creative applications like visual effects in filmmaking. Previous methods rely on establishing point-to-point correspondences and determining smooth deformation trajectories, which inherently restrict them to shape-only morphing on untextured, topologically aligned datasets. This restriction leads to labor-intensive preprocessing and poor generalization. To overcome these challenges, we propose a method for 3D regenerative morphing using a 3D diffusion prior. Unlike previous methods that depend on explicit correspondences and deformations, our method eliminates + +the additional need for obtaining correspondence and uses the 3D diffusion prior to generate morphing. Specifically, we first introduce a 3D diffusion model and interpolate the source and target information at three levels: initial noise, model parameters, and condition features. We then explore an Attention Fusion strategy to generate smoother morphing sequences. To further improve the plausibility of semantic interpolation and the generated 3D surfaces, we propose two strategies: (a) Token Reordering, where we match approximate tokens based on semantic analysis to guide implicit correspondences in the denoising process of the diffusion model, and (b) Low-Frequency Enhancement, where we enhance low-frequency signals in the tokens to improve the quality of generated surfaces. Experimental results show that our method achieves superior smoothness and + +plausibility in 3D morphing across diverse cross-category object pairs, offering a novel regenerative method for 3D morphing with textured representations. + +# 1. Introduction + +Morphing [35] generates an interpolation sequence between a source and a target, requiring smooth and plausible transitions. This fundamental technique is crucial for creative applications like visual effects in film and media. Depending on the data type, morphing is categorized into image morphing [4, 7, 18, 23, 34, 54, 80] and 3D morphing [21, 29, 64]. Compared to image morphing, 3D morphing aligns more naturally with visual effects production, where objects and transformations inherently operate in 3D space. However, 3D morphing is more challenging, requiring the interpolation of 3D objects holistically (i.e., image morphing can be viewed as a special case of 3D morphing from a specific viewpoint). Our work addresses the task of textured 3D morphing [64], which generates a sequence for two textured 3D representations, aiming for smooth and plausible transitions in shape and texture, as shown in Fig. 1. + +Previous 3D morphing methods mainly focused on morphing shapes, which can be summarized in two steps: first, establishing correspondence [12, 13, 31, 37] between the 3D representations of the source and target objects, and second, determining smooth and plausible deformation trajectories between corresponding 3D points [17, 61, 78]. Following these steps, previous methods blend the two 3D representations with respective weights to obtain a sequence of interpolated 3D representations. + +Due to the scarcity of topologically aligned and textured 3D datasets, previous 3D morphing methods [17, 61, 78] have mainly focused on in-domain untextured datasets, such as FAUST [9] (human shapes) and Shrec'20 [14] (quadruped animals). As a result, these methods are limited to shape-only morphing with two key generalization challenges: (a) Labor-Intensive Preprocessing: Morphing new in-domain 3D data with these methods [17, 78] requires domain-specific alignment with the training datasets through tedious registration [61] and matching [83] steps. (b) Limited Morphing Capacity: The methods [5, 17, 66, 78] suffer from limited object diversity and small datasets, leading to ambiguous and implausible interpolations. + +The limitations of previous methods inspire us to consider two critical questions: (a) Is explicit point-to-point correspondence truly necessary? (b) Can we enhance the generalization capability of textured 3D morphing via a generic generative prior? + +For correspondence, obtaining dense correspondences between textured 3D representations across categories remains underexplored [83] despite being foundational to previous methods, while explicit constraints can instead limit + +the creativity of morphing. Therefore, a promising method is using labor-saving implicit correspondences [31, 74] to guide the morphing. For instance, morphing could be formulated as an optimization problem, allowing correspondences to emerge automatically [64]. Alternatively, attention mechanisms in generative models could enable automatic alignment during object blending [23, 55, 80]. + +For generative priors, recent advancements in diffusion-based generation models offer two strategies for 3D morphing: using 3D diffusion models directly [10, 33, 70], or enhancing 2D models with 3D priors. However, those hybrid 2D-3D methods [22, 29, 47, 51] face significant challenges: 2D models lack a holistic 3D knowledge, and optimizing the 2D-3D mapping cannot guarantee that the morphing process will maintain 3D consistency across viewpoints. Therefore, directly using a 3D diffusion model to regenerate the interpolated 3D representations (i.e., regenerative morphing [54, 64]) enables authentic 3D morphing and has the potential to be scaled up by using a more state-of-the-art 3D generation model. + +Building on the above analysis, we adopt a generic 3D diffusion prior that leverages its implicit correspondence and 3D generation capabilities to blend source and target information, enabling the regeneration of interpolated textured 3D representations. + +Specifically, we first introduce a 3D diffusion model [33] and interpolate the information from the source and target at three levels: initial noises, model parameters, and condition features. The 3D representation for each interpolation is then regenerated using the 3D generation model. To improve the 3D diffusion model's ability to generate smooth morphing sequences, we explore an extended Attention Fusion strategy [23, 80]. However, fusing different information weakened the model's denoising capability, and aggressively applying Attention Fusion to all denoising steps for smoother morphing resulted in implausible outcomes. Therefore, to improve the plausibility of semantic interpolation and the generated 3D surfaces, we propose two strategies: (a) Token Reordering: After semantic analysis, we identify semantic correspondences in diffusion tokens and propose matching approximate tokens before attention computation to better guide implicit correspondences in the diffusion space. (b) Low-Frequency Enhancement: Frequency-domain analysis [57] reveals that boosting low-frequency signals improves the surface quality of the 3D diffusion model. Thus, we enhance low-frequency signals at key time steps to preserve the model's ability to generate 3D surfaces. + +Our contributions can be summarized as follows: (a) We are the first to use a generic 3D diffusion prior for morphing textured 3D representations, enabling 3D morphing without explicit correspondences. (b) We analyze the merging of source and target information during morphing with + +a 3D diffusion prior from semantic and frequency perspectives, proposing Token Reordering and Low-Frequency Enhancement to improve plausibility. (c) Extensive experimental results demonstrate that our method achieves superior smoothness and plausibility in performing 3D morphing across diverse cross-category object pairs. + +# 2. Related Work + +# 2.1. Image Morphing + +Image morphing is a long-standing challenge in computer vision and graphics [3, 67, 84]. Traditional methods [6, 7, 11, 34, 54] use feature-based warping and blending to create smooth transitions but struggle to generate new content, often leading to artifacts. Data-driven methods [4, 18] leverage large single-class datasets to achieve smoother results but are limited in cross-domain or personalized applications due to their reliance on specific data. The methods like DiffMorpher [80] and AID [23] address this by utilizing pre-trained diffusion models on diverse datasets, enabling flexible morphing across a wide range of object categories. Inputting multi-view images to image morphing methods enables 3D-aware morphing. + +# 2.2.3D Morphing + +Previous 3D morphing methods focus on shape-only correspondence [62], and realize 3D morphing through interpolation or deformation between corresponding 3D primitives (e.g., points, vertices, and faces). They can be divided into (a) Axiomatic Methods: They tend to rely on sparse landmarks [15, 30] or use functional maps [44] to address under-constrained mapping spaces, such as MapTree [50] and SmoothShells [16]. Energy-minimizing functions [59, 60] and skinning methods [19, 28] enable deformation control, while optimal transportation [58, 64] approximates shape correspondence. (b) Learning-Based Methods: This category of methods can be categorized into introducing other generative priors and using aligned data. SATR [1] introduces the semantic labels from multi-view images. NSSM [41] uses DINO [43] for sparse landmarks and SRIF [61] introduces large vision models [80] for semantic shape registration and morphing. A recent class of works leverages text prompts as user inputs for driving a deformation towards an arbitrary textual prompt [20, 39, 40], but CLIP [49] objective lacks a full understanding of object details. For data prior, category-specific training [76] on a topology-aligned dataset such as NeuroMorph [17] is also prevailing in learning-based 3D shape analysis. + +These methods focus on shape-only morphing or rely on rigging annotations, requiring extra effort due to 2D-3D mismatches. In contrast, our work enables textured 3D morphing without learning correspondence or deformation, using a 3D diffusion model to regenerate interpolated rep + +resentations, addressing previous challenges and offering a new direction for 3D morphing research. + +# 2.3. 3D Diffusion Model + +Generative 3D priors fall into two types: native 3D diffusion models and 3D-aware diffusion models. Due to the scarcity of 3D data, methods leveraging 2D priors to generate 3D or multi-view content have been proposed. For instance, Score Distillation Sampling [47] distills 3D information from a 2D diffusion model. 3D-aware generation can be divided into two steps: multi-view image generation [56] followed by feed-forward 3D reconstruction [71]. However, these methods inherently lack a 3D latent space. To address this, native 3D diffusion models [32, 77, 79, 81] that encode and learn 3D representations have been introduced. These models typically consist of two steps: training a VAE to encode 3D data and training a diffusion model based on corresponding latent codes. We use Gaussian Anything [33] as the 3D diffusion prior, encoding 3D information in a point cloud latent space. + +# 3. Method + +The framework for 3D regenerative morphing based on the 3D diffusion prior, as shown in Fig. 2, consists of three main steps: Basic Interpolation, where essential information is interpolated (Sec. 3.2); Smoothness Improvement, achieved through an Attention Fusion mechanism (Sec. 3.3); and Plausibility Improvement, which involves two strategies, Token Reordering (Sec. 3.4) and Low-Frequency Enhancement (Sec. 3.5). + +# 3.1. Preliminary + +# 3.1.1. 3D Diffusion Model + +We select Gaussian Anything [33] as our 3D diffusion prior, a two-stage native 3D diffusion model with a structured latent representation (i.e., point cloud) and the DiT [46] structure. It consists of a geometry generation model $\epsilon_{G}$ and a texture generation model $\epsilon_{T}$ . In the first stage, the model $\epsilon_{G}$ takes a Gaussian initial noise $\mathbf{z}_{G}$ and textual conditioning information $\mathbf{c}$ as inputs, to generate a structured point cloud representation $\mathbf{x}_{\text {point-cloud }} = \epsilon_{G}(\mathbf{z}_{G}, \mathbf{c})$ . In the second stage, $\mathbf{x}_{\text {point-cloud }}$ is added to a Gaussian initial noise $\mathbf{z}_{T}$ , and the resulting variable is denoised by $\epsilon_{T}$ with condition $\mathbf{c}$ to obtain the final texture feature $\mathbf{x}_{\text {feature }} = \epsilon_{T}(\mathbf{x}_{\text {point-cloud }} + \mathbf{z}_{T}, \mathbf{c})$ . Finally, $\mathbf{x}_{\text {point-cloud }}$ and $\mathbf{x}_{\text {feature }}$ are fed into a pre-trained decoder $\mathcal{D}$ to produce the final 3D Gaussian representation $\mathbf{x}_{3D} = \mathcal{D}(\mathbf{x}_{\text {point-cloud}}, \mathbf{x}_{\text {feature}})$ , which can be rendered as multi-view images. Notably, we denote the inherently unordered point-cloud-based token sequences between DiT blocks as $\{h_{j}\}_{j=1}^{M}$ , where $M$ is the length of token sequence. + +![](images/cf650e840532d63ca4a1a001466ed896c27bf33835f6e0f00260cf20ea6343b8.jpg) + +![](images/58e45fb395ccbe570fb1b0057a38d046a9cc323574d17958fcf3ac9b85d00759.jpg) +Figure 2. The framework of our method. Beyond basic interpolation, Attention Fusion is explored to improve smoothness, while Token Reordering and Low-Frequency Enhancement are proposed to improve plausibility. + +![](images/ad4b57356c64e20c0552fe9d21127835754d300b14fee184931e2e5150eb4cf9.jpg) + +![](images/4c801e05a74c58d48307c7ebab74d66fdcce3bd3b3d2d8208717aabdd1a1e80b.jpg) + +# 3.1.2. Attention + +The attention [65] mechanism is an important component for the current text-driven diffusion models [10, 46, 51, 52, 70], especially cross-attention and self-attention. Given a latent variable $z \in \mathbb{R}^{d_z}$ , a text condition $c \in \mathbb{R}^{d_c}$ , and the attention layer with matrices $W_{Q} \in \mathbb{R}^{d_{z} \times d_{q}}$ , $W_{K} \in \mathbb{R}^{d_{c} \times d_{k}}$ , and $W_{V} \in \mathbb{R}^{d_{c} \times d_{v}}$ , the cross-attention is computed as: + +$$ +A (z, c) = A t t n (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {\top}}{\sqrt {d _ {k}}}\right) V, \tag {1} +$$ + +where $Q = W_{Q}^{\top}z, K = W_{K}^{\top}c, V = W_{V}^{\top}c$ . Self-attention is a specific form of cross-attention, computed as $A(z,z)$ . + +# 3.2. Basic Interpolation + +# 3.2.1. Implementation + +The basic interpolation has three levels, with source and target weights given by $(1 - \alpha)$ and $\alpha$ , respectively, where $\alpha = 0$ generates the source $\mathbf{x}_{3D}^{src}$ , and $\alpha = 1$ generates the target $\mathbf{x}_{3D}^{tgt}$ . + +(a) Initial Noises: The textured 3D representations $\mathbf{x}_{3D}^{src}$ and $\mathbf{x}_{3D}^{tgt}$ , are inverted via diffusion inversion [36] to obtain their respective input noises, $[\mathbf{z}_T^{src},\mathbf{z}_G^{src}]$ and $[\mathbf{z}_T^{tgt},\mathbf{z}_G^{tgt}]$ . To ensure that Gaussian noise properties are preserved, spherical linear interpolation [53] is applied to these noises to generate the interpolated noises, $[\mathbf{z}_T^\alpha ,\mathbf{z}_G^\alpha ]$ +(b) Model Parameters: Given $\mathbf{x}_{3D}^{src}$ and $\mathbf{x}_{3D}^{tgt}$ , we fine-tune the model using LoRA (Low-Rank Adaptation) [26] to obtain two sets of LoRA parameters. These parameters are then linearly interpolated and fused to obtain the morphing models $\epsilon_G^\alpha$ and $\epsilon_T^\alpha$ . +(c) Condition Features: We adopt text conditions as they provide a more consistent semantic space and are layout-independent compared to DINO-based [43] image condi + +tions. The text prompts of the source and target 3D representations are encoded via a CLIP [49] encoder into $\mathbf{c}^{src}$ and $\mathbf{c}^{tgt}$ , which are linearly interpolated to produce $\mathbf{c}^{\alpha}$ . + +# 3.2.2. Problems + +Basic interpolation realizes the blending of basic information, but there are still the following problems: + +(a) Abrupt Changes (Smoothness): Nonlinear multi-step denoising in diffusion models introduces variability in noise-to-data mapping, while the CLIP encoder's space lacks guaranteed semantic smoothness, leading to abrupt changes (yellow-highlighted part in Fig. 8). +(b) Artifacts (Plausibility): Misalignment between conditioning and diffusion spaces disrupts learned mappings, causing artifacts like structure collapse or degraded surface quality (red and yellowhighlighted parts in Fig. 8). + +Therefore, we investigate Attention Fusion (Sec. 3.3) for smoothness and Token Reordering (Sec. 3.4) with Low-Frequency Enhancement (Sec. 3.5) for plausibility. + +# 3.3. Attention Fusion + +The fusion of attention has been proven effective in improving morphing smoothness in 2D diffusion-based image morphing. However, DiffMorpher [80] only explored self-attention interpolation and did not consider the smoothness of semantic features in conditioning. AID [23] did not address the differences between the original and LoRA models, neglecting conditioning feature interpolation with accurate inversion. Our method combines self-attention and cross-attention fusion, using unified attention features from fine-tuned models to enhance smoothness while encouraging plausible generation. + +Specifically, as shown in Fig. 2, we first feed $z^{src}$ , $z^{tgt}$ and $z^{\alpha}$ into blocks of the morphing model $\epsilon^{\alpha}$ to obtain three sets of $(Q^{src}, K^{src}, V^{src})$ , $(Q^{tgt}, K^{tgt}, V^{tgt})$ , and $(Q^{\alpha}, K^{\alpha}, V^{\alpha})$ . Then, based on Eq. (1), denoting concatenation as $[\cdot, \cdot]$ , we obtain the fused attention by: + +$$ +F u s e d - A t t n (Q ^ {\alpha}, K ^ {\alpha}, V ^ {\alpha}) = A t t n (Q ^ {\alpha}, +$$ + +$$ +\left[ (1 - \alpha) K ^ {\mathrm {s r c}} + \alpha K ^ {\mathrm {t g t}}, K ^ {\alpha} \right], \left[ (1 - \alpha) V ^ {\mathrm {s r c}} + \alpha V ^ {\mathrm {t g t}}, V ^ {\alpha} \right]). \tag {2} +$$ + +Using the interpolated model for all attention calculations reduces quality deterioration caused by Attention Fusion. However, excessive fusion across time steps harms plausibility, causing structural collapse and surface issues (see Fig. 8). + +# 3.4. Token Reordering + +# 3.4.1. Motivation + +The point clouds of 3D objects are tokenized into sequences $\{h_j\}_{j=1}^M$ , with each token $h_j$ representing a 3D point. However, point clouds are inherently unordered representations [77], and relying solely on the attention mechanism to match the points in different 3D objects may lead the model to make semantically implausible connections (e.g., the connected tokens may represent "chair legs" and "donut frosting", respectively). This vanilla use of the diffusion prior does not fully exploit its potential, as works like DIFT [63] show that 2D diffusion features or tokens represent semantics [24, 75, 82]. We infer that 3D diffusion features also capture semantic correspondences. Therefore, why not first pair points with similar semantics and then interpolate within this semantically plausible space? + +# 3.4.2. Experimental Analysis + +To test our hypothesis, we first present the problem of position-identical tokens in capturing correspondence (i.e., a problem of vanilla Attention Fusion). We then reveal the emergent correspondence between position-different tokens, supporting the validity of Token Reordering. + +(a) Problem of Vanilla Attention Fusion. The DiT models tend to have stronger semantic correspondence in the intermediate time steps [75], and shorter distances between two tokens indicate their higher semantic similarity. To test this in 3D diffusion, we first obtain fully aligned 3D objects by scaling to test our model's correspondence. For instance, as shown in Fig. 3, the right chair results from scaling the left chair's x-axis by 0.6 and z-axis by 1.2. This ensures that tokens at the identical sequence positions represent the originally same point-cloud points (i.e., the points shift positions after scaling while retaining the same color and semantics). Ideally, in geometry diffusion, token distance at the same position should first decrease (semantic learning) until step 200, then increase (point location learning). In texture diffusion, it should gradually decrease, ensuring the color information at the identical position consistent. But + +![](images/9c96dc898f1a2d206a32ebb1cbb9811a84095920a79f7f6600360799df986be4.jpg) +Figure 3. Mean squared error of token distances at identical positions. The solid line represents real data, while the red dashed line shows the ideal trend (i.e., if position-identical tokens can capture the correspondence). Light blue blocks highlight their conflicts. + +![](images/072599d6a2516dedcb71ec73e934ac11bb0b31909dd8b1447ab9a82777a7a7d0.jpg) + +![](images/986c9ee0e976479b4eb1c2c85dd4c169515a3a38aabd73e763b897003bf71453.jpg) +Figure 4. Mean squared error of point distances between token-distance-closest points. The red dashed line indicates the time step range where semantic correspondence emerges. + +![](images/7281a51f332585ab661e4035e74d0fe13156ce8b9f9498a2cdcdf3587e1176aa.jpg) + +the Fig. 3 indicates that position-identical semantic alignment is decreasing during denoising, especially the texture diffusion. The vanilla Attention Fusion links these positional tokens, which causes artifacts (see Fig. 8, 3rd row). + +(b) Emergence of Semantic Correspondence. By extracting the corresponding points in the final point cloud for two tokens and computing their distance, we can assess the semantic matching of the tokens. We obtain token sequences for the same 3D point cloud through different random seeds, meaning tokens at the identical sequence positions do not correspond to the same point-cloud points. For example, as shown in Fig. 4, the 3D point at the same position may correspond to the $1st$ token (the left chair) and the $M-th$ token (the right chair) respectively. As shown in Fig. 4, the distance between points based on geometry token distance decreases with increasing time steps, indicating emergent semantic correspondence. Conversely, the distance based on texture token distance first decreases and then increases, suggesting that the denoising process initially promotes semantic alignment, but later shifts towards learning texture details. This aligns with findings [75] in 2D diffusion. + +Therefore, by considering both geometry and texture denoising characteristics, we reorder tokens in the intermediate stages to better guide plausible morphing. + +# 3.4.3. Implementation + +Based on these observations, we reorder the token sequences before passing the output of the $i$ -th block to the $(i + 1)$ -th block, where $i \in \{1,2,\dots,N - 1\}$ and $N$ is the total number of blocks in the diffusion model. As shown in Fig. 2 and Alg. 1, we reorder the source and target token sequences $\{h_j^{\mathrm{src}}\}_{j = 1}^M$ , $\{h_j^{\mathrm{tgt}}\}_{j = 1}^M$ such that tokens with similar distances align at the same index, as follows: + +$$ +\underset {j = 1} {\text {m i n i m i z e}} \sum_ {j = 1} ^ {M} \left\| h _ {j} ^ {\mathrm {s r c}} - h _ {\sigma (j)} ^ {\mathrm {t g t}} \right\|, \tag {3} +$$ + +# Algorithm 1 Token Reordering + +1: Input: Source token sequence $\{h_j^{\mathrm{src}}\}_{j = 1}^M$ target token sequence $\{h_j^{\mathrm{tgt}}\}_{j = 1}^M$ +2: Output: Reordered sequences $\{\tilde{h}_j^{\mathrm{src}}\}_{j = 1}^M$ $\{\tilde{h}_j^{\mathrm{tgt}}\}_{j = 1}^M$ +3: for $i = 1$ to $N - 1$ do $\triangleright$ Efficient computation by PyTorch [45] operations +4: Compute $D_{j,k} = \| h_j^{\mathrm{src}} - h_k^{\mathrm{tgt}}\|$ and $\sigma (j) = \arg \min_{k}D_{j,k},\forall j$ +5: if $\alpha < 0.5$ then +6: $\{\tilde{h}_j^{\mathrm{tgt}}\}_{j = 1}^M\gets \{h_{\sigma (j)}^{\mathrm{tgt}}\} ,\{\tilde{h}_j^{\mathrm{src}}\}_{j = 1}^M\gets \{h_j^{\mathrm{src}}\}_{j = 1}^M$ +7:els +8: $\{\tilde{h}_j^{\mathrm{src}}\}_{j = 1}^M\leftarrow \{h_{\sigma -1(j)}^{\mathrm{src}}\} ,\{\tilde{h}_j^{\mathrm{tgt}}\}_{j = 1}^M\leftarrow \{h_j^{\mathrm{tgt}}\}_{j = 1}^M$ +9: end if +10: end for +11: Return $\{\tilde{h}_j^{\mathrm{src}}\}_{j = 1}^M$ $\{\tilde{h}_j^{\mathrm{tgt}}\}_{j = 1}^M$ + +where $M$ is the number of tokens, and $\sigma(j)$ denotes the index of the element in the target sequence that best corresponds to the $j$ -th element in the source sequence. To further refine the generation, we set different reordering strategies depending on $\alpha$ : For $\alpha \in [0, 0.5)$ , the target token sequence is reordered based on the source token sequence. For $\alpha \in [0.5, 1]$ , the source token sequence is reordered based on the target token sequence. + +# 3.5. Low-Frequency Enhancement + +# 3.5.1. Motivation + +Due to the additional Attention Fusion operations, the misalignment between the condition space and the diffusion space is exacerbated. Excessive Attention Fusion at later time steps significantly degrades the model's denoising capability, with the deterioration becoming more pronounced as the fusion time step increases (see Fig. 8, from the $4th$ to the $5th$ row). This raises the question: what part of the $3D$ diffusion model is influenced by the morphing operations, leading to this decline in performance? + +# 3.5.2. Experimental Analysis + +The deterioration problem is analyzed in the frequency domain to address the significant visual differences between 3D objects, as shown in Fig. 5. In 3D generation, low-frequency noise controls the overall layout, while high-frequency noise governs surface details. Excessive amplification of high-frequency components during denoising can interfere with the low-frequency components, degrading overall quality. Therefore, enhancing low-frequency signals during denoising is crucial to prevent the model from overemphasizing high-frequency noise, thus improving the quality of 3D surface generation. Similar patterns appear in 2D diffusion studies [48, 57, 69], with low frequencies tied to image layout and high frequencies to details. + +# 3.5.3. Implementation + +To address this, as shown in Fig. 2, we propose to enhance the low-frequency signal when generating the 3D interpolations. Specifically, the process is defined as: + +$$ +F (h) = \operatorname {F F T} (h), \tag {4} +$$ + +$$ +F _ {\omega < \omega_ {0}} ^ {\prime} (h) = F _ {\omega < \omega_ {0}} (h) \odot \text {s c a l e}, \tag {5} +$$ + +$$ +h ^ {\prime} = \operatorname {I F F T} \left(\left[ F _ {\omega < \omega_ {0}} ^ {\prime} (h), F _ {\omega \geq \omega_ {0}} (h) \right]\right), \tag {6} +$$ + +![](images/53ffca7b0d30eaded2579e00f4f29007da582c821af681bca5592e63b2dcc9a8.jpg) +Figure 5. The changes of high and low frequency signals in the diffusion model during the denoising process. Visualizing signal amplitudes at different time steps reveals that low-frequency noise varies less than high-frequency noise, with smaller gaps across denoising time steps. + +![](images/4b4f7cf3696832a2b5bf1f270d9a9f4014f3a46650eebb584120a1fc69cda84b.jpg) + +![](images/b952db4e63fc36a45322068f82a9c2568c1fd11e30fc130e4a5df44347460e2a.jpg) +Figure 6. The effects of scale for improving surface quality. + +where $h$ represents the tokens, $F(h)$ their Fourier features, and $h'$ the enhanced tokens, $\omega$ and $\omega_0$ denote Fourier frequencies and the threshold, with $\omega < \omega_0$ indicating low-frequency components. The scale is the enhancement coefficient, and $\mathrm{FFT}(\cdot)$ and $\mathrm{IFFT}(\cdot)$ are the Fourier transform and inverse Fourier transform. As shown in Fig. 6, increasing the scale continuously enhances the surface generation capability. However, beyond a certain value, the surface does not increase further with larger scale values. + +# 4. Experiments + +# 4.1. Implementation Details + +# 4.1.1. Morphing Configuration + +All experiments use 250 denoising time steps. Finetuning is performed using LoRA models implemented with PEFT [38] using 500 training steps (rank = 16, alpha = 20, and list = ['to_k', 'to_q', 'to_v', 'qkv']) The initial value of $\alpha$ is sampled from a Beta distribution, $Beta(5,5)$ , over the $[0,1]$ interval, with 20 points selected as interpolation weights. If multiple morphing is allowed, $\alpha$ can be concentrated in ranges with more variation by adjusting the Beta distribution or using the reschedule strategy from DiffMorpher [80]. For Attention Fusion, we begin at step 1, with geometry diffusion ending at step 160 and texture diffusion at step 5. Token Reordering starts at step 80 and ends at step 200. Low-Frequency Enhancement occurs between steps 200 and 230, with scale = 5 and $\omega_0 = 0.1\pi$ . Note that all our results share the same hyperparameters. + +# 4.1.2. Baselines + +(a) Task (Textured 3D Morphing): MorphFlow [64] directly implements morphing on textured 3D volumetric representations, which has the closest setting to ours; (b) Morphing Tricks: DiffMorpher [80] and AID [23] propose fusion + +![](images/b2415ebf24f9023a93bceb563a6c4a28292f4446788cfd42507882e57a9ca4d8.jpg) +Figure 7. Qualitative comparisons of methods in tasks, morphing tricks, and generative priors. More video results can be found here. + +Table 1. Quantitative comparisons of different methods. + +
Quantitative MetricsUser Study
FID↓KID↓STP-G↑SEP-G↑PPL↓PDV↓STP-U↑SEP-U↑
DiffMorpher218.070.3670.230.135.230.05350.4350.300
AID115.720.1240.670.704.680.01180.3800.505
MV-Adapter120.930.1290.630.577.290.01520.2250.350
Luma95.490.0880.830.777.370.00070.4150.330
MorphFlow147.700.1500.870.903.100.00010.5550.505
Ours6.360.0011.001.003.020.00010.9150.950
+ +strategies for attention to enhance smoothness; (c) Generative Priors: We compare the performance of using 2D diffusion (DiffMorpher, AID), multi-view diffusion (MV-Adapter [27]), video generation (Luma [2]), and 3D diffusion as prior in 3D (multi-view image) morphing task. + +# 4.1.3. Metrics + +(a) FID [25] and KID [8]: They are calculated by comparing 3,000 images rendered from original and interpolated 3D representations. (b) Structural Plausibility-GPT (STP-G) and Semantic Plausibility-GPT (SEP-G): GPT-4o [42, 68] evaluates structural and semantic plausibility based on testing results, providing scores and explanations. (c) Perceptual Path Length (PPL) and Perceptual Distance Variance (PDV) [80]: PPL sums perceptual losses over 20-frame sequences, reflecting smoothness and consistency, while PDV measures the variance, indicating transition ho + +mogeneity. (d) Structural Plausibility-User (STP-U) and Semantic Plausibility-User (SEP-U): Volunteers score morphing results based on structural and semantic plausibility. + +# 4.2. Evaluation + +# 4.2.1. Textured 3D Morphing + +Our method adopts a 3D generation prior for 3D morphing, offering two main benefits: direct morphing of textured 3D representations, and maintaining structural and semantic consistency in interpolated representations, as shown in Fig. 1. The closest baseline to our method is MorphFlow [64], which morphs between two 3D volumetric representations using optimal transport optimization. However, as shown in Fig. 7, MorphFlow has two main drawbacks. First, the generated quality is low (see Tab. 1). The advantage of volumetric representations for photorealism is diminished due to the dense interpolation process, which requires morphing even points with no color. Second, it lacks an effective generative prior to the intermediate process, limiting its ability to understand the semantic meaning of intermediate stages. As a result, it often introduces artifacts in morphing scenarios, such as generating six legs instead of four when morphing a bear's and a stool's legs (see the "Bear-Stool" case in Fig. 7). + +Based on the quantitative metrics and user study in the + +Tab. 1, ours not only significantly outperforms MorphFlow in quality but also demonstrates an impressive "understanding" capabilities: (a) the effective mapping of semantically similar parts, as shown when morphing a boot into a red teddy bear (see the "Boot-Teddy Bear" case in Fig. 1), where our model smoothly splits the boot into legs and transforms them into the bear's facial features, and (b) minimal disconnected artifacts. The regenerative nature of our method ensures the fusion of source and target while considering the distribution of the entire latent space, minimizing issues like 3D structure collapse or disconnections. + +# 4.2.2. 3D-Aware (Multi-View Image) Morphing + +We evaluate alternatives to our method for textured 3D morphing from two perspectives. First, 3D morphing can be viewed as multi-view image morphing, where multiview source and target images are fed into image morphing methods for pseudo-3D morphing. Second, many image and video generation models, beyond 3D priors, can perform interpolation tasks. Notably, 2D generation models [27, 56, 72-74], trained on larger datasets often produce 3D-consistent images with superior semantic, structural, and textural understanding compared to 3D models. Therefore, we explore three types of generative priors: 2D diffusion, multi-view diffusion, and video generation. + +As shown in Fig. 7 and Tab. 1, compared to state-of-the-art 2D image morphing methods, such as DiffMorpher [80] and AID [23], 2D diffusion models often suffer from mode collapse when influenced by non-object image regions (e.g., DiffMorpher is sensitive to white backgrounds). Their lack of 3D consistency leads to inconsistent morphing results for the same $\alpha$ across different viewpoints. When comparing with multi-view diffusion models, we observe that their image-based multi-view generation is limited by pixel-aligned morphing. This limitation becomes apparent when matching pixels across large spatial distances—such as aligning a point on the lower edge of a pumpkin's surface with a point on the stem of a mushroom in the image's center, resulting in interpolation errors (see the "Pumpkpin-Mushroom" case in Fig. 7). Lastly, video generation models, while demonstrating strong spatial understanding and achieving morphing by specifying the source and target images as the first and last frames, suffer from limited controllability during generation, often producing incomplete or out-of-frame content. Additionally, the structural consistency of intermediate frames remains suboptimal. + +# 4.3. Ablation Study + +We observe that texture diffusion requires minimal Attention Fusion (AF), while geometry diffusion offers a larger range, emphasizing shape understanding in 3D morphing. Thus, we incorporate Attention Fusion in texture diffusion from the first to the fifth time step, while progressively extending the final time step for geometry diffusion. As shown + +![](images/8eb74bb7f786c6684884505d273177cfbb261ea66435feb73bacba3fc182c7fa.jpg) +Figure 8. Ablation experiments of our proposed strategies. + +Table 2. Quantitative evaluation of ablation study. + +
Geometry Diffusion Time Step80 Steps120 Steps160 Steps
FID↓PPL↓PDV↓FID↓PPL↓PDV↓FID↓PPL↓PDV↓
Basic6.386.010.05726.386.010.05726.386.010.0572
Basic+AF9.975.650.054496.324.360.0121105.624.440.0110
Basic+AF+TR8.284.060.00709.703.300.00389.413.370.0024
Basic+AF+LE8.093.760.00497.293.240.00227.083.150.0009
Basic+AF+TR+LE6.403.540.00356.373.210.00196.363.020.0001
+ +in Fig. 8, adding Attention Fusion on top of basic interpolation improves smoothness, but extending the final time step too far causes 3D structural collapse. Token Reordering (TR) helps mitigate this issue, though pushing the time step further reduces generative quality. Ultimately, by applying the Low-Frequency Enhancement (LE) strategy, we balance the smoothness and structural plausibility and ensure all frequencies are maintained effectively. We quantitatively evaluate five different pairs under various settings, as shown in Tab. 2. The ablation study demonstrates that achieving a smooth and plausible morphing sequence requires increasing Attention Fusion steps, as well as applying Token Reordering and Low-Frequency Enhancement. + +# 5. Conclusions + +We propose a method to achieve smooth and plausible morphing sequences across diverse cross-category 3D object pairs, incorporating Attention Fusion, Token Reordering, and Low-Frequency Enhancement. This introduces a new paradigm for textured 3D morphing, extending beyond the limitations of previous research confined to shape-only morphing on topologically aligned datasets. + +Limitation and Future Work. Limited by the performance of existing open-source Text-to-3D models, complex 3D morphing remains challenging. In the future, we will enhance fidelity with more advanced 3D diffusion models [70, 81], enable few-shot motion interpolation [55] with temporal consistency, and explore 4D transitions [21]. Please check our supplementary materials for more implementation details and discussions. + +# 6. Acknowledgment + +This research is supported by the National Research Foundation, Singapore, under its NRF Fellowship Award (NRF-NRFF16-2024-0003). This research is also supported by MOE AcRF Tier 1 (RG97/23) and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAFICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Ahmed Abdelreheem, Ivan Skorokhodov, Maks Ovsjanikov, and Peter Wonka. Satr: Zero-shot semantic segmentation of 3d shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15166-15179, 2023. 3 +[2] Luma Labs AI. Luma dream machine: Ai-powered video content creation. https://dream-machine.lumalabs.ai/, 2025. Accessed: 2025-01-15. 7 +[3] Alyaa Qusay Aloraibi. Image morphing techniques: A review. 2023. 3 +[4] Hadar Averbuch-Elor, Daniel Cohen-Or, and Johannes Kopf. Smooth image sequences for data-driven morphing. In computer graphics forum, pages 203–213. Wiley Online Library, 2016. 2, 3 +[5] Melike Aydinlar and Yusuf Sahillioğlu. Part-based data-driven 3d shape interpolation. Computer-Aided Design, 136: 103027, 2021. 2 +[6] Thaddeus Beier and Shawn Neely. Feature-based image metamorphosis. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 529-536. 2023. 3 +[7] Bhumika G Bhatt. Comparative study of triangulation based and feature based image morphing. Signal & Image Processing, 2(4):235, 2011. 2, 3 +[8] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 7 +[9] Federica Bogo, Javier Romero, Matthew Loper, and Michael J Black. Faust: Dataset and evaluation for 3d mesh registration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3794-3801, 2014. 2 +[10] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024. 2, 4 +[11] Soheil Darabi, Eli Shechtman, Connelly Barnes, Dan B Goldman, and Pradeep Sen. Image melding: Combining inconsistent images using patch-based synthesis. ACM Transactions on graphics (TOG), 31(4):1-10, 2012. 3 +[12] Jiacheng Deng, Chuxin Wang, Jiahao Lu, Jianfeng He, Tianzhu Zhang, Jiyang Yu, and Zhe Zhang. Se-ornet: Self-ensembling orientation-aware network for unsupervised point cloud shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364–5373, 2023. 2 + +[13] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4494-4504, 2024. 2 +[14] Roberto M Dyke, Yu-Kun Lai, Paul L Rosin, Stefano Zappala, Seana Dykes, Daoliang Guo, Kun Li, Riccardo Marin, Simone Melzi, and Jingyu Yang. Shrec'20: Shape correspondence with non-isometric deformations. Computers & Graphics, 92:28-43, 2020. 2 +[15] Michal Edelstein, Danielle Ezuz, and Mirela Ben-Chen. Enigma: Evolutionary non-isometric geometry matching. arXiv preprint arXiv:1905.10763, 2019. 3 +[16] Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi-scale shape registration with functional maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12265-12274, 2020. 3 +[17] Marvin Eisenberger, David Novotny, Gael Kerchenbaum, Patrick Labatut, Natalia Neverova, Daniel Cremers, and Andrea Vedaldi. Neuromorph: Unsupervised shape interpolation and correspondence in one go. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7473-7483, 2021. 2, 3 +[18] Noa Fish, Richard Zhang, Lilach Perry, Daniel Cohen-Or, Eli Shechtman, and Connelly Barnes. Image morphing with perceptual constraints and stn alignment. In Computer Graphics Forum, pages 303–313. Wiley Online Library, 2020. 2, 3 +[19] Lawson Fulton, Vismay Modi, David Duvenaud, David IW Levin, and Alec Jacobson. Latent-space dynamics for reduced deformable simulation. In Computer graphics forum, pages 379-391. Wiley Online Library, 2019. 3 +[20] William Gao, Noam Aigerman, Thibault Groueix, Vova Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3 +[21] Chen Geng, Yunzhi Zhang, Shangzhe Wu, and Jiajun Wu. Birth and death of a rose, 2024. 2, 8 +[22] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19740–19750, 2023. 2 +[23] Qiyuan He, Jinghao Wang, Ziwei Liu, and Angela Yao. Aid: Attention interpolation of text-to-image diffusion. NeurIPS, 2024. 2, 3, 4, 6, 8 +[24] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 36, 2024. 5 +[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[26] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. + +Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4 +[27] Zehuan Huang, Yuanchen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024. 7, 8 +[28] Alec Jacobson, Zhigang Deng, Ladislav Kavan, and John P Lewis. Skinning: Real-time shape deformation (full text not available). In ACM SIGGRAPH 2014 Courses, pages 1-1. 2014. 3 +[29] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G Kim, and Rana Hanocka. Meshup: Multi-target mesh deformation via blended score distillation. arXiv preprint arXiv:2408.14899, 2024. 2 +[30] Vladimir G Kim, Yaron Lipman, and Thomas Funkhouser. Blended intrinsic maps. ACM transactions on graphics (TOG), 30(4):1-12, 2011. 3 +[31] Yushi Lan, Chen Change Loy, and Bo Dai. DDF: Correspondence distillation from nerf-based gan. IJCV, 2022. 2 +[32] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2025. 3 +[33] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. In ICLR, 2025. 2, 3 +[34] Jing Liao, Rodolfo S Lima, Diego Nehab, Hugues Hoppe, Pedro V Sander, and Jinhui Yu. Automating image morphing using structural similarity on a halfway domain. ACM Transactions on Graphics (TOG), 33(5):1-12, 2014. 2, 3 +[35] Jianchu Lin, Yinxi Gu, Guangxiao Du, Guoqiang Qu, Xiaobing Chen, Yudong Zhang, Shangbing Gao, Zhen Liu, and Nallappan Gunasekaran. 2d/3d image morphing technology from traditional to modern: A survey. Information Fusion, page 102913, 2024. 2 +[36] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 4 +[37] Robin Magnet and Maks Ovsjanikov. Scalable and efficient functional map computations on dense meshes. In Computer Graphics Forum, pages 89-101. Wiley Online Library, 2023. 2 +[38] Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, and Benjamin Bossan. Peft: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft, 2022.6 +[39] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3 +[40] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3 + +[41] Luca Morreale, Noam Aigerman, Vladimir G Kim, and Niloy J Mitra. Neural semantic surface maps. In Computer Graphics Forum, page e15005. Wiley Online Library, 2024. 3 +[42] OpenAI. Gpt-4: Openai's fourth-generation language model, 2023. 7 +[43] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 4 +[44] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. ACM Transactions on Graphics (ToG), 31(4):1-11, 2012. 3 +[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3, 4 +[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3 +[48] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. *ICLR*, 2024. 6 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4 +[50] Jing Ren, Simone Melzi, Maks Ovsjanikov, and Peter Wonka. Maptree: Recovering multiple solutions in the space of maps. ACM Transactions on Graphics, 39(6):1-17, 2020. 3 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4 +[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 4 +[53] Dvir Samuel, Rami Ben-Ari, Nir Darshan, Haggai Maron, and Gal Chechik. Norm-guided latent space exploration for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 4 + +[54] Eli Shechtman, Alex Rav-Acha, Michal Irani, and Steve Seitz. Regenerative morphing. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 615-622. IEEE, 2010. 2, 3 +[55] Liao Shen, Tianqi Liu, Huiqiang Sun, Xinyi Ye, Baopu Li, Jianming Zhang, and Zhiguo Cao. Dreammover: Leveraging the prior of diffusion models for image interpolation with large motion. arXiv preprint arXiv:2409.09605, 2, 2024. 2, 8 +[56] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3, 8 +[57] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2024. 2, 6 +[58] Justin Solomon, Fernando De Goes, Gabriel Peyre, Marco Cuturi, Adrian Butscher, Andy Nguyen, Tao Du, and Leonidas Guibas. Convolutional wasserstein distances: Efficient optimal transportation on geometric domains. ACM Transactions on Graphics (ToG), 34(4):1-11, 2015. 3 +[59] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, pages 109–116. CiteSeer, 2007. 3 +[60] Olga Sorkine, Daniel Cohen-Or, Yaron Lipman, Marc Alexa, Christian Rössl, and H-P Seidel. Laplacian surface editing. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH symposium on Geometry processing, pages 175–184, 2004. 3 +[61] Mingze Sun, Chen Guo, Puhua Jiang, Shiwei Mao, Yurun Chen, and Ruqi Huang. Srif: Semantic shape registration empowered by diffusion-based image morphing and flow estimation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 2, 3 +[62] Gary KL Tam, Zhi-Quan Cheng, Yu-Kun Lai, Frank C Langbein, Yonghuai Liu, David Marshall, Ralph R Martin, XianFang Sun, and Paul L Rosin. Registration of 3d point clouds and meshes: A survey from rigid to nonrigid. IEEE transactions on visualization and computer graphics, 19(7):1199-1217, 2012. 3 +[63] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 5 +[64] Chih-Jung Tsai, Cheng Sun, and Hwann-Tzong Chen. Multiview regenerative morphing with dual flows. In European Conference on Computer Vision, pages 492-509. Springer, 2022. 2, 3, 6, 7 +[65] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4 +[66] Shantanu Vyas, Ting-Ju Chen, Ronak R Mohanty, Peng Jiang, and Vinayak R Krishnamurthy. Latent embedded graphs for image and shape interpolation. Computer-Aided Design, 140:103091, 2021. 2 +[67] George Wolberg. Image morphing: a survey. The visual computer, 14(8-9):360-372, 1998. 3 + +[68] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v(ision) is a human-aligned evaluator for text-to-3d generation. In CVPR, 2024. 7 +[69] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In European Conference on Computer Vision, pages 378-394. Springer, 2025. 6 +[70] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2, 4, 8 +[71] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.3 +[72] Songlin Yang, Wei Wang, Jun Ling, Bo Peng, Xu Tan, and Jing Dong. Context-aware talking-head video editing. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7718-7727, 2023. 8 +[73] Songlin Yang, Wei Wang, Bo Peng, and Jing Dong. Designing a 3d-aware stylenerf encoder for face editing. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. +[74] Songlin Yang, Wei Wang, Yushi Lan, Xiangyu Fan, Bo Peng, Lei Yang, and Jing Dong. Learning dense correspondence for nef-based face reenactment. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6522-6530, 2024. 2, 8 +[75] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 5 +[76] Mehmet Ersin Yumer, Siddhartha Chaudhuri, Jessica K Hodgins, and Levent Burak Kara. Semantic shape editing using deformation handles. ACM Transactions on Graphics (TOG), 34(4):1-12, 2015. 3 +[77] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3, 5 +[78] Xiao Zhan, Rao Fu, and Daniel Ritchie. Charactermixer: Rig-aware interpolation of 3d characters. In Computer Graphics Forum, page e15047. Wiley Online Library, 2024. 2 +[79] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 3 +[80] Kaiwen Zhang, Yifan Zhou, Xudong Xu, Bo Dai, and Xinggang Pan. Diffmorpher: Unleashing the capability of diffusion models for image morphing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7912-7921, 2024. 2, 3, 4, 6, 7, 8 + +[81] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 3, 8 +[82] Zhengbo Zhang, Li Xu, Duo Peng, Hossein Rahmani, and Jun Liu. Diff-tracker: text-to-image diffusion models are unsupervised trackers. In European Conference on Computer Vision, pages 319-337. Springer, 2025. 5 +[83] Junzhe Zhu, Yuanchen Ju, Junyi Zhang, Muhan Wang, Zhecheng Yuan, Kaizhe Hu, and Huazhe Xu. Dense-matcher: Learning 3d semantic correspondence for category-level manipulation from a single demo. arXiv preprint arXiv:2412.05268, 2024. 2 +[84] Bhushan Zope and Soniya B Zope. A survey of morphing techniques. International Journal of Advanced Engineering, Management and Science, 3(2):239773, 2017. 3 \ No newline at end of file diff --git a/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/images.zip b/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8595e0a10302a59fc3c882525f2566934512c5b0 --- /dev/null +++ b/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:493dac49657e4c9ba3bcf07289bef2acaf6ad2f1f318ddac8e8c829b0c8a7840 +size 574498 diff --git a/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/layout.json b/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..941a26f6cc27069e4d6b0890a6b9651ef121afd8 --- /dev/null +++ b/2025/Textured 3D Regenerative Morphing with 3D Diffusion Prior/layout.json @@ -0,0 +1,10348 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 117, + 103, + 494, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 103, + 494, + 121 + ], + "spans": [ + { + "bbox": [ + 117, + 103, + 494, + 121 + ], + "type": "text", + "content": "Textured 3D Regenerative Morphing with 3D Diffusion Prior" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 170, + 143, + 439, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 143, + 439, + 172 + ], + "spans": [ + { + "bbox": [ + 170, + 143, + 439, + 172 + ], + "type": "text", + "content": "Songlin Yang, Yushi Lan, Honghua Chen, Xingang Pan \nS-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 109, + 174, + 498, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 174, + 498, + 185 + ], + "spans": [ + { + "bbox": [ + 109, + 174, + 498, + 185 + ], + "type": "text", + "content": "sl.yang888@outlook.com, {yushi001, honghua.chen, xingang.pan}@ntu.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 196, + 555, + 479 + ], + "blocks": [ + { + "bbox": [ + 59, + 196, + 555, + 479 + ], + "lines": [ + { + "bbox": [ + 59, + 196, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 59, + 196, + 555, + 479 + ], + "type": "image", + "image_path": "ae9c7ebaef496fc140b4e44516f54aaefcc49c649abbb99a12be1a5faed571bd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 485, + 555, + 520 + ], + "lines": [ + { + "bbox": [ + 54, + 485, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 555, + 520 + ], + "type": "text", + "content": "Figure 1. The textured 3D morphing sequences regenerated using our method with 3D diffusion prior. Our method requires no category-specific alignment data for model training and no labor-intensive preprocessing for explicit correspondence, enabling smooth and plausible morphing across diverse cross-category 3D object pairs. More video results can be found here." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 531, + 200, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 531, + 200, + 544 + ], + "spans": [ + { + "bbox": [ + 151, + 531, + 200, + 544 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 558, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 297, + 714 + ], + "type": "text", + "content": "Textured 3D morphing creates smooth and plausible interpolation sequences between two 3D objects, focusing on transitions in both shape and texture. This is important for creative applications like visual effects in filmmaking. Previous methods rely on establishing point-to-point correspondences and determining smooth deformation trajectories, which inherently restrict them to shape-only morphing on untextured, topologically aligned datasets. This restriction leads to labor-intensive preprocessing and poor generalization. To overcome these challenges, we propose a method for 3D regenerative morphing using a 3D diffusion prior. Unlike previous methods that depend on explicit correspondences and deformations, our method eliminates" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 712 + ], + "type": "text", + "content": "the additional need for obtaining correspondence and uses the 3D diffusion prior to generate morphing. Specifically, we first introduce a 3D diffusion model and interpolate the source and target information at three levels: initial noise, model parameters, and condition features. We then explore an Attention Fusion strategy to generate smoother morphing sequences. To further improve the plausibility of semantic interpolation and the generated 3D surfaces, we propose two strategies: (a) Token Reordering, where we match approximate tokens based on semantic analysis to guide implicit correspondences in the denoising process of the diffusion model, and (b) Low-Frequency Enhancement, where we enhance low-frequency signals in the tokens to improve the quality of generated surfaces. Experimental results show that our method achieves superior smoothness and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15159" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "plausibility in 3D morphing across diverse cross-category object pairs, offering a novel regenerative method for 3D morphing with textured representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 152, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 296, + 342 + ], + "type": "text", + "content": "Morphing [35] generates an interpolation sequence between a source and a target, requiring smooth and plausible transitions. This fundamental technique is crucial for creative applications like visual effects in film and media. Depending on the data type, morphing is categorized into image morphing [4, 7, 18, 23, 34, 54, 80] and 3D morphing [21, 29, 64]. Compared to image morphing, 3D morphing aligns more naturally with visual effects production, where objects and transformations inherently operate in 3D space. However, 3D morphing is more challenging, requiring the interpolation of 3D objects holistically (i.e., image morphing can be viewed as a special case of 3D morphing from a specific viewpoint). Our work addresses the task of textured 3D morphing [64], which generates a sequence for two textured 3D representations, aiming for smooth and plausible transitions in shape and texture, as shown in Fig. 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 450 + ], + "type": "text", + "content": "Previous 3D morphing methods mainly focused on morphing shapes, which can be summarized in two steps: first, establishing correspondence [12, 13, 31, 37] between the 3D representations of the source and target objects, and second, determining smooth and plausible deformation trajectories between corresponding 3D points [17, 61, 78]. Following these steps, previous methods blend the two 3D representations with respective weights to obtain a sequence of interpolated 3D representations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 605 + ], + "type": "text", + "content": "Due to the scarcity of topologically aligned and textured 3D datasets, previous 3D morphing methods [17, 61, 78] have mainly focused on in-domain untextured datasets, such as FAUST [9] (human shapes) and Shrec'20 [14] (quadruped animals). As a result, these methods are limited to shape-only morphing with two key generalization challenges: (a) Labor-Intensive Preprocessing: Morphing new in-domain 3D data with these methods [17, 78] requires domain-specific alignment with the training datasets through tedious registration [61] and matching [83] steps. (b) Limited Morphing Capacity: The methods [5, 17, 66, 78] suffer from limited object diversity and small datasets, leading to ambiguous and implausible interpolations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 606, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 296, + 666 + ], + "type": "text", + "content": "The limitations of previous methods inspire us to consider two critical questions: (a) Is explicit point-to-point correspondence truly necessary? (b) Can we enhance the generalization capability of textured 3D morphing via a generic generative prior?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": "For correspondence, obtaining dense correspondences between textured 3D representations across categories remains underexplored [83] despite being foundational to previous methods, while explicit constraints can instead limit" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "the creativity of morphing. Therefore, a promising method is using labor-saving implicit correspondences [31, 74] to guide the morphing. For instance, morphing could be formulated as an optimization problem, allowing correspondences to emerge automatically [64]. Alternatively, attention mechanisms in generative models could enable automatic alignment during object blending [23, 55, 80]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 158, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 553, + 312 + ], + "type": "text", + "content": "For generative priors, recent advancements in diffusion-based generation models offer two strategies for 3D morphing: using 3D diffusion models directly [10, 33, 70], or enhancing 2D models with 3D priors. However, those hybrid 2D-3D methods [22, 29, 47, 51] face significant challenges: 2D models lack a holistic 3D knowledge, and optimizing the 2D-3D mapping cannot guarantee that the morphing process will maintain 3D consistency across viewpoints. Therefore, directly using a 3D diffusion model to regenerate the interpolated 3D representations (i.e., regenerative morphing [54, 64]) enables authentic 3D morphing and has the potential to be scaled up by using a more state-of-the-art 3D generation model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 315, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 553, + 374 + ], + "type": "text", + "content": "Building on the above analysis, we adopt a generic 3D diffusion prior that leverages its implicit correspondence and 3D generation capabilities to blend source and target information, enabling the regeneration of interpolated textured 3D representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 553, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 651 + ], + "type": "text", + "content": "Specifically, we first introduce a 3D diffusion model [33] and interpolate the information from the source and target at three levels: initial noises, model parameters, and condition features. The 3D representation for each interpolation is then regenerated using the 3D generation model. To improve the 3D diffusion model's ability to generate smooth morphing sequences, we explore an extended Attention Fusion strategy [23, 80]. However, fusing different information weakened the model's denoising capability, and aggressively applying Attention Fusion to all denoising steps for smoother morphing resulted in implausible outcomes. Therefore, to improve the plausibility of semantic interpolation and the generated 3D surfaces, we propose two strategies: (a) Token Reordering: After semantic analysis, we identify semantic correspondences in diffusion tokens and propose matching approximate tokens before attention computation to better guide implicit correspondences in the diffusion space. (b) Low-Frequency Enhancement: Frequency-domain analysis [57] reveals that boosting low-frequency signals improves the surface quality of the 3D diffusion model. Thus, we enhance low-frequency signals at key time steps to preserve the model's ability to generate 3D surfaces." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Our contributions can be summarized as follows: (a) We are the first to use a generic 3D diffusion prior for morphing textured 3D representations, enabling 3D morphing without explicit correspondences. (b) We analyze the merging of source and target information during morphing with" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15160" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "content": "a 3D diffusion prior from semantic and frequency perspectives, proposing Token Reordering and Low-Frequency Enhancement to improve plausibility. (c) Extensive experimental results demonstrate that our method achieves superior smoothness and plausibility in performing 3D morphing across diverse cross-category object pairs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 142, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 142, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 142, + 168 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 177, + 157, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 157, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 157, + 190 + ], + "type": "text", + "content": "2.1. Image Morphing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 195, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 295, + 351 + ], + "type": "text", + "content": "Image morphing is a long-standing challenge in computer vision and graphics [3, 67, 84]. Traditional methods [6, 7, 11, 34, 54] use feature-based warping and blending to create smooth transitions but struggle to generate new content, often leading to artifacts. Data-driven methods [4, 18] leverage large single-class datasets to achieve smoother results but are limited in cross-domain or personalized applications due to their reliance on specific data. The methods like DiffMorpher [80] and AID [23] address this by utilizing pre-trained diffusion models on diverse datasets, enabling flexible morphing across a wide range of object categories. Inputting multi-view images to image morphing methods enables 3D-aware morphing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 142, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 142, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 142, + 372 + ], + "type": "text", + "content": "2.2.3D Morphing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 378, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 653 + ], + "type": "text", + "content": "Previous 3D morphing methods focus on shape-only correspondence [62], and realize 3D morphing through interpolation or deformation between corresponding 3D primitives (e.g., points, vertices, and faces). They can be divided into (a) Axiomatic Methods: They tend to rely on sparse landmarks [15, 30] or use functional maps [44] to address under-constrained mapping spaces, such as MapTree [50] and SmoothShells [16]. Energy-minimizing functions [59, 60] and skinning methods [19, 28] enable deformation control, while optimal transportation [58, 64] approximates shape correspondence. (b) Learning-Based Methods: This category of methods can be categorized into introducing other generative priors and using aligned data. SATR [1] introduces the semantic labels from multi-view images. NSSM [41] uses DINO [43] for sparse landmarks and SRIF [61] introduces large vision models [80] for semantic shape registration and morphing. A recent class of works leverages text prompts as user inputs for driving a deformation towards an arbitrary textual prompt [20, 39, 40], but CLIP [49] objective lacks a full understanding of object details. For data prior, category-specific training [76] on a topology-aligned dataset such as NeuroMorph [17] is also prevailing in learning-based 3D shape analysis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "These methods focus on shape-only morphing or rely on rigging annotations, requiring extra effort due to 2D-3D mismatches. In contrast, our work enables textured 3D morphing without learning correspondence or deformation, using a 3D diffusion model to regenerate interpolated rep" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "resentations, addressing previous challenges and offering a new direction for 3D morphing research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 108, + 429, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 429, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 429, + 119 + ], + "type": "text", + "content": "2.3. 3D Diffusion Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 127, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 555, + 319 + ], + "type": "text", + "content": "Generative 3D priors fall into two types: native 3D diffusion models and 3D-aware diffusion models. Due to the scarcity of 3D data, methods leveraging 2D priors to generate 3D or multi-view content have been proposed. For instance, Score Distillation Sampling [47] distills 3D information from a 2D diffusion model. 3D-aware generation can be divided into two steps: multi-view image generation [56] followed by feed-forward 3D reconstruction [71]. However, these methods inherently lack a 3D latent space. To address this, native 3D diffusion models [32, 77, 79, 81] that encode and learn 3D representations have been introduced. These models typically consist of two steps: training a VAE to encode 3D data and training a diffusion model based on corresponding latent codes. We use Gaussian Anything [33] as the 3D diffusion prior, encoding 3D information in a point cloud latent space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 334, + 370, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 370, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 370, + 346 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 355, + 555, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 451 + ], + "type": "text", + "content": "The framework for 3D regenerative morphing based on the 3D diffusion prior, as shown in Fig. 2, consists of three main steps: Basic Interpolation, where essential information is interpolated (Sec. 3.2); Smoothness Improvement, achieved through an Attention Fusion mechanism (Sec. 3.3); and Plausibility Improvement, which involves two strategies, Token Reordering (Sec. 3.4) and Low-Frequency Enhancement (Sec. 3.5)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 463, + 394, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 394, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 394, + 475 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 481, + 426, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 426, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 426, + 493 + ], + "type": "text", + "content": "3.1.1. 3D Diffusion Model" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": "We select Gaussian Anything [33] as our 3D diffusion prior, a two-stage native 3D diffusion model with a structured latent representation (i.e., point cloud) and the DiT [46] structure. It consists of a geometry generation model " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{G}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " and a texture generation model " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{T}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ". In the first stage, the model " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{G}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " takes a Gaussian initial noise " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{G}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " and textual conditioning information " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " as inputs, to generate a structured point cloud representation " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\text {point-cloud }} = \\epsilon_{G}(\\mathbf{z}_{G}, \\mathbf{c})" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ". In the second stage, " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\text {point-cloud }}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " is added to a Gaussian initial noise " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{T}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ", and the resulting variable is denoised by " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{T}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " with condition " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " to obtain the final texture feature " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\text {feature }} = \\epsilon_{T}(\\mathbf{x}_{\\text {point-cloud }} + \\mathbf{z}_{T}, \\mathbf{c})" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\text {point-cloud }}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\text {feature }}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " are fed into a pre-trained decoder " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " to produce the final 3D Gaussian representation " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{3D} = \\mathcal{D}(\\mathbf{x}_{\\text {point-cloud}}, \\mathbf{x}_{\\text {feature}})" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ", which can be rendered as multi-view images. Notably, we denote the inherently unordered point-cloud-based token sequences between DiT blocks as " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{h_{j}\\}_{j=1}^{M}" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 498, + 555, + 713 + ], + "type": "text", + "content": " is the length of token sequence." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15161" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 553, + 189 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 189 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 189 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 189 + ], + "type": "image", + "image_path": "cf650e840532d63ca4a1a001466ed896c27bf33835f6e0f00260cf20ea6343b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 191, + 211, + 290 + ], + "blocks": [ + { + "bbox": [ + 60, + 191, + 211, + 290 + ], + "lines": [ + { + "bbox": [ + 60, + 191, + 211, + 290 + ], + "spans": [ + { + "bbox": [ + 60, + 191, + 211, + 290 + ], + "type": "image", + "image_path": "58e45fb395ccbe570fb1b0057a38d046a9cc323574d17958fcf3ac9b85d00759.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 295, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 555, + 319 + ], + "type": "text", + "content": "Figure 2. The framework of our method. Beyond basic interpolation, Attention Fusion is explored to improve smoothness, while Token Reordering and Low-Frequency Enhancement are proposed to improve plausibility." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 213, + 190, + 411, + 290 + ], + "blocks": [ + { + "bbox": [ + 213, + 190, + 411, + 290 + ], + "lines": [ + { + "bbox": [ + 213, + 190, + 411, + 290 + ], + "spans": [ + { + "bbox": [ + 213, + 190, + 411, + 290 + ], + "type": "image", + "image_path": "ad4b57356c64e20c0552fe9d21127835754d300b14fee184931e2e5150eb4cf9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 414, + 191, + 550, + 290 + ], + "blocks": [ + { + "bbox": [ + 414, + 191, + 550, + 290 + ], + "lines": [ + { + "bbox": [ + 414, + 191, + 550, + 290 + ], + "spans": [ + { + "bbox": [ + 414, + 191, + 550, + 290 + ], + "type": "image", + "image_path": "4c801e05a74c58d48307c7ebab74d66fdcce3bd3b3d2d8208717aabdd1a1e80b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 324, + 124, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 124, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 124, + 335 + ], + "type": "text", + "content": "3.1.2. Attention" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "type": "text", + "content": "The attention [65] mechanism is an important component for the current text-driven diffusion models [10, 46, 51, 52, 70], especially cross-attention and self-attention. Given a latent variable " + }, + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{d_z}" + }, + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "type": "text", + "content": ", a text condition " + }, + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^{d_c}" + }, + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "type": "text", + "content": ", and the attention layer with matrices " + }, + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "type": "inline_equation", + "content": "W_{Q} \\in \\mathbb{R}^{d_{z} \\times d_{q}}" + }, + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "type": "inline_equation", + "content": "W_{K} \\in \\mathbb{R}^{d_{c} \\times d_{k}}" + }, + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "type": "inline_equation", + "content": "W_{V} \\in \\mathbb{R}^{d_{c} \\times d_{v}}" + }, + { + "bbox": [ + 55, + 338, + 296, + 407 + ], + "type": "text", + "content": ", the cross-attention is computed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 77, + 407, + 295, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 407, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 77, + 407, + 295, + 433 + ], + "type": "interline_equation", + "content": "A (z, c) = A t t n (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d _ {k}}}\\right) V, \\tag {1}", + "image_path": "cae9fae3b41905404efed1a5d6e692b4b1e2395b4fcd2e20fd05f59ff41fc6d6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 433, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 433, + 296, + 458 + ], + "type": "inline_equation", + "content": "Q = W_{Q}^{\\top}z, K = W_{K}^{\\top}c, V = W_{V}^{\\top}c" + }, + { + "bbox": [ + 55, + 433, + 296, + 458 + ], + "type": "text", + "content": ". Self-attention is a specific form of cross-attention, computed as " + }, + { + "bbox": [ + 55, + 433, + 296, + 458 + ], + "type": "inline_equation", + "content": "A(z,z)" + }, + { + "bbox": [ + 55, + 433, + 296, + 458 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 464, + 168, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 168, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 168, + 477 + ], + "type": "text", + "content": "3.2. Basic Interpolation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 482, + 152, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 152, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 152, + 494 + ], + "type": "text", + "content": "3.2.1. Implementation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "text", + "content": "The basic interpolation has three levels, with source and target weights given by " + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "inline_equation", + "content": "(1 - \\alpha)" + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "text", + "content": ", respectively, where " + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "text", + "content": " generates the source " + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{3D}^{src}" + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "text", + "content": " generates the target " + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{3D}^{tgt}" + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 544, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 55, + 544, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 618 + ], + "type": "text", + "content": "(a) Initial Noises: The textured 3D representations " + }, + { + "bbox": [ + 55, + 544, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{3D}^{src}" + }, + { + "bbox": [ + 55, + 544, + 295, + 618 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{3D}^{tgt}" + }, + { + "bbox": [ + 55, + 544, + 295, + 618 + ], + "type": "text", + "content": ", are inverted via diffusion inversion [36] to obtain their respective input noises, " + }, + { + "bbox": [ + 55, + 544, + 295, + 618 + ], + "type": "inline_equation", + "content": "[\\mathbf{z}_T^{src},\\mathbf{z}_G^{src}]" + }, + { + "bbox": [ + 55, + 544, + 295, + 618 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 544, + 295, + 618 + ], + "type": "inline_equation", + "content": "[\\mathbf{z}_T^{tgt},\\mathbf{z}_G^{tgt}]" + }, + { + "bbox": [ + 55, + 544, + 295, + 618 + ], + "type": "text", + "content": ". To ensure that Gaussian noise properties are preserved, spherical linear interpolation [53] is applied to these noises to generate the interpolated noises, " + }, + { + "bbox": [ + 55, + 544, + 295, + 618 + ], + "type": "inline_equation", + "content": "[\\mathbf{z}_T^\\alpha ,\\mathbf{z}_G^\\alpha ]" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": "(b) Model Parameters: Given " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{3D}^{src}" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{3D}^{tgt}" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": ", we fine-tune the model using LoRA (Low-Rank Adaptation) [26] to obtain two sets of LoRA parameters. These parameters are then linearly interpolated and fused to obtain the morphing models " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\epsilon_G^\\alpha" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\epsilon_T^\\alpha" + }, + { + "bbox": [ + 55, + 617, + 295, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "(c) Condition Features: We adopt text conditions as they provide a more consistent semantic space and are layout-independent compared to DINO-based [43] image condi" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 324, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 361 + ], + "type": "text", + "content": "tions. The text prompts of the source and target 3D representations are encoded via a CLIP [49] encoder into " + }, + { + "bbox": [ + 313, + 324, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{src}" + }, + { + "bbox": [ + 313, + 324, + 554, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 324, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{tgt}" + }, + { + "bbox": [ + 313, + 324, + 554, + 361 + ], + "type": "text", + "content": ", which are linearly interpolated to produce " + }, + { + "bbox": [ + 313, + 324, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{\\alpha}" + }, + { + "bbox": [ + 313, + 324, + 554, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 369, + 383, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 383, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 383, + 380 + ], + "type": "text", + "content": "3.2.2. Problems" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 384, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 408 + ], + "type": "text", + "content": "Basic interpolation realizes the blending of basic information, but there are still the following problems:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 409, + 554, + 518 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 313, + 409, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 554, + 468 + ], + "type": "text", + "content": "(a) Abrupt Changes (Smoothness): Nonlinear multi-step denoising in diffusion models introduces variability in noise-to-data mapping, while the CLIP encoder's space lacks guaranteed semantic smoothness, leading to abrupt changes (yellow-highlighted part in Fig. 8)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 469, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 554, + 518 + ], + "type": "text", + "content": "(b) Artifacts (Plausibility): Misalignment between conditioning and diffusion spaces disrupts learned mappings, causing artifacts like structure collapse or degraded surface quality (red and yellowhighlighted parts in Fig. 8)." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 518, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 554, + 555 + ], + "type": "text", + "content": "Therefore, we investigate Attention Fusion (Sec. 3.3) for smoothness and Token Reordering (Sec. 3.4) with Low-Frequency Enhancement (Sec. 3.5) for plausibility." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 563, + 416, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 416, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 416, + 575 + ], + "type": "text", + "content": "3.3. Attention Fusion" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "The fusion of attention has been proven effective in improving morphing smoothness in 2D diffusion-based image morphing. However, DiffMorpher [80] only explored self-attention interpolation and did not consider the smoothness of semantic features in conditioning. AID [23] did not address the differences between the original and LoRA models, neglecting conditioning feature interpolation with accurate inversion. Our method combines self-attention and cross-attention fusion, using unified attention features from fine-tuned models to enhance smoothness while encouraging plausible generation." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15162" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "Specifically, as shown in Fig. 2, we first feed " + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "inline_equation", + "content": "z^{src}" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "inline_equation", + "content": "z^{tgt}" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "inline_equation", + "content": "z^{\\alpha}" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": " into blocks of the morphing model " + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "inline_equation", + "content": "\\epsilon^{\\alpha}" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": " to obtain three sets of " + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "inline_equation", + "content": "(Q^{src}, K^{src}, V^{src})" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "inline_equation", + "content": "(Q^{tgt}, K^{tgt}, V^{tgt})" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "inline_equation", + "content": "(Q^{\\alpha}, K^{\\alpha}, V^{\\alpha})" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": ". Then, based on Eq. (1), denoting concatenation as " + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "inline_equation", + "content": "[\\cdot, \\cdot]" + }, + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": ", we obtain the fused attention by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 137, + 221, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 137, + 221, + 149 + ], + "spans": [ + { + "bbox": [ + 66, + 137, + 221, + 149 + ], + "type": "interline_equation", + "content": "F u s e d - A t t n (Q ^ {\\alpha}, K ^ {\\alpha}, V ^ {\\alpha}) = A t t n (Q ^ {\\alpha},", + "image_path": "be492c8f3e278d422a8b721290f259f457f7c4326deed634d3da2b30253646bf.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 152, + 295, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 152, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 69, + 152, + 295, + 174 + ], + "type": "interline_equation", + "content": "\\left[ (1 - \\alpha) K ^ {\\mathrm {s r c}} + \\alpha K ^ {\\mathrm {t g t}}, K ^ {\\alpha} \\right], \\left[ (1 - \\alpha) V ^ {\\mathrm {s r c}} + \\alpha V ^ {\\mathrm {t g t}}, V ^ {\\alpha} \\right]). \\tag {2}", + "image_path": "79aba58b4d1e1a88c47096cfc2ccf4e33c4085db5351a45a54d092dcb8e53c8e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 173, + 295, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 173, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 295, + 232 + ], + "type": "text", + "content": "Using the interpolated model for all attention calculations reduces quality deterioration caused by Attention Fusion. However, excessive fusion across time steps harms plausibility, causing structural collapse and surface issues (see Fig. 8)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 240, + 163, + 252 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 163, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 163, + 252 + ], + "type": "text", + "content": "3.4. Token Reordering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 257, + 130, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 130, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 130, + 268 + ], + "type": "text", + "content": "3.4.1. Motivation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 273, + 295, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 273, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 295, + 440 + ], + "type": "text", + "content": "The point clouds of 3D objects are tokenized into sequences " + }, + { + "bbox": [ + 54, + 273, + 295, + 440 + ], + "type": "inline_equation", + "content": "\\{h_j\\}_{j=1}^M" + }, + { + "bbox": [ + 54, + 273, + 295, + 440 + ], + "type": "text", + "content": ", with each token " + }, + { + "bbox": [ + 54, + 273, + 295, + 440 + ], + "type": "inline_equation", + "content": "h_j" + }, + { + "bbox": [ + 54, + 273, + 295, + 440 + ], + "type": "text", + "content": " representing a 3D point. However, point clouds are inherently unordered representations [77], and relying solely on the attention mechanism to match the points in different 3D objects may lead the model to make semantically implausible connections (e.g., the connected tokens may represent \"chair legs\" and \"donut frosting\", respectively). This vanilla use of the diffusion prior does not fully exploit its potential, as works like DIFT [63] show that 2D diffusion features or tokens represent semantics [24, 75, 82]. We infer that 3D diffusion features also capture semantic correspondences. Therefore, why not first pair points with similar semantics and then interpolate within this semantically plausible space?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 447, + 180, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 180, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 180, + 459 + ], + "type": "text", + "content": "3.4.2. Experimental Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 462, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 295, + 521 + ], + "type": "text", + "content": "To test our hypothesis, we first present the problem of position-identical tokens in capturing correspondence (i.e., a problem of vanilla Attention Fusion). We then reveal the emergent correspondence between position-different tokens, supporting the validity of Token Reordering." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "(a) Problem of Vanilla Attention Fusion. The DiT models tend to have stronger semantic correspondence in the intermediate time steps [75], and shorter distances between two tokens indicate their higher semantic similarity. To test this in 3D diffusion, we first obtain fully aligned 3D objects by scaling to test our model's correspondence. For instance, as shown in Fig. 3, the right chair results from scaling the left chair's x-axis by 0.6 and z-axis by 1.2. This ensures that tokens at the identical sequence positions represent the originally same point-cloud points (i.e., the points shift positions after scaling while retaining the same color and semantics). Ideally, in geometry diffusion, token distance at the same position should first decrease (semantic learning) until step 200, then increase (point location learning). In texture diffusion, it should gradually decrease, ensuring the color information at the identical position consistent. But" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 460, + 136 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 460, + 136 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 460, + 136 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 460, + 136 + ], + "type": "image", + "image_path": "9c96dc898f1a2d206a32ebb1cbb9811a84095920a79f7f6600360799df986be4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 141, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 313, + 141, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 141, + 553, + 186 + ], + "type": "text", + "content": "Figure 3. Mean squared error of token distances at identical positions. The solid line represents real data, while the red dashed line shows the ideal trend (i.e., if position-identical tokens can capture the correspondence). Light blue blocks highlight their conflicts." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 460, + 72, + 553, + 136 + ], + "blocks": [ + { + "bbox": [ + 460, + 72, + 553, + 136 + ], + "lines": [ + { + "bbox": [ + 460, + 72, + 553, + 136 + ], + "spans": [ + { + "bbox": [ + 460, + 72, + 553, + 136 + ], + "type": "image", + "image_path": "072599d6a2516dedcb71ec73e934ac11bb0b31909dd8b1447ab9a82777a7a7d0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 193, + 457, + 256 + ], + "blocks": [ + { + "bbox": [ + 317, + 193, + 457, + 256 + ], + "lines": [ + { + "bbox": [ + 317, + 193, + 457, + 256 + ], + "spans": [ + { + "bbox": [ + 317, + 193, + 457, + 256 + ], + "type": "image", + "image_path": "986c9ee0e976479b4eb1c2c85dd4c169515a3a38aabd73e763b897003bf71453.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 262, + 553, + 294 + ], + "lines": [ + { + "bbox": [ + 313, + 262, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 553, + 294 + ], + "type": "text", + "content": "Figure 4. Mean squared error of point distances between token-distance-closest points. The red dashed line indicates the time step range where semantic correspondence emerges." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 457, + 193, + 550, + 255 + ], + "blocks": [ + { + "bbox": [ + 457, + 193, + 550, + 255 + ], + "lines": [ + { + "bbox": [ + 457, + 193, + 550, + 255 + ], + "spans": [ + { + "bbox": [ + 457, + 193, + 550, + 255 + ], + "type": "image", + "image_path": "7281a51f332585ab661e4035e74d0fe13156ce8b9f9498a2cdcdf3587e1176aa.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 300, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 348 + ], + "type": "text", + "content": "the Fig. 3 indicates that position-identical semantic alignment is decreasing during denoising, especially the texture diffusion. The vanilla Attention Fusion links these positional tokens, which causes artifacts (see Fig. 8, 3rd row)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 349, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 552 + ], + "type": "text", + "content": "(b) Emergence of Semantic Correspondence. By extracting the corresponding points in the final point cloud for two tokens and computing their distance, we can assess the semantic matching of the tokens. We obtain token sequences for the same 3D point cloud through different random seeds, meaning tokens at the identical sequence positions do not correspond to the same point-cloud points. For example, as shown in Fig. 4, the 3D point at the same position may correspond to the " + }, + { + "bbox": [ + 313, + 349, + 554, + 552 + ], + "type": "inline_equation", + "content": "1st" + }, + { + "bbox": [ + 313, + 349, + 554, + 552 + ], + "type": "text", + "content": " token (the left chair) and the " + }, + { + "bbox": [ + 313, + 349, + 554, + 552 + ], + "type": "inline_equation", + "content": "M-th" + }, + { + "bbox": [ + 313, + 349, + 554, + 552 + ], + "type": "text", + "content": " token (the right chair) respectively. As shown in Fig. 4, the distance between points based on geometry token distance decreases with increasing time steps, indicating emergent semantic correspondence. Conversely, the distance based on texture token distance first decreases and then increases, suggesting that the denoising process initially promotes semantic alignment, but later shifts towards learning texture details. This aligns with findings [75] in 2D diffusion." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 552, + 553, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 553, + 589 + ], + "type": "text", + "content": "Therefore, by considering both geometry and texture denoising characteristics, we reorder tokens in the intermediate stages to better guide plausible morphing." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 595, + 410, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 595, + 410, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 410, + 607 + ], + "type": "text", + "content": "3.4.3. Implementation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "text", + "content": "Based on these observations, we reorder the token sequences before passing the output of the " + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "text", + "content": "-th block to the " + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "inline_equation", + "content": "(i + 1)" + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "text", + "content": "-th block, where " + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,\\dots,N - 1\\}" + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "text", + "content": " is the total number of blocks in the diffusion model. As shown in Fig. 2 and Alg. 1, we reorder the source and target token sequences " + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\{h_j^{\\mathrm{src}}\\}_{j = 1}^M" + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\{h_j^{\\mathrm{tgt}}\\}_{j = 1}^M" + }, + { + "bbox": [ + 313, + 609, + 554, + 687 + ], + "type": "text", + "content": " such that tokens with similar distances align at the same index, as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 379, + 686, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 686, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 379, + 686, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\underset {j = 1} {\\text {m i n i m i z e}} \\sum_ {j = 1} ^ {M} \\left\\| h _ {j} ^ {\\mathrm {s r c}} - h _ {\\sigma (j)} ^ {\\mathrm {t g t}} \\right\\|, \\tag {3}", + "image_path": "145226c1f764366e62c652a1b95539ab5e14a95dd99c132d731c9744742a8add.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15163" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 185, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 185, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 185, + 84 + ], + "type": "text", + "content": "Algorithm 1 Token Reordering" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 87, + 294, + 191 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 62, + 87, + 289, + 99 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 87, + 289, + 99 + ], + "spans": [ + { + "bbox": [ + 62, + 87, + 289, + 99 + ], + "type": "text", + "content": "1: Input: Source token sequence " + }, + { + "bbox": [ + 62, + 87, + 289, + 99 + ], + "type": "inline_equation", + "content": "\\{h_j^{\\mathrm{src}}\\}_{j = 1}^M" + }, + { + "bbox": [ + 62, + 87, + 289, + 99 + ], + "type": "text", + "content": " target token sequence " + }, + { + "bbox": [ + 62, + 87, + 289, + 99 + ], + "type": "inline_equation", + "content": "\\{h_j^{\\mathrm{tgt}}\\}_{j = 1}^M" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 99, + 228, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 99, + 228, + 110 + ], + "spans": [ + { + "bbox": [ + 63, + 99, + 228, + 110 + ], + "type": "text", + "content": "2: Output: Reordered sequences " + }, + { + "bbox": [ + 63, + 99, + 228, + 110 + ], + "type": "inline_equation", + "content": "\\{\\tilde{h}_j^{\\mathrm{src}}\\}_{j = 1}^M" + }, + { + "bbox": [ + 63, + 99, + 228, + 110 + ], + "type": "inline_equation", + "content": "\\{\\tilde{h}_j^{\\mathrm{tgt}}\\}_{j = 1}^M" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 110, + 294, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 110, + 294, + 118 + ], + "spans": [ + { + "bbox": [ + 63, + 110, + 294, + 118 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 63, + 110, + 294, + 118 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 63, + 110, + 294, + 118 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 63, + 110, + 294, + 118 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 63, + 110, + 294, + 118 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 63, + 110, + 294, + 118 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 63, + 110, + 294, + 118 + ], + "type": "text", + "content": " Efficient computation by PyTorch [45] operations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 118, + 279, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 118, + 279, + 127 + ], + "spans": [ + { + "bbox": [ + 63, + 118, + 279, + 127 + ], + "type": "text", + "content": "4: Compute " + }, + { + "bbox": [ + 63, + 118, + 279, + 127 + ], + "type": "inline_equation", + "content": "D_{j,k} = \\| h_j^{\\mathrm{src}} - h_k^{\\mathrm{tgt}}\\|" + }, + { + "bbox": [ + 63, + 118, + 279, + 127 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 63, + 118, + 279, + 127 + ], + "type": "inline_equation", + "content": "\\sigma (j) = \\arg \\min_{k}D_{j,k},\\forall j" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 128, + 133, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 128, + 133, + 135 + ], + "spans": [ + { + "bbox": [ + 63, + 128, + 133, + 135 + ], + "type": "text", + "content": "5: if " + }, + { + "bbox": [ + 63, + 128, + 133, + 135 + ], + "type": "inline_equation", + "content": "\\alpha < 0.5" + }, + { + "bbox": [ + 63, + 128, + 133, + 135 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 135, + 243, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 135, + 243, + 146 + ], + "spans": [ + { + "bbox": [ + 63, + 135, + 243, + 146 + ], + "type": "text", + "content": "6: " + }, + { + "bbox": [ + 63, + 135, + 243, + 146 + ], + "type": "inline_equation", + "content": "\\{\\tilde{h}_j^{\\mathrm{tgt}}\\}_{j = 1}^M\\gets \\{h_{\\sigma (j)}^{\\mathrm{tgt}}\\} ,\\{\\tilde{h}_j^{\\mathrm{src}}\\}_{j = 1}^M\\gets \\{h_j^{\\mathrm{src}}\\}_{j = 1}^M" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 146, + 92, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 146, + 92, + 152 + ], + "spans": [ + { + "bbox": [ + 63, + 146, + 92, + 152 + ], + "type": "text", + "content": "7:els" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 152, + 252, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 152, + 252, + 164 + ], + "spans": [ + { + "bbox": [ + 63, + 152, + 252, + 164 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 63, + 152, + 252, + 164 + ], + "type": "inline_equation", + "content": "\\{\\tilde{h}_j^{\\mathrm{src}}\\}_{j = 1}^M\\leftarrow \\{h_{\\sigma -1(j)}^{\\mathrm{src}}\\} ,\\{\\tilde{h}_j^{\\mathrm{tgt}}\\}_{j = 1}^M\\leftarrow \\{h_j^{\\mathrm{tgt}}\\}_{j = 1}^M" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 165, + 103, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 165, + 103, + 171 + ], + "spans": [ + { + "bbox": [ + 62, + 165, + 103, + 171 + ], + "type": "text", + "content": "9: end if" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 172, + 96, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 172, + 96, + 178 + ], + "spans": [ + { + "bbox": [ + 62, + 172, + 96, + 178 + ], + "type": "text", + "content": "10: end for" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 179, + 162, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 179, + 162, + 191 + ], + "spans": [ + { + "bbox": [ + 62, + 179, + 162, + 191 + ], + "type": "text", + "content": "11: Return " + }, + { + "bbox": [ + 62, + 179, + 162, + 191 + ], + "type": "inline_equation", + "content": "\\{\\tilde{h}_j^{\\mathrm{src}}\\}_{j = 1}^M" + }, + { + "bbox": [ + 62, + 179, + 162, + 191 + ], + "type": "inline_equation", + "content": "\\{\\tilde{h}_j^{\\mathrm{tgt}}\\}_{j = 1}^M" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "text", + "content": " is the number of tokens, and " + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\sigma(j)" + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "text", + "content": " denotes the index of the element in the target sequence that best corresponds to the " + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "text", + "content": "-th element in the source sequence. To further refine the generation, we set different reordering strategies depending on " + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "text", + "content": ": For " + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0, 0.5)" + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "text", + "content": ", the target token sequence is reordered based on the source token sequence. For " + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0.5, 1]" + }, + { + "bbox": [ + 55, + 200, + 296, + 297 + ], + "type": "text", + "content": ", the source token sequence is reordered based on the target token sequence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 302, + 219, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 219, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 219, + 315 + ], + "type": "text", + "content": "3.5. Low-Frequency Enhancement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 319, + 130, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 130, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 130, + 330 + ], + "type": "text", + "content": "3.5.1. Motivation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "text", + "content": "Due to the additional Attention Fusion operations, the misalignment between the condition space and the diffusion space is exacerbated. Excessive Attention Fusion at later time steps significantly degrades the model's denoising capability, with the deterioration becoming more pronounced as the fusion time step increases (see Fig. 8, from the " + }, + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "inline_equation", + "content": "4th" + }, + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "inline_equation", + "content": "5th" + }, + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "text", + "content": " row). This raises the question: what part of the " + }, + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "text", + "content": " diffusion model is influenced by the morphing operations, leading to this decline in performance?" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 443, + 180, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 180, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 180, + 455 + ], + "type": "text", + "content": "3.5.2. Experimental Analysis" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 458, + 296, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 296, + 614 + ], + "type": "text", + "content": "The deterioration problem is analyzed in the frequency domain to address the significant visual differences between 3D objects, as shown in Fig. 5. In 3D generation, low-frequency noise controls the overall layout, while high-frequency noise governs surface details. Excessive amplification of high-frequency components during denoising can interfere with the low-frequency components, degrading overall quality. Therefore, enhancing low-frequency signals during denoising is crucial to prevent the model from overemphasizing high-frequency noise, thus improving the quality of 3D surface generation. Similar patterns appear in 2D diffusion studies [48, 57, 69], with low frequencies tied to image layout and high frequencies to details." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 618, + 152, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 152, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 152, + 629 + ], + "type": "text", + "content": "3.5.3. Implementation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 632, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 295, + 666 + ], + "type": "text", + "content": "To address this, as shown in Fig. 2, we propose to enhance the low-frequency signal when generating the 3D interpolations. Specifically, the process is defined as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 111, + 670, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 670, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 111, + 670, + 295, + 682 + ], + "type": "interline_equation", + "content": "F (h) = \\operatorname {F F T} (h), \\tag {4}", + "image_path": "c69759fbca910c3696c90ecaece6763d97a087bb637eb51c5736745b9f3d071b.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 93, + 685, + 294, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 685, + 294, + 696 + ], + "spans": [ + { + "bbox": [ + 93, + 685, + 294, + 696 + ], + "type": "interline_equation", + "content": "F _ {\\omega < \\omega_ {0}} ^ {\\prime} (h) = F _ {\\omega < \\omega_ {0}} (h) \\odot \\text {s c a l e}, \\tag {5}", + "image_path": "7328e09873ea8cabf6685bf8f077069b6cb422c9c55eaf1b35ff57dda7ac1283.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 123, + 698, + 294, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 698, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 123, + 698, + 294, + 711 + ], + "type": "interline_equation", + "content": "h ^ {\\prime} = \\operatorname {I F F T} \\left(\\left[ F _ {\\omega < \\omega_ {0}} ^ {\\prime} (h), F _ {\\omega \\geq \\omega_ {0}} (h) \\right]\\right), \\tag {6}", + "image_path": "2f48b86c81e5e4afc8332eb8cbe187e2fb7760761486cbcabe5478f62245ec1b.jpg" + } + ] + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 435, + 150 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 435, + 150 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 435, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 435, + 150 + ], + "type": "image", + "image_path": "53ffca7b0d30eaded2579e00f4f29007da582c821af681bca5592e63b2dcc9a8.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 154, + 554, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 154, + 554, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 554, + 209 + ], + "type": "text", + "content": "Figure 5. The changes of high and low frequency signals in the diffusion model during the denoising process. Visualizing signal amplitudes at different time steps reveals that low-frequency noise varies less than high-frequency noise, with smaller gaps across denoising time steps." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 435, + 71, + 553, + 150 + ], + "blocks": [ + { + "bbox": [ + 435, + 71, + 553, + 150 + ], + "lines": [ + { + "bbox": [ + 435, + 71, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 435, + 71, + 553, + 150 + ], + "type": "image", + "image_path": "4b4f7cf3696832a2b5bf1f270d9a9f4014f3a46650eebb584120a1fc69cda84b.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 319, + 214, + 553, + 266 + ], + "blocks": [ + { + "bbox": [ + 319, + 214, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 319, + 214, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 319, + 214, + 553, + 266 + ], + "type": "image", + "image_path": "b952db4e63fc36a45322068f82a9c2568c1fd11e30fc130e4a5df44347460e2a.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 271, + 543, + 282 + ], + "lines": [ + { + "bbox": [ + 324, + 271, + 543, + 282 + ], + "spans": [ + { + "bbox": [ + 324, + 271, + 543, + 282 + ], + "type": "text", + "content": "Figure 6. The effects of scale for improving surface quality." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "text", + "content": " represents the tokens, " + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "inline_equation", + "content": "F(h)" + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "text", + "content": " their Fourier features, and " + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "inline_equation", + "content": "h'" + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "text", + "content": " the enhanced tokens, " + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\omega_0" + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "text", + "content": " denote Fourier frequencies and the threshold, with " + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\omega < \\omega_0" + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "text", + "content": " indicating low-frequency components. The scale is the enhancement coefficient, and " + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\mathrm{FFT}(\\cdot)" + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "inline_equation", + "content": "\\mathrm{IFFT}(\\cdot)" + }, + { + "bbox": [ + 313, + 288, + 554, + 396 + ], + "type": "text", + "content": " are the Fourier transform and inverse Fourier transform. As shown in Fig. 6, increasing the scale continuously enhances the surface generation capability. However, beyond a certain value, the surface does not increase further with larger scale values." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 409, + 394, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 394, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 394, + 422 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 429, + 446, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 446, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 446, + 441 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 447, + 448, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 448, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 448, + 459 + ], + "type": "text", + "content": "4.1.1. Morphing Configuration" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "type": "text", + "content": "All experiments use 250 denoising time steps. Finetuning is performed using LoRA models implemented with PEFT [38] using 500 training steps (rank = 16, alpha = 20, and list = ['to_k', 'to_q', 'to_v', 'qkv']) The initial value of " + }, + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "type": "text", + "content": " is sampled from a Beta distribution, " + }, + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "type": "inline_equation", + "content": "Beta(5,5)" + }, + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "type": "text", + "content": ", over the " + }, + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "type": "text", + "content": " interval, with 20 points selected as interpolation weights. If multiple morphing is allowed, " + }, + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "type": "text", + "content": " can be concentrated in ranges with more variation by adjusting the Beta distribution or using the reschedule strategy from DiffMorpher [80]. For Attention Fusion, we begin at step 1, with geometry diffusion ending at step 160 and texture diffusion at step 5. Token Reordering starts at step 80 and ends at step 200. Low-Frequency Enhancement occurs between steps 200 and 230, with scale = 5 and " + }, + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "type": "inline_equation", + "content": "\\omega_0 = 0.1\\pi" + }, + { + "bbox": [ + 313, + 463, + 555, + 643 + ], + "type": "text", + "content": ". Note that all our results share the same hyperparameters." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 650, + 382, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 650, + 382, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 650, + 382, + 660 + ], + "type": "text", + "content": "4.1.2. Baselines" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "(a) Task (Textured 3D Morphing): MorphFlow [64] directly implements morphing on textured 3D volumetric representations, which has the closest setting to ours; (b) Morphing Tricks: DiffMorpher [80] and AID [23] propose fusion" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15164" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 549, + 383 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 549, + 383 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 549, + 383 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 549, + 383 + ], + "type": "image", + "image_path": "b2415ebf24f9023a93bceb563a6c4a28292f4446788cfd42507882e57a9ca4d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 388, + 546, + 401 + ], + "lines": [ + { + "bbox": [ + 63, + 388, + 546, + 401 + ], + "spans": [ + { + "bbox": [ + 63, + 388, + 546, + 401 + ], + "type": "text", + "content": "Figure 7. Qualitative comparisons of methods in tasks, morphing tricks, and generative priors. More video results can be found here." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 422, + 296, + 505 + ], + "blocks": [ + { + "bbox": [ + 73, + 408, + 277, + 419 + ], + "lines": [ + { + "bbox": [ + 73, + 408, + 277, + 419 + ], + "spans": [ + { + "bbox": [ + 73, + 408, + 277, + 419 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons of different methods." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 422, + 296, + 505 + ], + "lines": [ + { + "bbox": [ + 57, + 422, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 57, + 422, + 296, + 505 + ], + "type": "table", + "html": "
Quantitative MetricsUser Study
FID↓KID↓STP-G↑SEP-G↑PPL↓PDV↓STP-U↑SEP-U↑
DiffMorpher218.070.3670.230.135.230.05350.4350.300
AID115.720.1240.670.704.680.01180.3800.505
MV-Adapter120.930.1290.630.577.290.01520.2250.350
Luma95.490.0880.830.777.370.00070.4150.330
MorphFlow147.700.1500.870.903.100.00010.5550.505
Ours6.360.0011.001.003.020.00010.9150.950
", + "image_path": "9ce5ac50ebd45b4ac82109fc648baeb2402e7537029fb4f29b81e09e153ee161.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 511, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 571 + ], + "type": "text", + "content": "strategies for attention to enhance smoothness; (c) Generative Priors: We compare the performance of using 2D diffusion (DiffMorpher, AID), multi-view diffusion (MV-Adapter [27]), video generation (Luma [2]), and 3D diffusion as prior in 3D (multi-view image) morphing task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 578, + 117, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 117, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 117, + 589 + ], + "type": "text", + "content": "4.1.3. Metrics" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "(a) FID [25] and KID [8]: They are calculated by comparing 3,000 images rendered from original and interpolated 3D representations. (b) Structural Plausibility-GPT (STP-G) and Semantic Plausibility-GPT (SEP-G): GPT-4o [42, 68] evaluates structural and semantic plausibility based on testing results, providing scores and explanations. (c) Perceptual Path Length (PPL) and Perceptual Distance Variance (PDV) [80]: PPL sums perceptual losses over 20-frame sequences, reflecting smoothness and consistency, while PDV measures the variance, indicating transition ho" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 410, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 447 + ], + "type": "text", + "content": "mogeneity. (d) Structural Plausibility-User (STP-U) and Semantic Plausibility-User (SEP-U): Volunteers score morphing results based on structural and semantic plausibility." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 453, + 389, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 389, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 389, + 464 + ], + "type": "text", + "content": "4.2. Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 471, + 441, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 441, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 441, + 483 + ], + "type": "text", + "content": "4.2.1. Textured 3D Morphing" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 700 + ], + "type": "text", + "content": "Our method adopts a 3D generation prior for 3D morphing, offering two main benefits: direct morphing of textured 3D representations, and maintaining structural and semantic consistency in interpolated representations, as shown in Fig. 1. The closest baseline to our method is MorphFlow [64], which morphs between two 3D volumetric representations using optimal transport optimization. However, as shown in Fig. 7, MorphFlow has two main drawbacks. First, the generated quality is low (see Tab. 1). The advantage of volumetric representations for photorealism is diminished due to the dense interpolation process, which requires morphing even points with no color. Second, it lacks an effective generative prior to the intermediate process, limiting its ability to understand the semantic meaning of intermediate stages. As a result, it often introduces artifacts in morphing scenarios, such as generating six legs instead of four when morphing a bear's and a stool's legs (see the \"Bear-Stool\" case in Fig. 7)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "content": "Based on the quantitative metrics and user study in the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15165" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 204 + ], + "type": "text", + "content": "Tab. 1, ours not only significantly outperforms MorphFlow in quality but also demonstrates an impressive \"understanding\" capabilities: (a) the effective mapping of semantically similar parts, as shown when morphing a boot into a red teddy bear (see the \"Boot-Teddy Bear\" case in Fig. 1), where our model smoothly splits the boot into legs and transforms them into the bear's facial features, and (b) minimal disconnected artifacts. The regenerative nature of our method ensures the fusion of source and target while considering the distribution of the entire latent space, minimizing issues like 3D structure collapse or disconnections." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 258, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 258, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 258, + 221 + ], + "type": "text", + "content": "4.2.2. 3D-Aware (Multi-View Image) Morphing" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 223, + 295, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 366 + ], + "type": "text", + "content": "We evaluate alternatives to our method for textured 3D morphing from two perspectives. First, 3D morphing can be viewed as multi-view image morphing, where multiview source and target images are fed into image morphing methods for pseudo-3D morphing. Second, many image and video generation models, beyond 3D priors, can perform interpolation tasks. Notably, 2D generation models [27, 56, 72-74], trained on larger datasets often produce 3D-consistent images with superior semantic, structural, and textural understanding compared to 3D models. Therefore, we explore three types of generative priors: 2D diffusion, multi-view diffusion, and video generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 367, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 618 + ], + "type": "text", + "content": "As shown in Fig. 7 and Tab. 1, compared to state-of-the-art 2D image morphing methods, such as DiffMorpher [80] and AID [23], 2D diffusion models often suffer from mode collapse when influenced by non-object image regions (e.g., DiffMorpher is sensitive to white backgrounds). Their lack of 3D consistency leads to inconsistent morphing results for the same " + }, + { + "bbox": [ + 55, + 367, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 367, + 295, + 618 + ], + "type": "text", + "content": " across different viewpoints. When comparing with multi-view diffusion models, we observe that their image-based multi-view generation is limited by pixel-aligned morphing. This limitation becomes apparent when matching pixels across large spatial distances—such as aligning a point on the lower edge of a pumpkin's surface with a point on the stem of a mushroom in the image's center, resulting in interpolation errors (see the \"Pumpkpin-Mushroom\" case in Fig. 7). Lastly, video generation models, while demonstrating strong spatial understanding and achieving morphing by specifying the source and target images as the first and last frames, suffer from limited controllability during generation, often producing incomplete or out-of-frame content. Additionally, the structural consistency of intermediate frames remains suboptimal." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 624, + 149, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 149, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 149, + 637 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": "We observe that texture diffusion requires minimal Attention Fusion (AF), while geometry diffusion offers a larger range, emphasizing shape understanding in 3D morphing. Thus, we incorporate Attention Fusion in texture diffusion from the first to the fifth time step, while progressively extending the final time step for geometry diffusion. As shown" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 560, + 205 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 560, + 205 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 560, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 560, + 205 + ], + "type": "image", + "image_path": "8eb74bb7f786c6684884505d273177cfbb261ea66435feb73bacba3fc182c7fa.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 215, + 539, + 226 + ], + "lines": [ + { + "bbox": [ + 327, + 215, + 539, + 226 + ], + "spans": [ + { + "bbox": [ + 327, + 215, + 539, + 226 + ], + "type": "text", + "content": "Figure 8. Ablation experiments of our proposed strategies." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 247, + 551, + 321 + ], + "blocks": [ + { + "bbox": [ + 343, + 232, + 525, + 243 + ], + "lines": [ + { + "bbox": [ + 343, + 232, + 525, + 243 + ], + "spans": [ + { + "bbox": [ + 343, + 232, + 525, + 243 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation of ablation study." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 247, + 551, + 321 + ], + "lines": [ + { + "bbox": [ + 315, + 247, + 551, + 321 + ], + "spans": [ + { + "bbox": [ + 315, + 247, + 551, + 321 + ], + "type": "table", + "html": "
Geometry Diffusion Time Step80 Steps120 Steps160 Steps
FID↓PPL↓PDV↓FID↓PPL↓PDV↓FID↓PPL↓PDV↓
Basic6.386.010.05726.386.010.05726.386.010.0572
Basic+AF9.975.650.054496.324.360.0121105.624.440.0110
Basic+AF+TR8.284.060.00709.703.300.00389.413.370.0024
Basic+AF+LE8.093.760.00497.293.240.00227.083.150.0009
Basic+AF+TR+LE6.403.540.00356.373.210.00196.363.020.0001
", + "image_path": "0ec6c892d0074bb40b78da4251dad5337163e16cd81cadff1ce9cc05a6ea9860.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 495 + ], + "type": "text", + "content": "in Fig. 8, adding Attention Fusion on top of basic interpolation improves smoothness, but extending the final time step too far causes 3D structural collapse. Token Reordering (TR) helps mitigate this issue, though pushing the time step further reduces generative quality. Ultimately, by applying the Low-Frequency Enhancement (LE) strategy, we balance the smoothness and structural plausibility and ensure all frequencies are maintained effectively. We quantitatively evaluate five different pairs under various settings, as shown in Tab. 2. The ablation study demonstrates that achieving a smooth and plausible morphing sequence requires increasing Attention Fusion steps, as well as applying Token Reordering and Low-Frequency Enhancement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 391, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 391, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 391, + 523 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 533, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 616 + ], + "type": "text", + "content": "We propose a method to achieve smooth and plausible morphing sequences across diverse cross-category 3D object pairs, incorporating Attention Fusion, Token Reordering, and Low-Frequency Enhancement. This introduces a new paradigm for textured 3D morphing, extending beyond the limitations of previous research confined to shape-only morphing on topologically aligned datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "type": "text", + "content": "Limitation and Future Work. Limited by the performance of existing open-source Text-to-3D models, complex 3D morphing remains challenging. In the future, we will enhance fidelity with more advanced 3D diffusion models [70, 81], enable few-shot motion interpolation [55] with temporal consistency, and explore 4D transitions [21]. Please check our supplementary materials for more implementation details and discussions." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15166" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 161, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This research is supported by the National Research Foundation, Singapore, under its NRF Fellowship Award (NRF-NRFF16-2024-0003). This research is also supported by MOE AcRF Tier 1 (RG97/23) and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAFICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 115, + 197 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 205, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 205, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 258 + ], + "type": "text", + "content": "[1] Ahmed Abdelreheem, Ivan Skorokhodov, Maks Ovsjanikov, and Peter Wonka. Satr: Zero-shot semantic segmentation of 3d shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15166-15179, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 260, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 296, + 293 + ], + "type": "text", + "content": "[2] Luma Labs AI. Luma dream machine: Ai-powered video content creation. https://dream-machine.lumalabs.ai/, 2025. Accessed: 2025-01-15. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 293, + 296, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 293, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 62, + 293, + 296, + 314 + ], + "type": "text", + "content": "[3] Alyaa Qusay Aloraibi. Image morphing techniques: A review. 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 316, + 296, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 296, + 358 + ], + "type": "text", + "content": "[4] Hadar Averbuch-Elor, Daniel Cohen-Or, and Johannes Kopf. Smooth image sequences for data-driven morphing. In computer graphics forum, pages 203–213. Wiley Online Library, 2016. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 360, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 296, + 392 + ], + "type": "text", + "content": "[5] Melike Aydinlar and Yusuf Sahillioğlu. Part-based data-driven 3d shape interpolation. Computer-Aided Design, 136: 103027, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 393, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 296, + 426 + ], + "type": "text", + "content": "[6] Thaddeus Beier and Shawn Neely. Feature-based image metamorphosis. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 529-536. 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 426, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 296, + 459 + ], + "type": "text", + "content": "[7] Bhumika G Bhatt. Comparative study of triangulation based and feature based image morphing. Signal & Image Processing, 2(4):235, 2011. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 460, + 296, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 492 + ], + "type": "text", + "content": "[8] Mikołaj Binkowski, Danica J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying mmd gans. arXiv preprint arXiv:1801.01401, 2018. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 493, + 296, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 296, + 546 + ], + "type": "text", + "content": "[9] Federica Bogo, Javier Romero, Matthew Loper, and Michael J Black. Faust: Dataset and evaluation for 3d mesh registration. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3794-3801, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 548, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 296, + 602 + ], + "type": "text", + "content": "[10] Zhaoxi Chen, Jiaxiang Tang, Yuhao Dong, Ziang Cao, Fangzhou Hong, Yushi Lan, Tengfei Wang, Haozhe Xie, Tong Wu, Shunsuke Saito, et al. 3dtopia-xl: Scaling high-quality 3d asset generation via primitive diffusion. arXiv preprint arXiv:2409.12957, 2024. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 603, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 296, + 647 + ], + "type": "text", + "content": "[11] Soheil Darabi, Eli Shechtman, Connelly Barnes, Dan B Goldman, and Pradeep Sen. Image melding: Combining inconsistent images using patch-based synthesis. ACM Transactions on graphics (TOG), 31(4):1-10, 2012. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "type": "text", + "content": "[12] Jiacheng Deng, Chuxin Wang, Jiahao Lu, Jianfeng He, Tianzhu Zhang, Jiyang Yu, and Zhe Zhang. Se-ornet: Self-ensembling orientation-aware network for unsupervised point cloud shape correspondence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5364–5373, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[13] Niladri Shekhar Dutt, Sanjeev Muralikrishnan, and Niloy J Mitra. Diffusion 3d features (diff3f): Decorating textured shapes with distilled semantic features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4494-4504, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 183 + ], + "type": "text", + "content": "[14] Roberto M Dyke, Yu-Kun Lai, Paul L Rosin, Stefano Zappala, Seana Dykes, Daoliang Guo, Kun Li, Riccardo Marin, Simone Melzi, and Jingyu Yang. Shrec'20: Shape correspondence with non-isometric deformations. Computers & Graphics, 92:28-43, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 184, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 554, + 217 + ], + "type": "text", + "content": "[15] Michal Edelstein, Danielle Ezuz, and Mirela Ben-Chen. Enigma: Evolutionary non-isometric geometry matching. arXiv preprint arXiv:1905.10763, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 217, + 555, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 555, + 270 + ], + "type": "text", + "content": "[16] Marvin Eisenberger, Zorah Lahner, and Daniel Cremers. Smooth shells: Multi-scale shape registration with functional maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12265-12274, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 271, + 555, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 271, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 271, + 555, + 338 + ], + "type": "text", + "content": "[17] Marvin Eisenberger, David Novotny, Gael Kerchenbaum, Patrick Labatut, Natalia Neverova, Daniel Cremers, and Andrea Vedaldi. Neuromorph: Unsupervised shape interpolation and correspondence in one go. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7473-7483, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 338, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 554, + 381 + ], + "type": "text", + "content": "[18] Noa Fish, Richard Zhang, Lilach Perry, Daniel Cohen-Or, Eli Shechtman, and Connelly Barnes. Image morphing with perceptual constraints and stn alignment. In Computer Graphics Forum, pages 303–313. Wiley Online Library, 2020. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 383, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 554, + 426 + ], + "type": "text", + "content": "[19] Lawson Fulton, Vismay Modi, David Duvenaud, David IW Levin, and Alec Jacobson. Latent-space dynamics for reduced deformable simulation. In Computer graphics forum, pages 379-391. Wiley Online Library, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 426, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 554, + 470 + ], + "type": "text", + "content": "[20] William Gao, Noam Aigerman, Thibault Groueix, Vova Kim, and Rana Hanocka. Textdeformer: Geometry manipulation using text guidance. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 471, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 491 + ], + "type": "text", + "content": "[21] Chen Geng, Yunzhi Zhang, Shangzhe Wu, and Jiajun Wu. Birth and death of a rose, 2024. 2, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 492, + 554, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 492, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 317, + 492, + 554, + 548 + ], + "type": "text", + "content": "[22] Ayaan Haque, Matthew Tancik, Alexei A Efros, Aleksander Holynski, and Angjoo Kanazawa. Instruct-nerf2nerf: Editing 3d scenes with instructions. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19740–19750, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 548, + 554, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 580 + ], + "type": "text", + "content": "[23] Qiyuan He, Jinghao Wang, Ziwei Liu, and Angela Yao. Aid: Attention interpolation of text-to-image diffusion. NeurIPS, 2024. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 581, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 581, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 581, + 554, + 635 + ], + "type": "text", + "content": "[24] Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 36, 2024. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 635, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 689 + ], + "type": "text", + "content": "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 690, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 554, + 713 + ], + "type": "text", + "content": "[26] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15167" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": "[27] Zehuan Huang, Yuanchen Guo, Haoran Wang, Ran Yi, Lizhuang Ma, Yan-Pei Cao, and Lu Sheng. Mv-adapter: Multi-view consistent image generation made easy. arXiv preprint arXiv:2412.03632, 2024. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[28] Alec Jacobson, Zhigang Deng, Ladislav Kavan, and John P Lewis. Skinning: Real-time shape deformation (full text not available). In ACM SIGGRAPH 2014 Courses, pages 1-1. 2014. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "text", + "content": "[29] Hyunwoo Kim, Itai Lang, Noam Aigerman, Thibault Groueix, Vladimir G Kim, and Rana Hanocka. Meshup: Multi-target mesh deformation via blended score distillation. arXiv preprint arXiv:2408.14899, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 259 + ], + "type": "text", + "content": "[30] Vladimir G Kim, Yaron Lipman, and Thomas Funkhouser. Blended intrinsic maps. ACM transactions on graphics (TOG), 30(4):1-12, 2011. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 282 + ], + "type": "text", + "content": "[31] Yushi Lan, Chen Change Loy, and Bo Dai. DDF: Correspondence distillation from nerf-based gan. IJCV, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 284, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 338 + ], + "type": "text", + "content": "[32] Yushi Lan, Fangzhou Hong, Shuai Yang, Shangchen Zhou, Xuyi Meng, Bo Dai, Xingang Pan, and Chen Change Loy. Ln3diff: Scalable latent neural fields diffusion for speedy 3d generation. In European Conference on Computer Vision, pages 112-130. Springer, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[33] Yushi Lan, Shangchen Zhou, Zhaoyang Lyu, Fangzhou Hong, Shuai Yang, Bo Dai, Xingang Pan, and Chen Change Loy. Gaussiananything: Interactive point cloud latent diffusion for 3d generation. In ICLR, 2025. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[34] Jing Liao, Rodolfo S Lima, Diego Nehab, Hugues Hoppe, Pedro V Sander, and Jinhui Yu. Automating image morphing using structural similarity on a halfway domain. ACM Transactions on Graphics (TOG), 33(5):1-12, 2014. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 481 + ], + "type": "text", + "content": "[35] Jianchu Lin, Yinxi Gu, Guangxiao Du, Guoqiang Qu, Xiaobing Chen, Yudong Zhang, Shangbing Gao, Zhen Liu, and Nallappan Gunasekaran. 2d/3d image morphing technology from traditional to modern: A survey. Information Fusion, page 102913, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 514 + ], + "type": "text", + "content": "[36] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 515, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 557 + ], + "type": "text", + "content": "[37] Robin Magnet and Maks Ovsjanikov. Scalable and efficient functional map computations on dense meshes. In Computer Graphics Forum, pages 89-101. Wiley Online Library, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 612 + ], + "type": "text", + "content": "[38] Sourab Mangrulkar, Sylvain Gugger, Lysandre Debut, Younes Belkada, Sayak Paul, and Benjamin Bossan. Peft: State-of-the-art parameter-efficient fine-tuning methods. https://github.com/huggingface/peft, 2022.6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "type": "text", + "content": "[39] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 713 + ], + "type": "text", + "content": "[40] Nasir Mohammad Khalid, Tianhao Xie, Eugene Belilovsky, and Tiberiu Popa. Clip-mesh: Generating textured meshes from text using pretrained image-text models. In SIGGRAPH Asia 2022 conference papers, pages 1-8, 2022. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 115 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 115 + ], + "type": "text", + "content": "[41] Luca Morreale, Noam Aigerman, Vladimir G Kim, and Niloy J Mitra. Neural semantic surface maps. In Computer Graphics Forum, page e15005. Wiley Online Library, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 118, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 139 + ], + "type": "text", + "content": "[42] OpenAI. Gpt-4: Openai's fourth-generation language model, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "text", + "content": "[43] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[44] Maks Ovsjanikov, Mirela Ben-Chen, Justin Solomon, Adrian Butscher, and Leonidas Guibas. Functional maps: a flexible representation of maps between shapes. ACM Transactions on Graphics (ToG), 31(4):1-11, 2012. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 306 + ], + "type": "text", + "content": "[45] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[46] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 388 + ], + "type": "text", + "content": "[47] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[48] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. *ICLR*, 2024. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 434, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 498 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "text", + "content": "[50] Jing Ren, Simone Melzi, Maks Ovsjanikov, and Peter Wonka. Maptree: Recovering multiple solutions in the space of maps. ACM Transactions on Graphics, 39(6):1-17, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 546, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 554, + 601 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 602, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 668 + ], + "type": "text", + "content": "[52] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[53] Dvir Samuel, Rami Ben-Ari, Nir Darshan, Haggai Maron, and Gal Chechik. Norm-guided latent space exploration for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 4" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15168" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[54] Eli Shechtman, Alex Rav-Acha, Michal Irani, and Steve Seitz. Regenerative morphing. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 615-622. IEEE, 2010. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 174 + ], + "type": "text", + "content": "[55] Liao Shen, Tianqi Liu, Huiqiang Sun, Xinyi Ye, Baopu Li, Jianming Zhang, and Zhiguo Cao. Dreammover: Leveraging the prior of diffusion models for image interpolation with large motion. arXiv preprint arXiv:2409.09605, 2, 2024. 2, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 209 + ], + "type": "text", + "content": "[56] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 211, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 295, + 255 + ], + "type": "text", + "content": "[57] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. Freeu: Free lunch in diffusion u-net. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4733-4743, 2024. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 256, + 294, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 294, + 311 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 294, + 311 + ], + "type": "text", + "content": "[58] Justin Solomon, Fernando De Goes, Gabriel Peyre, Marco Cuturi, Adrian Butscher, Andy Nguyen, Tao Du, and Leonidas Guibas. Convolutional wasserstein distances: Efficient optimal transportation on geometric domains. ACM Transactions on Graphics (ToG), 34(4):1-11, 2015. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 313, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 313, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 294, + 346 + ], + "type": "text", + "content": "[59] Olga Sorkine and Marc Alexa. As-rigid-as-possible surface modeling. In Symposium on Geometry processing, pages 109–116. CiteSeer, 2007. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 348, + 294, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 348, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 294, + 401 + ], + "type": "text", + "content": "[60] Olga Sorkine, Daniel Cohen-Or, Yaron Lipman, Marc Alexa, Christian Rössl, and H-P Seidel. Laplacian surface editing. In Proceedings of the 2004 Eurographics/ACM SIGGRAPH symposium on Geometry processing, pages 175–184, 2004. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 405, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 294, + 460 + ], + "type": "text", + "content": "[61] Mingze Sun, Chen Guo, Puhua Jiang, Shiwei Mao, Yurun Chen, and Ruqi Huang. Srif: Semantic shape registration empowered by diffusion-based image morphing and flow estimation. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 462, + 294, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 294, + 527 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 294, + 527 + ], + "type": "text", + "content": "[62] Gary KL Tam, Zhi-Quan Cheng, Yu-Kun Lai, Frank C Langbein, Yonghuai Liu, David Marshall, Ralph R Martin, XianFang Sun, and Paul L Rosin. Registration of 3d point clouds and meshes: A survey from rigid to nonrigid. IEEE transactions on visualization and computer graphics, 19(7):1199-1217, 2012. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 529, + 294, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 529, + 294, + 573 + ], + "spans": [ + { + "bbox": [ + 56, + 529, + 294, + 573 + ], + "type": "text", + "content": "[63] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent correspondence from image diffusion. Advances in Neural Information Processing Systems, 36:1363-1389, 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 575, + 294, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 575, + 294, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 575, + 294, + 619 + ], + "type": "text", + "content": "[64] Chih-Jung Tsai, Cheng Sun, and Hwann-Tzong Chen. Multiview regenerative morphing with dual flows. In European Conference on Computer Vision, pages 492-509. Springer, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 621, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 621, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 621, + 294, + 643 + ], + "type": "text", + "content": "[65] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "type": "text", + "content": "[66] Shantanu Vyas, Ting-Ju Chen, Ronak R Mohanty, Peng Jiang, and Vinayak R Krishnamurthy. Latent embedded graphs for image and shape interpolation. Computer-Aided Design, 140:103091, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[67] George Wolberg. Image morphing: a survey. The visual computer, 14(8-9):360-372, 1998. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[68] Tong Wu, Guandao Yang, Zhibing Li, Kai Zhang, Ziwei Liu, Leonidas Guibas, Dahua Lin, and Gordon Wetzstein. Gpt-4v(ision) is a human-aligned evaluator for text-to-3d generation. In CVPR, 2024. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[69] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In European Conference on Computer Vision, pages 378-394. Springer, 2025. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[70] Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. arXiv preprint arXiv:2412.01506, 2024. 2, 4, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 258 + ], + "type": "text", + "content": "[71] Jiale Xu, Weihao Cheng, Yiming Gao, Xintao Wang, Shenghua Gao, and Ying Shan. Instantmesh: Efficient 3d mesh generation from a single image with sparse-view large reconstruction models. arXiv preprint arXiv:2404.07191, 2024.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 260, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 260, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 260, + 553, + 304 + ], + "type": "text", + "content": "[72] Songlin Yang, Wei Wang, Jun Ling, Bo Peng, Xu Tan, and Jing Dong. Context-aware talking-head video editing. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7718-7727, 2023. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "type": "text", + "content": "[73] Songlin Yang, Wei Wang, Bo Peng, and Jing Dong. Designing a 3d-aware stylenerf encoder for face editing. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 360, + 555, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 555, + 414 + ], + "type": "text", + "content": "[74] Songlin Yang, Wei Wang, Yushi Lan, Xiangyu Fan, Bo Peng, Lei Yang, and Jing Dong. Learning dense correspondence for nef-based face reenactment. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6522-6530, 2024. 2, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 469 + ], + "type": "text", + "content": "[75] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[76] Mehmet Ersin Yumer, Siddhartha Chaudhuri, Jessica K Hodgins, and Levent Burak Kara. Semantic shape editing using deformation handles. ACM Transactions on Graphics (TOG), 34(4):1-12, 2015. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 568 + ], + "type": "text", + "content": "[77] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. Lion: Latent point diffusion models for 3d shape generation. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "type": "text", + "content": "[78] Xiao Zhan, Rao Fu, and Daniel Ritchie. Charactermixer: Rig-aware interpolation of 3d characters. In Computer Graphics Forum, page e15047. Wiley Online Library, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[79] Biao Zhang, Jiapeng Tang, Matthias Niessner, and Peter Wonka. 3dshape2vecset: A 3d shape representation for neural fields and generative diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-16, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "text", + "content": "[80] Kaiwen Zhang, Yifan Zhou, Xudong Xu, Bo Dai, and Xinggang Pan. Diffmorpher: Unleashing the capability of diffusion models for image morphing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7912-7921, 2024. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15169" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 262 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[81] Longwen Zhang, Ziyu Wang, Qixuan Zhang, Qiwei Qiu, Anqi Pang, Haoran Jiang, Wei Yang, Lan Xu, and Jingyi Yu. Clay: A controllable large-scale generative model for creating high-quality 3d assets. ACM Transactions on Graphics (TOG), 43(4):1-20, 2024. 3, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[82] Zhengbo Zhang, Li Xu, Duo Peng, Hossein Rahmani, and Jun Liu. Diff-tracker: text-to-image diffusion models are unsupervised trackers. In European Conference on Computer Vision, pages 319-337. Springer, 2025. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[83] Junzhe Zhu, Yuanchen Ju, Junyi Zhang, Muhan Wang, Zhecheng Yuan, Kaizhe Hu, and Huazhe Xu. Dense-matcher: Learning 3d semantic correspondence for category-level manipulation from a single demo. arXiv preprint arXiv:2412.05268, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[84] Bhushan Zope and Soniya B Zope. A survey of morphing techniques. International Journal of Advanced Engineering, Management and Science, 3(2):239773, 2017. 3" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15170" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/b1aafd05-162c-4fda-a00d-cb706731d3d7_content_list.json b/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/b1aafd05-162c-4fda-a00d-cb706731d3d7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..12f89f87539021238f0d405d7bea1a93fea3a2f9 --- /dev/null +++ b/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/b1aafd05-162c-4fda-a00d-cb706731d3d7_content_list.json @@ -0,0 +1,1440 @@ +[ + { + "type": "text", + "text": "The Best of Both Worlds: Integrating Language Models and Diffusion Models for Video Generation", + "text_level": 1, + "bbox": [ + 107, + 128, + 890, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aoxiong Yin $^{1*}$ , Xu Tan $^{2*†}$ , Kai Shen $^{2*}$ , Yichong Leng $^{2}$ , Xinyu Zhou $^{2}$ , Juncheng Li $^{1†}$ , Siliang Tang $^{1}$", + "bbox": [ + 104, + 202, + 890, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Zhejiang University, $^{2}$ Moonshot AI", + "bbox": [ + 390, + 223, + 607, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yinaoxiong@zju.edu.cn, tanxu2012@gmail.com, junchengli@zju.edu.cn", + "bbox": [ + 204, + 241, + 781, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in text-to-video (T2V) generation have been driven by two competing paradigms: autoregressive language models and diffusion models. However, each paradigm has intrinsic limitations: language models struggle with visual quality and error accumulation, while diffusion models lack semantic understanding and causal modeling. In this work, we propose LanDiff, a hybrid framework that synergizes the strengths of both paradigms through coarse-to-fine generation. Our architecture introduces three key innovations: (1) a semantic tokenizer that compresses 3D visual features into compact 1D discrete representations through efficient semantic compression, achieving $a \\sim 14,000 \\times$ compression ratio; (2) a language model that generates semantic tokens with high-level semantic relationships; (3) a streaming diffusion model that refines coarse semantics into high-fidelity videos. Experiments show that LanDiff, a 5B model, achieves a score of 85.43 on the VBenchmark T2V benchmark, surpassing the state-of-the-art open-source models Hunyuan Video (13B) and other commercial models such as Sora, Kling, and Hailuo. Furthermore, our model also achieves state-of-the-art performance in long video generation, surpassing other open-source models in this field. Our demo can be viewed at https://landiff.github.io/.", + "bbox": [ + 88, + 323, + 485, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 715, + 220, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-video (T2V) [6, 30, 55, 60] has made significant progress in recent years, becoming an important research direction in the fields of computer vision and artificial intelligence. Recent works in T2V models have primarily revolved around two predominant paradigms: autoregressive large language model (LLM)-based [30, 55] frameworks and diffusion-based architectures [6, 60]. However, each paradigm has their own advantages and limitations, as", + "bbox": [ + 89, + 742, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4de5e5549f22ffb211329d92de6ed929e5f0ad1bf570b367cd1192551fd840ef.jpg", + "image_caption": [ + "Figure 1. The rate-distortion curve illustrates how visual distortion decreases as the number of transmitted bits increases. With just a small number of bits representing high-level semantic features, we can already achieve relatively low visual distortion. Building on this information-theoretic insight, LanDiff combines the strengths of both paradigms: LLMs efficiently generate compact semantic features in the first stage, followed by diffusion models that add perceptual details in the second stage, before final decoding to pixels via VAE. Data from Ho et al. [20], illustration is conceptual." + ], + "image_footnote": [], + "bbox": [ + 517, + 291, + 903, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shown in Table 1.", + "text_level": 1, + "bbox": [ + 511, + 625, + 633, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "From a representation perspective, LLM-based methods leverage discrete tokenization to explicitly encode high-level semantics through vector quantization, effectively prioritizing conceptual abstraction and narrative coherence. However, this discretization inherently sacrifices low-level visual fidelity due to information compression, resulting in low reconstruction quality. In contrast, diffusion-based approaches employ continuous latent representations to preserve much more perceptual details, enabling superior reconstruction quality at the cost of diluted semantic interpretability, as hierarchical features remain entangled in the latent space. From a generative modeling perspective, LLM-based systems adopt autoregressive modeling to enforce causal dependencies between video frames, ensuring strong temporal coherence. However, the autoregressive generation inherently risks error propagation across time steps, where inaccuracies in early predictions amplify dur", + "bbox": [ + 509, + 643, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15604", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/11d349a931cbc4d1e87607388c671e212022c997daa23131f6bb298dfb7a988e.jpg", + "table_caption": [ + "Table 1. The comparison between LLM based and diffusion based T2V systems. The advantages and disadvantages are marked by $\\text{喜}$ and $\\text{喜}$ , respectively." + ], + "table_footnote": [], + "table_body": "
MethodsRepresentationsModeling
LLMDiscrete TokensAutoregressvie
Low Reconstruction QualityNo Refinement
Highlight Semantic InformationCausal Modeling
DiffusionContinuous VectorsNon-Autoregressvie
High Reconstruction QualityProgressive Refinement
Lack Semantic InformationNon-causal Modeling
", + "bbox": [ + 94, + 141, + 478, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing decoding. In contrast, diffusion-based methods employ non-autoregressive generation, refining outputs in parallel through iterative denoising steps. Although this design mitigates error accumulation and enhances generation flexibility, the absence of explicit causal constraints often leads to temporal inconsistencies or semantic hallucinations.", + "bbox": [ + 89, + 258, + 483, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a hybrid architecture that synergizes the strengths of both Language models and Diffusion models, named LanDiff, through a coarse-to-fine generation paradigm. As shown in Figure 1, inspired by the human creation of video which will generate the high-level storyline first and then add low-level visual details based on the storyline to form the video, we design a two-stage video generation process with the number of bits gradually increasing and carefully design the autoregressive model and the diffusion model to be responsible for different stages of T2V generation, so as to play to their strengths and avoid their weaknesses. In detail, 1) at the low-bit position \"semantic feature\", the low-bit information ensures that the token sequence is not too long, and the high-level information makes it easier for the model to capture the overall semantic entity motion of the video, so as to play to the strengths of the autoregressive model and avoid its weaknesses. Thus we use LLM to generate a coarse-grained video in the first stage; 2) at the high-bit position \"perceptual feature\", since we have already obtained the coarse-grained with rich semantic and time-serial information, we only need to focus on how to add details to the coarse-grained video. Thus we apply a diffusion model in the second stage. Finally, a VAE decoder transforms the generated \"perceptual feature\" into the final RGB video output. By unifying these complementary mechanisms, we demonstrate that hybrid architectures can overcome the inherent limitations of isolated approaches, enabling coherent, semantically faithful, and visually compelling video generation from textual descriptions, as shown in Table 2.", + "bbox": [ + 91, + 353, + 483, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With this design, the ideal semantic feature should contain high-level semantic information and motion information and only require a few bits to represent. We achieve this goal by performing extreme compression on video representations rich in high-level semantics. For video representation, we select the Theia model [46] as our visual rep", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/73f574e84f4cdee02ca331b688d84801cb835181e928cc8c84d91663c6759158.jpg", + "table_caption": [ + "Table 2. The comparison between LanDiff and previous largescale T2V systems. The compression rates of LLM-based and diffusion-based models are illustrated using VideoPoet [30] and CogVideoX [60] as examples, respectively." + ], + "table_footnote": [], + "table_body": "
ModelsLLMDiffusionLanDiff
Highlight Semantic Information?✓✗✓
Progressive Refinement?✗✓✓
Causal Modeling?✓✗✓
High Visual Quality?✗✓✓
Compression Ratio ↑~256~1024~14000
Long video Generation?✗✗✓
", + "bbox": [ + 517, + 155, + 898, + 266 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "resentation backbone, which has been distilled from multiple visual understanding and self-supervised representation models, ensuring the encoded features contain rich semantic information. To achieve extreme compression and reduce the number of bits, we design an efficient tokenizer to compress 3D visual features into 1D discrete representations. The tokenizer is based on the Transformer [54] structure, uses query embedding to aggregate visual features, and has a higher compression rate than CNN-based structures [63]. To further compress the video by fully utilizing the temporal redundancy of the video, inspired by the MP4 video encoding algorithm [32], we divide the video into keyframes and non-keyframes, and set more numbers of tokens for keyframes. The detailed design is shown in subsection 3.1. For the diffusion model, we use generated semantic tokens as conditions and generate the target video by gradually removing the noises. To better support the long-video generation, we train a chunk-wise streaming diffusion model that only uses a limited number of historical frames as conditions, thereby greatly reducing the computational cost of training and inference.", + "bbox": [ + 511, + 271, + 906, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thanks to these designs, our LanDiff has made significant progress in spatial relationship compliance, action coherence, visual quality, etc. Specifically, 1) for short video generation, our 5B model achieved a score of 85.43 on the VBenchmark, surpassing the state-of-the-art open-source models Hunyuan Video (13B) and other commercial models such as Sora, Kling, and Hailuo; 2) for long video generation, after testing by the VBenchmark T2V benchmark, our model also achieved state-of-the-art performance, surpassing other open-source models in this field. Our video examples can be viewed at https://landiff.github.io/.", + "bbox": [ + 511, + 589, + 908, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 784, + 658, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Tokenization. Video tokenization plays a crucial role in video understanding and generation tasks. Since a video can be represented as a sequence of continuous frames, some works directly use image tokenizers to process videos frame by frame. For example, Wang et al. [55] directly use SBER-MoVQGAN as the video tokenizer.", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15605", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, this method ignores the temporal redundancy in videos, resulting in a low compression rate. To reduce temporal redundancy, some works [58, 62] try to extend image tokenizers based on 2D convolution to 3D convolution, which can process both temporal and spatial information simultaneously. These methods encode videos in the original RGB space and perform video reconstruction tasks, which are more about perceptual compression. In addition, some works [15, 25] try to train video tokenizers on features extracted from pre-trained visual encoders. These tokenizers can achieve good performance in understanding and generation tasks while maintaining a high compression rate. The video tokenizer we propose belongs to this category. Unlike previous feature-based tokenizers that achieved limited compression rates and primarily focused on image processing, our video tokenizer delivers significantly higher compression while handling both images and videos in a unified framework.", + "bbox": [ + 89, + 90, + 480, + 361 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLM based Video Generation. LLM-based video generation methods are usually based on the Transformer [54] structure, learning the mapping from text to video through next token prediction. TATS [14] uses VQ-GAN as the tokenizer and predicts video tokens using a GPT-like model structure. VideoGPT [58] uses 3D convolution to extract features and quantize them, and predicts the quantized video discrete tokens using a GPT-like model. Recently, VideoPoet [62] uses MagViT2 [62] as the tokenizer and unifies multiple modalities as input to a large language model (LLM) to conditionally generate video tokens. In addition, Emu3 [55] uses SBER-MoVQGAN as the tokenizer to perform video understanding and generation by predicting the next token. These works all use LLMs to directly generate \"perceptual features\" that contain rich visual details with high bit rates. Recently, ARLON [33] attempts to discretize VAE features into a small number of tokens to reduce the bit rate required for LLMs' prediction. In this way, the tokens retain low-frequency visual information such as blurry contours rather than high-level semantic information. In contrast, our method employs tokens containing high-level semantic information as prediction targets for LLMs, which enables us to fully leverage LLMs' advantages in causal modeling to precisely generate high-quality videos.", + "bbox": [ + 91, + 368, + 482, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion based Video Generation. Diffusion-based methods have achieved great success in image generation, and recently many people [7, 21, 22, 49, 69] have tried to apply them to video generation tasks. VDM [22] extends the 3D U-Net structure for video generation. Wang et al. [56] propose to generate high-quality and aesthetically pleasing videos in a cascaded manner. Benefiting from the success of the text-to-image (T2I) field, some works such as Animatediff [16], SVD [6], and PixelDance [65] try to use pre-trained T2I models as initialization, and then add modules for temporal modeling to capture motion informa", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion for video generation. Ma et al. [39] explore the generation capabilities of multiple different structures of latent diffusion transformer. After the release of SORA, a series of video generation methods based on the DiT [41] model have been proposed, including OpenSora [68], OpenSoraPlan [34], Cogvideox [60], Hunyuan Video [31], Mira [27] and STIV [35] etc. These methods can only generate short videos of a few seconds. Recently, StreamingT2V [18] generates long videos by block-wise generation on a pre-trained short video generation model, and then uniformly performs mixed augmentation. In addition, some works improve the consistency of long video generation by leveraging noise rescheduling techniques [37, 43]. Our method employs diffusion models as renderers for semantic features, enabling us to leverage their superior visual generation quality while circumventing their limitations in causal modeling.", + "bbox": [ + 511, + 90, + 903, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 345, + 604, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose a novel text-to-video generation framework that synergistically integrates the strengths of autoregressive modeling and diffusion processes while circumventing their respective limitations. The framework mainly consists of the following components: 1) an efficient tokenizer that transforms 3D visual features into compact 1D discrete representations while preserving and enhancing their semantic information; 2) a language model that performs temporal sequence modeling to generate semantic tokens representing video blueprints from textual descriptions; 3) a streaming diffusion model that progressively refines coarse semantic videos by adding fine-grained details to produce high-quality VAE features; and 4) a VAE decoder that reconstructs the final video frames from the refined VAE features.", + "bbox": [ + 511, + 371, + 903, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Video Semantic Tokenizer", + "text_level": 1, + "bbox": [ + 511, + 606, + 750, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce a novel video semantic tokenizer that efficiently compresses the video into semantic information. Firstly, we will introduce the video semantic representation used in tokenization. Then, considering the high redundancy of video in both spatial and temporal dimensions, we introduce two compression strategies: 1) a query-based causal tokenization that efficiently reduces spatial redundancy while preserving essential semantic information through vector quantization; 2) inspired by MP4 [32], we implement video frame grouping to minimize temporal redundancy by treating grouped frames as a unit, where the first frame (I-frame) is fully encoded while subsequent frames (P-frames) only capture the temporal changes by referencing content from previous frames.", + "bbox": [ + 511, + 628, + 903, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Semantic Representation. Generally, video representations can be divided into two categories: 1) some methods [55, 62] directly utilize an autoencoder to learn the video representations. 2) Some works [25, 26, 29, 51] use", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15606", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bca7e08238517ff87e8b2fcc62116f2dd4eff394178801b9aed0b7c24836e234.jpg", + "image_caption": [ + "Figure 2. The architecture of LanDiff. Given text inputs, we first extract text embeddings and employ an LLM to generate semantic tokens in the first stage. Subsequently, we utilize a diffusion model to synthesize perceptual features conditioned on these semantic tokens, followed by a VAE decoder that transforms these features into the final video frames." + ], + "image_footnote": [], + "bbox": [ + 174, + 88, + 823, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pre-trained visual self-supervised learning features (SSL) as video representations. Compared with the first directly learned autoencoder latents which contain lots of visual details, the second SSL features maintain much more semantic features, which enable the LLM to focus more on the high-level semantic information of the video.", + "bbox": [ + 88, + 392, + 480, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on these thoughts, we choose the pretrained SSL features as the video representations. For better usage of different SSL models, we choose Theia model [46], which is a unified visual feature extractor distilled from multiple visual task models, including CLIP [44], SAM [28], DINOv2 [40], ViT [10], and Depth-Anything [59].", + "bbox": [ + 89, + 483, + 483, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tokenizer Design. In this part, we introduce the design of our tokenizer, which leverages the query tokens to compress the video semantic features and uses quantization to discrete the video semantic representation while minimizing the reconstruction loss.", + "bbox": [ + 89, + 574, + 483, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In detail, firstly, we extract the semantic features using the Theia model and flatten it to obtain $F \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}$ . Inspired by TiTok [63], we use N randomly initialized tokens as query tokens $Q \\in \\mathbb{R}^{N \\times D}$ and concatenate them with the semantic features $F$ . Then we use a transformer encoder to encode them and only take the encoded features of the query tokens.", + "bbox": [ + 89, + 648, + 483, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {Q} = \\operatorname {E n c} ([ F; Q ]), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 765, + 482, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[\\cdot ]$ represents the concatenation operation and $Z_{Q}\\in$ $\\mathbb{R}^{N\\times D}$ . We then apply vector quantization on $Z_{Q}$ by train a VQ-VAE model and obtain the quantized feature $\\hat{Z}_Q$ . In the decoding stage, the quantized feature $\\hat{Z}_Q$ is used as a condition, and then a sequence of mask tokens $M\\in \\mathbb{R}^{T\\times H\\times W}$ are added in front of the $\\hat{Z}_Q$ to form the inputs of decoder. Then we only take the features of the mask tokens as fol", + "bbox": [ + 89, + 790, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lows:", + "bbox": [ + 513, + 392, + 552, + 405 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {F} = \\operatorname {D e c} ([ M; \\hat {Z} _ {Q} ]), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 405, + 903, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{F} \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}$ represents the reconstructed feature, and $M$ represents the mask tokens. Inspired by Huang et al. [23], Wang et al. [57], we minimize the reconstruction loss of the video semantic feature during training the VQ-VAE.", + "bbox": [ + 511, + 428, + 903, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the VQ-VAE, we follow the method of Wang et al. [57], Yu et al. [61]. We update the codebook using exponential moving average (EMA). We also optimize the codebook with the video semantic feature reconstruction loss. The loss function is shown as follows:", + "bbox": [ + 511, + 505, + 903, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {r e c} \\| \\hat {F} - F \\| _ {2} + \\lambda_ {c o m m i t} \\| \\operatorname {s g} (\\hat {Z} _ {Q}) - Z _ {Q} \\| _ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 587, + 903, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{sg}()$ is the stop-gradient operation, $||\\cdot ||_2$ is the L2 loss.", + "bbox": [ + 511, + 613, + 903, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Frame Grouping. For video, a straightforward observation is the redundancy in time series (i.e., the difference between adjacent video frames is minimal). Intuitively, we can achieve a better compression rate by modeling the difference between adjacent video frames instead of tokenizing all frames equally. Inspired by the popular video compression method MP4 [32], given $N$ frames of a video, we first split them into $N / T$ groups (i.e., $T$ frames as a group. For clarification, as shown in Figure 3, we use $T = 3$ as an example.). Then: 1) we will model different groups independently; 2) for each group, we will fully encode the first frame (Intra-coded Frame, IFrame), while for the remaining $[1,T - 1]$ frames (the Predictive-coded Frame, PFrame), we encode them by referencing their previous frames.", + "bbox": [ + 511, + 643, + 906, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve this, within a group, the first frame (IFrame) will only see itself and have a large number of query tokens", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15607", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0255745df1d25767370eb3b425067e43c428c0dffbef7d7588a2b1b3cab18b30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 90, + 496, + 375 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/db8d53fd9c435c1e0f79d208bb4eb8b78c66c47877ce383b61fc231d9483919c.jpg", + "image_caption": [ + "Figure 3. Proposed architecture of the video semantic tokenizer. We use query tokens to compress the semantic sequence length. Furthermore, we group the frames into groups (3 frames in a group in this figure). In a group, the first frame is the IFrame and the rest frames are PFrames. We use different query token numbers for them. The attention mask design is shown in the right." + ], + "image_footnote": [], + "bbox": [ + 500, + 90, + 781, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(e.g., 3 query tokens in Figure 3) to achieve better reconstruction quality. For the rest frame (PFrame) $i \\in [1, T - 1]$ , it will see the previous frames (i.e., frames $j \\in [0, i - 1]$ ) and have a small number of query tokens (e.g., 1 query token which is $1/3$ of the IFrame) to force the model to learn the difference. Technically, as shown in the attention mask on the right side of Figure 3, we use frame-level causal masks for the feature sequence during encoding. The query tokens are also divided according to the frame, and each token can only attend to the features of the corresponding frame and the previous frames. During decoding, the mask token corresponding to each frame can see the previous features, the corresponding query tokens, and the previous query tokens.", + "bbox": [ + 88, + 426, + 485, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Language Model for Semantic Token Generation", + "text_level": 1, + "bbox": [ + 89, + 635, + 483, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 2, after training an efficient tokenizer, we use a language model to generate semantic tokens autoregressively based on text. Specifically, we first use a pretrained text encoder T5-XXL [45] to extract text features $X$ . We use the video tokenizer encoder trained in subsection 3.1 to convert the video into a 1D discrete token sequence $Y$ . To enhance the controllability of the generated results, we introduce additional control conditions $CC$ . It includes: 1) frames condition for requiring the model to generate videos with a specified number of frames; 2) motion score condition, which is a value between 0 and 1, used to control the degree of motion in the generated video. The discrete tokens are converted into embedding vectors during generation and added with positional encoding, and then concatenated with the control conditions as input. The structure of", + "bbox": [ + 89, + 674, + 485, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the language model follows the typical LLaMA model [53] network structure. We train the model from scratch, using cross-entropy loss as the loss function.", + "bbox": [ + 511, + 426, + 906, + 472 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L M}} = \\mathbb {E} \\left[ - \\log p \\left(Y _ {i} \\mid X, C C, Y _ {< i}\\right) \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 488, + 906, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Diffusion Model for Perceptual Feature Generation", + "text_level": 1, + "bbox": [ + 511, + 513, + 906, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The video detokenizer in LanDiff is responsible for converting semantic tokens into VAE latent vectors. As shown in Figure 2 we use a conditional diffusion model to complete this task. Especially to support long-video generation, we also design the streaming inference strategy. In this section, we will first introduce the architecture for a single semantic token chunk. Then we will introduce the chunk-wise streaming strategy.", + "bbox": [ + 511, + 551, + 908, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture. Our diffusion model employs an architecture similar to MMDiT[12, 60]. Specifically, 1) we use the video tokenizer decoder trained in subsection 3.1 to decode the semantic tokens into semantic features $\\hat{F}$ . Then we use the semantic features $\\hat{F}$ as a condition to guide the diffusion model to generate videos; 2) we inject the control signals into the model in a similar way to ControlNet [67]. In detail, during training, the parameters of the main model are not updated, the control module copies the parameters of the first half layers of the main model, and adds them to the output of the main model after a linear layer initialized with zeros; 3) to make the semantic features match the target VAE features in the space dimension, we additionally add an upsampling module. We show the model structure in the appendix Figure 8.", + "bbox": [ + 511, + 674, + 910, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15608", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, given the input video chunk $V \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}$ , where $T$ is the frame length of the VAE latent and $D$ is the VAE latent feature dimension, the diffusion algorithm progressively adds noise to the video and produces a noisy video $V^t$ , where $t$ represents the number of times noise is added. The diffusion algorithm uses time step $t$ , and semantic features $\\hat{F} \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}$ as conditions, and then uses a network $\\epsilon_{\\theta}$ to predict the noise added to the noisy video $V_t$ through:", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {V, t, c _ {s}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(V ^ {t}, t, \\hat {F}\\right) \\| _ {2} ^ {2} \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 232, + 482, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Chunk-wise Streaming Strategy. To support long video generation scenario whose semantic token sequence is very long and difficult to generate as a whole, we propose the chunk-wise streaming strategy. During training, given video latent chunk $V \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}$ , to maintain the appearance continuity in the video, we use the first half of $V(V_{l})$ as the prompt and generate the second half of $V_{r}$ . In detail, we do not add noise to the first half chunk and give time condition $t$ to 0.999. We also randomly mask the first half chunk with a ratio of $20\\%$ . The loss function for training at this time is:", + "bbox": [ + 89, + 261, + 483, + 425 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {V _ {r}, t, c _ {s}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(V _ {l}, V _ {r} ^ {t}, t, c _ {s}\\right) \\| _ {2} ^ {2} \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 429, + 482, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During inference, the first $L / 2$ VAE latents will be generated without the prompt. For the following VAE latents, we will accumulate $L / 2$ tokens to form a chunk and use the previous chunk as prompt.", + "bbox": [ + 89, + 459, + 483, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments and Results", + "text_level": 1, + "bbox": [ + 89, + 532, + 323, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 556, + 297, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For the video tokenizer and language model, we use an internal dataset with 200M video-text pairs for training. All videos with a duration of less than 6s are filtered, and the videos are kept in the aspect ratio and then scaled to around 480x720 resolution for center cropping. Consistent with CogVideoX [60], we set the fps of all videos to 8. For the diffusion model, we select a dataset containing 3M high-quality video-text pairs for training. To evaluate the performance of the text-to-video generation model, we use the prompts provided by the widely used VBench [24] T2V benchmark to generate videos.", + "bbox": [ + 89, + 578, + 482, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. For the video tokenizer, we set the group size $T$ mentioned in subsection 3.1 to 13. We set the number of tokens $I$ corresponding to the IFrame to 330, and the number of tokens $P$ corresponding to the PFrame to 74. The dimension of the tokenizer's codebook is set to 16, and the vocabulary size is set to 2048. For the language model, we use the LLaMA [53] structure, with 2B model parameters, and use 1D RoPE [50] positional encoding. We use the 2B version of the CogVideoX [60] model as the", + "bbox": [ + 89, + 744, + 483, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "base model for our video detokenizer. We list the structure settings and more implementation details of each module in the appendix subsection A.1, subsection A.2, and subsection A.3.", + "bbox": [ + 511, + 90, + 905, + 150 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate the text-to-video generation task, we use the metrics proposed in the VBench and VBench-Long [24] benchmarks.", + "bbox": [ + 511, + 152, + 905, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Baselines. We compare LanDiff with baselines: 1) Sora [5]. 2) Jimeng [3]. 3) Hailuo [2]. 4) OpenSoraPlan V1.1 [34]. 5) Kling [38]. 6) InstructVideo [64]. 7) Gen-3 [1]. 9) Latte-1 [39] 9) HiGen [42]. 10)AnimateDiffV2 [16]. 11) Show-1 [66]. 12) Pika [4]. 13) VideoCrafter2.0 [7]. 14) OpenSora V1.2 [68]. 15) LTX-Video [17]. 16) Mochi-1 [52]. 17) CogVideoX [60]. 18) Vchitect-2.0 [13]. 19) RepVideo [48]. 20) HunyuanVideo [31]. 21) ARLON [33]. 22) Emu3 [55].", + "bbox": [ + 511, + 198, + 905, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 513, + 345, + 715, + 361 + ], + "page_idx": 5 + }, + { + "type": "ref_text", + "text": "A life-sized ice sculpture of a playful dog, with intricate details and a joyful expression, stands in the middle of a sunlit, grassy field on a sweltering summer day. The ice dog, initially solid and vibrant, begins to melt under the relentless heat, with droplets of water forming on its surface. As the day progresses, the ice dog's form gradually diminishes, with its once sharp features becoming blurred and distorted. The melting process accelerates, and the ice dog's body starts to collapse, pooling into a puddle of water on the ground. By the end of the day, all that remains is a shallow puddle, reflecting the cloudless sky, with the memory of the once majestic ice dog now just a memory.", + "bbox": [ + 519, + 382, + 898, + 462 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df4b8ae44b75e13b80b941287681438324475d9ad202ba0dd8c7b1edd40b4ced.jpg", + "image_caption": [ + "Figure 4. Comparison of qualitative results for text-to-video generation." + ], + "image_footnote": [], + "bbox": [ + 521, + 459, + 629, + 547 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e8f2da8dfaeb6db258ae37e51ff71724171ceae3be3a1e4be6921efaf6da66e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 460, + 717, + 550 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d1b7e3b8e30e606bc38e7fc2b50a06b74b610c43d9cd413e1d6fc2a6b65018c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 460, + 805, + 550 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/35b80902c0c1627075823259dbcf504b84b22689e9f62362fc80fa09b7946671.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 808, + 460, + 893, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Text to Video Generation. As shown in Table 3, our 5B LanDiff model outperforms the state-of-the-art open-source models on VBenchmark, including the 13B HunyuanVideo, achieving a top score of 85.42. LanDiff exhibits superior semantic understanding while maintaining excellent generation quality, surpassing same-sized CogVideoX-5B across most metrics. To validate our approach's effectiveness independent of training data volume, we trained a 7B DiT-based diffusion model with identical training data (marked as DiT in the table). The results confirm that our task division strategy successfully leverages the complementary strengths of autoregressive and diffusion models. Figure 4 shows qualitative comparisons between LanDiff and CogVideoX-5B. Our model effectively captures temporal dynamics described in the prompt, accurately rendering a melting ice dog sculpture. LanDiff successfully depicts the progressive transformation from solid form to liquid state, while CogVideoX-5B produces a static ice sculpture that remains virtually unchanged throughout the sequence. This demonstrates our model's superior ability to compre", + "bbox": [ + 511, + 598, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://huggingface.co/THUDM/CogVideoX-2b", + "bbox": [ + 107, + 886, + 441, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15609", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/930cecbb8f0e7fa1fe94aae210f06c336fc013eddae364f837d729f7bae6c831.jpg", + "table_caption": [ + "Table 3. Performance comparison of Text-to-video (T2V) generation between our LanDiff and other state-of-the-art models on VBench benchmark. We selected 8 out of the 16 evaluation dimensions from VBench, along with Total Score, Quality Score, and Semantic Score, for presentation. The complete results with all 16 evaluation dimensions can be found in the appendix Table 8. The best and second-best scores are highlighted in bold and underline, respectively. $\\dagger$ indicates the scores we reproduced, while $\\ddagger$ indicates the scores from the original papers, and other scores are from the VBench benchmark." + ], + "table_footnote": [], + "table_body": "
Model NameTypeModel SizeTotal ScoreQuality ScoreSemantic Scorebackground consistencydynamic degreemotion smoothnessmultiple objectsobject classscenespatial relationshipsubject consistency
Open Sourced Models
InstructVideoDiffusion1.3B76.6181.5656.8196.9769.7296.6221.5773.2622.2143.4995.30
Latte-1Diffusion0.7B77.2979.7267.5895.4068.8994.6334.5386.5336.2641.5388.88
OpenSoraPlan V1.1Diffusion2.7B78.0080.9166.3896.7347.7298.2840.3576.3027.1753.1195.73
Show-1Diffusion6.3B78.9380.4272.9898.0244.4498.2445.4793.0747.0353.5095.53
OpenSora V1.2Diffusion1.1B79.7681.3573.3997.6142.3998.5051.8382.2242.4468.5696.75
LTX-VideoDiffusion1.9B80.0082.3070.7997.2054.3598.9645.4383.4551.0765.4396.56
Mochi-1Diffusion10B80.1382.6470.0897.2861.8599.0250.4786.5136.9969.2496.99
AnimateDiff-V2Diffusion1.3B80.2782.9069.7597.6840.8397.7636.8890.9050.1934.6095.30
VideoCrafter-2.0Diffusion1.7B80.4482.2073.4298.2242.5097.7340.6692.5555.2935.8696.85
CogVideoX-2BDiffusion2B80.9182.1875.8396.6359.8697.7362.6383.3751.1469.9096.78
Emu3 ‡LLM8B80.96N/AN/A97.6979.2798.9344.6486.1737.1168.7395.32
Vchitect-2.0-2BDiffusion2B81.5782.5177.7996.5358.3397.7669.3587.8157.5154.6496.42
CogVideoX-5BDiffusion5B81.6182.7577.0496.5270.9796.9262.1185.2353.2066.3596.23
DiT †Diffusion7B81.8582.7078.4297.6551.0298.9469.6393.4057.1060.8997.05
RepVideoDiffusion2B81.9482.7078.9196.5657.7898.1371.1887.8352.9674.7496.25
Vchitect-2.0[E]Diffusion2B82.2483.5477.0696.6663.8998.9868.8486.6156.5757.5596.83
HunyuanVideoDiffusion13B83.2485.0975.8297.7670.8398.9968.5586.1053.8868.6897.37
Close Sourced Models
Pika-1.0DiffusionN/A80.6982.9271.7797.3647.5099.5043.0888.7249.8361.0396.94
KlingDiffusionN/A81.8583.3975.6897.6046.9499.4068.0587.2450.8673.0398.33
JimengDiffusionN/A81.9783.2976.6998.3938.4398.0969.0889.6244.9477.4597.25
Gen-3DiffusionN/A82.3284.1175.1796.6260.1499.2353.6487.8154.5765.0997.10
HailuoDiffusionN/A83.4184.8577.6597.0564.9199.2276.0487.8350.6875.5097.51
SoraDiffusionN/A84.2885.5179.3596.3579.9198.7470.8593.9356.9574.2996.23
ARLON ‡LLM+Diffusion1.5BN/AN/AN/A97.1052.7798.92N/A89.8054.43N/A93.41
ARLON †LLM+Diffusion5B82.3183.5877.2798.6572.2297.5674.4990.6052.0762.5395.59
LanDiffLLM+Diffusion5B85.4386.1382.6198.7392.7197.0886.6994.9453.7973.7496.11
", + "bbox": [ + 133, + 160, + 861, + 534 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "hend and visualize complex temporal transformations and physical processes. For more video examples, please refer to our demo website2.", + "bbox": [ + 88, + 553, + 482, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Long Video Generation. We compare our model with other open-source text-to-long video generation models: FreeNoise [43], StreamingT2V [18], OpenSora-V1.2 [68], and ARLON [33]. As shown in Table 4, LanDiff achieves state-of-the-art performance with the highest Total Score of 68.34. While StreamingT2V shows superior Dynamic Degree (85.64), it sacrifices consistency metrics (Subject: 87.31, Background: 94.64). Conversely, models like ARLON excel in consistency but exhibit limited dynamism (50.42). Our approach strikes an optimal balance with strong dynamism (72.86) while maintaining high consistency scores and achieving the best results in Aesthetic Quality (60.96) and Overall Consistency (27.29). As illustrated in Figure 5, LanDiff effectively renders complex motion dynamics in the prompt \"a car accelerating to gain speed,\" successfully depicting increasing velocity while preserving visual consistency. In contrast, competing models struggle with this motion, producing either static", + "bbox": [ + 91, + 599, + 482, + 872 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vehicles, inconsistent appearances, or unrealistic acceleration patterns. Additional video samples are available on our demo site2.", + "bbox": [ + 511, + 553, + 906, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Tokenization. We present visualization results of our model's video reconstruction in Figure 6. We extract semantic tokens from the reference video and subsequently convert these tokens back into video using the video detokenizer. The results demonstrate that through our careful design, our video tokenizer can accurately reconstruct videos with accurate semantics and actions using only approximately 1/50th of the sequence length required by MagViT2 [62]. While some minor discrepancies in clothing details exist between the reconstructed and reference videos, these differences remain within acceptable limits for practical video generation applications. Additional video samples are available on our demo site2.", + "bbox": [ + 511, + 599, + 908, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 801, + 666, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conducted ablation experiments on the video tokenizer and classifier-free guidance. To fairly evaluate the effectiveness of our video tokenizer, we implemented an ARLON-like method as a comparison baseline. Specifically, for the control group without our proposed", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2https://landiff.github.io/", + "bbox": [ + 107, + 886, + 321, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15610", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Long video generation results of LanDiff and other models on VBench. The best and second-best scores are highlighted in bold and underline, respectively.", + "bbox": [ + 89, + 88, + 906, + 117 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e29716b0accdd512fa71a3676baa79660ee3bf6c05253ac7b973f9077370eed.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of long video generation results between LanDiff and other state-of-the-art models (FreeNoise, StreamingT2V and OpenSora-V1.2)." + ], + "image_footnote": [], + "bbox": [ + 173, + 119, + 818, + 492 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/24fe78fd9647fa699f7cde4132ec1be689b4e123c0d7cef6543f468d325716a6.jpg", + "image_caption": [ + "Figure 6. Visualization results of video reconstruction using video tokenizer." + ], + "image_footnote": [], + "bbox": [ + 94, + 530, + 472, + 664 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2775efe50e5bc5b8ffab6a2a1fbf01f0abb0e1f61ae71840c9757fafa37775b8.jpg", + "table_caption": [ + "Table 5. The ablation study of video tokenizer and classifier free guidance (cfg) on Vbench benchmark." + ], + "table_footnote": [], + "table_body": "
ModelsTotal ScoreQuality ScoreSemantic Score
LanDiff85.4386.1382.61
-video tokenizer82.3183.5877.27
-cfg81.0683.0473.13
", + "bbox": [ + 135, + 724, + 434, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "video tokenizer, we followed ARLON's architectural approach by training a large language model to predict quantized VAE features, then using CogVideoX-5B as the base diffusion model conditioned on these LLM-predicted VAE", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "features to generate videos. This baseline model was trained with exactly the same dataset and training recipe as Lan-Diff, ensuring strict experimental control variables and effectively eliminating interference factors such as model parameter count and training data scale. As shown in Table 5, the experimental results clearly demonstrate that our proposed video tokenizer significantly improves both the quality score and semantic score of generated videos, validating the importance of this component in our framework. In the case of not using classifier-free guidance for text, our model has decreased in quality and semantic scores.", + "bbox": [ + 511, + 531, + 906, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 709, + 633, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a new text-to-video generation model, LanDiff. It combines the advantages of autoregressive models and diffusion models, including: 1) an efficient 1D video tokenizer to extract videos into semantic tokens; 2) a language model to generate semantic tokens based on text; 3) a video detokenizer to convert semantic tokens into videos. LanDiff outperforms the state-of-the-art open-source models on the VBenchmark, surpassing other open-source models in quality and semantic scores. At the same time, LanDiff also achieves state-of-the-art performance in long video generation tasks.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15611", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the NSFC (62272411), the Key R&D Projects in Zhejiang Province (No. 2024C01106, 2025C01030), the Zhejiang NSF (LRG25F020001), Zhejiang University Education Foundation Qizhen Scholar Foundation. This work was also supported by the Moonshot AI.", + "bbox": [ + 89, + 113, + 483, + 198 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 226 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Gen-3. https://runwayml.com/research/introducing-gen-3-alpha, 2024.6", + "[2] Hailuo. https://platform.minimaxi.com/, 2024.6", + "[3] Jimeng. https://jimeng.jianying.com/, 2024.6", + "[4] Pika 1.0. https://www.pika.art/, 2024.6", + "[5] Openai sora. https://openai.com/index/sora, 2024.6", + "[6] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets, 2023. 1, 3", + "[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 7310-7320. IEEE, 2024. 3, 6", + "[8] Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High Fidelity Neural Audio Compression, 2022. 14", + "[9] Juechu Dong, Boyuan Feng, Driss Guessous, Yanbo Liang, and Horace He. Flex Attention: A Programming Model for Generating Optimized Attention Kernels, 2024. 14", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 4", + "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming Transformers for High-Resolution Image Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873–12883, 2021. 15", + "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. Open-Review.net, 2024. 5", + "[13] Weichen Fan, Chenyang Si, Junhao Song, Zhenyu Yang, Yinan He, Long Zhuo, Ziqi Huang, Ziyue Dong, Jingwen He," + ], + "bbox": [ + 93, + 236, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dongwei Pan, Yi Wang, Yuming Jiang, Yaohui Wang, Peng Gao, Xinyuan Chen, Hengjie Li, Dahua Lin, Yu Qiao, and Ziwei Liu. Vchitect-2.0: Parallel Transformer for Scaling Up Video Diffusion Models, 2025. 6", + "[14] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long Video Generation with Time-Agnostic VQGAN and Time-Sensitive Transformer, 2022. 3", + "[15] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a SEED of Vision in Large Language Model, 2023. 3", + "[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3, 6", + "[17] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. LTX-Video: Realtime Video Latent Diffusion, 2024. 6", + "[18] Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text, 2024. 3, 7", + "[19] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance, 2022. 15", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 1", + "[21] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. CoRR, abs/2210.02303, 2022. 3", + "[22] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models, 2022. 3", + "[23] Zhichao Huang, Chutong Meng, and Tom Ko. Repocodec: A speech representation codec for speech tokenization. arXiv preprint arXiv:2309.00169, 2023. 4", + "[24] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21807-21818. IEEE, 2024. 6", + "[25] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jian-choao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu," + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15612", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chengru Song, Xiaogiang Lei, Di Zhang, Wenwu Ou, Kun Gai, and Yadong Mu. Unified Language-Vision Pretraining in LLM with Dynamic Discrete Visual Tokenization, 2023. 3", + "[26] Yang Jin, Zhicheng Sun, Kun Xu, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, Kun Kai, and Yadong Mu. Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization, 2024. 3", + "[27] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. MiraData: A Large-Scale Video Dataset with Long Durations and Structured Captions, 2024. 3", + "[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment Anything, 2023. 4", + "[29] Jing Yu Koh, Daniel Fried, and Russ Salakhutdinov. Generating images with multimodal language models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 3", + "[30] Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, Krishna Somandepalli, Hassan Akbari, Yair Alon, Yong Cheng, Joshua V. Dillon, Agrim Gupta, Meera Hahn, Anja Hauth, David Hendon, Alonso Martinez, David Minnen, Mikhail Sirotenko, Kihyuk Sohn, Xuan Yang, Hartwig Adam, Ming-Hsuan Yang, Irfan Essa, Huisheng Wang, David A. Ross, Bryan Seybold, and Lu Jiang. VideoPoet: A large language model for zero-shot video generation. In *Forty-First International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 1, 2", + "[31] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Daquan Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. HunyuanVideo: A Systematic Framework For Large Video Generative Models, 2024. 3, 6", + "[32] Didier Le Gall. MPEG: A video compression standard for multimedia applications. Communications of the ACM, 34 (4):46-58, 1991. 2, 3, 4", + "[33] Zongyi Li, Shujie Hu, Shujie Liu, Long Zhou, Jeongsoo Choi, Lingwei Meng, Xun Guo, Jinyu Li, Hefei Ling, and Furu Wei. ARLON: Boosting Diffusion Transformers with Autoregressive Models for Long Video Generation, 2024. 3, 6, 7" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, Tanghui Jia, Junwu Zhang, Zhenyu Tang, Yatian Pang, Bin She, Cen Yan, Zhiheng Hu, Xiaoyi Dong, Lin Chen, Zhang Pan, Xing Zhou, Shaoling Dong, Yonghong Tian, and Li Yuan. Open-Sora Plan: Open-Source Large Video Generation Model, 2024. 3, 6", + "[35] Zongyu Lin, Wei Liu, Chen Chen, Jiasen Lu, Wenze Hu, Tsu-Jui Fu, Jesse Allardice, Zhengfeng Lai, Liangchen Song, Bowen Zhang, Cha Chen, Yiran Fei, Yifan Jiang, Lezhi Li, Yizhou Sun, Kai-Wei Chang, and Yinfei Yang. STIV: Scalable Text and Image Conditioned Video Generation, 2024. 3", + "[36] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 14", + "[37] Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. FreeLong: Training-Free Long Video Generation with Spectral-Blend Temporal Attention, 2024. 3", + "[38] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6", + "[39] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent Diffusion Transformer for Video Generation, 2024. 3, 6", + "[40] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINov2: Learning robust visual features without supervision. 2024, 2024. 4", + "[41] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers, 2023. 3", + "[42] Zhiwu Qing, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yujie Wei, Yingya Zhang, Changxin Gao, and Nong Sang. Hierarchical spatio-temporal decoupling for text-to-video generation. In CVPR, 2024. 6", + "[43] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. FreeNoise: Tuning-free longer video diffusion via noise rescheduling. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3, 7", + "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 4", + "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15613", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21:140:1-140:67, 2020. 5", + "[46] Jinghuan Shang, Karl Schmeckpeper, Brandon B. May, Maria Vittoria Minniti, Tarik Kelestemur, David Watkins, and Laura Herlant. Theia: Distilling Diverse Vision Foundation Models for Robot Learning, 2024. 2, 4, 13", + "[47] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1874-1883. IEEE Computer Society, 2016. 15", + "[48] Chenyang Si, Weichen Fan, Zhengyao Lv, Ziqi Huang, Yu Qiao, and Ziwei Liu. RepVideo: Rethinking Cross-Layer Representation for Video Generation, 2025. 6", + "[49] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3", + "[50] Jianlin Su, Murtadha H. M. Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 6, 13", + "[51] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 14398-14409. IEEE, 2024. 3", + "[52] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024.6", + "[53] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and Efficient Foundation Language Models, 2023. 5, 6, 15", + "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3", + "[55] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yulong Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin, Tiejun Huang, and Zhongyuan Wang. Emu3: Next-Token Prediction is All You Need, 2024. 1, 2, 3, 6", + "[56] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu," + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, pages 1-20, 2024. 3", + "[57] Yuancheng Wang, Haoyue Zhan, Liwei Liu, Ruihong Zeng, Haotian Guo, Jiachen Zheng, Qiang Zhang, Xueyao Zhang, Shunsi Zhang, and Zhizheng Wu. Maskgct: Zero-shot text-to-speech with masked generative codec transformer. arXiv preprint arXiv:2409.00750, 2024. 4", + "[58] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video Generation using VQ-VAE and Transformers, 2021. 3", + "[59] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 10371-10381. IEEE, 2024. 4", + "[60] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer, 2024. 1, 2, 3, 5, 6, 15", + "[61] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 4", + "[62] Lijun Yu, José Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G. Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A. Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3, 7, 13", + "[63] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An Image is Worth 32 Tokens for Reconstruction and Generation, 2024. 2, 4, 13", + "[64] Hangjie Yuan, Shiwei Zhang, Xiang Wang, Yujie Wei, Tao Feng, Yining Pan, Yingya Zhang, Ziwei Liu, Samuel Albanie, and Dong Ni. InstructVideo: Instructing video diffusion models with human feedback. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 6463-6474. IEEE, 2024. 6", + "[65] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make Pixels Dance: High-Dynamic Video Generation, 2023. 3", + "[66] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying Pixel and Latent Diffusion Models for Text-to-Video Generation, 2023. 6", + "[67] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15614", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ICCV 2023, Paris, France, October 1-6, 2023, pages 3813-3824. IEEE, 2023. 5", + "[68] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 3, 6, 7", + "[69] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient Video Generation With Latent Diffusion Models, 2023. 3" + ], + "bbox": [ + 91, + 90, + 483, + 219 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "15615", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/b1aafd05-162c-4fda-a00d-cb706731d3d7_model.json b/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/b1aafd05-162c-4fda-a00d-cb706731d3d7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..101023f696cc0142d2f8ec08dbc85d7d2b4e09ea --- /dev/null +++ b/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/b1aafd05-162c-4fda-a00d-cb706731d3d7_model.json @@ -0,0 +1,2237 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.13, + 0.892, + 0.175 + ], + "angle": 0, + "content": "The Best of Both Worlds: Integrating Language Models and Diffusion Models for Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.203, + 0.892, + 0.223 + ], + "angle": 0, + "content": "Aoxiong Yin\\(^{1*}\\), Xu Tan\\(^{2*†}\\), Kai Shen\\(^{2*}\\), Yichong Leng\\(^{2}\\), Xinyu Zhou\\(^{2}\\), Juncheng Li\\(^{1†}\\), Siliang Tang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.224, + 0.608, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\)Zhejiang University, \\(^{2}\\)Moonshot AI" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.242, + 0.782, + 0.256 + ], + "angle": 0, + "content": "yinaoxiong@zju.edu.cn, tanxu2012@gmail.com, junchengli@zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.324, + 0.486, + 0.687 + ], + "angle": 0, + "content": "Recent advancements in text-to-video (T2V) generation have been driven by two competing paradigms: autoregressive language models and diffusion models. However, each paradigm has intrinsic limitations: language models struggle with visual quality and error accumulation, while diffusion models lack semantic understanding and causal modeling. In this work, we propose LanDiff, a hybrid framework that synergizes the strengths of both paradigms through coarse-to-fine generation. Our architecture introduces three key innovations: (1) a semantic tokenizer that compresses 3D visual features into compact 1D discrete representations through efficient semantic compression, achieving \\(a \\sim 14,000 \\times\\) compression ratio; (2) a language model that generates semantic tokens with high-level semantic relationships; (3) a streaming diffusion model that refines coarse semantics into high-fidelity videos. Experiments show that LanDiff, a 5B model, achieves a score of 85.43 on the VBenchmark T2V benchmark, surpassing the state-of-the-art open-source models Hunyuan Video (13B) and other commercial models such as Sora, Kling, and Hailuo. Furthermore, our model also achieves state-of-the-art performance in long video generation, surpassing other open-source models in this field. Our demo can be viewed at https://landiff.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.717, + 0.222, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Text-to-video (T2V) [6, 30, 55, 60] has made significant progress in recent years, becoming an important research direction in the fields of computer vision and artificial intelligence. Recent works in T2V models have primarily revolved around two predominant paradigms: autoregressive large language model (LLM)-based [30, 55] frameworks and diffusion-based architectures [6, 60]. However, each paradigm has their own advantages and limitations, as" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.481, + 0.907, + 0.607 + ], + "angle": 0, + "content": "Figure 1. The rate-distortion curve illustrates how visual distortion decreases as the number of transmitted bits increases. With just a small number of bits representing high-level semantic features, we can already achieve relatively low visual distortion. Building on this information-theoretic insight, LanDiff combines the strengths of both paradigms: LLMs efficiently generate compact semantic features in the first stage, followed by diffusion models that add perceptual details in the second stage, before final decoding to pixels via VAE. Data from Ho et al. [20], illustration is conceptual." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.626, + 0.634, + 0.64 + ], + "angle": 0, + "content": "shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.903 + ], + "angle": 0, + "content": "From a representation perspective, LLM-based methods leverage discrete tokenization to explicitly encode high-level semantics through vector quantization, effectively prioritizing conceptual abstraction and narrative coherence. However, this discretization inherently sacrifices low-level visual fidelity due to information compression, resulting in low reconstruction quality. In contrast, diffusion-based approaches employ continuous latent representations to preserve much more perceptual details, enabling superior reconstruction quality at the cost of diluted semantic interpretability, as hierarchical features remain entangled in the latent space. From a generative modeling perspective, LLM-based systems adopt autoregressive modeling to enforce causal dependencies between video frames, ensuring strong temporal coherence. However, the autoregressive generation inherently risks error propagation across time steps, where inaccuracies in early predictions amplify dur" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15604" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Table 1. The comparison between LLM based and diffusion based T2V systems. The advantages and disadvantages are marked by \\(\\text{喜}\\) and \\(\\text{喜}\\), respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.142, + 0.479, + 0.242 + ], + "angle": 0, + "content": "
MethodsRepresentationsModeling
LLMDiscrete TokensAutoregressvie
Low Reconstruction QualityNo Refinement
Highlight Semantic InformationCausal Modeling
DiffusionContinuous VectorsNon-Autoregressvie
High Reconstruction QualityProgressive Refinement
Lack Semantic InformationNon-causal Modeling
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.485, + 0.351 + ], + "angle": 0, + "content": "ing decoding. In contrast, diffusion-based methods employ non-autoregressive generation, refining outputs in parallel through iterative denoising steps. Although this design mitigates error accumulation and enhances generation flexibility, the absence of explicit causal constraints often leads to temporal inconsistencies or semantic hallucinations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.354, + 0.485, + 0.807 + ], + "angle": 0, + "content": "In this work, we propose a hybrid architecture that synergizes the strengths of both Language models and Diffusion models, named LanDiff, through a coarse-to-fine generation paradigm. As shown in Figure 1, inspired by the human creation of video which will generate the high-level storyline first and then add low-level visual details based on the storyline to form the video, we design a two-stage video generation process with the number of bits gradually increasing and carefully design the autoregressive model and the diffusion model to be responsible for different stages of T2V generation, so as to play to their strengths and avoid their weaknesses. In detail, 1) at the low-bit position \"semantic feature\", the low-bit information ensures that the token sequence is not too long, and the high-level information makes it easier for the model to capture the overall semantic entity motion of the video, so as to play to the strengths of the autoregressive model and avoid its weaknesses. Thus we use LLM to generate a coarse-grained video in the first stage; 2) at the high-bit position \"perceptual feature\", since we have already obtained the coarse-grained with rich semantic and time-serial information, we only need to focus on how to add details to the coarse-grained video. Thus we apply a diffusion model in the second stage. Finally, a VAE decoder transforms the generated \"perceptual feature\" into the final RGB video output. By unifying these complementary mechanisms, we demonstrate that hybrid architectures can overcome the inherent limitations of isolated approaches, enabling coherent, semantically faithful, and visually compelling video generation from textual descriptions, as shown in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.903 + ], + "angle": 0, + "content": "With this design, the ideal semantic feature should contain high-level semantic information and motion information and only require a few bits to represent. We achieve this goal by performing extreme compression on video representations rich in high-level semantics. For video representation, we select the Theia model [46] as our visual rep" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.148 + ], + "angle": 0, + "content": "Table 2. The comparison between LanDiff and previous largescale T2V systems. The compression rates of LLM-based and diffusion-based models are illustrated using VideoPoet [30] and CogVideoX [60] as examples, respectively." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.156, + 0.9, + 0.267 + ], + "angle": 0, + "content": "
ModelsLLMDiffusionLanDiff
Highlight Semantic Information?✓✗✓
Progressive Refinement?✗✓✓
Causal Modeling?✓✗✓
High Visual Quality?✗✓✓
Compression Ratio ↑~256~1024~14000
Long video Generation?✗✗✓
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.272, + 0.908, + 0.59 + ], + "angle": 0, + "content": "resentation backbone, which has been distilled from multiple visual understanding and self-supervised representation models, ensuring the encoded features contain rich semantic information. To achieve extreme compression and reduce the number of bits, we design an efficient tokenizer to compress 3D visual features into 1D discrete representations. The tokenizer is based on the Transformer [54] structure, uses query embedding to aggregate visual features, and has a higher compression rate than CNN-based structures [63]. To further compress the video by fully utilizing the temporal redundancy of the video, inspired by the MP4 video encoding algorithm [32], we divide the video into keyframes and non-keyframes, and set more numbers of tokens for keyframes. The detailed design is shown in subsection 3.1. For the diffusion model, we use generated semantic tokens as conditions and generate the target video by gradually removing the noises. To better support the long-video generation, we train a chunk-wise streaming diffusion model that only uses a limited number of historical frames as conditions, thereby greatly reducing the computational cost of training and inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.909, + 0.772 + ], + "angle": 0, + "content": "Thanks to these designs, our LanDiff has made significant progress in spatial relationship compliance, action coherence, visual quality, etc. Specifically, 1) for short video generation, our 5B model achieved a score of 85.43 on the VBenchmark, surpassing the state-of-the-art open-source models Hunyuan Video (13B) and other commercial models such as Sora, Kling, and Hailuo; 2) for long video generation, after testing by the VBenchmark T2V benchmark, our model also achieved state-of-the-art performance, surpassing other open-source models in this field. Our video examples can be viewed at https://landiff.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.785, + 0.659, + 0.801 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Video Tokenization. Video tokenization plays a crucial role in video understanding and generation tasks. Since a video can be represented as a sequence of continuous frames, some works directly use image tokenizers to process videos frame by frame. For example, Wang et al. [55] directly use SBER-MoVQGAN as the video tokenizer." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15605" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.362 + ], + "angle": 0, + "content": "However, this method ignores the temporal redundancy in videos, resulting in a low compression rate. To reduce temporal redundancy, some works [58, 62] try to extend image tokenizers based on 2D convolution to 3D convolution, which can process both temporal and spatial information simultaneously. These methods encode videos in the original RGB space and perform video reconstruction tasks, which are more about perceptual compression. In addition, some works [15, 25] try to train video tokenizers on features extracted from pre-trained visual encoders. These tokenizers can achieve good performance in understanding and generation tasks while maintaining a high compression rate. The video tokenizer we propose belongs to this category. Unlike previous feature-based tokenizers that achieved limited compression rates and primarily focused on image processing, our video tokenizer delivers significantly higher compression while handling both images and videos in a unified framework." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.369, + 0.483, + 0.732 + ], + "angle": 0, + "content": "LLM based Video Generation. LLM-based video generation methods are usually based on the Transformer [54] structure, learning the mapping from text to video through next token prediction. TATS [14] uses VQ-GAN as the tokenizer and predicts video tokens using a GPT-like model structure. VideoGPT [58] uses 3D convolution to extract features and quantize them, and predicts the quantized video discrete tokens using a GPT-like model. Recently, VideoPoet [62] uses MagViT2 [62] as the tokenizer and unifies multiple modalities as input to a large language model (LLM) to conditionally generate video tokens. In addition, Emu3 [55] uses SBER-MoVQGAN as the tokenizer to perform video understanding and generation by predicting the next token. These works all use LLMs to directly generate \"perceptual features\" that contain rich visual details with high bit rates. Recently, ARLON [33] attempts to discretize VAE features into a small number of tokens to reduce the bit rate required for LLMs' prediction. In this way, the tokens retain low-frequency visual information such as blurry contours rather than high-level semantic information. In contrast, our method employs tokens containing high-level semantic information as prediction targets for LLMs, which enables us to fully leverage LLMs' advantages in causal modeling to precisely generate high-quality videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Diffusion based Video Generation. Diffusion-based methods have achieved great success in image generation, and recently many people [7, 21, 22, 49, 69] have tried to apply them to video generation tasks. VDM [22] extends the 3D U-Net structure for video generation. Wang et al. [56] propose to generate high-quality and aesthetically pleasing videos in a cascaded manner. Benefiting from the success of the text-to-image (T2I) field, some works such as Animatediff [16], SVD [6], and PixelDance [65] try to use pre-trained T2I models as initialization, and then add modules for temporal modeling to capture motion informa" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.332 + ], + "angle": 0, + "content": "tion for video generation. Ma et al. [39] explore the generation capabilities of multiple different structures of latent diffusion transformer. After the release of SORA, a series of video generation methods based on the DiT [41] model have been proposed, including OpenSora [68], OpenSoraPlan [34], Cogvideox [60], Hunyuan Video [31], Mira [27] and STIV [35] etc. These methods can only generate short videos of a few seconds. Recently, StreamingT2V [18] generates long videos by block-wise generation on a pre-trained short video generation model, and then uniformly performs mixed augmentation. In addition, some works improve the consistency of long video generation by leveraging noise rescheduling techniques [37, 43]. Our method employs diffusion models as renderers for semantic features, enabling us to leverage their superior visual generation quality while circumventing their limitations in causal modeling." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.346, + 0.605, + 0.361 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.905, + 0.597 + ], + "angle": 0, + "content": "In this work, we propose a novel text-to-video generation framework that synergistically integrates the strengths of autoregressive modeling and diffusion processes while circumventing their respective limitations. The framework mainly consists of the following components: 1) an efficient tokenizer that transforms 3D visual features into compact 1D discrete representations while preserving and enhancing their semantic information; 2) a language model that performs temporal sequence modeling to generate semantic tokens representing video blueprints from textual descriptions; 3) a streaming diffusion model that progressively refines coarse semantic videos by adding fine-grained details to produce high-quality VAE features; and 4) a VAE decoder that reconstructs the final video frames from the refined VAE features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.75, + 0.622 + ], + "angle": 0, + "content": "3.1. Video Semantic Tokenizer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.84 + ], + "angle": 0, + "content": "In this section, we introduce a novel video semantic tokenizer that efficiently compresses the video into semantic information. Firstly, we will introduce the video semantic representation used in tokenization. Then, considering the high redundancy of video in both spatial and temporal dimensions, we introduce two compression strategies: 1) a query-based causal tokenization that efficiently reduces spatial redundancy while preserving essential semantic information through vector quantization; 2) inspired by MP4 [32], we implement video frame grouping to minimize temporal redundancy by treating grouped frames as a unit, where the first frame (I-frame) is fully encoded while subsequent frames (P-frames) only capture the temporal changes by referencing content from previous frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Video Semantic Representation. Generally, video representations can be divided into two categories: 1) some methods [55, 62] directly utilize an autoencoder to learn the video representations. 2) Some works [25, 26, 29, 51] use" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15606" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.176, + 0.089, + 0.825, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.339, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 2. The architecture of LanDiff. Given text inputs, we first extract text embeddings and employ an LLM to generate semantic tokens in the first stage. Subsequently, we utilize a diffusion model to synthesize perceptual features conditioned on these semantic tokens, followed by a VAE decoder that transforms these features into the final video frames." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.393, + 0.482, + 0.482 + ], + "angle": 0, + "content": "pre-trained visual self-supervised learning features (SSL) as video representations. Compared with the first directly learned autoencoder latents which contain lots of visual details, the second SSL features maintain much more semantic features, which enable the LLM to focus more on the high-level semantic information of the video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.574 + ], + "angle": 0, + "content": "Based on these thoughts, we choose the pretrained SSL features as the video representations. For better usage of different SSL models, we choose Theia model [46], which is a unified visual feature extractor distilled from multiple visual task models, including CLIP [44], SAM [28], DINOv2 [40], ViT [10], and Depth-Anything [59]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.649 + ], + "angle": 0, + "content": "Tokenizer Design. In this part, we introduce the design of our tokenizer, which leverages the query tokens to compress the video semantic features and uses quantization to discrete the video semantic representation while minimizing the reconstruction loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.756 + ], + "angle": 0, + "content": "In detail, firstly, we extract the semantic features using the Theia model and flatten it to obtain \\( F \\in \\mathbb{R}^{(T \\times H \\times W) \\times D} \\). Inspired by TiTok [63], we use N randomly initialized tokens as query tokens \\( Q \\in \\mathbb{R}^{N \\times D} \\) and concatenate them with the semantic features \\( F \\). Then we use a transformer encoder to encode them and only take the encoded features of the query tokens." + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.766, + 0.483, + 0.784 + ], + "angle": 0, + "content": "\\[\nZ _ {Q} = \\operatorname {E n c} ([ F; Q ]), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.791, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\([\\cdot ]\\) represents the concatenation operation and \\(Z_{Q}\\in\\) \\(\\mathbb{R}^{N\\times D}\\) . We then apply vector quantization on \\(Z_{Q}\\) by train a VQ-VAE model and obtain the quantized feature \\(\\hat{Z}_Q\\) . In the decoding stage, the quantized feature \\(\\hat{Z}_Q\\) is used as a condition, and then a sequence of mask tokens \\(M\\in \\mathbb{R}^{T\\times H\\times W}\\) are added in front of the \\(\\hat{Z}_Q\\) to form the inputs of decoder. Then we only take the features of the mask tokens as fol" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.393, + 0.553, + 0.406 + ], + "angle": 0, + "content": "lows:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.406, + 0.905, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\hat {F} = \\operatorname {D e c} ([ M; \\hat {Z} _ {Q} ]), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.905, + 0.504 + ], + "angle": 0, + "content": "where \\(\\hat{F} \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}\\) represents the reconstructed feature, and \\(M\\) represents the mask tokens. Inspired by Huang et al. [23], Wang et al. [57], we minimize the reconstruction loss of the video semantic feature during training the VQ-VAE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.905, + 0.58 + ], + "angle": 0, + "content": "For the VQ-VAE, we follow the method of Wang et al. [57], Yu et al. [61]. We update the codebook using exponential moving average (EMA). We also optimize the codebook with the video semantic feature reconstruction loss. The loss function is shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.588, + 0.905, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {r e c} \\| \\hat {F} - F \\| _ {2} + \\lambda_ {c o m m i t} \\| \\operatorname {s g} (\\hat {Z} _ {Q}) - Z _ {Q} \\| _ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.642 + ], + "angle": 0, + "content": "where \\(\\mathrm{sg}()\\) is the stop-gradient operation, \\(||\\cdot ||_2\\) is the L2 loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Video Frame Grouping. For video, a straightforward observation is the redundancy in time series (i.e., the difference between adjacent video frames is minimal). Intuitively, we can achieve a better compression rate by modeling the difference between adjacent video frames instead of tokenizing all frames equally. Inspired by the popular video compression method MP4 [32], given \\( N \\) frames of a video, we first split them into \\( N / T \\) groups (i.e., \\( T \\) frames as a group. For clarification, as shown in Figure 3, we use \\( T = 3 \\) as an example.). Then: 1) we will model different groups independently; 2) for each group, we will fully encode the first frame (Intra-coded Frame, IFrame), while for the remaining \\( [1,T - 1] \\) frames (the Predictive-coded Frame, PFrame), we encode them by referencing their previous frames." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To achieve this, within a group, the first frame (IFrame) will only see itself and have a large number of query tokens" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15607" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.214, + 0.092, + 0.498, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.092, + 0.782, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.38, + 0.908, + 0.424 + ], + "angle": 0, + "content": "Figure 3. Proposed architecture of the video semantic tokenizer. We use query tokens to compress the semantic sequence length. Furthermore, we group the frames into groups (3 frames in a group in this figure). In a group, the first frame is the IFrame and the rest frames are PFrames. We use different query token numbers for them. The attention mask design is shown in the right." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.428, + 0.486, + 0.626 + ], + "angle": 0, + "content": "(e.g., 3 query tokens in Figure 3) to achieve better reconstruction quality. For the rest frame (PFrame) \\( i \\in [1, T - 1] \\), it will see the previous frames (i.e., frames \\( j \\in [0, i - 1] \\)) and have a small number of query tokens (e.g., 1 query token which is \\( 1/3 \\) of the IFrame) to force the model to learn the difference. Technically, as shown in the attention mask on the right side of Figure 3, we use frame-level causal masks for the feature sequence during encoding. The query tokens are also divided according to the frame, and each token can only attend to the features of the corresponding frame and the previous frames. During decoding, the mask token corresponding to each frame can see the previous features, the corresponding query tokens, and the previous query tokens." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.636, + 0.485, + 0.666 + ], + "angle": 0, + "content": "3.2. Language Model for Semantic Token Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.486, + 0.901 + ], + "angle": 0, + "content": "As shown in Figure 2, after training an efficient tokenizer, we use a language model to generate semantic tokens autoregressively based on text. Specifically, we first use a pretrained text encoder T5-XXL [45] to extract text features \\( X \\). We use the video tokenizer encoder trained in subsection 3.1 to convert the video into a 1D discrete token sequence \\( Y \\). To enhance the controllability of the generated results, we introduce additional control conditions \\( CC \\). It includes: 1) frames condition for requiring the model to generate videos with a specified number of frames; 2) motion score condition, which is a value between 0 and 1, used to control the degree of motion in the generated video. The discrete tokens are converted into embedding vectors during generation and added with positional encoding, and then concatenated with the control conditions as input. The structure of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.908, + 0.473 + ], + "angle": 0, + "content": "the language model follows the typical LLaMA model [53] network structure. We train the model from scratch, using cross-entropy loss as the loss function." + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.489, + 0.907, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L M}} = \\mathbb {E} \\left[ - \\log p \\left(Y _ {i} \\mid X, C C, Y _ {< i}\\right) \\right] \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.907, + 0.546 + ], + "angle": 0, + "content": "3.3. Diffusion Model for Perceptual Feature Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.674 + ], + "angle": 0, + "content": "The video detokenizer in LanDiff is responsible for converting semantic tokens into VAE latent vectors. As shown in Figure 2 we use a conditional diffusion model to complete this task. Especially to support long-video generation, we also design the streaming inference strategy. In this section, we will first introduce the architecture for a single semantic token chunk. Then we will introduce the chunk-wise streaming strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Architecture. Our diffusion model employs an architecture similar to MMDiT[12, 60]. Specifically, 1) we use the video tokenizer decoder trained in subsection 3.1 to decode the semantic tokens into semantic features \\(\\hat{F}\\). Then we use the semantic features \\(\\hat{F}\\) as a condition to guide the diffusion model to generate videos; 2) we inject the control signals into the model in a similar way to ControlNet [67]. In detail, during training, the parameters of the main model are not updated, the control module copies the parameters of the first half layers of the main model, and adds them to the output of the main model after a linear layer initialized with zeros; 3) to make the semantic features match the target VAE features in the space dimension, we additionally add an upsampling module. We show the model structure in the appendix Figure 8." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15608" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "During training, given the input video chunk \\(V \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}\\), where \\(T\\) is the frame length of the VAE latent and \\(D\\) is the VAE latent feature dimension, the diffusion algorithm progressively adds noise to the video and produces a noisy video \\(V^t\\), where \\(t\\) represents the number of times noise is added. The diffusion algorithm uses time step \\(t\\), and semantic features \\(\\hat{F} \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}\\) as conditions, and then uses a network \\(\\epsilon_{\\theta}\\) to predict the noise added to the noisy video \\(V_t\\) through:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.233, + 0.483, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {V, t, c _ {s}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(V ^ {t}, t, \\hat {F}\\right) \\| _ {2} ^ {2} \\right] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.485, + 0.426 + ], + "angle": 0, + "content": "Chunk-wise Streaming Strategy. To support long video generation scenario whose semantic token sequence is very long and difficult to generate as a whole, we propose the chunk-wise streaming strategy. During training, given video latent chunk \\( V \\in \\mathbb{R}^{(T \\times H \\times W) \\times D} \\), to maintain the appearance continuity in the video, we use the first half of \\( V(V_{l}) \\) as the prompt and generate the second half of \\( V_{r} \\). In detail, we do not add noise to the first half chunk and give time condition \\( t \\) to 0.999. We also randomly mask the first half chunk with a ratio of \\( 20\\% \\). The loss function for training at this time is:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.43, + 0.483, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {V _ {r}, t, c _ {s}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(V _ {l}, V _ {r} ^ {t}, t, c _ {s}\\right) \\| _ {2} ^ {2} \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.522 + ], + "angle": 0, + "content": "During inference, the first \\( L / 2 \\) VAE latents will be generated without the prompt. For the following VAE latents, we will accumulate \\( L / 2 \\) tokens to form a chunk and use the previous chunk as prompt." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.533, + 0.325, + 0.55 + ], + "angle": 0, + "content": "4. Experiments and Results" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.557, + 0.299, + 0.573 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Datasets. For the video tokenizer and language model, we use an internal dataset with 200M video-text pairs for training. All videos with a duration of less than 6s are filtered, and the videos are kept in the aspect ratio and then scaled to around 480x720 resolution for center cropping. Consistent with CogVideoX [60], we set the fps of all videos to 8. For the diffusion model, we select a dataset containing 3M high-quality video-text pairs for training. To evaluate the performance of the text-to-video generation model, we use the prompts provided by the widely used VBench [24] T2V benchmark to generate videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.746, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Implementation Details. For the video tokenizer, we set the group size \\( T \\) mentioned in subsection 3.1 to 13. We set the number of tokens \\( I \\) corresponding to the IFrame to 330, and the number of tokens \\( P \\) corresponding to the PFrame to 74. The dimension of the tokenizer's codebook is set to 16, and the vocabulary size is set to 2048. For the language model, we use the LLaMA [53] structure, with 2B model parameters, and use 1D RoPE [50] positional encoding. We use the 2B version of the CogVideoX [60] model as the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.151 + ], + "angle": 0, + "content": "base model for our video detokenizer. We list the structure settings and more implementation details of each module in the appendix subsection A.1, subsection A.2, and subsection A.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.906, + 0.198 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate the text-to-video generation task, we use the metrics proposed in the VBench and VBench-Long [24] benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.906, + 0.335 + ], + "angle": 0, + "content": "Evaluation Baselines. We compare LanDiff with baselines: 1) Sora [5]. 2) Jimeng [3]. 3) Hailuo [2]. 4) OpenSoraPlan V1.1 [34]. 5) Kling [38]. 6) InstructVideo [64]. 7) Gen-3 [1]. 9) Latte-1 [39] 9) HiGen [42]. 10)AnimateDiffV2 [16]. 11) Show-1 [66]. 12) Pika [4]. 13) VideoCrafter2.0 [7]. 14) OpenSora V1.2 [68]. 15) LTX-Video [17]. 16) Mochi-1 [52]. 17) CogVideoX [60]. 18) Vchitect-2.0 [13]. 19) RepVideo [48]. 20) HunyuanVideo [31]. 21) ARLON [33]. 22) Emu3 [55]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.346, + 0.716, + 0.362 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "ref_text", + "bbox": [ + 0.52, + 0.383, + 0.9, + 0.463 + ], + "angle": 0, + "content": "A life-sized ice sculpture of a playful dog, with intricate details and a joyful expression, stands in the middle of a sunlit, grassy field on a sweltering summer day. The ice dog, initially solid and vibrant, begins to melt under the relentless heat, with droplets of water forming on its surface. As the day progresses, the ice dog's form gradually diminishes, with its once sharp features becoming blurred and distorted. The melting process accelerates, and the ice dog's body starts to collapse, pooling into a puddle of water on the ground. By the end of the day, all that remains is a shallow puddle, reflecting the cloudless sky, with the memory of the once majestic ice dog now just a memory." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.46, + 0.63, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.462, + 0.718, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.462, + 0.807, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.462, + 0.894, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.565, + 0.906, + 0.592 + ], + "angle": 0, + "content": "Figure 4. Comparison of qualitative results for text-to-video generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Text to Video Generation. As shown in Table 3, our 5B LanDiff model outperforms the state-of-the-art open-source models on VBenchmark, including the 13B HunyuanVideo, achieving a top score of 85.42. LanDiff exhibits superior semantic understanding while maintaining excellent generation quality, surpassing same-sized CogVideoX-5B across most metrics. To validate our approach's effectiveness independent of training data volume, we trained a 7B DiT-based diffusion model with identical training data (marked as DiT in the table). The results confirm that our task division strategy successfully leverages the complementary strengths of autoregressive and diffusion models. Figure 4 shows qualitative comparisons between LanDiff and CogVideoX-5B. Our model effectively captures temporal dynamics described in the prompt, accurately rendering a melting ice dog sculpture. LanDiff successfully depicts the progressive transformation from solid form to liquid state, while CogVideoX-5B produces a static ice sculpture that remains virtually unchanged throughout the sequence. This demonstrates our model's superior ability to compre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.442, + 0.9 + ], + "angle": 0, + "content": "1https://huggingface.co/THUDM/CogVideoX-2b" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15609" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.159 + ], + "angle": 0, + "content": "Table 3. Performance comparison of Text-to-video (T2V) generation between our LanDiff and other state-of-the-art models on VBench benchmark. We selected 8 out of the 16 evaluation dimensions from VBench, along with Total Score, Quality Score, and Semantic Score, for presentation. The complete results with all 16 evaluation dimensions can be found in the appendix Table 8. The best and second-best scores are highlighted in bold and underline, respectively. \\(\\dagger\\) indicates the scores we reproduced, while \\(\\ddagger\\) indicates the scores from the original papers, and other scores are from the VBench benchmark." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.161, + 0.862, + 0.535 + ], + "angle": 0, + "content": "
Model NameTypeModel SizeTotal ScoreQuality ScoreSemantic Scorebackground consistencydynamic degreemotion smoothnessmultiple objectsobject classscenespatial relationshipsubject consistency
Open Sourced Models
InstructVideoDiffusion1.3B76.6181.5656.8196.9769.7296.6221.5773.2622.2143.4995.30
Latte-1Diffusion0.7B77.2979.7267.5895.4068.8994.6334.5386.5336.2641.5388.88
OpenSoraPlan V1.1Diffusion2.7B78.0080.9166.3896.7347.7298.2840.3576.3027.1753.1195.73
Show-1Diffusion6.3B78.9380.4272.9898.0244.4498.2445.4793.0747.0353.5095.53
OpenSora V1.2Diffusion1.1B79.7681.3573.3997.6142.3998.5051.8382.2242.4468.5696.75
LTX-VideoDiffusion1.9B80.0082.3070.7997.2054.3598.9645.4383.4551.0765.4396.56
Mochi-1Diffusion10B80.1382.6470.0897.2861.8599.0250.4786.5136.9969.2496.99
AnimateDiff-V2Diffusion1.3B80.2782.9069.7597.6840.8397.7636.8890.9050.1934.6095.30
VideoCrafter-2.0Diffusion1.7B80.4482.2073.4298.2242.5097.7340.6692.5555.2935.8696.85
CogVideoX-2BDiffusion2B80.9182.1875.8396.6359.8697.7362.6383.3751.1469.9096.78
Emu3 ‡LLM8B80.96N/AN/A97.6979.2798.9344.6486.1737.1168.7395.32
Vchitect-2.0-2BDiffusion2B81.5782.5177.7996.5358.3397.7669.3587.8157.5154.6496.42
CogVideoX-5BDiffusion5B81.6182.7577.0496.5270.9796.9262.1185.2353.2066.3596.23
DiT †Diffusion7B81.8582.7078.4297.6551.0298.9469.6393.4057.1060.8997.05
RepVideoDiffusion2B81.9482.7078.9196.5657.7898.1371.1887.8352.9674.7496.25
Vchitect-2.0[E]Diffusion2B82.2483.5477.0696.6663.8998.9868.8486.6156.5757.5596.83
HunyuanVideoDiffusion13B83.2485.0975.8297.7670.8398.9968.5586.1053.8868.6897.37
Close Sourced Models
Pika-1.0DiffusionN/A80.6982.9271.7797.3647.5099.5043.0888.7249.8361.0396.94
KlingDiffusionN/A81.8583.3975.6897.6046.9499.4068.0587.2450.8673.0398.33
JimengDiffusionN/A81.9783.2976.6998.3938.4398.0969.0889.6244.9477.4597.25
Gen-3DiffusionN/A82.3284.1175.1796.6260.1499.2353.6487.8154.5765.0997.10
HailuoDiffusionN/A83.4184.8577.6597.0564.9199.2276.0487.8350.6875.5097.51
SoraDiffusionN/A84.2885.5179.3596.3579.9198.7470.8593.9356.9574.2996.23
ARLON ‡LLM+Diffusion1.5BN/AN/AN/A97.1052.7798.92N/A89.8054.43N/A93.41
ARLON †LLM+Diffusion5B82.3183.5877.2798.6572.2297.5674.4990.6052.0762.5395.59
LanDiffLLM+Diffusion5B85.4386.1382.6198.7392.7197.0886.6994.9453.7973.7496.11
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.554, + 0.483, + 0.599 + ], + "angle": 0, + "content": "hend and visualize complex temporal transformations and physical processes. For more video examples, please refer to our demo website2." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.601, + 0.483, + 0.873 + ], + "angle": 0, + "content": "Long Video Generation. We compare our model with other open-source text-to-long video generation models: FreeNoise [43], StreamingT2V [18], OpenSora-V1.2 [68], and ARLON [33]. As shown in Table 4, LanDiff achieves state-of-the-art performance with the highest Total Score of 68.34. While StreamingT2V shows superior Dynamic Degree (85.64), it sacrifices consistency metrics (Subject: 87.31, Background: 94.64). Conversely, models like ARLON excel in consistency but exhibit limited dynamism (50.42). Our approach strikes an optimal balance with strong dynamism (72.86) while maintaining high consistency scores and achieving the best results in Aesthetic Quality (60.96) and Overall Consistency (27.29). As illustrated in Figure 5, LanDiff effectively renders complex motion dynamics in the prompt \"a car accelerating to gain speed,\" successfully depicting increasing velocity while preserving visual consistency. In contrast, competing models struggle with this motion, producing either static" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.597 + ], + "angle": 0, + "content": "vehicles, inconsistent appearances, or unrealistic acceleration patterns. Additional video samples are available on our demo site2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Video Tokenization. We present visualization results of our model's video reconstruction in Figure 6. We extract semantic tokens from the reference video and subsequently convert these tokens back into video using the video detokenizer. The results demonstrate that through our careful design, our video tokenizer can accurately reconstruct videos with accurate semantics and actions using only approximately 1/50th of the sequence length required by MagViT2 [62]. While some minor discrepancies in clothing details exist between the reconstructed and reference videos, these differences remain within acceptable limits for practical video generation applications. Additional video samples are available on our demo site2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.667, + 0.82 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this section, we conducted ablation experiments on the video tokenizer and classifier-free guidance. To fairly evaluate the effectiveness of our video tokenizer, we implemented an ARLON-like method as a comparison baseline. Specifically, for the control group without our proposed" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.322, + 0.9 + ], + "angle": 0, + "content": "2https://landiff.github.io/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15610" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 4. Long video generation results of LanDiff and other models on VBench. The best and second-best scores are highlighted in bold and underline, respectively." + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.12, + 0.819, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.498, + 0.908, + 0.527 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of long video generation results between LanDiff and other state-of-the-art models (FreeNoise, StreamingT2V and OpenSora-V1.2)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.531, + 0.473, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.667, + 0.483, + 0.693 + ], + "angle": 0, + "content": "Figure 6. Visualization results of video reconstruction using video tokenizer." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.697, + 0.483, + 0.723 + ], + "angle": 0, + "content": "Table 5. The ablation study of video tokenizer and classifier free guidance (cfg) on Vbench benchmark." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.725, + 0.436, + 0.822 + ], + "angle": 0, + "content": "
ModelsTotal ScoreQuality ScoreSemantic Score
LanDiff85.4386.1382.61
-video tokenizer82.3183.5877.27
-cfg81.0683.0473.13
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "video tokenizer, we followed ARLON's architectural approach by training a large language model to predict quantized VAE features, then using CogVideoX-5B as the base diffusion model conditioned on these LLM-predicted VAE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.907, + 0.699 + ], + "angle": 0, + "content": "features to generate videos. This baseline model was trained with exactly the same dataset and training recipe as Lan-Diff, ensuring strict experimental control variables and effectively eliminating interference factors such as model parameter count and training data scale. As shown in Table 5, the experimental results clearly demonstrate that our proposed video tokenizer significantly improves both the quality score and semantic score of generated videos, validating the importance of this component in our framework. In the case of not using classifier-free guidance for text, our model has decreased in quality and semantic scores." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.634, + 0.725 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a new text-to-video generation model, LanDiff. It combines the advantages of autoregressive models and diffusion models, including: 1) an efficient 1D video tokenizer to extract videos into semantic tokens; 2) a language model to generate semantic tokens based on text; 3) a video detokenizer to convert semantic tokens into videos. LanDiff outperforms the state-of-the-art open-source models on the VBenchmark, surpassing other open-source models in quality and semantic scores. At the same time, LanDiff also achieves state-of-the-art performance in long video generation tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15611" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.199 + ], + "angle": 0, + "content": "This work was supported by the NSFC (62272411), the Key R&D Projects in Zhejiang Province (No. 2024C01106, 2025C01030), the Zhejiang NSF (LRG25F020001), Zhejiang University Education Foundation Qizhen Scholar Foundation. This work was also supported by the Moonshot AI." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.211, + 0.188, + 0.227 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.237, + 0.484, + 0.265 + ], + "angle": 0, + "content": "[1] Gen-3. https://runwayml.com/research/introducing-gen-3-alpha, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.266, + 0.484, + 0.293 + ], + "angle": 0, + "content": "[2] Hailuo. https://platform.minimaxi.com/, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.295, + 0.48, + 0.309 + ], + "angle": 0, + "content": "[3] Jimeng. https://jimeng.jianying.com/, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.31, + 0.425, + 0.324 + ], + "angle": 0, + "content": "[4] Pika 1.0. https://www.pika.art/, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.351 + ], + "angle": 0, + "content": "[5] Openai sora. https://openai.com/index/sora, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.354, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[6] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.424, + 0.484, + 0.507 + ], + "angle": 0, + "content": "[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 7310-7320. IEEE, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.483, + 0.548 + ], + "angle": 0, + "content": "[8] Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High Fidelity Neural Audio Compression, 2022. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[9] Juechu Dong, Boyuan Feng, Driss Guessous, Yanbo Liang, and Horace He. Flex Attention: A Programming Model for Generating Optimized Attention Kernels, 2024. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.593, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming Transformers for High-Resolution Image Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873–12883, 2021. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. Open-Review.net, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[13] Weichen Fan, Chenyang Si, Junhao Song, Zhenyu Yang, Yinan He, Long Zhuo, Ziqi Huang, Ziyue Dong, Jingwen He," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.237, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "Dongwei Pan, Yi Wang, Yuming Jiang, Yaohui Wang, Peng Gao, Xinyuan Chen, Hengjie Li, Dahua Lin, Yu Qiao, and Ziwei Liu. Vchitect-2.0: Parallel Transformer for Scaling Up Video Diffusion Models, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[14] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long Video Generation with Time-Agnostic VQGAN and Time-Sensitive Transformer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[15] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a SEED of Vision in Large Language Model, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.43 + ], + "angle": 0, + "content": "[17] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. LTX-Video: Realtime Video Latent Diffusion, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[18] Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance, 2022. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.906, + 0.602 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.673 + ], + "angle": 0, + "content": "[21] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. CoRR, abs/2210.02303, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[23] Zhichao Huang, Chutong Meng, and Tom Ko. Repocodec: A speech representation codec for speech tokenization. arXiv preprint arXiv:2309.00169, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[24] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21807-21818. IEEE, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[25] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jian-choao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15612" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "Chengru Song, Xiaogiang Lei, Di Zhang, Wenwu Ou, Kun Gai, and Yadong Mu. Unified Language-Vision Pretraining in LLM with Dynamic Discrete Visual Tokenization, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[26] Yang Jin, Zhicheng Sun, Kun Xu, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, Kun Kai, and Yadong Mu. Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.277 + ], + "angle": 0, + "content": "[27] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. MiraData: A Large-Scale Video Dataset with Long Durations and Structured Captions, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.28, + 0.482, + 0.335 + ], + "angle": 0, + "content": "[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment Anything, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.337, + 0.482, + 0.419 + ], + "angle": 0, + "content": "[29] Jing Yu Koh, Daniel Fried, and Russ Salakhutdinov. Generating images with multimodal language models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.423, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[30] Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, Krishna Somandepalli, Hassan Akbari, Yair Alon, Yong Cheng, Joshua V. Dillon, Agrim Gupta, Meera Hahn, Anja Hauth, David Hendon, Alonso Martinez, David Minnen, Mikhail Sirotenko, Kihyuk Sohn, Xuan Yang, Hartwig Adam, Ming-Hsuan Yang, Irfan Essa, Huisheng Wang, David A. Ross, Bryan Seybold, and Lu Jiang. VideoPoet: A large language model for zero-shot video generation. In *Forty-First International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.606, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[31] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Daquan Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. HunyuanVideo: A Systematic Framework For Large Video Generative Models, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[32] Didier Le Gall. MPEG: A video compression standard for multimedia applications. Communications of the ACM, 34 (4):46-58, 1991. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[33] Zongyi Li, Shujie Hu, Shujie Liu, Long Zhou, Jeongsoo Choi, Lingwei Meng, Xun Guo, Jinyu Li, Hefei Ling, and Furu Wei. ARLON: Boosting Diffusion Transformers with Autoregressive Models for Long Video Generation, 2024. 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[34] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, Tanghui Jia, Junwu Zhang, Zhenyu Tang, Yatian Pang, Bin She, Cen Yan, Zhiheng Hu, Xiaoyi Dong, Lin Chen, Zhang Pan, Xing Zhou, Shaoling Dong, Yonghong Tian, and Li Yuan. Open-Sora Plan: Open-Source Large Video Generation Model, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[35] Zongyu Lin, Wei Liu, Chen Chen, Jiasen Lu, Wenze Hu, Tsu-Jui Fu, Jesse Allardice, Zhengfeng Lai, Liangchen Song, Bowen Zhang, Cha Chen, Yiran Fei, Yifan Jiang, Lezhi Li, Yizhou Sun, Kai-Wei Chang, and Yinfei Yang. STIV: Scalable Text and Image Conditioned Video Generation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[36] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[37] Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. FreeLong: Training-Free Long Video Generation with Spectral-Blend Temporal Attention, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.907, + 0.426 + ], + "angle": 0, + "content": "[38] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.48 + ], + "angle": 0, + "content": "[39] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent Diffusion Transformer for Video Generation, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[40] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINov2: Learning robust visual features without supervision. 2024, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[41] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[42] Zhiwu Qing, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yujie Wei, Yingya Zhang, Changxin Gao, and Nong Sang. Hierarchical spatio-temporal decoupling for text-to-video generation. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[43] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. FreeNoise: Tuning-free longer video diffusion via noise rescheduling. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15613" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21:140:1-140:67, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.191 + ], + "angle": 0, + "content": "[46] Jinghuan Shang, Karl Schmeckpeper, Brandon B. May, Maria Vittoria Minniti, Tarik Kelestemur, David Watkins, and Laura Herlant. Theia: Distilling Diverse Vision Foundation Models for Robot Learning, 2024. 2, 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[47] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1874-1883. IEEE Computer Society, 2016. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[48] Chenyang Si, Weichen Fan, Zhengyao Lv, Ziqi Huang, Yu Qiao, and Ziwei Liu. RepVideo: Rethinking Cross-Layer Representation for Video Generation, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[49] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[50] Jianlin Su, Murtadha H. M. Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 6, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[51] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 14398-14409. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[52] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[53] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and Efficient Foundation Language Models, 2023. 5, 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[55] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yulong Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin, Tiejun Huang, and Zhongyuan Wang. Emu3: Next-Token Prediction is All You Need, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[56] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, pages 1-20, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[57] Yuancheng Wang, Haoyue Zhan, Liwei Liu, Ruihong Zeng, Haotian Guo, Jiachen Zheng, Qiang Zhang, Xueyao Zhang, Shunsi Zhang, and Zhizheng Wu. Maskgct: Zero-shot text-to-speech with masked generative codec transformer. arXiv preprint arXiv:2409.00750, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[58] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video Generation using VQ-VAE and Transformers, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[59] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 10371-10381. IEEE, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[60] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer, 2024. 1, 2, 3, 5, 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[61] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[62] Lijun Yu, José Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G. Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A. Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[63] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An Image is Worth 32 Tokens for Reconstruction and Generation, 2024. 2, 4, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[64] Hangjie Yuan, Shiwei Zhang, Xiang Wang, Yujie Wei, Tao Feng, Yining Pan, Yingya Zhang, Ziwei Liu, Samuel Albanie, and Dong Ni. InstructVideo: Instructing video diffusion models with human feedback. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 6463-6474. IEEE, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[65] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make Pixels Dance: High-Dynamic Video Generation, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[66] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying Pixel and Latent Diffusion Models for Text-to-Video Generation, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[67] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15614" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "ICCV 2023, Paris, France, October 1-6, 2023, pages 3813-3824. IEEE, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[68] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.22 + ], + "angle": 0, + "content": "[69] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient Video Generation With Latent Diffusion Models, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15615" + } + ] +] \ No newline at end of file diff --git a/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/b1aafd05-162c-4fda-a00d-cb706731d3d7_origin.pdf b/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/b1aafd05-162c-4fda-a00d-cb706731d3d7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..327bbd2b6452a6e415faaedeff7cf41df220197d --- /dev/null +++ b/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/b1aafd05-162c-4fda-a00d-cb706731d3d7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3a4dffdc9595c8a790767b040e3fc5462cd5b481876912940ccc0969e4d19ab +size 8745118 diff --git a/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/full.md b/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1158996952c43276fb55c3212f0beed58e399199 --- /dev/null +++ b/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/full.md @@ -0,0 +1,288 @@ +# The Best of Both Worlds: Integrating Language Models and Diffusion Models for Video Generation + +Aoxiong Yin $^{1*}$ , Xu Tan $^{2*†}$ , Kai Shen $^{2*}$ , Yichong Leng $^{2}$ , Xinyu Zhou $^{2}$ , Juncheng Li $^{1†}$ , Siliang Tang $^{1}$ + +$^{1}$ Zhejiang University, $^{2}$ Moonshot AI + +yinaoxiong@zju.edu.cn, tanxu2012@gmail.com, junchengli@zju.edu.cn + +# Abstract + +Recent advancements in text-to-video (T2V) generation have been driven by two competing paradigms: autoregressive language models and diffusion models. However, each paradigm has intrinsic limitations: language models struggle with visual quality and error accumulation, while diffusion models lack semantic understanding and causal modeling. In this work, we propose LanDiff, a hybrid framework that synergizes the strengths of both paradigms through coarse-to-fine generation. Our architecture introduces three key innovations: (1) a semantic tokenizer that compresses 3D visual features into compact 1D discrete representations through efficient semantic compression, achieving $a \sim 14,000 \times$ compression ratio; (2) a language model that generates semantic tokens with high-level semantic relationships; (3) a streaming diffusion model that refines coarse semantics into high-fidelity videos. Experiments show that LanDiff, a 5B model, achieves a score of 85.43 on the VBenchmark T2V benchmark, surpassing the state-of-the-art open-source models Hunyuan Video (13B) and other commercial models such as Sora, Kling, and Hailuo. Furthermore, our model also achieves state-of-the-art performance in long video generation, surpassing other open-source models in this field. Our demo can be viewed at https://landiff.github.io/. + +# 1. Introduction + +Text-to-video (T2V) [6, 30, 55, 60] has made significant progress in recent years, becoming an important research direction in the fields of computer vision and artificial intelligence. Recent works in T2V models have primarily revolved around two predominant paradigms: autoregressive large language model (LLM)-based [30, 55] frameworks and diffusion-based architectures [6, 60]. However, each paradigm has their own advantages and limitations, as + +![](images/4de5e5549f22ffb211329d92de6ed929e5f0ad1bf570b367cd1192551fd840ef.jpg) +Figure 1. The rate-distortion curve illustrates how visual distortion decreases as the number of transmitted bits increases. With just a small number of bits representing high-level semantic features, we can already achieve relatively low visual distortion. Building on this information-theoretic insight, LanDiff combines the strengths of both paradigms: LLMs efficiently generate compact semantic features in the first stage, followed by diffusion models that add perceptual details in the second stage, before final decoding to pixels via VAE. Data from Ho et al. [20], illustration is conceptual. + +# shown in Table 1. + +From a representation perspective, LLM-based methods leverage discrete tokenization to explicitly encode high-level semantics through vector quantization, effectively prioritizing conceptual abstraction and narrative coherence. However, this discretization inherently sacrifices low-level visual fidelity due to information compression, resulting in low reconstruction quality. In contrast, diffusion-based approaches employ continuous latent representations to preserve much more perceptual details, enabling superior reconstruction quality at the cost of diluted semantic interpretability, as hierarchical features remain entangled in the latent space. From a generative modeling perspective, LLM-based systems adopt autoregressive modeling to enforce causal dependencies between video frames, ensuring strong temporal coherence. However, the autoregressive generation inherently risks error propagation across time steps, where inaccuracies in early predictions amplify dur + +Table 1. The comparison between LLM based and diffusion based T2V systems. The advantages and disadvantages are marked by $\text{喜}$ and $\text{喜}$ , respectively. + +
MethodsRepresentationsModeling
LLMDiscrete TokensAutoregressvie
Low Reconstruction QualityNo Refinement
Highlight Semantic InformationCausal Modeling
DiffusionContinuous VectorsNon-Autoregressvie
High Reconstruction QualityProgressive Refinement
Lack Semantic InformationNon-causal Modeling
+ +ing decoding. In contrast, diffusion-based methods employ non-autoregressive generation, refining outputs in parallel through iterative denoising steps. Although this design mitigates error accumulation and enhances generation flexibility, the absence of explicit causal constraints often leads to temporal inconsistencies or semantic hallucinations. + +In this work, we propose a hybrid architecture that synergizes the strengths of both Language models and Diffusion models, named LanDiff, through a coarse-to-fine generation paradigm. As shown in Figure 1, inspired by the human creation of video which will generate the high-level storyline first and then add low-level visual details based on the storyline to form the video, we design a two-stage video generation process with the number of bits gradually increasing and carefully design the autoregressive model and the diffusion model to be responsible for different stages of T2V generation, so as to play to their strengths and avoid their weaknesses. In detail, 1) at the low-bit position "semantic feature", the low-bit information ensures that the token sequence is not too long, and the high-level information makes it easier for the model to capture the overall semantic entity motion of the video, so as to play to the strengths of the autoregressive model and avoid its weaknesses. Thus we use LLM to generate a coarse-grained video in the first stage; 2) at the high-bit position "perceptual feature", since we have already obtained the coarse-grained with rich semantic and time-serial information, we only need to focus on how to add details to the coarse-grained video. Thus we apply a diffusion model in the second stage. Finally, a VAE decoder transforms the generated "perceptual feature" into the final RGB video output. By unifying these complementary mechanisms, we demonstrate that hybrid architectures can overcome the inherent limitations of isolated approaches, enabling coherent, semantically faithful, and visually compelling video generation from textual descriptions, as shown in Table 2. + +With this design, the ideal semantic feature should contain high-level semantic information and motion information and only require a few bits to represent. We achieve this goal by performing extreme compression on video representations rich in high-level semantics. For video representation, we select the Theia model [46] as our visual rep + +Table 2. The comparison between LanDiff and previous largescale T2V systems. The compression rates of LLM-based and diffusion-based models are illustrated using VideoPoet [30] and CogVideoX [60] as examples, respectively. + +
ModelsLLMDiffusionLanDiff
Highlight Semantic Information?✓✗✓
Progressive Refinement?✗✓✓
Causal Modeling?✓✗✓
High Visual Quality?✗✓✓
Compression Ratio ↑~256~1024~14000
Long video Generation?✗✗✓
+ +resentation backbone, which has been distilled from multiple visual understanding and self-supervised representation models, ensuring the encoded features contain rich semantic information. To achieve extreme compression and reduce the number of bits, we design an efficient tokenizer to compress 3D visual features into 1D discrete representations. The tokenizer is based on the Transformer [54] structure, uses query embedding to aggregate visual features, and has a higher compression rate than CNN-based structures [63]. To further compress the video by fully utilizing the temporal redundancy of the video, inspired by the MP4 video encoding algorithm [32], we divide the video into keyframes and non-keyframes, and set more numbers of tokens for keyframes. The detailed design is shown in subsection 3.1. For the diffusion model, we use generated semantic tokens as conditions and generate the target video by gradually removing the noises. To better support the long-video generation, we train a chunk-wise streaming diffusion model that only uses a limited number of historical frames as conditions, thereby greatly reducing the computational cost of training and inference. + +Thanks to these designs, our LanDiff has made significant progress in spatial relationship compliance, action coherence, visual quality, etc. Specifically, 1) for short video generation, our 5B model achieved a score of 85.43 on the VBenchmark, surpassing the state-of-the-art open-source models Hunyuan Video (13B) and other commercial models such as Sora, Kling, and Hailuo; 2) for long video generation, after testing by the VBenchmark T2V benchmark, our model also achieved state-of-the-art performance, surpassing other open-source models in this field. Our video examples can be viewed at https://landiff.github.io/. + +# 2. Related Work + +Video Tokenization. Video tokenization plays a crucial role in video understanding and generation tasks. Since a video can be represented as a sequence of continuous frames, some works directly use image tokenizers to process videos frame by frame. For example, Wang et al. [55] directly use SBER-MoVQGAN as the video tokenizer. + +However, this method ignores the temporal redundancy in videos, resulting in a low compression rate. To reduce temporal redundancy, some works [58, 62] try to extend image tokenizers based on 2D convolution to 3D convolution, which can process both temporal and spatial information simultaneously. These methods encode videos in the original RGB space and perform video reconstruction tasks, which are more about perceptual compression. In addition, some works [15, 25] try to train video tokenizers on features extracted from pre-trained visual encoders. These tokenizers can achieve good performance in understanding and generation tasks while maintaining a high compression rate. The video tokenizer we propose belongs to this category. Unlike previous feature-based tokenizers that achieved limited compression rates and primarily focused on image processing, our video tokenizer delivers significantly higher compression while handling both images and videos in a unified framework. + +LLM based Video Generation. LLM-based video generation methods are usually based on the Transformer [54] structure, learning the mapping from text to video through next token prediction. TATS [14] uses VQ-GAN as the tokenizer and predicts video tokens using a GPT-like model structure. VideoGPT [58] uses 3D convolution to extract features and quantize them, and predicts the quantized video discrete tokens using a GPT-like model. Recently, VideoPoet [62] uses MagViT2 [62] as the tokenizer and unifies multiple modalities as input to a large language model (LLM) to conditionally generate video tokens. In addition, Emu3 [55] uses SBER-MoVQGAN as the tokenizer to perform video understanding and generation by predicting the next token. These works all use LLMs to directly generate "perceptual features" that contain rich visual details with high bit rates. Recently, ARLON [33] attempts to discretize VAE features into a small number of tokens to reduce the bit rate required for LLMs' prediction. In this way, the tokens retain low-frequency visual information such as blurry contours rather than high-level semantic information. In contrast, our method employs tokens containing high-level semantic information as prediction targets for LLMs, which enables us to fully leverage LLMs' advantages in causal modeling to precisely generate high-quality videos. + +Diffusion based Video Generation. Diffusion-based methods have achieved great success in image generation, and recently many people [7, 21, 22, 49, 69] have tried to apply them to video generation tasks. VDM [22] extends the 3D U-Net structure for video generation. Wang et al. [56] propose to generate high-quality and aesthetically pleasing videos in a cascaded manner. Benefiting from the success of the text-to-image (T2I) field, some works such as Animatediff [16], SVD [6], and PixelDance [65] try to use pre-trained T2I models as initialization, and then add modules for temporal modeling to capture motion informa + +tion for video generation. Ma et al. [39] explore the generation capabilities of multiple different structures of latent diffusion transformer. After the release of SORA, a series of video generation methods based on the DiT [41] model have been proposed, including OpenSora [68], OpenSoraPlan [34], Cogvideox [60], Hunyuan Video [31], Mira [27] and STIV [35] etc. These methods can only generate short videos of a few seconds. Recently, StreamingT2V [18] generates long videos by block-wise generation on a pre-trained short video generation model, and then uniformly performs mixed augmentation. In addition, some works improve the consistency of long video generation by leveraging noise rescheduling techniques [37, 43]. Our method employs diffusion models as renderers for semantic features, enabling us to leverage their superior visual generation quality while circumventing their limitations in causal modeling. + +# 3. Method + +In this work, we propose a novel text-to-video generation framework that synergistically integrates the strengths of autoregressive modeling and diffusion processes while circumventing their respective limitations. The framework mainly consists of the following components: 1) an efficient tokenizer that transforms 3D visual features into compact 1D discrete representations while preserving and enhancing their semantic information; 2) a language model that performs temporal sequence modeling to generate semantic tokens representing video blueprints from textual descriptions; 3) a streaming diffusion model that progressively refines coarse semantic videos by adding fine-grained details to produce high-quality VAE features; and 4) a VAE decoder that reconstructs the final video frames from the refined VAE features. + +# 3.1. Video Semantic Tokenizer + +In this section, we introduce a novel video semantic tokenizer that efficiently compresses the video into semantic information. Firstly, we will introduce the video semantic representation used in tokenization. Then, considering the high redundancy of video in both spatial and temporal dimensions, we introduce two compression strategies: 1) a query-based causal tokenization that efficiently reduces spatial redundancy while preserving essential semantic information through vector quantization; 2) inspired by MP4 [32], we implement video frame grouping to minimize temporal redundancy by treating grouped frames as a unit, where the first frame (I-frame) is fully encoded while subsequent frames (P-frames) only capture the temporal changes by referencing content from previous frames. + +Video Semantic Representation. Generally, video representations can be divided into two categories: 1) some methods [55, 62] directly utilize an autoencoder to learn the video representations. 2) Some works [25, 26, 29, 51] use + +![](images/bca7e08238517ff87e8b2fcc62116f2dd4eff394178801b9aed0b7c24836e234.jpg) +Figure 2. The architecture of LanDiff. Given text inputs, we first extract text embeddings and employ an LLM to generate semantic tokens in the first stage. Subsequently, we utilize a diffusion model to synthesize perceptual features conditioned on these semantic tokens, followed by a VAE decoder that transforms these features into the final video frames. + +pre-trained visual self-supervised learning features (SSL) as video representations. Compared with the first directly learned autoencoder latents which contain lots of visual details, the second SSL features maintain much more semantic features, which enable the LLM to focus more on the high-level semantic information of the video. + +Based on these thoughts, we choose the pretrained SSL features as the video representations. For better usage of different SSL models, we choose Theia model [46], which is a unified visual feature extractor distilled from multiple visual task models, including CLIP [44], SAM [28], DINOv2 [40], ViT [10], and Depth-Anything [59]. + +Tokenizer Design. In this part, we introduce the design of our tokenizer, which leverages the query tokens to compress the video semantic features and uses quantization to discrete the video semantic representation while minimizing the reconstruction loss. + +In detail, firstly, we extract the semantic features using the Theia model and flatten it to obtain $F \in \mathbb{R}^{(T \times H \times W) \times D}$ . Inspired by TiTok [63], we use N randomly initialized tokens as query tokens $Q \in \mathbb{R}^{N \times D}$ and concatenate them with the semantic features $F$ . Then we use a transformer encoder to encode them and only take the encoded features of the query tokens. + +$$ +Z _ {Q} = \operatorname {E n c} ([ F; Q ]), \tag {1} +$$ + +where $[\cdot ]$ represents the concatenation operation and $Z_{Q}\in$ $\mathbb{R}^{N\times D}$ . We then apply vector quantization on $Z_{Q}$ by train a VQ-VAE model and obtain the quantized feature $\hat{Z}_Q$ . In the decoding stage, the quantized feature $\hat{Z}_Q$ is used as a condition, and then a sequence of mask tokens $M\in \mathbb{R}^{T\times H\times W}$ are added in front of the $\hat{Z}_Q$ to form the inputs of decoder. Then we only take the features of the mask tokens as fol + +lows: + +$$ +\hat {F} = \operatorname {D e c} ([ M; \hat {Z} _ {Q} ]), \tag {2} +$$ + +where $\hat{F} \in \mathbb{R}^{(T \times H \times W) \times D}$ represents the reconstructed feature, and $M$ represents the mask tokens. Inspired by Huang et al. [23], Wang et al. [57], we minimize the reconstruction loss of the video semantic feature during training the VQ-VAE. + +For the VQ-VAE, we follow the method of Wang et al. [57], Yu et al. [61]. We update the codebook using exponential moving average (EMA). We also optimize the codebook with the video semantic feature reconstruction loss. The loss function is shown as follows: + +$$ +\mathcal {L} = \lambda_ {r e c} \| \hat {F} - F \| _ {2} + \lambda_ {c o m m i t} \| \operatorname {s g} (\hat {Z} _ {Q}) - Z _ {Q} \| _ {2}, \tag {3} +$$ + +where $\mathrm{sg}()$ is the stop-gradient operation, $||\cdot ||_2$ is the L2 loss. + +Video Frame Grouping. For video, a straightforward observation is the redundancy in time series (i.e., the difference between adjacent video frames is minimal). Intuitively, we can achieve a better compression rate by modeling the difference between adjacent video frames instead of tokenizing all frames equally. Inspired by the popular video compression method MP4 [32], given $N$ frames of a video, we first split them into $N / T$ groups (i.e., $T$ frames as a group. For clarification, as shown in Figure 3, we use $T = 3$ as an example.). Then: 1) we will model different groups independently; 2) for each group, we will fully encode the first frame (Intra-coded Frame, IFrame), while for the remaining $[1,T - 1]$ frames (the Predictive-coded Frame, PFrame), we encode them by referencing their previous frames. + +To achieve this, within a group, the first frame (IFrame) will only see itself and have a large number of query tokens + +![](images/0255745df1d25767370eb3b425067e43c428c0dffbef7d7588a2b1b3cab18b30.jpg) + +![](images/db8d53fd9c435c1e0f79d208bb4eb8b78c66c47877ce383b61fc231d9483919c.jpg) +Figure 3. Proposed architecture of the video semantic tokenizer. We use query tokens to compress the semantic sequence length. Furthermore, we group the frames into groups (3 frames in a group in this figure). In a group, the first frame is the IFrame and the rest frames are PFrames. We use different query token numbers for them. The attention mask design is shown in the right. + +(e.g., 3 query tokens in Figure 3) to achieve better reconstruction quality. For the rest frame (PFrame) $i \in [1, T - 1]$ , it will see the previous frames (i.e., frames $j \in [0, i - 1]$ ) and have a small number of query tokens (e.g., 1 query token which is $1/3$ of the IFrame) to force the model to learn the difference. Technically, as shown in the attention mask on the right side of Figure 3, we use frame-level causal masks for the feature sequence during encoding. The query tokens are also divided according to the frame, and each token can only attend to the features of the corresponding frame and the previous frames. During decoding, the mask token corresponding to each frame can see the previous features, the corresponding query tokens, and the previous query tokens. + +# 3.2. Language Model for Semantic Token Generation + +As shown in Figure 2, after training an efficient tokenizer, we use a language model to generate semantic tokens autoregressively based on text. Specifically, we first use a pretrained text encoder T5-XXL [45] to extract text features $X$ . We use the video tokenizer encoder trained in subsection 3.1 to convert the video into a 1D discrete token sequence $Y$ . To enhance the controllability of the generated results, we introduce additional control conditions $CC$ . It includes: 1) frames condition for requiring the model to generate videos with a specified number of frames; 2) motion score condition, which is a value between 0 and 1, used to control the degree of motion in the generated video. The discrete tokens are converted into embedding vectors during generation and added with positional encoding, and then concatenated with the control conditions as input. The structure of + +the language model follows the typical LLaMA model [53] network structure. We train the model from scratch, using cross-entropy loss as the loss function. + +$$ +\mathcal {L} _ {\mathrm {L M}} = \mathbb {E} \left[ - \log p \left(Y _ {i} \mid X, C C, Y _ {< i}\right) \right] \tag {4} +$$ + +# 3.3. Diffusion Model for Perceptual Feature Generation + +The video detokenizer in LanDiff is responsible for converting semantic tokens into VAE latent vectors. As shown in Figure 2 we use a conditional diffusion model to complete this task. Especially to support long-video generation, we also design the streaming inference strategy. In this section, we will first introduce the architecture for a single semantic token chunk. Then we will introduce the chunk-wise streaming strategy. + +Architecture. Our diffusion model employs an architecture similar to MMDiT[12, 60]. Specifically, 1) we use the video tokenizer decoder trained in subsection 3.1 to decode the semantic tokens into semantic features $\hat{F}$ . Then we use the semantic features $\hat{F}$ as a condition to guide the diffusion model to generate videos; 2) we inject the control signals into the model in a similar way to ControlNet [67]. In detail, during training, the parameters of the main model are not updated, the control module copies the parameters of the first half layers of the main model, and adds them to the output of the main model after a linear layer initialized with zeros; 3) to make the semantic features match the target VAE features in the space dimension, we additionally add an upsampling module. We show the model structure in the appendix Figure 8. + +During training, given the input video chunk $V \in \mathbb{R}^{(T \times H \times W) \times D}$ , where $T$ is the frame length of the VAE latent and $D$ is the VAE latent feature dimension, the diffusion algorithm progressively adds noise to the video and produces a noisy video $V^t$ , where $t$ represents the number of times noise is added. The diffusion algorithm uses time step $t$ , and semantic features $\hat{F} \in \mathbb{R}^{(T \times H \times W) \times D}$ as conditions, and then uses a network $\epsilon_{\theta}$ to predict the noise added to the noisy video $V_t$ through: + +$$ +\mathcal {L} = \mathbb {E} _ {V, t, c _ {s}, \epsilon \sim \mathcal {N} (0, 1)} \left[ \| \epsilon - \epsilon_ {\theta} \left(V ^ {t}, t, \hat {F}\right) \| _ {2} ^ {2} \right] \tag {5} +$$ + +Chunk-wise Streaming Strategy. To support long video generation scenario whose semantic token sequence is very long and difficult to generate as a whole, we propose the chunk-wise streaming strategy. During training, given video latent chunk $V \in \mathbb{R}^{(T \times H \times W) \times D}$ , to maintain the appearance continuity in the video, we use the first half of $V(V_{l})$ as the prompt and generate the second half of $V_{r}$ . In detail, we do not add noise to the first half chunk and give time condition $t$ to 0.999. We also randomly mask the first half chunk with a ratio of $20\%$ . The loss function for training at this time is: + +$$ +\mathcal {L} = \mathbb {E} _ {V _ {r}, t, c _ {s}, \epsilon \sim \mathcal {N} (0, 1)} \left[ \| \epsilon - \epsilon_ {\theta} \left(V _ {l}, V _ {r} ^ {t}, t, c _ {s}\right) \| _ {2} ^ {2} \right]. \tag {6} +$$ + +During inference, the first $L / 2$ VAE latents will be generated without the prompt. For the following VAE latents, we will accumulate $L / 2$ tokens to form a chunk and use the previous chunk as prompt. + +# 4. Experiments and Results + +# 4.1. Experimental Settings + +Datasets. For the video tokenizer and language model, we use an internal dataset with 200M video-text pairs for training. All videos with a duration of less than 6s are filtered, and the videos are kept in the aspect ratio and then scaled to around 480x720 resolution for center cropping. Consistent with CogVideoX [60], we set the fps of all videos to 8. For the diffusion model, we select a dataset containing 3M high-quality video-text pairs for training. To evaluate the performance of the text-to-video generation model, we use the prompts provided by the widely used VBench [24] T2V benchmark to generate videos. + +Implementation Details. For the video tokenizer, we set the group size $T$ mentioned in subsection 3.1 to 13. We set the number of tokens $I$ corresponding to the IFrame to 330, and the number of tokens $P$ corresponding to the PFrame to 74. The dimension of the tokenizer's codebook is set to 16, and the vocabulary size is set to 2048. For the language model, we use the LLaMA [53] structure, with 2B model parameters, and use 1D RoPE [50] positional encoding. We use the 2B version of the CogVideoX [60] model as the + +base model for our video detokenizer. We list the structure settings and more implementation details of each module in the appendix subsection A.1, subsection A.2, and subsection A.3. + +Evaluation Metrics. To evaluate the text-to-video generation task, we use the metrics proposed in the VBench and VBench-Long [24] benchmarks. + +Evaluation Baselines. We compare LanDiff with baselines: 1) Sora [5]. 2) Jimeng [3]. 3) Hailuo [2]. 4) OpenSoraPlan V1.1 [34]. 5) Kling [38]. 6) InstructVideo [64]. 7) Gen-3 [1]. 9) Latte-1 [39] 9) HiGen [42]. 10)AnimateDiffV2 [16]. 11) Show-1 [66]. 12) Pika [4]. 13) VideoCrafter2.0 [7]. 14) OpenSora V1.2 [68]. 15) LTX-Video [17]. 16) Mochi-1 [52]. 17) CogVideoX [60]. 18) Vchitect-2.0 [13]. 19) RepVideo [48]. 20) HunyuanVideo [31]. 21) ARLON [33]. 22) Emu3 [55]. + +# 4.2. Experimental Results + +A life-sized ice sculpture of a playful dog, with intricate details and a joyful expression, stands in the middle of a sunlit, grassy field on a sweltering summer day. The ice dog, initially solid and vibrant, begins to melt under the relentless heat, with droplets of water forming on its surface. As the day progresses, the ice dog's form gradually diminishes, with its once sharp features becoming blurred and distorted. The melting process accelerates, and the ice dog's body starts to collapse, pooling into a puddle of water on the ground. By the end of the day, all that remains is a shallow puddle, reflecting the cloudless sky, with the memory of the once majestic ice dog now just a memory. + +![](images/df4b8ae44b75e13b80b941287681438324475d9ad202ba0dd8c7b1edd40b4ced.jpg) +Figure 4. Comparison of qualitative results for text-to-video generation. + +![](images/e8f2da8dfaeb6db258ae37e51ff71724171ceae3be3a1e4be6921efaf6da66e3.jpg) + +![](images/5d1b7e3b8e30e606bc38e7fc2b50a06b74b610c43d9cd413e1d6fc2a6b65018c.jpg) + +![](images/35b80902c0c1627075823259dbcf504b84b22689e9f62362fc80fa09b7946671.jpg) + +Text to Video Generation. As shown in Table 3, our 5B LanDiff model outperforms the state-of-the-art open-source models on VBenchmark, including the 13B HunyuanVideo, achieving a top score of 85.42. LanDiff exhibits superior semantic understanding while maintaining excellent generation quality, surpassing same-sized CogVideoX-5B across most metrics. To validate our approach's effectiveness independent of training data volume, we trained a 7B DiT-based diffusion model with identical training data (marked as DiT in the table). The results confirm that our task division strategy successfully leverages the complementary strengths of autoregressive and diffusion models. Figure 4 shows qualitative comparisons between LanDiff and CogVideoX-5B. Our model effectively captures temporal dynamics described in the prompt, accurately rendering a melting ice dog sculpture. LanDiff successfully depicts the progressive transformation from solid form to liquid state, while CogVideoX-5B produces a static ice sculpture that remains virtually unchanged throughout the sequence. This demonstrates our model's superior ability to compre + +Table 3. Performance comparison of Text-to-video (T2V) generation between our LanDiff and other state-of-the-art models on VBench benchmark. We selected 8 out of the 16 evaluation dimensions from VBench, along with Total Score, Quality Score, and Semantic Score, for presentation. The complete results with all 16 evaluation dimensions can be found in the appendix Table 8. The best and second-best scores are highlighted in bold and underline, respectively. $\dagger$ indicates the scores we reproduced, while $\ddagger$ indicates the scores from the original papers, and other scores are from the VBench benchmark. + +
Model NameTypeModel SizeTotal ScoreQuality ScoreSemantic Scorebackground consistencydynamic degreemotion smoothnessmultiple objectsobject classscenespatial relationshipsubject consistency
Open Sourced Models
InstructVideoDiffusion1.3B76.6181.5656.8196.9769.7296.6221.5773.2622.2143.4995.30
Latte-1Diffusion0.7B77.2979.7267.5895.4068.8994.6334.5386.5336.2641.5388.88
OpenSoraPlan V1.1Diffusion2.7B78.0080.9166.3896.7347.7298.2840.3576.3027.1753.1195.73
Show-1Diffusion6.3B78.9380.4272.9898.0244.4498.2445.4793.0747.0353.5095.53
OpenSora V1.2Diffusion1.1B79.7681.3573.3997.6142.3998.5051.8382.2242.4468.5696.75
LTX-VideoDiffusion1.9B80.0082.3070.7997.2054.3598.9645.4383.4551.0765.4396.56
Mochi-1Diffusion10B80.1382.6470.0897.2861.8599.0250.4786.5136.9969.2496.99
AnimateDiff-V2Diffusion1.3B80.2782.9069.7597.6840.8397.7636.8890.9050.1934.6095.30
VideoCrafter-2.0Diffusion1.7B80.4482.2073.4298.2242.5097.7340.6692.5555.2935.8696.85
CogVideoX-2BDiffusion2B80.9182.1875.8396.6359.8697.7362.6383.3751.1469.9096.78
Emu3 ‡LLM8B80.96N/AN/A97.6979.2798.9344.6486.1737.1168.7395.32
Vchitect-2.0-2BDiffusion2B81.5782.5177.7996.5358.3397.7669.3587.8157.5154.6496.42
CogVideoX-5BDiffusion5B81.6182.7577.0496.5270.9796.9262.1185.2353.2066.3596.23
DiT †Diffusion7B81.8582.7078.4297.6551.0298.9469.6393.4057.1060.8997.05
RepVideoDiffusion2B81.9482.7078.9196.5657.7898.1371.1887.8352.9674.7496.25
Vchitect-2.0[E]Diffusion2B82.2483.5477.0696.6663.8998.9868.8486.6156.5757.5596.83
HunyuanVideoDiffusion13B83.2485.0975.8297.7670.8398.9968.5586.1053.8868.6897.37
Close Sourced Models
Pika-1.0DiffusionN/A80.6982.9271.7797.3647.5099.5043.0888.7249.8361.0396.94
KlingDiffusionN/A81.8583.3975.6897.6046.9499.4068.0587.2450.8673.0398.33
JimengDiffusionN/A81.9783.2976.6998.3938.4398.0969.0889.6244.9477.4597.25
Gen-3DiffusionN/A82.3284.1175.1796.6260.1499.2353.6487.8154.5765.0997.10
HailuoDiffusionN/A83.4184.8577.6597.0564.9199.2276.0487.8350.6875.5097.51
SoraDiffusionN/A84.2885.5179.3596.3579.9198.7470.8593.9356.9574.2996.23
ARLON ‡LLM+Diffusion1.5BN/AN/AN/A97.1052.7798.92N/A89.8054.43N/A93.41
ARLON †LLM+Diffusion5B82.3183.5877.2798.6572.2297.5674.4990.6052.0762.5395.59
LanDiffLLM+Diffusion5B85.4386.1382.6198.7392.7197.0886.6994.9453.7973.7496.11
+ +hend and visualize complex temporal transformations and physical processes. For more video examples, please refer to our demo website2. + +Long Video Generation. We compare our model with other open-source text-to-long video generation models: FreeNoise [43], StreamingT2V [18], OpenSora-V1.2 [68], and ARLON [33]. As shown in Table 4, LanDiff achieves state-of-the-art performance with the highest Total Score of 68.34. While StreamingT2V shows superior Dynamic Degree (85.64), it sacrifices consistency metrics (Subject: 87.31, Background: 94.64). Conversely, models like ARLON excel in consistency but exhibit limited dynamism (50.42). Our approach strikes an optimal balance with strong dynamism (72.86) while maintaining high consistency scores and achieving the best results in Aesthetic Quality (60.96) and Overall Consistency (27.29). As illustrated in Figure 5, LanDiff effectively renders complex motion dynamics in the prompt "a car accelerating to gain speed," successfully depicting increasing velocity while preserving visual consistency. In contrast, competing models struggle with this motion, producing either static + +vehicles, inconsistent appearances, or unrealistic acceleration patterns. Additional video samples are available on our demo site2. + +Video Tokenization. We present visualization results of our model's video reconstruction in Figure 6. We extract semantic tokens from the reference video and subsequently convert these tokens back into video using the video detokenizer. The results demonstrate that through our careful design, our video tokenizer can accurately reconstruct videos with accurate semantics and actions using only approximately 1/50th of the sequence length required by MagViT2 [62]. While some minor discrepancies in clothing details exist between the reconstructed and reference videos, these differences remain within acceptable limits for practical video generation applications. Additional video samples are available on our demo site2. + +# 4.3. Ablation Study + +In this section, we conducted ablation experiments on the video tokenizer and classifier-free guidance. To fairly evaluate the effectiveness of our video tokenizer, we implemented an ARLON-like method as a comparison baseline. Specifically, for the control group without our proposed + +Table 4. Long video generation results of LanDiff and other models on VBench. The best and second-best scores are highlighted in bold and underline, respectively. + +![](images/2e29716b0accdd512fa71a3676baa79660ee3bf6c05253ac7b973f9077370eed.jpg) +Figure 5. Qualitative comparison of long video generation results between LanDiff and other state-of-the-art models (FreeNoise, StreamingT2V and OpenSora-V1.2). + +![](images/24fe78fd9647fa699f7cde4132ec1be689b4e123c0d7cef6543f468d325716a6.jpg) +Figure 6. Visualization results of video reconstruction using video tokenizer. + +Table 5. The ablation study of video tokenizer and classifier free guidance (cfg) on Vbench benchmark. + +
ModelsTotal ScoreQuality ScoreSemantic Score
LanDiff85.4386.1382.61
-video tokenizer82.3183.5877.27
-cfg81.0683.0473.13
+ +video tokenizer, we followed ARLON's architectural approach by training a large language model to predict quantized VAE features, then using CogVideoX-5B as the base diffusion model conditioned on these LLM-predicted VAE + +features to generate videos. This baseline model was trained with exactly the same dataset and training recipe as Lan-Diff, ensuring strict experimental control variables and effectively eliminating interference factors such as model parameter count and training data scale. As shown in Table 5, the experimental results clearly demonstrate that our proposed video tokenizer significantly improves both the quality score and semantic score of generated videos, validating the importance of this component in our framework. In the case of not using classifier-free guidance for text, our model has decreased in quality and semantic scores. + +# 5. Conclusion + +In this paper, we propose a new text-to-video generation model, LanDiff. It combines the advantages of autoregressive models and diffusion models, including: 1) an efficient 1D video tokenizer to extract videos into semantic tokens; 2) a language model to generate semantic tokens based on text; 3) a video detokenizer to convert semantic tokens into videos. LanDiff outperforms the state-of-the-art open-source models on the VBenchmark, surpassing other open-source models in quality and semantic scores. At the same time, LanDiff also achieves state-of-the-art performance in long video generation tasks. + +# 6. Acknowledgments + +This work was supported by the NSFC (62272411), the Key R&D Projects in Zhejiang Province (No. 2024C01106, 2025C01030), the Zhejiang NSF (LRG25F020001), Zhejiang University Education Foundation Qizhen Scholar Foundation. This work was also supported by the Moonshot AI. + +# References + +[1] Gen-3. https://runwayml.com/research/introducing-gen-3-alpha, 2024.6 +[2] Hailuo. https://platform.minimaxi.com/, 2024.6 +[3] Jimeng. https://jimeng.jianying.com/, 2024.6 +[4] Pika 1.0. https://www.pika.art/, 2024.6 +[5] Openai sora. https://openai.com/index/sora, 2024.6 +[6] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets, 2023. 1, 3 +[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 7310-7320. IEEE, 2024. 3, 6 +[8] Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High Fidelity Neural Audio Compression, 2022. 14 +[9] Juechu Dong, Boyuan Feng, Driss Guessous, Yanbo Liang, and Horace He. Flex Attention: A Programming Model for Generating Optimized Attention Kernels, 2024. 14 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 4 +[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming Transformers for High-Resolution Image Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873–12883, 2021. 15 +[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. Open-Review.net, 2024. 5 +[13] Weichen Fan, Chenyang Si, Junhao Song, Zhenyu Yang, Yinan He, Long Zhuo, Ziqi Huang, Ziyue Dong, Jingwen He, + +Dongwei Pan, Yi Wang, Yuming Jiang, Yaohui Wang, Peng Gao, Xinyuan Chen, Hengjie Li, Dahua Lin, Yu Qiao, and Ziwei Liu. Vchitect-2.0: Parallel Transformer for Scaling Up Video Diffusion Models, 2025. 6 +[14] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long Video Generation with Time-Agnostic VQGAN and Time-Sensitive Transformer, 2022. 3 +[15] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a SEED of Vision in Large Language Model, 2023. 3 +[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3, 6 +[17] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. LTX-Video: Realtime Video Latent Diffusion, 2024. 6 +[18] Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text, 2024. 3, 7 +[19] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance, 2022. 15 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 1 +[21] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. CoRR, abs/2210.02303, 2022. 3 +[22] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models, 2022. 3 +[23] Zhichao Huang, Chutong Meng, and Tom Ko. Repocodec: A speech representation codec for speech tokenization. arXiv preprint arXiv:2309.00169, 2023. 4 +[24] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21807-21818. IEEE, 2024. 6 +[25] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jian-choao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu, + +Chengru Song, Xiaogiang Lei, Di Zhang, Wenwu Ou, Kun Gai, and Yadong Mu. Unified Language-Vision Pretraining in LLM with Dynamic Discrete Visual Tokenization, 2023. 3 +[26] Yang Jin, Zhicheng Sun, Kun Xu, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, Kun Kai, and Yadong Mu. Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization, 2024. 3 +[27] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. MiraData: A Large-Scale Video Dataset with Long Durations and Structured Captions, 2024. 3 +[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment Anything, 2023. 4 +[29] Jing Yu Koh, Daniel Fried, and Russ Salakhutdinov. Generating images with multimodal language models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 3 +[30] Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, Krishna Somandepalli, Hassan Akbari, Yair Alon, Yong Cheng, Joshua V. Dillon, Agrim Gupta, Meera Hahn, Anja Hauth, David Hendon, Alonso Martinez, David Minnen, Mikhail Sirotenko, Kihyuk Sohn, Xuan Yang, Hartwig Adam, Ming-Hsuan Yang, Irfan Essa, Huisheng Wang, David A. Ross, Bryan Seybold, and Lu Jiang. VideoPoet: A large language model for zero-shot video generation. In *Forty-First International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 1, 2 +[31] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Daquan Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. HunyuanVideo: A Systematic Framework For Large Video Generative Models, 2024. 3, 6 +[32] Didier Le Gall. MPEG: A video compression standard for multimedia applications. Communications of the ACM, 34 (4):46-58, 1991. 2, 3, 4 +[33] Zongyi Li, Shujie Hu, Shujie Liu, Long Zhou, Jeongsoo Choi, Lingwei Meng, Xun Guo, Jinyu Li, Hefei Ling, and Furu Wei. ARLON: Boosting Diffusion Transformers with Autoregressive Models for Long Video Generation, 2024. 3, 6, 7 + +[34] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, Tanghui Jia, Junwu Zhang, Zhenyu Tang, Yatian Pang, Bin She, Cen Yan, Zhiheng Hu, Xiaoyi Dong, Lin Chen, Zhang Pan, Xing Zhou, Shaoling Dong, Yonghong Tian, and Li Yuan. Open-Sora Plan: Open-Source Large Video Generation Model, 2024. 3, 6 +[35] Zongyu Lin, Wei Liu, Chen Chen, Jiasen Lu, Wenze Hu, Tsu-Jui Fu, Jesse Allardice, Zhengfeng Lai, Liangchen Song, Bowen Zhang, Cha Chen, Yiran Fei, Yifan Jiang, Lezhi Li, Yizhou Sun, Kai-Wei Chang, and Yinfei Yang. STIV: Scalable Text and Image Conditioned Video Generation, 2024. 3 +[36] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 14 +[37] Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. FreeLong: Training-Free Long Video Generation with Spectral-Blend Temporal Attention, 2024. 3 +[38] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6 +[39] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent Diffusion Transformer for Video Generation, 2024. 3, 6 +[40] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINov2: Learning robust visual features without supervision. 2024, 2024. 4 +[41] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers, 2023. 3 +[42] Zhiwu Qing, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yujie Wei, Yingya Zhang, Changxin Gao, and Nong Sang. Hierarchical spatio-temporal decoupling for text-to-video generation. In CVPR, 2024. 6 +[43] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. FreeNoise: Tuning-free longer video diffusion via noise rescheduling. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3, 7 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 4 +[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and + +Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21:140:1-140:67, 2020. 5 +[46] Jinghuan Shang, Karl Schmeckpeper, Brandon B. May, Maria Vittoria Minniti, Tarik Kelestemur, David Watkins, and Laura Herlant. Theia: Distilling Diverse Vision Foundation Models for Robot Learning, 2024. 2, 4, 13 +[47] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1874-1883. IEEE Computer Society, 2016. 15 +[48] Chenyang Si, Weichen Fan, Zhengyao Lv, Ziqi Huang, Yu Qiao, and Ziwei Liu. RepVideo: Rethinking Cross-Layer Representation for Video Generation, 2025. 6 +[49] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3 +[50] Jianlin Su, Murtadha H. M. Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 6, 13 +[51] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 14398-14409. IEEE, 2024. 3 +[52] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024.6 +[53] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and Efficient Foundation Language Models, 2023. 5, 6, 15 +[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3 +[55] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yulong Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin, Tiejun Huang, and Zhongyuan Wang. Emu3: Next-Token Prediction is All You Need, 2024. 1, 2, 3, 6 +[56] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, + +Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, pages 1-20, 2024. 3 +[57] Yuancheng Wang, Haoyue Zhan, Liwei Liu, Ruihong Zeng, Haotian Guo, Jiachen Zheng, Qiang Zhang, Xueyao Zhang, Shunsi Zhang, and Zhizheng Wu. Maskgct: Zero-shot text-to-speech with masked generative codec transformer. arXiv preprint arXiv:2409.00750, 2024. 4 +[58] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video Generation using VQ-VAE and Transformers, 2021. 3 +[59] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 10371-10381. IEEE, 2024. 4 +[60] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer, 2024. 1, 2, 3, 5, 6, 15 +[61] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 4 +[62] Lijun Yu, José Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G. Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A. Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3, 7, 13 +[63] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An Image is Worth 32 Tokens for Reconstruction and Generation, 2024. 2, 4, 13 +[64] Hangjie Yuan, Shiwei Zhang, Xiang Wang, Yujie Wei, Tao Feng, Yining Pan, Yingya Zhang, Ziwei Liu, Samuel Albanie, and Dong Ni. InstructVideo: Instructing video diffusion models with human feedback. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 6463-6474. IEEE, 2024. 6 +[65] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make Pixels Dance: High-Dynamic Video Generation, 2023. 3 +[66] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying Pixel and Latent Diffusion Models for Text-to-Video Generation, 2023. 6 +[67] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision, + +ICCV 2023, Paris, France, October 1-6, 2023, pages 3813-3824. IEEE, 2023. 5 +[68] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 3, 6, 7 +[69] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient Video Generation With Latent Diffusion Models, 2023. 3 \ No newline at end of file diff --git a/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/images.zip b/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..47acf689f76949f95272e0167487da36560749e5 --- /dev/null +++ b/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57c5c6a7af4fa5179426e64c28e40246fffe4867e403cf09059aea5ec75c8ac5 +size 659320 diff --git a/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/layout.json b/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ce6db48a5955328beba8031e95f8b18423216446 --- /dev/null +++ b/2025/The Best of Both Worlds_ Integrating Language Models and Diffusion Models for Video Generation/layout.json @@ -0,0 +1,7973 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 545, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 545, + 138 + ], + "type": "text", + "content": "The Best of Both Worlds: Integrating Language Models and Diffusion Models for Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "text", + "content": "Aoxiong Yin" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "text", + "content": ", Xu Tan" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "inline_equation", + "content": "^{2*†}" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "text", + "content": ", Kai Shen" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "text", + "content": ", Yichong Leng" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "text", + "content": ", Xinyu Zhou" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "text", + "content": ", Juncheng Li" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "inline_equation", + "content": "^{1†}" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "text", + "content": ", Siliang Tang" + }, + { + "bbox": [ + 64, + 160, + 545, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 239, + 177, + 372, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 177, + 372, + 189 + ], + "spans": [ + { + "bbox": [ + 239, + 177, + 372, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 239, + 177, + 372, + 189 + ], + "type": "text", + "content": "Zhejiang University, " + }, + { + "bbox": [ + 239, + 177, + 372, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 239, + 177, + 372, + 189 + ], + "type": "text", + "content": "Moonshot AI" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 191, + 478, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 191, + 478, + 202 + ], + "spans": [ + { + "bbox": [ + 125, + 191, + 478, + 202 + ], + "type": "text", + "content": "yinaoxiong@zju.edu.cn, tanxu2012@gmail.com, junchengli@zju.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 256, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 256, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 297, + 544 + ], + "type": "text", + "content": "Recent advancements in text-to-video (T2V) generation have been driven by two competing paradigms: autoregressive language models and diffusion models. However, each paradigm has intrinsic limitations: language models struggle with visual quality and error accumulation, while diffusion models lack semantic understanding and causal modeling. In this work, we propose LanDiff, a hybrid framework that synergizes the strengths of both paradigms through coarse-to-fine generation. Our architecture introduces three key innovations: (1) a semantic tokenizer that compresses 3D visual features into compact 1D discrete representations through efficient semantic compression, achieving " + }, + { + "bbox": [ + 54, + 256, + 297, + 544 + ], + "type": "inline_equation", + "content": "a \\sim 14,000 \\times" + }, + { + "bbox": [ + 54, + 256, + 297, + 544 + ], + "type": "text", + "content": " compression ratio; (2) a language model that generates semantic tokens with high-level semantic relationships; (3) a streaming diffusion model that refines coarse semantics into high-fidelity videos. Experiments show that LanDiff, a 5B model, achieves a score of 85.43 on the VBenchmark T2V benchmark, surpassing the state-of-the-art open-source models Hunyuan Video (13B) and other commercial models such as Sora, Kling, and Hailuo. Furthermore, our model also achieves state-of-the-art performance in long video generation, surpassing other open-source models in this field. Our demo can be viewed at https://landiff.github.io/." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 567, + 135, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 135, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 135, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 588, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 296, + 684 + ], + "type": "text", + "content": "Text-to-video (T2V) [6, 30, 55, 60] has made significant progress in recent years, becoming an important research direction in the fields of computer vision and artificial intelligence. Recent works in T2V models have primarily revolved around two predominant paradigms: autoregressive large language model (LLM)-based [30, 55] frameworks and diffusion-based architectures [6, 60]. However, each paradigm has their own advantages and limitations, as" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 231, + 553, + 371 + ], + "blocks": [ + { + "bbox": [ + 317, + 231, + 553, + 371 + ], + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 371 + ], + "type": "image", + "image_path": "4de5e5549f22ffb211329d92de6ed929e5f0ad1bf570b367cd1192551fd840ef.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 380, + 555, + 480 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 480 + ], + "type": "text", + "content": "Figure 1. The rate-distortion curve illustrates how visual distortion decreases as the number of transmitted bits increases. With just a small number of bits representing high-level semantic features, we can already achieve relatively low visual distortion. Building on this information-theoretic insight, LanDiff combines the strengths of both paradigms: LLMs efficiently generate compact semantic features in the first stage, followed by diffusion models that add perceptual details in the second stage, before final decoding to pixels via VAE. Data from Ho et al. [20], illustration is conceptual." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 495, + 388, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 388, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 388, + 506 + ], + "type": "text", + "content": "shown in Table 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "type": "text", + "content": "From a representation perspective, LLM-based methods leverage discrete tokenization to explicitly encode high-level semantics through vector quantization, effectively prioritizing conceptual abstraction and narrative coherence. However, this discretization inherently sacrifices low-level visual fidelity due to information compression, resulting in low reconstruction quality. In contrast, diffusion-based approaches employ continuous latent representations to preserve much more perceptual details, enabling superior reconstruction quality at the cost of diluted semantic interpretability, as hierarchical features remain entangled in the latent space. From a generative modeling perspective, LLM-based systems adopt autoregressive modeling to enforce causal dependencies between video frames, ensuring strong temporal coherence. However, the autoregressive generation inherently risks error propagation across time steps, where inaccuracies in early predictions amplify dur" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15604" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 112, + 293, + 191 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": "Table 1. The comparison between LLM based and diffusion based T2V systems. The advantages and disadvantages are marked by " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "\\text{喜}" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "\\text{喜}" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 112, + 293, + 191 + ], + "lines": [ + { + "bbox": [ + 58, + 112, + 293, + 191 + ], + "spans": [ + { + "bbox": [ + 58, + 112, + 293, + 191 + ], + "type": "table", + "html": "
MethodsRepresentationsModeling
LLMDiscrete TokensAutoregressvie
Low Reconstruction QualityNo Refinement
Highlight Semantic InformationCausal Modeling
DiffusionContinuous VectorsNon-Autoregressvie
High Reconstruction QualityProgressive Refinement
Lack Semantic InformationNon-causal Modeling
", + "image_path": "11d349a931cbc4d1e87607388c671e212022c997daa23131f6bb298dfb7a988e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 277 + ], + "type": "text", + "content": "ing decoding. In contrast, diffusion-based methods employ non-autoregressive generation, refining outputs in parallel through iterative denoising steps. Although this design mitigates error accumulation and enhances generation flexibility, the absence of explicit causal constraints often leads to temporal inconsistencies or semantic hallucinations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 280, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 296, + 639 + ], + "type": "text", + "content": "In this work, we propose a hybrid architecture that synergizes the strengths of both Language models and Diffusion models, named LanDiff, through a coarse-to-fine generation paradigm. As shown in Figure 1, inspired by the human creation of video which will generate the high-level storyline first and then add low-level visual details based on the storyline to form the video, we design a two-stage video generation process with the number of bits gradually increasing and carefully design the autoregressive model and the diffusion model to be responsible for different stages of T2V generation, so as to play to their strengths and avoid their weaknesses. In detail, 1) at the low-bit position \"semantic feature\", the low-bit information ensures that the token sequence is not too long, and the high-level information makes it easier for the model to capture the overall semantic entity motion of the video, so as to play to the strengths of the autoregressive model and avoid its weaknesses. Thus we use LLM to generate a coarse-grained video in the first stage; 2) at the high-bit position \"perceptual feature\", since we have already obtained the coarse-grained with rich semantic and time-serial information, we only need to focus on how to add details to the coarse-grained video. Thus we apply a diffusion model in the second stage. Finally, a VAE decoder transforms the generated \"perceptual feature\" into the final RGB video output. By unifying these complementary mechanisms, we demonstrate that hybrid architectures can overcome the inherent limitations of isolated approaches, enabling coherent, semantically faithful, and visually compelling video generation from textual descriptions, as shown in Table 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "With this design, the ideal semantic feature should contain high-level semantic information and motion information and only require a few bits to represent. We achieve this goal by performing extreme compression on video representations rich in high-level semantics. For video representation, we select the Theia model [46] as our visual rep" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 123, + 550, + 211 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 117 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 117 + ], + "type": "text", + "content": "Table 2. The comparison between LanDiff and previous largescale T2V systems. The compression rates of LLM-based and diffusion-based models are illustrated using VideoPoet [30] and CogVideoX [60] as examples, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 123, + 550, + 211 + ], + "lines": [ + { + "bbox": [ + 317, + 123, + 550, + 211 + ], + "spans": [ + { + "bbox": [ + 317, + 123, + 550, + 211 + ], + "type": "table", + "html": "
ModelsLLMDiffusionLanDiff
Highlight Semantic Information?✓✗✓
Progressive Refinement?✗✓✓
Causal Modeling?✓✗✓
High Visual Quality?✗✓✓
Compression Ratio ↑~256~1024~14000
Long video Generation?✗✗✓
", + "image_path": "73f574e84f4cdee02ca331b688d84801cb835181e928cc8c84d91663c6759158.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 215, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 467 + ], + "type": "text", + "content": "resentation backbone, which has been distilled from multiple visual understanding and self-supervised representation models, ensuring the encoded features contain rich semantic information. To achieve extreme compression and reduce the number of bits, we design an efficient tokenizer to compress 3D visual features into 1D discrete representations. The tokenizer is based on the Transformer [54] structure, uses query embedding to aggregate visual features, and has a higher compression rate than CNN-based structures [63]. To further compress the video by fully utilizing the temporal redundancy of the video, inspired by the MP4 video encoding algorithm [32], we divide the video into keyframes and non-keyframes, and set more numbers of tokens for keyframes. The detailed design is shown in subsection 3.1. For the diffusion model, we use generated semantic tokens as conditions and generate the target video by gradually removing the noises. To better support the long-video generation, we train a chunk-wise streaming diffusion model that only uses a limited number of historical frames as conditions, thereby greatly reducing the computational cost of training and inference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 467, + 556, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 556, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 556, + 611 + ], + "type": "text", + "content": "Thanks to these designs, our LanDiff has made significant progress in spatial relationship compliance, action coherence, visual quality, etc. Specifically, 1) for short video generation, our 5B model achieved a score of 85.43 on the VBenchmark, surpassing the state-of-the-art open-source models Hunyuan Video (13B) and other commercial models such as Sora, Kling, and Hailuo; 2) for long video generation, after testing by the VBenchmark T2V benchmark, our model also achieved state-of-the-art performance, surpassing other open-source models in this field. Our video examples can be viewed at https://landiff.github.io/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 621, + 403, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 621, + 403, + 634 + ], + "spans": [ + { + "bbox": [ + 314, + 621, + 403, + 634 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Video Tokenization. Video tokenization plays a crucial role in video understanding and generation tasks. Since a video can be represented as a sequence of continuous frames, some works directly use image tokenizers to process videos frame by frame. For example, Wang et al. [55] directly use SBER-MoVQGAN as the video tokenizer." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15605" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "type": "text", + "content": "However, this method ignores the temporal redundancy in videos, resulting in a low compression rate. To reduce temporal redundancy, some works [58, 62] try to extend image tokenizers based on 2D convolution to 3D convolution, which can process both temporal and spatial information simultaneously. These methods encode videos in the original RGB space and perform video reconstruction tasks, which are more about perceptual compression. In addition, some works [15, 25] try to train video tokenizers on features extracted from pre-trained visual encoders. These tokenizers can achieve good performance in understanding and generation tasks while maintaining a high compression rate. The video tokenizer we propose belongs to this category. Unlike previous feature-based tokenizers that achieved limited compression rates and primarily focused on image processing, our video tokenizer delivers significantly higher compression while handling both images and videos in a unified framework." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 292, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 295, + 579 + ], + "type": "text", + "content": "LLM based Video Generation. LLM-based video generation methods are usually based on the Transformer [54] structure, learning the mapping from text to video through next token prediction. TATS [14] uses VQ-GAN as the tokenizer and predicts video tokens using a GPT-like model structure. VideoGPT [58] uses 3D convolution to extract features and quantize them, and predicts the quantized video discrete tokens using a GPT-like model. Recently, VideoPoet [62] uses MagViT2 [62] as the tokenizer and unifies multiple modalities as input to a large language model (LLM) to conditionally generate video tokens. In addition, Emu3 [55] uses SBER-MoVQGAN as the tokenizer to perform video understanding and generation by predicting the next token. These works all use LLMs to directly generate \"perceptual features\" that contain rich visual details with high bit rates. Recently, ARLON [33] attempts to discretize VAE features into a small number of tokens to reduce the bit rate required for LLMs' prediction. In this way, the tokens retain low-frequency visual information such as blurry contours rather than high-level semantic information. In contrast, our method employs tokens containing high-level semantic information as prediction targets for LLMs, which enables us to fully leverage LLMs' advantages in causal modeling to precisely generate high-quality videos." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Diffusion based Video Generation. Diffusion-based methods have achieved great success in image generation, and recently many people [7, 21, 22, 49, 69] have tried to apply them to video generation tasks. VDM [22] extends the 3D U-Net structure for video generation. Wang et al. [56] propose to generate high-quality and aesthetically pleasing videos in a cascaded manner. Benefiting from the success of the text-to-image (T2I) field, some works such as Animatediff [16], SVD [6], and PixelDance [65] try to use pre-trained T2I models as initialization, and then add modules for temporal modeling to capture motion informa" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": "tion for video generation. Ma et al. [39] explore the generation capabilities of multiple different structures of latent diffusion transformer. After the release of SORA, a series of video generation methods based on the DiT [41] model have been proposed, including OpenSora [68], OpenSoraPlan [34], Cogvideox [60], Hunyuan Video [31], Mira [27] and STIV [35] etc. These methods can only generate short videos of a few seconds. Recently, StreamingT2V [18] generates long videos by block-wise generation on a pre-trained short video generation model, and then uniformly performs mixed augmentation. In addition, some works improve the consistency of long video generation by leveraging noise rescheduling techniques [37, 43]. Our method employs diffusion models as renderers for semantic features, enabling us to leverage their superior visual generation quality while circumventing their limitations in causal modeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 274, + 370, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 370, + 285 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 370, + 285 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 294, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 472 + ], + "type": "text", + "content": "In this work, we propose a novel text-to-video generation framework that synergistically integrates the strengths of autoregressive modeling and diffusion processes while circumventing their respective limitations. The framework mainly consists of the following components: 1) an efficient tokenizer that transforms 3D visual features into compact 1D discrete representations while preserving and enhancing their semantic information; 2) a language model that performs temporal sequence modeling to generate semantic tokens representing video blueprints from textual descriptions; 3) a streaming diffusion model that progressively refines coarse semantic videos by adding fine-grained details to produce high-quality VAE features; and 4) a VAE decoder that reconstructs the final video frames from the refined VAE features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 480, + 459, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 459, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 459, + 492 + ], + "type": "text", + "content": "3.1. Video Semantic Tokenizer" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 498, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 665 + ], + "type": "text", + "content": "In this section, we introduce a novel video semantic tokenizer that efficiently compresses the video into semantic information. Firstly, we will introduce the video semantic representation used in tokenization. Then, considering the high redundancy of video in both spatial and temporal dimensions, we introduce two compression strategies: 1) a query-based causal tokenization that efficiently reduces spatial redundancy while preserving essential semantic information through vector quantization; 2) inspired by MP4 [32], we implement video frame grouping to minimize temporal redundancy by treating grouped frames as a unit, where the first frame (I-frame) is fully encoded while subsequent frames (P-frames) only capture the temporal changes by referencing content from previous frames." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "Video Semantic Representation. Generally, video representations can be divided into two categories: 1) some methods [55, 62] directly utilize an autoencoder to learn the video representations. 2) Some works [25, 26, 29, 51] use" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15606" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 70, + 504, + 266 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 504, + 266 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 504, + 266 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 504, + 266 + ], + "type": "image", + "image_path": "bca7e08238517ff87e8b2fcc62116f2dd4eff394178801b9aed0b7c24836e234.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 268, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 268, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 555, + 301 + ], + "type": "text", + "content": "Figure 2. The architecture of LanDiff. Given text inputs, we first extract text embeddings and employ an LLM to generate semantic tokens in the first stage. Subsequently, we utilize a diffusion model to synthesize perceptual features conditioned on these semantic tokens, followed by a VAE decoder that transforms these features into the final video frames." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 311, + 294, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 311, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 294, + 381 + ], + "type": "text", + "content": "pre-trained visual self-supervised learning features (SSL) as video representations. Compared with the first directly learned autoencoder latents which contain lots of visual details, the second SSL features maintain much more semantic features, which enable the LLM to focus more on the high-level semantic information of the video." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 454 + ], + "type": "text", + "content": "Based on these thoughts, we choose the pretrained SSL features as the video representations. For better usage of different SSL models, we choose Theia model [46], which is a unified visual feature extractor distilled from multiple visual task models, including CLIP [44], SAM [28], DINOv2 [40], ViT [10], and Depth-Anything [59]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 455, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 514 + ], + "type": "text", + "content": "Tokenizer Design. In this part, we introduce the design of our tokenizer, which leverages the query tokens to compress the video semantic features and uses quantization to discrete the video semantic representation while minimizing the reconstruction loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 598 + ], + "type": "text", + "content": "In detail, firstly, we extract the semantic features using the Theia model and flatten it to obtain " + }, + { + "bbox": [ + 55, + 514, + 296, + 598 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}" + }, + { + "bbox": [ + 55, + 514, + 296, + 598 + ], + "type": "text", + "content": ". Inspired by TiTok [63], we use N randomly initialized tokens as query tokens " + }, + { + "bbox": [ + 55, + 514, + 296, + 598 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 55, + 514, + 296, + 598 + ], + "type": "text", + "content": " and concatenate them with the semantic features " + }, + { + "bbox": [ + 55, + 514, + 296, + 598 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 514, + 296, + 598 + ], + "type": "text", + "content": ". Then we use a transformer encoder to encode them and only take the encoded features of the query tokens." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 606, + 295, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 606, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 134, + 606, + 295, + 620 + ], + "type": "interline_equation", + "content": "Z _ {Q} = \\operatorname {E n c} ([ F; Q ]), \\tag {1}", + "image_path": "adb1579eb82c062ff56d623a668e06d9d2a4dd51b9d305c8416512be0636c0e6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "inline_equation", + "content": "[\\cdot ]" + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "text", + "content": " represents the concatenation operation and " + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "inline_equation", + "content": "Z_{Q}\\in" + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "text", + "content": " . We then apply vector quantization on " + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "inline_equation", + "content": "Z_{Q}" + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "text", + "content": " by train a VQ-VAE model and obtain the quantized feature " + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{Z}_Q" + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "text", + "content": " . In the decoding stage, the quantized feature " + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{Z}_Q" + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "text", + "content": " is used as a condition, and then a sequence of mask tokens " + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "inline_equation", + "content": "M\\in \\mathbb{R}^{T\\times H\\times W}" + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "text", + "content": " are added in front of the " + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{Z}_Q" + }, + { + "bbox": [ + 55, + 626, + 296, + 713 + ], + "type": "text", + "content": " to form the inputs of decoder. Then we only take the features of the mask tokens as fol" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 311, + 338, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 311, + 338, + 321 + ], + "spans": [ + { + "bbox": [ + 314, + 311, + 338, + 321 + ], + "type": "text", + "content": "lows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 391, + 321, + 553, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 321, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 391, + 321, + 553, + 336 + ], + "type": "interline_equation", + "content": "\\hat {F} = \\operatorname {D e c} ([ M; \\hat {Z} _ {Q} ]), \\tag {2}", + "image_path": "feb1755dcfdce86991795a755bf5d7e5f76b2616f46a1220b4e4ce4d6ca22616.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "type": "inline_equation", + "content": "\\hat{F} \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}" + }, + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "type": "text", + "content": " represents the reconstructed feature, and " + }, + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 339, + 553, + 399 + ], + "type": "text", + "content": " represents the mask tokens. Inspired by Huang et al. [23], Wang et al. [57], we minimize the reconstruction loss of the video semantic feature during training the VQ-VAE." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 400, + 553, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 553, + 459 + ], + "type": "text", + "content": "For the VQ-VAE, we follow the method of Wang et al. [57], Yu et al. [61]. We update the codebook using exponential moving average (EMA). We also optimize the codebook with the video semantic feature reconstruction loss. The loss function is shown as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 465, + 553, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 465, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 326, + 465, + 553, + 481 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {r e c} \\| \\hat {F} - F \\| _ {2} + \\lambda_ {c o m m i t} \\| \\operatorname {s g} (\\hat {Z} _ {Q}) - Z _ {Q} \\| _ {2}, \\tag {3}", + "image_path": "bd97a8a6e2802a3c2aa0fcc21e3f3281e0fa51f64c685d9ebfbfb609e3c77e4f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 553, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 508 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 486, + 553, + 508 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}()" + }, + { + "bbox": [ + 313, + 486, + 553, + 508 + ], + "type": "text", + "content": " is the stop-gradient operation, " + }, + { + "bbox": [ + 313, + 486, + 553, + 508 + ], + "type": "inline_equation", + "content": "||\\cdot ||_2" + }, + { + "bbox": [ + 313, + 486, + 553, + 508 + ], + "type": "text", + "content": " is the L2 loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "type": "text", + "content": "Video Frame Grouping. For video, a straightforward observation is the redundancy in time series (i.e., the difference between adjacent video frames is minimal). Intuitively, we can achieve a better compression rate by modeling the difference between adjacent video frames instead of tokenizing all frames equally. Inspired by the popular video compression method MP4 [32], given " + }, + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "type": "text", + "content": " frames of a video, we first split them into " + }, + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "type": "inline_equation", + "content": "N / T" + }, + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "type": "text", + "content": " groups (i.e., " + }, + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "type": "text", + "content": " frames as a group. For clarification, as shown in Figure 3, we use " + }, + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "type": "text", + "content": " as an example.). Then: 1) we will model different groups independently; 2) for each group, we will fully encode the first frame (Intra-coded Frame, IFrame), while for the remaining " + }, + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "type": "inline_equation", + "content": "[1,T - 1]" + }, + { + "bbox": [ + 313, + 510, + 555, + 688 + ], + "type": "text", + "content": " frames (the Predictive-coded Frame, PFrame), we encode them by referencing their previous frames." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "To achieve this, within a group, the first frame (IFrame) will only see itself and have a large number of query tokens" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15607" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 130, + 72, + 304, + 297 + ], + "blocks": [ + { + "bbox": [ + 130, + 72, + 304, + 297 + ], + "lines": [ + { + "bbox": [ + 130, + 72, + 304, + 297 + ], + "spans": [ + { + "bbox": [ + 130, + 72, + 304, + 297 + ], + "type": "image", + "image_path": "0255745df1d25767370eb3b425067e43c428c0dffbef7d7588a2b1b3cab18b30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 478, + 300 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 478, + 300 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 478, + 300 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 478, + 300 + ], + "type": "image", + "image_path": "db8d53fd9c435c1e0f79d208bb4eb8b78c66c47877ce383b61fc231d9483919c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 300, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 54, + 300, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 555, + 335 + ], + "type": "text", + "content": "Figure 3. Proposed architecture of the video semantic tokenizer. We use query tokens to compress the semantic sequence length. Furthermore, we group the frames into groups (3 frames in a group in this figure). In a group, the first frame is the IFrame and the rest frames are PFrames. We use different query token numbers for them. The attention mask design is shown in the right." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 338, + 297, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 338, + 297, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 297, + 495 + ], + "type": "text", + "content": "(e.g., 3 query tokens in Figure 3) to achieve better reconstruction quality. For the rest frame (PFrame) " + }, + { + "bbox": [ + 54, + 338, + 297, + 495 + ], + "type": "inline_equation", + "content": "i \\in [1, T - 1]" + }, + { + "bbox": [ + 54, + 338, + 297, + 495 + ], + "type": "text", + "content": ", it will see the previous frames (i.e., frames " + }, + { + "bbox": [ + 54, + 338, + 297, + 495 + ], + "type": "inline_equation", + "content": "j \\in [0, i - 1]" + }, + { + "bbox": [ + 54, + 338, + 297, + 495 + ], + "type": "text", + "content": ") and have a small number of query tokens (e.g., 1 query token which is " + }, + { + "bbox": [ + 54, + 338, + 297, + 495 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 54, + 338, + 297, + 495 + ], + "type": "text", + "content": " of the IFrame) to force the model to learn the difference. Technically, as shown in the attention mask on the right side of Figure 3, we use frame-level causal masks for the feature sequence during encoding. The query tokens are also divided according to the frame, and each token can only attend to the features of the corresponding frame and the previous frames. During decoding, the mask token corresponding to each frame can see the previous features, the corresponding query tokens, and the previous query tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 503, + 296, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 296, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 296, + 527 + ], + "type": "text", + "content": "3.2. Language Model for Semantic Token Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "content": "As shown in Figure 2, after training an efficient tokenizer, we use a language model to generate semantic tokens autoregressively based on text. Specifically, we first use a pretrained text encoder T5-XXL [45] to extract text features " + }, + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "content": ". We use the video tokenizer encoder trained in subsection 3.1 to convert the video into a 1D discrete token sequence " + }, + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "content": ". To enhance the controllability of the generated results, we introduce additional control conditions " + }, + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "inline_equation", + "content": "CC" + }, + { + "bbox": [ + 55, + 534, + 297, + 713 + ], + "type": "text", + "content": ". It includes: 1) frames condition for requiring the model to generate videos with a specified number of frames; 2) motion score condition, which is a value between 0 and 1, used to control the degree of motion in the generated video. The discrete tokens are converted into embedding vectors during generation and added with positional encoding, and then concatenated with the control conditions as input. The structure of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 338, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 374 + ], + "type": "text", + "content": "the language model follows the typical LLaMA model [53] network structure. We train the model from scratch, using cross-entropy loss as the loss function." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 362, + 387, + 555, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 387, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 362, + 387, + 555, + 401 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L M}} = \\mathbb {E} \\left[ - \\log p \\left(Y _ {i} \\mid X, C C, Y _ {< i}\\right) \\right] \\tag {4}", + "image_path": "3f8074f65fde2600537c2f8b6bf92a2ae8f8b6648c1dd10bc291067be08f518b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 407, + 555, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 432 + ], + "type": "text", + "content": "3.3. Diffusion Model for Perceptual Feature Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 533 + ], + "type": "text", + "content": "The video detokenizer in LanDiff is responsible for converting semantic tokens into VAE latent vectors. As shown in Figure 2 we use a conditional diffusion model to complete this task. Especially to support long-video generation, we also design the streaming inference strategy. In this section, we will first introduce the architecture for a single semantic token chunk. Then we will introduce the chunk-wise streaming strategy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": "Architecture. Our diffusion model employs an architecture similar to MMDiT[12, 60]. Specifically, 1) we use the video tokenizer decoder trained in subsection 3.1 to decode the semantic tokens into semantic features " + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\hat{F}" + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": ". Then we use the semantic features " + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\hat{F}" + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": " as a condition to guide the diffusion model to generate videos; 2) we inject the control signals into the model in a similar way to ControlNet [67]. In detail, during training, the parameters of the main model are not updated, the control module copies the parameters of the first half layers of the main model, and adds them to the output of the main model after a linear layer initialized with zeros; 3) to make the semantic features match the target VAE features in the space dimension, we additionally add an upsampling module. We show the model structure in the appendix Figure 8." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15608" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "During training, given the input video chunk " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " is the frame length of the VAE latent and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " is the VAE latent feature dimension, the diffusion algorithm progressively adds noise to the video and produces a noisy video " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "V^t" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " represents the number of times noise is added. The diffusion algorithm uses time step " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", and semantic features " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\hat{F} \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " as conditions, and then uses a network " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " to predict the noise added to the noisy video " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "V_t" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " through:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 184, + 295, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 184, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 85, + 184, + 295, + 205 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {V, t, c _ {s}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(V ^ {t}, t, \\hat {F}\\right) \\| _ {2} ^ {2} \\right] \\tag {5}", + "image_path": "ff1c19255de29be52996ce3ce0bf96f2a8a86c90fe77d27bb3f058ea07750904.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "type": "text", + "content": "Chunk-wise Streaming Strategy. To support long video generation scenario whose semantic token sequence is very long and difficult to generate as a whole, we propose the chunk-wise streaming strategy. During training, given video latent chunk " + }, + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{(T \\times H \\times W) \\times D}" + }, + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "type": "text", + "content": ", to maintain the appearance continuity in the video, we use the first half of " + }, + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "type": "inline_equation", + "content": "V(V_{l})" + }, + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "type": "text", + "content": " as the prompt and generate the second half of " + }, + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "type": "inline_equation", + "content": "V_{r}" + }, + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "type": "text", + "content": ". In detail, we do not add noise to the first half chunk and give time condition " + }, + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "type": "text", + "content": " to 0.999. We also randomly mask the first half chunk with a ratio of " + }, + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 207, + 296, + 337 + ], + "type": "text", + "content": ". The loss function for training at this time is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 340, + 295, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 340, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 69, + 340, + 295, + 361 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {V _ {r}, t, c _ {s}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(V _ {l}, V _ {r} ^ {t}, t, c _ {s}\\right) \\| _ {2} ^ {2} \\right]. \\tag {6}", + "image_path": "8bea50a74da9b8d260d0826669a983da28a1afd8c52dd65b50bd8d347cdbb43c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 364, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 413 + ], + "type": "text", + "content": "During inference, the first " + }, + { + "bbox": [ + 55, + 364, + 296, + 413 + ], + "type": "inline_equation", + "content": "L / 2" + }, + { + "bbox": [ + 55, + 364, + 296, + 413 + ], + "type": "text", + "content": " VAE latents will be generated without the prompt. For the following VAE latents, we will accumulate " + }, + { + "bbox": [ + 55, + 364, + 296, + 413 + ], + "type": "inline_equation", + "content": "L / 2" + }, + { + "bbox": [ + 55, + 364, + 296, + 413 + ], + "type": "text", + "content": " tokens to form a chunk and use the previous chunk as prompt." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 198, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 198, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 198, + 435 + ], + "type": "text", + "content": "4. Experiments and Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 441, + 182, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 182, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 182, + 453 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 458, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 295, + 589 + ], + "type": "text", + "content": "Datasets. For the video tokenizer and language model, we use an internal dataset with 200M video-text pairs for training. All videos with a duration of less than 6s are filtered, and the videos are kept in the aspect ratio and then scaled to around 480x720 resolution for center cropping. Consistent with CogVideoX [60], we set the fps of all videos to 8. For the diffusion model, we select a dataset containing 3M high-quality video-text pairs for training. To evaluate the performance of the text-to-video generation model, we use the prompts provided by the widely used VBench [24] T2V benchmark to generate videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 590, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 696 + ], + "type": "text", + "content": "Implementation Details. For the video tokenizer, we set the group size " + }, + { + "bbox": [ + 55, + 590, + 296, + 696 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 590, + 296, + 696 + ], + "type": "text", + "content": " mentioned in subsection 3.1 to 13. We set the number of tokens " + }, + { + "bbox": [ + 55, + 590, + 296, + 696 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 590, + 296, + 696 + ], + "type": "text", + "content": " corresponding to the IFrame to 330, and the number of tokens " + }, + { + "bbox": [ + 55, + 590, + 296, + 696 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 590, + 296, + 696 + ], + "type": "text", + "content": " corresponding to the PFrame to 74. The dimension of the tokenizer's codebook is set to 16, and the vocabulary size is set to 2048. For the language model, we use the LLaMA [53] structure, with 2B model parameters, and use 1D RoPE [50] positional encoding. We use the 2B version of the CogVideoX [60] model as the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": "base model for our video detokenizer. We list the structure settings and more implementation details of each module in the appendix subsection A.1, subsection A.2, and subsection A.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 121, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 554, + 156 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate the text-to-video generation task, we use the metrics proposed in the VBench and VBench-Long [24] benchmarks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 157, + 554, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 554, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 554, + 265 + ], + "type": "text", + "content": "Evaluation Baselines. We compare LanDiff with baselines: 1) Sora [5]. 2) Jimeng [3]. 3) Hailuo [2]. 4) OpenSoraPlan V1.1 [34]. 5) Kling [38]. 6) InstructVideo [64]. 7) Gen-3 [1]. 9) Latte-1 [39] 9) HiGen [42]. 10)AnimateDiffV2 [16]. 11) Show-1 [66]. 12) Pika [4]. 13) VideoCrafter2.0 [7]. 14) OpenSora V1.2 [68]. 15) LTX-Video [17]. 16) Mochi-1 [52]. 17) CogVideoX [60]. 18) Vchitect-2.0 [13]. 19) RepVideo [48]. 20) HunyuanVideo [31]. 21) ARLON [33]. 22) Emu3 [55]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 274, + 438, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 438, + 286 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 438, + 286 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 303, + 550, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 303, + 550, + 366 + ], + "spans": [ + { + "bbox": [ + 318, + 303, + 550, + 366 + ], + "type": "text", + "content": "A life-sized ice sculpture of a playful dog, with intricate details and a joyful expression, stands in the middle of a sunlit, grassy field on a sweltering summer day. The ice dog, initially solid and vibrant, begins to melt under the relentless heat, with droplets of water forming on its surface. As the day progresses, the ice dog's form gradually diminishes, with its once sharp features becoming blurred and distorted. The melting process accelerates, and the ice dog's body starts to collapse, pooling into a puddle of water on the ground. By the end of the day, all that remains is a shallow puddle, reflecting the cloudless sky, with the memory of the once majestic ice dog now just a memory." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 319, + 364, + 385, + 434 + ], + "blocks": [ + { + "bbox": [ + 319, + 364, + 385, + 434 + ], + "lines": [ + { + "bbox": [ + 319, + 364, + 385, + 434 + ], + "spans": [ + { + "bbox": [ + 319, + 364, + 385, + 434 + ], + "type": "image", + "image_path": "df4b8ae44b75e13b80b941287681438324475d9ad202ba0dd8c7b1edd40b4ced.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 447, + 554, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 468 + ], + "type": "text", + "content": "Figure 4. Comparison of qualitative results for text-to-video generation." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 387, + 365, + 439, + 436 + ], + "blocks": [ + { + "bbox": [ + 387, + 365, + 439, + 436 + ], + "lines": [ + { + "bbox": [ + 387, + 365, + 439, + 436 + ], + "spans": [ + { + "bbox": [ + 387, + 365, + 439, + 436 + ], + "type": "image", + "image_path": "e8f2da8dfaeb6db258ae37e51ff71724171ceae3be3a1e4be6921efaf6da66e3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 441, + 365, + 493, + 436 + ], + "blocks": [ + { + "bbox": [ + 441, + 365, + 493, + 436 + ], + "lines": [ + { + "bbox": [ + 441, + 365, + 493, + 436 + ], + "spans": [ + { + "bbox": [ + 441, + 365, + 493, + 436 + ], + "type": "image", + "image_path": "5d1b7e3b8e30e606bc38e7fc2b50a06b74b610c43d9cd413e1d6fc2a6b65018c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 495, + 365, + 547, + 436 + ], + "blocks": [ + { + "bbox": [ + 495, + 365, + 547, + 436 + ], + "lines": [ + { + "bbox": [ + 495, + 365, + 547, + 436 + ], + "spans": [ + { + "bbox": [ + 495, + 365, + 547, + 436 + ], + "type": "image", + "image_path": "35b80902c0c1627075823259dbcf504b84b22689e9f62362fc80fa09b7946671.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": "Text to Video Generation. As shown in Table 3, our 5B LanDiff model outperforms the state-of-the-art open-source models on VBenchmark, including the 13B HunyuanVideo, achieving a top score of 85.42. LanDiff exhibits superior semantic understanding while maintaining excellent generation quality, surpassing same-sized CogVideoX-5B across most metrics. To validate our approach's effectiveness independent of training data volume, we trained a 7B DiT-based diffusion model with identical training data (marked as DiT in the table). The results confirm that our task division strategy successfully leverages the complementary strengths of autoregressive and diffusion models. Figure 4 shows qualitative comparisons between LanDiff and CogVideoX-5B. Our model effectively captures temporal dynamics described in the prompt, accurately rendering a melting ice dog sculpture. LanDiff successfully depicts the progressive transformation from solid form to liquid state, while CogVideoX-5B produces a static ice sculpture that remains virtually unchanged throughout the sequence. This demonstrates our model's superior ability to compre" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 270, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 270, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 270, + 712 + ], + "type": "text", + "content": "1https://huggingface.co/THUDM/CogVideoX-2b" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15609" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 127, + 527, + 423 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "text", + "content": "Table 3. Performance comparison of Text-to-video (T2V) generation between our LanDiff and other state-of-the-art models on VBench benchmark. We selected 8 out of the 16 evaluation dimensions from VBench, along with Total Score, Quality Score, and Semantic Score, for presentation. The complete results with all 16 evaluation dimensions can be found in the appendix Table 8. The best and second-best scores are highlighted in bold and underline, respectively. " + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "text", + "content": " indicates the scores we reproduced, while " + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "text", + "content": " indicates the scores from the original papers, and other scores are from the VBench benchmark." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 127, + 527, + 423 + ], + "lines": [ + { + "bbox": [ + 82, + 127, + 527, + 423 + ], + "spans": [ + { + "bbox": [ + 82, + 127, + 527, + 423 + ], + "type": "table", + "html": "
Model NameTypeModel SizeTotal ScoreQuality ScoreSemantic Scorebackground consistencydynamic degreemotion smoothnessmultiple objectsobject classscenespatial relationshipsubject consistency
Open Sourced Models
InstructVideoDiffusion1.3B76.6181.5656.8196.9769.7296.6221.5773.2622.2143.4995.30
Latte-1Diffusion0.7B77.2979.7267.5895.4068.8994.6334.5386.5336.2641.5388.88
OpenSoraPlan V1.1Diffusion2.7B78.0080.9166.3896.7347.7298.2840.3576.3027.1753.1195.73
Show-1Diffusion6.3B78.9380.4272.9898.0244.4498.2445.4793.0747.0353.5095.53
OpenSora V1.2Diffusion1.1B79.7681.3573.3997.6142.3998.5051.8382.2242.4468.5696.75
LTX-VideoDiffusion1.9B80.0082.3070.7997.2054.3598.9645.4383.4551.0765.4396.56
Mochi-1Diffusion10B80.1382.6470.0897.2861.8599.0250.4786.5136.9969.2496.99
AnimateDiff-V2Diffusion1.3B80.2782.9069.7597.6840.8397.7636.8890.9050.1934.6095.30
VideoCrafter-2.0Diffusion1.7B80.4482.2073.4298.2242.5097.7340.6692.5555.2935.8696.85
CogVideoX-2BDiffusion2B80.9182.1875.8396.6359.8697.7362.6383.3751.1469.9096.78
Emu3 ‡LLM8B80.96N/AN/A97.6979.2798.9344.6486.1737.1168.7395.32
Vchitect-2.0-2BDiffusion2B81.5782.5177.7996.5358.3397.7669.3587.8157.5154.6496.42
CogVideoX-5BDiffusion5B81.6182.7577.0496.5270.9796.9262.1185.2353.2066.3596.23
DiT †Diffusion7B81.8582.7078.4297.6551.0298.9469.6393.4057.1060.8997.05
RepVideoDiffusion2B81.9482.7078.9196.5657.7898.1371.1887.8352.9674.7496.25
Vchitect-2.0[E]Diffusion2B82.2483.5477.0696.6663.8998.9868.8486.6156.5757.5596.83
HunyuanVideoDiffusion13B83.2485.0975.8297.7670.8398.9968.5586.1053.8868.6897.37
Close Sourced Models
Pika-1.0DiffusionN/A80.6982.9271.7797.3647.5099.5043.0888.7249.8361.0396.94
KlingDiffusionN/A81.8583.3975.6897.6046.9499.4068.0587.2450.8673.0398.33
JimengDiffusionN/A81.9783.2976.6998.3938.4398.0969.0889.6244.9477.4597.25
Gen-3DiffusionN/A82.3284.1175.1796.6260.1499.2353.6487.8154.5765.0997.10
HailuoDiffusionN/A83.4184.8577.6597.0564.9199.2276.0487.8350.6875.5097.51
SoraDiffusionN/A84.2885.5179.3596.3579.9198.7470.8593.9356.9574.2996.23
ARLON ‡LLM+Diffusion1.5BN/AN/AN/A97.1052.7798.92N/A89.8054.43N/A93.41
ARLON †LLM+Diffusion5B82.3183.5877.2798.6572.2297.5674.4990.6052.0762.5395.59
LanDiffLLM+Diffusion5B85.4386.1382.6198.7392.7197.0886.6994.9453.7973.7496.11
", + "image_path": "930cecbb8f0e7fa1fe94aae210f06c336fc013eddae364f837d729f7bae6c831.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 438, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 438, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 438, + 295, + 474 + ], + "type": "text", + "content": "hend and visualize complex temporal transformations and physical processes. For more video examples, please refer to our demo website2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 475, + 295, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 295, + 691 + ], + "type": "text", + "content": "Long Video Generation. We compare our model with other open-source text-to-long video generation models: FreeNoise [43], StreamingT2V [18], OpenSora-V1.2 [68], and ARLON [33]. As shown in Table 4, LanDiff achieves state-of-the-art performance with the highest Total Score of 68.34. While StreamingT2V shows superior Dynamic Degree (85.64), it sacrifices consistency metrics (Subject: 87.31, Background: 94.64). Conversely, models like ARLON excel in consistency but exhibit limited dynamism (50.42). Our approach strikes an optimal balance with strong dynamism (72.86) while maintaining high consistency scores and achieving the best results in Aesthetic Quality (60.96) and Overall Consistency (27.29). As illustrated in Figure 5, LanDiff effectively renders complex motion dynamics in the prompt \"a car accelerating to gain speed,\" successfully depicting increasing velocity while preserving visual consistency. In contrast, competing models struggle with this motion, producing either static" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 438, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 472 + ], + "type": "text", + "content": "vehicles, inconsistent appearances, or unrealistic acceleration patterns. Additional video samples are available on our demo site2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 475, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 629 + ], + "type": "text", + "content": "Video Tokenization. We present visualization results of our model's video reconstruction in Figure 6. We extract semantic tokens from the reference video and subsequently convert these tokens back into video using the video detokenizer. The results demonstrate that through our careful design, our video tokenizer can accurately reconstruct videos with accurate semantics and actions using only approximately 1/50th of the sequence length required by MagViT2 [62]. While some minor discrepancies in clothing details exist between the reconstructed and reference videos, these differences remain within acceptable limits for practical video generation applications. Additional video samples are available on our demo site2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 635, + 408, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 408, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 408, + 649 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "In this section, we conducted ablation experiments on the video tokenizer and classifier-free guidance. To fairly evaluate the effectiveness of our video tokenizer, we implemented an ARLON-like method as a comparison baseline. Specifically, for the control group without our proposed" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 197, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 197, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 197, + 712 + ], + "type": "text", + "content": "2https://landiff.github.io/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15610" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 4. Long video generation results of LanDiff and other models on VBench. The best and second-best scores are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 106, + 95, + 501, + 390 + ], + "blocks": [ + { + "bbox": [ + 106, + 95, + 501, + 390 + ], + "lines": [ + { + "bbox": [ + 106, + 95, + 501, + 390 + ], + "spans": [ + { + "bbox": [ + 106, + 95, + 501, + 390 + ], + "type": "image", + "image_path": "2e29716b0accdd512fa71a3676baa79660ee3bf6c05253ac7b973f9077370eed.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 394, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 55, + 394, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 555, + 417 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of long video generation results between LanDiff and other state-of-the-art models (FreeNoise, StreamingT2V and OpenSora-V1.2)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 420, + 289, + 526 + ], + "blocks": [ + { + "bbox": [ + 58, + 420, + 289, + 526 + ], + "lines": [ + { + "bbox": [ + 58, + 420, + 289, + 526 + ], + "spans": [ + { + "bbox": [ + 58, + 420, + 289, + 526 + ], + "type": "image", + "image_path": "24fe78fd9647fa699f7cde4132ec1be689b4e123c0d7cef6543f468d325716a6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 528, + 295, + 548 + ], + "lines": [ + { + "bbox": [ + 55, + 528, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 295, + 548 + ], + "type": "text", + "content": "Figure 6. Visualization results of video reconstruction using video tokenizer." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 83, + 574, + 266, + 651 + ], + "blocks": [ + { + "bbox": [ + 55, + 552, + 295, + 572 + ], + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 572 + ], + "type": "text", + "content": "Table 5. The ablation study of video tokenizer and classifier free guidance (cfg) on Vbench benchmark." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 574, + 266, + 651 + ], + "lines": [ + { + "bbox": [ + 83, + 574, + 266, + 651 + ], + "spans": [ + { + "bbox": [ + 83, + 574, + 266, + 651 + ], + "type": "table", + "html": "
ModelsTotal ScoreQuality ScoreSemantic Score
LanDiff85.4386.1382.61
-video tokenizer82.3183.5877.27
-cfg81.0683.0473.13
", + "image_path": "2775efe50e5bc5b8ffab6a2a1fbf01f0abb0e1f61ae71840c9757fafa37775b8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "video tokenizer, we followed ARLON's architectural approach by training a large language model to predict quantized VAE features, then using CogVideoX-5B as the base diffusion model conditioned on these LLM-predicted VAE" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 421, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 553 + ], + "type": "text", + "content": "features to generate videos. This baseline model was trained with exactly the same dataset and training recipe as Lan-Diff, ensuring strict experimental control variables and effectively eliminating interference factors such as model parameter count and training data scale. As shown in Table 5, the experimental results clearly demonstrate that our proposed video tokenizer significantly improves both the quality score and semantic score of generated videos, validating the importance of this component in our framework. In the case of not using classifier-free guidance for text, our model has decreased in quality and semantic scores." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 388, + 574 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a new text-to-video generation model, LanDiff. It combines the advantages of autoregressive models and diffusion models, including: 1) an efficient 1D video tokenizer to extract videos into semantic tokens; 2) a language model to generate semantic tokens based on text; 3) a video detokenizer to convert semantic tokens into videos. LanDiff outperforms the state-of-the-art open-source models on the VBenchmark, surpassing other open-source models in quality and semantic scores. At the same time, LanDiff also achieves state-of-the-art performance in long video generation tasks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15611" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 157 + ], + "type": "text", + "content": "This work was supported by the NSFC (62272411), the Key R&D Projects in Zhejiang Province (No. 2024C01106, 2025C01030), the Zhejiang NSF (LRG25F020001), Zhejiang University Education Foundation Qizhen Scholar Foundation. This work was also supported by the Moonshot AI." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 187, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 187, + 296, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 187, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 187, + 296, + 209 + ], + "type": "text", + "content": "[1] Gen-3. https://runwayml.com/research/introducing-gen-3-alpha, 2024.6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 210, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 210, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 210, + 296, + 232 + ], + "type": "text", + "content": "[2] Hailuo. https://platform.minimaxi.com/, 2024.6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 233, + 293, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 233, + 293, + 244 + ], + "spans": [ + { + "bbox": [ + 61, + 233, + 293, + 244 + ], + "type": "text", + "content": "[3] Jimeng. https://jimeng.jianying.com/, 2024.6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 245, + 260, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 245, + 260, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 245, + 260, + 256 + ], + "type": "text", + "content": "[4] Pika 1.0. https://www.pika.art/, 2024.6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 258, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 277 + ], + "type": "text", + "content": "[5] Openai sora. https://openai.com/index/sora, 2024.6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 280, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 296, + 334 + ], + "type": "text", + "content": "[6] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets, 2023. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 335, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 335, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 61, + 335, + 296, + 401 + ], + "type": "text", + "content": "[7] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCrafter2: Overcoming data limitations for high-quality video diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 7310-7320. IEEE, 2024. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 402, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 295, + 434 + ], + "type": "text", + "content": "[8] Alexandre Défossez, Jade Copet, Gabriel Synnaeve, and Yossi Adi. High Fidelity Neural Audio Compression, 2022. 14" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 436, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 296, + 468 + ], + "type": "text", + "content": "[9] Juechu Dong, Boyuan Feng, Driss Guessous, Yanbo Liang, and Horace He. Flex Attention: A Programming Model for Generating Optimized Attention Kernels, 2024. 14" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 469, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 469, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 469, + 296, + 557 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 558, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 296, + 602 + ], + "type": "text", + "content": "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming Transformers for High-Resolution Image Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873–12883, 2021. 15" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 603, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 296, + 689 + ], + "type": "text", + "content": "[12] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. Open-Review.net, 2024. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[13] Weichen Fan, Chenyang Si, Junhao Song, Zhenyu Yang, Yinan He, Long Zhuo, Ziqi Huang, Ziyue Dong, Jingwen He," + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "text", + "content": "Dongwei Pan, Yi Wang, Yuming Jiang, Yaohui Wang, Peng Gao, Xinyuan Chen, Hengjie Li, Dahua Lin, Yu Qiao, and Ziwei Liu. Vchitect-2.0: Parallel Transformer for Scaling Up Video Diffusion Models, 2025. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "text", + "content": "[14] Songwei Ge, Thomas Hayes, Harry Yang, Xi Yin, Guan Pang, David Jacobs, Jia-Bin Huang, and Devi Parikh. Long Video Generation with Time-Agnostic VQGAN and Time-Sensitive Transformer, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[15] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a SEED of Vision in Large Language Model, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 554, + 274 + ], + "type": "text", + "content": "[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. AnimateDiff: Animate your personalized text-to-image diffusion models without specific tuning. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 276, + 554, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 340 + ], + "type": "text", + "content": "[17] Yoav HaCohen, Nisan Chiprut, Benny Brazowski, Daniel Shalem, Dudu Moshe, Eitan Richardson, Eran Levin, Guy Shiran, Nir Zabari, Ori Gordon, Poriya Panet, Sapir Weissbuch, Victor Kulikov, Yaki Bitterman, Zeev Melumian, and Ofir Bibi. LTX-Video: Realtime Video Latent Diffusion, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 342, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 397 + ], + "type": "text", + "content": "[18] Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text, 2024. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 399, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 420 + ], + "type": "text", + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-Free Diffusion Guidance, 2022. 15" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 422, + 554, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 476 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "type": "text", + "content": "[21] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey A. Gritsenko, Diederik P. Kingma, Ben Poole, Mohammad Norouzi, David J. Fleet, and Tim Salimans. Imagen video: High definition video generation with diffusion models. CoRR, abs/2210.02303, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 534, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 554, + 567 + ], + "type": "text", + "content": "[22] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video Diffusion Models, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 568, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 554, + 602 + ], + "type": "text", + "content": "[23] Zhichao Huang, Chutong Meng, and Tom Ko. Repocodec: A speech representation codec for speech tokenization. arXiv preprint arXiv:2309.00169, 2023. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 602, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 554, + 689 + ], + "type": "text", + "content": "[24] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 21807-21818. IEEE, 2024. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 714 + ], + "type": "text", + "content": "[25] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jian-choao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15612" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 115 + ], + "type": "text", + "content": "Chengru Song, Xiaogiang Lei, Di Zhang, Wenwu Ou, Kun Gai, and Yadong Mu. Unified Language-Vision Pretraining in LLM with Dynamic Discrete Visual Tokenization, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 173 + ], + "type": "text", + "content": "[26] Yang Jin, Zhicheng Sun, Kun Xu, Kun Xu, Liwei Chen, Hao Jiang, Quzhe Huang, Chengru Song, Yuliang Liu, Di Zhang, Yang Song, Kun Kai, and Yadong Mu. Video-LaVIT: Unified Video-Language Pre-training with Decoupled Visual-Motional Tokenization, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 219 + ], + "type": "text", + "content": "[27] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. MiraData: A Large-Scale Video Dataset with Long Durations and Structured Captions, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 221, + 294, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 221, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 56, + 221, + 294, + 265 + ], + "type": "text", + "content": "[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment Anything, 2023. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 266, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 294, + 331 + ], + "type": "text", + "content": "[29] Jing Yu Koh, Daniel Fried, and Russ Salakhutdinov. Generating images with multimodal language models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 335, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 335, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 294, + 475 + ], + "type": "text", + "content": "[30] Dan Kondratyuk, Lijun Yu, Xiuye Gu, Jose Lezama, Jonathan Huang, Grant Schindler, Rachel Hornung, Vighnesh Birodkar, Jimmy Yan, Ming-Chang Chiu, Krishna Somandepalli, Hassan Akbari, Yair Alon, Yong Cheng, Joshua V. Dillon, Agrim Gupta, Meera Hahn, Anja Hauth, David Hendon, Alonso Martinez, David Minnen, Mikhail Sirotenko, Kihyuk Sohn, Xuan Yang, Hartwig Adam, Ming-Hsuan Yang, Irfan Essa, Huisheng Wang, David A. Ross, Bryan Seybold, and Lu Jiang. VideoPoet: A large language model for zero-shot video generation. In *Forty-First International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 479, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 621 + ], + "type": "text", + "content": "[31] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, Kathrina Wu, Qin Lin, Junkun Yuan, Yanxin Long, Aladdin Wang, Andong Wang, Changlin Li, Duojun Huang, Fang Yang, Hao Tan, Hongmei Wang, Jacob Song, Jiawang Bai, Jianbing Wu, Jinbao Xue, Joey Wang, Kai Wang, Mengyang Liu, Pengyu Li, Shuai Li, Weiyan Wang, Wenqing Yu, Xinchi Deng, Yang Li, Yi Chen, Yutao Cui, Yuanbo Peng, Zhentao Yu, Zhiyu He, Zhiyong Xu, Zixiang Zhou, Zunnan Xu, Yangyu Tao, Qinglin Lu, Songtao Liu, Daquan Zhou, Hongfa Wang, Yong Yang, Di Wang, Yuhong Liu, Jie Jiang, and Caesar Zhong. HunyuanVideo: A Systematic Framework For Large Video Generative Models, 2024. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "type": "text", + "content": "[32] Didier Le Gall. MPEG: A video compression standard for multimedia applications. Communications of the ACM, 34 (4):46-58, 1991. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[33] Zongyi Li, Shujie Hu, Shujie Liu, Long Zhou, Jeongsoo Choi, Lingwei Meng, Xun Guo, Jinyu Li, Hefei Ling, and Furu Wei. ARLON: Boosting Diffusion Transformers with Autoregressive Models for Long Video Generation, 2024. 3, 6, 7" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 148 + ], + "type": "text", + "content": "[34] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, Tanghui Jia, Junwu Zhang, Zhenyu Tang, Yatian Pang, Bin She, Cen Yan, Zhiheng Hu, Xiaoyi Dong, Lin Chen, Zhang Pan, Xing Zhou, Shaoling Dong, Yonghong Tian, and Li Yuan. Open-Sora Plan: Open-Source Large Video Generation Model, 2024. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 150, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 215 + ], + "type": "text", + "content": "[35] Zongyu Lin, Wei Liu, Chen Chen, Jiasen Lu, Wenze Hu, Tsu-Jui Fu, Jesse Allardice, Zhengfeng Lai, Liangchen Song, Bowen Zhang, Cha Chen, Yiran Fei, Yifan Jiang, Lezhi Li, Yizhou Sun, Kai-Wei Chang, and Yinfei Yang. STIV: Scalable Text and Image Conditioned Video Generation, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 217, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 238 + ], + "type": "text", + "content": "[36] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 14" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "type": "text", + "content": "[37] Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. FreeLong: Training-Free Long Video Generation with Spectral-Blend Temporal Attention, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 272, + 555, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 555, + 337 + ], + "type": "text", + "content": "[38] Zhengxiong Luo, Dayou Chen, Yingya Zhang, Yan Huang, Liang Wang, Yujun Shen, Deli Zhao, Jingren Zhou, and Tieniu Tan. Videofusion: Decomposed diffusion models for high-quality video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 338, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 380 + ], + "type": "text", + "content": "[39] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent Diffusion Transformer for Video Generation, 2024. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 382, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 480 + ], + "type": "text", + "content": "[40] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINov2: Learning robust visual features without supervision. 2024, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 482, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 503 + ], + "type": "text", + "content": "[41] William Peebles and Saining Xie. Scalable Diffusion Models with Transformers, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 504, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 547 + ], + "type": "text", + "content": "[42] Zhiwu Qing, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yujie Wei, Yingya Zhang, Changxin Gao, and Nong Sang. Hierarchical spatio-temporal decoupling for text-to-video generation. In CVPR, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 548, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 613 + ], + "type": "text", + "content": "[43] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. FreeNoise: Tuning-free longer video diffusion via noise rescheduling. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 614, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 690 + ], + "type": "text", + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15613" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21:140:1-140:67, 2020. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 151 + ], + "type": "text", + "content": "[46] Jinghuan Shang, Karl Schmeckpeper, Brandon B. May, Maria Vittoria Minniti, Tarik Kelestemur, David Watkins, and Laura Herlant. Theia: Distilling Diverse Vision Foundation Models for Robot Learning, 2024. 2, 4, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 229 + ], + "type": "text", + "content": "[47] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 1874-1883. IEEE Computer Society, 2016. 15" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "text", + "content": "[48] Chenyang Si, Weichen Fan, Zhengyao Lv, Ziqi Huang, Yu Qiao, and Ziwei Liu. RepVideo: Rethinking Cross-Layer Representation for Video Generation, 2025. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 341 + ], + "type": "text", + "content": "[49] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, Devi Parikh, Sonal Gupta, and Yaniv Taigman. Make-a-video: Text-to-video generation without text-video data. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 386 + ], + "type": "text", + "content": "[50] Jianlin Su, Murtadha H. M. Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. RoFormer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 6, 13" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 464 + ], + "type": "text", + "content": "[51] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 14398-14409. IEEE, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 466, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 488 + ], + "type": "text", + "content": "[52] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024.6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 490, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 555 + ], + "type": "text", + "content": "[53] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and Efficient Foundation Language Models, 2023. 5, 6, 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "text", + "content": "[54] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is All you Need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 689 + ], + "type": "text", + "content": "[55] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, Yingli Zhao, Yulong Ao, Xuebin Min, Tao Li, Boya Wu, Bo Zhao, Bowen Zhang, Liangdong Wang, Guang Liu, Zheqi He, Xi Yang, Jingjing Liu, Yonghua Lin, Tiejun Huang, and Zhongyuan Wang. Emu3: Next-Token Prediction is All You Need, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[56] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. International Journal of Computer Vision, pages 1-20, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[57] Yuancheng Wang, Haoyue Zhan, Liwei Liu, Ruihong Zeng, Haotian Guo, Jiachen Zheng, Qiang Zhang, Xueyao Zhang, Shunsi Zhang, and Zhizheng Wu. Maskgct: Zero-shot text-to-speech with masked generative codec transformer. arXiv preprint arXiv:2409.00750, 2024. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[58] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. VideoGPT: Video Generation using VQ-VAE and Transformers, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "type": "text", + "content": "[59] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 10371-10381. IEEE, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 265, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 330 + ], + "type": "text", + "content": "[60] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, Da Yin, Xiaotao Gu, Yuxuan Zhang, Weihan Wang, Yean Cheng, Ting Liu, Bin Xu, Yuxiao Dong, and Jie Tang. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer, 2024. 1, 2, 3, 5, 6, 15" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 332, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 397 + ], + "type": "text", + "content": "[61] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved VQGAN. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 399, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 487 + ], + "type": "text", + "content": "[62] Lijun Yu, José Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G. Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A. Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 3, 7, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 488, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 520 + ], + "type": "text", + "content": "[63] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An Image is Worth 32 Tokens for Reconstruction and Generation, 2024. 2, 4, 13" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 522, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 598 + ], + "type": "text", + "content": "[64] Hangjie Yuan, Shiwei Zhang, Xiang Wang, Yujie Wei, Tao Feng, Yining Pan, Yingya Zhang, Ziwei Liu, Samuel Albanie, and Dong Ni. InstructVideo: Instructing video diffusion models with human feedback. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 6463-6474. IEEE, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "text", + "content": "[65] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make Pixels Dance: High-Dynamic Video Generation, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[66] David Junhao Zhang, Jay Zhangjie Wu, Jia-Wei Liu, Rui Zhao, Lingmin Ran, Yuchao Gu, Difei Gao, and Mike Zheng Shou. Show-1: Marrying Pixel and Latent Diffusion Models for Text-to-Video Generation, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[67] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision," + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15614" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 174 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "type": "text", + "content": "ICCV 2023, Paris, France, October 1-6, 2023, pages 3813-3824. IEEE, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[68] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 174 + ], + "type": "text", + "content": "[69] Daquan Zhou, Weimin Wang, Hanshu Yan, Weiwei Lv, Yizhe Zhu, and Jiashi Feng. MagicVideo: Efficient Video Generation With Latent Diffusion Models, 2023. 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15615" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/2ff56f19-5f85-4cca-a5d4-ab26816c514c_content_list.json b/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/2ff56f19-5f85-4cca-a5d4-ab26816c514c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a0c539eff6303395a6b72e7c693175d756a6a321 --- /dev/null +++ b/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/2ff56f19-5f85-4cca-a5d4-ab26816c514c_content_list.json @@ -0,0 +1,1556 @@ +[ + { + "type": "text", + "text": "The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation", + "text_level": 1, + "bbox": [ + 122, + 128, + 875, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ho Kei Cheng Alexander Schwing \nUniversity of Illinois Urbana-Champaign", + "bbox": [ + 334, + 203, + 661, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hokeikc2,aschwing}@illinois.edu", + "bbox": [ + 382, + 243, + 604, + 255 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/38678b1cb76cbb3b6d3e7c503a10fc6a5665cb3960e9edee95567a426ef7c902.jpg", + "image_caption": [ + "Figure 1. We visualize the flows learned by different algorithms using the 8gaussians $\\rightarrow$ moons dataset. We compare flow matching (FM), minibatch optimal transport FM (OT), and our proposed conditional optimal transport FM ( $C^2 OT$ ) using one Euler step and an adaptive ODE solver, respectively. Below each plot, we show the 2-Wasserstein distance (lower is better; mean $\\pm$ std over 10 runs). For unconditional generation, OT achieves significantly straighter flows, outperforming FM. However, paradoxically, OT performs worse when conditions are introduced. This paper analyzes this degradation and finds that it occurs because optimal transport disregards conditions, leading to a train-test gap. To address this, we propose a simple fix ( $C^2 OT$ ) that outperforms FM and OT in conditional generation." + ], + "image_footnote": [], + "bbox": [ + 106, + 270, + 890, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 607, + 326, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Minibatch optimal transport coupling straightens paths in unconditional flow matching. This leads to computationally less demanding inference as fewer integration steps and less complex numerical solvers can be employed when numerically solving an ordinary differential equation at test time. However, in the conditional setting, minibatch optimal transport falls short. This is because the default optimal transport mapping disregards conditions, resulting in a conditionally skewed prior distribution during training. In contrast, at test time, we have no access to the skewed prior, and instead sample from the full, unbiased prior distribution. This gap between training and testing leads to a subpar performance. To bridge this gap, we propose conditional optimal transport $(C^2 OT)$ that adds a conditional weighting term in the cost matrix when computing the optimal transport assignment. Experiments demonstrate that this simple fix works with both discrete and continuous conditions in 8gaussians→moons,", + "bbox": [ + 88, + 643, + 485, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CIFAR-10, ImageNet-32×32, and ImageNet-256×256. Our method performs better overall compared to the existing baselines across different function evaluation budgets. Code is available at hkchengrex.github.io/C2OT.", + "bbox": [ + 511, + 608, + 908, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 705, + 645, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We focus on flow-matching-based conditional generative models, i.e., generation guided by an input condition. Examples of such conditions include class labels, text, or video. Recently, flow matching has been applied in many areas of computer vision, including text-to-image/video [9, 28], vision-language applications [3], image restoration [27], and video-to-audio [4]. However, these methods can be slow at", + "bbox": [ + 509, + 734, + 908, + 843 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1We use condition to denote an input condition, not to be confused with the 'condition' in conditional flow matching (CFM) [23] (where the second 'C' in our acronym comes from), which refers to a data sample. To avoid ambiguity, we refer to CFM simply as flow matching (FM) in this paper.", + "bbox": [ + 511, + 851, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15875", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "test-time – obtaining a solution requires numerically integrating the flow with an ODE solver, typically involving many steps, each requiring a deep network forward pass. Naively reducing the number of steps degrades performance, as the underlying flow is often curved (Figure 1, leftmost column), necessitating a small step size for accurate numerical integration. One way to address this issue is by straightening the flow. In the context of unconditional generation, Tong et al. [35] and Pooladian et al. [29] concurrently proposed \"minibatch optimal transport\" (OT), which deterministically couples data and samples from the prior within a minibatch via optimal transport (replacing random coupling) to minimize flow path lengths and to straighten the flow.", + "bbox": [ + 88, + 90, + 482, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While OT improves unconditional generation, we find that it paradoxically and consistently harms conditional generation (Figure 1). This occurs because OT disregards the conditions when computing the coupling. As a result, during training, OT samples from a prior distribution that is skewed by the condition, as we show in Section 3.2. However, at test time, we have no access to this skewed distribution and instead sample from the full prior distribution. This mismatch creates a gap between training and testing, leading to performance degradation. To bridge this gap, we propose conditional optimal transport FM $(\\mathrm{C}^2\\mathrm{OT})$ which introduces a simple yet effective condition-aware weighting term when computing the cost matrix for OT. Additionally, we propose two techniques: adaptive weight finding to simplify hyperparameter tuning, and efficient oversampling of OT batches to counteract the reduced effective OT batch size introduced by the weighting term.", + "bbox": [ + 91, + 287, + 485, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments on both two-dimensional synthetic data and high-dimensional image data, including CIFAR-10, ImageNet-32 $\\times$ 32, and ImageNet-256 $\\times$ 256. The results demonstrate that our method performs well across different condition types (discrete and continuous), datasets, network architectures (UNets and transformers), and data spaces (image space and latent space). $C^2\\mathrm{OT}$ achieves better overall performance than the existing baselines across different inference computation budgets, including with few-steps Euler's method and with an adaptive ordinary differential equation (ODE) solver [8].", + "bbox": [ + 89, + 545, + 485, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 734, + 241, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Flow Matching. Recently, flow matching [1, 2, 23] has become a popular choice for generative modeling, in part due to its simple training objective and its ability to generate high-quality samples. However, at test time, flow-based methods typically require many computationally expensive forward passes through a deep net to numerically integrate the flow. This is because the underlying flow is usually curved and therefore cannot be approximated well with a few integration steps.", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Optimal Transport Coupling. To straighten the flow, in the context of unconditional generation, Tong et al. [35] and Pooladian et al. [29] concurrently proposed minibatch optimal transport (OT). While OT has proven effective in the unconditional setting, we show in Section 3.2 that it skews the prior distribution and fails in the conditional setting if used naively. Our method, $\\mathrm{C}^2\\mathrm{OT}$ , specifically addresses this failure, aiming to extend the success of OT in unconditional generation to conditional generation. Note, OT has also been used in equivariant flow matching [18, 34], which exploits domain-specific data symmetry (e.g., in molecules). In contrast, our method targets general data. Further, Hui et al. [15] find that mixing of OT and independent coupling benefits shape completion, which is an orthogonal contribution to this paper. Concurrently, Davtyan et al. [6] improve data-noise coupling for unconditional generation, while leaving data coupling for conditional generation to future work.", + "bbox": [ + 511, + 90, + 906, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Straightening Flow Paths. Other methods of straightening flow paths exist. Reflow [25] achieves this by retraining a flow matching network multiple times, at the cost of additional training overhead. Variational flow matching [12] reduces flow ambiguity with a latent variable and hierarchical flow matching [40] straightens flows by modeling higher-order flows. These approaches are orthogonal to our focus on prior-data coupling and, in principle, can be combined with our method.", + "bbox": [ + 511, + 349, + 908, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning the Prior Distribution. An alternative approach to improving flow is to jointly learn a prior distribution with the flow matching network. However, these methods [22, 24, 32] typically rely on a variational autoencoder (VAE) [14, 17] to learn the prior. This introduces yet another density estimation problem, which complicates training. In this work, we focus on improving training-free coupling plans without learning a new prior.", + "bbox": [ + 511, + 488, + 908, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Consistency Models. Consistency models [33] have been proposed to accelerate diffusion model sampling, and can be extended to flow matching [38]. Recent advancements have improved training efficiency [11], stability [26], and quality [16]. These methods are orthogonal to our contribution, which focuses on improving prior-data coupling, and can be integrated, as demonstrated by Silvestri et al. [32] in their combination of consistency models with OT. In this paper, we focus on the base form of flow matching to avoid distractions from other formulations.", + "bbox": [ + 511, + 612, + 908, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 784, + 604, + 799 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 814, + 666, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Flow Matching (FM). We base our discussion on FM [23, 35] (also called rectified flow [25]) for generative modeling and refer readers to [25, 35] for details. At test-time, a sample $\\tilde{x}_1$ is characterized by an ordinary differential equation", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15876", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/95849445438a1dfb8948167facc0d9d74ca8c25a9fa9634fd62430bcfaca1c11.jpg", + "image_caption": [ + "Figure 2. We visualize the coupling during training and the learned flow from $t = 0$ to $t = 1$ during testing for FM, OT, and our proposed method $C^2 OT$ . The prior is defined as $q(x_0) = \\mathcal{N}(0,1)$ , while the target distribution is given by $q(x_{1}) = \\frac{1}{2} q(x_{1}|c = 0) + \\frac{1}{2} q(x_{1}|c = 1)) = \\frac{1}{2}\\mathcal{N}(-2,0.5) + \\frac{1}{2}\\mathcal{N}(2,0.5)$ . Note, FM results in curved flows during testing, while OT degenerates due to skewed training samples ( $q(x_0|c) \\neq q(x_0)$ ). In contrast, our method successfully learns straight flows without degeneration." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 370, + 277 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c0f240adc569ff6e5bdcd6d1c715f4ddad481eab94b6fb224163c0c69f1ff9fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 88, + 638, + 277 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/88eea11d8895ec1f0a725be32cf2417094fde3431a1482f27456da5547c6ece1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 88, + 898, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(ODE) with an initial boundary value specified via noise $x_0$ , randomly drawn from a prior distribution (e.g., the standard normal). To compute the sample $\\tilde{x}_1$ , this ODE is commonly solved by numerically integrating from time $t = 0$ to time $t = 1$ a learned flow/velocity field $v_{\\theta}$ , where $\\theta$ denotes the set of trainable deep net parameters. For conditional generation, we additionally obtain a condition $c$ , e.g., a class label from user input, which adjusts the flow. We use $v_{\\theta,c}$ to denote this conditional flow. At training time, we find the deep net parameters $\\theta$ by minimizing the objective", + "bbox": [ + 88, + 347, + 483, + 500 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {t, \\left(x _ {0}, x _ {1}, c\\right) \\sim \\pi \\left(x _ {0}, x _ {1}, c\\right)} \\| v _ {\\theta , c} (t, x _ {t}) - u \\left(x _ {t} \\mid x _ {0}, x _ {1}\\right) \\| ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 511, + 483, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t$ is uniformly drawn from the interval $[0,1]$ , and where $\\pi (x_0,x_1,c)$ denotes the joint distribution of the prior and the training data. In practice, often, independent coupling is used, i.e., the prior and the training data are sampled independently, such that", + "bbox": [ + 88, + 541, + 483, + 617 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\pi \\left(x _ {0}, x _ {1}, c\\right) = q \\left(x _ {0}\\right) q \\left(x _ {1}, c\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 632, + 482, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Further,", + "bbox": [ + 89, + 662, + 147, + 676 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = t x _ {1} + (1 - t) x _ {0} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 680, + 482, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "defines a linear interpolation between noise and data, and", + "bbox": [ + 89, + 704, + 470, + 720 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nu \\left(x _ {t} \\mid x _ {0}, x _ {1}\\right) = x _ {1} - x _ {0} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 734, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "denotes its corresponding flow velocity at $x_{t}$ . Note, simply dropping the condition $c$ leads to the unconditional flow matching formulation.", + "bbox": [ + 89, + 763, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although rectified flow matching is trained with straight paths, the learned flow is often curved (Figure 1, bottom-left) since there are many possible paths induced by the independently sampled couplings $x_0, x_1$ through any $x_t$ at time $t$ [29]. Curved paths require more numerical integration and network evaluation steps, which is not ideal.", + "bbox": [ + 88, + 810, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimal Transport (OT) Coupling. To address this issue for unconditional generation, Tong et al. [35] and Pooladian et al. [29] concurrently proposed to use minibatch optimal transport to deterministically couple the sampled prior and a sampled batch of the data. This minimizes flow path lengths and straightens the flow. Concretely, given a minibatch of $b$ $e$ -dimensional samples, an OT coupling seeks a $b \\times b$ permutation matrix $P$ such that the squared Euclidean distance is minimized, i.e.,", + "bbox": [ + 511, + 347, + 906, + 482 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {P} \\| X _ {0} - P X _ {1} \\| _ {2} ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 489, + 906, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $X_0, X_1 \\in \\mathbb{R}^{b \\times e}$ are matrices containing a minibatch of samples from the prior and the data. Instead of an \"independent coupling\", in the unconditional setting we then optimize a variant of Equation (1) using the coupled tuple $(X_0, PX_1)$ . Note that this algorithm corresponds to sampling tuples $(x_0, x_1)$ from a joint distribution $\\pi_{\\mathrm{ot}}(x_0, x_1)$ .", + "bbox": [ + 511, + 520, + 906, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Importantly, for unconditional generation, i.e., when using the joint distribution $\\pi_{\\mathrm{ot}}(x_0,x_1)$ , the marginals remain unchanged from independent coupling (see [35]):", + "bbox": [ + 511, + 612, + 905, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\int \\pi_ {\\mathrm {o t}} (x _ {0}, x _ {1}) q (x _ {0}) \\mathrm {d} x _ {0} = q (x _ {1}), \\text {a n d}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 664, + 834, + 694 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\int \\pi_ {\\mathrm {o t}} \\left(x _ {0}, x _ {1}\\right) q \\left(x _ {1}\\right) \\mathrm {d} x _ {1} = q \\left(x _ {0}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 696, + 906, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hence, during test-time, we can sample from the prior $q(x_0)$ and integrate the flow $v_{\\theta}$ to simulate data from the data distribution $q(x_{1})$ .", + "bbox": [ + 511, + 734, + 905, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unfortunately, as shown in the next section, this does not hold for conditional generation. Said differently, naively extending the unconditional joint distribution $\\pi_{\\mathrm{ot}}(x_0,x_1)$ to a conditional joint distribution $\\pi_{\\mathrm{ot}}(x_0,x_1,c)$ by using the coupled tuple $(X_0,PX_1,PC)$ , where $C\\in \\mathbb{R}^{b\\times h}$ is a batch of $h$ -dimensional conditions, does not maintain marginals and leads to a gap between training time and test time. We discuss this next.", + "bbox": [ + 511, + 780, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15877", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. OT Coupling Skews The Prior", + "text_level": 1, + "bbox": [ + 89, + 89, + 382, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intuition. Consider the one-dimensional example illustrated in Figure 2: the prior distribution $q(x_0)$ follows a Gaussian distribution, and the target distribution $q(x_1)$ consists of a mixture of two Gaussians translated left and right, with the translation direction indicated by a binary condition $c$ . When training with a $\\pi_{\\mathrm{ot}}(x_0,x_1,c)$ that represents the coupled tuple $(X_0, PX_1,PC)$ , samples from the prior and the sampled conditions become correlated through the OT coupling – samples from the left half of the prior distribution are always associated with $c = 0$ , while those from the right half always correspond to $c = 1$ . As a result, the model overfits on data pairs $(x_0,c)$ where either $(x_0 < 0,c = 0)$ or $(x_0 > 0,c = 1)$ . However, during testing, this skewed distribution is no longer accessible, and instead, $x_0$ and $c$ are sampled independently. Consequently, the model encounters previously unseen data pairs, i.e., $(x_0 > 0,c = 0)$ and $(x_0 < 0,c = 1)$ – leading to failure. A similar phenomenon can be observed in the two-dimensional example illustrated in Figure 1.", + "bbox": [ + 89, + 114, + 483, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Skewed Prior. Formally, imagine that we consider the most general OT coupling joint distribution $\\pi_{\\mathrm{ot}}(x_0,x_1,c)$ at training time. At test time, the user wants to specify a condition $c_{1}$ to obtain samples from the conditional distribution $q(x_{1}|c = c_{1})$ . What is the prior distribution that we need to sample from such that we arrive at the correct conditional distribution $q(x_{1}|c = c_{1})$ ? To compute this, imagine, we are given an arbitrary condition $c_{1}$ , i.e., we look at the induced distribution $\\pi_{\\mathrm{ot}}(x_0,x_1,c|c = c_1)$ . Marginalizing this induced distribution over $x_{1}$ yields", + "bbox": [ + 89, + 402, + 483, + 554 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\int \\pi_ {\\mathrm {o t}} \\left(x _ {0}, x _ {1}, c \\mid c = c _ {1}\\right) \\mathrm {d} x _ {1} = \\int \\pi_ {\\mathrm {o t}} \\left(x _ {0}, x _ {1} \\mid c = c _ {1}\\right) \\mathrm {d} x _ {1} \\\\ = q \\left(x _ {0} \\mid c = c _ {1}\\right), \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 561, + 483, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which implies that the prior required to arrive at the conditional distribution $q(x_{1}|c = c_{1})$ is $q(x_{0}|c = c_{1})$ - a distribution that we cannot access at test time. In general, the prior that we need to sample from, given condition $c$ , is skewed from $q(x_{0})$ to $q(x_{0}|c)$ which is inaccessible at test time. As a result, accurately capturing $q(x_{1}|c)$ at test time becomes infeasible unless $q(x_{0})$ and $q(c)$ are independent, i.e., if $q(x_{0}|c) = q(x_{0})$ . However, this condition is generally not satisfied by the most general OT coupling – for instance, in Figure 2, clearly $q(x_{0}|c) \\neq q(x_{0})$ . To address this issue, we propose a method $C^2\\mathrm{OT}$ to unkew the prior. We discuss this in the next section.", + "bbox": [ + 89, + 619, + 483, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Conditional Optimal Transport FM", + "text_level": 1, + "bbox": [ + 89, + 819, + 428, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Formulation", + "text_level": 1, + "bbox": [ + 89, + 845, + 223, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose conditional optimal transport FM (C²OT) to unkew the prior $q(x_0|c)$ such that $q(x_0|c) = q(x_0)$ . This", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ensures that at test time, we can sample from the full prior $q(x_0)$ , irrespective of the condition $c$ . Importantly, at the same time, we also aim to preserve the straight flow paths provided by OT. Conceptually, we construct a prior distribution independently for each condition, i.e., $q(x_0|c_1) = q(x_0|c_2) = q(x_0), \\forall c_1, c_2$ . This can be achieved by sampling from the prior and computing OT independently for each condition, as shown in Figure 2 (right).", + "bbox": [ + 511, + 90, + 906, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, we construct the joint distribution as", + "bbox": [ + 532, + 212, + 841, + 226 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {\\mathrm {c} ^ {2} \\mathrm {o t}} (x _ {0}, x _ {1}, c) := q (x _ {0}) q (c) \\pi_ {\\mathrm {o t}} (x _ {1} | x _ {0}, c). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 238, + 906, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, the prior $q(x_0)$ and the condition $q(c)$ are sampled independently, while the data $x_1$ is conditioned on $x_0$ (via OT) and $c$ (via the dataset construction). Different from independent coupling in Equation (2), $x_1$ and $x_0$ are associated through OT and therefore lead to straighter flow paths. Also different from the most general OT coupling $\\pi_{\\mathrm{ot}}(x_0,x_1,c)$ , we explicitly enforce the independence between $x_0$ and $c$ .", + "bbox": [ + 511, + 263, + 906, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To see that this joint distribution provides an unskewed prior, imagine we are given an arbitrary input condition $c_{1}$ , marginalizing over $x_{1}$ yields", + "bbox": [ + 511, + 369, + 906, + 415 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\int q (x _ {0}) q (c | c = c _ {1}) \\pi_ {\\mathrm {o t}} \\left(x _ {1} \\mid x _ {0}, c = c _ {1}\\right) \\mathrm {d} x _ {1} \\\\ = \\int q \\left(x _ {0}\\right) \\pi_ {\\mathrm {o t}} \\left(x _ {1} \\mid x _ {0}, c = c _ {1}\\right) \\mathrm {d} x _ {1} = q \\left(x _ {0}\\right). \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 425, + 906, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This implies that, at test time, we can sample from the entire prior $q(x_0)$ , regardless of the condition $c$ , to arrive at the desired data distribution $q(x_1|c)$ . During training, in practice, we sample from $\\pi_{\\mathrm{c}^2\\mathrm{ot}}$ by modifying the optimal transport cost function in Equation (5). This process is exact for discrete conditions (Section 3.3.2) and approximate for continuous conditions (Section 3.3.3). Specifically, given a minibatch of $b$ $e$ -dimensional samples and a minibatch of $b$ $h$ -dimensional conditions $C \\in \\mathbb{R}^{b \\times h}$ , $\\mathbf{C}^2\\mathrm{OT}$ seeks a $b \\times b$ permutation matrix $P$ that minimizes the following cost function:", + "bbox": [ + 511, + 496, + 905, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {P} \\| X _ {0} - P X _ {1} \\| _ {2} ^ {2} + \\sum_ {i = 1} ^ {b} f \\left(C _ {i}, [ P C ] _ {i}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 657, + 906, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f$ is a symmetric and non-negative cost function, satisfying the property $f(c, c) = 0 \\forall c$ . The key differences between independent coupling, OT coupling, and $\\mathrm{C}^2\\mathrm{OT}$ coupling are highlighted in Algorithms 1 to 3. In the following two sections, we discuss our choice of $f$ for both discrete and continuous conditions.", + "bbox": [ + 511, + 707, + 905, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Discrete Conditions", + "text_level": 1, + "bbox": [ + 511, + 816, + 696, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For discrete conditions, we assume the conditions correspond to class labels from a finite, discrete set, i.e., $c \\in \\{0, 1, \\ldots, k - 1\\}$ . For example, in Figure 2, the labels 0 and 1 represent the left and right groups in the target distribution,", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15878", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 1 Independent Coupling \n1: function INDEPENDENT(X1, C) \n2: X0 ~ N ▷ Sample from prior \n3: \n4: \n5: return X0, X1, C", + "bbox": [ + 93, + 83, + 349, + 189 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 2 OT Coupling \n1: function OT(X1, C) \n2: X0 ~ N ▷ Sample from prior \n3: Cost ← pairwise_dist(X0, X1) \n4: i, j ← Hungarian_MATCHing(Cost) \n5: return X0[i], X1[j], C[j]", + "bbox": [ + 369, + 83, + 625, + 189 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 3 $\\mathrm{C}^2\\mathrm{OT}$ Coupling (Ours) \n1: function $\\mathrm{C}^2\\mathrm{OT}(X_1,C)$ \n2: $X_0\\sim \\mathcal{N}$ ▷ Sample from prior \n3: Cost $\\leftarrow$ pairwise_dist(Xo,X1)+ pairwise_dist(C,C) \n4: $i,j\\gets$ Hungarian_MATCHing(Cost) \n5: return $X_0[i],X_1[j],C[j]$", + "bbox": [ + 643, + 83, + 900, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "respectively. We define $f = f_{\\mathrm{d}}$ such that no transport is allowed to modify the conditions (i.e., $PC = C$ ). This is achieved via", + "bbox": [ + 89, + 202, + 482, + 247 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {d}} \\left(c _ {1}, c _ {2}\\right) = \\left\\{ \\begin{array}{l} \\infty , c _ {1} \\neq c _ {2}, \\\\ 0, \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 257, + 483, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a result, the unordered set of prior samples for any given condition $c$ remains unchanged: $\\{X_{0i}|c_i = c\\} \\stackrel{\\mathrm{iid}}{\\sim} q(x_0)$ . This formulation samples exactly from our ideal (Equation (8)), as OT is computed independently within each class, and the data sample from each class is exposed to the full prior without skew. This effect is illustrated in Figure 1 (middle) and Figure 2 (right). As expected, the results differ from FM's curved flow paths and OT's inability to properly capture the target distribution.", + "bbox": [ + 88, + 309, + 483, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is important to note that enforcing $PC = C$ does not necessarily imply $P = I$ since we can still swap row $i$ and $j$ in $P$ as long as $c_{i} = c_{j}$ . Setting $P = I$ would correspond to a degeneration back to independent coupling in Equation (2) and forfeit the benefits of straightened flow provided by OT.", + "bbox": [ + 89, + 450, + 483, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unfortunately, using $f_{\\mathrm{d}}$ with continuous conditions leads to exactly this degeneration, since, in general, no $c_{i} = c_{j}$ unless $i = j$ . Hence, in the next section, we devise a relaxed cost for handling continuous conditions.", + "bbox": [ + 89, + 526, + 483, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3. Continuous Conditions", + "text_level": 1, + "bbox": [ + 89, + 609, + 297, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For continuous conditions, the condition is typically a feature embedding, e.g., computed from a text prompt. Directly applying $f_{\\mathrm{d}}$ in Equation (11) leads to a degeneration back to independent coupling as discussed earlier. To avoid this, we propose a relaxed penalty function $f_{\\mathrm{c}}$ based on the cosine distance:", + "bbox": [ + 89, + 633, + 483, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {c}} \\left(c _ {1}, c _ {2}\\right) = w \\left(1 - \\frac {c _ {1} \\cdot c _ {2}}{\\| c _ {1} \\| \\| c _ {2} \\|}\\right) = w \\cdot \\operatorname {c d i s t} \\left(c _ {1}, c _ {2}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 734, + 483, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $w > 0$ is a scaling hyperparameter. Note, when $w = 0$ , this formulation degenerates to regular OT; when $w \\to \\infty$ , this formulation approaches Equation (11), i.e., independent coupling if no two conditions are equal. The right section of Figure 1 illustrates an example where the $x$ -coordinate of the target data point serves as the condition. Notably, $C^2\\mathrm{OT}$ preserves straight flow paths without exhibiting OT's apparent degradation.", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finding $w$ . The optimal $w$ is highly dependent on the data distribution and the magnitude of features. To alleviate the need for tuning the hyperparameter $w$ , we propose to find $w$ adaptively for each minibatch. For this, we develop a ratio $r(w)$ which measures the proportion of samples that are considered \"potential optimal transport candidates\", i.e.,", + "bbox": [ + 511, + 202, + 906, + 294 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} r (w) = \\frac {1}{B ^ {2}} \\sum_ {i, j} \\mathbb {1} \\left(\\left\\| X _ {0 i} - X _ {1 j} \\right\\| _ {2} ^ {2} + w \\cdot \\operatorname {c d i s t} \\left(c _ {i}, c _ {j}\\right) \\right. \\\\ \\leq \\left\\| X _ {0 i} - X _ {1 i} \\right\\| _ {2} ^ {2}), \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 305, + 906, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ is an indicator function that evaluates to one if the condition is satisfied and to zero otherwise. Then, we introduce a hyperparameter $r_{\\mathrm{tar}}$ as the target ratio and find $w$ such that $r(w) \\approx r_{\\mathrm{tar}}$ . Note that setting $r_{\\mathrm{tar}}$ is invariant to the scaling of distances between $x_0$ and $x_1$ . Namely, if $\\|X_{0i} - X_{1j}\\|_2^2$ is scaled by $s$ , $w$ would also be scaled by $s$ to preserve the same $r_{\\mathrm{tar}}$ . This invariance is particularly useful when transitioning to a new dataset, such as changing image resolutions in image generation tasks. In our implementation, we search for $w$ in two steps: first with exponential search to establish the upper/ lower bounds, then with binary search to locate a precise value. The final $w$ is used as the initial value in the subsequent minibatch. In practice, we observe $w$ differs very little between minibatches ( $< 1\\%$ ) and the search converges quickly within 10 iterations with a negligible overhead. Next, we introduce oversampling, a simple technique to obtain more accurate OT couplings.", + "bbox": [ + 511, + 372, + 908, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.4. Oversampling OT Batches", + "text_level": 1, + "bbox": [ + 511, + 647, + 743, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "OT vs. Deep Net Batch Sizes. Typically, the \"OT batch size\" $b$ used to compute optimal transport is set equal to the \"deep net batch size\" $b_{n}$ used to compute forward/backward passes of the deep net [29, 35]. While the OT batch size controls the dynamics of prior-to-data assignments, the deep net batch size affects gradient variances, training efficiency, and memory usage. There is no reason why setting them equal should be optimal. Particularly, with our proposed $C^2\\mathrm{OT}$ , the effective OT batch size is reduced since we restrict transport between samples with divergent conditions, which calls for an increased OT batch size $b$ . At the same time, keeping the deep net batch size $b_{n}$ unchanged prevents unwarranted side effects in other aspects of training.", + "bbox": [ + 509, + 671, + 906, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reduced Effective OT Batch Size. In the case of discrete conditions, OT is effectively performed within", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15879", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/36e58a6faa0c5d2dead8cc694df8a17251a414385c639f00d23af41b9c046cf7.jpg", + "table_caption": [ + "Table 1. Results on the 8gaussians $\\rightarrow$ moons dataset with different training coupling methods and different conditioning. Note that our method $C^2\\mathrm{OT}$ does not apply in the unconditional setting." + ], + "table_footnote": [], + "table_body": "
MethodEuler-1 (W2↓)Adaptive (W2↓)NFE ↓
Unconditional
FM6.232±0.1860.120±0.04593.22±1.81
OT0.072±0.0250.050±0.04036.08±2.57
Binary conditions
FM2.573±0.1120.059±0.02991.10±2.77
OT0.483±0.0590.462±0.02532.89±0.95
C2OT (ours)0.048±0.0100.018±0.00658.88±1.48
Continuous conditions
FM0.732±0.0520.028±0.01093.86±1.76
OT8.276±6.5102.143±1.99342.07±2.59
C2OT (ours)0.077±0.0240.013±0.00389.87±1.53
", + "bbox": [ + 91, + 133, + 478, + 343 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\{X_{0i}, X_{1i} | c_i = c\\}$ for each condition $c$ . With $k$ classes, the expected effective OT batch size is reduced to $\\mathbb{E}(|\\{i | c_i = c\\}|) = b / k$ . Empirically, we find that using a much larger OT batch size $b$ than the network batch size $b_n$ is helpful. This observation motivates oversampling.", + "bbox": [ + 89, + 353, + 483, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Efficiency. A natural concern with increasing $b$ is the computational overhead, since optimal transport with Hungarian matching [20] has a time complexity of $O(b^{3})$ . However,", + "bbox": [ + 89, + 434, + 483, + 479 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Since each OT batch can be used across $(b / b_{n})$ forward/backward passes, the amortized cost per minibatch reduces to $O(b^{2}b_{n})$ .", + "- We offload OT computation from the main process to the data workers, enabling data preparation on the CPU in parallel with forward/backward passes on the GPU." + ], + "bbox": [ + 89, + 479, + 483, + 570 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With a modest choice of 8 data workers per GPU, we observe no wall-clock overhead for OT batch sizes up to 6,400 when training on CIFAR-10 and ImageNet, as data preparation is completed faster than the GPU can process batches.", + "bbox": [ + 89, + 570, + 483, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 654, + 223, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We experimentally verify our proposed method $\\mathrm{C}^2\\mathrm{OT}$ in two sections: first on the two-dimensional synthetic dataset 8gaussians $\\rightarrow$ moons, and then on high-dimensional image data, including CIFAR-10 [19] and ImageNet-1K [7]. For ImageNet-1K, we conduct experiments on both $32\\times 32$ images in image space and $256\\times 256$ images in latent space. Implementation details are provided in the appendix.", + "bbox": [ + 89, + 686, + 485, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Two-Dimensional Data", + "text_level": 1, + "bbox": [ + 89, + 814, + 321, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The two-dimensional 8gaussians $\\rightarrow$ moons dataset maps a mixture of eight Gaussian distributions to two interleaving half-circles, as visualized in Figure 1. We consider three conditioning scenarios: (a) unconditional; (b) binary", + "bbox": [ + 89, + 840, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "class conditions, where each class corresponds to one of the half-circles; and (c) continuous conditions, where the $x$ -coordinate of the desired target point is given as the condition. For (b), we apply our setup for discrete conditions (Section 3.3.2); for (c), we apply our setup for continuous conditions (Section 3.3.3) with $r_{\\mathrm{tar}} = 0.01$ .", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We report the 2-Wasserstein distance $(W_2^2)$ (following [35]) between 10K generated data points and 10K samples from the ground-truth moons distribution. We experiment with single-step Euler's method (Euler-1) and the adaptive Dormand-Prince method [8] (dopri5, referred to as adaptive) for numerical integration. Additionally, we report the number of function evaluations (NFE) in the adaptive method, which is the number of forward passes through the neural network. We average the results over ten training runs with different random seeds and report mean±std.", + "bbox": [ + 511, + 184, + 908, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. The results are summarized in Table 1 and visualized in Figure 1. In the unconditional setting, OT produces straighter flows, leading to better distribution matching with fewer NFEs, and our method does not apply. However, in the conditional setting, OT degrades and performs worse than regular FM in the adaptive setting due to the aforementioned train-test gap in OT coupling. Our method, $\\mathrm{C}^2\\mathrm{OT}$ , effectively models the target distribution with both Euler-1 and the adaptive method, improving upon the baselines.", + "bbox": [ + 511, + 337, + 908, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. High-Dimensional Image Data", + "text_level": 1, + "bbox": [ + 511, + 493, + 808, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on CIFAR-10 [19] (image space, class-conditioned), ImageNet- $32 \\times 32$ [7] (image space, caption-conditioned), and ImageNet- $256 \\times 256$ [7] (latent space, caption-conditioned). For ImageNet, we use the captions provided by [21] and encode text features using a CLIP [30]-like DFN [10] text encoder as input conditions. Our results are summarized in Table 2.", + "bbox": [ + 511, + 518, + 906, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We report Fréchet inception distance (FID) [13] and the number of function evaluations (NFE) in the adaptive solver for all three datasets. To assess how well the generations adhere to the input conditions, we additionally compute condition adherence metrics. For CIFAR-10, we run a pretrained classifier [5] on the generated images to obtain logits and report the average cross entropy (CE) with the input class label. For ImageNet, we compute CLIP features [30] using SigLIP-2 [36] from the generated images and the input captions respectively, and report the average cosine similarities (CLIP).", + "bbox": [ + 511, + 626, + 908, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIFAR-10. We base our UNet architecture on Tong et al. [35], with details provided in the appendix. The CIFAR-10 training set contains 50,000 $32 \\times 32$ color images across 10 object classes. During testing, we generate 50,000 images evenly distributed among classes and assess FID against the training set following [35]. Each model is trained five times with different random seeds and the results are reported", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15880", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/85472c3671ea35873ca2343477c6da6ab0131c3f22520e50b17ca48b1aa4cc8d.jpg", + "table_caption": [ + "Table 2. Results of different training algorithms on image generation tasks. We bold the best-performing entry and underline the second-best. Our proposed method, $C^2 OT$ , achieves the best overall performance – better than FM with few sampling steps and better than OT with more sampling steps. In cases where $C^2 OT$ performs second, our performance is often comparable to the best entry. Additionally, $C^2 OT$ has the most stable performance, with the smallest deviations across runs in most entries. Variations in CLIP scores are minimal." + ], + "table_footnote": [], + "table_body": "
CIFAR-10 Class-Conditioned Generation
MethodEuler-2Euler-5Euler-10Euler-25Euler-50Euler-100AdaptiveNFE ↓
FID → FM105.481±1.61921.968±0.40010.479±0.2375.640±0.1174.128±0.0693.429±0.0542.922±0.039124.0±2.5
OT64.417±0.43418.194±0.40910.778±0.2706.821±0.1865.368±0.1134.671±0.0664.144±0.018125.7±0.4
C²OT (ours)64.694±0.37917.565±0.2029.762±0.1365.644±0.0814.152±0.0613.447±0.0492.904±0.035127.7±0.7
CE → FM3.343±0.0220.546±0.0200.323±0.0070.272±0.0040.267±0.0020.269±0.0030.276±0.004124.0±2.5
OT2.529±0.0100.624±0.0120.426±0.0050.369±0.0040.363±0.0040.364±0.0050.372±0.005125.7±0.4
C²OT (ours)2.264±0.0220.475±0.0090.322±0.0040.276±0.0050.271±0.0030.272±0.0030.279±0.004127.7±0.7
ImageNet 32×32 Caption-Conditioned Generation
FID → FM113.250±0.79323.015±0.12311.141±0.0517.165±0.0876.142±0.0855.673±0.0745.358±0.059146.9±3.5
OT81.317±0.36920.763±0.15411.360±0.0897.854±0.0776.899±0.0796.469±0.0716.220±0.065137.7±0.2
C²OT (ours)102.380±0.27921.965±0.03510.897±0.0337.069±0.0276.084±0.0165.638±0.0185.350±0.017134.2±1.4
CLIP → FM0.067±0.0000.076±0.0000.074±0.0000.071±0.0000.071±0.0000.070±0.0000.069±0.000146.9±3.5
OT0.067±0.0000.071±0.0000.070±0.0000.068±0.0000.068±0.0000.067±0.0000.067±0.000137.7±0.2
C²OT (ours)0.068±0.0000.075±0.0000.073±0.0000.071±0.0000.070±0.0000.070±0.0000.069±0.000134.2±1.4
ImageNet 256×256 Caption-Conditioned Generation in Latent Space
FID → FM202.519±0.60631.266±0.3559.996±0.1065.189±0.0823.886±0.0663.484±0.0553.498±0.054131.1±2.3
OT193.376±1.66542.780±2.13215.638±1.5107.514±1.0345.693±0.9985.360±1.0165.726±1.074125.6±6.5
C²OT (ours)200.811±0.38530.715±0.22110.000±0.0725.069±0.0373.722±0.0373.320±0.0323.369±0.034128.2±2.6
CLIP → FM0.028±0.0000.120±0.0000.134±0.0000.137±0.0000.138±0.0000.139±0.0000.140±0.000131.2±2.3
OT0.031±0.0000.107±0.0030.120±0.0030.123±0.0030.124±0.0030.124±0.0040.125±0.004125.6±6.5
C²OT (ours)0.029±0.0000.120±0.0000.133±0.0000.137±0.0000.138±0.0000.138±0.0000.139±0.000128.2±2.6
", + "bbox": [ + 91, + 148, + 906, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as mean±std. We use an OT batch size $b$ of 640 and a ratio $r_{\\mathrm{tar}}$ of 0.01. Compared to FM, our method achieves better performance (FID and CE) in few-steps settings and comparable performance in many-steps settings. Compared to OT, our method significantly outperforms in the many-step setting. Although OT achieves slightly better FID in Euler-2, it performs worse in CE, indicating that it fails to follow the condition.", + "bbox": [ + 88, + 526, + 480, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ImageNet $32 \\times 32$ . We base our UNet architecture on Pooladian et al. [29], and list details in the appendix. We train and evaluate on the face-blurred [37] ImageNet, following [29]. In contrast to prior works that assess FID against the training set, we assess FID against the validation set (49,997 images), as the fine-grained nature of captions may lead to overfitting on the training set. Each model is trained three times with different random seeds and the results are reported as mean $\\pm$ std. Compared to CIFAR-10, there are much more variations in this dataset (e.g., 1000 classes vs. 10 classes in CIFAR-10). Thus, we adopt a larger OT batch size $b$ of 6400 while keeping the same target ratio $r_{\\mathrm{tar}}$ 0.01 as before. Similar to our findings in CIFAR-10, our method performs better or is comparable to FM; in few-steps settings, OT achieves better FID but worse condition following (CLIP). Our method strikes the best overall balance.", + "bbox": [ + 89, + 657, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ImageNet $256 \\times 256$ in Latent Space. Here, we explore latent flow matching models [31]. During training, the flow matching model learns to generate data in the latent space, which is encoded from images. At test-time, the generated latents are decoded into images using a pretrained decoder. We base our experiment on a recent transformer-based model, LightningDiT [39], demonstrating that our method is effective across different network architectures and data spaces. We use the officially provided \"64 epochs\" configuration for all experiments (due to computational constraints), which is not directly comparable to methods trained with 800 epochs. We follow the same evaluation protocol as ImageNet- $32 \\times 32$ , and adopt the same OT batch size $b$ of 6400 and the target ratio $r_{\\mathrm{tar}}$ of 0.01. Each model is trained five times with different random seeds, and the results are reported as mean $\\pm$ std. Again, $\\mathrm{C}^2\\mathrm{OT}$ has the best overall performance compared to both FM and OT. We visualize the generations in Figure 3.", + "bbox": [ + 511, + 526, + 908, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Limitations in High-Dimensional Data We observed that the benefits of $\\mathrm{C}^2\\mathrm{OT}$ are less pronounced, though still visible, in higher-dimensional image data compared to low-dimensional data. We hypothesize this is because image data is more entangled, making the prior skew induced by OT less detrimental. We also observe statistically straighter paths in $\\mathrm{C}^2\\mathrm{OT}$ compared to FM, with details given in the appendix.", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15881", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1bdbf61a6d1473856e6f3ed0ddb35d54c0dcb10afd6a447fdf531862d4a28611.jpg", + "image_caption": [ + "Figure 3. Visual comparisons of $256 \\times 256$ images generated by the baselines and our approach with different amounts of sampling steps. Our approach converges faster and produces cleaner outputs. The input caption is \"a small dog sitting on a carpet\"." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 483, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 311, + 269, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.1. Varying OT Batch Size $b$", + "text_level": 1, + "bbox": [ + 89, + 340, + 308, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, we analyze the effect of varying the OT batch size $b$ . We conduct these experiments on the CIFAR-10 dataset, training each model three times with different random seeds for every choice of $b$ . We plot the results in Figure 4. Increasing the OT batch size clearly improves FID in the few-sampling-steps setting (Euler-2). This is intuitive, because a larger OT batch size enables the coupling to approach the true OT results, i.e., dataset-level optimal transport. In turn, this leads to straighter paths. However, in the adaptive setting, increasing the OT batch size slightly worsens FID, though the effect is much less significant (the range of the blue y-axis is much smaller than the range of the red y-axis). We think this occurs because a more accurate OT coupling reduces variations in the training data (similar to having less data augmentations) which harms performance as the benefit of straighter flows diminishes in the adaptive setting. This observation aligns with the findings of Pooladian et al. [29]: computing OT coupling across GPUs (i.e., having a larger effective OT batch size) slightly harms results. In contrast to Pooladian et al. [29], though, we seek to counteract the reduced effective OT batch size induced by conditional optimal transport (as discussed in Section 3.3.4), and we do find reasonably large (640 for CIFAR, 6400 for ImageNet) OT batch sizes useful while not adding wall-clock overhead.", + "bbox": [ + 88, + 364, + 486, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2. Varying Target Ratio $r_{\\mathrm{tar}}$", + "text_level": 1, + "bbox": [ + 89, + 750, + 312, + 766 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here, we analyze the effects of varying the target ratio $r_{\\mathrm{tar}}$ , as defined in Equation (13). We conduct these experiments on the ImageNet- $32 \\times 32$ dataset, training each model three times with different random seeds for every choice of $r_{\\mathrm{tar}}$ . We plot the results in Figure 5. Recall that when $r \\to 0$ , $w \\to \\infty$ and the method approaches FM; when $r \\to 0.5$ ,", + "bbox": [ + 89, + 773, + 485, + 864 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97a9d202d6d745a3acf16bec93aa699b9e76f37042bdf144a38e357b02329111.jpg", + "image_caption": [ + "Figure 4. Changes in FID with respect to varying OT batch sizes $b$ . We plot mean±std over three runs and represent std with a shaded region. OT (adaptive) and FM (Euler-2) perform worse than all other results in this plot, hence results are not shown (full plot in the appendix)." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 903, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/021a2927ace7652996892924796d7b1493956ce09caf1ec68644a6c328feefd0.jpg", + "image_caption": [ + "Figure 5. Changes in FID with respect to varying target ratio $r_{\\mathrm{tar}}$ . We plot mean±std over three runs and represent std with a shaded region. OT (adaptive) performs worse than all other results in this plot, hence results are not shown (full plot in the appendix)." + ], + "image_footnote": [], + "bbox": [ + 516, + 369, + 903, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$w \\to 0$ and the method approaches OT (Section 3.3.3). Our findings are similar to those in Section 4.3.1: as we approach true OT, whether by increasing OT batch size or increasing $r_{\\mathrm{tar}}$ , FID improves in the few-step setting and worsens in the adaptive setting.", + "bbox": [ + 511, + 633, + 908, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 724, + 633, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We first investigate and formalize the struggle of minibatch optimal transport in conditional generation. Based on the findings, we propose $\\mathrm{C}^2\\mathrm{OT}$ , a simple yet effective addition that corrects the degeneration of OT while maintaining straight integration paths. Extensive experiments verify the effectiveness of our method. We believe that these findings, along with the proposed simple technique, close a gap in current research on prior-data coupling.", + "bbox": [ + 511, + 757, + 908, + 877 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "There is a 0.5 chance that the distance between a random prior-data pair is closer to that of another random prior-data pair.", + "bbox": [ + 89, + 875, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "Acknowledgment. This work is supported in part by NSF grants 2008387, 2045586, 2106825, NIFA award 2020-67021-32799, and OAC 2320345.", + "bbox": [ + 511, + 878, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15882", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In Proc. ICLR, 2023. 2", + "[2] Michael Albergo, Nicholas Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. 2", + "[3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi0: A vision-language-action flow model for general robot control. arXiv, 2024. 1", + "[4] Ho Kei Cheng, Masato Ishii, Akio Hayakawa, Takashi Shibuya, Alexander Schwing, and Yuki Mitsufuji. Taming multimodal joint training for high-quality video-to-audio synthesis. CVPR, 2025. 1", + "[5] chenyaofo. Pytorch cifar models. https://github.com/chenyaofo/pytorch-cifar-models, 2023.6", + "[6] Aram Davtyan, Leello Tadesse Dadi, Volkan Cevher, and Paolo Favaro. Faster inference of flow-based generative models via improved data-noise coupling. In ICLR, 2025. 2", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6", + "[8] John R Dormand and Peter J Prince. A family of embedded runge-kutta formulae. Journal of computational and applied mathematics, 1980. 2, 6", + "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1", + "[10] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. ICLR, 2024. 6", + "[11] Zhengyang Geng, Ashwini Pokle, William Luo, Justin Lin, and J Zico Kolter. Consistency models made easy. ICLR, 2025. 2", + "[12] Pengsheng Guo and Alexander G Schwing. Variational rectified flow matching. arXiv, 2025. 2", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6", + "[14] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In ICLR, 2017. 2", + "[15] Ka-Hei Hui, Chao Liu, Xiaohui Zeng, Chi-Wing Fu, and Arash Vahdat. Not-so-optimal transport flows for 3d point cloud generation. arXiv, 2025. 2", + "[16] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. *ICLR*, 2024. 2" + ], + "bbox": [ + 93, + 118, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. In ICLR, 2014. 2", + "[18] Leon Klein, Andreas Kramer, and Frank Noé. Equivariant flow matching. NeurIPS, 36, 2023. 2", + "[19] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images, 2009. 6", + "[20] Harold W. Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 6", + "[21] Visual Layer. Imagenet-1k-vl-enriched. https://huggingface.co/datasets/visual-layer/imagenet-1k-vl-enriched, 2024.6", + "[22] Sangyun Lee, Beomsu Kim, and Jong Chul Ye. Minimizing trajectory curvature of ode-based generative models. In ICML, 2023. 2", + "[23] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. *ICLR*, 2023. 1, 2", + "[24] Qihao Liu, Xi Yin, Alan Yuille, Andrew Brown, and Mannat Singh. Flowing from words to pixels: A framework for cross-modality evolution. arXiv preprint arXiv:2412.15213, 2024. 2", + "[25] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. ICLR, 2023. 2", + "[26] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. *ICLR*, 2024. 2", + "[27] Ségolène Martin, Anne Gagneux, Paul Hagemann, and Gabriele Steidl. Pnp-flow: Plug-and-play image restoration with flow matching. *ICLR*, 2025. 1", + "[28] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models. arXiv, 2024. 1", + "[29] Aram-Alexandre Pooladian, Heli Ben-Hamu, Carles Domingo-Enrich, Brandon Amos, Yaron Lipman, and Ricky TQ Chen. Multisample flow matching: Straightening flows with minibatch couplings. ICML, 2023. 2, 3, 5, 7, 8" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15883", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 7", + "[32] Gianluigi Silvestri, Luca Ambrogioni, Chieh-Hsin Lai, Yuhta Takida, and Yuki Mitsufuji. Training consistency models with variational noise coupling. arXiv preprint arXiv:2502.18197, 2025. 2", + "[33] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2", + "[34] Yuxuan Song, Jingjing Gong, Minkai Xu, Ziyao Cao, Yanyan Lan, Stefano Ermon, Hao Zhou, and Wei-Ying Ma. Equivariant flow matching with hybrid probability transport for 3d molecule generation. NeurIPS, 36, 2023. 2", + "[35] Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. TMLR, 2024. 2, 3, 5, 6", + "[36] Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. arXiv preprint arXiv:2502.14786, 2025. 6", + "[37] Kaiyu Yang, Jacqueline H Yau, Li Fei-Fei, Jia Deng, and Olga Russakovsky. A study of face obfuscation in imagenet. In ICML, 2022. 7", + "[38] Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining straight flows with velocity consistency. arXiv, 2024. 2", + "[39] Jingfeng Yao and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. CVPR, 2025. 7", + "[40] Yichi Zhang, Yici Yan, Alex Schwing, and Zhizhen Zhao. Towards hierarchical rectified flow. *ICLR*, 2025. 2" + ], + "bbox": [ + 91, + 92, + 486, + 684 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15884", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/2ff56f19-5f85-4cca-a5d4-ab26816c514c_model.json b/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/2ff56f19-5f85-4cca-a5d4-ab26816c514c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2a318696b77502d8f7369fc1731d39836c8fc6a3 --- /dev/null +++ b/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/2ff56f19-5f85-4cca-a5d4-ab26816c514c_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.13, + 0.877, + 0.175 + ], + "angle": 0, + "content": "The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.204, + 0.662, + 0.24 + ], + "angle": 0, + "content": "Ho Kei Cheng Alexander Schwing \nUniversity of Illinois Urbana-Champaign" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.244, + 0.605, + 0.256 + ], + "angle": 0, + "content": "{hokeikc2,aschwing}@illinois.edu" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.271, + 0.891, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.513, + 0.908, + 0.598 + ], + "angle": 0, + "content": "Figure 1. We visualize the flows learned by different algorithms using the 8gaussians \\(\\rightarrow\\) moons dataset. We compare flow matching (FM), minibatch optimal transport FM (OT), and our proposed conditional optimal transport FM (\\(C^2 OT\\)) using one Euler step and an adaptive ODE solver, respectively. Below each plot, we show the 2-Wasserstein distance (lower is better; mean \\(\\pm\\) std over 10 runs). For unconditional generation, OT achieves significantly straighter flows, outperforming FM. However, paradoxically, OT performs worse when conditions are introduced. This paper analyzes this degradation and finds that it occurs because optimal transport disregards conditions, leading to a train-test gap. To address this, we propose a simple fix (\\(C^2 OT\\)) that outperforms FM and OT in conditional generation." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.608, + 0.327, + 0.624 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Minibatch optimal transport coupling straightens paths in unconditional flow matching. This leads to computationally less demanding inference as fewer integration steps and less complex numerical solvers can be employed when numerically solving an ordinary differential equation at test time. However, in the conditional setting, minibatch optimal transport falls short. This is because the default optimal transport mapping disregards conditions, resulting in a conditionally skewed prior distribution during training. In contrast, at test time, we have no access to the skewed prior, and instead sample from the full, unbiased prior distribution. This gap between training and testing leads to a subpar performance. To bridge this gap, we propose conditional optimal transport \\((C^2 OT)\\) that adds a conditional weighting term in the cost matrix when computing the optimal transport assignment. Experiments demonstrate that this simple fix works with both discrete and continuous conditions in 8gaussians→moons," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.609, + 0.909, + 0.671 + ], + "angle": 0, + "content": "CIFAR-10, ImageNet-32×32, and ImageNet-256×256. Our method performs better overall compared to the existing baselines across different function evaluation budgets. Code is available at hkchengrex.github.io/C2OT." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.707, + 0.646, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.736, + 0.909, + 0.844 + ], + "angle": 0, + "content": "We focus on flow-matching-based conditional generative models, i.e., generation guided by an input condition. Examples of such conditions include class labels, text, or video. Recently, flow matching has been applied in many areas of computer vision, including text-to-image/video [9, 28], vision-language applications [3], image restoration [27], and video-to-audio [4]. However, these methods can be slow at" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.852, + 0.908, + 0.903 + ], + "angle": 0, + "content": "1We use condition to denote an input condition, not to be confused with the 'condition' in conditional flow matching (CFM) [23] (where the second 'C' in our acronym comes from), which refers to a data sample. To avoid ambiguity, we refer to CFM simply as flow matching (FM) in this paper." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15875" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.483, + 0.288 + ], + "angle": 0, + "content": "test-time – obtaining a solution requires numerically integrating the flow with an ODE solver, typically involving many steps, each requiring a deep network forward pass. Naively reducing the number of steps degrades performance, as the underlying flow is often curved (Figure 1, leftmost column), necessitating a small step size for accurate numerical integration. One way to address this issue is by straightening the flow. In the context of unconditional generation, Tong et al. [35] and Pooladian et al. [29] concurrently proposed \"minibatch optimal transport\" (OT), which deterministically couples data and samples from the prior within a minibatch via optimal transport (replacing random coupling) to minimize flow path lengths and to straighten the flow." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.289, + 0.486, + 0.545 + ], + "angle": 0, + "content": "While OT improves unconditional generation, we find that it paradoxically and consistently harms conditional generation (Figure 1). This occurs because OT disregards the conditions when computing the coupling. As a result, during training, OT samples from a prior distribution that is skewed by the condition, as we show in Section 3.2. However, at test time, we have no access to this skewed distribution and instead sample from the full prior distribution. This mismatch creates a gap between training and testing, leading to performance degradation. To bridge this gap, we propose conditional optimal transport FM \\((\\mathrm{C}^2\\mathrm{OT})\\) which introduces a simple yet effective condition-aware weighting term when computing the cost matrix for OT. Additionally, we propose two techniques: adaptive weight finding to simplify hyperparameter tuning, and efficient oversampling of OT batches to counteract the reduced effective OT batch size introduced by the weighting term." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.486, + 0.712 + ], + "angle": 0, + "content": "We conduct extensive experiments on both two-dimensional synthetic data and high-dimensional image data, including CIFAR-10, ImageNet-32 \\(\\times\\) 32, and ImageNet-256 \\(\\times\\) 256. The results demonstrate that our method performs well across different condition types (discrete and continuous), datasets, network architectures (UNets and transformers), and data spaces (image space and latent space). \\(C^2\\mathrm{OT}\\) achieves better overall performance than the existing baselines across different inference computation budgets, including with few-steps Euler's method and with an adaptive ordinary differential equation (ODE) solver [8]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.735, + 0.242, + 0.751 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Flow Matching. Recently, flow matching [1, 2, 23] has become a popular choice for generative modeling, in part due to its simple training objective and its ability to generate high-quality samples. However, at test time, flow-based methods typically require many computationally expensive forward passes through a deep net to numerically integrate the flow. This is because the underlying flow is usually curved and therefore cannot be approximated well with a few integration steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.349 + ], + "angle": 0, + "content": "Optimal Transport Coupling. To straighten the flow, in the context of unconditional generation, Tong et al. [35] and Pooladian et al. [29] concurrently proposed minibatch optimal transport (OT). While OT has proven effective in the unconditional setting, we show in Section 3.2 that it skews the prior distribution and fails in the conditional setting if used naively. Our method, \\(\\mathrm{C}^2\\mathrm{OT}\\), specifically addresses this failure, aiming to extend the success of OT in unconditional generation to conditional generation. Note, OT has also been used in equivariant flow matching [18, 34], which exploits domain-specific data symmetry (e.g., in molecules). In contrast, our method targets general data. Further, Hui et al. [15] find that mixing of OT and independent coupling benefits shape completion, which is an orthogonal contribution to this paper. Concurrently, Davtyan et al. [6] improve data-noise coupling for unconditional generation, while leaving data coupling for conditional generation to future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.909, + 0.487 + ], + "angle": 0, + "content": "Straightening Flow Paths. Other methods of straightening flow paths exist. Reflow [25] achieves this by retraining a flow matching network multiple times, at the cost of additional training overhead. Variational flow matching [12] reduces flow ambiguity with a latent variable and hierarchical flow matching [40] straightens flows by modeling higher-order flows. These approaches are orthogonal to our focus on prior-data coupling and, in principle, can be combined with our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.611 + ], + "angle": 0, + "content": "Learning the Prior Distribution. An alternative approach to improving flow is to jointly learn a prior distribution with the flow matching network. However, these methods [22, 24, 32] typically rely on a variational autoencoder (VAE) [14, 17] to learn the prior. This introduces yet another density estimation problem, which complicates training. In this work, we focus on improving training-free coupling plans without learning a new prior." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.909, + 0.765 + ], + "angle": 0, + "content": "Consistency Models. Consistency models [33] have been proposed to accelerate diffusion model sampling, and can be extended to flow matching [38]. Recent advancements have improved training efficiency [11], stability [26], and quality [16]. These methods are orthogonal to our contribution, which focuses on improving prior-data coupling, and can be integrated, as demonstrated by Silvestri et al. [32] in their combination of consistency models with OT. In this paper, we focus on the base form of flow matching to avoid distractions from other formulations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.785, + 0.605, + 0.8 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.667, + 0.831 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Flow Matching (FM). We base our discussion on FM [23, 35] (also called rectified flow [25]) for generative modeling and refer readers to [25, 35] for details. At test-time, a sample \\(\\tilde{x}_1\\) is characterized by an ordinary differential equation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15876" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.372, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.089, + 0.639, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.089, + 0.9, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.28, + 0.907, + 0.337 + ], + "angle": 0, + "content": "Figure 2. We visualize the coupling during training and the learned flow from \\( t = 0 \\) to \\( t = 1 \\) during testing for FM, OT, and our proposed method \\( C^2 OT \\). The prior is defined as \\( q(x_0) = \\mathcal{N}(0,1) \\), while the target distribution is given by \\( q(x_{1}) = \\frac{1}{2} q(x_{1}|c = 0) + \\frac{1}{2} q(x_{1}|c = 1)) = \\frac{1}{2}\\mathcal{N}(-2,0.5) + \\frac{1}{2}\\mathcal{N}(2,0.5) \\). Note, FM results in curved flows during testing, while OT degenerates due to skewed training samples (\\( q(x_0|c) \\neq q(x_0) \\)). In contrast, our method successfully learns straight flows without degeneration." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.348, + 0.484, + 0.501 + ], + "angle": 0, + "content": "(ODE) with an initial boundary value specified via noise \\( x_0 \\), randomly drawn from a prior distribution (e.g., the standard normal). To compute the sample \\( \\tilde{x}_1 \\), this ODE is commonly solved by numerically integrating from time \\( t = 0 \\) to time \\( t = 1 \\) a learned flow/velocity field \\( v_{\\theta} \\), where \\( \\theta \\) denotes the set of trainable deep net parameters. For conditional generation, we additionally obtain a condition \\( c \\), e.g., a class label from user input, which adjusts the flow. We use \\( v_{\\theta,c} \\) to denote this conditional flow. At training time, we find the deep net parameters \\( \\theta \\) by minimizing the objective" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.512, + 0.484, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {t, \\left(x _ {0}, x _ {1}, c\\right) \\sim \\pi \\left(x _ {0}, x _ {1}, c\\right)} \\| v _ {\\theta , c} (t, x _ {t}) - u \\left(x _ {t} \\mid x _ {0}, x _ {1}\\right) \\| ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.542, + 0.485, + 0.618 + ], + "angle": 0, + "content": "where \\(t\\) is uniformly drawn from the interval \\([0,1]\\), and where \\(\\pi (x_0,x_1,c)\\) denotes the joint distribution of the prior and the training data. In practice, often, independent coupling is used, i.e., the prior and the training data are sampled independently, such that" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.633, + 0.483, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\pi \\left(x _ {0}, x _ {1}, c\\right) = q \\left(x _ {0}\\right) q \\left(x _ {1}, c\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.663, + 0.148, + 0.677 + ], + "angle": 0, + "content": "Further," + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.681, + 0.483, + 0.696 + ], + "angle": 0, + "content": "\\[\nx _ {t} = t x _ {1} + (1 - t) x _ {0} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.472, + 0.721 + ], + "angle": 0, + "content": "defines a linear interpolation between noise and data, and" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.735, + 0.484, + 0.751 + ], + "angle": 0, + "content": "\\[\nu \\left(x _ {t} \\mid x _ {0}, x _ {1}\\right) = x _ {1} - x _ {0} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.809 + ], + "angle": 0, + "content": "denotes its corresponding flow velocity at \\( x_{t} \\). Note, simply dropping the condition \\( c \\) leads to the unconditional flow matching formulation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Although rectified flow matching is trained with straight paths, the learned flow is often curved (Figure 1, bottom-left) since there are many possible paths induced by the independently sampled couplings \\( x_0, x_1 \\) through any \\( x_t \\) at time \\( t \\) [29]. Curved paths require more numerical integration and network evaluation steps, which is not ideal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.483 + ], + "angle": 0, + "content": "Optimal Transport (OT) Coupling. To address this issue for unconditional generation, Tong et al. [35] and Pooladian et al. [29] concurrently proposed to use minibatch optimal transport to deterministically couple the sampled prior and a sampled batch of the data. This minimizes flow path lengths and straightens the flow. Concretely, given a minibatch of \\( b \\) \\( e \\)-dimensional samples, an OT coupling seeks a \\( b \\times b \\) permutation matrix \\( P \\) such that the squared Euclidean distance is minimized, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.49, + 0.907, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\min _ {P} \\| X _ {0} - P X _ {1} \\| _ {2} ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.612 + ], + "angle": 0, + "content": "Here, \\(X_0, X_1 \\in \\mathbb{R}^{b \\times e}\\) are matrices containing a minibatch of samples from the prior and the data. Instead of an \"independent coupling\", in the unconditional setting we then optimize a variant of Equation (1) using the coupled tuple \\((X_0, PX_1)\\). Note that this algorithm corresponds to sampling tuples \\((x_0, x_1)\\) from a joint distribution \\(\\pi_{\\mathrm{ot}}(x_0, x_1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Importantly, for unconditional generation, i.e., when using the joint distribution \\(\\pi_{\\mathrm{ot}}(x_0,x_1)\\), the marginals remain unchanged from independent coupling (see [35]):" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.665, + 0.836, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\int \\pi_ {\\mathrm {o t}} (x _ {0}, x _ {1}) q (x _ {0}) \\mathrm {d} x _ {0} = q (x _ {1}), \\text {a n d}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.697, + 0.907, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\int \\pi_ {\\mathrm {o t}} \\left(x _ {0}, x _ {1}\\right) q \\left(x _ {1}\\right) \\mathrm {d} x _ {1} = q \\left(x _ {0}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Hence, during test-time, we can sample from the prior \\( q(x_0) \\) and integrate the flow \\( v_{\\theta} \\) to simulate data from the data distribution \\( q(x_{1}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Unfortunately, as shown in the next section, this does not hold for conditional generation. Said differently, naively extending the unconditional joint distribution \\(\\pi_{\\mathrm{ot}}(x_0,x_1)\\) to a conditional joint distribution \\(\\pi_{\\mathrm{ot}}(x_0,x_1,c)\\) by using the coupled tuple \\((X_0,PX_1,PC)\\), where \\(C\\in \\mathbb{R}^{b\\times h}\\) is a batch of \\(h\\)-dimensional conditions, does not maintain marginals and leads to a gap between training time and test time. We discuss this next." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15877" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.383, + 0.108 + ], + "angle": 0, + "content": "3.2. OT Coupling Skews The Prior" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.401 + ], + "angle": 0, + "content": "Intuition. Consider the one-dimensional example illustrated in Figure 2: the prior distribution \\( q(x_0) \\) follows a Gaussian distribution, and the target distribution \\( q(x_1) \\) consists of a mixture of two Gaussians translated left and right, with the translation direction indicated by a binary condition \\( c \\). When training with a \\( \\pi_{\\mathrm{ot}}(x_0,x_1,c) \\) that represents the coupled tuple \\( (X_0, PX_1,PC) \\), samples from the prior and the sampled conditions become correlated through the OT coupling – samples from the left half of the prior distribution are always associated with \\( c = 0 \\), while those from the right half always correspond to \\( c = 1 \\). As a result, the model overfits on data pairs \\( (x_0,c) \\) where either \\( (x_0 < 0,c = 0) \\) or \\( (x_0 > 0,c = 1) \\). However, during testing, this skewed distribution is no longer accessible, and instead, \\( x_0 \\) and \\( c \\) are sampled independently. Consequently, the model encounters previously unseen data pairs, i.e., \\( (x_0 > 0,c = 0) \\) and \\( (x_0 < 0,c = 1) \\) – leading to failure. A similar phenomenon can be observed in the two-dimensional example illustrated in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.484, + 0.555 + ], + "angle": 0, + "content": "Skewed Prior. Formally, imagine that we consider the most general OT coupling joint distribution \\(\\pi_{\\mathrm{ot}}(x_0,x_1,c)\\) at training time. At test time, the user wants to specify a condition \\(c_{1}\\) to obtain samples from the conditional distribution \\(q(x_{1}|c = c_{1})\\). What is the prior distribution that we need to sample from such that we arrive at the correct conditional distribution \\(q(x_{1}|c = c_{1})\\)? To compute this, imagine, we are given an arbitrary condition \\(c_{1}\\), i.e., we look at the induced distribution \\(\\pi_{\\mathrm{ot}}(x_0,x_1,c|c = c_1)\\). Marginalizing this induced distribution over \\(x_{1}\\) yields" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.562, + 0.484, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\int \\pi_ {\\mathrm {o t}} \\left(x _ {0}, x _ {1}, c \\mid c = c _ {1}\\right) \\mathrm {d} x _ {1} = \\int \\pi_ {\\mathrm {o t}} \\left(x _ {0}, x _ {1} \\mid c = c _ {1}\\right) \\mathrm {d} x _ {1} \\\\ = q \\left(x _ {0} \\mid c = c _ {1}\\right), \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.801 + ], + "angle": 0, + "content": "which implies that the prior required to arrive at the conditional distribution \\( q(x_{1}|c = c_{1}) \\) is \\( q(x_{0}|c = c_{1}) \\) - a distribution that we cannot access at test time. In general, the prior that we need to sample from, given condition \\( c \\), is skewed from \\( q(x_{0}) \\) to \\( q(x_{0}|c) \\) which is inaccessible at test time. As a result, accurately capturing \\( q(x_{1}|c) \\) at test time becomes infeasible unless \\( q(x_{0}) \\) and \\( q(c) \\) are independent, i.e., if \\( q(x_{0}|c) = q(x_{0}) \\). However, this condition is generally not satisfied by the most general OT coupling – for instance, in Figure 2, clearly \\( q(x_{0}|c) \\neq q(x_{0}) \\). To address this issue, we propose a method \\( C^2\\mathrm{OT} \\) to unkew the prior. We discuss this in the next section." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.82, + 0.429, + 0.838 + ], + "angle": 0, + "content": "3.3. Conditional Optimal Transport FM" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.225, + 0.861 + ], + "angle": 0, + "content": "3.3.1. Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We propose conditional optimal transport FM (C²OT) to unkew the prior \\( q(x_0|c) \\) such that \\( q(x_0|c) = q(x_0) \\). This" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.212 + ], + "angle": 0, + "content": "ensures that at test time, we can sample from the full prior \\( q(x_0) \\), irrespective of the condition \\( c \\). Importantly, at the same time, we also aim to preserve the straight flow paths provided by OT. Conceptually, we construct a prior distribution independently for each condition, i.e., \\( q(x_0|c_1) = q(x_0|c_2) = q(x_0), \\forall c_1, c_2 \\). This can be achieved by sampling from the prior and computing OT independently for each condition, as shown in Figure 2 (right)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.213, + 0.843, + 0.227 + ], + "angle": 0, + "content": "Formally, we construct the joint distribution as" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.239, + 0.907, + 0.255 + ], + "angle": 0, + "content": "\\[\n\\pi_ {\\mathrm {c} ^ {2} \\mathrm {o t}} (x _ {0}, x _ {1}, c) := q (x _ {0}) q (c) \\pi_ {\\mathrm {o t}} (x _ {1} | x _ {0}, c). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.265, + 0.907, + 0.37 + ], + "angle": 0, + "content": "Here, the prior \\( q(x_0) \\) and the condition \\( q(c) \\) are sampled independently, while the data \\( x_1 \\) is conditioned on \\( x_0 \\) (via OT) and \\( c \\) (via the dataset construction). Different from independent coupling in Equation (2), \\( x_1 \\) and \\( x_0 \\) are associated through OT and therefore lead to straighter flow paths. Also different from the most general OT coupling \\( \\pi_{\\mathrm{ot}}(x_0,x_1,c) \\), we explicitly enforce the independence between \\( x_0 \\) and \\( c \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.907, + 0.416 + ], + "angle": 0, + "content": "To see that this joint distribution provides an unskewed prior, imagine we are given an arbitrary input condition \\( c_{1} \\), marginalizing over \\( x_{1} \\) yields" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.426, + 0.907, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\int q (x _ {0}) q (c | c = c _ {1}) \\pi_ {\\mathrm {o t}} \\left(x _ {1} \\mid x _ {0}, c = c _ {1}\\right) \\mathrm {d} x _ {1} \\\\ = \\int q \\left(x _ {0}\\right) \\pi_ {\\mathrm {o t}} \\left(x _ {1} \\mid x _ {0}, c = c _ {1}\\right) \\mathrm {d} x _ {1} = q \\left(x _ {0}\\right). \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.649 + ], + "angle": 0, + "content": "This implies that, at test time, we can sample from the entire prior \\( q(x_0) \\), regardless of the condition \\( c \\), to arrive at the desired data distribution \\( q(x_1|c) \\). During training, in practice, we sample from \\( \\pi_{\\mathrm{c}^2\\mathrm{ot}} \\) by modifying the optimal transport cost function in Equation (5). This process is exact for discrete conditions (Section 3.3.2) and approximate for continuous conditions (Section 3.3.3). Specifically, given a minibatch of \\( b \\) \\( e \\)-dimensional samples and a minibatch of \\( b \\) \\( h \\)-dimensional conditions \\( C \\in \\mathbb{R}^{b \\times h} \\), \\( \\mathbf{C}^2\\mathrm{OT} \\) seeks a \\( b \\times b \\) permutation matrix \\( P \\) that minimizes the following cost function:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.659, + 0.907, + 0.699 + ], + "angle": 0, + "content": "\\[\n\\min _ {P} \\| X _ {0} - P X _ {1} \\| _ {2} ^ {2} + \\sum_ {i = 1} ^ {b} f \\left(C _ {i}, [ P C ] _ {i}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.906, + 0.799 + ], + "angle": 0, + "content": "where \\( f \\) is a symmetric and non-negative cost function, satisfying the property \\( f(c, c) = 0 \\forall c \\). The key differences between independent coupling, OT coupling, and \\( \\mathrm{C}^2\\mathrm{OT} \\) coupling are highlighted in Algorithms 1 to 3. In the following two sections, we discuss our choice of \\( f \\) for both discrete and continuous conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.697, + 0.831 + ], + "angle": 0, + "content": "3.3.2. Discrete Conditions" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "For discrete conditions, we assume the conditions correspond to class labels from a finite, discrete set, i.e., \\( c \\in \\{0, 1, \\ldots, k - 1\\} \\). For example, in Figure 2, the labels 0 and 1 represent the left and right groups in the target distribution," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15878" + } + ], + [ + { + "type": "algorithm", + "bbox": [ + 0.094, + 0.084, + 0.351, + 0.19 + ], + "angle": 0, + "content": "Algorithm 1 Independent Coupling \n1: function INDEPENDENT(X1, C) \n2: X0 ~ N ▷ Sample from prior \n3: \n4: \n5: return X0, X1, C" + }, + { + "type": "algorithm", + "bbox": [ + 0.37, + 0.084, + 0.626, + 0.19 + ], + "angle": 0, + "content": "Algorithm 2 OT Coupling \n1: function OT(X1, C) \n2: X0 ~ N ▷ Sample from prior \n3: Cost ← pairwise_dist(X0, X1) \n4: i, j ← Hungarian_MATCHing(Cost) \n5: return X0[i], X1[j], C[j]" + }, + { + "type": "algorithm", + "bbox": [ + 0.645, + 0.084, + 0.901, + 0.192 + ], + "angle": 0, + "content": "Algorithm 3 \\(\\mathrm{C}^2\\mathrm{OT}\\) Coupling (Ours) \n1: function \\(\\mathrm{C}^2\\mathrm{OT}(X_1,C)\\) \n2: \\(X_0\\sim \\mathcal{N}\\) ▷ Sample from prior \n3: Cost \\(\\leftarrow\\) pairwise_dist(Xo,X1)+ pairwise_dist(C,C) \n4: \\(i,j\\gets\\) Hungarian_MATCHing(Cost) \n5: return \\(X_0[i],X_1[j],C[j]\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.203, + 0.483, + 0.248 + ], + "angle": 0, + "content": "respectively. We define \\( f = f_{\\mathrm{d}} \\) such that no transport is allowed to modify the conditions (i.e., \\( PC = C \\)). This is achieved via" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.258, + 0.484, + 0.299 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {d}} \\left(c _ {1}, c _ {2}\\right) = \\left\\{ \\begin{array}{l} \\infty , c _ {1} \\neq c _ {2}, \\\\ 0, \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.31, + 0.484, + 0.449 + ], + "angle": 0, + "content": "As a result, the unordered set of prior samples for any given condition \\( c \\) remains unchanged: \\( \\{X_{0i}|c_i = c\\} \\stackrel{\\mathrm{iid}}{\\sim} q(x_0) \\). This formulation samples exactly from our ideal (Equation (8)), as OT is computed independently within each class, and the data sample from each class is exposed to the full prior without skew. This effect is illustrated in Figure 1 (middle) and Figure 2 (right). As expected, the results differ from FM's curved flow paths and OT's inability to properly capture the target distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.526 + ], + "angle": 0, + "content": "It is important to note that enforcing \\(PC = C\\) does not necessarily imply \\(P = I\\) since we can still swap row \\(i\\) and \\(j\\) in \\(P\\) as long as \\(c_{i} = c_{j}\\). Setting \\(P = I\\) would correspond to a degeneration back to independent coupling in Equation (2) and forfeit the benefits of straightened flow provided by OT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.587 + ], + "angle": 0, + "content": "Unfortunately, using \\( f_{\\mathrm{d}} \\) with continuous conditions leads to exactly this degeneration, since, in general, no \\( c_{i} = c_{j} \\) unless \\( i = j \\). Hence, in the next section, we devise a relaxed cost for handling continuous conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.61, + 0.299, + 0.624 + ], + "angle": 0, + "content": "3.3.3. Continuous Conditions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.484, + 0.725 + ], + "angle": 0, + "content": "For continuous conditions, the condition is typically a feature embedding, e.g., computed from a text prompt. Directly applying \\( f_{\\mathrm{d}} \\) in Equation (11) leads to a degeneration back to independent coupling as discussed earlier. To avoid this, we propose a relaxed penalty function \\( f_{\\mathrm{c}} \\) based on the cosine distance:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.736, + 0.484, + 0.769 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {c}} \\left(c _ {1}, c _ {2}\\right) = w \\left(1 - \\frac {c _ {1} \\cdot c _ {2}}{\\| c _ {1} \\| \\| c _ {2} \\|}\\right) = w \\cdot \\operatorname {c d i s t} \\left(c _ {1}, c _ {2}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Here, \\( w > 0 \\) is a scaling hyperparameter. Note, when \\( w = 0 \\), this formulation degenerates to regular OT; when \\( w \\to \\infty \\), this formulation approaches Equation (11), i.e., independent coupling if no two conditions are equal. The right section of Figure 1 illustrates an example where the \\( x \\)-coordinate of the target data point serves as the condition. Notably, \\( C^2\\mathrm{OT} \\) preserves straight flow paths without exhibiting OT's apparent degradation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.203, + 0.907, + 0.295 + ], + "angle": 0, + "content": "Finding \\( w \\). The optimal \\( w \\) is highly dependent on the data distribution and the magnitude of features. To alleviate the need for tuning the hyperparameter \\( w \\), we propose to find \\( w \\) adaptively for each minibatch. For this, we develop a ratio \\( r(w) \\) which measures the proportion of samples that are considered \"potential optimal transport candidates\", i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.306, + 0.907, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} r (w) = \\frac {1}{B ^ {2}} \\sum_ {i, j} \\mathbb {1} \\left(\\left\\| X _ {0 i} - X _ {1 j} \\right\\| _ {2} ^ {2} + w \\cdot \\operatorname {c d i s t} \\left(c _ {i}, c _ {j}\\right) \\right. \\\\ \\leq \\left\\| X _ {0 i} - X _ {1 i} \\right\\| _ {2} ^ {2}), \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.909, + 0.631 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) is an indicator function that evaluates to one if the condition is satisfied and to zero otherwise. Then, we introduce a hyperparameter \\(r_{\\mathrm{tar}}\\) as the target ratio and find \\(w\\) such that \\(r(w) \\approx r_{\\mathrm{tar}}\\). Note that setting \\(r_{\\mathrm{tar}}\\) is invariant to the scaling of distances between \\(x_0\\) and \\(x_1\\). Namely, if \\(\\|X_{0i} - X_{1j}\\|_2^2\\) is scaled by \\(s\\), \\(w\\) would also be scaled by \\(s\\) to preserve the same \\(r_{\\mathrm{tar}}\\). This invariance is particularly useful when transitioning to a new dataset, such as changing image resolutions in image generation tasks. In our implementation, we search for \\(w\\) in two steps: first with exponential search to establish the upper/ lower bounds, then with binary search to locate a precise value. The final \\(w\\) is used as the initial value in the subsequent minibatch. In practice, we observe \\(w\\) differs very little between minibatches (\\(< 1\\%\\)) and the search converges quickly within 10 iterations with a negligible overhead. Next, we introduce oversampling, a simple technique to obtain more accurate OT couplings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.744, + 0.664 + ], + "angle": 0, + "content": "3.3.4. Oversampling OT Batches" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.672, + 0.907, + 0.868 + ], + "angle": 0, + "content": "OT vs. Deep Net Batch Sizes. Typically, the \"OT batch size\" \\( b \\) used to compute optimal transport is set equal to the \"deep net batch size\" \\( b_{n} \\) used to compute forward/backward passes of the deep net [29, 35]. While the OT batch size controls the dynamics of prior-to-data assignments, the deep net batch size affects gradient variances, training efficiency, and memory usage. There is no reason why setting them equal should be optimal. Particularly, with our proposed \\( C^2\\mathrm{OT} \\), the effective OT batch size is reduced since we restrict transport between samples with divergent conditions, which calls for an increased OT batch size \\( b \\). At the same time, keeping the deep net batch size \\( b_{n} \\) unchanged prevents unwarranted side effects in other aspects of training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Reduced Effective OT Batch Size. In the case of discrete conditions, OT is effectively performed within" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15879" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Table 1. Results on the 8gaussians \\(\\rightarrow\\) moons dataset with different training coupling methods and different conditioning. Note that our method \\(C^2\\mathrm{OT}\\) does not apply in the unconditional setting." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.135, + 0.479, + 0.344 + ], + "angle": 0, + "content": "
MethodEuler-1 (W2↓)Adaptive (W2↓)NFE ↓
Unconditional
FM6.232±0.1860.120±0.04593.22±1.81
OT0.072±0.0250.050±0.04036.08±2.57
Binary conditions
FM2.573±0.1120.059±0.02991.10±2.77
OT0.483±0.0590.462±0.02532.89±0.95
C2OT (ours)0.048±0.0100.018±0.00658.88±1.48
Continuous conditions
FM0.732±0.0520.028±0.01093.86±1.76
OT8.276±6.5102.143±1.99342.07±2.59
C2OT (ours)0.077±0.0240.013±0.00389.87±1.53
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.485, + 0.432 + ], + "angle": 0, + "content": "\\(\\{X_{0i}, X_{1i} | c_i = c\\}\\) for each condition \\(c\\). With \\(k\\) classes, the expected effective OT batch size is reduced to \\(\\mathbb{E}(|\\{i | c_i = c\\}|) = b / k\\). Empirically, we find that using a much larger OT batch size \\(b\\) than the network batch size \\(b_n\\) is helpful. This observation motivates oversampling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.48 + ], + "angle": 0, + "content": "Efficiency. A natural concern with increasing \\( b \\) is the computational overhead, since optimal transport with Hungarian matching [20] has a time complexity of \\( O(b^{3}) \\). However," + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.48, + 0.484, + 0.525 + ], + "angle": 0, + "content": "- Since each OT batch can be used across \\((b / b_{n})\\) forward/backward passes, the amortized cost per minibatch reduces to \\(O(b^{2}b_{n})\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.526, + 0.483, + 0.571 + ], + "angle": 0, + "content": "- We offload OT computation from the main process to the data workers, enabling data preparation on the CPU in parallel with forward/backward passes on the GPU." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.48, + 0.484, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.571, + 0.484, + 0.632 + ], + "angle": 0, + "content": "With a modest choice of 8 data workers per GPU, we observe no wall-clock overhead for OT batch sizes up to 6,400 when training on CIFAR-10 and ImageNet, as data preparation is completed faster than the GPU can process batches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.655, + 0.224, + 0.673 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.486, + 0.794 + ], + "angle": 0, + "content": "We experimentally verify our proposed method \\(\\mathrm{C}^2\\mathrm{OT}\\) in two sections: first on the two-dimensional synthetic dataset 8gaussians \\(\\rightarrow\\) moons, and then on high-dimensional image data, including CIFAR-10 [19] and ImageNet-1K [7]. For ImageNet-1K, we conduct experiments on both \\(32\\times 32\\) images in image space and \\(256\\times 256\\) images in latent space. Implementation details are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.322, + 0.831 + ], + "angle": 0, + "content": "4.1. Two-Dimensional Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.486, + 0.903 + ], + "angle": 0, + "content": "The two-dimensional 8gaussians \\(\\rightarrow\\) moons dataset maps a mixture of eight Gaussian distributions to two interleaving half-circles, as visualized in Figure 1. We consider three conditioning scenarios: (a) unconditional; (b) binary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.182 + ], + "angle": 0, + "content": "class conditions, where each class corresponds to one of the half-circles; and (c) continuous conditions, where the \\(x\\)-coordinate of the desired target point is given as the condition. For (b), we apply our setup for discrete conditions (Section 3.3.2); for (c), we apply our setup for continuous conditions (Section 3.3.3) with \\(r_{\\mathrm{tar}} = 0.01\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.185, + 0.909, + 0.336 + ], + "angle": 0, + "content": "Metrics. We report the 2-Wasserstein distance \\((W_2^2)\\) (following [35]) between 10K generated data points and 10K samples from the ground-truth moons distribution. We experiment with single-step Euler's method (Euler-1) and the adaptive Dormand-Prince method [8] (dopri5, referred to as adaptive) for numerical integration. Additionally, we report the number of function evaluations (NFE) in the adaptive method, which is the number of forward passes through the neural network. We average the results over ten training runs with different random seeds and report mean±std." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.909, + 0.475 + ], + "angle": 0, + "content": "Results. The results are summarized in Table 1 and visualized in Figure 1. In the unconditional setting, OT produces straighter flows, leading to better distribution matching with fewer NFEs, and our method does not apply. However, in the conditional setting, OT degrades and performs worse than regular FM in the adaptive setting due to the aforementioned train-test gap in OT coupling. Our method, \\( \\mathrm{C}^2\\mathrm{OT} \\), effectively models the target distribution with both Euler-1 and the adaptive method, improving upon the baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.494, + 0.809, + 0.511 + ], + "angle": 0, + "content": "4.2. High-Dimensional Image Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.908, + 0.624 + ], + "angle": 0, + "content": "We conduct experiments on CIFAR-10 [19] (image space, class-conditioned), ImageNet- \\(32 \\times 32\\) [7] (image space, caption-conditioned), and ImageNet- \\(256 \\times 256\\) [7] (latent space, caption-conditioned). For ImageNet, we use the captions provided by [21] and encode text features using a CLIP [30]-like DFN [10] text encoder as input conditions. Our results are summarized in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Metrics. We report Fréchet inception distance (FID) [13] and the number of function evaluations (NFE) in the adaptive solver for all three datasets. To assess how well the generations adhere to the input conditions, we additionally compute condition adherence metrics. For CIFAR-10, we run a pretrained classifier [5] on the generated images to obtain logits and report the average cross entropy (CE) with the input class label. For ImageNet, we compute CLIP features [30] using SigLIP-2 [36] from the generated images and the input captions respectively, and report the average cosine similarities (CLIP)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "CIFAR-10. We base our UNet architecture on Tong et al. [35], with details provided in the appendix. The CIFAR-10 training set contains 50,000 \\(32 \\times 32\\) color images across 10 object classes. During testing, we generate 50,000 images evenly distributed among classes and assess FID against the training set following [35]. Each model is trained five times with different random seeds and the results are reported" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15880" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.146 + ], + "angle": 0, + "content": "Table 2. Results of different training algorithms on image generation tasks. We bold the best-performing entry and underline the second-best. Our proposed method, \\( C^2 OT \\), achieves the best overall performance – better than FM with few sampling steps and better than OT with more sampling steps. In cases where \\( C^2 OT \\) performs second, our performance is often comparable to the best entry. Additionally, \\( C^2 OT \\) has the most stable performance, with the smallest deviations across runs in most entries. Variations in CLIP scores are minimal." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.15, + 0.907, + 0.515 + ], + "angle": 0, + "content": "
CIFAR-10 Class-Conditioned Generation
MethodEuler-2Euler-5Euler-10Euler-25Euler-50Euler-100AdaptiveNFE ↓
FID → FM105.481±1.61921.968±0.40010.479±0.2375.640±0.1174.128±0.0693.429±0.0542.922±0.039124.0±2.5
OT64.417±0.43418.194±0.40910.778±0.2706.821±0.1865.368±0.1134.671±0.0664.144±0.018125.7±0.4
C²OT (ours)64.694±0.37917.565±0.2029.762±0.1365.644±0.0814.152±0.0613.447±0.0492.904±0.035127.7±0.7
CE → FM3.343±0.0220.546±0.0200.323±0.0070.272±0.0040.267±0.0020.269±0.0030.276±0.004124.0±2.5
OT2.529±0.0100.624±0.0120.426±0.0050.369±0.0040.363±0.0040.364±0.0050.372±0.005125.7±0.4
C²OT (ours)2.264±0.0220.475±0.0090.322±0.0040.276±0.0050.271±0.0030.272±0.0030.279±0.004127.7±0.7
ImageNet 32×32 Caption-Conditioned Generation
FID → FM113.250±0.79323.015±0.12311.141±0.0517.165±0.0876.142±0.0855.673±0.0745.358±0.059146.9±3.5
OT81.317±0.36920.763±0.15411.360±0.0897.854±0.0776.899±0.0796.469±0.0716.220±0.065137.7±0.2
C²OT (ours)102.380±0.27921.965±0.03510.897±0.0337.069±0.0276.084±0.0165.638±0.0185.350±0.017134.2±1.4
CLIP → FM0.067±0.0000.076±0.0000.074±0.0000.071±0.0000.071±0.0000.070±0.0000.069±0.000146.9±3.5
OT0.067±0.0000.071±0.0000.070±0.0000.068±0.0000.068±0.0000.067±0.0000.067±0.000137.7±0.2
C²OT (ours)0.068±0.0000.075±0.0000.073±0.0000.071±0.0000.070±0.0000.070±0.0000.069±0.000134.2±1.4
ImageNet 256×256 Caption-Conditioned Generation in Latent Space
FID → FM202.519±0.60631.266±0.3559.996±0.1065.189±0.0823.886±0.0663.484±0.0553.498±0.054131.1±2.3
OT193.376±1.66542.780±2.13215.638±1.5107.514±1.0345.693±0.9985.360±1.0165.726±1.074125.6±6.5
C²OT (ours)200.811±0.38530.715±0.22110.000±0.0725.069±0.0373.722±0.0373.320±0.0323.369±0.034128.2±2.6
CLIP → FM0.028±0.0000.120±0.0000.134±0.0000.137±0.0000.138±0.0000.139±0.0000.140±0.000131.2±2.3
OT0.031±0.0000.107±0.0030.120±0.0030.123±0.0030.124±0.0030.124±0.0040.125±0.004125.6±6.5
C²OT (ours)0.029±0.0000.120±0.0000.133±0.0000.137±0.0000.138±0.0000.138±0.0000.139±0.000128.2±2.6
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.527, + 0.482, + 0.646 + ], + "angle": 0, + "content": "as mean±std. We use an OT batch size \\( b \\) of 640 and a ratio \\( r_{\\mathrm{tar}} \\) of 0.01. Compared to FM, our method achieves better performance (FID and CE) in few-steps settings and comparable performance in many-steps settings. Compared to OT, our method significantly outperforms in the many-step setting. Although OT achieves slightly better FID in Euler-2, it performs worse in CE, indicating that it fails to follow the condition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.486, + 0.901 + ], + "angle": 0, + "content": "ImageNet \\(32 \\times 32\\). We base our UNet architecture on Pooladian et al. [29], and list details in the appendix. We train and evaluate on the face-blurred [37] ImageNet, following [29]. In contrast to prior works that assess FID against the training set, we assess FID against the validation set (49,997 images), as the fine-grained nature of captions may lead to overfitting on the training set. Each model is trained three times with different random seeds and the results are reported as mean \\(\\pm\\) std. Compared to CIFAR-10, there are much more variations in this dataset (e.g., 1000 classes vs. 10 classes in CIFAR-10). Thus, we adopt a larger OT batch size \\(b\\) of 6400 while keeping the same target ratio \\(r_{\\mathrm{tar}}\\) 0.01 as before. Similar to our findings in CIFAR-10, our method performs better or is comparable to FM; in few-steps settings, OT achieves better FID but worse condition following (CLIP). Our method strikes the best overall balance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.909, + 0.785 + ], + "angle": 0, + "content": "ImageNet \\(256 \\times 256\\) in Latent Space. Here, we explore latent flow matching models [31]. During training, the flow matching model learns to generate data in the latent space, which is encoded from images. At test-time, the generated latents are decoded into images using a pretrained decoder. We base our experiment on a recent transformer-based model, LightningDiT [39], demonstrating that our method is effective across different network architectures and data spaces. We use the officially provided \"64 epochs\" configuration for all experiments (due to computational constraints), which is not directly comparable to methods trained with 800 epochs. We follow the same evaluation protocol as ImageNet- \\(32 \\times 32\\), and adopt the same OT batch size \\(b\\) of 6400 and the target ratio \\(r_{\\mathrm{tar}}\\) of 0.01. Each model is trained five times with different random seeds, and the results are reported as mean \\(\\pm\\) std. Again, \\(\\mathrm{C}^2\\mathrm{OT}\\) has the best overall performance compared to both FM and OT. We visualize the generations in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Limitations in High-Dimensional Data We observed that the benefits of \\(\\mathrm{C}^2\\mathrm{OT}\\) are less pronounced, though still visible, in higher-dimensional image data compared to low-dimensional data. We hypothesize this is because image data is more entangled, making the prior skew induced by OT less detrimental. We also observe statistically straighter paths in \\(\\mathrm{C}^2\\mathrm{OT}\\) compared to FM, with details given in the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "15881" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.485, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.246, + 0.486, + 0.304 + ], + "angle": 0, + "content": "Figure 3. Visual comparisons of \\(256 \\times 256\\) images generated by the baselines and our approach with different amounts of sampling steps. Our approach converges faster and produces cleaner outputs. The input caption is \"a small dog sitting on a carpet\"." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.313, + 0.271, + 0.33 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.341, + 0.31, + 0.358 + ], + "angle": 0, + "content": "4.3.1. Varying OT Batch Size \\( b \\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.366, + 0.487, + 0.73 + ], + "angle": 0, + "content": "Here, we analyze the effect of varying the OT batch size \\( b \\). We conduct these experiments on the CIFAR-10 dataset, training each model three times with different random seeds for every choice of \\( b \\). We plot the results in Figure 4. Increasing the OT batch size clearly improves FID in the few-sampling-steps setting (Euler-2). This is intuitive, because a larger OT batch size enables the coupling to approach the true OT results, i.e., dataset-level optimal transport. In turn, this leads to straighter paths. However, in the adaptive setting, increasing the OT batch size slightly worsens FID, though the effect is much less significant (the range of the blue y-axis is much smaller than the range of the red y-axis). We think this occurs because a more accurate OT coupling reduces variations in the training data (similar to having less data augmentations) which harms performance as the benefit of straighter flows diminishes in the adaptive setting. This observation aligns with the findings of Pooladian et al. [29]: computing OT coupling across GPUs (i.e., having a larger effective OT batch size) slightly harms results. In contrast to Pooladian et al. [29], though, we seek to counteract the reduced effective OT batch size induced by conditional optimal transport (as discussed in Section 3.3.4), and we do find reasonably large (640 for CIFAR, 6400 for ImageNet) OT batch sizes useful while not adding wall-clock overhead." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.75, + 0.313, + 0.767 + ], + "angle": 0, + "content": "4.3.2. Varying Target Ratio \\( r_{\\mathrm{tar}} \\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.486, + 0.866 + ], + "angle": 0, + "content": "Here, we analyze the effects of varying the target ratio \\( r_{\\mathrm{tar}} \\), as defined in Equation (13). We conduct these experiments on the ImageNet- \\( 32 \\times 32 \\) dataset, training each model three times with different random seeds for every choice of \\( r_{\\mathrm{tar}} \\). We plot the results in Figure 5. Recall that when \\( r \\to 0 \\), \\( w \\to \\infty \\) and the method approaches FM; when \\( r \\to 0.5 \\)," + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.485, + 0.903 + ], + "angle": 0, + "content": "There is a 0.5 chance that the distance between a random prior-data pair is closer to that of another random prior-data pair." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.284, + 0.908, + 0.355 + ], + "angle": 0, + "content": "Figure 4. Changes in FID with respect to varying OT batch sizes \\( b \\). We plot mean±std over three runs and represent std with a shaded region. OT (adaptive) and FM (Euler-2) perform worse than all other results in this plot, hence results are not shown (full plot in the appendix)." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.563, + 0.909, + 0.621 + ], + "angle": 0, + "content": "Figure 5. Changes in FID with respect to varying target ratio \\( r_{\\mathrm{tar}} \\). We plot mean±std over three runs and represent std with a shaded region. OT (adaptive) performs worse than all other results in this plot, hence results are not shown (full plot in the appendix)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.909, + 0.71 + ], + "angle": 0, + "content": "\\(w \\to 0\\) and the method approaches OT (Section 3.3.3). Our findings are similar to those in Section 4.3.1: as we approach true OT, whether by increasing OT batch size or increasing \\(r_{\\mathrm{tar}}\\), FID improves in the few-step setting and worsens in the adaptive setting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.634, + 0.743 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.909, + 0.878 + ], + "angle": 0, + "content": "We first investigate and formalize the struggle of minibatch optimal transport in conditional generation. Based on the findings, we propose \\(\\mathrm{C}^2\\mathrm{OT}\\), a simple yet effective addition that corrects the degeneration of OT while maintaining straight integration paths. Extensive experiments verify the effectiveness of our method. We believe that these findings, along with the proposed simple technique, close a gap in current research on prior-data coupling." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.879, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This work is supported in part by NSF grants 2008387, 2045586, 2106825, NIFA award 2020-67021-32799, and OAC 2320345." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15882" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.119, + 0.486, + 0.158 + ], + "angle": 0, + "content": "[1] Michael Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In Proc. ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.484, + 0.203 + ], + "angle": 0, + "content": "[2] Michael Albergo, Nicholas Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.205, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi0: A vision-language-action flow model for general robot control. arXiv, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.276, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[4] Ho Kei Cheng, Masato Ishii, Akio Hayakawa, Takashi Shibuya, Alexander Schwing, and Yuki Mitsufuji. Taming multimodal joint training for high-quality video-to-audio synthesis. CVPR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.333, + 0.485, + 0.36 + ], + "angle": 0, + "content": "[5] chenyaofo. Pytorch cifar models. https://github.com/chenyaofo/pytorch-cifar-models, 2023.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.484, + 0.402 + ], + "angle": 0, + "content": "[6] Aram Davtyan, Leello Tadesse Dadi, Volkan Cevher, and Paolo Favaro. Faster inference of flow-based generative models via improved data-noise coupling. In ICLR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.484, + 0.444 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[8] John R Dormand and Peter J Prince. A family of embedded runge-kutta formulae. Journal of computational and applied mathematics, 1980. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.484, + 0.557 + ], + "angle": 0, + "content": "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.561, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[10] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. ICLR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.604, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[11] Zhengyang Geng, Ashwini Pokle, William Luo, Justin Lin, and J Zico Kolter. Consistency models made easy. ICLR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[12] Pengsheng Guo and Alexander G Schwing. Variational rectified flow matching. arXiv, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.676, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[14] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In ICLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[15] Ka-Hei Hui, Chao Liu, Xiaohui Zeng, Chi-Wing Fu, and Arash Vahdat. Not-so-optimal transport flows for 3d point cloud generation. arXiv, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[16] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. *ICLR*, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.119, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[17] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. In ICLR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[18] Leon Klein, Andreas Kramer, and Frank Noé. Equivariant flow matching. NeurIPS, 36, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[19] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[20] Harold W. Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[21] Visual Layer. Imagenet-1k-vl-enriched. https://huggingface.co/datasets/visual-layer/imagenet-1k-vl-enriched, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.248, + 0.907, + 0.286 + ], + "angle": 0, + "content": "[22] Sangyun Lee, Beomsu Kim, and Jong Chul Ye. Minimizing trajectory curvature of ode-based generative models. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.907, + 0.329 + ], + "angle": 0, + "content": "[23] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. *ICLR*, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.907, + 0.385 + ], + "angle": 0, + "content": "[24] Qihao Liu, Xi Yin, Alan Yuille, Andrew Brown, and Mannat Singh. Flowing from words to pixels: A framework for cross-modality evolution. arXiv preprint arXiv:2412.15213, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.388, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[25] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.456 + ], + "angle": 0, + "content": "[26] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. *ICLR*, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.458, + 0.905, + 0.498 + ], + "angle": 0, + "content": "[27] Ségolène Martin, Anne Gagneux, Paul Hagemann, and Gabriele Steidl. Pnp-flow: Plug-and-play image restoration with flow matching. *ICLR*, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.5, + 0.908, + 0.83 + ], + "angle": 0, + "content": "[28] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models. arXiv, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.899 + ], + "angle": 0, + "content": "[29] Aram-Alexandre Pooladian, Heli Ben-Hamu, Carles Domingo-Enrich, Brandon Amos, Yaron Lipman, and Ricky TQ Chen. Multisample flow matching: Straightening flows with minibatch couplings. ICML, 2023. 2, 3, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15883" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.161 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.486, + 0.204 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[32] Gianluigi Silvestri, Luca Ambrogioni, Chieh-Hsin Lai, Yuhta Takida, and Yuki Mitsufuji. Training consistency models with variational noise coupling. arXiv preprint arXiv:2502.18197, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.263, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[33] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.292, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[34] Yuxuan Song, Jingjing Gong, Minkai Xu, Ziyao Cao, Yanyan Lan, Stefano Ermon, Hao Zhou, and Wei-Ying Ma. Equivariant flow matching with hybrid probability transport for 3d molecule generation. NeurIPS, 36, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.416 + ], + "angle": 0, + "content": "[35] Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. TMLR, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.486, + 0.514 + ], + "angle": 0, + "content": "[36] Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. arXiv preprint arXiv:2502.14786, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.484, + 0.557 + ], + "angle": 0, + "content": "[37] Kaiyu Yang, Jacqueline H Yau, Li Fei-Fei, Jia Deng, and Olga Russakovsky. A study of face obfuscation in imagenet. In ICML, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.484, + 0.615 + ], + "angle": 0, + "content": "[38] Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining straight flows with velocity consistency. arXiv, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.484, + 0.656 + ], + "angle": 0, + "content": "[39] Jingfeng Yao and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. CVPR, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[40] Yichi Zhang, Yici Yan, Alex Schwing, and Zhizhen Zhao. Towards hierarchical rectified flow. *ICLR*, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15884" + } + ] +] \ No newline at end of file diff --git a/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/2ff56f19-5f85-4cca-a5d4-ab26816c514c_origin.pdf b/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/2ff56f19-5f85-4cca-a5d4-ab26816c514c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..00cd8f9d53d1941f6a0007710030016b1e2e6a0d --- /dev/null +++ b/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/2ff56f19-5f85-4cca-a5d4-ab26816c514c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6b7d5385307201babc1ba51116ba6b391ad08aa25e481be813400f852b631b5 +size 8008856 diff --git a/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/full.md b/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f2e0cf0aaf0ee53176937d3ae3aaf865aa8fca29 --- /dev/null +++ b/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/full.md @@ -0,0 +1,320 @@ +# The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation + +Ho Kei Cheng Alexander Schwing +University of Illinois Urbana-Champaign + +{hokeikc2,aschwing}@illinois.edu + +![](images/38678b1cb76cbb3b6d3e7c503a10fc6a5665cb3960e9edee95567a426ef7c902.jpg) +Figure 1. We visualize the flows learned by different algorithms using the 8gaussians $\rightarrow$ moons dataset. We compare flow matching (FM), minibatch optimal transport FM (OT), and our proposed conditional optimal transport FM ( $C^2 OT$ ) using one Euler step and an adaptive ODE solver, respectively. Below each plot, we show the 2-Wasserstein distance (lower is better; mean $\pm$ std over 10 runs). For unconditional generation, OT achieves significantly straighter flows, outperforming FM. However, paradoxically, OT performs worse when conditions are introduced. This paper analyzes this degradation and finds that it occurs because optimal transport disregards conditions, leading to a train-test gap. To address this, we propose a simple fix ( $C^2 OT$ ) that outperforms FM and OT in conditional generation. + +# Abstract + +Minibatch optimal transport coupling straightens paths in unconditional flow matching. This leads to computationally less demanding inference as fewer integration steps and less complex numerical solvers can be employed when numerically solving an ordinary differential equation at test time. However, in the conditional setting, minibatch optimal transport falls short. This is because the default optimal transport mapping disregards conditions, resulting in a conditionally skewed prior distribution during training. In contrast, at test time, we have no access to the skewed prior, and instead sample from the full, unbiased prior distribution. This gap between training and testing leads to a subpar performance. To bridge this gap, we propose conditional optimal transport $(C^2 OT)$ that adds a conditional weighting term in the cost matrix when computing the optimal transport assignment. Experiments demonstrate that this simple fix works with both discrete and continuous conditions in 8gaussians→moons, + +CIFAR-10, ImageNet-32×32, and ImageNet-256×256. Our method performs better overall compared to the existing baselines across different function evaluation budgets. Code is available at hkchengrex.github.io/C2OT. + +# 1. Introduction + +We focus on flow-matching-based conditional generative models, i.e., generation guided by an input condition. Examples of such conditions include class labels, text, or video. Recently, flow matching has been applied in many areas of computer vision, including text-to-image/video [9, 28], vision-language applications [3], image restoration [27], and video-to-audio [4]. However, these methods can be slow at + +test-time – obtaining a solution requires numerically integrating the flow with an ODE solver, typically involving many steps, each requiring a deep network forward pass. Naively reducing the number of steps degrades performance, as the underlying flow is often curved (Figure 1, leftmost column), necessitating a small step size for accurate numerical integration. One way to address this issue is by straightening the flow. In the context of unconditional generation, Tong et al. [35] and Pooladian et al. [29] concurrently proposed "minibatch optimal transport" (OT), which deterministically couples data and samples from the prior within a minibatch via optimal transport (replacing random coupling) to minimize flow path lengths and to straighten the flow. + +While OT improves unconditional generation, we find that it paradoxically and consistently harms conditional generation (Figure 1). This occurs because OT disregards the conditions when computing the coupling. As a result, during training, OT samples from a prior distribution that is skewed by the condition, as we show in Section 3.2. However, at test time, we have no access to this skewed distribution and instead sample from the full prior distribution. This mismatch creates a gap between training and testing, leading to performance degradation. To bridge this gap, we propose conditional optimal transport FM $(\mathrm{C}^2\mathrm{OT})$ which introduces a simple yet effective condition-aware weighting term when computing the cost matrix for OT. Additionally, we propose two techniques: adaptive weight finding to simplify hyperparameter tuning, and efficient oversampling of OT batches to counteract the reduced effective OT batch size introduced by the weighting term. + +We conduct extensive experiments on both two-dimensional synthetic data and high-dimensional image data, including CIFAR-10, ImageNet-32 $\times$ 32, and ImageNet-256 $\times$ 256. The results demonstrate that our method performs well across different condition types (discrete and continuous), datasets, network architectures (UNets and transformers), and data spaces (image space and latent space). $C^2\mathrm{OT}$ achieves better overall performance than the existing baselines across different inference computation budgets, including with few-steps Euler's method and with an adaptive ordinary differential equation (ODE) solver [8]. + +# 2. Related Works + +Flow Matching. Recently, flow matching [1, 2, 23] has become a popular choice for generative modeling, in part due to its simple training objective and its ability to generate high-quality samples. However, at test time, flow-based methods typically require many computationally expensive forward passes through a deep net to numerically integrate the flow. This is because the underlying flow is usually curved and therefore cannot be approximated well with a few integration steps. + +Optimal Transport Coupling. To straighten the flow, in the context of unconditional generation, Tong et al. [35] and Pooladian et al. [29] concurrently proposed minibatch optimal transport (OT). While OT has proven effective in the unconditional setting, we show in Section 3.2 that it skews the prior distribution and fails in the conditional setting if used naively. Our method, $\mathrm{C}^2\mathrm{OT}$ , specifically addresses this failure, aiming to extend the success of OT in unconditional generation to conditional generation. Note, OT has also been used in equivariant flow matching [18, 34], which exploits domain-specific data symmetry (e.g., in molecules). In contrast, our method targets general data. Further, Hui et al. [15] find that mixing of OT and independent coupling benefits shape completion, which is an orthogonal contribution to this paper. Concurrently, Davtyan et al. [6] improve data-noise coupling for unconditional generation, while leaving data coupling for conditional generation to future work. + +Straightening Flow Paths. Other methods of straightening flow paths exist. Reflow [25] achieves this by retraining a flow matching network multiple times, at the cost of additional training overhead. Variational flow matching [12] reduces flow ambiguity with a latent variable and hierarchical flow matching [40] straightens flows by modeling higher-order flows. These approaches are orthogonal to our focus on prior-data coupling and, in principle, can be combined with our method. + +Learning the Prior Distribution. An alternative approach to improving flow is to jointly learn a prior distribution with the flow matching network. However, these methods [22, 24, 32] typically rely on a variational autoencoder (VAE) [14, 17] to learn the prior. This introduces yet another density estimation problem, which complicates training. In this work, we focus on improving training-free coupling plans without learning a new prior. + +Consistency Models. Consistency models [33] have been proposed to accelerate diffusion model sampling, and can be extended to flow matching [38]. Recent advancements have improved training efficiency [11], stability [26], and quality [16]. These methods are orthogonal to our contribution, which focuses on improving prior-data coupling, and can be integrated, as demonstrated by Silvestri et al. [32] in their combination of consistency models with OT. In this paper, we focus on the base form of flow matching to avoid distractions from other formulations. + +# 3. Method + +# 3.1. Preliminaries + +Flow Matching (FM). We base our discussion on FM [23, 35] (also called rectified flow [25]) for generative modeling and refer readers to [25, 35] for details. At test-time, a sample $\tilde{x}_1$ is characterized by an ordinary differential equation + +![](images/95849445438a1dfb8948167facc0d9d74ca8c25a9fa9634fd62430bcfaca1c11.jpg) +Figure 2. We visualize the coupling during training and the learned flow from $t = 0$ to $t = 1$ during testing for FM, OT, and our proposed method $C^2 OT$ . The prior is defined as $q(x_0) = \mathcal{N}(0,1)$ , while the target distribution is given by $q(x_{1}) = \frac{1}{2} q(x_{1}|c = 0) + \frac{1}{2} q(x_{1}|c = 1)) = \frac{1}{2}\mathcal{N}(-2,0.5) + \frac{1}{2}\mathcal{N}(2,0.5)$ . Note, FM results in curved flows during testing, while OT degenerates due to skewed training samples ( $q(x_0|c) \neq q(x_0)$ ). In contrast, our method successfully learns straight flows without degeneration. + +![](images/c0f240adc569ff6e5bdcd6d1c715f4ddad481eab94b6fb224163c0c69f1ff9fc.jpg) + +![](images/88eea11d8895ec1f0a725be32cf2417094fde3431a1482f27456da5547c6ece1.jpg) + +(ODE) with an initial boundary value specified via noise $x_0$ , randomly drawn from a prior distribution (e.g., the standard normal). To compute the sample $\tilde{x}_1$ , this ODE is commonly solved by numerically integrating from time $t = 0$ to time $t = 1$ a learned flow/velocity field $v_{\theta}$ , where $\theta$ denotes the set of trainable deep net parameters. For conditional generation, we additionally obtain a condition $c$ , e.g., a class label from user input, which adjusts the flow. We use $v_{\theta,c}$ to denote this conditional flow. At training time, we find the deep net parameters $\theta$ by minimizing the objective + +$$ +\mathbb {E} _ {t, \left(x _ {0}, x _ {1}, c\right) \sim \pi \left(x _ {0}, x _ {1}, c\right)} \| v _ {\theta , c} (t, x _ {t}) - u \left(x _ {t} \mid x _ {0}, x _ {1}\right) \| ^ {2}, \tag {1} +$$ + +where $t$ is uniformly drawn from the interval $[0,1]$ , and where $\pi (x_0,x_1,c)$ denotes the joint distribution of the prior and the training data. In practice, often, independent coupling is used, i.e., the prior and the training data are sampled independently, such that + +$$ +\pi \left(x _ {0}, x _ {1}, c\right) = q \left(x _ {0}\right) q \left(x _ {1}, c\right). \tag {2} +$$ + +Further, + +$$ +x _ {t} = t x _ {1} + (1 - t) x _ {0} \tag {3} +$$ + +defines a linear interpolation between noise and data, and + +$$ +u \left(x _ {t} \mid x _ {0}, x _ {1}\right) = x _ {1} - x _ {0} \tag {4} +$$ + +denotes its corresponding flow velocity at $x_{t}$ . Note, simply dropping the condition $c$ leads to the unconditional flow matching formulation. + +Although rectified flow matching is trained with straight paths, the learned flow is often curved (Figure 1, bottom-left) since there are many possible paths induced by the independently sampled couplings $x_0, x_1$ through any $x_t$ at time $t$ [29]. Curved paths require more numerical integration and network evaluation steps, which is not ideal. + +Optimal Transport (OT) Coupling. To address this issue for unconditional generation, Tong et al. [35] and Pooladian et al. [29] concurrently proposed to use minibatch optimal transport to deterministically couple the sampled prior and a sampled batch of the data. This minimizes flow path lengths and straightens the flow. Concretely, given a minibatch of $b$ $e$ -dimensional samples, an OT coupling seeks a $b \times b$ permutation matrix $P$ such that the squared Euclidean distance is minimized, i.e., + +$$ +\min _ {P} \| X _ {0} - P X _ {1} \| _ {2} ^ {2}. \tag {5} +$$ + +Here, $X_0, X_1 \in \mathbb{R}^{b \times e}$ are matrices containing a minibatch of samples from the prior and the data. Instead of an "independent coupling", in the unconditional setting we then optimize a variant of Equation (1) using the coupled tuple $(X_0, PX_1)$ . Note that this algorithm corresponds to sampling tuples $(x_0, x_1)$ from a joint distribution $\pi_{\mathrm{ot}}(x_0, x_1)$ . + +Importantly, for unconditional generation, i.e., when using the joint distribution $\pi_{\mathrm{ot}}(x_0,x_1)$ , the marginals remain unchanged from independent coupling (see [35]): + +$$ +\int \pi_ {\mathrm {o t}} (x _ {0}, x _ {1}) q (x _ {0}) \mathrm {d} x _ {0} = q (x _ {1}), \text {a n d} +$$ + +$$ +\int \pi_ {\mathrm {o t}} \left(x _ {0}, x _ {1}\right) q \left(x _ {1}\right) \mathrm {d} x _ {1} = q \left(x _ {0}\right). \tag {6} +$$ + +Hence, during test-time, we can sample from the prior $q(x_0)$ and integrate the flow $v_{\theta}$ to simulate data from the data distribution $q(x_{1})$ . + +Unfortunately, as shown in the next section, this does not hold for conditional generation. Said differently, naively extending the unconditional joint distribution $\pi_{\mathrm{ot}}(x_0,x_1)$ to a conditional joint distribution $\pi_{\mathrm{ot}}(x_0,x_1,c)$ by using the coupled tuple $(X_0,PX_1,PC)$ , where $C\in \mathbb{R}^{b\times h}$ is a batch of $h$ -dimensional conditions, does not maintain marginals and leads to a gap between training time and test time. We discuss this next. + +# 3.2. OT Coupling Skews The Prior + +Intuition. Consider the one-dimensional example illustrated in Figure 2: the prior distribution $q(x_0)$ follows a Gaussian distribution, and the target distribution $q(x_1)$ consists of a mixture of two Gaussians translated left and right, with the translation direction indicated by a binary condition $c$ . When training with a $\pi_{\mathrm{ot}}(x_0,x_1,c)$ that represents the coupled tuple $(X_0, PX_1,PC)$ , samples from the prior and the sampled conditions become correlated through the OT coupling – samples from the left half of the prior distribution are always associated with $c = 0$ , while those from the right half always correspond to $c = 1$ . As a result, the model overfits on data pairs $(x_0,c)$ where either $(x_0 < 0,c = 0)$ or $(x_0 > 0,c = 1)$ . However, during testing, this skewed distribution is no longer accessible, and instead, $x_0$ and $c$ are sampled independently. Consequently, the model encounters previously unseen data pairs, i.e., $(x_0 > 0,c = 0)$ and $(x_0 < 0,c = 1)$ – leading to failure. A similar phenomenon can be observed in the two-dimensional example illustrated in Figure 1. + +Skewed Prior. Formally, imagine that we consider the most general OT coupling joint distribution $\pi_{\mathrm{ot}}(x_0,x_1,c)$ at training time. At test time, the user wants to specify a condition $c_{1}$ to obtain samples from the conditional distribution $q(x_{1}|c = c_{1})$ . What is the prior distribution that we need to sample from such that we arrive at the correct conditional distribution $q(x_{1}|c = c_{1})$ ? To compute this, imagine, we are given an arbitrary condition $c_{1}$ , i.e., we look at the induced distribution $\pi_{\mathrm{ot}}(x_0,x_1,c|c = c_1)$ . Marginalizing this induced distribution over $x_{1}$ yields + +$$ +\begin{array}{l} \int \pi_ {\mathrm {o t}} \left(x _ {0}, x _ {1}, c \mid c = c _ {1}\right) \mathrm {d} x _ {1} = \int \pi_ {\mathrm {o t}} \left(x _ {0}, x _ {1} \mid c = c _ {1}\right) \mathrm {d} x _ {1} \\ = q \left(x _ {0} \mid c = c _ {1}\right), \tag {7} \\ \end{array} +$$ + +which implies that the prior required to arrive at the conditional distribution $q(x_{1}|c = c_{1})$ is $q(x_{0}|c = c_{1})$ - a distribution that we cannot access at test time. In general, the prior that we need to sample from, given condition $c$ , is skewed from $q(x_{0})$ to $q(x_{0}|c)$ which is inaccessible at test time. As a result, accurately capturing $q(x_{1}|c)$ at test time becomes infeasible unless $q(x_{0})$ and $q(c)$ are independent, i.e., if $q(x_{0}|c) = q(x_{0})$ . However, this condition is generally not satisfied by the most general OT coupling – for instance, in Figure 2, clearly $q(x_{0}|c) \neq q(x_{0})$ . To address this issue, we propose a method $C^2\mathrm{OT}$ to unkew the prior. We discuss this in the next section. + +# 3.3. Conditional Optimal Transport FM + +# 3.3.1. Formulation + +We propose conditional optimal transport FM (C²OT) to unkew the prior $q(x_0|c)$ such that $q(x_0|c) = q(x_0)$ . This + +ensures that at test time, we can sample from the full prior $q(x_0)$ , irrespective of the condition $c$ . Importantly, at the same time, we also aim to preserve the straight flow paths provided by OT. Conceptually, we construct a prior distribution independently for each condition, i.e., $q(x_0|c_1) = q(x_0|c_2) = q(x_0), \forall c_1, c_2$ . This can be achieved by sampling from the prior and computing OT independently for each condition, as shown in Figure 2 (right). + +Formally, we construct the joint distribution as + +$$ +\pi_ {\mathrm {c} ^ {2} \mathrm {o t}} (x _ {0}, x _ {1}, c) := q (x _ {0}) q (c) \pi_ {\mathrm {o t}} (x _ {1} | x _ {0}, c). \tag {8} +$$ + +Here, the prior $q(x_0)$ and the condition $q(c)$ are sampled independently, while the data $x_1$ is conditioned on $x_0$ (via OT) and $c$ (via the dataset construction). Different from independent coupling in Equation (2), $x_1$ and $x_0$ are associated through OT and therefore lead to straighter flow paths. Also different from the most general OT coupling $\pi_{\mathrm{ot}}(x_0,x_1,c)$ , we explicitly enforce the independence between $x_0$ and $c$ . + +To see that this joint distribution provides an unskewed prior, imagine we are given an arbitrary input condition $c_{1}$ , marginalizing over $x_{1}$ yields + +$$ +\begin{array}{l} \int q (x _ {0}) q (c | c = c _ {1}) \pi_ {\mathrm {o t}} \left(x _ {1} \mid x _ {0}, c = c _ {1}\right) \mathrm {d} x _ {1} \\ = \int q \left(x _ {0}\right) \pi_ {\mathrm {o t}} \left(x _ {1} \mid x _ {0}, c = c _ {1}\right) \mathrm {d} x _ {1} = q \left(x _ {0}\right). \tag {9} \\ \end{array} +$$ + +This implies that, at test time, we can sample from the entire prior $q(x_0)$ , regardless of the condition $c$ , to arrive at the desired data distribution $q(x_1|c)$ . During training, in practice, we sample from $\pi_{\mathrm{c}^2\mathrm{ot}}$ by modifying the optimal transport cost function in Equation (5). This process is exact for discrete conditions (Section 3.3.2) and approximate for continuous conditions (Section 3.3.3). Specifically, given a minibatch of $b$ $e$ -dimensional samples and a minibatch of $b$ $h$ -dimensional conditions $C \in \mathbb{R}^{b \times h}$ , $\mathbf{C}^2\mathrm{OT}$ seeks a $b \times b$ permutation matrix $P$ that minimizes the following cost function: + +$$ +\min _ {P} \| X _ {0} - P X _ {1} \| _ {2} ^ {2} + \sum_ {i = 1} ^ {b} f \left(C _ {i}, [ P C ] _ {i}\right), \tag {10} +$$ + +where $f$ is a symmetric and non-negative cost function, satisfying the property $f(c, c) = 0 \forall c$ . The key differences between independent coupling, OT coupling, and $\mathrm{C}^2\mathrm{OT}$ coupling are highlighted in Algorithms 1 to 3. In the following two sections, we discuss our choice of $f$ for both discrete and continuous conditions. + +# 3.3.2. Discrete Conditions + +For discrete conditions, we assume the conditions correspond to class labels from a finite, discrete set, i.e., $c \in \{0, 1, \ldots, k - 1\}$ . For example, in Figure 2, the labels 0 and 1 represent the left and right groups in the target distribution, + +Algorithm 1 Independent Coupling +1: function INDEPENDENT(X1, C) +2: X0 ~ N ▷ Sample from prior +3: +4: +5: return X0, X1, C + +Algorithm 2 OT Coupling +1: function OT(X1, C) +2: X0 ~ N ▷ Sample from prior +3: Cost ← pairwise_dist(X0, X1) +4: i, j ← Hungarian_MATCHing(Cost) +5: return X0[i], X1[j], C[j] + +Algorithm 3 $\mathrm{C}^2\mathrm{OT}$ Coupling (Ours) +1: function $\mathrm{C}^2\mathrm{OT}(X_1,C)$ +2: $X_0\sim \mathcal{N}$ ▷ Sample from prior +3: Cost $\leftarrow$ pairwise_dist(Xo,X1)+ pairwise_dist(C,C) +4: $i,j\gets$ Hungarian_MATCHing(Cost) +5: return $X_0[i],X_1[j],C[j]$ + +respectively. We define $f = f_{\mathrm{d}}$ such that no transport is allowed to modify the conditions (i.e., $PC = C$ ). This is achieved via + +$$ +f _ {\mathrm {d}} \left(c _ {1}, c _ {2}\right) = \left\{ \begin{array}{l} \infty , c _ {1} \neq c _ {2}, \\ 0, \text {o t h e r w i s e .} \end{array} \right. \tag {11} +$$ + +As a result, the unordered set of prior samples for any given condition $c$ remains unchanged: $\{X_{0i}|c_i = c\} \stackrel{\mathrm{iid}}{\sim} q(x_0)$ . This formulation samples exactly from our ideal (Equation (8)), as OT is computed independently within each class, and the data sample from each class is exposed to the full prior without skew. This effect is illustrated in Figure 1 (middle) and Figure 2 (right). As expected, the results differ from FM's curved flow paths and OT's inability to properly capture the target distribution. + +It is important to note that enforcing $PC = C$ does not necessarily imply $P = I$ since we can still swap row $i$ and $j$ in $P$ as long as $c_{i} = c_{j}$ . Setting $P = I$ would correspond to a degeneration back to independent coupling in Equation (2) and forfeit the benefits of straightened flow provided by OT. + +Unfortunately, using $f_{\mathrm{d}}$ with continuous conditions leads to exactly this degeneration, since, in general, no $c_{i} = c_{j}$ unless $i = j$ . Hence, in the next section, we devise a relaxed cost for handling continuous conditions. + +# 3.3.3. Continuous Conditions + +For continuous conditions, the condition is typically a feature embedding, e.g., computed from a text prompt. Directly applying $f_{\mathrm{d}}$ in Equation (11) leads to a degeneration back to independent coupling as discussed earlier. To avoid this, we propose a relaxed penalty function $f_{\mathrm{c}}$ based on the cosine distance: + +$$ +f _ {\mathrm {c}} \left(c _ {1}, c _ {2}\right) = w \left(1 - \frac {c _ {1} \cdot c _ {2}}{\| c _ {1} \| \| c _ {2} \|}\right) = w \cdot \operatorname {c d i s t} \left(c _ {1}, c _ {2}\right). \tag {12} +$$ + +Here, $w > 0$ is a scaling hyperparameter. Note, when $w = 0$ , this formulation degenerates to regular OT; when $w \to \infty$ , this formulation approaches Equation (11), i.e., independent coupling if no two conditions are equal. The right section of Figure 1 illustrates an example where the $x$ -coordinate of the target data point serves as the condition. Notably, $C^2\mathrm{OT}$ preserves straight flow paths without exhibiting OT's apparent degradation. + +Finding $w$ . The optimal $w$ is highly dependent on the data distribution and the magnitude of features. To alleviate the need for tuning the hyperparameter $w$ , we propose to find $w$ adaptively for each minibatch. For this, we develop a ratio $r(w)$ which measures the proportion of samples that are considered "potential optimal transport candidates", i.e., + +$$ +\begin{array}{l} r (w) = \frac {1}{B ^ {2}} \sum_ {i, j} \mathbb {1} \left(\left\| X _ {0 i} - X _ {1 j} \right\| _ {2} ^ {2} + w \cdot \operatorname {c d i s t} \left(c _ {i}, c _ {j}\right) \right. \\ \leq \left\| X _ {0 i} - X _ {1 i} \right\| _ {2} ^ {2}), \tag {13} \\ \end{array} +$$ + +where $\mathbb{1}(\cdot)$ is an indicator function that evaluates to one if the condition is satisfied and to zero otherwise. Then, we introduce a hyperparameter $r_{\mathrm{tar}}$ as the target ratio and find $w$ such that $r(w) \approx r_{\mathrm{tar}}$ . Note that setting $r_{\mathrm{tar}}$ is invariant to the scaling of distances between $x_0$ and $x_1$ . Namely, if $\|X_{0i} - X_{1j}\|_2^2$ is scaled by $s$ , $w$ would also be scaled by $s$ to preserve the same $r_{\mathrm{tar}}$ . This invariance is particularly useful when transitioning to a new dataset, such as changing image resolutions in image generation tasks. In our implementation, we search for $w$ in two steps: first with exponential search to establish the upper/ lower bounds, then with binary search to locate a precise value. The final $w$ is used as the initial value in the subsequent minibatch. In practice, we observe $w$ differs very little between minibatches ( $< 1\%$ ) and the search converges quickly within 10 iterations with a negligible overhead. Next, we introduce oversampling, a simple technique to obtain more accurate OT couplings. + +# 3.3.4. Oversampling OT Batches + +OT vs. Deep Net Batch Sizes. Typically, the "OT batch size" $b$ used to compute optimal transport is set equal to the "deep net batch size" $b_{n}$ used to compute forward/backward passes of the deep net [29, 35]. While the OT batch size controls the dynamics of prior-to-data assignments, the deep net batch size affects gradient variances, training efficiency, and memory usage. There is no reason why setting them equal should be optimal. Particularly, with our proposed $C^2\mathrm{OT}$ , the effective OT batch size is reduced since we restrict transport between samples with divergent conditions, which calls for an increased OT batch size $b$ . At the same time, keeping the deep net batch size $b_{n}$ unchanged prevents unwarranted side effects in other aspects of training. + +Reduced Effective OT Batch Size. In the case of discrete conditions, OT is effectively performed within + +Table 1. Results on the 8gaussians $\rightarrow$ moons dataset with different training coupling methods and different conditioning. Note that our method $C^2\mathrm{OT}$ does not apply in the unconditional setting. + +
MethodEuler-1 (W2↓)Adaptive (W2↓)NFE ↓
Unconditional
FM6.232±0.1860.120±0.04593.22±1.81
OT0.072±0.0250.050±0.04036.08±2.57
Binary conditions
FM2.573±0.1120.059±0.02991.10±2.77
OT0.483±0.0590.462±0.02532.89±0.95
C2OT (ours)0.048±0.0100.018±0.00658.88±1.48
Continuous conditions
FM0.732±0.0520.028±0.01093.86±1.76
OT8.276±6.5102.143±1.99342.07±2.59
C2OT (ours)0.077±0.0240.013±0.00389.87±1.53
+ +$\{X_{0i}, X_{1i} | c_i = c\}$ for each condition $c$ . With $k$ classes, the expected effective OT batch size is reduced to $\mathbb{E}(|\{i | c_i = c\}|) = b / k$ . Empirically, we find that using a much larger OT batch size $b$ than the network batch size $b_n$ is helpful. This observation motivates oversampling. + +Efficiency. A natural concern with increasing $b$ is the computational overhead, since optimal transport with Hungarian matching [20] has a time complexity of $O(b^{3})$ . However, + +- Since each OT batch can be used across $(b / b_{n})$ forward/backward passes, the amortized cost per minibatch reduces to $O(b^{2}b_{n})$ . +- We offload OT computation from the main process to the data workers, enabling data preparation on the CPU in parallel with forward/backward passes on the GPU. + +With a modest choice of 8 data workers per GPU, we observe no wall-clock overhead for OT batch sizes up to 6,400 when training on CIFAR-10 and ImageNet, as data preparation is completed faster than the GPU can process batches. + +# 4. Experiments + +We experimentally verify our proposed method $\mathrm{C}^2\mathrm{OT}$ in two sections: first on the two-dimensional synthetic dataset 8gaussians $\rightarrow$ moons, and then on high-dimensional image data, including CIFAR-10 [19] and ImageNet-1K [7]. For ImageNet-1K, we conduct experiments on both $32\times 32$ images in image space and $256\times 256$ images in latent space. Implementation details are provided in the appendix. + +# 4.1. Two-Dimensional Data + +The two-dimensional 8gaussians $\rightarrow$ moons dataset maps a mixture of eight Gaussian distributions to two interleaving half-circles, as visualized in Figure 1. We consider three conditioning scenarios: (a) unconditional; (b) binary + +class conditions, where each class corresponds to one of the half-circles; and (c) continuous conditions, where the $x$ -coordinate of the desired target point is given as the condition. For (b), we apply our setup for discrete conditions (Section 3.3.2); for (c), we apply our setup for continuous conditions (Section 3.3.3) with $r_{\mathrm{tar}} = 0.01$ . + +Metrics. We report the 2-Wasserstein distance $(W_2^2)$ (following [35]) between 10K generated data points and 10K samples from the ground-truth moons distribution. We experiment with single-step Euler's method (Euler-1) and the adaptive Dormand-Prince method [8] (dopri5, referred to as adaptive) for numerical integration. Additionally, we report the number of function evaluations (NFE) in the adaptive method, which is the number of forward passes through the neural network. We average the results over ten training runs with different random seeds and report mean±std. + +Results. The results are summarized in Table 1 and visualized in Figure 1. In the unconditional setting, OT produces straighter flows, leading to better distribution matching with fewer NFEs, and our method does not apply. However, in the conditional setting, OT degrades and performs worse than regular FM in the adaptive setting due to the aforementioned train-test gap in OT coupling. Our method, $\mathrm{C}^2\mathrm{OT}$ , effectively models the target distribution with both Euler-1 and the adaptive method, improving upon the baselines. + +# 4.2. High-Dimensional Image Data + +We conduct experiments on CIFAR-10 [19] (image space, class-conditioned), ImageNet- $32 \times 32$ [7] (image space, caption-conditioned), and ImageNet- $256 \times 256$ [7] (latent space, caption-conditioned). For ImageNet, we use the captions provided by [21] and encode text features using a CLIP [30]-like DFN [10] text encoder as input conditions. Our results are summarized in Table 2. + +Metrics. We report Fréchet inception distance (FID) [13] and the number of function evaluations (NFE) in the adaptive solver for all three datasets. To assess how well the generations adhere to the input conditions, we additionally compute condition adherence metrics. For CIFAR-10, we run a pretrained classifier [5] on the generated images to obtain logits and report the average cross entropy (CE) with the input class label. For ImageNet, we compute CLIP features [30] using SigLIP-2 [36] from the generated images and the input captions respectively, and report the average cosine similarities (CLIP). + +CIFAR-10. We base our UNet architecture on Tong et al. [35], with details provided in the appendix. The CIFAR-10 training set contains 50,000 $32 \times 32$ color images across 10 object classes. During testing, we generate 50,000 images evenly distributed among classes and assess FID against the training set following [35]. Each model is trained five times with different random seeds and the results are reported + +Table 2. Results of different training algorithms on image generation tasks. We bold the best-performing entry and underline the second-best. Our proposed method, $C^2 OT$ , achieves the best overall performance – better than FM with few sampling steps and better than OT with more sampling steps. In cases where $C^2 OT$ performs second, our performance is often comparable to the best entry. Additionally, $C^2 OT$ has the most stable performance, with the smallest deviations across runs in most entries. Variations in CLIP scores are minimal. + +
CIFAR-10 Class-Conditioned Generation
MethodEuler-2Euler-5Euler-10Euler-25Euler-50Euler-100AdaptiveNFE ↓
FID → FM105.481±1.61921.968±0.40010.479±0.2375.640±0.1174.128±0.0693.429±0.0542.922±0.039124.0±2.5
OT64.417±0.43418.194±0.40910.778±0.2706.821±0.1865.368±0.1134.671±0.0664.144±0.018125.7±0.4
C²OT (ours)64.694±0.37917.565±0.2029.762±0.1365.644±0.0814.152±0.0613.447±0.0492.904±0.035127.7±0.7
CE → FM3.343±0.0220.546±0.0200.323±0.0070.272±0.0040.267±0.0020.269±0.0030.276±0.004124.0±2.5
OT2.529±0.0100.624±0.0120.426±0.0050.369±0.0040.363±0.0040.364±0.0050.372±0.005125.7±0.4
C²OT (ours)2.264±0.0220.475±0.0090.322±0.0040.276±0.0050.271±0.0030.272±0.0030.279±0.004127.7±0.7
ImageNet 32×32 Caption-Conditioned Generation
FID → FM113.250±0.79323.015±0.12311.141±0.0517.165±0.0876.142±0.0855.673±0.0745.358±0.059146.9±3.5
OT81.317±0.36920.763±0.15411.360±0.0897.854±0.0776.899±0.0796.469±0.0716.220±0.065137.7±0.2
C²OT (ours)102.380±0.27921.965±0.03510.897±0.0337.069±0.0276.084±0.0165.638±0.0185.350±0.017134.2±1.4
CLIP → FM0.067±0.0000.076±0.0000.074±0.0000.071±0.0000.071±0.0000.070±0.0000.069±0.000146.9±3.5
OT0.067±0.0000.071±0.0000.070±0.0000.068±0.0000.068±0.0000.067±0.0000.067±0.000137.7±0.2
C²OT (ours)0.068±0.0000.075±0.0000.073±0.0000.071±0.0000.070±0.0000.070±0.0000.069±0.000134.2±1.4
ImageNet 256×256 Caption-Conditioned Generation in Latent Space
FID → FM202.519±0.60631.266±0.3559.996±0.1065.189±0.0823.886±0.0663.484±0.0553.498±0.054131.1±2.3
OT193.376±1.66542.780±2.13215.638±1.5107.514±1.0345.693±0.9985.360±1.0165.726±1.074125.6±6.5
C²OT (ours)200.811±0.38530.715±0.22110.000±0.0725.069±0.0373.722±0.0373.320±0.0323.369±0.034128.2±2.6
CLIP → FM0.028±0.0000.120±0.0000.134±0.0000.137±0.0000.138±0.0000.139±0.0000.140±0.000131.2±2.3
OT0.031±0.0000.107±0.0030.120±0.0030.123±0.0030.124±0.0030.124±0.0040.125±0.004125.6±6.5
C²OT (ours)0.029±0.0000.120±0.0000.133±0.0000.137±0.0000.138±0.0000.138±0.0000.139±0.000128.2±2.6
+ +as mean±std. We use an OT batch size $b$ of 640 and a ratio $r_{\mathrm{tar}}$ of 0.01. Compared to FM, our method achieves better performance (FID and CE) in few-steps settings and comparable performance in many-steps settings. Compared to OT, our method significantly outperforms in the many-step setting. Although OT achieves slightly better FID in Euler-2, it performs worse in CE, indicating that it fails to follow the condition. + +ImageNet $32 \times 32$ . We base our UNet architecture on Pooladian et al. [29], and list details in the appendix. We train and evaluate on the face-blurred [37] ImageNet, following [29]. In contrast to prior works that assess FID against the training set, we assess FID against the validation set (49,997 images), as the fine-grained nature of captions may lead to overfitting on the training set. Each model is trained three times with different random seeds and the results are reported as mean $\pm$ std. Compared to CIFAR-10, there are much more variations in this dataset (e.g., 1000 classes vs. 10 classes in CIFAR-10). Thus, we adopt a larger OT batch size $b$ of 6400 while keeping the same target ratio $r_{\mathrm{tar}}$ 0.01 as before. Similar to our findings in CIFAR-10, our method performs better or is comparable to FM; in few-steps settings, OT achieves better FID but worse condition following (CLIP). Our method strikes the best overall balance. + +ImageNet $256 \times 256$ in Latent Space. Here, we explore latent flow matching models [31]. During training, the flow matching model learns to generate data in the latent space, which is encoded from images. At test-time, the generated latents are decoded into images using a pretrained decoder. We base our experiment on a recent transformer-based model, LightningDiT [39], demonstrating that our method is effective across different network architectures and data spaces. We use the officially provided "64 epochs" configuration for all experiments (due to computational constraints), which is not directly comparable to methods trained with 800 epochs. We follow the same evaluation protocol as ImageNet- $32 \times 32$ , and adopt the same OT batch size $b$ of 6400 and the target ratio $r_{\mathrm{tar}}$ of 0.01. Each model is trained five times with different random seeds, and the results are reported as mean $\pm$ std. Again, $\mathrm{C}^2\mathrm{OT}$ has the best overall performance compared to both FM and OT. We visualize the generations in Figure 3. + +Limitations in High-Dimensional Data We observed that the benefits of $\mathrm{C}^2\mathrm{OT}$ are less pronounced, though still visible, in higher-dimensional image data compared to low-dimensional data. We hypothesize this is because image data is more entangled, making the prior skew induced by OT less detrimental. We also observe statistically straighter paths in $\mathrm{C}^2\mathrm{OT}$ compared to FM, with details given in the appendix. + +![](images/1bdbf61a6d1473856e6f3ed0ddb35d54c0dcb10afd6a447fdf531862d4a28611.jpg) +Figure 3. Visual comparisons of $256 \times 256$ images generated by the baselines and our approach with different amounts of sampling steps. Our approach converges faster and produces cleaner outputs. The input caption is "a small dog sitting on a carpet". + +# 4.3. Ablation Studies + +# 4.3.1. Varying OT Batch Size $b$ + +Here, we analyze the effect of varying the OT batch size $b$ . We conduct these experiments on the CIFAR-10 dataset, training each model three times with different random seeds for every choice of $b$ . We plot the results in Figure 4. Increasing the OT batch size clearly improves FID in the few-sampling-steps setting (Euler-2). This is intuitive, because a larger OT batch size enables the coupling to approach the true OT results, i.e., dataset-level optimal transport. In turn, this leads to straighter paths. However, in the adaptive setting, increasing the OT batch size slightly worsens FID, though the effect is much less significant (the range of the blue y-axis is much smaller than the range of the red y-axis). We think this occurs because a more accurate OT coupling reduces variations in the training data (similar to having less data augmentations) which harms performance as the benefit of straighter flows diminishes in the adaptive setting. This observation aligns with the findings of Pooladian et al. [29]: computing OT coupling across GPUs (i.e., having a larger effective OT batch size) slightly harms results. In contrast to Pooladian et al. [29], though, we seek to counteract the reduced effective OT batch size induced by conditional optimal transport (as discussed in Section 3.3.4), and we do find reasonably large (640 for CIFAR, 6400 for ImageNet) OT batch sizes useful while not adding wall-clock overhead. + +# 4.3.2. Varying Target Ratio $r_{\mathrm{tar}}$ + +Here, we analyze the effects of varying the target ratio $r_{\mathrm{tar}}$ , as defined in Equation (13). We conduct these experiments on the ImageNet- $32 \times 32$ dataset, training each model three times with different random seeds for every choice of $r_{\mathrm{tar}}$ . We plot the results in Figure 5. Recall that when $r \to 0$ , $w \to \infty$ and the method approaches FM; when $r \to 0.5$ , + +![](images/97a9d202d6d745a3acf16bec93aa699b9e76f37042bdf144a38e357b02329111.jpg) +Figure 4. Changes in FID with respect to varying OT batch sizes $b$ . We plot mean±std over three runs and represent std with a shaded region. OT (adaptive) and FM (Euler-2) perform worse than all other results in this plot, hence results are not shown (full plot in the appendix). + +![](images/021a2927ace7652996892924796d7b1493956ce09caf1ec68644a6c328feefd0.jpg) +Figure 5. Changes in FID with respect to varying target ratio $r_{\mathrm{tar}}$ . We plot mean±std over three runs and represent std with a shaded region. OT (adaptive) performs worse than all other results in this plot, hence results are not shown (full plot in the appendix). + +$w \to 0$ and the method approaches OT (Section 3.3.3). Our findings are similar to those in Section 4.3.1: as we approach true OT, whether by increasing OT batch size or increasing $r_{\mathrm{tar}}$ , FID improves in the few-step setting and worsens in the adaptive setting. + +# 5. Conclusion + +We first investigate and formalize the struggle of minibatch optimal transport in conditional generation. Based on the findings, we propose $\mathrm{C}^2\mathrm{OT}$ , a simple yet effective addition that corrects the degeneration of OT while maintaining straight integration paths. Extensive experiments verify the effectiveness of our method. We believe that these findings, along with the proposed simple technique, close a gap in current research on prior-data coupling. + +# References + +[1] Michael Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In Proc. ICLR, 2023. 2 +[2] Michael Albergo, Nicholas Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. 2 +[3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi0: A vision-language-action flow model for general robot control. arXiv, 2024. 1 +[4] Ho Kei Cheng, Masato Ishii, Akio Hayakawa, Takashi Shibuya, Alexander Schwing, and Yuki Mitsufuji. Taming multimodal joint training for high-quality video-to-audio synthesis. CVPR, 2025. 1 +[5] chenyaofo. Pytorch cifar models. https://github.com/chenyaofo/pytorch-cifar-models, 2023.6 +[6] Aram Davtyan, Leello Tadesse Dadi, Volkan Cevher, and Paolo Favaro. Faster inference of flow-based generative models via improved data-noise coupling. In ICLR, 2025. 2 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6 +[8] John R Dormand and Peter J Prince. A family of embedded runge-kutta formulae. Journal of computational and applied mathematics, 1980. 2, 6 +[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1 +[10] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. ICLR, 2024. 6 +[11] Zhengyang Geng, Ashwini Pokle, William Luo, Justin Lin, and J Zico Kolter. Consistency models made easy. ICLR, 2025. 2 +[12] Pengsheng Guo and Alexander G Schwing. Variational rectified flow matching. arXiv, 2025. 2 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6 +[14] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In ICLR, 2017. 2 +[15] Ka-Hei Hui, Chao Liu, Xiaohui Zeng, Chi-Wing Fu, and Arash Vahdat. Not-so-optimal transport flows for 3d point cloud generation. arXiv, 2025. 2 +[16] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. *ICLR*, 2024. 2 + +[17] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. In ICLR, 2014. 2 +[18] Leon Klein, Andreas Kramer, and Frank Noé. Equivariant flow matching. NeurIPS, 36, 2023. 2 +[19] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images, 2009. 6 +[20] Harold W. Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 6 +[21] Visual Layer. Imagenet-1k-vl-enriched. https://huggingface.co/datasets/visual-layer/imagenet-1k-vl-enriched, 2024.6 +[22] Sangyun Lee, Beomsu Kim, and Jong Chul Ye. Minimizing trajectory curvature of ode-based generative models. In ICML, 2023. 2 +[23] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. *ICLR*, 2023. 1, 2 +[24] Qihao Liu, Xi Yin, Alan Yuille, Andrew Brown, and Mannat Singh. Flowing from words to pixels: A framework for cross-modality evolution. arXiv preprint arXiv:2412.15213, 2024. 2 +[25] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. ICLR, 2023. 2 +[26] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. *ICLR*, 2024. 2 +[27] Ségolène Martin, Anne Gagneux, Paul Hagemann, and Gabriele Steidl. Pnp-flow: Plug-and-play image restoration with flow matching. *ICLR*, 2025. 1 +[28] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models. arXiv, 2024. 1 +[29] Aram-Alexandre Pooladian, Heli Ben-Hamu, Carles Domingo-Enrich, Brandon Amos, Yaron Lipman, and Ricky TQ Chen. Multisample flow matching: Straightening flows with minibatch couplings. ICML, 2023. 2, 3, 5, 7, 8 + +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 7 +[32] Gianluigi Silvestri, Luca Ambrogioni, Chieh-Hsin Lai, Yuhta Takida, and Yuki Mitsufuji. Training consistency models with variational noise coupling. arXiv preprint arXiv:2502.18197, 2025. 2 +[33] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2 +[34] Yuxuan Song, Jingjing Gong, Minkai Xu, Ziyao Cao, Yanyan Lan, Stefano Ermon, Hao Zhou, and Wei-Ying Ma. Equivariant flow matching with hybrid probability transport for 3d molecule generation. NeurIPS, 36, 2023. 2 +[35] Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. TMLR, 2024. 2, 3, 5, 6 +[36] Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. arXiv preprint arXiv:2502.14786, 2025. 6 +[37] Kaiyu Yang, Jacqueline H Yau, Li Fei-Fei, Jia Deng, and Olga Russakovsky. A study of face obfuscation in imagenet. In ICML, 2022. 7 +[38] Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining straight flows with velocity consistency. arXiv, 2024. 2 +[39] Jingfeng Yao and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. CVPR, 2025. 7 +[40] Yichi Zhang, Yici Yan, Alex Schwing, and Zhizhen Zhao. Towards hierarchical rectified flow. *ICLR*, 2025. 2 \ No newline at end of file diff --git a/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/images.zip b/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..80312f1d9afaf184b9fc0802ace0eed0021e989f --- /dev/null +++ b/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f06b8728b8a482ed7397d49aac8c60eeff488188ee2e362438578223f7f5185 +size 713505 diff --git a/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/layout.json b/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9dc51cbbf831a09e45461d03937b01c48e42c039 --- /dev/null +++ b/2025/The Curse of Conditions_ Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation/layout.json @@ -0,0 +1,10570 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 536, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 536, + 138 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 536, + 138 + ], + "type": "text", + "content": "The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 205, + 161, + 405, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 161, + 405, + 190 + ], + "spans": [ + { + "bbox": [ + 205, + 161, + 405, + 190 + ], + "type": "text", + "content": "Ho Kei Cheng Alexander Schwing \nUniversity of Illinois Urbana-Champaign" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 234, + 193, + 370, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 193, + 370, + 202 + ], + "spans": [ + { + "bbox": [ + 234, + 193, + 370, + 202 + ], + "type": "text", + "content": "{hokeikc2,aschwing}@illinois.edu" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 65, + 214, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 65, + 214, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 65, + 214, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 65, + 214, + 545, + 403 + ], + "type": "image", + "image_path": "38678b1cb76cbb3b6d3e7c503a10fc6a5665cb3960e9edee95567a426ef7c902.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 406, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 54, + 406, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 406, + 555, + 473 + ], + "type": "text", + "content": "Figure 1. We visualize the flows learned by different algorithms using the 8gaussians " + }, + { + "bbox": [ + 54, + 406, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 54, + 406, + 555, + 473 + ], + "type": "text", + "content": " moons dataset. We compare flow matching (FM), minibatch optimal transport FM (OT), and our proposed conditional optimal transport FM (" + }, + { + "bbox": [ + 54, + 406, + 555, + 473 + ], + "type": "inline_equation", + "content": "C^2 OT" + }, + { + "bbox": [ + 54, + 406, + 555, + 473 + ], + "type": "text", + "content": ") using one Euler step and an adaptive ODE solver, respectively. Below each plot, we show the 2-Wasserstein distance (lower is better; mean " + }, + { + "bbox": [ + 54, + 406, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 54, + 406, + 555, + 473 + ], + "type": "text", + "content": " std over 10 runs). For unconditional generation, OT achieves significantly straighter flows, outperforming FM. However, paradoxically, OT performs worse when conditions are introduced. This paper analyzes this degradation and finds that it occurs because optimal transport disregards conditions, leading to a train-test gap. To address this, we propose a simple fix (" + }, + { + "bbox": [ + 54, + 406, + 555, + 473 + ], + "type": "inline_equation", + "content": "C^2 OT" + }, + { + "bbox": [ + 54, + 406, + 555, + 473 + ], + "type": "text", + "content": ") that outperforms FM and OT in conditional generation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 481, + 200, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 481, + 200, + 494 + ], + "spans": [ + { + "bbox": [ + 152, + 481, + 200, + 494 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "text", + "content": "Minibatch optimal transport coupling straightens paths in unconditional flow matching. This leads to computationally less demanding inference as fewer integration steps and less complex numerical solvers can be employed when numerically solving an ordinary differential equation at test time. However, in the conditional setting, minibatch optimal transport falls short. This is because the default optimal transport mapping disregards conditions, resulting in a conditionally skewed prior distribution during training. In contrast, at test time, we have no access to the skewed prior, and instead sample from the full, unbiased prior distribution. This gap between training and testing leads to a subpar performance. To bridge this gap, we propose conditional optimal transport " + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "inline_equation", + "content": "(C^2 OT)" + }, + { + "bbox": [ + 54, + 510, + 297, + 713 + ], + "type": "text", + "content": " that adds a conditional weighting term in the cost matrix when computing the optimal transport assignment. Experiments demonstrate that this simple fix works with both discrete and continuous conditions in 8gaussians→moons," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 482, + 556, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 556, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 556, + 531 + ], + "type": "text", + "content": "CIFAR-10, ImageNet-32×32, and ImageNet-256×256. Our method performs better overall compared to the existing baselines across different function evaluation budgets. Code is available at hkchengrex.github.io/C2OT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 559, + 395, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 559, + 395, + 572 + ], + "spans": [ + { + "bbox": [ + 314, + 559, + 395, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 582, + 556, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 556, + 668 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 556, + 668 + ], + "type": "text", + "content": "We focus on flow-matching-based conditional generative models, i.e., generation guided by an input condition. Examples of such conditions include class labels, text, or video. Recently, flow matching has been applied in many areas of computer vision, including text-to-image/video [9, 28], vision-language applications [3], image restoration [27], and video-to-audio [4]. However, these methods can be slow at" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 674, + 555, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 555, + 715 + ], + "type": "text", + "content": "1We use condition to denote an input condition, not to be confused with the 'condition' in conditional flow matching (CFM) [23] (where the second 'C' in our acronym comes from), which refers to a data sample. To avoid ambiguity, we refer to CFM simply as flow matching (FM) in this paper." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15875" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 295, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 295, + 228 + ], + "type": "text", + "content": "test-time – obtaining a solution requires numerically integrating the flow with an ODE solver, typically involving many steps, each requiring a deep network forward pass. Naively reducing the number of steps degrades performance, as the underlying flow is often curved (Figure 1, leftmost column), necessitating a small step size for accurate numerical integration. One way to address this issue is by straightening the flow. In the context of unconditional generation, Tong et al. [35] and Pooladian et al. [29] concurrently proposed \"minibatch optimal transport\" (OT), which deterministically couples data and samples from the prior within a minibatch via optimal transport (replacing random coupling) to minimize flow path lengths and to straighten the flow." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 228, + 297, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 297, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 297, + 431 + ], + "type": "text", + "content": "While OT improves unconditional generation, we find that it paradoxically and consistently harms conditional generation (Figure 1). This occurs because OT disregards the conditions when computing the coupling. As a result, during training, OT samples from a prior distribution that is skewed by the condition, as we show in Section 3.2. However, at test time, we have no access to this skewed distribution and instead sample from the full prior distribution. This mismatch creates a gap between training and testing, leading to performance degradation. To bridge this gap, we propose conditional optimal transport FM " + }, + { + "bbox": [ + 56, + 228, + 297, + 431 + ], + "type": "inline_equation", + "content": "(\\mathrm{C}^2\\mathrm{OT})" + }, + { + "bbox": [ + 56, + 228, + 297, + 431 + ], + "type": "text", + "content": " which introduces a simple yet effective condition-aware weighting term when computing the cost matrix for OT. Additionally, we propose two techniques: adaptive weight finding to simplify hyperparameter tuning, and efficient oversampling of OT batches to counteract the reduced effective OT batch size introduced by the weighting term." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 432, + 297, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 297, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 297, + 563 + ], + "type": "text", + "content": "We conduct extensive experiments on both two-dimensional synthetic data and high-dimensional image data, including CIFAR-10, ImageNet-32 " + }, + { + "bbox": [ + 55, + 432, + 297, + 563 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 432, + 297, + 563 + ], + "type": "text", + "content": " 32, and ImageNet-256 " + }, + { + "bbox": [ + 55, + 432, + 297, + 563 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 432, + 297, + 563 + ], + "type": "text", + "content": " 256. The results demonstrate that our method performs well across different condition types (discrete and continuous), datasets, network architectures (UNets and transformers), and data spaces (image space and latent space). " + }, + { + "bbox": [ + 55, + 432, + 297, + 563 + ], + "type": "inline_equation", + "content": "C^2\\mathrm{OT}" + }, + { + "bbox": [ + 55, + 432, + 297, + 563 + ], + "type": "text", + "content": " achieves better overall performance than the existing baselines across different inference computation budgets, including with few-steps Euler's method and with an adaptive ordinary differential equation (ODE) solver [8]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 148, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 148, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 148, + 594 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "Flow Matching. Recently, flow matching [1, 2, 23] has become a popular choice for generative modeling, in part due to its simple training objective and its ability to generate high-quality samples. However, at test time, flow-based methods typically require many computationally expensive forward passes through a deep net to numerically integrate the flow. This is because the underlying flow is usually curved and therefore cannot be approximated well with a few integration steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "text", + "content": "Optimal Transport Coupling. To straighten the flow, in the context of unconditional generation, Tong et al. [35] and Pooladian et al. [29] concurrently proposed minibatch optimal transport (OT). While OT has proven effective in the unconditional setting, we show in Section 3.2 that it skews the prior distribution and fails in the conditional setting if used naively. Our method, " + }, + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{OT}" + }, + { + "bbox": [ + 313, + 72, + 555, + 276 + ], + "type": "text", + "content": ", specifically addresses this failure, aiming to extend the success of OT in unconditional generation to conditional generation. Note, OT has also been used in equivariant flow matching [18, 34], which exploits domain-specific data symmetry (e.g., in molecules). In contrast, our method targets general data. Further, Hui et al. [15] find that mixing of OT and independent coupling benefits shape completion, which is an orthogonal contribution to this paper. Concurrently, Davtyan et al. [6] improve data-noise coupling for unconditional generation, while leaving data coupling for conditional generation to future work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 277, + 556, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 556, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 556, + 385 + ], + "type": "text", + "content": "Straightening Flow Paths. Other methods of straightening flow paths exist. Reflow [25] achieves this by retraining a flow matching network multiple times, at the cost of additional training overhead. Variational flow matching [12] reduces flow ambiguity with a latent variable and hierarchical flow matching [40] straightens flows by modeling higher-order flows. These approaches are orthogonal to our focus on prior-data coupling and, in principle, can be combined with our method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 387, + 556, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 483 + ], + "type": "text", + "content": "Learning the Prior Distribution. An alternative approach to improving flow is to jointly learn a prior distribution with the flow matching network. However, these methods [22, 24, 32] typically rely on a variational autoencoder (VAE) [14, 17] to learn the prior. This introduces yet another density estimation problem, which complicates training. In this work, we focus on improving training-free coupling plans without learning a new prior." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 485, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 556, + 605 + ], + "type": "text", + "content": "Consistency Models. Consistency models [33] have been proposed to accelerate diffusion model sampling, and can be extended to flow matching [38]. Recent advancements have improved training efficiency [11], stability [26], and quality [16]. These methods are orthogonal to our contribution, which focuses on improving prior-data coupling, and can be integrated, as demonstrated by Silvestri et al. [32] in their combination of consistency models with OT. In this paper, we focus on the base form of flow matching to avoid distractions from other formulations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 621, + 370, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 621, + 370, + 633 + ], + "spans": [ + { + "bbox": [ + 314, + 621, + 370, + 633 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 645, + 408, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 408, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 408, + 658 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Flow Matching (FM). We base our discussion on FM [23, 35] (also called rectified flow [25]) for generative modeling and refer readers to [25, 35] for details. At test-time, a sample " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{x}_1" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": " is characterized by an ordinary differential equation" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15876" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 227, + 220 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 227, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 227, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 227, + 220 + ], + "type": "image", + "image_path": "95849445438a1dfb8948167facc0d9d74ca8c25a9fa9634fd62430bcfaca1c11.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "text", + "content": "Figure 2. We visualize the coupling during training and the learned flow from " + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "text", + "content": " during testing for FM, OT, and our proposed method " + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "inline_equation", + "content": "C^2 OT" + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "text", + "content": ". The prior is defined as " + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "inline_equation", + "content": "q(x_0) = \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "text", + "content": ", while the target distribution is given by " + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "inline_equation", + "content": "q(x_{1}) = \\frac{1}{2} q(x_{1}|c = 0) + \\frac{1}{2} q(x_{1}|c = 1)) = \\frac{1}{2}\\mathcal{N}(-2,0.5) + \\frac{1}{2}\\mathcal{N}(2,0.5)" + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "text", + "content": ". Note, FM results in curved flows during testing, while OT degenerates due to skewed training samples (" + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "inline_equation", + "content": "q(x_0|c) \\neq q(x_0)" + }, + { + "bbox": [ + 54, + 221, + 555, + 266 + ], + "type": "text", + "content": "). In contrast, our method successfully learns straight flows without degeneration." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 233, + 70, + 391, + 220 + ], + "blocks": [ + { + "bbox": [ + 233, + 70, + 391, + 220 + ], + "lines": [ + { + "bbox": [ + 233, + 70, + 391, + 220 + ], + "spans": [ + { + "bbox": [ + 233, + 70, + 391, + 220 + ], + "type": "image", + "image_path": "c0f240adc569ff6e5bdcd6d1c715f4ddad481eab94b6fb224163c0c69f1ff9fc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 392, + 70, + 550, + 220 + ], + "blocks": [ + { + "bbox": [ + 392, + 70, + 550, + 220 + ], + "lines": [ + { + "bbox": [ + 392, + 70, + 550, + 220 + ], + "spans": [ + { + "bbox": [ + 392, + 70, + 550, + 220 + ], + "type": "image", + "image_path": "88eea11d8895ec1f0a725be32cf2417094fde3431a1482f27456da5547c6ece1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "text", + "content": "(ODE) with an initial boundary value specified via noise " + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "text", + "content": ", randomly drawn from a prior distribution (e.g., the standard normal). To compute the sample " + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\tilde{x}_1" + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "text", + "content": ", this ODE is commonly solved by numerically integrating from time " + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "text", + "content": " to time " + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "text", + "content": " a learned flow/velocity field " + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "text", + "content": " denotes the set of trainable deep net parameters. For conditional generation, we additionally obtain a condition " + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "text", + "content": ", e.g., a class label from user input, which adjusts the flow. We use " + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "inline_equation", + "content": "v_{\\theta,c}" + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "text", + "content": " to denote this conditional flow. At training time, we find the deep net parameters " + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 275, + 296, + 396 + ], + "type": "text", + "content": " by minimizing the objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 405, + 296, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 405, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 65, + 405, + 296, + 420 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {t, \\left(x _ {0}, x _ {1}, c\\right) \\sim \\pi \\left(x _ {0}, x _ {1}, c\\right)} \\| v _ {\\theta , c} (t, x _ {t}) - u \\left(x _ {t} \\mid x _ {0}, x _ {1}\\right) \\| ^ {2}, \\tag {1}", + "image_path": "bbb0b59efc70b0f90122be441b60e7844da597ae78c2117ccf0f115ff5e52ac1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 429, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 429, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 54, + 429, + 296, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 429, + 296, + 489 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 429, + 296, + 489 + ], + "type": "text", + "content": " is uniformly drawn from the interval " + }, + { + "bbox": [ + 54, + 429, + 296, + 489 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 54, + 429, + 296, + 489 + ], + "type": "text", + "content": ", and where " + }, + { + "bbox": [ + 54, + 429, + 296, + 489 + ], + "type": "inline_equation", + "content": "\\pi (x_0,x_1,c)" + }, + { + "bbox": [ + 54, + 429, + 296, + 489 + ], + "type": "text", + "content": " denotes the joint distribution of the prior and the training data. In practice, often, independent coupling is used, i.e., the prior and the training data are sampled independently, such that" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 115, + 501, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 501, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 115, + 501, + 295, + 514 + ], + "type": "interline_equation", + "content": "\\pi \\left(x _ {0}, x _ {1}, c\\right) = q \\left(x _ {0}\\right) q \\left(x _ {1}, c\\right). \\tag {2}", + "image_path": "ef32845db4047a9d2cb4207cfcef78198e50aecd7495605d660d5f6ad402ba7f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 525, + 90, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 90, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 90, + 536 + ], + "type": "text", + "content": "Further," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 539, + 295, + 551 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 539, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 129, + 539, + 295, + 551 + ], + "type": "interline_equation", + "content": "x _ {t} = t x _ {1} + (1 - t) x _ {0} \\tag {3}", + "image_path": "3092af985b00c809e6501365724c7d0ba3aae2cd623a5b0f3bbdd8fdde8e935c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 558, + 288, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 288, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 288, + 571 + ], + "type": "text", + "content": "defines a linear interpolation between noise and data, and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 126, + 582, + 296, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 582, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 126, + 582, + 296, + 594 + ], + "type": "interline_equation", + "content": "u \\left(x _ {t} \\mid x _ {0}, x _ {1}\\right) = x _ {1} - x _ {0} \\tag {4}", + "image_path": "fd78e21ef7bde7fc7c7adb0c77c3ba642e9455a8c0333ff774b7954331d3c261.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "type": "text", + "content": "denotes its corresponding flow velocity at " + }, + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "type": "text", + "content": ". Note, simply dropping the condition " + }, + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 605, + 295, + 640 + ], + "type": "text", + "content": " leads to the unconditional flow matching formulation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 296, + 713 + ], + "type": "text", + "content": "Although rectified flow matching is trained with straight paths, the learned flow is often curved (Figure 1, bottom-left) since there are many possible paths induced by the independently sampled couplings " + }, + { + "bbox": [ + 54, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "x_0, x_1" + }, + { + "bbox": [ + 54, + 642, + 296, + 713 + ], + "type": "text", + "content": " through any " + }, + { + "bbox": [ + 54, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 54, + 642, + 296, + 713 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 54, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 642, + 296, + 713 + ], + "type": "text", + "content": " [29]. Curved paths require more numerical integration and network evaluation steps, which is not ideal." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 275, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 382 + ], + "type": "text", + "content": "Optimal Transport (OT) Coupling. To address this issue for unconditional generation, Tong et al. [35] and Pooladian et al. [29] concurrently proposed to use minibatch optimal transport to deterministically couple the sampled prior and a sampled batch of the data. This minimizes flow path lengths and straightens the flow. Concretely, given a minibatch of " + }, + { + "bbox": [ + 313, + 275, + 555, + 382 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 275, + 555, + 382 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 275, + 555, + 382 + ], + "type": "text", + "content": "-dimensional samples, an OT coupling seeks a " + }, + { + "bbox": [ + 313, + 275, + 555, + 382 + ], + "type": "inline_equation", + "content": "b \\times b" + }, + { + "bbox": [ + 313, + 275, + 555, + 382 + ], + "type": "text", + "content": " permutation matrix " + }, + { + "bbox": [ + 313, + 275, + 555, + 382 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 275, + 555, + 382 + ], + "type": "text", + "content": " such that the squared Euclidean distance is minimized, i.e.," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 391, + 388, + 555, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 388, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 391, + 388, + 555, + 407 + ], + "type": "interline_equation", + "content": "\\min _ {P} \\| X _ {0} - P X _ {1} \\| _ {2} ^ {2}. \\tag {5}", + "image_path": "56f8df6e231dd6efd1e617e732e22a23e9392312e44ed8aceb5d02318b09af81.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 412, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 484 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 412, + 555, + 484 + ], + "type": "inline_equation", + "content": "X_0, X_1 \\in \\mathbb{R}^{b \\times e}" + }, + { + "bbox": [ + 313, + 412, + 555, + 484 + ], + "type": "text", + "content": " are matrices containing a minibatch of samples from the prior and the data. Instead of an \"independent coupling\", in the unconditional setting we then optimize a variant of Equation (1) using the coupled tuple " + }, + { + "bbox": [ + 313, + 412, + 555, + 484 + ], + "type": "inline_equation", + "content": "(X_0, PX_1)" + }, + { + "bbox": [ + 313, + 412, + 555, + 484 + ], + "type": "text", + "content": ". Note that this algorithm corresponds to sampling tuples " + }, + { + "bbox": [ + 313, + 412, + 555, + 484 + ], + "type": "inline_equation", + "content": "(x_0, x_1)" + }, + { + "bbox": [ + 313, + 412, + 555, + 484 + ], + "type": "text", + "content": " from a joint distribution " + }, + { + "bbox": [ + 313, + 412, + 555, + 484 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ot}}(x_0, x_1)" + }, + { + "bbox": [ + 313, + 412, + 555, + 484 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "content": "Importantly, for unconditional generation, i.e., when using the joint distribution " + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ot}}(x_0,x_1)" + }, + { + "bbox": [ + 313, + 485, + 554, + 521 + ], + "type": "text", + "content": ", the marginals remain unchanged from independent coupling (see [35]):" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 353, + 526, + 511, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 526, + 511, + 550 + ], + "spans": [ + { + "bbox": [ + 353, + 526, + 511, + 550 + ], + "type": "interline_equation", + "content": "\\int \\pi_ {\\mathrm {o t}} (x _ {0}, x _ {1}) q (x _ {0}) \\mathrm {d} x _ {0} = q (x _ {1}), \\text {a n d}", + "image_path": "9dd9ccbf68ccb24209c57a6bf7e23960a03e7f445699977e2705f2baf5016a1b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 356, + 552, + 555, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 552, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 356, + 552, + 555, + 577 + ], + "type": "interline_equation", + "content": "\\int \\pi_ {\\mathrm {o t}} \\left(x _ {0}, x _ {1}\\right) q \\left(x _ {1}\\right) \\mathrm {d} x _ {1} = q \\left(x _ {0}\\right). \\tag {6}", + "image_path": "d4d6ced057b3ae157d9483d4d51280203f2820f6f271d22d6a3a81b258ae6d67.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "type": "text", + "content": "Hence, during test-time, we can sample from the prior " + }, + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "type": "text", + "content": " and integrate the flow " + }, + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "type": "inline_equation", + "content": "v_{\\theta}" + }, + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "type": "text", + "content": " to simulate data from the data distribution " + }, + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "type": "inline_equation", + "content": "q(x_{1})" + }, + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "text", + "content": "Unfortunately, as shown in the next section, this does not hold for conditional generation. Said differently, naively extending the unconditional joint distribution " + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ot}}(x_0,x_1)" + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "text", + "content": " to a conditional joint distribution " + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ot}}(x_0,x_1,c)" + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "text", + "content": " by using the coupled tuple " + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "inline_equation", + "content": "(X_0,PX_1,PC)" + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "inline_equation", + "content": "C\\in \\mathbb{R}^{b\\times h}" + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "text", + "content": " is a batch of " + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 618, + 555, + 712 + ], + "type": "text", + "content": "-dimensional conditions, does not maintain marginals and leads to a gap between training time and test time. We discuss this next." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15877" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 234, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 234, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 234, + 85 + ], + "type": "text", + "content": "3.2. OT Coupling Skews The Prior" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": "Intuition. Consider the one-dimensional example illustrated in Figure 2: the prior distribution " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": " follows a Gaussian distribution, and the target distribution " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "q(x_1)" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": " consists of a mixture of two Gaussians translated left and right, with the translation direction indicated by a binary condition " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": ". When training with a " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ot}}(x_0,x_1,c)" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": " that represents the coupled tuple " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "(X_0, PX_1,PC)" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": ", samples from the prior and the sampled conditions become correlated through the OT coupling – samples from the left half of the prior distribution are always associated with " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "c = 0" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": ", while those from the right half always correspond to " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "c = 1" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": ". As a result, the model overfits on data pairs " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "(x_0,c)" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": " where either " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "(x_0 < 0,c = 0)" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "(x_0 > 0,c = 1)" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": ". However, during testing, this skewed distribution is no longer accessible, and instead, " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": " are sampled independently. Consequently, the model encounters previously unseen data pairs, i.e., " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "(x_0 > 0,c = 0)" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "inline_equation", + "content": "(x_0 < 0,c = 1)" + }, + { + "bbox": [ + 55, + 91, + 296, + 317 + ], + "type": "text", + "content": " – leading to failure. A similar phenomenon can be observed in the two-dimensional example illustrated in Figure 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": "Skewed Prior. Formally, imagine that we consider the most general OT coupling joint distribution " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ot}}(x_0,x_1,c)" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": " at training time. At test time, the user wants to specify a condition " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "c_{1}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": " to obtain samples from the conditional distribution " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "q(x_{1}|c = c_{1})" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ". What is the prior distribution that we need to sample from such that we arrive at the correct conditional distribution " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "q(x_{1}|c = c_{1})" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": "? To compute this, imagine, we are given an arbitrary condition " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "c_{1}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ", i.e., we look at the induced distribution " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ot}}(x_0,x_1,c|c = c_1)" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": ". Marginalizing this induced distribution over " + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 319, + 296, + 439 + ], + "type": "text", + "content": " yields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 445, + 296, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 445, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 61, + 445, + 296, + 484 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\int \\pi_ {\\mathrm {o t}} \\left(x _ {0}, x _ {1}, c \\mid c = c _ {1}\\right) \\mathrm {d} x _ {1} = \\int \\pi_ {\\mathrm {o t}} \\left(x _ {0}, x _ {1} \\mid c = c _ {1}\\right) \\mathrm {d} x _ {1} \\\\ = q \\left(x _ {0} \\mid c = c _ {1}\\right), \\tag {7} \\\\ \\end{array}", + "image_path": "6de935beb742d15be76574d91d273e6d1ca8f922214e6b301e85e25b06cf81e1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": "which implies that the prior required to arrive at the conditional distribution " + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "inline_equation", + "content": "q(x_{1}|c = c_{1})" + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "inline_equation", + "content": "q(x_{0}|c = c_{1})" + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": " - a distribution that we cannot access at test time. In general, the prior that we need to sample from, given condition " + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": ", is skewed from " + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "inline_equation", + "content": "q(x_{0})" + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "inline_equation", + "content": "q(x_{0}|c)" + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": " which is inaccessible at test time. As a result, accurately capturing " + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "inline_equation", + "content": "q(x_{1}|c)" + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": " at test time becomes infeasible unless " + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "inline_equation", + "content": "q(x_{0})" + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "inline_equation", + "content": "q(c)" + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": " are independent, i.e., if " + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "inline_equation", + "content": "q(x_{0}|c) = q(x_{0})" + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": ". However, this condition is generally not satisfied by the most general OT coupling – for instance, in Figure 2, clearly " + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "inline_equation", + "content": "q(x_{0}|c) \\neq q(x_{0})" + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": ". To address this issue, we propose a method " + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "inline_equation", + "content": "C^2\\mathrm{OT}" + }, + { + "bbox": [ + 55, + 491, + 296, + 634 + ], + "type": "text", + "content": " to unkew the prior. We discuss this in the next section." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 649, + 262, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 262, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 262, + 663 + ], + "type": "text", + "content": "3.3. Conditional Optimal Transport FM" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 670, + 137, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 137, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 137, + 681 + ], + "type": "text", + "content": "3.3.1. Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "We propose conditional optimal transport FM (C²OT) to unkew the prior " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "q(x_0|c)" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "q(x_0|c) = q(x_0)" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": ". This" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": "ensures that at test time, we can sample from the full prior " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ", irrespective of the condition " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ". Importantly, at the same time, we also aim to preserve the straight flow paths provided by OT. Conceptually, we construct a prior distribution independently for each condition, i.e., " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "q(x_0|c_1) = q(x_0|c_2) = q(x_0), \\forall c_1, c_2" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ". This can be achieved by sampling from the prior and computing OT independently for each condition, as shown in Figure 2 (right)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 168, + 515, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 168, + 515, + 179 + ], + "spans": [ + { + "bbox": [ + 326, + 168, + 515, + 179 + ], + "type": "text", + "content": "Formally, we construct the joint distribution as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 189, + 555, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 189, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 348, + 189, + 555, + 201 + ], + "type": "interline_equation", + "content": "\\pi_ {\\mathrm {c} ^ {2} \\mathrm {o t}} (x _ {0}, x _ {1}, c) := q (x _ {0}) q (c) \\pi_ {\\mathrm {o t}} (x _ {1} | x _ {0}, c). \\tag {8}", + "image_path": "6c0f2563bb9ee49d9c18b9fae392328dd8c46e9ef92152f6cbdd5a281a32d353.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "text", + "content": "Here, the prior " + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "text", + "content": " and the condition " + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "inline_equation", + "content": "q(c)" + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "text", + "content": " are sampled independently, while the data " + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "text", + "content": " is conditioned on " + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "text", + "content": " (via OT) and " + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "text", + "content": " (via the dataset construction). Different from independent coupling in Equation (2), " + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "text", + "content": " are associated through OT and therefore lead to straighter flow paths. Also different from the most general OT coupling " + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ot}}(x_0,x_1,c)" + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "text", + "content": ", we explicitly enforce the independence between " + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 209, + 555, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "text", + "content": "To see that this joint distribution provides an unskewed prior, imagine we are given an arbitrary input condition " + }, + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "c_{1}" + }, + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "text", + "content": ", marginalizing over " + }, + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 313, + 293, + 555, + 329 + ], + "type": "text", + "content": " yields" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 343, + 337, + 555, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 337, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 343, + 337, + 555, + 387 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\int q (x _ {0}) q (c | c = c _ {1}) \\pi_ {\\mathrm {o t}} \\left(x _ {1} \\mid x _ {0}, c = c _ {1}\\right) \\mathrm {d} x _ {1} \\\\ = \\int q \\left(x _ {0}\\right) \\pi_ {\\mathrm {o t}} \\left(x _ {1} \\mid x _ {0}, c = c _ {1}\\right) \\mathrm {d} x _ {1} = q \\left(x _ {0}\\right). \\tag {9} \\\\ \\end{array}", + "image_path": "2fa8c463c4a10c7584b0e63b96f1be6494cf09f5260dfcd5a324a163a3e9b6ad.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "text", + "content": "This implies that, at test time, we can sample from the entire prior " + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "inline_equation", + "content": "q(x_0)" + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "text", + "content": ", regardless of the condition " + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "text", + "content": ", to arrive at the desired data distribution " + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "inline_equation", + "content": "q(x_1|c)" + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "text", + "content": ". During training, in practice, we sample from " + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{c}^2\\mathrm{ot}}" + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "text", + "content": " by modifying the optimal transport cost function in Equation (5). This process is exact for discrete conditions (Section 3.3.2) and approximate for continuous conditions (Section 3.3.3). Specifically, given a minibatch of " + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "text", + "content": "-dimensional samples and a minibatch of " + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "text", + "content": "-dimensional conditions " + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "inline_equation", + "content": "C \\in \\mathbb{R}^{b \\times h}" + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^2\\mathrm{OT}" + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "text", + "content": " seeks a " + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "inline_equation", + "content": "b \\times b" + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "text", + "content": " permutation matrix " + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 393, + 554, + 514 + ], + "type": "text", + "content": " that minimizes the following cost function:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 521, + 555, + 553 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 521, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 351, + 521, + 555, + 553 + ], + "type": "interline_equation", + "content": "\\min _ {P} \\| X _ {0} - P X _ {1} \\| _ {2} ^ {2} + \\sum_ {i = 1} ^ {b} f \\left(C _ {i}, [ P C ] _ {i}\\right), \\tag {10}", + "image_path": "dac4ec6099885811d1b668f48446d9c774cbc7a4720ffc77d7a12a45cfec460c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 560, + 554, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 554, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 560, + 554, + 632 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 560, + 554, + 632 + ], + "type": "text", + "content": " is a symmetric and non-negative cost function, satisfying the property " + }, + { + "bbox": [ + 313, + 560, + 554, + 632 + ], + "type": "inline_equation", + "content": "f(c, c) = 0 \\forall c" + }, + { + "bbox": [ + 313, + 560, + 554, + 632 + ], + "type": "text", + "content": ". The key differences between independent coupling, OT coupling, and " + }, + { + "bbox": [ + 313, + 560, + 554, + 632 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{OT}" + }, + { + "bbox": [ + 313, + 560, + 554, + 632 + ], + "type": "text", + "content": " coupling are highlighted in Algorithms 1 to 3. In the following two sections, we discuss our choice of " + }, + { + "bbox": [ + 313, + 560, + 554, + 632 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 560, + 554, + 632 + ], + "type": "text", + "content": " for both discrete and continuous conditions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 647, + 426, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 426, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 426, + 658 + ], + "type": "text", + "content": "3.3.2. Discrete Conditions" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "For discrete conditions, we assume the conditions correspond to class labels from a finite, discrete set, i.e., " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "c \\in \\{0, 1, \\ldots, k - 1\\}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ". For example, in Figure 2, the labels 0 and 1 represent the left and right groups in the target distribution," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15878" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 57, + 66, + 214, + 150 + ], + "blocks": [ + { + "bbox": [ + 57, + 66, + 214, + 150 + ], + "lines": [ + { + "bbox": [ + 57, + 66, + 214, + 150 + ], + "spans": [ + { + "bbox": [ + 57, + 66, + 214, + 150 + ], + "type": "text", + "content": "Algorithm 1 Independent Coupling \n1: function INDEPENDENT(X1, C) \n2: X0 ~ N ▷ Sample from prior \n3: \n4: \n5: return X0, X1, C" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_body" + } + ], + "index": 0, + "sub_type": "algorithm" + }, + { + "type": "code", + "bbox": [ + 226, + 66, + 383, + 150 + ], + "blocks": [ + { + "bbox": [ + 226, + 66, + 383, + 150 + ], + "lines": [ + { + "bbox": [ + 226, + 66, + 383, + 150 + ], + "spans": [ + { + "bbox": [ + 226, + 66, + 383, + 150 + ], + "type": "text", + "content": "Algorithm 2 OT Coupling \n1: function OT(X1, C) \n2: X0 ~ N ▷ Sample from prior \n3: Cost ← pairwise_dist(X0, X1) \n4: i, j ← Hungarian_MATCHing(Cost) \n5: return X0[i], X1[j], C[j]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "type": "code", + "bbox": [ + 394, + 66, + 551, + 152 + ], + "blocks": [ + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "lines": [ + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "spans": [ + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "type": "text", + "content": "Algorithm 3 " + }, + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{OT}" + }, + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "type": "text", + "content": " Coupling (Ours) \n1: function " + }, + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{OT}(X_1,C)" + }, + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "type": "inline_equation", + "content": "X_0\\sim \\mathcal{N}" + }, + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "type": "text", + "content": " ▷ Sample from prior \n3: Cost " + }, + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "type": "text", + "content": " pairwise_dist(Xo,X1)+ pairwise_dist(C,C) \n4: " + }, + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "type": "inline_equation", + "content": "i,j\\gets" + }, + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "type": "text", + "content": " Hungarian_MATCHing(Cost) \n5: return " + }, + { + "bbox": [ + 394, + 66, + 551, + 152 + ], + "type": "inline_equation", + "content": "X_0[i],X_1[j],C[j]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_body" + } + ], + "index": 2, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 160, + 295, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 160, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 295, + 196 + ], + "type": "text", + "content": "respectively. We define " + }, + { + "bbox": [ + 55, + 160, + 295, + 196 + ], + "type": "inline_equation", + "content": "f = f_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 160, + 295, + 196 + ], + "type": "text", + "content": " such that no transport is allowed to modify the conditions (i.e., " + }, + { + "bbox": [ + 55, + 160, + 295, + 196 + ], + "type": "inline_equation", + "content": "PC = C" + }, + { + "bbox": [ + 55, + 160, + 295, + 196 + ], + "type": "text", + "content": "). This is achieved via" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 112, + 204, + 296, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 204, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 112, + 204, + 296, + 236 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {d}} \\left(c _ {1}, c _ {2}\\right) = \\left\\{ \\begin{array}{l} \\infty , c _ {1} \\neq c _ {2}, \\\\ 0, \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {11}", + "image_path": "4c7d17d6910240eafbe2b0f1586ab5696c75da89cc4498f95f2fa214a2b675df.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 245, + 296, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 245, + 296, + 355 + ], + "spans": [ + { + "bbox": [ + 54, + 245, + 296, + 355 + ], + "type": "text", + "content": "As a result, the unordered set of prior samples for any given condition " + }, + { + "bbox": [ + 54, + 245, + 296, + 355 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 245, + 296, + 355 + ], + "type": "text", + "content": " remains unchanged: " + }, + { + "bbox": [ + 54, + 245, + 296, + 355 + ], + "type": "inline_equation", + "content": "\\{X_{0i}|c_i = c\\} \\stackrel{\\mathrm{iid}}{\\sim} q(x_0)" + }, + { + "bbox": [ + 54, + 245, + 296, + 355 + ], + "type": "text", + "content": ". This formulation samples exactly from our ideal (Equation (8)), as OT is computed independently within each class, and the data sample from each class is exposed to the full prior without skew. This effect is illustrated in Figure 1 (middle) and Figure 2 (right). As expected, the results differ from FM's curved flow paths and OT's inability to properly capture the target distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "text", + "content": "It is important to note that enforcing " + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "inline_equation", + "content": "PC = C" + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "text", + "content": " does not necessarily imply " + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "inline_equation", + "content": "P = I" + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "text", + "content": " since we can still swap row " + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "text", + "content": " as long as " + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "inline_equation", + "content": "c_{i} = c_{j}" + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "text", + "content": ". Setting " + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "inline_equation", + "content": "P = I" + }, + { + "bbox": [ + 55, + 357, + 296, + 416 + ], + "type": "text", + "content": " would correspond to a degeneration back to independent coupling in Equation (2) and forfeit the benefits of straightened flow provided by OT." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "type": "text", + "content": "Unfortunately, using " + }, + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "type": "text", + "content": " with continuous conditions leads to exactly this degeneration, since, in general, no " + }, + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "type": "inline_equation", + "content": "c_{i} = c_{j}" + }, + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "type": "text", + "content": " unless " + }, + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "type": "inline_equation", + "content": "i = j" + }, + { + "bbox": [ + 55, + 417, + 296, + 464 + ], + "type": "text", + "content": ". Hence, in the next section, we devise a relaxed cost for handling continuous conditions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 483, + 182, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 182, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 182, + 494 + ], + "type": "text", + "content": "3.3.3. Continuous Conditions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "text", + "content": "For continuous conditions, the condition is typically a feature embedding, e.g., computed from a text prompt. Directly applying " + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "text", + "content": " in Equation (11) leads to a degeneration back to independent coupling as discussed earlier. To avoid this, we propose a relaxed penalty function " + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 502, + 296, + 574 + ], + "type": "text", + "content": " based on the cosine distance:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 582, + 296, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 582, + 296, + 609 + ], + "spans": [ + { + "bbox": [ + 61, + 582, + 296, + 609 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {c}} \\left(c _ {1}, c _ {2}\\right) = w \\left(1 - \\frac {c _ {1} \\cdot c _ {2}}{\\| c _ {1} \\| \\| c _ {2} \\|}\\right) = w \\cdot \\operatorname {c d i s t} \\left(c _ {1}, c _ {2}\\right). \\tag {12}", + "image_path": "e85fba4b3bdf2374a327463f72042256ae0542d73c957d83338273afb3031874.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "w > 0" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " is a scaling hyperparameter. Note, when " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "w = 0" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ", this formulation degenerates to regular OT; when " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "w \\to \\infty" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": ", this formulation approaches Equation (11), i.e., independent coupling if no two conditions are equal. The right section of Figure 1 illustrates an example where the " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "-coordinate of the target data point serves as the condition. Notably, " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "C^2\\mathrm{OT}" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " preserves straight flow paths without exhibiting OT's apparent degradation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "type": "text", + "content": "Finding " + }, + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "type": "text", + "content": ". The optimal " + }, + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "type": "text", + "content": " is highly dependent on the data distribution and the magnitude of features. To alleviate the need for tuning the hyperparameter " + }, + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "type": "text", + "content": ", we propose to find " + }, + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "type": "text", + "content": " adaptively for each minibatch. For this, we develop a ratio " + }, + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "type": "inline_equation", + "content": "r(w)" + }, + { + "bbox": [ + 313, + 160, + 555, + 233 + ], + "type": "text", + "content": " which measures the proportion of samples that are considered \"potential optimal transport candidates\", i.e.," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 323, + 242, + 555, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 242, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 323, + 242, + 555, + 289 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} r (w) = \\frac {1}{B ^ {2}} \\sum_ {i, j} \\mathbb {1} \\left(\\left\\| X _ {0 i} - X _ {1 j} \\right\\| _ {2} ^ {2} + w \\cdot \\operatorname {c d i s t} \\left(c _ {i}, c _ {j}\\right) \\right. \\\\ \\leq \\left\\| X _ {0 i} - X _ {1 i} \\right\\| _ {2} ^ {2}), \\tag {13} \\\\ \\end{array}", + "image_path": "12f3790c8ce6a5a774e6d536102e7c9a74087274f88c2351d4665b781a850bf2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": " is an indicator function that evaluates to one if the condition is satisfied and to zero otherwise. Then, we introduce a hyperparameter " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{tar}}" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": " as the target ratio and find " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "r(w) \\approx r_{\\mathrm{tar}}" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": ". Note that setting " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{tar}}" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": " is invariant to the scaling of distances between " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "x_1" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": ". Namely, if " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "\\|X_{0i} - X_{1j}\\|_2^2" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": " is scaled by " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": " would also be scaled by " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": " to preserve the same " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{tar}}" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": ". This invariance is particularly useful when transitioning to a new dataset, such as changing image resolutions in image generation tasks. In our implementation, we search for " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": " in two steps: first with exponential search to establish the upper/ lower bounds, then with binary search to locate a precise value. The final " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": " is used as the initial value in the subsequent minibatch. In practice, we observe " + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": " differs very little between minibatches (" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "inline_equation", + "content": "< 1\\%" + }, + { + "bbox": [ + 313, + 295, + 556, + 499 + ], + "type": "text", + "content": ") and the search converges quickly within 10 iterations with a negligible overhead. Next, we introduce oversampling, a simple technique to obtain more accurate OT couplings." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 513, + 455, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 455, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 455, + 525 + ], + "type": "text", + "content": "3.3.4. Oversampling OT Batches" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "text", + "content": "OT vs. Deep Net Batch Sizes. Typically, the \"OT batch size\" " + }, + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "text", + "content": " used to compute optimal transport is set equal to the \"deep net batch size\" " + }, + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "inline_equation", + "content": "b_{n}" + }, + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "text", + "content": " used to compute forward/backward passes of the deep net [29, 35]. While the OT batch size controls the dynamics of prior-to-data assignments, the deep net batch size affects gradient variances, training efficiency, and memory usage. There is no reason why setting them equal should be optimal. Particularly, with our proposed " + }, + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "inline_equation", + "content": "C^2\\mathrm{OT}" + }, + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "text", + "content": ", the effective OT batch size is reduced since we restrict transport between samples with divergent conditions, which calls for an increased OT batch size " + }, + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "text", + "content": ". At the same time, keeping the deep net batch size " + }, + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "inline_equation", + "content": "b_{n}" + }, + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "text", + "content": " unchanged prevents unwarranted side effects in other aspects of training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Reduced Effective OT Batch Size. In the case of discrete conditions, OT is effectively performed within" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15879" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 106, + 293, + 272 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": "Table 1. Results on the 8gaussians " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": " moons dataset with different training coupling methods and different conditioning. Note that our method " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "C^2\\mathrm{OT}" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": " does not apply in the unconditional setting." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 106, + 293, + 272 + ], + "lines": [ + { + "bbox": [ + 56, + 106, + 293, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 293, + 272 + ], + "type": "table", + "html": "
MethodEuler-1 (W2↓)Adaptive (W2↓)NFE ↓
Unconditional
FM6.232±0.1860.120±0.04593.22±1.81
OT0.072±0.0250.050±0.04036.08±2.57
Binary conditions
FM2.573±0.1120.059±0.02991.10±2.77
OT0.483±0.0590.462±0.02532.89±0.95
C2OT (ours)0.048±0.0100.018±0.00658.88±1.48
Continuous conditions
FM0.732±0.0520.028±0.01093.86±1.76
OT8.276±6.5102.143±1.99342.07±2.59
C2OT (ours)0.077±0.0240.013±0.00389.87±1.53
", + "image_path": "36e58a6faa0c5d2dead8cc694df8a17251a414385c639f00d23af41b9c046cf7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\{X_{0i}, X_{1i} | c_i = c\\}" + }, + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "text", + "content": " for each condition " + }, + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "text", + "content": ". With " + }, + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "text", + "content": " classes, the expected effective OT batch size is reduced to " + }, + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\mathbb{E}(|\\{i | c_i = c\\}|) = b / k" + }, + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "text", + "content": ". Empirically, we find that using a much larger OT batch size " + }, + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "text", + "content": " than the network batch size " + }, + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "inline_equation", + "content": "b_n" + }, + { + "bbox": [ + 55, + 280, + 296, + 342 + ], + "type": "text", + "content": " is helpful. This observation motivates oversampling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 344, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 380 + ], + "type": "text", + "content": "Efficiency. A natural concern with increasing " + }, + { + "bbox": [ + 55, + 344, + 296, + 380 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 344, + 296, + 380 + ], + "type": "text", + "content": " is the computational overhead, since optimal transport with Hungarian matching [20] has a time complexity of " + }, + { + "bbox": [ + 55, + 344, + 296, + 380 + ], + "type": "inline_equation", + "content": "O(b^{3})" + }, + { + "bbox": [ + 55, + 344, + 296, + 380 + ], + "type": "text", + "content": ". However," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 380, + 296, + 452 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "text", + "content": "- Since each OT batch can be used across " + }, + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "inline_equation", + "content": "(b / b_{n})" + }, + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "text", + "content": " forward/backward passes, the amortized cost per minibatch reduces to " + }, + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "inline_equation", + "content": "O(b^{2}b_{n})" + }, + { + "bbox": [ + 55, + 380, + 296, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 416, + 295, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 295, + 452 + ], + "type": "text", + "content": "- We offload OT computation from the main process to the data workers, enabling data preparation on the CPU in parallel with forward/backward passes on the GPU." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 452, + 296, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 452, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 296, + 500 + ], + "type": "text", + "content": "With a modest choice of 8 data workers per GPU, we observe no wall-clock overhead for OT batch sizes up to 6,400 when training on CIFAR-10 and ImageNet, as data preparation is completed faster than the GPU can process batches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 518, + 137, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 137, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 137, + 533 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 544, + 297, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 297, + 628 + ], + "type": "text", + "content": "We experimentally verify our proposed method " + }, + { + "bbox": [ + 55, + 544, + 297, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{OT}" + }, + { + "bbox": [ + 55, + 544, + 297, + 628 + ], + "type": "text", + "content": " in two sections: first on the two-dimensional synthetic dataset 8gaussians " + }, + { + "bbox": [ + 55, + 544, + 297, + 628 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 544, + 297, + 628 + ], + "type": "text", + "content": " moons, and then on high-dimensional image data, including CIFAR-10 [19] and ImageNet-1K [7]. For ImageNet-1K, we conduct experiments on both " + }, + { + "bbox": [ + 55, + 544, + 297, + 628 + ], + "type": "inline_equation", + "content": "32\\times 32" + }, + { + "bbox": [ + 55, + 544, + 297, + 628 + ], + "type": "text", + "content": " images in image space and " + }, + { + "bbox": [ + 55, + 544, + 297, + 628 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 55, + 544, + 297, + 628 + ], + "type": "text", + "content": " images in latent space. Implementation details are provided in the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 645, + 197, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 197, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 197, + 658 + ], + "type": "text", + "content": "4.1. Two-Dimensional Data" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 666, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 297, + 715 + ], + "type": "text", + "content": "The two-dimensional 8gaussians " + }, + { + "bbox": [ + 55, + 666, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 666, + 297, + 715 + ], + "type": "text", + "content": " moons dataset maps a mixture of eight Gaussian distributions to two interleaving half-circles, as visualized in Figure 1. We consider three conditioning scenarios: (a) unconditional; (b) binary" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "class conditions, where each class corresponds to one of the half-circles; and (c) continuous conditions, where the " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "-coordinate of the desired target point is given as the condition. For (b), we apply our setup for discrete conditions (Section 3.3.2); for (c), we apply our setup for continuous conditions (Section 3.3.3) with " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{tar}} = 0.01" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 146, + 556, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 556, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 556, + 266 + ], + "type": "text", + "content": "Metrics. We report the 2-Wasserstein distance " + }, + { + "bbox": [ + 313, + 146, + 556, + 266 + ], + "type": "inline_equation", + "content": "(W_2^2)" + }, + { + "bbox": [ + 313, + 146, + 556, + 266 + ], + "type": "text", + "content": " (following [35]) between 10K generated data points and 10K samples from the ground-truth moons distribution. We experiment with single-step Euler's method (Euler-1) and the adaptive Dormand-Prince method [8] (dopri5, referred to as adaptive) for numerical integration. Additionally, we report the number of function evaluations (NFE) in the adaptive method, which is the number of forward passes through the neural network. We average the results over ten training runs with different random seeds and report mean±std." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 267, + 556, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 556, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 556, + 376 + ], + "type": "text", + "content": "Results. The results are summarized in Table 1 and visualized in Figure 1. In the unconditional setting, OT produces straighter flows, leading to better distribution matching with fewer NFEs, and our method does not apply. However, in the conditional setting, OT degrades and performs worse than regular FM in the adaptive setting due to the aforementioned train-test gap in OT coupling. Our method, " + }, + { + "bbox": [ + 313, + 267, + 556, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{OT}" + }, + { + "bbox": [ + 313, + 267, + 556, + 376 + ], + "type": "text", + "content": ", effectively models the target distribution with both Euler-1 and the adaptive method, improving upon the baselines." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 391, + 495, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 495, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 495, + 404 + ], + "type": "text", + "content": "4.2. High-Dimensional Image Data" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 411, + 555, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 494 + ], + "type": "text", + "content": "We conduct experiments on CIFAR-10 [19] (image space, class-conditioned), ImageNet- " + }, + { + "bbox": [ + 313, + 411, + 555, + 494 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 411, + 555, + 494 + ], + "type": "text", + "content": " [7] (image space, caption-conditioned), and ImageNet- " + }, + { + "bbox": [ + 313, + 411, + 555, + 494 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 411, + 555, + 494 + ], + "type": "text", + "content": " [7] (latent space, caption-conditioned). For ImageNet, we use the captions provided by [21] and encode text features using a CLIP [30]-like DFN [10] text encoder as input conditions. Our results are summarized in Table 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 496, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 556, + 628 + ], + "type": "text", + "content": "Metrics. We report Fréchet inception distance (FID) [13] and the number of function evaluations (NFE) in the adaptive solver for all three datasets. To assess how well the generations adhere to the input conditions, we additionally compute condition adherence metrics. For CIFAR-10, we run a pretrained classifier [5] on the generated images to obtain logits and report the average cross entropy (CE) with the input class label. For ImageNet, we compute CLIP features [30] using SigLIP-2 [36] from the generated images and the input captions respectively, and report the average cosine similarities (CLIP)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "CIFAR-10. We base our UNet architecture on Tong et al. [35], with details provided in the appendix. The CIFAR-10 training set contains 50,000 " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " color images across 10 object classes. During testing, we generate 50,000 images evenly distributed among classes and assess FID against the training set following [35]. Each model is trained five times with different random seeds and the results are reported" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15880" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 118, + 555, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 2. Results of different training algorithms on image generation tasks. We bold the best-performing entry and underline the second-best. Our proposed method, " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "C^2 OT" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": ", achieves the best overall performance – better than FM with few sampling steps and better than OT with more sampling steps. In cases where " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "C^2 OT" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " performs second, our performance is often comparable to the best entry. Additionally, " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "C^2 OT" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " has the most stable performance, with the smallest deviations across runs in most entries. Variations in CLIP scores are minimal." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 118, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 56, + 118, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 555, + 407 + ], + "type": "table", + "html": "
CIFAR-10 Class-Conditioned Generation
MethodEuler-2Euler-5Euler-10Euler-25Euler-50Euler-100AdaptiveNFE ↓
FID → FM105.481±1.61921.968±0.40010.479±0.2375.640±0.1174.128±0.0693.429±0.0542.922±0.039124.0±2.5
OT64.417±0.43418.194±0.40910.778±0.2706.821±0.1865.368±0.1134.671±0.0664.144±0.018125.7±0.4
C²OT (ours)64.694±0.37917.565±0.2029.762±0.1365.644±0.0814.152±0.0613.447±0.0492.904±0.035127.7±0.7
CE → FM3.343±0.0220.546±0.0200.323±0.0070.272±0.0040.267±0.0020.269±0.0030.276±0.004124.0±2.5
OT2.529±0.0100.624±0.0120.426±0.0050.369±0.0040.363±0.0040.364±0.0050.372±0.005125.7±0.4
C²OT (ours)2.264±0.0220.475±0.0090.322±0.0040.276±0.0050.271±0.0030.272±0.0030.279±0.004127.7±0.7
ImageNet 32×32 Caption-Conditioned Generation
FID → FM113.250±0.79323.015±0.12311.141±0.0517.165±0.0876.142±0.0855.673±0.0745.358±0.059146.9±3.5
OT81.317±0.36920.763±0.15411.360±0.0897.854±0.0776.899±0.0796.469±0.0716.220±0.065137.7±0.2
C²OT (ours)102.380±0.27921.965±0.03510.897±0.0337.069±0.0276.084±0.0165.638±0.0185.350±0.017134.2±1.4
CLIP → FM0.067±0.0000.076±0.0000.074±0.0000.071±0.0000.071±0.0000.070±0.0000.069±0.000146.9±3.5
OT0.067±0.0000.071±0.0000.070±0.0000.068±0.0000.068±0.0000.067±0.0000.067±0.000137.7±0.2
C²OT (ours)0.068±0.0000.075±0.0000.073±0.0000.071±0.0000.070±0.0000.070±0.0000.069±0.000134.2±1.4
ImageNet 256×256 Caption-Conditioned Generation in Latent Space
FID → FM202.519±0.60631.266±0.3559.996±0.1065.189±0.0823.886±0.0663.484±0.0553.498±0.054131.1±2.3
OT193.376±1.66542.780±2.13215.638±1.5107.514±1.0345.693±0.9985.360±1.0165.726±1.074125.6±6.5
C²OT (ours)200.811±0.38530.715±0.22110.000±0.0725.069±0.0373.722±0.0373.320±0.0323.369±0.034128.2±2.6
CLIP → FM0.028±0.0000.120±0.0000.134±0.0000.137±0.0000.138±0.0000.139±0.0000.140±0.000131.2±2.3
OT0.031±0.0000.107±0.0030.120±0.0030.123±0.0030.124±0.0030.124±0.0040.125±0.004125.6±6.5
C²OT (ours)0.029±0.0000.120±0.0000.133±0.0000.137±0.0000.138±0.0000.138±0.0000.139±0.000128.2±2.6
", + "image_path": "85472c3671ea35873ca2343477c6da6ab0131c3f22520e50b17ca48b1aa4cc8d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 417, + 294, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 417, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 54, + 417, + 294, + 511 + ], + "type": "text", + "content": "as mean±std. We use an OT batch size " + }, + { + "bbox": [ + 54, + 417, + 294, + 511 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 54, + 417, + 294, + 511 + ], + "type": "text", + "content": " of 640 and a ratio " + }, + { + "bbox": [ + 54, + 417, + 294, + 511 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{tar}}" + }, + { + "bbox": [ + 54, + 417, + 294, + 511 + ], + "type": "text", + "content": " of 0.01. Compared to FM, our method achieves better performance (FID and CE) in few-steps settings and comparable performance in many-steps settings. Compared to OT, our method significantly outperforms in the many-step setting. Although OT achieves slightly better FID in Euler-2, it performs worse in CE, indicating that it fails to follow the condition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "text", + "content": "ImageNet " + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "text", + "content": ". We base our UNet architecture on Pooladian et al. [29], and list details in the appendix. We train and evaluate on the face-blurred [37] ImageNet, following [29]. In contrast to prior works that assess FID against the training set, we assess FID against the validation set (49,997 images), as the fine-grained nature of captions may lead to overfitting on the training set. Each model is trained three times with different random seeds and the results are reported as mean " + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "text", + "content": " std. Compared to CIFAR-10, there are much more variations in this dataset (e.g., 1000 classes vs. 10 classes in CIFAR-10). Thus, we adopt a larger OT batch size " + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "text", + "content": " of 6400 while keeping the same target ratio " + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{tar}}" + }, + { + "bbox": [ + 55, + 521, + 297, + 713 + ], + "type": "text", + "content": " 0.01 as before. Similar to our findings in CIFAR-10, our method performs better or is comparable to FM; in few-steps settings, OT achieves better FID but worse condition following (CLIP). Our method strikes the best overall balance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "text", + "content": "ImageNet " + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "text", + "content": " in Latent Space. Here, we explore latent flow matching models [31]. During training, the flow matching model learns to generate data in the latent space, which is encoded from images. At test-time, the generated latents are decoded into images using a pretrained decoder. We base our experiment on a recent transformer-based model, LightningDiT [39], demonstrating that our method is effective across different network architectures and data spaces. We use the officially provided \"64 epochs\" configuration for all experiments (due to computational constraints), which is not directly comparable to methods trained with 800 epochs. We follow the same evaluation protocol as ImageNet- " + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "text", + "content": ", and adopt the same OT batch size " + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "text", + "content": " of 6400 and the target ratio " + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{tar}}" + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "text", + "content": " of 0.01. Each model is trained five times with different random seeds, and the results are reported as mean " + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "text", + "content": " std. Again, " + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{OT}" + }, + { + "bbox": [ + 313, + 417, + 556, + 621 + ], + "type": "text", + "content": " has the best overall performance compared to both FM and OT. We visualize the generations in Figure 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Limitations in High-Dimensional Data We observed that the benefits of " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{OT}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " are less pronounced, though still visible, in higher-dimensional image data compared to low-dimensional data. We hypothesize this is because image data is more entangled, making the prior skew induced by OT less detrimental. We also observe statistically straighter paths in " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{OT}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " compared to FM, with details given in the appendix." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15881" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 296, + 193 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 296, + 193 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 296, + 193 + ], + "type": "image", + "image_path": "1bdbf61a6d1473856e6f3ed0ddb35d54c0dcb10afd6a447fdf531862d4a28611.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 297, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 297, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 297, + 240 + ], + "type": "text", + "content": "Figure 3. Visual comparisons of " + }, + { + "bbox": [ + 55, + 194, + 297, + 240 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 194, + 297, + 240 + ], + "type": "text", + "content": " images generated by the baselines and our approach with different amounts of sampling steps. Our approach converges faster and produces cleaner outputs. The input caption is \"a small dog sitting on a carpet\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 165, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 165, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 165, + 261 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 189, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 189, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 189, + 283 + ], + "type": "text", + "content": "4.3.1. Varying OT Batch Size " + }, + { + "bbox": [ + 55, + 270, + 189, + 283 + ], + "type": "inline_equation", + "content": "b" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 289, + 298, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 298, + 578 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 298, + 578 + ], + "type": "text", + "content": "Here, we analyze the effect of varying the OT batch size " + }, + { + "bbox": [ + 54, + 289, + 298, + 578 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 54, + 289, + 298, + 578 + ], + "type": "text", + "content": ". We conduct these experiments on the CIFAR-10 dataset, training each model three times with different random seeds for every choice of " + }, + { + "bbox": [ + 54, + 289, + 298, + 578 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 54, + 289, + 298, + 578 + ], + "type": "text", + "content": ". We plot the results in Figure 4. Increasing the OT batch size clearly improves FID in the few-sampling-steps setting (Euler-2). This is intuitive, because a larger OT batch size enables the coupling to approach the true OT results, i.e., dataset-level optimal transport. In turn, this leads to straighter paths. However, in the adaptive setting, increasing the OT batch size slightly worsens FID, though the effect is much less significant (the range of the blue y-axis is much smaller than the range of the red y-axis). We think this occurs because a more accurate OT coupling reduces variations in the training data (similar to having less data augmentations) which harms performance as the benefit of straighter flows diminishes in the adaptive setting. This observation aligns with the findings of Pooladian et al. [29]: computing OT coupling across GPUs (i.e., having a larger effective OT batch size) slightly harms results. In contrast to Pooladian et al. [29], though, we seek to counteract the reduced effective OT batch size induced by conditional optimal transport (as discussed in Section 3.3.4), and we do find reasonably large (640 for CIFAR, 6400 for ImageNet) OT batch sizes useful while not adding wall-clock overhead." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 191, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 191, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 191, + 607 + ], + "type": "text", + "content": "4.3.2. Varying Target Ratio " + }, + { + "bbox": [ + 55, + 594, + 191, + 607 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{tar}}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "text", + "content": "Here, we analyze the effects of varying the target ratio " + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{tar}}" + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "text", + "content": ", as defined in Equation (13). We conduct these experiments on the ImageNet- " + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "text", + "content": " dataset, training each model three times with different random seeds for every choice of " + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{tar}}" + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "text", + "content": ". We plot the results in Figure 5. Recall that when " + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "inline_equation", + "content": "r \\to 0" + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "inline_equation", + "content": "w \\to \\infty" + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "text", + "content": " and the method approaches FM; when " + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "inline_equation", + "content": "r \\to 0.5" + }, + { + "bbox": [ + 55, + 613, + 297, + 685 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 220 + ], + "type": "image", + "image_path": "97a9d202d6d745a3acf16bec93aa699b9e76f37042bdf144a38e357b02329111.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 224, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 281 + ], + "type": "text", + "content": "Figure 4. Changes in FID with respect to varying OT batch sizes " + }, + { + "bbox": [ + 313, + 224, + 555, + 281 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 224, + 555, + 281 + ], + "type": "text", + "content": ". We plot mean±std over three runs and represent std with a shaded region. OT (adaptive) and FM (Euler-2) perform worse than all other results in this plot, hence results are not shown (full plot in the appendix)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 293, + 553, + 441 + ], + "blocks": [ + { + "bbox": [ + 316, + 293, + 553, + 441 + ], + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 441 + ], + "type": "image", + "image_path": "021a2927ace7652996892924796d7b1493956ce09caf1ec68644a6c328feefd0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 445, + 556, + 491 + ], + "lines": [ + { + "bbox": [ + 313, + 445, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 556, + 491 + ], + "type": "text", + "content": "Figure 5. Changes in FID with respect to varying target ratio " + }, + { + "bbox": [ + 313, + 445, + 556, + 491 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{tar}}" + }, + { + "bbox": [ + 313, + 445, + 556, + 491 + ], + "type": "text", + "content": ". We plot mean±std over three runs and represent std with a shaded region. OT (adaptive) performs worse than all other results in this plot, hence results are not shown (full plot in the appendix)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 502, + 556, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 556, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 556, + 562 + ], + "type": "inline_equation", + "content": "w \\to 0" + }, + { + "bbox": [ + 313, + 502, + 556, + 562 + ], + "type": "text", + "content": " and the method approaches OT (Section 3.3.3). Our findings are similar to those in Section 4.3.1: as we approach true OT, whether by increasing OT batch size or increasing " + }, + { + "bbox": [ + 313, + 502, + 556, + 562 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{tar}}" + }, + { + "bbox": [ + 313, + 502, + 556, + 562 + ], + "type": "text", + "content": ", FID improves in the few-step setting and worsens in the adaptive setting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 574, + 388, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 388, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 388, + 588 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 600, + 556, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 556, + 695 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 556, + 695 + ], + "type": "text", + "content": "We first investigate and formalize the struggle of minibatch optimal transport in conditional generation. Based on the findings, we propose " + }, + { + "bbox": [ + 313, + 600, + 556, + 695 + ], + "type": "inline_equation", + "content": "\\mathrm{C}^2\\mathrm{OT}" + }, + { + "bbox": [ + 313, + 600, + 556, + 695 + ], + "type": "text", + "content": ", a simple yet effective addition that corrects the degeneration of OT while maintaining straight integration paths. Extensive experiments verify the effectiveness of our method. We believe that these findings, along with the proposed simple technique, close a gap in current research on prior-data coupling." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 296, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 296, + 715 + ], + "type": "text", + "content": "There is a 0.5 chance that the distance between a random prior-data pair is closer to that of another random prior-data pair." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 696, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 696, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 696, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgment. This work is supported in part by NSF grants 2008387, 2045586, 2106825, NIFA award 2020-67021-32799, and OAC 2320345." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15882" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 94, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 94, + 297, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 94, + 297, + 125 + ], + "spans": [ + { + "bbox": [ + 61, + 94, + 297, + 125 + ], + "type": "text", + "content": "[1] Michael Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In Proc. ICLR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 128, + 296, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 296, + 160 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 296, + 160 + ], + "type": "text", + "content": "[2] Michael Albergo, Nicholas Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 162, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 162, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 162, + 296, + 215 + ], + "type": "text", + "content": "[3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi0: A vision-language-action flow model for general robot control. arXiv, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 218, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 218, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 62, + 218, + 296, + 261 + ], + "type": "text", + "content": "[4] Ho Kei Cheng, Masato Ishii, Akio Hayakawa, Takashi Shibuya, Alexander Schwing, and Yuki Mitsufuji. Taming multimodal joint training for high-quality video-to-audio synthesis. CVPR, 2025. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 263, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 263, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 296, + 285 + ], + "type": "text", + "content": "[5] chenyaofo. Pytorch cifar models. https://github.com/chenyaofo/pytorch-cifar-models, 2023.6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 286, + 296, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 296, + 318 + ], + "type": "text", + "content": "[6] Aram Davtyan, Leello Tadesse Dadi, Volkan Cevher, and Paolo Favaro. Faster inference of flow-based generative models via improved data-noise coupling. In ICLR, 2025. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 319, + 296, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 351 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 354, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 296, + 386 + ], + "type": "text", + "content": "[8] John R Dormand and Peter J Prince. A family of embedded runge-kutta formulae. Journal of computational and applied mathematics, 1980. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 388, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 441 + ], + "type": "text", + "content": "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 444, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 444, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 57, + 444, + 295, + 476 + ], + "type": "text", + "content": "[10] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. ICLR, 2024. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 478, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 478, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 57, + 478, + 296, + 509 + ], + "type": "text", + "content": "[11] Zhengyang Geng, Ashwini Pokle, William Luo, Justin Lin, and J Zico Kolter. Consistency models made easy. ICLR, 2025. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 511, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 296, + 533 + ], + "type": "text", + "content": "[12] Pengsheng Guo and Alexander G Schwing. Variational rectified flow matching. arXiv, 2025. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 535, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 535, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 535, + 296, + 578 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 580, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 623 + ], + "type": "text", + "content": "[14] Irina Higgins, Loic Matthew, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. In ICLR, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "type": "text", + "content": "[15] Ka-Hei Hui, Chao Liu, Xiaohui Zeng, Chi-Wing Fu, and Arash Vahdat. Not-so-optimal transport flows for 3d point cloud generation. arXiv, 2025. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 712 + ], + "type": "text", + "content": "[16] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. *ICLR*, 2024. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[17] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. In ICLR, 2014. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 96, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 116 + ], + "type": "text", + "content": "[18] Leon Klein, Andreas Kramer, and Frank Noé. Equivariant flow matching. NeurIPS, 36, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 118, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 139 + ], + "type": "text", + "content": "[19] Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images, 2009. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 140, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 161 + ], + "type": "text", + "content": "[20] Harold W. Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 1955. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 163, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 555, + 194 + ], + "type": "text", + "content": "[21] Visual Layer. Imagenet-1k-vl-enriched. https://huggingface.co/datasets/visual-layer/imagenet-1k-vl-enriched, 2024.6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 196, + 555, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 196, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 196, + 555, + 226 + ], + "type": "text", + "content": "[22] Sangyun Lee, Beomsu Kim, and Jong Chul Ye. Minimizing trajectory curvature of ode-based generative models. In ICML, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 229, + 555, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 555, + 260 + ], + "type": "text", + "content": "[23] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. *ICLR*, 2023. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 262, + 555, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 555, + 304 + ], + "type": "text", + "content": "[24] Qihao Liu, Xi Yin, Alan Yuille, Andrew Brown, and Mannat Singh. Flowing from words to pixels: A framework for cross-modality evolution. arXiv preprint arXiv:2412.15213, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 307, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 307, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 307, + 553, + 338 + ], + "type": "text", + "content": "[25] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. ICLR, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 339, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 361 + ], + "type": "text", + "content": "[26] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. *ICLR*, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 362, + 553, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 362, + 553, + 394 + ], + "spans": [ + { + "bbox": [ + 316, + 362, + 553, + 394 + ], + "type": "text", + "content": "[27] Ségolène Martin, Anne Gagneux, Paul Hagemann, and Gabriele Steidl. Pnp-flow: Plug-and-play image restoration with flow matching. *ICLR*, 2025. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 396, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 396, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 396, + 555, + 657 + ], + "type": "text", + "content": "[28] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, David Yan, Dhruv Choudhary, Dingkang Wang, Geet Sethi, Guan Pang, Haoyu Ma, Ishan Misra, Ji Hou, Jialiang Wang, Kiran Jagadeesh, Kunpeng Li, Luxin Zhang, Mannat Singh, Mary Williamson, Matt Le, Matthew Yu, Mitesh Kumar Singh, Peizhao Zhang, Peter Vajda, Quentin Duval, Rohit Girdhar, Roshan Sumbaly, Sai Saketh Rambhatla, Sam Tsai, Samaneh Azadi, Samyak Datta, Sanyuan Chen, Sean Bell, Sharadh Ramaswamy, Shelly Sheynin, Siddharth Bhattacharya, Simran Motwani, Tao Xu, Tianhe Li, Tingbo Hou, Wei-Ning Hsu, Xi Yin, Xiaoliang Dai, Yaniv Taigman, Yaqiao Luo, Yen-Cheng Liu, Yi-Chiao Wu, Yue Zhao, Yuval Kirstain, Zecheng He, Zijian He, Albert Pumarola, Ali Thabet, Artsiom Sanakoyeu, Arun Mallya, Baishan Guo, Boris Araya, Breena Kerr, Carleigh Wood, Ce Liu, Cen Peng, Dimitry Vengertsev, Edgar Schonfeld, Elliot Blanchard, Felix Juefei-Xu, Fraylie Nord, Jeff Liang, John Hoffman, Jonas Kohler, Kaolin Fire, Karthik Sivakumar, Lawrence Chen, Licheng Yu, Luya Gao, Markos Georgopoulos, Rashel Moritz, Sara K. Sampson, Shikai Li, Simone Parmeggiani, Steve Fine, Tara Fowler, Vladan Petrovic, and Yuming Du. Movie gen: A cast of media foundation models. arXiv, 2024. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[29] Aram-Alexandre Pooladian, Heli Ben-Hamu, Carles Domingo-Enrich, Brandon Amos, Yaron Lipman, and Ricky TQ Chen. Multisample flow matching: Straightening flows with minibatch couplings. ICML, 2023. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15883" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 542 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 298, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 298, + 127 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 297, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 297, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 297, + 161 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 205 + ], + "type": "text", + "content": "[32] Gianluigi Silvestri, Luca Ambrogioni, Chieh-Hsin Lai, Yuhta Takida, and Yuki Mitsufuji. Training consistency models with variational noise coupling. arXiv preprint arXiv:2502.18197, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 208, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 208, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 57, + 208, + 296, + 229 + ], + "type": "text", + "content": "[33] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 231, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 231, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 231, + 296, + 274 + ], + "type": "text", + "content": "[34] Yuxuan Song, Jingjing Gong, Minkai Xu, Ziyao Cao, Yanyan Lan, Stefano Ermon, Hao Zhou, and Wei-Ying Ma. Equivariant flow matching with hybrid probability transport for 3d molecule generation. NeurIPS, 36, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 296, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 329 + ], + "type": "text", + "content": "[35] Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. TMLR, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 297, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 297, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 297, + 407 + ], + "type": "text", + "content": "[36] Michael Tschannen, Alexey Gritsenko, Xiao Wang, Muhammad Ferjad Naeem, Ibrahim Alabdulmohsin, Nikhil Parthasarathy, Talfan Evans, Lucas Beyer, Ye Xia, Basil Mustafa, et al. Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features. arXiv preprint arXiv:2502.14786, 2025. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 296, + 441 + ], + "type": "text", + "content": "[37] Kaiyu Yang, Jacqueline H Yau, Li Fei-Fei, Jia Deng, and Olga Russakovsky. A study of face obfuscation in imagenet. In ICML, 2022. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 296, + 487 + ], + "type": "text", + "content": "[38] Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining straight flows with velocity consistency. arXiv, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 488, + 296, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 296, + 519 + ], + "type": "text", + "content": "[39] Jingfeng Yao and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. CVPR, 2025. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 521, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 296, + 542 + ], + "type": "text", + "content": "[40] Yichi Zhang, Yici Yan, Alex Schwing, and Zhizhen Zhao. Towards hierarchical rectified flow. *ICLR*, 2025. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "15884" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/10bd97eb-949b-4ebe-ab75-25cc73578707_content_list.json b/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/10bd97eb-949b-4ebe-ab75-25cc73578707_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0cc761851a0f55ea82cec1a9693780c136423318 --- /dev/null +++ b/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/10bd97eb-949b-4ebe-ab75-25cc73578707_content_list.json @@ -0,0 +1,1604 @@ +[ + { + "type": "text", + "text": "The Devil is in the Spurious Correlations: Boosting Moment Retrieval with Dynamic Learning", + "text_level": 1, + "bbox": [ + 235, + 130, + 759, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyang Zhou $^{1\\star}$ Fanyue Wei $^{2\\star}$ Lixin Duan $^{1}$ Angela Yao $^{2}$ Wen Li $^{1\\dagger}$", + "bbox": [ + 187, + 210, + 797, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Electronic Science and Technology of China 2National University of Singapore", + "bbox": [ + 181, + 233, + 815, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{x.y.zhou2020，lxduan，liwenbnu}@gmail.com", + "bbox": [ + 313, + 253, + 683, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wfanyue,ayao}@comp.nus.edu.sg", + "bbox": [ + 362, + 270, + 635, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Query: \"The blonde woman works out in a gym with red lighting.\"", + "bbox": [ + 96, + 292, + 501, + 305 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb97a2924d77f0090dbfd6c50cb7ac69c98eb9d08d9f29618aa97a903884e18d.jpg", + "image_caption": [ + "Figure 1. (a)" + ], + "image_footnote": [], + "bbox": [ + 96, + 306, + 571, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7793ea090ed810e53aa8fbe8687d45d682ffc86acef3ba1311fe1d5a1fa501a.jpg", + "image_caption": [ + "Figure 1. (b)", + "Figure 1. (a) Comparison of moment retrieval models under normal and spurious correlation videos by masking the content of target clips of video. We found the existing works are suffering from a crucial reason stems from the spurious correlation between the text queries and the moment context. Baselines predict the Spurious GT even if the target moments are masked. In contrast, TD-DETR predicts the segment near the mask with lower confidence. (b) To verify the issue of spurious correlation, we introduce the Spurious mAP as the metric. Our model achieves the best ratio of mAP to Spurious mAP." + ], + "image_footnote": [], + "bbox": [ + 594, + 297, + 906, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 613, + 326, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Given a textual query along with a corresponding video, the objective of moment retrieval aims to localize the moments relevant to the query within the video. While commendable results have been demonstrated by existing transformer-based approaches, predicting the accurate temporal span of the target moment is still a major challenge. This paper reveals that a crucial reason stems from the spurious correlation between the text query and the moment context. Namely, the model makes predictions by overly associating queries with background frames rather than distinguishing target moments. To address this issue, we propose a dynamic learning approach for moment retrieval, where two strategies are designed to mitigate the spurious correlation. First, we introduce a novel video synthesis approach to con", + "bbox": [ + 88, + 647, + 485, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "struct a dynamic context for the queried moment, enabling the model to attend to the target moment of the corresponding query across dynamic backgrounds. Second, to alleviate the over-association with backgrounds, we enhance representations temporally by incorporating text-dynamics interaction, which encourages the model to align text with target moments through complementary dynamic representations. With the proposed method, our model significantly alleviates the spurious correlation issue in moment retrieval and establishes new state-of-the-art performance on two popular benchmarks, i.e., QVHighlights and Charades-STA. In addition, detailed ablation studies and evaluations across different architectures demonstrate the generalization and effectiveness of the proposed strategies. Our code is available at https://github.com/xyangzhou/TD-DETR.", + "bbox": [ + 511, + 614, + 908, + 857 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 114, + 875, + 225, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 114, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20981", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 90, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Videos are a ubiquitous media format, but it can be time-consuming to browse through videos to localize specific moments. Using text to retrieve corresponding moments within a lengthy video[2, 54] plays an important role for entertainment, search, etc.", + "bbox": [ + 89, + 114, + 480, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing moment retrieval approaches [15, 17, 19, 30, 49] leverage the DETR [4] detection transformer architecture to fuse text query into video representation. They show remarkable results in detecting the target moment by text-video alignment, but accurately predicting the temporal span remains a significant challenge. The learned video representations [17, 19] are aligned with the semantics of the text, thus generally biased to subjects or spatial objects. Consequently, the semantic elements in background frames are overly associated with textual queries. Such inappropriate association between irrelevant contexts and labels is the manifestation of Spurious correlation [1, 43], which are under-explored in moment retrieval tasks.", + "bbox": [ + 89, + 190, + 482, + 385 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work identifies such spurious correlations between the text queries and moment visual context as the root cause for this performance gap in moment retrieval. Specifically, learned SOTA moment retrievers [17, 19, 30] tend to associate text queries with the background frames rather than distinguishing the target moment. Figure 1 illustrates how such spurious issues lead to a sub-optimal performance for moment retrieval. For example, the SOTA method, BAM-DETR [17], predicts a similar temporal span for an 'The blonde woman works out in a gym with red lighting.' in both the original video and an altered version where the target moment is masked out. In the masked case, the segment near the mask is a more reasonable output but the baseline methods [17, 29, 30, 49] predict the masked segment from background frames with the learned spurious correlation, which is the most unrelated output to the text query.", + "bbox": [ + 89, + 387, + 482, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the spurious issues, we propose a dynamic learning method with two novel strategies to mitigate spurious correlation. First, we propose a synthesis strategy that dynamically contextualizes the target moments for retrieval. We select similar video pairs and augment the samples by synthesizing target moments within similar contexts. This encourages the model to distinguish target moments from queries across more dynamic contexts, alleviating the over-association with background frames. Second, we propose to enhance video-text representations to reduce spurious dependencies on background frames by aligning text queries with temporal dynamics. To achieve this, we design a simple yet efficient temporal tokenizer to extract dynamic representations. Additionally, a text-dynamics interaction module is introduced to align dynamics with text queries, enabling the model to establish a stand-up correlation with complementary dynamic representations while", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mitigating background bias.", + "bbox": [ + 511, + 90, + 700, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments on two challenging benchmarks, i.e., QVHighlights and Charades-STA, show that our method outperforms existing methods by clear margins. Besides, spurious correlation in moment retrieval is validated to be largely alleviated by spurious evaluation. The detailed ablation analyses demonstrate the effectiveness of the proposed strategies to resolve the spurious issues.", + "bbox": [ + 511, + 107, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In conclusion, our contributions to this work are summarized as follows:", + "bbox": [ + 511, + 213, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to investigate the spurious correlation in moment retrieval.", + "- We propose a dynamic learning framework for moment retrieval that mitigates spurious correlations by dynamically contextualizing target moments through novel video synthesis and enhancing representations with text-dynamics alignment.", + "- The proposed method achieves SOTA performance across all benchmarks. Besides, our model also provides a strong interpretation of alleviating spurious correlations." + ], + "bbox": [ + 513, + 244, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 410, + 653, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Moment Retrieval", + "text_level": 1, + "bbox": [ + 511, + 435, + 689, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moment Retrieval aims to predict a temporal span within a video corresponding to a given natural language query [54]. Early methods had two stages: first, sampling candidate moments as proposals, and then scoring the proposals to obtain the final predictions [2, 8, 9, 23, 24]. However, more recent approaches are one-shot [18, 37, 47, 50-52, 56, 59] to predict the temporal span directly. Simultaneously, hybrid frameworks that integrate transformers for feature learning and fusion with CNNs for prediction have been widely studied. SnAG [31] explored the impact of cross-modal fusion on the scalability of video grounding models and proposed an efficient training scheme.", + "bbox": [ + 511, + 458, + 906, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although many moment retrieval methods have achieved remarkable performance, the majority primarily focus on architectural design, such as exploring new networks [16, 21, 27, 50, 52], improving contrastive learning, enhancing video-text alignment[18, 26], or refining attention mechanisms[31]. Generally, these methods focus on how to fuse text and video representations to improve the temporal span prediction but the overall performance are limited by the inefficient alignment architecture.", + "bbox": [ + 511, + 640, + 905, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Detection Transformers", + "text_level": 1, + "bbox": [ + 511, + 787, + 732, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the past several years, the wide adoption of detection transformers (DETR) [4] in video moment retrieval has gained remarkable performance [14, 17, 19, 25, 30, 48, 49]. While introducing the DETR framework to the task of moment retrieval, it demonstrated the effectiveness on QVHighlights dataset [19]. Furthermore, DETR-based ap", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20982", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/54e94479e0177e6462ac6ada01e629a369d91e600fdd8c54d9b3b3df030c9041.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed TD-DETR. Video pairs are sampled by their similarity and then forwarded into the Video Synthesizer to generate Dynamic Context (Section 3.1). Then, the Dynamic Context, i.e., synthesized pairs, are enhanced by Temporal Dynamic Tokenization and interact with textual information (Section 3.2). Finally, the enhanced video tokens are sent to the transformer encoder-decoder with prediction heads to predict moments (Section 3.3)." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 908, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "proaches have explored the relationship between moment retrieval and highlight detection, achieving significant advancements in both tasks. However, they still struggle with inaccurate span predictions [33, 54], limiting their overall reliability. To address this issue, Zhang et al. [54] analyzes the distribution of target moments and highlights the problem of imbalance, while Otani et al. [33] investigate the skewed distribution of query representations and expose the overly query-dependent phenomenon of existing models. Unlike previous studies, we reveal that spurious correlations is the key cause of inaccurate span predictions, arising from the over-association between the text query and the background frames.", + "bbox": [ + 88, + 460, + 483, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Spurious Correlation", + "text_level": 1, + "bbox": [ + 89, + 667, + 290, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One reason for the poor generalization of vision algorithms is that they are prone to memorizing patterns or contextual cues [3, 10, 11, 13, 39, 41]. These patterns or cues are often spurious correlations - misleading heuristics of the training data correlated with the majority of examples but does not hold in general [11]. While numerous approaches have been proposed to address this challenge [33, 42, 44, 45, 54, 55, 57, 58] in various domains, their impact on video understanding remains largely unexplored. For example, in image-based tasks, spurious correlations are often tied to spatial biases [1, 43], whereas in video-related tasks, such spatial biases are redundant in frames. This distinction is particularly crucial in retrieving moments from lengthy videos, where the goal is not merely", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to recognize objects but to precisely localize temporal spans with text queries, making the study of spurious correlations in this context even more critical.", + "bbox": [ + 511, + 460, + 906, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite in Moment Retrieval, Otani et al. [33] investigated the biases introduced by text queries and the neglect of semantic information, while Zhang et al.[54] explores the distribution of start and end moments. However, the issue of spurious correlation between the text and background frames is a key challenge of moment retrieval. In this work, we focus on this point and develop a new method by learning dynamics to alleviate this problem.", + "bbox": [ + 511, + 507, + 908, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 645, + 648, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective of moment retrieval is to localize a temporal span that semantically corresponds to a given text query. Consider a video represented by $L$ vision tokens, denoted by $V = \\{v_{1}, v_{2}, \\ldots, v_{L}\\}$ , along with a natural language description of $W$ words, as $\\{q_{1}, q_{2}, \\ldots, q_{W}\\}$ . A moment retrieval model predicts a temporal span $m$ with centre $m_{c}$ and duration $m_{\\sigma}$ which is relevant to the textual query, and token-wise saliency scores $\\{s_{1}, s_{2}, \\ldots, s_{L}\\}$ of the text.", + "bbox": [ + 511, + 672, + 906, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, we propose a novel framework to learn Temporal Dynamics utilizing DEtection Transformer (TD-DETR). The video pairs are sampled by similarity and then forwarded into the video synthesizer to generate dynamic contexts. Then, the representations of synthesized pairs are enhanced through temporal dynamic interaction with texts. Finally, dynamic-enhanced representations are", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20983", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "fed into the transformer encoder-decoder with prediction heads to predict target moments.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Video Synthesizer for Dynamic Context", + "text_level": 1, + "bbox": [ + 89, + 133, + 429, + 148 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce a novel video synthesizer to infuse the target moments with dynamic context for moment retrieval. Spurious correlations often stem from linking the moment's context to the text query. We address the issue by synthesizing new samples for the target moments with more dynamic contextual variations.", + "bbox": [ + 89, + 156, + 483, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spurious Pair Selection. In order to synthesize a dynamic context aligned with the target moment, we construct a spurious pair of the target moment in a given video $V$ . To ensure the challenge and rationality of the synthesized video, we select a video which is similar to $V$ contextually.", + "bbox": [ + 89, + 247, + 483, + 323 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In a training batch of $N$ video samples, we sample the most similar video $\\{V_k\\}$ for every $\\{V_i\\}$ where $k \\in [1, N], k \\neq i$ , for each $i \\in [1, N]$ to construct a spurious video pair $p_i$ . For any videos $V_j$ and $V_l$ from a batch, we employ cosine similarity to model the similarity relationship between videos as follows:", + "bbox": [ + 89, + 324, + 483, + 414 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {j, l} = \\frac {1}{L _ {j} \\cdot L _ {l}} \\sum_ {p = 0} ^ {L _ {j}} \\sum_ {q = 0, p \\neq q} ^ {L _ {l}} \\frac {\\left. v _ {p} ^ {j} ^ {T} \\cdot v _ {q} ^ {l} \\right|}{\\left\\| v _ {p} ^ {j} \\right\\| \\cdot \\left\\| v _ {q} ^ {l} \\right\\|} \\quad j, l \\in [ 1, N ]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 428, + 482, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq. 1, we denote $v_{q}^{j}$ and $v_{p}^{l}$ q-th clip of $V_{j}$ and q-th clip of $V_{l}$ , $L_{j}$ and $L_{l}$ for the length of $V_{j}$ and $V_{l}$ , $a_{j,l}$ for the average similarity between $V_{j}$ and $V_{l}$ . We select the most similar video $V_{k}$ for every video $V_{i}$ by $k = \\arg \\max_{k\\in [0,N]}s_{[i,:\\backslash \\{i\\}]}$ in the batch.", + "bbox": [ + 89, + 486, + 483, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Synthesis with Dynamic Context. After obtaining the spurious pairs, we synthesize a new video $\\tilde{V}_i$ using the target moment $m$ and the spurious pair $p_i$ , where $p_i = \\{V_i, V_k\\}$ . We then compose $\\{V_i, V_k\\}$ and dynamically refine the ground truth of the target moment $m_i$ . Illustrated by Figure 3, given video $V_i$ with $L_i$ tokens and $V_k$ with $L_k$ tokens, on the one hand, we expect the sample from $V_i$ and $V_k$ with the completeness of target moment $m_i$ of $V_i$ . We first sample from $V_i$ without ground truth tokens $\\mathrm{NG} \\in [0, L_i] \\setminus m_i$ with sampling ratio $\\alpha$ , i.e., every token in NG has the same ratio $\\alpha$ to be selected. We sample from the paired $V_k$ with sampling ratio $1 - \\alpha$ , since $V_k$ is irrelevant to ground truth $\\mathrm{GT}_i$ .", + "bbox": [ + 89, + 563, + 483, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we concatenate the two sets of sampled video tokens into $\\tilde{V}_i$ . On the other hand, we have the same process, except now we focus on $V_k$ and $\\mathrm{GT}_k$ with sampling ratio $\\alpha$ and ratio $1 - \\alpha$ to sample from the whole $V_i$ and concatenate sampled tokens into $\\tilde{V}_k$ .", + "bbox": [ + 89, + 760, + 483, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dynamics Enhancement", + "text_level": 1, + "bbox": [ + 89, + 847, + 313, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The attention of DETR-like architecture associates text queries with visual representations. Unfortunately, this", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/adce30a8d306c1003640b9a965a0983caf13509e2bf91c4bb0f6b506296bd049.jpg", + "image_caption": [ + "Figure 3. Illustration of Video Synthesis with Dynamic Context. The numbers in the boxes represent the token indices, indicating their sequential order. Vision tokens are selected based on their selection probabilities and concatenated while preserving the completeness of the ground truth spans, i.e., selection probabilities of ground truth spans are set to 1. The synthesized videos maintain their original length with a random bias." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 707, + 349 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0dd1edb110489cf5ad2162d8f83461b9c7e67c937801e41db34be8336affac99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 88, + 903, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tends to emphasize background frames, thereby exacerbating the spurious correlation. To this end, besides the dynamic context synthesis, we introduce a dynamics enhancement module to encourage our model to align text queries with temporal dynamic representations. Thus, the model considers not only the association of background frames but also background-independent dynamics. This alignment enables our model to establish a stand-up correlation between the query-related moment and its context.", + "bbox": [ + 511, + 478, + 905, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic Tokenizer. To model the dynamic temporality, we introduce a simple yet effective strategy to tokenize the temporal dynamics, which is nearly cost-free to obtain temporal dynamic representations.", + "bbox": [ + 511, + 614, + 905, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given spurious pair $p_i = \\{\\tilde{V}_i, \\tilde{V}_k\\}$ , the temporal dynamic tokenizer processes all video pairs in the same manner. Then, we use $\\tilde{V}_i$ to explain the processing steps without loss of generality.", + "bbox": [ + 511, + 675, + 905, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we concatenate a learnable token to the start of the video as follows:", + "bbox": [ + 511, + 737, + 905, + 766 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {V} _ {i} = \\left\\{s t, \\tilde {v} _ {1}, \\tilde {v} _ {2}, \\dots , \\tilde {v} _ {L} \\right\\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 779, + 903, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $st$ is a learnable token that captures the start signal of the video, preventing the loss of information from the first clip in the video representations. We then employ an element-wise difference to model the temporal dynamics.", + "bbox": [ + 511, + 811, + 905, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT = \\left\\{\\tilde {v} _ {1} - s t, \\tilde {v} _ {2} - \\tilde {v} _ {1}, \\tilde {v} _ {3} - \\tilde {v} _ {2}, \\dots , \\tilde {v} _ {L _ {i}} - \\tilde {v} _ {L _ {i} - 1} \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 883, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20984", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ is a learnable dynamic representation, it learns and focuses on the dynamic information in nearby video clips.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-Dynamics Interaction. After obtaining dynamic tokens $T$ in Eq. 3 from Dynamic Tokenizer, we utilize cross-attention for both text-vision and text-dynamic representations with the same text query to mitigate the over-emphasis on video backgrounds.", + "bbox": [ + 89, + 121, + 482, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In detail, the query for each cross-attention layer is prepared by linear projection of the video and temporal clips as follows:", + "bbox": [ + 89, + 196, + 482, + 241 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {\\delta} = \\left[ p _ {\\delta} (\\delta), \\dots , p _ {\\delta} \\left(\\delta_ {L _ {i}}\\right) \\right], \\delta \\in \\{\\tilde {V} _ {i}, T \\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 253, + 482, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The key and value are computed with the query text representation as $K_{q} = [p_{k}(q_{1}),\\dots,p_{k}(q_{W})]$ and $V_{q} = [p_{v}(q_{1}),\\dots,p_{v}(q_{L})]$ , $p_{\\delta}, p_{k}, p_{v}$ are linear projection layers for dynamic-enhanced query, temporal query, key and value. Then the cross-attention layer operates as follows:", + "bbox": [ + 89, + 282, + 483, + 359 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta^ {\\prime} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {\\delta} K _ {q} ^ {T}}{\\sqrt {d}}\\right) V _ {q}, \\delta \\in \\left\\{\\tilde {V} _ {i}, T \\right\\}, \\delta^ {\\prime} \\in \\left\\{\\tilde {V} _ {i} ^ {\\prime}, T ^ {\\prime} \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 369, + 482, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d$ is the dimension of the projected key, value and query. To emphasize the learned temporal information, we utilize the weighted element-wise addition to inject text-guided temporal representation into text-guided dynamic-enhanced representation as follows:", + "bbox": [ + 89, + 417, + 483, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {V} _ {i} ^ {\\prime} = \\beta \\cdot \\tilde {V} _ {i} + (1 - \\beta) \\cdot T ^ {\\prime}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 503, + 482, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is a hyper-parameter to adapt the addition ratio between $\\tilde{V}_i'$ and $T'$ . Hence, from the spurious pair $p_i = \\{\\tilde{V}_i, \\tilde{V}_k\\}$ , we obtain a new spurious pair $p_i' = \\{\\tilde{V}_i', \\tilde{V}_k'\\}$ .", + "bbox": [ + 89, + 534, + 483, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Network and Objectives", + "text_level": 1, + "bbox": [ + 89, + 589, + 313, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our network structure (see Figure 2) follows previous work [19, 30]. It features the synthesis of dynamic contextual videos and the enhancement of dynamic representations.", + "bbox": [ + 89, + 611, + 482, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transformer encoder-decoder with prediction heads. Given spurious pair $p_i' = \\{\\tilde{V}_i', \\tilde{V}_k'\\}$ , this module processes all video pairs in the same manner. Therefore, we use $\\tilde{V}_i'$ to explain the processing steps without loss of generality. The encoder consists of $T$ stacked typical transformer encoder layers, as in [19, 30], producing encoded representations $E_{enc}$ . Our decoder, following [30], also uses $T$ stacked typical transformer decoder layers, along with $N$ learnable moment queries representing the centre $m_c$ and duration $m_\\sigma$ . The decoder processes $E_{enc}$ with the moment queries.", + "bbox": [ + 89, + 657, + 482, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt prediction heads following [30]. A linear layer predicts saliency scores from the encoded representations, and another linear layer handles negative pairs. From the decoder outputs, a 3-layer MLP with ReLU predicts the normalized moment center and duration, while a linear layer with softmax predicts the foreground.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hungarian matching. Following [4, 19, 30], we perform Hungarian matching between the two predictions and two corresponding labels respectively. Give prediction $\\hat{y}$ of $\\tilde{V}_i'$ and corresponding to ground truth $y$ , the optimal matching results between predictions and ground truths $\\hat{\\sigma}$ can be written as $\\hat{\\sigma} = \\arg \\min_{\\sigma \\in G_N} \\sum_i^N \\mathcal{C}_{\\mathrm{match}}(y, \\hat{y}_{\\sigma(i)})$ , where $G$ is a permutation of predictions and ground truths pairs and $\\mathcal{C}_{\\mathrm{match}}$ is the matching cost.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Functions. We calculate the loss between $\\hat{\\sigma}$ and the ground truth $y$ corresponding to $\\tilde{V}_i'$ . Following [19], the $L_1$ loss $\\mathcal{L}_{L_1}$ and the gIoU [38] loss $\\mathcal{L}_{gIoU}$ are used to measure the distance and overlapping between the predictions and a cross-entropy loss $\\mathcal{L}_{cls}$ is used to measure classification. For highlight detection, we also use three loss functions which are margin ranking loss $\\mathcal{L}_{margin}$ , rank-aware contrastive loss $\\mathcal{L}_{margin}$ and negative loss $\\mathcal{L}_{neg}$ .", + "bbox": [ + 511, + 212, + 903, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss is shown as follows with $\\lambda_{*}$ as balancing coefficient:", + "bbox": [ + 511, + 334, + 903, + 362 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m o m e n t}} = \\lambda_ {L _ {1}} \\mathcal {L} _ {L _ {1}} + \\lambda_ {i o u} \\mathcal {L} _ {g I o U} + \\lambda_ {c l s} \\mathcal {L} _ {c l s}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 377, + 903, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h l} = \\lambda_ {\\text {m a r g i n}} \\mathcal {L} _ {\\text {m a r g i n}} + \\lambda_ {\\text {c o n t}} \\mathcal {L} _ {\\text {c o n t}} + \\lambda_ {\\text {n e g}} \\mathcal {L} _ {\\text {n e g}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 396, + 903, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {h l} + \\mathcal {L} _ {\\text {m o m e n t}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 415, + 903, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 444, + 645, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 469, + 702, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on two widely-used benchmarks: QVHighlights [19] and Charades-STA [8], following the setup of prior works [19, 30]. QVHighlights has over 10,000 video-query pairs primarily from vlog and news content. We use the splits defined in MomentDETR [19] and report results for both the val and test splits. Charades-STA, derived from the Charades dataset [40], has more than 18,000 video-sentence pairs. We follow the standard evaluation protocol from [8] and use two separate training and testing splits.", + "bbox": [ + 511, + 491, + 903, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Standard Evaluation. We adopt the same metrics of Moment-DETR [17, 19, 30] for evaluation. Specifically, for Moment Retrieval, we report the mean average precision (mAP) at Intersection over Union (IoU) thresholds of 0.5 and 0.75, as well as the average mAP over IoU thresholds ranging from 0.5 to 0.95 with a step size of 0.05. We also include Recall@1 (R@1) at IoU thresholds of 0.5 and 0.75. For a fair comparison, we also include the test split of the QVHighlights dataset, which is evaluated on the CodaLab competition platform [35]. Notably, performance at higher IoU thresholds, such as 0.7, serves as an indicator of more precise alignments between predicted moments and the ground truth.", + "bbox": [ + 511, + 643, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spurious Evaluation. To verify the spurious correlation with overly association the text with background, we replace the target moments in video content with random-valued masks while keeping the video duration unchanged,", + "bbox": [ + 511, + 839, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20985", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/52861c8339750c82453e5b69f2ab41642c55118535b281a1f21c5c0cfbcaba92.jpg", + "table_caption": [], + "table_footnote": [ + "$\\dagger$ reproduced by the official code", + "Table 1. Performance comparison on QVHighlights test split. Our experimental results are averaged over three runs and $\\pm$ denotes the standard deviation. For the compared methods, the results are copied from their original papers and we reproduce SnAG by the official code via slowfast+clip features." + ], + "table_body": "
MethodMR-R1MR-mAP
@0.5@0.7@0.5@0.75Average
MCN [2] ICCV'1711.412.7224.948.2210.67
CAL [6] arXiv'1925.4911.5423.407.659.89
XML [18] ECCV'2041.8330.3544.6331.7332.14
XML+ [19] NIPS'2146.6933.4647.8934.6734.90
SnAG† [31] CVPR'2459.7948.1058.6344.3742.71
SnAG /w TD-DETR66.4852.9363.7149.1146.75
Moment-DETR [19] NIPS'2152.8933.0254.8229.4030.73
UMT [25] CVPR'2256.2341.1853.8337.0136.12
MomentDiff [20] NIPS'2357.4239.6654.0235.7335.95
QD-DETR [30] CVPR'2362.4044.9862.5239.8839.86
UniVTG [22] ICCV'2358.8640.8657.6035.5935.47
CG-DETR [29] arXiv'2365.4048.4064.5042.8042.90
UVCOM [46] CVPR'2463.5548.7064.4744.0143.27
BAM-DETR [17] ECCV'2464.5348.6464.5746.3345.36
TD-DETR (Ours)64.53±0.6250.37±0.5366.21±0.2147.32±0.5346.69±0.26
", + "bbox": [ + 114, + 88, + 879, + 351 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6dae8998099b6c4aa6e3c531b57342ff43f813eb6023772f5300a4db6ee4b683.jpg", + "table_caption": [], + "table_footnote": [ + "$\\dagger$ reproduced by the official code", + "Table 2. Performance comparison on Charades-STA test split. Our experimental results are averaged over three runs. For all the compared methods, the results are taken from their original papers." + ], + "table_body": "
MethodR1@0.5R1@0.7
CAL [6]44.9024.37
2D TAN [56]39.7023.31
VSLNet [53]47.3130.19
IVG-DCL [32]50.2432.88
SnAG†[31]65.7237.32
SnAG /w TD-DETR70.1442.35
Moment-DETR [19]53.6331.37
Moment-Diff [20]55.5732.42
UMT [25]48.3129.25
QD-DETR [30]57.3132.55
CG-DETR [29]58.4036.30
BAM-DETR [17]59.9539.38
TD-DETR (Ours)60.8940.35
", + "bbox": [ + 93, + 438, + 480, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "making the masked region the most unrelated moments to text queries. We then assess the model's performance under these conditions and report the spurious metric of Spurious $R@1$ and Spurious mAP, analogous to the standard R@1 and mAP. In this case, lower values are better, as they indicate less spurious correlation.", + "bbox": [ + 89, + 750, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our method is implemented in PyTorch [34]. We follow the implementation of QD-DETR [30]. For all datasets, we use video features both extracted from SlowFast [7] pre-trained on Kinetics [5] and", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3963ae56407e89751f67fe2b6c4e9595d1d46cac3371b49528b201aab7248854.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Performance comparison on QVHighlights val split with random masks. The target clips in the video content are replaced with random-valued masks while keeping the video duration unchanged. We compare the baseline models—QD-DETR, CG-DETR, and BAM-DETR—against their respective versions integrated with our proposed TD-DETR, marked Ours w/, under the same experimental settings. The metrics here have been explained in section 4.1." + ], + "table_body": "
MethodSpurious R1 ↓Spurious mAP ↓Standard mAP ↑
@0.7@0.9@0.75Avg.@0.75Avg.
QD-DETR9.355.299.9010.4041.8241.22
Ours w/ QD8.263.687.468.1549.8649.05
CG-DETR4.651.295.556.1445.7044.90
Ours w/ CG2.580.393.384.4149.1648.38
BAM-DETR7.161.876.306.7248.5647.61
Ours w/ BAM1.610.521.731.9849.6248.67
", + "bbox": [ + 514, + 439, + 929, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pre-trained CLIP [36] vision encoder, and text feature extracted from pre-trained CLIP [36] text encoder, following Moment-DETR.", + "bbox": [ + 511, + 700, + 906, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- For QVHighlights dataset, we set batch size to 32 and an initial learning rate of $1e - 4$ with weight decay of $1e - 4$ . We set the hidden size $d = 256$ , layers of encoder/decoder $T = 3$ , and moment queries $N = 10$ . The model is trained for 200 epochs.", + "bbox": [ + 513, + 750, + 906, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- For Charades-STA, we set batch size to 32 and use an initial learning rate of $1e - 4$ with weight decay of $1e - 4$ . We set the hidden size $d = 256$ , layers of encoder/decoder $T = 3$ , and moment queries $N = 10$ . The model is trained for 100 epochs and the learning rate is decayed to", + "bbox": [ + 511, + 825, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20986", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\frac{1}{10}$ every 40 epochs.", + "bbox": [ + 107, + 90, + 238, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use mini-batch and AdamW [28] gradient descent algorithm to optimize the network parameters, initialize weights with Xavier init [12], use dropout of 0.1 for transformer layers and 0.5 for input projection layers and set cost coefficient as the same as loss balancing coefficient on all datasets. For Dynamic Video Synthesizer, the selection Probability $\\alpha$ is set to 0.5 and for Dynamic Temporal Identifier the $\\beta$ is set to 0.7 on all datasets. The total training time is approximately 8 hours on QVHighlight and 6 hours for Charades-STA, with a single NVIDIA RTX 2080 GPU.", + "bbox": [ + 89, + 107, + 480, + 257 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results on Standard Evaluation", + "text_level": 1, + "bbox": [ + 89, + 268, + 372, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Table 1 and Table 2, we compare $TD$ -DETR with several baseline methods on QVHighlights and Charades-STA benchmarks.", + "bbox": [ + 89, + 291, + 482, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on QVHighlights. On QVHighlights test split, we compare our model against proposal-based, proposal-free, and DETR-based methods, as shown in Table 1. As observed, our TD-DETR achieves state-of-the-art performance across all evaluation metrics. Notably, TD-DETR shows remarkable improvements compared to the other methods. Performance at higher IoU thresholds (e.g., 0.7) provides a clearer indication of the alignment between predicted moments and the ground truth. Our TD-DETR outperforms the previous state-of-the-art model by a substantial margin, with improvements of up to $3.56\\%$ in $\\mathbf{R}@\\mathbf{1}@\\mathbf{0.7}$ and $2.14\\%$ in $\\mathrm{mAP}@0.75$ .", + "bbox": [ + 89, + 337, + 482, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on Charades-STA. We evaluate our method methods on Charades-STA test split, compared with proposal-based, proposal-free and DETR-based methods. As shown in Table 2, TD-DETR achieves state-of-the-art performance across all evaluation metrics. Notably, our method achieves improvements of up to $2.46\\%$ in R1@0.7 and $1.57\\%$ in R1@0.5 on this challenging benchmark. Recall that performance at high IoU thresholds stands for the precision with which predictions match the ground truth data.", + "bbox": [ + 89, + 518, + 482, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Results on Spurious Correlation Evaluation", + "text_level": 1, + "bbox": [ + 89, + 666, + 464, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moment retrieval models are often prone to inaccurate predictions due to spurious correlations, a typical example being when predictions rely overly on background cues. To investigate this issue, we adopt the spurious evaluation setting pre-defined in Sec 4.1.", + "bbox": [ + 89, + 688, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our methods on several strong baselines, including QD-DETR [30], CG-DETR [29] and BAM-DETR [17]. The results in Table 3 demonstrate the alleviation of spurious correlation issue, as indicated by the improved Spurious $R1$ and Spurious $mAP$ across all compared baselines while boosting the standard $mAP$ . For instance, our methods improve QD-DETR [30] in Spurious $R1@0.9$ from 10.40 to 8.15 (30.43%), with standard $mAP@0.75$ increasing from 41.22 to 49.05 (18.99%). While for BAM", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DETR [17], it achieves the most substantial improvement, improving Spurious R1 to 1.61 and Spurious mAP to 1.73. The results on spurious correlation evaluation demonstrate the effectiveness of our proposed approaches to alleviate the spurious correlation issue and also the generalization across different baselines.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 511, + 191, + 705, + 208 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we analyze the impact of mitigating spurious correlations in our TD-DETR model. As illustrated in Figure 4, the entire video depicts a man speaking to the camera, while the target moment is described as: \"A man wearing a cap backwards talking while some other videos appear in the bottom left corner of the screen.\" Notably, baseline models fail to distinguish between \"The man is talking\" and \"The man is talking while other videos appear in the corner,\" indicating that background elements are overly correlated with the textual query. In contrast, our TD-DETR accurately predicts the target moment without being misled by background frames, demonstrating its ability to associate the information between text and the proper frames rather than relying on spurious correlations.", + "bbox": [ + 511, + 215, + 906, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 436, + 679, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies on both QVHighlights and Charades-STA benchmarks to validate the effectiveness of each proposed component. VSDC and TDEM refer to Video Synthesizer for Dynamic Context and Temporal Dynamics Enhancement Module, respectively. As illustrated in Table 4, rows (b) to (c) show the effectiveness of each component compared to the baseline model (a), and (d) demonstrate the overall effectiveness of all the components. In detail, on QVHighlight, Video Synthesizer for Dynamic Context contributes improvement of $5.60\\%$ in R1@0.7, $9.18\\%$ in mAP@0.75 and $9.05\\%$ in average mAP while Temporal Dynamics Enhancement Module contributes improvement of $3.16\\%$ in R1@0.7, $4.19\\%$ in mAP@0.75 and $5.41\\%$ in average mAP for Moment Retrieval. With all components integrated, we observe a substantial $14.75\\%$ , $14.20\\%$ and $15.30\\%$ improvement in the R1@0.7, mAP@0.75 and average mAP, respectively. Moreover, the spurious correlation is also improved, where $30.43\\%$ in Spurious R1@0.9 and $21.63\\%$ in Spurious mAP. The same phenomenon can be observed in Charades-STA as well. More experiments and analyses are included in the supplementary materials.", + "bbox": [ + 511, + 459, + 906, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Generalization across Different Architectures.", + "text_level": 1, + "bbox": [ + 511, + 787, + 901, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We validate the generalization by incorporating our proposed approaches into different architectures. SnAG [31] is a hybrid framework, which integrates transformer for feature fusion and CNN for prediction, while QD-DETR [30] and BAM-DETR [17] follow DETR-like architecture.", + "bbox": [ + 511, + 809, + 903, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As illustrated in Table 5, for the hybrid architecture, our", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20987", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b57c6f4367b9490e4acaa37292ac57fab2f77f4c8a507699f08d88d293347a15.jpg", + "image_caption": [ + "Figure 4. Visualization of results comparison between the baseline models and our TD-DETR. We display the relative position of ground truth and prediction moments in colourful boxes, where the green with GT stands for ground truth moment, the orange box with QD-DETR stands for our baseline predictions and the purple box with ours stands for our TD-DETR. To quantify the quality of predictions, we mark the start and end timestamps in the box and mark the IoU with the same colour as the models. Best read in colour." + ], + "image_footnote": [], + "bbox": [ + 102, + 98, + 895, + 297 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1c51bf30ddbf1414ee76d8894cf4f40e845f1a71bc161ec2d5382931c518ddde.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VSDCTDEMQVHighlightCharades-STA
R1↑mAP↑Spurious R1↓Spurious mAP↓R1↑Spurious R1↓
@0.5@0.7@0.5@0.75Avg.@0.7@0.9@0.75Avg.@0.5@0.7@0.7@0.9
(a)61.1246.7762.4543.6642.549.355.299.9010.4057.3132.5525.726.31
(b)✓63.4749.3964.8247.6746.398.773.878.648.9139.1263.6723.155.42
(c)✓62.9348.2564.2245.4944.848.844.09.109.5638.5160.8024.035.73
(d)✓✓65.8853.6766.4349.8649.058.263.687.468.1560.8940.3522.134.82
", + "bbox": [ + 91, + 385, + 911, + 488 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8eee95c6198ce104db6e86905dfb646331a72677112203ce5def50f63451ffea.jpg", + "table_caption": [ + "Table 4. Ablation studies on QVHighlights val split and Charades-STA test split. VSDC and TDEM stand for Video Synthesizer for Dynamic Context and Temporal Dynamics Enhancement Module respectively. All the ablation results are averaged over three runs. We also reveal spurious metrics of the proposed modules." + ], + "table_footnote": [], + "table_body": "
MethodQVHighlights valCharades-STA test
R1@0.7mAP@0.75mAPR1@0.5R1@0.7
CG52.1045.7044.9058.4036.30
Ours w/ CG53.25+1.1549.16+3.4648.38+3.4859.35+0.9537.84+1.54
BAM51.6148.5647.6159.9539.38
Ours w/BAM52.87+1.2649.62+1.0648.82+1.2160.92+0.9740.25+0.87
QD46.6641.8241.2257.3132.55
Ours w/ QD53.67+7.0149.86+8.0449.00+7.7860.89+3.5840.35+7.80
SnAG48.1044.3742.7165.7237.32
Ours w/ SnAG52.93+4.8349.11+4.7446.75+4.0470.14+4.4242.35+5.03
", + "bbox": [ + 91, + 544, + 500, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Generalization across different architectures on both QVHighlight val split and Charades-STA test split. CG stands for CG-DETR, while QD stands for QD-DETR and BAM stands for BAM-DETR.", + "bbox": [ + 89, + 681, + 482, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "method via SnAG achieves improvements of up to $10.04\\%$ in R1@0.7 and $9.46\\%$ in mAP on QVHighlights and $6.73\\%$ in R1@0.5 and $13.48\\%$ in R1@0.7 on Charades-STA. For DETR-like architecture, our approach enhances QD-DETR with significant gains of $19.23\\%$ in R1@0.7 and $18.87\\%$ in mAP on QVHighlights and $6.25\\%$ in R1@0.5 and $23.96\\%$ in R1@0.7 on Charades-STA. Our strategies significantly improve the performance of both hybrid and DETR-like baselines on both benchmarks, demonstrating the generalization of our approach.", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 558, + 633, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although existing transformer-based approaches have demonstrated remarkable performance, they still struggle with spurious correlation, which overly associates the textual query with the background frames rather than the target moment. To address this issue in moment retrieval, we introduce a novel dynamic learning method with two strategies. First, we propose a novel video synthesis strategy that constructs a dynamic context for the relevant moment. This synthesis strategy enables our model to attend to the target moment corresponding to the query across various dynamic video contexts. Second, we enhance the representation by learning temporal dynamics aligned with texts. In addition to visual features, text queries are aligned with dynamic representations, encouraging our model to establish a non-spurious correlation between the query-related moment and its context. Extensive experiments and detailed ablation studies on both the QVHighlights and Charades-STA benchmarks validate the effectiveness and generalization of the proposed TD-DETR, demonstrating significant alleviation of spurious correlation and superior performance in moment retrieval.", + "bbox": [ + 511, + 583, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20988", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (No. 62176047, No. 62476051), the Sichuan Natural Science Foundation (No. 2024NSFTD0041), and the Sichuan Science and Technology Program (No. 2021YFS0374).", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vedika Agarwal, Rakshith Shetty, and Mario Fritz. Towards causal vqa: Revealing and reducing spurious correlations by invariant and covariant semantic editing. In CVPR, pages 9690-9698, 2020. 2, 3", + "[2] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 2, 6", + "[3] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, pages 456-473, 2018. 3", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 2, 5", + "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 6", + "[6] Victor Escorcia, Mattia Soldan, Josef Sivic, Bernard Ghanem, and Bryan Russell. Finding moments in video collections using natural language. arXiv preprint arXiv:1907.12763, 2019. 6", + "[7] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 6", + "[8] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, pages 5267-5275, 2017. 2, 5", + "[9] Runzhou Ge, Jiyang Gao, Kan Chen, and Ram Nevatia. Mac: Mining activity concepts for language-based temporal localization. In WACV, pages 245-253. IEEE, 2019. 2", + "[10] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. 3", + "[11] Soumya Suva Ghosal and Yixuan Li. Are vision transformers robust to spurious correlations? In IJCV, pages 689-709. Springer, 2024. 3", + "[12] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In AISTATS, 2010. 7", + "[13] Karan Goel, Albert Gu, Yixuan Li, and Christopher Ré. Model patching: Closing the subgroup performance gap with data augmentation. In ICLR, 2021. 3", + "[14] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In ICCV, pages 13846-13856, 2023. 2" + ], + "bbox": [ + 93, + 227, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Minjoon Jung, Youwon Jang, Seongho Choi, Joochan Kim, Jin-Hwa Kim, and Byoung-Tak Zhang. Background-aware moment detection for video moment retrieval. In WACV, pages 8586-8596, 2025. 2", + "[16] Minjoon Jung, Junbin Xiao, Byoung-Tak Zhang, and Angela Yao. On the consistency of video large language models in temporal comprehension. In CVPR, pages 13713-13722, 2025. 2", + "[17] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In ECCV, pages 220-238. Springer, 2024. 2, 5, 6, 7", + "[18] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In ECCV, pages 447-463. Springer, 2020. 2, 6", + "[19] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In NeurIPS, pages 11846-11858, 2021. 2, 5, 6", + "[20] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. In NeurlPS, 2023. 6", + "[21] Xunsong Li, Pengzhan Sun, Yangcen Liu, Lixin Duan, and Wen Li. Simultaneous detection and interaction reasoning for object-centric action recognition. IEEE Transactions on Multimedia, pages 1-13, 2025. 2", + "[22] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 6", + "[23] Meng Liu, Xiang Wang, Liqiang Nie, Xiangnan He, Baoquan Chen, and Tat-Seng Chua. Attentive moment retrieval in videos. In SIGIR, pages 15-24, 2018. 2", + "[24] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In ACM MM, pages 843-851, 2018. 2", + "[25] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, pages 3042-3051, 2022. 2, 6", + "[26] Yangcen Liu, Ziyi Liu, Yuanhao Zhai, Wen Li, David Doerman, and Junsong Yuan. Stat: Towards generalizable temporal action localization. arXiv preprint arXiv:2404.13311, 2024. 2", + "[27] Ziyi Liu and Yangcen Liu. Bridge the gap: From weak to full supervision for temporal action localization with pseudoformer. In CVPR, pages 8711-8720, 2025. 2", + "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7", + "[29] WonJun Moon, Sangeek Hyun, SuBeen Lee, and JaePil Heo. Correlation-guided query-dependency calibration for video temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 2, 6, 7", + "[30] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 2, 5, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20989", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. In CVPR, pages 18930-18940, 2024. 2, 6, 7", + "[32] Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. Interventional video grounding with dual contrastive learning. In CVPR, 2021. 6", + "[33] Mayu Otani, Yuta Nakashima, Esa Rahtu, and Janne Heikkilä. Uncovering hidden challenges in query-based video moment retrieval. In BMVC, 2020. 3", + "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6", + "[35] Adrien Pavao, Isabelle Guyon, Anne-Catherine Letournel, Dinh-Tuan Tran, Xavier Baro, Hugo Jair Escalante, Sergio Escalera, Tyler Thomas, and Zhen Xu. Codalab competitions: An open source platform to organize scientific challenges. JMLR, 24(198):1-6, 2023. 5", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 6", + "[37] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. In Trans. Assoc. Comput. Linguistics, pages 25-36, 2013. 2", + "[38] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5", + "[39] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 3", + "[40] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, pages 510-526. Springer, 2016. 5", + "[41] Lifu Tu, Garima Lalwani, Spandana Gella, and He He. An empirical study on robustness to spurious correlations using pre-trained language models. In ACL, pages 621-633, 2020. 3", + "[42] Yipei Wang and Xiaogian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In AISTATS, pages 3745-3753. PMLR, 2024. 3", + "[43] Yipei Wang and Xiaoqian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In International Conference on Artificial Intelligence and Statistics, pages 3745-3753. PMLR, 2024. 2, 3", + "[44] Zhao Wang and Aron Culotta. Identifying spurious correlations for robust text classification. In EMNLP, 2020. 3", + "[45] Fanyue Wei, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Learning pixel-level distinctions for video highlight detection. In CVPR, pages 3073-3082, 2022." + ], + "bbox": [ + 91, + 92, + 480, + 887 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In CVPR, pages 18709-18719, 2024. 6", + "[47] Mengmeng Xu, Mattia Soldan, Jialin Gao, Shuming Liu, Juan-Manuel Pérez-Rúa, and Bernard Ghanem. Boundary-denoising for video activity localization. arXiv preprint arXiv:2304.02934, 2023. 2", + "[48] Yifang Xu, Yunzhuo Sun, Yang Li, Yilei Shi, Xiaoxiang Zhu, and Sidan Du. Mh-detr: Video moment and highlight detection with cross-modal transformer. In ACM MM, 2023. 2", + "[49] Jin Yang, Ping Wei, Huan Li, and Ziyang Ren. Task-driven exploration: Decoupling and inter-task feedback for joint moment retrieval and highlight detection. In CVPR, pages 18308-18318, 2024. 2", + "[50] Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos. NeurlPS, 32, 2019. 2", + "[51] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In CVPR, pages 10287-10296, 2020.", + "[52] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In CVPR, pages 1247-1257, 2019. 2", + "[53] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In ACL, 2020. 6", + "[54] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Temporal sentence grounding in videos: A survey and future directions. In IEEE TPAMI, pages 10443-10465. IEEE, 2023. 2, 3", + "[55] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. In ICML, 2022. 3", + "[56] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, pages 12870-12877, 2020. 2, 6", + "[57] Xing Zhang, Zuxuan Wu, Zejia Weng, Huazhu Fu, Jingjing Chen, Yu-Gang Jiang, and Larry S Davis. Videolt: Large-scale long-tailed video recognition. In ICCV, pages 7960-7969, 2021. 3", + "[58] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. In IEEE TPAMI, pages 10795-10816. IEEE, 2023. 3", + "[59] Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In ICCV, pages 2914-2923, 2017. 2" + ], + "bbox": [ + 516, + 92, + 903, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20990", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/10bd97eb-949b-4ebe-ab75-25cc73578707_model.json b/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/10bd97eb-949b-4ebe-ab75-25cc73578707_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bb7c02eb58eba4830a4d718e63fc3cc6f08f41a2 --- /dev/null +++ b/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/10bd97eb-949b-4ebe-ab75-25cc73578707_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.236, + 0.131, + 0.761, + 0.177 + ], + "angle": 0, + "content": "The Devil is in the Spurious Correlations: Boosting Moment Retrieval with Dynamic Learning" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.211, + 0.798, + 0.231 + ], + "angle": 0, + "content": "Xinyang Zhou\\(^{1\\star}\\) Fanyue Wei\\(^{2\\star}\\) Lixin Duan\\(^{1}\\) Angela Yao\\(^{2}\\) Wen Li\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.234, + 0.816, + 0.252 + ], + "angle": 0, + "content": "1University of Electronic Science and Technology of China 2National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.254, + 0.684, + 0.269 + ], + "angle": 0, + "content": "{x.y.zhou2020，lxduan，liwenbnu}@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.363, + 0.271, + 0.637, + 0.287 + ], + "angle": 0, + "content": "{wfanyue,ayao}@comp.nus.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.293, + 0.502, + 0.306 + ], + "angle": 0, + "content": "Query: \"The blonde woman works out in a gym with red lighting.\"" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.308, + 0.573, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.506, + 0.378, + 0.522 + ], + "angle": 0, + "content": "Figure 1. (a)" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.298, + 0.907, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.506, + 0.824, + 0.521 + ], + "angle": 0, + "content": "Figure 1. (b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.532, + 0.907, + 0.603 + ], + "angle": 0, + "content": "Figure 1. (a) Comparison of moment retrieval models under normal and spurious correlation videos by masking the content of target clips of video. We found the existing works are suffering from a crucial reason stems from the spurious correlation between the text queries and the moment context. Baselines predict the Spurious GT even if the target moments are masked. In contrast, TD-DETR predicts the segment near the mask with lower confidence. (b) To verify the issue of spurious correlation, we introduce the Spurious mAP as the metric. Our model achieves the best ratio of mAP to Spurious mAP." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.614, + 0.327, + 0.63 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.648, + 0.486, + 0.86 + ], + "angle": 0, + "content": "Given a textual query along with a corresponding video, the objective of moment retrieval aims to localize the moments relevant to the query within the video. While commendable results have been demonstrated by existing transformer-based approaches, predicting the accurate temporal span of the target moment is still a major challenge. This paper reveals that a crucial reason stems from the spurious correlation between the text query and the moment context. Namely, the model makes predictions by overly associating queries with background frames rather than distinguishing target moments. To address this issue, we propose a dynamic learning approach for moment retrieval, where two strategies are designed to mitigate the spurious correlation. First, we introduce a novel video synthesis approach to con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.909, + 0.858 + ], + "angle": 0, + "content": "struct a dynamic context for the queried moment, enabling the model to attend to the target moment of the corresponding query across dynamic backgrounds. Second, to alleviate the over-association with backgrounds, we enhance representations temporally by incorporating text-dynamics interaction, which encourages the model to align text with target moments through complementary dynamic representations. With the proposed method, our model significantly alleviates the spurious correlation issue in moment retrieval and establishes new state-of-the-art performance on two popular benchmarks, i.e., QVHighlights and Charades-STA. In addition, detailed ablation studies and evaluations across different architectures demonstrate the generalization and effectiveness of the proposed strategies. Our code is available at https://github.com/xyangzhou/TD-DETR." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.875, + 0.227, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.115, + 0.875, + 0.242, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20981" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.19 + ], + "angle": 0, + "content": "Videos are a ubiquitous media format, but it can be time-consuming to browse through videos to localize specific moments. Using text to retrieve corresponding moments within a lengthy video[2, 54] plays an important role for entertainment, search, etc." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.191, + 0.483, + 0.386 + ], + "angle": 0, + "content": "Existing moment retrieval approaches [15, 17, 19, 30, 49] leverage the DETR [4] detection transformer architecture to fuse text query into video representation. They show remarkable results in detecting the target moment by text-video alignment, but accurately predicting the temporal span remains a significant challenge. The learned video representations [17, 19] are aligned with the semantics of the text, thus generally biased to subjects or spatial objects. Consequently, the semantic elements in background frames are overly associated with textual queries. Such inappropriate association between irrelevant contexts and labels is the manifestation of Spurious correlation [1, 43], which are under-explored in moment retrieval tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.483, + 0.644 + ], + "angle": 0, + "content": "This work identifies such spurious correlations between the text queries and moment visual context as the root cause for this performance gap in moment retrieval. Specifically, learned SOTA moment retrievers [17, 19, 30] tend to associate text queries with the background frames rather than distinguishing the target moment. Figure 1 illustrates how such spurious issues lead to a sub-optimal performance for moment retrieval. For example, the SOTA method, BAM-DETR [17], predicts a similar temporal span for an 'The blonde woman works out in a gym with red lighting.' in both the original video and an altered version where the target moment is masked out. In the masked case, the segment near the mask is a more reasonable output but the baseline methods [17, 29, 30, 49] predict the masked segment from background frames with the learned spurious correlation, which is the most unrelated output to the text query." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address the spurious issues, we propose a dynamic learning method with two novel strategies to mitigate spurious correlation. First, we propose a synthesis strategy that dynamically contextualizes the target moments for retrieval. We select similar video pairs and augment the samples by synthesizing target moments within similar contexts. This encourages the model to distinguish target moments from queries across more dynamic contexts, alleviating the over-association with background frames. Second, we propose to enhance video-text representations to reduce spurious dependencies on background frames by aligning text queries with temporal dynamics. To achieve this, we design a simple yet efficient temporal tokenizer to extract dynamic representations. Additionally, a text-dynamics interaction module is introduced to align dynamics with text queries, enabling the model to establish a stand-up correlation with complementary dynamic representations while" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.701, + 0.107 + ], + "angle": 0, + "content": "mitigating background bias." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.108, + 0.905, + 0.212 + ], + "angle": 0, + "content": "Experiments on two challenging benchmarks, i.e., QVHighlights and Charades-STA, show that our method outperforms existing methods by clear margins. Besides, spurious correlation in moment retrieval is validated to be largely alleviated by spurious evaluation. The detailed ablation analyses demonstrate the effectiveness of the proposed strategies to resolve the spurious issues." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.214, + 0.905, + 0.243 + ], + "angle": 0, + "content": "In conclusion, our contributions to this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.245, + 0.904, + 0.274 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to investigate the spurious correlation in moment retrieval." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.275, + 0.905, + 0.349 + ], + "angle": 0, + "content": "- We propose a dynamic learning framework for moment retrieval that mitigates spurious correlations by dynamically contextualizing target moments through novel video synthesis and enhancing representations with text-dynamics alignment." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.35, + 0.905, + 0.395 + ], + "angle": 0, + "content": "- The proposed method achieves SOTA performance across all benchmarks. Besides, our model also provides a strong interpretation of alleviating spurious correlations." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.245, + 0.905, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.411, + 0.655, + 0.426 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.436, + 0.691, + 0.451 + ], + "angle": 0, + "content": "2.1. Moment Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.639 + ], + "angle": 0, + "content": "Moment Retrieval aims to predict a temporal span within a video corresponding to a given natural language query [54]. Early methods had two stages: first, sampling candidate moments as proposals, and then scoring the proposals to obtain the final predictions [2, 8, 9, 23, 24]. However, more recent approaches are one-shot [18, 37, 47, 50-52, 56, 59] to predict the temporal span directly. Simultaneously, hybrid frameworks that integrate transformers for feature learning and fusion with CNNs for prediction have been widely studied. SnAG [31] explored the impact of cross-modal fusion on the scalability of video grounding models and proposed an efficient training scheme." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.906, + 0.777 + ], + "angle": 0, + "content": "Although many moment retrieval methods have achieved remarkable performance, the majority primarily focus on architectural design, such as exploring new networks [16, 21, 27, 50, 52], improving contrastive learning, enhancing video-text alignment[18, 26], or refining attention mechanisms[31]. Generally, these methods focus on how to fuse text and video representations to improve the temporal span prediction but the overall performance are limited by the inefficient alignment architecture." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.733, + 0.803 + ], + "angle": 0, + "content": "2.2. Detection Transformers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In the past several years, the wide adoption of detection transformers (DETR) [4] in video moment retrieval has gained remarkable performance [14, 17, 19, 25, 30, 48, 49]. While introducing the DETR framework to the task of moment retrieval, it demonstrated the effectiveness on QVHighlights dataset [19]. Furthermore, DETR-based ap" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20982" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.91, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.385, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed TD-DETR. Video pairs are sampled by their similarity and then forwarded into the Video Synthesizer to generate Dynamic Context (Section 3.1). Then, the Dynamic Context, i.e., synthesized pairs, are enhanced by Temporal Dynamic Tokenization and interact with textual information (Section 3.2). Finally, the enhanced video tokens are sent to the transformer encoder-decoder with prediction heads to predict moments (Section 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.484, + 0.658 + ], + "angle": 0, + "content": "proaches have explored the relationship between moment retrieval and highlight detection, achieving significant advancements in both tasks. However, they still struggle with inaccurate span predictions [33, 54], limiting their overall reliability. To address this issue, Zhang et al. [54] analyzes the distribution of target moments and highlights the problem of imbalance, while Otani et al. [33] investigate the skewed distribution of query representations and expose the overly query-dependent phenomenon of existing models. Unlike previous studies, we reveal that spurious correlations is the key cause of inaccurate span predictions, arising from the over-association between the text query and the background frames." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.292, + 0.684 + ], + "angle": 0, + "content": "2.3. Spurious Correlation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "One reason for the poor generalization of vision algorithms is that they are prone to memorizing patterns or contextual cues [3, 10, 11, 13, 39, 41]. These patterns or cues are often spurious correlations - misleading heuristics of the training data correlated with the majority of examples but does not hold in general [11]. While numerous approaches have been proposed to address this challenge [33, 42, 44, 45, 54, 55, 57, 58] in various domains, their impact on video understanding remains largely unexplored. For example, in image-based tasks, spurious correlations are often tied to spatial biases [1, 43], whereas in video-related tasks, such spatial biases are redundant in frames. This distinction is particularly crucial in retrieving moments from lengthy videos, where the goal is not merely" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.506 + ], + "angle": 0, + "content": "to recognize objects but to precisely localize temporal spans with text queries, making the study of spurious correlations in this context even more critical." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.909, + 0.63 + ], + "angle": 0, + "content": "Despite in Moment Retrieval, Otani et al. [33] investigated the biases introduced by text queries and the neglect of semantic information, while Zhang et al.[54] explores the distribution of start and end moments. However, the issue of spurious correlation between the text and background frames is a key challenge of moment retrieval. In this work, we focus on this point and develop a new method by learning dynamics to alleviate this problem." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.646, + 0.649, + 0.663 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.794 + ], + "angle": 0, + "content": "The objective of moment retrieval is to localize a temporal span that semantically corresponds to a given text query. Consider a video represented by \\( L \\) vision tokens, denoted by \\( V = \\{v_{1}, v_{2}, \\ldots, v_{L}\\} \\), along with a natural language description of \\( W \\) words, as \\( \\{q_{1}, q_{2}, \\ldots, q_{W}\\} \\). A moment retrieval model predicts a temporal span \\( m \\) with centre \\( m_{c} \\) and duration \\( m_{\\sigma} \\) which is relevant to the textual query, and token-wise saliency scores \\( \\{s_{1}, s_{2}, \\ldots, s_{L}\\} \\) of the text." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "As shown in Figure 2, we propose a novel framework to learn Temporal Dynamics utilizing DEtection Transformer (TD-DETR). The video pairs are sampled by similarity and then forwarded into the video synthesizer to generate dynamic contexts. Then, the representations of synthesized pairs are enhanced through temporal dynamic interaction with texts. Finally, dynamic-enhanced representations are" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20983" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "fed into the transformer encoder-decoder with prediction heads to predict target moments." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.134, + 0.431, + 0.15 + ], + "angle": 0, + "content": "3.1. Video Synthesizer for Dynamic Context" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.157, + 0.484, + 0.246 + ], + "angle": 0, + "content": "We introduce a novel video synthesizer to infuse the target moments with dynamic context for moment retrieval. Spurious correlations often stem from linking the moment's context to the text query. We address the issue by synthesizing new samples for the target moments with more dynamic contextual variations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.484, + 0.324 + ], + "angle": 0, + "content": "Spurious Pair Selection. In order to synthesize a dynamic context aligned with the target moment, we construct a spurious pair of the target moment in a given video \\( V \\). To ensure the challenge and rationality of the synthesized video, we select a video which is similar to \\( V \\) contextually." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.415 + ], + "angle": 0, + "content": "In a training batch of \\(N\\) video samples, we sample the most similar video \\(\\{V_k\\}\\) for every \\(\\{V_i\\}\\) where \\(k \\in [1, N], k \\neq i\\), for each \\(i \\in [1, N]\\) to construct a spurious video pair \\(p_i\\). For any videos \\(V_j\\) and \\(V_l\\) from a batch, we employ cosine similarity to model the similarity relationship between videos as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.429, + 0.483, + 0.485 + ], + "angle": 0, + "content": "\\[\ns _ {j, l} = \\frac {1}{L _ {j} \\cdot L _ {l}} \\sum_ {p = 0} ^ {L _ {j}} \\sum_ {q = 0, p \\neq q} ^ {L _ {l}} \\frac {\\left. v _ {p} ^ {j} ^ {T} \\cdot v _ {q} ^ {l} \\right|}{\\left\\| v _ {p} ^ {j} \\right\\| \\cdot \\left\\| v _ {q} ^ {l} \\right\\|} \\quad j, l \\in [ 1, N ]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.561 + ], + "angle": 0, + "content": "In Eq. 1, we denote \\( v_{q}^{j} \\) and \\( v_{p}^{l} \\) q-th clip of \\( V_{j} \\) and q-th clip of \\( V_{l} \\), \\( L_{j} \\) and \\( L_{l} \\) for the length of \\( V_{j} \\) and \\( V_{l} \\), \\( a_{j,l} \\) for the average similarity between \\( V_{j} \\) and \\( V_{l} \\). We select the most similar video \\( V_{k} \\) for every video \\( V_{i} \\) by \\( k = \\arg \\max_{k\\in [0,N]}s_{[i,:\\backslash \\{i\\}]}\\) in the batch." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.484, + 0.759 + ], + "angle": 0, + "content": "Video Synthesis with Dynamic Context. After obtaining the spurious pairs, we synthesize a new video \\(\\tilde{V}_i\\) using the target moment \\(m\\) and the spurious pair \\(p_i\\), where \\(p_i = \\{V_i, V_k\\}\\). We then compose \\(\\{V_i, V_k\\}\\) and dynamically refine the ground truth of the target moment \\(m_i\\). Illustrated by Figure 3, given video \\(V_i\\) with \\(L_i\\) tokens and \\(V_k\\) with \\(L_k\\) tokens, on the one hand, we expect the sample from \\(V_i\\) and \\(V_k\\) with the completeness of target moment \\(m_i\\) of \\(V_i\\). We first sample from \\(V_i\\) without ground truth tokens \\(\\mathrm{NG} \\in [0, L_i] \\setminus m_i\\) with sampling ratio \\(\\alpha\\), i.e., every token in NG has the same ratio \\(\\alpha\\) to be selected. We sample from the paired \\(V_k\\) with sampling ratio \\(1 - \\alpha\\), since \\(V_k\\) is irrelevant to ground truth \\(\\mathrm{GT}_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.484, + 0.837 + ], + "angle": 0, + "content": "Finally, we concatenate the two sets of sampled video tokens into \\(\\tilde{V}_i\\). On the other hand, we have the same process, except now we focus on \\(V_k\\) and \\(\\mathrm{GT}_k\\) with sampling ratio \\(\\alpha\\) and ratio \\(1 - \\alpha\\) to sample from the whole \\(V_i\\) and concatenate sampled tokens into \\(\\tilde{V}_k\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.315, + 0.864 + ], + "angle": 0, + "content": "3.2. Dynamics Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The attention of DETR-like architecture associates text queries with visual representations. Unfortunately, this" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.709, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.089, + 0.904, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.36, + 0.907, + 0.458 + ], + "angle": 0, + "content": "Figure 3. Illustration of Video Synthesis with Dynamic Context. The numbers in the boxes represent the token indices, indicating their sequential order. Vision tokens are selected based on their selection probabilities and concatenated while preserving the completeness of the ground truth spans, i.e., selection probabilities of ground truth spans are set to 1. The synthesized videos maintain their original length with a random bias." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.906, + 0.614 + ], + "angle": 0, + "content": "tends to emphasize background frames, thereby exacerbating the spurious correlation. To this end, besides the dynamic context synthesis, we introduce a dynamics enhancement module to encourage our model to align text queries with temporal dynamic representations. Thus, the model considers not only the association of background frames but also background-independent dynamics. This alignment enables our model to establish a stand-up correlation between the query-related moment and its context." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.906, + 0.675 + ], + "angle": 0, + "content": "Dynamic Tokenizer. To model the dynamic temporality, we introduce a simple yet effective strategy to tokenize the temporal dynamics, which is nearly cost-free to obtain temporal dynamic representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.676, + 0.906, + 0.737 + ], + "angle": 0, + "content": "Given spurious pair \\( p_i = \\{\\tilde{V}_i, \\tilde{V}_k\\} \\), the temporal dynamic tokenizer processes all video pairs in the same manner. Then, we use \\( \\tilde{V}_i \\) to explain the processing steps without loss of generality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.738, + 0.906, + 0.767 + ], + "angle": 0, + "content": "First, we concatenate a learnable token to the start of the video as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.78, + 0.905, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\tilde {V} _ {i} = \\left\\{s t, \\tilde {v} _ {1}, \\tilde {v} _ {2}, \\dots , \\tilde {v} _ {L} \\right\\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.812, + 0.906, + 0.872 + ], + "angle": 0, + "content": "where \\( st \\) is a learnable token that captures the start signal of the video, preventing the loss of information from the first clip in the video representations. We then employ an element-wise difference to model the temporal dynamics." + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.885, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\nT = \\left\\{\\tilde {v} _ {1} - s t, \\tilde {v} _ {2} - \\tilde {v} _ {1}, \\tilde {v} _ {3} - \\tilde {v} _ {2}, \\dots , \\tilde {v} _ {L _ {i}} - \\tilde {v} _ {L _ {i} - 1} \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20984" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "where \\( T \\) is a learnable dynamic representation, it learns and focuses on the dynamic information in nearby video clips." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.483, + 0.197 + ], + "angle": 0, + "content": "Text-Dynamics Interaction. After obtaining dynamic tokens \\( T \\) in Eq. 3 from Dynamic Tokenizer, we utilize cross-attention for both text-vision and text-dynamic representations with the same text query to mitigate the over-emphasis on video backgrounds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.483, + 0.242 + ], + "angle": 0, + "content": "In detail, the query for each cross-attention layer is prepared by linear projection of the video and temporal clips as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.254, + 0.483, + 0.273 + ], + "angle": 0, + "content": "\\[\nQ _ {\\delta} = \\left[ p _ {\\delta} (\\delta), \\dots , p _ {\\delta} \\left(\\delta_ {L _ {i}}\\right) \\right], \\delta \\in \\{\\tilde {V} _ {i}, T \\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.284, + 0.484, + 0.361 + ], + "angle": 0, + "content": "The key and value are computed with the query text representation as \\( K_{q} = [p_{k}(q_{1}),\\dots,p_{k}(q_{W})] \\) and \\( V_{q} = [p_{v}(q_{1}),\\dots,p_{v}(q_{L})] \\), \\( p_{\\delta}, p_{k}, p_{v} \\) are linear projection layers for dynamic-enhanced query, temporal query, key and value. Then the cross-attention layer operates as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.371, + 0.483, + 0.418 + ], + "angle": 0, + "content": "\\[\n\\delta^ {\\prime} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {\\delta} K _ {q} ^ {T}}{\\sqrt {d}}\\right) V _ {q}, \\delta \\in \\left\\{\\tilde {V} _ {i}, T \\right\\}, \\delta^ {\\prime} \\in \\left\\{\\tilde {V} _ {i} ^ {\\prime}, T ^ {\\prime} \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.495 + ], + "angle": 0, + "content": "where \\(d\\) is the dimension of the projected key, value and query. To emphasize the learned temporal information, we utilize the weighted element-wise addition to inject text-guided temporal representation into text-guided dynamic-enhanced representation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.505, + 0.483, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\tilde {V} _ {i} ^ {\\prime} = \\beta \\cdot \\tilde {V} _ {i} + (1 - \\beta) \\cdot T ^ {\\prime}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.583 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a hyper-parameter to adapt the addition ratio between \\(\\tilde{V}_i'\\) and \\(T'\\). Hence, from the spurious pair \\(p_i = \\{\\tilde{V}_i, \\tilde{V}_k\\}\\), we obtain a new spurious pair \\(p_i' = \\{\\tilde{V}_i', \\tilde{V}_k'\\}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.59, + 0.314, + 0.607 + ], + "angle": 0, + "content": "3.3. Network and Objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.658 + ], + "angle": 0, + "content": "Our network structure (see Figure 2) follows previous work [19, 30]. It features the synthesis of dynamic contextual videos and the enhancement of dynamic representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.811 + ], + "angle": 0, + "content": "Transformer encoder-decoder with prediction heads. Given spurious pair \\( p_i' = \\{\\tilde{V}_i', \\tilde{V}_k'\\} \\), this module processes all video pairs in the same manner. Therefore, we use \\( \\tilde{V}_i' \\) to explain the processing steps without loss of generality. The encoder consists of \\( T \\) stacked typical transformer encoder layers, as in [19, 30], producing encoded representations \\( E_{enc} \\). Our decoder, following [30], also uses \\( T \\) stacked typical transformer decoder layers, along with \\( N \\) learnable moment queries representing the centre \\( m_c \\) and duration \\( m_\\sigma \\). The decoder processes \\( E_{enc} \\) with the moment queries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We adopt prediction heads following [30]. A linear layer predicts saliency scores from the encoded representations, and another linear layer handles negative pairs. From the decoder outputs, a 3-layer MLP with ReLU predicts the normalized moment center and duration, while a linear layer with softmax predicts the foreground." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "Hungarian matching. Following [4, 19, 30], we perform Hungarian matching between the two predictions and two corresponding labels respectively. Give prediction \\(\\hat{y}\\) of \\(\\tilde{V}_i'\\) and corresponding to ground truth \\(y\\), the optimal matching results between predictions and ground truths \\(\\hat{\\sigma}\\) can be written as \\(\\hat{\\sigma} = \\arg \\min_{\\sigma \\in G_N} \\sum_i^N \\mathcal{C}_{\\mathrm{match}}(y, \\hat{y}_{\\sigma(i)})\\), where \\(G\\) is a permutation of predictions and ground truths pairs and \\(\\mathcal{C}_{\\mathrm{match}}\\) is the matching cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.335 + ], + "angle": 0, + "content": "Loss Functions. We calculate the loss between \\(\\hat{\\sigma}\\) and the ground truth \\(y\\) corresponding to \\(\\tilde{V}_i'\\). Following [19], the \\(L_1\\) loss \\(\\mathcal{L}_{L_1}\\) and the gIoU [38] loss \\(\\mathcal{L}_{gIoU}\\) are used to measure the distance and overlapping between the predictions and a cross-entropy loss \\(\\mathcal{L}_{cls}\\) is used to measure classification. For highlight detection, we also use three loss functions which are margin ranking loss \\(\\mathcal{L}_{margin}\\), rank-aware contrastive loss \\(\\mathcal{L}_{margin}\\) and negative loss \\(\\mathcal{L}_{neg}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.905, + 0.363 + ], + "angle": 0, + "content": "The overall loss is shown as follows with \\(\\lambda_{*}\\) as balancing coefficient:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.378, + 0.905, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m o m e n t}} = \\lambda_ {L _ {1}} \\mathcal {L} _ {L _ {1}} + \\lambda_ {i o u} \\mathcal {L} _ {g I o U} + \\lambda_ {c l s} \\mathcal {L} _ {c l s}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.397, + 0.905, + 0.413 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h l} = \\lambda_ {\\text {m a r g i n}} \\mathcal {L} _ {\\text {m a r g i n}} + \\lambda_ {\\text {c o n t}} \\mathcal {L} _ {\\text {c o n t}} + \\lambda_ {\\text {n e g}} \\mathcal {L} _ {\\text {n e g}}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.416, + 0.905, + 0.431 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {h l} + \\mathcal {L} _ {\\text {m o m e n t}}. \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.445, + 0.646, + 0.462 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.704, + 0.487 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.905, + 0.644 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on two widely-used benchmarks: QVHighlights [19] and Charades-STA [8], following the setup of prior works [19, 30]. QVHighlights has over 10,000 video-query pairs primarily from vlog and news content. We use the splits defined in MomentDETR [19] and report results for both the val and test splits. Charades-STA, derived from the Charades dataset [40], has more than 18,000 video-sentence pairs. We follow the standard evaluation protocol from [8] and use two separate training and testing splits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Standard Evaluation. We adopt the same metrics of Moment-DETR [17, 19, 30] for evaluation. Specifically, for Moment Retrieval, we report the mean average precision (mAP) at Intersection over Union (IoU) thresholds of 0.5 and 0.75, as well as the average mAP over IoU thresholds ranging from 0.5 to 0.95 with a step size of 0.05. We also include Recall@1 (R@1) at IoU thresholds of 0.5 and 0.75. For a fair comparison, we also include the test split of the QVHighlights dataset, which is evaluated on the CodaLab competition platform [35]. Notably, performance at higher IoU thresholds, such as 0.7, serves as an indicator of more precise alignments between predicted moments and the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Spurious Evaluation. To verify the spurious correlation with overly association the text with background, we replace the target moments in video content with random-valued masks while keeping the video duration unchanged," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20985" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.88, + 0.352 + ], + "angle": 0, + "content": "
MethodMR-R1MR-mAP
@0.5@0.7@0.5@0.75Average
MCN [2] ICCV'1711.412.7224.948.2210.67
CAL [6] arXiv'1925.4911.5423.407.659.89
XML [18] ECCV'2041.8330.3544.6331.7332.14
XML+ [19] NIPS'2146.6933.4647.8934.6734.90
SnAG† [31] CVPR'2459.7948.1058.6344.3742.71
SnAG /w TD-DETR66.4852.9363.7149.1146.75
Moment-DETR [19] NIPS'2152.8933.0254.8229.4030.73
UMT [25] CVPR'2256.2341.1853.8337.0136.12
MomentDiff [20] NIPS'2357.4239.6654.0235.7335.95
QD-DETR [30] CVPR'2362.4044.9862.5239.8839.86
UniVTG [22] ICCV'2358.8640.8657.6035.5935.47
CG-DETR [29] arXiv'2365.4048.4064.5042.8042.90
UVCOM [46] CVPR'2463.5548.7064.4744.0143.27
BAM-DETR [17] ECCV'2464.5348.6464.5746.3345.36
TD-DETR (Ours)64.53±0.6250.37±0.5366.21±0.2147.32±0.5346.69±0.26
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.114, + 0.354, + 0.329, + 0.371 + ], + "angle": 0, + "content": "\\(\\dagger\\) reproduced by the official code" + }, + { + "type": "table_footnote", + "bbox": [ + 0.111, + 0.38, + 0.88, + 0.424 + ], + "angle": 0, + "content": "Table 1. Performance comparison on QVHighlights test split. Our experimental results are averaged over three runs and \\(\\pm\\) denotes the standard deviation. For the compared methods, the results are copied from their original papers and we reproduce SnAG by the official code via slowfast+clip features." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.439, + 0.482, + 0.658 + ], + "angle": 0, + "content": "
MethodR1@0.5R1@0.7
CAL [6]44.9024.37
2D TAN [56]39.7023.31
VSLNet [53]47.3130.19
IVG-DCL [32]50.2432.88
SnAG†[31]65.7237.32
SnAG /w TD-DETR70.1442.35
Moment-DETR [19]53.6331.37
Moment-Diff [20]55.5732.42
UMT [25]48.3129.25
QD-DETR [30]57.3132.55
CG-DETR [29]58.4036.30
BAM-DETR [17]59.9539.38
TD-DETR (Ours)60.8940.35
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.661, + 0.308, + 0.677 + ], + "angle": 0, + "content": "\\(\\dagger\\) reproduced by the official code" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.731 + ], + "angle": 0, + "content": "Table 2. Performance comparison on Charades-STA test split. Our experimental results are averaged over three runs. For all the compared methods, the results are taken from their original papers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.84 + ], + "angle": 0, + "content": "making the masked region the most unrelated moments to text queries. We then assess the model's performance under these conditions and report the spurious metric of Spurious \\( R@1 \\) and Spurious mAP, analogous to the standard R@1 and mAP. In this case, lower values are better, as they indicate less spurious correlation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. Our method is implemented in PyTorch [34]. We follow the implementation of QD-DETR [30]. For all datasets, we use video features both extracted from SlowFast [7] pre-trained on Kinetics [5] and" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.44, + 0.93, + 0.556 + ], + "angle": 0, + "content": "
MethodSpurious R1 ↓Spurious mAP ↓Standard mAP ↑
@0.7@0.9@0.75Avg.@0.75Avg.
QD-DETR9.355.299.9010.4041.8241.22
Ours w/ QD8.263.687.468.1549.8649.05
CG-DETR4.651.295.556.1445.7044.90
Ours w/ CG2.580.393.384.4149.1648.38
BAM-DETR7.161.876.306.7248.5647.61
Ours w/ BAM1.610.521.731.9849.6248.67
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.512, + 0.566, + 0.907, + 0.678 + ], + "angle": 0, + "content": "Table 3. Performance comparison on QVHighlights val split with random masks. The target clips in the video content are replaced with random-valued masks while keeping the video duration unchanged. We compare the baseline models—QD-DETR, CG-DETR, and BAM-DETR—against their respective versions integrated with our proposed TD-DETR, marked Ours w/, under the same experimental settings. The metrics here have been explained in section 4.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.907, + 0.746 + ], + "angle": 0, + "content": "pre-trained CLIP [36] vision encoder, and text feature extracted from pre-trained CLIP [36] text encoder, following Moment-DETR." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.75, + 0.907, + 0.826 + ], + "angle": 0, + "content": "- For QVHighlights dataset, we set batch size to 32 and an initial learning rate of \\(1e - 4\\) with weight decay of \\(1e - 4\\). We set the hidden size \\(d = 256\\), layers of encoder/decoder \\(T = 3\\), and moment queries \\(N = 10\\). The model is trained for 200 epochs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.909, + 0.902 + ], + "angle": 0, + "content": "- For Charades-STA, we set batch size to 32 and use an initial learning rate of \\(1e - 4\\) with weight decay of \\(1e - 4\\). We set the hidden size \\(d = 256\\), layers of encoder/decoder \\(T = 3\\), and moment queries \\(N = 10\\). The model is trained for 100 epochs and the learning rate is decayed to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20986" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.108, + 0.091, + 0.24, + 0.108 + ], + "angle": 0, + "content": "\\(\\frac{1}{10}\\) every 40 epochs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.482, + 0.258 + ], + "angle": 0, + "content": "We use mini-batch and AdamW [28] gradient descent algorithm to optimize the network parameters, initialize weights with Xavier init [12], use dropout of 0.1 for transformer layers and 0.5 for input projection layers and set cost coefficient as the same as loss balancing coefficient on all datasets. For Dynamic Video Synthesizer, the selection Probability \\(\\alpha\\) is set to 0.5 and for Dynamic Temporal Identifier the \\(\\beta\\) is set to 0.7 on all datasets. The total training time is approximately 8 hours on QVHighlight and 6 hours for Charades-STA, with a single NVIDIA RTX 2080 GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.269, + 0.373, + 0.285 + ], + "angle": 0, + "content": "4.2. Results on Standard Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.483, + 0.336 + ], + "angle": 0, + "content": "As illustrated in Table 1 and Table 2, we compare \\(TD\\)-DETR with several baseline methods on QVHighlights and Charades-STA benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.483, + 0.518 + ], + "angle": 0, + "content": "Results on QVHighlights. On QVHighlights test split, we compare our model against proposal-based, proposal-free, and DETR-based methods, as shown in Table 1. As observed, our TD-DETR achieves state-of-the-art performance across all evaluation metrics. Notably, TD-DETR shows remarkable improvements compared to the other methods. Performance at higher IoU thresholds (e.g., 0.7) provides a clearer indication of the alignment between predicted moments and the ground truth. Our TD-DETR outperforms the previous state-of-the-art model by a substantial margin, with improvements of up to \\(3.56\\%\\) in \\(\\mathbf{R}@\\mathbf{1}@\\mathbf{0.7}\\) and \\(2.14\\%\\) in \\(\\mathrm{mAP}@0.75\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Results on Charades-STA. We evaluate our method methods on Charades-STA test split, compared with proposal-based, proposal-free and DETR-based methods. As shown in Table 2, TD-DETR achieves state-of-the-art performance across all evaluation metrics. Notably, our method achieves improvements of up to \\(2.46\\%\\) in R1@0.7 and \\(1.57\\%\\) in R1@0.5 on this challenging benchmark. Recall that performance at high IoU thresholds stands for the precision with which predictions match the ground truth data." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.667, + 0.465, + 0.683 + ], + "angle": 0, + "content": "4.3. Results on Spurious Correlation Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Moment retrieval models are often prone to inaccurate predictions due to spurious correlations, a typical example being when predictions rely overly on background cues. To investigate this issue, we adopt the spurious evaluation setting pre-defined in Sec 4.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate our methods on several strong baselines, including QD-DETR [30], CG-DETR [29] and BAM-DETR [17]. The results in Table 3 demonstrate the alleviation of spurious correlation issue, as indicated by the improved Spurious \\( R1 \\) and Spurious \\( mAP \\) across all compared baselines while boosting the standard \\( mAP \\). For instance, our methods improve QD-DETR [30] in Spurious \\( R1@0.9 \\) from 10.40 to 8.15 (30.43%), with standard \\( mAP@0.75 \\) increasing from 41.22 to 49.05 (18.99%). While for BAM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "DETR [17], it achieves the most substantial improvement, improving Spurious R1 to 1.61 and Spurious mAP to 1.73. The results on spurious correlation evaluation demonstrate the effectiveness of our proposed approaches to alleviate the spurious correlation issue and also the generalization across different baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.193, + 0.707, + 0.209 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.216, + 0.907, + 0.427 + ], + "angle": 0, + "content": "In this subsection, we analyze the impact of mitigating spurious correlations in our TD-DETR model. As illustrated in Figure 4, the entire video depicts a man speaking to the camera, while the target moment is described as: \"A man wearing a cap backwards talking while some other videos appear in the bottom left corner of the screen.\" Notably, baseline models fail to distinguish between \"The man is talking\" and \"The man is talking while other videos appear in the corner,\" indicating that background elements are overly correlated with the textual query. In contrast, our TD-DETR accurately predicts the target moment without being misled by background frames, demonstrating its ability to associate the information between text and the proper frames rather than relying on spurious correlations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.68, + 0.453 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.779 + ], + "angle": 0, + "content": "We conduct ablation studies on both QVHighlights and Charades-STA benchmarks to validate the effectiveness of each proposed component. VSDC and TDEM refer to Video Synthesizer for Dynamic Context and Temporal Dynamics Enhancement Module, respectively. As illustrated in Table 4, rows (b) to (c) show the effectiveness of each component compared to the baseline model (a), and (d) demonstrate the overall effectiveness of all the components. In detail, on QVHighlight, Video Synthesizer for Dynamic Context contributes improvement of \\(5.60\\%\\) in R1@0.7, \\(9.18\\%\\) in mAP@0.75 and \\(9.05\\%\\) in average mAP while Temporal Dynamics Enhancement Module contributes improvement of \\(3.16\\%\\) in R1@0.7, \\(4.19\\%\\) in mAP@0.75 and \\(5.41\\%\\) in average mAP for Moment Retrieval. With all components integrated, we observe a substantial \\(14.75\\%\\), \\(14.20\\%\\) and \\(15.30\\%\\) improvement in the R1@0.7, mAP@0.75 and average mAP, respectively. Moreover, the spurious correlation is also improved, where \\(30.43\\%\\) in Spurious R1@0.9 and \\(21.63\\%\\) in Spurious mAP. The same phenomenon can be observed in Charades-STA as well. More experiments and analyses are included in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.902, + 0.803 + ], + "angle": 0, + "content": "4.6. Generalization across Different Architectures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.885 + ], + "angle": 0, + "content": "We validate the generalization by incorporating our proposed approaches into different architectures. SnAG [31] is a hybrid framework, which integrates transformer for feature fusion and CNN for prediction, while QD-DETR [30] and BAM-DETR [17] follow DETR-like architecture." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Table 5, for the hybrid architecture, our" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20987" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.099, + 0.896, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.316, + 0.908, + 0.375 + ], + "angle": 0, + "content": "Figure 4. Visualization of results comparison between the baseline models and our TD-DETR. We display the relative position of ground truth and prediction moments in colourful boxes, where the green with GT stands for ground truth moment, the orange box with QD-DETR stands for our baseline predictions and the purple box with ours stands for our TD-DETR. To quantify the quality of predictions, we mark the start and end timestamps in the box and mark the IoU with the same colour as the models. Best read in colour." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.386, + 0.912, + 0.489 + ], + "angle": 0, + "content": "
VSDCTDEMQVHighlightCharades-STA
R1↑mAP↑Spurious R1↓Spurious mAP↓R1↑Spurious R1↓
@0.5@0.7@0.5@0.75Avg.@0.7@0.9@0.75Avg.@0.5@0.7@0.7@0.9
(a)61.1246.7762.4543.6642.549.355.299.9010.4057.3132.5525.726.31
(b)✓63.4749.3964.8247.6746.398.773.878.648.9139.1263.6723.155.42
(c)✓62.9348.2564.2245.4944.848.844.09.109.5638.5160.8024.035.73
(d)✓✓65.8853.6766.4349.8649.058.263.687.468.1560.8940.3522.134.82
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.499, + 0.908, + 0.542 + ], + "angle": 0, + "content": "Table 4. Ablation studies on QVHighlights val split and Charades-STA test split. VSDC and TDEM stand for Video Synthesizer for Dynamic Context and Temporal Dynamics Enhancement Module respectively. All the ablation results are averaged over three runs. We also reveal spurious metrics of the proposed modules." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.545, + 0.501, + 0.672 + ], + "angle": 0, + "content": "
MethodQVHighlights valCharades-STA test
R1@0.7mAP@0.75mAPR1@0.5R1@0.7
CG52.1045.7044.9058.4036.30
Ours w/ CG53.25+1.1549.16+3.4648.38+3.4859.35+0.9537.84+1.54
BAM51.6148.5647.6159.9539.38
Ours w/BAM52.87+1.2649.62+1.0648.82+1.2160.92+0.9740.25+0.87
QD46.6641.8241.2257.3132.55
Ours w/ QD53.67+7.0149.86+8.0449.00+7.7860.89+3.5840.35+7.80
SnAG48.1044.3742.7165.7237.32
Ours w/ SnAG52.93+4.8349.11+4.7446.75+4.0470.14+4.4242.35+5.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.683, + 0.483, + 0.738 + ], + "angle": 0, + "content": "Table 5. Generalization across different architectures on both QVHighlight val split and Charades-STA test split. CG stands for CG-DETR, while QD stands for QD-DETR and BAM stands for BAM-DETR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "method via SnAG achieves improvements of up to \\(10.04\\%\\) in R1@0.7 and \\(9.46\\%\\) in mAP on QVHighlights and \\(6.73\\%\\) in R1@0.5 and \\(13.48\\%\\) in R1@0.7 on Charades-STA. For DETR-like architecture, our approach enhances QD-DETR with significant gains of \\(19.23\\%\\) in R1@0.7 and \\(18.87\\%\\) in mAP on QVHighlights and \\(6.25\\%\\) in R1@0.5 and \\(23.96\\%\\) in R1@0.7 on Charades-STA. Our strategies significantly improve the performance of both hybrid and DETR-like baselines on both benchmarks, demonstrating the generalization of our approach." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.559, + 0.634, + 0.575 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Although existing transformer-based approaches have demonstrated remarkable performance, they still struggle with spurious correlation, which overly associates the textual query with the background frames rather than the target moment. To address this issue in moment retrieval, we introduce a novel dynamic learning method with two strategies. First, we propose a novel video synthesis strategy that constructs a dynamic context for the relevant moment. This synthesis strategy enables our model to attend to the target moment corresponding to the query across various dynamic video contexts. Second, we enhance the representation by learning temporal dynamics aligned with texts. In addition to visual features, text queries are aligned with dynamic representations, encouraging our model to establish a non-spurious correlation between the query-related moment and its context. Extensive experiments and detailed ablation studies on both the QVHighlights and Charades-STA benchmarks validate the effectiveness and generalization of the proposed TD-DETR, demonstrating significant alleviation of spurious correlation and superior performance in moment retrieval." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20988" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.192 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (No. 62176047, No. 62476051), the Sichuan Natural Science Foundation (No. 2024NSFTD0041), and the Sichuan Science and Technology Program (No. 2021YFS0374)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[1] Vedika Agarwal, Rakshith Shetty, and Mario Fritz. Towards causal vqa: Revealing and reducing spurious correlations by invariant and covariant semantic editing. In CVPR, pages 9690-9698, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.285, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[2] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.341, + 0.483, + 0.369 + ], + "angle": 0, + "content": "[3] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, pages 456-473, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.37, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.483, + 0.467 + ], + "angle": 0, + "content": "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.468, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[6] Victor Escorcia, Mattia Soldan, Josef Sivic, Bernard Ghanem, and Bryan Russell. Finding moments in video collections using natural language. arXiv preprint arXiv:1907.12763, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[7] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.566, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[8] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, pages 5267-5275, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[9] Runzhou Ge, Jiyang Gao, Kan Chen, and Ram Nevatia. Mac: Mining activity concepts for language-based temporal localization. In WACV, pages 245-253. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.482, + 0.717 + ], + "angle": 0, + "content": "[10] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[11] Soumya Suva Ghosal and Yixuan Li. Are vision transformers robust to spurious correlations? In IJCV, pages 689-709. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[12] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In AISTATS, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[13] Karan Goel, Albert Gu, Yixuan Li, and Christopher Ré. Model patching: Closing the subgroup performance gap with data augmentation. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[14] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In ICCV, pages 13846-13856, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] Minjoon Jung, Youwon Jang, Seongho Choi, Joochan Kim, Jin-Hwa Kim, and Byoung-Tak Zhang. Background-aware moment detection for video moment retrieval. In WACV, pages 8586-8596, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[16] Minjoon Jung, Junbin Xiao, Byoung-Tak Zhang, and Angela Yao. On the consistency of video large language models in temporal comprehension. In CVPR, pages 13713-13722, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[17] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In ECCV, pages 220-238. Springer, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[18] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In ECCV, pages 447-463. Springer, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[19] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In NeurIPS, pages 11846-11858, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[20] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. In NeurlPS, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[21] Xunsong Li, Pengzhan Sun, Yangcen Liu, Lixin Duan, and Wen Li. Simultaneous detection and interaction reasoning for object-centric action recognition. IEEE Transactions on Multimedia, pages 1-13, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[22] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[23] Meng Liu, Xiang Wang, Liqiang Nie, Xiangnan He, Baoquan Chen, and Tat-Seng Chua. Attentive moment retrieval in videos. In SIGIR, pages 15-24, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[24] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In ACM MM, pages 843-851, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[25] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, pages 3042-3051, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[26] Yangcen Liu, Ziyi Liu, Yuanhao Zhai, Wen Li, David Doerman, and Junsong Yuan. Stat: Towards generalizable temporal action localization. arXiv preprint arXiv:2404.13311, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[27] Ziyi Liu and Yangcen Liu. Bridge the gap: From weak to full supervision for temporal action localization with pseudoformer. In CVPR, pages 8711-8720, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[29] WonJun Moon, Sangeek Hyun, SuBeen Lee, and JaePil Heo. Correlation-guided query-dependency calibration for video temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20989" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[31] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. In CVPR, pages 18930-18940, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[32] Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. Interventional video grounding with dual contrastive learning. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[33] Mayu Otani, Yuta Nakashima, Esa Rahtu, and Janne Heikkilä. Uncovering hidden challenges in query-based video moment retrieval. In BMVC, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[35] Adrien Pavao, Isabelle Guyon, Anne-Catherine Letournel, Dinh-Tuan Tran, Xavier Baro, Hugo Jair Escalante, Sergio Escalera, Tyler Thomas, and Zhen Xu. Codalab competitions: An open source platform to organize scientific challenges. JMLR, 24(198):1-6, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[37] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. In Trans. Assoc. Comput. Linguistics, pages 25-36, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[38] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[39] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[40] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, pages 510-526. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[41] Lifu Tu, Garima Lalwani, Spandana Gella, and He He. An empirical study on robustness to spurious correlations using pre-trained language models. In ACL, pages 621-633, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[42] Yipei Wang and Xiaogian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In AISTATS, pages 3745-3753. PMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[43] Yipei Wang and Xiaoqian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In International Conference on Artificial Intelligence and Statistics, pages 3745-3753. PMLR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[44] Zhao Wang and Aron Culotta. Identifying spurious correlations for robust text classification. In EMNLP, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.888 + ], + "angle": 0, + "content": "[45] Fanyue Wei, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Learning pixel-level distinctions for video highlight detection. In CVPR, pages 3073-3082, 2022." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[46] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In CVPR, pages 18709-18719, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[47] Mengmeng Xu, Mattia Soldan, Jialin Gao, Shuming Liu, Juan-Manuel Pérez-Rúa, and Bernard Ghanem. Boundary-denoising for video activity localization. arXiv preprint arXiv:2304.02934, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[48] Yifang Xu, Yunzhuo Sun, Yang Li, Yilei Shi, Xiaoxiang Zhu, and Sidan Du. Mh-detr: Video moment and highlight detection with cross-modal transformer. In ACM MM, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[49] Jin Yang, Ping Wei, Huan Li, and Ziyang Ren. Task-driven exploration: Decoupling and inter-task feedback for joint moment retrieval and highlight detection. In CVPR, pages 18308-18318, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[50] Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos. NeurlPS, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[51] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In CVPR, pages 10287-10296, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[52] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In CVPR, pages 1247-1257, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[53] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In ACL, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[54] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Temporal sentence grounding in videos: A survey and future directions. In IEEE TPAMI, pages 10443-10465. IEEE, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[55] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. In ICML, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[56] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, pages 12870-12877, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[57] Xing Zhang, Zuxuan Wu, Zejia Weng, Huazhu Fu, Jingjing Chen, Yu-Gang Jiang, and Larry S Davis. Videolt: Large-scale long-tailed video recognition. In ICCV, pages 7960-7969, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[58] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. In IEEE TPAMI, pages 10795-10816. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.827 + ], + "angle": 0, + "content": "[59] Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In ICCV, pages 2914-2923, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20990" + } + ] +] \ No newline at end of file diff --git a/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/10bd97eb-949b-4ebe-ab75-25cc73578707_origin.pdf b/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/10bd97eb-949b-4ebe-ab75-25cc73578707_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..217d5e6d0d063deb7ed28fd77e95eb107db70a11 --- /dev/null +++ b/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/10bd97eb-949b-4ebe-ab75-25cc73578707_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:301d0366215954265cd0303eb8939c3b139f11b3399c62628f3ae9e46e9951f0 +size 1630954 diff --git a/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/full.md b/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c091eff5d39d8daf8c27c42c08eef44f9899b10c --- /dev/null +++ b/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/full.md @@ -0,0 +1,326 @@ +# The Devil is in the Spurious Correlations: Boosting Moment Retrieval with Dynamic Learning + +Xinyang Zhou $^{1\star}$ Fanyue Wei $^{2\star}$ Lixin Duan $^{1}$ Angela Yao $^{2}$ Wen Li $^{1\dagger}$ + +1University of Electronic Science and Technology of China 2National University of Singapore + +{x.y.zhou2020，lxduan，liwenbnu}@gmail.com + +{wfanyue,ayao}@comp.nus.edu.sg + +Query: "The blonde woman works out in a gym with red lighting." + +![](images/fb97a2924d77f0090dbfd6c50cb7ac69c98eb9d08d9f29618aa97a903884e18d.jpg) +Figure 1. (a) + +![](images/d7793ea090ed810e53aa8fbe8687d45d682ffc86acef3ba1311fe1d5a1fa501a.jpg) +Figure 1. (b) +Figure 1. (a) Comparison of moment retrieval models under normal and spurious correlation videos by masking the content of target clips of video. We found the existing works are suffering from a crucial reason stems from the spurious correlation between the text queries and the moment context. Baselines predict the Spurious GT even if the target moments are masked. In contrast, TD-DETR predicts the segment near the mask with lower confidence. (b) To verify the issue of spurious correlation, we introduce the Spurious mAP as the metric. Our model achieves the best ratio of mAP to Spurious mAP. + +# Abstract + +Given a textual query along with a corresponding video, the objective of moment retrieval aims to localize the moments relevant to the query within the video. While commendable results have been demonstrated by existing transformer-based approaches, predicting the accurate temporal span of the target moment is still a major challenge. This paper reveals that a crucial reason stems from the spurious correlation between the text query and the moment context. Namely, the model makes predictions by overly associating queries with background frames rather than distinguishing target moments. To address this issue, we propose a dynamic learning approach for moment retrieval, where two strategies are designed to mitigate the spurious correlation. First, we introduce a novel video synthesis approach to con + +struct a dynamic context for the queried moment, enabling the model to attend to the target moment of the corresponding query across dynamic backgrounds. Second, to alleviate the over-association with backgrounds, we enhance representations temporally by incorporating text-dynamics interaction, which encourages the model to align text with target moments through complementary dynamic representations. With the proposed method, our model significantly alleviates the spurious correlation issue in moment retrieval and establishes new state-of-the-art performance on two popular benchmarks, i.e., QVHighlights and Charades-STA. In addition, detailed ablation studies and evaluations across different architectures demonstrate the generalization and effectiveness of the proposed strategies. Our code is available at https://github.com/xyangzhou/TD-DETR. + +# 1. Introduction + +Videos are a ubiquitous media format, but it can be time-consuming to browse through videos to localize specific moments. Using text to retrieve corresponding moments within a lengthy video[2, 54] plays an important role for entertainment, search, etc. + +Existing moment retrieval approaches [15, 17, 19, 30, 49] leverage the DETR [4] detection transformer architecture to fuse text query into video representation. They show remarkable results in detecting the target moment by text-video alignment, but accurately predicting the temporal span remains a significant challenge. The learned video representations [17, 19] are aligned with the semantics of the text, thus generally biased to subjects or spatial objects. Consequently, the semantic elements in background frames are overly associated with textual queries. Such inappropriate association between irrelevant contexts and labels is the manifestation of Spurious correlation [1, 43], which are under-explored in moment retrieval tasks. + +This work identifies such spurious correlations between the text queries and moment visual context as the root cause for this performance gap in moment retrieval. Specifically, learned SOTA moment retrievers [17, 19, 30] tend to associate text queries with the background frames rather than distinguishing the target moment. Figure 1 illustrates how such spurious issues lead to a sub-optimal performance for moment retrieval. For example, the SOTA method, BAM-DETR [17], predicts a similar temporal span for an 'The blonde woman works out in a gym with red lighting.' in both the original video and an altered version where the target moment is masked out. In the masked case, the segment near the mask is a more reasonable output but the baseline methods [17, 29, 30, 49] predict the masked segment from background frames with the learned spurious correlation, which is the most unrelated output to the text query. + +To address the spurious issues, we propose a dynamic learning method with two novel strategies to mitigate spurious correlation. First, we propose a synthesis strategy that dynamically contextualizes the target moments for retrieval. We select similar video pairs and augment the samples by synthesizing target moments within similar contexts. This encourages the model to distinguish target moments from queries across more dynamic contexts, alleviating the over-association with background frames. Second, we propose to enhance video-text representations to reduce spurious dependencies on background frames by aligning text queries with temporal dynamics. To achieve this, we design a simple yet efficient temporal tokenizer to extract dynamic representations. Additionally, a text-dynamics interaction module is introduced to align dynamics with text queries, enabling the model to establish a stand-up correlation with complementary dynamic representations while + +mitigating background bias. + +Experiments on two challenging benchmarks, i.e., QVHighlights and Charades-STA, show that our method outperforms existing methods by clear margins. Besides, spurious correlation in moment retrieval is validated to be largely alleviated by spurious evaluation. The detailed ablation analyses demonstrate the effectiveness of the proposed strategies to resolve the spurious issues. + +In conclusion, our contributions to this work are summarized as follows: + +- To the best of our knowledge, we are the first to investigate the spurious correlation in moment retrieval. +- We propose a dynamic learning framework for moment retrieval that mitigates spurious correlations by dynamically contextualizing target moments through novel video synthesis and enhancing representations with text-dynamics alignment. +- The proposed method achieves SOTA performance across all benchmarks. Besides, our model also provides a strong interpretation of alleviating spurious correlations. + +# 2. Related Work + +# 2.1. Moment Retrieval + +Moment Retrieval aims to predict a temporal span within a video corresponding to a given natural language query [54]. Early methods had two stages: first, sampling candidate moments as proposals, and then scoring the proposals to obtain the final predictions [2, 8, 9, 23, 24]. However, more recent approaches are one-shot [18, 37, 47, 50-52, 56, 59] to predict the temporal span directly. Simultaneously, hybrid frameworks that integrate transformers for feature learning and fusion with CNNs for prediction have been widely studied. SnAG [31] explored the impact of cross-modal fusion on the scalability of video grounding models and proposed an efficient training scheme. + +Although many moment retrieval methods have achieved remarkable performance, the majority primarily focus on architectural design, such as exploring new networks [16, 21, 27, 50, 52], improving contrastive learning, enhancing video-text alignment[18, 26], or refining attention mechanisms[31]. Generally, these methods focus on how to fuse text and video representations to improve the temporal span prediction but the overall performance are limited by the inefficient alignment architecture. + +# 2.2. Detection Transformers + +In the past several years, the wide adoption of detection transformers (DETR) [4] in video moment retrieval has gained remarkable performance [14, 17, 19, 25, 30, 48, 49]. While introducing the DETR framework to the task of moment retrieval, it demonstrated the effectiveness on QVHighlights dataset [19]. Furthermore, DETR-based ap + +![](images/54e94479e0177e6462ac6ada01e629a369d91e600fdd8c54d9b3b3df030c9041.jpg) +Figure 2. Overview of the proposed TD-DETR. Video pairs are sampled by their similarity and then forwarded into the Video Synthesizer to generate Dynamic Context (Section 3.1). Then, the Dynamic Context, i.e., synthesized pairs, are enhanced by Temporal Dynamic Tokenization and interact with textual information (Section 3.2). Finally, the enhanced video tokens are sent to the transformer encoder-decoder with prediction heads to predict moments (Section 3.3). + +proaches have explored the relationship between moment retrieval and highlight detection, achieving significant advancements in both tasks. However, they still struggle with inaccurate span predictions [33, 54], limiting their overall reliability. To address this issue, Zhang et al. [54] analyzes the distribution of target moments and highlights the problem of imbalance, while Otani et al. [33] investigate the skewed distribution of query representations and expose the overly query-dependent phenomenon of existing models. Unlike previous studies, we reveal that spurious correlations is the key cause of inaccurate span predictions, arising from the over-association between the text query and the background frames. + +# 2.3. Spurious Correlation + +One reason for the poor generalization of vision algorithms is that they are prone to memorizing patterns or contextual cues [3, 10, 11, 13, 39, 41]. These patterns or cues are often spurious correlations - misleading heuristics of the training data correlated with the majority of examples but does not hold in general [11]. While numerous approaches have been proposed to address this challenge [33, 42, 44, 45, 54, 55, 57, 58] in various domains, their impact on video understanding remains largely unexplored. For example, in image-based tasks, spurious correlations are often tied to spatial biases [1, 43], whereas in video-related tasks, such spatial biases are redundant in frames. This distinction is particularly crucial in retrieving moments from lengthy videos, where the goal is not merely + +to recognize objects but to precisely localize temporal spans with text queries, making the study of spurious correlations in this context even more critical. + +Despite in Moment Retrieval, Otani et al. [33] investigated the biases introduced by text queries and the neglect of semantic information, while Zhang et al.[54] explores the distribution of start and end moments. However, the issue of spurious correlation between the text and background frames is a key challenge of moment retrieval. In this work, we focus on this point and develop a new method by learning dynamics to alleviate this problem. + +# 3. Methodology + +The objective of moment retrieval is to localize a temporal span that semantically corresponds to a given text query. Consider a video represented by $L$ vision tokens, denoted by $V = \{v_{1}, v_{2}, \ldots, v_{L}\}$ , along with a natural language description of $W$ words, as $\{q_{1}, q_{2}, \ldots, q_{W}\}$ . A moment retrieval model predicts a temporal span $m$ with centre $m_{c}$ and duration $m_{\sigma}$ which is relevant to the textual query, and token-wise saliency scores $\{s_{1}, s_{2}, \ldots, s_{L}\}$ of the text. + +As shown in Figure 2, we propose a novel framework to learn Temporal Dynamics utilizing DEtection Transformer (TD-DETR). The video pairs are sampled by similarity and then forwarded into the video synthesizer to generate dynamic contexts. Then, the representations of synthesized pairs are enhanced through temporal dynamic interaction with texts. Finally, dynamic-enhanced representations are + +fed into the transformer encoder-decoder with prediction heads to predict target moments. + +# 3.1. Video Synthesizer for Dynamic Context + +We introduce a novel video synthesizer to infuse the target moments with dynamic context for moment retrieval. Spurious correlations often stem from linking the moment's context to the text query. We address the issue by synthesizing new samples for the target moments with more dynamic contextual variations. + +Spurious Pair Selection. In order to synthesize a dynamic context aligned with the target moment, we construct a spurious pair of the target moment in a given video $V$ . To ensure the challenge and rationality of the synthesized video, we select a video which is similar to $V$ contextually. + +In a training batch of $N$ video samples, we sample the most similar video $\{V_k\}$ for every $\{V_i\}$ where $k \in [1, N], k \neq i$ , for each $i \in [1, N]$ to construct a spurious video pair $p_i$ . For any videos $V_j$ and $V_l$ from a batch, we employ cosine similarity to model the similarity relationship between videos as follows: + +$$ +s _ {j, l} = \frac {1}{L _ {j} \cdot L _ {l}} \sum_ {p = 0} ^ {L _ {j}} \sum_ {q = 0, p \neq q} ^ {L _ {l}} \frac {\left. v _ {p} ^ {j} ^ {T} \cdot v _ {q} ^ {l} \right|}{\left\| v _ {p} ^ {j} \right\| \cdot \left\| v _ {q} ^ {l} \right\|} \quad j, l \in [ 1, N ]. \tag {1} +$$ + +In Eq. 1, we denote $v_{q}^{j}$ and $v_{p}^{l}$ q-th clip of $V_{j}$ and q-th clip of $V_{l}$ , $L_{j}$ and $L_{l}$ for the length of $V_{j}$ and $V_{l}$ , $a_{j,l}$ for the average similarity between $V_{j}$ and $V_{l}$ . We select the most similar video $V_{k}$ for every video $V_{i}$ by $k = \arg \max_{k\in [0,N]}s_{[i,:\backslash \{i\}]}$ in the batch. + +Video Synthesis with Dynamic Context. After obtaining the spurious pairs, we synthesize a new video $\tilde{V}_i$ using the target moment $m$ and the spurious pair $p_i$ , where $p_i = \{V_i, V_k\}$ . We then compose $\{V_i, V_k\}$ and dynamically refine the ground truth of the target moment $m_i$ . Illustrated by Figure 3, given video $V_i$ with $L_i$ tokens and $V_k$ with $L_k$ tokens, on the one hand, we expect the sample from $V_i$ and $V_k$ with the completeness of target moment $m_i$ of $V_i$ . We first sample from $V_i$ without ground truth tokens $\mathrm{NG} \in [0, L_i] \setminus m_i$ with sampling ratio $\alpha$ , i.e., every token in NG has the same ratio $\alpha$ to be selected. We sample from the paired $V_k$ with sampling ratio $1 - \alpha$ , since $V_k$ is irrelevant to ground truth $\mathrm{GT}_i$ . + +Finally, we concatenate the two sets of sampled video tokens into $\tilde{V}_i$ . On the other hand, we have the same process, except now we focus on $V_k$ and $\mathrm{GT}_k$ with sampling ratio $\alpha$ and ratio $1 - \alpha$ to sample from the whole $V_i$ and concatenate sampled tokens into $\tilde{V}_k$ . + +# 3.2. Dynamics Enhancement + +The attention of DETR-like architecture associates text queries with visual representations. Unfortunately, this + +![](images/adce30a8d306c1003640b9a965a0983caf13509e2bf91c4bb0f6b506296bd049.jpg) +Figure 3. Illustration of Video Synthesis with Dynamic Context. The numbers in the boxes represent the token indices, indicating their sequential order. Vision tokens are selected based on their selection probabilities and concatenated while preserving the completeness of the ground truth spans, i.e., selection probabilities of ground truth spans are set to 1. The synthesized videos maintain their original length with a random bias. + +![](images/0dd1edb110489cf5ad2162d8f83461b9c7e67c937801e41db34be8336affac99.jpg) + +tends to emphasize background frames, thereby exacerbating the spurious correlation. To this end, besides the dynamic context synthesis, we introduce a dynamics enhancement module to encourage our model to align text queries with temporal dynamic representations. Thus, the model considers not only the association of background frames but also background-independent dynamics. This alignment enables our model to establish a stand-up correlation between the query-related moment and its context. + +Dynamic Tokenizer. To model the dynamic temporality, we introduce a simple yet effective strategy to tokenize the temporal dynamics, which is nearly cost-free to obtain temporal dynamic representations. + +Given spurious pair $p_i = \{\tilde{V}_i, \tilde{V}_k\}$ , the temporal dynamic tokenizer processes all video pairs in the same manner. Then, we use $\tilde{V}_i$ to explain the processing steps without loss of generality. + +First, we concatenate a learnable token to the start of the video as follows: + +$$ +\tilde {V} _ {i} = \left\{s t, \tilde {v} _ {1}, \tilde {v} _ {2}, \dots , \tilde {v} _ {L} \right\}, \tag {2} +$$ + +where $st$ is a learnable token that captures the start signal of the video, preventing the loss of information from the first clip in the video representations. We then employ an element-wise difference to model the temporal dynamics. + +$$ +T = \left\{\tilde {v} _ {1} - s t, \tilde {v} _ {2} - \tilde {v} _ {1}, \tilde {v} _ {3} - \tilde {v} _ {2}, \dots , \tilde {v} _ {L _ {i}} - \tilde {v} _ {L _ {i} - 1} \right\}, \tag {3} +$$ + +where $T$ is a learnable dynamic representation, it learns and focuses on the dynamic information in nearby video clips. + +Text-Dynamics Interaction. After obtaining dynamic tokens $T$ in Eq. 3 from Dynamic Tokenizer, we utilize cross-attention for both text-vision and text-dynamic representations with the same text query to mitigate the over-emphasis on video backgrounds. + +In detail, the query for each cross-attention layer is prepared by linear projection of the video and temporal clips as follows: + +$$ +Q _ {\delta} = \left[ p _ {\delta} (\delta), \dots , p _ {\delta} \left(\delta_ {L _ {i}}\right) \right], \delta \in \{\tilde {V} _ {i}, T \}. \tag {4} +$$ + +The key and value are computed with the query text representation as $K_{q} = [p_{k}(q_{1}),\dots,p_{k}(q_{W})]$ and $V_{q} = [p_{v}(q_{1}),\dots,p_{v}(q_{L})]$ , $p_{\delta}, p_{k}, p_{v}$ are linear projection layers for dynamic-enhanced query, temporal query, key and value. Then the cross-attention layer operates as follows: + +$$ +\delta^ {\prime} = \operatorname {s o f t m a x} \left(\frac {Q _ {\delta} K _ {q} ^ {T}}{\sqrt {d}}\right) V _ {q}, \delta \in \left\{\tilde {V} _ {i}, T \right\}, \delta^ {\prime} \in \left\{\tilde {V} _ {i} ^ {\prime}, T ^ {\prime} \right\}, \tag {5} +$$ + +where $d$ is the dimension of the projected key, value and query. To emphasize the learned temporal information, we utilize the weighted element-wise addition to inject text-guided temporal representation into text-guided dynamic-enhanced representation as follows: + +$$ +\tilde {V} _ {i} ^ {\prime} = \beta \cdot \tilde {V} _ {i} + (1 - \beta) \cdot T ^ {\prime}, \tag {6} +$$ + +where $\beta$ is a hyper-parameter to adapt the addition ratio between $\tilde{V}_i'$ and $T'$ . Hence, from the spurious pair $p_i = \{\tilde{V}_i, \tilde{V}_k\}$ , we obtain a new spurious pair $p_i' = \{\tilde{V}_i', \tilde{V}_k'\}$ . + +# 3.3. Network and Objectives + +Our network structure (see Figure 2) follows previous work [19, 30]. It features the synthesis of dynamic contextual videos and the enhancement of dynamic representations. + +Transformer encoder-decoder with prediction heads. Given spurious pair $p_i' = \{\tilde{V}_i', \tilde{V}_k'\}$ , this module processes all video pairs in the same manner. Therefore, we use $\tilde{V}_i'$ to explain the processing steps without loss of generality. The encoder consists of $T$ stacked typical transformer encoder layers, as in [19, 30], producing encoded representations $E_{enc}$ . Our decoder, following [30], also uses $T$ stacked typical transformer decoder layers, along with $N$ learnable moment queries representing the centre $m_c$ and duration $m_\sigma$ . The decoder processes $E_{enc}$ with the moment queries. + +We adopt prediction heads following [30]. A linear layer predicts saliency scores from the encoded representations, and another linear layer handles negative pairs. From the decoder outputs, a 3-layer MLP with ReLU predicts the normalized moment center and duration, while a linear layer with softmax predicts the foreground. + +Hungarian matching. Following [4, 19, 30], we perform Hungarian matching between the two predictions and two corresponding labels respectively. Give prediction $\hat{y}$ of $\tilde{V}_i'$ and corresponding to ground truth $y$ , the optimal matching results between predictions and ground truths $\hat{\sigma}$ can be written as $\hat{\sigma} = \arg \min_{\sigma \in G_N} \sum_i^N \mathcal{C}_{\mathrm{match}}(y, \hat{y}_{\sigma(i)})$ , where $G$ is a permutation of predictions and ground truths pairs and $\mathcal{C}_{\mathrm{match}}$ is the matching cost. + +Loss Functions. We calculate the loss between $\hat{\sigma}$ and the ground truth $y$ corresponding to $\tilde{V}_i'$ . Following [19], the $L_1$ loss $\mathcal{L}_{L_1}$ and the gIoU [38] loss $\mathcal{L}_{gIoU}$ are used to measure the distance and overlapping between the predictions and a cross-entropy loss $\mathcal{L}_{cls}$ is used to measure classification. For highlight detection, we also use three loss functions which are margin ranking loss $\mathcal{L}_{margin}$ , rank-aware contrastive loss $\mathcal{L}_{margin}$ and negative loss $\mathcal{L}_{neg}$ . + +The overall loss is shown as follows with $\lambda_{*}$ as balancing coefficient: + +$$ +\mathcal {L} _ {\text {m o m e n t}} = \lambda_ {L _ {1}} \mathcal {L} _ {L _ {1}} + \lambda_ {i o u} \mathcal {L} _ {g I o U} + \lambda_ {c l s} \mathcal {L} _ {c l s}, \tag {7} +$$ + +$$ +\mathcal {L} _ {h l} = \lambda_ {\text {m a r g i n}} \mathcal {L} _ {\text {m a r g i n}} + \lambda_ {\text {c o n t}} \mathcal {L} _ {\text {c o n t}} + \lambda_ {\text {n e g}} \mathcal {L} _ {\text {n e g}}, \tag {8} +$$ + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {h l} + \mathcal {L} _ {\text {m o m e n t}}. \tag {9} +$$ + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We evaluate our method on two widely-used benchmarks: QVHighlights [19] and Charades-STA [8], following the setup of prior works [19, 30]. QVHighlights has over 10,000 video-query pairs primarily from vlog and news content. We use the splits defined in MomentDETR [19] and report results for both the val and test splits. Charades-STA, derived from the Charades dataset [40], has more than 18,000 video-sentence pairs. We follow the standard evaluation protocol from [8] and use two separate training and testing splits. + +Standard Evaluation. We adopt the same metrics of Moment-DETR [17, 19, 30] for evaluation. Specifically, for Moment Retrieval, we report the mean average precision (mAP) at Intersection over Union (IoU) thresholds of 0.5 and 0.75, as well as the average mAP over IoU thresholds ranging from 0.5 to 0.95 with a step size of 0.05. We also include Recall@1 (R@1) at IoU thresholds of 0.5 and 0.75. For a fair comparison, we also include the test split of the QVHighlights dataset, which is evaluated on the CodaLab competition platform [35]. Notably, performance at higher IoU thresholds, such as 0.7, serves as an indicator of more precise alignments between predicted moments and the ground truth. + +Spurious Evaluation. To verify the spurious correlation with overly association the text with background, we replace the target moments in video content with random-valued masks while keeping the video duration unchanged, + +
MethodMR-R1MR-mAP
@0.5@0.7@0.5@0.75Average
MCN [2] ICCV'1711.412.7224.948.2210.67
CAL [6] arXiv'1925.4911.5423.407.659.89
XML [18] ECCV'2041.8330.3544.6331.7332.14
XML+ [19] NIPS'2146.6933.4647.8934.6734.90
SnAG† [31] CVPR'2459.7948.1058.6344.3742.71
SnAG /w TD-DETR66.4852.9363.7149.1146.75
Moment-DETR [19] NIPS'2152.8933.0254.8229.4030.73
UMT [25] CVPR'2256.2341.1853.8337.0136.12
MomentDiff [20] NIPS'2357.4239.6654.0235.7335.95
QD-DETR [30] CVPR'2362.4044.9862.5239.8839.86
UniVTG [22] ICCV'2358.8640.8657.6035.5935.47
CG-DETR [29] arXiv'2365.4048.4064.5042.8042.90
UVCOM [46] CVPR'2463.5548.7064.4744.0143.27
BAM-DETR [17] ECCV'2464.5348.6464.5746.3345.36
TD-DETR (Ours)64.53±0.6250.37±0.5366.21±0.2147.32±0.5346.69±0.26
+ +$\dagger$ reproduced by the official code +Table 1. Performance comparison on QVHighlights test split. Our experimental results are averaged over three runs and $\pm$ denotes the standard deviation. For the compared methods, the results are copied from their original papers and we reproduce SnAG by the official code via slowfast+clip features. + +
MethodR1@0.5R1@0.7
CAL [6]44.9024.37
2D TAN [56]39.7023.31
VSLNet [53]47.3130.19
IVG-DCL [32]50.2432.88
SnAG†[31]65.7237.32
SnAG /w TD-DETR70.1442.35
Moment-DETR [19]53.6331.37
Moment-Diff [20]55.5732.42
UMT [25]48.3129.25
QD-DETR [30]57.3132.55
CG-DETR [29]58.4036.30
BAM-DETR [17]59.9539.38
TD-DETR (Ours)60.8940.35
+ +$\dagger$ reproduced by the official code +Table 2. Performance comparison on Charades-STA test split. Our experimental results are averaged over three runs. For all the compared methods, the results are taken from their original papers. + +making the masked region the most unrelated moments to text queries. We then assess the model's performance under these conditions and report the spurious metric of Spurious $R@1$ and Spurious mAP, analogous to the standard R@1 and mAP. In this case, lower values are better, as they indicate less spurious correlation. + +Implementation Details. Our method is implemented in PyTorch [34]. We follow the implementation of QD-DETR [30]. For all datasets, we use video features both extracted from SlowFast [7] pre-trained on Kinetics [5] and + +
MethodSpurious R1 ↓Spurious mAP ↓Standard mAP ↑
@0.7@0.9@0.75Avg.@0.75Avg.
QD-DETR9.355.299.9010.4041.8241.22
Ours w/ QD8.263.687.468.1549.8649.05
CG-DETR4.651.295.556.1445.7044.90
Ours w/ CG2.580.393.384.4149.1648.38
BAM-DETR7.161.876.306.7248.5647.61
Ours w/ BAM1.610.521.731.9849.6248.67
+ +Table 3. Performance comparison on QVHighlights val split with random masks. The target clips in the video content are replaced with random-valued masks while keeping the video duration unchanged. We compare the baseline models—QD-DETR, CG-DETR, and BAM-DETR—against their respective versions integrated with our proposed TD-DETR, marked Ours w/, under the same experimental settings. The metrics here have been explained in section 4.1. + +pre-trained CLIP [36] vision encoder, and text feature extracted from pre-trained CLIP [36] text encoder, following Moment-DETR. + +- For QVHighlights dataset, we set batch size to 32 and an initial learning rate of $1e - 4$ with weight decay of $1e - 4$ . We set the hidden size $d = 256$ , layers of encoder/decoder $T = 3$ , and moment queries $N = 10$ . The model is trained for 200 epochs. + +- For Charades-STA, we set batch size to 32 and use an initial learning rate of $1e - 4$ with weight decay of $1e - 4$ . We set the hidden size $d = 256$ , layers of encoder/decoder $T = 3$ , and moment queries $N = 10$ . The model is trained for 100 epochs and the learning rate is decayed to + +$\frac{1}{10}$ every 40 epochs. + +We use mini-batch and AdamW [28] gradient descent algorithm to optimize the network parameters, initialize weights with Xavier init [12], use dropout of 0.1 for transformer layers and 0.5 for input projection layers and set cost coefficient as the same as loss balancing coefficient on all datasets. For Dynamic Video Synthesizer, the selection Probability $\alpha$ is set to 0.5 and for Dynamic Temporal Identifier the $\beta$ is set to 0.7 on all datasets. The total training time is approximately 8 hours on QVHighlight and 6 hours for Charades-STA, with a single NVIDIA RTX 2080 GPU. + +# 4.2. Results on Standard Evaluation + +As illustrated in Table 1 and Table 2, we compare $TD$ -DETR with several baseline methods on QVHighlights and Charades-STA benchmarks. + +Results on QVHighlights. On QVHighlights test split, we compare our model against proposal-based, proposal-free, and DETR-based methods, as shown in Table 1. As observed, our TD-DETR achieves state-of-the-art performance across all evaluation metrics. Notably, TD-DETR shows remarkable improvements compared to the other methods. Performance at higher IoU thresholds (e.g., 0.7) provides a clearer indication of the alignment between predicted moments and the ground truth. Our TD-DETR outperforms the previous state-of-the-art model by a substantial margin, with improvements of up to $3.56\%$ in $\mathbf{R}@\mathbf{1}@\mathbf{0.7}$ and $2.14\%$ in $\mathrm{mAP}@0.75$ . + +Results on Charades-STA. We evaluate our method methods on Charades-STA test split, compared with proposal-based, proposal-free and DETR-based methods. As shown in Table 2, TD-DETR achieves state-of-the-art performance across all evaluation metrics. Notably, our method achieves improvements of up to $2.46\%$ in R1@0.7 and $1.57\%$ in R1@0.5 on this challenging benchmark. Recall that performance at high IoU thresholds stands for the precision with which predictions match the ground truth data. + +# 4.3. Results on Spurious Correlation Evaluation + +Moment retrieval models are often prone to inaccurate predictions due to spurious correlations, a typical example being when predictions rely overly on background cues. To investigate this issue, we adopt the spurious evaluation setting pre-defined in Sec 4.1. + +We evaluate our methods on several strong baselines, including QD-DETR [30], CG-DETR [29] and BAM-DETR [17]. The results in Table 3 demonstrate the alleviation of spurious correlation issue, as indicated by the improved Spurious $R1$ and Spurious $mAP$ across all compared baselines while boosting the standard $mAP$ . For instance, our methods improve QD-DETR [30] in Spurious $R1@0.9$ from 10.40 to 8.15 (30.43%), with standard $mAP@0.75$ increasing from 41.22 to 49.05 (18.99%). While for BAM + +DETR [17], it achieves the most substantial improvement, improving Spurious R1 to 1.61 and Spurious mAP to 1.73. The results on spurious correlation evaluation demonstrate the effectiveness of our proposed approaches to alleviate the spurious correlation issue and also the generalization across different baselines. + +# 4.4. Qualitative Analysis + +In this subsection, we analyze the impact of mitigating spurious correlations in our TD-DETR model. As illustrated in Figure 4, the entire video depicts a man speaking to the camera, while the target moment is described as: "A man wearing a cap backwards talking while some other videos appear in the bottom left corner of the screen." Notably, baseline models fail to distinguish between "The man is talking" and "The man is talking while other videos appear in the corner," indicating that background elements are overly correlated with the textual query. In contrast, our TD-DETR accurately predicts the target moment without being misled by background frames, demonstrating its ability to associate the information between text and the proper frames rather than relying on spurious correlations. + +# 4.5. Ablation Studies + +We conduct ablation studies on both QVHighlights and Charades-STA benchmarks to validate the effectiveness of each proposed component. VSDC and TDEM refer to Video Synthesizer for Dynamic Context and Temporal Dynamics Enhancement Module, respectively. As illustrated in Table 4, rows (b) to (c) show the effectiveness of each component compared to the baseline model (a), and (d) demonstrate the overall effectiveness of all the components. In detail, on QVHighlight, Video Synthesizer for Dynamic Context contributes improvement of $5.60\%$ in R1@0.7, $9.18\%$ in mAP@0.75 and $9.05\%$ in average mAP while Temporal Dynamics Enhancement Module contributes improvement of $3.16\%$ in R1@0.7, $4.19\%$ in mAP@0.75 and $5.41\%$ in average mAP for Moment Retrieval. With all components integrated, we observe a substantial $14.75\%$ , $14.20\%$ and $15.30\%$ improvement in the R1@0.7, mAP@0.75 and average mAP, respectively. Moreover, the spurious correlation is also improved, where $30.43\%$ in Spurious R1@0.9 and $21.63\%$ in Spurious mAP. The same phenomenon can be observed in Charades-STA as well. More experiments and analyses are included in the supplementary materials. + +# 4.6. Generalization across Different Architectures. + +We validate the generalization by incorporating our proposed approaches into different architectures. SnAG [31] is a hybrid framework, which integrates transformer for feature fusion and CNN for prediction, while QD-DETR [30] and BAM-DETR [17] follow DETR-like architecture. + +As illustrated in Table 5, for the hybrid architecture, our + +![](images/b57c6f4367b9490e4acaa37292ac57fab2f77f4c8a507699f08d88d293347a15.jpg) +Figure 4. Visualization of results comparison between the baseline models and our TD-DETR. We display the relative position of ground truth and prediction moments in colourful boxes, where the green with GT stands for ground truth moment, the orange box with QD-DETR stands for our baseline predictions and the purple box with ours stands for our TD-DETR. To quantify the quality of predictions, we mark the start and end timestamps in the box and mark the IoU with the same colour as the models. Best read in colour. + +
VSDCTDEMQVHighlightCharades-STA
R1↑mAP↑Spurious R1↓Spurious mAP↓R1↑Spurious R1↓
@0.5@0.7@0.5@0.75Avg.@0.7@0.9@0.75Avg.@0.5@0.7@0.7@0.9
(a)61.1246.7762.4543.6642.549.355.299.9010.4057.3132.5525.726.31
(b)✓63.4749.3964.8247.6746.398.773.878.648.9139.1263.6723.155.42
(c)✓62.9348.2564.2245.4944.848.844.09.109.5638.5160.8024.035.73
(d)✓✓65.8853.6766.4349.8649.058.263.687.468.1560.8940.3522.134.82
+ +Table 4. Ablation studies on QVHighlights val split and Charades-STA test split. VSDC and TDEM stand for Video Synthesizer for Dynamic Context and Temporal Dynamics Enhancement Module respectively. All the ablation results are averaged over three runs. We also reveal spurious metrics of the proposed modules. + +
MethodQVHighlights valCharades-STA test
R1@0.7mAP@0.75mAPR1@0.5R1@0.7
CG52.1045.7044.9058.4036.30
Ours w/ CG53.25+1.1549.16+3.4648.38+3.4859.35+0.9537.84+1.54
BAM51.6148.5647.6159.9539.38
Ours w/BAM52.87+1.2649.62+1.0648.82+1.2160.92+0.9740.25+0.87
QD46.6641.8241.2257.3132.55
Ours w/ QD53.67+7.0149.86+8.0449.00+7.7860.89+3.5840.35+7.80
SnAG48.1044.3742.7165.7237.32
Ours w/ SnAG52.93+4.8349.11+4.7446.75+4.0470.14+4.4242.35+5.03
+ +Table 5. Generalization across different architectures on both QVHighlight val split and Charades-STA test split. CG stands for CG-DETR, while QD stands for QD-DETR and BAM stands for BAM-DETR. + +method via SnAG achieves improvements of up to $10.04\%$ in R1@0.7 and $9.46\%$ in mAP on QVHighlights and $6.73\%$ in R1@0.5 and $13.48\%$ in R1@0.7 on Charades-STA. For DETR-like architecture, our approach enhances QD-DETR with significant gains of $19.23\%$ in R1@0.7 and $18.87\%$ in mAP on QVHighlights and $6.25\%$ in R1@0.5 and $23.96\%$ in R1@0.7 on Charades-STA. Our strategies significantly improve the performance of both hybrid and DETR-like baselines on both benchmarks, demonstrating the generalization of our approach. + +# 5. Conclusion + +Although existing transformer-based approaches have demonstrated remarkable performance, they still struggle with spurious correlation, which overly associates the textual query with the background frames rather than the target moment. To address this issue in moment retrieval, we introduce a novel dynamic learning method with two strategies. First, we propose a novel video synthesis strategy that constructs a dynamic context for the relevant moment. This synthesis strategy enables our model to attend to the target moment corresponding to the query across various dynamic video contexts. Second, we enhance the representation by learning temporal dynamics aligned with texts. In addition to visual features, text queries are aligned with dynamic representations, encouraging our model to establish a non-spurious correlation between the query-related moment and its context. Extensive experiments and detailed ablation studies on both the QVHighlights and Charades-STA benchmarks validate the effectiveness and generalization of the proposed TD-DETR, demonstrating significant alleviation of spurious correlation and superior performance in moment retrieval. + +# Acknowledgement + +This work is supported by the National Natural Science Foundation of China (No. 62176047, No. 62476051), the Sichuan Natural Science Foundation (No. 2024NSFTD0041), and the Sichuan Science and Technology Program (No. 2021YFS0374). + +# References + +[1] Vedika Agarwal, Rakshith Shetty, and Mario Fritz. Towards causal vqa: Revealing and reducing spurious correlations by invariant and covariant semantic editing. In CVPR, pages 9690-9698, 2020. 2, 3 +[2] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 2, 6 +[3] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, pages 456-473, 2018. 3 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 2, 5 +[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 6 +[6] Victor Escorcia, Mattia Soldan, Josef Sivic, Bernard Ghanem, and Bryan Russell. Finding moments in video collections using natural language. arXiv preprint arXiv:1907.12763, 2019. 6 +[7] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 6 +[8] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, pages 5267-5275, 2017. 2, 5 +[9] Runzhou Ge, Jiyang Gao, Kan Chen, and Ram Nevatia. Mac: Mining activity concepts for language-based temporal localization. In WACV, pages 245-253. IEEE, 2019. 2 +[10] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. 3 +[11] Soumya Suva Ghosal and Yixuan Li. Are vision transformers robust to spurious correlations? In IJCV, pages 689-709. Springer, 2024. 3 +[12] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In AISTATS, 2010. 7 +[13] Karan Goel, Albert Gu, Yixuan Li, and Christopher Ré. Model patching: Closing the subgroup performance gap with data augmentation. In ICLR, 2021. 3 +[14] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In ICCV, pages 13846-13856, 2023. 2 + +[15] Minjoon Jung, Youwon Jang, Seongho Choi, Joochan Kim, Jin-Hwa Kim, and Byoung-Tak Zhang. Background-aware moment detection for video moment retrieval. In WACV, pages 8586-8596, 2025. 2 +[16] Minjoon Jung, Junbin Xiao, Byoung-Tak Zhang, and Angela Yao. On the consistency of video large language models in temporal comprehension. In CVPR, pages 13713-13722, 2025. 2 +[17] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In ECCV, pages 220-238. Springer, 2024. 2, 5, 6, 7 +[18] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In ECCV, pages 447-463. Springer, 2020. 2, 6 +[19] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In NeurIPS, pages 11846-11858, 2021. 2, 5, 6 +[20] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. In NeurlPS, 2023. 6 +[21] Xunsong Li, Pengzhan Sun, Yangcen Liu, Lixin Duan, and Wen Li. Simultaneous detection and interaction reasoning for object-centric action recognition. IEEE Transactions on Multimedia, pages 1-13, 2025. 2 +[22] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 6 +[23] Meng Liu, Xiang Wang, Liqiang Nie, Xiangnan He, Baoquan Chen, and Tat-Seng Chua. Attentive moment retrieval in videos. In SIGIR, pages 15-24, 2018. 2 +[24] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In ACM MM, pages 843-851, 2018. 2 +[25] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, pages 3042-3051, 2022. 2, 6 +[26] Yangcen Liu, Ziyi Liu, Yuanhao Zhai, Wen Li, David Doerman, and Junsong Yuan. Stat: Towards generalizable temporal action localization. arXiv preprint arXiv:2404.13311, 2024. 2 +[27] Ziyi Liu and Yangcen Liu. Bridge the gap: From weak to full supervision for temporal action localization with pseudoformer. In CVPR, pages 8711-8720, 2025. 2 +[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7 +[29] WonJun Moon, Sangeek Hyun, SuBeen Lee, and JaePil Heo. Correlation-guided query-dependency calibration for video temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 2, 6, 7 +[30] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 2, 5, 6, 7 + +[31] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. In CVPR, pages 18930-18940, 2024. 2, 6, 7 +[32] Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. Interventional video grounding with dual contrastive learning. In CVPR, 2021. 6 +[33] Mayu Otani, Yuta Nakashima, Esa Rahtu, and Janne Heikkilä. Uncovering hidden challenges in query-based video moment retrieval. In BMVC, 2020. 3 +[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6 +[35] Adrien Pavao, Isabelle Guyon, Anne-Catherine Letournel, Dinh-Tuan Tran, Xavier Baro, Hugo Jair Escalante, Sergio Escalera, Tyler Thomas, and Zhen Xu. Codalab competitions: An open source platform to organize scientific challenges. JMLR, 24(198):1-6, 2023. 5 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 6 +[37] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. In Trans. Assoc. Comput. Linguistics, pages 25-36, 2013. 2 +[38] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5 +[39] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 3 +[40] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, pages 510-526. Springer, 2016. 5 +[41] Lifu Tu, Garima Lalwani, Spandana Gella, and He He. An empirical study on robustness to spurious correlations using pre-trained language models. In ACL, pages 621-633, 2020. 3 +[42] Yipei Wang and Xiaogian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In AISTATS, pages 3745-3753. PMLR, 2024. 3 +[43] Yipei Wang and Xiaoqian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In International Conference on Artificial Intelligence and Statistics, pages 3745-3753. PMLR, 2024. 2, 3 +[44] Zhao Wang and Aron Culotta. Identifying spurious correlations for robust text classification. In EMNLP, 2020. 3 +[45] Fanyue Wei, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Learning pixel-level distinctions for video highlight detection. In CVPR, pages 3073-3082, 2022. + +[46] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In CVPR, pages 18709-18719, 2024. 6 +[47] Mengmeng Xu, Mattia Soldan, Jialin Gao, Shuming Liu, Juan-Manuel Pérez-Rúa, and Bernard Ghanem. Boundary-denoising for video activity localization. arXiv preprint arXiv:2304.02934, 2023. 2 +[48] Yifang Xu, Yunzhuo Sun, Yang Li, Yilei Shi, Xiaoxiang Zhu, and Sidan Du. Mh-detr: Video moment and highlight detection with cross-modal transformer. In ACM MM, 2023. 2 +[49] Jin Yang, Ping Wei, Huan Li, and Ziyang Ren. Task-driven exploration: Decoupling and inter-task feedback for joint moment retrieval and highlight detection. In CVPR, pages 18308-18318, 2024. 2 +[50] Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos. NeurlPS, 32, 2019. 2 +[51] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In CVPR, pages 10287-10296, 2020. +[52] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In CVPR, pages 1247-1257, 2019. 2 +[53] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In ACL, 2020. 6 +[54] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Temporal sentence grounding in videos: A survey and future directions. In IEEE TPAMI, pages 10443-10465. IEEE, 2023. 2, 3 +[55] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. In ICML, 2022. 3 +[56] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, pages 12870-12877, 2020. 2, 6 +[57] Xing Zhang, Zuxuan Wu, Zejia Weng, Huazhu Fu, Jingjing Chen, Yu-Gang Jiang, and Larry S Davis. Videolt: Large-scale long-tailed video recognition. In ICCV, pages 7960-7969, 2021. 3 +[58] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. In IEEE TPAMI, pages 10795-10816. IEEE, 2023. 3 +[59] Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In ICCV, pages 2914-2923, 2017. 2 \ No newline at end of file diff --git a/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/images.zip b/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2e89704ee074b2655e8a1161b4cb7a5a3a5e2a21 --- /dev/null +++ b/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63e9fde182918566a8b75901ab9e4646284463c4cce500cafd6b5c3863a4180a +size 653260 diff --git a/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/layout.json b/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d35669c370ed7a1be2948a782ff92836be2b0747 --- /dev/null +++ b/2025/The Devil is in the Spurious Correlations_ Boosting Moment Retrieval with Dynamic Learning/layout.json @@ -0,0 +1,10025 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 144, + 103, + 465, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 103, + 465, + 140 + ], + "spans": [ + { + "bbox": [ + 144, + 103, + 465, + 140 + ], + "type": "text", + "content": "The Devil is in the Spurious Correlations: Boosting Moment Retrieval with Dynamic Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 167, + 488, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 167, + 488, + 182 + ], + "spans": [ + { + "bbox": [ + 115, + 167, + 488, + 182 + ], + "type": "text", + "content": "Xinyang Zhou" + }, + { + "bbox": [ + 115, + 167, + 488, + 182 + ], + "type": "inline_equation", + "content": "^{1\\star}" + }, + { + "bbox": [ + 115, + 167, + 488, + 182 + ], + "type": "text", + "content": " Fanyue Wei" + }, + { + "bbox": [ + 115, + 167, + 488, + 182 + ], + "type": "inline_equation", + "content": "^{2\\star}" + }, + { + "bbox": [ + 115, + 167, + 488, + 182 + ], + "type": "text", + "content": " Lixin Duan" + }, + { + "bbox": [ + 115, + 167, + 488, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 167, + 488, + 182 + ], + "type": "text", + "content": " Angela Yao" + }, + { + "bbox": [ + 115, + 167, + 488, + 182 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 167, + 488, + 182 + ], + "type": "text", + "content": " Wen Li" + }, + { + "bbox": [ + 115, + 167, + 488, + 182 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 185, + 499, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 185, + 499, + 199 + ], + "spans": [ + { + "bbox": [ + 111, + 185, + 499, + 199 + ], + "type": "text", + "content": "1University of Electronic Science and Technology of China 2National University of Singapore" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 201, + 418, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 201, + 418, + 213 + ], + "spans": [ + { + "bbox": [ + 192, + 201, + 418, + 213 + ], + "type": "text", + "content": "{x.y.zhou2020，lxduan，liwenbnu}@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 222, + 214, + 389, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 214, + 389, + 227 + ], + "spans": [ + { + "bbox": [ + 222, + 214, + 389, + 227 + ], + "type": "text", + "content": "{wfanyue,ayao}@comp.nus.edu.sg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 232, + 307, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 232, + 307, + 242 + ], + "spans": [ + { + "bbox": [ + 59, + 232, + 307, + 242 + ], + "type": "text", + "content": "Query: \"The blonde woman works out in a gym with red lighting.\"" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 243, + 350, + 388 + ], + "blocks": [ + { + "bbox": [ + 59, + 243, + 350, + 388 + ], + "lines": [ + { + "bbox": [ + 59, + 243, + 350, + 388 + ], + "spans": [ + { + "bbox": [ + 59, + 243, + 350, + 388 + ], + "type": "image", + "image_path": "fb97a2924d77f0090dbfd6c50cb7ac69c98eb9d08d9f29618aa97a903884e18d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 400, + 231, + 413 + ], + "lines": [ + { + "bbox": [ + 175, + 400, + 231, + 413 + ], + "spans": [ + { + "bbox": [ + 175, + 400, + 231, + 413 + ], + "type": "text", + "content": "Figure 1. (a)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 364, + 236, + 555, + 395 + ], + "blocks": [ + { + "bbox": [ + 364, + 236, + 555, + 395 + ], + "lines": [ + { + "bbox": [ + 364, + 236, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 364, + 236, + 555, + 395 + ], + "type": "image", + "image_path": "d7793ea090ed810e53aa8fbe8687d45d682ffc86acef3ba1311fe1d5a1fa501a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 400, + 504, + 412 + ], + "lines": [ + { + "bbox": [ + 447, + 400, + 504, + 412 + ], + "spans": [ + { + "bbox": [ + 447, + 400, + 504, + 412 + ], + "type": "text", + "content": "Figure 1. (b)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 421, + 555, + 477 + ], + "lines": [ + { + "bbox": [ + 54, + 421, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 54, + 421, + 555, + 477 + ], + "type": "text", + "content": "Figure 1. (a) Comparison of moment retrieval models under normal and spurious correlation videos by masking the content of target clips of video. We found the existing works are suffering from a crucial reason stems from the spurious correlation between the text queries and the moment context. Baselines predict the Spurious GT even if the target moments are masked. In contrast, TD-DETR predicts the segment near the mask with lower confidence. (b) To verify the issue of spurious correlation, we introduce the Spurious mAP as the metric. Our model achieves the best ratio of mAP to Spurious mAP." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 486, + 200, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 486, + 200, + 498 + ], + "spans": [ + { + "bbox": [ + 151, + 486, + 200, + 498 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 513, + 297, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 513, + 297, + 681 + ], + "spans": [ + { + "bbox": [ + 54, + 513, + 297, + 681 + ], + "type": "text", + "content": "Given a textual query along with a corresponding video, the objective of moment retrieval aims to localize the moments relevant to the query within the video. While commendable results have been demonstrated by existing transformer-based approaches, predicting the accurate temporal span of the target moment is still a major challenge. This paper reveals that a crucial reason stems from the spurious correlation between the text query and the moment context. Namely, the model makes predictions by overly associating queries with background frames rather than distinguishing target moments. To address this issue, we propose a dynamic learning approach for moment retrieval, where two strategies are designed to mitigate the spurious correlation. First, we introduce a novel video synthesis approach to con" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 487, + 556, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 556, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 556, + 679 + ], + "type": "text", + "content": "struct a dynamic context for the queried moment, enabling the model to attend to the target moment of the corresponding query across dynamic backgrounds. Second, to alleviate the over-association with backgrounds, we enhance representations temporally by incorporating text-dynamics interaction, which encourages the model to align text with target moments through complementary dynamic representations. With the proposed method, our model significantly alleviates the spurious correlation issue in moment retrieval and establishes new state-of-the-art performance on two popular benchmarks, i.e., QVHighlights and Charades-STA. In addition, detailed ablation studies and evaluations across different architectures demonstrate the generalization and effectiveness of the proposed strategies. Our code is available at https://github.com/xyangzhou/TD-DETR." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 693, + 138, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 693, + 138, + 703 + ], + "spans": [ + { + "bbox": [ + 70, + 693, + 138, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 70, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 148, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20981" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 150 + ], + "type": "text", + "content": "Videos are a ubiquitous media format, but it can be time-consuming to browse through videos to localize specific moments. Using text to retrieve corresponding moments within a lengthy video[2, 54] plays an important role for entertainment, search, etc." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 151, + 295, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 295, + 305 + ], + "type": "text", + "content": "Existing moment retrieval approaches [15, 17, 19, 30, 49] leverage the DETR [4] detection transformer architecture to fuse text query into video representation. They show remarkable results in detecting the target moment by text-video alignment, but accurately predicting the temporal span remains a significant challenge. The learned video representations [17, 19] are aligned with the semantics of the text, thus generally biased to subjects or spatial objects. Consequently, the semantic elements in background frames are overly associated with textual queries. Such inappropriate association between irrelevant contexts and labels is the manifestation of Spurious correlation [1, 43], which are under-explored in moment retrieval tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 307, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 295, + 510 + ], + "type": "text", + "content": "This work identifies such spurious correlations between the text queries and moment visual context as the root cause for this performance gap in moment retrieval. Specifically, learned SOTA moment retrievers [17, 19, 30] tend to associate text queries with the background frames rather than distinguishing the target moment. Figure 1 illustrates how such spurious issues lead to a sub-optimal performance for moment retrieval. For example, the SOTA method, BAM-DETR [17], predicts a similar temporal span for an 'The blonde woman works out in a gym with red lighting.' in both the original video and an altered version where the target moment is masked out. In the masked case, the segment near the mask is a more reasonable output but the baseline methods [17, 29, 30, 49] predict the masked segment from background frames with the learned spurious correlation, which is the most unrelated output to the text query." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "To address the spurious issues, we propose a dynamic learning method with two novel strategies to mitigate spurious correlation. First, we propose a synthesis strategy that dynamically contextualizes the target moments for retrieval. We select similar video pairs and augment the samples by synthesizing target moments within similar contexts. This encourages the model to distinguish target moments from queries across more dynamic contexts, alleviating the over-association with background frames. Second, we propose to enhance video-text representations to reduce spurious dependencies on background frames by aligning text queries with temporal dynamics. To achieve this, we design a simple yet efficient temporal tokenizer to extract dynamic representations. Additionally, a text-dynamics interaction module is introduced to align dynamics with text queries, enabling the model to establish a stand-up correlation with complementary dynamic representations while" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 429, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 429, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 429, + 84 + ], + "type": "text", + "content": "mitigating background bias." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 85, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 553, + 167 + ], + "type": "text", + "content": "Experiments on two challenging benchmarks, i.e., QVHighlights and Charades-STA, show that our method outperforms existing methods by clear margins. Besides, spurious correlation in moment retrieval is validated to be largely alleviated by spurious evaluation. The detailed ablation analyses demonstrate the effectiveness of the proposed strategies to resolve the spurious issues." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 169, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 553, + 192 + ], + "type": "text", + "content": "In conclusion, our contributions to this work are summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 194, + 553, + 312 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 194, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 194, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 314, + 194, + 553, + 217 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to investigate the spurious correlation in moment retrieval." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 217, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 217, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 315, + 217, + 553, + 276 + ], + "type": "text", + "content": "- We propose a dynamic learning framework for moment retrieval that mitigates spurious correlations by dynamically contextualizing target moments through novel video synthesis and enhancing representations with text-dynamics alignment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 277, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 277, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 277, + 553, + 312 + ], + "type": "text", + "content": "- The proposed method achieves SOTA performance across all benchmarks. Besides, our model also provides a strong interpretation of alleviating spurious correlations." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 325, + 400, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 400, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 400, + 337 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 345, + 422, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 422, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 422, + 357 + ], + "type": "text", + "content": "2.1. Moment Retrieval" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 506 + ], + "type": "text", + "content": "Moment Retrieval aims to predict a temporal span within a video corresponding to a given natural language query [54]. Early methods had two stages: first, sampling candidate moments as proposals, and then scoring the proposals to obtain the final predictions [2, 8, 9, 23, 24]. However, more recent approaches are one-shot [18, 37, 47, 50-52, 56, 59] to predict the temporal span directly. Simultaneously, hybrid frameworks that integrate transformers for feature learning and fusion with CNNs for prediction have been widely studied. SnAG [31] explored the impact of cross-modal fusion on the scalability of video grounding models and proposed an efficient training scheme." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 507, + 554, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 554, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 554, + 615 + ], + "type": "text", + "content": "Although many moment retrieval methods have achieved remarkable performance, the majority primarily focus on architectural design, such as exploring new networks [16, 21, 27, 50, 52], improving contrastive learning, enhancing video-text alignment[18, 26], or refining attention mechanisms[31]. Generally, these methods focus on how to fuse text and video representations to improve the temporal span prediction but the overall performance are limited by the inefficient alignment architecture." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 624, + 448, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 448, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 448, + 635 + ], + "type": "text", + "content": "2.2. Detection Transformers" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "In the past several years, the wide adoption of detection transformers (DETR) [4] in video moment retrieval has gained remarkable performance [14, 17, 19, 25, 30, 48, 49]. While introducing the DETR framework to the task of moment retrieval, it demonstrated the effectiveness on QVHighlights dataset [19]. Furthermore, DETR-based ap" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20982" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 556, + 296 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 556, + 296 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 556, + 296 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 556, + 296 + ], + "type": "image", + "image_path": "54e94479e0177e6462ac6ada01e629a369d91e600fdd8c54d9b3b3df030c9041.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 304, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 54, + 304, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 555, + 350 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed TD-DETR. Video pairs are sampled by their similarity and then forwarded into the Video Synthesizer to generate Dynamic Context (Section 3.1). Then, the Dynamic Context, i.e., synthesized pairs, are enhanced by Temporal Dynamic Tokenization and interact with textual information (Section 3.2). Finally, the enhanced video tokens are sent to the transformer encoder-decoder with prediction heads to predict moments (Section 3.3)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 365, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 296, + 521 + ], + "type": "text", + "content": "proaches have explored the relationship between moment retrieval and highlight detection, achieving significant advancements in both tasks. However, they still struggle with inaccurate span predictions [33, 54], limiting their overall reliability. To address this issue, Zhang et al. [54] analyzes the distribution of target moments and highlights the problem of imbalance, while Otani et al. [33] investigate the skewed distribution of query representations and expose the overly query-dependent phenomenon of existing models. Unlike previous studies, we reveal that spurious correlations is the key cause of inaccurate span predictions, arising from the over-association between the text query and the background frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 529, + 178, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 178, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 178, + 541 + ], + "type": "text", + "content": "2.3. Spurious Correlation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "One reason for the poor generalization of vision algorithms is that they are prone to memorizing patterns or contextual cues [3, 10, 11, 13, 39, 41]. These patterns or cues are often spurious correlations - misleading heuristics of the training data correlated with the majority of examples but does not hold in general [11]. While numerous approaches have been proposed to address this challenge [33, 42, 44, 45, 54, 55, 57, 58] in various domains, their impact on video understanding remains largely unexplored. For example, in image-based tasks, spurious correlations are often tied to spatial biases [1, 43], whereas in video-related tasks, such spatial biases are redundant in frames. This distinction is particularly crucial in retrieving moments from lengthy videos, where the goal is not merely" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 365, + 555, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 400 + ], + "type": "text", + "content": "to recognize objects but to precisely localize temporal spans with text queries, making the study of spurious correlations in this context even more critical." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 402, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 556, + 498 + ], + "type": "text", + "content": "Despite in Moment Retrieval, Otani et al. [33] investigated the biases introduced by text queries and the neglect of semantic information, while Zhang et al.[54] explores the distribution of start and end moments. However, the issue of spurious correlation between the text and background frames is a key challenge of moment retrieval. In this work, we focus on this point and develop a new method by learning dynamics to alleviate this problem." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 511, + 397, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 397, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 397, + 525 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": "The objective of moment retrieval is to localize a temporal span that semantically corresponds to a given text query. Consider a video represented by " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " vision tokens, denoted by " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "V = \\{v_{1}, v_{2}, \\ldots, v_{L}\\}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": ", along with a natural language description of " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " words, as " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\{q_{1}, q_{2}, \\ldots, q_{W}\\}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": ". A moment retrieval model predicts a temporal span " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " with centre " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "m_{c}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " and duration " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "m_{\\sigma}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " which is relevant to the textual query, and token-wise saliency scores " + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\{s_{1}, s_{2}, \\ldots, s_{L}\\}" + }, + { + "bbox": [ + 313, + 533, + 555, + 628 + ], + "type": "text", + "content": " of the text." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "As shown in Figure 2, we propose a novel framework to learn Temporal Dynamics utilizing DEtection Transformer (TD-DETR). The video pairs are sampled by similarity and then forwarded into the video synthesizer to generate dynamic contexts. Then, the representations of synthesized pairs are enhanced through temporal dynamic interaction with texts. Finally, dynamic-enhanced representations are" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20983" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "fed into the transformer encoder-decoder with prediction heads to predict target moments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 106, + 263, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 106, + 263, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 106, + 263, + 118 + ], + "type": "text", + "content": "3.1. Video Synthesizer for Dynamic Context" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 124, + 296, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 124, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 124, + 296, + 194 + ], + "type": "text", + "content": "We introduce a novel video synthesizer to infuse the target moments with dynamic context for moment retrieval. Spurious correlations often stem from linking the moment's context to the text query. We address the issue by synthesizing new samples for the target moments with more dynamic contextual variations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 196, + 296, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 296, + 256 + ], + "type": "text", + "content": "Spurious Pair Selection. In order to synthesize a dynamic context aligned with the target moment, we construct a spurious pair of the target moment in a given video " + }, + { + "bbox": [ + 55, + 196, + 296, + 256 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 196, + 296, + 256 + ], + "type": "text", + "content": ". To ensure the challenge and rationality of the synthesized video, we select a video which is similar to " + }, + { + "bbox": [ + 55, + 196, + 296, + 256 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 196, + 296, + 256 + ], + "type": "text", + "content": " contextually." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "content": "In a training batch of " + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "content": " video samples, we sample the most similar video " + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "inline_equation", + "content": "\\{V_k\\}" + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "content": " for every " + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "inline_equation", + "content": "\\{V_i\\}" + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "inline_equation", + "content": "k \\in [1, N], k \\neq i" + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "content": ", for each " + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "inline_equation", + "content": "i \\in [1, N]" + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "content": " to construct a spurious video pair " + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "content": ". For any videos " + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "inline_equation", + "content": "V_j" + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "inline_equation", + "content": "V_l" + }, + { + "bbox": [ + 55, + 257, + 296, + 328 + ], + "type": "text", + "content": " from a batch, we employ cosine similarity to model the similarity relationship between videos as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 339, + 295, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 339, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 65, + 339, + 295, + 384 + ], + "type": "interline_equation", + "content": "s _ {j, l} = \\frac {1}{L _ {j} \\cdot L _ {l}} \\sum_ {p = 0} ^ {L _ {j}} \\sum_ {q = 0, p \\neq q} ^ {L _ {l}} \\frac {\\left. v _ {p} ^ {j} ^ {T} \\cdot v _ {q} ^ {l} \\right|}{\\left\\| v _ {p} ^ {j} \\right\\| \\cdot \\left\\| v _ {q} ^ {l} \\right\\|} \\quad j, l \\in [ 1, N ]. \\tag {1}", + "image_path": "12678e308bdd6295d3fa065a77cbb8a24c67ab14a844f8e8935127aa89543df2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": "In Eq. 1, we denote " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "v_{q}^{j}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "v_{p}^{l}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": " q-th clip of " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": " and q-th clip of " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "L_{j}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "L_{l}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": " for the length of " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "a_{j,l}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": " for the average similarity between " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": ". We select the most similar video " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "V_{k}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": " for every video " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "V_{i}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "inline_equation", + "content": "k = \\arg \\max_{k\\in [0,N]}s_{[i,:\\backslash \\{i\\}]}" + }, + { + "bbox": [ + 55, + 385, + 296, + 444 + ], + "type": "text", + "content": " in the batch." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": "Video Synthesis with Dynamic Context. After obtaining the spurious pairs, we synthesize a new video " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\tilde{V}_i" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " using the target moment " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " and the spurious pair " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "p_i = \\{V_i, V_k\\}" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": ". We then compose " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\{V_i, V_k\\}" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " and dynamically refine the ground truth of the target moment " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": ". Illustrated by Figure 3, given video " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "L_i" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " tokens and " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "V_k" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "L_k" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " tokens, on the one hand, we expect the sample from " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "V_k" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " with the completeness of target moment " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": ". We first sample from " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " without ground truth tokens " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{NG} \\in [0, L_i] \\setminus m_i" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " with sampling ratio " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": ", i.e., every token in NG has the same ratio " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " to be selected. We sample from the paired " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "V_k" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " with sampling ratio " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "1 - \\alpha" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": ", since " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "V_k" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": " is irrelevant to ground truth " + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{GT}_i" + }, + { + "bbox": [ + 55, + 446, + 296, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "text", + "content": "Finally, we concatenate the two sets of sampled video tokens into " + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\tilde{V}_i" + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "text", + "content": ". On the other hand, we have the same process, except now we focus on " + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "inline_equation", + "content": "V_k" + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{GT}_k" + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "text", + "content": " with sampling ratio " + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "text", + "content": " and ratio " + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "inline_equation", + "content": "1 - \\alpha" + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "text", + "content": " to sample from the whole " + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "text", + "content": " and concatenate sampled tokens into " + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\tilde{V}_k" + }, + { + "bbox": [ + 55, + 602, + 296, + 662 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 671, + 192, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 192, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 192, + 684 + ], + "type": "text", + "content": "3.2. Dynamics Enhancement" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "The attention of DETR-like architecture associates text queries with visual representations. Unfortunately, this" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 433, + 277 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 433, + 277 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 433, + 277 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 433, + 277 + ], + "type": "image", + "image_path": "adce30a8d306c1003640b9a965a0983caf13509e2bf91c4bb0f6b506296bd049.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 285, + 555, + 362 + ], + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 362 + ], + "type": "text", + "content": "Figure 3. Illustration of Video Synthesis with Dynamic Context. The numbers in the boxes represent the token indices, indicating their sequential order. Vision tokens are selected based on their selection probabilities and concatenated while preserving the completeness of the ground truth spans, i.e., selection probabilities of ground truth spans are set to 1. The synthesized videos maintain their original length with a random bias." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 434, + 70, + 553, + 277 + ], + "blocks": [ + { + "bbox": [ + 434, + 70, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 434, + 70, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 434, + 70, + 553, + 277 + ], + "type": "image", + "image_path": "0dd1edb110489cf5ad2162d8f83461b9c7e67c937801e41db34be8336affac99.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 379, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 554, + 486 + ], + "type": "text", + "content": "tends to emphasize background frames, thereby exacerbating the spurious correlation. To this end, besides the dynamic context synthesis, we introduce a dynamics enhancement module to encourage our model to align text queries with temporal dynamic representations. Thus, the model considers not only the association of background frames but also background-independent dynamics. This alignment enables our model to establish a stand-up correlation between the query-related moment and its context." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 487, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 534 + ], + "type": "text", + "content": "Dynamic Tokenizer. To model the dynamic temporality, we introduce a simple yet effective strategy to tokenize the temporal dynamics, which is nearly cost-free to obtain temporal dynamic representations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 535, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 554, + 583 + ], + "type": "text", + "content": "Given spurious pair " + }, + { + "bbox": [ + 313, + 535, + 554, + 583 + ], + "type": "inline_equation", + "content": "p_i = \\{\\tilde{V}_i, \\tilde{V}_k\\}" + }, + { + "bbox": [ + 313, + 535, + 554, + 583 + ], + "type": "text", + "content": ", the temporal dynamic tokenizer processes all video pairs in the same manner. Then, we use " + }, + { + "bbox": [ + 313, + 535, + 554, + 583 + ], + "type": "inline_equation", + "content": "\\tilde{V}_i" + }, + { + "bbox": [ + 313, + 535, + 554, + 583 + ], + "type": "text", + "content": " to explain the processing steps without loss of generality." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 584, + 554, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 554, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 554, + 607 + ], + "type": "text", + "content": "First, we concatenate a learnable token to the start of the video as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 380, + 617, + 553, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 617, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 380, + 617, + 553, + 632 + ], + "type": "interline_equation", + "content": "\\tilde {V} _ {i} = \\left\\{s t, \\tilde {v} _ {1}, \\tilde {v} _ {2}, \\dots , \\tilde {v} _ {L} \\right\\}, \\tag {2}", + "image_path": "12fb43d9f8e193c4e8ac102961842c75cdf184b5d3e5ab1adfa8d064f5880d5f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 643, + 554, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 554, + 690 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 643, + 554, + 690 + ], + "type": "inline_equation", + "content": "st" + }, + { + "bbox": [ + 313, + 643, + 554, + 690 + ], + "type": "text", + "content": " is a learnable token that captures the start signal of the video, preventing the loss of information from the first clip in the video representations. We then employ an element-wise difference to model the temporal dynamics." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 323, + 700, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 700, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 323, + 700, + 553, + 715 + ], + "type": "interline_equation", + "content": "T = \\left\\{\\tilde {v} _ {1} - s t, \\tilde {v} _ {2} - \\tilde {v} _ {1}, \\tilde {v} _ {3} - \\tilde {v} _ {2}, \\dots , \\tilde {v} _ {L _ {i}} - \\tilde {v} _ {L _ {i} - 1} \\right\\}, \\tag {3}", + "image_path": "95065e94206b4a04f3690e8a3fab1f6e24234acc452f1349ab36eac363a564d5.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20984" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " is a learnable dynamic representation, it learns and focuses on the dynamic information in nearby video clips." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 295, + 156 + ], + "type": "text", + "content": "Text-Dynamics Interaction. After obtaining dynamic tokens " + }, + { + "bbox": [ + 55, + 96, + 295, + 156 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 96, + 295, + 156 + ], + "type": "text", + "content": " in Eq. 3 from Dynamic Tokenizer, we utilize cross-attention for both text-vision and text-dynamic representations with the same text query to mitigate the over-emphasis on video backgrounds." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 295, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 295, + 191 + ], + "type": "text", + "content": "In detail, the query for each cross-attention layer is prepared by linear projection of the video and temporal clips as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 201, + 295, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 201, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 94, + 201, + 295, + 216 + ], + "type": "interline_equation", + "content": "Q _ {\\delta} = \\left[ p _ {\\delta} (\\delta), \\dots , p _ {\\delta} \\left(\\delta_ {L _ {i}}\\right) \\right], \\delta \\in \\{\\tilde {V} _ {i}, T \\}. \\tag {4}", + "image_path": "eb13770dc5a3f5d1b42d6dbfbb5bebf719c868255216681644b078c07df6c6e5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 224, + 296, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 285 + ], + "type": "text", + "content": "The key and value are computed with the query text representation as " + }, + { + "bbox": [ + 55, + 224, + 296, + 285 + ], + "type": "inline_equation", + "content": "K_{q} = [p_{k}(q_{1}),\\dots,p_{k}(q_{W})]" + }, + { + "bbox": [ + 55, + 224, + 296, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 224, + 296, + 285 + ], + "type": "inline_equation", + "content": "V_{q} = [p_{v}(q_{1}),\\dots,p_{v}(q_{L})]" + }, + { + "bbox": [ + 55, + 224, + 296, + 285 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 224, + 296, + 285 + ], + "type": "inline_equation", + "content": "p_{\\delta}, p_{k}, p_{v}" + }, + { + "bbox": [ + 55, + 224, + 296, + 285 + ], + "type": "text", + "content": " are linear projection layers for dynamic-enhanced query, temporal query, key and value. Then the cross-attention layer operates as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 293, + 295, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 293, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 65, + 293, + 295, + 331 + ], + "type": "interline_equation", + "content": "\\delta^ {\\prime} = \\operatorname {s o f t m a x} \\left(\\frac {Q _ {\\delta} K _ {q} ^ {T}}{\\sqrt {d}}\\right) V _ {q}, \\delta \\in \\left\\{\\tilde {V} _ {i}, T \\right\\}, \\delta^ {\\prime} \\in \\left\\{\\tilde {V} _ {i} ^ {\\prime}, T ^ {\\prime} \\right\\}, \\tag {5}", + "image_path": "d1eff6e0657252646b4c6a530d379650df2dfb0fb9f0d43f2b508e96bd5b4dcb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 331, + 296, + 392 + ], + "type": "text", + "content": " is the dimension of the projected key, value and query. To emphasize the learned temporal information, we utilize the weighted element-wise addition to inject text-guided temporal representation into text-guided dynamic-enhanced representation as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 119, + 399, + 295, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 399, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 119, + 399, + 295, + 414 + ], + "type": "interline_equation", + "content": "\\tilde {V} _ {i} ^ {\\prime} = \\beta \\cdot \\tilde {V} _ {i} + (1 - \\beta) \\cdot T ^ {\\prime}, \\tag {6}", + "image_path": "ceba1c4e1fbcaa72074d22af1dadbbc4940cb610732d35475b609e306aae8e58.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "type": "text", + "content": " is a hyper-parameter to adapt the addition ratio between " + }, + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "type": "inline_equation", + "content": "\\tilde{V}_i'" + }, + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "type": "inline_equation", + "content": "T'" + }, + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "type": "text", + "content": ". Hence, from the spurious pair " + }, + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "type": "inline_equation", + "content": "p_i = \\{\\tilde{V}_i, \\tilde{V}_k\\}" + }, + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "type": "text", + "content": ", we obtain a new spurious pair " + }, + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "type": "inline_equation", + "content": "p_i' = \\{\\tilde{V}_i', \\tilde{V}_k'\\}" + }, + { + "bbox": [ + 55, + 423, + 296, + 461 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 467, + 192, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 192, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 192, + 480 + ], + "type": "text", + "content": "3.3. Network and Objectives" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 484, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 521 + ], + "type": "text", + "content": "Our network structure (see Figure 2) follows previous work [19, 30]. It features the synthesis of dynamic contextual videos and the enhancement of dynamic representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "text", + "content": "Transformer encoder-decoder with prediction heads. Given spurious pair " + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "inline_equation", + "content": "p_i' = \\{\\tilde{V}_i', \\tilde{V}_k'\\}" + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "text", + "content": ", this module processes all video pairs in the same manner. Therefore, we use " + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "inline_equation", + "content": "\\tilde{V}_i'" + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "text", + "content": " to explain the processing steps without loss of generality. The encoder consists of " + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "text", + "content": " stacked typical transformer encoder layers, as in [19, 30], producing encoded representations " + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "inline_equation", + "content": "E_{enc}" + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "text", + "content": ". Our decoder, following [30], also uses " + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "text", + "content": " stacked typical transformer decoder layers, along with " + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "text", + "content": " learnable moment queries representing the centre " + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "inline_equation", + "content": "m_c" + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "text", + "content": " and duration " + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "inline_equation", + "content": "m_\\sigma" + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "text", + "content": ". The decoder processes " + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "inline_equation", + "content": "E_{enc}" + }, + { + "bbox": [ + 55, + 521, + 295, + 642 + ], + "type": "text", + "content": " with the moment queries." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "We adopt prediction heads following [30]. A linear layer predicts saliency scores from the encoded representations, and another linear layer handles negative pairs. From the decoder outputs, a 3-layer MLP with ReLU predicts the normalized moment center and duration, while a linear layer with softmax predicts the foreground." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "Hungarian matching. Following [4, 19, 30], we perform Hungarian matching between the two predictions and two corresponding labels respectively. Give prediction " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\tilde{V}_i'" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " and corresponding to ground truth " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": ", the optimal matching results between predictions and ground truths " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " can be written as " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma} = \\arg \\min_{\\sigma \\in G_N} \\sum_i^N \\mathcal{C}_{\\mathrm{match}}(y, \\hat{y}_{\\sigma(i)})" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " is a permutation of predictions and ground truths pairs and " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{match}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " is the matching cost." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "text", + "content": "Loss Functions. We calculate the loss between " + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}" + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "text", + "content": " and the ground truth " + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\tilde{V}_i'" + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "text", + "content": ". Following [19], the " + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "text", + "content": " loss " + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{L_1}" + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "text", + "content": " and the gIoU [38] loss " + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{gIoU}" + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "text", + "content": " are used to measure the distance and overlapping between the predictions and a cross-entropy loss " + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cls}" + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "text", + "content": " is used to measure classification. For highlight detection, we also use three loss functions which are margin ranking loss " + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{margin}" + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "text", + "content": ", rank-aware contrastive loss " + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{margin}" + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "text", + "content": " and negative loss " + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{neg}" + }, + { + "bbox": [ + 313, + 168, + 553, + 265 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 265, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 553, + 287 + ], + "type": "text", + "content": "The overall loss is shown as follows with " + }, + { + "bbox": [ + 313, + 265, + 553, + 287 + ], + "type": "inline_equation", + "content": "\\lambda_{*}" + }, + { + "bbox": [ + 313, + 265, + 553, + 287 + ], + "type": "text", + "content": " as balancing coefficient:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 339, + 299, + 553, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 299, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 339, + 299, + 553, + 312 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m o m e n t}} = \\lambda_ {L _ {1}} \\mathcal {L} _ {L _ {1}} + \\lambda_ {i o u} \\mathcal {L} _ {g I o U} + \\lambda_ {c l s} \\mathcal {L} _ {c l s}, \\tag {7}", + "image_path": "f3c18151cb37448ce0e254df704a185b78aac49ba936696bddafebaf3a396706.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 314, + 553, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 314, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 324, + 314, + 553, + 327 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h l} = \\lambda_ {\\text {m a r g i n}} \\mathcal {L} _ {\\text {m a r g i n}} + \\lambda_ {\\text {c o n t}} \\mathcal {L} _ {\\text {c o n t}} + \\lambda_ {\\text {n e g}} \\mathcal {L} _ {\\text {n e g}}, \\tag {8}", + "image_path": "fb23130b10842a7db94ed49dd9d213982ef402a4be7b54cc0295e3305cd64fad.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 381, + 329, + 553, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 329, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 381, + 329, + 553, + 341 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {h l} + \\mathcal {L} _ {\\text {m o m e n t}}. \\tag {9}", + "image_path": "dfb612efca694f4095ba1fd4c893c1da13db59ef1ab86a3967205145e38b9742.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 352, + 395, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 395, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 395, + 365 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 372, + 430, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 430, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 430, + 385 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 389, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 553, + 510 + ], + "type": "text", + "content": "Datasets. We evaluate our method on two widely-used benchmarks: QVHighlights [19] and Charades-STA [8], following the setup of prior works [19, 30]. QVHighlights has over 10,000 video-query pairs primarily from vlog and news content. We use the splits defined in MomentDETR [19] and report results for both the val and test splits. Charades-STA, derived from the Charades dataset [40], has more than 18,000 video-sentence pairs. We follow the standard evaluation protocol from [8] and use two separate training and testing splits." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "type": "text", + "content": "Standard Evaluation. We adopt the same metrics of Moment-DETR [17, 19, 30] for evaluation. Specifically, for Moment Retrieval, we report the mean average precision (mAP) at Intersection over Union (IoU) thresholds of 0.5 and 0.75, as well as the average mAP over IoU thresholds ranging from 0.5 to 0.95 with a step size of 0.05. We also include Recall@1 (R@1) at IoU thresholds of 0.5 and 0.75. For a fair comparison, we also include the test split of the QVHighlights dataset, which is evaluated on the CodaLab competition platform [35]. Notably, performance at higher IoU thresholds, such as 0.7, serves as an indicator of more precise alignments between predicted moments and the ground truth." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": "Spurious Evaluation. To verify the spurious correlation with overly association the text with background, we replace the target moments in video content with random-valued masks while keeping the video duration unchanged," + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20985" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 538, + 278 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 538, + 278 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 538, + 278 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 538, + 278 + ], + "type": "table", + "html": "
MethodMR-R1MR-mAP
@0.5@0.7@0.5@0.75Average
MCN [2] ICCV'1711.412.7224.948.2210.67
CAL [6] arXiv'1925.4911.5423.407.659.89
XML [18] ECCV'2041.8330.3544.6331.7332.14
XML+ [19] NIPS'2146.6933.4647.8934.6734.90
SnAG† [31] CVPR'2459.7948.1058.6344.3742.71
SnAG /w TD-DETR66.4852.9363.7149.1146.75
Moment-DETR [19] NIPS'2152.8933.0254.8229.4030.73
UMT [25] CVPR'2256.2341.1853.8337.0136.12
MomentDiff [20] NIPS'2357.4239.6654.0235.7335.95
QD-DETR [30] CVPR'2362.4044.9862.5239.8839.86
UniVTG [22] ICCV'2358.8640.8657.6035.5935.47
CG-DETR [29] arXiv'2365.4048.4064.5042.8042.90
UVCOM [46] CVPR'2463.5548.7064.4744.0143.27
BAM-DETR [17] ECCV'2464.5348.6464.5746.3345.36
TD-DETR (Ours)64.53±0.6250.37±0.5366.21±0.2147.32±0.5346.69±0.26
", + "image_path": "52861c8339750c82453e5b69f2ab41642c55118535b281a1f21c5c0cfbcaba92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 69, + 280, + 201, + 293 + ], + "lines": [ + { + "bbox": [ + 69, + 280, + 201, + 293 + ], + "spans": [ + { + "bbox": [ + 69, + 280, + 201, + 293 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 69, + 280, + 201, + 293 + ], + "type": "text", + "content": " reproduced by the official code" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 67, + 300, + 538, + 335 + ], + "lines": [ + { + "bbox": [ + 67, + 300, + 538, + 335 + ], + "spans": [ + { + "bbox": [ + 67, + 300, + 538, + 335 + ], + "type": "text", + "content": "Table 1. Performance comparison on QVHighlights test split. Our experimental results are averaged over three runs and " + }, + { + "bbox": [ + 67, + 300, + 538, + 335 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 67, + 300, + 538, + 335 + ], + "type": "text", + "content": " denotes the standard deviation. For the compared methods, the results are copied from their original papers and we reproduce SnAG by the official code via slowfast+clip features." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 347, + 294, + 521 + ], + "blocks": [ + { + "bbox": [ + 57, + 347, + 294, + 521 + ], + "lines": [ + { + "bbox": [ + 57, + 347, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 347, + 294, + 521 + ], + "type": "table", + "html": "
MethodR1@0.5R1@0.7
CAL [6]44.9024.37
2D TAN [56]39.7023.31
VSLNet [53]47.3130.19
IVG-DCL [32]50.2432.88
SnAG†[31]65.7237.32
SnAG /w TD-DETR70.1442.35
Moment-DETR [19]53.6331.37
Moment-Diff [20]55.5732.42
UMT [25]48.3129.25
QD-DETR [30]57.3132.55
CG-DETR [29]58.4036.30
BAM-DETR [17]59.9539.38
TD-DETR (Ours)60.8940.35
", + "image_path": "6dae8998099b6c4aa6e3c531b57342ff43f813eb6023772f5300a4db6ee4b683.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 56, + 523, + 188, + 536 + ], + "lines": [ + { + "bbox": [ + 56, + 523, + 188, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 188, + 536 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 56, + 523, + 188, + 536 + ], + "type": "text", + "content": " reproduced by the official code" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 55, + 544, + 295, + 578 + ], + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 578 + ], + "type": "text", + "content": "Table 2. Performance comparison on Charades-STA test split. Our experimental results are averaged over three runs. For all the compared methods, the results are taken from their original papers." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "text", + "content": "making the masked region the most unrelated moments to text queries. We then assess the model's performance under these conditions and report the spurious metric of Spurious " + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "inline_equation", + "content": "R@1" + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "text", + "content": " and Spurious mAP, analogous to the standard R@1 and mAP. In this case, lower values are better, as they indicate less spurious correlation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Implementation Details. Our method is implemented in PyTorch [34]. We follow the implementation of QD-DETR [30]. For all datasets, we use video features both extracted from SlowFast [7] pre-trained on Kinetics [5] and" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 348, + 569, + 440 + ], + "blocks": [ + { + "bbox": [ + 315, + 348, + 569, + 440 + ], + "lines": [ + { + "bbox": [ + 315, + 348, + 569, + 440 + ], + "spans": [ + { + "bbox": [ + 315, + 348, + 569, + 440 + ], + "type": "table", + "html": "
MethodSpurious R1 ↓Spurious mAP ↓Standard mAP ↑
@0.7@0.9@0.75Avg.@0.75Avg.
QD-DETR9.355.299.9010.4041.8241.22
Ours w/ QD8.263.687.468.1549.8649.05
CG-DETR4.651.295.556.1445.7044.90
Ours w/ CG2.580.393.384.4149.1648.38
BAM-DETR7.161.876.306.7248.5647.61
Ours w/ BAM1.610.521.731.9849.6248.67
", + "image_path": "3963ae56407e89751f67fe2b6c4e9595d1d46cac3371b49528b201aab7248854.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 448, + 555, + 536 + ], + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 536 + ], + "type": "text", + "content": "Table 3. Performance comparison on QVHighlights val split with random masks. The target clips in the video content are replaced with random-valued masks while keeping the video duration unchanged. We compare the baseline models—QD-DETR, CG-DETR, and BAM-DETR—against their respective versions integrated with our proposed TD-DETR, marked Ours w/, under the same experimental settings. The metrics here have been explained in section 4.1." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 555, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 590 + ], + "type": "text", + "content": "pre-trained CLIP [36] vision encoder, and text feature extracted from pre-trained CLIP [36] text encoder, following Moment-DETR." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "text", + "content": "- For QVHighlights dataset, we set batch size to 32 and an initial learning rate of " + }, + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "text", + "content": " with weight decay of " + }, + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "text", + "content": ". We set the hidden size " + }, + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "inline_equation", + "content": "d = 256" + }, + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "text", + "content": ", layers of encoder/decoder " + }, + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "text", + "content": ", and moment queries " + }, + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 314, + 594, + 555, + 654 + ], + "type": "text", + "content": ". The model is trained for 200 epochs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": "- For Charades-STA, we set batch size to 32 and use an initial learning rate of " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": " with weight decay of " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": ". We set the hidden size " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "d = 256" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": ", layers of encoder/decoder " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "T = 3" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": ", and moment queries " + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "inline_equation", + "content": "N = 10" + }, + { + "bbox": [ + 313, + 654, + 556, + 714 + ], + "type": "text", + "content": ". The model is trained for 100 epochs and the learning rate is decayed to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20986" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 66, + 72, + 146, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 146, + 85 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 146, + 85 + ], + "type": "inline_equation", + "content": "\\frac{1}{10}" + }, + { + "bbox": [ + 66, + 72, + 146, + 85 + ], + "type": "text", + "content": " every 40 epochs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 294, + 204 + ], + "type": "text", + "content": "We use mini-batch and AdamW [28] gradient descent algorithm to optimize the network parameters, initialize weights with Xavier init [12], use dropout of 0.1 for transformer layers and 0.5 for input projection layers and set cost coefficient as the same as loss balancing coefficient on all datasets. For Dynamic Video Synthesizer, the selection Probability " + }, + { + "bbox": [ + 55, + 85, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 85, + 294, + 204 + ], + "type": "text", + "content": " is set to 0.5 and for Dynamic Temporal Identifier the " + }, + { + "bbox": [ + 55, + 85, + 294, + 204 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 85, + 294, + 204 + ], + "type": "text", + "content": " is set to 0.7 on all datasets. The total training time is approximately 8 hours on QVHighlight and 6 hours for Charades-STA, with a single NVIDIA RTX 2080 GPU." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 228, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 228, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 228, + 225 + ], + "type": "text", + "content": "4.2. Results on Standard Evaluation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 231, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 295, + 266 + ], + "type": "text", + "content": "As illustrated in Table 1 and Table 2, we compare " + }, + { + "bbox": [ + 55, + 231, + 295, + 266 + ], + "type": "inline_equation", + "content": "TD" + }, + { + "bbox": [ + 55, + 231, + 295, + 266 + ], + "type": "text", + "content": "-DETR with several baseline methods on QVHighlights and Charades-STA benchmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 267, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 295, + 410 + ], + "type": "text", + "content": "Results on QVHighlights. On QVHighlights test split, we compare our model against proposal-based, proposal-free, and DETR-based methods, as shown in Table 1. As observed, our TD-DETR achieves state-of-the-art performance across all evaluation metrics. Notably, TD-DETR shows remarkable improvements compared to the other methods. Performance at higher IoU thresholds (e.g., 0.7) provides a clearer indication of the alignment between predicted moments and the ground truth. Our TD-DETR outperforms the previous state-of-the-art model by a substantial margin, with improvements of up to " + }, + { + "bbox": [ + 55, + 267, + 295, + 410 + ], + "type": "inline_equation", + "content": "3.56\\%" + }, + { + "bbox": [ + 55, + 267, + 295, + 410 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 267, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{R}@\\mathbf{1}@\\mathbf{0.7}" + }, + { + "bbox": [ + 55, + 267, + 295, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 267, + 295, + 410 + ], + "type": "inline_equation", + "content": "2.14\\%" + }, + { + "bbox": [ + 55, + 267, + 295, + 410 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 267, + 295, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}@0.75" + }, + { + "bbox": [ + 55, + 267, + 295, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 411, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 519 + ], + "type": "text", + "content": "Results on Charades-STA. We evaluate our method methods on Charades-STA test split, compared with proposal-based, proposal-free and DETR-based methods. As shown in Table 2, TD-DETR achieves state-of-the-art performance across all evaluation metrics. Notably, our method achieves improvements of up to " + }, + { + "bbox": [ + 55, + 411, + 295, + 519 + ], + "type": "inline_equation", + "content": "2.46\\%" + }, + { + "bbox": [ + 55, + 411, + 295, + 519 + ], + "type": "text", + "content": " in R1@0.7 and " + }, + { + "bbox": [ + 55, + 411, + 295, + 519 + ], + "type": "inline_equation", + "content": "1.57\\%" + }, + { + "bbox": [ + 55, + 411, + 295, + 519 + ], + "type": "text", + "content": " in R1@0.5 on this challenging benchmark. Recall that performance at high IoU thresholds stands for the precision with which predictions match the ground truth data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 528, + 284, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 284, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 284, + 540 + ], + "type": "text", + "content": "4.3. Results on Spurious Correlation Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 605 + ], + "type": "text", + "content": "Moment retrieval models are often prone to inaccurate predictions due to spurious correlations, a typical example being when predictions rely overly on background cues. To investigate this issue, we adopt the spurious evaluation setting pre-defined in Sec 4.1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "We evaluate our methods on several strong baselines, including QD-DETR [30], CG-DETR [29] and BAM-DETR [17]. The results in Table 3 demonstrate the alleviation of spurious correlation issue, as indicated by the improved Spurious " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "R1" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " and Spurious " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "mAP" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " across all compared baselines while boosting the standard " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "mAP" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": ". For instance, our methods improve QD-DETR [30] in Spurious " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "R1@0.9" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " from 10.40 to 8.15 (30.43%), with standard " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "mAP@0.75" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " increasing from 41.22 to 49.05 (18.99%). While for BAM" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "DETR [17], it achieves the most substantial improvement, improving Spurious R1 to 1.61 and Spurious mAP to 1.73. The results on spurious correlation evaluation demonstrate the effectiveness of our proposed approaches to alleviate the spurious correlation issue and also the generalization across different baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 152, + 432, + 165 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 432, + 165 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 432, + 165 + ], + "type": "text", + "content": "4.4. Qualitative Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 171, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 338 + ], + "type": "text", + "content": "In this subsection, we analyze the impact of mitigating spurious correlations in our TD-DETR model. As illustrated in Figure 4, the entire video depicts a man speaking to the camera, while the target moment is described as: \"A man wearing a cap backwards talking while some other videos appear in the bottom left corner of the screen.\" Notably, baseline models fail to distinguish between \"The man is talking\" and \"The man is talking while other videos appear in the corner,\" indicating that background elements are overly correlated with the textual query. In contrast, our TD-DETR accurately predicts the target moment without being misled by background frames, demonstrating its ability to associate the information between text and the proper frames rather than relying on spurious correlations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 346, + 416, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 416, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 416, + 358 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "content": "We conduct ablation studies on both QVHighlights and Charades-STA benchmarks to validate the effectiveness of each proposed component. VSDC and TDEM refer to Video Synthesizer for Dynamic Context and Temporal Dynamics Enhancement Module, respectively. As illustrated in Table 4, rows (b) to (c) show the effectiveness of each component compared to the baseline model (a), and (d) demonstrate the overall effectiveness of all the components. In detail, on QVHighlight, Video Synthesizer for Dynamic Context contributes improvement of " + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "inline_equation", + "content": "5.60\\%" + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "content": " in R1@0.7, " + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "inline_equation", + "content": "9.18\\%" + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "content": " in mAP@0.75 and " + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "inline_equation", + "content": "9.05\\%" + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "content": " in average mAP while Temporal Dynamics Enhancement Module contributes improvement of " + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "inline_equation", + "content": "3.16\\%" + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "content": " in R1@0.7, " + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "inline_equation", + "content": "4.19\\%" + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "content": " in mAP@0.75 and " + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "inline_equation", + "content": "5.41\\%" + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "content": " in average mAP for Moment Retrieval. With all components integrated, we observe a substantial " + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "inline_equation", + "content": "14.75\\%" + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "inline_equation", + "content": "14.20\\%" + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "inline_equation", + "content": "15.30\\%" + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "content": " improvement in the R1@0.7, mAP@0.75 and average mAP, respectively. Moreover, the spurious correlation is also improved, where " + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "inline_equation", + "content": "30.43\\%" + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "content": " in Spurious R1@0.9 and " + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "inline_equation", + "content": "21.63\\%" + }, + { + "bbox": [ + 313, + 364, + 555, + 616 + ], + "type": "text", + "content": " in Spurious mAP. The same phenomenon can be observed in Charades-STA as well. More experiments and analyses are included in the supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 624, + 552, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 552, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 552, + 635 + ], + "type": "text", + "content": "4.6. Generalization across Different Architectures." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 700 + ], + "type": "text", + "content": "We validate the generalization by incorporating our proposed approaches into different architectures. SnAG [31] is a hybrid framework, which integrates transformer for feature fusion and CNN for prediction, while QD-DETR [30] and BAM-DETR [17] follow DETR-like architecture." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "As illustrated in Table 5, for the hybrid architecture, our" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20987" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 78, + 548, + 236 + ], + "blocks": [ + { + "bbox": [ + 63, + 78, + 548, + 236 + ], + "lines": [ + { + "bbox": [ + 63, + 78, + 548, + 236 + ], + "spans": [ + { + "bbox": [ + 63, + 78, + 548, + 236 + ], + "type": "image", + "image_path": "b57c6f4367b9490e4acaa37292ac57fab2f77f4c8a507699f08d88d293347a15.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 250, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 555, + 297 + ], + "type": "text", + "content": "Figure 4. Visualization of results comparison between the baseline models and our TD-DETR. We display the relative position of ground truth and prediction moments in colourful boxes, where the green with GT stands for ground truth moment, the orange box with QD-DETR stands for our baseline predictions and the purple box with ours stands for our TD-DETR. To quantify the quality of predictions, we mark the start and end timestamps in the box and mark the IoU with the same colour as the models. Best read in colour." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 305, + 558, + 387 + ], + "blocks": [ + { + "bbox": [ + 56, + 305, + 558, + 387 + ], + "lines": [ + { + "bbox": [ + 56, + 305, + 558, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 558, + 387 + ], + "type": "table", + "html": "
VSDCTDEMQVHighlightCharades-STA
R1↑mAP↑Spurious R1↓Spurious mAP↓R1↑Spurious R1↓
@0.5@0.7@0.5@0.75Avg.@0.7@0.9@0.75Avg.@0.5@0.7@0.7@0.9
(a)61.1246.7762.4543.6642.549.355.299.9010.4057.3132.5525.726.31
(b)✓63.4749.3964.8247.6746.398.773.878.648.9139.1263.6723.155.42
(c)✓62.9348.2564.2245.4944.848.844.09.109.5638.5160.8024.035.73
(d)✓✓65.8853.6766.4349.8649.058.263.687.468.1560.8940.3522.134.82
", + "image_path": "1c51bf30ddbf1414ee76d8894cf4f40e845f1a71bc161ec2d5382931c518ddde.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 431, + 306, + 532 + ], + "blocks": [ + { + "bbox": [ + 55, + 395, + 555, + 429 + ], + "lines": [ + { + "bbox": [ + 55, + 395, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 555, + 429 + ], + "type": "text", + "content": "Table 4. Ablation studies on QVHighlights val split and Charades-STA test split. VSDC and TDEM stand for Video Synthesizer for Dynamic Context and Temporal Dynamics Enhancement Module respectively. All the ablation results are averaged over three runs. We also reveal spurious metrics of the proposed modules." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 431, + 306, + 532 + ], + "lines": [ + { + "bbox": [ + 56, + 431, + 306, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 306, + 532 + ], + "type": "table", + "html": "
MethodQVHighlights valCharades-STA test
R1@0.7mAP@0.75mAPR1@0.5R1@0.7
CG52.1045.7044.9058.4036.30
Ours w/ CG53.25+1.1549.16+3.4648.38+3.4859.35+0.9537.84+1.54
BAM51.6148.5647.6159.9539.38
Ours w/BAM52.87+1.2649.62+1.0648.82+1.2160.92+0.9740.25+0.87
QD46.6641.8241.2257.3132.55
Ours w/ QD53.67+7.0149.86+8.0449.00+7.7860.89+3.5840.35+7.80
SnAG48.1044.3742.7165.7237.32
Ours w/ SnAG52.93+4.8349.11+4.7446.75+4.0470.14+4.4242.35+5.03
", + "image_path": "8eee95c6198ce104db6e86905dfb646331a72677112203ce5def50f63451ffea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 295, + 584 + ], + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 584 + ], + "type": "text", + "content": "Table 5. Generalization across different architectures on both QVHighlight val split and Charades-STA test split. CG stands for CG-DETR, while QD stands for QD-DETR and BAM stands for BAM-DETR." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "method via SnAG achieves improvements of up to " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "10.04\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " in R1@0.7 and " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "9.46\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " in mAP on QVHighlights and " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "6.73\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " in R1@0.5 and " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "13.48\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " in R1@0.7 on Charades-STA. For DETR-like architecture, our approach enhances QD-DETR with significant gains of " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "19.23\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " in R1@0.7 and " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "18.87\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " in mAP on QVHighlights and " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "6.25\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " in R1@0.5 and " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "23.96\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " in R1@0.7 on Charades-STA. Our strategies significantly improve the performance of both hybrid and DETR-like baselines on both benchmarks, demonstrating the generalization of our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 442, + 388, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 442, + 388, + 455 + ], + "spans": [ + { + "bbox": [ + 314, + 442, + 388, + 455 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": "Although existing transformer-based approaches have demonstrated remarkable performance, they still struggle with spurious correlation, which overly associates the textual query with the background frames rather than the target moment. To address this issue in moment retrieval, we introduce a novel dynamic learning method with two strategies. First, we propose a novel video synthesis strategy that constructs a dynamic context for the relevant moment. This synthesis strategy enables our model to attend to the target moment corresponding to the query across various dynamic video contexts. Second, we enhance the representation by learning temporal dynamics aligned with texts. In addition to visual features, text queries are aligned with dynamic representations, encouraging our model to establish a non-spurious correlation between the query-related moment and its context. Extensive experiments and detailed ablation studies on both the QVHighlights and Charades-STA benchmarks validate the effectiveness and generalization of the proposed TD-DETR, demonstrating significant alleviation of spurious correlation and superior performance in moment retrieval." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20988" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (No. 62176047, No. 62476051), the Sichuan Natural Science Foundation (No. 2024NSFTD0041), and the Sichuan Science and Technology Program (No. 2021YFS0374)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 180, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 180, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 296, + 224 + ], + "type": "text", + "content": "[1] Vedika Agarwal, Rakshith Shetty, and Mario Fritz. Towards causal vqa: Revealing and reducing spurious correlations by invariant and covariant semantic editing. In CVPR, pages 9690-9698, 2020. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 225, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 225, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 225, + 296, + 269 + ], + "type": "text", + "content": "[2] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 270, + 295, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 295, + 292 + ], + "type": "text", + "content": "[3] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, pages 456-473, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 293, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 293, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 293, + 295, + 335 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229. Springer, 2020. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 336, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 295, + 369 + ], + "type": "text", + "content": "[5] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In CVPR, pages 6299-6308, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 370, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 370, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 370, + 296, + 413 + ], + "type": "text", + "content": "[6] Victor Escorcia, Mattia Soldan, Josef Sivic, Bernard Ghanem, and Bryan Russell. Finding moments in video collections using natural language. arXiv preprint arXiv:1907.12763, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 415, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 446 + ], + "type": "text", + "content": "[7] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In ICCV, pages 6202-6211, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 448, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 448, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 448, + 294, + 480 + ], + "type": "text", + "content": "[8] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, pages 5267-5275, 2017. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 481, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 294, + 513 + ], + "type": "text", + "content": "[9] Runzhou Ge, Jiyang Gao, Kan Chen, and Ram Nevatia. Mac: Mining activity concepts for language-based temporal localization. In WACV, pages 245-253. IEEE, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 514, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 567 + ], + "type": "text", + "content": "[10] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 294, + 602 + ], + "type": "text", + "content": "[11] Soumya Suva Ghosal and Yixuan Li. Are vision transformers robust to spurious correlations? In IJCV, pages 689-709. Springer, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 603, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 294, + 635 + ], + "type": "text", + "content": "[12] Xavier Glorot and Yoshua Bengio. Understanding the difficulty of training deep feedforward neural networks. In AISTATS, 2010. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 636, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 669 + ], + "type": "text", + "content": "[13] Karan Goel, Albert Gu, Yixuan Li, and Christopher Ré. Model patching: Closing the subgroup performance gap with data augmentation. In ICLR, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[14] Jinhyun Jang, Jungin Park, Jin Kim, Hyeongjun Kwon, and Kwanghoon Sohn. Knowing where to focus: Event-aware transformer for video grounding. In ICCV, pages 13846-13856, 2023. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] Minjoon Jung, Youwon Jang, Seongho Choi, Joochan Kim, Jin-Hwa Kim, and Byoung-Tak Zhang. Background-aware moment detection for video moment retrieval. In WACV, pages 8586-8596, 2025. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 160 + ], + "type": "text", + "content": "[16] Minjoon Jung, Junbin Xiao, Byoung-Tak Zhang, and Angela Yao. On the consistency of video large language models in temporal comprehension. In CVPR, pages 13713-13722, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "text", + "content": "[17] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In ECCV, pages 220-238. Springer, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "text", + "content": "[18] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In ECCV, pages 447-463. Springer, 2020. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "text", + "content": "[19] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In NeurIPS, pages 11846-11858, 2021. 2, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 272, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 315 + ], + "type": "text", + "content": "[20] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. In NeurlPS, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "text", + "content": "[21] Xunsong Li, Pengzhan Sun, Yangcen Liu, Lixin Duan, and Wen Li. Simultaneous detection and interaction reasoning for object-centric action recognition. IEEE Transactions on Multimedia, pages 1-13, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 414 + ], + "type": "text", + "content": "[22] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 448 + ], + "type": "text", + "content": "[23] Meng Liu, Xiang Wang, Liqiang Nie, Xiangnan He, Baoquan Chen, and Tat-Seng Chua. Attentive moment retrieval in videos. In SIGIR, pages 15-24, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 481 + ], + "type": "text", + "content": "[24] Meng Liu, Xiang Wang, Liqiang Nie, Qi Tian, Baoquan Chen, and Tat-Seng Chua. Cross-modal moment localization in videos. In ACM MM, pages 843-851, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "text", + "content": "[25] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, pages 3042-3051, 2022. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 568 + ], + "type": "text", + "content": "[26] Yangcen Liu, Ziyi Liu, Yuanhao Zhai, Wen Li, David Doerman, and Junsong Yuan. Stat: Towards generalizable temporal action localization. arXiv preprint arXiv:2404.13311, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 570, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 602 + ], + "type": "text", + "content": "[27] Ziyi Liu and Yangcen Liu. Bridge the gap: From weak to full supervision for temporal action localization with pseudoformer. In CVPR, pages 8711-8720, 2025. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 603, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 624 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 668 + ], + "type": "text", + "content": "[29] WonJun Moon, Sangeek Hyun, SuBeen Lee, and JaePil Heo. Correlation-guided query-dependency calibration for video temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 2, 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[30] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20989" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 703 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[31] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. In CVPR, pages 18930-18940, 2024. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "text", + "content": "[32] Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. Interventional video grounding with dual contrastive learning. In CVPR, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "type": "text", + "content": "[33] Mayu Otani, Yuta Nakashima, Esa Rahtu, and Janne Heikkilä. Uncovering hidden challenges in query-based video moment retrieval. In BMVC, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "type": "text", + "content": "[34] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "text", + "content": "[35] Adrien Pavao, Isabelle Guyon, Anne-Catherine Letournel, Dinh-Tuan Tran, Xavier Baro, Hugo Jair Escalante, Sergio Escalera, Tyler Thomas, and Zhen Xu. Codalab competitions: An open source platform to organize scientific challenges. JMLR, 24(198):1-6, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[37] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. In Trans. Assoc. Comput. Linguistics, pages 25-36, 2013. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[38] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "text", + "content": "[39] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "text", + "content": "[40] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In ECCV, pages 510-526. Springer, 2016. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 568 + ], + "type": "text", + "content": "[41] Lifu Tu, Garima Lalwani, Spandana Gella, and He He. An empirical study on robustness to spurious correlations using pre-trained language models. In ACL, pages 621-633, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "text", + "content": "[42] Yipei Wang and Xiaogian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In AISTATS, pages 3745-3753. PMLR, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 647 + ], + "type": "text", + "content": "[43] Yipei Wang and Xiaoqian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In International Conference on Artificial Intelligence and Statistics, pages 3745-3753. PMLR, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 669 + ], + "type": "text", + "content": "[44] Zhao Wang and Aron Culotta. Identifying spurious correlations for robust text classification. In EMNLP, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 703 + ], + "type": "text", + "content": "[45] Fanyue Wei, Biao Wang, Tiezheng Ge, Yuning Jiang, Wen Li, and Lixin Duan. Learning pixel-level distinctions for video highlight detection. In CVPR, pages 3073-3082, 2022." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 654 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[46] Yicheng Xiao, Zhuoyan Luo, Yong Liu, Yue Ma, Hengwei Bian, Yatai Ji, Yujiu Yang, and Xiu Li. Bridging the gap: A unified video comprehension framework for moment retrieval and highlight detection. In CVPR, pages 18709-18719, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[47] Mengmeng Xu, Mattia Soldan, Jialin Gao, Shuming Liu, Juan-Manuel Pérez-Rúa, and Bernard Ghanem. Boundary-denoising for video activity localization. arXiv preprint arXiv:2304.02934, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 173, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 206 + ], + "type": "text", + "content": "[48] Yifang Xu, Yunzhuo Sun, Yang Li, Yilei Shi, Xiaoxiang Zhu, and Sidan Du. Mh-detr: Video moment and highlight detection with cross-modal transformer. In ACM MM, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "text", + "content": "[49] Jin Yang, Ping Wei, Huan Li, and Ziyang Ren. Task-driven exploration: Decoupling and inter-task feedback for joint moment retrieval and highlight detection. In CVPR, pages 18308-18318, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "text", + "content": "[50] Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos. NeurlPS, 32, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[51] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In CVPR, pages 10287-10296, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 320, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 364 + ], + "type": "text", + "content": "[52] Da Zhang, Xiyang Dai, Xin Wang, Yuan-Fang Wang, and Larry S Davis. Man: Moment alignment network for natural language moment retrieval via iterative graph adjustment. In CVPR, pages 1247-1257, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "text", + "content": "[53] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In ACL, 2020. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 441 + ], + "type": "text", + "content": "[54] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Temporal sentence grounding in videos: A survey and future directions. In IEEE TPAMI, pages 10443-10465. IEEE, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 487 + ], + "type": "text", + "content": "[55] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Ré. Correct-n-contrast: A contrastive approach for improving robustness to spurious correlations. In ICML, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 488, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 531 + ], + "type": "text", + "content": "[56] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, pages 12870-12877, 2020. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 533, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 577 + ], + "type": "text", + "content": "[57] Xing Zhang, Zuxuan Wu, Zejia Weng, Huazhu Fu, Jingjing Chen, Yu-Gang Jiang, and Larry S Davis. Videolt: Large-scale long-tailed video recognition. In ICCV, pages 7960-7969, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 578, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 611 + ], + "type": "text", + "content": "[58] Yifan Zhang, Bingyi Kang, Bryan Hooi, Shuicheng Yan, and Jiashi Feng. Deep long-tailed learning: A survey. In IEEE TPAMI, pages 10795-10816. IEEE, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 612, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 654 + ], + "type": "text", + "content": "[59] Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In ICCV, pages 2914-2923, 2017. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20990" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/f4964a00-88d3-4539-871c-39c0ed82d2c6_content_list.json b/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/f4964a00-88d3-4539-871c-39c0ed82d2c6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2407bccfc90a0547e299d36f8c0d95ffb8e4f84f --- /dev/null +++ b/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/f4964a00-88d3-4539-871c-39c0ed82d2c6_content_list.json @@ -0,0 +1,1623 @@ +[ + { + "type": "text", + "text": "The Inter-Intra Modal Measure: A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models", + "text_level": 1, + "bbox": [ + 94, + 128, + 901, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Laura Niss* Kevin Vogt-Lowell* Theodoros Tsiligkaridis", + "bbox": [ + 254, + 203, + 743, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{laura.niss, kevin.vogt-lowell, ttsili}@ll.mit.edu", + "bbox": [ + 277, + 223, + 720, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MIT Lincoln Laboratory", + "bbox": [ + 398, + 239, + 596, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "244 Wood St, Lexington, MA 02421", + "bbox": [ + 351, + 256, + 640, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The fine-tuning of large vision-language foundation models remains an underexplored area, particularly regarding its impact on learning gains and catastrophic forgetting. Inspired by the significance of modality gaps in contrastive dual-encoders, we introduce the Inter-Intra Modal Measure (IIMM)—a predictive metric that quantifies the relationship between intra-modal image embedding similarity and inter-modal misalignment. Through extensive empirical analysis across four state-of-the-art vision-language models and five fine-tuning techniques, we establish a strong linear relationship: tasks with higher IIMM scores yield greater in-domain performance improvements but suffer from more pronounced out-of-domain degradation, with some parameter-efficient fine-tuning (PEFT) methods exhibiting severe forgetting. Compared to existing transferability measures, the IIMM demonstrates significantly stronger predictive power for accuracy changes post fine-tuning in dual-encoder models. Moreover, we provide a theoretical bound, proving that changes in IIMM are limited by the Wasserstein distance between pre- and post-fine-tuning embedding distributions, ensuring its stability and robustness as a predictive measure. With only a single forward pass of the target data, practitioners can leverage this key insight to evaluate the degree to which a model can be expected to improve following fine-tuning. When combined with prior knowledge of a model's performance across diverse tasks, the IIMM further enhances transferability predictions for novel tasks, offering a lightweight yet effective tool for guiding model adaptation strategies. Our code is provided at https://github.com/mit-ll/IIMM.", + "bbox": [ + 88, + 340, + 485, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 821, + 223, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large vision-language foundation models have demonstrated superior zero-shot generalization compared to tra", + "bbox": [ + 89, + 847, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ditional vision models, leveraging web-scale pre-training to achieve state-of-the-art performance across diverse tasks without task-specific supervision [39]. Advances in enhancing zero-shot generalization have also been proposed [31, 41, 57]. Fine-tuning can further enhance performance on specific downstream applications [4, 27, 28, 46, 61]. However, fine-tuned foundation models still struggle on certain tasks, particularly in specialized fields such as satellite and medical image classification [1, 18, 49, 53]. Given the computational cost of fine-tuning, determining whether adaptation will lead to meaningful improvements remains a crucial challenge.", + "bbox": [ + 511, + 310, + 906, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Beyond resource efficiency, fine-tuning also introduces the risk of catastrophic forgetting—where a model loses previously learned knowledge while improving on the new task. This effect is particularly pronounced in dual-encoder vision-language models, where fine-tuning disrupts the alignment between image and text embedding spaces, degrading the model's ability to generalize to out-of-domain tasks [8, 36]. This creates an inherent trade-off between learning and forgetting, making it critical for practitioners to anticipate the effects of fine-tuning before allocating computational resources [32, 58, 63]. A predictive metric that captures both learning gains and forgetting risks would provide a valuable tool for optimizing fine-tuning decisions.", + "bbox": [ + 511, + 492, + 908, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The challenge of estimating a model's fine-tuned performance a priori falls under the domain of model transferability [38, 51]. Common transferability measures assess a pretrained model's suitability for a given task by analyzing its latent representations. However, most traditional methods rely on source data and task labels, making them incompatible with contrastively trained foundation models that lack explicit supervision [43]. Furthermore, newer transferability metrics are designed for uni-modal settings, failing to account for inter-modal interactions in vision-language models. Among approaches that estimate post-fine-tuning performance, evaluation is often limited to rank correlation rather than directly predicting performance changes [17, 42].", + "bbox": [ + 511, + 689, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 886, + 220, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2396", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3f2363d156a0bd9ba19e94948c4788ac2a0d7bb192676307db92febb5ada5b9c.jpg", + "image_caption": [ + "Figure 1. The Inter-Intra Modal Measure (IIMM) combines two key aspects of embedding geometry: intra-modal image embedding similarity, and inter-modal alignment between image and text embeddings. This measure captures information about the embedding geometry that demonstrates a strong relationship to performance gains observed with fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 197, + 88, + 802, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Present work To address these limitations, we introduce the Inter-Intra Modal Measure (IIMM), a novel metric that quantifies both intra-modal representation spread and inter-modal alignment in contrastive vision-language models, capturing the fundamental trade-off between learning and forgetting without the need for fine-tuning. We compare our measure against several standard and recent transferability measures in both rank correlation and strength of predictive fit to performance gains after fine-tuning. To evaluate our measure, we perform fine-tuning across nine well-known vision classification tasks using four state-of-the-art vision-language models and five fine-tuning techniques, analyzing in-domain improvements as well as induced catastrophic forgetting across eight out-of-domain tasks. We compare our measure against six transferability measures applicable to vision-language encoders, and show that our measure has a significantly stronger monotonic relationship with both fine-tuned in-domain accuracy and in-domain performance gains scaled by the zero-shot error. When fitting a simple linear regression of the IIMM to the scaled in-domain accuracy gains, we achieve $R^2 > 0.85$ for all base models. This offers a practical and theoretically grounded tool for model adaptation in vision-language tasks.", + "bbox": [ + 91, + 428, + 483, + 775 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions In summary, our work provides the following contributions:", + "bbox": [ + 89, + 794, + 482, + 824 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce the Inter-Intra Modal Measure (IIMM), a standalone metric that integrates inter-modal misalignment and intra-modal uniformity to predict fine-tuning outcomes in dual-encoder vision-language models.", + "2. We derive the IIMM directly from contrastive learning" + ], + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "principles, establishing a clear connection between the geometry of the embedding space and transferability estimation.", + "bbox": [ + 531, + 428, + 906, + 470 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "3. We prove a theoretical bound showing that changes in the IIMM are constrained by the 1-Wasserstein distance between pre- and post-fine-tuning embedding distributions. This result guarantees that significant shifts in the representation space will lead to substantial changes in IIMM, underscoring its stability as a predictive measure.", + "4. We show that the zero-shot IIMM has a strong, positive linear relationship with learning and a negative relationship with forgetting post fine-tuning. We also provide empirical evidence that for any vision-language dual-encoder, fine-tuning on only a few standard vision benchmarks is sufficient to fit a strong linear predictor of fine-tuning accuracy gains with the IIMM for transfer estimation of novel tasks." + ], + "bbox": [ + 511, + 473, + 908, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 705, + 651, + 722 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dual-encoders and the modality gap Multi-modal dual-encoders trained using contrastive learning consist of two models jointly trained to map each modality into a shared representation space [19, 25, 40, 54, 62]. Previous research has highlighted the existence of a modality gap within this joint embedding space that results from the encoders initializing embeddings within random cones, and that contrastive loss maintains this initialized gap [26]. Additionally, empirical evidence shows that manipulating this gap can impact zero-shot performance. While catastrophic forgetting has been proven to significantly affect traditional su", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2397", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pervised learning techniques [20, 30], recent research has shown that self-supervised, uni-modal models do not suffer from such severe forgetting [15, 35]. Yet, with multimodal foundation models, such as CLIP, continuous training can cause misalignment of the image and text embeddings, an outcome highlighted and explored in several works [8, 11, 29, 36, 44].", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transferability estimation A significant amount of research has focused on addressing transferability estimation with respect to uni-modal encoders. Estimation of model transferability often involves transferability metrics, or measures that predict how well knowledge learned from source tasks will transfer to new target tasks. [9] The approaches for quantifying these metrics vary greatly, with some approaches simply defining them as the log-likelihood of a target task. Methods that rely upon a log-likelihood definition are inherently well-suited for model selection because the definition relies directly upon the size of the target dataset, and within model selection the target dataset is held constant. Regardless of approach, nearly all metrics are evaluated via rank correlation across a model set [17, 42, 43, 51].", + "bbox": [ + 89, + 218, + 483, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some modern approaches include training small neural networks and comparing the similarity of resultant task embeddings [10], training a full neural network on large amounts of models trained across many tasks to learn a similarity score function [24], and comparing source training data to new target task data to estimate similarity [7]. Some classic methods, such as LEEP [34] and NCE [48], rely upon information from the source training data and are thus inapplicable to self-supervised foundation models; they also cannot be applied to dual-encoders. However, there are a range of uni-modal methods that can be easily applied to the vision encoder of a dual-encoder model [17, 43, 56].", + "bbox": [ + 89, + 445, + 483, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For multi-modal models, some recent research has involved zero-shot model selection using language to infer the transferability of a dataset, such as LOVM [64] and SWAB [55]. These methods do not claim to estimate model performance after fine-tuning, and are specific to the zero-shot setting.", + "bbox": [ + 89, + 627, + 483, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Inter-Intra Modal Measure", + "text_level": 1, + "bbox": [ + 89, + 732, + 344, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we provide a principled derivation of the IIMM metric from contrastive learning principles. In addition, we prove a theoretical bound and provide intuition behind the metric for predicting fine-tuning outcomes.", + "bbox": [ + 89, + 758, + 483, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notation Let $t$ denote a target task, and let $\\mathcal{T}$ be the set of tasks under evaluation. Consider a training dataset for task $t$ composed of image-text pairs. Let $X$ denote the set of image embeddings, where each $x \\in X \\subset \\mathbb{R}^d$ . Similarly, let", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$Y$ denote the set of text embeddings, where each $y \\in Y \\subset \\mathbb{R}^d$ . For every image $x$ , let $y(x) \\in Y$ be the text embedding corresponding to its correct label.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mu$ and $\\nu$ be two probability measures on a metric space $(\\mathcal{X},d)$ . The 1-Wasserstein distance between $\\mu$ and $\\nu$ is defined as:", + "bbox": [ + 511, + 136, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW _ {1} (\\mu , \\nu) = \\inf _ {\\gamma \\in \\Gamma (\\mu , \\nu)} \\int_ {\\mathcal {X} \\times \\mathcal {X}} d (x, y) d \\gamma (x, y),\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 185, + 859, + 217 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Gamma (\\mu ,\\nu)$ denotes the set of all couplings (i.e., joint distributions) with marginals $\\mu$ and $\\nu$", + "bbox": [ + 511, + 223, + 903, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Derivation via Contrastive Learning Principles", + "text_level": 1, + "bbox": [ + 511, + 261, + 903, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin with the widely used InfoNCE loss in contrastive learning [37]. For an image embedding $x \\in X$ and its corresponding (correct) text embedding $y \\in Y$ , the loss is defined as", + "bbox": [ + 511, + 282, + 903, + 340 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL (x, y, \\{y ^ {\\prime} \\}) = - \\log \\frac {\\exp \\left(\\frac {x ^ {\\top} y}{\\tau}\\right)}{\\exp \\left(\\frac {x ^ {\\top} y}{\\tau}\\right) + \\sum_ {y ^ {\\prime} \\in Y \\backslash \\{y (x) \\}} \\exp \\left(\\frac {x ^ {\\top} y ^ {\\prime}}{\\tau}\\right)},\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 345, + 942, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tau > 0$ is the temperature parameter and the summation is over all incorrect text embeddings $y'$ (i.e. $y' \\neq y(x)$ ). Minimizing the InfoNCE loss encourages two main effects:", + "bbox": [ + 511, + 411, + 903, + 455 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Positive Pair Alignment: The numerator is maximized when the cosine similarity $x^{\\top}y$ is high.", + "2. Negative Pair Separation: The denominator penalizes high similarities between $x$ and any incorrect text embedding $y'$ , thus pushing down $x^\\top y'$ for $y' \\neq y(x)$ ." + ], + "bbox": [ + 513, + 455, + 903, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The resulting changes in embedding geometry can be decomposed into two components: inter-modal misalignment and intra-modal uniformity.", + "bbox": [ + 511, + 532, + 903, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inter-modal misalignment The effect on incorrect pairs can be quantified by the average similarity between an image embedding $x$ and the set of incorrect text embeddings:", + "bbox": [ + 511, + 594, + 903, + 640 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\text {i n t e r}} = \\frac {1}{| X |} \\sum_ {x \\in \\mathcal {X}} \\frac {1}{| Y | - 1} \\sum_ {y ^ {\\prime} \\in Y \\backslash \\{y (x) \\}} x ^ {\\top} y ^ {\\prime}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 648, + 903, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A high $S_{\\mathrm{inter}}$ indicates that the model has not sufficiently separated the image embedding from non-matching text embeddings. While indicative of potential for improvement, this can also be problematic, as fine-tuning may lead to over-alignment and subsequent catastrophic forgetting.", + "bbox": [ + 511, + 691, + 903, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intra-modal uniformity In addition to inter-modal alignment, contrastive learning also encourages a good spread, or uniformity, of embeddings within each modality [50]. For image embeddings, we quantify uniformity by computing the average pairwise similarity:", + "bbox": [ + 511, + 784, + 906, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\text {i n t r a}} = \\frac {2}{| X | (| X | - 1)} \\sum_ {1 \\leq i < j \\leq | X |} x _ {i} ^ {\\top} x _ {j}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 864, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2398", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A high $S_{\\mathrm{intra}}$ implies that the image embeddings are overly clustered, suggesting that there is significant potential for further separation (or re-organization) during fine-tuning.", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formulation and intuition Based on the above principles, we define the Inter-Intra Modal Measure (IIMM) as the average of the inter-modal and intra-modal terms:", + "bbox": [ + 89, + 154, + 483, + 200 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {I I M M} = \\frac {1}{2} \\left(S _ {\\text {i n t e r}} + S _ {\\text {i n t r a}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 210, + 483, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A high IIMM value indicates that the average similarity between an image and incorrect text embeddings is high and/or the image embeddings are highly clustered. This situation suggests that while there is significant capacity for learning (i.e., re-aligning the modalities), the risk of catastrophic forgetting is also high. Conversely, a low IIMM value implies that the image and text modalities are well separated (low inter-modal similarity for negatives) and the image embeddings are sufficiently dispersed. In this case, fine-tuning is expected to yield minimal performance changes.", + "bbox": [ + 89, + 250, + 483, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The IIMM metric arises naturally from the objectives of contrastive learning:", + "bbox": [ + 89, + 416, + 483, + 445 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. The InfoNCE loss drives positive pairs to be closely aligned and negative pairs to be well separated.", + "2. The average similarity over incorrect pairs, $S_{\\mathrm{inter}}$ , captures the degree of inter-modal misalignment.", + "3. The average similarity among image embeddings, $S_{\\mathrm{intra}}$ , measures intra-modal uniformity." + ], + "bbox": [ + 89, + 446, + 482, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By combining these two measures as in Equation (4), the IIMM provides a principled, predictive metric that reflects both the learning potential and the risk of catastrophic forgetting in vision-language models.", + "bbox": [ + 89, + 537, + 483, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We provide further exploration of this measure and its relation to learning, forgetting, and the uniformity of the embedding space in Sec. 4.2. Additionally, ?? provides an analysis of various convex combinations of the intra- and inter-modal sub-scores, highlighting modification of component weights as a means of evaluating each component's contributions to the predicted outcomes.", + "bbox": [ + 89, + 597, + 483, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Theoretical Bound", + "text_level": 1, + "bbox": [ + 89, + 712, + 272, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we provide a theoretical result that links the change in IIMM to the Wasserstein distance between the joint distributions of incorrect image-text pairs (and image-image pairs) before and after fine-tuning. This bound offers a rigorous guarantee: if fine-tuning only induces a small distributional shift as measured by the Wasserstein distance, then the IIMM metric will change by no more than a proportional amount. Such a result is important as it justifies the use of IIMM as a robust indicator—ensuring that minor, perhaps noisy, changes in the embeddings do not lead to erratic changes in the IIMM, and that significant changes in", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the IIMM are indeed indicative of a meaningful shift in the model's representation.", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. Assume that all image and text embeddings lie on the unit sphere $S^{d - 1} \\subset \\mathbb{R}^d$ . Let the IIMM metric be defined by", + "bbox": [ + 511, + 130, + 906, + 176 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {I I M M} (P) = \\frac {1}{2} \\Big (A (P) + B (P) \\Big),\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 186, + 823, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 224, + 558, + 238 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA (P) = \\mathbb {E} _ {(x, y ^ {\\prime}) \\sim P _ {X Y}} \\left[ x ^ {\\top} y ^ {\\prime} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 237, + 808, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $P_{XY}$ being the joint distribution of incorrect image-text pairs, and", + "bbox": [ + 511, + 262, + 906, + 292 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nB (P) = \\mathbb {E} _ {(x, x ^ {\\prime}) \\sim P _ {X} \\otimes P _ {X}} \\left[ x ^ {\\top} x ^ {\\prime} \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 301, + 821, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $P_{X}$ being the marginal distribution of image embeddings. Consider perturbed versions of these distributions $Q_{XY}$ and $Q_{X}$ , respectively. Assume that the 1-Wasserstein distances satisfy", + "bbox": [ + 511, + 332, + 905, + 392 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {1} \\left(P _ {X Y}, Q _ {X Y}\\right) \\leq \\delta_ {A} \\quad a n d \\quad W _ {1} \\left(P _ {X}, Q _ {X}\\right) \\leq \\delta_ {B}.\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 402, + 883, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then the following bound holds:", + "bbox": [ + 513, + 431, + 730, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\operatorname {I I M M} (Q) - \\operatorname {I I M M} (P) \\right| \\leq \\frac {\\delta_ {A}}{2} + \\delta_ {B}.\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 457, + 836, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, if one computes empirical estimates $\\widehat{\\mathrm{IMM}}(P)$ and $\\widehat{\\mathrm{IMM}}(Q)$ using $N$ independent samples, then with probability at least $1 - \\eta$ ,", + "bbox": [ + 511, + 494, + 905, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\widehat {\\operatorname {I M M}} (Q) - \\widehat {\\operatorname {I M M}} (P) \\right| \\leq \\frac {\\delta_ {A}}{2} + \\delta_ {B} + \\epsilon_ {N},\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 551, + 857, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 592, + 558, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon_ {N} = O \\Big (\\sqrt {\\frac {\\log (1 / \\eta)}{N}} \\Big)\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 603, + 785, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "accounts for the sampling error.", + "bbox": [ + 511, + 641, + 728, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our empirical results have indicated that IIMM is strongly correlated with performance gains and catastrophic forgetting. The bound provided here offers a theoretical explanation: only significant reorganization of the embedding space (reflected by a large Wasserstein distance) can cause a substantial shift in IIMM. Therefore, a high IIMM change can be interpreted as evidence of a considerable shift in the model's internal representation, likely correlating with notable performance changes.", + "bbox": [ + 511, + 665, + 906, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 815, + 645, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets To evaluate the relationships between transferability measures and model performance changes following fine-tuning, we compiled a diverse collection of nine commonly used, publicly available computer vision benchmarks", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2399", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "provided by PyTorch [26, 39, 64] as our testing dataset, with the goal of minimizing domain overlap and covering a range of coarse and granular classification tasks. These datasets included Stanford Cars [21], CIFAR100 [22], the Describable Textures Dataset (DTD) [6], satellite imagery datasets EuroSAT [14] and RESISC45 [5], the German Traffic Sign Recognition Benchmark (GTSRB) [45], MNIST [23], the Scene Understanding Benchmark (SUN397) [52], and the Street View House Numbers (SVHN) dataset [33]. For each dataset, we generated validation splits by randomly sampling $10\\%$ of the respective training data, capping the validation split size at 5000 samples for larger datasets.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models and fine-tuning methods To investigate the applicability of IIMM to vision-language encoders, we conducted all experiments using two distinct groups of evaluations schema: one in which the base pre-trained model varies while the fine-tuning strategy is held constant, and another group in which the base pre-trained model is held constant while the fine-tuning strategy varies. We will refer to these evaluation schema as the base model evaluation and PEFT evaluation, respectively.", + "bbox": [ + 89, + 291, + 480, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the base model evaluation, we perform full finetuning over all parameters for the following four pre-trained models: CLIP, EVA-02-CLIP [12], SigLIP [60], and CoCa [59], using a ViT-B/32 image backbone for CLIP and CoCa and a ViT-B/16 image backbone for SigLIP and EVA-02-CLIP.", + "bbox": [ + 89, + 428, + 480, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the PEFT evaluation, we utilized a CLIP model with a ViT-B/32 image backbone as the base model and evaluate four common parameter-efficient fine-tuning (PEFT) methods: CLIP-Adapter [13], BitFit (bias tuning) [3], LoRA [16], and a method we call Attention-Weight Tuning, which simply refines the fine-tuning of all attention layer weights in Touvron et al. [47] to just the attention layers' inprojection weights.", + "bbox": [ + 89, + 518, + 480, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines While no methods currently exist for explicitly predicting gains in accuracy from fine-tuning for dual-encoder vision-language models, many related methods exist for stand-alone vision encoders. In particular, we compare against several recent or established measures self-described as transferability estimators. The six comparison measure are: TransRate [17], Self-challenging Fisher Discriminant Analysis (SFDA) [42], Neural Collapse informed Transferability Index (NCTI) [51], Efficient Multitask Model Selector EMMS [43], Gaussian Bhattacharyya Coefficient (GBC) [38], and H-score [2].", + "bbox": [ + 89, + 657, + 480, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We chose these comparison measures to span the definitions and applications of self-described transferability measures. Those that estimate the log-likelihood are H-Score, SFDA, TransRate, and EMMS. NCTI functions as an estimator of the gap between the embedding geometry of the", + "bbox": [ + 89, + 825, + 480, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pre-trained model and the embedding geometry of a hypothetical fine-tuned model [51]. The GBC quantifies transferability as a measure of overlap in the feature space between embedding clusters of different class labels [38].", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation measures We evaluate performance of the transferability measures through the strength of their relationship to three variables: accuracy on the target test data after fine-tuning on the target task training data; gain in performance from the pre-trained model to the fine-tuned model scaled by the pre-trained model error, which we refer to as gain over zero-shot error; and average accuracy across the remaining eight non-target datasets to measure loss in performance from the pre-trained model on nontarget tasks.", + "bbox": [ + 511, + 167, + 903, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Definition 1 (Gain over zero-shot error). For a given task, we define $A_{zs}$ as the zero-shot accuracy and $A_{ft}$ as the fine-tuned accuracy. We then define the gain over zero-shot error as,", + "bbox": [ + 511, + 325, + 903, + 385 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {A _ {f t} - A _ {z s}}{1 - A _ {z s}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 669, + 402, + 903, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This value normalizes for differences in the zero-shot accuracy across models and tasks.", + "bbox": [ + 511, + 439, + 903, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first evaluate all measures' monotonic relationship to both fine-tuned accuracy and gain over zero-shot error using Kendall's $\\tau$ , which is a standard evaluation measure within the literature. We then look at the top performing measures of this rank correlation and fit a least-squares model to evaluate their predictive strength, additionally looking at a least-squares model fit of the top measure against the average accuracy on off-target tasks.", + "bbox": [ + 511, + 474, + 903, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training details To establish the ground truth for changes in model in-domain (ID) and out-of-domain (OOD) accuracy following fine-tuning on each dataset, we started by separately fine-tuning each of the base models on the nine datasets in our dataset suite. Following training on a given dataset, each model was tested on all nine datasets, with in-domain accuracy calculated as its accuracy on the test split of the training dataset and out-of-domain accuracies as the accuracies obtained on the remaining datasets test split. Each fine-tuning run consisted of 30 epochs of gradient descent using an SGD optimizer and a batch size of 128 and incorporated an exhaustive grid search over all relevant model hyper-parameters, examining values $\\{1\\mathrm{e} - 2,1\\mathrm{e} - 3,1\\mathrm{e} - 4,1\\mathrm{e} - 5\\}$ for the learning rate, $\\{1\\mathrm{e} - 3,1\\mathrm{e} - 4,1\\mathrm{e} - 5\\}$ for the weight decay, $\\{2,4,8,16\\}$ for LoRA model rank, and $\\{4,8,16,32\\}$ for CLIP-Adapter reduction. The initial image and text embeddings were then generated per dataset by a forward pass of each dataset's train split through the pre-trained base models and normalizing to unit vectors.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2400", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2923caa44ee53cf31a762f28adf4c5a680595e8747187182d4ca195ed0ddee0a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueτp-valueτp-valueτp-valueτ
IIMM< 10-30.890.0010.830.0020.780.0020.78
GBC0.0020.780.0060.720.0020.780.0010.83
TransRate0.0130.670.0250.610.0130.670.0060.72
EMMS0.0250.610.3580.280.0250.610.1190.44
NCTI0.919-0.060.028-0.590.8180.070.827-0.06
SFDA0.2600.330.919-0.060.1800.390.3580.28
H-Score0.7610.110.002-0.780.0750.500.013-0.67
", + "bbox": [ + 228, + 88, + 767, + 234 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/64ba5033f20bb312374e42b6322741c94dbf9de60e9b60df615a8ef49478beb5.jpg", + "table_caption": [ + "Table 1. Kendall's $\\tau$ correlation of score and gain over zero-shot error. Larger $\\tau$ values are better.*" + ], + "table_footnote": [], + "table_body": "
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueτp-valueτp-valueτp-valueτ
IIMM0.0060.720.0060.720.0130.670.0250.61
GBC0.0250.610.0060.720.0130.670.0130.67
TransRate0.0750.500.0250.610.0450.560.0450.56
EMMS0.0130.670.0250.610.0750.500.0250.61
NCTI1.0000.000.173-0.370.818-0.070.8270.06
SFDA0.0750.500.6120.170.0750.500.2600.33
H-Score0.6120.170.045-0.560.0250.610.075-0.50
", + "bbox": [ + 232, + 273, + 766, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Kendall's $\\tau$ correlation of score and accuracy. Larger $\\tau$ values are better.*", + "bbox": [ + 253, + 433, + 740, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Results", + "text_level": 1, + "bbox": [ + 89, + 473, + 183, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first look at Kendall's $\\tau$ correlation to determine the strength of the monotonic relationship of our proposed and comparison measures to both the fine-tuned gain over zero-shot error and fine-tuned accuracy.", + "bbox": [ + 89, + 501, + 482, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The IIMM and the GBC have significantly stronger monotonic relationships than the other measures for both gain over zero-shot error, as shown in Tab. 1, and accuracy, as shown in Tab. 2. In both tables, the IIMM and the GBC are the most strongly correlated measures and display the greatest significance. Upon further investigation, the IIMM displays a mostly linear relationship with gain over zero-shot error, while the GBC displays a mostly exponential relationship with gain over zero-shot error. This allows us to fit and evaluate a predictive model using least-squares regression of the IIMM with gain over zero-shot error, and the GBC with the log of gain over zero-shot error. The results of these evaluations can be seen in Tab. 3, which shows the significance of the slope coefficient and the $R^2$ value, an estimate of the amount of variation within the data explained by its linear relationship to the measure. The IIMM has a stronger fit across all base pre-trained models compared to the GBC, and indeed shows itself to have a strong linear relationship to gain over zero-shot error with $R^2 > 0.89$ for all base models. Without the ability to capture cross-modal alignment, the compared uni-modal metrics expectedly demonstrate worse performance when com", + "bbox": [ + 89, + 568, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pared to the IIMM, underscoring the value of an integrated vision-language metric.", + "bbox": [ + 511, + 474, + 903, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This positive linear relationship between the IIMM and in-domain gain over zero-shot error is clearly seen in the top plots of Fig. 2. The bottom plots of Fig. 2 show the negative linear relationship of the IIMM with the average change in accuracy across a suite of eight out-of-domain tasks. For all base pre-trained models aside from CLIP, the IIMM also shows a strong linear relationship to forgetting, as evidenced by the significance of the slope coefficient and high $R^2$ values. These relationships hold true across parameter-efficient fine-tuning methods, as shown in Fig. 3. We see with CLIP-Adapter that its overall worse in-domain performance attenuates the strength of the relationship between IIMM and gain over zero-shot error, and predictably, that a lower in-domain performance gain resulted in significantly less catastrophic forgetting. CLIP-Adapter's design — injecting task kernels atop a frozen backbone — results in smaller, localized representational shifts, resulting in weaker in-distribution performance gains while preventing greater out-of-distribution performance loss. Since the IIMM is correlated with near-optimal performance gains, fine-tuning methods that result in sub-optimal gains may yield lower IIMM correlations.", + "bbox": [ + 509, + 505, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "These results provide evidence that the IIMM is strongly correlated with fine-tuning gains over zero-shot error. Un", + "bbox": [ + 511, + 835, + 903, + 867 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "*P-values reflect the likelihood of observing results by random chance; smaller values suggest stronger associations.", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2401", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f093eeb5dde358e31a2e56dd671ebc2af45ff05d95987ba287aba677b32cb306.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueR²p-valueR²p-valueR²p-valueR²
IIMM< 10-30.92< 10-30.89< 10-30.92< 10-30.92
GBC0.0040.720.0060.680.0150.590.0050.71
", + "bbox": [ + 235, + 88, + 761, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Evaluation of linear fit of the IIMM to gain over zero-shot error and the GBC to the log of gain over zero-shot error. The significance of the slope coefficient is given by the p-value, and the strength of the fit is evaluated with the $R^2$ value.*", + "bbox": [ + 89, + 176, + 906, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/217baa5082a3ae5e039d150e69a16730c276d98da8b4e205abc790382876a615.jpg", + "image_caption": [ + "Figure 2. Performance gains and losses versus the IIMM per pre-trained model, with linear regression fit and $96\\%$ confidence intervals. The top plots show the relationship between the on-target (ID) task IIMM to the on-target task performance gains over zero-shot error, while the bottom plots show the relationship between on-target task IIMM and the average performance loss on off-target (OOD) tasks." + ], + "image_footnote": [], + "bbox": [ + 122, + 220, + 872, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "like GBC, the second-highest correlated measure, the IIMM is bounded in $[-1,1]$ and extreme values are informative in isolation.", + "bbox": [ + 89, + 599, + 482, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Exploring embedding features", + "text_level": 1, + "bbox": [ + 89, + 662, + 359, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To empirically validate the provided theoretical justification for leveraging multi-modal information to improve transferability estimation in vision-language models, we independently evaluated all relevant components of CLIP's embedding geometry, known to manifest as a hypersphere. Using (3), we separately computed the average pairwise similarities for the image and text modalities and fit each to the accuracy gain over zero-shot error. We then applied (2) to compute the degree of incorrect label alignment across the hypersphere and also fit the measure to the accuracy gain over zero-shot error. The final results of these evaluations can be seen in Tab. 4.", + "bbox": [ + 88, + 686, + 482, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is observed that the features of the embedding structure most predictive of accuracy gains are the intra-images dis", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tance and the incorrect label alignment. Based on these results, the correct label alignment and intra-texts distance are generally uninformative of accuracy gains with fine-tuning. As described in Sec. 3, each of these terms corresponds to a feature of the shared embedding space related to learning and forgetting. We include additional results of the Pearson correlations between different convex combinations of our inter- and intra- terms and gain over zero-shot error in the supplementary material. This empirical evidence successfully verifies that the terms combined have a stronger relationship with learning and forgetting than they do independently.", + "bbox": [ + 511, + 599, + 906, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 797, + 633, + 814 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this work, we propose the Inter-Intra Modal Measure (IIMM) as a novel metric for predicting fine-tuning performance in vision-language models. We demonstrate that the IIMM strongly correlates with both learning gains and catastrophic forgetting, offering a simple yet powerful met", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2402", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eeff330e2b2dafa96fe48e0df5e5d8b56bba36b7bca4780eb94cf566c76799ec.jpg", + "image_caption": [ + "Figure 3. Performance gains and losses versus the IIMM per PEFT method, with linear regression fit and $96\\%$ confidence intervals. The top plots show the relationship between the on-target task (ID) IIMM and on-target task performance gains over zero-shot error, while the bottom plots show the relationship between on-target task IIMM and the average performance loss on off-target (OOD) tasks." + ], + "image_footnote": [], + "bbox": [ + 122, + 88, + 872, + 385 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2ea7ed7fbfc58db062c7eb009f92dd74bf02e3a43be66110676af96d6bd23dad.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueR²p-valueR²p-valueR²p-valueR²
Incorrect Alignment0.0010.820.0040.720.0120.620.0010.80
Intra-Images Distance< 10-30.88< 10-30.860.0010.81< 10-30.88
Intra-Texts Distance0.0840.370.0340.500.0190.570.2650.17
", + "bbox": [ + 192, + 454, + 805, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Linear fit of measure to gain over zero-shot error with full fine-tuning by pre-trained model. The significance of the slope coefficient is given by the p-value, and the $R^2$ values represent the proportion of variation within the data modeled by the linear relationship.", + "bbox": [ + 88, + 558, + 906, + 587 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ric for estimating fine-tuning outcomes. Because the IIMM is bounded, an extreme value in isolation can inform practitioners whether a model will have limited or significant gains from fine-tuning. In addition, our results show that the IIMM can be leveraged to fit predictive linear models for estimating fine-tuning accuracy gains across diverse tasks with minimal computational overhead. We demonstrate that our proposed measure has a stronger relationship to both fine-tuned accuracy and accuracy changes with fine-tuning than baselines. We further show the IIMM is informative of performance changes across parameter efficient fine-tuning methods, highlighting some methods are particularly susceptible to catastrophic forgetting when the zero-shot IIMM is high. Beyond empirical validation, we provide a theoretical bound linking changes in IIMM to the Wasserstein distance between pre- and post-fine-tuning embedding distributions, ensuring stability in IIMM predictions with respect to changes in representation space. Given the IIMM measures changes in embedding geometry prior to any task", + "bbox": [ + 88, + 612, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "specific heads, the metric is also task-agnostic, enabling its application to a variety of downstream tasks. We encourage future research to validate such use in broader applications. Overall, our work provides a practical and theoretically grounded tool for optimizing fine-tuning decisions in vision-language models, paving the way for more efficient and informed adaptation of foundation models to new tasks.", + "bbox": [ + 511, + 612, + 908, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations It is possible that the IIMM could show more variation when compared across a more diverse set of base models. While extreme values of the IIMM are heuristically informative in isolation, fitting a model for transfer accuracy estimation requires first fine-tuning across a small but diverse set of benchmarks. This could be an avenue of future research, as there might be additional information within the embedding geometry that allows for fine-tuning performance prediction without requiring information from fine-tuning on other tasks.", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2403", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 256, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors acknowledge the MIT Lincoln Laboratory Supercomputing Center for providing the high-performance computing resources that have contributed to the research results reported in this paper.", + "bbox": [ + 89, + 114, + 480, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "DISTRIBUTION STATEMENT A. Approved for public release. Distribution is unlimited. This material is based upon work supported by the Under Secretary of Defense for Research and Engineering under Air Force Contract No. FA8702-15-D-0001 or FA8702-25-D-B002. Any opinions, findings, conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Under Secretary of Defense for Research and Engineering.", + "bbox": [ + 89, + 175, + 482, + 311 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 325, + 186, + 340 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational Models Defining a New Era in Vision: A Survey and Outlook, 2023. 1", + "[2] Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An Information-Theoretic Approach to Transferability in Task Transfer Learning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2309-2313, Taipei, Taiwan, 2019. IEEE. 5", + "[3] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1-9, Dublin, Ireland, 2022. Association for Computational Linguistics. 5", + "[4] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1", + "[5] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 5", + "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing Textures in the Wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, Columbus, OH, USA, 2014. IEEE. 5", + "[7] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large Scale Fine-Grained Categorization and Domain-Specific Transfer Learning, 2018. 3", + "[8] Yuxuan Ding, Lingqiao Liu, Chunna Tian, Jingyuan Yang, and Haoxuan Ding. Don't Stop Learning: Towards Continual Learning for the CLIP Model, 2022. 1, 3", + "[9] Yuhe Ding, Bo Jiang, Aijing Yu, Aihua Zheng, and Jian Liang. Which model to transfer? a survey on transferability estimation. arXiv preprint arXiv:2402.15231, 2024. 3" + ], + "bbox": [ + 99, + 349, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Kshitij Dwivedi and Gemma Roig. Representation Similarity Analysis for Efficient Task Taxonomy & Transfer Learning. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12379-12388, Long Beach, CA, USA, 2019. IEEE. 3", + "[11] Zhihao Fan, Zhongyu Wei, Jingjing Chen, Siyuan Wang, Ze-jun Li, Jiarong Xu, and Xuanjing Huang. A unified continuous learning framework for multi-modal knowledge discovery and pre-training, 2022. 3", + "[12] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331, 2023.5", + "[13] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. CLIP-Adapter: Better Vision-Language Models with Feature Adapters, 2021. 5", + "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Introducing Eurosat: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. In IGARSS 2018 - 2018 IEEE International Geoscience and Remote Sensing Symposium, pages 204-207, Valencia, 2018. IEEE. 5", + "[15] Dapeng Hu, Shipeng Yan, Qizhengqiu Lu, Lanqing Hong, Hailin Hu, Yifan Zhang, Zhenguo Li, Xinchao Wang, and Jiashi Feng. How well does self-supervised pre-training perform with streaming data?, 2022. 3", + "[16] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In ICLR. arXiv, 2022. 5", + "[17] Long-Kai Huang, Junzhou Huang, Yu Rong, Qiang Yang, and Ying Wei. Frustratingly easy transferability estimation. In International Conference on Machine Learning, pages 9201-9225. PMLR, 2022. 1, 3, 5", + "[18] Joana Palés Huix, Adithya Raju Ganeshan, Johan Fredin Haslum, Magnus Söderberg, Christos Matsoukas, and Kevin Smith. Are Natural Domain Foundation Models Useful for Medical Image Classification? In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 7619-7628. IEEE, 2024. 1", + "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2", + "[20] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 3", + "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D Object Representations for Fine-Grained Categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, Sydney, Australia, 2013. IEEE. 5" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2404", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Alex Krizhevsky. Learning Multiple Layers of Features from Tiny Images, 2009. 5", + "[23] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, Nov./1998. 5", + "[24] Hao Li, Charless Fowlkes, Hao Yang, Onkar Dabeer, Zhuowen Tu, and Stefano Soatto. Guided Recommendation for Model Fine-Tuning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3633-3642, Vancouver, BC, Canada, 2023. IEEE. 3", + "[25] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before Fuse: Vision and Language Representation Learning with Momentum Distillation. In Advances in Neural Information Processing Systems, pages 9694-9705. Curran Associates, Inc., 2021. 2", + "[26] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning. In First Workshop of Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022, 2022. 2, 5", + "[27] C. Liao, T. Tsiligkaridis, and B. Kulis. Descriptor and word soups: Overcoming the parameter efficiency accuracy trade-off for out-of-distribution few-shot learning. In CVPR, 2024. 1", + "[28] C. Liao, C. So, T. Tsiligkaridis, and B. Kulis. Multimodal unsupervised domain generalization by retrieving across the modality gap. In ICLR, 2025. 1", + "[29] Yong Lin, Lu Tan, Hangyu Lin, Zeming Zheng, Renjie Pi, Jipeng Zhang, Shizhe Diao, Haoxiang Wang, Han Zhao, Yuan Yao, and Tong Zhang. Speciality vs Generality: An Empirical Study on Catastrophic Forgetting in Fine-tuning Foundation Models, 2023. 3", + "[30] Michael McCloskey and Neal J. Cohen. Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem. In *Psychology of Learning and Motivation*, pages 109–165. Elsevier, 1989. 3", + "[31] S. Menon and C. Vondrick. Visual classification via description from large language models. In ICLR, 2023. 1", + "[32] Jishnu Mukhoti, Yarin Gal, Philip HS Torr, and Puneet K Dokania. Fine-tuning can cripple your foundation model; preserving features may be the solution. arXiv preprint arXiv:2308.13320, 2023. 1", + "[33] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading Digits in Natural Images with Unsupervised Feature Learning. NeurIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 5", + "[34] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. LEEP: A New Measure to Evaluate Transferability of Learned Representations. In Proceedings of the 37th International Conference on Machine Learning, pages 7294-7305. PMLR, 2020. 3", + "[35] Zixuan Ni, Siliang Tang, and Yueting Zhuang. Self-supervised class incremental learning, 2021. 3", + "[36] Zixuan Ni, Longhui Wei, Siliang Tang, Yueting Zhuang, and Qi Tian. Continual Vision-Language Representation Learning with Off-Diagonal Information, 2023. 1, 3" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3", + "[38] Michal Páncy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Transferability Estimation using Bhattacharyya Class Separability, 2022. 1, 5", + "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 5", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2", + "[41] K. Roth, J. M. Kim, A. S. Koepke, O. Vinyals, C. Schmid, and Z. Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 1", + "[42] Wenqi Shao, Xun Zhao, Yixiao Ge, Zhaoyang Zhang, Lei Yang, Xiaogang Wang, Ying Shan, and Ping Luo. Not all models are equal: Predicting model transferability in a self-challenging fisher space. In European Conference on Computer Vision, pages 286-302. Springer, 2022. 1, 3, 5", + "[43] Wenqi Shao, Chonghe Jiang, Kaipeng Zhang, Yu Qiao, Ping Luo, et al. Foundation model is efficient multimodal multitask model selector. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 5", + "[44] Tejas Srinivasan, Ting-Yun Chang, Leticia Pinto Alva, Georgios Chochlakis, Mohammad Rostami, and Jesse Thomason. CLiMB: A continual learning benchmark for vision-and-language tasks. In Advances in Neural Information Processing Systems, pages 29440-29453. Curran Associates, Inc., 2022. 3", + "[45] J. Stallkamp, M. Schlipsing, J. Salmen, and C. Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. *Neural Networks*, 32:323–332, 2012. 5", + "[46] P. Teterwak, K. Saito, T. Tsiligkaridis, B. A. Plummer, and K. Saenko. Is large-scale pretraining the secret to good domain generalization? In ICLR, 2025. 1", + "[47] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about Vision Transformers, 2022. 5", + "[48] Anh Tran, Cuong Nguyen, and Tal Hassner. Transferability and Hardness of Supervised Classification Tasks. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1395-1405, Seoul, Korea (South), 2019. IEEE. 3", + "[49] Kevin Vogt-Lowell, Noah Lee, Theodoros Tsiligkaridis, and Marc Vaillant. Robust fine-tuning of vision-language models for domain generalization. In 2023 IEEE High Performance" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2405", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Extreme Computing Conference (HPEC), pages 1-7. IEEE, 2023. 1", + "[50] Tongzhou Wang and Phillip Isola. Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere. In Proceedings of the 37th International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 3", + "[51] Zijian Wang, Yadan Luo, Liang Zheng, Zi Huang, and Mahsa Baktashmotlagh. How far pre-trained models are from neural collapse on the target dataset informs their transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5549-5558, 2023. 1, 3, 5", + "[52] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, San Francisco, CA, USA, 2010. IEEE. 5", + "[53] Yiqun Xie, Zhihao Wang, Weiye Chen, Zhili Li, Xiaowei Jia, Yanhua Li, Ruichen Wang, Kangyang Chai, Ruohan Li, and Sergii Skakun. When are foundation models effective? understanding the suitability for pixel-level classification using multispectral imagery. CoRR, abs/2404.11797, 2024. 1", + "[54] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive Pretraining for Zero-shot Video-Text Understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787–6800, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2", + "[55] Chao Yi, De-Chuan Zhan, and Han-Jia Ye. Bridge the Modality and Capacity Gaps in Vision-Language Model Selection, 2024. 3", + "[56] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In International Conference on Machine Learning, pages 12133-12143. PMLR, 2021. 3", + "[57] E. Yu, C. Liao, S. Ravi, T. Tsiligkaridis, and B. Kulis. Image-capture encoding for improving zero-shot generalization. In WACV, 2025. 1", + "[58] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23219-23230. IEEE. 1", + "[59] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 5", + "[60] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 5", + "[61] Dan Zhang, Tao Feng, Lilong Xue, Yuandong Wang, Yuxiao Dong, and Jie Tang. Parameter-efficient fine-tuning for" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "foundation models. arXiv preprint arXiv:2501.13787, 2025. 1", + "[62] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and Curtis P. Langlotz. Contrastive Learning of Medical Visual Representations from Paired Images and Text. In Proceedings of the 7th Machine Learning for Healthcare Conference, pages 2-25. PMLR, 2022. 2", + "[63] Zangwei Zheng, Mingyuan Ma, Kai Wang, Ziheng Qin, Xi-angyu Yue, and Yang You. Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 19068-19079. IEEE. 1", + "[64] Orr Zohar, Kuan-Chieh Wang, Shih-Cheng Huang, and Serena Yeung. LOVM: Language-Only Vision Model Selection. In Conference on Neural Information Processing Systems (Neurips), 2023. 3, 5" + ], + "bbox": [ + 516, + 92, + 903, + 318 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2406", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/f4964a00-88d3-4539-871c-39c0ed82d2c6_model.json b/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/f4964a00-88d3-4539-871c-39c0ed82d2c6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5ff9cfa7896b8a615b88e20af31a4ba35ed22287 --- /dev/null +++ b/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/f4964a00-88d3-4539-871c-39c0ed82d2c6_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.13, + 0.903, + 0.177 + ], + "angle": 0, + "content": "The Inter-Intra Modal Measure: A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.256, + 0.204, + 0.744, + 0.222 + ], + "angle": 0, + "content": "Laura Niss* Kevin Vogt-Lowell* Theodoros Tsiligkaridis" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.224, + 0.721, + 0.239 + ], + "angle": 0, + "content": "{laura.niss, kevin.vogt-lowell, ttsili}@ll.mit.edu" + }, + { + "type": "text", + "bbox": [ + 0.399, + 0.24, + 0.598, + 0.257 + ], + "angle": 0, + "content": "MIT Lincoln Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.257, + 0.641, + 0.275 + ], + "angle": 0, + "content": "244 Wood St, Lexington, MA 02421" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.486, + 0.795 + ], + "angle": 0, + "content": "The fine-tuning of large vision-language foundation models remains an underexplored area, particularly regarding its impact on learning gains and catastrophic forgetting. Inspired by the significance of modality gaps in contrastive dual-encoders, we introduce the Inter-Intra Modal Measure (IIMM)—a predictive metric that quantifies the relationship between intra-modal image embedding similarity and inter-modal misalignment. Through extensive empirical analysis across four state-of-the-art vision-language models and five fine-tuning techniques, we establish a strong linear relationship: tasks with higher IIMM scores yield greater in-domain performance improvements but suffer from more pronounced out-of-domain degradation, with some parameter-efficient fine-tuning (PEFT) methods exhibiting severe forgetting. Compared to existing transferability measures, the IIMM demonstrates significantly stronger predictive power for accuracy changes post fine-tuning in dual-encoder models. Moreover, we provide a theoretical bound, proving that changes in IIMM are limited by the Wasserstein distance between pre- and post-fine-tuning embedding distributions, ensuring its stability and robustness as a predictive measure. With only a single forward pass of the target data, practitioners can leverage this key insight to evaluate the degree to which a model can be expected to improve following fine-tuning. When combined with prior knowledge of a model's performance across diverse tasks, the IIMM further enhances transferability predictions for novel tasks, offering a lightweight yet effective tool for guiding model adaptation strategies. Our code is provided at https://github.com/mit-ll/IIMM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.224, + 0.837 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.848, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Large vision-language foundation models have demonstrated superior zero-shot generalization compared to tra" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.908, + 0.492 + ], + "angle": 0, + "content": "ditional vision models, leveraging web-scale pre-training to achieve state-of-the-art performance across diverse tasks without task-specific supervision [39]. Advances in enhancing zero-shot generalization have also been proposed [31, 41, 57]. Fine-tuning can further enhance performance on specific downstream applications [4, 27, 28, 46, 61]. However, fine-tuned foundation models still struggle on certain tasks, particularly in specialized fields such as satellite and medical image classification [1, 18, 49, 53]. Given the computational cost of fine-tuning, determining whether adaptation will lead to meaningful improvements remains a crucial challenge." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.909, + 0.689 + ], + "angle": 0, + "content": "Beyond resource efficiency, fine-tuning also introduces the risk of catastrophic forgetting—where a model loses previously learned knowledge while improving on the new task. This effect is particularly pronounced in dual-encoder vision-language models, where fine-tuning disrupts the alignment between image and text embedding spaces, degrading the model's ability to generalize to out-of-domain tasks [8, 36]. This creates an inherent trade-off between learning and forgetting, making it critical for practitioners to anticipate the effects of fine-tuning before allocating computational resources [32, 58, 63]. A predictive metric that captures both learning gains and forgetting risks would provide a valuable tool for optimizing fine-tuning decisions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.911, + 0.902 + ], + "angle": 0, + "content": "The challenge of estimating a model's fine-tuned performance a priori falls under the domain of model transferability [38, 51]. Common transferability measures assess a pretrained model's suitability for a given task by analyzing its latent representations. However, most traditional methods rely on source data and task labels, making them incompatible with contrastively trained foundation models that lack explicit supervision [43]. Furthermore, newer transferability metrics are designed for uni-modal settings, failing to account for inter-modal interactions in vision-language models. Among approaches that estimate post-fine-tuning performance, evaluation is often limited to rank correlation rather than directly predicting performance changes [17, 42]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.221, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2396" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.198, + 0.089, + 0.803, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.359, + 0.908, + 0.403 + ], + "angle": 0, + "content": "Figure 1. The Inter-Intra Modal Measure (IIMM) combines two key aspects of embedding geometry: intra-modal image embedding similarity, and inter-modal alignment between image and text embeddings. This measure captures information about the embedding geometry that demonstrates a strong relationship to performance gains observed with fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.429, + 0.485, + 0.776 + ], + "angle": 0, + "content": "Present work To address these limitations, we introduce the Inter-Intra Modal Measure (IIMM), a novel metric that quantifies both intra-modal representation spread and inter-modal alignment in contrastive vision-language models, capturing the fundamental trade-off between learning and forgetting without the need for fine-tuning. We compare our measure against several standard and recent transferability measures in both rank correlation and strength of predictive fit to performance gains after fine-tuning. To evaluate our measure, we perform fine-tuning across nine well-known vision classification tasks using four state-of-the-art vision-language models and five fine-tuning techniques, analyzing in-domain improvements as well as induced catastrophic forgetting across eight out-of-domain tasks. We compare our measure against six transferability measures applicable to vision-language encoders, and show that our measure has a significantly stronger monotonic relationship with both fine-tuned in-domain accuracy and in-domain performance gains scaled by the zero-shot error. When fitting a simple linear regression of the IIMM to the scaled in-domain accuracy gains, we achieve \\( R^2 > 0.85 \\) for all base models. This offers a practical and theoretically grounded tool for model adaptation in vision-language tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Contributions In summary, our work provides the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.826, + 0.483, + 0.886 + ], + "angle": 0, + "content": "1. We introduce the Inter-Intra Modal Measure (IIMM), a standalone metric that integrates inter-modal misalignment and intra-modal uniformity to predict fine-tuning outcomes in dual-encoder vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "2. We derive the IIMM directly from contrastive learning" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.429, + 0.907, + 0.472 + ], + "angle": 0, + "content": "principles, establishing a clear connection between the geometry of the embedding space and transferability estimation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.474, + 0.907, + 0.565 + ], + "angle": 0, + "content": "3. We prove a theoretical bound showing that changes in the IIMM are constrained by the 1-Wasserstein distance between pre- and post-fine-tuning embedding distributions. This result guarantees that significant shifts in the representation space will lead to substantial changes in IIMM, underscoring its stability as a predictive measure." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.565, + 0.909, + 0.685 + ], + "angle": 0, + "content": "4. We show that the zero-shot IIMM has a strong, positive linear relationship with learning and a negative relationship with forgetting post fine-tuning. We also provide empirical evidence that for any vision-language dual-encoder, fine-tuning on only a few standard vision benchmarks is sufficient to fit a strong linear predictor of fine-tuning accuracy gains with the IIMM for transfer estimation of novel tasks." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.474, + 0.909, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.707, + 0.652, + 0.723 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Dual-encoders and the modality gap Multi-modal dual-encoders trained using contrastive learning consist of two models jointly trained to map each modality into a shared representation space [19, 25, 40, 54, 62]. Previous research has highlighted the existence of a modality gap within this joint embedding space that results from the encoders initializing embeddings within random cones, and that contrastive loss maintains this initialized gap [26]. Additionally, empirical evidence shows that manipulating this gap can impact zero-shot performance. While catastrophic forgetting has been proven to significantly affect traditional su" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2397" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "pervised learning techniques [20, 30], recent research has shown that self-supervised, uni-modal models do not suffer from such severe forgetting [15, 35]. Yet, with multimodal foundation models, such as CLIP, continuous training can cause misalignment of the image and text embeddings, an outcome highlighted and explored in several works [8, 11, 29, 36, 44]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.219, + 0.485, + 0.445 + ], + "angle": 0, + "content": "Transferability estimation A significant amount of research has focused on addressing transferability estimation with respect to uni-modal encoders. Estimation of model transferability often involves transferability metrics, or measures that predict how well knowledge learned from source tasks will transfer to new target tasks. [9] The approaches for quantifying these metrics vary greatly, with some approaches simply defining them as the log-likelihood of a target task. Methods that rely upon a log-likelihood definition are inherently well-suited for model selection because the definition relies directly upon the size of the target dataset, and within model selection the target dataset is held constant. Regardless of approach, nearly all metrics are evaluated via rank correlation across a model set [17, 42, 43, 51]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.446, + 0.484, + 0.627 + ], + "angle": 0, + "content": "Some modern approaches include training small neural networks and comparing the similarity of resultant task embeddings [10], training a full neural network on large amounts of models trained across many tasks to learn a similarity score function [24], and comparing source training data to new target task data to estimate similarity [7]. Some classic methods, such as LEEP [34] and NCE [48], rely upon information from the source training data and are thus inapplicable to self-supervised foundation models; they also cannot be applied to dual-encoders. However, there are a range of uni-modal methods that can be easily applied to the vision encoder of a dual-encoder model [17, 43, 56]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.719 + ], + "angle": 0, + "content": "For multi-modal models, some recent research has involved zero-shot model selection using language to infer the transferability of a dataset, such as LOVM [64] and SWAB [55]. These methods do not claim to estimate model performance after fine-tuning, and are specific to the zero-shot setting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.733, + 0.345, + 0.75 + ], + "angle": 0, + "content": "3. Inter-Intra Modal Measure" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.484, + 0.82 + ], + "angle": 0, + "content": "In this section, we provide a principled derivation of the IIMM metric from contrastive learning principles. In addition, we prove a theoretical bound and provide intuition behind the metric for predicting fine-tuning outcomes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Notation Let \\( t \\) denote a target task, and let \\( \\mathcal{T} \\) be the set of tasks under evaluation. Consider a training dataset for task \\( t \\) composed of image-text pairs. Let \\( X \\) denote the set of image embeddings, where each \\( x \\in X \\subset \\mathbb{R}^d \\). Similarly, let" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "\\(Y\\) denote the set of text embeddings, where each \\(y \\in Y \\subset \\mathbb{R}^d\\). For every image \\(x\\), let \\(y(x) \\in Y\\) be the text embedding corresponding to its correct label." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.182 + ], + "angle": 0, + "content": "Let \\(\\mu\\) and \\(\\nu\\) be two probability measures on a metric space \\((\\mathcal{X},d)\\). The 1-Wasserstein distance between \\(\\mu\\) and \\(\\nu\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.186, + 0.86, + 0.218 + ], + "angle": 0, + "content": "\\[\nW _ {1} (\\mu , \\nu) = \\inf _ {\\gamma \\in \\Gamma (\\mu , \\nu)} \\int_ {\\mathcal {X} \\times \\mathcal {X}} d (x, y) d \\gamma (x, y),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.224, + 0.905, + 0.255 + ], + "angle": 0, + "content": "where \\(\\Gamma (\\mu ,\\nu)\\) denotes the set of all couplings (i.e., joint distributions) with marginals \\(\\mu\\) and \\(\\nu\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.262, + 0.905, + 0.278 + ], + "angle": 0, + "content": "3.1. Derivation via Contrastive Learning Principles" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.284, + 0.905, + 0.342 + ], + "angle": 0, + "content": "We begin with the widely used InfoNCE loss in contrastive learning [37]. For an image embedding \\( x \\in X \\) and its corresponding (correct) text embedding \\( y \\in Y \\), the loss is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.347, + 0.944, + 0.412 + ], + "angle": 0, + "content": "\\[\nL (x, y, \\{y ^ {\\prime} \\}) = - \\log \\frac {\\exp \\left(\\frac {x ^ {\\top} y}{\\tau}\\right)}{\\exp \\left(\\frac {x ^ {\\top} y}{\\tau}\\right) + \\sum_ {y ^ {\\prime} \\in Y \\backslash \\{y (x) \\}} \\exp \\left(\\frac {x ^ {\\top} y ^ {\\prime}}{\\tau}\\right)},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.412, + 0.905, + 0.457 + ], + "angle": 0, + "content": "where \\(\\tau > 0\\) is the temperature parameter and the summation is over all incorrect text embeddings \\(y'\\) (i.e. \\(y' \\neq y(x)\\)). Minimizing the InfoNCE loss encourages two main effects:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.457, + 0.905, + 0.487 + ], + "angle": 0, + "content": "1. Positive Pair Alignment: The numerator is maximized when the cosine similarity \\( x^{\\top}y \\) is high." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.487, + 0.905, + 0.533 + ], + "angle": 0, + "content": "2. Negative Pair Separation: The denominator penalizes high similarities between \\( x \\) and any incorrect text embedding \\( y' \\), thus pushing down \\( x^\\top y' \\) for \\( y' \\neq y(x) \\)." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.457, + 0.905, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.533, + 0.905, + 0.578 + ], + "angle": 0, + "content": "The resulting changes in embedding geometry can be decomposed into two components: inter-modal misalignment and intra-modal uniformity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.595, + 0.905, + 0.641 + ], + "angle": 0, + "content": "Inter-modal misalignment The effect on incorrect pairs can be quantified by the average similarity between an image embedding \\( x \\) and the set of incorrect text embeddings:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.65, + 0.905, + 0.689 + ], + "angle": 0, + "content": "\\[\nS _ {\\text {i n t e r}} = \\frac {1}{| X |} \\sum_ {x \\in \\mathcal {X}} \\frac {1}{| Y | - 1} \\sum_ {y ^ {\\prime} \\in Y \\backslash \\{y (x) \\}} x ^ {\\top} y ^ {\\prime}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.692, + 0.905, + 0.768 + ], + "angle": 0, + "content": "A high \\( S_{\\mathrm{inter}} \\) indicates that the model has not sufficiently separated the image embedding from non-matching text embeddings. While indicative of potential for improvement, this can also be problematic, as fine-tuning may lead to over-alignment and subsequent catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.785, + 0.907, + 0.861 + ], + "angle": 0, + "content": "Intra-modal uniformity In addition to inter-modal alignment, contrastive learning also encourages a good spread, or uniformity, of embeddings within each modality [50]. For image embeddings, we quantify uniformity by computing the average pairwise similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.865, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\nS _ {\\text {i n t r a}} = \\frac {2}{| X | (| X | - 1)} \\sum_ {1 \\leq i < j \\leq | X |} x _ {i} ^ {\\top} x _ {j}. \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2398" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "A high \\( S_{\\mathrm{intra}} \\) implies that the image embeddings are overly clustered, suggesting that there is significant potential for further separation (or re-organization) during fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.485, + 0.202 + ], + "angle": 0, + "content": "Formulation and intuition Based on the above principles, we define the Inter-Intra Modal Measure (IIMM) as the average of the inter-modal and intra-modal terms:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.211, + 0.484, + 0.242 + ], + "angle": 0, + "content": "\\[\n\\mathrm {I I M M} = \\frac {1}{2} \\left(S _ {\\text {i n t e r}} + S _ {\\text {i n t r a}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.25, + 0.485, + 0.416 + ], + "angle": 0, + "content": "A high IIMM value indicates that the average similarity between an image and incorrect text embeddings is high and/or the image embeddings are highly clustered. This situation suggests that while there is significant capacity for learning (i.e., re-aligning the modalities), the risk of catastrophic forgetting is also high. Conversely, a low IIMM value implies that the image and text modalities are well separated (low inter-modal similarity for negatives) and the image embeddings are sufficiently dispersed. In this case, fine-tuning is expected to yield minimal performance changes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.446 + ], + "angle": 0, + "content": "The IIMM metric arises naturally from the objectives of contrastive learning:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.447, + 0.483, + 0.478 + ], + "angle": 0, + "content": "1. The InfoNCE loss drives positive pairs to be closely aligned and negative pairs to be well separated." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.478, + 0.483, + 0.507 + ], + "angle": 0, + "content": "2. The average similarity over incorrect pairs, \\( S_{\\mathrm{inter}} \\), captures the degree of inter-modal misalignment." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.508, + 0.483, + 0.537 + ], + "angle": 0, + "content": "3. The average similarity among image embeddings, \\( S_{\\mathrm{intra}} \\), measures intra-modal uniformity." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.447, + 0.483, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.598 + ], + "angle": 0, + "content": "By combining these two measures as in Equation (4), the IIMM provides a principled, predictive metric that reflects both the learning potential and the risk of catastrophic forgetting in vision-language models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.705 + ], + "angle": 0, + "content": "We provide further exploration of this measure and its relation to learning, forgetting, and the uniformity of the embedding space in Sec. 4.2. Additionally, ?? provides an analysis of various convex combinations of the intra- and inter-modal sub-scores, highlighting modification of component weights as a means of evaluating each component's contributions to the predicted outcomes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.273, + 0.728 + ], + "angle": 0, + "content": "3.2. Theoretical Bound" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this section, we provide a theoretical result that links the change in IIMM to the Wasserstein distance between the joint distributions of incorrect image-text pairs (and image-image pairs) before and after fine-tuning. This bound offers a rigorous guarantee: if fine-tuning only induces a small distributional shift as measured by the Wasserstein distance, then the IIMM metric will change by no more than a proportional amount. Such a result is important as it justifies the use of IIMM as a robust indicator—ensuring that minor, perhaps noisy, changes in the embeddings do not lead to erratic changes in the IIMM, and that significant changes in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "the IIMM are indeed indicative of a meaningful shift in the model's representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.131, + 0.907, + 0.178 + ], + "angle": 0, + "content": "Theorem 1. Assume that all image and text embeddings lie on the unit sphere \\( S^{d - 1} \\subset \\mathbb{R}^d \\). Let the IIMM metric be defined by" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.187, + 0.824, + 0.217 + ], + "angle": 0, + "content": "\\[\n\\operatorname {I I M M} (P) = \\frac {1}{2} \\Big (A (P) + B (P) \\Big),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.226, + 0.56, + 0.239 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.238, + 0.81, + 0.258 + ], + "angle": 0, + "content": "\\[\nA (P) = \\mathbb {E} _ {(x, y ^ {\\prime}) \\sim P _ {X Y}} \\left[ x ^ {\\top} y ^ {\\prime} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.907, + 0.294 + ], + "angle": 0, + "content": "with \\( P_{XY} \\) being the joint distribution of incorrect image-text pairs, and" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.303, + 0.822, + 0.323 + ], + "angle": 0, + "content": "\\[\nB (P) = \\mathbb {E} _ {(x, x ^ {\\prime}) \\sim P _ {X} \\otimes P _ {X}} \\left[ x ^ {\\top} x ^ {\\prime} \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.333, + 0.906, + 0.393 + ], + "angle": 0, + "content": "with \\(P_{X}\\) being the marginal distribution of image embeddings. Consider perturbed versions of these distributions \\(Q_{XY}\\) and \\(Q_{X}\\), respectively. Assume that the 1-Wasserstein distances satisfy" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.404, + 0.884, + 0.422 + ], + "angle": 0, + "content": "\\[\nW _ {1} \\left(P _ {X Y}, Q _ {X Y}\\right) \\leq \\delta_ {A} \\quad a n d \\quad W _ {1} \\left(P _ {X}, Q _ {X}\\right) \\leq \\delta_ {B}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.432, + 0.732, + 0.448 + ], + "angle": 0, + "content": "Then the following bound holds:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.458, + 0.838, + 0.488 + ], + "angle": 0, + "content": "\\[\n\\left| \\operatorname {I I M M} (Q) - \\operatorname {I I M M} (P) \\right| \\leq \\frac {\\delta_ {A}}{2} + \\delta_ {B}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.496, + 0.906, + 0.544 + ], + "angle": 0, + "content": "Furthermore, if one computes empirical estimates \\(\\widehat{\\mathrm{IMM}}(P)\\) and \\(\\widehat{\\mathrm{IMM}}(Q)\\) using \\(N\\) independent samples, then with probability at least \\(1 - \\eta\\)," + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.553, + 0.858, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\left| \\widehat {\\operatorname {I M M}} (Q) - \\widehat {\\operatorname {I M M}} (P) \\right| \\leq \\frac {\\delta_ {A}}{2} + \\delta_ {B} + \\epsilon_ {N},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.593, + 0.56, + 0.605 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.604, + 0.786, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\epsilon_ {N} = O \\Big (\\sqrt {\\frac {\\log (1 / \\eta)}{N}} \\Big)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.642, + 0.729, + 0.657 + ], + "angle": 0, + "content": "accounts for the sampling error." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.907, + 0.803 + ], + "angle": 0, + "content": "Our empirical results have indicated that IIMM is strongly correlated with performance gains and catastrophic forgetting. The bound provided here offers a theoretical explanation: only significant reorganization of the embedding space (reflected by a large Wasserstein distance) can cause a substantial shift in IIMM. Therefore, a high IIMM change can be interpreted as evidence of a considerable shift in the model's internal representation, likely correlating with notable performance changes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.647, + 0.833 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Datasets To evaluate the relationships between transferability measures and model performance changes following fine-tuning, we compiled a diverse collection of nine commonly used, publicly available computer vision benchmarks" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2399" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "provided by PyTorch [26, 39, 64] as our testing dataset, with the goal of minimizing domain overlap and covering a range of coarse and granular classification tasks. These datasets included Stanford Cars [21], CIFAR100 [22], the Describable Textures Dataset (DTD) [6], satellite imagery datasets EuroSAT [14] and RESISC45 [5], the German Traffic Sign Recognition Benchmark (GTSRB) [45], MNIST [23], the Scene Understanding Benchmark (SUN397) [52], and the Street View House Numbers (SVHN) dataset [33]. For each dataset, we generated validation splits by randomly sampling \\(10\\%\\) of the respective training data, capping the validation split size at 5000 samples for larger datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.482, + 0.428 + ], + "angle": 0, + "content": "Models and fine-tuning methods To investigate the applicability of IIMM to vision-language encoders, we conducted all experiments using two distinct groups of evaluations schema: one in which the base pre-trained model varies while the fine-tuning strategy is held constant, and another group in which the base pre-trained model is held constant while the fine-tuning strategy varies. We will refer to these evaluation schema as the base model evaluation and PEFT evaluation, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.482, + 0.517 + ], + "angle": 0, + "content": "For the base model evaluation, we perform full finetuning over all parameters for the following four pre-trained models: CLIP, EVA-02-CLIP [12], SigLIP [60], and CoCa [59], using a ViT-B/32 image backbone for CLIP and CoCa and a ViT-B/16 image backbone for SigLIP and EVA-02-CLIP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.482, + 0.64 + ], + "angle": 0, + "content": "For the PEFT evaluation, we utilized a CLIP model with a ViT-B/32 image backbone as the base model and evaluate four common parameter-efficient fine-tuning (PEFT) methods: CLIP-Adapter [13], BitFit (bias tuning) [3], LoRA [16], and a method we call Attention-Weight Tuning, which simply refines the fine-tuning of all attention layer weights in Touvron et al. [47] to just the attention layers' inprojection weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.482, + 0.825 + ], + "angle": 0, + "content": "Baselines While no methods currently exist for explicitly predicting gains in accuracy from fine-tuning for dual-encoder vision-language models, many related methods exist for stand-alone vision encoders. In particular, we compare against several recent or established measures self-described as transferability estimators. The six comparison measure are: TransRate [17], Self-challenging Fisher Discriminant Analysis (SFDA) [42], Neural Collapse informed Transferability Index (NCTI) [51], Efficient Multitask Model Selector EMMS [43], Gaussian Bhattacharyya Coefficient (GBC) [38], and H-score [2]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.482, + 0.901 + ], + "angle": 0, + "content": "We chose these comparison measures to span the definitions and applications of self-described transferability measures. Those that estimate the log-likelihood are H-Score, SFDA, TransRate, and EMMS. NCTI functions as an estimator of the gap between the embedding geometry of the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "pre-trained model and the embedding geometry of a hypothetical fine-tuned model [51]. The GBC quantifies transferability as a measure of overlap in the feature space between embedding clusters of different class labels [38]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.905, + 0.32 + ], + "angle": 0, + "content": "Evaluation measures We evaluate performance of the transferability measures through the strength of their relationship to three variables: accuracy on the target test data after fine-tuning on the target task training data; gain in performance from the pre-trained model to the fine-tuned model scaled by the pre-trained model error, which we refer to as gain over zero-shot error; and average accuracy across the remaining eight non-target datasets to measure loss in performance from the pre-trained model on nontarget tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.326, + 0.905, + 0.386 + ], + "angle": 0, + "content": "Definition 1 (Gain over zero-shot error). For a given task, we define \\( A_{zs} \\) as the zero-shot accuracy and \\( A_{ft} \\) as the fine-tuned accuracy. We then define the gain over zero-shot error as," + }, + { + "type": "equation", + "bbox": [ + 0.67, + 0.404, + 0.905, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\frac {A _ {f t} - A _ {z s}}{1 - A _ {z s}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.904, + 0.469 + ], + "angle": 0, + "content": "This value normalizes for differences in the zero-shot accuracy across models and tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.905, + 0.597 + ], + "angle": 0, + "content": "We first evaluate all measures' monotonic relationship to both fine-tuned accuracy and gain over zero-shot error using Kendall's \\(\\tau\\), which is a standard evaluation measure within the literature. We then look at the top performing measures of this rank correlation and fit a least-squares model to evaluate their predictive strength, additionally looking at a least-squares model fit of the top measure against the average accuracy on off-target tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Training details To establish the ground truth for changes in model in-domain (ID) and out-of-domain (OOD) accuracy following fine-tuning on each dataset, we started by separately fine-tuning each of the base models on the nine datasets in our dataset suite. Following training on a given dataset, each model was tested on all nine datasets, with in-domain accuracy calculated as its accuracy on the test split of the training dataset and out-of-domain accuracies as the accuracies obtained on the remaining datasets test split. Each fine-tuning run consisted of 30 epochs of gradient descent using an SGD optimizer and a batch size of 128 and incorporated an exhaustive grid search over all relevant model hyper-parameters, examining values \\(\\{1\\mathrm{e} - 2,1\\mathrm{e} - 3,1\\mathrm{e} - 4,1\\mathrm{e} - 5\\}\\) for the learning rate, \\(\\{1\\mathrm{e} - 3,1\\mathrm{e} - 4,1\\mathrm{e} - 5\\}\\) for the weight decay, \\(\\{2,4,8,16\\}\\) for LoRA model rank, and \\(\\{4,8,16,32\\}\\) for CLIP-Adapter reduction. The initial image and text embeddings were then generated per dataset by a forward pass of each dataset's train split through the pre-trained base models and normalizing to unit vectors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2400" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.23, + 0.089, + 0.768, + 0.235 + ], + "angle": 0, + "content": "
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueτp-valueτp-valueτp-valueτ
IIMM< 10-30.890.0010.830.0020.780.0020.78
GBC0.0020.780.0060.720.0020.780.0010.83
TransRate0.0130.670.0250.610.0130.670.0060.72
EMMS0.0250.610.3580.280.0250.610.1190.44
NCTI0.919-0.060.028-0.590.8180.070.827-0.06
SFDA0.2600.330.919-0.060.1800.390.3580.28
H-Score0.7610.110.002-0.780.0750.500.013-0.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.208, + 0.247, + 0.787, + 0.262 + ], + "angle": 0, + "content": "Table 1. Kendall's \\( \\tau \\) correlation of score and gain over zero-shot error. Larger \\( \\tau \\) values are better.*" + }, + { + "type": "table", + "bbox": [ + 0.233, + 0.275, + 0.767, + 0.423 + ], + "angle": 0, + "content": "
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueτp-valueτp-valueτp-valueτ
IIMM0.0060.720.0060.720.0130.670.0250.61
GBC0.0250.610.0060.720.0130.670.0130.67
TransRate0.0750.500.0250.610.0450.560.0450.56
EMMS0.0130.670.0250.610.0750.500.0250.61
NCTI1.0000.000.173-0.370.818-0.070.8270.06
SFDA0.0750.500.6120.170.0750.500.2600.33
H-Score0.6120.170.045-0.560.0250.610.075-0.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.254, + 0.434, + 0.741, + 0.448 + ], + "angle": 0, + "content": "Table 2. Kendall's \\( \\tau \\) correlation of score and accuracy. Larger \\( \\tau \\) values are better.*" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.474, + 0.184, + 0.489 + ], + "angle": 0, + "content": "4.1. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.483, + 0.563 + ], + "angle": 0, + "content": "We first look at Kendall's \\(\\tau\\) correlation to determine the strength of the monotonic relationship of our proposed and comparison measures to both the fine-tuned gain over zero-shot error and fine-tuned accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The IIMM and the GBC have significantly stronger monotonic relationships than the other measures for both gain over zero-shot error, as shown in Tab. 1, and accuracy, as shown in Tab. 2. In both tables, the IIMM and the GBC are the most strongly correlated measures and display the greatest significance. Upon further investigation, the IIMM displays a mostly linear relationship with gain over zero-shot error, while the GBC displays a mostly exponential relationship with gain over zero-shot error. This allows us to fit and evaluate a predictive model using least-squares regression of the IIMM with gain over zero-shot error, and the GBC with the log of gain over zero-shot error. The results of these evaluations can be seen in Tab. 3, which shows the significance of the slope coefficient and the \\( R^2 \\) value, an estimate of the amount of variation within the data explained by its linear relationship to the measure. The IIMM has a stronger fit across all base pre-trained models compared to the GBC, and indeed shows itself to have a strong linear relationship to gain over zero-shot error with \\( R^2 > 0.89 \\) for all base models. Without the ability to capture cross-modal alignment, the compared uni-modal metrics expectedly demonstrate worse performance when com" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.475, + 0.905, + 0.504 + ], + "angle": 0, + "content": "pared to the IIMM, underscoring the value of an integrated vision-language metric." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.506, + 0.907, + 0.836 + ], + "angle": 0, + "content": "This positive linear relationship between the IIMM and in-domain gain over zero-shot error is clearly seen in the top plots of Fig. 2. The bottom plots of Fig. 2 show the negative linear relationship of the IIMM with the average change in accuracy across a suite of eight out-of-domain tasks. For all base pre-trained models aside from CLIP, the IIMM also shows a strong linear relationship to forgetting, as evidenced by the significance of the slope coefficient and high \\( R^2 \\) values. These relationships hold true across parameter-efficient fine-tuning methods, as shown in Fig. 3. We see with CLIP-Adapter that its overall worse in-domain performance attenuates the strength of the relationship between IIMM and gain over zero-shot error, and predictably, that a lower in-domain performance gain resulted in significantly less catastrophic forgetting. CLIP-Adapter's design — injecting task kernels atop a frozen backbone — results in smaller, localized representational shifts, resulting in weaker in-distribution performance gains while preventing greater out-of-distribution performance loss. Since the IIMM is correlated with near-optimal performance gains, fine-tuning methods that result in sub-optimal gains may yield lower IIMM correlations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.837, + 0.905, + 0.868 + ], + "angle": 0, + "content": "These results provide evidence that the IIMM is strongly correlated with fine-tuning gains over zero-shot error. Un" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.876, + 0.905, + 0.901 + ], + "angle": 0, + "content": "*P-values reflect the likelihood of observing results by random chance; smaller values suggest stronger associations." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2401" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.236, + 0.089, + 0.763, + 0.168 + ], + "angle": 0, + "content": "
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueR²p-valueR²p-valueR²p-valueR²
IIMM< 10-30.92< 10-30.89< 10-30.92< 10-30.92
GBC0.0040.720.0060.680.0150.590.0050.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.178, + 0.907, + 0.207 + ], + "angle": 0, + "content": "Table 3. Evaluation of linear fit of the IIMM to gain over zero-shot error and the GBC to the log of gain over zero-shot error. The significance of the slope coefficient is given by the p-value, and the strength of the fit is evaluated with the \\( R^2 \\) value.*" + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.221, + 0.874, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.532, + 0.908, + 0.577 + ], + "angle": 0, + "content": "Figure 2. Performance gains and losses versus the IIMM per pre-trained model, with linear regression fit and \\(96\\%\\) confidence intervals. The top plots show the relationship between the on-target (ID) task IIMM to the on-target task performance gains over zero-shot error, while the bottom plots show the relationship between on-target task IIMM and the average performance loss on off-target (OOD) tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.483, + 0.645 + ], + "angle": 0, + "content": "like GBC, the second-highest correlated measure, the IIMM is bounded in \\([-1,1]\\) and extreme values are informative in isolation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.663, + 0.36, + 0.68 + ], + "angle": 0, + "content": "4.2. Exploring embedding features" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.687, + 0.483, + 0.867 + ], + "angle": 0, + "content": "To empirically validate the provided theoretical justification for leveraging multi-modal information to improve transferability estimation in vision-language models, we independently evaluated all relevant components of CLIP's embedding geometry, known to manifest as a hypersphere. Using (3), we separately computed the average pairwise similarities for the image and text modalities and fit each to the accuracy gain over zero-shot error. We then applied (2) to compute the degree of incorrect label alignment across the hypersphere and also fit the measure to the accuracy gain over zero-shot error. The final results of these evaluations can be seen in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "It is observed that the features of the embedding structure most predictive of accuracy gains are the intra-images dis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.907, + 0.783 + ], + "angle": 0, + "content": "tance and the incorrect label alignment. Based on these results, the correct label alignment and intra-texts distance are generally uninformative of accuracy gains with fine-tuning. As described in Sec. 3, each of these terms corresponds to a feature of the shared embedding space related to learning and forgetting. We include additional results of the Pearson correlations between different convex combinations of our inter- and intra- terms and gain over zero-shot error in the supplementary material. This empirical evidence successfully verifies that the terms combined have a stronger relationship with learning and forgetting than they do independently." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.799, + 0.634, + 0.815 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose the Inter-Intra Modal Measure (IIMM) as a novel metric for predicting fine-tuning performance in vision-language models. We demonstrate that the IIMM strongly correlates with both learning gains and catastrophic forgetting, offering a simple yet powerful met" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2402" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.089, + 0.874, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.399, + 0.908, + 0.444 + ], + "angle": 0, + "content": "Figure 3. Performance gains and losses versus the IIMM per PEFT method, with linear regression fit and \\(96\\%\\) confidence intervals. The top plots show the relationship between the on-target task (ID) IIMM and on-target task performance gains over zero-shot error, while the bottom plots show the relationship between on-target task IIMM and the average performance loss on off-target (OOD) tasks." + }, + { + "type": "table", + "bbox": [ + 0.194, + 0.455, + 0.806, + 0.548 + ], + "angle": 0, + "content": "
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueR²p-valueR²p-valueR²p-valueR²
Incorrect Alignment0.0010.820.0040.720.0120.620.0010.80
Intra-Images Distance< 10-30.88< 10-30.860.0010.81< 10-30.88
Intra-Texts Distance0.0840.370.0340.500.0190.570.2650.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.559, + 0.908, + 0.588 + ], + "angle": 0, + "content": "Table 4. Linear fit of measure to gain over zero-shot error with full fine-tuning by pre-trained model. The significance of the slope coefficient is given by the p-value, and the \\( R^2 \\) values represent the proportion of variation within the data modeled by the linear relationship." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.613, + 0.486, + 0.901 + ], + "angle": 0, + "content": "ric for estimating fine-tuning outcomes. Because the IIMM is bounded, an extreme value in isolation can inform practitioners whether a model will have limited or significant gains from fine-tuning. In addition, our results show that the IIMM can be leveraged to fit predictive linear models for estimating fine-tuning accuracy gains across diverse tasks with minimal computational overhead. We demonstrate that our proposed measure has a stronger relationship to both fine-tuned accuracy and accuracy changes with fine-tuning than baselines. We further show the IIMM is informative of performance changes across parameter efficient fine-tuning methods, highlighting some methods are particularly susceptible to catastrophic forgetting when the zero-shot IIMM is high. Beyond empirical validation, we provide a theoretical bound linking changes in IIMM to the Wasserstein distance between pre- and post-fine-tuning embedding distributions, ensuring stability in IIMM predictions with respect to changes in representation space. Given the IIMM measures changes in embedding geometry prior to any task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.909, + 0.72 + ], + "angle": 0, + "content": "specific heads, the metric is also task-agnostic, enabling its application to a variety of downstream tasks. We encourage future research to validate such use in broader applications. Overall, our work provides a practical and theoretically grounded tool for optimizing fine-tuning decisions in vision-language models, paving the way for more efficient and informed adaptation of foundation models to new tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Limitations It is possible that the IIMM could show more variation when compared across a more diverse set of base models. While extreme values of the IIMM are heuristically informative in isolation, fitting a model for transfer accuracy estimation requires first fine-tuning across a small but diverse set of benchmarks. This could be an avenue of future research, as there might be additional information within the embedding geometry that allows for fine-tuning performance prediction without requiring information from fine-tuning on other tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2403" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.176 + ], + "angle": 0, + "content": "The authors acknowledge the MIT Lincoln Laboratory Supercomputing Center for providing the high-performance computing resources that have contributed to the research results reported in this paper." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.176, + 0.483, + 0.312 + ], + "angle": 0, + "content": "DISTRIBUTION STATEMENT A. Approved for public release. Distribution is unlimited. This material is based upon work supported by the Under Secretary of Defense for Research and Engineering under Air Force Contract No. FA8702-15-D-0001 or FA8702-25-D-B002. Any opinions, findings, conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Under Secretary of Defense for Research and Engineering." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.326, + 0.187, + 0.341 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.351, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[1] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational Models Defining a New Era in Vision: A Survey and Outlook, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.421, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[2] Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An Information-Theoretic Approach to Transferability in Task Transfer Learning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2309-2313, Taipei, Taiwan, 2019. IEEE. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.506, + 0.484, + 0.589 + ], + "angle": 0, + "content": "[3] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1-9, Dublin, Ireland, 2022. Association for Computational Linguistics. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[4] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.661, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[5] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing Textures in the Wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, Columbus, OH, USA, 2014. IEEE. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.774, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[7] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large Scale Fine-Grained Categorization and Domain-Specific Transfer Learning, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[8] Yuxuan Ding, Lingqiao Liu, Chunna Tian, Jingyuan Yang, and Haoxuan Ding. Don't Stop Learning: Towards Continual Learning for the CLIP Model, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[9] Yuhe Ding, Bo Jiang, Aijing Yu, Aihua Zheng, and Jian Liang. Which model to transfer? a survey on transferability estimation. arXiv preprint arXiv:2402.15231, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.351, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[10] Kshitij Dwivedi and Gemma Roig. Representation Similarity Analysis for Efficient Task Taxonomy & Transfer Learning. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12379-12388, Long Beach, CA, USA, 2019. IEEE. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[11] Zhihao Fan, Zhongyu Wei, Jingjing Chen, Siyuan Wang, Ze-jun Li, Jiarong Xu, and Xuanjing Huang. A unified continuous learning framework for multi-modal knowledge discovery and pre-training, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[12] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[13] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. CLIP-Adapter: Better Vision-Language Models with Feature Adapters, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Introducing Eurosat: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. In IGARSS 2018 - 2018 IEEE International Geoscience and Remote Sensing Symposium, pages 204-207, Valencia, 2018. IEEE. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[15] Dapeng Hu, Shipeng Yan, Qizhengqiu Lu, Lanqing Hong, Hailin Hu, Yifan Zhang, Zhenguo Li, Xinchao Wang, and Jiashi Feng. How well does self-supervised pre-training perform with streaming data?, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[16] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In ICLR. arXiv, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[17] Long-Kai Huang, Junzhou Huang, Yu Rong, Qiang Yang, and Ying Wei. Frustratingly easy transferability estimation. In International Conference on Machine Learning, pages 9201-9225. PMLR, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[18] Joana Palés Huix, Adithya Raju Ganeshan, Johan Fredin Haslum, Magnus Söderberg, Christos Matsoukas, and Kevin Smith. Are Natural Domain Foundation Models Useful for Medical Image Classification? In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 7619-7628. IEEE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[20] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D Object Representations for Fine-Grained Categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, Sydney, Australia, 2013. IEEE. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2404" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[22] Alex Krizhevsky. Learning Multiple Layers of Features from Tiny Images, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[23] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, Nov./1998. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[24] Hao Li, Charless Fowlkes, Hao Yang, Onkar Dabeer, Zhuowen Tu, and Stefano Soatto. Guided Recommendation for Model Fine-Tuning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3633-3642, Vancouver, BC, Canada, 2023. IEEE. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[25] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before Fuse: Vision and Language Representation Learning with Momentum Distillation. In Advances in Neural Information Processing Systems, pages 9694-9705. Curran Associates, Inc., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[26] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning. In First Workshop of Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[27] C. Liao, T. Tsiligkaridis, and B. Kulis. Descriptor and word soups: Overcoming the parameter efficiency accuracy trade-off for out-of-distribution few-shot learning. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[28] C. Liao, C. So, T. Tsiligkaridis, and B. Kulis. Multimodal unsupervised domain generalization by retrieving across the modality gap. In ICLR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.551 + ], + "angle": 0, + "content": "[29] Yong Lin, Lu Tan, Hangyu Lin, Zeming Zheng, Renjie Pi, Jipeng Zhang, Shizhe Diao, Haoxiang Wang, Han Zhao, Yuan Yao, and Tong Zhang. Speciality vs Generality: An Empirical Study on Catastrophic Forgetting in Fine-tuning Foundation Models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[30] Michael McCloskey and Neal J. Cohen. Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem. In *Psychology of Learning and Motivation*, pages 109–165. Elsevier, 1989. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[31] S. Menon and C. Vondrick. Visual classification via description from large language models. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[32] Jishnu Mukhoti, Yarin Gal, Philip HS Torr, and Puneet K Dokania. Fine-tuning can cripple your foundation model; preserving features may be the solution. arXiv preprint arXiv:2308.13320, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[33] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading Digits in Natural Images with Unsupervised Feature Learning. NeurIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[34] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. LEEP: A New Measure to Evaluate Transferability of Learned Representations. In Proceedings of the 37th International Conference on Machine Learning, pages 7294-7305. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[35] Zixuan Ni, Siliang Tang, and Yueting Zhuang. Self-supervised class incremental learning, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Zixuan Ni, Longhui Wei, Siliang Tang, Yueting Zhuang, and Qi Tian. Continual Vision-Language Representation Learning with Off-Diagonal Information, 2023. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[38] Michal Páncy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Transferability Estimation using Bhattacharyya Class Separability, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[41] K. Roth, J. M. Kim, A. S. Koepke, O. Vinyals, C. Schmid, and Z. Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[42] Wenqi Shao, Xun Zhao, Yixiao Ge, Zhaoyang Zhang, Lei Yang, Xiaogang Wang, Ying Shan, and Ping Luo. Not all models are equal: Predicting model transferability in a self-challenging fisher space. In European Conference on Computer Vision, pages 286-302. Springer, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[43] Wenqi Shao, Chonghe Jiang, Kaipeng Zhang, Yu Qiao, Ping Luo, et al. Foundation model is efficient multimodal multitask model selector. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[44] Tejas Srinivasan, Ting-Yun Chang, Leticia Pinto Alva, Georgios Chochlakis, Mohammad Rostami, and Jesse Thomason. CLiMB: A continual learning benchmark for vision-and-language tasks. In Advances in Neural Information Processing Systems, pages 29440-29453. Curran Associates, Inc., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[45] J. Stallkamp, M. Schlipsing, J. Salmen, and C. Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. *Neural Networks*, 32:323–332, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[46] P. Teterwak, K. Saito, T. Tsiligkaridis, B. A. Plummer, and K. Saenko. Is large-scale pretraining the secret to good domain generalization? In ICLR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[47] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about Vision Transformers, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[48] Anh Tran, Cuong Nguyen, and Tal Hassner. Transferability and Hardness of Supervised Classification Tasks. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1395-1405, Seoul, Korea (South), 2019. IEEE. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Kevin Vogt-Lowell, Noah Lee, Theodoros Tsiligkaridis, and Marc Vaillant. Robust fine-tuning of vision-language models for domain generalization. In 2023 IEEE High Performance" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2405" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Extreme Computing Conference (HPEC), pages 1-7. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[50] Tongzhou Wang and Phillip Isola. Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere. In Proceedings of the 37th International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[51] Zijian Wang, Yadan Luo, Liang Zheng, Zi Huang, and Mahsa Baktashmotlagh. How far pre-trained models are from neural collapse on the target dataset informs their transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5549-5558, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[52] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, San Francisco, CA, USA, 2010. IEEE. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[53] Yiqun Xie, Zhihao Wang, Weiye Chen, Zhili Li, Xiaowei Jia, Yanhua Li, Ruichen Wang, Kangyang Chai, Ruohan Li, and Sergii Skakun. When are foundation models effective? understanding the suitability for pixel-level classification using multispectral imagery. CoRR, abs/2404.11797, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[54] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive Pretraining for Zero-shot Video-Text Understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787–6800, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.483, + 0.586 + ], + "angle": 0, + "content": "[55] Chao Yi, De-Chuan Zhan, and Han-Jia Ye. Bridge the Modality and Capacity Gaps in Vision-Language Model Selection, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[56] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In International Conference on Machine Learning, pages 12133-12143. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[57] E. Yu, C. Liao, S. Ravi, T. Tsiligkaridis, and B. Kulis. Image-capture encoding for improving zero-shot generalization. In WACV, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[58] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23219-23230. IEEE. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[59] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[60] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[61] Dan Zhang, Tao Feng, Lilong Xue, Yuandong Wang, Yuxiao Dong, and Jie Tang. Parameter-efficient fine-tuning for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "foundation models. arXiv preprint arXiv:2501.13787, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[62] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and Curtis P. Langlotz. Contrastive Learning of Medical Visual Representations from Paired Images and Text. In Proceedings of the 7th Machine Learning for Healthcare Conference, pages 2-25. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[63] Zangwei Zheng, Mingyuan Ma, Kai Wang, Ziheng Qin, Xi-angyu Yue, and Yang You. Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 19068-19079. IEEE. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[64] Orr Zohar, Kuan-Chieh Wang, Shih-Cheng Huang, and Serena Yeung. LOVM: Language-Only Vision Model Selection. In Conference on Neural Information Processing Systems (Neurips), 2023. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "2406" + } + ] +] \ No newline at end of file diff --git a/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/f4964a00-88d3-4539-871c-39c0ed82d2c6_origin.pdf b/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/f4964a00-88d3-4539-871c-39c0ed82d2c6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..edd1167440112aebbdda09b6e76ff62ec42f751e --- /dev/null +++ b/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/f4964a00-88d3-4539-871c-39c0ed82d2c6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7039b1b28ee9ce9e19b8fb1220378a7a0eb5d2dd12e5fa60b5ac3209ba4beb18 +size 1124731 diff --git a/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/full.md b/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..91baead2dbf9b10ba6e67cc20ee4877861c3bf85 --- /dev/null +++ b/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/full.md @@ -0,0 +1,333 @@ +# The Inter-Intra Modal Measure: A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models + +Laura Niss* Kevin Vogt-Lowell* Theodoros Tsiligkaridis + +{laura.niss, kevin.vogt-lowell, ttsili}@ll.mit.edu + +MIT Lincoln Laboratory + +244 Wood St, Lexington, MA 02421 + +# Abstract + +The fine-tuning of large vision-language foundation models remains an underexplored area, particularly regarding its impact on learning gains and catastrophic forgetting. Inspired by the significance of modality gaps in contrastive dual-encoders, we introduce the Inter-Intra Modal Measure (IIMM)—a predictive metric that quantifies the relationship between intra-modal image embedding similarity and inter-modal misalignment. Through extensive empirical analysis across four state-of-the-art vision-language models and five fine-tuning techniques, we establish a strong linear relationship: tasks with higher IIMM scores yield greater in-domain performance improvements but suffer from more pronounced out-of-domain degradation, with some parameter-efficient fine-tuning (PEFT) methods exhibiting severe forgetting. Compared to existing transferability measures, the IIMM demonstrates significantly stronger predictive power for accuracy changes post fine-tuning in dual-encoder models. Moreover, we provide a theoretical bound, proving that changes in IIMM are limited by the Wasserstein distance between pre- and post-fine-tuning embedding distributions, ensuring its stability and robustness as a predictive measure. With only a single forward pass of the target data, practitioners can leverage this key insight to evaluate the degree to which a model can be expected to improve following fine-tuning. When combined with prior knowledge of a model's performance across diverse tasks, the IIMM further enhances transferability predictions for novel tasks, offering a lightweight yet effective tool for guiding model adaptation strategies. Our code is provided at https://github.com/mit-ll/IIMM. + +# 1. Introduction + +Large vision-language foundation models have demonstrated superior zero-shot generalization compared to tra + +ditional vision models, leveraging web-scale pre-training to achieve state-of-the-art performance across diverse tasks without task-specific supervision [39]. Advances in enhancing zero-shot generalization have also been proposed [31, 41, 57]. Fine-tuning can further enhance performance on specific downstream applications [4, 27, 28, 46, 61]. However, fine-tuned foundation models still struggle on certain tasks, particularly in specialized fields such as satellite and medical image classification [1, 18, 49, 53]. Given the computational cost of fine-tuning, determining whether adaptation will lead to meaningful improvements remains a crucial challenge. + +Beyond resource efficiency, fine-tuning also introduces the risk of catastrophic forgetting—where a model loses previously learned knowledge while improving on the new task. This effect is particularly pronounced in dual-encoder vision-language models, where fine-tuning disrupts the alignment between image and text embedding spaces, degrading the model's ability to generalize to out-of-domain tasks [8, 36]. This creates an inherent trade-off between learning and forgetting, making it critical for practitioners to anticipate the effects of fine-tuning before allocating computational resources [32, 58, 63]. A predictive metric that captures both learning gains and forgetting risks would provide a valuable tool for optimizing fine-tuning decisions. + +The challenge of estimating a model's fine-tuned performance a priori falls under the domain of model transferability [38, 51]. Common transferability measures assess a pretrained model's suitability for a given task by analyzing its latent representations. However, most traditional methods rely on source data and task labels, making them incompatible with contrastively trained foundation models that lack explicit supervision [43]. Furthermore, newer transferability metrics are designed for uni-modal settings, failing to account for inter-modal interactions in vision-language models. Among approaches that estimate post-fine-tuning performance, evaluation is often limited to rank correlation rather than directly predicting performance changes [17, 42]. + +![](images/3f2363d156a0bd9ba19e94948c4788ac2a0d7bb192676307db92febb5ada5b9c.jpg) +Figure 1. The Inter-Intra Modal Measure (IIMM) combines two key aspects of embedding geometry: intra-modal image embedding similarity, and inter-modal alignment between image and text embeddings. This measure captures information about the embedding geometry that demonstrates a strong relationship to performance gains observed with fine-tuning. + +Present work To address these limitations, we introduce the Inter-Intra Modal Measure (IIMM), a novel metric that quantifies both intra-modal representation spread and inter-modal alignment in contrastive vision-language models, capturing the fundamental trade-off between learning and forgetting without the need for fine-tuning. We compare our measure against several standard and recent transferability measures in both rank correlation and strength of predictive fit to performance gains after fine-tuning. To evaluate our measure, we perform fine-tuning across nine well-known vision classification tasks using four state-of-the-art vision-language models and five fine-tuning techniques, analyzing in-domain improvements as well as induced catastrophic forgetting across eight out-of-domain tasks. We compare our measure against six transferability measures applicable to vision-language encoders, and show that our measure has a significantly stronger monotonic relationship with both fine-tuned in-domain accuracy and in-domain performance gains scaled by the zero-shot error. When fitting a simple linear regression of the IIMM to the scaled in-domain accuracy gains, we achieve $R^2 > 0.85$ for all base models. This offers a practical and theoretically grounded tool for model adaptation in vision-language tasks. + +Contributions In summary, our work provides the following contributions: + +1. We introduce the Inter-Intra Modal Measure (IIMM), a standalone metric that integrates inter-modal misalignment and intra-modal uniformity to predict fine-tuning outcomes in dual-encoder vision-language models. +2. We derive the IIMM directly from contrastive learning + +principles, establishing a clear connection between the geometry of the embedding space and transferability estimation. + +3. We prove a theoretical bound showing that changes in the IIMM are constrained by the 1-Wasserstein distance between pre- and post-fine-tuning embedding distributions. This result guarantees that significant shifts in the representation space will lead to substantial changes in IIMM, underscoring its stability as a predictive measure. +4. We show that the zero-shot IIMM has a strong, positive linear relationship with learning and a negative relationship with forgetting post fine-tuning. We also provide empirical evidence that for any vision-language dual-encoder, fine-tuning on only a few standard vision benchmarks is sufficient to fit a strong linear predictor of fine-tuning accuracy gains with the IIMM for transfer estimation of novel tasks. + +# 2. Related work + +Dual-encoders and the modality gap Multi-modal dual-encoders trained using contrastive learning consist of two models jointly trained to map each modality into a shared representation space [19, 25, 40, 54, 62]. Previous research has highlighted the existence of a modality gap within this joint embedding space that results from the encoders initializing embeddings within random cones, and that contrastive loss maintains this initialized gap [26]. Additionally, empirical evidence shows that manipulating this gap can impact zero-shot performance. While catastrophic forgetting has been proven to significantly affect traditional su + +pervised learning techniques [20, 30], recent research has shown that self-supervised, uni-modal models do not suffer from such severe forgetting [15, 35]. Yet, with multimodal foundation models, such as CLIP, continuous training can cause misalignment of the image and text embeddings, an outcome highlighted and explored in several works [8, 11, 29, 36, 44]. + +Transferability estimation A significant amount of research has focused on addressing transferability estimation with respect to uni-modal encoders. Estimation of model transferability often involves transferability metrics, or measures that predict how well knowledge learned from source tasks will transfer to new target tasks. [9] The approaches for quantifying these metrics vary greatly, with some approaches simply defining them as the log-likelihood of a target task. Methods that rely upon a log-likelihood definition are inherently well-suited for model selection because the definition relies directly upon the size of the target dataset, and within model selection the target dataset is held constant. Regardless of approach, nearly all metrics are evaluated via rank correlation across a model set [17, 42, 43, 51]. + +Some modern approaches include training small neural networks and comparing the similarity of resultant task embeddings [10], training a full neural network on large amounts of models trained across many tasks to learn a similarity score function [24], and comparing source training data to new target task data to estimate similarity [7]. Some classic methods, such as LEEP [34] and NCE [48], rely upon information from the source training data and are thus inapplicable to self-supervised foundation models; they also cannot be applied to dual-encoders. However, there are a range of uni-modal methods that can be easily applied to the vision encoder of a dual-encoder model [17, 43, 56]. + +For multi-modal models, some recent research has involved zero-shot model selection using language to infer the transferability of a dataset, such as LOVM [64] and SWAB [55]. These methods do not claim to estimate model performance after fine-tuning, and are specific to the zero-shot setting. + +# 3. Inter-Intra Modal Measure + +In this section, we provide a principled derivation of the IIMM metric from contrastive learning principles. In addition, we prove a theoretical bound and provide intuition behind the metric for predicting fine-tuning outcomes. + +Notation Let $t$ denote a target task, and let $\mathcal{T}$ be the set of tasks under evaluation. Consider a training dataset for task $t$ composed of image-text pairs. Let $X$ denote the set of image embeddings, where each $x \in X \subset \mathbb{R}^d$ . Similarly, let + +$Y$ denote the set of text embeddings, where each $y \in Y \subset \mathbb{R}^d$ . For every image $x$ , let $y(x) \in Y$ be the text embedding corresponding to its correct label. + +Let $\mu$ and $\nu$ be two probability measures on a metric space $(\mathcal{X},d)$ . The 1-Wasserstein distance between $\mu$ and $\nu$ is defined as: + +$$ +W _ {1} (\mu , \nu) = \inf _ {\gamma \in \Gamma (\mu , \nu)} \int_ {\mathcal {X} \times \mathcal {X}} d (x, y) d \gamma (x, y), +$$ + +where $\Gamma (\mu ,\nu)$ denotes the set of all couplings (i.e., joint distributions) with marginals $\mu$ and $\nu$ + +# 3.1. Derivation via Contrastive Learning Principles + +We begin with the widely used InfoNCE loss in contrastive learning [37]. For an image embedding $x \in X$ and its corresponding (correct) text embedding $y \in Y$ , the loss is defined as + +$$ +L (x, y, \{y ^ {\prime} \}) = - \log \frac {\exp \left(\frac {x ^ {\top} y}{\tau}\right)}{\exp \left(\frac {x ^ {\top} y}{\tau}\right) + \sum_ {y ^ {\prime} \in Y \backslash \{y (x) \}} \exp \left(\frac {x ^ {\top} y ^ {\prime}}{\tau}\right)}, +$$ + +where $\tau > 0$ is the temperature parameter and the summation is over all incorrect text embeddings $y'$ (i.e. $y' \neq y(x)$ ). Minimizing the InfoNCE loss encourages two main effects: + +1. Positive Pair Alignment: The numerator is maximized when the cosine similarity $x^{\top}y$ is high. +2. Negative Pair Separation: The denominator penalizes high similarities between $x$ and any incorrect text embedding $y'$ , thus pushing down $x^\top y'$ for $y' \neq y(x)$ . + +The resulting changes in embedding geometry can be decomposed into two components: inter-modal misalignment and intra-modal uniformity. + +Inter-modal misalignment The effect on incorrect pairs can be quantified by the average similarity between an image embedding $x$ and the set of incorrect text embeddings: + +$$ +S _ {\text {i n t e r}} = \frac {1}{| X |} \sum_ {x \in \mathcal {X}} \frac {1}{| Y | - 1} \sum_ {y ^ {\prime} \in Y \backslash \{y (x) \}} x ^ {\top} y ^ {\prime}. \tag {2} +$$ + +A high $S_{\mathrm{inter}}$ indicates that the model has not sufficiently separated the image embedding from non-matching text embeddings. While indicative of potential for improvement, this can also be problematic, as fine-tuning may lead to over-alignment and subsequent catastrophic forgetting. + +Intra-modal uniformity In addition to inter-modal alignment, contrastive learning also encourages a good spread, or uniformity, of embeddings within each modality [50]. For image embeddings, we quantify uniformity by computing the average pairwise similarity: + +$$ +S _ {\text {i n t r a}} = \frac {2}{| X | (| X | - 1)} \sum_ {1 \leq i < j \leq | X |} x _ {i} ^ {\top} x _ {j}. \tag {3} +$$ + +A high $S_{\mathrm{intra}}$ implies that the image embeddings are overly clustered, suggesting that there is significant potential for further separation (or re-organization) during fine-tuning. + +Formulation and intuition Based on the above principles, we define the Inter-Intra Modal Measure (IIMM) as the average of the inter-modal and intra-modal terms: + +$$ +\mathrm {I I M M} = \frac {1}{2} \left(S _ {\text {i n t e r}} + S _ {\text {i n t r a}}\right). \tag {4} +$$ + +A high IIMM value indicates that the average similarity between an image and incorrect text embeddings is high and/or the image embeddings are highly clustered. This situation suggests that while there is significant capacity for learning (i.e., re-aligning the modalities), the risk of catastrophic forgetting is also high. Conversely, a low IIMM value implies that the image and text modalities are well separated (low inter-modal similarity for negatives) and the image embeddings are sufficiently dispersed. In this case, fine-tuning is expected to yield minimal performance changes. + +The IIMM metric arises naturally from the objectives of contrastive learning: + +1. The InfoNCE loss drives positive pairs to be closely aligned and negative pairs to be well separated. +2. The average similarity over incorrect pairs, $S_{\mathrm{inter}}$ , captures the degree of inter-modal misalignment. +3. The average similarity among image embeddings, $S_{\mathrm{intra}}$ , measures intra-modal uniformity. + +By combining these two measures as in Equation (4), the IIMM provides a principled, predictive metric that reflects both the learning potential and the risk of catastrophic forgetting in vision-language models. + +We provide further exploration of this measure and its relation to learning, forgetting, and the uniformity of the embedding space in Sec. 4.2. Additionally, ?? provides an analysis of various convex combinations of the intra- and inter-modal sub-scores, highlighting modification of component weights as a means of evaluating each component's contributions to the predicted outcomes. + +# 3.2. Theoretical Bound + +In this section, we provide a theoretical result that links the change in IIMM to the Wasserstein distance between the joint distributions of incorrect image-text pairs (and image-image pairs) before and after fine-tuning. This bound offers a rigorous guarantee: if fine-tuning only induces a small distributional shift as measured by the Wasserstein distance, then the IIMM metric will change by no more than a proportional amount. Such a result is important as it justifies the use of IIMM as a robust indicator—ensuring that minor, perhaps noisy, changes in the embeddings do not lead to erratic changes in the IIMM, and that significant changes in + +the IIMM are indeed indicative of a meaningful shift in the model's representation. + +Theorem 1. Assume that all image and text embeddings lie on the unit sphere $S^{d - 1} \subset \mathbb{R}^d$ . Let the IIMM metric be defined by + +$$ +\operatorname {I I M M} (P) = \frac {1}{2} \Big (A (P) + B (P) \Big), +$$ + +where + +$$ +A (P) = \mathbb {E} _ {(x, y ^ {\prime}) \sim P _ {X Y}} \left[ x ^ {\top} y ^ {\prime} \right], +$$ + +with $P_{XY}$ being the joint distribution of incorrect image-text pairs, and + +$$ +B (P) = \mathbb {E} _ {(x, x ^ {\prime}) \sim P _ {X} \otimes P _ {X}} \left[ x ^ {\top} x ^ {\prime} \right], +$$ + +with $P_{X}$ being the marginal distribution of image embeddings. Consider perturbed versions of these distributions $Q_{XY}$ and $Q_{X}$ , respectively. Assume that the 1-Wasserstein distances satisfy + +$$ +W _ {1} \left(P _ {X Y}, Q _ {X Y}\right) \leq \delta_ {A} \quad a n d \quad W _ {1} \left(P _ {X}, Q _ {X}\right) \leq \delta_ {B}. +$$ + +Then the following bound holds: + +$$ +\left| \operatorname {I I M M} (Q) - \operatorname {I I M M} (P) \right| \leq \frac {\delta_ {A}}{2} + \delta_ {B}. +$$ + +Furthermore, if one computes empirical estimates $\widehat{\mathrm{IMM}}(P)$ and $\widehat{\mathrm{IMM}}(Q)$ using $N$ independent samples, then with probability at least $1 - \eta$ , + +$$ +\left| \widehat {\operatorname {I M M}} (Q) - \widehat {\operatorname {I M M}} (P) \right| \leq \frac {\delta_ {A}}{2} + \delta_ {B} + \epsilon_ {N}, +$$ + +where + +$$ +\epsilon_ {N} = O \Big (\sqrt {\frac {\log (1 / \eta)}{N}} \Big) +$$ + +accounts for the sampling error. + +Our empirical results have indicated that IIMM is strongly correlated with performance gains and catastrophic forgetting. The bound provided here offers a theoretical explanation: only significant reorganization of the embedding space (reflected by a large Wasserstein distance) can cause a substantial shift in IIMM. Therefore, a high IIMM change can be interpreted as evidence of a considerable shift in the model's internal representation, likely correlating with notable performance changes. + +# 4. Experiments + +Datasets To evaluate the relationships between transferability measures and model performance changes following fine-tuning, we compiled a diverse collection of nine commonly used, publicly available computer vision benchmarks + +provided by PyTorch [26, 39, 64] as our testing dataset, with the goal of minimizing domain overlap and covering a range of coarse and granular classification tasks. These datasets included Stanford Cars [21], CIFAR100 [22], the Describable Textures Dataset (DTD) [6], satellite imagery datasets EuroSAT [14] and RESISC45 [5], the German Traffic Sign Recognition Benchmark (GTSRB) [45], MNIST [23], the Scene Understanding Benchmark (SUN397) [52], and the Street View House Numbers (SVHN) dataset [33]. For each dataset, we generated validation splits by randomly sampling $10\%$ of the respective training data, capping the validation split size at 5000 samples for larger datasets. + +Models and fine-tuning methods To investigate the applicability of IIMM to vision-language encoders, we conducted all experiments using two distinct groups of evaluations schema: one in which the base pre-trained model varies while the fine-tuning strategy is held constant, and another group in which the base pre-trained model is held constant while the fine-tuning strategy varies. We will refer to these evaluation schema as the base model evaluation and PEFT evaluation, respectively. + +For the base model evaluation, we perform full finetuning over all parameters for the following four pre-trained models: CLIP, EVA-02-CLIP [12], SigLIP [60], and CoCa [59], using a ViT-B/32 image backbone for CLIP and CoCa and a ViT-B/16 image backbone for SigLIP and EVA-02-CLIP. + +For the PEFT evaluation, we utilized a CLIP model with a ViT-B/32 image backbone as the base model and evaluate four common parameter-efficient fine-tuning (PEFT) methods: CLIP-Adapter [13], BitFit (bias tuning) [3], LoRA [16], and a method we call Attention-Weight Tuning, which simply refines the fine-tuning of all attention layer weights in Touvron et al. [47] to just the attention layers' inprojection weights. + +Baselines While no methods currently exist for explicitly predicting gains in accuracy from fine-tuning for dual-encoder vision-language models, many related methods exist for stand-alone vision encoders. In particular, we compare against several recent or established measures self-described as transferability estimators. The six comparison measure are: TransRate [17], Self-challenging Fisher Discriminant Analysis (SFDA) [42], Neural Collapse informed Transferability Index (NCTI) [51], Efficient Multitask Model Selector EMMS [43], Gaussian Bhattacharyya Coefficient (GBC) [38], and H-score [2]. + +We chose these comparison measures to span the definitions and applications of self-described transferability measures. Those that estimate the log-likelihood are H-Score, SFDA, TransRate, and EMMS. NCTI functions as an estimator of the gap between the embedding geometry of the + +pre-trained model and the embedding geometry of a hypothetical fine-tuned model [51]. The GBC quantifies transferability as a measure of overlap in the feature space between embedding clusters of different class labels [38]. + +Evaluation measures We evaluate performance of the transferability measures through the strength of their relationship to three variables: accuracy on the target test data after fine-tuning on the target task training data; gain in performance from the pre-trained model to the fine-tuned model scaled by the pre-trained model error, which we refer to as gain over zero-shot error; and average accuracy across the remaining eight non-target datasets to measure loss in performance from the pre-trained model on nontarget tasks. + +Definition 1 (Gain over zero-shot error). For a given task, we define $A_{zs}$ as the zero-shot accuracy and $A_{ft}$ as the fine-tuned accuracy. We then define the gain over zero-shot error as, + +$$ +\frac {A _ {f t} - A _ {z s}}{1 - A _ {z s}} \tag {5} +$$ + +This value normalizes for differences in the zero-shot accuracy across models and tasks. + +We first evaluate all measures' monotonic relationship to both fine-tuned accuracy and gain over zero-shot error using Kendall's $\tau$ , which is a standard evaluation measure within the literature. We then look at the top performing measures of this rank correlation and fit a least-squares model to evaluate their predictive strength, additionally looking at a least-squares model fit of the top measure against the average accuracy on off-target tasks. + +Training details To establish the ground truth for changes in model in-domain (ID) and out-of-domain (OOD) accuracy following fine-tuning on each dataset, we started by separately fine-tuning each of the base models on the nine datasets in our dataset suite. Following training on a given dataset, each model was tested on all nine datasets, with in-domain accuracy calculated as its accuracy on the test split of the training dataset and out-of-domain accuracies as the accuracies obtained on the remaining datasets test split. Each fine-tuning run consisted of 30 epochs of gradient descent using an SGD optimizer and a batch size of 128 and incorporated an exhaustive grid search over all relevant model hyper-parameters, examining values $\{1\mathrm{e} - 2,1\mathrm{e} - 3,1\mathrm{e} - 4,1\mathrm{e} - 5\}$ for the learning rate, $\{1\mathrm{e} - 3,1\mathrm{e} - 4,1\mathrm{e} - 5\}$ for the weight decay, $\{2,4,8,16\}$ for LoRA model rank, and $\{4,8,16,32\}$ for CLIP-Adapter reduction. The initial image and text embeddings were then generated per dataset by a forward pass of each dataset's train split through the pre-trained base models and normalizing to unit vectors. + +
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueτp-valueτp-valueτp-valueτ
IIMM< 10-30.890.0010.830.0020.780.0020.78
GBC0.0020.780.0060.720.0020.780.0010.83
TransRate0.0130.670.0250.610.0130.670.0060.72
EMMS0.0250.610.3580.280.0250.610.1190.44
NCTI0.919-0.060.028-0.590.8180.070.827-0.06
SFDA0.2600.330.919-0.060.1800.390.3580.28
H-Score0.7610.110.002-0.780.0750.500.013-0.67
+ +Table 1. Kendall's $\tau$ correlation of score and gain over zero-shot error. Larger $\tau$ values are better.* + +
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueτp-valueτp-valueτp-valueτ
IIMM0.0060.720.0060.720.0130.670.0250.61
GBC0.0250.610.0060.720.0130.670.0130.67
TransRate0.0750.500.0250.610.0450.560.0450.56
EMMS0.0130.670.0250.610.0750.500.0250.61
NCTI1.0000.000.173-0.370.818-0.070.8270.06
SFDA0.0750.500.6120.170.0750.500.2600.33
H-Score0.6120.170.045-0.560.0250.610.075-0.50
+ +Table 2. Kendall's $\tau$ correlation of score and accuracy. Larger $\tau$ values are better.* + +# 4.1. Results + +We first look at Kendall's $\tau$ correlation to determine the strength of the monotonic relationship of our proposed and comparison measures to both the fine-tuned gain over zero-shot error and fine-tuned accuracy. + +The IIMM and the GBC have significantly stronger monotonic relationships than the other measures for both gain over zero-shot error, as shown in Tab. 1, and accuracy, as shown in Tab. 2. In both tables, the IIMM and the GBC are the most strongly correlated measures and display the greatest significance. Upon further investigation, the IIMM displays a mostly linear relationship with gain over zero-shot error, while the GBC displays a mostly exponential relationship with gain over zero-shot error. This allows us to fit and evaluate a predictive model using least-squares regression of the IIMM with gain over zero-shot error, and the GBC with the log of gain over zero-shot error. The results of these evaluations can be seen in Tab. 3, which shows the significance of the slope coefficient and the $R^2$ value, an estimate of the amount of variation within the data explained by its linear relationship to the measure. The IIMM has a stronger fit across all base pre-trained models compared to the GBC, and indeed shows itself to have a strong linear relationship to gain over zero-shot error with $R^2 > 0.89$ for all base models. Without the ability to capture cross-modal alignment, the compared uni-modal metrics expectedly demonstrate worse performance when com + +pared to the IIMM, underscoring the value of an integrated vision-language metric. + +This positive linear relationship between the IIMM and in-domain gain over zero-shot error is clearly seen in the top plots of Fig. 2. The bottom plots of Fig. 2 show the negative linear relationship of the IIMM with the average change in accuracy across a suite of eight out-of-domain tasks. For all base pre-trained models aside from CLIP, the IIMM also shows a strong linear relationship to forgetting, as evidenced by the significance of the slope coefficient and high $R^2$ values. These relationships hold true across parameter-efficient fine-tuning methods, as shown in Fig. 3. We see with CLIP-Adapter that its overall worse in-domain performance attenuates the strength of the relationship between IIMM and gain over zero-shot error, and predictably, that a lower in-domain performance gain resulted in significantly less catastrophic forgetting. CLIP-Adapter's design — injecting task kernels atop a frozen backbone — results in smaller, localized representational shifts, resulting in weaker in-distribution performance gains while preventing greater out-of-distribution performance loss. Since the IIMM is correlated with near-optimal performance gains, fine-tuning methods that result in sub-optimal gains may yield lower IIMM correlations. + +These results provide evidence that the IIMM is strongly correlated with fine-tuning gains over zero-shot error. Un + +
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueR²p-valueR²p-valueR²p-valueR²
IIMM< 10-30.92< 10-30.89< 10-30.92< 10-30.92
GBC0.0040.720.0060.680.0150.590.0050.71
+ +Table 3. Evaluation of linear fit of the IIMM to gain over zero-shot error and the GBC to the log of gain over zero-shot error. The significance of the slope coefficient is given by the p-value, and the strength of the fit is evaluated with the $R^2$ value.* + +![](images/217baa5082a3ae5e039d150e69a16730c276d98da8b4e205abc790382876a615.jpg) +Figure 2. Performance gains and losses versus the IIMM per pre-trained model, with linear regression fit and $96\%$ confidence intervals. The top plots show the relationship between the on-target (ID) task IIMM to the on-target task performance gains over zero-shot error, while the bottom plots show the relationship between on-target task IIMM and the average performance loss on off-target (OOD) tasks. + +like GBC, the second-highest correlated measure, the IIMM is bounded in $[-1,1]$ and extreme values are informative in isolation. + +# 4.2. Exploring embedding features + +To empirically validate the provided theoretical justification for leveraging multi-modal information to improve transferability estimation in vision-language models, we independently evaluated all relevant components of CLIP's embedding geometry, known to manifest as a hypersphere. Using (3), we separately computed the average pairwise similarities for the image and text modalities and fit each to the accuracy gain over zero-shot error. We then applied (2) to compute the degree of incorrect label alignment across the hypersphere and also fit the measure to the accuracy gain over zero-shot error. The final results of these evaluations can be seen in Tab. 4. + +It is observed that the features of the embedding structure most predictive of accuracy gains are the intra-images dis + +tance and the incorrect label alignment. Based on these results, the correct label alignment and intra-texts distance are generally uninformative of accuracy gains with fine-tuning. As described in Sec. 3, each of these terms corresponds to a feature of the shared embedding space related to learning and forgetting. We include additional results of the Pearson correlations between different convex combinations of our inter- and intra- terms and gain over zero-shot error in the supplementary material. This empirical evidence successfully verifies that the terms combined have a stronger relationship with learning and forgetting than they do independently. + +# 5. Conclusion + +In this work, we propose the Inter-Intra Modal Measure (IIMM) as a novel metric for predicting fine-tuning performance in vision-language models. We demonstrate that the IIMM strongly correlates with both learning gains and catastrophic forgetting, offering a simple yet powerful met + +![](images/eeff330e2b2dafa96fe48e0df5e5d8b56bba36b7bca4780eb94cf566c76799ec.jpg) +Figure 3. Performance gains and losses versus the IIMM per PEFT method, with linear regression fit and $96\%$ confidence intervals. The top plots show the relationship between the on-target task (ID) IIMM and on-target task performance gains over zero-shot error, while the bottom plots show the relationship between on-target task IIMM and the average performance loss on off-target (OOD) tasks. + +
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueR²p-valueR²p-valueR²p-valueR²
Incorrect Alignment0.0010.820.0040.720.0120.620.0010.80
Intra-Images Distance< 10-30.88< 10-30.860.0010.81< 10-30.88
Intra-Texts Distance0.0840.370.0340.500.0190.570.2650.17
+ +Table 4. Linear fit of measure to gain over zero-shot error with full fine-tuning by pre-trained model. The significance of the slope coefficient is given by the p-value, and the $R^2$ values represent the proportion of variation within the data modeled by the linear relationship. + +ric for estimating fine-tuning outcomes. Because the IIMM is bounded, an extreme value in isolation can inform practitioners whether a model will have limited or significant gains from fine-tuning. In addition, our results show that the IIMM can be leveraged to fit predictive linear models for estimating fine-tuning accuracy gains across diverse tasks with minimal computational overhead. We demonstrate that our proposed measure has a stronger relationship to both fine-tuned accuracy and accuracy changes with fine-tuning than baselines. We further show the IIMM is informative of performance changes across parameter efficient fine-tuning methods, highlighting some methods are particularly susceptible to catastrophic forgetting when the zero-shot IIMM is high. Beyond empirical validation, we provide a theoretical bound linking changes in IIMM to the Wasserstein distance between pre- and post-fine-tuning embedding distributions, ensuring stability in IIMM predictions with respect to changes in representation space. Given the IIMM measures changes in embedding geometry prior to any task + +specific heads, the metric is also task-agnostic, enabling its application to a variety of downstream tasks. We encourage future research to validate such use in broader applications. Overall, our work provides a practical and theoretically grounded tool for optimizing fine-tuning decisions in vision-language models, paving the way for more efficient and informed adaptation of foundation models to new tasks. + +Limitations It is possible that the IIMM could show more variation when compared across a more diverse set of base models. While extreme values of the IIMM are heuristically informative in isolation, fitting a model for transfer accuracy estimation requires first fine-tuning across a small but diverse set of benchmarks. This could be an avenue of future research, as there might be additional information within the embedding geometry that allows for fine-tuning performance prediction without requiring information from fine-tuning on other tasks. + +# Acknowledgements + +The authors acknowledge the MIT Lincoln Laboratory Supercomputing Center for providing the high-performance computing resources that have contributed to the research results reported in this paper. + +DISTRIBUTION STATEMENT A. Approved for public release. Distribution is unlimited. This material is based upon work supported by the Under Secretary of Defense for Research and Engineering under Air Force Contract No. FA8702-15-D-0001 or FA8702-25-D-B002. Any opinions, findings, conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Under Secretary of Defense for Research and Engineering. + +# References + +[1] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational Models Defining a New Era in Vision: A Survey and Outlook, 2023. 1 +[2] Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An Information-Theoretic Approach to Transferability in Task Transfer Learning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2309-2313, Taipei, Taiwan, 2019. IEEE. 5 +[3] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1-9, Dublin, Ireland, 2022. Association for Computational Linguistics. 5 +[4] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1 +[5] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 5 +[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing Textures in the Wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, Columbus, OH, USA, 2014. IEEE. 5 +[7] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large Scale Fine-Grained Categorization and Domain-Specific Transfer Learning, 2018. 3 +[8] Yuxuan Ding, Lingqiao Liu, Chunna Tian, Jingyuan Yang, and Haoxuan Ding. Don't Stop Learning: Towards Continual Learning for the CLIP Model, 2022. 1, 3 +[9] Yuhe Ding, Bo Jiang, Aijing Yu, Aihua Zheng, and Jian Liang. Which model to transfer? a survey on transferability estimation. arXiv preprint arXiv:2402.15231, 2024. 3 + +[10] Kshitij Dwivedi and Gemma Roig. Representation Similarity Analysis for Efficient Task Taxonomy & Transfer Learning. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12379-12388, Long Beach, CA, USA, 2019. IEEE. 3 +[11] Zhihao Fan, Zhongyu Wei, Jingjing Chen, Siyuan Wang, Ze-jun Li, Jiarong Xu, and Xuanjing Huang. A unified continuous learning framework for multi-modal knowledge discovery and pre-training, 2022. 3 +[12] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331, 2023.5 +[13] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. CLIP-Adapter: Better Vision-Language Models with Feature Adapters, 2021. 5 +[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Introducing Eurosat: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. In IGARSS 2018 - 2018 IEEE International Geoscience and Remote Sensing Symposium, pages 204-207, Valencia, 2018. IEEE. 5 +[15] Dapeng Hu, Shipeng Yan, Qizhengqiu Lu, Lanqing Hong, Hailin Hu, Yifan Zhang, Zhenguo Li, Xinchao Wang, and Jiashi Feng. How well does self-supervised pre-training perform with streaming data?, 2022. 3 +[16] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In ICLR. arXiv, 2022. 5 +[17] Long-Kai Huang, Junzhou Huang, Yu Rong, Qiang Yang, and Ying Wei. Frustratingly easy transferability estimation. In International Conference on Machine Learning, pages 9201-9225. PMLR, 2022. 1, 3, 5 +[18] Joana Palés Huix, Adithya Raju Ganeshan, Johan Fredin Haslum, Magnus Söderberg, Christos Matsoukas, and Kevin Smith. Are Natural Domain Foundation Models Useful for Medical Image Classification? In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 7619-7628. IEEE, 2024. 1 +[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2 +[20] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 3 +[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D Object Representations for Fine-Grained Categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, Sydney, Australia, 2013. IEEE. 5 + +[22] Alex Krizhevsky. Learning Multiple Layers of Features from Tiny Images, 2009. 5 +[23] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, Nov./1998. 5 +[24] Hao Li, Charless Fowlkes, Hao Yang, Onkar Dabeer, Zhuowen Tu, and Stefano Soatto. Guided Recommendation for Model Fine-Tuning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3633-3642, Vancouver, BC, Canada, 2023. IEEE. 3 +[25] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before Fuse: Vision and Language Representation Learning with Momentum Distillation. In Advances in Neural Information Processing Systems, pages 9694-9705. Curran Associates, Inc., 2021. 2 +[26] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning. In First Workshop of Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022, 2022. 2, 5 +[27] C. Liao, T. Tsiligkaridis, and B. Kulis. Descriptor and word soups: Overcoming the parameter efficiency accuracy trade-off for out-of-distribution few-shot learning. In CVPR, 2024. 1 +[28] C. Liao, C. So, T. Tsiligkaridis, and B. Kulis. Multimodal unsupervised domain generalization by retrieving across the modality gap. In ICLR, 2025. 1 +[29] Yong Lin, Lu Tan, Hangyu Lin, Zeming Zheng, Renjie Pi, Jipeng Zhang, Shizhe Diao, Haoxiang Wang, Han Zhao, Yuan Yao, and Tong Zhang. Speciality vs Generality: An Empirical Study on Catastrophic Forgetting in Fine-tuning Foundation Models, 2023. 3 +[30] Michael McCloskey and Neal J. Cohen. Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem. In *Psychology of Learning and Motivation*, pages 109–165. Elsevier, 1989. 3 +[31] S. Menon and C. Vondrick. Visual classification via description from large language models. In ICLR, 2023. 1 +[32] Jishnu Mukhoti, Yarin Gal, Philip HS Torr, and Puneet K Dokania. Fine-tuning can cripple your foundation model; preserving features may be the solution. arXiv preprint arXiv:2308.13320, 2023. 1 +[33] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading Digits in Natural Images with Unsupervised Feature Learning. NeurIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 5 +[34] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. LEEP: A New Measure to Evaluate Transferability of Learned Representations. In Proceedings of the 37th International Conference on Machine Learning, pages 7294-7305. PMLR, 2020. 3 +[35] Zixuan Ni, Siliang Tang, and Yueting Zhuang. Self-supervised class incremental learning, 2021. 3 +[36] Zixuan Ni, Longhui Wei, Siliang Tang, Yueting Zhuang, and Qi Tian. Continual Vision-Language Representation Learning with Off-Diagonal Information, 2023. 1, 3 + +[37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3 +[38] Michal Páncy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Transferability Estimation using Bhattacharyya Class Separability, 2022. 1, 5 +[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 5 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2 +[41] K. Roth, J. M. Kim, A. S. Koepke, O. Vinyals, C. Schmid, and Z. Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 1 +[42] Wenqi Shao, Xun Zhao, Yixiao Ge, Zhaoyang Zhang, Lei Yang, Xiaogang Wang, Ying Shan, and Ping Luo. Not all models are equal: Predicting model transferability in a self-challenging fisher space. In European Conference on Computer Vision, pages 286-302. Springer, 2022. 1, 3, 5 +[43] Wenqi Shao, Chonghe Jiang, Kaipeng Zhang, Yu Qiao, Ping Luo, et al. Foundation model is efficient multimodal multitask model selector. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 5 +[44] Tejas Srinivasan, Ting-Yun Chang, Leticia Pinto Alva, Georgios Chochlakis, Mohammad Rostami, and Jesse Thomason. CLiMB: A continual learning benchmark for vision-and-language tasks. In Advances in Neural Information Processing Systems, pages 29440-29453. Curran Associates, Inc., 2022. 3 +[45] J. Stallkamp, M. Schlipsing, J. Salmen, and C. Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. *Neural Networks*, 32:323–332, 2012. 5 +[46] P. Teterwak, K. Saito, T. Tsiligkaridis, B. A. Plummer, and K. Saenko. Is large-scale pretraining the secret to good domain generalization? In ICLR, 2025. 1 +[47] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about Vision Transformers, 2022. 5 +[48] Anh Tran, Cuong Nguyen, and Tal Hassner. Transferability and Hardness of Supervised Classification Tasks. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1395-1405, Seoul, Korea (South), 2019. IEEE. 3 +[49] Kevin Vogt-Lowell, Noah Lee, Theodoros Tsiligkaridis, and Marc Vaillant. Robust fine-tuning of vision-language models for domain generalization. In 2023 IEEE High Performance + +Extreme Computing Conference (HPEC), pages 1-7. IEEE, 2023. 1 +[50] Tongzhou Wang and Phillip Isola. Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere. In Proceedings of the 37th International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 3 +[51] Zijian Wang, Yadan Luo, Liang Zheng, Zi Huang, and Mahsa Baktashmotlagh. How far pre-trained models are from neural collapse on the target dataset informs their transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5549-5558, 2023. 1, 3, 5 +[52] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, San Francisco, CA, USA, 2010. IEEE. 5 +[53] Yiqun Xie, Zhihao Wang, Weiye Chen, Zhili Li, Xiaowei Jia, Yanhua Li, Ruichen Wang, Kangyang Chai, Ruohan Li, and Sergii Skakun. When are foundation models effective? understanding the suitability for pixel-level classification using multispectral imagery. CoRR, abs/2404.11797, 2024. 1 +[54] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive Pretraining for Zero-shot Video-Text Understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787–6800, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2 +[55] Chao Yi, De-Chuan Zhan, and Han-Jia Ye. Bridge the Modality and Capacity Gaps in Vision-Language Model Selection, 2024. 3 +[56] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In International Conference on Machine Learning, pages 12133-12143. PMLR, 2021. 3 +[57] E. Yu, C. Liao, S. Ravi, T. Tsiligkaridis, and B. Kulis. Image-capture encoding for improving zero-shot generalization. In WACV, 2025. 1 +[58] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23219-23230. IEEE. 1 +[59] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 5 +[60] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 5 +[61] Dan Zhang, Tao Feng, Lilong Xue, Yuandong Wang, Yuxiao Dong, and Jie Tang. Parameter-efficient fine-tuning for + +foundation models. arXiv preprint arXiv:2501.13787, 2025. 1 +[62] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and Curtis P. Langlotz. Contrastive Learning of Medical Visual Representations from Paired Images and Text. In Proceedings of the 7th Machine Learning for Healthcare Conference, pages 2-25. PMLR, 2022. 2 +[63] Zangwei Zheng, Mingyuan Ma, Kai Wang, Ziheng Qin, Xi-angyu Yue, and Yang You. Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 19068-19079. IEEE. 1 +[64] Orr Zohar, Kuan-Chieh Wang, Shih-Cheng Huang, and Serena Yeung. LOVM: Language-Only Vision Model Selection. In Conference on Neural Information Processing Systems (Neurips), 2023. 3, 5 \ No newline at end of file diff --git a/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/images.zip b/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cb57b2cc163c23a77eed335030024289bd3b96fe --- /dev/null +++ b/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:353c40cfbc4882e9f0261f980e6fa0632c0ab1fd7bac643ac8c17eed9b2aaaa2 +size 407897 diff --git a/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/layout.json b/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1657f10eb6590c12aa8690c4101383dec0aedd4b --- /dev/null +++ b/2025/The Inter-Intra Modal Measure_ A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models/layout.json @@ -0,0 +1,8307 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 552, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 552, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 552, + 140 + ], + "type": "text", + "content": "The Inter-Intra Modal Measure: A Predictive Lens on Fine-Tuning Outcomes in Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 156, + 161, + 455, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 161, + 455, + 175 + ], + "spans": [ + { + "bbox": [ + 156, + 161, + 455, + 175 + ], + "type": "text", + "content": "Laura Niss* Kevin Vogt-Lowell* Theodoros Tsiligkaridis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 170, + 177, + 441, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 177, + 441, + 189 + ], + "spans": [ + { + "bbox": [ + 170, + 177, + 441, + 189 + ], + "type": "text", + "content": "{laura.niss, kevin.vogt-lowell, ttsili}@ll.mit.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 244, + 190, + 365, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 244, + 190, + 365, + 203 + ], + "spans": [ + { + "bbox": [ + 244, + 190, + 365, + 203 + ], + "type": "text", + "content": "MIT Lincoln Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 215, + 203, + 392, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 203, + 392, + 217 + ], + "spans": [ + { + "bbox": [ + 215, + 203, + 392, + 217 + ], + "type": "text", + "content": "244 Wood St, Lexington, MA 02421" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 270, + 297, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 629 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 629 + ], + "type": "text", + "content": "The fine-tuning of large vision-language foundation models remains an underexplored area, particularly regarding its impact on learning gains and catastrophic forgetting. Inspired by the significance of modality gaps in contrastive dual-encoders, we introduce the Inter-Intra Modal Measure (IIMM)—a predictive metric that quantifies the relationship between intra-modal image embedding similarity and inter-modal misalignment. Through extensive empirical analysis across four state-of-the-art vision-language models and five fine-tuning techniques, we establish a strong linear relationship: tasks with higher IIMM scores yield greater in-domain performance improvements but suffer from more pronounced out-of-domain degradation, with some parameter-efficient fine-tuning (PEFT) methods exhibiting severe forgetting. Compared to existing transferability measures, the IIMM demonstrates significantly stronger predictive power for accuracy changes post fine-tuning in dual-encoder models. Moreover, we provide a theoretical bound, proving that changes in IIMM are limited by the Wasserstein distance between pre- and post-fine-tuning embedding distributions, ensuring its stability and robustness as a predictive measure. With only a single forward pass of the target data, practitioners can leverage this key insight to evaluate the degree to which a model can be expected to improve following fine-tuning. When combined with prior knowledge of a model's performance across diverse tasks, the IIMM further enhances transferability predictions for novel tasks, offering a lightweight yet effective tool for guiding model adaptation strategies. Our code is provided at https://github.com/mit-ll/IIMM." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 651, + 137, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 137, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 137, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 671, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 296, + 696 + ], + "type": "text", + "content": "Large vision-language foundation models have demonstrated superior zero-shot generalization compared to tra" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 555, + 389 + ], + "type": "text", + "content": "ditional vision models, leveraging web-scale pre-training to achieve state-of-the-art performance across diverse tasks without task-specific supervision [39]. Advances in enhancing zero-shot generalization have also been proposed [31, 41, 57]. Fine-tuning can further enhance performance on specific downstream applications [4, 27, 28, 46, 61]. However, fine-tuned foundation models still struggle on certain tasks, particularly in specialized fields such as satellite and medical image classification [1, 18, 49, 53]. Given the computational cost of fine-tuning, determining whether adaptation will lead to meaningful improvements remains a crucial challenge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 390, + 556, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 556, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 556, + 545 + ], + "type": "text", + "content": "Beyond resource efficiency, fine-tuning also introduces the risk of catastrophic forgetting—where a model loses previously learned knowledge while improving on the new task. This effect is particularly pronounced in dual-encoder vision-language models, where fine-tuning disrupts the alignment between image and text embedding spaces, degrading the model's ability to generalize to out-of-domain tasks [8, 36]. This creates an inherent trade-off between learning and forgetting, making it critical for practitioners to anticipate the effects of fine-tuning before allocating computational resources [32, 58, 63]. A predictive metric that captures both learning gains and forgetting risks would provide a valuable tool for optimizing fine-tuning decisions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 557, + 714 + ], + "type": "text", + "content": "The challenge of estimating a model's fine-tuned performance a priori falls under the domain of model transferability [38, 51]. Common transferability measures assess a pretrained model's suitability for a given task by analyzing its latent representations. However, most traditional methods rely on source data and task labels, making them incompatible with contrastively trained foundation models that lack explicit supervision [43]. Furthermore, newer transferability metrics are designed for uni-modal settings, failing to account for inter-modal interactions in vision-language models. Among approaches that estimate post-fine-tuning performance, evaluation is often limited to rank correlation rather than directly predicting performance changes [17, 42]." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 135, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 135, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 135, + 713 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2396" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 121, + 70, + 491, + 266 + ], + "blocks": [ + { + "bbox": [ + 121, + 70, + 491, + 266 + ], + "lines": [ + { + "bbox": [ + 121, + 70, + 491, + 266 + ], + "spans": [ + { + "bbox": [ + 121, + 70, + 491, + 266 + ], + "type": "image", + "image_path": "3f2363d156a0bd9ba19e94948c4788ac2a0d7bb192676307db92febb5ada5b9c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 284, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 54, + 284, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 555, + 319 + ], + "type": "text", + "content": "Figure 1. The Inter-Intra Modal Measure (IIMM) combines two key aspects of embedding geometry: intra-modal image embedding similarity, and inter-modal alignment between image and text embeddings. This measure captures information about the embedding geometry that demonstrates a strong relationship to performance gains observed with fine-tuning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 339, + 296, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 339, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 56, + 339, + 296, + 614 + ], + "type": "text", + "content": "Present work To address these limitations, we introduce the Inter-Intra Modal Measure (IIMM), a novel metric that quantifies both intra-modal representation spread and inter-modal alignment in contrastive vision-language models, capturing the fundamental trade-off between learning and forgetting without the need for fine-tuning. We compare our measure against several standard and recent transferability measures in both rank correlation and strength of predictive fit to performance gains after fine-tuning. To evaluate our measure, we perform fine-tuning across nine well-known vision classification tasks using four state-of-the-art vision-language models and five fine-tuning techniques, analyzing in-domain improvements as well as induced catastrophic forgetting across eight out-of-domain tasks. We compare our measure against six transferability measures applicable to vision-language encoders, and show that our measure has a significantly stronger monotonic relationship with both fine-tuned in-domain accuracy and in-domain performance gains scaled by the zero-shot error. When fitting a simple linear regression of the IIMM to the scaled in-domain accuracy gains, we achieve " + }, + { + "bbox": [ + 56, + 339, + 296, + 614 + ], + "type": "inline_equation", + "content": "R^2 > 0.85" + }, + { + "bbox": [ + 56, + 339, + 296, + 614 + ], + "type": "text", + "content": " for all base models. This offers a practical and theoretically grounded tool for model adaptation in vision-language tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 653 + ], + "type": "text", + "content": "Contributions In summary, our work provides the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 654, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 654, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 56, + 654, + 295, + 701 + ], + "type": "text", + "content": "1. We introduce the Inter-Intra Modal Measure (IIMM), a standalone metric that integrates inter-modal misalignment and intra-modal uniformity to predict fine-tuning outcomes in dual-encoder vision-language models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "2. We derive the IIMM directly from contrastive learning" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 325, + 339, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 339, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 325, + 339, + 555, + 373 + ], + "type": "text", + "content": "principles, establishing a clear connection between the geometry of the embedding space and transferability estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 375, + 556, + 542 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 375, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 375, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 314, + 375, + 555, + 447 + ], + "type": "text", + "content": "3. We prove a theoretical bound showing that changes in the IIMM are constrained by the 1-Wasserstein distance between pre- and post-fine-tuning embedding distributions. This result guarantees that significant shifts in the representation space will lead to substantial changes in IIMM, underscoring its stability as a predictive measure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 447, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 542 + ], + "type": "text", + "content": "4. We show that the zero-shot IIMM has a strong, positive linear relationship with learning and a negative relationship with forgetting post fine-tuning. We also provide empirical evidence that for any vision-language dual-encoder, fine-tuning on only a few standard vision benchmarks is sufficient to fit a strong linear predictor of fine-tuning accuracy gains with the IIMM for transfer estimation of novel tasks." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 559, + 399, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 559, + 399, + 572 + ], + "spans": [ + { + "bbox": [ + 314, + 559, + 399, + 572 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Dual-encoders and the modality gap Multi-modal dual-encoders trained using contrastive learning consist of two models jointly trained to map each modality into a shared representation space [19, 25, 40, 54, 62]. Previous research has highlighted the existence of a modality gap within this joint embedding space that results from the encoders initializing embeddings within random cones, and that contrastive loss maintains this initialized gap [26]. Additionally, empirical evidence shows that manipulating this gap can impact zero-shot performance. While catastrophic forgetting has been proven to significantly affect traditional su" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2397" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "pervised learning techniques [20, 30], recent research has shown that self-supervised, uni-modal models do not suffer from such severe forgetting [15, 35]. Yet, with multimodal foundation models, such as CLIP, continuous training can cause misalignment of the image and text embeddings, an outcome highlighted and explored in several works [8, 11, 29, 36, 44]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 173, + 296, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 173, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 296, + 352 + ], + "type": "text", + "content": "Transferability estimation A significant amount of research has focused on addressing transferability estimation with respect to uni-modal encoders. Estimation of model transferability often involves transferability metrics, or measures that predict how well knowledge learned from source tasks will transfer to new target tasks. [9] The approaches for quantifying these metrics vary greatly, with some approaches simply defining them as the log-likelihood of a target task. Methods that rely upon a log-likelihood definition are inherently well-suited for model selection because the definition relies directly upon the size of the target dataset, and within model selection the target dataset is held constant. Regardless of approach, nearly all metrics are evaluated via rank correlation across a model set [17, 42, 43, 51]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 353, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 353, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 353, + 296, + 496 + ], + "type": "text", + "content": "Some modern approaches include training small neural networks and comparing the similarity of resultant task embeddings [10], training a full neural network on large amounts of models trained across many tasks to learn a similarity score function [24], and comparing source training data to new target task data to estimate similarity [7]. Some classic methods, such as LEEP [34] and NCE [48], rely upon information from the source training data and are thus inapplicable to self-supervised foundation models; they also cannot be applied to dual-encoders. However, there are a range of uni-modal methods that can be easily applied to the vision encoder of a dual-encoder model [17, 43, 56]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 497, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 569 + ], + "type": "text", + "content": "For multi-modal models, some recent research has involved zero-shot model selection using language to infer the transferability of a dataset, such as LOVM [64] and SWAB [55]. These methods do not claim to estimate model performance after fine-tuning, and are specific to the zero-shot setting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 580, + 211, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 211, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 211, + 594 + ], + "type": "text", + "content": "3. Inter-Intra Modal Measure" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 601, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 649 + ], + "type": "text", + "content": "In this section, we provide a principled derivation of the IIMM metric from contrastive learning principles. In addition, we prove a theoretical bound and provide intuition behind the metric for predicting fine-tuning outcomes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Notation Let " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " denote a target task, and let " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " be the set of tasks under evaluation. Consider a training dataset for task " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " composed of image-text pairs. Let " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " denote the set of image embeddings, where each " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "x \\in X \\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ". Similarly, let" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " denote the set of text embeddings, where each " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "y \\in Y \\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": ". For every image " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "y(x) \\in Y" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " be the text embedding corresponding to its correct label." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "content": " be two probability measures on a metric space " + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "inline_equation", + "content": "(\\mathcal{X},d)" + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "content": ". The 1-Wasserstein distance between " + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 342, + 147, + 526, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 147, + 526, + 172 + ], + "spans": [ + { + "bbox": [ + 342, + 147, + 526, + 172 + ], + "type": "interline_equation", + "content": "W _ {1} (\\mu , \\nu) = \\inf _ {\\gamma \\in \\Gamma (\\mu , \\nu)} \\int_ {\\mathcal {X} \\times \\mathcal {X}} d (x, y) d \\gamma (x, y),", + "image_path": "c0531bb71e1efc07beb0b0131de9b757fb2dedc6ac5400d45e3b6a8a6ad21736.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "inline_equation", + "content": "\\Gamma (\\mu ,\\nu)" + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "text", + "content": " denotes the set of all couplings (i.e., joint distributions) with marginals " + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "inline_equation", + "content": "\\nu" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 207, + 553, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 553, + 220 + ], + "type": "text", + "content": "3.1. Derivation via Contrastive Learning Principles" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "text", + "content": "We begin with the widely used InfoNCE loss in contrastive learning [37]. For an image embedding " + }, + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "inline_equation", + "content": "x \\in X" + }, + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "text", + "content": " and its corresponding (correct) text embedding " + }, + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "inline_equation", + "content": "y \\in Y" + }, + { + "bbox": [ + 313, + 224, + 553, + 270 + ], + "type": "text", + "content": ", the loss is defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 274, + 577, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 577, + 326 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 577, + 326 + ], + "type": "interline_equation", + "content": "L (x, y, \\{y ^ {\\prime} \\}) = - \\log \\frac {\\exp \\left(\\frac {x ^ {\\top} y}{\\tau}\\right)}{\\exp \\left(\\frac {x ^ {\\top} y}{\\tau}\\right) + \\sum_ {y ^ {\\prime} \\in Y \\backslash \\{y (x) \\}} \\exp \\left(\\frac {x ^ {\\top} y ^ {\\prime}}{\\tau}\\right)},", + "image_path": "6bd3decda0e5aa427bb545bc8294c10385c21fc79ea642ceac34bf46c45d1662.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 326, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 553, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 326, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\tau > 0" + }, + { + "bbox": [ + 313, + 326, + 553, + 361 + ], + "type": "text", + "content": " is the temperature parameter and the summation is over all incorrect text embeddings " + }, + { + "bbox": [ + 313, + 326, + 553, + 361 + ], + "type": "inline_equation", + "content": "y'" + }, + { + "bbox": [ + 313, + 326, + 553, + 361 + ], + "type": "text", + "content": " (i.e. " + }, + { + "bbox": [ + 313, + 326, + 553, + 361 + ], + "type": "inline_equation", + "content": "y' \\neq y(x)" + }, + { + "bbox": [ + 313, + 326, + 553, + 361 + ], + "type": "text", + "content": "). Minimizing the InfoNCE loss encourages two main effects:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 361, + 553, + 422 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 315, + 361, + 553, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 361, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 315, + 361, + 553, + 385 + ], + "type": "text", + "content": "1. Positive Pair Alignment: The numerator is maximized when the cosine similarity " + }, + { + "bbox": [ + 315, + 361, + 553, + 385 + ], + "type": "inline_equation", + "content": "x^{\\top}y" + }, + { + "bbox": [ + 315, + 361, + 553, + 385 + ], + "type": "text", + "content": " is high." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 385, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 385, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 314, + 385, + 553, + 422 + ], + "type": "text", + "content": "2. Negative Pair Separation: The denominator penalizes high similarities between " + }, + { + "bbox": [ + 314, + 385, + 553, + 422 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 314, + 385, + 553, + 422 + ], + "type": "text", + "content": " and any incorrect text embedding " + }, + { + "bbox": [ + 314, + 385, + 553, + 422 + ], + "type": "inline_equation", + "content": "y'" + }, + { + "bbox": [ + 314, + 385, + 553, + 422 + ], + "type": "text", + "content": ", thus pushing down " + }, + { + "bbox": [ + 314, + 385, + 553, + 422 + ], + "type": "inline_equation", + "content": "x^\\top y'" + }, + { + "bbox": [ + 314, + 385, + 553, + 422 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 314, + 385, + 553, + 422 + ], + "type": "inline_equation", + "content": "y' \\neq y(x)" + }, + { + "bbox": [ + 314, + 385, + 553, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 422, + 553, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 553, + 457 + ], + "type": "text", + "content": "The resulting changes in embedding geometry can be decomposed into two components: inter-modal misalignment and intra-modal uniformity." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 471, + 553, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 553, + 507 + ], + "type": "text", + "content": "Inter-modal misalignment The effect on incorrect pairs can be quantified by the average similarity between an image embedding " + }, + { + "bbox": [ + 313, + 471, + 553, + 507 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 471, + 553, + 507 + ], + "type": "text", + "content": " and the set of incorrect text embeddings:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 344, + 514, + 553, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 514, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 344, + 514, + 553, + 545 + ], + "type": "interline_equation", + "content": "S _ {\\text {i n t e r}} = \\frac {1}{| X |} \\sum_ {x \\in \\mathcal {X}} \\frac {1}{| Y | - 1} \\sum_ {y ^ {\\prime} \\in Y \\backslash \\{y (x) \\}} x ^ {\\top} y ^ {\\prime}. \\tag {2}", + "image_path": "1a564c94743a6a12f8a67bce27f122c3023d602dfaeb527df6174094636762f6.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 548, + 553, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 553, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 553, + 608 + ], + "type": "text", + "content": "A high " + }, + { + "bbox": [ + 313, + 548, + 553, + 608 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{inter}}" + }, + { + "bbox": [ + 313, + 548, + 553, + 608 + ], + "type": "text", + "content": " indicates that the model has not sufficiently separated the image embedding from non-matching text embeddings. While indicative of potential for improvement, this can also be problematic, as fine-tuning may lead to over-alignment and subsequent catastrophic forgetting." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 621, + 555, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 555, + 681 + ], + "type": "text", + "content": "Intra-modal uniformity In addition to inter-modal alignment, contrastive learning also encourages a good spread, or uniformity, of embeddings within each modality [50]. For image embeddings, we quantify uniformity by computing the average pairwise similarity:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 353, + 685, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 685, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 353, + 685, + 553, + 715 + ], + "type": "interline_equation", + "content": "S _ {\\text {i n t r a}} = \\frac {2}{| X | (| X | - 1)} \\sum_ {1 \\leq i < j \\leq | X |} x _ {i} ^ {\\top} x _ {j}. \\tag {3}", + "image_path": "28beb89a5cd351a6dea48899ffd65e6e2bc8e6350c82baa1218432f82480d84b.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2398" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "A high " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{intra}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": " implies that the image embeddings are overly clustered, suggesting that there is significant potential for further separation (or re-organization) during fine-tuning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 296, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 296, + 159 + ], + "type": "text", + "content": "Formulation and intuition Based on the above principles, we define the Inter-Intra Modal Measure (IIMM) as the average of the inter-modal and intra-modal terms:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 118, + 167, + 296, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 167, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 118, + 167, + 296, + 191 + ], + "type": "interline_equation", + "content": "\\mathrm {I I M M} = \\frac {1}{2} \\left(S _ {\\text {i n t e r}} + S _ {\\text {i n t r a}}\\right). \\tag {4}", + "image_path": "6b0c875ea602a42d3fa77bc1b26b9331f2e899180a0846ebbe2e98daf810ed71.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 198, + 296, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 329 + ], + "type": "text", + "content": "A high IIMM value indicates that the average similarity between an image and incorrect text embeddings is high and/or the image embeddings are highly clustered. This situation suggests that while there is significant capacity for learning (i.e., re-aligning the modalities), the risk of catastrophic forgetting is also high. Conversely, a low IIMM value implies that the image and text modalities are well separated (low inter-modal similarity for negatives) and the image embeddings are sufficiently dispersed. In this case, fine-tuning is expected to yield minimal performance changes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 330, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 353 + ], + "type": "text", + "content": "The IIMM metric arises naturally from the objectives of contrastive learning:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 354, + 295, + 425 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 378 + ], + "type": "text", + "content": "1. The InfoNCE loss drives positive pairs to be closely aligned and negative pairs to be well separated." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 378, + 295, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 401 + ], + "type": "text", + "content": "2. The average similarity over incorrect pairs, " + }, + { + "bbox": [ + 55, + 378, + 295, + 401 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{inter}}" + }, + { + "bbox": [ + 55, + 378, + 295, + 401 + ], + "type": "text", + "content": ", captures the degree of inter-modal misalignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 402, + 295, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 295, + 425 + ], + "type": "text", + "content": "3. The average similarity among image embeddings, " + }, + { + "bbox": [ + 55, + 402, + 295, + 425 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{intra}}" + }, + { + "bbox": [ + 55, + 402, + 295, + 425 + ], + "type": "text", + "content": ", measures intra-modal uniformity." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 426, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 473 + ], + "type": "text", + "content": "By combining these two measures as in Equation (4), the IIMM provides a principled, predictive metric that reflects both the learning potential and the risk of catastrophic forgetting in vision-language models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "type": "text", + "content": "We provide further exploration of this measure and its relation to learning, forgetting, and the uniformity of the embedding space in Sec. 4.2. Additionally, ?? provides an analysis of various convex combinations of the intra- and inter-modal sub-scores, highlighting modification of component weights as a means of evaluating each component's contributions to the predicted outcomes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 564, + 167, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 167, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 167, + 576 + ], + "type": "text", + "content": "3.2. Theoretical Bound" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "In this section, we provide a theoretical result that links the change in IIMM to the Wasserstein distance between the joint distributions of incorrect image-text pairs (and image-image pairs) before and after fine-tuning. This bound offers a rigorous guarantee: if fine-tuning only induces a small distributional shift as measured by the Wasserstein distance, then the IIMM metric will change by no more than a proportional amount. Such a result is important as it justifies the use of IIMM as a robust indicator—ensuring that minor, perhaps noisy, changes in the embeddings do not lead to erratic changes in the IIMM, and that significant changes in" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "the IIMM are indeed indicative of a meaningful shift in the model's representation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 103, + 555, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 103, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 103, + 555, + 140 + ], + "type": "text", + "content": "Theorem 1. Assume that all image and text embeddings lie on the unit sphere " + }, + { + "bbox": [ + 313, + 103, + 555, + 140 + ], + "type": "inline_equation", + "content": "S^{d - 1} \\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 103, + 555, + 140 + ], + "type": "text", + "content": ". Let the IIMM metric be defined by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 148, + 504, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 148, + 504, + 171 + ], + "spans": [ + { + "bbox": [ + 364, + 148, + 504, + 171 + ], + "type": "interline_equation", + "content": "\\operatorname {I I M M} (P) = \\frac {1}{2} \\Big (A (P) + B (P) \\Big),", + "image_path": "6258cb086e874d6cf855ea93ebd62d581820b84b07cf484145bca1aca59d6cad.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 178, + 342, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 178, + 342, + 189 + ], + "spans": [ + { + "bbox": [ + 314, + 178, + 342, + 189 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 373, + 188, + 495, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 188, + 495, + 204 + ], + "spans": [ + { + "bbox": [ + 373, + 188, + 495, + 204 + ], + "type": "interline_equation", + "content": "A (P) = \\mathbb {E} _ {(x, y ^ {\\prime}) \\sim P _ {X Y}} \\left[ x ^ {\\top} y ^ {\\prime} \\right],", + "image_path": "9d2c76c42ff357eb8fc12754c2683c27a9e11d1ff5f7c54d4001ebf9c160aeb2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 208, + 555, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 555, + 232 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 208, + 555, + 232 + ], + "type": "inline_equation", + "content": "P_{XY}" + }, + { + "bbox": [ + 313, + 208, + 555, + 232 + ], + "type": "text", + "content": " being the joint distribution of incorrect image-text pairs, and" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 367, + 239, + 503, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 239, + 503, + 255 + ], + "spans": [ + { + "bbox": [ + 367, + 239, + 503, + 255 + ], + "type": "interline_equation", + "content": "B (P) = \\mathbb {E} _ {(x, x ^ {\\prime}) \\sim P _ {X} \\otimes P _ {X}} \\left[ x ^ {\\top} x ^ {\\prime} \\right],", + "image_path": "a79a60bbf5d08498b99c5b856d79daa6e4e43a827fbb7318f684a59922c6cec3.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "inline_equation", + "content": "P_{X}" + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": " being the marginal distribution of image embeddings. Consider perturbed versions of these distributions " + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "inline_equation", + "content": "Q_{XY}" + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "inline_equation", + "content": "Q_{X}" + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": ", respectively. Assume that the 1-Wasserstein distances satisfy" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 327, + 319, + 541, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 319, + 541, + 334 + ], + "spans": [ + { + "bbox": [ + 327, + 319, + 541, + 334 + ], + "type": "interline_equation", + "content": "W _ {1} \\left(P _ {X Y}, Q _ {X Y}\\right) \\leq \\delta_ {A} \\quad a n d \\quad W _ {1} \\left(P _ {X}, Q _ {X}\\right) \\leq \\delta_ {B}.", + "image_path": "d3b70de31add3d50a4e3b267c3a8d5e96dc294dfda4a1b5f67f32e0e361775eb.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 342, + 447, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 342, + 447, + 354 + ], + "spans": [ + { + "bbox": [ + 314, + 342, + 447, + 354 + ], + "type": "text", + "content": "Then the following bound holds:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 355, + 362, + 512, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 362, + 512, + 386 + ], + "spans": [ + { + "bbox": [ + 355, + 362, + 512, + 386 + ], + "type": "interline_equation", + "content": "\\left| \\operatorname {I I M M} (Q) - \\operatorname {I I M M} (P) \\right| \\leq \\frac {\\delta_ {A}}{2} + \\delta_ {B}.", + "image_path": "ba2bf0684db6e35f07503246ac4ca8a0d7830e7a3feda545fef024363a105633.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 392, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 430 + ], + "type": "text", + "content": "Furthermore, if one computes empirical estimates " + }, + { + "bbox": [ + 313, + 392, + 554, + 430 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathrm{IMM}}(P)" + }, + { + "bbox": [ + 313, + 392, + 554, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 392, + 554, + 430 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathrm{IMM}}(Q)" + }, + { + "bbox": [ + 313, + 392, + 554, + 430 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 392, + 554, + 430 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 392, + 554, + 430 + ], + "type": "text", + "content": " independent samples, then with probability at least " + }, + { + "bbox": [ + 313, + 392, + 554, + 430 + ], + "type": "inline_equation", + "content": "1 - \\eta" + }, + { + "bbox": [ + 313, + 392, + 554, + 430 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 343, + 437, + 525, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 437, + 525, + 461 + ], + "spans": [ + { + "bbox": [ + 343, + 437, + 525, + 461 + ], + "type": "interline_equation", + "content": "\\left| \\widehat {\\operatorname {I M M}} (Q) - \\widehat {\\operatorname {I M M}} (P) \\right| \\leq \\frac {\\delta_ {A}}{2} + \\delta_ {B} + \\epsilon_ {N},", + "image_path": "6788f2a9a7481317887c535928fc9e0856e64e81230f7b59882af525c031656b.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 469, + 342, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 469, + 342, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 469, + 342, + 479 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 386, + 478, + 481, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 478, + 481, + 504 + ], + "spans": [ + { + "bbox": [ + 386, + 478, + 481, + 504 + ], + "type": "interline_equation", + "content": "\\epsilon_ {N} = O \\Big (\\sqrt {\\frac {\\log (1 / \\eta)}{N}} \\Big)", + "image_path": "59a00377f52e6d1c6a629e470eca2267cb2b97382f9266aba15d1facb8b339f2.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 508, + 446, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 446, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 446, + 520 + ], + "type": "text", + "content": "accounts for the sampling error." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 527, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 555, + 635 + ], + "type": "text", + "content": "Our empirical results have indicated that IIMM is strongly correlated with performance gains and catastrophic forgetting. The bound provided here offers a theoretical explanation: only significant reorganization of the embedding space (reflected by a large Wasserstein distance) can cause a substantial shift in IIMM. Therefore, a high IIMM change can be interpreted as evidence of a considerable shift in the model's internal representation, likely correlating with notable performance changes." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 646, + 395, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 395, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 395, + 659 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Datasets To evaluate the relationships between transferability measures and model performance changes following fine-tuning, we compiled a diverse collection of nine commonly used, publicly available computer vision benchmarks" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2399" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "provided by PyTorch [26, 39, 64] as our testing dataset, with the goal of minimizing domain overlap and covering a range of coarse and granular classification tasks. These datasets included Stanford Cars [21], CIFAR100 [22], the Describable Textures Dataset (DTD) [6], satellite imagery datasets EuroSAT [14] and RESISC45 [5], the German Traffic Sign Recognition Benchmark (GTSRB) [45], MNIST [23], the Scene Understanding Benchmark (SUN397) [52], and the Street View House Numbers (SVHN) dataset [33]. For each dataset, we generated validation splits by randomly sampling " + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": " of the respective training data, capping the validation split size at 5000 samples for larger datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 231, + 294, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 294, + 338 + ], + "type": "text", + "content": "Models and fine-tuning methods To investigate the applicability of IIMM to vision-language encoders, we conducted all experiments using two distinct groups of evaluations schema: one in which the base pre-trained model varies while the fine-tuning strategy is held constant, and another group in which the base pre-trained model is held constant while the fine-tuning strategy varies. We will refer to these evaluation schema as the base model evaluation and PEFT evaluation, respectively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 339, + 294, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 294, + 409 + ], + "type": "text", + "content": "For the base model evaluation, we perform full finetuning over all parameters for the following four pre-trained models: CLIP, EVA-02-CLIP [12], SigLIP [60], and CoCa [59], using a ViT-B/32 image backbone for CLIP and CoCa and a ViT-B/16 image backbone for SigLIP and EVA-02-CLIP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 411, + 294, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 294, + 506 + ], + "type": "text", + "content": "For the PEFT evaluation, we utilized a CLIP model with a ViT-B/32 image backbone as the base model and evaluate four common parameter-efficient fine-tuning (PEFT) methods: CLIP-Adapter [13], BitFit (bias tuning) [3], LoRA [16], and a method we call Attention-Weight Tuning, which simply refines the fine-tuning of all attention layer weights in Touvron et al. [47] to just the attention layers' inprojection weights." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 294, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 294, + 653 + ], + "type": "text", + "content": "Baselines While no methods currently exist for explicitly predicting gains in accuracy from fine-tuning for dual-encoder vision-language models, many related methods exist for stand-alone vision encoders. In particular, we compare against several recent or established measures self-described as transferability estimators. The six comparison measure are: TransRate [17], Self-challenging Fisher Discriminant Analysis (SFDA) [42], Neural Collapse informed Transferability Index (NCTI) [51], Efficient Multitask Model Selector EMMS [43], Gaussian Bhattacharyya Coefficient (GBC) [38], and H-score [2]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "content": "We chose these comparison measures to span the definitions and applications of self-described transferability measures. Those that estimate the log-likelihood are H-Score, SFDA, TransRate, and EMMS. NCTI functions as an estimator of the gap between the embedding geometry of the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "pre-trained model and the embedding geometry of a hypothetical fine-tuned model [51]. The GBC quantifies transferability as a measure of overlap in the feature space between embedding clusters of different class labels [38]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 133, + 553, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 253 + ], + "type": "text", + "content": "Evaluation measures We evaluate performance of the transferability measures through the strength of their relationship to three variables: accuracy on the target test data after fine-tuning on the target task training data; gain in performance from the pre-trained model to the fine-tuned model scaled by the pre-trained model error, which we refer to as gain over zero-shot error; and average accuracy across the remaining eight non-target datasets to measure loss in performance from the pre-trained model on nontarget tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 258, + 553, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 305 + ], + "type": "text", + "content": "Definition 1 (Gain over zero-shot error). For a given task, we define " + }, + { + "bbox": [ + 313, + 258, + 553, + 305 + ], + "type": "inline_equation", + "content": "A_{zs}" + }, + { + "bbox": [ + 313, + 258, + 553, + 305 + ], + "type": "text", + "content": " as the zero-shot accuracy and " + }, + { + "bbox": [ + 313, + 258, + 553, + 305 + ], + "type": "inline_equation", + "content": "A_{ft}" + }, + { + "bbox": [ + 313, + 258, + 553, + 305 + ], + "type": "text", + "content": " as the fine-tuned accuracy. We then define the gain over zero-shot error as," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 410, + 319, + 553, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 319, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 410, + 319, + 553, + 345 + ], + "type": "interline_equation", + "content": "\\frac {A _ {f t} - A _ {z s}}{1 - A _ {z s}} \\tag {5}", + "image_path": "9cfe8b5136075b38b8036a1ac937ad6c615df16a383e5047780f220dff7eab01.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 348, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 553, + 371 + ], + "type": "text", + "content": "This value normalizes for differences in the zero-shot accuracy across models and tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 376, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 472 + ], + "type": "text", + "content": "We first evaluate all measures' monotonic relationship to both fine-tuned accuracy and gain over zero-shot error using Kendall's " + }, + { + "bbox": [ + 313, + 376, + 553, + 472 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 376, + 553, + 472 + ], + "type": "text", + "content": ", which is a standard evaluation measure within the literature. We then look at the top performing measures of this rank correlation and fit a least-squares model to evaluate their predictive strength, additionally looking at a least-squares model fit of the top measure against the average accuracy on off-target tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "Training details To establish the ground truth for changes in model in-domain (ID) and out-of-domain (OOD) accuracy following fine-tuning on each dataset, we started by separately fine-tuning each of the base models on the nine datasets in our dataset suite. Following training on a given dataset, each model was tested on all nine datasets, with in-domain accuracy calculated as its accuracy on the test split of the training dataset and out-of-domain accuracies as the accuracies obtained on the remaining datasets test split. Each fine-tuning run consisted of 30 epochs of gradient descent using an SGD optimizer and a batch size of 128 and incorporated an exhaustive grid search over all relevant model hyper-parameters, examining values " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{1\\mathrm{e} - 2,1\\mathrm{e} - 3,1\\mathrm{e} - 4,1\\mathrm{e} - 5\\}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " for the learning rate, " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{1\\mathrm{e} - 3,1\\mathrm{e} - 4,1\\mathrm{e} - 5\\}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " for the weight decay, " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{2,4,8,16\\}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " for LoRA model rank, and " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{4,8,16,32\\}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " for CLIP-Adapter reduction. The initial image and text embeddings were then generated per dataset by a forward pass of each dataset's train split through the pre-trained base models and normalizing to unit vectors." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2400" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 140, + 70, + 470, + 186 + ], + "blocks": [ + { + "bbox": [ + 140, + 70, + 470, + 186 + ], + "lines": [ + { + "bbox": [ + 140, + 70, + 470, + 186 + ], + "spans": [ + { + "bbox": [ + 140, + 70, + 470, + 186 + ], + "type": "table", + "html": "
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueτp-valueτp-valueτp-valueτ
IIMM< 10-30.890.0010.830.0020.780.0020.78
GBC0.0020.780.0060.720.0020.780.0010.83
TransRate0.0130.670.0250.610.0130.670.0060.72
EMMS0.0250.610.3580.280.0250.610.1190.44
NCTI0.919-0.060.028-0.590.8180.070.827-0.06
SFDA0.2600.330.919-0.060.1800.390.3580.28
H-Score0.7610.110.002-0.780.0750.500.013-0.67
", + "image_path": "2923caa44ee53cf31a762f28adf4c5a680595e8747187182d4ca195ed0ddee0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 142, + 217, + 469, + 335 + ], + "blocks": [ + { + "bbox": [ + 127, + 195, + 481, + 207 + ], + "lines": [ + { + "bbox": [ + 127, + 195, + 481, + 207 + ], + "spans": [ + { + "bbox": [ + 127, + 195, + 481, + 207 + ], + "type": "text", + "content": "Table 1. Kendall's " + }, + { + "bbox": [ + 127, + 195, + 481, + 207 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 127, + 195, + 481, + 207 + ], + "type": "text", + "content": " correlation of score and gain over zero-shot error. Larger " + }, + { + "bbox": [ + 127, + 195, + 481, + 207 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 127, + 195, + 481, + 207 + ], + "type": "text", + "content": " values are better.*" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 142, + 217, + 469, + 335 + ], + "lines": [ + { + "bbox": [ + 142, + 217, + 469, + 335 + ], + "spans": [ + { + "bbox": [ + 142, + 217, + 469, + 335 + ], + "type": "table", + "html": "
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueτp-valueτp-valueτp-valueτ
IIMM0.0060.720.0060.720.0130.670.0250.61
GBC0.0250.610.0060.720.0130.670.0130.67
TransRate0.0750.500.0250.610.0450.560.0450.56
EMMS0.0130.670.0250.610.0750.500.0250.61
NCTI1.0000.000.173-0.370.818-0.070.8270.06
SFDA0.0750.500.6120.170.0750.500.2600.33
H-Score0.6120.170.045-0.560.0250.610.075-0.50
", + "image_path": "64ba5033f20bb312374e42b6322741c94dbf9de60e9b60df615a8ef49478beb5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 343, + 453, + 354 + ], + "lines": [ + { + "bbox": [ + 155, + 343, + 453, + 354 + ], + "spans": [ + { + "bbox": [ + 155, + 343, + 453, + 354 + ], + "type": "text", + "content": "Table 2. Kendall's " + }, + { + "bbox": [ + 155, + 343, + 453, + 354 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 155, + 343, + 453, + 354 + ], + "type": "text", + "content": " correlation of score and accuracy. Larger " + }, + { + "bbox": [ + 155, + 343, + 453, + 354 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 155, + 343, + 453, + 354 + ], + "type": "text", + "content": " values are better.*" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 375, + 112, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 112, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 112, + 387 + ], + "type": "text", + "content": "4.1. Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 397, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 295, + 445 + ], + "type": "text", + "content": "We first look at Kendall's " + }, + { + "bbox": [ + 55, + 397, + 295, + 445 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 397, + 295, + 445 + ], + "type": "text", + "content": " correlation to determine the strength of the monotonic relationship of our proposed and comparison measures to both the fine-tuned gain over zero-shot error and fine-tuned accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": "The IIMM and the GBC have significantly stronger monotonic relationships than the other measures for both gain over zero-shot error, as shown in Tab. 1, and accuracy, as shown in Tab. 2. In both tables, the IIMM and the GBC are the most strongly correlated measures and display the greatest significance. Upon further investigation, the IIMM displays a mostly linear relationship with gain over zero-shot error, while the GBC displays a mostly exponential relationship with gain over zero-shot error. This allows us to fit and evaluate a predictive model using least-squares regression of the IIMM with gain over zero-shot error, and the GBC with the log of gain over zero-shot error. The results of these evaluations can be seen in Tab. 3, which shows the significance of the slope coefficient and the " + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " value, an estimate of the amount of variation within the data explained by its linear relationship to the measure. The IIMM has a stronger fit across all base pre-trained models compared to the GBC, and indeed shows itself to have a strong linear relationship to gain over zero-shot error with " + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "inline_equation", + "content": "R^2 > 0.89" + }, + { + "bbox": [ + 55, + 450, + 295, + 713 + ], + "type": "text", + "content": " for all base models. Without the ability to capture cross-modal alignment, the compared uni-modal metrics expectedly demonstrate worse performance when com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 376, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 399 + ], + "type": "text", + "content": "pared to the IIMM, underscoring the value of an integrated vision-language metric." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 400, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 400, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 312, + 400, + 555, + 662 + ], + "type": "text", + "content": "This positive linear relationship between the IIMM and in-domain gain over zero-shot error is clearly seen in the top plots of Fig. 2. The bottom plots of Fig. 2 show the negative linear relationship of the IIMM with the average change in accuracy across a suite of eight out-of-domain tasks. For all base pre-trained models aside from CLIP, the IIMM also shows a strong linear relationship to forgetting, as evidenced by the significance of the slope coefficient and high " + }, + { + "bbox": [ + 312, + 400, + 555, + 662 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 312, + 400, + 555, + 662 + ], + "type": "text", + "content": " values. These relationships hold true across parameter-efficient fine-tuning methods, as shown in Fig. 3. We see with CLIP-Adapter that its overall worse in-domain performance attenuates the strength of the relationship between IIMM and gain over zero-shot error, and predictably, that a lower in-domain performance gain resulted in significantly less catastrophic forgetting. CLIP-Adapter's design — injecting task kernels atop a frozen backbone — results in smaller, localized representational shifts, resulting in weaker in-distribution performance gains while preventing greater out-of-distribution performance loss. Since the IIMM is correlated with near-optimal performance gains, fine-tuning methods that result in sub-optimal gains may yield lower IIMM correlations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 662, + 553, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 553, + 687 + ], + "type": "text", + "content": "These results provide evidence that the IIMM is strongly correlated with fine-tuning gains over zero-shot error. Un" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "*P-values reflect the likelihood of observing results by random chance; smaller values suggest stronger associations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2401" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 144, + 70, + 466, + 133 + ], + "blocks": [ + { + "bbox": [ + 144, + 70, + 466, + 133 + ], + "lines": [ + { + "bbox": [ + 144, + 70, + 466, + 133 + ], + "spans": [ + { + "bbox": [ + 144, + 70, + 466, + 133 + ], + "type": "table", + "html": "
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueR²p-valueR²p-valueR²p-valueR²
IIMM< 10-30.92< 10-30.89< 10-30.92< 10-30.92
GBC0.0040.720.0060.680.0150.590.0050.71
", + "image_path": "f093eeb5dde358e31a2e56dd671ebc2af45ff05d95987ba287aba677b32cb306.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 140, + 555, + 163 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 555, + 163 + ], + "type": "text", + "content": "Table 3. Evaluation of linear fit of the IIMM to gain over zero-shot error and the GBC to the log of gain over zero-shot error. The significance of the slope coefficient is given by the p-value, and the strength of the fit is evaluated with the " + }, + { + "bbox": [ + 55, + 140, + 555, + 163 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 55, + 140, + 555, + 163 + ], + "type": "text", + "content": " value.*" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 75, + 175, + 534, + 411 + ], + "blocks": [ + { + "bbox": [ + 75, + 175, + 534, + 411 + ], + "lines": [ + { + "bbox": [ + 75, + 175, + 534, + 411 + ], + "spans": [ + { + "bbox": [ + 75, + 175, + 534, + 411 + ], + "type": "image", + "image_path": "217baa5082a3ae5e039d150e69a16730c276d98da8b4e205abc790382876a615.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 421, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 421, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 555, + 456 + ], + "type": "text", + "content": "Figure 2. Performance gains and losses versus the IIMM per pre-trained model, with linear regression fit and " + }, + { + "bbox": [ + 55, + 421, + 555, + 456 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 55, + 421, + 555, + 456 + ], + "type": "text", + "content": " confidence intervals. The top plots show the relationship between the on-target (ID) task IIMM to the on-target task performance gains over zero-shot error, while the bottom plots show the relationship between on-target task IIMM and the average performance loss on off-target (OOD) tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": "like GBC, the second-highest correlated measure, the IIMM is bounded in " + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 55, + 475, + 295, + 510 + ], + "type": "text", + "content": " and extreme values are informative in isolation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 525, + 220, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 220, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 220, + 538 + ], + "type": "text", + "content": "4.2. Exploring embedding features" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 544, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 544, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 54, + 544, + 295, + 686 + ], + "type": "text", + "content": "To empirically validate the provided theoretical justification for leveraging multi-modal information to improve transferability estimation in vision-language models, we independently evaluated all relevant components of CLIP's embedding geometry, known to manifest as a hypersphere. Using (3), we separately computed the average pairwise similarities for the image and text modalities and fit each to the accuracy gain over zero-shot error. We then applied (2) to compute the degree of incorrect label alignment across the hypersphere and also fit the measure to the accuracy gain over zero-shot error. The final results of these evaluations can be seen in Tab. 4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "It is observed that the features of the embedding structure most predictive of accuracy gains are the intra-images dis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 475, + 555, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 620 + ], + "type": "text", + "content": "tance and the incorrect label alignment. Based on these results, the correct label alignment and intra-texts distance are generally uninformative of accuracy gains with fine-tuning. As described in Sec. 3, each of these terms corresponds to a feature of the shared embedding space related to learning and forgetting. We include additional results of the Pearson correlations between different convex combinations of our inter- and intra- terms and gain over zero-shot error in the supplementary material. This empirical evidence successfully verifies that the terms combined have a stronger relationship with learning and forgetting than they do independently." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 632, + 388, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 388, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 388, + 645 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "In this work, we propose the Inter-Intra Modal Measure (IIMM) as a novel metric for predicting fine-tuning performance in vision-language models. We demonstrate that the IIMM strongly correlates with both learning gains and catastrophic forgetting, offering a simple yet powerful met" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2402" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 70, + 534, + 305 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 534, + 305 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 534, + 305 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 534, + 305 + ], + "type": "image", + "image_path": "eeff330e2b2dafa96fe48e0df5e5d8b56bba36b7bca4780eb94cf566c76799ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 316, + 555, + 351 + ], + "lines": [ + { + "bbox": [ + 54, + 316, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 555, + 351 + ], + "type": "text", + "content": "Figure 3. Performance gains and losses versus the IIMM per PEFT method, with linear regression fit and " + }, + { + "bbox": [ + 54, + 316, + 555, + 351 + ], + "type": "inline_equation", + "content": "96\\%" + }, + { + "bbox": [ + 54, + 316, + 555, + 351 + ], + "type": "text", + "content": " confidence intervals. The top plots show the relationship between the on-target task (ID) IIMM and on-target task performance gains over zero-shot error, while the bottom plots show the relationship between on-target task IIMM and the average performance loss on off-target (OOD) tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 118, + 360, + 493, + 434 + ], + "blocks": [ + { + "bbox": [ + 118, + 360, + 493, + 434 + ], + "lines": [ + { + "bbox": [ + 118, + 360, + 493, + 434 + ], + "spans": [ + { + "bbox": [ + 118, + 360, + 493, + 434 + ], + "type": "table", + "html": "
ModelCoCaEVA-02CLIPSigLIP
Measurep-valueR²p-valueR²p-valueR²p-valueR²
Incorrect Alignment0.0010.820.0040.720.0120.620.0010.80
Intra-Images Distance< 10-30.88< 10-30.860.0010.81< 10-30.88
Intra-Texts Distance0.0840.370.0340.500.0190.570.2650.17
", + "image_path": "2ea7ed7fbfc58db062c7eb009f92dd74bf02e3a43be66110676af96d6bd23dad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 442, + 555, + 465 + ], + "lines": [ + { + "bbox": [ + 54, + 442, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 54, + 442, + 555, + 465 + ], + "type": "text", + "content": "Table 4. Linear fit of measure to gain over zero-shot error with full fine-tuning by pre-trained model. The significance of the slope coefficient is given by the p-value, and the " + }, + { + "bbox": [ + 54, + 442, + 555, + 465 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 54, + 442, + 555, + 465 + ], + "type": "text", + "content": " values represent the proportion of variation within the data modeled by the linear relationship." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 485, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 485, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 297, + 713 + ], + "type": "text", + "content": "ric for estimating fine-tuning outcomes. Because the IIMM is bounded, an extreme value in isolation can inform practitioners whether a model will have limited or significant gains from fine-tuning. In addition, our results show that the IIMM can be leveraged to fit predictive linear models for estimating fine-tuning accuracy gains across diverse tasks with minimal computational overhead. We demonstrate that our proposed measure has a stronger relationship to both fine-tuned accuracy and accuracy changes with fine-tuning than baselines. We further show the IIMM is informative of performance changes across parameter efficient fine-tuning methods, highlighting some methods are particularly susceptible to catastrophic forgetting when the zero-shot IIMM is high. Beyond empirical validation, we provide a theoretical bound linking changes in IIMM to the Wasserstein distance between pre- and post-fine-tuning embedding distributions, ensuring stability in IIMM predictions with respect to changes in representation space. Given the IIMM measures changes in embedding geometry prior to any task" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 485, + 556, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 556, + 570 + ], + "type": "text", + "content": "specific heads, the metric is also task-agnostic, enabling its application to a variety of downstream tasks. We encourage future research to validate such use in broader applications. Overall, our work provides a practical and theoretically grounded tool for optimizing fine-tuning decisions in vision-language models, paving the way for more efficient and informed adaptation of foundation models to new tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "Limitations It is possible that the IIMM could show more variation when compared across a more diverse set of base models. While extreme values of the IIMM are heuristically informative in isolation, fitting a model for transfer accuracy estimation requires first fine-tuning across a small but diverse set of benchmarks. This could be an avenue of future research, as there might be additional information within the embedding geometry that allows for fine-tuning performance prediction without requiring information from fine-tuning on other tasks." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2403" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 139 + ], + "type": "text", + "content": "The authors acknowledge the MIT Lincoln Laboratory Supercomputing Center for providing the high-performance computing resources that have contributed to the research results reported in this paper." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 139, + 295, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 139, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 295, + 247 + ], + "type": "text", + "content": "DISTRIBUTION STATEMENT A. Approved for public release. Distribution is unlimited. This material is based upon work supported by the Under Secretary of Defense for Research and Engineering under Air Force Contract No. FA8702-15-D-0001 or FA8702-25-D-B002. Any opinions, findings, conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Under Secretary of Defense for Research and Engineering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 258, + 114, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 114, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 114, + 270 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 277, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 277, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 277, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 61, + 277, + 295, + 331 + ], + "type": "text", + "content": "[1] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational Models Defining a New Era in Vision: A Survey and Outlook, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 333, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 333, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 61, + 333, + 295, + 399 + ], + "type": "text", + "content": "[2] Yajie Bao, Yang Li, Shao-Lun Huang, Lin Zhang, Lizhong Zheng, Amir Zamir, and Leonidas Guibas. An Information-Theoretic Approach to Transferability in Task Transfer Learning. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2309-2313, Taipei, Taiwan, 2019. IEEE. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 400, + 296, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 400, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 61, + 400, + 296, + 466 + ], + "type": "text", + "content": "[3] Elad Ben Zaken, Yoav Goldberg, and Shauli Ravfogel. BitFit: Simple Parameter-efficient Fine-tuning for Transformer-based Masked Language-models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 1-9, Dublin, Ireland, 2022. Association for Computational Linguistics. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 467, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 295, + 521 + ], + "type": "text", + "content": "[4] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 523, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 523, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 523, + 295, + 555 + ], + "type": "text", + "content": "[5] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 557, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 294, + 610 + ], + "type": "text", + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing Textures in the Wild. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 3606-3613, Columbus, OH, USA, 2014. IEEE. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 613, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 295, + 645 + ], + "type": "text", + "content": "[7] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large Scale Fine-Grained Categorization and Domain-Specific Transfer Learning, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 295, + 679 + ], + "type": "text", + "content": "[8] Yuxuan Ding, Lingqiao Liu, Chunna Tian, Jingyuan Yang, and Haoxuan Ding. Don't Stop Learning: Towards Continual Learning for the CLIP Model, 2022. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 294, + 713 + ], + "type": "text", + "content": "[9] Yuhe Ding, Bo Jiang, Aijing Yu, Aihua Zheng, and Jian Liang. Which model to transfer? a survey on transferability estimation. arXiv preprint arXiv:2402.15231, 2024. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[10] Kshitij Dwivedi and Gemma Roig. Representation Similarity Analysis for Efficient Task Taxonomy & Transfer Learning. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12379-12388, Long Beach, CA, USA, 2019. IEEE. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[11] Zhihao Fan, Zhongyu Wei, Jingjing Chen, Siyuan Wang, Ze-jun Li, Jiarong Xu, and Xuanjing Huang. A unified continuous learning framework for multi-modal knowledge discovery and pre-training, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "text", + "content": "[12] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. arXiv preprint arXiv:2303.11331, 2023.5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "type": "text", + "content": "[13] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. CLIP-Adapter: Better Vision-Language Models with Feature Adapters, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 315 + ], + "type": "text", + "content": "[14] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Introducing Eurosat: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. In IGARSS 2018 - 2018 IEEE International Geoscience and Remote Sensing Symposium, pages 204-207, Valencia, 2018. IEEE. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "text", + "content": "[15] Dapeng Hu, Shipeng Yan, Qizhengqiu Lu, Lanqing Hong, Hailin Hu, Yifan Zhang, Zhenguo Li, Xinchao Wang, and Jiashi Feng. How well does self-supervised pre-training perform with streaming data?, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "text", + "content": "[16] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In ICLR. arXiv, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 405, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 448 + ], + "type": "text", + "content": "[17] Long-Kai Huang, Junzhou Huang, Yu Rong, Qiang Yang, and Ying Wei. Frustratingly easy transferability estimation. In International Conference on Machine Learning, pages 9201-9225. PMLR, 2022. 1, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 449, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 514 + ], + "type": "text", + "content": "[18] Joana Palés Huix, Adithya Raju Ganeshan, Johan Fredin Haslum, Magnus Söderberg, Christos Matsoukas, and Kevin Smith. Are Natural Domain Foundation Models Useful for Medical Image Classification? In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 7619-7628. IEEE, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 515, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 581 + ], + "type": "text", + "content": "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 582, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 658 + ], + "type": "text", + "content": "[20] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[21] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3D Object Representations for Fine-Grained Categorization. In 2013 IEEE International Conference on Computer Vision Workshops, pages 554–561, Sydney, Australia, 2013. IEEE. 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2404" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[22] Alex Krizhevsky. Learning Multiple Layers of Features from Tiny Images, 2009. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 128 + ], + "type": "text", + "content": "[23] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, Nov./1998. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[24] Hao Li, Charless Fowlkes, Hao Yang, Onkar Dabeer, Zhuowen Tu, and Stefano Soatto. Guided Recommendation for Model Fine-Tuning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3633-3642, Vancouver, BC, Canada, 2023. IEEE. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 249 + ], + "type": "text", + "content": "[25] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before Fuse: Vision and Language Representation Learning with Momentum Distillation. In Advances in Neural Information Processing Systems, pages 9694-9705. Curran Associates, Inc., 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "text", + "content": "[26] Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Zou. Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learning. In First Workshop of Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022, 2022. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "text", + "content": "[27] C. Liao, T. Tsiligkaridis, and B. Kulis. Descriptor and word soups: Overcoming the parameter efficiency accuracy trade-off for out-of-distribution few-shot learning. In CVPR, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 294, + 381 + ], + "type": "text", + "content": "[28] C. Liao, C. So, T. Tsiligkaridis, and B. Kulis. Multimodal unsupervised domain generalization by retrieving across the modality gap. In ICLR, 2025. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 436 + ], + "type": "text", + "content": "[29] Yong Lin, Lu Tan, Hangyu Lin, Zeming Zheng, Renjie Pi, Jipeng Zhang, Shizhe Diao, Haoxiang Wang, Han Zhao, Yuan Yao, and Tong Zhang. Speciality vs Generality: An Empirical Study on Catastrophic Forgetting in Fine-tuning Foundation Models, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 480 + ], + "type": "text", + "content": "[30] Michael McCloskey and Neal J. Cohen. Catastrophic Interference in Connectionist Networks: The Sequential Learning Problem. In *Psychology of Learning and Motivation*, pages 109–165. Elsevier, 1989. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 503 + ], + "type": "text", + "content": "[31] S. Menon and C. Vondrick. Visual classification via description from large language models. In ICLR, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 547 + ], + "type": "text", + "content": "[32] Jishnu Mukhoti, Yarin Gal, Philip HS Torr, and Puneet K Dokania. Fine-tuning can cripple your foundation model; preserving features may be the solution. arXiv preprint arXiv:2308.13320, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[33] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading Digits in Natural Images with Unsupervised Feature Learning. NeurIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "type": "text", + "content": "[34] Cuong Nguyen, Tal Hassner, Matthias Seeger, and Cedric Archambeau. LEEP: A New Measure to Evaluate Transferability of Learned Representations. In Proceedings of the 37th International Conference on Machine Learning, pages 7294-7305. PMLR, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 680 + ], + "type": "text", + "content": "[35] Zixuan Ni, Siliang Tang, and Yueting Zhuang. Self-supervised class incremental learning, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[36] Zixuan Ni, Longhui Wei, Siliang Tang, Yueting Zhuang, and Qi Tian. Continual Vision-Language Representation Learning with Off-Diagonal Information, 2023. 1, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[37] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[38] Michal Páncy, Andrea Agostinelli, Jasper Uijlings, Vittorio Ferrari, and Thomas Mensink. Transferability Estimation using Bhattacharyya Class Separability, 2022. 1, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 217 + ], + "type": "text", + "content": "[39] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 219, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 297 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "text", + "content": "[41] K. Roth, J. M. Kim, A. S. Koepke, O. Vinyals, C. Schmid, and Z. Akata. Waffling around for performance: Visual classification with random words and broad concepts. In ICCV, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 398 + ], + "type": "text", + "content": "[42] Wenqi Shao, Xun Zhao, Yixiao Ge, Zhaoyang Zhang, Lei Yang, Xiaogang Wang, Ying Shan, and Ping Luo. Not all models are equal: Predicting model transferability in a self-challenging fisher space. In European Conference on Computer Vision, pages 286-302. Springer, 2022. 1, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "text", + "content": "[43] Wenqi Shao, Chonghe Jiang, Kaipeng Zhang, Yu Qiao, Ping Luo, et al. Foundation model is efficient multimodal multitask model selector. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 443, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 553, + 508 + ], + "type": "text", + "content": "[44] Tejas Srinivasan, Ting-Yun Chang, Leticia Pinto Alva, Georgios Chochlakis, Mohammad Rostami, and Jesse Thomason. CLiMB: A continual learning benchmark for vision-and-language tasks. In Advances in Neural Information Processing Systems, pages 29440-29453. Curran Associates, Inc., 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 509, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 553 + ], + "type": "text", + "content": "[45] J. Stallkamp, M. Schlipsing, J. Salmen, and C. Igel. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. *Neural Networks*, 32:323–332, 2012. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 555, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 589 + ], + "type": "text", + "content": "[46] P. Teterwak, K. Saito, T. Tsiligkaridis, B. A. Plummer, and K. Saenko. Is large-scale pretraining the secret to good domain generalization? In ICLR, 2025. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 622 + ], + "type": "text", + "content": "[47] Hugo Touvron, Matthieu Cord, Alaaeldin El-Nouby, Jakob Verbeek, and Hervé Jégou. Three things everyone should know about Vision Transformers, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[48] Anh Tran, Cuong Nguyen, and Tal Hassner. Transferability and Hardness of Supervised Classification Tasks. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1395-1405, Seoul, Korea (South), 2019. IEEE. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "text", + "content": "[49] Kevin Vogt-Lowell, Noah Lee, Theodoros Tsiligkaridis, and Marc Vaillant. Robust fine-tuning of vision-language models for domain generalization. In 2023 IEEE High Performance" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2405" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Extreme Computing Conference (HPEC), pages 1-7. IEEE, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[50] Tongzhou Wang and Phillip Isola. Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere. In Proceedings of the 37th International Conference on Machine Learning, pages 9929-9939. PMLR, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 217 + ], + "type": "text", + "content": "[51] Zijian Wang, Yadan Luo, Liang Zheng, Zi Huang, and Mahsa Baktashmotlagh. How far pre-trained models are from neural collapse on the target dataset informs their transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5549-5558, 2023. 1, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 284 + ], + "type": "text", + "content": "[52] Jianxiong Xiao, James Hays, Krista A. Ehinger, Aude Oliva, and Antonio Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 3485-3492, San Francisco, CA, USA, 2010. IEEE. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "text", + "content": "[53] Yiqun Xie, Zhihao Wang, Weiye Chen, Zhili Li, Xiaowei Jia, Yanhua Li, Ruichen Wang, Kangyang Chai, Ruohan Li, and Sergii Skakun. When are foundation models effective? understanding the suitability for pixel-level classification using multispectral imagery. CoRR, abs/2404.11797, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 430 + ], + "type": "text", + "content": "[54] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive Pretraining for Zero-shot Video-Text Understanding. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 6787–6800, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 432, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 295, + 464 + ], + "type": "text", + "content": "[55] Chao Yi, De-Chuan Zhan, and Han-Jia Ye. Bridge the Modality and Capacity Gaps in Vision-Language Model Selection, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 466, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 295, + 510 + ], + "type": "text", + "content": "[56] Kaichao You, Yong Liu, Jianmin Wang, and Mingsheng Long. Logme: Practical assessment of pre-trained models for transfer learning. In International Conference on Machine Learning, pages 12133-12143. PMLR, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 511, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 295, + 543 + ], + "type": "text", + "content": "[57] E. Yu, C. Liao, S. Ravi, T. Tsiligkaridis, and B. Kulis. Image-capture encoding for improving zero-shot generalization. In WACV, 2025. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 600 + ], + "type": "text", + "content": "[58] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23219-23230. IEEE. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 644 + ], + "type": "text", + "content": "[59] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 690 + ], + "type": "text", + "content": "[60] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[61] Dan Zhang, Tao Feng, Lilong Xue, Yuandong Wang, Yuxiao Dong, and Jie Tang. Parameter-efficient fine-tuning for" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 252 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "foundation models. arXiv preprint arXiv:2501.13787, 2025. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 151 + ], + "type": "text", + "content": "[62] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and Curtis P. Langlotz. Contrastive Learning of Medical Visual Representations from Paired Images and Text. In Proceedings of the 7th Machine Learning for Healthcare Conference, pages 2-25. PMLR, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "text", + "content": "[63] Zangwei Zheng, Mingyuan Ma, Kai Wang, Ziheng Qin, Xi-angyu Yue, and Yang You. Preventing Zero-Shot Transfer Degradation in Continual Learning of Vision-Language Models. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 19068-19079. IEEE. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[64] Orr Zohar, Kuan-Chieh Wang, Shih-Cheng Huang, and Serena Yeung. LOVM: Language-Only Vision Model Selection. In Conference on Neural Information Processing Systems (Neurips), 2023. 3, 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2406" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/2b8a94f1-28e8-4218-82c7-da4dba2c45e1_content_list.json b/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/2b8a94f1-28e8-4218-82c7-da4dba2c45e1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..515bf7eff55ecbbc0efc38b2251f85183043b39c --- /dev/null +++ b/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/2b8a94f1-28e8-4218-82c7-da4dba2c45e1_content_list.json @@ -0,0 +1,1566 @@ +[ + { + "type": "text", + "text": "The Scalability of Simplicity: Empirical Analysis of Vision-Language Learning with a Single Transformer", + "text_level": 1, + "bbox": [ + 117, + 130, + 880, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weixian Lei* Jiacong Wang* Haochen Wang* \nXiangtai Li Jun Hao Liew Jiashi Feng Zilong Huang† \n*Equal contribution, † Corresponding author \nBytedance Seed", + "bbox": [ + 264, + 202, + 730, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces SAIL, a single transformer unified multimodal large language model (MLLM) that integrates raw pixel encoding and language decoding within a singular architecture. Unlike existing modular MLLMs, which rely on a pre-trained vision transformer (ViT), SAIL eliminates the need for a separate vision encoder, presenting a more minimalist architecture design. Instead of introducing novel architectural components, SAIL adapts mix-attention mechanisms and multimodal rotary position embedding to better align with the distinct characteristics of visual and textual modalities. We systematically compare SAIL's properties—including scalability, cross-modal information flow patterns, and visual representation capabilities—with those of modular MLLMs. By scaling both training data and model size, SAIL achieves performance comparable to modular MLLMs. Notably, the removal of pretrained ViT components enhances SAIL's scalability and results in significantly different cross-modal information flow patterns. Moreover, SAIL demonstrates strong visual representation capabilities, achieving results on par with ViT22B in vision tasks such as semantic segmentation. Code and models are available at https://github.com/bytedance/SAIL.", + "bbox": [ + 88, + 340, + 483, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 722, + 220, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The pursuit of multimodal intelligence has driven the development of Multimodal Large Language Models (MLLMs) [48, 58, 65], which typically adopt a modular design: a pre-trained vision encoder (e.g., CLIPViT [15, 59]) extracts image features, a Large Language Model (LLM) [2, 16, 37, 55-57] processes text, and a lightweight projector aligns the two modalities. This framework achieves strong performance through multi-stage pretraining, supervised fine-tuning (SFT), and post-training on multimodal datasets [3, 18, 48, 70, 71, 84]. While effective,", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b3550ba22a6bb0225a56ba63ea2b9e2fb22e444614f6b4e1100358509dc9c52.jpg", + "image_caption": [ + "(A)", + "Figure 1. (A) Data scaling curve for Modular Multimodal Large Language Model (MLLM) and SAIL, our Single Transformer-based MLLM. As pretraining data increases, the single transformer SAIL shows a sharper performance gain, demonstrating its superior data scalability. Experiment settings are described in Section 4.3. (B) Comparison to existing Single Transformer-based MLLMs: our SAIL pushes the performance boundaries on both vision tasks and vision-language tasks." + ], + "image_footnote": [], + "bbox": [ + 514, + 316, + 714, + 458 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4aaec2709c4af572be5932a3c719e9c28bd183f82a5bf5c8f87e81fb1ce455b9.jpg", + "image_caption": [ + "(B)" + ], + "image_footnote": [], + "bbox": [ + 715, + 316, + 903, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "this modular MLLM paradigm inherently fragments multimodal processing, reinforces reliance on pretrained visual encoders, which may limit deployment flexibility and scalability [11, 21, 53].", + "bbox": [ + 511, + 608, + 905, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A promising alternative is to eliminate the visual encoder entirely and process raw image patches and text tokens within a single Transformer. This unified architecture removes modality-specific modules, enabling parameter sharing and end-to-end learning of vision-language interactions. Previous works [11, 21, 53] have primarily explored the architecture design, training data, and methods of Single Transformer-based MLLMs. However, little exploration has been given to their fundamental properties, such as scalability, cross-modal information flow patterns, and visual representation capabilities. A deeper understanding of these properties is crucial for unlocking the full potential of Single Transformer-based MLLMs.", + "bbox": [ + 509, + 671, + 906, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we present an experimental analysis of the fundamental properties of Single Transformer-based", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20758", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MLLMs and compare them to modular MLLMs (e.g. LLaVA [48]). Additionally, in the absence of a pre-trained visual encoder, Single Transformers have to learn visual representations from scratch. Thus, an intriguing question arises: can a trained Single Transformer emerge as a strong vision encoder?", + "bbox": [ + 89, + 90, + 480, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct a series of experiments to train and study our Single trAnformer model for vIsion and Language (SAIL). While we do not propose novel architecture designs, we introduce necessary modifications to enable the model to process different modalities in a unified architecture. In its micro architecture design, we address the different spatial characteristics of 2D images and 1D text data by employing a mixed attention mechanism: bidirectional attention for image patches and causal attention for text tokens, combined with multimodal rotary position embedding. Through model and data scaling, SAIL achieves performance on vision-language benchmarks comparable to modular MLLMs, while also functioning as a high-performing vision backbone, as shown in Figure 1.", + "bbox": [ + 89, + 183, + 482, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More concretely, our empirical analysis uncovers following striking advantages of Single Transformer architectures:", + "bbox": [ + 89, + 396, + 482, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(i) Superior Data Scaling: In controlled experiments, SAIL exhibits steeper performance gains as pretraining data scales. While LLaVA-style modular MLLMs initially perform well, our model's performance becomes very close to theirs when pretrained on 512M samples, as shown in Figure 1(A). This suggests that unified architectures can effectively leverage large-scale data and potentially match the performance of modular MLLMs.", + "bbox": [ + 89, + 428, + 482, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(ii) Vision Centric Information Flow Pattern: Through analysis of attention distributions, we observe that Single Transformers assign significantly higher attention scores to image tokens during token prediction compared to modular MLLMs. This indicates that the information flow in Single Transformer MLLMs is more direct, with visual tokens influencing prediction tokens more prominently, highlighting a vision-centric approach to decision-making.", + "bbox": [ + 89, + 550, + 482, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(iii) Vision Encoder Functioning: Our experiments further demonstrate that the pretrained Single Transformer inherently serves as a powerful vision encoder. Comprehensive evaluations on vision-centric tasks, such as image classification and semantic segmentation, show that the model learns rich visual representations during multimodal pretraining. These representations enhance its capacity for both semantic-level comprehension (e.g., object categorization) and pixel-level understanding (e.g., fine-grained segmentation masks), bridging high-level abstraction and low-level visual reasoning within a unified architecture.", + "bbox": [ + 89, + 672, + 482, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our findings indicate that Single Transformer-based MLLMs hold great promise in surpassing modular MLLMs in terms of leveraging large-scale data, forming direct vision-centric information pathways,", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and functioning as effective vision encoders. We hope our empirical findings inspire further research to refine and enhance Single Transformer architecture, ultimately driving advancements in multimodal intelligence from a new perspective.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 184, + 653, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Paradigms in Vision-Language Model Design", + "text_level": 1, + "bbox": [ + 511, + 210, + 895, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Modular MLLMs with Visual Encoders. The prevailing approach in MLLM design employs modular architectures [3, 48, 58, 65, 71] that rely on pretrained vision encoders (e.g., CLIP-ViT [15, 59], InternViT [14]) to process visual inputs. The visual features extracted from these frozen encoders are then aligned with LLM input spaces via linear [7, 48, 69, 84] or cross-attention layers [1, 32]. While this module design enables effective transfer of pretrained visual-language knowledge, it also introduces several limitations. First, incorporating a separate ViT encoder significantly slows down both training and inference, increasing deployment complexity and requiring costly infrastructure—especially when compared to a single transformer unified model. Second, common strategies for integrating visual features, such as direct mapping into LLM inputs [7, 48, 84] or sharing them across LLM layers [1, 32], often struggle to reconcile the inherent differences between images and text representations. Finally, as model scale, balancing the interactions between the encoder, LLM, and alignment layers becomes increasingly challenging [11, 53]. Thus, in this work, we explore a single transformer-based MLLM architecture that eliminates the ViT encoder and alignment components to overcome these challenges.", + "bbox": [ + 511, + 234, + 903, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single Transformer-based MLLMs Without Visual Encoders. Emerging research explores end-to-end architectures that process raw image patches and text tokens through a single Transformer, bypassing visual encoders entirely. These monolithic designs fall into two categories: continuous tokenization and discrete tokenization. Continuous tokenization, exemplified by Fuyu-8B [5] and SOLO [11], directly maps patches to LLM embeddings via linear projections, enabling flexible resolution handling but requiring massive pretraining data. Discrete tokenization, adopted by Chameleon [64] and Emu3 [73], employs VQ-VAE tokenizers to compress images into discrete tokens, trading pixel-level fidelity for generation capabilities. While later efforts such as EVE [21] and MonoInternVL [53] demonstrate the feasibility of encoder-free training, critical gaps remain: (1) Existing methods rely on extra designs and auxiliary loss [21], complicating training pipelines; (2) The scaling laws and fundamental properties of purely end-to-end trained models remain poorly understood; (3) Vision-language interaction in shared parame", + "bbox": [ + 511, + 598, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20759", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ter spaces lacks systematic analysis—most prior MLLMs default to causal attention for processing image-text sequences. In this work, we reveal that enabling bidirectional attention between image patches significantly enhances visual representation learning, addressing a key limitation in current designs. More importantly, our study bridges these gaps by establishing foundational principles for training scalable, self-contained single-transformer MLLMs.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Vision Representation Learning", + "text_level": 1, + "bbox": [ + 89, + 223, + 372, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning effective vision representations is a core challenge in computer vision research, with extensive works [4, 19, 20, 25, 30, 74] dedicated to this problem. With the proliferation of large-scale web-sourced image-text datasets [6, 26, 60], recent methods leverage this data to train deep visual representations via three primary paradigms:", + "bbox": [ + 89, + 244, + 482, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text as Classification Labels. Early methods used textual descriptions as weak supervision by extracting categorical labels from captions. For example, frameworks like Tag2Text [33] and RAM [82] used ViTs [23] to predict noun-based pseudo-labels from datasets like CC12M [6]. CatLIP [54] scaled labels to millions using object-centric supervision, and SuperClass [34] directly used tokenized text tokens as classification categories.", + "bbox": [ + 89, + 335, + 482, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-Text Contrastive Learning Contrastive pretraining, as exemplified by CLIP [15, 59] and ALIGN [36], aligns global image-text embeddings within a shared latent space. Subsequent works [8, 14, 41, 44, 72, 81] focused on enhancing performance and improving training efficiency.", + "bbox": [ + 89, + 458, + 482, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text as Autoregressive Targets. Caption generation as a pretext task is another approach for visual representation learning. SimVLM [74] trains encoder-decoder architectures to autoregressively predict captions, while CapPa [67] trains vision encoders through sequence prediction. These methods often retain modular designs or auxiliary components like contrastive losses [77]. Our work aligns with this category but removes architectural fragmentation by jointly modeling image patches and text tokens in a single Transformer. We find that the pre-trained Single Transformer learns transferable vision representations, enabling it to handle downstream multimodal understanding tasks and function as a vision encoder without modifications.", + "bbox": [ + 89, + 534, + 482, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SAIL: Training a Single Transformer for Vision and Language", + "text_level": 1, + "bbox": [ + 89, + 744, + 482, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model Architecture", + "text_level": 1, + "bbox": [ + 89, + 787, + 277, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SAIL is built upon a unified Transformer architecture (Figure 2(A)) that processes multimodal inputs through streamlined, modality-specific preprocessing. For text, raw input is tokenized using the language model's tokenizer and then transformed into embeddings via the textual embedding module. For images, we partition the input into fixed-", + "bbox": [ + 89, + 809, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "size patches and project them into continuous embeddings via a linear projection. Additionally, we maintain a list of special tokens explicitly designed for visual modality encoding: and tokens mark the beginning and end of an image patch span, respectively. In multimodal scenarios, such as image-text pairs, these embeddings are concatenated into a single sequence and fed into the Transformer, enabling joint cross-modal interactions through unified self-attention layers. This design eliminates the need for modality-specific encoders, which efficiently processess heterogeneous data within a Single Transformer framework.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bidirectional attention within image patches. While existing multimodal large language models (MLLMs) [40, 47, 48, 84] predominantly adopt causal attention for autoregressive sequence modeling, our experiments reveal that enabling full bidirectional attention among tokens from the same image significantly enhances visual representation learning and boosts downstream vision-language task performance. Note that previous Single Transformers [11, 21, 53, 73] have only utilized causal attention, without exploring the potential of mixed attention mechanisms.", + "bbox": [ + 511, + 273, + 903, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 2(B), for SAIL we implement a mixed attention scheme: (1) For text tokens, we preserve causal attention to maintain autoregressive generation capabilities, allowing each token to attend only to its predecessors. (2) For image tokens, we activate full bidirectional attention within each image patch group, empowering every visual token to interact with all others in the same image. This design captures holistic spatial relationships and contextual dependencies among visual elements, addressing the under-explored potential of attention mechanisms in cross-modal alignment. The improved interaction paradigm not only refines vision-language feature fusion but also provides stronger inductive biases for complex reasoning tasks.", + "bbox": [ + 511, + 425, + 903, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Rotary Position Embeddings. Following [71], we implement Multimodal RoPE (M-RoPE) in SAIL to harmonize positional modeling for multimodal inputs. The method decomposes positional encoding into two axes: height, and width. For text tokens, all axes share uniform position IDs (aligned with 1D-RoPE), whereas for images, height/width IDs adaptively map to token coordinates, as is shown in Figure 2(C). Notably, position indexing is sequentially initialized across modalities (e.g., starting from images before extending to subsequent text), preserving inter-modal consistency. M-RoPE not only improves positional sensitivity but also constrains absolute position magnitudes for visual tokens, facilitating robust generalization to extended sequences in inference.", + "bbox": [ + 511, + 623, + 903, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Pretraining", + "text_level": 1, + "bbox": [ + 511, + 847, + 638, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We apply a two-stage curriculum to progressively strengthen the visual perception of SAIL while preserving", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20760", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7e4de715e61dd3d9bcd8398615d081679a59a412f7a74115d4387dad3353f861.jpg", + "image_caption": [ + "(A) Model Architecture" + ], + "image_footnote": [], + "bbox": [ + 125, + 92, + 503, + 339 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f386a7e5da052dd8a7f2235a610b1e87504e6c81edfbb8fae052f42dfc7eeac1.jpg", + "image_caption": [ + "(C) Multimodal RoPE for SAIL", + "Figure 2. Model architecture and micro-designs for SAIL. (A) Model Architecture: SAIL is a unified transformer that processes both images and texts without extra module designs. (B) Mixed Attention Mechanism: we adopt bidirectional attention for image patches from the same image and causal attention for text tokens. Examples for a multimodal sequence and a text sequence are provided. Colored squares represent \"allow to attend\" and white squares indicate \"prevent from attending\". (C) Multimodal RoPE: an illustration of the multimodal rotary position embedding for SAIL, with examples for a multimodal sequence and a text sequence." + ], + "image_footnote": [], + "bbox": [ + 535, + 89, + 867, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "its inherent language capabilities.", + "bbox": [ + 89, + 463, + 313, + 479 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/b4bbd09bbe43a0d8800395fdac7badde95b4f539f571c1f0d4392350a14d9637.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StageDatasetImg.ResNum
Pretraining S1Recap-DataComp-1B [42]224x224512M
SlimPajama [63]-
Pretraining S2Capfusion [78]AnyRes60M
OCR from LAION COCO [60]7M
InifinityMM Stage 2 subset [28]19M
SlimPajama [63]-
SFTInifinityMM Stage3 [28]AnyRes6M
", + "bbox": [ + 94, + 492, + 478, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Details of training datasets used across all stages. \"Img.Res\" refers to the image resolution settings applied during each training stage. All datasets listed are publicly available. Note that these settings represent the default configuration for standard SAIL training, while separate settings are used for scaling experiments and ablation studies.", + "bbox": [ + 89, + 619, + 482, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 1: Accelerated Visual Knowledge Acquisition. In this stage, we pretrain SAIL on large-scale image-text pairs to rapidly bootstrap its visual understanding. To maximize data throughput, we uniformly resize all images to a lower resolution (e.g., $224 \\times 224$ ), reducing multimodal sequence lengths and enabling the model to process more samples within fixed training time. To prevent catastrophic forgetting of linguistic knowledge, we interleave pure text corpora with multimodal data during training. This hybrid approach ensures efficient exposure to visual patterns while maintaining robust language proficiency.", + "bbox": [ + 88, + 703, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stage 2: Enhancing Any-Resolution Image Understanding. Real-world applications require robustness to images", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of varying resolutions and aspect ratios, such as documents, charts, or infographics. Following prior works [11, 21], we extend pretraining with an any-resolution strategy: images retain their native resolutions during processing, and positional embeddings adapt dynamically to arbitrary spatial dimensions. This stage further refines SAIL's ability to model fine-grained visual details (e.g., tabular structures, text-rich graphics) while continuing to incorporate text-only data for language capability preservation.", + "bbox": [ + 511, + 463, + 906, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pretraining Objective. Throughout both stages, we optimize the standard language modeling loss only on text tokens. Image patches and special visual tokens are excluded from loss computation.", + "bbox": [ + 511, + 599, + 906, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Supervised Fine-tuning", + "text_level": 1, + "bbox": [ + 511, + 667, + 730, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During the Supervised Fine-tuning (SFT) stage, we train SAIL on publicly available, multi-source instruction datasets to enhance its understanding of complex linguistic instructions and diverse dialogue patterns critical for real-world deployment. This phase fine-tunes the entire network architecture, focusing on aligning the model's responses with human intent through exposure to varied instructional formats and multimodal interactions.", + "bbox": [ + 511, + 688, + 905, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 shows the details of training datasets for pretraining and supervised fine-tuning (SFT) across all stages. During Stage 1 pretraining, we utilized mixed multimodal and pure text datasets including Recap-DataComp-1B [42] and SlimPajama [63], with images at a resolution of $224 \\times 224$ , totaling 512M image-text pairs. In Stage 2, the pretrain", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20761", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing datasets include Capfusion [78], self-curated OCR data from LAION COCO [60], InfinityMM Stage 2 subset [28], and SlimPajama, utilizing the any resolution (AnyRes) strategy, with a combined total of 86M image-text pairs along with text data. The SFT stage employed the InfinityMM Stage 3 subset, processed at any resolution, containing 6M image-text pairs.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 213, + 215, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 239, + 297, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Benchmarks. For evaluation of vision and language tasks, we evaluate SAIL and existing MLLMs on a broad range of multimodal benchmarks. Specifically, MLLM benchmarks encompass MMBench-EN [49], SEEDBench-IMG [39], MMVet [79], MME [45], HallusionBench [29], MathVistaMINI [52], and OCR-Bench [50]. Visual question answering benchmarks include TextVQA [62], ScienceQA-IMG [51], AI2D [38], MMStar [9], RealWorldQA [75].", + "bbox": [ + 89, + 262, + 482, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For evaluation of vision representation learning, we conduct experiments on ImageNet-1K [20] for image classification, ADE20K for semantic segmentation [83], and ARO [80] for attribute, relation, and ordering.", + "bbox": [ + 89, + 398, + 482, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. For pretraining, we initialize SAIL from the Mistral-7B-v0.1 base LLM and set the patch size to 14. We modify Megatron [61] to support SAIL's multimodal input. Pretraining uses 128 NVIDIA A100 80G GPUs with 2-way tensor parallelism and 64-way data parallelism. The learning rate is set at 5e-5 and decays cosinely to a minimum of 5e-6. For training efficiency, we concatenate sequences from different data samples into one long sequence of 32,768 tokens, adjusting the attention mask to ensure that tokens from different samples do not attend to each other. We use a round-robin approach to interleave image-text packed sequences and pure text packed sequences, configuring the global batch to contain approximately 16K image-text pairs.", + "bbox": [ + 89, + 460, + 482, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For Supervised Fine-Tuning (SFT), the global batch size is set to 512. Training is performed for one epoch with a maximum learning rate of 1e-5, following a linear warm-up phase and then transitioning to a cosine decay schedule.", + "bbox": [ + 89, + 672, + 482, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For vision, we load the checkpoint after Stage 1 pretraining and keep it frozen for downstream evaluations, including (1) image classification on ImageNet-1K [20], (2) semantic segmentation on ADE20K [83], and (3) attribute, relation and, ordering on the ARO benchmark [80]. Specifically, (1) for image classification, we utilize a linear classifier with 90 epochs for linear probing, where detailed configurations are mostly obtained from common practices [31]. Images are resized to $224 \\times 224$ and the global batch size is 8,192 across 8 A100 (80G) GPUs. (2) For semantic segmentation, we adopt ViT-Adapter [12] with UperNet [76] as the segmenta", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion decoder. The implementation is based on MMSegmentation [17] with 80k training iterations. The input resolution is $512 \\times 512$ and the global batch size is 16 across 8 A100 (80G) GPUs. (3) For attribute, relation and, ordering, we regard the negative of the caption loss over each image-text pair as the similarity metric for retrieval.", + "bbox": [ + 511, + 90, + 903, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 193, + 714, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Results on Vision Language Tasks", + "text_level": 1, + "bbox": [ + 511, + 215, + 795, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Table 2, we compare SAIL against existing MLLMs across 13 vision-language benchmarks. SAIL consistently outperforms other Single Transformer-based models like Fuyu [5], EVE [21], SOLO [11], MonoInternVL [53], and EVE2 [22] across diverse vision-language tasks. This demonstrates that SAIL can achieve significant performance gains and push the boundaries of Single Transformer-based MLLMs without needing extra component designs or auxiliary training losses. Moreover, when compared to methods employing discrete vision tokens (e.g., Chameleon and Emu3), SAIL demonstrates superior performance. These results validate that scaling up single-transformer pretraining effectively enhances cross-modal alignment between images and text. Compared to the state-of-the-art modular MLLM LLaVA-OneVision [40], SAIL achieves comparable performance on some benchmarks, such as MMStar, SEEDBench-IMG, and RealWorldQA. While the performance of Single Transformer-based MLLMs still lags behind modular MLLMs in certain areas, we hypothesize that scaling the pretraining data volume or incorporating higher-quality instruction-tuning data will bridge the remaining performance gap.", + "bbox": [ + 511, + 234, + 906, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2. Results on Vision Representation Learning", + "text_level": 1, + "bbox": [ + 511, + 577, + 857, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we compare the quality of learned visual representations of our SAIL with other Single Transformers, including EVE [21], EVE2 [22], and SOLO [11].", + "bbox": [ + 511, + 597, + 903, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Classification and Segmentation. As demonstrated in Table 3, our method, SAIL, achieves a Top-1 accuracy of $84.06\\%$ and a Top-5 accuracy of $97.48\\%$ on the validation set of ImageNet-1K [20], significantly outperforming state-of-the-art alternatives [11, 21, 22]. In the segmentation task, SAIL also demonstrates superior performance with an mIoU of $55.30\\%$ , an mAcc of $67.24\\%$ , and an aAcc of $84.87\\%$ , illustrated in Table 3. These results indicate that SAIL is effective in both classification and segmentation tasks, offering substantial improvements over existing methods. In Table 4, even when comparing with other state-of-the-art vision backbones, our SAIL manages to achieve remarkable competitive performance with significantly less training data, demonstrating the scaling property of SAIL.", + "bbox": [ + 511, + 643, + 906, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attribute, Relation, and Ordering. To systematically evaluate the ability of SAIL to understand different types of relationships, attributes, and order information, we conduct", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20762", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/898288e5affcc9e3a84ca8645bd174c549a22bdb25207ef3a93f6c98e728f478.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method#Param#Data#VtokenGeneral VQAHallucinationMath&knowledgeOCR VQA
MMS*MMBenSEEDIMMVMMERWQAPOPEHalluSQA1MathVTQAAI2DOCRB
Modular MLLMs:
InternVL-1.5 [13]2.2B-/-332846.770.969.839.3190257.988.337.384.941.370.569.8654
QwenVL-Chat [3]7B7.2B / 50M25634.560.658.2-184849.3-36.868.235.361.545.9488
LLVA-1.5 [46]7B0.4B+ / 665K57633.164.364.330.5185954.885.927.666.825.546.154.8318
LLVA-1.6 [47]7B0.4B+ / 760K288037.667.464.743.9184257.886.427.670.232.564.966.6532
Cambrian-1 [66]8B10B+ / 7M57650.775.974.7--64.2-30.680.448.171.773.0-
LLVA-OneVision [40]7B10B+ / 3.2M729060.981.774.858.8199865.5--96.656.1-81.6-
Single Transformer-based MLLMs:
Fuyu [5]8B-/--34.410.759.321.4-43.78429.856.830.2-46.8366
Chameleon [64]7B1.4B+ / 1.8M102431.131.130.68.31703919.417.147.222.54.846.07.0
EVE [21]7B33M / 1.8M2304-52.364.625.71628-85.0-64.9-56.861.0398
SOLO [11]8B43.7M / 2M102435.867.764.430.4126044.778.640.473.332.925.061.4126
Mono-InternVL [53]3B1.3B / 7M6400-65.567.440.11875--45.793.645.772.668.6767
Emu3 [73]8B-/-16K46.658.568.237.2-57.485.231.789.231.364.770.0687
EVE2 [22]7B92M / 7.3M2500-66.371.445.0170962.487.6-96.2-71.174.8702
SAIL7B600M / 6M360053.170.172.946.3171963.985.854.293.357.077.176.7783
", + "bbox": [ + 94, + 88, + 903, + 320 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/26c8b3c48f5b96c37a8a141991a6911b2cf9f3cccfee847afdfc86f1222c139d.jpg", + "table_caption": [ + "Table 2. Comparison with existing vision-language models on various vision-language benchmarks, including MMS*: MMStar [9]; MMBench-EN [49]; SEED: SEEDBench-IMG [39]; MMV: MMVet [79]; MME [45]; POPE [43]; Hallu: HallusionBench [29]; SQA: ScienceQA-Img [51]; TVQA: TextVQA [62]; MathV: MathVistaMINI [52]; AI2D [38]; RWQA: RealWorldQA [75]; OCRB: OCR-Bench [50]. Note that #A-Param denotes the number of activated parameters; #Data represents the pre-training / fine-tuning data volume; #Vtoken indicates the maximum image patch tokens. For MME, we report the sum of perception and cognition scores. The top two results are highlighted in bold and underline, respectively. All results are derived from those reported in other papers and the official reproduction results from the OpenCompass leaderboard [24]. Our results are obtained from VLMEvalKit [24]." + ], + "table_footnote": [], + "table_body": "
MethodClassificationSegmentation
Top-1Top-5mIoUmAccaAcc
EVE [21]42.0365.7727.1235.8972.91
EVE2 [22]44.8669.4140.8553.5379.31
SOLO [11]59.1080.8935.1144.8176.02
SAIL84.0697.4855.3067.2484.87
", + "bbox": [ + 91, + 457, + 480, + 563 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ac58cf334df9fa18cc1e9a1c5e1ae9744ecdeef63f2ac21834885222f2d93fc5.jpg", + "table_caption": [ + "Table 3. Comparison on image classification and semantic segmentation with other encoder-free approaches. Our SAIL outperforms other alternatives by a large margin." + ], + "table_footnote": [], + "table_body": "
Method#Data#ParamImageNet-1KADE20K
OpenCLIP-H [15]2B0.6B84.4-
OpenCLIP-G [15]2B1.8B86.239.3†
ViT-22B [19]3B22B89.555.3
InternViT [14]6B6B88.258.7
SAIL0.5B7B84.155.3
", + "bbox": [ + 99, + 637, + 475, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "experiments on the ARO benchmark [80]. As demonstrated in Table 5, SAIL encodes compositional relationships between objects and attributes almost perfectly, significantly surpassing other state-of-the-art vision backbones.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/58247d54c3914871d0dcab7932bae06effc80ba180323469d40c7908c71ca038.jpg", + "table_caption": [ + "Table 4. Comparison on image classification and semantic segmentation with other vision backbones. $\\dagger$ indicates training with head tuning using UperNet [76], while others are based on ViT-Adapter [12]. SAIL, with significantly less training data, achieves competitive performance." + ], + "table_footnote": [], + "table_body": "
MethodRelationAttributeOrder
COCOFlickr30K
OpenCLIP-H [15]49.964.632.640.4
OpenCLIP-G [15]49.965.633.038.3
CLIP-B/32 [59]59.262.948.157.9
CLIP-L/14 [59]61.261.746.856.8
InternViT [14]59.666.073.476.3
NegCLIP [80]81.071.086.091.0
CapPa [68]86.785.798.899.2
SAIL100.099.5100.0100.0
", + "bbox": [ + 516, + 457, + 903, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 5. Comparison on attribute, relation, and ordering (ARO) with other vision backbones. SAIL almost encodes compositional relationships between objects and attributes perfectly.", + "bbox": [ + 511, + 628, + 906, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Properties of Single Transformer", + "text_level": 1, + "bbox": [ + 511, + 688, + 803, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1. Scaling Properties.", + "text_level": 1, + "bbox": [ + 511, + 710, + 692, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Scaling: We selected models of different sizes: SAIL-0.5B, SAIL-3B, and SAIL-7B (SAIL by default) for our experiments. Each model underwent Stage 1 pretraining on a mixed multimodal and pure text dataset, encountering 512M image-text pairs. Subsequently, they were finetuned on the LLaVA-mix-665K dataset using the any resolution (anyres) strategy. We evaluated the models based on their performance on vision and language benchmarks after supervised fine-tuning.", + "bbox": [ + 511, + 732, + 906, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The normalized performance of SAIL against model size is plotted in Figure 3. As the model size scales up, we ob", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20763", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f63de2d668ebe946af061ff2b559117f9fbab47d3ccc1f41eb89bcd60f9949a.jpg", + "image_caption": [ + "Figure 3. Model scaling of SAIL. Left: As the model size increases, the training language modeling loss gradually decreases. Right: As the model size increases, performance on downstream VLM tasks progressively improves." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 480, + 213 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "serve a corresponding enhancement in performance. Additionally, as shown on the left side of Figure 3, the training language modeling loss decreases with increasing model size. This reduction in training loss indicates that larger models have a greater capacity to learn multimodal alignments effectively, enabling them to capture complex relationships between vision and language more accurately. The improved learning capacity directly translates to better performance on downstream VLM tasks, showcasing the benefits of scaling up the Single Transformer architecture.", + "bbox": [ + 88, + 295, + 482, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data Scaling: we compared SAIL with its modular MLLM counterpart. For the modular MLLM, we used SigLIP-SO [81] as the vision encoder, and the language model shared the same architecture and initialization parameters as SAIL. Both models were pre-trained using Pretraining stage-1 setting, with SAIL encountering 32M, 128M, and 512M image-text pairs during training, followed by finetuning on the LLaVA-mix-665K dataset. All parameters of both models are trainable. Both models employ an identical number of input tokens for images and text. The normalized performance of both models is plotted in Figure 1(A). The results show that in the low-data regime (32M), SAIL's performance lags behind the modular MLLM, likely due to SigLIP's prior training on 40B samples. However, as the data scales, SAIL exhibits a steeper performance curve, indicating more promising data scaling properties. At 512M image-text pairs, SAIL achieves performance comparable to the modular MLLM in our evaluation subset. This demonstrates the single transformer's superior data scalability, even without a pretrained vision encoder.", + "bbox": [ + 91, + 448, + 482, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results on evaluated benchmark tasks of model scaling and data scaling are tabulated in Supp.", + "bbox": [ + 89, + 750, + 482, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2. Information Flow Pattern", + "text_level": 1, + "bbox": [ + 89, + 790, + 316, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different attention pattern compared to modular MLLM: since our comparative experiments show that the Single Transformer model exhibits more promising data scaling properties, we conducted an analysis of the trained SAIL model and its modular MLLM counterpart. Specifically, we followed the methodology from FastV [10] to", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc3e29b6eb89c82f900839c5c0ca1a9eeef2287bd2b5759db834896d812bde8e.jpg", + "image_caption": [ + "Figure 4. Image Attention Score Allocation: The figure shows the proportion of image attention scores across different transformer layers for Single Transformer-based MLLM and modular MLLM when predicting tokens. Single Transformer-based MLLM generally allocates higher attention weights to image tokens compared to modular MLLM." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "analyze the attention score distribution for each predicted token given an image and a user query. This analysis focuses on how much attention is allocated to image tokens during token prediction. We selected 1000 samples from various datasets including VQAv2, GQA, TextVQA, DocVQA, MME, SEEDBench-IMG, MMBench, and some self-curated dialog examples. For each model prediction, we computed the average attention scores assigned to previous image tokens by the output token.", + "bbox": [ + 511, + 393, + 905, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted a comparative experiment between Single Transformer-based MLLMs (SAIL, SOLO [11], EVE [21]) and modular MLLMs (Qwen2-VL [71], LLaVA-OneVision [40], LLaVA-1.5 [46]).", + "bbox": [ + 511, + 532, + 905, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results are depicted in Figure 4. Single Transformer-based MLLMs allocate between $60\\%$ and $80\\%$ of attention scores to image tokens across all layers when predicting tokens. In contrast, modular MLLMs such as Qwen2-VL and LLaVA-OneVision allocate only $10\\%$ to $30\\%$ of attention scores to image tokens across different layers. For LLaVA-1.5, which does not update the ViT parameters during supervised fine-tuning (SFT), the image attention score is relatively high in the first two transformer layers but declines sharply in subsequent layers.", + "bbox": [ + 511, + 595, + 905, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From this experiment, we can conclude that Single Transformer-based MLLMs tend to allocate a significant portion of attention to previous image tokens during prediction. In contrast, modular MLLMs allocate a smaller portion of their attention directly to image tokens, indicating a less image-centric approach in their prediction mechanism.", + "bbox": [ + 511, + 748, + 905, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These findings indicate that the Single Transformer model places more emphasis on grounding its predictions in the visual information. As the model undergoes data scaling, it allocates more effective computation to image tokens,", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20764", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6048758e170d955de58ba2977175f5060395adc72e1ee17b1cfadaa6e71c5d41.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMMBench [49]MME [45]
Physical RelationCelebrity RelationPositionPostersCelebrity
Modular MLLM30.450.598.3134.0100.3
SAIL52.288.9160.0108.275.0
", + "bbox": [ + 94, + 88, + 478, + 176 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "thereby enhancing its capability as a vision-centric model.", + "bbox": [ + 89, + 266, + 475, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, the attention pattern analysis underscores the Single Transformer's ability to robustly integrate visual context, enabling it to scale efficiently and potentially outperform modular MLLMs in vision-language tasks.", + "bbox": [ + 89, + 281, + 482, + 342 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.3. Task-Specific Performance Analysis", + "text_level": 1, + "bbox": [ + 89, + 349, + 385, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We dissect SAIL through case studies:", + "bbox": [ + 89, + 369, + 344, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Strengths: Spatial Reasoning. SAIL excels at tasks requiring precise spatial location. As shown in Table 6, under the setting of our data scaling experiment, it outperforms the modular counterpart by 61.7 points on the MME Position split and $21.8\\%$ on MMBench Physical Relation questions. The unified architecture likely enables tighter coupling between visual geometry and linguistic descriptions.", + "bbox": [ + 89, + 383, + 482, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Weaknesses: World Knowledge. Conversely, SAIL falls short in tasks that demand extensive world knowledge. As shown in Table 6, SAIL underperforms in the MME celebrity and art splits compared to the modular MLLM. This underperformance can be attributed to SAIL's lack of diverse domain-specific data during pretraining, a gap that was not sufficiently addressed during supervised finetuning. Modular MLLMs, with their pretrained vision encoders like CLIP [15, 59] or SigLIP [81], have a broader knowledge base and therefore handle such tasks more effectively. We hypothesize that scaling up SAIL's pretraining data diversity could help bridge this gap, enhancing its performance on knowledge-intensive tasks.", + "bbox": [ + 89, + 489, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Empirical Observations on Basic Factors", + "text_level": 1, + "bbox": [ + 89, + 696, + 441, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To guide scalable training of single-transformer MLLMs, we conduct ablation studies on two critical design choices using SAIL-0.5B pretrained on 128M image-text pairs and fine-tuned on LLaVA-mix-665K. Performance is evaluated through zero-shot image classification after pretraining [67] and vision-language benchmarks after SFT.", + "bbox": [ + 89, + 719, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Bidirectional Attention for Image Patches with Multimodal Position Encoding. We compare two approaches for integrating image patches into the transformer: (1) Causal attention with 1D positional encoding, using a token to demarcate image rows. (2) Full bidirectional attention for image patches paired with mul", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fbc15c2fc7e771856c3eeca0b13e830272fdf699943a1cc5ef90cfa85321b81c.jpg", + "table_caption": [ + "Table 6. Performance Comparison of SAIL and Modular MLLM in MMBench and MME Tasks: the strengths of SAIL in spatial reasoning tasks (MMBench Physical Relation and MME Position split) and its weaknesses in world knowledge tasks (MMBench Celebrity Relation and MME Celebrity and Posters splits)." + ], + "table_footnote": [], + "table_body": "
Exp. SettingVQAv2GQASQATQASEED-I
Default59.146.959.620.135.1
#1 No Img full attn57.845.258.716.233.8
#2 No pure text in PT56.342.148.618.332.4
", + "bbox": [ + 522, + 89, + 901, + 154 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation Study on Basic Factors for SAIL: We presents the impact of different ablation settings on the performance of SAIL across VQAv2 [27], GQA [35], SQA [51], TQA [62], and SEED-I [39]. The default setting includes image full attention and including pure text data in pretraining. Ablation #1 removes image full attention, and ablation #2 excludes pure text in pretraining.", + "bbox": [ + 511, + 159, + 906, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "timodal rotary position embeddings (RoPE), which jointly encode spatial coordinates (e.g., 2D grid positions) and text token positions. As shown in Table 7, using bidirectional attention with multimodal RoPE significantly improves performance on vision-language tasks, with a particularly notable gain of $3.1\\%$ on TextVQA. This suggests that enabling cross-patch interactions during pretraining enhances visual representation learning and tightens cross-modal alignment. Interleaving Pure Text Data During Pretraining. We analyze the impact of mixing SlimPajama text data with image-text pairs during pretraining. The results, as presented in Table 7 #2, reveal that mixing in pure text data consistently improves performance across vision and language benchmarks. This finding underscores the importance of preserving language capabilities in the LLM when training Single Transformer models, as maintaining strong language skills is crucial for building a multimodal model capable of complex reasoning. Currently, incorporating text data in training is one of the effective methods to maintain the language abilities of the model.", + "bbox": [ + 511, + 250, + 906, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our ablation studies identify key designs for training SAIL effectively. Using bi-directional attention with multimodal rotary position embeddings enhances visual perception, while incorporating pure text data preserves language capabilities for robust multimodal performance.", + "bbox": [ + 511, + 551, + 905, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we conducted an extensive analysis of Single Transformer-based MLLMs compared to modular MLLMs. Our investigation explored the unique properties of Single Transformers, including scalability, cross-modal information flow patterns, and visual representation capabilities. A series of experiments on our trained SAIL model demonstrated that this unified architecture achieves performance on vision-language benchmarks comparable to modular MLLMs while also functioning effectively as a vision backbone. Our findings highlight several advantages of Single Transformer architectures, such as superior data scalability, vision-centric information flow, and inherent capabilities as a powerful vision encoder. We hope our empirical findings will inspire further research to refine and enhance Single Transformer architectures, advancing the field of multimodal intelligence.", + "bbox": [ + 511, + 657, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20765", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2", + "[2] AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 1", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2, 6", + "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3", + "[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Taşirlar. Introducing our multimodal models, 2023. 2, 5, 6", + "[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 3", + "[7] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2", + "[8] Jieneng Chen, Qihang Yu, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Vitamin: Designing scalable vision models in the vision-language era. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3", + "[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 5, 6", + "[10] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 7", + "[11] Yangyi Chen, Xingyao Wang, Hao Peng, and Heng Ji. A single transformer for scalable vision-language modeling. Transactions on Machine Learning Research, 2024. 1, 2, 3, 4, 5, 6, 7", + "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 5, 6", + "[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 6", + "[14] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2, 3, 6", + "[15] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 1, 2, 3, 6, 8", + "[16] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality, 2023. 1", + "[17] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.5", + "[18] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1", + "[19] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdul-mohsin, et al. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 3, 6", + "[20] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3, 5", + "[21] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. In Advances in Neural Information Processing Systems, pages 52545-52567. Curran Associates, Inc., 2024. 1, 2, 3, 4, 5, 6, 7", + "[22] Haiwen Diao, Xiaotong Li, Yufeng Cui, Yueze Wang, Haoge Deng, Ting Pan, Wenxuan Wang, Huchuan Lu, and Xinlong Wang. *Evev2: Improved baselines for encoder-free vision-language models.* arXiv preprint arXiv:2502.06788, 2025. 5, 6", + "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20766", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11198-11201, 2024. 6", + "[25] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19358-19369, 2023. 3", + "[26] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 36:27092-27112, 2023. 3", + "[27] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 8", + "[28] Shuhao Gu, Jialing Zhang, Siyuan Zhou, Kevin Yu, Zhaohu Xing, Liangdong Wang, Zhou Cao, Jintao Jia, Zhuoyi Zhang, Yixuan Wang, et al. Infinity-mm: Scaling multimodal performance with large-scale and high-quality instruction data. arXiv preprint arXiv:2410.18558, 2024. 4, 5", + "[29] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 5, 6", + "[30] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3", + "[31] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5", + "[32] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. arXiv preprint arXiv:2312.08914, 2023. 2", + "[33] Xinyu Huang, Youcai Zhang, Jinyu Ma, Weiwei Tian, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Lei Zhang. Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657, 2023. 3", + "[34] Zilong Huang, Qinghao Ye, Bingyi Kang, Jiashi Feng, and Haoqi Fan. Classification done right for vision-language pretraining. In NeurIPS, 2024. 3", + "[35] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 8", + "[36] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 3", + "[37] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1", + "[38] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images, 2016. 5, 6", + "[39] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5, 6, 8", + "[40] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 5, 6, 7", + "[41] Xianhang Li, Zeyu Wang, and Cihang Xie. An inverse scaling law for clip training. In NeurIPS, 2023. 3", + "[42] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, Yuyin Zhou, and Cihang Xie. What if we recapture billions of web images with llama-3? arXiv preprint arXiv:2406.08478, 2024. 4", + "[43] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 6", + "[44] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23390-23400, 2023. 3", + "[45] Zijing Liang, Yanjie Xu, Yifan Hong, Penghui Shang, Qi Wang, Qiang Fu, and Ke Liu. A survey of multimodel large language models. In Proceedings of the 3rd International Conference on Computer, Artificial Intelligence and Control Engineering, pages 405-409, 2024. 5, 6, 8", + "[46] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 6, 7", + "[47] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.3,6", + "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3", + "[49] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20767", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5, 6, 8", + "[50] Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xu-Cheng Yin, Cheng-Lin Liu, Lianwen Jin, and Xiang Bai. Ocrbench: on the hidden mystery ofOCR in large multimodal models. Science China Information Sciences, 67(12):220102, 2024. 5, 6", + "[51] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5, 6, 8", + "[52] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 5, 6", + "[53] Gen Luo, Xue Yang, Wenhan Dou, Zhaokai Wang, Jiawen Liu, Jifeng Dai, Yu Qiao, and Xizhou Zhu. Mono-internvl: Pushing the boundaries of monolithic multimodal large language models with endogenous visual pre-training. arXiv preprint arXiv:2410.08202, 2024. 1, 2, 3, 5, 6", + "[54] Sachin Mehta, Maxwell Horton, Fartash Faghri, Mohammad Hossein Sekhavat, Mahyar Najibi, Mehrdad Farajtabar, Oncel Tuzel, and Mohammad Rastegari. Catlip: Cliplevel visual recognition accuracy with 2.7 x faster pretraining on web-scale image-text data. arXiv preprint arXiv:2404.15653, 2024. 3", + "[55] Meta. Introducing meta llama 3: The most capable openly available llm to date, 2024. Accessed: 2024-04-18. 1", + "[56] OpenAI. Introducing chatgpt. OpenAI Blog, 2021.", + "[57] OpenAI. Gpt-4 technical report, 2023. 1", + "[58] OpenAI. Gpt-4v(ision) system card, 2023. 1, 2", + "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 3, 6, 8", + "[60] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 3, 4, 5", + "[61] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatronlm: Training multi-billion parameter language models using model parallelism. arXiv preprint arXiv:1909.08053, 2019. 5", + "[62] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 6, 8" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Daria Soboleva, Faisal Al-Khateeb, Robert Myers, Jacob R Steeves, Joel Hestness, and Nolan Dey. SlimPajama: A 627B token cleaned and deduplicated version of RedPajama. https://cerebras.ai/blog/slimpajama-a-627b-token-cleaned-and-deduplicated-version-of-redpajama, 2023.4", + "[64] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2, 6", + "[65] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2", + "[66] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2024. 6", + "[67] Michael Tschannen, Manoj Kumar, Andreas Steiner, Xiaohua Zhai, Neil Houlsby, and Lucas Beyer. Image captioners are scalable vision learners too. Advances in Neural Information Processing Systems, 36:46830-46855, 2023. 3, 8", + "[68] Michael Tschannen, Manoj Kumar, Andreas Steiner, Xiaohua Zhai, Neil Houlsby, and Lucas Beyer. Image captioners are scalable vision learners too. Advances in Neural Information Processing Systems, 36:46830-46855, 2023. 6", + "[69] Haochen Wang, Anlin Zheng, Yucheng Zhao, Tiancai Wang, Ge Zheng, Xiangyu Zhang, and Zhaoxiang Zhang. Reconstructive visual instruction tuning. In International Conference on Learning Representations, 2025. 2", + "[70] Jiacong Wang, Bohong Wu, Haiyong Jiang, Zhou Xun, Xin Xiao, Haoyuan Guo, and Jun Xiao. World to code: Multimodal data generation via self-instructed compositional captioning and filtering. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4608-4623, 2024. 1", + "[71] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution, 2024. 1, 2, 3, 7", + "[72] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion feedback helps clip see better. arXiv preprint arXiv:2407.20171, 2024. 3", + "[73] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 6", + "[74] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In International Conference on Learning Representations, 2022. 3", + "[75] X.ai. Grok-1.5 vision preview. https://x.ai/blog/grok-1.5v, 2024.5, 6" + ], + "bbox": [ + 516, + 90, + 905, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20768", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 5, 6", + "[77] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3", + "[78] Qiying Yu, Quan Sun, Xiaosong Zhang, Yufeng Cui, Fan Zhang, Yue Cao, Xinlong Wang, and Jingjing Liu. Capsfusion: Rethinking image-text data at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14022-14032, 2024. 4, 5", + "[79] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 5, 6", + "[80] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936, 2022. 5, 6", + "[81] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3, 7, 8", + "[82] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 3", + "[83] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5", + "[84] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 3" + ], + "bbox": [ + 91, + 92, + 482, + 656 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20769", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/2b8a94f1-28e8-4218-82c7-da4dba2c45e1_model.json b/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/2b8a94f1-28e8-4218-82c7-da4dba2c45e1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1f21a6429373e0004f43e73dfc31c45ace1a9336 --- /dev/null +++ b/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/2b8a94f1-28e8-4218-82c7-da4dba2c45e1_model.json @@ -0,0 +1,2501 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.131, + 0.882, + 0.177 + ], + "angle": 0, + "content": "The Scalability of Simplicity: Empirical Analysis of Vision-Language Learning with a Single Transformer" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.203, + 0.731, + 0.275 + ], + "angle": 0, + "content": "Weixian Lei* Jiacong Wang* Haochen Wang* \nXiangtai Li Jun Hao Liew Jiashi Feng Zilong Huang† \n*Equal contribution, † Corresponding author \nBytedance Seed" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.485, + 0.691 + ], + "angle": 0, + "content": "This paper introduces SAIL, a single transformer unified multimodal large language model (MLLM) that integrates raw pixel encoding and language decoding within a singular architecture. Unlike existing modular MLLMs, which rely on a pre-trained vision transformer (ViT), SAIL eliminates the need for a separate vision encoder, presenting a more minimalist architecture design. Instead of introducing novel architectural components, SAIL adapts mix-attention mechanisms and multimodal rotary position embedding to better align with the distinct characteristics of visual and textual modalities. We systematically compare SAIL's properties—including scalability, cross-modal information flow patterns, and visual representation capabilities—with those of modular MLLMs. By scaling both training data and model size, SAIL achieves performance comparable to modular MLLMs. Notably, the removal of pretrained ViT components enhances SAIL's scalability and results in significantly different cross-modal information flow patterns. Moreover, SAIL demonstrates strong visual representation capabilities, achieving results on par with ViT22B in vision tasks such as semantic segmentation. Code and models are available at https://github.com/bytedance/SAIL." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.723, + 0.222, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The pursuit of multimodal intelligence has driven the development of Multimodal Large Language Models (MLLMs) [48, 58, 65], which typically adopt a modular design: a pre-trained vision encoder (e.g., CLIPViT [15, 59]) extracts image features, a Large Language Model (LLM) [2, 16, 37, 55-57] processes text, and a lightweight projector aligns the two modalities. This framework achieves strong performance through multi-stage pretraining, supervised fine-tuning (SFT), and post-training on multimodal datasets [3, 18, 48, 70, 71, 84]. While effective," + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.318, + 0.715, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.463, + 0.635, + 0.475 + ], + "angle": 0, + "content": "(A)" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.318, + 0.905, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.463, + 0.823, + 0.474 + ], + "angle": 0, + "content": "(B)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Figure 1. (A) Data scaling curve for Modular Multimodal Large Language Model (MLLM) and SAIL, our Single Transformer-based MLLM. As pretraining data increases, the single transformer SAIL shows a sharper performance gain, demonstrating its superior data scalability. Experiment settings are described in Section 4.3. (B) Comparison to existing Single Transformer-based MLLMs: our SAIL pushes the performance boundaries on both vision tasks and vision-language tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.906, + 0.67 + ], + "angle": 0, + "content": "this modular MLLM paradigm inherently fragments multimodal processing, reinforces reliance on pretrained visual encoders, which may limit deployment flexibility and scalability [11, 21, 53]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.672, + 0.907, + 0.868 + ], + "angle": 0, + "content": "A promising alternative is to eliminate the visual encoder entirely and process raw image patches and text tokens within a single Transformer. This unified architecture removes modality-specific modules, enabling parameter sharing and end-to-end learning of vision-language interactions. Previous works [11, 21, 53] have primarily explored the architecture design, training data, and methods of Single Transformer-based MLLMs. However, little exploration has been given to their fundamental properties, such as scalability, cross-modal information flow patterns, and visual representation capabilities. A deeper understanding of these properties is crucial for unlocking the full potential of Single Transformer-based MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this work, we present an experimental analysis of the fundamental properties of Single Transformer-based" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20758" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.181 + ], + "angle": 0, + "content": "MLLMs and compare them to modular MLLMs (e.g. LLaVA [48]). Additionally, in the absence of a pre-trained visual encoder, Single Transformers have to learn visual representations from scratch. Thus, an intriguing question arises: can a trained Single Transformer emerge as a strong vision encoder?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.483, + 0.395 + ], + "angle": 0, + "content": "We conduct a series of experiments to train and study our Single trAnformer model for vIsion and Language (SAIL). While we do not propose novel architecture designs, we introduce necessary modifications to enable the model to process different modalities in a unified architecture. In its micro architecture design, we address the different spatial characteristics of 2D images and 1D text data by employing a mixed attention mechanism: bidirectional attention for image patches and causal attention for text tokens, combined with multimodal rotary position embedding. Through model and data scaling, SAIL achieves performance on vision-language benchmarks comparable to modular MLLMs, while also functioning as a high-performing vision backbone, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.483, + 0.426 + ], + "angle": 0, + "content": "More concretely, our empirical analysis uncovers following striking advantages of Single Transformer architectures:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.483, + 0.548 + ], + "angle": 0, + "content": "(i) Superior Data Scaling: In controlled experiments, SAIL exhibits steeper performance gains as pretraining data scales. While LLaVA-style modular MLLMs initially perform well, our model's performance becomes very close to theirs when pretrained on 512M samples, as shown in Figure 1(A). This suggests that unified architectures can effectively leverage large-scale data and potentially match the performance of modular MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.671 + ], + "angle": 0, + "content": "(ii) Vision Centric Information Flow Pattern: Through analysis of attention distributions, we observe that Single Transformers assign significantly higher attention scores to image tokens during token prediction compared to modular MLLMs. This indicates that the information flow in Single Transformer MLLMs is more direct, with visual tokens influencing prediction tokens more prominently, highlighting a vision-centric approach to decision-making." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.839 + ], + "angle": 0, + "content": "(iii) Vision Encoder Functioning: Our experiments further demonstrate that the pretrained Single Transformer inherently serves as a powerful vision encoder. Comprehensive evaluations on vision-centric tasks, such as image classification and semantic segmentation, show that the model learns rich visual representations during multimodal pretraining. These representations enhance its capacity for both semantic-level comprehension (e.g., object categorization) and pixel-level understanding (e.g., fine-grained segmentation masks), bridging high-level abstraction and low-level visual reasoning within a unified architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In summary, our findings indicate that Single Transformer-based MLLMs hold great promise in surpassing modular MLLMs in terms of leveraging large-scale data, forming direct vision-centric information pathways," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "and functioning as effective vision encoders. We hope our empirical findings inspire further research to refine and enhance Single Transformer architecture, ultimately driving advancements in multimodal intelligence from a new perspective." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.185, + 0.655, + 0.201 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.212, + 0.897, + 0.228 + ], + "angle": 0, + "content": "2.1. Paradigms in Vision-Language Model Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.235, + 0.905, + 0.597 + ], + "angle": 0, + "content": "Modular MLLMs with Visual Encoders. The prevailing approach in MLLM design employs modular architectures [3, 48, 58, 65, 71] that rely on pretrained vision encoders (e.g., CLIP-ViT [15, 59], InternViT [14]) to process visual inputs. The visual features extracted from these frozen encoders are then aligned with LLM input spaces via linear [7, 48, 69, 84] or cross-attention layers [1, 32]. While this module design enables effective transfer of pretrained visual-language knowledge, it also introduces several limitations. First, incorporating a separate ViT encoder significantly slows down both training and inference, increasing deployment complexity and requiring costly infrastructure—especially when compared to a single transformer unified model. Second, common strategies for integrating visual features, such as direct mapping into LLM inputs [7, 48, 84] or sharing them across LLM layers [1, 32], often struggle to reconcile the inherent differences between images and text representations. Finally, as model scale, balancing the interactions between the encoder, LLM, and alignment layers becomes increasingly challenging [11, 53]. Thus, in this work, we explore a single transformer-based MLLM architecture that eliminates the ViT encoder and alignment components to overcome these challenges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Single Transformer-based MLLMs Without Visual Encoders. Emerging research explores end-to-end architectures that process raw image patches and text tokens through a single Transformer, bypassing visual encoders entirely. These monolithic designs fall into two categories: continuous tokenization and discrete tokenization. Continuous tokenization, exemplified by Fuyu-8B [5] and SOLO [11], directly maps patches to LLM embeddings via linear projections, enabling flexible resolution handling but requiring massive pretraining data. Discrete tokenization, adopted by Chameleon [64] and Emu3 [73], employs VQ-VAE tokenizers to compress images into discrete tokens, trading pixel-level fidelity for generation capabilities. While later efforts such as EVE [21] and MonoInternVL [53] demonstrate the feasibility of encoder-free training, critical gaps remain: (1) Existing methods rely on extra designs and auxiliary loss [21], complicating training pipelines; (2) The scaling laws and fundamental properties of purely end-to-end trained models remain poorly understood; (3) Vision-language interaction in shared parame" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20759" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "ter spaces lacks systematic analysis—most prior MLLMs default to causal attention for processing image-text sequences. In this work, we reveal that enabling bidirectional attention between image patches significantly enhances visual representation learning, addressing a key limitation in current designs. More importantly, our study bridges these gaps by establishing foundational principles for training scalable, self-contained single-transformer MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.224, + 0.373, + 0.24 + ], + "angle": 0, + "content": "2.2. Vision Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.483, + 0.336 + ], + "angle": 0, + "content": "Learning effective vision representations is a core challenge in computer vision research, with extensive works [4, 19, 20, 25, 30, 74] dedicated to this problem. With the proliferation of large-scale web-sourced image-text datasets [6, 26, 60], recent methods leverage this data to train deep visual representations via three primary paradigms:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.483, + 0.457 + ], + "angle": 0, + "content": "Text as Classification Labels. Early methods used textual descriptions as weak supervision by extracting categorical labels from captions. For example, frameworks like Tag2Text [33] and RAM [82] used ViTs [23] to predict noun-based pseudo-labels from datasets like CC12M [6]. CatLIP [54] scaled labels to millions using object-centric supervision, and SuperClass [34] directly used tokenized text tokens as classification categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Image-Text Contrastive Learning Contrastive pretraining, as exemplified by CLIP [15, 59] and ALIGN [36], aligns global image-text embeddings within a shared latent space. Subsequent works [8, 14, 41, 44, 72, 81] focused on enhancing performance and improving training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.73 + ], + "angle": 0, + "content": "Text as Autoregressive Targets. Caption generation as a pretext task is another approach for visual representation learning. SimVLM [74] trains encoder-decoder architectures to autoregressively predict captions, while CapPa [67] trains vision encoders through sequence prediction. These methods often retain modular designs or auxiliary components like contrastive losses [77]. Our work aligns with this category but removes architectural fragmentation by jointly modeling image patches and text tokens in a single Transformer. We find that the pre-trained Single Transformer learns transferable vision representations, enabling it to handle downstream multimodal understanding tasks and function as a vision encoder without modifications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.745, + 0.483, + 0.78 + ], + "angle": 0, + "content": "3. SAIL: Training a Single Transformer for Vision and Language" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.279, + 0.803 + ], + "angle": 0, + "content": "3.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.901 + ], + "angle": 0, + "content": "SAIL is built upon a unified Transformer architecture (Figure 2(A)) that processes multimodal inputs through streamlined, modality-specific preprocessing. For text, raw input is tokenized using the language model's tokenizer and then transformed into embeddings via the textual embedding module. For images, we partition the input into fixed-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "size patches and project them into continuous embeddings via a linear projection. Additionally, we maintain a list of special tokens explicitly designed for visual modality encoding: and tokens mark the beginning and end of an image patch span, respectively. In multimodal scenarios, such as image-text pairs, these embeddings are concatenated into a single sequence and fed into the Transformer, enabling joint cross-modal interactions through unified self-attention layers. This design eliminates the need for modality-specific encoders, which efficiently processess heterogeneous data within a Single Transformer framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.905, + 0.426 + ], + "angle": 0, + "content": "Bidirectional attention within image patches. While existing multimodal large language models (MLLMs) [40, 47, 48, 84] predominantly adopt causal attention for autoregressive sequence modeling, our experiments reveal that enabling full bidirectional attention among tokens from the same image significantly enhances visual representation learning and boosts downstream vision-language task performance. Note that previous Single Transformers [11, 21, 53, 73] have only utilized causal attention, without exploring the potential of mixed attention mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.905, + 0.622 + ], + "angle": 0, + "content": "As illustrated in Figure 2(B), for SAIL we implement a mixed attention scheme: (1) For text tokens, we preserve causal attention to maintain autoregressive generation capabilities, allowing each token to attend only to its predecessors. (2) For image tokens, we activate full bidirectional attention within each image patch group, empowering every visual token to interact with all others in the same image. This design captures holistic spatial relationships and contextual dependencies among visual elements, addressing the under-explored potential of attention mechanisms in cross-modal alignment. The improved interaction paradigm not only refines vision-language feature fusion but also provides stronger inductive biases for complex reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.905, + 0.835 + ], + "angle": 0, + "content": "Multimodal Rotary Position Embeddings. Following [71], we implement Multimodal RoPE (M-RoPE) in SAIL to harmonize positional modeling for multimodal inputs. The method decomposes positional encoding into two axes: height, and width. For text tokens, all axes share uniform position IDs (aligned with 1D-RoPE), whereas for images, height/width IDs adaptively map to token coordinates, as is shown in Figure 2(C). Notably, position indexing is sequentially initialized across modalities (e.g., starting from images before extending to subsequent text), preserving inter-modal consistency. M-RoPE not only improves positional sensitivity but also constrains absolute position magnitudes for visual tokens, facilitating robust generalization to extended sequences in inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.64, + 0.864 + ], + "angle": 0, + "content": "3.2. Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "We apply a two-stage curriculum to progressively strengthen the visual perception of SAIL while preserving" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20760" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.093, + 0.504, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.349, + 0.393, + 0.363 + ], + "angle": 0, + "content": "(A) Model Architecture" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.09, + 0.869, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.348, + 0.818, + 0.364 + ], + "angle": 0, + "content": "(C) Multimodal RoPE for SAIL" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.38, + 0.907, + 0.451 + ], + "angle": 0, + "content": "Figure 2. Model architecture and micro-designs for SAIL. (A) Model Architecture: SAIL is a unified transformer that processes both images and texts without extra module designs. (B) Mixed Attention Mechanism: we adopt bidirectional attention for image patches from the same image and causal attention for text tokens. Examples for a multimodal sequence and a text sequence are provided. Colored squares represent \"allow to attend\" and white squares indicate \"prevent from attending\". (C) Multimodal RoPE: an illustration of the multimodal rotary position embedding for SAIL, with examples for a multimodal sequence and a text sequence." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.464, + 0.315, + 0.48 + ], + "angle": 0, + "content": "its inherent language capabilities." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.493, + 0.48, + 0.615 + ], + "angle": 0, + "content": "
StageDatasetImg.ResNum
Pretraining S1Recap-DataComp-1B [42]224x224512M
SlimPajama [63]-
Pretraining S2Capfusion [78]AnyRes60M
OCR from LAION COCO [60]7M
InifinityMM Stage 2 subset [28]19M
SlimPajama [63]-
SFTInifinityMM Stage3 [28]AnyRes6M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.702 + ], + "angle": 0, + "content": "Table 1. Details of training datasets used across all stages. \"Img.Res\" refers to the image resolution settings applied during each training stage. All datasets listed are publicly available. Note that these settings represent the default configuration for standard SAIL training, while separate settings are used for scaling experiments and ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.704, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Stage 1: Accelerated Visual Knowledge Acquisition. In this stage, we pretrain SAIL on large-scale image-text pairs to rapidly bootstrap its visual understanding. To maximize data throughput, we uniformly resize all images to a lower resolution (e.g., \\(224 \\times 224\\)), reducing multimodal sequence lengths and enabling the model to process more samples within fixed training time. To prevent catastrophic forgetting of linguistic knowledge, we interleave pure text corpora with multimodal data during training. This hybrid approach ensures efficient exposure to visual patterns while maintaining robust language proficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Stage 2: Enhancing Any-Resolution Image Understanding. Real-world applications require robustness to images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.6 + ], + "angle": 0, + "content": "of varying resolutions and aspect ratios, such as documents, charts, or infographics. Following prior works [11, 21], we extend pretraining with an any-resolution strategy: images retain their native resolutions during processing, and positional embeddings adapt dynamically to arbitrary spatial dimensions. This stage further refines SAIL's ability to model fine-grained visual details (e.g., tabular structures, text-rich graphics) while continuing to incorporate text-only data for language capability preservation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.661 + ], + "angle": 0, + "content": "Pretraining Objective. Throughout both stages, we optimize the standard language modeling loss only on text tokens. Image patches and special visual tokens are excluded from loss computation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.731, + 0.684 + ], + "angle": 0, + "content": "3.3. Supervised Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.809 + ], + "angle": 0, + "content": "During the Supervised Fine-tuning (SFT) stage, we train SAIL on publicly available, multi-source instruction datasets to enhance its understanding of complex linguistic instructions and diverse dialogue patterns critical for real-world deployment. This phase fine-tunes the entire network architecture, focusing on aligning the model's responses with human intent through exposure to varied instructional formats and multimodal interactions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Table 1 shows the details of training datasets for pretraining and supervised fine-tuning (SFT) across all stages. During Stage 1 pretraining, we utilized mixed multimodal and pure text datasets including Recap-DataComp-1B [42] and SlimPajama [63], with images at a resolution of \\(224 \\times 224\\), totaling 512M image-text pairs. In Stage 2, the pretrain" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20761" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "ing datasets include Capfusion [78], self-curated OCR data from LAION COCO [60], InfinityMM Stage 2 subset [28], and SlimPajama, utilizing the any resolution (AnyRes) strategy, with a combined total of 86M image-text pairs along with text data. The SFT stage employed the InfinityMM Stage 3 subset, processed at any resolution, containing 6M image-text pairs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.214, + 0.216, + 0.231 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.24, + 0.298, + 0.258 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.399 + ], + "angle": 0, + "content": "Evaluation Benchmarks. For evaluation of vision and language tasks, we evaluate SAIL and existing MLLMs on a broad range of multimodal benchmarks. Specifically, MLLM benchmarks encompass MMBench-EN [49], SEEDBench-IMG [39], MMVet [79], MME [45], HallusionBench [29], MathVistaMINI [52], and OCR-Bench [50]. Visual question answering benchmarks include TextVQA [62], ScienceQA-IMG [51], AI2D [38], MMStar [9], RealWorldQA [75]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.461 + ], + "angle": 0, + "content": "For evaluation of vision representation learning, we conduct experiments on ImageNet-1K [20] for image classification, ADE20K for semantic segmentation [83], and ARO [80] for attribute, relation, and ordering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.483, + 0.673 + ], + "angle": 0, + "content": "Implementation Details. For pretraining, we initialize SAIL from the Mistral-7B-v0.1 base LLM and set the patch size to 14. We modify Megatron [61] to support SAIL's multimodal input. Pretraining uses 128 NVIDIA A100 80G GPUs with 2-way tensor parallelism and 64-way data parallelism. The learning rate is set at 5e-5 and decays cosinely to a minimum of 5e-6. For training efficiency, we concatenate sequences from different data samples into one long sequence of 32,768 tokens, adjusting the attention mask to ensure that tokens from different samples do not attend to each other. We use a round-robin approach to interleave image-text packed sequences and pure text packed sequences, configuring the global batch to contain approximately 16K image-text pairs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.735 + ], + "angle": 0, + "content": "For Supervised Fine-Tuning (SFT), the global batch size is set to 512. Training is performed for one epoch with a maximum learning rate of 1e-5, following a linear warm-up phase and then transitioning to a cosine decay schedule." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For vision, we load the checkpoint after Stage 1 pretraining and keep it frozen for downstream evaluations, including (1) image classification on ImageNet-1K [20], (2) semantic segmentation on ADE20K [83], and (3) attribute, relation and, ordering on the ARO benchmark [80]. Specifically, (1) for image classification, we utilize a linear classifier with 90 epochs for linear probing, where detailed configurations are mostly obtained from common practices [31]. Images are resized to \\(224 \\times 224\\) and the global batch size is 8,192 across 8 A100 (80G) GPUs. (2) For semantic segmentation, we adopt ViT-Adapter [12] with UperNet [76] as the segmenta" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.184 + ], + "angle": 0, + "content": "tion decoder. The implementation is based on MMSegmentation [17] with 80k training iterations. The input resolution is \\(512 \\times 512\\) and the global batch size is 16 across 8 A100 (80G) GPUs. (3) For attribute, relation and, ordering, we regard the negative of the caption loss over each image-text pair as the similarity metric for retrieval." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.194, + 0.715, + 0.21 + ], + "angle": 0, + "content": "4.2. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.217, + 0.796, + 0.233 + ], + "angle": 0, + "content": "4.2.1. Results on Vision Language Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.236, + 0.907, + 0.569 + ], + "angle": 0, + "content": "As shown in Table 2, we compare SAIL against existing MLLMs across 13 vision-language benchmarks. SAIL consistently outperforms other Single Transformer-based models like Fuyu [5], EVE [21], SOLO [11], MonoInternVL [53], and EVE2 [22] across diverse vision-language tasks. This demonstrates that SAIL can achieve significant performance gains and push the boundaries of Single Transformer-based MLLMs without needing extra component designs or auxiliary training losses. Moreover, when compared to methods employing discrete vision tokens (e.g., Chameleon and Emu3), SAIL demonstrates superior performance. These results validate that scaling up single-transformer pretraining effectively enhances cross-modal alignment between images and text. Compared to the state-of-the-art modular MLLM LLaVA-OneVision [40], SAIL achieves comparable performance on some benchmarks, such as MMStar, SEEDBench-IMG, and RealWorldQA. While the performance of Single Transformer-based MLLMs still lags behind modular MLLMs in certain areas, we hypothesize that scaling the pretraining data volume or incorporating higher-quality instruction-tuning data will bridge the remaining performance gap." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.578, + 0.859, + 0.594 + ], + "angle": 0, + "content": "4.2.2. Results on Vision Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.905, + 0.643 + ], + "angle": 0, + "content": "In this section, we compare the quality of learned visual representations of our SAIL with other Single Transformers, including EVE [21], EVE2 [22], and SOLO [11]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Classification and Segmentation. As demonstrated in Table 3, our method, SAIL, achieves a Top-1 accuracy of \\(84.06\\%\\) and a Top-5 accuracy of \\(97.48\\%\\) on the validation set of ImageNet-1K [20], significantly outperforming state-of-the-art alternatives [11, 21, 22]. In the segmentation task, SAIL also demonstrates superior performance with an mIoU of \\(55.30\\%\\), an mAcc of \\(67.24\\%\\), and an aAcc of \\(84.87\\%\\), illustrated in Table 3. These results indicate that SAIL is effective in both classification and segmentation tasks, offering substantial improvements over existing methods. In Table 4, even when comparing with other state-of-the-art vision backbones, our SAIL manages to achieve remarkable competitive performance with significantly less training data, demonstrating the scaling property of SAIL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Attribute, Relation, and Ordering. To systematically evaluate the ability of SAIL to understand different types of relationships, attributes, and order information, we conduct" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20762" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.321 + ], + "angle": 0, + "content": "
Method#Param#Data#VtokenGeneral VQAHallucinationMath&knowledgeOCR VQA
MMS*MMBenSEEDIMMVMMERWQAPOPEHalluSQA1MathVTQAAI2DOCRB
Modular MLLMs:
InternVL-1.5 [13]2.2B-/-332846.770.969.839.3190257.988.337.384.941.370.569.8654
QwenVL-Chat [3]7B7.2B / 50M25634.560.658.2-184849.3-36.868.235.361.545.9488
LLVA-1.5 [46]7B0.4B+ / 665K57633.164.364.330.5185954.885.927.666.825.546.154.8318
LLVA-1.6 [47]7B0.4B+ / 760K288037.667.464.743.9184257.886.427.670.232.564.966.6532
Cambrian-1 [66]8B10B+ / 7M57650.775.974.7--64.2-30.680.448.171.773.0-
LLVA-OneVision [40]7B10B+ / 3.2M729060.981.774.858.8199865.5--96.656.1-81.6-
Single Transformer-based MLLMs:
Fuyu [5]8B-/--34.410.759.321.4-43.78429.856.830.2-46.8366
Chameleon [64]7B1.4B+ / 1.8M102431.131.130.68.31703919.417.147.222.54.846.07.0
EVE [21]7B33M / 1.8M2304-52.364.625.71628-85.0-64.9-56.861.0398
SOLO [11]8B43.7M / 2M102435.867.764.430.4126044.778.640.473.332.925.061.4126
Mono-InternVL [53]3B1.3B / 7M6400-65.567.440.11875--45.793.645.772.668.6767
Emu3 [73]8B-/-16K46.658.568.237.2-57.485.231.789.231.364.770.0687
EVE2 [22]7B92M / 7.3M2500-66.371.445.0170962.487.6-96.2-71.174.8702
SAIL7B600M / 6M360053.170.172.946.3171963.985.854.293.357.077.176.7783
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.337, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Table 2. Comparison with existing vision-language models on various vision-language benchmarks, including MMS*: MMStar [9]; MMBench-EN [49]; SEED: SEEDBench-IMG [39]; MMV: MMVet [79]; MME [45]; POPE [43]; Hallu: HallusionBench [29]; SQA: ScienceQA-Img [51]; TVQA: TextVQA [62]; MathV: MathVistaMINI [52]; AI2D [38]; RWQA: RealWorldQA [75]; OCRB: OCR-Bench [50]. Note that #A-Param denotes the number of activated parameters; #Data represents the pre-training / fine-tuning data volume; #Vtoken indicates the maximum image patch tokens. For MME, we report the sum of perception and cognition scores. The top two results are highlighted in bold and underline, respectively. All results are derived from those reported in other papers and the official reproduction results from the OpenCompass leaderboard [24]. Our results are obtained from VLMEvalKit [24]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.458, + 0.482, + 0.564 + ], + "angle": 0, + "content": "
MethodClassificationSegmentation
Top-1Top-5mIoUmAccaAcc
EVE [21]42.0365.7727.1235.8972.91
EVE2 [22]44.8669.4140.8553.5379.31
SOLO [11]59.1080.8935.1144.8176.02
SAIL84.0697.4855.3067.2484.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.618 + ], + "angle": 0, + "content": "Table 3. Comparison on image classification and semantic segmentation with other encoder-free approaches. Our SAIL outperforms other alternatives by a large margin." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.638, + 0.476, + 0.737 + ], + "angle": 0, + "content": "
Method#Data#ParamImageNet-1KADE20K
OpenCLIP-H [15]2B0.6B84.4-
OpenCLIP-G [15]2B1.8B86.239.3†
ViT-22B [19]3B22B89.555.3
InternViT [14]6B6B88.258.7
SAIL0.5B7B84.155.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.819 + ], + "angle": 0, + "content": "Table 4. Comparison on image classification and semantic segmentation with other vision backbones. \\(\\dagger\\) indicates training with head tuning using UperNet [76], while others are based on ViT-Adapter [12]. SAIL, with significantly less training data, achieves competitive performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "experiments on the ARO benchmark [80]. As demonstrated in Table 5, SAIL encodes compositional relationships between objects and attributes almost perfectly, significantly surpassing other state-of-the-art vision backbones." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.458, + 0.905, + 0.62 + ], + "angle": 0, + "content": "
MethodRelationAttributeOrder
COCOFlickr30K
OpenCLIP-H [15]49.964.632.640.4
OpenCLIP-G [15]49.965.633.038.3
CLIP-B/32 [59]59.262.948.157.9
CLIP-L/14 [59]61.261.746.856.8
InternViT [14]59.666.073.476.3
NegCLIP [80]81.071.086.091.0
CapPa [68]86.785.798.899.2
SAIL100.099.5100.0100.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.629, + 0.907, + 0.673 + ], + "angle": 0, + "content": "Table 5. Comparison on attribute, relation, and ordering (ARO) with other vision backbones. SAIL almost encodes compositional relationships between objects and attributes perfectly." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.689, + 0.804, + 0.706 + ], + "angle": 0, + "content": "4.3. Properties of Single Transformer" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.693, + 0.728 + ], + "angle": 0, + "content": "4.3.1. Scaling Properties." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Model Scaling: We selected models of different sizes: SAIL-0.5B, SAIL-3B, and SAIL-7B (SAIL by default) for our experiments. Each model underwent Stage 1 pretraining on a mixed multimodal and pure text dataset, encountering 512M image-text pairs. Subsequently, they were finetuned on the LLaVA-mix-665K dataset using the any resolution (anyres) strategy. We evaluated the models based on their performance on vision and language benchmarks after supervised fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The normalized performance of SAIL against model size is plotted in Figure 3. As the model size scales up, we ob" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20763" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.482, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.225, + 0.483, + 0.281 + ], + "angle": 0, + "content": "Figure 3. Model scaling of SAIL. Left: As the model size increases, the training language modeling loss gradually decreases. Right: As the model size increases, performance on downstream VLM tasks progressively improves." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.296, + 0.483, + 0.447 + ], + "angle": 0, + "content": "serve a corresponding enhancement in performance. Additionally, as shown on the left side of Figure 3, the training language modeling loss decreases with increasing model size. This reduction in training loss indicates that larger models have a greater capacity to learn multimodal alignments effectively, enabling them to capture complex relationships between vision and language more accurately. The improved learning capacity directly translates to better performance on downstream VLM tasks, showcasing the benefits of scaling up the Single Transformer architecture." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.75 + ], + "angle": 0, + "content": "Data Scaling: we compared SAIL with its modular MLLM counterpart. For the modular MLLM, we used SigLIP-SO [81] as the vision encoder, and the language model shared the same architecture and initialization parameters as SAIL. Both models were pre-trained using Pretraining stage-1 setting, with SAIL encountering 32M, 128M, and 512M image-text pairs during training, followed by finetuning on the LLaVA-mix-665K dataset. All parameters of both models are trainable. Both models employ an identical number of input tokens for images and text. The normalized performance of both models is plotted in Figure 1(A). The results show that in the low-data regime (32M), SAIL's performance lags behind the modular MLLM, likely due to SigLIP's prior training on 40B samples. However, as the data scales, SAIL exhibits a steeper performance curve, indicating more promising data scaling properties. At 512M image-text pairs, SAIL achieves performance comparable to the modular MLLM in our evaluation subset. This demonstrates the single transformer's superior data scalability, even without a pretrained vision encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.483, + 0.782 + ], + "angle": 0, + "content": "Quantitative results on evaluated benchmark tasks of model scaling and data scaling are tabulated in Supp." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.791, + 0.317, + 0.805 + ], + "angle": 0, + "content": "4.3.2. Information Flow Pattern" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Different attention pattern compared to modular MLLM: since our comparative experiments show that the Single Transformer model exhibits more promising data scaling properties, we conducted an analysis of the trained SAIL model and its modular MLLM counterpart. Specifically, we followed the methodology from FastV [10] to" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.3, + 0.906, + 0.385 + ], + "angle": 0, + "content": "Figure 4. Image Attention Score Allocation: The figure shows the proportion of image attention scores across different transformer layers for Single Transformer-based MLLM and modular MLLM when predicting tokens. Single Transformer-based MLLM generally allocates higher attention weights to image tokens compared to modular MLLM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.906, + 0.531 + ], + "angle": 0, + "content": "analyze the attention score distribution for each predicted token given an image and a user query. This analysis focuses on how much attention is allocated to image tokens during token prediction. We selected 1000 samples from various datasets including VQAv2, GQA, TextVQA, DocVQA, MME, SEEDBench-IMG, MMBench, and some self-curated dialog examples. For each model prediction, we computed the average attention scores assigned to previous image tokens by the output token." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.594 + ], + "angle": 0, + "content": "We conducted a comparative experiment between Single Transformer-based MLLMs (SAIL, SOLO [11], EVE [21]) and modular MLLMs (Qwen2-VL [71], LLaVA-OneVision [40], LLaVA-1.5 [46])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.746 + ], + "angle": 0, + "content": "The results are depicted in Figure 4. Single Transformer-based MLLMs allocate between \\(60\\%\\) and \\(80\\%\\) of attention scores to image tokens across all layers when predicting tokens. In contrast, modular MLLMs such as Qwen2-VL and LLaVA-OneVision allocate only \\(10\\%\\) to \\(30\\%\\) of attention scores to image tokens across different layers. For LLaVA-1.5, which does not update the ViT parameters during supervised fine-tuning (SFT), the image attention score is relatively high in the first two transformer layers but declines sharply in subsequent layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.906, + 0.839 + ], + "angle": 0, + "content": "From this experiment, we can conclude that Single Transformer-based MLLMs tend to allocate a significant portion of attention to previous image tokens during prediction. In contrast, modular MLLMs allocate a smaller portion of their attention directly to image tokens, indicating a less image-centric approach in their prediction mechanism." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "These findings indicate that the Single Transformer model places more emphasis on grounding its predictions in the visual information. As the model undergoes data scaling, it allocates more effective computation to image tokens," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20764" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.177 + ], + "angle": 0, + "content": "
MethodMMBench [49]MME [45]
Physical RelationCelebrity RelationPositionPostersCelebrity
Modular MLLM30.450.598.3134.0100.3
SAIL52.288.9160.0108.275.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.182, + 0.483, + 0.254 + ], + "angle": 0, + "content": "Table 6. Performance Comparison of SAIL and Modular MLLM in MMBench and MME Tasks: the strengths of SAIL in spatial reasoning tasks (MMBench Physical Relation and MME Position split) and its weaknesses in world knowledge tasks (MMBench Celebrity Relation and MME Celebrity and Posters splits)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.476, + 0.281 + ], + "angle": 0, + "content": "thereby enhancing its capability as a vision-centric model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.343 + ], + "angle": 0, + "content": "In summary, the attention pattern analysis underscores the Single Transformer's ability to robustly integrate visual context, enabling it to scale efficiently and potentially outperform modular MLLMs in vision-language tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.351, + 0.387, + 0.366 + ], + "angle": 0, + "content": "4.3.3. Task-Specific Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.345, + 0.384 + ], + "angle": 0, + "content": "We dissect SAIL through case studies:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.49 + ], + "angle": 0, + "content": "Strengths: Spatial Reasoning. SAIL excels at tasks requiring precise spatial location. As shown in Table 6, under the setting of our data scaling experiment, it outperforms the modular counterpart by 61.7 points on the MME Position split and \\(21.8\\%\\) on MMBench Physical Relation questions. The unified architecture likely enables tighter coupling between visual geometry and linguistic descriptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Weaknesses: World Knowledge. Conversely, SAIL falls short in tasks that demand extensive world knowledge. As shown in Table 6, SAIL underperforms in the MME celebrity and art splits compared to the modular MLLM. This underperformance can be attributed to SAIL's lack of diverse domain-specific data during pretraining, a gap that was not sufficiently addressed during supervised finetuning. Modular MLLMs, with their pretrained vision encoders like CLIP [15, 59] or SigLIP [81], have a broader knowledge base and therefore handle such tasks more effectively. We hypothesize that scaling up SAIL's pretraining data diversity could help bridge this gap, enhancing its performance on knowledge-intensive tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.442, + 0.714 + ], + "angle": 0, + "content": "4.4. Empirical Observations on Basic Factors" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.809 + ], + "angle": 0, + "content": "To guide scalable training of single-transformer MLLMs, we conduct ablation studies on two critical design choices using SAIL-0.5B pretrained on 128M image-text pairs and fine-tuned on LLaVA-mix-665K. Performance is evaluated through zero-shot image classification after pretraining [67] and vision-language benchmarks after SFT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Bidirectional Attention for Image Patches with Multimodal Position Encoding. We compare two approaches for integrating image patches into the transformer: (1) Causal attention with 1D positional encoding, using a token to demarcate image rows. (2) Full bidirectional attention for image patches paired with mul" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.09, + 0.903, + 0.155 + ], + "angle": 0, + "content": "
Exp. SettingVQAv2GQASQATQASEED-I
Default59.146.959.620.135.1
#1 No Img full attn57.845.258.716.233.8
#2 No pure text in PT56.342.148.618.332.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.16, + 0.907, + 0.245 + ], + "angle": 0, + "content": "Table 7. Ablation Study on Basic Factors for SAIL: We presents the impact of different ablation settings on the performance of SAIL across VQAv2 [27], GQA [35], SQA [51], TQA [62], and SEED-I [39]. The default setting includes image full attention and including pure text data in pretraining. Ablation #1 removes image full attention, and ablation #2 excludes pure text in pretraining." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.25, + 0.907, + 0.552 + ], + "angle": 0, + "content": "timodal rotary position embeddings (RoPE), which jointly encode spatial coordinates (e.g., 2D grid positions) and text token positions. As shown in Table 7, using bidirectional attention with multimodal RoPE significantly improves performance on vision-language tasks, with a particularly notable gain of \\(3.1\\%\\) on TextVQA. This suggests that enabling cross-patch interactions during pretraining enhances visual representation learning and tightens cross-modal alignment. Interleaving Pure Text Data During Pretraining. We analyze the impact of mixing SlimPajama text data with image-text pairs during pretraining. The results, as presented in Table 7 #2, reveal that mixing in pure text data consistently improves performance across vision and language benchmarks. This finding underscores the importance of preserving language capabilities in the LLM when training Single Transformer models, as maintaining strong language skills is crucial for building a multimodal model capable of complex reasoning. Currently, incorporating text data in training is one of the effective methods to maintain the language abilities of the model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.553, + 0.906, + 0.629 + ], + "angle": 0, + "content": "Our ablation studies identify key designs for training SAIL effectively. Using bi-directional attention with multimodal rotary position embeddings enhances visual perception, while incorporating pure text data preserves language capabilities for robust multimodal performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.634, + 0.65 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In this work, we conducted an extensive analysis of Single Transformer-based MLLMs compared to modular MLLMs. Our investigation explored the unique properties of Single Transformers, including scalability, cross-modal information flow patterns, and visual representation capabilities. A series of experiments on our trained SAIL model demonstrated that this unified architecture achieves performance on vision-language benchmarks comparable to modular MLLMs while also functioning effectively as a vision backbone. Our findings highlight several advantages of Single Transformer architectures, such as superior data scalability, vision-centric information flow, and inherent capabilities as a powerful vision encoder. We hope our empirical findings will inspire further research to refine and enhance Single Transformer architectures, advancing the field of multimodal intelligence." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20765" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.2, + 0.483, + 0.226 + ], + "angle": 0, + "content": "[2] AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.483, + 0.295 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.298, + 0.482, + 0.339 + ], + "angle": 0, + "content": "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Taşirlar. Introducing our multimodal models, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.484, + 0.451 + ], + "angle": 0, + "content": "[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.484, + 0.535 + ], + "angle": 0, + "content": "[7] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.484, + 0.606 + ], + "angle": 0, + "content": "[8] Jieneng Chen, Qihang Yu, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Vitamin: Designing scalable vision models in the vision-language era. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[11] Yangyi Chen, Xingyao Wang, Hao Peng, and Heng Ji. A single transformer for scalable vision-language modeling. Transactions on Machine Learning Research, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[14] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[15] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[16] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[17] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[18] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[19] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdul-mohsin, et al. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[20] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[21] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. In Advances in Neural Information Processing Systems, pages 52545-52567. Curran Associates, Inc., 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[22] Haiwen Diao, Xiaotong Li, Yufeng Cui, Yueze Wang, Haoge Deng, Ting Pan, Wenxuan Wang, Huchuan Lu, and Xinlong Wang. *Evev2: Improved baselines for encoder-free vision-language models.* arXiv preprint arXiv:2502.06788, 2025. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20766" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[24] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11198-11201, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[25] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19358-19369, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[26] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 36:27092-27112, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[27] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[28] Shuhao Gu, Jialing Zhang, Siyuan Zhou, Kevin Yu, Zhaohu Xing, Liangdong Wang, Zhou Cao, Jintao Jia, Zhuoyi Zhang, Yixuan Wang, et al. Infinity-mm: Scaling multimodal performance with large-scale and high-quality instruction data. arXiv preprint arXiv:2410.18558, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[29] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[30] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[31] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[32] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. arXiv preprint arXiv:2312.08914, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[33] Xinyu Huang, Youcai Zhang, Jinyu Ma, Weiwei Tian, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Lei Zhang. Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[34] Zilong Huang, Qinghao Ye, Bingyi Kang, Jiashi Feng, and Haoqi Fan. Classification done right for vision-language pretraining. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[36] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[37] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[38] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[39] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[40] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[41] Xianhang Li, Zeyu Wang, and Cihang Xie. An inverse scaling law for clip training. In NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[42] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, Yuyin Zhou, and Cihang Xie. What if we recapture billions of web images with llama-3? arXiv preprint arXiv:2406.08478, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[43] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[44] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23390-23400, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[45] Zijing Liang, Yanjie Xu, Yifan Hong, Penghui Shang, Qi Wang, Qiang Fu, and Ke Liu. A survey of multimodel large language models. In Proceedings of the 3rd International Conference on Computer, Artificial Intelligence and Control Engineering, pages 405-409, 2024. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[46] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[47] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.3,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20767" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[50] Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xu-Cheng Yin, Cheng-Lin Liu, Lianwen Jin, and Xiang Bai. Ocrbench: on the hidden mystery ofOCR in large multimodal models. Science China Information Sciences, 67(12):220102, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[51] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[52] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[53] Gen Luo, Xue Yang, Wenhan Dou, Zhaokai Wang, Jiawen Liu, Jifeng Dai, Yu Qiao, and Xizhou Zhu. Mono-internvl: Pushing the boundaries of monolithic multimodal large language models with endogenous visual pre-training. arXiv preprint arXiv:2410.08202, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.483, + 0.515 + ], + "angle": 0, + "content": "[54] Sachin Mehta, Maxwell Horton, Fartash Faghri, Mohammad Hossein Sekhavat, Mahyar Najibi, Mehrdad Farajtabar, Oncel Tuzel, and Mohammad Rastegari. Catlip: Cliplevel visual recognition accuracy with 2.7 x faster pretraining on web-scale image-text data. arXiv preprint arXiv:2404.15653, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[55] Meta. Introducing meta llama 3: The most capable openly available llm to date, 2024. Accessed: 2024-04-18. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.428, + 0.56 + ], + "angle": 0, + "content": "[56] OpenAI. Introducing chatgpt. OpenAI Blog, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.368, + 0.575 + ], + "angle": 0, + "content": "[57] OpenAI. Gpt-4 technical report, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.406, + 0.59 + ], + "angle": 0, + "content": "[58] OpenAI. Gpt-4v(ision) system card, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.592, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[60] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[61] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatronlm: Training multi-billion parameter language models using model parallelism. arXiv preprint arXiv:1909.08053, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[62] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[63] Daria Soboleva, Faisal Al-Khateeb, Robert Myers, Jacob R Steeves, Joel Hestness, and Nolan Dey. SlimPajama: A 627B token cleaned and deduplicated version of RedPajama. https://cerebras.ai/blog/slimpajama-a-627b-token-cleaned-and-deduplicated-version-of-redpajama, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[64] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.906, + 0.286 + ], + "angle": 0, + "content": "[65] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[66] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.426 + ], + "angle": 0, + "content": "[67] Michael Tschannen, Manoj Kumar, Andreas Steiner, Xiaohua Zhai, Neil Houlsby, and Lucas Beyer. Image captioners are scalable vision learners too. Advances in Neural Information Processing Systems, 36:46830-46855, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[68] Michael Tschannen, Manoj Kumar, Andreas Steiner, Xiaohua Zhai, Neil Houlsby, and Lucas Beyer. Image captioners are scalable vision learners too. Advances in Neural Information Processing Systems, 36:46830-46855, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[69] Haochen Wang, Anlin Zheng, Yucheng Zhao, Tiancai Wang, Ge Zheng, Xiangyu Zhang, and Zhaoxiang Zhang. Reconstructive visual instruction tuning. In International Conference on Learning Representations, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.906, + 0.621 + ], + "angle": 0, + "content": "[70] Jiacong Wang, Bohong Wu, Haiyong Jiang, Zhou Xun, Xin Xiao, Haoyuan Guo, and Jun Xiao. World to code: Multimodal data generation via self-instructed compositional captioning and filtering. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4608-4623, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.906, + 0.705 + ], + "angle": 0, + "content": "[71] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.906, + 0.747 + ], + "angle": 0, + "content": "[72] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion feedback helps clip see better. arXiv preprint arXiv:2407.20171, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[73] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[74] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[75] X.ai. Grok-1.5 vision preview. https://x.ai/blog/grok-1.5v, 2024.5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20768" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[76] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[77] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[78] Qiying Yu, Quan Sun, Xiaosong Zhang, Yufeng Cui, Fan Zhang, Yue Cao, Xinlong Wang, and Jingjing Liu. Capsfusion: Rethinking image-text data at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14022-14032, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[79] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[80] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[81] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.529 + ], + "angle": 0, + "content": "[82] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[83] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[84] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "20769" + } + ] +] \ No newline at end of file diff --git a/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/2b8a94f1-28e8-4218-82c7-da4dba2c45e1_origin.pdf b/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/2b8a94f1-28e8-4218-82c7-da4dba2c45e1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5a1e75d1b5d81ab3521ff76f5b54207f625bc440 --- /dev/null +++ b/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/2b8a94f1-28e8-4218-82c7-da4dba2c45e1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5ab5e6d419addc22c968213ed36ccfa1c08023cb66e37aa84e3454942a325276 +size 1010010 diff --git a/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/full.md b/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dbd07e9302ff2408b9d9ba094ebd8f85cda65370 --- /dev/null +++ b/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/full.md @@ -0,0 +1,322 @@ +# The Scalability of Simplicity: Empirical Analysis of Vision-Language Learning with a Single Transformer + +Weixian Lei* Jiacong Wang* Haochen Wang* +Xiangtai Li Jun Hao Liew Jiashi Feng Zilong Huang† +*Equal contribution, † Corresponding author +Bytedance Seed + +# Abstract + +This paper introduces SAIL, a single transformer unified multimodal large language model (MLLM) that integrates raw pixel encoding and language decoding within a singular architecture. Unlike existing modular MLLMs, which rely on a pre-trained vision transformer (ViT), SAIL eliminates the need for a separate vision encoder, presenting a more minimalist architecture design. Instead of introducing novel architectural components, SAIL adapts mix-attention mechanisms and multimodal rotary position embedding to better align with the distinct characteristics of visual and textual modalities. We systematically compare SAIL's properties—including scalability, cross-modal information flow patterns, and visual representation capabilities—with those of modular MLLMs. By scaling both training data and model size, SAIL achieves performance comparable to modular MLLMs. Notably, the removal of pretrained ViT components enhances SAIL's scalability and results in significantly different cross-modal information flow patterns. Moreover, SAIL demonstrates strong visual representation capabilities, achieving results on par with ViT22B in vision tasks such as semantic segmentation. Code and models are available at https://github.com/bytedance/SAIL. + +# 1. Introduction + +The pursuit of multimodal intelligence has driven the development of Multimodal Large Language Models (MLLMs) [48, 58, 65], which typically adopt a modular design: a pre-trained vision encoder (e.g., CLIPViT [15, 59]) extracts image features, a Large Language Model (LLM) [2, 16, 37, 55-57] processes text, and a lightweight projector aligns the two modalities. This framework achieves strong performance through multi-stage pretraining, supervised fine-tuning (SFT), and post-training on multimodal datasets [3, 18, 48, 70, 71, 84]. While effective, + +![](images/5b3550ba22a6bb0225a56ba63ea2b9e2fb22e444614f6b4e1100358509dc9c52.jpg) +(A) +Figure 1. (A) Data scaling curve for Modular Multimodal Large Language Model (MLLM) and SAIL, our Single Transformer-based MLLM. As pretraining data increases, the single transformer SAIL shows a sharper performance gain, demonstrating its superior data scalability. Experiment settings are described in Section 4.3. (B) Comparison to existing Single Transformer-based MLLMs: our SAIL pushes the performance boundaries on both vision tasks and vision-language tasks. + +![](images/4aaec2709c4af572be5932a3c719e9c28bd183f82a5bf5c8f87e81fb1ce455b9.jpg) +(B) + +this modular MLLM paradigm inherently fragments multimodal processing, reinforces reliance on pretrained visual encoders, which may limit deployment flexibility and scalability [11, 21, 53]. + +A promising alternative is to eliminate the visual encoder entirely and process raw image patches and text tokens within a single Transformer. This unified architecture removes modality-specific modules, enabling parameter sharing and end-to-end learning of vision-language interactions. Previous works [11, 21, 53] have primarily explored the architecture design, training data, and methods of Single Transformer-based MLLMs. However, little exploration has been given to their fundamental properties, such as scalability, cross-modal information flow patterns, and visual representation capabilities. A deeper understanding of these properties is crucial for unlocking the full potential of Single Transformer-based MLLMs. + +In this work, we present an experimental analysis of the fundamental properties of Single Transformer-based + +MLLMs and compare them to modular MLLMs (e.g. LLaVA [48]). Additionally, in the absence of a pre-trained visual encoder, Single Transformers have to learn visual representations from scratch. Thus, an intriguing question arises: can a trained Single Transformer emerge as a strong vision encoder? + +We conduct a series of experiments to train and study our Single trAnformer model for vIsion and Language (SAIL). While we do not propose novel architecture designs, we introduce necessary modifications to enable the model to process different modalities in a unified architecture. In its micro architecture design, we address the different spatial characteristics of 2D images and 1D text data by employing a mixed attention mechanism: bidirectional attention for image patches and causal attention for text tokens, combined with multimodal rotary position embedding. Through model and data scaling, SAIL achieves performance on vision-language benchmarks comparable to modular MLLMs, while also functioning as a high-performing vision backbone, as shown in Figure 1. + +More concretely, our empirical analysis uncovers following striking advantages of Single Transformer architectures: + +(i) Superior Data Scaling: In controlled experiments, SAIL exhibits steeper performance gains as pretraining data scales. While LLaVA-style modular MLLMs initially perform well, our model's performance becomes very close to theirs when pretrained on 512M samples, as shown in Figure 1(A). This suggests that unified architectures can effectively leverage large-scale data and potentially match the performance of modular MLLMs. + +(ii) Vision Centric Information Flow Pattern: Through analysis of attention distributions, we observe that Single Transformers assign significantly higher attention scores to image tokens during token prediction compared to modular MLLMs. This indicates that the information flow in Single Transformer MLLMs is more direct, with visual tokens influencing prediction tokens more prominently, highlighting a vision-centric approach to decision-making. + +(iii) Vision Encoder Functioning: Our experiments further demonstrate that the pretrained Single Transformer inherently serves as a powerful vision encoder. Comprehensive evaluations on vision-centric tasks, such as image classification and semantic segmentation, show that the model learns rich visual representations during multimodal pretraining. These representations enhance its capacity for both semantic-level comprehension (e.g., object categorization) and pixel-level understanding (e.g., fine-grained segmentation masks), bridging high-level abstraction and low-level visual reasoning within a unified architecture. + +In summary, our findings indicate that Single Transformer-based MLLMs hold great promise in surpassing modular MLLMs in terms of leveraging large-scale data, forming direct vision-centric information pathways, + +and functioning as effective vision encoders. We hope our empirical findings inspire further research to refine and enhance Single Transformer architecture, ultimately driving advancements in multimodal intelligence from a new perspective. + +# 2. Related Work + +# 2.1. Paradigms in Vision-Language Model Design + +Modular MLLMs with Visual Encoders. The prevailing approach in MLLM design employs modular architectures [3, 48, 58, 65, 71] that rely on pretrained vision encoders (e.g., CLIP-ViT [15, 59], InternViT [14]) to process visual inputs. The visual features extracted from these frozen encoders are then aligned with LLM input spaces via linear [7, 48, 69, 84] or cross-attention layers [1, 32]. While this module design enables effective transfer of pretrained visual-language knowledge, it also introduces several limitations. First, incorporating a separate ViT encoder significantly slows down both training and inference, increasing deployment complexity and requiring costly infrastructure—especially when compared to a single transformer unified model. Second, common strategies for integrating visual features, such as direct mapping into LLM inputs [7, 48, 84] or sharing them across LLM layers [1, 32], often struggle to reconcile the inherent differences between images and text representations. Finally, as model scale, balancing the interactions between the encoder, LLM, and alignment layers becomes increasingly challenging [11, 53]. Thus, in this work, we explore a single transformer-based MLLM architecture that eliminates the ViT encoder and alignment components to overcome these challenges. + +Single Transformer-based MLLMs Without Visual Encoders. Emerging research explores end-to-end architectures that process raw image patches and text tokens through a single Transformer, bypassing visual encoders entirely. These monolithic designs fall into two categories: continuous tokenization and discrete tokenization. Continuous tokenization, exemplified by Fuyu-8B [5] and SOLO [11], directly maps patches to LLM embeddings via linear projections, enabling flexible resolution handling but requiring massive pretraining data. Discrete tokenization, adopted by Chameleon [64] and Emu3 [73], employs VQ-VAE tokenizers to compress images into discrete tokens, trading pixel-level fidelity for generation capabilities. While later efforts such as EVE [21] and MonoInternVL [53] demonstrate the feasibility of encoder-free training, critical gaps remain: (1) Existing methods rely on extra designs and auxiliary loss [21], complicating training pipelines; (2) The scaling laws and fundamental properties of purely end-to-end trained models remain poorly understood; (3) Vision-language interaction in shared parame + +ter spaces lacks systematic analysis—most prior MLLMs default to causal attention for processing image-text sequences. In this work, we reveal that enabling bidirectional attention between image patches significantly enhances visual representation learning, addressing a key limitation in current designs. More importantly, our study bridges these gaps by establishing foundational principles for training scalable, self-contained single-transformer MLLMs. + +# 2.2. Vision Representation Learning + +Learning effective vision representations is a core challenge in computer vision research, with extensive works [4, 19, 20, 25, 30, 74] dedicated to this problem. With the proliferation of large-scale web-sourced image-text datasets [6, 26, 60], recent methods leverage this data to train deep visual representations via three primary paradigms: + +Text as Classification Labels. Early methods used textual descriptions as weak supervision by extracting categorical labels from captions. For example, frameworks like Tag2Text [33] and RAM [82] used ViTs [23] to predict noun-based pseudo-labels from datasets like CC12M [6]. CatLIP [54] scaled labels to millions using object-centric supervision, and SuperClass [34] directly used tokenized text tokens as classification categories. + +Image-Text Contrastive Learning Contrastive pretraining, as exemplified by CLIP [15, 59] and ALIGN [36], aligns global image-text embeddings within a shared latent space. Subsequent works [8, 14, 41, 44, 72, 81] focused on enhancing performance and improving training efficiency. + +Text as Autoregressive Targets. Caption generation as a pretext task is another approach for visual representation learning. SimVLM [74] trains encoder-decoder architectures to autoregressively predict captions, while CapPa [67] trains vision encoders through sequence prediction. These methods often retain modular designs or auxiliary components like contrastive losses [77]. Our work aligns with this category but removes architectural fragmentation by jointly modeling image patches and text tokens in a single Transformer. We find that the pre-trained Single Transformer learns transferable vision representations, enabling it to handle downstream multimodal understanding tasks and function as a vision encoder without modifications. + +# 3. SAIL: Training a Single Transformer for Vision and Language + +# 3.1. Model Architecture + +SAIL is built upon a unified Transformer architecture (Figure 2(A)) that processes multimodal inputs through streamlined, modality-specific preprocessing. For text, raw input is tokenized using the language model's tokenizer and then transformed into embeddings via the textual embedding module. For images, we partition the input into fixed- + +size patches and project them into continuous embeddings via a linear projection. Additionally, we maintain a list of special tokens explicitly designed for visual modality encoding: and tokens mark the beginning and end of an image patch span, respectively. In multimodal scenarios, such as image-text pairs, these embeddings are concatenated into a single sequence and fed into the Transformer, enabling joint cross-modal interactions through unified self-attention layers. This design eliminates the need for modality-specific encoders, which efficiently processess heterogeneous data within a Single Transformer framework. + +Bidirectional attention within image patches. While existing multimodal large language models (MLLMs) [40, 47, 48, 84] predominantly adopt causal attention for autoregressive sequence modeling, our experiments reveal that enabling full bidirectional attention among tokens from the same image significantly enhances visual representation learning and boosts downstream vision-language task performance. Note that previous Single Transformers [11, 21, 53, 73] have only utilized causal attention, without exploring the potential of mixed attention mechanisms. + +As illustrated in Figure 2(B), for SAIL we implement a mixed attention scheme: (1) For text tokens, we preserve causal attention to maintain autoregressive generation capabilities, allowing each token to attend only to its predecessors. (2) For image tokens, we activate full bidirectional attention within each image patch group, empowering every visual token to interact with all others in the same image. This design captures holistic spatial relationships and contextual dependencies among visual elements, addressing the under-explored potential of attention mechanisms in cross-modal alignment. The improved interaction paradigm not only refines vision-language feature fusion but also provides stronger inductive biases for complex reasoning tasks. + +Multimodal Rotary Position Embeddings. Following [71], we implement Multimodal RoPE (M-RoPE) in SAIL to harmonize positional modeling for multimodal inputs. The method decomposes positional encoding into two axes: height, and width. For text tokens, all axes share uniform position IDs (aligned with 1D-RoPE), whereas for images, height/width IDs adaptively map to token coordinates, as is shown in Figure 2(C). Notably, position indexing is sequentially initialized across modalities (e.g., starting from images before extending to subsequent text), preserving inter-modal consistency. M-RoPE not only improves positional sensitivity but also constrains absolute position magnitudes for visual tokens, facilitating robust generalization to extended sequences in inference. + +# 3.2. Pretraining + +We apply a two-stage curriculum to progressively strengthen the visual perception of SAIL while preserving + +![](images/7e4de715e61dd3d9bcd8398615d081679a59a412f7a74115d4387dad3353f861.jpg) +(A) Model Architecture + +![](images/f386a7e5da052dd8a7f2235a610b1e87504e6c81edfbb8fae052f42dfc7eeac1.jpg) +(C) Multimodal RoPE for SAIL +Figure 2. Model architecture and micro-designs for SAIL. (A) Model Architecture: SAIL is a unified transformer that processes both images and texts without extra module designs. (B) Mixed Attention Mechanism: we adopt bidirectional attention for image patches from the same image and causal attention for text tokens. Examples for a multimodal sequence and a text sequence are provided. Colored squares represent "allow to attend" and white squares indicate "prevent from attending". (C) Multimodal RoPE: an illustration of the multimodal rotary position embedding for SAIL, with examples for a multimodal sequence and a text sequence. + +its inherent language capabilities. + +
StageDatasetImg.ResNum
Pretraining S1Recap-DataComp-1B [42]224x224512M
SlimPajama [63]-
Pretraining S2Capfusion [78]AnyRes60M
OCR from LAION COCO [60]7M
InifinityMM Stage 2 subset [28]19M
SlimPajama [63]-
SFTInifinityMM Stage3 [28]AnyRes6M
+ +Table 1. Details of training datasets used across all stages. "Img.Res" refers to the image resolution settings applied during each training stage. All datasets listed are publicly available. Note that these settings represent the default configuration for standard SAIL training, while separate settings are used for scaling experiments and ablation studies. + +Stage 1: Accelerated Visual Knowledge Acquisition. In this stage, we pretrain SAIL on large-scale image-text pairs to rapidly bootstrap its visual understanding. To maximize data throughput, we uniformly resize all images to a lower resolution (e.g., $224 \times 224$ ), reducing multimodal sequence lengths and enabling the model to process more samples within fixed training time. To prevent catastrophic forgetting of linguistic knowledge, we interleave pure text corpora with multimodal data during training. This hybrid approach ensures efficient exposure to visual patterns while maintaining robust language proficiency. + +Stage 2: Enhancing Any-Resolution Image Understanding. Real-world applications require robustness to images + +of varying resolutions and aspect ratios, such as documents, charts, or infographics. Following prior works [11, 21], we extend pretraining with an any-resolution strategy: images retain their native resolutions during processing, and positional embeddings adapt dynamically to arbitrary spatial dimensions. This stage further refines SAIL's ability to model fine-grained visual details (e.g., tabular structures, text-rich graphics) while continuing to incorporate text-only data for language capability preservation. + +Pretraining Objective. Throughout both stages, we optimize the standard language modeling loss only on text tokens. Image patches and special visual tokens are excluded from loss computation. + +# 3.3. Supervised Fine-tuning + +During the Supervised Fine-tuning (SFT) stage, we train SAIL on publicly available, multi-source instruction datasets to enhance its understanding of complex linguistic instructions and diverse dialogue patterns critical for real-world deployment. This phase fine-tunes the entire network architecture, focusing on aligning the model's responses with human intent through exposure to varied instructional formats and multimodal interactions. + +Table 1 shows the details of training datasets for pretraining and supervised fine-tuning (SFT) across all stages. During Stage 1 pretraining, we utilized mixed multimodal and pure text datasets including Recap-DataComp-1B [42] and SlimPajama [63], with images at a resolution of $224 \times 224$ , totaling 512M image-text pairs. In Stage 2, the pretrain + +ing datasets include Capfusion [78], self-curated OCR data from LAION COCO [60], InfinityMM Stage 2 subset [28], and SlimPajama, utilizing the any resolution (AnyRes) strategy, with a combined total of 86M image-text pairs along with text data. The SFT stage employed the InfinityMM Stage 3 subset, processed at any resolution, containing 6M image-text pairs. + +# 4. Experiment + +# 4.1. Experimental Settings + +Evaluation Benchmarks. For evaluation of vision and language tasks, we evaluate SAIL and existing MLLMs on a broad range of multimodal benchmarks. Specifically, MLLM benchmarks encompass MMBench-EN [49], SEEDBench-IMG [39], MMVet [79], MME [45], HallusionBench [29], MathVistaMINI [52], and OCR-Bench [50]. Visual question answering benchmarks include TextVQA [62], ScienceQA-IMG [51], AI2D [38], MMStar [9], RealWorldQA [75]. + +For evaluation of vision representation learning, we conduct experiments on ImageNet-1K [20] for image classification, ADE20K for semantic segmentation [83], and ARO [80] for attribute, relation, and ordering. + +Implementation Details. For pretraining, we initialize SAIL from the Mistral-7B-v0.1 base LLM and set the patch size to 14. We modify Megatron [61] to support SAIL's multimodal input. Pretraining uses 128 NVIDIA A100 80G GPUs with 2-way tensor parallelism and 64-way data parallelism. The learning rate is set at 5e-5 and decays cosinely to a minimum of 5e-6. For training efficiency, we concatenate sequences from different data samples into one long sequence of 32,768 tokens, adjusting the attention mask to ensure that tokens from different samples do not attend to each other. We use a round-robin approach to interleave image-text packed sequences and pure text packed sequences, configuring the global batch to contain approximately 16K image-text pairs. + +For Supervised Fine-Tuning (SFT), the global batch size is set to 512. Training is performed for one epoch with a maximum learning rate of 1e-5, following a linear warm-up phase and then transitioning to a cosine decay schedule. + +For vision, we load the checkpoint after Stage 1 pretraining and keep it frozen for downstream evaluations, including (1) image classification on ImageNet-1K [20], (2) semantic segmentation on ADE20K [83], and (3) attribute, relation and, ordering on the ARO benchmark [80]. Specifically, (1) for image classification, we utilize a linear classifier with 90 epochs for linear probing, where detailed configurations are mostly obtained from common practices [31]. Images are resized to $224 \times 224$ and the global batch size is 8,192 across 8 A100 (80G) GPUs. (2) For semantic segmentation, we adopt ViT-Adapter [12] with UperNet [76] as the segmenta + +tion decoder. The implementation is based on MMSegmentation [17] with 80k training iterations. The input resolution is $512 \times 512$ and the global batch size is 16 across 8 A100 (80G) GPUs. (3) For attribute, relation and, ordering, we regard the negative of the caption loss over each image-text pair as the similarity metric for retrieval. + +# 4.2. Experimental Results + +# 4.2.1. Results on Vision Language Tasks + +As shown in Table 2, we compare SAIL against existing MLLMs across 13 vision-language benchmarks. SAIL consistently outperforms other Single Transformer-based models like Fuyu [5], EVE [21], SOLO [11], MonoInternVL [53], and EVE2 [22] across diverse vision-language tasks. This demonstrates that SAIL can achieve significant performance gains and push the boundaries of Single Transformer-based MLLMs without needing extra component designs or auxiliary training losses. Moreover, when compared to methods employing discrete vision tokens (e.g., Chameleon and Emu3), SAIL demonstrates superior performance. These results validate that scaling up single-transformer pretraining effectively enhances cross-modal alignment between images and text. Compared to the state-of-the-art modular MLLM LLaVA-OneVision [40], SAIL achieves comparable performance on some benchmarks, such as MMStar, SEEDBench-IMG, and RealWorldQA. While the performance of Single Transformer-based MLLMs still lags behind modular MLLMs in certain areas, we hypothesize that scaling the pretraining data volume or incorporating higher-quality instruction-tuning data will bridge the remaining performance gap. + +# 4.2.2. Results on Vision Representation Learning + +In this section, we compare the quality of learned visual representations of our SAIL with other Single Transformers, including EVE [21], EVE2 [22], and SOLO [11]. + +Classification and Segmentation. As demonstrated in Table 3, our method, SAIL, achieves a Top-1 accuracy of $84.06\%$ and a Top-5 accuracy of $97.48\%$ on the validation set of ImageNet-1K [20], significantly outperforming state-of-the-art alternatives [11, 21, 22]. In the segmentation task, SAIL also demonstrates superior performance with an mIoU of $55.30\%$ , an mAcc of $67.24\%$ , and an aAcc of $84.87\%$ , illustrated in Table 3. These results indicate that SAIL is effective in both classification and segmentation tasks, offering substantial improvements over existing methods. In Table 4, even when comparing with other state-of-the-art vision backbones, our SAIL manages to achieve remarkable competitive performance with significantly less training data, demonstrating the scaling property of SAIL. + +Attribute, Relation, and Ordering. To systematically evaluate the ability of SAIL to understand different types of relationships, attributes, and order information, we conduct + +
Method#Param#Data#VtokenGeneral VQAHallucinationMath&knowledgeOCR VQA
MMS*MMBenSEEDIMMVMMERWQAPOPEHalluSQA1MathVTQAAI2DOCRB
Modular MLLMs:
InternVL-1.5 [13]2.2B-/-332846.770.969.839.3190257.988.337.384.941.370.569.8654
QwenVL-Chat [3]7B7.2B / 50M25634.560.658.2-184849.3-36.868.235.361.545.9488
LLVA-1.5 [46]7B0.4B+ / 665K57633.164.364.330.5185954.885.927.666.825.546.154.8318
LLVA-1.6 [47]7B0.4B+ / 760K288037.667.464.743.9184257.886.427.670.232.564.966.6532
Cambrian-1 [66]8B10B+ / 7M57650.775.974.7--64.2-30.680.448.171.773.0-
LLVA-OneVision [40]7B10B+ / 3.2M729060.981.774.858.8199865.5--96.656.1-81.6-
Single Transformer-based MLLMs:
Fuyu [5]8B-/--34.410.759.321.4-43.78429.856.830.2-46.8366
Chameleon [64]7B1.4B+ / 1.8M102431.131.130.68.31703919.417.147.222.54.846.07.0
EVE [21]7B33M / 1.8M2304-52.364.625.71628-85.0-64.9-56.861.0398
SOLO [11]8B43.7M / 2M102435.867.764.430.4126044.778.640.473.332.925.061.4126
Mono-InternVL [53]3B1.3B / 7M6400-65.567.440.11875--45.793.645.772.668.6767
Emu3 [73]8B-/-16K46.658.568.237.2-57.485.231.789.231.364.770.0687
EVE2 [22]7B92M / 7.3M2500-66.371.445.0170962.487.6-96.2-71.174.8702
SAIL7B600M / 6M360053.170.172.946.3171963.985.854.293.357.077.176.7783
+ +Table 2. Comparison with existing vision-language models on various vision-language benchmarks, including MMS*: MMStar [9]; MMBench-EN [49]; SEED: SEEDBench-IMG [39]; MMV: MMVet [79]; MME [45]; POPE [43]; Hallu: HallusionBench [29]; SQA: ScienceQA-Img [51]; TVQA: TextVQA [62]; MathV: MathVistaMINI [52]; AI2D [38]; RWQA: RealWorldQA [75]; OCRB: OCR-Bench [50]. Note that #A-Param denotes the number of activated parameters; #Data represents the pre-training / fine-tuning data volume; #Vtoken indicates the maximum image patch tokens. For MME, we report the sum of perception and cognition scores. The top two results are highlighted in bold and underline, respectively. All results are derived from those reported in other papers and the official reproduction results from the OpenCompass leaderboard [24]. Our results are obtained from VLMEvalKit [24]. + +
MethodClassificationSegmentation
Top-1Top-5mIoUmAccaAcc
EVE [21]42.0365.7727.1235.8972.91
EVE2 [22]44.8669.4140.8553.5379.31
SOLO [11]59.1080.8935.1144.8176.02
SAIL84.0697.4855.3067.2484.87
+ +Table 3. Comparison on image classification and semantic segmentation with other encoder-free approaches. Our SAIL outperforms other alternatives by a large margin. + +
Method#Data#ParamImageNet-1KADE20K
OpenCLIP-H [15]2B0.6B84.4-
OpenCLIP-G [15]2B1.8B86.239.3†
ViT-22B [19]3B22B89.555.3
InternViT [14]6B6B88.258.7
SAIL0.5B7B84.155.3
+ +experiments on the ARO benchmark [80]. As demonstrated in Table 5, SAIL encodes compositional relationships between objects and attributes almost perfectly, significantly surpassing other state-of-the-art vision backbones. + +Table 4. Comparison on image classification and semantic segmentation with other vision backbones. $\dagger$ indicates training with head tuning using UperNet [76], while others are based on ViT-Adapter [12]. SAIL, with significantly less training data, achieves competitive performance. + +
MethodRelationAttributeOrder
COCOFlickr30K
OpenCLIP-H [15]49.964.632.640.4
OpenCLIP-G [15]49.965.633.038.3
CLIP-B/32 [59]59.262.948.157.9
CLIP-L/14 [59]61.261.746.856.8
InternViT [14]59.666.073.476.3
NegCLIP [80]81.071.086.091.0
CapPa [68]86.785.798.899.2
SAIL100.099.5100.0100.0
+ +Table 5. Comparison on attribute, relation, and ordering (ARO) with other vision backbones. SAIL almost encodes compositional relationships between objects and attributes perfectly. + +# 4.3. Properties of Single Transformer + +# 4.3.1. Scaling Properties. + +Model Scaling: We selected models of different sizes: SAIL-0.5B, SAIL-3B, and SAIL-7B (SAIL by default) for our experiments. Each model underwent Stage 1 pretraining on a mixed multimodal and pure text dataset, encountering 512M image-text pairs. Subsequently, they were finetuned on the LLaVA-mix-665K dataset using the any resolution (anyres) strategy. We evaluated the models based on their performance on vision and language benchmarks after supervised fine-tuning. + +The normalized performance of SAIL against model size is plotted in Figure 3. As the model size scales up, we ob + +![](images/9f63de2d668ebe946af061ff2b559117f9fbab47d3ccc1f41eb89bcd60f9949a.jpg) +Figure 3. Model scaling of SAIL. Left: As the model size increases, the training language modeling loss gradually decreases. Right: As the model size increases, performance on downstream VLM tasks progressively improves. + +serve a corresponding enhancement in performance. Additionally, as shown on the left side of Figure 3, the training language modeling loss decreases with increasing model size. This reduction in training loss indicates that larger models have a greater capacity to learn multimodal alignments effectively, enabling them to capture complex relationships between vision and language more accurately. The improved learning capacity directly translates to better performance on downstream VLM tasks, showcasing the benefits of scaling up the Single Transformer architecture. + +Data Scaling: we compared SAIL with its modular MLLM counterpart. For the modular MLLM, we used SigLIP-SO [81] as the vision encoder, and the language model shared the same architecture and initialization parameters as SAIL. Both models were pre-trained using Pretraining stage-1 setting, with SAIL encountering 32M, 128M, and 512M image-text pairs during training, followed by finetuning on the LLaVA-mix-665K dataset. All parameters of both models are trainable. Both models employ an identical number of input tokens for images and text. The normalized performance of both models is plotted in Figure 1(A). The results show that in the low-data regime (32M), SAIL's performance lags behind the modular MLLM, likely due to SigLIP's prior training on 40B samples. However, as the data scales, SAIL exhibits a steeper performance curve, indicating more promising data scaling properties. At 512M image-text pairs, SAIL achieves performance comparable to the modular MLLM in our evaluation subset. This demonstrates the single transformer's superior data scalability, even without a pretrained vision encoder. + +Quantitative results on evaluated benchmark tasks of model scaling and data scaling are tabulated in Supp. + +# 4.3.2. Information Flow Pattern + +Different attention pattern compared to modular MLLM: since our comparative experiments show that the Single Transformer model exhibits more promising data scaling properties, we conducted an analysis of the trained SAIL model and its modular MLLM counterpart. Specifically, we followed the methodology from FastV [10] to + +![](images/cc3e29b6eb89c82f900839c5c0ca1a9eeef2287bd2b5759db834896d812bde8e.jpg) +Figure 4. Image Attention Score Allocation: The figure shows the proportion of image attention scores across different transformer layers for Single Transformer-based MLLM and modular MLLM when predicting tokens. Single Transformer-based MLLM generally allocates higher attention weights to image tokens compared to modular MLLM. + +analyze the attention score distribution for each predicted token given an image and a user query. This analysis focuses on how much attention is allocated to image tokens during token prediction. We selected 1000 samples from various datasets including VQAv2, GQA, TextVQA, DocVQA, MME, SEEDBench-IMG, MMBench, and some self-curated dialog examples. For each model prediction, we computed the average attention scores assigned to previous image tokens by the output token. + +We conducted a comparative experiment between Single Transformer-based MLLMs (SAIL, SOLO [11], EVE [21]) and modular MLLMs (Qwen2-VL [71], LLaVA-OneVision [40], LLaVA-1.5 [46]). + +The results are depicted in Figure 4. Single Transformer-based MLLMs allocate between $60\%$ and $80\%$ of attention scores to image tokens across all layers when predicting tokens. In contrast, modular MLLMs such as Qwen2-VL and LLaVA-OneVision allocate only $10\%$ to $30\%$ of attention scores to image tokens across different layers. For LLaVA-1.5, which does not update the ViT parameters during supervised fine-tuning (SFT), the image attention score is relatively high in the first two transformer layers but declines sharply in subsequent layers. + +From this experiment, we can conclude that Single Transformer-based MLLMs tend to allocate a significant portion of attention to previous image tokens during prediction. In contrast, modular MLLMs allocate a smaller portion of their attention directly to image tokens, indicating a less image-centric approach in their prediction mechanism. + +These findings indicate that the Single Transformer model places more emphasis on grounding its predictions in the visual information. As the model undergoes data scaling, it allocates more effective computation to image tokens, + +
MethodMMBench [49]MME [45]
Physical RelationCelebrity RelationPositionPostersCelebrity
Modular MLLM30.450.598.3134.0100.3
SAIL52.288.9160.0108.275.0
+ +thereby enhancing its capability as a vision-centric model. + +In summary, the attention pattern analysis underscores the Single Transformer's ability to robustly integrate visual context, enabling it to scale efficiently and potentially outperform modular MLLMs in vision-language tasks. + +# 4.3.3. Task-Specific Performance Analysis + +We dissect SAIL through case studies: + +Strengths: Spatial Reasoning. SAIL excels at tasks requiring precise spatial location. As shown in Table 6, under the setting of our data scaling experiment, it outperforms the modular counterpart by 61.7 points on the MME Position split and $21.8\%$ on MMBench Physical Relation questions. The unified architecture likely enables tighter coupling between visual geometry and linguistic descriptions. + +Weaknesses: World Knowledge. Conversely, SAIL falls short in tasks that demand extensive world knowledge. As shown in Table 6, SAIL underperforms in the MME celebrity and art splits compared to the modular MLLM. This underperformance can be attributed to SAIL's lack of diverse domain-specific data during pretraining, a gap that was not sufficiently addressed during supervised finetuning. Modular MLLMs, with their pretrained vision encoders like CLIP [15, 59] or SigLIP [81], have a broader knowledge base and therefore handle such tasks more effectively. We hypothesize that scaling up SAIL's pretraining data diversity could help bridge this gap, enhancing its performance on knowledge-intensive tasks. + +# 4.4. Empirical Observations on Basic Factors + +To guide scalable training of single-transformer MLLMs, we conduct ablation studies on two critical design choices using SAIL-0.5B pretrained on 128M image-text pairs and fine-tuned on LLaVA-mix-665K. Performance is evaluated through zero-shot image classification after pretraining [67] and vision-language benchmarks after SFT. + +Bidirectional Attention for Image Patches with Multimodal Position Encoding. We compare two approaches for integrating image patches into the transformer: (1) Causal attention with 1D positional encoding, using a token to demarcate image rows. (2) Full bidirectional attention for image patches paired with mul + +Table 6. Performance Comparison of SAIL and Modular MLLM in MMBench and MME Tasks: the strengths of SAIL in spatial reasoning tasks (MMBench Physical Relation and MME Position split) and its weaknesses in world knowledge tasks (MMBench Celebrity Relation and MME Celebrity and Posters splits). + +
Exp. SettingVQAv2GQASQATQASEED-I
Default59.146.959.620.135.1
#1 No Img full attn57.845.258.716.233.8
#2 No pure text in PT56.342.148.618.332.4
+ +Table 7. Ablation Study on Basic Factors for SAIL: We presents the impact of different ablation settings on the performance of SAIL across VQAv2 [27], GQA [35], SQA [51], TQA [62], and SEED-I [39]. The default setting includes image full attention and including pure text data in pretraining. Ablation #1 removes image full attention, and ablation #2 excludes pure text in pretraining. + +timodal rotary position embeddings (RoPE), which jointly encode spatial coordinates (e.g., 2D grid positions) and text token positions. As shown in Table 7, using bidirectional attention with multimodal RoPE significantly improves performance on vision-language tasks, with a particularly notable gain of $3.1\%$ on TextVQA. This suggests that enabling cross-patch interactions during pretraining enhances visual representation learning and tightens cross-modal alignment. Interleaving Pure Text Data During Pretraining. We analyze the impact of mixing SlimPajama text data with image-text pairs during pretraining. The results, as presented in Table 7 #2, reveal that mixing in pure text data consistently improves performance across vision and language benchmarks. This finding underscores the importance of preserving language capabilities in the LLM when training Single Transformer models, as maintaining strong language skills is crucial for building a multimodal model capable of complex reasoning. Currently, incorporating text data in training is one of the effective methods to maintain the language abilities of the model. + +Our ablation studies identify key designs for training SAIL effectively. Using bi-directional attention with multimodal rotary position embeddings enhances visual perception, while incorporating pure text data preserves language capabilities for robust multimodal performance. + +# 5. Conclusion + +In this work, we conducted an extensive analysis of Single Transformer-based MLLMs compared to modular MLLMs. Our investigation explored the unique properties of Single Transformers, including scalability, cross-modal information flow patterns, and visual representation capabilities. A series of experiments on our trained SAIL model demonstrated that this unified architecture achieves performance on vision-language benchmarks comparable to modular MLLMs while also functioning effectively as a vision backbone. Our findings highlight several advantages of Single Transformer architectures, such as superior data scalability, vision-centric information flow, and inherent capabilities as a powerful vision encoder. We hope our empirical findings will inspire further research to refine and enhance Single Transformer architectures, advancing the field of multimodal intelligence. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2 +[2] AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 1 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2, 6 +[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3 +[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Taşirlar. Introducing our multimodal models, 2023. 2, 5, 6 +[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 3 +[7] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2 +[8] Jieneng Chen, Qihang Yu, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Vitamin: Designing scalable vision models in the vision-language era. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3 +[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 5, 6 +[10] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 7 +[11] Yangyi Chen, Xingyao Wang, Hao Peng, and Heng Ji. A single transformer for scalable vision-language modeling. Transactions on Machine Learning Research, 2024. 1, 2, 3, 4, 5, 6, 7 +[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 5, 6 +[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing + +the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 6 +[14] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2, 3, 6 +[15] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 1, 2, 3, 6, 8 +[16] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality, 2023. 1 +[17] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.5 +[18] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1 +[19] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdul-mohsin, et al. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 3, 6 +[20] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3, 5 +[21] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. In Advances in Neural Information Processing Systems, pages 52545-52567. Curran Associates, Inc., 2024. 1, 2, 3, 4, 5, 6, 7 +[22] Haiwen Diao, Xiaotong Li, Yufeng Cui, Yueze Wang, Haoge Deng, Ting Pan, Wenxuan Wang, Huchuan Lu, and Xinlong Wang. *Evev2: Improved baselines for encoder-free vision-language models.* arXiv preprint arXiv:2502.06788, 2025. 5, 6 +[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3 + +[24] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11198-11201, 2024. 6 +[25] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19358-19369, 2023. 3 +[26] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 36:27092-27112, 2023. 3 +[27] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 8 +[28] Shuhao Gu, Jialing Zhang, Siyuan Zhou, Kevin Yu, Zhaohu Xing, Liangdong Wang, Zhou Cao, Jintao Jia, Zhuoyi Zhang, Yixuan Wang, et al. Infinity-mm: Scaling multimodal performance with large-scale and high-quality instruction data. arXiv preprint arXiv:2410.18558, 2024. 4, 5 +[29] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 5, 6 +[30] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3 +[31] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5 +[32] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. arXiv preprint arXiv:2312.08914, 2023. 2 +[33] Xinyu Huang, Youcai Zhang, Jinyu Ma, Weiwei Tian, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Lei Zhang. Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657, 2023. 3 +[34] Zilong Huang, Qinghao Ye, Bingyi Kang, Jiashi Feng, and Haoqi Fan. Classification done right for vision-language pretraining. In NeurIPS, 2024. 3 +[35] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional + +question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 8 +[36] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 3 +[37] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1 +[38] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images, 2016. 5, 6 +[39] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5, 6, 8 +[40] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 5, 6, 7 +[41] Xianhang Li, Zeyu Wang, and Cihang Xie. An inverse scaling law for clip training. In NeurIPS, 2023. 3 +[42] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, Yuyin Zhou, and Cihang Xie. What if we recapture billions of web images with llama-3? arXiv preprint arXiv:2406.08478, 2024. 4 +[43] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 6 +[44] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23390-23400, 2023. 3 +[45] Zijing Liang, Yanjie Xu, Yifan Hong, Penghui Shang, Qi Wang, Qiang Fu, and Ke Liu. A survey of multimodel large language models. In Proceedings of the 3rd International Conference on Computer, Artificial Intelligence and Control Engineering, pages 405-409, 2024. 5, 6, 8 +[46] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 6, 7 +[47] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.3,6 +[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3 +[49] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, + +Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5, 6, 8 +[50] Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xu-Cheng Yin, Cheng-Lin Liu, Lianwen Jin, and Xiang Bai. Ocrbench: on the hidden mystery ofOCR in large multimodal models. Science China Information Sciences, 67(12):220102, 2024. 5, 6 +[51] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5, 6, 8 +[52] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 5, 6 +[53] Gen Luo, Xue Yang, Wenhan Dou, Zhaokai Wang, Jiawen Liu, Jifeng Dai, Yu Qiao, and Xizhou Zhu. Mono-internvl: Pushing the boundaries of monolithic multimodal large language models with endogenous visual pre-training. arXiv preprint arXiv:2410.08202, 2024. 1, 2, 3, 5, 6 +[54] Sachin Mehta, Maxwell Horton, Fartash Faghri, Mohammad Hossein Sekhavat, Mahyar Najibi, Mehrdad Farajtabar, Oncel Tuzel, and Mohammad Rastegari. Catlip: Cliplevel visual recognition accuracy with 2.7 x faster pretraining on web-scale image-text data. arXiv preprint arXiv:2404.15653, 2024. 3 +[55] Meta. Introducing meta llama 3: The most capable openly available llm to date, 2024. Accessed: 2024-04-18. 1 +[56] OpenAI. Introducing chatgpt. OpenAI Blog, 2021. +[57] OpenAI. Gpt-4 technical report, 2023. 1 +[58] OpenAI. Gpt-4v(ision) system card, 2023. 1, 2 +[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 3, 6, 8 +[60] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 3, 4, 5 +[61] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatronlm: Training multi-billion parameter language models using model parallelism. arXiv preprint arXiv:1909.08053, 2019. 5 +[62] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 6, 8 + +[63] Daria Soboleva, Faisal Al-Khateeb, Robert Myers, Jacob R Steeves, Joel Hestness, and Nolan Dey. SlimPajama: A 627B token cleaned and deduplicated version of RedPajama. https://cerebras.ai/blog/slimpajama-a-627b-token-cleaned-and-deduplicated-version-of-redpajama, 2023.4 +[64] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2, 6 +[65] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2 +[66] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2024. 6 +[67] Michael Tschannen, Manoj Kumar, Andreas Steiner, Xiaohua Zhai, Neil Houlsby, and Lucas Beyer. Image captioners are scalable vision learners too. Advances in Neural Information Processing Systems, 36:46830-46855, 2023. 3, 8 +[68] Michael Tschannen, Manoj Kumar, Andreas Steiner, Xiaohua Zhai, Neil Houlsby, and Lucas Beyer. Image captioners are scalable vision learners too. Advances in Neural Information Processing Systems, 36:46830-46855, 2023. 6 +[69] Haochen Wang, Anlin Zheng, Yucheng Zhao, Tiancai Wang, Ge Zheng, Xiangyu Zhang, and Zhaoxiang Zhang. Reconstructive visual instruction tuning. In International Conference on Learning Representations, 2025. 2 +[70] Jiacong Wang, Bohong Wu, Haiyong Jiang, Zhou Xun, Xin Xiao, Haoyuan Guo, and Jun Xiao. World to code: Multimodal data generation via self-instructed compositional captioning and filtering. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4608-4623, 2024. 1 +[71] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution, 2024. 1, 2, 3, 7 +[72] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion feedback helps clip see better. arXiv preprint arXiv:2407.20171, 2024. 3 +[73] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 6 +[74] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In International Conference on Learning Representations, 2022. 3 +[75] X.ai. Grok-1.5 vision preview. https://x.ai/blog/grok-1.5v, 2024.5, 6 + +[76] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 5, 6 +[77] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3 +[78] Qiying Yu, Quan Sun, Xiaosong Zhang, Yufeng Cui, Fan Zhang, Yue Cao, Xinlong Wang, and Jingjing Liu. Capsfusion: Rethinking image-text data at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14022-14032, 2024. 4, 5 +[79] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 5, 6 +[80] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936, 2022. 5, 6 +[81] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3, 7, 8 +[82] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 3 +[83] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5 +[84] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 3 \ No newline at end of file diff --git a/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/images.zip b/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f1edc955599b88d51a1a4718ded9f92de1569925 --- /dev/null +++ b/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c62a7fe2487afdc1100d854d67bf25de144af16b97e147e31a3fae5df4b8572 +size 446419 diff --git a/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/layout.json b/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c33d25c78e8eaec696f8bebe40e85ac5a78f5032 --- /dev/null +++ b/2025/The Scalability of Simplicity_ Empirical Analysis of Vision-Language Learning with a Single Transformer/layout.json @@ -0,0 +1,7933 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 103, + 539, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 103, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 72, + 103, + 539, + 140 + ], + "type": "text", + "content": "The Scalability of Simplicity: Empirical Analysis of Vision-Language Learning with a Single Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 160, + 447, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 160, + 447, + 217 + ], + "spans": [ + { + "bbox": [ + 162, + 160, + 447, + 217 + ], + "type": "text", + "content": "Weixian Lei* Jiacong Wang* Haochen Wang* \nXiangtai Li Jun Hao Liew Jiashi Feng Zilong Huang† \n*Equal contribution, † Corresponding author \nBytedance Seed" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 296, + 547 + ], + "type": "text", + "content": "This paper introduces SAIL, a single transformer unified multimodal large language model (MLLM) that integrates raw pixel encoding and language decoding within a singular architecture. Unlike existing modular MLLMs, which rely on a pre-trained vision transformer (ViT), SAIL eliminates the need for a separate vision encoder, presenting a more minimalist architecture design. Instead of introducing novel architectural components, SAIL adapts mix-attention mechanisms and multimodal rotary position embedding to better align with the distinct characteristics of visual and textual modalities. We systematically compare SAIL's properties—including scalability, cross-modal information flow patterns, and visual representation capabilities—with those of modular MLLMs. By scaling both training data and model size, SAIL achieves performance comparable to modular MLLMs. Notably, the removal of pretrained ViT components enhances SAIL's scalability and results in significantly different cross-modal information flow patterns. Moreover, SAIL demonstrates strong visual representation capabilities, achieving results on par with ViT22B in vision tasks such as semantic segmentation. Code and models are available at https://github.com/bytedance/SAIL." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 135, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "The pursuit of multimodal intelligence has driven the development of Multimodal Large Language Models (MLLMs) [48, 58, 65], which typically adopt a modular design: a pre-trained vision encoder (e.g., CLIPViT [15, 59]) extracts image features, a Large Language Model (LLM) [2, 16, 37, 55-57] processes text, and a lightweight projector aligns the two modalities. This framework achieves strong performance through multi-stage pretraining, supervised fine-tuning (SFT), and post-training on multimodal datasets [3, 18, 48, 70, 71, 84]. While effective," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 251, + 437, + 363 + ], + "blocks": [ + { + "bbox": [ + 315, + 251, + 437, + 363 + ], + "lines": [ + { + "bbox": [ + 315, + 251, + 437, + 363 + ], + "spans": [ + { + "bbox": [ + 315, + 251, + 437, + 363 + ], + "type": "image", + "image_path": "5b3550ba22a6bb0225a56ba63ea2b9e2fb22e444614f6b4e1100358509dc9c52.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 366, + 388, + 376 + ], + "lines": [ + { + "bbox": [ + 376, + 366, + 388, + 376 + ], + "spans": [ + { + "bbox": [ + 376, + 366, + 388, + 376 + ], + "type": "text", + "content": "(A)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 384, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 473 + ], + "type": "text", + "content": "Figure 1. (A) Data scaling curve for Modular Multimodal Large Language Model (MLLM) and SAIL, our Single Transformer-based MLLM. As pretraining data increases, the single transformer SAIL shows a sharper performance gain, demonstrating its superior data scalability. Experiment settings are described in Section 4.3. (B) Comparison to existing Single Transformer-based MLLMs: our SAIL pushes the performance boundaries on both vision tasks and vision-language tasks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 438, + 251, + 553, + 363 + ], + "blocks": [ + { + "bbox": [ + 438, + 251, + 553, + 363 + ], + "lines": [ + { + "bbox": [ + 438, + 251, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 438, + 251, + 553, + 363 + ], + "type": "image", + "image_path": "4aaec2709c4af572be5932a3c719e9c28bd183f82a5bf5c8f87e81fb1ce455b9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 366, + 503, + 375 + ], + "lines": [ + { + "bbox": [ + 492, + 366, + 503, + 375 + ], + "spans": [ + { + "bbox": [ + 492, + 366, + 503, + 375 + ], + "type": "text", + "content": "(B)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 482, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 554, + 530 + ], + "type": "text", + "content": "this modular MLLM paradigm inherently fragments multimodal processing, reinforces reliance on pretrained visual encoders, which may limit deployment flexibility and scalability [11, 21, 53]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 312, + 532, + 555, + 687 + ], + "type": "text", + "content": "A promising alternative is to eliminate the visual encoder entirely and process raw image patches and text tokens within a single Transformer. This unified architecture removes modality-specific modules, enabling parameter sharing and end-to-end learning of vision-language interactions. Previous works [11, 21, 53] have primarily explored the architecture design, training data, and methods of Single Transformer-based MLLMs. However, little exploration has been given to their fundamental properties, such as scalability, cross-modal information flow patterns, and visual representation capabilities. A deeper understanding of these properties is crucial for unlocking the full potential of Single Transformer-based MLLMs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "In this work, we present an experimental analysis of the fundamental properties of Single Transformer-based" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20758" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "content": "MLLMs and compare them to modular MLLMs (e.g. LLaVA [48]). Additionally, in the absence of a pre-trained visual encoder, Single Transformers have to learn visual representations from scratch. Thus, an intriguing question arises: can a trained Single Transformer emerge as a strong vision encoder?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 145, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 295, + 312 + ], + "type": "text", + "content": "We conduct a series of experiments to train and study our Single trAnformer model for vIsion and Language (SAIL). While we do not propose novel architecture designs, we introduce necessary modifications to enable the model to process different modalities in a unified architecture. In its micro architecture design, we address the different spatial characteristics of 2D images and 1D text data by employing a mixed attention mechanism: bidirectional attention for image patches and causal attention for text tokens, combined with multimodal rotary position embedding. Through model and data scaling, SAIL achieves performance on vision-language benchmarks comparable to modular MLLMs, while also functioning as a high-performing vision backbone, as shown in Figure 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 314, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 295, + 337 + ], + "type": "text", + "content": "More concretely, our empirical analysis uncovers following striking advantages of Single Transformer architectures:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 339, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 295, + 434 + ], + "type": "text", + "content": "(i) Superior Data Scaling: In controlled experiments, SAIL exhibits steeper performance gains as pretraining data scales. While LLaVA-style modular MLLMs initially perform well, our model's performance becomes very close to theirs when pretrained on 512M samples, as shown in Figure 1(A). This suggests that unified architectures can effectively leverage large-scale data and potentially match the performance of modular MLLMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 436, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 531 + ], + "type": "text", + "content": "(ii) Vision Centric Information Flow Pattern: Through analysis of attention distributions, we observe that Single Transformers assign significantly higher attention scores to image tokens during token prediction compared to modular MLLMs. This indicates that the information flow in Single Transformer MLLMs is more direct, with visual tokens influencing prediction tokens more prominently, highlighting a vision-centric approach to decision-making." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 664 + ], + "type": "text", + "content": "(iii) Vision Encoder Functioning: Our experiments further demonstrate that the pretrained Single Transformer inherently serves as a powerful vision encoder. Comprehensive evaluations on vision-centric tasks, such as image classification and semantic segmentation, show that the model learns rich visual representations during multimodal pretraining. These representations enhance its capacity for both semantic-level comprehension (e.g., object categorization) and pixel-level understanding (e.g., fine-grained segmentation masks), bridging high-level abstraction and low-level visual reasoning within a unified architecture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "In summary, our findings indicate that Single Transformer-based MLLMs hold great promise in surpassing modular MLLMs in terms of leveraging large-scale data, forming direct vision-centric information pathways," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "and functioning as effective vision encoders. We hope our empirical findings inspire further research to refine and enhance Single Transformer architecture, ultimately driving advancements in multimodal intelligence from a new perspective." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 146, + 400, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 400, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 400, + 159 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 167, + 548, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 548, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 548, + 180 + ], + "type": "text", + "content": "2.1. Paradigms in Vision-Language Model Design" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 186, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 472 + ], + "type": "text", + "content": "Modular MLLMs with Visual Encoders. The prevailing approach in MLLM design employs modular architectures [3, 48, 58, 65, 71] that rely on pretrained vision encoders (e.g., CLIP-ViT [15, 59], InternViT [14]) to process visual inputs. The visual features extracted from these frozen encoders are then aligned with LLM input spaces via linear [7, 48, 69, 84] or cross-attention layers [1, 32]. While this module design enables effective transfer of pretrained visual-language knowledge, it also introduces several limitations. First, incorporating a separate ViT encoder significantly slows down both training and inference, increasing deployment complexity and requiring costly infrastructure—especially when compared to a single transformer unified model. Second, common strategies for integrating visual features, such as direct mapping into LLM inputs [7, 48, 84] or sharing them across LLM layers [1, 32], often struggle to reconcile the inherent differences between images and text representations. Finally, as model scale, balancing the interactions between the encoder, LLM, and alignment layers becomes increasingly challenging [11, 53]. Thus, in this work, we explore a single transformer-based MLLM architecture that eliminates the ViT encoder and alignment components to overcome these challenges." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 553, + 713 + ], + "type": "text", + "content": "Single Transformer-based MLLMs Without Visual Encoders. Emerging research explores end-to-end architectures that process raw image patches and text tokens through a single Transformer, bypassing visual encoders entirely. These monolithic designs fall into two categories: continuous tokenization and discrete tokenization. Continuous tokenization, exemplified by Fuyu-8B [5] and SOLO [11], directly maps patches to LLM embeddings via linear projections, enabling flexible resolution handling but requiring massive pretraining data. Discrete tokenization, adopted by Chameleon [64] and Emu3 [73], employs VQ-VAE tokenizers to compress images into discrete tokens, trading pixel-level fidelity for generation capabilities. While later efforts such as EVE [21] and MonoInternVL [53] demonstrate the feasibility of encoder-free training, critical gaps remain: (1) Existing methods rely on extra designs and auxiliary loss [21], complicating training pipelines; (2) The scaling laws and fundamental properties of purely end-to-end trained models remain poorly understood; (3) Vision-language interaction in shared parame" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20759" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "ter spaces lacks systematic analysis—most prior MLLMs default to causal attention for processing image-text sequences. In this work, we reveal that enabling bidirectional attention between image patches significantly enhances visual representation learning, addressing a key limitation in current designs. More importantly, our study bridges these gaps by establishing foundational principles for training scalable, self-contained single-transformer MLLMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 177, + 228, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 228, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 228, + 190 + ], + "type": "text", + "content": "2.2. Vision Representation Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 266 + ], + "type": "text", + "content": "Learning effective vision representations is a core challenge in computer vision research, with extensive works [4, 19, 20, 25, 30, 74] dedicated to this problem. With the proliferation of large-scale web-sourced image-text datasets [6, 26, 60], recent methods leverage this data to train deep visual representations via three primary paradigms:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 361 + ], + "type": "text", + "content": "Text as Classification Labels. Early methods used textual descriptions as weak supervision by extracting categorical labels from captions. For example, frameworks like Tag2Text [33] and RAM [82] used ViTs [23] to predict noun-based pseudo-labels from datasets like CC12M [6]. CatLIP [54] scaled labels to millions using object-centric supervision, and SuperClass [34] directly used tokenized text tokens as classification categories." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 363, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 422 + ], + "type": "text", + "content": "Image-Text Contrastive Learning Contrastive pretraining, as exemplified by CLIP [15, 59] and ALIGN [36], aligns global image-text embeddings within a shared latent space. Subsequent works [8, 14, 41, 44, 72, 81] focused on enhancing performance and improving training efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 578 + ], + "type": "text", + "content": "Text as Autoregressive Targets. Caption generation as a pretext task is another approach for visual representation learning. SimVLM [74] trains encoder-decoder architectures to autoregressively predict captions, while CapPa [67] trains vision encoders through sequence prediction. These methods often retain modular designs or auxiliary components like contrastive losses [77]. Our work aligns with this category but removes architectural fragmentation by jointly modeling image patches and text tokens in a single Transformer. We find that the pre-trained Single Transformer learns transferable vision representations, enabling it to handle downstream multimodal understanding tasks and function as a vision encoder without modifications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 590, + 295, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 617 + ], + "type": "text", + "content": "3. SAIL: Training a Single Transformer for Vision and Language" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 170, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 170, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 170, + 635 + ], + "type": "text", + "content": "3.1. Model Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": "SAIL is built upon a unified Transformer architecture (Figure 2(A)) that processes multimodal inputs through streamlined, modality-specific preprocessing. For text, raw input is tokenized using the language model's tokenizer and then transformed into embeddings via the textual embedding module. For images, we partition the input into fixed-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "size patches and project them into continuous embeddings via a linear projection. Additionally, we maintain a list of special tokens explicitly designed for visual modality encoding: and tokens mark the beginning and end of an image patch span, respectively. In multimodal scenarios, such as image-text pairs, these embeddings are concatenated into a single sequence and fed into the Transformer, enabling joint cross-modal interactions through unified self-attention layers. This design eliminates the need for modality-specific encoders, which efficiently processess heterogeneous data within a Single Transformer framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 217, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 337 + ], + "type": "text", + "content": "Bidirectional attention within image patches. While existing multimodal large language models (MLLMs) [40, 47, 48, 84] predominantly adopt causal attention for autoregressive sequence modeling, our experiments reveal that enabling full bidirectional attention among tokens from the same image significantly enhances visual representation learning and boosts downstream vision-language task performance. Note that previous Single Transformers [11, 21, 53, 73] have only utilized causal attention, without exploring the potential of mixed attention mechanisms." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 337, + 553, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 553, + 492 + ], + "type": "text", + "content": "As illustrated in Figure 2(B), for SAIL we implement a mixed attention scheme: (1) For text tokens, we preserve causal attention to maintain autoregressive generation capabilities, allowing each token to attend only to its predecessors. (2) For image tokens, we activate full bidirectional attention within each image patch group, empowering every visual token to interact with all others in the same image. This design captures holistic spatial relationships and contextual dependencies among visual elements, addressing the under-explored potential of attention mechanisms in cross-modal alignment. The improved interaction paradigm not only refines vision-language feature fusion but also provides stronger inductive biases for complex reasoning tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 494, + 553, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 553, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 553, + 661 + ], + "type": "text", + "content": "Multimodal Rotary Position Embeddings. Following [71], we implement Multimodal RoPE (M-RoPE) in SAIL to harmonize positional modeling for multimodal inputs. The method decomposes positional encoding into two axes: height, and width. For text tokens, all axes share uniform position IDs (aligned with 1D-RoPE), whereas for images, height/width IDs adaptively map to token coordinates, as is shown in Figure 2(C). Notably, position indexing is sequentially initialized across modalities (e.g., starting from images before extending to subsequent text), preserving inter-modal consistency. M-RoPE not only improves positional sensitivity but also constrains absolute position magnitudes for visual tokens, facilitating robust generalization to extended sequences in inference." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 671, + 391, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 391, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 391, + 684 + ], + "type": "text", + "content": "3.2. Pretraining" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "We apply a two-stage curriculum to progressively strengthen the visual perception of SAIL while preserving" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20760" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 73, + 308, + 269 + ], + "blocks": [ + { + "bbox": [ + 77, + 73, + 308, + 269 + ], + "lines": [ + { + "bbox": [ + 77, + 73, + 308, + 269 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 308, + 269 + ], + "type": "image", + "image_path": "7e4de715e61dd3d9bcd8398615d081679a59a412f7a74115d4387dad3353f861.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 140, + 276, + 240, + 287 + ], + "lines": [ + { + "bbox": [ + 140, + 276, + 240, + 287 + ], + "spans": [ + { + "bbox": [ + 140, + 276, + 240, + 287 + ], + "type": "text", + "content": "(A) Model Architecture" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 328, + 71, + 531, + 273 + ], + "blocks": [ + { + "bbox": [ + 328, + 71, + 531, + 273 + ], + "lines": [ + { + "bbox": [ + 328, + 71, + 531, + 273 + ], + "spans": [ + { + "bbox": [ + 328, + 71, + 531, + 273 + ], + "type": "image", + "image_path": "f386a7e5da052dd8a7f2235a610b1e87504e6c81edfbb8fae052f42dfc7eeac1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 275, + 500, + 288 + ], + "lines": [ + { + "bbox": [ + 365, + 275, + 500, + 288 + ], + "spans": [ + { + "bbox": [ + 365, + 275, + 500, + 288 + ], + "type": "text", + "content": "(C) Multimodal RoPE for SAIL" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 300, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 357 + ], + "type": "text", + "content": "Figure 2. Model architecture and micro-designs for SAIL. (A) Model Architecture: SAIL is a unified transformer that processes both images and texts without extra module designs. (B) Mixed Attention Mechanism: we adopt bidirectional attention for image patches from the same image and causal attention for text tokens. Examples for a multimodal sequence and a text sequence are provided. Colored squares represent \"allow to attend\" and white squares indicate \"prevent from attending\". (C) Multimodal RoPE: an illustration of the multimodal rotary position embedding for SAIL, with examples for a multimodal sequence and a text sequence." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 367, + 192, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 192, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 192, + 380 + ], + "type": "text", + "content": "its inherent language capabilities." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 390, + 293, + 487 + ], + "blocks": [ + { + "bbox": [ + 58, + 390, + 293, + 487 + ], + "lines": [ + { + "bbox": [ + 58, + 390, + 293, + 487 + ], + "spans": [ + { + "bbox": [ + 58, + 390, + 293, + 487 + ], + "type": "table", + "html": "
StageDatasetImg.ResNum
Pretraining S1Recap-DataComp-1B [42]224x224512M
SlimPajama [63]-
Pretraining S2Capfusion [78]AnyRes60M
OCR from LAION COCO [60]7M
InifinityMM Stage 2 subset [28]19M
SlimPajama [63]-
SFTInifinityMM Stage3 [28]AnyRes6M
", + "image_path": "b4bbd09bbe43a0d8800395fdac7badde95b4f539f571c1f0d4392350a14d9637.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 491, + 295, + 555 + ], + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 555 + ], + "type": "text", + "content": "Table 1. Details of training datasets used across all stages. \"Img.Res\" refers to the image resolution settings applied during each training stage. All datasets listed are publicly available. Note that these settings represent the default configuration for standard SAIL training, while separate settings are used for scaling experiments and ablation studies." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 557, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 295, + 689 + ], + "type": "text", + "content": "Stage 1: Accelerated Visual Knowledge Acquisition. In this stage, we pretrain SAIL on large-scale image-text pairs to rapidly bootstrap its visual understanding. To maximize data throughput, we uniformly resize all images to a lower resolution (e.g., " + }, + { + "bbox": [ + 54, + 557, + 295, + 689 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 54, + 557, + 295, + 689 + ], + "type": "text", + "content": "), reducing multimodal sequence lengths and enabling the model to process more samples within fixed training time. To prevent catastrophic forgetting of linguistic knowledge, we interleave pure text corpora with multimodal data during training. This hybrid approach ensures efficient exposure to visual patterns while maintaining robust language proficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Stage 2: Enhancing Any-Resolution Image Understanding. Real-world applications require robustness to images" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 367, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 475 + ], + "type": "text", + "content": "of varying resolutions and aspect ratios, such as documents, charts, or infographics. Following prior works [11, 21], we extend pretraining with an any-resolution strategy: images retain their native resolutions during processing, and positional embeddings adapt dynamically to arbitrary spatial dimensions. This stage further refines SAIL's ability to model fine-grained visual details (e.g., tabular structures, text-rich graphics) while continuing to incorporate text-only data for language capability preservation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 523 + ], + "type": "text", + "content": "Pretraining Objective. Throughout both stages, we optimize the standard language modeling loss only on text tokens. Image patches and special visual tokens are excluded from loss computation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 529, + 447, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 447, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 447, + 541 + ], + "type": "text", + "content": "3.3. Supervised Fine-tuning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 640 + ], + "type": "text", + "content": "During the Supervised Fine-tuning (SFT) stage, we train SAIL on publicly available, multi-source instruction datasets to enhance its understanding of complex linguistic instructions and diverse dialogue patterns critical for real-world deployment. This phase fine-tunes the entire network architecture, focusing on aligning the model's responses with human intent through exposure to varied instructional formats and multimodal interactions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "Table 1 shows the details of training datasets for pretraining and supervised fine-tuning (SFT) across all stages. During Stage 1 pretraining, we utilized mixed multimodal and pure text datasets including Recap-DataComp-1B [42] and SlimPajama [63], with images at a resolution of " + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": ", totaling 512M image-text pairs. In Stage 2, the pretrain" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20761" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "ing datasets include Capfusion [78], self-curated OCR data from LAION COCO [60], InfinityMM Stage 2 subset [28], and SlimPajama, utilizing the any resolution (AnyRes) strategy, with a combined total of 86M image-text pairs along with text data. The SFT stage employed the InfinityMM Stage 3 subset, processed at any resolution, containing 6M image-text pairs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 132, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 132, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 132, + 182 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 190, + 182, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 182, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 182, + 204 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 208, + 295, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 316 + ], + "type": "text", + "content": "Evaluation Benchmarks. For evaluation of vision and language tasks, we evaluate SAIL and existing MLLMs on a broad range of multimodal benchmarks. Specifically, MLLM benchmarks encompass MMBench-EN [49], SEEDBench-IMG [39], MMVet [79], MME [45], HallusionBench [29], MathVistaMINI [52], and OCR-Bench [50]. Visual question answering benchmarks include TextVQA [62], ScienceQA-IMG [51], AI2D [38], MMStar [9], RealWorldQA [75]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 365 + ], + "type": "text", + "content": "For evaluation of vision representation learning, we conduct experiments on ImageNet-1K [20] for image classification, ADE20K for semantic segmentation [83], and ARO [80] for attribute, relation, and ordering." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 533 + ], + "type": "text", + "content": "Implementation Details. For pretraining, we initialize SAIL from the Mistral-7B-v0.1 base LLM and set the patch size to 14. We modify Megatron [61] to support SAIL's multimodal input. Pretraining uses 128 NVIDIA A100 80G GPUs with 2-way tensor parallelism and 64-way data parallelism. The learning rate is set at 5e-5 and decays cosinely to a minimum of 5e-6. For training efficiency, we concatenate sequences from different data samples into one long sequence of 32,768 tokens, adjusting the attention mask to ensure that tokens from different samples do not attend to each other. We use a round-robin approach to interleave image-text packed sequences and pure text packed sequences, configuring the global batch to contain approximately 16K image-text pairs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": "For Supervised Fine-Tuning (SFT), the global batch size is set to 512. Training is performed for one epoch with a maximum learning rate of 1e-5, following a linear warm-up phase and then transitioning to a cosine decay schedule." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "For vision, we load the checkpoint after Stage 1 pretraining and keep it frozen for downstream evaluations, including (1) image classification on ImageNet-1K [20], (2) semantic segmentation on ADE20K [83], and (3) attribute, relation and, ordering on the ARO benchmark [80]. Specifically, (1) for image classification, we utilize a linear classifier with 90 epochs for linear probing, where detailed configurations are mostly obtained from common practices [31]. Images are resized to " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " and the global batch size is 8,192 across 8 A100 (80G) GPUs. (2) For semantic segmentation, we adopt ViT-Adapter [12] with UperNet [76] as the segmenta" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "content": "tion decoder. The implementation is based on MMSegmentation [17] with 80k training iterations. The input resolution is " + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 72, + 553, + 145 + ], + "type": "text", + "content": " and the global batch size is 16 across 8 A100 (80G) GPUs. (3) For attribute, relation and, ordering, we regard the negative of the caption loss over each image-text pair as the similarity metric for retrieval." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 153, + 437, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 437, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 437, + 166 + ], + "type": "text", + "content": "4.2. Experimental Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 171, + 487, + 184 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 487, + 184 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 487, + 184 + ], + "type": "text", + "content": "4.2.1. Results on Vision Language Tasks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 186, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 450 + ], + "type": "text", + "content": "As shown in Table 2, we compare SAIL against existing MLLMs across 13 vision-language benchmarks. SAIL consistently outperforms other Single Transformer-based models like Fuyu [5], EVE [21], SOLO [11], MonoInternVL [53], and EVE2 [22] across diverse vision-language tasks. This demonstrates that SAIL can achieve significant performance gains and push the boundaries of Single Transformer-based MLLMs without needing extra component designs or auxiliary training losses. Moreover, when compared to methods employing discrete vision tokens (e.g., Chameleon and Emu3), SAIL demonstrates superior performance. These results validate that scaling up single-transformer pretraining effectively enhances cross-modal alignment between images and text. Compared to the state-of-the-art modular MLLM LLaVA-OneVision [40], SAIL achieves comparable performance on some benchmarks, such as MMStar, SEEDBench-IMG, and RealWorldQA. While the performance of Single Transformer-based MLLMs still lags behind modular MLLMs in certain areas, we hypothesize that scaling the pretraining data volume or incorporating higher-quality instruction-tuning data will bridge the remaining performance gap." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 457, + 525, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 525, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 525, + 470 + ], + "type": "text", + "content": "4.2.2. Results on Vision Representation Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 473, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 553, + 509 + ], + "type": "text", + "content": "In this section, we compare the quality of learned visual representations of our SAIL with other Single Transformers, including EVE [21], EVE2 [22], and SOLO [11]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "type": "text", + "content": "Classification and Segmentation. As demonstrated in Table 3, our method, SAIL, achieves a Top-1 accuracy of " + }, + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "type": "inline_equation", + "content": "84.06\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "type": "text", + "content": " and a Top-5 accuracy of " + }, + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "type": "inline_equation", + "content": "97.48\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "type": "text", + "content": " on the validation set of ImageNet-1K [20], significantly outperforming state-of-the-art alternatives [11, 21, 22]. In the segmentation task, SAIL also demonstrates superior performance with an mIoU of " + }, + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "type": "inline_equation", + "content": "55.30\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "type": "text", + "content": ", an mAcc of " + }, + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "type": "inline_equation", + "content": "67.24\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "type": "text", + "content": ", and an aAcc of " + }, + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "type": "inline_equation", + "content": "84.87\\%" + }, + { + "bbox": [ + 313, + 510, + 555, + 677 + ], + "type": "text", + "content": ", illustrated in Table 3. These results indicate that SAIL is effective in both classification and segmentation tasks, offering substantial improvements over existing methods. In Table 4, even when comparing with other state-of-the-art vision backbones, our SAIL manages to achieve remarkable competitive performance with significantly less training data, demonstrating the scaling property of SAIL." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Attribute, Relation, and Ordering. To systematically evaluate the ability of SAIL to understand different types of relationships, attributes, and order information, we conduct" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20762" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 254 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 254 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 254 + ], + "type": "table", + "html": "
Method#Param#Data#VtokenGeneral VQAHallucinationMath&knowledgeOCR VQA
MMS*MMBenSEEDIMMVMMERWQAPOPEHalluSQA1MathVTQAAI2DOCRB
Modular MLLMs:
InternVL-1.5 [13]2.2B-/-332846.770.969.839.3190257.988.337.384.941.370.569.8654
QwenVL-Chat [3]7B7.2B / 50M25634.560.658.2-184849.3-36.868.235.361.545.9488
LLVA-1.5 [46]7B0.4B+ / 665K57633.164.364.330.5185954.885.927.666.825.546.154.8318
LLVA-1.6 [47]7B0.4B+ / 760K288037.667.464.743.9184257.886.427.670.232.564.966.6532
Cambrian-1 [66]8B10B+ / 7M57650.775.974.7--64.2-30.680.448.171.773.0-
LLVA-OneVision [40]7B10B+ / 3.2M729060.981.774.858.8199865.5--96.656.1-81.6-
Single Transformer-based MLLMs:
Fuyu [5]8B-/--34.410.759.321.4-43.78429.856.830.2-46.8366
Chameleon [64]7B1.4B+ / 1.8M102431.131.130.68.31703919.417.147.222.54.846.07.0
EVE [21]7B33M / 1.8M2304-52.364.625.71628-85.0-64.9-56.861.0398
SOLO [11]8B43.7M / 2M102435.867.764.430.4126044.778.640.473.332.925.061.4126
Mono-InternVL [53]3B1.3B / 7M6400-65.567.440.11875--45.793.645.772.668.6767
Emu3 [73]8B-/-16K46.658.568.237.2-57.485.231.789.231.364.770.0687
EVE2 [22]7B92M / 7.3M2500-66.371.445.0170962.487.6-96.2-71.174.8702
SAIL7B600M / 6M360053.170.172.946.3171963.985.854.293.357.077.176.7783
", + "image_path": "898288e5affcc9e3a84ca8645bd174c549a22bdb25207ef3a93f6c98e728f478.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 362, + 294, + 446 + ], + "blocks": [ + { + "bbox": [ + 55, + 266, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 555, + 344 + ], + "type": "text", + "content": "Table 2. Comparison with existing vision-language models on various vision-language benchmarks, including MMS*: MMStar [9]; MMBench-EN [49]; SEED: SEEDBench-IMG [39]; MMV: MMVet [79]; MME [45]; POPE [43]; Hallu: HallusionBench [29]; SQA: ScienceQA-Img [51]; TVQA: TextVQA [62]; MathV: MathVistaMINI [52]; AI2D [38]; RWQA: RealWorldQA [75]; OCRB: OCR-Bench [50]. Note that #A-Param denotes the number of activated parameters; #Data represents the pre-training / fine-tuning data volume; #Vtoken indicates the maximum image patch tokens. For MME, we report the sum of perception and cognition scores. The top two results are highlighted in bold and underline, respectively. All results are derived from those reported in other papers and the official reproduction results from the OpenCompass leaderboard [24]. Our results are obtained from VLMEvalKit [24]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 362, + 294, + 446 + ], + "lines": [ + { + "bbox": [ + 56, + 362, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 362, + 294, + 446 + ], + "type": "table", + "html": "
MethodClassificationSegmentation
Top-1Top-5mIoUmAccaAcc
EVE [21]42.0365.7727.1235.8972.91
EVE2 [22]44.8669.4140.8553.5379.31
SOLO [11]59.1080.8935.1144.8176.02
SAIL84.0697.4855.3067.2484.87
", + "image_path": "26c8b3c48f5b96c37a8a141991a6911b2cf9f3cccfee847afdfc86f1222c139d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 61, + 505, + 291, + 583 + ], + "blocks": [ + { + "bbox": [ + 55, + 455, + 295, + 489 + ], + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 489 + ], + "type": "text", + "content": "Table 3. Comparison on image classification and semantic segmentation with other encoder-free approaches. Our SAIL outperforms other alternatives by a large margin." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 505, + 291, + 583 + ], + "lines": [ + { + "bbox": [ + 61, + 505, + 291, + 583 + ], + "spans": [ + { + "bbox": [ + 61, + 505, + 291, + 583 + ], + "type": "table", + "html": "
Method#Data#ParamImageNet-1KADE20K
OpenCLIP-H [15]2B0.6B84.4-
OpenCLIP-G [15]2B1.8B86.239.3†
ViT-22B [19]3B22B89.555.3
InternViT [14]6B6B88.258.7
SAIL0.5B7B84.155.3
", + "image_path": "ac58cf334df9fa18cc1e9a1c5e1ae9744ecdeef63f2ac21834885222f2d93fc5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "experiments on the ARO benchmark [80]. As demonstrated in Table 5, SAIL encodes compositional relationships between objects and attributes almost perfectly, significantly surpassing other state-of-the-art vision backbones." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 362, + 553, + 491 + ], + "blocks": [ + { + "bbox": [ + 55, + 592, + 295, + 648 + ], + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 648 + ], + "type": "text", + "content": "Table 4. Comparison on image classification and semantic segmentation with other vision backbones. " + }, + { + "bbox": [ + 55, + 592, + 295, + 648 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 592, + 295, + 648 + ], + "type": "text", + "content": " indicates training with head tuning using UperNet [76], while others are based on ViT-Adapter [12]. SAIL, with significantly less training data, achieves competitive performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 362, + 553, + 491 + ], + "lines": [ + { + "bbox": [ + 316, + 362, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 362, + 553, + 491 + ], + "type": "table", + "html": "
MethodRelationAttributeOrder
COCOFlickr30K
OpenCLIP-H [15]49.964.632.640.4
OpenCLIP-G [15]49.965.633.038.3
CLIP-B/32 [59]59.262.948.157.9
CLIP-L/14 [59]61.261.746.856.8
InternViT [14]59.666.073.476.3
NegCLIP [80]81.071.086.091.0
CapPa [68]86.785.798.899.2
SAIL100.099.5100.0100.0
", + "image_path": "58247d54c3914871d0dcab7932bae06effc80ba180323469d40c7908c71ca038.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 498, + 555, + 533 + ], + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 533 + ], + "type": "text", + "content": "Table 5. Comparison on attribute, relation, and ordering (ARO) with other vision backbones. SAIL almost encodes compositional relationships between objects and attributes perfectly." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 545, + 492, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 492, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 492, + 559 + ], + "type": "text", + "content": "4.3. Properties of Single Transformer" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 563, + 424, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 424, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 424, + 576 + ], + "type": "text", + "content": "4.3.1. Scaling Properties." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 580, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 688 + ], + "type": "text", + "content": "Model Scaling: We selected models of different sizes: SAIL-0.5B, SAIL-3B, and SAIL-7B (SAIL by default) for our experiments. Each model underwent Stage 1 pretraining on a mixed multimodal and pure text dataset, encountering 512M image-text pairs. Subsequently, they were finetuned on the LLaVA-mix-665K dataset using the any resolution (anyres) strategy. We evaluated the models based on their performance on vision and language benchmarks after supervised fine-tuning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "The normalized performance of SAIL against model size is plotted in Figure 3. As the model size scales up, we ob" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20763" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 294, + 169 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 294, + 169 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 294, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 294, + 169 + ], + "type": "image", + "image_path": "9f63de2d668ebe946af061ff2b559117f9fbab47d3ccc1f41eb89bcd60f9949a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 178, + 295, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 295, + 222 + ], + "type": "text", + "content": "Figure 3. Model scaling of SAIL. Left: As the model size increases, the training language modeling loss gradually decreases. Right: As the model size increases, performance on downstream VLM tasks progressively improves." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 234, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 234, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 234, + 295, + 354 + ], + "type": "text", + "content": "serve a corresponding enhancement in performance. Additionally, as shown on the left side of Figure 3, the training language modeling loss decreases with increasing model size. This reduction in training loss indicates that larger models have a greater capacity to learn multimodal alignments effectively, enabling them to capture complex relationships between vision and language more accurately. The improved learning capacity directly translates to better performance on downstream VLM tasks, showcasing the benefits of scaling up the Single Transformer architecture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 355, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 594 + ], + "type": "text", + "content": "Data Scaling: we compared SAIL with its modular MLLM counterpart. For the modular MLLM, we used SigLIP-SO [81] as the vision encoder, and the language model shared the same architecture and initialization parameters as SAIL. Both models were pre-trained using Pretraining stage-1 setting, with SAIL encountering 32M, 128M, and 512M image-text pairs during training, followed by finetuning on the LLaVA-mix-665K dataset. All parameters of both models are trainable. Both models employ an identical number of input tokens for images and text. The normalized performance of both models is plotted in Figure 1(A). The results show that in the low-data regime (32M), SAIL's performance lags behind the modular MLLM, likely due to SigLIP's prior training on 40B samples. However, as the data scales, SAIL exhibits a steeper performance curve, indicating more promising data scaling properties. At 512M image-text pairs, SAIL achieves performance comparable to the modular MLLM in our evaluation subset. This demonstrates the single transformer's superior data scalability, even without a pretrained vision encoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 619 + ], + "type": "text", + "content": "Quantitative results on evaluated benchmark tasks of model scaling and data scaling are tabulated in Supp." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 626, + 194, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 194, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 194, + 637 + ], + "type": "text", + "content": "4.3.2. Information Flow Pattern" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Different attention pattern compared to modular MLLM: since our comparative experiments show that the Single Transformer model exhibits more promising data scaling properties, we conducted an analysis of the trained SAIL model and its modular MLLM counterpart. Specifically, we followed the methodology from FastV [10] to" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 228 + ], + "type": "image", + "image_path": "cc3e29b6eb89c82f900839c5c0ca1a9eeef2287bd2b5759db834896d812bde8e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 237, + 554, + 304 + ], + "lines": [ + { + "bbox": [ + 313, + 237, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 554, + 304 + ], + "type": "text", + "content": "Figure 4. Image Attention Score Allocation: The figure shows the proportion of image attention scores across different transformer layers for Single Transformer-based MLLM and modular MLLM when predicting tokens. Single Transformer-based MLLM generally allocates higher attention weights to image tokens compared to modular MLLM." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 554, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 420 + ], + "type": "text", + "content": "analyze the attention score distribution for each predicted token given an image and a user query. This analysis focuses on how much attention is allocated to image tokens during token prediction. We selected 1000 samples from various datasets including VQAv2, GQA, TextVQA, DocVQA, MME, SEEDBench-IMG, MMBench, and some self-curated dialog examples. For each model prediction, we computed the average attention scores assigned to previous image tokens by the output token." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 422, + 554, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 470 + ], + "type": "text", + "content": "We conducted a comparative experiment between Single Transformer-based MLLMs (SAIL, SOLO [11], EVE [21]) and modular MLLMs (Qwen2-VL [71], LLaVA-OneVision [40], LLaVA-1.5 [46])." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 590 + ], + "type": "text", + "content": "The results are depicted in Figure 4. Single Transformer-based MLLMs allocate between " + }, + { + "bbox": [ + 313, + 472, + 554, + 590 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 472, + 554, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 554, + 590 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 472, + 554, + 590 + ], + "type": "text", + "content": " of attention scores to image tokens across all layers when predicting tokens. In contrast, modular MLLMs such as Qwen2-VL and LLaVA-OneVision allocate only " + }, + { + "bbox": [ + 313, + 472, + 554, + 590 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 472, + 554, + 590 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 472, + 554, + 590 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 472, + 554, + 590 + ], + "type": "text", + "content": " of attention scores to image tokens across different layers. For LLaVA-1.5, which does not update the ViT parameters during supervised fine-tuning (SFT), the image attention score is relatively high in the first two transformer layers but declines sharply in subsequent layers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 593, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 554, + 664 + ], + "type": "text", + "content": "From this experiment, we can conclude that Single Transformer-based MLLMs tend to allocate a significant portion of attention to previous image tokens during prediction. In contrast, modular MLLMs allocate a smaller portion of their attention directly to image tokens, indicating a less image-centric approach in their prediction mechanism." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "These findings indicate that the Single Transformer model places more emphasis on grounding its predictions in the visual information. As the model undergoes data scaling, it allocates more effective computation to image tokens," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20764" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 140 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 140 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 140 + ], + "type": "table", + "html": "
MethodMMBench [49]MME [45]
Physical RelationCelebrity RelationPositionPostersCelebrity
Modular MLLM30.450.598.3134.0100.3
SAIL52.288.9160.0108.275.0
", + "image_path": "6048758e170d955de58ba2977175f5060395adc72e1ee17b1cfadaa6e71c5d41.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 211, + 291, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 291, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 291, + 222 + ], + "type": "text", + "content": "thereby enhancing its capability as a vision-centric model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 223, + 295, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 271 + ], + "type": "text", + "content": "In summary, the attention pattern analysis underscores the Single Transformer's ability to robustly integrate visual context, enabling it to scale efficiently and potentially outperform modular MLLMs in vision-language tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 236, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 236, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 236, + 289 + ], + "type": "text", + "content": "4.3.3. Task-Specific Performance Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 293, + 211, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 211, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 211, + 304 + ], + "type": "text", + "content": "We dissect SAIL through case studies:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 304, + 295, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 388 + ], + "type": "text", + "content": "Strengths: Spatial Reasoning. SAIL excels at tasks requiring precise spatial location. As shown in Table 6, under the setting of our data scaling experiment, it outperforms the modular counterpart by 61.7 points on the MME Position split and " + }, + { + "bbox": [ + 55, + 304, + 295, + 388 + ], + "type": "inline_equation", + "content": "21.8\\%" + }, + { + "bbox": [ + 55, + 304, + 295, + 388 + ], + "type": "text", + "content": " on MMBench Physical Relation questions. The unified architecture likely enables tighter coupling between visual geometry and linguistic descriptions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 388, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 544 + ], + "type": "text", + "content": "Weaknesses: World Knowledge. Conversely, SAIL falls short in tasks that demand extensive world knowledge. As shown in Table 6, SAIL underperforms in the MME celebrity and art splits compared to the modular MLLM. This underperformance can be attributed to SAIL's lack of diverse domain-specific data during pretraining, a gap that was not sufficiently addressed during supervised finetuning. Modular MLLMs, with their pretrained vision encoders like CLIP [15, 59] or SigLIP [81], have a broader knowledge base and therefore handle such tasks more effectively. We hypothesize that scaling up SAIL's pretraining data diversity could help bridge this gap, enhancing its performance on knowledge-intensive tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 270, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 270, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 270, + 565 + ], + "type": "text", + "content": "4.4. Empirical Observations on Basic Factors" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 640 + ], + "type": "text", + "content": "To guide scalable training of single-transformer MLLMs, we conduct ablation studies on two critical design choices using SAIL-0.5B pretrained on 128M image-text pairs and fine-tuned on LLaVA-mix-665K. Performance is evaluated through zero-shot image classification after pretraining [67] and vision-language benchmarks after SFT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Bidirectional Attention for Image Patches with Multimodal Position Encoding. We compare two approaches for integrating image patches into the transformer: (1) Causal attention with 1D positional encoding, using a token to demarcate image rows. (2) Full bidirectional attention for image patches paired with mul" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 320, + 71, + 552, + 122 + ], + "blocks": [ + { + "bbox": [ + 55, + 144, + 295, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 201 + ], + "type": "text", + "content": "Table 6. Performance Comparison of SAIL and Modular MLLM in MMBench and MME Tasks: the strengths of SAIL in spatial reasoning tasks (MMBench Physical Relation and MME Position split) and its weaknesses in world knowledge tasks (MMBench Celebrity Relation and MME Celebrity and Posters splits)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 71, + 552, + 122 + ], + "lines": [ + { + "bbox": [ + 320, + 71, + 552, + 122 + ], + "spans": [ + { + "bbox": [ + 320, + 71, + 552, + 122 + ], + "type": "table", + "html": "
Exp. SettingVQAv2GQASQATQASEED-I
Default59.146.959.620.135.1
#1 No Img full attn57.845.258.716.233.8
#2 No pure text in PT56.342.148.618.332.4
", + "image_path": "fbc15c2fc7e771856c3eeca0b13e830272fdf699943a1cc5ef90cfa85321b81c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 126, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 126, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 555, + 194 + ], + "type": "text", + "content": "Table 7. Ablation Study on Basic Factors for SAIL: We presents the impact of different ablation settings on the performance of SAIL across VQAv2 [27], GQA [35], SQA [51], TQA [62], and SEED-I [39]. The default setting includes image full attention and including pure text data in pretraining. Ablation #1 removes image full attention, and ablation #2 excludes pure text in pretraining." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 198, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 437 + ], + "type": "text", + "content": "timodal rotary position embeddings (RoPE), which jointly encode spatial coordinates (e.g., 2D grid positions) and text token positions. As shown in Table 7, using bidirectional attention with multimodal RoPE significantly improves performance on vision-language tasks, with a particularly notable gain of " + }, + { + "bbox": [ + 313, + 198, + 555, + 437 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 313, + 198, + 555, + 437 + ], + "type": "text", + "content": " on TextVQA. This suggests that enabling cross-patch interactions during pretraining enhances visual representation learning and tightens cross-modal alignment. Interleaving Pure Text Data During Pretraining. We analyze the impact of mixing SlimPajama text data with image-text pairs during pretraining. The results, as presented in Table 7 #2, reveal that mixing in pure text data consistently improves performance across vision and language benchmarks. This finding underscores the importance of preserving language capabilities in the LLM when training Single Transformer models, as maintaining strong language skills is crucial for building a multimodal model capable of complex reasoning. Currently, incorporating text data in training is one of the effective methods to maintain the language abilities of the model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 498 + ], + "type": "text", + "content": "Our ablation studies identify key designs for training SAIL effectively. Using bi-directional attention with multimodal rotary position embeddings enhances visual perception, while incorporating pure text data preserves language capabilities for robust multimodal performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": "In this work, we conducted an extensive analysis of Single Transformer-based MLLMs compared to modular MLLMs. Our investigation explored the unique properties of Single Transformers, including scalability, cross-modal information flow patterns, and visual representation capabilities. A series of experiments on our trained SAIL model demonstrated that this unified architecture achieves performance on vision-language benchmarks comparable to modular MLLMs while also functioning effectively as a vision backbone. Our findings highlight several advantages of Single Transformer architectures, such as superior data scalability, vision-centric information flow, and inherent capabilities as a powerful vision encoder. We hope our empirical findings will inspire further research to refine and enhance Single Transformer architectures, advancing the field of multimodal intelligence." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20765" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 295, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 295, + 178 + ], + "type": "text", + "content": "[2] AI Anthropic. The claude 3 model family: Opus, sonnet, haiku. Claude-3 Model Card, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 180, + 295, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 295, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 295, + 233 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 236, + 294, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 236, + 294, + 268 + ], + "spans": [ + { + "bbox": [ + 61, + 236, + 294, + 268 + ], + "type": "text", + "content": "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 302 + ], + "type": "text", + "content": "[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sagnak Taşirlar. Introducing our multimodal models, 2023. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 303, + 296, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 296, + 357 + ], + "type": "text", + "content": "[6] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558-3568, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 358, + 296, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 296, + 423 + ], + "type": "text", + "content": "[7] Jun Chen, Deyao Zhu, Xiaogian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 296, + 479 + ], + "type": "text", + "content": "[8] Jieneng Chen, Qihang Yu, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Vitamin: Designing scalable vision models in the vision-language era. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 534 + ], + "type": "text", + "content": "[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, et al. Are we on the right way for evaluating large vision-language models? arXiv preprint arXiv:2403.20330, 2024. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "text", + "content": "[10] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19-35. Springer, 2024. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[11] Yangyi Chen, Xingyao Wang, Hao Peng, and Heng Ji. A single transformer for scalable vision-language modeling. Transactions on Machine Learning Research, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "text", + "content": "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision transformer adapter for dense predictions. arXiv preprint arXiv:2205.08534, 2022. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 714 + ], + "type": "text", + "content": "[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 106, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 172 + ], + "type": "text", + "content": "[14] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 248 + ], + "type": "text", + "content": "[15] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[16] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": " * chatgpt quality, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 349 + ], + "type": "text", + "content": "[17] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 415 + ], + "type": "text", + "content": "[18] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. InstructBLIP: Towards general-purpose vision-language models with instruction tuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 481 + ], + "type": "text", + "content": "[19] Mostafa Dehghani, Josip Djolonga, Basil Mustafa, Piotr Padlewski, Jonathan Heek, Justin Gilmer, Andreas Peter Steiner, Mathilde Caron, Robert Geirhos, Ibrahim Alabdul-mohsin, et al. Scaling vision transformers to 22 billion parameters. In International Conference on Machine Learning, pages 7480-7512. PMLR, 2023. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "text", + "content": "[20] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 581 + ], + "type": "text", + "content": "[21] Haiwen Diao, Yufeng Cui, Xiaotong Li, Yueze Wang, Huchuan Lu, and Xinlong Wang. Unveiling encoder-free vision-language models. In Advances in Neural Information Processing Systems, pages 52545-52567. Curran Associates, Inc., 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[22] Haiwen Diao, Xiaotong Li, Yufeng Cui, Yueze Wang, Haoge Deng, Ting Pan, Wenxuan Wang, Huchuan Lu, and Xinlong Wang. *Evev2: Improved baselines for encoder-free vision-language models.* arXiv preprint arXiv:2502.06788, 2025. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 636, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 712 + ], + "type": "text", + "content": "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20766" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[24] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 11198-11201, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "text", + "content": "[25] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19358-19369, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 205, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 270 + ], + "type": "text", + "content": "[26] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal datasets. Advances in Neural Information Processing Systems, 36:27092-27112, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 271, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 294, + 327 + ], + "type": "text", + "content": "[27] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "text", + "content": "[28] Shuhao Gu, Jialing Zhang, Siyuan Zhou, Kevin Yu, Zhaohu Xing, Liangdong Wang, Zhou Cao, Jintao Jia, Zhuoyi Zhang, Yixuan Wang, et al. Infinity-mm: Scaling multimodal performance with large-scale and high-quality instruction data. arXiv preprint arXiv:2410.18558, 2024. 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 383, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 459 + ], + "type": "text", + "content": "[29] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, et al. Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14375-14385, 2024. 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 459, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 294, + 514 + ], + "type": "text", + "content": "[30] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 514, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 568 + ], + "type": "text", + "content": "[31] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[32] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. arXiv preprint arXiv:2312.08914, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "text", + "content": "[33] Xinyu Huang, Youcai Zhang, Jinyu Ma, Weiwei Tian, Rui Feng, Yuejie Zhang, Yaqian Li, Yandong Guo, and Lei Zhang. Tag2text: Guiding vision-language model via image tagging. arXiv preprint arXiv:2303.05657, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "text", + "content": "[34] Zilong Huang, Qinghao Ye, Bingyi Kang, Jiashi Feng, and Haoqi Fan. Classification done right for vision-language pretraining. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[35] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 106, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 171 + ], + "type": "text", + "content": "[36] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 172, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 227 + ], + "type": "text", + "content": "[37] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 259 + ], + "type": "text", + "content": "[38] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images, 2016. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 304 + ], + "type": "text", + "content": "[39] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 5, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 304, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 350 + ], + "type": "text", + "content": "[40] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 371 + ], + "type": "text", + "content": "[41] Xianhang Li, Zeyu Wang, and Cihang Xie. An inverse scaling law for clip training. In NeurIPS, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "type": "text", + "content": "[42] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, Yuyin Zhou, and Cihang Xie. What if we recapture billions of web images with llama-3? arXiv preprint arXiv:2406.08478, 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "text", + "content": "[43] Yifan Li, Yifan Du, Kun Zhou, Jinping Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 471, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 524 + ], + "type": "text", + "content": "[44] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pre-training via masking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23390-23400, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 525, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 581 + ], + "type": "text", + "content": "[45] Zijing Liang, Yanjie Xu, Yifan Hong, Penghui Shang, Qi Wang, Qiang Fu, and Ke Liu. A survey of multimodel large language models. In Proceedings of the 3rd International Conference on Computer, Artificial Intelligence and Control Engineering, pages 405-409, 2024. 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[46] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "text", + "content": "[47] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.3,6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 691 + ], + "type": "text", + "content": "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[49] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He," + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20767" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European conference on computer vision, pages 216-233. Springer, 2024. 5, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[50] Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xu-Cheng Yin, Cheng-Lin Liu, Lianwen Jin, and Xiang Bai. Ocrbench: on the hidden mystery ofOCR in large multimodal models. Science China Information Sciences, 67(12):220102, 2024. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 228 + ], + "type": "text", + "content": "[51] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 285 + ], + "type": "text", + "content": "[52] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 295, + 341 + ], + "type": "text", + "content": "[53] Gen Luo, Xue Yang, Wenhan Dou, Zhaokai Wang, Jiawen Liu, Jifeng Dai, Yu Qiao, and Xizhou Zhu. Mono-internvl: Pushing the boundaries of monolithic multimodal large language models with endogenous visual pre-training. arXiv preprint arXiv:2410.08202, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 295, + 407 + ], + "type": "text", + "content": "[54] Sachin Mehta, Maxwell Horton, Fartash Faghri, Mohammad Hossein Sekhavat, Mahyar Najibi, Mehrdad Farajtabar, Oncel Tuzel, and Mohammad Rastegari. Catlip: Cliplevel visual recognition accuracy with 2.7 x faster pretraining on web-scale image-text data. arXiv preprint arXiv:2404.15653, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 409, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 295, + 430 + ], + "type": "text", + "content": "[55] Meta. Introducing meta llama 3: The most capable openly available llm to date, 2024. Accessed: 2024-04-18. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 432, + 261, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 261, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 261, + 443 + ], + "type": "text", + "content": "[56] OpenAI. Introducing chatgpt. OpenAI Blog, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 225, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 225, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 225, + 455 + ], + "type": "text", + "content": "[57] OpenAI. Gpt-4 technical report, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 248, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 248, + 467 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 248, + 467 + ], + "type": "text", + "content": "[58] OpenAI. Gpt-4v(ision) system card, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 468, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 533 + ], + "type": "text", + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 600 + ], + "type": "text", + "content": "[60] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 3, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 295, + 656 + ], + "type": "text", + "content": "[61] Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper, and Bryan Catanzaro. Megatronlm: Training multi-billion parameter language models using model parallelism. arXiv preprint arXiv:1909.08053, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[62] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 6, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 139 + ], + "type": "text", + "content": "[63] Daria Soboleva, Faisal Al-Khateeb, Robert Myers, Jacob R Steeves, Joel Hestness, and Nolan Dey. SlimPajama: A 627B token cleaned and deduplicated version of RedPajama. https://cerebras.ai/blog/slimpajama-a-627b-token-cleaned-and-deduplicated-version-of-redpajama, 2023.4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 171 + ], + "type": "text", + "content": "[64] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 554, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 554, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 554, + 226 + ], + "type": "text", + "content": "[65] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 293 + ], + "type": "text", + "content": "[66] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 554, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 337 + ], + "type": "text", + "content": "[67] Michael Tschannen, Manoj Kumar, Andreas Steiner, Xiaohua Zhai, Neil Houlsby, and Lucas Beyer. Image captioners are scalable vision learners too. Advances in Neural Information Processing Systems, 36:46830-46855, 2023. 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "text", + "content": "[68] Michael Tschannen, Manoj Kumar, Andreas Steiner, Xiaohua Zhai, Neil Houlsby, and Lucas Beyer. Image captioners are scalable vision learners too. Advances in Neural Information Processing Systems, 36:46830-46855, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 426 + ], + "type": "text", + "content": "[69] Haochen Wang, Anlin Zheng, Yucheng Zhao, Tiancai Wang, Ge Zheng, Xiangyu Zhang, and Zhaoxiang Zhang. Reconstructive visual instruction tuning. In International Conference on Learning Representations, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 426, + 554, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 554, + 491 + ], + "type": "text", + "content": "[70] Jiacong Wang, Bohong Wu, Haiyong Jiang, Zhou Xun, Xin Xiao, Haoyuan Guo, and Jun Xiao. World to code: Multimodal data generation via self-instructed compositional captioning and filtering. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4608-4623, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 554, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 554, + 558 + ], + "type": "text", + "content": "[71] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 554, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 554, + 591 + ], + "type": "text", + "content": "[72] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion feedback helps clip see better. arXiv preprint arXiv:2407.20171, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 593, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 554, + 645 + ], + "type": "text", + "content": "[73] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "text", + "content": "[74] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. SimVLM: Simple visual language model pretraining with weak supervision. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[75] X.ai. Grok-1.5 vision preview. https://x.ai/blog/grok-1.5v, 2024.5, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20768" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 520 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[76] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[77] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "text", + "content": "[78] Qiying Yu, Quan Sun, Xiaosong Zhang, Yufeng Cui, Fan Zhang, Yue Cao, Xinlong Wang, and Jingjing Liu. Capsfusion: Rethinking image-text data at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14022-14032, 2024. 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[79] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv preprint arXiv:2308.02490, 2023. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[80] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936, 2022. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 352 + ], + "type": "text", + "content": "[81] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 3, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 418 + ], + "type": "text", + "content": "[82] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 474 + ], + "type": "text", + "content": "[83] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 520 + ], + "type": "text", + "content": "[84] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1, 2, 3" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20769" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/76b9c942-da58-47e6-8ca6-ac0446eb367e_content_list.json b/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/76b9c942-da58-47e6-8ca6-ac0446eb367e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..445b538b6a5777b60bc0dd9768d89ba74e2799ad --- /dev/null +++ b/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/76b9c942-da58-47e6-8ca6-ac0446eb367e_content_list.json @@ -0,0 +1,1592 @@ +[ + { + "type": "text", + "text": "The Silent Assistant: NoiseQuery as Implicit Guidance for Goal-Driven Image Generation", + "text_level": 1, + "bbox": [ + 207, + 130, + 790, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruoyu Wang", + "bbox": [ + 142, + 202, + 254, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huayang Huang", + "bbox": [ + 297, + 203, + 434, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ye Zhu", + "bbox": [ + 478, + 203, + 547, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Olga Russakovsky", + "bbox": [ + 589, + 203, + 746, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu $\\mathbf{W}\\mathbf{u}^{1*}$", + "bbox": [ + 787, + 203, + 861, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Wuhan University $^{2}$ Department of Computer Science, Princeton University", + "bbox": [ + 276, + 238, + 720, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wangruoyu, hyhuang, wuyucs}@whu.edu.cn", + "bbox": [ + 326, + 277, + 669, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yezhu, olgarus}@princeton.edu", + "bbox": [ + 364, + 295, + 627, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we introduce NoiseQuery as a novel method for enhanced noise initialization in versatile goal-driven text-to-image (T2I) generation. Specifically, we propose to leverage an aligned Gaussian noise as implicit guidance to complement explicit user-defined inputs, such as text prompts, for better generation quality and controllability. Unlike existing noise optimization methods designed for specific models, our approach is grounded in a fundamental examination of the generic finite-step noise scheduler design in diffusion formulation, allowing better generalization across different diffusion-based architectures in a tuning-free manner. This model-agnostic nature allows us to construct a reusable noise library compatible with multiple T2I models and enhancement techniques, serving as a foundational layer for more effective generation. Extensive experiments demonstrate that NoiseQuery enables fine-grained control and yields significant performance boosts not only over high-level semantics but also over low-level visual attributes, which are typically difficult to specify through text alone, with seamless integration into current workflows with minimal computational overhead. Code is available at https://github.com/wangruoyu02/NoiseQuery.", + "bbox": [ + 86, + 378, + 483, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 741, + 220, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image synthesis (T2I) has gained significant research attention with the rising popularity of diffusion generative models [20, 39, 50, 51]. While these models demonstrate impressive generative capabilities, they still struggle to faithfully align generated content with user prompts, leading to suboptimal outputs. To enrich the text-based conditioning capabilities of diffusion models, researchers have", + "bbox": [ + 89, + 767, + 482, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6bf33591d212fda3a8f71277ee39ccc73e80d4093a26b2fa0f0da3a67e83c151.jpg", + "image_caption": [ + "Figure 1. Each row shows a set of images generated by different diffusion models starting from the same initial noise, using the same sampler, and with NULL as the text prompt. The generated images demonstrate striking both high-level semantic and low-level visual similarities across models." + ], + "image_footnote": [], + "bbox": [ + 517, + 340, + 903, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "explored various techniques, including fine-tuning diffusion UNet [2, 12, 53], enhancing language encoders [23, 67], refining inference strategies [3, 10], and extra additional image conditions [15, 45, 66]. A recent popular line of research focuses on optimizing the initial noise [4, 11, 18, 41], motivated by the observation that images generated from the same noise often exhibit high similarities within a single model, even when prompted differently. In this work, we extend this understanding by revealing that such noise-induced consistency persists across T2I models with different diffusion backbones, as illustrated in Fig. 1.", + "bbox": [ + 511, + 685, + 906, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To better understand this phenomenon, we examine the formulation of diffusion models and reveal that cross-model similarities stem from the finite-step noise scheduler design", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17618", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and persist across different diffusion variants. Specifically, one of the fundamental assumptions in diffusion models lies in a gradual stepwise perturbation process that transforms a real-world distribution into a standard Gaussian. This transition requires, in theory, an infinite number of diffusion steps to perfectly obscure the raw images, contradicting the practical implementation where we can only apply a finite number (i.e., non-asymptotic) of degradations through predefined noise schedulers [7, 28]. As a result, the real-world diffusion models inadvertently learn to leverage residual information embedded in image data, creating an unintended link between noise and data regardless of the model architecture during inference. However, rather than being a limitation, we leverage this seemingly imperfect model design and unintentional connection as an unexpected advantage by laying the foundation for using stochastic noise as a generic \"silent assistant\", which implicitly encodes clues about the image that will be generated.", + "bbox": [ + 89, + 90, + 483, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We note that it is critical to align the implicit noise clues with the user-provided input, such as text prompts, in order to achieve optimal generation performance. As shown in Fig. 2, when the inherent tendencies of the noise guidance and text prompt diverge significantly, the generation process leads to instability or semantically inconsistent results. Furthermore, unlike text prompts that convey high-level semantics, initial noise inherently encodes low-level visual primitives—such as color gradients, texture patterns, and spatial frequencies—directly tied to the pixel space. This makes it particularly suitable for controlling fine-grained visual properties that are difficult to specify through text alone. By leveraging initial noise as a complementary control signal, we produce more nuanced and precise outputs, enhancing the overall generative capability of T2I models.", + "bbox": [ + 89, + 367, + 483, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on this insight, we propose NoiseQuery, a novel method that exploits initial noise as a shared generative asset across diverse T2I models. The key challenge lies in establishing an interpretable mapping between stochastic noise patterns and their concrete generative clues. To bridge this gap, we utilize images generated with empty NULL text prompts as proxies to reveal the hidden information encoded in the initial noise (details in Sec. 3.3), denoted as the implicit generative posterior. When deprived of text guidance, the generation process becomes purely noise-driven, forcing the model to excavate and decode the residual signals preserved in the initial noise. This process enables the construction of a comprehensive Noise Library, built offline by generating a diverse set of such noise-driven images from a Gaussian distribution and extracting their multi-grained features, spanning high-level semantics and low-level visual attributes. During inference, given a specific user request (single or combined goals), we extract its features, perform a quick lookup in the pre-built library, and select the best-matching noise for initialization. Notably,", + "bbox": [ + 89, + 598, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a640fc88760112795e8545a15c6c692a469507d4d32647ae37f9e926c2d0e5a3.jpg", + "image_caption": [ + "Figure 2. Compared to mismatched noise, noise aligned with text prompts can produce more accurate images in complex T2I generation. A NULL prompt represents unconditional generation." + ], + "image_footnote": [], + "bbox": [ + 532, + 89, + 883, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NoiseQuery is model-agnostic and serves as a reusable resource across multiple diffusion architectures. This ability makes it naturally compatible with diverse enhancement strategies including recent noise optimization approaches, providing a foundational layer for more effective generation. Moreover, NoiseQuery operates at a negligible online computational cost (0.002 seconds per prompt), ensuring seamless integration into existing workflows while enhancing the controllability and quality of T2I generation.", + "bbox": [ + 511, + 387, + 906, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate the effectiveness of NoiseQuery, we conduct experiments on low-level property preferences, high-level semantic consistency, and their combinations. Extensive evaluations with six diffusion-based T2I models [6, 43, 49], four types of T2I enhancement methods [10, 11, 53, 67] and two controllable T2I models [15, 66] show that NoiseQuery not only improve the base model's performance but also complements multiple T2I enhancement methods. Our approach is also among the first to address the challenge of subtle low-level preference control with diffusion models, highlighting a new avenue for fine-grained generative tasks. Furthermore, by progressively matching the noise, our method is capable of supporting combinations of generation goals, offering more nuanced control for users.", + "bbox": [ + 511, + 523, + 908, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 532, + 734, + 834, + 748 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reveal the generic generative tendencies within the initial noise and leverage this implicit guidance as a universal generative asset that can be seamlessly integrated with existing T2I models and enhancement methods.", + "- We introduce NoiseQuery, a novel method that retrieves an optimal initial noise from a pre-built large-scale noise library, fulfilling versatile user-specified requirements encompassing both semantic and low-level visual attributes—a relatively under-explored area.", + "- We conduct extensive experiments to demonstrate that" + ], + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17619", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "our method can improve the general quality of diffusion-based generators and can significantly improve multiple T2I enhancement methods while maintaining a minimal cost and without requiring additional learning.", + "bbox": [ + 102, + 90, + 482, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 165, + 230, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Improving Base Text-to-Image Generation. Recent breakthroughs in text-to-image (T2I) generation have been driven by diffusion-based models, yet these models still suffer from artifacts and misalignment with textual prompts. Current solutions primarily fall into two categories: tuning-based and inference-time methods. Tuning-based methods [2, 12, 30, 53, 62] optimize model parameters through techniques like direct preference optimization (DPO) [30, 53, 62]. Additionally, LaVi-Bridge [67] enhances textual understanding of T2I models by integrating advanced language models with parameter-efficient adapters. In contrast, inference-time methods [3, 10, 21, 31, 69] dynamically adjust the generation process. For example, CFG++ [10] improves image quality by addressing non-manifold issues in classifier-free guidance. Our approach complements both strategies by offering improved initialization for generation. Enhancing T2I with Additional Conditions. To address the limitations of text-only conditions in meeting diverse user needs, numerous works [5, 8, 9, 13, 15, 40, 45, 59, 64, 66] have expanded T2I generation through multi-modal conditions. For instance, methods like ControlNet [64] introduce structural signals (e.g., edge and depth maps) to control spatial layouts, while IC-Light [65] adjusts the illumination effects. Meanwhile, customization generation approaches focus on learning user-provided concepts like identity [15, 45, 46, 55, 63], style [14, 54], and interaction [25]. In this paper, we highlight that the initial noise for diffusion sampling can also serve as a conditional resource. Furthermore, compared to efforts that focus primarily on high-level semantic control, we show that initial noise holds untapped potential for guiding various low-level visual properties (e.g., color, texture, and sharpness).", + "bbox": [ + 89, + 190, + 483, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Initial Noise Optimization for Diffusion Sampling. A growing body of research has focused on optimizing the initial noise in diffusion models, driven by the consensus that it significantly impacts the generated output [1, 24, 36, 56, 61]. These methods iteratively refine initial noise through backpropagation at inference, guided by different generative objectives, such as semantic alignment [4, 11, 18, 41, 52], rare concept generation [48] and image translation [17]. For instance, ReNO [11] optimizes noises using various human preference rewards like HPS v2 [58], to enhance image quality and align with user intent. However, these approaches rely on iterative gradient updates during test-time, which are computationally expensive and prone to numerical instability (e.g., gradient explosion). In this work, we observe that initial noise exhibits cross-model consistency", + "bbox": [ + 89, + 674, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/232390aa8abb55d73837b546e102a652f165cea1c9b189971bdfbcb28318e54d.jpg", + "image_caption": [ + "Figure 3. Using the same initial noise, the generative posteriors (left) exhibit similarity with the text-conditioned images (right)." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 903, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(Fig. 1), making it an exploitable shared asset. Leveraging this property, we construct a reusable noise library that not only facilitates efficient noise selection but also helps existing noise optimization methods improve generation quality.", + "bbox": [ + 511, + 315, + 906, + 377 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 387, + 648, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 412, + 715, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a pre-trained generative diffusion model $\\mathcal{M}$ , which is capable of generating images $\\mathcal{I}$ based on user-defined conditions $c$ , typically a text prompt. The generative process starts with the initial noise $\\epsilon$ sampled from a Gaussian distribution $\\mathcal{N}(0, I)$ and gradually transforms this noise into a clean image through $T$ denoising steps1. Our motivation is to construct a large-scale noise sample library, denoted as $\\mathcal{N}_{set} = \\{\\epsilon_1, \\epsilon_2, \\dots, \\epsilon_n\\}$ , where each noise sample $\\epsilon_i$ has its hidden features pre-computed from the generative posteriors. The goal is to identify the most suitable noise $\\epsilon^* \\in \\mathcal{N}_{set}$ that aligns with a specific generation objective $\\mathcal{O}$ . Notably, this generation objective is inherently open-ended and encompasses various dimensions such as semantic consistency, color properties, and other user-defined attributes.", + "bbox": [ + 511, + 434, + 906, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Implicit Guidance in Initial Noise", + "text_level": 1, + "bbox": [ + 511, + 667, + 805, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To investigate the influence of initial noise on generated images, we analyze the forward noise addition process in diffusion models [20, 37, 50, 51]. The forward diffusion process gradually perturbs an image $x_0$ over $T$ timesteps, transforming it into a \"pure\" Gaussian noise $x_T$ . The noise addition at each step follows:", + "bbox": [ + 511, + 690, + 905, + 780 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 789, + 903, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta_{t}$ is a predefined variances schedule. Defining $\\alpha_{t} = 1 - \\beta$ , the cumulative effect of noise over $t$ steps results in:", + "bbox": [ + 511, + 814, + 903, + 844 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 851, + 903, + 869 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1For simplicity, we omit the process of transitioning from pixel space to latent space in latent diffusion models (LDM).", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17620", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9403d1b910c1e857a0739e9aa87fbc8bc7ebecfbaf92e68ef57c475d2403bd3b.jpg", + "image_caption": [ + "Offline Library Construction", + "Figure 4. Illustration of our NoiseQuery pipeline which involves offline library construction and real-time inference." + ], + "image_footnote": [], + "bbox": [ + 153, + 104, + 844, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i$ and $\\epsilon \\sim \\mathcal{N}(0, I)$ . A fundamental assumption in diffusion theory is that $\\bar{\\alpha}_T \\to 0$ as $T \\to \\infty$ , ensuring that the raw image is fully obscured. However, practical implementations can only apply a finite number of degradations (e.g., $T = 1000$ ) to approximate this ideal. As analyzed in prior work [32], common schedules (e.g., Linear [20], Cosine [37], Stable Diffusion [43]) fail to achieve zero signal-to-noise ratio (SNR) at $T$ , leaving residual signal from $x_0$ in $x_T$ . For instance, Stable Diffusion's final timestep retains:", + "bbox": [ + 89, + 333, + 483, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {T} = 0. 0 6 8 2 6 5 \\cdot x _ {0} + 0. 9 9 7 6 6 7 \\cdot \\epsilon , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 496, + 480, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sqrt{\\overline{\\alpha}_T} = 0.068265$ quantifies the leakage. During training, diffusion models minimize reconstruction loss by utilizing these residuals as shortcuts rather than predicting the original image from scratch. This creates an unexpected link between noise and data, regardless of the model architecture. As a result, during inference, models continue to rely on the implicit knowledge gained during training to interpret pure Gaussian noise. We take advantage of this seemingly flawed model design and the unintended connection, using stochastic noise as a generic \"silent assistant\" that encodes clues about the image to be generated.", + "bbox": [ + 89, + 522, + 483, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dissect Initial Noise via Generation Posterior", + "text_level": 1, + "bbox": [ + 89, + 696, + 472, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While recent works [4, 11, 17, 18, 41] optimize initial noise via gradient back-propagation for text-guided generation, they often overlook the inherent generative tendencies embedded in different noises. In other words, if the initial noise deviates significantly from the desired objective, iterative optimization struggles to bridge this gap, resulting in suboptimal convergence despite extensive tuning. In this subsection, we propose a more efficient alternative: directly analyzing the initial noise itself without the text prompt. Additionally, our method also yields better starting points for optimization, leading to faster convergence.", + "bbox": [ + 89, + 719, + 483, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generative Posterior. When the model generates images", + "bbox": [ + 89, + 885, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1: NoiseQuery for Goal-Driven Generation", + "text_level": 1, + "bbox": [ + 514, + 338, + 875, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Diffusion model $\\mathcal{M}$ ; Noise library $\\mathcal{N}_{set} = \\{\\epsilon_1, \\epsilon_2, \\dots, \\epsilon_n\\}$ ; Generation objective $\\mathcal{O}$ and corresponding matching function $S(\\cdot)$ .", + "bbox": [ + 514, + 354, + 877, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: Optimal noise sample $\\epsilon^{*}$ for objective $\\mathcal{O}$ .", + "bbox": [ + 514, + 396, + 815, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Offline Library Construction", + "text_level": 1, + "bbox": [ + 514, + 412, + 705, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "foreach $\\epsilon_{i}\\in \\mathcal{N}_{set}$ do $\\mathcal{I}_i^{\\mathrm{uncond}} = \\mathcal{M}(\\epsilon_i,c = \\emptyset)$ Extract features $\\mathcal{F}_i$ (as detailed in Tab. 1) from $\\mathcal{I}_i^{\\mathrm{uncond}}$ Store $(\\epsilon_i,\\mathcal{F}_i)$ in the noise library.", + "bbox": [ + 514, + 428, + 869, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end", + "bbox": [ + 517, + 484, + 545, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Real-time Inference", + "text_level": 1, + "bbox": [ + 519, + 510, + 647, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Extract desired features $\\mathcal{F}_O$ from objective $\\mathcal{O}$ $\\epsilon^{*} = \\arg \\max_{\\epsilon_{i}\\in \\mathcal{N}_{set}}\\{S(\\mathcal{F}_{i},\\mathcal{F}_{O})\\mid \\epsilon_{i}\\in \\mathcal{N}_{set}\\}$", + "bbox": [ + 517, + 523, + 813, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "return $\\epsilon^{*}$", + "bbox": [ + 514, + 558, + 578, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "without any specific prompt, it depends solely on the noise characteristics, producing what we term \"generative posterior\". As shown in Fig. 3, generative posteriors exhibit clear generative tendencies (e.g., semantics, spatial layouts, and color palettes) similar to those of text-guided outputs. These similarities suggest that the initial noise carries significant latent information, influencing both unconditional generation and conditional generation in a similar manner, acting as an additional form of guidance alongside the text prompt. Thus, these purely noise-guided images can serve as an effective proxy for interpreting the initial noise.", + "bbox": [ + 511, + 609, + 906, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-model Consistency. We also observe another meaningful phenomenon: the impact of initial noise is universal and model-agnostic. As illustrated in Fig. 1, generative posteriors generated from the same noise remain highly similar across models with different architectures and parameters. For example, PixArt- $\\alpha$ [6] is a Transformer-based T2I diffusion model, while Stable Diffusion [43] is UNet-based. More discussion of models and samplers can be found in", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17621", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/db9bb8ae672a9ba00462cdf90948160b29c1459175165bfa04f8663b28c70e30.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Generation GoalsFeature TypeMatch Function
SemanticsCLIP [42], BLIP [29]Cosine Similarity
StyleGram Matrix [16]MSE
ColorRGB, HSV, LABAbsolute Difference
TextureGLCM [19]Euclidean Distance
ShapeHu Moments [22]Euclidean Distance
SharpnessHigh Frequency Energy (HFE)Absolute Difference
", + "bbox": [ + 94, + 88, + 496, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Feature types and matching functions.", + "bbox": [ + 145, + 188, + 426, + 203 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Appendix A. This finding empowers us to create a large-scale, model-agnostic noise library, where each noise sample can be associated with diverse latent features and then shared across different models. The universality of noise-driven priors positions initial noise as a unified control resource for various applications.", + "bbox": [ + 89, + 229, + 482, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. NoiseQuery for Goal-Driven Generation", + "text_level": 1, + "bbox": [ + 89, + 329, + 434, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building upon the insights regarding initial noise characteristics and cross-model consistency, we propose NoiseQuery, an efficient approach for selecting appropriate noise samples tailored to diverse generation objectives. As illustrated in Fig. 4, our method consists of two main phases: offline library construction and real-time inference. We also provide pseudo-code in Algorithm 1.", + "bbox": [ + 89, + 349, + 482, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Offline Library Construction. In the first stage, we build a noise library $\\mathcal{N}_{\\text {set }}$ , which contains many initial noise samples (e.g., 100k) sampled from standard Gaussian distribution. These samples are utilized to generate generative posteriors by running a diffusion model without any user-defined prompts. Subsequently, we extract a variety of features $\\mathcal{F}_i$ from these generative posteriors that correspond to various generative objectives $\\mathcal{O}$ , such as semantic representations (e.g., CLIP features), color properties (e.g., saturation), etc. These extracted features are then stored as keys alongside their corresponding noise samples (values) within the noise library, creating a comprehensive key-value repository that can be queried during the inference phase.", + "bbox": [ + 89, + 455, + 482, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Real-time Inference. When performing real-time inference, we specify desired features $\\mathcal{F}_O$ following the user's generation objective $\\mathcal{O}$ . These features could be semantic, color, spatial, or style-based attributes, depending on the specific goal. Using a match function $S(\\cdot)$ , we evaluate the alignment between each feature $\\mathcal{F}_i$ in our pre-computed library and the objective features $\\mathcal{F}_O$ . The optimal noise sample $\\epsilon^*$ is selected by maximizing the matching scores:", + "bbox": [ + 89, + 652, + 482, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\epsilon^ {*} = \\arg \\max _ {\\epsilon_ {i} \\in \\mathcal {N} _ {s e t}} \\left\\{S \\left(\\mathcal {F} _ {i}, \\mathcal {F} _ {O}\\right) \\mid \\epsilon_ {i} \\in \\mathcal {N} _ {s e t} \\right\\}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 780, + 482, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the optimal noise sample $\\epsilon^{*}$ is identified, it is fed into the generative model to produce an image that closely matches the user intent. In Tab. 1, we summarize several common generation goals achievable with NoiseQuery, including both high-level semantics and low-level visual properties. For each goal, we specify the corresponding", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "feature types and matching functions. Detailed explanations of these features can be found in Appendix C. Notably, these goals can be applied individually or in combination. Users can specify multiple constraints simultaneously (e.g., generating a \"sunset\" image with a \"warm color palette\" while ensuring \"high sharpness\"). This flexibility makes NoiseQuery highly versatile and generalizable, enabling fine-grained control over the generated output.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 227, + 645, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 253, + 702, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our approach on two popular benchmarks: MSCOCO [33] and DrawBench [47]. MSCOCO is a standard benchmark for text-to-image generation, providing textual descriptions of diverse natural scenes and object classes. We randomly select 10k captions as user prompts. DrawBench consists of 200 carefully designed prompts intended to challenge generative scenarios on various aspects, including long-form text, complex compositions, conflicting instructions, and rare vocabulary.", + "bbox": [ + 511, + 275, + 903, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model Zoo. We conduct experiments on six popular diffusion generation models, including Stable diffusion (SD) series [43] (1.4, 1.5, 2.0, 2.1), PixArt- $\\alpha$ [6], and SD-Turbo [49]. Stable Diffusion is a Latent Diffusion Model with UNet [44] backbone. By default, we use 50 DDIM denoising steps with a CFG scale of 7.0. PixArt- $\\alpha$ is a Transformer Latent Diffusion Model, and SD-Turbo is a distillation model based on Stable Diffusion 2.1, generating high-quality images even with only 1-2 sampling steps without CFG. For PixArt- $\\alpha$ , we use the default CFG scale of 4.5.", + "bbox": [ + 511, + 412, + 903, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. For semantic similarity, we use ClipScore [42], which directly measures the alignment between the generated images and the corresponding textual prompts. Additionally, we incorporate several popular metrics trained on large-scale human preference data to quantify human preferences for AI-generated images, including HPS v2 [58], PickScore [27], and ImageReward [60]. These metrics offer a thorough evaluation of image quality, aesthetics, and alignment with user intent.", + "bbox": [ + 511, + 563, + 905, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Improvement on High-level Semantics", + "text_level": 1, + "bbox": [ + 511, + 710, + 841, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For semantic consistency goals, we utilize features like CLIP [42] and BLIP [29] as the query to identify the semantically matched noise for each text prompt. Such aligned text and noise yield two benefits: (i) prevents the model from generating irrelevant outputs or falling into unsolvable regions. (ii) minimizes generation difficulty and maximizes the likelihood for semantic consistency.", + "bbox": [ + 511, + 733, + 903, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Analysis. In Tab. 2, we evaluate our method on standard text-to-image (T2I) benchmarks, demonstrating consistent performance gains across all base models by offering better noise-text alignment. NoiseQuery", + "bbox": [ + 511, + 839, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17622", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b2a6fc8ebbb9800730de9f25559782c0858391fade64feb040d23b9a2e471133.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Base ModelMethodDrawBench [47]MSCOCO [33]Time Cost
ImageRewardPickScoreHPS v2CLIPScoreImageRewardPickScoreHPS v2CLIPScore
SD 1.5Base Model0.0421.1124.5730.900.1521.4125.6531.081.334 s
+ NoiseQuery0.0821.1625.0231.410.2721.4826.0731.471.336 s
+ Diffusion-DPO [53]0.0921.2925.0231.190.2521.6426.3131.261.350 s
+ Diffusion-DPO [53] + NoiseQuery0.1721.3325.2531.410.3521.6826.6031.551.352 s
SD 2.1Base Model0.1221.3324.9331.130.3621.7226.5831.401.301 s
+ NoiseQuery0.2621.4625.3931.680.4421.7626.8231.501.303 s
+ CFG++ [10]0.1221.3324.8331.130.3721.7226.6631.313.724 s
+ CFG++ [10] + NoiseQuery0.2721.4325.5531.610.4721.7626.9731.673.726 s
SD-TurboBase Model0.2621.7825.2331.290.4722.0726.2231.510.072 s
+ NoiseQuery0.4121.8725.6631.580.5022.1726.8231.760.074 s
+ ReNO [11]1.6723.4032.4832.55----23.56 s
+ ReNO [11] + NoiseQuery1.7123.5232.9232.78----23.56 s
PixArt-αBase Model0.7022.0828.2730.830.7822.2429.3331.484.327 s
+ NoiseQuery0.8222.1128.4531.270.7922.3329.5631.644.328 s
+ LaVi-Bridge [67]0.6322.0828.3530.920.7522.3129.4931.865.092 s
+ LaVi-Bridge [67] + NoiseQuery0.7222.2428.6131.350.7822.3529.6732.015.094 s
", + "bbox": [ + 94, + 88, + 903, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Evaluation of objective metrics on different datasets and models. Higher is better for all metrics. Our approach enhances the base model's performance and complements a wide range of T2I enhancement methods, including DPO [53] (reward-based fine-tuning), CFG++ [10] (guidance refinement), ReNO [11] (gradient-based initial noise optimization), and LaVi-Bridge [67] (text encoding enhancement via LORA/adapters). $\\cdot$ : ReNO results on MSCOCO are excluded due to excessive inference time.", + "bbox": [ + 89, + 319, + 906, + 376 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/507911a5084e97e7d2c18ebdec3bacd774e3943bfecd99089ef7ffa969a75f86.jpg", + "image_caption": [ + "Figure 5. Generated images of the base model, base model with T2I enhancement methods, and base model with enhancement and NoiseQuery. The four rows, from top to bottom, respectively correspond to the method groups listed in Table 2: Diffusion-DPO [53], CFG++ [10], ReNO [11], and LaVi-Bridge [67]." + ], + "image_footnote": [], + "bbox": [ + 91, + 387, + 480, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "also serves as a universal foundation layer, complementing T2I enhancement techniques like Diffusion-DPO [53], CFG++ [10], ReNO [11], and LaVi-Bridge [67]. Notably, these methods face inherent limitations: model-specific dependencies (e.g., Diffusion-DPO and LaVi-Bridge require retraining for new models; ReNO is restricted to one-step models due to heavy computation, excluding com", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f014cb793cc04e6966c74a1b523c31602db45805889a79962dfe3d5bd2f12bd.jpg", + "image_caption": [ + "Figure 6. Queried noise v.s. random noise." + ], + "image_footnote": [], + "bbox": [ + 532, + 388, + 883, + 518 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9786536d46520ccf67ca7e9fb178aefe4c79983cad0fdf1da93dd27f3d668e24.jpg", + "image_caption": [ + "Figure 7. CLIP and FID evaluations across CFG scales on MSCOCO [33], with the CFG scale increasing from left to right." + ], + "image_footnote": [], + "bbox": [ + 531, + 542, + 883, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mon multi-step models like SD) and high latency. In contrast, our method seamlessly integrates with them, enhancing their performance while adding near-zero time costs. We also present qualitative results in Fig. 5 to illustrate its effectiveness in assisting text-to-image synthesis. Additionally, we show the zero-shot transferability of NoiseQuery across various generative models in Appendix B.1. This model-agnostic compatibility, combined with minimal computational cost, establishes NoiseQuery as a versatile and efficient enhancement tool for diverse pipelines.", + "bbox": [ + 511, + 715, + 906, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diversity Analysis. We report the performance of selecting the top 20 matched noise samples per prompt instead of", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17623", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3829f9940f18480604f7a8e324a6b834b2ff3bd5ecc49dc4d023673efa04d36c.jpg", + "image_caption": [ + "(a) Color properties guidance." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 486, + 362 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d182bdc33a273b2555ab7508296ae040025e02acf5f3f67e8c4ff71962ca9f65.jpg", + "image_caption": [ + "(b) Structural features guidance." + ], + "image_footnote": [], + "bbox": [ + 506, + 88, + 901, + 362 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1e45f2df2667c30c560b88ca7202b6a4dd5602225639d8a07ff488a09ee1070e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Library Size0.5k1k2k5k10k50k100k
Matching Function Cost (×10-4s)1.391.391.391.391.401.401.51
Argmax Selection Cost (×10-4s)0.250.250.250.250.376.1613.25
CLIP Score31.5131.5331.5731.5931.6631.7331.74
", + "bbox": [ + 94, + 429, + 483, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Retrieval time breakdown across various library sizes.", + "bbox": [ + 99, + 487, + 472, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the single best match. As shown in Fig. 6, these samples consistently outperform the standard random noise baseline on DrawBench [47], demonstrating that NoiseQuery provides not only optimal but also diverse high-quality noise candidates. To quantify diversity, we use the Diversity Metric (DIV) [57], computed as the standard deviation of DINOV2 [38] embeddings across 20 generated images per prompt (see Fig. 6 legend). Our method achieves comparable diversity to random noise, confirming that NoiseQuery enhances image quality without sacrificing diversity.", + "bbox": [ + 88, + 532, + 482, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Different Classifier-free Guidance Scale Analysis. We assess our method's performance across various classifier-free guidance (CFG) scales (1.5, 2.0, 3.0, ..., 15.0) on 10k MSCOCO [33] samples. Our NoiseQuery achieves high-quality outputs even with low guidance scales, as demonstrated in Fig. 7. This supports our hypothesis that by selecting more compatible initial noise, we can effectively reduce the generation difficulty. Therefore, we can obtain high-quality output at low scales and avoid the over-saturation and instability that usually come with high scales.", + "bbox": [ + 88, + 686, + 482, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Time Cost. We also provide a cost breakdown for the retrieval process in Tab. 3, which includes the computation of the matching function (utilizing query CLIP features at no extra cost since CLIP is embedded in SD) and argmax se", + "bbox": [ + 88, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f31b3e9b0aa0fa9e7f2001036b111dd3d953ec3909ae0044680d96629130150.jpg", + "image_caption": [ + "Figure 8. Multiple low-level visual properties control through NoiseQuery. All images use minimal text prompts containing only semantic concepts (e.g., object labels), lacking explicit visual specifications.", + "Figure 9. Text prompts often fail to align generated images with user low-level attribute preferences, while initial noise plays a dominant role in controlling them." + ], + "image_footnote": [], + "bbox": [ + 513, + 417, + 906, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lection. The results demonstrate that both steps are efficient, even with large libraries (e.g., 0.0015 seconds for 100,000 entries), indicating that the additional cost is negligible.", + "bbox": [ + 511, + 579, + 906, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Controllability on Low-Level Visual Properties", + "text_level": 1, + "bbox": [ + 511, + 636, + 903, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text prompts effectively steer high-level semantics in text-to-image (T2I) generation but often fail to provide sufficient control over low-level visual properties that affect visual perception. This limitation stems from two factors: (1) text is inherently inadequate for conveying detailed visual reference input, and (2) text encoders like CLIP [42], trained primarily on high-level semantics, struggle to capture and control fine-grained low-level properties such as color, texture, shape, and sharpness.", + "bbox": [ + 509, + 657, + 906, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In contrast, initial noise plays a dominant role in determining these properties, making it a valuable resource for precise control, as illustrated in Fig. 9. Our method leverages this by enabling fine-grained control of low-level attributes through initial noise, without requiring additional training. The following experiments demonstrate the effectiveness of our approach in achieving these targets, with fur", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17624", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7a2d3de948f45cbf79e86d5414c8a7595de0bebb8affc1d444dd8563c462966.jpg", + "image_caption": [ + "Diverse User Demands", + "Stable Diffusion", + "Stable Diffusion (enhanced prompt)", + "NoiseQuery (ours)", + "Figure 10. Examples of combining high-level semantics (red) and low-level properties (blue). Enhanced prompts are created by fusing semantic objects and visual details via a large language model." + ], + "image_footnote": [], + "bbox": [ + 93, + 108, + 480, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02bc7ed3fb8653715ab9059a20cbc8579d9479ccf964c508d5655e57e8ffff1a.jpg", + "image_caption": [ + "Figure 11. Compatibility with controllable models such as TI [15] and ControlNet [64]. Our method provides a better initial noise, enhancing semantic consistency and low-level attribute control." + ], + "image_footnote": [], + "bbox": [ + 91, + 335, + 482, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ther implementation details provided in Appendix C.", + "bbox": [ + 89, + 584, + 439, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Color Properties. Color attributes such as brightness, saturation, contrast, and hue are crucial for an image's visual appeal. We assess the overall color hue by averaging each RGB channel, measure saturation and brightness in the HSV color space, and compute contrast as the standard deviation of grayscale intensities. As shown in Fig. 8a, selecting initial noise that matches target color distributions enables precise control over the generated images' color tendencies, aligning them with user preferences.", + "bbox": [ + 89, + 606, + 483, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Structural Features. For spatial control, we implement: (1) Texture through Gray-Level Co-occurrence Matrix (GLCM) [19] features that quantify intensity co-occurrence patterns; (2) Shape consistency via Hu Moments [22] as rotation-invariant descriptors; and (3) Sharpness by prioritizing high-frequency energy (HFE) in Fourier spectra. Fig. 8b demonstrates how noise selection governs these structural features—preserving geometric fidelity in object shapes, modulating surface textures, and enhancing edge crispness—all achieved without prompt engineering.", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Enhancement on Combined Scenarios", + "text_level": 1, + "bbox": [ + 513, + 90, + 841, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Building upon the results in previous sections, which demonstrate our method's effectiveness in high-level semantic alignment and low-level attribute control, this section highlights its ability to combine these diverse goals and seamlessly integrate with powerful existing controllable generation models.", + "bbox": [ + 511, + 114, + 905, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Combined Goals. Beyond individual tasks, our method excels in simultaneously achieving multiple generation goals. By sequentially reranking noise candidates—first selecting top candidates based on one goal (e.g., high-level semantics) and then refining within this subset for another goal (e.g., low-level attributes)—we progressively narrow down the optimal noise pool. As shown in Fig. 10, this approach allows the combination of high-level semantic alignment with various low-level attribute controls, enabling the generation of images that meet diverse user preferences.", + "bbox": [ + 511, + 209, + 906, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compatibility with Controlling Methods. Our method is designed to be instant-time and model-agnostic, making it highly compatible with existing controllable generation models such as Textual Inversion (TI) [15] and ControlNet [64]. By providing a better initial noise, our approach establishes a foundational layer for controllable generation, which synergizes with subsequent refinement processes. As shown in Fig. 11, this integration enhances semantic consistency while enabling precise control over challenging low-level attributes, such as color and texture.", + "bbox": [ + 511, + 363, + 908, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 537, + 633, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose NoiseQuery, a tuning-free method that enables versatile goal-driven T2I generation through enhanced noise initialization, supported by a generic and reusable pre-built noise library. By selecting Gaussian initialization in the noise library better aligned with the text prompt or specified attributes, NoiseQuery achieves fine-grained control over both high-level semantic and intricate low-level details. In addition, the model-agnostic nature of our method allows it to be seamlessly integrated with existing T2I augmentation and controllable generation methods without model adjustments or significant computational overhead. Experiments validate its effectiveness in improving output fidelity and precision, particularly for visual attributes that are challenging to articulate textually. The lightweight and scalability of the framework make it adaptable to real-world applications.", + "bbox": [ + 511, + 565, + 906, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although NoiseQuery provides flexibility for aligning high-level semantics and controlling low-level attributes, it remains bounded by the base model's ability, and may still underperform in extremely complex cases. Additionally, it has limitations in achieving fine-grained control, such as Canny edge control.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17625", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the National Natural Science Foundation of China under grant 62372341. This work was also partially supported by an Amazon Research Award 2023 to OR and YZ. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not reflect the views of Amazon.", + "bbox": [ + 89, + 114, + 485, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 243, + 187, + 258 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Donghoon Ahn, Jiwon Kang, Sanghyun Lee, Jaewon Min, Minjae Kim, Wooseok Jang, Hyoungwon Cho, Sayak Paul, SeonHwa Kim, Eunju Cha, et al. A noise is worth diffusion guidance. arXiv, 2024. 3", + "[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In ICLR, 2024. 1, 3", + "[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG), 2023. 1, 3", + "[4] Changgu Chen, Libing Yang, Xiaoyan Yang, Lianggangxu Chen, Gaoqi He, Changbo Wang, and Yang Li. Find: Finetuning initial noise distribution with policy optimization for diffusion models. In ACM MM, 2024. 1, 3, 4", + "[5] Haoyang Chen, Dongfang Sun, Caoyuan Ma, Shiqin Wang, Kewei Zhang, Zheng Wang, and Zhixiang Wang. Subjective camera: Bridging human cognition and visual reconstruction through sequence-aware sketch-guided diffusion. arXiv, 2025. 3", + "[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2, 4, 5, 1", + "[7] Sitan Chen, Giannis Daras, and Alex Dimakis. Restoration-degradation beyond linear diffusions: A non-asymptotic analysis for ddim-type samplers. In ICML, 2023. 2", + "[8] Zijiao Chen, Jiaxin Qing, Tiange Xiang, Wan Lin Yue, and Juan Helen Zhou. Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding. In CVPR, 2023. 3", + "[9] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv, 2023. 3", + "[10] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv, 2024. 1, 2, 3, 6", + "[11] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. NeurIPS, 37:125487-125519, 2024. 1, 2, 3, 4, 6" + ], + "bbox": [ + 93, + 267, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. NeurIPS, 2023. 1, 3", + "[13] Yingying Fan, Quanwei Yang, Kaisiyuan Wang, Hang Zhou, Yingying Li, Haocheng Feng, Errui Ding, Yu Wu, and Jingdong Wang. Re-hold: Video hand object interaction reenactment via adaptive layout-instructed diffusion model. In CVPR, 2025. 3", + "[14] Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. In ECCV, 2024. 3", + "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 1, 2, 3, 8", + "[16] Leon A Gatys. A neural algorithm of artistic style. arXiv, 2015. 5", + "[17] Or Greenberg, Eran Kishon, and Dani Lischinski. S2st: Image-to-image translation in the seed space of latent diffusion. arXiv, 2023. 3, 4", + "[18] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024. 1, 3, 4", + "[19] Robert M Haralick, Karthikeyan Shanmugam, and Its' Hak Dinstein. Textural features for image classification. IEEE SMCS, 1973. 5, 8, 2", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 1, 3, 4", + "[21] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In ICCV, 2023. 3", + "[22] Ming-Kuei Hu. Visual pattern recognition by moment invariants. IRE transactions on information theory, 1962. 5, 8, 2", + "[23] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv, 2024. 1", + "[24] Huayang Huang, Yu Wu, and Qian Wang. Robin: Robust and invisible watermarks for diffusion models with adversarial optimization. NIPS, 2024. 3", + "[25] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan, and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. In SIGGRAPH Asia, 2024. 3", + "[26] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. 1", + "[27] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. NeurIPS, 2023. 5", + "[28] Gen Li, Yuting Wei, Yuxin Chen, and Yuejie Chi. Towards non-asymptotic convergence for diffusion-based generative models. In ICLR, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17626", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML. PMLR, 2022. 5", + "[30] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. NeurIPS, 2024. 3", + "[31] Yumeng Li, Margret Keuper, Dan Zhang, and Anna Khoreva. Divide & bind your attention for improved generative semantic nursing. arXiv, 2023. 3", + "[32] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In WACV, 2024. 4, 5", + "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 6, 7, 1", + "[34] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In ICLR, 2022. 1", + "[35] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. NeurIPS, 2022. 1", + "[36] Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, et al. Inference-time scaling for diffusion models beyond scaling denoising steps. arXiv, 2025. 3", + "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3, 4", + "[38] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024. 7", + "[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 1", + "[40] Yuhan Pei, Ruoyu Wang, Yongqi Yang, Ye Zhu, Olga Russakovsky, and Yu Wu. Sowing information: Cultivating contextual coherence with mllms in image generation. arXiv, 2024. 3", + "[41] Zipeng Qi, Lichen Bai, Haoyi Xiong, et al. Not all noises are created equally: Diffusion noise selection and optimization. arXiv, 2024. 1, 3, 4", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML. PMLR, 2021. 5, 7", + "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4, 5, 1", + "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI. Springer, 2015. 5" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 3", + "[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, 2024. 3", + "[47] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 5, 6, 7, 1", + "[48] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In AAAI, 2024. 3", + "[49] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 2, 5, 1", + "[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 3", + "[51] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 3", + "[52] Zhiwei Tang, Jiangweizhi Peng, Jiasheng Tang, Mingyi Hong, Fan Wang, and Tsung-Hui Chang. Inference-time alignment of diffusion models with direct noise optimization. arXiv, 2024. 3", + "[53] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, pages 8228-8238, 2024. 1, 2, 3, 6", + "[54] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv, 2024. 3", + "[55] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv, 2024. 3", + "[56] Ruoyu Wang, Yongqi Yang, Zhihao Qian, Ye Zhu, and Yu Wu. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, 2024. 3", + "[57] Xi Wang, Nicolas Dufour, Nefeli Andreou, Marie-Paule Cani, Victoria Fernández Abrevaya, David Picard, and Vicky Kalogeiton. Analysis of classifier-free guidance weight schedulers. arXiv, 2024. 7", + "[58] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv, 2023. 3, 5", + "[59] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. *IJCV*, 2024. 3" + ], + "bbox": [ + 514, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17627", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. NeurIPS, 2024. 5", + "[61] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models. arXiv, 2024. 3", + "[62] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3", + "[63] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv, 2023. 3", + "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 8", + "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Ic-light github page, 2024. 3", + "[66] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2024. 1, 2, 3", + "[67] Shihao Zhao, Shaozhe Hao, Bojia Zi, Huaizhe Xu, and Kwan-Yee K Wong. Bridging different language models and generative vision models for text-to-image generation. In ECCV, 2024. 1, 2, 3, 6", + "[68] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. NeurIPS, 2024. 1", + "[69] Chenyi Zhuang, Ying Hu, and Pan Gao. Magnet: We never know how text-to-image diffusion models work, until we learn how vision-language models function. NeurIPS, 2024. 3" + ], + "bbox": [ + 91, + 92, + 480, + 571 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17628", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/76b9c942-da58-47e6-8ca6-ac0446eb367e_model.json b/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/76b9c942-da58-47e6-8ca6-ac0446eb367e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a0211d7491a4af5ff42d2c9a473ea74a9e946763 --- /dev/null +++ b/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/76b9c942-da58-47e6-8ca6-ac0446eb367e_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.208, + 0.131, + 0.792, + 0.177 + ], + "angle": 0, + "content": "The Silent Assistant: NoiseQuery as Implicit Guidance for Goal-Driven Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.143, + 0.203, + 0.256, + 0.223 + ], + "angle": 0, + "content": "Ruoyu Wang" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.204, + 0.436, + 0.223 + ], + "angle": 0, + "content": "Huayang Huang" + }, + { + "type": "text", + "bbox": [ + 0.48, + 0.204, + 0.549, + 0.221 + ], + "angle": 0, + "content": "Ye Zhu" + }, + { + "type": "text", + "bbox": [ + 0.59, + 0.204, + 0.747, + 0.223 + ], + "angle": 0, + "content": "Olga Russakovsky" + }, + { + "type": "text", + "bbox": [ + 0.789, + 0.204, + 0.862, + 0.221 + ], + "angle": 0, + "content": "Yu \\(\\mathbf{W}\\mathbf{u}^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.239, + 0.721, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science, Wuhan University \\(^{2}\\)Department of Computer Science, Princeton University" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.278, + 0.671, + 0.293 + ], + "angle": 0, + "content": "{wangruoyu, hyhuang, wuyucs}@whu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.296, + 0.629, + 0.31 + ], + "angle": 0, + "content": "{yezhu, olgarus}@princeton.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.379, + 0.485, + 0.711 + ], + "angle": 0, + "content": "In this work, we introduce NoiseQuery as a novel method for enhanced noise initialization in versatile goal-driven text-to-image (T2I) generation. Specifically, we propose to leverage an aligned Gaussian noise as implicit guidance to complement explicit user-defined inputs, such as text prompts, for better generation quality and controllability. Unlike existing noise optimization methods designed for specific models, our approach is grounded in a fundamental examination of the generic finite-step noise scheduler design in diffusion formulation, allowing better generalization across different diffusion-based architectures in a tuning-free manner. This model-agnostic nature allows us to construct a reusable noise library compatible with multiple T2I models and enhancement techniques, serving as a foundational layer for more effective generation. Extensive experiments demonstrate that NoiseQuery enables fine-grained control and yields significant performance boosts not only over high-level semantics but also over low-level visual attributes, which are typically difficult to specify through text alone, with seamless integration into current workflows with minimal computational overhead. Code is available at https://github.com/wangruoyu02/NoiseQuery." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.742, + 0.222, + 0.758 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Text-to-image synthesis (T2I) has gained significant research attention with the rising popularity of diffusion generative models [20, 39, 50, 51]. While these models demonstrate impressive generative capabilities, they still struggle to faithfully align generated content with user prompts, leading to suboptimal outputs. To enrich the text-based conditioning capabilities of diffusion models, researchers have" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.342, + 0.905, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.597, + 0.908, + 0.667 + ], + "angle": 0, + "content": "Figure 1. Each row shows a set of images generated by different diffusion models starting from the same initial noise, using the same sampler, and with NULL as the text prompt. The generated images demonstrate striking both high-level semantic and low-level visual similarities across models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.907, + 0.852 + ], + "angle": 0, + "content": "explored various techniques, including fine-tuning diffusion UNet [2, 12, 53], enhancing language encoders [23, 67], refining inference strategies [3, 10], and extra additional image conditions [15, 45, 66]. A recent popular line of research focuses on optimizing the initial noise [4, 11, 18, 41], motivated by the observation that images generated from the same noise often exhibit high similarities within a single model, even when prompted differently. In this work, we extend this understanding by revealing that such noise-induced consistency persists across T2I models with different diffusion backbones, as illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To better understand this phenomenon, we examine the formulation of diffusion models and reveal that cross-model similarities stem from the finite-step noise scheduler design" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17618" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.364 + ], + "angle": 0, + "content": "and persist across different diffusion variants. Specifically, one of the fundamental assumptions in diffusion models lies in a gradual stepwise perturbation process that transforms a real-world distribution into a standard Gaussian. This transition requires, in theory, an infinite number of diffusion steps to perfectly obscure the raw images, contradicting the practical implementation where we can only apply a finite number (i.e., non-asymptotic) of degradations through predefined noise schedulers [7, 28]. As a result, the real-world diffusion models inadvertently learn to leverage residual information embedded in image data, creating an unintended link between noise and data regardless of the model architecture during inference. However, rather than being a limitation, we leverage this seemingly imperfect model design and unintentional connection as an unexpected advantage by laying the foundation for using stochastic noise as a generic \"silent assistant\", which implicitly encodes clues about the image that will be generated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.484, + 0.595 + ], + "angle": 0, + "content": "We note that it is critical to align the implicit noise clues with the user-provided input, such as text prompts, in order to achieve optimal generation performance. As shown in Fig. 2, when the inherent tendencies of the noise guidance and text prompt diverge significantly, the generation process leads to instability or semantically inconsistent results. Furthermore, unlike text prompts that convey high-level semantics, initial noise inherently encodes low-level visual primitives—such as color gradients, texture patterns, and spatial frequencies—directly tied to the pixel space. This makes it particularly suitable for controlling fine-grained visual properties that are difficult to specify through text alone. By leveraging initial noise as a complementary control signal, we produce more nuanced and precise outputs, enhancing the overall generative capability of T2I models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Building on this insight, we propose NoiseQuery, a novel method that exploits initial noise as a shared generative asset across diverse T2I models. The key challenge lies in establishing an interpretable mapping between stochastic noise patterns and their concrete generative clues. To bridge this gap, we utilize images generated with empty NULL text prompts as proxies to reveal the hidden information encoded in the initial noise (details in Sec. 3.3), denoted as the implicit generative posterior. When deprived of text guidance, the generation process becomes purely noise-driven, forcing the model to excavate and decode the residual signals preserved in the initial noise. This process enables the construction of a comprehensive Noise Library, built offline by generating a diverse set of such noise-driven images from a Gaussian distribution and extracting their multi-grained features, spanning high-level semantics and low-level visual attributes. During inference, given a specific user request (single or combined goals), we extract its features, perform a quick lookup in the pre-built library, and select the best-matching noise for initialization. Notably," + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.09, + 0.885, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.319, + 0.908, + 0.362 + ], + "angle": 0, + "content": "Figure 2. Compared to mismatched noise, noise aligned with text prompts can produce more accurate images in complex T2I generation. A NULL prompt represents unconditional generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.524 + ], + "angle": 0, + "content": "NoiseQuery is model-agnostic and serves as a reusable resource across multiple diffusion architectures. This ability makes it naturally compatible with diverse enhancement strategies including recent noise optimization approaches, providing a foundational layer for more effective generation. Moreover, NoiseQuery operates at a negligible online computational cost (0.002 seconds per prompt), ensuring seamless integration into existing workflows while enhancing the controllability and quality of T2I generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.909, + 0.735 + ], + "angle": 0, + "content": "To validate the effectiveness of NoiseQuery, we conduct experiments on low-level property preferences, high-level semantic consistency, and their combinations. Extensive evaluations with six diffusion-based T2I models [6, 43, 49], four types of T2I enhancement methods [10, 11, 53, 67] and two controllable T2I models [15, 66] show that NoiseQuery not only improve the base model's performance but also complements multiple T2I enhancement methods. Our approach is also among the first to address the challenge of subtle low-level preference control with diffusion models, highlighting a new avenue for fine-grained generative tasks. Furthermore, by progressively matching the noise, our method is capable of supporting combinations of generation goals, offering more nuanced control for users." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.736, + 0.835, + 0.749 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.75, + 0.905, + 0.81 + ], + "angle": 0, + "content": "- We reveal the generic generative tendencies within the initial noise and leverage this implicit guidance as a universal generative asset that can be seamlessly integrated with existing T2I models and enhancement methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.886 + ], + "angle": 0, + "content": "- We introduce NoiseQuery, a novel method that retrieves an optimal initial noise from a pre-built large-scale noise library, fulfilling versatile user-specified requirements encompassing both semantic and low-level visual attributes—a relatively under-explored area." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.887, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- We conduct extensive experiments to demonstrate that" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17619" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "our method can improve the general quality of diffusion-based generators and can significantly improve multiple T2I enhancement methods while maintaining a minimal cost and without requiring additional learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.166, + 0.232, + 0.181 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.191, + 0.484, + 0.675 + ], + "angle": 0, + "content": "Improving Base Text-to-Image Generation. Recent breakthroughs in text-to-image (T2I) generation have been driven by diffusion-based models, yet these models still suffer from artifacts and misalignment with textual prompts. Current solutions primarily fall into two categories: tuning-based and inference-time methods. Tuning-based methods [2, 12, 30, 53, 62] optimize model parameters through techniques like direct preference optimization (DPO) [30, 53, 62]. Additionally, LaVi-Bridge [67] enhances textual understanding of T2I models by integrating advanced language models with parameter-efficient adapters. In contrast, inference-time methods [3, 10, 21, 31, 69] dynamically adjust the generation process. For example, CFG++ [10] improves image quality by addressing non-manifold issues in classifier-free guidance. Our approach complements both strategies by offering improved initialization for generation. Enhancing T2I with Additional Conditions. To address the limitations of text-only conditions in meeting diverse user needs, numerous works [5, 8, 9, 13, 15, 40, 45, 59, 64, 66] have expanded T2I generation through multi-modal conditions. For instance, methods like ControlNet [64] introduce structural signals (e.g., edge and depth maps) to control spatial layouts, while IC-Light [65] adjusts the illumination effects. Meanwhile, customization generation approaches focus on learning user-provided concepts like identity [15, 45, 46, 55, 63], style [14, 54], and interaction [25]. In this paper, we highlight that the initial noise for diffusion sampling can also serve as a conditional resource. Furthermore, compared to efforts that focus primarily on high-level semantic control, we show that initial noise holds untapped potential for guiding various low-level visual properties (e.g., color, texture, and sharpness)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Initial Noise Optimization for Diffusion Sampling. A growing body of research has focused on optimizing the initial noise in diffusion models, driven by the consensus that it significantly impacts the generated output [1, 24, 36, 56, 61]. These methods iteratively refine initial noise through backpropagation at inference, guided by different generative objectives, such as semantic alignment [4, 11, 18, 41, 52], rare concept generation [48] and image translation [17]. For instance, ReNO [11] optimizes noises using various human preference rewards like HPS v2 [58], to enhance image quality and align with user intent. However, these approaches rely on iterative gradient updates during test-time, which are computationally expensive and prone to numerical instability (e.g., gradient explosion). In this work, we observe that initial noise exhibits cross-model consistency" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.089, + 0.905, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.268, + 0.905, + 0.297 + ], + "angle": 0, + "content": "Figure 3. Using the same initial noise, the generative posteriors (left) exhibit similarity with the text-conditioned images (right)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.907, + 0.378 + ], + "angle": 0, + "content": "(Fig. 1), making it an exploitable shared asset. Leveraging this property, we construct a reusable noise library that not only facilitates efficient noise selection but also helps existing noise optimization methods improve generation quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.388, + 0.649, + 0.405 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.413, + 0.717, + 0.428 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.907, + 0.661 + ], + "angle": 0, + "content": "Given a pre-trained generative diffusion model \\(\\mathcal{M}\\), which is capable of generating images \\(\\mathcal{I}\\) based on user-defined conditions \\(c\\), typically a text prompt. The generative process starts with the initial noise \\(\\epsilon\\) sampled from a Gaussian distribution \\(\\mathcal{N}(0, I)\\) and gradually transforms this noise into a clean image through \\(T\\) denoising steps1. Our motivation is to construct a large-scale noise sample library, denoted as \\(\\mathcal{N}_{set} = \\{\\epsilon_1, \\epsilon_2, \\dots, \\epsilon_n\\}\\), where each noise sample \\(\\epsilon_i\\) has its hidden features pre-computed from the generative posteriors. The goal is to identify the most suitable noise \\(\\epsilon^* \\in \\mathcal{N}_{set}\\) that aligns with a specific generation objective \\(\\mathcal{O}\\). Notably, this generation objective is inherently open-ended and encompasses various dimensions such as semantic consistency, color properties, and other user-defined attributes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.669, + 0.807, + 0.685 + ], + "angle": 0, + "content": "3.2. Implicit Guidance in Initial Noise" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.906, + 0.781 + ], + "angle": 0, + "content": "To investigate the influence of initial noise on generated images, we analyze the forward noise addition process in diffusion models [20, 37, 50, 51]. The forward diffusion process gradually perturbs an image \\( x_0 \\) over \\( T \\) timesteps, transforming it into a \"pure\" Gaussian noise \\( x_T \\). The noise addition at each step follows:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.79, + 0.905, + 0.808 + ], + "angle": 0, + "content": "\\[\nq \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.815, + 0.905, + 0.845 + ], + "angle": 0, + "content": "where \\(\\beta_{t}\\) is a predefined variances schedule. Defining \\(\\alpha_{t} = 1 - \\beta\\), the cumulative effect of noise over \\(t\\) steps results in:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.852, + 0.905, + 0.871 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {2}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.876, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1For simplicity, we omit the process of transitioning from pixel space to latent space in latent diffusion models (LDM)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17620" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.089, + 0.506, + 0.103 + ], + "angle": 0, + "content": "Offline Library Construction" + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.106, + 0.845, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.293, + 0.842, + 0.308 + ], + "angle": 0, + "content": "Figure 4. Illustration of our NoiseQuery pipeline which involves offline library construction and real-time inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.484, + 0.485 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i\\) and \\(\\epsilon \\sim \\mathcal{N}(0, I)\\). A fundamental assumption in diffusion theory is that \\(\\bar{\\alpha}_T \\to 0\\) as \\(T \\to \\infty\\), ensuring that the raw image is fully obscured. However, practical implementations can only apply a finite number of degradations (e.g., \\(T = 1000\\)) to approximate this ideal. As analyzed in prior work [32], common schedules (e.g., Linear [20], Cosine [37], Stable Diffusion [43]) fail to achieve zero signal-to-noise ratio (SNR) at \\(T\\), leaving residual signal from \\(x_0\\) in \\(x_T\\). For instance, Stable Diffusion's final timestep retains:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.497, + 0.482, + 0.512 + ], + "angle": 0, + "content": "\\[\nx _ {T} = 0. 0 6 8 2 6 5 \\cdot x _ {0} + 0. 9 9 7 6 6 7 \\cdot \\epsilon , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.69 + ], + "angle": 0, + "content": "where \\(\\sqrt{\\overline{\\alpha}_T} = 0.068265\\) quantifies the leakage. During training, diffusion models minimize reconstruction loss by utilizing these residuals as shortcuts rather than predicting the original image from scratch. This creates an unexpected link between noise and data, regardless of the model architecture. As a result, during inference, models continue to rely on the implicit knowledge gained during training to interpret pure Gaussian noise. We take advantage of this seemingly flawed model design and the unintended connection, using stochastic noise as a generic \"silent assistant\" that encodes clues about the image to be generated." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.473, + 0.713 + ], + "angle": 0, + "content": "3.3. Dissect Initial Noise via Generation Posterior" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.885 + ], + "angle": 0, + "content": "While recent works [4, 11, 17, 18, 41] optimize initial noise via gradient back-propagation for text-guided generation, they often overlook the inherent generative tendencies embedded in different noises. In other words, if the initial noise deviates significantly from the desired objective, iterative optimization struggles to bridge this gap, resulting in suboptimal convergence despite extensive tuning. In this subsection, we propose a more efficient alternative: directly analyzing the initial noise itself without the text prompt. Additionally, our method also yields better starting points for optimization, leading to faster convergence." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Generative Posterior. When the model generates images" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.339, + 0.877, + 0.353 + ], + "angle": 0, + "content": "Algorithm 1: NoiseQuery for Goal-Driven Generation" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.356, + 0.879, + 0.397 + ], + "angle": 0, + "content": "Input: Diffusion model \\(\\mathcal{M}\\); Noise library \\(\\mathcal{N}_{set} = \\{\\epsilon_1, \\epsilon_2, \\dots, \\epsilon_n\\}\\); Generation objective \\(\\mathcal{O}\\) and corresponding matching function \\(S(\\cdot)\\)." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.397, + 0.816, + 0.411 + ], + "angle": 0, + "content": "Output: Optimal noise sample \\(\\epsilon^{*}\\) for objective \\(\\mathcal{O}\\)." + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.414, + 0.706, + 0.427 + ], + "angle": 0, + "content": "Offline Library Construction" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.429, + 0.87, + 0.484 + ], + "angle": 0, + "content": "foreach \\(\\epsilon_{i}\\in \\mathcal{N}_{set}\\) do \\(\\mathcal{I}_i^{\\mathrm{uncond}} = \\mathcal{M}(\\epsilon_i,c = \\emptyset)\\) Extract features \\(\\mathcal{F}_i\\) (as detailed in Tab. 1) from \\(\\mathcal{I}_i^{\\mathrm{uncond}}\\) Store \\((\\epsilon_i,\\mathcal{F}_i)\\) in the noise library." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.485, + 0.547, + 0.496 + ], + "angle": 0, + "content": "end" + }, + { + "type": "title", + "bbox": [ + 0.521, + 0.511, + 0.648, + 0.522 + ], + "angle": 0, + "content": "Real-time Inference" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.524, + 0.814, + 0.553 + ], + "angle": 0, + "content": "Extract desired features \\(\\mathcal{F}_O\\) from objective \\(\\mathcal{O}\\) \\(\\epsilon^{*} = \\arg \\max_{\\epsilon_{i}\\in \\mathcal{N}_{set}}\\{S(\\mathcal{F}_{i},\\mathcal{F}_{O})\\mid \\epsilon_{i}\\in \\mathcal{N}_{set}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.559, + 0.58, + 0.571 + ], + "angle": 0, + "content": "return \\(\\epsilon^{*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.776 + ], + "angle": 0, + "content": "without any specific prompt, it depends solely on the noise characteristics, producing what we term \"generative posterior\". As shown in Fig. 3, generative posteriors exhibit clear generative tendencies (e.g., semantics, spatial layouts, and color palettes) similar to those of text-guided outputs. These similarities suggest that the initial noise carries significant latent information, influencing both unconditional generation and conditional generation in a similar manner, acting as an additional form of guidance alongside the text prompt. Thus, these purely noise-guided images can serve as an effective proxy for interpreting the initial noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Cross-model Consistency. We also observe another meaningful phenomenon: the impact of initial noise is universal and model-agnostic. As illustrated in Fig. 1, generative posteriors generated from the same noise remain highly similar across models with different architectures and parameters. For example, PixArt-\\(\\alpha\\) [6] is a Transformer-based T2I diffusion model, while Stable Diffusion [43] is UNet-based. More discussion of models and samplers can be found in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17621" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.498, + 0.184 + ], + "angle": 0, + "content": "
Generation GoalsFeature TypeMatch Function
SemanticsCLIP [42], BLIP [29]Cosine Similarity
StyleGram Matrix [16]MSE
ColorRGB, HSV, LABAbsolute Difference
TextureGLCM [19]Euclidean Distance
ShapeHu Moments [22]Euclidean Distance
SharpnessHigh Frequency Energy (HFE)Absolute Difference
" + }, + { + "type": "table_caption", + "bbox": [ + 0.146, + 0.189, + 0.427, + 0.204 + ], + "angle": 0, + "content": "Table 1. Feature types and matching functions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.483, + 0.32 + ], + "angle": 0, + "content": "Appendix A. This finding empowers us to create a large-scale, model-agnostic noise library, where each noise sample can be associated with diverse latent features and then shared across different models. The universality of noise-driven priors positions initial noise as a unified control resource for various applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.33, + 0.436, + 0.345 + ], + "angle": 0, + "content": "3.4. NoiseQuery for Goal-Driven Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.456 + ], + "angle": 0, + "content": "Building upon the insights regarding initial noise characteristics and cross-model consistency, we propose NoiseQuery, an efficient approach for selecting appropriate noise samples tailored to diverse generation objectives. As illustrated in Fig. 4, our method consists of two main phases: offline library construction and real-time inference. We also provide pseudo-code in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.653 + ], + "angle": 0, + "content": "Offline Library Construction. In the first stage, we build a noise library \\(\\mathcal{N}_{\\text {set }}\\), which contains many initial noise samples (e.g., 100k) sampled from standard Gaussian distribution. These samples are utilized to generate generative posteriors by running a diffusion model without any user-defined prompts. Subsequently, we extract a variety of features \\(\\mathcal{F}_i\\) from these generative posteriors that correspond to various generative objectives \\(\\mathcal{O}\\), such as semantic representations (e.g., CLIP features), color properties (e.g., saturation), etc. These extracted features are then stored as keys alongside their corresponding noise samples (values) within the noise library, creating a comprehensive key-value repository that can be queried during the inference phase." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.774 + ], + "angle": 0, + "content": "Real-time Inference. When performing real-time inference, we specify desired features \\(\\mathcal{F}_O\\) following the user's generation objective \\(\\mathcal{O}\\). These features could be semantic, color, spatial, or style-based attributes, depending on the specific goal. Using a match function \\(S(\\cdot)\\), we evaluate the alignment between each feature \\(\\mathcal{F}_i\\) in our pre-computed library and the objective features \\(\\mathcal{F}_O\\). The optimal noise sample \\(\\epsilon^*\\) is selected by maximizing the matching scores:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.781, + 0.483, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\epsilon^ {*} = \\arg \\max _ {\\epsilon_ {i} \\in \\mathcal {N} _ {s e t}} \\left\\{S \\left(\\mathcal {F} _ {i}, \\mathcal {F} _ {O}\\right) \\mid \\epsilon_ {i} \\in \\mathcal {N} _ {s e t} \\right\\}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Once the optimal noise sample \\(\\epsilon^{*}\\) is identified, it is fed into the generative model to produce an image that closely matches the user intent. In Tab. 1, we summarize several common generation goals achievable with NoiseQuery, including both high-level semantics and low-level visual properties. For each goal, we specify the corresponding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "feature types and matching functions. Detailed explanations of these features can be found in Appendix C. Notably, these goals can be applied individually or in combination. Users can specify multiple constraints simultaneously (e.g., generating a \"sunset\" image with a \"warm color palette\" while ensuring \"high sharpness\"). This flexibility makes NoiseQuery highly versatile and generalizable, enabling fine-grained control over the generated output." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.228, + 0.646, + 0.244 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.254, + 0.704, + 0.27 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.905, + 0.412 + ], + "angle": 0, + "content": "Datasets. We evaluate our approach on two popular benchmarks: MSCOCO [33] and DrawBench [47]. MSCOCO is a standard benchmark for text-to-image generation, providing textual descriptions of diverse natural scenes and object classes. We randomly select 10k captions as user prompts. DrawBench consists of 200 carefully designed prompts intended to challenge generative scenarios on various aspects, including long-form text, complex compositions, conflicting instructions, and rare vocabulary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.905, + 0.563 + ], + "angle": 0, + "content": "Model Zoo. We conduct experiments on six popular diffusion generation models, including Stable diffusion (SD) series [43] (1.4, 1.5, 2.0, 2.1), PixArt-\\(\\alpha\\) [6], and SD-Turbo [49]. Stable Diffusion is a Latent Diffusion Model with UNet [44] backbone. By default, we use 50 DDIM denoising steps with a CFG scale of 7.0. PixArt-\\(\\alpha\\) is a Transformer Latent Diffusion Model, and SD-Turbo is a distillation model based on Stable Diffusion 2.1, generating high-quality images even with only 1-2 sampling steps without CFG. For PixArt-\\(\\alpha\\), we use the default CFG scale of 4.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.906, + 0.701 + ], + "angle": 0, + "content": "Evaluation Metrics. For semantic similarity, we use ClipScore [42], which directly measures the alignment between the generated images and the corresponding textual prompts. Additionally, we incorporate several popular metrics trained on large-scale human preference data to quantify human preferences for AI-generated images, including HPS v2 [58], PickScore [27], and ImageReward [60]. These metrics offer a thorough evaluation of image quality, aesthetics, and alignment with user intent." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.843, + 0.728 + ], + "angle": 0, + "content": "4.2. Improvement on High-level Semantics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.905, + 0.84 + ], + "angle": 0, + "content": "For semantic consistency goals, we utilize features like CLIP [42] and BLIP [29] as the query to identify the semantically matched noise for each text prompt. Such aligned text and noise yield two benefits: (i) prevents the model from generating irrelevant outputs or falling into unsolvable regions. (ii) minimizes generation difficulty and maximizes the likelihood for semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Quantitative Analysis. In Tab. 2, we evaluate our method on standard text-to-image (T2I) benchmarks, demonstrating consistent performance gains across all base models by offering better noise-text alignment. NoiseQuery" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17622" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.31 + ], + "angle": 0, + "content": "
Base ModelMethodDrawBench [47]MSCOCO [33]Time Cost
ImageRewardPickScoreHPS v2CLIPScoreImageRewardPickScoreHPS v2CLIPScore
SD 1.5Base Model0.0421.1124.5730.900.1521.4125.6531.081.334 s
+ NoiseQuery0.0821.1625.0231.410.2721.4826.0731.471.336 s
+ Diffusion-DPO [53]0.0921.2925.0231.190.2521.6426.3131.261.350 s
+ Diffusion-DPO [53] + NoiseQuery0.1721.3325.2531.410.3521.6826.6031.551.352 s
SD 2.1Base Model0.1221.3324.9331.130.3621.7226.5831.401.301 s
+ NoiseQuery0.2621.4625.3931.680.4421.7626.8231.501.303 s
+ CFG++ [10]0.1221.3324.8331.130.3721.7226.6631.313.724 s
+ CFG++ [10] + NoiseQuery0.2721.4325.5531.610.4721.7626.9731.673.726 s
SD-TurboBase Model0.2621.7825.2331.290.4722.0726.2231.510.072 s
+ NoiseQuery0.4121.8725.6631.580.5022.1726.8231.760.074 s
+ ReNO [11]1.6723.4032.4832.55----23.56 s
+ ReNO [11] + NoiseQuery1.7123.5232.9232.78----23.56 s
PixArt-αBase Model0.7022.0828.2730.830.7822.2429.3331.484.327 s
+ NoiseQuery0.8222.1128.4531.270.7922.3329.5631.644.328 s
+ LaVi-Bridge [67]0.6322.0828.3530.920.7522.3129.4931.865.092 s
+ LaVi-Bridge [67] + NoiseQuery0.7222.2428.6131.350.7822.3529.6732.015.094 s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.32, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Table 2. Evaluation of objective metrics on different datasets and models. Higher is better for all metrics. Our approach enhances the base model's performance and complements a wide range of T2I enhancement methods, including DPO [53] (reward-based fine-tuning), CFG++ [10] (guidance refinement), ReNO [11] (gradient-based initial noise optimization), and LaVi-Bridge [67] (text encoding enhancement via LORA/adapters). \\(\\cdot\\): ReNO results on MSCOCO are excluded due to excessive inference time." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.388, + 0.482, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.707, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Figure 5. Generated images of the base model, base model with T2I enhancement methods, and base model with enhancement and NoiseQuery. The four rows, from top to bottom, respectively correspond to the method groups listed in Table 2: Diffusion-DPO [53], CFG++ [10], ReNO [11], and LaVi-Bridge [67]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "also serves as a universal foundation layer, complementing T2I enhancement techniques like Diffusion-DPO [53], CFG++ [10], ReNO [11], and LaVi-Bridge [67]. Notably, these methods face inherent limitations: model-specific dependencies (e.g., Diffusion-DPO and LaVi-Bridge require retraining for new models; ReNO is restricted to one-step models due to heavy computation, excluding com" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.389, + 0.885, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.522, + 0.838, + 0.535 + ], + "angle": 0, + "content": "Figure 6. Queried noise v.s. random noise." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.544, + 0.885, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.675, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Figure 7. CLIP and FID evaluations across CFG scales on MSCOCO [33], with the CFG scale increasing from left to right." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.907, + 0.868 + ], + "angle": 0, + "content": "mon multi-step models like SD) and high latency. In contrast, our method seamlessly integrates with them, enhancing their performance while adding near-zero time costs. We also present qualitative results in Fig. 5 to illustrate its effectiveness in assisting text-to-image synthesis. Additionally, we show the zero-shot transferability of NoiseQuery across various generative models in Appendix B.1. This model-agnostic compatibility, combined with minimal computational cost, establishes NoiseQuery as a versatile and efficient enhancement tool for diverse pipelines." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Diversity Analysis. We report the performance of selecting the top 20 matched noise samples per prompt instead of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17623" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.488, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.365, + 0.368, + 0.377 + ], + "angle": 0, + "content": "(a) Color properties guidance." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.089, + 0.903, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.365, + 0.791, + 0.377 + ], + "angle": 0, + "content": "(b) Structural features guidance." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.378, + 0.907, + 0.407 + ], + "angle": 0, + "content": "Figure 8. Multiple low-level visual properties control through NoiseQuery. All images use minimal text prompts containing only semantic concepts (e.g., object labels), lacking explicit visual specifications." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.43, + 0.484, + 0.484 + ], + "angle": 0, + "content": "
Library Size0.5k1k2k5k10k50k100k
Matching Function Cost (×10-4s)1.391.391.391.391.401.401.51
Argmax Selection Cost (×10-4s)0.250.250.250.250.376.1613.25
CLIP Score31.5131.5331.5731.5931.6631.7331.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.488, + 0.473, + 0.501 + ], + "angle": 0, + "content": "Table 3. Retrieval time breakdown across various library sizes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.534, + 0.483, + 0.685 + ], + "angle": 0, + "content": "the single best match. As shown in Fig. 6, these samples consistently outperform the standard random noise baseline on DrawBench [47], demonstrating that NoiseQuery provides not only optimal but also diverse high-quality noise candidates. To quantify diversity, we use the Diversity Metric (DIV) [57], computed as the standard deviation of DINOV2 [38] embeddings across 20 generated images per prompt (see Fig. 6 legend). Our method achieves comparable diversity to random noise, confirming that NoiseQuery enhances image quality without sacrificing diversity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.687, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Different Classifier-free Guidance Scale Analysis. We assess our method's performance across various classifier-free guidance (CFG) scales (1.5, 2.0, 3.0, ..., 15.0) on 10k MSCOCO [33] samples. Our NoiseQuery achieves high-quality outputs even with low guidance scales, as demonstrated in Fig. 7. This supports our hypothesis that by selecting more compatible initial noise, we can effectively reduce the generation difficulty. Therefore, we can obtain high-quality output at low scales and avoid the over-saturation and instability that usually come with high scales." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Time Cost. We also provide a cost breakdown for the retrieval process in Tab. 3, which includes the computation of the matching function (utilizing query CLIP features at no extra cost since CLIP is embedded in SD) and argmax se" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.419, + 0.907, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.51, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Figure 9. Text prompts often fail to align generated images with user low-level attribute preferences, while initial noise plays a dominant role in controlling them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.907, + 0.627 + ], + "angle": 0, + "content": "lection. The results demonstrate that both steps are efficient, even with large libraries (e.g., 0.0015 seconds for 100,000 entries), indicating that the additional cost is negligible." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.905, + 0.654 + ], + "angle": 0, + "content": "4.3. Controllability on Low-Level Visual Properties" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Text prompts effectively steer high-level semantics in text-to-image (T2I) generation but often fail to provide sufficient control over low-level visual properties that affect visual perception. This limitation stems from two factors: (1) text is inherently inadequate for conveying detailed visual reference input, and (2) text encoders like CLIP [42], trained primarily on high-level semantics, struggle to capture and control fine-grained low-level properties such as color, texture, shape, and sharpness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In contrast, initial noise plays a dominant role in determining these properties, making it a valuable resource for precise control, as illustrated in Fig. 9. Our method leverages this by enabling fine-grained control of low-level attributes through initial noise, without requiring additional training. The following experiments demonstrate the effectiveness of our approach in achieving these targets, with fur" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17624" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.089, + 0.177, + 0.108 + ], + "angle": 0, + "content": "Diverse User Demands" + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.089, + 0.264, + 0.108 + ], + "angle": 0, + "content": "Stable Diffusion" + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.089, + 0.385, + 0.109 + ], + "angle": 0, + "content": "Stable Diffusion (enhanced prompt)" + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.089, + 0.485, + 0.109 + ], + "angle": 0, + "content": "NoiseQuery (ours)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.109, + 0.482, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.268, + 0.483, + 0.311 + ], + "angle": 0, + "content": "Figure 10. Examples of combining high-level semantics (red) and low-level properties (blue). Enhanced prompts are created by fusing semantic objects and visual details via a large language model." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.337, + 0.483, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.502, + 0.483, + 0.545 + ], + "angle": 0, + "content": "Figure 11. Compatibility with controllable models such as TI [15] and ControlNet [64]. Our method provides a better initial noise, enhancing semantic consistency and low-level attribute control." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.585, + 0.441, + 0.601 + ], + "angle": 0, + "content": "ther implementation details provided in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.744 + ], + "angle": 0, + "content": "Color Properties. Color attributes such as brightness, saturation, contrast, and hue are crucial for an image's visual appeal. We assess the overall color hue by averaging each RGB channel, measure saturation and brightness in the HSV color space, and compute contrast as the standard deviation of grayscale intensities. As shown in Fig. 8a, selecting initial noise that matches target color distributions enables precise control over the generated images' color tendencies, aligning them with user preferences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Structural Features. For spatial control, we implement: (1) Texture through Gray-Level Co-occurrence Matrix (GLCM) [19] features that quantify intensity co-occurrence patterns; (2) Shape consistency via Hu Moments [22] as rotation-invariant descriptors; and (3) Sharpness by prioritizing high-frequency energy (HFE) in Fourier spectra. Fig. 8b demonstrates how noise selection governs these structural features—preserving geometric fidelity in object shapes, modulating surface textures, and enhancing edge crispness—all achieved without prompt engineering." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.843, + 0.107 + ], + "angle": 0, + "content": "4.4. Enhancement on Combined Scenarios" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.116, + 0.906, + 0.207 + ], + "angle": 0, + "content": "Building upon the results in previous sections, which demonstrate our method's effectiveness in high-level semantic alignment and low-level attribute control, this section highlights its ability to combine these diverse goals and seamlessly integrate with powerful existing controllable generation models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.21, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Combined Goals. Beyond individual tasks, our method excels in simultaneously achieving multiple generation goals. By sequentially reranking noise candidates—first selecting top candidates based on one goal (e.g., high-level semantics) and then refining within this subset for another goal (e.g., low-level attributes)—we progressively narrow down the optimal noise pool. As shown in Fig. 10, this approach allows the combination of high-level semantic alignment with various low-level attribute controls, enabling the generation of images that meet diverse user preferences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.909, + 0.514 + ], + "angle": 0, + "content": "Compatibility with Controlling Methods. Our method is designed to be instant-time and model-agnostic, making it highly compatible with existing controllable generation models such as Textual Inversion (TI) [15] and ControlNet [64]. By providing a better initial noise, our approach establishes a foundational layer for controllable generation, which synergizes with subsequent refinement processes. As shown in Fig. 11, this integration enhances semantic consistency while enabling precise control over challenging low-level attributes, such as color and texture." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.538, + 0.634, + 0.553 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.907, + 0.808 + ], + "angle": 0, + "content": "In this work, we propose NoiseQuery, a tuning-free method that enables versatile goal-driven T2I generation through enhanced noise initialization, supported by a generic and reusable pre-built noise library. By selecting Gaussian initialization in the noise library better aligned with the text prompt or specified attributes, NoiseQuery achieves fine-grained control over both high-level semantic and intricate low-level details. In addition, the model-agnostic nature of our method allows it to be seamlessly integrated with existing T2I augmentation and controllable generation methods without model adjustments or significant computational overhead. Experiments validate its effectiveness in improving output fidelity and precision, particularly for visual attributes that are challenging to articulate textually. The lightweight and scalability of the framework make it adaptable to real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Limitations. Although NoiseQuery provides flexibility for aligning high-level semantics and controlling low-level attributes, it remains bounded by the base model's ability, and may still underperform in extremely complex cases. Additionally, it has limitations in achieving fine-grained control, such as Canny edge control." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17625" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.213 + ], + "angle": 0, + "content": "This work was partially supported by the National Natural Science Foundation of China under grant 62372341. This work was also partially supported by an Amazon Research Award 2023 to OR and YZ. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not reflect the views of Amazon." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.244, + 0.188, + 0.259 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.268, + 0.482, + 0.325 + ], + "angle": 0, + "content": "[1] Donghoon Ahn, Jiwon Kang, Sanghyun Lee, Jaewon Min, Minjae Kim, Wooseok Jang, Hyoungwon Cho, Sayak Paul, SeonHwa Kim, Eunju Cha, et al. A noise is worth diffusion guidance. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.327, + 0.484, + 0.368 + ], + "angle": 0, + "content": "[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In ICLR, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.371, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG), 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.429, + 0.484, + 0.483 + ], + "angle": 0, + "content": "[4] Changgu Chen, Libing Yang, Xiaoyan Yang, Lianggangxu Chen, Gaoqi He, Changbo Wang, and Yang Li. Find: Finetuning initial noise distribution with policy optimization for diffusion models. In ACM MM, 2024. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.484, + 0.554 + ], + "angle": 0, + "content": "[5] Haoyang Chen, Dongfang Sun, Caoyuan Ma, Shiqin Wang, Kewei Zhang, Zheng Wang, and Zhixiang Wang. Subjective camera: Bridging human cognition and visual reconstruction through sequence-aware sketch-guided diffusion. arXiv, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.557, + 0.484, + 0.626 + ], + "angle": 0, + "content": "[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2, 4, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.629, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[7] Sitan Chen, Giannis Daras, and Alex Dimakis. Restoration-degradation beyond linear diffusions: A non-asymptotic analysis for ddim-type samplers. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[8] Zijiao Chen, Jiaxin Qing, Tiange Xiang, Wan Lin Yue, and Juan Helen Zhou. Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.484, + 0.785 + ], + "angle": 0, + "content": "[9] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.484, + 0.842 + ], + "angle": 0, + "content": "[10] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. NeurIPS, 37:125487-125519, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.268, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[12] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. NeurIPS, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[13] Yingying Fan, Quanwei Yang, Kaisiyuan Wang, Hang Zhou, Yingying Li, Haocheng Feng, Errui Ding, Yu Wu, and Jingdong Wang. Re-hold: Video hand object interaction reenactment via adaptive layout-instructed diffusion model. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[14] Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.334 + ], + "angle": 0, + "content": "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.336, + 0.904, + 0.363 + ], + "angle": 0, + "content": "[16] Leon A Gatys. A neural algorithm of artistic style. arXiv, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.366, + 0.905, + 0.407 + ], + "angle": 0, + "content": "[17] Or Greenberg, Eran Kishon, and Dani Lischinski. S2st: Image-to-image translation in the seed space of latent diffusion. arXiv, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.41, + 0.905, + 0.464 + ], + "angle": 0, + "content": "[18] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.467, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[19] Robert M Haralick, Karthikeyan Shanmugam, and Its' Hak Dinstein. Textural features for image classification. IEEE SMCS, 1973. 5, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.905, + 0.582 + ], + "angle": 0, + "content": "[21] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.584, + 0.905, + 0.625 + ], + "angle": 0, + "content": "[22] Ming-Kuei Hu. Visual pattern recognition by moment invariants. IRE transactions on information theory, 1962. 5, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.628, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[23] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[24] Huayang Huang, Yu Wu, and Qian Wang. Robin: Robust and invisible watermarks for diffusion models with adversarial optimization. NIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[25] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan, and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. In SIGGRAPH Asia, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[26] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[27] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. NeurIPS, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Gen Li, Yuting Wei, Yuxin Chen, and Yuejie Chi. Towards non-asymptotic convergence for diffusion-based generative models. In ICLR, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17626" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML. PMLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[30] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[31] Yumeng Li, Margret Keuper, Dan Zhang, and Anna Khoreva. Divide & bind your attention for improved generative semantic nursing. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[32] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In WACV, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[34] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In ICLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[35] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[36] Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, et al. Inference-time scaling for diffusion models beyond scaling denoising steps. arXiv, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[38] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[40] Yuhan Pei, Ruoyu Wang, Yongqi Yang, Ye Zhu, Olga Russakovsky, and Yu Wu. Sowing information: Cultivating contextual coherence with mllms in image generation. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[41] Zipeng Qi, Lichen Bai, Haoyi Xiong, et al. Not all noises are created equally: Diffusion noise selection and optimization. arXiv, 2024. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML. PMLR, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI. Springer, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.219, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[47] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 5, 6, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.291, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[48] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In AAAI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.333, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[49] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 2, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.376, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.405, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[51] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.462, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[52] Zhiwei Tang, Jiangweizhi Peng, Jiasheng Tang, Mingyi Hong, Fan Wang, and Tsung-Hui Chang. Inference-time alignment of diffusion models with direct noise optimization. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[53] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, pages 8228-8238, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.589, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[54] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.633, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[55] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.689, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[56] Ruoyu Wang, Yongqi Yang, Zhihao Qian, Ye Zhu, and Yu Wu. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[57] Xi Wang, Nicolas Dufour, Nefeli Andreou, Marie-Paule Cani, Victoria Fernández Abrevaya, David Picard, and Vicky Kalogeiton. Analysis of classifier-free guidance weight schedulers. arXiv, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[58] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[59] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. *IJCV*, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17627" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[60] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. NeurIPS, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[61] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models. arXiv, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[62] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[63] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Ic-light github page, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[66] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[67] Shihao Zhao, Shaozhe Hao, Bojia Zi, Huaizhe Xu, and Kwan-Yee K Wong. Bridging different language models and generative vision models for text-to-image generation. In ECCV, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[68] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[69] Chenyi Zhuang, Ying Hu, and Pan Gao. Magnet: We never know how text-to-image diffusion models work, until we learn how vision-language models function. NeurIPS, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17628" + } + ] +] \ No newline at end of file diff --git a/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/76b9c942-da58-47e6-8ca6-ac0446eb367e_origin.pdf b/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/76b9c942-da58-47e6-8ca6-ac0446eb367e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f21023346896bdc1ee533fde8e23de4bd086eaa1 --- /dev/null +++ b/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/76b9c942-da58-47e6-8ca6-ac0446eb367e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:042a1abebc65c1ed1863aa63292f3fc6eafb3403fc9247e352982dc240f8152f +size 7604041 diff --git a/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/full.md b/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e3b603d2ae1a0cac0fcfd4ecb6c9353f9ec4e50d --- /dev/null +++ b/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/full.md @@ -0,0 +1,319 @@ +# The Silent Assistant: NoiseQuery as Implicit Guidance for Goal-Driven Image Generation + +Ruoyu Wang + +Huayang Huang + +Ye Zhu + +Olga Russakovsky + +Yu $\mathbf{W}\mathbf{u}^{1*}$ + +$^{1}$ School of Computer Science, Wuhan University $^{2}$ Department of Computer Science, Princeton University + +{wangruoyu, hyhuang, wuyucs}@whu.edu.cn + +{yezhu, olgarus}@princeton.edu + +# Abstract + +In this work, we introduce NoiseQuery as a novel method for enhanced noise initialization in versatile goal-driven text-to-image (T2I) generation. Specifically, we propose to leverage an aligned Gaussian noise as implicit guidance to complement explicit user-defined inputs, such as text prompts, for better generation quality and controllability. Unlike existing noise optimization methods designed for specific models, our approach is grounded in a fundamental examination of the generic finite-step noise scheduler design in diffusion formulation, allowing better generalization across different diffusion-based architectures in a tuning-free manner. This model-agnostic nature allows us to construct a reusable noise library compatible with multiple T2I models and enhancement techniques, serving as a foundational layer for more effective generation. Extensive experiments demonstrate that NoiseQuery enables fine-grained control and yields significant performance boosts not only over high-level semantics but also over low-level visual attributes, which are typically difficult to specify through text alone, with seamless integration into current workflows with minimal computational overhead. Code is available at https://github.com/wangruoyu02/NoiseQuery. + +# 1. Introduction + +Text-to-image synthesis (T2I) has gained significant research attention with the rising popularity of diffusion generative models [20, 39, 50, 51]. While these models demonstrate impressive generative capabilities, they still struggle to faithfully align generated content with user prompts, leading to suboptimal outputs. To enrich the text-based conditioning capabilities of diffusion models, researchers have + +![](images/6bf33591d212fda3a8f71277ee39ccc73e80d4093a26b2fa0f0da3a67e83c151.jpg) +Figure 1. Each row shows a set of images generated by different diffusion models starting from the same initial noise, using the same sampler, and with NULL as the text prompt. The generated images demonstrate striking both high-level semantic and low-level visual similarities across models. + +explored various techniques, including fine-tuning diffusion UNet [2, 12, 53], enhancing language encoders [23, 67], refining inference strategies [3, 10], and extra additional image conditions [15, 45, 66]. A recent popular line of research focuses on optimizing the initial noise [4, 11, 18, 41], motivated by the observation that images generated from the same noise often exhibit high similarities within a single model, even when prompted differently. In this work, we extend this understanding by revealing that such noise-induced consistency persists across T2I models with different diffusion backbones, as illustrated in Fig. 1. + +To better understand this phenomenon, we examine the formulation of diffusion models and reveal that cross-model similarities stem from the finite-step noise scheduler design + +and persist across different diffusion variants. Specifically, one of the fundamental assumptions in diffusion models lies in a gradual stepwise perturbation process that transforms a real-world distribution into a standard Gaussian. This transition requires, in theory, an infinite number of diffusion steps to perfectly obscure the raw images, contradicting the practical implementation where we can only apply a finite number (i.e., non-asymptotic) of degradations through predefined noise schedulers [7, 28]. As a result, the real-world diffusion models inadvertently learn to leverage residual information embedded in image data, creating an unintended link between noise and data regardless of the model architecture during inference. However, rather than being a limitation, we leverage this seemingly imperfect model design and unintentional connection as an unexpected advantage by laying the foundation for using stochastic noise as a generic "silent assistant", which implicitly encodes clues about the image that will be generated. + +We note that it is critical to align the implicit noise clues with the user-provided input, such as text prompts, in order to achieve optimal generation performance. As shown in Fig. 2, when the inherent tendencies of the noise guidance and text prompt diverge significantly, the generation process leads to instability or semantically inconsistent results. Furthermore, unlike text prompts that convey high-level semantics, initial noise inherently encodes low-level visual primitives—such as color gradients, texture patterns, and spatial frequencies—directly tied to the pixel space. This makes it particularly suitable for controlling fine-grained visual properties that are difficult to specify through text alone. By leveraging initial noise as a complementary control signal, we produce more nuanced and precise outputs, enhancing the overall generative capability of T2I models. + +Building on this insight, we propose NoiseQuery, a novel method that exploits initial noise as a shared generative asset across diverse T2I models. The key challenge lies in establishing an interpretable mapping between stochastic noise patterns and their concrete generative clues. To bridge this gap, we utilize images generated with empty NULL text prompts as proxies to reveal the hidden information encoded in the initial noise (details in Sec. 3.3), denoted as the implicit generative posterior. When deprived of text guidance, the generation process becomes purely noise-driven, forcing the model to excavate and decode the residual signals preserved in the initial noise. This process enables the construction of a comprehensive Noise Library, built offline by generating a diverse set of such noise-driven images from a Gaussian distribution and extracting their multi-grained features, spanning high-level semantics and low-level visual attributes. During inference, given a specific user request (single or combined goals), we extract its features, perform a quick lookup in the pre-built library, and select the best-matching noise for initialization. Notably, + +![](images/a640fc88760112795e8545a15c6c692a469507d4d32647ae37f9e926c2d0e5a3.jpg) +Figure 2. Compared to mismatched noise, noise aligned with text prompts can produce more accurate images in complex T2I generation. A NULL prompt represents unconditional generation. + +NoiseQuery is model-agnostic and serves as a reusable resource across multiple diffusion architectures. This ability makes it naturally compatible with diverse enhancement strategies including recent noise optimization approaches, providing a foundational layer for more effective generation. Moreover, NoiseQuery operates at a negligible online computational cost (0.002 seconds per prompt), ensuring seamless integration into existing workflows while enhancing the controllability and quality of T2I generation. + +To validate the effectiveness of NoiseQuery, we conduct experiments on low-level property preferences, high-level semantic consistency, and their combinations. Extensive evaluations with six diffusion-based T2I models [6, 43, 49], four types of T2I enhancement methods [10, 11, 53, 67] and two controllable T2I models [15, 66] show that NoiseQuery not only improve the base model's performance but also complements multiple T2I enhancement methods. Our approach is also among the first to address the challenge of subtle low-level preference control with diffusion models, highlighting a new avenue for fine-grained generative tasks. Furthermore, by progressively matching the noise, our method is capable of supporting combinations of generation goals, offering more nuanced control for users. + +In summary, our contributions are as follows: + +- We reveal the generic generative tendencies within the initial noise and leverage this implicit guidance as a universal generative asset that can be seamlessly integrated with existing T2I models and enhancement methods. +- We introduce NoiseQuery, a novel method that retrieves an optimal initial noise from a pre-built large-scale noise library, fulfilling versatile user-specified requirements encompassing both semantic and low-level visual attributes—a relatively under-explored area. +- We conduct extensive experiments to demonstrate that + +our method can improve the general quality of diffusion-based generators and can significantly improve multiple T2I enhancement methods while maintaining a minimal cost and without requiring additional learning. + +# 2. Related Work + +Improving Base Text-to-Image Generation. Recent breakthroughs in text-to-image (T2I) generation have been driven by diffusion-based models, yet these models still suffer from artifacts and misalignment with textual prompts. Current solutions primarily fall into two categories: tuning-based and inference-time methods. Tuning-based methods [2, 12, 30, 53, 62] optimize model parameters through techniques like direct preference optimization (DPO) [30, 53, 62]. Additionally, LaVi-Bridge [67] enhances textual understanding of T2I models by integrating advanced language models with parameter-efficient adapters. In contrast, inference-time methods [3, 10, 21, 31, 69] dynamically adjust the generation process. For example, CFG++ [10] improves image quality by addressing non-manifold issues in classifier-free guidance. Our approach complements both strategies by offering improved initialization for generation. Enhancing T2I with Additional Conditions. To address the limitations of text-only conditions in meeting diverse user needs, numerous works [5, 8, 9, 13, 15, 40, 45, 59, 64, 66] have expanded T2I generation through multi-modal conditions. For instance, methods like ControlNet [64] introduce structural signals (e.g., edge and depth maps) to control spatial layouts, while IC-Light [65] adjusts the illumination effects. Meanwhile, customization generation approaches focus on learning user-provided concepts like identity [15, 45, 46, 55, 63], style [14, 54], and interaction [25]. In this paper, we highlight that the initial noise for diffusion sampling can also serve as a conditional resource. Furthermore, compared to efforts that focus primarily on high-level semantic control, we show that initial noise holds untapped potential for guiding various low-level visual properties (e.g., color, texture, and sharpness). + +Initial Noise Optimization for Diffusion Sampling. A growing body of research has focused on optimizing the initial noise in diffusion models, driven by the consensus that it significantly impacts the generated output [1, 24, 36, 56, 61]. These methods iteratively refine initial noise through backpropagation at inference, guided by different generative objectives, such as semantic alignment [4, 11, 18, 41, 52], rare concept generation [48] and image translation [17]. For instance, ReNO [11] optimizes noises using various human preference rewards like HPS v2 [58], to enhance image quality and align with user intent. However, these approaches rely on iterative gradient updates during test-time, which are computationally expensive and prone to numerical instability (e.g., gradient explosion). In this work, we observe that initial noise exhibits cross-model consistency + +![](images/232390aa8abb55d73837b546e102a652f165cea1c9b189971bdfbcb28318e54d.jpg) +Figure 3. Using the same initial noise, the generative posteriors (left) exhibit similarity with the text-conditioned images (right). + +(Fig. 1), making it an exploitable shared asset. Leveraging this property, we construct a reusable noise library that not only facilitates efficient noise selection but also helps existing noise optimization methods improve generation quality. + +# 3. Methodology + +# 3.1. Problem Formulation + +Given a pre-trained generative diffusion model $\mathcal{M}$ , which is capable of generating images $\mathcal{I}$ based on user-defined conditions $c$ , typically a text prompt. The generative process starts with the initial noise $\epsilon$ sampled from a Gaussian distribution $\mathcal{N}(0, I)$ and gradually transforms this noise into a clean image through $T$ denoising steps1. Our motivation is to construct a large-scale noise sample library, denoted as $\mathcal{N}_{set} = \{\epsilon_1, \epsilon_2, \dots, \epsilon_n\}$ , where each noise sample $\epsilon_i$ has its hidden features pre-computed from the generative posteriors. The goal is to identify the most suitable noise $\epsilon^* \in \mathcal{N}_{set}$ that aligns with a specific generation objective $\mathcal{O}$ . Notably, this generation objective is inherently open-ended and encompasses various dimensions such as semantic consistency, color properties, and other user-defined attributes. + +# 3.2. Implicit Guidance in Initial Noise + +To investigate the influence of initial noise on generated images, we analyze the forward noise addition process in diffusion models [20, 37, 50, 51]. The forward diffusion process gradually perturbs an image $x_0$ over $T$ timesteps, transforming it into a "pure" Gaussian noise $x_T$ . The noise addition at each step follows: + +$$ +q \left(x _ {t} \mid x _ {t - 1}\right) = \mathcal {N} \left(\sqrt {1 - \beta_ {t}} x _ {t - 1}, \beta_ {t} \mathbf {I}\right), \tag {1} +$$ + +where $\beta_{t}$ is a predefined variances schedule. Defining $\alpha_{t} = 1 - \beta$ , the cumulative effect of noise over $t$ steps results in: + +$$ +x _ {t} = \sqrt {\bar {\alpha} _ {t}} x _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \tag {2} +$$ + +![](images/9403d1b910c1e857a0739e9aa87fbc8bc7ebecfbaf92e68ef57c475d2403bd3b.jpg) +Offline Library Construction +Figure 4. Illustration of our NoiseQuery pipeline which involves offline library construction and real-time inference. + +where $\bar{\alpha}_t = \prod_{i=1}^t \alpha_i$ and $\epsilon \sim \mathcal{N}(0, I)$ . A fundamental assumption in diffusion theory is that $\bar{\alpha}_T \to 0$ as $T \to \infty$ , ensuring that the raw image is fully obscured. However, practical implementations can only apply a finite number of degradations (e.g., $T = 1000$ ) to approximate this ideal. As analyzed in prior work [32], common schedules (e.g., Linear [20], Cosine [37], Stable Diffusion [43]) fail to achieve zero signal-to-noise ratio (SNR) at $T$ , leaving residual signal from $x_0$ in $x_T$ . For instance, Stable Diffusion's final timestep retains: + +$$ +x _ {T} = 0. 0 6 8 2 6 5 \cdot x _ {0} + 0. 9 9 7 6 6 7 \cdot \epsilon , \tag {3} +$$ + +where $\sqrt{\overline{\alpha}_T} = 0.068265$ quantifies the leakage. During training, diffusion models minimize reconstruction loss by utilizing these residuals as shortcuts rather than predicting the original image from scratch. This creates an unexpected link between noise and data, regardless of the model architecture. As a result, during inference, models continue to rely on the implicit knowledge gained during training to interpret pure Gaussian noise. We take advantage of this seemingly flawed model design and the unintended connection, using stochastic noise as a generic "silent assistant" that encodes clues about the image to be generated. + +# 3.3. Dissect Initial Noise via Generation Posterior + +While recent works [4, 11, 17, 18, 41] optimize initial noise via gradient back-propagation for text-guided generation, they often overlook the inherent generative tendencies embedded in different noises. In other words, if the initial noise deviates significantly from the desired objective, iterative optimization struggles to bridge this gap, resulting in suboptimal convergence despite extensive tuning. In this subsection, we propose a more efficient alternative: directly analyzing the initial noise itself without the text prompt. Additionally, our method also yields better starting points for optimization, leading to faster convergence. + +Generative Posterior. When the model generates images + +# Algorithm 1: NoiseQuery for Goal-Driven Generation + +Input: Diffusion model $\mathcal{M}$ ; Noise library $\mathcal{N}_{set} = \{\epsilon_1, \epsilon_2, \dots, \epsilon_n\}$ ; Generation objective $\mathcal{O}$ and corresponding matching function $S(\cdot)$ . + +Output: Optimal noise sample $\epsilon^{*}$ for objective $\mathcal{O}$ . + +# Offline Library Construction + +foreach $\epsilon_{i}\in \mathcal{N}_{set}$ do $\mathcal{I}_i^{\mathrm{uncond}} = \mathcal{M}(\epsilon_i,c = \emptyset)$ Extract features $\mathcal{F}_i$ (as detailed in Tab. 1) from $\mathcal{I}_i^{\mathrm{uncond}}$ Store $(\epsilon_i,\mathcal{F}_i)$ in the noise library. + +end + +# Real-time Inference + +Extract desired features $\mathcal{F}_O$ from objective $\mathcal{O}$ $\epsilon^{*} = \arg \max_{\epsilon_{i}\in \mathcal{N}_{set}}\{S(\mathcal{F}_{i},\mathcal{F}_{O})\mid \epsilon_{i}\in \mathcal{N}_{set}\}$ + +return $\epsilon^{*}$ + +without any specific prompt, it depends solely on the noise characteristics, producing what we term "generative posterior". As shown in Fig. 3, generative posteriors exhibit clear generative tendencies (e.g., semantics, spatial layouts, and color palettes) similar to those of text-guided outputs. These similarities suggest that the initial noise carries significant latent information, influencing both unconditional generation and conditional generation in a similar manner, acting as an additional form of guidance alongside the text prompt. Thus, these purely noise-guided images can serve as an effective proxy for interpreting the initial noise. + +Cross-model Consistency. We also observe another meaningful phenomenon: the impact of initial noise is universal and model-agnostic. As illustrated in Fig. 1, generative posteriors generated from the same noise remain highly similar across models with different architectures and parameters. For example, PixArt- $\alpha$ [6] is a Transformer-based T2I diffusion model, while Stable Diffusion [43] is UNet-based. More discussion of models and samplers can be found in + +
Generation GoalsFeature TypeMatch Function
SemanticsCLIP [42], BLIP [29]Cosine Similarity
StyleGram Matrix [16]MSE
ColorRGB, HSV, LABAbsolute Difference
TextureGLCM [19]Euclidean Distance
ShapeHu Moments [22]Euclidean Distance
SharpnessHigh Frequency Energy (HFE)Absolute Difference
+ +Table 1. Feature types and matching functions. + +Appendix A. This finding empowers us to create a large-scale, model-agnostic noise library, where each noise sample can be associated with diverse latent features and then shared across different models. The universality of noise-driven priors positions initial noise as a unified control resource for various applications. + +# 3.4. NoiseQuery for Goal-Driven Generation + +Building upon the insights regarding initial noise characteristics and cross-model consistency, we propose NoiseQuery, an efficient approach for selecting appropriate noise samples tailored to diverse generation objectives. As illustrated in Fig. 4, our method consists of two main phases: offline library construction and real-time inference. We also provide pseudo-code in Algorithm 1. + +Offline Library Construction. In the first stage, we build a noise library $\mathcal{N}_{\text {set }}$ , which contains many initial noise samples (e.g., 100k) sampled from standard Gaussian distribution. These samples are utilized to generate generative posteriors by running a diffusion model without any user-defined prompts. Subsequently, we extract a variety of features $\mathcal{F}_i$ from these generative posteriors that correspond to various generative objectives $\mathcal{O}$ , such as semantic representations (e.g., CLIP features), color properties (e.g., saturation), etc. These extracted features are then stored as keys alongside their corresponding noise samples (values) within the noise library, creating a comprehensive key-value repository that can be queried during the inference phase. + +Real-time Inference. When performing real-time inference, we specify desired features $\mathcal{F}_O$ following the user's generation objective $\mathcal{O}$ . These features could be semantic, color, spatial, or style-based attributes, depending on the specific goal. Using a match function $S(\cdot)$ , we evaluate the alignment between each feature $\mathcal{F}_i$ in our pre-computed library and the objective features $\mathcal{F}_O$ . The optimal noise sample $\epsilon^*$ is selected by maximizing the matching scores: + +$$ +\epsilon^ {*} = \arg \max _ {\epsilon_ {i} \in \mathcal {N} _ {s e t}} \left\{S \left(\mathcal {F} _ {i}, \mathcal {F} _ {O}\right) \mid \epsilon_ {i} \in \mathcal {N} _ {s e t} \right\}. \tag {4} +$$ + +Once the optimal noise sample $\epsilon^{*}$ is identified, it is fed into the generative model to produce an image that closely matches the user intent. In Tab. 1, we summarize several common generation goals achievable with NoiseQuery, including both high-level semantics and low-level visual properties. For each goal, we specify the corresponding + +feature types and matching functions. Detailed explanations of these features can be found in Appendix C. Notably, these goals can be applied individually or in combination. Users can specify multiple constraints simultaneously (e.g., generating a "sunset" image with a "warm color palette" while ensuring "high sharpness"). This flexibility makes NoiseQuery highly versatile and generalizable, enabling fine-grained control over the generated output. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We evaluate our approach on two popular benchmarks: MSCOCO [33] and DrawBench [47]. MSCOCO is a standard benchmark for text-to-image generation, providing textual descriptions of diverse natural scenes and object classes. We randomly select 10k captions as user prompts. DrawBench consists of 200 carefully designed prompts intended to challenge generative scenarios on various aspects, including long-form text, complex compositions, conflicting instructions, and rare vocabulary. + +Model Zoo. We conduct experiments on six popular diffusion generation models, including Stable diffusion (SD) series [43] (1.4, 1.5, 2.0, 2.1), PixArt- $\alpha$ [6], and SD-Turbo [49]. Stable Diffusion is a Latent Diffusion Model with UNet [44] backbone. By default, we use 50 DDIM denoising steps with a CFG scale of 7.0. PixArt- $\alpha$ is a Transformer Latent Diffusion Model, and SD-Turbo is a distillation model based on Stable Diffusion 2.1, generating high-quality images even with only 1-2 sampling steps without CFG. For PixArt- $\alpha$ , we use the default CFG scale of 4.5. + +Evaluation Metrics. For semantic similarity, we use ClipScore [42], which directly measures the alignment between the generated images and the corresponding textual prompts. Additionally, we incorporate several popular metrics trained on large-scale human preference data to quantify human preferences for AI-generated images, including HPS v2 [58], PickScore [27], and ImageReward [60]. These metrics offer a thorough evaluation of image quality, aesthetics, and alignment with user intent. + +# 4.2. Improvement on High-level Semantics + +For semantic consistency goals, we utilize features like CLIP [42] and BLIP [29] as the query to identify the semantically matched noise for each text prompt. Such aligned text and noise yield two benefits: (i) prevents the model from generating irrelevant outputs or falling into unsolvable regions. (ii) minimizes generation difficulty and maximizes the likelihood for semantic consistency. + +Quantitative Analysis. In Tab. 2, we evaluate our method on standard text-to-image (T2I) benchmarks, demonstrating consistent performance gains across all base models by offering better noise-text alignment. NoiseQuery + +
Base ModelMethodDrawBench [47]MSCOCO [33]Time Cost
ImageRewardPickScoreHPS v2CLIPScoreImageRewardPickScoreHPS v2CLIPScore
SD 1.5Base Model0.0421.1124.5730.900.1521.4125.6531.081.334 s
+ NoiseQuery0.0821.1625.0231.410.2721.4826.0731.471.336 s
+ Diffusion-DPO [53]0.0921.2925.0231.190.2521.6426.3131.261.350 s
+ Diffusion-DPO [53] + NoiseQuery0.1721.3325.2531.410.3521.6826.6031.551.352 s
SD 2.1Base Model0.1221.3324.9331.130.3621.7226.5831.401.301 s
+ NoiseQuery0.2621.4625.3931.680.4421.7626.8231.501.303 s
+ CFG++ [10]0.1221.3324.8331.130.3721.7226.6631.313.724 s
+ CFG++ [10] + NoiseQuery0.2721.4325.5531.610.4721.7626.9731.673.726 s
SD-TurboBase Model0.2621.7825.2331.290.4722.0726.2231.510.072 s
+ NoiseQuery0.4121.8725.6631.580.5022.1726.8231.760.074 s
+ ReNO [11]1.6723.4032.4832.55----23.56 s
+ ReNO [11] + NoiseQuery1.7123.5232.9232.78----23.56 s
PixArt-αBase Model0.7022.0828.2730.830.7822.2429.3331.484.327 s
+ NoiseQuery0.8222.1128.4531.270.7922.3329.5631.644.328 s
+ LaVi-Bridge [67]0.6322.0828.3530.920.7522.3129.4931.865.092 s
+ LaVi-Bridge [67] + NoiseQuery0.7222.2428.6131.350.7822.3529.6732.015.094 s
+ +Table 2. Evaluation of objective metrics on different datasets and models. Higher is better for all metrics. Our approach enhances the base model's performance and complements a wide range of T2I enhancement methods, including DPO [53] (reward-based fine-tuning), CFG++ [10] (guidance refinement), ReNO [11] (gradient-based initial noise optimization), and LaVi-Bridge [67] (text encoding enhancement via LORA/adapters). $\cdot$ : ReNO results on MSCOCO are excluded due to excessive inference time. + +![](images/507911a5084e97e7d2c18ebdec3bacd774e3943bfecd99089ef7ffa969a75f86.jpg) +Figure 5. Generated images of the base model, base model with T2I enhancement methods, and base model with enhancement and NoiseQuery. The four rows, from top to bottom, respectively correspond to the method groups listed in Table 2: Diffusion-DPO [53], CFG++ [10], ReNO [11], and LaVi-Bridge [67]. + +also serves as a universal foundation layer, complementing T2I enhancement techniques like Diffusion-DPO [53], CFG++ [10], ReNO [11], and LaVi-Bridge [67]. Notably, these methods face inherent limitations: model-specific dependencies (e.g., Diffusion-DPO and LaVi-Bridge require retraining for new models; ReNO is restricted to one-step models due to heavy computation, excluding com + +![](images/5f014cb793cc04e6966c74a1b523c31602db45805889a79962dfe3d5bd2f12bd.jpg) +Figure 6. Queried noise v.s. random noise. + +![](images/9786536d46520ccf67ca7e9fb178aefe4c79983cad0fdf1da93dd27f3d668e24.jpg) +Figure 7. CLIP and FID evaluations across CFG scales on MSCOCO [33], with the CFG scale increasing from left to right. + +mon multi-step models like SD) and high latency. In contrast, our method seamlessly integrates with them, enhancing their performance while adding near-zero time costs. We also present qualitative results in Fig. 5 to illustrate its effectiveness in assisting text-to-image synthesis. Additionally, we show the zero-shot transferability of NoiseQuery across various generative models in Appendix B.1. This model-agnostic compatibility, combined with minimal computational cost, establishes NoiseQuery as a versatile and efficient enhancement tool for diverse pipelines. + +Diversity Analysis. We report the performance of selecting the top 20 matched noise samples per prompt instead of + +![](images/3829f9940f18480604f7a8e324a6b834b2ff3bd5ecc49dc4d023673efa04d36c.jpg) +(a) Color properties guidance. + +![](images/d182bdc33a273b2555ab7508296ae040025e02acf5f3f67e8c4ff71962ca9f65.jpg) +(b) Structural features guidance. + +
Library Size0.5k1k2k5k10k50k100k
Matching Function Cost (×10-4s)1.391.391.391.391.401.401.51
Argmax Selection Cost (×10-4s)0.250.250.250.250.376.1613.25
CLIP Score31.5131.5331.5731.5931.6631.7331.74
+ +Table 3. Retrieval time breakdown across various library sizes. + +the single best match. As shown in Fig. 6, these samples consistently outperform the standard random noise baseline on DrawBench [47], demonstrating that NoiseQuery provides not only optimal but also diverse high-quality noise candidates. To quantify diversity, we use the Diversity Metric (DIV) [57], computed as the standard deviation of DINOV2 [38] embeddings across 20 generated images per prompt (see Fig. 6 legend). Our method achieves comparable diversity to random noise, confirming that NoiseQuery enhances image quality without sacrificing diversity. + +Different Classifier-free Guidance Scale Analysis. We assess our method's performance across various classifier-free guidance (CFG) scales (1.5, 2.0, 3.0, ..., 15.0) on 10k MSCOCO [33] samples. Our NoiseQuery achieves high-quality outputs even with low guidance scales, as demonstrated in Fig. 7. This supports our hypothesis that by selecting more compatible initial noise, we can effectively reduce the generation difficulty. Therefore, we can obtain high-quality output at low scales and avoid the over-saturation and instability that usually come with high scales. + +Time Cost. We also provide a cost breakdown for the retrieval process in Tab. 3, which includes the computation of the matching function (utilizing query CLIP features at no extra cost since CLIP is embedded in SD) and argmax se + +![](images/7f31b3e9b0aa0fa9e7f2001036b111dd3d953ec3909ae0044680d96629130150.jpg) +Figure 8. Multiple low-level visual properties control through NoiseQuery. All images use minimal text prompts containing only semantic concepts (e.g., object labels), lacking explicit visual specifications. +Figure 9. Text prompts often fail to align generated images with user low-level attribute preferences, while initial noise plays a dominant role in controlling them. + +lection. The results demonstrate that both steps are efficient, even with large libraries (e.g., 0.0015 seconds for 100,000 entries), indicating that the additional cost is negligible. + +# 4.3. Controllability on Low-Level Visual Properties + +Text prompts effectively steer high-level semantics in text-to-image (T2I) generation but often fail to provide sufficient control over low-level visual properties that affect visual perception. This limitation stems from two factors: (1) text is inherently inadequate for conveying detailed visual reference input, and (2) text encoders like CLIP [42], trained primarily on high-level semantics, struggle to capture and control fine-grained low-level properties such as color, texture, shape, and sharpness. + +In contrast, initial noise plays a dominant role in determining these properties, making it a valuable resource for precise control, as illustrated in Fig. 9. Our method leverages this by enabling fine-grained control of low-level attributes through initial noise, without requiring additional training. The following experiments demonstrate the effectiveness of our approach in achieving these targets, with fur + +![](images/b7a2d3de948f45cbf79e86d5414c8a7595de0bebb8affc1d444dd8563c462966.jpg) +Diverse User Demands +Stable Diffusion +Stable Diffusion (enhanced prompt) +NoiseQuery (ours) +Figure 10. Examples of combining high-level semantics (red) and low-level properties (blue). Enhanced prompts are created by fusing semantic objects and visual details via a large language model. + +![](images/02bc7ed3fb8653715ab9059a20cbc8579d9479ccf964c508d5655e57e8ffff1a.jpg) +Figure 11. Compatibility with controllable models such as TI [15] and ControlNet [64]. Our method provides a better initial noise, enhancing semantic consistency and low-level attribute control. + +ther implementation details provided in Appendix C. + +Color Properties. Color attributes such as brightness, saturation, contrast, and hue are crucial for an image's visual appeal. We assess the overall color hue by averaging each RGB channel, measure saturation and brightness in the HSV color space, and compute contrast as the standard deviation of grayscale intensities. As shown in Fig. 8a, selecting initial noise that matches target color distributions enables precise control over the generated images' color tendencies, aligning them with user preferences. + +Structural Features. For spatial control, we implement: (1) Texture through Gray-Level Co-occurrence Matrix (GLCM) [19] features that quantify intensity co-occurrence patterns; (2) Shape consistency via Hu Moments [22] as rotation-invariant descriptors; and (3) Sharpness by prioritizing high-frequency energy (HFE) in Fourier spectra. Fig. 8b demonstrates how noise selection governs these structural features—preserving geometric fidelity in object shapes, modulating surface textures, and enhancing edge crispness—all achieved without prompt engineering. + +# 4.4. Enhancement on Combined Scenarios + +Building upon the results in previous sections, which demonstrate our method's effectiveness in high-level semantic alignment and low-level attribute control, this section highlights its ability to combine these diverse goals and seamlessly integrate with powerful existing controllable generation models. + +Combined Goals. Beyond individual tasks, our method excels in simultaneously achieving multiple generation goals. By sequentially reranking noise candidates—first selecting top candidates based on one goal (e.g., high-level semantics) and then refining within this subset for another goal (e.g., low-level attributes)—we progressively narrow down the optimal noise pool. As shown in Fig. 10, this approach allows the combination of high-level semantic alignment with various low-level attribute controls, enabling the generation of images that meet diverse user preferences. + +Compatibility with Controlling Methods. Our method is designed to be instant-time and model-agnostic, making it highly compatible with existing controllable generation models such as Textual Inversion (TI) [15] and ControlNet [64]. By providing a better initial noise, our approach establishes a foundational layer for controllable generation, which synergizes with subsequent refinement processes. As shown in Fig. 11, this integration enhances semantic consistency while enabling precise control over challenging low-level attributes, such as color and texture. + +# 5. Conclusion + +In this work, we propose NoiseQuery, a tuning-free method that enables versatile goal-driven T2I generation through enhanced noise initialization, supported by a generic and reusable pre-built noise library. By selecting Gaussian initialization in the noise library better aligned with the text prompt or specified attributes, NoiseQuery achieves fine-grained control over both high-level semantic and intricate low-level details. In addition, the model-agnostic nature of our method allows it to be seamlessly integrated with existing T2I augmentation and controllable generation methods without model adjustments or significant computational overhead. Experiments validate its effectiveness in improving output fidelity and precision, particularly for visual attributes that are challenging to articulate textually. The lightweight and scalability of the framework make it adaptable to real-world applications. + +Limitations. Although NoiseQuery provides flexibility for aligning high-level semantics and controlling low-level attributes, it remains bounded by the base model's ability, and may still underperform in extremely complex cases. Additionally, it has limitations in achieving fine-grained control, such as Canny edge control. + +# Acknowledgements + +This work was partially supported by the National Natural Science Foundation of China under grant 62372341. This work was also partially supported by an Amazon Research Award 2023 to OR and YZ. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not reflect the views of Amazon. + +# References + +[1] Donghoon Ahn, Jiwon Kang, Sanghyun Lee, Jaewon Min, Minjae Kim, Wooseok Jang, Hyoungwon Cho, Sayak Paul, SeonHwa Kim, Eunju Cha, et al. A noise is worth diffusion guidance. arXiv, 2024. 3 +[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In ICLR, 2024. 1, 3 +[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG), 2023. 1, 3 +[4] Changgu Chen, Libing Yang, Xiaoyan Yang, Lianggangxu Chen, Gaoqi He, Changbo Wang, and Yang Li. Find: Finetuning initial noise distribution with policy optimization for diffusion models. In ACM MM, 2024. 1, 3, 4 +[5] Haoyang Chen, Dongfang Sun, Caoyuan Ma, Shiqin Wang, Kewei Zhang, Zheng Wang, and Zhixiang Wang. Subjective camera: Bridging human cognition and visual reconstruction through sequence-aware sketch-guided diffusion. arXiv, 2025. 3 +[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2, 4, 5, 1 +[7] Sitan Chen, Giannis Daras, and Alex Dimakis. Restoration-degradation beyond linear diffusions: A non-asymptotic analysis for ddim-type samplers. In ICML, 2023. 2 +[8] Zijiao Chen, Jiaxin Qing, Tiange Xiang, Wan Lin Yue, and Juan Helen Zhou. Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding. In CVPR, 2023. 3 +[9] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv, 2023. 3 +[10] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv, 2024. 1, 2, 3, 6 +[11] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. NeurIPS, 37:125487-125519, 2024. 1, 2, 3, 4, 6 + +[12] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. NeurIPS, 2023. 1, 3 +[13] Yingying Fan, Quanwei Yang, Kaisiyuan Wang, Hang Zhou, Yingying Li, Haocheng Feng, Errui Ding, Yu Wu, and Jingdong Wang. Re-hold: Video hand object interaction reenactment via adaptive layout-instructed diffusion model. In CVPR, 2025. 3 +[14] Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. In ECCV, 2024. 3 +[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 1, 2, 3, 8 +[16] Leon A Gatys. A neural algorithm of artistic style. arXiv, 2015. 5 +[17] Or Greenberg, Eran Kishon, and Dani Lischinski. S2st: Image-to-image translation in the seed space of latent diffusion. arXiv, 2023. 3, 4 +[18] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024. 1, 3, 4 +[19] Robert M Haralick, Karthikeyan Shanmugam, and Its' Hak Dinstein. Textural features for image classification. IEEE SMCS, 1973. 5, 8, 2 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 1, 3, 4 +[21] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In ICCV, 2023. 3 +[22] Ming-Kuei Hu. Visual pattern recognition by moment invariants. IRE transactions on information theory, 1962. 5, 8, 2 +[23] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv, 2024. 1 +[24] Huayang Huang, Yu Wu, and Qian Wang. Robin: Robust and invisible watermarks for diffusion models with adversarial optimization. NIPS, 2024. 3 +[25] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan, and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. In SIGGRAPH Asia, 2024. 3 +[26] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. 1 +[27] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. NeurIPS, 2023. 5 +[28] Gen Li, Yuting Wei, Yuxin Chen, and Yuejie Chi. Towards non-asymptotic convergence for diffusion-based generative models. In ICLR, 2024. 2 + +[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML. PMLR, 2022. 5 +[30] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. NeurIPS, 2024. 3 +[31] Yumeng Li, Margret Keuper, Dan Zhang, and Anna Khoreva. Divide & bind your attention for improved generative semantic nursing. arXiv, 2023. 3 +[32] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In WACV, 2024. 4, 5 +[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 6, 7, 1 +[34] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In ICLR, 2022. 1 +[35] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. NeurIPS, 2022. 1 +[36] Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, et al. Inference-time scaling for diffusion models beyond scaling denoising steps. arXiv, 2025. 3 +[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3, 4 +[38] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024. 7 +[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 1 +[40] Yuhan Pei, Ruoyu Wang, Yongqi Yang, Ye Zhu, Olga Russakovsky, and Yu Wu. Sowing information: Cultivating contextual coherence with mllms in image generation. arXiv, 2024. 3 +[41] Zipeng Qi, Lichen Bai, Haoyi Xiong, et al. Not all noises are created equally: Diffusion noise selection and optimization. arXiv, 2024. 1, 3, 4 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML. PMLR, 2021. 5, 7 +[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4, 5, 1 +[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI. Springer, 2015. 5 + +[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 3 +[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, 2024. 3 +[47] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 5, 6, 7, 1 +[48] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In AAAI, 2024. 3 +[49] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 2, 5, 1 +[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 3 +[51] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 3 +[52] Zhiwei Tang, Jiangweizhi Peng, Jiasheng Tang, Mingyi Hong, Fan Wang, and Tsung-Hui Chang. Inference-time alignment of diffusion models with direct noise optimization. arXiv, 2024. 3 +[53] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, pages 8228-8238, 2024. 1, 2, 3, 6 +[54] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv, 2024. 3 +[55] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv, 2024. 3 +[56] Ruoyu Wang, Yongqi Yang, Zhihao Qian, Ye Zhu, and Yu Wu. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, 2024. 3 +[57] Xi Wang, Nicolas Dufour, Nefeli Andreou, Marie-Paule Cani, Victoria Fernández Abrevaya, David Picard, and Vicky Kalogeiton. Analysis of classifier-free guidance weight schedulers. arXiv, 2024. 7 +[58] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv, 2023. 3, 5 +[59] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. *IJCV*, 2024. 3 + +[60] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. NeurIPS, 2024. 5 +[61] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models. arXiv, 2024. 3 +[62] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3 +[63] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv, 2023. 3 +[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 8 +[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Ic-light github page, 2024. 3 +[66] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2024. 1, 2, 3 +[67] Shihao Zhao, Shaozhe Hao, Bojia Zi, Huaizhe Xu, and Kwan-Yee K Wong. Bridging different language models and generative vision models for text-to-image generation. In ECCV, 2024. 1, 2, 3, 6 +[68] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. NeurIPS, 2024. 1 +[69] Chenyi Zhuang, Ying Hu, and Pan Gao. Magnet: We never know how text-to-image diffusion models work, until we learn how vision-language models function. NeurIPS, 2024. 3 \ No newline at end of file diff --git a/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/images.zip b/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4555936364b208f3ad1f4a32d3399fc2c07a7826 --- /dev/null +++ b/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41529092cb9fed4d627578a1b1ae0b1df84a796470f875ed14cb21c512110e6f +size 877274 diff --git a/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/layout.json b/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..caf01f4cb5d2ab65d1493c09990dd87928a7d8a5 --- /dev/null +++ b/2025/The Silent Assistant_ NoiseQuery as Implicit Guidance for Goal-Driven Image Generation/layout.json @@ -0,0 +1,8537 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 484, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 484, + 140 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 484, + 140 + ], + "type": "text", + "content": "The Silent Assistant: NoiseQuery as Implicit Guidance for Goal-Driven Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 160, + 156, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 160, + 156, + 176 + ], + "spans": [ + { + "bbox": [ + 87, + 160, + 156, + 176 + ], + "type": "text", + "content": "Ruoyu Wang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 161, + 266, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 161, + 266, + 176 + ], + "spans": [ + { + "bbox": [ + 182, + 161, + 266, + 176 + ], + "type": "text", + "content": "Huayang Huang" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 293, + 161, + 335, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 161, + 335, + 175 + ], + "spans": [ + { + "bbox": [ + 293, + 161, + 335, + 175 + ], + "type": "text", + "content": "Ye Zhu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 361, + 161, + 457, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 161, + 457, + 176 + ], + "spans": [ + { + "bbox": [ + 361, + 161, + 457, + 176 + ], + "type": "text", + "content": "Olga Russakovsky" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 482, + 161, + 527, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 482, + 161, + 527, + 175 + ], + "spans": [ + { + "bbox": [ + 482, + 161, + 527, + 175 + ], + "type": "text", + "content": "Yu " + }, + { + "bbox": [ + 482, + 161, + 527, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{W}\\mathbf{u}^{1*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 169, + 189, + 441, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 189, + 441, + 217 + ], + "spans": [ + { + "bbox": [ + 169, + 189, + 441, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 169, + 189, + 441, + 217 + ], + "type": "text", + "content": "School of Computer Science, Wuhan University " + }, + { + "bbox": [ + 169, + 189, + 441, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 169, + 189, + 441, + 217 + ], + "type": "text", + "content": "Department of Computer Science, Princeton University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 200, + 220, + 410, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 220, + 410, + 232 + ], + "spans": [ + { + "bbox": [ + 200, + 220, + 410, + 232 + ], + "type": "text", + "content": "{wangruoyu, hyhuang, wuyucs}@whu.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 223, + 234, + 384, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 234, + 384, + 245 + ], + "spans": [ + { + "bbox": [ + 223, + 234, + 384, + 245 + ], + "type": "text", + "content": "{yezhu, olgarus}@princeton.edu" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 300, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 300, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 53, + 300, + 296, + 563 + ], + "type": "text", + "content": "In this work, we introduce NoiseQuery as a novel method for enhanced noise initialization in versatile goal-driven text-to-image (T2I) generation. Specifically, we propose to leverage an aligned Gaussian noise as implicit guidance to complement explicit user-defined inputs, such as text prompts, for better generation quality and controllability. Unlike existing noise optimization methods designed for specific models, our approach is grounded in a fundamental examination of the generic finite-step noise scheduler design in diffusion formulation, allowing better generalization across different diffusion-based architectures in a tuning-free manner. This model-agnostic nature allows us to construct a reusable noise library compatible with multiple T2I models and enhancement techniques, serving as a foundational layer for more effective generation. Extensive experiments demonstrate that NoiseQuery enables fine-grained control and yields significant performance boosts not only over high-level semantics but also over low-level visual attributes, which are typically difficult to specify through text alone, with seamless integration into current workflows with minimal computational overhead. Code is available at https://github.com/wangruoyu02/NoiseQuery." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 587, + 135, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 135, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 135, + 600 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 608, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 295, + 693 + ], + "type": "text", + "content": "Text-to-image synthesis (T2I) has gained significant research attention with the rising popularity of diffusion generative models [20, 39, 50, 51]. While these models demonstrate impressive generative capabilities, they still struggle to faithfully align generated content with user prompts, leading to suboptimal outputs. To enrich the text-based conditioning capabilities of diffusion models, researchers have" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 270, + 553, + 466 + ], + "blocks": [ + { + "bbox": [ + 317, + 270, + 553, + 466 + ], + "lines": [ + { + "bbox": [ + 317, + 270, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 270, + 553, + 466 + ], + "type": "image", + "image_path": "6bf33591d212fda3a8f71277ee39ccc73e80d4093a26b2fa0f0da3a67e83c151.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 472, + 555, + 528 + ], + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 528 + ], + "type": "text", + "content": "Figure 1. Each row shows a set of images generated by different diffusion models starting from the same initial noise, using the same sampler, and with NULL as the text prompt. The generated images demonstrate striking both high-level semantic and low-level visual similarities across models." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 543, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 555, + 674 + ], + "type": "text", + "content": "explored various techniques, including fine-tuning diffusion UNet [2, 12, 53], enhancing language encoders [23, 67], refining inference strategies [3, 10], and extra additional image conditions [15, 45, 66]. A recent popular line of research focuses on optimizing the initial noise [4, 11, 18, 41], motivated by the observation that images generated from the same noise often exhibit high similarities within a single model, even when prompted differently. In this work, we extend this understanding by revealing that such noise-induced consistency persists across T2I models with different diffusion backbones, as illustrated in Fig. 1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "To better understand this phenomenon, we examine the formulation of diffusion models and reveal that cross-model similarities stem from the finite-step noise scheduler design" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17618" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "content": "and persist across different diffusion variants. Specifically, one of the fundamental assumptions in diffusion models lies in a gradual stepwise perturbation process that transforms a real-world distribution into a standard Gaussian. This transition requires, in theory, an infinite number of diffusion steps to perfectly obscure the raw images, contradicting the practical implementation where we can only apply a finite number (i.e., non-asymptotic) of degradations through predefined noise schedulers [7, 28]. As a result, the real-world diffusion models inadvertently learn to leverage residual information embedded in image data, creating an unintended link between noise and data regardless of the model architecture during inference. However, rather than being a limitation, we leverage this seemingly imperfect model design and unintentional connection as an unexpected advantage by laying the foundation for using stochastic noise as a generic \"silent assistant\", which implicitly encodes clues about the image that will be generated." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 291, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 471 + ], + "type": "text", + "content": "We note that it is critical to align the implicit noise clues with the user-provided input, such as text prompts, in order to achieve optimal generation performance. As shown in Fig. 2, when the inherent tendencies of the noise guidance and text prompt diverge significantly, the generation process leads to instability or semantically inconsistent results. Furthermore, unlike text prompts that convey high-level semantics, initial noise inherently encodes low-level visual primitives—such as color gradients, texture patterns, and spatial frequencies—directly tied to the pixel space. This makes it particularly suitable for controlling fine-grained visual properties that are difficult to specify through text alone. By leveraging initial noise as a complementary control signal, we produce more nuanced and precise outputs, enhancing the overall generative capability of T2I models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "content": "Building on this insight, we propose NoiseQuery, a novel method that exploits initial noise as a shared generative asset across diverse T2I models. The key challenge lies in establishing an interpretable mapping between stochastic noise patterns and their concrete generative clues. To bridge this gap, we utilize images generated with empty NULL text prompts as proxies to reveal the hidden information encoded in the initial noise (details in Sec. 3.3), denoted as the implicit generative posterior. When deprived of text guidance, the generation process becomes purely noise-driven, forcing the model to excavate and decode the residual signals preserved in the initial noise. This process enables the construction of a comprehensive Noise Library, built offline by generating a diverse set of such noise-driven images from a Gaussian distribution and extracting their multi-grained features, spanning high-level semantics and low-level visual attributes. During inference, given a specific user request (single or combined goals), we extract its features, perform a quick lookup in the pre-built library, and select the best-matching noise for initialization. Notably," + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 326, + 71, + 541, + 247 + ], + "blocks": [ + { + "bbox": [ + 326, + 71, + 541, + 247 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 541, + 247 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 541, + 247 + ], + "type": "image", + "image_path": "a640fc88760112795e8545a15c6c692a469507d4d32647ae37f9e926c2d0e5a3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 252, + 555, + 286 + ], + "lines": [ + { + "bbox": [ + 313, + 252, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 555, + 286 + ], + "type": "text", + "content": "Figure 2. Compared to mismatched noise, noise aligned with text prompts can produce more accurate images in complex T2I generation. A NULL prompt represents unconditional generation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 307, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 415 + ], + "type": "text", + "content": "NoiseQuery is model-agnostic and serves as a reusable resource across multiple diffusion architectures. This ability makes it naturally compatible with diverse enhancement strategies including recent noise optimization approaches, providing a foundational layer for more effective generation. Moreover, NoiseQuery operates at a negligible online computational cost (0.002 seconds per prompt), ensuring seamless integration into existing workflows while enhancing the controllability and quality of T2I generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 415, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 582 + ], + "type": "text", + "content": "To validate the effectiveness of NoiseQuery, we conduct experiments on low-level property preferences, high-level semantic consistency, and their combinations. Extensive evaluations with six diffusion-based T2I models [6, 43, 49], four types of T2I enhancement methods [10, 11, 53, 67] and two controllable T2I models [15, 66] show that NoiseQuery not only improve the base model's performance but also complements multiple T2I enhancement methods. Our approach is also among the first to address the challenge of subtle low-level preference control with diffusion models, highlighting a new avenue for fine-grained generative tasks. Furthermore, by progressively matching the noise, our method is capable of supporting combinations of generation goals, offering more nuanced control for users." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 582, + 511, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 582, + 511, + 593 + ], + "spans": [ + { + "bbox": [ + 326, + 582, + 511, + 593 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 594, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 553, + 641 + ], + "type": "text", + "content": "- We reveal the generic generative tendencies within the initial noise and leverage this implicit guidance as a universal generative asset that can be seamlessly integrated with existing T2I models and enhancement methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 701 + ], + "type": "text", + "content": "- We introduce NoiseQuery, a novel method that retrieves an optimal initial noise from a pre-built large-scale noise library, fulfilling versatile user-specified requirements encompassing both semantic and low-level visual attributes—a relatively under-explored area." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 702, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 702, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 702, + 553, + 713 + ], + "type": "text", + "content": "- We conduct extensive experiments to demonstrate that" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17619" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 120 + ], + "type": "text", + "content": "our method can improve the general quality of diffusion-based generators and can significantly improve multiple T2I enhancement methods while maintaining a minimal cost and without requiring additional learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 131, + 141, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 141, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 141, + 143 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 151, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 296, + 534 + ], + "type": "text", + "content": "Improving Base Text-to-Image Generation. Recent breakthroughs in text-to-image (T2I) generation have been driven by diffusion-based models, yet these models still suffer from artifacts and misalignment with textual prompts. Current solutions primarily fall into two categories: tuning-based and inference-time methods. Tuning-based methods [2, 12, 30, 53, 62] optimize model parameters through techniques like direct preference optimization (DPO) [30, 53, 62]. Additionally, LaVi-Bridge [67] enhances textual understanding of T2I models by integrating advanced language models with parameter-efficient adapters. In contrast, inference-time methods [3, 10, 21, 31, 69] dynamically adjust the generation process. For example, CFG++ [10] improves image quality by addressing non-manifold issues in classifier-free guidance. Our approach complements both strategies by offering improved initialization for generation. Enhancing T2I with Additional Conditions. To address the limitations of text-only conditions in meeting diverse user needs, numerous works [5, 8, 9, 13, 15, 40, 45, 59, 64, 66] have expanded T2I generation through multi-modal conditions. For instance, methods like ControlNet [64] introduce structural signals (e.g., edge and depth maps) to control spatial layouts, while IC-Light [65] adjusts the illumination effects. Meanwhile, customization generation approaches focus on learning user-provided concepts like identity [15, 45, 46, 55, 63], style [14, 54], and interaction [25]. In this paper, we highlight that the initial noise for diffusion sampling can also serve as a conditional resource. Furthermore, compared to efforts that focus primarily on high-level semantic control, we show that initial noise holds untapped potential for guiding various low-level visual properties (e.g., color, texture, and sharpness)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": "Initial Noise Optimization for Diffusion Sampling. A growing body of research has focused on optimizing the initial noise in diffusion models, driven by the consensus that it significantly impacts the generated output [1, 24, 36, 56, 61]. These methods iteratively refine initial noise through backpropagation at inference, guided by different generative objectives, such as semantic alignment [4, 11, 18, 41, 52], rare concept generation [48] and image translation [17]. For instance, ReNO [11] optimizes noises using various human preference rewards like HPS v2 [58], to enhance image quality and align with user intent. However, these approaches rely on iterative gradient updates during test-time, which are computationally expensive and prone to numerical instability (e.g., gradient explosion). In this work, we observe that initial noise exhibits cross-model consistency" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 205 + ], + "type": "image", + "image_path": "232390aa8abb55d73837b546e102a652f165cea1c9b189971bdfbcb28318e54d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 212, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 313, + 212, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 553, + 235 + ], + "type": "text", + "content": "Figure 3. Using the same initial noise, the generative posteriors (left) exhibit similarity with the text-conditioned images (right)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 250, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 555, + 299 + ], + "type": "text", + "content": "(Fig. 1), making it an exploitable shared asset. Leveraging this property, we construct a reusable noise library that not only facilitates efficient noise selection but also helps existing noise optimization methods improve generation quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 307, + 397, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 397, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 397, + 320 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 327, + 438, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 438, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 438, + 338 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "text", + "content": "Given a pre-trained generative diffusion model " + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "text", + "content": ", which is capable of generating images " + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "text", + "content": " based on user-defined conditions " + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "text", + "content": ", typically a text prompt. The generative process starts with the initial noise " + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "text", + "content": " sampled from a Gaussian distribution " + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "text", + "content": " and gradually transforms this noise into a clean image through " + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "text", + "content": " denoising steps1. Our motivation is to construct a large-scale noise sample library, denoted as " + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{set} = \\{\\epsilon_1, \\epsilon_2, \\dots, \\epsilon_n\\}" + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "text", + "content": ", where each noise sample " + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\epsilon_i" + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "text", + "content": " has its hidden features pre-computed from the generative posteriors. The goal is to identify the most suitable noise " + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\epsilon^* \\in \\mathcal{N}_{set}" + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "text", + "content": " that aligns with a specific generation objective " + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 313, + 344, + 555, + 523 + ], + "type": "text", + "content": ". Notably, this generation objective is inherently open-ended and encompasses various dimensions such as semantic consistency, color properties, and other user-defined attributes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 529, + 493, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 493, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 493, + 542 + ], + "type": "text", + "content": "3.2. Implicit Guidance in Initial Noise" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 547, + 554, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 554, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 554, + 618 + ], + "type": "text", + "content": "To investigate the influence of initial noise on generated images, we analyze the forward noise addition process in diffusion models [20, 37, 50, 51]. The forward diffusion process gradually perturbs an image " + }, + { + "bbox": [ + 313, + 547, + 554, + 618 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 547, + 554, + 618 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 313, + 547, + 554, + 618 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 547, + 554, + 618 + ], + "type": "text", + "content": " timesteps, transforming it into a \"pure\" Gaussian noise " + }, + { + "bbox": [ + 313, + 547, + 554, + 618 + ], + "type": "inline_equation", + "content": "x_T" + }, + { + "bbox": [ + 313, + 547, + 554, + 618 + ], + "type": "text", + "content": ". The noise addition at each step follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 356, + 625, + 553, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 625, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 356, + 625, + 553, + 639 + ], + "type": "interline_equation", + "content": "q \\left(x _ {t} \\mid x _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {1 - \\beta_ {t}} x _ {t - 1}, \\beta_ {t} \\mathbf {I}\\right), \\tag {1}", + "image_path": "cdc79e0585f5799e5ce0c298cf5318541bd3a0f458fa0124a5b1b648174eee70.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "type": "text", + "content": " is a predefined variances schedule. Defining " + }, + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta" + }, + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "type": "text", + "content": ", the cumulative effect of noise over " + }, + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "type": "text", + "content": " steps results in:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 379, + 674, + 553, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 674, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 379, + 674, + 553, + 689 + ], + "type": "interline_equation", + "content": "x _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} x _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\tag {2}", + "image_path": "f865769c6b4061141ebcc7cf5954186d6bf7a1227d81f7ed040364a973b451dd.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": "1For simplicity, we omit the process of transitioning from pixel space to latent space in latent diffusion models (LDM)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17620" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 83, + 517, + 230 + ], + "blocks": [ + { + "bbox": [ + 184, + 70, + 309, + 81 + ], + "lines": [ + { + "bbox": [ + 184, + 70, + 309, + 81 + ], + "spans": [ + { + "bbox": [ + 184, + 70, + 309, + 81 + ], + "type": "text", + "content": "Offline Library Construction" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 94, + 83, + 517, + 230 + ], + "lines": [ + { + "bbox": [ + 94, + 83, + 517, + 230 + ], + "spans": [ + { + "bbox": [ + 94, + 83, + 517, + 230 + ], + "type": "image", + "image_path": "9403d1b910c1e857a0739e9aa87fbc8bc7ebecfbaf92e68ef57c475d2403bd3b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 232, + 515, + 243 + ], + "lines": [ + { + "bbox": [ + 94, + 232, + 515, + 243 + ], + "spans": [ + { + "bbox": [ + 94, + 232, + 515, + 243 + ], + "type": "text", + "content": "Figure 4. Illustration of our NoiseQuery pipeline which involves offline library construction and real-time inference." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{i=1}^t \\alpha_i" + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": ". A fundamental assumption in diffusion theory is that " + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_T \\to 0" + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "inline_equation", + "content": "T \\to \\infty" + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": ", ensuring that the raw image is fully obscured. However, practical implementations can only apply a finite number of degradations (e.g., " + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": ") to approximate this ideal. As analyzed in prior work [32], common schedules (e.g., Linear [20], Cosine [37], Stable Diffusion [43]) fail to achieve zero signal-to-noise ratio (SNR) at " + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": ", leaving residual signal from " + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "inline_equation", + "content": "x_T" + }, + { + "bbox": [ + 55, + 264, + 296, + 384 + ], + "type": "text", + "content": ". For instance, Stable Diffusion's final timestep retains:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 393, + 294, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 393, + 294, + 405 + ], + "spans": [ + { + "bbox": [ + 101, + 393, + 294, + 405 + ], + "type": "interline_equation", + "content": "x _ {T} = 0. 0 6 8 2 6 5 \\cdot x _ {0} + 0. 9 9 7 6 6 7 \\cdot \\epsilon , \\tag {3}", + "image_path": "0bc71e6460fd909dcde5b904f142f44a709a732accb5dc1ed956d3eaa030fbe5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 414, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 546 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 414, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\sqrt{\\overline{\\alpha}_T} = 0.068265" + }, + { + "bbox": [ + 55, + 414, + 296, + 546 + ], + "type": "text", + "content": " quantifies the leakage. During training, diffusion models minimize reconstruction loss by utilizing these residuals as shortcuts rather than predicting the original image from scratch. This creates an unexpected link between noise and data, regardless of the model architecture. As a result, during inference, models continue to rely on the implicit knowledge gained during training to interpret pure Gaussian noise. We take advantage of this seemingly flawed model design and the unintended connection, using stochastic noise as a generic \"silent assistant\" that encodes clues about the image to be generated." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 289, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 289, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 289, + 564 + ], + "type": "text", + "content": "3.3. Dissect Initial Noise via Generation Posterior" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 700 + ], + "type": "text", + "content": "While recent works [4, 11, 17, 18, 41] optimize initial noise via gradient back-propagation for text-guided generation, they often overlook the inherent generative tendencies embedded in different noises. In other words, if the initial noise deviates significantly from the desired objective, iterative optimization struggles to bridge this gap, resulting in suboptimal convergence despite extensive tuning. In this subsection, we propose a more efficient alternative: directly analyzing the initial noise itself without the text prompt. Additionally, our method also yields better starting points for optimization, leading to faster convergence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": "Generative Posterior. When the model generates images" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 268, + 536, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 268, + 536, + 279 + ], + "spans": [ + { + "bbox": [ + 315, + 268, + 536, + 279 + ], + "type": "text", + "content": "Algorithm 1: NoiseQuery for Goal-Driven Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 281, + 537, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 281, + 537, + 314 + ], + "spans": [ + { + "bbox": [ + 315, + 281, + 537, + 314 + ], + "type": "text", + "content": "Input: Diffusion model " + }, + { + "bbox": [ + 315, + 281, + 537, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 315, + 281, + 537, + 314 + ], + "type": "text", + "content": "; Noise library " + }, + { + "bbox": [ + 315, + 281, + 537, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{set} = \\{\\epsilon_1, \\epsilon_2, \\dots, \\epsilon_n\\}" + }, + { + "bbox": [ + 315, + 281, + 537, + 314 + ], + "type": "text", + "content": "; Generation objective " + }, + { + "bbox": [ + 315, + 281, + 537, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 315, + 281, + 537, + 314 + ], + "type": "text", + "content": " and corresponding matching function " + }, + { + "bbox": [ + 315, + 281, + 537, + 314 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 315, + 281, + 537, + 314 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 314, + 499, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 314, + 499, + 325 + ], + "spans": [ + { + "bbox": [ + 315, + 314, + 499, + 325 + ], + "type": "text", + "content": "Output: Optimal noise sample " + }, + { + "bbox": [ + 315, + 314, + 499, + 325 + ], + "type": "inline_equation", + "content": "\\epsilon^{*}" + }, + { + "bbox": [ + 315, + 314, + 499, + 325 + ], + "type": "text", + "content": " for objective " + }, + { + "bbox": [ + 315, + 314, + 499, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 315, + 314, + 499, + 325 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 327, + 432, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 327, + 432, + 338 + ], + "spans": [ + { + "bbox": [ + 315, + 327, + 432, + 338 + ], + "type": "text", + "content": "Offline Library Construction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "spans": [ + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "type": "text", + "content": "foreach " + }, + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "type": "inline_equation", + "content": "\\epsilon_{i}\\in \\mathcal{N}_{set}" + }, + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "type": "text", + "content": " do " + }, + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i^{\\mathrm{uncond}} = \\mathcal{M}(\\epsilon_i,c = \\emptyset)" + }, + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "type": "text", + "content": " Extract features " + }, + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "type": "text", + "content": " (as detailed in Tab. 1) from " + }, + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_i^{\\mathrm{uncond}}" + }, + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "type": "text", + "content": " Store " + }, + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "type": "inline_equation", + "content": "(\\epsilon_i,\\mathcal{F}_i)" + }, + { + "bbox": [ + 315, + 339, + 532, + 383 + ], + "type": "text", + "content": " in the noise library." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 384, + 334, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 384, + 334, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 384, + 334, + 392 + ], + "type": "text", + "content": "end" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 404, + 396, + 413 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 404, + 396, + 413 + ], + "spans": [ + { + "bbox": [ + 318, + 404, + 396, + 413 + ], + "type": "text", + "content": "Real-time Inference" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 415, + 498, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 498, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 498, + 437 + ], + "type": "text", + "content": "Extract desired features " + }, + { + "bbox": [ + 317, + 415, + 498, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_O" + }, + { + "bbox": [ + 317, + 415, + 498, + 437 + ], + "type": "text", + "content": " from objective " + }, + { + "bbox": [ + 317, + 415, + 498, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 317, + 415, + 498, + 437 + ], + "type": "inline_equation", + "content": "\\epsilon^{*} = \\arg \\max_{\\epsilon_{i}\\in \\mathcal{N}_{set}}\\{S(\\mathcal{F}_{i},\\mathcal{F}_{O})\\mid \\epsilon_{i}\\in \\mathcal{N}_{set}\\}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 442, + 354, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 442, + 354, + 452 + ], + "spans": [ + { + "bbox": [ + 315, + 442, + 354, + 452 + ], + "type": "text", + "content": "return " + }, + { + "bbox": [ + 315, + 442, + 354, + 452 + ], + "type": "inline_equation", + "content": "\\epsilon^{*}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 483, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 614 + ], + "type": "text", + "content": "without any specific prompt, it depends solely on the noise characteristics, producing what we term \"generative posterior\". As shown in Fig. 3, generative posteriors exhibit clear generative tendencies (e.g., semantics, spatial layouts, and color palettes) similar to those of text-guided outputs. These similarities suggest that the initial noise carries significant latent information, influencing both unconditional generation and conditional generation in a similar manner, acting as an additional form of guidance alongside the text prompt. Thus, these purely noise-guided images can serve as an effective proxy for interpreting the initial noise." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Cross-model Consistency. We also observe another meaningful phenomenon: the impact of initial noise is universal and model-agnostic. As illustrated in Fig. 1, generative posteriors generated from the same noise remain highly similar across models with different architectures and parameters. For example, PixArt-" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": " [6] is a Transformer-based T2I diffusion model, while Stable Diffusion [43] is UNet-based. More discussion of models and samplers can be found in" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17621" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 304, + 145 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 304, + 145 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 304, + 145 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 304, + 145 + ], + "type": "table", + "html": "
Generation GoalsFeature TypeMatch Function
SemanticsCLIP [42], BLIP [29]Cosine Similarity
StyleGram Matrix [16]MSE
ColorRGB, HSV, LABAbsolute Difference
TextureGLCM [19]Euclidean Distance
ShapeHu Moments [22]Euclidean Distance
SharpnessHigh Frequency Energy (HFE)Absolute Difference
", + "image_path": "db9bb8ae672a9ba00462cdf90948160b29c1459175165bfa04f8663b28c70e30.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 89, + 149, + 261, + 161 + ], + "lines": [ + { + "bbox": [ + 89, + 149, + 261, + 161 + ], + "spans": [ + { + "bbox": [ + 89, + 149, + 261, + 161 + ], + "type": "text", + "content": "Table 1. Feature types and matching functions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 182, + 295, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 253 + ], + "type": "text", + "content": "Appendix A. This finding empowers us to create a large-scale, model-agnostic noise library, where each noise sample can be associated with diverse latent features and then shared across different models. The universality of noise-driven priors positions initial noise as a unified control resource for various applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 261, + 266, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 266, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 266, + 273 + ], + "type": "text", + "content": "3.4. NoiseQuery for Goal-Driven Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 361 + ], + "type": "text", + "content": "Building upon the insights regarding initial noise characteristics and cross-model consistency, we propose NoiseQuery, an efficient approach for selecting appropriate noise samples tailored to diverse generation objectives. As illustrated in Fig. 4, our method consists of two main phases: offline library construction and real-time inference. We also provide pseudo-code in Algorithm 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 361, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 517 + ], + "type": "text", + "content": "Offline Library Construction. In the first stage, we build a noise library " + }, + { + "bbox": [ + 55, + 361, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{\\text {set }}" + }, + { + "bbox": [ + 55, + 361, + 295, + 517 + ], + "type": "text", + "content": ", which contains many initial noise samples (e.g., 100k) sampled from standard Gaussian distribution. These samples are utilized to generate generative posteriors by running a diffusion model without any user-defined prompts. Subsequently, we extract a variety of features " + }, + { + "bbox": [ + 55, + 361, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 55, + 361, + 295, + 517 + ], + "type": "text", + "content": " from these generative posteriors that correspond to various generative objectives " + }, + { + "bbox": [ + 55, + 361, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 361, + 295, + 517 + ], + "type": "text", + "content": ", such as semantic representations (e.g., CLIP features), color properties (e.g., saturation), etc. These extracted features are then stored as keys alongside their corresponding noise samples (values) within the noise library, creating a comprehensive key-value repository that can be queried during the inference phase." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "text", + "content": "Real-time Inference. When performing real-time inference, we specify desired features " + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_O" + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "text", + "content": " following the user's generation objective " + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "text", + "content": ". These features could be semantic, color, spatial, or style-based attributes, depending on the specific goal. Using a match function " + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "inline_equation", + "content": "S(\\cdot)" + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "text", + "content": ", we evaluate the alignment between each feature " + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_i" + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "text", + "content": " in our pre-computed library and the objective features " + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_O" + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "text", + "content": ". The optimal noise sample " + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "inline_equation", + "content": "\\epsilon^*" + }, + { + "bbox": [ + 55, + 517, + 295, + 613 + ], + "type": "text", + "content": " is selected by maximizing the matching scores:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 618, + 295, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 618, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 88, + 618, + 295, + 637 + ], + "type": "interline_equation", + "content": "\\epsilon^ {*} = \\arg \\max _ {\\epsilon_ {i} \\in \\mathcal {N} _ {s e t}} \\left\\{S \\left(\\mathcal {F} _ {i}, \\mathcal {F} _ {O}\\right) \\mid \\epsilon_ {i} \\in \\mathcal {N} _ {s e t} \\right\\}. \\tag {4}", + "image_path": "53b8f5dfb1d1c21384c3b3dbaa2204afd83bf4417ac2d2bb55aa20fdc14e5656.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Once the optimal noise sample " + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon^{*}" + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": " is identified, it is fed into the generative model to produce an image that closely matches the user intent. In Tab. 1, we summarize several common generation goals achievable with NoiseQuery, including both high-level semantics and low-level visual properties. For each goal, we specify the corresponding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "feature types and matching functions. Detailed explanations of these features can be found in Appendix C. Notably, these goals can be applied individually or in combination. Users can specify multiple constraints simultaneously (e.g., generating a \"sunset\" image with a \"warm color palette\" while ensuring \"high sharpness\"). This flexibility makes NoiseQuery highly versatile and generalizable, enabling fine-grained control over the generated output." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 180, + 395, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 395, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 395, + 193 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 201, + 430, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 430, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 430, + 213 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 218, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 553, + 326 + ], + "type": "text", + "content": "Datasets. We evaluate our approach on two popular benchmarks: MSCOCO [33] and DrawBench [47]. MSCOCO is a standard benchmark for text-to-image generation, providing textual descriptions of diverse natural scenes and object classes. We randomly select 10k captions as user prompts. DrawBench consists of 200 carefully designed prompts intended to challenge generative scenarios on various aspects, including long-form text, complex compositions, conflicting instructions, and rare vocabulary." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 327, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 553, + 445 + ], + "type": "text", + "content": "Model Zoo. We conduct experiments on six popular diffusion generation models, including Stable diffusion (SD) series [43] (1.4, 1.5, 2.0, 2.1), PixArt-" + }, + { + "bbox": [ + 313, + 327, + 553, + 445 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 327, + 553, + 445 + ], + "type": "text", + "content": " [6], and SD-Turbo [49]. Stable Diffusion is a Latent Diffusion Model with UNet [44] backbone. By default, we use 50 DDIM denoising steps with a CFG scale of 7.0. PixArt-" + }, + { + "bbox": [ + 313, + 327, + 553, + 445 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 327, + 553, + 445 + ], + "type": "text", + "content": " is a Transformer Latent Diffusion Model, and SD-Turbo is a distillation model based on Stable Diffusion 2.1, generating high-quality images even with only 1-2 sampling steps without CFG. For PixArt-" + }, + { + "bbox": [ + 313, + 327, + 553, + 445 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 327, + 553, + 445 + ], + "type": "text", + "content": ", we use the default CFG scale of 4.5." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 446, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 554, + 555 + ], + "type": "text", + "content": "Evaluation Metrics. For semantic similarity, we use ClipScore [42], which directly measures the alignment between the generated images and the corresponding textual prompts. Additionally, we incorporate several popular metrics trained on large-scale human preference data to quantify human preferences for AI-generated images, including HPS v2 [58], PickScore [27], and ImageReward [60]. These metrics offer a thorough evaluation of image quality, aesthetics, and alignment with user intent." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 563, + 515, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 515, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 515, + 576 + ], + "type": "text", + "content": "4.2. Improvement on High-level Semantics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 581, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 553, + 665 + ], + "type": "text", + "content": "For semantic consistency goals, we utilize features like CLIP [42] and BLIP [29] as the query to identify the semantically matched noise for each text prompt. Such aligned text and noise yield two benefits: (i) prevents the model from generating irrelevant outputs or falling into unsolvable regions. (ii) minimizes generation difficulty and maximizes the likelihood for semantic consistency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 714 + ], + "type": "text", + "content": "Quantitative Analysis. In Tab. 2, we evaluate our method on standard text-to-image (T2I) benchmarks, demonstrating consistent performance gains across all base models by offering better noise-text alignment. NoiseQuery" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17622" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 245 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 245 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 245 + ], + "type": "table", + "html": "
Base ModelMethodDrawBench [47]MSCOCO [33]Time Cost
ImageRewardPickScoreHPS v2CLIPScoreImageRewardPickScoreHPS v2CLIPScore
SD 1.5Base Model0.0421.1124.5730.900.1521.4125.6531.081.334 s
+ NoiseQuery0.0821.1625.0231.410.2721.4826.0731.471.336 s
+ Diffusion-DPO [53]0.0921.2925.0231.190.2521.6426.3131.261.350 s
+ Diffusion-DPO [53] + NoiseQuery0.1721.3325.2531.410.3521.6826.6031.551.352 s
SD 2.1Base Model0.1221.3324.9331.130.3621.7226.5831.401.301 s
+ NoiseQuery0.2621.4625.3931.680.4421.7626.8231.501.303 s
+ CFG++ [10]0.1221.3324.8331.130.3721.7226.6631.313.724 s
+ CFG++ [10] + NoiseQuery0.2721.4325.5531.610.4721.7626.9731.673.726 s
SD-TurboBase Model0.2621.7825.2331.290.4722.0726.2231.510.072 s
+ NoiseQuery0.4121.8725.6631.580.5022.1726.8231.760.074 s
+ ReNO [11]1.6723.4032.4832.55----23.56 s
+ ReNO [11] + NoiseQuery1.7123.5232.9232.78----23.56 s
PixArt-αBase Model0.7022.0828.2730.830.7822.2429.3331.484.327 s
+ NoiseQuery0.8222.1128.4531.270.7922.3329.5631.644.328 s
+ LaVi-Bridge [67]0.6322.0828.3530.920.7522.3129.4931.865.092 s
+ LaVi-Bridge [67] + NoiseQuery0.7222.2428.6131.350.7822.3529.6732.015.094 s
", + "image_path": "b2a6fc8ebbb9800730de9f25559782c0858391fade64feb040d23b9a2e471133.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 253, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 555, + 298 + ], + "type": "text", + "content": "Table 2. Evaluation of objective metrics on different datasets and models. Higher is better for all metrics. Our approach enhances the base model's performance and complements a wide range of T2I enhancement methods, including DPO [53] (reward-based fine-tuning), CFG++ [10] (guidance refinement), ReNO [11] (gradient-based initial noise optimization), and LaVi-Bridge [67] (text encoding enhancement via LORA/adapters). " + }, + { + "bbox": [ + 55, + 253, + 555, + 298 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 253, + 555, + 298 + ], + "type": "text", + "content": ": ReNO results on MSCOCO are excluded due to excessive inference time." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 307, + 294, + 553 + ], + "blocks": [ + { + "bbox": [ + 56, + 307, + 294, + 553 + ], + "lines": [ + { + "bbox": [ + 56, + 307, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 307, + 294, + 553 + ], + "type": "image", + "image_path": "507911a5084e97e7d2c18ebdec3bacd774e3943bfecd99089ef7ffa969a75f86.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 559, + 295, + 616 + ], + "lines": [ + { + "bbox": [ + 55, + 559, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 295, + 616 + ], + "type": "text", + "content": "Figure 5. Generated images of the base model, base model with T2I enhancement methods, and base model with enhancement and NoiseQuery. The four rows, from top to bottom, respectively correspond to the method groups listed in Table 2: Diffusion-DPO [53], CFG++ [10], ReNO [11], and LaVi-Bridge [67]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "also serves as a universal foundation layer, complementing T2I enhancement techniques like Diffusion-DPO [53], CFG++ [10], ReNO [11], and LaVi-Bridge [67]. Notably, these methods face inherent limitations: model-specific dependencies (e.g., Diffusion-DPO and LaVi-Bridge require retraining for new models; ReNO is restricted to one-step models due to heavy computation, excluding com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 326, + 308, + 541, + 411 + ], + "blocks": [ + { + "bbox": [ + 326, + 308, + 541, + 411 + ], + "lines": [ + { + "bbox": [ + 326, + 308, + 541, + 411 + ], + "spans": [ + { + "bbox": [ + 326, + 308, + 541, + 411 + ], + "type": "image", + "image_path": "5f014cb793cc04e6966c74a1b523c31602db45805889a79962dfe3d5bd2f12bd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 413, + 512, + 423 + ], + "lines": [ + { + "bbox": [ + 356, + 413, + 512, + 423 + ], + "spans": [ + { + "bbox": [ + 356, + 413, + 512, + 423 + ], + "type": "text", + "content": "Figure 6. Queried noise v.s. random noise." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 325, + 430, + 541, + 533 + ], + "blocks": [ + { + "bbox": [ + 325, + 430, + 541, + 533 + ], + "lines": [ + { + "bbox": [ + 325, + 430, + 541, + 533 + ], + "spans": [ + { + "bbox": [ + 325, + 430, + 541, + 533 + ], + "type": "image", + "image_path": "9786536d46520ccf67ca7e9fb178aefe4c79983cad0fdf1da93dd27f3d668e24.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 534, + 555, + 556 + ], + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 556 + ], + "type": "text", + "content": "Figure 7. CLIP and FID evaluations across CFG scales on MSCOCO [33], with the CFG scale increasing from left to right." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 687 + ], + "type": "text", + "content": "mon multi-step models like SD) and high latency. In contrast, our method seamlessly integrates with them, enhancing their performance while adding near-zero time costs. We also present qualitative results in Fig. 5 to illustrate its effectiveness in assisting text-to-image synthesis. Additionally, we show the zero-shot transferability of NoiseQuery across various generative models in Appendix B.1. This model-agnostic compatibility, combined with minimal computational cost, establishes NoiseQuery as a versatile and efficient enhancement tool for diverse pipelines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Diversity Analysis. We report the performance of selecting the top 20 matched noise samples per prompt instead of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17623" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 298, + 287 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 298, + 287 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 298, + 287 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 298, + 287 + ], + "type": "image", + "image_path": "3829f9940f18480604f7a8e324a6b834b2ff3bd5ecc49dc4d023673efa04d36c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 289, + 225, + 298 + ], + "lines": [ + { + "bbox": [ + 127, + 289, + 225, + 298 + ], + "spans": [ + { + "bbox": [ + 127, + 289, + 225, + 298 + ], + "type": "text", + "content": "(a) Color properties guidance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 552, + 287 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 552, + 287 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 552, + 287 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 552, + 287 + ], + "type": "image", + "image_path": "d182bdc33a273b2555ab7508296ae040025e02acf5f3f67e8c4ff71962ca9f65.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 289, + 484, + 298 + ], + "lines": [ + { + "bbox": [ + 379, + 289, + 484, + 298 + ], + "spans": [ + { + "bbox": [ + 379, + 289, + 484, + 298 + ], + "type": "text", + "content": "(b) Structural features guidance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 340, + 296, + 383 + ], + "blocks": [ + { + "bbox": [ + 58, + 340, + 296, + 383 + ], + "lines": [ + { + "bbox": [ + 58, + 340, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 58, + 340, + 296, + 383 + ], + "type": "table", + "html": "
Library Size0.5k1k2k5k10k50k100k
Matching Function Cost (×10-4s)1.391.391.391.391.401.401.51
Argmax Selection Cost (×10-4s)0.250.250.250.250.376.1613.25
CLIP Score31.5131.5331.5731.5931.6631.7331.74
", + "image_path": "1e45f2df2667c30c560b88ca7202b6a4dd5602225639d8a07ff488a09ee1070e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 386, + 289, + 396 + ], + "lines": [ + { + "bbox": [ + 61, + 386, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 61, + 386, + 289, + 396 + ], + "type": "text", + "content": "Table 3. Retrieval time breakdown across various library sizes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 422, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 422, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 422, + 295, + 542 + ], + "type": "text", + "content": "the single best match. As shown in Fig. 6, these samples consistently outperform the standard random noise baseline on DrawBench [47], demonstrating that NoiseQuery provides not only optimal but also diverse high-quality noise candidates. To quantify diversity, we use the Diversity Metric (DIV) [57], computed as the standard deviation of DINOV2 [38] embeddings across 20 generated images per prompt (see Fig. 6 legend). Our method achieves comparable diversity to random noise, confirming that NoiseQuery enhances image quality without sacrificing diversity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 544, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 544, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 54, + 544, + 295, + 663 + ], + "type": "text", + "content": "Different Classifier-free Guidance Scale Analysis. We assess our method's performance across various classifier-free guidance (CFG) scales (1.5, 2.0, 3.0, ..., 15.0) on 10k MSCOCO [33] samples. Our NoiseQuery achieves high-quality outputs even with low guidance scales, as demonstrated in Fig. 7. This supports our hypothesis that by selecting more compatible initial noise, we can effectively reduce the generation difficulty. Therefore, we can obtain high-quality output at low scales and avoid the over-saturation and instability that usually come with high scales." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 296, + 713 + ], + "type": "text", + "content": "Time Cost. We also provide a cost breakdown for the retrieval process in Tab. 3, which includes the computation of the matching function (utilizing query CLIP features at no extra cost since CLIP is embedded in SD) and argmax se" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 314, + 331, + 555, + 401 + ], + "blocks": [ + { + "bbox": [ + 55, + 299, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 55, + 299, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 555, + 322 + ], + "type": "text", + "content": "Figure 8. Multiple low-level visual properties control through NoiseQuery. All images use minimal text prompts containing only semantic concepts (e.g., object labels), lacking explicit visual specifications." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 331, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 314, + 331, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 314, + 331, + 555, + 401 + ], + "type": "image", + "image_path": "7f31b3e9b0aa0fa9e7f2001036b111dd3d953ec3909ae0044680d96629130150.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 403, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 437 + ], + "type": "text", + "content": "Figure 9. Text prompts often fail to align generated images with user low-level attribute preferences, while initial noise plays a dominant role in controlling them." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 459, + 555, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 496 + ], + "type": "text", + "content": "lection. The results demonstrate that both steps are efficient, even with large libraries (e.g., 0.0015 seconds for 100,000 entries), indicating that the additional cost is negligible." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 504, + 553, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 553, + 517 + ], + "type": "text", + "content": "4.3. Controllability on Low-Level Visual Properties" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 521, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 629 + ], + "type": "text", + "content": "Text prompts effectively steer high-level semantics in text-to-image (T2I) generation but often fail to provide sufficient control over low-level visual properties that affect visual perception. This limitation stems from two factors: (1) text is inherently inadequate for conveying detailed visual reference input, and (2) text encoders like CLIP [42], trained primarily on high-level semantics, struggle to capture and control fine-grained low-level properties such as color, texture, shape, and sharpness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "In contrast, initial noise plays a dominant role in determining these properties, making it a valuable resource for precise control, as illustrated in Fig. 9. Our method leverages this by enabling fine-grained control of low-level attributes through initial noise, without requiring additional training. The following experiments demonstrate the effectiveness of our approach in achieving these targets, with fur" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17624" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 86, + 294, + 205 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 108, + 85 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 108, + 85 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 108, + 85 + ], + "type": "text", + "content": "Diverse User Demands" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 70, + 161, + 85 + ], + "lines": [ + { + "bbox": [ + 133, + 70, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 133, + 70, + 161, + 85 + ], + "type": "text", + "content": "Stable Diffusion" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 70, + 235, + 86 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 235, + 86 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 235, + 86 + ], + "type": "text", + "content": "Stable Diffusion (enhanced prompt)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 232, + 70, + 296, + 86 + ], + "lines": [ + { + "bbox": [ + 232, + 70, + 296, + 86 + ], + "spans": [ + { + "bbox": [ + 232, + 70, + 296, + 86 + ], + "type": "text", + "content": "NoiseQuery (ours)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 86, + 294, + 205 + ], + "lines": [ + { + "bbox": [ + 57, + 86, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 86, + 294, + 205 + ], + "type": "image", + "image_path": "b7a2d3de948f45cbf79e86d5414c8a7595de0bebb8affc1d444dd8563c462966.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 212, + 295, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 246 + ], + "type": "text", + "content": "Figure 10. Examples of combining high-level semantics (red) and low-level properties (blue). Enhanced prompts are created by fusing semantic objects and visual details via a large language model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 266, + 295, + 392 + ], + "blocks": [ + { + "bbox": [ + 56, + 266, + 295, + 392 + ], + "lines": [ + { + "bbox": [ + 56, + 266, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 295, + 392 + ], + "type": "image", + "image_path": "02bc7ed3fb8653715ab9059a20cbc8579d9479ccf964c508d5655e57e8ffff1a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 397, + 295, + 431 + ], + "lines": [ + { + "bbox": [ + 55, + 397, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 295, + 431 + ], + "type": "text", + "content": "Figure 11. Compatibility with controllable models such as TI [15] and ControlNet [64]. Our method provides a better initial noise, enhancing semantic consistency and low-level attribute control." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 463, + 269, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 269, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 269, + 475 + ], + "type": "text", + "content": "ther implementation details provided in Appendix C." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 480, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 589 + ], + "type": "text", + "content": "Color Properties. Color attributes such as brightness, saturation, contrast, and hue are crucial for an image's visual appeal. We assess the overall color hue by averaging each RGB channel, measure saturation and brightness in the HSV color space, and compute contrast as the standard deviation of grayscale intensities. As shown in Fig. 8a, selecting initial noise that matches target color distributions enables precise control over the generated images' color tendencies, aligning them with user preferences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "Structural Features. For spatial control, we implement: (1) Texture through Gray-Level Co-occurrence Matrix (GLCM) [19] features that quantify intensity co-occurrence patterns; (2) Shape consistency via Hu Moments [22] as rotation-invariant descriptors; and (3) Sharpness by prioritizing high-frequency energy (HFE) in Fourier spectra. Fig. 8b demonstrates how noise selection governs these structural features—preserving geometric fidelity in object shapes, modulating surface textures, and enhancing edge crispness—all achieved without prompt engineering." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 515, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 515, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 515, + 84 + ], + "type": "text", + "content": "4.4. Enhancement on Combined Scenarios" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 91, + 554, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 554, + 163 + ], + "type": "text", + "content": "Building upon the results in previous sections, which demonstrate our method's effectiveness in high-level semantic alignment and low-level attribute control, this section highlights its ability to combine these diverse goals and seamlessly integrate with powerful existing controllable generation models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 166, + 555, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 555, + 285 + ], + "type": "text", + "content": "Combined Goals. Beyond individual tasks, our method excels in simultaneously achieving multiple generation goals. By sequentially reranking noise candidates—first selecting top candidates based on one goal (e.g., high-level semantics) and then refining within this subset for another goal (e.g., low-level attributes)—we progressively narrow down the optimal noise pool. As shown in Fig. 10, this approach allows the combination of high-level semantic alignment with various low-level attribute controls, enabling the generation of images that meet diverse user preferences." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 288, + 556, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 556, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 556, + 407 + ], + "type": "text", + "content": "Compatibility with Controlling Methods. Our method is designed to be instant-time and model-agnostic, making it highly compatible with existing controllable generation models such as Textual Inversion (TI) [15] and ControlNet [64]. By providing a better initial noise, our approach establishes a foundational layer for controllable generation, which synergizes with subsequent refinement processes. As shown in Fig. 11, this integration enhances semantic consistency while enabling precise control over challenging low-level attributes, such as color and texture." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 426, + 388, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 388, + 437 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 388, + 437 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 448, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 639 + ], + "type": "text", + "content": "In this work, we propose NoiseQuery, a tuning-free method that enables versatile goal-driven T2I generation through enhanced noise initialization, supported by a generic and reusable pre-built noise library. By selecting Gaussian initialization in the noise library better aligned with the text prompt or specified attributes, NoiseQuery achieves fine-grained control over both high-level semantic and intricate low-level details. In addition, the model-agnostic nature of our method allows it to be seamlessly integrated with existing T2I augmentation and controllable generation methods without model adjustments or significant computational overhead. Experiments validate its effectiveness in improving output fidelity and precision, particularly for visual attributes that are challenging to articulate textually. The lightweight and scalability of the framework make it adaptable to real-world applications." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Limitations. Although NoiseQuery provides flexibility for aligning high-level semantics and controlling low-level attributes, it remains bounded by the base model's ability, and may still underperform in extremely complex cases. Additionally, it has limitations in achieving fine-grained control, such as Canny edge control." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17625" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 168 + ], + "type": "text", + "content": "This work was partially supported by the National Natural Science Foundation of China under grant 62372341. This work was also partially supported by an Amazon Research Award 2023 to OR and YZ. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not reflect the views of Amazon." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 193, + 115, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 193, + 115, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 115, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 212, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 212, + 294, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 212, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 61, + 212, + 294, + 257 + ], + "type": "text", + "content": "[1] Donghoon Ahn, Jiwon Kang, Sanghyun Lee, Jaewon Min, Minjae Kim, Wooseok Jang, Hyoungwon Cho, Sayak Paul, SeonHwa Kim, Eunju Cha, et al. A noise is worth diffusion guidance. arXiv, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 258, + 296, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 258, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 61, + 258, + 296, + 291 + ], + "type": "text", + "content": "[2] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. In ICLR, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 293, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 293, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 61, + 293, + 296, + 337 + ], + "type": "text", + "content": "[3] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM transactions on Graphics (TOG), 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 339, + 296, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 339, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 339, + 296, + 382 + ], + "type": "text", + "content": "[4] Changgu Chen, Libing Yang, Xiaoyan Yang, Lianggangxu Chen, Gaoqi He, Changbo Wang, and Yang Li. Find: Finetuning initial noise distribution with policy optimization for diffusion models. In ACM MM, 2024. 1, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 384, + 296, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 296, + 438 + ], + "type": "text", + "content": "[5] Haoyang Chen, Dongfang Sun, Caoyuan Ma, Shiqin Wang, Kewei Zhang, Zheng Wang, and Zhixiang Wang. Subjective camera: Bridging human cognition and visual reconstruction through sequence-aware sketch-guided diffusion. arXiv, 2025. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 441, + 296, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 296, + 495 + ], + "type": "text", + "content": "[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, 2024. 2, 4, 5, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 498, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 296, + 531 + ], + "type": "text", + "content": "[7] Sitan Chen, Giannis Daras, and Alex Dimakis. Restoration-degradation beyond linear diffusions: A non-asymptotic analysis for ddim-type samplers. In ICML, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 576 + ], + "type": "text", + "content": "[8] Zijiao Chen, Jiaxin Qing, Tiange Xiang, Wan Lin Yue, and Juan Helen Zhou. Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding. In CVPR, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 578, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 296, + 621 + ], + "type": "text", + "content": "[9] Jiaxin Cheng, Xiao Liang, Xingjian Shi, Tong He, Tianjun Xiao, and Mu Li. Layoutdiffuse: Adapting foundational diffusion models for layout-to-image generation. arXiv, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 666 + ], + "type": "text", + "content": "[10] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[11] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. NeurIPS, 37:125487-125519, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[12] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. NeurIPS, 2023. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[13] Yingying Fan, Quanwei Yang, Kaisiyuan Wang, Hang Zhou, Yingying Li, Haocheng Feng, Errui Ding, Yu Wu, and Jingdong Wang. Re-hold: Video hand object interaction reenactment via adaptive layout-instructed diffusion model. In CVPR, 2025. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 218 + ], + "type": "text", + "content": "[14] Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or. Implicit style-content separation using b-lora. In ECCV, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 264 + ], + "type": "text", + "content": "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023. 1, 2, 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 266, + 553, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 553, + 287 + ], + "type": "text", + "content": "[16] Leon A Gatys. A neural algorithm of artistic style. arXiv, 2015. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 289, + 553, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 289, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 316, + 289, + 553, + 322 + ], + "type": "text", + "content": "[17] Or Greenberg, Eran Kishon, and Dani Lischinski. S2st: Image-to-image translation in the seed space of latent diffusion. arXiv, 2023. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 324, + 553, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 324, + 553, + 367 + ], + "spans": [ + { + "bbox": [ + 316, + 324, + 553, + 367 + ], + "type": "text", + "content": "[18] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024. 1, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 369, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 369, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 369, + 553, + 402 + ], + "type": "text", + "content": "[19] Robert M Haralick, Karthikeyan Shanmugam, and Its' Hak Dinstein. Textural features for image classification. IEEE SMCS, 1973. 5, 8, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 404, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 426 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 1, 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 428, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 553, + 460 + ], + "type": "text", + "content": "[21] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In ICCV, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 462, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 462, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 462, + 553, + 495 + ], + "type": "text", + "content": "[22] Ming-Kuei Hu. Visual pattern recognition by moment invariants. IRE transactions on information theory, 1962. 5, 8, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 497, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 497, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 553, + 529 + ], + "type": "text", + "content": "[23] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 531, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 553, + 564 + ], + "type": "text", + "content": "[24] Huayang Huang, Yu Wu, and Qian Wang. Robin: Robust and invisible watermarks for diffusion models with adversarial optimization. NIPS, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 566, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 553, + 599 + ], + "type": "text", + "content": "[25] Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin CK Chan, and Ziwei Liu. Reversion: Diffusion-based relation inversion from images. In SIGGRAPH Asia, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "text", + "content": "[26] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[27] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. NeurIPS, 2023. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[28] Gen Li, Yuting Wei, Yuxin Chen, and Yuejie Chi. Towards non-asymptotic convergence for diffusion-based generative models. In ICLR, 2024. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17626" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[29] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML. PMLR, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[30] Shufan Li, Konstantinos Kallidromitis, Akash Gokul, Yusuke Kato, and Kazuki Kozuka. Aligning diffusion models by optimizing human utility. NeurIPS, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[31] Yumeng Li, Margret Keuper, Dan Zhang, and Anna Khoreva. Divide & bind your attention for improved generative semantic nursing. arXiv, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "text", + "content": "[32] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In WACV, 2024. 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 5, 6, 7, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[34] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In ICLR, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "text", + "content": "[35] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. NeurIPS, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 386 + ], + "type": "text", + "content": "[36] Nanye Ma, Shangyuan Tong, Haolin Jia, Hexiang Hu, Yu-Chuan Su, Mingda Zhang, Xuan Yang, Yandong Li, Tommi Jaakkola, Xuhui Jia, et al. Inference-time scaling for diffusion models beyond scaling denoising steps. arXiv, 2025. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "text", + "content": "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "text", + "content": "[38] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Trans. Mach. Learn. Res., 2024. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 478, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 499 + ], + "type": "text", + "content": "[39] William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "text", + "content": "[40] Yuhan Pei, Ruoyu Wang, Yongqi Yang, Ye Zhu, Olga Russakovsky, and Yu Wu. Sowing information: Cultivating contextual coherence with mllms in image generation. arXiv, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "text", + "content": "[41] Zipeng Qi, Lichen Bai, Haoyi Xiong, et al. Not all noises are created equally: Diffusion noise selection and optimization. arXiv, 2024. 1, 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML. PMLR, 2021. 5, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 677 + ], + "type": "text", + "content": "[43] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 4, 5, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[44] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI. Springer, 2015. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[45] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 171 + ], + "type": "text", + "content": "[46] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In CVPR, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 173, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 173, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 173, + 553, + 228 + ], + "type": "text", + "content": "[47] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 5, 6, 7, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 230, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 230, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 315, + 230, + 553, + 262 + ], + "type": "text", + "content": "[48] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In AAAI, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 263, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 263, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 315, + 263, + 553, + 296 + ], + "type": "text", + "content": "[49] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In ECCV, 2024. 2, 5, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 297, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 297, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 315, + 297, + 553, + 319 + ], + "type": "text", + "content": "[50] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 320, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 320, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 315, + 320, + 553, + 364 + ], + "type": "text", + "content": "[51] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 315, + 365, + 553, + 408 + ], + "type": "text", + "content": "[52] Zhiwei Tang, Jiangweizhi Peng, Jiasheng Tang, Mingyi Hong, Fan Wang, and Tsung-Hui Chang. Inference-time alignment of diffusion models with direct noise optimization. arXiv, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 315, + 411, + 553, + 464 + ], + "type": "text", + "content": "[53] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, pages 8228-8238, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 466, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 466, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 315, + 466, + 553, + 499 + ], + "type": "text", + "content": "[54] Haofan Wang, Matteo Spinelli, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 501, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 501, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 315, + 501, + 553, + 542 + ], + "type": "text", + "content": "[55] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 315, + 545, + 553, + 578 + ], + "type": "text", + "content": "[56] Ruoyu Wang, Yongqi Yang, Zhihao Qian, Ye Zhu, and Yu Wu. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 315, + 579, + 553, + 623 + ], + "type": "text", + "content": "[57] Xi Wang, Nicolas Dufour, Nefeli Andreou, Marie-Paule Cani, Victoria Fernández Abrevaya, David Picard, and Vicky Kalogeiton. Analysis of classifier-free guidance weight schedulers. arXiv, 2024. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 315, + 624, + 553, + 668 + ], + "type": "text", + "content": "[58] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv, 2023. 3, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 670, + 553, + 712 + ], + "type": "text", + "content": "[59] Guangxuan Xiao, Tianwei Yin, William T Freeman, Frédo Durand, and Song Han. Fastcomposer: Tuning-free multisubject image generation with localized attention. *IJCV*, 2024. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17627" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 453 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[60] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. NeurIPS, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[61] Katherine Xu, Lingzhi Zhang, and Jianbo Shi. Good seed makes a good crop: Discovering secret seeds in text-to-image diffusion models. arXiv, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[62] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "text", + "content": "[63] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 262 + ], + "type": "text", + "content": "[64] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 286 + ], + "type": "text", + "content": "[65] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Ic-light github page, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 288, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 330 + ], + "type": "text", + "content": "[66] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2024. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[67] Shihao Zhao, Shaozhe Hao, Bojia Zi, Huaizhe Xu, and Kwan-Yee K Wong. Bridging different language models and generative vision models for text-to-image generation. In ECCV, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 410 + ], + "type": "text", + "content": "[68] Wenliang Zhao, Lujia Bai, Yongming Rao, Jie Zhou, and Jiwen Lu. Unipc: A unified predictor-corrector framework for fast sampling of diffusion models. NeurIPS, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[69] Chenyi Zhuang, Ying Hu, and Pan Gao. Magnet: We never know how text-to-image diffusion models work, until we learn how vision-language models function. NeurIPS, 2024. 3" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17628" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/cab6b62f-c3b3-45e2-8877-ae8b2379307d_content_list.json b/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/cab6b62f-c3b3-45e2-8877-ae8b2379307d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ce1996cf17c92cd10686c3f8503c8ed387c2b46c --- /dev/null +++ b/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/cab6b62f-c3b3-45e2-8877-ae8b2379307d_content_list.json @@ -0,0 +1,1858 @@ +[ + { + "type": "text", + "text": "The Source Image is the Best Attention for Infrared and Visible Image Fusion", + "text_level": 1, + "bbox": [ + 107, + 130, + 888, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Song Wang Xie Han Liquun Kuang\\* Boying Wang Zhongyu Chen Zherui Qiao Fan Yang Xiaoxia Liu Bingyu Zhang Zhixun Wang", + "bbox": [ + 204, + 180, + 790, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of Computer Science and Technology, North University of China, Taiyuan 030051, China", + "bbox": [ + 209, + 219, + 785, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shanxi Key Laboratory of Machine Vision & Virtual Reality, Taiyuan 030051, China", + "bbox": [ + 245, + 234, + 750, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shanxi Vision Information Processing and Intelligent Robot Engineering Research Center, Taiyuan 030051, China", + "bbox": [ + 158, + 253, + 834, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "b20230704@st.nuc.edu.cn,kuang@nuc.edu.cn", + "bbox": [ + 313, + 272, + 679, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Infrared and visible image fusion (IVF) endeavors to engineer composite outputs by blending optimal virtues of divergent modalities. This paper reveals, unprecedentedly, the intrinsic \"attention properties\" of infrared images, which directly arise from their physical characteristics (i.e., heat distribution) and can be linked to attention mechanisms naturally, as observed in the gradient-weighted class activation mapping (Grad-CAM) visualization analysis of image classification models. To incorporate this property into IVF for better fusion, we propose the source infrared cross attention (I-SCA) and further extend it to the visible modality, subsequently introducing the source visible cross attention (V-SCA). The joint use of I-SCA and V-SCA greatly alleviate longstanding issues, such as insufficient and incomplete multimodal feature interaction and fusion, in IVF. Moreover, an auxiliary component for I-SCA and V-SCA, termed CBSM, is employed to boost the channel, map space, and suppress redundancy and misleading information of the source images. Specifically, we directly treat the CBSM-processed raw image as the query, while the intermediate features of another modality are treated as keys and values in I-SCA and V-SCA. Unlike attention mechanisms that divide images into patches or limit computations to local windows, our cross attention modules achieve smoother and more robust IVF through true global modeling across the entire image space with linear complexity. Comparison with current SOTA methods on three popular public datasets confirms its superiority.", + "bbox": [ + 89, + 353, + 483, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 806, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "IVF, a bedrock of image processing, with the objective of engendering a composite that encapsulates the considerable thermal radiation from infrared and the intricacies de", + "bbox": [ + 89, + 832, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cf6a6a3a69c705248c33627af0920a354fa0205b3e693fab823f005c4e66d760.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 553, + 319, + 704, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/42d9e79675dcb052c7be455df5e9004fad7d2174bd45e51c9827735cba04ff01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 319, + 864, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/60f756b556ad51bd892a8590a1c3e0e75c420a167306e77564c381641e919d87.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 553, + 428, + 702, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/94df5b43c0a9fbd1255f09420f8b5c323a4452d5b3fbecebe99263e0fc68ff88.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 715, + 426, + 864, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b7e700277990f92bce909bb11eedcfae37e9a0ef4d900f865c63a74fcef4cfba.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 553, + 525, + 702, + 609 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/41fffc37ad8b5d59409848e620541e1864e5188e612d5bdf874955d285424cc9.jpg", + "image_caption": [ + "(d)", + "(f)", + "Figure 1. (a, b) Grad-CAM visualization results for two images (00022004, 00028021), both labeled as \"dog\", predicted by the final layer of ResNet-50 on the ImageNet validation set. The color gradient from red to blue indicates the decreasing level of attention paid by the model to different regions. (c) An infrared image example (00519D) consists of highlighted thermal radiation regions and remaining ordinary regions from the MSRS dataset. (d) Visible image of the same scene corresponding to subplot (c). (e) Subplot (c) is displayed by applying the Jet colormap. (f) The negative transformation of subplot (c) is visualized with the Jet colormap." + ], + "image_footnote": [], + "bbox": [ + 715, + 523, + 864, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tail from visible. Prior to the surge in deep learning, this process was typically achieved by traditional algorithms in spatial and transform domains [11]. However, thanks to the development of neural networks and attention mechanisms, heuristic algorithms have become more widely adopted.", + "bbox": [ + 511, + 806, + 906, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we draw inspiration from the Grad-CAM", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13513", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[22] visualization results of classic image classification models, such as ResNet-50 [9], and discover the inherent \"attention properties\" of infrared images. These Grad-CAM visualizations, as shown in Fig. 1(a) and Fig. 1(b), reveal how the model assigns varying weights to different regions during the prediction process: regions marked in red represent \"important areas\" that exert a substantial impact on the classification result, while regions marked in blue correspond to \"general areas\" with lesser contribution. Interestingly, we observe that the structural characteristics of infrared images (e.g., as shown in Fig. 1(c)) are strikingly similar to the Grad-CAM images, as both consist of important areas and general areas (in infrared images, these regions correspond to the highlighted areas with higher thermal energy and the darker zones with lower energy). To more intuitively showcase this similarity, we re-map the infrared image with the Jet colormap (see Fig. 1(e)).", + "bbox": [ + 89, + 90, + 480, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these observations, we believe that infrared images inherently possess an attention property similar to the one exhibited by Grad-CAM. Unlike image classification models, which require be trained to learn attention patterns, the attention properties of infrared images directly stem from their physical characteristics—the heat distribution. Moreover, this attribute is not only more specific and accurate but also avoids issues such as insufficient or misdirected attention that may occur in image classification models. For instance, when ResNet-50 processes the dog in Fig. 1(a), it focuses only on the dog's eyes, and when processing the dog in Fig. 1(b), it not only focuses on the dog itself but also incorrectly attends to irrelevant backgrounds. In contrast, infrared images provide precise localization of heat sources, ensuring that all attention is concentrated on specific areas.", + "bbox": [ + 89, + 351, + 482, + 592 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the above finding, an immediate question arises: how can we fully leverage the inherent attention properties of infrared images to obtain better-fused images in IVF? Intuitively, we think that it can be linked to the attention mechanism [26], as both essentially represent the importance of different regions. Furthermore, the interaction of cross-modal features plays a pivotal role in image fusion. Therefore, our goal is to introduce the attention properties at this stage to counteract the persistent obstacles in image fusion, such as inadequate cross-domain information exchange and incomplete multimodal feature fusion, thereby enhancing the caliber of the merged images.", + "bbox": [ + 89, + 595, + 482, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we propose the source infrared cross attention (I-SCA) based on a modified Restormer [36], which can perform global feature modeling in the image space with linear complexity. I-SCA first takes the original infrared image and the intermediate features of the visible image as inputs. It then treats the infrared image as the query (Q), thanks to the properties of infrared images, which allow them to be directly used as queries without processing,", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "while the visible features serve as the keys (K) and values (V), enabling an infrared heat prior-guided cross-domain information exchange. However, this approach can only prioritize on \"important areas\", whereas image fusion is a global pixel-wise task that demands equal treatment for all areas. Fortunately, a facile solution is to treat the negative transformation of the infrared image (as shown in Fig. 1(f)) as another query to allow significant attention complementarity.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, using I-SCA only also faces a critical problem, i.e., it can only capture \"infrared-guided\" cross-modal interaction features. Theoretically, the optimal fused image could be obtained by simultaneously fusing features dominated by both infrared and visible images. Therefore, we make a bold inference: we extend the discovery to visible and, based on this, propose the source visible cross attention (V-SCA). While visible images lack explicit physical saliency, their visually significant regions (e.g., edges, textures, and contrast) can still serve as effective priors.", + "bbox": [ + 511, + 232, + 903, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, since both I-SCA and V-SCA require inputs to have the same channels, but the channels of the source image and the intermediate representation from the other modality are not consistent, we deploy an adaptive channel boosting and intelligent space mapping module (CBSM), which also serves crucial roles in transforming the feature space of the source image from its initial state to one more suitable for attention mechanisms to reduce discrepancies between features at different stages and suppressing redundancy or misleading content contained in the initial images. To validate the effectiveness of I-SCA, V-SCA, and CBSM, extensive analysis are conducted on three popular public datasets. Experiments demonstrate that both qualitative and quantitative results outperform state-of-the-art (SOTA) methods.", + "bbox": [ + 511, + 388, + 903, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The essence of our contributions is as follows:", + "bbox": [ + 532, + 621, + 839, + 635 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We observe a strong structural correlation between infrared images and Grad-CAM, leading us to hypothesize, for the first time, that infrared images possess an inherently exceptionally precise attention property.", + "- Based on this attention property, we propose a heat-distribution-prior-guided cross-modal attention mechanism, which can be further extend to the visible modality. Additionally, an auxiliary module, CBSM, is deployed to complement the proposed attention mechanism for further enhancing fusion quality.", + "- Our findings inspire a novel paradigm guided by source image priors for image fusion, significantly alleviating challenges in cross-domain feature interaction and fusion faced by current methods.", + "- Comprehensive experiments instituted on three prominent available datasets with SOTA works demonstrate the superiority." + ], + "bbox": [ + 532, + 641, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13514", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 91, + 89, + 240, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Attention-based Fusion Methods", + "text_level": 1, + "bbox": [ + 89, + 114, + 377, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The triumph of transformer [26] in NLP has sparked considerable interest across the academic community. To establish a foundation for the integration of this technology in computer vision, Alexey et al. proposed Vision Transformer (ViT) [6], which segments an image into multiple units, then flattens the units into a shape compatible with text sequences, allowing direct application of the original transformer for long-range modeling. Subsequently, works such as TUFusion [41], DATfuse [24], HitFusion [4], AFT-Fusion [3], etc., extended ViT to the image fusion task. On the one hand, TUFusion adopted an encoder-decoder pipeline, featuring a dual-branch structure that merged parallel DenseBlock [12] and ViT in the encoder. DATFuse, on the other hand, incorporated channel and spatial attention mechanisms similar to CBAM [31] before the self-attention mechanism and performed fusion at the image level. Although transformer-based methods have shown great effectiveness, they are consistently constrained by the computational burden of the self-attention in image space. To remedy this issue, Liu et al. proposed SwinTransformer [18], which divides images into patches and further splits the patches into multiple windows, restricting self-attention to each window. Later, SwinFuse [29] used a full SwinTransformer approach to obtain fused images. SwinFusion [20] further extended this by introducing a cross-modal SwinTransformer, which generalizes window attention to cross-modal feature fusion. However, the practice of restricting attention to local regions contradicts the fundamental idea of attention mechanisms, which is to model global information. To achieve true overall modeling linearly, Zamir et al. proposed Restormer, which introduces multi-dconv head transposed attention (MDTA) and gated dconv feedforward network (GDFN) to replace the self-attention and feed-forward network [36]. Later, CDDFuse [42] further validated the effectiveness of the Restormer in image fusion. However, it only applied the Restormer to a single modality and overlooked the potential for feature interactions across multiple modalities. This paper narrowed down the challenge of interaction between different modalities by proposing I-SCA and V-SCA, and further enabled Restormer to be used for the multimodal information interaction problem.", + "bbox": [ + 89, + 138, + 485, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. CNN-based Fusion Methods", + "text_level": 1, + "bbox": [ + 89, + 786, + 344, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Currently, the realm of deep learning-driven fusion is segmented into various types, encompassing techniques reliant on autoencoders, end-to-end models, unified models, amalgamations with downstream or other specialized tasks, and approaches utilizing generative adversarial networks (GAN). Autoencoder-based methods [14, 32] first", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "train an encoder-decoder pair on datasets like COCO [15] and then directly generalize them for image fusion. However, the performance of such methods is markedly hampered by the quality of the fusion rules designed subjectively, and developing a fusion strategy that works well with the encoder-decoder framework is not trivial. Consequently, a series of end-to-end methods [8, 13] were later proposed. These methods train all modules, including the feature fusion layer, with learning strategies, which greatly reduces the challenges related to constructing fusion strategies in autoencoder-based methods. Furthermore, as image fusion has evolved, the community has become dissatisfied with using a single model to solve just one specific fusion task, leading to unified models [5, 17, 45] for multiple tasks. However, these methods adopt a re-training strategy for each new task, rather than using a single weight parameter to handle multiple tasks simultaneously. Meanwhile, along with the development of downstream tasks such as object detection [2, 27, 28], scene reconstruction [39], etc., a major trend has been to combine image fusion with these computer vision sub-tasks, facilitating the mutual promotion of both [37, 40]. Another category of methods [34, 44] leverages large language models such as GPT [1] and CLIP [21] to guide the fusion process, achieving cross-domain integration for different tasks. Finally, GAN-based methods [30, 38] transform the fusion into an adversarial problem. Through continuous antagonistic training, the generator's output progressively converges toward the target distribution. Nevertheless, these techniques face hurdles like mode collapse, erratic training, and a lack of clear understanding. Later works proposed denoising diffusion model-based methods [35, 43] to address these issues. Although the core innovation of this study lies in source images prior-guided cross-modal attention, approximately half of the components are still based on the convolutional neural network (CNN). To ensure a comprehensive evaluation, we undertake a comparative analysis of it with several recent CNN-based SOTA methods discussed above.", + "bbox": [ + 511, + 90, + 906, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 683, + 604, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 511, + 710, + 709, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, a hybrid of CNN and attention mechanisms is employed to achieve the IVF. The overall architecture is shown in Fig. 2. Specifically, we construct a serial pipeline, consisting of SE-ResNet Block [10], Restormer Block [36], and I-SCA/V-SCA for infrared and visible branches. For each modality, we first employ the SE-ResNet Block (the SE acts as a gating mechanism to suppress information redundancy by refine features) for initial feature extraction while increasing the feature channel from ch_in to ch_mid. The Restormer Block then performs further deep extraction with another key function of transforming the feature space", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13515", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/64735d6cc1ae6d0f6815b6935e058903fd7407df15ddb76dc1ca766692751859.jpg", + "image_caption": [ + "Figure 2. The overall architecture of our method, which incorporates SE-ResNet Block, Restormer Block, I-SCA/V-SCA, and CBSM. ir, vi denote infrared and visible images, and $ir\\_n$ , $vi\\_n$ signify negative transformations of infrared and visible images. ch_in, ch_mid, and ch_out represent input, middle, and output channels, respectively. “C” indicates a channel-wise concatenation operation. “Q” stands for the query in the attention mechanism, while “K” and “V” symbolize the key and value, respectively." + ], + "image_footnote": [], + "bbox": [ + 106, + 95, + 911, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from the domain where convolutional operations excel to the one where attention is more effective, thereby facilitating the subsequent cross-modal feature interaction. The I-SCA/V-SCA subsequently takes the features of the current modality along with either the image of another modality processed by CBSM or the negative transformation of that image for feature interaction. Post processing via I-SCA/V-SCA, a quartet of feature maps, steered by the originating image, is acquired. Then they will merge across the channel dimension and be processed through an additional SE-ResNet Block before being output.", + "bbox": [ + 88, + 441, + 480, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. CBSM", + "text_level": 1, + "bbox": [ + 89, + 618, + 181, + 633 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6d87b58cc64b314abb4deff6f780c7e4b6301f6c4e3af48ebc832590aeca096f.jpg", + "image_caption": [ + "Figure 3. The composition of adaptive channel boosting and intelligent space mapping module (CBSM). The SE is employed to suppress redundancy and misleading information in the source images. $\\odot$ denotes a channel-wise multiplication operation." + ], + "image_footnote": [], + "bbox": [ + 106, + 652, + 473, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since attention calculation in I-SCA and V-SCA requires all channels of inputs to be consistent, it is imperative to augment the channel dimension of the source image", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from ch_in to ch_mid. In general, this can be achieved by Parameters, $Conv_{1\\times 1}$ , or $Conv_{3\\times 3}$ . However, all three have a common problem: after channel expansion, the multi-channel feature space still remains in the initial state. In cross-modal attention, directly using feature values from the raw space for interaction with deep features will significantly degrade the information exchange effect due to spatial discrepancy. Therefore, a key issue is how to reduce the gap between them. Additionally, the source image may introduce redundancy or misleading cues in certain cases, making it crucial to mitigate potential negative effects. To address this, CMSM module, which consists of convolutional layers (for channel boosting), activation layers (for spatial mapping), and Squeeze-and-Excitation (SE) [10] layer (for suppressing negative influences) is employed, as shown in Fig. 3. The structure of CMSM is intentionally simple, because it aligns with our goal of using the raw image priors to guide feature interactions. A more complex design could significantly alter the initial pattern of the source image, conflicting with our motivation. For the input source image, denoted as $I_m^{raw} \\in \\mathbb{R}^{H \\times W \\times 1}$ , this module performs the following operations:", + "bbox": [ + 511, + 441, + 906, + 777 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {m} ^ {t r s} = \\alpha \\left(S E \\left(C o n v _ {3 \\times 3} \\left(\\alpha \\left(C o n v _ {3 \\times 3} \\left(I _ {m} ^ {r a w}\\right)\\right)\\right)\\right)\\right). (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 782, + 906, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The result of Eq. (1), denoted as $I_{m}^{trs} \\in \\mathbb{R}^{H \\times W \\times 48}$ , represents the image after channel boosting and space transformation. The variable $m$ represents different input modalities, and desirable values include $ir$ , $vi$ , $ir\\_n$ , and $vi\\_n$ . $ir\\_n$ and $vi\\_n$ denote the negative transforms of the infrared and the visible image, respectively. $\\alpha(\\cdot)$ is the parametric", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13516", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rectified linear unit (PReLU).", + "bbox": [ + 89, + 90, + 287, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Cross Attention Based on Source Images", + "text_level": 1, + "bbox": [ + 89, + 119, + 437, + 137 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c150dac778b77550090f62882d409b099f4d9be16fc252f26c34f8425f17b5fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 160, + 450, + 297 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/979bb7d27141db02e313efd5ad12869d49efee98302463e8265147179aca27e1.jpg", + "image_caption": [ + "Figure 4. Our proposed I-SCA and V-SCA, both of which consist of two modules: I-DCA/V-DCA and GDFN. Structurally, the two are identical, differing only in their input content." + ], + "image_footnote": [], + "bbox": [ + 112, + 311, + 447, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The introduction of I-SCA and V-SCA aims to reduce the difficulty of multimodal feature interactions. The structures of these two components are illustrated in Fig. 4. In fact, both share an identical structural composition, differing only in the content of their inputs. Although the symmetrical design is used between I-SCA and V-SCA, this does not imply an inability to reflect modal heterogeneity. On the contrary, I-SCA emphasizes the heat distribution, contours, and structural priors of infrared images, while V-SCA focuses on edges, textures, and high-contrast guidance. Specifically, I-SCA comprises the dual-dconv infrared cross attention (I-DCA) and the gated dconv feedforward network (GDFN). Its inputs include $I_{ir}^{trs}$ (or $I_{ir,n}^{trs}$ ) and the visible features $feat_{vi}^{mid}$ , which are processed by the Restormer Block in the visible branch. In contrast, V-SCA is composed of the dual-dconv visible cross attention (V-DCA) and the GDFN, with inputs $I_{vi}^{trs}$ (or $I_{vi,n}^{trs}$ ) and the infrared features $feat_{ir}^{mid}$ .", + "bbox": [ + 89, + 484, + 482, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To illustrate the working mechanism of I-SCA and V-SCA, we demonstrate I-SCA, with inputs $I_{ir}^{trs}$ and $feat_{vi}^{mid}$ , as an example. In I-SCA, both $I_{ir}^{trs}$ and $feat_{vi}^{mid}$ first undergo the layer normalization (LN) in I-DCA. Following this, $I_{ir}^{trs}$ is directly used as the query (Q) after LN, while $feat_{vi}^{mid}$ is further processed to generate the key (K) and value (V). The process is described as follows:", + "bbox": [ + 89, + 758, + 482, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK, V = D \\operatorname {c o n v} \\left(C o n v _ {1 \\times 1} \\left(L N \\left(f e a t _ {v i} ^ {m i d}\\right)\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 877, + 482, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Dconv(\\cdot)$ denotes the depthwise separable convolution, consisting of two steps: depthwise convolution and pointwise convolution. This operation preserves spatial dimensions and independently processes the channels at each spatial position. Subsequently, the Q $(I_{ir}^{trs})$ , K, and V are reshaped for compatibility and undergo the following operation:", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f e a t _ {i r} ^ {o r i} = C o n v _ {1 \\times 1} \\left(V \\cdot S o f t m a x \\left(\\left(I _ {i r} ^ {t r s} \\cdot K\\right) \\odot s\\right)\\right) \\\\ + f e a t _ {v i} ^ {m i d}, (3) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 210, + 903, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "resulting in infrared-oriented intermediate features $\\text{feat}_{ir}^{\\text{ori}}$ , where $\\odot$ represents element-wise multiplication, and $s$ is a learnable parameter. These features are then processed through the GDFN module as follows:", + "bbox": [ + 511, + 261, + 906, + 323 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY _ {1}, Y _ {2} = D \\operatorname {c o n v} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(L N \\left(f e a t _ {i r} ^ {\\text {o r i}}\\right)\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 333, + 906, + 352 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf e a t _ {i r} ^ {o r i} ^ {\\prime} = C o n v _ {1 \\times 1} \\left(G E L U \\left(Y _ {1}\\right) \\odot Y _ {2}\\right) + f e a t _ {i r} ^ {o r i}, (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 383, + 905, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where GELU refers to the Gaussian Error Linear Unit.", + "bbox": [ + 511, + 416, + 872, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the above process, four source images-oriented features, which are further refined by GDFN, denoted as $\\text{feat}_m^{\\text{ori}'}$ , are obtained. These features, corresponding to different modalities ( $m = ir$ , $vi$ , $ir\\_n$ , and $vi\\_n$ ), are arranged along the axis of the channel and further refined by another SE-ResNet Block. The transformation of selfattention from the spatial dimension to the channel dimension, implemented by $Dconv(\\cdot)$ and $Conv_{1\\times 1}(\\cdot)$ , ensures the linear computational characteristic of I-SCA and V-SCA. Additionally, the gating mechanism within the GDFN effectively suppresses redundant features, allowing only beneficial features to propagate through the module.", + "bbox": [ + 511, + 431, + 906, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Function", + "text_level": 1, + "bbox": [ + 511, + 626, + 660, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach utilizes intensity $(L_{int})$ and gradient $(L_{grad})$ loss techniques to guarantee comprehensive and detailed data for the combined image. The complete loss is characterized as:", + "bbox": [ + 511, + 648, + 905, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t o t a l}} = \\alpha_ {1} L _ {\\text {i n t}} + L _ {\\text {g r a d}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 720, + 903, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{1}$ represents the intensity loss factor. The specific definition of intensity loss is as follows:", + "bbox": [ + 511, + 746, + 903, + 776 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i n t} = \\frac {1}{H W} \\| \\max \\left(I _ {i r} ^ {r a w}, I _ {v i} ^ {r a w}\\right) - f e a t _ {o u t} \\| _ {1}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 786, + 903, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\max(\\cdot)$ represents the element-wise maximum operation, $\\text{feat}_{\\text{out}}$ signifies the final output of our model, and $\\| \\cdot \\|_1$ denotes the $l_1$ -norm. Inspired by [42] and [7], the gradient term combines Sobel for edge preservation and Laplacian for noise suppression, guiding the model to capture", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13517", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "valuable information while avoiding distractions. The formula is defined as follows:", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {g r a d} = \\alpha_ {2} L _ {S} + \\alpha_ {3} L _ {L}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 133, + 482, + 150 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {S} = \\frac {1}{H W} \\| \\max \\left(\\left| \\nabla_ {S} I _ {i r} ^ {\\text {r a w}} \\right|, \\left| \\nabla_ {S} I _ {v i} ^ {\\text {r a w}} \\right|\\right) - \\left| \\nabla_ {S} f e a t _ {o u t} \\right| \\| _ {1}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 181, + 495, + 223 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {L} = \\| O \\left(\\max \\left(| \\nabla_ {L} I _ {i r} ^ {\\text {r a w}} |, | \\nabla_ {L} I _ {v i} ^ {\\text {r a w}} |\\right)\\right) - \\nabla_ {L} f e a t _ {o u t} \\| _ {1}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 250, + 482, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha_{2}$ and $\\alpha_{3}$ are the coefficients of the Sobel gradient loss and the Laplace gradient loss, respectively. $|\\cdot |$ refers to the absolute operation. $\\nabla_S$ and $\\nabla_L$ respectively refer to the Sobel and Laplacian gradient operator. $O(|x|) = x$ represents the initial gradient value retrieved before the absolute value process is executed.", + "bbox": [ + 89, + 281, + 483, + 371 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 385, + 223, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 409, + 173, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. During the training stage, the MSRS [23] training set is utilized. For additional enhancement of the training set, 200 pairs are randomly chosen from the RoadScene [33] dataset as a supplement. During the evaluation stage, we use the MSRS test set. Furthermore, to assess the generalization ability of our method, we conduct tests on the $\\mathrm{M}^3\\mathrm{FD}$ [16] and TNO [25] datasets no any re-training or fine-tuning. It's important to recognize that due to the high resolution of images in the $\\mathrm{M}^3\\mathrm{FD}$ dataset, some contrast methods have extremely prolonged runtimes. To expedite testing, the images are resized to half their initial resolution, and the precision for all methods on $\\mathrm{M}^3\\mathrm{FD}$ is reported based on the resized resolution.", + "bbox": [ + 89, + 431, + 483, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implement Details. The hardware setup includes an NVIDIA TITAN RTX and a 3.6 GHz Intel Core i9-9900K, with the code implemented in PyTorch. The Adam optimizer, without weight decay, are employed to optimize the network parameters. The training phase consists of 60 epochs, a batch size of 4, an initial learning rate of $10^{-4}$ , and the learning rate is halved every 25 epochs. In the training phase, input images undergo random cropping into 128 $\\times$ 128-sized segments and are transformed from RGB format to Y, Cb, and Cr. The coefficients for the terms in loss are configured as: $\\alpha_{1} = 0.1$ , $\\alpha_{2} = 1$ , and $\\alpha_{3} = 10$ , which are based on the experience from our previous work. The channels at various stages is respectively defined as $ch\\_in = 1$ , $ch\\_mid = 48$ , and $ch\\_out = 1$ .", + "bbox": [ + 89, + 628, + 483, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 89, + 848, + 382, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Competing Methods. To validate the effectiveness of our method, we compare it with seven state-of-the-art works:", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/40893697326cd8b3a50e328d502c922ce3372b43fca724b25cc65eb396795885.jpg", + "table_caption": [ + "Table 1. Quantitative results of our method compared to seven state-of-the-art methods on the MSRS, $\\mathbf{M}^3\\mathrm{FD}$ , and TNO datasets. Bold indicates the best result, and underlining represents the second best." + ], + "table_footnote": [], + "table_body": "
MSRS
PublisherVIF ↑SCD ↑MI ↑Qabf ↑SSIM ↑MS_SSIM ↑FMI ↑
LRRNet [13]23'TPAMI0.5440.7912.9450.4570.8590.8480.919
CoCoNet [17]24'IJCV0.6601.3872.2190.2790.7330.7470.884
MetaFusion [40]23'CVPR0.7261.4861.6840.4780.9120.9040.908
TUFusion [41]24'TCSVT0.5691.2982.5750.2290.8850.8980.909
DATFuse [24]23'TCSVT0.9061.4103.8970.6400.9550.9560.912
SwinFusion [20]22'JAS0.9901.6904.5490.6540.9680.9690.931
CDDFuse [42]23'CVPR1.0511.6215.0150.6930.9670.9660.930
Ours25'ICCV1.0611.7005.1110.7150.9810.9750.933
M²FD
PublisherVIF ↑SCD ↑MI ↑Qabf ↑SSIM ↑MS_SSIM ↑FMI ↑
LRRNet [13]23'TPAMI0.4931.4602.8750.4660.8910.8700.862
CoCoNet [17]24'IJCV0.6241.7692.8130.3990.8750.8540.853
MetaFusion [40]23'CVPR0.7011.6652.5700.3870.8540.8610.850
TUFusion [41]24'TCSVT0.4001.5022.9650.1890.7880.8640.856
DATFuse [24]23'TCSVT0.5901.2804.0380.4890.8550.8310.849
SwinFusion [20]22'JAS0.6901.5744.0970.5830.9470.9030.872
CDDFuse [42]23'CVPR0.7201.6473.9990.5960.9510.9130.871
Ours25'ICCV0.7591.6603.7710.6540.9640.9230.883
TNO
PublisherVIF ↑SCD ↑MI ↑Qabf ↑SSIM ↑MS_SSIM ↑FMI ↑
LRRNet [13]23'TPAMI0.5821.5942.5930.3630.8700.8700.886
CoCoNet [17]24'IJCV0.6551.7172.1610.2970.8070.8180.871
MetaFusion [40]23'CVPR0.6611.7301.6520.2980.7880.8150.877
TUFusion [41]24'TCSVT0.4651.5992.0940.2240.8250.8670.893
DATFuse [24]23'TCSVT0.6901.4103.1920.5030.8490.8190.878
SwinFusion [20]22'JAS0.7431.6873.2730.5160.9370.8990.900
CDDFuse [42]23'CVPR0.7831.7703.0610.5170.9290.9080.899
Ours25'ICCV0.8361.7243.5080.5880.9320.9040.914
", + "bbox": [ + 514, + 155, + 903, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LRRNet [13], CoCoNet [17], MetaFusion [40], TUFusion [41], DATFuse [24], SwinFusion [20], and CDDFuse [42]. CoCoNet, LRRNet, and MetaFusion are primarily CNN-based methods, while SwinFusion is dominated by the SwinTransformer architecture. TUFusion, DATFuse, CDDFuse, and our method adopt a hybrid mode that combines CNN and attention mechanisms.", + "bbox": [ + 511, + 445, + 906, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We conduct a comprehensive analysis from both qualitative and quantitative perspectives. The qualitative comparison mainly relies on human subjective evaluation of the visual effects of the fused images, primarily concentrating on the preservation of texture details and the retention of salient thermal radiation targets. The quantitative evaluation employs seven mainstream image quality assessment metrics: visual information fidelity (VIF), the sum of the correlations of differences (SCD), mutual information (MI), gradient-based similarity measurement $(\\mathrm{Q_{abf}})$ , structural similarity index measure (SSIM), multiscale structural similarity (MS-SSIM), and feature mutual information (FMI). All selected metrics are positive, with higher values indicating better performance. Detailed explanations can be found in [19].", + "bbox": [ + 511, + 551, + 908, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison. The qualitative comparison results are shown in Fig. 5, which comprises three groups of images arranged from top to bottom, based on the MSRS, $\\mathrm{M}^{3}\\mathrm{FD}$ , and TNO datasets, respectively. The first group of images demonstrates the ability of different methods to preserve thermal radiation information. DATFuse, LRRNet, and TUFusion fail to effectively retain thermal radiation targets, while MetaFusion, CoCoNet, and SwinFusion pro", + "bbox": [ + 511, + 780, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13518", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/effaa20c8c022ff21038661abb44550f753743e79850b925fe2c9e316bf01400.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 85, + 261, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cc0d3a547bd58ee3dab0c2fe34ee53de1aaafd3b0b59c6e47484179aead6c830.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 266, + 87, + 415, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/40f396a4bc400b5aac6470028b527f72bda26d1576c342e7f0f73161d08c5f34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 87, + 568, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a2109d5b234f714b70badb6b9820970cca00aa6c246dab2cff363aaac97d4c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 87, + 725, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce4c5b9602381a3ec87d6399d156a7b137fbd8cdb1d61646754403f914bf6adf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 87, + 882, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ca3ed4cfd5276f720bc8775224568bce96b3892c27943d2ef99c829bad9b746.jpg", + "image_caption": [ + "Infrared", + "CDDFuse" + ], + "image_footnote": [], + "bbox": [ + 114, + 186, + 261, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1da1f1808a81971f07582c57b29eae74ead53665347c64a3787a8076bbd64e0.jpg", + "image_caption": [ + "Visible", + "MetaFusion" + ], + "image_footnote": [], + "bbox": [ + 266, + 186, + 413, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca1fa735f25c3e4120ab77cee5af368349b0101fb2af628bd75c8db61e86dc7a.jpg", + "image_caption": [ + "DATFuse", + "SwinFusion" + ], + "image_footnote": [], + "bbox": [ + 421, + 186, + 568, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/772ccb091d965bb76b9acee8e53585dae1bad537dd41b97ab0f6c9073a98651a.jpg", + "image_caption": [ + "CoCoNet", + "TUFusion" + ], + "image_footnote": [], + "bbox": [ + 578, + 186, + 725, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e6977d8b246d9e104eba81852fb031d76ddff48593c4b7df0cf39f8fe1da95f.jpg", + "image_caption": [ + "LRRNet", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 733, + 186, + 882, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/60face0e156a3d77a5b30d98ff6c3352c7233a6caa1c0141e3fcb32f978fd70f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 295, + 263, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f37d0199dc6b631174df4457d2b28d77f4ae8a9fdcbf49ef41078f5271f5c3b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 295, + 415, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f24e00e8fe6a55285de1652dba2a38fbdae25feaec925ac4cc2ba3b50fff96f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 295, + 568, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8cf6ae43485f692d28aa462a811db6560e1070e36362fb2efc1a1e2cc7a708ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 295, + 725, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/482246027b02c2de08313fd27c1d679b894bcea2de85a718500b8712a9f322ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 295, + 882, + 380 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b94447b5ed31ea6703bb6ab11a636285a1a46da6996a1080d7923b568ed1a52.jpg", + "image_caption": [ + "Infrared", + "CDDFuse" + ], + "image_footnote": [], + "bbox": [ + 116, + 393, + 263, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/371b840f96f31b604d3ebbca952afc44c32691780218907f805653c969daf3da.jpg", + "image_caption": [ + "Visible", + "MetaFusion" + ], + "image_footnote": [], + "bbox": [ + 267, + 393, + 415, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81f49eca2289bf05a458ae9c9d9cac1e874993d7ed1f59f0c8e71a37ef39c834.jpg", + "image_caption": [ + "DATFuse", + "SwinFusion" + ], + "image_footnote": [], + "bbox": [ + 421, + 393, + 568, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/458a400c635fabe7d1a2d7df1273b2cd2a4a753b77f5c5fbf4e430d67e32c433.jpg", + "image_caption": [ + "CoCoNet", + "TUFusion" + ], + "image_footnote": [], + "bbox": [ + 578, + 393, + 725, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2adb9d10a2b425cfdf480b30b35382b9aabc50a5693e7d820603b154fe409b22.jpg", + "image_caption": [ + "LRRNet", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 735, + 393, + 882, + 479 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/229cfc223ce944daafce7c5cb61952c67931e90423825728a8261ed3aae18f91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 116, + 502, + 263, + 587 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/88c3b0fb1403bca2df640933c2e5a1fcc85cec58b08b226ce0a9295d71eb1e97.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 502, + 415, + 587 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf23d8aed191fb9492e734a4c6069bdbe6823b6045190e2c77c45d7baa2b8479.jpg", + "image_caption": [ + "DATFuse" + ], + "image_footnote": [], + "bbox": [ + 424, + 503, + 571, + 588 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/337668a7a80278204ae3ab722fd5307a0d499988e61de6c37ac840403aa8af99.jpg", + "image_caption": [ + "CoCoNet" + ], + "image_footnote": [], + "bbox": [ + 576, + 503, + 723, + 588 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8e17613a1daf2722791cfec22d8ab9669a1853b3a80ba20755ef25d32c0933f.jpg", + "image_caption": [ + "LRRNet" + ], + "image_footnote": [], + "bbox": [ + 733, + 503, + 882, + 588 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/239e8bda70cd922385111cb3a3f7ce5754db7f8f8e2bddd653a31d93705b0e60.jpg", + "image_caption": [ + "Infrared", + "CDDFuse", + "Figure 5. Visual comparison results of our method with seven state-of-the-art methods on typical image pairs from MSRS (the first group), M $^3$ FD (the second group), and TNO (the third group) datasets." + ], + "image_footnote": [], + "bbox": [ + 114, + 603, + 261, + 686 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9154111fd9a2bff71b6416310e8d9fe16e1f394843648c54f1fffe9078a9e38b.jpg", + "image_caption": [ + "Visible", + "MetaFusion" + ], + "image_footnote": [], + "bbox": [ + 267, + 603, + 415, + 686 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9cc03f50a44e3589d34a3b402d596b6d1d3912b70e6e92170c904697d3074447.jpg", + "image_caption": [ + "SwinFusion" + ], + "image_footnote": [], + "bbox": [ + 424, + 603, + 570, + 686 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55dc447b27e0fd04b75b2fdce73441c8004605a10d5abb604b37cf0a65d4484a.jpg", + "image_caption": [ + "TUFusion" + ], + "image_footnote": [], + "bbox": [ + 578, + 603, + 725, + 686 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1136693fbf960b67e9d1a31eb37ce92212e859aa8646e9fff2b45bf574f7c5a.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 733, + 603, + 882, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "duce distortions in the legs of the highlighted person. The second and third groups of images illustrate the capability of different methods to preserve details. For instance, neither the clouds in the second group nor the shrubberies in the third group are effectively preserved by DATFuse, LRR-Net, SwinFusion, and TUFusion, while CoCoNet and MetaFusion introduce significant noise. Only CDDFuse and our proposed method perform well in both aspects.", + "bbox": [ + 89, + 773, + 485, + 895 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison. The quantitative comparison results are presented in Tab. 1. In the context of the MSRS dataset, our method ranks first across all evaluation metrics, demonstrating the following advantages. 1) The optimal VIF indicates that our technique yields fused outputs with enhanced visual quality, further confirming its qualitative superiority. 2) The ideal SCD, MI, and FMI suggest that our method is most effective in conveying data from the", + "bbox": [ + 511, + 773, + 906, + 895 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13519", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "original input to the outcome. 3) The superior SSIM and MS-SSIM show that our model minimizes information loss and image distortion. 4) The perfect Qabf illustrates that our solution successfully preserves the maximum edge details. Furthermore, to verify the generalization competence of our method, we perform tests on the $\\mathbf{M}^3\\mathrm{FD}$ and TNO datasets without any fine-tuning. In the context of the $\\mathbf{M}^3\\mathrm{FD}$ dataset, our technique leads in every metric, with the sole exceptions being SCD and MI. Within the TNO dataset, it secures four top positions and two runner-up spots.", + "bbox": [ + 89, + 90, + 480, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 253, + 254, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Within this part, we undertake multiple ablations. Each one is executed using the MSRS test dataset as a foundation. Extension. To explore the correctness of extending the attention properties of infrared images to the visible modality, two sets of controlled experiments are performed: the first set utilize only I-SCA, while the second set employ both I-SCA and V-SCA. The experimental results are shown in Tab. 2. The integration of V-SCA led to enhancements in all measurements, with the sole exception of a minor reduction in SCD and SSIM, which confirms the validity of this extension.", + "bbox": [ + 89, + 275, + 482, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "I/V-SCA. To validate the effectiveness of I-SCA and V-SCA, we design two strategies in this subsection. 1) Directly removing the proposed modules, which is denoted as w/o I/V-SCA. 2) To prevent inconsistencies in model complexity before and after the removal of I/V-SCA, which could introduce additional experimental bias, we further replace I/V-SCA with Restormer Blocks. This ablation is denoted as \"to Restormer\". Findings from the experiment are detailed in Tab. 3, which illustrates that our technique holds the top position in five of the seven assessment measures, affirming the efficacy of I/V-SCA.", + "bbox": [ + 89, + 441, + 482, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CBSM. For confirming CBSM's efficiency, we substitute CBSM with Parameters, $Conv_{1 \\times 1}$ , and $Conv_{3 \\times 3}$ , respectively. The experiment's findings are outlined in Tab. 4. Compared to the other three schemes, the introduction of CBSM results in more than half of the evaluated metrics being ranked first, which suggests that CBSM is capable of efficiently diminishing the disparities between the source image and the intermediate features of another modality.", + "bbox": [ + 89, + 609, + 482, + 731 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a93543c0dfea1b81987b445ce8dbee486441bf7baff254fe65b980d1da101c63.jpg", + "table_caption": [ + "Table 2. Ablation before and after the extension of I-SCA." + ], + "table_footnote": [], + "table_body": "
ConfigurationsVIFSCDMIQabfSSIMMS_SSIMFMI
I-SCAV-SCA
✓✗1.0391.7214.5710.7030.9820.9750.932
✓✓1.0611.7005.1110.7150.9810.9750.933
", + "bbox": [ + 91, + 768, + 480, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Efficiency Comparison", + "text_level": 1, + "bbox": [ + 89, + 847, + 302, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Within this part, we evaluate the parameters, FLOPs, and FPS of our work with the most competitive methods", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3df4181b1649a537982b7ae52d31f679c1f4aefabf436c48e83e950491c5cb72.jpg", + "table_caption": [ + "Table 3. Ablation of the I-SCA and V-SCA modules." + ], + "table_footnote": [], + "table_body": "
ConfigurationsVIFSCDMIQabfSSIMMS_SSIMFMI
w/o I/V-SCA1.0651.6695.6510.7080.9680.9670.933
to Restormer1.0631.6815.4630.7120.9750.9710.933
Ours1.0611.7005.1110.7150.9810.9750.933
", + "bbox": [ + 514, + 114, + 903, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/966c5190c88307c67d0b1ff430aeccfe9b04d620d161950696eca4d76e1f604a.jpg", + "table_caption": [ + "Table 4. Ablation of the CBSM module." + ], + "table_footnote": [], + "table_body": "
ConfigurationsVIFSCDMIQabfSSIMMS_SSIMFMI
Parameters1.0601.7085.3030.7120.9770.9730.932
Conv1×11.0261.6914.6710.6910.9750.9670.932
Conv3×31.0661.7055.2540.7110.9770.9730.932
CBSM1.0611.7005.1110.7150.9810.9750.933
", + "bbox": [ + 514, + 205, + 903, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to highlight its speed advantage. Findings are presented in Tab. 5, where the FLOPs are determined based on an input dimension of $128 \\times 128 \\times 1$ , and the FPS on $\\mathbf{M}^3\\mathrm{FD}$ is measured with images of reduced resolution by half. Our method achieves the highest FPS, enabling seamless integration with both low- and high-level computer vision tasks.", + "bbox": [ + 511, + 295, + 906, + 387 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/19b6a1b42a694f5218ced497841d1471a8a4d7ff8353a839e5134577928b1848.jpg", + "table_caption": [ + "Table 5. The parameters, FLOPs, and FPS." + ], + "table_footnote": [], + "table_body": "
Params (M) ↓FLOPs (G) ↓FPS ↑
MSRSM3FDTNO
SwinFusion [20]0.92717.9860.6241.0160.651
CDDFuse [42]1.18629.21339.72340.48040.179
Ours0.5659.207132.303137.271126.537
", + "bbox": [ + 514, + 436, + 903, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 532, + 633, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper reveals the inherent \"attention properties\" of infrared images from the Grad-CAM of ResNet-50. To incorporate the property into IVF, we propose I-SCA and extend it to the visible modality, subsequently introducing V-SCA. The auxiliary module (CBSM) further ensures fusion quality by enhancing feature channels, enabling spatial transformations, and suppressing negative information in the source images. Additionally, the linear computations of I-SCA and V-SCA guarantee efficiency, achieving FPS of 132, 137, and 127 on MSRS, $\\mathbf{M}^3\\mathrm{FD}$ , and TNO, respectively. Moving forward, our focus will shift to crafting more specialized modules and undertaking an in-depth methodological examination to enhance the refinement of our approach.", + "bbox": [ + 511, + 556, + 906, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 768, + 671, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (62272426), the Shanxi Province Science and Technology Major Special Project (202201150401021), and the Fundamental Research Program of Shanxi Province (TZLH20230818005, 202303021212189, 202303021212372, 202403021212166).", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13520", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3", + "[2] Yaming Cao, Lei Guo, Fengguang Xiong, Liqun Kuang, and Xie Han. Physical-simulation-based dynamic template matching method for remote sensing small object detection. IEEE Transactions on Geoscience and Remote Sensing, 62: 1-14, 2023. 3", + "[3] Zhihao Chang, Zhixi Feng, Shuyuan Yang, and Quanwei Gao. Aft: Adaptive fusion transformer for visible and infrared images. IEEE Transactions on Image Processing, 32: 2077-2092, 2023. 3", + "[4] Jun Chen, Jianfeng Ding, and Jiayi Ma. Hitfusion: Infrared and visible image fusion for high-level vision tasks using transformer. IEEE Transactions on Multimedia, 26: 10145-10159, 2024. 3", + "[5] Chunyang Cheng, Tianyang Xu, and Xiao-Jun Wu. Mufusion: A general unsupervised image fusion network based on memory unit. Information Fusion, 92:80-92, 2023. 3", + "[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3", + "[7] Keying Du, Huafeng Li, Yafei Zhang, and Zhengtao Yu. Chitnet: A complementary to harmonious information transfer network for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 74: 1-17, 2025. 5", + "[8] Xueyan Gao and Shiguang Liu. Bcmfuse: A bilateral cross-modal feature interaction-based network for infrared and visible image fusion. Remote Sensing, 16(17):3136, 2024. 3", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2016. 2", + "[10] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3, 4", + "[11] Shahid Karim, Geng Tong, Jinyang Li, Akeel Qadir, Umar Farooq, and Yiting Yu. Current advances and future perspectives of image fusion: A comprehensive review. Information Fusion, 90:185-217, 2023. 1", + "[12] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2019. 3", + "[13] Hui Li, Tianyang Xu, Xiao-Jun Wu, Jiwen Lu, and Josef Kittler. Lrrnet: A novel representation learning guided fusion network for infrared and visible images. IEEE transactions on pattern analysis and machine intelligence, 45(9):11040-11052, 2023. 3, 6", + "[14] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In European Conference on Computer Vision, pages 719-735, 2022. 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3", + "[16] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5811, 2022. 6", + "[17] Jinyuan Liu, Runjia Lin, Guanyao Wu, Risheng Liu, Zhongxuan Luo, and Xin Fan. Coconet: Coupled contrastive learning network with multi-level feature ensemble for multi-modality image fusion. International Journal of Computer Vision, 132(5):1748-1775, 2024. 3, 6", + "[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3", + "[19] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information fusion, 45:153-178, 2019. 6", + "[20] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swim transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200-1217, 2022. 3, 6, 8", + "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[22] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), page 618-626. IEEE, 2017. 2", + "[23] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Information Fusion, 83:79-92, 2022. 6", + "[24] Wei Tang, Fazhi He, Yu Liu, Yansong Duan, and Tongzhen Si. Datfuse: Infrared and visible image fusion via dual attention transformer. IEEE Transactions on Circuits and Systems for Video Technology, 33(7):3159-3172, 2023. 3, 6", + "[25] Alexander Toet. The tno multiband image data collection. Data in brief, 15:249-251, 2017. 6", + "[26] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3", + "[27] Boying Wang, Libo Zhang, Longyin Wen, Xianglong Liu, and Yanjun Wu. Towards real-world prohibited item detection: A large-scale x-ray benchmark. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13521", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF international conference on computer vision, pages 5412-5421, 2021.3", + "[28] Boying Wang, Ruyi Ji, Libo Zhang, Yanjun Wu, and Jing Liu. Learning to zoom: Exploiting mixed-scale contextual information for object detection. Expert Systems with Applications, 264:125871, 2025. 3", + "[29] Zhishe Wang, Yanlin Chen, Wenyu Shao, Hui Li, and Lei Zhang. Swinfuse: A residual swim transformer fusion network for infrared and visible images. IEEE Transactions on Instrumentation and Measurement, 71:1-12, 2022. 3", + "[30] Zhishe Wang, Wenyu Shao, Yanlin Chen, Jiawei Xu, and Xiaoqin Zhang. Infrared and visible image fusion via interactive compensatory attention adversarial learning. IEEE Transactions on Multimedia, 25:7800-7813, 2022. 3", + "[31] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3", + "[32] Yongbiao Xiao, Hui Li, Chunyang Cheng, and Xiaoning Song. Le2fusion: A novel local edge enhancement module for infrared and visible image fusion. In International conference on image and graphics, pages 293-305. Springer, 2023. 3", + "[33] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 6", + "[34] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Text-if: Leveraging semantic text guidance for degradation-aware and interactive image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27026-27035, 2024. 3", + "[35] Jun Yue, Leyuan Fang, Shaobo Xia, Yue Deng, and Jiayi Ma. Dif-fusion: Toward high color fidelity in infrared and visible image fusion with diffusion models. IEEE Transactions on Image Processing, 32:5705-5720, 2023. 3", + "[36] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 3", + "[37] Hao Zhang, Xuhui Zuo, Jie Jiang, Chunchao Guo, and Jiayi Ma. Mrfs: Mutually reinforcing image fusion and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26974-26983, 2024. 3", + "[38] Cheng Zhao, Peng Yang, Feng Zhou, Guanghui Yue, Shuigen Wang, Huisi Wu, Guoliang Chen, Tianfu Wang, and Baiying Lei. Mhw-gan: multidiscriminator hierarchical wavelet generative adversarial network for multimodal image fusion. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3", + "[39] Rong Zhao, Xie Han, Xindong Guo, Liquun Kuang, Xiaowen Yang, and Fusheng Sun. Exploring the point feature relation on point cloud for multi-view stereo. IEEE Transactions on Circuits and Systems for Video Technology, 33(11): 6747-6763, 2023. 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13955-13965, 2023. 3, 6", + "[41] Yangyang Zhao, Qingchun Zheng, Peihao Zhu, Xu Zhang, and Wenpeng Ma. Tufusion: A transformer-based universal fusion algorithm for multimodal images. IEEE Transactions on Circuits and Systems for Video Technology, 34(3): 1712-1725, 2024. 3, 6", + "[42] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5906-5916, 2023. 3, 5, 6, 8", + "[43] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: denoising diffusion model for multi-modality image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8082-8093, 2023. 3", + "[44] Zixiang Zhao, Lilun Deng, Haowen Bai, Yukun Cui, Zhipeng Zhang, Yulun Zhang, Haotong Qin, Dongdong Chen, Jiangshe Zhang, Peng Wang, et al. Image fusion via vision-language model. In Proceedings of the 41st International Conference on Machine Learning, pages 60749-60765, 2024. 3", + "[45] Pengfei Zhu, Yang Sun, Bing Cao, and Qinghua Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099-7108, 2024. 3" + ], + "bbox": [ + 516, + 90, + 906, + 542 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13522", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/cab6b62f-c3b3-45e2-8877-ae8b2379307d_model.json b/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/cab6b62f-c3b3-45e2-8877-ae8b2379307d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..72a595b947b1e7889fa3c05c614a71b19231fc9f --- /dev/null +++ b/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/cab6b62f-c3b3-45e2-8877-ae8b2379307d_model.json @@ -0,0 +1,2673 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.131, + 0.89, + 0.153 + ], + "angle": 0, + "content": "The Source Image is the Best Attention for Infrared and Visible Image Fusion" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.181, + 0.792, + 0.218 + ], + "angle": 0, + "content": "Song Wang Xie Han Liquun Kuang\\* Boying Wang Zhongyu Chen Zherui Qiao Fan Yang Xiaoxia Liu Bingyu Zhang Zhixun Wang" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.22, + 0.786, + 0.234 + ], + "angle": 0, + "content": "School of Computer Science and Technology, North University of China, Taiyuan 030051, China" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.236, + 0.751, + 0.251 + ], + "angle": 0, + "content": "Shanxi Key Laboratory of Machine Vision & Virtual Reality, Taiyuan 030051, China" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.254, + 0.836, + 0.269 + ], + "angle": 0, + "content": "Shanxi Vision Information Processing and Intelligent Robot Engineering Research Center, Taiyuan 030051, China" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.273, + 0.681, + 0.287 + ], + "angle": 0, + "content": "b20230704@st.nuc.edu.cn,kuang@nuc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.485, + 0.779 + ], + "angle": 0, + "content": "Infrared and visible image fusion (IVF) endeavors to engineer composite outputs by blending optimal virtues of divergent modalities. This paper reveals, unprecedentedly, the intrinsic \"attention properties\" of infrared images, which directly arise from their physical characteristics (i.e., heat distribution) and can be linked to attention mechanisms naturally, as observed in the gradient-weighted class activation mapping (Grad-CAM) visualization analysis of image classification models. To incorporate this property into IVF for better fusion, we propose the source infrared cross attention (I-SCA) and further extend it to the visible modality, subsequently introducing the source visible cross attention (V-SCA). The joint use of I-SCA and V-SCA greatly alleviate longstanding issues, such as insufficient and incomplete multimodal feature interaction and fusion, in IVF. Moreover, an auxiliary component for I-SCA and V-SCA, termed CBSM, is employed to boost the channel, map space, and suppress redundancy and misleading information of the source images. Specifically, we directly treat the CBSM-processed raw image as the query, while the intermediate features of another modality are treated as keys and values in I-SCA and V-SCA. Unlike attention mechanisms that divide images into patches or limit computations to local windows, our cross attention modules achieve smoother and more robust IVF through true global modeling across the entire image space with linear complexity. Comparison with current SOTA methods on three popular public datasets confirms its superiority." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.808, + 0.222, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.484, + 0.879 + ], + "angle": 0, + "content": "IVF, a bedrock of image processing, with the objective of engendering a composite that encapsulates the considerable thermal radiation from infrared and the intricacies de" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.32, + 0.705, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.42, + 0.635, + 0.428 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.32, + 0.866, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.42, + 0.798, + 0.427 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.429, + 0.703, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.517, + 0.635, + 0.524 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.428, + 0.865, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.517, + 0.798, + 0.524 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.526, + 0.703, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.613, + 0.636, + 0.621 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.525, + 0.866, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.613, + 0.798, + 0.621 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.636, + 0.907, + 0.775 + ], + "angle": 0, + "content": "Figure 1. (a, b) Grad-CAM visualization results for two images (00022004, 00028021), both labeled as \"dog\", predicted by the final layer of ResNet-50 on the ImageNet validation set. The color gradient from red to blue indicates the decreasing level of attention paid by the model to different regions. (c) An infrared image example (00519D) consists of highlighted thermal radiation regions and remaining ordinary regions from the MSRS dataset. (d) Visible image of the same scene corresponding to subplot (c). (e) Subplot (c) is displayed by applying the Jet colormap. (f) The negative transformation of subplot (c) is visualized with the Jet colormap." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.907, + 0.883 + ], + "angle": 0, + "content": "tail from visible. Prior to the surge in deep learning, this process was typically achieved by traditional algorithms in spatial and transform domains [11]. However, thanks to the development of neural networks and attention mechanisms, heuristic algorithms have become more widely adopted." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In this paper, we draw inspiration from the Grad-CAM" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13513" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.349 + ], + "angle": 0, + "content": "[22] visualization results of classic image classification models, such as ResNet-50 [9], and discover the inherent \"attention properties\" of infrared images. These Grad-CAM visualizations, as shown in Fig. 1(a) and Fig. 1(b), reveal how the model assigns varying weights to different regions during the prediction process: regions marked in red represent \"important areas\" that exert a substantial impact on the classification result, while regions marked in blue correspond to \"general areas\" with lesser contribution. Interestingly, we observe that the structural characteristics of infrared images (e.g., as shown in Fig. 1(c)) are strikingly similar to the Grad-CAM images, as both consist of important areas and general areas (in infrared images, these regions correspond to the highlighted areas with higher thermal energy and the darker zones with lower energy). To more intuitively showcase this similarity, we re-map the infrared image with the Jet colormap (see Fig. 1(e))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Based on these observations, we believe that infrared images inherently possess an attention property similar to the one exhibited by Grad-CAM. Unlike image classification models, which require be trained to learn attention patterns, the attention properties of infrared images directly stem from their physical characteristics—the heat distribution. Moreover, this attribute is not only more specific and accurate but also avoids issues such as insufficient or misdirected attention that may occur in image classification models. For instance, when ResNet-50 processes the dog in Fig. 1(a), it focuses only on the dog's eyes, and when processing the dog in Fig. 1(b), it not only focuses on the dog itself but also incorrectly attends to irrelevant backgrounds. In contrast, infrared images provide precise localization of heat sources, ensuring that all attention is concentrated on specific areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Based on the above finding, an immediate question arises: how can we fully leverage the inherent attention properties of infrared images to obtain better-fused images in IVF? Intuitively, we think that it can be linked to the attention mechanism [26], as both essentially represent the importance of different regions. Furthermore, the interaction of cross-modal features plays a pivotal role in image fusion. Therefore, our goal is to introduce the attention properties at this stage to counteract the persistent obstacles in image fusion, such as inadequate cross-domain information exchange and incomplete multimodal feature fusion, thereby enhancing the caliber of the merged images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Specifically, we propose the source infrared cross attention (I-SCA) based on a modified Restormer [36], which can perform global feature modeling in the image space with linear complexity. I-SCA first takes the original infrared image and the intermediate features of the visible image as inputs. It then treats the infrared image as the query (Q), thanks to the properties of infrared images, which allow them to be directly used as queries without processing," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "while the visible features serve as the keys (K) and values (V), enabling an infrared heat prior-guided cross-domain information exchange. However, this approach can only prioritize on \"important areas\", whereas image fusion is a global pixel-wise task that demands equal treatment for all areas. Fortunately, a facile solution is to treat the negative transformation of the infrared image (as shown in Fig. 1(f)) as another query to allow significant attention complementarity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.905, + 0.385 + ], + "angle": 0, + "content": "Nevertheless, using I-SCA only also faces a critical problem, i.e., it can only capture \"infrared-guided\" cross-modal interaction features. Theoretically, the optimal fused image could be obtained by simultaneously fusing features dominated by both infrared and visible images. Therefore, we make a bold inference: we extend the discovery to visible and, based on this, propose the source visible cross attention (V-SCA). While visible images lack explicit physical saliency, their visually significant regions (e.g., edges, textures, and contrast) can still serve as effective priors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.905, + 0.617 + ], + "angle": 0, + "content": "Finally, since both I-SCA and V-SCA require inputs to have the same channels, but the channels of the source image and the intermediate representation from the other modality are not consistent, we deploy an adaptive channel boosting and intelligent space mapping module (CBSM), which also serves crucial roles in transforming the feature space of the source image from its initial state to one more suitable for attention mechanisms to reduce discrepancies between features at different stages and suppressing redundancy or misleading content contained in the initial images. To validate the effectiveness of I-SCA, V-SCA, and CBSM, extensive analysis are conducted on three popular public datasets. Experiments demonstrate that both qualitative and quantitative results outperform state-of-the-art (SOTA) methods." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.622, + 0.84, + 0.636 + ], + "angle": 0, + "content": "The essence of our contributions is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.642, + 0.905, + 0.702 + ], + "angle": 0, + "content": "- We observe a strong structural correlation between infrared images and Grad-CAM, leading us to hypothesize, for the first time, that infrared images possess an inherently exceptionally precise attention property." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.703, + 0.905, + 0.793 + ], + "angle": 0, + "content": "- Based on this attention property, we propose a heat-distribution-prior-guided cross-modal attention mechanism, which can be further extend to the visible modality. Additionally, an auxiliary module, CBSM, is deployed to complement the proposed attention mechanism for further enhancing fusion quality." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.793, + 0.905, + 0.853 + ], + "angle": 0, + "content": "- Our findings inspire a novel paradigm guided by source image priors for image fusion, significantly alleviating challenges in cross-domain feature interaction and fusion faced by current methods." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.853, + 0.905, + 0.899 + ], + "angle": 0, + "content": "- Comprehensive experiments instituted on three prominent available datasets with SOTA works demonstrate the superiority." + }, + { + "type": "list", + "bbox": [ + 0.533, + 0.642, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13514" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.241, + 0.106 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.116, + 0.378, + 0.131 + ], + "angle": 0, + "content": "2.1. Attention-based Fusion Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.486, + 0.775 + ], + "angle": 0, + "content": "The triumph of transformer [26] in NLP has sparked considerable interest across the academic community. To establish a foundation for the integration of this technology in computer vision, Alexey et al. proposed Vision Transformer (ViT) [6], which segments an image into multiple units, then flattens the units into a shape compatible with text sequences, allowing direct application of the original transformer for long-range modeling. Subsequently, works such as TUFusion [41], DATfuse [24], HitFusion [4], AFT-Fusion [3], etc., extended ViT to the image fusion task. On the one hand, TUFusion adopted an encoder-decoder pipeline, featuring a dual-branch structure that merged parallel DenseBlock [12] and ViT in the encoder. DATFuse, on the other hand, incorporated channel and spatial attention mechanisms similar to CBAM [31] before the self-attention mechanism and performed fusion at the image level. Although transformer-based methods have shown great effectiveness, they are consistently constrained by the computational burden of the self-attention in image space. To remedy this issue, Liu et al. proposed SwinTransformer [18], which divides images into patches and further splits the patches into multiple windows, restricting self-attention to each window. Later, SwinFuse [29] used a full SwinTransformer approach to obtain fused images. SwinFusion [20] further extended this by introducing a cross-modal SwinTransformer, which generalizes window attention to cross-modal feature fusion. However, the practice of restricting attention to local regions contradicts the fundamental idea of attention mechanisms, which is to model global information. To achieve true overall modeling linearly, Zamir et al. proposed Restormer, which introduces multi-dconv head transposed attention (MDTA) and gated dconv feedforward network (GDFN) to replace the self-attention and feed-forward network [36]. Later, CDDFuse [42] further validated the effectiveness of the Restormer in image fusion. However, it only applied the Restormer to a single modality and overlooked the potential for feature interactions across multiple modalities. This paper narrowed down the challenge of interaction between different modalities by proposing I-SCA and V-SCA, and further enabled Restormer to be used for the multimodal information interaction problem." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.787, + 0.345, + 0.803 + ], + "angle": 0, + "content": "2.2. CNN-based Fusion Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Currently, the realm of deep learning-driven fusion is segmented into various types, encompassing techniques reliant on autoencoders, end-to-end models, unified models, amalgamations with downstream or other specialized tasks, and approaches utilizing generative adversarial networks (GAN). Autoencoder-based methods [14, 32] first" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.665 + ], + "angle": 0, + "content": "train an encoder-decoder pair on datasets like COCO [15] and then directly generalize them for image fusion. However, the performance of such methods is markedly hampered by the quality of the fusion rules designed subjectively, and developing a fusion strategy that works well with the encoder-decoder framework is not trivial. Consequently, a series of end-to-end methods [8, 13] were later proposed. These methods train all modules, including the feature fusion layer, with learning strategies, which greatly reduces the challenges related to constructing fusion strategies in autoencoder-based methods. Furthermore, as image fusion has evolved, the community has become dissatisfied with using a single model to solve just one specific fusion task, leading to unified models [5, 17, 45] for multiple tasks. However, these methods adopt a re-training strategy for each new task, rather than using a single weight parameter to handle multiple tasks simultaneously. Meanwhile, along with the development of downstream tasks such as object detection [2, 27, 28], scene reconstruction [39], etc., a major trend has been to combine image fusion with these computer vision sub-tasks, facilitating the mutual promotion of both [37, 40]. Another category of methods [34, 44] leverages large language models such as GPT [1] and CLIP [21] to guide the fusion process, achieving cross-domain integration for different tasks. Finally, GAN-based methods [30, 38] transform the fusion into an adversarial problem. Through continuous antagonistic training, the generator's output progressively converges toward the target distribution. Nevertheless, these techniques face hurdles like mode collapse, erratic training, and a lack of clear understanding. Later works proposed denoising diffusion model-based methods [35, 43] to address these issues. Although the core innovation of this study lies in source images prior-guided cross-modal attention, approximately half of the components are still based on the convolutional neural network (CNN). To ensure a comprehensive evaluation, we undertake a comparative analysis of it with several recent CNN-based SOTA methods discussed above." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.684, + 0.606, + 0.7 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.711, + 0.71, + 0.726 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, a hybrid of CNN and attention mechanisms is employed to achieve the IVF. The overall architecture is shown in Fig. 2. Specifically, we construct a serial pipeline, consisting of SE-ResNet Block [10], Restormer Block [36], and I-SCA/V-SCA for infrared and visible branches. For each modality, we first employ the SE-ResNet Block (the SE acts as a gating mechanism to suppress information redundancy by refine features) for initial feature extraction while increasing the feature channel from ch_in to ch_mid. The Restormer Block then performs further deep extraction with another key function of transforming the feature space" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13515" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.096, + 0.912, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.361, + 0.908, + 0.418 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of our method, which incorporates SE-ResNet Block, Restormer Block, I-SCA/V-SCA, and CBSM. ir, vi denote infrared and visible images, and \\(ir\\_n\\), \\(vi\\_n\\) signify negative transformations of infrared and visible images. ch_in, ch_mid, and ch_out represent input, middle, and output channels, respectively. “C” indicates a channel-wise concatenation operation. “Q” stands for the query in the attention mechanism, while “K” and “V” symbolize the key and value, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.443, + 0.482, + 0.61 + ], + "angle": 0, + "content": "from the domain where convolutional operations excel to the one where attention is more effective, thereby facilitating the subsequent cross-modal feature interaction. The I-SCA/V-SCA subsequently takes the features of the current modality along with either the image of another modality processed by CBSM or the negative transformation of that image for feature interaction. Post processing via I-SCA/V-SCA, a quartet of feature maps, steered by the originating image, is acquired. Then they will merge across the channel dimension and be processed through an additional SE-ResNet Block before being output." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.182, + 0.634 + ], + "angle": 0, + "content": "3.2. CBSM" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.654, + 0.475, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Figure 3. The composition of adaptive channel boosting and intelligent space mapping module (CBSM). The SE is employed to suppress redundancy and misleading information in the source images. \\(\\odot\\) denotes a channel-wise multiplication operation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Since attention calculation in I-SCA and V-SCA requires all channels of inputs to be consistent, it is imperative to augment the channel dimension of the source image" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.908, + 0.779 + ], + "angle": 0, + "content": "from ch_in to ch_mid. In general, this can be achieved by Parameters, \\(Conv_{1\\times 1}\\), or \\(Conv_{3\\times 3}\\). However, all three have a common problem: after channel expansion, the multi-channel feature space still remains in the initial state. In cross-modal attention, directly using feature values from the raw space for interaction with deep features will significantly degrade the information exchange effect due to spatial discrepancy. Therefore, a key issue is how to reduce the gap between them. Additionally, the source image may introduce redundancy or misleading cues in certain cases, making it crucial to mitigate potential negative effects. To address this, CMSM module, which consists of convolutional layers (for channel boosting), activation layers (for spatial mapping), and Squeeze-and-Excitation (SE) [10] layer (for suppressing negative influences) is employed, as shown in Fig. 3. The structure of CMSM is intentionally simple, because it aligns with our goal of using the raw image priors to guide feature interactions. A more complex design could significantly alter the initial pattern of the source image, conflicting with our motivation. For the input source image, denoted as \\(I_m^{raw} \\in \\mathbb{R}^{H \\times W \\times 1}\\), this module performs the following operations:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.784, + 0.907, + 0.803 + ], + "angle": 0, + "content": "\\[\nI _ {m} ^ {t r s} = \\alpha \\left(S E \\left(C o n v _ {3 \\times 3} \\left(\\alpha \\left(C o n v _ {3 \\times 3} \\left(I _ {m} ^ {r a w}\\right)\\right)\\right)\\right)\\right). (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The result of Eq. (1), denoted as \\( I_{m}^{trs} \\in \\mathbb{R}^{H \\times W \\times 48} \\), represents the image after channel boosting and space transformation. The variable \\( m \\) represents different input modalities, and desirable values include \\( ir \\), \\( vi \\), \\( ir\\_n \\), and \\( vi\\_n \\). \\( ir\\_n \\) and \\( vi\\_n \\) denote the negative transforms of the infrared and the visible image, respectively. \\( \\alpha(\\cdot) \\) is the parametric" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13516" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.289, + 0.106 + ], + "angle": 0, + "content": "rectified linear unit (PReLU)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.12, + 0.439, + 0.138 + ], + "angle": 0, + "content": "3.3. Cross Attention Based on Source Images" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.161, + 0.452, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.313, + 0.449, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.425, + 0.483, + 0.468 + ], + "angle": 0, + "content": "Figure 4. Our proposed I-SCA and V-SCA, both of which consist of two modules: I-DCA/V-DCA and GDFN. Structurally, the two are identical, differing only in their input content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.759 + ], + "angle": 0, + "content": "The introduction of I-SCA and V-SCA aims to reduce the difficulty of multimodal feature interactions. The structures of these two components are illustrated in Fig. 4. In fact, both share an identical structural composition, differing only in the content of their inputs. Although the symmetrical design is used between I-SCA and V-SCA, this does not imply an inability to reflect modal heterogeneity. On the contrary, I-SCA emphasizes the heat distribution, contours, and structural priors of infrared images, while V-SCA focuses on edges, textures, and high-contrast guidance. Specifically, I-SCA comprises the dual-dconv infrared cross attention (I-DCA) and the gated dconv feedforward network (GDFN). Its inputs include \\( I_{ir}^{trs} \\) (or \\( I_{ir,n}^{trs} \\)) and the visible features \\( feat_{vi}^{mid} \\), which are processed by the Restormer Block in the visible branch. In contrast, V-SCA is composed of the dual-dconv visible cross attention (V-DCA) and the GDFN, with inputs \\( I_{vi}^{trs} \\) (or \\( I_{vi,n}^{trs} \\)) and the infrared features \\( feat_{ir}^{mid} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.483, + 0.866 + ], + "angle": 0, + "content": "To illustrate the working mechanism of I-SCA and V-SCA, we demonstrate I-SCA, with inputs \\(I_{ir}^{trs}\\) and \\(feat_{vi}^{mid}\\), as an example. In I-SCA, both \\(I_{ir}^{trs}\\) and \\(feat_{vi}^{mid}\\) first undergo the layer normalization (LN) in I-DCA. Following this, \\(I_{ir}^{trs}\\) is directly used as the query (Q) after LN, while \\(feat_{vi}^{mid}\\) is further processed to generate the key (K) and value (V). The process is described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.878, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\nK, V = D \\operatorname {c o n v} \\left(C o n v _ {1 \\times 1} \\left(L N \\left(f e a t _ {v i} ^ {m i d}\\right)\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "where \\( Dconv(\\cdot) \\) denotes the depthwise separable convolution, consisting of two steps: depthwise convolution and pointwise convolution. This operation preserves spatial dimensions and independently processes the channels at each spatial position. Subsequently, the Q \\( (I_{ir}^{trs}) \\), K, and V are reshaped for compatibility and undergo the following operation:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.211, + 0.905, + 0.251 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f e a t _ {i r} ^ {o r i} = C o n v _ {1 \\times 1} \\left(V \\cdot S o f t m a x \\left(\\left(I _ {i r} ^ {t r s} \\cdot K\\right) \\odot s\\right)\\right) \\\\ + f e a t _ {v i} ^ {m i d}, (3) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.262, + 0.907, + 0.324 + ], + "angle": 0, + "content": "resulting in infrared-oriented intermediate features \\( \\text{feat}_{ir}^{\\text{ori}} \\), where \\( \\odot \\) represents element-wise multiplication, and \\( s \\) is a learnable parameter. These features are then processed through the GDFN module as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.334, + 0.907, + 0.353 + ], + "angle": 0, + "content": "\\[\nY _ {1}, Y _ {2} = D \\operatorname {c o n v} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(L N \\left(f e a t _ {i r} ^ {\\text {o r i}}\\right)\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.385, + 0.906, + 0.408 + ], + "angle": 0, + "content": "\\[\nf e a t _ {i r} ^ {o r i} ^ {\\prime} = C o n v _ {1 \\times 1} \\left(G E L U \\left(Y _ {1}\\right) \\odot Y _ {2}\\right) + f e a t _ {i r} ^ {o r i}, (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.417, + 0.874, + 0.431 + ], + "angle": 0, + "content": "where GELU refers to the Gaussian Error Linear Unit." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.618 + ], + "angle": 0, + "content": "After the above process, four source images-oriented features, which are further refined by GDFN, denoted as \\( \\text{feat}_m^{\\text{ori}'} \\), are obtained. These features, corresponding to different modalities (\\( m = ir \\), \\( vi \\), \\( ir\\_n \\), and \\( vi\\_n \\)), are arranged along the axis of the channel and further refined by another SE-ResNet Block. The transformation of selfattention from the spatial dimension to the channel dimension, implemented by \\( Dconv(\\cdot) \\) and \\( Conv_{1\\times 1}(\\cdot) \\), ensures the linear computational characteristic of I-SCA and V-SCA. Additionally, the gating mechanism within the GDFN effectively suppresses redundant features, allowing only beneficial features to propagate through the module." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.627, + 0.661, + 0.642 + ], + "angle": 0, + "content": "3.4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.906, + 0.707 + ], + "angle": 0, + "content": "Our approach utilizes intensity \\((L_{int})\\) and gradient \\((L_{grad})\\) loss techniques to guarantee comprehensive and detailed data for the combined image. The complete loss is characterized as:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.721, + 0.905, + 0.737 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t o t a l}} = \\alpha_ {1} L _ {\\text {i n t}} + L _ {\\text {g r a d}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.905, + 0.777 + ], + "angle": 0, + "content": "where \\(\\alpha_{1}\\) represents the intensity loss factor. The specific definition of intensity loss is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.787, + 0.905, + 0.817 + ], + "angle": 0, + "content": "\\[\nL _ {i n t} = \\frac {1}{H W} \\| \\max \\left(I _ {i r} ^ {r a w}, I _ {v i} ^ {r a w}\\right) - f e a t _ {o u t} \\| _ {1}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\( \\max(\\cdot) \\) represents the element-wise maximum operation, \\( \\text{feat}_{\\text{out}} \\) signifies the final output of our model, and \\( \\| \\cdot \\|_1 \\) denotes the \\( l_1 \\)-norm. Inspired by [42] and [7], the gradient term combines Sobel for edge preservation and Laplacian for noise suppression, guiding the model to capture" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13517" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "valuable information while avoiding distractions. The formula is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.134, + 0.483, + 0.151 + ], + "angle": 0, + "content": "\\[\nL _ {g r a d} = \\alpha_ {2} L _ {S} + \\alpha_ {3} L _ {L}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.183, + 0.496, + 0.224 + ], + "angle": 0, + "content": "\\[\nL _ {S} = \\frac {1}{H W} \\| \\max \\left(\\left| \\nabla_ {S} I _ {i r} ^ {\\text {r a w}} \\right|, \\left| \\nabla_ {S} I _ {v i} ^ {\\text {r a w}} \\right|\\right) - \\left| \\nabla_ {S} f e a t _ {o u t} \\right| \\| _ {1}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.281 + ], + "angle": 0, + "content": "\\[\nL _ {L} = \\| O \\left(\\max \\left(| \\nabla_ {L} I _ {i r} ^ {\\text {r a w}} |, | \\nabla_ {L} I _ {v i} ^ {\\text {r a w}} |\\right)\\right) - \\nabla_ {L} f e a t _ {o u t} \\| _ {1}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.282, + 0.484, + 0.372 + ], + "angle": 0, + "content": "where \\(\\alpha_{2}\\) and \\(\\alpha_{3}\\) are the coefficients of the Sobel gradient loss and the Laplace gradient loss, respectively. \\(|\\cdot |\\) refers to the absolute operation. \\(\\nabla_S\\) and \\(\\nabla_L\\) respectively refer to the Sobel and Laplacian gradient operator. \\(O(|x|) = x\\) represents the initial gradient value retrieved before the absolute value process is executed." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.386, + 0.224, + 0.403 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.41, + 0.174, + 0.426 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.627 + ], + "angle": 0, + "content": "Datasets. During the training stage, the MSRS [23] training set is utilized. For additional enhancement of the training set, 200 pairs are randomly chosen from the RoadScene [33] dataset as a supplement. During the evaluation stage, we use the MSRS test set. Furthermore, to assess the generalization ability of our method, we conduct tests on the \\(\\mathrm{M}^3\\mathrm{FD}\\) [16] and TNO [25] datasets no any re-training or fine-tuning. It's important to recognize that due to the high resolution of images in the \\(\\mathrm{M}^3\\mathrm{FD}\\) dataset, some contrast methods have extremely prolonged runtimes. To expedite testing, the images are resized to half their initial resolution, and the precision for all methods on \\(\\mathrm{M}^3\\mathrm{FD}\\) is reported based on the resized resolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Implement Details. The hardware setup includes an NVIDIA TITAN RTX and a 3.6 GHz Intel Core i9-9900K, with the code implemented in PyTorch. The Adam optimizer, without weight decay, are employed to optimize the network parameters. The training phase consists of 60 epochs, a batch size of 4, an initial learning rate of \\(10^{-4}\\), and the learning rate is halved every 25 epochs. In the training phase, input images undergo random cropping into 128 \\(\\times\\) 128-sized segments and are transformed from RGB format to Y, Cb, and Cr. The coefficients for the terms in loss are configured as: \\(\\alpha_{1} = 0.1\\), \\(\\alpha_{2} = 1\\), and \\(\\alpha_{3} = 10\\), which are based on the experience from our previous work. The channels at various stages is respectively defined as \\(ch\\_in = 1\\), \\(ch\\_mid = 48\\), and \\(ch\\_out = 1\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.383, + 0.866 + ], + "angle": 0, + "content": "4.2. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Competing Methods. To validate the effectiveness of our method, we compare it with seven state-of-the-art works:" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.145 + ], + "angle": 0, + "content": "Table 1. Quantitative results of our method compared to seven state-of-the-art methods on the MSRS, \\(\\mathbf{M}^3\\mathrm{FD}\\), and TNO datasets. Bold indicates the best result, and underlining represents the second best." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.156, + 0.905, + 0.419 + ], + "angle": 0, + "content": "
MSRS
PublisherVIF ↑SCD ↑MI ↑Qabf ↑SSIM ↑MS_SSIM ↑FMI ↑
LRRNet [13]23'TPAMI0.5440.7912.9450.4570.8590.8480.919
CoCoNet [17]24'IJCV0.6601.3872.2190.2790.7330.7470.884
MetaFusion [40]23'CVPR0.7261.4861.6840.4780.9120.9040.908
TUFusion [41]24'TCSVT0.5691.2982.5750.2290.8850.8980.909
DATFuse [24]23'TCSVT0.9061.4103.8970.6400.9550.9560.912
SwinFusion [20]22'JAS0.9901.6904.5490.6540.9680.9690.931
CDDFuse [42]23'CVPR1.0511.6215.0150.6930.9670.9660.930
Ours25'ICCV1.0611.7005.1110.7150.9810.9750.933
M²FD
PublisherVIF ↑SCD ↑MI ↑Qabf ↑SSIM ↑MS_SSIM ↑FMI ↑
LRRNet [13]23'TPAMI0.4931.4602.8750.4660.8910.8700.862
CoCoNet [17]24'IJCV0.6241.7692.8130.3990.8750.8540.853
MetaFusion [40]23'CVPR0.7011.6652.5700.3870.8540.8610.850
TUFusion [41]24'TCSVT0.4001.5022.9650.1890.7880.8640.856
DATFuse [24]23'TCSVT0.5901.2804.0380.4890.8550.8310.849
SwinFusion [20]22'JAS0.6901.5744.0970.5830.9470.9030.872
CDDFuse [42]23'CVPR0.7201.6473.9990.5960.9510.9130.871
Ours25'ICCV0.7591.6603.7710.6540.9640.9230.883
TNO
PublisherVIF ↑SCD ↑MI ↑Qabf ↑SSIM ↑MS_SSIM ↑FMI ↑
LRRNet [13]23'TPAMI0.5821.5942.5930.3630.8700.8700.886
CoCoNet [17]24'IJCV0.6551.7172.1610.2970.8070.8180.871
MetaFusion [40]23'CVPR0.6611.7301.6520.2980.7880.8150.877
TUFusion [41]24'TCSVT0.4651.5992.0940.2240.8250.8670.893
DATFuse [24]23'TCSVT0.6901.4103.1920.5030.8490.8190.878
SwinFusion [20]22'JAS0.7431.6873.2730.5160.9370.8990.900
CDDFuse [42]23'CVPR0.7831.7703.0610.5170.9290.9080.899
Ours25'ICCV0.8361.7243.5080.5880.9320.9040.914
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.907, + 0.55 + ], + "angle": 0, + "content": "LRRNet [13], CoCoNet [17], MetaFusion [40], TUFusion [41], DATFuse [24], SwinFusion [20], and CDDFuse [42]. CoCoNet, LRRNet, and MetaFusion are primarily CNN-based methods, while SwinFusion is dominated by the SwinTransformer architecture. TUFusion, DATFuse, CDDFuse, and our method adopt a hybrid mode that combines CNN and attention mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.78 + ], + "angle": 0, + "content": "Evaluation Metrics. We conduct a comprehensive analysis from both qualitative and quantitative perspectives. The qualitative comparison mainly relies on human subjective evaluation of the visual effects of the fused images, primarily concentrating on the preservation of texture details and the retention of salient thermal radiation targets. The quantitative evaluation employs seven mainstream image quality assessment metrics: visual information fidelity (VIF), the sum of the correlations of differences (SCD), mutual information (MI), gradient-based similarity measurement \\((\\mathrm{Q_{abf}})\\), structural similarity index measure (SSIM), multiscale structural similarity (MS-SSIM), and feature mutual information (FMI). All selected metrics are positive, with higher values indicating better performance. Detailed explanations can be found in [19]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Qualitative Comparison. The qualitative comparison results are shown in Fig. 5, which comprises three groups of images arranged from top to bottom, based on the MSRS, \\(\\mathrm{M}^{3}\\mathrm{FD}\\), and TNO datasets, respectively. The first group of images demonstrates the ability of different methods to preserve thermal radiation information. DATFuse, LRRNet, and TUFusion fail to effectively retain thermal radiation targets, while MetaFusion, CoCoNet, and SwinFusion pro" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13518" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.087, + 0.262, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.178, + 0.205, + 0.187 + ], + "angle": 0, + "content": "Infrared" + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.088, + 0.416, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.178, + 0.359, + 0.187 + ], + "angle": 0, + "content": "Visible" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.088, + 0.569, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.178, + 0.517, + 0.187 + ], + "angle": 0, + "content": "DATFuse" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.088, + 0.726, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.178, + 0.674, + 0.187 + ], + "angle": 0, + "content": "CoCoNet" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.088, + 0.883, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.178, + 0.828, + 0.187 + ], + "angle": 0, + "content": "LRRNet" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.188, + 0.262, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.277, + 0.212, + 0.285 + ], + "angle": 0, + "content": "CDDFuse" + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.188, + 0.414, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.277, + 0.369, + 0.286 + ], + "angle": 0, + "content": "MetaFusion" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.188, + 0.569, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.277, + 0.522, + 0.286 + ], + "angle": 0, + "content": "SwinFusion" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.188, + 0.726, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.277, + 0.675, + 0.286 + ], + "angle": 0, + "content": "TUFusion" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.188, + 0.883, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.277, + 0.82, + 0.286 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.296, + 0.264, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.385, + 0.209, + 0.394 + ], + "angle": 0, + "content": "Infrared" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.296, + 0.416, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.385, + 0.359, + 0.394 + ], + "angle": 0, + "content": "Visible" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.296, + 0.57, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.474, + 0.385, + 0.518, + 0.394 + ], + "angle": 0, + "content": "DATFuse" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.296, + 0.726, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.385, + 0.675, + 0.394 + ], + "angle": 0, + "content": "CoCoNet" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.296, + 0.883, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.385, + 0.829, + 0.394 + ], + "angle": 0, + "content": "LRRNet" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.395, + 0.264, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.484, + 0.213, + 0.493 + ], + "angle": 0, + "content": "CDDFuse" + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.395, + 0.416, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.484, + 0.37, + 0.493 + ], + "angle": 0, + "content": "MetaFusion" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.395, + 0.57, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.484, + 0.524, + 0.493 + ], + "angle": 0, + "content": "SwinFusion" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.395, + 0.726, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.484, + 0.675, + 0.493 + ], + "angle": 0, + "content": "TUFusion" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.395, + 0.883, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.484, + 0.822, + 0.493 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.503, + 0.264, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.592, + 0.209, + 0.602 + ], + "angle": 0, + "content": "Infrared" + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.503, + 0.416, + 0.588 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.326, + 0.592, + 0.36, + 0.602 + ], + "angle": 0, + "content": "Visible" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.504, + 0.572, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.475, + 0.592, + 0.521, + 0.602 + ], + "angle": 0, + "content": "DATFuse" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.504, + 0.725, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.592, + 0.675, + 0.602 + ], + "angle": 0, + "content": "CoCoNet" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.504, + 0.883, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.592, + 0.828, + 0.602 + ], + "angle": 0, + "content": "LRRNet" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.604, + 0.263, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.692, + 0.213, + 0.702 + ], + "angle": 0, + "content": "CDDFuse" + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.604, + 0.416, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.692, + 0.37, + 0.702 + ], + "angle": 0, + "content": "MetaFusion" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.604, + 0.571, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.692, + 0.526, + 0.702 + ], + "angle": 0, + "content": "SwinFusion" + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.604, + 0.726, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.692, + 0.676, + 0.702 + ], + "angle": 0, + "content": "TUFusion" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.604, + 0.883, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.796, + 0.692, + 0.82, + 0.701 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.72, + 0.907, + 0.749 + ], + "angle": 0, + "content": "Figure 5. Visual comparison results of our method with seven state-of-the-art methods on typical image pairs from MSRS (the first group), M\\(^3\\)FD (the second group), and TNO (the third group) datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.486, + 0.896 + ], + "angle": 0, + "content": "duce distortions in the legs of the highlighted person. The second and third groups of images illustrate the capability of different methods to preserve details. For instance, neither the clouds in the second group nor the shrubberies in the third group are effectively preserved by DATFuse, LRR-Net, SwinFusion, and TUFusion, while CoCoNet and MetaFusion introduce significant noise. Only CDDFuse and our proposed method perform well in both aspects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.774, + 0.907, + 0.896 + ], + "angle": 0, + "content": "Quantitative Comparison. The quantitative comparison results are presented in Tab. 1. In the context of the MSRS dataset, our method ranks first across all evaluation metrics, demonstrating the following advantages. 1) The optimal VIF indicates that our technique yields fused outputs with enhanced visual quality, further confirming its qualitative superiority. 2) The ideal SCD, MI, and FMI suggest that our method is most effective in conveying data from the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13519" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.244 + ], + "angle": 0, + "content": "original input to the outcome. 3) The superior SSIM and MS-SSIM show that our model minimizes information loss and image distortion. 4) The perfect Qabf illustrates that our solution successfully preserves the maximum edge details. Furthermore, to verify the generalization competence of our method, we perform tests on the \\(\\mathbf{M}^3\\mathrm{FD}\\) and TNO datasets without any fine-tuning. In the context of the \\(\\mathbf{M}^3\\mathrm{FD}\\) dataset, our technique leads in every metric, with the sole exceptions being SCD and MI. Within the TNO dataset, it secures four top positions and two runner-up spots." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.254, + 0.255, + 0.269 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.483, + 0.442 + ], + "angle": 0, + "content": "Within this part, we undertake multiple ablations. Each one is executed using the MSRS test dataset as a foundation. Extension. To explore the correctness of extending the attention properties of infrared images to the visible modality, two sets of controlled experiments are performed: the first set utilize only I-SCA, while the second set employ both I-SCA and V-SCA. The experimental results are shown in Tab. 2. The integration of V-SCA led to enhancements in all measurements, with the sole exception of a minor reduction in SCD and SSIM, which confirms the validity of this extension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.483, + 0.61 + ], + "angle": 0, + "content": "I/V-SCA. To validate the effectiveness of I-SCA and V-SCA, we design two strategies in this subsection. 1) Directly removing the proposed modules, which is denoted as w/o I/V-SCA. 2) To prevent inconsistencies in model complexity before and after the removal of I/V-SCA, which could introduce additional experimental bias, we further replace I/V-SCA with Restormer Blocks. This ablation is denoted as \"to Restormer\". Findings from the experiment are detailed in Tab. 3, which illustrates that our technique holds the top position in five of the seven assessment measures, affirming the efficacy of I/V-SCA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.483, + 0.732 + ], + "angle": 0, + "content": "CBSM. For confirming CBSM's efficiency, we substitute CBSM with Parameters, \\(Conv_{1 \\times 1}\\), and \\(Conv_{3 \\times 3}\\), respectively. The experiment's findings are outlined in Tab. 4. Compared to the other three schemes, the introduction of CBSM results in more than half of the evaluated metrics being ranked first, which suggests that CBSM is capable of efficiently diminishing the disparities between the source image and the intermediate features of another modality." + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.745, + 0.46, + 0.759 + ], + "angle": 0, + "content": "Table 2. Ablation before and after the extension of I-SCA." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.77, + 0.482, + 0.822 + ], + "angle": 0, + "content": "
ConfigurationsVIFSCDMIQabfSSIMMS_SSIMFMI
I-SCAV-SCA
✓✗1.0391.7214.5710.7030.9820.9750.932
✓✓1.0611.7005.1110.7150.9810.9750.933
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.303, + 0.865 + ], + "angle": 0, + "content": "4.4. Efficiency Comparison" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Within this part, we evaluate the parameters, FLOPs, and FPS of our work with the most competitive methods" + }, + { + "type": "table_caption", + "bbox": [ + 0.552, + 0.09, + 0.868, + 0.104 + ], + "angle": 0, + "content": "Table 3. Ablation of the I-SCA and V-SCA modules." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.115, + 0.905, + 0.168 + ], + "angle": 0, + "content": "
ConfigurationsVIFSCDMIQabfSSIMMS_SSIMFMI
w/o I/V-SCA1.0651.6695.6510.7080.9680.9670.933
to Restormer1.0631.6815.4630.7120.9750.9710.933
Ours1.0611.7005.1110.7150.9810.9750.933
" + }, + { + "type": "table_caption", + "bbox": [ + 0.589, + 0.182, + 0.831, + 0.195 + ], + "angle": 0, + "content": "Table 4. Ablation of the CBSM module." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.207, + 0.905, + 0.27 + ], + "angle": 0, + "content": "
ConfigurationsVIFSCDMIQabfSSIMMS_SSIMFMI
Parameters1.0601.7085.3030.7120.9770.9730.932
Conv1×11.0261.6914.6710.6910.9750.9670.932
Conv3×31.0661.7055.2540.7110.9770.9730.932
CBSM1.0611.7005.1110.7150.9810.9750.933
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.907, + 0.388 + ], + "angle": 0, + "content": "to highlight its speed advantage. Findings are presented in Tab. 5, where the FLOPs are determined based on an input dimension of \\(128 \\times 128 \\times 1\\), and the FPS on \\(\\mathbf{M}^3\\mathrm{FD}\\) is measured with images of reduced resolution by half. Our method achieves the highest FPS, enabling seamless integration with both low- and high-level computer vision tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.581, + 0.413, + 0.838, + 0.427 + ], + "angle": 0, + "content": "Table 5. The parameters, FLOPs, and FPS." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.438, + 0.905, + 0.501 + ], + "angle": 0, + "content": "
Params (M) ↓FLOPs (G) ↓FPS ↑
MSRSM3FDTNO
SwinFusion [20]0.92717.9860.6241.0160.651
CDDFuse [42]1.18629.21339.72340.48040.179
Ours0.5659.207132.303137.271126.537
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.533, + 0.634, + 0.549 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.907, + 0.756 + ], + "angle": 0, + "content": "This paper reveals the inherent \"attention properties\" of infrared images from the Grad-CAM of ResNet-50. To incorporate the property into IVF, we propose I-SCA and extend it to the visible modality, subsequently introducing V-SCA. The auxiliary module (CBSM) further ensures fusion quality by enhancing feature channels, enabling spatial transformations, and suppressing negative information in the source images. Additionally, the linear computations of I-SCA and V-SCA guarantee efficiency, achieving FPS of 132, 137, and 127 on MSRS, \\(\\mathbf{M}^3\\mathrm{FD}\\), and TNO, respectively. Moving forward, our focus will shift to crafting more specialized modules and undertaking an in-depth methodological examination to enhance the refinement of our approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.673, + 0.787 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (62272426), the Shanxi Province Science and Technology Major Special Project (202201150401021), and the Fundamental Research Program of Shanxi Province (TZLH20230818005, 202303021212189, 202303021212372, 202403021212166)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13520" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.187, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Yaming Cao, Lei Guo, Fengguang Xiong, Liqun Kuang, and Xie Han. Physical-simulation-based dynamic template matching method for remote sensing small object detection. IEEE Transactions on Geoscience and Remote Sensing, 62: 1-14, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.258, + 0.483, + 0.312 + ], + "angle": 0, + "content": "[3] Zhihao Chang, Zhixi Feng, Shuyuan Yang, and Quanwei Gao. Aft: Adaptive fusion transformer for visible and infrared images. IEEE Transactions on Image Processing, 32: 2077-2092, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.315, + 0.483, + 0.369 + ], + "angle": 0, + "content": "[4] Jun Chen, Jianfeng Ding, and Jiayi Ma. Hitfusion: Infrared and visible image fusion for high-level vision tasks using transformer. IEEE Transactions on Multimedia, 26: 10145-10159, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.372, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[5] Chunyang Cheng, Tianyang Xu, and Xiao-Jun Wu. Mufusion: A general unsupervised image fusion network based on memory unit. Information Fusion, 92:80-92, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.459, + 0.483, + 0.527 + ], + "angle": 0, + "content": "[7] Keying Du, Huafeng Li, Yafei Zhang, and Zhengtao Yu. Chitnet: A complementary to harmonious information transfer network for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 74: 1-17, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.571 + ], + "angle": 0, + "content": "[8] Xueyan Gao and Shiguang Liu. Bcmfuse: A bilateral cross-modal feature interaction-based network for infrared and visible image fusion. Remote Sensing, 16(17):3136, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.631, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[10] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.483, + 0.728 + ], + "angle": 0, + "content": "[11] Shahid Karim, Geng Tong, Jinyang Li, Akeel Qadir, Umar Farooq, and Yiting Yu. Current advances and future perspectives of image fusion: A comprehensive review. Information Fusion, 90:185-217, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[12] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[13] Hui Li, Tianyang Xu, Xiao-Jun Wu, Jiwen Lu, and Josef Kittler. Lrrnet: A novel representation learning guided fusion network for infrared and visible images. IEEE transactions on pattern analysis and machine intelligence, 45(9):11040-11052, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In European Conference on Computer Vision, pages 719-735, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[16] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5811, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.263, + 0.907, + 0.331 + ], + "angle": 0, + "content": "[17] Jinyuan Liu, Runjia Lin, Guanyao Wu, Risheng Liu, Zhongxuan Luo, and Xin Fan. Coconet: Coupled contrastive learning network with multi-level feature ensemble for multi-modality image fusion. International Journal of Computer Vision, 132(5):1748-1775, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[19] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information fusion, 45:153-178, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.515 + ], + "angle": 0, + "content": "[20] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swim transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200-1217, 2022. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.6 + ], + "angle": 0, + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.603, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[22] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), page 618-626. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[23] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Information Fusion, 83:79-92, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[24] Wei Tang, Fazhi He, Yu Liu, Yansong Duan, and Tongzhen Si. Datfuse: Infrared and visible image fusion via dual attention transformer. IEEE Transactions on Circuits and Systems for Video Technology, 33(7):3159-3172, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.801, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[25] Alexander Toet. The tno multiband image data collection. Data in brief, 15:249-251, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[26] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[27] Boying Wang, Libo Zhang, Longyin Wen, Xianglong Liu, and Yanjun Wu. Towards real-world prohibited item detection: A large-scale x-ray benchmark. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13521" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF international conference on computer vision, pages 5412-5421, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[28] Boying Wang, Ruyi Ji, Libo Zhang, Yanjun Wu, and Jing Liu. Learning to zoom: Exploiting mixed-scale contextual information for object detection. Expert Systems with Applications, 264:125871, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[29] Zhishe Wang, Yanlin Chen, Wenyu Shao, Hui Li, and Lei Zhang. Swinfuse: A residual swim transformer fusion network for infrared and visible images. IEEE Transactions on Instrumentation and Measurement, 71:1-12, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[30] Zhishe Wang, Wenyu Shao, Yanlin Chen, Jiawei Xu, and Xiaoqin Zhang. Infrared and visible image fusion via interactive compensatory attention adversarial learning. IEEE Transactions on Multimedia, 25:7800-7813, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[31] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.345, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[32] Yongbiao Xiao, Hui Li, Chunyang Cheng, and Xiaoning Song. Le2fusion: A novel local edge enhancement module for infrared and visible image fusion. In International conference on image and graphics, pages 293-305. Springer, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.483, + 0.469 + ], + "angle": 0, + "content": "[33] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.483, + 0.539 + ], + "angle": 0, + "content": "[34] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Text-if: Leveraging semantic text guidance for degradation-aware and interactive image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27026-27035, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.54, + 0.483, + 0.595 + ], + "angle": 0, + "content": "[35] Jun Yue, Leyuan Fang, Shaobo Xia, Yue Deng, and Jiayi Ma. Dif-fusion: Toward high color fidelity in infrared and visible image fusion with diffusion models. IEEE Transactions on Image Processing, 32:5705-5720, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[36] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[37] Hao Zhang, Xuhui Zuo, Jie Jiang, Chunchao Guo, and Jiayi Ma. Mrfs: Mutually reinforcing image fusion and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26974-26983, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[38] Cheng Zhao, Peng Yang, Feng Zhou, Guanghui Yue, Shuigen Wang, Huisi Wu, Guoliang Chen, Tianfu Wang, and Baiying Lei. Mhw-gan: multidiscriminator hierarchical wavelet generative adversarial network for multimodal image fusion. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[39] Rong Zhao, Xie Han, Xindong Guo, Liquun Kuang, Xiaowen Yang, and Fusheng Sun. Exploring the point feature relation on point cloud for multi-view stereo. IEEE Transactions on Circuits and Systems for Video Technology, 33(11): 6747-6763, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[40] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13955-13965, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[41] Yangyang Zhao, Qingchun Zheng, Peihao Zhu, Xu Zhang, and Wenpeng Ma. Tufusion: A transformer-based universal fusion algorithm for multimodal images. IEEE Transactions on Circuits and Systems for Video Technology, 34(3): 1712-1725, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.908, + 0.317 + ], + "angle": 0, + "content": "[42] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5906-5916, 2023. 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.908, + 0.401 + ], + "angle": 0, + "content": "[43] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: denoising diffusion model for multi-modality image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8082-8093, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.402, + 0.907, + 0.485 + ], + "angle": 0, + "content": "[44] Zixiang Zhao, Lilun Deng, Haowen Bai, Yukun Cui, Zhipeng Zhang, Yulun Zhang, Haotong Qin, Dongdong Chen, Jiangshe Zhang, Peng Wang, et al. Image fusion via vision-language model. In Proceedings of the 41st International Conference on Machine Learning, pages 60749-60765, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.487, + 0.907, + 0.543 + ], + "angle": 0, + "content": "[45] Pengfei Zhu, Yang Sun, Bing Cao, and Qinghua Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099-7108, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13522" + } + ] +] \ No newline at end of file diff --git a/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/cab6b62f-c3b3-45e2-8877-ae8b2379307d_origin.pdf b/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/cab6b62f-c3b3-45e2-8877-ae8b2379307d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..612d4a4d6dbd3fdca448b60903c67b0196178342 --- /dev/null +++ b/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/cab6b62f-c3b3-45e2-8877-ae8b2379307d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8aa565f0a15cf320bfd25afb78a44b348a9123e22a4d711d53bd4aa45ff9aab3 +size 15714203 diff --git a/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/full.md b/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ab5f0d20a18088d4a37f9650d1ea1a81a6ba1126 --- /dev/null +++ b/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/full.md @@ -0,0 +1,379 @@ +# The Source Image is the Best Attention for Infrared and Visible Image Fusion + +Song Wang Xie Han Liquun Kuang\* Boying Wang Zhongyu Chen Zherui Qiao Fan Yang Xiaoxia Liu Bingyu Zhang Zhixun Wang + +School of Computer Science and Technology, North University of China, Taiyuan 030051, China + +Shanxi Key Laboratory of Machine Vision & Virtual Reality, Taiyuan 030051, China + +Shanxi Vision Information Processing and Intelligent Robot Engineering Research Center, Taiyuan 030051, China + +b20230704@st.nuc.edu.cn,kuang@nuc.edu.cn + +# Abstract + +Infrared and visible image fusion (IVF) endeavors to engineer composite outputs by blending optimal virtues of divergent modalities. This paper reveals, unprecedentedly, the intrinsic "attention properties" of infrared images, which directly arise from their physical characteristics (i.e., heat distribution) and can be linked to attention mechanisms naturally, as observed in the gradient-weighted class activation mapping (Grad-CAM) visualization analysis of image classification models. To incorporate this property into IVF for better fusion, we propose the source infrared cross attention (I-SCA) and further extend it to the visible modality, subsequently introducing the source visible cross attention (V-SCA). The joint use of I-SCA and V-SCA greatly alleviate longstanding issues, such as insufficient and incomplete multimodal feature interaction and fusion, in IVF. Moreover, an auxiliary component for I-SCA and V-SCA, termed CBSM, is employed to boost the channel, map space, and suppress redundancy and misleading information of the source images. Specifically, we directly treat the CBSM-processed raw image as the query, while the intermediate features of another modality are treated as keys and values in I-SCA and V-SCA. Unlike attention mechanisms that divide images into patches or limit computations to local windows, our cross attention modules achieve smoother and more robust IVF through true global modeling across the entire image space with linear complexity. Comparison with current SOTA methods on three popular public datasets confirms its superiority. + +# 1. Introduction + +IVF, a bedrock of image processing, with the objective of engendering a composite that encapsulates the considerable thermal radiation from infrared and the intricacies de + +![](images/cf6a6a3a69c705248c33627af0920a354fa0205b3e693fab823f005c4e66d760.jpg) +(a) + +![](images/42d9e79675dcb052c7be455df5e9004fad7d2174bd45e51c9827735cba04ff01.jpg) + +![](images/60f756b556ad51bd892a8590a1c3e0e75c420a167306e77564c381641e919d87.jpg) +(c) + +![](images/94df5b43c0a9fbd1255f09420f8b5c323a4452d5b3fbecebe99263e0fc68ff88.jpg) +(b) + +![](images/b7e700277990f92bce909bb11eedcfae37e9a0ef4d900f865c63a74fcef4cfba.jpg) +(e) + +![](images/41fffc37ad8b5d59409848e620541e1864e5188e612d5bdf874955d285424cc9.jpg) +(d) +(f) +Figure 1. (a, b) Grad-CAM visualization results for two images (00022004, 00028021), both labeled as "dog", predicted by the final layer of ResNet-50 on the ImageNet validation set. The color gradient from red to blue indicates the decreasing level of attention paid by the model to different regions. (c) An infrared image example (00519D) consists of highlighted thermal radiation regions and remaining ordinary regions from the MSRS dataset. (d) Visible image of the same scene corresponding to subplot (c). (e) Subplot (c) is displayed by applying the Jet colormap. (f) The negative transformation of subplot (c) is visualized with the Jet colormap. + +tail from visible. Prior to the surge in deep learning, this process was typically achieved by traditional algorithms in spatial and transform domains [11]. However, thanks to the development of neural networks and attention mechanisms, heuristic algorithms have become more widely adopted. + +In this paper, we draw inspiration from the Grad-CAM + +[22] visualization results of classic image classification models, such as ResNet-50 [9], and discover the inherent "attention properties" of infrared images. These Grad-CAM visualizations, as shown in Fig. 1(a) and Fig. 1(b), reveal how the model assigns varying weights to different regions during the prediction process: regions marked in red represent "important areas" that exert a substantial impact on the classification result, while regions marked in blue correspond to "general areas" with lesser contribution. Interestingly, we observe that the structural characteristics of infrared images (e.g., as shown in Fig. 1(c)) are strikingly similar to the Grad-CAM images, as both consist of important areas and general areas (in infrared images, these regions correspond to the highlighted areas with higher thermal energy and the darker zones with lower energy). To more intuitively showcase this similarity, we re-map the infrared image with the Jet colormap (see Fig. 1(e)). + +Based on these observations, we believe that infrared images inherently possess an attention property similar to the one exhibited by Grad-CAM. Unlike image classification models, which require be trained to learn attention patterns, the attention properties of infrared images directly stem from their physical characteristics—the heat distribution. Moreover, this attribute is not only more specific and accurate but also avoids issues such as insufficient or misdirected attention that may occur in image classification models. For instance, when ResNet-50 processes the dog in Fig. 1(a), it focuses only on the dog's eyes, and when processing the dog in Fig. 1(b), it not only focuses on the dog itself but also incorrectly attends to irrelevant backgrounds. In contrast, infrared images provide precise localization of heat sources, ensuring that all attention is concentrated on specific areas. + +Based on the above finding, an immediate question arises: how can we fully leverage the inherent attention properties of infrared images to obtain better-fused images in IVF? Intuitively, we think that it can be linked to the attention mechanism [26], as both essentially represent the importance of different regions. Furthermore, the interaction of cross-modal features plays a pivotal role in image fusion. Therefore, our goal is to introduce the attention properties at this stage to counteract the persistent obstacles in image fusion, such as inadequate cross-domain information exchange and incomplete multimodal feature fusion, thereby enhancing the caliber of the merged images. + +Specifically, we propose the source infrared cross attention (I-SCA) based on a modified Restormer [36], which can perform global feature modeling in the image space with linear complexity. I-SCA first takes the original infrared image and the intermediate features of the visible image as inputs. It then treats the infrared image as the query (Q), thanks to the properties of infrared images, which allow them to be directly used as queries without processing, + +while the visible features serve as the keys (K) and values (V), enabling an infrared heat prior-guided cross-domain information exchange. However, this approach can only prioritize on "important areas", whereas image fusion is a global pixel-wise task that demands equal treatment for all areas. Fortunately, a facile solution is to treat the negative transformation of the infrared image (as shown in Fig. 1(f)) as another query to allow significant attention complementarity. + +Nevertheless, using I-SCA only also faces a critical problem, i.e., it can only capture "infrared-guided" cross-modal interaction features. Theoretically, the optimal fused image could be obtained by simultaneously fusing features dominated by both infrared and visible images. Therefore, we make a bold inference: we extend the discovery to visible and, based on this, propose the source visible cross attention (V-SCA). While visible images lack explicit physical saliency, their visually significant regions (e.g., edges, textures, and contrast) can still serve as effective priors. + +Finally, since both I-SCA and V-SCA require inputs to have the same channels, but the channels of the source image and the intermediate representation from the other modality are not consistent, we deploy an adaptive channel boosting and intelligent space mapping module (CBSM), which also serves crucial roles in transforming the feature space of the source image from its initial state to one more suitable for attention mechanisms to reduce discrepancies between features at different stages and suppressing redundancy or misleading content contained in the initial images. To validate the effectiveness of I-SCA, V-SCA, and CBSM, extensive analysis are conducted on three popular public datasets. Experiments demonstrate that both qualitative and quantitative results outperform state-of-the-art (SOTA) methods. + +The essence of our contributions is as follows: + +- We observe a strong structural correlation between infrared images and Grad-CAM, leading us to hypothesize, for the first time, that infrared images possess an inherently exceptionally precise attention property. +- Based on this attention property, we propose a heat-distribution-prior-guided cross-modal attention mechanism, which can be further extend to the visible modality. Additionally, an auxiliary module, CBSM, is deployed to complement the proposed attention mechanism for further enhancing fusion quality. +- Our findings inspire a novel paradigm guided by source image priors for image fusion, significantly alleviating challenges in cross-domain feature interaction and fusion faced by current methods. +- Comprehensive experiments instituted on three prominent available datasets with SOTA works demonstrate the superiority. + +# 2. Related Works + +# 2.1. Attention-based Fusion Methods + +The triumph of transformer [26] in NLP has sparked considerable interest across the academic community. To establish a foundation for the integration of this technology in computer vision, Alexey et al. proposed Vision Transformer (ViT) [6], which segments an image into multiple units, then flattens the units into a shape compatible with text sequences, allowing direct application of the original transformer for long-range modeling. Subsequently, works such as TUFusion [41], DATfuse [24], HitFusion [4], AFT-Fusion [3], etc., extended ViT to the image fusion task. On the one hand, TUFusion adopted an encoder-decoder pipeline, featuring a dual-branch structure that merged parallel DenseBlock [12] and ViT in the encoder. DATFuse, on the other hand, incorporated channel and spatial attention mechanisms similar to CBAM [31] before the self-attention mechanism and performed fusion at the image level. Although transformer-based methods have shown great effectiveness, they are consistently constrained by the computational burden of the self-attention in image space. To remedy this issue, Liu et al. proposed SwinTransformer [18], which divides images into patches and further splits the patches into multiple windows, restricting self-attention to each window. Later, SwinFuse [29] used a full SwinTransformer approach to obtain fused images. SwinFusion [20] further extended this by introducing a cross-modal SwinTransformer, which generalizes window attention to cross-modal feature fusion. However, the practice of restricting attention to local regions contradicts the fundamental idea of attention mechanisms, which is to model global information. To achieve true overall modeling linearly, Zamir et al. proposed Restormer, which introduces multi-dconv head transposed attention (MDTA) and gated dconv feedforward network (GDFN) to replace the self-attention and feed-forward network [36]. Later, CDDFuse [42] further validated the effectiveness of the Restormer in image fusion. However, it only applied the Restormer to a single modality and overlooked the potential for feature interactions across multiple modalities. This paper narrowed down the challenge of interaction between different modalities by proposing I-SCA and V-SCA, and further enabled Restormer to be used for the multimodal information interaction problem. + +# 2.2. CNN-based Fusion Methods + +Currently, the realm of deep learning-driven fusion is segmented into various types, encompassing techniques reliant on autoencoders, end-to-end models, unified models, amalgamations with downstream or other specialized tasks, and approaches utilizing generative adversarial networks (GAN). Autoencoder-based methods [14, 32] first + +train an encoder-decoder pair on datasets like COCO [15] and then directly generalize them for image fusion. However, the performance of such methods is markedly hampered by the quality of the fusion rules designed subjectively, and developing a fusion strategy that works well with the encoder-decoder framework is not trivial. Consequently, a series of end-to-end methods [8, 13] were later proposed. These methods train all modules, including the feature fusion layer, with learning strategies, which greatly reduces the challenges related to constructing fusion strategies in autoencoder-based methods. Furthermore, as image fusion has evolved, the community has become dissatisfied with using a single model to solve just one specific fusion task, leading to unified models [5, 17, 45] for multiple tasks. However, these methods adopt a re-training strategy for each new task, rather than using a single weight parameter to handle multiple tasks simultaneously. Meanwhile, along with the development of downstream tasks such as object detection [2, 27, 28], scene reconstruction [39], etc., a major trend has been to combine image fusion with these computer vision sub-tasks, facilitating the mutual promotion of both [37, 40]. Another category of methods [34, 44] leverages large language models such as GPT [1] and CLIP [21] to guide the fusion process, achieving cross-domain integration for different tasks. Finally, GAN-based methods [30, 38] transform the fusion into an adversarial problem. Through continuous antagonistic training, the generator's output progressively converges toward the target distribution. Nevertheless, these techniques face hurdles like mode collapse, erratic training, and a lack of clear understanding. Later works proposed denoising diffusion model-based methods [35, 43] to address these issues. Although the core innovation of this study lies in source images prior-guided cross-modal attention, approximately half of the components are still based on the convolutional neural network (CNN). To ensure a comprehensive evaluation, we undertake a comparative analysis of it with several recent CNN-based SOTA methods discussed above. + +# 3. Method + +# 3.1. Overall Architecture + +In this work, a hybrid of CNN and attention mechanisms is employed to achieve the IVF. The overall architecture is shown in Fig. 2. Specifically, we construct a serial pipeline, consisting of SE-ResNet Block [10], Restormer Block [36], and I-SCA/V-SCA for infrared and visible branches. For each modality, we first employ the SE-ResNet Block (the SE acts as a gating mechanism to suppress information redundancy by refine features) for initial feature extraction while increasing the feature channel from ch_in to ch_mid. The Restormer Block then performs further deep extraction with another key function of transforming the feature space + +![](images/64735d6cc1ae6d0f6815b6935e058903fd7407df15ddb76dc1ca766692751859.jpg) +Figure 2. The overall architecture of our method, which incorporates SE-ResNet Block, Restormer Block, I-SCA/V-SCA, and CBSM. ir, vi denote infrared and visible images, and $ir\_n$ , $vi\_n$ signify negative transformations of infrared and visible images. ch_in, ch_mid, and ch_out represent input, middle, and output channels, respectively. “C” indicates a channel-wise concatenation operation. “Q” stands for the query in the attention mechanism, while “K” and “V” symbolize the key and value, respectively. + +from the domain where convolutional operations excel to the one where attention is more effective, thereby facilitating the subsequent cross-modal feature interaction. The I-SCA/V-SCA subsequently takes the features of the current modality along with either the image of another modality processed by CBSM or the negative transformation of that image for feature interaction. Post processing via I-SCA/V-SCA, a quartet of feature maps, steered by the originating image, is acquired. Then they will merge across the channel dimension and be processed through an additional SE-ResNet Block before being output. + +# 3.2. CBSM + +![](images/6d87b58cc64b314abb4deff6f780c7e4b6301f6c4e3af48ebc832590aeca096f.jpg) +Figure 3. The composition of adaptive channel boosting and intelligent space mapping module (CBSM). The SE is employed to suppress redundancy and misleading information in the source images. $\odot$ denotes a channel-wise multiplication operation. + +Since attention calculation in I-SCA and V-SCA requires all channels of inputs to be consistent, it is imperative to augment the channel dimension of the source image + +from ch_in to ch_mid. In general, this can be achieved by Parameters, $Conv_{1\times 1}$ , or $Conv_{3\times 3}$ . However, all three have a common problem: after channel expansion, the multi-channel feature space still remains in the initial state. In cross-modal attention, directly using feature values from the raw space for interaction with deep features will significantly degrade the information exchange effect due to spatial discrepancy. Therefore, a key issue is how to reduce the gap between them. Additionally, the source image may introduce redundancy or misleading cues in certain cases, making it crucial to mitigate potential negative effects. To address this, CMSM module, which consists of convolutional layers (for channel boosting), activation layers (for spatial mapping), and Squeeze-and-Excitation (SE) [10] layer (for suppressing negative influences) is employed, as shown in Fig. 3. The structure of CMSM is intentionally simple, because it aligns with our goal of using the raw image priors to guide feature interactions. A more complex design could significantly alter the initial pattern of the source image, conflicting with our motivation. For the input source image, denoted as $I_m^{raw} \in \mathbb{R}^{H \times W \times 1}$ , this module performs the following operations: + +$$ +I _ {m} ^ {t r s} = \alpha \left(S E \left(C o n v _ {3 \times 3} \left(\alpha \left(C o n v _ {3 \times 3} \left(I _ {m} ^ {r a w}\right)\right)\right)\right)\right). (1) +$$ + +The result of Eq. (1), denoted as $I_{m}^{trs} \in \mathbb{R}^{H \times W \times 48}$ , represents the image after channel boosting and space transformation. The variable $m$ represents different input modalities, and desirable values include $ir$ , $vi$ , $ir\_n$ , and $vi\_n$ . $ir\_n$ and $vi\_n$ denote the negative transforms of the infrared and the visible image, respectively. $\alpha(\cdot)$ is the parametric + +rectified linear unit (PReLU). + +# 3.3. Cross Attention Based on Source Images + +![](images/c150dac778b77550090f62882d409b099f4d9be16fc252f26c34f8425f17b5fc.jpg) + +![](images/979bb7d27141db02e313efd5ad12869d49efee98302463e8265147179aca27e1.jpg) +Figure 4. Our proposed I-SCA and V-SCA, both of which consist of two modules: I-DCA/V-DCA and GDFN. Structurally, the two are identical, differing only in their input content. + +The introduction of I-SCA and V-SCA aims to reduce the difficulty of multimodal feature interactions. The structures of these two components are illustrated in Fig. 4. In fact, both share an identical structural composition, differing only in the content of their inputs. Although the symmetrical design is used between I-SCA and V-SCA, this does not imply an inability to reflect modal heterogeneity. On the contrary, I-SCA emphasizes the heat distribution, contours, and structural priors of infrared images, while V-SCA focuses on edges, textures, and high-contrast guidance. Specifically, I-SCA comprises the dual-dconv infrared cross attention (I-DCA) and the gated dconv feedforward network (GDFN). Its inputs include $I_{ir}^{trs}$ (or $I_{ir,n}^{trs}$ ) and the visible features $feat_{vi}^{mid}$ , which are processed by the Restormer Block in the visible branch. In contrast, V-SCA is composed of the dual-dconv visible cross attention (V-DCA) and the GDFN, with inputs $I_{vi}^{trs}$ (or $I_{vi,n}^{trs}$ ) and the infrared features $feat_{ir}^{mid}$ . + +To illustrate the working mechanism of I-SCA and V-SCA, we demonstrate I-SCA, with inputs $I_{ir}^{trs}$ and $feat_{vi}^{mid}$ , as an example. In I-SCA, both $I_{ir}^{trs}$ and $feat_{vi}^{mid}$ first undergo the layer normalization (LN) in I-DCA. Following this, $I_{ir}^{trs}$ is directly used as the query (Q) after LN, while $feat_{vi}^{mid}$ is further processed to generate the key (K) and value (V). The process is described as follows: + +$$ +K, V = D \operatorname {c o n v} \left(C o n v _ {1 \times 1} \left(L N \left(f e a t _ {v i} ^ {m i d}\right)\right)\right), \tag {2} +$$ + +where $Dconv(\cdot)$ denotes the depthwise separable convolution, consisting of two steps: depthwise convolution and pointwise convolution. This operation preserves spatial dimensions and independently processes the channels at each spatial position. Subsequently, the Q $(I_{ir}^{trs})$ , K, and V are reshaped for compatibility and undergo the following operation: + +$$ +\begin{array}{l} f e a t _ {i r} ^ {o r i} = C o n v _ {1 \times 1} \left(V \cdot S o f t m a x \left(\left(I _ {i r} ^ {t r s} \cdot K\right) \odot s\right)\right) \\ + f e a t _ {v i} ^ {m i d}, (3) \\ \end{array} +$$ + +resulting in infrared-oriented intermediate features $\text{feat}_{ir}^{\text{ori}}$ , where $\odot$ represents element-wise multiplication, and $s$ is a learnable parameter. These features are then processed through the GDFN module as follows: + +$$ +Y _ {1}, Y _ {2} = D \operatorname {c o n v} \left(\operatorname {C o n v} _ {1 \times 1} \left(L N \left(f e a t _ {i r} ^ {\text {o r i}}\right)\right)\right), \tag {4} +$$ + +$$ +f e a t _ {i r} ^ {o r i} ^ {\prime} = C o n v _ {1 \times 1} \left(G E L U \left(Y _ {1}\right) \odot Y _ {2}\right) + f e a t _ {i r} ^ {o r i}, (5) +$$ + +where GELU refers to the Gaussian Error Linear Unit. + +After the above process, four source images-oriented features, which are further refined by GDFN, denoted as $\text{feat}_m^{\text{ori}'}$ , are obtained. These features, corresponding to different modalities ( $m = ir$ , $vi$ , $ir\_n$ , and $vi\_n$ ), are arranged along the axis of the channel and further refined by another SE-ResNet Block. The transformation of selfattention from the spatial dimension to the channel dimension, implemented by $Dconv(\cdot)$ and $Conv_{1\times 1}(\cdot)$ , ensures the linear computational characteristic of I-SCA and V-SCA. Additionally, the gating mechanism within the GDFN effectively suppresses redundant features, allowing only beneficial features to propagate through the module. + +# 3.4. Loss Function + +Our approach utilizes intensity $(L_{int})$ and gradient $(L_{grad})$ loss techniques to guarantee comprehensive and detailed data for the combined image. The complete loss is characterized as: + +$$ +L _ {\text {t o t a l}} = \alpha_ {1} L _ {\text {i n t}} + L _ {\text {g r a d}}, \tag {6} +$$ + +where $\alpha_{1}$ represents the intensity loss factor. The specific definition of intensity loss is as follows: + +$$ +L _ {i n t} = \frac {1}{H W} \| \max \left(I _ {i r} ^ {r a w}, I _ {v i} ^ {r a w}\right) - f e a t _ {o u t} \| _ {1}, \tag {7} +$$ + +where $\max(\cdot)$ represents the element-wise maximum operation, $\text{feat}_{\text{out}}$ signifies the final output of our model, and $\| \cdot \|_1$ denotes the $l_1$ -norm. Inspired by [42] and [7], the gradient term combines Sobel for edge preservation and Laplacian for noise suppression, guiding the model to capture + +valuable information while avoiding distractions. The formula is defined as follows: + +$$ +L _ {g r a d} = \alpha_ {2} L _ {S} + \alpha_ {3} L _ {L}, \tag {8} +$$ + +$$ +L _ {S} = \frac {1}{H W} \| \max \left(\left| \nabla_ {S} I _ {i r} ^ {\text {r a w}} \right|, \left| \nabla_ {S} I _ {v i} ^ {\text {r a w}} \right|\right) - \left| \nabla_ {S} f e a t _ {o u t} \right| \| _ {1}, \tag {9} +$$ + +$$ +L _ {L} = \| O \left(\max \left(| \nabla_ {L} I _ {i r} ^ {\text {r a w}} |, | \nabla_ {L} I _ {v i} ^ {\text {r a w}} |\right)\right) - \nabla_ {L} f e a t _ {o u t} \| _ {1}, \tag {10} +$$ + +where $\alpha_{2}$ and $\alpha_{3}$ are the coefficients of the Sobel gradient loss and the Laplace gradient loss, respectively. $|\cdot |$ refers to the absolute operation. $\nabla_S$ and $\nabla_L$ respectively refer to the Sobel and Laplacian gradient operator. $O(|x|) = x$ represents the initial gradient value retrieved before the absolute value process is executed. + +# 4. Experiments + +# 4.1. Setup + +Datasets. During the training stage, the MSRS [23] training set is utilized. For additional enhancement of the training set, 200 pairs are randomly chosen from the RoadScene [33] dataset as a supplement. During the evaluation stage, we use the MSRS test set. Furthermore, to assess the generalization ability of our method, we conduct tests on the $\mathrm{M}^3\mathrm{FD}$ [16] and TNO [25] datasets no any re-training or fine-tuning. It's important to recognize that due to the high resolution of images in the $\mathrm{M}^3\mathrm{FD}$ dataset, some contrast methods have extremely prolonged runtimes. To expedite testing, the images are resized to half their initial resolution, and the precision for all methods on $\mathrm{M}^3\mathrm{FD}$ is reported based on the resized resolution. + +Implement Details. The hardware setup includes an NVIDIA TITAN RTX and a 3.6 GHz Intel Core i9-9900K, with the code implemented in PyTorch. The Adam optimizer, without weight decay, are employed to optimize the network parameters. The training phase consists of 60 epochs, a batch size of 4, an initial learning rate of $10^{-4}$ , and the learning rate is halved every 25 epochs. In the training phase, input images undergo random cropping into 128 $\times$ 128-sized segments and are transformed from RGB format to Y, Cb, and Cr. The coefficients for the terms in loss are configured as: $\alpha_{1} = 0.1$ , $\alpha_{2} = 1$ , and $\alpha_{3} = 10$ , which are based on the experience from our previous work. The channels at various stages is respectively defined as $ch\_in = 1$ , $ch\_mid = 48$ , and $ch\_out = 1$ . + +# 4.2. Comparison with SOTA Methods + +Competing Methods. To validate the effectiveness of our method, we compare it with seven state-of-the-art works: + +Table 1. Quantitative results of our method compared to seven state-of-the-art methods on the MSRS, $\mathbf{M}^3\mathrm{FD}$ , and TNO datasets. Bold indicates the best result, and underlining represents the second best. + +
MSRS
PublisherVIF ↑SCD ↑MI ↑Qabf ↑SSIM ↑MS_SSIM ↑FMI ↑
LRRNet [13]23'TPAMI0.5440.7912.9450.4570.8590.8480.919
CoCoNet [17]24'IJCV0.6601.3872.2190.2790.7330.7470.884
MetaFusion [40]23'CVPR0.7261.4861.6840.4780.9120.9040.908
TUFusion [41]24'TCSVT0.5691.2982.5750.2290.8850.8980.909
DATFuse [24]23'TCSVT0.9061.4103.8970.6400.9550.9560.912
SwinFusion [20]22'JAS0.9901.6904.5490.6540.9680.9690.931
CDDFuse [42]23'CVPR1.0511.6215.0150.6930.9670.9660.930
Ours25'ICCV1.0611.7005.1110.7150.9810.9750.933
M²FD
PublisherVIF ↑SCD ↑MI ↑Qabf ↑SSIM ↑MS_SSIM ↑FMI ↑
LRRNet [13]23'TPAMI0.4931.4602.8750.4660.8910.8700.862
CoCoNet [17]24'IJCV0.6241.7692.8130.3990.8750.8540.853
MetaFusion [40]23'CVPR0.7011.6652.5700.3870.8540.8610.850
TUFusion [41]24'TCSVT0.4001.5022.9650.1890.7880.8640.856
DATFuse [24]23'TCSVT0.5901.2804.0380.4890.8550.8310.849
SwinFusion [20]22'JAS0.6901.5744.0970.5830.9470.9030.872
CDDFuse [42]23'CVPR0.7201.6473.9990.5960.9510.9130.871
Ours25'ICCV0.7591.6603.7710.6540.9640.9230.883
TNO
PublisherVIF ↑SCD ↑MI ↑Qabf ↑SSIM ↑MS_SSIM ↑FMI ↑
LRRNet [13]23'TPAMI0.5821.5942.5930.3630.8700.8700.886
CoCoNet [17]24'IJCV0.6551.7172.1610.2970.8070.8180.871
MetaFusion [40]23'CVPR0.6611.7301.6520.2980.7880.8150.877
TUFusion [41]24'TCSVT0.4651.5992.0940.2240.8250.8670.893
DATFuse [24]23'TCSVT0.6901.4103.1920.5030.8490.8190.878
SwinFusion [20]22'JAS0.7431.6873.2730.5160.9370.8990.900
CDDFuse [42]23'CVPR0.7831.7703.0610.5170.9290.9080.899
Ours25'ICCV0.8361.7243.5080.5880.9320.9040.914
+ +LRRNet [13], CoCoNet [17], MetaFusion [40], TUFusion [41], DATFuse [24], SwinFusion [20], and CDDFuse [42]. CoCoNet, LRRNet, and MetaFusion are primarily CNN-based methods, while SwinFusion is dominated by the SwinTransformer architecture. TUFusion, DATFuse, CDDFuse, and our method adopt a hybrid mode that combines CNN and attention mechanisms. + +Evaluation Metrics. We conduct a comprehensive analysis from both qualitative and quantitative perspectives. The qualitative comparison mainly relies on human subjective evaluation of the visual effects of the fused images, primarily concentrating on the preservation of texture details and the retention of salient thermal radiation targets. The quantitative evaluation employs seven mainstream image quality assessment metrics: visual information fidelity (VIF), the sum of the correlations of differences (SCD), mutual information (MI), gradient-based similarity measurement $(\mathrm{Q_{abf}})$ , structural similarity index measure (SSIM), multiscale structural similarity (MS-SSIM), and feature mutual information (FMI). All selected metrics are positive, with higher values indicating better performance. Detailed explanations can be found in [19]. + +Qualitative Comparison. The qualitative comparison results are shown in Fig. 5, which comprises three groups of images arranged from top to bottom, based on the MSRS, $\mathrm{M}^{3}\mathrm{FD}$ , and TNO datasets, respectively. The first group of images demonstrates the ability of different methods to preserve thermal radiation information. DATFuse, LRRNet, and TUFusion fail to effectively retain thermal radiation targets, while MetaFusion, CoCoNet, and SwinFusion pro + +![](images/effaa20c8c022ff21038661abb44550f753743e79850b925fe2c9e316bf01400.jpg) + +![](images/cc0d3a547bd58ee3dab0c2fe34ee53de1aaafd3b0b59c6e47484179aead6c830.jpg) + +![](images/40f396a4bc400b5aac6470028b527f72bda26d1576c342e7f0f73161d08c5f34.jpg) + +![](images/7a2109d5b234f714b70badb6b9820970cca00aa6c246dab2cff363aaac97d4c2.jpg) + +![](images/ce4c5b9602381a3ec87d6399d156a7b137fbd8cdb1d61646754403f914bf6adf.jpg) + +![](images/4ca3ed4cfd5276f720bc8775224568bce96b3892c27943d2ef99c829bad9b746.jpg) +Infrared +CDDFuse + +![](images/c1da1f1808a81971f07582c57b29eae74ead53665347c64a3787a8076bbd64e0.jpg) +Visible +MetaFusion + +![](images/ca1fa735f25c3e4120ab77cee5af368349b0101fb2af628bd75c8db61e86dc7a.jpg) +DATFuse +SwinFusion + +![](images/772ccb091d965bb76b9acee8e53585dae1bad537dd41b97ab0f6c9073a98651a.jpg) +CoCoNet +TUFusion + +![](images/7e6977d8b246d9e104eba81852fb031d76ddff48593c4b7df0cf39f8fe1da95f.jpg) +LRRNet +Ours + +![](images/60face0e156a3d77a5b30d98ff6c3352c7233a6caa1c0141e3fcb32f978fd70f.jpg) + +![](images/f37d0199dc6b631174df4457d2b28d77f4ae8a9fdcbf49ef41078f5271f5c3b5.jpg) + +![](images/f24e00e8fe6a55285de1652dba2a38fbdae25feaec925ac4cc2ba3b50fff96f5.jpg) + +![](images/8cf6ae43485f692d28aa462a811db6560e1070e36362fb2efc1a1e2cc7a708ca.jpg) + +![](images/482246027b02c2de08313fd27c1d679b894bcea2de85a718500b8712a9f322ff.jpg) + +![](images/1b94447b5ed31ea6703bb6ab11a636285a1a46da6996a1080d7923b568ed1a52.jpg) +Infrared +CDDFuse + +![](images/371b840f96f31b604d3ebbca952afc44c32691780218907f805653c969daf3da.jpg) +Visible +MetaFusion + +![](images/81f49eca2289bf05a458ae9c9d9cac1e874993d7ed1f59f0c8e71a37ef39c834.jpg) +DATFuse +SwinFusion + +![](images/458a400c635fabe7d1a2d7df1273b2cd2a4a753b77f5c5fbf4e430d67e32c433.jpg) +CoCoNet +TUFusion + +![](images/2adb9d10a2b425cfdf480b30b35382b9aabc50a5693e7d820603b154fe409b22.jpg) +LRRNet +Ours + +![](images/229cfc223ce944daafce7c5cb61952c67931e90423825728a8261ed3aae18f91.jpg) + +![](images/88c3b0fb1403bca2df640933c2e5a1fcc85cec58b08b226ce0a9295d71eb1e97.jpg) + +![](images/cf23d8aed191fb9492e734a4c6069bdbe6823b6045190e2c77c45d7baa2b8479.jpg) +DATFuse + +![](images/337668a7a80278204ae3ab722fd5307a0d499988e61de6c37ac840403aa8af99.jpg) +CoCoNet + +![](images/a8e17613a1daf2722791cfec22d8ab9669a1853b3a80ba20755ef25d32c0933f.jpg) +LRRNet + +![](images/239e8bda70cd922385111cb3a3f7ce5754db7f8f8e2bddd653a31d93705b0e60.jpg) +Infrared +CDDFuse +Figure 5. Visual comparison results of our method with seven state-of-the-art methods on typical image pairs from MSRS (the first group), M $^3$ FD (the second group), and TNO (the third group) datasets. + +![](images/9154111fd9a2bff71b6416310e8d9fe16e1f394843648c54f1fffe9078a9e38b.jpg) +Visible +MetaFusion + +![](images/9cc03f50a44e3589d34a3b402d596b6d1d3912b70e6e92170c904697d3074447.jpg) +SwinFusion + +![](images/55dc447b27e0fd04b75b2fdce73441c8004605a10d5abb604b37cf0a65d4484a.jpg) +TUFusion + +![](images/d1136693fbf960b67e9d1a31eb37ce92212e859aa8646e9fff2b45bf574f7c5a.jpg) +Ours + +duce distortions in the legs of the highlighted person. The second and third groups of images illustrate the capability of different methods to preserve details. For instance, neither the clouds in the second group nor the shrubberies in the third group are effectively preserved by DATFuse, LRR-Net, SwinFusion, and TUFusion, while CoCoNet and MetaFusion introduce significant noise. Only CDDFuse and our proposed method perform well in both aspects. + +Quantitative Comparison. The quantitative comparison results are presented in Tab. 1. In the context of the MSRS dataset, our method ranks first across all evaluation metrics, demonstrating the following advantages. 1) The optimal VIF indicates that our technique yields fused outputs with enhanced visual quality, further confirming its qualitative superiority. 2) The ideal SCD, MI, and FMI suggest that our method is most effective in conveying data from the + +original input to the outcome. 3) The superior SSIM and MS-SSIM show that our model minimizes information loss and image distortion. 4) The perfect Qabf illustrates that our solution successfully preserves the maximum edge details. Furthermore, to verify the generalization competence of our method, we perform tests on the $\mathbf{M}^3\mathrm{FD}$ and TNO datasets without any fine-tuning. In the context of the $\mathbf{M}^3\mathrm{FD}$ dataset, our technique leads in every metric, with the sole exceptions being SCD and MI. Within the TNO dataset, it secures four top positions and two runner-up spots. + +# 4.3. Ablation Studies + +Within this part, we undertake multiple ablations. Each one is executed using the MSRS test dataset as a foundation. Extension. To explore the correctness of extending the attention properties of infrared images to the visible modality, two sets of controlled experiments are performed: the first set utilize only I-SCA, while the second set employ both I-SCA and V-SCA. The experimental results are shown in Tab. 2. The integration of V-SCA led to enhancements in all measurements, with the sole exception of a minor reduction in SCD and SSIM, which confirms the validity of this extension. + +I/V-SCA. To validate the effectiveness of I-SCA and V-SCA, we design two strategies in this subsection. 1) Directly removing the proposed modules, which is denoted as w/o I/V-SCA. 2) To prevent inconsistencies in model complexity before and after the removal of I/V-SCA, which could introduce additional experimental bias, we further replace I/V-SCA with Restormer Blocks. This ablation is denoted as "to Restormer". Findings from the experiment are detailed in Tab. 3, which illustrates that our technique holds the top position in five of the seven assessment measures, affirming the efficacy of I/V-SCA. + +CBSM. For confirming CBSM's efficiency, we substitute CBSM with Parameters, $Conv_{1 \times 1}$ , and $Conv_{3 \times 3}$ , respectively. The experiment's findings are outlined in Tab. 4. Compared to the other three schemes, the introduction of CBSM results in more than half of the evaluated metrics being ranked first, which suggests that CBSM is capable of efficiently diminishing the disparities between the source image and the intermediate features of another modality. + +Table 2. Ablation before and after the extension of I-SCA. + +
ConfigurationsVIFSCDMIQabfSSIMMS_SSIMFMI
I-SCAV-SCA
✓✗1.0391.7214.5710.7030.9820.9750.932
✓✓1.0611.7005.1110.7150.9810.9750.933
+ +# 4.4. Efficiency Comparison + +Within this part, we evaluate the parameters, FLOPs, and FPS of our work with the most competitive methods + +Table 3. Ablation of the I-SCA and V-SCA modules. + +
ConfigurationsVIFSCDMIQabfSSIMMS_SSIMFMI
w/o I/V-SCA1.0651.6695.6510.7080.9680.9670.933
to Restormer1.0631.6815.4630.7120.9750.9710.933
Ours1.0611.7005.1110.7150.9810.9750.933
+ +Table 4. Ablation of the CBSM module. + +
ConfigurationsVIFSCDMIQabfSSIMMS_SSIMFMI
Parameters1.0601.7085.3030.7120.9770.9730.932
Conv1×11.0261.6914.6710.6910.9750.9670.932
Conv3×31.0661.7055.2540.7110.9770.9730.932
CBSM1.0611.7005.1110.7150.9810.9750.933
+ +to highlight its speed advantage. Findings are presented in Tab. 5, where the FLOPs are determined based on an input dimension of $128 \times 128 \times 1$ , and the FPS on $\mathbf{M}^3\mathrm{FD}$ is measured with images of reduced resolution by half. Our method achieves the highest FPS, enabling seamless integration with both low- and high-level computer vision tasks. + +Table 5. The parameters, FLOPs, and FPS. + +
Params (M) ↓FLOPs (G) ↓FPS ↑
MSRSM3FDTNO
SwinFusion [20]0.92717.9860.6241.0160.651
CDDFuse [42]1.18629.21339.72340.48040.179
Ours0.5659.207132.303137.271126.537
+ +# 5. Conclusion + +This paper reveals the inherent "attention properties" of infrared images from the Grad-CAM of ResNet-50. To incorporate the property into IVF, we propose I-SCA and extend it to the visible modality, subsequently introducing V-SCA. The auxiliary module (CBSM) further ensures fusion quality by enhancing feature channels, enabling spatial transformations, and suppressing negative information in the source images. Additionally, the linear computations of I-SCA and V-SCA guarantee efficiency, achieving FPS of 132, 137, and 127 on MSRS, $\mathbf{M}^3\mathrm{FD}$ , and TNO, respectively. Moving forward, our focus will shift to crafting more specialized modules and undertaking an in-depth methodological examination to enhance the refinement of our approach. + +# Acknowledgement + +This work was supported by the National Natural Science Foundation of China (62272426), the Shanxi Province Science and Technology Major Special Project (202201150401021), and the Fundamental Research Program of Shanxi Province (TZLH20230818005, 202303021212189, 202303021212372, 202403021212166). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[2] Yaming Cao, Lei Guo, Fengguang Xiong, Liqun Kuang, and Xie Han. Physical-simulation-based dynamic template matching method for remote sensing small object detection. IEEE Transactions on Geoscience and Remote Sensing, 62: 1-14, 2023. 3 +[3] Zhihao Chang, Zhixi Feng, Shuyuan Yang, and Quanwei Gao. Aft: Adaptive fusion transformer for visible and infrared images. IEEE Transactions on Image Processing, 32: 2077-2092, 2023. 3 +[4] Jun Chen, Jianfeng Ding, and Jiayi Ma. Hitfusion: Infrared and visible image fusion for high-level vision tasks using transformer. IEEE Transactions on Multimedia, 26: 10145-10159, 2024. 3 +[5] Chunyang Cheng, Tianyang Xu, and Xiao-Jun Wu. Mufusion: A general unsupervised image fusion network based on memory unit. Information Fusion, 92:80-92, 2023. 3 +[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3 +[7] Keying Du, Huafeng Li, Yafei Zhang, and Zhengtao Yu. Chitnet: A complementary to harmonious information transfer network for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 74: 1-17, 2025. 5 +[8] Xueyan Gao and Shiguang Liu. Bcmfuse: A bilateral cross-modal feature interaction-based network for infrared and visible image fusion. Remote Sensing, 16(17):3136, 2024. 3 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2016. 2 +[10] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3, 4 +[11] Shahid Karim, Geng Tong, Jinyang Li, Akeel Qadir, Umar Farooq, and Yiting Yu. Current advances and future perspectives of image fusion: A comprehensive review. Information Fusion, 90:185-217, 2023. 1 +[12] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2019. 3 +[13] Hui Li, Tianyang Xu, Xiao-Jun Wu, Jiwen Lu, and Josef Kittler. Lrrnet: A novel representation learning guided fusion network for infrared and visible images. IEEE transactions on pattern analysis and machine intelligence, 45(9):11040-11052, 2023. 3, 6 +[14] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In European Conference on Computer Vision, pages 719-735, 2022. 3 + +[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3 +[16] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5811, 2022. 6 +[17] Jinyuan Liu, Runjia Lin, Guanyao Wu, Risheng Liu, Zhongxuan Luo, and Xin Fan. Coconet: Coupled contrastive learning network with multi-level feature ensemble for multi-modality image fusion. International Journal of Computer Vision, 132(5):1748-1775, 2024. 3, 6 +[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3 +[19] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information fusion, 45:153-178, 2019. 6 +[20] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swim transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200-1217, 2022. 3, 6, 8 +[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[22] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), page 618-626. IEEE, 2017. 2 +[23] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Information Fusion, 83:79-92, 2022. 6 +[24] Wei Tang, Fazhi He, Yu Liu, Yansong Duan, and Tongzhen Si. Datfuse: Infrared and visible image fusion via dual attention transformer. IEEE Transactions on Circuits and Systems for Video Technology, 33(7):3159-3172, 2023. 3, 6 +[25] Alexander Toet. The tno multiband image data collection. Data in brief, 15:249-251, 2017. 6 +[26] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3 +[27] Boying Wang, Libo Zhang, Longyin Wen, Xianglong Liu, and Yanjun Wu. Towards real-world prohibited item detection: A large-scale x-ray benchmark. In Proceedings of + +the IEEE/CVF international conference on computer vision, pages 5412-5421, 2021.3 +[28] Boying Wang, Ruyi Ji, Libo Zhang, Yanjun Wu, and Jing Liu. Learning to zoom: Exploiting mixed-scale contextual information for object detection. Expert Systems with Applications, 264:125871, 2025. 3 +[29] Zhishe Wang, Yanlin Chen, Wenyu Shao, Hui Li, and Lei Zhang. Swinfuse: A residual swim transformer fusion network for infrared and visible images. IEEE Transactions on Instrumentation and Measurement, 71:1-12, 2022. 3 +[30] Zhishe Wang, Wenyu Shao, Yanlin Chen, Jiawei Xu, and Xiaoqin Zhang. Infrared and visible image fusion via interactive compensatory attention adversarial learning. IEEE Transactions on Multimedia, 25:7800-7813, 2022. 3 +[31] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3 +[32] Yongbiao Xiao, Hui Li, Chunyang Cheng, and Xiaoning Song. Le2fusion: A novel local edge enhancement module for infrared and visible image fusion. In International conference on image and graphics, pages 293-305. Springer, 2023. 3 +[33] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 6 +[34] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Text-if: Leveraging semantic text guidance for degradation-aware and interactive image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27026-27035, 2024. 3 +[35] Jun Yue, Leyuan Fang, Shaobo Xia, Yue Deng, and Jiayi Ma. Dif-fusion: Toward high color fidelity in infrared and visible image fusion with diffusion models. IEEE Transactions on Image Processing, 32:5705-5720, 2023. 3 +[36] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 3 +[37] Hao Zhang, Xuhui Zuo, Jie Jiang, Chunchao Guo, and Jiayi Ma. Mrfs: Mutually reinforcing image fusion and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26974-26983, 2024. 3 +[38] Cheng Zhao, Peng Yang, Feng Zhou, Guanghui Yue, Shuigen Wang, Huisi Wu, Guoliang Chen, Tianfu Wang, and Baiying Lei. Mhw-gan: multidiscriminator hierarchical wavelet generative adversarial network for multimodal image fusion. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3 +[39] Rong Zhao, Xie Han, Xindong Guo, Liquun Kuang, Xiaowen Yang, and Fusheng Sun. Exploring the point feature relation on point cloud for multi-view stereo. IEEE Transactions on Circuits and Systems for Video Technology, 33(11): 6747-6763, 2023. 3 + +[40] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13955-13965, 2023. 3, 6 +[41] Yangyang Zhao, Qingchun Zheng, Peihao Zhu, Xu Zhang, and Wenpeng Ma. Tufusion: A transformer-based universal fusion algorithm for multimodal images. IEEE Transactions on Circuits and Systems for Video Technology, 34(3): 1712-1725, 2024. 3, 6 +[42] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5906-5916, 2023. 3, 5, 6, 8 +[43] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: denoising diffusion model for multi-modality image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8082-8093, 2023. 3 +[44] Zixiang Zhao, Lilun Deng, Haowen Bai, Yukun Cui, Zhipeng Zhang, Yulun Zhang, Haotong Qin, Dongdong Chen, Jiangshe Zhang, Peng Wang, et al. Image fusion via vision-language model. In Proceedings of the 41st International Conference on Machine Learning, pages 60749-60765, 2024. 3 +[45] Pengfei Zhu, Yang Sun, Bing Cao, and Qinghua Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099-7108, 2024. 3 \ No newline at end of file diff --git a/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/images.zip b/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3aa9644ae7cfc053078c3c8a090053f64651a4cf --- /dev/null +++ b/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b496d5928b5a051afd3319fc033a0ff574608a69df16289bbdac83e1a1516c6 +size 631876 diff --git a/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/layout.json b/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3b38a41d868da219ab087752a6dc48365cb107dd --- /dev/null +++ b/2025/The Source Image is the Best Attention for Infrared and Visible Image Fusion/layout.json @@ -0,0 +1,10009 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 544, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 544, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 544, + 121 + ], + "type": "text", + "content": "The Source Image is the Best Attention for Infrared and Visible Image Fusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 125, + 143, + 484, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 143, + 484, + 172 + ], + "spans": [ + { + "bbox": [ + 125, + 143, + 484, + 172 + ], + "type": "text", + "content": "Song Wang Xie Han Liquun Kuang\\* Boying Wang Zhongyu Chen Zherui Qiao Fan Yang Xiaoxia Liu Bingyu Zhang Zhixun Wang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 174, + 481, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 174, + 481, + 185 + ], + "spans": [ + { + "bbox": [ + 128, + 174, + 481, + 185 + ], + "type": "text", + "content": "School of Computer Science and Technology, North University of China, Taiyuan 030051, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 150, + 186, + 459, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 186, + 459, + 198 + ], + "spans": [ + { + "bbox": [ + 150, + 186, + 459, + 198 + ], + "type": "text", + "content": "Shanxi Key Laboratory of Machine Vision & Virtual Reality, Taiyuan 030051, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 201, + 511, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 201, + 511, + 213 + ], + "spans": [ + { + "bbox": [ + 97, + 201, + 511, + 213 + ], + "type": "text", + "content": "Shanxi Vision Information Processing and Intelligent Robot Engineering Research Center, Taiyuan 030051, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 192, + 216, + 416, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 216, + 416, + 227 + ], + "spans": [ + { + "bbox": [ + 192, + 216, + 416, + 227 + ], + "type": "text", + "content": "b20230704@st.nuc.edu.cn,kuang@nuc.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 280, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 616 + ], + "type": "text", + "content": "Infrared and visible image fusion (IVF) endeavors to engineer composite outputs by blending optimal virtues of divergent modalities. This paper reveals, unprecedentedly, the intrinsic \"attention properties\" of infrared images, which directly arise from their physical characteristics (i.e., heat distribution) and can be linked to attention mechanisms naturally, as observed in the gradient-weighted class activation mapping (Grad-CAM) visualization analysis of image classification models. To incorporate this property into IVF for better fusion, we propose the source infrared cross attention (I-SCA) and further extend it to the visible modality, subsequently introducing the source visible cross attention (V-SCA). The joint use of I-SCA and V-SCA greatly alleviate longstanding issues, such as insufficient and incomplete multimodal feature interaction and fusion, in IVF. Moreover, an auxiliary component for I-SCA and V-SCA, termed CBSM, is employed to boost the channel, map space, and suppress redundancy and misleading information of the source images. Specifically, we directly treat the CBSM-processed raw image as the query, while the intermediate features of another modality are treated as keys and values in I-SCA and V-SCA. Unlike attention mechanisms that divide images into patches or limit computations to local windows, our cross attention modules achieve smoother and more robust IVF through true global modeling across the entire image space with linear complexity. Comparison with current SOTA methods on three popular public datasets confirms its superiority." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": "IVF, a bedrock of image processing, with the objective of engendering a composite that encapsulates the considerable thermal radiation from infrared and the intricacies de" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 339, + 253, + 431, + 331 + ], + "blocks": [ + { + "bbox": [ + 339, + 253, + 431, + 331 + ], + "lines": [ + { + "bbox": [ + 339, + 253, + 431, + 331 + ], + "spans": [ + { + "bbox": [ + 339, + 253, + 431, + 331 + ], + "type": "image", + "image_path": "cf6a6a3a69c705248c33627af0920a354fa0205b3e693fab823f005c4e66d760.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 332, + 388, + 338 + ], + "lines": [ + { + "bbox": [ + 381, + 332, + 388, + 338 + ], + "spans": [ + { + "bbox": [ + 381, + 332, + 388, + 338 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 438, + 253, + 529, + 331 + ], + "blocks": [ + { + "bbox": [ + 438, + 253, + 529, + 331 + ], + "lines": [ + { + "bbox": [ + 438, + 253, + 529, + 331 + ], + "spans": [ + { + "bbox": [ + 438, + 253, + 529, + 331 + ], + "type": "image", + "image_path": "42d9e79675dcb052c7be455df5e9004fad7d2174bd45e51c9827735cba04ff01.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 339, + 339, + 430, + 407 + ], + "blocks": [ + { + "bbox": [ + 339, + 339, + 430, + 407 + ], + "lines": [ + { + "bbox": [ + 339, + 339, + 430, + 407 + ], + "spans": [ + { + "bbox": [ + 339, + 339, + 430, + 407 + ], + "type": "image", + "image_path": "60f756b556ad51bd892a8590a1c3e0e75c420a167306e77564c381641e919d87.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 409, + 388, + 415 + ], + "lines": [ + { + "bbox": [ + 381, + 409, + 388, + 415 + ], + "spans": [ + { + "bbox": [ + 381, + 409, + 388, + 415 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 438, + 338, + 529, + 407 + ], + "blocks": [ + { + "bbox": [ + 480, + 332, + 488, + 338 + ], + "lines": [ + { + "bbox": [ + 480, + 332, + 488, + 338 + ], + "spans": [ + { + "bbox": [ + 480, + 332, + 488, + 338 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 438, + 338, + 529, + 407 + ], + "lines": [ + { + "bbox": [ + 438, + 338, + 529, + 407 + ], + "spans": [ + { + "bbox": [ + 438, + 338, + 529, + 407 + ], + "type": "image", + "image_path": "94df5b43c0a9fbd1255f09420f8b5c323a4452d5b3fbecebe99263e0fc68ff88.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 339, + 416, + 430, + 483 + ], + "blocks": [ + { + "bbox": [ + 339, + 416, + 430, + 483 + ], + "lines": [ + { + "bbox": [ + 339, + 416, + 430, + 483 + ], + "spans": [ + { + "bbox": [ + 339, + 416, + 430, + 483 + ], + "type": "image", + "image_path": "b7e700277990f92bce909bb11eedcfae37e9a0ef4d900f865c63a74fcef4cfba.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 485, + 389, + 491 + ], + "lines": [ + { + "bbox": [ + 381, + 485, + 389, + 491 + ], + "spans": [ + { + "bbox": [ + 381, + 485, + 389, + 491 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 438, + 415, + 529, + 484 + ], + "blocks": [ + { + "bbox": [ + 480, + 409, + 488, + 415 + ], + "lines": [ + { + "bbox": [ + 480, + 409, + 488, + 415 + ], + "spans": [ + { + "bbox": [ + 480, + 409, + 488, + 415 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 438, + 415, + 529, + 484 + ], + "lines": [ + { + "bbox": [ + 438, + 415, + 529, + 484 + ], + "spans": [ + { + "bbox": [ + 438, + 415, + 529, + 484 + ], + "type": "image", + "image_path": "41fffc37ad8b5d59409848e620541e1864e5188e612d5bdf874955d285424cc9.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 485, + 488, + 491 + ], + "lines": [ + { + "bbox": [ + 481, + 485, + 488, + 491 + ], + "spans": [ + { + "bbox": [ + 481, + 485, + 488, + 491 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 503, + 555, + 613 + ], + "lines": [ + { + "bbox": [ + 313, + 503, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 555, + 613 + ], + "type": "text", + "content": "Figure 1. (a, b) Grad-CAM visualization results for two images (00022004, 00028021), both labeled as \"dog\", predicted by the final layer of ResNet-50 on the ImageNet validation set. The color gradient from red to blue indicates the decreasing level of attention paid by the model to different regions. (c) An infrared image example (00519D) consists of highlighted thermal radiation regions and remaining ordinary regions from the MSRS dataset. (d) Visible image of the same scene corresponding to subplot (c). (e) Subplot (c) is displayed by applying the Jet colormap. (f) The negative transformation of subplot (c) is visualized with the Jet colormap." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 639, + 555, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 555, + 699 + ], + "type": "text", + "content": "tail from visible. Prior to the surge in deep learning, this process was typically achieved by traditional algorithms in spatial and transform domains [11]. However, thanks to the development of neural networks and attention mechanisms, heuristic algorithms have become more widely adopted." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "In this paper, we draw inspiration from the Grad-CAM" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13513" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 276 + ], + "type": "text", + "content": "[22] visualization results of classic image classification models, such as ResNet-50 [9], and discover the inherent \"attention properties\" of infrared images. These Grad-CAM visualizations, as shown in Fig. 1(a) and Fig. 1(b), reveal how the model assigns varying weights to different regions during the prediction process: regions marked in red represent \"important areas\" that exert a substantial impact on the classification result, while regions marked in blue correspond to \"general areas\" with lesser contribution. Interestingly, we observe that the structural characteristics of infrared images (e.g., as shown in Fig. 1(c)) are strikingly similar to the Grad-CAM images, as both consist of important areas and general areas (in infrared images, these regions correspond to the highlighted areas with higher thermal energy and the darker zones with lower energy). To more intuitively showcase this similarity, we re-map the infrared image with the Jet colormap (see Fig. 1(e))." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 278, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 469 + ], + "type": "text", + "content": "Based on these observations, we believe that infrared images inherently possess an attention property similar to the one exhibited by Grad-CAM. Unlike image classification models, which require be trained to learn attention patterns, the attention properties of infrared images directly stem from their physical characteristics—the heat distribution. Moreover, this attribute is not only more specific and accurate but also avoids issues such as insufficient or misdirected attention that may occur in image classification models. For instance, when ResNet-50 processes the dog in Fig. 1(a), it focuses only on the dog's eyes, and when processing the dog in Fig. 1(b), it not only focuses on the dog itself but also incorrectly attends to irrelevant backgrounds. In contrast, infrared images provide precise localization of heat sources, ensuring that all attention is concentrated on specific areas." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 472, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 616 + ], + "type": "text", + "content": "Based on the above finding, an immediate question arises: how can we fully leverage the inherent attention properties of infrared images to obtain better-fused images in IVF? Intuitively, we think that it can be linked to the attention mechanism [26], as both essentially represent the importance of different regions. Furthermore, the interaction of cross-modal features plays a pivotal role in image fusion. Therefore, our goal is to introduce the attention properties at this stage to counteract the persistent obstacles in image fusion, such as inadequate cross-domain information exchange and incomplete multimodal feature fusion, thereby enhancing the caliber of the merged images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "Specifically, we propose the source infrared cross attention (I-SCA) based on a modified Restormer [36], which can perform global feature modeling in the image space with linear complexity. I-SCA first takes the original infrared image and the intermediate features of the visible image as inputs. It then treats the infrared image as the query (Q), thanks to the properties of infrared images, which allow them to be directly used as queries without processing," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "while the visible features serve as the keys (K) and values (V), enabling an infrared heat prior-guided cross-domain information exchange. However, this approach can only prioritize on \"important areas\", whereas image fusion is a global pixel-wise task that demands equal treatment for all areas. Fortunately, a facile solution is to treat the negative transformation of the infrared image (as shown in Fig. 1(f)) as another query to allow significant attention complementarity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 184, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 553, + 304 + ], + "type": "text", + "content": "Nevertheless, using I-SCA only also faces a critical problem, i.e., it can only capture \"infrared-guided\" cross-modal interaction features. Theoretically, the optimal fused image could be obtained by simultaneously fusing features dominated by both infrared and visible images. Therefore, we make a bold inference: we extend the discovery to visible and, based on this, propose the source visible cross attention (V-SCA). While visible images lack explicit physical saliency, their visually significant regions (e.g., edges, textures, and contrast) can still serve as effective priors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 308, + 553, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 553, + 488 + ], + "type": "text", + "content": "Finally, since both I-SCA and V-SCA require inputs to have the same channels, but the channels of the source image and the intermediate representation from the other modality are not consistent, we deploy an adaptive channel boosting and intelligent space mapping module (CBSM), which also serves crucial roles in transforming the feature space of the source image from its initial state to one more suitable for attention mechanisms to reduce discrepancies between features at different stages and suppressing redundancy or misleading content contained in the initial images. To validate the effectiveness of I-SCA, V-SCA, and CBSM, extensive analysis are conducted on three popular public datasets. Experiments demonstrate that both qualitative and quantitative results outperform state-of-the-art (SOTA) methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 492, + 514, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 492, + 514, + 503 + ], + "spans": [ + { + "bbox": [ + 326, + 492, + 514, + 503 + ], + "type": "text", + "content": "The essence of our contributions is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 508, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 326, + 508, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 508, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 326, + 508, + 553, + 555 + ], + "type": "text", + "content": "- We observe a strong structural correlation between infrared images and Grad-CAM, leading us to hypothesize, for the first time, that infrared images possess an inherently exceptionally precise attention property." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 556, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 556, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 326, + 556, + 553, + 628 + ], + "type": "text", + "content": "- Based on this attention property, we propose a heat-distribution-prior-guided cross-modal attention mechanism, which can be further extend to the visible modality. Additionally, an auxiliary module, CBSM, is deployed to complement the proposed attention mechanism for further enhancing fusion quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 628, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 628, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 326, + 628, + 553, + 675 + ], + "type": "text", + "content": "- Our findings inspire a novel paradigm guided by source image priors for image fusion, significantly alleviating challenges in cross-domain feature interaction and fusion faced by current methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 675, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 675, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 326, + 675, + 553, + 712 + ], + "type": "text", + "content": "- Comprehensive experiments instituted on three prominent available datasets with SOTA works demonstrate the superiority." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13514" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 147, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 147, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 147, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 231, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 231, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 231, + 103 + ], + "type": "text", + "content": "2.1. Attention-based Fusion Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 110, + 297, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 297, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 297, + 613 + ], + "type": "text", + "content": "The triumph of transformer [26] in NLP has sparked considerable interest across the academic community. To establish a foundation for the integration of this technology in computer vision, Alexey et al. proposed Vision Transformer (ViT) [6], which segments an image into multiple units, then flattens the units into a shape compatible with text sequences, allowing direct application of the original transformer for long-range modeling. Subsequently, works such as TUFusion [41], DATfuse [24], HitFusion [4], AFT-Fusion [3], etc., extended ViT to the image fusion task. On the one hand, TUFusion adopted an encoder-decoder pipeline, featuring a dual-branch structure that merged parallel DenseBlock [12] and ViT in the encoder. DATFuse, on the other hand, incorporated channel and spatial attention mechanisms similar to CBAM [31] before the self-attention mechanism and performed fusion at the image level. Although transformer-based methods have shown great effectiveness, they are consistently constrained by the computational burden of the self-attention in image space. To remedy this issue, Liu et al. proposed SwinTransformer [18], which divides images into patches and further splits the patches into multiple windows, restricting self-attention to each window. Later, SwinFuse [29] used a full SwinTransformer approach to obtain fused images. SwinFusion [20] further extended this by introducing a cross-modal SwinTransformer, which generalizes window attention to cross-modal feature fusion. However, the practice of restricting attention to local regions contradicts the fundamental idea of attention mechanisms, which is to model global information. To achieve true overall modeling linearly, Zamir et al. proposed Restormer, which introduces multi-dconv head transposed attention (MDTA) and gated dconv feedforward network (GDFN) to replace the self-attention and feed-forward network [36]. Later, CDDFuse [42] further validated the effectiveness of the Restormer in image fusion. However, it only applied the Restormer to a single modality and overlooked the potential for feature interactions across multiple modalities. This paper narrowed down the challenge of interaction between different modalities by proposing I-SCA and V-SCA, and further enabled Restormer to be used for the multimodal information interaction problem." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 623, + 211, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 211, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 211, + 635 + ], + "type": "text", + "content": "2.2. CNN-based Fusion Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Currently, the realm of deep learning-driven fusion is segmented into various types, encompassing techniques reliant on autoencoders, end-to-end models, unified models, amalgamations with downstream or other specialized tasks, and approaches utilizing generative adversarial networks (GAN). Autoencoder-based methods [14, 32] first" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 526 + ], + "type": "text", + "content": "train an encoder-decoder pair on datasets like COCO [15] and then directly generalize them for image fusion. However, the performance of such methods is markedly hampered by the quality of the fusion rules designed subjectively, and developing a fusion strategy that works well with the encoder-decoder framework is not trivial. Consequently, a series of end-to-end methods [8, 13] were later proposed. These methods train all modules, including the feature fusion layer, with learning strategies, which greatly reduces the challenges related to constructing fusion strategies in autoencoder-based methods. Furthermore, as image fusion has evolved, the community has become dissatisfied with using a single model to solve just one specific fusion task, leading to unified models [5, 17, 45] for multiple tasks. However, these methods adopt a re-training strategy for each new task, rather than using a single weight parameter to handle multiple tasks simultaneously. Meanwhile, along with the development of downstream tasks such as object detection [2, 27, 28], scene reconstruction [39], etc., a major trend has been to combine image fusion with these computer vision sub-tasks, facilitating the mutual promotion of both [37, 40]. Another category of methods [34, 44] leverages large language models such as GPT [1] and CLIP [21] to guide the fusion process, achieving cross-domain integration for different tasks. Finally, GAN-based methods [30, 38] transform the fusion into an adversarial problem. Through continuous antagonistic training, the generator's output progressively converges toward the target distribution. Nevertheless, these techniques face hurdles like mode collapse, erratic training, and a lack of clear understanding. Later works proposed denoising diffusion model-based methods [35, 43] to address these issues. Although the core innovation of this study lies in source images prior-guided cross-modal attention, approximately half of the components are still based on the convolutional neural network (CNN). To ensure a comprehensive evaluation, we undertake a comparative analysis of it with several recent CNN-based SOTA methods discussed above." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 541, + 370, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 541, + 370, + 554 + ], + "spans": [ + { + "bbox": [ + 314, + 541, + 370, + 554 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 563, + 434, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 434, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 434, + 574 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "In this work, a hybrid of CNN and attention mechanisms is employed to achieve the IVF. The overall architecture is shown in Fig. 2. Specifically, we construct a serial pipeline, consisting of SE-ResNet Block [10], Restormer Block [36], and I-SCA/V-SCA for infrared and visible branches. For each modality, we first employ the SE-ResNet Block (the SE acts as a gating mechanism to suppress information redundancy by refine features) for initial feature extraction while increasing the feature channel from ch_in to ch_mid. The Restormer Block then performs further deep extraction with another key function of transforming the feature space" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13515" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 76, + 558, + 273 + ], + "blocks": [ + { + "bbox": [ + 65, + 76, + 558, + 273 + ], + "lines": [ + { + "bbox": [ + 65, + 76, + 558, + 273 + ], + "spans": [ + { + "bbox": [ + 65, + 76, + 558, + 273 + ], + "type": "image", + "image_path": "64735d6cc1ae6d0f6815b6935e058903fd7407df15ddb76dc1ca766692751859.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 285, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 54, + 285, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 555, + 331 + ], + "type": "text", + "content": "Figure 2. The overall architecture of our method, which incorporates SE-ResNet Block, Restormer Block, I-SCA/V-SCA, and CBSM. ir, vi denote infrared and visible images, and " + }, + { + "bbox": [ + 54, + 285, + 555, + 331 + ], + "type": "inline_equation", + "content": "ir\\_n" + }, + { + "bbox": [ + 54, + 285, + 555, + 331 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 285, + 555, + 331 + ], + "type": "inline_equation", + "content": "vi\\_n" + }, + { + "bbox": [ + 54, + 285, + 555, + 331 + ], + "type": "text", + "content": " signify negative transformations of infrared and visible images. ch_in, ch_mid, and ch_out represent input, middle, and output channels, respectively. “C” indicates a channel-wise concatenation operation. “Q” stands for the query in the attention mechanism, while “K” and “V” symbolize the key and value, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 350, + 294, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 294, + 483 + ], + "type": "text", + "content": "from the domain where convolutional operations excel to the one where attention is more effective, thereby facilitating the subsequent cross-modal feature interaction. The I-SCA/V-SCA subsequently takes the features of the current modality along with either the image of another modality processed by CBSM or the negative transformation of that image for feature interaction. Post processing via I-SCA/V-SCA, a quartet of feature maps, steered by the originating image, is acquired. Then they will merge across the channel dimension and be processed through an additional SE-ResNet Block before being output." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 490, + 111, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 111, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 111, + 502 + ], + "type": "text", + "content": "3.2. CBSM" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 65, + 517, + 290, + 605 + ], + "blocks": [ + { + "bbox": [ + 65, + 517, + 290, + 605 + ], + "lines": [ + { + "bbox": [ + 65, + 517, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 65, + 517, + 290, + 605 + ], + "type": "image", + "image_path": "6d87b58cc64b314abb4deff6f780c7e4b6301f6c4e3af48ebc832590aeca096f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 618, + 295, + 664 + ], + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 664 + ], + "type": "text", + "content": "Figure 3. The composition of adaptive channel boosting and intelligent space mapping module (CBSM). The SE is employed to suppress redundancy and misleading information in the source images. " + }, + { + "bbox": [ + 55, + 618, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 618, + 295, + 664 + ], + "type": "text", + "content": " denotes a channel-wise multiplication operation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Since attention calculation in I-SCA and V-SCA requires all channels of inputs to be consistent, it is imperative to augment the channel dimension of the source image" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 350, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 616 + ], + "type": "text", + "content": "from ch_in to ch_mid. In general, this can be achieved by Parameters, " + }, + { + "bbox": [ + 313, + 350, + 555, + 616 + ], + "type": "inline_equation", + "content": "Conv_{1\\times 1}" + }, + { + "bbox": [ + 313, + 350, + 555, + 616 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 313, + 350, + 555, + 616 + ], + "type": "inline_equation", + "content": "Conv_{3\\times 3}" + }, + { + "bbox": [ + 313, + 350, + 555, + 616 + ], + "type": "text", + "content": ". However, all three have a common problem: after channel expansion, the multi-channel feature space still remains in the initial state. In cross-modal attention, directly using feature values from the raw space for interaction with deep features will significantly degrade the information exchange effect due to spatial discrepancy. Therefore, a key issue is how to reduce the gap between them. Additionally, the source image may introduce redundancy or misleading cues in certain cases, making it crucial to mitigate potential negative effects. To address this, CMSM module, which consists of convolutional layers (for channel boosting), activation layers (for spatial mapping), and Squeeze-and-Excitation (SE) [10] layer (for suppressing negative influences) is employed, as shown in Fig. 3. The structure of CMSM is intentionally simple, because it aligns with our goal of using the raw image priors to guide feature interactions. A more complex design could significantly alter the initial pattern of the source image, conflicting with our motivation. For the input source image, denoted as " + }, + { + "bbox": [ + 313, + 350, + 555, + 616 + ], + "type": "inline_equation", + "content": "I_m^{raw} \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 313, + 350, + 555, + 616 + ], + "type": "text", + "content": ", this module performs the following operations:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 323, + 620, + 555, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 620, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 323, + 620, + 555, + 635 + ], + "type": "interline_equation", + "content": "I _ {m} ^ {t r s} = \\alpha \\left(S E \\left(C o n v _ {3 \\times 3} \\left(\\alpha \\left(C o n v _ {3 \\times 3} \\left(I _ {m} ^ {r a w}\\right)\\right)\\right)\\right)\\right). (1)", + "image_path": "8331babbc90b6efe3c2c978a084bb87ca2e428dc6fbd713bcb098d1a03ab9eba.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "The result of Eq. (1), denoted as " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "I_{m}^{trs} \\in \\mathbb{R}^{H \\times W \\times 48}" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ", represents the image after channel boosting and space transformation. The variable " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " represents different input modalities, and desirable values include " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "ir" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "vi" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "ir\\_n" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "vi\\_n" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "ir\\_n" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "vi\\_n" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " denote the negative transforms of the infrared and the visible image, respectively. " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha(\\cdot)" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " is the parametric" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13516" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 176, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 176, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 176, + 83 + ], + "type": "text", + "content": "rectified linear unit (PReLU)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 95, + 268, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 95, + 268, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 95, + 268, + 109 + ], + "type": "text", + "content": "3.3. Cross Attention Based on Source Images" + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 69, + 127, + 276, + 236 + ], + "blocks": [ + { + "bbox": [ + 69, + 127, + 276, + 236 + ], + "lines": [ + { + "bbox": [ + 69, + 127, + 276, + 236 + ], + "spans": [ + { + "bbox": [ + 69, + 127, + 276, + 236 + ], + "type": "image", + "image_path": "c150dac778b77550090f62882d409b099f4d9be16fc252f26c34f8425f17b5fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 247, + 274, + 326 + ], + "blocks": [ + { + "bbox": [ + 69, + 247, + 274, + 326 + ], + "lines": [ + { + "bbox": [ + 69, + 247, + 274, + 326 + ], + "spans": [ + { + "bbox": [ + 69, + 247, + 274, + 326 + ], + "type": "image", + "image_path": "979bb7d27141db02e313efd5ad12869d49efee98302463e8265147179aca27e1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 336, + 295, + 370 + ], + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 370 + ], + "type": "text", + "content": "Figure 4. Our proposed I-SCA and V-SCA, both of which consist of two modules: I-DCA/V-DCA and GDFN. Structurally, the two are identical, differing only in their input content." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "text", + "content": "The introduction of I-SCA and V-SCA aims to reduce the difficulty of multimodal feature interactions. The structures of these two components are illustrated in Fig. 4. In fact, both share an identical structural composition, differing only in the content of their inputs. Although the symmetrical design is used between I-SCA and V-SCA, this does not imply an inability to reflect modal heterogeneity. On the contrary, I-SCA emphasizes the heat distribution, contours, and structural priors of infrared images, while V-SCA focuses on edges, textures, and high-contrast guidance. Specifically, I-SCA comprises the dual-dconv infrared cross attention (I-DCA) and the gated dconv feedforward network (GDFN). Its inputs include " + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "inline_equation", + "content": "I_{ir}^{trs}" + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "inline_equation", + "content": "I_{ir,n}^{trs}" + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "text", + "content": ") and the visible features " + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "inline_equation", + "content": "feat_{vi}^{mid}" + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "text", + "content": ", which are processed by the Restormer Block in the visible branch. In contrast, V-SCA is composed of the dual-dconv visible cross attention (V-DCA) and the GDFN, with inputs " + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "inline_equation", + "content": "I_{vi}^{trs}" + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "inline_equation", + "content": "I_{vi,n}^{trs}" + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "text", + "content": ") and the infrared features " + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "inline_equation", + "content": "feat_{ir}^{mid}" + }, + { + "bbox": [ + 55, + 384, + 295, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "content": "To illustrate the working mechanism of I-SCA and V-SCA, we demonstrate I-SCA, with inputs " + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "inline_equation", + "content": "I_{ir}^{trs}" + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "inline_equation", + "content": "feat_{vi}^{mid}" + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "content": ", as an example. In I-SCA, both " + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "inline_equation", + "content": "I_{ir}^{trs}" + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "inline_equation", + "content": "feat_{vi}^{mid}" + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "content": " first undergo the layer normalization (LN) in I-DCA. Following this, " + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "inline_equation", + "content": "I_{ir}^{trs}" + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "content": " is directly used as the query (Q) after LN, while " + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "inline_equation", + "content": "feat_{vi}^{mid}" + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "content": " is further processed to generate the key (K) and value (V). The process is described as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 695, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 695, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 69, + 695, + 295, + 717 + ], + "type": "interline_equation", + "content": "K, V = D \\operatorname {c o n v} \\left(C o n v _ {1 \\times 1} \\left(L N \\left(f e a t _ {v i} ^ {m i d}\\right)\\right)\\right), \\tag {2}", + "image_path": "645ab20bbac0628d0d61e30564b5c5303b6b9d663ca6f36ec1578fbbeba479c5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "Dconv(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " denotes the depthwise separable convolution, consisting of two steps: depthwise convolution and pointwise convolution. This operation preserves spatial dimensions and independently processes the channels at each spatial position. Subsequently, the Q " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "(I_{ir}^{trs})" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", K, and V are reshaped for compatibility and undergo the following operation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 323, + 167, + 553, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 167, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 323, + 167, + 553, + 198 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f e a t _ {i r} ^ {o r i} = C o n v _ {1 \\times 1} \\left(V \\cdot S o f t m a x \\left(\\left(I _ {i r} ^ {t r s} \\cdot K\\right) \\odot s\\right)\\right) \\\\ + f e a t _ {v i} ^ {m i d}, (3) \\\\ \\end{array}", + "image_path": "549b22d85dd88c2fe7fcc94096316d485572bf638d5ae57ace8fa9d07f633584.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 207, + 555, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 555, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 555, + 256 + ], + "type": "text", + "content": "resulting in infrared-oriented intermediate features " + }, + { + "bbox": [ + 313, + 207, + 555, + 256 + ], + "type": "inline_equation", + "content": "\\text{feat}_{ir}^{\\text{ori}}" + }, + { + "bbox": [ + 313, + 207, + 555, + 256 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 207, + 555, + 256 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 207, + 555, + 256 + ], + "type": "text", + "content": " represents element-wise multiplication, and " + }, + { + "bbox": [ + 313, + 207, + 555, + 256 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 207, + 555, + 256 + ], + "type": "text", + "content": " is a learnable parameter. These features are then processed through the GDFN module as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 337, + 264, + 555, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 264, + 555, + 279 + ], + "spans": [ + { + "bbox": [ + 337, + 264, + 555, + 279 + ], + "type": "interline_equation", + "content": "Y _ {1}, Y _ {2} = D \\operatorname {c o n v} \\left(\\operatorname {C o n v} _ {1 \\times 1} \\left(L N \\left(f e a t _ {i r} ^ {\\text {o r i}}\\right)\\right)\\right), \\tag {4}", + "image_path": "3bdcbf65b56f96769acf7153561fb4d712c34e232c0ed4580e487703fb30bf6f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 304, + 554, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 304, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 320, + 304, + 554, + 323 + ], + "type": "interline_equation", + "content": "f e a t _ {i r} ^ {o r i} ^ {\\prime} = C o n v _ {1 \\times 1} \\left(G E L U \\left(Y _ {1}\\right) \\odot Y _ {2}\\right) + f e a t _ {i r} ^ {o r i}, (5)", + "image_path": "6d193ef2266fa009238c3125b956fdaf2dd6702e04232b349c77e8066fce36f9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 330, + 534, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 534, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 534, + 341 + ], + "type": "text", + "content": "where GELU refers to the Gaussian Error Linear Unit." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "text", + "content": "After the above process, four source images-oriented features, which are further refined by GDFN, denoted as " + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "inline_equation", + "content": "\\text{feat}_m^{\\text{ori}'}" + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "text", + "content": ", are obtained. These features, corresponding to different modalities (" + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "inline_equation", + "content": "m = ir" + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "inline_equation", + "content": "vi" + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "inline_equation", + "content": "ir\\_n" + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "inline_equation", + "content": "vi\\_n" + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "text", + "content": "), are arranged along the axis of the channel and further refined by another SE-ResNet Block. The transformation of selfattention from the spatial dimension to the channel dimension, implemented by " + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "inline_equation", + "content": "Dconv(\\cdot)" + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "inline_equation", + "content": "Conv_{1\\times 1}(\\cdot)" + }, + { + "bbox": [ + 313, + 342, + 555, + 489 + ], + "type": "text", + "content": ", ensures the linear computational characteristic of I-SCA and V-SCA. Additionally, the gating mechanism within the GDFN effectively suppresses redundant features, allowing only beneficial features to propagate through the module." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 496, + 404, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 404, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 404, + 508 + ], + "type": "text", + "content": "3.4. Loss Function" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 514, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 559 + ], + "type": "text", + "content": "Our approach utilizes intensity " + }, + { + "bbox": [ + 313, + 514, + 554, + 559 + ], + "type": "inline_equation", + "content": "(L_{int})" + }, + { + "bbox": [ + 313, + 514, + 554, + 559 + ], + "type": "text", + "content": " and gradient " + }, + { + "bbox": [ + 313, + 514, + 554, + 559 + ], + "type": "inline_equation", + "content": "(L_{grad})" + }, + { + "bbox": [ + 313, + 514, + 554, + 559 + ], + "type": "text", + "content": " loss techniques to guarantee comprehensive and detailed data for the combined image. The complete loss is characterized as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 380, + 571, + 553, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 571, + 553, + 583 + ], + "spans": [ + { + "bbox": [ + 380, + 571, + 553, + 583 + ], + "type": "interline_equation", + "content": "L _ {\\text {t o t a l}} = \\alpha_ {1} L _ {\\text {i n t}} + L _ {\\text {g r a d}}, \\tag {6}", + "image_path": "be47a3a4d3448acc359b2a285ec5d81f0fbcb2851c1abb6bff89931b93f6d2ed.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 591, + 553, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 553, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 553, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 591, + 553, + 615 + ], + "type": "inline_equation", + "content": "\\alpha_{1}" + }, + { + "bbox": [ + 313, + 591, + 553, + 615 + ], + "type": "text", + "content": " represents the intensity loss factor. The specific definition of intensity loss is as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 338, + 623, + 553, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 623, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 338, + 623, + 553, + 647 + ], + "type": "interline_equation", + "content": "L _ {i n t} = \\frac {1}{H W} \\| \\max \\left(I _ {i r} ^ {r a w}, I _ {v i} ^ {r a w}\\right) - f e a t _ {o u t} \\| _ {1}, \\tag {7}", + "image_path": "bb92ff85e19d007e50637ba94220901ce43836ac30ea1c58bfe35a3f19ead3e0.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\max(\\cdot)" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " represents the element-wise maximum operation, " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\text{feat}_{\\text{out}}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " signifies the final output of our model, and " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_1" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "l_1" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "-norm. Inspired by [42] and [7], the gradient term combines Sobel for edge preservation and Laplacian for noise suppression, guiding the model to capture" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "13517" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "valuable information while avoiding distractions. The formula is defined as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 124, + 106, + 295, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 106, + 295, + 119 + ], + "spans": [ + { + "bbox": [ + 124, + 106, + 295, + 119 + ], + "type": "interline_equation", + "content": "L _ {g r a d} = \\alpha_ {2} L _ {S} + \\alpha_ {3} L _ {L}, \\tag {8}", + "image_path": "cc8f26870f30ec91d7bd453f07287e8b03e69e06836c4d769394bebf76b81139.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 144, + 303, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 303, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 303, + 177 + ], + "type": "interline_equation", + "content": "L _ {S} = \\frac {1}{H W} \\| \\max \\left(\\left| \\nabla_ {S} I _ {i r} ^ {\\text {r a w}} \\right|, \\left| \\nabla_ {S} I _ {v i} ^ {\\text {r a w}} \\right|\\right) - \\left| \\nabla_ {S} f e a t _ {o u t} \\right| \\| _ {1}, \\tag {9}", + "image_path": "b89a320bd2f60df37bd3db34ee43e15df9b8d999410837a3b984b28f93aea956.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 222 + ], + "type": "interline_equation", + "content": "L _ {L} = \\| O \\left(\\max \\left(| \\nabla_ {L} I _ {i r} ^ {\\text {r a w}} |, | \\nabla_ {L} I _ {v i} ^ {\\text {r a w}} |\\right)\\right) - \\nabla_ {L} f e a t _ {o u t} \\| _ {1}, \\tag {10}", + "image_path": "202f72c388294121a14f8f70596e4063e4b47461ddb3b9106edbcc28858d9f81.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\alpha_{3}" + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "text", + "content": " are the coefficients of the Sobel gradient loss and the Laplace gradient loss, respectively. " + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "inline_equation", + "content": "|\\cdot |" + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "text", + "content": " refers to the absolute operation. " + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\nabla_S" + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\nabla_L" + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "text", + "content": " respectively refer to the Sobel and Laplacian gradient operator. " + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "inline_equation", + "content": "O(|x|) = x" + }, + { + "bbox": [ + 55, + 223, + 296, + 294 + ], + "type": "text", + "content": " represents the initial gradient value retrieved before the absolute value process is executed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 305, + 137, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 137, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 137, + 319 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 324, + 106, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 106, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 106, + 337 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 342, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 496 + ], + "type": "text", + "content": "Datasets. During the training stage, the MSRS [23] training set is utilized. For additional enhancement of the training set, 200 pairs are randomly chosen from the RoadScene [33] dataset as a supplement. During the evaluation stage, we use the MSRS test set. Furthermore, to assess the generalization ability of our method, we conduct tests on the " + }, + { + "bbox": [ + 55, + 342, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 55, + 342, + 296, + 496 + ], + "type": "text", + "content": " [16] and TNO [25] datasets no any re-training or fine-tuning. It's important to recognize that due to the high resolution of images in the " + }, + { + "bbox": [ + 55, + 342, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 55, + 342, + 296, + 496 + ], + "type": "text", + "content": " dataset, some contrast methods have extremely prolonged runtimes. To expedite testing, the images are resized to half their initial resolution, and the precision for all methods on " + }, + { + "bbox": [ + 55, + 342, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 55, + 342, + 296, + 496 + ], + "type": "text", + "content": " is reported based on the resized resolution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": "Implement Details. The hardware setup includes an NVIDIA TITAN RTX and a 3.6 GHz Intel Core i9-9900K, with the code implemented in PyTorch. The Adam optimizer, without weight decay, are employed to optimize the network parameters. The training phase consists of 60 epochs, a batch size of 4, an initial learning rate of " + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": ", and the learning rate is halved every 25 epochs. In the training phase, input images undergo random cropping into 128 " + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": " 128-sized segments and are transformed from RGB format to Y, Cb, and Cr. The coefficients for the terms in loss are configured as: " + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\alpha_{1} = 0.1" + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\alpha_{2} = 1" + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\alpha_{3} = 10" + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": ", which are based on the experience from our previous work. The channels at various stages is respectively defined as " + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "inline_equation", + "content": "ch\\_in = 1" + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "inline_equation", + "content": "ch\\_mid = 48" + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "inline_equation", + "content": "ch\\_out = 1" + }, + { + "bbox": [ + 55, + 498, + 296, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 234, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 234, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 234, + 685 + ], + "type": "text", + "content": "4.2. Comparison with SOTA Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Competing Methods. To validate the effectiveness of our method, we compare it with seven state-of-the-art works:" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 123, + 553, + 331 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 114 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 114 + ], + "type": "text", + "content": "Table 1. Quantitative results of our method compared to seven state-of-the-art methods on the MSRS, " + }, + { + "bbox": [ + 313, + 71, + 555, + 114 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 313, + 71, + 555, + 114 + ], + "type": "text", + "content": ", and TNO datasets. Bold indicates the best result, and underlining represents the second best." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 123, + 553, + 331 + ], + "lines": [ + { + "bbox": [ + 315, + 123, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 315, + 123, + 553, + 331 + ], + "type": "table", + "html": "
MSRS
PublisherVIF ↑SCD ↑MI ↑Qabf ↑SSIM ↑MS_SSIM ↑FMI ↑
LRRNet [13]23'TPAMI0.5440.7912.9450.4570.8590.8480.919
CoCoNet [17]24'IJCV0.6601.3872.2190.2790.7330.7470.884
MetaFusion [40]23'CVPR0.7261.4861.6840.4780.9120.9040.908
TUFusion [41]24'TCSVT0.5691.2982.5750.2290.8850.8980.909
DATFuse [24]23'TCSVT0.9061.4103.8970.6400.9550.9560.912
SwinFusion [20]22'JAS0.9901.6904.5490.6540.9680.9690.931
CDDFuse [42]23'CVPR1.0511.6215.0150.6930.9670.9660.930
Ours25'ICCV1.0611.7005.1110.7150.9810.9750.933
M²FD
PublisherVIF ↑SCD ↑MI ↑Qabf ↑SSIM ↑MS_SSIM ↑FMI ↑
LRRNet [13]23'TPAMI0.4931.4602.8750.4660.8910.8700.862
CoCoNet [17]24'IJCV0.6241.7692.8130.3990.8750.8540.853
MetaFusion [40]23'CVPR0.7011.6652.5700.3870.8540.8610.850
TUFusion [41]24'TCSVT0.4001.5022.9650.1890.7880.8640.856
DATFuse [24]23'TCSVT0.5901.2804.0380.4890.8550.8310.849
SwinFusion [20]22'JAS0.6901.5744.0970.5830.9470.9030.872
CDDFuse [42]23'CVPR0.7201.6473.9990.5960.9510.9130.871
Ours25'ICCV0.7591.6603.7710.6540.9640.9230.883
TNO
PublisherVIF ↑SCD ↑MI ↑Qabf ↑SSIM ↑MS_SSIM ↑FMI ↑
LRRNet [13]23'TPAMI0.5821.5942.5930.3630.8700.8700.886
CoCoNet [17]24'IJCV0.6551.7172.1610.2970.8070.8180.871
MetaFusion [40]23'CVPR0.6611.7301.6520.2980.7880.8150.877
TUFusion [41]24'TCSVT0.4651.5992.0940.2240.8250.8670.893
DATFuse [24]23'TCSVT0.6901.4103.1920.5030.8490.8190.878
SwinFusion [20]22'JAS0.7431.6873.2730.5160.9370.8990.900
CDDFuse [42]23'CVPR0.7831.7703.0610.5170.9290.9080.899
Ours25'ICCV0.8361.7243.5080.5880.9320.9040.914
", + "image_path": "40893697326cd8b3a50e328d502c922ce3372b43fca724b25cc65eb396795885.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 353, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 435 + ], + "type": "text", + "content": "LRRNet [13], CoCoNet [17], MetaFusion [40], TUFusion [41], DATFuse [24], SwinFusion [20], and CDDFuse [42]. CoCoNet, LRRNet, and MetaFusion are primarily CNN-based methods, while SwinFusion is dominated by the SwinTransformer architecture. TUFusion, DATFuse, CDDFuse, and our method adopt a hybrid mode that combines CNN and attention mechanisms." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "text", + "content": "Evaluation Metrics. We conduct a comprehensive analysis from both qualitative and quantitative perspectives. The qualitative comparison mainly relies on human subjective evaluation of the visual effects of the fused images, primarily concentrating on the preservation of texture details and the retention of salient thermal radiation targets. The quantitative evaluation employs seven mainstream image quality assessment metrics: visual information fidelity (VIF), the sum of the correlations of differences (SCD), mutual information (MI), gradient-based similarity measurement " + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "inline_equation", + "content": "(\\mathrm{Q_{abf}})" + }, + { + "bbox": [ + 313, + 437, + 556, + 617 + ], + "type": "text", + "content": ", structural similarity index measure (SSIM), multiscale structural similarity (MS-SSIM), and feature mutual information (FMI). All selected metrics are positive, with higher values indicating better performance. Detailed explanations can be found in [19]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 557, + 715 + ], + "type": "text", + "content": "Qualitative Comparison. The qualitative comparison results are shown in Fig. 5, which comprises three groups of images arranged from top to bottom, based on the MSRS, " + }, + { + "bbox": [ + 313, + 618, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{M}^{3}\\mathrm{FD}" + }, + { + "bbox": [ + 313, + 618, + 557, + 715 + ], + "type": "text", + "content": ", and TNO datasets, respectively. The first group of images demonstrates the ability of different methods to preserve thermal radiation information. DATFuse, LRRNet, and TUFusion fail to effectively retain thermal radiation targets, while MetaFusion, CoCoNet, and SwinFusion pro" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13518" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 160, + 137 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 160, + 137 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 160, + 137 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 160, + 137 + ], + "type": "image", + "image_path": "effaa20c8c022ff21038661abb44550f753743e79850b925fe2c9e316bf01400.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 163, + 69, + 254, + 137 + ], + "blocks": [ + { + "bbox": [ + 163, + 69, + 254, + 137 + ], + "lines": [ + { + "bbox": [ + 163, + 69, + 254, + 137 + ], + "spans": [ + { + "bbox": [ + 163, + 69, + 254, + 137 + ], + "type": "image", + "image_path": "cc0d3a547bd58ee3dab0c2fe34ee53de1aaafd3b0b59c6e47484179aead6c830.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 258, + 69, + 348, + 137 + ], + "blocks": [ + { + "bbox": [ + 258, + 69, + 348, + 137 + ], + "lines": [ + { + "bbox": [ + 258, + 69, + 348, + 137 + ], + "spans": [ + { + "bbox": [ + 258, + 69, + 348, + 137 + ], + "type": "image", + "image_path": "40f396a4bc400b5aac6470028b527f72bda26d1576c342e7f0f73161d08c5f34.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 353, + 69, + 444, + 137 + ], + "blocks": [ + { + "bbox": [ + 353, + 69, + 444, + 137 + ], + "lines": [ + { + "bbox": [ + 353, + 69, + 444, + 137 + ], + "spans": [ + { + "bbox": [ + 353, + 69, + 444, + 137 + ], + "type": "image", + "image_path": "7a2109d5b234f714b70badb6b9820970cca00aa6c246dab2cff363aaac97d4c2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 449, + 69, + 540, + 137 + ], + "blocks": [ + { + "bbox": [ + 449, + 69, + 540, + 137 + ], + "lines": [ + { + "bbox": [ + 449, + 69, + 540, + 137 + ], + "spans": [ + { + "bbox": [ + 449, + 69, + 540, + 137 + ], + "type": "image", + "image_path": "ce4c5b9602381a3ec87d6399d156a7b137fbd8cdb1d61646754403f914bf6adf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 70, + 148, + 160, + 217 + ], + "blocks": [ + { + "bbox": [ + 102, + 140, + 125, + 148 + ], + "lines": [ + { + "bbox": [ + 102, + 140, + 125, + 148 + ], + "spans": [ + { + "bbox": [ + 102, + 140, + 125, + 148 + ], + "type": "text", + "content": "Infrared" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 148, + 160, + 217 + ], + "lines": [ + { + "bbox": [ + 70, + 148, + 160, + 217 + ], + "spans": [ + { + "bbox": [ + 70, + 148, + 160, + 217 + ], + "type": "image", + "image_path": "4ca3ed4cfd5276f720bc8775224568bce96b3892c27943d2ef99c829bad9b746.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 219, + 129, + 225 + ], + "lines": [ + { + "bbox": [ + 100, + 219, + 129, + 225 + ], + "spans": [ + { + "bbox": [ + 100, + 219, + 129, + 225 + ], + "type": "text", + "content": "CDDFuse" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 163, + 148, + 253, + 217 + ], + "blocks": [ + { + "bbox": [ + 198, + 140, + 219, + 148 + ], + "lines": [ + { + "bbox": [ + 198, + 140, + 219, + 148 + ], + "spans": [ + { + "bbox": [ + 198, + 140, + 219, + 148 + ], + "type": "text", + "content": "Visible" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 163, + 148, + 253, + 217 + ], + "lines": [ + { + "bbox": [ + 163, + 148, + 253, + 217 + ], + "spans": [ + { + "bbox": [ + 163, + 148, + 253, + 217 + ], + "type": "image", + "image_path": "c1da1f1808a81971f07582c57b29eae74ead53665347c64a3787a8076bbd64e0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 219, + 225, + 226 + ], + "lines": [ + { + "bbox": [ + 192, + 219, + 225, + 226 + ], + "spans": [ + { + "bbox": [ + 192, + 219, + 225, + 226 + ], + "type": "text", + "content": "MetaFusion" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 258, + 148, + 348, + 217 + ], + "blocks": [ + { + "bbox": [ + 289, + 140, + 316, + 148 + ], + "lines": [ + { + "bbox": [ + 289, + 140, + 316, + 148 + ], + "spans": [ + { + "bbox": [ + 289, + 140, + 316, + 148 + ], + "type": "text", + "content": "DATFuse" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 148, + 348, + 217 + ], + "lines": [ + { + "bbox": [ + 258, + 148, + 348, + 217 + ], + "spans": [ + { + "bbox": [ + 258, + 148, + 348, + 217 + ], + "type": "image", + "image_path": "ca1fa735f25c3e4120ab77cee5af368349b0101fb2af628bd75c8db61e86dc7a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 219, + 319, + 226 + ], + "lines": [ + { + "bbox": [ + 286, + 219, + 319, + 226 + ], + "spans": [ + { + "bbox": [ + 286, + 219, + 319, + 226 + ], + "type": "text", + "content": "SwinFusion" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 354, + 148, + 444, + 217 + ], + "blocks": [ + { + "bbox": [ + 385, + 140, + 412, + 148 + ], + "lines": [ + { + "bbox": [ + 385, + 140, + 412, + 148 + ], + "spans": [ + { + "bbox": [ + 385, + 140, + 412, + 148 + ], + "type": "text", + "content": "CoCoNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 354, + 148, + 444, + 217 + ], + "lines": [ + { + "bbox": [ + 354, + 148, + 444, + 217 + ], + "spans": [ + { + "bbox": [ + 354, + 148, + 444, + 217 + ], + "type": "image", + "image_path": "772ccb091d965bb76b9acee8e53585dae1bad537dd41b97ab0f6c9073a98651a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 219, + 413, + 226 + ], + "lines": [ + { + "bbox": [ + 384, + 219, + 413, + 226 + ], + "spans": [ + { + "bbox": [ + 384, + 219, + 413, + 226 + ], + "type": "text", + "content": "TUFusion" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 449, + 148, + 540, + 217 + ], + "blocks": [ + { + "bbox": [ + 482, + 140, + 506, + 148 + ], + "lines": [ + { + "bbox": [ + 482, + 140, + 506, + 148 + ], + "spans": [ + { + "bbox": [ + 482, + 140, + 506, + 148 + ], + "type": "text", + "content": "LRRNet" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 148, + 540, + 217 + ], + "lines": [ + { + "bbox": [ + 449, + 148, + 540, + 217 + ], + "spans": [ + { + "bbox": [ + 449, + 148, + 540, + 217 + ], + "type": "image", + "image_path": "7e6977d8b246d9e104eba81852fb031d76ddff48593c4b7df0cf39f8fe1da95f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 219, + 501, + 226 + ], + "lines": [ + { + "bbox": [ + 487, + 219, + 501, + 226 + ], + "spans": [ + { + "bbox": [ + 487, + 219, + 501, + 226 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 71, + 234, + 161, + 301 + ], + "blocks": [ + { + "bbox": [ + 71, + 234, + 161, + 301 + ], + "lines": [ + { + "bbox": [ + 71, + 234, + 161, + 301 + ], + "spans": [ + { + "bbox": [ + 71, + 234, + 161, + 301 + ], + "type": "image", + "image_path": "60face0e156a3d77a5b30d98ff6c3352c7233a6caa1c0141e3fcb32f978fd70f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 164, + 234, + 254, + 301 + ], + "blocks": [ + { + "bbox": [ + 164, + 234, + 254, + 301 + ], + "lines": [ + { + "bbox": [ + 164, + 234, + 254, + 301 + ], + "spans": [ + { + "bbox": [ + 164, + 234, + 254, + 301 + ], + "type": "image", + "image_path": "f37d0199dc6b631174df4457d2b28d77f4ae8a9fdcbf49ef41078f5271f5c3b5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 258, + 234, + 348, + 301 + ], + "blocks": [ + { + "bbox": [ + 258, + 234, + 348, + 301 + ], + "lines": [ + { + "bbox": [ + 258, + 234, + 348, + 301 + ], + "spans": [ + { + "bbox": [ + 258, + 234, + 348, + 301 + ], + "type": "image", + "image_path": "f24e00e8fe6a55285de1652dba2a38fbdae25feaec925ac4cc2ba3b50fff96f5.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 354, + 234, + 444, + 301 + ], + "blocks": [ + { + "bbox": [ + 354, + 234, + 444, + 301 + ], + "lines": [ + { + "bbox": [ + 354, + 234, + 444, + 301 + ], + "spans": [ + { + "bbox": [ + 354, + 234, + 444, + 301 + ], + "type": "image", + "image_path": "8cf6ae43485f692d28aa462a811db6560e1070e36362fb2efc1a1e2cc7a708ca.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 451, + 234, + 540, + 301 + ], + "blocks": [ + { + "bbox": [ + 451, + 234, + 540, + 301 + ], + "lines": [ + { + "bbox": [ + 451, + 234, + 540, + 301 + ], + "spans": [ + { + "bbox": [ + 451, + 234, + 540, + 301 + ], + "type": "image", + "image_path": "482246027b02c2de08313fd27c1d679b894bcea2de85a718500b8712a9f322ff.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 71, + 312, + 161, + 380 + ], + "blocks": [ + { + "bbox": [ + 104, + 304, + 127, + 312 + ], + "lines": [ + { + "bbox": [ + 104, + 304, + 127, + 312 + ], + "spans": [ + { + "bbox": [ + 104, + 304, + 127, + 312 + ], + "type": "text", + "content": "Infrared" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 312, + 161, + 380 + ], + "lines": [ + { + "bbox": [ + 71, + 312, + 161, + 380 + ], + "spans": [ + { + "bbox": [ + 71, + 312, + 161, + 380 + ], + "type": "image", + "image_path": "1b94447b5ed31ea6703bb6ab11a636285a1a46da6996a1080d7923b568ed1a52.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 383, + 130, + 390 + ], + "lines": [ + { + "bbox": [ + 101, + 383, + 130, + 390 + ], + "spans": [ + { + "bbox": [ + 101, + 383, + 130, + 390 + ], + "type": "text", + "content": "CDDFuse" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 164, + 312, + 254, + 380 + ], + "blocks": [ + { + "bbox": [ + 199, + 304, + 219, + 312 + ], + "lines": [ + { + "bbox": [ + 199, + 304, + 219, + 312 + ], + "spans": [ + { + "bbox": [ + 199, + 304, + 219, + 312 + ], + "type": "text", + "content": "Visible" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 164, + 312, + 254, + 380 + ], + "lines": [ + { + "bbox": [ + 164, + 312, + 254, + 380 + ], + "spans": [ + { + "bbox": [ + 164, + 312, + 254, + 380 + ], + "type": "image", + "image_path": "371b840f96f31b604d3ebbca952afc44c32691780218907f805653c969daf3da.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 383, + 226, + 390 + ], + "lines": [ + { + "bbox": [ + 192, + 383, + 226, + 390 + ], + "spans": [ + { + "bbox": [ + 192, + 383, + 226, + 390 + ], + "type": "text", + "content": "MetaFusion" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 258, + 312, + 348, + 380 + ], + "blocks": [ + { + "bbox": [ + 290, + 304, + 317, + 312 + ], + "lines": [ + { + "bbox": [ + 290, + 304, + 317, + 312 + ], + "spans": [ + { + "bbox": [ + 290, + 304, + 317, + 312 + ], + "type": "text", + "content": "DATFuse" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 312, + 348, + 380 + ], + "lines": [ + { + "bbox": [ + 258, + 312, + 348, + 380 + ], + "spans": [ + { + "bbox": [ + 258, + 312, + 348, + 380 + ], + "type": "image", + "image_path": "81f49eca2289bf05a458ae9c9d9cac1e874993d7ed1f59f0c8e71a37ef39c834.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 383, + 320, + 390 + ], + "lines": [ + { + "bbox": [ + 286, + 383, + 320, + 390 + ], + "spans": [ + { + "bbox": [ + 286, + 383, + 320, + 390 + ], + "type": "text", + "content": "SwinFusion" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 354, + 312, + 444, + 380 + ], + "blocks": [ + { + "bbox": [ + 386, + 304, + 413, + 312 + ], + "lines": [ + { + "bbox": [ + 386, + 304, + 413, + 312 + ], + "spans": [ + { + "bbox": [ + 386, + 304, + 413, + 312 + ], + "type": "text", + "content": "CoCoNet" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 354, + 312, + 444, + 380 + ], + "lines": [ + { + "bbox": [ + 354, + 312, + 444, + 380 + ], + "spans": [ + { + "bbox": [ + 354, + 312, + 444, + 380 + ], + "type": "image", + "image_path": "458a400c635fabe7d1a2d7df1273b2cd2a4a753b77f5c5fbf4e430d67e32c433.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 385, + 383, + 413, + 390 + ], + "lines": [ + { + "bbox": [ + 385, + 383, + 413, + 390 + ], + "spans": [ + { + "bbox": [ + 385, + 383, + 413, + 390 + ], + "type": "text", + "content": "TUFusion" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 450, + 312, + 540, + 380 + ], + "blocks": [ + { + "bbox": [ + 482, + 304, + 507, + 312 + ], + "lines": [ + { + "bbox": [ + 482, + 304, + 507, + 312 + ], + "spans": [ + { + "bbox": [ + 482, + 304, + 507, + 312 + ], + "type": "text", + "content": "LRRNet" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 450, + 312, + 540, + 380 + ], + "lines": [ + { + "bbox": [ + 450, + 312, + 540, + 380 + ], + "spans": [ + { + "bbox": [ + 450, + 312, + 540, + 380 + ], + "type": "image", + "image_path": "2adb9d10a2b425cfdf480b30b35382b9aabc50a5693e7d820603b154fe409b22.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 383, + 503, + 390 + ], + "lines": [ + { + "bbox": [ + 488, + 383, + 503, + 390 + ], + "spans": [ + { + "bbox": [ + 488, + 383, + 503, + 390 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 71, + 398, + 161, + 465 + ], + "blocks": [ + { + "bbox": [ + 71, + 398, + 161, + 465 + ], + "lines": [ + { + "bbox": [ + 71, + 398, + 161, + 465 + ], + "spans": [ + { + "bbox": [ + 71, + 398, + 161, + 465 + ], + "type": "image", + "image_path": "229cfc223ce944daafce7c5cb61952c67931e90423825728a8261ed3aae18f91.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 164, + 398, + 254, + 465 + ], + "blocks": [ + { + "bbox": [ + 164, + 398, + 254, + 465 + ], + "lines": [ + { + "bbox": [ + 164, + 398, + 254, + 465 + ], + "spans": [ + { + "bbox": [ + 164, + 398, + 254, + 465 + ], + "type": "image", + "image_path": "88c3b0fb1403bca2df640933c2e5a1fcc85cec58b08b226ce0a9295d71eb1e97.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 260, + 399, + 350, + 466 + ], + "blocks": [ + { + "bbox": [ + 260, + 399, + 350, + 466 + ], + "lines": [ + { + "bbox": [ + 260, + 399, + 350, + 466 + ], + "spans": [ + { + "bbox": [ + 260, + 399, + 350, + 466 + ], + "type": "image", + "image_path": "cf23d8aed191fb9492e734a4c6069bdbe6823b6045190e2c77c45d7baa2b8479.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 290, + 468, + 318, + 476 + ], + "lines": [ + { + "bbox": [ + 290, + 468, + 318, + 476 + ], + "spans": [ + { + "bbox": [ + 290, + 468, + 318, + 476 + ], + "type": "text", + "content": "DATFuse" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 353, + 399, + 443, + 466 + ], + "blocks": [ + { + "bbox": [ + 353, + 399, + 443, + 466 + ], + "lines": [ + { + "bbox": [ + 353, + 399, + 443, + 466 + ], + "spans": [ + { + "bbox": [ + 353, + 399, + 443, + 466 + ], + "type": "image", + "image_path": "337668a7a80278204ae3ab722fd5307a0d499988e61de6c37ac840403aa8af99.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 468, + 413, + 476 + ], + "lines": [ + { + "bbox": [ + 386, + 468, + 413, + 476 + ], + "spans": [ + { + "bbox": [ + 386, + 468, + 413, + 476 + ], + "type": "text", + "content": "CoCoNet" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 449, + 399, + 540, + 466 + ], + "blocks": [ + { + "bbox": [ + 449, + 399, + 540, + 466 + ], + "lines": [ + { + "bbox": [ + 449, + 399, + 540, + 466 + ], + "spans": [ + { + "bbox": [ + 449, + 399, + 540, + 466 + ], + "type": "image", + "image_path": "a8e17613a1daf2722791cfec22d8ab9669a1853b3a80ba20755ef25d32c0933f.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 468, + 506, + 476 + ], + "lines": [ + { + "bbox": [ + 482, + 468, + 506, + 476 + ], + "spans": [ + { + "bbox": [ + 482, + 468, + 506, + 476 + ], + "type": "text", + "content": "LRRNet" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 70, + 478, + 160, + 544 + ], + "blocks": [ + { + "bbox": [ + 104, + 468, + 127, + 476 + ], + "lines": [ + { + "bbox": [ + 104, + 468, + 127, + 476 + ], + "spans": [ + { + "bbox": [ + 104, + 468, + 127, + 476 + ], + "type": "text", + "content": "Infrared" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 70, + 478, + 160, + 544 + ], + "lines": [ + { + "bbox": [ + 70, + 478, + 160, + 544 + ], + "spans": [ + { + "bbox": [ + 70, + 478, + 160, + 544 + ], + "type": "image", + "image_path": "239e8bda70cd922385111cb3a3f7ce5754db7f8f8e2bddd653a31d93705b0e60.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 548, + 130, + 555 + ], + "lines": [ + { + "bbox": [ + 101, + 548, + 130, + 555 + ], + "spans": [ + { + "bbox": [ + 101, + 548, + 130, + 555 + ], + "type": "text", + "content": "CDDFuse" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 570, + 555, + 593 + ], + "lines": [ + { + "bbox": [ + 55, + 570, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 555, + 593 + ], + "type": "text", + "content": "Figure 5. Visual comparison results of our method with seven state-of-the-art methods on typical image pairs from MSRS (the first group), M" + }, + { + "bbox": [ + 55, + 570, + 555, + 593 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 55, + 570, + 555, + 593 + ], + "type": "text", + "content": "FD (the second group), and TNO (the third group) datasets." + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 164, + 478, + 254, + 544 + ], + "blocks": [ + { + "bbox": [ + 199, + 468, + 220, + 476 + ], + "lines": [ + { + "bbox": [ + 199, + 468, + 220, + 476 + ], + "spans": [ + { + "bbox": [ + 199, + 468, + 220, + 476 + ], + "type": "text", + "content": "Visible" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 164, + 478, + 254, + 544 + ], + "lines": [ + { + "bbox": [ + 164, + 478, + 254, + 544 + ], + "spans": [ + { + "bbox": [ + 164, + 478, + 254, + 544 + ], + "type": "image", + "image_path": "9154111fd9a2bff71b6416310e8d9fe16e1f394843648c54f1fffe9078a9e38b.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 548, + 226, + 555 + ], + "lines": [ + { + "bbox": [ + 192, + 548, + 226, + 555 + ], + "spans": [ + { + "bbox": [ + 192, + 548, + 226, + 555 + ], + "type": "text", + "content": "MetaFusion" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 260, + 478, + 349, + 544 + ], + "blocks": [ + { + "bbox": [ + 260, + 478, + 349, + 544 + ], + "lines": [ + { + "bbox": [ + 260, + 478, + 349, + 544 + ], + "spans": [ + { + "bbox": [ + 260, + 478, + 349, + 544 + ], + "type": "image", + "image_path": "9cc03f50a44e3589d34a3b402d596b6d1d3912b70e6e92170c904697d3074447.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 548, + 321, + 555 + ], + "lines": [ + { + "bbox": [ + 288, + 548, + 321, + 555 + ], + "spans": [ + { + "bbox": [ + 288, + 548, + 321, + 555 + ], + "type": "text", + "content": "SwinFusion" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_caption" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 354, + 478, + 444, + 544 + ], + "blocks": [ + { + "bbox": [ + 354, + 478, + 444, + 544 + ], + "lines": [ + { + "bbox": [ + 354, + 478, + 444, + 544 + ], + "spans": [ + { + "bbox": [ + 354, + 478, + 444, + 544 + ], + "type": "image", + "image_path": "55dc447b27e0fd04b75b2fdce73441c8004605a10d5abb604b37cf0a65d4484a.jpg" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 548, + 413, + 555 + ], + "lines": [ + { + "bbox": [ + 384, + 548, + 413, + 555 + ], + "spans": [ + { + "bbox": [ + 384, + 548, + 413, + 555 + ], + "type": "text", + "content": "TUFusion" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_caption" + } + ], + "index": 56 + }, + { + "type": "image", + "bbox": [ + 449, + 478, + 540, + 544 + ], + "blocks": [ + { + "bbox": [ + 449, + 478, + 540, + 544 + ], + "lines": [ + { + "bbox": [ + 449, + 478, + 540, + 544 + ], + "spans": [ + { + "bbox": [ + 449, + 478, + 540, + 544 + ], + "type": "image", + "image_path": "d1136693fbf960b67e9d1a31eb37ce92212e859aa8646e9fff2b45bf574f7c5a.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 487, + 548, + 501, + 555 + ], + "lines": [ + { + "bbox": [ + 487, + 548, + 501, + 555 + ], + "spans": [ + { + "bbox": [ + 487, + 548, + 501, + 555 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "bbox": [ + 55, + 613, + 297, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 297, + 709 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 297, + 709 + ], + "type": "text", + "content": "duce distortions in the legs of the highlighted person. The second and third groups of images illustrate the capability of different methods to preserve details. For instance, neither the clouds in the second group nor the shrubberies in the third group are effectively preserved by DATFuse, LRR-Net, SwinFusion, and TUFusion, while CoCoNet and MetaFusion introduce significant noise. Only CDDFuse and our proposed method perform well in both aspects." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 313, + 613, + 555, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 555, + 709 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 555, + 709 + ], + "type": "text", + "content": "Quantitative Comparison. The quantitative comparison results are presented in Tab. 1. In the context of the MSRS dataset, our method ranks first across all evaluation metrics, demonstrating the following advantages. 1) The optimal VIF indicates that our technique yields fused outputs with enhanced visual quality, further confirming its qualitative superiority. 2) The ideal SCD, MI, and FMI suggest that our method is most effective in conveying data from the" + } + ] + } + ], + "index": 62 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13519" + } + ] + } + ], + "index": 63 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": "original input to the outcome. 3) The superior SSIM and MS-SSIM show that our model minimizes information loss and image distortion. 4) The perfect Qabf illustrates that our solution successfully preserves the maximum edge details. Furthermore, to verify the generalization competence of our method, we perform tests on the " + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": " and TNO datasets without any fine-tuning. In the context of the " + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 55, + 72, + 294, + 193 + ], + "type": "text", + "content": " dataset, our technique leads in every metric, with the sole exceptions being SCD and MI. Within the TNO dataset, it secures four top positions and two runner-up spots." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 156, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 156, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 156, + 213 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 218, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 295, + 350 + ], + "type": "text", + "content": "Within this part, we undertake multiple ablations. Each one is executed using the MSRS test dataset as a foundation. Extension. To explore the correctness of extending the attention properties of infrared images to the visible modality, two sets of controlled experiments are performed: the first set utilize only I-SCA, while the second set employ both I-SCA and V-SCA. The experimental results are shown in Tab. 2. The integration of V-SCA led to enhancements in all measurements, with the sole exception of a minor reduction in SCD and SSIM, which confirms the validity of this extension." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 483 + ], + "type": "text", + "content": "I/V-SCA. To validate the effectiveness of I-SCA and V-SCA, we design two strategies in this subsection. 1) Directly removing the proposed modules, which is denoted as w/o I/V-SCA. 2) To prevent inconsistencies in model complexity before and after the removal of I/V-SCA, which could introduce additional experimental bias, we further replace I/V-SCA with Restormer Blocks. This ablation is denoted as \"to Restormer\". Findings from the experiment are detailed in Tab. 3, which illustrates that our technique holds the top position in five of the seven assessment measures, affirming the efficacy of I/V-SCA." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 579 + ], + "type": "text", + "content": "CBSM. For confirming CBSM's efficiency, we substitute CBSM with Parameters, " + }, + { + "bbox": [ + 55, + 483, + 295, + 579 + ], + "type": "inline_equation", + "content": "Conv_{1 \\times 1}" + }, + { + "bbox": [ + 55, + 483, + 295, + 579 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 483, + 295, + 579 + ], + "type": "inline_equation", + "content": "Conv_{3 \\times 3}" + }, + { + "bbox": [ + 55, + 483, + 295, + 579 + ], + "type": "text", + "content": ", respectively. The experiment's findings are outlined in Tab. 4. Compared to the other three schemes, the introduction of CBSM results in more than half of the evaluated metrics being ranked first, which suggests that CBSM is capable of efficiently diminishing the disparities between the source image and the intermediate features of another modality." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 609, + 294, + 651 + ], + "blocks": [ + { + "bbox": [ + 69, + 590, + 281, + 601 + ], + "lines": [ + { + "bbox": [ + 69, + 590, + 281, + 601 + ], + "spans": [ + { + "bbox": [ + 69, + 590, + 281, + 601 + ], + "type": "text", + "content": "Table 2. Ablation before and after the extension of I-SCA." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 609, + 294, + 651 + ], + "lines": [ + { + "bbox": [ + 56, + 609, + 294, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 609, + 294, + 651 + ], + "type": "table", + "html": "
ConfigurationsVIFSCDMIQabfSSIMMS_SSIMFMI
I-SCAV-SCA
✓✗1.0391.7214.5710.7030.9820.9750.932
✓✓1.0611.7005.1110.7150.9810.9750.933
", + "image_path": "a93543c0dfea1b81987b445ce8dbee486441bf7baff254fe65b980d1da101c63.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 671, + 185, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 185, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 185, + 685 + ], + "type": "text", + "content": "4.4. Efficiency Comparison" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Within this part, we evaluate the parameters, FLOPs, and FPS of our work with the most competitive methods" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 91, + 553, + 133 + ], + "blocks": [ + { + "bbox": [ + 337, + 71, + 531, + 82 + ], + "lines": [ + { + "bbox": [ + 337, + 71, + 531, + 82 + ], + "spans": [ + { + "bbox": [ + 337, + 71, + 531, + 82 + ], + "type": "text", + "content": "Table 3. Ablation of the I-SCA and V-SCA modules." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 91, + 553, + 133 + ], + "lines": [ + { + "bbox": [ + 315, + 91, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 315, + 91, + 553, + 133 + ], + "type": "table", + "html": "
ConfigurationsVIFSCDMIQabfSSIMMS_SSIMFMI
w/o I/V-SCA1.0651.6695.6510.7080.9680.9670.933
to Restormer1.0631.6815.4630.7120.9750.9710.933
Ours1.0611.7005.1110.7150.9810.9750.933
", + "image_path": "3df4181b1649a537982b7ae52d31f679c1f4aefabf436c48e83e950491c5cb72.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 163, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 360, + 144, + 508, + 154 + ], + "lines": [ + { + "bbox": [ + 360, + 144, + 508, + 154 + ], + "spans": [ + { + "bbox": [ + 360, + 144, + 508, + 154 + ], + "type": "text", + "content": "Table 4. Ablation of the CBSM module." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 163, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 315, + 163, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 553, + 213 + ], + "type": "table", + "html": "
ConfigurationsVIFSCDMIQabfSSIMMS_SSIMFMI
Parameters1.0601.7085.3030.7120.9770.9730.932
Conv1×11.0261.6914.6710.6910.9750.9670.932
Conv3×31.0661.7055.2540.7110.9770.9730.932
CBSM1.0611.7005.1110.7150.9810.9750.933
", + "image_path": "966c5190c88307c67d0b1ff430aeccfe9b04d620d161950696eca4d76e1f604a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 234, + 555, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 307 + ], + "type": "text", + "content": "to highlight its speed advantage. Findings are presented in Tab. 5, where the FLOPs are determined based on an input dimension of " + }, + { + "bbox": [ + 313, + 234, + 555, + 307 + ], + "type": "inline_equation", + "content": "128 \\times 128 \\times 1" + }, + { + "bbox": [ + 313, + 234, + 555, + 307 + ], + "type": "text", + "content": ", and the FPS on " + }, + { + "bbox": [ + 313, + 234, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 313, + 234, + 555, + 307 + ], + "type": "text", + "content": " is measured with images of reduced resolution by half. Our method achieves the highest FPS, enabling seamless integration with both low- and high-level computer vision tasks." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 315, + 346, + 553, + 396 + ], + "blocks": [ + { + "bbox": [ + 355, + 327, + 512, + 338 + ], + "lines": [ + { + "bbox": [ + 355, + 327, + 512, + 338 + ], + "spans": [ + { + "bbox": [ + 355, + 327, + 512, + 338 + ], + "type": "text", + "content": "Table 5. The parameters, FLOPs, and FPS." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 346, + 553, + 396 + ], + "lines": [ + { + "bbox": [ + 315, + 346, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 315, + 346, + 553, + 396 + ], + "type": "table", + "html": "
Params (M) ↓FLOPs (G) ↓FPS ↑
MSRSM3FDTNO
SwinFusion [20]0.92717.9860.6241.0160.651
CDDFuse [42]1.18629.21339.72340.48040.179
Ours0.5659.207132.303137.271126.537
", + "image_path": "19b6a1b42a694f5218ced497841d1471a8a4d7ff8353a839e5134577928b1848.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 422, + 388, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 388, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 388, + 434 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 441, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 598 + ], + "type": "text", + "content": "This paper reveals the inherent \"attention properties\" of infrared images from the Grad-CAM of ResNet-50. To incorporate the property into IVF, we propose I-SCA and extend it to the visible modality, subsequently introducing V-SCA. The auxiliary module (CBSM) further ensures fusion quality by enhancing feature channels, enabling spatial transformations, and suppressing negative information in the source images. Additionally, the linear computations of I-SCA and V-SCA guarantee efficiency, achieving FPS of 132, 137, and 127 on MSRS, " + }, + { + "bbox": [ + 313, + 441, + 555, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{FD}" + }, + { + "bbox": [ + 313, + 441, + 555, + 598 + ], + "type": "text", + "content": ", and TNO, respectively. Moving forward, our focus will shift to crafting more specialized modules and undertaking an in-depth methodological examination to enhance the refinement of our approach." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 609, + 411, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 411, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 411, + 623 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (62272426), the Shanxi Province Science and Technology Major Special Project (202201150401021), and the Fundamental Research Program of Shanxi Province (TZLH20230818005, 202303021212189, 202303021212372, 202403021212166)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13520" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 201 + ], + "type": "text", + "content": "[2] Yaming Cao, Lei Guo, Fengguang Xiong, Liqun Kuang, and Xie Han. Physical-simulation-based dynamic template matching method for remote sensing small object detection. IEEE Transactions on Geoscience and Remote Sensing, 62: 1-14, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 204, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 204, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 204, + 295, + 247 + ], + "type": "text", + "content": "[3] Zhihao Chang, Zhixi Feng, Shuyuan Yang, and Quanwei Gao. Aft: Adaptive fusion transformer for visible and infrared images. IEEE Transactions on Image Processing, 32: 2077-2092, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 249, + 295, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 249, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 61, + 249, + 295, + 292 + ], + "type": "text", + "content": "[4] Jun Chen, Jianfeng Ding, and Jiayi Ma. Hitfusion: Infrared and visible image fusion for high-level vision tasks using transformer. IEEE Transactions on Multimedia, 26: 10145-10159, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 294, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 294, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 61, + 294, + 294, + 327 + ], + "type": "text", + "content": "[5] Chunyang Cheng, Tianyang Xu, and Xiao-Jun Wu. Mufusion: A general unsupervised image fusion network based on memory unit. Information Fusion, 92:80-92, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 295, + 360 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 363, + 295, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 363, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 61, + 363, + 295, + 417 + ], + "type": "text", + "content": "[7] Keying Du, Huafeng Li, Yafei Zhang, and Zhengtao Yu. Chitnet: A complementary to harmonious information transfer network for infrared and visible image fusion. IEEE Transactions on Instrumentation and Measurement, 74: 1-17, 2025. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 420, + 295, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 452 + ], + "type": "text", + "content": "[8] Xueyan Gao and Shiguang Liu. Bcmfuse: A bilateral cross-modal feature interaction-based network for infrared and visible image fusion. Remote Sensing, 16(17):3136, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 454, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 295, + 497 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 499, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 499, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 57, + 499, + 295, + 532 + ], + "type": "text", + "content": "[10] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 295, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 576 + ], + "type": "text", + "content": "[11] Shahid Karim, Geng Tong, Jinyang Li, Akeel Qadir, Umar Farooq, and Yiting Yu. Current advances and future perspectives of image fusion: A comprehensive review. Information Fusion, 90:185-217, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 611 + ], + "type": "text", + "content": "[12] Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE Transactions on Image Processing, 28(5):2614-2623, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "text", + "content": "[13] Hui Li, Tianyang Xu, Xiao-Jun Wu, Jiwen Lu, and Josef Kittler. Lrrnet: A novel representation learning guided fusion network for infrared and visible images. IEEE transactions on pattern analysis and machine intelligence, 45(9):11040-11052, 2023. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[14] Pengwei Liang, Junjun Jiang, Xianming Liu, and Jiayi Ma. Fusion from decomposition: A self-supervised decomposition approach for image fusion. In European Conference on Computer Vision, pages 719-735, 2022. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[15] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 206 + ], + "type": "text", + "content": "[16] Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi-modality benchmark to fuse infrared and visible for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5811, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 208, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 208, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 208, + 555, + 262 + ], + "type": "text", + "content": "[17] Jinyuan Liu, Runjia Lin, Guanyao Wu, Risheng Liu, Zhongxuan Luo, and Xin Fan. Coconet: Coupled contrastive learning network with multi-level feature ensemble for multi-modality image fusion. International Journal of Computer Vision, 132(5):1748-1775, 2024. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 320, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 352 + ], + "type": "text", + "content": "[19] Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey. Information fusion, 45:153-178, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 354, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 407 + ], + "type": "text", + "content": "[20] Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfusion: Cross-domain long-range learning for general image fusion via swim transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200-1217, 2022. 3, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 410, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 475 + ], + "type": "text", + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 477, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 477, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 317, + 477, + 553, + 542 + ], + "type": "text", + "content": "[22] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In 2017 IEEE International Conference on Computer Vision (ICCV), page 618-626. IEEE, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 544, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 586 + ], + "type": "text", + "content": "[23] Linfeng Tang, Jiteng Yuan, Hao Zhang, Xingyu Jiang, and Jiayi Ma. Piafusion: A progressive infrared and visible image fusion network based on illumination aware. Information Fusion, 83:79-92, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 589, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 633 + ], + "type": "text", + "content": "[24] Wei Tang, Fazhi He, Yu Liu, Yansong Duan, and Tongzhen Si. Datfuse: Infrared and visible image fusion via dual attention transformer. IEEE Transactions on Circuits and Systems for Video Technology, 33(7):3159-3172, 2023. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 634, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 634, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 634, + 553, + 655 + ], + "type": "text", + "content": "[25] Alexander Toet. The tno multiband image data collection. Data in brief, 15:249-251, 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 679 + ], + "type": "text", + "content": "[26] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "text", + "content": "[27] Boying Wang, Libo Zhang, Longyin Wen, Xianglong Liu, and Yanjun Wu. Towards real-world prohibited item detection: A large-scale x-ray benchmark. In Proceedings of" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13521" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "the IEEE/CVF international conference on computer vision, pages 5412-5421, 2021.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "text", + "content": "[28] Boying Wang, Ruyi Ji, Libo Zhang, Yanjun Wu, and Jing Liu. Learning to zoom: Exploiting mixed-scale contextual information for object detection. Expert Systems with Applications, 264:125871, 2025. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[29] Zhishe Wang, Yanlin Chen, Wenyu Shao, Hui Li, and Lei Zhang. Swinfuse: A residual swim transformer fusion network for infrared and visible images. IEEE Transactions on Instrumentation and Measurement, 71:1-12, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 227 + ], + "type": "text", + "content": "[30] Zhishe Wang, Wenyu Shao, Yanlin Chen, Jiawei Xu, and Xiaoqin Zhang. Infrared and visible image fusion via interactive compensatory attention adversarial learning. IEEE Transactions on Multimedia, 25:7800-7813, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "text", + "content": "[31] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 273, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 273, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 273, + 295, + 326 + ], + "type": "text", + "content": "[32] Yongbiao Xiao, Hui Li, Chunyang Cheng, and Xiaoning Song. Le2fusion: A novel local edge enhancement module for infrared and visible image fusion. In International conference on image and graphics, pages 293-305. Springer, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 295, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 371 + ], + "type": "text", + "content": "[33] Han Xu, Jiayi Ma, Junjun Jiang, Xiaojie Guo, and Haibin Ling. U2fusion: A unified unsupervised image fusion network. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):502-518, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 295, + 426 + ], + "type": "text", + "content": "[34] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Text-if: Leveraging semantic text guidance for degradation-aware and interactive image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27026-27035, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 427, + 295, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 427, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 295, + 471 + ], + "type": "text", + "content": "[35] Jun Yue, Leyuan Fang, Shaobo Xia, Yue Deng, and Jiayi Ma. Dif-fusion: Toward high color fidelity in infrared and visible image fusion with diffusion models. IEEE Transactions on Image Processing, 32:5705-5720, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 472, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 295, + 536 + ], + "type": "text", + "content": "[36] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5728-5739, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 591 + ], + "type": "text", + "content": "[37] Hao Zhang, Xuhui Zuo, Jie Jiang, Chunchao Guo, and Jiayi Ma. Mrfs: Mutually reinforcing image fusion and segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26974-26983, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 593, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 295, + 658 + ], + "type": "text", + "content": "[38] Cheng Zhao, Peng Yang, Feng Zhou, Guanghui Yue, Shuigen Wang, Huisi Wu, Guoliang Chen, Tianfu Wang, and Baiying Lei. Mhw-gan: multidiscriminator hierarchical wavelet generative adversarial network for multimodal image fusion. IEEE Transactions on Neural Networks and Learning Systems, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[39] Rong Zhao, Xie Han, Xindong Guo, Liquun Kuang, Xiaowen Yang, and Fusheng Sun. Exploring the point feature relation on point cloud for multi-view stereo. IEEE Transactions on Circuits and Systems for Video Technology, 33(11): 6747-6763, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 430 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[40] Wenda Zhao, Shigeng Xie, Fan Zhao, You He, and Huchuan Lu. Metafusion: Infrared and visible image fusion via metafeature embedding from object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13955-13965, 2023. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 183 + ], + "type": "text", + "content": "[41] Yangyang Zhao, Qingchun Zheng, Peihao Zhu, Xu Zhang, and Wenpeng Ma. Tufusion: A transformer-based universal fusion algorithm for multimodal images. IEEE Transactions on Circuits and Systems for Video Technology, 34(3): 1712-1725, 2024. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 555, + 251 + ], + "type": "text", + "content": "[42] Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5906-5916, 2023. 3, 5, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 251, + 555, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 317 + ], + "type": "text", + "content": "[43] Zixiang Zhao, Haowen Bai, Yuanzhi Zhu, Jiangshe Zhang, Shuang Xu, Yulun Zhang, Kai Zhang, Deyu Meng, Radu Timofte, and Luc Van Gool. Ddfm: denoising diffusion model for multi-modality image fusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8082-8093, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 318, + 555, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 318, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 318, + 555, + 384 + ], + "type": "text", + "content": "[44] Zixiang Zhao, Lilun Deng, Haowen Bai, Yukun Cui, Zhipeng Zhang, Yulun Zhang, Haotong Qin, Dongdong Chen, Jiangshe Zhang, Peng Wang, et al. Image fusion via vision-language model. In Proceedings of the 41st International Conference on Machine Learning, pages 60749-60765, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 385, + 555, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 385, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 385, + 555, + 430 + ], + "type": "text", + "content": "[45] Pengfei Zhu, Yang Sun, Bing Cao, and Qinghua Hu. Task-customized mixture of adapters for general image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7099-7108, 2024. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13522" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Thermal Polarimetric Multi-view Stereo/493fd467-1c08-4433-ac25-63db3b58338a_content_list.json b/2025/Thermal Polarimetric Multi-view Stereo/493fd467-1c08-4433-ac25-63db3b58338a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..79603fa8bd0445e7ba4052a0b1a16f620db017e5 --- /dev/null +++ b/2025/Thermal Polarimetric Multi-view Stereo/493fd467-1c08-4433-ac25-63db3b58338a_content_list.json @@ -0,0 +1,1853 @@ +[ + { + "type": "text", + "text": "Thermal Polarimetric Multi-view Stereo", + "text_level": 1, + "bbox": [ + 295, + 130, + 702, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Takahiro Kushida Kenichiro Tanaka Ritsumeikan University 2-150 Iwakura-cho, Ibaraki, Osaka, Japan {tkushida,ken-t}@fc.ritsumei.ac.jp", + "bbox": [ + 331, + 181, + 665, + 253 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces a novel method for detailed 3D shape reconstruction utilizing thermal polarization cues. Unlike state-of-the-art methods, the proposed approach is independent of illumination and material properties. In this paper, we formulate a general theory of polarization observation and show that long-wave infrared (LWIR) polarimetric imaging is free from the ambiguities that affect visible polarization analyses. Subsequently, we propose a method for recovering detailed 3D shapes using multi-view thermal polarimetric images. Experimental results demonstrate that our approach effectively reconstructs fine details in transparent, translucent, and heterogeneous objects, outperforming existing techniques.", + "bbox": [ + 89, + 316, + 485, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 542, + 220, + 558 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D shape reconstruction is a fundamental problem in computer vision, with applications spanning both academia and industry. Over the years, various 3D reconstruction methods using RGB cameras have been proposed, but many of these methods rely on strong assumptions about illumination and reflection models. For instance, multi-view stereo [19, 44, 45, 47] depends on surface texture to find correspondences, while structured light [16, 43], photometric stereo [10, 20, 41, 50], and shape from polarization [1, 2, 32, 34, 46, 49] require specific lighting conditions and assume opaque surfaces, limiting their applicability to a narrow range of materials.", + "bbox": [ + 89, + 568, + 482, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thermal imaging presents an attractive alternative for 3D reconstruction in challenging scenarios because it does not rely on environmental illumination; any object with heat emits long-wave infrared (LWIR) light, effectively serving as its own light source. Furthermore, most materials, except for those designed for thermal optics, are opaque in the LWIR spectrum, eliminating the need for many of the assumptions required by traditional reflection models. Building on this concept, several methods, such as thermal multi-view stereo [53] and absorption-based depth es", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "timation [9, 14, 24, 35] have been proposed. Although these methods demonstrate potential applicability, their results also imply the limitation inherent in model-based approaches, especially with regard to accuracy.", + "bbox": [ + 511, + 287, + 903, + 348 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photometric cues in the LWIR spectrum are useful for recovering the accurate shape, similar to the visible-light spectrum. For example, thermal photometric stereo [48] and shape from heat conduction [36] have demonstrated detailed surface reconstruction. However, these methods require active heating and cooling of the object, making them both time-consuming and impractical. In this paper, we aim to reconstruct fine 3D shapes in a steady state using polarimetric cues, a type of photometric cue that can be obtained without monitoring the heating or cooling processes.", + "bbox": [ + 511, + 349, + 906, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Table 1 presents a comparison of various 3D reconstruction approaches. Our approach is illumination-independent, material-independent, free from heating or cooling process, and highly accurate. Our contributions are twofold. First, we establish a unified theoretical framework for LWIR polarization, showing that, unlike visible-light polarization, it is not affected by specular-diffuse ambiguities. Second, we show that LWIR polarization serves as a powerful cue for detailed 3D shape reconstruction. Experimental results confirm that our method outperforms the existing approaches on heterogeneous materials.", + "bbox": [ + 511, + 501, + 908, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 684, + 655, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2.1. Shape from visible polarization", + "text_level": 1, + "bbox": [ + 511, + 709, + 790, + 726 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shape from Polarization (SfP), which recovers surface normals from the polarization state of reflected light, has a long-standing history in computer vision [1, 2, 32, 34, 46, 49]. Recently, the advent of snapshot polarization cameras has significantly boosted this area of research [40].", + "bbox": [ + 511, + 732, + 903, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SfP approaches continue to tackle ambiguities in surface normals that arise from the mixture of specular and diffuse polarizations [46]. One effective way to resolve these ambiguities is to use multi-view polarization images, which provide a richer set of constraints on the surface normals and lead to more accurate and robust surface reconstruc", + "bbox": [ + 511, + 810, + 906, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27390", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/9d66793122a34d89ecbe9640ea973fc3e7fbbbbf5e596ba516918561714f5992.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIlluminationMaterialAccuracyMeasurement
Visible Multi-view Stereo [19, 44, 45, 47]DependentTextured onlyHighEasy
Visible SL / PS [10, 16, 20, 41, 43, 50]DependentOpaque onlyHighEasy
Visible Polarization [1, 2, 32, 34, 46, 49]DependentDependentHighEasy
Visible NeRFs [30, 51]DependentWide range of materialsLowEasy
Thermal Multi-view Stereo [53]IndependentTextured onlyHighEasy
Thermal SL [11]IndependentDiffuse reflective onlyHighEasy
Thermal PS [48]IndependentWide range of materialsModerateHeating & cooling
Thermal NeRFs [28, 52]IndependentWide range of materialsLowEasy
Heat Conduction [36]IndependentWide range of materialsModerateHeating
Thermal Polarization (Ours)IndependentWide range of materialsHighEasy
", + "bbox": [ + 168, + 87, + 828, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison table of shape reconstruction methods. While other methods have pros and cons, our approach is independent of illumination and material's optical properties as well as accurate shape recovery is possible in steady state measurements.", + "bbox": [ + 89, + 234, + 906, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions [8]. Another approach involves using polarimetric BRDF models, which can handle complex material properties [3, 13, 21, 55]. Recent approaches incorporate neural scene representations and differentiable rendering, enabling the joint estimation of shape, environment maps, and reflectance [18, 27]. Nevertheless, SfP in the visible light spectrum remains fundamentally sensitive to the illumination conditions, such as the assumption of uniform, unpolarized ambient light, and to variations in material's optical properties, including transparency, translucency, matte, or shiny finishes [46].", + "bbox": [ + 88, + 289, + 480, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce the use of polarization in the LWIR spectrum. The LWIR light is emitted directly from the object and is independent of external illumination. Moreover, the polarization of LWIR emission is not mixed up with the diffuse and specular reflections or transmission effects. These properties make LWIR polarization advantageous for robust shape reconstruction.", + "bbox": [ + 89, + 455, + 482, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. LWIR imaging and physics-based vision", + "text_level": 1, + "bbox": [ + 89, + 574, + 434, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond 3D reconstruction, thermal imaging is utilized for a variety of physics-based vision tasks. The analysis of heat conduction, for instance, has been leveraged for material classification by exploiting differences in thermal diffusivity and conductivity [42]. Reflection analysis of LWIR light has facilitated non-line-of-sight (NLoS) imaging, expanding scene understanding by examining how thermal rays propagate around the corner [23, 29]. Joint analyses of both visible and LWIR light transport have further enriched scene understanding [38]. We also advance the area of physics-based thermal vision through a shape reconstruction using LWIR polarization.", + "bbox": [ + 89, + 597, + 482, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thermal imaging is an effective approach for 3D shape reconstruction in some challenging scenarios. For example, it is effective for stealth observation as the thermal radiation does not rely on external illumination [6, 35, 52, 53]. Another example is that it is feasible to observe the shape of challenging materials in the visible light spectrum because many materials that appear transparent or translucent are opaque in the LWIR spectrum [48].", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing 3D reconstruction techniques using thermal cameras can be broadly categorized into geometric and photometric approaches. Geometric methods [12, 31, 53] such as based on stereo/multi-view observations recover surface depth by matching corresponding points across thermal images, commonly assuming near-Lambertian emission or sufficient texture. Photometric approaches infer depth, shape, or surface normal from intensity cues, which include thermal photometric stereo [48], absorption-based method [9, 14, 24, 35], and heat conduction-based method [36]. While they show strong possibilities of shape reconstruction using a thermal camera, their accuracy is generally limited. There are hybrid approaches that combine visible-light and LWIR polarization [34], where the LWIR polarization is used only as an auxiliary cue to resolve the specular-diffuse ambiguity in visible polarization rather than directly contributing to shape recovery.", + "bbox": [ + 511, + 289, + 903, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, neural rendering methods inspired by Neural Radiance Fields (NeRF) have been explored in the thermal domain [6, 28, 52] as well. By jointly optimizing a volumetric scene representation and a rendering function to satisfy the rendered images and align observed images from multiple viewpoints, these thermal NeRF techniques recover both geometry and radiometric properties. We are also inspired by this neural representation and model the object's surface by a neural implicit surface.", + "bbox": [ + 511, + 546, + 905, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we leverage the polarization state of thermal radiation to enrich the information available for 3D reconstruction. By incorporating the polarization cues, geometric rotation, and neural representation, our method enables more accurate shape estimation.", + "bbox": [ + 511, + 681, + 905, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Polarimetric LWIR Observation", + "text_level": 1, + "bbox": [ + 511, + 768, + 808, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we present a general theory of polarization observation and discuss the advantages of thermal polarization over visible polarization.", + "bbox": [ + 511, + 795, + 903, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Physics of Polarization", + "text_level": 1, + "bbox": [ + 511, + 848, + 723, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The polarization state of light, including both visible and LWIR spectra, changes when it is reflected from or trans", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27391", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/862fcd43ab96a685b0f2e6a38c8d26efb11173cd27ce0f73f8bca444bdb22e57.jpg", + "image_caption": [ + "Figure 1. Coordinate systems and the zenith and projected azimuth angles of the surface normal. While Mueller calculus is performed in the object's surface coordinate system, the observed Stokes parameter and projected azimuth angle are represented in the image coordinate system." + ], + "image_footnote": [], + "bbox": [ + 140, + 88, + 433, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mitted through an object's surface. The transport of polarized light can be described using the Stokes parameters $\\boldsymbol{s} = [s_0,s_1,s_2,s_3]^{\\mathrm{T}}$ and the Mueller matrix $M$ as", + "bbox": [ + 89, + 392, + 483, + 439 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {o} = \\boldsymbol {M} \\boldsymbol {s} _ {i}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 246, + 449, + 480, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $s_i$ and $s_o$ denote the Stokes parameters of the incoming and outgoing light, respectively, both defined in the surface coordinate system [7].", + "bbox": [ + 89, + 474, + 483, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The observed scene is a combination of specular reflection, diffuse reflection, transmission, and emission. Consequently, the overall polarization state $s$ can be expressed as", + "bbox": [ + 89, + 520, + 483, + 579 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} = \\boldsymbol {s} _ {s} + \\boldsymbol {s} _ {d} + \\boldsymbol {s} _ {t} + \\boldsymbol {s} _ {e}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 593, + 480, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $s_s, s_d, s_t$ , and $s_e$ represent the Stokes vectors corresponding to the specular, diffuse, transmission, and emission components, respectively.", + "bbox": [ + 89, + 618, + 483, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specular polarization The change in the polarization state due to specular reflection [3, 22] is described by", + "bbox": [ + 89, + 681, + 483, + 712 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {s} = \\boldsymbol {C} (\\varphi) \\boldsymbol {F} ^ {\\mathrm {R}} (\\theta) \\boldsymbol {s} _ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 720, + 480, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $s_i$ denotes the Stokes vector of the incident light in the object's surface coordinate system, $s_s$ denotes the corresponding Stokes vector in the image coordinate system, $C(\\varphi)$ is a rotation matrix that transforms coordinates from the image to the object's surface, $\\varphi$ is the projected azimuth angle (i.e. the angle between the vertical axis of the image plane and the surface normal projected into the image plane), $F^{\\mathrm{R}}(\\theta)$ is the Mueller matrix for the specular component, and $\\theta$ is the zenith angle of the reflection (i.e. the angle between the surface normal and the viewing direction). The", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "surface coordinate system, image coordinate system, zenith angle, and projected azimuth angle are illus in Fig. 1. The rotation matrix $C(\\varphi)$ is given by", + "bbox": [ + 511, + 92, + 905, + 137 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {C} (\\varphi) = \\left[ \\begin{array}{c c c c} 1 & 0 & 0 & 0 \\\\ 0 & \\cos 2 \\varphi & - \\sin 2 \\varphi & 0 \\\\ 0 & \\sin 2 \\varphi & \\cos 2 \\varphi & 0 \\\\ 0 & 0 & 0 & 1 \\end{array} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 145, + 905, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Mueller matrix $\\pmb{F}^{\\mathrm{R}}(\\theta)$ is derived from Fresnel's law and expressed as", + "bbox": [ + 511, + 219, + 905, + 250 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} ^ {\\mathrm {R}} (\\theta) = \\left[ \\begin{array}{l l l l} r _ {+} & r _ {-} & 0 & 0 \\\\ r _ {-} & r _ {+} & 0 & 0 \\\\ 0 & 0 & r _ {\\times} \\cos \\delta & r _ {\\times} \\sin \\delta \\\\ 0 & 0 & - r _ {\\times} \\sin \\delta & r _ {\\times} \\cos \\delta \\end{array} \\right], \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 258, + 905, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $r_{\\pm} = \\frac{r_s \\pm r_p}{2}$ , $r_{\\times} = \\sqrt{r_p r_s}$ , $\\theta'$ denotes the refractive angle. Here, $\\cos \\delta$ equals $-1$ when $\\theta$ is less than Brewster's angle and $1$ otherwise, with the sign of $\\sin \\delta$ reversed accordingly [3]. The Fresnel reflectances for s- and p-polarization, $r_s$ and $r_p$ , are given by", + "bbox": [ + 511, + 333, + 905, + 411 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nr _ {s} (\\theta) = \\left(\\frac {\\cos \\theta - \\eta \\cos \\theta^ {\\prime}}{\\cos \\theta + \\eta \\cos \\theta^ {\\prime}}\\right) ^ {2}, r _ {p} (\\theta) = \\left(\\frac {\\cos \\theta^ {\\prime} - \\eta \\cos \\theta}{\\cos \\theta^ {\\prime} + \\eta \\cos \\theta}\\right) ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 419, + 921, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to Snell's law, the refractive angle is given as", + "bbox": [ + 513, + 479, + 883, + 496 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {\\prime} = \\sin^ {- 1} \\left(\\frac {1}{\\eta} \\sin \\theta\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 505, + 905, + 539 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\eta$ represents the refractive index of the object.", + "bbox": [ + 511, + 547, + 859, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffuse polarization The change in the polarization state due to diffuse reflection or subsurface scattering [1, 22] is described by", + "bbox": [ + 511, + 580, + 905, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {d} = \\boldsymbol {C} (\\varphi) \\boldsymbol {F} ^ {\\mathrm {T}} (\\theta) \\int_ {\\Omega} \\boldsymbol {D} \\boldsymbol {F} ^ {\\mathrm {T}} (\\theta) \\boldsymbol {s} _ {i, \\omega} d \\omega , \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 633, + 905, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $s_{i,\\omega}$ denotes the Stokes vector of incident light coming from the direction $\\omega$ . The matrix $D$ , which describes the depolarization, is defined as", + "bbox": [ + 511, + 674, + 905, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {D} = \\left[ \\begin{array}{l l l l} \\rho & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\end{array} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 728, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\rho$ denotes the proportion of the depolarization. The Mueller matrix for transmittance $\\pmb{F}^{\\mathrm{T}}(\\theta)$ is given by", + "bbox": [ + 511, + 800, + 905, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} ^ {\\mathrm {T}} (\\theta) = \\left[ \\begin{array}{c c c c} t _ {+} & t _ {-} & 0 & 0 \\\\ t _ {-} & t _ {+} & 0 & 0 \\\\ 0 & 0 & t _ {\\times} \\cos \\delta & t _ {\\times} \\sin \\delta \\\\ 0 & 0 & - t _ {\\times} \\sin \\delta & t _ {\\times} \\cos \\delta \\end{array} \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 840, + 903, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27392", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t_{\\pm} = \\frac{t_s \\pm t_p}{2}$ , $t_{\\times} = \\sqrt{t_p t_s}$ . $t_p$ and $t_s$ denote the Fresnel transmittance coefficients for s- and p- polarization, respectively, and are expressed as", + "bbox": [ + 89, + 89, + 483, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {s} (\\theta) = \\left(\\frac {2 \\cos \\theta}{\\cos \\theta + \\eta \\cos \\theta^ {\\prime}}\\right) ^ {2}, t _ {p} (\\theta) = \\left(\\frac {2 \\cos \\theta}{\\cos \\theta^ {\\prime} + \\eta \\cos \\theta}\\right) ^ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 141, + 496, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transmission polarization The polarization state also changes due to light transmission. According to the Fresnel's law, the transmission polarization is described by", + "bbox": [ + 89, + 208, + 483, + 253 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {t} = \\boldsymbol {C} (\\varphi) \\boldsymbol {F} ^ {\\mathrm {T}} (\\theta) \\boldsymbol {s} _ {i}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 258, + 483, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_i$ denotes the Stokes vector of the light coming from under the surface.", + "bbox": [ + 89, + 284, + 483, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Emission polarization The polarization state of emitted light is described by", + "bbox": [ + 89, + 330, + 483, + 362 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {e} = \\boldsymbol {C} (\\varphi) \\boldsymbol {E} (\\theta) \\boldsymbol {s} _ {c}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 369, + 482, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_c$ denotes the Stokes vector of the emitted light including thermal radiation as well as any chemical or electrical emissions and $E(\\theta)$ is the Mueller matrix corresponding to the emission component. Since there are many types of light emission, the elements of the emission Mueller matrix are not specified explicitly in general.", + "bbox": [ + 89, + 392, + 483, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. LWIR Polarization Observation", + "text_level": 1, + "bbox": [ + 89, + 491, + 372, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Polarization of visible light Since objects other than light sources do not emit light in the visible spectrum, the emission component is generally ignorable. Therefore, the observed polarization can be modeled as", + "bbox": [ + 89, + 512, + 483, + 573 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} = \\boldsymbol {s} _ {d} + \\boldsymbol {s} _ {s} + \\boldsymbol {s} _ {t}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 582, + 480, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The magnitude of each component varies in complex ways depending on the object's material properties and the illumination environment. This complexity makes polarization analyses more challenging, leading existing visible polarization studies to rely on various assumptions to facilitate the analysis.", + "bbox": [ + 89, + 603, + 483, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Polarization of LWIR light Most objects are known to be opaque in the LWIR spectrum; therefore, the transmission component can be safely ignored. Moreover, an interesting property of LWIR light polarization is that the relative contributions of each component can be controlled by the object's temperature. In a typical environments, where no objects are extremely hot in the surroundings (i.e., where there are no significant LWIR light sources in the surroundings) the magnitude of reflection components is small compared to that of the emission. In such cases, the LWIR polarization observation can be modeled as", + "bbox": [ + 89, + 710, + 483, + 877 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns = s _ {e}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 259, + 886, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since LWIR emission primarily arises from black body radiation, Eq. (13) can be rewritten as", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {e} = \\boldsymbol {C} (\\varphi) \\boldsymbol {\\mathcal {E}} (\\theta) \\boldsymbol {s} _ {B} (T), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 128, + 903, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{E}(\\theta)$ denotes the Mueller matrix for black body radiation, and $s_B(T)$ represents the Stokes vector of the radiation, which can be expressed as", + "bbox": [ + 511, + 150, + 903, + 195 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {B} (T) = \\left[ \\begin{array}{c} B (T) \\\\ 0 \\\\ 0 \\\\ 0 \\end{array} \\right], \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 200, + 903, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $B$ denotes the black body radiation. According to Planck's law, $B$ is given by", + "bbox": [ + 511, + 268, + 903, + 299 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nB (T) = \\frac {2 h c ^ {2}}{\\lambda^ {5}} \\frac {1}{e ^ {\\frac {h c}{\\lambda k T}} - 1}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 303, + 903, + 337 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ is the speed of light, $k$ is the Boltzmann constant, $h$ is the Planck constant, $T$ is the object's absolute temperature, and $\\lambda$ is the observation wavelength. Equation (17) demonstrates that the object's temperature $T$ controls the power of emission.", + "bbox": [ + 511, + 340, + 903, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Mueller matrix of black body radiation $\\mathcal{E}(\\theta)$ [26, 33] is given by", + "bbox": [ + 511, + 417, + 903, + 446 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mathcal {E}} (\\theta) = \\left[ \\begin{array}{l l l l} \\varepsilon_ {+} & 0 & 0 & 0 \\\\ \\varepsilon_ {-} & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\end{array} \\right], \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 446, + 903, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\varepsilon_{\\pm} = \\frac{\\varepsilon_s \\pm \\varepsilon_p}{2}$ . $\\varepsilon_p$ and $\\varepsilon_s$ denote the emissivities for s- and p-polarization, respectively. According to the energy conservation law, also known as Kirchhoff's law, the sum of reflectance, transmittance, and emissivity is equal to 1. Since the transmittance is negligible, as discussed above, we obtain", + "bbox": [ + 511, + 513, + 905, + 603 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\varepsilon_ {s} (\\theta) = 1 - r _ {s} (\\theta) \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 611, + 903, + 627 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\varepsilon_ {p} (\\theta) = 1 - r _ {p} (\\theta). \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 630, + 903, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, in the LWIR spectrum, the Stokes observation $s$ can be analytically represented without ambiguity, thereby simplifying the analysis of the polarization state, which is a significant advantage for shape reconstruction.", + "bbox": [ + 511, + 652, + 905, + 713 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Polarization to surface normal: LWIR polarization provides cues for an object's surface normal, which can be determined analytically. The degree of linear polarization (DoLP) is related to the zenith angle $\\theta$ as follows:", + "bbox": [ + 511, + 729, + 905, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {D o L P} = \\frac {\\sqrt {s _ {1} ^ {2} + s _ {2} ^ {2}}}{s _ {0}} (22) \\\\ = \\frac {\\left(\\eta - \\frac {1}{\\eta}\\right) ^ {2} \\sin^ {2} \\theta}{2 + 2 \\eta^ {2} - \\left(\\eta + \\frac {1}{\\eta}\\right) \\sin^ {2} \\theta + 4 \\cos \\theta \\sqrt {\\eta^ {2} - \\sin^ {2} \\theta}}. (23) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 795, + 911, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27393", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0a83ce182410d44b86e138c9b93a2f722fbf8ac0ba166d6d29ef2fd27e0d956e.jpg", + "image_caption": [ + "Figure 2. Visible and thermal polarization observations of spheres made from various materials under natural room lighting. While visible polarization is unstable due to variations in a material's optical properties, thermal polarimetric cues, especially AoLP images, remain consistent across different materials." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 480, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The angle of linear polarization (AoLP) is equivalent to the azimuth angle $\\varphi$ as follows:", + "bbox": [ + 89, + 407, + 483, + 439 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A o L P} = \\frac {1}{2} \\tan^ {- 1} \\left(\\frac {s _ {2}}{s _ {1}}\\right) = \\varphi . \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 449, + 482, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These correspondences enable the reconstruction of the surface normal from polarization observations. Especially, AoLP serves as a robust cue because it directly reflects the azimuth angle of the surface normal and is independent of material properties, whereas DoLP is material-dependent due to the inclusion of the refractive index term.", + "bbox": [ + 89, + 491, + 483, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Examples of Visible vs. LWIR Polarization", + "text_level": 1, + "bbox": [ + 89, + 590, + 455, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the visible light spectrum, specular, diffuse, and transmission polarization components are often mixed [2, 25, 46], resulting in complicated observation. Fig. 2 shows the polarimetric observations of spheres made from various materials; in the visible spectrum, unwanted polarization observations frequently appear. For example, because black stone exhibits strong specular polarization, the DoLP image is heavily affected by reflections from the surrounding environment. In the case of transparent glass sphere, the observation is notably contaminated by background textures. For translucent plastic materials, both DoLP and AoLP images are affected by surface texture and subsurface scattering. These unstable observations significantly complicate subsequent shape recovery.", + "bbox": [ + 89, + 613, + 482, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast, in the LWIR spectrum, polarimetric observations remain consistent and independent of both material properties and environmental reflections, particularly in AoLP images. This consistency is a key advantage of thermal polarization cues, as it provides robustness against vari", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ations in materials and surrounding reflections.", + "bbox": [ + 513, + 90, + 823, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Thermal Polarimetric Multi-view Stereo", + "text_level": 1, + "bbox": [ + 511, + 119, + 875, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our goal is to recover the 3D shape using thermal polarimetric cues. As discussed in the previous section, the AoLP observation is invariant to material properties. To leverage this strong cue, we only use the AoLP images for material-independent 3D shape reconstruction.", + "bbox": [ + 511, + 146, + 905, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach employs multi-view thermal AoLP images to estimate the Signed Distance Function (SDF), which is implicitly represented by a Multi-Layer Perceptron (MLP) optimized via differentiable rendering. To effectively incorporate the surface normal information present in AoLP images, we adopt the concept of tangent space constraint (TSC) proposed in a multi-view azimuth stereo (MVAS) [5].", + "bbox": [ + 511, + 222, + 906, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Shape representation", + "text_level": 1, + "bbox": [ + 511, + 354, + 681, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The SDF is implicitly represented by an MLP $f(\\pmb{x}; \\pmb{\\theta}) : \\mathbb{R}^3 \\times \\mathbb{R}^d \\to \\mathbb{R}$ , where $\\pmb{x} \\in \\mathbb{R}$ denotes a point in space and $\\pmb{\\theta} \\in \\mathbb{R}^d$ are the learnable parameters of the MLP. The zero-level set of the SDF defines the object's surface $\\mathcal{M}$ as [37]", + "bbox": [ + 511, + 375, + 905, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} (\\boldsymbol {\\theta}) = \\{\\boldsymbol {x} | f (\\boldsymbol {x}; \\boldsymbol {\\theta}) = 0 \\}. \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 452, + 903, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The MLP is optimized within the framework of implicit differentiable renderer (IDR) [51].", + "bbox": [ + 511, + 476, + 905, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss function", + "text_level": 1, + "bbox": [ + 511, + 517, + 622, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt the loss function proposed in MVAS [5], which is defined as,", + "bbox": [ + 511, + 539, + 905, + 568 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {T S C}} + \\lambda_ {1} \\mathcal {L} _ {\\text {s i l h o u e t t e}} + \\lambda_ {2} \\mathcal {L} _ {\\text {E i k o n a l}}, \\tag {26}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 582, + 903, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{TSC}}$ is the tangent space consistency loss, $\\mathcal{L}_{\\mathrm{silhouette}}$ is the silhouette loss and $\\mathcal{L}_{\\mathrm{Eikonal}}$ is the Eikonal regularization. $\\lambda_{1}$ and $\\lambda_{2}$ are the weights assigned to the respective loss terms.", + "bbox": [ + 511, + 611, + 905, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tangent space consistency loss [5] The AoLP captured from different camera positions enforces a strong constraint on surface normal estimation. For a unit normal vector $\\pmb{n} \\subset S^2 \\in \\mathbb{R}^3$ at a surface point $\\pmb{x} \\in \\mathbb{R}^3$ and the projected azimuth angle $\\varphi$ at the corresponding pixel, the following relationship holds:", + "bbox": [ + 511, + 691, + 905, + 782 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {r} _ {1} ^ {\\top} \\boldsymbol {n} \\cos \\varphi - \\boldsymbol {r} _ {2} ^ {\\top} \\boldsymbol {n} \\sin \\varphi = 0, \\tag {27}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 792, + 903, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{R} = [r_1, r_2, r_3]^\\top$ denotes the rotation matrix of the camera pose. This equation can be rearranged as", + "bbox": [ + 511, + 823, + 903, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {n} ^ {\\top} \\underbrace {\\left(\\boldsymbol {r} _ {1} \\cos \\varphi - \\boldsymbol {r} _ {2} \\sin \\varphi\\right)} _ {\\boldsymbol {t} (\\varphi)} = 0, \\tag {28}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 864, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27394", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t(\\varphi)$ is called a projected tangent vector and the above equation is called tangent space constraint. The TSC loss is the sum of the square error of the tangent space constraint and is defined as", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T S C}} = \\frac {1}{P} \\sum_ {\\boldsymbol {x} \\in \\boldsymbol {X}} \\frac {\\sum_ {i = 1} ^ {N} \\Phi_ {i} \\left(\\boldsymbol {n} ^ {\\top} (\\boldsymbol {x} ; \\boldsymbol {\\theta}) \\boldsymbol {t} _ {i} (\\boldsymbol {x})\\right) ^ {2}}{\\sum_ {i = 1} ^ {N} \\Phi_ {i}} \\tag {29}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 154, + 482, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P$ denotes the number of pixel samples, $\\Phi_{i}$ is a binary indicator of the visibility of point $\\pmb{x}$ from the $i$ -th view, $\\pmb{n}^{\\top}(\\pmb{x};\\pmb{\\theta})$ is the estimated surface normal at the point $\\pmb{x}$ given network parameters $\\pmb{\\theta}$ , and $\\pmb{t}_{i}(\\pmb{x})$ is the tangent vector corresponding to the $i$ -th view's pixel onto which the point $\\pmb{x}$ is projected.", + "bbox": [ + 89, + 200, + 483, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Silhouette loss [5, 51] The silhouette loss constrains the visual hull of the shape and is defined as", + "bbox": [ + 89, + 309, + 483, + 339 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i l h o u e t t e}} = \\frac {1}{\\alpha P} \\sum_ {\\boldsymbol {x} \\in \\hat {\\boldsymbol {X}}} \\Psi (O (\\Pi (\\boldsymbol {x})), \\sigma (\\alpha f ^ {*})), \\tag {30}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 344, + 482, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\Psi$ denotes the cross entropy and $O$ is a binary mask indicating whether a pixel lies inside the object silhouette (as provided by the input). $P$ is the number of pixel samples, $\\Pi$ is the projection from the world coordinate to the image coordinate, $\\tilde{\\pmb{X}}$ represents the pixels with no ray-surface interactions or outside the silhouette, $f^{*}$ represents the minimum SDF distance between the ray that does not have intersections and the object surface, $\\sigma$ is a sigmoid function, and $\\alpha$ controls the sharpness of the sigmoid function.", + "bbox": [ + 89, + 388, + 483, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Eikonal regularization [17] The Eikonal loss regularizes the gradient of SDF, enforcing the SDF to maintain its property of having a gradient close to 1:", + "bbox": [ + 89, + 556, + 483, + 602 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {E i k o n a l}} = \\mathbb {E} _ {\\boldsymbol {x}} \\left(\\left(\\left\\| \\boldsymbol {n} (\\boldsymbol {x}; \\boldsymbol {\\theta}) \\right\\| _ {2} - 1\\right) ^ {2}\\right). \\tag {31}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 607, + 482, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbb{E}_x(\\cdot)$ denotes the expected value computed over sample points $\\pmb{x}$ that are randomly sampled within the bounding box.", + "bbox": [ + 89, + 631, + 482, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 686, + 223, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 712, + 171, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Following MVAS [5], we use an 8-layer MLP with softmax activations, combined with positional encoding (dimension 10). The input 3D position and encoded features are skip-connected to the 4th layer of the network. The optimization is performed using the Adam optimizer with a learning rate of 0.001. The weights for the loss terms are set to $\\lambda_{1} = 50$ (silhouette loss) and $\\lambda_{2} = 0.1$ (Eikonal loss). Training is conducted for 50 epochs with a batch size of 4096 pixels, and the silhouette loss weight is halved every 10 epochs. Training takes approximately 2 hours on a single RTX A6000 GPU.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3641a6f9bb07ad68ad30c19c074f61a0cd205f1688ea98c3f7372abdb2989620.jpg", + "image_caption": [ + "Figure 3. Camera system and setup. The system consists of a rotation stage to rotate the target and a thermal camera with a wire-grid polarizer. A visible polarization camera is used for camera pose estimation and for comparison with the visible MVAS." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 905, + 299 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Camera system Figure 3 shows our camera system and experimental setup. Our thermal polarization camera system consists of a thermal camera (FLIR Boson 320) with $15\\mathrm{mm}$ lens and a wire-grid polarizer (Thorlabs WP25M-IRC) mounted on a motorized rotation mount (Thorlabs K10CR2). The polarizer is installed between the image sensor and imaging lens to mitigate the narcissus reflections. Thermal polarization images are captured by rotating the polarizer using the rotation mount. We captured four polarization images at 0, 45, 90, and 135 degrees. A monochrome visible-light polarization camera (FLIR Blackfly S BFS-U3-51S5P) is also placed adjacent to the thermal camera. The visible-light polarization camera is used both to capture images for a baseline method and to get the poses of both visible and thermal cameras.", + "bbox": [ + 511, + 392, + 906, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Camera poses We employed a marker-based approach to estimate the camera poses. We put Aruco markers [15] on the rotation stage to estimate the pose of visible-light camera. Since printed markers are not visible in the thermal images, the poses of the thermal camera were obtained by transforming the camera poses of the visible-light camera. For stereo calibration between the visible-light and thermal cameras, we used a calibration target that can be observed by both cameras. The calibration board consists of a white-painted aluminum pegboard placed in front of a black surface heater. By heating the back panel, the circle grid pattern became visible to both cameras. The calibration parameters were estimated using Zhang's method [54].", + "bbox": [ + 511, + 638, + 908, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset We prepared a real-world dataset with ground-truth shapes. We captured 7 objects with various shapes and materials. The target objects include transparent and", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27395", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3dc83f94bbd3d1bc082ad55d0fdc720bf6979f8c241ef121a8acf3b9bdb3ded5.jpg", + "image_caption": [ + "Figure 4. Experimental results. AoLP images in both visible and LWIR spectra are shown below the scene photograph. Each method's result includes the recovered shape, surface normals, and angular error maps. The Chamfer distance is displayed below the estimated shape and the mean angular error is displayed to the right of the angular error map. For both metrics, lower values indicate better performance. The smallest errors are shown in magenta. It demonstrates that the proposed method outperforms the other methods." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 906, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "low-reflective objects, which pose challenges for existing reconstruction methods. The objects were placed on a rotating table and captured from about 20–30 views by rotating the table. To robustly capture thermal polarization, the objects were kept warm during the measurements. Masks for thermal images were generated by thresholding pixel values, while masks for visible images were generated by using Segment Anything Model 2 [39]. Ground-truth shapes were obtained using a structured light 3D scanner (EinScan-SP V2). To facilitate the ground-truth measurement, the surfaces of the objects were painted white to make them opaque.", + "bbox": [ + 89, + 599, + 483, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines We compare our proposed method with two state-of-the-art methods: MVAS [5] based on visible-light polarization cues and IDR [51], which is one of the neural implicit surface reconstruction methods, with thermal image input. For MVAS with visible-light polarization images, there exists $\\pm \\frac{\\pi}{2}$ ambiguity in surface normals due to", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the mixture of specular and diffuse polarization. Therefore, TSC loss is modified to account for the $\\pm \\frac{\\pi}{2}$ ambiguity, as originally proposed in MVAS. Since our method does not exhibit this ambiguity, the modified loss function is applied only to visible light observation.", + "bbox": [ + 511, + 599, + 906, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics We use two metrics to evaluate both the surface points and surface normals. Chamfer distance is the metric for evaluating the geometric error, which calculates the point cloud distance between the estimate and the ground-truth. The mean angular error is the metric for surface normal, which calculates the angular error of normal maps. For both metrics, a smaller value indicates a better result. Prior to computing these metrics, the estimated and the ground-truth meshes are aligned using ICP algorithm [4].", + "bbox": [ + 511, + 699, + 906, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 511, + 847, + 607, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 shows the results. The target objects are an owl made of ceramic with a clear coating, a glass container with", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27396", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/999293604978cbac00a51ab636cb2162f1933fb058b2908e72623c1e9200977d.jpg", + "image_caption": [ + "Figure 5. Qualitative results for other objects. AoLP images are shown below the scene photos. For each method, the upper image is the estimated shape and the lower is the corresponding normal map. It is shown that our method successfully reconstructs fine details of the object, particularly the concave parts in the transparent vase and bottle, while other methods fail to reconstruct such details." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 486, + 506 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and without interiors, and a black tumbler, some of which are challenging objects for the existing approaches, especially a container with cutlery. In AoLP observations, while visible light is affected by a mixture of reflections and transmission components, thermal observations remain consistent and align with the azimuth angle of the surface normal. This consistency leads to a clear difference in detailed 3D reconstruction results between visible polarization and thermal polarization.", + "bbox": [ + 89, + 642, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the owl's result, while the detailed shape such as the eyes and the beak is missing in both visible MVAS and thermal IDR, they are successfully recovered in our result. Similarly, the relief pattern on the glass container is clearly visible in our results. For a black tumbler with a white lid shown in the fourth row, the AoLP observation on the lid in the visible-light spectrum is too noisy. In such a case, the surface normal of the lid is not reconstructible, resulting in", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a wavy surface. In contrast, our method recovers a reasonable shape due to the presence of a unique AoLP on the lid in LWIR polarization. Overall, visible MVAS tends to reconstruct with wavy artifacts as AoLP is noisy, and thermal IDR tends to recover swollen shapes as it does not consider the surface normal.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To quantitatively compare the results, the estimated shapes are evaluated using Chamfer distance, and the surface normals are evaluated by mean angular error. Our method outperforms the other methods in both Chamfer distance and the mean angular error for all objects.", + "bbox": [ + 511, + 186, + 906, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5 shows additional qualitative results. Our results are visually better than the other methods. The concave details of the vase and bottle are especially notable, as they are precisely reconstructed in our method while other results appear blurred. In addition, the surface of the container's lid is only reconstructed by our method while there is a big hole in the results of the others. Some artifacts are seen in thin parts in our result, particularly in the neck of the bottle. This artifact comes from the low spatial resolution of the thermal camera, which can be reduced by using a higher-resolution camera or increasing camera views.", + "bbox": [ + 511, + 270, + 908, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 469, + 633, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a 3D shape reconstruction using multi-view thermal polarimetric observations. We theoretically show that thermal polarimetric cues, especially AoLP images, are independent of material properties and illumination environment. Based on this property, we demonstrate a detailed 3D shape recovery using multi-view thermal AoLP images and show that our approach is applicable to heterogeneous objects including transparent and low-reflective objects.", + "bbox": [ + 511, + 501, + 906, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although the proposed method performs well, some challenges remain in expanding the applicable scenes. Our approach assumes the sufficient power of the emission component compared to the reflection component. This assumption is not valid when the hot objects are placed near the target object or when the target's temperature is lower than the environment. Moreover, our method is less effective on metallic surfaces or rough surfaces, where the polarization signal can become unstable or noisy. In such cases, decomposing the emission and reflection components [48] before applying the shape recovery or integrating additional cues may be a good option. Improving the observation system is also an important factor since our capturing system needs some time to rotate the polarization filter. A single-shot polarized LWIR camera could accelerate the thermal polarization analysis in the near future, as in the visible-light spectrum.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27397", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by JSPS KAKENHI Grant Numbers JP24K20802, JP22K18420, JP24K02965.", + "bbox": [ + 91, + 114, + 480, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 157, + 186, + 172 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] G.A. Atkinson and E.R. Hancock. Recovery of surface orientation from diffuse polarization. IEEE TIP, 15(6):1653-1664, 2006.", + "[2] Yunhao Ba, Alex Gilbert, Franklin Wang, Jina Yang, Rui Chen, Yiqin Wang, Lei Yan, Boxin Shi, and Achuta Kadambi. Deep Shape from Polarization. In ECCV, pages 554-571, 2020.", + "[3] Seung-Hwan Baek, Daniel S. Jeon, Xin Tong, and Min H. Kim. Simultaneous acquisition of polarimetric SVBRDF and normals. ACM TOG, 37(6):1-15, 2018.", + "[4] P.J. Besl and Neil D. McKay. A method for registration of 3-D shapes. IEEE TPAMI, 14(2):239-256, 1992.", + "[5] Xu Cao, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita. Multi-View Azimuth Stereo via Tangent Space Consistency. In CVPR, pages 825–834, 2023.", + "[6] Qian Chen, Shihao Shu, and Xiangzhi Bai. Thermal3D-GS: Physics-Induced 3D Gaussians for Thermal Infrared Novel-View Synthesis. In ECCV, pages 253–269, 2025.", + "[7] Charly Collin, Sumanta Pattanaik, Patrick LiKamWa, and Kadi Bouatouch. Computation of polarized subsurface BRDF for rendering. In Proceedings of Graphics Interface 2014, pages 201-208. Canadian Information Processing Society, 2014.", + "[8] Zhaopeng Cui, Jinwei Gu, Boxin Shi, Ping Tan, and Jan Kautz. Polarimetric Multi-view Stereo. In CVPR, pages 369-378, 2017.", + "[9] Unay Dorken Gallastegi, Hoover Rueda-Chacón, Martin J. Stevens, and Vivek K Goyal. Absorption-based hyperspectral thermal ranging: Performance analyses, optimization, and simulations. Optics Express, 32(1):151, 2024.", + "[10] Jean-Denis Durou, Maurizio Falcone, Yvain Quéau, and Silvia Tozza. A Comprehensive Introduction to Photometric 3D-Reconstruction. In Advances in Photometric 3D-Reconstruction, pages 1-29. Springer International Publishing, 2020.", + "[11] Jack Erdozain, Kazuto Ichimaru, Tomohiro Maeda, Hiroshi Kawasaki, Ramesh Raskar, and Achuta Kadambi. 3d Imaging For Thermal Cameras Using Structured Light. In ICIP, pages 2795-2799, 2020.", + "[12] Gonen Eren, Olivier Aubreton, Fabrice Meriaudeau, L.A. Sanchez Secades, David Fofi, A. Teoman Naskali, Frederic Truchetet, and Aytul Ercil. Scanning from heating: 3D shape estimation of transparent objects from local surface heating. Optics Express, 17(14):11457, 2009.", + "[13] Yoshiki Fukao, Ryo Kawahara, Shohei Nobuhara, and Ko Nishino. Polarimetric Normal Stereo. In CVPR, pages 682-690, 2021.", + "[14] Unay Dorken Gallastegi, Hoover Rueda-Chacon, Martin J. Stevens, and Vivek K Goyal. Absorption-Based, Passive Range Imaging from Hyperspectral Thermal Measurements. IEEE TPAMI, pages 1-18, 2025." + ], + "bbox": [ + 93, + 184, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] S. Garrido-Jurado, R. Muñoz-Salinas, F.J. Madrid-Cuevas, and M.J. Marín-Jiménez. Automatic generation and detection of highly reliable fiducial markers under occlusion. PR, 47(6):2280-2292, 2014.", + "[16] Jason Geng. Structured-light 3D surface imaging: A tutorial. Advances in Optics and Photonics, 3(2):128, 2011.", + "[17] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit Geometric Regularization for Learning Shapes, 2020.", + "[18] Yufei Han, Heng Guo, Koki Fukai, Hiroaki Santo, Boxin Shi, Fumio Okura, Zhanyu Ma, and Yunpeng Jia. NeRSP: Neural 3D Reconstruction for Reflective Objects with Sparse Polarized Images. In CVPR, pages 11821-11830, 2024.", + "[19] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2nd ed edition, 2003.", + "[20] A. Hertzmann and S.M. Seitz. Example-based photometric stereo: Shape reconstruction with general, varying BRDFs. IEEE TPAMI, 27(8):1254-1264, 2005.", + "[21] Inseung Hwang, Daniel S. Jeon, Adolfo Muñoz, Diego Gutierrez, Xin Tong, and Min H. Kim. Sparse ellipsometry: Portable acquisition of polarimetric SVBRDF and shape with unstructured flash photography. ACM TOG, 41(4):1-14, 2022.", + "[22] Tomoki Ichikawa, Matthew Purri, Ryo Kawahara, Shohei Nobuhara, Kristin Dana, and Ko Nishino. Shape from Sky: Polarimetric Normal Recovery Under The Sky. In CVPR, pages 14827-14836, 2021.", + "[23] Masaki Kaga, Takahiro Kushida, Tsuyoshi Takatani, Kenichiro Tanaka, Takuya Funatomi, and Yasuhiro Mukaigawa. Thermal non-line-of-sight imaging from specular and diffuse reflections. IPSJ Transactions on Computer Vision and Applications, 11(1):8, 2019.", + "[24] Takahiro Kushida, Ryutaro Nakamura, Hiroaki Matsuda, Wenhao Chen, and Kenichiro Tanaka. Affine transform representation for reducing calibration cost on absorption-based LWIR depth sensing. Scientific Reports, 14(1):26429, 2024.", + "[25] Chenyang Lei, Chenyang Qi, Jiaxin Xie, Na Fan, Vladlen Koltun, and Qifeng Chen. Shape from Polarization for Complex Scenes in the Wild. In CVPR, pages 12622-12631, 2022.", + "[26] Daniel A. LeMaster and Michael T. Eismann. Passive Polarimetric Imaging. In Multi-Dimensional Imaging, pages 391-428. Wiley, 1 edition, 2014.", + "[27] Chenhao Li, Taishi Ono, Takeshi Uemori, Hajime Mihara, Alexander Gatto, Hajime Nagahara, and Yusuke Moriuchi. NeISF: Neural Incident Stokes Field for Geometry and Material Estimation. In CVPR, pages 21434-21445, 2024.", + "[28] Yvette Y. Lin, Xin-Yi Pan, Sara Fridovich-Keil, and Gordon Wetzstein. ThermalNeRF: Thermal Radiance Fields. In ICCP, pages 1–12, 2024.", + "[29] Tomohiro Maeda, Yiqin Wang, Ramesh Raskar, and Achuta Kadambi. Thermal Non-Line-of-Sight Imaging. In ICCP, pages 1-11, 2019.", + "[30] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020." + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27398", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Seiichi Mita, Xu Yuquan, Kazuhisa Ishimaru, and Sakiko Nishino. Robust 3D Perception for any Environment and any Weather Condition using Thermal Stereo. In IEEE Intelligent Vehicles Symposium (IV), pages 2569-2574, 2019.", + "[32] Miyazaki, Tan, Hara, and Ikeuchi. Polarization-based inverse rendering from a single view. In ICCV, pages 982-987 vol.2, 2003.", + "[33] Daisuke Miyazaki, Megumi Saito, Yoichi Sato, and Katsushi Ikeuchi. Determining surface orientations of transparent objects based on polarization degrees in visible and infrared wavelengths. J. Opt. Soc. Am. A, 19(4):687, 2002.", + "[34] D. Miyazaki, M. Kagesawa, and K. Ikeuchi. Transparent surface modeling from a pair of polarization images. IEEE TPAMI, 26(1):73-82, 2004.", + "[35] Yasuto Nagase, Takahiro Kushida, Kenichiro Tanaka, Takuya Funatomi, and Yasuhiro Mukaigawa. Shape from Thermal Radiation: Passive Ranging Using Multi-spectral LWIR Measurements. In CVPR, pages 12651-12661, 2022.", + "[36] Sriram Narayanan, Mani Ramanagopal, Mark Sheinin, Aswin C. Sankaranarayanan, and Srinivasa G. Narasimhan. Shape from heat conduction. In ECCV, pages 426-444, 2025.", + "[37] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation. In CVPR, pages 165-174, 2019.", + "[38] Mani Ramanagopal, Sriram Narayanan, Aswin C. Sankaranarayanan, and Srinivasa G. Narasimhan. A Theory of Joint Light and Heat Transport for Lambertian Scenes. In CVPR, pages 11924-11933, 2024.", + "[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024.", + "[40] Joaquin Rodriguez, Lew-Fock-Chong Lew-Yan-Voon, Renato Martins, and Olivier Morel. Pola4All: Survey of polarimetric applications and an open-source toolkit to analyze polarization. Journal of Electronic Imaging, 33(01), 2024.", + "[41] Hiroaki Santo, Masaki Samejima, Yusuke Sugano, Boxin Shi, and Yasuyuki Matsushita. Deep Photometric Stereo Networks for Determining Surface Normal and Reflectances. IEEE TPAMI, 44(1):114-128, 2022.", + "[42] Philip Saponaro, Scott Sorensen, Abhishek Kolagunda, and Chandra Kambhametty. Material classification with thermal imagery. In CVPR, pages 4649-4656, 2015.", + "[43] D. Scharstein and R. Szeliski. High-accuracy stereo depth maps using structured light. In CVPR, pages I-195-I-202, 2003.", + "[44] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In CVPR, pages 4104-4113, 2016.", + "[45] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise View Selection for Unstructured Multi-View Stereo. In ECCV, pages 501–518, 2016." + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Boxin Shi, Jina Yang, Jinwei Chen, Ruihua Zhang, and Rui Chen. Recent Progress in Shape from Polarization. In Advances in Photometric 3D-Reconstruction, pages 177-203. Springer International Publishing, 2020.", + "[47] Richard Szeliski. Computer Vision: Algorithms and Applications. Springer International Publishing, 2022.", + "[48] Kenichiro Tanaka, Nobuhiro Ikeya, Tsuyoshi Takatani, Hiroyuki Kubo, Takuya Funatomi, Vijay Ravi, Achuta Kadambi, and Yasuhiro Mukaigawa. Time-Resolved Far Infrared Light Transport Decomposition for Thermal Photometric Stereo. IEEE TPAMI, 43(6):2075–2085, 2021.", + "[49] L.B. Wolff and T.E. Boult. Constraining object features using a polarization reflectance model. IEEE TPAMI, 13(7): 635-657, 1991.", + "[50] Robert J. Woodham. Photometric Method For Determining Surface Orientation From Multiple Images. Optical Engineering, 19(1), 1980.", + "[51] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. NeurIPS, 33, 2020.", + "[52] Tianxiang Ye, Qi Wu, Junyuan Deng, Guoqing Liu, Liu Liu, Songpengcheng Xia, Liang Pang, Wenxian Yu, and Ling Pei. Thermal-NeRF: Neural Radiance Fields from an Infrared Camera. In IROS, pages 1046–1053, 2024.", + "[53] Jincheng Zhang, Andrew R Willis, and Kevin Brink. Photometric Correction for Infrared Sensors. In CVPRW, pages 431-439, 2023.", + "[54] Z. Zhang. A flexible new technique for camera calibration. IEEE TPAMI, 22(11):1330-1334, 2000.", + "[55] Jinyu Zhao, Yusuke Monno, and Masatoshi Okutomi. Polarimetric Multi-View Inverse Rendering. IEEE TPAMI, 45 (7):8798-8812, 2023." + ], + "bbox": [ + 516, + 92, + 903, + 560 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27399", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Thermal Polarimetric Multi-view Stereo/493fd467-1c08-4433-ac25-63db3b58338a_model.json b/2025/Thermal Polarimetric Multi-view Stereo/493fd467-1c08-4433-ac25-63db3b58338a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4258d34a4b329adfaa416ebbcd2947543d5d6317 --- /dev/null +++ b/2025/Thermal Polarimetric Multi-view Stereo/493fd467-1c08-4433-ac25-63db3b58338a_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.296, + 0.131, + 0.704, + 0.151 + ], + "angle": 0, + "content": "Thermal Polarimetric Multi-view Stereo" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.182, + 0.666, + 0.254 + ], + "angle": 0, + "content": "Takahiro Kushida Kenichiro Tanaka Ritsumeikan University 2-150 Iwakura-cho, Ibaraki, Osaka, Japan {tkushida,ken-t}@fc.ritsumei.ac.jp" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.486, + 0.516 + ], + "angle": 0, + "content": "This paper introduces a novel method for detailed 3D shape reconstruction utilizing thermal polarization cues. Unlike state-of-the-art methods, the proposed approach is independent of illumination and material properties. In this paper, we formulate a general theory of polarization observation and show that long-wave infrared (LWIR) polarimetric imaging is free from the ambiguities that affect visible polarization analyses. Subsequently, we propose a method for recovering detailed 3D shapes using multi-view thermal polarimetric images. Experimental results demonstrate that our approach effectively reconstructs fine details in transparent, translucent, and heterogeneous objects, outperforming existing techniques." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.543, + 0.222, + 0.559 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.75 + ], + "angle": 0, + "content": "3D shape reconstruction is a fundamental problem in computer vision, with applications spanning both academia and industry. Over the years, various 3D reconstruction methods using RGB cameras have been proposed, but many of these methods rely on strong assumptions about illumination and reflection models. For instance, multi-view stereo [19, 44, 45, 47] depends on surface texture to find correspondences, while structured light [16, 43], photometric stereo [10, 20, 41, 50], and shape from polarization [1, 2, 32, 34, 46, 49] require specific lighting conditions and assume opaque surfaces, limiting their applicability to a narrow range of materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Thermal imaging presents an attractive alternative for 3D reconstruction in challenging scenarios because it does not rely on environmental illumination; any object with heat emits long-wave infrared (LWIR) light, effectively serving as its own light source. Furthermore, most materials, except for those designed for thermal optics, are opaque in the LWIR spectrum, eliminating the need for many of the assumptions required by traditional reflection models. Building on this concept, several methods, such as thermal multi-view stereo [53] and absorption-based depth es" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.349 + ], + "angle": 0, + "content": "timation [9, 14, 24, 35] have been proposed. Although these methods demonstrate potential applicability, their results also imply the limitation inherent in model-based approaches, especially with regard to accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Photometric cues in the LWIR spectrum are useful for recovering the accurate shape, similar to the visible-light spectrum. For example, thermal photometric stereo [48] and shape from heat conduction [36] have demonstrated detailed surface reconstruction. However, these methods require active heating and cooling of the object, making them both time-consuming and impractical. In this paper, we aim to reconstruct fine 3D shapes in a steady state using polarimetric cues, a type of photometric cue that can be obtained without monitoring the heating or cooling processes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.909, + 0.669 + ], + "angle": 0, + "content": "Table 1 presents a comparison of various 3D reconstruction approaches. Our approach is illumination-independent, material-independent, free from heating or cooling process, and highly accurate. Our contributions are twofold. First, we establish a unified theoretical framework for LWIR polarization, showing that, unlike visible-light polarization, it is not affected by specular-diffuse ambiguities. Second, we show that LWIR polarization serves as a powerful cue for detailed 3D shape reconstruction. Experimental results confirm that our method outperforms the existing approaches on heterogeneous materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.685, + 0.656, + 0.7 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.791, + 0.727 + ], + "angle": 0, + "content": "2.1. Shape from visible polarization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Shape from Polarization (SfP), which recovers surface normals from the polarization state of reflected light, has a long-standing history in computer vision [1, 2, 32, 34, 46, 49]. Recently, the advent of snapshot polarization cameras has significantly boosted this area of research [40]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.9 + ], + "angle": 0, + "content": "SfP approaches continue to tackle ambiguities in surface normals that arise from the mixture of specular and diffuse polarizations [46]. One effective way to resolve these ambiguities is to use multi-view polarization images, which provide a richer set of constraints on the surface normals and lead to more accurate and robust surface reconstruc" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27390" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.169, + 0.088, + 0.83, + 0.226 + ], + "angle": 0, + "content": "
MethodIlluminationMaterialAccuracyMeasurement
Visible Multi-view Stereo [19, 44, 45, 47]DependentTextured onlyHighEasy
Visible SL / PS [10, 16, 20, 41, 43, 50]DependentOpaque onlyHighEasy
Visible Polarization [1, 2, 32, 34, 46, 49]DependentDependentHighEasy
Visible NeRFs [30, 51]DependentWide range of materialsLowEasy
Thermal Multi-view Stereo [53]IndependentTextured onlyHighEasy
Thermal SL [11]IndependentDiffuse reflective onlyHighEasy
Thermal PS [48]IndependentWide range of materialsModerateHeating & cooling
Thermal NeRFs [28, 52]IndependentWide range of materialsLowEasy
Heat Conduction [36]IndependentWide range of materialsModerateHeating
Thermal Polarization (Ours)IndependentWide range of materialsHighEasy
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.235, + 0.908, + 0.264 + ], + "angle": 0, + "content": "Table 1. Comparison table of shape reconstruction methods. While other methods have pros and cons, our approach is independent of illumination and material's optical properties as well as accurate shape recovery is possible in steady state measurements." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.29, + 0.482, + 0.456 + ], + "angle": 0, + "content": "tions [8]. Another approach involves using polarimetric BRDF models, which can handle complex material properties [3, 13, 21, 55]. Recent approaches incorporate neural scene representations and differentiable rendering, enabling the joint estimation of shape, environment maps, and reflectance [18, 27]. Nevertheless, SfP in the visible light spectrum remains fundamentally sensitive to the illumination conditions, such as the assumption of uniform, unpolarized ambient light, and to variations in material's optical properties, including transparency, translucency, matte, or shiny finishes [46]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.563 + ], + "angle": 0, + "content": "In this paper, we introduce the use of polarization in the LWIR spectrum. The LWIR light is emitted directly from the object and is independent of external illumination. Moreover, the polarization of LWIR emission is not mixed up with the diffuse and specular reflections or transmission effects. These properties make LWIR polarization advantageous for robust shape reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.435, + 0.592 + ], + "angle": 0, + "content": "2.2. LWIR imaging and physics-based vision" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Beyond 3D reconstruction, thermal imaging is utilized for a variety of physics-based vision tasks. The analysis of heat conduction, for instance, has been leveraged for material classification by exploiting differences in thermal diffusivity and conductivity [42]. Reflection analysis of LWIR light has facilitated non-line-of-sight (NLoS) imaging, expanding scene understanding by examining how thermal rays propagate around the corner [23, 29]. Joint analyses of both visible and LWIR light transport have further enriched scene understanding [38]. We also advance the area of physics-based thermal vision through a shape reconstruction using LWIR polarization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Thermal imaging is an effective approach for 3D shape reconstruction in some challenging scenarios. For example, it is effective for stealth observation as the thermal radiation does not rely on external illumination [6, 35, 52, 53]. Another example is that it is feasible to observe the shape of challenging materials in the visible light spectrum because many materials that appear transparent or translucent are opaque in the LWIR spectrum [48]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.905, + 0.547 + ], + "angle": 0, + "content": "Existing 3D reconstruction techniques using thermal cameras can be broadly categorized into geometric and photometric approaches. Geometric methods [12, 31, 53] such as based on stereo/multi-view observations recover surface depth by matching corresponding points across thermal images, commonly assuming near-Lambertian emission or sufficient texture. Photometric approaches infer depth, shape, or surface normal from intensity cues, which include thermal photometric stereo [48], absorption-based method [9, 14, 24, 35], and heat conduction-based method [36]. While they show strong possibilities of shape reconstruction using a thermal camera, their accuracy is generally limited. There are hybrid approaches that combine visible-light and LWIR polarization [34], where the LWIR polarization is used only as an auxiliary cue to resolve the specular-diffuse ambiguity in visible polarization rather than directly contributing to shape recovery." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.906, + 0.682 + ], + "angle": 0, + "content": "Recently, neural rendering methods inspired by Neural Radiance Fields (NeRF) have been explored in the thermal domain [6, 28, 52] as well. By jointly optimizing a volumetric scene representation and a rendering function to satisfy the rendered images and align observed images from multiple viewpoints, these thermal NeRF techniques recover both geometry and radiometric properties. We are also inspired by this neural representation and model the object's surface by a neural implicit surface." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.906, + 0.758 + ], + "angle": 0, + "content": "In this paper, we leverage the polarization state of thermal radiation to enrich the information available for 3D reconstruction. By incorporating the polarization cues, geometric rotation, and neural representation, our method enables more accurate shape estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.809, + 0.787 + ], + "angle": 0, + "content": "3. Polarimetric LWIR Observation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.841 + ], + "angle": 0, + "content": "In this section, we present a general theory of polarization observation and discuss the advantages of thermal polarization over visible polarization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.724, + 0.865 + ], + "angle": 0, + "content": "3.1. Physics of Polarization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The polarization state of light, including both visible and LWIR spectra, changes when it is reflected from or trans" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27391" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.142, + 0.089, + 0.434, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.298, + 0.485, + 0.368 + ], + "angle": 0, + "content": "Figure 1. Coordinate systems and the zenith and projected azimuth angles of the surface normal. While Mueller calculus is performed in the object's surface coordinate system, the observed Stokes parameter and projected azimuth angle are represented in the image coordinate system." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.44 + ], + "angle": 0, + "content": "mitted through an object's surface. The transport of polarized light can be described using the Stokes parameters \\(\\boldsymbol{s} = [s_0,s_1,s_2,s_3]^{\\mathrm{T}}\\) and the Mueller matrix \\(M\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.247, + 0.45, + 0.482, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {o} = \\boldsymbol {M} \\boldsymbol {s} _ {i}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.521 + ], + "angle": 0, + "content": "where \\( s_i \\) and \\( s_o \\) denote the Stokes parameters of the incoming and outgoing light, respectively, both defined in the surface coordinate system [7]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.484, + 0.58 + ], + "angle": 0, + "content": "The observed scene is a combination of specular reflection, diffuse reflection, transmission, and emission. Consequently, the overall polarization state \\( s \\) can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.594, + 0.482, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} = \\boldsymbol {s} _ {s} + \\boldsymbol {s} _ {d} + \\boldsymbol {s} _ {t} + \\boldsymbol {s} _ {e}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.665 + ], + "angle": 0, + "content": "where \\( s_s, s_d, s_t \\), and \\( s_e \\) represent the Stokes vectors corresponding to the specular, diffuse, transmission, and emission components, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.484, + 0.713 + ], + "angle": 0, + "content": "Specular polarization The change in the polarization state due to specular reflection [3, 22] is described by" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.722, + 0.482, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {s} = \\boldsymbol {C} (\\varphi) \\boldsymbol {F} ^ {\\mathrm {R}} (\\theta) \\boldsymbol {s} _ {i}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\( s_i \\) denotes the Stokes vector of the incident light in the object's surface coordinate system, \\( s_s \\) denotes the corresponding Stokes vector in the image coordinate system, \\( C(\\varphi) \\) is a rotation matrix that transforms coordinates from the image to the object's surface, \\( \\varphi \\) is the projected azimuth angle (i.e. the angle between the vertical axis of the image plane and the surface normal projected into the image plane), \\( F^{\\mathrm{R}}(\\theta) \\) is the Mueller matrix for the specular component, and \\( \\theta \\) is the zenith angle of the reflection (i.e. the angle between the surface normal and the viewing direction). The" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.093, + 0.906, + 0.138 + ], + "angle": 0, + "content": "surface coordinate system, image coordinate system, zenith angle, and projected azimuth angle are illus in Fig. 1. The rotation matrix \\( C(\\varphi) \\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.146, + 0.906, + 0.211 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {C} (\\varphi) = \\left[ \\begin{array}{c c c c} 1 & 0 & 0 & 0 \\\\ 0 & \\cos 2 \\varphi & - \\sin 2 \\varphi & 0 \\\\ 0 & \\sin 2 \\varphi & \\cos 2 \\varphi & 0 \\\\ 0 & 0 & 0 & 1 \\end{array} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.22, + 0.906, + 0.251 + ], + "angle": 0, + "content": "The Mueller matrix \\( \\pmb{F}^{\\mathrm{R}}(\\theta) \\) is derived from Fresnel's law and expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.259, + 0.906, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} ^ {\\mathrm {R}} (\\theta) = \\left[ \\begin{array}{l l l l} r _ {+} & r _ {-} & 0 & 0 \\\\ r _ {-} & r _ {+} & 0 & 0 \\\\ 0 & 0 & r _ {\\times} \\cos \\delta & r _ {\\times} \\sin \\delta \\\\ 0 & 0 & - r _ {\\times} \\sin \\delta & r _ {\\times} \\cos \\delta \\end{array} \\right], \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.906, + 0.412 + ], + "angle": 0, + "content": "where \\(r_{\\pm} = \\frac{r_s \\pm r_p}{2}\\), \\(r_{\\times} = \\sqrt{r_p r_s}\\), \\(\\theta'\\) denotes the refractive angle. Here, \\(\\cos \\delta\\) equals \\(-1\\) when \\(\\theta\\) is less than Brewster's angle and \\(1\\) otherwise, with the sign of \\(\\sin \\delta\\) reversed accordingly [3]. The Fresnel reflectances for s- and p-polarization, \\(r_s\\) and \\(r_p\\), are given by" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.42, + 0.922, + 0.472 + ], + "angle": 0, + "content": "\\[\nr _ {s} (\\theta) = \\left(\\frac {\\cos \\theta - \\eta \\cos \\theta^ {\\prime}}{\\cos \\theta + \\eta \\cos \\theta^ {\\prime}}\\right) ^ {2}, r _ {p} (\\theta) = \\left(\\frac {\\cos \\theta^ {\\prime} - \\eta \\cos \\theta}{\\cos \\theta^ {\\prime} + \\eta \\cos \\theta}\\right) ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.481, + 0.885, + 0.497 + ], + "angle": 0, + "content": "According to Snell's law, the refractive angle is given as" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.506, + 0.906, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\theta^ {\\prime} = \\sin^ {- 1} \\left(\\frac {1}{\\eta} \\sin \\theta\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.548, + 0.861, + 0.563 + ], + "angle": 0, + "content": "where \\(\\eta\\) represents the refractive index of the object." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.581, + 0.906, + 0.626 + ], + "angle": 0, + "content": "Diffuse polarization The change in the polarization state due to diffuse reflection or subsurface scattering [1, 22] is described by" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.635, + 0.906, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {d} = \\boldsymbol {C} (\\varphi) \\boldsymbol {F} ^ {\\mathrm {T}} (\\theta) \\int_ {\\Omega} \\boldsymbol {D} \\boldsymbol {F} ^ {\\mathrm {T}} (\\theta) \\boldsymbol {s} _ {i, \\omega} d \\omega , \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.906, + 0.72 + ], + "angle": 0, + "content": "where \\( s_{i,\\omega} \\) denotes the Stokes vector of incident light coming from the direction \\( \\omega \\). The matrix \\( D \\), which describes the depolarization, is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.729, + 0.906, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {D} = \\left[ \\begin{array}{l l l l} \\rho & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\end{array} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.801, + 0.906, + 0.832 + ], + "angle": 0, + "content": "where \\(\\rho\\) denotes the proportion of the depolarization. The Mueller matrix for transmittance \\(\\pmb{F}^{\\mathrm{T}}(\\theta)\\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.841, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} ^ {\\mathrm {T}} (\\theta) = \\left[ \\begin{array}{c c c c} t _ {+} & t _ {-} & 0 & 0 \\\\ t _ {-} & t _ {+} & 0 & 0 \\\\ 0 & 0 & t _ {\\times} \\cos \\delta & t _ {\\times} \\sin \\delta \\\\ 0 & 0 & - t _ {\\times} \\sin \\delta & t _ {\\times} \\cos \\delta \\end{array} \\right], \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27392" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.485, + 0.138 + ], + "angle": 0, + "content": "where \\( t_{\\pm} = \\frac{t_s \\pm t_p}{2} \\), \\( t_{\\times} = \\sqrt{t_p t_s} \\). \\( t_p \\) and \\( t_s \\) denote the Fresnel transmittance coefficients for s- and p- polarization, respectively, and are expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.142, + 0.498, + 0.192 + ], + "angle": 0, + "content": "\\[\nt _ {s} (\\theta) = \\left(\\frac {2 \\cos \\theta}{\\cos \\theta + \\eta \\cos \\theta^ {\\prime}}\\right) ^ {2}, t _ {p} (\\theta) = \\left(\\frac {2 \\cos \\theta}{\\cos \\theta^ {\\prime} + \\eta \\cos \\theta}\\right) ^ {2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.209, + 0.485, + 0.255 + ], + "angle": 0, + "content": "Transmission polarization The polarization state also changes due to light transmission. According to the Fresnel's law, the transmission polarization is described by" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.26, + 0.484, + 0.279 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {t} = \\boldsymbol {C} (\\varphi) \\boldsymbol {F} ^ {\\mathrm {T}} (\\theta) \\boldsymbol {s} _ {i}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.285, + 0.484, + 0.314 + ], + "angle": 0, + "content": "where \\( s_i \\) denotes the Stokes vector of the light coming from under the surface." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.484, + 0.363 + ], + "angle": 0, + "content": "Emission polarization The polarization state of emitted light is described by" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.37, + 0.483, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {e} = \\boldsymbol {C} (\\varphi) \\boldsymbol {E} (\\theta) \\boldsymbol {s} _ {c}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.485 + ], + "angle": 0, + "content": "where \\( s_c \\) denotes the Stokes vector of the emitted light including thermal radiation as well as any chemical or electrical emissions and \\( E(\\theta) \\) is the Mueller matrix corresponding to the emission component. Since there are many types of light emission, the elements of the emission Mueller matrix are not specified explicitly in general." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.492, + 0.373, + 0.506 + ], + "angle": 0, + "content": "3.2. LWIR Polarization Observation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.513, + 0.484, + 0.574 + ], + "angle": 0, + "content": "Polarization of visible light Since objects other than light sources do not emit light in the visible spectrum, the emission component is generally ignorable. Therefore, the observed polarization can be modeled as" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.583, + 0.482, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} = \\boldsymbol {s} _ {d} + \\boldsymbol {s} _ {s} + \\boldsymbol {s} _ {t}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.484, + 0.695 + ], + "angle": 0, + "content": "The magnitude of each component varies in complex ways depending on the object's material properties and the illumination environment. This complexity makes polarization analyses more challenging, leading existing visible polarization studies to rely on various assumptions to facilitate the analysis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.484, + 0.878 + ], + "angle": 0, + "content": "Polarization of LWIR light Most objects are known to be opaque in the LWIR spectrum; therefore, the transmission component can be safely ignored. Moreover, an interesting property of LWIR light polarization is that the relative contributions of each component can be controlled by the object's temperature. In a typical environments, where no objects are extremely hot in the surroundings (i.e., where there are no significant LWIR light sources in the surroundings) the magnitude of reflection components is small compared to that of the emission. In such cases, the LWIR polarization observation can be modeled as" + }, + { + "type": "equation", + "bbox": [ + 0.26, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\[\ns = s _ {e}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "Since LWIR emission primarily arises from black body radiation, Eq. (13) can be rewritten as" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.129, + 0.905, + 0.144 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {e} = \\boldsymbol {C} (\\varphi) \\boldsymbol {\\mathcal {E}} (\\theta) \\boldsymbol {s} _ {B} (T), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.151, + 0.905, + 0.196 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}(\\theta)\\) denotes the Mueller matrix for black body radiation, and \\(s_B(T)\\) represents the Stokes vector of the radiation, which can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.201, + 0.905, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {B} (T) = \\left[ \\begin{array}{c} B (T) \\\\ 0 \\\\ 0 \\\\ 0 \\end{array} \\right], \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.269, + 0.905, + 0.3 + ], + "angle": 0, + "content": "where \\(B\\) denotes the black body radiation. According to Planck's law, \\(B\\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.304, + 0.905, + 0.338 + ], + "angle": 0, + "content": "\\[\nB (T) = \\frac {2 h c ^ {2}}{\\lambda^ {5}} \\frac {1}{e ^ {\\frac {h c}{\\lambda k T}} - 1}, \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.905, + 0.417 + ], + "angle": 0, + "content": "where \\(c\\) is the speed of light, \\(k\\) is the Boltzmann constant, \\(h\\) is the Planck constant, \\(T\\) is the object's absolute temperature, and \\(\\lambda\\) is the observation wavelength. Equation (17) demonstrates that the object's temperature \\(T\\) controls the power of emission." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.418, + 0.904, + 0.447 + ], + "angle": 0, + "content": "The Mueller matrix of black body radiation \\(\\mathcal{E}(\\theta)\\) [26, 33] is given by" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.447, + 0.905, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mathcal {E}} (\\theta) = \\left[ \\begin{array}{l l l l} \\varepsilon_ {+} & 0 & 0 & 0 \\\\ \\varepsilon_ {-} & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\end{array} \\right], \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.514, + 0.906, + 0.604 + ], + "angle": 0, + "content": "where \\(\\varepsilon_{\\pm} = \\frac{\\varepsilon_s \\pm \\varepsilon_p}{2}\\). \\(\\varepsilon_p\\) and \\(\\varepsilon_s\\) denote the emissivities for s- and p-polarization, respectively. According to the energy conservation law, also known as Kirchhoff's law, the sum of reflectance, transmittance, and emissivity is equal to 1. Since the transmittance is negligible, as discussed above, we obtain" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.612, + 0.905, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\varepsilon_ {s} (\\theta) = 1 - r _ {s} (\\theta) \\tag {20}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.631, + 0.905, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\varepsilon_ {p} (\\theta) = 1 - r _ {p} (\\theta). \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.653, + 0.906, + 0.714 + ], + "angle": 0, + "content": "In summary, in the LWIR spectrum, the Stokes observation \\( s \\) can be analytically represented without ambiguity, thereby simplifying the analysis of the polarization state, which is a significant advantage for shape reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.731, + 0.906, + 0.791 + ], + "angle": 0, + "content": "Polarization to surface normal: LWIR polarization provides cues for an object's surface normal, which can be determined analytically. The degree of linear polarization (DoLP) is related to the zenith angle \\(\\theta\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.796, + 0.912, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {D o L P} = \\frac {\\sqrt {s _ {1} ^ {2} + s _ {2} ^ {2}}}{s _ {0}} (22) \\\\ = \\frac {\\left(\\eta - \\frac {1}{\\eta}\\right) ^ {2} \\sin^ {2} \\theta}{2 + 2 \\eta^ {2} - \\left(\\eta + \\frac {1}{\\eta}\\right) \\sin^ {2} \\theta + 4 \\cos \\theta \\sqrt {\\eta^ {2} - \\sin^ {2} \\theta}}. (23) \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27393" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.09, + 0.482, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.313, + 0.485, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Visible and thermal polarization observations of spheres made from various materials under natural room lighting. While visible polarization is unstable due to variations in a material's optical properties, thermal polarimetric cues, especially AoLP images, remain consistent across different materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.44 + ], + "angle": 0, + "content": "The angle of linear polarization (AoLP) is equivalent to the azimuth angle \\(\\varphi\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.45, + 0.483, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A o L P} = \\frac {1}{2} \\tan^ {- 1} \\left(\\frac {s _ {2}}{s _ {1}}\\right) = \\varphi . \\tag {24}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.584 + ], + "angle": 0, + "content": "These correspondences enable the reconstruction of the surface normal from polarization observations. Especially, AoLP serves as a robust cue because it directly reflects the azimuth angle of the surface normal and is independent of material properties, whereas DoLP is material-dependent due to the inclusion of the refractive index term." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.457, + 0.608 + ], + "angle": 0, + "content": "3.3. Examples of Visible vs. LWIR Polarization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.825 + ], + "angle": 0, + "content": "In the visible light spectrum, specular, diffuse, and transmission polarization components are often mixed [2, 25, 46], resulting in complicated observation. Fig. 2 shows the polarimetric observations of spheres made from various materials; in the visible spectrum, unwanted polarization observations frequently appear. For example, because black stone exhibits strong specular polarization, the DoLP image is heavily affected by reflections from the surrounding environment. In the case of transparent glass sphere, the observation is notably contaminated by background textures. For translucent plastic materials, both DoLP and AoLP images are affected by surface texture and subsurface scattering. These unstable observations significantly complicate subsequent shape recovery." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In contrast, in the LWIR spectrum, polarimetric observations remain consistent and independent of both material properties and environmental reflections, particularly in AoLP images. This consistency is a key advantage of thermal polarization cues, as it provides robustness against vari" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.825, + 0.107 + ], + "angle": 0, + "content": "ations in materials and surrounding reflections." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.121, + 0.877, + 0.137 + ], + "angle": 0, + "content": "4. Thermal Polarimetric Multi-view Stereo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.147, + 0.906, + 0.222 + ], + "angle": 0, + "content": "Our goal is to recover the 3D shape using thermal polarimetric cues. As discussed in the previous section, the AoLP observation is invariant to material properties. To leverage this strong cue, we only use the AoLP images for material-independent 3D shape reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Our approach employs multi-view thermal AoLP images to estimate the Signed Distance Function (SDF), which is implicitly represented by a Multi-Layer Perceptron (MLP) optimized via differentiable rendering. To effectively incorporate the surface normal information present in AoLP images, we adopt the concept of tangent space constraint (TSC) proposed in a multi-view azimuth stereo (MVAS) [5]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.355, + 0.683, + 0.371 + ], + "angle": 0, + "content": "Shape representation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.376, + 0.906, + 0.451 + ], + "angle": 0, + "content": "The SDF is implicitly represented by an MLP \\( f(\\pmb{x}; \\pmb{\\theta}) : \\mathbb{R}^3 \\times \\mathbb{R}^d \\to \\mathbb{R} \\), where \\( \\pmb{x} \\in \\mathbb{R} \\) denotes a point in space and \\( \\pmb{\\theta} \\in \\mathbb{R}^d \\) are the learnable parameters of the MLP. The zero-level set of the SDF defines the object's surface \\( \\mathcal{M} \\) as [37]" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.453, + 0.905, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} (\\boldsymbol {\\theta}) = \\{\\boldsymbol {x} | f (\\boldsymbol {x}; \\boldsymbol {\\theta}) = 0 \\}. \\tag {25}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.477, + 0.906, + 0.507 + ], + "angle": 0, + "content": "The MLP is optimized within the framework of implicit differentiable renderer (IDR) [51]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.518, + 0.624, + 0.532 + ], + "angle": 0, + "content": "Loss function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.906, + 0.569 + ], + "angle": 0, + "content": "We adopt the loss function proposed in MVAS [5], which is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.583, + 0.905, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {T S C}} + \\lambda_ {1} \\mathcal {L} _ {\\text {s i l h o u e t t e}} + \\lambda_ {2} \\mathcal {L} _ {\\text {E i k o n a l}}, \\tag {26}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.612, + 0.906, + 0.672 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{TSC}}\\) is the tangent space consistency loss, \\(\\mathcal{L}_{\\mathrm{silhouette}}\\) is the silhouette loss and \\(\\mathcal{L}_{\\mathrm{Eikonal}}\\) is the Eikonal regularization. \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are the weights assigned to the respective loss terms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.783 + ], + "angle": 0, + "content": "Tangent space consistency loss [5] The AoLP captured from different camera positions enforces a strong constraint on surface normal estimation. For a unit normal vector \\( \\pmb{n} \\subset S^2 \\in \\mathbb{R}^3 \\) at a surface point \\( \\pmb{x} \\in \\mathbb{R}^3 \\) and the projected azimuth angle \\( \\varphi \\) at the corresponding pixel, the following relationship holds:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.794, + 0.905, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {r} _ {1} ^ {\\top} \\boldsymbol {n} \\cos \\varphi - \\boldsymbol {r} _ {2} ^ {\\top} \\boldsymbol {n} \\sin \\varphi = 0, \\tag {27}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.824, + 0.905, + 0.856 + ], + "angle": 0, + "content": "where \\(\\pmb{R} = [r_1, r_2, r_3]^\\top\\) denotes the rotation matrix of the camera pose. This equation can be rearranged as" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.866, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {n} ^ {\\top} \\underbrace {\\left(\\boldsymbol {r} _ {1} \\cos \\varphi - \\boldsymbol {r} _ {2} \\sin \\varphi\\right)} _ {\\boldsymbol {t} (\\varphi)} = 0, \\tag {28}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27394" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "where \\( t(\\varphi) \\) is called a projected tangent vector and the above equation is called tangent space constraint. The TSC loss is the sum of the square error of the tangent space constraint and is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.155, + 0.483, + 0.197 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T S C}} = \\frac {1}{P} \\sum_ {\\boldsymbol {x} \\in \\boldsymbol {X}} \\frac {\\sum_ {i = 1} ^ {N} \\Phi_ {i} \\left(\\boldsymbol {n} ^ {\\top} (\\boldsymbol {x} ; \\boldsymbol {\\theta}) \\boldsymbol {t} _ {i} (\\boldsymbol {x})\\right) ^ {2}}{\\sum_ {i = 1} ^ {N} \\Phi_ {i}} \\tag {29}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.202, + 0.484, + 0.295 + ], + "angle": 0, + "content": "where \\(P\\) denotes the number of pixel samples, \\(\\Phi_{i}\\) is a binary indicator of the visibility of point \\(\\pmb{x}\\) from the \\(i\\)-th view, \\(\\pmb{n}^{\\top}(\\pmb{x};\\pmb{\\theta})\\) is the estimated surface normal at the point \\(\\pmb{x}\\) given network parameters \\(\\pmb{\\theta}\\), and \\(\\pmb{t}_{i}(\\pmb{x})\\) is the tangent vector corresponding to the \\(i\\)-th view's pixel onto which the point \\(\\pmb{x}\\) is projected." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.34 + ], + "angle": 0, + "content": "Silhouette loss [5, 51] The silhouette loss constrains the visual hull of the shape and is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.345, + 0.483, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i l h o u e t t e}} = \\frac {1}{\\alpha P} \\sum_ {\\boldsymbol {x} \\in \\hat {\\boldsymbol {X}}} \\Psi (O (\\Pi (\\boldsymbol {x})), \\sigma (\\alpha f ^ {*})), \\tag {30}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.54 + ], + "angle": 0, + "content": "where \\(\\Psi\\) denotes the cross entropy and \\(O\\) is a binary mask indicating whether a pixel lies inside the object silhouette (as provided by the input). \\(P\\) is the number of pixel samples, \\(\\Pi\\) is the projection from the world coordinate to the image coordinate, \\(\\tilde{\\pmb{X}}\\) represents the pixels with no ray-surface interactions or outside the silhouette, \\(f^{*}\\) represents the minimum SDF distance between the ray that does not have intersections and the object surface, \\(\\sigma\\) is a sigmoid function, and \\(\\alpha\\) controls the sharpness of the sigmoid function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.603 + ], + "angle": 0, + "content": "Eikonal regularization [17] The Eikonal loss regularizes the gradient of SDF, enforcing the SDF to maintain its property of having a gradient close to 1:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.608, + 0.483, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {E i k o n a l}} = \\mathbb {E} _ {\\boldsymbol {x}} \\left(\\left(\\left\\| \\boldsymbol {n} (\\boldsymbol {x}; \\boldsymbol {\\theta}) \\right\\| _ {2} - 1\\right) ^ {2}\\right). \\tag {31}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.483, + 0.678 + ], + "angle": 0, + "content": "\\(\\mathbb{E}_x(\\cdot)\\) denotes the expected value computed over sample points \\(\\pmb{x}\\) that are randomly sampled within the bounding box." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.688, + 0.224, + 0.705 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.173, + 0.729 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Implementation Following MVAS [5], we use an 8-layer MLP with softmax activations, combined with positional encoding (dimension 10). The input 3D position and encoded features are skip-connected to the 4th layer of the network. The optimization is performed using the Adam optimizer with a learning rate of 0.001. The weights for the loss terms are set to \\(\\lambda_{1} = 50\\) (silhouette loss) and \\(\\lambda_{2} = 0.1\\) (Eikonal loss). Training is conducted for 50 epochs with a batch size of 4096 pixels, and the silhouette loss weight is halved every 10 epochs. Training takes approximately 2 hours on a single RTX A6000 GPU." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.906, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.31, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Camera system and setup. The system consists of a rotation stage to rotate the target and a thermal camera with a wire-grid polarizer. A visible polarization camera is used for camera pose estimation and for comparison with the visible MVAS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.62 + ], + "angle": 0, + "content": "Camera system Figure 3 shows our camera system and experimental setup. Our thermal polarization camera system consists of a thermal camera (FLIR Boson 320) with \\(15\\mathrm{mm}\\) lens and a wire-grid polarizer (Thorlabs WP25M-IRC) mounted on a motorized rotation mount (Thorlabs K10CR2). The polarizer is installed between the image sensor and imaging lens to mitigate the narcissus reflections. Thermal polarization images are captured by rotating the polarizer using the rotation mount. We captured four polarization images at 0, 45, 90, and 135 degrees. A monochrome visible-light polarization camera (FLIR Blackfly S BFS-U3-51S5P) is also placed adjacent to the thermal camera. The visible-light polarization camera is used both to capture images for a baseline method and to get the poses of both visible and thermal cameras." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.909, + 0.837 + ], + "angle": 0, + "content": "Camera poses We employed a marker-based approach to estimate the camera poses. We put Aruco markers [15] on the rotation stage to estimate the pose of visible-light camera. Since printed markers are not visible in the thermal images, the poses of the thermal camera were obtained by transforming the camera poses of the visible-light camera. For stereo calibration between the visible-light and thermal cameras, we used a calibration target that can be observed by both cameras. The calibration board consists of a white-painted aluminum pegboard placed in front of a black surface heater. By heating the back panel, the circle grid pattern became visible to both cameras. The calibration parameters were estimated using Zhang's method [54]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Dataset We prepared a real-world dataset with ground-truth shapes. We captured 7 objects with various shapes and materials. The target objects include transparent and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "27395" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.907, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.518, + 0.907, + 0.575 + ], + "angle": 0, + "content": "Figure 4. Experimental results. AoLP images in both visible and LWIR spectra are shown below the scene photograph. Each method's result includes the recovered shape, surface normals, and angular error maps. The Chamfer distance is displayed below the estimated shape and the mean angular error is displayed to the right of the angular error map. For both metrics, lower values indicate better performance. The smallest errors are shown in magenta. It demonstrates that the proposed method outperforms the other methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.484, + 0.783 + ], + "angle": 0, + "content": "low-reflective objects, which pose challenges for existing reconstruction methods. The objects were placed on a rotating table and captured from about 20–30 views by rotating the table. To robustly capture thermal polarization, the objects were kept warm during the measurements. Masks for thermal images were generated by thresholding pixel values, while masks for visible images were generated by using Segment Anything Model 2 [39]. Ground-truth shapes were obtained using a structured light 3D scanner (EinScan-SP V2). To facilitate the ground-truth measurement, the surfaces of the objects were painted white to make them opaque." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Baselines We compare our proposed method with two state-of-the-art methods: MVAS [5] based on visible-light polarization cues and IDR [51], which is one of the neural implicit surface reconstruction methods, with thermal image input. For MVAS with visible-light polarization images, there exists \\(\\pm \\frac{\\pi}{2}\\) ambiguity in surface normals due to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.907, + 0.678 + ], + "angle": 0, + "content": "the mixture of specular and diffuse polarization. Therefore, TSC loss is modified to account for the \\(\\pm \\frac{\\pi}{2}\\) ambiguity, as originally proposed in MVAS. Since our method does not exhibit this ambiguity, the modified loss function is applied only to visible light observation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.907, + 0.837 + ], + "angle": 0, + "content": "Evaluation metrics We use two metrics to evaluate both the surface points and surface normals. Chamfer distance is the metric for evaluating the geometric error, which calculates the point cloud distance between the estimate and the ground-truth. The mean angular error is the metric for surface normal, which calculates the angular error of normal maps. For both metrics, a smaller value indicates a better result. Prior to computing these metrics, the estimated and the ground-truth meshes are aligned using ICP algorithm [4]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.608, + 0.862 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Figure 4 shows the results. The target objects are an owl made of ceramic with a clear coating, a glass container with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27396" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.488, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.517, + 0.486, + 0.617 + ], + "angle": 0, + "content": "Figure 5. Qualitative results for other objects. AoLP images are shown below the scene photos. For each method, the upper image is the estimated shape and the lower is the corresponding normal map. It is shown that our method successfully reconstructs fine details of the object, particularly the concave parts in the transparent vase and bottle, while other methods fail to reconstruct such details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.779 + ], + "angle": 0, + "content": "and without interiors, and a black tumbler, some of which are challenging objects for the existing approaches, especially a container with cutlery. In AoLP observations, while visible light is affected by a mixture of reflections and transmission components, thermal observations remain consistent and align with the azimuth angle of the surface normal. This consistency leads to a clear difference in detailed 3D reconstruction results between visible polarization and thermal polarization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In the owl's result, while the detailed shape such as the eyes and the beak is missing in both visible MVAS and thermal IDR, they are successfully recovered in our result. Similarly, the relief pattern on the glass container is clearly visible in our results. For a black tumbler with a white lid shown in the fourth row, the AoLP observation on the lid in the visible-light spectrum is too noisy. In such a case, the surface normal of the lid is not reconstructible, resulting in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "a wavy surface. In contrast, our method recovers a reasonable shape due to the presence of a unique AoLP on the lid in LWIR polarization. Overall, visible MVAS tends to reconstruct with wavy artifacts as AoLP is noisy, and thermal IDR tends to recover swollen shapes as it does not consider the surface normal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.188, + 0.907, + 0.265 + ], + "angle": 0, + "content": "To quantitatively compare the results, the estimated shapes are evaluated using Chamfer distance, and the surface normals are evaluated by mean angular error. Our method outperforms the other methods in both Chamfer distance and the mean angular error for all objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.909, + 0.438 + ], + "angle": 0, + "content": "Figure 5 shows additional qualitative results. Our results are visually better than the other methods. The concave details of the vase and bottle are especially notable, as they are precisely reconstructed in our method while other results appear blurred. In addition, the surface of the container's lid is only reconstructed by our method while there is a big hole in the results of the others. Some artifacts are seen in thin parts in our result, particularly in the neck of the bottle. This artifact comes from the low spatial resolution of the thermal camera, which can be reduced by using a higher-resolution camera or increasing camera views." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.47, + 0.634, + 0.485 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.638 + ], + "angle": 0, + "content": "In this paper, we propose a 3D shape reconstruction using multi-view thermal polarimetric observations. We theoretically show that thermal polarimetric cues, especially AoLP images, are independent of material properties and illumination environment. Based on this property, we demonstrate a detailed 3D shape recovery using multi-view thermal AoLP images and show that our approach is applicable to heterogeneous objects including transparent and low-reflective objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Although the proposed method performs well, some challenges remain in expanding the applicable scenes. Our approach assumes the sufficient power of the emission component compared to the reflection component. This assumption is not valid when the hot objects are placed near the target object or when the target's temperature is lower than the environment. Moreover, our method is less effective on metallic surfaces or rough surfaces, where the polarization signal can become unstable or noisy. In such cases, decomposing the emission and reflection components [48] before applying the shape recovery or integrating additional cues may be a good option. Improving the observation system is also an important factor since our capturing system needs some time to rotate the polarization filter. A single-shot polarized LWIR camera could accelerate the thermal polarization analysis in the near future, as in the visible-light spectrum." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27397" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.115, + 0.482, + 0.145 + ], + "angle": 0, + "content": "This work was supported by JSPS KAKENHI Grant Numbers JP24K20802, JP22K18420, JP24K02965." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.159, + 0.187, + 0.174 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.185, + 0.483, + 0.224 + ], + "angle": 0, + "content": "[1] G.A. Atkinson and E.R. Hancock. Recovery of surface orientation from diffuse polarization. IEEE TIP, 15(6):1653-1664, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.226, + 0.483, + 0.28 + ], + "angle": 0, + "content": "[2] Yunhao Ba, Alex Gilbert, Franklin Wang, Jina Yang, Rui Chen, Yiqin Wang, Lei Yan, Boxin Shi, and Achuta Kadambi. Deep Shape from Polarization. In ECCV, pages 554-571, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.283, + 0.482, + 0.323 + ], + "angle": 0, + "content": "[3] Seung-Hwan Baek, Daniel S. Jeon, Xin Tong, and Min H. Kim. Simultaneous acquisition of polarimetric SVBRDF and normals. ACM TOG, 37(6):1-15, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[4] P.J. Besl and Neil D. McKay. A method for registration of 3-D shapes. IEEE TPAMI, 14(2):239-256, 1992." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.482, + 0.394 + ], + "angle": 0, + "content": "[5] Xu Cao, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita. Multi-View Azimuth Stereo via Tangent Space Consistency. In CVPR, pages 825–834, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[6] Qian Chen, Shihao Shu, and Xiangzhi Bai. Thermal3D-GS: Physics-Induced 3D Gaussians for Thermal Infrared Novel-View Synthesis. In ECCV, pages 253–269, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[7] Charly Collin, Sumanta Pattanaik, Patrick LiKamWa, and Kadi Bouatouch. Computation of polarized subsurface BRDF for rendering. In Proceedings of Graphics Interface 2014, pages 201-208. Canadian Information Processing Society, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.482, + 0.548 + ], + "angle": 0, + "content": "[8] Zhaopeng Cui, Jinwei Gu, Boxin Shi, Ping Tan, and Jan Kautz. Polarimetric Multi-view Stereo. In CVPR, pages 369-378, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[9] Unay Dorken Gallastegi, Hoover Rueda-Chacón, Martin J. Stevens, and Vivek K Goyal. Absorption-based hyperspectral thermal ranging: Performance analyses, optimization, and simulations. Optics Express, 32(1):151, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.482, + 0.675 + ], + "angle": 0, + "content": "[10] Jean-Denis Durou, Maurizio Falcone, Yvain Quéau, and Silvia Tozza. A Comprehensive Introduction to Photometric 3D-Reconstruction. In Advances in Photometric 3D-Reconstruction, pages 1-29. Springer International Publishing, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.678, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[11] Jack Erdozain, Kazuto Ichimaru, Tomohiro Maeda, Hiroshi Kawasaki, Ramesh Raskar, and Achuta Kadambi. 3d Imaging For Thermal Cameras Using Structured Light. In ICIP, pages 2795-2799, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[12] Gonen Eren, Olivier Aubreton, Fabrice Meriaudeau, L.A. Sanchez Secades, David Fofi, A. Teoman Naskali, Frederic Truchetet, and Aytul Ercil. Scanning from heating: 3D shape estimation of transparent objects from local surface heating. Optics Express, 17(14):11457, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[13] Yoshiki Fukao, Ryo Kawahara, Shohei Nobuhara, and Ko Nishino. Polarimetric Normal Stereo. In CVPR, pages 682-690, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Unay Dorken Gallastegi, Hoover Rueda-Chacon, Martin J. Stevens, and Vivek K Goyal. Absorption-Based, Passive Range Imaging from Hyperspectral Thermal Measurements. IEEE TPAMI, pages 1-18, 2025." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.185, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] S. Garrido-Jurado, R. Muñoz-Salinas, F.J. Madrid-Cuevas, and M.J. Marín-Jiménez. Automatic generation and detection of highly reliable fiducial markers under occlusion. PR, 47(6):2280-2292, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.904, + 0.175 + ], + "angle": 0, + "content": "[16] Jason Geng. Structured-light 3D surface imaging: A tutorial. Advances in Optics and Photonics, 3(2):128, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[17] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit Geometric Regularization for Learning Shapes, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[18] Yufei Han, Heng Guo, Koki Fukai, Hiroaki Santo, Boxin Shi, Fumio Okura, Zhanyu Ma, and Yunpeng Jia. NeRSP: Neural 3D Reconstruction for Reflective Objects with Sparse Polarized Images. In CVPR, pages 11821-11830, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.904, + 0.314 + ], + "angle": 0, + "content": "[19] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2nd ed edition, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.904, + 0.356 + ], + "angle": 0, + "content": "[20] A. Hertzmann and S.M. Seitz. Example-based photometric stereo: Shape reconstruction with general, varying BRDFs. IEEE TPAMI, 27(8):1254-1264, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.904, + 0.425 + ], + "angle": 0, + "content": "[21] Inseung Hwang, Daniel S. Jeon, Adolfo Muñoz, Diego Gutierrez, Xin Tong, and Min H. Kim. Sparse ellipsometry: Portable acquisition of polarimetric SVBRDF and shape with unstructured flash photography. ACM TOG, 41(4):1-14, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.904, + 0.482 + ], + "angle": 0, + "content": "[22] Tomoki Ichikawa, Matthew Purri, Ryo Kawahara, Shohei Nobuhara, Kristin Dana, and Ko Nishino. Shape from Sky: Polarimetric Normal Recovery Under The Sky. In CVPR, pages 14827-14836, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[23] Masaki Kaga, Takahiro Kushida, Tsuyoshi Takatani, Kenichiro Tanaka, Takuya Funatomi, and Yasuhiro Mukaigawa. Thermal non-line-of-sight imaging from specular and diffuse reflections. IPSJ Transactions on Computer Vision and Applications, 11(1):8, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[24] Takahiro Kushida, Ryutaro Nakamura, Hiroaki Matsuda, Wenhao Chen, and Kenichiro Tanaka. Affine transform representation for reducing calibration cost on absorption-based LWIR depth sensing. Scientific Reports, 14(1):26429, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.661 + ], + "angle": 0, + "content": "[25] Chenyang Lei, Chenyang Qi, Jiaxin Xie, Na Fan, Vladlen Koltun, and Qifeng Chen. Shape from Polarization for Complex Scenes in the Wild. In CVPR, pages 12622-12631, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[26] Daniel A. LeMaster and Michael T. Eismann. Passive Polarimetric Imaging. In Multi-Dimensional Imaging, pages 391-428. Wiley, 1 edition, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[27] Chenhao Li, Taishi Ono, Takeshi Uemori, Hajime Mihara, Alexander Gatto, Hajime Nagahara, and Yusuke Moriuchi. NeISF: Neural Incident Stokes Field for Geometry and Material Estimation. In CVPR, pages 21434-21445, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.904, + 0.802 + ], + "angle": 0, + "content": "[28] Yvette Y. Lin, Xin-Yi Pan, Sara Fridovich-Keil, and Gordon Wetzstein. ThermalNeRF: Thermal Radiance Fields. In ICCP, pages 1–12, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[29] Tomohiro Maeda, Yiqin Wang, Ramesh Raskar, and Achuta Kadambi. Thermal Non-Line-of-Sight Imaging. In ICCP, pages 1-11, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27398" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[31] Seiichi Mita, Xu Yuquan, Kazuhisa Ishimaru, and Sakiko Nishino. Robust 3D Perception for any Environment and any Weather Condition using Thermal Stereo. In IEEE Intelligent Vehicles Symposium (IV), pages 2569-2574, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[32] Miyazaki, Tan, Hara, and Ikeuchi. Polarization-based inverse rendering from a single view. In ICCV, pages 982-987 vol.2, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[33] Daisuke Miyazaki, Megumi Saito, Yoichi Sato, and Katsushi Ikeuchi. Determining surface orientations of transparent objects based on polarization degrees in visible and infrared wavelengths. J. Opt. Soc. Am. A, 19(4):687, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[34] D. Miyazaki, M. Kagesawa, and K. Ikeuchi. Transparent surface modeling from a pair of polarization images. IEEE TPAMI, 26(1):73-82, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[35] Yasuto Nagase, Takahiro Kushida, Kenichiro Tanaka, Takuya Funatomi, and Yasuhiro Mukaigawa. Shape from Thermal Radiation: Passive Ranging Using Multi-spectral LWIR Measurements. In CVPR, pages 12651-12661, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[36] Sriram Narayanan, Mani Ramanagopal, Mark Sheinin, Aswin C. Sankaranarayanan, and Srinivasa G. Narasimhan. Shape from heat conduction. In ECCV, pages 426-444, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[37] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation. In CVPR, pages 165-174, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[38] Mani Ramanagopal, Sriram Narayanan, Aswin C. Sankaranarayanan, and Srinivasa G. Narasimhan. A Theory of Joint Light and Heat Transport for Lambertian Scenes. In CVPR, pages 11924-11933, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[40] Joaquin Rodriguez, Lew-Fock-Chong Lew-Yan-Voon, Renato Martins, and Olivier Morel. Pola4All: Survey of polarimetric applications and an open-source toolkit to analyze polarization. Journal of Electronic Imaging, 33(01), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.674, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[41] Hiroaki Santo, Masaki Samejima, Yusuke Sugano, Boxin Shi, and Yasuyuki Matsushita. Deep Photometric Stereo Networks for Determining Surface Normal and Reflectances. IEEE TPAMI, 44(1):114-128, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.73, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[42] Philip Saponaro, Scott Sorensen, Abhishek Kolagunda, and Chandra Kambhametty. Material classification with thermal imagery. In CVPR, pages 4649-4656, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[43] D. Scharstein and R. Szeliski. High-accuracy stereo depth maps using structured light. In CVPR, pages I-195-I-202, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[44] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In CVPR, pages 4104-4113, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[45] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise View Selection for Unstructured Multi-View Stereo. In ECCV, pages 501–518, 2016." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[46] Boxin Shi, Jina Yang, Jinwei Chen, Ruihua Zhang, and Rui Chen. Recent Progress in Shape from Polarization. In Advances in Photometric 3D-Reconstruction, pages 177-203. Springer International Publishing, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[47] Richard Szeliski. Computer Vision: Algorithms and Applications. Springer International Publishing, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[48] Kenichiro Tanaka, Nobuhiro Ikeya, Tsuyoshi Takatani, Hiroyuki Kubo, Takuya Funatomi, Vijay Ravi, Achuta Kadambi, and Yasuhiro Mukaigawa. Time-Resolved Far Infrared Light Transport Decomposition for Thermal Photometric Stereo. IEEE TPAMI, 43(6):2075–2085, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[49] L.B. Wolff and T.E. Boult. Constraining object features using a polarization reflectance model. IEEE TPAMI, 13(7): 635-657, 1991." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[50] Robert J. Woodham. Photometric Method For Determining Surface Orientation From Multiple Images. Optical Engineering, 19(1), 1980." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[51] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. NeurIPS, 33, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[52] Tianxiang Ye, Qi Wu, Junyuan Deng, Guoqing Liu, Liu Liu, Songpengcheng Xia, Liang Pang, Wenxian Yu, and Ling Pei. Thermal-NeRF: Neural Radiance Fields from an Infrared Camera. In IROS, pages 1046–1053, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[53] Jincheng Zhang, Andrew R Willis, and Kevin Brink. Photometric Correction for Infrared Sensors. In CVPRW, pages 431-439, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[54] Z. Zhang. A flexible new technique for camera calibration. IEEE TPAMI, 22(11):1330-1334, 2000." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.561 + ], + "angle": 0, + "content": "[55] Jinyu Zhao, Yusuke Monno, and Masatoshi Okutomi. Polarimetric Multi-View Inverse Rendering. IEEE TPAMI, 45 (7):8798-8812, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27399" + } + ] +] \ No newline at end of file diff --git a/2025/Thermal Polarimetric Multi-view Stereo/493fd467-1c08-4433-ac25-63db3b58338a_origin.pdf b/2025/Thermal Polarimetric Multi-view Stereo/493fd467-1c08-4433-ac25-63db3b58338a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d204cc40ae58038f3ff6237b2b3db01bc41a5b22 --- /dev/null +++ b/2025/Thermal Polarimetric Multi-view Stereo/493fd467-1c08-4433-ac25-63db3b58338a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bcf874a07fc710de916aa32b73daaef7edcdd5dc1687296e18bb1009759dfe09 +size 3303719 diff --git a/2025/Thermal Polarimetric Multi-view Stereo/full.md b/2025/Thermal Polarimetric Multi-view Stereo/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8cec2154c5bdf11de138ae3f233855fcd19b299f --- /dev/null +++ b/2025/Thermal Polarimetric Multi-view Stereo/full.md @@ -0,0 +1,403 @@ +# Thermal Polarimetric Multi-view Stereo + +Takahiro Kushida Kenichiro Tanaka Ritsumeikan University 2-150 Iwakura-cho, Ibaraki, Osaka, Japan {tkushida,ken-t}@fc.ritsumei.ac.jp + +# Abstract + +This paper introduces a novel method for detailed 3D shape reconstruction utilizing thermal polarization cues. Unlike state-of-the-art methods, the proposed approach is independent of illumination and material properties. In this paper, we formulate a general theory of polarization observation and show that long-wave infrared (LWIR) polarimetric imaging is free from the ambiguities that affect visible polarization analyses. Subsequently, we propose a method for recovering detailed 3D shapes using multi-view thermal polarimetric images. Experimental results demonstrate that our approach effectively reconstructs fine details in transparent, translucent, and heterogeneous objects, outperforming existing techniques. + +# 1. Introduction + +3D shape reconstruction is a fundamental problem in computer vision, with applications spanning both academia and industry. Over the years, various 3D reconstruction methods using RGB cameras have been proposed, but many of these methods rely on strong assumptions about illumination and reflection models. For instance, multi-view stereo [19, 44, 45, 47] depends on surface texture to find correspondences, while structured light [16, 43], photometric stereo [10, 20, 41, 50], and shape from polarization [1, 2, 32, 34, 46, 49] require specific lighting conditions and assume opaque surfaces, limiting their applicability to a narrow range of materials. + +Thermal imaging presents an attractive alternative for 3D reconstruction in challenging scenarios because it does not rely on environmental illumination; any object with heat emits long-wave infrared (LWIR) light, effectively serving as its own light source. Furthermore, most materials, except for those designed for thermal optics, are opaque in the LWIR spectrum, eliminating the need for many of the assumptions required by traditional reflection models. Building on this concept, several methods, such as thermal multi-view stereo [53] and absorption-based depth es + +timation [9, 14, 24, 35] have been proposed. Although these methods demonstrate potential applicability, their results also imply the limitation inherent in model-based approaches, especially with regard to accuracy. + +Photometric cues in the LWIR spectrum are useful for recovering the accurate shape, similar to the visible-light spectrum. For example, thermal photometric stereo [48] and shape from heat conduction [36] have demonstrated detailed surface reconstruction. However, these methods require active heating and cooling of the object, making them both time-consuming and impractical. In this paper, we aim to reconstruct fine 3D shapes in a steady state using polarimetric cues, a type of photometric cue that can be obtained without monitoring the heating or cooling processes. + +Table 1 presents a comparison of various 3D reconstruction approaches. Our approach is illumination-independent, material-independent, free from heating or cooling process, and highly accurate. Our contributions are twofold. First, we establish a unified theoretical framework for LWIR polarization, showing that, unlike visible-light polarization, it is not affected by specular-diffuse ambiguities. Second, we show that LWIR polarization serves as a powerful cue for detailed 3D shape reconstruction. Experimental results confirm that our method outperforms the existing approaches on heterogeneous materials. + +# 2. Related Work + +# 2.1. Shape from visible polarization + +Shape from Polarization (SfP), which recovers surface normals from the polarization state of reflected light, has a long-standing history in computer vision [1, 2, 32, 34, 46, 49]. Recently, the advent of snapshot polarization cameras has significantly boosted this area of research [40]. + +SfP approaches continue to tackle ambiguities in surface normals that arise from the mixture of specular and diffuse polarizations [46]. One effective way to resolve these ambiguities is to use multi-view polarization images, which provide a richer set of constraints on the surface normals and lead to more accurate and robust surface reconstruc + +
MethodIlluminationMaterialAccuracyMeasurement
Visible Multi-view Stereo [19, 44, 45, 47]DependentTextured onlyHighEasy
Visible SL / PS [10, 16, 20, 41, 43, 50]DependentOpaque onlyHighEasy
Visible Polarization [1, 2, 32, 34, 46, 49]DependentDependentHighEasy
Visible NeRFs [30, 51]DependentWide range of materialsLowEasy
Thermal Multi-view Stereo [53]IndependentTextured onlyHighEasy
Thermal SL [11]IndependentDiffuse reflective onlyHighEasy
Thermal PS [48]IndependentWide range of materialsModerateHeating & cooling
Thermal NeRFs [28, 52]IndependentWide range of materialsLowEasy
Heat Conduction [36]IndependentWide range of materialsModerateHeating
Thermal Polarization (Ours)IndependentWide range of materialsHighEasy
+ +Table 1. Comparison table of shape reconstruction methods. While other methods have pros and cons, our approach is independent of illumination and material's optical properties as well as accurate shape recovery is possible in steady state measurements. + +tions [8]. Another approach involves using polarimetric BRDF models, which can handle complex material properties [3, 13, 21, 55]. Recent approaches incorporate neural scene representations and differentiable rendering, enabling the joint estimation of shape, environment maps, and reflectance [18, 27]. Nevertheless, SfP in the visible light spectrum remains fundamentally sensitive to the illumination conditions, such as the assumption of uniform, unpolarized ambient light, and to variations in material's optical properties, including transparency, translucency, matte, or shiny finishes [46]. + +In this paper, we introduce the use of polarization in the LWIR spectrum. The LWIR light is emitted directly from the object and is independent of external illumination. Moreover, the polarization of LWIR emission is not mixed up with the diffuse and specular reflections or transmission effects. These properties make LWIR polarization advantageous for robust shape reconstruction. + +# 2.2. LWIR imaging and physics-based vision + +Beyond 3D reconstruction, thermal imaging is utilized for a variety of physics-based vision tasks. The analysis of heat conduction, for instance, has been leveraged for material classification by exploiting differences in thermal diffusivity and conductivity [42]. Reflection analysis of LWIR light has facilitated non-line-of-sight (NLoS) imaging, expanding scene understanding by examining how thermal rays propagate around the corner [23, 29]. Joint analyses of both visible and LWIR light transport have further enriched scene understanding [38]. We also advance the area of physics-based thermal vision through a shape reconstruction using LWIR polarization. + +Thermal imaging is an effective approach for 3D shape reconstruction in some challenging scenarios. For example, it is effective for stealth observation as the thermal radiation does not rely on external illumination [6, 35, 52, 53]. Another example is that it is feasible to observe the shape of challenging materials in the visible light spectrum because many materials that appear transparent or translucent are opaque in the LWIR spectrum [48]. + +Existing 3D reconstruction techniques using thermal cameras can be broadly categorized into geometric and photometric approaches. Geometric methods [12, 31, 53] such as based on stereo/multi-view observations recover surface depth by matching corresponding points across thermal images, commonly assuming near-Lambertian emission or sufficient texture. Photometric approaches infer depth, shape, or surface normal from intensity cues, which include thermal photometric stereo [48], absorption-based method [9, 14, 24, 35], and heat conduction-based method [36]. While they show strong possibilities of shape reconstruction using a thermal camera, their accuracy is generally limited. There are hybrid approaches that combine visible-light and LWIR polarization [34], where the LWIR polarization is used only as an auxiliary cue to resolve the specular-diffuse ambiguity in visible polarization rather than directly contributing to shape recovery. + +Recently, neural rendering methods inspired by Neural Radiance Fields (NeRF) have been explored in the thermal domain [6, 28, 52] as well. By jointly optimizing a volumetric scene representation and a rendering function to satisfy the rendered images and align observed images from multiple viewpoints, these thermal NeRF techniques recover both geometry and radiometric properties. We are also inspired by this neural representation and model the object's surface by a neural implicit surface. + +In this paper, we leverage the polarization state of thermal radiation to enrich the information available for 3D reconstruction. By incorporating the polarization cues, geometric rotation, and neural representation, our method enables more accurate shape estimation. + +# 3. Polarimetric LWIR Observation + +In this section, we present a general theory of polarization observation and discuss the advantages of thermal polarization over visible polarization. + +# 3.1. Physics of Polarization + +The polarization state of light, including both visible and LWIR spectra, changes when it is reflected from or trans + +![](images/862fcd43ab96a685b0f2e6a38c8d26efb11173cd27ce0f73f8bca444bdb22e57.jpg) +Figure 1. Coordinate systems and the zenith and projected azimuth angles of the surface normal. While Mueller calculus is performed in the object's surface coordinate system, the observed Stokes parameter and projected azimuth angle are represented in the image coordinate system. + +mitted through an object's surface. The transport of polarized light can be described using the Stokes parameters $\boldsymbol{s} = [s_0,s_1,s_2,s_3]^{\mathrm{T}}$ and the Mueller matrix $M$ as + +$$ +\boldsymbol {s} _ {o} = \boldsymbol {M} \boldsymbol {s} _ {i}, \tag {1} +$$ + +where $s_i$ and $s_o$ denote the Stokes parameters of the incoming and outgoing light, respectively, both defined in the surface coordinate system [7]. + +The observed scene is a combination of specular reflection, diffuse reflection, transmission, and emission. Consequently, the overall polarization state $s$ can be expressed as + +$$ +\boldsymbol {s} = \boldsymbol {s} _ {s} + \boldsymbol {s} _ {d} + \boldsymbol {s} _ {t} + \boldsymbol {s} _ {e}, \tag {2} +$$ + +where $s_s, s_d, s_t$ , and $s_e$ represent the Stokes vectors corresponding to the specular, diffuse, transmission, and emission components, respectively. + +Specular polarization The change in the polarization state due to specular reflection [3, 22] is described by + +$$ +\boldsymbol {s} _ {s} = \boldsymbol {C} (\varphi) \boldsymbol {F} ^ {\mathrm {R}} (\theta) \boldsymbol {s} _ {i}, \tag {3} +$$ + +where $s_i$ denotes the Stokes vector of the incident light in the object's surface coordinate system, $s_s$ denotes the corresponding Stokes vector in the image coordinate system, $C(\varphi)$ is a rotation matrix that transforms coordinates from the image to the object's surface, $\varphi$ is the projected azimuth angle (i.e. the angle between the vertical axis of the image plane and the surface normal projected into the image plane), $F^{\mathrm{R}}(\theta)$ is the Mueller matrix for the specular component, and $\theta$ is the zenith angle of the reflection (i.e. the angle between the surface normal and the viewing direction). The + +surface coordinate system, image coordinate system, zenith angle, and projected azimuth angle are illus in Fig. 1. The rotation matrix $C(\varphi)$ is given by + +$$ +\boldsymbol {C} (\varphi) = \left[ \begin{array}{c c c c} 1 & 0 & 0 & 0 \\ 0 & \cos 2 \varphi & - \sin 2 \varphi & 0 \\ 0 & \sin 2 \varphi & \cos 2 \varphi & 0 \\ 0 & 0 & 0 & 1 \end{array} \right]. \tag {4} +$$ + +The Mueller matrix $\pmb{F}^{\mathrm{R}}(\theta)$ is derived from Fresnel's law and expressed as + +$$ +\boldsymbol {F} ^ {\mathrm {R}} (\theta) = \left[ \begin{array}{l l l l} r _ {+} & r _ {-} & 0 & 0 \\ r _ {-} & r _ {+} & 0 & 0 \\ 0 & 0 & r _ {\times} \cos \delta & r _ {\times} \sin \delta \\ 0 & 0 & - r _ {\times} \sin \delta & r _ {\times} \cos \delta \end{array} \right], \quad (5) +$$ + +where $r_{\pm} = \frac{r_s \pm r_p}{2}$ , $r_{\times} = \sqrt{r_p r_s}$ , $\theta'$ denotes the refractive angle. Here, $\cos \delta$ equals $-1$ when $\theta$ is less than Brewster's angle and $1$ otherwise, with the sign of $\sin \delta$ reversed accordingly [3]. The Fresnel reflectances for s- and p-polarization, $r_s$ and $r_p$ , are given by + +$$ +r _ {s} (\theta) = \left(\frac {\cos \theta - \eta \cos \theta^ {\prime}}{\cos \theta + \eta \cos \theta^ {\prime}}\right) ^ {2}, r _ {p} (\theta) = \left(\frac {\cos \theta^ {\prime} - \eta \cos \theta}{\cos \theta^ {\prime} + \eta \cos \theta}\right) ^ {2}. \tag {6} +$$ + +According to Snell's law, the refractive angle is given as + +$$ +\theta^ {\prime} = \sin^ {- 1} \left(\frac {1}{\eta} \sin \theta\right), \tag {7} +$$ + +where $\eta$ represents the refractive index of the object. + +Diffuse polarization The change in the polarization state due to diffuse reflection or subsurface scattering [1, 22] is described by + +$$ +\boldsymbol {s} _ {d} = \boldsymbol {C} (\varphi) \boldsymbol {F} ^ {\mathrm {T}} (\theta) \int_ {\Omega} \boldsymbol {D} \boldsymbol {F} ^ {\mathrm {T}} (\theta) \boldsymbol {s} _ {i, \omega} d \omega , \tag {8} +$$ + +where $s_{i,\omega}$ denotes the Stokes vector of incident light coming from the direction $\omega$ . The matrix $D$ , which describes the depolarization, is defined as + +$$ +\boldsymbol {D} = \left[ \begin{array}{l l l l} \rho & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \end{array} \right], \tag {9} +$$ + +where $\rho$ denotes the proportion of the depolarization. The Mueller matrix for transmittance $\pmb{F}^{\mathrm{T}}(\theta)$ is given by + +$$ +\boldsymbol {F} ^ {\mathrm {T}} (\theta) = \left[ \begin{array}{c c c c} t _ {+} & t _ {-} & 0 & 0 \\ t _ {-} & t _ {+} & 0 & 0 \\ 0 & 0 & t _ {\times} \cos \delta & t _ {\times} \sin \delta \\ 0 & 0 & - t _ {\times} \sin \delta & t _ {\times} \cos \delta \end{array} \right], \tag {10} +$$ + +where $t_{\pm} = \frac{t_s \pm t_p}{2}$ , $t_{\times} = \sqrt{t_p t_s}$ . $t_p$ and $t_s$ denote the Fresnel transmittance coefficients for s- and p- polarization, respectively, and are expressed as + +$$ +t _ {s} (\theta) = \left(\frac {2 \cos \theta}{\cos \theta + \eta \cos \theta^ {\prime}}\right) ^ {2}, t _ {p} (\theta) = \left(\frac {2 \cos \theta}{\cos \theta^ {\prime} + \eta \cos \theta}\right) ^ {2}. \tag {11} +$$ + +Transmission polarization The polarization state also changes due to light transmission. According to the Fresnel's law, the transmission polarization is described by + +$$ +\boldsymbol {s} _ {t} = \boldsymbol {C} (\varphi) \boldsymbol {F} ^ {\mathrm {T}} (\theta) \boldsymbol {s} _ {i}, \tag {12} +$$ + +where $s_i$ denotes the Stokes vector of the light coming from under the surface. + +Emission polarization The polarization state of emitted light is described by + +$$ +\boldsymbol {s} _ {e} = \boldsymbol {C} (\varphi) \boldsymbol {E} (\theta) \boldsymbol {s} _ {c}, \tag {13} +$$ + +where $s_c$ denotes the Stokes vector of the emitted light including thermal radiation as well as any chemical or electrical emissions and $E(\theta)$ is the Mueller matrix corresponding to the emission component. Since there are many types of light emission, the elements of the emission Mueller matrix are not specified explicitly in general. + +# 3.2. LWIR Polarization Observation + +Polarization of visible light Since objects other than light sources do not emit light in the visible spectrum, the emission component is generally ignorable. Therefore, the observed polarization can be modeled as + +$$ +\boldsymbol {s} = \boldsymbol {s} _ {d} + \boldsymbol {s} _ {s} + \boldsymbol {s} _ {t}. \tag {14} +$$ + +The magnitude of each component varies in complex ways depending on the object's material properties and the illumination environment. This complexity makes polarization analyses more challenging, leading existing visible polarization studies to rely on various assumptions to facilitate the analysis. + +Polarization of LWIR light Most objects are known to be opaque in the LWIR spectrum; therefore, the transmission component can be safely ignored. Moreover, an interesting property of LWIR light polarization is that the relative contributions of each component can be controlled by the object's temperature. In a typical environments, where no objects are extremely hot in the surroundings (i.e., where there are no significant LWIR light sources in the surroundings) the magnitude of reflection components is small compared to that of the emission. In such cases, the LWIR polarization observation can be modeled as + +$$ +s = s _ {e}. \tag {15} +$$ + +Since LWIR emission primarily arises from black body radiation, Eq. (13) can be rewritten as + +$$ +\boldsymbol {s} _ {e} = \boldsymbol {C} (\varphi) \boldsymbol {\mathcal {E}} (\theta) \boldsymbol {s} _ {B} (T), \tag {16} +$$ + +where $\mathcal{E}(\theta)$ denotes the Mueller matrix for black body radiation, and $s_B(T)$ represents the Stokes vector of the radiation, which can be expressed as + +$$ +\boldsymbol {s} _ {B} (T) = \left[ \begin{array}{c} B (T) \\ 0 \\ 0 \\ 0 \end{array} \right], \tag {17} +$$ + +where $B$ denotes the black body radiation. According to Planck's law, $B$ is given by + +$$ +B (T) = \frac {2 h c ^ {2}}{\lambda^ {5}} \frac {1}{e ^ {\frac {h c}{\lambda k T}} - 1}, \tag {18} +$$ + +where $c$ is the speed of light, $k$ is the Boltzmann constant, $h$ is the Planck constant, $T$ is the object's absolute temperature, and $\lambda$ is the observation wavelength. Equation (17) demonstrates that the object's temperature $T$ controls the power of emission. + +The Mueller matrix of black body radiation $\mathcal{E}(\theta)$ [26, 33] is given by + +$$ +\boldsymbol {\mathcal {E}} (\theta) = \left[ \begin{array}{l l l l} \varepsilon_ {+} & 0 & 0 & 0 \\ \varepsilon_ {-} & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \\ 0 & 0 & 0 & 0 \end{array} \right], \tag {19} +$$ + +where $\varepsilon_{\pm} = \frac{\varepsilon_s \pm \varepsilon_p}{2}$ . $\varepsilon_p$ and $\varepsilon_s$ denote the emissivities for s- and p-polarization, respectively. According to the energy conservation law, also known as Kirchhoff's law, the sum of reflectance, transmittance, and emissivity is equal to 1. Since the transmittance is negligible, as discussed above, we obtain + +$$ +\varepsilon_ {s} (\theta) = 1 - r _ {s} (\theta) \tag {20} +$$ + +$$ +\varepsilon_ {p} (\theta) = 1 - r _ {p} (\theta). \tag {21} +$$ + +In summary, in the LWIR spectrum, the Stokes observation $s$ can be analytically represented without ambiguity, thereby simplifying the analysis of the polarization state, which is a significant advantage for shape reconstruction. + +Polarization to surface normal: LWIR polarization provides cues for an object's surface normal, which can be determined analytically. The degree of linear polarization (DoLP) is related to the zenith angle $\theta$ as follows: + +$$ +\begin{array}{l} \mathrm {D o L P} = \frac {\sqrt {s _ {1} ^ {2} + s _ {2} ^ {2}}}{s _ {0}} (22) \\ = \frac {\left(\eta - \frac {1}{\eta}\right) ^ {2} \sin^ {2} \theta}{2 + 2 \eta^ {2} - \left(\eta + \frac {1}{\eta}\right) \sin^ {2} \theta + 4 \cos \theta \sqrt {\eta^ {2} - \sin^ {2} \theta}}. (23) \\ \end{array} +$$ + +![](images/0a83ce182410d44b86e138c9b93a2f722fbf8ac0ba166d6d29ef2fd27e0d956e.jpg) +Figure 2. Visible and thermal polarization observations of spheres made from various materials under natural room lighting. While visible polarization is unstable due to variations in a material's optical properties, thermal polarimetric cues, especially AoLP images, remain consistent across different materials. + +The angle of linear polarization (AoLP) is equivalent to the azimuth angle $\varphi$ as follows: + +$$ +\mathrm {A o L P} = \frac {1}{2} \tan^ {- 1} \left(\frac {s _ {2}}{s _ {1}}\right) = \varphi . \tag {24} +$$ + +These correspondences enable the reconstruction of the surface normal from polarization observations. Especially, AoLP serves as a robust cue because it directly reflects the azimuth angle of the surface normal and is independent of material properties, whereas DoLP is material-dependent due to the inclusion of the refractive index term. + +# 3.3. Examples of Visible vs. LWIR Polarization + +In the visible light spectrum, specular, diffuse, and transmission polarization components are often mixed [2, 25, 46], resulting in complicated observation. Fig. 2 shows the polarimetric observations of spheres made from various materials; in the visible spectrum, unwanted polarization observations frequently appear. For example, because black stone exhibits strong specular polarization, the DoLP image is heavily affected by reflections from the surrounding environment. In the case of transparent glass sphere, the observation is notably contaminated by background textures. For translucent plastic materials, both DoLP and AoLP images are affected by surface texture and subsurface scattering. These unstable observations significantly complicate subsequent shape recovery. + +In contrast, in the LWIR spectrum, polarimetric observations remain consistent and independent of both material properties and environmental reflections, particularly in AoLP images. This consistency is a key advantage of thermal polarization cues, as it provides robustness against vari + +ations in materials and surrounding reflections. + +# 4. Thermal Polarimetric Multi-view Stereo + +Our goal is to recover the 3D shape using thermal polarimetric cues. As discussed in the previous section, the AoLP observation is invariant to material properties. To leverage this strong cue, we only use the AoLP images for material-independent 3D shape reconstruction. + +Our approach employs multi-view thermal AoLP images to estimate the Signed Distance Function (SDF), which is implicitly represented by a Multi-Layer Perceptron (MLP) optimized via differentiable rendering. To effectively incorporate the surface normal information present in AoLP images, we adopt the concept of tangent space constraint (TSC) proposed in a multi-view azimuth stereo (MVAS) [5]. + +# Shape representation + +The SDF is implicitly represented by an MLP $f(\pmb{x}; \pmb{\theta}) : \mathbb{R}^3 \times \mathbb{R}^d \to \mathbb{R}$ , where $\pmb{x} \in \mathbb{R}$ denotes a point in space and $\pmb{\theta} \in \mathbb{R}^d$ are the learnable parameters of the MLP. The zero-level set of the SDF defines the object's surface $\mathcal{M}$ as [37] + +$$ +\mathcal {M} (\boldsymbol {\theta}) = \{\boldsymbol {x} | f (\boldsymbol {x}; \boldsymbol {\theta}) = 0 \}. \tag {25} +$$ + +The MLP is optimized within the framework of implicit differentiable renderer (IDR) [51]. + +# Loss function + +We adopt the loss function proposed in MVAS [5], which is defined as, + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {T S C}} + \lambda_ {1} \mathcal {L} _ {\text {s i l h o u e t t e}} + \lambda_ {2} \mathcal {L} _ {\text {E i k o n a l}}, \tag {26} +$$ + +where $\mathcal{L}_{\mathrm{TSC}}$ is the tangent space consistency loss, $\mathcal{L}_{\mathrm{silhouette}}$ is the silhouette loss and $\mathcal{L}_{\mathrm{Eikonal}}$ is the Eikonal regularization. $\lambda_{1}$ and $\lambda_{2}$ are the weights assigned to the respective loss terms. + +Tangent space consistency loss [5] The AoLP captured from different camera positions enforces a strong constraint on surface normal estimation. For a unit normal vector $\pmb{n} \subset S^2 \in \mathbb{R}^3$ at a surface point $\pmb{x} \in \mathbb{R}^3$ and the projected azimuth angle $\varphi$ at the corresponding pixel, the following relationship holds: + +$$ +\boldsymbol {r} _ {1} ^ {\top} \boldsymbol {n} \cos \varphi - \boldsymbol {r} _ {2} ^ {\top} \boldsymbol {n} \sin \varphi = 0, \tag {27} +$$ + +where $\pmb{R} = [r_1, r_2, r_3]^\top$ denotes the rotation matrix of the camera pose. This equation can be rearranged as + +$$ +\boldsymbol {n} ^ {\top} \underbrace {\left(\boldsymbol {r} _ {1} \cos \varphi - \boldsymbol {r} _ {2} \sin \varphi\right)} _ {\boldsymbol {t} (\varphi)} = 0, \tag {28} +$$ + +where $t(\varphi)$ is called a projected tangent vector and the above equation is called tangent space constraint. The TSC loss is the sum of the square error of the tangent space constraint and is defined as + +$$ +\mathcal {L} _ {\mathrm {T S C}} = \frac {1}{P} \sum_ {\boldsymbol {x} \in \boldsymbol {X}} \frac {\sum_ {i = 1} ^ {N} \Phi_ {i} \left(\boldsymbol {n} ^ {\top} (\boldsymbol {x} ; \boldsymbol {\theta}) \boldsymbol {t} _ {i} (\boldsymbol {x})\right) ^ {2}}{\sum_ {i = 1} ^ {N} \Phi_ {i}} \tag {29} +$$ + +where $P$ denotes the number of pixel samples, $\Phi_{i}$ is a binary indicator of the visibility of point $\pmb{x}$ from the $i$ -th view, $\pmb{n}^{\top}(\pmb{x};\pmb{\theta})$ is the estimated surface normal at the point $\pmb{x}$ given network parameters $\pmb{\theta}$ , and $\pmb{t}_{i}(\pmb{x})$ is the tangent vector corresponding to the $i$ -th view's pixel onto which the point $\pmb{x}$ is projected. + +Silhouette loss [5, 51] The silhouette loss constrains the visual hull of the shape and is defined as + +$$ +\mathcal {L} _ {\text {s i l h o u e t t e}} = \frac {1}{\alpha P} \sum_ {\boldsymbol {x} \in \hat {\boldsymbol {X}}} \Psi (O (\Pi (\boldsymbol {x})), \sigma (\alpha f ^ {*})), \tag {30} +$$ + +where $\Psi$ denotes the cross entropy and $O$ is a binary mask indicating whether a pixel lies inside the object silhouette (as provided by the input). $P$ is the number of pixel samples, $\Pi$ is the projection from the world coordinate to the image coordinate, $\tilde{\pmb{X}}$ represents the pixels with no ray-surface interactions or outside the silhouette, $f^{*}$ represents the minimum SDF distance between the ray that does not have intersections and the object surface, $\sigma$ is a sigmoid function, and $\alpha$ controls the sharpness of the sigmoid function. + +Eikonal regularization [17] The Eikonal loss regularizes the gradient of SDF, enforcing the SDF to maintain its property of having a gradient close to 1: + +$$ +\mathcal {L} _ {\text {E i k o n a l}} = \mathbb {E} _ {\boldsymbol {x}} \left(\left(\left\| \boldsymbol {n} (\boldsymbol {x}; \boldsymbol {\theta}) \right\| _ {2} - 1\right) ^ {2}\right). \tag {31} +$$ + +$\mathbb{E}_x(\cdot)$ denotes the expected value computed over sample points $\pmb{x}$ that are randomly sampled within the bounding box. + +# 5. Experiments + +# 5.1. Setup + +Implementation Following MVAS [5], we use an 8-layer MLP with softmax activations, combined with positional encoding (dimension 10). The input 3D position and encoded features are skip-connected to the 4th layer of the network. The optimization is performed using the Adam optimizer with a learning rate of 0.001. The weights for the loss terms are set to $\lambda_{1} = 50$ (silhouette loss) and $\lambda_{2} = 0.1$ (Eikonal loss). Training is conducted for 50 epochs with a batch size of 4096 pixels, and the silhouette loss weight is halved every 10 epochs. Training takes approximately 2 hours on a single RTX A6000 GPU. + +![](images/3641a6f9bb07ad68ad30c19c074f61a0cd205f1688ea98c3f7372abdb2989620.jpg) +Figure 3. Camera system and setup. The system consists of a rotation stage to rotate the target and a thermal camera with a wire-grid polarizer. A visible polarization camera is used for camera pose estimation and for comparison with the visible MVAS. + +Camera system Figure 3 shows our camera system and experimental setup. Our thermal polarization camera system consists of a thermal camera (FLIR Boson 320) with $15\mathrm{mm}$ lens and a wire-grid polarizer (Thorlabs WP25M-IRC) mounted on a motorized rotation mount (Thorlabs K10CR2). The polarizer is installed between the image sensor and imaging lens to mitigate the narcissus reflections. Thermal polarization images are captured by rotating the polarizer using the rotation mount. We captured four polarization images at 0, 45, 90, and 135 degrees. A monochrome visible-light polarization camera (FLIR Blackfly S BFS-U3-51S5P) is also placed adjacent to the thermal camera. The visible-light polarization camera is used both to capture images for a baseline method and to get the poses of both visible and thermal cameras. + +Camera poses We employed a marker-based approach to estimate the camera poses. We put Aruco markers [15] on the rotation stage to estimate the pose of visible-light camera. Since printed markers are not visible in the thermal images, the poses of the thermal camera were obtained by transforming the camera poses of the visible-light camera. For stereo calibration between the visible-light and thermal cameras, we used a calibration target that can be observed by both cameras. The calibration board consists of a white-painted aluminum pegboard placed in front of a black surface heater. By heating the back panel, the circle grid pattern became visible to both cameras. The calibration parameters were estimated using Zhang's method [54]. + +Dataset We prepared a real-world dataset with ground-truth shapes. We captured 7 objects with various shapes and materials. The target objects include transparent and + +![](images/3dc83f94bbd3d1bc082ad55d0fdc720bf6979f8c241ef121a8acf3b9bdb3ded5.jpg) +Figure 4. Experimental results. AoLP images in both visible and LWIR spectra are shown below the scene photograph. Each method's result includes the recovered shape, surface normals, and angular error maps. The Chamfer distance is displayed below the estimated shape and the mean angular error is displayed to the right of the angular error map. For both metrics, lower values indicate better performance. The smallest errors are shown in magenta. It demonstrates that the proposed method outperforms the other methods. + +low-reflective objects, which pose challenges for existing reconstruction methods. The objects were placed on a rotating table and captured from about 20–30 views by rotating the table. To robustly capture thermal polarization, the objects were kept warm during the measurements. Masks for thermal images were generated by thresholding pixel values, while masks for visible images were generated by using Segment Anything Model 2 [39]. Ground-truth shapes were obtained using a structured light 3D scanner (EinScan-SP V2). To facilitate the ground-truth measurement, the surfaces of the objects were painted white to make them opaque. + +Baselines We compare our proposed method with two state-of-the-art methods: MVAS [5] based on visible-light polarization cues and IDR [51], which is one of the neural implicit surface reconstruction methods, with thermal image input. For MVAS with visible-light polarization images, there exists $\pm \frac{\pi}{2}$ ambiguity in surface normals due to + +the mixture of specular and diffuse polarization. Therefore, TSC loss is modified to account for the $\pm \frac{\pi}{2}$ ambiguity, as originally proposed in MVAS. Since our method does not exhibit this ambiguity, the modified loss function is applied only to visible light observation. + +Evaluation metrics We use two metrics to evaluate both the surface points and surface normals. Chamfer distance is the metric for evaluating the geometric error, which calculates the point cloud distance between the estimate and the ground-truth. The mean angular error is the metric for surface normal, which calculates the angular error of normal maps. For both metrics, a smaller value indicates a better result. Prior to computing these metrics, the estimated and the ground-truth meshes are aligned using ICP algorithm [4]. + +# 5.2. Results + +Figure 4 shows the results. The target objects are an owl made of ceramic with a clear coating, a glass container with + +![](images/999293604978cbac00a51ab636cb2162f1933fb058b2908e72623c1e9200977d.jpg) +Figure 5. Qualitative results for other objects. AoLP images are shown below the scene photos. For each method, the upper image is the estimated shape and the lower is the corresponding normal map. It is shown that our method successfully reconstructs fine details of the object, particularly the concave parts in the transparent vase and bottle, while other methods fail to reconstruct such details. + +and without interiors, and a black tumbler, some of which are challenging objects for the existing approaches, especially a container with cutlery. In AoLP observations, while visible light is affected by a mixture of reflections and transmission components, thermal observations remain consistent and align with the azimuth angle of the surface normal. This consistency leads to a clear difference in detailed 3D reconstruction results between visible polarization and thermal polarization. + +In the owl's result, while the detailed shape such as the eyes and the beak is missing in both visible MVAS and thermal IDR, they are successfully recovered in our result. Similarly, the relief pattern on the glass container is clearly visible in our results. For a black tumbler with a white lid shown in the fourth row, the AoLP observation on the lid in the visible-light spectrum is too noisy. In such a case, the surface normal of the lid is not reconstructible, resulting in + +a wavy surface. In contrast, our method recovers a reasonable shape due to the presence of a unique AoLP on the lid in LWIR polarization. Overall, visible MVAS tends to reconstruct with wavy artifacts as AoLP is noisy, and thermal IDR tends to recover swollen shapes as it does not consider the surface normal. + +To quantitatively compare the results, the estimated shapes are evaluated using Chamfer distance, and the surface normals are evaluated by mean angular error. Our method outperforms the other methods in both Chamfer distance and the mean angular error for all objects. + +Figure 5 shows additional qualitative results. Our results are visually better than the other methods. The concave details of the vase and bottle are especially notable, as they are precisely reconstructed in our method while other results appear blurred. In addition, the surface of the container's lid is only reconstructed by our method while there is a big hole in the results of the others. Some artifacts are seen in thin parts in our result, particularly in the neck of the bottle. This artifact comes from the low spatial resolution of the thermal camera, which can be reduced by using a higher-resolution camera or increasing camera views. + +# 6. Conclusion + +In this paper, we propose a 3D shape reconstruction using multi-view thermal polarimetric observations. We theoretically show that thermal polarimetric cues, especially AoLP images, are independent of material properties and illumination environment. Based on this property, we demonstrate a detailed 3D shape recovery using multi-view thermal AoLP images and show that our approach is applicable to heterogeneous objects including transparent and low-reflective objects. + +Although the proposed method performs well, some challenges remain in expanding the applicable scenes. Our approach assumes the sufficient power of the emission component compared to the reflection component. This assumption is not valid when the hot objects are placed near the target object or when the target's temperature is lower than the environment. Moreover, our method is less effective on metallic surfaces or rough surfaces, where the polarization signal can become unstable or noisy. In such cases, decomposing the emission and reflection components [48] before applying the shape recovery or integrating additional cues may be a good option. Improving the observation system is also an important factor since our capturing system needs some time to rotate the polarization filter. A single-shot polarized LWIR camera could accelerate the thermal polarization analysis in the near future, as in the visible-light spectrum. + +# Acknowledgment + +This work was supported by JSPS KAKENHI Grant Numbers JP24K20802, JP22K18420, JP24K02965. + +# References + +[1] G.A. Atkinson and E.R. Hancock. Recovery of surface orientation from diffuse polarization. IEEE TIP, 15(6):1653-1664, 2006. +[2] Yunhao Ba, Alex Gilbert, Franklin Wang, Jina Yang, Rui Chen, Yiqin Wang, Lei Yan, Boxin Shi, and Achuta Kadambi. Deep Shape from Polarization. In ECCV, pages 554-571, 2020. +[3] Seung-Hwan Baek, Daniel S. Jeon, Xin Tong, and Min H. Kim. Simultaneous acquisition of polarimetric SVBRDF and normals. ACM TOG, 37(6):1-15, 2018. +[4] P.J. Besl and Neil D. McKay. A method for registration of 3-D shapes. IEEE TPAMI, 14(2):239-256, 1992. +[5] Xu Cao, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita. Multi-View Azimuth Stereo via Tangent Space Consistency. In CVPR, pages 825–834, 2023. +[6] Qian Chen, Shihao Shu, and Xiangzhi Bai. Thermal3D-GS: Physics-Induced 3D Gaussians for Thermal Infrared Novel-View Synthesis. In ECCV, pages 253–269, 2025. +[7] Charly Collin, Sumanta Pattanaik, Patrick LiKamWa, and Kadi Bouatouch. Computation of polarized subsurface BRDF for rendering. In Proceedings of Graphics Interface 2014, pages 201-208. Canadian Information Processing Society, 2014. +[8] Zhaopeng Cui, Jinwei Gu, Boxin Shi, Ping Tan, and Jan Kautz. Polarimetric Multi-view Stereo. In CVPR, pages 369-378, 2017. +[9] Unay Dorken Gallastegi, Hoover Rueda-Chacón, Martin J. Stevens, and Vivek K Goyal. Absorption-based hyperspectral thermal ranging: Performance analyses, optimization, and simulations. Optics Express, 32(1):151, 2024. +[10] Jean-Denis Durou, Maurizio Falcone, Yvain Quéau, and Silvia Tozza. A Comprehensive Introduction to Photometric 3D-Reconstruction. In Advances in Photometric 3D-Reconstruction, pages 1-29. Springer International Publishing, 2020. +[11] Jack Erdozain, Kazuto Ichimaru, Tomohiro Maeda, Hiroshi Kawasaki, Ramesh Raskar, and Achuta Kadambi. 3d Imaging For Thermal Cameras Using Structured Light. In ICIP, pages 2795-2799, 2020. +[12] Gonen Eren, Olivier Aubreton, Fabrice Meriaudeau, L.A. Sanchez Secades, David Fofi, A. Teoman Naskali, Frederic Truchetet, and Aytul Ercil. Scanning from heating: 3D shape estimation of transparent objects from local surface heating. Optics Express, 17(14):11457, 2009. +[13] Yoshiki Fukao, Ryo Kawahara, Shohei Nobuhara, and Ko Nishino. Polarimetric Normal Stereo. In CVPR, pages 682-690, 2021. +[14] Unay Dorken Gallastegi, Hoover Rueda-Chacon, Martin J. Stevens, and Vivek K Goyal. Absorption-Based, Passive Range Imaging from Hyperspectral Thermal Measurements. IEEE TPAMI, pages 1-18, 2025. + +[15] S. Garrido-Jurado, R. Muñoz-Salinas, F.J. Madrid-Cuevas, and M.J. Marín-Jiménez. Automatic generation and detection of highly reliable fiducial markers under occlusion. PR, 47(6):2280-2292, 2014. +[16] Jason Geng. Structured-light 3D surface imaging: A tutorial. Advances in Optics and Photonics, 3(2):128, 2011. +[17] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit Geometric Regularization for Learning Shapes, 2020. +[18] Yufei Han, Heng Guo, Koki Fukai, Hiroaki Santo, Boxin Shi, Fumio Okura, Zhanyu Ma, and Yunpeng Jia. NeRSP: Neural 3D Reconstruction for Reflective Objects with Sparse Polarized Images. In CVPR, pages 11821-11830, 2024. +[19] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2nd ed edition, 2003. +[20] A. Hertzmann and S.M. Seitz. Example-based photometric stereo: Shape reconstruction with general, varying BRDFs. IEEE TPAMI, 27(8):1254-1264, 2005. +[21] Inseung Hwang, Daniel S. Jeon, Adolfo Muñoz, Diego Gutierrez, Xin Tong, and Min H. Kim. Sparse ellipsometry: Portable acquisition of polarimetric SVBRDF and shape with unstructured flash photography. ACM TOG, 41(4):1-14, 2022. +[22] Tomoki Ichikawa, Matthew Purri, Ryo Kawahara, Shohei Nobuhara, Kristin Dana, and Ko Nishino. Shape from Sky: Polarimetric Normal Recovery Under The Sky. In CVPR, pages 14827-14836, 2021. +[23] Masaki Kaga, Takahiro Kushida, Tsuyoshi Takatani, Kenichiro Tanaka, Takuya Funatomi, and Yasuhiro Mukaigawa. Thermal non-line-of-sight imaging from specular and diffuse reflections. IPSJ Transactions on Computer Vision and Applications, 11(1):8, 2019. +[24] Takahiro Kushida, Ryutaro Nakamura, Hiroaki Matsuda, Wenhao Chen, and Kenichiro Tanaka. Affine transform representation for reducing calibration cost on absorption-based LWIR depth sensing. Scientific Reports, 14(1):26429, 2024. +[25] Chenyang Lei, Chenyang Qi, Jiaxin Xie, Na Fan, Vladlen Koltun, and Qifeng Chen. Shape from Polarization for Complex Scenes in the Wild. In CVPR, pages 12622-12631, 2022. +[26] Daniel A. LeMaster and Michael T. Eismann. Passive Polarimetric Imaging. In Multi-Dimensional Imaging, pages 391-428. Wiley, 1 edition, 2014. +[27] Chenhao Li, Taishi Ono, Takeshi Uemori, Hajime Mihara, Alexander Gatto, Hajime Nagahara, and Yusuke Moriuchi. NeISF: Neural Incident Stokes Field for Geometry and Material Estimation. In CVPR, pages 21434-21445, 2024. +[28] Yvette Y. Lin, Xin-Yi Pan, Sara Fridovich-Keil, and Gordon Wetzstein. ThermalNeRF: Thermal Radiance Fields. In ICCP, pages 1–12, 2024. +[29] Tomohiro Maeda, Yiqin Wang, Ramesh Raskar, and Achuta Kadambi. Thermal Non-Line-of-Sight Imaging. In ICCP, pages 1-11, 2019. +[30] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. + +[31] Seiichi Mita, Xu Yuquan, Kazuhisa Ishimaru, and Sakiko Nishino. Robust 3D Perception for any Environment and any Weather Condition using Thermal Stereo. In IEEE Intelligent Vehicles Symposium (IV), pages 2569-2574, 2019. +[32] Miyazaki, Tan, Hara, and Ikeuchi. Polarization-based inverse rendering from a single view. In ICCV, pages 982-987 vol.2, 2003. +[33] Daisuke Miyazaki, Megumi Saito, Yoichi Sato, and Katsushi Ikeuchi. Determining surface orientations of transparent objects based on polarization degrees in visible and infrared wavelengths. J. Opt. Soc. Am. A, 19(4):687, 2002. +[34] D. Miyazaki, M. Kagesawa, and K. Ikeuchi. Transparent surface modeling from a pair of polarization images. IEEE TPAMI, 26(1):73-82, 2004. +[35] Yasuto Nagase, Takahiro Kushida, Kenichiro Tanaka, Takuya Funatomi, and Yasuhiro Mukaigawa. Shape from Thermal Radiation: Passive Ranging Using Multi-spectral LWIR Measurements. In CVPR, pages 12651-12661, 2022. +[36] Sriram Narayanan, Mani Ramanagopal, Mark Sheinin, Aswin C. Sankaranarayanan, and Srinivasa G. Narasimhan. Shape from heat conduction. In ECCV, pages 426-444, 2025. +[37] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation. In CVPR, pages 165-174, 2019. +[38] Mani Ramanagopal, Sriram Narayanan, Aswin C. Sankaranarayanan, and Srinivasa G. Narasimhan. A Theory of Joint Light and Heat Transport for Lambertian Scenes. In CVPR, pages 11924-11933, 2024. +[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. +[40] Joaquin Rodriguez, Lew-Fock-Chong Lew-Yan-Voon, Renato Martins, and Olivier Morel. Pola4All: Survey of polarimetric applications and an open-source toolkit to analyze polarization. Journal of Electronic Imaging, 33(01), 2024. +[41] Hiroaki Santo, Masaki Samejima, Yusuke Sugano, Boxin Shi, and Yasuyuki Matsushita. Deep Photometric Stereo Networks for Determining Surface Normal and Reflectances. IEEE TPAMI, 44(1):114-128, 2022. +[42] Philip Saponaro, Scott Sorensen, Abhishek Kolagunda, and Chandra Kambhametty. Material classification with thermal imagery. In CVPR, pages 4649-4656, 2015. +[43] D. Scharstein and R. Szeliski. High-accuracy stereo depth maps using structured light. In CVPR, pages I-195-I-202, 2003. +[44] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In CVPR, pages 4104-4113, 2016. +[45] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise View Selection for Unstructured Multi-View Stereo. In ECCV, pages 501–518, 2016. + +[46] Boxin Shi, Jina Yang, Jinwei Chen, Ruihua Zhang, and Rui Chen. Recent Progress in Shape from Polarization. In Advances in Photometric 3D-Reconstruction, pages 177-203. Springer International Publishing, 2020. +[47] Richard Szeliski. Computer Vision: Algorithms and Applications. Springer International Publishing, 2022. +[48] Kenichiro Tanaka, Nobuhiro Ikeya, Tsuyoshi Takatani, Hiroyuki Kubo, Takuya Funatomi, Vijay Ravi, Achuta Kadambi, and Yasuhiro Mukaigawa. Time-Resolved Far Infrared Light Transport Decomposition for Thermal Photometric Stereo. IEEE TPAMI, 43(6):2075–2085, 2021. +[49] L.B. Wolff and T.E. Boult. Constraining object features using a polarization reflectance model. IEEE TPAMI, 13(7): 635-657, 1991. +[50] Robert J. Woodham. Photometric Method For Determining Surface Orientation From Multiple Images. Optical Engineering, 19(1), 1980. +[51] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. NeurIPS, 33, 2020. +[52] Tianxiang Ye, Qi Wu, Junyuan Deng, Guoqing Liu, Liu Liu, Songpengcheng Xia, Liang Pang, Wenxian Yu, and Ling Pei. Thermal-NeRF: Neural Radiance Fields from an Infrared Camera. In IROS, pages 1046–1053, 2024. +[53] Jincheng Zhang, Andrew R Willis, and Kevin Brink. Photometric Correction for Infrared Sensors. In CVPRW, pages 431-439, 2023. +[54] Z. Zhang. A flexible new technique for camera calibration. IEEE TPAMI, 22(11):1330-1334, 2000. +[55] Jinyu Zhao, Yusuke Monno, and Masatoshi Okutomi. Polarimetric Multi-View Inverse Rendering. IEEE TPAMI, 45 (7):8798-8812, 2023. \ No newline at end of file diff --git a/2025/Thermal Polarimetric Multi-view Stereo/images.zip b/2025/Thermal Polarimetric Multi-view Stereo/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f30e083995affee5d45825f7b77d904790435a7b --- /dev/null +++ b/2025/Thermal Polarimetric Multi-view Stereo/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de3def0599b0de782f043205a140aa31baa5fcac8cfd06b30a65b78c365df321 +size 598901 diff --git a/2025/Thermal Polarimetric Multi-view Stereo/layout.json b/2025/Thermal Polarimetric Multi-view Stereo/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4463d4d61904f19ba369e94ef0cdd74d0d3497a2 --- /dev/null +++ b/2025/Thermal Polarimetric Multi-view Stereo/layout.json @@ -0,0 +1,9141 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 181, + 103, + 430, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 103, + 430, + 119 + ], + "spans": [ + { + "bbox": [ + 181, + 103, + 430, + 119 + ], + "type": "text", + "content": "Thermal Polarimetric Multi-view Stereo" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 203, + 144, + 407, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 144, + 407, + 201 + ], + "spans": [ + { + "bbox": [ + 203, + 144, + 407, + 201 + ], + "type": "text", + "content": "Takahiro Kushida Kenichiro Tanaka Ritsumeikan University 2-150 Iwakura-cho, Ibaraki, Osaka, Japan {tkushida,ken-t}@fc.ritsumei.ac.jp" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 251, + 297, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 297, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 297, + 408 + ], + "type": "text", + "content": "This paper introduces a novel method for detailed 3D shape reconstruction utilizing thermal polarization cues. Unlike state-of-the-art methods, the proposed approach is independent of illumination and material properties. In this paper, we formulate a general theory of polarization observation and show that long-wave infrared (LWIR) polarimetric imaging is free from the ambiguities that affect visible polarization analyses. Subsequently, we propose a method for recovering detailed 3D shapes using multi-view thermal polarimetric images. Experimental results demonstrate that our approach effectively reconstructs fine details in transparent, translucent, and heterogeneous objects, outperforming existing techniques." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 430, + 135, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 135, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 135, + 442 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 450, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 594 + ], + "type": "text", + "content": "3D shape reconstruction is a fundamental problem in computer vision, with applications spanning both academia and industry. Over the years, various 3D reconstruction methods using RGB cameras have been proposed, but many of these methods rely on strong assumptions about illumination and reflection models. For instance, multi-view stereo [19, 44, 45, 47] depends on surface texture to find correspondences, while structured light [16, 43], photometric stereo [10, 20, 41, 50], and shape from polarization [1, 2, 32, 34, 46, 49] require specific lighting conditions and assume opaque surfaces, limiting their applicability to a narrow range of materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Thermal imaging presents an attractive alternative for 3D reconstruction in challenging scenarios because it does not rely on environmental illumination; any object with heat emits long-wave infrared (LWIR) light, effectively serving as its own light source. Furthermore, most materials, except for those designed for thermal optics, are opaque in the LWIR spectrum, eliminating the need for many of the assumptions required by traditional reflection models. Building on this concept, several methods, such as thermal multi-view stereo [53] and absorption-based depth es" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 276 + ], + "type": "text", + "content": "timation [9, 14, 24, 35] have been proposed. Although these methods demonstrate potential applicability, their results also imply the limitation inherent in model-based approaches, especially with regard to accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 277, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 396 + ], + "type": "text", + "content": "Photometric cues in the LWIR spectrum are useful for recovering the accurate shape, similar to the visible-light spectrum. For example, thermal photometric stereo [48] and shape from heat conduction [36] have demonstrated detailed surface reconstruction. However, these methods require active heating and cooling of the object, making them both time-consuming and impractical. In this paper, we aim to reconstruct fine 3D shapes in a steady state using polarimetric cues, a type of photometric cue that can be obtained without monitoring the heating or cooling processes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 397, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 556, + 529 + ], + "type": "text", + "content": "Table 1 presents a comparison of various 3D reconstruction approaches. Our approach is illumination-independent, material-independent, free from heating or cooling process, and highly accurate. Our contributions are twofold. First, we establish a unified theoretical framework for LWIR polarization, showing that, unlike visible-light polarization, it is not affected by specular-diffuse ambiguities. Second, we show that LWIR polarization serves as a powerful cue for detailed 3D shape reconstruction. Experimental results confirm that our method outperforms the existing approaches on heterogeneous materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 542, + 401, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 401, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 401, + 554 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 562, + 484, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 484, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 484, + 575 + ], + "type": "text", + "content": "2.1. Shape from visible polarization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 580, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 553, + 640 + ], + "type": "text", + "content": "Shape from Polarization (SfP), which recovers surface normals from the polarization state of reflected light, has a long-standing history in computer vision [1, 2, 32, 34, 46, 49]. Recently, the advent of snapshot polarization cameras has significantly boosted this area of research [40]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 712 + ], + "type": "text", + "content": "SfP approaches continue to tackle ambiguities in surface normals that arise from the mixture of specular and diffuse polarizations [46]. One effective way to resolve these ambiguities is to use multi-view polarization images, which provide a richer set of constraints on the surface normals and lead to more accurate and robust surface reconstruc" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27390" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 103, + 69, + 507, + 178 + ], + "blocks": [ + { + "bbox": [ + 103, + 69, + 507, + 178 + ], + "lines": [ + { + "bbox": [ + 103, + 69, + 507, + 178 + ], + "spans": [ + { + "bbox": [ + 103, + 69, + 507, + 178 + ], + "type": "table", + "html": "
MethodIlluminationMaterialAccuracyMeasurement
Visible Multi-view Stereo [19, 44, 45, 47]DependentTextured onlyHighEasy
Visible SL / PS [10, 16, 20, 41, 43, 50]DependentOpaque onlyHighEasy
Visible Polarization [1, 2, 32, 34, 46, 49]DependentDependentHighEasy
Visible NeRFs [30, 51]DependentWide range of materialsLowEasy
Thermal Multi-view Stereo [53]IndependentTextured onlyHighEasy
Thermal SL [11]IndependentDiffuse reflective onlyHighEasy
Thermal PS [48]IndependentWide range of materialsModerateHeating & cooling
Thermal NeRFs [28, 52]IndependentWide range of materialsLowEasy
Heat Conduction [36]IndependentWide range of materialsModerateHeating
Thermal Polarization (Ours)IndependentWide range of materialsHighEasy
", + "image_path": "9d66793122a34d89ecbe9640ea973fc3e7fbbbbf5e596ba516918561714f5992.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "type": "text", + "content": "Table 1. Comparison table of shape reconstruction methods. While other methods have pros and cons, our approach is independent of illumination and material's optical properties as well as accurate shape recovery is possible in steady state measurements." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 229, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 229, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 229, + 294, + 361 + ], + "type": "text", + "content": "tions [8]. Another approach involves using polarimetric BRDF models, which can handle complex material properties [3, 13, 21, 55]. Recent approaches incorporate neural scene representations and differentiable rendering, enabling the joint estimation of shape, environment maps, and reflectance [18, 27]. Nevertheless, SfP in the visible light spectrum remains fundamentally sensitive to the illumination conditions, such as the assumption of uniform, unpolarized ambient light, and to variations in material's optical properties, including transparency, translucency, matte, or shiny finishes [46]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 445 + ], + "type": "text", + "content": "In this paper, we introduce the use of polarization in the LWIR spectrum. The LWIR light is emitted directly from the object and is independent of external illumination. Moreover, the polarization of LWIR emission is not mixed up with the diffuse and specular reflections or transmission effects. These properties make LWIR polarization advantageous for robust shape reconstruction." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 455, + 266, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 266, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 266, + 468 + ], + "type": "text", + "content": "2.2. LWIR imaging and physics-based vision" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 616 + ], + "type": "text", + "content": "Beyond 3D reconstruction, thermal imaging is utilized for a variety of physics-based vision tasks. The analysis of heat conduction, for instance, has been leveraged for material classification by exploiting differences in thermal diffusivity and conductivity [42]. Reflection analysis of LWIR light has facilitated non-line-of-sight (NLoS) imaging, expanding scene understanding by examining how thermal rays propagate around the corner [23, 29]. Joint analyses of both visible and LWIR light transport have further enriched scene understanding [38]. We also advance the area of physics-based thermal vision through a shape reconstruction using LWIR polarization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Thermal imaging is an effective approach for 3D shape reconstruction in some challenging scenarios. For example, it is effective for stealth observation as the thermal radiation does not rely on external illumination [6, 35, 52, 53]. Another example is that it is feasible to observe the shape of challenging materials in the visible light spectrum because many materials that appear transparent or translucent are opaque in the LWIR spectrum [48]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 229, + 553, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 553, + 433 + ], + "type": "text", + "content": "Existing 3D reconstruction techniques using thermal cameras can be broadly categorized into geometric and photometric approaches. Geometric methods [12, 31, 53] such as based on stereo/multi-view observations recover surface depth by matching corresponding points across thermal images, commonly assuming near-Lambertian emission or sufficient texture. Photometric approaches infer depth, shape, or surface normal from intensity cues, which include thermal photometric stereo [48], absorption-based method [9, 14, 24, 35], and heat conduction-based method [36]. While they show strong possibilities of shape reconstruction using a thermal camera, their accuracy is generally limited. There are hybrid approaches that combine visible-light and LWIR polarization [34], where the LWIR polarization is used only as an auxiliary cue to resolve the specular-diffuse ambiguity in visible polarization rather than directly contributing to shape recovery." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 433, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 554, + 540 + ], + "type": "text", + "content": "Recently, neural rendering methods inspired by Neural Radiance Fields (NeRF) have been explored in the thermal domain [6, 28, 52] as well. By jointly optimizing a volumetric scene representation and a rendering function to satisfy the rendered images and align observed images from multiple viewpoints, these thermal NeRF techniques recover both geometry and radiometric properties. We are also inspired by this neural representation and model the object's surface by a neural implicit surface." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 540, + 554, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 600 + ], + "type": "text", + "content": "In this paper, we leverage the polarization state of thermal radiation to enrich the information available for 3D reconstruction. By incorporating the polarization cues, geometric rotation, and neural representation, our method enables more accurate shape estimation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 609, + 495, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 495, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 495, + 623 + ], + "type": "text", + "content": "3. Polarimetric LWIR Observation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 666 + ], + "type": "text", + "content": "In this section, we present a general theory of polarization observation and discuss the advantages of thermal polarization over visible polarization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 672, + 443, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 443, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 443, + 685 + ], + "type": "text", + "content": "3.1. Physics of Polarization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "The polarization state of light, including both visible and LWIR spectra, changes when it is reflected from or trans" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27391" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 70, + 265, + 228 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 265, + 228 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 265, + 228 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 265, + 228 + ], + "type": "image", + "image_path": "862fcd43ab96a685b0f2e6a38c8d26efb11173cd27ce0f73f8bca444bdb22e57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 296, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 291 + ], + "type": "text", + "content": "Figure 1. Coordinate systems and the zenith and projected azimuth angles of the surface normal. While Mueller calculus is performed in the object's surface coordinate system, the observed Stokes parameter and projected azimuth angle are represented in the image coordinate system." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 311, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 348 + ], + "type": "text", + "content": "mitted through an object's surface. The transport of polarized light can be described using the Stokes parameters " + }, + { + "bbox": [ + 55, + 311, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\boldsymbol{s} = [s_0,s_1,s_2,s_3]^{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 311, + 296, + 348 + ], + "type": "text", + "content": " and the Mueller matrix " + }, + { + "bbox": [ + 55, + 311, + 296, + 348 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 311, + 296, + 348 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 356, + 294, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 356, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 151, + 356, + 294, + 369 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {o} = \\boldsymbol {M} \\boldsymbol {s} _ {i}, \\tag {1}", + "image_path": "3f01d83f078c34de358ce97e25aaaf64ecd9c0a54d642e91d50871cd6a4f6347.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "type": "inline_equation", + "content": "s_o" + }, + { + "bbox": [ + 55, + 376, + 296, + 412 + ], + "type": "text", + "content": " denote the Stokes parameters of the incoming and outgoing light, respectively, both defined in the surface coordinate system [7]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 412, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 296, + 459 + ], + "type": "text", + "content": "The observed scene is a combination of specular reflection, diffuse reflection, transmission, and emission. Consequently, the overall polarization state " + }, + { + "bbox": [ + 55, + 412, + 296, + 459 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 412, + 296, + 459 + ], + "type": "text", + "content": " can be expressed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 126, + 470, + 294, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 470, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 126, + 470, + 294, + 482 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} = \\boldsymbol {s} _ {s} + \\boldsymbol {s} _ {d} + \\boldsymbol {s} _ {t} + \\boldsymbol {s} _ {e}, \\tag {2}", + "image_path": "809e77d15eb54efb5eaf1615d2cd5d647b705e557cbfe42f2f410202dd668584.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 490, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 526 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 490, + 296, + 526 + ], + "type": "inline_equation", + "content": "s_s, s_d, s_t" + }, + { + "bbox": [ + 55, + 490, + 296, + 526 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 490, + 296, + 526 + ], + "type": "inline_equation", + "content": "s_e" + }, + { + "bbox": [ + 55, + 490, + 296, + 526 + ], + "type": "text", + "content": " represent the Stokes vectors corresponding to the specular, diffuse, transmission, and emission components, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 564 + ], + "type": "text", + "content": "Specular polarization The change in the polarization state due to specular reflection [3, 22] is described by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 131, + 571, + 294, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 571, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 131, + 571, + 294, + 586 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {s} = \\boldsymbol {C} (\\varphi) \\boldsymbol {F} ^ {\\mathrm {R}} (\\theta) \\boldsymbol {s} _ {i}, \\tag {3}", + "image_path": "2b026a2d19b466210c41bb2f1908dbf075034e3010ea4062ff1570c68ba1131c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " denotes the Stokes vector of the incident light in the object's surface coordinate system, " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "s_s" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " denotes the corresponding Stokes vector in the image coordinate system, " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "C(\\varphi)" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " is a rotation matrix that transforms coordinates from the image to the object's surface, " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " is the projected azimuth angle (i.e. the angle between the vertical axis of the image plane and the surface normal projected into the image plane), " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{R}}(\\theta)" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " is the Mueller matrix for the specular component, and " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " is the zenith angle of the reflection (i.e. the angle between the surface normal and the viewing direction). The" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 73, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 554, + 109 + ], + "type": "text", + "content": "surface coordinate system, image coordinate system, zenith angle, and projected azimuth angle are illus in Fig. 1. The rotation matrix " + }, + { + "bbox": [ + 313, + 73, + 554, + 109 + ], + "type": "inline_equation", + "content": "C(\\varphi)" + }, + { + "bbox": [ + 313, + 73, + 554, + 109 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 115, + 554, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 115, + 554, + 167 + ], + "spans": [ + { + "bbox": [ + 354, + 115, + 554, + 167 + ], + "type": "interline_equation", + "content": "\\boldsymbol {C} (\\varphi) = \\left[ \\begin{array}{c c c c} 1 & 0 & 0 & 0 \\\\ 0 & \\cos 2 \\varphi & - \\sin 2 \\varphi & 0 \\\\ 0 & \\sin 2 \\varphi & \\cos 2 \\varphi & 0 \\\\ 0 & 0 & 0 & 1 \\end{array} \\right]. \\tag {4}", + "image_path": "c7fe7c21bd7e22c5a350dbcb90c291a84f12e5e347b94597ce60f86607a54ae1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 174, + 554, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 554, + 198 + ], + "type": "text", + "content": "The Mueller matrix " + }, + { + "bbox": [ + 313, + 174, + 554, + 198 + ], + "type": "inline_equation", + "content": "\\pmb{F}^{\\mathrm{R}}(\\theta)" + }, + { + "bbox": [ + 313, + 174, + 554, + 198 + ], + "type": "text", + "content": " is derived from Fresnel's law and expressed as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 205, + 554, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 205, + 554, + 257 + ], + "spans": [ + { + "bbox": [ + 340, + 205, + 554, + 257 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} ^ {\\mathrm {R}} (\\theta) = \\left[ \\begin{array}{l l l l} r _ {+} & r _ {-} & 0 & 0 \\\\ r _ {-} & r _ {+} & 0 & 0 \\\\ 0 & 0 & r _ {\\times} \\cos \\delta & r _ {\\times} \\sin \\delta \\\\ 0 & 0 & - r _ {\\times} \\sin \\delta & r _ {\\times} \\cos \\delta \\end{array} \\right], \\quad (5)", + "image_path": "e532441f0a58742aa91490f559fc6a61edb28af2ca0b876275b601245ed638eb.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "inline_equation", + "content": "r_{\\pm} = \\frac{r_s \\pm r_p}{2}" + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "inline_equation", + "content": "r_{\\times} = \\sqrt{r_p r_s}" + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "text", + "content": " denotes the refractive angle. Here, " + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "inline_equation", + "content": "\\cos \\delta" + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "text", + "content": " equals " + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "text", + "content": " is less than Brewster's angle and " + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "inline_equation", + "content": "1" + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "text", + "content": " otherwise, with the sign of " + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "inline_equation", + "content": "\\sin \\delta" + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "text", + "content": " reversed accordingly [3]. The Fresnel reflectances for s- and p-polarization, " + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "inline_equation", + "content": "r_s" + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "inline_equation", + "content": "r_p" + }, + { + "bbox": [ + 313, + 264, + 554, + 326 + ], + "type": "text", + "content": ", are given by" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 332, + 564, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 564, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 564, + 373 + ], + "type": "interline_equation", + "content": "r _ {s} (\\theta) = \\left(\\frac {\\cos \\theta - \\eta \\cos \\theta^ {\\prime}}{\\cos \\theta + \\eta \\cos \\theta^ {\\prime}}\\right) ^ {2}, r _ {p} (\\theta) = \\left(\\frac {\\cos \\theta^ {\\prime} - \\eta \\cos \\theta}{\\cos \\theta^ {\\prime} + \\eta \\cos \\theta}\\right) ^ {2}. \\tag {6}", + "image_path": "cc0ffffb1f6598cee87413d41542604d67adf1f66da9406c9db86611830f6773.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 380, + 541, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 380, + 541, + 393 + ], + "spans": [ + { + "bbox": [ + 314, + 380, + 541, + 393 + ], + "type": "text", + "content": "According to Snell's law, the refractive angle is given as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 385, + 400, + 554, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 400, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 385, + 400, + 554, + 427 + ], + "type": "interline_equation", + "content": "\\theta^ {\\prime} = \\sin^ {- 1} \\left(\\frac {1}{\\eta} \\sin \\theta\\right), \\tag {7}", + "image_path": "d2721a0c390e22a27cd1fcc499868e6c08eeeb8f87b48014ac4e9046f425e29d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 434, + 526, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 526, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 526, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 434, + 526, + 445 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 434, + 526, + 445 + ], + "type": "text", + "content": " represents the refractive index of the object." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 460, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 495 + ], + "type": "text", + "content": "Diffuse polarization The change in the polarization state due to diffuse reflection or subsurface scattering [1, 22] is described by" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 354, + 502, + 554, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 502, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 354, + 502, + 554, + 529 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {d} = \\boldsymbol {C} (\\varphi) \\boldsymbol {F} ^ {\\mathrm {T}} (\\theta) \\int_ {\\Omega} \\boldsymbol {D} \\boldsymbol {F} ^ {\\mathrm {T}} (\\theta) \\boldsymbol {s} _ {i, \\omega} d \\omega , \\tag {8}", + "image_path": "58ffb37233d261c6bbd9d114ee4ab0bdc77867c0a92ad9dd146d49ba56b83e9c.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 534, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 570 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 534, + 554, + 570 + ], + "type": "inline_equation", + "content": "s_{i,\\omega}" + }, + { + "bbox": [ + 313, + 534, + 554, + 570 + ], + "type": "text", + "content": " denotes the Stokes vector of incident light coming from the direction " + }, + { + "bbox": [ + 313, + 534, + 554, + 570 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 534, + 554, + 570 + ], + "type": "text", + "content": ". The matrix " + }, + { + "bbox": [ + 313, + 534, + 554, + 570 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 534, + 554, + 570 + ], + "type": "text", + "content": ", which describes the depolarization, is defined as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 387, + 577, + 554, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 577, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 387, + 577, + 554, + 628 + ], + "type": "interline_equation", + "content": "\\boldsymbol {D} = \\left[ \\begin{array}{l l l l} \\rho & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\end{array} \\right], \\tag {9}", + "image_path": "f35f291ee35d872dcb7bb0c978d84e9125e3f76355eae333979c923f82eda0ba.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 634, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 554, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 634, + 554, + 658 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 634, + 554, + 658 + ], + "type": "text", + "content": " denotes the proportion of the depolarization. The Mueller matrix for transmittance " + }, + { + "bbox": [ + 313, + 634, + 554, + 658 + ], + "type": "inline_equation", + "content": "\\pmb{F}^{\\mathrm{T}}(\\theta)" + }, + { + "bbox": [ + 313, + 634, + 554, + 658 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 334, + 666, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 666, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 334, + 666, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} ^ {\\mathrm {T}} (\\theta) = \\left[ \\begin{array}{c c c c} t _ {+} & t _ {-} & 0 & 0 \\\\ t _ {-} & t _ {+} & 0 & 0 \\\\ 0 & 0 & t _ {\\times} \\cos \\delta & t _ {\\times} \\sin \\delta \\\\ 0 & 0 & - t _ {\\times} \\sin \\delta & t _ {\\times} \\cos \\delta \\end{array} \\right], \\tag {10}", + "image_path": "0e28f90c5cdf5a34e3663831972799200f05cb99e6b131d51ae9737285872045.jpg" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27392" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 296, + 109 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 71, + 296, + 109 + ], + "type": "inline_equation", + "content": "t_{\\pm} = \\frac{t_s \\pm t_p}{2}" + }, + { + "bbox": [ + 55, + 71, + 296, + 109 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 71, + 296, + 109 + ], + "type": "inline_equation", + "content": "t_{\\times} = \\sqrt{t_p t_s}" + }, + { + "bbox": [ + 55, + 71, + 296, + 109 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 71, + 296, + 109 + ], + "type": "inline_equation", + "content": "t_p" + }, + { + "bbox": [ + 55, + 71, + 296, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 296, + 109 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 55, + 71, + 296, + 109 + ], + "type": "text", + "content": " denote the Fresnel transmittance coefficients for s- and p- polarization, respectively, and are expressed as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 112, + 304, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 112, + 304, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 112, + 304, + 152 + ], + "type": "interline_equation", + "content": "t _ {s} (\\theta) = \\left(\\frac {2 \\cos \\theta}{\\cos \\theta + \\eta \\cos \\theta^ {\\prime}}\\right) ^ {2}, t _ {p} (\\theta) = \\left(\\frac {2 \\cos \\theta}{\\cos \\theta^ {\\prime} + \\eta \\cos \\theta}\\right) ^ {2}. \\tag {11}", + "image_path": "edae47b480a0f5b28bd3695f810dc5f694a227b546042420aa663ebd7e3f1d83.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 165, + 296, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 165, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 165, + 296, + 201 + ], + "type": "text", + "content": "Transmission polarization The polarization state also changes due to light transmission. According to the Fresnel's law, the transmission polarization is described by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 205, + 296, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 205, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 132, + 205, + 296, + 220 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {t} = \\boldsymbol {C} (\\varphi) \\boldsymbol {F} ^ {\\mathrm {T}} (\\theta) \\boldsymbol {s} _ {i}, \\tag {12}", + "image_path": "25e181c5aaf7dc7664f339ecd4c9a9294e3f3594d78060dd709bff12633ef14e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 225, + 296, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 296, + 248 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 225, + 296, + 248 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 225, + 296, + 248 + ], + "type": "text", + "content": " denotes the Stokes vector of the light coming from under the surface." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 262, + 296, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 287 + ], + "type": "text", + "content": "Emission polarization The polarization state of emitted light is described by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 293, + 295, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 293, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 134, + 293, + 295, + 306 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {e} = \\boldsymbol {C} (\\varphi) \\boldsymbol {E} (\\theta) \\boldsymbol {s} _ {c}, \\tag {13}", + "image_path": "a1cd11839b45ee08bd2a806557d43c8d694f7974218408c6365a1b706d43f783.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 311, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 384 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 311, + 296, + 384 + ], + "type": "inline_equation", + "content": "s_c" + }, + { + "bbox": [ + 55, + 311, + 296, + 384 + ], + "type": "text", + "content": " denotes the Stokes vector of the emitted light including thermal radiation as well as any chemical or electrical emissions and " + }, + { + "bbox": [ + 55, + 311, + 296, + 384 + ], + "type": "inline_equation", + "content": "E(\\theta)" + }, + { + "bbox": [ + 55, + 311, + 296, + 384 + ], + "type": "text", + "content": " is the Mueller matrix corresponding to the emission component. Since there are many types of light emission, the elements of the emission Mueller matrix are not specified explicitly in general." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 389, + 228, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 228, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 228, + 400 + ], + "type": "text", + "content": "3.2. LWIR Polarization Observation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 406, + 296, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 296, + 454 + ], + "type": "text", + "content": "Polarization of visible light Since objects other than light sources do not emit light in the visible spectrum, the emission component is generally ignorable. Therefore, the observed polarization can be modeled as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 137, + 461, + 294, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 461, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 137, + 461, + 294, + 472 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} = \\boldsymbol {s} _ {d} + \\boldsymbol {s} _ {s} + \\boldsymbol {s} _ {t}. \\tag {14}", + "image_path": "0b7bb439df7c48d7cc9e2d2e7b54f3a1ead7a9bfa9b4a8c378bc17a14b41ae5d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 478, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 550 + ], + "type": "text", + "content": "The magnitude of each component varies in complex ways depending on the object's material properties and the illumination environment. This complexity makes polarization analyses more challenging, leading existing visible polarization studies to rely on various assumptions to facilitate the analysis." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 563, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 695 + ], + "type": "text", + "content": "Polarization of LWIR light Most objects are known to be opaque in the LWIR spectrum; therefore, the transmission component can be safely ignored. Moreover, an interesting property of LWIR light polarization is that the relative contributions of each component can be controlled by the object's temperature. In a typical environments, where no objects are extremely hot in the surroundings (i.e., where there are no significant LWIR light sources in the surroundings) the magnitude of reflection components is small compared to that of the emission. In such cases, the LWIR polarization observation can be modeled as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 159, + 702, + 295, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 159, + 702, + 295, + 713 + ], + "type": "interline_equation", + "content": "s = s _ {e}. \\tag {15}", + "image_path": "769b2073bcd9777cd3557d55454bb90db5d4b6ec8583c5dede78292393f6af59.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "Since LWIR emission primarily arises from black body radiation, Eq. (13) can be rewritten as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 384, + 102, + 553, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 102, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 384, + 102, + 553, + 114 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {e} = \\boldsymbol {C} (\\varphi) \\boldsymbol {\\mathcal {E}} (\\theta) \\boldsymbol {s} _ {B} (T), \\tag {16}", + "image_path": "30e0de105ba2086e2c72420e8b4421d0a49aa165e515719d3ff6da64d0d1e222.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 119, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 119, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 119, + 553, + 155 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 119, + 553, + 155 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\theta)" + }, + { + "bbox": [ + 313, + 119, + 553, + 155 + ], + "type": "text", + "content": " denotes the Mueller matrix for black body radiation, and " + }, + { + "bbox": [ + 313, + 119, + 553, + 155 + ], + "type": "inline_equation", + "content": "s_B(T)" + }, + { + "bbox": [ + 313, + 119, + 553, + 155 + ], + "type": "text", + "content": " represents the Stokes vector of the radiation, which can be expressed as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 392, + 159, + 553, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 159, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 392, + 159, + 553, + 209 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {B} (T) = \\left[ \\begin{array}{c} B (T) \\\\ 0 \\\\ 0 \\\\ 0 \\end{array} \\right], \\tag {17}", + "image_path": "2a2c2dea270df19faeccb9590128f03e176899943e4ecc2b1e2dbc3b28d62fc5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 213, + 553, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 237 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 213, + 553, + 237 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 213, + 553, + 237 + ], + "type": "text", + "content": " denotes the black body radiation. According to Planck's law, " + }, + { + "bbox": [ + 313, + 213, + 553, + 237 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 213, + 553, + 237 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 382, + 240, + 553, + 267 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 240, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 382, + 240, + 553, + 267 + ], + "type": "interline_equation", + "content": "B (T) = \\frac {2 h c ^ {2}}{\\lambda^ {5}} \\frac {1}{e ^ {\\frac {h c}{\\lambda k T}} - 1}, \\tag {18}", + "image_path": "dc1bdd9bbc015cf73ba9b4c79f8fafcfa42b4f47874ffe0dea373fb8ae8a2608.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "text", + "content": " is the speed of light, " + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "text", + "content": " is the Boltzmann constant, " + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "text", + "content": " is the Planck constant, " + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "text", + "content": " is the object's absolute temperature, and " + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "text", + "content": " is the observation wavelength. Equation (17) demonstrates that the object's temperature " + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 270, + 553, + 330 + ], + "type": "text", + "content": " controls the power of emission." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 331, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 354 + ], + "type": "text", + "content": "The Mueller matrix of black body radiation " + }, + { + "bbox": [ + 313, + 331, + 553, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(\\theta)" + }, + { + "bbox": [ + 313, + 331, + 553, + 354 + ], + "type": "text", + "content": " [26, 33] is given by" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 378, + 354, + 553, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 354, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 378, + 354, + 553, + 403 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mathcal {E}} (\\theta) = \\left[ \\begin{array}{l l l l} \\varepsilon_ {+} & 0 & 0 & 0 \\\\ \\varepsilon_ {-} & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\\\ 0 & 0 & 0 & 0 \\end{array} \\right], \\tag {19}", + "image_path": "9ab4a4fdbccc4c4e1e767844de635090407cc99df2f5e2a1b0cff0c3fd413408.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 407, + 554, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 478 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 407, + 554, + 478 + ], + "type": "inline_equation", + "content": "\\varepsilon_{\\pm} = \\frac{\\varepsilon_s \\pm \\varepsilon_p}{2}" + }, + { + "bbox": [ + 313, + 407, + 554, + 478 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 407, + 554, + 478 + ], + "type": "inline_equation", + "content": "\\varepsilon_p" + }, + { + "bbox": [ + 313, + 407, + 554, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 407, + 554, + 478 + ], + "type": "inline_equation", + "content": "\\varepsilon_s" + }, + { + "bbox": [ + 313, + 407, + 554, + 478 + ], + "type": "text", + "content": " denote the emissivities for s- and p-polarization, respectively. According to the energy conservation law, also known as Kirchhoff's law, the sum of reflectance, transmittance, and emissivity is equal to 1. Since the transmittance is negligible, as discussed above, we obtain" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 394, + 484, + 553, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 484, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 394, + 484, + 553, + 497 + ], + "type": "interline_equation", + "content": "\\varepsilon_ {s} (\\theta) = 1 - r _ {s} (\\theta) \\tag {20}", + "image_path": "7d00f06dfd94f7a9cb2868fb3767712f8a2dd9cd38ce4b0f5bc114eda689c8e8.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 394, + 499, + 553, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 499, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 394, + 499, + 553, + 512 + ], + "type": "interline_equation", + "content": "\\varepsilon_ {p} (\\theta) = 1 - r _ {p} (\\theta). \\tag {21}", + "image_path": "cd5137a588624b26942d282d51da2729e520a6beaaf0e9d91e5243e3aa278ef0.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 517, + 554, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 565 + ], + "type": "text", + "content": "In summary, in the LWIR spectrum, the Stokes observation " + }, + { + "bbox": [ + 313, + 517, + 554, + 565 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 517, + 554, + 565 + ], + "type": "text", + "content": " can be analytically represented without ambiguity, thereby simplifying the analysis of the polarization state, which is a significant advantage for shape reconstruction." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 578, + 554, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 554, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 554, + 626 + ], + "type": "text", + "content": "Polarization to surface normal: LWIR polarization provides cues for an object's surface normal, which can be determined analytically. The degree of linear polarization (DoLP) is related to the zenith angle " + }, + { + "bbox": [ + 313, + 578, + 554, + 626 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 578, + 554, + 626 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 630, + 558, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 630, + 558, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 630, + 558, + 713 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {D o L P} = \\frac {\\sqrt {s _ {1} ^ {2} + s _ {2} ^ {2}}}{s _ {0}} (22) \\\\ = \\frac {\\left(\\eta - \\frac {1}{\\eta}\\right) ^ {2} \\sin^ {2} \\theta}{2 + 2 \\eta^ {2} - \\left(\\eta + \\frac {1}{\\eta}\\right) \\sin^ {2} \\theta + 4 \\cos \\theta \\sqrt {\\eta^ {2} - \\sin^ {2} \\theta}}. (23) \\\\ \\end{array}", + "image_path": "2d4aa694a4bcd3f35d0116e480da26b52ff2791475db925438fbd47cecf8021f.jpg" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27393" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 294, + 239 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 294, + 239 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 294, + 239 + ], + "type": "image", + "image_path": "0a83ce182410d44b86e138c9b93a2f722fbf8ac0ba166d6d29ef2fd27e0d956e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 296, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 304 + ], + "type": "text", + "content": "Figure 2. Visible and thermal polarization observations of spheres made from various materials under natural room lighting. While visible polarization is unstable due to variations in a material's optical properties, thermal polarimetric cues, especially AoLP images, remain consistent across different materials." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 323, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 348 + ], + "type": "text", + "content": "The angle of linear polarization (AoLP) is equivalent to the azimuth angle " + }, + { + "bbox": [ + 55, + 323, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 323, + 296, + 348 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 356, + 295, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 356, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 112, + 356, + 295, + 383 + ], + "type": "interline_equation", + "content": "\\mathrm {A o L P} = \\frac {1}{2} \\tan^ {- 1} \\left(\\frac {s _ {2}}{s _ {1}}\\right) = \\varphi . \\tag {24}", + "image_path": "0a5d75dcaaf7fe12ed6a789641b8df33fe0f16ff615f19fe49fda23fe49c9fd0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 389, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 462 + ], + "type": "text", + "content": "These correspondences enable the reconstruction of the surface normal from polarization observations. Especially, AoLP serves as a robust cue because it directly reflects the azimuth angle of the surface normal and is independent of material properties, whereas DoLP is material-dependent due to the inclusion of the refractive index term." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 279, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 279, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 279, + 481 + ], + "type": "text", + "content": "3.3. Examples of Visible vs. LWIR Polarization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 653 + ], + "type": "text", + "content": "In the visible light spectrum, specular, diffuse, and transmission polarization components are often mixed [2, 25, 46], resulting in complicated observation. Fig. 2 shows the polarimetric observations of spheres made from various materials; in the visible spectrum, unwanted polarization observations frequently appear. For example, because black stone exhibits strong specular polarization, the DoLP image is heavily affected by reflections from the surrounding environment. In the case of transparent glass sphere, the observation is notably contaminated by background textures. For translucent plastic materials, both DoLP and AoLP images are affected by surface texture and subsurface scattering. These unstable observations significantly complicate subsequent shape recovery." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "In contrast, in the LWIR spectrum, polarimetric observations remain consistent and independent of both material properties and environmental reflections, particularly in AoLP images. This consistency is a key advantage of thermal polarization cues, as it provides robustness against vari" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 504, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 504, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 504, + 84 + ], + "type": "text", + "content": "ations in materials and surrounding reflections." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 95, + 536, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 95, + 536, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 95, + 536, + 108 + ], + "type": "text", + "content": "4. Thermal Polarimetric Multi-view Stereo" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 116, + 554, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 116, + 554, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 554, + 175 + ], + "type": "text", + "content": "Our goal is to recover the 3D shape using thermal polarimetric cues. As discussed in the previous section, the AoLP observation is invariant to material properties. To leverage this strong cue, we only use the AoLP images for material-independent 3D shape reconstruction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 176, + 555, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 273 + ], + "type": "text", + "content": "Our approach employs multi-view thermal AoLP images to estimate the Signed Distance Function (SDF), which is implicitly represented by a Multi-Layer Perceptron (MLP) optimized via differentiable rendering. To effectively incorporate the surface normal information present in AoLP images, we adopt the concept of tangent space constraint (TSC) proposed in a multi-view azimuth stereo (MVAS) [5]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 281, + 417, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 417, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 417, + 293 + ], + "type": "text", + "content": "Shape representation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 297, + 554, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 554, + 357 + ], + "type": "text", + "content": "The SDF is implicitly represented by an MLP " + }, + { + "bbox": [ + 313, + 297, + 554, + 357 + ], + "type": "inline_equation", + "content": "f(\\pmb{x}; \\pmb{\\theta}) : \\mathbb{R}^3 \\times \\mathbb{R}^d \\to \\mathbb{R}" + }, + { + "bbox": [ + 313, + 297, + 554, + 357 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 297, + 554, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 297, + 554, + 357 + ], + "type": "text", + "content": " denotes a point in space and " + }, + { + "bbox": [ + 313, + 297, + 554, + 357 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 297, + 554, + 357 + ], + "type": "text", + "content": " are the learnable parameters of the MLP. The zero-level set of the SDF defines the object's surface " + }, + { + "bbox": [ + 313, + 297, + 554, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 297, + 554, + 357 + ], + "type": "text", + "content": " as [37]" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 377, + 358, + 553, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 358, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 377, + 358, + 553, + 371 + ], + "type": "interline_equation", + "content": "\\mathcal {M} (\\boldsymbol {\\theta}) = \\{\\boldsymbol {x} | f (\\boldsymbol {x}; \\boldsymbol {\\theta}) = 0 \\}. \\tag {25}", + "image_path": "78e725862f6ecaa2c86da33386189446a5c47f5f4fa84c5c769626517cff695a.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 377, + 554, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 401 + ], + "type": "text", + "content": "The MLP is optimized within the framework of implicit differentiable renderer (IDR) [51]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 410, + 381, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 381, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 381, + 421 + ], + "type": "text", + "content": "Loss function" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 427, + 554, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 554, + 450 + ], + "type": "text", + "content": "We adopt the loss function proposed in MVAS [5], which is defined as," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 350, + 461, + 553, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 461, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 350, + 461, + 553, + 474 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {T S C}} + \\lambda_ {1} \\mathcal {L} _ {\\text {s i l h o u e t t e}} + \\lambda_ {2} \\mathcal {L} _ {\\text {E i k o n a l}}, \\tag {26}", + "image_path": "fbfcd1280a459ce240ded55f11d7ba4556404a3b151a2bd6a6d01dac090b472a.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{TSC}}" + }, + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "type": "text", + "content": " is the tangent space consistency loss, " + }, + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{silhouette}}" + }, + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "type": "text", + "content": " is the silhouette loss and " + }, + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Eikonal}}" + }, + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "type": "text", + "content": " is the Eikonal regularization. " + }, + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 484, + 554, + 532 + ], + "type": "text", + "content": " are the weights assigned to the respective loss terms." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 548, + 554, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 620 + ], + "type": "text", + "content": "Tangent space consistency loss [5] The AoLP captured from different camera positions enforces a strong constraint on surface normal estimation. For a unit normal vector " + }, + { + "bbox": [ + 313, + 548, + 554, + 620 + ], + "type": "inline_equation", + "content": "\\pmb{n} \\subset S^2 \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 548, + 554, + 620 + ], + "type": "text", + "content": " at a surface point " + }, + { + "bbox": [ + 313, + 548, + 554, + 620 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 548, + 554, + 620 + ], + "type": "text", + "content": " and the projected azimuth angle " + }, + { + "bbox": [ + 313, + 548, + 554, + 620 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 548, + 554, + 620 + ], + "type": "text", + "content": " at the corresponding pixel, the following relationship holds:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 375, + 628, + 553, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 628, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 375, + 628, + 553, + 643 + ], + "type": "interline_equation", + "content": "\\boldsymbol {r} _ {1} ^ {\\top} \\boldsymbol {n} \\cos \\varphi - \\boldsymbol {r} _ {2} ^ {\\top} \\boldsymbol {n} \\sin \\varphi = 0, \\tag {27}", + "image_path": "85ea6956f5edb4de045a4b2764f0efa30e1f764f107d37d5089585ccd1fdfce7.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 652, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 553, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 652, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{R} = [r_1, r_2, r_3]^\\top" + }, + { + "bbox": [ + 313, + 652, + 553, + 677 + ], + "type": "text", + "content": " denotes the rotation matrix of the camera pose. This equation can be rearranged as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 372, + 685, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 685, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 372, + 685, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {n} ^ {\\top} \\underbrace {\\left(\\boldsymbol {r} _ {1} \\cos \\varphi - \\boldsymbol {r} _ {2} \\sin \\varphi\\right)} _ {\\boldsymbol {t} (\\varphi)} = 0, \\tag {28}", + "image_path": "14ab8177fc73ad1e8cc781f390f14a29b8f808f54a1fa6fe7730972eb5e991ba.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27394" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "t(\\varphi)" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " is called a projected tangent vector and the above equation is called tangent space constraint. The TSC loss is the sum of the square error of the tangent space constraint and is defined as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 122, + 295, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 122, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 74, + 122, + 295, + 156 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T S C}} = \\frac {1}{P} \\sum_ {\\boldsymbol {x} \\in \\boldsymbol {X}} \\frac {\\sum_ {i = 1} ^ {N} \\Phi_ {i} \\left(\\boldsymbol {n} ^ {\\top} (\\boldsymbol {x} ; \\boldsymbol {\\theta}) \\boldsymbol {t} _ {i} (\\boldsymbol {x})\\right) ^ {2}}{\\sum_ {i = 1} ^ {N} \\Phi_ {i}} \\tag {29}", + "image_path": "37aeeeee191e4f891b010516b8151a081e85c8c034855236175e1f2543e1a911.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": " denotes the number of pixel samples, " + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\Phi_{i}" + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": " is a binary indicator of the visibility of point " + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": "-th view, " + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\pmb{n}^{\\top}(\\pmb{x};\\pmb{\\theta})" + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": " is the estimated surface normal at the point " + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": " given network parameters " + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\pmb{\\theta}" + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{i}(\\pmb{x})" + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": " is the tangent vector corresponding to the " + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": "-th view's pixel onto which the point " + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 159, + 296, + 233 + ], + "type": "text", + "content": " is projected." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 245, + 296, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 269 + ], + "type": "text", + "content": "Silhouette loss [5, 51] The silhouette loss constrains the visual hull of the shape and is defined as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 273, + 295, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 273, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 71, + 273, + 295, + 304 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i l h o u e t t e}} = \\frac {1}{\\alpha P} \\sum_ {\\boldsymbol {x} \\in \\hat {\\boldsymbol {X}}} \\Psi (O (\\Pi (\\boldsymbol {x})), \\sigma (\\alpha f ^ {*})), \\tag {30}", + "image_path": "b7fd5eaa4bddc4f78e07096f6a4f2231108afbffff72027c041e522a5f5f6424.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "text", + "content": " denotes the cross entropy and " + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "text", + "content": " is a binary mask indicating whether a pixel lies inside the object silhouette (as provided by the input). " + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "text", + "content": " is the number of pixel samples, " + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "text", + "content": " is the projection from the world coordinate to the image coordinate, " + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{X}}" + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "text", + "content": " represents the pixels with no ray-surface interactions or outside the silhouette, " + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "inline_equation", + "content": "f^{*}" + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "text", + "content": " represents the minimum SDF distance between the ray that does not have intersections and the object surface, " + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "text", + "content": " is a sigmoid function, and " + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 308, + 296, + 427 + ], + "type": "text", + "content": " controls the sharpness of the sigmoid function." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 441, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 477 + ], + "type": "text", + "content": "Eikonal regularization [17] The Eikonal loss regularizes the gradient of SDF, enforcing the SDF to maintain its property of having a gradient close to 1:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 97, + 481, + 295, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 481, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 97, + 481, + 295, + 497 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {E i k o n a l}} = \\mathbb {E} _ {\\boldsymbol {x}} \\left(\\left(\\left\\| \\boldsymbol {n} (\\boldsymbol {x}; \\boldsymbol {\\theta}) \\right\\| _ {2} - 1\\right) ^ {2}\\right). \\tag {31}", + "image_path": "84c58e873ef31fd574b88e831c987cb67832542b6914f096d3aa4fe48c421dd7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 500, + 295, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 295, + 536 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_x(\\cdot)" + }, + { + "bbox": [ + 55, + 500, + 295, + 536 + ], + "type": "text", + "content": " denotes the expected value computed over sample points " + }, + { + "bbox": [ + 55, + 500, + 295, + 536 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 500, + 295, + 536 + ], + "type": "text", + "content": " that are randomly sampled within the bounding box." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 544, + 137, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 137, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 137, + 558 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 564, + 105, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 105, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 105, + 577 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Implementation Following MVAS [5], we use an 8-layer MLP with softmax activations, combined with positional encoding (dimension 10). The input 3D position and encoded features are skip-connected to the 4th layer of the network. The optimization is performed using the Adam optimizer with a learning rate of 0.001. The weights for the loss terms are set to " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 50" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " (silhouette loss) and " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 0.1" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " (Eikonal loss). Training is conducted for 50 epochs with a batch size of 4096 pixels, and the silhouette loss weight is halved every 10 epochs. Training takes approximately 2 hours on a single RTX A6000 GPU." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 554, + 237 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 554, + 237 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 554, + 237 + ], + "type": "image", + "image_path": "3641a6f9bb07ad68ad30c19c074f61a0cd205f1688ea98c3f7372abdb2989620.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 245, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 291 + ], + "type": "text", + "content": "Figure 3. Camera system and setup. The system consists of a rotation stage to rotate the target and a thermal camera with a wire-grid polarizer. A visible polarization camera is used for camera pose estimation and for comparison with the visible MVAS." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 311, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 491 + ], + "type": "text", + "content": "Camera system Figure 3 shows our camera system and experimental setup. Our thermal polarization camera system consists of a thermal camera (FLIR Boson 320) with " + }, + { + "bbox": [ + 313, + 311, + 555, + 491 + ], + "type": "inline_equation", + "content": "15\\mathrm{mm}" + }, + { + "bbox": [ + 313, + 311, + 555, + 491 + ], + "type": "text", + "content": " lens and a wire-grid polarizer (Thorlabs WP25M-IRC) mounted on a motorized rotation mount (Thorlabs K10CR2). The polarizer is installed between the image sensor and imaging lens to mitigate the narcissus reflections. Thermal polarization images are captured by rotating the polarizer using the rotation mount. We captured four polarization images at 0, 45, 90, and 135 degrees. A monochrome visible-light polarization camera (FLIR Blackfly S BFS-U3-51S5P) is also placed adjacent to the thermal camera. The visible-light polarization camera is used both to capture images for a baseline method and to get the poses of both visible and thermal cameras." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 506, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 662 + ], + "type": "text", + "content": "Camera poses We employed a marker-based approach to estimate the camera poses. We put Aruco markers [15] on the rotation stage to estimate the pose of visible-light camera. Since printed markers are not visible in the thermal images, the poses of the thermal camera were obtained by transforming the camera poses of the visible-light camera. For stereo calibration between the visible-light and thermal cameras, we used a calibration target that can be observed by both cameras. The calibration board consists of a white-painted aluminum pegboard placed in front of a black surface heater. By heating the back panel, the circle grid pattern became visible to both cameras. The calibration parameters were estimated using Zhang's method [54]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Dataset We prepared a real-world dataset with ground-truth shapes. We captured 7 objects with various shapes and materials. The target objects include transparent and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "27395" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 555, + 403 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 555, + 403 + ], + "type": "image", + "image_path": "3dc83f94bbd3d1bc082ad55d0fdc720bf6979f8c241ef121a8acf3b9bdb3ded5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 410, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 55, + 410, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 555, + 455 + ], + "type": "text", + "content": "Figure 4. Experimental results. AoLP images in both visible and LWIR spectra are shown below the scene photograph. Each method's result includes the recovered shape, surface normals, and angular error maps. The Chamfer distance is displayed below the estimated shape and the mean angular error is displayed to the right of the angular error map. For both metrics, lower values indicate better performance. The smallest errors are shown in magenta. It demonstrates that the proposed method outperforms the other methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 475, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 620 + ], + "type": "text", + "content": "low-reflective objects, which pose challenges for existing reconstruction methods. The objects were placed on a rotating table and captured from about 20–30 views by rotating the table. To robustly capture thermal polarization, the objects were kept warm during the measurements. Masks for thermal images were generated by thresholding pixel values, while masks for visible images were generated by using Segment Anything Model 2 [39]. Ground-truth shapes were obtained using a structured light 3D scanner (EinScan-SP V2). To facilitate the ground-truth measurement, the surfaces of the objects were painted white to make them opaque." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "Baselines We compare our proposed method with two state-of-the-art methods: MVAS [5] based on visible-light polarization cues and IDR [51], which is one of the neural implicit surface reconstruction methods, with thermal image input. For MVAS with visible-light polarization images, there exists " + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pm \\frac{\\pi}{2}" + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": " ambiguity in surface normals due to" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 475, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 536 + ], + "type": "text", + "content": "the mixture of specular and diffuse polarization. Therefore, TSC loss is modified to account for the " + }, + { + "bbox": [ + 313, + 475, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\pm \\frac{\\pi}{2}" + }, + { + "bbox": [ + 313, + 475, + 555, + 536 + ], + "type": "text", + "content": " ambiguity, as originally proposed in MVAS. Since our method does not exhibit this ambiguity, the modified loss function is applied only to visible light observation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 662 + ], + "type": "text", + "content": "Evaluation metrics We use two metrics to evaluate both the surface points and surface normals. Chamfer distance is the metric for evaluating the geometric error, which calculates the point cloud distance between the estimate and the ground-truth. The mean angular error is the metric for surface normal, which calculates the angular error of normal maps. For both metrics, a smaller value indicates a better result. Prior to computing these metrics, the estimated and the ground-truth meshes are aligned using ICP algorithm [4]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 671, + 372, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 372, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 372, + 682 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Figure 4 shows the results. The target objects are an owl made of ceramic with a clear coating, a glass container with" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27396" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 298, + 401 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 298, + 401 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 298, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 298, + 401 + ], + "type": "image", + "image_path": "999293604978cbac00a51ab636cb2162f1933fb058b2908e72623c1e9200977d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 409, + 297, + 488 + ], + "lines": [ + { + "bbox": [ + 55, + 409, + 297, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 297, + 488 + ], + "type": "text", + "content": "Figure 5. Qualitative results for other objects. AoLP images are shown below the scene photos. For each method, the upper image is the estimated shape and the lower is the corresponding normal map. It is shown that our method successfully reconstructs fine details of the object, particularly the concave parts in the transparent vase and bottle, while other methods fail to reconstruct such details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 616 + ], + "type": "text", + "content": "and without interiors, and a black tumbler, some of which are challenging objects for the existing approaches, especially a container with cutlery. In AoLP observations, while visible light is affected by a mixture of reflections and transmission components, thermal observations remain consistent and align with the azimuth angle of the surface normal. This consistency leads to a clear difference in detailed 3D reconstruction results between visible polarization and thermal polarization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "In the owl's result, while the detailed shape such as the eyes and the beak is missing in both visible MVAS and thermal IDR, they are successfully recovered in our result. Similarly, the relief pattern on the glass container is clearly visible in our results. For a black tumbler with a white lid shown in the fourth row, the AoLP observation on the lid in the visible-light spectrum is too noisy. In such a case, the surface normal of the lid is not reconstructible, resulting in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "a wavy surface. In contrast, our method recovers a reasonable shape due to the presence of a unique AoLP on the lid in LWIR polarization. Overall, visible MVAS tends to reconstruct with wavy artifacts as AoLP is noisy, and thermal IDR tends to recover swollen shapes as it does not consider the surface normal." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 148, + 555, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 209 + ], + "type": "text", + "content": "To quantitatively compare the results, the estimated shapes are evaluated using Chamfer distance, and the surface normals are evaluated by mean angular error. Our method outperforms the other methods in both Chamfer distance and the mean angular error for all objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 214, + 556, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 556, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 556, + 346 + ], + "type": "text", + "content": "Figure 5 shows additional qualitative results. Our results are visually better than the other methods. The concave details of the vase and bottle are especially notable, as they are precisely reconstructed in our method while other results appear blurred. In addition, the surface of the container's lid is only reconstructed by our method while there is a big hole in the results of the others. Some artifacts are seen in thin parts in our result, particularly in the neck of the bottle. This artifact comes from the low spatial resolution of the thermal camera, which can be reduced by using a higher-resolution camera or increasing camera views." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 372, + 388, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 372, + 388, + 384 + ], + "spans": [ + { + "bbox": [ + 314, + 372, + 388, + 384 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 397, + 555, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 505 + ], + "type": "text", + "content": "In this paper, we propose a 3D shape reconstruction using multi-view thermal polarimetric observations. We theoretically show that thermal polarimetric cues, especially AoLP images, are independent of material properties and illumination environment. Based on this property, we demonstrate a detailed 3D shape recovery using multi-view thermal AoLP images and show that our approach is applicable to heterogeneous objects including transparent and low-reflective objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Although the proposed method performs well, some challenges remain in expanding the applicable scenes. Our approach assumes the sufficient power of the emission component compared to the reflection component. This assumption is not valid when the hot objects are placed near the target object or when the target's temperature is lower than the environment. Moreover, our method is less effective on metallic surfaces or rough surfaces, where the polarization signal can become unstable or noisy. In such cases, decomposing the emission and reflection components [48] before applying the shape recovery or integrating additional cues may be a good option. Improving the observation system is also an important factor since our capturing system needs some time to rotate the polarization filter. A single-shot polarized LWIR camera could accelerate the thermal polarization analysis in the near future, as in the visible-light spectrum." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27397" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 294, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 294, + 114 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 294, + 114 + ], + "type": "text", + "content": "This work was supported by JSPS KAKENHI Grant Numbers JP24K20802, JP22K18420, JP24K02965." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 125, + 114, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 125, + 114, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 114, + 137 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 146, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 62, + 146, + 295, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 146, + 295, + 177 + ], + "spans": [ + { + "bbox": [ + 62, + 146, + 295, + 177 + ], + "type": "text", + "content": "[1] G.A. Atkinson and E.R. Hancock. Recovery of surface orientation from diffuse polarization. IEEE TIP, 15(6):1653-1664, 2006." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 178, + 295, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 178, + 295, + 221 + ], + "spans": [ + { + "bbox": [ + 62, + 178, + 295, + 221 + ], + "type": "text", + "content": "[2] Yunhao Ba, Alex Gilbert, Franklin Wang, Jina Yang, Rui Chen, Yiqin Wang, Lei Yan, Boxin Shi, and Achuta Kadambi. Deep Shape from Polarization. In ECCV, pages 554-571, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 224, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 255 + ], + "type": "text", + "content": "[3] Seung-Hwan Baek, Daniel S. Jeon, Xin Tong, and Min H. Kim. Simultaneous acquisition of polarimetric SVBRDF and normals. ACM TOG, 37(6):1-15, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 258, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 278 + ], + "type": "text", + "content": "[4] P.J. Besl and Neil D. McKay. A method for registration of 3-D shapes. IEEE TPAMI, 14(2):239-256, 1992." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 280, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 294, + 312 + ], + "type": "text", + "content": "[5] Xu Cao, Hiroaki Santo, Fumio Okura, and Yasuyuki Matsushita. Multi-View Azimuth Stereo via Tangent Space Consistency. In CVPR, pages 825–834, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 314, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 294, + 345 + ], + "type": "text", + "content": "[6] Qian Chen, Shihao Shu, and Xiangzhi Bai. Thermal3D-GS: Physics-Induced 3D Gaussians for Thermal Infrared Novel-View Synthesis. In ECCV, pages 253–269, 2025." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 347, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 401 + ], + "type": "text", + "content": "[7] Charly Collin, Sumanta Pattanaik, Patrick LiKamWa, and Kadi Bouatouch. Computation of polarized subsurface BRDF for rendering. In Proceedings of Graphics Interface 2014, pages 201-208. Canadian Information Processing Society, 2014." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 403, + 294, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 434 + ], + "type": "text", + "content": "[8] Zhaopeng Cui, Jinwei Gu, Boxin Shi, Ping Tan, and Jan Kautz. Polarimetric Multi-view Stereo. In CVPR, pages 369-378, 2017." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 479 + ], + "type": "text", + "content": "[9] Unay Dorken Gallastegi, Hoover Rueda-Chacón, Martin J. Stevens, and Vivek K Goyal. Absorption-based hyperspectral thermal ranging: Performance analyses, optimization, and simulations. Optics Express, 32(1):151, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 480, + 294, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 294, + 534 + ], + "type": "text", + "content": "[10] Jean-Denis Durou, Maurizio Falcone, Yvain Quéau, and Silvia Tozza. A Comprehensive Introduction to Photometric 3D-Reconstruction. In Advances in Photometric 3D-Reconstruction, pages 1-29. Springer International Publishing, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 536, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 579 + ], + "type": "text", + "content": "[11] Jack Erdozain, Kazuto Ichimaru, Tomohiro Maeda, Hiroshi Kawasaki, Ramesh Raskar, and Achuta Kadambi. 3d Imaging For Thermal Cameras Using Structured Light. In ICIP, pages 2795-2799, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 581, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 294, + 635 + ], + "type": "text", + "content": "[12] Gonen Eren, Olivier Aubreton, Fabrice Meriaudeau, L.A. Sanchez Secades, David Fofi, A. Teoman Naskali, Frederic Truchetet, and Aytul Ercil. Scanning from heating: 3D shape estimation of transparent objects from local surface heating. Optics Express, 17(14):11457, 2009." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 636, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 667 + ], + "type": "text", + "content": "[13] Yoshiki Fukao, Ryo Kawahara, Shohei Nobuhara, and Ko Nishino. Polarimetric Normal Stereo. In CVPR, pages 682-690, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[14] Unay Dorken Gallastegi, Hoover Rueda-Chacon, Martin J. Stevens, and Vivek K Goyal. Absorption-Based, Passive Range Imaging from Hyperspectral Thermal Measurements. IEEE TPAMI, pages 1-18, 2025." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] S. Garrido-Jurado, R. Muñoz-Salinas, F.J. Madrid-Cuevas, and M.J. Marín-Jiménez. Automatic generation and detection of highly reliable fiducial markers under occlusion. PR, 47(6):2280-2292, 2014." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 118, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 138 + ], + "type": "text", + "content": "[16] Jason Geng. Structured-light 3D surface imaging: A tutorial. Advances in Optics and Photonics, 3(2):128, 2011." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "text", + "content": "[17] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit Geometric Regularization for Learning Shapes, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "text", + "content": "[18] Yufei Han, Heng Guo, Koki Fukai, Hiroaki Santo, Boxin Shi, Fumio Okura, Zhanyu Ma, and Yunpeng Jia. NeRSP: Neural 3D Reconstruction for Reflective Objects with Sparse Polarized Images. In CVPR, pages 11821-11830, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 217, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 248 + ], + "type": "text", + "content": "[19] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2nd ed edition, 2003." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 251, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 281 + ], + "type": "text", + "content": "[20] A. Hertzmann and S.M. Seitz. Example-based photometric stereo: Shape reconstruction with general, varying BRDFs. IEEE TPAMI, 27(8):1254-1264, 2005." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 284, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 336 + ], + "type": "text", + "content": "[21] Inseung Hwang, Daniel S. Jeon, Adolfo Muñoz, Diego Gutierrez, Xin Tong, and Min H. Kim. Sparse ellipsometry: Portable acquisition of polarimetric SVBRDF and shape with unstructured flash photography. ACM TOG, 41(4):1-14, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[22] Tomoki Ichikawa, Matthew Purri, Ryo Kawahara, Shohei Nobuhara, Kristin Dana, and Ko Nishino. Shape from Sky: Polarimetric Normal Recovery Under The Sky. In CVPR, pages 14827-14836, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 383, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 436 + ], + "type": "text", + "content": "[23] Masaki Kaga, Takahiro Kushida, Tsuyoshi Takatani, Kenichiro Tanaka, Takuya Funatomi, and Yasuhiro Mukaigawa. Thermal non-line-of-sight imaging from specular and diffuse reflections. IPSJ Transactions on Computer Vision and Applications, 11(1):8, 2019." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "text", + "content": "[24] Takahiro Kushida, Ryutaro Nakamura, Hiroaki Matsuda, Wenhao Chen, and Kenichiro Tanaka. Affine transform representation for reducing calibration cost on absorption-based LWIR depth sensing. Scientific Reports, 14(1):26429, 2024." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 482, + 553, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 523 + ], + "type": "text", + "content": "[25] Chenyang Lei, Chenyang Qi, Jiaxin Xie, Na Fan, Vladlen Koltun, and Qifeng Chen. Shape from Polarization for Complex Scenes in the Wild. In CVPR, pages 12622-12631, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 526, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 558 + ], + "type": "text", + "content": "[26] Daniel A. LeMaster and Michael T. Eismann. Passive Polarimetric Imaging. In Multi-Dimensional Imaging, pages 391-428. Wiley, 1 edition, 2014." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "text", + "content": "[27] Chenhao Li, Taishi Ono, Takeshi Uemori, Hajime Mihara, Alexander Gatto, Hajime Nagahara, and Yusuke Moriuchi. NeISF: Neural Incident Stokes Field for Geometry and Material Estimation. In CVPR, pages 21434-21445, 2024." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "type": "text", + "content": "[28] Yvette Y. Lin, Xin-Yi Pan, Sara Fridovich-Keil, and Gordon Wetzstein. ThermalNeRF: Thermal Radiance Fields. In ICCP, pages 1–12, 2024." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 636, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 668 + ], + "type": "text", + "content": "[29] Tomohiro Maeda, Yiqin Wang, Ramesh Raskar, and Achuta Kadambi. Thermal Non-Line-of-Sight Imaging. In ICCP, pages 1-11, 2019." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[30] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27398" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "text", + "content": "[31] Seiichi Mita, Xu Yuquan, Kazuhisa Ishimaru, and Sakiko Nishino. Robust 3D Perception for any Environment and any Weather Condition using Thermal Stereo. In IEEE Intelligent Vehicles Symposium (IV), pages 2569-2574, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[32] Miyazaki, Tan, Hara, and Ikeuchi. Polarization-based inverse rendering from a single view. In ICCV, pages 982-987 vol.2, 2003." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[33] Daisuke Miyazaki, Megumi Saito, Yoichi Sato, and Katsushi Ikeuchi. Determining surface orientations of transparent objects based on polarization degrees in visible and infrared wavelengths. J. Opt. Soc. Am. A, 19(4):687, 2002." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "text", + "content": "[34] D. Miyazaki, M. Kagesawa, and K. Ikeuchi. Transparent surface modeling from a pair of polarization images. IEEE TPAMI, 26(1):73-82, 2004." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 275 + ], + "type": "text", + "content": "[35] Yasuto Nagase, Takahiro Kushida, Kenichiro Tanaka, Takuya Funatomi, and Yasuhiro Mukaigawa. Shape from Thermal Radiation: Passive Ranging Using Multi-spectral LWIR Measurements. In CVPR, pages 12651-12661, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[36] Sriram Narayanan, Mani Ramanagopal, Mark Sheinin, Aswin C. Sankaranarayanan, and Srinivasa G. Narasimhan. Shape from heat conduction. In ECCV, pages 426-444, 2025." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 365 + ], + "type": "text", + "content": "[37] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation. In CVPR, pages 165-174, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 410 + ], + "type": "text", + "content": "[38] Mani Ramanagopal, Sriram Narayanan, Aswin C. Sankaranarayanan, and Srinivasa G. Narasimhan. A Theory of Joint Light and Heat Transport for Lambertian Scenes. In CVPR, pages 11924-11933, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 487 + ], + "type": "text", + "content": "[39] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenhofer. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 488, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 533 + ], + "type": "text", + "content": "[40] Joaquin Rodriguez, Lew-Fock-Chong Lew-Yan-Voon, Renato Martins, and Olivier Morel. Pola4All: Survey of polarimetric applications and an open-source toolkit to analyze polarization. Journal of Electronic Imaging, 33(01), 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 533, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 577 + ], + "type": "text", + "content": "[41] Hiroaki Santo, Masaki Samejima, Yusuke Sugano, Boxin Shi, and Yasuyuki Matsushita. Deep Photometric Stereo Networks for Determining Surface Normal and Reflectances. IEEE TPAMI, 44(1):114-128, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 578, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 611 + ], + "type": "text", + "content": "[42] Philip Saponaro, Scott Sorensen, Abhishek Kolagunda, and Chandra Kambhametty. Material classification with thermal imagery. In CVPR, pages 4649-4656, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 294, + 644 + ], + "type": "text", + "content": "[43] D. Scharstein and R. Szeliski. High-accuracy stereo depth maps using structured light. In CVPR, pages I-195-I-202, 2003." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 646, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 669 + ], + "type": "text", + "content": "[44] Johannes L. Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In CVPR, pages 4104-4113, 2016." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[45] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise View Selection for Unstructured Multi-View Stereo. In ECCV, pages 501–518, 2016." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 444 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[46] Boxin Shi, Jina Yang, Jinwei Chen, Ruihua Zhang, and Rui Chen. Recent Progress in Shape from Polarization. In Advances in Photometric 3D-Reconstruction, pages 177-203. Springer International Publishing, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 140 + ], + "type": "text", + "content": "[47] Richard Szeliski. Computer Vision: Algorithms and Applications. Springer International Publishing, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[48] Kenichiro Tanaka, Nobuhiro Ikeya, Tsuyoshi Takatani, Hiroyuki Kubo, Takuya Funatomi, Vijay Ravi, Achuta Kadambi, and Yasuhiro Mukaigawa. Time-Resolved Far Infrared Light Transport Decomposition for Thermal Photometric Stereo. IEEE TPAMI, 43(6):2075–2085, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 228 + ], + "type": "text", + "content": "[49] L.B. Wolff and T.E. Boult. Constraining object features using a polarization reflectance model. IEEE TPAMI, 13(7): 635-657, 1991." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 230, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 263 + ], + "type": "text", + "content": "[50] Robert J. Woodham. Photometric Method For Determining Surface Orientation From Multiple Images. Optical Engineering, 19(1), 1980." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[51] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. NeurIPS, 33, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 353 + ], + "type": "text", + "content": "[52] Tianxiang Ye, Qi Wu, Junyuan Deng, Guoqing Liu, Liu Liu, Songpengcheng Xia, Liang Pang, Wenxian Yu, and Ling Pei. Thermal-NeRF: Neural Radiance Fields from an Infrared Camera. In IROS, pages 1046–1053, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[53] Jincheng Zhang, Andrew R Willis, and Kevin Brink. Photometric Correction for Infrared Sensors. In CVPRW, pages 431-439, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 410 + ], + "type": "text", + "content": "[54] Z. Zhang. A flexible new technique for camera calibration. IEEE TPAMI, 22(11):1330-1334, 2000." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 444 + ], + "type": "text", + "content": "[55] Jinyu Zhao, Yusuke Monno, and Masatoshi Okutomi. Polarimetric Multi-View Inverse Rendering. IEEE TPAMI, 45 (7):8798-8812, 2023." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27399" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/87c172f1-4134-4138-9793-b4455c2d3c37_content_list.json b/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/87c172f1-4134-4138-9793-b4455c2d3c37_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4d8f3a0fd6d7fb694130d22241d768dc64075ae8 --- /dev/null +++ b/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/87c172f1-4134-4138-9793-b4455c2d3c37_content_list.json @@ -0,0 +1,1481 @@ +[ + { + "type": "text", + "text": "Think Twice: Test-Time Reasoning for Robust CLIP Zero-Shot Classification", + "text_level": 1, + "bbox": [ + 109, + 128, + 887, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shenyu Lu Zhaoying Pan Xiaoqian Wang* Elmore Family School of Electrical and Computer Engineering Purdue University", + "bbox": [ + 248, + 180, + 748, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{lu876, pan433, joywang}@purdue.edu", + "bbox": [ + 341, + 236, + 651, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive Language-Image Pre-training (CLIP) models exhibit intriguing properties, particularly in their zero-shot classification capability. However, the reliability of CLIP zero-shot classification is severely undermined by spurious correlations. Existing efforts to enhance the robustness of zero-shot CLIP models often rely on prior knowledge or annotations of spurious correlations, limiting real-world applicability due to the unavailability of such information. Alternative methods attempt to detect distribution shift at test time but require training statistics whose access is often restricted or computationally expensive. To address the challenges brought by spurious correlation under zero-shot settings, we propose a novel test-time reasoning approach. Our method, inspired by human recognition, localizes the object and refines the classification accordingly. The inherent capacity of CLIP for semantic understanding allows us to isolate the object of interest without auxiliary models. Zero-shot classification is then performed exclusively on the localized objects, effectively mitigating the influence of spurious correlation. The proposed approach is interpretable and flexible as it requires no spurious annotations or prior knowledge, making it widely applicable. The substantial improvements across multiple benchmark datasets validated the effectiveness of our approach.", + "bbox": [ + 89, + 318, + 485, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 712, + 220, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive Language-Image Pre-training (CLIP) [38] has achieved significant success as a powerful pre-training paradigm. It leverages a unified multimodal framework that integrates images and text to achieve strong generalizability. The zero-shot capability of CLIP allows it to classify unseen test samples, making it highly applicable in real-world scenarios. Despite its impressive generalizability, CLIP remains susceptible to spurious correlations, undermining the reliability of its zero-shot predictions. A spurious correla", + "bbox": [ + 89, + 738, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f27f81a57019aa5dd09b24b83ef4991dbfc47fcef1b022a75c01a057f21f8e5d.jpg", + "image_caption": [ + "Figure 1. Attention map visualization based on [CLS] for randomly selected samples. Left: Vanilla zero-shot prediction with different vision transformers. Right: Proposed method applied to the same vision transformers." + ], + "image_footnote": [], + "bbox": [ + 535, + 284, + 740, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d58c89fcb03e3876238a1c489e2a83b1ab9ec1920ef6490baac584ef12c820fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 285, + 883, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion occurs when a feature is correlated with the target label but lacks a causal relationship [19]. This issue is particularly critical in high-stakes settings where reliance on such correlations can lead to detrimental consequences. For instance, Seyyed-Kalantari et al. [43] found that pneumonia diagnosis systems using chest X-rays may mistakenly base their predictions on patient demographics rather than the actual medical conditions, highlighting the potential risks of these correlations.", + "bbox": [ + 511, + 542, + 908, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing works have attempted to address spurious correlations in zero-shot CLIP models, with many focusing on debiasing the text space. However, as noted by [31], these approaches require domain expertise to identify spurious features and carry the risk of causing misalignment with the image modality. Alternative methods, such as ROBOSHOT [1] and TIE* [31], propose robust solutions but inherently rely on prior identification and annotations of spurious features, which significantly restricts their applicability. As different spurious text prompts yield varying robustness outcomes, identifying the optimal spurious prompts remains challenging. More importantly, the effectiveness of these methods is often compromised in real-world scenarios where spurious features are not predefined,", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2919", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "thereby limiting their practical deployment.", + "bbox": [ + 89, + 90, + 379, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent literature on Test-Time Adaptation explores methods to adapt models at test time using unlabeled test samples. While minimizing the entropy of the predictions has been shown to enhance test-time adaptation, Lee et al. [22] demonstrates that minimizing entropy fails to improve worst-group accuracy with the presence of severe spurious correlation and even degrades robustness sometimes, performing worse than models without adaptation. Test-Time Label Shift Adaptation (TTLSA) [47] addresses spurious label shifts by adapting predictions to the target distribution through calibration on the joint distribution of the target and spurious labels, which corresponds to the differences between the training and target sets. However, this method requires the training set annotations, which are often unavailable given the large-scale and diverse training sources of CLIP models, making it challenging to apply in zero-shot classification.", + "bbox": [ + 89, + 106, + 483, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A critical yet underexplored question remains: Can unknown spurious correlations be mitigated in zero-shot CLIP predictions without access to the information of training statistics or spurious attributes?", + "bbox": [ + 89, + 363, + 482, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we set out to understand spurious correlations in CLIP models. We observe that, given a prompt, CLIP models allocate significant attention to spurious features (e.g., background), as shown in the left panel of Figure 1. Motivated by this observation, we conduct a semantic analysis of CLIP's image encoder. Surprisingly, we find that a pre-trained CLIP model exhibits an emergent object detection capability, allowing it to recognize objects in test images. Building on this insight, we propose test-time reasoning, an approach that mimics human perception by first localizing the core object and then reducing the influence of spurious features to perform zero-shot classification. The proposed approach operates in a purely test-time setting and only requires the test samples and the corresponding text prompt for the classification task, without requiring additional information on spurious attributes or training statistics. We summarize our contribution as follows:", + "bbox": [ + 89, + 424, + 482, + 678 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We discover that the CLIP model, a text-guided training paradigm, acquires semantic understanding and demonstrates a strong capability to identify objects. Notably, it can segment objects effectively even without a dedicated segmentation training procedure.", + "- We propose an interpretable and easy-to-use approach named Test-Time Reasoning (TTR) to mitigate spurious correlation in zero-shot classification with CLIP without requiring expertise or annotations of spurious attributes.", + "- We demonstrated the effectiveness of TTR via comprehensive experiments, achieving a $5.05\\%$ improvement in worst-group accuracy over the leading existing method that relies on spurious information." + ], + "bbox": [ + 89, + 680, + 483, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Spurious correlation mitigation", + "text_level": 1, + "bbox": [ + 513, + 113, + 790, + 128 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spurious correlation refers to the context-dependent clues that are learned by the models during training but might be broken or not exist in test data [3, 13, 39]. There have been various works in spurious correlation mitigation, requiring or without requiring group annotation. We categorize these methods according to their focus on the data or optimization aspects. The data-based methods [11, 16, 18, 25, 40, 58, 61] rely on data balancing or augmentation to improve the robustness against spurious correlation, while optimization-based methods [3, 20, 26, 33, 39, 53] aim to learn more invariant or robust feature representations during model training. Despite the effectiveness, these methods have primarily been studied in architectures like ResNet and often rely on retraining models, making it challenging to directly integrate them into large-scale pretrained models such as CLIP. GC-CLIP [41] improves zero-shot CLIP classification of small objects by cropping images with an auxiliary model. In contrast, our method focuses on spurious correlation and localizes objects using CLIP's own semantic cues without external supervision, making it more broadly applicable.", + "bbox": [ + 511, + 136, + 906, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Spurious correlation mitigation for VLMs", + "text_level": 1, + "bbox": [ + 513, + 450, + 870, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Similar to ResNet and other vision-based models, VLMs also suffer from performance degradation from spurious correlation. However, VLMs, known for their large scale and multimodality, bring new challenges for spurious correlation mitigation. Current work can be mainly categorized depending on whether it modifies the original models. Several methods [15, 56] utilize contrastive learning to fine-tune VLMs with spurious attributes either through text descriptions or explicit detection. Additionally, VLMs' multimodal capabilities can be used to generate supervisory signals, such as saliency maps, which then guide the fine-tuning process [37]. In contrast, zero-shot approaches avoid model modification by contextual prompt engineering, such as Perception CLIP [2], or projection-based debiasing, including Orth-Cali [8], ROBOSHOT [1], and TIE* [31]. More details on the methods can be found in Sec. 9.", + "bbox": [ + 511, + 473, + 906, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Test time adaptation", + "text_level": 1, + "bbox": [ + 513, + 726, + 705, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There is growing research interest in Test Time Adaptation (TTA), which adapts trained models at inference time using unlabeled test samples. TTA closely aligns with the scope of this work, as adapting zero-shot classification tasks also involves unlabeled data. A major branch of TTA is Test Time Training (TTT) [4, 28, 47]. TTT modifies training losses to incorporate self-supervised contrastive objectives, then reapplies these objectives to test samples, enhancing adaptation to unseen data. TTT is challenging to apply in zero-shot prediction since the training procedure", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2920", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cannot be modified in this setting. Another category of TTA approaches focuses on minimizing prediction entropy [32, 34, 50, 60]. However, minimizing entropy fails to address spurious correlations because of the inherent group imbalance in the dataset [22]. Test Time Label Shift Adaptation (TTLSA) [46] addresses spurious correlations via test-time adaptation. However, this method requires training set annotations to detect label shifts during testing. Applying TTLSA to CLIP models is challenging due to the difficulty in obtaining spurious correlation information from the training data, given the complexity inherent in foundation model datasets.", + "bbox": [ + 89, + 90, + 482, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem formulation and observation", + "text_level": 1, + "bbox": [ + 89, + 287, + 429, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we investigate why the CLIP model exhibits less robustness in the presence of spurious features. Our analysis reveals that spurious features receive disproportionately higher attention weights compared to core features. This misallocated attention contributes to the model's decreased robustness in zero-shot classification. We further analyze the semantic representations learned by CLIP and reveal its remarkable ability to identify core objects in images. This powerful capability can be harnessed to enhance the robustness of zero-shot classification, offering new insights into CLIP's potential.", + "bbox": [ + 89, + 310, + 482, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To understand the poor worst-group performance of CLIP on datasets with spurious correlations, we hypothesize two possible explanations: (1) CLIP model fails to localize the core object due to the influence of spurious features, or (2) CLIP model fails to correctly identify the core object. To test these hypotheses, we conduct experiments on the well-established Waterbirds dataset [39], where the bird serves as the core feature and the background as the spurious feature. A detailed introduction is in Section 9.", + "bbox": [ + 89, + 477, + 482, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spurious feature dominates attention weights. We first investigate the wrong attention allocation hypothesis. Darcet et al. [10] report that final attention maps highlight semantically meaningful regions, Simeoni et al. [44] utilize attention maps for object detection. Additionally, Chefer et al. [6] developed an explanation tool to analyze how vision transformers allocate attention. These works suggest that attention maps provide strong evidence for analyzing transformer-based models.", + "bbox": [ + 89, + 628, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on these insights, we apply the attention-based explanation method from Chefer et al. [6] to CLIP ViT/L-14 and ViT/B-32 [7], visualizing the attention map using [CLS] tokens for image embeddings, as shown in Figure 1. We experiment with transformers of varying sizes, as [44] reports that large-scale models potentially show outlier artifacts, whereas base models show fewer effects. Across both models, we observe that attention maps predominantly highlight background regions, indicating that CLIP models", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lack robustness in the presence of spurious features, failing to accurately localize the core features.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP accurately classifies the core object in a spurious feature-free setting. We then validate the CLIP model's zero-shot ability to classify core features. To eliminate the effect of spurious features on performance, we construct the Core Object Only datasets. This dataset consists solely of birds, ensuring that the bird objects are identical to those in the Waterbirds original test set. The backgrounds are uniformly black (BLK) or white (Wht). For baseline comparison, we create the Original Variant (OV) dataset, where the core objects remain identical to those in the Core Object Only datasets, but spurious features (backgrounds) are reintroduced. We provide the details of dataset preparation and experimental settings in Section 7. We test the zero-shot classification with the same text prompts with the original Waterbirds test set and the Core Object Only sets with ViT/L-14 backbone [7]. The results are shown in Table 1.", + "bbox": [ + 511, + 132, + 906, + 375 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/ae125d6233d3821fdc56d05d25d8fd36d5838ab1cddb1f8e28d3fce2617a27fe.jpg", + "table_caption": [ + "Table 1. Zero-shot results on a variant of waterbirds datasets." + ], + "table_footnote": [], + "table_body": "
OVCore BlkCore Wht
WGAvgWGAvgWGAvg
34.1387.6291.3992.3790.9693.02
", + "bbox": [ + 531, + 402, + 885, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We observe that the worst-group accuracy significantly improves by $57.26\\%$ and $56.83\\%$ compared to the dataset containing spurious features. This significant improvement indicates the CLIP model's capability to accurately classify the core object in an image. We conclude that, in the absence of spurious features, the CLIP model achieves strong performance in zero-shot classification.", + "bbox": [ + 511, + 465, + 905, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-trained CLIP acquired semantic meaning. We further analyze the image embeddings from the CLIP model, as they are deterministic. Although the text space is equally important, it is significantly impacted by text prompts, and optimal prompts are challenging to obtain.", + "bbox": [ + 511, + 583, + 905, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To analyze the embedding structure, we extract image embeddings excluding the [CLS] token and represent them as $\\mathbf{z}_{img} \\in \\mathbb{R}^{N \\times d}$ , where $N$ denotes the number of patches and $d$ is the latent dimension. We apply Principal Component Analysis (PCA) to reduce the $d$ -dimensional embeddings to three components, extracting the top principal components for each image and obtaining $\\mathbf{z}_{reduced} \\in \\mathbb{R}^{N \\times 3}$ . We map these three dimensions to RGB values for visualization [35], as shown in the middle row of Figure 2. In each plot, patches with similar hues indicate higher similarity. For instance, patches with pink hues show strong semantic grouping. This visualization demonstrates that the CLIP model effectively captures semantic representations in its latent space.", + "bbox": [ + 511, + 659, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, we apply PCA to a batch of 10 images to extract shared information across the batch. K-means clus", + "bbox": [ + 511, + 869, + 903, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2921", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ad9746e1d401f4f74e3a18163837b9d9acfc00cef69d74f763c79b7d9eea2bde.jpg", + "image_caption": [ + "Figure 2. Semantic analysis results. Middle: We visualize PCA features using different colors, where patches with similar hues indicate semantic similarity. Bottom: Semantic masks are shown in different colors. In the last column, multiple birds within a single image are accurately identified." + ], + "image_footnote": [], + "bbox": [ + 114, + 90, + 460, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tering is then performed on the PCA-reduced features to group image patches based on similarity. We set the number of clusters $k = 3$ for this experiment. Based on the clustering results, we generate masks to segment the object, as illustrated in Figure 2. The results show that the CLIP model effectively segments core objects, such as birds. Additional results on various datasets are presented in Figure 5, demonstrating that CLIP captures semantically meaningful representations across diverse datasets.", + "bbox": [ + 89, + 316, + 483, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. TTR: Test Time robustification via reasoning", + "text_level": 1, + "bbox": [ + 89, + 465, + 483, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The analysis in Section 3 reveals that the vanilla zero-shot CLIP model disproportionately allocates attention to spurious features, thereby compromising its generalizability. However, the study also validates that the CLIP model possesses strong semantic identification capabilities, enabling it to effectively recognize semantically meaningful objects in the latent space.", + "bbox": [ + 89, + 491, + 482, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate the impact of spurious correlations in zero-shot CLIP models, we propose a test-time reasoning framework inspired by human visual cognition. Conceptually, the model first localizes the object before performing classification, aligning with the human perceptual process [5]. Before presenting our solution, we conduct a study on semantic identification and robustness enhancement.", + "bbox": [ + 89, + 598, + 482, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Improved semantic identification recipe", + "text_level": 1, + "bbox": [ + 89, + 712, + 431, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the CLIP model demonstrates robust performance on images containing only core objects, we conduct the following study to improve core object identification. Specifically, we ablate the use of auxiliary models or the pure CLIP model to perform zero-shot core feature segmentation. Regarding the auxiliary models, we leverage state-of-the-art world models, including Grounding DINO (GDINO) [27] and the combination of Segment Anything (SAM) [21].", + "bbox": [ + 89, + 734, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In terms of the pure CLIP model, we explore the following techniques: (1) dimension reduction, (2) high-dimensional clustering, and (3) a hybrid approach combin", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing dimension reduction with low-dimensional clustering. Due to the page limits, a detailed introduction to the experiment setting is provided in Section 8. We conduct experiments on the Waterbirds using the original test set. The proposed approaches are evaluated from multiple perspectives, including utility (worst-group accuracy (WG), average accuracy (Avg), and DICE score), test-time latency (measured in minutes), and peak GPU memory consumption (VRAM usage). The results are shown in Table 2. The visualization results for segmentation, object detection are shown in Figure 4.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 presents a plot of the relationship between DICE scores and worst-group accuracy, illustrating that precise localization of the core object positively impacts prediction robustness. In other words, an improved core object identification algorithm directly enhances the robustness of CLIP's zero-shot predictions. From Table 2, we observe that GDINO improves worst-group accuracy. As it only provides bounding boxes, spurious features may still be retained alongside the core object. Notably, combining GDINO with SAM to accurately identify the core object and using only that object for zero-shot classification yields the highest robust performance with CLIP models. Under this approach, worst-group accuracy improves significantly and nearly reaches average accuracy. This result echoes the observation from Section 3 and indicates a uniform high performance across groups. However, this robustness comes at a substantial cost of higher latency and significant GPU memory requirements.", + "bbox": [ + 511, + 258, + 906, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By contrast, K-means clustering on the image representations also elevates worst-group accuracy. This suggests that the learned representations inherently capture semantic meaning, as clustering algorithms can effectively group meaningful features and produce well-segmented results. Combining PCA with K-means provides the best trade-off between robustness and efficiency. Reducing features to a 3-dimensional space before clustering significantly improves worst-group accuracy while maintaining low latency. Furthermore, this approach does not require auxiliary models, achieving strong performance while maintaining low GPU memory consumption. As a resource-efficient method, it provides an effective balance between performance and computational efficiency.", + "bbox": [ + 511, + 532, + 908, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Core feature identification and Zero-Shot robustification.", + "text_level": 1, + "bbox": [ + 511, + 756, + 906, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now present our solution TTR for enhancing the robustness of zero-shot classification on CLIP. Based on the insights from Section 4.1, we utilize PCA for dimensionality reduction, followed by K-means clustering for object identification. As discussed in Section 3, processing a batch of images yields more robust results. However, we expect to enable inference on any number of samples. To address this", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2922", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/859119eee0d0a4c11bf9838a2888cbb37f40cf9beac3d36781292927e70db60a.jpg", + "table_caption": [ + "Table 2. Ablation study on different methods. Latency (Lat) is reported in minutes, and GPU memory usage (VRAM) is reported in gigabytes (GB). ref stands for the reference." + ], + "table_footnote": [], + "table_body": "
WGAvgDICELatVRAM
Zero-Shot47.3673.70-2.52.0
+GDINO66.80 (19.44 ↑)82.22 (8.52 ↑)-18.64.6
+GDINO+ SAM87.79 (40.43 ↑)88.18 (14.48 ↑)ref81.511.9
+PCA49.93 (2.57 ↑)71.83 (1.87 ↓)40.535.52.5
+TSNE60.11 (12.75 ↑)76.65 (2.95 ↑)38.7159.02.5
+UMAP36.01(11.35 ↓)63.41 (10.29 ↓)32.2121.52.5
+K-means (d = 768)69.43(22.07 ↑)82.24 (8.54 ↑)61.6315.02.0
+TSNE+K-means64.31(16.95 ↑)78.99 (5.29 ↑)42.4659.62.5
+PCA+K-means74.14 (26.78 ↑)87.21(13.51 ↑)65.486.02.5
", + "bbox": [ + 99, + 127, + 638, + 286 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/252c439eb3c08548ca256f90cbeee47800670fd4ed69fc5e682c44b3202f49e9.jpg", + "image_caption": [ + "Figure 3. Characteristics curve of CLIP for waterbirds dataset: DICE vs. WG" + ], + "image_footnote": [], + "bbox": [ + 661, + 92, + 901, + 241 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59cbc6ba9d649efe8d5efc7751a760a80e38f8358f0069d5c357c60e48417daf.jpg", + "image_caption": [ + "Figure 4. Visualization of core object identification for each method. GDINO provides bounding boxes. Other methods generate core object masks, which are shown in black. Identifiable artifacts are highlighted with an orange circle. We perform zero-shot classification based on the core object identified by each method." + ], + "image_footnote": [], + "bbox": [ + 93, + 308, + 480, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "limitation, we introduce a queue to store historical data. The queue length $l \\in R^{+}$ is a hyperparameter. A more detailed discussion and experiment are presented in section 5.4. For each test image, we pop up the oldest image from the queue and append the new image. We perform PCA on the queue to find the semantic structure in each image. To improve the continuity of the feature recognition, we apply a Gaussian filter to smooth the PCA features before performing k-means clustering. We employ k-means clustering to image patches based on the smoothed PCA features. We evaluated alternative clustering methods, including Gaussian Mixture Models and hierarchical clustering, and found that K-means clustering yields the best results. In our experiment, we set all datasets with $k = 3$ for all datasets to ensure consistency and ease of use. An ablation study on the number of clusters is presented in Section 5.4.", + "bbox": [ + 88, + 516, + 482, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the clustering results, we generate segmentation masks and apply them to the original test images. To eliminate any influence from unmasked regions, we set them to zero, ensuring they do not contribute to the final prediction. This process produces $k$ segmented images.", + "bbox": [ + 89, + 761, + 482, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To identify the segmentation containing the core object, we employ zero-shot classification. As this approach relies on text prompts, we introduce a prompt selection strategy that captures the core object of interest. Instead of using a", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "specific prompt such as \"a photo of a waterbird,\" we adopt a more general prompt, \"a photo of a bird,\" which provides a higher-level abstraction of the query. Using the general text embeddings, we compute image embeddings for each segmentation and measure their inner product with the text prompt. Drawing insights from Section 3, where the core features can be accurately classified without the influence of spurious features, we consider the segment closest to the prompt as the one containing core objects and perform zero-shot classification solely on this segmentation. The TTR procedure does not require information about spurious features, highlighting its ease of use and broad applicability.", + "bbox": [ + 511, + 309, + 906, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 513, + 505, + 638, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 513, + 531, + 596, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We conduct a comprehensive evaluation on six well-established benchmark datasets commonly used in spurious correlation studies [1, 20, 30, 39, 54]. These datasets include Waterbirds [39], CelebA [29], Metashifts [24], UrbanCars [23], ISIC [9], and CXR [51]. A detailed introduction to these datasets is provided in Appendix 9.", + "bbox": [ + 511, + 551, + 905, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our method against baselines in two categories and a detailed introduction to each baseline is provided in Appendix 9:", + "bbox": [ + 511, + 648, + 906, + 695 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Robustifying vision-language models: Orth-Cali [8], Perception CLIP [2], ROBOSHOT [1], TIE* [31], and BEND-VLM [14].", + "- Test-time adaptation methods: MeanShift for Test-time Augmentation (MTA) [59], Test-Time Low-Rank Adaptation (TTLRA) [17], and Test-Time Label-Shift Adaptation (TTLSA) [46]." + ], + "bbox": [ + 513, + 700, + 906, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backbone and Metrics. Lu et al. [31] demonstrate that ViT-L/14 [12] shows a significant gap between worst-group accuracy and average accuracy, highlighting the need for robustification in zero-shot prediction for CLIP ViT-L/14. Based on this observation, we employ CLIP ViT-L/14 [7]", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2923", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for general tasks. For ISIC and CXR datasets, we follow [1] and utilize Biomed CLIP [62] as the zero-shot prediction backbone. Additionally, we present results for various Vision Transformer architectures in Section 5.4.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the protocol in the robust learning study [39], we report three key metrics: Worst-group accuracy (WG), Average accuracy (Avg), and the gap between these metrics (Gap). Ideally, both metrics should be high and close in value to ensure accurate and robust performance. We bold the best results and underline the second-best results. Detailed implementation and reproducible text prompts are provided in Section 10.", + "bbox": [ + 89, + 152, + 480, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Core object identification", + "text_level": 1, + "bbox": [ + 89, + 285, + 321, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on our analysis in Section 3, high-quality core object identification significantly enhances robustness. To validate our approach TTR, we first assess its ability to correctly identify the core object. We visualize the segments with the highest similarity to the general text prompts across all test datasets, as shown in Figure 5.", + "bbox": [ + 89, + 309, + 480, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our observations indicate that across all evaluated datasets, TTR consistently identifies the core object in the test set successfully. Interestingly, in the CelebA dataset, the spurious feature (gender) is closely located near the core feature (hair color). TTR effectively distinguishes and segments the core feature, demonstrating CLIP's strong reasoning capabilities. For the Metashifts dataset, we observe that the color of a cat and a boot are similar (see the top-right corner of Metashifts in Figure 5). Despite this, TTR effectively segments the cat, further validating CLIP's capability for semantic understanding. For the CXR dataset, chest X-ray orientations vary in the test set (e.g., frontal vs. lateral views). TTR successfully extracts these view differences and highlights the lung region, ensuring robustness across different imaging perspectives. In the ISIC dataset, more than two spurious features exist, such as color patches, rulers, hairs, and bubbles [54]. However, TTR does not rely on these artifacts, as it focuses solely on segmenting the core feature while ensuring accurate identification of the lesion region with integrity.", + "bbox": [ + 89, + 400, + 482, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Main result", + "text_level": 1, + "bbox": [ + 89, + 714, + 217, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The main results on general datasets are presented in Table 3. Additional requirements beyond the task-related text prompt for running each method are detailed in the table. \"Prompt $a$ \" indicates the need for spurious text prompts, with most methods relying on large language models (LLMs) to provide spurious feature descriptions. The \"label $a$ \" denotes that the method requires spurious labels for the test set. \"Train $y, a$ \" indicates the requirement for accessing training set annotations.", + "bbox": [ + 89, + 737, + 480, + 873 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TTR consistently achieves robust results. For the Waterbirds dataset, TTR improves worst-group performance by $5.3\\%$ compared to the second-best method, despite requiring fewer resources. For the UrbanCars dataset, the zero-shot baseline exhibits a significant performance gap between WG and Avg, indicating a strong spurious correlation between cars and background. After implementing our method, this gap is reduced from $43.3\\%$ to $22.3\\%$ , demonstrating its effectiveness in mitigating spurious correlations. We evaluate the average performance across the general dataset, where Orth-Cali, TIE*, and TTLSA achieve performance improvements. However, these methods require prior knowledge of spurious correlations. In contrast, TTR outperforms all other methods, exceeding the second-best approach by $5.05\\%$ in worst-group accuracy while requiring fewer annotations.", + "bbox": [ + 511, + 90, + 903, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TTR achieves higher average accuracy compared to benchmark methods, highlighting its strong capability to capture core object features and classify them accurately. The gap between average accuracy and worst-group accuracy serves as a measure of robustness. Across the average of general datasets, TTR shows the smallest gap, demonstrating its robustness in zero-shot prediction.", + "bbox": [ + 511, + 333, + 903, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4 reports the results for the high-stakes biomedical domain datasets. The results demonstrate that our method achieves the highest worst group accuracy while maintaining a small gap between WG and average accuracy. TTLSA achieves the highest average accuracy on both datasets. However, the large gap between WG and average accuracy is concerning in class-imbalanced test sets, as it suggests that the model predominantly predicts a single class, resulting in high overall accuracy while failing to perform well across all groups. In contrast, our method achieves a more balanced performance across different groups. While it may reduce average accuracy, this is often due to the removal of the strong spurious correlation (such as color patches in ISIC), which correlates with the majority class [55]. By encouraging the model to focus on semantically meaningful regions, our method improves reliability by avoiding short cues that might lead to clinically critical cases.", + "bbox": [ + 511, + 438, + 903, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Ablation study", + "text_level": 1, + "bbox": [ + 513, + 703, + 663, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we explore two types of ablation studies. The parameter ablation examines the effect of hyperparameters within TTR on the results. The structure ablation investigates the impact of different types of masks and variations in vision transformers on performance.", + "bbox": [ + 511, + 724, + 903, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4.1. Parameters ablation", + "text_level": 1, + "bbox": [ + 511, + 806, + 700, + 820 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The length of the queue We first examine the effect of queue length on performance. A queue stores historical samples, which can potentially enhance robustness. To analyze this, we conduct an ablation study on the Waterbirds dataset, varying the queue length from 0 (i.e., no queue) to", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Code at https://github.com/lu876/TTR", + "bbox": [ + 107, + 886, + 382, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2924", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/37532b170e90fd63135486d7b8e7f34e771ca5d7a66fa170dddde91689e98c45.jpg", + "image_caption": [ + "Figure 5. Core object identification results across benchmark datasets." + ], + "image_footnote": [], + "bbox": [ + 171, + 90, + 823, + 369 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8dd043401862b1d84161d19a1c2daacd8bb4a8fecf2a7a0e4b00023b48e44d37.jpg", + "table_caption": [ + "Table 3. Zero-Shot classification result on the general dataset." + ], + "table_footnote": [], + "table_body": "
RequirementWaterbirdsCelebAMetashiftUrbanCarsAverage
WG↑Avg↑WG↑Avg↑WG↑Avg↑WG↑Avg↑WG↑Acc↑
Orth-CaliPrompt a68.5186.3176.1486.2481.7081.7024.4054.7062.6977.24
PerceptionCLIPPrompt a54.1286.7474.3180.3090.8191.6727.6056.7061.7178.85
ROBOSHOTPrompt a45.1764.4382.6185.5490.6492.6114.0052.7058.1173.82
TIE*Prompt a61.6078.9881.9884.2791.7592.9726.8052.4065.5377.16
BEND-VLMLabel a39.8253.3479.1284.6790.6492.051.2049.7052.6969.94
TTLSATrain {y, a}53.1585.4082.1084.3990.5292.2322.8052.3062.1478.58
MTANone41.4271.1979.1683.3489.8792.057.6051.2054.5174.45
TENTNone52.5975.7774.4490.4988.6891.405.2051.9055.2377.39
TTLRANone41.6471.7876.8981.9091.5093.354.8050.9153.7174.49
Zero-ShotNone47.3673.7073.3581.2089.5491.688.0051.3054.5674.47
TTR (Ours)None74.1487.2183.3386.0090.8591.1334.0056.3070.5880.16
", + "bbox": [ + 93, + 421, + 906, + 633 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/785533a0f9461df7d56cd1d081515b61fc64cdf6bd562e773cc801241f6bd358.jpg", + "table_caption": [ + "Table 4. Zero-Shot classification results on ISIC and CXR datasets" + ], + "table_footnote": [], + "table_body": "
ISICCXR
WG↑Avg↑Gap↓WG↑Avg↑Gap↓
Orth-Cali21.4372.5451.1142.1153.8511.74
PerceptionCLIP41.5552.7411.1915.7974.7959.00
ROBOSHOT53.3059.846.5441.6256.2314.61
TIE*61.1171.6810.5742.1173.9231.81
BEND-VLM46.8376.0529.2226.3265.3939.07
TTLSA48.4179.7731.3610.5380.7770.24
MTA44.4465.2120.7747.3761.5414.17
TENT51.2657.616.3521.0570.5149.46
TTLRA60.1866.996.8118.4279.4961.07
Zero-Shot42.2170.2128.0028.9055.2826.38
TTR (Ours)64.2868.614.3355.2663.257.99
", + "bbox": [ + 91, + 670, + 483, + 881 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "30. We report both worst-group accuracy and average accuracy to assess the impact of different queue lengths, as shown in the left panel of Figure 6. The results clearly demonstrate that employing a queue mechanism significantly improves both worst-group and average accuracy. Additionally, performance remains relatively stable across queue lengths that are larger than 5.", + "bbox": [ + 511, + 647, + 906, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The number of clusters We then analyze the effect of the number of clusters used to generate masks. A small number of clusters may cause core object features and spurious features to be grouped together, reducing the effectiveness of TTR. Conversely, an excessive number of clusters can fragment the core features, making them less intact and ultimately diminishing zero-shot performance. The results are shown in the right panel of Figure 6. When the number of clusters is set to 1, the method is equivalent to standard zero", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2925", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8b9dc50fef27ab70310a165067bd0992ee6150e911bca4a22d66ba59545abac8.jpg", + "image_caption": [ + "Figure 6. Left: WG and Avg change w.r.t. the length of the queue. Right: WG and Avg change w.r.t. the number of cluster." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 285, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ea068f796e0181a081c48d50dce58fda23abcfcfbe6e18d82bc447f9a7965f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 90, + 478, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shot classification. With 2 or 3 clusters, TTR significantly improves both worst-group and average accuracy. However, an excessively high number of clusters negatively impacts performance. Based on these findings, we recommend using 2 or 3 clusters to achieve optimal performance.", + "bbox": [ + 89, + 268, + 483, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4.2. Structure ablation", + "text_level": 1, + "bbox": [ + 89, + 352, + 264, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mask types Wang and Wang [52] demonstrated that the method used to mask task-unrelated features significantly impacts the final prediction. Since masking operations alter the distribution of test samples, choosing an appropriate masking strategy is crucial. The major masking methods include filling with zero, filling with the mean value of all masked patches, and applying Gaussian blurring to the masked region. To determine the most effective masking method for TTR in CLIP models, we conduct an ablation study comparing these approaches.", + "bbox": [ + 89, + 371, + 483, + 523 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45e5c7595e28323a0bcfe4571c7330b500a11f260a74cdd048c901632f6599f1.jpg", + "image_caption": [ + "Figure 7. Comparison of the effectiveness with different masking types applied for the region of task-unrelated features. Left: worst group accuracy; Right: average accuracy." + ], + "image_footnote": [], + "bbox": [ + 94, + 536, + 285, + 631 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4049c670da3c576a06d0ad4bc83fd957e56b73654fc2456483a0e347865716ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 535, + 480, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results from Figure 7 indicate that the zero-mask approach achieves the highest worst-group accuracy among all methods. We hypothesize that this is because zero masking effectively eliminates spurious feature information and disrupts reliance on spurious features more drastically. The effectiveness of zero masking is further supported by findings in Section 3, where a black-background version of the Waterbirds dataset also achieved the highest worst-group accuracy, aligning with our observations. Based on these results, we recommend using the zero-mask strategy.", + "bbox": [ + 89, + 696, + 482, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More results on different CLIP models We extend TTR to additional CLIP models to evaluate its generalizability. Specifically, we apply TTR to ViT-B/32 and ViT-H/14 to", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "assess its effectiveness across different scales of CLIP models. The results are presented in Figure 8, where we compare the baseline zero-shot performance of each architecture with the performance change after applying TTR. The findings clearly show that TTR improves worst-group accuracy across all tested datasets, demonstrating its adaptability to different vision transformer architectures. The right panel of Figure 8 illustrates the changes in average accuracy. While one dataset with ViT-L exhibits a slight performance decrease, all others show varying degrees of improvement. This suggests that TTR enhances generalizability in zero-shot prediction.", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b6a993051bae7a1f3e7103430a1a94cdae2a619e6132e0e9abba0d667cfcc23b.jpg", + "image_caption": [ + "Figure 8. Ablation on different ViT architectures. The pink bar represents TTR improvement, the white bar indicates reduction. TTR enhances performance across various vision transformers." + ], + "image_footnote": [], + "bbox": [ + 519, + 286, + 712, + 386 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/beb264dedb684bb623e55af61ac97c00c8864d0b0dbaaac27fb3c6015955d72e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 286, + 901, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 513, + 457, + 627, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce TTR, a test-time robustification method to effectively mitigate the influence of spurious correlation on CLIP for robust zero-shot classification, without requiring annotations or prior knowledge on spurious attributes. Motivated by human reasoning for recognition, we examined the actual zero-shot classification capacity of CLIP in the absence of spurious correlations and further proposed TTR on top of that. We reveal and harness the semantic understanding capability of the CLIP image encoder to localize the area of the objects for refined classification. Compared to the standard zero-shot scheme, TTR not only improves classification performance but also enhances the interpretability of predictions. We demonstrate the effectiveness and robustness of TTR, which significantly boosts both worst-group and average accuracy on real-world datasets with challenging spurious correlations.", + "bbox": [ + 511, + 483, + 906, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and Future direction. Despite mitigating spurious correlation effectively, TTR depends on CLIP's ability to segment the core features. When spurious attributes are abstract, such as texture, the segmentation can become infeasible, limiting TTR's applicability. Additionally, while we demonstrated its effectiveness with simple text prompts based on the classification task, the performance of TTR could be influenced by the quality of the prompts in complex scenarios. Addressing these challenges or exploring the applicability in other multimodal tasks can be promising avenues for future research.", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2926", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 198, + 187, + 214 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dyah Adila, Changho Shin, Linrong Cai, and Frederic Sala. Zero-shot robustification of zero-shot models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 6, 3", + "[2] Bang An, Sicheng Zhu, Michael-Andrei Panaitescu-Liess, Chaithanya Kumar Mummadi, and Furong Huang. Perceptionclip: Visual classification by inferring and conditioning on contexts. In The Twelfth International Conference on Learning Representations. 2, 5, 3", + "[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2", + "[4] Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In International Conference on Artificial Intelligence and Statistics, pages 3080-3090. PMLR, 2022. 2", + "[5] Irving Biederman. Recognition-by-components: a theory of human image understanding. Psychological review, 94(2): 115, 1987. 4", + "[6] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 397-406, 2021. 3", + "[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 3, 5, 1", + "[8] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2, 5, 3", + "[9] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), pages 168-172. IEEE, 2018. 5", + "[10] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 3" + ], + "bbox": [ + 93, + 223, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was partially supported by the EMBRIO Institute, contract #2120200, a National Science Foundation (NSF) Biology Integration Institute, and NSF IIS #1955890, IIS #2146091, IIS #2345235.", + "[11] Yihe Deng, Yu Yang, Baharan Mirzasoleiman, and Quanquan Gu. Robust learning with progressive data expansion against spurious correlation. Advances in neural information processing systems, 36:1390-1402, 2023. 2", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5, 1", + "[13] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 2", + "[14] Walter Gerych, Haoran Zhang, Kimia Hamidieh, Eileen Pan, Maanas Sharma, Thomas Hartvigsen, and Marzyeh Ghassemi. Bendvlm: Test-time debiasing of vision-language embeddings. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 5, 3", + "[15] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19338-19347, 2023. 2", + "[16] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. 2", + "[17] Raza Imam, Hanan Gani, Muhammad Huzaifa, and Karthik Nandakumar. Test-time low rank adaptation via confidence maximization for zero-shot generalization of vision-language models. arXiv preprint arXiv:2407.15913, 2024. 5, 3", + "[18] Taeuk Jang and Xiaoqian Wang. Difficulty-based sampling for debiased contrastive representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24039-24048, 2023. 2", + "[19] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. In The Eleventh International Conference on Learning Representations. 1", + "[20] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. In The Eleventh International Conference on Learning Representations, 2023. 2, 5", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 4, 1", + "[22] Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. arXiv preprint arXiv:2403.07366, 2024. 2, 3", + "[23] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and" + ], + "bbox": [ + 93, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2927", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20071-20082, 2023. 5", + "[24] Weixin Liang and James Zou. Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. In International Conference on Learning Representations, 2022. 5", + "[25] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. 2", + "[26] Sheng Liu, Xu Zhang, Nitesh Sekhar, Yue Wu, Prateek Singhal, and Carlos Fernandez-Granda. Avoiding spurious correlations via logit correction. In The Eleventh International Conference on Learning Representations, 2022. 2", + "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 4, 1", + "[28] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 2", + "[29] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5", + "[30] Shenyu Lu, Junyi Chai, and Xiaoqian Wang. Neural collapse inspired debiased representation learning for min-max fairness. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2048-2059, 2024. 5", + "[31] Shenyu Lu, Junyi Chai, and Xiaogian Wang. Mitigating spurious correlations in zero-shot multimodal models. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2, 5", + "[32] Chaithanya Kumar Mummadi, Robin Hutmacher, Kilian Rambach, Evgeny Levinkov, Thomas Brox, and Jan Hendrik Metzen. Test-time adaptation to distribution shift by confidence maximization and input transformation. arXiv preprint arXiv:2106.14999, 2021. 3", + "[33] Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: Training debiased classifier from biased classifier. In Advances in Neural Information Processing Systems, 2020. 2", + "[34] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 3", + "[35] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez," + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3", + "[36] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournaepau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830, 2011. 4", + "[37] Suzanne Petryk, Lisa Dunlap, Keyan Nasseri, Joseph Gonzalez, Trevor Darrell, and Anna Rohrbach. On guiding visual attention with language specification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18092-18102, 2022. 2", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1", + "[39] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 2, 3, 5, 6", + "[40] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. 2", + "[41] Piyapat Saranrittichai, Mauricio Munoz, Volker Fischer, and Chaithanya Kumar Mummadi. Zero-shot recognition with guided cropping. In ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models, 2024. 2", + "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 1", + "[43] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA McDermott, Irene Y Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12):2176-2182, 2021. 1", + "[44] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 3", + "[45] Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, and Christopher Ré. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems, 33: 19339-19352, 2020. 1", + "[46] Qingyao Sun, Kevin P Murphy, Sayna Ebrahimi, and Alexander D'Amour. Beyond invariance: test-time label" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2928", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "shift adaptation for addressing \"spurious\" correlations. Advances in Neural Information Processing Systems, 36: 23789-23812, 2023. 3, 5", + "[47] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2", + "[48] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 1", + "[49] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations. 3", + "[50] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020.3", + "[51] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 5", + "[52] Yipei Wang and Xiaogian Wang. Benchmarking deletion metrics with the principled explanations. In *Forty-first International Conference on Machine Learning*, 2024. 8", + "[53] Yipei Wang and Xiaoqian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In International Conference on Artificial Intelligence and Statistics, pages 3745-3753. PMLR, 2024. 2", + "[54] Shirley Wu, Mert Yuksekgonul, Linjun Zhang, and James Zou. Discover and cure: Concept-aware mitigation of spurious correlation. In International Conference on Machine Learning, pages 37765-37786. PMLR, 2023. 5, 6, 2", + "[55] Xiwei Xuan, Ziquan Deng, Hsuan-Tien Lin, and Kwan-Liu Ma. Slim: Spuriousness mitigation with minimal human annotations. In European Conference on Computer Vision, pages 215–231. Springer, 2024. 6", + "[56] Yu Yang, Besmira Nushi, Hamid Palangi, and Baharan Mirzasoleiman. Mitigating spurious correlations in multimodal models during fine-tuning. In International Conference on Machine Learning, pages 39365-39379. PMLR, 2023. 2", + "[57] Yuzhe Yang, Haoran Zhang, Dina Katabi, and Marzyeh Ghassemi. Change is hard: A closer look at subpopulation shift. arXiv preprint arXiv:2302.12254, 2023. 2", + "[58] Huaxiu Yao, Yu Wang, Sai Li, Linjun Zhang, Weixin Liang, James Zou, and Chelsea Finn. Improving out-of-distribution robustness via selective augmentation. In International Conference on Machine Learning, pages 25407-25437. PMLR, 2022. 2", + "[59] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 23783-23793, 2024. 5, 3", + "[60] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in neural information processing systems, 35:38629-38642, 2022. 3", + "[61] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Re. Correct-n-contrast: a contrastive approach for improving robustness to spurious correlations. In International Conference on Machine Learning, pages 26484–26516. PMLR, 2022. 2", + "[62] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 6", + "[63] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 1" + ], + "bbox": [ + 516, + 92, + 903, + 388 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2929", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/87c172f1-4134-4138-9793-b4455c2d3c37_model.json b/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/87c172f1-4134-4138-9793-b4455c2d3c37_model.json new file mode 100644 index 0000000000000000000000000000000000000000..47450166d009f0c11995be749ee1f43f1a0ad6ce --- /dev/null +++ b/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/87c172f1-4134-4138-9793-b4455c2d3c37_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.13, + 0.888, + 0.152 + ], + "angle": 0, + "content": "Think Twice: Test-Time Reasoning for Robust CLIP Zero-Shot Classification" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.181, + 0.749, + 0.234 + ], + "angle": 0, + "content": "Shenyu Lu Zhaoying Pan Xiaoqian Wang* Elmore Family School of Electrical and Computer Engineering Purdue University" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.237, + 0.652, + 0.253 + ], + "angle": 0, + "content": "{lu876, pan433, joywang}@purdue.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.486, + 0.683 + ], + "angle": 0, + "content": "Contrastive Language-Image Pre-training (CLIP) models exhibit intriguing properties, particularly in their zero-shot classification capability. However, the reliability of CLIP zero-shot classification is severely undermined by spurious correlations. Existing efforts to enhance the robustness of zero-shot CLIP models often rely on prior knowledge or annotations of spurious correlations, limiting real-world applicability due to the unavailability of such information. Alternative methods attempt to detect distribution shift at test time but require training statistics whose access is often restricted or computationally expensive. To address the challenges brought by spurious correlation under zero-shot settings, we propose a novel test-time reasoning approach. Our method, inspired by human recognition, localizes the object and refines the classification accordingly. The inherent capacity of CLIP for semantic understanding allows us to isolate the object of interest without auxiliary models. Zero-shot classification is then performed exclusively on the localized objects, effectively mitigating the influence of spurious correlation. The proposed approach is interpretable and flexible as it requires no spurious annotations or prior knowledge, making it widely applicable. The substantial improvements across multiple benchmark datasets validated the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.713, + 0.222, + 0.729 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.739, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Contrastive Language-Image Pre-training (CLIP) [38] has achieved significant success as a powerful pre-training paradigm. It leverages a unified multimodal framework that integrates images and text to achieve strong generalizability. The zero-shot capability of CLIP allows it to classify unseen test samples, making it highly applicable in real-world scenarios. Despite its impressive generalizability, CLIP remains susceptible to spurious correlations, undermining the reliability of its zero-shot predictions. A spurious correla" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.285, + 0.741, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.286, + 0.884, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.464, + 0.908, + 0.518 + ], + "angle": 0, + "content": "Figure 1. Attention map visualization based on [CLS] for randomly selected samples. Left: Vanilla zero-shot prediction with different vision transformers. Right: Proposed method applied to the same vision transformers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.909, + 0.68 + ], + "angle": 0, + "content": "tion occurs when a feature is correlated with the target label but lacks a causal relationship [19]. This issue is particularly critical in high-stakes settings where reliance on such correlations can lead to detrimental consequences. For instance, Seyyed-Kalantari et al. [43] found that pneumonia diagnosis systems using chest X-rays may mistakenly base their predictions on patient demographics rather than the actual medical conditions, highlighting the potential risks of these correlations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Existing works have attempted to address spurious correlations in zero-shot CLIP models, with many focusing on debiasing the text space. However, as noted by [31], these approaches require domain expertise to identify spurious features and carry the risk of causing misalignment with the image modality. Alternative methods, such as ROBOSHOT [1] and TIE* [31], propose robust solutions but inherently rely on prior identification and annotations of spurious features, which significantly restricts their applicability. As different spurious text prompts yield varying robustness outcomes, identifying the optimal spurious prompts remains challenging. More importantly, the effectiveness of these methods is often compromised in real-world scenarios where spurious features are not predefined," + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2919" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.38, + 0.106 + ], + "angle": 0, + "content": "thereby limiting their practical deployment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.362 + ], + "angle": 0, + "content": "Recent literature on Test-Time Adaptation explores methods to adapt models at test time using unlabeled test samples. While minimizing the entropy of the predictions has been shown to enhance test-time adaptation, Lee et al. [22] demonstrates that minimizing entropy fails to improve worst-group accuracy with the presence of severe spurious correlation and even degrades robustness sometimes, performing worse than models without adaptation. Test-Time Label Shift Adaptation (TTLSA) [47] addresses spurious label shifts by adapting predictions to the target distribution through calibration on the joint distribution of the target and spurious labels, which corresponds to the differences between the training and target sets. However, this method requires the training set annotations, which are often unavailable given the large-scale and diverse training sources of CLIP models, making it challenging to apply in zero-shot classification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.424 + ], + "angle": 0, + "content": "A critical yet underexplored question remains: Can unknown spurious correlations be mitigated in zero-shot CLIP predictions without access to the information of training statistics or spurious attributes?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.483, + 0.679 + ], + "angle": 0, + "content": "In this work, we set out to understand spurious correlations in CLIP models. We observe that, given a prompt, CLIP models allocate significant attention to spurious features (e.g., background), as shown in the left panel of Figure 1. Motivated by this observation, we conduct a semantic analysis of CLIP's image encoder. Surprisingly, we find that a pre-trained CLIP model exhibits an emergent object detection capability, allowing it to recognize objects in test images. Building on this insight, we propose test-time reasoning, an approach that mimics human perception by first localizing the core object and then reducing the influence of spurious features to perform zero-shot classification. The proposed approach operates in a purely test-time setting and only requires the test samples and the corresponding text prompt for the classification task, without requiring additional information on spurious attributes or training statistics. We summarize our contribution as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.681, + 0.483, + 0.756 + ], + "angle": 0, + "content": "- We discover that the CLIP model, a text-guided training paradigm, acquires semantic understanding and demonstrates a strong capability to identify objects. Notably, it can segment objects effectively even without a dedicated segmentation training procedure." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.767, + 0.483, + 0.829 + ], + "angle": 0, + "content": "- We propose an interpretable and easy-to-use approach named Test-Time Reasoning (TTR) to mitigate spurious correlation in zero-shot classification with CLIP without requiring expertise or annotations of spurious attributes." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.838, + 0.484, + 0.898 + ], + "angle": 0, + "content": "- We demonstrated the effectiveness of TTR via comprehensive experiments, achieving a \\(5.05\\%\\) improvement in worst-group accuracy over the leading existing method that relies on spurious information." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.681, + 0.484, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.114, + 0.791, + 0.13 + ], + "angle": 0, + "content": "2.1. Spurious correlation mitigation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Spurious correlation refers to the context-dependent clues that are learned by the models during training but might be broken or not exist in test data [3, 13, 39]. There have been various works in spurious correlation mitigation, requiring or without requiring group annotation. We categorize these methods according to their focus on the data or optimization aspects. The data-based methods [11, 16, 18, 25, 40, 58, 61] rely on data balancing or augmentation to improve the robustness against spurious correlation, while optimization-based methods [3, 20, 26, 33, 39, 53] aim to learn more invariant or robust feature representations during model training. Despite the effectiveness, these methods have primarily been studied in architectures like ResNet and often rely on retraining models, making it challenging to directly integrate them into large-scale pretrained models such as CLIP. GC-CLIP [41] improves zero-shot CLIP classification of small objects by cropping images with an auxiliary model. In contrast, our method focuses on spurious correlation and localizes objects using CLIP's own semantic cues without external supervision, making it more broadly applicable." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.451, + 0.871, + 0.467 + ], + "angle": 0, + "content": "2.2. Spurious correlation mitigation for VLMs" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.474, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Similar to ResNet and other vision-based models, VLMs also suffer from performance degradation from spurious correlation. However, VLMs, known for their large scale and multimodality, bring new challenges for spurious correlation mitigation. Current work can be mainly categorized depending on whether it modifies the original models. Several methods [15, 56] utilize contrastive learning to fine-tune VLMs with spurious attributes either through text descriptions or explicit detection. Additionally, VLMs' multimodal capabilities can be used to generate supervisory signals, such as saliency maps, which then guide the fine-tuning process [37]. In contrast, zero-shot approaches avoid model modification by contextual prompt engineering, such as Perception CLIP [2], or projection-based debiasing, including Orth-Cali [8], ROBOSHOT [1], and TIE* [31]. More details on the methods can be found in Sec. 9." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.727, + 0.707, + 0.743 + ], + "angle": 0, + "content": "2.3. Test time adaptation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "There is growing research interest in Test Time Adaptation (TTA), which adapts trained models at inference time using unlabeled test samples. TTA closely aligns with the scope of this work, as adapting zero-shot classification tasks also involves unlabeled data. A major branch of TTA is Test Time Training (TTT) [4, 28, 47]. TTT modifies training losses to incorporate self-supervised contrastive objectives, then reapplies these objectives to test samples, enhancing adaptation to unseen data. TTT is challenging to apply in zero-shot prediction since the training procedure" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2920" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.272 + ], + "angle": 0, + "content": "cannot be modified in this setting. Another category of TTA approaches focuses on minimizing prediction entropy [32, 34, 50, 60]. However, minimizing entropy fails to address spurious correlations because of the inherent group imbalance in the dataset [22]. Test Time Label Shift Adaptation (TTLSA) [46] addresses spurious correlations via test-time adaptation. However, this method requires training set annotations to detect label shifts during testing. Applying TTLSA to CLIP models is challenging due to the difficulty in obtaining spurious correlation information from the training data, given the complexity inherent in foundation model datasets." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.289, + 0.431, + 0.305 + ], + "angle": 0, + "content": "3. Problem formulation and observation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.311, + 0.483, + 0.476 + ], + "angle": 0, + "content": "In this section, we investigate why the CLIP model exhibits less robustness in the presence of spurious features. Our analysis reveals that spurious features receive disproportionately higher attention weights compared to core features. This misallocated attention contributes to the model's decreased robustness in zero-shot classification. We further analyze the semantic representations learned by CLIP and reveal its remarkable ability to identify core objects in images. This powerful capability can be harnessed to enhance the robustness of zero-shot classification, offering new insights into CLIP's potential." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.613 + ], + "angle": 0, + "content": "To understand the poor worst-group performance of CLIP on datasets with spurious correlations, we hypothesize two possible explanations: (1) CLIP model fails to localize the core object due to the influence of spurious features, or (2) CLIP model fails to correctly identify the core object. To test these hypotheses, we conduct experiments on the well-established Waterbirds dataset [39], where the bird serves as the core feature and the background as the spurious feature. A detailed introduction is in Section 9." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Spurious feature dominates attention weights. We first investigate the wrong attention allocation hypothesis. Darcet et al. [10] report that final attention maps highlight semantically meaningful regions, Simeoni et al. [44] utilize attention maps for object detection. Additionally, Chefer et al. [6] developed an explanation tool to analyze how vision transformers allocate attention. These works suggest that attention maps provide strong evidence for analyzing transformer-based models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Building on these insights, we apply the attention-based explanation method from Chefer et al. [6] to CLIP ViT/L-14 and ViT/B-32 [7], visualizing the attention map using [CLS] tokens for image embeddings, as shown in Figure 1. We experiment with transformers of varying sizes, as [44] reports that large-scale models potentially show outlier artifacts, whereas base models show fewer effects. Across both models, we observe that attention maps predominantly highlight background regions, indicating that CLIP models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "lack robustness in the presence of spurious features, failing to accurately localize the core features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.133, + 0.907, + 0.376 + ], + "angle": 0, + "content": "CLIP accurately classifies the core object in a spurious feature-free setting. We then validate the CLIP model's zero-shot ability to classify core features. To eliminate the effect of spurious features on performance, we construct the Core Object Only datasets. This dataset consists solely of birds, ensuring that the bird objects are identical to those in the Waterbirds original test set. The backgrounds are uniformly black (BLK) or white (Wht). For baseline comparison, we create the Original Variant (OV) dataset, where the core objects remain identical to those in the Core Object Only datasets, but spurious features (backgrounds) are reintroduced. We provide the details of dataset preparation and experimental settings in Section 7. We test the zero-shot classification with the same text prompts with the original Waterbirds test set and the Core Object Only sets with ViT/L-14 backbone [7]. The results are shown in Table 1." + }, + { + "type": "table_caption", + "bbox": [ + 0.529, + 0.384, + 0.89, + 0.397 + ], + "angle": 0, + "content": "Table 1. Zero-shot results on a variant of waterbirds datasets." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.403, + 0.887, + 0.46 + ], + "angle": 0, + "content": "
OVCore BlkCore Wht
WGAvgWGAvgWGAvg
34.1387.6291.3992.3790.9693.02
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.906, + 0.572 + ], + "angle": 0, + "content": "We observe that the worst-group accuracy significantly improves by \\(57.26\\%\\) and \\(56.83\\%\\) compared to the dataset containing spurious features. This significant improvement indicates the CLIP model's capability to accurately classify the core object in an image. We conclude that, in the absence of spurious features, the CLIP model achieves strong performance in zero-shot classification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.659 + ], + "angle": 0, + "content": "Pre-trained CLIP acquired semantic meaning. We further analyze the image embeddings from the CLIP model, as they are deterministic. Although the text space is equally important, it is significantly impacted by text prompts, and optimal prompts are challenging to obtain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.87 + ], + "angle": 0, + "content": "To analyze the embedding structure, we extract image embeddings excluding the [CLS] token and represent them as \\(\\mathbf{z}_{img} \\in \\mathbb{R}^{N \\times d}\\), where \\(N\\) denotes the number of patches and \\(d\\) is the latent dimension. We apply Principal Component Analysis (PCA) to reduce the \\(d\\)-dimensional embeddings to three components, extracting the top principal components for each image and obtaining \\(\\mathbf{z}_{reduced} \\in \\mathbb{R}^{N \\times 3}\\). We map these three dimensions to RGB values for visualization [35], as shown in the middle row of Figure 2. In each plot, patches with similar hues indicate higher similarity. For instance, patches with pink hues show strong semantic grouping. This visualization demonstrates that the CLIP model effectively captures semantic representations in its latent space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Furthermore, we apply PCA to a batch of 10 images to extract shared information across the batch. K-means clus" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2921" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.091, + 0.462, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.234, + 0.485, + 0.306 + ], + "angle": 0, + "content": "Figure 2. Semantic analysis results. Middle: We visualize PCA features using different colors, where patches with similar hues indicate semantic similarity. Bottom: Semantic masks are shown in different colors. In the last column, multiple birds within a single image are accurately identified." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.455 + ], + "angle": 0, + "content": "tering is then performed on the PCA-reduced features to group image patches based on similarity. We set the number of clusters \\( k = 3 \\) for this experiment. Based on the clustering results, we generate masks to segment the object, as illustrated in Figure 2. The results show that the CLIP model effectively segments core objects, such as birds. Additional results on various datasets are presented in Figure 5, demonstrating that CLIP captures semantically meaningful representations across diverse datasets." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.485 + ], + "angle": 0, + "content": "4. TTR: Test Time robustification via reasoning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.598 + ], + "angle": 0, + "content": "The analysis in Section 3 reveals that the vanilla zero-shot CLIP model disproportionately allocates attention to spurious features, thereby compromising its generalizability. However, the study also validates that the CLIP model possesses strong semantic identification capabilities, enabling it to effectively recognize semantically meaningful objects in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.704 + ], + "angle": 0, + "content": "To mitigate the impact of spurious correlations in zero-shot CLIP models, we propose a test-time reasoning framework inspired by human visual cognition. Conceptually, the model first localizes the object before performing classification, aligning with the human perceptual process [5]. Before presenting our solution, we conduct a study on semantic identification and robustness enhancement." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.713, + 0.433, + 0.73 + ], + "angle": 0, + "content": "4.1. Improved semantic identification recipe" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.856 + ], + "angle": 0, + "content": "As the CLIP model demonstrates robust performance on images containing only core objects, we conduct the following study to improve core object identification. Specifically, we ablate the use of auxiliary models or the pure CLIP model to perform zero-shot core feature segmentation. Regarding the auxiliary models, we leverage state-of-the-art world models, including Grounding DINO (GDINO) [27] and the combination of Segment Anything (SAM) [21]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In terms of the pure CLIP model, we explore the following techniques: (1) dimension reduction, (2) high-dimensional clustering, and (3) a hybrid approach combin" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "ing dimension reduction with low-dimensional clustering. Due to the page limits, a detailed introduction to the experiment setting is provided in Section 8. We conduct experiments on the Waterbirds using the original test set. The proposed approaches are evaluated from multiple perspectives, including utility (worst-group accuracy (WG), average accuracy (Avg), and DICE score), test-time latency (measured in minutes), and peak GPU memory consumption (VRAM usage). The results are shown in Table 2. The visualization results for segmentation, object detection are shown in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Figure 3 presents a plot of the relationship between DICE scores and worst-group accuracy, illustrating that precise localization of the core object positively impacts prediction robustness. In other words, an improved core object identification algorithm directly enhances the robustness of CLIP's zero-shot predictions. From Table 2, we observe that GDINO improves worst-group accuracy. As it only provides bounding boxes, spurious features may still be retained alongside the core object. Notably, combining GDINO with SAM to accurately identify the core object and using only that object for zero-shot classification yields the highest robust performance with CLIP models. Under this approach, worst-group accuracy improves significantly and nearly reaches average accuracy. This result echoes the observation from Section 3 and indicates a uniform high performance across groups. However, this robustness comes at a substantial cost of higher latency and significant GPU memory requirements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.909, + 0.744 + ], + "angle": 0, + "content": "By contrast, K-means clustering on the image representations also elevates worst-group accuracy. This suggests that the learned representations inherently capture semantic meaning, as clustering algorithms can effectively group meaningful features and produce well-segmented results. Combining PCA with K-means provides the best trade-off between robustness and efficiency. Reducing features to a 3-dimensional space before clustering significantly improves worst-group accuracy while maintaining low latency. Furthermore, this approach does not require auxiliary models, achieving strong performance while maintaining low GPU memory consumption. As a resource-efficient method, it provides an effective balance between performance and computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.908, + 0.788 + ], + "angle": 0, + "content": "4.2. Core feature identification and Zero-Shot robustification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We now present our solution TTR for enhancing the robustness of zero-shot classification on CLIP. Based on the insights from Section 4.1, we utilize PCA for dimensionality reduction, followed by K-means clustering for object identification. As discussed in Section 3, processing a batch of images yields more robust results. However, we expect to enable inference on any number of samples. To address this" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2922" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.647, + 0.119 + ], + "angle": 0, + "content": "Table 2. Ablation study on different methods. Latency (Lat) is reported in minutes, and GPU memory usage (VRAM) is reported in gigabytes (GB). ref stands for the reference." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.128, + 0.64, + 0.287 + ], + "angle": 0, + "content": "
WGAvgDICELatVRAM
Zero-Shot47.3673.70-2.52.0
+GDINO66.80 (19.44 ↑)82.22 (8.52 ↑)-18.64.6
+GDINO+ SAM87.79 (40.43 ↑)88.18 (14.48 ↑)ref81.511.9
+PCA49.93 (2.57 ↑)71.83 (1.87 ↓)40.535.52.5
+TSNE60.11 (12.75 ↑)76.65 (2.95 ↑)38.7159.02.5
+UMAP36.01(11.35 ↓)63.41 (10.29 ↓)32.2121.52.5
+K-means (d = 768)69.43(22.07 ↑)82.24 (8.54 ↑)61.6315.02.0
+TSNE+K-means64.31(16.95 ↑)78.99 (5.29 ↑)42.4659.62.5
+PCA+K-means74.14 (26.78 ↑)87.21(13.51 ↑)65.486.02.5
" + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.093, + 0.903, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.255, + 0.906, + 0.283 + ], + "angle": 0, + "content": "Figure 3. Characteristics curve of CLIP for waterbirds dataset: DICE vs. WG" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.309, + 0.481, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.504 + ], + "angle": 0, + "content": "Figure 4. Visualization of core object identification for each method. GDINO provides bounding boxes. Other methods generate core object masks, which are shown in black. Identifiable artifacts are highlighted with an orange circle. We perform zero-shot classification based on the core object identified by each method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.517, + 0.483, + 0.759 + ], + "angle": 0, + "content": "limitation, we introduce a queue to store historical data. The queue length \\( l \\in R^{+} \\) is a hyperparameter. A more detailed discussion and experiment are presented in section 5.4. For each test image, we pop up the oldest image from the queue and append the new image. We perform PCA on the queue to find the semantic structure in each image. To improve the continuity of the feature recognition, we apply a Gaussian filter to smooth the PCA features before performing k-means clustering. We employ k-means clustering to image patches based on the smoothed PCA features. We evaluated alternative clustering methods, including Gaussian Mixture Models and hierarchical clustering, and found that K-means clustering yields the best results. In our experiment, we set all datasets with \\( k = 3 \\) for all datasets to ensure consistency and ease of use. An ablation study on the number of clusters is presented in Section 5.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Based on the clustering results, we generate segmentation masks and apply them to the original test images. To eliminate any influence from unmasked regions, we set them to zero, ensuring they do not contribute to the final prediction. This process produces \\( k \\) segmented images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To identify the segmentation containing the core object, we employ zero-shot classification. As this approach relies on text prompts, we introduce a prompt selection strategy that captures the core object of interest. Instead of using a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.493 + ], + "angle": 0, + "content": "specific prompt such as \"a photo of a waterbird,\" we adopt a more general prompt, \"a photo of a bird,\" which provides a higher-level abstraction of the query. Using the general text embeddings, we compute image embeddings for each segmentation and measure their inner product with the text prompt. Drawing insights from Section 3, where the core features can be accurately classified without the influence of spurious features, we consider the segment closest to the prompt as the one containing core objects and perform zero-shot classification solely on this segmentation. The TTR procedure does not require information about spurious features, highlighting its ease of use and broad applicability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.506, + 0.64, + 0.525 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.532, + 0.597, + 0.549 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.645 + ], + "angle": 0, + "content": "Dataset. We conduct a comprehensive evaluation on six well-established benchmark datasets commonly used in spurious correlation studies [1, 20, 30, 39, 54]. These datasets include Waterbirds [39], CelebA [29], Metashifts [24], UrbanCars [23], ISIC [9], and CXR [51]. A detailed introduction to these datasets is provided in Appendix 9." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.65, + 0.907, + 0.696 + ], + "angle": 0, + "content": "Baselines. We compare our method against baselines in two categories and a detailed introduction to each baseline is provided in Appendix 9:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.701, + 0.906, + 0.747 + ], + "angle": 0, + "content": "- Robustifying vision-language models: Orth-Cali [8], Perception CLIP [2], ROBOSHOT [1], TIE* [31], and BEND-VLM [14]." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.751, + 0.907, + 0.813 + ], + "angle": 0, + "content": "- Test-time adaptation methods: MeanShift for Test-time Augmentation (MTA) [59], Test-Time Low-Rank Adaptation (TTLRA) [17], and Test-Time Label-Shift Adaptation (TTLSA) [46]." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.701, + 0.907, + 0.813 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Backbone and Metrics. Lu et al. [31] demonstrate that ViT-L/14 [12] shows a significant gap between worst-group accuracy and average accuracy, highlighting the need for robustification in zero-shot prediction for CLIP ViT-L/14. Based on this observation, we employ CLIP ViT-L/14 [7]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2923" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "for general tasks. For ISIC and CXR datasets, we follow [1] and utilize Biomed CLIP [62] as the zero-shot prediction backbone. Additionally, we present results for various Vision Transformer architectures in Section 5.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.482, + 0.273 + ], + "angle": 0, + "content": "Following the protocol in the robust learning study [39], we report three key metrics: Worst-group accuracy (WG), Average accuracy (Avg), and the gap between these metrics (Gap). Ideally, both metrics should be high and close in value to ensure accurate and robust performance. We bold the best results and underline the second-best results. Detailed implementation and reproducible text prompts are provided in Section 10." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.286, + 0.322, + 0.303 + ], + "angle": 0, + "content": "5.2. Core object identification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.482, + 0.4 + ], + "angle": 0, + "content": "Based on our analysis in Section 3, high-quality core object identification significantly enhances robustness. To validate our approach TTR, we first assess its ability to correctly identify the core object. We visualize the segments with the highest similarity to the general text prompts across all test datasets, as shown in Figure 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Our observations indicate that across all evaluated datasets, TTR consistently identifies the core object in the test set successfully. Interestingly, in the CelebA dataset, the spurious feature (gender) is closely located near the core feature (hair color). TTR effectively distinguishes and segments the core feature, demonstrating CLIP's strong reasoning capabilities. For the Metashifts dataset, we observe that the color of a cat and a boot are similar (see the top-right corner of Metashifts in Figure 5). Despite this, TTR effectively segments the cat, further validating CLIP's capability for semantic understanding. For the CXR dataset, chest X-ray orientations vary in the test set (e.g., frontal vs. lateral views). TTR successfully extracts these view differences and highlights the lung region, ensuring robustness across different imaging perspectives. In the ISIC dataset, more than two spurious features exist, such as color patches, rulers, hairs, and bubbles [54]. However, TTR does not rely on these artifacts, as it focuses solely on segmenting the core feature while ensuring accurate identification of the lesion region with integrity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.715, + 0.218, + 0.73 + ], + "angle": 0, + "content": "5.3. Main result" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.482, + 0.874 + ], + "angle": 0, + "content": "The main results on general datasets are presented in Table 3. Additional requirements beyond the task-related text prompt for running each method are detailed in the table. \"Prompt \\( a \\)\" indicates the need for spurious text prompts, with most methods relying on large language models (LLMs) to provide spurious feature descriptions. The \"label \\( a \\)\" denotes that the method requires spurious labels for the test set. \"Train \\( y, a \\)\" indicates the requirement for accessing training set annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.332 + ], + "angle": 0, + "content": "TTR consistently achieves robust results. For the Waterbirds dataset, TTR improves worst-group performance by \\(5.3\\%\\) compared to the second-best method, despite requiring fewer resources. For the UrbanCars dataset, the zero-shot baseline exhibits a significant performance gap between WG and Avg, indicating a strong spurious correlation between cars and background. After implementing our method, this gap is reduced from \\(43.3\\%\\) to \\(22.3\\%\\), demonstrating its effectiveness in mitigating spurious correlations. We evaluate the average performance across the general dataset, where Orth-Cali, TIE*, and TTLSA achieve performance improvements. However, these methods require prior knowledge of spurious correlations. In contrast, TTR outperforms all other methods, exceeding the second-best approach by \\(5.05\\%\\) in worst-group accuracy while requiring fewer annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.905, + 0.438 + ], + "angle": 0, + "content": "TTR achieves higher average accuracy compared to benchmark methods, highlighting its strong capability to capture core object features and classify them accurately. The gap between average accuracy and worst-group accuracy serves as a measure of robustness. Across the average of general datasets, TTR shows the smallest gap, demonstrating its robustness in zero-shot prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.905, + 0.695 + ], + "angle": 0, + "content": "Table 4 reports the results for the high-stakes biomedical domain datasets. The results demonstrate that our method achieves the highest worst group accuracy while maintaining a small gap between WG and average accuracy. TTLSA achieves the highest average accuracy on both datasets. However, the large gap between WG and average accuracy is concerning in class-imbalanced test sets, as it suggests that the model predominantly predicts a single class, resulting in high overall accuracy while failing to perform well across all groups. In contrast, our method achieves a more balanced performance across different groups. While it may reduce average accuracy, this is often due to the removal of the strong spurious correlation (such as color patches in ISIC), which correlates with the majority class [55]. By encouraging the model to focus on semantically meaningful regions, our method improves reliability by avoiding short cues that might lead to clinically critical cases." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.704, + 0.664, + 0.719 + ], + "angle": 0, + "content": "5.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.905, + 0.801 + ], + "angle": 0, + "content": "In this section, we explore two types of ablation studies. The parameter ablation examines the effect of hyperparameters within TTR on the results. The structure ablation investigates the impact of different types of masks and variations in vision transformers on performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.807, + 0.702, + 0.821 + ], + "angle": 0, + "content": "5.4.1. Parameters ablation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The length of the queue We first examine the effect of queue length on performance. A queue stores historical samples, which can potentially enhance robustness. To analyze this, we conduct an ablation study on the Waterbirds dataset, varying the queue length from 0 (i.e., no queue) to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.383, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Code at https://github.com/lu876/TTR" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2924" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.173, + 0.092, + 0.825, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.374, + 0.706, + 0.387 + ], + "angle": 0, + "content": "Figure 5. Core object identification results across benchmark datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.315, + 0.403, + 0.68, + 0.416 + ], + "angle": 0, + "content": "Table 3. Zero-Shot classification result on the general dataset." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.422, + 0.907, + 0.635 + ], + "angle": 0, + "content": "
RequirementWaterbirdsCelebAMetashiftUrbanCarsAverage
WG↑Avg↑WG↑Avg↑WG↑Avg↑WG↑Avg↑WG↑Acc↑
Orth-CaliPrompt a68.5186.3176.1486.2481.7081.7024.4054.7062.6977.24
PerceptionCLIPPrompt a54.1286.7474.3180.3090.8191.6727.6056.7061.7178.85
ROBOSHOTPrompt a45.1764.4382.6185.5490.6492.6114.0052.7058.1173.82
TIE*Prompt a61.6078.9881.9884.2791.7592.9726.8052.4065.5377.16
BEND-VLMLabel a39.8253.3479.1284.6790.6492.051.2049.7052.6969.94
TTLSATrain {y, a}53.1585.4082.1084.3990.5292.2322.8052.3062.1478.58
MTANone41.4271.1979.1683.3489.8792.057.6051.2054.5174.45
TENTNone52.5975.7774.4490.4988.6891.405.2051.9055.2377.39
TTLRANone41.6471.7876.8981.9091.5093.354.8050.9153.7174.49
Zero-ShotNone47.3673.7073.3581.2089.5491.688.0051.3054.5674.47
TTR (Ours)None74.1487.2183.3386.0090.8591.1334.0056.3070.5880.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.646, + 0.483, + 0.659 + ], + "angle": 0, + "content": "Table 4. Zero-Shot classification results on ISIC and CXR datasets" + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.671, + 0.485, + 0.882 + ], + "angle": 0, + "content": "
ISICCXR
WG↑Avg↑Gap↓WG↑Avg↑Gap↓
Orth-Cali21.4372.5451.1142.1153.8511.74
PerceptionCLIP41.5552.7411.1915.7974.7959.00
ROBOSHOT53.3059.846.5441.6256.2314.61
TIE*61.1171.6810.5742.1173.9231.81
BEND-VLM46.8376.0529.2226.3265.3939.07
TTLSA48.4179.7731.3610.5380.7770.24
MTA44.4465.2120.7747.3761.5414.17
TENT51.2657.616.3521.0570.5149.46
TTLRA60.1866.996.8118.4279.4961.07
Zero-Shot42.2170.2128.0028.9055.2826.38
TTR (Ours)64.2868.614.3355.2663.257.99
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.648, + 0.907, + 0.754 + ], + "angle": 0, + "content": "30. We report both worst-group accuracy and average accuracy to assess the impact of different queue lengths, as shown in the left panel of Figure 6. The results clearly demonstrate that employing a queue mechanism significantly improves both worst-group and average accuracy. Additionally, performance remains relatively stable across queue lengths that are larger than 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The number of clusters We then analyze the effect of the number of clusters used to generate masks. A small number of clusters may cause core object features and spurious features to be grouped together, reducing the effectiveness of TTR. Conversely, an excessive number of clusters can fragment the core features, making them less intact and ultimately diminishing zero-shot performance. The results are shown in the right panel of Figure 6. When the number of clusters is set to 1, the method is equivalent to standard zero" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2925" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.091, + 0.287, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.091, + 0.48, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.231, + 0.483, + 0.259 + ], + "angle": 0, + "content": "Figure 6. Left: WG and Avg change w.r.t. the length of the queue. Right: WG and Avg change w.r.t. the number of cluster." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.27, + 0.484, + 0.348 + ], + "angle": 0, + "content": "shot classification. With 2 or 3 clusters, TTR significantly improves both worst-group and average accuracy. However, an excessively high number of clusters negatively impacts performance. Based on these findings, we recommend using 2 or 3 clusters to achieve optimal performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.353, + 0.266, + 0.367 + ], + "angle": 0, + "content": "5.4.2. Structure ablation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.484, + 0.524 + ], + "angle": 0, + "content": "Mask types Wang and Wang [52] demonstrated that the method used to mask task-unrelated features significantly impacts the final prediction. Since masking operations alter the distribution of test samples, choosing an appropriate masking strategy is crucial. The major masking methods include filling with zero, filling with the mean value of all masked patches, and applying Gaussian blurring to the masked region. To determine the most effective masking method for TTR in CLIP models, we conduct an ablation study comparing these approaches." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.537, + 0.287, + 0.632 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.536, + 0.482, + 0.631 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.687 + ], + "angle": 0, + "content": "Figure 7. Comparison of the effectiveness with different masking types applied for the region of task-unrelated features. Left: worst group accuracy; Right: average accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.849 + ], + "angle": 0, + "content": "The results from Figure 7 indicate that the zero-mask approach achieves the highest worst-group accuracy among all methods. We hypothesize that this is because zero masking effectively eliminates spurious feature information and disrupts reliance on spurious features more drastically. The effectiveness of zero masking is further supported by findings in Section 3, where a black-background version of the Waterbirds dataset also achieved the highest worst-group accuracy, aligning with our observations. Based on these results, we recommend using the zero-mask strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "More results on different CLIP models We extend TTR to additional CLIP models to evaluate its generalizability. Specifically, we apply TTR to ViT-B/32 and ViT-H/14 to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.273 + ], + "angle": 0, + "content": "assess its effectiveness across different scales of CLIP models. The results are presented in Figure 8, where we compare the baseline zero-shot performance of each architecture with the performance change after applying TTR. The findings clearly show that TTR improves worst-group accuracy across all tested datasets, demonstrating its adaptability to different vision transformer architectures. The right panel of Figure 8 illustrates the changes in average accuracy. While one dataset with ViT-L exhibits a slight performance decrease, all others show varying degrees of improvement. This suggests that TTR enhances generalizability in zero-shot prediction." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.287, + 0.713, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.287, + 0.902, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.4, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Figure 8. Ablation on different ViT architectures. The pink bar represents TTR improvement, the white bar indicates reduction. TTR enhances performance across various vision transformers." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.458, + 0.628, + 0.473 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.907, + 0.727 + ], + "angle": 0, + "content": "We introduce TTR, a test-time robustification method to effectively mitigate the influence of spurious correlation on CLIP for robust zero-shot classification, without requiring annotations or prior knowledge on spurious attributes. Motivated by human reasoning for recognition, we examined the actual zero-shot classification capacity of CLIP in the absence of spurious correlations and further proposed TTR on top of that. We reveal and harness the semantic understanding capability of the CLIP image encoder to localize the area of the objects for refined classification. Compared to the standard zero-shot scheme, TTR not only improves classification performance but also enhances the interpretability of predictions. We demonstrate the effectiveness and robustness of TTR, which significantly boosts both worst-group and average accuracy on real-world datasets with challenging spurious correlations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Limitation and Future direction. Despite mitigating spurious correlation effectively, TTR depends on CLIP's ability to segment the core features. When spurious attributes are abstract, such as texture, the segmentation can become infeasible, limiting TTR's applicability. Additionally, while we demonstrated its effectiveness with simple text prompts based on the classification task, the performance of TTR could be influenced by the quality of the prompts in complex scenarios. Addressing these challenges or exploring the applicability in other multimodal tasks can be promising avenues for future research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2926" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.171 + ], + "angle": 0, + "content": "This work was partially supported by the EMBRIO Institute, contract #2120200, a National Science Foundation (NSF) Biology Integration Institute, and NSF IIS #1955890, IIS #2146091, IIS #2345235." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.199, + 0.188, + 0.215 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.224, + 0.484, + 0.279 + ], + "angle": 0, + "content": "[1] Dyah Adila, Changho Shin, Linrong Cai, and Frederic Sala. Zero-shot robustification of zero-shot models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.281, + 0.484, + 0.35 + ], + "angle": 0, + "content": "[2] Bang An, Sicheng Zhu, Michael-Andrei Panaitescu-Liess, Chaithanya Kumar Mummadi, and Furong Huang. Perceptionclip: Visual classification by inferring and conditioning on contexts. In The Twelfth International Conference on Learning Representations. 2, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.352, + 0.483, + 0.393 + ], + "angle": 0, + "content": "[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.395, + 0.484, + 0.463 + ], + "angle": 0, + "content": "[4] Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In International Conference on Artificial Intelligence and Statistics, pages 3080-3090. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.465, + 0.484, + 0.506 + ], + "angle": 0, + "content": "[5] Irving Biederman. Recognition-by-components: a theory of human image understanding. Psychological review, 94(2): 115, 1987. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.484, + 0.576 + ], + "angle": 0, + "content": "[6] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 397-406, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.484, + 0.661 + ], + "angle": 0, + "content": "[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 3, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[8] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[9] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), pages 168-172. IEEE, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[10] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.224, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[11] Yihe Deng, Yu Yang, Baharan Mirzasoleiman, and Quanquan Gu. Robust learning with progressive data expansion against spurious correlation. Advances in neural information processing systems, 36:1390-1402, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.244 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[13] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[14] Walter Gerych, Haoran Zhang, Kimia Hamidieh, Eileen Pan, Maanas Sharma, Thomas Hartvigsen, and Marzyeh Ghassemi. Bendvlm: Test-time debiasing of vision-language embeddings. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.44 + ], + "angle": 0, + "content": "[15] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19338-19347, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.906, + 0.496 + ], + "angle": 0, + "content": "[16] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.906, + 0.564 + ], + "angle": 0, + "content": "[17] Raza Imam, Hanan Gani, Muhammad Huzaifa, and Karthik Nandakumar. Test-time low rank adaptation via confidence maximization for zero-shot generalization of vision-language models. arXiv preprint arXiv:2407.15913, 2024. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.567, + 0.906, + 0.621 + ], + "angle": 0, + "content": "[18] Taeuk Jang and Xiaoqian Wang. Difficulty-based sampling for debiased contrastive representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24039-24048, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.906, + 0.677 + ], + "angle": 0, + "content": "[19] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. In The Eleventh International Conference on Learning Representations. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[20] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. In The Eleventh International Conference on Learning Representations, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.906, + 0.803 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[22] Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. arXiv preprint arXiv:2403.07366, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2927" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20071-20082, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.204 + ], + "angle": 0, + "content": "[24] Weixin Liang and James Zou. Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. In International Conference on Learning Representations, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.484, + 0.285 + ], + "angle": 0, + "content": "[25] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[26] Sheng Liu, Xu Zhang, Nitesh Sekhar, Yue Wu, Prateek Singhal, and Carlos Fernandez-Granda. Avoiding spurious correlations via logit correction. In The Eleventh International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[28] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[29] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[30] Shenyu Lu, Junyi Chai, and Xiaoqian Wang. Neural collapse inspired debiased representation learning for min-max fairness. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2048-2059, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.484, + 0.677 + ], + "angle": 0, + "content": "[31] Shenyu Lu, Junyi Chai, and Xiaogian Wang. Mitigating spurious correlations in zero-shot multimodal models. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[32] Chaithanya Kumar Mummadi, Robin Hutmacher, Kilian Rambach, Evgeny Levinkov, Thomas Brox, and Jan Hendrik Metzen. Test-time adaptation to distribution shift by confidence maximization and input transformation. arXiv preprint arXiv:2106.14999, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[33] Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: Training debiased classifier from biased classifier. In Advances in Neural Information Processing Systems, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[34] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[35] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[36] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournaepau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[37] Suzanne Petryk, Lisa Dunlap, Keyan Nasseri, Joseph Gonzalez, Trevor Darrell, and Anna Rohrbach. On guiding visual attention with language specification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18092-18102, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[39] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.508 + ], + "angle": 0, + "content": "[40] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.578 + ], + "angle": 0, + "content": "[41] Piyapat Saranrittichai, Mauricio Munoz, Volker Fischer, and Chaithanya Kumar Mummadi. Zero-shot recognition with guided cropping. In ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[43] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA McDermott, Irene Y Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12):2176-2182, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[44] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[45] Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, and Christopher Ré. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems, 33: 19339-19352, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Qingyao Sun, Kevin P Murphy, Sayna Ebrahimi, and Alexander D'Amour. Beyond invariance: test-time label" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2928" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "shift adaptation for addressing \"spurious\" correlations. Advances in Neural Information Processing Systems, 36: 23789-23812, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[47] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[48] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[49] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[50] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[51] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[52] Yipei Wang and Xiaogian Wang. Benchmarking deletion metrics with the principled explanations. In *Forty-first International Conference on Machine Learning*, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[53] Yipei Wang and Xiaoqian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In International Conference on Artificial Intelligence and Statistics, pages 3745-3753. PMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[54] Shirley Wu, Mert Yuksekgonul, Linjun Zhang, and James Zou. Discover and cure: Concept-aware mitigation of spurious correlation. In International Conference on Machine Learning, pages 37765-37786. PMLR, 2023. 5, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[55] Xiwei Xuan, Ziquan Deng, Hsuan-Tien Lin, and Kwan-Liu Ma. Slim: Spuriousness mitigation with minimal human annotations. In European Conference on Computer Vision, pages 215–231. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[56] Yu Yang, Besmira Nushi, Hamid Palangi, and Baharan Mirzasoleiman. Mitigating spurious correlations in multimodal models during fine-tuning. In International Conference on Machine Learning, pages 39365-39379. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[57] Yuzhe Yang, Haoran Zhang, Dina Katabi, and Marzyeh Ghassemi. Change is hard: A closer look at subpopulation shift. arXiv preprint arXiv:2302.12254, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[58] Huaxiu Yao, Yu Wang, Sai Li, Linjun Zhang, Weixin Liang, James Zou, and Chelsea Finn. Improving out-of-distribution robustness via selective augmentation. In International Conference on Machine Learning, pages 25407-25437. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[59] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 23783-23793, 2024. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[60] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in neural information processing systems, 35:38629-38642, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[61] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Re. Correct-n-contrast: a contrastive approach for improving robustness to spurious correlations. In International Conference on Machine Learning, pages 26484–26516. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[62] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[63] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2929" + } + ] +] \ No newline at end of file diff --git a/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/87c172f1-4134-4138-9793-b4455c2d3c37_origin.pdf b/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/87c172f1-4134-4138-9793-b4455c2d3c37_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c77423aa5cebe801d110b16a083a1a8614eef7ee --- /dev/null +++ b/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/87c172f1-4134-4138-9793-b4455c2d3c37_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3252160040a7d0fb5ae6f30fc72f241480840f597b83e6a4fd7a0b935fe1dd66 +size 8589968 diff --git a/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/full.md b/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d24cc5f9560af79fda8a8ebb5beebc3a5343bb43 --- /dev/null +++ b/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/full.md @@ -0,0 +1,288 @@ +# Think Twice: Test-Time Reasoning for Robust CLIP Zero-Shot Classification + +Shenyu Lu Zhaoying Pan Xiaoqian Wang* Elmore Family School of Electrical and Computer Engineering Purdue University + +{lu876, pan433, joywang}@purdue.edu + +# Abstract + +Contrastive Language-Image Pre-training (CLIP) models exhibit intriguing properties, particularly in their zero-shot classification capability. However, the reliability of CLIP zero-shot classification is severely undermined by spurious correlations. Existing efforts to enhance the robustness of zero-shot CLIP models often rely on prior knowledge or annotations of spurious correlations, limiting real-world applicability due to the unavailability of such information. Alternative methods attempt to detect distribution shift at test time but require training statistics whose access is often restricted or computationally expensive. To address the challenges brought by spurious correlation under zero-shot settings, we propose a novel test-time reasoning approach. Our method, inspired by human recognition, localizes the object and refines the classification accordingly. The inherent capacity of CLIP for semantic understanding allows us to isolate the object of interest without auxiliary models. Zero-shot classification is then performed exclusively on the localized objects, effectively mitigating the influence of spurious correlation. The proposed approach is interpretable and flexible as it requires no spurious annotations or prior knowledge, making it widely applicable. The substantial improvements across multiple benchmark datasets validated the effectiveness of our approach. + +# 1. Introduction + +Contrastive Language-Image Pre-training (CLIP) [38] has achieved significant success as a powerful pre-training paradigm. It leverages a unified multimodal framework that integrates images and text to achieve strong generalizability. The zero-shot capability of CLIP allows it to classify unseen test samples, making it highly applicable in real-world scenarios. Despite its impressive generalizability, CLIP remains susceptible to spurious correlations, undermining the reliability of its zero-shot predictions. A spurious correla + +![](images/f27f81a57019aa5dd09b24b83ef4991dbfc47fcef1b022a75c01a057f21f8e5d.jpg) +Figure 1. Attention map visualization based on [CLS] for randomly selected samples. Left: Vanilla zero-shot prediction with different vision transformers. Right: Proposed method applied to the same vision transformers. + +![](images/d58c89fcb03e3876238a1c489e2a83b1ab9ec1920ef6490baac584ef12c820fa.jpg) + +tion occurs when a feature is correlated with the target label but lacks a causal relationship [19]. This issue is particularly critical in high-stakes settings where reliance on such correlations can lead to detrimental consequences. For instance, Seyyed-Kalantari et al. [43] found that pneumonia diagnosis systems using chest X-rays may mistakenly base their predictions on patient demographics rather than the actual medical conditions, highlighting the potential risks of these correlations. + +Existing works have attempted to address spurious correlations in zero-shot CLIP models, with many focusing on debiasing the text space. However, as noted by [31], these approaches require domain expertise to identify spurious features and carry the risk of causing misalignment with the image modality. Alternative methods, such as ROBOSHOT [1] and TIE* [31], propose robust solutions but inherently rely on prior identification and annotations of spurious features, which significantly restricts their applicability. As different spurious text prompts yield varying robustness outcomes, identifying the optimal spurious prompts remains challenging. More importantly, the effectiveness of these methods is often compromised in real-world scenarios where spurious features are not predefined, + +thereby limiting their practical deployment. + +Recent literature on Test-Time Adaptation explores methods to adapt models at test time using unlabeled test samples. While minimizing the entropy of the predictions has been shown to enhance test-time adaptation, Lee et al. [22] demonstrates that minimizing entropy fails to improve worst-group accuracy with the presence of severe spurious correlation and even degrades robustness sometimes, performing worse than models without adaptation. Test-Time Label Shift Adaptation (TTLSA) [47] addresses spurious label shifts by adapting predictions to the target distribution through calibration on the joint distribution of the target and spurious labels, which corresponds to the differences between the training and target sets. However, this method requires the training set annotations, which are often unavailable given the large-scale and diverse training sources of CLIP models, making it challenging to apply in zero-shot classification. + +A critical yet underexplored question remains: Can unknown spurious correlations be mitigated in zero-shot CLIP predictions without access to the information of training statistics or spurious attributes? + +In this work, we set out to understand spurious correlations in CLIP models. We observe that, given a prompt, CLIP models allocate significant attention to spurious features (e.g., background), as shown in the left panel of Figure 1. Motivated by this observation, we conduct a semantic analysis of CLIP's image encoder. Surprisingly, we find that a pre-trained CLIP model exhibits an emergent object detection capability, allowing it to recognize objects in test images. Building on this insight, we propose test-time reasoning, an approach that mimics human perception by first localizing the core object and then reducing the influence of spurious features to perform zero-shot classification. The proposed approach operates in a purely test-time setting and only requires the test samples and the corresponding text prompt for the classification task, without requiring additional information on spurious attributes or training statistics. We summarize our contribution as follows: + +- We discover that the CLIP model, a text-guided training paradigm, acquires semantic understanding and demonstrates a strong capability to identify objects. Notably, it can segment objects effectively even without a dedicated segmentation training procedure. +- We propose an interpretable and easy-to-use approach named Test-Time Reasoning (TTR) to mitigate spurious correlation in zero-shot classification with CLIP without requiring expertise or annotations of spurious attributes. +- We demonstrated the effectiveness of TTR via comprehensive experiments, achieving a $5.05\%$ improvement in worst-group accuracy over the leading existing method that relies on spurious information. + +# 2. Related Work + +# 2.1. Spurious correlation mitigation + +Spurious correlation refers to the context-dependent clues that are learned by the models during training but might be broken or not exist in test data [3, 13, 39]. There have been various works in spurious correlation mitigation, requiring or without requiring group annotation. We categorize these methods according to their focus on the data or optimization aspects. The data-based methods [11, 16, 18, 25, 40, 58, 61] rely on data balancing or augmentation to improve the robustness against spurious correlation, while optimization-based methods [3, 20, 26, 33, 39, 53] aim to learn more invariant or robust feature representations during model training. Despite the effectiveness, these methods have primarily been studied in architectures like ResNet and often rely on retraining models, making it challenging to directly integrate them into large-scale pretrained models such as CLIP. GC-CLIP [41] improves zero-shot CLIP classification of small objects by cropping images with an auxiliary model. In contrast, our method focuses on spurious correlation and localizes objects using CLIP's own semantic cues without external supervision, making it more broadly applicable. + +# 2.2. Spurious correlation mitigation for VLMs + +Similar to ResNet and other vision-based models, VLMs also suffer from performance degradation from spurious correlation. However, VLMs, known for their large scale and multimodality, bring new challenges for spurious correlation mitigation. Current work can be mainly categorized depending on whether it modifies the original models. Several methods [15, 56] utilize contrastive learning to fine-tune VLMs with spurious attributes either through text descriptions or explicit detection. Additionally, VLMs' multimodal capabilities can be used to generate supervisory signals, such as saliency maps, which then guide the fine-tuning process [37]. In contrast, zero-shot approaches avoid model modification by contextual prompt engineering, such as Perception CLIP [2], or projection-based debiasing, including Orth-Cali [8], ROBOSHOT [1], and TIE* [31]. More details on the methods can be found in Sec. 9. + +# 2.3. Test time adaptation + +There is growing research interest in Test Time Adaptation (TTA), which adapts trained models at inference time using unlabeled test samples. TTA closely aligns with the scope of this work, as adapting zero-shot classification tasks also involves unlabeled data. A major branch of TTA is Test Time Training (TTT) [4, 28, 47]. TTT modifies training losses to incorporate self-supervised contrastive objectives, then reapplies these objectives to test samples, enhancing adaptation to unseen data. TTT is challenging to apply in zero-shot prediction since the training procedure + +cannot be modified in this setting. Another category of TTA approaches focuses on minimizing prediction entropy [32, 34, 50, 60]. However, minimizing entropy fails to address spurious correlations because of the inherent group imbalance in the dataset [22]. Test Time Label Shift Adaptation (TTLSA) [46] addresses spurious correlations via test-time adaptation. However, this method requires training set annotations to detect label shifts during testing. Applying TTLSA to CLIP models is challenging due to the difficulty in obtaining spurious correlation information from the training data, given the complexity inherent in foundation model datasets. + +# 3. Problem formulation and observation + +In this section, we investigate why the CLIP model exhibits less robustness in the presence of spurious features. Our analysis reveals that spurious features receive disproportionately higher attention weights compared to core features. This misallocated attention contributes to the model's decreased robustness in zero-shot classification. We further analyze the semantic representations learned by CLIP and reveal its remarkable ability to identify core objects in images. This powerful capability can be harnessed to enhance the robustness of zero-shot classification, offering new insights into CLIP's potential. + +To understand the poor worst-group performance of CLIP on datasets with spurious correlations, we hypothesize two possible explanations: (1) CLIP model fails to localize the core object due to the influence of spurious features, or (2) CLIP model fails to correctly identify the core object. To test these hypotheses, we conduct experiments on the well-established Waterbirds dataset [39], where the bird serves as the core feature and the background as the spurious feature. A detailed introduction is in Section 9. + +Spurious feature dominates attention weights. We first investigate the wrong attention allocation hypothesis. Darcet et al. [10] report that final attention maps highlight semantically meaningful regions, Simeoni et al. [44] utilize attention maps for object detection. Additionally, Chefer et al. [6] developed an explanation tool to analyze how vision transformers allocate attention. These works suggest that attention maps provide strong evidence for analyzing transformer-based models. + +Building on these insights, we apply the attention-based explanation method from Chefer et al. [6] to CLIP ViT/L-14 and ViT/B-32 [7], visualizing the attention map using [CLS] tokens for image embeddings, as shown in Figure 1. We experiment with transformers of varying sizes, as [44] reports that large-scale models potentially show outlier artifacts, whereas base models show fewer effects. Across both models, we observe that attention maps predominantly highlight background regions, indicating that CLIP models + +lack robustness in the presence of spurious features, failing to accurately localize the core features. + +CLIP accurately classifies the core object in a spurious feature-free setting. We then validate the CLIP model's zero-shot ability to classify core features. To eliminate the effect of spurious features on performance, we construct the Core Object Only datasets. This dataset consists solely of birds, ensuring that the bird objects are identical to those in the Waterbirds original test set. The backgrounds are uniformly black (BLK) or white (Wht). For baseline comparison, we create the Original Variant (OV) dataset, where the core objects remain identical to those in the Core Object Only datasets, but spurious features (backgrounds) are reintroduced. We provide the details of dataset preparation and experimental settings in Section 7. We test the zero-shot classification with the same text prompts with the original Waterbirds test set and the Core Object Only sets with ViT/L-14 backbone [7]. The results are shown in Table 1. + +Table 1. Zero-shot results on a variant of waterbirds datasets. + +
OVCore BlkCore Wht
WGAvgWGAvgWGAvg
34.1387.6291.3992.3790.9693.02
+ +We observe that the worst-group accuracy significantly improves by $57.26\%$ and $56.83\%$ compared to the dataset containing spurious features. This significant improvement indicates the CLIP model's capability to accurately classify the core object in an image. We conclude that, in the absence of spurious features, the CLIP model achieves strong performance in zero-shot classification. + +Pre-trained CLIP acquired semantic meaning. We further analyze the image embeddings from the CLIP model, as they are deterministic. Although the text space is equally important, it is significantly impacted by text prompts, and optimal prompts are challenging to obtain. + +To analyze the embedding structure, we extract image embeddings excluding the [CLS] token and represent them as $\mathbf{z}_{img} \in \mathbb{R}^{N \times d}$ , where $N$ denotes the number of patches and $d$ is the latent dimension. We apply Principal Component Analysis (PCA) to reduce the $d$ -dimensional embeddings to three components, extracting the top principal components for each image and obtaining $\mathbf{z}_{reduced} \in \mathbb{R}^{N \times 3}$ . We map these three dimensions to RGB values for visualization [35], as shown in the middle row of Figure 2. In each plot, patches with similar hues indicate higher similarity. For instance, patches with pink hues show strong semantic grouping. This visualization demonstrates that the CLIP model effectively captures semantic representations in its latent space. + +Furthermore, we apply PCA to a batch of 10 images to extract shared information across the batch. K-means clus + +![](images/ad9746e1d401f4f74e3a18163837b9d9acfc00cef69d74f763c79b7d9eea2bde.jpg) +Figure 2. Semantic analysis results. Middle: We visualize PCA features using different colors, where patches with similar hues indicate semantic similarity. Bottom: Semantic masks are shown in different colors. In the last column, multiple birds within a single image are accurately identified. + +tering is then performed on the PCA-reduced features to group image patches based on similarity. We set the number of clusters $k = 3$ for this experiment. Based on the clustering results, we generate masks to segment the object, as illustrated in Figure 2. The results show that the CLIP model effectively segments core objects, such as birds. Additional results on various datasets are presented in Figure 5, demonstrating that CLIP captures semantically meaningful representations across diverse datasets. + +# 4. TTR: Test Time robustification via reasoning + +The analysis in Section 3 reveals that the vanilla zero-shot CLIP model disproportionately allocates attention to spurious features, thereby compromising its generalizability. However, the study also validates that the CLIP model possesses strong semantic identification capabilities, enabling it to effectively recognize semantically meaningful objects in the latent space. + +To mitigate the impact of spurious correlations in zero-shot CLIP models, we propose a test-time reasoning framework inspired by human visual cognition. Conceptually, the model first localizes the object before performing classification, aligning with the human perceptual process [5]. Before presenting our solution, we conduct a study on semantic identification and robustness enhancement. + +# 4.1. Improved semantic identification recipe + +As the CLIP model demonstrates robust performance on images containing only core objects, we conduct the following study to improve core object identification. Specifically, we ablate the use of auxiliary models or the pure CLIP model to perform zero-shot core feature segmentation. Regarding the auxiliary models, we leverage state-of-the-art world models, including Grounding DINO (GDINO) [27] and the combination of Segment Anything (SAM) [21]. + +In terms of the pure CLIP model, we explore the following techniques: (1) dimension reduction, (2) high-dimensional clustering, and (3) a hybrid approach combin + +ing dimension reduction with low-dimensional clustering. Due to the page limits, a detailed introduction to the experiment setting is provided in Section 8. We conduct experiments on the Waterbirds using the original test set. The proposed approaches are evaluated from multiple perspectives, including utility (worst-group accuracy (WG), average accuracy (Avg), and DICE score), test-time latency (measured in minutes), and peak GPU memory consumption (VRAM usage). The results are shown in Table 2. The visualization results for segmentation, object detection are shown in Figure 4. + +Figure 3 presents a plot of the relationship between DICE scores and worst-group accuracy, illustrating that precise localization of the core object positively impacts prediction robustness. In other words, an improved core object identification algorithm directly enhances the robustness of CLIP's zero-shot predictions. From Table 2, we observe that GDINO improves worst-group accuracy. As it only provides bounding boxes, spurious features may still be retained alongside the core object. Notably, combining GDINO with SAM to accurately identify the core object and using only that object for zero-shot classification yields the highest robust performance with CLIP models. Under this approach, worst-group accuracy improves significantly and nearly reaches average accuracy. This result echoes the observation from Section 3 and indicates a uniform high performance across groups. However, this robustness comes at a substantial cost of higher latency and significant GPU memory requirements. + +By contrast, K-means clustering on the image representations also elevates worst-group accuracy. This suggests that the learned representations inherently capture semantic meaning, as clustering algorithms can effectively group meaningful features and produce well-segmented results. Combining PCA with K-means provides the best trade-off between robustness and efficiency. Reducing features to a 3-dimensional space before clustering significantly improves worst-group accuracy while maintaining low latency. Furthermore, this approach does not require auxiliary models, achieving strong performance while maintaining low GPU memory consumption. As a resource-efficient method, it provides an effective balance between performance and computational efficiency. + +# 4.2. Core feature identification and Zero-Shot robustification. + +We now present our solution TTR for enhancing the robustness of zero-shot classification on CLIP. Based on the insights from Section 4.1, we utilize PCA for dimensionality reduction, followed by K-means clustering for object identification. As discussed in Section 3, processing a batch of images yields more robust results. However, we expect to enable inference on any number of samples. To address this + +Table 2. Ablation study on different methods. Latency (Lat) is reported in minutes, and GPU memory usage (VRAM) is reported in gigabytes (GB). ref stands for the reference. + +
WGAvgDICELatVRAM
Zero-Shot47.3673.70-2.52.0
+GDINO66.80 (19.44 ↑)82.22 (8.52 ↑)-18.64.6
+GDINO+ SAM87.79 (40.43 ↑)88.18 (14.48 ↑)ref81.511.9
+PCA49.93 (2.57 ↑)71.83 (1.87 ↓)40.535.52.5
+TSNE60.11 (12.75 ↑)76.65 (2.95 ↑)38.7159.02.5
+UMAP36.01(11.35 ↓)63.41 (10.29 ↓)32.2121.52.5
+K-means (d = 768)69.43(22.07 ↑)82.24 (8.54 ↑)61.6315.02.0
+TSNE+K-means64.31(16.95 ↑)78.99 (5.29 ↑)42.4659.62.5
+PCA+K-means74.14 (26.78 ↑)87.21(13.51 ↑)65.486.02.5
+ +![](images/252c439eb3c08548ca256f90cbeee47800670fd4ed69fc5e682c44b3202f49e9.jpg) +Figure 3. Characteristics curve of CLIP for waterbirds dataset: DICE vs. WG + +![](images/59cbc6ba9d649efe8d5efc7751a760a80e38f8358f0069d5c357c60e48417daf.jpg) +Figure 4. Visualization of core object identification for each method. GDINO provides bounding boxes. Other methods generate core object masks, which are shown in black. Identifiable artifacts are highlighted with an orange circle. We perform zero-shot classification based on the core object identified by each method. + +limitation, we introduce a queue to store historical data. The queue length $l \in R^{+}$ is a hyperparameter. A more detailed discussion and experiment are presented in section 5.4. For each test image, we pop up the oldest image from the queue and append the new image. We perform PCA on the queue to find the semantic structure in each image. To improve the continuity of the feature recognition, we apply a Gaussian filter to smooth the PCA features before performing k-means clustering. We employ k-means clustering to image patches based on the smoothed PCA features. We evaluated alternative clustering methods, including Gaussian Mixture Models and hierarchical clustering, and found that K-means clustering yields the best results. In our experiment, we set all datasets with $k = 3$ for all datasets to ensure consistency and ease of use. An ablation study on the number of clusters is presented in Section 5.4. + +Based on the clustering results, we generate segmentation masks and apply them to the original test images. To eliminate any influence from unmasked regions, we set them to zero, ensuring they do not contribute to the final prediction. This process produces $k$ segmented images. + +To identify the segmentation containing the core object, we employ zero-shot classification. As this approach relies on text prompts, we introduce a prompt selection strategy that captures the core object of interest. Instead of using a + +specific prompt such as "a photo of a waterbird," we adopt a more general prompt, "a photo of a bird," which provides a higher-level abstraction of the query. Using the general text embeddings, we compute image embeddings for each segmentation and measure their inner product with the text prompt. Drawing insights from Section 3, where the core features can be accurately classified without the influence of spurious features, we consider the segment closest to the prompt as the one containing core objects and perform zero-shot classification solely on this segmentation. The TTR procedure does not require information about spurious features, highlighting its ease of use and broad applicability. + +# 5. Experiment + +# 5.1. Setup + +Dataset. We conduct a comprehensive evaluation on six well-established benchmark datasets commonly used in spurious correlation studies [1, 20, 30, 39, 54]. These datasets include Waterbirds [39], CelebA [29], Metashifts [24], UrbanCars [23], ISIC [9], and CXR [51]. A detailed introduction to these datasets is provided in Appendix 9. + +Baselines. We compare our method against baselines in two categories and a detailed introduction to each baseline is provided in Appendix 9: + +- Robustifying vision-language models: Orth-Cali [8], Perception CLIP [2], ROBOSHOT [1], TIE* [31], and BEND-VLM [14]. +- Test-time adaptation methods: MeanShift for Test-time Augmentation (MTA) [59], Test-Time Low-Rank Adaptation (TTLRA) [17], and Test-Time Label-Shift Adaptation (TTLSA) [46]. + +Backbone and Metrics. Lu et al. [31] demonstrate that ViT-L/14 [12] shows a significant gap between worst-group accuracy and average accuracy, highlighting the need for robustification in zero-shot prediction for CLIP ViT-L/14. Based on this observation, we employ CLIP ViT-L/14 [7] + +for general tasks. For ISIC and CXR datasets, we follow [1] and utilize Biomed CLIP [62] as the zero-shot prediction backbone. Additionally, we present results for various Vision Transformer architectures in Section 5.4. + +Following the protocol in the robust learning study [39], we report three key metrics: Worst-group accuracy (WG), Average accuracy (Avg), and the gap between these metrics (Gap). Ideally, both metrics should be high and close in value to ensure accurate and robust performance. We bold the best results and underline the second-best results. Detailed implementation and reproducible text prompts are provided in Section 10. + +# 5.2. Core object identification + +Based on our analysis in Section 3, high-quality core object identification significantly enhances robustness. To validate our approach TTR, we first assess its ability to correctly identify the core object. We visualize the segments with the highest similarity to the general text prompts across all test datasets, as shown in Figure 5. + +Our observations indicate that across all evaluated datasets, TTR consistently identifies the core object in the test set successfully. Interestingly, in the CelebA dataset, the spurious feature (gender) is closely located near the core feature (hair color). TTR effectively distinguishes and segments the core feature, demonstrating CLIP's strong reasoning capabilities. For the Metashifts dataset, we observe that the color of a cat and a boot are similar (see the top-right corner of Metashifts in Figure 5). Despite this, TTR effectively segments the cat, further validating CLIP's capability for semantic understanding. For the CXR dataset, chest X-ray orientations vary in the test set (e.g., frontal vs. lateral views). TTR successfully extracts these view differences and highlights the lung region, ensuring robustness across different imaging perspectives. In the ISIC dataset, more than two spurious features exist, such as color patches, rulers, hairs, and bubbles [54]. However, TTR does not rely on these artifacts, as it focuses solely on segmenting the core feature while ensuring accurate identification of the lesion region with integrity. + +# 5.3. Main result + +The main results on general datasets are presented in Table 3. Additional requirements beyond the task-related text prompt for running each method are detailed in the table. "Prompt $a$ " indicates the need for spurious text prompts, with most methods relying on large language models (LLMs) to provide spurious feature descriptions. The "label $a$ " denotes that the method requires spurious labels for the test set. "Train $y, a$ " indicates the requirement for accessing training set annotations. + +TTR consistently achieves robust results. For the Waterbirds dataset, TTR improves worst-group performance by $5.3\%$ compared to the second-best method, despite requiring fewer resources. For the UrbanCars dataset, the zero-shot baseline exhibits a significant performance gap between WG and Avg, indicating a strong spurious correlation between cars and background. After implementing our method, this gap is reduced from $43.3\%$ to $22.3\%$ , demonstrating its effectiveness in mitigating spurious correlations. We evaluate the average performance across the general dataset, where Orth-Cali, TIE*, and TTLSA achieve performance improvements. However, these methods require prior knowledge of spurious correlations. In contrast, TTR outperforms all other methods, exceeding the second-best approach by $5.05\%$ in worst-group accuracy while requiring fewer annotations. + +TTR achieves higher average accuracy compared to benchmark methods, highlighting its strong capability to capture core object features and classify them accurately. The gap between average accuracy and worst-group accuracy serves as a measure of robustness. Across the average of general datasets, TTR shows the smallest gap, demonstrating its robustness in zero-shot prediction. + +Table 4 reports the results for the high-stakes biomedical domain datasets. The results demonstrate that our method achieves the highest worst group accuracy while maintaining a small gap between WG and average accuracy. TTLSA achieves the highest average accuracy on both datasets. However, the large gap between WG and average accuracy is concerning in class-imbalanced test sets, as it suggests that the model predominantly predicts a single class, resulting in high overall accuracy while failing to perform well across all groups. In contrast, our method achieves a more balanced performance across different groups. While it may reduce average accuracy, this is often due to the removal of the strong spurious correlation (such as color patches in ISIC), which correlates with the majority class [55]. By encouraging the model to focus on semantically meaningful regions, our method improves reliability by avoiding short cues that might lead to clinically critical cases. + +# 5.4. Ablation study + +In this section, we explore two types of ablation studies. The parameter ablation examines the effect of hyperparameters within TTR on the results. The structure ablation investigates the impact of different types of masks and variations in vision transformers on performance. + +# 5.4.1. Parameters ablation + +The length of the queue We first examine the effect of queue length on performance. A queue stores historical samples, which can potentially enhance robustness. To analyze this, we conduct an ablation study on the Waterbirds dataset, varying the queue length from 0 (i.e., no queue) to + +![](images/37532b170e90fd63135486d7b8e7f34e771ca5d7a66fa170dddde91689e98c45.jpg) +Figure 5. Core object identification results across benchmark datasets. + +Table 3. Zero-Shot classification result on the general dataset. + +
RequirementWaterbirdsCelebAMetashiftUrbanCarsAverage
WG↑Avg↑WG↑Avg↑WG↑Avg↑WG↑Avg↑WG↑Acc↑
Orth-CaliPrompt a68.5186.3176.1486.2481.7081.7024.4054.7062.6977.24
PerceptionCLIPPrompt a54.1286.7474.3180.3090.8191.6727.6056.7061.7178.85
ROBOSHOTPrompt a45.1764.4382.6185.5490.6492.6114.0052.7058.1173.82
TIE*Prompt a61.6078.9881.9884.2791.7592.9726.8052.4065.5377.16
BEND-VLMLabel a39.8253.3479.1284.6790.6492.051.2049.7052.6969.94
TTLSATrain {y, a}53.1585.4082.1084.3990.5292.2322.8052.3062.1478.58
MTANone41.4271.1979.1683.3489.8792.057.6051.2054.5174.45
TENTNone52.5975.7774.4490.4988.6891.405.2051.9055.2377.39
TTLRANone41.6471.7876.8981.9091.5093.354.8050.9153.7174.49
Zero-ShotNone47.3673.7073.3581.2089.5491.688.0051.3054.5674.47
TTR (Ours)None74.1487.2183.3386.0090.8591.1334.0056.3070.5880.16
+ +Table 4. Zero-Shot classification results on ISIC and CXR datasets + +
ISICCXR
WG↑Avg↑Gap↓WG↑Avg↑Gap↓
Orth-Cali21.4372.5451.1142.1153.8511.74
PerceptionCLIP41.5552.7411.1915.7974.7959.00
ROBOSHOT53.3059.846.5441.6256.2314.61
TIE*61.1171.6810.5742.1173.9231.81
BEND-VLM46.8376.0529.2226.3265.3939.07
TTLSA48.4179.7731.3610.5380.7770.24
MTA44.4465.2120.7747.3761.5414.17
TENT51.2657.616.3521.0570.5149.46
TTLRA60.1866.996.8118.4279.4961.07
Zero-Shot42.2170.2128.0028.9055.2826.38
TTR (Ours)64.2868.614.3355.2663.257.99
+ +30. We report both worst-group accuracy and average accuracy to assess the impact of different queue lengths, as shown in the left panel of Figure 6. The results clearly demonstrate that employing a queue mechanism significantly improves both worst-group and average accuracy. Additionally, performance remains relatively stable across queue lengths that are larger than 5. + +The number of clusters We then analyze the effect of the number of clusters used to generate masks. A small number of clusters may cause core object features and spurious features to be grouped together, reducing the effectiveness of TTR. Conversely, an excessive number of clusters can fragment the core features, making them less intact and ultimately diminishing zero-shot performance. The results are shown in the right panel of Figure 6. When the number of clusters is set to 1, the method is equivalent to standard zero + +![](images/8b9dc50fef27ab70310a165067bd0992ee6150e911bca4a22d66ba59545abac8.jpg) +Figure 6. Left: WG and Avg change w.r.t. the length of the queue. Right: WG and Avg change w.r.t. the number of cluster. + +![](images/6ea068f796e0181a081c48d50dce58fda23abcfcfbe6e18d82bc447f9a7965f1.jpg) + +shot classification. With 2 or 3 clusters, TTR significantly improves both worst-group and average accuracy. However, an excessively high number of clusters negatively impacts performance. Based on these findings, we recommend using 2 or 3 clusters to achieve optimal performance. + +# 5.4.2. Structure ablation + +Mask types Wang and Wang [52] demonstrated that the method used to mask task-unrelated features significantly impacts the final prediction. Since masking operations alter the distribution of test samples, choosing an appropriate masking strategy is crucial. The major masking methods include filling with zero, filling with the mean value of all masked patches, and applying Gaussian blurring to the masked region. To determine the most effective masking method for TTR in CLIP models, we conduct an ablation study comparing these approaches. + +![](images/45e5c7595e28323a0bcfe4571c7330b500a11f260a74cdd048c901632f6599f1.jpg) +Figure 7. Comparison of the effectiveness with different masking types applied for the region of task-unrelated features. Left: worst group accuracy; Right: average accuracy. + +![](images/4049c670da3c576a06d0ad4bc83fd957e56b73654fc2456483a0e347865716ff.jpg) + +The results from Figure 7 indicate that the zero-mask approach achieves the highest worst-group accuracy among all methods. We hypothesize that this is because zero masking effectively eliminates spurious feature information and disrupts reliance on spurious features more drastically. The effectiveness of zero masking is further supported by findings in Section 3, where a black-background version of the Waterbirds dataset also achieved the highest worst-group accuracy, aligning with our observations. Based on these results, we recommend using the zero-mask strategy. + +More results on different CLIP models We extend TTR to additional CLIP models to evaluate its generalizability. Specifically, we apply TTR to ViT-B/32 and ViT-H/14 to + +assess its effectiveness across different scales of CLIP models. The results are presented in Figure 8, where we compare the baseline zero-shot performance of each architecture with the performance change after applying TTR. The findings clearly show that TTR improves worst-group accuracy across all tested datasets, demonstrating its adaptability to different vision transformer architectures. The right panel of Figure 8 illustrates the changes in average accuracy. While one dataset with ViT-L exhibits a slight performance decrease, all others show varying degrees of improvement. This suggests that TTR enhances generalizability in zero-shot prediction. + +![](images/b6a993051bae7a1f3e7103430a1a94cdae2a619e6132e0e9abba0d667cfcc23b.jpg) +Figure 8. Ablation on different ViT architectures. The pink bar represents TTR improvement, the white bar indicates reduction. TTR enhances performance across various vision transformers. + +![](images/beb264dedb684bb623e55af61ac97c00c8864d0b0dbaaac27fb3c6015955d72e.jpg) + +# 6. Discussion + +We introduce TTR, a test-time robustification method to effectively mitigate the influence of spurious correlation on CLIP for robust zero-shot classification, without requiring annotations or prior knowledge on spurious attributes. Motivated by human reasoning for recognition, we examined the actual zero-shot classification capacity of CLIP in the absence of spurious correlations and further proposed TTR on top of that. We reveal and harness the semantic understanding capability of the CLIP image encoder to localize the area of the objects for refined classification. Compared to the standard zero-shot scheme, TTR not only improves classification performance but also enhances the interpretability of predictions. We demonstrate the effectiveness and robustness of TTR, which significantly boosts both worst-group and average accuracy on real-world datasets with challenging spurious correlations. + +Limitation and Future direction. Despite mitigating spurious correlation effectively, TTR depends on CLIP's ability to segment the core features. When spurious attributes are abstract, such as texture, the segmentation can become infeasible, limiting TTR's applicability. Additionally, while we demonstrated its effectiveness with simple text prompts based on the classification task, the performance of TTR could be influenced by the quality of the prompts in complex scenarios. Addressing these challenges or exploring the applicability in other multimodal tasks can be promising avenues for future research. + +# Acknowledgements + +# References + +[1] Dyah Adila, Changho Shin, Linrong Cai, and Frederic Sala. Zero-shot robustification of zero-shot models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 6, 3 +[2] Bang An, Sicheng Zhu, Michael-Andrei Panaitescu-Liess, Chaithanya Kumar Mummadi, and Furong Huang. Perceptionclip: Visual classification by inferring and conditioning on contexts. In The Twelfth International Conference on Learning Representations. 2, 5, 3 +[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2 +[4] Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In International Conference on Artificial Intelligence and Statistics, pages 3080-3090. PMLR, 2022. 2 +[5] Irving Biederman. Recognition-by-components: a theory of human image understanding. Psychological review, 94(2): 115, 1987. 4 +[6] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 397-406, 2021. 3 +[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 3, 5, 1 +[8] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2, 5, 3 +[9] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), pages 168-172. IEEE, 2018. 5 +[10] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 3 + +This work was partially supported by the EMBRIO Institute, contract #2120200, a National Science Foundation (NSF) Biology Integration Institute, and NSF IIS #1955890, IIS #2146091, IIS #2345235. +[11] Yihe Deng, Yu Yang, Baharan Mirzasoleiman, and Quanquan Gu. Robust learning with progressive data expansion against spurious correlation. Advances in neural information processing systems, 36:1390-1402, 2023. 2 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5, 1 +[13] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 2 +[14] Walter Gerych, Haoran Zhang, Kimia Hamidieh, Eileen Pan, Maanas Sharma, Thomas Hartvigsen, and Marzyeh Ghassemi. Bendvlm: Test-time debiasing of vision-language embeddings. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 5, 3 +[15] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19338-19347, 2023. 2 +[16] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. 2 +[17] Raza Imam, Hanan Gani, Muhammad Huzaifa, and Karthik Nandakumar. Test-time low rank adaptation via confidence maximization for zero-shot generalization of vision-language models. arXiv preprint arXiv:2407.15913, 2024. 5, 3 +[18] Taeuk Jang and Xiaoqian Wang. Difficulty-based sampling for debiased contrastive representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24039-24048, 2023. 2 +[19] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. In The Eleventh International Conference on Learning Representations. 1 +[20] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. In The Eleventh International Conference on Learning Representations, 2023. 2, 5 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 4, 1 +[22] Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. arXiv preprint arXiv:2403.07366, 2024. 2, 3 +[23] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and + +Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20071-20082, 2023. 5 +[24] Weixin Liang and James Zou. Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. In International Conference on Learning Representations, 2022. 5 +[25] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. 2 +[26] Sheng Liu, Xu Zhang, Nitesh Sekhar, Yue Wu, Prateek Singhal, and Carlos Fernandez-Granda. Avoiding spurious correlations via logit correction. In The Eleventh International Conference on Learning Representations, 2022. 2 +[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 4, 1 +[28] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 2 +[29] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5 +[30] Shenyu Lu, Junyi Chai, and Xiaoqian Wang. Neural collapse inspired debiased representation learning for min-max fairness. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2048-2059, 2024. 5 +[31] Shenyu Lu, Junyi Chai, and Xiaogian Wang. Mitigating spurious correlations in zero-shot multimodal models. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2, 5 +[32] Chaithanya Kumar Mummadi, Robin Hutmacher, Kilian Rambach, Evgeny Levinkov, Thomas Brox, and Jan Hendrik Metzen. Test-time adaptation to distribution shift by confidence maximization and input transformation. arXiv preprint arXiv:2106.14999, 2021. 3 +[33] Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: Training debiased classifier from biased classifier. In Advances in Neural Information Processing Systems, 2020. 2 +[34] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 3 +[35] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, + +Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3 +[36] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournaepau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830, 2011. 4 +[37] Suzanne Petryk, Lisa Dunlap, Keyan Nasseri, Joseph Gonzalez, Trevor Darrell, and Anna Rohrbach. On guiding visual attention with language specification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18092-18102, 2022. 2 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1 +[39] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 2, 3, 5, 6 +[40] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. 2 +[41] Piyapat Saranrittichai, Mauricio Munoz, Volker Fischer, and Chaithanya Kumar Mummadi. Zero-shot recognition with guided cropping. In ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models, 2024. 2 +[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 1 +[43] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA McDermott, Irene Y Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12):2176-2182, 2021. 1 +[44] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 3 +[45] Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, and Christopher Ré. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems, 33: 19339-19352, 2020. 1 +[46] Qingyao Sun, Kevin P Murphy, Sayna Ebrahimi, and Alexander D'Amour. Beyond invariance: test-time label + +shift adaptation for addressing "spurious" correlations. Advances in Neural Information Processing Systems, 36: 23789-23812, 2023. 3, 5 +[47] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2 +[48] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 1 +[49] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations. 3 +[50] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020.3 +[51] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 5 +[52] Yipei Wang and Xiaogian Wang. Benchmarking deletion metrics with the principled explanations. In *Forty-first International Conference on Machine Learning*, 2024. 8 +[53] Yipei Wang and Xiaoqian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In International Conference on Artificial Intelligence and Statistics, pages 3745-3753. PMLR, 2024. 2 +[54] Shirley Wu, Mert Yuksekgonul, Linjun Zhang, and James Zou. Discover and cure: Concept-aware mitigation of spurious correlation. In International Conference on Machine Learning, pages 37765-37786. PMLR, 2023. 5, 6, 2 +[55] Xiwei Xuan, Ziquan Deng, Hsuan-Tien Lin, and Kwan-Liu Ma. Slim: Spuriousness mitigation with minimal human annotations. In European Conference on Computer Vision, pages 215–231. Springer, 2024. 6 +[56] Yu Yang, Besmira Nushi, Hamid Palangi, and Baharan Mirzasoleiman. Mitigating spurious correlations in multimodal models during fine-tuning. In International Conference on Machine Learning, pages 39365-39379. PMLR, 2023. 2 +[57] Yuzhe Yang, Haoran Zhang, Dina Katabi, and Marzyeh Ghassemi. Change is hard: A closer look at subpopulation shift. arXiv preprint arXiv:2302.12254, 2023. 2 +[58] Huaxiu Yao, Yu Wang, Sai Li, Linjun Zhang, Weixin Liang, James Zou, and Chelsea Finn. Improving out-of-distribution robustness via selective augmentation. In International Conference on Machine Learning, pages 25407-25437. PMLR, 2022. 2 +[59] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 23783-23793, 2024. 5, 3 +[60] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in neural information processing systems, 35:38629-38642, 2022. 3 +[61] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Re. Correct-n-contrast: a contrastive approach for improving robustness to spurious correlations. In International Conference on Machine Learning, pages 26484–26516. PMLR, 2022. 2 +[62] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 6 +[63] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 1 \ No newline at end of file diff --git a/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/images.zip b/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..69a21320e36f8c5932b9ea0b5672a38f6f9323a6 --- /dev/null +++ b/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a30dbb19d996d2dfb10fc96c6b3de0c8148516c4c2fc1a755cc201b0a68c7ffb +size 597871 diff --git a/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/layout.json b/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f15f868890b3d24e3ae948bd5f71368b6dd9671d --- /dev/null +++ b/2025/Think Twice_ Test-Time Reasoning for Robust CLIP Zero-Shot Classification/layout.json @@ -0,0 +1,7293 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 543, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 543, + 120 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 543, + 120 + ], + "type": "text", + "content": "Think Twice: Test-Time Reasoning for Robust CLIP Zero-Shot Classification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 152, + 143, + 458, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 143, + 458, + 185 + ], + "spans": [ + { + "bbox": [ + 152, + 143, + 458, + 185 + ], + "type": "text", + "content": "Shenyu Lu Zhaoying Pan Xiaoqian Wang* Elmore Family School of Electrical and Computer Engineering Purdue University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 209, + 187, + 399, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 187, + 399, + 200 + ], + "spans": [ + { + "bbox": [ + 209, + 187, + 399, + 200 + ], + "type": "text", + "content": "{lu876, pan433, joywang}@purdue.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 252, + 297, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 297, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 297, + 540 + ], + "type": "text", + "content": "Contrastive Language-Image Pre-training (CLIP) models exhibit intriguing properties, particularly in their zero-shot classification capability. However, the reliability of CLIP zero-shot classification is severely undermined by spurious correlations. Existing efforts to enhance the robustness of zero-shot CLIP models often rely on prior knowledge or annotations of spurious correlations, limiting real-world applicability due to the unavailability of such information. Alternative methods attempt to detect distribution shift at test time but require training statistics whose access is often restricted or computationally expensive. To address the challenges brought by spurious correlation under zero-shot settings, we propose a novel test-time reasoning approach. Our method, inspired by human recognition, localizes the object and refines the classification accordingly. The inherent capacity of CLIP for semantic understanding allows us to isolate the object of interest without auxiliary models. Zero-shot classification is then performed exclusively on the localized objects, effectively mitigating the influence of spurious correlation. The proposed approach is interpretable and flexible as it requires no spurious annotations or prior knowledge, making it widely applicable. The substantial improvements across multiple benchmark datasets validated the effectiveness of our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 564, + 135, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 564, + 135, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 564, + 135, + 577 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 585, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 296, + 693 + ], + "type": "text", + "content": "Contrastive Language-Image Pre-training (CLIP) [38] has achieved significant success as a powerful pre-training paradigm. It leverages a unified multimodal framework that integrates images and text to achieve strong generalizability. The zero-shot capability of CLIP allows it to classify unseen test samples, making it highly applicable in real-world scenarios. Despite its impressive generalizability, CLIP remains susceptible to spurious correlations, undermining the reliability of its zero-shot predictions. A spurious correla" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 225, + 453, + 358 + ], + "blocks": [ + { + "bbox": [ + 328, + 225, + 453, + 358 + ], + "lines": [ + { + "bbox": [ + 328, + 225, + 453, + 358 + ], + "spans": [ + { + "bbox": [ + 328, + 225, + 453, + 358 + ], + "type": "image", + "image_path": "f27f81a57019aa5dd09b24b83ef4991dbfc47fcef1b022a75c01a057f21f8e5d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 367, + 555, + 410 + ], + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 410 + ], + "type": "text", + "content": "Figure 1. Attention map visualization based on [CLS] for randomly selected samples. Left: Vanilla zero-shot prediction with different vision transformers. Right: Proposed method applied to the same vision transformers." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 457, + 226, + 541, + 358 + ], + "blocks": [ + { + "bbox": [ + 457, + 226, + 541, + 358 + ], + "lines": [ + { + "bbox": [ + 457, + 226, + 541, + 358 + ], + "spans": [ + { + "bbox": [ + 457, + 226, + 541, + 358 + ], + "type": "image", + "image_path": "d58c89fcb03e3876238a1c489e2a83b1ab9ec1920ef6490baac584ef12c820fa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 430, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 538 + ], + "type": "text", + "content": "tion occurs when a feature is correlated with the target label but lacks a causal relationship [19]. This issue is particularly critical in high-stakes settings where reliance on such correlations can lead to detrimental consequences. For instance, Seyyed-Kalantari et al. [43] found that pneumonia diagnosis systems using chest X-rays may mistakenly base their predictions on patient demographics rather than the actual medical conditions, highlighting the potential risks of these correlations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "Existing works have attempted to address spurious correlations in zero-shot CLIP models, with many focusing on debiasing the text space. However, as noted by [31], these approaches require domain expertise to identify spurious features and carry the risk of causing misalignment with the image modality. Alternative methods, such as ROBOSHOT [1] and TIE* [31], propose robust solutions but inherently rely on prior identification and annotations of spurious features, which significantly restricts their applicability. As different spurious text prompts yield varying robustness outcomes, identifying the optimal spurious prompts remains challenging. More importantly, the effectiveness of these methods is often compromised in real-world scenarios where spurious features are not predefined," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2919" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 232, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 232, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 232, + 83 + ], + "type": "text", + "content": "thereby limiting their practical deployment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 286 + ], + "type": "text", + "content": "Recent literature on Test-Time Adaptation explores methods to adapt models at test time using unlabeled test samples. While minimizing the entropy of the predictions has been shown to enhance test-time adaptation, Lee et al. [22] demonstrates that minimizing entropy fails to improve worst-group accuracy with the presence of severe spurious correlation and even degrades robustness sometimes, performing worse than models without adaptation. Test-Time Label Shift Adaptation (TTLSA) [47] addresses spurious label shifts by adapting predictions to the target distribution through calibration on the joint distribution of the target and spurious labels, which corresponds to the differences between the training and target sets. However, this method requires the training set annotations, which are often unavailable given the large-scale and diverse training sources of CLIP models, making it challenging to apply in zero-shot classification." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 288, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 335 + ], + "type": "text", + "content": "A critical yet underexplored question remains: Can unknown spurious correlations be mitigated in zero-shot CLIP predictions without access to the information of training statistics or spurious attributes?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 336, + 295, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 537 + ], + "type": "text", + "content": "In this work, we set out to understand spurious correlations in CLIP models. We observe that, given a prompt, CLIP models allocate significant attention to spurious features (e.g., background), as shown in the left panel of Figure 1. Motivated by this observation, we conduct a semantic analysis of CLIP's image encoder. Surprisingly, we find that a pre-trained CLIP model exhibits an emergent object detection capability, allowing it to recognize objects in test images. Building on this insight, we propose test-time reasoning, an approach that mimics human perception by first localizing the core object and then reducing the influence of spurious features to perform zero-shot classification. The proposed approach operates in a purely test-time setting and only requires the test samples and the corresponding text prompt for the classification task, without requiring additional information on spurious attributes or training statistics. We summarize our contribution as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 539, + 296, + 711 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 539, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 295, + 598 + ], + "type": "text", + "content": "- We discover that the CLIP model, a text-guided training paradigm, acquires semantic understanding and demonstrates a strong capability to identify objects. Notably, it can segment objects effectively even without a dedicated segmentation training procedure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 607, + 295, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 607, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 607, + 295, + 656 + ], + "type": "text", + "content": "- We propose an interpretable and easy-to-use approach named Test-Time Reasoning (TTR) to mitigate spurious correlation in zero-shot classification with CLIP without requiring expertise or annotations of spurious attributes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 663, + 296, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 663, + 296, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 663, + 296, + 711 + ], + "type": "text", + "content": "- We demonstrated the effectiveness of TTR via comprehensive experiments, achieving a " + }, + { + "bbox": [ + 55, + 663, + 296, + 711 + ], + "type": "inline_equation", + "content": "5.05\\%" + }, + { + "bbox": [ + 55, + 663, + 296, + 711 + ], + "type": "text", + "content": " improvement in worst-group accuracy over the leading existing method that relies on spurious information." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 90, + 484, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 90, + 484, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 90, + 484, + 102 + ], + "type": "text", + "content": "2.1. Spurious correlation mitigation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 108, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 555, + 348 + ], + "type": "text", + "content": "Spurious correlation refers to the context-dependent clues that are learned by the models during training but might be broken or not exist in test data [3, 13, 39]. There have been various works in spurious correlation mitigation, requiring or without requiring group annotation. We categorize these methods according to their focus on the data or optimization aspects. The data-based methods [11, 16, 18, 25, 40, 58, 61] rely on data balancing or augmentation to improve the robustness against spurious correlation, while optimization-based methods [3, 20, 26, 33, 39, 53] aim to learn more invariant or robust feature representations during model training. Despite the effectiveness, these methods have primarily been studied in architectures like ResNet and often rely on retraining models, making it challenging to directly integrate them into large-scale pretrained models such as CLIP. GC-CLIP [41] improves zero-shot CLIP classification of small objects by cropping images with an auxiliary model. In contrast, our method focuses on spurious correlation and localizes objects using CLIP's own semantic cues without external supervision, making it more broadly applicable." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 357, + 533, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 357, + 533, + 369 + ], + "spans": [ + { + "bbox": [ + 314, + 357, + 533, + 369 + ], + "type": "text", + "content": "2.2. Spurious correlation mitigation for VLMs" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 375, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 567 + ], + "type": "text", + "content": "Similar to ResNet and other vision-based models, VLMs also suffer from performance degradation from spurious correlation. However, VLMs, known for their large scale and multimodality, bring new challenges for spurious correlation mitigation. Current work can be mainly categorized depending on whether it modifies the original models. Several methods [15, 56] utilize contrastive learning to fine-tune VLMs with spurious attributes either through text descriptions or explicit detection. Additionally, VLMs' multimodal capabilities can be used to generate supervisory signals, such as saliency maps, which then guide the fine-tuning process [37]. In contrast, zero-shot approaches avoid model modification by contextual prompt engineering, such as Perception CLIP [2], or projection-based debiasing, including Orth-Cali [8], ROBOSHOT [1], and TIE* [31]. More details on the methods can be found in Sec. 9." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 575, + 432, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 575, + 432, + 588 + ], + "spans": [ + { + "bbox": [ + 314, + 575, + 432, + 588 + ], + "type": "text", + "content": "2.3. Test time adaptation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "There is growing research interest in Test Time Adaptation (TTA), which adapts trained models at inference time using unlabeled test samples. TTA closely aligns with the scope of this work, as adapting zero-shot classification tasks also involves unlabeled data. A major branch of TTA is Test Time Training (TTT) [4, 28, 47]. TTT modifies training losses to incorporate self-supervised contrastive objectives, then reapplies these objectives to test samples, enhancing adaptation to unseen data. TTT is challenging to apply in zero-shot prediction since the training procedure" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2920" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 215 + ], + "type": "text", + "content": "cannot be modified in this setting. Another category of TTA approaches focuses on minimizing prediction entropy [32, 34, 50, 60]. However, minimizing entropy fails to address spurious correlations because of the inherent group imbalance in the dataset [22]. Test Time Label Shift Adaptation (TTLSA) [46] addresses spurious correlations via test-time adaptation. However, this method requires training set annotations to detect label shifts during testing. Applying TTLSA to CLIP models is challenging due to the difficulty in obtaining spurious correlation information from the training data, given the complexity inherent in foundation model datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 263, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 263, + 241 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 263, + 241 + ], + "type": "text", + "content": "3. Problem formulation and observation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 246, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 295, + 376 + ], + "type": "text", + "content": "In this section, we investigate why the CLIP model exhibits less robustness in the presence of spurious features. Our analysis reveals that spurious features receive disproportionately higher attention weights compared to core features. This misallocated attention contributes to the model's decreased robustness in zero-shot classification. We further analyze the semantic representations learned by CLIP and reveal its remarkable ability to identify core objects in images. This powerful capability can be harnessed to enhance the robustness of zero-shot classification, offering new insights into CLIP's potential." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 378, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 485 + ], + "type": "text", + "content": "To understand the poor worst-group performance of CLIP on datasets with spurious correlations, we hypothesize two possible explanations: (1) CLIP model fails to localize the core object due to the influence of spurious features, or (2) CLIP model fails to correctly identify the core object. To test these hypotheses, we conduct experiments on the well-established Waterbirds dataset [39], where the bird serves as the core feature and the background as the spurious feature. A detailed introduction is in Section 9." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 604 + ], + "type": "text", + "content": "Spurious feature dominates attention weights. We first investigate the wrong attention allocation hypothesis. Darcet et al. [10] report that final attention maps highlight semantically meaningful regions, Simeoni et al. [44] utilize attention maps for object detection. Additionally, Chefer et al. [6] developed an explanation tool to analyze how vision transformers allocate attention. These works suggest that attention maps provide strong evidence for analyzing transformer-based models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Building on these insights, we apply the attention-based explanation method from Chefer et al. [6] to CLIP ViT/L-14 and ViT/B-32 [7], visualizing the attention map using [CLS] tokens for image embeddings, as shown in Figure 1. We experiment with transformers of varying sizes, as [44] reports that large-scale models potentially show outlier artifacts, whereas base models show fewer effects. Across both models, we observe that attention maps predominantly highlight background regions, indicating that CLIP models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "lack robustness in the presence of spurious features, failing to accurately localize the core features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 105, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 555, + 297 + ], + "type": "text", + "content": "CLIP accurately classifies the core object in a spurious feature-free setting. We then validate the CLIP model's zero-shot ability to classify core features. To eliminate the effect of spurious features on performance, we construct the Core Object Only datasets. This dataset consists solely of birds, ensuring that the bird objects are identical to those in the Waterbirds original test set. The backgrounds are uniformly black (BLK) or white (Wht). For baseline comparison, we create the Original Variant (OV) dataset, where the core objects remain identical to those in the Core Object Only datasets, but spurious features (backgrounds) are reintroduced. We provide the details of dataset preparation and experimental settings in Section 7. We test the zero-shot classification with the same text prompts with the original Waterbirds test set and the Core Object Only sets with ViT/L-14 backbone [7]. The results are shown in Table 1." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 325, + 319, + 542, + 364 + ], + "blocks": [ + { + "bbox": [ + 323, + 304, + 544, + 314 + ], + "lines": [ + { + "bbox": [ + 323, + 304, + 544, + 314 + ], + "spans": [ + { + "bbox": [ + 323, + 304, + 544, + 314 + ], + "type": "text", + "content": "Table 1. Zero-shot results on a variant of waterbirds datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 319, + 542, + 364 + ], + "lines": [ + { + "bbox": [ + 325, + 319, + 542, + 364 + ], + "spans": [ + { + "bbox": [ + 325, + 319, + 542, + 364 + ], + "type": "table", + "html": "
OVCore BlkCore Wht
WGAvgWGAvgWGAvg
34.1387.6291.3992.3790.9693.02
", + "image_path": "ae125d6233d3821fdc56d05d25d8fd36d5838ab1cddb1f8e28d3fce2617a27fe.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 369, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 453 + ], + "type": "text", + "content": "We observe that the worst-group accuracy significantly improves by " + }, + { + "bbox": [ + 313, + 369, + 554, + 453 + ], + "type": "inline_equation", + "content": "57.26\\%" + }, + { + "bbox": [ + 313, + 369, + 554, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 369, + 554, + 453 + ], + "type": "inline_equation", + "content": "56.83\\%" + }, + { + "bbox": [ + 313, + 369, + 554, + 453 + ], + "type": "text", + "content": " compared to the dataset containing spurious features. This significant improvement indicates the CLIP model's capability to accurately classify the core object in an image. We conclude that, in the absence of spurious features, the CLIP model achieves strong performance in zero-shot classification." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 521 + ], + "type": "text", + "content": "Pre-trained CLIP acquired semantic meaning. We further analyze the image embeddings from the CLIP model, as they are deterministic. Although the text space is equally important, it is significantly impacted by text prompts, and optimal prompts are challenging to obtain." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "type": "text", + "content": "To analyze the embedding structure, we extract image embeddings excluding the [CLS] token and represent them as " + }, + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{img} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "type": "text", + "content": " denotes the number of patches and " + }, + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "type": "text", + "content": " is the latent dimension. We apply Principal Component Analysis (PCA) to reduce the " + }, + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "type": "text", + "content": "-dimensional embeddings to three components, extracting the top principal components for each image and obtaining " + }, + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{reduced} \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 313, + 522, + 554, + 689 + ], + "type": "text", + "content": ". We map these three dimensions to RGB values for visualization [35], as shown in the middle row of Figure 2. In each plot, patches with similar hues indicate higher similarity. For instance, patches with pink hues show strong semantic grouping. This visualization demonstrates that the CLIP model effectively captures semantic representations in its latent space." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": "Furthermore, we apply PCA to a batch of 10 images to extract shared information across the batch. K-means clus" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2921" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 72, + 282, + 178 + ], + "blocks": [ + { + "bbox": [ + 70, + 72, + 282, + 178 + ], + "lines": [ + { + "bbox": [ + 70, + 72, + 282, + 178 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 282, + 178 + ], + "type": "image", + "image_path": "ad9746e1d401f4f74e3a18163837b9d9acfc00cef69d74f763c79b7d9eea2bde.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 185, + 296, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 296, + 242 + ], + "type": "text", + "content": "Figure 2. Semantic analysis results. Middle: We visualize PCA features using different colors, where patches with similar hues indicate semantic similarity. Bottom: Semantic masks are shown in different colors. In the last column, multiple birds within a single image are accurately identified." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 360 + ], + "type": "text", + "content": "tering is then performed on the PCA-reduced features to group image patches based on similarity. We set the number of clusters " + }, + { + "bbox": [ + 55, + 251, + 296, + 360 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 55, + 251, + 296, + 360 + ], + "type": "text", + "content": " for this experiment. Based on the clustering results, we generate masks to segment the object, as illustrated in Figure 2. The results show that the CLIP model effectively segments core objects, such as birds. Additional results on various datasets are presented in Figure 5, demonstrating that CLIP captures semantically meaningful representations across diverse datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 296, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 384 + ], + "type": "text", + "content": "4. TTR: Test Time robustification via reasoning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 473 + ], + "type": "text", + "content": "The analysis in Section 3 reveals that the vanilla zero-shot CLIP model disproportionately allocates attention to spurious features, thereby compromising its generalizability. However, the study also validates that the CLIP model possesses strong semantic identification capabilities, enabling it to effectively recognize semantically meaningful objects in the latent space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 557 + ], + "type": "text", + "content": "To mitigate the impact of spurious correlations in zero-shot CLIP models, we propose a test-time reasoning framework inspired by human visual cognition. Conceptually, the model first localizes the object before performing classification, aligning with the human perceptual process [5]. Before presenting our solution, we conduct a study on semantic identification and robustness enhancement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 564, + 264, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 264, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 264, + 578 + ], + "type": "text", + "content": "4.1. Improved semantic identification recipe" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 677 + ], + "type": "text", + "content": "As the CLIP model demonstrates robust performance on images containing only core objects, we conduct the following study to improve core object identification. Specifically, we ablate the use of auxiliary models or the pure CLIP model to perform zero-shot core feature segmentation. Regarding the auxiliary models, we leverage state-of-the-art world models, including Grounding DINO (GDINO) [27] and the combination of Segment Anything (SAM) [21]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "In terms of the pure CLIP model, we explore the following techniques: (1) dimension reduction, (2) high-dimensional clustering, and (3) a hybrid approach combin" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "ing dimension reduction with low-dimensional clustering. Due to the page limits, a detailed introduction to the experiment setting is provided in Section 8. We conduct experiments on the Waterbirds using the original test set. The proposed approaches are evaluated from multiple perspectives, including utility (worst-group accuracy (WG), average accuracy (Avg), and DICE score), test-time latency (measured in minutes), and peak GPU memory consumption (VRAM usage). The results are shown in Table 2. The visualization results for segmentation, object detection are shown in Figure 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 205, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 419 + ], + "type": "text", + "content": "Figure 3 presents a plot of the relationship between DICE scores and worst-group accuracy, illustrating that precise localization of the core object positively impacts prediction robustness. In other words, an improved core object identification algorithm directly enhances the robustness of CLIP's zero-shot predictions. From Table 2, we observe that GDINO improves worst-group accuracy. As it only provides bounding boxes, spurious features may still be retained alongside the core object. Notably, combining GDINO with SAM to accurately identify the core object and using only that object for zero-shot classification yields the highest robust performance with CLIP models. Under this approach, worst-group accuracy improves significantly and nearly reaches average accuracy. This result echoes the observation from Section 3 and indicates a uniform high performance across groups. However, this robustness comes at a substantial cost of higher latency and significant GPU memory requirements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 422, + 556, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 556, + 589 + ], + "type": "text", + "content": "By contrast, K-means clustering on the image representations also elevates worst-group accuracy. This suggests that the learned representations inherently capture semantic meaning, as clustering algorithms can effectively group meaningful features and produce well-segmented results. Combining PCA with K-means provides the best trade-off between robustness and efficiency. Reducing features to a 3-dimensional space before clustering significantly improves worst-group accuracy while maintaining low latency. Furthermore, this approach does not require auxiliary models, achieving strong performance while maintaining low GPU memory consumption. As a resource-efficient method, it provides an effective balance between performance and computational efficiency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 599, + 555, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 555, + 624 + ], + "type": "text", + "content": "4.2. Core feature identification and Zero-Shot robustification." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "We now present our solution TTR for enhancing the robustness of zero-shot classification on CLIP. Based on the insights from Section 4.1, we utilize PCA for dimensionality reduction, followed by K-means clustering for object identification. As discussed in Section 3, processing a batch of images yields more robust results. However, we expect to enable inference on any number of samples. To address this" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2922" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 101, + 391, + 227 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 395, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 395, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 395, + 94 + ], + "type": "text", + "content": "Table 2. Ablation study on different methods. Latency (Lat) is reported in minutes, and GPU memory usage (VRAM) is reported in gigabytes (GB). ref stands for the reference." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 101, + 391, + 227 + ], + "lines": [ + { + "bbox": [ + 61, + 101, + 391, + 227 + ], + "spans": [ + { + "bbox": [ + 61, + 101, + 391, + 227 + ], + "type": "table", + "html": "
WGAvgDICELatVRAM
Zero-Shot47.3673.70-2.52.0
+GDINO66.80 (19.44 ↑)82.22 (8.52 ↑)-18.64.6
+GDINO+ SAM87.79 (40.43 ↑)88.18 (14.48 ↑)ref81.511.9
+PCA49.93 (2.57 ↑)71.83 (1.87 ↓)40.535.52.5
+TSNE60.11 (12.75 ↑)76.65 (2.95 ↑)38.7159.02.5
+UMAP36.01(11.35 ↓)63.41 (10.29 ↓)32.2121.52.5
+K-means (d = 768)69.43(22.07 ↑)82.24 (8.54 ↑)61.6315.02.0
+TSNE+K-means64.31(16.95 ↑)78.99 (5.29 ↑)42.4659.62.5
+PCA+K-means74.14 (26.78 ↑)87.21(13.51 ↑)65.486.02.5
", + "image_path": "859119eee0d0a4c11bf9838a2888cbb37f40cf9beac3d36781292927e70db60a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 405, + 73, + 552, + 191 + ], + "blocks": [ + { + "bbox": [ + 405, + 73, + 552, + 191 + ], + "lines": [ + { + "bbox": [ + 405, + 73, + 552, + 191 + ], + "spans": [ + { + "bbox": [ + 405, + 73, + 552, + 191 + ], + "type": "image", + "image_path": "252c439eb3c08548ca256f90cbeee47800670fd4ed69fc5e682c44b3202f49e9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 201, + 554, + 224 + ], + "lines": [ + { + "bbox": [ + 402, + 201, + 554, + 224 + ], + "spans": [ + { + "bbox": [ + 402, + 201, + 554, + 224 + ], + "type": "text", + "content": "Figure 3. Characteristics curve of CLIP for waterbirds dataset: DICE vs. WG" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 244, + 294, + 338 + ], + "blocks": [ + { + "bbox": [ + 57, + 244, + 294, + 338 + ], + "lines": [ + { + "bbox": [ + 57, + 244, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 57, + 244, + 294, + 338 + ], + "type": "image", + "image_path": "59cbc6ba9d649efe8d5efc7751a760a80e38f8358f0069d5c357c60e48417daf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 342, + 295, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 399 + ], + "type": "text", + "content": "Figure 4. Visualization of core object identification for each method. GDINO provides bounding boxes. Other methods generate core object masks, which are shown in black. Identifiable artifacts are highlighted with an orange circle. We perform zero-shot classification based on the core object identified by each method." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 409, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 409, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 54, + 409, + 295, + 601 + ], + "type": "text", + "content": "limitation, we introduce a queue to store historical data. The queue length " + }, + { + "bbox": [ + 54, + 409, + 295, + 601 + ], + "type": "inline_equation", + "content": "l \\in R^{+}" + }, + { + "bbox": [ + 54, + 409, + 295, + 601 + ], + "type": "text", + "content": " is a hyperparameter. A more detailed discussion and experiment are presented in section 5.4. For each test image, we pop up the oldest image from the queue and append the new image. We perform PCA on the queue to find the semantic structure in each image. To improve the continuity of the feature recognition, we apply a Gaussian filter to smooth the PCA features before performing k-means clustering. We employ k-means clustering to image patches based on the smoothed PCA features. We evaluated alternative clustering methods, including Gaussian Mixture Models and hierarchical clustering, and found that K-means clustering yields the best results. In our experiment, we set all datasets with " + }, + { + "bbox": [ + 54, + 409, + 295, + 601 + ], + "type": "inline_equation", + "content": "k = 3" + }, + { + "bbox": [ + 54, + 409, + 295, + 601 + ], + "type": "text", + "content": " for all datasets to ensure consistency and ease of use. An ablation study on the number of clusters is presented in Section 5.4." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 603, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 664 + ], + "type": "text", + "content": "Based on the clustering results, we generate segmentation masks and apply them to the original test images. To eliminate any influence from unmasked regions, we set them to zero, ensuring they do not contribute to the final prediction. This process produces " + }, + { + "bbox": [ + 55, + 603, + 295, + 664 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 603, + 295, + 664 + ], + "type": "text", + "content": " segmented images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "To identify the segmentation containing the core object, we employ zero-shot classification. As this approach relies on text prompts, we introduce a prompt selection strategy that captures the core object of interest. Instead of using a" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 245, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 390 + ], + "type": "text", + "content": "specific prompt such as \"a photo of a waterbird,\" we adopt a more general prompt, \"a photo of a bird,\" which provides a higher-level abstraction of the query. Using the general text embeddings, we compute image embeddings for each segmentation and measure their inner product with the text prompt. Drawing insights from Section 3, where the core features can be accurately classified without the influence of spurious features, we consider the segment closest to the prompt as the one containing core objects and perform zero-shot classification solely on this segmentation. The TTR procedure does not require information about spurious features, highlighting its ease of use and broad applicability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 400, + 391, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 400, + 391, + 415 + ], + "spans": [ + { + "bbox": [ + 314, + 400, + 391, + 415 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 421, + 365, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 365, + 434 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 365, + 434 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": "Dataset. We conduct a comprehensive evaluation on six well-established benchmark datasets commonly used in spurious correlation studies [1, 20, 30, 39, 54]. These datasets include Waterbirds [39], CelebA [29], Metashifts [24], UrbanCars [23], ISIC [9], and CXR [51]. A detailed introduction to these datasets is provided in Appendix 9." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 514, + 555, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 551 + ], + "type": "text", + "content": "Baselines. We compare our method against baselines in two categories and a detailed introduction to each baseline is provided in Appendix 9:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 555, + 555, + 643 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 555, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 555, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 314, + 555, + 554, + 591 + ], + "type": "text", + "content": "- Robustifying vision-language models: Orth-Cali [8], Perception CLIP [2], ROBOSHOT [1], TIE* [31], and BEND-VLM [14]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 594, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 555, + 643 + ], + "type": "text", + "content": "- Test-time adaptation methods: MeanShift for Test-time Augmentation (MTA) [59], Test-Time Low-Rank Adaptation (TTLRA) [17], and Test-Time Label-Shift Adaptation (TTLSA) [46]." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "Backbone and Metrics. Lu et al. [31] demonstrate that ViT-L/14 [12] shows a significant gap between worst-group accuracy and average accuracy, highlighting the need for robustification in zero-shot prediction for CLIP ViT-L/14. Based on this observation, we employ CLIP ViT-L/14 [7]" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2923" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "for general tasks. For ISIC and CXR datasets, we follow [1] and utilize Biomed CLIP [62] as the zero-shot prediction backbone. Additionally, we present results for various Vision Transformer architectures in Section 5.4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 294, + 216 + ], + "type": "text", + "content": "Following the protocol in the robust learning study [39], we report three key metrics: Worst-group accuracy (WG), Average accuracy (Avg), and the gap between these metrics (Gap). Ideally, both metrics should be high and close in value to ensure accurate and robust performance. We bold the best results and underline the second-best results. Detailed implementation and reproducible text prompts are provided in Section 10." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 226, + 197, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 197, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 197, + 239 + ], + "type": "text", + "content": "5.2. Core object identification" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 245, + 294, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 294, + 316 + ], + "type": "text", + "content": "Based on our analysis in Section 3, high-quality core object identification significantly enhances robustness. To validate our approach TTR, we first assess its ability to correctly identify the core object. We visualize the segments with the highest similarity to the general text prompts across all test datasets, as shown in Figure 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 317, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 295, + 556 + ], + "type": "text", + "content": "Our observations indicate that across all evaluated datasets, TTR consistently identifies the core object in the test set successfully. Interestingly, in the CelebA dataset, the spurious feature (gender) is closely located near the core feature (hair color). TTR effectively distinguishes and segments the core feature, demonstrating CLIP's strong reasoning capabilities. For the Metashifts dataset, we observe that the color of a cat and a boot are similar (see the top-right corner of Metashifts in Figure 5). Despite this, TTR effectively segments the cat, further validating CLIP's capability for semantic understanding. For the CXR dataset, chest X-ray orientations vary in the test set (e.g., frontal vs. lateral views). TTR successfully extracts these view differences and highlights the lung region, ensuring robustness across different imaging perspectives. In the ISIC dataset, more than two spurious features exist, such as color patches, rulers, hairs, and bubbles [54]. However, TTR does not rely on these artifacts, as it focuses solely on segmenting the core feature while ensuring accurate identification of the lesion region with integrity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 566, + 133, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 133, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 133, + 578 + ], + "type": "text", + "content": "5.3. Main result" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 584, + 294, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 294, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 294, + 692 + ], + "type": "text", + "content": "The main results on general datasets are presented in Table 3. Additional requirements beyond the task-related text prompt for running each method are detailed in the table. \"Prompt " + }, + { + "bbox": [ + 55, + 584, + 294, + 692 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 584, + 294, + 692 + ], + "type": "text", + "content": "\" indicates the need for spurious text prompts, with most methods relying on large language models (LLMs) to provide spurious feature descriptions. The \"label " + }, + { + "bbox": [ + 55, + 584, + 294, + 692 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 55, + 584, + 294, + 692 + ], + "type": "text", + "content": "\" denotes that the method requires spurious labels for the test set. \"Train " + }, + { + "bbox": [ + 55, + 584, + 294, + 692 + ], + "type": "inline_equation", + "content": "y, a" + }, + { + "bbox": [ + 55, + 584, + 294, + 692 + ], + "type": "text", + "content": "\" indicates the requirement for accessing training set annotations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": "TTR consistently achieves robust results. For the Waterbirds dataset, TTR improves worst-group performance by " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": " compared to the second-best method, despite requiring fewer resources. For the UrbanCars dataset, the zero-shot baseline exhibits a significant performance gap between WG and Avg, indicating a strong spurious correlation between cars and background. After implementing our method, this gap is reduced from " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "43.3\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "22.3\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": ", demonstrating its effectiveness in mitigating spurious correlations. We evaluate the average performance across the general dataset, where Orth-Cali, TIE*, and TTLSA achieve performance improvements. However, these methods require prior knowledge of spurious correlations. In contrast, TTR outperforms all other methods, exceeding the second-best approach by " + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "inline_equation", + "content": "5.05\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 262 + ], + "type": "text", + "content": " in worst-group accuracy while requiring fewer annotations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 264, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 346 + ], + "type": "text", + "content": "TTR achieves higher average accuracy compared to benchmark methods, highlighting its strong capability to capture core object features and classify them accurately. The gap between average accuracy and worst-group accuracy serves as a measure of robustness. Across the average of general datasets, TTR shows the smallest gap, demonstrating its robustness in zero-shot prediction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 347, + 553, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 553, + 550 + ], + "type": "text", + "content": "Table 4 reports the results for the high-stakes biomedical domain datasets. The results demonstrate that our method achieves the highest worst group accuracy while maintaining a small gap between WG and average accuracy. TTLSA achieves the highest average accuracy on both datasets. However, the large gap between WG and average accuracy is concerning in class-imbalanced test sets, as it suggests that the model predominantly predicts a single class, resulting in high overall accuracy while failing to perform well across all groups. In contrast, our method achieves a more balanced performance across different groups. While it may reduce average accuracy, this is often due to the removal of the strong spurious correlation (such as color patches in ISIC), which correlates with the majority class [55]. By encouraging the model to focus on semantically meaningful regions, our method improves reliability by avoiding short cues that might lead to clinically critical cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 557, + 406, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 557, + 406, + 569 + ], + "spans": [ + { + "bbox": [ + 314, + 557, + 406, + 569 + ], + "type": "text", + "content": "5.4. Ablation study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 574, + 553, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 553, + 634 + ], + "type": "text", + "content": "In this section, we explore two types of ablation studies. The parameter ablation examines the effect of hyperparameters within TTR on the results. The structure ablation investigates the impact of different types of masks and variations in vision transformers on performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 639, + 429, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 429, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 429, + 650 + ], + "type": "text", + "content": "5.4.1. Parameters ablation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "The length of the queue We first examine the effect of queue length on performance. A queue stores historical samples, which can potentially enhance robustness. To analyze this, we conduct an ablation study on the Waterbirds dataset, varying the queue length from 0 (i.e., no queue) to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 234, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 234, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 234, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 702, + 234, + 712 + ], + "type": "text", + "content": "Code at https://github.com/lu876/TTR" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2924" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 72, + 504, + 293 + ], + "blocks": [ + { + "bbox": [ + 105, + 72, + 504, + 293 + ], + "lines": [ + { + "bbox": [ + 105, + 72, + 504, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 72, + 504, + 293 + ], + "type": "image", + "image_path": "37532b170e90fd63135486d7b8e7f34e771ca5d7a66fa170dddde91689e98c45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 177, + 296, + 432, + 306 + ], + "lines": [ + { + "bbox": [ + 177, + 296, + 432, + 306 + ], + "spans": [ + { + "bbox": [ + 177, + 296, + 432, + 306 + ], + "type": "text", + "content": "Figure 5. Core object identification results across benchmark datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 334, + 555, + 502 + ], + "blocks": [ + { + "bbox": [ + 192, + 319, + 416, + 329 + ], + "lines": [ + { + "bbox": [ + 192, + 319, + 416, + 329 + ], + "spans": [ + { + "bbox": [ + 192, + 319, + 416, + 329 + ], + "type": "text", + "content": "Table 3. Zero-Shot classification result on the general dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 334, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 57, + 334, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 57, + 334, + 555, + 502 + ], + "type": "table", + "html": "
RequirementWaterbirdsCelebAMetashiftUrbanCarsAverage
WG↑Avg↑WG↑Avg↑WG↑Avg↑WG↑Avg↑WG↑Acc↑
Orth-CaliPrompt a68.5186.3176.1486.2481.7081.7024.4054.7062.6977.24
PerceptionCLIPPrompt a54.1286.7474.3180.3090.8191.6727.6056.7061.7178.85
ROBOSHOTPrompt a45.1764.4382.6185.5490.6492.6114.0052.7058.1173.82
TIE*Prompt a61.6078.9881.9884.2791.7592.9726.8052.4065.5377.16
BEND-VLMLabel a39.8253.3479.1284.6790.6492.051.2049.7052.6969.94
TTLSATrain {y, a}53.1585.4082.1084.3990.5292.2322.8052.3062.1478.58
MTANone41.4271.1979.1683.3489.8792.057.6051.2054.5174.45
TENTNone52.5975.7774.4490.4988.6891.405.2051.9055.2377.39
TTLRANone41.6471.7876.8981.9091.5093.354.8050.9153.7174.49
Zero-ShotNone47.3673.7073.3581.2089.5491.688.0051.3054.5674.47
TTR (Ours)None74.1487.2183.3386.0090.8591.1334.0056.3070.5880.16
", + "image_path": "8dd043401862b1d84161d19a1c2daacd8bb4a8fecf2a7a0e4b00023b48e44d37.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 531, + 296, + 698 + ], + "blocks": [ + { + "bbox": [ + 56, + 511, + 295, + 521 + ], + "lines": [ + { + "bbox": [ + 56, + 511, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 295, + 521 + ], + "type": "text", + "content": "Table 4. Zero-Shot classification results on ISIC and CXR datasets" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 531, + 296, + 698 + ], + "lines": [ + { + "bbox": [ + 56, + 531, + 296, + 698 + ], + "spans": [ + { + "bbox": [ + 56, + 531, + 296, + 698 + ], + "type": "table", + "html": "
ISICCXR
WG↑Avg↑Gap↓WG↑Avg↑Gap↓
Orth-Cali21.4372.5451.1142.1153.8511.74
PerceptionCLIP41.5552.7411.1915.7974.7959.00
ROBOSHOT53.3059.846.5441.6256.2314.61
TIE*61.1171.6810.5742.1173.9231.81
BEND-VLM46.8376.0529.2226.3265.3939.07
TTLSA48.4179.7731.3610.5380.7770.24
MTA44.4465.2120.7747.3761.5414.17
TENT51.2657.616.3521.0570.5149.46
TTLRA60.1866.996.8118.4279.4961.07
Zero-Shot42.2170.2128.0028.9055.2826.38
TTR (Ours)64.2868.614.3355.2663.257.99
", + "image_path": "785533a0f9461df7d56cd1d081515b61fc64cdf6bd562e773cc801241f6bd358.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 513, + 555, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 555, + 597 + ], + "type": "text", + "content": "30. We report both worst-group accuracy and average accuracy to assess the impact of different queue lengths, as shown in the left panel of Figure 6. The results clearly demonstrate that employing a queue mechanism significantly improves both worst-group and average accuracy. Additionally, performance remains relatively stable across queue lengths that are larger than 5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "The number of clusters We then analyze the effect of the number of clusters used to generate masks. A small number of clusters may cause core object features and spurious features to be grouped together, reducing the effectiveness of TTR. Conversely, an excessive number of clusters can fragment the core features, making them less intact and ultimately diminishing zero-shot performance. The results are shown in the right panel of Figure 6. When the number of clusters is set to 1, the method is equivalent to standard zero" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2925" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 175, + 171 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 175, + 171 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 175, + 171 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 175, + 171 + ], + "type": "image", + "image_path": "8b9dc50fef27ab70310a165067bd0992ee6150e911bca4a22d66ba59545abac8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 295, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 205 + ], + "type": "text", + "content": "Figure 6. Left: WG and Avg change w.r.t. the length of the queue. Right: WG and Avg change w.r.t. the number of cluster." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 72, + 293, + 171 + ], + "blocks": [ + { + "bbox": [ + 179, + 72, + 293, + 171 + ], + "lines": [ + { + "bbox": [ + 179, + 72, + 293, + 171 + ], + "spans": [ + { + "bbox": [ + 179, + 72, + 293, + 171 + ], + "type": "image", + "image_path": "6ea068f796e0181a081c48d50dce58fda23abcfcfbe6e18d82bc447f9a7965f1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 275 + ], + "type": "text", + "content": "shot classification. With 2 or 3 clusters, TTR significantly improves both worst-group and average accuracy. However, an excessively high number of clusters negatively impacts performance. Based on these findings, we recommend using 2 or 3 clusters to achieve optimal performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 279, + 162, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 162, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 162, + 290 + ], + "type": "text", + "content": "5.4.2. Structure ablation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 294, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 296, + 415 + ], + "type": "text", + "content": "Mask types Wang and Wang [52] demonstrated that the method used to mask task-unrelated features significantly impacts the final prediction. Since masking operations alter the distribution of test samples, choosing an appropriate masking strategy is crucial. The major masking methods include filling with zero, filling with the mean value of all masked patches, and applying Gaussian blurring to the masked region. To determine the most effective masking method for TTR in CLIP models, we conduct an ablation study comparing these approaches." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 425, + 175, + 500 + ], + "blocks": [ + { + "bbox": [ + 58, + 425, + 175, + 500 + ], + "lines": [ + { + "bbox": [ + 58, + 425, + 175, + 500 + ], + "spans": [ + { + "bbox": [ + 58, + 425, + 175, + 500 + ], + "type": "image", + "image_path": "45e5c7595e28323a0bcfe4571c7330b500a11f260a74cdd048c901632f6599f1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 510, + 296, + 544 + ], + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 544 + ], + "type": "text", + "content": "Figure 7. Comparison of the effectiveness with different masking types applied for the region of task-unrelated features. Left: worst group accuracy; Right: average accuracy." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 177, + 424, + 294, + 499 + ], + "blocks": [ + { + "bbox": [ + 177, + 424, + 294, + 499 + ], + "lines": [ + { + "bbox": [ + 177, + 424, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 177, + 424, + 294, + 499 + ], + "type": "image", + "image_path": "4049c670da3c576a06d0ad4bc83fd957e56b73654fc2456483a0e347865716ff.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 672 + ], + "type": "text", + "content": "The results from Figure 7 indicate that the zero-mask approach achieves the highest worst-group accuracy among all methods. We hypothesize that this is because zero masking effectively eliminates spurious feature information and disrupts reliance on spurious features more drastically. The effectiveness of zero masking is further supported by findings in Section 3, where a black-background version of the Waterbirds dataset also achieved the highest worst-group accuracy, aligning with our observations. Based on these results, we recommend using the zero-mask strategy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "More results on different CLIP models We extend TTR to additional CLIP models to evaluate its generalizability. Specifically, we apply TTR to ViT-B/32 and ViT-H/14 to" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "assess its effectiveness across different scales of CLIP models. The results are presented in Figure 8, where we compare the baseline zero-shot performance of each architecture with the performance change after applying TTR. The findings clearly show that TTR improves worst-group accuracy across all tested datasets, demonstrating its adaptability to different vision transformer architectures. The right panel of Figure 8 illustrates the changes in average accuracy. While one dataset with ViT-L exhibits a slight performance decrease, all others show varying degrees of improvement. This suggests that TTR enhances generalizability in zero-shot prediction." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 227, + 436, + 306 + ], + "blocks": [ + { + "bbox": [ + 318, + 227, + 436, + 306 + ], + "lines": [ + { + "bbox": [ + 318, + 227, + 436, + 306 + ], + "spans": [ + { + "bbox": [ + 318, + 227, + 436, + 306 + ], + "type": "image", + "image_path": "b6a993051bae7a1f3e7103430a1a94cdae2a619e6132e0e9abba0d667cfcc23b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 316, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 350 + ], + "type": "text", + "content": "Figure 8. Ablation on different ViT architectures. The pink bar represents TTR improvement, the white bar indicates reduction. TTR enhances performance across various vision transformers." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 439, + 227, + 552, + 306 + ], + "blocks": [ + { + "bbox": [ + 439, + 227, + 552, + 306 + ], + "lines": [ + { + "bbox": [ + 439, + 227, + 552, + 306 + ], + "spans": [ + { + "bbox": [ + 439, + 227, + 552, + 306 + ], + "type": "image", + "image_path": "beb264dedb684bb623e55af61ac97c00c8864d0b0dbaaac27fb3c6015955d72e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 362, + 384, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 362, + 384, + 374 + ], + "spans": [ + { + "bbox": [ + 314, + 362, + 384, + 374 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 383, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 575 + ], + "type": "text", + "content": "We introduce TTR, a test-time robustification method to effectively mitigate the influence of spurious correlation on CLIP for robust zero-shot classification, without requiring annotations or prior knowledge on spurious attributes. Motivated by human reasoning for recognition, we examined the actual zero-shot classification capacity of CLIP in the absence of spurious correlations and further proposed TTR on top of that. We reveal and harness the semantic understanding capability of the CLIP image encoder to localize the area of the objects for refined classification. Compared to the standard zero-shot scheme, TTR not only improves classification performance but also enhances the interpretability of predictions. We demonstrate the effectiveness and robustness of TTR, which significantly boosts both worst-group and average accuracy on real-world datasets with challenging spurious correlations." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "Limitation and Future direction. Despite mitigating spurious correlation effectively, TTR depends on CLIP's ability to segment the core features. When spurious attributes are abstract, such as texture, the segmentation can become infeasible, limiting TTR's applicability. Additionally, while we demonstrated its effectiveness with simple text prompts based on the classification task, the performance of TTR could be influenced by the quality of the prompts in complex scenarios. Addressing these challenges or exploring the applicability in other multimodal tasks can be promising avenues for future research." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2926" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 157, + 115, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 157, + 115, + 170 + ], + "spans": [ + { + "bbox": [ + 56, + 157, + 115, + 170 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 177, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 177, + 296, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 177, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 61, + 177, + 296, + 220 + ], + "type": "text", + "content": "[1] Dyah Adila, Changho Shin, Linrong Cai, and Frederic Sala. Zero-shot robustification of zero-shot models. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 5, 6, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 222, + 296, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 222, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 222, + 296, + 277 + ], + "type": "text", + "content": "[2] Bang An, Sicheng Zhu, Michael-Andrei Panaitescu-Liess, Chaithanya Kumar Mummadi, and Furong Huang. Perceptionclip: Visual classification by inferring and conditioning on contexts. In The Twelfth International Conference on Learning Representations. 2, 5, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 278, + 295, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 278, + 295, + 311 + ], + "spans": [ + { + "bbox": [ + 61, + 278, + 295, + 311 + ], + "type": "text", + "content": "[3] Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. arXiv preprint arXiv:1907.02893, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 312, + 296, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 312, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 61, + 312, + 296, + 366 + ], + "type": "text", + "content": "[4] Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In International Conference on Artificial Intelligence and Statistics, pages 3080-3090. PMLR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 368, + 296, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 368, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 368, + 296, + 400 + ], + "type": "text", + "content": "[5] Irving Biederman. Recognition-by-components: a theory of human image understanding. Psychological review, 94(2): 115, 1987. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 402, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 296, + 456 + ], + "type": "text", + "content": "[6] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 397-406, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 457, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 296, + 523 + ], + "type": "text", + "content": "[7] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2818-2829, 2023. 3, 5, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 525, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 296, + 567 + ], + "type": "text", + "content": "[8] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 2, 5, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 570, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 296, + 667 + ], + "type": "text", + "content": "[9] Noel CF Codella, David Gutman, M Emre Celebi, Brian Helba, Michael A Marchetti, Stephen W Dusza, Aadi Kalloo, Konstantinos Liopyris, Nabin Mishra, Harald Kittler, et al. Skin lesion analysis toward melanoma detection: A challenge at the 2017 international symposium on biomedical imaging (isbi), hosted by the international skin imaging collaboration (isic). In 2018 IEEE 15th international symposium on biomedical imaging (ISBI 2018), pages 168-172. IEEE, 2018. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[10] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 55, + 90, + 297, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 135 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 135 + ], + "type": "text", + "content": "This work was partially supported by the EMBRIO Institute, contract #2120200, a National Science Foundation (NSF) Biology Integration Institute, and NSF IIS #1955890, IIS #2146091, IIS #2345235." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[11] Yihe Deng, Yu Yang, Baharan Mirzasoleiman, and Quanquan Gu. Robust learning with progressive data expansion against spurious correlation. Advances in neural information processing systems, 36:1390-1402, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 193 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 193 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 5, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "text", + "content": "[13] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 293 + ], + "type": "text", + "content": "[14] Walter Gerych, Haoran Zhang, Kimia Hamidieh, Eileen Pan, Maanas Sharma, Thomas Hartvigsen, and Marzyeh Ghassemi. Bendvlm: Test-time debiasing of vision-language embeddings. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 5, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 294, + 554, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 348 + ], + "type": "text", + "content": "[15] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19338-19347, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 392 + ], + "type": "text", + "content": "[16] Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy. In Conference on Causal Learning and Reasoning, pages 336–351. PMLR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 394, + 554, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 554, + 446 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 554, + 446 + ], + "type": "text", + "content": "[17] Raza Imam, Hanan Gani, Muhammad Huzaifa, and Karthik Nandakumar. Test-time low rank adaptation via confidence maximization for zero-shot generalization of vision-language models. arXiv preprint arXiv:2407.15913, 2024. 5, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 449, + 554, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 554, + 491 + ], + "type": "text", + "content": "[18] Taeuk Jang and Xiaoqian Wang. Difficulty-based sampling for debiased contrastive representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24039-24048, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 493, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 554, + 536 + ], + "type": "text", + "content": "[19] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. In The Eleventh International Conference on Learning Representations. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 537, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 581 + ], + "type": "text", + "content": "[20] Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations. In The Eleventh International Conference on Learning Representations, 2023. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 582, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 554, + 635 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 4, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 636, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 554, + 689 + ], + "type": "text", + "content": "[22] Jonghyun Lee, Dahuin Jung, Saehyung Lee, Junsung Park, Juhyeon Shin, Uiwon Hwang, and Sungroh Yoon. Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. arXiv preprint arXiv:2403.07366, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[23] Zhiheng Li, Ivan Evtimov, Albert Gordo, Caner Hazirbas, Tal Hassner, Cristian Canton Ferrer, Chenliang Xu, and" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2927" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Mark Ibrahim. A whac-a-mole dilemma: Shortcuts come in multiples where mitigating one amplifies others. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20071-20082, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "type": "text", + "content": "[24] Weixin Liang and James Zou. Metashift: A dataset of datasets for evaluating contextual distribution shifts and training conflicts. In International Conference on Learning Representations, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 225 + ], + "type": "text", + "content": "[25] Evan Z Liu, Behzad Haghgoo, Annie S Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information. In International Conference on Machine Learning, pages 6781-6792. PMLR, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 271 + ], + "type": "text", + "content": "[26] Sheng Liu, Xu Zhang, Nitesh Sekhar, Yue Wu, Prateek Singhal, and Carlos Fernandez-Granda. Avoiding spurious correlations via logit correction. In The Eleventh International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 296, + 337 + ], + "type": "text", + "content": "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38-55. Springer, 2024. 4, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 392 + ], + "type": "text", + "content": "[28] Yuejiang Liu, Parth Kothari, Bastien Van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? Advances in Neural Information Processing Systems, 34: 21808-21820, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 296, + 435 + ], + "type": "text", + "content": "[29] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "type": "text", + "content": "[30] Shenyu Lu, Junyi Chai, and Xiaoqian Wang. Neural collapse inspired debiased representation learning for min-max fairness. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 2048-2059, 2024. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 296, + 536 + ], + "type": "text", + "content": "[31] Shenyu Lu, Junyi Chai, and Xiaogian Wang. Mitigating spurious correlations in zero-shot multimodal models. In The Thirteenth International Conference on Learning Representations, 2025. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 590 + ], + "type": "text", + "content": "[32] Chaithanya Kumar Mummadi, Robin Hutmacher, Kilian Rambach, Evgeny Levinkov, Thomas Brox, and Jan Hendrik Metzen. Test-time adaptation to distribution shift by confidence maximization and input transformation. arXiv preprint arXiv:2106.14999, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 635 + ], + "type": "text", + "content": "[33] Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: Training debiased classifier from biased classifier. In Advances in Neural Information Processing Systems, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 689 + ], + "type": "text", + "content": "[34] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In International conference on machine learning, pages 16888-16905. PMLR, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[35] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "text", + "content": "[36] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournaepau, M. Brucher, M. Perrot, and E. Duchesnay. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830, 2011. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 172, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 555, + 227 + ], + "type": "text", + "content": "[37] Suzanne Petryk, Lisa Dunlap, Keyan Nasseri, Joseph Gonzalez, Trevor Darrell, and Anna Rohrbach. On guiding visual attention with language specification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18092-18102, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 293 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 348 + ], + "type": "text", + "content": "[39] Shiori Sagawa, Pang Wei Koh, Tatsunori B Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization. arXiv preprint arXiv:1911.08731, 2019. 2, 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 402 + ], + "type": "text", + "content": "[40] Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, and Percy Liang. An investigation of why overparameterization exacerbates spurious correlations. In International Conference on Machine Learning, pages 8346-8356. PMLR, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 404, + 553, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 457 + ], + "type": "text", + "content": "[41] Piyapat Saranrittichai, Mauricio Munoz, Volker Fischer, and Chaithanya Kumar Mummadi. Zero-shot recognition with guided cropping. In ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 460, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 525 + ], + "type": "text", + "content": "[42] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 35:25278-25294, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 580 + ], + "type": "text", + "content": "[43] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA McDermott, Irene Y Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12):2176-2182, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[44] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. arXiv preprint arXiv:2109.14279, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "text", + "content": "[45] Nimit Sohoni, Jared Dunnmon, Geoffrey Angus, Albert Gu, and Christopher Ré. No subclass left behind: Fine-grained robustness in coarse-grained classification problems. Advances in Neural Information Processing Systems, 33: 19339-19352, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[46] Qingyao Sun, Kevin P Murphy, Sayna Ebrahimi, and Alexander D'Amour. Beyond invariance: test-time label" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2928" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "shift adaptation for addressing \"spurious\" correlations. Advances in Neural Information Processing Systems, 36: 23789-23812, 2023. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[47] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In International conference on machine learning, pages 9229-9248. PMLR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 196 + ], + "type": "text", + "content": "[48] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[49] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In International Conference on Learning Representations. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 285 + ], + "type": "text", + "content": "[50] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726, 2020.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 363 + ], + "type": "text", + "content": "[51] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 399 + ], + "type": "text", + "content": "[52] Yipei Wang and Xiaogian Wang. Benchmarking deletion metrics with the principled explanations. In *Forty-first International Conference on Machine Learning*, 2024. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "text", + "content": "[53] Yipei Wang and Xiaoqian Wang. On the effect of key factors in spurious correlation: A theoretical perspective. In International Conference on Artificial Intelligence and Statistics, pages 3745-3753. PMLR, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "type": "text", + "content": "[54] Shirley Wu, Mert Yuksekgonul, Linjun Zhang, and James Zou. Discover and cure: Concept-aware mitigation of spurious correlation. In International Conference on Machine Learning, pages 37765-37786. PMLR, 2023. 5, 6, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[55] Xiwei Xuan, Ziquan Deng, Hsuan-Tien Lin, and Kwan-Liu Ma. Slim: Spuriousness mitigation with minimal human annotations. In European Conference on Computer Vision, pages 215–231. Springer, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 588 + ], + "type": "text", + "content": "[56] Yu Yang, Besmira Nushi, Hamid Palangi, and Baharan Mirzasoleiman. Mitigating spurious correlations in multimodal models during fine-tuning. In International Conference on Machine Learning, pages 39365-39379. PMLR, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 624 + ], + "type": "text", + "content": "[57] Yuzhe Yang, Haoran Zhang, Dina Katabi, and Marzyeh Ghassemi. Change is hard: A closer look at subpopulation shift. arXiv preprint arXiv:2302.12254, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "type": "text", + "content": "[58] Huaxiu Yao, Yu Wang, Sai Li, Linjun Zhang, Weixin Liang, James Zou, and Chelsea Finn. Improving out-of-distribution robustness via selective augmentation. In International Conference on Machine Learning, pages 25407-25437. PMLR, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "text", + "content": "[59] Maxime Zanella and Ismail Ben Ayed. On the test-time zero-shot generalization of vision-language models: Do we really need prompt learning? In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 308 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 23783-23793, 2024. 5, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": "[60] Marvin Zhang, Sergey Levine, and Chelsea Finn. Memo: Test time robustness via adaptation and augmentation. Advances in neural information processing systems, 35:38629-38642, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "text", + "content": "[61] Michael Zhang, Nimit S Sohoni, Hongyang R Zhang, Chelsea Finn, and Christopher Re. Correct-n-contrast: a contrastive approach for improving robustness to spurious correlations. In International Conference on Machine Learning, pages 26484–26516. PMLR, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 261 + ], + "type": "text", + "content": "[62] Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 263, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 308 + ], + "type": "text", + "content": "[63] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 1" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2929" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/5102f011-579e-4213-9591-ec3e5e763b92_content_list.json b/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/5102f011-579e-4213-9591-ec3e5e763b92_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aa9e735180bd19454c85fcbbee9b9eccd805d921 --- /dev/null +++ b/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/5102f011-579e-4213-9591-ec3e5e763b92_content_list.json @@ -0,0 +1,1429 @@ +[ + { + "type": "text", + "text": "TikZero: Zero-Shot Text-Guided Graphics Program Synthesis", + "text_level": 1, + "bbox": [ + 179, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jonas Belouadi* Eddy Ilg† Margret Keuper*,‡ Hideki Tanaka§ Masao Utiyama§ Raj Dabre§ Steffen Eger† Simone Ponzetto* \nUniversity of Mannheim, Germany* University of Technology Nuremberg, Germany† Max Planck Institute for Informatics, Saarland Informatics Campus, Germany‡ National Institute of Information and Communications Technology, Japan§ jonas.belouadi@uni-mannheim.de", + "bbox": [ + 133, + 178, + 870, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 310, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automatically synthesizing figures from text captions is a compelling capability. However, achieving high geometric precision and editability requires representing figures as graphics programs in languages like TikZ, and aligned training data (i.e., graphics programs with captions) remains scarce. Meanwhile, large amounts of unaligned graphics programs and captioned raster images are more readily available. We reconcile these disparate data sources by presenting TikZERO, which decouples graphics program generation from text understanding by using image representations as an intermediary bridge. It enables independent training on graphics programs and captioned images and allows for zero-shot text-guided graphics program synthesis during inference. We show that our method substantially outperforms baselines that can only operate with caption-aligned graphics programs. Furthermore, when leveraging caption-aligned graphics programs as a complementary training signal, TikZERO matches or exceeds the performance of much larger models, including commercial systems like GPT-4o. Our code, datasets, and select models are publicly available.1", + "bbox": [ + 89, + 342, + 483, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 672, + 220, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graphics programming languages offer distinct advantages over low-level vector formats (PDF, SVG) or raster image formats by representing visual concepts as high-level programs that preserve semantics, remain human-interpretable, and allow manual editing. These properties are particularly valuable in academia, where specialized graphics programming languages like TikZ [1] are popular for creating complex figures with high expressivity. However, this comes with a steep learning curve, as seen on the TEX Stack Exchange $^2$ (TEX.SE), where nearly $10\\%$ of questions concern TikZ and make it the most frequently discussed topic on the platform [2, 3].", + "bbox": [ + 89, + 699, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3847c540f76ba4bdd94cb752b78fe3ba6f97ed7688ef35cbf45137413f14e349.jpg", + "image_caption": [ + "Figure 1. Qualitative comparison of our $\\mathrm{TikZero}^+$ model (last two columns) and the end-to-end trained baseline $\\mathrm{AUTOMATikZ}_{\\mathrm{v2}}$ (LLM; first two columns) on text-guided graphics program synthesis with TikZ. Our method generates outputs that more closely follow the given captions. Example program listings are in Supp. Mat. F." + ], + "image_footnote": [], + "bbox": [ + 517, + 308, + 901, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With recent advances in generative AI, simplifying the creation of graphics programs has become increasingly feasible. Belouadi et al. [2] introduce $\\mathrm{D}\\mathsf{E}\\mathrm{T}\\mathsf{1}\\kappa \\mathrm{Z}\\mathsf{I}\\mathsf{F}\\mathsf{Y}$ , an inverse graphics model that generates TikZ programs from images and hand-drawn sketches. However, creating these visual inputs stays cumbersome, motivating alternative input modalities such as natural language. While Belouadi et al. [3] propose AutomATikZ, a text-guided synthesis model for TikZ programs trained end-to-end on an aligned caption-program corpus, its performance remains limited (cf. Fig. 1) [4, 5].", + "bbox": [ + 511, + 672, + 908, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We identify insufficient training data as the primary limitation. Unlike inverse graphics models such as $\\mathrm{DETikZIFY}$ , which are inherently self-supervised (trained by being conditioned on compiled representations of their output programs) and can access sufficient training data (cf. Fig. 2), end-to", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/potamides/DeTikZify", + "bbox": [ + 107, + 875, + 379, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "2https://tex.stackexchange.com", + "bbox": [ + 109, + 888, + 318, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17793", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a8bc9b60538868e0b4edd0b582689f6259d0cf86d8c62a48a5a0dd71b324d30a.jpg", + "image_caption": [ + "Figure 2. Illustration of training data availability for graphics program synthesis. $\\mathrm{DETIKZIFY}$ can leverage all graphics programs for training but lacks text guidance, while $\\mathrm{AUTOMATIKZ}$ is constrained to the small intersection of captioned graphics programs, resulting in limited performance. Our approach, $\\mathrm{TIKZERO}$ , trains independently on both graphics programs and captioned images, enabling more effective use of available data and yielding superior results." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 482, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "end text-guided models like AUTOMATIKZ require graphics programs paired with captions, substantially reducing the available data pool (cf. Fig. 2).", + "bbox": [ + 89, + 417, + 483, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we decouple the graphics program generation component from text understanding, enabling independent training on graphics programs and captioned images without requiring paired data (cf. Fig. 2). Our approach first trains an inverse graphics model conditioned on image patch embeddings from a vision encoder [6]. We then train an adapter network that generates synthetic image patch embeddings from captions. This adapter training relies solely on captioned images, effectively circumventing resource limitations and enabling zero-shot (in the sense that no aligned caption-program examples are involved in the training process) text-guided graphics program synthesis [7]. We demonstrate that this approach, to which we refer as $\\mathrm{T_{IK}ZERO}$ , outperforms previous state-of-the-art methods (cf. Fig. 1). Our key contributions are:", + "bbox": [ + 88, + 463, + 483, + 688 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(i) A novel two-stage architecture, $\\mathrm{TikZERO}$ , which addresses the low-resource challenge in text-guided graphics program synthesis by aligning representation spaces rather than relying on aligned data.", + "(ii) The $\\mathrm{DATiKZ_{v3}}$ dataset, comprising over 450k TiKZ graphics programs with roughly 170k captioned samples. Using this dataset, we train both $\\mathrm{TiKZERO}$ and $\\mathrm{AUTOMATiKZ_{v2}}$ (an updated version of $\\mathrm{AUTOMATiKZ}$ ) on the same source data and show that $\\mathrm{TiKZERO}$ outperforms $\\mathrm{AUTOMATiKZ}$ , $\\mathrm{AUTOMATiKZ_{v2}}$ , and other end-to-end trained baselines.", + "(iii) An enhanced model, $\\mathrm{T_{IK}Z_{ZERO} + }$ , combining $\\mathrm{T_{IK}Z_{ZERO}}$ with the end-to-end fine-tuning of $\\mathrm{AUTOMAT_{IK}Z_{v2}}$ , which surpasses larger baselines and matches the performance of commercial models like GPT-4o [8] on key metrics." + ], + "bbox": [ + 78, + 689, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inverse Graphics Program Synthesis Inverse graphics, i.e., synthesizing a graphics program to reproduce a visual target, represents a specialized instance of neural program synthesis [9-11]. Deep learning models have shown remarkable success in this domain [12-14], with Vision-Language Models (VLMs) increasingly gaining prominence [2, 15-17]. While controlled experimental studies often rely on synthetic datasets [13, 15, 17-20], real-world applications typically leverage more complex and diverse human-created data [2, 21-25], highlighting the importance of data availability. In scientific contexts, TikZ has emerged as a popular choice due to its versatility, expressiveness, and widespread adoption in academic circles [2, 21-23, 25]. Although these approaches are not tailored to text-guided generation, we incorporate key elements from them into our approach.", + "bbox": [ + 511, + 114, + 906, + 342 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-Guided Graphics Program Synthesis Current text-guided approaches to graphics program synthesis remain limited, mainly because of the scarcity of captioned graphics programs outlined in Sec. 1, but also because of the difficulty of generating synthetic data with human-like captions [26, 27]. Researchers interested in this capability currently rely on the emerging capabilities of large commercial models such as GPT-4o [8, 28–30], which raises concerns about accessibility, reproducibility, and computational cost [31]. In contrast, related domains like vector graphics generation [24, 32–35] and NL2Vis [36–40] have shown more progress. Similar to inverse graphics, these fields increasingly incorporate large language models (LLMs) [24, 36, 40]. However, vector graphics approaches typically generate only low-level Bezier curves, limiting output complexity [2, 32, 33], and NL2Vis focuses exclusively on data visualization with a restricted set of visualization types [40]. More complex applications, such as generating arbitrary scientific figures from captions with TikZ, remain underexplored—a gap we address in this work.", + "bbox": [ + 511, + 349, + 908, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Image Generation $\\mathrm{TiKZero}$ shares conceptual and architectural similarities with several text-to-image generation methods [41-45]. Rodriguez et al. [46, 47] explore generative adversarial networks [48, 49] and diffusion models [50, 51] for scientific figure generation, but these approaches are tied to raster images, which are not ideal for representing scientific figures. Ramesh et al. [43] propose a two-stage model with independently trained prior and decoder components to generate raster images from text. Although prior networks resemble our adapters and have been used with inverse graphics models [52], they target global image embeddings containing only abstract information, which degrades performance when used with inverse graphics architectures that work best with patch-level details [53]. In contrast, our adapters specifically operate on patch-level embeddings, and we demonstrate that this improves performance compared to end-to-end trained baselines.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17794", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5a4db9ff5a15611160ed325154439f80a3c320b123f535faeb2122769ccb72be.jpg", + "image_caption": [ + "Figure 3. Architecture overview of $\\mathrm{T1KZERO}$ during inference. Solid lines represent the standard caption-conditioned path, which flows through the text encoder into the adapter network of $\\mathrm{T1KZERO}$ before connecting to the vision encoder. In certain configurations (cf. Sec. 5.2), the caption also feeds into the text decoder (depicted by dotted lines and “ $\\bullet$ ” markers representing shortcuts). The self- and cross-attention layers (yellow) are simplified representations, omitting internal feed-forward layers and residual connections [54]. An exception is the explicit residual connection between the cross-attention and self-attention layers of the vision encoder, visualizing the gating mechanism $\\gamma$ (purple). Additionally, the dashed path illustrates how the inverse graphics model generates graphics programs when conditioned on images." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 906, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The TIKZERO Model & Architecture", + "text_level": 1, + "bbox": [ + 89, + 440, + 419, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the foundation of $\\mathrm{TiKZero}$ , we first develop a state-of-the-art inverse graphics model for graphics program synthesis. We then incorporate a cross-attention adapter network [55] for text guidance. Fig. 3 provides an overview of our method.", + "bbox": [ + 89, + 465, + 485, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Inverse Graphics Model Due to their demonstrated effectiveness (cf. Sec. 2), we adopt a VLM architecture for the inverse graphics model of $\\mathrm{T_{IK}Zero}$ . Figure 3 illustrates its inner workings (dashed lines): the model processes rasterized images and autoregressively generates their corresponding programs without involving captions at this stage.", + "bbox": [ + 89, + 532, + 483, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Adapter Network VLMs consist of two primary components: a vision encoder that produces image embeddings and a text decoder that, in our case, generates graphics programs conditioned on these embeddings. The unidirectional and localized flow of information between these components allows us to inject additional information solely into the vision encoder, thereby influencing the output of the text decoder. We exploit this property by introducing a trainable, text-conditioned adapter network that mimics the outputs of the original vision encoder. This effectively enables zero-shot generation of graphics programs conditioned on text when its outputs are fed into the decoder. In addition to circumventing the resource limitations discussed in Sec. 1, this architecture has other welcome implications: During adapter training, the text decoder, usually the largest component of the model, does not need to be loaded, resulting in efficient and fast training even with large datasets. Our adapter incorporates a lightweight text encoder for embedding captions and intro-", + "bbox": [ + 89, + 628, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "duces newly initialized gated cross-attention layers [56, 57] before each vision encoder layer (cf., Fig. 3). The keys and values derive from the final text encoder representations, while the queries originate from a trainable probe used instead of image inputs. The gates $\\gamma$ allow the model to learn at which layers and to what extent information from the text encoder should flow into the vision encoder. Contrary to existing literature, which often employs tanh gates [58] that initialize to zero (indicating no information flow), we find that using sigmoid gates (0.5 at initialization) accelerates training convergence since, in our case, only little information originates from the vision encoder inputs (i.e., the probe). We ablate the gates and the probe in Supp. Mat. A.", + "bbox": [ + 511, + 441, + 908, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training Objective Given a caption-image dataset for training, we first embed the patches $p \\in \\mathbf{p}$ of image $i$ using the unmodified vision encoder $\\mathbf{M}$ of our VLM. Subsequently, we incorporate the cross-attention adapter to obtain the modified encoder $\\widehat{\\mathbf{M}}$ , which we then distill on these image patch embeddings conditioned solely on the caption $t$ and probe $\\hat{t}$ [59]. This leads to the following objective:", + "bbox": [ + 511, + 645, + 908, + 752 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d i s t}} = \\frac {1}{| p |} \\sum_ {p \\in p} \\operatorname {d i s t} \\left(\\mathrm {M} _ {\\theta} (p \\mid i), \\widehat {\\mathrm {M}} _ {\\theta , \\hat {\\theta}} (p \\mid \\hat {t}, t)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 762, + 906, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{dist}(\\pmb{x},\\pmb{y})$ represents a distance metric. Following common practices in model distillation [60, 61], we experiment with cosine distance and mean squared error. Here, $\\theta$ denotes the original model parameters that remain fully frozen, while $\\hat{\\theta}$ represents the adapter parameters of which the cross-attention layers and the image probe are trainable.", + "bbox": [ + 511, + 809, + 910, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17795", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/a57abde527289ba1c77e24470e3d10042b4adc73c2562d1645786e418c9fb559.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SourceDATIKZDATIKZv2DATIKZv3
curated98115663646
TEX.SE292383060942654
arXiv85656326450407851
artificial195719582256
all117832360583456469
", + "bbox": [ + 120, + 88, + 452, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Breakdown of the number of unique $\\mathrm{Ti}k\\mathrm{Z}$ graphics in $\\mathrm{DATI}\\kappa \\mathrm{Z}_{\\mathrm{v3}}$ compared to its predecessors $\\mathrm{DATI}\\kappa \\mathrm{Z}$ and $\\mathrm{DATI}\\kappa \\mathrm{Z}_{\\mathrm{v2}}$ . Qualitative examples can be found in Supp. Mat. F.", + "bbox": [ + 89, + 212, + 485, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Datasets & Model Training", + "text_level": 1, + "bbox": [ + 89, + 284, + 344, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce $\\mathrm{DATikZ_{v3}}$ , a novel dataset of TikZ graphics programs designed to support the training and evaluation of $\\mathrm{TIKZERO}$ . Additionally, we train $\\mathrm{AUTOMATikZ_{v2}}$ as a directly comparable baseline operating on the same data source.", + "bbox": [ + 89, + 311, + 483, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The $\\mathrm{DATiKZ_{v3}}$ Dataset $\\mathrm{DATiKZ_{v3}}$ expands upon its predecessors $\\mathrm{DATiKZ_{v2}}$ [2, 3], incorporating programs from curated repositories, TEX.SE, arXiv papers, and artificial samples (cf. Tab. 1). While previous versions focused exclusively on TikZ graphics with (v1) or without (v2) captions, $\\mathrm{DATiKZ_{v3}}$ systematically extracts captions alongside TikZ graphics whenever possible to support our claims. From over 450k instances, fewer than 170k include captions, underscoring the challenges discussed in Sec. 1.", + "bbox": [ + 89, + 387, + 483, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training TIKZERO $\\mathrm{T_{IK}Zero}$ 's VLM builds upon DE $\\mathrm{T_{IK}ZIFY}$ [2] by conditioning a LAMA-based text decoder [62] on patch embeddings from a S1G1P vision encoder [63]. Specifically, we combine $\\mathrm{LLAMA}_{3.1}$ (8B) [57] with S1G1P SoViT (0.4B). Unlike $\\mathrm{DE_{IK}ZIFY}$ and inspired by the continued ViT pretraining approach of INTERNVL 1.5 [64], we initialize the vision encoder with weights from the fine-tuned encoder of PALIGEMMA [65] and increase the input resolution to $420\\times 420$ pixels. Furthermore, we fully fine-tune the vision encoder alongside the rest of the model instead of freezing it. We train on $\\mathrm{DAT_{IK}Z_{v3}}$ for 5 epochs with a learning rate of 5e-5 and a batch size of 128. $\\mathrm{T_{IK}ZERO}$ 's VLM consistently outperforms $\\mathrm{DE_{IK}ZIFY}$ , with detailed evaluation results provided in Supp. Mat. B. For the adapter network, we initialize with $\\mathrm{LLAMA}_{3.2}$ (1B) as the text encoder [57] and leverage ARxivCAP [26], a dataset comprising 6.4 million scientific caption-image pairs for training. The adapter accounts for 2 billion of TIKZERO's 10 billion total parameters, with only 400 million being trainable. We train for 3 epochs with a learning rate of 1e-4 and a batch size of 512. We emphasize that this two-stage training process does not access caption-program pairs. However, we demonstrate that incorporating such aligned data in a subsequent fine-tuning step (Sec. 5.2) further enhances performance.", + "bbox": [ + 89, + 537, + 485, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training AUTOMATIKZv2 Similar to its predecessor, AUTOMATIKZv2 is a token-conditioned LLM that uses tokenized captions as conditioning information for graphics prediction (rather than patch embeddings). We initialize AUTOMATIKZv2 in two different ways: (i) AUTOMATIKZv2 (LLM), which starts from vanilla LLAMA3.1 (8B) weights, and (ii) AUTOMATIKZv2 (VLM), which leverages TIKZERO's trained VLM (minus the vision encoder) to benefit from transfer learning [66] on its larger training corpus. Both variants employ the same hyperparameters as TIKZERO's VLM but can only utilize the caption-annotated subset of DATIKZv3 for training. Despite having access to less caption-aligned data than TIKZERO's adapter network, AUTOMATIKZv2 requires a longer training period primarily due to fine-tuning the large decoder (8 billion trainable parameters versus the adapter network's 400 million). Training requires more than two days for AUTOMATIKZv2 and 1.5 days for TIKZERO's adapter network when using eight Nvidia A100 40GB GPUs.", + "bbox": [ + 511, + 90, + 906, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 378, + 645, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before training models on $\\mathrm{DAT}_{\\mathrm{IK}}\\mathrm{Z}_{\\mathrm{v3}}$ , we extract 1k samples from its captioned subset to form our test set. To mitigate data leakage from pretraining to testing, we only include instances created after the cut-off date specified by $\\mathrm{LLAMA}_{3.2}$ and $\\mathrm{ARXIVCAP}$ . We also employ an $n$ -gram matching algorithm to avoid cross-contamination with our training split [8]. For all models, the temperature is set to 0.8 and top-p to 0.95. Example outputs are provided in Fig. 1 and Supp. Mat. F.", + "bbox": [ + 511, + 406, + 908, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation Metrics The multimodal nature of our task allows for various evaluation metrics in our automatic evaluations. We assess perceptual image similarity between generated outputs and references by computing DREAMSIM (DSIM) [67, 68], which correlates highly with human judgments for scientific images [2]. We also calculate the Kernel Inception Distance (KID) [69] using SigLIP image features, which evaluates the overall quality of generated figures by comparing to the distribution of reference figures. We evaluate caption similarity between generated outputs and reference captions using CLIPSCORE (CLIP) [70] with SigLIP features. To measure code similarity between generated and reference TikZ programs, we use CRYSTALBLEU (cBLEU), a BLEU variant optimized for code evaluation [71, 72], and TEX Edit Distance (TED) [2], a variant of the Extended Edit Distance [73] utilizing a TEX tokenizer. Since some metrics require that generated programs compile to images, resampling is necessary if the output contains irrecoverable errors. To quantify this, we compute the Mean Token Efficiency (MTE), defined as the $10\\%$ winsorized mean of the ratio between the number of tokens in the final TikZ program and the total number of tokens generated to produce that program. For a comprehensive view of model performance, we calculate the arithmetic mean (AVG) of all previous", + "bbox": [ + 511, + 537, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17796", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/7325da1a0578251ed24aa247c5f1d746a5c3af9c4fab3b4cd8f4cfaf2ce9cbbe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsOriginal TextRedacted Text
DSIM↑KID↓CLIP↑cBLEU↑TED↓MTE↑AVG↑CLIP↑Ratio↑
IDEFICS 3 (8B)45.47511.42614.3270.65663.17569.55866.6284.85133.858
AUTOMATIKZ (13B)46.0331.2943.9550.38662.2485.86663.0932.96574.975
AUTOMATIKZv2 (VLM)38.31333.2030.7750.32876.98521.5950.00.28436.597
AUTOMATIKZv2 (LLM)50.5483.49115.7660.65862.30781.77582.3758.00250.753
TIKZERO (MSE)52.0245.66410.5831.72366.0779.31885.0048.23777.831
TIKZERO (Cos)52.8295.10310.0511.60365.5182.29185.5997.22671.893
", + "bbox": [ + 107, + 88, + 890, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. System-level scores $\\times {100}$ for ${\\mathrm{{T\\kappa Zero}}}_{0}$ and baselines of comparable size and training setup. Bold and underlined values denote the best and second-best scores for each metric column, respectively. Cell shading illustrates relative score magnitudes. Arrows indicate metric directionality. Overall, ${\\mathrm{{T\\kappa Zero}}}_{0}$ achieves the strongest average performance across metrics.", + "bbox": [ + 88, + 258, + 906, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "metrics. As these metrics operate on different scales, we apply min-max normalization before computing the average. Additionally, some metrics are recomputed with redacted text in the outputs as part of our analysis, cf. Sec. 6.1.", + "bbox": [ + 88, + 327, + 485, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Comparison against End-to-End Fine-Tuning", + "text_level": 1, + "bbox": [ + 89, + 393, + 478, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our initial experiment, we evaluate the zero-shot performance of $\\mathrm{TiKZero}$ , trained as described in Secs. 3 & 4 using either cosine distance (Cos) or mean squared error (MSE), and compare it against end-to-end trained baselines.", + "bbox": [ + 88, + 416, + 485, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines Besides $\\mathrm{AUTOMATIKZ_{v2}}$ (LLM & VLM), which we designed as directly comparable baselines, we assess other token-conditioned models of similar and slightly larger sizes trained on $\\mathrm{TikZ}$ . Specifically, we evaluate $\\mathrm{AUTOMATIKZ}$ (13B) $^3$ the strongest original $\\mathrm{AUTOMATIKZ}$ baseline [3], and the general-purpose chatbot IDEFICS 3 (8B) [22]. Additional models and details are available in Supp. Mats. A & C.", + "bbox": [ + 88, + 482, + 483, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results We present the system-level metric scores in Tab. 2 (Original Text). On average, $\\mathrm{T_{IK}Z_{ero}}$ , trained with cosine distance, achieves the best performance with an AVG score of 85.599, closely followed by the MSE variant at 85.004. The next best model, $\\mathrm{AUTOMAT_{IK}Z_{v2}}$ (LLM), scores 82.375, which is 3 percentage points (pp) lower. The remaining models exhibit a substantial performance gap, with IDEFICS 3 (8B) and $\\mathrm{AUTOMAT_{IK}Z}$ (13B) falling behind by approximately 20pp and $\\mathrm{AUTOMAT_{IK}Z_{v2}}$ (VLM) showing the weakest performance across all metrics, resulting in an AVG score of 0. The surprisingly poor results of $\\mathrm{AUTOMAT_{IK}Z_{v2}}$ (VLM) are likely due to catastrophic forgetting [74], as the removal of the vision encoder from $\\mathrm{T_{IK}Z_{ERO}}$ ’s VLM necessitates reacquisition of conditioning based solely on text.", + "bbox": [ + 88, + 593, + 485, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As for individual metrics, our adapter-based models perform particularly well in perceptual image similarity, with $\\mathrm{T_{IK}Z_{\\mathrm{ero}}}$ (Cos) outperforming the best baseline, AUTOMATIKZv2 (LLM), by 3pp on DREAMSIM. Although AUTOMATIKZv2 (LLM) outperforms $\\mathrm{T_{IK}Z_{\\mathrm{ero}}}$ by 1.5pp on KID, this", + "bbox": [ + 88, + 805, + 485, + 881 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "indicates in this context that such token-conditioned models (compared to those using patch embeddings) capture the general appearance of scientific figures well but fall short in inferring visual specifics from captions. They do, however, have an edge in reproducing text from captions, which we identify as the primary reason for up to 5pp higher CLIP-Score, as noted in Sec. 6.1. Regarding code similarity, both $\\mathrm{TIKZERO}$ models considerably outperform others on cBLEU. Interestingly, we observe a mild inverse correlation between cBLEU and TED. Models conditioned solely on tokenized captions tend to generate shorter, often simplified programs [3], potentially resulting in a reduced edit distance to the reference. In terms of efficiency, all models achieve an MTE of 80-85, indicating that only 2 out of 10 inferences require resampling. $\\mathrm{TIKZERO}$ (Cos) is 3pp more efficient than MSE, while AUTOMATIKZ (13B), likely benefiting from its larger model size, exceeds it by another 3pp.", + "bbox": [ + 511, + 327, + 908, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, training $\\mathrm{AUTOMATIKZ_{v2}}$ on top of a VLM yields worse performance than training based on vanilla $\\mathrm{LLAMA}_{3.1}$ , indicating that effective end-to-end training can only leverage the small intersection of graphics programs and images with captions, as illustrated in Fig. 2. However, even without access to this intersection, $\\mathrm{TikZERO}$ surpasses both $\\mathrm{AUTOMATIKZ}$ (13B) and $\\mathrm{AUTOMATIKZ_{v2}}$ (LLM) on average by being able to train on images with captions independently of graphics programs. Moreover, using a loss function based on cosine distance proves more effective than using MSE.", + "bbox": [ + 511, + 584, + 908, + 737 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Combining Adapters with Fine-Tuning", + "text_level": 1, + "bbox": [ + 511, + 742, + 849, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we investigate whether explicitly incorporating the subset of $\\mathrm{DAT}\\kappa \\mathrm{Z}_{\\mathrm{v3}}$ that includes captions into the training process of $\\mathrm{T}\\mathbf{I}\\kappa \\mathbf{Z}\\mathbf{E}\\mathbf{R}\\mathbf{O}$ enhances performance. Our approach involves three incremental stages: (i) We perform a light fine-tuning of $\\mathrm{T}\\mathbf{I}\\kappa \\mathbf{Z}\\mathbf{E}\\mathbf{R}\\mathbf{O}$ (Cos) end-to-end on caption-program pairs for one epoch with a low learning rate of 1e-5. Extending the training duration or increasing the learning rate does not yield further performance gains, likely due to the decoder having already reached its saturation point;", + "bbox": [ + 511, + 763, + 910, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3Belouadi et al. [3] refer to this model as CLiMA (13B).", + "bbox": [ + 107, + 886, + 408, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17797", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ab5fb95989f476e4f28695960b72b52bf9b287b7a9437d10256381ee6e3512fc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsOriginal TextRedacted Text
DSIM↑KID↓CLIP↑cBLEU↑TED↓MTE↑AVG↑CLIP↑Ratio↑
QWEN2.5 CODER (32B)54.4735.49324.870.28559.85697.26948.59312.16448.911
GPT-4o56.4642.84431.7870.32758.51197.67579.01913.3241.905
TikZero (Cos)52.8295.10310.0511.60365.5182.29114.6587.22671.893
+ Fine-tuning (i)53.2031.79410.6870.75961.57294.85146.4976.51260.931
+ Separate Captions (ii)52.9832.90515.720.80461.3295.72246.3268.74155.608
+ Weight Resetting (iii)56.2951.83124.1771.98859.00893.05887.04311.47947.478
", + "bbox": [ + 109, + 88, + 887, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. System-level scores $\\times {100}$ for additional baselines and ${\\mathrm{{TikZero}}}$ combined with fine-tuning and token-conditioning. The scores for TIKZERO (Cos) are replicated from Tab. 2 for convenience. Bold and underlined values denote the best and second-best scores for each metric column, respectively. Cell shading illustrates relative score magnitudes. Arrows indicate metric directionality. Overall, ${\\mathrm{{TikZero}}}_{\\mathrm{O}}\\left( \\mathrm{{Cos}}\\right)$ with Weight Resetting (iii) demonstrates the strongest average performance across metrics.", + "bbox": [ + 89, + 258, + 906, + 315 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(ii) Alongside feeding captions into the adapter, we provide them separately to the text decoder in tokenized form (cf., Fig. 3); (iii) Prior to fine-tuning, we reset the decoder to its initial weights to overcome saturation, enabling us to fine-tune using the setup described in Sec. 4, which involves 5 epochs and a learning rate of $5\\mathrm{e} - 5$", + "bbox": [ + 89, + 340, + 480, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines In addition to the baselines in Tab. 2, which remain comparable, we also evaluate larger and commercial models that serve as stronger baselines (cf. Supp. Mat. C). Specifically, we assess GPT-4o [8], which has demonstrated strong performance in generating TikZ [3, 28, 30] and QwEN2.5 CODER (32B) [75] as an open-weights model.", + "bbox": [ + 89, + 439, + 482, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results In Tab. 3 (Original Text), all fine-tuning setups of $\\mathrm{TiKZero}$ show considerable improvement over the base version. Approaches (i) and (ii) each enhance performance by over 30pp on AVG, while approach (iii) surpasses them with an improvement of over 70pp, positioning it as the best-performing model on average, even when compared to our new baselines, with GPT-4o being 8pp lower and $\\mathrm{QWEN}_{2.5}$ CODER (32B) approximately 40pp lower. Approach (i) demonstrates that direct fine-tuning yields positive effects across nearly all metrics, notably improving MTE by 12pp, TED by 4pp, and KID by 3.5pp. Approach (ii) shows similar trends but, by also incorporating tokenized captions, further improves CLIPSCORE by 5pp, closing the gap to AUTOMATIKZv2 (LLM). Interestingly, both (i) and (ii) slightly decrease performance on cBLEU, potentially due to similar reasons discussed in Sec. 5.1. However, the same cannot be said for (iii), which not only achieves the highest score on cBLEU but also ranks as the second-best on TED, trailing only 0.5pp behind GPT-4o and showcasing that it is possible to perform well on both metrics. Additionally, it increases DREAMSIM by another 3pp and CLIPSCORE by 8.5pp, competing with the much stronger baselines QWEN2.5 CODER (32B) and GPT-4o. In KID, it even surpasses them by 3.5pp and 1pp, respectively.", + "bbox": [ + 89, + 537, + 482, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, fine-tuning $\\mathrm{T_{IK}Zero}$ , especially when com", + "bbox": [ + 109, + 886, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "bined with a separate caption input and weight resetting, greatly improves performance. This illustrates that the intersection of graphics programs and images with captions, though small, provides a valuable training signal, and best performance can be achieved by making full use of both sets. The best-performing $\\mathrm{TiKZero}$ model even competes with and often surpasses QwEN2.5 CODER (32B) and GPT-4o on several key metrics. Notably, the former model is more than three times larger, and the latter is often estimated at around 1.8 trillion parameters [76], making it 180 times larger.", + "bbox": [ + 511, + 340, + 906, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Human Evaluation", + "text_level": 1, + "bbox": [ + 513, + 512, + 694, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To corroborate our findings from automatic evaluation, we conduct a human annotation campaign focusing on two key properties: caption and image similarity. We employ BestWorst Scaling (BWS) [77], a comparative annotation method that yields high-quality results even with few annotators [78, 79]. We sample 100 instances from our test set and present annotators with $n$ -tuples of generated figures, asking them to identify the most and least similar figure to either the reference caption or reference image. This data is then transformed into scores from -1 (poor) to 1 (excellent) by subtracting the proportion of times a figure is selected as the best from the proportion of times it is chosen as the worst [80]. For a manageable workload, we focus on $n = 4$ key models: $\\mathrm{TikZERO}(\\mathrm{Cos})$ , our best-performing model from Sec. 5.1; $\\mathrm{AUTOMATIKZv2}$ (LLM), its direct end-to-end trained competitor; GPT-4o, our strongest baseline; and $\\mathrm{TikZERO}(\\mathrm{Cos})$ fine-tuned using approach (iii) from Sec. 5.2, our best model overall, henceforth referred to as $\\mathrm{TikZERO}^+$ for convenience. We engage thirteen annotators and obtain six fully annotated sets per task (cf. Supp. Mat. E for more details). To assess annotator consistency, we calculate the split-half reliability (SHR) [79]. This method randomly divides all annotations into two sets, calculates scores independently, and then determines their correlation using Spearman's $\\rho$ .", + "bbox": [ + 511, + 537, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17798", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f5c20c6dddaabcd9f5290421346d7a7e02abe869dceed211d4680fe1f6259f7.jpg", + "image_caption": [ + "Figure 4. Bivariate distributions of BWS scores (higher is better) using kernel density estimation for caption and image similarity. Along the diagonal, $\\mathrm{T_{IK}Z_{ERO}}$ (Cos) achieves higher scores than $\\mathrm{AUTOMAT_{IK}Z_{v2}}$ (LLM), while $\\mathrm{T_{IK}Z_{ERO+}}$ and GPT-4o demonstrate superior performance compared to both." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 472, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results Fig. 4 presents kernel density estimates for the BWS scores, showing generally consistent rankings with automatic evaluations but revealing notable differences in the magnitude of gaps. For caption similarity, the ranking aligns with CLIPSCORE evaluations $(\\rho = 1.0)$ , with $\\mathrm{TikZero}$ (Cos), $\\mathrm{AUTOMATIKZ_{v2}}$ (LLM), $\\mathrm{TikZero}+$ , and GPT-4o achieving mean scores $\\mu$ of -0.25, -0.18, 0.03, and 0.4, respectively. Interestingly, humans perceive a $40\\%$ smaller gap between $\\mathrm{AUTOMATIKZ_{v2}}$ (LLM) and $\\mathrm{TikZero}$ (Cos) than suggested by CLIPSCORE values, where $\\mathrm{AUTOMATIKZ_{v2}}$ (LLM) outperforms $\\mathrm{TikZero}$ (Cos) by $50\\%$ . This indicates humans may evaluate caption similarity differently than CLIPSCORE (cf. Sec. 6.1). For image similarity, the system order remains consistent with our DREAMSIM metric $(\\rho = 1.0)$ , with $\\mathrm{AUTOMATIKZ_{v2}}$ (LLM), $\\mathrm{TikZero}$ (Cos), $\\mathrm{TikZero}+$ , and GPT-4o achieving $\\mu$ of -0.26, -0.02, 0.01, and 0.27, respectively. However, the relative gaps between models differ: the separation between $\\mathrm{AUTOMATIKZ_{v2}}$ (LLM) and $\\mathrm{TikZero}$ (Cos), as well as between $\\mathrm{TikZero}+$ and GPT-4o, appear more pronounced than observed with DREAMSIM. This discrepancy likely stems from BWS capturing relative preferences rather than absolute performance differences. GPT-4o is selected $20\\%$ more often as the best model than $\\mathrm{TikZero}+$ , and $\\mathrm{AUTOMATIKZ_{v2}}$ (LLM) $15\\%$ more often as the worst model than $\\mathrm{TikZero}$ (Cos), creating larger perceived gaps even when qualitative differences may be subtle.", + "bbox": [ + 91, + 474, + 483, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The SHR values of 0.68 for caption similarity and 0.76 for image similarity indicate moderate to strong inter-annotator", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "agreement. We also observe a correlation between these two tasks, with segment-level $\\rho = 0.62$ and system-level $\\rho = 0.8$ , suggesting that both evaluation dimensions capture related aspects of model performance. GPT-4o emerges as the best-performing model, aligning with its superior performance on the corresponding automatic metrics, CLIPSCORE and DSIM. Among open-source models, $\\mathrm{T_{IK}Z_{v2}}$ performs best, while AUTOMATIKZv2 (LLM) ranks lowest overall.", + "bbox": [ + 511, + 90, + 908, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Analysis", + "text_level": 1, + "bbox": [ + 511, + 224, + 609, + 242 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present a comprehensive analysis, investigating the influence of typographic attacks on CLIPSCORE and examining the effectiveness of our architecture in low-resource settings, both in terms of training data and trainable parameters.", + "bbox": [ + 511, + 250, + 906, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. CLIPSCORE Limitations & Typographic Attacks", + "text_level": 1, + "bbox": [ + 511, + 319, + 906, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A known limitation of CLIPSCORE with text-rich images is its susceptibility to typographic attacks, where scores are disproportionately influenced by string similarity between images and captions [3, 43]. We suspect that token-conditioned models like AUTOMATIKZv2 (LLM) achieve higher CLIPSCORE values than models such as TIKZERO (Cos & MSE) primarily because they tend to visibly copy more substrings from the caption in the output image. To test this hypothesis, we apply the ROT13 substitution cipher [81] to all visible strings in the generated figures and recompute CLIPSCORE. This basic cipher replaces each letter with the 13th letter after it in the Latin alphabet. While not cryptographically secure, the ratio between the original and recomputed CLIPSCORE values should indicate the influence of string matching, i.e., higher ratios suggest less copied text and vice versa.", + "bbox": [ + 511, + 340, + 908, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tabs. 2 & 3 (Redacted Text) present the recomputed CLIP-Score values and ratios for all evaluated models. The results reveal that most $\\mathrm{T_{IK}Zero}$ models, except for fine-tuning approaches (ii) and (iii), which also condition on tokenized captions, have considerably higher ratios $(61\\% -78\\%)$ compared to strictly token-conditioned models $(34\\% -51\\%)$ , supporting our hypothesis. AUTOMATIKZ (13B) is an exception, possibly due to its initially low score. Further analysis shows that with redacted text, $\\mathrm{AUTOMATIKZ}_{\\mathrm{V2}}$ (LLM)'s CLIPSCORE performance drops to the same level as $\\mathrm{T_{IK}Zero}$ (Cos & MSE), suggesting that string matching is the primary factor in its superior performance rather than producing better visuals—arguably a more difficult task. Nevertheless, reproducing strings is still somewhat desirable. The human oracle of our test set achieves a ratio of $50.8\\%$ , close to the $47.5\\%$ of $\\mathrm{T_{IK}Zero+}$ , our best-performing model. In contrast, models like GPT-4o, with a lower ratio of $41.9\\%$ , may overfit to caption copying, artificially inflating the CLIPSCORE values.", + "bbox": [ + 511, + 566, + 910, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Low-Resource Training", + "text_level": 1, + "bbox": [ + 511, + 848, + 732, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While our adapters train efficiently on large-scale datasets, we investigate whether such extensive data is necessary for opti", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17799", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bd86657fbf675f2526105b7fa9175380a6d0849e7faf0a66fd8ddbf74c2b5c12.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. AVG scores for $\\mathrm{TiKZero}$ (Cos) trained on varying fractions of data and intervals of cross-attention layers. Higher scores indicate better performance. Bold and underlined values denote the best and second-best scores for the whole table, respectively. Cell shading illustrates score magnitudes." + ], + "table_body": "
Intv.Training Data
100%50%25%12.5%
192.41177.47849.96756.055
287.55785.24954.94233.817
482.25447.38132.1237.914
876.54540.81629.77416.25
", + "bbox": [ + 130, + 88, + 444, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mal performance. Along the same vein, we examine the impact of reducing the amount of cross-attention layers inserted into the vision encoder. We retrain $\\mathrm{TIKZERO}$ (Cos) using varying fractions of the training data ( $d \\in \\{1, \\frac{1}{2}, \\frac{1}{4}, \\frac{1}{8}\\}$ ) and insert cross-attention layers at different intervals ( $i \\in \\{1, 2, 4, 8\\}$ ). Table 4 presents the AVG scores from this parameter grid, with detailed scores in Supp. Mat. D. Our findings reveal that utilizing the full dataset and inserting cross-attention at every layer yields the highest average performance, highlighting the benefits of maximizing both variables. Interestingly, the model's performance appears more robust to a reduction in the number of layers compared to a decrease in training data. For instance, training on only $\\frac{1}{8}$ th of the data leads to a substantial performance drop of $36\\mathrm{pp}$ , whereas inserting cross-attention layers every 8 layers (resulting in only 3 cross-attention layers in total) causes a more modest decline of $16\\mathrm{pp}$ . Minimizing both variables leads to the most severe drop of over $75\\mathrm{pp}$ . These results validate our training setup while suggesting that incorporating additional data might further enhance performance. Given that $\\mathrm{ARXIVCAP}$ extracts figures from only $572\\mathrm{k}$ papers, whereas some corpora index over 200 million papers [82], there remains a lot of potential for leveraging larger datasets in future work.", + "bbox": [ + 91, + 316, + 483, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 678, + 209, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we demonstrate the potential of $\\mathrm{TiKZero}$ and its variants for generating TiKZ graphics programs from captions. Notably, $\\mathrm{TiKZero}$ does not require aligned caption-program pairs in its original formulation but instead aligns representation spaces of unaligned graphics programs and captioned images. This enables our model to leverage substantially more training data compared to end-to-end trained models that operate solely on caption-image pairs (cf. Fig. 2) while maintaining training efficiency. $\\mathrm{TiKZero}$ outperforms strong end-to-end trained baselines, including our independently trained AUTOMATiKZv2 models, which use the same data pool, excluding instances they cannot process, illustrating the strengths of our approach. When extending the $\\mathrm{TiKZero}$ ap", + "bbox": [ + 89, + 703, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "proach with additional end-to-end training, it also compares favorably to much larger baselines and commercial systems like GPT-4o. While this enhanced approach, $\\mathrm{T_{IK}ZERO^{+}}$ , is no longer zero-shot by our definition, it remains a $\\mathrm{T_{IK}ZERO}$ model in the sense that it operates on both sets of graphics programs and captioned images, with the added advantage of explicitly utilizing their intersection (cf. Fig. 2).", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These results demonstrate the benefits of designing architectures around available data and validate the approach of decoupling graphics program generation from text understanding (with optional later reconciliation through $\\mathrm{TiKZero^{+}}$ ). Although we demonstrate our method specifically on $\\mathrm{TiKZ}$ , we believe its general principles will inspire future work on related graphics program synthesis tasks.", + "bbox": [ + 511, + 198, + 908, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future Work Beyond scaling up our training data to explore convergence limits (cf. Sec. 6.2), we plan to investigate automatic methods for improving the quality and alignment of caption-image or caption-program pairs. This includes rewriting potentially noisy captions with LLMs and enhancing them with the visual understanding capabilities of VLMs [83-85]. We believe our approach to aligning textual and image modalities enables other promising applications for graphics program synthesis, such as editing images in latent space via textual instructions to generate modified graphics programs. Additionally, we intend to explore alternative alignment strategies beyond model distillation, including contrastive learning [86], which has successfully aligned modalities in discriminative models [6, 87, 88].", + "bbox": [ + 511, + 311, + 908, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 513, + 537, + 617, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our evaluations include proprietary systems that operate as black boxes; their training data is unknown, and they offer no guarantees of consistent performance over time. This (i) makes addressing data leakage and cross-contamination impossible and (ii) limits the fairness and reproducibility of our experiments. Nevertheless, even under these unfavorable conditions, our open models remain competitive. Users should be aware, however, that our models may behave unpredictably, and outputs might differ from expectations. Additionally, our models do not include safeguards against potential misuse, e.g., for generating fake scientific content.", + "bbox": [ + 511, + 564, + 908, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Regarding licensing of our training data, a large portion of the TikZ programs in $\\mathrm{DAT}_{\\mathrm{IK}}\\mathrm{Z}_{\\mathrm{v3}}$ are licensed under permissive terms that allow redistribution. The remaining programs are distributed under the arXiv.org perpetual, non-exclusive license, which prohibits redistribution, which is why we exclude them from the public release of $\\mathrm{DAT}_{\\mathrm{IK}}\\mathrm{Z}_{\\mathrm{v3}}$ . However, since we release our dataset creation scripts, we encourage others to reproduce the full version independently.", + "bbox": [ + 511, + 732, + 908, + 852 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "4https://creativecommons.org/licenses; https://opensource.org/licenses/mit; https://www.gnu.org/licenses/fdl-1.3.en.html; https://openai.com/policies/terms-of-use", + "bbox": [ + 511, + 862, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17800", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 246, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We extend our sincere gratitude to the following individuals (in no particular order) for their valuable contributions: Christian Greisinger, Hour Kaing, Ran Zhang, Tejaswini Medi, Yanran Chen, Sotaro Takeshita, Katharina Prasse, JiWoo Kim, Christoph Leiter, Haiyue Song, and Aida Kostikova. Their assistance with our human evaluation campaign, proofreading, insightful discussions, and constructive feedback has been instrumental to our work. The first author conducted part of this research during an internship at the National Institute of Information and Communications Technology (NICT), Japan. The second to last author is supported by the Federal Ministry of Education and Research (BMBF) via the research grant METRICS4NLG and the German Research Foundation (DFG) via the Heisenberg Grant EG 375/5-1. We acknowledge computing resources provided by the state of Baden-Württemberg through bwHPC and the German Research Foundation (DFG) through grant INST 35/1597-1 FUGG. Finally, we thank the OpenMoji project for the open-source icons used throughout this work.", + "bbox": [ + 89, + 113, + 485, + 380 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 406, + 186, + 422 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Till Tantau. The TikZ and PGF Packages, 2023. 1", + "[2] Jonas Belouadi, Simone Paolo Ponzetto, and Steffen Eger. DeTikZify: Synthesizing graphics programs for scientific figures and sketches with TikZ. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 4, 15", + "[3] Jonas Belouadi, Anne Lauscher, and Steffen Eger. AutomataTikZ: Text-guided synthesis of scientific vector graphics with TikZ. In The Twelfth International Conference on Learning Representations, 2024. 1, 4, 5, 6, 7, 16, 18", + "[4] Leixin Zhang, Yinjie Cheng, Weihe Zhai, Steffen Eger, Jonas Belouadi, Fahimeh Moafian, and Zhixue Zhao. ScImage: How good are multimodal large language models at scientific text-to-image generation? In The Thirteenth International Conference on Learning Representations, 2025. 1, 15", + "[5] Abhay Zala, Han Lin, Jaemin Cho, and Mohit Bansal. DiagrammerGPT: Generating open-domain, open-platform diagrams via LLM planning. In First Conference on Language Modeling, 2024. 1", + "[6] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 8", + "[7] Mark Palatucci, Dean Pomerleau, Geoffrey E Hinton, and Tom M Mitchell. Zero-shot learning with semantic output codes. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2009. 2", + "[8] OpenAI. GPT-4 technical report, 2023. 2, 4, 6", + "[9] Emilio Parisotto, Abdel rahman Mohamed, Rishabh Singh, Lihong Li, Dengyong Zhou, and Pushmeet Kohli. Neuro" + ], + "bbox": [ + 99, + 431, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "symbolic program synthesis. In International Conference on Learning Representations, 2017. 2", + "[10] Jacob Devlin, Jonathan Uesato, Surya Bhupatiraju, Rishabh Singh, Abdel rahman Mohamed, and Pushmeet Kohli. Robust-Fill: Neural program learning under noisy I/O. In Proceedings of the 34th International Conference on Machine Learning, pages 990-998. PMLR, 2017.", + "[11] Kevin Ellis, Catherine Wong, Maxwell Nye, Mathias Sable-Meyer, Lucas Morales, Luke Hewitt, Luc Cary, Armando Solar-Lezama, and Joshua B. Tenenbaum. DreamCoder: bootstrapping inductive program synthesis with wake-sleep library learning. In Proceedings of the 42nd ACM SIGPLAN International Conference on Programming Language Design and Implementation, page 835-850, New York, NY, USA, 2021. Association for Computing Machinery. 2", + "[12] Yaroslav Ganin, Tejas Kulkarni, Igor Babuschkin, S. M. Ali Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. In Proceedings of the 35th International Conference on Machine Learning, pages 1666-1675. PMLR, 2018. 2", + "[13] Kevin Ellis, Daniel Ritchie, Armando Solar-Lezama, and Josh Tenenbaum. Learning to infer graphics programs from hand-drawn images. In Thirty-second Conference on Neural Information Processing Systems, pages 6062–6071, 2018. 2", + "[14] Kevin Ellis, Maxwell Nye, Yewen Pu, Felix Sosa, Josh Tenenbaum, and Armando Solar-Lezama. Write, execute, assess: Program synthesis with a REPL. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2", + "[15] Peter Kulits, Haiwen Feng, Weiyang Liu, Victoria Fernandez Abrevaya, and Michael J. Black. Re-thinking inverse graphics with large language models. Transactions on Machine Learning Research, 2024. 2", + "[16] Wen-Ding Li and Kevin Ellis. Is programming by example solved by LLMs? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024.", + "[17] Shreyas Kapur, Erik Jenner, and Stuart Russell. Diffusion on syntax trees for program synthesis. In The Thirteenth International Conference on Learning Representations, 2025. 2", + "[18] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5515-5523. Computer Vision Foundation / IEEE Computer Society, 2018.", + "[19] Yonglong Tian, Andrew Luo, Xingyuan Sun, Kevin Ellis, William T. Freeman, Joshua B. Tenenbaum, and Jiajun Wu. Learning to infer and execute 3D shape programs. In International Conference on Learning Representations, 2019.", + "[20] Javier Camara, Javier Troya, Lola Burgueño, and Antonio Vallecillo. On the assessment of generative AI in modeling tasks: an experience report with chatgpt and UML. Softw. Syst. Model., 22(3):781-793, 2023. 2", + "[21] Hugo Laurençon, Leo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17801", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions, 2024. 5", + "[23] Shengbang Tong, Ellis L Brown II, Penghao Wu, Sanghyun Woo, Adithya Jairam Iyer, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, Xichen Pan, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian1: A fully open, vision-centric exploration of multimodal LLMs. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2", + "[24] Juan A. Rodriguez, Abhay Puri, Shubham Agarwal, Issam H. Laradji, Pau Rodriguez, Sai Rajeswar, David Vazquez, Christopher Pal, and Marco Pedersoli. StarVector: Generating scalable vector graphics code from images and text. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 16175-16186, 2025. 2", + "[25] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, Sam Dodge, Keen You, Zhen Yang, Aleksei Timofeev, Mingze Xu, Hong-You Chen, Jean-Philippe Fauconnier, Zhengfeng Lai, Haoxuan You, Zirui Wang, Afshin Dehghan, Peter Grasch, and Yinfei Yang. MM1.5: Methods, analysis & insights from multimodal LLM fine-tuning. In The Thirteenth International Conference on Learning Representations, 2025. 2", + "[26] Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal ArXiv: A dataset for improving scientific comprehension of large vision-language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14369-14387, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 4", + "[27] Jaeyoung Kim, Jongho Lee, Hong-Jun Choi, Ting-Yao Hsu, Chieh-Yang Huang, Sungchul Kim, Ryan Rossi, Tong Yu, Clyde Lee Giles, Ting-Hao 'Kenneth' Huang, and Sungchul Choi. Multi-LLM collaborative caption generation in scientific documents. In AI for Research and Scalable, Efficient Systems, pages 142–160, Singapore, 2025. Springer Nature Singapore. 2", + "[28] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. Sparks of artificial general intelligence: Early experiments with GPT-4, 2023. 2, 6", + "[29] Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad, Stephanie Fu, Adrian Rodriguez-Munoz, Shivam Duggal, Phillip Isola, and Antonio Torralba. A vision check-up for language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14410-14419, 2024.", + "[30] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with GPT-4, 2023. 2, 6", + "[31] Lingjiao Chen, Matei Zaharia, and James Zou. How Is ChatGPT's Behavior Changing Over Time? Harvard Data Science Review, 6(2), 2024. https://hdsr.mitpress.mit.edu/pub/y95zitmz.2" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Sagi Polaczek, Yuval Alaluf, Elad Richardson, Yael Vinker, and Daniel Cohen-Or. NeuralSVG: An implicit representation for text-to-vector generation, 2025. 2", + "[33] Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. IconShop: Text-guided vector icon synthesis with autoregressive transformers. ACM Trans. Graph., 42(6), 2023. 2", + "[34] Ajay Jain, Amber Xie, and Pieter Abbeel. VectorFusion: Text-to-SVG by abstracting pixel-based diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1911-1920, 2023.", + "[35] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. CLIPDraw: Exploring text-to-drawing synthesis through language-image encoders. In NeurIPS, 2022. 2", + "[36] Henrik Voigt, Kai Lawonn, and Sina Zarrieß. Plots made quickly: An efficient approach for generating visualizations from natural language queries. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 12787-12793, Torino, Italia, 2024. ELRA and ICCL. 2", + "[37] Yuyu Luo, Nan Tang, Guoliang Li, Chengliang Chai, Wenbo Li, and Xuedi Qin. Synthesizing natural language to visualization (nl2vis) benchmarks from NL2SQL benchmarks. In Proceedings of the 2021 International Conference on Management of Data, page 1235-1247, New York, NY, USA, 2021. Association for Computing Machinery.", + "[38] Jock Mackinlay. Automating the design of graphical presentations of relational information. ACM Trans. Graph., 5(2): 110-141, 1986.", + "[39] Steven F. Roth, John Kolojejchick, Joe Mattis, and Jade Goldstein. Interactive graphic design using automatic presentation knowledge. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, page 112-117, New York, NY, USA, 1994. Association for Computing Machinery.", + "[40] Yang Wu, Yao Wan, Hongyu Zhang, Yulei Sui, Wucai Wei, Wei Zhao, Guandong Xu, and Hai Jin. Automated data visualization from natural language via large language models: An exploratory study. Proc. ACM Manag. Data, 2(3), 2024. 2", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 2", + "[42] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021.", + "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents, 2022. 2, 7", + "[44] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. CogView: Mastering text-to-image generation via transformers. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 19822-19835, 2021." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17802", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. CogView2: Faster and better text-to-image generation via hierarchical transformers. In NeurIPS, 2022. 2", + "[46] Juan A. Rodriguez, David Vázquez, Issam H. Laradji, Marco Pedersoli, and Pau Rodríguez. FigGen: Text to scientific figure generation. In The First Tiny Papers Track at ICLR 2023, Tiny Papers @ ICLR 2023, Kigali, Rwanda, May 5, 2023. OpenReview.net, 2023. 2", + "[47] Juan A. Rodriguez, David Vázquez, Issam H. Laradji, Marco Pedersoli, and Pau Rodríguez. OCR-VQGAN: Taming textwithin-image generation. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2023, Waikoloa, HI, USA, January 2-7, 2023, pages 3678-3687. IEEE, 2023. 2", + "[48] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139-144, 2020. 2", + "[49] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 2", + "[50] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, Lille, France, 2015. PMLR. 2", + "[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2", + "[52] Yiwei Hu, Paul Guerrero, Milos Hasan, Holly Rushmeier, and Valentin Deschaintre. Generating procedural materials from text or image prompts. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2", + "[53] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12):nwae403, 2024. 2", + "[54] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3", + "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3", + "[56] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 3" + ], + "bbox": [ + 89, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[57] Aaron Grattaftiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, Danny Wyatt, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Francisco Guzmán, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Govind Thattai, Graeme Nail, Gregoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel Kloumann, Ishan Misra, Ivan Evtimov, Jack Zhang, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala, Karthik Prasad, Kartikeya Upasani, Kate Plawiak, Ke Li, Kenneth Heafield, Kevin Stone, Khalid El-Arini, Krithika Iyer, Kshitiz Malik, Kuenley Chiu, Kunal Bhalla, Kushal Lakhotia, Lauren Rantala-Yeary, Laurens van der Maaten, Lawrence Chen, Liang Tan, Liz Jenkins, Louis Martin, Lovish Madaan, Lubo Malo, Lukas Blecher, Lukas Landzaat, Luke de Oliveira, Madeline Muzzi, Mahesh Pasupuleti, Mannat Singh, Manohar Paluri, Marcin Kardas, Maria Tsimpoukelli, Mathew Oldham, Mathieu Rita, Maya Pavlova, Melanie Kambadur, Mike Lewis, Min Si, Mitesh Kumar Singh, Mona Hassan Naman Goyal Narjes Torabi Nikolay Bashlykov Nikolay Bogoychev,Niladri Chatterji Ning Zhang,Olivier Duchenne Onur Celebi Patrick Alrassy Pengchuan Zhang,Pengwei Li Petar Vasic Peter WengPrajjwal BhargavaPratik Dubal Praveen Krishnan Punit Singh Koura Puxin Xu Qing He Qingxiao Dong Ragavan Srinivasan Raj Ganapathy Ramon Calderer Ricardo Silveira Cabral Roberta Raileanu Rohan Maheswari Rohit Girdhar Rohit Patel Romain SauvestreRonnie Polidoro Roshan Sumbaly Ross Taylor Ruan Silva Rui Hou Rui Wang Saghar Hosseini Sahana Chennabasappa Sanjay Singh Sean Bell Seohyun Sonia Kim,Sergey Edunov Shaoliang Nie Sharan Narang Sharath Rarpathy Sheng Shen Shengye Wan Shruti BhosaleShun ZhangSimon Vandenhende Soumya Batra Spencer WhitmanSten Sootla Stephane Collot Suchin Gururangan Sydney Borodinsky Tamar Herman Tara Fowler Tarek Sheasha Thomas Georgiou Thomas Scialom Tobias Speckbacher Todor Mihaylov Tong Xiao Ujjwal Karn,Vedanuj Goswami Vibhor Gupta,Vignesh Ramanathan,Viktor Kerkez Vincent Gonguet,Virginie Do,Vish VogetiVitor Albiero,Vladan Petrovic Weiwei Chu Wenhan Xiong,Wenyin Fu Whitney", + "bbox": [ + 516, + 90, + 906, + 897 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17803", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "Meers, Xavier Martinet, Xiaodong Wang, Xiaofang Wang, Xiaoqing Ellen Tan, Xide Xia, Xinfeng Xie, Xuchao Jia, Xuewei Wang, Yaelle Goldschlag, Yashesh Gaur, Yasmine Babaei, Yi Wen, Yiwen Song, Yuchen Zhang, Yue Li, Yuning Mao, Zacharie Delpierre Coudert, Zheng Yan, Zhengxing Chen, Zoe Papakipos, Aaditya Singh, Aayushi Srivastava, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adithya Gangidi, Adolfo Victoria, Ahuva Goldstand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alexei Baevski, Allie Feinstein, Amanda Kallet, Amit Sangani, Amos Teo, Anam Yunus, Andrei Lupu, Andres Alvarado, Andrew Caples, Andrew Gu, Andrew Ho, Andrew Poulton, Andrew Ryan, Ankit Ramchandani, Annie Dong, Annie Franco, Anuj Goyal, Aparajita Saraf, Arkabandhu Chowdhury, Ashley Gabriel, Ashwin Bharambe, Assaf Eisenman, Azadeh Yazdan, Beau James, Ben Maurer, Benjamin Leonhardi, Bernie Huang, Beth Loyd, Beto De Paola, Bhargavi Paranjape, Bing Liu, Bo Wu, Boyu Ni, Braden Hancock, Bram Wasti, Brandon Spence, Brani Stojkovic, Brian Gamido, Britt Montalvo, Carl Parker, Carly Burton, Catalina Mejia, Ce Liu, Changhan Wang, Changkyu Kim, Chao Zhou, Chester Hu, Ching-Hsiang Chu, Chris Cai, Chris Tindal, Christoph Feichtenhofer, Cynthia Gao, Damon Civin, Dana Beaty, Daniel Kreymer, Daniel Li, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le, Dustin Holland, Edward Dowling, Eissa Jamil, Elaine Montgomery. Eleonora Presani, Emily Hahn, Emily Wood, Eric-Tuan Le. Erik Brinkman Esteban Arcaute Evan Dunbar Evan Smothers. Fei Sun Felix Kreuk Feng Tian Filippos Kokkinos. Firat Ozgenel Francesco Caggioni Frank Kanayet Frank Seide,Gabriela Medina Florez,Gabriella Schwarz,Gada Badeer Georgia Swee Gil Halpern Grant Herman Grigory SizovGuangyi,ZhangGuna LakshminarayananHakan Inan Hamid Shojanazeri Han Zou Hannah Wang Hanwen Zha Haroun Habeeb Harrison Rudolph Helen Suk Henry Aspegren Hunter Goldman Hongyuan Zhan Ibrahim Damlaj Igor Molybog,Igor Tufanov Ilias Leontiadis Irina-Elena Veliche Itai Gat Jake WeissmanJames GeboskiJames Kohli Janice LamJaphet AsherJean-Baptiste GayaJeff MarcusJeff TangJennifer Chan Jenny Zhen Jeremy Reizenstein Jeremy Teboul Jessica Zhong Jian Jin Jingyi Yang Joe Cummings Jon Carvill Jon Shepard Jonathan McPhie Jonathan Torres Josh Ginsburg Junjie Wang Kai Wu Kam Hou U Karan Saxena Kartikay Khandelwal Katayoun Zand Kathy Matosich Kaushik Veeraraghavan Kelly Michelena Keqian Li Kiran Jagadeesh Kun HuangKunal ChawlaKyle HuangLailin Chen Lakshya Garg Lavender A Leandro Silva Lee Bell Lei Zhang Liangpeng Guo Licheng Yu Liron Moshkovich Luca Wehrstedt Madian Khabsa Manav Avalani Manish Bhatt Martynas Mankus Matan Hasson Matthew Lennie Matthias Reso Maxim Groshev Maxim Naumov Maya Lathi. Meghan Keneally Miao Liu Michael L. Seltzer Michal Valko Michelle Restrepo Mihir Patel Mik Vyatskov Mikayel Samvelyan Mike Clark Mike Macey Mike Wang Miquel Jubert Hermoso Mo Metanat Mohammad Rastegari Munish Bansal Nandhini Santhanam Natascha Parks Natasha White Navyata Bawa Nayan Singhal Nick Egebo Nicolas Usunier Nikhil Mehta Nikolay Pavlovich Laptev Ning Dong Nor", + "bbox": [ + 122, + 90, + 485, + 897 + ], + "page_idx": 11 + }, + { + "type": "text", + "text": "man Cheng, Oleg Chernoguz, Olivia Hart, Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pavan Balaji, Pedro Rittner, Philip Bontrager, Pierre Roux, Piotr Dollar, Polina Zvyagina, Prashant Ratanchandani, Pritish Yuvraj, Qian Liang, Rachad Alao, Rachel Rodriguez, Rafi Ayub, Ragtoham Murthy, Raghu Nayani, Rahul Mitra, Rangaprabhu Parthasarathy, Raymond Li, Rebekkah Hogan, Robin Battey, Rocky Wang, Russ Howes, Rudy Rinott, Sachin Mehta, Sachin Siby, Sai Jayesh Bondu, Samyak Datta, Sara Chugh, Sara Hunt, Sargun Dhillon, Sasha Sidorov, Satadru Pan, Saurabh Mahajan, Saurabh Verma, Seiji Yamamoto, Sharadh Ramaswamy, Shaun Lindsay, Shaun Lindsay, Sheng Feng, Shenghao Lin, Shengxin Cindy Zha, Shishir Patil, Shiva Shankar, Shuqiang Zhang, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe, Soumith Chintala, Stephanie Max, Stephen Chen, Steve Kehoe, Steve Satterfield, Sudarshan Govindaprasad, Sumit Gupta, Summer Deng, Sungmin Cho, Sunny Virk, Suraj Subramanian, Sy Choudhury, Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Koehler, Thomas Robinson, Tianhe Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victoria Ajayi, Victoria Montanez, Vijai Mohan, Vinay Satish Kumar, Vishal Mangla, Vlad Ionescu, Vlad Poenaru, Vlad Tiberiu Mihailescu, Vladimir Ivanov, Wei Li, Wenchen Wang, Wenwen Jiang, Wes Bouaziz, Will Constable, Xiaocheng Tang, Xiaojian Wu, Xiaolan Wang, Xilun Wu, Xinbo Gao, Yaniv Kleinman, Yanjun Chen, Ye Hu, Ye Jia, Ye Qi, Yenda Li, Yilin Zhang, Ying Zhang, Yossi Adi, Youngjin Nam, Yu, Wang, Yu Zhao, Yuchen Hao, Yundi Qian, Yunlu Li, Yuzi He, Zach Rait. Zachary DeVito, Zef Rosnbrick. Zhaoduo Wen. Zhenyu Yang. Zhiwei Zhao and Zhiyu Ma. The llama 3 herd of models. 2024. 3. 4", + "bbox": [ + 545, + 90, + 908, + 532 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735-1780, 1997. 3", + "[59] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network, 2015. 3", + "[60] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. CLIP-KD: An empirical study of CLIP model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15952-15962, 2024. 3", + "[61] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter, 2020. 3", + "[62] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and efficient foundation language models, 2023. 4", + "[63] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 4", + "[64] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, Ji Ma, Jiaqi Wang, Xiaoyi Dong, Hang Yan, Hewei Guo, Conghui He, Botian Shi, Zhenjiang Jin, Chao Xu, Bin" + ], + "bbox": [ + 514, + 536, + 906, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "17804", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Wang, Xingjian Wei, Wei Li, Wenjian Zhang, Bo Zhang, Pinlong Cai, Licheng Wen, Xiangchao Yan, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. How far are we to GPT-4V? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 4", + "[65] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, Thomas Unterthiner, Daniel Keysers, Skanda Koppula, Fangyu Liu, Adam Grycner, Alexey Gritsenko, Neil Houlsby, Manoj Kumar, Keran Rong, Julian Eisenschlos, Rishabh Kabra, Matthias Bauer, Matko Bošnjak, Xi Chen, Matthias Minderer, Paul Voigtlaender, Ioana Bica, Ivana Balazevic, Joan Puigcerver, Pinelopi Papalampidi, Olivier Henaff, Xi Xiong, Radu Soricut, Jeremiah Harmsen, and Xiaohua Zhai. PaliGemma: A versatile 3B VLM for transfer, 2024. 4", + "[66] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 4", + "[67] Stephanie Fu, Netanel Yakir Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. DreamSim: Learning new dimensions of human visual similarity using synthetic data. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4, 18", + "[68] Shobhita Sundaram, Stephanie Fu, Lukas Muttenthaler, Netei Yakir Tamir, Lucy Chai, Simon Kornblith, Trevor Darrell, and Phillip Isola. When does perceptual alignment benefit vision representations? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 4, 18", + "[69] Mikołaj Binkowski, Dougal J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations, 2018. 4", + "[70] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 4", + "[71] Aryaz Eghbali and Michael Pradel. CrystalBLEU: Precisely and efficiently measuring the similarity of code. In Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering, New York, NY, USA, 2023. Association for Computing Machinery. 4", + "[72] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 4", + "[73] Peter Stanchev, Weiyue Wang, and Hermann Ney. EED: Extended edit distance measure for machine translation. In Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1), pages 514-520, Florence, Italy, 2019. Association for Computational Linguistics. 4" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 5", + "[75] Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, Kai Dang, Yang Fan, Yichang Zhang, An Yang, Rui Men, Fei Huang, Bo Zheng, Yibo Miao, Shanghaoran Quan, Yunlong Feng, Xingzhang Ren, Xuancheng Ren, Jingren Zhou, and Junyang Lin. Qwen2.5-Coder technical report, 2024. 6, 15", + "[76] Michael Saxon, Fatima Jahara, Mahsa Khoshnoodi, Yujie Lu, Aditya Sharma, and William Yang Wang. Who evaluates the evaluations? objectively scoring text-to-image prompt coherence metrics with t2IScore score (TS2). In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 6", + "[77] Jordan J. Louviere, Terry N. Flynn, and A. A. J. Marley. Best—Worst Scaling: Theory, Methods and Applications. Cambridge University Press, 2015. 6", + "[78] Svetlana Kiritchenko and Saif M. Mohammad. Capturing reliable fine-grained sentiment associations by crowdsourcing and best-worst scaling. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 811-817, San Diego, California, 2016. Association for Computational Linguistics. 6", + "[79] Svetlana Kiritchenko and Saif Mohammad. Best-worst scaling more reliable than rating scales: A case study on sentiment intensity annotation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 465-470, Vancouver, Canada, 2017. Association for Computational Linguistics. 6", + "[80] Bryan K. Orme. MaxDiff analysis: Simple counting, individual-level logit, and HB. Sawtooth Software Research Paper Series, 2009. 6", + "[81] Bruce Schneier. Applied cryptography - protocols, algorithms, and source code in C, 2nd Edition. Wiley, 1996. 7", + "[82] Kyle Lo, Lucy Lu Wang, Mark Neumann, Rodney Kinney, and Daniel Weld. S2ORC: The semantic scholar open research corpus. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4969-4983, Online, 2020. Association for Computational Linguistics. 8", + "[83] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 8", + "[84] Anas Awadalla, Le Xue, Manli Shu, An Yan, Jun Wang, Senthil Purushwalkam, Sheng Shen, Hannah Lee, Oscar Lo, Jae Sung Park, Etash Kumar Guha, Silvio Savarese, Ludwig Schmidt, Yejin Choi, Caiming Xiong, and Ran Xu. BLIP3-KALE: Knowledge augmented large-scale dense captions. In Synthetic Data for Computer Vision Workshop @ CVPR 2025, 2025." + ], + "bbox": [ + 514, + 90, + 906, + 898 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "17805", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, Yuyin Zhou, and Cihang Xie. What if we recapture billions of web images with LLaMA-3?, 2024. 8", + "[86] R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 1735-1742, 2006. 8", + "[87] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. ImageBind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15180-15190, 2023. 8", + "[88] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, Cai Wan Zhang, Zhifeng Li, Wei Liu, and Li Yuan. LanguageBind: Extending video-language pretraining to n-modality by language-based semantic alignment. In The Twelfth International Conference on Learning Representations, 2024. 8", + "[89] OpenAI. Learning to reason with LLMs, 2024. 15", + "[90] DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu, Shengfeng Ye, Tao Yun, Tian Pei, Tianyu Sun, T. Wang, Wangding Zeng, Wanjia Zhao, Wen Liu, Wenfeng Liang, Wenjun Gao, Wenqin Yu, Wentao Zhang, W. L. Xiao, Wei An, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaotao Nie, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xinyu Yang, Xinyuan Li, Xuecheng Su, Xuheng Lin, X. Q. Li, Xiangyue Jin, Xiaojin Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxiang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y. K. Li, Y. Q. Wang, Y. X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yuan Ou, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 13 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zhou, Y. X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning, 2025. 15", + "[91] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 15", + "[92] Urbano Lorenzo-Seva and Jos M. F. ten Berge. Tucker's congruence coefficient as a meaningful index of factor similarity. Methodology: European Journal of Research Methods for the Behavioral and Social Sciences, 2(2):57-64, 2006. 15", + "[93] Jonas Belouadi and Steffen Eger. UScore: An effective approach to fully unsupervised evaluation metrics for machine translation. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 358-374, Dubrovnik, Croatia, 2023. Association for Computational Linguistics. 15", + "[94] Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M. Meyer, and Steffen Eger. MoverScore: Text generation evaluating with contextualized embeddings and earth mover distance. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 563-578, Hong Kong, China, 2019. Association for Computational Linguistics.", + "[95] Wei Zhao, Goran Glavaš, Maxime Peyrard, Yang Gao, Robert West, and Steffen Eger. On the limitations of cross-lingual encoders as exposed by reference-free machine translation evaluation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1656–1671, Online, 2020. Association for Computational Linguistics.", + "[96] Yurun Song, Junchen Zhao, and Lucia Specia. SentSim: Crosslingual semantic evaluation of machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3143-3156, Online, 2021. Association for Computational Linguistics. 15", + "[97] Y. Rubner, C. Tomasi, and L.J. Guibas. A metric for distributions with applications to image databases. In Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271), pages 59-66, 1998. 15", + "[98] Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to document distances. In Proceedings of the 32nd International Conference on Machine Learning, pages 957–966, Lille, France, 2015. PMLR. 15" + ], + "bbox": [ + 516, + 92, + 906, + 835 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "17806", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/5102f011-579e-4213-9591-ec3e5e763b92_model.json b/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/5102f011-579e-4213-9591-ec3e5e763b92_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2b79109d1e45627294c496db1549135d01ce91a1 --- /dev/null +++ b/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/5102f011-579e-4213-9591-ec3e5e763b92_model.json @@ -0,0 +1,2483 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.819, + 0.154 + ], + "angle": 0, + "content": "TikZero: Zero-Shot Text-Guided Graphics Program Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.179, + 0.871, + 0.278 + ], + "angle": 0, + "content": "Jonas Belouadi* Eddy Ilg† Margret Keuper*,‡ Hideki Tanaka§ Masao Utiyama§ Raj Dabre§ Steffen Eger† Simone Ponzetto* \nUniversity of Mannheim, Germany* University of Technology Nuremberg, Germany† Max Planck Institute for Informatics, Saarland Informatics Campus, Germany‡ National Institute of Information and Communications Technology, Japan§ jonas.belouadi@uni-mannheim.de" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.311, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.485, + 0.645 + ], + "angle": 0, + "content": "Automatically synthesizing figures from text captions is a compelling capability. However, achieving high geometric precision and editability requires representing figures as graphics programs in languages like TikZ, and aligned training data (i.e., graphics programs with captions) remains scarce. Meanwhile, large amounts of unaligned graphics programs and captioned raster images are more readily available. We reconcile these disparate data sources by presenting TikZERO, which decouples graphics program generation from text understanding by using image representations as an intermediary bridge. It enables independent training on graphics programs and captioned images and allows for zero-shot text-guided graphics program synthesis during inference. We show that our method substantially outperforms baselines that can only operate with caption-aligned graphics programs. Furthermore, when leveraging caption-aligned graphics programs as a complementary training signal, TikZERO matches or exceeds the performance of much larger models, including commercial systems like GPT-4o. Our code, datasets, and select models are publicly available.1" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.674, + 0.222, + 0.691 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Graphics programming languages offer distinct advantages over low-level vector formats (PDF, SVG) or raster image formats by representing visual concepts as high-level programs that preserve semantics, remain human-interpretable, and allow manual editing. These properties are particularly valuable in academia, where specialized graphics programming languages like TikZ [1] are popular for creating complex figures with high expressivity. However, this comes with a steep learning curve, as seen on the TEX Stack Exchange\\(^2\\) (TEX.SE), where nearly \\(10\\%\\) of questions concern TikZ and make it the most frequently discussed topic on the platform [2, 3]." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.309, + 0.902, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.577, + 0.907, + 0.648 + ], + "angle": 0, + "content": "Figure 1. Qualitative comparison of our \\(\\mathrm{TikZero}^+\\) model (last two columns) and the end-to-end trained baseline \\(\\mathrm{AUTOMATikZ}_{\\mathrm{v2}}\\) (LLM; first two columns) on text-guided graphics program synthesis with TikZ. Our method generates outputs that more closely follow the given captions. Example program listings are in Supp. Mat. F." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.825 + ], + "angle": 0, + "content": "With recent advances in generative AI, simplifying the creation of graphics programs has become increasingly feasible. Belouadi et al. [2] introduce \\(\\mathrm{D}\\mathsf{E}\\mathrm{T}\\mathsf{1}\\kappa \\mathrm{Z}\\mathsf{I}\\mathsf{F}\\mathsf{Y}\\), an inverse graphics model that generates TikZ programs from images and hand-drawn sketches. However, creating these visual inputs stays cumbersome, motivating alternative input modalities such as natural language. While Belouadi et al. [3] propose AutomATikZ, a text-guided synthesis model for TikZ programs trained end-to-end on an aligned caption-program corpus, its performance remains limited (cf. Fig. 1) [4, 5]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We identify insufficient training data as the primary limitation. Unlike inverse graphics models such as \\(\\mathrm{DETikZIFY}\\), which are inherently self-supervised (trained by being conditioned on compiled representations of their output programs) and can access sufficient training data (cf. Fig. 2), end-to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.38, + 0.888 + ], + "angle": 0, + "content": "1https://github.com/potamides/DeTikZify" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.889, + 0.32, + 0.9 + ], + "angle": 0, + "content": "2https://tex.stackexchange.com" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.38, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17793" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.483, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.293, + 0.485, + 0.391 + ], + "angle": 0, + "content": "Figure 2. Illustration of training data availability for graphics program synthesis. \\(\\mathrm{DETIKZIFY}\\) can leverage all graphics programs for training but lacks text guidance, while \\(\\mathrm{AUTOMATIKZ}\\) is constrained to the small intersection of captioned graphics programs, resulting in limited performance. Our approach, \\(\\mathrm{TIKZERO}\\), trains independently on both graphics programs and captioned images, enabling more effective use of available data and yielding superior results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.463 + ], + "angle": 0, + "content": "end text-guided models like AUTOMATIKZ require graphics programs paired with captions, substantially reducing the available data pool (cf. Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.464, + 0.485, + 0.689 + ], + "angle": 0, + "content": "To address this challenge, we decouple the graphics program generation component from text understanding, enabling independent training on graphics programs and captioned images without requiring paired data (cf. Fig. 2). Our approach first trains an inverse graphics model conditioned on image patch embeddings from a vision encoder [6]. We then train an adapter network that generates synthetic image patch embeddings from captions. This adapter training relies solely on captioned images, effectively circumventing resource limitations and enabling zero-shot (in the sense that no aligned caption-program examples are involved in the training process) text-guided graphics program synthesis [7]. We demonstrate that this approach, to which we refer as \\(\\mathrm{T_{IK}ZERO}\\), outperforms previous state-of-the-art methods (cf. Fig. 1). Our key contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.69, + 0.484, + 0.75 + ], + "angle": 0, + "content": "(i) A novel two-stage architecture, \\(\\mathrm{TikZERO}\\), which addresses the low-resource challenge in text-guided graphics program synthesis by aligning representation spaces rather than relying on aligned data." + }, + { + "type": "text", + "bbox": [ + 0.084, + 0.75, + 0.484, + 0.841 + ], + "angle": 0, + "content": "(ii) The \\(\\mathrm{DATiKZ_{v3}}\\) dataset, comprising over 450k TiKZ graphics programs with roughly 170k captioned samples. Using this dataset, we train both \\(\\mathrm{TiKZERO}\\) and \\(\\mathrm{AUTOMATiKZ_{v2}}\\) (an updated version of \\(\\mathrm{AUTOMATiKZ}\\)) on the same source data and show that \\(\\mathrm{TiKZERO}\\) outperforms \\(\\mathrm{AUTOMATiKZ}\\), \\(\\mathrm{AUTOMATiKZ_{v2}}\\), and other end-to-end trained baselines." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.841, + 0.482, + 0.902 + ], + "angle": 0, + "content": "(iii) An enhanced model, \\(\\mathrm{T_{IK}Z_{ZERO} + }\\), combining \\(\\mathrm{T_{IK}Z_{ZERO}}\\) with the end-to-end fine-tuning of \\(\\mathrm{AUTOMAT_{IK}Z_{v2}}\\), which surpasses larger baselines and matches the performance of commercial models like GPT-4o [8] on key metrics." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.69, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.908, + 0.343 + ], + "angle": 0, + "content": "Inverse Graphics Program Synthesis Inverse graphics, i.e., synthesizing a graphics program to reproduce a visual target, represents a specialized instance of neural program synthesis [9-11]. Deep learning models have shown remarkable success in this domain [12-14], with Vision-Language Models (VLMs) increasingly gaining prominence [2, 15-17]. While controlled experimental studies often rely on synthetic datasets [13, 15, 17-20], real-world applications typically leverage more complex and diverse human-created data [2, 21-25], highlighting the importance of data availability. In scientific contexts, TikZ has emerged as a popular choice due to its versatility, expressiveness, and widespread adoption in academic circles [2, 21-23, 25]. Although these approaches are not tailored to text-guided generation, we incorporate key elements from them into our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.909, + 0.637 + ], + "angle": 0, + "content": "Text-Guided Graphics Program Synthesis Current text-guided approaches to graphics program synthesis remain limited, mainly because of the scarcity of captioned graphics programs outlined in Sec. 1, but also because of the difficulty of generating synthetic data with human-like captions [26, 27]. Researchers interested in this capability currently rely on the emerging capabilities of large commercial models such as GPT-4o [8, 28–30], which raises concerns about accessibility, reproducibility, and computational cost [31]. In contrast, related domains like vector graphics generation [24, 32–35] and NL2Vis [36–40] have shown more progress. Similar to inverse graphics, these fields increasingly incorporate large language models (LLMs) [24, 36, 40]. However, vector graphics approaches typically generate only low-level Bezier curves, limiting output complexity [2, 32, 33], and NL2Vis focuses exclusively on data visualization with a restricted set of visualization types [40]. More complex applications, such as generating arbitrary scientific figures from captions with TikZ, remain underexplored—a gap we address in this work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Text-to-Image Generation \\(\\mathrm{TiKZero}\\) shares conceptual and architectural similarities with several text-to-image generation methods [41-45]. Rodriguez et al. [46, 47] explore generative adversarial networks [48, 49] and diffusion models [50, 51] for scientific figure generation, but these approaches are tied to raster images, which are not ideal for representing scientific figures. Ramesh et al. [43] propose a two-stage model with independently trained prior and decoder components to generate raster images from text. Although prior networks resemble our adapters and have been used with inverse graphics models [52], they target global image embeddings containing only abstract information, which degrades performance when used with inverse graphics architectures that work best with patch-level details [53]. In contrast, our adapters specifically operate on patch-level embeddings, and we demonstrate that this improves performance compared to end-to-end trained baselines." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17794" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.907, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.319, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Figure 3. Architecture overview of \\(\\mathrm{T1KZERO}\\) during inference. Solid lines represent the standard caption-conditioned path, which flows through the text encoder into the adapter network of \\(\\mathrm{T1KZERO}\\) before connecting to the vision encoder. In certain configurations (cf. Sec. 5.2), the caption also feeds into the text decoder (depicted by dotted lines and “\\(\\bullet\\)” markers representing shortcuts). The self- and cross-attention layers (yellow) are simplified representations, omitting internal feed-forward layers and residual connections [54]. An exception is the explicit residual connection between the cross-attention and self-attention layers of the vision encoder, visualizing the gating mechanism \\(\\gamma\\) (purple). Additionally, the dashed path illustrates how the inverse graphics model generates graphics programs when conditioned on images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.441, + 0.42, + 0.457 + ], + "angle": 0, + "content": "3. The TIKZERO Model & Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.486, + 0.528 + ], + "angle": 0, + "content": "As the foundation of \\(\\mathrm{TiKZero}\\), we first develop a state-of-the-art inverse graphics model for graphics program synthesis. We then incorporate a cross-attention adapter network [55] for text guidance. Fig. 3 provides an overview of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.625 + ], + "angle": 0, + "content": "The Inverse Graphics Model Due to their demonstrated effectiveness (cf. Sec. 2), we adopt a VLM architecture for the inverse graphics model of \\(\\mathrm{T_{IK}Zero}\\). Figure 3 illustrates its inner workings (dashed lines): the model processes rasterized images and autoregressively generates their corresponding programs without involving captions at this stage." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The Adapter Network VLMs consist of two primary components: a vision encoder that produces image embeddings and a text decoder that, in our case, generates graphics programs conditioned on these embeddings. The unidirectional and localized flow of information between these components allows us to inject additional information solely into the vision encoder, thereby influencing the output of the text decoder. We exploit this property by introducing a trainable, text-conditioned adapter network that mimics the outputs of the original vision encoder. This effectively enables zero-shot generation of graphics programs conditioned on text when its outputs are fed into the decoder. In addition to circumventing the resource limitations discussed in Sec. 1, this architecture has other welcome implications: During adapter training, the text decoder, usually the largest component of the model, does not need to be loaded, resulting in efficient and fast training even with large datasets. Our adapter incorporates a lightweight text encoder for embedding captions and intro-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.909, + 0.64 + ], + "angle": 0, + "content": "duces newly initialized gated cross-attention layers [56, 57] before each vision encoder layer (cf., Fig. 3). The keys and values derive from the final text encoder representations, while the queries originate from a trainable probe used instead of image inputs. The gates \\(\\gamma\\) allow the model to learn at which layers and to what extent information from the text encoder should flow into the vision encoder. Contrary to existing literature, which often employs tanh gates [58] that initialize to zero (indicating no information flow), we find that using sigmoid gates (0.5 at initialization) accelerates training convergence since, in our case, only little information originates from the vision encoder inputs (i.e., the probe). We ablate the gates and the probe in Supp. Mat. A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.909, + 0.753 + ], + "angle": 0, + "content": "Training Objective Given a caption-image dataset for training, we first embed the patches \\( p \\in \\mathbf{p} \\) of image \\( i \\) using the unmodified vision encoder \\( \\mathbf{M} \\) of our VLM. Subsequently, we incorporate the cross-attention adapter to obtain the modified encoder \\( \\widehat{\\mathbf{M}} \\), which we then distill on these image patch embeddings conditioned solely on the caption \\( t \\) and probe \\( \\hat{t} \\) [59]. This leads to the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.763, + 0.907, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d i s t}} = \\frac {1}{| p |} \\sum_ {p \\in p} \\operatorname {d i s t} \\left(\\mathrm {M} _ {\\theta} (p \\mid i), \\widehat {\\mathrm {M}} _ {\\theta , \\hat {\\theta}} (p \\mid \\hat {t}, t)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.911, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathrm{dist}(\\pmb{x},\\pmb{y})\\) represents a distance metric. Following common practices in model distillation [60, 61], we experiment with cosine distance and mean squared error. Here, \\(\\theta\\) denotes the original model parameters that remain fully frozen, while \\(\\hat{\\theta}\\) represents the adapter parameters of which the cross-attention layers and the image probe are trainable." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17795" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.089, + 0.454, + 0.203 + ], + "angle": 0, + "content": "
SourceDATIKZDATIKZv2DATIKZv3
curated98115663646
TEX.SE292383060942654
arXiv85656326450407851
artificial195719582256
all117832360583456469
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.213, + 0.486, + 0.258 + ], + "angle": 0, + "content": "Table 1. Breakdown of the number of unique \\(\\mathrm{Ti}k\\mathrm{Z}\\) graphics in \\(\\mathrm{DATI}\\kappa \\mathrm{Z}_{\\mathrm{v3}}\\) compared to its predecessors \\(\\mathrm{DATI}\\kappa \\mathrm{Z}\\) and \\(\\mathrm{DATI}\\kappa \\mathrm{Z}_{\\mathrm{v2}}\\). Qualitative examples can be found in Supp. Mat. F." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.285, + 0.345, + 0.304 + ], + "angle": 0, + "content": "4. Datasets & Model Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.485, + 0.375 + ], + "angle": 0, + "content": "We introduce \\(\\mathrm{DATikZ_{v3}}\\), a novel dataset of TikZ graphics programs designed to support the training and evaluation of \\(\\mathrm{TIKZERO}\\). Additionally, we train \\(\\mathrm{AUTOMATikZ_{v2}}\\) as a directly comparable baseline operating on the same data source." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.484, + 0.525 + ], + "angle": 0, + "content": "The \\(\\mathrm{DATiKZ_{v3}}\\) Dataset \\(\\mathrm{DATiKZ_{v3}}\\) expands upon its predecessors \\(\\mathrm{DATiKZ_{v2}}\\) [2, 3], incorporating programs from curated repositories, TEX.SE, arXiv papers, and artificial samples (cf. Tab. 1). While previous versions focused exclusively on TikZ graphics with (v1) or without (v2) captions, \\(\\mathrm{DATiKZ_{v3}}\\) systematically extracts captions alongside TikZ graphics whenever possible to support our claims. From over 450k instances, fewer than 170k include captions, underscoring the challenges discussed in Sec. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Training TIKZERO \\(\\mathrm{T_{IK}Zero}\\) 's VLM builds upon DE\\(\\mathrm{T_{IK}ZIFY}\\) [2] by conditioning a LAMA-based text decoder [62] on patch embeddings from a S1G1P vision encoder [63]. Specifically, we combine \\(\\mathrm{LLAMA}_{3.1}\\) (8B) [57] with S1G1P SoViT (0.4B). Unlike \\(\\mathrm{DE_{IK}ZIFY}\\) and inspired by the continued ViT pretraining approach of INTERNVL 1.5 [64], we initialize the vision encoder with weights from the fine-tuned encoder of PALIGEMMA [65] and increase the input resolution to \\(420\\times 420\\) pixels. Furthermore, we fully fine-tune the vision encoder alongside the rest of the model instead of freezing it. We train on \\(\\mathrm{DAT_{IK}Z_{v3}}\\) for 5 epochs with a learning rate of 5e-5 and a batch size of 128. \\(\\mathrm{T_{IK}ZERO}\\) 's VLM consistently outperforms \\(\\mathrm{DE_{IK}ZIFY}\\), with detailed evaluation results provided in Supp. Mat. B. For the adapter network, we initialize with \\(\\mathrm{LLAMA}_{3.2}\\) (1B) as the text encoder [57] and leverage ARxivCAP [26], a dataset comprising 6.4 million scientific caption-image pairs for training. The adapter accounts for 2 billion of TIKZERO's 10 billion total parameters, with only 400 million being trainable. We train for 3 epochs with a learning rate of 1e-4 and a batch size of 512. We emphasize that this two-stage training process does not access caption-program pairs. However, we demonstrate that incorporating such aligned data in a subsequent fine-tuning step (Sec. 5.2) further enhances performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.365 + ], + "angle": 0, + "content": "Training AUTOMATIKZv2 Similar to its predecessor, AUTOMATIKZv2 is a token-conditioned LLM that uses tokenized captions as conditioning information for graphics prediction (rather than patch embeddings). We initialize AUTOMATIKZv2 in two different ways: (i) AUTOMATIKZv2 (LLM), which starts from vanilla LLAMA3.1 (8B) weights, and (ii) AUTOMATIKZv2 (VLM), which leverages TIKZERO's trained VLM (minus the vision encoder) to benefit from transfer learning [66] on its larger training corpus. Both variants employ the same hyperparameters as TIKZERO's VLM but can only utilize the caption-annotated subset of DATIKZv3 for training. Despite having access to less caption-aligned data than TIKZERO's adapter network, AUTOMATIKZv2 requires a longer training period primarily due to fine-tuning the large decoder (8 billion trainable parameters versus the adapter network's 400 million). Training requires more than two days for AUTOMATIKZv2 and 1.5 days for TIKZERO's adapter network when using eight Nvidia A100 40GB GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.38, + 0.646, + 0.398 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.909, + 0.529 + ], + "angle": 0, + "content": "Before training models on \\(\\mathrm{DAT}_{\\mathrm{IK}}\\mathrm{Z}_{\\mathrm{v3}}\\), we extract 1k samples from its captioned subset to form our test set. To mitigate data leakage from pretraining to testing, we only include instances created after the cut-off date specified by \\(\\mathrm{LLAMA}_{3.2}\\) and \\(\\mathrm{ARXIVCAP}\\). We also employ an \\(n\\)-gram matching algorithm to avoid cross-contamination with our training split [8]. For all models, the temperature is set to 0.8 and top-p to 0.95. Example outputs are provided in Fig. 1 and Supp. Mat. F." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics The multimodal nature of our task allows for various evaluation metrics in our automatic evaluations. We assess perceptual image similarity between generated outputs and references by computing DREAMSIM (DSIM) [67, 68], which correlates highly with human judgments for scientific images [2]. We also calculate the Kernel Inception Distance (KID) [69] using SigLIP image features, which evaluates the overall quality of generated figures by comparing to the distribution of reference figures. We evaluate caption similarity between generated outputs and reference captions using CLIPSCORE (CLIP) [70] with SigLIP features. To measure code similarity between generated and reference TikZ programs, we use CRYSTALBLEU (cBLEU), a BLEU variant optimized for code evaluation [71, 72], and TEX Edit Distance (TED) [2], a variant of the Extended Edit Distance [73] utilizing a TEX tokenizer. Since some metrics require that generated programs compile to images, resampling is necessary if the output contains irrecoverable errors. To quantify this, we compute the Mean Token Efficiency (MTE), defined as the \\(10\\%\\) winsorized mean of the ratio between the number of tokens in the final TikZ program and the total number of tokens generated to produce that program. For a comprehensive view of model performance, we calculate the arithmetic mean (AVG) of all previous" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "17796" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.891, + 0.25 + ], + "angle": 0, + "content": "
ModelsOriginal TextRedacted Text
DSIM↑KID↓CLIP↑cBLEU↑TED↓MTE↑AVG↑CLIP↑Ratio↑
IDEFICS 3 (8B)45.47511.42614.3270.65663.17569.55866.6284.85133.858
AUTOMATIKZ (13B)46.0331.2943.9550.38662.2485.86663.0932.96574.975
AUTOMATIKZv2 (VLM)38.31333.2030.7750.32876.98521.5950.00.28436.597
AUTOMATIKZv2 (LLM)50.5483.49115.7660.65862.30781.77582.3758.00250.753
TIKZERO (MSE)52.0245.66410.5831.72366.0779.31885.0048.23777.831
TIKZERO (Cos)52.8295.10310.0511.60365.5182.29185.5997.22671.893
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.259, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Table 2. System-level scores \\( \\times {100} \\) for \\( {\\mathrm{{T\\kappa Zero}}}_{0} \\) and baselines of comparable size and training setup. Bold and underlined values denote the best and second-best scores for each metric column, respectively. Cell shading illustrates relative score magnitudes. Arrows indicate metric directionality. Overall, \\( {\\mathrm{{T\\kappa Zero}}}_{0} \\) achieves the strongest average performance across metrics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.328, + 0.486, + 0.39 + ], + "angle": 0, + "content": "metrics. As these metrics operate on different scales, we apply min-max normalization before computing the average. Additionally, some metrics are recomputed with redacted text in the outputs as part of our analysis, cf. Sec. 6.1." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.395, + 0.479, + 0.414 + ], + "angle": 0, + "content": "5.1. Comparison against End-to-End Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.417, + 0.486, + 0.479 + ], + "angle": 0, + "content": "In our initial experiment, we evaluate the zero-shot performance of \\(\\mathrm{TiKZero}\\), trained as described in Secs. 3 & 4 using either cosine distance (Cos) or mean squared error (MSE), and compare it against end-to-end trained baselines." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.483, + 0.485, + 0.59 + ], + "angle": 0, + "content": "Baselines Besides \\(\\mathrm{AUTOMATIKZ_{v2}}\\) (LLM & VLM), which we designed as directly comparable baselines, we assess other token-conditioned models of similar and slightly larger sizes trained on \\(\\mathrm{TikZ}\\). Specifically, we evaluate \\(\\mathrm{AUTOMATIKZ}\\) (13B)\\(^3\\) the strongest original \\(\\mathrm{AUTOMATIKZ}\\) baseline [3], and the general-purpose chatbot IDEFICS 3 (8B) [22]. Additional models and details are available in Supp. Mats. A & C." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.594, + 0.486, + 0.806 + ], + "angle": 0, + "content": "Results We present the system-level metric scores in Tab. 2 (Original Text). On average, \\(\\mathrm{T_{IK}Z_{ero}}\\), trained with cosine distance, achieves the best performance with an AVG score of 85.599, closely followed by the MSE variant at 85.004. The next best model, \\(\\mathrm{AUTOMAT_{IK}Z_{v2}}\\) (LLM), scores 82.375, which is 3 percentage points (pp) lower. The remaining models exhibit a substantial performance gap, with IDEFICS 3 (8B) and \\(\\mathrm{AUTOMAT_{IK}Z}\\) (13B) falling behind by approximately 20pp and \\(\\mathrm{AUTOMAT_{IK}Z_{v2}}\\) (VLM) showing the weakest performance across all metrics, resulting in an AVG score of 0. The surprisingly poor results of \\(\\mathrm{AUTOMAT_{IK}Z_{v2}}\\) (VLM) are likely due to catastrophic forgetting [74], as the removal of the vision encoder from \\(\\mathrm{T_{IK}Z_{ERO}}\\)’s VLM necessitates reacquisition of conditioning based solely on text." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.806, + 0.486, + 0.882 + ], + "angle": 0, + "content": "As for individual metrics, our adapter-based models perform particularly well in perceptual image similarity, with \\(\\mathrm{T_{IK}Z_{\\mathrm{ero}}}\\) (Cos) outperforming the best baseline, AUTOMATIKZv2 (LLM), by 3pp on DREAMSIM. Although AUTOMATIKZv2 (LLM) outperforms \\(\\mathrm{T_{IK}Z_{\\mathrm{ero}}}\\) by 1.5pp on KID, this" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.909, + 0.585 + ], + "angle": 0, + "content": "indicates in this context that such token-conditioned models (compared to those using patch embeddings) capture the general appearance of scientific figures well but fall short in inferring visual specifics from captions. They do, however, have an edge in reproducing text from captions, which we identify as the primary reason for up to 5pp higher CLIP-Score, as noted in Sec. 6.1. Regarding code similarity, both \\(\\mathrm{TIKZERO}\\) models considerably outperform others on cBLEU. Interestingly, we observe a mild inverse correlation between cBLEU and TED. Models conditioned solely on tokenized captions tend to generate shorter, often simplified programs [3], potentially resulting in a reduced edit distance to the reference. In terms of efficiency, all models achieve an MTE of 80-85, indicating that only 2 out of 10 inferences require resampling. \\(\\mathrm{TIKZERO}\\) (Cos) is 3pp more efficient than MSE, while AUTOMATIKZ (13B), likely benefiting from its larger model size, exceeds it by another 3pp." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.909, + 0.738 + ], + "angle": 0, + "content": "In summary, training \\(\\mathrm{AUTOMATIKZ_{v2}}\\) on top of a VLM yields worse performance than training based on vanilla \\(\\mathrm{LLAMA}_{3.1}\\), indicating that effective end-to-end training can only leverage the small intersection of graphics programs and images with captions, as illustrated in Fig. 2. However, even without access to this intersection, \\(\\mathrm{TikZERO}\\) surpasses both \\(\\mathrm{AUTOMATIKZ}\\) (13B) and \\(\\mathrm{AUTOMATIKZ_{v2}}\\) (LLM) on average by being able to train on images with captions independently of graphics programs. Moreover, using a loss function based on cosine distance proves more effective than using MSE." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.851, + 0.76 + ], + "angle": 0, + "content": "5.2. Combining Adapters with Fine-Tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.911, + 0.903 + ], + "angle": 0, + "content": "In this section, we investigate whether explicitly incorporating the subset of \\(\\mathrm{DAT}\\kappa \\mathrm{Z}_{\\mathrm{v3}}\\) that includes captions into the training process of \\(\\mathrm{T}\\mathbf{I}\\kappa \\mathbf{Z}\\mathbf{E}\\mathbf{R}\\mathbf{O}\\) enhances performance. Our approach involves three incremental stages: (i) We perform a light fine-tuning of \\(\\mathrm{T}\\mathbf{I}\\kappa \\mathbf{Z}\\mathbf{E}\\mathbf{R}\\mathbf{O}\\) (Cos) end-to-end on caption-program pairs for one epoch with a low learning rate of 1e-5. Extending the training duration or increasing the learning rate does not yield further performance gains, likely due to the decoder having already reached its saturation point;" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.409, + 0.901 + ], + "angle": 0, + "content": "3Belouadi et al. [3] refer to this model as CLiMA (13B)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17797" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.089, + 0.888, + 0.248 + ], + "angle": 0, + "content": "
ModelsOriginal TextRedacted Text
DSIM↑KID↓CLIP↑cBLEU↑TED↓MTE↑AVG↑CLIP↑Ratio↑
QWEN2.5 CODER (32B)54.4735.49324.870.28559.85697.26948.59312.16448.911
GPT-4o56.4642.84431.7870.32758.51197.67579.01913.3241.905
TikZero (Cos)52.8295.10310.0511.60365.5182.29114.6587.22671.893
+ Fine-tuning (i)53.2031.79410.6870.75961.57294.85146.4976.51260.931
+ Separate Captions (ii)52.9832.90515.720.80461.3295.72246.3268.74155.608
+ Weight Resetting (iii)56.2951.83124.1771.98859.00893.05887.04311.47947.478
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.259, + 0.908, + 0.316 + ], + "angle": 0, + "content": "Table 3. System-level scores \\( \\times {100} \\) for additional baselines and \\( {\\mathrm{{TikZero}}} \\) combined with fine-tuning and token-conditioning. The scores for TIKZERO (Cos) are replicated from Tab. 2 for convenience. Bold and underlined values denote the best and second-best scores for each metric column, respectively. Cell shading illustrates relative score magnitudes. Arrows indicate metric directionality. Overall, \\( {\\mathrm{{TikZero}}}_{\\mathrm{O}}\\left( \\mathrm{{Cos}}\\right) \\) with Weight Resetting (iii) demonstrates the strongest average performance across metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.482, + 0.435 + ], + "angle": 0, + "content": "(ii) Alongside feeding captions into the adapter, we provide them separately to the text decoder in tokenized form (cf., Fig. 3); (iii) Prior to fine-tuning, we reset the decoder to its initial weights to overcome saturation, enabling us to fine-tune using the setup described in Sec. 4, which involves 5 epochs and a learning rate of \\(5\\mathrm{e} - 5\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.483, + 0.532 + ], + "angle": 0, + "content": "Baselines In addition to the baselines in Tab. 2, which remain comparable, we also evaluate larger and commercial models that serve as stronger baselines (cf. Supp. Mat. C). Specifically, we assess GPT-4o [8], which has demonstrated strong performance in generating TikZ [3, 28, 30] and QwEN2.5 CODER (32B) [75] as an open-weights model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.886 + ], + "angle": 0, + "content": "Results In Tab. 3 (Original Text), all fine-tuning setups of \\(\\mathrm{TiKZero}\\) show considerable improvement over the base version. Approaches (i) and (ii) each enhance performance by over 30pp on AVG, while approach (iii) surpasses them with an improvement of over 70pp, positioning it as the best-performing model on average, even when compared to our new baselines, with GPT-4o being 8pp lower and \\(\\mathrm{QWEN}_{2.5}\\) CODER (32B) approximately 40pp lower. Approach (i) demonstrates that direct fine-tuning yields positive effects across nearly all metrics, notably improving MTE by 12pp, TED by 4pp, and KID by 3.5pp. Approach (ii) shows similar trends but, by also incorporating tokenized captions, further improves CLIPSCORE by 5pp, closing the gap to AUTOMATIKZv2 (LLM). Interestingly, both (i) and (ii) slightly decrease performance on cBLEU, potentially due to similar reasons discussed in Sec. 5.1. However, the same cannot be said for (iii), which not only achieves the highest score on cBLEU but also ranks as the second-best on TED, trailing only 0.5pp behind GPT-4o and showcasing that it is possible to perform well on both metrics. Additionally, it increases DREAMSIM by another 3pp and CLIPSCORE by 8.5pp, competing with the much stronger baselines QWEN2.5 CODER (32B) and GPT-4o. In KID, it even surpasses them by 3.5pp and 1pp, respectively." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.887, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In summary, fine-tuning \\(\\mathrm{T_{IK}Zero}\\), especially when com" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.908, + 0.493 + ], + "angle": 0, + "content": "bined with a separate caption input and weight resetting, greatly improves performance. This illustrates that the intersection of graphics programs and images with captions, though small, provides a valuable training signal, and best performance can be achieved by making full use of both sets. The best-performing \\(\\mathrm{TiKZero}\\) model even competes with and often surpasses QwEN2.5 CODER (32B) and GPT-4o on several key metrics. Notably, the former model is more than three times larger, and the latter is often estimated at around 1.8 trillion parameters [76], making it 180 times larger." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.513, + 0.696, + 0.527 + ], + "angle": 0, + "content": "5.3. Human Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To corroborate our findings from automatic evaluation, we conduct a human annotation campaign focusing on two key properties: caption and image similarity. We employ BestWorst Scaling (BWS) [77], a comparative annotation method that yields high-quality results even with few annotators [78, 79]. We sample 100 instances from our test set and present annotators with \\( n \\)-tuples of generated figures, asking them to identify the most and least similar figure to either the reference caption or reference image. This data is then transformed into scores from -1 (poor) to 1 (excellent) by subtracting the proportion of times a figure is selected as the best from the proportion of times it is chosen as the worst [80]. For a manageable workload, we focus on \\( n = 4 \\) key models: \\( \\mathrm{TikZERO}(\\mathrm{Cos}) \\), our best-performing model from Sec. 5.1; \\( \\mathrm{AUTOMATIKZv2} \\) (LLM), its direct end-to-end trained competitor; GPT-4o, our strongest baseline; and \\( \\mathrm{TikZERO}(\\mathrm{Cos}) \\) fine-tuned using approach (iii) from Sec. 5.2, our best model overall, henceforth referred to as \\( \\mathrm{TikZERO}^+ \\) for convenience. We engage thirteen annotators and obtain six fully annotated sets per task (cf. Supp. Mat. E for more details). To assess annotator consistency, we calculate the split-half reliability (SHR) [79]. This method randomly divides all annotations into two sets, calculates scores independently, and then determines their correlation using Spearman's \\( \\rho \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "17798" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.473, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.373, + 0.487, + 0.445 + ], + "angle": 0, + "content": "Figure 4. Bivariate distributions of BWS scores (higher is better) using kernel density estimation for caption and image similarity. Along the diagonal, \\(\\mathrm{T_{IK}Z_{ERO}}\\) (Cos) achieves higher scores than \\(\\mathrm{AUTOMAT_{IK}Z_{v2}}\\) (LLM), while \\(\\mathrm{T_{IK}Z_{ERO+}}\\) and GPT-4o demonstrate superior performance compared to both." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.476, + 0.485, + 0.868 + ], + "angle": 0, + "content": "Results Fig. 4 presents kernel density estimates for the BWS scores, showing generally consistent rankings with automatic evaluations but revealing notable differences in the magnitude of gaps. For caption similarity, the ranking aligns with CLIPSCORE evaluations \\((\\rho = 1.0)\\), with \\(\\mathrm{TikZero}\\) (Cos), \\(\\mathrm{AUTOMATIKZ_{v2}}\\) (LLM), \\(\\mathrm{TikZero}+\\), and GPT-4o achieving mean scores \\(\\mu\\) of -0.25, -0.18, 0.03, and 0.4, respectively. Interestingly, humans perceive a \\(40\\%\\) smaller gap between \\(\\mathrm{AUTOMATIKZ_{v2}}\\) (LLM) and \\(\\mathrm{TikZero}\\) (Cos) than suggested by CLIPSCORE values, where \\(\\mathrm{AUTOMATIKZ_{v2}}\\) (LLM) outperforms \\(\\mathrm{TikZero}\\) (Cos) by \\(50\\%\\). This indicates humans may evaluate caption similarity differently than CLIPSCORE (cf. Sec. 6.1). For image similarity, the system order remains consistent with our DREAMSIM metric \\((\\rho = 1.0)\\), with \\(\\mathrm{AUTOMATIKZ_{v2}}\\) (LLM), \\(\\mathrm{TikZero}\\) (Cos), \\(\\mathrm{TikZero}+\\), and GPT-4o achieving \\(\\mu\\) of -0.26, -0.02, 0.01, and 0.27, respectively. However, the relative gaps between models differ: the separation between \\(\\mathrm{AUTOMATIKZ_{v2}}\\) (LLM) and \\(\\mathrm{TikZero}\\) (Cos), as well as between \\(\\mathrm{TikZero}+\\) and GPT-4o, appear more pronounced than observed with DREAMSIM. This discrepancy likely stems from BWS capturing relative preferences rather than absolute performance differences. GPT-4o is selected \\(20\\%\\) more often as the best model than \\(\\mathrm{TikZero}+\\), and \\(\\mathrm{AUTOMATIKZ_{v2}}\\) (LLM) \\(15\\%\\) more often as the worst model than \\(\\mathrm{TikZero}\\) (Cos), creating larger perceived gaps even when qualitative differences may be subtle." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "The SHR values of 0.68 for caption similarity and 0.76 for image similarity indicate moderate to strong inter-annotator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.213 + ], + "angle": 0, + "content": "agreement. We also observe a correlation between these two tasks, with segment-level \\(\\rho = 0.62\\) and system-level \\(\\rho = 0.8\\), suggesting that both evaluation dimensions capture related aspects of model performance. GPT-4o emerges as the best-performing model, aligning with its superior performance on the corresponding automatic metrics, CLIPSCORE and DSIM. Among open-source models, \\(\\mathrm{T_{IK}Z_{v2}}\\) performs best, while AUTOMATIKZv2 (LLM) ranks lowest overall." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.226, + 0.611, + 0.243 + ], + "angle": 0, + "content": "6. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.908, + 0.312 + ], + "angle": 0, + "content": "We present a comprehensive analysis, investigating the influence of typographic attacks on CLIPSCORE and examining the effectiveness of our architecture in low-resource settings, both in terms of training data and trainable parameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.32, + 0.907, + 0.337 + ], + "angle": 0, + "content": "6.1. CLIPSCORE Limitations & Typographic Attacks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.909, + 0.568 + ], + "angle": 0, + "content": "A known limitation of CLIPSCORE with text-rich images is its susceptibility to typographic attacks, where scores are disproportionately influenced by string similarity between images and captions [3, 43]. We suspect that token-conditioned models like AUTOMATIKZv2 (LLM) achieve higher CLIPSCORE values than models such as TIKZERO (Cos & MSE) primarily because they tend to visibly copy more substrings from the caption in the output image. To test this hypothesis, we apply the ROT13 substitution cipher [81] to all visible strings in the generated figures and recompute CLIPSCORE. This basic cipher replaces each letter with the 13th letter after it in the Latin alphabet. While not cryptographically secure, the ratio between the original and recomputed CLIPSCORE values should indicate the influence of string matching, i.e., higher ratios suggest less copied text and vice versa." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.911, + 0.841 + ], + "angle": 0, + "content": "Tabs. 2 & 3 (Redacted Text) present the recomputed CLIP-Score values and ratios for all evaluated models. The results reveal that most \\(\\mathrm{T_{IK}Zero}\\) models, except for fine-tuning approaches (ii) and (iii), which also condition on tokenized captions, have considerably higher ratios \\((61\\% -78\\%)\\) compared to strictly token-conditioned models \\((34\\% -51\\%)\\), supporting our hypothesis. AUTOMATIKZ (13B) is an exception, possibly due to its initially low score. Further analysis shows that with redacted text, \\(\\mathrm{AUTOMATIKZ}_{\\mathrm{V2}}\\) (LLM)'s CLIPSCORE performance drops to the same level as \\(\\mathrm{T_{IK}Zero}\\) (Cos & MSE), suggesting that string matching is the primary factor in its superior performance rather than producing better visuals—arguably a more difficult task. Nevertheless, reproducing strings is still somewhat desirable. The human oracle of our test set achieves a ratio of \\(50.8\\%\\), close to the \\(47.5\\%\\) of \\(\\mathrm{T_{IK}Zero+}\\), our best-performing model. In contrast, models like GPT-4o, with a lower ratio of \\(41.9\\%\\), may overfit to caption copying, artificially inflating the CLIPSCORE values." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.733, + 0.866 + ], + "angle": 0, + "content": "6.2. Low-Resource Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "While our adapters train efficiently on large-scale datasets, we investigate whether such extensive data is necessary for opti" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17799" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.131, + 0.089, + 0.445, + 0.21 + ], + "angle": 0, + "content": "
Intv.Training Data
100%50%25%12.5%
192.41177.47849.96756.055
287.55785.24954.94233.817
482.25447.38132.1237.914
876.54540.81629.77416.25
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.221, + 0.482, + 0.292 + ], + "angle": 0, + "content": "Table 4. AVG scores for \\(\\mathrm{TiKZero}\\) (Cos) trained on varying fractions of data and intervals of cross-attention layers. Higher scores indicate better performance. Bold and underlined values denote the best and second-best scores for the whole table, respectively. Cell shading illustrates score magnitudes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.318, + 0.485, + 0.665 + ], + "angle": 0, + "content": "mal performance. Along the same vein, we examine the impact of reducing the amount of cross-attention layers inserted into the vision encoder. We retrain \\(\\mathrm{TIKZERO}\\) (Cos) using varying fractions of the training data (\\(d \\in \\{1, \\frac{1}{2}, \\frac{1}{4}, \\frac{1}{8}\\}\\)) and insert cross-attention layers at different intervals (\\(i \\in \\{1, 2, 4, 8\\}\\)). Table 4 presents the AVG scores from this parameter grid, with detailed scores in Supp. Mat. D. Our findings reveal that utilizing the full dataset and inserting cross-attention at every layer yields the highest average performance, highlighting the benefits of maximizing both variables. Interestingly, the model's performance appears more robust to a reduction in the number of layers compared to a decrease in training data. For instance, training on only \\(\\frac{1}{8}\\)th of the data leads to a substantial performance drop of \\(36\\mathrm{pp}\\), whereas inserting cross-attention layers every 8 layers (resulting in only 3 cross-attention layers in total) causes a more modest decline of \\(16\\mathrm{pp}\\). Minimizing both variables leads to the most severe drop of over \\(75\\mathrm{pp}\\). These results validate our training setup while suggesting that incorporating additional data might further enhance performance. Given that \\(\\mathrm{ARXIVCAP}\\) extracts figures from only \\(572\\mathrm{k}\\) papers, whereas some corpora index over 200 million papers [82], there remains a lot of potential for leveraging larger datasets in future work." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.21, + 0.695 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.486, + 0.902 + ], + "angle": 0, + "content": "In this work, we demonstrate the potential of \\(\\mathrm{TiKZero}\\) and its variants for generating TiKZ graphics programs from captions. Notably, \\(\\mathrm{TiKZero}\\) does not require aligned caption-program pairs in its original formulation but instead aligns representation spaces of unaligned graphics programs and captioned images. This enables our model to leverage substantially more training data compared to end-to-end trained models that operate solely on caption-image pairs (cf. Fig. 2) while maintaining training efficiency. \\(\\mathrm{TiKZero}\\) outperforms strong end-to-end trained baselines, including our independently trained AUTOMATiKZv2 models, which use the same data pool, excluding instances they cannot process, illustrating the strengths of our approach. When extending the \\(\\mathrm{TiKZero}\\) ap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "proach with additional end-to-end training, it also compares favorably to much larger baselines and commercial systems like GPT-4o. While this enhanced approach, \\(\\mathrm{T_{IK}ZERO^{+}}\\), is no longer zero-shot by our definition, it remains a \\(\\mathrm{T_{IK}ZERO}\\) model in the sense that it operates on both sets of graphics programs and captioned images, with the added advantage of explicitly utilizing their intersection (cf. Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.909, + 0.304 + ], + "angle": 0, + "content": "These results demonstrate the benefits of designing architectures around available data and validate the approach of decoupling graphics program generation from text understanding (with optional later reconciliation through \\(\\mathrm{TiKZero^{+}}\\)). Although we demonstrate our method specifically on \\(\\mathrm{TiKZ}\\), we believe its general principles will inspire future work on related graphics program synthesis tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.909, + 0.525 + ], + "angle": 0, + "content": "Future Work Beyond scaling up our training data to explore convergence limits (cf. Sec. 6.2), we plan to investigate automatic methods for improving the quality and alignment of caption-image or caption-program pairs. This includes rewriting potentially noisy captions with LLMs and enhancing them with the visual understanding capabilities of VLMs [83-85]. We believe our approach to aligning textual and image modalities enables other promising applications for graphics program synthesis, such as editing images in latent space via textual instructions to generate modified graphics programs. Additionally, we intend to explore alternative alignment strategies beyond model distillation, including contrastive learning [86], which has successfully aligned modalities in discriminative models [6, 87, 88]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.539, + 0.618, + 0.555 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.909, + 0.732 + ], + "angle": 0, + "content": "Our evaluations include proprietary systems that operate as black boxes; their training data is unknown, and they offer no guarantees of consistent performance over time. This (i) makes addressing data leakage and cross-contamination impossible and (ii) limits the fairness and reproducibility of our experiments. Nevertheless, even under these unfavorable conditions, our open models remain competitive. Users should be aware, however, that our models may behave unpredictably, and outputs might differ from expectations. Additionally, our models do not include safeguards against potential misuse, e.g., for generating fake scientific content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.909, + 0.853 + ], + "angle": 0, + "content": "Regarding licensing of our training data, a large portion of the TikZ programs in \\(\\mathrm{DAT}_{\\mathrm{IK}}\\mathrm{Z}_{\\mathrm{v3}}\\) are licensed under permissive terms that allow redistribution. The remaining programs are distributed under the arXiv.org perpetual, non-exclusive license, which prohibits redistribution, which is why we exclude them from the public release of \\(\\mathrm{DAT}_{\\mathrm{IK}}\\mathrm{Z}_{\\mathrm{v3}}\\). However, since we release our dataset creation scripts, we encourage others to reproduce the full version independently." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.908, + 0.901 + ], + "angle": 0, + "content": "4https://creativecommons.org/licenses; https://opensource.org/licenses/mit; https://www.gnu.org/licenses/fdl-1.3.en.html; https://openai.com/policies/terms-of-use" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17800" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.248, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.381 + ], + "angle": 0, + "content": "We extend our sincere gratitude to the following individuals (in no particular order) for their valuable contributions: Christian Greisinger, Hour Kaing, Ran Zhang, Tejaswini Medi, Yanran Chen, Sotaro Takeshita, Katharina Prasse, JiWoo Kim, Christoph Leiter, Haiyue Song, and Aida Kostikova. Their assistance with our human evaluation campaign, proofreading, insightful discussions, and constructive feedback has been instrumental to our work. The first author conducted part of this research during an internship at the National Institute of Information and Communications Technology (NICT), Japan. The second to last author is supported by the Federal Ministry of Education and Research (BMBF) via the research grant METRICS4NLG and the German Research Foundation (DFG) via the Heisenberg Grant EG 375/5-1. We acknowledge computing resources provided by the state of Baden-Württemberg through bwHPC and the German Research Foundation (DFG) through grant INST 35/1597-1 FUGG. Finally, we thank the OpenMoji project for the open-source icons used throughout this work." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.407, + 0.187, + 0.423 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.432, + 0.422, + 0.446 + ], + "angle": 0, + "content": "[1] Till Tantau. The TikZ and PGF Packages, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.448, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[2] Jonas Belouadi, Simone Paolo Ponzetto, and Steffen Eger. DeTikZify: Synthesizing graphics programs for scientific figures and sketches with TikZ. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 4, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.484, + 0.573 + ], + "angle": 0, + "content": "[3] Jonas Belouadi, Anne Lauscher, and Steffen Eger. AutomataTikZ: Text-guided synthesis of scientific vector graphics with TikZ. In The Twelfth International Conference on Learning Representations, 2024. 1, 4, 5, 6, 7, 16, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.576, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[4] Leixin Zhang, Yinjie Cheng, Weihe Zhai, Steffen Eger, Jonas Belouadi, Fahimeh Moafian, and Zhixue Zhao. ScImage: How good are multimodal large language models at scientific text-to-image generation? In The Thirteenth International Conference on Learning Representations, 2025. 1, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[5] Abhay Zala, Han Lin, Jaemin Cho, and Mohit Bansal. DiagrammerGPT: Generating open-domain, open-platform diagrams via LLM planning. In First Conference on Language Modeling, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[6] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[7] Mark Palatucci, Dean Pomerleau, Geoffrey E Hinton, and Tom M Mitchell. Zero-shot learning with semantic output codes. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.402, + 0.872 + ], + "angle": 0, + "content": "[8] OpenAI. GPT-4 technical report, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Emilio Parisotto, Abdel rahman Mohamed, Rishabh Singh, Lihong Li, Dengyong Zhou, and Pushmeet Kohli. Neuro" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.432, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "symbolic program synthesis. In International Conference on Learning Representations, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Jonathan Uesato, Surya Bhupatiraju, Rishabh Singh, Abdel rahman Mohamed, and Pushmeet Kohli. Robust-Fill: Neural program learning under noisy I/O. In Proceedings of the 34th International Conference on Machine Learning, pages 990-998. PMLR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.907, + 0.302 + ], + "angle": 0, + "content": "[11] Kevin Ellis, Catherine Wong, Maxwell Nye, Mathias Sable-Meyer, Lucas Morales, Luke Hewitt, Luc Cary, Armando Solar-Lezama, and Joshua B. Tenenbaum. DreamCoder: bootstrapping inductive program synthesis with wake-sleep library learning. In Proceedings of the 42nd ACM SIGPLAN International Conference on Programming Language Design and Implementation, page 835-850, New York, NY, USA, 2021. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.303, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[12] Yaroslav Ganin, Tejas Kulkarni, Igor Babuschkin, S. M. Ali Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. In Proceedings of the 35th International Conference on Machine Learning, pages 1666-1675. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.372, + 0.906, + 0.427 + ], + "angle": 0, + "content": "[13] Kevin Ellis, Daniel Ritchie, Armando Solar-Lezama, and Josh Tenenbaum. Learning to infer graphics programs from hand-drawn images. In Thirty-second Conference on Neural Information Processing Systems, pages 6062–6071, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.428, + 0.907, + 0.494 + ], + "angle": 0, + "content": "[14] Kevin Ellis, Maxwell Nye, Yewen Pu, Felix Sosa, Josh Tenenbaum, and Armando Solar-Lezama. Write, execute, assess: Program synthesis with a REPL. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.497, + 0.906, + 0.552 + ], + "angle": 0, + "content": "[15] Peter Kulits, Haiwen Feng, Weiyang Liu, Victoria Fernandez Abrevaya, and Michael J. Black. Re-thinking inverse graphics with large language models. Transactions on Machine Learning Research, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[16] Wen-Ding Li and Kevin Ellis. Is programming by example solved by LLMs? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.907, + 0.648 + ], + "angle": 0, + "content": "[17] Shreyas Kapur, Erik Jenner, and Stuart Russell. Diffusion on syntax trees for program synthesis. In The Thirteenth International Conference on Learning Representations, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.907, + 0.734 + ], + "angle": 0, + "content": "[18] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5515-5523. Computer Vision Foundation / IEEE Computer Society, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.907, + 0.789 + ], + "angle": 0, + "content": "[19] Yonglong Tian, Andrew Luo, Xingyuan Sun, Kevin Ellis, William T. Freeman, Joshua B. Tenenbaum, and Jiajun Wu. Learning to infer and execute 3D shape programs. In International Conference on Learning Representations, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[20] Javier Camara, Javier Troya, Lola Burgueño, and Antonio Vallecillo. On the assessment of generative AI in modeling tasks: an experience report with chatgpt and UML. Softw. Syst. Model., 22(3):781-793, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[21] Hugo Laurençon, Leo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17801" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[22] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[23] Shengbang Tong, Ellis L Brown II, Penghao Wu, Sanghyun Woo, Adithya Jairam Iyer, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, Xichen Pan, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian1: A fully open, vision-centric exploration of multimodal LLMs. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[24] Juan A. Rodriguez, Abhay Puri, Shubham Agarwal, Issam H. Laradji, Pau Rodriguez, Sai Rajeswar, David Vazquez, Christopher Pal, and Marco Pedersoli. StarVector: Generating scalable vector graphics code from images and text. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 16175-16186, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.483, + 0.439 + ], + "angle": 0, + "content": "[25] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, Sam Dodge, Keen You, Zhen Yang, Aleksei Timofeev, Mingze Xu, Hong-You Chen, Jean-Philippe Fauconnier, Zhengfeng Lai, Haoxuan You, Zirui Wang, Afshin Dehghan, Peter Grasch, and Yinfei Yang. MM1.5: Methods, analysis & insights from multimodal LLM fine-tuning. In The Thirteenth International Conference on Learning Representations, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.483, + 0.538 + ], + "angle": 0, + "content": "[26] Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal ArXiv: A dataset for improving scientific comprehension of large vision-language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14369-14387, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[27] Jaeyoung Kim, Jongho Lee, Hong-Jun Choi, Ting-Yao Hsu, Chieh-Yang Huang, Sungchul Kim, Ryan Rossi, Tong Yu, Clyde Lee Giles, Ting-Hao 'Kenneth' Huang, and Sungchul Choi. Multi-LLM collaborative caption generation in scientific documents. In AI for Research and Scalable, Efficient Systems, pages 142–160, Singapore, 2025. Springer Nature Singapore. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[28] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. Sparks of artificial general intelligence: Early experiments with GPT-4, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[29] Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad, Stephanie Fu, Adrian Rodriguez-Munoz, Shivam Duggal, Phillip Isola, and Antonio Torralba. A vision check-up for language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14410-14419, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[30] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with GPT-4, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[31] Lingjiao Chen, Matei Zaharia, and James Zou. How Is ChatGPT's Behavior Changing Over Time? Harvard Data Science Review, 6(2), 2024. https://hdsr.mitpress.mit.edu/pub/y95zitmz.2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[32] Sagi Polaczek, Yuval Alaluf, Elad Richardson, Yael Vinker, and Daniel Cohen-Or. NeuralSVG: An implicit representation for text-to-vector generation, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[33] Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. IconShop: Text-guided vector icon synthesis with autoregressive transformers. ACM Trans. Graph., 42(6), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[34] Ajay Jain, Amber Xie, and Pieter Abbeel. VectorFusion: Text-to-SVG by abstracting pixel-based diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1911-1920, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.906, + 0.273 + ], + "angle": 0, + "content": "[35] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. CLIPDraw: Exploring text-to-drawing synthesis through language-image encoders. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.357 + ], + "angle": 0, + "content": "[36] Henrik Voigt, Kai Lawonn, and Sina Zarrieß. Plots made quickly: An efficient approach for generating visualizations from natural language queries. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 12787-12793, Torino, Italia, 2024. ELRA and ICCL. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.906, + 0.441 + ], + "angle": 0, + "content": "[37] Yuyu Luo, Nan Tang, Guoliang Li, Chengliang Chai, Wenbo Li, and Xuedi Qin. Synthesizing natural language to visualization (nl2vis) benchmarks from NL2SQL benchmarks. In Proceedings of the 2021 International Conference on Management of Data, page 1235-1247, New York, NY, USA, 2021. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[38] Jock Mackinlay. Automating the design of graphical presentations of relational information. ACM Trans. Graph., 5(2): 110-141, 1986." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.552 + ], + "angle": 0, + "content": "[39] Steven F. Roth, John Kolojejchick, Joe Mattis, and Jade Goldstein. Interactive graphic design using automatic presentation knowledge. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, page 112-117, New York, NY, USA, 1994. Association for Computing Machinery." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.906, + 0.609 + ], + "angle": 0, + "content": "[40] Yang Wu, Yao Wan, Hongyu Zhang, Yulei Sui, Wucai Wei, Wei Zhao, Guandong Xu, and Hai Jin. Automated data visualization from natural language via large language models: An exploratory study. Proc. ACM Manag. Data, 2(3), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.61, + 0.906, + 0.69 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[42] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[44] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. CogView: Mastering text-to-image generation via transformers. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 19822-19835, 2021." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17802" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.486, + 0.134 + ], + "angle": 0, + "content": "[45] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. CogView2: Faster and better text-to-image generation via hierarchical transformers. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.486, + 0.207 + ], + "angle": 0, + "content": "[46] Juan A. Rodriguez, David Vázquez, Issam H. Laradji, Marco Pedersoli, and Pau Rodríguez. FigGen: Text to scientific figure generation. In The First Tiny Papers Track at ICLR 2023, Tiny Papers @ ICLR 2023, Kigali, Rwanda, May 5, 2023. OpenReview.net, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.485, + 0.277 + ], + "angle": 0, + "content": "[47] Juan A. Rodriguez, David Vázquez, Issam H. Laradji, Marco Pedersoli, and Pau Rodríguez. OCR-VQGAN: Taming textwithin-image generation. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2023, Waikoloa, HI, USA, January 2-7, 2023, pages 3678-3687. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.278, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[48] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139-144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.335, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[49] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.392, + 0.485, + 0.461 + ], + "angle": 0, + "content": "[50] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, Lille, France, 2015. PMLR. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.463, + 0.484, + 0.504 + ], + "angle": 0, + "content": "[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.507, + 0.484, + 0.576 + ], + "angle": 0, + "content": "[52] Yiwei Hu, Paul Guerrero, Milos Hasan, Holly Rushmeier, and Valentin Deschaintre. Generating procedural materials from text or image prompts. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.578, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[53] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12):nwae403, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.634, + 0.484, + 0.69 + ], + "angle": 0, + "content": "[54] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.484, + 0.76 + ], + "angle": 0, + "content": "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.763, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[56] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.898 + ], + "angle": 0, + "content": "[57] Aaron Grattaftiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, Danny Wyatt, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Francisco Guzmán, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Govind Thattai, Graeme Nail, Gregoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel Kloumann, Ishan Misra, Ivan Evtimov, Jack Zhang, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala, Karthik Prasad, Kartikeya Upasani, Kate Plawiak, Ke Li, Kenneth Heafield, Kevin Stone, Khalid El-Arini, Krithika Iyer, Kshitiz Malik, Kuenley Chiu, Kunal Bhalla, Kushal Lakhotia, Lauren Rantala-Yeary, Laurens van der Maaten, Lawrence Chen, Liang Tan, Liz Jenkins, Louis Martin, Lovish Madaan, Lubo Malo, Lukas Blecher, Lukas Landzaat, Luke de Oliveira, Madeline Muzzi, Mahesh Pasupuleti, Mannat Singh, Manohar Paluri, Marcin Kardas, Maria Tsimpoukelli, Mathew Oldham, Mathieu Rita, Maya Pavlova, Melanie Kambadur, Mike Lewis, Min Si, Mitesh Kumar Singh, Mona Hassan Naman Goyal Narjes Torabi Nikolay Bashlykov Nikolay Bogoychev,Niladri Chatterji Ning Zhang,Olivier Duchenne Onur Celebi Patrick Alrassy Pengchuan Zhang,Pengwei Li Petar Vasic Peter WengPrajjwal BhargavaPratik Dubal Praveen Krishnan Punit Singh Koura Puxin Xu Qing He Qingxiao Dong Ragavan Srinivasan Raj Ganapathy Ramon Calderer Ricardo Silveira Cabral Roberta Raileanu Rohan Maheswari Rohit Girdhar Rohit Patel Romain SauvestreRonnie Polidoro Roshan Sumbaly Ross Taylor Ruan Silva Rui Hou Rui Wang Saghar Hosseini Sahana Chennabasappa Sanjay Singh Sean Bell Seohyun Sonia Kim,Sergey Edunov Shaoliang Nie Sharan Narang Sharath Rarpathy Sheng Shen Shengye Wan Shruti BhosaleShun ZhangSimon Vandenhende Soumya Batra Spencer WhitmanSten Sootla Stephane Collot Suchin Gururangan Sydney Borodinsky Tamar Herman Tara Fowler Tarek Sheasha Thomas Georgiou Thomas Scialom Tobias Speckbacher Todor Mihaylov Tong Xiao Ujjwal Karn,Vedanuj Goswami Vibhor Gupta,Vignesh Ramanathan,Viktor Kerkez Vincent Gonguet,Virginie Do,Vish VogetiVitor Albiero,Vladan Petrovic Weiwei Chu Wenhan Xiong,Wenyin Fu Whitney" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17803" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.123, + 0.092, + 0.486, + 0.898 + ], + "angle": 0, + "content": "Meers, Xavier Martinet, Xiaodong Wang, Xiaofang Wang, Xiaoqing Ellen Tan, Xide Xia, Xinfeng Xie, Xuchao Jia, Xuewei Wang, Yaelle Goldschlag, Yashesh Gaur, Yasmine Babaei, Yi Wen, Yiwen Song, Yuchen Zhang, Yue Li, Yuning Mao, Zacharie Delpierre Coudert, Zheng Yan, Zhengxing Chen, Zoe Papakipos, Aaditya Singh, Aayushi Srivastava, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adithya Gangidi, Adolfo Victoria, Ahuva Goldstand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alexei Baevski, Allie Feinstein, Amanda Kallet, Amit Sangani, Amos Teo, Anam Yunus, Andrei Lupu, Andres Alvarado, Andrew Caples, Andrew Gu, Andrew Ho, Andrew Poulton, Andrew Ryan, Ankit Ramchandani, Annie Dong, Annie Franco, Anuj Goyal, Aparajita Saraf, Arkabandhu Chowdhury, Ashley Gabriel, Ashwin Bharambe, Assaf Eisenman, Azadeh Yazdan, Beau James, Ben Maurer, Benjamin Leonhardi, Bernie Huang, Beth Loyd, Beto De Paola, Bhargavi Paranjape, Bing Liu, Bo Wu, Boyu Ni, Braden Hancock, Bram Wasti, Brandon Spence, Brani Stojkovic, Brian Gamido, Britt Montalvo, Carl Parker, Carly Burton, Catalina Mejia, Ce Liu, Changhan Wang, Changkyu Kim, Chao Zhou, Chester Hu, Ching-Hsiang Chu, Chris Cai, Chris Tindal, Christoph Feichtenhofer, Cynthia Gao, Damon Civin, Dana Beaty, Daniel Kreymer, Daniel Li, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le, Dustin Holland, Edward Dowling, Eissa Jamil, Elaine Montgomery. Eleonora Presani, Emily Hahn, Emily Wood, Eric-Tuan Le. Erik Brinkman Esteban Arcaute Evan Dunbar Evan Smothers. Fei Sun Felix Kreuk Feng Tian Filippos Kokkinos. Firat Ozgenel Francesco Caggioni Frank Kanayet Frank Seide,Gabriela Medina Florez,Gabriella Schwarz,Gada Badeer Georgia Swee Gil Halpern Grant Herman Grigory SizovGuangyi,ZhangGuna LakshminarayananHakan Inan Hamid Shojanazeri Han Zou Hannah Wang Hanwen Zha Haroun Habeeb Harrison Rudolph Helen Suk Henry Aspegren Hunter Goldman Hongyuan Zhan Ibrahim Damlaj Igor Molybog,Igor Tufanov Ilias Leontiadis Irina-Elena Veliche Itai Gat Jake WeissmanJames GeboskiJames Kohli Janice LamJaphet AsherJean-Baptiste GayaJeff MarcusJeff TangJennifer Chan Jenny Zhen Jeremy Reizenstein Jeremy Teboul Jessica Zhong Jian Jin Jingyi Yang Joe Cummings Jon Carvill Jon Shepard Jonathan McPhie Jonathan Torres Josh Ginsburg Junjie Wang Kai Wu Kam Hou U Karan Saxena Kartikay Khandelwal Katayoun Zand Kathy Matosich Kaushik Veeraraghavan Kelly Michelena Keqian Li Kiran Jagadeesh Kun HuangKunal ChawlaKyle HuangLailin Chen Lakshya Garg Lavender A Leandro Silva Lee Bell Lei Zhang Liangpeng Guo Licheng Yu Liron Moshkovich Luca Wehrstedt Madian Khabsa Manav Avalani Manish Bhatt Martynas Mankus Matan Hasson Matthew Lennie Matthias Reso Maxim Groshev Maxim Naumov Maya Lathi. Meghan Keneally Miao Liu Michael L. Seltzer Michal Valko Michelle Restrepo Mihir Patel Mik Vyatskov Mikayel Samvelyan Mike Clark Mike Macey Mike Wang Miquel Jubert Hermoso Mo Metanat Mohammad Rastegari Munish Bansal Nandhini Santhanam Natascha Parks Natasha White Navyata Bawa Nayan Singhal Nick Egebo Nicolas Usunier Nikhil Mehta Nikolay Pavlovich Laptev Ning Dong Nor" + }, + { + "type": "text", + "bbox": [ + 0.546, + 0.092, + 0.909, + 0.534 + ], + "angle": 0, + "content": "man Cheng, Oleg Chernoguz, Olivia Hart, Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pavan Balaji, Pedro Rittner, Philip Bontrager, Pierre Roux, Piotr Dollar, Polina Zvyagina, Prashant Ratanchandani, Pritish Yuvraj, Qian Liang, Rachad Alao, Rachel Rodriguez, Rafi Ayub, Ragtoham Murthy, Raghu Nayani, Rahul Mitra, Rangaprabhu Parthasarathy, Raymond Li, Rebekkah Hogan, Robin Battey, Rocky Wang, Russ Howes, Rudy Rinott, Sachin Mehta, Sachin Siby, Sai Jayesh Bondu, Samyak Datta, Sara Chugh, Sara Hunt, Sargun Dhillon, Sasha Sidorov, Satadru Pan, Saurabh Mahajan, Saurabh Verma, Seiji Yamamoto, Sharadh Ramaswamy, Shaun Lindsay, Shaun Lindsay, Sheng Feng, Shenghao Lin, Shengxin Cindy Zha, Shishir Patil, Shiva Shankar, Shuqiang Zhang, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe, Soumith Chintala, Stephanie Max, Stephen Chen, Steve Kehoe, Steve Satterfield, Sudarshan Govindaprasad, Sumit Gupta, Summer Deng, Sungmin Cho, Sunny Virk, Suraj Subramanian, Sy Choudhury, Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Koehler, Thomas Robinson, Tianhe Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victoria Ajayi, Victoria Montanez, Vijai Mohan, Vinay Satish Kumar, Vishal Mangla, Vlad Ionescu, Vlad Poenaru, Vlad Tiberiu Mihailescu, Vladimir Ivanov, Wei Li, Wenchen Wang, Wenwen Jiang, Wes Bouaziz, Will Constable, Xiaocheng Tang, Xiaojian Wu, Xiaolan Wang, Xilun Wu, Xinbo Gao, Yaniv Kleinman, Yanjun Chen, Ye Hu, Ye Jia, Ye Qi, Yenda Li, Yilin Zhang, Ying Zhang, Yossi Adi, Youngjin Nam, Yu, Wang, Yu Zhao, Yuchen Hao, Yundi Qian, Yunlu Li, Yuzi He, Zach Rait. Zachary DeVito, Zef Rosnbrick. Zhaoduo Wen. Zhenyu Yang. Zhiwei Zhao and Zhiyu Ma. The llama 3 herd of models. 2024. 3. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.537, + 0.905, + 0.563 + ], + "angle": 0, + "content": "[58] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735-1780, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.565, + 0.905, + 0.592 + ], + "angle": 0, + "content": "[59] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.594, + 0.907, + 0.661 + ], + "angle": 0, + "content": "[60] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. CLIP-KD: An empirical study of CLIP model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15952-15962, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[61] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.907, + 0.787 + ], + "angle": 0, + "content": "[62] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and efficient foundation language models, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.907, + 0.844 + ], + "angle": 0, + "content": "[63] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[64] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, Ji Ma, Jiaqi Wang, Xiaoyi Dong, Hang Yan, Hewei Guo, Conghui He, Botian Shi, Zhenjiang Jin, Chao Xu, Bin" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.537, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17804" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.176 + ], + "angle": 0, + "content": "Wang, Xingjian Wei, Wei Li, Wenjian Zhang, Bo Zhang, Pinlong Cai, Licheng Wen, Xiangchao Yan, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. How far are we to GPT-4V? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[65] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, Thomas Unterthiner, Daniel Keysers, Skanda Koppula, Fangyu Liu, Adam Grycner, Alexey Gritsenko, Neil Houlsby, Manoj Kumar, Keran Rong, Julian Eisenschlos, Rishabh Kabra, Matthias Bauer, Matko Bošnjak, Xi Chen, Matthias Minderer, Paul Voigtlaender, Ioana Bica, Ivana Balazevic, Joan Puigcerver, Pinelopi Papalampidi, Olivier Henaff, Xi Xiong, Radu Soricut, Jeremiah Harmsen, and Xiaohua Zhai. PaliGemma: A versatile 3B VLM for transfer, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.329, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[66] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[67] Stephanie Fu, Netanel Yakir Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. DreamSim: Learning new dimensions of human visual similarity using synthetic data. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.484, + 0.537 + ], + "angle": 0, + "content": "[68] Shobhita Sundaram, Stephanie Fu, Lukas Muttenthaler, Netei Yakir Tamir, Lucy Chai, Simon Kornblith, Trevor Darrell, and Phillip Isola. When does perceptual alignment benefit vision representations? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 4, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.538, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[69] Mikołaj Binkowski, Dougal J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.484, + 0.664 + ], + "angle": 0, + "content": "[70] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[71] Aryaz Eghbali and Michael Pradel. CrystalBLEU: Precisely and efficiently measuring the similarity of code. In Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering, New York, NY, USA, 2023. Association for Computing Machinery. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[72] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[73] Peter Stanchev, Weiyue Wang, and Hermann Ney. EED: Extended edit distance measure for machine translation. In Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1), pages 514-520, Florence, Italy, 2019. Association for Computational Linguistics. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.189 + ], + "angle": 0, + "content": "[74] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.19, + 0.908, + 0.275 + ], + "angle": 0, + "content": "[75] Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, Kai Dang, Yang Fan, Yichang Zhang, An Yang, Rui Men, Fei Huang, Bo Zheng, Yibo Miao, Shanghaoran Quan, Yunlong Feng, Xingzhang Ren, Xuancheng Ren, Jingren Zhou, and Junyang Lin. Qwen2.5-Coder technical report, 2024. 6, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.276, + 0.908, + 0.359 + ], + "angle": 0, + "content": "[76] Michael Saxon, Fatima Jahara, Mahsa Khoshnoodi, Yujie Lu, Aditya Sharma, and William Yang Wang. Who evaluates the evaluations? objectively scoring text-to-image prompt coherence metrics with t2IScore score (TS2). In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.361, + 0.908, + 0.403 + ], + "angle": 0, + "content": "[77] Jordan J. Louviere, Terry N. Flynn, and A. A. J. Marley. Best—Worst Scaling: Theory, Methods and Applications. Cambridge University Press, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.403, + 0.908, + 0.502 + ], + "angle": 0, + "content": "[78] Svetlana Kiritchenko and Saif M. Mohammad. Capturing reliable fine-grained sentiment associations by crowdsourcing and best-worst scaling. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 811-817, San Diego, California, 2016. Association for Computational Linguistics. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.502, + 0.908, + 0.587 + ], + "angle": 0, + "content": "[79] Svetlana Kiritchenko and Saif Mohammad. Best-worst scaling more reliable than rating scales: A case study on sentiment intensity annotation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 465-470, Vancouver, Canada, 2017. Association for Computational Linguistics. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.588, + 0.908, + 0.63 + ], + "angle": 0, + "content": "[80] Bryan K. Orme. MaxDiff analysis: Simple counting, individual-level logit, and HB. Sawtooth Software Research Paper Series, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.631, + 0.908, + 0.66 + ], + "angle": 0, + "content": "[81] Bruce Schneier. Applied cryptography - protocols, algorithms, and source code in C, 2nd Edition. Wiley, 1996. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.661, + 0.908, + 0.731 + ], + "angle": 0, + "content": "[82] Kyle Lo, Lucy Lu Wang, Mark Neumann, Rodney Kinney, and Daniel Weld. S2ORC: The semantic scholar open research corpus. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4969-4983, Online, 2020. Association for Computational Linguistics. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.731, + 0.908, + 0.801 + ], + "angle": 0, + "content": "[83] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.802, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[84] Anas Awadalla, Le Xue, Manli Shu, An Yan, Jun Wang, Senthil Purushwalkam, Sheng Shen, Hannah Lee, Oscar Lo, Jae Sung Park, Etash Kumar Guha, Silvio Savarese, Ludwig Schmidt, Yejin Choi, Caiming Xiong, and Ran Xu. BLIP3-KALE: Knowledge augmented large-scale dense captions. In Synthetic Data for Computer Vision Workshop @ CVPR 2025, 2025." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17805" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.148 + ], + "angle": 0, + "content": "[85] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, Yuyin Zhou, and Cihang Xie. What if we recapture billions of web images with LLaMA-3?, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[86] R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 1735-1742, 2006. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[87] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. ImageBind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15180-15190, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[88] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, Cai Wan Zhang, Zhifeng Li, Wei Liu, and Li Yuan. LanguageBind: Extending video-language pretraining to n-modality by language-based semantic alignment. In The Twelfth International Conference on Learning Representations, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.388, + 0.425, + 0.402 + ], + "angle": 0, + "content": "[89] OpenAI. Learning to reason with LLMs, 2024. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[90] DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu, Shengfeng Ye, Tao Yun, Tian Pei, Tianyu Sun, T. Wang, Wangding Zeng, Wanjia Zhao, Wen Liu, Wenfeng Liang, Wenjun Gao, Wenqin Yu, Wentao Zhang, W. L. Xiao, Wei An, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaotao Nie, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xinyu Yang, Xinyuan Li, Xuecheng Su, Xuheng Lin, X. Q. Li, Xiangyue Jin, Xiaojin Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxiang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y. K. Li, Y. Q. Wang, Y. X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yuan Ou, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.217 + ], + "angle": 0, + "content": "Zhou, Y. X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning, 2025. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.907, + 0.3 + ], + "angle": 0, + "content": "[91] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.907, + 0.357 + ], + "angle": 0, + "content": "[92] Urbano Lorenzo-Seva and Jos M. F. ten Berge. Tucker's congruence coefficient as a meaningful index of factor similarity. Methodology: European Journal of Research Methods for the Behavioral and Social Sciences, 2(2):57-64, 2006. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.36, + 0.907, + 0.442 + ], + "angle": 0, + "content": "[93] Jonas Belouadi and Steffen Eger. UScore: An effective approach to fully unsupervised evaluation metrics for machine translation. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 358-374, Dubrovnik, Croatia, 2023. Association for Computational Linguistics. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.444, + 0.907, + 0.554 + ], + "angle": 0, + "content": "[94] Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M. Meyer, and Steffen Eger. MoverScore: Text generation evaluating with contextualized embeddings and earth mover distance. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 563-578, Hong Kong, China, 2019. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.556, + 0.907, + 0.639 + ], + "angle": 0, + "content": "[95] Wei Zhao, Goran Glavaš, Maxime Peyrard, Yang Gao, Robert West, and Steffen Eger. On the limitations of cross-lingual encoders as exposed by reference-free machine translation evaluation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1656–1671, Online, 2020. Association for Computational Linguistics." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.641, + 0.907, + 0.723 + ], + "angle": 0, + "content": "[96] Yurun Song, Junchen Zhao, and Lucia Specia. SentSim: Crosslingual semantic evaluation of machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3143-3156, Online, 2021. Association for Computational Linguistics. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.725, + 0.907, + 0.78 + ], + "angle": 0, + "content": "[97] Y. Rubner, C. Tomasi, and L.J. Guibas. A metric for distributions with applications to image databases. In Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271), pages 59-66, 1998. 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.781, + 0.907, + 0.836 + ], + "angle": 0, + "content": "[98] Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to document distances. In Proceedings of the 32nd International Conference on Machine Learning, pages 957–966, Lille, France, 2015. PMLR. 15" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.836 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17806" + } + ] +] \ No newline at end of file diff --git a/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/5102f011-579e-4213-9591-ec3e5e763b92_origin.pdf b/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/5102f011-579e-4213-9591-ec3e5e763b92_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d0c2c8208fcd58411b7d6995244da619cc0996d7 --- /dev/null +++ b/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/5102f011-579e-4213-9591-ec3e5e763b92_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c02868e0d229a4bc124562e7b68f1a8b7962c709e320e615c22295242469e8b +size 1633207 diff --git a/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/full.md b/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..95e592faa8da66d49b1b11483833a81c71850eb7 --- /dev/null +++ b/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/full.md @@ -0,0 +1,288 @@ +# TikZero: Zero-Shot Text-Guided Graphics Program Synthesis + +Jonas Belouadi* Eddy Ilg† Margret Keuper*,‡ Hideki Tanaka§ Masao Utiyama§ Raj Dabre§ Steffen Eger† Simone Ponzetto* +University of Mannheim, Germany* University of Technology Nuremberg, Germany† Max Planck Institute for Informatics, Saarland Informatics Campus, Germany‡ National Institute of Information and Communications Technology, Japan§ jonas.belouadi@uni-mannheim.de + +# Abstract + +Automatically synthesizing figures from text captions is a compelling capability. However, achieving high geometric precision and editability requires representing figures as graphics programs in languages like TikZ, and aligned training data (i.e., graphics programs with captions) remains scarce. Meanwhile, large amounts of unaligned graphics programs and captioned raster images are more readily available. We reconcile these disparate data sources by presenting TikZERO, which decouples graphics program generation from text understanding by using image representations as an intermediary bridge. It enables independent training on graphics programs and captioned images and allows for zero-shot text-guided graphics program synthesis during inference. We show that our method substantially outperforms baselines that can only operate with caption-aligned graphics programs. Furthermore, when leveraging caption-aligned graphics programs as a complementary training signal, TikZERO matches or exceeds the performance of much larger models, including commercial systems like GPT-4o. Our code, datasets, and select models are publicly available.1 + +# 1. Introduction + +Graphics programming languages offer distinct advantages over low-level vector formats (PDF, SVG) or raster image formats by representing visual concepts as high-level programs that preserve semantics, remain human-interpretable, and allow manual editing. These properties are particularly valuable in academia, where specialized graphics programming languages like TikZ [1] are popular for creating complex figures with high expressivity. However, this comes with a steep learning curve, as seen on the TEX Stack Exchange $^2$ (TEX.SE), where nearly $10\%$ of questions concern TikZ and make it the most frequently discussed topic on the platform [2, 3]. + +![](images/3847c540f76ba4bdd94cb752b78fe3ba6f97ed7688ef35cbf45137413f14e349.jpg) +Figure 1. Qualitative comparison of our $\mathrm{TikZero}^+$ model (last two columns) and the end-to-end trained baseline $\mathrm{AUTOMATikZ}_{\mathrm{v2}}$ (LLM; first two columns) on text-guided graphics program synthesis with TikZ. Our method generates outputs that more closely follow the given captions. Example program listings are in Supp. Mat. F. + +With recent advances in generative AI, simplifying the creation of graphics programs has become increasingly feasible. Belouadi et al. [2] introduce $\mathrm{D}\mathsf{E}\mathrm{T}\mathsf{1}\kappa \mathrm{Z}\mathsf{I}\mathsf{F}\mathsf{Y}$ , an inverse graphics model that generates TikZ programs from images and hand-drawn sketches. However, creating these visual inputs stays cumbersome, motivating alternative input modalities such as natural language. While Belouadi et al. [3] propose AutomATikZ, a text-guided synthesis model for TikZ programs trained end-to-end on an aligned caption-program corpus, its performance remains limited (cf. Fig. 1) [4, 5]. + +We identify insufficient training data as the primary limitation. Unlike inverse graphics models such as $\mathrm{DETikZIFY}$ , which are inherently self-supervised (trained by being conditioned on compiled representations of their output programs) and can access sufficient training data (cf. Fig. 2), end-to + +![](images/a8bc9b60538868e0b4edd0b582689f6259d0cf86d8c62a48a5a0dd71b324d30a.jpg) +Figure 2. Illustration of training data availability for graphics program synthesis. $\mathrm{DETIKZIFY}$ can leverage all graphics programs for training but lacks text guidance, while $\mathrm{AUTOMATIKZ}$ is constrained to the small intersection of captioned graphics programs, resulting in limited performance. Our approach, $\mathrm{TIKZERO}$ , trains independently on both graphics programs and captioned images, enabling more effective use of available data and yielding superior results. + +end text-guided models like AUTOMATIKZ require graphics programs paired with captions, substantially reducing the available data pool (cf. Fig. 2). + +To address this challenge, we decouple the graphics program generation component from text understanding, enabling independent training on graphics programs and captioned images without requiring paired data (cf. Fig. 2). Our approach first trains an inverse graphics model conditioned on image patch embeddings from a vision encoder [6]. We then train an adapter network that generates synthetic image patch embeddings from captions. This adapter training relies solely on captioned images, effectively circumventing resource limitations and enabling zero-shot (in the sense that no aligned caption-program examples are involved in the training process) text-guided graphics program synthesis [7]. We demonstrate that this approach, to which we refer as $\mathrm{T_{IK}ZERO}$ , outperforms previous state-of-the-art methods (cf. Fig. 1). Our key contributions are: + +(i) A novel two-stage architecture, $\mathrm{TikZERO}$ , which addresses the low-resource challenge in text-guided graphics program synthesis by aligning representation spaces rather than relying on aligned data. +(ii) The $\mathrm{DATiKZ_{v3}}$ dataset, comprising over 450k TiKZ graphics programs with roughly 170k captioned samples. Using this dataset, we train both $\mathrm{TiKZERO}$ and $\mathrm{AUTOMATiKZ_{v2}}$ (an updated version of $\mathrm{AUTOMATiKZ}$ ) on the same source data and show that $\mathrm{TiKZERO}$ outperforms $\mathrm{AUTOMATiKZ}$ , $\mathrm{AUTOMATiKZ_{v2}}$ , and other end-to-end trained baselines. +(iii) An enhanced model, $\mathrm{T_{IK}Z_{ZERO} + }$ , combining $\mathrm{T_{IK}Z_{ZERO}}$ with the end-to-end fine-tuning of $\mathrm{AUTOMAT_{IK}Z_{v2}}$ , which surpasses larger baselines and matches the performance of commercial models like GPT-4o [8] on key metrics. + +# 2. Related Work + +Inverse Graphics Program Synthesis Inverse graphics, i.e., synthesizing a graphics program to reproduce a visual target, represents a specialized instance of neural program synthesis [9-11]. Deep learning models have shown remarkable success in this domain [12-14], with Vision-Language Models (VLMs) increasingly gaining prominence [2, 15-17]. While controlled experimental studies often rely on synthetic datasets [13, 15, 17-20], real-world applications typically leverage more complex and diverse human-created data [2, 21-25], highlighting the importance of data availability. In scientific contexts, TikZ has emerged as a popular choice due to its versatility, expressiveness, and widespread adoption in academic circles [2, 21-23, 25]. Although these approaches are not tailored to text-guided generation, we incorporate key elements from them into our approach. + +Text-Guided Graphics Program Synthesis Current text-guided approaches to graphics program synthesis remain limited, mainly because of the scarcity of captioned graphics programs outlined in Sec. 1, but also because of the difficulty of generating synthetic data with human-like captions [26, 27]. Researchers interested in this capability currently rely on the emerging capabilities of large commercial models such as GPT-4o [8, 28–30], which raises concerns about accessibility, reproducibility, and computational cost [31]. In contrast, related domains like vector graphics generation [24, 32–35] and NL2Vis [36–40] have shown more progress. Similar to inverse graphics, these fields increasingly incorporate large language models (LLMs) [24, 36, 40]. However, vector graphics approaches typically generate only low-level Bezier curves, limiting output complexity [2, 32, 33], and NL2Vis focuses exclusively on data visualization with a restricted set of visualization types [40]. More complex applications, such as generating arbitrary scientific figures from captions with TikZ, remain underexplored—a gap we address in this work. + +Text-to-Image Generation $\mathrm{TiKZero}$ shares conceptual and architectural similarities with several text-to-image generation methods [41-45]. Rodriguez et al. [46, 47] explore generative adversarial networks [48, 49] and diffusion models [50, 51] for scientific figure generation, but these approaches are tied to raster images, which are not ideal for representing scientific figures. Ramesh et al. [43] propose a two-stage model with independently trained prior and decoder components to generate raster images from text. Although prior networks resemble our adapters and have been used with inverse graphics models [52], they target global image embeddings containing only abstract information, which degrades performance when used with inverse graphics architectures that work best with patch-level details [53]. In contrast, our adapters specifically operate on patch-level embeddings, and we demonstrate that this improves performance compared to end-to-end trained baselines. + +![](images/5a4db9ff5a15611160ed325154439f80a3c320b123f535faeb2122769ccb72be.jpg) +Figure 3. Architecture overview of $\mathrm{T1KZERO}$ during inference. Solid lines represent the standard caption-conditioned path, which flows through the text encoder into the adapter network of $\mathrm{T1KZERO}$ before connecting to the vision encoder. In certain configurations (cf. Sec. 5.2), the caption also feeds into the text decoder (depicted by dotted lines and “ $\bullet$ ” markers representing shortcuts). The self- and cross-attention layers (yellow) are simplified representations, omitting internal feed-forward layers and residual connections [54]. An exception is the explicit residual connection between the cross-attention and self-attention layers of the vision encoder, visualizing the gating mechanism $\gamma$ (purple). Additionally, the dashed path illustrates how the inverse graphics model generates graphics programs when conditioned on images. + +# 3. The TIKZERO Model & Architecture + +As the foundation of $\mathrm{TiKZero}$ , we first develop a state-of-the-art inverse graphics model for graphics program synthesis. We then incorporate a cross-attention adapter network [55] for text guidance. Fig. 3 provides an overview of our method. + +The Inverse Graphics Model Due to their demonstrated effectiveness (cf. Sec. 2), we adopt a VLM architecture for the inverse graphics model of $\mathrm{T_{IK}Zero}$ . Figure 3 illustrates its inner workings (dashed lines): the model processes rasterized images and autoregressively generates their corresponding programs without involving captions at this stage. + +The Adapter Network VLMs consist of two primary components: a vision encoder that produces image embeddings and a text decoder that, in our case, generates graphics programs conditioned on these embeddings. The unidirectional and localized flow of information between these components allows us to inject additional information solely into the vision encoder, thereby influencing the output of the text decoder. We exploit this property by introducing a trainable, text-conditioned adapter network that mimics the outputs of the original vision encoder. This effectively enables zero-shot generation of graphics programs conditioned on text when its outputs are fed into the decoder. In addition to circumventing the resource limitations discussed in Sec. 1, this architecture has other welcome implications: During adapter training, the text decoder, usually the largest component of the model, does not need to be loaded, resulting in efficient and fast training even with large datasets. Our adapter incorporates a lightweight text encoder for embedding captions and intro- + +duces newly initialized gated cross-attention layers [56, 57] before each vision encoder layer (cf., Fig. 3). The keys and values derive from the final text encoder representations, while the queries originate from a trainable probe used instead of image inputs. The gates $\gamma$ allow the model to learn at which layers and to what extent information from the text encoder should flow into the vision encoder. Contrary to existing literature, which often employs tanh gates [58] that initialize to zero (indicating no information flow), we find that using sigmoid gates (0.5 at initialization) accelerates training convergence since, in our case, only little information originates from the vision encoder inputs (i.e., the probe). We ablate the gates and the probe in Supp. Mat. A. + +Training Objective Given a caption-image dataset for training, we first embed the patches $p \in \mathbf{p}$ of image $i$ using the unmodified vision encoder $\mathbf{M}$ of our VLM. Subsequently, we incorporate the cross-attention adapter to obtain the modified encoder $\widehat{\mathbf{M}}$ , which we then distill on these image patch embeddings conditioned solely on the caption $t$ and probe $\hat{t}$ [59]. This leads to the following objective: + +$$ +\mathcal {L} _ {\mathrm {d i s t}} = \frac {1}{| p |} \sum_ {p \in p} \operatorname {d i s t} \left(\mathrm {M} _ {\theta} (p \mid i), \widehat {\mathrm {M}} _ {\theta , \hat {\theta}} (p \mid \hat {t}, t)\right), \tag {1} +$$ + +where $\mathrm{dist}(\pmb{x},\pmb{y})$ represents a distance metric. Following common practices in model distillation [60, 61], we experiment with cosine distance and mean squared error. Here, $\theta$ denotes the original model parameters that remain fully frozen, while $\hat{\theta}$ represents the adapter parameters of which the cross-attention layers and the image probe are trainable. + +
SourceDATIKZDATIKZv2DATIKZv3
curated98115663646
TEX.SE292383060942654
arXiv85656326450407851
artificial195719582256
all117832360583456469
+ +Table 1. Breakdown of the number of unique $\mathrm{Ti}k\mathrm{Z}$ graphics in $\mathrm{DATI}\kappa \mathrm{Z}_{\mathrm{v3}}$ compared to its predecessors $\mathrm{DATI}\kappa \mathrm{Z}$ and $\mathrm{DATI}\kappa \mathrm{Z}_{\mathrm{v2}}$ . Qualitative examples can be found in Supp. Mat. F. + +# 4. Datasets & Model Training + +We introduce $\mathrm{DATikZ_{v3}}$ , a novel dataset of TikZ graphics programs designed to support the training and evaluation of $\mathrm{TIKZERO}$ . Additionally, we train $\mathrm{AUTOMATikZ_{v2}}$ as a directly comparable baseline operating on the same data source. + +The $\mathrm{DATiKZ_{v3}}$ Dataset $\mathrm{DATiKZ_{v3}}$ expands upon its predecessors $\mathrm{DATiKZ_{v2}}$ [2, 3], incorporating programs from curated repositories, TEX.SE, arXiv papers, and artificial samples (cf. Tab. 1). While previous versions focused exclusively on TikZ graphics with (v1) or without (v2) captions, $\mathrm{DATiKZ_{v3}}$ systematically extracts captions alongside TikZ graphics whenever possible to support our claims. From over 450k instances, fewer than 170k include captions, underscoring the challenges discussed in Sec. 1. + +Training TIKZERO $\mathrm{T_{IK}Zero}$ 's VLM builds upon DE $\mathrm{T_{IK}ZIFY}$ [2] by conditioning a LAMA-based text decoder [62] on patch embeddings from a S1G1P vision encoder [63]. Specifically, we combine $\mathrm{LLAMA}_{3.1}$ (8B) [57] with S1G1P SoViT (0.4B). Unlike $\mathrm{DE_{IK}ZIFY}$ and inspired by the continued ViT pretraining approach of INTERNVL 1.5 [64], we initialize the vision encoder with weights from the fine-tuned encoder of PALIGEMMA [65] and increase the input resolution to $420\times 420$ pixels. Furthermore, we fully fine-tune the vision encoder alongside the rest of the model instead of freezing it. We train on $\mathrm{DAT_{IK}Z_{v3}}$ for 5 epochs with a learning rate of 5e-5 and a batch size of 128. $\mathrm{T_{IK}ZERO}$ 's VLM consistently outperforms $\mathrm{DE_{IK}ZIFY}$ , with detailed evaluation results provided in Supp. Mat. B. For the adapter network, we initialize with $\mathrm{LLAMA}_{3.2}$ (1B) as the text encoder [57] and leverage ARxivCAP [26], a dataset comprising 6.4 million scientific caption-image pairs for training. The adapter accounts for 2 billion of TIKZERO's 10 billion total parameters, with only 400 million being trainable. We train for 3 epochs with a learning rate of 1e-4 and a batch size of 512. We emphasize that this two-stage training process does not access caption-program pairs. However, we demonstrate that incorporating such aligned data in a subsequent fine-tuning step (Sec. 5.2) further enhances performance. + +Training AUTOMATIKZv2 Similar to its predecessor, AUTOMATIKZv2 is a token-conditioned LLM that uses tokenized captions as conditioning information for graphics prediction (rather than patch embeddings). We initialize AUTOMATIKZv2 in two different ways: (i) AUTOMATIKZv2 (LLM), which starts from vanilla LLAMA3.1 (8B) weights, and (ii) AUTOMATIKZv2 (VLM), which leverages TIKZERO's trained VLM (minus the vision encoder) to benefit from transfer learning [66] on its larger training corpus. Both variants employ the same hyperparameters as TIKZERO's VLM but can only utilize the caption-annotated subset of DATIKZv3 for training. Despite having access to less caption-aligned data than TIKZERO's adapter network, AUTOMATIKZv2 requires a longer training period primarily due to fine-tuning the large decoder (8 billion trainable parameters versus the adapter network's 400 million). Training requires more than two days for AUTOMATIKZv2 and 1.5 days for TIKZERO's adapter network when using eight Nvidia A100 40GB GPUs. + +# 5. Experiments + +Before training models on $\mathrm{DAT}_{\mathrm{IK}}\mathrm{Z}_{\mathrm{v3}}$ , we extract 1k samples from its captioned subset to form our test set. To mitigate data leakage from pretraining to testing, we only include instances created after the cut-off date specified by $\mathrm{LLAMA}_{3.2}$ and $\mathrm{ARXIVCAP}$ . We also employ an $n$ -gram matching algorithm to avoid cross-contamination with our training split [8]. For all models, the temperature is set to 0.8 and top-p to 0.95. Example outputs are provided in Fig. 1 and Supp. Mat. F. + +Evaluation Metrics The multimodal nature of our task allows for various evaluation metrics in our automatic evaluations. We assess perceptual image similarity between generated outputs and references by computing DREAMSIM (DSIM) [67, 68], which correlates highly with human judgments for scientific images [2]. We also calculate the Kernel Inception Distance (KID) [69] using SigLIP image features, which evaluates the overall quality of generated figures by comparing to the distribution of reference figures. We evaluate caption similarity between generated outputs and reference captions using CLIPSCORE (CLIP) [70] with SigLIP features. To measure code similarity between generated and reference TikZ programs, we use CRYSTALBLEU (cBLEU), a BLEU variant optimized for code evaluation [71, 72], and TEX Edit Distance (TED) [2], a variant of the Extended Edit Distance [73] utilizing a TEX tokenizer. Since some metrics require that generated programs compile to images, resampling is necessary if the output contains irrecoverable errors. To quantify this, we compute the Mean Token Efficiency (MTE), defined as the $10\%$ winsorized mean of the ratio between the number of tokens in the final TikZ program and the total number of tokens generated to produce that program. For a comprehensive view of model performance, we calculate the arithmetic mean (AVG) of all previous + +
ModelsOriginal TextRedacted Text
DSIM↑KID↓CLIP↑cBLEU↑TED↓MTE↑AVG↑CLIP↑Ratio↑
IDEFICS 3 (8B)45.47511.42614.3270.65663.17569.55866.6284.85133.858
AUTOMATIKZ (13B)46.0331.2943.9550.38662.2485.86663.0932.96574.975
AUTOMATIKZv2 (VLM)38.31333.2030.7750.32876.98521.5950.00.28436.597
AUTOMATIKZv2 (LLM)50.5483.49115.7660.65862.30781.77582.3758.00250.753
TIKZERO (MSE)52.0245.66410.5831.72366.0779.31885.0048.23777.831
TIKZERO (Cos)52.8295.10310.0511.60365.5182.29185.5997.22671.893
+ +Table 2. System-level scores $\times {100}$ for ${\mathrm{{T\kappa Zero}}}_{0}$ and baselines of comparable size and training setup. Bold and underlined values denote the best and second-best scores for each metric column, respectively. Cell shading illustrates relative score magnitudes. Arrows indicate metric directionality. Overall, ${\mathrm{{T\kappa Zero}}}_{0}$ achieves the strongest average performance across metrics. + +metrics. As these metrics operate on different scales, we apply min-max normalization before computing the average. Additionally, some metrics are recomputed with redacted text in the outputs as part of our analysis, cf. Sec. 6.1. + +# 5.1. Comparison against End-to-End Fine-Tuning + +In our initial experiment, we evaluate the zero-shot performance of $\mathrm{TiKZero}$ , trained as described in Secs. 3 & 4 using either cosine distance (Cos) or mean squared error (MSE), and compare it against end-to-end trained baselines. + +Baselines Besides $\mathrm{AUTOMATIKZ_{v2}}$ (LLM & VLM), which we designed as directly comparable baselines, we assess other token-conditioned models of similar and slightly larger sizes trained on $\mathrm{TikZ}$ . Specifically, we evaluate $\mathrm{AUTOMATIKZ}$ (13B) $^3$ the strongest original $\mathrm{AUTOMATIKZ}$ baseline [3], and the general-purpose chatbot IDEFICS 3 (8B) [22]. Additional models and details are available in Supp. Mats. A & C. + +Results We present the system-level metric scores in Tab. 2 (Original Text). On average, $\mathrm{T_{IK}Z_{ero}}$ , trained with cosine distance, achieves the best performance with an AVG score of 85.599, closely followed by the MSE variant at 85.004. The next best model, $\mathrm{AUTOMAT_{IK}Z_{v2}}$ (LLM), scores 82.375, which is 3 percentage points (pp) lower. The remaining models exhibit a substantial performance gap, with IDEFICS 3 (8B) and $\mathrm{AUTOMAT_{IK}Z}$ (13B) falling behind by approximately 20pp and $\mathrm{AUTOMAT_{IK}Z_{v2}}$ (VLM) showing the weakest performance across all metrics, resulting in an AVG score of 0. The surprisingly poor results of $\mathrm{AUTOMAT_{IK}Z_{v2}}$ (VLM) are likely due to catastrophic forgetting [74], as the removal of the vision encoder from $\mathrm{T_{IK}Z_{ERO}}$ ’s VLM necessitates reacquisition of conditioning based solely on text. + +As for individual metrics, our adapter-based models perform particularly well in perceptual image similarity, with $\mathrm{T_{IK}Z_{\mathrm{ero}}}$ (Cos) outperforming the best baseline, AUTOMATIKZv2 (LLM), by 3pp on DREAMSIM. Although AUTOMATIKZv2 (LLM) outperforms $\mathrm{T_{IK}Z_{\mathrm{ero}}}$ by 1.5pp on KID, this + +indicates in this context that such token-conditioned models (compared to those using patch embeddings) capture the general appearance of scientific figures well but fall short in inferring visual specifics from captions. They do, however, have an edge in reproducing text from captions, which we identify as the primary reason for up to 5pp higher CLIP-Score, as noted in Sec. 6.1. Regarding code similarity, both $\mathrm{TIKZERO}$ models considerably outperform others on cBLEU. Interestingly, we observe a mild inverse correlation between cBLEU and TED. Models conditioned solely on tokenized captions tend to generate shorter, often simplified programs [3], potentially resulting in a reduced edit distance to the reference. In terms of efficiency, all models achieve an MTE of 80-85, indicating that only 2 out of 10 inferences require resampling. $\mathrm{TIKZERO}$ (Cos) is 3pp more efficient than MSE, while AUTOMATIKZ (13B), likely benefiting from its larger model size, exceeds it by another 3pp. + +In summary, training $\mathrm{AUTOMATIKZ_{v2}}$ on top of a VLM yields worse performance than training based on vanilla $\mathrm{LLAMA}_{3.1}$ , indicating that effective end-to-end training can only leverage the small intersection of graphics programs and images with captions, as illustrated in Fig. 2. However, even without access to this intersection, $\mathrm{TikZERO}$ surpasses both $\mathrm{AUTOMATIKZ}$ (13B) and $\mathrm{AUTOMATIKZ_{v2}}$ (LLM) on average by being able to train on images with captions independently of graphics programs. Moreover, using a loss function based on cosine distance proves more effective than using MSE. + +# 5.2. Combining Adapters with Fine-Tuning + +In this section, we investigate whether explicitly incorporating the subset of $\mathrm{DAT}\kappa \mathrm{Z}_{\mathrm{v3}}$ that includes captions into the training process of $\mathrm{T}\mathbf{I}\kappa \mathbf{Z}\mathbf{E}\mathbf{R}\mathbf{O}$ enhances performance. Our approach involves three incremental stages: (i) We perform a light fine-tuning of $\mathrm{T}\mathbf{I}\kappa \mathbf{Z}\mathbf{E}\mathbf{R}\mathbf{O}$ (Cos) end-to-end on caption-program pairs for one epoch with a low learning rate of 1e-5. Extending the training duration or increasing the learning rate does not yield further performance gains, likely due to the decoder having already reached its saturation point; + +
ModelsOriginal TextRedacted Text
DSIM↑KID↓CLIP↑cBLEU↑TED↓MTE↑AVG↑CLIP↑Ratio↑
QWEN2.5 CODER (32B)54.4735.49324.870.28559.85697.26948.59312.16448.911
GPT-4o56.4642.84431.7870.32758.51197.67579.01913.3241.905
TikZero (Cos)52.8295.10310.0511.60365.5182.29114.6587.22671.893
+ Fine-tuning (i)53.2031.79410.6870.75961.57294.85146.4976.51260.931
+ Separate Captions (ii)52.9832.90515.720.80461.3295.72246.3268.74155.608
+ Weight Resetting (iii)56.2951.83124.1771.98859.00893.05887.04311.47947.478
+ +Table 3. System-level scores $\times {100}$ for additional baselines and ${\mathrm{{TikZero}}}$ combined with fine-tuning and token-conditioning. The scores for TIKZERO (Cos) are replicated from Tab. 2 for convenience. Bold and underlined values denote the best and second-best scores for each metric column, respectively. Cell shading illustrates relative score magnitudes. Arrows indicate metric directionality. Overall, ${\mathrm{{TikZero}}}_{\mathrm{O}}\left( \mathrm{{Cos}}\right)$ with Weight Resetting (iii) demonstrates the strongest average performance across metrics. + +(ii) Alongside feeding captions into the adapter, we provide them separately to the text decoder in tokenized form (cf., Fig. 3); (iii) Prior to fine-tuning, we reset the decoder to its initial weights to overcome saturation, enabling us to fine-tune using the setup described in Sec. 4, which involves 5 epochs and a learning rate of $5\mathrm{e} - 5$ + +Baselines In addition to the baselines in Tab. 2, which remain comparable, we also evaluate larger and commercial models that serve as stronger baselines (cf. Supp. Mat. C). Specifically, we assess GPT-4o [8], which has demonstrated strong performance in generating TikZ [3, 28, 30] and QwEN2.5 CODER (32B) [75] as an open-weights model. + +Results In Tab. 3 (Original Text), all fine-tuning setups of $\mathrm{TiKZero}$ show considerable improvement over the base version. Approaches (i) and (ii) each enhance performance by over 30pp on AVG, while approach (iii) surpasses them with an improvement of over 70pp, positioning it as the best-performing model on average, even when compared to our new baselines, with GPT-4o being 8pp lower and $\mathrm{QWEN}_{2.5}$ CODER (32B) approximately 40pp lower. Approach (i) demonstrates that direct fine-tuning yields positive effects across nearly all metrics, notably improving MTE by 12pp, TED by 4pp, and KID by 3.5pp. Approach (ii) shows similar trends but, by also incorporating tokenized captions, further improves CLIPSCORE by 5pp, closing the gap to AUTOMATIKZv2 (LLM). Interestingly, both (i) and (ii) slightly decrease performance on cBLEU, potentially due to similar reasons discussed in Sec. 5.1. However, the same cannot be said for (iii), which not only achieves the highest score on cBLEU but also ranks as the second-best on TED, trailing only 0.5pp behind GPT-4o and showcasing that it is possible to perform well on both metrics. Additionally, it increases DREAMSIM by another 3pp and CLIPSCORE by 8.5pp, competing with the much stronger baselines QWEN2.5 CODER (32B) and GPT-4o. In KID, it even surpasses them by 3.5pp and 1pp, respectively. + +In summary, fine-tuning $\mathrm{T_{IK}Zero}$ , especially when com + +bined with a separate caption input and weight resetting, greatly improves performance. This illustrates that the intersection of graphics programs and images with captions, though small, provides a valuable training signal, and best performance can be achieved by making full use of both sets. The best-performing $\mathrm{TiKZero}$ model even competes with and often surpasses QwEN2.5 CODER (32B) and GPT-4o on several key metrics. Notably, the former model is more than three times larger, and the latter is often estimated at around 1.8 trillion parameters [76], making it 180 times larger. + +# 5.3. Human Evaluation + +To corroborate our findings from automatic evaluation, we conduct a human annotation campaign focusing on two key properties: caption and image similarity. We employ BestWorst Scaling (BWS) [77], a comparative annotation method that yields high-quality results even with few annotators [78, 79]. We sample 100 instances from our test set and present annotators with $n$ -tuples of generated figures, asking them to identify the most and least similar figure to either the reference caption or reference image. This data is then transformed into scores from -1 (poor) to 1 (excellent) by subtracting the proportion of times a figure is selected as the best from the proportion of times it is chosen as the worst [80]. For a manageable workload, we focus on $n = 4$ key models: $\mathrm{TikZERO}(\mathrm{Cos})$ , our best-performing model from Sec. 5.1; $\mathrm{AUTOMATIKZv2}$ (LLM), its direct end-to-end trained competitor; GPT-4o, our strongest baseline; and $\mathrm{TikZERO}(\mathrm{Cos})$ fine-tuned using approach (iii) from Sec. 5.2, our best model overall, henceforth referred to as $\mathrm{TikZERO}^+$ for convenience. We engage thirteen annotators and obtain six fully annotated sets per task (cf. Supp. Mat. E for more details). To assess annotator consistency, we calculate the split-half reliability (SHR) [79]. This method randomly divides all annotations into two sets, calculates scores independently, and then determines their correlation using Spearman's $\rho$ . + +![](images/5f5c20c6dddaabcd9f5290421346d7a7e02abe869dceed211d4680fe1f6259f7.jpg) +Figure 4. Bivariate distributions of BWS scores (higher is better) using kernel density estimation for caption and image similarity. Along the diagonal, $\mathrm{T_{IK}Z_{ERO}}$ (Cos) achieves higher scores than $\mathrm{AUTOMAT_{IK}Z_{v2}}$ (LLM), while $\mathrm{T_{IK}Z_{ERO+}}$ and GPT-4o demonstrate superior performance compared to both. + +Results Fig. 4 presents kernel density estimates for the BWS scores, showing generally consistent rankings with automatic evaluations but revealing notable differences in the magnitude of gaps. For caption similarity, the ranking aligns with CLIPSCORE evaluations $(\rho = 1.0)$ , with $\mathrm{TikZero}$ (Cos), $\mathrm{AUTOMATIKZ_{v2}}$ (LLM), $\mathrm{TikZero}+$ , and GPT-4o achieving mean scores $\mu$ of -0.25, -0.18, 0.03, and 0.4, respectively. Interestingly, humans perceive a $40\%$ smaller gap between $\mathrm{AUTOMATIKZ_{v2}}$ (LLM) and $\mathrm{TikZero}$ (Cos) than suggested by CLIPSCORE values, where $\mathrm{AUTOMATIKZ_{v2}}$ (LLM) outperforms $\mathrm{TikZero}$ (Cos) by $50\%$ . This indicates humans may evaluate caption similarity differently than CLIPSCORE (cf. Sec. 6.1). For image similarity, the system order remains consistent with our DREAMSIM metric $(\rho = 1.0)$ , with $\mathrm{AUTOMATIKZ_{v2}}$ (LLM), $\mathrm{TikZero}$ (Cos), $\mathrm{TikZero}+$ , and GPT-4o achieving $\mu$ of -0.26, -0.02, 0.01, and 0.27, respectively. However, the relative gaps between models differ: the separation between $\mathrm{AUTOMATIKZ_{v2}}$ (LLM) and $\mathrm{TikZero}$ (Cos), as well as between $\mathrm{TikZero}+$ and GPT-4o, appear more pronounced than observed with DREAMSIM. This discrepancy likely stems from BWS capturing relative preferences rather than absolute performance differences. GPT-4o is selected $20\%$ more often as the best model than $\mathrm{TikZero}+$ , and $\mathrm{AUTOMATIKZ_{v2}}$ (LLM) $15\%$ more often as the worst model than $\mathrm{TikZero}$ (Cos), creating larger perceived gaps even when qualitative differences may be subtle. + +The SHR values of 0.68 for caption similarity and 0.76 for image similarity indicate moderate to strong inter-annotator + +agreement. We also observe a correlation between these two tasks, with segment-level $\rho = 0.62$ and system-level $\rho = 0.8$ , suggesting that both evaluation dimensions capture related aspects of model performance. GPT-4o emerges as the best-performing model, aligning with its superior performance on the corresponding automatic metrics, CLIPSCORE and DSIM. Among open-source models, $\mathrm{T_{IK}Z_{v2}}$ performs best, while AUTOMATIKZv2 (LLM) ranks lowest overall. + +# 6. Analysis + +We present a comprehensive analysis, investigating the influence of typographic attacks on CLIPSCORE and examining the effectiveness of our architecture in low-resource settings, both in terms of training data and trainable parameters. + +# 6.1. CLIPSCORE Limitations & Typographic Attacks + +A known limitation of CLIPSCORE with text-rich images is its susceptibility to typographic attacks, where scores are disproportionately influenced by string similarity between images and captions [3, 43]. We suspect that token-conditioned models like AUTOMATIKZv2 (LLM) achieve higher CLIPSCORE values than models such as TIKZERO (Cos & MSE) primarily because they tend to visibly copy more substrings from the caption in the output image. To test this hypothesis, we apply the ROT13 substitution cipher [81] to all visible strings in the generated figures and recompute CLIPSCORE. This basic cipher replaces each letter with the 13th letter after it in the Latin alphabet. While not cryptographically secure, the ratio between the original and recomputed CLIPSCORE values should indicate the influence of string matching, i.e., higher ratios suggest less copied text and vice versa. + +Tabs. 2 & 3 (Redacted Text) present the recomputed CLIP-Score values and ratios for all evaluated models. The results reveal that most $\mathrm{T_{IK}Zero}$ models, except for fine-tuning approaches (ii) and (iii), which also condition on tokenized captions, have considerably higher ratios $(61\% -78\%)$ compared to strictly token-conditioned models $(34\% -51\%)$ , supporting our hypothesis. AUTOMATIKZ (13B) is an exception, possibly due to its initially low score. Further analysis shows that with redacted text, $\mathrm{AUTOMATIKZ}_{\mathrm{V2}}$ (LLM)'s CLIPSCORE performance drops to the same level as $\mathrm{T_{IK}Zero}$ (Cos & MSE), suggesting that string matching is the primary factor in its superior performance rather than producing better visuals—arguably a more difficult task. Nevertheless, reproducing strings is still somewhat desirable. The human oracle of our test set achieves a ratio of $50.8\%$ , close to the $47.5\%$ of $\mathrm{T_{IK}Zero+}$ , our best-performing model. In contrast, models like GPT-4o, with a lower ratio of $41.9\%$ , may overfit to caption copying, artificially inflating the CLIPSCORE values. + +# 6.2. Low-Resource Training + +While our adapters train efficiently on large-scale datasets, we investigate whether such extensive data is necessary for opti + +
Intv.Training Data
100%50%25%12.5%
192.41177.47849.96756.055
287.55785.24954.94233.817
482.25447.38132.1237.914
876.54540.81629.77416.25
+ +Table 4. AVG scores for $\mathrm{TiKZero}$ (Cos) trained on varying fractions of data and intervals of cross-attention layers. Higher scores indicate better performance. Bold and underlined values denote the best and second-best scores for the whole table, respectively. Cell shading illustrates score magnitudes. + +mal performance. Along the same vein, we examine the impact of reducing the amount of cross-attention layers inserted into the vision encoder. We retrain $\mathrm{TIKZERO}$ (Cos) using varying fractions of the training data ( $d \in \{1, \frac{1}{2}, \frac{1}{4}, \frac{1}{8}\}$ ) and insert cross-attention layers at different intervals ( $i \in \{1, 2, 4, 8\}$ ). Table 4 presents the AVG scores from this parameter grid, with detailed scores in Supp. Mat. D. Our findings reveal that utilizing the full dataset and inserting cross-attention at every layer yields the highest average performance, highlighting the benefits of maximizing both variables. Interestingly, the model's performance appears more robust to a reduction in the number of layers compared to a decrease in training data. For instance, training on only $\frac{1}{8}$ th of the data leads to a substantial performance drop of $36\mathrm{pp}$ , whereas inserting cross-attention layers every 8 layers (resulting in only 3 cross-attention layers in total) causes a more modest decline of $16\mathrm{pp}$ . Minimizing both variables leads to the most severe drop of over $75\mathrm{pp}$ . These results validate our training setup while suggesting that incorporating additional data might further enhance performance. Given that $\mathrm{ARXIVCAP}$ extracts figures from only $572\mathrm{k}$ papers, whereas some corpora index over 200 million papers [82], there remains a lot of potential for leveraging larger datasets in future work. + +# 7. Conclusion + +In this work, we demonstrate the potential of $\mathrm{TiKZero}$ and its variants for generating TiKZ graphics programs from captions. Notably, $\mathrm{TiKZero}$ does not require aligned caption-program pairs in its original formulation but instead aligns representation spaces of unaligned graphics programs and captioned images. This enables our model to leverage substantially more training data compared to end-to-end trained models that operate solely on caption-image pairs (cf. Fig. 2) while maintaining training efficiency. $\mathrm{TiKZero}$ outperforms strong end-to-end trained baselines, including our independently trained AUTOMATiKZv2 models, which use the same data pool, excluding instances they cannot process, illustrating the strengths of our approach. When extending the $\mathrm{TiKZero}$ ap + +proach with additional end-to-end training, it also compares favorably to much larger baselines and commercial systems like GPT-4o. While this enhanced approach, $\mathrm{T_{IK}ZERO^{+}}$ , is no longer zero-shot by our definition, it remains a $\mathrm{T_{IK}ZERO}$ model in the sense that it operates on both sets of graphics programs and captioned images, with the added advantage of explicitly utilizing their intersection (cf. Fig. 2). + +These results demonstrate the benefits of designing architectures around available data and validate the approach of decoupling graphics program generation from text understanding (with optional later reconciliation through $\mathrm{TiKZero^{+}}$ ). Although we demonstrate our method specifically on $\mathrm{TiKZ}$ , we believe its general principles will inspire future work on related graphics program synthesis tasks. + +Future Work Beyond scaling up our training data to explore convergence limits (cf. Sec. 6.2), we plan to investigate automatic methods for improving the quality and alignment of caption-image or caption-program pairs. This includes rewriting potentially noisy captions with LLMs and enhancing them with the visual understanding capabilities of VLMs [83-85]. We believe our approach to aligning textual and image modalities enables other promising applications for graphics program synthesis, such as editing images in latent space via textual instructions to generate modified graphics programs. Additionally, we intend to explore alternative alignment strategies beyond model distillation, including contrastive learning [86], which has successfully aligned modalities in discriminative models [6, 87, 88]. + +# Limitations + +Our evaluations include proprietary systems that operate as black boxes; their training data is unknown, and they offer no guarantees of consistent performance over time. This (i) makes addressing data leakage and cross-contamination impossible and (ii) limits the fairness and reproducibility of our experiments. Nevertheless, even under these unfavorable conditions, our open models remain competitive. Users should be aware, however, that our models may behave unpredictably, and outputs might differ from expectations. Additionally, our models do not include safeguards against potential misuse, e.g., for generating fake scientific content. + +Regarding licensing of our training data, a large portion of the TikZ programs in $\mathrm{DAT}_{\mathrm{IK}}\mathrm{Z}_{\mathrm{v3}}$ are licensed under permissive terms that allow redistribution. The remaining programs are distributed under the arXiv.org perpetual, non-exclusive license, which prohibits redistribution, which is why we exclude them from the public release of $\mathrm{DAT}_{\mathrm{IK}}\mathrm{Z}_{\mathrm{v3}}$ . However, since we release our dataset creation scripts, we encourage others to reproduce the full version independently. + +# Acknowledgments + +We extend our sincere gratitude to the following individuals (in no particular order) for their valuable contributions: Christian Greisinger, Hour Kaing, Ran Zhang, Tejaswini Medi, Yanran Chen, Sotaro Takeshita, Katharina Prasse, JiWoo Kim, Christoph Leiter, Haiyue Song, and Aida Kostikova. Their assistance with our human evaluation campaign, proofreading, insightful discussions, and constructive feedback has been instrumental to our work. The first author conducted part of this research during an internship at the National Institute of Information and Communications Technology (NICT), Japan. The second to last author is supported by the Federal Ministry of Education and Research (BMBF) via the research grant METRICS4NLG and the German Research Foundation (DFG) via the Heisenberg Grant EG 375/5-1. We acknowledge computing resources provided by the state of Baden-Württemberg through bwHPC and the German Research Foundation (DFG) through grant INST 35/1597-1 FUGG. Finally, we thank the OpenMoji project for the open-source icons used throughout this work. + +# References + +[1] Till Tantau. The TikZ and PGF Packages, 2023. 1 +[2] Jonas Belouadi, Simone Paolo Ponzetto, and Steffen Eger. DeTikZify: Synthesizing graphics programs for scientific figures and sketches with TikZ. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 4, 15 +[3] Jonas Belouadi, Anne Lauscher, and Steffen Eger. AutomataTikZ: Text-guided synthesis of scientific vector graphics with TikZ. In The Twelfth International Conference on Learning Representations, 2024. 1, 4, 5, 6, 7, 16, 18 +[4] Leixin Zhang, Yinjie Cheng, Weihe Zhai, Steffen Eger, Jonas Belouadi, Fahimeh Moafian, and Zhixue Zhao. ScImage: How good are multimodal large language models at scientific text-to-image generation? In The Thirteenth International Conference on Learning Representations, 2025. 1, 15 +[5] Abhay Zala, Han Lin, Jaemin Cho, and Mohit Bansal. DiagrammerGPT: Generating open-domain, open-platform diagrams via LLM planning. In First Conference on Language Modeling, 2024. 1 +[6] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 8 +[7] Mark Palatucci, Dean Pomerleau, Geoffrey E Hinton, and Tom M Mitchell. Zero-shot learning with semantic output codes. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2009. 2 +[8] OpenAI. GPT-4 technical report, 2023. 2, 4, 6 +[9] Emilio Parisotto, Abdel rahman Mohamed, Rishabh Singh, Lihong Li, Dengyong Zhou, and Pushmeet Kohli. Neuro + +symbolic program synthesis. In International Conference on Learning Representations, 2017. 2 +[10] Jacob Devlin, Jonathan Uesato, Surya Bhupatiraju, Rishabh Singh, Abdel rahman Mohamed, and Pushmeet Kohli. Robust-Fill: Neural program learning under noisy I/O. In Proceedings of the 34th International Conference on Machine Learning, pages 990-998. PMLR, 2017. +[11] Kevin Ellis, Catherine Wong, Maxwell Nye, Mathias Sable-Meyer, Lucas Morales, Luke Hewitt, Luc Cary, Armando Solar-Lezama, and Joshua B. Tenenbaum. DreamCoder: bootstrapping inductive program synthesis with wake-sleep library learning. In Proceedings of the 42nd ACM SIGPLAN International Conference on Programming Language Design and Implementation, page 835-850, New York, NY, USA, 2021. Association for Computing Machinery. 2 +[12] Yaroslav Ganin, Tejas Kulkarni, Igor Babuschkin, S. M. Ali Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. In Proceedings of the 35th International Conference on Machine Learning, pages 1666-1675. PMLR, 2018. 2 +[13] Kevin Ellis, Daniel Ritchie, Armando Solar-Lezama, and Josh Tenenbaum. Learning to infer graphics programs from hand-drawn images. In Thirty-second Conference on Neural Information Processing Systems, pages 6062–6071, 2018. 2 +[14] Kevin Ellis, Maxwell Nye, Yewen Pu, Felix Sosa, Josh Tenenbaum, and Armando Solar-Lezama. Write, execute, assess: Program synthesis with a REPL. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2 +[15] Peter Kulits, Haiwen Feng, Weiyang Liu, Victoria Fernandez Abrevaya, and Michael J. Black. Re-thinking inverse graphics with large language models. Transactions on Machine Learning Research, 2024. 2 +[16] Wen-Ding Li and Kevin Ellis. Is programming by example solved by LLMs? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. +[17] Shreyas Kapur, Erik Jenner, and Stuart Russell. Diffusion on syntax trees for program synthesis. In The Thirteenth International Conference on Learning Representations, 2025. 2 +[18] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5515-5523. Computer Vision Foundation / IEEE Computer Society, 2018. +[19] Yonglong Tian, Andrew Luo, Xingyuan Sun, Kevin Ellis, William T. Freeman, Joshua B. Tenenbaum, and Jiajun Wu. Learning to infer and execute 3D shape programs. In International Conference on Learning Representations, 2019. +[20] Javier Camara, Javier Troya, Lola Burgueño, and Antonio Vallecillo. On the assessment of generative AI in modeling tasks: an experience report with chatgpt and UML. Softw. Syst. Model., 22(3):781-793, 2023. 2 +[21] Hugo Laurençon, Leo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2 + +[22] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions, 2024. 5 +[23] Shengbang Tong, Ellis L Brown II, Penghao Wu, Sanghyun Woo, Adithya Jairam Iyer, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, Xichen Pan, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian1: A fully open, vision-centric exploration of multimodal LLMs. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2 +[24] Juan A. Rodriguez, Abhay Puri, Shubham Agarwal, Issam H. Laradji, Pau Rodriguez, Sai Rajeswar, David Vazquez, Christopher Pal, and Marco Pedersoli. StarVector: Generating scalable vector graphics code from images and text. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 16175-16186, 2025. 2 +[25] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, Sam Dodge, Keen You, Zhen Yang, Aleksei Timofeev, Mingze Xu, Hong-You Chen, Jean-Philippe Fauconnier, Zhengfeng Lai, Haoxuan You, Zirui Wang, Afshin Dehghan, Peter Grasch, and Yinfei Yang. MM1.5: Methods, analysis & insights from multimodal LLM fine-tuning. In The Thirteenth International Conference on Learning Representations, 2025. 2 +[26] Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal ArXiv: A dataset for improving scientific comprehension of large vision-language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14369-14387, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 4 +[27] Jaeyoung Kim, Jongho Lee, Hong-Jun Choi, Ting-Yao Hsu, Chieh-Yang Huang, Sungchul Kim, Ryan Rossi, Tong Yu, Clyde Lee Giles, Ting-Hao 'Kenneth' Huang, and Sungchul Choi. Multi-LLM collaborative caption generation in scientific documents. In AI for Research and Scalable, Efficient Systems, pages 142–160, Singapore, 2025. Springer Nature Singapore. 2 +[28] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. Sparks of artificial general intelligence: Early experiments with GPT-4, 2023. 2, 6 +[29] Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad, Stephanie Fu, Adrian Rodriguez-Munoz, Shivam Duggal, Phillip Isola, and Antonio Torralba. A vision check-up for language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14410-14419, 2024. +[30] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with GPT-4, 2023. 2, 6 +[31] Lingjiao Chen, Matei Zaharia, and James Zou. How Is ChatGPT's Behavior Changing Over Time? Harvard Data Science Review, 6(2), 2024. https://hdsr.mitpress.mit.edu/pub/y95zitmz.2 + +[32] Sagi Polaczek, Yuval Alaluf, Elad Richardson, Yael Vinker, and Daniel Cohen-Or. NeuralSVG: An implicit representation for text-to-vector generation, 2025. 2 +[33] Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. IconShop: Text-guided vector icon synthesis with autoregressive transformers. ACM Trans. Graph., 42(6), 2023. 2 +[34] Ajay Jain, Amber Xie, and Pieter Abbeel. VectorFusion: Text-to-SVG by abstracting pixel-based diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1911-1920, 2023. +[35] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. CLIPDraw: Exploring text-to-drawing synthesis through language-image encoders. In NeurIPS, 2022. 2 +[36] Henrik Voigt, Kai Lawonn, and Sina Zarrieß. Plots made quickly: An efficient approach for generating visualizations from natural language queries. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 12787-12793, Torino, Italia, 2024. ELRA and ICCL. 2 +[37] Yuyu Luo, Nan Tang, Guoliang Li, Chengliang Chai, Wenbo Li, and Xuedi Qin. Synthesizing natural language to visualization (nl2vis) benchmarks from NL2SQL benchmarks. In Proceedings of the 2021 International Conference on Management of Data, page 1235-1247, New York, NY, USA, 2021. Association for Computing Machinery. +[38] Jock Mackinlay. Automating the design of graphical presentations of relational information. ACM Trans. Graph., 5(2): 110-141, 1986. +[39] Steven F. Roth, John Kolojejchick, Joe Mattis, and Jade Goldstein. Interactive graphic design using automatic presentation knowledge. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, page 112-117, New York, NY, USA, 1994. Association for Computing Machinery. +[40] Yang Wu, Yao Wan, Hongyu Zhang, Yulei Sui, Wucai Wei, Wei Zhao, Guandong Xu, and Hai Jin. Automated data visualization from natural language via large language models: An exploratory study. Proc. ACM Manag. Data, 2(3), 2024. 2 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 2 +[42] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. +[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents, 2022. 2, 7 +[44] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. CogView: Mastering text-to-image generation via transformers. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 19822-19835, 2021. + +[45] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. CogView2: Faster and better text-to-image generation via hierarchical transformers. In NeurIPS, 2022. 2 +[46] Juan A. Rodriguez, David Vázquez, Issam H. Laradji, Marco Pedersoli, and Pau Rodríguez. FigGen: Text to scientific figure generation. In The First Tiny Papers Track at ICLR 2023, Tiny Papers @ ICLR 2023, Kigali, Rwanda, May 5, 2023. OpenReview.net, 2023. 2 +[47] Juan A. Rodriguez, David Vázquez, Issam H. Laradji, Marco Pedersoli, and Pau Rodríguez. OCR-VQGAN: Taming textwithin-image generation. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2023, Waikoloa, HI, USA, January 2-7, 2023, pages 3678-3687. IEEE, 2023. 2 +[48] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139-144, 2020. 2 +[49] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 2 +[50] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, Lille, France, 2015. PMLR. 2 +[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2 +[52] Yiwei Hu, Paul Guerrero, Milos Hasan, Holly Rushmeier, and Valentin Deschaintre. Generating procedural materials from text or image prompts. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2 +[53] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12):nwae403, 2024. 2 +[54] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3 +[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3 +[56] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 3 + +[57] Aaron Grattaftiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, Danny Wyatt, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Francisco Guzmán, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Govind Thattai, Graeme Nail, Gregoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel Kloumann, Ishan Misra, Ivan Evtimov, Jack Zhang, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala, Karthik Prasad, Kartikeya Upasani, Kate Plawiak, Ke Li, Kenneth Heafield, Kevin Stone, Khalid El-Arini, Krithika Iyer, Kshitiz Malik, Kuenley Chiu, Kunal Bhalla, Kushal Lakhotia, Lauren Rantala-Yeary, Laurens van der Maaten, Lawrence Chen, Liang Tan, Liz Jenkins, Louis Martin, Lovish Madaan, Lubo Malo, Lukas Blecher, Lukas Landzaat, Luke de Oliveira, Madeline Muzzi, Mahesh Pasupuleti, Mannat Singh, Manohar Paluri, Marcin Kardas, Maria Tsimpoukelli, Mathew Oldham, Mathieu Rita, Maya Pavlova, Melanie Kambadur, Mike Lewis, Min Si, Mitesh Kumar Singh, Mona Hassan Naman Goyal Narjes Torabi Nikolay Bashlykov Nikolay Bogoychev,Niladri Chatterji Ning Zhang,Olivier Duchenne Onur Celebi Patrick Alrassy Pengchuan Zhang,Pengwei Li Petar Vasic Peter WengPrajjwal BhargavaPratik Dubal Praveen Krishnan Punit Singh Koura Puxin Xu Qing He Qingxiao Dong Ragavan Srinivasan Raj Ganapathy Ramon Calderer Ricardo Silveira Cabral Roberta Raileanu Rohan Maheswari Rohit Girdhar Rohit Patel Romain SauvestreRonnie Polidoro Roshan Sumbaly Ross Taylor Ruan Silva Rui Hou Rui Wang Saghar Hosseini Sahana Chennabasappa Sanjay Singh Sean Bell Seohyun Sonia Kim,Sergey Edunov Shaoliang Nie Sharan Narang Sharath Rarpathy Sheng Shen Shengye Wan Shruti BhosaleShun ZhangSimon Vandenhende Soumya Batra Spencer WhitmanSten Sootla Stephane Collot Suchin Gururangan Sydney Borodinsky Tamar Herman Tara Fowler Tarek Sheasha Thomas Georgiou Thomas Scialom Tobias Speckbacher Todor Mihaylov Tong Xiao Ujjwal Karn,Vedanuj Goswami Vibhor Gupta,Vignesh Ramanathan,Viktor Kerkez Vincent Gonguet,Virginie Do,Vish VogetiVitor Albiero,Vladan Petrovic Weiwei Chu Wenhan Xiong,Wenyin Fu Whitney + +Meers, Xavier Martinet, Xiaodong Wang, Xiaofang Wang, Xiaoqing Ellen Tan, Xide Xia, Xinfeng Xie, Xuchao Jia, Xuewei Wang, Yaelle Goldschlag, Yashesh Gaur, Yasmine Babaei, Yi Wen, Yiwen Song, Yuchen Zhang, Yue Li, Yuning Mao, Zacharie Delpierre Coudert, Zheng Yan, Zhengxing Chen, Zoe Papakipos, Aaditya Singh, Aayushi Srivastava, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adithya Gangidi, Adolfo Victoria, Ahuva Goldstand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alexei Baevski, Allie Feinstein, Amanda Kallet, Amit Sangani, Amos Teo, Anam Yunus, Andrei Lupu, Andres Alvarado, Andrew Caples, Andrew Gu, Andrew Ho, Andrew Poulton, Andrew Ryan, Ankit Ramchandani, Annie Dong, Annie Franco, Anuj Goyal, Aparajita Saraf, Arkabandhu Chowdhury, Ashley Gabriel, Ashwin Bharambe, Assaf Eisenman, Azadeh Yazdan, Beau James, Ben Maurer, Benjamin Leonhardi, Bernie Huang, Beth Loyd, Beto De Paola, Bhargavi Paranjape, Bing Liu, Bo Wu, Boyu Ni, Braden Hancock, Bram Wasti, Brandon Spence, Brani Stojkovic, Brian Gamido, Britt Montalvo, Carl Parker, Carly Burton, Catalina Mejia, Ce Liu, Changhan Wang, Changkyu Kim, Chao Zhou, Chester Hu, Ching-Hsiang Chu, Chris Cai, Chris Tindal, Christoph Feichtenhofer, Cynthia Gao, Damon Civin, Dana Beaty, Daniel Kreymer, Daniel Li, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le, Dustin Holland, Edward Dowling, Eissa Jamil, Elaine Montgomery. Eleonora Presani, Emily Hahn, Emily Wood, Eric-Tuan Le. Erik Brinkman Esteban Arcaute Evan Dunbar Evan Smothers. Fei Sun Felix Kreuk Feng Tian Filippos Kokkinos. Firat Ozgenel Francesco Caggioni Frank Kanayet Frank Seide,Gabriela Medina Florez,Gabriella Schwarz,Gada Badeer Georgia Swee Gil Halpern Grant Herman Grigory SizovGuangyi,ZhangGuna LakshminarayananHakan Inan Hamid Shojanazeri Han Zou Hannah Wang Hanwen Zha Haroun Habeeb Harrison Rudolph Helen Suk Henry Aspegren Hunter Goldman Hongyuan Zhan Ibrahim Damlaj Igor Molybog,Igor Tufanov Ilias Leontiadis Irina-Elena Veliche Itai Gat Jake WeissmanJames GeboskiJames Kohli Janice LamJaphet AsherJean-Baptiste GayaJeff MarcusJeff TangJennifer Chan Jenny Zhen Jeremy Reizenstein Jeremy Teboul Jessica Zhong Jian Jin Jingyi Yang Joe Cummings Jon Carvill Jon Shepard Jonathan McPhie Jonathan Torres Josh Ginsburg Junjie Wang Kai Wu Kam Hou U Karan Saxena Kartikay Khandelwal Katayoun Zand Kathy Matosich Kaushik Veeraraghavan Kelly Michelena Keqian Li Kiran Jagadeesh Kun HuangKunal ChawlaKyle HuangLailin Chen Lakshya Garg Lavender A Leandro Silva Lee Bell Lei Zhang Liangpeng Guo Licheng Yu Liron Moshkovich Luca Wehrstedt Madian Khabsa Manav Avalani Manish Bhatt Martynas Mankus Matan Hasson Matthew Lennie Matthias Reso Maxim Groshev Maxim Naumov Maya Lathi. Meghan Keneally Miao Liu Michael L. Seltzer Michal Valko Michelle Restrepo Mihir Patel Mik Vyatskov Mikayel Samvelyan Mike Clark Mike Macey Mike Wang Miquel Jubert Hermoso Mo Metanat Mohammad Rastegari Munish Bansal Nandhini Santhanam Natascha Parks Natasha White Navyata Bawa Nayan Singhal Nick Egebo Nicolas Usunier Nikhil Mehta Nikolay Pavlovich Laptev Ning Dong Nor + +man Cheng, Oleg Chernoguz, Olivia Hart, Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pavan Balaji, Pedro Rittner, Philip Bontrager, Pierre Roux, Piotr Dollar, Polina Zvyagina, Prashant Ratanchandani, Pritish Yuvraj, Qian Liang, Rachad Alao, Rachel Rodriguez, Rafi Ayub, Ragtoham Murthy, Raghu Nayani, Rahul Mitra, Rangaprabhu Parthasarathy, Raymond Li, Rebekkah Hogan, Robin Battey, Rocky Wang, Russ Howes, Rudy Rinott, Sachin Mehta, Sachin Siby, Sai Jayesh Bondu, Samyak Datta, Sara Chugh, Sara Hunt, Sargun Dhillon, Sasha Sidorov, Satadru Pan, Saurabh Mahajan, Saurabh Verma, Seiji Yamamoto, Sharadh Ramaswamy, Shaun Lindsay, Shaun Lindsay, Sheng Feng, Shenghao Lin, Shengxin Cindy Zha, Shishir Patil, Shiva Shankar, Shuqiang Zhang, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe, Soumith Chintala, Stephanie Max, Stephen Chen, Steve Kehoe, Steve Satterfield, Sudarshan Govindaprasad, Sumit Gupta, Summer Deng, Sungmin Cho, Sunny Virk, Suraj Subramanian, Sy Choudhury, Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Koehler, Thomas Robinson, Tianhe Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victoria Ajayi, Victoria Montanez, Vijai Mohan, Vinay Satish Kumar, Vishal Mangla, Vlad Ionescu, Vlad Poenaru, Vlad Tiberiu Mihailescu, Vladimir Ivanov, Wei Li, Wenchen Wang, Wenwen Jiang, Wes Bouaziz, Will Constable, Xiaocheng Tang, Xiaojian Wu, Xiaolan Wang, Xilun Wu, Xinbo Gao, Yaniv Kleinman, Yanjun Chen, Ye Hu, Ye Jia, Ye Qi, Yenda Li, Yilin Zhang, Ying Zhang, Yossi Adi, Youngjin Nam, Yu, Wang, Yu Zhao, Yuchen Hao, Yundi Qian, Yunlu Li, Yuzi He, Zach Rait. Zachary DeVito, Zef Rosnbrick. Zhaoduo Wen. Zhenyu Yang. Zhiwei Zhao and Zhiyu Ma. The llama 3 herd of models. 2024. 3. 4 + +[58] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735-1780, 1997. 3 +[59] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network, 2015. 3 +[60] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. CLIP-KD: An empirical study of CLIP model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15952-15962, 2024. 3 +[61] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter, 2020. 3 +[62] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and efficient foundation language models, 2023. 4 +[63] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 4 +[64] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, Ji Ma, Jiaqi Wang, Xiaoyi Dong, Hang Yan, Hewei Guo, Conghui He, Botian Shi, Zhenjiang Jin, Chao Xu, Bin + +Wang, Xingjian Wei, Wei Li, Wenjian Zhang, Bo Zhang, Pinlong Cai, Licheng Wen, Xiangchao Yan, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. How far are we to GPT-4V? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 4 +[65] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, Thomas Unterthiner, Daniel Keysers, Skanda Koppula, Fangyu Liu, Adam Grycner, Alexey Gritsenko, Neil Houlsby, Manoj Kumar, Keran Rong, Julian Eisenschlos, Rishabh Kabra, Matthias Bauer, Matko Bošnjak, Xi Chen, Matthias Minderer, Paul Voigtlaender, Ioana Bica, Ivana Balazevic, Joan Puigcerver, Pinelopi Papalampidi, Olivier Henaff, Xi Xiong, Radu Soricut, Jeremiah Harmsen, and Xiaohua Zhai. PaliGemma: A versatile 3B VLM for transfer, 2024. 4 +[66] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 4 +[67] Stephanie Fu, Netanel Yakir Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. DreamSim: Learning new dimensions of human visual similarity using synthetic data. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4, 18 +[68] Shobhita Sundaram, Stephanie Fu, Lukas Muttenthaler, Netei Yakir Tamir, Lucy Chai, Simon Kornblith, Trevor Darrell, and Phillip Isola. When does perceptual alignment benefit vision representations? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 4, 18 +[69] Mikołaj Binkowski, Dougal J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations, 2018. 4 +[70] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 4 +[71] Aryaz Eghbali and Michael Pradel. CrystalBLEU: Precisely and efficiently measuring the similarity of code. In Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering, New York, NY, USA, 2023. Association for Computing Machinery. 4 +[72] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 4 +[73] Peter Stanchev, Weiyue Wang, and Hermann Ney. EED: Extended edit distance measure for machine translation. In Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1), pages 514-520, Florence, Italy, 2019. Association for Computational Linguistics. 4 + +[74] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 5 +[75] Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, Kai Dang, Yang Fan, Yichang Zhang, An Yang, Rui Men, Fei Huang, Bo Zheng, Yibo Miao, Shanghaoran Quan, Yunlong Feng, Xingzhang Ren, Xuancheng Ren, Jingren Zhou, and Junyang Lin. Qwen2.5-Coder technical report, 2024. 6, 15 +[76] Michael Saxon, Fatima Jahara, Mahsa Khoshnoodi, Yujie Lu, Aditya Sharma, and William Yang Wang. Who evaluates the evaluations? objectively scoring text-to-image prompt coherence metrics with t2IScore score (TS2). In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 6 +[77] Jordan J. Louviere, Terry N. Flynn, and A. A. J. Marley. Best—Worst Scaling: Theory, Methods and Applications. Cambridge University Press, 2015. 6 +[78] Svetlana Kiritchenko and Saif M. Mohammad. Capturing reliable fine-grained sentiment associations by crowdsourcing and best-worst scaling. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 811-817, San Diego, California, 2016. Association for Computational Linguistics. 6 +[79] Svetlana Kiritchenko and Saif Mohammad. Best-worst scaling more reliable than rating scales: A case study on sentiment intensity annotation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 465-470, Vancouver, Canada, 2017. Association for Computational Linguistics. 6 +[80] Bryan K. Orme. MaxDiff analysis: Simple counting, individual-level logit, and HB. Sawtooth Software Research Paper Series, 2009. 6 +[81] Bruce Schneier. Applied cryptography - protocols, algorithms, and source code in C, 2nd Edition. Wiley, 1996. 7 +[82] Kyle Lo, Lucy Lu Wang, Mark Neumann, Rodney Kinney, and Daniel Weld. S2ORC: The semantic scholar open research corpus. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4969-4983, Online, 2020. Association for Computational Linguistics. 8 +[83] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 8 +[84] Anas Awadalla, Le Xue, Manli Shu, An Yan, Jun Wang, Senthil Purushwalkam, Sheng Shen, Hannah Lee, Oscar Lo, Jae Sung Park, Etash Kumar Guha, Silvio Savarese, Ludwig Schmidt, Yejin Choi, Caiming Xiong, and Ran Xu. BLIP3-KALE: Knowledge augmented large-scale dense captions. In Synthetic Data for Computer Vision Workshop @ CVPR 2025, 2025. + +[85] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, Yuyin Zhou, and Cihang Xie. What if we recapture billions of web images with LLaMA-3?, 2024. 8 +[86] R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 1735-1742, 2006. 8 +[87] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. ImageBind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15180-15190, 2023. 8 +[88] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, Cai Wan Zhang, Zhifeng Li, Wei Liu, and Li Yuan. LanguageBind: Extending video-language pretraining to n-modality by language-based semantic alignment. In The Twelfth International Conference on Learning Representations, 2024. 8 +[89] OpenAI. Learning to reason with LLMs, 2024. 15 +[90] DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu, Shengfeng Ye, Tao Yun, Tian Pei, Tianyu Sun, T. Wang, Wangding Zeng, Wanjia Zhao, Wen Liu, Wenfeng Liang, Wenjun Gao, Wenqin Yu, Wentao Zhang, W. L. Xiao, Wei An, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaotao Nie, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xinyu Yang, Xinyuan Li, Xuecheng Su, Xuheng Lin, X. Q. Li, Xiangyue Jin, Xiaojin Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxiang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y. K. Li, Y. Q. Wang, Y. X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yuan Ou, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang + +Zhou, Y. X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning, 2025. 15 +[91] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 15 +[92] Urbano Lorenzo-Seva and Jos M. F. ten Berge. Tucker's congruence coefficient as a meaningful index of factor similarity. Methodology: European Journal of Research Methods for the Behavioral and Social Sciences, 2(2):57-64, 2006. 15 +[93] Jonas Belouadi and Steffen Eger. UScore: An effective approach to fully unsupervised evaluation metrics for machine translation. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 358-374, Dubrovnik, Croatia, 2023. Association for Computational Linguistics. 15 +[94] Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M. Meyer, and Steffen Eger. MoverScore: Text generation evaluating with contextualized embeddings and earth mover distance. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 563-578, Hong Kong, China, 2019. Association for Computational Linguistics. +[95] Wei Zhao, Goran Glavaš, Maxime Peyrard, Yang Gao, Robert West, and Steffen Eger. On the limitations of cross-lingual encoders as exposed by reference-free machine translation evaluation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1656–1671, Online, 2020. Association for Computational Linguistics. +[96] Yurun Song, Junchen Zhao, and Lucia Specia. SentSim: Crosslingual semantic evaluation of machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3143-3156, Online, 2021. Association for Computational Linguistics. 15 +[97] Y. Rubner, C. Tomasi, and L.J. Guibas. A metric for distributions with applications to image databases. In Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271), pages 59-66, 1998. 15 +[98] Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to document distances. In Proceedings of the 32nd International Conference on Machine Learning, pages 957–966, Lille, France, 2015. PMLR. 15 \ No newline at end of file diff --git a/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/images.zip b/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d850af9f8ea7fd9d50094bf9ae809359a524531e --- /dev/null +++ b/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:240f742b3739fa14cb4e82e9fd45d1b3d7a0d33274647fc733c343ed03735ef0 +size 420997 diff --git a/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/layout.json b/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fcda537b2b33ca6a2b49b80080697dc968a84ee9 --- /dev/null +++ b/2025/TikZero_ Zero-Shot Text-Guided Graphics Program Synthesis/layout.json @@ -0,0 +1,10648 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "type": "text", + "content": "TikZero: Zero-Shot Text-Guided Graphics Program Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 141, + 533, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 141, + 533, + 220 + ], + "spans": [ + { + "bbox": [ + 82, + 141, + 533, + 220 + ], + "type": "text", + "content": "Jonas Belouadi* Eddy Ilg† Margret Keuper*,‡ Hideki Tanaka§ Masao Utiyama§ Raj Dabre§ Steffen Eger† Simone Ponzetto* \nUniversity of Mannheim, Germany* University of Technology Nuremberg, Germany† Max Planck Institute for Informatics, Saarland Informatics Campus, Germany‡ National Institute of Information and Communications Technology, Japan§ jonas.belouadi@uni-mannheim.de" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 246, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 246, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 152, + 246, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 271, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 510 + ], + "type": "text", + "content": "Automatically synthesizing figures from text captions is a compelling capability. However, achieving high geometric precision and editability requires representing figures as graphics programs in languages like TikZ, and aligned training data (i.e., graphics programs with captions) remains scarce. Meanwhile, large amounts of unaligned graphics programs and captioned raster images are more readily available. We reconcile these disparate data sources by presenting TikZERO, which decouples graphics program generation from text understanding by using image representations as an intermediary bridge. It enables independent training on graphics programs and captioned images and allows for zero-shot text-guided graphics program synthesis during inference. We show that our method substantially outperforms baselines that can only operate with caption-aligned graphics programs. Furthermore, when leveraging caption-aligned graphics programs as a complementary training signal, TikZERO matches or exceeds the performance of much larger models, including commercial systems like GPT-4o. Our code, datasets, and select models are publicly available.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 533, + 135, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 135, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 135, + 547 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "type": "text", + "content": "Graphics programming languages offer distinct advantages over low-level vector formats (PDF, SVG) or raster image formats by representing visual concepts as high-level programs that preserve semantics, remain human-interpretable, and allow manual editing. These properties are particularly valuable in academia, where specialized graphics programming languages like TikZ [1] are popular for creating complex figures with high expressivity. However, this comes with a steep learning curve, as seen on the TEX Stack Exchange" + }, + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "type": "text", + "content": " (TEX.SE), where nearly " + }, + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 554, + 296, + 685 + ], + "type": "text", + "content": " of questions concern TikZ and make it the most frequently discussed topic on the platform [2, 3]." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 244, + 552, + 448 + ], + "blocks": [ + { + "bbox": [ + 317, + 244, + 552, + 448 + ], + "lines": [ + { + "bbox": [ + 317, + 244, + 552, + 448 + ], + "spans": [ + { + "bbox": [ + 317, + 244, + 552, + 448 + ], + "type": "image", + "image_path": "3847c540f76ba4bdd94cb752b78fe3ba6f97ed7688ef35cbf45137413f14e349.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 456, + 555, + 513 + ], + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 513 + ], + "type": "text", + "content": "Figure 1. Qualitative comparison of our " + }, + { + "bbox": [ + 313, + 456, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZero}^+" + }, + { + "bbox": [ + 313, + 456, + 555, + 513 + ], + "type": "text", + "content": " model (last two columns) and the end-to-end trained baseline " + }, + { + "bbox": [ + 313, + 456, + 555, + 513 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATikZ}_{\\mathrm{v2}}" + }, + { + "bbox": [ + 313, + 456, + 555, + 513 + ], + "type": "text", + "content": " (LLM; first two columns) on text-guided graphics program synthesis with TikZ. Our method generates outputs that more closely follow the given captions. Example program listings are in Supp. Mat. F." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": "With recent advances in generative AI, simplifying the creation of graphics programs has become increasingly feasible. Belouadi et al. [2] introduce " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{D}\\mathsf{E}\\mathrm{T}\\mathsf{1}\\kappa \\mathrm{Z}\\mathsf{I}\\mathsf{F}\\mathsf{Y}" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": ", an inverse graphics model that generates TikZ programs from images and hand-drawn sketches. However, creating these visual inputs stays cumbersome, motivating alternative input modalities such as natural language. While Belouadi et al. [3] propose AutomATikZ, a text-guided synthesis model for TikZ programs trained end-to-end on an aligned caption-program corpus, its performance remains limited (cf. Fig. 1) [4, 5]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "We identify insufficient training data as the primary limitation. Unlike inverse graphics models such as " + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{DETikZIFY}" + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": ", which are inherently self-supervised (trained by being conditioned on compiled representations of their output programs) and can access sufficient training data (cf. Fig. 2), end-to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 232, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 232, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 232, + 703 + ], + "type": "text", + "content": "1https://github.com/potamides/DeTikZify" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 704, + 195, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 195, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 195, + 712 + ], + "type": "text", + "content": "2https://tex.stackexchange.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17793" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 295, + 224 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 295, + 224 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 295, + 224 + ], + "type": "image", + "image_path": "a8bc9b60538868e0b4edd0b582689f6259d0cf86d8c62a48a5a0dd71b324d30a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 296, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 309 + ], + "type": "text", + "content": "Figure 2. Illustration of training data availability for graphics program synthesis. " + }, + { + "bbox": [ + 55, + 232, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\mathrm{DETIKZIFY}" + }, + { + "bbox": [ + 55, + 232, + 296, + 309 + ], + "type": "text", + "content": " can leverage all graphics programs for training but lacks text guidance, while " + }, + { + "bbox": [ + 55, + 232, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ}" + }, + { + "bbox": [ + 55, + 232, + 296, + 309 + ], + "type": "text", + "content": " is constrained to the small intersection of captioned graphics programs, resulting in limited performance. Our approach, " + }, + { + "bbox": [ + 55, + 232, + 296, + 309 + ], + "type": "inline_equation", + "content": "\\mathrm{TIKZERO}" + }, + { + "bbox": [ + 55, + 232, + 296, + 309 + ], + "type": "text", + "content": ", trains independently on both graphics programs and captioned images, enabling more effective use of available data and yielding superior results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 331, + 296, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 366 + ], + "type": "text", + "content": "end text-guided models like AUTOMATIKZ require graphics programs paired with captions, substantially reducing the available data pool (cf. Fig. 2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 367, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 367, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 54, + 367, + 296, + 545 + ], + "type": "text", + "content": "To address this challenge, we decouple the graphics program generation component from text understanding, enabling independent training on graphics programs and captioned images without requiring paired data (cf. Fig. 2). Our approach first trains an inverse graphics model conditioned on image patch embeddings from a vision encoder [6]. We then train an adapter network that generates synthetic image patch embeddings from captions. This adapter training relies solely on captioned images, effectively circumventing resource limitations and enabling zero-shot (in the sense that no aligned caption-program examples are involved in the training process) text-guided graphics program synthesis [7]. We demonstrate that this approach, to which we refer as " + }, + { + "bbox": [ + 54, + 367, + 296, + 545 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}ZERO}" + }, + { + "bbox": [ + 54, + 367, + 296, + 545 + ], + "type": "text", + "content": ", outperforms previous state-of-the-art methods (cf. Fig. 1). Our key contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 546, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 53, + 546, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 546, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 53, + 546, + 296, + 594 + ], + "type": "text", + "content": "(i) A novel two-stage architecture, " + }, + { + "bbox": [ + 53, + 546, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZERO}" + }, + { + "bbox": [ + 53, + 546, + 296, + 594 + ], + "type": "text", + "content": ", which addresses the low-resource challenge in text-guided graphics program synthesis by aligning representation spaces rather than relying on aligned data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "text", + "content": "(ii) The " + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{DATiKZ_{v3}}" + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "text", + "content": " dataset, comprising over 450k TiKZ graphics programs with roughly 170k captioned samples. Using this dataset, we train both " + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZERO}" + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATiKZ_{v2}}" + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "text", + "content": " (an updated version of " + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATiKZ}" + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "text", + "content": ") on the same source data and show that " + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZERO}" + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "text", + "content": " outperforms " + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATiKZ}" + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATiKZ_{v2}}" + }, + { + "bbox": [ + 51, + 594, + 296, + 666 + ], + "type": "text", + "content": ", and other end-to-end trained baselines." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 666, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 666, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 666, + 294, + 714 + ], + "type": "text", + "content": "(iii) An enhanced model, " + }, + { + "bbox": [ + 48, + 666, + 294, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Z_{ZERO} + }" + }, + { + "bbox": [ + 48, + 666, + 294, + 714 + ], + "type": "text", + "content": ", combining " + }, + { + "bbox": [ + 48, + 666, + 294, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Z_{ZERO}}" + }, + { + "bbox": [ + 48, + 666, + 294, + 714 + ], + "type": "text", + "content": " with the end-to-end fine-tuning of " + }, + { + "bbox": [ + 48, + 666, + 294, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMAT_{IK}Z_{v2}}" + }, + { + "bbox": [ + 48, + 666, + 294, + 714 + ], + "type": "text", + "content": ", which surpasses larger baselines and matches the performance of commercial models like GPT-4o [8] on key metrics." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 555, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 271 + ], + "type": "text", + "content": "Inverse Graphics Program Synthesis Inverse graphics, i.e., synthesizing a graphics program to reproduce a visual target, represents a specialized instance of neural program synthesis [9-11]. Deep learning models have shown remarkable success in this domain [12-14], with Vision-Language Models (VLMs) increasingly gaining prominence [2, 15-17]. While controlled experimental studies often rely on synthetic datasets [13, 15, 17-20], real-world applications typically leverage more complex and diverse human-created data [2, 21-25], highlighting the importance of data availability. In scientific contexts, TikZ has emerged as a popular choice due to its versatility, expressiveness, and widespread adoption in academic circles [2, 21-23, 25]. Although these approaches are not tailored to text-guided generation, we incorporate key elements from them into our approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 277, + 556, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 556, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 556, + 504 + ], + "type": "text", + "content": "Text-Guided Graphics Program Synthesis Current text-guided approaches to graphics program synthesis remain limited, mainly because of the scarcity of captioned graphics programs outlined in Sec. 1, but also because of the difficulty of generating synthetic data with human-like captions [26, 27]. Researchers interested in this capability currently rely on the emerging capabilities of large commercial models such as GPT-4o [8, 28–30], which raises concerns about accessibility, reproducibility, and computational cost [31]. In contrast, related domains like vector graphics generation [24, 32–35] and NL2Vis [36–40] have shown more progress. Similar to inverse graphics, these fields increasingly incorporate large language models (LLMs) [24, 36, 40]. However, vector graphics approaches typically generate only low-level Bezier curves, limiting output complexity [2, 32, 33], and NL2Vis focuses exclusively on data visualization with a restricted set of visualization types [40]. More complex applications, such as generating arbitrary scientific figures from captions with TikZ, remain underexplored—a gap we address in this work." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Text-to-Image Generation " + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZero}" + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": " shares conceptual and architectural similarities with several text-to-image generation methods [41-45]. Rodriguez et al. [46, 47] explore generative adversarial networks [48, 49] and diffusion models [50, 51] for scientific figure generation, but these approaches are tied to raster images, which are not ideal for representing scientific figures. Ramesh et al. [43] propose a two-stage model with independently trained prior and decoder components to generate raster images from text. Although prior networks resemble our adapters and have been used with inverse graphics models [52], they target global image embeddings containing only abstract information, which degrades performance when used with inverse graphics architectures that work best with patch-level details [53]. In contrast, our adapters specifically operate on patch-level embeddings, and we demonstrate that this improves performance compared to end-to-end trained baselines." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17794" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 555, + 248 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 555, + 248 + ], + "type": "image", + "image_path": "5a4db9ff5a15611160ed325154439f80a3c320b123f535faeb2122769ccb72be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 252, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 54, + 252, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 555, + 321 + ], + "type": "text", + "content": "Figure 3. Architecture overview of " + }, + { + "bbox": [ + 54, + 252, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\mathrm{T1KZERO}" + }, + { + "bbox": [ + 54, + 252, + 555, + 321 + ], + "type": "text", + "content": " during inference. Solid lines represent the standard caption-conditioned path, which flows through the text encoder into the adapter network of " + }, + { + "bbox": [ + 54, + 252, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\mathrm{T1KZERO}" + }, + { + "bbox": [ + 54, + 252, + 555, + 321 + ], + "type": "text", + "content": " before connecting to the vision encoder. In certain configurations (cf. Sec. 5.2), the caption also feeds into the text decoder (depicted by dotted lines and “" + }, + { + "bbox": [ + 54, + 252, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 54, + 252, + 555, + 321 + ], + "type": "text", + "content": "” markers representing shortcuts). The self- and cross-attention layers (yellow) are simplified representations, omitting internal feed-forward layers and residual connections [54]. An exception is the explicit residual connection between the cross-attention and self-attention layers of the vision encoder, visualizing the gating mechanism " + }, + { + "bbox": [ + 54, + 252, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 252, + 555, + 321 + ], + "type": "text", + "content": " (purple). Additionally, the dashed path illustrates how the inverse graphics model generates graphics programs when conditioned on images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 349, + 257, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 257, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 257, + 361 + ], + "type": "text", + "content": "3. The TIKZERO Model & Architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 297, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 297, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 297, + 418 + ], + "type": "text", + "content": "As the foundation of " + }, + { + "bbox": [ + 55, + 369, + 297, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZero}" + }, + { + "bbox": [ + 55, + 369, + 297, + 418 + ], + "type": "text", + "content": ", we first develop a state-of-the-art inverse graphics model for graphics program synthesis. We then incorporate a cross-attention adapter network [55] for text guidance. Fig. 3 provides an overview of our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "text", + "content": "The Inverse Graphics Model Due to their demonstrated effectiveness (cf. Sec. 2), we adopt a VLM architecture for the inverse graphics model of " + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Zero}" + }, + { + "bbox": [ + 55, + 422, + 296, + 495 + ], + "type": "text", + "content": ". Figure 3 illustrates its inner workings (dashed lines): the model processes rasterized images and autoregressively generates their corresponding programs without involving captions at this stage." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 297, + 714 + ], + "type": "text", + "content": "The Adapter Network VLMs consist of two primary components: a vision encoder that produces image embeddings and a text decoder that, in our case, generates graphics programs conditioned on these embeddings. The unidirectional and localized flow of information between these components allows us to inject additional information solely into the vision encoder, thereby influencing the output of the text decoder. We exploit this property by introducing a trainable, text-conditioned adapter network that mimics the outputs of the original vision encoder. This effectively enables zero-shot generation of graphics programs conditioned on text when its outputs are fed into the decoder. In addition to circumventing the resource limitations discussed in Sec. 1, this architecture has other welcome implications: During adapter training, the text decoder, usually the largest component of the model, does not need to be loaded, resulting in efficient and fast training even with large datasets. Our adapter incorporates a lightweight text encoder for embedding captions and intro-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 350, + 556, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 556, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 556, + 506 + ], + "type": "text", + "content": "duces newly initialized gated cross-attention layers [56, 57] before each vision encoder layer (cf., Fig. 3). The keys and values derive from the final text encoder representations, while the queries originate from a trainable probe used instead of image inputs. The gates " + }, + { + "bbox": [ + 313, + 350, + 556, + 506 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 350, + 556, + 506 + ], + "type": "text", + "content": " allow the model to learn at which layers and to what extent information from the text encoder should flow into the vision encoder. Contrary to existing literature, which often employs tanh gates [58] that initialize to zero (indicating no information flow), we find that using sigmoid gates (0.5 at initialization) accelerates training convergence since, in our case, only little information originates from the vision encoder inputs (i.e., the probe). We ablate the gates and the probe in Supp. Mat. A." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "text", + "content": "Training Objective Given a caption-image dataset for training, we first embed the patches " + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "inline_equation", + "content": "p \\in \\mathbf{p}" + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "text", + "content": " of image " + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "text", + "content": " using the unmodified vision encoder " + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "text", + "content": " of our VLM. Subsequently, we incorporate the cross-attention adapter to obtain the modified encoder " + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{M}}" + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "text", + "content": ", which we then distill on these image patch embeddings conditioned solely on the caption " + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "text", + "content": " and probe " + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "inline_equation", + "content": "\\hat{t}" + }, + { + "bbox": [ + 313, + 511, + 556, + 596 + ], + "type": "text", + "content": " [59]. This leads to the following objective:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 329, + 604, + 555, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 604, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 329, + 604, + 555, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d i s t}} = \\frac {1}{| p |} \\sum_ {p \\in p} \\operatorname {d i s t} \\left(\\mathrm {M} _ {\\theta} (p \\mid i), \\widehat {\\mathrm {M}} _ {\\theta , \\hat {\\theta}} (p \\mid \\hat {t}, t)\\right), \\tag {1}", + "image_path": "3ea169cef01e6d5c5333348e8bfd8fedc761de671ac5d8eeeef5772e57df3872.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 641, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 557, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{dist}(\\pmb{x},\\pmb{y})" + }, + { + "bbox": [ + 313, + 641, + 557, + 714 + ], + "type": "text", + "content": " represents a distance metric. Following common practices in model distillation [60, 61], we experiment with cosine distance and mean squared error. Here, " + }, + { + "bbox": [ + 313, + 641, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 641, + 557, + 714 + ], + "type": "text", + "content": " denotes the original model parameters that remain fully frozen, while " + }, + { + "bbox": [ + 313, + 641, + 557, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\theta}" + }, + { + "bbox": [ + 313, + 641, + 557, + 714 + ], + "type": "text", + "content": " represents the adapter parameters of which the cross-attention layers and the image probe are trainable." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17795" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 70, + 277, + 160 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 277, + 160 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 277, + 160 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 277, + 160 + ], + "type": "table", + "html": "
SourceDATIKZDATIKZv2DATIKZv3
curated98115663646
TEX.SE292383060942654
arXiv85656326450407851
artificial195719582256
all117832360583456469
", + "image_path": "a57abde527289ba1c77e24470e3d10042b4adc73c2562d1645786e418c9fb559.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 297, + 204 + ], + "lines": [ + { + "bbox": [ + 55, + 168, + 297, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 297, + 204 + ], + "type": "text", + "content": "Table 1. Breakdown of the number of unique " + }, + { + "bbox": [ + 55, + 168, + 297, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{Ti}k\\mathrm{Z}" + }, + { + "bbox": [ + 55, + 168, + 297, + 204 + ], + "type": "text", + "content": " graphics in " + }, + { + "bbox": [ + 55, + 168, + 297, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{DATI}\\kappa \\mathrm{Z}_{\\mathrm{v3}}" + }, + { + "bbox": [ + 55, + 168, + 297, + 204 + ], + "type": "text", + "content": " compared to its predecessors " + }, + { + "bbox": [ + 55, + 168, + 297, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{DATI}\\kappa \\mathrm{Z}" + }, + { + "bbox": [ + 55, + 168, + 297, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 168, + 297, + 204 + ], + "type": "inline_equation", + "content": "\\mathrm{DATI}\\kappa \\mathrm{Z}_{\\mathrm{v2}}" + }, + { + "bbox": [ + 55, + 168, + 297, + 204 + ], + "type": "text", + "content": ". Qualitative examples can be found in Supp. Mat. F." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 225, + 211, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 211, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 211, + 240 + ], + "type": "text", + "content": "4. Datasets & Model Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "text", + "content": "We introduce " + }, + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\mathrm{DATikZ_{v3}}" + }, + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "text", + "content": ", a novel dataset of TikZ graphics programs designed to support the training and evaluation of " + }, + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\mathrm{TIKZERO}" + }, + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "text", + "content": ". Additionally, we train " + }, + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATikZ_{v2}}" + }, + { + "bbox": [ + 55, + 247, + 296, + 297 + ], + "type": "text", + "content": " as a directly comparable baseline operating on the same data source." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{DATiKZ_{v3}}" + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "text", + "content": " Dataset " + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{DATiKZ_{v3}}" + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "text", + "content": " expands upon its predecessors " + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{DATiKZ_{v2}}" + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "text", + "content": " [2, 3], incorporating programs from curated repositories, TEX.SE, arXiv papers, and artificial samples (cf. Tab. 1). While previous versions focused exclusively on TikZ graphics with (v1) or without (v2) captions, " + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathrm{DATiKZ_{v3}}" + }, + { + "bbox": [ + 55, + 307, + 296, + 415 + ], + "type": "text", + "content": " systematically extracts captions alongside TikZ graphics whenever possible to support our claims. From over 450k instances, fewer than 170k include captions, underscoring the challenges discussed in Sec. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": "Training TIKZERO " + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Zero}" + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": " 's VLM builds upon DE" + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}ZIFY}" + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": " [2] by conditioning a LAMA-based text decoder [62] on patch embeddings from a S1G1P vision encoder [63]. Specifically, we combine " + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{LLAMA}_{3.1}" + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": " (8B) [57] with S1G1P SoViT (0.4B). Unlike " + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{DE_{IK}ZIFY}" + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": " and inspired by the continued ViT pretraining approach of INTERNVL 1.5 [64], we initialize the vision encoder with weights from the fine-tuned encoder of PALIGEMMA [65] and increase the input resolution to " + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "inline_equation", + "content": "420\\times 420" + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": " pixels. Furthermore, we fully fine-tune the vision encoder alongside the rest of the model instead of freezing it. We train on " + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{DAT_{IK}Z_{v3}}" + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": " for 5 epochs with a learning rate of 5e-5 and a batch size of 128. " + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}ZERO}" + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": " 's VLM consistently outperforms " + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{DE_{IK}ZIFY}" + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": ", with detailed evaluation results provided in Supp. Mat. B. For the adapter network, we initialize with " + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{LLAMA}_{3.2}" + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": " (1B) as the text encoder [57] and leverage ARxivCAP [26], a dataset comprising 6.4 million scientific caption-image pairs for training. The adapter accounts for 2 billion of TIKZERO's 10 billion total parameters, with only 400 million being trainable. We train for 3 epochs with a learning rate of 1e-4 and a batch size of 512. We emphasize that this two-stage training process does not access caption-program pairs. However, we demonstrate that incorporating such aligned data in a subsequent fine-tuning step (Sec. 5.2) further enhances performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 289 + ], + "type": "text", + "content": "Training AUTOMATIKZv2 Similar to its predecessor, AUTOMATIKZv2 is a token-conditioned LLM that uses tokenized captions as conditioning information for graphics prediction (rather than patch embeddings). We initialize AUTOMATIKZv2 in two different ways: (i) AUTOMATIKZv2 (LLM), which starts from vanilla LLAMA3.1 (8B) weights, and (ii) AUTOMATIKZv2 (VLM), which leverages TIKZERO's trained VLM (minus the vision encoder) to benefit from transfer learning [66] on its larger training corpus. Both variants employ the same hyperparameters as TIKZERO's VLM but can only utilize the caption-annotated subset of DATIKZv3 for training. Despite having access to less caption-aligned data than TIKZERO's adapter network, AUTOMATIKZv2 requires a longer training period primarily due to fine-tuning the large decoder (8 billion trainable parameters versus the adapter network's 400 million). Training requires more than two days for AUTOMATIKZv2 and 1.5 days for TIKZERO's adapter network when using eight Nvidia A100 40GB GPUs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 300, + 395, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 395, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 395, + 315 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 322, + 556, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 556, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 556, + 418 + ], + "type": "text", + "content": "Before training models on " + }, + { + "bbox": [ + 313, + 322, + 556, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{DAT}_{\\mathrm{IK}}\\mathrm{Z}_{\\mathrm{v3}}" + }, + { + "bbox": [ + 313, + 322, + 556, + 418 + ], + "type": "text", + "content": ", we extract 1k samples from its captioned subset to form our test set. To mitigate data leakage from pretraining to testing, we only include instances created after the cut-off date specified by " + }, + { + "bbox": [ + 313, + 322, + 556, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{LLAMA}_{3.2}" + }, + { + "bbox": [ + 313, + 322, + 556, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 322, + 556, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{ARXIVCAP}" + }, + { + "bbox": [ + 313, + 322, + 556, + 418 + ], + "type": "text", + "content": ". We also employ an " + }, + { + "bbox": [ + 313, + 322, + 556, + 418 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 322, + 556, + 418 + ], + "type": "text", + "content": "-gram matching algorithm to avoid cross-contamination with our training split [8]. For all models, the temperature is set to 0.8 and top-p to 0.95. Example outputs are provided in Fig. 1 and Supp. Mat. F." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "Evaluation Metrics The multimodal nature of our task allows for various evaluation metrics in our automatic evaluations. We assess perceptual image similarity between generated outputs and references by computing DREAMSIM (DSIM) [67, 68], which correlates highly with human judgments for scientific images [2]. We also calculate the Kernel Inception Distance (KID) [69] using SigLIP image features, which evaluates the overall quality of generated figures by comparing to the distribution of reference figures. We evaluate caption similarity between generated outputs and reference captions using CLIPSCORE (CLIP) [70] with SigLIP features. To measure code similarity between generated and reference TikZ programs, we use CRYSTALBLEU (cBLEU), a BLEU variant optimized for code evaluation [71, 72], and TEX Edit Distance (TED) [2], a variant of the Extended Edit Distance [73] utilizing a TEX tokenizer. Since some metrics require that generated programs compile to images, resampling is necessary if the output contains irrecoverable errors. To quantify this, we compute the Mean Token Efficiency (MTE), defined as the " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " winsorized mean of the ratio between the number of tokens in the final TikZ program and the total number of tokens generated to produce that program. For a comprehensive view of model performance, we calculate the arithmetic mean (AVG) of all previous" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17796" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 545, + 198 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 545, + 198 + ], + "type": "table", + "html": "
ModelsOriginal TextRedacted Text
DSIM↑KID↓CLIP↑cBLEU↑TED↓MTE↑AVG↑CLIP↑Ratio↑
IDEFICS 3 (8B)45.47511.42614.3270.65663.17569.55866.6284.85133.858
AUTOMATIKZ (13B)46.0331.2943.9550.38662.2485.86663.0932.96574.975
AUTOMATIKZv2 (VLM)38.31333.2030.7750.32876.98521.5950.00.28436.597
AUTOMATIKZv2 (LLM)50.5483.49115.7660.65862.30781.77582.3758.00250.753
TIKZERO (MSE)52.0245.66410.5831.72366.0779.31885.0048.23777.831
TIKZERO (Cos)52.8295.10310.0511.60365.5182.29185.5997.22671.893
", + "image_path": "7325da1a0578251ed24aa247c5f1d746a5c3af9c4fab3b4cd8f4cfaf2ce9cbbe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 205, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 54, + 205, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 205, + 555, + 240 + ], + "type": "text", + "content": "Table 2. System-level scores " + }, + { + "bbox": [ + 54, + 205, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\times {100}" + }, + { + "bbox": [ + 54, + 205, + 555, + 240 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 54, + 205, + 555, + 240 + ], + "type": "inline_equation", + "content": "{\\mathrm{{T\\kappa Zero}}}_{0}" + }, + { + "bbox": [ + 54, + 205, + 555, + 240 + ], + "type": "text", + "content": " and baselines of comparable size and training setup. Bold and underlined values denote the best and second-best scores for each metric column, respectively. Cell shading illustrates relative score magnitudes. Arrows indicate metric directionality. Overall, " + }, + { + "bbox": [ + 54, + 205, + 555, + 240 + ], + "type": "inline_equation", + "content": "{\\mathrm{{T\\kappa Zero}}}_{0}" + }, + { + "bbox": [ + 54, + 205, + 555, + 240 + ], + "type": "text", + "content": " achieves the strongest average performance across metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 259, + 297, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 259, + 297, + 308 + ], + "spans": [ + { + "bbox": [ + 54, + 259, + 297, + 308 + ], + "type": "text", + "content": "metrics. As these metrics operate on different scales, we apply min-max normalization before computing the average. Additionally, some metrics are recomputed with redacted text in the outputs as part of our analysis, cf. Sec. 6.1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 293, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 293, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 293, + 327 + ], + "type": "text", + "content": "5.1. Comparison against End-to-End Fine-Tuning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 330, + 297, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 330, + 297, + 379 + ], + "spans": [ + { + "bbox": [ + 54, + 330, + 297, + 379 + ], + "type": "text", + "content": "In our initial experiment, we evaluate the zero-shot performance of " + }, + { + "bbox": [ + 54, + 330, + 297, + 379 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZero}" + }, + { + "bbox": [ + 54, + 330, + 297, + 379 + ], + "type": "text", + "content": ", trained as described in Secs. 3 & 4 using either cosine distance (Cos) or mean squared error (MSE), and compare it against end-to-end trained baselines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "type": "text", + "content": "Baselines Besides " + }, + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ_{v2}}" + }, + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "type": "text", + "content": " (LLM & VLM), which we designed as directly comparable baselines, we assess other token-conditioned models of similar and slightly larger sizes trained on " + }, + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZ}" + }, + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "type": "text", + "content": ". Specifically, we evaluate " + }, + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ}" + }, + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "type": "text", + "content": " (13B)" + }, + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "type": "text", + "content": " the strongest original " + }, + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ}" + }, + { + "bbox": [ + 54, + 382, + 296, + 467 + ], + "type": "text", + "content": " baseline [3], and the general-purpose chatbot IDEFICS 3 (8B) [22]. Additional models and details are available in Supp. Mats. A & C." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "spans": [ + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "text", + "content": "Results We present the system-level metric scores in Tab. 2 (Original Text). On average, " + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Z_{ero}}" + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "text", + "content": ", trained with cosine distance, achieves the best performance with an AVG score of 85.599, closely followed by the MSE variant at 85.004. The next best model, " + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMAT_{IK}Z_{v2}}" + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "text", + "content": " (LLM), scores 82.375, which is 3 percentage points (pp) lower. The remaining models exhibit a substantial performance gap, with IDEFICS 3 (8B) and " + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMAT_{IK}Z}" + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "text", + "content": " (13B) falling behind by approximately 20pp and " + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMAT_{IK}Z_{v2}}" + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "text", + "content": " (VLM) showing the weakest performance across all metrics, resulting in an AVG score of 0. The surprisingly poor results of " + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMAT_{IK}Z_{v2}}" + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "text", + "content": " (VLM) are likely due to catastrophic forgetting [74], as the removal of the vision encoder from " + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Z_{ERO}}" + }, + { + "bbox": [ + 54, + 470, + 297, + 638 + ], + "type": "text", + "content": "’s VLM necessitates reacquisition of conditioning based solely on text." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 638, + 297, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 638, + 297, + 698 + ], + "spans": [ + { + "bbox": [ + 54, + 638, + 297, + 698 + ], + "type": "text", + "content": "As for individual metrics, our adapter-based models perform particularly well in perceptual image similarity, with " + }, + { + "bbox": [ + 54, + 638, + 297, + 698 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Z_{\\mathrm{ero}}}" + }, + { + "bbox": [ + 54, + 638, + 297, + 698 + ], + "type": "text", + "content": " (Cos) outperforming the best baseline, AUTOMATIKZv2 (LLM), by 3pp on DREAMSIM. Although AUTOMATIKZv2 (LLM) outperforms " + }, + { + "bbox": [ + 54, + 638, + 297, + 698 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Z_{\\mathrm{ero}}}" + }, + { + "bbox": [ + 54, + 638, + 297, + 698 + ], + "type": "text", + "content": " by 1.5pp on KID, this" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 259, + 556, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 556, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 556, + 463 + ], + "type": "text", + "content": "indicates in this context that such token-conditioned models (compared to those using patch embeddings) capture the general appearance of scientific figures well but fall short in inferring visual specifics from captions. They do, however, have an edge in reproducing text from captions, which we identify as the primary reason for up to 5pp higher CLIP-Score, as noted in Sec. 6.1. Regarding code similarity, both " + }, + { + "bbox": [ + 313, + 259, + 556, + 463 + ], + "type": "inline_equation", + "content": "\\mathrm{TIKZERO}" + }, + { + "bbox": [ + 313, + 259, + 556, + 463 + ], + "type": "text", + "content": " models considerably outperform others on cBLEU. Interestingly, we observe a mild inverse correlation between cBLEU and TED. Models conditioned solely on tokenized captions tend to generate shorter, often simplified programs [3], potentially resulting in a reduced edit distance to the reference. In terms of efficiency, all models achieve an MTE of 80-85, indicating that only 2 out of 10 inferences require resampling. " + }, + { + "bbox": [ + 313, + 259, + 556, + 463 + ], + "type": "inline_equation", + "content": "\\mathrm{TIKZERO}" + }, + { + "bbox": [ + 313, + 259, + 556, + 463 + ], + "type": "text", + "content": " (Cos) is 3pp more efficient than MSE, while AUTOMATIKZ (13B), likely benefiting from its larger model size, exceeds it by another 3pp." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "type": "text", + "content": "In summary, training " + }, + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ_{v2}}" + }, + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "type": "text", + "content": " on top of a VLM yields worse performance than training based on vanilla " + }, + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{LLAMA}_{3.1}" + }, + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "type": "text", + "content": ", indicating that effective end-to-end training can only leverage the small intersection of graphics programs and images with captions, as illustrated in Fig. 2. However, even without access to this intersection, " + }, + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZERO}" + }, + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "type": "text", + "content": " surpasses both " + }, + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ}" + }, + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "type": "text", + "content": " (13B) and " + }, + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ_{v2}}" + }, + { + "bbox": [ + 313, + 463, + 556, + 584 + ], + "type": "text", + "content": " (LLM) on average by being able to train on images with captions independently of graphics programs. Moreover, using a loss function based on cosine distance proves more effective than using MSE." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 588, + 520, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 520, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 520, + 601 + ], + "type": "text", + "content": "5.2. Combining Adapters with Fine-Tuning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": "In this section, we investigate whether explicitly incorporating the subset of " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{DAT}\\kappa \\mathrm{Z}_{\\mathrm{v3}}" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": " that includes captions into the training process of " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{T}\\mathbf{I}\\kappa \\mathbf{Z}\\mathbf{E}\\mathbf{R}\\mathbf{O}" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": " enhances performance. Our approach involves three incremental stages: (i) We perform a light fine-tuning of " + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{T}\\mathbf{I}\\kappa \\mathbf{Z}\\mathbf{E}\\mathbf{R}\\mathbf{O}" + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": " (Cos) end-to-end on caption-program pairs for one epoch with a low learning rate of 1e-5. Extending the training duration or increasing the learning rate does not yield further performance gains, likely due to the decoder having already reached its saturation point;" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 250, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 250, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 250, + 713 + ], + "type": "text", + "content": "3Belouadi et al. [3] refer to this model as CLiMA (13B)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17797" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 543, + 196 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 543, + 196 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 543, + 196 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 543, + 196 + ], + "type": "table", + "html": "
ModelsOriginal TextRedacted Text
DSIM↑KID↓CLIP↑cBLEU↑TED↓MTE↑AVG↑CLIP↑Ratio↑
QWEN2.5 CODER (32B)54.4735.49324.870.28559.85697.26948.59312.16448.911
GPT-4o56.4642.84431.7870.32758.51197.67579.01913.3241.905
TikZero (Cos)52.8295.10310.0511.60365.5182.29114.6587.22671.893
+ Fine-tuning (i)53.2031.79410.6870.75961.57294.85146.4976.51260.931
+ Separate Captions (ii)52.9832.90515.720.80461.3295.72246.3268.74155.608
+ Weight Resetting (iii)56.2951.83124.1771.98859.00893.05887.04311.47947.478
", + "image_path": "ab5fb95989f476e4f28695960b72b52bf9b287b7a9437d10256381ee6e3512fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 555, + 250 + ], + "type": "text", + "content": "Table 3. System-level scores " + }, + { + "bbox": [ + 55, + 205, + 555, + 250 + ], + "type": "inline_equation", + "content": "\\times {100}" + }, + { + "bbox": [ + 55, + 205, + 555, + 250 + ], + "type": "text", + "content": " for additional baselines and " + }, + { + "bbox": [ + 55, + 205, + 555, + 250 + ], + "type": "inline_equation", + "content": "{\\mathrm{{TikZero}}}" + }, + { + "bbox": [ + 55, + 205, + 555, + 250 + ], + "type": "text", + "content": " combined with fine-tuning and token-conditioning. The scores for TIKZERO (Cos) are replicated from Tab. 2 for convenience. Bold and underlined values denote the best and second-best scores for each metric column, respectively. Cell shading illustrates relative score magnitudes. Arrows indicate metric directionality. Overall, " + }, + { + "bbox": [ + 55, + 205, + 555, + 250 + ], + "type": "inline_equation", + "content": "{\\mathrm{{TikZero}}}_{\\mathrm{O}}\\left( \\mathrm{{Cos}}\\right)" + }, + { + "bbox": [ + 55, + 205, + 555, + 250 + ], + "type": "text", + "content": " with Weight Resetting (iii) demonstrates the strongest average performance across metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 270, + 294, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 294, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 294, + 344 + ], + "type": "text", + "content": "(ii) Alongside feeding captions into the adapter, we provide them separately to the text decoder in tokenized form (cf., Fig. 3); (iii) Prior to fine-tuning, we reset the decoder to its initial weights to overcome saturation, enabling us to fine-tune using the setup described in Sec. 4, which involves 5 epochs and a learning rate of " + }, + { + "bbox": [ + 55, + 270, + 294, + 344 + ], + "type": "inline_equation", + "content": "5\\mathrm{e} - 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 295, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 295, + 421 + ], + "type": "text", + "content": "Baselines In addition to the baselines in Tab. 2, which remain comparable, we also evaluate larger and commercial models that serve as stronger baselines (cf. Supp. Mat. C). Specifically, we assess GPT-4o [8], which has demonstrated strong performance in generating TikZ [3, 28, 30] and QwEN2.5 CODER (32B) [75] as an open-weights model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 701 + ], + "type": "text", + "content": "Results In Tab. 3 (Original Text), all fine-tuning setups of " + }, + { + "bbox": [ + 55, + 426, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZero}" + }, + { + "bbox": [ + 55, + 426, + 295, + 701 + ], + "type": "text", + "content": " show considerable improvement over the base version. Approaches (i) and (ii) each enhance performance by over 30pp on AVG, while approach (iii) surpasses them with an improvement of over 70pp, positioning it as the best-performing model on average, even when compared to our new baselines, with GPT-4o being 8pp lower and " + }, + { + "bbox": [ + 55, + 426, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\mathrm{QWEN}_{2.5}" + }, + { + "bbox": [ + 55, + 426, + 295, + 701 + ], + "type": "text", + "content": " CODER (32B) approximately 40pp lower. Approach (i) demonstrates that direct fine-tuning yields positive effects across nearly all metrics, notably improving MTE by 12pp, TED by 4pp, and KID by 3.5pp. Approach (ii) shows similar trends but, by also incorporating tokenized captions, further improves CLIPSCORE by 5pp, closing the gap to AUTOMATIKZv2 (LLM). Interestingly, both (i) and (ii) slightly decrease performance on cBLEU, potentially due to similar reasons discussed in Sec. 5.1. However, the same cannot be said for (iii), which not only achieves the highest score on cBLEU but also ranks as the second-best on TED, trailing only 0.5pp behind GPT-4o and showcasing that it is possible to perform well on both metrics. Additionally, it increases DREAMSIM by another 3pp and CLIPSCORE by 8.5pp, competing with the much stronger baselines QWEN2.5 CODER (32B) and GPT-4o. In KID, it even surpasses them by 3.5pp and 1pp, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "type": "text", + "content": "In summary, fine-tuning " + }, + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Zero}" + }, + { + "bbox": [ + 67, + 702, + 295, + 713 + ], + "type": "text", + "content": ", especially when com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 270, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 390 + ], + "type": "text", + "content": "bined with a separate caption input and weight resetting, greatly improves performance. This illustrates that the intersection of graphics programs and images with captions, though small, provides a valuable training signal, and best performance can be achieved by making full use of both sets. The best-performing " + }, + { + "bbox": [ + 313, + 270, + 555, + 390 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZero}" + }, + { + "bbox": [ + 313, + 270, + 555, + 390 + ], + "type": "text", + "content": " model even competes with and often surpasses QwEN2.5 CODER (32B) and GPT-4o on several key metrics. Notably, the former model is more than three times larger, and the latter is often estimated at around 1.8 trillion parameters [76], making it 180 times larger." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 406, + 425, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 406, + 425, + 417 + ], + "spans": [ + { + "bbox": [ + 314, + 406, + 425, + 417 + ], + "type": "text", + "content": "5.3. Human Evaluation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "To corroborate our findings from automatic evaluation, we conduct a human annotation campaign focusing on two key properties: caption and image similarity. We employ BestWorst Scaling (BWS) [77], a comparative annotation method that yields high-quality results even with few annotators [78, 79]. We sample 100 instances from our test set and present annotators with " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "-tuples of generated figures, asking them to identify the most and least similar figure to either the reference caption or reference image. This data is then transformed into scores from -1 (poor) to 1 (excellent) by subtracting the proportion of times a figure is selected as the best from the proportion of times it is chosen as the worst [80]. For a manageable workload, we focus on " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " key models: " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZERO}(\\mathrm{Cos})" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": ", our best-performing model from Sec. 5.1; " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZv2}" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " (LLM), its direct end-to-end trained competitor; GPT-4o, our strongest baseline; and " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZERO}(\\mathrm{Cos})" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " fine-tuned using approach (iii) from Sec. 5.2, our best model overall, henceforth referred to as " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZERO}^+" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " for convenience. We engage thirteen annotators and obtain six fully annotated sets per task (cf. Supp. Mat. E for more details). To assess annotator consistency, we calculate the split-half reliability (SHR) [79]. This method randomly divides all annotations into two sets, calculates scores independently, and then determines their correlation using Spearman's " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17798" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 289, + 286 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 289, + 286 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 289, + 286 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 289, + 286 + ], + "type": "image", + "image_path": "5f5c20c6dddaabcd9f5290421346d7a7e02abe869dceed211d4680fe1f6259f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 295, + 298, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 298, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 298, + 352 + ], + "type": "text", + "content": "Figure 4. Bivariate distributions of BWS scores (higher is better) using kernel density estimation for caption and image similarity. Along the diagonal, " + }, + { + "bbox": [ + 55, + 295, + 298, + 352 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Z_{ERO}}" + }, + { + "bbox": [ + 55, + 295, + 298, + 352 + ], + "type": "text", + "content": " (Cos) achieves higher scores than " + }, + { + "bbox": [ + 55, + 295, + 298, + 352 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMAT_{IK}Z_{v2}}" + }, + { + "bbox": [ + 55, + 295, + 298, + 352 + ], + "type": "text", + "content": " (LLM), while " + }, + { + "bbox": [ + 55, + 295, + 298, + 352 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Z_{ERO+}}" + }, + { + "bbox": [ + 55, + 295, + 298, + 352 + ], + "type": "text", + "content": " and GPT-4o demonstrate superior performance compared to both." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": "Results Fig. 4 presents kernel density estimates for the BWS scores, showing generally consistent rankings with automatic evaluations but revealing notable differences in the magnitude of gaps. For caption similarity, the ranking aligns with CLIPSCORE evaluations " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "(\\rho = 1.0)" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZero}" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " (Cos), " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ_{v2}}" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " (LLM), " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZero}+" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": ", and GPT-4o achieving mean scores " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " of -0.25, -0.18, 0.03, and 0.4, respectively. Interestingly, humans perceive a " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " smaller gap between " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ_{v2}}" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " (LLM) and " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZero}" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " (Cos) than suggested by CLIPSCORE values, where " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ_{v2}}" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " (LLM) outperforms " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZero}" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " (Cos) by " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": ". This indicates humans may evaluate caption similarity differently than CLIPSCORE (cf. Sec. 6.1). For image similarity, the system order remains consistent with our DREAMSIM metric " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "(\\rho = 1.0)" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ_{v2}}" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " (LLM), " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZero}" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " (Cos), " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZero}+" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": ", and GPT-4o achieving " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " of -0.26, -0.02, 0.01, and 0.27, respectively. However, the relative gaps between models differ: the separation between " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ_{v2}}" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " (LLM) and " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZero}" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " (Cos), as well as between " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZero}+" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " and GPT-4o, appear more pronounced than observed with DREAMSIM. This discrepancy likely stems from BWS capturing relative preferences rather than absolute performance differences. GPT-4o is selected " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " more often as the best model than " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZero}+" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ_{v2}}" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " (LLM) " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " more often as the worst model than " + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\mathrm{TikZero}" + }, + { + "bbox": [ + 56, + 376, + 296, + 687 + ], + "type": "text", + "content": " (Cos), creating larger perceived gaps even when qualitative differences may be subtle." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "The SHR values of 0.68 for caption similarity and 0.76 for image similarity indicate moderate to strong inter-annotator" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "content": "agreement. We also observe a correlation between these two tasks, with segment-level " + }, + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "inline_equation", + "content": "\\rho = 0.62" + }, + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "content": " and system-level " + }, + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "inline_equation", + "content": "\\rho = 0.8" + }, + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "content": ", suggesting that both evaluation dimensions capture related aspects of model performance. GPT-4o emerges as the best-performing model, aligning with its superior performance on the corresponding automatic metrics, CLIPSCORE and DSIM. Among open-source models, " + }, + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Z_{v2}}" + }, + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "content": " performs best, while AUTOMATIKZv2 (LLM) ranks lowest overall." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 178, + 373, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 373, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 373, + 192 + ], + "type": "text", + "content": "6. Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 198, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 247 + ], + "type": "text", + "content": "We present a comprehensive analysis, investigating the influence of typographic attacks on CLIPSCORE and examining the effectiveness of our architecture in low-resource settings, both in terms of training data and trainable parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 253, + 555, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 266 + ], + "type": "text", + "content": "6.1. CLIPSCORE Limitations & Typographic Attacks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 270, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 556, + 449 + ], + "type": "text", + "content": "A known limitation of CLIPSCORE with text-rich images is its susceptibility to typographic attacks, where scores are disproportionately influenced by string similarity between images and captions [3, 43]. We suspect that token-conditioned models like AUTOMATIKZv2 (LLM) achieve higher CLIPSCORE values than models such as TIKZERO (Cos & MSE) primarily because they tend to visibly copy more substrings from the caption in the output image. To test this hypothesis, we apply the ROT13 substitution cipher [81] to all visible strings in the generated figures and recompute CLIPSCORE. This basic cipher replaces each letter with the 13th letter after it in the Latin alphabet. While not cryptographically secure, the ratio between the original and recomputed CLIPSCORE values should indicate the influence of string matching, i.e., higher ratios suggest less copied text and vice versa." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "text", + "content": "Tabs. 2 & 3 (Redacted Text) present the recomputed CLIP-Score values and ratios for all evaluated models. The results reveal that most " + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Zero}" + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "text", + "content": " models, except for fine-tuning approaches (ii) and (iii), which also condition on tokenized captions, have considerably higher ratios " + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "inline_equation", + "content": "(61\\% -78\\%)" + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "text", + "content": " compared to strictly token-conditioned models " + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "inline_equation", + "content": "(34\\% -51\\%)" + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "text", + "content": ", supporting our hypothesis. AUTOMATIKZ (13B) is an exception, possibly due to its initially low score. Further analysis shows that with redacted text, " + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{AUTOMATIKZ}_{\\mathrm{V2}}" + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "text", + "content": " (LLM)'s CLIPSCORE performance drops to the same level as " + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Zero}" + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "text", + "content": " (Cos & MSE), suggesting that string matching is the primary factor in its superior performance rather than producing better visuals—arguably a more difficult task. Nevertheless, reproducing strings is still somewhat desirable. The human oracle of our test set achieves a ratio of " + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "inline_equation", + "content": "50.8\\%" + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "text", + "content": ", close to the " + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "inline_equation", + "content": "47.5\\%" + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}Zero+}" + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "text", + "content": ", our best-performing model. In contrast, models like GPT-4o, with a lower ratio of " + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "inline_equation", + "content": "41.9\\%" + }, + { + "bbox": [ + 313, + 449, + 557, + 666 + ], + "type": "text", + "content": ", may overfit to caption copying, artificially inflating the CLIPSCORE values." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 672, + 448, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 448, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 448, + 685 + ], + "type": "text", + "content": "6.2. Low-Resource Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "While our adapters train efficiently on large-scale datasets, we investigate whether such extensive data is necessary for opti" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17799" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 272, + 166 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 272, + 166 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 272, + 166 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 272, + 166 + ], + "type": "table", + "html": "
Intv.Training Data
100%50%25%12.5%
192.41177.47849.96756.055
287.55785.24954.94233.817
482.25447.38132.1237.914
876.54540.81629.77416.25
", + "image_path": "bd86657fbf675f2526105b7fa9175380a6d0849e7faf0a66fd8ddbf74c2b5c12.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 175, + 294, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 294, + 231 + ], + "type": "text", + "content": "Table 4. AVG scores for " + }, + { + "bbox": [ + 55, + 175, + 294, + 231 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZero}" + }, + { + "bbox": [ + 55, + 175, + 294, + 231 + ], + "type": "text", + "content": " (Cos) trained on varying fractions of data and intervals of cross-attention layers. Higher scores indicate better performance. Bold and underlined values denote the best and second-best scores for the whole table, respectively. Cell shading illustrates score magnitudes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "text", + "content": "mal performance. Along the same vein, we examine the impact of reducing the amount of cross-attention layers inserted into the vision encoder. We retrain " + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathrm{TIKZERO}" + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "text", + "content": " (Cos) using varying fractions of the training data (" + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "inline_equation", + "content": "d \\in \\{1, \\frac{1}{2}, \\frac{1}{4}, \\frac{1}{8}\\}" + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "text", + "content": ") and insert cross-attention layers at different intervals (" + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "inline_equation", + "content": "i \\in \\{1, 2, 4, 8\\}" + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "text", + "content": "). Table 4 presents the AVG scores from this parameter grid, with detailed scores in Supp. Mat. D. Our findings reveal that utilizing the full dataset and inserting cross-attention at every layer yields the highest average performance, highlighting the benefits of maximizing both variables. Interestingly, the model's performance appears more robust to a reduction in the number of layers compared to a decrease in training data. For instance, training on only " + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\frac{1}{8}" + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "text", + "content": "th of the data leads to a substantial performance drop of " + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "inline_equation", + "content": "36\\mathrm{pp}" + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "text", + "content": ", whereas inserting cross-attention layers every 8 layers (resulting in only 3 cross-attention layers in total) causes a more modest decline of " + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "inline_equation", + "content": "16\\mathrm{pp}" + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "text", + "content": ". Minimizing both variables leads to the most severe drop of over " + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "inline_equation", + "content": "75\\mathrm{pp}" + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "text", + "content": ". These results validate our training setup while suggesting that incorporating additional data might further enhance performance. Given that " + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathrm{ARXIVCAP}" + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "text", + "content": " extracts figures from only " + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "inline_equation", + "content": "572\\mathrm{k}" + }, + { + "bbox": [ + 56, + 251, + 296, + 526 + ], + "type": "text", + "content": " papers, whereas some corpora index over 200 million papers [82], there remains a lot of potential for leveraging larger datasets in future work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 537, + 128, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 128, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 128, + 550 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "text", + "content": "In this work, we demonstrate the potential of " + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZero}" + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "text", + "content": " and its variants for generating TiKZ graphics programs from captions. Notably, " + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZero}" + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "text", + "content": " does not require aligned caption-program pairs in its original formulation but instead aligns representation spaces of unaligned graphics programs and captioned images. This enables our model to leverage substantially more training data compared to end-to-end trained models that operate solely on caption-image pairs (cf. Fig. 2) while maintaining training efficiency. " + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZero}" + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "text", + "content": " outperforms strong end-to-end trained baselines, including our independently trained AUTOMATiKZv2 models, which use the same data pool, excluding instances they cannot process, illustrating the strengths of our approach. When extending the " + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZero}" + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "text", + "content": " ap" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "proach with additional end-to-end training, it also compares favorably to much larger baselines and commercial systems like GPT-4o. While this enhanced approach, " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}ZERO^{+}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", is no longer zero-shot by our definition, it remains a " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\mathrm{T_{IK}ZERO}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " model in the sense that it operates on both sets of graphics programs and captioned images, with the added advantage of explicitly utilizing their intersection (cf. Fig. 2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 157, + 556, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 556, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 556, + 240 + ], + "type": "text", + "content": "These results demonstrate the benefits of designing architectures around available data and validate the approach of decoupling graphics program generation from text understanding (with optional later reconciliation through " + }, + { + "bbox": [ + 313, + 157, + 556, + 240 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZero^{+}}" + }, + { + "bbox": [ + 313, + 157, + 556, + 240 + ], + "type": "text", + "content": "). Although we demonstrate our method specifically on " + }, + { + "bbox": [ + 313, + 157, + 556, + 240 + ], + "type": "inline_equation", + "content": "\\mathrm{TiKZ}" + }, + { + "bbox": [ + 313, + 157, + 556, + 240 + ], + "type": "text", + "content": ", we believe its general principles will inspire future work on related graphics program synthesis tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 247, + 556, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 556, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 556, + 415 + ], + "type": "text", + "content": "Future Work Beyond scaling up our training data to explore convergence limits (cf. Sec. 6.2), we plan to investigate automatic methods for improving the quality and alignment of caption-image or caption-program pairs. This includes rewriting potentially noisy captions with LLMs and enhancing them with the visual understanding capabilities of VLMs [83-85]. We believe our approach to aligning textual and image modalities enables other promising applications for graphics program synthesis, such as editing images in latent space via textual instructions to generate modified graphics programs. Additionally, we intend to explore alternative alignment strategies beyond model distillation, including contrastive learning [86], which has successfully aligned modalities in discriminative models [6, 87, 88]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 426, + 378, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 378, + 439 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 378, + 439 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 447, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 579 + ], + "type": "text", + "content": "Our evaluations include proprietary systems that operate as black boxes; their training data is unknown, and they offer no guarantees of consistent performance over time. This (i) makes addressing data leakage and cross-contamination impossible and (ii) limits the fairness and reproducibility of our experiments. Nevertheless, even under these unfavorable conditions, our open models remain competitive. Users should be aware, however, that our models may behave unpredictably, and outputs might differ from expectations. Additionally, our models do not include safeguards against potential misuse, e.g., for generating fake scientific content." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 580, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 556, + 675 + ], + "type": "text", + "content": "Regarding licensing of our training data, a large portion of the TikZ programs in " + }, + { + "bbox": [ + 313, + 580, + 556, + 675 + ], + "type": "inline_equation", + "content": "\\mathrm{DAT}_{\\mathrm{IK}}\\mathrm{Z}_{\\mathrm{v3}}" + }, + { + "bbox": [ + 313, + 580, + 556, + 675 + ], + "type": "text", + "content": " are licensed under permissive terms that allow redistribution. The remaining programs are distributed under the arXiv.org perpetual, non-exclusive license, which prohibits redistribution, which is why we exclude them from the public release of " + }, + { + "bbox": [ + 313, + 580, + 556, + 675 + ], + "type": "inline_equation", + "content": "\\mathrm{DAT}_{\\mathrm{IK}}\\mathrm{Z}_{\\mathrm{v3}}" + }, + { + "bbox": [ + 313, + 580, + 556, + 675 + ], + "type": "text", + "content": ". However, since we release our dataset creation scripts, we encourage others to reproduce the full version independently." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": "4https://creativecommons.org/licenses; https://opensource.org/licenses/mit; https://www.gnu.org/licenses/fdl-1.3.en.html; https://openai.com/policies/terms-of-use" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17800" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 151, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 151, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 151, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 301 + ], + "type": "text", + "content": "We extend our sincere gratitude to the following individuals (in no particular order) for their valuable contributions: Christian Greisinger, Hour Kaing, Ran Zhang, Tejaswini Medi, Yanran Chen, Sotaro Takeshita, Katharina Prasse, JiWoo Kim, Christoph Leiter, Haiyue Song, and Aida Kostikova. Their assistance with our human evaluation campaign, proofreading, insightful discussions, and constructive feedback has been instrumental to our work. The first author conducted part of this research during an internship at the National Institute of Information and Communications Technology (NICT), Japan. The second to last author is supported by the Federal Ministry of Education and Research (BMBF) via the research grant METRICS4NLG and the German Research Foundation (DFG) via the Heisenberg Grant EG 375/5-1. We acknowledge computing resources provided by the state of Baden-Württemberg through bwHPC and the German Research Foundation (DFG) through grant INST 35/1597-1 FUGG. Finally, we thank the OpenMoji project for the open-source icons used throughout this work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 322, + 114, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 114, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 114, + 335 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 342, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 342, + 258, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 342, + 258, + 353 + ], + "spans": [ + { + "bbox": [ + 61, + 342, + 258, + 353 + ], + "type": "text", + "content": "[1] Till Tantau. The TikZ and PGF Packages, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 354, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 354, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 61, + 354, + 296, + 408 + ], + "type": "text", + "content": "[2] Jonas Belouadi, Simone Paolo Ponzetto, and Steffen Eger. DeTikZify: Synthesizing graphics programs for scientific figures and sketches with TikZ. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 4, 15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 411, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 296, + 453 + ], + "type": "text", + "content": "[3] Jonas Belouadi, Anne Lauscher, and Steffen Eger. AutomataTikZ: Text-guided synthesis of scientific vector graphics with TikZ. In The Twelfth International Conference on Learning Representations, 2024. 1, 4, 5, 6, 7, 16, 18" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 456, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 296, + 510 + ], + "type": "text", + "content": "[4] Leixin Zhang, Yinjie Cheng, Weihe Zhai, Steffen Eger, Jonas Belouadi, Fahimeh Moafian, and Zhixue Zhao. ScImage: How good are multimodal large language models at scientific text-to-image generation? In The Thirteenth International Conference on Learning Representations, 2025. 1, 15" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "type": "text", + "content": "[5] Abhay Zala, Han Lin, Jaemin Cho, and Mohit Bansal. DiagrammerGPT: Generating open-domain, open-platform diagrams via LLM planning. In First Conference on Language Modeling, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 557, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 632 + ], + "type": "text", + "content": "[6] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 635, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 677 + ], + "type": "text", + "content": "[7] Mark Palatucci, Dean Pomerleau, Geoffrey E Hinton, and Tom M Mitchell. Zero-shot learning with semantic output codes. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2009. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 680, + 246, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 246, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 246, + 690 + ], + "type": "text", + "content": "[8] OpenAI. GPT-4 technical report, 2023. 2, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "type": "text", + "content": "[9] Emilio Parisotto, Abdel rahman Mohamed, Rishabh Singh, Lihong Li, Dengyong Zhou, and Pushmeet Kohli. Neuro" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "text", + "content": "symbolic program synthesis. In International Conference on Learning Representations, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 96, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 555, + 149 + ], + "type": "text", + "content": "[10] Jacob Devlin, Jonathan Uesato, Surya Bhupatiraju, Rishabh Singh, Abdel rahman Mohamed, and Pushmeet Kohli. Robust-Fill: Neural program learning under noisy I/O. In Proceedings of the 34th International Conference on Machine Learning, pages 990-998. PMLR, 2017." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 151, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 555, + 239 + ], + "type": "text", + "content": "[11] Kevin Ellis, Catherine Wong, Maxwell Nye, Mathias Sable-Meyer, Lucas Morales, Luke Hewitt, Luc Cary, Armando Solar-Lezama, and Joshua B. Tenenbaum. DreamCoder: bootstrapping inductive program synthesis with wake-sleep library learning. In Proceedings of the 42nd ACM SIGPLAN International Conference on Programming Language Design and Implementation, page 835-850, New York, NY, USA, 2021. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 239, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 554, + 293 + ], + "type": "text", + "content": "[12] Yaroslav Ganin, Tejas Kulkarni, Igor Babuschkin, S. M. Ali Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. In Proceedings of the 35th International Conference on Machine Learning, pages 1666-1675. PMLR, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 294, + 554, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 554, + 338 + ], + "type": "text", + "content": "[13] Kevin Ellis, Daniel Ritchie, Armando Solar-Lezama, and Josh Tenenbaum. Learning to infer graphics programs from hand-drawn images. In Thirty-second Conference on Neural Information Processing Systems, pages 6062–6071, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 338, + 555, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 555, + 391 + ], + "type": "text", + "content": "[14] Kevin Ellis, Maxwell Nye, Yewen Pu, Felix Sosa, Josh Tenenbaum, and Armando Solar-Lezama. Write, execute, assess: Program synthesis with a REPL. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 393, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 393, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 393, + 554, + 437 + ], + "type": "text", + "content": "[15] Peter Kulits, Haiwen Feng, Weiyang Liu, Victoria Fernandez Abrevaya, and Michael J. Black. Re-thinking inverse graphics with large language models. Transactions on Machine Learning Research, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 438, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 554, + 470 + ], + "type": "text", + "content": "[16] Wen-Ding Li and Kevin Ellis. Is programming by example solved by LLMs? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 472, + 555, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 555, + 513 + ], + "type": "text", + "content": "[17] Shreyas Kapur, Erik Jenner, and Stuart Russell. Diffusion on syntax trees for program synthesis. In The Thirteenth International Conference on Learning Representations, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 515, + 555, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 555, + 581 + ], + "type": "text", + "content": "[18] Gopal Sharma, Rishabh Goyal, Difan Liu, Evangelos Kalogerakis, and Subhransu Maji. CSGNet: Neural shape parser for constructive solid geometry. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 5515-5523. Computer Vision Foundation / IEEE Computer Society, 2018." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 582, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 555, + 624 + ], + "type": "text", + "content": "[19] Yonglong Tian, Andrew Luo, Xingyuan Sun, Kevin Ellis, William T. Freeman, Joshua B. Tenenbaum, and Jiajun Wu. Learning to infer and execute 3D shape programs. In International Conference on Learning Representations, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 625, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 554, + 669 + ], + "type": "text", + "content": "[20] Javier Camara, Javier Troya, Lola Burgueño, and Antonio Vallecillo. On the assessment of generative AI in modeling tasks: an experience report with chatgpt and UML. Softw. Syst. Model., 22(3):781-793, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[21] Hugo Laurençon, Leo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17801" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[22] Hugo Laurençon, Andrés Marafioti, Victor Sanh, and Léo Tronchon. Building and better understanding vision-language models: insights and future directions, 2024. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 183 + ], + "type": "text", + "content": "[23] Shengbang Tong, Ellis L Brown II, Penghao Wu, Sanghyun Woo, Adithya Jairam Iyer, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, Xichen Pan, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian1: A fully open, vision-centric exploration of multimodal LLMs. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 249 + ], + "type": "text", + "content": "[24] Juan A. Rodriguez, Abhay Puri, Shubham Agarwal, Issam H. Laradji, Pau Rodriguez, Sai Rajeswar, David Vazquez, Christopher Pal, and Marco Pedersoli. StarVector: Generating scalable vector graphics code from images and text. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 16175-16186, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 295, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 295, + 347 + ], + "type": "text", + "content": "[25] Haotian Zhang, Mingfei Gao, Zhe Gan, Philipp Dufter, Nina Wenzel, Forrest Huang, Dhruti Shah, Xianzhi Du, Bowen Zhang, Yanghao Li, Sam Dodge, Keen You, Zhen Yang, Aleksei Timofeev, Mingze Xu, Hong-You Chen, Jean-Philippe Fauconnier, Zhengfeng Lai, Haoxuan You, Zirui Wang, Afshin Dehghan, Peter Grasch, and Yinfei Yang. MM1.5: Methods, analysis & insights from multimodal LLM fine-tuning. In The Thirteenth International Conference on Learning Representations, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 349, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 295, + 426 + ], + "type": "text", + "content": "[26] Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal ArXiv: A dataset for improving scientific comprehension of large vision-language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 14369-14387, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 426, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 502 + ], + "type": "text", + "content": "[27] Jaeyoung Kim, Jongho Lee, Hong-Jun Choi, Ting-Yao Hsu, Chieh-Yang Huang, Sungchul Kim, Ryan Rossi, Tong Yu, Clyde Lee Giles, Ting-Hao 'Kenneth' Huang, and Sungchul Choi. Multi-LLM collaborative caption generation in scientific documents. In AI for Research and Scalable, Efficient Systems, pages 142–160, Singapore, 2025. Springer Nature Singapore. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 503, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 295, + 568 + ], + "type": "text", + "content": "[28] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, Harsha Nori, Hamid Palangi, Marco Tulio Ribeiro, and Yi Zhang. Sparks of artificial general intelligence: Early experiments with GPT-4, 2023. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 570, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 635 + ], + "type": "text", + "content": "[29] Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad, Stephanie Fu, Adrian Rodriguez-Munoz, Shivam Duggal, Phillip Isola, and Antonio Torralba. A vision check-up for language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14410-14419, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 636, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 667 + ], + "type": "text", + "content": "[30] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with GPT-4, 2023. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "text", + "content": "[31] Lingjiao Chen, Matei Zaharia, and James Zou. How Is ChatGPT's Behavior Changing Over Time? Harvard Data Science Review, 6(2), 2024. https://hdsr.mitpress.mit.edu/pub/y95zitmz.2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[32] Sagi Polaczek, Yuval Alaluf, Elad Richardson, Yael Vinker, and Daniel Cohen-Or. NeuralSVG: An implicit representation for text-to-vector generation, 2025. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "type": "text", + "content": "[33] Ronghuan Wu, Wanchao Su, Kede Ma, and Jing Liao. IconShop: Text-guided vector icon synthesis with autoregressive transformers. ACM Trans. Graph., 42(6), 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "type": "text", + "content": "[34] Ajay Jain, Amber Xie, and Pieter Abbeel. VectorFusion: Text-to-SVG by abstracting pixel-based diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1911-1920, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 184, + 554, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 554, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 554, + 216 + ], + "type": "text", + "content": "[35] Kevin Frans, Lisa B. Soros, and Olaf Witkowski. CLIPDraw: Exploring text-to-drawing synthesis through language-image encoders. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 217, + 554, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 282 + ], + "type": "text", + "content": "[36] Henrik Voigt, Kai Lawonn, and Sina Zarrieß. Plots made quickly: An efficient approach for generating visualizations from natural language queries. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 12787-12793, Torino, Italia, 2024. ELRA and ICCL. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 284, + 554, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 554, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 554, + 349 + ], + "type": "text", + "content": "[37] Yuyu Luo, Nan Tang, Guoliang Li, Chengliang Chai, Wenbo Li, and Xuedi Qin. Synthesizing natural language to visualization (nl2vis) benchmarks from NL2SQL benchmarks. In Proceedings of the 2021 International Conference on Management of Data, page 1235-1247, New York, NY, USA, 2021. Association for Computing Machinery." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 350, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 381 + ], + "type": "text", + "content": "[38] Jock Mackinlay. Automating the design of graphical presentations of relational information. ACM Trans. Graph., 5(2): 110-141, 1986." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 383, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 437 + ], + "type": "text", + "content": "[39] Steven F. Roth, John Kolojejchick, Joe Mattis, and Jade Goldstein. Interactive graphic design using automatic presentation knowledge. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, page 112-117, New York, NY, USA, 1994. Association for Computing Machinery." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 438, + 554, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 554, + 482 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 554, + 482 + ], + "type": "text", + "content": "[40] Yang Wu, Yao Wan, Hongyu Zhang, Yulei Sui, Wucai Wei, Wei Zhao, Guandong Xu, and Hai Jin. Automated data visualization from natural language via large language models: An exploratory study. Proc. ACM Manag. Data, 2(3), 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 483, + 554, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 483, + 554, + 546 + ], + "spans": [ + { + "bbox": [ + 316, + 483, + 554, + 546 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 548, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 602 + ], + "type": "text", + "content": "[42] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 635 + ], + "type": "text", + "content": "[43] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents, 2022. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 636, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 554, + 713 + ], + "type": "text", + "content": "[44] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, and Jie Tang. CogView: Mastering text-to-image generation via transformers. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 19822-19835, 2021." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17802" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "text", + "content": "[45] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. CogView2: Faster and better text-to-image generation via hierarchical transformers. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 297, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 297, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 297, + 163 + ], + "type": "text", + "content": "[46] Juan A. Rodriguez, David Vázquez, Issam H. Laradji, Marco Pedersoli, and Pau Rodríguez. FigGen: Text to scientific figure generation. In The First Tiny Papers Track at ICLR 2023, Tiny Papers @ ICLR 2023, Kigali, Rwanda, May 5, 2023. OpenReview.net, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 296, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 296, + 219 + ], + "type": "text", + "content": "[47] Juan A. Rodriguez, David Vázquez, Issam H. Laradji, Marco Pedersoli, and Pau Rodríguez. OCR-VQGAN: Taming textwithin-image generation. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2023, Waikoloa, HI, USA, January 2-7, 2023, pages 3678-3687. IEEE, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 220, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 296, + 262 + ], + "type": "text", + "content": "[48] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139-144, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 265, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 265, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 57, + 265, + 296, + 308 + ], + "type": "text", + "content": "[49] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12873-12883, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 310, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 310, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 57, + 310, + 296, + 365 + ], + "type": "text", + "content": "[50] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning, pages 2256-2265, Lille, France, 2015. PMLR. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 366, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 366, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 57, + 366, + 296, + 399 + ], + "type": "text", + "content": "[51] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 401, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 401, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 401, + 296, + 456 + ], + "type": "text", + "content": "[52] Yiwei Hu, Paul Guerrero, Milos Hasan, Holly Rushmeier, and Valentin Deschaintre. Generating procedural materials from text or image prompts. In ACM SIGGRAPH 2023 Conference Proceedings, New York, NY, USA, 2023. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 457, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 457, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 57, + 457, + 296, + 499 + ], + "type": "text", + "content": "[53] Shukang Yin, Chaoyou Fu, Sirui Zhao, Ke Li, Xing Sun, Tong Xu, and Enhong Chen. A survey on multimodal large language models. National Science Review, 11(12):nwae403, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 502, + 296, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 296, + 546 + ], + "type": "text", + "content": "[54] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 296, + 601 + ], + "type": "text", + "content": "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 604, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 604, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 604, + 296, + 713 + ], + "type": "text", + "content": "[56] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 711 + ], + "type": "text", + "content": "[57] Aaron Grattaftiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava Spataru, Baptiste Roziere, Bethany Biron, Binh Tang, Bobbie Chern, Charlotte Caucheteux, Chaya Nayak, Chloe Bi, Chris Marra, Chris McConnell, Christian Keller, Christophe Touret, Chunyang Wu, Corinne Wong, Cristian Canton Ferrer, Cyrus Nikolaidis, Damien Allonsius, Daniel Song, Danielle Pintz, Danny Livshits, Danny Wyatt, David Esiobu, Dhruv Choudhary, Dhruv Mahajan, Diego Garcia-Olano, Diego Perino, Dieuwke Hupkes, Egor Lakomkin, Ehab AlBadawy, Elina Lobanova, Emily Dinan, Eric Michael Smith, Filip Radenovic, Francisco Guzmán, Frank Zhang, Gabriel Synnaeve, Gabrielle Lee, Georgia Lewis Anderson, Govind Thattai, Graeme Nail, Gregoire Mialon, Guan Pang, Guillem Cucurell, Hailey Nguyen, Hannah Korevaar, Hu Xu, Hugo Touvron, Iliyan Zarov, Imanol Arrieta Ibarra, Isabel Kloumann, Ishan Misra, Ivan Evtimov, Jack Zhang, Jade Copet, Jaewon Lee, Jan Geffert, Jana Vranes, Jason Park, Jay Mahadeokar, Jeet Shah, Jelmer van der Linde, Jennifer Billock, Jenny Hong, Jenya Lee, Jeremy Fu, Jianfeng Chi, Jianyu Huang, Jiawen Liu, Jie Wang, Jiecao Yu, Joanna Bitton, Joe Spisak, Jongsoo Park, Joseph Rocca, Joshua Johnstun, Joshua Saxe, Junteng Jia, Kalyan Vasuden Alwala, Karthik Prasad, Kartikeya Upasani, Kate Plawiak, Ke Li, Kenneth Heafield, Kevin Stone, Khalid El-Arini, Krithika Iyer, Kshitiz Malik, Kuenley Chiu, Kunal Bhalla, Kushal Lakhotia, Lauren Rantala-Yeary, Laurens van der Maaten, Lawrence Chen, Liang Tan, Liz Jenkins, Louis Martin, Lovish Madaan, Lubo Malo, Lukas Blecher, Lukas Landzaat, Luke de Oliveira, Madeline Muzzi, Mahesh Pasupuleti, Mannat Singh, Manohar Paluri, Marcin Kardas, Maria Tsimpoukelli, Mathew Oldham, Mathieu Rita, Maya Pavlova, Melanie Kambadur, Mike Lewis, Min Si, Mitesh Kumar Singh, Mona Hassan Naman Goyal Narjes Torabi Nikolay Bashlykov Nikolay Bogoychev,Niladri Chatterji Ning Zhang,Olivier Duchenne Onur Celebi Patrick Alrassy Pengchuan Zhang,Pengwei Li Petar Vasic Peter WengPrajjwal BhargavaPratik Dubal Praveen Krishnan Punit Singh Koura Puxin Xu Qing He Qingxiao Dong Ragavan Srinivasan Raj Ganapathy Ramon Calderer Ricardo Silveira Cabral Roberta Raileanu Rohan Maheswari Rohit Girdhar Rohit Patel Romain SauvestreRonnie Polidoro Roshan Sumbaly Ross Taylor Ruan Silva Rui Hou Rui Wang Saghar Hosseini Sahana Chennabasappa Sanjay Singh Sean Bell Seohyun Sonia Kim,Sergey Edunov Shaoliang Nie Sharan Narang Sharath Rarpathy Sheng Shen Shengye Wan Shruti BhosaleShun ZhangSimon Vandenhende Soumya Batra Spencer WhitmanSten Sootla Stephane Collot Suchin Gururangan Sydney Borodinsky Tamar Herman Tara Fowler Tarek Sheasha Thomas Georgiou Thomas Scialom Tobias Speckbacher Todor Mihaylov Tong Xiao Ujjwal Karn,Vedanuj Goswami Vibhor Gupta,Vignesh Ramanathan,Viktor Kerkez Vincent Gonguet,Virginie Do,Vish VogetiVitor Albiero,Vladan Petrovic Weiwei Chu Wenhan Xiong,Wenyin Fu Whitney" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17803" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 75, + 72, + 297, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 297, + 711 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 297, + 711 + ], + "type": "text", + "content": "Meers, Xavier Martinet, Xiaodong Wang, Xiaofang Wang, Xiaoqing Ellen Tan, Xide Xia, Xinfeng Xie, Xuchao Jia, Xuewei Wang, Yaelle Goldschlag, Yashesh Gaur, Yasmine Babaei, Yi Wen, Yiwen Song, Yuchen Zhang, Yue Li, Yuning Mao, Zacharie Delpierre Coudert, Zheng Yan, Zhengxing Chen, Zoe Papakipos, Aaditya Singh, Aayushi Srivastava, Abha Jain, Adam Kelsey, Adam Shajnfeld, Adithya Gangidi, Adolfo Victoria, Ahuva Goldstand, Ajay Menon, Ajay Sharma, Alex Boesenberg, Alexei Baevski, Allie Feinstein, Amanda Kallet, Amit Sangani, Amos Teo, Anam Yunus, Andrei Lupu, Andres Alvarado, Andrew Caples, Andrew Gu, Andrew Ho, Andrew Poulton, Andrew Ryan, Ankit Ramchandani, Annie Dong, Annie Franco, Anuj Goyal, Aparajita Saraf, Arkabandhu Chowdhury, Ashley Gabriel, Ashwin Bharambe, Assaf Eisenman, Azadeh Yazdan, Beau James, Ben Maurer, Benjamin Leonhardi, Bernie Huang, Beth Loyd, Beto De Paola, Bhargavi Paranjape, Bing Liu, Bo Wu, Boyu Ni, Braden Hancock, Bram Wasti, Brandon Spence, Brani Stojkovic, Brian Gamido, Britt Montalvo, Carl Parker, Carly Burton, Catalina Mejia, Ce Liu, Changhan Wang, Changkyu Kim, Chao Zhou, Chester Hu, Ching-Hsiang Chu, Chris Cai, Chris Tindal, Christoph Feichtenhofer, Cynthia Gao, Damon Civin, Dana Beaty, Daniel Kreymer, Daniel Li, David Adkins, David Xu, Davide Testuggine, Delia David, Devi Parikh, Diana Liskovich, Didem Foss, Dingkang Wang, Duc Le, Dustin Holland, Edward Dowling, Eissa Jamil, Elaine Montgomery. Eleonora Presani, Emily Hahn, Emily Wood, Eric-Tuan Le. Erik Brinkman Esteban Arcaute Evan Dunbar Evan Smothers. Fei Sun Felix Kreuk Feng Tian Filippos Kokkinos. Firat Ozgenel Francesco Caggioni Frank Kanayet Frank Seide,Gabriela Medina Florez,Gabriella Schwarz,Gada Badeer Georgia Swee Gil Halpern Grant Herman Grigory SizovGuangyi,ZhangGuna LakshminarayananHakan Inan Hamid Shojanazeri Han Zou Hannah Wang Hanwen Zha Haroun Habeeb Harrison Rudolph Helen Suk Henry Aspegren Hunter Goldman Hongyuan Zhan Ibrahim Damlaj Igor Molybog,Igor Tufanov Ilias Leontiadis Irina-Elena Veliche Itai Gat Jake WeissmanJames GeboskiJames Kohli Janice LamJaphet AsherJean-Baptiste GayaJeff MarcusJeff TangJennifer Chan Jenny Zhen Jeremy Reizenstein Jeremy Teboul Jessica Zhong Jian Jin Jingyi Yang Joe Cummings Jon Carvill Jon Shepard Jonathan McPhie Jonathan Torres Josh Ginsburg Junjie Wang Kai Wu Kam Hou U Karan Saxena Kartikay Khandelwal Katayoun Zand Kathy Matosich Kaushik Veeraraghavan Kelly Michelena Keqian Li Kiran Jagadeesh Kun HuangKunal ChawlaKyle HuangLailin Chen Lakshya Garg Lavender A Leandro Silva Lee Bell Lei Zhang Liangpeng Guo Licheng Yu Liron Moshkovich Luca Wehrstedt Madian Khabsa Manav Avalani Manish Bhatt Martynas Mankus Matan Hasson Matthew Lennie Matthias Reso Maxim Groshev Maxim Naumov Maya Lathi. Meghan Keneally Miao Liu Michael L. Seltzer Michal Valko Michelle Restrepo Mihir Patel Mik Vyatskov Mikayel Samvelyan Mike Clark Mike Macey Mike Wang Miquel Jubert Hermoso Mo Metanat Mohammad Rastegari Munish Bansal Nandhini Santhanam Natascha Parks Natasha White Navyata Bawa Nayan Singhal Nick Egebo Nicolas Usunier Nikhil Mehta Nikolay Pavlovich Laptev Ning Dong Nor" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 334, + 72, + 556, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 72, + 556, + 422 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 556, + 422 + ], + "type": "text", + "content": "man Cheng, Oleg Chernoguz, Olivia Hart, Omkar Salpekar, Ozlem Kalinli, Parkin Kent, Parth Parekh, Paul Saab, Pavan Balaji, Pedro Rittner, Philip Bontrager, Pierre Roux, Piotr Dollar, Polina Zvyagina, Prashant Ratanchandani, Pritish Yuvraj, Qian Liang, Rachad Alao, Rachel Rodriguez, Rafi Ayub, Ragtoham Murthy, Raghu Nayani, Rahul Mitra, Rangaprabhu Parthasarathy, Raymond Li, Rebekkah Hogan, Robin Battey, Rocky Wang, Russ Howes, Rudy Rinott, Sachin Mehta, Sachin Siby, Sai Jayesh Bondu, Samyak Datta, Sara Chugh, Sara Hunt, Sargun Dhillon, Sasha Sidorov, Satadru Pan, Saurabh Mahajan, Saurabh Verma, Seiji Yamamoto, Sharadh Ramaswamy, Shaun Lindsay, Shaun Lindsay, Sheng Feng, Shenghao Lin, Shengxin Cindy Zha, Shishir Patil, Shiva Shankar, Shuqiang Zhang, Shuqiang Zhang, Sinong Wang, Sneha Agarwal, Soji Sajuyigbe, Soumith Chintala, Stephanie Max, Stephen Chen, Steve Kehoe, Steve Satterfield, Sudarshan Govindaprasad, Sumit Gupta, Summer Deng, Sungmin Cho, Sunny Virk, Suraj Subramanian, Sy Choudhury, Sydney Goldman, Tal Remez, Tamar Glaser, Tamara Best, Thilo Koehler, Thomas Robinson, Tianhe Li, Tianjun Zhang, Tim Matthews, Timothy Chou, Tzook Shaked, Varun Vontimitta, Victoria Ajayi, Victoria Montanez, Vijai Mohan, Vinay Satish Kumar, Vishal Mangla, Vlad Ionescu, Vlad Poenaru, Vlad Tiberiu Mihailescu, Vladimir Ivanov, Wei Li, Wenchen Wang, Wenwen Jiang, Wes Bouaziz, Will Constable, Xiaocheng Tang, Xiaojian Wu, Xiaolan Wang, Xilun Wu, Xinbo Gao, Yaniv Kleinman, Yanjun Chen, Ye Hu, Ye Jia, Ye Qi, Yenda Li, Yilin Zhang, Ying Zhang, Yossi Adi, Youngjin Nam, Yu, Wang, Yu Zhao, Yuchen Hao, Yundi Qian, Yunlu Li, Yuzi He, Zach Rait. Zachary DeVito, Zef Rosnbrick. Zhaoduo Wen. Zhenyu Yang. Zhiwei Zhao and Zhiyu Ma. The llama 3 herd of models. 2024. 3. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 315, + 425, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 315, + 425, + 553, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 425, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 315, + 425, + 553, + 445 + ], + "type": "text", + "content": "[58] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735-1780, 1997. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 315, + 447, + 553, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 447, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 315, + 447, + 553, + 468 + ], + "type": "text", + "content": "[59] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 315, + 470, + 555, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 470, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 315, + 470, + 555, + 523 + ], + "type": "text", + "content": "[60] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Han Yang, Boyu Diao, and Yongjun Xu. CLIP-KD: An empirical study of CLIP model distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15952-15962, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 525, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 557 + ], + "type": "text", + "content": "[61] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 559, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 623 + ], + "type": "text", + "content": "[62] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and efficient foundation language models, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 625, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 555, + 668 + ], + "type": "text", + "content": "[63] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 11975-11986, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[64] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, Ji Ma, Jiaqi Wang, Xiaoyi Dong, Hang Yan, Hewei Guo, Conghui He, Botian Shi, Zhenjiang Jin, Chao Xu, Bin" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17804" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 139 + ], + "type": "text", + "content": "Wang, Xingjian Wei, Wei Li, Wenjian Zhang, Bo Zhang, Pinlong Cai, Licheng Wen, Xiangchao Yan, Min Dou, Lewei Lu, Xizhou Zhu, Tong Lu, Dahua Lin, Yu Qiao, Jifeng Dai, and Wenhai Wang. How far are we to GPT-4V? closing the gap to commercial multimodal models with open-source suites. Science China Information Sciences, 67(12):220101, 2024. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 259 + ], + "type": "text", + "content": "[65] Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, Thomas Unterthiner, Daniel Keysers, Skanda Koppula, Fangyu Liu, Adam Grycner, Alexey Gritsenko, Neil Houlsby, Manoj Kumar, Keran Rong, Julian Eisenschlos, Rishabh Kabra, Matthias Bauer, Matko Bošnjak, Xi Chen, Matthias Minderer, Paul Voigtlaender, Ioana Bica, Ivana Balazevic, Joan Puigcerver, Pinelopi Papalampidi, Olivier Henaff, Xi Xiong, Radu Soricut, Jeremiah Harmsen, and Xiaohua Zhai. PaliGemma: A versatile 3B VLM for transfer, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 260, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 260, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 260, + 296, + 304 + ], + "type": "text", + "content": "[66] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proceedings of the IEEE, 109(1):43-76, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 304, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 296, + 360 + ], + "type": "text", + "content": "[67] Stephanie Fu, Netanel Yakir Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. DreamSim: Learning new dimensions of human visual similarity using synthetic data. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 4, 18" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 360, + 296, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 296, + 425 + ], + "type": "text", + "content": "[68] Shobhita Sundaram, Stephanie Fu, Lukas Muttenthaler, Netei Yakir Tamir, Lucy Chai, Simon Kornblith, Trevor Darrell, and Phillip Isola. When does perceptual alignment benefit vision representations? In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 4, 18" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 426, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 459 + ], + "type": "text", + "content": "[69] Mikołaj Binkowski, Dougal J. Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. In International Conference on Learning Representations, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 459, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 296, + 525 + ], + "type": "text", + "content": "[70] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 525, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 296, + 581 + ], + "type": "text", + "content": "[71] Aryaz Eghbali and Michael Pradel. CrystalBLEU: Precisely and efficiently measuring the similarity of code. In Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineering, New York, NY, USA, 2023. Association for Computing Machinery. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 581, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 296, + 647 + ], + "type": "text", + "content": "[72] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "type": "text", + "content": "[73] Peter Stanchev, Weiyue Wang, and Hermann Ney. EED: Extended edit distance measure for machine translation. In Proceedings of the Fourth Conference on Machine Translation (Volume 2: Shared Task Papers, Day 1), pages 514-520, Florence, Italy, 2019. Association for Computational Linguistics. 4" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 149 + ], + "type": "text", + "content": "[74] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A. Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, Demis Hassabis, Claudia Clopath, Dharshan Kumaran, and Raia Hadsell. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114(13):3521-3526, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 150, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 150, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 315, + 150, + 555, + 217 + ], + "type": "text", + "content": "[75] Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, Kai Dang, Yang Fan, Yichang Zhang, An Yang, Rui Men, Fei Huang, Bo Zheng, Yibo Miao, Shanghaoran Quan, Yunlong Feng, Xingzhang Ren, Xuancheng Ren, Jingren Zhou, and Junyang Lin. Qwen2.5-Coder technical report, 2024. 6, 15" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 218, + 555, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 218, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 315, + 218, + 555, + 284 + ], + "type": "text", + "content": "[76] Michael Saxon, Fatima Jahara, Mahsa Khoshnoodi, Yujie Lu, Aditya Sharma, and William Yang Wang. Who evaluates the evaluations? objectively scoring text-to-image prompt coherence metrics with t2IScore score (TS2). In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 285, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 285, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 315, + 285, + 555, + 319 + ], + "type": "text", + "content": "[77] Jordan J. Louviere, Terry N. Flynn, and A. A. J. Marley. Best—Worst Scaling: Theory, Methods and Applications. Cambridge University Press, 2015. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 319, + 555, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 319, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 315, + 319, + 555, + 397 + ], + "type": "text", + "content": "[78] Svetlana Kiritchenko and Saif M. Mohammad. Capturing reliable fine-grained sentiment associations by crowdsourcing and best-worst scaling. In Proceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 811-817, San Diego, California, 2016. Association for Computational Linguistics. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 397, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 397, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 315, + 397, + 555, + 464 + ], + "type": "text", + "content": "[79] Svetlana Kiritchenko and Saif Mohammad. Best-worst scaling more reliable than rating scales: A case study on sentiment intensity annotation. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 465-470, Vancouver, Canada, 2017. Association for Computational Linguistics. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 465, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 465, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 315, + 465, + 555, + 498 + ], + "type": "text", + "content": "[80] Bryan K. Orme. MaxDiff analysis: Simple counting, individual-level logit, and HB. Sawtooth Software Research Paper Series, 2009. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 499, + 555, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 499, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 315, + 499, + 555, + 522 + ], + "type": "text", + "content": "[81] Bruce Schneier. Applied cryptography - protocols, algorithms, and source code in C, 2nd Edition. Wiley, 1996. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 523, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 523, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 315, + 523, + 555, + 578 + ], + "type": "text", + "content": "[82] Kyle Lo, Lucy Lu Wang, Mark Neumann, Rodney Kinney, and Daniel Weld. S2ORC: The semantic scholar open research corpus. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4969-4983, Online, 2020. Association for Computational Linguistics. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 578, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 578, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 315, + 578, + 555, + 634 + ], + "type": "text", + "content": "[83] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. In Thirty-seventh Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 635, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 635, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 635, + 555, + 712 + ], + "type": "text", + "content": "[84] Anas Awadalla, Le Xue, Manli Shu, An Yan, Jun Wang, Senthil Purushwalkam, Sheng Shen, Hannah Lee, Oscar Lo, Jae Sung Park, Etash Kumar Guha, Silvio Savarese, Ludwig Schmidt, Yejin Choi, Caiming Xiong, and Ran Xu. BLIP3-KALE: Knowledge augmented large-scale dense captions. In Synthetic Data for Computer Vision Workshop @ CVPR 2025, 2025." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17805" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 117 + ], + "type": "text", + "content": "[85] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, Yuyin Zhou, and Cihang Xie. What if we recapture billions of web images with LLaMA-3?, 2024. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 162 + ], + "type": "text", + "content": "[86] R. Hadsell, S. Chopra, and Y. LeCun. Dimensionality reduction by learning an invariant mapping. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 1735-1742, 2006. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 228 + ], + "type": "text", + "content": "[87] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. ImageBind: One embedding space to bind them all. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15180-15190, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 296, + 305 + ], + "type": "text", + "content": "[88] Bin Zhu, Bin Lin, Munan Ning, Yang Yan, Jiaxi Cui, WANG HongFa, Yatian Pang, Wenhao Jiang, Junwu Zhang, Zongwei Li, Cai Wan Zhang, Zhifeng Li, Wei Liu, and Li Yuan. LanguageBind: Extending video-language pretraining to n-modality by language-based semantic alignment. In The Twelfth International Conference on Learning Representations, 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 307, + 260, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 307, + 260, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 307, + 260, + 318 + ], + "type": "text", + "content": "[89] OpenAI. Learning to reason with LLMs, 2024. 15" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 713 + ], + "type": "text", + "content": "[90] DeepSeek-AI, Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, Xiaokang Zhang, Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhihong Shao, Zhuoshu Li, Ziyi Gao, Aixin Liu, Bing Xue, Bingxuan Wang, Bochao Wu, Bei Feng, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Honghui Ding, Huajian Xin, Huazuo Gao, Hui Qu, Hui Li, Jianzhong Guo, Jiashi Li, Jiawei Wang, Jingchang Chen, Jingyang Yuan, Junjie Qiu, Junlong Li, J. L. Cai, Jiaqi Ni, Jian Liang, Jin Chen, Kai Dong, Kai Hu, Kaige Gao, Kang Guan, Kexin Huang, Kuai Yu, Lean Wang, Lecong Zhang, Liang Zhao, Litong Wang, Liyue Zhang, Lei Xu, Leyi Xia, Mingchuan Zhang, Minghua Zhang, Minghui Tang, Meng Li, Miaojun Wang, Mingming Li, Ning Tian, Panpan Huang, Peng Zhang, Qiancheng Wang, Qinyu Chen, Qiushi Du, Ruiqi Ge, Ruisong Zhang, Ruizhe Pan, Runji Wang, R. J. Chen, R. L. Jin, Ruyi Chen, Shanghao Lu, Shangyan Zhou, Shanhuang Chen, Shengfeng Ye, Shiyu Wang, Shuiping Yu, Shunfeng Zhou, Shuting Pan, S. S. Li, Shuang Zhou, Shaoqing Wu, Shengfeng Ye, Tao Yun, Tian Pei, Tianyu Sun, T. Wang, Wangding Zeng, Wanjia Zhao, Wen Liu, Wenfeng Liang, Wenjun Gao, Wenqin Yu, Wentao Zhang, W. L. Xiao, Wei An, Xiaodong Liu, Xiaohan Wang, Xiaokang Chen, Xiaotao Nie, Xin Cheng, Xin Liu, Xin Xie, Xingchao Liu, Xinyu Yang, Xinyuan Li, Xuecheng Su, Xuheng Lin, X. Q. Li, Xiangyue Jin, Xiaojin Shen, Xiaosha Chen, Xiaowen Sun, Xiaoxiang Wang, Xinnan Song, Xinyi Zhou, Xianzu Wang, Xinxia Shan, Y. K. Li, Y. Q. Wang, Y. X. Wei, Yang Zhang, Yanhong Xu, Yao Li, Yao Zhao, Yaofeng Sun, Yaohui Wang, Yi Yu, Yichao Zhang, Yifan Shi, Yiliang Xiong, Ying He, Yishi Piao, Yisong Wang, Yixuan Tan, Yiyang Ma, Yiyuan Liu, Yongqiang Guo, Yuan Ou, Yuduan Wang, Yue Gong, Yuheng Zou, Yujia He, Yunfan Xiong, Yuxiang Luo, Yuxiang You, Yuxuan Liu, Yuyang" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 662 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 171 + ], + "type": "text", + "content": "Zhou, Y. X. Zhu, Yanhong Xu, Yanping Huang, Yaohui Li, Yi Zheng, Yuchen Zhu, Yunxian Ma, Ying Tang, Yukun Zha, Yuting Yan, Z. Z. Ren, Zehui Ren, Zhangli Sha, Zhe Fu, Zhean Xu, Zhenda Xie, Zhengyan Zhang, Zhewen Hao, Zhicheng Ma, Zhigang Yan, Zhiyu Wu, Zihui Gu, Zijia Zhu, Zijun Liu, Zilin Li, Ziwei Xie, Ziyang Song, Zizheng Pan, Zhen Huang, Zhipeng Xu, Zhongyu Zhang, and Zhen Zhang. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning, 2025. 15" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 172, + 555, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 555, + 237 + ], + "type": "text", + "content": "[91] Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024. 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 239, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 555, + 282 + ], + "type": "text", + "content": "[92] Urbano Lorenzo-Seva and Jos M. F. ten Berge. Tucker's congruence coefficient as a meaningful index of factor similarity. Methodology: European Journal of Research Methods for the Behavioral and Social Sciences, 2(2):57-64, 2006. 15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 285, + 555, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 555, + 350 + ], + "type": "text", + "content": "[93] Jonas Belouadi and Steffen Eger. UScore: An effective approach to fully unsupervised evaluation metrics for machine translation. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 358-374, Dubrovnik, Croatia, 2023. Association for Computational Linguistics. 15" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 351, + 555, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 351, + 555, + 438 + ], + "spans": [ + { + "bbox": [ + 316, + 351, + 555, + 438 + ], + "type": "text", + "content": "[94] Wei Zhao, Maxime Peyrard, Fei Liu, Yang Gao, Christian M. Meyer, and Steffen Eger. MoverScore: Text generation evaluating with contextualized embeddings and earth mover distance. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 563-578, Hong Kong, China, 2019. Association for Computational Linguistics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 440, + 555, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 440, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 440, + 555, + 506 + ], + "type": "text", + "content": "[95] Wei Zhao, Goran Glavaš, Maxime Peyrard, Yang Gao, Robert West, and Steffen Eger. On the limitations of cross-lingual encoders as exposed by reference-free machine translation evaluation. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1656–1671, Online, 2020. Association for Computational Linguistics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 507, + 555, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 507, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 316, + 507, + 555, + 572 + ], + "type": "text", + "content": "[96] Yurun Song, Junchen Zhao, and Lucia Specia. SentSim: Crosslingual semantic evaluation of machine translation. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3143-3156, Online, 2021. Association for Computational Linguistics. 15" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 574, + 555, + 617 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 574, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 316, + 574, + 555, + 617 + ], + "type": "text", + "content": "[97] Y. Rubner, C. Tomasi, and L.J. Guibas. A metric for distributions with applications to image databases. In Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271), pages 59-66, 1998. 15" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 618, + 555, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 618, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 316, + 618, + 555, + 662 + ], + "type": "text", + "content": "[98] Matt Kusner, Yu Sun, Nicholas Kolkin, and Kilian Weinberger. From word embeddings to document distances. In Proceedings of the 32nd International Conference on Machine Learning, pages 957–966, Lille, France, 2015. PMLR. 15" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17806" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/e220fb2d-345d-4c10-96de-761db897dc9b_content_list.json b/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/e220fb2d-345d-4c10-96de-761db897dc9b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..27b0b6935509b889abad36528b8d05d4462df69b --- /dev/null +++ b/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/e220fb2d-345d-4c10-96de-761db897dc9b_content_list.json @@ -0,0 +1,1499 @@ +[ + { + "type": "text", + "text": "Tile-wise vs. Image-wise: Random-Tile Loss and Training Paradigm for Gaussian Splatting", + "text_level": 1, + "bbox": [ + 137, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyu Zhang $^{1*}$ Weihong Pan $^{2*}$ Xiaojun Xiang $^{1}$ Hongjia Zhai $^{2}$ Liyang Zhou $^{1}$ $^{1}$ SenseTime Research ${}^{2}$ State Key Lab of CAD&CG, Zhejiang University", + "bbox": [ + 120, + 200, + 872, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ebcf5affed1d17c580fa56a2a3e5f1db03ea669d7f7ca3838945301ef832f074.jpg", + "image_caption": [ + "Figure 1. Random-Tile Loss integrates multi-view appearance and structural constraints during each optimization step. Under the same batch size, it consistently improves convergence efficiency and rendering quality while preserving compactness and real-time capability. The results on test views across different training iterations suggest a more stable and efficient convergence compared to the image-wise loss in Gaussian Splatting. Please refer to the Supplementary Video for more detailed comparisons on convergence performance." + ], + "image_footnote": [], + "bbox": [ + 112, + 266, + 890, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 587, + 326, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) has drawn significant attention for its advantages in rendering speed and quality. Most existing methods still rely on the image-wise loss and training paradigm because of its intuitive nature in the Splatting algorithm. However, image-wise loss lacks multi-view constraints, which are generally essential for optimizing 3D appearance and geometry. To address this, we propose RT-Loss along with a tile-based training paradigm, which uses randomly sampled tiles to integrate multi-view appearance and structural constraints in 3DGS. Additionally, we introduce an tile-based adaptive densification control strategy tailored for our training paradigm. Extensive experiments show that our approach consistently improves performance metrics while maintaining efficiency across various benchmark datasets.", + "bbox": [ + 88, + 619, + 485, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 587, + 643, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF) [33] has brought revolutionary advances to photorealistic novel view synthesis (NVS) through exceptional rendering quality, attracting significant attention from the computer vision, graphics, and robotics communities [44]. Recently, 3D Gaussian Splating (3DGS) [19] has significantly enhanced rendering efficiency by replacing traditional ray marching with the splating algorithm [22]. Numerous studies have focused on optimizing the Gaussian feature or heuristic split algorithm in 3DGS to improve the appearance quality and geometric accuracy [6, 24, 31, 51].", + "bbox": [ + 509, + 612, + 906, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, while the splatting algorithm significantly improves rendering speed, its potential for incorporating Tile-based multi-view constraints remains underexplored. Specifically, 3DGS initializes with point clouds derived from Structure from Motion (SfM) [1, 12, 36] and optimizes a set of anisotropic 3D Gaussians to represent the scene. Tile-based splatting plays a crucial role in achieving real-time rendering by directly splatting visible Gaussians onto", + "bbox": [ + 509, + 780, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Authors contributed equally.", + "bbox": [ + 107, + 875, + 267, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26923", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the camera view, rather than querying every ray point. As a result, defining the loss function on each training image $\\mathcal{I}$ is a intuitive and straightforward choice:", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\Theta) = (1 - \\lambda) L _ {\\mathrm {M A E}} (\\Theta , \\mathcal {I}) + \\lambda L _ {\\mathrm {D - S S I M}} (\\Theta , \\mathcal {I}), \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 148, + 482, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $L_{MAE}$ and $L_{D-SSIM}$ are computed image-wise on the entire training image. Although this approach benefits the calculation of structural loss SSIM across the dense image, it also means that gradients from multiple viewpoints cannot be integrated in a single optimization step. In contrast, NeRF typically optimizes a pixel-wise loss:", + "bbox": [ + 89, + 179, + 483, + 270 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\Theta) = \\frac {1}{\\| \\mathcal {R} \\|} \\sum_ {\\boldsymbol {r} \\in \\mathcal {R}} L _ {\\mathrm {M S E}} (\\Theta , \\boldsymbol {r}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 281, + 483, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where each individual ray $r = (x, d, c)$ in the mini-batch $\\mathcal{R}$ is randomly sampled from the entire training dataset. Pointwise sampling from random viewpoints provides more comprehensive gradients. However, it fails to utilize the collective supervision of distant pixels, thus inevitably overlooking structural information in the overall scene [42].", + "bbox": [ + 89, + 330, + 483, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we propose a Tile-based training paradigm that integrates both multi-view constraints and structural information within the splatting process. We sample a set of tiles $\\mathcal{T}$ from multiple training views, each individual tile $t$ is the minimal unit in splatting. The loss is formulated as follows:", + "bbox": [ + 89, + 422, + 483, + 512 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R T} (\\Theta) = \\frac {1}{\\| \\mathcal {T} \\|} \\sum_ {\\boldsymbol {t} \\in \\mathcal {T}} L _ {\\mathrm {M A E}} (\\Theta , \\boldsymbol {t}) + \\lambda L _ {\\mathrm {R T - S S I M}} (\\Theta , \\mathcal {T}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 521, + 483, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which is computed over a random-tile batch $\\mathcal{T}$ that contains multi-view appearance and structure. Our Tile-based training paradigm maintains the training efficiency of the 3DGS rendering process, with detailed definitions and calculations provided in Sec. 3.2. Overall, our contributions are as follows:", + "bbox": [ + 89, + 585, + 483, + 674 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Tile-based training paradigm and loss formulation, incorporating both multi-view appearance constraints and structural similarity constraints with random tile sampling. To the best of our knowledge, we are the first to introduce a random-tile approach as a replacement for image-wise training paradigms in 3DGS.", + "- Extensive experiments indicate that this Tile-based training paradigm yields notable improvements in training performance while maintaining model compactness." + ], + "bbox": [ + 89, + 676, + 482, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 829, + 233, + 844 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel View Synthesis. Neural Radiance Field (NeRF) [33] has significantly advanced novel view synthesis by employing a multilayer perceptron (MLP) to model scenes", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with volumetric rendering, enabling the generation of realistic images from different perspectives. NeRF introduced a transformative approach to scene representation, sparking extensive research focused on enhancing rendering quality [2, 3, 27, 37, 50, 54, 55], speeding up computation [5, 14, 17, 34, 35], enabling 3D content generation [38], handling dynamic scenes [9, 15, 25], and improving generalization to novel scenes [45]. Due to the ray marching rendering technique, most NeRF-based methods are trained by utilizing batches of randomly sampled rays, with mean squared error (MSE) as the primary loss function. To integrate structural similarity, S3IM [42] enhances NeRF by incorporating patches of random rays in the loss calculation with a structural similarity index (SSIM), improving performance in tasks such as sparse view synthesis.", + "bbox": [ + 511, + 90, + 906, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gaussian Splatting Training Paradigm. 3D Gaussian Splatting (3DGS) has gained significant attention in novel view synthesis tasks due to its impressive performance. Using a rasterization-based rendering approach [20], 3DGS avoids the computational overhead of ray marching. With its high rendering speed and quality, 3DGS has found applications across various fields [48], including dynamic scene reconstruction [7, 10, 28, 29, 32, 41, 47], sparse view synthesis [43, 52], reflective object modeling [53], and geometric surface recovery [24, 30, 40]. Additional optimizations target specific issues, such as the geometric improvements seen in Scaffold-GS [31], which combines neural field models with Gaussian Splatting to achieve state-of-the-art surface accuracy. [18] introduces adaptive strategies to reduce dependency on SfM initialization. While most approaches adopt an image-wise training paradigm, some recent research explores multi-view constraints to enhance Gaussian Splatting. For example, [46] leverages multi-view stereo (MVS) to enhance initialization geometry, using multi-view consistency to optimize Gaussian geometry. [8, 10, 26] utilize batch accumulation to alleviate data pressure from the temporal dimension by rendering multiple images per iteration and accumulating gradients. However, this straightforward image-wise accumulation not only impacts performance but also fails to fully leverage the benefits of multiview constraints with the same batch size.", + "bbox": [ + 511, + 324, + 908, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 513, + 750, + 611, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 781, + 653, + 795 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gaussian Splatting. 3DGS [19] represents scenes using anisotropic 3D Gaussians, which retains the differential properties of volumetric representations while enabling efficient rendering through Tile-based rasterization. Each Gaussian is assigned a color $c$ , represented by spherical harmonics, and an opacity $\\alpha$ , which modulates $G(x)$ during", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26924", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "blending. Its spatial distribution is then defined by:", + "bbox": [ + 89, + 90, + 428, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (x) = \\exp \\left(- \\frac {1}{2} (x - \\mu) ^ {T} \\Sigma^ {- 1} (x - \\mu)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 114, + 482, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x$ denotes a spatial location in the 3D scene and $\\mu$ , initialized from Structure-from-Motion (SfM) points, specifies the Gaussian center. The covariance matrix $\\Sigma$ is constructed using a scaling matrix $S$ and rotation matrix $R$ as $\\Sigma = RSS^T R^T$ to ensure positive semi-definiteness.", + "bbox": [ + 89, + 155, + 482, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tile-based Rendering. Tile-based rendering is the key to achieving high-resolution and real-time performance in 3DGS. Unlike conventional volumetric methods, it projects 3D Gaussians into 2D Gaussians $G'(x)$ on the image plane rather than relying on resource-intensive ray marching. The image is divided into tiles of size $16 \\times 16$ to process all sorted 2D Gaussians in parallel. Each pixel within a tile applies $\\alpha$ -blending:", + "bbox": [ + 89, + 231, + 483, + 352 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC \\left(x ^ {\\prime}\\right) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {j}\\right), \\quad \\sigma_ {i} = \\alpha_ {i} G _ {i} ^ {\\prime} \\left(x ^ {\\prime}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 361, + 482, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $x$ represents the queried pixel position, and $\\mathcal{N}$ denotes the number of sorted 2D Gaussians linked to that tile. By leveraging the Tile-based differentiable rasterizer, all 3D Gaussians projected onto the training view can be directly optimized in an end-to-end manner.", + "bbox": [ + 89, + 411, + 483, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MAE, MSE, and SSIM. These three loss functions are widely used for image quality assessment. The Mean Absolute Error (MAE) and Mean Squared Error (MSE) are two methods for calculating RGB loss:", + "bbox": [ + 89, + 486, + 483, + 547 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M A E} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| y _ {i} - \\hat {y} _ {i} \\right|, \\mathrm {M S E} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(y _ {i} - \\hat {y} _ {i}\\right) ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 555, + 482, + 595 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S S I M} (x, y) = \\frac {\\left(2 \\mu_ {x} \\mu_ {y} + C _ {1}\\right) \\left(2 \\sigma_ {x y} + C _ {2}\\right)}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + C _ {1}\\right) \\left(\\sigma_ {x} ^ {2} + \\sigma_ {y} ^ {2} + C _ {2}\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 614, + 482, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu_{x}$ and $\\mu_{y}$ are the means of the images $x$ and $y$ , $\\sigma_x^2$ and $\\sigma_y^2$ are the variances, $\\sigma_{xy}$ is the covariance between the images, and $C_1, C_2$ are constants to stabilize the division when the denominator is small. MAE measures the average absolute error, treating all deviations equally. MSE squares each error before averaging, making it more sensitive to large errors and outliers. The Structural Similarity (SSIM) focuses on perceptual quality by comparing structural patterns, luminance, and contrast between images.", + "bbox": [ + 89, + 657, + 483, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Tile-based Training via RT-Loss", + "text_level": 1, + "bbox": [ + 89, + 801, + 375, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce RT-Loss and our Tile-based training paradigm. Our goal is to incorporate random viewpoint observations within each training batch, while preserving structural information to enforce more comprehensive constraints. In contrast to the individual rays used", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in NeRFs, Gaussian Splatting utilizes a Tile-based splatting process that projects Gaussians onto the entire view screen, rather than performing active ray marching. Hence, the smallest unit in the loss calculation must align with the splatting rendering and is set as a $16 \\times 16$ tile $t$ . We load the datasets as tiles instead of images, randomly sample $N$ tiles as a training batch $\\mathcal{T}$ and obtain the rendered tiles set $\\mathcal{T}(\\hat{\\mathcal{C}})$ along with their corresponding ground truth $\\mathcal{T}(\\mathcal{C})$ , where $\\hat{\\mathcal{C}} = \\{\\hat{\\mathcal{C}}(t) \\mid t \\in \\mathcal{T}\\}$ and $\\mathcal{C} = \\{\\mathcal{C}(t) \\mid t \\in \\mathcal{T}\\}$ . We use MAE as the color loss, which can be written as:", + "bbox": [ + 511, + 90, + 906, + 243 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {R T} - \\mathrm {M A E}} = \\frac {1}{\\| \\mathcal {T} \\|} \\sum_ {\\boldsymbol {t} \\in \\mathcal {T}} \\| \\hat {\\boldsymbol {C}} (\\boldsymbol {t}) - \\boldsymbol {C} (\\boldsymbol {t}) \\| _ {1}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 252, + 906, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The tiles contain viewpoint's structural information. If we solely rely on rgb loss $L_{\\mathrm{MAE}}$ , the positional relationship contained within the tiles would be completely lost [42]. Inspired by S3IM's adaptation of SSIM on random rays, we propose RT-SSIM for 3DGS to retain the structural information embedded in random tiles. The concept is straightforward: while keeping the total batch size unchanged, we compute and average the SSIM across multiple independent and non-overlapping random tiles from different views, with kernel size and stride set to $K \\times K$ (e.g., $9 \\times 9$ for overall structtrue or $3 \\times 3$ for fine detail) with a stride of $s$ . RT-SSIM can be written as:", + "bbox": [ + 511, + 297, + 908, + 479 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R T - S S I M} (\\hat {\\mathcal {T}}, \\mathcal {T}) = \\frac {1}{N} \\sum_ {\\boldsymbol {t} \\in \\mathcal {T}, v \\in \\boldsymbol {V}} \\operatorname {S S I M} \\left(\\boldsymbol {t} ^ {(v)} (\\hat {\\mathcal {C}}), \\boldsymbol {t} ^ {(v)} (\\mathcal {C})\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 486, + 905, + 540 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $V$ controls the number of views in all $N$ tiles. In particular, when $V = N$ , it implies all tiles are selected from different views. As RT-SSIM positively correlated with image similarity, we define the structure loss $L_{\\mathrm{RT - SSIM}}$ as", + "bbox": [ + 511, + 550, + 905, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {R T} - \\mathrm {S S I M}} (\\Theta , \\mathcal {T}) = 1 - \\mathrm {R T} - \\mathrm {S S I M} (\\hat {\\mathcal {T}}, \\mathcal {T}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 636, + 903, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We note that tiles are particularly suitable for parallel computation, thus the additional computational cost of Tile-based training is minimal. The final hybrid loss $\\mathcal{L}_{\\mathrm{RT}}$ can be written as:", + "bbox": [ + 511, + 665, + 905, + 724 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R T}} (\\Theta , \\mathcal {T}) = (1 - \\lambda_ {s}) L _ {\\mathrm {R T} - \\mathrm {M A E}} + \\lambda_ {s} L _ {\\mathrm {R T} - \\mathrm {S S I M}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 737, + 903, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda_{s}$ balances structure and color losses.", + "bbox": [ + 511, + 763, + 812, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present the pseudocode in Algorithm 1. In the standard 3DGS training paradigm, tiles are rendered in parallel using the same view parameters. In contrast, we sample tiles from multiple different views to integrate multi-view constraints. In our random-tile training paradigm, We introduce two additional hyperparameters, $\\lambda_{s}$ and $V$ , where $\\lambda_{s}$ controls the loss weight and $V$ determines the number of viewpoints. When $V = N$ , each tile is sampled from a", + "bbox": [ + 511, + 780, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26925", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Algorithm 1: Tile-based Training via RT-loss \n1 Let $\\mathcal{A}$ be an SGD training algorithm; \n2 while the termination criteria has not been satisfied do \n3 Sample a training batch $\\mathcal{T}$ consisting of $N$ tiles from $V$ views in dataset $\\mathcal{D}$ ; \n4 for $i = 1$ to $V$ do \n5 Randomly select tiles $\\mathcal{T}_i = [t_1, t_2 \\dots t_{V/N}]$ , where $t_j \\in \\mathcal{T}_i \\subseteq \\mathcal{T}$ ; \n6 VisibilityCheck(); \n7 end \n8 for each tile $t$ in $\\mathcal{T}$ do \n9 Compute the forward splating result after $\\hat{\\mathcal{C}} = \\{\\hat{C}(t) \\mid t \\in \\mathcal{T}\\}$ ; \n10 Obtain the ground-truth $\\mathcal{C} = \\{C(t) \\mid t \\in \\mathcal{T}\\}$ ; \n11 end \n12 for $i = 1$ to $V$ do \n13 let rendered pixels $\\hat{\\mathcal{C}}$ form the pixel patch $\\pmb{t}^{(v)}(\\hat{\\mathcal{C}})$ in view $V_i$ ; \n14 Compute SSIM $(\\pmb{t}^{(v)}(\\hat{\\mathcal{C}}), \\pmb{t}^{(v)}(\\mathcal{C}))$ with the given kernel $K \\times K$ ; \n15 end \n16 Obtain the RT-SSIM loss $L_{\\mathrm{RT - SSIM}}(\\Theta, \\mathcal{T}) = 1 - \\frac{1}{N} \\sum \\mathrm{SSIM}(\\pmb{t}^{(v)}(\\hat{\\mathcal{C}}), \\pmb{t}^{(v)}(\\mathcal{C}))$ ; \n17 Obtain the conventional $L_{MAE}$ loss as color loss $L_{\\mathrm{RT - MAE}}(\\Theta) = \\frac{1}{||\\mathcal{T}||} \\sum_{t \\in \\mathcal{T}} ||\\hat{\\mathcal{C}}(t) - C(t)||_1$ ; \n18 $\\mathcal{L}(\\Theta, \\mathcal{T}) = (1 - \\lambda_s)L_{\\mathrm{RT - MAE}}(\\Theta, \\mathcal{T}) + \\lambda_s L_{\\mathrm{RT - SSIM}}(\\Theta, \\mathcal{T})$ ; \n19 Compute the gradient $\\nabla \\mathcal{L}(\\Theta)$ ; \n20 Update model parameters $\\Theta$ by $\\mathcal{A}$ \n21 end", + "bbox": [ + 91, + 93, + 480, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "distinct view, and their visibility is computed in parallel during preprocessing to accelerate the rendering process. For a fair comparison, we set the total tiles count $N$ to match that in image-wise training, ensuring a consistent batch size in gradient descent algorithm $\\mathcal{A}$ . Our random-tile strategy uniformly samples tiles across each view, balancing observations for each perspective. Our ablation study in Sec. 4.4 shows that this strategy achieves slightly better results than purely random selection.", + "bbox": [ + 89, + 655, + 483, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tile-based Adaptive Density Control", + "text_level": 1, + "bbox": [ + 89, + 801, + 406, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In 3D Gaussian Splating, a Gaussian is split or cloned depending on whether the average cumulative 2D gradient on the NDC coordinates $\\sum \\| \\mathbf{g}_i\\|$ across interval iterations exceeds a specified threshold $\\tau_{\\mathrm{densify}}$ , as shown in Fig. 2. During our Tile-based training, Gaussians in one iteration will", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a076c6dcc119548e866a1570437a3f7d52a44658c2e20de84f1f8e9ebbecbea8.jpg", + "image_caption": [ + "Figure 2. Image-wise vs Tile-wise Adaptive Density Control. In Tile-based training, NDC gradients for densification are computed as a weighted sum across all observations, balancing multi-view contributions, where $S1$ represents the area of the Gaussian projection onto the tile, and $S2$ is the total projected area." + ], + "image_footnote": [], + "bbox": [ + 526, + 99, + 671, + 316 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bb3fca209b35109499f4ef69e35f33f058d8d5b8d800197aec5a16773fef25d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 103, + 880, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "have multiple NDC gradients from different training views. Thus, there are two methods for computing density gradients: Tile-Count Densification and Iteration-Count Densification.", + "bbox": [ + 511, + 431, + 906, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\sum_ {k = 1} ^ {\\text {I t e r}} \\sum_ {m = 1} ^ {M _ {k}} \\| \\mathbf {g} _ {k m} \\|}{\\sum_ {k = 1} ^ {\\text {I t e r}} M _ {k}} > \\tau_ {\\text {d e n s i f y}} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 497, + 905, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\sum_ {k = 1} ^ {\\text {I t e r}} \\sum_ {m = 1} ^ {M _ {k}} \\left(\\left\\| \\mathbf {g} _ {k m} \\right\\| \\times r _ {m}\\right)}{\\sum_ {k = 1} ^ {\\text {I t e r}} 1} > \\tau_ {\\text {d e n s i f y}}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 553, + 903, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\text{Iter}$ represents the number of iterations in which a Gaussian participates in $\\alpha$ -blending within each densification interval. In Tile-count Densification ( Eq. (12)), the NDC gradient is averaged over the number of observed tiles $M_{k}$ , while in Iteration-Count Densification ( Eq. (13)), it is averaged over the number of observed iterations. Additionally, in Iteration-Count Densification, the NDC gradient is weighted by $r_{m}$ , the ratio of the projected area shown in Fig. 2. To maintain consistency with the split parameters used in the original 3DGS and to emphasize the effect of RT-Loss, we default to using Iteration-Count for adaptive density control. Further details on NDC gradients and densification algorithm can be found in the Supplementary Material.", + "bbox": [ + 511, + 599, + 906, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 823, + 645, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 848, + 702, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset and Metrics. We conduct a comprehensive evaluation on 26 publicly available scenes widely used", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26926", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/07180de8892f5710ab3ca20ed72c83f68a7e12926427b9bf384c7b2c2a62672d.jpg", + "table_caption": [ + "Table 1. Quantitative comparison to previous methods on benchmark datasets. We report results on three commonly used datasets. The best, second best, and third best results are denoted by red, orange, and yellow, respectively." + ], + "table_footnote": [], + "table_body": "
Dataset\nMethod | MetricsMip-NeRF360Tanks&TemplesDeep Blending
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Plenoxels [13]23.620.6700.44322.130.7620.33423.060.7950.510
Instant-NGP [34]26.430.7250.33922.940.7520.30423.620.7970.423
Mip-NeRF360 [3]29.230.8440.20723.120.7710.25029.400.9010.245
3D-GS [19]28.690.8700.18224.360.8310.21329.410.9030.243
Scaffold-GS [31]28.840.8480.22024.810.8400.20530.210.9060.254
Ours29.660.8920.17125.130.8520.19230.150.9110.224
Ours (+ Scaffold-GS)29.860.8850.20625.320.8570.18930.490.9130.231
", + "bbox": [ + 150, + 128, + 848, + 261 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1ba36238fca19915e7c3b35d8f16d1ed327a9c034b61561d95b3cccd8f40be24.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison across benchmark datasets [3, 11, 21]. We highlight the differences in the image using patches for clearer visibility. Our approach consistently outperforms 3D-GS on these scenes, demonstrating clear advantages in challenging scenarios, such as fine-scale details (MIP360-ROOM(a), TANDT-BARN), thin geometry (SMERF-NYC(a), SMERF-NYC(b)), texture-less regions (TANDT-TRAIN), light effects (MIP360-COUNTER, TANDT-TRUCK), and insufficient observations (TANDT-BARN, SMERF-NYC(b))." + ], + "image_footnote": [], + "bbox": [ + 106, + 277, + 897, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in the field of 3D reconstruction across multiple diverse tasks. Specifically, we evaluate our method on seven scenes from Mip-NeRF360 [3], another seven from Tanks&Temples [21], and two additional scenes from DeepBlending [16]. To further demonstrate the limitations of image-wise loss in capturing global structure, our evaluation incorporated all four scenes from [4], with $5 - 10 \\times$ more training views. We additionally evaluated dynamic benchmarks, specifically using the Neural 3D Video Dataset containing six scenes, to demonstrate the effectiveness of RT-Loss in 4D reconstruction. In addition to the commonly used metrics (PSNR, SSIM [39], and LPIPS [49]), we fur", + "bbox": [ + 89, + 707, + 485, + 888 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ther provide the storage size (MB) and rendering speed (FPS) to assess model compactness and efficiency, with all experiments performed on an NVIDIA RTX 4090 GPU. The main paper reports the average metrics across all scenes in each dataset, while detailed per-scene quantitative results are presented in the supplementary materials.", + "bbox": [ + 511, + 707, + 908, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline and Implementation. 3DGS[19], Scaffold-GS[31] and 4D-GS [10] are selected as our baselines due to their state-of-the-art performance. We also present the results of Mip-NeRF360 [3], iNGP [34] and Plenoxels [13] as in [19]. To demonstrate the generality of the tile-based ap", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26927", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/13d751e8e657b1770978728860ef8a278143b67ab58d0941ecd1adf8ab9a47c3.jpg", + "image_caption": [ + "Figure 4. Convergence performance. We evaluated the convergence of PSNR for both the training and testing views during the training process. With the same batch size, the convergence of image-wise loss in 3DGS is slower than RT-Loss, with the gap widening on larger datasets, e.g., COURTHOUSE and BERLIN" + ], + "image_footnote": [], + "bbox": [ + 102, + 89, + 367, + 277 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/022bc578cbbd7af7a70a6439ccf4adf392cfcec02a7324e64f459960cee6d63b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 89, + 625, + 282 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ba3d1e74b5577f324a2ae4ad3745555784ef85925c3e046fbfd0224e204d501.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 90, + 890, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "proach, we implemented the tile-wise method in Scaffolds for comparison. We trained for $30k$ iterations and rendered the same number of pixels per iteration, ensuring the total pixels from random tiles matched a complete image. To independently demonstrate the effectiveness of the tile-based training paradigm and RT-loss, we modified only the sampling and loss calculations for random tile training while keeping the default 3DGS parameters unchanged, including the densification threshold $\\tau_{\\mathrm{densify}} = 0.0002$ and all learning rates. For RT-loss, we sample random tiles from $V = 5$ views and apply the same structure loss weight, $\\lambda_s = 0.2$ , as used in the original 3DGS.", + "bbox": [ + 88, + 364, + 483, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results Analysis", + "text_level": 1, + "bbox": [ + 89, + 559, + 253, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark Evaluation. The quantitative results on Mip-NeRF360 and Tanks&Temples are presented in Tab. 1. Results show that we outperform existing methods on standard datasets. Fig. 3 illustrates that tile-wise training achieves enhanced visual quality, with improved geometry accuracy and richer texture details. Moreover, our method retains the efficiency of Gaussian methods, as shown in Tab. 2. We compared the rendering speed and storage size with 3DGS. We achieve improved performance while using a comparable or fewer number of Gaussians, demonstrating high-quality reconstruction without compromising speed. By incorporating our tile-based approach, Scaffold-GS also achieved notable improvements, suggesting that the image-wise training paradigm, which lacks multi-view constraints, limits the reconstruction quality of various GS methods.", + "bbox": [ + 89, + 582, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Convergence Performance. We evaluated our convergence performance across several representative scenarios. As shown in Fig. 4, compared to the original image-wise loss in 3DGS, the RT-Loss based on multi-view observations provides more effective constraints, resulting in significantly faster convergence. The difference in convergence speed is", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f93cf6752b9e3190523c6d2e452b51aa8c281bc20dc6227099309ea9d59fabe9.jpg", + "table_caption": [ + "Table 2. Rendering FPS and storage comparison. The rendering speed of both methods is measured on the same machine." + ], + "table_footnote": [], + "table_body": "
ScenesMip-NeRF360ROOMKITCHEN
MemFPSPSNRMemFPSPSNRMemFPS
3DGS6717831.7931712331.63372118
Ours6259332.8131812532.35320135
ScenesTanks&TemplesCourthOUSEMEETINGROOM
MemFPSPSNRMemFPSPSNRMemFPS
3DGS31712623.1317414626.27276120
Ours30213024.1416813926.98260133
", + "bbox": [ + 517, + 402, + 903, + 511 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a11e154853d5f07d396f858ba0a7eaa866b1f966bf3c4e0135290509b9b87b60.jpg", + "table_caption": [ + "Table 3. PSNR evaluation for scenes with increased training views. The convergence capability of RT-Loss is particularly evident on the SMERF dataset[11], which includes approximately 10 times more training views than Mipnerf360. Additionally, we recorded the PSNR gap at 4k and 30k iterations." + ], + "table_footnote": [], + "table_body": "
Scene.views test iterationAlameda(1734)Berlin(1511)London(1874)
4k30k4k30k4k30k
3DGS17.6222.2722.6327.5322.3225.75
Ours20.1123.91624.8028.7924.0726.626
", + "bbox": [ + 517, + 604, + 901, + 669 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/33b16967fda4dde009619f50b83af835e695bf9cb425e10db153f6ce5605c6b3.jpg", + "table_caption": [ + "Table 4. Quantitative comparison on dynamic datasets. We applied RT loss for dynamic scene reconstruction and compared it with 4D-GS[10] on the challenging real-world dynamic dataset Neu3D[23]." + ], + "table_footnote": [], + "table_body": "
ModelPSNR↑D-SSIM↓LPIPS↓FPS↑Storage (MB)↓
4D-GS[10]31.150.9740.0493090
Ours31.740.9730.0433478
", + "bbox": [ + 517, + 748, + 901, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "visually evident in Fig. 1. This advantage in convergence is even more pronounced in scenes with a large number of training views, e.g., COURTHOUSE and BERLIN.", + "bbox": [ + 511, + 824, + 906, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scene with More Training Views. We compared our method with the original 3DGS in scenarios with more", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26928", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/90541314fd9e72cfe853e4ded18acf5f31110dd2a786fd58303c0c1a8a5f0ea0.jpg", + "image_caption": [ + "Figure 5. Gaussians structure comparison. RT-Loss demonstrates stronger structural reconstruction ability in scenes with more training views. In contrast, 3DGS not only loses details but also fails to recover the geometric structure of the scene." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12290bde3171d39c0a76a115e98ef6a414861af77d76583485dc54b62c29df93.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons on the Neural 3D Video Dataset. In 4D reconstruction, constraints from multiple viewpoints and time also help generate more accurate details." + ], + "image_footnote": [], + "bbox": [ + 94, + 329, + 478, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training views. In Tab. 3, our method achieved a significant improvement in metrics. The gap is particularly noticeable in the early training stages (4k iterations), as image-wise training naturally requires more iterations to fully capture the scene. Fig. 5 further illustrates that our method produces a more refined geometric structure on large datasets. This supports our hypothesis about the current training paradigm: as the number of training views increases, image-wise training struggles to comprehensively capture scene information, resulting in less reliable optimization and densification. Furthermore, through experiments with extended training durations, we demonstrate that naively increasing the iteration count for image-wise training does not effectively address this limitation. This is because once incorrect geometry is formed during Gaussian optimization, it is difficult to correct by simply extending the training process. For further details, please refer to the supplementary material.", + "bbox": [ + 88, + 566, + 482, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dynamic Scenes. We evaluated our method on the real-world dynamic scene dataset, the Neural 3D Video Dataset, using 4D-GS [10] as the baseline. The expansion of the time dimension in the dynamic dataset resulted in a larger", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/adfb8d7efd72ac251a2e2a69e964aaa8c80e7ed193171c1133863bddcdc94c30.jpg", + "table_caption": [ + "Table 5. Ablation study on RT-Loss components. 'Img-' refers to image-wise loss, and 'RT-' refers to random-tile loss." + ], + "table_footnote": [], + "table_body": "
Scene(views)KITCHEN(279)CORTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
LImg-MAE + LImg-SSIM (3DGS)31.6737223.11173
LImg-MAE31.0913121.3544
LRT-MAE31.6512722.6363
LRT-MAE + LImg-SSIM32.0235223.50186
LRT-MAE + LRT-SSIM (Ours)32.3532024.15168
", + "bbox": [ + 516, + 125, + 898, + 215 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f93a3f79901d2859bca8764b8f8f7216b3f9b7f8c54bbbd10c8be1b5adb09ad9.jpg", + "table_caption": [ + "Table 6. Effect of Adaptive Density Control. 'Tile-Count' and 'Iteration-Count' are densification strategies described in Sec. 3.3, with 'Iteration-Count' being the default." + ], + "table_footnote": [], + "table_body": "
Scene(views)KICHEN(279)COURTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
3DGS w/o Densification29.845822.96118
Ours w/o Densification30.325823.83118
Ours w/ Tile-Count31.9517523.56125
Ours w/ Iteration-Count32.3532024.15168
", + "bbox": [ + 532, + 279, + 880, + 353 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/74d7ed9d12947bf4430e6336044e8680b88e0cb0434ec152ac5a5b277a0502ef.jpg", + "table_caption": [ + "Table 7. Effect of Tile Sampling and Batch Training. 'V' refers to the number of randomly selected viewpoints, determining the spread of observations. 'B' refers to Batch train, which means using multiple images as a batch for training. 'Non-Uniform' and 'Uniform' indicate whether the tiles are distributed evenly." + ], + "table_footnote": [], + "table_body": "
Scene.views)KICHEN(279)COURTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
tile-wise V = 231.8034323.76182
tile-wise V = 432.2233723.92175
tile-wise V = 632.4033224.29161
image-wise B = 231.6735223.82214
image-wise B = 431.2234823.69207
Non-Uniform Sample32.3031024.02192
Uniform Sample32.3832024.14168
", + "bbox": [ + 534, + 445, + 880, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "volume of training data. Although most 4D Gaussian methods address this issue by using batch training and rendering multiple images together as a training batch, our results show that, with the same batch size, tile-based training provides more effective constraints across multiple viewpoints and time frames. The results show that RT-Loss is also effective for dynamic scene reconstruction. As depicted in Tab. 4, we notably improved the rendering performance on dynamic scenes without sacrificing efficiency. As shown in Fig. 6, our method reconstructs finer details, attributed to our randomized approach, which provides stronger constraints across both viewpoints and time frames compared to image-wise training. Details are provided in the supplementary material.", + "bbox": [ + 511, + 582, + 906, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 801, + 679, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of RT-Loss components. We conducted an ablation on the loss components of Eq. (11), which includes the base color loss $L_{\\mathrm{RT - MAE}}$ and the structural loss $L_{\\mathrm{RT - SSIM}}$ . As shown in the $2^{nd}$ and $3^{rd}$ row in Tab. 5, when only MAE loss is used, the PSNR significantly decreases, with a lower", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26929", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2141aae8c323a480c32e98b5946b71910c7133473a04e988f9195bc5cafec40d.jpg", + "image_caption": [ + "Figure 7. Effect of initialization on reconstruction quality. (Left) Gaussians optimized from SfM point initialization. (Right) Gaussians optimized from random points. Our method reconstructs better structures and finer details even with random initialization, demonstrating superior robustness to initialization issues." + ], + "image_footnote": [], + "bbox": [ + 94, + 95, + 222, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e2a0f2ecf5a0597b516e8951e4c3ba6c92cb1969ad768c9b32e12232e2df3eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 95, + 352, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c29dda0bb4a84c4a9fcaafe348d54de7ec83b886b3c270b65ecf048ce8623e56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 95, + 478, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fdf1e83b03d6abf6e02e6518a5434ead7809fa29b27d9bd73a73151fa7471906.jpg", + "table_caption": [ + "Table 8. Sensitivity to Initialization. In the 'w/o SfM initialization' setting, Gaussians are initialized with random positions. RT-Loss demonstrates stronger robustness in addressing initialization issues." + ], + "table_footnote": [], + "table_body": "
Scene(views)KICHEN(279)COURTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
3DGS w/ SfM initialization31.6337223.13174
3DGS w/o SfM initialization29.8926921.77121
Ours w/ SfM initialization32.3532024.14168
Ours w/o SfM initialization31.2227223.21186
", + "bbox": [ + 96, + 410, + 478, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tendency for Gaussian densification. However, our $L_{\\mathrm{RT-MAE}}$ still outperforms the image-wise constraints, and this effect becomes more pronounced in the CourtHouse scene with more training views, where the PSNR difference reaches up to 1.3 dB. Notably, in the Kitchen scene, the $L_{\\mathrm{RT-MAE}}$ constraint achieves PSNR on par with 3DGS, despite using only one-third of the Gaussian points, as shown in the $1^{st}$ and $3^{rd}$ row in Tab. 5. Additionally, incorporating SSIM improves the PSNR as the number of Gaussians increases. Furthermore, to demonstrate the improvement of structure loss brought by RT-SSIM, we rendered the complete image and incorporated image-wise SSIM as a structural loss during random tile training. In contrast, the metrics using RT-SSIM consistently yield higher values, showing that the multi-view observations in RT-SSIM provide more accurate reconstruction constraints than image-wise SSIM.", + "bbox": [ + 89, + 506, + 482, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Adaptive Density Control. We ablate our densification strategy described in Sec. 3.3 on Tab. 6. Without any densification, the optimization of GS relies solely on the loss function without any splitting or cloning. Our improvements demonstrate the effectiveness of multi-view constraints in enhancing performance. Tile-Count densification, which averages the NDC gradient across tiles, causes the Gaussians to resist densifying and lower PSNR. This is because random-tile training increases the frequency with which a Gaussian is observed, leading to a smaller av", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "erage gradient.", + "bbox": [ + 511, + 92, + 614, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Tile Sampling and Batch Training. We analyzed the impact of viewpoint count $V$ and tile uniformity described in Sec. 3.2 on Tab. 7. Uniform sampling better distributes viewpoint constraints in the loss, yielding slight improvements and serving as our default. In data-rich scenes like Courthouse, more viewpoints provide greater benefits, as larger scenes require sufficient and evenly distributed observations. This highlights the adaptability of our random-tile training paradigm to different scene requirements. We also tested the comparative effect of image-wise batch training. Under a similar total training volume (iterations = 15k), introducing multiple images as a batch to incorporate multi-view observations does not improve quality as effectively as the tile-wise approach. Moreover, our observations indicate that such dense inclusion of multiple images significantly increases computational overhead.", + "bbox": [ + 511, + 106, + 903, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sensitivity to Initialization. The initialization problem has long been considered a critical factor impacting 3DGS performance [18]. We conduct experiments under two initialization conditions. Fig. 7 shows that our method recovers improved structure and appearance even with random initialization. As shown in Tab. 8, our method demonstrates greater robustness to initialization variations. This suggests that part of the initialization issue may arise from uneven and insufficient training constraints, and RT-Loss has the potential to address this limitation.", + "bbox": [ + 511, + 347, + 906, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Discussions and Limitations", + "text_level": 1, + "bbox": [ + 511, + 506, + 764, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a practical and effective approach by incorporating tile-wise constraints into Gaussian optimization to replace image-wise constraints. We have demonstrated its clear effectiveness and universality through both the main experiments and ablation studies. However, in experiments with large texture-less regions, our method fails to improve the geometry. To further enhance the tile-wise optimization, finer strategies could be explored in practice, such as further allocating multi-view constraints to identical content, in order to provide more accurate and uniform multi-view constraints for improvement.", + "bbox": [ + 511, + 527, + 906, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 705, + 633, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce a Tile-based training paradigm that integrates multi-view appearance and structural constraints. Extensive experiments demonstrate its effectiveness across various scenarios. We expect tile-wise training to hold strong potential as a promising paradigm for future research and applications in Gaussian Splatting.", + "bbox": [ + 511, + 731, + 905, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 833, + 671, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partially supported by Key R&D Program of Zhejiang Province (No. 2023C01039).", + "bbox": [ + 511, + 857, + 890, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26930", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 1", + "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 2, 5, 3", + "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697–19705, 2023. 5", + "[5] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16569-16578, 2023. 2", + "[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splattering with progressive propagation, 2024. 1", + "[7] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2", + "[8] Xiaobiao Du, Yida Wang, and Xin Yu. Mvgs: Multi-view-regulated gaussian splatting for novel view synthesis, 2024. 2", + "[9] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing, 2021. 2", + "[10] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. arXiv preprint arXiv:2402.03307, 2024. 2, 5, 6, 7", + "[11] Daniel Duckworth, Peter Hedman, Christian Reiser, Peter Zhizhin, Jean-François Thibert, Mario Lučić, Richard Szeliski, and Jonathan T Barron. Smerf: Streamable memory efficient radiance fields for real-time large-scene exploration. ACM Transactions on Graphics (TOG), 43(4): 1-13, 2024. 5, 6, 3", + "[12] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building rome on a cloudless day. In European Conference on Computer Vision, pages 368–381, 2010. 1", + "[13] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of" + ], + "bbox": [ + 93, + 114, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 5501-5510, 2022. 5, 3", + "[14] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2", + "[15] Xiang Guo, Guanying Chen, Yuchao Dai, Xiaqing Ye, Jiadai Sun, Xiao Tan, and Errui Ding. Neural deformable voxel grid for fast optimization of dynamic view synthesis, 2022. 2", + "[16] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 5, 3", + "[17] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf: Efficient neural radiance fields, 2022. 2", + "[18] Jaewoo Jung, Jisang Han, Honggyu An, Jiwon Kang, Seonghoon Park, and Seungryong Kim. Relaxing accurate initialization constraint for 3d gaussian splatting. arXiv preprint arXiv:2403.09413, 2024. 2, 8", + "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 5, 3", + "[20] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces, 2022. 2", + "[21] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 5, 3", + "[22] Christoph Lassner and Michael Zollhöfer. Pulsar: Efficient sphere-based neural rendering, 2020. 1", + "[23] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 6", + "[24] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering, 2024. 1, 2", + "[25] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2", + "[26] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splattering for real-time dynamic view synthesis, 2024. 2", + "[27] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 2", + "[28] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. High-fidelity and real-time novel view synthesis for dynamic scenes. In SIGGRAPH Asia Conference Proceedings, 2023. 2" + ], + "bbox": [ + 516, + 93, + 900, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26931", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763, 2023. 2", + "[30] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. arXiv preprint arXiv:2311.17061, 2023. 2", + "[31] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 1, 2, 5, 3", + "[32] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2", + "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2", + "[34] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2, 5, 3", + "[35] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2", + "[36] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1", + "[37] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2", + "[38] Dan Wang, Xinrui Cui, Septimiu Salcudean, and Z. Jane Wang. Generalizable neural radiance fields for novel view synthesis with transformer, 2022. 2", + "[39] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[40] Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. arXiv preprint arXiv:2403.16292, 2024. 2", + "[41] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Zeke Xie, Xindi Yang, Yujie Yang, Qi Sun, Yixiang Jiang, Haoran Wang, Yunfeng Cai, and Mingming Sun. S3im: Stochastic structural similarity and its unreasonable effectiveness for neural fields, 2023. 2, 3", + "[43] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Real-time 360sparse view synthesis using gaussian splatting, 2024. 2", + "[44] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, Dahua Lin, Michael Zollhöfer, and Christian Richardt. Vr-nerf: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia 2023 Conference Papers, page 1–12. ACM, 2023. 1", + "[45] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields, 2023. 2", + "[46] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. arXiv preprint arXiv:2409.14316, 2024. 2", + "[47] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splattering. arXiv preprint arXiv 2310.10642, 2023. 2", + "[48] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 2", + "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5", + "[50] Xiaoyu Zhang, Weihong Pan, Chong Bao, Xiyu Zhang, Xiaojun Xiang, Hanqing Jiang, and Hujun Bao. Lookcloser: Frequency-aware radiance field for tiny-detail scene. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16122-16132, 2025. 2", + "[51] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 1", + "[52] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. 2", + "[53] Zuo-Liang Zhu, Beibei Wang, and Jian Yang. Gs-ror: 3d gaussian splattering for reflective object relighting via sdf priors, 2024. 2", + "[54] C Lawrence Zitnick, Sing Bing Kang, Matthew Uysettadaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM TOG, 2004. 2", + "[55] Yiming Zuo and Jia Deng. View synthesis with sculpted neural points. arXiv preprint arXiv:2205.05869, 2022. 2" + ], + "bbox": [ + 516, + 92, + 905, + 883 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26932", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/e220fb2d-345d-4c10-96de-761db897dc9b_model.json b/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/e220fb2d-345d-4c10-96de-761db897dc9b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..199c206856206f7d14797bb2954b447b4db59d57 --- /dev/null +++ b/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/e220fb2d-345d-4c10-96de-761db897dc9b_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.13, + 0.861, + 0.178 + ], + "angle": 0, + "content": "Tile-wise vs. Image-wise: Random-Tile Loss and Training Paradigm for Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.202, + 0.874, + 0.259 + ], + "angle": 0, + "content": "Xiaoyu Zhang\\(^{1*}\\) Weihong Pan\\(^{2*}\\) Xiaojun Xiang\\(^{1}\\) Hongjia Zhai\\(^{2}\\) Liyang Zhou\\(^{1}\\) \n\\(^{1}\\)SenseTime Research \\({}^{2}\\)State Key Lab of CAD&CG, Zhejiang University" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.267, + 0.891, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.525, + 0.908, + 0.582 + ], + "angle": 0, + "content": "Figure 1. Random-Tile Loss integrates multi-view appearance and structural constraints during each optimization step. Under the same batch size, it consistently improves convergence efficiency and rendering quality while preserving compactness and real-time capability. The results on test views across different training iterations suggest a more stable and efficient convergence compared to the image-wise loss in Gaussian Splatting. Please refer to the Supplementary Video for more detailed comparisons on convergence performance." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.588, + 0.327, + 0.603 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.621, + 0.486, + 0.846 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) has drawn significant attention for its advantages in rendering speed and quality. Most existing methods still rely on the image-wise loss and training paradigm because of its intuitive nature in the Splatting algorithm. However, image-wise loss lacks multi-view constraints, which are generally essential for optimizing 3D appearance and geometry. To address this, we propose RT-Loss along with a tile-based training paradigm, which uses randomly sampled tiles to integrate multi-view appearance and structural constraints in 3DGS. Additionally, we introduce an tile-based adaptive densification control strategy tailored for our training paradigm. Extensive experiments show that our approach consistently improves performance metrics while maintaining efficiency across various benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.588, + 0.645, + 0.603 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.613, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF) [33] has brought revolutionary advances to photorealistic novel view synthesis (NVS) through exceptional rendering quality, attracting significant attention from the computer vision, graphics, and robotics communities [44]. Recently, 3D Gaussian Splating (3DGS) [19] has significantly enhanced rendering efficiency by replacing traditional ray marching with the splating algorithm [22]. Numerous studies have focused on optimizing the Gaussian feature or heuristic split algorithm in 3DGS to improve the appearance quality and geometric accuracy [6, 24, 31, 51]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "However, while the splatting algorithm significantly improves rendering speed, its potential for incorporating Tile-based multi-view constraints remains underexplored. Specifically, 3DGS initializes with point clouds derived from Structure from Motion (SfM) [1, 12, 36] and optimizes a set of anisotropic 3D Gaussians to represent the scene. Tile-based splatting plays a crucial role in achieving real-time rendering by directly splatting visible Gaussians onto" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.268, + 0.888 + ], + "angle": 0, + "content": "*Authors contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.268, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26923" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "the camera view, rather than querying every ray point. As a result, defining the loss function on each training image \\(\\mathcal{I}\\) is a intuitive and straightforward choice:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.15, + 0.483, + 0.169 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\Theta) = (1 - \\lambda) L _ {\\mathrm {M A E}} (\\Theta , \\mathcal {I}) + \\lambda L _ {\\mathrm {D - S S I M}} (\\Theta , \\mathcal {I}), \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.18, + 0.485, + 0.271 + ], + "angle": 0, + "content": "where \\( L_{MAE} \\) and \\( L_{D-SSIM} \\) are computed image-wise on the entire training image. Although this approach benefits the calculation of structural loss SSIM across the dense image, it also means that gradients from multiple viewpoints cannot be integrated in a single optimization step. In contrast, NeRF typically optimizes a pixel-wise loss:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.282, + 0.484, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\Theta) = \\frac {1}{\\| \\mathcal {R} \\|} \\sum_ {\\boldsymbol {r} \\in \\mathcal {R}} L _ {\\mathrm {M S E}} (\\Theta , \\boldsymbol {r}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.331, + 0.484, + 0.422 + ], + "angle": 0, + "content": "where each individual ray \\(r = (x, d, c)\\) in the mini-batch \\(\\mathcal{R}\\) is randomly sampled from the entire training dataset. Pointwise sampling from random viewpoints provides more comprehensive gradients. However, it fails to utilize the collective supervision of distant pixels, thus inevitably overlooking structural information in the overall scene [42]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.423, + 0.484, + 0.513 + ], + "angle": 0, + "content": "To address this issue, we propose a Tile-based training paradigm that integrates both multi-view constraints and structural information within the splatting process. We sample a set of tiles \\(\\mathcal{T}\\) from multiple training views, each individual tile \\(t\\) is the minimal unit in splatting. The loss is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.522, + 0.484, + 0.574 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R T} (\\Theta) = \\frac {1}{\\| \\mathcal {T} \\|} \\sum_ {\\boldsymbol {t} \\in \\mathcal {T}} L _ {\\mathrm {M A E}} (\\Theta , \\boldsymbol {t}) + \\lambda L _ {\\mathrm {R T - S S I M}} (\\Theta , \\mathcal {T}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.587, + 0.484, + 0.675 + ], + "angle": 0, + "content": "which is computed over a random-tile batch \\(\\mathcal{T}\\) that contains multi-view appearance and structure. Our Tile-based training paradigm maintains the training efficiency of the 3DGS rendering process, with detailed definitions and calculations provided in Sec. 3.2. Overall, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.678, + 0.483, + 0.769 + ], + "angle": 0, + "content": "- We propose a Tile-based training paradigm and loss formulation, incorporating both multi-view appearance constraints and structural similarity constraints with random tile sampling. To the best of our knowledge, we are the first to introduce a random-tile approach as a replacement for image-wise training paradigms in 3DGS." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.769, + 0.483, + 0.815 + ], + "angle": 0, + "content": "- Extensive experiments indicate that this Tile-based training paradigm yields notable improvements in training performance while maintaining model compactness." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.678, + 0.483, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.234, + 0.845 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Novel View Synthesis. Neural Radiance Field (NeRF) [33] has significantly advanced novel view synthesis by employing a multilayer perceptron (MLP) to model scenes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.319 + ], + "angle": 0, + "content": "with volumetric rendering, enabling the generation of realistic images from different perspectives. NeRF introduced a transformative approach to scene representation, sparking extensive research focused on enhancing rendering quality [2, 3, 27, 37, 50, 54, 55], speeding up computation [5, 14, 17, 34, 35], enabling 3D content generation [38], handling dynamic scenes [9, 15, 25], and improving generalization to novel scenes [45]. Due to the ray marching rendering technique, most NeRF-based methods are trained by utilizing batches of randomly sampled rays, with mean squared error (MSE) as the primary loss function. To integrate structural similarity, S3IM [42] enhances NeRF by incorporating patches of random rays in the loss calculation with a structural similarity index (SSIM), improving performance in tasks such as sparse view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.909, + 0.718 + ], + "angle": 0, + "content": "Gaussian Splatting Training Paradigm. 3D Gaussian Splatting (3DGS) has gained significant attention in novel view synthesis tasks due to its impressive performance. Using a rasterization-based rendering approach [20], 3DGS avoids the computational overhead of ray marching. With its high rendering speed and quality, 3DGS has found applications across various fields [48], including dynamic scene reconstruction [7, 10, 28, 29, 32, 41, 47], sparse view synthesis [43, 52], reflective object modeling [53], and geometric surface recovery [24, 30, 40]. Additional optimizations target specific issues, such as the geometric improvements seen in Scaffold-GS [31], which combines neural field models with Gaussian Splatting to achieve state-of-the-art surface accuracy. [18] introduces adaptive strategies to reduce dependency on SfM initialization. While most approaches adopt an image-wise training paradigm, some recent research explores multi-view constraints to enhance Gaussian Splatting. For example, [46] leverages multi-view stereo (MVS) to enhance initialization geometry, using multi-view consistency to optimize Gaussian geometry. [8, 10, 26] utilize batch accumulation to alleviate data pressure from the temporal dimension by rendering multiple images per iteration and accumulating gradients. However, this straightforward image-wise accumulation not only impacts performance but also fails to fully leverage the benefits of multiview constraints with the same batch size." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.75, + 0.612, + 0.765 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.782, + 0.655, + 0.796 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Gaussian Splatting. 3DGS [19] represents scenes using anisotropic 3D Gaussians, which retains the differential properties of volumetric representations while enabling efficient rendering through Tile-based rasterization. Each Gaussian is assigned a color \\(c\\), represented by spherical harmonics, and an opacity \\(\\alpha\\), which modulates \\(G(x)\\) during" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26924" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.429, + 0.107 + ], + "angle": 0, + "content": "blending. Its spatial distribution is then defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.115, + 0.483, + 0.149 + ], + "angle": 0, + "content": "\\[\nG (x) = \\exp \\left(- \\frac {1}{2} (x - \\mu) ^ {T} \\Sigma^ {- 1} (x - \\mu)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.156, + 0.483, + 0.231 + ], + "angle": 0, + "content": "where \\( x \\) denotes a spatial location in the 3D scene and \\( \\mu \\), initialized from Structure-from-Motion (SfM) points, specifies the Gaussian center. The covariance matrix \\( \\Sigma \\) is constructed using a scaling matrix \\( S \\) and rotation matrix \\( R \\) as \\( \\Sigma = RSS^T R^T \\) to ensure positive semi-definiteness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.232, + 0.484, + 0.353 + ], + "angle": 0, + "content": "Tile-based Rendering. Tile-based rendering is the key to achieving high-resolution and real-time performance in 3DGS. Unlike conventional volumetric methods, it projects 3D Gaussians into 2D Gaussians \\( G'(x) \\) on the image plane rather than relying on resource-intensive ray marching. The image is divided into tiles of size \\( 16 \\times 16 \\) to process all sorted 2D Gaussians in parallel. Each pixel within a tile applies \\( \\alpha \\)-blending:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.362, + 0.483, + 0.405 + ], + "angle": 0, + "content": "\\[\nC \\left(x ^ {\\prime}\\right) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {j}\\right), \\quad \\sigma_ {i} = \\alpha_ {i} G _ {i} ^ {\\prime} \\left(x ^ {\\prime}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.487 + ], + "angle": 0, + "content": "Here, \\( x \\) represents the queried pixel position, and \\( \\mathcal{N} \\) denotes the number of sorted 2D Gaussians linked to that tile. By leveraging the Tile-based differentiable rasterizer, all 3D Gaussians projected onto the training view can be directly optimized in an end-to-end manner." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.548 + ], + "angle": 0, + "content": "MAE, MSE, and SSIM. These three loss functions are widely used for image quality assessment. The Mean Absolute Error (MAE) and Mean Squared Error (MSE) are two methods for calculating RGB loss:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.556, + 0.483, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M A E} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| y _ {i} - \\hat {y} _ {i} \\right|, \\mathrm {M S E} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(y _ {i} - \\hat {y} _ {i}\\right) ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.615, + 0.483, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S S I M} (x, y) = \\frac {\\left(2 \\mu_ {x} \\mu_ {y} + C _ {1}\\right) \\left(2 \\sigma_ {x y} + C _ {2}\\right)}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + C _ {1}\\right) \\left(\\sigma_ {x} ^ {2} + \\sigma_ {y} ^ {2} + C _ {2}\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.795 + ], + "angle": 0, + "content": "where \\(\\mu_{x}\\) and \\(\\mu_{y}\\) are the means of the images \\(x\\) and \\(y\\), \\(\\sigma_x^2\\) and \\(\\sigma_y^2\\) are the variances, \\(\\sigma_{xy}\\) is the covariance between the images, and \\(C_1, C_2\\) are constants to stabilize the division when the denominator is small. MAE measures the average absolute error, treating all deviations equally. MSE squares each error before averaging, making it more sensitive to large errors and outliers. The Structural Similarity (SSIM) focuses on perceptual quality by comparing structural patterns, luminance, and contrast between images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.376, + 0.82 + ], + "angle": 0, + "content": "3.2. Tile-based Training via RT-Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce RT-Loss and our Tile-based training paradigm. Our goal is to incorporate random viewpoint observations within each training batch, while preserving structural information to enforce more comprehensive constraints. In contrast to the individual rays used" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.244 + ], + "angle": 0, + "content": "in NeRFs, Gaussian Splatting utilizes a Tile-based splatting process that projects Gaussians onto the entire view screen, rather than performing active ray marching. Hence, the smallest unit in the loss calculation must align with the splatting rendering and is set as a \\(16 \\times 16\\) tile \\(t\\). We load the datasets as tiles instead of images, randomly sample \\(N\\) tiles as a training batch \\(\\mathcal{T}\\) and obtain the rendered tiles set \\(\\mathcal{T}(\\hat{\\mathcal{C}})\\) along with their corresponding ground truth \\(\\mathcal{T}(\\mathcal{C})\\), where \\(\\hat{\\mathcal{C}} = \\{\\hat{\\mathcal{C}}(t) \\mid t \\in \\mathcal{T}\\}\\) and \\(\\mathcal{C} = \\{\\mathcal{C}(t) \\mid t \\in \\mathcal{T}\\}\\). We use MAE as the color loss, which can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.253, + 0.907, + 0.289 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {R T} - \\mathrm {M A E}} = \\frac {1}{\\| \\mathcal {T} \\|} \\sum_ {\\boldsymbol {t} \\in \\mathcal {T}} \\| \\hat {\\boldsymbol {C}} (\\boldsymbol {t}) - \\boldsymbol {C} (\\boldsymbol {t}) \\| _ {1}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.909, + 0.48 + ], + "angle": 0, + "content": "The tiles contain viewpoint's structural information. If we solely rely on rgb loss \\( L_{\\mathrm{MAE}} \\), the positional relationship contained within the tiles would be completely lost [42]. Inspired by S3IM's adaptation of SSIM on random rays, we propose RT-SSIM for 3DGS to retain the structural information embedded in random tiles. The concept is straightforward: while keeping the total batch size unchanged, we compute and average the SSIM across multiple independent and non-overlapping random tiles from different views, with kernel size and stride set to \\( K \\times K \\) (e.g., \\( 9 \\times 9 \\) for overall structtrue or \\( 3 \\times 3 \\) for fine detail) with a stride of \\( s \\). RT-SSIM can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.487, + 0.906, + 0.541 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R T - S S I M} (\\hat {\\mathcal {T}}, \\mathcal {T}) = \\frac {1}{N} \\sum_ {\\boldsymbol {t} \\in \\mathcal {T}, v \\in \\boldsymbol {V}} \\operatorname {S S I M} \\left(\\boldsymbol {t} ^ {(v)} (\\hat {\\mathcal {C}}), \\boldsymbol {t} ^ {(v)} (\\mathcal {C})\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.906, + 0.626 + ], + "angle": 0, + "content": "where \\(V\\) controls the number of views in all \\(N\\) tiles. In particular, when \\(V = N\\), it implies all tiles are selected from different views. As RT-SSIM positively correlated with image similarity, we define the structure loss \\(L_{\\mathrm{RT - SSIM}}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.637, + 0.905, + 0.656 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {R T} - \\mathrm {S S I M}} (\\Theta , \\mathcal {T}) = 1 - \\mathrm {R T} - \\mathrm {S S I M} (\\hat {\\mathcal {T}}, \\mathcal {T}) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.906, + 0.725 + ], + "angle": 0, + "content": "We note that tiles are particularly suitable for parallel computation, thus the additional computational cost of Tile-based training is minimal. The final hybrid loss \\(\\mathcal{L}_{\\mathrm{RT}}\\) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.738, + 0.905, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R T}} (\\Theta , \\mathcal {T}) = (1 - \\lambda_ {s}) L _ {\\mathrm {R T} - \\mathrm {M A E}} + \\lambda_ {s} L _ {\\mathrm {R T} - \\mathrm {S S I M}} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.813, + 0.779 + ], + "angle": 0, + "content": "where \\(\\lambda_{s}\\) balances structure and color losses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We present the pseudocode in Algorithm 1. In the standard 3DGS training paradigm, tiles are rendered in parallel using the same view parameters. In contrast, we sample tiles from multiple different views to integrate multi-view constraints. In our random-tile training paradigm, We introduce two additional hyperparameters, \\(\\lambda_{s}\\) and \\(V\\), where \\(\\lambda_{s}\\) controls the loss weight and \\(V\\) determines the number of viewpoints. When \\(V = N\\), each tile is sampled from a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26925" + } + ], + [ + { + "type": "algorithm", + "bbox": [ + 0.092, + 0.094, + 0.482, + 0.625 + ], + "angle": 0, + "content": "Algorithm 1: Tile-based Training via RT-loss \n1 Let \\(\\mathcal{A}\\) be an SGD training algorithm; \n2 while the termination criteria has not been satisfied do \n3 Sample a training batch \\(\\mathcal{T}\\) consisting of \\(N\\) tiles from \\(V\\) views in dataset \\(\\mathcal{D}\\); \n4 for \\(i = 1\\) to \\(V\\) do \n5 Randomly select tiles \\(\\mathcal{T}_i = [t_1, t_2 \\dots t_{V/N}]\\), where \\(t_j \\in \\mathcal{T}_i \\subseteq \\mathcal{T}\\); \n6 VisibilityCheck(); \n7 end \n8 for each tile \\(t\\) in \\(\\mathcal{T}\\) do \n9 Compute the forward splating result after \\(\\hat{\\mathcal{C}} = \\{\\hat{C}(t) \\mid t \\in \\mathcal{T}\\}\\); \n10 Obtain the ground-truth \\(\\mathcal{C} = \\{C(t) \\mid t \\in \\mathcal{T}\\}\\); \n11 end \n12 for \\(i = 1\\) to \\(V\\) do \n13 let rendered pixels \\(\\hat{\\mathcal{C}}\\) form the pixel patch \\(\\pmb{t}^{(v)}(\\hat{\\mathcal{C}})\\) in view \\(V_i\\); \n14 Compute SSIM \\((\\pmb{t}^{(v)}(\\hat{\\mathcal{C}}), \\pmb{t}^{(v)}(\\mathcal{C}))\\) with the given kernel \\(K \\times K\\); \n15 end \n16 Obtain the RT-SSIM loss \\(L_{\\mathrm{RT - SSIM}}(\\Theta, \\mathcal{T}) = 1 - \\frac{1}{N} \\sum \\mathrm{SSIM}(\\pmb{t}^{(v)}(\\hat{\\mathcal{C}}), \\pmb{t}^{(v)}(\\mathcal{C}))\\); \n17 Obtain the conventional \\(L_{MAE}\\) loss as color loss \\(L_{\\mathrm{RT - MAE}}(\\Theta) = \\frac{1}{||\\mathcal{T}||} \\sum_{t \\in \\mathcal{T}} ||\\hat{\\mathcal{C}}(t) - C(t)||_1\\); \n18 \\(\\mathcal{L}(\\Theta, \\mathcal{T}) = (1 - \\lambda_s)L_{\\mathrm{RT - MAE}}(\\Theta, \\mathcal{T}) + \\lambda_s L_{\\mathrm{RT - SSIM}}(\\Theta, \\mathcal{T})\\); \n19 Compute the gradient \\(\\nabla \\mathcal{L}(\\Theta)\\); \n20 Update model parameters \\(\\Theta\\) by \\(\\mathcal{A}\\) \n21 end" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.792 + ], + "angle": 0, + "content": "distinct view, and their visibility is computed in parallel during preprocessing to accelerate the rendering process. For a fair comparison, we set the total tiles count \\(N\\) to match that in image-wise training, ensuring a consistent batch size in gradient descent algorithm \\(\\mathcal{A}\\). Our random-tile strategy uniformly samples tiles across each view, balancing observations for each perspective. Our ablation study in Sec. 4.4 shows that this strategy achieves slightly better results than purely random selection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.408, + 0.819 + ], + "angle": 0, + "content": "3.3. Tile-based Adaptive Density Control" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In 3D Gaussian Splating, a Gaussian is split or cloned depending on whether the average cumulative 2D gradient on the NDC coordinates \\(\\sum \\| \\mathbf{g}_i\\|\\) across interval iterations exceeds a specified threshold \\(\\tau_{\\mathrm{densify}}\\), as shown in Fig. 2. During our Tile-based training, Gaussians in one iteration will" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.1, + 0.672, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.104, + 0.882, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.339, + 0.907, + 0.409 + ], + "angle": 0, + "content": "Figure 2. Image-wise vs Tile-wise Adaptive Density Control. In Tile-based training, NDC gradients for densification are computed as a weighted sum across all observations, balancing multi-view contributions, where \\( S1 \\) represents the area of the Gaussian projection onto the tile, and \\( S2 \\) is the total projected area." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.492 + ], + "angle": 0, + "content": "have multiple NDC gradients from different training views. Thus, there are two methods for computing density gradients: Tile-Count Densification and Iteration-Count Densification." + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.498, + 0.906, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\frac {\\sum_ {k = 1} ^ {\\text {I t e r}} \\sum_ {m = 1} ^ {M _ {k}} \\| \\mathbf {g} _ {k m} \\|}{\\sum_ {k = 1} ^ {\\text {I t e r}} M _ {k}} > \\tau_ {\\text {d e n s i f y}} \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.554, + 0.905, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\frac {\\sum_ {k = 1} ^ {\\text {I t e r}} \\sum_ {m = 1} ^ {M _ {k}} \\left(\\left\\| \\mathbf {g} _ {k m} \\right\\| \\times r _ {m}\\right)}{\\sum_ {k = 1} ^ {\\text {I t e r}} 1} > \\tau_ {\\text {d e n s i f y}}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.907, + 0.811 + ], + "angle": 0, + "content": "where \\( \\text{Iter} \\) represents the number of iterations in which a Gaussian participates in \\( \\alpha \\)-blending within each densification interval. In Tile-count Densification ( Eq. (12)), the NDC gradient is averaged over the number of observed tiles \\( M_{k} \\), while in Iteration-Count Densification ( Eq. (13)), it is averaged over the number of observed iterations. Additionally, in Iteration-Count Densification, the NDC gradient is weighted by \\( r_{m} \\), the ratio of the projected area shown in Fig. 2. To maintain consistency with the split parameters used in the original 3DGS and to emphasize the effect of RT-Loss, we default to using Iteration-Count for adaptive density control. Further details on NDC gradients and densification algorithm can be found in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.646, + 0.841 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.704, + 0.866 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Dataset and Metrics. We conduct a comprehensive evaluation on 26 publicly available scenes widely used" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "26926" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison to previous methods on benchmark datasets. We report results on three commonly used datasets. The best, second best, and third best results are denoted by red, orange, and yellow, respectively." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.13, + 0.849, + 0.262 + ], + "angle": 0, + "content": "
Dataset\nMethod | MetricsMip-NeRF360Tanks&TemplesDeep Blending
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Plenoxels [13]23.620.6700.44322.130.7620.33423.060.7950.510
Instant-NGP [34]26.430.7250.33922.940.7520.30423.620.7970.423
Mip-NeRF360 [3]29.230.8440.20723.120.7710.25029.400.9010.245
3D-GS [19]28.690.8700.18224.360.8310.21329.410.9030.243
Scaffold-GS [31]28.840.8480.22024.810.8400.20530.210.9060.254
Ours29.660.8920.17125.130.8520.19230.150.9110.224
Ours (+ Scaffold-GS)29.860.8850.20625.320.8570.18930.490.9130.231
" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.279, + 0.898, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.638, + 0.908, + 0.696 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison across benchmark datasets [3, 11, 21]. We highlight the differences in the image using patches for clearer visibility. Our approach consistently outperforms 3D-GS on these scenes, demonstrating clear advantages in challenging scenarios, such as fine-scale details (MIP360-ROOM(a), TANDT-BARN), thin geometry (SMERF-NYC(a), SMERF-NYC(b)), texture-less regions (TANDT-TRAIN), light effects (MIP360-COUNTER, TANDT-TRUCK), and insufficient observations (TANDT-BARN, SMERF-NYC(b))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.486, + 0.89 + ], + "angle": 0, + "content": "in the field of 3D reconstruction across multiple diverse tasks. Specifically, we evaluate our method on seven scenes from Mip-NeRF360 [3], another seven from Tanks&Temples [21], and two additional scenes from DeepBlending [16]. To further demonstrate the limitations of image-wise loss in capturing global structure, our evaluation incorporated all four scenes from [4], with \\(5 - 10 \\times\\) more training views. We additionally evaluated dynamic benchmarks, specifically using the Neural 3D Video Dataset containing six scenes, to demonstrate the effectiveness of RT-Loss in 4D reconstruction. In addition to the commonly used metrics (PSNR, SSIM [39], and LPIPS [49]), we fur" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.909, + 0.8 + ], + "angle": 0, + "content": "ther provide the storage size (MB) and rendering speed (FPS) to assess model compactness and efficiency, with all experiments performed on an NVIDIA RTX 4090 GPU. The main paper reports the average metrics across all scenes in each dataset, while detailed per-scene quantitative results are presented in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Baseline and Implementation. 3DGS[19], Scaffold-GS[31] and 4D-GS [10] are selected as our baselines due to their state-of-the-art performance. We also present the results of Mip-NeRF360 [3], iNGP [34] and Plenoxels [13] as in [19]. To demonstrate the generality of the tile-based ap" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26927" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.368, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.09, + 0.627, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.091, + 0.891, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.297, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 4. Convergence performance. We evaluated the convergence of PSNR for both the training and testing views during the training process. With the same batch size, the convergence of image-wise loss in 3DGS is slower than RT-Loss, with the gap widening on larger datasets, e.g., COURTHOUSE and BERLIN" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.365, + 0.485, + 0.548 + ], + "angle": 0, + "content": "proach, we implemented the tile-wise method in Scaffolds for comparison. We trained for \\(30k\\) iterations and rendered the same number of pixels per iteration, ensuring the total pixels from random tiles matched a complete image. To independently demonstrate the effectiveness of the tile-based training paradigm and RT-loss, we modified only the sampling and loss calculations for random tile training while keeping the default 3DGS parameters unchanged, including the densification threshold \\(\\tau_{\\mathrm{densify}} = 0.0002\\) and all learning rates. For RT-loss, we sample random tiles from \\(V = 5\\) views and apply the same structure loss weight, \\(\\lambda_s = 0.2\\), as used in the original 3DGS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.56, + 0.254, + 0.575 + ], + "angle": 0, + "content": "4.2. Results Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Benchmark Evaluation. The quantitative results on Mip-NeRF360 and Tanks&Temples are presented in Tab. 1. Results show that we outperform existing methods on standard datasets. Fig. 3 illustrates that tile-wise training achieves enhanced visual quality, with improved geometry accuracy and richer texture details. Moreover, our method retains the efficiency of Gaussian methods, as shown in Tab. 2. We compared the rendering speed and storage size with 3DGS. We achieve improved performance while using a comparable or fewer number of Gaussians, demonstrating high-quality reconstruction without compromising speed. By incorporating our tile-based approach, Scaffold-GS also achieved notable improvements, suggesting that the image-wise training paradigm, which lacks multi-view constraints, limits the reconstruction quality of various GS methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Convergence Performance. We evaluated our convergence performance across several representative scenarios. As shown in Fig. 4, compared to the original image-wise loss in 3DGS, the RT-Loss based on multi-view observations provides more effective constraints, resulting in significantly faster convergence. The difference in convergence speed is" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.364, + 0.907, + 0.391 + ], + "angle": 0, + "content": "Table 2. Rendering FPS and storage comparison. The rendering speed of both methods is measured on the same machine." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.403, + 0.905, + 0.512 + ], + "angle": 0, + "content": "
ScenesMip-NeRF360ROOMKITCHEN
MemFPSPSNRMemFPSPSNRMemFPS
3DGS6717831.7931712331.63372118
Ours6259332.8131812532.35320135
ScenesTanks&TemplesCourthOUSEMEETINGROOM
MemFPSPSNRMemFPSPSNRMemFPS
3DGS31712623.1317414626.27276120
Ours30213024.1416813926.98260133
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.527, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Table 3. PSNR evaluation for scenes with increased training views. The convergence capability of RT-Loss is particularly evident on the SMERF dataset[11], which includes approximately 10 times more training views than Mipnerf360. Additionally, we recorded the PSNR gap at 4k and 30k iterations." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.606, + 0.903, + 0.67 + ], + "angle": 0, + "content": "
Scene.views test iterationAlameda(1734)Berlin(1511)London(1874)
4k30k4k30k4k30k
3DGS17.6222.2722.6327.5322.3225.75
Ours20.1123.91624.8028.7924.0726.626
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.685, + 0.907, + 0.741 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison on dynamic datasets. We applied RT loss for dynamic scene reconstruction and compared it with 4D-GS[10] on the challenging real-world dynamic dataset Neu3D[23]." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.749, + 0.903, + 0.798 + ], + "angle": 0, + "content": "
ModelPSNR↑D-SSIM↓LPIPS↓FPS↑Storage (MB)↓
4D-GS[10]31.150.9740.0493090
Ours31.740.9730.0433478
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.87 + ], + "angle": 0, + "content": "visually evident in Fig. 1. This advantage in convergence is even more pronounced in scenes with a large number of training views, e.g., COURTHOUSE and BERLIN." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Scene with More Training Views. We compared our method with the original 3DGS in scenarios with more" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26928" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.481, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.255, + 0.483, + 0.311 + ], + "angle": 0, + "content": "Figure 5. Gaussians structure comparison. RT-Loss demonstrates stronger structural reconstruction ability in scenes with more training views. In contrast, 3DGS not only loses details but also fails to recover the geometric structure of the scene." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.33, + 0.48, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.494, + 0.483, + 0.536 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons on the Neural 3D Video Dataset. In 4D reconstruction, constraints from multiple viewpoints and time also help generate more accurate details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.567, + 0.483, + 0.838 + ], + "angle": 0, + "content": "training views. In Tab. 3, our method achieved a significant improvement in metrics. The gap is particularly noticeable in the early training stages (4k iterations), as image-wise training naturally requires more iterations to fully capture the scene. Fig. 5 further illustrates that our method produces a more refined geometric structure on large datasets. This supports our hypothesis about the current training paradigm: as the number of training views increases, image-wise training struggles to comprehensively capture scene information, resulting in less reliable optimization and densification. Furthermore, through experiments with extended training durations, we demonstrate that naively increasing the iteration count for image-wise training does not effectively address this limitation. This is because once incorrect geometry is formed during Gaussian optimization, it is difficult to correct by simply extending the training process. For further details, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Dynamic Scenes. We evaluated our method on the real-world dynamic scene dataset, the Neural 3D Video Dataset, using 4D-GS [10] as the baseline. The expansion of the time dimension in the dynamic dataset resulted in a larger" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 5. Ablation study on RT-Loss components. 'Img-' refers to image-wise loss, and 'RT-' refers to random-tile loss." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.126, + 0.9, + 0.216 + ], + "angle": 0, + "content": "
Scene(views)KITCHEN(279)CORTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
LImg-MAE + LImg-SSIM (3DGS)31.6737223.11173
LImg-MAE31.0913121.3544
LRT-MAE31.6512722.6363
LRT-MAE + LImg-SSIM32.0235223.50186
LRT-MAE + LRT-SSIM (Ours)32.3532024.15168
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.23, + 0.907, + 0.272 + ], + "angle": 0, + "content": "Table 6. Effect of Adaptive Density Control. 'Tile-Count' and 'Iteration-Count' are densification strategies described in Sec. 3.3, with 'Iteration-Count' being the default." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.28, + 0.882, + 0.354 + ], + "angle": 0, + "content": "
Scene(views)KICHEN(279)COURTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
3DGS w/o Densification29.845822.96118
Ours w/o Densification30.325823.83118
Ours w/ Tile-Count31.9517523.56125
Ours w/ Iteration-Count32.3532024.15168
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.368, + 0.907, + 0.438 + ], + "angle": 0, + "content": "Table 7. Effect of Tile Sampling and Batch Training. 'V' refers to the number of randomly selected viewpoints, determining the spread of observations. 'B' refers to Batch train, which means using multiple images as a batch for training. 'Non-Uniform' and 'Uniform' indicate whether the tiles are distributed evenly." + }, + { + "type": "table", + "bbox": [ + 0.535, + 0.446, + 0.882, + 0.558 + ], + "angle": 0, + "content": "
Scene.views)KICHEN(279)COURTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
tile-wise V = 231.8034323.76182
tile-wise V = 432.2233723.92175
tile-wise V = 632.4033224.29161
image-wise B = 231.6735223.82214
image-wise B = 431.2234823.69207
Non-Uniform Sample32.3031024.02192
Uniform Sample32.3832024.14168
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.794 + ], + "angle": 0, + "content": "volume of training data. Although most 4D Gaussian methods address this issue by using batch training and rendering multiple images together as a training batch, our results show that, with the same batch size, tile-based training provides more effective constraints across multiple viewpoints and time frames. The results show that RT-Loss is also effective for dynamic scene reconstruction. As depicted in Tab. 4, we notably improved the rendering performance on dynamic scenes without sacrificing efficiency. As shown in Fig. 6, our method reconstructs finer details, attributed to our randomized approach, which provides stronger constraints across both viewpoints and time frames compared to image-wise training. Details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.68, + 0.818 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Effect of RT-Loss components. We conducted an ablation on the loss components of Eq. (11), which includes the base color loss \\( L_{\\mathrm{RT - MAE}} \\) and the structural loss \\( L_{\\mathrm{RT - SSIM}} \\). As shown in the \\( 2^{nd} \\) and \\( 3^{rd} \\) row in Tab. 5, when only MAE loss is used, the PSNR significantly decreases, with a lower" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26929" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.096, + 0.223, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.097, + 0.354, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.097, + 0.479, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.329 + ], + "angle": 0, + "content": "Figure 7. Effect of initialization on reconstruction quality. (Left) Gaussians optimized from SfM point initialization. (Right) Gaussians optimized from random points. Our method reconstructs better structures and finer details even with random initialization, demonstrating superior robustness to initialization issues." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.346, + 0.483, + 0.402 + ], + "angle": 0, + "content": "Table 8. Sensitivity to Initialization. In the 'w/o SfM initialization' setting, Gaussians are initialized with random positions. RT-Loss demonstrates stronger robustness in addressing initialization issues." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.411, + 0.48, + 0.49 + ], + "angle": 0, + "content": "
Scene(views)KICHEN(279)COURTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
3DGS w/ SfM initialization31.6337223.13174
3DGS w/o SfM initialization29.8926921.77121
Ours w/ SfM initialization32.3532024.14168
Ours w/o SfM initialization31.2227223.21186
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.748 + ], + "angle": 0, + "content": "tendency for Gaussian densification. However, our \\( L_{\\mathrm{RT-MAE}} \\) still outperforms the image-wise constraints, and this effect becomes more pronounced in the CourtHouse scene with more training views, where the PSNR difference reaches up to 1.3 dB. Notably, in the Kitchen scene, the \\( L_{\\mathrm{RT-MAE}} \\) constraint achieves PSNR on par with 3DGS, despite using only one-third of the Gaussian points, as shown in the \\( 1^{st} \\) and \\( 3^{rd} \\) row in Tab. 5. Additionally, incorporating SSIM improves the PSNR as the number of Gaussians increases. Furthermore, to demonstrate the improvement of structure loss brought by RT-SSIM, we rendered the complete image and incorporated image-wise SSIM as a structural loss during random tile training. In contrast, the metrics using RT-SSIM consistently yield higher values, showing that the multi-view observations in RT-SSIM provide more accurate reconstruction constraints than image-wise SSIM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effect of Adaptive Density Control. We ablate our densification strategy described in Sec. 3.3 on Tab. 6. Without any densification, the optimization of GS relies solely on the loss function without any splitting or cloning. Our improvements demonstrate the effectiveness of multi-view constraints in enhancing performance. Tile-Count densification, which averages the NDC gradient across tiles, causes the Gaussians to resist densifying and lower PSNR. This is because random-tile training increases the frequency with which a Gaussian is observed, leading to a smaller av" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.093, + 0.616, + 0.106 + ], + "angle": 0, + "content": "erage gradient." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.348 + ], + "angle": 0, + "content": "Effect of Tile Sampling and Batch Training. We analyzed the impact of viewpoint count \\( V \\) and tile uniformity described in Sec. 3.2 on Tab. 7. Uniform sampling better distributes viewpoint constraints in the loss, yielding slight improvements and serving as our default. In data-rich scenes like Courthouse, more viewpoints provide greater benefits, as larger scenes require sufficient and evenly distributed observations. This highlights the adaptability of our random-tile training paradigm to different scene requirements. We also tested the comparative effect of image-wise batch training. Under a similar total training volume (iterations = 15k), introducing multiple images as a batch to incorporate multi-view observations does not improve quality as effectively as the tile-wise approach. Moreover, our observations indicate that such dense inclusion of multiple images significantly increases computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Sensitivity to Initialization. The initialization problem has long been considered a critical factor impacting 3DGS performance [18]. We conduct experiments under two initialization conditions. Fig. 7 shows that our method recovers improved structure and appearance even with random initialization. As shown in Tab. 8, our method demonstrates greater robustness to initialization variations. This suggests that part of the initialization issue may arise from uneven and insufficient training constraints, and RT-Loss has the potential to address this limitation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.507, + 0.766, + 0.521 + ], + "angle": 0, + "content": "4.4. Discussions and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.907, + 0.695 + ], + "angle": 0, + "content": "We propose a practical and effective approach by incorporating tile-wise constraints into Gaussian optimization to replace image-wise constraints. We have demonstrated its clear effectiveness and universality through both the main experiments and ablation studies. However, in experiments with large texture-less regions, our method fails to improve the geometry. To further enhance the tile-wise optimization, finer strategies could be explored in practice, such as further allocating multi-view constraints to identical content, in order to provide more accurate and uniform multi-view constraints for improvement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.707, + 0.634, + 0.722 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.906, + 0.824 + ], + "angle": 0, + "content": "In this work, we introduce a Tile-based training paradigm that integrates multi-view appearance and structural constraints. Extensive experiments demonstrate its effectiveness across various scenarios. We expect tile-wise training to hold strong potential as a promising paradigm for future research and applications in Gaussian Splatting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.673, + 0.851 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.858, + 0.892, + 0.888 + ], + "angle": 0, + "content": "This work was partially supported by Key R&D Program of Zhejiang Province (No. 2023C01039)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26930" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.17 + ], + "angle": 0, + "content": "[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.48, + 0.227 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.471, + 0.299 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 2, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.468, + 0.37 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697–19705, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.473, + 0.455 + ], + "angle": 0, + "content": "[5] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16569-16578, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splattering with progressive propagation, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.468, + 0.556 + ], + "angle": 0, + "content": "[7] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.45, + 0.599 + ], + "angle": 0, + "content": "[8] Xiaobiao Du, Yida Wang, and Xin Yu. Mvgs: Multi-view-regulated gaussian splatting for novel view synthesis, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.458, + 0.643 + ], + "angle": 0, + "content": "[9] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.478, + 0.7 + ], + "angle": 0, + "content": "[10] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. arXiv preprint arXiv:2402.03307, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.475, + 0.785 + ], + "angle": 0, + "content": "[11] Daniel Duckworth, Peter Hedman, Christian Reiser, Peter Zhizhin, Jean-François Thibert, Mario Lučić, Richard Szeliski, and Jonathan T Barron. Smerf: Streamable memory efficient radiance fields for real-time large-scene exploration. ACM Transactions on Graphics (TOG), 43(4): 1-13, 2024. 5, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.48, + 0.857 + ], + "angle": 0, + "content": "[12] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building rome on a cloudless day. In European Conference on Computer Vision, pages 368–381, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.479, + 0.901 + ], + "angle": 0, + "content": "[13] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 5501-5510, 2022. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.895, + 0.189 + ], + "angle": 0, + "content": "[14] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.897, + 0.243 + ], + "angle": 0, + "content": "[15] Xiang Guo, Guanying Chen, Yuchao Dai, Xiaqing Ye, Jiadai Sun, Xiao Tan, and Errui Ding. Neural deformable voxel grid for fast optimization of dynamic view synthesis, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.891, + 0.3 + ], + "angle": 0, + "content": "[16] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.891, + 0.328 + ], + "angle": 0, + "content": "[17] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf: Efficient neural radiance fields, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.895, + 0.385 + ], + "angle": 0, + "content": "[18] Jaewoo Jung, Jisang Han, Honggyu An, Jiwon Kang, Seonghoon Park, and Seungryong Kim. Relaxing accurate initialization constraint for 3d gaussian splatting. arXiv preprint arXiv:2403.09413, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.9, + 0.44 + ], + "angle": 0, + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.882, + 0.469 + ], + "angle": 0, + "content": "[20] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.886, + 0.524 + ], + "angle": 0, + "content": "[21] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.898, + 0.552 + ], + "angle": 0, + "content": "[22] Christoph Lassner and Michael Zollhöfer. Pulsar: Efficient sphere-based neural rendering, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.897, + 0.636 + ], + "angle": 0, + "content": "[23] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.9, + 0.678 + ], + "angle": 0, + "content": "[24] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.68, + 0.89, + 0.747 + ], + "angle": 0, + "content": "[25] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.878, + 0.788 + ], + "angle": 0, + "content": "[26] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splattering for real-time dynamic view synthesis, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.901, + 0.845 + ], + "angle": 0, + "content": "[27] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.898, + 0.9 + ], + "angle": 0, + "content": "[28] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. High-fidelity and real-time novel view synthesis for dynamic scenes. In SIGGRAPH Asia Conference Proceedings, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.901, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26931" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.48, + 0.147 + ], + "angle": 0, + "content": "[29] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.478, + 0.219 + ], + "angle": 0, + "content": "[30] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. arXiv preprint arXiv:2311.17061, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.477, + 0.29 + ], + "angle": 0, + "content": "[31] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 1, 2, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.457, + 0.346 + ], + "angle": 0, + "content": "[32] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.473, + 0.416 + ], + "angle": 0, + "content": "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.473, + 0.473 + ], + "angle": 0, + "content": "[34] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2, 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.475, + 0.545 + ], + "angle": 0, + "content": "[35] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.45, + 0.602 + ], + "angle": 0, + "content": "[36] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.468, + 0.687 + ], + "angle": 0, + "content": "[37] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[38] Dan Wang, Xinrui Cui, Septimiu Salcudean, and Z. Jane Wang. Generalizable neural radiance fields for novel view synthesis with transformer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.475, + 0.787 + ], + "angle": 0, + "content": "[39] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[40] Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. arXiv preprint arXiv:2403.16292, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[41] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[42] Zeke Xie, Xindi Yang, Yujie Yang, Qi Sun, Yixiang Jiang, Haoran Wang, Yunfeng Cai, and Mingming Sun. S3im: Stochastic structural similarity and its unreasonable effectiveness for neural fields, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.871, + 0.205 + ], + "angle": 0, + "content": "[43] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Real-time 360sparse view synthesis using gaussian splatting, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.903, + 0.289 + ], + "angle": 0, + "content": "[44] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, Dahua Lin, Michael Zollhöfer, and Christian Richardt. Vr-nerf: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia 2023 Conference Papers, page 1–12. ACM, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[45] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.9, + 0.388 + ], + "angle": 0, + "content": "[46] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. arXiv preprint arXiv:2409.14316, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.887, + 0.445 + ], + "angle": 0, + "content": "[47] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splattering. arXiv preprint arXiv 2310.10642, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.906, + 0.514 + ], + "angle": 0, + "content": "[48] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.906, + 0.585 + ], + "angle": 0, + "content": "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[50] Xiaoyu Zhang, Weihong Pan, Chong Bao, Xiyu Zhang, Xiaojun Xiang, Hanqing Jiang, and Hujun Bao. Lookcloser: Frequency-aware radiance field for tiny-detail scene. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16122-16132, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.861, + 0.713 + ], + "angle": 0, + "content": "[51] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.901, + 0.756 + ], + "angle": 0, + "content": "[52] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.885, + 0.798 + ], + "angle": 0, + "content": "[53] Zuo-Liang Zhu, Beibei Wang, and Jian Yang. Gs-ror: 3d gaussian splattering for reflective object relighting via sdf priors, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.8, + 0.866, + 0.854 + ], + "angle": 0, + "content": "[54] C Lawrence Zitnick, Sing Bing Kang, Matthew Uysettadaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM TOG, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.856, + 0.884, + 0.884 + ], + "angle": 0, + "content": "[55] Yiming Zuo and Jia Deng. View synthesis with sculpted neural points. arXiv preprint arXiv:2205.05869, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26932" + } + ] +] \ No newline at end of file diff --git a/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/e220fb2d-345d-4c10-96de-761db897dc9b_origin.pdf b/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/e220fb2d-345d-4c10-96de-761db897dc9b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8a30101f64ea993d8b400f77f176338bd6da67b7 --- /dev/null +++ b/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/e220fb2d-345d-4c10-96de-761db897dc9b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b23e0f19d49e5c25917e6be99720709e8e6bdac8428778e80b7703a13a95f4db +size 6912059 diff --git a/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/full.md b/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fa6c55d5b50a9b91f7cefa5902ba8ae92598967c --- /dev/null +++ b/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/full.md @@ -0,0 +1,335 @@ +# Tile-wise vs. Image-wise: Random-Tile Loss and Training Paradigm for Gaussian Splatting + +Xiaoyu Zhang $^{1*}$ Weihong Pan $^{2*}$ Xiaojun Xiang $^{1}$ Hongjia Zhai $^{2}$ Liyang Zhou $^{1}$ $^{1}$ SenseTime Research ${}^{2}$ State Key Lab of CAD&CG, Zhejiang University + +![](images/ebcf5affed1d17c580fa56a2a3e5f1db03ea669d7f7ca3838945301ef832f074.jpg) +Figure 1. Random-Tile Loss integrates multi-view appearance and structural constraints during each optimization step. Under the same batch size, it consistently improves convergence efficiency and rendering quality while preserving compactness and real-time capability. The results on test views across different training iterations suggest a more stable and efficient convergence compared to the image-wise loss in Gaussian Splatting. Please refer to the Supplementary Video for more detailed comparisons on convergence performance. + +# Abstract + +3D Gaussian Splatting (3DGS) has drawn significant attention for its advantages in rendering speed and quality. Most existing methods still rely on the image-wise loss and training paradigm because of its intuitive nature in the Splatting algorithm. However, image-wise loss lacks multi-view constraints, which are generally essential for optimizing 3D appearance and geometry. To address this, we propose RT-Loss along with a tile-based training paradigm, which uses randomly sampled tiles to integrate multi-view appearance and structural constraints in 3DGS. Additionally, we introduce an tile-based adaptive densification control strategy tailored for our training paradigm. Extensive experiments show that our approach consistently improves performance metrics while maintaining efficiency across various benchmark datasets. + +# 1. Introduction + +Neural Radiance Fields (NeRF) [33] has brought revolutionary advances to photorealistic novel view synthesis (NVS) through exceptional rendering quality, attracting significant attention from the computer vision, graphics, and robotics communities [44]. Recently, 3D Gaussian Splating (3DGS) [19] has significantly enhanced rendering efficiency by replacing traditional ray marching with the splating algorithm [22]. Numerous studies have focused on optimizing the Gaussian feature or heuristic split algorithm in 3DGS to improve the appearance quality and geometric accuracy [6, 24, 31, 51]. + +However, while the splatting algorithm significantly improves rendering speed, its potential for incorporating Tile-based multi-view constraints remains underexplored. Specifically, 3DGS initializes with point clouds derived from Structure from Motion (SfM) [1, 12, 36] and optimizes a set of anisotropic 3D Gaussians to represent the scene. Tile-based splatting plays a crucial role in achieving real-time rendering by directly splatting visible Gaussians onto + +the camera view, rather than querying every ray point. As a result, defining the loss function on each training image $\mathcal{I}$ is a intuitive and straightforward choice: + +$$ +\mathcal {L} (\Theta) = (1 - \lambda) L _ {\mathrm {M A E}} (\Theta , \mathcal {I}) + \lambda L _ {\mathrm {D - S S I M}} (\Theta , \mathcal {I}), \quad (1) +$$ + +where $L_{MAE}$ and $L_{D-SSIM}$ are computed image-wise on the entire training image. Although this approach benefits the calculation of structural loss SSIM across the dense image, it also means that gradients from multiple viewpoints cannot be integrated in a single optimization step. In contrast, NeRF typically optimizes a pixel-wise loss: + +$$ +\mathcal {L} (\Theta) = \frac {1}{\| \mathcal {R} \|} \sum_ {\boldsymbol {r} \in \mathcal {R}} L _ {\mathrm {M S E}} (\Theta , \boldsymbol {r}), \tag {2} +$$ + +where each individual ray $r = (x, d, c)$ in the mini-batch $\mathcal{R}$ is randomly sampled from the entire training dataset. Pointwise sampling from random viewpoints provides more comprehensive gradients. However, it fails to utilize the collective supervision of distant pixels, thus inevitably overlooking structural information in the overall scene [42]. + +To address this issue, we propose a Tile-based training paradigm that integrates both multi-view constraints and structural information within the splatting process. We sample a set of tiles $\mathcal{T}$ from multiple training views, each individual tile $t$ is the minimal unit in splatting. The loss is formulated as follows: + +$$ +\mathcal {L} _ {R T} (\Theta) = \frac {1}{\| \mathcal {T} \|} \sum_ {\boldsymbol {t} \in \mathcal {T}} L _ {\mathrm {M A E}} (\Theta , \boldsymbol {t}) + \lambda L _ {\mathrm {R T - S S I M}} (\Theta , \mathcal {T}), \tag {3} +$$ + +which is computed over a random-tile batch $\mathcal{T}$ that contains multi-view appearance and structure. Our Tile-based training paradigm maintains the training efficiency of the 3DGS rendering process, with detailed definitions and calculations provided in Sec. 3.2. Overall, our contributions are as follows: + +- We propose a Tile-based training paradigm and loss formulation, incorporating both multi-view appearance constraints and structural similarity constraints with random tile sampling. To the best of our knowledge, we are the first to introduce a random-tile approach as a replacement for image-wise training paradigms in 3DGS. +- Extensive experiments indicate that this Tile-based training paradigm yields notable improvements in training performance while maintaining model compactness. + +# 2. Related Work + +Novel View Synthesis. Neural Radiance Field (NeRF) [33] has significantly advanced novel view synthesis by employing a multilayer perceptron (MLP) to model scenes + +with volumetric rendering, enabling the generation of realistic images from different perspectives. NeRF introduced a transformative approach to scene representation, sparking extensive research focused on enhancing rendering quality [2, 3, 27, 37, 50, 54, 55], speeding up computation [5, 14, 17, 34, 35], enabling 3D content generation [38], handling dynamic scenes [9, 15, 25], and improving generalization to novel scenes [45]. Due to the ray marching rendering technique, most NeRF-based methods are trained by utilizing batches of randomly sampled rays, with mean squared error (MSE) as the primary loss function. To integrate structural similarity, S3IM [42] enhances NeRF by incorporating patches of random rays in the loss calculation with a structural similarity index (SSIM), improving performance in tasks such as sparse view synthesis. + +Gaussian Splatting Training Paradigm. 3D Gaussian Splatting (3DGS) has gained significant attention in novel view synthesis tasks due to its impressive performance. Using a rasterization-based rendering approach [20], 3DGS avoids the computational overhead of ray marching. With its high rendering speed and quality, 3DGS has found applications across various fields [48], including dynamic scene reconstruction [7, 10, 28, 29, 32, 41, 47], sparse view synthesis [43, 52], reflective object modeling [53], and geometric surface recovery [24, 30, 40]. Additional optimizations target specific issues, such as the geometric improvements seen in Scaffold-GS [31], which combines neural field models with Gaussian Splatting to achieve state-of-the-art surface accuracy. [18] introduces adaptive strategies to reduce dependency on SfM initialization. While most approaches adopt an image-wise training paradigm, some recent research explores multi-view constraints to enhance Gaussian Splatting. For example, [46] leverages multi-view stereo (MVS) to enhance initialization geometry, using multi-view consistency to optimize Gaussian geometry. [8, 10, 26] utilize batch accumulation to alleviate data pressure from the temporal dimension by rendering multiple images per iteration and accumulating gradients. However, this straightforward image-wise accumulation not only impacts performance but also fails to fully leverage the benefits of multiview constraints with the same batch size. + +# 3. Methods + +# 3.1. Preliminaries + +Gaussian Splatting. 3DGS [19] represents scenes using anisotropic 3D Gaussians, which retains the differential properties of volumetric representations while enabling efficient rendering through Tile-based rasterization. Each Gaussian is assigned a color $c$ , represented by spherical harmonics, and an opacity $\alpha$ , which modulates $G(x)$ during + +blending. Its spatial distribution is then defined by: + +$$ +G (x) = \exp \left(- \frac {1}{2} (x - \mu) ^ {T} \Sigma^ {- 1} (x - \mu)\right), \tag {4} +$$ + +where $x$ denotes a spatial location in the 3D scene and $\mu$ , initialized from Structure-from-Motion (SfM) points, specifies the Gaussian center. The covariance matrix $\Sigma$ is constructed using a scaling matrix $S$ and rotation matrix $R$ as $\Sigma = RSS^T R^T$ to ensure positive semi-definiteness. + +Tile-based Rendering. Tile-based rendering is the key to achieving high-resolution and real-time performance in 3DGS. Unlike conventional volumetric methods, it projects 3D Gaussians into 2D Gaussians $G'(x)$ on the image plane rather than relying on resource-intensive ray marching. The image is divided into tiles of size $16 \times 16$ to process all sorted 2D Gaussians in parallel. Each pixel within a tile applies $\alpha$ -blending: + +$$ +C \left(x ^ {\prime}\right) = \sum_ {i \in \mathcal {N}} c _ {i} \sigma_ {i} \prod_ {j = 1} ^ {i - 1} \left(1 - \sigma_ {j}\right), \quad \sigma_ {i} = \alpha_ {i} G _ {i} ^ {\prime} \left(x ^ {\prime}\right). \tag {5} +$$ + +Here, $x$ represents the queried pixel position, and $\mathcal{N}$ denotes the number of sorted 2D Gaussians linked to that tile. By leveraging the Tile-based differentiable rasterizer, all 3D Gaussians projected onto the training view can be directly optimized in an end-to-end manner. + +MAE, MSE, and SSIM. These three loss functions are widely used for image quality assessment. The Mean Absolute Error (MAE) and Mean Squared Error (MSE) are two methods for calculating RGB loss: + +$$ +\mathrm {M A E} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left| y _ {i} - \hat {y} _ {i} \right|, \mathrm {M S E} = \frac {1}{n} \sum_ {i = 1} ^ {n} \left(y _ {i} - \hat {y} _ {i}\right) ^ {2}, \tag {6} +$$ + +$$ +\operatorname {S S I M} (x, y) = \frac {\left(2 \mu_ {x} \mu_ {y} + C _ {1}\right) \left(2 \sigma_ {x y} + C _ {2}\right)}{\left(\mu_ {x} ^ {2} + \mu_ {y} ^ {2} + C _ {1}\right) \left(\sigma_ {x} ^ {2} + \sigma_ {y} ^ {2} + C _ {2}\right)}, \tag {7} +$$ + +where $\mu_{x}$ and $\mu_{y}$ are the means of the images $x$ and $y$ , $\sigma_x^2$ and $\sigma_y^2$ are the variances, $\sigma_{xy}$ is the covariance between the images, and $C_1, C_2$ are constants to stabilize the division when the denominator is small. MAE measures the average absolute error, treating all deviations equally. MSE squares each error before averaging, making it more sensitive to large errors and outliers. The Structural Similarity (SSIM) focuses on perceptual quality by comparing structural patterns, luminance, and contrast between images. + +# 3.2. Tile-based Training via RT-Loss + +In this section, we introduce RT-Loss and our Tile-based training paradigm. Our goal is to incorporate random viewpoint observations within each training batch, while preserving structural information to enforce more comprehensive constraints. In contrast to the individual rays used + +in NeRFs, Gaussian Splatting utilizes a Tile-based splatting process that projects Gaussians onto the entire view screen, rather than performing active ray marching. Hence, the smallest unit in the loss calculation must align with the splatting rendering and is set as a $16 \times 16$ tile $t$ . We load the datasets as tiles instead of images, randomly sample $N$ tiles as a training batch $\mathcal{T}$ and obtain the rendered tiles set $\mathcal{T}(\hat{\mathcal{C}})$ along with their corresponding ground truth $\mathcal{T}(\mathcal{C})$ , where $\hat{\mathcal{C}} = \{\hat{\mathcal{C}}(t) \mid t \in \mathcal{T}\}$ and $\mathcal{C} = \{\mathcal{C}(t) \mid t \in \mathcal{T}\}$ . We use MAE as the color loss, which can be written as: + +$$ +L _ {\mathrm {R T} - \mathrm {M A E}} = \frac {1}{\| \mathcal {T} \|} \sum_ {\boldsymbol {t} \in \mathcal {T}} \| \hat {\boldsymbol {C}} (\boldsymbol {t}) - \boldsymbol {C} (\boldsymbol {t}) \| _ {1}. \tag {8} +$$ + +The tiles contain viewpoint's structural information. If we solely rely on rgb loss $L_{\mathrm{MAE}}$ , the positional relationship contained within the tiles would be completely lost [42]. Inspired by S3IM's adaptation of SSIM on random rays, we propose RT-SSIM for 3DGS to retain the structural information embedded in random tiles. The concept is straightforward: while keeping the total batch size unchanged, we compute and average the SSIM across multiple independent and non-overlapping random tiles from different views, with kernel size and stride set to $K \times K$ (e.g., $9 \times 9$ for overall structtrue or $3 \times 3$ for fine detail) with a stride of $s$ . RT-SSIM can be written as: + +$$ +\operatorname {R T - S S I M} (\hat {\mathcal {T}}, \mathcal {T}) = \frac {1}{N} \sum_ {\boldsymbol {t} \in \mathcal {T}, v \in \boldsymbol {V}} \operatorname {S S I M} \left(\boldsymbol {t} ^ {(v)} (\hat {\mathcal {C}}), \boldsymbol {t} ^ {(v)} (\mathcal {C})\right), \tag {9} +$$ + +where $V$ controls the number of views in all $N$ tiles. In particular, when $V = N$ , it implies all tiles are selected from different views. As RT-SSIM positively correlated with image similarity, we define the structure loss $L_{\mathrm{RT - SSIM}}$ as + +$$ +L _ {\mathrm {R T} - \mathrm {S S I M}} (\Theta , \mathcal {T}) = 1 - \mathrm {R T} - \mathrm {S S I M} (\hat {\mathcal {T}}, \mathcal {T}) \tag {10} +$$ + +We note that tiles are particularly suitable for parallel computation, thus the additional computational cost of Tile-based training is minimal. The final hybrid loss $\mathcal{L}_{\mathrm{RT}}$ can be written as: + +$$ +\mathcal {L} _ {\mathrm {R T}} (\Theta , \mathcal {T}) = (1 - \lambda_ {s}) L _ {\mathrm {R T} - \mathrm {M A E}} + \lambda_ {s} L _ {\mathrm {R T} - \mathrm {S S I M}} \tag {11} +$$ + +where $\lambda_{s}$ balances structure and color losses. + +We present the pseudocode in Algorithm 1. In the standard 3DGS training paradigm, tiles are rendered in parallel using the same view parameters. In contrast, we sample tiles from multiple different views to integrate multi-view constraints. In our random-tile training paradigm, We introduce two additional hyperparameters, $\lambda_{s}$ and $V$ , where $\lambda_{s}$ controls the loss weight and $V$ determines the number of viewpoints. When $V = N$ , each tile is sampled from a + +Algorithm 1: Tile-based Training via RT-loss +1 Let $\mathcal{A}$ be an SGD training algorithm; +2 while the termination criteria has not been satisfied do +3 Sample a training batch $\mathcal{T}$ consisting of $N$ tiles from $V$ views in dataset $\mathcal{D}$ ; +4 for $i = 1$ to $V$ do +5 Randomly select tiles $\mathcal{T}_i = [t_1, t_2 \dots t_{V/N}]$ , where $t_j \in \mathcal{T}_i \subseteq \mathcal{T}$ ; +6 VisibilityCheck(); +7 end +8 for each tile $t$ in $\mathcal{T}$ do +9 Compute the forward splating result after $\hat{\mathcal{C}} = \{\hat{C}(t) \mid t \in \mathcal{T}\}$ ; +10 Obtain the ground-truth $\mathcal{C} = \{C(t) \mid t \in \mathcal{T}\}$ ; +11 end +12 for $i = 1$ to $V$ do +13 let rendered pixels $\hat{\mathcal{C}}$ form the pixel patch $\pmb{t}^{(v)}(\hat{\mathcal{C}})$ in view $V_i$ ; +14 Compute SSIM $(\pmb{t}^{(v)}(\hat{\mathcal{C}}), \pmb{t}^{(v)}(\mathcal{C}))$ with the given kernel $K \times K$ ; +15 end +16 Obtain the RT-SSIM loss $L_{\mathrm{RT - SSIM}}(\Theta, \mathcal{T}) = 1 - \frac{1}{N} \sum \mathrm{SSIM}(\pmb{t}^{(v)}(\hat{\mathcal{C}}), \pmb{t}^{(v)}(\mathcal{C}))$ ; +17 Obtain the conventional $L_{MAE}$ loss as color loss $L_{\mathrm{RT - MAE}}(\Theta) = \frac{1}{||\mathcal{T}||} \sum_{t \in \mathcal{T}} ||\hat{\mathcal{C}}(t) - C(t)||_1$ ; +18 $\mathcal{L}(\Theta, \mathcal{T}) = (1 - \lambda_s)L_{\mathrm{RT - MAE}}(\Theta, \mathcal{T}) + \lambda_s L_{\mathrm{RT - SSIM}}(\Theta, \mathcal{T})$ ; +19 Compute the gradient $\nabla \mathcal{L}(\Theta)$ ; +20 Update model parameters $\Theta$ by $\mathcal{A}$ +21 end + +distinct view, and their visibility is computed in parallel during preprocessing to accelerate the rendering process. For a fair comparison, we set the total tiles count $N$ to match that in image-wise training, ensuring a consistent batch size in gradient descent algorithm $\mathcal{A}$ . Our random-tile strategy uniformly samples tiles across each view, balancing observations for each perspective. Our ablation study in Sec. 4.4 shows that this strategy achieves slightly better results than purely random selection. + +# 3.3. Tile-based Adaptive Density Control + +In 3D Gaussian Splating, a Gaussian is split or cloned depending on whether the average cumulative 2D gradient on the NDC coordinates $\sum \| \mathbf{g}_i\|$ across interval iterations exceeds a specified threshold $\tau_{\mathrm{densify}}$ , as shown in Fig. 2. During our Tile-based training, Gaussians in one iteration will + +![](images/a076c6dcc119548e866a1570437a3f7d52a44658c2e20de84f1f8e9ebbecbea8.jpg) +Figure 2. Image-wise vs Tile-wise Adaptive Density Control. In Tile-based training, NDC gradients for densification are computed as a weighted sum across all observations, balancing multi-view contributions, where $S1$ represents the area of the Gaussian projection onto the tile, and $S2$ is the total projected area. + +![](images/bb3fca209b35109499f4ef69e35f33f058d8d5b8d800197aec5a16773fef25d3.jpg) + +have multiple NDC gradients from different training views. Thus, there are two methods for computing density gradients: Tile-Count Densification and Iteration-Count Densification. + +$$ +\frac {\sum_ {k = 1} ^ {\text {I t e r}} \sum_ {m = 1} ^ {M _ {k}} \| \mathbf {g} _ {k m} \|}{\sum_ {k = 1} ^ {\text {I t e r}} M _ {k}} > \tau_ {\text {d e n s i f y}} \tag {12} +$$ + +$$ +\frac {\sum_ {k = 1} ^ {\text {I t e r}} \sum_ {m = 1} ^ {M _ {k}} \left(\left\| \mathbf {g} _ {k m} \right\| \times r _ {m}\right)}{\sum_ {k = 1} ^ {\text {I t e r}} 1} > \tau_ {\text {d e n s i f y}}, \tag {13} +$$ + +where $\text{Iter}$ represents the number of iterations in which a Gaussian participates in $\alpha$ -blending within each densification interval. In Tile-count Densification ( Eq. (12)), the NDC gradient is averaged over the number of observed tiles $M_{k}$ , while in Iteration-Count Densification ( Eq. (13)), it is averaged over the number of observed iterations. Additionally, in Iteration-Count Densification, the NDC gradient is weighted by $r_{m}$ , the ratio of the projected area shown in Fig. 2. To maintain consistency with the split parameters used in the original 3DGS and to emphasize the effect of RT-Loss, we default to using Iteration-Count for adaptive density control. Further details on NDC gradients and densification algorithm can be found in the Supplementary Material. + +# 4. Experiments + +# 4.1. Experimental Setup + +Dataset and Metrics. We conduct a comprehensive evaluation on 26 publicly available scenes widely used + +Table 1. Quantitative comparison to previous methods on benchmark datasets. We report results on three commonly used datasets. The best, second best, and third best results are denoted by red, orange, and yellow, respectively. + +
Dataset +Method | MetricsMip-NeRF360Tanks&TemplesDeep Blending
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Plenoxels [13]23.620.6700.44322.130.7620.33423.060.7950.510
Instant-NGP [34]26.430.7250.33922.940.7520.30423.620.7970.423
Mip-NeRF360 [3]29.230.8440.20723.120.7710.25029.400.9010.245
3D-GS [19]28.690.8700.18224.360.8310.21329.410.9030.243
Scaffold-GS [31]28.840.8480.22024.810.8400.20530.210.9060.254
Ours29.660.8920.17125.130.8520.19230.150.9110.224
Ours (+ Scaffold-GS)29.860.8850.20625.320.8570.18930.490.9130.231
+ +![](images/1ba36238fca19915e7c3b35d8f16d1ed327a9c034b61561d95b3cccd8f40be24.jpg) +Figure 3. Qualitative comparison across benchmark datasets [3, 11, 21]. We highlight the differences in the image using patches for clearer visibility. Our approach consistently outperforms 3D-GS on these scenes, demonstrating clear advantages in challenging scenarios, such as fine-scale details (MIP360-ROOM(a), TANDT-BARN), thin geometry (SMERF-NYC(a), SMERF-NYC(b)), texture-less regions (TANDT-TRAIN), light effects (MIP360-COUNTER, TANDT-TRUCK), and insufficient observations (TANDT-BARN, SMERF-NYC(b)). + +in the field of 3D reconstruction across multiple diverse tasks. Specifically, we evaluate our method on seven scenes from Mip-NeRF360 [3], another seven from Tanks&Temples [21], and two additional scenes from DeepBlending [16]. To further demonstrate the limitations of image-wise loss in capturing global structure, our evaluation incorporated all four scenes from [4], with $5 - 10 \times$ more training views. We additionally evaluated dynamic benchmarks, specifically using the Neural 3D Video Dataset containing six scenes, to demonstrate the effectiveness of RT-Loss in 4D reconstruction. In addition to the commonly used metrics (PSNR, SSIM [39], and LPIPS [49]), we fur + +ther provide the storage size (MB) and rendering speed (FPS) to assess model compactness and efficiency, with all experiments performed on an NVIDIA RTX 4090 GPU. The main paper reports the average metrics across all scenes in each dataset, while detailed per-scene quantitative results are presented in the supplementary materials. + +Baseline and Implementation. 3DGS[19], Scaffold-GS[31] and 4D-GS [10] are selected as our baselines due to their state-of-the-art performance. We also present the results of Mip-NeRF360 [3], iNGP [34] and Plenoxels [13] as in [19]. To demonstrate the generality of the tile-based ap + +![](images/13d751e8e657b1770978728860ef8a278143b67ab58d0941ecd1adf8ab9a47c3.jpg) +Figure 4. Convergence performance. We evaluated the convergence of PSNR for both the training and testing views during the training process. With the same batch size, the convergence of image-wise loss in 3DGS is slower than RT-Loss, with the gap widening on larger datasets, e.g., COURTHOUSE and BERLIN + +![](images/022bc578cbbd7af7a70a6439ccf4adf392cfcec02a7324e64f459960cee6d63b.jpg) + +![](images/1ba3d1e74b5577f324a2ae4ad3745555784ef85925c3e046fbfd0224e204d501.jpg) + +proach, we implemented the tile-wise method in Scaffolds for comparison. We trained for $30k$ iterations and rendered the same number of pixels per iteration, ensuring the total pixels from random tiles matched a complete image. To independently demonstrate the effectiveness of the tile-based training paradigm and RT-loss, we modified only the sampling and loss calculations for random tile training while keeping the default 3DGS parameters unchanged, including the densification threshold $\tau_{\mathrm{densify}} = 0.0002$ and all learning rates. For RT-loss, we sample random tiles from $V = 5$ views and apply the same structure loss weight, $\lambda_s = 0.2$ , as used in the original 3DGS. + +# 4.2. Results Analysis + +Benchmark Evaluation. The quantitative results on Mip-NeRF360 and Tanks&Temples are presented in Tab. 1. Results show that we outperform existing methods on standard datasets. Fig. 3 illustrates that tile-wise training achieves enhanced visual quality, with improved geometry accuracy and richer texture details. Moreover, our method retains the efficiency of Gaussian methods, as shown in Tab. 2. We compared the rendering speed and storage size with 3DGS. We achieve improved performance while using a comparable or fewer number of Gaussians, demonstrating high-quality reconstruction without compromising speed. By incorporating our tile-based approach, Scaffold-GS also achieved notable improvements, suggesting that the image-wise training paradigm, which lacks multi-view constraints, limits the reconstruction quality of various GS methods. + +Convergence Performance. We evaluated our convergence performance across several representative scenarios. As shown in Fig. 4, compared to the original image-wise loss in 3DGS, the RT-Loss based on multi-view observations provides more effective constraints, resulting in significantly faster convergence. The difference in convergence speed is + +Table 2. Rendering FPS and storage comparison. The rendering speed of both methods is measured on the same machine. + +
ScenesMip-NeRF360ROOMKITCHEN
MemFPSPSNRMemFPSPSNRMemFPS
3DGS6717831.7931712331.63372118
Ours6259332.8131812532.35320135
ScenesTanks&TemplesCourthOUSEMEETINGROOM
MemFPSPSNRMemFPSPSNRMemFPS
3DGS31712623.1317414626.27276120
Ours30213024.1416813926.98260133
+ +Table 3. PSNR evaluation for scenes with increased training views. The convergence capability of RT-Loss is particularly evident on the SMERF dataset[11], which includes approximately 10 times more training views than Mipnerf360. Additionally, we recorded the PSNR gap at 4k and 30k iterations. + +
Scene.views test iterationAlameda(1734)Berlin(1511)London(1874)
4k30k4k30k4k30k
3DGS17.6222.2722.6327.5322.3225.75
Ours20.1123.91624.8028.7924.0726.626
+ +Table 4. Quantitative comparison on dynamic datasets. We applied RT loss for dynamic scene reconstruction and compared it with 4D-GS[10] on the challenging real-world dynamic dataset Neu3D[23]. + +
ModelPSNR↑D-SSIM↓LPIPS↓FPS↑Storage (MB)↓
4D-GS[10]31.150.9740.0493090
Ours31.740.9730.0433478
+ +visually evident in Fig. 1. This advantage in convergence is even more pronounced in scenes with a large number of training views, e.g., COURTHOUSE and BERLIN. + +Scene with More Training Views. We compared our method with the original 3DGS in scenarios with more + +![](images/90541314fd9e72cfe853e4ded18acf5f31110dd2a786fd58303c0c1a8a5f0ea0.jpg) +Figure 5. Gaussians structure comparison. RT-Loss demonstrates stronger structural reconstruction ability in scenes with more training views. In contrast, 3DGS not only loses details but also fails to recover the geometric structure of the scene. + +![](images/12290bde3171d39c0a76a115e98ef6a414861af77d76583485dc54b62c29df93.jpg) +Figure 6. Qualitative comparisons on the Neural 3D Video Dataset. In 4D reconstruction, constraints from multiple viewpoints and time also help generate more accurate details. + +training views. In Tab. 3, our method achieved a significant improvement in metrics. The gap is particularly noticeable in the early training stages (4k iterations), as image-wise training naturally requires more iterations to fully capture the scene. Fig. 5 further illustrates that our method produces a more refined geometric structure on large datasets. This supports our hypothesis about the current training paradigm: as the number of training views increases, image-wise training struggles to comprehensively capture scene information, resulting in less reliable optimization and densification. Furthermore, through experiments with extended training durations, we demonstrate that naively increasing the iteration count for image-wise training does not effectively address this limitation. This is because once incorrect geometry is formed during Gaussian optimization, it is difficult to correct by simply extending the training process. For further details, please refer to the supplementary material. + +Dynamic Scenes. We evaluated our method on the real-world dynamic scene dataset, the Neural 3D Video Dataset, using 4D-GS [10] as the baseline. The expansion of the time dimension in the dynamic dataset resulted in a larger + +Table 5. Ablation study on RT-Loss components. 'Img-' refers to image-wise loss, and 'RT-' refers to random-tile loss. + +
Scene(views)KITCHEN(279)CORTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
LImg-MAE + LImg-SSIM (3DGS)31.6737223.11173
LImg-MAE31.0913121.3544
LRT-MAE31.6512722.6363
LRT-MAE + LImg-SSIM32.0235223.50186
LRT-MAE + LRT-SSIM (Ours)32.3532024.15168
+ +Table 6. Effect of Adaptive Density Control. 'Tile-Count' and 'Iteration-Count' are densification strategies described in Sec. 3.3, with 'Iteration-Count' being the default. + +
Scene(views)KICHEN(279)COURTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
3DGS w/o Densification29.845822.96118
Ours w/o Densification30.325823.83118
Ours w/ Tile-Count31.9517523.56125
Ours w/ Iteration-Count32.3532024.15168
+ +Table 7. Effect of Tile Sampling and Batch Training. 'V' refers to the number of randomly selected viewpoints, determining the spread of observations. 'B' refers to Batch train, which means using multiple images as a batch for training. 'Non-Uniform' and 'Uniform' indicate whether the tiles are distributed evenly. + +
Scene.views)KICHEN(279)COURTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
tile-wise V = 231.8034323.76182
tile-wise V = 432.2233723.92175
tile-wise V = 632.4033224.29161
image-wise B = 231.6735223.82214
image-wise B = 431.2234823.69207
Non-Uniform Sample32.3031024.02192
Uniform Sample32.3832024.14168
+ +volume of training data. Although most 4D Gaussian methods address this issue by using batch training and rendering multiple images together as a training batch, our results show that, with the same batch size, tile-based training provides more effective constraints across multiple viewpoints and time frames. The results show that RT-Loss is also effective for dynamic scene reconstruction. As depicted in Tab. 4, we notably improved the rendering performance on dynamic scenes without sacrificing efficiency. As shown in Fig. 6, our method reconstructs finer details, attributed to our randomized approach, which provides stronger constraints across both viewpoints and time frames compared to image-wise training. Details are provided in the supplementary material. + +# 4.3. Ablation Studies + +Effect of RT-Loss components. We conducted an ablation on the loss components of Eq. (11), which includes the base color loss $L_{\mathrm{RT - MAE}}$ and the structural loss $L_{\mathrm{RT - SSIM}}$ . As shown in the $2^{nd}$ and $3^{rd}$ row in Tab. 5, when only MAE loss is used, the PSNR significantly decreases, with a lower + +![](images/2141aae8c323a480c32e98b5946b71910c7133473a04e988f9195bc5cafec40d.jpg) +Figure 7. Effect of initialization on reconstruction quality. (Left) Gaussians optimized from SfM point initialization. (Right) Gaussians optimized from random points. Our method reconstructs better structures and finer details even with random initialization, demonstrating superior robustness to initialization issues. + +![](images/7e2a0f2ecf5a0597b516e8951e4c3ba6c92cb1969ad768c9b32e12232e2df3eb.jpg) + +![](images/c29dda0bb4a84c4a9fcaafe348d54de7ec83b886b3c270b65ecf048ce8623e56.jpg) + +Table 8. Sensitivity to Initialization. In the 'w/o SfM initialization' setting, Gaussians are initialized with random positions. RT-Loss demonstrates stronger robustness in addressing initialization issues. + +
Scene(views)KICHEN(279)COURTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
3DGS w/ SfM initialization31.6337223.13174
3DGS w/o SfM initialization29.8926921.77121
Ours w/ SfM initialization32.3532024.14168
Ours w/o SfM initialization31.2227223.21186
+ +tendency for Gaussian densification. However, our $L_{\mathrm{RT-MAE}}$ still outperforms the image-wise constraints, and this effect becomes more pronounced in the CourtHouse scene with more training views, where the PSNR difference reaches up to 1.3 dB. Notably, in the Kitchen scene, the $L_{\mathrm{RT-MAE}}$ constraint achieves PSNR on par with 3DGS, despite using only one-third of the Gaussian points, as shown in the $1^{st}$ and $3^{rd}$ row in Tab. 5. Additionally, incorporating SSIM improves the PSNR as the number of Gaussians increases. Furthermore, to demonstrate the improvement of structure loss brought by RT-SSIM, we rendered the complete image and incorporated image-wise SSIM as a structural loss during random tile training. In contrast, the metrics using RT-SSIM consistently yield higher values, showing that the multi-view observations in RT-SSIM provide more accurate reconstruction constraints than image-wise SSIM. + +Effect of Adaptive Density Control. We ablate our densification strategy described in Sec. 3.3 on Tab. 6. Without any densification, the optimization of GS relies solely on the loss function without any splitting or cloning. Our improvements demonstrate the effectiveness of multi-view constraints in enhancing performance. Tile-Count densification, which averages the NDC gradient across tiles, causes the Gaussians to resist densifying and lower PSNR. This is because random-tile training increases the frequency with which a Gaussian is observed, leading to a smaller av + +erage gradient. + +Effect of Tile Sampling and Batch Training. We analyzed the impact of viewpoint count $V$ and tile uniformity described in Sec. 3.2 on Tab. 7. Uniform sampling better distributes viewpoint constraints in the loss, yielding slight improvements and serving as our default. In data-rich scenes like Courthouse, more viewpoints provide greater benefits, as larger scenes require sufficient and evenly distributed observations. This highlights the adaptability of our random-tile training paradigm to different scene requirements. We also tested the comparative effect of image-wise batch training. Under a similar total training volume (iterations = 15k), introducing multiple images as a batch to incorporate multi-view observations does not improve quality as effectively as the tile-wise approach. Moreover, our observations indicate that such dense inclusion of multiple images significantly increases computational overhead. + +Sensitivity to Initialization. The initialization problem has long been considered a critical factor impacting 3DGS performance [18]. We conduct experiments under two initialization conditions. Fig. 7 shows that our method recovers improved structure and appearance even with random initialization. As shown in Tab. 8, our method demonstrates greater robustness to initialization variations. This suggests that part of the initialization issue may arise from uneven and insufficient training constraints, and RT-Loss has the potential to address this limitation. + +# 4.4. Discussions and Limitations + +We propose a practical and effective approach by incorporating tile-wise constraints into Gaussian optimization to replace image-wise constraints. We have demonstrated its clear effectiveness and universality through both the main experiments and ablation studies. However, in experiments with large texture-less regions, our method fails to improve the geometry. To further enhance the tile-wise optimization, finer strategies could be explored in practice, such as further allocating multi-view constraints to identical content, in order to provide more accurate and uniform multi-view constraints for improvement. + +# 5. Conclusion + +In this work, we introduce a Tile-based training paradigm that integrates multi-view appearance and structural constraints. Extensive experiments demonstrate its effectiveness across various scenarios. We expect tile-wise training to hold strong potential as a promising paradigm for future research and applications in Gaussian Splatting. + +# Acknowledgments + +This work was partially supported by Key R&D Program of Zhejiang Province (No. 2023C01039). + +# References + +[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 1 +[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 2, 5, 3 +[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697–19705, 2023. 5 +[5] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16569-16578, 2023. 2 +[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splattering with progressive propagation, 2024. 1 +[7] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2 +[8] Xiaobiao Du, Yida Wang, and Xin Yu. Mvgs: Multi-view-regulated gaussian splatting for novel view synthesis, 2024. 2 +[9] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing, 2021. 2 +[10] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. arXiv preprint arXiv:2402.03307, 2024. 2, 5, 6, 7 +[11] Daniel Duckworth, Peter Hedman, Christian Reiser, Peter Zhizhin, Jean-François Thibert, Mario Lučić, Richard Szeliski, and Jonathan T Barron. Smerf: Streamable memory efficient radiance fields for real-time large-scene exploration. ACM Transactions on Graphics (TOG), 43(4): 1-13, 2024. 5, 6, 3 +[12] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building rome on a cloudless day. In European Conference on Computer Vision, pages 368–381, 2010. 1 +[13] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 5501-5510, 2022. 5, 3 +[14] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2 +[15] Xiang Guo, Guanying Chen, Yuchao Dai, Xiaqing Ye, Jiadai Sun, Xiao Tan, and Errui Ding. Neural deformable voxel grid for fast optimization of dynamic view synthesis, 2022. 2 +[16] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 5, 3 +[17] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf: Efficient neural radiance fields, 2022. 2 +[18] Jaewoo Jung, Jisang Han, Honggyu An, Jiwon Kang, Seonghoon Park, and Seungryong Kim. Relaxing accurate initialization constraint for 3d gaussian splatting. arXiv preprint arXiv:2403.09413, 2024. 2, 8 +[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 5, 3 +[20] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces, 2022. 2 +[21] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 5, 3 +[22] Christoph Lassner and Michael Zollhöfer. Pulsar: Efficient sphere-based neural rendering, 2020. 1 +[23] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 6 +[24] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering, 2024. 1, 2 +[25] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2 +[26] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splattering for real-time dynamic view synthesis, 2024. 2 +[27] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 2 +[28] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. High-fidelity and real-time novel view synthesis for dynamic scenes. In SIGGRAPH Asia Conference Proceedings, 2023. 2 + +[29] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763, 2023. 2 +[30] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. arXiv preprint arXiv:2311.17061, 2023. 2 +[31] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 1, 2, 5, 3 +[32] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2 +[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2 +[34] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2, 5, 3 +[35] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2 +[36] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1 +[37] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2 +[38] Dan Wang, Xinrui Cui, Septimiu Salcudean, and Z. Jane Wang. Generalizable neural radiance fields for novel view synthesis with transformer, 2022. 2 +[39] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[40] Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. arXiv preprint arXiv:2403.16292, 2024. 2 +[41] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2 + +[42] Zeke Xie, Xindi Yang, Yujie Yang, Qi Sun, Yixiang Jiang, Haoran Wang, Yunfeng Cai, and Mingming Sun. S3im: Stochastic structural similarity and its unreasonable effectiveness for neural fields, 2023. 2, 3 +[43] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Real-time 360sparse view synthesis using gaussian splatting, 2024. 2 +[44] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, Dahua Lin, Michael Zollhöfer, and Christian Richardt. Vr-nerf: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia 2023 Conference Papers, page 1–12. ACM, 2023. 1 +[45] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields, 2023. 2 +[46] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. arXiv preprint arXiv:2409.14316, 2024. 2 +[47] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splattering. arXiv preprint arXiv 2310.10642, 2023. 2 +[48] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 2 +[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5 +[50] Xiaoyu Zhang, Weihong Pan, Chong Bao, Xiyu Zhang, Xiaojun Xiang, Hanqing Jiang, and Hujun Bao. Lookcloser: Frequency-aware radiance field for tiny-detail scene. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16122-16132, 2025. 2 +[51] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 1 +[52] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. 2 +[53] Zuo-Liang Zhu, Beibei Wang, and Jian Yang. Gs-ror: 3d gaussian splattering for reflective object relighting via sdf priors, 2024. 2 +[54] C Lawrence Zitnick, Sing Bing Kang, Matthew Uysettadaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM TOG, 2004. 2 +[55] Yiming Zuo and Jia Deng. View synthesis with sculpted neural points. arXiv preprint arXiv:2205.05869, 2022. 2 \ No newline at end of file diff --git a/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/images.zip b/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8c7778bb7db50f4d7e5f1376c05f7c4733a69c63 --- /dev/null +++ b/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09582df41b76f5d9273b6e4e784ff530faaee48ac43d3ac7a35c24ad6bb2d3e6 +size 889132 diff --git a/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/layout.json b/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0febafb483ad18b85ea58fc80fedf6ae288467c6 --- /dev/null +++ b/2025/Tile-wise vs. Image-wise_ Random-Tile Loss and Training Paradigm for Gaussian Splatting/layout.json @@ -0,0 +1,8894 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 84, + 102, + 526, + 140 + ], + "type": "text", + "content": "Tile-wise vs. Image-wise: Random-Tile Loss and Training Paradigm for Gaussian Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "spans": [ + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "text", + "content": "Xiaoyu Zhang" + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "text", + "content": " Weihong Pan" + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "text", + "content": " Xiaojun Xiang" + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "text", + "content": " Hongjia Zhai" + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "text", + "content": " Liyang Zhou" + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "text", + "content": "SenseTime Research " + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 74, + 159, + 534, + 205 + ], + "type": "text", + "content": "State Key Lab of CAD&CG, Zhejiang University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 69, + 211, + 545, + 409 + ], + "blocks": [ + { + "bbox": [ + 69, + 211, + 545, + 409 + ], + "lines": [ + { + "bbox": [ + 69, + 211, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 69, + 211, + 545, + 409 + ], + "type": "image", + "image_path": "ebcf5affed1d17c580fa56a2a3e5f1db03ea669d7f7ca3838945301ef832f074.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 415, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 54, + 415, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 555, + 460 + ], + "type": "text", + "content": "Figure 1. Random-Tile Loss integrates multi-view appearance and structural constraints during each optimization step. Under the same batch size, it consistently improves convergence efficiency and rendering quality while preserving compactness and real-time capability. The results on test views across different training iterations suggest a more stable and efficient convergence compared to the image-wise loss in Gaussian Splatting. Please refer to the Supplementary Video for more detailed comparisons on convergence performance." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 465, + 200, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 465, + 200, + 477 + ], + "spans": [ + { + "bbox": [ + 151, + 465, + 200, + 477 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 491, + 297, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 491, + 297, + 670 + ], + "spans": [ + { + "bbox": [ + 54, + 491, + 297, + 670 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) has drawn significant attention for its advantages in rendering speed and quality. Most existing methods still rely on the image-wise loss and training paradigm because of its intuitive nature in the Splatting algorithm. However, image-wise loss lacks multi-view constraints, which are generally essential for optimizing 3D appearance and geometry. To address this, we propose RT-Loss along with a tile-based training paradigm, which uses randomly sampled tiles to integrate multi-view appearance and structural constraints in 3DGS. Additionally, we introduce an tile-based adaptive densification control strategy tailored for our training paradigm. Extensive experiments show that our approach consistently improves performance metrics while maintaining efficiency across various benchmark datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 465, + 394, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 465, + 394, + 477 + ], + "spans": [ + { + "bbox": [ + 315, + 465, + 394, + 477 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 485, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 485, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 312, + 485, + 555, + 617 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF) [33] has brought revolutionary advances to photorealistic novel view synthesis (NVS) through exceptional rendering quality, attracting significant attention from the computer vision, graphics, and robotics communities [44]. Recently, 3D Gaussian Splating (3DGS) [19] has significantly enhanced rendering efficiency by replacing traditional ray marching with the splating algorithm [22]. Numerous studies have focused on optimizing the Gaussian feature or heuristic split algorithm in 3DGS to improve the appearance quality and geometric accuracy [6, 24, 31, 51]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 618, + 556, + 715 + ], + "type": "text", + "content": "However, while the splatting algorithm significantly improves rendering speed, its potential for incorporating Tile-based multi-view constraints remains underexplored. Specifically, 3DGS initializes with point clouds derived from Structure from Motion (SfM) [1, 12, 36] and optimizes a set of anisotropic 3D Gaussians to represent the scene. Tile-based splatting plays a crucial role in achieving real-time rendering by directly splatting visible Gaussians onto" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 164, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 164, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 164, + 703 + ], + "type": "text", + "content": "*Authors contributed equally." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26923" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "the camera view, rather than querying every ray point. As a result, defining the loss function on each training image " + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": " is a intuitive and straightforward choice:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 118, + 295, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 118, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 67, + 118, + 295, + 133 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\Theta) = (1 - \\lambda) L _ {\\mathrm {M A E}} (\\Theta , \\mathcal {I}) + \\lambda L _ {\\mathrm {D - S S I M}} (\\Theta , \\mathcal {I}), \\quad (1)", + "image_path": "8516c82e8640ab5fd1ffb31c2887436bf94df754e95d9cec84cf0e420cf1bd89.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 142, + 296, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 296, + 214 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 142, + 296, + 214 + ], + "type": "inline_equation", + "content": "L_{MAE}" + }, + { + "bbox": [ + 55, + 142, + 296, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 142, + 296, + 214 + ], + "type": "inline_equation", + "content": "L_{D-SSIM}" + }, + { + "bbox": [ + 55, + 142, + 296, + 214 + ], + "type": "text", + "content": " are computed image-wise on the entire training image. Although this approach benefits the calculation of structural loss SSIM across the dense image, it also means that gradients from multiple viewpoints cannot be integrated in a single optimization step. In contrast, NeRF typically optimizes a pixel-wise loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 223, + 296, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 223, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 109, + 223, + 296, + 253 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\Theta) = \\frac {1}{\\| \\mathcal {R} \\|} \\sum_ {\\boldsymbol {r} \\in \\mathcal {R}} L _ {\\mathrm {M S E}} (\\Theta , \\boldsymbol {r}), \\tag {2}", + "image_path": "c1e3c92f5041eeb39807830fc4ea1049caf3d17d801cc7266ddb7d870714444b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "text", + "content": "where each individual ray " + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "inline_equation", + "content": "r = (x, d, c)" + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "text", + "content": " in the mini-batch " + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 55, + 262, + 296, + 334 + ], + "type": "text", + "content": " is randomly sampled from the entire training dataset. Pointwise sampling from random viewpoints provides more comprehensive gradients. However, it fails to utilize the collective supervision of distant pixels, thus inevitably overlooking structural information in the overall scene [42]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 335, + 296, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 406 + ], + "type": "text", + "content": "To address this issue, we propose a Tile-based training paradigm that integrates both multi-view constraints and structural information within the splatting process. We sample a set of tiles " + }, + { + "bbox": [ + 55, + 335, + 296, + 406 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 335, + 296, + 406 + ], + "type": "text", + "content": " from multiple training views, each individual tile " + }, + { + "bbox": [ + 55, + 335, + 296, + 406 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 335, + 296, + 406 + ], + "type": "text", + "content": " is the minimal unit in splatting. The loss is formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 413, + 296, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 413, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 59, + 413, + 296, + 454 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R T} (\\Theta) = \\frac {1}{\\| \\mathcal {T} \\|} \\sum_ {\\boldsymbol {t} \\in \\mathcal {T}} L _ {\\mathrm {M A E}} (\\Theta , \\boldsymbol {t}) + \\lambda L _ {\\mathrm {R T - S S I M}} (\\Theta , \\mathcal {T}), \\tag {3}", + "image_path": "3b84df7d4f90acfbabb2d493486ff88b687d59131f51d70a32396b9e94de5c80.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 464, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 534 + ], + "type": "text", + "content": "which is computed over a random-tile batch " + }, + { + "bbox": [ + 55, + 464, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 464, + 296, + 534 + ], + "type": "text", + "content": " that contains multi-view appearance and structure. Our Tile-based training paradigm maintains the training efficiency of the 3DGS rendering process, with detailed definitions and calculations provided in Sec. 3.2. Overall, our contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 536, + 295, + 645 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 536, + 295, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 609 + ], + "type": "text", + "content": "- We propose a Tile-based training paradigm and loss formulation, incorporating both multi-view appearance constraints and structural similarity constraints with random tile sampling. To the best of our knowledge, we are the first to introduce a random-tile approach as a replacement for image-wise training paradigms in 3DGS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 609, + 295, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 645 + ], + "type": "text", + "content": "- Extensive experiments indicate that this Tile-based training paradigm yields notable improvements in training performance while maintaining model compactness." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 657, + 143, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 143, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 143, + 669 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Novel View Synthesis. Neural Radiance Field (NeRF) [33] has significantly advanced novel view synthesis by employing a multilayer perceptron (MLP) to model scenes" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 252 + ], + "type": "text", + "content": "with volumetric rendering, enabling the generation of realistic images from different perspectives. NeRF introduced a transformative approach to scene representation, sparking extensive research focused on enhancing rendering quality [2, 3, 27, 37, 50, 54, 55], speeding up computation [5, 14, 17, 34, 35], enabling 3D content generation [38], handling dynamic scenes [9, 15, 25], and improving generalization to novel scenes [45]. Due to the ray marching rendering technique, most NeRF-based methods are trained by utilizing batches of randomly sampled rays, with mean squared error (MSE) as the primary loss function. To integrate structural similarity, S3IM [42] enhances NeRF by incorporating patches of random rays in the loss calculation with a structural similarity index (SSIM), improving performance in tasks such as sparse view synthesis." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 257, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 556, + 568 + ], + "type": "text", + "content": "Gaussian Splatting Training Paradigm. 3D Gaussian Splatting (3DGS) has gained significant attention in novel view synthesis tasks due to its impressive performance. Using a rasterization-based rendering approach [20], 3DGS avoids the computational overhead of ray marching. With its high rendering speed and quality, 3DGS has found applications across various fields [48], including dynamic scene reconstruction [7, 10, 28, 29, 32, 41, 47], sparse view synthesis [43, 52], reflective object modeling [53], and geometric surface recovery [24, 30, 40]. Additional optimizations target specific issues, such as the geometric improvements seen in Scaffold-GS [31], which combines neural field models with Gaussian Splatting to achieve state-of-the-art surface accuracy. [18] introduces adaptive strategies to reduce dependency on SfM initialization. While most approaches adopt an image-wise training paradigm, some recent research explores multi-view constraints to enhance Gaussian Splatting. For example, [46] leverages multi-view stereo (MVS) to enhance initialization geometry, using multi-view consistency to optimize Gaussian geometry. [8, 10, 26] utilize batch accumulation to alleviate data pressure from the temporal dimension by rendering multiple images per iteration and accumulating gradients. However, this straightforward image-wise accumulation not only impacts performance but also fails to fully leverage the benefits of multiview constraints with the same batch size." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 594, + 374, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 374, + 605 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 374, + 605 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 619, + 400, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 400, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 400, + 630 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "Gaussian Splatting. 3DGS [19] represents scenes using anisotropic 3D Gaussians, which retains the differential properties of volumetric representations while enabling efficient rendering through Tile-based rasterization. Each Gaussian is assigned a color " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": ", represented by spherical harmonics, and an opacity " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": ", which modulates " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "G(x)" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": " during" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26924" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 262, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 262, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 262, + 84 + ], + "type": "text", + "content": "blending. Its spatial distribution is then defined by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 86, + 91, + 295, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 91, + 295, + 118 + ], + "spans": [ + { + "bbox": [ + 86, + 91, + 295, + 118 + ], + "type": "interline_equation", + "content": "G (x) = \\exp \\left(- \\frac {1}{2} (x - \\mu) ^ {T} \\Sigma^ {- 1} (x - \\mu)\\right), \\tag {4}", + "image_path": "75180f6ee8f11a80a198001414453645937e10af5ca1b307bff8213cf2e0c3b8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "text", + "content": " denotes a spatial location in the 3D scene and " + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "text", + "content": ", initialized from Structure-from-Motion (SfM) points, specifies the Gaussian center. The covariance matrix " + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "text", + "content": " is constructed using a scaling matrix " + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "text", + "content": " and rotation matrix " + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "inline_equation", + "content": "\\Sigma = RSS^T R^T" + }, + { + "bbox": [ + 55, + 123, + 295, + 182 + ], + "type": "text", + "content": " to ensure positive semi-definiteness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 183, + 296, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 183, + 296, + 279 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 296, + 279 + ], + "type": "text", + "content": "Tile-based Rendering. Tile-based rendering is the key to achieving high-resolution and real-time performance in 3DGS. Unlike conventional volumetric methods, it projects 3D Gaussians into 2D Gaussians " + }, + { + "bbox": [ + 55, + 183, + 296, + 279 + ], + "type": "inline_equation", + "content": "G'(x)" + }, + { + "bbox": [ + 55, + 183, + 296, + 279 + ], + "type": "text", + "content": " on the image plane rather than relying on resource-intensive ray marching. The image is divided into tiles of size " + }, + { + "bbox": [ + 55, + 183, + 296, + 279 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 183, + 296, + 279 + ], + "type": "text", + "content": " to process all sorted 2D Gaussians in parallel. Each pixel within a tile applies " + }, + { + "bbox": [ + 55, + 183, + 296, + 279 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 183, + 296, + 279 + ], + "type": "text", + "content": "-blending:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 286, + 295, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 286, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 64, + 286, + 295, + 320 + ], + "type": "interline_equation", + "content": "C \\left(x ^ {\\prime}\\right) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\sigma_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\sigma_ {j}\\right), \\quad \\sigma_ {i} = \\alpha_ {i} G _ {i} ^ {\\prime} \\left(x ^ {\\prime}\\right). \\tag {5}", + "image_path": "cbbffce636be4a5b26abf2688f97d7a1af415349f32433b104d40ecf15ccc18a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "type": "text", + "content": " represents the queried pixel position, and " + }, + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "type": "text", + "content": " denotes the number of sorted 2D Gaussians linked to that tile. By leveraging the Tile-based differentiable rasterizer, all 3D Gaussians projected onto the training view can be directly optimized in an end-to-end manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 385, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 434 + ], + "type": "text", + "content": "MAE, MSE, and SSIM. These three loss functions are widely used for image quality assessment. The Mean Absolute Error (MAE) and Mean Squared Error (MSE) are two methods for calculating RGB loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 440, + 295, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 440, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 63, + 440, + 295, + 472 + ], + "type": "interline_equation", + "content": "\\mathrm {M A E} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left| y _ {i} - \\hat {y} _ {i} \\right|, \\mathrm {M S E} = \\frac {1}{n} \\sum_ {i = 1} ^ {n} \\left(y _ {i} - \\hat {y} _ {i}\\right) ^ {2}, \\tag {6}", + "image_path": "e7d97929dad1e04977ef1c6fc0fe27974deb12eb56ab0efa6cde1c0aaa233b8a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 487, + 295, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 487, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 67, + 487, + 295, + 515 + ], + "type": "interline_equation", + "content": "\\operatorname {S S I M} (x, y) = \\frac {\\left(2 \\mu_ {x} \\mu_ {y} + C _ {1}\\right) \\left(2 \\sigma_ {x y} + C _ {2}\\right)}{\\left(\\mu_ {x} ^ {2} + \\mu_ {y} ^ {2} + C _ {1}\\right) \\left(\\sigma_ {x} ^ {2} + \\sigma_ {y} ^ {2} + C _ {2}\\right)}, \\tag {7}", + "image_path": "b26781f69009c3ed36c72efa097681e71a19bacee1c57b9d7654a912cda8dcee.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\mu_{x}" + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\mu_{y}" + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": " are the means of the images " + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\sigma_x^2" + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\sigma_y^2" + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": " are the variances, " + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\sigma_{xy}" + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": " is the covariance between the images, and " + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "inline_equation", + "content": "C_1, C_2" + }, + { + "bbox": [ + 55, + 521, + 296, + 629 + ], + "type": "text", + "content": " are constants to stabilize the division when the denominator is small. MAE measures the average absolute error, treating all deviations equally. MSE squares each error before averaging, making it more sensitive to large errors and outliers. The Structural Similarity (SSIM) focuses on perceptual quality by comparing structural patterns, luminance, and contrast between images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 635, + 230, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 230, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 230, + 649 + ], + "type": "text", + "content": "3.2. Tile-based Training via RT-Loss" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "In this section, we introduce RT-Loss and our Tile-based training paradigm. Our goal is to incorporate random viewpoint observations within each training batch, while preserving structural information to enforce more comprehensive constraints. In contrast to the individual rays used" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": "in NeRFs, Gaussian Splatting utilizes a Tile-based splatting process that projects Gaussians onto the entire view screen, rather than performing active ray marching. Hence, the smallest unit in the loss calculation must align with the splatting rendering and is set as a " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " tile " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ". We load the datasets as tiles instead of images, randomly sample " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " tiles as a training batch " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " and obtain the rendered tiles set " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\hat{\\mathcal{C}})" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " along with their corresponding ground truth " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\mathcal{C})" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}} = \\{\\hat{\\mathcal{C}}(t) \\mid t \\in \\mathcal{T}\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\mathcal{C}(t) \\mid t \\in \\mathcal{T}\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 193 + ], + "type": "text", + "content": ". We use MAE as the color loss, which can be written as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 349, + 200, + 555, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 200, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 349, + 200, + 555, + 228 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {R T} - \\mathrm {M A E}} = \\frac {1}{\\| \\mathcal {T} \\|} \\sum_ {\\boldsymbol {t} \\in \\mathcal {T}} \\| \\hat {\\boldsymbol {C}} (\\boldsymbol {t}) - \\boldsymbol {C} (\\boldsymbol {t}) \\| _ {1}. \\tag {8}", + "image_path": "b963190ec3a14ed0344eb3f7ee2259e623b7db709bf519d1270117e81fccfc2f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "text", + "content": "The tiles contain viewpoint's structural information. If we solely rely on rgb loss " + }, + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{MAE}}" + }, + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "text", + "content": ", the positional relationship contained within the tiles would be completely lost [42]. Inspired by S3IM's adaptation of SSIM on random rays, we propose RT-SSIM for 3DGS to retain the structural information embedded in random tiles. The concept is straightforward: while keeping the total batch size unchanged, we compute and average the SSIM across multiple independent and non-overlapping random tiles from different views, with kernel size and stride set to " + }, + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "text", + "content": " (e.g., " + }, + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "inline_equation", + "content": "9 \\times 9" + }, + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "text", + "content": " for overall structtrue or " + }, + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "text", + "content": " for fine detail) with a stride of " + }, + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 236, + 556, + 380 + ], + "type": "text", + "content": ". RT-SSIM can be written as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 385, + 554, + 428 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 385, + 554, + 428 + ], + "spans": [ + { + "bbox": [ + 315, + 385, + 554, + 428 + ], + "type": "interline_equation", + "content": "\\operatorname {R T - S S I M} (\\hat {\\mathcal {T}}, \\mathcal {T}) = \\frac {1}{N} \\sum_ {\\boldsymbol {t} \\in \\mathcal {T}, v \\in \\boldsymbol {V}} \\operatorname {S S I M} \\left(\\boldsymbol {t} ^ {(v)} (\\hat {\\mathcal {C}}), \\boldsymbol {t} ^ {(v)} (\\mathcal {C})\\right), \\tag {9}", + "image_path": "f00dd6f8af8ef74150a7ed8959db5fc485ee1f7d3a6916d57b3bc141001c6c12.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 436, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 554, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 436, + 554, + 495 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 436, + 554, + 495 + ], + "type": "text", + "content": " controls the number of views in all " + }, + { + "bbox": [ + 313, + 436, + 554, + 495 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 436, + 554, + 495 + ], + "type": "text", + "content": " tiles. In particular, when " + }, + { + "bbox": [ + 313, + 436, + 554, + 495 + ], + "type": "inline_equation", + "content": "V = N" + }, + { + "bbox": [ + 313, + 436, + 554, + 495 + ], + "type": "text", + "content": ", it implies all tiles are selected from different views. As RT-SSIM positively correlated with image similarity, we define the structure loss " + }, + { + "bbox": [ + 313, + 436, + 554, + 495 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{RT - SSIM}}" + }, + { + "bbox": [ + 313, + 436, + 554, + 495 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 339, + 504, + 553, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 504, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 339, + 504, + 553, + 519 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {R T} - \\mathrm {S S I M}} (\\Theta , \\mathcal {T}) = 1 - \\mathrm {R T} - \\mathrm {S S I M} (\\hat {\\mathcal {T}}, \\mathcal {T}) \\tag {10}", + "image_path": "de759b97f4c94724490aea88175e16ea5a140d82b9a4b95261843118a66f41be.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 527, + 554, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 554, + 574 + ], + "type": "text", + "content": "We note that tiles are particularly suitable for parallel computation, thus the additional computational cost of Tile-based training is minimal. The final hybrid loss " + }, + { + "bbox": [ + 313, + 527, + 554, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RT}}" + }, + { + "bbox": [ + 313, + 527, + 554, + 574 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 323, + 584, + 553, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 584, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 323, + 584, + 553, + 597 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R T}} (\\Theta , \\mathcal {T}) = (1 - \\lambda_ {s}) L _ {\\mathrm {R T} - \\mathrm {M A E}} + \\lambda_ {s} L _ {\\mathrm {R T} - \\mathrm {S S I M}} \\tag {11}", + "image_path": "ce3c59a53c4370a1025027e515786843616cda880ad583c8656d4ed6ba6cd893.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 497, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 497, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 497, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 497, + 616 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 313, + 605, + 497, + 616 + ], + "type": "text", + "content": " balances structure and color losses." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": "We present the pseudocode in Algorithm 1. In the standard 3DGS training paradigm, tiles are rendered in parallel using the same view parameters. In contrast, we sample tiles from multiple different views to integrate multi-view constraints. In our random-tile training paradigm, We introduce two additional hyperparameters, " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " controls the loss weight and " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": " determines the number of viewpoints. When " + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "inline_equation", + "content": "V = N" + }, + { + "bbox": [ + 313, + 618, + 555, + 713 + ], + "type": "text", + "content": ", each tile is sampled from a" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26925" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 56, + 74, + 294, + 495 + ], + "blocks": [ + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "lines": [ + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": "Algorithm 1: Tile-based Training via RT-loss \n1 Let " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " be an SGD training algorithm; \n2 while the termination criteria has not been satisfied do \n3 Sample a training batch " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " consisting of " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " tiles from " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " views in dataset " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": "; \n4 for " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " do \n5 Randomly select tiles " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_i = [t_1, t_2 \\dots t_{V/N}]" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "t_j \\in \\mathcal{T}_i \\subseteq \\mathcal{T}" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": "; \n6 VisibilityCheck(); \n7 end \n8 for each tile " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " do \n9 Compute the forward splating result after " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}} = \\{\\hat{C}(t) \\mid t \\in \\mathcal{T}\\}" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": "; \n10 Obtain the ground-truth " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{C(t) \\mid t \\in \\mathcal{T}\\}" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": "; \n11 end \n12 for " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " do \n13 let rendered pixels " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}}" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " form the pixel patch " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\pmb{t}^{(v)}(\\hat{\\mathcal{C}})" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " in view " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "V_i" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": "; \n14 Compute SSIM " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "(\\pmb{t}^{(v)}(\\hat{\\mathcal{C}}), \\pmb{t}^{(v)}(\\mathcal{C}))" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " with the given kernel " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "K \\times K" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": "; \n15 end \n16 Obtain the RT-SSIM loss " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{RT - SSIM}}(\\Theta, \\mathcal{T}) = 1 - \\frac{1}{N} \\sum \\mathrm{SSIM}(\\pmb{t}^{(v)}(\\hat{\\mathcal{C}}), \\pmb{t}^{(v)}(\\mathcal{C}))" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": "; \n17 Obtain the conventional " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "L_{MAE}" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " loss as color loss " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{RT - MAE}}(\\Theta) = \\frac{1}{||\\mathcal{T}||} \\sum_{t \\in \\mathcal{T}} ||\\hat{\\mathcal{C}}(t) - C(t)||_1" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": "; \n18 " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\Theta, \\mathcal{T}) = (1 - \\lambda_s)L_{\\mathrm{RT - MAE}}(\\Theta, \\mathcal{T}) + \\lambda_s L_{\\mathrm{RT - SSIM}}(\\Theta, \\mathcal{T})" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": "; \n19 Compute the gradient " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\nabla \\mathcal{L}(\\Theta)" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": "; \n20 Update model parameters " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 56, + 74, + 294, + 495 + ], + "type": "text", + "content": " \n21 end" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_body" + } + ], + "index": 0, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 519, + 296, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 627 + ], + "type": "text", + "content": "distinct view, and their visibility is computed in parallel during preprocessing to accelerate the rendering process. For a fair comparison, we set the total tiles count " + }, + { + "bbox": [ + 55, + 519, + 296, + 627 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 519, + 296, + 627 + ], + "type": "text", + "content": " to match that in image-wise training, ensuring a consistent batch size in gradient descent algorithm " + }, + { + "bbox": [ + 55, + 519, + 296, + 627 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 519, + 296, + 627 + ], + "type": "text", + "content": ". Our random-tile strategy uniformly samples tiles across each view, balancing observations for each perspective. Our ablation study in Sec. 4.4 shows that this strategy achieves slightly better results than purely random selection." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 635, + 249, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 249, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 249, + 648 + ], + "type": "text", + "content": "3.3. Tile-based Adaptive Density Control" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "In 3D Gaussian Splating, a Gaussian is split or cloned depending on whether the average cumulative 2D gradient on the NDC coordinates " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sum \\| \\mathbf{g}_i\\|" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " across interval iterations exceeds a specified threshold " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{densify}}" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": ", as shown in Fig. 2. During our Tile-based training, Gaussians in one iteration will" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 322, + 79, + 411, + 251 + ], + "blocks": [ + { + "bbox": [ + 322, + 79, + 411, + 251 + ], + "lines": [ + { + "bbox": [ + 322, + 79, + 411, + 251 + ], + "spans": [ + { + "bbox": [ + 322, + 79, + 411, + 251 + ], + "type": "image", + "image_path": "a076c6dcc119548e866a1570437a3f7d52a44658c2e20de84f1f8e9ebbecbea8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 268, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 323 + ], + "type": "text", + "content": "Figure 2. Image-wise vs Tile-wise Adaptive Density Control. In Tile-based training, NDC gradients for densification are computed as a weighted sum across all observations, balancing multi-view contributions, where " + }, + { + "bbox": [ + 313, + 268, + 555, + 323 + ], + "type": "inline_equation", + "content": "S1" + }, + { + "bbox": [ + 313, + 268, + 555, + 323 + ], + "type": "text", + "content": " represents the area of the Gaussian projection onto the tile, and " + }, + { + "bbox": [ + 313, + 268, + 555, + 323 + ], + "type": "inline_equation", + "content": "S2" + }, + { + "bbox": [ + 313, + 268, + 555, + 323 + ], + "type": "text", + "content": " is the total projected area." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 416, + 82, + 539, + 251 + ], + "blocks": [ + { + "bbox": [ + 416, + 82, + 539, + 251 + ], + "lines": [ + { + "bbox": [ + 416, + 82, + 539, + 251 + ], + "spans": [ + { + "bbox": [ + 416, + 82, + 539, + 251 + ], + "type": "image", + "image_path": "bb3fca209b35109499f4ef69e35f33f058d8d5b8d800197aec5a16773fef25d3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 342, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 389 + ], + "type": "text", + "content": "have multiple NDC gradients from different training views. Thus, there are two methods for computing density gradients: Tile-Count Densification and Iteration-Count Densification." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 368, + 394, + 554, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 394, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 368, + 394, + 554, + 426 + ], + "type": "interline_equation", + "content": "\\frac {\\sum_ {k = 1} ^ {\\text {I t e r}} \\sum_ {m = 1} ^ {M _ {k}} \\| \\mathbf {g} _ {k m} \\|}{\\sum_ {k = 1} ^ {\\text {I t e r}} M _ {k}} > \\tau_ {\\text {d e n s i f y}} \\tag {12}", + "image_path": "6726145a5f4f02b8f833472ce172d0029a3361d9c9f4954389b4ddb7219455b8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 351, + 438, + 553, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 438, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 351, + 438, + 553, + 471 + ], + "type": "interline_equation", + "content": "\\frac {\\sum_ {k = 1} ^ {\\text {I t e r}} \\sum_ {m = 1} ^ {M _ {k}} \\left(\\left\\| \\mathbf {g} _ {k m} \\right\\| \\times r _ {m}\\right)}{\\sum_ {k = 1} ^ {\\text {I t e r}} 1} > \\tau_ {\\text {d e n s i f y}}, \\tag {13}", + "image_path": "130e2e642fd1271814f2ea30283da7fb0b26a7301d61ffba8e51f28fb6f11f00.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 475, + 555, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 475, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\text{Iter}" + }, + { + "bbox": [ + 313, + 475, + 555, + 642 + ], + "type": "text", + "content": " represents the number of iterations in which a Gaussian participates in " + }, + { + "bbox": [ + 313, + 475, + 555, + 642 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 475, + 555, + 642 + ], + "type": "text", + "content": "-blending within each densification interval. In Tile-count Densification ( Eq. (12)), the NDC gradient is averaged over the number of observed tiles " + }, + { + "bbox": [ + 313, + 475, + 555, + 642 + ], + "type": "inline_equation", + "content": "M_{k}" + }, + { + "bbox": [ + 313, + 475, + 555, + 642 + ], + "type": "text", + "content": ", while in Iteration-Count Densification ( Eq. (13)), it is averaged over the number of observed iterations. Additionally, in Iteration-Count Densification, the NDC gradient is weighted by " + }, + { + "bbox": [ + 313, + 475, + 555, + 642 + ], + "type": "inline_equation", + "content": "r_{m}" + }, + { + "bbox": [ + 313, + 475, + 555, + 642 + ], + "type": "text", + "content": ", the ratio of the projected area shown in Fig. 2. To maintain consistency with the split parameters used in the original 3DGS and to emphasize the effect of RT-Loss, we default to using Iteration-Count for adaptive density control. Further details on NDC gradients and densification algorithm can be found in the Supplementary Material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 672, + 430, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 430, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 430, + 685 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Dataset and Metrics. We conduct a comprehensive evaluation on 26 publicly available scenes widely used" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26926" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 92, + 102, + 519, + 207 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 1. Quantitative comparison to previous methods on benchmark datasets. We report results on three commonly used datasets. The best, second best, and third best results are denoted by red, orange, and yellow, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 102, + 519, + 207 + ], + "lines": [ + { + "bbox": [ + 92, + 102, + 519, + 207 + ], + "spans": [ + { + "bbox": [ + 92, + 102, + 519, + 207 + ], + "type": "table", + "html": "
Dataset\nMethod | MetricsMip-NeRF360Tanks&TemplesDeep Blending
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Plenoxels [13]23.620.6700.44322.130.7620.33423.060.7950.510
Instant-NGP [34]26.430.7250.33922.940.7520.30423.620.7970.423
Mip-NeRF360 [3]29.230.8440.20723.120.7710.25029.400.9010.245
3D-GS [19]28.690.8700.18224.360.8310.21329.410.9030.243
Scaffold-GS [31]28.840.8480.22024.810.8400.20530.210.9060.254
Ours29.660.8920.17125.130.8520.19230.150.9110.224
Ours (+ Scaffold-GS)29.860.8850.20625.320.8570.18930.490.9130.231
", + "image_path": "07180de8892f5710ab3ca20ed72c83f68a7e12926427b9bf384c7b2c2a62672d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 65, + 220, + 549, + 502 + ], + "blocks": [ + { + "bbox": [ + 65, + 220, + 549, + 502 + ], + "lines": [ + { + "bbox": [ + 65, + 220, + 549, + 502 + ], + "spans": [ + { + "bbox": [ + 65, + 220, + 549, + 502 + ], + "type": "image", + "image_path": "1ba36238fca19915e7c3b35d8f16d1ed327a9c034b61561d95b3cccd8f40be24.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 505, + 555, + 551 + ], + "lines": [ + { + "bbox": [ + 55, + 505, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 555, + 551 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison across benchmark datasets [3, 11, 21]. We highlight the differences in the image using patches for clearer visibility. Our approach consistently outperforms 3D-GS on these scenes, demonstrating clear advantages in challenging scenarios, such as fine-scale details (MIP360-ROOM(a), TANDT-BARN), thin geometry (SMERF-NYC(a), SMERF-NYC(b)), texture-less regions (TANDT-TRAIN), light effects (MIP360-COUNTER, TANDT-TRUCK), and insufficient observations (TANDT-BARN, SMERF-NYC(b))." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "content": "in the field of 3D reconstruction across multiple diverse tasks. Specifically, we evaluate our method on seven scenes from Mip-NeRF360 [3], another seven from Tanks&Temples [21], and two additional scenes from DeepBlending [16]. To further demonstrate the limitations of image-wise loss in capturing global structure, our evaluation incorporated all four scenes from [4], with " + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "inline_equation", + "content": "5 - 10 \\times" + }, + { + "bbox": [ + 55, + 560, + 297, + 704 + ], + "type": "text", + "content": " more training views. We additionally evaluated dynamic benchmarks, specifically using the Neural 3D Video Dataset containing six scenes, to demonstrate the effectiveness of RT-Loss in 4D reconstruction. In addition to the commonly used metrics (PSNR, SSIM [39], and LPIPS [49]), we fur" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 560, + 556, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 556, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 556, + 633 + ], + "type": "text", + "content": "ther provide the storage size (MB) and rendering speed (FPS) to assess model compactness and efficiency, with all experiments performed on an NVIDIA RTX 4090 GPU. The main paper reports the average metrics across all scenes in each dataset, while detailed per-scene quantitative results are presented in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Baseline and Implementation. 3DGS[19], Scaffold-GS[31] and 4D-GS [10] are selected as our baselines due to their state-of-the-art performance. We also present the results of Mip-NeRF360 [3], iNGP [34] and Plenoxels [13] as in [19]. To demonstrate the generality of the tile-based ap" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26927" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 225, + 220 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 225, + 220 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 225, + 220 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 225, + 220 + ], + "type": "image", + "image_path": "13d751e8e657b1770978728860ef8a278143b67ab58d0941ecd1adf8ab9a47c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 235, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 54, + 235, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 235, + 555, + 269 + ], + "type": "text", + "content": "Figure 4. Convergence performance. We evaluated the convergence of PSNR for both the training and testing views during the training process. With the same batch size, the convergence of image-wise loss in 3DGS is slower than RT-Loss, with the gap widening on larger datasets, e.g., COURTHOUSE and BERLIN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 71, + 383, + 224 + ], + "blocks": [ + { + "bbox": [ + 225, + 71, + 383, + 224 + ], + "lines": [ + { + "bbox": [ + 225, + 71, + 383, + 224 + ], + "spans": [ + { + "bbox": [ + 225, + 71, + 383, + 224 + ], + "type": "image", + "image_path": "022bc578cbbd7af7a70a6439ccf4adf392cfcec02a7324e64f459960cee6d63b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 383, + 72, + 545, + 224 + ], + "blocks": [ + { + "bbox": [ + 383, + 72, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 383, + 72, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 383, + 72, + 545, + 224 + ], + "type": "image", + "image_path": "1ba3d1e74b5577f324a2ae4ad3745555784ef85925c3e046fbfd0224e204d501.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "text", + "content": "proach, we implemented the tile-wise method in Scaffolds for comparison. We trained for " + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "inline_equation", + "content": "30k" + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "text", + "content": " iterations and rendered the same number of pixels per iteration, ensuring the total pixels from random tiles matched a complete image. To independently demonstrate the effectiveness of the tile-based training paradigm and RT-loss, we modified only the sampling and loss calculations for random tile training while keeping the default 3DGS parameters unchanged, including the densification threshold " + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{densify}} = 0.0002" + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "text", + "content": " and all learning rates. For RT-loss, we sample random tiles from " + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "inline_equation", + "content": "V = 5" + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "text", + "content": " views and apply the same structure loss weight, " + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\lambda_s = 0.2" + }, + { + "bbox": [ + 54, + 289, + 296, + 434 + ], + "type": "text", + "content": ", as used in the original 3DGS." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 443, + 155, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 155, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 155, + 455 + ], + "type": "text", + "content": "4.2. Results Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 461, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 640 + ], + "type": "text", + "content": "Benchmark Evaluation. The quantitative results on Mip-NeRF360 and Tanks&Temples are presented in Tab. 1. Results show that we outperform existing methods on standard datasets. Fig. 3 illustrates that tile-wise training achieves enhanced visual quality, with improved geometry accuracy and richer texture details. Moreover, our method retains the efficiency of Gaussian methods, as shown in Tab. 2. We compared the rendering speed and storage size with 3DGS. We achieve improved performance while using a comparable or fewer number of Gaussians, demonstrating high-quality reconstruction without compromising speed. By incorporating our tile-based approach, Scaffold-GS also achieved notable improvements, suggesting that the image-wise training paradigm, which lacks multi-view constraints, limits the reconstruction quality of various GS methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Convergence Performance. We evaluated our convergence performance across several representative scenarios. As shown in Fig. 4, compared to the original image-wise loss in 3DGS, the RT-Loss based on multi-view observations provides more effective constraints, resulting in significantly faster convergence. The difference in convergence speed is" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 319, + 553, + 405 + ], + "blocks": [ + { + "bbox": [ + 313, + 288, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 309 + ], + "type": "text", + "content": "Table 2. Rendering FPS and storage comparison. The rendering speed of both methods is measured on the same machine." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 319, + 553, + 405 + ], + "lines": [ + { + "bbox": [ + 317, + 319, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 317, + 319, + 553, + 405 + ], + "type": "table", + "html": "
ScenesMip-NeRF360ROOMKITCHEN
MemFPSPSNRMemFPSPSNRMemFPS
3DGS6717831.7931712331.63372118
Ours6259332.8131812532.35320135
ScenesTanks&TemplesCourthOUSEMEETINGROOM
MemFPSPSNRMemFPSPSNRMemFPS
3DGS31712623.1317414626.27276120
Ours30213024.1416813926.98260133
", + "image_path": "f93cf6752b9e3190523c6d2e452b51aa8c281bc20dc6227099309ea9d59fabe9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 479, + 552, + 530 + ], + "blocks": [ + { + "bbox": [ + 313, + 417, + 555, + 473 + ], + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 473 + ], + "type": "text", + "content": "Table 3. PSNR evaluation for scenes with increased training views. The convergence capability of RT-Loss is particularly evident on the SMERF dataset[11], which includes approximately 10 times more training views than Mipnerf360. Additionally, we recorded the PSNR gap at 4k and 30k iterations." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 479, + 552, + 530 + ], + "lines": [ + { + "bbox": [ + 317, + 479, + 552, + 530 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 552, + 530 + ], + "type": "table", + "html": "
Scene.views test iterationAlameda(1734)Berlin(1511)London(1874)
4k30k4k30k4k30k
3DGS17.6222.2722.6327.5322.3225.75
Ours20.1123.91624.8028.7924.0726.626
", + "image_path": "a11e154853d5f07d396f858ba0a7eaa866b1f966bf3c4e0135290509b9b87b60.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 593, + 552, + 632 + ], + "blocks": [ + { + "bbox": [ + 313, + 542, + 555, + 586 + ], + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 586 + ], + "type": "text", + "content": "Table 4. Quantitative comparison on dynamic datasets. We applied RT loss for dynamic scene reconstruction and compared it with 4D-GS[10] on the challenging real-world dynamic dataset Neu3D[23]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 593, + 552, + 632 + ], + "lines": [ + { + "bbox": [ + 317, + 593, + 552, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 552, + 632 + ], + "type": "table", + "html": "
ModelPSNR↑D-SSIM↓LPIPS↓FPS↑Storage (MB)↓
4D-GS[10]31.150.9740.0493090
Ours31.740.9730.0433478
", + "image_path": "33b16967fda4dde009619f50b83af835e695bf9cb425e10db153f6ce5605c6b3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 653, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 689 + ], + "type": "text", + "content": "visually evident in Fig. 1. This advantage in convergence is even more pronounced in scenes with a large number of training views, e.g., COURTHOUSE and BERLIN." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Scene with More Training Views. We compared our method with the original 3DGS in scenarios with more" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26928" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 193 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 193 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 193 + ], + "type": "image", + "image_path": "90541314fd9e72cfe853e4ded18acf5f31110dd2a786fd58303c0c1a8a5f0ea0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 295, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 246 + ], + "type": "text", + "content": "Figure 5. Gaussians structure comparison. RT-Loss demonstrates stronger structural reconstruction ability in scenes with more training views. In contrast, 3DGS not only loses details but also fails to recover the geometric structure of the scene." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 261, + 293, + 381 + ], + "blocks": [ + { + "bbox": [ + 58, + 261, + 293, + 381 + ], + "lines": [ + { + "bbox": [ + 58, + 261, + 293, + 381 + ], + "spans": [ + { + "bbox": [ + 58, + 261, + 293, + 381 + ], + "type": "image", + "image_path": "12290bde3171d39c0a76a115e98ef6a414861af77d76583485dc54b62c29df93.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 391, + 295, + 424 + ], + "lines": [ + { + "bbox": [ + 55, + 391, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 295, + 424 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons on the Neural 3D Video Dataset. In 4D reconstruction, constraints from multiple viewpoints and time also help generate more accurate details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 449, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 449, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 295, + 663 + ], + "type": "text", + "content": "training views. In Tab. 3, our method achieved a significant improvement in metrics. The gap is particularly noticeable in the early training stages (4k iterations), as image-wise training naturally requires more iterations to fully capture the scene. Fig. 5 further illustrates that our method produces a more refined geometric structure on large datasets. This supports our hypothesis about the current training paradigm: as the number of training views increases, image-wise training struggles to comprehensively capture scene information, resulting in less reliable optimization and densification. Furthermore, through experiments with extended training durations, we demonstrate that naively increasing the iteration count for image-wise training does not effectively address this limitation. This is because once incorrect geometry is formed during Gaussian optimization, it is difficult to correct by simply extending the training process. For further details, please refer to the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Dynamic Scenes. We evaluated our method on the real-world dynamic scene dataset, the Neural 3D Video Dataset, using 4D-GS [10] as the baseline. The expansion of the time dimension in the dynamic dataset resulted in a larger" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 99, + 550, + 171 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 5. Ablation study on RT-Loss components. 'Img-' refers to image-wise loss, and 'RT-' refers to random-tile loss." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 99, + 550, + 171 + ], + "lines": [ + { + "bbox": [ + 316, + 99, + 550, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 99, + 550, + 171 + ], + "type": "table", + "html": "
Scene(views)KITCHEN(279)CORTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
LImg-MAE + LImg-SSIM (3DGS)31.6737223.11173
LImg-MAE31.0913121.3544
LRT-MAE31.6512722.6363
LRT-MAE + LImg-SSIM32.0235223.50186
LRT-MAE + LRT-SSIM (Ours)32.3532024.15168
", + "image_path": "adfb8d7efd72ac251a2e2a69e964aaa8c80e7ed193171c1133863bddcdc94c30.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 221, + 539, + 280 + ], + "blocks": [ + { + "bbox": [ + 313, + 182, + 555, + 215 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 215 + ], + "type": "text", + "content": "Table 6. Effect of Adaptive Density Control. 'Tile-Count' and 'Iteration-Count' are densification strategies described in Sec. 3.3, with 'Iteration-Count' being the default." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 221, + 539, + 280 + ], + "lines": [ + { + "bbox": [ + 326, + 221, + 539, + 280 + ], + "spans": [ + { + "bbox": [ + 326, + 221, + 539, + 280 + ], + "type": "table", + "html": "
Scene(views)KICHEN(279)COURTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
3DGS w/o Densification29.845822.96118
Ours w/o Densification30.325823.83118
Ours w/ Tile-Count31.9517523.56125
Ours w/ Iteration-Count32.3532024.15168
", + "image_path": "f93a3f79901d2859bca8764b8f8f7216b3f9b7f8c54bbbd10c8be1b5adb09ad9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 327, + 353, + 539, + 441 + ], + "blocks": [ + { + "bbox": [ + 313, + 291, + 555, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 346 + ], + "type": "text", + "content": "Table 7. Effect of Tile Sampling and Batch Training. 'V' refers to the number of randomly selected viewpoints, determining the spread of observations. 'B' refers to Batch train, which means using multiple images as a batch for training. 'Non-Uniform' and 'Uniform' indicate whether the tiles are distributed evenly." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 327, + 353, + 539, + 441 + ], + "lines": [ + { + "bbox": [ + 327, + 353, + 539, + 441 + ], + "spans": [ + { + "bbox": [ + 327, + 353, + 539, + 441 + ], + "type": "table", + "html": "
Scene.views)KICHEN(279)COURTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
tile-wise V = 231.8034323.76182
tile-wise V = 432.2233723.92175
tile-wise V = 632.4033224.29161
image-wise B = 231.6735223.82214
image-wise B = 431.2234823.69207
Non-Uniform Sample32.3031024.02192
Uniform Sample32.3832024.14168
", + "image_path": "74d7ed9d12947bf4430e6336044e8680b88e0cb0434ec152ac5a5b277a0502ef.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 461, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 628 + ], + "type": "text", + "content": "volume of training data. Although most 4D Gaussian methods address this issue by using batch training and rendering multiple images together as a training batch, our results show that, with the same batch size, tile-based training provides more effective constraints across multiple viewpoints and time frames. The results show that RT-Loss is also effective for dynamic scene reconstruction. As depicted in Tab. 4, we notably improved the rendering performance on dynamic scenes without sacrificing efficiency. As shown in Fig. 6, our method reconstructs finer details, attributed to our randomized approach, which provides stronger constraints across both viewpoints and time frames compared to image-wise training. Details are provided in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 635, + 416, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 416, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 416, + 647 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "Effect of RT-Loss components. We conducted an ablation on the loss components of Eq. (11), which includes the base color loss " + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{RT - MAE}}" + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": " and the structural loss " + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{RT - SSIM}}" + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": ". As shown in the " + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": " row in Tab. 5, when only MAE loss is used, the PSNR significantly decreases, with a lower" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26929" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 76, + 136, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 76, + 136, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 76, + 136, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 76, + 136, + 194 + ], + "type": "image", + "image_path": "2141aae8c323a480c32e98b5946b71910c7133473a04e988f9195bc5cafec40d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 295, + 260 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 260 + ], + "type": "text", + "content": "Figure 7. Effect of initialization on reconstruction quality. (Left) Gaussians optimized from SfM point initialization. (Right) Gaussians optimized from random points. Our method reconstructs better structures and finer details even with random initialization, demonstrating superior robustness to initialization issues." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 138, + 76, + 216, + 194 + ], + "blocks": [ + { + "bbox": [ + 138, + 76, + 216, + 194 + ], + "lines": [ + { + "bbox": [ + 138, + 76, + 216, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 76, + 216, + 194 + ], + "type": "image", + "image_path": "7e2a0f2ecf5a0597b516e8951e4c3ba6c92cb1969ad768c9b32e12232e2df3eb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 216, + 76, + 293, + 194 + ], + "blocks": [ + { + "bbox": [ + 216, + 76, + 293, + 194 + ], + "lines": [ + { + "bbox": [ + 216, + 76, + 293, + 194 + ], + "spans": [ + { + "bbox": [ + 216, + 76, + 293, + 194 + ], + "type": "image", + "image_path": "c29dda0bb4a84c4a9fcaafe348d54de7ec83b886b3c270b65ecf048ce8623e56.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 325, + 293, + 388 + ], + "blocks": [ + { + "bbox": [ + 55, + 274, + 295, + 318 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 318 + ], + "type": "text", + "content": "Table 8. Sensitivity to Initialization. In the 'w/o SfM initialization' setting, Gaussians are initialized with random positions. RT-Loss demonstrates stronger robustness in addressing initialization issues." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 325, + 293, + 388 + ], + "lines": [ + { + "bbox": [ + 59, + 325, + 293, + 388 + ], + "spans": [ + { + "bbox": [ + 59, + 325, + 293, + 388 + ], + "type": "table", + "html": "
Scene(views)KICHEN(279)COURTHOUSE(1106)
PSNRMem(Mb)PSNRMem(Mb)
3DGS w/ SfM initialization31.6337223.13174
3DGS w/o SfM initialization29.8926921.77121
Ours w/ SfM initialization32.3532024.14168
Ours w/o SfM initialization31.2227223.21186
", + "image_path": "fdf1e83b03d6abf6e02e6518a5434ead7809fa29b27d9bd73a73151fa7471906.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 401, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 592 + ], + "type": "text", + "content": "tendency for Gaussian densification. However, our " + }, + { + "bbox": [ + 55, + 401, + 295, + 592 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{RT-MAE}}" + }, + { + "bbox": [ + 55, + 401, + 295, + 592 + ], + "type": "text", + "content": " still outperforms the image-wise constraints, and this effect becomes more pronounced in the CourtHouse scene with more training views, where the PSNR difference reaches up to 1.3 dB. Notably, in the Kitchen scene, the " + }, + { + "bbox": [ + 55, + 401, + 295, + 592 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{RT-MAE}}" + }, + { + "bbox": [ + 55, + 401, + 295, + 592 + ], + "type": "text", + "content": " constraint achieves PSNR on par with 3DGS, despite using only one-third of the Gaussian points, as shown in the " + }, + { + "bbox": [ + 55, + 401, + 295, + 592 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 55, + 401, + 295, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 401, + 295, + 592 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 55, + 401, + 295, + 592 + ], + "type": "text", + "content": " row in Tab. 5. Additionally, incorporating SSIM improves the PSNR as the number of Gaussians increases. Furthermore, to demonstrate the improvement of structure loss brought by RT-SSIM, we rendered the complete image and incorporated image-wise SSIM as a structural loss during random tile training. In contrast, the metrics using RT-SSIM consistently yield higher values, showing that the multi-view observations in RT-SSIM provide more accurate reconstruction constraints than image-wise SSIM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Effect of Adaptive Density Control. We ablate our densification strategy described in Sec. 3.3 on Tab. 6. Without any densification, the optimization of GS relies solely on the loss function without any splitting or cloning. Our improvements demonstrate the effectiveness of multi-view constraints in enhancing performance. Tile-Count densification, which averages the NDC gradient across tiles, causes the Gaussians to resist densifying and lower PSNR. This is because random-tile training increases the frequency with which a Gaussian is observed, leading to a smaller av" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 73, + 376, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 376, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 376, + 83 + ], + "type": "text", + "content": "erage gradient." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 84, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 275 + ], + "type": "text", + "content": "Effect of Tile Sampling and Batch Training. We analyzed the impact of viewpoint count " + }, + { + "bbox": [ + 313, + 84, + 553, + 275 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 84, + 553, + 275 + ], + "type": "text", + "content": " and tile uniformity described in Sec. 3.2 on Tab. 7. Uniform sampling better distributes viewpoint constraints in the loss, yielding slight improvements and serving as our default. In data-rich scenes like Courthouse, more viewpoints provide greater benefits, as larger scenes require sufficient and evenly distributed observations. This highlights the adaptability of our random-tile training paradigm to different scene requirements. We also tested the comparative effect of image-wise batch training. Under a similar total training volume (iterations = 15k), introducing multiple images as a batch to incorporate multi-view observations does not improve quality as effectively as the tile-wise approach. Moreover, our observations indicate that such dense inclusion of multiple images significantly increases computational overhead." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 275, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 396 + ], + "type": "text", + "content": "Sensitivity to Initialization. The initialization problem has long been considered a critical factor impacting 3DGS performance [18]. We conduct experiments under two initialization conditions. Fig. 7 shows that our method recovers improved structure and appearance even with random initialization. As shown in Tab. 8, our method demonstrates greater robustness to initialization variations. This suggests that part of the initialization issue may arise from uneven and insufficient training constraints, and RT-Loss has the potential to address this limitation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 401, + 468, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 468, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 468, + 412 + ], + "type": "text", + "content": "4.4. Discussions and Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 418, + 555, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 550 + ], + "type": "text", + "content": "We propose a practical and effective approach by incorporating tile-wise constraints into Gaussian optimization to replace image-wise constraints. We have demonstrated its clear effectiveness and universality through both the main experiments and ablation studies. However, in experiments with large texture-less regions, our method fails to improve the geometry. To further enhance the tile-wise optimization, finer strategies could be explored in practice, such as further allocating multi-view constraints to identical content, in order to provide more accurate and uniform multi-view constraints for improvement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 388, + 571 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 579, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 554, + 652 + ], + "type": "text", + "content": "In this work, we introduce a Tile-based training paradigm that integrates multi-view appearance and structural constraints. Extensive experiments demonstrate its effectiveness across various scenarios. We expect tile-wise training to hold strong potential as a promising paradigm for future research and applications in Gaussian Splatting." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 660, + 411, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 411, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 411, + 673 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 679, + 545, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 679, + 545, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 679, + 545, + 703 + ], + "type": "text", + "content": "This work was partially supported by Key R&D Program of Zhejiang Province (No. 2023C01039)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26930" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 134 + ], + "type": "text", + "content": "[1] Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day. Communications of the ACM, 54 (10):105-112, 2011. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 293, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 293, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 293, + 179 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 288, + 236 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470–5479, 2022. 2, 5, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 286, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 286, + 293 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19697–19705, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 289, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 289, + 360 + ], + "type": "text", + "content": "[5] Zhiqin Chen, Thomas Funkhouser, Peter Hedman, and Andrea Tagliasacchi. Mobilenerf: Exploiting the polygon rasterization pipeline for efficient neural field rendering on mobile architectures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16569-16578, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 362, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 294, + 396 + ], + "type": "text", + "content": "[6] Kai Cheng, Xiaoxiao Long, Kaizhi Yang, Yao Yao, Wei Yin, Yuexin Ma, Wenping Wang, and Xuejin Chen. Gaussianpro: 3d gaussian splattering with progressive propagation, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 397, + 286, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 286, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 286, + 440 + ], + "type": "text", + "content": "[7] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 442, + 275, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 275, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 275, + 474 + ], + "type": "text", + "content": "[8] Xiaobiao Du, Yida Wang, and Xin Yu. Mvgs: Multi-view-regulated gaussian splatting for novel view synthesis, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 477, + 280, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 280, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 280, + 509 + ], + "type": "text", + "content": "[9] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B. Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 511, + 292, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 292, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 292, + 554 + ], + "type": "text", + "content": "[10] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. arXiv preprint arXiv:2402.03307, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 556, + 290, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 290, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 290, + 621 + ], + "type": "text", + "content": "[11] Daniel Duckworth, Peter Hedman, Christian Reiser, Peter Zhizhin, Jean-François Thibert, Mario Lučić, Richard Szeliski, and Jonathan T Barron. Smerf: Streamable memory efficient radiance fields for real-time large-scene exploration. ACM Transactions on Graphics (TOG), 43(4): 1-13, 2024. 5, 6, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 293, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 293, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 293, + 678 + ], + "type": "text", + "content": "[12] Jan-Michael Frahm, Pierre Fite-Georgel, David Gallup, Tim Johnson, Rahul Raguram, Changchang Wu, Yi-Hung Jen, Enrique Dunn, Brian Clipp, Svetlana Lazebnik, et al. Building rome on a cloudless day. In European Conference on Computer Vision, pages 368–381, 2010. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 293, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 293, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 293, + 713 + ], + "type": "text", + "content": "[13] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 551, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 74, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 5501-5510, 2022. 5, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 547, + 149 + ], + "type": "text", + "content": "[14] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14346-14355, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 151, + 548, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 548, + 192 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 548, + 192 + ], + "type": "text", + "content": "[15] Xiang Guo, Guanying Chen, Yuchao Dai, Xiaqing Ye, Jiadai Sun, Xiao Tan, and Errui Ding. Neural deformable voxel grid for fast optimization of dynamic view synthesis, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 545, + 237 + ], + "type": "text", + "content": "[16] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 5, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 545, + 259 + ], + "type": "text", + "content": "[17] Tao Hu, Shu Liu, Yilun Chen, Tiancheng Shen, and Jiaya Jia. Efficientnerf: Efficient neural radiance fields, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 262, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 547, + 304 + ], + "type": "text", + "content": "[18] Jaewoo Jung, Jisang Han, Honggyu An, Jiwon Kang, Seonghoon Park, and Seungryong Kim. Relaxing accurate initialization constraint for 3d gaussian splatting. arXiv preprint arXiv:2403.09413, 2024. 2, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 306, + 550, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 550, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 550, + 348 + ], + "type": "text", + "content": "[19] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 5, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 539, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 539, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 539, + 371 + ], + "type": "text", + "content": "[20] Leonid Keselman and Martial Hebert. Approximate differentiable rendering with algebraic surfaces, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 372, + 542, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 542, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 542, + 415 + ], + "type": "text", + "content": "[21] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 5, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 416, + 549, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 549, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 549, + 437 + ], + "type": "text", + "content": "[22] Christoph Lassner and Michael Zollhöfer. Pulsar: Efficient sphere-based neural rendering, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 438, + 548, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 548, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 548, + 503 + ], + "type": "text", + "content": "[23] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 505, + 550, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 550, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 550, + 536 + ], + "type": "text", + "content": "[24] Yanyan Li, Chenyu Lyu, Yan Di, Guangyao Zhai, Gim Hee Lee, and Federico Tombari. Geogaussian: Geometry-aware gaussian splatting for scene rendering, 2024. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 538, + 544, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 538, + 544, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 538, + 544, + 591 + ], + "type": "text", + "content": "[25] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 593, + 537, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 537, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 537, + 624 + ], + "type": "text", + "content": "[26] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splattering for real-time dynamic view synthesis, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 625, + 551, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 551, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 551, + 669 + ], + "type": "text", + "content": "[27] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 549, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 549, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 549, + 712 + ], + "type": "text", + "content": "[28] Haotong Lin, Sida Peng, Zhen Xu, Tao Xie, Xingyi He, Hujun Bao, and Xiaowei Zhou. High-fidelity and real-time novel view synthesis for dynamic scenes. In SIGGRAPH Asia Conference Proceedings, 2023. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26931" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 293, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 293, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 293, + 116 + ], + "type": "text", + "content": "[29] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. arXiv preprint arXiv:2312.13763, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 292, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 292, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 292, + 173 + ], + "type": "text", + "content": "[30] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. arXiv preprint arXiv:2311.17061, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 291, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 291, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 291, + 229 + ], + "type": "text", + "content": "[31] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 1, 2, 5, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 279, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 279, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 279, + 274 + ], + "type": "text", + "content": "[32] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 289, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 289, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 289, + 329 + ], + "type": "text", + "content": "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 289, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 289, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 289, + 374 + ], + "type": "text", + "content": "[34] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2, 5, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 290, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 290, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 290, + 431 + ], + "type": "text", + "content": "[35] Christian Reiser, Rick Szeliski, Dor Verbin, Pratul Srinivasan, Ben Mildenhall, Andreas Geiger, Jon Barron, and Peter Hedman. Merf: Memory-efficient radiance fields for real-time view synthesis in unbounded scenes. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 275, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 275, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 275, + 476 + ], + "type": "text", + "content": "[36] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 286, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 286, + 544 + ], + "type": "text", + "content": "[37] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 287, + 577 + ], + "type": "text", + "content": "[38] Dan Wang, Xinrui Cui, Septimiu Salcudean, and Z. Jane Wang. Generalizable neural radiance fields for novel view synthesis with transformer, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 290, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 290, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 290, + 623 + ], + "type": "text", + "content": "[39] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[40] Christopher Wewer, Kevin Raj, Eddy Ilg, Bernt Schiele, and Jan Eric Lenssen. latentsplat: Autoencoding variational gaussians for fast generalizable 3d reconstruction. arXiv preprint arXiv:2403.16292, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[41] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Wang Xinggang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 700 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 545, + 116 + ], + "type": "text", + "content": "[42] Zeke Xie, Xindi Yang, Yujie Yang, Qi Sun, Yixiang Jiang, Haoran Wang, Yunfeng Cai, and Mingming Sun. S3im: Stochastic structural similarity and its unreasonable effectiveness for neural fields, 2023. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 533, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 533, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 533, + 162 + ], + "type": "text", + "content": "[43] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Real-time 360sparse view synthesis using gaussian splatting, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 552, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 552, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 552, + 228 + ], + "type": "text", + "content": "[44] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, Dahua Lin, Michael Zollhöfer, and Christian Richardt. Vr-nerf: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia 2023 Conference Papers, page 1–12. ACM, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 230, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 545, + 261 + ], + "type": "text", + "content": "[45] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-nerf: Point-based neural radiance fields, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 263, + 550, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 550, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 550, + 307 + ], + "type": "text", + "content": "[46] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. Mvpgs: Excavating multi-view priors for gaussian splatting from sparse input views. arXiv preprint arXiv:2409.14316, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 542, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 542, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 542, + 352 + ], + "type": "text", + "content": "[47] Zeyu Yang, Hongye Yang, Zijie Pan, Xiatian Zhu, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splattering. arXiv preprint arXiv 2310.10642, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 554, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 407 + ], + "type": "text", + "content": "[48] Taoran Yi, Jiemin Fang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussian splatting with point cloud priors. arXiv preprint arXiv:2310.08529, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 409, + 554, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 554, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 554, + 463 + ], + "type": "text", + "content": "[49] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 464, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 520 + ], + "type": "text", + "content": "[50] Xiaoyu Zhang, Weihong Pan, Chong Bao, Xiyu Zhang, Xiaojun Xiang, Hanqing Jiang, and Hujun Bao. Lookcloser: Frequency-aware radiance field for tiny-detail scene. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 16122-16132, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 521, + 526, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 526, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 526, + 564 + ], + "type": "text", + "content": "[51] Zheng Zhang, Wenbo Hu, Yixing Lao, Tong He, and Hengshuang Zhao. Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting. arXiv preprint arXiv:2403.15530, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 566, + 551, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 551, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 551, + 598 + ], + "type": "text", + "content": "[52] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 600, + 541, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 541, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 541, + 632 + ], + "type": "text", + "content": "[53] Zuo-Liang Zhu, Beibei Wang, and Jian Yang. Gs-ror: 3d gaussian splattering for reflective object relighting via sdf priors, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 633, + 529, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 633, + 529, + 676 + ], + "spans": [ + { + "bbox": [ + 316, + 633, + 529, + 676 + ], + "type": "text", + "content": "[54] C Lawrence Zitnick, Sing Bing Kang, Matthew Uysettadaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM TOG, 2004. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 677, + 541, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 541, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 541, + 700 + ], + "type": "text", + "content": "[55] Yiming Zuo and Jia Deng. View synthesis with sculpted neural points. arXiv preprint arXiv:2205.05869, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26932" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/eaeeea1f-eecf-4d2e-8480-0e89bb127b12_content_list.json b/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/eaeeea1f-eecf-4d2e-8480-0e89bb127b12_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3c1bda2702f2f7a5dc25c2df30d1df09b383dd3f --- /dev/null +++ b/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/eaeeea1f-eecf-4d2e-8480-0e89bb127b12_content_list.json @@ -0,0 +1,1479 @@ +[ + { + "type": "text", + "text": "Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images", + "text_level": 1, + "bbox": [ + 111, + 128, + 888, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Elena Buglakova1, Anwai Archit2, Edoardo D'Imprime1,4, Julia Mahamid1, Constantin Pape2,3, Anna Kreshuk1", + "bbox": [ + 199, + 202, + 795, + 238 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ European Molecular Biology Laboratory, Heidelberg", + "$^{2}$ Institute of Computer Science, University of Göttingen" + ], + "bbox": [ + 328, + 238, + 668, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Cluster of Excellence 'Multiscale Bioimaging: from Molecular Machines to Networks of Excitable Cells' (MBExC), Georg-August-University Göttingen", + "bbox": [ + 179, + 273, + 815, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ IRCCS Humanitas Research Hospital, via Manzoni 56, 20089 Rozzano, Milan, Italy", + "bbox": [ + 238, + 309, + 751, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 361, + 326, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Segmentation of very large images is a common problem in microscopy, medical imaging or remote sensing. The problem is usually addressed by sliding window inference, which can theoretically lead to seamlessly stitched predictions. However, in practice many of the popular pipelines still suffer from tiling artifacts. We investigate the root cause of these issues and show that they stem from the normalization layers within the neural networks. We propose indicators to detect normalization issues and further explore the trade-offs between artifact-free and high-quality predictions, using three diverse microscopy datasets as examples. Finally, we propose to use BatchRenorm as the most suitable normalization strategy, which effectively removes tiling artifacts and enhances transfer performance, thereby improving the reusability of trained networks for new datasets.", + "bbox": [ + 89, + 393, + 483, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 648, + 220, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Segmentation is an essential first step of pipelines in biological image analysis, a diverse domain of computer vision that enables automated quantification for various types of microscopy images. While this domain has benefited a lot from the developments in natural image analysis, it also poses unique challenges. Investigation of biological diversity, discovery of unexpected phenotypes and development of novel imaging protocols leads to high variability of datasets. This explains the popularity of human-in-the-loop approaches to segmentation, where a small amount of data is either annotated from scratch or from the output of pretrained models [1, 22, 26], then a new model is either trained from scratch or fine-tuned. After that the next rounds of correction and training happen, until the desired quality is reached. Being able to quickly train a smaller model for an", + "bbox": [ + 89, + 672, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "unseen image type is crucial in this setting.", + "bbox": [ + 511, + 363, + 797, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Another important feature of biological images, especially for volumetric imaging, is their large size. Electron microscopy volumes for connectomics [6] or light sheet microscopy volumes [20] can reach the scale of multiple terabytes. It is routine to work with 10-100 GB images of $\\sim 1000^3$ pixels, which exceed GPU memory.", + "bbox": [ + 511, + 380, + 905, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As shown in Fig. 1, during training random patches are sampled from the annotated images and combined into a batch. During inference/evaluation, the image is split into a grid of tiles which are predicted independently and then stitched back to form the full prediction. Here we refer both to 2D and 3D image patches as tiles. In this setup the size of the tiles is limited from above by GPU memory and from below by the receptive field of the network [21]. Typical tile size is around $96 \\times 96 \\times 96 - 256 \\times 256 \\times 256$ pixels and the training batch size is 2-3 tiles per batch. During inference it is not necessary to store gradients, therefore either tile size or batch size can be larger.", + "bbox": [ + 511, + 473, + 905, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tile-wise prediction and stitching might look like a purely technical step, however for microscopy patch sampling and stitching can significantly affect prediction quality and create artifacts. In contrast, big natural image datasets, such as ImageNet [5], CityScapes [4] and SAM dataset [14], usually contain relatively small 2D images, making tiling and stitching unnecessary, as well as allowing for larger batch size.", + "bbox": [ + 511, + 656, + 905, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "It has been shown that seamless stitching with CNNs is possible if edge effects are taken into account [17]. For CNNs only a limited area of the input image called the receptive field affects the prediction value in a given pixel [16]. Pixels on the edge of the tile do not get full context for prediction, resulting in a mismatch between neighboring tiles. To deal with this issue the tiles are sampled with overlap and only the central part of each tile is used", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13109", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80ca87eae8407c0424cf6bca006a04415534dedf2e8a6091d6659dd4edc04aaf.jpg", + "image_caption": [ + "Figure 1. Illustration of the pipeline for processing images larger than GPU memory: random sampling of tiles during training and sliding window inference. Example of artifacts caused by tiling: hallucinations in the low signal areas and discontinuous predictions at the tile borders." + ], + "image_footnote": [], + "bbox": [ + 140, + 88, + 433, + 448 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for stitching. We refer to the removed area as halo. We find that even with sufficient halo sliding window inference can cause tiling artifacts manifesting in two ways: as abrupt discrepancy between predictions in neighboring tiles or as background hallucinations, as shown in Fig. 1 and Fig. 2. This problem is rarely discussed but it affects many widely used tools, including nnU-Net [13], a popular baseline method for biomedical image segmentation.", + "bbox": [ + 89, + 564, + 482, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A common way to deal with the tiling artifacts is to apply additional postprocessing, for example, to average neighboring tiles with gaussian weights to avoid discontinuous predictions [13], or to use additional thresholding by intensity and filtering by the object size to remove artifacts in the background like in the Fig. 1. However, adding heuristic postprocessing steps makes the whole pipeline harder to use and apply to new datasets. Moreover, we show that tiling artifacts become more pronounced in the transfer setting, further reducing reusability of the trained networks.", + "bbox": [ + 89, + 686, + 482, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "According to our findings, the main cause of the artifacts is the tile-wise feature normalization inside the network. If InstanceNorm [23] is used in the CNN architecture, both during training and evaluation the statistics of", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the current input are used for normalization. This makes predictions for every pixel dependent on the content of the whole tile, therefore the limited receptive field assumption becomes invalid and, regardless of the size of the halo, it is impossible to make the stitching seamless.", + "bbox": [ + 511, + 90, + 903, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike InstanceNorm, another popular normalization method BatchNorm [12] uses the statistics of the current input during training but switches to saved running average statistics in the evaluation mode, allowing to achieve seamless stitching. However, we show that the performance of the network drastically reduces in evaluation mode compared to training mode because the statistics used during training are not stable with small batch size, making the running average not representative of the statistics of individual batches. This effect has previously been observed for natural image classification [10, 11], but we find it to be significantly more pronounced in biological image segmentation.", + "bbox": [ + 511, + 167, + 906, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work we demonstrate the trade-off between the artifact-free stitching ensured by the global normalization with BatchNorm and prediction quality achieved by tilewise normalization with InstanceNorm. We suggest BatchRenorm [11] which uses the same global normalization statistics both during training and inference as a solution which produces artifact-free predictions without compromising on the prediction quality.", + "bbox": [ + 511, + 363, + 905, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we make the following contributions:", + "bbox": [ + 529, + 484, + 864, + 498 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We show the existence of a normalization trade-off in segmentation of large biological images: using tile-wise statistics for feature normalization in the network during inference (InstanceNorm) leads to tiling artifacts even if edge effects were taken into consideration. Using tile-wise statistics in training and global accumulated statistics during inference (BatchNorm) leads to an unexpected difference between network performance during training and inference due to small batch size and high variability of the data;", + "- We suggest metrics for detecting normalization issues in networks used for sliding window inference pipelines: tile mismatch and train/eval disparity;", + "- We propose to use the Batch Renormalization method to achieve seamless stitching without sacrificing prediction accuracy to the train/eval disparity;", + "- We validate our findings on three volumetric biological datasets, with two network architectures: CNN-based 3D U-Net and ViT-based UNETR." + ], + "bbox": [ + 513, + 500, + 903, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 799, + 650, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sliding window inference is used in different domains, including medical [19, 24], biological [8] and aerial [9] image analysis. It was shown that receptive field $(RF)$ [15, 16] plays a key role in the seamless stitching [17]. Often in case of dense prediction the input and output have the same", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13110", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a0b2b4b552c58510840e0f70472d860965bf9ffcebad373fc8816d941b8832b6.jpg", + "image_caption": [ + "Figure 2. Examples of predictions with and without tiling artifacts" + ], + "image_footnote": [], + "bbox": [ + 133, + 87, + 310, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1ee204b8bcb3325498477b93443645858a02c07ade0fbcef4fbed9200f08f27e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 85, + 496, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/33977a809b71602cac52cf4d63a82fb40d183dd7bf70366efa592ca16a1baac4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 85, + 681, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/aacde65da1b4d3e5d1ff0502a707b57f7c30de3f03cf62bc2c2f8306866cb64c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 85, + 862, + 345 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08eca0c9df4f2e5e91f0f137c444c31d4a54d77d14071f8c31d2c2f2eb23fc2a.jpg", + "image_caption": [ + "Figure 3. Illustration of the tile mismatch metric." + ], + "image_footnote": [], + "bbox": [ + 148, + 398, + 410, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "size thanks to the zero-padding in convolutional layers, so the coordinates in both spaces can be used interchangeably. For the output pixels which are closer to the edge of the tile than $\\frac{1}{2} RF$ part of the input values within the receptive field is replaced by zeros, creating mismatch in the predictions between tiles. Proposed solution is to crop the prediction and only use the part of the output where each pixel has full context for stitching. The minimal sufficient size of the cropped area is $\\frac{1}{2} RF$ . The same concept was discussed from the point of view of translational equivariance [9, 21]. Although convolutions and pooling are translation-equivariant, zero-padding makes the CNN overall lose this property. With valid convolutions the size of the output is smaller than the input but it allows to keep the translational", + "bbox": [ + 89, + 688, + 485, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "equivariance and therefore seamless stitching.", + "bbox": [ + 511, + 397, + 816, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice many pipelines using sliding window inference do not aim to completely eliminate the tiling artifacts but rather minimize them by using the largest possible tile size in inference and averaging. For example, nnU-Net [13] and Cellpose [22] compute weighted average of the tiles, with the weight decreasing from 1 in the middle of the tile to 0 at the border, smoothing the border between tiles.", + "bbox": [ + 511, + 411, + 905, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These recommendations help against the discontinuity in predictions caused by edge effects but our experiments show that even in the valid convolution area predictions do not necessarily match (see Fig. 3). In addition, this does not explain hallucination-like artifacts in the background like in Fig. 1. It has been shown that prediction accuracy can depend significantly on the tile size [3, 18]. To our knowledge, the connection between these effects and feature normalization inside the network was not studied before.", + "bbox": [ + 511, + 517, + 906, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 511, + 664, + 612, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section is structured as follows: we define how we perform sliding window inference and which normalization strategies we consider. We then propose two metrics that can identify normalization issues within the network, pointing out both tiling artifacts (\"tile mismatch\" metric) and performance drop in evaluation mode (\"train/eval disparity\" metric).", + "bbox": [ + 511, + 689, + 905, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Sliding window inference with halo", + "text_level": 1, + "bbox": [ + 511, + 801, + 820, + 818 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input data can be a 2D image or a 3D volume with one or multiple channels. Here both images and volumes are referred to as images. For inference the image is split into a grid of tiles, incomplete tiles are padded with zeros. The stitched prediction is calculated from by tile-wise", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13111", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c9505d6eff9dcc13148bbc44a5a2471ec0739bce650fad4814ba2751dcd1805.jpg", + "image_caption": [ + "Figure 4. Receptive field, shown as Log10 of mean gradient of the central pixel output with respect to the input" + ], + "image_footnote": [], + "bbox": [ + 200, + 89, + 803, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "predictions. To ensure that observed tiling artifacts were not caused by edge effects, we calculated theoretical (TRF) and effective receptive field (ERF) of the networks following [15]. We define TRF as all pixels which belong to the computation graph for the target pixel in the output space. ERF is defined as width $(\\frac{3\\sigma}{2})$ of a Gaussian distribution fitted to the averaged gradient distribution During sliding window inference the overlap was set to the size of halo, minimizing the number of tiles needed to cover the whole volume. The halo part of the tile was completely removed so there was no averaging of predictions or other postprocessing.", + "bbox": [ + 89, + 364, + 483, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Normalization strategies", + "text_level": 1, + "bbox": [ + 89, + 541, + 316, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Normalization can be inserted in two parts of the pipeline: preprocessing of the input data and feature map normalization inside the network.", + "bbox": [ + 89, + 564, + 483, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Input normalization", + "text_level": 1, + "bbox": [ + 89, + 617, + 277, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quantile normalization with $q_{min} = 0.01$ and $q_{max} = 0.98$ and clipping to [0..1] were applied to each image as a whole to convert the data type and value range from the microscope output to float32 and [0..1] range. In addition to the global normalization, quantile normalization can be applied to each tile separately to make the input of the network always have the same range and make tiles more comparable. Based on this we considered two strategies:", + "bbox": [ + 89, + 636, + 483, + 756 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Global: quantile normalization only of the whole image.", + "- Tile-wise: quantile normalization of each sampled tile, both during training and during inference." + ], + "bbox": [ + 89, + 758, + 480, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Feature normalization", + "text_level": 1, + "bbox": [ + 89, + 813, + 292, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "General formula for the normalization operation with input $x$ , output $y$ is:", + "bbox": [ + 89, + 830, + 483, + 863 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny = \\frac {x - \\mu}{\\sqrt {\\sigma^ {2} + \\epsilon}},\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 875, + 338, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu$ and $\\sigma$ are normalization parameters and $\\epsilon$ is a small constant used for numerical stability. Parameters $\\mu$ and $\\sigma$ can be estimated directly from the input $\\mu = \\operatorname{E}[x]$ and $\\sigma^2 = \\operatorname{Var}[x]$ , where average can be taken either over each sample independently or over the whole batch. Alternatively, global normalization parameters independent of the current input can be used. A common strategy is to estimate the parameters as a running average over multiple samples: $p_{new} = (1 - momentum) \\times p_{old} + momentum \\times p_t$ . The update speed is determined by the momentum.", + "bbox": [ + 511, + 364, + 906, + 517 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- BatchNorm [12] Training: use statistics of the current batch and collect running average. Inference: use saved running average.", + "- InstanceNorm [23] Both training and inference: use statistics of each input sample. In Pytorch this behavior can be changed by setting trackRunning.stats=True. Then during inference the statistics collected during training will be utilized, same as in BatchNorm.", + "- BatchRenorm [11] Both training and inference: use running average statistics.", + "- Identity Skip feature normalization." + ], + "bbox": [ + 511, + 525, + 906, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Proposed evaluation metrics", + "text_level": 1, + "bbox": [ + 526, + 723, + 782, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Train/eval disparity When global normalization with running average statistics is used during inference, a disparity can arise between the predictions of the same tile in train and eval mode as the train mode always uses statistics of the current batch. This disparity is an indicator of a potential drop in performance which is not related to model overfitting. To measure it quantitatively, we introduce the following metric:", + "bbox": [ + 524, + 744, + 906, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "train/eval disparity $= 1 - \\text{Dice}(P_{\\text{train}}, P_{\\text{eval}})$ ,", + "bbox": [ + 545, + 885, + 883, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13112", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{train}$ is the prediction done with model.train() and $P_{eval}$ - with model.eval().", + "bbox": [ + 102, + 90, + 480, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tile mismatch", + "text_level": 1, + "bbox": [ + 104, + 121, + 209, + 133 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantification of tiling artifacts is challenging because their magnitude depends on the similarity of the content of neighboring tiles in different parts of the image. We propose to split the image into overlapping tiles and compare the predictions in the overlap areas. To avoid edge effects, we only take the valid part of the overlap, as shown in Fig. 3. The following metrics are proposed:", + "bbox": [ + 102, + 136, + 483, + 243 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm a x d i s t = \\max _ {i = 1} ^ {M} (| O _ {i 1} - O _ {i 2} |)\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 253, + 400, + 280 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nt i l e m i s m a t c h = \\underset {i = 1} {\\overset {M} {\\operatorname {m e d i a n}}} (1 - D i c e (O _ {i 1}, O _ {i 2})),\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 301, + 457, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $O_{i1}$ and $O_{i2}$ are predictions in the valid overlap region in tile $i$ and $M$ is number of sampled tiles. In max dist max is taken over all channels and all tiles, making it an indicator of wether given setup produces tiling artifacts. In cases when max dist = 0 we explicitly report \"no\" instead of tile mismatch to emphasize that the predictions perfectly match. tile mismatch is calculated per channel and characterizes the magnitude of the artifacts.", + "bbox": [ + 102, + 340, + 483, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 491, + 223, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 516, + 192, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing cell segmentation benchmarks often operate at the same image size as natural images and contain mostly the object-centered parts. Here the datasets were chosen to demonstrate the common practical challenges of microscopy segmentation such as large image size and variability of class distributions within the volume.", + "bbox": [ + 89, + 537, + 483, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "organoids [7]: semantic segmentation of tissue in patient-derived colorectal cancer organoids, electron microscopy (FIB/SEM), $40 \\times 30 \\times 30 \\mathrm{~nm}^3$ voxel size. The volume of size $1350 \\times 1506 \\times 1647$ pixels was split 70/30 into train and validation dataset. Classes: cell boundaries, foreground (pixels inside the cells), extracellular matrix (space between cells and background).", + "bbox": [ + 89, + 628, + 483, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "plants [25]: instance segmentation of nuclei in Arabidopsis ovules, confocal fluorescent microscopy, isotropic resolution of $0.13\\mu m^3$ per pixel, 4 train volumes and 1 validation volume $\\sim 500\\times 1000\\times 1000$ pixels each.", + "bbox": [ + 89, + 734, + 482, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "embryo [2]: instance segmentation of nuclei in mouse embryos, confocal fluorescent microscopy, isotropic resolution of $0.2\\mu m^3$ per pixel, 22 train volumes and 13 validation volumes, $\\sim 600\\times 1000\\times 700$ px each.", + "bbox": [ + 89, + 795, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For plants and embryo datasets the original instance segmentation groundtruth was converted to semantic segmentation of three classes: object boundaries, foreground", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/25de00239db4f38cf74f0a6a8ea8005a7a1b5828fdeef28f6a0690ca61ff9823.jpg", + "image_caption": [ + "Figure 5. Tiling artifacts become more pronounced in a transfer setting. For quantitative evaluation see STab. 1." + ], + "image_footnote": [], + "bbox": [ + 555, + 89, + 861, + 468 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e512195231098beb6a7a82913eaec4db07e017029fb627ab51a1893fcd107152.jpg", + "image_caption": [ + "Figure 6. Performance vs tile size." + ], + "image_footnote": [], + "bbox": [ + 542, + 518, + 877, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(objects with subtracted boundaries) and background. For the transfer setup, networks trained for embryo dataset were used to segment plants dataset and vice versa.", + "bbox": [ + 511, + 672, + 906, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Architectures", + "text_level": 1, + "bbox": [ + 511, + 727, + 653, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D U-Net: 3D U-Net with 2 downsampling steps and 2 convolutional blocks per level was used for all datasets. Each convolutional block consists of a convolutional layer, a normalization layer and a non-linearity. The normalization layers in all the blocks were replaced with one of the normalization layers listed above. Tile size (192, 192, 192) was used for training and (128, 128, 128) for sliding window inference.", + "bbox": [ + 511, + 750, + 906, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D UNETR: UNETR consists of a ViT encoder and a convolutional decoder. In the encoder the volume is split", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13113", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0d245ad8b35f61ff9f5c00ad69c6303228b3aca6658280a0d07a2b1954cf0c0d.jpg", + "table_caption": [ + "Table 1. U-Net results table" + ], + "table_footnote": [], + "table_body": "
Normalization layerBatchNorm globalBatchNorm tile-wise globalInstanceNorm global tile-wiseInstanceNorm tile-wiseBatchRenorm globalIdentity global
Input norm
Feature norm
dice, eval modeorganoidsforeground0.950.880.940.930.950.95
boundaries0.490.150.750.750.730.72
matrix0.820.610.610.610.850.83
plantsforeground0.850.880.890.890.890.52
boundaries0.690.730.750.740.750.26
embryoforeground0.660.680.670.720.740.59
boundaries0.550.550.550.530.580.44
dice, train modeorganoidsforeground0.940.930.940.930.950.95
boundaries0.750.750.750.750.730.72
matrix0.610.610.610.610.850.83
plantsforeground0.890.890.890.890.890.52
boundaries0.750.740.750.740.750.26
embryoforeground0.670.720.670.720.740.59
boundaries0.550.520.550.530.580.44
train/eval disparityorganoids0.480.840000
plants0.170.110000
embryo0.370.290000
tile mismatchorganoidsno0.030.110.11nono
plantsno0.010.050.05nono
embryono0.090.160.20nono
", + "bbox": [ + 109, + 114, + 887, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "into $16 \\times 16 \\times 16$ patches and embeddings in the ViT are normalized using LayerNorm after each round of attention. With LayerNorm all features are normalized using average and variance of the features in each patch independently. Other normalization strategies can be used in the ViT [27], but here LayerNorm was kept in all experiments as a standard approach. Convolutional decoder consists of 4 upsampling steps with 2 convolutional blocks at each level of the U-Net-like architecture. The normalization layers in all the blocks were replaced with one of the normalization layers listed above.", + "bbox": [ + 89, + 508, + 480, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Reported metrics", + "text_level": 1, + "bbox": [ + 89, + 694, + 261, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall prediction quality: Prediction was done on the whole volume using sliding window inference as described above. If the setup included tile-wise input normalization, each tile was normalized separately before being passed through the network. The median Dice score over all samples in the validation dataset was reported for each class. Train/eval disparity: We report the Dice score both for training and evaluation mode to demonstrate the disparity in Tab. 1 and Tab. 2. Inference in both modes was done on the same images from the validation set, so the difference comes only from the behavior of the model in different modes and not overfitting to the training data. Tile mis", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "match: The tiles of $192 \\times 192 \\times 288\\mathrm{px}$ were sampled in a grid with a stride of 64 pixels to test the mismatch in the whole volume. Sampled tiles were split into two overlapping tiles of size $192 \\times 192 \\times 192$ . Both tiles were then processed to compare the predictions in the overlapping region. Tile mismatch values for boundary channel are reported in Tab. 1, Tab. 2 and STab. 1.", + "bbox": [ + 511, + 508, + 903, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Further details can be found in the Supplementary.", + "bbox": [ + 532, + 614, + 866, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 513, + 642, + 601, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Receptive field", + "text_level": 1, + "bbox": [ + 511, + 667, + 661, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We calculated TRF and ERF for the architectures with different normalization layers. As expected, for U-Net with the global feature normalization (BatchNorm, BatchRenorm, Identity) only a limited region around the central pixel belongs to TRF, ERF constitutes the effective receptive field only the small portion as shown in Fig. 4. For the U-Net with tile-wise feature normalization (InstanceNorm) ERF is the same, which is expected since it is determined by the parameters of the convolutional architecture, but the TRF takes the whole tile as values of all pixels in the tile make a contribution to the prediction for the central pixel via normalization. Therefore, regardless of the size of the overlap between tiles in the sliding window prediction, the stitched prediction will not be exactly seam", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13114", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "less. In all sliding window experiments with the U-Net we used halo calculated using TRF of the U-Net with global feature normalization (Fig. 4).", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For UNETR TRF also takes the whole input tile due to normalization in the convolutional decoder and the attention mechanism. The averaged gradients image Fig. 4 has a clear square pattern because in the encoder the image is downsampled by 16 and then upsampled back to the original resolution. Despite the attention layers, ERF is the same as for the U-Net. Since only 3 central patches have large effect on the central pixel prediction, we set $halo = 32$ for all UNETR experiments (Fig. 4).", + "bbox": [ + 89, + 137, + 483, + 273 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Tile-wise feature normalization with InstanceNorm causes tiling artifacts", + "text_level": 1, + "bbox": [ + 89, + 284, + 483, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using InstanceNorm allows to reach high quality predictions, however the tile-wise feature normalization is a source of artifacts both for U-Net (Tab. 1) and for UNETR (Tab. 2) as shown in Fig. 2. We observe that tile-wise input normalization (BatchNorm, tile-wise input norm) causes less severe tile mismatch than tile-wise network feature normalization (InstanceNorm, global input norm). Although it is impossible to reach completely seamless stitching with UNETR, experiments with InstanceNorm show strong tile mismatch, while global feature normalization and sufficient halo based on ERF size allow to reduce mismatch and make it almost unnoticeable.", + "bbox": [ + 89, + 321, + 483, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although InstanceNorm performs better than BatchNorm for boundaries, networks with tile-wise normalization struggle with classes which cover areas larger than the tile size, such as foreground and matrix in organoids dataset. Experiments with inference using different tile sizes Fig. 6 show that the prediction depends strongly on the tile size, reaching the best quality when the tiles become large enough for the feature distribution in each tile to reproduce the distribution in the whole image.", + "bbox": [ + 89, + 503, + 483, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 shows how artifacts get worse in the transfer setting: the model originally trained on the embryo dataset produces less artifacts than the model transferred from the plants dataset. We observe the same effect in transfer in both directions (STab. 1).", + "bbox": [ + 89, + 640, + 483, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Mismatch in batch statistics causes train and eval mode disparity for BatchNorm", + "text_level": 1, + "bbox": [ + 89, + 727, + 482, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Global feature normalization with BatchNorm allows to eliminate the tile mismatch but predictions for organoids boundary and other datasets are worse than with InstanceNorm. The reason for this is the disparity between the performance of the network in train and evaluation mode. Estimation of normalization statistics with a running average in BatchNorm relies on the batch statistics being stable enough to represent the whole dataset. If batches are too small, the running average can be very dif", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3a86a71116a9c1e17a10236de253d8a906a33506b922698cce53d3e6c91cb912.jpg", + "table_caption": [ + "Table 2. UNETR results table" + ], + "table_footnote": [], + "table_body": "
Norm layerBatchNormInstanceNormBatchRenormIdentity
Input normglobalglobalglobalglobal
Feature normglobaltile-wiseglobalglobal
dice,eval modeorganoids0.930.890.950.91
0.390.600.720.60
plants0.030.100.870.28
foreground0.840.850.850.52
boundaries0.640.690.660.21
foreground0.670.560.670.00
boundaries0.490.490.530.00
dice,train modeorganoids0.890.890.950.91
0.600.600.720.60
boundaries0.100.100.870.28
matrix0.850.850.850.52
foreground0.690.690.660.21
boundaries0.560.560.670.00
boundaries0.490.490.530.00
t/e disparityorganoids0.36000
plants0.20000
embryo0.33000
tile mismatchorganoids0.000.100.000.00
plants0.000.040.000.00
embryo0.000.060.000.00
", + "bbox": [ + 516, + 114, + 903, + 565 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ef99ac44483fe539c5193224e0f09484458018ac6b4bc432cfe96a4d66bc126.jpg", + "image_caption": [ + "Figure 7. Example of train/eval disparity" + ], + "image_footnote": [], + "bbox": [ + 563, + 583, + 856, + 813 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ferent from the batch-wise statistics used during training, leading to deteriorating performance. Doing inference with", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13115", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the network in train mode supports this hypothesis: the performance recovers, at the cost of tiling artifacts. In our experiments train and evaluation mode predictions can be quite different, with train/eval disparity reaching 0.84 for the organoids dataset. Qualitatively, the changes in the predictions occur in the areas of low prediction certainty. For example, in Fig. 7 the eval mode differs from the train mode in the areas where the background has a bright line artifact or in the noisier areas.", + "bbox": [ + 89, + 90, + 480, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This effect is not the same as the usual difference between train and evaluation performance caused by overfitting. Here, the network's output differs even for the exact same images thus violating the usual assumption that the prediction quality is approximately the same in train and eval mode. This issue would not be remedied by using augmentations or more training data as it does not address the high variance of statistics.", + "bbox": [ + 89, + 227, + 480, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. BatchRenorm corrects for train/eval disparity and provides seamless stitching", + "text_level": 1, + "bbox": [ + 89, + 357, + 480, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "InstanceNorm avoids the statistics mismatch by always using the statistics of the input both during training and during inference at the cost of causing tiling artifacts. BatchNorm does not introduce the tiling artifacts but can lose performance due to train/eval disparity.", + "bbox": [ + 89, + 393, + 480, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A potential solution is to remove the normalization altogether. In our experiments not using normalization (Identity in Tab. 1 and Tab. 2) has led to worse performance for plants dataset and similar performance for organoids and embryo. It could be possible to improve the performance of networks with no normalization by finding more suitable training parameters and reducing the learning rate, however the necessity to adjust training parameters for each dataset makes the pipeline harder to use.", + "bbox": [ + 89, + 469, + 480, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another solution is to use the same global normalization statistics both in training and in inference. Directly using running average statistics during training causes network's weights to explode [11], which is why we propose to use Batch Renormalization method. We found that networks with BatchRenorm produce same or better predictions as InstanceNorm (Tab. 1, Tab. 2) including the transfer setup (STab. 1) while allowing for seamless stitching due to global feature normalization (last column of Fig. 2, Fig. 5).", + "bbox": [ + 89, + 604, + 480, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 89, + 753, + 204, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The severity of tiling artifacts depends on the data and on the tile size. For example, the tile mismatch can be negligible in case the images are very homogeneous in content and the tile size is big enough that each tile mostly reproduces the global statistics of the image, such as for plants dataset in our experiments. However, even in that case the inference performance depends on the tile size and can change unpredictably in the transfer setting, putting higher", + "bbox": [ + 89, + 779, + 480, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "requirements on the hardware for inference and making the pipeline harder to use, especially for non-expert users. Tracing low validation accuracy to the instability in batch statistics is not straightforward either, as it is normally expected that the performance on the validation dataset should be worse than on the training data. If the training loss is low but the accuracy on validation data does not improve, it is often attributed to the insufficient amount of ground truth, however, we show that it can also be caused by train/eval disparity preventing efficient utilization of the ground truth.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Both of the described effects are highly dependent on the particular data, input normalization, sampling strategy, postprocessing and other parameters of the image analysis pipeline. For any such pipeline, we introduce two simple metrics: tile mismatch and train/eval disparity to help determine if it has issues with normalization. Both metrics are easy to use for any sliding window inference pipeline, including other dense prediction tasks if the Dice score is replaced with another, task-appropriate metric.", + "bbox": [ + 511, + 242, + 903, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We found that BatchRenorm, originally introduced for natural image classification, performs well for segmenting large biological images. The challenge with images larger than GPU memory is ensuring consistent normalization parameters across the entire image for seamless stitching, despite the inability to process the full image during training due to time constraints. BatchRenorm solves this by splitting training into two stages. First, the network is trained with BatchNorm. Then the normalization layer switches to using the accumulated running average statistics. In our experience it is very important that the training converges before switching and that the statistics change from tile-wise to global gradually to achieve the best final accuracy. While the BatchRenorm adds more hyperparameters, in our experiments the right moment to switch could easily be found from looking at the loss curves.", + "bbox": [ + 511, + 378, + 903, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we demonstrate that tiling artifacts in sliding window inference pipelines can be caused by the tilewise feature normalization in the network. We also find that switching from tile-wise normalization in training to global normalization in evaluation can lead to even worse performance, including hallucinations, due to small batch size. As these effects strongly depend on the data and tile size, we introduce tile mismatch and train/eval disparity metrics to easily check if a specific pipeline and dataset have these issues. We propose that Batch Renormalization technique will yield artifact-free predictions even with small batch size and demonstrate its good performance through extensive experiments on various microscopy datasets and network architectures.", + "bbox": [ + 511, + 659, + 903, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. We would like to thank Ricardo Sánchez Loayza and Qin Yu for helpful discussions. E.B.", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13116", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "was supported by the Joachim Herz Foundation through an Add-on Fellowship for Interdisciplinary Life Science. E.D'I. was supported by a fellowship from the EMBL Interdisciplinary (EI4POD) program under Marie Sk1odowska-Curie Actions COFUND (847543). J.M. acknowledges funding from the EMBL and the European Research Council (ERC 3DCellPhase-760067). This work was supported by the European Commission through the Horizon Europe program (IMAGINE project, grant agreement 101094250-IMAGINE). The work of A.A. was funded by the German Research Foundation through PA 4341/2-1. C.P. received support under Germany's Excellence Strategy - EXC 2067/1-390729940. We would like to thank the EMBL IT Services Department for providing computational infrastructure and support.", + "bbox": [ + 89, + 90, + 485, + 318 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 330, + 187, + 345 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anwai Archit, Luca Freckmann, Sushmita Nair, Nabeel Khalid, Paul Hilt, Vikas Rajashekar, Marei Freitag, Carolin Teuber, Genevieve Buckley, Sebastian von Haaren, Sagnik Gupta, Andreas Dengel, Sheraz Ahmed, and Constantin Pape. Segment anything for microscopy. Nature Methods, 2025. 1", + "[2] Vladyslav Bondarenko, Mikhail Nikolaev, Dimitri Kromm, Roman Belousov, Adrian Wolny, Marloes Blotenburg, Peter Zeller, Saba Rezakhani, Johannes Hugger, Virginie Uhlmann, Lars Hufnagel, Anna Kreshuk, Jan Ellenberg, Alexander van Oudenaarden, Anna Erzberger, Matthias P Lutolf, and Takashi Hiiragi. Embryo-uterine interaction coordinates mouse embryogenesis during implantation. The EMBO Journal, 42(17), 2023. 5", + "[3] Calimanut-Ionut Cira, Miguel-Ángel Manso-Callejo, Naoto Yokoya, Tudor Sălăgean, and Ana-Cornelia Badea. Impact of tile size and tile overlap on the prediction performance of convolutional neural networks trained for road classification. Remote Sensing, 16(15):2818, 2024. 3", + "[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 1", + "[6] Sven Dorkenwald, Arie Matslieh, Amy R Sterling, Philipp Schlegel, Szi-Chieh Yu, Claire E McKellar, Albert Lin, Marta Costa, Katharina Eichler, Yijie Yin, Will Silversmith, Casey Schneider-Mizell, Chris S Jordan, Derrick Brittain, Akhilesh Halageri, Kai Kuehner, Oluwaseun Ogedengbe, Ryan Morey, Jay Gager, Krzysztof Kruk, Eric Perlman, Runzhe Yang, David Deutsch, Doug Bland, Marissa Sorek, Ran Lu, Thomas Macrina, Kisuk Lee, J Alexander Bae, Shang Mu, Barak Nehoran, Eric Mitchell, Sergiy Popovych, Jingpeng Wu, Zhen Jia, Manuel A Castro, Nico Kemnitz, Dodam" + ], + "bbox": [ + 99, + 356, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ih, Alexander Shakeel Bates, Nils Eckstein, Jan Funke, Forrest Collman, Davi D Bock, Gregory S X E Jefferis, H Sebastian Seung, Mala Murthy, and FlyWire Consortium. Neuronal wiring diagram of an adult brain. Nature, 634(8032): 124-138, 2024. 1", + "[7] Edoardo D'Imprima, Marta Garcia Montero, Sylwia Gawrzak, Paolo Ronchi, Ievgeniia Zagoriy, Yannick Schwab, Martin Jechlinger, and Julia Mahamid. Light and electron microscopy continuum-resolution imaging of 3d cell cultures. Developmental Cell, 58(7):616-632.e6, 2023. 5", + "[8] T. Falk, D. Mai, R. Bensch, Ö. Ciçek, A. Abdulkadir, Y. Marrakchi, A. Böhm, J. Deubner, Z. Jäckel, K. Seiwald, A. Dovzhenko, O. Tietz, C. Dal Bosco, S. Walsh, D. Saltukoglu, T. L. Tay, M. Prinz, K. Palme, M. Simons, I. Diester, T. Brox, and O. Ronneberger. U-net - deep learning for cell counting, detection, and morphometry. Nature Methods, 16: 67-70, 2019. 2", + "[9] Bohao Huang, Daniel Reichman, Leslie M Collins, Kyle Bradbury, and Jordan M Malof. Tiling and stitching segmentation output for remote sensing: Basic challenges and recommendations. arXiv:1805.12219, 2018. 2, 3", + "[10] Lei Huang, Jie Qin, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Normalization techniques in training dnns: Methodology, analysis and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10173-10196, 2023. 2", + "[11] Sergey Ioffe. Batch renormalization: Towards reducing minibatch dependence in batch-normalized models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 4, 8", + "[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning, pages 448-456, Lille, France, 2015. PMLR. 2, 4", + "[13] Fabian Isensee and et al. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2):203-211, 2020. 2, 3", + "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 1", + "[15] Vincent Loos, Rohit Pardasani, and Navchetan Awasthi. Demystifying the effect of receptive field size in u-net models for medical image segmentation. Journal of Medical Imaging, 2024. 2, 4", + "[16] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 1, 2", + "[17] Michael Majurski and Peter Bajcsy. Exact tile-based segmentation inference for images larger thangpu memory. Journal of Research of the National Institute of Standards and Technology, 126, 2021. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13117", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] G. Anthony Reina, Ravi Panchumarthy, Siddhesh Pravin Thakur, Alexei Bastidas, and Spyridon Bakas. Systematic evaluation of image tiling adverse effects on deep learning semantic segmentation. Frontiers in Neuroscience, 14, 2020. 3", + "[19] O. Ronneberger, P.Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241. Springer, 2015. (available on arXiv:1505.04597 [cs.CV]). 2", + "[20] Xiongtao Ruan, Matthew Mueller, Gaoxiang Liu, Frederik Gorlitz, Tian-Ming Fu, Daniel E Milkie, Joshua L Lillvis, Alexander Kuhn, Johnny Gan Chong, Jason Li Hong, Chu Yi Aaron Herr, Wilmene Hercule, Marc Nienhaus, Alison N Killilea, Eric Betzig, and Srigokul Upadhyayula. Image processing tools for petabyte-scale light sheet microscopy data. Nat. Methods, 21(12):2342-2352, 2024. 1", + "[21] Josef Lorenz Rumberger, Xiaoyan Yu, Peter Hirsch, Melanie Dohmen, Vanessa Emanuela Guarino, Ashkan Mokarian, Lisa Mais, Jan Funke, and Dagmar Kainmueller. How shift equivariance impacts metric learning for instance segmentation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), page 7108-7116. IEEE, 2021. 1, 3", + "[22] Carsen Stringer, Tim Wang, Michalis Michaelos, and Marius Pachitariu. Cellpose: a generalist algorithm for cellular segmentation. Nature Methods, 18(1):100-106, 2020. 1, 3", + "[23] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization, 2016. 2, 4", + "[24] Özgün undefinedigkeit, Ahmed Abdulkadir, Soeren S. Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation, page 424-432. Springer International Publishing, 2016. 2", + "[25] Athul Vijayan, Tejasvinee Atul Mody, Qin Yu, Adrian Wolny, Lorenzo Cerrone, Soeren Strauss, Miltos Tsiantis, Richard S. Smith, Fred A. Hamprecht, Anna Kreshuk, and Kay Schneitz. A deep learning-based toolkit for 3d nuclei segmentation and quantitative analysis in cellular and tissue context. Development, 151(14), 2024. 5", + "[26] Martin Weigert, Uwe Schmidt, Robert Haase, Ko Sugawara, and Gene Myers. Star-convex polyhedra for 3d object detection and segmentation in microscopy. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), page 3655–3662. IEEE, 2020. 1", + "[27] Zhuliang Yao, Yue Cao, Yutong Lin, Ze Liu, Zheng Zhang, and Han Hu. Leveraging batch normalization for vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). IEEE, 2021. 6" + ], + "bbox": [ + 91, + 92, + 480, + 781 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13118", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/eaeeea1f-eecf-4d2e-8480-0e89bb127b12_model.json b/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/eaeeea1f-eecf-4d2e-8480-0e89bb127b12_model.json new file mode 100644 index 0000000000000000000000000000000000000000..de932a60c879c9858cffbbe2e360dc43111e4e11 --- /dev/null +++ b/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/eaeeea1f-eecf-4d2e-8480-0e89bb127b12_model.json @@ -0,0 +1,1914 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.112, + 0.13, + 0.89, + 0.177 + ], + "angle": 0, + "content": "Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.203, + 0.796, + 0.239 + ], + "angle": 0, + "content": "Elena Buglakova1, Anwai Archit2, Edoardo D'Imprime1,4, Julia Mahamid1, Constantin Pape2,3, Anna Kreshuk1" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.239, + 0.663, + 0.256 + ], + "angle": 0, + "content": "\\(^{1}\\) European Molecular Biology Laboratory, Heidelberg" + }, + { + "type": "text", + "bbox": [ + 0.33, + 0.257, + 0.669, + 0.274 + ], + "angle": 0, + "content": "\\(^{2}\\) Institute of Computer Science, University of Göttingen" + }, + { + "type": "list", + "bbox": [ + 0.33, + 0.239, + 0.669, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.181, + 0.275, + 0.816, + 0.309 + ], + "angle": 0, + "content": "3 Cluster of Excellence 'Multiscale Bioimaging: from Molecular Machines to Networks of Excitable Cells' (MBExC), Georg-August-University Göttingen" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.31, + 0.752, + 0.327 + ], + "angle": 0, + "content": "\\(^{4}\\) IRCCS Humanitas Research Hospital, via Manzoni 56, 20089 Rozzano, Milan, Italy" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.362, + 0.327, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Segmentation of very large images is a common problem in microscopy, medical imaging or remote sensing. The problem is usually addressed by sliding window inference, which can theoretically lead to seamlessly stitched predictions. However, in practice many of the popular pipelines still suffer from tiling artifacts. We investigate the root cause of these issues and show that they stem from the normalization layers within the neural networks. We propose indicators to detect normalization issues and further explore the trade-offs between artifact-free and high-quality predictions, using three diverse microscopy datasets as examples. Finally, we propose to use BatchRenorm as the most suitable normalization strategy, which effectively removes tiling artifacts and enhances transfer performance, thereby improving the reusability of trained networks for new datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.222, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Segmentation is an essential first step of pipelines in biological image analysis, a diverse domain of computer vision that enables automated quantification for various types of microscopy images. While this domain has benefited a lot from the developments in natural image analysis, it also poses unique challenges. Investigation of biological diversity, discovery of unexpected phenotypes and development of novel imaging protocols leads to high variability of datasets. This explains the popularity of human-in-the-loop approaches to segmentation, where a small amount of data is either annotated from scratch or from the output of pretrained models [1, 22, 26], then a new model is either trained from scratch or fine-tuned. After that the next rounds of correction and training happen, until the desired quality is reached. Being able to quickly train a smaller model for an" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.364, + 0.799, + 0.379 + ], + "angle": 0, + "content": "unseen image type is crucial in this setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.906, + 0.472 + ], + "angle": 0, + "content": "Another important feature of biological images, especially for volumetric imaging, is their large size. Electron microscopy volumes for connectomics [6] or light sheet microscopy volumes [20] can reach the scale of multiple terabytes. It is routine to work with 10-100 GB images of \\(\\sim 1000^3\\) pixels, which exceed GPU memory." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.906, + 0.655 + ], + "angle": 0, + "content": "As shown in Fig. 1, during training random patches are sampled from the annotated images and combined into a batch. During inference/evaluation, the image is split into a grid of tiles which are predicted independently and then stitched back to form the full prediction. Here we refer both to 2D and 3D image patches as tiles. In this setup the size of the tiles is limited from above by GPU memory and from below by the receptive field of the network [21]. Typical tile size is around \\(96 \\times 96 \\times 96 - 256 \\times 256 \\times 256\\) pixels and the training batch size is 2-3 tiles per batch. During inference it is not necessary to store gradients, therefore either tile size or batch size can be larger." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.778 + ], + "angle": 0, + "content": "Tile-wise prediction and stitching might look like a purely technical step, however for microscopy patch sampling and stitching can significantly affect prediction quality and create artifacts. In contrast, big natural image datasets, such as ImageNet [5], CityScapes [4] and SAM dataset [14], usually contain relatively small 2D images, making tiling and stitching unnecessary, as well as allowing for larger batch size." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "It has been shown that seamless stitching with CNNs is possible if edge effects are taken into account [17]. For CNNs only a limited area of the input image called the receptive field affects the prediction value in a given pixel [16]. Pixels on the edge of the tile do not get full context for prediction, resulting in a mismatch between neighboring tiles. To deal with this issue the tiles are sampled with overlap and only the central part of each tile is used" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13109" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.089, + 0.434, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Figure 1. Illustration of the pipeline for processing images larger than GPU memory: random sampling of tiles during training and sliding window inference. Example of artifacts caused by tiling: hallucinations in the low signal areas and discontinuous predictions at the tile borders." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.686 + ], + "angle": 0, + "content": "for stitching. We refer to the removed area as halo. We find that even with sufficient halo sliding window inference can cause tiling artifacts manifesting in two ways: as abrupt discrepancy between predictions in neighboring tiles or as background hallucinations, as shown in Fig. 1 and Fig. 2. This problem is rarely discussed but it affects many widely used tools, including nnU-Net [13], a popular baseline method for biomedical image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.839 + ], + "angle": 0, + "content": "A common way to deal with the tiling artifacts is to apply additional postprocessing, for example, to average neighboring tiles with gaussian weights to avoid discontinuous predictions [13], or to use additional thresholding by intensity and filtering by the object size to remove artifacts in the background like in the Fig. 1. However, adding heuristic postprocessing steps makes the whole pipeline harder to use and apply to new datasets. Moreover, we show that tiling artifacts become more pronounced in the transfer setting, further reducing reusability of the trained networks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "According to our findings, the main cause of the artifacts is the tile-wise feature normalization inside the network. If InstanceNorm [23] is used in the CNN architecture, both during training and evaluation the statistics of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.166 + ], + "angle": 0, + "content": "the current input are used for normalization. This makes predictions for every pixel dependent on the content of the whole tile, therefore the limited receptive field assumption becomes invalid and, regardless of the size of the halo, it is impossible to make the stitching seamless." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Unlike InstanceNorm, another popular normalization method BatchNorm [12] uses the statistics of the current input during training but switches to saved running average statistics in the evaluation mode, allowing to achieve seamless stitching. However, we show that the performance of the network drastically reduces in evaluation mode compared to training mode because the statistics used during training are not stable with small batch size, making the running average not representative of the statistics of individual batches. This effect has previously been observed for natural image classification [10, 11], but we find it to be significantly more pronounced in biological image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.484 + ], + "angle": 0, + "content": "In this work we demonstrate the trade-off between the artifact-free stitching ensured by the global normalization with BatchNorm and prediction quality achieved by tilewise normalization with InstanceNorm. We suggest BatchRenorm [11] which uses the same global normalization statistics both during training and inference as a solution which produces artifact-free predictions without compromising on the prediction quality." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.485, + 0.866, + 0.499 + ], + "angle": 0, + "content": "In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.5, + 0.905, + 0.65 + ], + "angle": 0, + "content": "- We show the existence of a normalization trade-off in segmentation of large biological images: using tile-wise statistics for feature normalization in the network during inference (InstanceNorm) leads to tiling artifacts even if edge effects were taken into consideration. Using tile-wise statistics in training and global accumulated statistics during inference (BatchNorm) leads to an unexpected difference between network performance during training and inference due to small batch size and high variability of the data;" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.651, + 0.905, + 0.695 + ], + "angle": 0, + "content": "- We suggest metrics for detecting normalization issues in networks used for sliding window inference pipelines: tile mismatch and train/eval disparity;" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.697, + 0.905, + 0.74 + ], + "angle": 0, + "content": "- We propose to use the Batch Renormalization method to achieve seamless stitching without sacrificing prediction accuracy to the train/eval disparity;" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.741, + 0.905, + 0.785 + ], + "angle": 0, + "content": "- We validate our findings on three volumetric biological datasets, with two network architectures: CNN-based 3D U-Net and ViT-based UNETR." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.5, + 0.905, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.8, + 0.651, + 0.815 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Sliding window inference is used in different domains, including medical [19, 24], biological [8] and aerial [9] image analysis. It was shown that receptive field \\((RF)\\) [15, 16] plays a key role in the seamless stitching [17]. Often in case of dense prediction the input and output have the same" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13110" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.088, + 0.312, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.087, + 0.498, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.087, + 0.682, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.087, + 0.864, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.356, + 0.696, + 0.371 + ], + "angle": 0, + "content": "Figure 2. Examples of predictions with and without tiling artifacts" + }, + { + "type": "image", + "bbox": [ + 0.15, + 0.399, + 0.411, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.646, + 0.434, + 0.661 + ], + "angle": 0, + "content": "Figure 3. Illustration of the tile mismatch metric." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.901 + ], + "angle": 0, + "content": "size thanks to the zero-padding in convolutional layers, so the coordinates in both spaces can be used interchangeably. For the output pixels which are closer to the edge of the tile than \\(\\frac{1}{2} RF\\) part of the input values within the receptive field is replaced by zeros, creating mismatch in the predictions between tiles. Proposed solution is to crop the prediction and only use the part of the output where each pixel has full context for stitching. The minimal sufficient size of the cropped area is \\(\\frac{1}{2} RF\\). The same concept was discussed from the point of view of translational equivariance [9, 21]. Although convolutions and pooling are translation-equivariant, zero-padding makes the CNN overall lose this property. With valid convolutions the size of the output is smaller than the input but it allows to keep the translational" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.398, + 0.818, + 0.412 + ], + "angle": 0, + "content": "equivariance and therefore seamless stitching." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.906, + 0.517 + ], + "angle": 0, + "content": "In practice many pipelines using sliding window inference do not aim to completely eliminate the tiling artifacts but rather minimize them by using the largest possible tile size in inference and averaging. For example, nnU-Net [13] and Cellpose [22] compute weighted average of the tiles, with the weight decreasing from 1 in the middle of the tile to 0 at the border, smoothing the border between tiles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.907, + 0.653 + ], + "angle": 0, + "content": "These recommendations help against the discontinuity in predictions caused by edge effects but our experiments show that even in the valid convolution area predictions do not necessarily match (see Fig. 3). In addition, this does not explain hallucination-like artifacts in the background like in Fig. 1. It has been shown that prediction accuracy can depend significantly on the tile size [3, 18]. To our knowledge, the connection between these effects and feature normalization inside the network was not studied before." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.613, + 0.681 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.797 + ], + "angle": 0, + "content": "This section is structured as follows: we define how we perform sliding window inference and which normalization strategies we consider. We then propose two metrics that can identify normalization issues within the network, pointing out both tiling artifacts (\"tile mismatch\" metric) and performance drop in evaluation mode (\"train/eval disparity\" metric)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.821, + 0.819 + ], + "angle": 0, + "content": "3.1. Sliding window inference with halo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Input data can be a 2D image or a 3D volume with one or multiple channels. Here both images and volumes are referred to as images. For inference the image is split into a grid of tiles, incomplete tiles are padded with zeros. The stitched prediction is calculated from by tile-wise" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13111" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.201, + 0.09, + 0.805, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.325, + 0.828, + 0.34 + ], + "angle": 0, + "content": "Figure 4. Receptive field, shown as Log10 of mean gradient of the central pixel output with respect to the input" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.534 + ], + "angle": 0, + "content": "predictions. To ensure that observed tiling artifacts were not caused by edge effects, we calculated theoretical (TRF) and effective receptive field (ERF) of the networks following [15]. We define TRF as all pixels which belong to the computation graph for the target pixel in the output space. ERF is defined as width \\((\\frac{3\\sigma}{2})\\) of a Gaussian distribution fitted to the averaged gradient distribution During sliding window inference the overlap was set to the size of halo, minimizing the number of tiles needed to cover the whole volume. The halo part of the tile was completely removed so there was no averaging of predictions or other postprocessing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.542, + 0.317, + 0.558 + ], + "angle": 0, + "content": "3.2. Normalization strategies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.609 + ], + "angle": 0, + "content": "Normalization can be inserted in two parts of the pipeline: preprocessing of the input data and feature map normalization inside the network." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.618, + 0.279, + 0.632 + ], + "angle": 0, + "content": "3.2.1. Input normalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.484, + 0.757 + ], + "angle": 0, + "content": "Quantile normalization with \\( q_{min} = 0.01 \\) and \\( q_{max} = 0.98 \\) and clipping to [0..1] were applied to each image as a whole to convert the data type and value range from the microscope output to float32 and [0..1] range. In addition to the global normalization, quantile normalization can be applied to each tile separately to make the input of the network always have the same range and make tiles more comparable. Based on this we considered two strategies:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.76, + 0.48, + 0.775 + ], + "angle": 0, + "content": "- Global: quantile normalization only of the whole image." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.776, + 0.482, + 0.806 + ], + "angle": 0, + "content": "- Tile-wise: quantile normalization of each sampled tile, both during training and during inference." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.76, + 0.482, + 0.806 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.814, + 0.294, + 0.828 + ], + "angle": 0, + "content": "3.2.2. Feature normalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.832, + 0.484, + 0.864 + ], + "angle": 0, + "content": "General formula for the normalization operation with input \\(x\\), output \\(y\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.875, + 0.339, + 0.906 + ], + "angle": 0, + "content": "\\[\ny = \\frac {x - \\mu}{\\sqrt {\\sigma^ {2} + \\epsilon}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.518 + ], + "angle": 0, + "content": "where \\(\\mu\\) and \\(\\sigma\\) are normalization parameters and \\(\\epsilon\\) is a small constant used for numerical stability. Parameters \\(\\mu\\) and \\(\\sigma\\) can be estimated directly from the input \\(\\mu = \\operatorname{E}[x]\\) and \\(\\sigma^2 = \\operatorname{Var}[x]\\), where average can be taken either over each sample independently or over the whole batch. Alternatively, global normalization parameters independent of the current input can be used. A common strategy is to estimate the parameters as a running average over multiple samples: \\(p_{new} = (1 - momentum) \\times p_{old} + momentum \\times p_t\\). The update speed is determined by the momentum." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.526, + 0.907, + 0.571 + ], + "angle": 0, + "content": "- BatchNorm [12] Training: use statistics of the current batch and collect running average. Inference: use saved running average." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.571, + 0.906, + 0.661 + ], + "angle": 0, + "content": "- InstanceNorm [23] Both training and inference: use statistics of each input sample. In Pytorch this behavior can be changed by setting trackRunning.stats=True. Then during inference the statistics collected during training will be utilized, same as in BatchNorm." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.662, + 0.905, + 0.692 + ], + "angle": 0, + "content": "- BatchRenorm [11] Both training and inference: use running average statistics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.693, + 0.792, + 0.707 + ], + "angle": 0, + "content": "- Identity Skip feature normalization." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.526, + 0.907, + 0.707 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.527, + 0.724, + 0.783, + 0.739 + ], + "angle": 0, + "content": "3.3. Proposed evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.745, + 0.907, + 0.867 + ], + "angle": 0, + "content": "Train/eval disparity When global normalization with running average statistics is used during inference, a disparity can arise between the predictions of the same tile in train and eval mode as the train mode always uses statistics of the current batch. This disparity is an indicator of a potential drop in performance which is not related to model overfitting. To measure it quantitatively, we introduce the following metric:" + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.886, + 0.885, + 0.902 + ], + "angle": 0, + "content": "train/eval disparity \\(= 1 - \\text{Dice}(P_{\\text{train}}, P_{\\text{eval}})\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13112" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "where \\( P_{train} \\) is the prediction done with model.train() and \\( P_{eval} \\) - with model.eval()." + }, + { + "type": "title", + "bbox": [ + 0.106, + 0.122, + 0.21, + 0.135 + ], + "angle": 0, + "content": "Tile mismatch" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.137, + 0.484, + 0.244 + ], + "angle": 0, + "content": "Quantification of tiling artifacts is challenging because their magnitude depends on the similarity of the content of neighboring tiles in different parts of the image. We propose to split the image into overlapping tiles and compare the predictions in the overlap areas. To avoid edge effects, we only take the valid part of the overlap, as shown in Fig. 3. The following metrics are proposed:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.254, + 0.401, + 0.281 + ], + "angle": 0, + "content": "\\[\nm a x d i s t = \\max _ {i = 1} ^ {M} (| O _ {i 1} - O _ {i 2} |)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.303, + 0.458, + 0.333 + ], + "angle": 0, + "content": "\\[\nt i l e m i s m a t c h = \\underset {i = 1} {\\overset {M} {\\operatorname {m e d i a n}}} (1 - D i c e (O _ {i 1}, O _ {i 2})),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.341, + 0.484, + 0.476 + ], + "angle": 0, + "content": "where \\( O_{i1} \\) and \\( O_{i2} \\) are predictions in the valid overlap region in tile \\( i \\) and \\( M \\) is number of sampled tiles. In max dist max is taken over all channels and all tiles, making it an indicator of wether given setup produces tiling artifacts. In cases when max dist = 0 we explicitly report \"no\" instead of tile mismatch to emphasize that the predictions perfectly match. tile mismatch is calculated per channel and characterizes the magnitude of the artifacts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.492, + 0.224, + 0.51 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.517, + 0.194, + 0.531 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.629 + ], + "angle": 0, + "content": "Existing cell segmentation benchmarks often operate at the same image size as natural images and contain mostly the object-centered parts. Here the datasets were chosen to demonstrate the common practical challenges of microscopy segmentation such as large image size and variability of class distributions within the volume." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.734 + ], + "angle": 0, + "content": "organoids [7]: semantic segmentation of tissue in patient-derived colorectal cancer organoids, electron microscopy (FIB/SEM), \\(40 \\times 30 \\times 30 \\mathrm{~nm}^3\\) voxel size. The volume of size \\(1350 \\times 1506 \\times 1647\\) pixels was split 70/30 into train and validation dataset. Classes: cell boundaries, foreground (pixels inside the cells), extracellular matrix (space between cells and background)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.796 + ], + "angle": 0, + "content": "plants [25]: instance segmentation of nuclei in Arabidopsis ovules, confocal fluorescent microscopy, isotropic resolution of \\(0.13\\mu m^3\\) per pixel, 4 train volumes and 1 validation volume \\(\\sim 500\\times 1000\\times 1000\\) pixels each." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.856 + ], + "angle": 0, + "content": "embryo [2]: instance segmentation of nuclei in mouse embryos, confocal fluorescent microscopy, isotropic resolution of \\(0.2\\mu m^3\\) per pixel, 22 train volumes and 13 validation volumes, \\(\\sim 600\\times 1000\\times 700\\) px each." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "For plants and embryo datasets the original instance segmentation groundtruth was converted to semantic segmentation of three classes: object boundaries, foreground" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.09, + 0.862, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.479, + 0.907, + 0.507 + ], + "angle": 0, + "content": "Figure 5. Tiling artifacts become more pronounced in a transfer setting. For quantitative evaluation see STab. 1." + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.519, + 0.878, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.634, + 0.813, + 0.647 + ], + "angle": 0, + "content": "Figure 6. Performance vs tile size." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.72 + ], + "angle": 0, + "content": "(objects with subtracted boundaries) and background. For the transfer setup, networks trained for embryo dataset were used to segment plants dataset and vice versa." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.655, + 0.743 + ], + "angle": 0, + "content": "4.2. Architectures" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.869 + ], + "angle": 0, + "content": "3D U-Net: 3D U-Net with 2 downsampling steps and 2 convolutional blocks per level was used for all datasets. Each convolutional block consists of a convolutional layer, a normalization layer and a non-linearity. The normalization layers in all the blocks were replaced with one of the normalization layers listed above. Tile size (192, 192, 192) was used for training and (128, 128, 128) for sliding window inference." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "3D UNETR: UNETR consists of a ViT encoder and a convolutional decoder. In the encoder the volume is split" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13113" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.416, + 0.09, + 0.582, + 0.103 + ], + "angle": 0, + "content": "Table 1. U-Net results table" + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.116, + 0.888, + 0.484 + ], + "angle": 0, + "content": "
Normalization layerBatchNorm globalBatchNorm tile-wise globalInstanceNorm global tile-wiseInstanceNorm tile-wiseBatchRenorm globalIdentity global
Input norm
Feature norm
dice, eval modeorganoidsforeground0.950.880.940.930.950.95
boundaries0.490.150.750.750.730.72
matrix0.820.610.610.610.850.83
plantsforeground0.850.880.890.890.890.52
boundaries0.690.730.750.740.750.26
embryoforeground0.660.680.670.720.740.59
boundaries0.550.550.550.530.580.44
dice, train modeorganoidsforeground0.940.930.940.930.950.95
boundaries0.750.750.750.750.730.72
matrix0.610.610.610.610.850.83
plantsforeground0.890.890.890.890.890.52
boundaries0.750.740.750.740.750.26
embryoforeground0.670.720.670.720.740.59
boundaries0.550.520.550.530.580.44
train/eval disparityorganoids0.480.840000
plants0.170.110000
embryo0.370.290000
tile mismatchorganoidsno0.030.110.11nono
plantsno0.010.050.05nono
embryono0.090.160.20nono
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.482, + 0.676 + ], + "angle": 0, + "content": "into \\(16 \\times 16 \\times 16\\) patches and embeddings in the ViT are normalized using LayerNorm after each round of attention. With LayerNorm all features are normalized using average and variance of the features in each patch independently. Other normalization strategies can be used in the ViT [27], but here LayerNorm was kept in all experiments as a standard approach. Convolutional decoder consists of 4 upsampling steps with 2 convolutional blocks at each level of the U-Net-like architecture. The normalization layers in all the blocks were replaced with one of the normalization layers listed above." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.262, + 0.712 + ], + "angle": 0, + "content": "4.3. Reported metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Overall prediction quality: Prediction was done on the whole volume using sliding window inference as described above. If the setup included tile-wise input normalization, each tile was normalized separately before being passed through the network. The median Dice score over all samples in the validation dataset was reported for each class. Train/eval disparity: We report the Dice score both for training and evaluation mode to demonstrate the disparity in Tab. 1 and Tab. 2. Inference in both modes was done on the same images from the validation set, so the difference comes only from the behavior of the model in different modes and not overfitting to the training data. Tile mis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.905, + 0.615 + ], + "angle": 0, + "content": "match: The tiles of \\(192 \\times 192 \\times 288\\mathrm{px}\\) were sampled in a grid with a stride of 64 pixels to test the mismatch in the whole volume. Sampled tiles were split into two overlapping tiles of size \\(192 \\times 192 \\times 192\\). Both tiles were then processed to compare the predictions in the overlapping region. Tile mismatch values for boundary channel are reported in Tab. 1, Tab. 2 and STab. 1." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.616, + 0.867, + 0.632 + ], + "angle": 0, + "content": "Further details can be found in the Supplementary." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.643, + 0.602, + 0.658 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.663, + 0.684 + ], + "angle": 0, + "content": "5.1. Receptive field" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We calculated TRF and ERF for the architectures with different normalization layers. As expected, for U-Net with the global feature normalization (BatchNorm, BatchRenorm, Identity) only a limited region around the central pixel belongs to TRF, ERF constitutes the effective receptive field only the small portion as shown in Fig. 4. For the U-Net with tile-wise feature normalization (InstanceNorm) ERF is the same, which is expected since it is determined by the parameters of the convolutional architecture, but the TRF takes the whole tile as values of all pixels in the tile make a contribution to the prediction for the central pixel via normalization. Therefore, regardless of the size of the overlap between tiles in the sliding window prediction, the stitched prediction will not be exactly seam" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13114" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "less. In all sliding window experiments with the U-Net we used halo calculated using TRF of the U-Net with global feature normalization (Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.484, + 0.274 + ], + "angle": 0, + "content": "For UNETR TRF also takes the whole input tile due to normalization in the convolutional decoder and the attention mechanism. The averaged gradients image Fig. 4 has a clear square pattern because in the encoder the image is downsampled by 16 and then upsampled back to the original resolution. Despite the attention layers, ERF is the same as for the U-Net. Since only 3 central patches have large effect on the central pixel prediction, we set \\( halo = 32 \\) for all UNETR experiments (Fig. 4)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.285, + 0.484, + 0.316 + ], + "angle": 0, + "content": "5.2. Tile-wise feature normalization with InstanceNorm causes tiling artifacts" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.502 + ], + "angle": 0, + "content": "Using InstanceNorm allows to reach high quality predictions, however the tile-wise feature normalization is a source of artifacts both for U-Net (Tab. 1) and for UNETR (Tab. 2) as shown in Fig. 2. We observe that tile-wise input normalization (BatchNorm, tile-wise input norm) causes less severe tile mismatch than tile-wise network feature normalization (InstanceNorm, global input norm). Although it is impossible to reach completely seamless stitching with UNETR, experiments with InstanceNorm show strong tile mismatch, while global feature normalization and sufficient halo based on ERF size allow to reduce mismatch and make it almost unnoticeable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.64 + ], + "angle": 0, + "content": "Although InstanceNorm performs better than BatchNorm for boundaries, networks with tile-wise normalization struggle with classes which cover areas larger than the tile size, such as foreground and matrix in organoids dataset. Experiments with inference using different tile sizes Fig. 6 show that the prediction depends strongly on the tile size, reaching the best quality when the tiles become large enough for the feature distribution in each tile to reproduce the distribution in the whole image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Fig. 5 shows how artifacts get worse in the transfer setting: the model originally trained on the embryo dataset produces less artifacts than the model transferred from the plants dataset. We observe the same effect in transfer in both directions (STab. 1)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.483, + 0.759 + ], + "angle": 0, + "content": "5.3. Mismatch in batch statistics causes train and eval mode disparity for BatchNorm" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Global feature normalization with BatchNorm allows to eliminate the tile mismatch but predictions for organoids boundary and other datasets are worse than with InstanceNorm. The reason for this is the disparity between the performance of the network in train and evaluation mode. Estimation of normalization statistics with a running average in BatchNorm relies on the batch statistics being stable enough to represent the whole dataset. If batches are too small, the running average can be very dif" + }, + { + "type": "table_caption", + "bbox": [ + 0.621, + 0.09, + 0.8, + 0.103 + ], + "angle": 0, + "content": "Table 2. UNETR results table" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.115, + 0.905, + 0.566 + ], + "angle": 0, + "content": "
Norm layerBatchNormInstanceNormBatchRenormIdentity
Input normglobalglobalglobalglobal
Feature normglobaltile-wiseglobalglobal
dice,eval modeorganoids0.930.890.950.91
0.390.600.720.60
plants0.030.100.870.28
foreground0.840.850.850.52
boundaries0.640.690.660.21
foreground0.670.560.670.00
boundaries0.490.490.530.00
dice,train modeorganoids0.890.890.950.91
0.600.600.720.60
boundaries0.100.100.870.28
matrix0.850.850.850.52
foreground0.690.690.660.21
boundaries0.560.560.670.00
boundaries0.490.490.530.00
t/e disparityorganoids0.36000
plants0.20000
embryo0.33000
tile mismatchorganoids0.000.100.000.00
plants0.000.040.000.00
embryo0.000.060.000.00
" + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.584, + 0.857, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.823, + 0.833, + 0.838 + ], + "angle": 0, + "content": "Figure 7. Example of train/eval disparity" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "ferent from the batch-wise statistics used during training, leading to deteriorating performance. Doing inference with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13115" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.226 + ], + "angle": 0, + "content": "the network in train mode supports this hypothesis: the performance recovers, at the cost of tiling artifacts. In our experiments train and evaluation mode predictions can be quite different, with train/eval disparity reaching 0.84 for the organoids dataset. Qualitatively, the changes in the predictions occur in the areas of low prediction certainty. For example, in Fig. 7 the eval mode differs from the train mode in the areas where the background has a bright line artifact or in the noisier areas." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.482, + 0.348 + ], + "angle": 0, + "content": "This effect is not the same as the usual difference between train and evaluation performance caused by overfitting. Here, the network's output differs even for the exact same images thus violating the usual assumption that the prediction quality is approximately the same in train and eval mode. This issue would not be remedied by using augmentations or more training data as it does not address the high variance of statistics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.358, + 0.482, + 0.389 + ], + "angle": 0, + "content": "5.4. BatchRenorm corrects for train/eval disparity and provides seamless stitching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.482, + 0.47 + ], + "angle": 0, + "content": "InstanceNorm avoids the statistics mismatch by always using the statistics of the input both during training and during inference at the cost of causing tiling artifacts. BatchNorm does not introduce the tiling artifacts but can lose performance due to train/eval disparity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.482, + 0.606 + ], + "angle": 0, + "content": "A potential solution is to remove the normalization altogether. In our experiments not using normalization (Identity in Tab. 1 and Tab. 2) has led to worse performance for plants dataset and similar performance for organoids and embryo. It could be possible to improve the performance of networks with no normalization by finding more suitable training parameters and reducing the learning rate, however the necessity to adjust training parameters for each dataset makes the pipeline harder to use." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.482, + 0.743 + ], + "angle": 0, + "content": "Another solution is to use the same global normalization statistics both in training and in inference. Directly using running average statistics during training causes network's weights to explode [11], which is why we propose to use Batch Renormalization method. We found that networks with BatchRenorm produce same or better predictions as InstanceNorm (Tab. 1, Tab. 2) including the transfer setup (STab. 1) while allowing for seamless stitching due to global feature normalization (last column of Fig. 2, Fig. 5)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.205, + 0.77 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.482, + 0.901 + ], + "angle": 0, + "content": "The severity of tiling artifacts depends on the data and on the tile size. For example, the tile mismatch can be negligible in case the images are very homogeneous in content and the tile size is big enough that each tile mostly reproduces the global statistics of the image, such as for plants dataset in our experiments. However, even in that case the inference performance depends on the tile size and can change unpredictably in the transfer setting, putting higher" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "requirements on the hardware for inference and making the pipeline harder to use, especially for non-expert users. Tracing low validation accuracy to the instability in batch statistics is not straightforward either, as it is normally expected that the performance on the validation dataset should be worse than on the training data. If the training loss is low but the accuracy on validation data does not improve, it is often attributed to the insufficient amount of ground truth, however, we show that it can also be caused by train/eval disparity preventing efficient utilization of the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.905, + 0.378 + ], + "angle": 0, + "content": "Both of the described effects are highly dependent on the particular data, input normalization, sampling strategy, postprocessing and other parameters of the image analysis pipeline. For any such pipeline, we introduce two simple metrics: tile mismatch and train/eval disparity to help determine if it has issues with normalization. Both metrics are easy to use for any sliding window inference pipeline, including other dense prediction tasks if the Dice score is replaced with another, task-appropriate metric." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.905, + 0.621 + ], + "angle": 0, + "content": "We found that BatchRenorm, originally introduced for natural image classification, performs well for segmenting large biological images. The challenge with images larger than GPU memory is ensuring consistent normalization parameters across the entire image for seamless stitching, despite the inability to process the full image during training due to time constraints. BatchRenorm solves this by splitting training into two stages. First, the network is trained with BatchNorm. Then the normalization layer switches to using the accumulated running average statistics. In our experience it is very important that the training converges before switching and that the statistics change from tile-wise to global gradually to achieve the best final accuracy. While the BatchRenorm adds more hyperparameters, in our experiments the right moment to switch could easily be found from looking at the loss curves." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.634, + 0.65 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.905, + 0.87 + ], + "angle": 0, + "content": "In this paper, we demonstrate that tiling artifacts in sliding window inference pipelines can be caused by the tilewise feature normalization in the network. We also find that switching from tile-wise normalization in training to global normalization in evaluation can lead to even worse performance, including hallucinations, due to small batch size. As these effects strongly depend on the data and tile size, we introduce tile mismatch and train/eval disparity metrics to easily check if a specific pipeline and dataset have these issues. We propose that Batch Renormalization technique will yield artifact-free predictions even with small batch size and demonstrate its good performance through extensive experiments on various microscopy datasets and network architectures." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. We would like to thank Ricardo Sánchez Loayza and Qin Yu for helpful discussions. E.B." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13116" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.319 + ], + "angle": 0, + "content": "was supported by the Joachim Herz Foundation through an Add-on Fellowship for Interdisciplinary Life Science. E.D'I. was supported by a fellowship from the EMBL Interdisciplinary (EI4POD) program under Marie Sk1odowska-Curie Actions COFUND (847543). J.M. acknowledges funding from the EMBL and the European Research Council (ERC 3DCellPhase-760067). This work was supported by the European Commission through the Horizon Europe program (IMAGINE project, grant agreement 101094250-IMAGINE). The work of A.A. was funded by the German Research Foundation through PA 4341/2-1. C.P. received support under Germany's Excellence Strategy - EXC 2067/1-390729940. We would like to thank the EMBL IT Services Department for providing computational infrastructure and support." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.331, + 0.188, + 0.347 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.357, + 0.483, + 0.438 + ], + "angle": 0, + "content": "[1] Anwai Archit, Luca Freckmann, Sushmita Nair, Nabeel Khalid, Paul Hilt, Vikas Rajashekar, Marei Freitag, Carolin Teuber, Genevieve Buckley, Sebastian von Haaren, Sagnik Gupta, Andreas Dengel, Sheraz Ahmed, and Constantin Pape. Segment anything for microscopy. Nature Methods, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.44, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[2] Vladyslav Bondarenko, Mikhail Nikolaev, Dimitri Kromm, Roman Belousov, Adrian Wolny, Marloes Blotenburg, Peter Zeller, Saba Rezakhani, Johannes Hugger, Virginie Uhlmann, Lars Hufnagel, Anna Kreshuk, Jan Ellenberg, Alexander van Oudenaarden, Anna Erzberger, Matthias P Lutolf, and Takashi Hiiragi. Embryo-uterine interaction coordinates mouse embryogenesis during implantation. The EMBO Journal, 42(17), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[3] Calimanut-Ionut Cira, Miguel-Ángel Manso-Callejo, Naoto Yokoya, Tudor Sălăgean, and Ana-Cornelia Badea. Impact of tile size and tile overlap on the prediction performance of convolutional neural networks trained for road classification. Remote Sensing, 16(15):2818, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.622, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[6] Sven Dorkenwald, Arie Matslieh, Amy R Sterling, Philipp Schlegel, Szi-Chieh Yu, Claire E McKellar, Albert Lin, Marta Costa, Katharina Eichler, Yijie Yin, Will Silversmith, Casey Schneider-Mizell, Chris S Jordan, Derrick Brittain, Akhilesh Halageri, Kai Kuehner, Oluwaseun Ogedengbe, Ryan Morey, Jay Gager, Krzysztof Kruk, Eric Perlman, Runzhe Yang, David Deutsch, Doug Bland, Marissa Sorek, Ran Lu, Thomas Macrina, Kisuk Lee, J Alexander Bae, Shang Mu, Barak Nehoran, Eric Mitchell, Sergiy Popovych, Jingpeng Wu, Zhen Jia, Manuel A Castro, Nico Kemnitz, Dodam" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.357, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "Ih, Alexander Shakeel Bates, Nils Eckstein, Jan Funke, Forrest Collman, Davi D Bock, Gregory S X E Jefferis, H Sebastian Seung, Mala Murthy, and FlyWire Consortium. Neuronal wiring diagram of an adult brain. Nature, 634(8032): 124-138, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.164, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[7] Edoardo D'Imprima, Marta Garcia Montero, Sylwia Gawrzak, Paolo Ronchi, Ievgeniia Zagoriy, Yannick Schwab, Martin Jechlinger, and Julia Mahamid. Light and electron microscopy continuum-resolution imaging of 3d cell cultures. Developmental Cell, 58(7):616-632.e6, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.249, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[8] T. Falk, D. Mai, R. Bensch, Ö. Ciçek, A. Abdulkadir, Y. Marrakchi, A. Böhm, J. Deubner, Z. Jäckel, K. Seiwald, A. Dovzhenko, O. Tietz, C. Dal Bosco, S. Walsh, D. Saltukoglu, T. L. Tay, M. Prinz, K. Palme, M. Simons, I. Diester, T. Brox, and O. Ronneberger. U-net - deep learning for cell counting, detection, and morphometry. Nature Methods, 16: 67-70, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.347, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[9] Bohao Huang, Daniel Reichman, Leslie M Collins, Kyle Bradbury, and Jordan M Malof. Tiling and stitching segmentation output for remote sensing: Basic challenges and recommendations. arXiv:1805.12219, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[10] Lei Huang, Jie Qin, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Normalization techniques in training dnns: Methodology, analysis and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10173-10196, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[11] Sergey Ioffe. Batch renormalization: Towards reducing minibatch dependence in batch-normalized models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.6 + ], + "angle": 0, + "content": "[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning, pages 448-456, Lille, France, 2015. PMLR. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.603, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[13] Fabian Isensee and et al. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2):203-211, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[15] Vincent Loos, Rohit Pardasani, and Navchetan Awasthi. Demystifying the effect of receptive field size in u-net models for medical image segmentation. Journal of Medical Imaging, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[16] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[17] Michael Majurski and Peter Bajcsy. Exact tile-based segmentation inference for images larger thangpu memory. Journal of Research of the National Institute of Standards and Technology, 126, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13117" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[18] G. Anthony Reina, Ravi Panchumarthy, Siddhesh Pravin Thakur, Alexei Bastidas, and Spyridon Bakas. Systematic evaluation of image tiling adverse effects on deep learning semantic segmentation. Frontiers in Neuroscience, 14, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[19] O. Ronneberger, P.Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241. Springer, 2015. (available on arXiv:1505.04597 [cs.CV]). 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[20] Xiongtao Ruan, Matthew Mueller, Gaoxiang Liu, Frederik Gorlitz, Tian-Ming Fu, Daniel E Milkie, Joshua L Lillvis, Alexander Kuhn, Johnny Gan Chong, Jason Li Hong, Chu Yi Aaron Herr, Wilmene Hercule, Marc Nienhaus, Alison N Killilea, Eric Betzig, and Srigokul Upadhyayula. Image processing tools for petabyte-scale light sheet microscopy data. Nat. Methods, 21(12):2342-2352, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.332, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[21] Josef Lorenz Rumberger, Xiaoyan Yu, Peter Hirsch, Melanie Dohmen, Vanessa Emanuela Guarino, Ashkan Mokarian, Lisa Mais, Jan Funke, and Dagmar Kainmueller. How shift equivariance impacts metric learning for instance segmentation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), page 7108-7116. IEEE, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.417, + 0.482, + 0.458 + ], + "angle": 0, + "content": "[22] Carsen Stringer, Tim Wang, Michalis Michaelos, and Marius Pachitariu. Cellpose: a generalist algorithm for cellular segmentation. Nature Methods, 18(1):100-106, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[23] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization, 2016. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.502, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[24] Özgün undefinedigkeit, Ahmed Abdulkadir, Soeren S. Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation, page 424-432. Springer International Publishing, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[25] Athul Vijayan, Tejasvinee Atul Mody, Qin Yu, Adrian Wolny, Lorenzo Cerrone, Soeren Strauss, Miltos Tsiantis, Richard S. Smith, Fred A. Hamprecht, Anna Kreshuk, and Kay Schneitz. A deep learning-based toolkit for 3d nuclei segmentation and quantitative analysis in cellular and tissue context. Development, 151(14), 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.725 + ], + "angle": 0, + "content": "[26] Martin Weigert, Uwe Schmidt, Robert Haase, Ko Sugawara, and Gene Myers. Star-convex polyhedra for 3d object detection and segmentation in microscopy. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), page 3655–3662. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.727, + 0.482, + 0.782 + ], + "angle": 0, + "content": "[27] Zhuliang Yao, Yue Cao, Yutong Lin, Ze Liu, Zheng Zhang, and Han Hu. Leveraging batch normalization for vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). IEEE, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13118" + } + ] +] \ No newline at end of file diff --git a/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/eaeeea1f-eecf-4d2e-8480-0e89bb127b12_origin.pdf b/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/eaeeea1f-eecf-4d2e-8480-0e89bb127b12_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6f52be1eb896833818d59213dd04f876a1c6e6d4 --- /dev/null +++ b/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/eaeeea1f-eecf-4d2e-8480-0e89bb127b12_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae5ea2140a9e1c3d94032913a63ac6bb5d11b32657ff34c45c7c7befec46972d +size 2091446 diff --git a/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/full.md b/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..77ecb565ca47b34ba918e25ce355dc3f49050a43 --- /dev/null +++ b/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/full.md @@ -0,0 +1,271 @@ +# Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images + +Elena Buglakova1, Anwai Archit2, Edoardo D'Imprime1,4, Julia Mahamid1, Constantin Pape2,3, Anna Kreshuk1 + +$^{1}$ European Molecular Biology Laboratory, Heidelberg +$^{2}$ Institute of Computer Science, University of Göttingen + +3 Cluster of Excellence 'Multiscale Bioimaging: from Molecular Machines to Networks of Excitable Cells' (MBExC), Georg-August-University Göttingen + +$^{4}$ IRCCS Humanitas Research Hospital, via Manzoni 56, 20089 Rozzano, Milan, Italy + +# Abstract + +Segmentation of very large images is a common problem in microscopy, medical imaging or remote sensing. The problem is usually addressed by sliding window inference, which can theoretically lead to seamlessly stitched predictions. However, in practice many of the popular pipelines still suffer from tiling artifacts. We investigate the root cause of these issues and show that they stem from the normalization layers within the neural networks. We propose indicators to detect normalization issues and further explore the trade-offs between artifact-free and high-quality predictions, using three diverse microscopy datasets as examples. Finally, we propose to use BatchRenorm as the most suitable normalization strategy, which effectively removes tiling artifacts and enhances transfer performance, thereby improving the reusability of trained networks for new datasets. + +# 1. Introduction + +Segmentation is an essential first step of pipelines in biological image analysis, a diverse domain of computer vision that enables automated quantification for various types of microscopy images. While this domain has benefited a lot from the developments in natural image analysis, it also poses unique challenges. Investigation of biological diversity, discovery of unexpected phenotypes and development of novel imaging protocols leads to high variability of datasets. This explains the popularity of human-in-the-loop approaches to segmentation, where a small amount of data is either annotated from scratch or from the output of pretrained models [1, 22, 26], then a new model is either trained from scratch or fine-tuned. After that the next rounds of correction and training happen, until the desired quality is reached. Being able to quickly train a smaller model for an + +unseen image type is crucial in this setting. + +Another important feature of biological images, especially for volumetric imaging, is their large size. Electron microscopy volumes for connectomics [6] or light sheet microscopy volumes [20] can reach the scale of multiple terabytes. It is routine to work with 10-100 GB images of $\sim 1000^3$ pixels, which exceed GPU memory. + +As shown in Fig. 1, during training random patches are sampled from the annotated images and combined into a batch. During inference/evaluation, the image is split into a grid of tiles which are predicted independently and then stitched back to form the full prediction. Here we refer both to 2D and 3D image patches as tiles. In this setup the size of the tiles is limited from above by GPU memory and from below by the receptive field of the network [21]. Typical tile size is around $96 \times 96 \times 96 - 256 \times 256 \times 256$ pixels and the training batch size is 2-3 tiles per batch. During inference it is not necessary to store gradients, therefore either tile size or batch size can be larger. + +Tile-wise prediction and stitching might look like a purely technical step, however for microscopy patch sampling and stitching can significantly affect prediction quality and create artifacts. In contrast, big natural image datasets, such as ImageNet [5], CityScapes [4] and SAM dataset [14], usually contain relatively small 2D images, making tiling and stitching unnecessary, as well as allowing for larger batch size. + +It has been shown that seamless stitching with CNNs is possible if edge effects are taken into account [17]. For CNNs only a limited area of the input image called the receptive field affects the prediction value in a given pixel [16]. Pixels on the edge of the tile do not get full context for prediction, resulting in a mismatch between neighboring tiles. To deal with this issue the tiles are sampled with overlap and only the central part of each tile is used + +![](images/80ca87eae8407c0424cf6bca006a04415534dedf2e8a6091d6659dd4edc04aaf.jpg) +Figure 1. Illustration of the pipeline for processing images larger than GPU memory: random sampling of tiles during training and sliding window inference. Example of artifacts caused by tiling: hallucinations in the low signal areas and discontinuous predictions at the tile borders. + +for stitching. We refer to the removed area as halo. We find that even with sufficient halo sliding window inference can cause tiling artifacts manifesting in two ways: as abrupt discrepancy between predictions in neighboring tiles or as background hallucinations, as shown in Fig. 1 and Fig. 2. This problem is rarely discussed but it affects many widely used tools, including nnU-Net [13], a popular baseline method for biomedical image segmentation. + +A common way to deal with the tiling artifacts is to apply additional postprocessing, for example, to average neighboring tiles with gaussian weights to avoid discontinuous predictions [13], or to use additional thresholding by intensity and filtering by the object size to remove artifacts in the background like in the Fig. 1. However, adding heuristic postprocessing steps makes the whole pipeline harder to use and apply to new datasets. Moreover, we show that tiling artifacts become more pronounced in the transfer setting, further reducing reusability of the trained networks. + +According to our findings, the main cause of the artifacts is the tile-wise feature normalization inside the network. If InstanceNorm [23] is used in the CNN architecture, both during training and evaluation the statistics of + +the current input are used for normalization. This makes predictions for every pixel dependent on the content of the whole tile, therefore the limited receptive field assumption becomes invalid and, regardless of the size of the halo, it is impossible to make the stitching seamless. + +Unlike InstanceNorm, another popular normalization method BatchNorm [12] uses the statistics of the current input during training but switches to saved running average statistics in the evaluation mode, allowing to achieve seamless stitching. However, we show that the performance of the network drastically reduces in evaluation mode compared to training mode because the statistics used during training are not stable with small batch size, making the running average not representative of the statistics of individual batches. This effect has previously been observed for natural image classification [10, 11], but we find it to be significantly more pronounced in biological image segmentation. + +In this work we demonstrate the trade-off between the artifact-free stitching ensured by the global normalization with BatchNorm and prediction quality achieved by tilewise normalization with InstanceNorm. We suggest BatchRenorm [11] which uses the same global normalization statistics both during training and inference as a solution which produces artifact-free predictions without compromising on the prediction quality. + +In summary, we make the following contributions: + +- We show the existence of a normalization trade-off in segmentation of large biological images: using tile-wise statistics for feature normalization in the network during inference (InstanceNorm) leads to tiling artifacts even if edge effects were taken into consideration. Using tile-wise statistics in training and global accumulated statistics during inference (BatchNorm) leads to an unexpected difference between network performance during training and inference due to small batch size and high variability of the data; +- We suggest metrics for detecting normalization issues in networks used for sliding window inference pipelines: tile mismatch and train/eval disparity; +- We propose to use the Batch Renormalization method to achieve seamless stitching without sacrificing prediction accuracy to the train/eval disparity; +- We validate our findings on three volumetric biological datasets, with two network architectures: CNN-based 3D U-Net and ViT-based UNETR. + +# 2. Related work + +Sliding window inference is used in different domains, including medical [19, 24], biological [8] and aerial [9] image analysis. It was shown that receptive field $(RF)$ [15, 16] plays a key role in the seamless stitching [17]. Often in case of dense prediction the input and output have the same + +![](images/a0b2b4b552c58510840e0f70472d860965bf9ffcebad373fc8816d941b8832b6.jpg) +Figure 2. Examples of predictions with and without tiling artifacts + +![](images/1ee204b8bcb3325498477b93443645858a02c07ade0fbcef4fbed9200f08f27e.jpg) + +![](images/33977a809b71602cac52cf4d63a82fb40d183dd7bf70366efa592ca16a1baac4.jpg) + +![](images/aacde65da1b4d3e5d1ff0502a707b57f7c30de3f03cf62bc2c2f8306866cb64c.jpg) + +![](images/08eca0c9df4f2e5e91f0f137c444c31d4a54d77d14071f8c31d2c2f2eb23fc2a.jpg) +Figure 3. Illustration of the tile mismatch metric. + +size thanks to the zero-padding in convolutional layers, so the coordinates in both spaces can be used interchangeably. For the output pixels which are closer to the edge of the tile than $\frac{1}{2} RF$ part of the input values within the receptive field is replaced by zeros, creating mismatch in the predictions between tiles. Proposed solution is to crop the prediction and only use the part of the output where each pixel has full context for stitching. The minimal sufficient size of the cropped area is $\frac{1}{2} RF$ . The same concept was discussed from the point of view of translational equivariance [9, 21]. Although convolutions and pooling are translation-equivariant, zero-padding makes the CNN overall lose this property. With valid convolutions the size of the output is smaller than the input but it allows to keep the translational + +equivariance and therefore seamless stitching. + +In practice many pipelines using sliding window inference do not aim to completely eliminate the tiling artifacts but rather minimize them by using the largest possible tile size in inference and averaging. For example, nnU-Net [13] and Cellpose [22] compute weighted average of the tiles, with the weight decreasing from 1 in the middle of the tile to 0 at the border, smoothing the border between tiles. + +These recommendations help against the discontinuity in predictions caused by edge effects but our experiments show that even in the valid convolution area predictions do not necessarily match (see Fig. 3). In addition, this does not explain hallucination-like artifacts in the background like in Fig. 1. It has been shown that prediction accuracy can depend significantly on the tile size [3, 18]. To our knowledge, the connection between these effects and feature normalization inside the network was not studied before. + +# 3. Methods + +This section is structured as follows: we define how we perform sliding window inference and which normalization strategies we consider. We then propose two metrics that can identify normalization issues within the network, pointing out both tiling artifacts ("tile mismatch" metric) and performance drop in evaluation mode ("train/eval disparity" metric). + +# 3.1. Sliding window inference with halo + +Input data can be a 2D image or a 3D volume with one or multiple channels. Here both images and volumes are referred to as images. For inference the image is split into a grid of tiles, incomplete tiles are padded with zeros. The stitched prediction is calculated from by tile-wise + +![](images/2c9505d6eff9dcc13148bbc44a5a2471ec0739bce650fad4814ba2751dcd1805.jpg) +Figure 4. Receptive field, shown as Log10 of mean gradient of the central pixel output with respect to the input + +predictions. To ensure that observed tiling artifacts were not caused by edge effects, we calculated theoretical (TRF) and effective receptive field (ERF) of the networks following [15]. We define TRF as all pixels which belong to the computation graph for the target pixel in the output space. ERF is defined as width $(\frac{3\sigma}{2})$ of a Gaussian distribution fitted to the averaged gradient distribution During sliding window inference the overlap was set to the size of halo, minimizing the number of tiles needed to cover the whole volume. The halo part of the tile was completely removed so there was no averaging of predictions or other postprocessing. + +# 3.2. Normalization strategies + +Normalization can be inserted in two parts of the pipeline: preprocessing of the input data and feature map normalization inside the network. + +# 3.2.1. Input normalization + +Quantile normalization with $q_{min} = 0.01$ and $q_{max} = 0.98$ and clipping to [0..1] were applied to each image as a whole to convert the data type and value range from the microscope output to float32 and [0..1] range. In addition to the global normalization, quantile normalization can be applied to each tile separately to make the input of the network always have the same range and make tiles more comparable. Based on this we considered two strategies: + +- Global: quantile normalization only of the whole image. +- Tile-wise: quantile normalization of each sampled tile, both during training and during inference. + +# 3.2.2. Feature normalization + +General formula for the normalization operation with input $x$ , output $y$ is: + +$$ +y = \frac {x - \mu}{\sqrt {\sigma^ {2} + \epsilon}}, +$$ + +where $\mu$ and $\sigma$ are normalization parameters and $\epsilon$ is a small constant used for numerical stability. Parameters $\mu$ and $\sigma$ can be estimated directly from the input $\mu = \operatorname{E}[x]$ and $\sigma^2 = \operatorname{Var}[x]$ , where average can be taken either over each sample independently or over the whole batch. Alternatively, global normalization parameters independent of the current input can be used. A common strategy is to estimate the parameters as a running average over multiple samples: $p_{new} = (1 - momentum) \times p_{old} + momentum \times p_t$ . The update speed is determined by the momentum. + +- BatchNorm [12] Training: use statistics of the current batch and collect running average. Inference: use saved running average. +- InstanceNorm [23] Both training and inference: use statistics of each input sample. In Pytorch this behavior can be changed by setting trackRunning.stats=True. Then during inference the statistics collected during training will be utilized, same as in BatchNorm. +- BatchRenorm [11] Both training and inference: use running average statistics. +- Identity Skip feature normalization. + +# 3.3. Proposed evaluation metrics + +Train/eval disparity When global normalization with running average statistics is used during inference, a disparity can arise between the predictions of the same tile in train and eval mode as the train mode always uses statistics of the current batch. This disparity is an indicator of a potential drop in performance which is not related to model overfitting. To measure it quantitatively, we introduce the following metric: + +train/eval disparity $= 1 - \text{Dice}(P_{\text{train}}, P_{\text{eval}})$ , + +where $P_{train}$ is the prediction done with model.train() and $P_{eval}$ - with model.eval(). + +# Tile mismatch + +Quantification of tiling artifacts is challenging because their magnitude depends on the similarity of the content of neighboring tiles in different parts of the image. We propose to split the image into overlapping tiles and compare the predictions in the overlap areas. To avoid edge effects, we only take the valid part of the overlap, as shown in Fig. 3. The following metrics are proposed: + +$$ +m a x d i s t = \max _ {i = 1} ^ {M} (| O _ {i 1} - O _ {i 2} |) +$$ + +$$ +t i l e m i s m a t c h = \underset {i = 1} {\overset {M} {\operatorname {m e d i a n}}} (1 - D i c e (O _ {i 1}, O _ {i 2})), +$$ + +where $O_{i1}$ and $O_{i2}$ are predictions in the valid overlap region in tile $i$ and $M$ is number of sampled tiles. In max dist max is taken over all channels and all tiles, making it an indicator of wether given setup produces tiling artifacts. In cases when max dist = 0 we explicitly report "no" instead of tile mismatch to emphasize that the predictions perfectly match. tile mismatch is calculated per channel and characterizes the magnitude of the artifacts. + +# 4. Experiments + +# 4.1. Datasets + +Existing cell segmentation benchmarks often operate at the same image size as natural images and contain mostly the object-centered parts. Here the datasets were chosen to demonstrate the common practical challenges of microscopy segmentation such as large image size and variability of class distributions within the volume. + +organoids [7]: semantic segmentation of tissue in patient-derived colorectal cancer organoids, electron microscopy (FIB/SEM), $40 \times 30 \times 30 \mathrm{~nm}^3$ voxel size. The volume of size $1350 \times 1506 \times 1647$ pixels was split 70/30 into train and validation dataset. Classes: cell boundaries, foreground (pixels inside the cells), extracellular matrix (space between cells and background). + +plants [25]: instance segmentation of nuclei in Arabidopsis ovules, confocal fluorescent microscopy, isotropic resolution of $0.13\mu m^3$ per pixel, 4 train volumes and 1 validation volume $\sim 500\times 1000\times 1000$ pixels each. + +embryo [2]: instance segmentation of nuclei in mouse embryos, confocal fluorescent microscopy, isotropic resolution of $0.2\mu m^3$ per pixel, 22 train volumes and 13 validation volumes, $\sim 600\times 1000\times 700$ px each. + +For plants and embryo datasets the original instance segmentation groundtruth was converted to semantic segmentation of three classes: object boundaries, foreground + +![](images/25de00239db4f38cf74f0a6a8ea8005a7a1b5828fdeef28f6a0690ca61ff9823.jpg) +Figure 5. Tiling artifacts become more pronounced in a transfer setting. For quantitative evaluation see STab. 1. + +![](images/e512195231098beb6a7a82913eaec4db07e017029fb627ab51a1893fcd107152.jpg) +Figure 6. Performance vs tile size. + +(objects with subtracted boundaries) and background. For the transfer setup, networks trained for embryo dataset were used to segment plants dataset and vice versa. + +# 4.2. Architectures + +3D U-Net: 3D U-Net with 2 downsampling steps and 2 convolutional blocks per level was used for all datasets. Each convolutional block consists of a convolutional layer, a normalization layer and a non-linearity. The normalization layers in all the blocks were replaced with one of the normalization layers listed above. Tile size (192, 192, 192) was used for training and (128, 128, 128) for sliding window inference. + +3D UNETR: UNETR consists of a ViT encoder and a convolutional decoder. In the encoder the volume is split + +Table 1. U-Net results table + +
Normalization layerBatchNorm globalBatchNorm tile-wise globalInstanceNorm global tile-wiseInstanceNorm tile-wiseBatchRenorm globalIdentity global
Input norm
Feature norm
dice, eval modeorganoidsforeground0.950.880.940.930.950.95
boundaries0.490.150.750.750.730.72
matrix0.820.610.610.610.850.83
plantsforeground0.850.880.890.890.890.52
boundaries0.690.730.750.740.750.26
embryoforeground0.660.680.670.720.740.59
boundaries0.550.550.550.530.580.44
dice, train modeorganoidsforeground0.940.930.940.930.950.95
boundaries0.750.750.750.750.730.72
matrix0.610.610.610.610.850.83
plantsforeground0.890.890.890.890.890.52
boundaries0.750.740.750.740.750.26
embryoforeground0.670.720.670.720.740.59
boundaries0.550.520.550.530.580.44
train/eval disparityorganoids0.480.840000
plants0.170.110000
embryo0.370.290000
tile mismatchorganoidsno0.030.110.11nono
plantsno0.010.050.05nono
embryono0.090.160.20nono
+ +into $16 \times 16 \times 16$ patches and embeddings in the ViT are normalized using LayerNorm after each round of attention. With LayerNorm all features are normalized using average and variance of the features in each patch independently. Other normalization strategies can be used in the ViT [27], but here LayerNorm was kept in all experiments as a standard approach. Convolutional decoder consists of 4 upsampling steps with 2 convolutional blocks at each level of the U-Net-like architecture. The normalization layers in all the blocks were replaced with one of the normalization layers listed above. + +# 4.3. Reported metrics + +Overall prediction quality: Prediction was done on the whole volume using sliding window inference as described above. If the setup included tile-wise input normalization, each tile was normalized separately before being passed through the network. The median Dice score over all samples in the validation dataset was reported for each class. Train/eval disparity: We report the Dice score both for training and evaluation mode to demonstrate the disparity in Tab. 1 and Tab. 2. Inference in both modes was done on the same images from the validation set, so the difference comes only from the behavior of the model in different modes and not overfitting to the training data. Tile mis + +match: The tiles of $192 \times 192 \times 288\mathrm{px}$ were sampled in a grid with a stride of 64 pixels to test the mismatch in the whole volume. Sampled tiles were split into two overlapping tiles of size $192 \times 192 \times 192$ . Both tiles were then processed to compare the predictions in the overlapping region. Tile mismatch values for boundary channel are reported in Tab. 1, Tab. 2 and STab. 1. + +Further details can be found in the Supplementary. + +# 5. Results + +# 5.1. Receptive field + +We calculated TRF and ERF for the architectures with different normalization layers. As expected, for U-Net with the global feature normalization (BatchNorm, BatchRenorm, Identity) only a limited region around the central pixel belongs to TRF, ERF constitutes the effective receptive field only the small portion as shown in Fig. 4. For the U-Net with tile-wise feature normalization (InstanceNorm) ERF is the same, which is expected since it is determined by the parameters of the convolutional architecture, but the TRF takes the whole tile as values of all pixels in the tile make a contribution to the prediction for the central pixel via normalization. Therefore, regardless of the size of the overlap between tiles in the sliding window prediction, the stitched prediction will not be exactly seam + +less. In all sliding window experiments with the U-Net we used halo calculated using TRF of the U-Net with global feature normalization (Fig. 4). + +For UNETR TRF also takes the whole input tile due to normalization in the convolutional decoder and the attention mechanism. The averaged gradients image Fig. 4 has a clear square pattern because in the encoder the image is downsampled by 16 and then upsampled back to the original resolution. Despite the attention layers, ERF is the same as for the U-Net. Since only 3 central patches have large effect on the central pixel prediction, we set $halo = 32$ for all UNETR experiments (Fig. 4). + +# 5.2. Tile-wise feature normalization with InstanceNorm causes tiling artifacts + +Using InstanceNorm allows to reach high quality predictions, however the tile-wise feature normalization is a source of artifacts both for U-Net (Tab. 1) and for UNETR (Tab. 2) as shown in Fig. 2. We observe that tile-wise input normalization (BatchNorm, tile-wise input norm) causes less severe tile mismatch than tile-wise network feature normalization (InstanceNorm, global input norm). Although it is impossible to reach completely seamless stitching with UNETR, experiments with InstanceNorm show strong tile mismatch, while global feature normalization and sufficient halo based on ERF size allow to reduce mismatch and make it almost unnoticeable. + +Although InstanceNorm performs better than BatchNorm for boundaries, networks with tile-wise normalization struggle with classes which cover areas larger than the tile size, such as foreground and matrix in organoids dataset. Experiments with inference using different tile sizes Fig. 6 show that the prediction depends strongly on the tile size, reaching the best quality when the tiles become large enough for the feature distribution in each tile to reproduce the distribution in the whole image. + +Fig. 5 shows how artifacts get worse in the transfer setting: the model originally trained on the embryo dataset produces less artifacts than the model transferred from the plants dataset. We observe the same effect in transfer in both directions (STab. 1). + +# 5.3. Mismatch in batch statistics causes train and eval mode disparity for BatchNorm + +Global feature normalization with BatchNorm allows to eliminate the tile mismatch but predictions for organoids boundary and other datasets are worse than with InstanceNorm. The reason for this is the disparity between the performance of the network in train and evaluation mode. Estimation of normalization statistics with a running average in BatchNorm relies on the batch statistics being stable enough to represent the whole dataset. If batches are too small, the running average can be very dif + +Table 2. UNETR results table + +
Norm layerBatchNormInstanceNormBatchRenormIdentity
Input normglobalglobalglobalglobal
Feature normglobaltile-wiseglobalglobal
dice,eval modeorganoids0.930.890.950.91
0.390.600.720.60
plants0.030.100.870.28
foreground0.840.850.850.52
boundaries0.640.690.660.21
foreground0.670.560.670.00
boundaries0.490.490.530.00
dice,train modeorganoids0.890.890.950.91
0.600.600.720.60
boundaries0.100.100.870.28
matrix0.850.850.850.52
foreground0.690.690.660.21
boundaries0.560.560.670.00
boundaries0.490.490.530.00
t/e disparityorganoids0.36000
plants0.20000
embryo0.33000
tile mismatchorganoids0.000.100.000.00
plants0.000.040.000.00
embryo0.000.060.000.00
+ +![](images/5ef99ac44483fe539c5193224e0f09484458018ac6b4bc432cfe96a4d66bc126.jpg) +Figure 7. Example of train/eval disparity + +ferent from the batch-wise statistics used during training, leading to deteriorating performance. Doing inference with + +the network in train mode supports this hypothesis: the performance recovers, at the cost of tiling artifacts. In our experiments train and evaluation mode predictions can be quite different, with train/eval disparity reaching 0.84 for the organoids dataset. Qualitatively, the changes in the predictions occur in the areas of low prediction certainty. For example, in Fig. 7 the eval mode differs from the train mode in the areas where the background has a bright line artifact or in the noisier areas. + +This effect is not the same as the usual difference between train and evaluation performance caused by overfitting. Here, the network's output differs even for the exact same images thus violating the usual assumption that the prediction quality is approximately the same in train and eval mode. This issue would not be remedied by using augmentations or more training data as it does not address the high variance of statistics. + +# 5.4. BatchRenorm corrects for train/eval disparity and provides seamless stitching + +InstanceNorm avoids the statistics mismatch by always using the statistics of the input both during training and during inference at the cost of causing tiling artifacts. BatchNorm does not introduce the tiling artifacts but can lose performance due to train/eval disparity. + +A potential solution is to remove the normalization altogether. In our experiments not using normalization (Identity in Tab. 1 and Tab. 2) has led to worse performance for plants dataset and similar performance for organoids and embryo. It could be possible to improve the performance of networks with no normalization by finding more suitable training parameters and reducing the learning rate, however the necessity to adjust training parameters for each dataset makes the pipeline harder to use. + +Another solution is to use the same global normalization statistics both in training and in inference. Directly using running average statistics during training causes network's weights to explode [11], which is why we propose to use Batch Renormalization method. We found that networks with BatchRenorm produce same or better predictions as InstanceNorm (Tab. 1, Tab. 2) including the transfer setup (STab. 1) while allowing for seamless stitching due to global feature normalization (last column of Fig. 2, Fig. 5). + +# 6. Discussion + +The severity of tiling artifacts depends on the data and on the tile size. For example, the tile mismatch can be negligible in case the images are very homogeneous in content and the tile size is big enough that each tile mostly reproduces the global statistics of the image, such as for plants dataset in our experiments. However, even in that case the inference performance depends on the tile size and can change unpredictably in the transfer setting, putting higher + +requirements on the hardware for inference and making the pipeline harder to use, especially for non-expert users. Tracing low validation accuracy to the instability in batch statistics is not straightforward either, as it is normally expected that the performance on the validation dataset should be worse than on the training data. If the training loss is low but the accuracy on validation data does not improve, it is often attributed to the insufficient amount of ground truth, however, we show that it can also be caused by train/eval disparity preventing efficient utilization of the ground truth. + +Both of the described effects are highly dependent on the particular data, input normalization, sampling strategy, postprocessing and other parameters of the image analysis pipeline. For any such pipeline, we introduce two simple metrics: tile mismatch and train/eval disparity to help determine if it has issues with normalization. Both metrics are easy to use for any sliding window inference pipeline, including other dense prediction tasks if the Dice score is replaced with another, task-appropriate metric. + +We found that BatchRenorm, originally introduced for natural image classification, performs well for segmenting large biological images. The challenge with images larger than GPU memory is ensuring consistent normalization parameters across the entire image for seamless stitching, despite the inability to process the full image during training due to time constraints. BatchRenorm solves this by splitting training into two stages. First, the network is trained with BatchNorm. Then the normalization layer switches to using the accumulated running average statistics. In our experience it is very important that the training converges before switching and that the statistics change from tile-wise to global gradually to achieve the best final accuracy. While the BatchRenorm adds more hyperparameters, in our experiments the right moment to switch could easily be found from looking at the loss curves. + +# 7. Conclusion + +In this paper, we demonstrate that tiling artifacts in sliding window inference pipelines can be caused by the tilewise feature normalization in the network. We also find that switching from tile-wise normalization in training to global normalization in evaluation can lead to even worse performance, including hallucinations, due to small batch size. As these effects strongly depend on the data and tile size, we introduce tile mismatch and train/eval disparity metrics to easily check if a specific pipeline and dataset have these issues. We propose that Batch Renormalization technique will yield artifact-free predictions even with small batch size and demonstrate its good performance through extensive experiments on various microscopy datasets and network architectures. + +Acknowledgment. We would like to thank Ricardo Sánchez Loayza and Qin Yu for helpful discussions. E.B. + +was supported by the Joachim Herz Foundation through an Add-on Fellowship for Interdisciplinary Life Science. E.D'I. was supported by a fellowship from the EMBL Interdisciplinary (EI4POD) program under Marie Sk1odowska-Curie Actions COFUND (847543). J.M. acknowledges funding from the EMBL and the European Research Council (ERC 3DCellPhase-760067). This work was supported by the European Commission through the Horizon Europe program (IMAGINE project, grant agreement 101094250-IMAGINE). The work of A.A. was funded by the German Research Foundation through PA 4341/2-1. C.P. received support under Germany's Excellence Strategy - EXC 2067/1-390729940. We would like to thank the EMBL IT Services Department for providing computational infrastructure and support. + +# References + +[1] Anwai Archit, Luca Freckmann, Sushmita Nair, Nabeel Khalid, Paul Hilt, Vikas Rajashekar, Marei Freitag, Carolin Teuber, Genevieve Buckley, Sebastian von Haaren, Sagnik Gupta, Andreas Dengel, Sheraz Ahmed, and Constantin Pape. Segment anything for microscopy. Nature Methods, 2025. 1 +[2] Vladyslav Bondarenko, Mikhail Nikolaev, Dimitri Kromm, Roman Belousov, Adrian Wolny, Marloes Blotenburg, Peter Zeller, Saba Rezakhani, Johannes Hugger, Virginie Uhlmann, Lars Hufnagel, Anna Kreshuk, Jan Ellenberg, Alexander van Oudenaarden, Anna Erzberger, Matthias P Lutolf, and Takashi Hiiragi. Embryo-uterine interaction coordinates mouse embryogenesis during implantation. The EMBO Journal, 42(17), 2023. 5 +[3] Calimanut-Ionut Cira, Miguel-Ángel Manso-Callejo, Naoto Yokoya, Tudor Sălăgean, and Ana-Cornelia Badea. Impact of tile size and tile overlap on the prediction performance of convolutional neural networks trained for road classification. Remote Sensing, 16(15):2818, 2024. 3 +[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 1 +[6] Sven Dorkenwald, Arie Matslieh, Amy R Sterling, Philipp Schlegel, Szi-Chieh Yu, Claire E McKellar, Albert Lin, Marta Costa, Katharina Eichler, Yijie Yin, Will Silversmith, Casey Schneider-Mizell, Chris S Jordan, Derrick Brittain, Akhilesh Halageri, Kai Kuehner, Oluwaseun Ogedengbe, Ryan Morey, Jay Gager, Krzysztof Kruk, Eric Perlman, Runzhe Yang, David Deutsch, Doug Bland, Marissa Sorek, Ran Lu, Thomas Macrina, Kisuk Lee, J Alexander Bae, Shang Mu, Barak Nehoran, Eric Mitchell, Sergiy Popovych, Jingpeng Wu, Zhen Jia, Manuel A Castro, Nico Kemnitz, Dodam + +Ih, Alexander Shakeel Bates, Nils Eckstein, Jan Funke, Forrest Collman, Davi D Bock, Gregory S X E Jefferis, H Sebastian Seung, Mala Murthy, and FlyWire Consortium. Neuronal wiring diagram of an adult brain. Nature, 634(8032): 124-138, 2024. 1 +[7] Edoardo D'Imprima, Marta Garcia Montero, Sylwia Gawrzak, Paolo Ronchi, Ievgeniia Zagoriy, Yannick Schwab, Martin Jechlinger, and Julia Mahamid. Light and electron microscopy continuum-resolution imaging of 3d cell cultures. Developmental Cell, 58(7):616-632.e6, 2023. 5 +[8] T. Falk, D. Mai, R. Bensch, Ö. Ciçek, A. Abdulkadir, Y. Marrakchi, A. Böhm, J. Deubner, Z. Jäckel, K. Seiwald, A. Dovzhenko, O. Tietz, C. Dal Bosco, S. Walsh, D. Saltukoglu, T. L. Tay, M. Prinz, K. Palme, M. Simons, I. Diester, T. Brox, and O. Ronneberger. U-net - deep learning for cell counting, detection, and morphometry. Nature Methods, 16: 67-70, 2019. 2 +[9] Bohao Huang, Daniel Reichman, Leslie M Collins, Kyle Bradbury, and Jordan M Malof. Tiling and stitching segmentation output for remote sensing: Basic challenges and recommendations. arXiv:1805.12219, 2018. 2, 3 +[10] Lei Huang, Jie Qin, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Normalization techniques in training dnns: Methodology, analysis and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10173-10196, 2023. 2 +[11] Sergey Ioffe. Batch renormalization: Towards reducing minibatch dependence in batch-normalized models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 4, 8 +[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning, pages 448-456, Lille, France, 2015. PMLR. 2, 4 +[13] Fabian Isensee and et al. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2):203-211, 2020. 2, 3 +[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 1 +[15] Vincent Loos, Rohit Pardasani, and Navchetan Awasthi. Demystifying the effect of receptive field size in u-net models for medical image segmentation. Journal of Medical Imaging, 2024. 2, 4 +[16] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 1, 2 +[17] Michael Majurski and Peter Bajcsy. Exact tile-based segmentation inference for images larger thangpu memory. Journal of Research of the National Institute of Standards and Technology, 126, 2021. 1, 2 + +[18] G. Anthony Reina, Ravi Panchumarthy, Siddhesh Pravin Thakur, Alexei Bastidas, and Spyridon Bakas. Systematic evaluation of image tiling adverse effects on deep learning semantic segmentation. Frontiers in Neuroscience, 14, 2020. 3 +[19] O. Ronneberger, P.Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241. Springer, 2015. (available on arXiv:1505.04597 [cs.CV]). 2 +[20] Xiongtao Ruan, Matthew Mueller, Gaoxiang Liu, Frederik Gorlitz, Tian-Ming Fu, Daniel E Milkie, Joshua L Lillvis, Alexander Kuhn, Johnny Gan Chong, Jason Li Hong, Chu Yi Aaron Herr, Wilmene Hercule, Marc Nienhaus, Alison N Killilea, Eric Betzig, and Srigokul Upadhyayula. Image processing tools for petabyte-scale light sheet microscopy data. Nat. Methods, 21(12):2342-2352, 2024. 1 +[21] Josef Lorenz Rumberger, Xiaoyan Yu, Peter Hirsch, Melanie Dohmen, Vanessa Emanuela Guarino, Ashkan Mokarian, Lisa Mais, Jan Funke, and Dagmar Kainmueller. How shift equivariance impacts metric learning for instance segmentation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), page 7108-7116. IEEE, 2021. 1, 3 +[22] Carsen Stringer, Tim Wang, Michalis Michaelos, and Marius Pachitariu. Cellpose: a generalist algorithm for cellular segmentation. Nature Methods, 18(1):100-106, 2020. 1, 3 +[23] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization, 2016. 2, 4 +[24] Özgün undefinedigkeit, Ahmed Abdulkadir, Soeren S. Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation, page 424-432. Springer International Publishing, 2016. 2 +[25] Athul Vijayan, Tejasvinee Atul Mody, Qin Yu, Adrian Wolny, Lorenzo Cerrone, Soeren Strauss, Miltos Tsiantis, Richard S. Smith, Fred A. Hamprecht, Anna Kreshuk, and Kay Schneitz. A deep learning-based toolkit for 3d nuclei segmentation and quantitative analysis in cellular and tissue context. Development, 151(14), 2024. 5 +[26] Martin Weigert, Uwe Schmidt, Robert Haase, Ko Sugawara, and Gene Myers. Star-convex polyhedra for 3d object detection and segmentation in microscopy. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), page 3655–3662. IEEE, 2020. 1 +[27] Zhuliang Yao, Yue Cao, Yutong Lin, Ze Liu, Zheng Zhang, and Han Hu. Leveraging batch normalization for vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). IEEE, 2021. 6 \ No newline at end of file diff --git a/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/images.zip b/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b697e4736e16f9cd6d13d266a26aba41642990b7 --- /dev/null +++ b/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f8f3fa2faaec73b4f06fc08194aaddf213ff4f2cbdd416b68358400a60592fe8 +size 534610 diff --git a/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/layout.json b/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c612270e3932802f1af72aec8618254d34ebc016 --- /dev/null +++ b/2025/Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images/layout.json @@ -0,0 +1,6538 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 68, + 102, + 544, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 102, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 68, + 102, + 544, + 140 + ], + "type": "text", + "content": "Tiling artifacts and trade-offs of feature normalization in the segmentation of large biological images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 160, + 487, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 160, + 487, + 189 + ], + "spans": [ + { + "bbox": [ + 122, + 160, + 487, + 189 + ], + "type": "text", + "content": "Elena Buglakova1, Anwai Archit2, Edoardo D'Imprime1,4, Julia Mahamid1, Constantin Pape2,3, Anna Kreshuk1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 201, + 189, + 409, + 217 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 204, + 189, + 405, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 189, + 405, + 202 + ], + "spans": [ + { + "bbox": [ + 204, + 189, + 405, + 202 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 204, + 189, + 405, + 202 + ], + "type": "text", + "content": " European Molecular Biology Laboratory, Heidelberg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 201, + 203, + 409, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 203, + 409, + 217 + ], + "spans": [ + { + "bbox": [ + 201, + 203, + 409, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 201, + 203, + 409, + 217 + ], + "type": "text", + "content": " Institute of Computer Science, University of Göttingen" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 110, + 217, + 499, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 217, + 499, + 244 + ], + "spans": [ + { + "bbox": [ + 110, + 217, + 499, + 244 + ], + "type": "text", + "content": "3 Cluster of Excellence 'Multiscale Bioimaging: from Molecular Machines to Networks of Excitable Cells' (MBExC), Georg-August-University Göttingen" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 245, + 460, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 245, + 460, + 258 + ], + "spans": [ + { + "bbox": [ + 146, + 245, + 460, + 258 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 146, + 245, + 460, + 258 + ], + "type": "text", + "content": " IRCCS Humanitas Research Hospital, via Manzoni 56, 20089 Rozzano, Milan, Italy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 312, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 491 + ], + "type": "text", + "content": "Segmentation of very large images is a common problem in microscopy, medical imaging or remote sensing. The problem is usually addressed by sliding window inference, which can theoretically lead to seamlessly stitched predictions. However, in practice many of the popular pipelines still suffer from tiling artifacts. We investigate the root cause of these issues and show that they stem from the normalization layers within the neural networks. We propose indicators to detect normalization issues and further explore the trade-offs between artifact-free and high-quality predictions, using three diverse microscopy datasets as examples. Finally, we propose to use BatchRenorm as the most suitable normalization strategy, which effectively removes tiling artifacts and enhances transfer performance, thereby improving the reusability of trained networks for new datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 514, + 135, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 135, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 135, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 533, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 713 + ], + "type": "text", + "content": "Segmentation is an essential first step of pipelines in biological image analysis, a diverse domain of computer vision that enables automated quantification for various types of microscopy images. While this domain has benefited a lot from the developments in natural image analysis, it also poses unique challenges. Investigation of biological diversity, discovery of unexpected phenotypes and development of novel imaging protocols leads to high variability of datasets. This explains the popularity of human-in-the-loop approaches to segmentation, where a small amount of data is either annotated from scratch or from the output of pretrained models [1, 22, 26], then a new model is either trained from scratch or fine-tuned. After that the next rounds of correction and training happen, until the desired quality is reached. Being able to quickly train a smaller model for an" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 288, + 488, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 488, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 488, + 300 + ], + "type": "text", + "content": "unseen image type is crucial in this setting." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 301, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 554, + 373 + ], + "type": "text", + "content": "Another important feature of biological images, especially for volumetric imaging, is their large size. Electron microscopy volumes for connectomics [6] or light sheet microscopy volumes [20] can reach the scale of multiple terabytes. It is routine to work with 10-100 GB images of " + }, + { + "bbox": [ + 313, + 301, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\sim 1000^3" + }, + { + "bbox": [ + 313, + 301, + 554, + 373 + ], + "type": "text", + "content": " pixels, which exceed GPU memory." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 375, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 554, + 518 + ], + "type": "text", + "content": "As shown in Fig. 1, during training random patches are sampled from the annotated images and combined into a batch. During inference/evaluation, the image is split into a grid of tiles which are predicted independently and then stitched back to form the full prediction. Here we refer both to 2D and 3D image patches as tiles. In this setup the size of the tiles is limited from above by GPU memory and from below by the receptive field of the network [21]. Typical tile size is around " + }, + { + "bbox": [ + 313, + 375, + 554, + 518 + ], + "type": "inline_equation", + "content": "96 \\times 96 \\times 96 - 256 \\times 256 \\times 256" + }, + { + "bbox": [ + 313, + 375, + 554, + 518 + ], + "type": "text", + "content": " pixels and the training batch size is 2-3 tiles per batch. During inference it is not necessary to store gradients, therefore either tile size or batch size can be larger." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 616 + ], + "type": "text", + "content": "Tile-wise prediction and stitching might look like a purely technical step, however for microscopy patch sampling and stitching can significantly affect prediction quality and create artifacts. In contrast, big natural image datasets, such as ImageNet [5], CityScapes [4] and SAM dataset [14], usually contain relatively small 2D images, making tiling and stitching unnecessary, as well as allowing for larger batch size." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "It has been shown that seamless stitching with CNNs is possible if edge effects are taken into account [17]. For CNNs only a limited area of the input image called the receptive field affects the prediction value in a given pixel [16]. Pixels on the edge of the tile do not get full context for prediction, resulting in a mismatch between neighboring tiles. To deal with this issue the tiles are sampled with overlap and only the central part of each tile is used" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13109" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 70, + 265, + 355 + ], + "blocks": [ + { + "bbox": [ + 86, + 70, + 265, + 355 + ], + "lines": [ + { + "bbox": [ + 86, + 70, + 265, + 355 + ], + "spans": [ + { + "bbox": [ + 86, + 70, + 265, + 355 + ], + "type": "image", + "image_path": "80ca87eae8407c0424cf6bca006a04415534dedf2e8a6091d6659dd4edc04aaf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 367, + 295, + 422 + ], + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 422 + ], + "type": "text", + "content": "Figure 1. Illustration of the pipeline for processing images larger than GPU memory: random sampling of tiles during training and sliding window inference. Example of artifacts caused by tiling: hallucinations in the low signal areas and discontinuous predictions at the tile borders." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 447, + 295, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 543 + ], + "type": "text", + "content": "for stitching. We refer to the removed area as halo. We find that even with sufficient halo sliding window inference can cause tiling artifacts manifesting in two ways: as abrupt discrepancy between predictions in neighboring tiles or as background hallucinations, as shown in Fig. 1 and Fig. 2. This problem is rarely discussed but it affects many widely used tools, including nnU-Net [13], a popular baseline method for biomedical image segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 664 + ], + "type": "text", + "content": "A common way to deal with the tiling artifacts is to apply additional postprocessing, for example, to average neighboring tiles with gaussian weights to avoid discontinuous predictions [13], or to use additional thresholding by intensity and filtering by the object size to remove artifacts in the background like in the Fig. 1. However, adding heuristic postprocessing steps makes the whole pipeline harder to use and apply to new datasets. Moreover, we show that tiling artifacts become more pronounced in the transfer setting, further reducing reusability of the trained networks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "According to our findings, the main cause of the artifacts is the tile-wise feature normalization inside the network. If InstanceNorm [23] is used in the CNN architecture, both during training and evaluation the statistics of" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": "the current input are used for normalization. This makes predictions for every pixel dependent on the content of the whole tile, therefore the limited receptive field assumption becomes invalid and, regardless of the size of the halo, it is impossible to make the stitching seamless." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 133, + 555, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 285 + ], + "type": "text", + "content": "Unlike InstanceNorm, another popular normalization method BatchNorm [12] uses the statistics of the current input during training but switches to saved running average statistics in the evaluation mode, allowing to achieve seamless stitching. However, we show that the performance of the network drastically reduces in evaluation mode compared to training mode because the statistics used during training are not stable with small batch size, making the running average not representative of the statistics of individual batches. This effect has previously been observed for natural image classification [10, 11], but we find it to be significantly more pronounced in biological image segmentation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 288, + 554, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 383 + ], + "type": "text", + "content": "In this work we demonstrate the trade-off between the artifact-free stitching ensured by the global normalization with BatchNorm and prediction quality achieved by tilewise normalization with InstanceNorm. We suggest BatchRenorm [11] which uses the same global normalization statistics both during training and inference as a solution which produces artifact-free predictions without compromising on the prediction quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 324, + 384, + 529, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 384, + 529, + 395 + ], + "spans": [ + { + "bbox": [ + 324, + 384, + 529, + 395 + ], + "type": "text", + "content": "In summary, we make the following contributions:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 396, + 553, + 621 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 396, + 553, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 553, + 514 + ], + "type": "text", + "content": "- We show the existence of a normalization trade-off in segmentation of large biological images: using tile-wise statistics for feature normalization in the network during inference (InstanceNorm) leads to tiling artifacts even if edge effects were taken into consideration. Using tile-wise statistics in training and global accumulated statistics during inference (BatchNorm) leads to an unexpected difference between network performance during training and inference due to small batch size and high variability of the data;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 515, + 553, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 515, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 515, + 553, + 550 + ], + "type": "text", + "content": "- We suggest metrics for detecting normalization issues in networks used for sliding window inference pipelines: tile mismatch and train/eval disparity;" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 552, + 553, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 553, + 586 + ], + "type": "text", + "content": "- We propose to use the Batch Renormalization method to achieve seamless stitching without sacrificing prediction accuracy to the train/eval disparity;" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 586, + 553, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 553, + 621 + ], + "type": "text", + "content": "- We validate our findings on three volumetric biological datasets, with two network architectures: CNN-based 3D U-Net and ViT-based UNETR." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 633, + 398, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 633, + 398, + 645 + ], + "spans": [ + { + "bbox": [ + 314, + 633, + 398, + 645 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": "Sliding window inference is used in different domains, including medical [19, 24], biological [8] and aerial [9] image analysis. It was shown that receptive field " + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "inline_equation", + "content": "(RF)" + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": " [15, 16] plays a key role in the seamless stitching [17]. Often in case of dense prediction the input and output have the same" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13110" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 69, + 190, + 274 + ], + "blocks": [ + { + "bbox": [ + 82, + 69, + 190, + 274 + ], + "lines": [ + { + "bbox": [ + 82, + 69, + 190, + 274 + ], + "spans": [ + { + "bbox": [ + 82, + 69, + 190, + 274 + ], + "type": "image", + "image_path": "a0b2b4b552c58510840e0f70472d860965bf9ffcebad373fc8816d941b8832b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 281, + 425, + 293 + ], + "lines": [ + { + "bbox": [ + 184, + 281, + 425, + 293 + ], + "spans": [ + { + "bbox": [ + 184, + 281, + 425, + 293 + ], + "type": "text", + "content": "Figure 2. Examples of predictions with and without tiling artifacts" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 209, + 68, + 304, + 274 + ], + "blocks": [ + { + "bbox": [ + 209, + 68, + 304, + 274 + ], + "lines": [ + { + "bbox": [ + 209, + 68, + 304, + 274 + ], + "spans": [ + { + "bbox": [ + 209, + 68, + 304, + 274 + ], + "type": "image", + "image_path": "1ee204b8bcb3325498477b93443645858a02c07ade0fbcef4fbed9200f08f27e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 323, + 68, + 417, + 274 + ], + "blocks": [ + { + "bbox": [ + 323, + 68, + 417, + 274 + ], + "lines": [ + { + "bbox": [ + 323, + 68, + 417, + 274 + ], + "spans": [ + { + "bbox": [ + 323, + 68, + 417, + 274 + ], + "type": "image", + "image_path": "33977a809b71602cac52cf4d63a82fb40d183dd7bf70366efa592ca16a1baac4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 434, + 68, + 528, + 274 + ], + "blocks": [ + { + "bbox": [ + 434, + 68, + 528, + 274 + ], + "lines": [ + { + "bbox": [ + 434, + 68, + 528, + 274 + ], + "spans": [ + { + "bbox": [ + 434, + 68, + 528, + 274 + ], + "type": "image", + "image_path": "aacde65da1b4d3e5d1ff0502a707b57f7c30de3f03cf62bc2c2f8306866cb64c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 91, + 316, + 251, + 504 + ], + "blocks": [ + { + "bbox": [ + 91, + 316, + 251, + 504 + ], + "lines": [ + { + "bbox": [ + 91, + 316, + 251, + 504 + ], + "spans": [ + { + "bbox": [ + 91, + 316, + 251, + 504 + ], + "type": "image", + "image_path": "08eca0c9df4f2e5e91f0f137c444c31d4a54d77d14071f8c31d2c2f2eb23fc2a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 511, + 265, + 523 + ], + "lines": [ + { + "bbox": [ + 86, + 511, + 265, + 523 + ], + "spans": [ + { + "bbox": [ + 86, + 511, + 265, + 523 + ], + "type": "text", + "content": "Figure 3. Illustration of the tile mismatch metric." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": "size thanks to the zero-padding in convolutional layers, so the coordinates in both spaces can be used interchangeably. For the output pixels which are closer to the edge of the tile than " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\frac{1}{2} RF" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": " part of the input values within the receptive field is replaced by zeros, creating mismatch in the predictions between tiles. Proposed solution is to crop the prediction and only use the part of the output where each pixel has full context for stitching. The minimal sufficient size of the cropped area is " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\frac{1}{2} RF" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": ". The same concept was discussed from the point of view of translational equivariance [9, 21]. Although convolutions and pooling are translation-equivariant, zero-padding makes the CNN overall lose this property. With valid convolutions the size of the output is smaller than the input but it allows to keep the translational" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 315, + 500, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 500, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 500, + 326 + ], + "type": "text", + "content": "equivariance and therefore seamless stitching." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 326, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 554, + 409 + ], + "type": "text", + "content": "In practice many pipelines using sliding window inference do not aim to completely eliminate the tiling artifacts but rather minimize them by using the largest possible tile size in inference and averaging. For example, nnU-Net [13] and Cellpose [22] compute weighted average of the tiles, with the weight decreasing from 1 in the middle of the tile to 0 at the border, smoothing the border between tiles." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 410, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 517 + ], + "type": "text", + "content": "These recommendations help against the discontinuity in predictions caused by edge effects but our experiments show that even in the valid convolution area predictions do not necessarily match (see Fig. 3). In addition, this does not explain hallucination-like artifacts in the background like in Fig. 1. It has been shown that prediction accuracy can depend significantly on the tile size [3, 18]. To our knowledge, the connection between these effects and feature normalization inside the network was not studied before." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 526, + 375, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 375, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 375, + 539 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 546, + 554, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 631 + ], + "type": "text", + "content": "This section is structured as follows: we define how we perform sliding window inference and which normalization strategies we consider. We then propose two metrics that can identify normalization issues within the network, pointing out both tiling artifacts (\"tile mismatch\" metric) and performance drop in evaluation mode (\"train/eval disparity\" metric)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 635, + 502, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 502, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 502, + 648 + ], + "type": "text", + "content": "3.1. Sliding window inference with halo" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "Input data can be a 2D image or a 3D volume with one or multiple channels. Here both images and volumes are referred to as images. For inference the image is split into a grid of tiles, incomplete tiles are padded with zeros. The stitched prediction is calculated from by tile-wise" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13111" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 71, + 492, + 249 + ], + "blocks": [ + { + "bbox": [ + 123, + 71, + 492, + 249 + ], + "lines": [ + { + "bbox": [ + 123, + 71, + 492, + 249 + ], + "spans": [ + { + "bbox": [ + 123, + 71, + 492, + 249 + ], + "type": "image", + "image_path": "2c9505d6eff9dcc13148bbc44a5a2471ec0739bce650fad4814ba2751dcd1805.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 103, + 257, + 506, + 269 + ], + "lines": [ + { + "bbox": [ + 103, + 257, + 506, + 269 + ], + "spans": [ + { + "bbox": [ + 103, + 257, + 506, + 269 + ], + "type": "text", + "content": "Figure 4. Receptive field, shown as Log10 of mean gradient of the central pixel output with respect to the input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 422 + ], + "type": "text", + "content": "predictions. To ensure that observed tiling artifacts were not caused by edge effects, we calculated theoretical (TRF) and effective receptive field (ERF) of the networks following [15]. We define TRF as all pixels which belong to the computation graph for the target pixel in the output space. ERF is defined as width " + }, + { + "bbox": [ + 55, + 289, + 296, + 422 + ], + "type": "inline_equation", + "content": "(\\frac{3\\sigma}{2})" + }, + { + "bbox": [ + 55, + 289, + 296, + 422 + ], + "type": "text", + "content": " of a Gaussian distribution fitted to the averaged gradient distribution During sliding window inference the overlap was set to the size of halo, minimizing the number of tiles needed to cover the whole volume. The halo part of the tile was completely removed so there was no averaging of predictions or other postprocessing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 429, + 194, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 194, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 194, + 441 + ], + "type": "text", + "content": "3.2. Normalization strategies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 447, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 482 + ], + "type": "text", + "content": "Normalization can be inserted in two parts of the pipeline: preprocessing of the input data and feature map normalization inside the network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 489, + 170, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 170, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 170, + 500 + ], + "type": "text", + "content": "3.2.1. Input normalization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 504, + 296, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 296, + 599 + ], + "type": "text", + "content": "Quantile normalization with " + }, + { + "bbox": [ + 55, + 504, + 296, + 599 + ], + "type": "inline_equation", + "content": "q_{min} = 0.01" + }, + { + "bbox": [ + 55, + 504, + 296, + 599 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 504, + 296, + 599 + ], + "type": "inline_equation", + "content": "q_{max} = 0.98" + }, + { + "bbox": [ + 55, + 504, + 296, + 599 + ], + "type": "text", + "content": " and clipping to [0..1] were applied to each image as a whole to convert the data type and value range from the microscope output to float32 and [0..1] range. In addition to the global normalization, quantile normalization can be applied to each tile separately to make the input of the network always have the same range and make tiles more comparable. Based on this we considered two strategies:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 601, + 294, + 638 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 601, + 293, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 293, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 293, + 613 + ], + "type": "text", + "content": "- Global: quantile normalization only of the whole image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 614, + 294, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 294, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 294, + 638 + ], + "type": "text", + "content": "- Tile-wise: quantile normalization of each sampled tile, both during training and during inference." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 644, + 179, + 655 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 179, + 655 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 179, + 655 + ], + "type": "text", + "content": "3.2.2. Feature normalization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 658, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 296, + 684 + ], + "type": "text", + "content": "General formula for the normalization operation with input " + }, + { + "bbox": [ + 55, + 658, + 296, + 684 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 658, + 296, + 684 + ], + "type": "text", + "content": ", output " + }, + { + "bbox": [ + 55, + 658, + 296, + 684 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 658, + 296, + 684 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 143, + 693, + 207, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 693, + 207, + 717 + ], + "spans": [ + { + "bbox": [ + 143, + 693, + 207, + 717 + ], + "type": "interline_equation", + "content": "y = \\frac {x - \\mu}{\\sqrt {\\sigma^ {2} + \\epsilon}},", + "image_path": "4fa5ec3cfa4c2074d3e973ec3404ba3955242d834932c54964f66847b0371c61.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "text", + "content": " are normalization parameters and " + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "text", + "content": " is a small constant used for numerical stability. Parameters " + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "text", + "content": " can be estimated directly from the input " + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\mu = \\operatorname{E}[x]" + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\sigma^2 = \\operatorname{Var}[x]" + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "text", + "content": ", where average can be taken either over each sample independently or over the whole batch. Alternatively, global normalization parameters independent of the current input can be used. A common strategy is to estimate the parameters as a running average over multiple samples: " + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "inline_equation", + "content": "p_{new} = (1 - momentum) \\times p_{old} + momentum \\times p_t" + }, + { + "bbox": [ + 313, + 289, + 555, + 410 + ], + "type": "text", + "content": ". The update speed is determined by the momentum." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 416, + 555, + 559 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 416, + 555, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 555, + 452 + ], + "type": "text", + "content": "- BatchNorm [12] Training: use statistics of the current batch and collect running average. Inference: use saved running average." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 452, + 554, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 554, + 523 + ], + "type": "text", + "content": "- InstanceNorm [23] Both training and inference: use statistics of each input sample. In Pytorch this behavior can be changed by setting trackRunning.stats=True. Then during inference the statistics collected during training will be utilized, same as in BatchNorm." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 524, + 553, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 553, + 548 + ], + "type": "text", + "content": "- BatchRenorm [11] Both training and inference: use running average statistics." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 548, + 484, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 484, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 484, + 559 + ], + "type": "text", + "content": "- Identity Skip feature normalization." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 573, + 479, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 573, + 479, + 585 + ], + "spans": [ + { + "bbox": [ + 322, + 573, + 479, + 585 + ], + "type": "text", + "content": "3.3. Proposed evaluation metrics" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 590, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 590, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 321, + 590, + 555, + 686 + ], + "type": "text", + "content": "Train/eval disparity When global normalization with running average statistics is used during inference, a disparity can arise between the predictions of the same tile in train and eval mode as the train mode always uses statistics of the current batch. This disparity is an indicator of a potential drop in performance which is not related to model overfitting. To measure it quantitatively, we introduce the following metric:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 334, + 701, + 541, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 701, + 541, + 714 + ], + "spans": [ + { + "bbox": [ + 334, + 701, + 541, + 714 + ], + "type": "text", + "content": "train/eval disparity " + }, + { + "bbox": [ + 334, + 701, + 541, + 714 + ], + "type": "inline_equation", + "content": "= 1 - \\text{Dice}(P_{\\text{train}}, P_{\\text{eval}})" + }, + { + "bbox": [ + 334, + 701, + 541, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13112" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 294, + 95 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 63, + 72, + 294, + 95 + ], + "type": "inline_equation", + "content": "P_{train}" + }, + { + "bbox": [ + 63, + 72, + 294, + 95 + ], + "type": "text", + "content": " is the prediction done with model.train() and " + }, + { + "bbox": [ + 63, + 72, + 294, + 95 + ], + "type": "inline_equation", + "content": "P_{eval}" + }, + { + "bbox": [ + 63, + 72, + 294, + 95 + ], + "type": "text", + "content": " - with model.eval()." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 96, + 128, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 96, + 128, + 106 + ], + "spans": [ + { + "bbox": [ + 64, + 96, + 128, + 106 + ], + "type": "text", + "content": "Tile mismatch" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 108, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 108, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 63, + 108, + 296, + 193 + ], + "type": "text", + "content": "Quantification of tiling artifacts is challenging because their magnitude depends on the similarity of the content of neighboring tiles in different parts of the image. We propose to split the image into overlapping tiles and compare the predictions in the overlap areas. To avoid edge effects, we only take the valid part of the overlap, as shown in Fig. 3. The following metrics are proposed:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 201, + 245, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 201, + 245, + 222 + ], + "spans": [ + { + "bbox": [ + 115, + 201, + 245, + 222 + ], + "type": "interline_equation", + "content": "m a x d i s t = \\max _ {i = 1} ^ {M} (| O _ {i 1} - O _ {i 2} |)", + "image_path": "c4afe0e54e80f7da53f8ca919146efda555dd74614b036a2c2071373eb9e9be7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 239, + 280, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 239, + 280, + 263 + ], + "spans": [ + { + "bbox": [ + 79, + 239, + 280, + 263 + ], + "type": "interline_equation", + "content": "t i l e m i s m a t c h = \\underset {i = 1} {\\overset {M} {\\operatorname {m e d i a n}}} (1 - D i c e (O _ {i 1}, O _ {i 2})),", + "image_path": "aa0938fe88f0341085c31adde8aea2b6c8b9db3f2d7b32487e04ebe5bff5bdd9.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 270, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 270, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 63, + 270, + 296, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 63, + 270, + 296, + 376 + ], + "type": "inline_equation", + "content": "O_{i1}" + }, + { + "bbox": [ + 63, + 270, + 296, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 63, + 270, + 296, + 376 + ], + "type": "inline_equation", + "content": "O_{i2}" + }, + { + "bbox": [ + 63, + 270, + 296, + 376 + ], + "type": "text", + "content": " are predictions in the valid overlap region in tile " + }, + { + "bbox": [ + 63, + 270, + 296, + 376 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 63, + 270, + 296, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 63, + 270, + 296, + 376 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 63, + 270, + 296, + 376 + ], + "type": "text", + "content": " is number of sampled tiles. In max dist max is taken over all channels and all tiles, making it an indicator of wether given setup produces tiling artifacts. In cases when max dist = 0 we explicitly report \"no\" instead of tile mismatch to emphasize that the predictions perfectly match. tile mismatch is calculated per channel and characterizes the magnitude of the artifacts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 137, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 137, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 137, + 403 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 409, + 118, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 118, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 118, + 420 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 498 + ], + "type": "text", + "content": "Existing cell segmentation benchmarks often operate at the same image size as natural images and contain mostly the object-centered parts. Here the datasets were chosen to demonstrate the common practical challenges of microscopy segmentation such as large image size and variability of class distributions within the volume." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "content": "organoids [7]: semantic segmentation of tissue in patient-derived colorectal cancer organoids, electron microscopy (FIB/SEM), " + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "inline_equation", + "content": "40 \\times 30 \\times 30 \\mathrm{~nm}^3" + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "content": " voxel size. The volume of size " + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "inline_equation", + "content": "1350 \\times 1506 \\times 1647" + }, + { + "bbox": [ + 55, + 498, + 296, + 581 + ], + "type": "text", + "content": " pixels was split 70/30 into train and validation dataset. Classes: cell boundaries, foreground (pixels inside the cells), extracellular matrix (space between cells and background)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": "plants [25]: instance segmentation of nuclei in Arabidopsis ovules, confocal fluorescent microscopy, isotropic resolution of " + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "inline_equation", + "content": "0.13\\mu m^3" + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": " per pixel, 4 train volumes and 1 validation volume " + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\sim 500\\times 1000\\times 1000" + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": " pixels each." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "content": "embryo [2]: instance segmentation of nuclei in mouse embryos, confocal fluorescent microscopy, isotropic resolution of " + }, + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "inline_equation", + "content": "0.2\\mu m^3" + }, + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "content": " per pixel, 22 train volumes and 13 validation volumes, " + }, + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\sim 600\\times 1000\\times 700" + }, + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "content": " px each." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "For plants and embryo datasets the original instance segmentation groundtruth was converted to semantic segmentation of three classes: object boundaries, foreground" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 340, + 71, + 527, + 371 + ], + "blocks": [ + { + "bbox": [ + 340, + 71, + 527, + 371 + ], + "lines": [ + { + "bbox": [ + 340, + 71, + 527, + 371 + ], + "spans": [ + { + "bbox": [ + 340, + 71, + 527, + 371 + ], + "type": "image", + "image_path": "25de00239db4f38cf74f0a6a8ea8005a7a1b5828fdeef28f6a0690ca61ff9823.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 379, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 401 + ], + "type": "text", + "content": "Figure 5. Tiling artifacts become more pronounced in a transfer setting. For quantitative evaluation see STab. 1." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 332, + 411, + 537, + 494 + ], + "blocks": [ + { + "bbox": [ + 332, + 411, + 537, + 494 + ], + "lines": [ + { + "bbox": [ + 332, + 411, + 537, + 494 + ], + "spans": [ + { + "bbox": [ + 332, + 411, + 537, + 494 + ], + "type": "image", + "image_path": "e512195231098beb6a7a82913eaec4db07e017029fb627ab51a1893fcd107152.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 502, + 497, + 512 + ], + "lines": [ + { + "bbox": [ + 370, + 502, + 497, + 512 + ], + "spans": [ + { + "bbox": [ + 370, + 502, + 497, + 512 + ], + "type": "text", + "content": "Figure 6. Performance vs tile size." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "text", + "content": "(objects with subtracted boundaries) and background. For the transfer setup, networks trained for embryo dataset were used to segment plants dataset and vice versa." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 576, + 400, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 400, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 400, + 588 + ], + "type": "text", + "content": "4.2. Architectures" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 688 + ], + "type": "text", + "content": "3D U-Net: 3D U-Net with 2 downsampling steps and 2 convolutional blocks per level was used for all datasets. Each convolutional block consists of a convolutional layer, a normalization layer and a non-linearity. The normalization layers in all the blocks were replaced with one of the normalization layers listed above. Tile size (192, 192, 192) was used for training and (128, 128, 128) for sliding window inference." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "3D UNETR: UNETR consists of a ViT encoder and a convolutional decoder. In the encoder the volume is split" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13113" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 91, + 543, + 383 + ], + "blocks": [ + { + "bbox": [ + 254, + 71, + 356, + 81 + ], + "lines": [ + { + "bbox": [ + 254, + 71, + 356, + 81 + ], + "spans": [ + { + "bbox": [ + 254, + 71, + 356, + 81 + ], + "type": "text", + "content": "Table 1. U-Net results table" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 91, + 543, + 383 + ], + "lines": [ + { + "bbox": [ + 67, + 91, + 543, + 383 + ], + "spans": [ + { + "bbox": [ + 67, + 91, + 543, + 383 + ], + "type": "table", + "html": "
Normalization layerBatchNorm globalBatchNorm tile-wise globalInstanceNorm global tile-wiseInstanceNorm tile-wiseBatchRenorm globalIdentity global
Input norm
Feature norm
dice, eval modeorganoidsforeground0.950.880.940.930.950.95
boundaries0.490.150.750.750.730.72
matrix0.820.610.610.610.850.83
plantsforeground0.850.880.890.890.890.52
boundaries0.690.730.750.740.750.26
embryoforeground0.660.680.670.720.740.59
boundaries0.550.550.550.530.580.44
dice, train modeorganoidsforeground0.940.930.940.930.950.95
boundaries0.750.750.750.750.730.72
matrix0.610.610.610.610.850.83
plantsforeground0.890.890.890.890.890.52
boundaries0.750.740.750.740.750.26
embryoforeground0.670.720.670.720.740.59
boundaries0.550.520.550.530.580.44
train/eval disparityorganoids0.480.840000
plants0.170.110000
embryo0.370.290000
tile mismatchorganoidsno0.030.110.11nono
plantsno0.010.050.05nono
embryono0.090.160.20nono
", + "image_path": "0d245ad8b35f61ff9f5c00ad69c6303228b3aca6658280a0d07a2b1954cf0c0d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 403, + 294, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 294, + 535 + ], + "type": "text", + "content": "into " + }, + { + "bbox": [ + 55, + 403, + 294, + 535 + ], + "type": "inline_equation", + "content": "16 \\times 16 \\times 16" + }, + { + "bbox": [ + 55, + 403, + 294, + 535 + ], + "type": "text", + "content": " patches and embeddings in the ViT are normalized using LayerNorm after each round of attention. With LayerNorm all features are normalized using average and variance of the features in each patch independently. Other normalization strategies can be used in the ViT [27], but here LayerNorm was kept in all experiments as a standard approach. Convolutional decoder consists of 4 upsampling steps with 2 convolutional blocks at each level of the U-Net-like architecture. The normalization layers in all the blocks were replaced with one of the normalization layers listed above." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 550, + 160, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 160, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 160, + 563 + ], + "type": "text", + "content": "4.3. Reported metrics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Overall prediction quality: Prediction was done on the whole volume using sliding window inference as described above. If the setup included tile-wise input normalization, each tile was normalized separately before being passed through the network. The median Dice score over all samples in the validation dataset was reported for each class. Train/eval disparity: We report the Dice score both for training and evaluation mode to demonstrate the disparity in Tab. 1 and Tab. 2. Inference in both modes was done on the same images from the validation set, so the difference comes only from the behavior of the model in different modes and not overfitting to the training data. Tile mis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 403, + 553, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 487 + ], + "type": "text", + "content": "match: The tiles of " + }, + { + "bbox": [ + 313, + 403, + 553, + 487 + ], + "type": "inline_equation", + "content": "192 \\times 192 \\times 288\\mathrm{px}" + }, + { + "bbox": [ + 313, + 403, + 553, + 487 + ], + "type": "text", + "content": " were sampled in a grid with a stride of 64 pixels to test the mismatch in the whole volume. Sampled tiles were split into two overlapping tiles of size " + }, + { + "bbox": [ + 313, + 403, + 553, + 487 + ], + "type": "inline_equation", + "content": "192 \\times 192 \\times 192" + }, + { + "bbox": [ + 313, + 403, + 553, + 487 + ], + "type": "text", + "content": ". Both tiles were then processed to compare the predictions in the overlapping region. Tile mismatch values for boundary channel are reported in Tab. 1, Tab. 2 and STab. 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 487, + 530, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 487, + 530, + 500 + ], + "spans": [ + { + "bbox": [ + 326, + 487, + 530, + 500 + ], + "type": "text", + "content": "Further details can be found in the Supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 509, + 368, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 509, + 368, + 521 + ], + "spans": [ + { + "bbox": [ + 314, + 509, + 368, + 521 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 529, + 405, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 405, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 405, + 541 + ], + "type": "text", + "content": "5.1. Receptive field" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "We calculated TRF and ERF for the architectures with different normalization layers. As expected, for U-Net with the global feature normalization (BatchNorm, BatchRenorm, Identity) only a limited region around the central pixel belongs to TRF, ERF constitutes the effective receptive field only the small portion as shown in Fig. 4. For the U-Net with tile-wise feature normalization (InstanceNorm) ERF is the same, which is expected since it is determined by the parameters of the convolutional architecture, but the TRF takes the whole tile as values of all pixels in the tile make a contribution to the prediction for the central pixel via normalization. Therefore, regardless of the size of the overlap between tiles in the sliding window prediction, the stitched prediction will not be exactly seam" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13114" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "less. In all sliding window experiments with the U-Net we used halo calculated using TRF of the U-Net with global feature normalization (Fig. 4)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 296, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 296, + 217 + ], + "type": "text", + "content": "For UNETR TRF also takes the whole input tile due to normalization in the convolutional decoder and the attention mechanism. The averaged gradients image Fig. 4 has a clear square pattern because in the encoder the image is downsampled by 16 and then upsampled back to the original resolution. Despite the attention layers, ERF is the same as for the U-Net. Since only 3 central patches have large effect on the central pixel prediction, we set " + }, + { + "bbox": [ + 55, + 109, + 296, + 217 + ], + "type": "inline_equation", + "content": "halo = 32" + }, + { + "bbox": [ + 55, + 109, + 296, + 217 + ], + "type": "text", + "content": " for all UNETR experiments (Fig. 4)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 225, + 296, + 250 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 296, + 250 + ], + "type": "text", + "content": "5.2. Tile-wise feature normalization with InstanceNorm causes tiling artifacts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 397 + ], + "type": "text", + "content": "Using InstanceNorm allows to reach high quality predictions, however the tile-wise feature normalization is a source of artifacts both for U-Net (Tab. 1) and for UNETR (Tab. 2) as shown in Fig. 2. We observe that tile-wise input normalization (BatchNorm, tile-wise input norm) causes less severe tile mismatch than tile-wise network feature normalization (InstanceNorm, global input norm). Although it is impossible to reach completely seamless stitching with UNETR, experiments with InstanceNorm show strong tile mismatch, while global feature normalization and sufficient halo based on ERF size allow to reduce mismatch and make it almost unnoticeable." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 506 + ], + "type": "text", + "content": "Although InstanceNorm performs better than BatchNorm for boundaries, networks with tile-wise normalization struggle with classes which cover areas larger than the tile size, such as foreground and matrix in organoids dataset. Experiments with inference using different tile sizes Fig. 6 show that the prediction depends strongly on the tile size, reaching the best quality when the tiles become large enough for the feature distribution in each tile to reproduce the distribution in the whole image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 507, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 296, + 567 + ], + "type": "text", + "content": "Fig. 5 shows how artifacts get worse in the transfer setting: the model originally trained on the embryo dataset produces less artifacts than the model transferred from the plants dataset. We observe the same effect in transfer in both directions (STab. 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 576, + 295, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 601 + ], + "type": "text", + "content": "5.3. Mismatch in batch statistics causes train and eval mode disparity for BatchNorm" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Global feature normalization with BatchNorm allows to eliminate the tile mismatch but predictions for organoids boundary and other datasets are worse than with InstanceNorm. The reason for this is the disparity between the performance of the network in train and evaluation mode. Estimation of normalization statistics with a running average in BatchNorm relies on the batch statistics being stable enough to represent the whole dataset. If batches are too small, the running average can be very dif" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 91, + 553, + 448 + ], + "blocks": [ + { + "bbox": [ + 380, + 71, + 489, + 81 + ], + "lines": [ + { + "bbox": [ + 380, + 71, + 489, + 81 + ], + "spans": [ + { + "bbox": [ + 380, + 71, + 489, + 81 + ], + "type": "text", + "content": "Table 2. UNETR results table" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 91, + 553, + 448 + ], + "lines": [ + { + "bbox": [ + 316, + 91, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 91, + 553, + 448 + ], + "type": "table", + "html": "
Norm layerBatchNormInstanceNormBatchRenormIdentity
Input normglobalglobalglobalglobal
Feature normglobaltile-wiseglobalglobal
dice,eval modeorganoids0.930.890.950.91
0.390.600.720.60
plants0.030.100.870.28
foreground0.840.850.850.52
boundaries0.640.690.660.21
foreground0.670.560.670.00
boundaries0.490.490.530.00
dice,train modeorganoids0.890.890.950.91
0.600.600.720.60
boundaries0.100.100.870.28
matrix0.850.850.850.52
foreground0.690.690.660.21
boundaries0.560.560.670.00
boundaries0.490.490.530.00
t/e disparityorganoids0.36000
plants0.20000
embryo0.33000
tile mismatchorganoids0.000.100.000.00
plants0.000.040.000.00
embryo0.000.060.000.00
", + "image_path": "3a86a71116a9c1e17a10236de253d8a906a33506b922698cce53d3e6c91cb912.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 345, + 462, + 524, + 644 + ], + "blocks": [ + { + "bbox": [ + 345, + 462, + 524, + 644 + ], + "lines": [ + { + "bbox": [ + 345, + 462, + 524, + 644 + ], + "spans": [ + { + "bbox": [ + 345, + 462, + 524, + 644 + ], + "type": "image", + "image_path": "5ef99ac44483fe539c5193224e0f09484458018ac6b4bc432cfe96a4d66bc126.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 651, + 509, + 663 + ], + "lines": [ + { + "bbox": [ + 359, + 651, + 509, + 663 + ], + "spans": [ + { + "bbox": [ + 359, + 651, + 509, + 663 + ], + "type": "text", + "content": "Figure 7. Example of train/eval disparity" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "ferent from the batch-wise statistics used during training, leading to deteriorating performance. Doing inference with" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13115" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": "the network in train mode supports this hypothesis: the performance recovers, at the cost of tiling artifacts. In our experiments train and evaluation mode predictions can be quite different, with train/eval disparity reaching 0.84 for the organoids dataset. Qualitatively, the changes in the predictions occur in the areas of low prediction certainty. For example, in Fig. 7 the eval mode differs from the train mode in the areas where the background has a bright line artifact or in the noisier areas." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 275 + ], + "type": "text", + "content": "This effect is not the same as the usual difference between train and evaluation performance caused by overfitting. Here, the network's output differs even for the exact same images thus violating the usual assumption that the prediction quality is approximately the same in train and eval mode. This issue would not be remedied by using augmentations or more training data as it does not address the high variance of statistics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 283, + 294, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 294, + 308 + ], + "type": "text", + "content": "5.4. BatchRenorm corrects for train/eval disparity and provides seamless stitching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 294, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 294, + 372 + ], + "type": "text", + "content": "InstanceNorm avoids the statistics mismatch by always using the statistics of the input both during training and during inference at the cost of causing tiling artifacts. BatchNorm does not introduce the tiling artifacts but can lose performance due to train/eval disparity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 294, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 294, + 479 + ], + "type": "text", + "content": "A potential solution is to remove the normalization altogether. In our experiments not using normalization (Identity in Tab. 1 and Tab. 2) has led to worse performance for plants dataset and similar performance for organoids and embryo. It could be possible to improve the performance of networks with no normalization by finding more suitable training parameters and reducing the learning rate, however the necessity to adjust training parameters for each dataset makes the pipeline harder to use." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 294, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 294, + 588 + ], + "type": "text", + "content": "Another solution is to use the same global normalization statistics both in training and in inference. Directly using running average statistics during training causes network's weights to explode [11], which is why we propose to use Batch Renormalization method. We found that networks with BatchRenorm produce same or better predictions as InstanceNorm (Tab. 1, Tab. 2) including the transfer setup (STab. 1) while allowing for seamless stitching due to global feature normalization (last column of Fig. 2, Fig. 5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 597, + 125, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 125, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 125, + 609 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 294, + 713 + ], + "type": "text", + "content": "The severity of tiling artifacts depends on the data and on the tile size. For example, the tile mismatch can be negligible in case the images are very homogeneous in content and the tile size is big enough that each tile mostly reproduces the global statistics of the image, such as for plants dataset in our experiments. However, even in that case the inference performance depends on the tile size and can change unpredictably in the transfer setting, putting higher" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "requirements on the hardware for inference and making the pipeline harder to use, especially for non-expert users. Tracing low validation accuracy to the instability in batch statistics is not straightforward either, as it is normally expected that the performance on the validation dataset should be worse than on the training data. If the training loss is low but the accuracy on validation data does not improve, it is often attributed to the insufficient amount of ground truth, however, we show that it can also be caused by train/eval disparity preventing efficient utilization of the ground truth." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 192, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 299 + ], + "type": "text", + "content": "Both of the described effects are highly dependent on the particular data, input normalization, sampling strategy, postprocessing and other parameters of the image analysis pipeline. For any such pipeline, we introduce two simple metrics: tile mismatch and train/eval disparity to help determine if it has issues with normalization. Both metrics are easy to use for any sliding window inference pipeline, including other dense prediction tasks if the Dice score is replaced with another, task-appropriate metric." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 300, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 491 + ], + "type": "text", + "content": "We found that BatchRenorm, originally introduced for natural image classification, performs well for segmenting large biological images. The challenge with images larger than GPU memory is ensuring consistent normalization parameters across the entire image for seamless stitching, despite the inability to process the full image during training due to time constraints. BatchRenorm solves this by splitting training into two stages. First, the network is trained with BatchNorm. Then the normalization layer switches to using the accumulated running average statistics. In our experience it is very important that the training converges before switching and that the statistics change from tile-wise to global gradually to achieve the best final accuracy. While the BatchRenorm adds more hyperparameters, in our experiments the right moment to switch could easily be found from looking at the loss curves." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 689 + ], + "type": "text", + "content": "In this paper, we demonstrate that tiling artifacts in sliding window inference pipelines can be caused by the tilewise feature normalization in the network. We also find that switching from tile-wise normalization in training to global normalization in evaluation can lead to even worse performance, including hallucinations, due to small batch size. As these effects strongly depend on the data and tile size, we introduce tile mismatch and train/eval disparity metrics to easily check if a specific pipeline and dataset have these issues. We propose that Batch Renormalization technique will yield artifact-free predictions even with small batch size and demonstrate its good performance through extensive experiments on various microscopy datasets and network architectures." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Acknowledgment. We would like to thank Ricardo Sánchez Loayza and Qin Yu for helpful discussions. E.B." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13116" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 252 + ], + "type": "text", + "content": "was supported by the Joachim Herz Foundation through an Add-on Fellowship for Interdisciplinary Life Science. E.D'I. was supported by a fellowship from the EMBL Interdisciplinary (EI4POD) program under Marie Sk1odowska-Curie Actions COFUND (847543). J.M. acknowledges funding from the EMBL and the European Research Council (ERC 3DCellPhase-760067). This work was supported by the European Commission through the Horizon Europe program (IMAGINE project, grant agreement 101094250-IMAGINE). The work of A.A. was funded by the German Research Foundation through PA 4341/2-1. C.P. received support under Germany's Excellence Strategy - EXC 2067/1-390729940. We would like to thank the EMBL IT Services Department for providing computational infrastructure and support." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 262, + 115, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 115, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 115, + 274 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 282, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 61, + 282, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 282, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 282, + 295, + 346 + ], + "type": "text", + "content": "[1] Anwai Archit, Luca Freckmann, Sushmita Nair, Nabeel Khalid, Paul Hilt, Vikas Rajashekar, Marei Freitag, Carolin Teuber, Genevieve Buckley, Sebastian von Haaren, Sagnik Gupta, Andreas Dengel, Sheraz Ahmed, and Constantin Pape. Segment anything for microscopy. Nature Methods, 2025. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 348, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 348, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 61, + 348, + 295, + 435 + ], + "type": "text", + "content": "[2] Vladyslav Bondarenko, Mikhail Nikolaev, Dimitri Kromm, Roman Belousov, Adrian Wolny, Marloes Blotenburg, Peter Zeller, Saba Rezakhani, Johannes Hugger, Virginie Uhlmann, Lars Hufnagel, Anna Kreshuk, Jan Ellenberg, Alexander van Oudenaarden, Anna Erzberger, Matthias P Lutolf, and Takashi Hiiragi. Embryo-uterine interaction coordinates mouse embryogenesis during implantation. The EMBO Journal, 42(17), 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 437, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 295, + 491 + ], + "type": "text", + "content": "[3] Calimanut-Ionut Cira, Miguel-Ángel Manso-Callejo, Naoto Yokoya, Tudor Sălăgean, and Ana-Cornelia Badea. Impact of tile size and tile overlap on the prediction performance of convolutional neural networks trained for road classification. Remote Sensing, 16(15):2818, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 492, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 492, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 492, + 295, + 557 + ], + "type": "text", + "content": "[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 559, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 295, + 602 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 604, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 295, + 713 + ], + "type": "text", + "content": "[6] Sven Dorkenwald, Arie Matslieh, Amy R Sterling, Philipp Schlegel, Szi-Chieh Yu, Claire E McKellar, Albert Lin, Marta Costa, Katharina Eichler, Yijie Yin, Will Silversmith, Casey Schneider-Mizell, Chris S Jordan, Derrick Brittain, Akhilesh Halageri, Kai Kuehner, Oluwaseun Ogedengbe, Ryan Morey, Jay Gager, Krzysztof Kruk, Eric Perlman, Runzhe Yang, David Deutsch, Doug Bland, Marissa Sorek, Ran Lu, Thomas Macrina, Kisuk Lee, J Alexander Bae, Shang Mu, Barak Nehoran, Eric Mitchell, Sergiy Popovych, Jingpeng Wu, Zhen Jia, Manuel A Castro, Nico Kemnitz, Dodam" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 127 + ], + "type": "text", + "content": "Ih, Alexander Shakeel Bates, Nils Eckstein, Jan Funke, Forrest Collman, Davi D Bock, Gregory S X E Jefferis, H Sebastian Seung, Mala Murthy, and FlyWire Consortium. Neuronal wiring diagram of an adult brain. Nature, 634(8032): 124-138, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 194 + ], + "type": "text", + "content": "[7] Edoardo D'Imprima, Marta Garcia Montero, Sylwia Gawrzak, Paolo Ronchi, Ievgeniia Zagoriy, Yannick Schwab, Martin Jechlinger, and Julia Mahamid. Light and electron microscopy continuum-resolution imaging of 3d cell cultures. Developmental Cell, 58(7):616-632.e6, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 197, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 197, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 320, + 197, + 553, + 272 + ], + "type": "text", + "content": "[8] T. Falk, D. Mai, R. Bensch, Ö. Ciçek, A. Abdulkadir, Y. Marrakchi, A. Böhm, J. Deubner, Z. Jäckel, K. Seiwald, A. Dovzhenko, O. Tietz, C. Dal Bosco, S. Walsh, D. Saltukoglu, T. L. Tay, M. Prinz, K. Palme, M. Simons, I. Diester, T. Brox, and O. Ronneberger. U-net - deep learning for cell counting, detection, and morphometry. Nature Methods, 16: 67-70, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 274, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 274, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 321, + 274, + 553, + 318 + ], + "type": "text", + "content": "[9] Bohao Huang, Daniel Reichman, Leslie M Collins, Kyle Bradbury, and Jordan M Malof. Tiling and stitching segmentation output for remote sensing: Basic challenges and recommendations. arXiv:1805.12219, 2018. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 373 + ], + "type": "text", + "content": "[10] Lei Huang, Jie Qin, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Normalization techniques in training dnns: Methodology, analysis and application. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10173-10196, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "type": "text", + "content": "[11] Sergey Ioffe. Batch renormalization: Towards reducing minibatch dependence in batch-normalized models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 4, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "text", + "content": "[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Proceedings of the 32nd International Conference on Machine Learning, pages 448-456, Lille, France, 2015. PMLR. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 477, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 477, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 477, + 553, + 510 + ], + "type": "text", + "content": "[13] Fabian Isensee and et al. nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 18(2):203-211, 2020. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 512, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 577 + ], + "type": "text", + "content": "[14] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 579, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 623 + ], + "type": "text", + "content": "[15] Vincent Loos, Rohit Pardasani, and Navchetan Awasthi. Demystifying the effect of receptive field size in u-net models for medical image segmentation. Journal of Medical Imaging, 2024. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 668 + ], + "type": "text", + "content": "[16] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[17] Michael Majurski and Peter Bajcsy. Exact tile-based segmentation inference for images larger thangpu memory. Journal of Research of the National Institute of Standards and Technology, 126, 2021. 1, 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13117" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 619 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[18] G. Anthony Reina, Ravi Panchumarthy, Siddhesh Pravin Thakur, Alexei Bastidas, and Spyridon Bakas. Systematic evaluation of image tiling adverse effects on deep learning semantic segmentation. Frontiers in Neuroscience, 14, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[19] O. Ronneberger, P.Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234-241. Springer, 2015. (available on arXiv:1505.04597 [cs.CV]). 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 261 + ], + "type": "text", + "content": "[20] Xiongtao Ruan, Matthew Mueller, Gaoxiang Liu, Frederik Gorlitz, Tian-Ming Fu, Daniel E Milkie, Joshua L Lillvis, Alexander Kuhn, Johnny Gan Chong, Jason Li Hong, Chu Yi Aaron Herr, Wilmene Hercule, Marc Nienhaus, Alison N Killilea, Eric Betzig, and Srigokul Upadhyayula. Image processing tools for petabyte-scale light sheet microscopy data. Nat. Methods, 21(12):2342-2352, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 262, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 294, + 328 + ], + "type": "text", + "content": "[21] Josef Lorenz Rumberger, Xiaoyan Yu, Peter Hirsch, Melanie Dohmen, Vanessa Emanuela Guarino, Ashkan Mokarian, Lisa Mais, Jan Funke, and Dagmar Kainmueller. How shift equivariance impacts metric learning for instance segmentation. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), page 7108-7116. IEEE, 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 330, + 294, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 330, + 294, + 362 + ], + "spans": [ + { + "bbox": [ + 56, + 330, + 294, + 362 + ], + "type": "text", + "content": "[22] Carsen Stringer, Tim Wang, Michalis Michaelos, and Marius Pachitariu. Cellpose: a generalist algorithm for cellular segmentation. Nature Methods, 18(1):100-106, 2020. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 364, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 396 + ], + "type": "text", + "content": "[23] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Instance normalization: The missing ingredient for fast stylization, 2016. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 397, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 397, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 56, + 397, + 294, + 451 + ], + "type": "text", + "content": "[24] Özgün undefinedigkeit, Ahmed Abdulkadir, Soeren S. Lienkamp, Thomas Brox, and Olaf Ronneberger. 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation, page 424-432. Springer International Publishing, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 453, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 518 + ], + "type": "text", + "content": "[25] Athul Vijayan, Tejasvinee Atul Mody, Qin Yu, Adrian Wolny, Lorenzo Cerrone, Soeren Strauss, Miltos Tsiantis, Richard S. Smith, Fred A. Hamprecht, Anna Kreshuk, and Kay Schneitz. A deep learning-based toolkit for 3d nuclei segmentation and quantitative analysis in cellular and tissue context. Development, 151(14), 2024. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 574 + ], + "type": "text", + "content": "[26] Martin Weigert, Uwe Schmidt, Robert Haase, Ko Sugawara, and Gene Myers. Star-convex polyhedra for 3d object detection and segmentation in microscopy. In 2020 IEEE Winter Conference on Applications of Computer Vision (WACV), page 3655–3662. IEEE, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 575, + 294, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 575, + 294, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 575, + 294, + 619 + ], + "type": "text", + "content": "[27] Zhuliang Yao, Yue Cao, Yutong Lin, Ze Liu, Zheng Zhang, and Han Hu. Leveraging batch normalization for vision transformers. In 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). IEEE, 2021. 6" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13118" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Time-Aware Auto White Balance in Mobile Photography/59296f3d-511f-4955-b6aa-60113b360f05_content_list.json b/2025/Time-Aware Auto White Balance in Mobile Photography/59296f3d-511f-4955-b6aa-60113b360f05_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f643c2b835f35178a4bc0af08c57719b9b7ad5a7 --- /dev/null +++ b/2025/Time-Aware Auto White Balance in Mobile Photography/59296f3d-511f-4955-b6aa-60113b360f05_content_list.json @@ -0,0 +1,1584 @@ +[ + { + "type": "text", + "text": "Time-Aware Auto White Balance in Mobile Photography", + "text_level": 1, + "bbox": [ + 212, + 130, + 782, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mahmoud Affifi*", + "bbox": [ + 264, + 172, + 395, + 189 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Luxi Zhao*", + "bbox": [ + 426, + 172, + 519, + 189 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abhijith Punnappurath", + "bbox": [ + 550, + 172, + 732, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohammed A. Abdelsalam", + "bbox": [ + 238, + 191, + 460, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ran Zhang Michael S. Brown", + "bbox": [ + 491, + 191, + 756, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "AI Center-Toronto, Samsung Electronics", + "bbox": [ + 334, + 209, + 661, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{m.afifil, lucy.zhao, abhijith.p, m.abdelsalam, ran.zhang, michael.b1}@samsung.com", + "bbox": [ + 135, + 229, + 859, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7fc4c1df2b7ca5e35050456c912a700f35890770c4ff1ee1c02be792a3fb1784.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 256, + 295, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/20de2c620390972ff2c6eaea908e4ac94af663fb2c890494f3aef6ca9751b23c.jpg", + "image_caption": [ + "Figure 1. The time of day influences scene illumination, making it a valuable cue for improving illuminant estimation. Shown are white-balanced results (gamma-corrected for visualization) using our method with (1) colors only, (2) colors plus contextual (timestamp and geolocation) and capture data, and (3) ground truth (from a color chart). Angular errors show improvements when time-capture information is used." + ], + "image_footnote": [], + "bbox": [ + 114, + 462, + 145, + 486 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/34933caa09610af20c0f945853961c95383be92bbb7d66865e91ff02329cd952.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 462, + 194, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/494610d5708b103912b8b3e91058b1c46fa8e0035839c2e65549f75d07e62175.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 462, + 225, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0c0110915450ac8e4be1ea77794ccf5a754fce231ba0fdd1491162392c00be3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 257, + 496, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/804e0117b967e293d9741042ea051a11d7801395b6e4320f7e6ecfd0d683aa6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 315, + 462, + 346, + 486 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f4234c86efb8c3db926f7d3948c6430d4f02764cf7391275e0dbe93abd58933c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 462, + 426, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/890626b0b999577fb0a45a5a233fea3a624a3e705b69a7695a4c09f0cee8444f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 257, + 699, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ab03fe008aea03315fc8bdcb39fa05d8d813d3261cbf357c35aab9c1ba117e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 462, + 547, + 486 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c8e35235cab070dd208c92e266b5e941db2ca4d6881db3c51cb5e93dabefb27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 565, + 462, + 627, + 487 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b75febfbbb93d44c1e182766884ab821dc57151fc76fc0c04a60cc4f9a1100d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 258, + 901, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d2bbeef182343bfae2d9943f268bc9fc8e9c603a81f1c3e213ebe50da9ee833a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 462, + 746, + 486 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e1502e8ef95c143df89633f3b5cf4f114bf3ff14e1af5a61e699f95bd8d5952.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 462, + 828, + 487 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 555, + 326, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cameras rely on auto white balance (AWB) to correct undesirable color casts caused by scene illumination and the camera's spectral sensitivity. This is typically achieved using an illuminant estimator that determines the global color cast solely from the color information in the camera's raw sensor image. Mobile devices provide valuable additional metadata—such as capture timestamp and geolocation—that offers strong contextual clues to help narrow down the possible illumination solutions. This paper proposes a lightweight illuminant estimation method that incorporates such contextual metadata, along with additional capture information and image colors, into a lightweight model ( $\\sim 5K$ parameters), achieving promising results, matching or surpassing larger models. To validate our method, we introduce a dataset of 3,224 smartphone images with contextual metadata collected at various times of day and under diverse lighting conditions. The dataset includes ground-truth illuminant colors, determined using a color chart, and user-preferred illuminants validated through a user study, providing a comprehensive benchmark for AWB evaluation.", + "bbox": [ + 88, + 577, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction and related work", + "text_level": 1, + "bbox": [ + 513, + 602, + 792, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Color constancy refers to the ability of the human visual system to maintain stable object colors despite variations in lighting conditions by leveraging contextual cues within the scene [27, 33, 51]. Cameras approximate this effect using auto white balance (AWB) correction, which aims to partially neutralize color casts introduced by scene illumination and the camera's spectral sensitivity [3]. AWB first estimates the illumination color as an RGB vector in the camera's raw color space. The raw image is then corrected by scaling its color channels according to the estimated illumination, typically under the assumption of a single global light source [8, 31].", + "bbox": [ + 509, + 627, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conventional illuminant estimation methods primarily rely on image colors, either by directly processing the raw image (e.g., [13, 32, 41, 49, 53, 58, 62, 66]) or by analyzing color histograms (e.g., [2, 6, 7, 10]). These methods can be broadly categorized into two groups: (1) classical statistical-based methods (e.g., [15, 26, 32, 56, 57, 62]),", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 109, + 886, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5038", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f233ecfc8dca891a11308e394d6b4a4823f290d025912e67d087ddffdabfae92.jpg", + "image_caption": [ + "Figure 2. The time of day at which an image is captured provides valuable information about the possible range of illuminant colors in outdoor scenes. The figure presents the correlated color temperature (CCT) of illuminant colors in our dataset (Sec. 3) for images captured at various times throughout the day and night. As shown, excluding images taken under artificial light, those captured at noon, for example, exhibit a different range of illuminant CCTs compared to images captured during sunset or sunrise." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 473, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which estimate the illuminant color based on image statistics, and (2) learning-based methods (e.g., [6, 10, 30, 42, 59]), which map image colors to their corresponding scene illuminant through data-driven models.", + "bbox": [ + 89, + 484, + 482, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While image colors are a key cue for estimating the scene's illuminant, mobile devices provide an opportunity to integrate additional contextual information. For instance, the device's location, along with the date and time, offer valuable cues about outdoor lighting conditions (e.g., sunrise, noon, sunset), thereby improving illuminant estimation for outdoor scenes (see Fig. 1).", + "bbox": [ + 89, + 547, + 482, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Intuitively, knowing the time of day when an outdoor scene is captured can help estimate the lighting conditions. Figure 2 further illustrates that the time of day, derived from contextual metadata (i.e., timestamp and geolocation) readily available on mobile devices, provides valuable insights into the likely range of illuminant correlated color temperature (CCT) in outdoor scenes. This information serves to narrow the range of possible illuminant colors. For instance, images taken at noon exhibit a different CCT range than those captured at sunrise or sunset. When combined with additional capture information to distinguish between environments (e.g., indoors vs. outdoors), such metadata can complement image colors to improve illuminant estimation accuracy.", + "bbox": [ + 89, + 657, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite its potential, only a few attempts have explored leveraging additional information available to the camera's", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image signal processor (ISP) for illuminant estimation, such as metadata-based model control [11] or data augmentation to enhance generalization [6]. However, to our knowledge, no previous work has investigated using contextual information—specifically, mobile device timestamps and geolocation—to refine illuminant estimation.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce a method that leverages contextual metadata from mobile phones, along with additional capture information available in camera ISPs, to train a lightweight illuminant estimator model with $\\sim 5\\mathrm{K}$ parameters. Our model delivers promising results, matching or surpassing that of larger models, while maintaining efficiency. Our model runs on a typical flagship mobile digital signal processor (DSP) and CPU in $0.25\\mathrm{ms}$ and $0.80\\mathrm{ms}$ , respectively. This compact and efficient design is especially advantageous for mobile devices, where minimizing power consumption and memory usage is critical [1, 7, 11].", + "bbox": [ + 511, + 202, + 906, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing white-balance datasets (e.g., [19, 23, 28, 43]) lack the contextual information needed to validate our method. The absence of contextual information is because most available datasets (e.g., [19, 23]) used DSLR cameras, which typically lack built-in GPS functionality or accurate timestamps. To address this, we captured a new dataset of 3,224 images using a consumer smartphone camera, accompanied by contextual and capture information. The ground truth for the dataset was established using two approaches: (1) the conventional approach, selecting gray patches from a calibration color chart, and (2) a manually selected user-preference white-balance target, which enhances the image's aesthetic appeal and accounts for incomplete chromatic adaptation [60]. The user-preference white-balance \"ground truth\" was validated through a user study. The dataset spans a variety of lighting conditions, including non-standard artificial illuminants [61], and covers different times of day—sunrise, noon, sunset, and night.", + "bbox": [ + 511, + 388, + 906, + 661 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contribution: We propose an AWB method that utilizes smartphone contextual metadata (i.e., timestamp and geolocation) alongside capture information to enhance illuminant estimation. We demonstrate that integrating this additional data with conventional color information into a lightweight neural model leads to significant improvements in accuracy. Additionally, we introduce a large-scale dataset captured with a consumer smartphone at various times of day, with ground truth derived from both a calibration color chart and user-preference-based white balance. Benchmarking results show that our method achieves state-of-the-art results in most standard metrics (i.e., angular error statistics [19]), while maintaining computational efficiency on mobile devices.", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5039", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1591a830cb7446514d9e26f8a58a76bccbdae499c6a48f3d34c47ded2a49bdbd.jpg", + "image_caption": [ + "Figure 3. Our method includes a lightweight model consisting of a convolutional network that processes the histogram feature (derived from raw image colors and edge histograms concatenated with the $u / v$ coordinates) to produce a latent feature $\\mathbf{v}_h\\in \\mathbb{R}^{16}$ . This is then concatenated with the latent feature of the processed time-capture feature, $\\mathbf{v}_t\\in \\mathbb{R}^{16}$ . The combined feature, $\\mathbf{v}\\in \\mathbb{R}^{32}$ , is passed through a lightweight MLP to produce the chromaticity of the scene illuminant, $\\ell_c\\in \\mathbb{R}^2$ , which is finally converted into normalized RGB illuminant color." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 480, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Method", + "text_level": 1, + "bbox": [ + 89, + 417, + 181, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of our method is shown in Fig. 3. Our method employs a learnable model that processes two distinct inputs: (1) the time-capture feature (Sec. 2.1), which combines the contextual and capture information available on mobile devices and accessible by their camera ISPs, and (2) the histogram feature (Sec. 2.2), which represents the image's R/G and B/G chromaticity values. The time-capture feature is first processed to project it into a latent space, producing the time-capture latent feature vector, $\\mathbf{v}_t \\in \\mathbb{R}^{16}$ . The model processes the histogram feature to produce the histogram's latent feature vector, $\\mathbf{v}_h \\in \\mathbb{R}^{16}$ . Both the histogram and time-capture latent feature vectors are concatenated and processed by a set of learnable layers to output an R/G, B/G chromaticity vector, $\\ell_c \\in \\mathbb{R}^2$ , representing the scene illuminant. This 2D vector is then converted into the illuminant RGB color used to perform white balancing.", + "bbox": [ + 89, + 444, + 483, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Time-capture feature", + "text_level": 1, + "bbox": [ + 89, + 696, + 290, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model incorporates contextual metadata, available on mobile devices, as one of its input features. Specifically, we use the geolocation and timestamp of image capture to compute the \"probability\" of the time of day (e.g., sunset, noon, etc.). Our approach converts conventional clock time (hour::minutes) to its corresponding time of day condition based on the date and geolocation of the scene. This approach allows the model to generalize across different time zones and prevents it from being influenced by the capturing location. The time probability vector represents the likelihood that the captured image corresponds to one of the solar event times (i.e., dawn, sunrise, noon, sunset, dusk, or mid", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "night) and is computed as follows:", + "bbox": [ + 513, + 90, + 743, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {g} = 1 - \\frac {\\left| t _ {c} - t _ {g} \\right|}{t _ {s}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 118, + 903, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t_c$ is the capture time in seconds (adjusted to the local time zone based on geolocation information), and $t_g$ is the local time in seconds of the solar event, $g \\in \\{\\text{dawn,sunrise,noon,sunset,dusk,midnight}\\}$ , computed using geolocation-based standard algorithms [52, 63, 64]. The scalar $t_s$ represents the total number of seconds in a day (i.e., 86,400).", + "bbox": [ + 511, + 156, + 906, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We pre-process the probability of each solar event, $p_{g}$ , in our time probability vector by computing the square root to enhance feature representation—compressing high probabilities and amplifying lower ones, which is intended to help create a more balanced and smooth distribution for the model to leverage. We then augment this time probability vector with a one-hot vector, $\\mathbf{b}$ , that indicates whether the capture time $t_{c}$ occurs before each solar event. This distinction helps the model account for expected variations in CCT, as illuminant colors can differ before and after certain solar events, such as sunset and sunrise. The value of this one-hot vector for a given solar event $g$ is computed as:", + "bbox": [ + 511, + 263, + 906, + 444 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {b} _ {g} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} t _ {c} \\leq t _ {g} \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 455, + 903, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{b}_g$ corresponds to the entry in the one-hot vector for the solar event, $g$ . Both the time probability vector and the one-hot vector together form our time feature, $\\mathbf{p} \\in \\mathbb{R}^{12}$ .", + "bbox": [ + 511, + 507, + 903, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enrich our feature set with additional capture information available from the camera ISP and help distinguish the capturing environment (e.g., indoor vs. outdoor), we include the following features in our final time-capture feature, c:", + "bbox": [ + 511, + 553, + 905, + 626 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- ISO (i): The sensitivity of the camera's image sensor to light, where lower values indicate good lighting conditions (e.g., bright scenes) and higher values suggest low-light environments, such as poorly lit indoor scenes.", + "- Shutter speed $(s)$ : The amount of time the camera's shutter remains open, allowing light to hit the image sensor. It provides an indication of lighting conditions, alongside the ISO value, $i$ .", + "- Flash status $(f)$ : A binary value indicating whether flash light was used during capturing.", + "- Noise information (optional): Since image denoising is typically applied before or in parallel with illuminant estimation in camera ISPs [22, 34, 45], we also explore the optional use of explicit noise information from the captured scene. More noise typically indicates low-light conditions, which can provide clues about the lighting color range of the scene. Unfortunately, while noise information is accessible within camera ISPs, obtaining accurate" + ], + "bbox": [ + 513, + 628, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5040", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "noise information for public use is challenging, as the noise profiles in DNG files are not always reliable [68]. To address this, we simulate the noise information using two approaches: noise statistics (stats) and/or signal-to-noise ratio (SNR) stats, which are described below.", + "bbox": [ + 102, + 90, + 482, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Noise stats (n): This represents the noise statistics in the captured raw image. We simulate this by denoising each raw image using Adobe Lighthroom and computing the noise stats as the mean and standard deviation of each color channel of the absolute difference between the denoised and noisy raw images. This approach provides a simplified method for estimating noise stats, as the denoised images are typically available within the camera ISPs, but difficult to extract from the DNG files.", + "bbox": [ + 89, + 172, + 482, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SNR stats (r): This alternative approach measures the noise information in the captured image without the need of a denoised reference. The SNR is computed by applying a $15 \\times 15$ sliding window over the raw image and calculated as: $10\\log_{10}\\left(\\frac{\\mu}{\\sigma + \\epsilon}\\right)$ , where $\\mu$ represents the mean RGB of the $15 \\times 15$ patch, $\\sigma$ is the standard deviation, and $\\epsilon$ is a small value added for numerical stability.", + "bbox": [ + 89, + 316, + 483, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that while these two approaches—namely, computing noise stats and SNR stats—yield satisfactory results, there is a wide body of research on noise estimation (e.g., [47, 54, 55]), which is beyond the scope of this paper.", + "bbox": [ + 89, + 429, + 482, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our complete time-capture feature is the combination of these inputs and can be expressed as follows:", + "bbox": [ + 89, + 489, + 483, + 520 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {c} = \\left[ \\mathbf {p} ^ {T}, \\log (i), \\log (s), f, \\mathbf {w} ^ {T} \\right] ^ {T}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 532, + 482, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{w}$ represents optional noise-related features, which can include either noise stats $\\mathbf{n}$ , SNR stats $\\mathbf{r}$ , or both. Alternatively, $\\mathbf{w}$ can be omitted if noise information is not considered.", + "bbox": [ + 89, + 566, + 482, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our time-capture feature is first normalized using min-max normalization (with the min and max values computed from the training data), and then processed through a learnable function, $f_{t}$ , as follows:", + "bbox": [ + 89, + 627, + 483, + 686 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {t} = f _ {t} (\\mathbf {c}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 245, + 705, + 482, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{t}$ is a learnable linear layer that transforms the time-capture feature, $\\mathbf{c}$ , into its latent representation, $\\mathbf{v}_t\\in \\mathbb{R}^{16}$ (see Fig. 3).", + "bbox": [ + 89, + 729, + 482, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2. Histogram feature", + "text_level": 1, + "bbox": [ + 89, + 787, + 267, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the time-capture feature, c, we provide our neural model with a histogram feature, $\\mathbf{H}$ which represents the colors of the raw image. Inspired by prior work [6, 10, 11], we use a 2D histogram to represent the R/G and B/G chromaticities of the input raw image, $\\mathbf{I} \\in \\mathbb{R}^{K \\times 3}$ , where $K$ denotes the total number of pixels in the raw im", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "age. Specifically, we compute the 2D chromaticity histogram, $\\mathbf{H}_c\\in \\mathbb{R}^{h\\times h}$ , as follows:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {c} ^ {(m, n)} = \\sum_ {k} \\| \\mathbf {I} ^ {(k)} \\| _ {2} \\cdot \\delta_ {m, n} ^ {(k)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 133, + 903, + 162 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\delta_ {m, n} ^ {(k)} = \\left[ u _ {m} \\leq r g ^ {(k)} < u _ {m + 1} \\right] \\wedge \\left[ v _ {n} \\leq b g ^ {(k)} < v _ {n + 1} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 191, + 903, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $rg^{(k)}$ and $bg^{(k)}$ are the R/G and B/G chromaticity values of pixel $k$ in I, and $\\| \\mathbf{I}^{(k)} \\|_2$ represents the intensity of pixel $k$ (i.e., Euclidean norm of the pixel's RGB values). The notation $\\wedge$ denotes the logical AND operator, and $[\\cdot]$ is the Iverson bracket, which evaluates to 1 when the condition is true and 0 otherwise. The histogram bins are defined by the edges $\\{u_m\\}$ and $\\{v_n\\}$ , where $u_m$ and $v_n$ are the lower edges of the bins, and $u_{m+1}$ and $v_{n+1}$ are the corresponding upper edges. The histogram accumulates the brightness values of pixels whose chromaticity values fall within the range $[u_m, u_{m+1})$ along the R/G axis (i.e., the $u$ -axis) and $[v_n, v_{n+1})$ along the B/G axis (i.e., the $v$ -axis). The number of histogram bins along each axis is denoted as $h$ . Following [10], we compute the square root of the histogram to enhance the utility of the histogram feature [9].", + "bbox": [ + 511, + 223, + 906, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that this histogram differs from the $\\log -uv$ histogram used in prior work [6, 10, 11], which operates in the logarithmic space of G/R and G/B [25]. We found that the R/G, B/G chromaticity histogram performs better, as it aligns with our model's output space. See the supplemental material for an ablation study.", + "bbox": [ + 511, + 450, + 905, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the chromaticity histogram, $\\mathbf{H}_c$ , of the image colors, and building on prior work [6, 11], we augment it with the square-rooted chromaticity histogram of the image's edges, $\\mathbf{H}_e$ , where the image edges are computed as follows:", + "bbox": [ + 511, + 541, + 905, + 614 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} ^ {(x, y)} = \\frac {1}{8} \\sum_ {\\Delta x, \\Delta y} \\left| \\mathbf {I} ^ {\\prime (x, y)} - \\mathbf {I} ^ {\\prime (x + \\Delta x, y + \\Delta y)} \\right|, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 640, + 903, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{E}$ represents the image edges, $\\mathbf{I}'$ refers to the raw image in 3D tensor form (height, width, channels), and $\\Delta x, \\Delta y \\in \\{-1,0,1\\}$ with $(\\Delta x, \\Delta y) \\neq (0,0)$ . The edge histogram, $\\mathbf{H}_e$ , is computed from $\\mathbf{E}$ using Eqs. 5 and 6.", + "bbox": [ + 511, + 689, + 903, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our histogram feature is constructed by concatenating these two histograms, $\\mathbf{H}_c$ and $\\mathbf{H}_e$ . Since this histogram feature is first processed by convolutional (conv) layers, as shown in Fig. 3, we follow [6] by appending additional channels that encode the positional information of the $u / v$ coordinates in histogram space, which helps capture spatial relationships within the histogram feature. Consequently, our final histogram feature, $\\mathbf{H} \\in \\mathbb{R}^{h \\times h \\times 4}$ , consists of 2 channels representing the chromaticity of the image and its edges, along with the additional $u / v$ coordinate channels.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5041", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This feature is processed through a series of conv layers with ELU activation [21]. The resulting latent representation undergoes adaptive average pooling before passing through a linear layer to produce the histogram's latent feature vector, $\\mathbf{v}_h\\in \\mathbb{R}^{16}$ , as follows:", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {h} = f _ {h} (\\mathbf {H}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 181, + 482, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{h}$ denotes the sub-model (i.e., conv layers, ELU activation, pooling, and linear layer) that maps the histogram feature into its latent space, as shown in Fig. 3.", + "bbox": [ + 89, + 205, + 483, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "2.3. Illuminant estimation", + "text_level": 1, + "bbox": [ + 89, + 261, + 294, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both feature vectors, $\\mathbf{v}_t$ and $\\mathbf{v}_h$ , are concatenated to produce the latent vector $\\mathbf{v}$ and processed by an illuminant estimation sub-model as follows:", + "bbox": [ + 89, + 282, + 483, + 329 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} = \\left[ \\mathbf {v} _ {t}; \\mathbf {v} _ {h} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 344, + 482, + 362 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {c} = f _ {\\ell} (\\mathbf {v}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 245, + 383, + 482, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{\\ell}$ consists of a set of linear layers, with batch normalization applied to the first layer, followed by activation functions—except for the final layer, which outputs $\\ell_{c}$ , the chromaticity vector of the scene illuminant. This vector is then transformed into an unnormalized RGB illuminant color by mapping $[\\mathrm{R / G},\\mathrm{B / G}]^T\\rightarrow [\\mathrm{R / G},1,\\mathrm{B / G}]^T$ , followed by normalization via division by its L2 norm to produce the final illuminant color. We optimize $f_{t}$ , $f_{h}$ , and $f_{\\ell}$ to minimize the angular error [36] between the predicted RGB illuminant color and the ground-truth RGB illuminant color.", + "bbox": [ + 89, + 407, + 483, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Dataset", + "text_level": 1, + "bbox": [ + 89, + 573, + 181, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train and validate our method, we require a dataset that includes contextual information (i.e., timestamp and geolocation) h h image. Existing datasets (e.g., [17, 19, 23, 28, 43, 44]) lack this essential information, motivating us to collect a new dataset using a smartphone camera that provides contextual metadata for each image. Specifically, we captured 3,224 linear raw images with the Samsung S24 Ultra's main camera, covering a wide range of scenes both indoors and outdoors, at various times of day (e.g., sunset, sunrise, noon, night). Our dataset includes images captured under various light sources (e.g., sunlight, incandescent, LED), as well as non-standard illuminant colors (e.g., colored LED light), which are not present in existing datasets [61]; see Fig. 4. Additionally, our dataset captures scenes under different weather conditions (sunny, cloudy, rain, snow, etc.). Example scenes can be found in the supplemental material.", + "bbox": [ + 89, + 598, + 482, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We follow prior work [7, 43] in collecting ground-truth illuminant colors for each scene. Specifically, for each scene, we first capture an image with a calibration color chart, which is used to extract the illuminant color from the", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/37f97bafb85ba114e5ec6509bbbf0db7564b7396fd3d3680416eea20a931d984.jpg", + "image_caption": [ + "Figure 4. A plot of the $rg$ chromaticity distribution of ground-truth illuminant colors for neutral white balance in our dataset. Example images in raw and sRGB spaces are shown (raw images are gamma-corrected for visualization). See supplemental material the $rg$ chromaticity distribution of user-preference illuminants." + ], + "image_footnote": [], + "bbox": [ + 519, + 94, + 901, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "gray patches. Next, we capture an image of the same scene without the color chart (see Fig. 5), resulting in $\\sim 6\\mathrm{K}$ images. After obtaining the ground-truth illuminant color, we discard all color chart images, leaving us with 3,224 images in our dataset. This approach allows us to test with natural images that mimic real-world scenarios, without the need to mask out the color chart patch.", + "bbox": [ + 511, + 425, + 905, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our dataset includes a wide variety of lighting conditions, such as sunset/sunrise, night scenes, and artificial light, we generate a \"user-preference\" ground truth in addition to the neutral ground truth obtained from the color chart for each scene in our dataset. This is to account for human incomplete chromatic adaptation in such scenes [50, 60] and user preference [20, 24]. Specifically, an expert photographer was asked to assign a ground truth illuminant to each scene to make it appear more natural, reflect real-world observations, and enhance the aesthetics of the image. Notably, the same person who captured the scene also performed the annotation, ensuring that the user-preference selection was based on real-world observations of how the scene should appear.", + "bbox": [ + 511, + 534, + 906, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We validated the user-preference ground truth through a study with 20 participants, who selected the most natural image from pairs of white-balanced images corresponding to the user-preference ground truth and the neutral ground truth (obtained from the color chart). The user-preference ground truth was selected in $71.95\\%$ of the trials, confirming its preference. See the supplemental material for additional details. In our experiments, both ground-truth types—neutral and user-preference—were used for training and evaluation (Sec. 4).", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5042", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5cbfde5f3f2a93b14d0a501e0f559052dccf4bed8a822c976be0457aad762bfe.jpg", + "image_caption": [ + "Figure 5. For each scene in our dataset, an image with a color chart placed in the scene was captured. The gray patches on the color chart were used to measure the ground-truth illuminant. These color-chart images were discarded, and only the images of the scenes without the color chart were used for the training, validation, and testing sets. For each example, we show both raw and sRGB images, with the raw images gamma-corrected to enhance visualization." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we created a mask for regions illuminated by non-dominant illuminants in each scene. These masks ensure that all scenes have only one dominant illuminant, matching the color of the neutral ground truth, without confounding effects from other illuminants. We applied these masks to the training images when training our method and others. To preserve privacy, sensitive information, such as car plates and faces, has been blurred.", + "bbox": [ + 88, + 277, + 480, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to raw images, both ground-truth illuminant types, and contextual and capture information, we provide additional valuable auxiliary data to broaden the impact of our dataset. This includes locally tone-mapped sRGB images rendered by an expert, that can serve as ground-truth for applications beyond white-balance correction, such as neural ISPs [35, 38, 39, 65]. More information about the dataset can be found in the supplemental material. We organize our dataset into 2,619 raw images for training, 205 raw images for validation, and 400 raw images for testing.", + "bbox": [ + 88, + 401, + 482, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 571, + 223, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details: We train the model using the Adam optimizer [40], for 400 epochs with betas set to (0.9, 0.999) and a weight decay factor of $10^{-9}$ . A warm-up strategy is applied to gradually increase the learning rate from $10^{-6}$ to $10^{-3}$ over the first 5 epochs. After this, we use a cosine annealing schedule [48]. Following [6], we employ a batch-size increment strategy during training, starting with a batch size of 8 and doubling it every 100 epochs. Following prior work [6, 7], we use images of size $384 \\times 256$ pixels in all experiments, which is a reasonable size for camera pipelines to reduce computational overhead. For our method, we use histograms with $h = 48$ bins. The histogram boundaries are determined by computing the 10th and 95th percentiles of the chromaticity values along each chromaticity axis from the training set. We report the results of our method both with and without the inclusion of noise stats, n, and SNR stats, r.", + "bbox": [ + 91, + 604, + 482, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results: We report the results of our method alongside several others benchmarked on the proposed dataset.", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method is compared with various statistical-based approaches [12, 14, 15, 18, 19, 26, 32, 56, 57, 62], camera-specific learning-based methods [4, 5, 11, 16, 29, 30, 37, 41, 42, 53, 59, 66, 67], and camera-independent techniques [2, 6, 13]. For the camera-specific learning-based approaches, including our method, all models were trained on our training set.", + "bbox": [ + 511, + 277, + 903, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the camera-independent methods [2, 6], we present results from models trained on the NUS [19] and Cube++ [23] datasets. Additionally, we report results from finetuned versions of these models [2, 6], incorporating our proposed dataset. We also evaluate camera-specific (CS) models of these methods [2, 6], which were trained exclusively on our training data, without any additional datasets. For further details, refer to the supplemental material.", + "bbox": [ + 511, + 385, + 905, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows the results on the testing set of our dataset. We report the mean, median, best $25\\%$ , worst $25\\%$ , worst $5\\%$ , tri-mean, and maximum angular errors between the estimated illuminant colors and the ground truth colors for each method. Results are provided for both neutral and user-preference ground truth, with two models, one trained for each ground-truth type, except for the models SIIE [2], SIIE (tuned) [2], C5 [6], and C5 (tuned) [6], which were trained on images from the NUS [19] and Cube++ [23] datasets that do not include user-preference ground truth.", + "bbox": [ + 511, + 507, + 905, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we report the total number of parameters for methods that involve tunable or learnable parameters. The results in Table 1 are based on images without masking regions illuminated by a light source other than the dominant one used to obtain the ground truth, reflecting real-world scenarios where scenes with a single light source are not always guaranteed. For completeness, additional results are provided in the supplementary material, where regions illuminated by light sources different from the ground-truth illuminant color are masked out.", + "bbox": [ + 511, + 657, + 905, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further demonstrate the effectiveness of our method on existing DSLR datasets, where contextual metadata is typically unavailable, we report our results on the Simple Cube++ dataset [23]. Although geolocation metadata is absent, the dataset provides capture settings such as ISO and exposure time from the DSLR camera ISP.", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5043", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c3a0513e498cf13147d7f2739646f9ba91aefcf5fd5565513ae77bf3d767306e.jpg", + "table_caption": [ + "Table 1. Results on the testing set. We report the mean, median, best $25\\%$ , worst $25\\%$ , tri-mean, and maximum angular errors for each method on neutral and user-preference white-balance ground-truth illuminants, presented in the format (neutral / user-preference). Symbols n and r represent noise stats and SNR stats, respectively. The best and second-best results are highlighted." + ], + "table_footnote": [], + "table_body": "
MethodMeanMed.Best 25%Worst 25%Worst 5%Tri.Max#params (K)
GW [15]6.23 / 5.546.01 / 4.521.02 / 1.0012.43 / 11.9118.05 / 19.815.76 / 4.6528.09 / 31.70-
SoG [26]4.45 / 3.593.54 / 2.170.74 / 0.679.61 / 8.8114.94 / 15.613.77 / 2.6223.22 / 29.36-
GE-1st [62]4.21 / 3.333.29 / 2.120.71 / 0.559.34 / 8.2915.34 / 15.193.49 / 2.3627.81 / 28.80-
GE-2nd [62]4.11 / 3.183.17 / 1.890.70 / 0.589.09 / 7.7914.90 / 14.853.35 / 2.1825.09 / 29.42-
Max-RGB [14]4.01 / 2.612.92 / 1.881.06 / 0.978.57 / 5.6914.01 / 11.073.30 / 1.9434.22 / 23.63-
wGE [32]3.96 / 3.072.96 / 1.910.62 / 0.498.95 / 7.7914.94 / 15.083.21 / 2.1031.33 / 30.54-
PCA [19]4.42 / 3.633.54 / 1.950.70 / 0.619.59 / 9.1914.46 / 16.163.76 / 2.5222.87 / 31.58-
MSGP [57]6.39 / 5.695.72 / 4.380.94 / 0.9813.39 / 12.7320.97 / 22.335.64 / 4.6837.25 / 36.92-
GI [56]4.70 / 4.933.19 / 2.970.44 / 0.7811.63 / 12.2220.48 / 21.353.48 / 3.5336.34 / 36.02-
TECC [12]4.12 / 3.173.23 / 1.910.74 / 0.559.10 / 7.8314.73 / 14.483.46 / 2.1727.08 / 28.87-
Gamut (pixels) [30]3.77 / 2.402.81 / 1.490.77 / 0.638.31 / 5.8412.93 / 11.183.16 / 1.6321.53 / 23.310.636
Gamut (edges) [30]4.45 / 3.943.52 / 3.041.08 / 1.019.51 / 8.5015.00 / 15.253.70 / 3.1828.55 / 29.46324
Gamut (1st) [30]4.10 / 3.653.08 / 2.670.72 / 0.909.26 / 8.2514.75 / 14.693.28 / 2.9121.38 / 26.35279
NIS [29]4.58 / 3.903.76 / 2.570.77 / 0.759.81 / 9.1014.70 / 15.663.87 / 2.9720.76 / 31.800.078
Classification-CC [53]2.73 / 1.611.98 / 1.160.61 / 0.366.03 / 3.609.37 / 6.072.18 / 1.2719.34 / 9.5358,384
FFCC [11]2.62 / 1.501.46 / 0.810.37 / 0.246.89 / 3.9916.59 / 8.431.66 / 0.9548.97 / 18.6012
FFCC (capture info) [11]2.31 / 1.351.38 / 0.800.34 / 0.235.82 / 3.5112.44 / 7.361.60 / 0.8847.67 / 16.9636.9
FC4 [37]3.80 / 2.652.78 / 2.250.85 / 0.858.61 / 5.1415.06 / 7.522.86 / 2.3725.74 / 11.421,705
APAP (GW) [5]3.74 / 2.093.14 / 1.670.93 / 0.497.72 / 4.4411.09 / 6.893.26 / 1.7616.43 / 9.020.289
SIIE [2]4.09 / -3.25 / -0.91 / -8.97 / -15.96 / -3.37 / -43.24 / -1,008
SIIE (tuned) [2]3.15 / -2.22 / -0.51 / -7.28 / -12.14 / -2.46 / -34.52 / -1,008
SIIE (tuned-CS) [2]3.14 / 1.742.20 / 1.200.50 / 0.327.39 / 4.0613.61 / 6.732.41 / 1.3038.96 / 9.441,008
KNN (raw) [4]2.44 / 1.411.51 / 0.830.36 / 0.206.13 / 3.6611.64 / 7.011.66 / 0.9328.95 / 13.49757
Quasi-U-CC [13]3.85 / 3.262.97 / 1.810.55 / 0.588.47 / 8.2413.30 / 15.783.25 / 2.2124.74 / 32.1754,421
Quasi-U-CC (tuned) [13]3.11 / 2.542.27 / 1.430.49 / 0.447.14 / 6.5711.53 / 13.622.44 / 1.6322.67 / 33.7454,421
BoCF [41]3.54 / 2.142.68 / 1.590.96 / 0.487.31 / 4.7411.40 / 7.852.96 / 1.7222.31 / 19.6859
C4 [66]1.92 / 1.491.30 / 0.900.36 / 0.244.64 / 3.829.08 / 7.531.40 / 1.0321.55 / 18.095,116
CWCC [42]3.65 / 2.302.71 / 1.720.82 / 0.677.96 / 4.9512.52 / 9.652.99 / 1.8118.66 / 20.81101
C5 [6]3.22 / -2.51 / -0.78 / -6.97 / -10.54 / -2.68 / -16.38 / -412
C5 (tuned) [6]1.91 / -1.24 / -0.38 / -4.57 / -8.43 / -1.38 / -17.22 / -412
C5 (tuned-CS) [6]1.95 / 1.251.32 / 0.840.37 / 0.234.72 / 2.948.15 / 4.981.44 / 0.9316.78 / 9.23172
TLCC [59]2.71 / 2.742.06 / 1.830.66 / 0.695.89 / 6.3010.30 / 12.902.17 / 1.9921.44 / 33.3332,910
PCC [67]3.03 / 1.672.13 / 1.200.53 / 0.407.08 / 3.7911.20 / 6.842.34 / 1.2716.82 / 12.330.378
RGP [18]4.59 / 4.563.13 / 2.810.43 / 0.7011.13 / 11.2718.79 / 20.253.53 / 3.3132.11 / 33.98-
CFCC [16]3.07 / 1.542.20 / 1.050.73 / 0.396.87 / 3.5512.55 / 6.982.36 / 1.1323.34 / 14.400.283
Ours (w/o n, w/o r)1.93 / 1.261.35 / 0.770.38 / 0.234.56 / 3.139.48 / 5.881.43 / 0.8822.63 / 15.904.83
Ours (w/o n, w/ r)1.89 / 1.231.18 / 0.790.32 / 0.244.74 / 3.0110.10 / 5.071.30 / 0.9024.99 / 12.914.93
Ours (w/ n, w/o r)1.87 / 1.201.24 / 0.720.37 / 0.244.58 / 2.999.82 / 5.581.30 / 0.8229.46 / 15.124.93
Ours (w/ n, w/ r)1.84 / 1.201.24 / 0.770.35 / 0.194.41 / 2.959.17 / 5.121.32 / 0.8735.42 / 12.715.03
", + "bbox": [ + 93, + 138, + 910, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We trained our model using the histogram feature $\\mathbf{H}$ , along with capture features including ISO $(i)$ , exposure time $(e)$ , and SNR stats $(\\mathbf{r})$ , deliberately excluding the contextual metadata that is not available in the Simple Cube++ DSLR dataset. The results of our method, both with and without the SNR stats, are compared to other methods in Table 2. As shown, our method outperforms competing methods across most evaluation metrics.", + "bbox": [ + 89, + 671, + 482, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inference time: As shown in Table 1, our method performs comparably to or outperforms prior methods, while maintaining a compact model with only approximately 5K parameters. This lightweight design results in faster runtimes compared to other methods that achieve competitive results, namely C4 [66] and tuned C5 [6].", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Results on the testing set of the 'Simple Cube++' dataset [23]. We report the mean, median, best $25\\%$ , worst $25\\%$ , tri-mean, and maximum angular errors for each method. Symbols c, i, e, and r refer to the time-capture feature, ISO, exposure time, and SNR stats, respectively. The best and second-best results are highlighted.", + "bbox": [ + 511, + 669, + 906, + 755 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d67481ecb5d5df94a46120793116406a89df2975219ce6c792742523988d2219.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMeanMed.Best 25%Worst 25%Worst 5%Tri.Max
FFCC [11]1.380.580.184.0510.300.6845.20
FC4 [37]3.722.020.479.9317.182.4029.92
C4 [66]1.160.650.243.006.720.7313.95
C5 (tuned-CS) [6]1.190.600.183.267.610.6815.62
TLCC [59]1.821.150.394.499.631.2019.02
Ours (c = [i, e]T)1.080.590.172.845.530.6911.03
Ours (c = [i, e, rT]T)1.010.600.172.685.710.6510.89
", + "bbox": [ + 517, + 762, + 901, + 861 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5044", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8c50fac65325ec32da390d5d62047bb19702b1c3549669fed5e38994e8d777da.jpg", + "table_caption": [ + "Table 3. Processing time on an AMD Ryzen Threadripper PRO 3975WX CPU460 and an NVIDIA RTX A6000 GPU." + ], + "table_footnote": [], + "table_body": "
MethodCPU (ms)GPU (ms)FLOPs
C4 [66]49.0311.282.28G
C5 [6]8.284.50103.54M
Ours0.550.4516.78M
", + "bbox": [ + 161, + 106, + 416, + 159 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 3, we present the processing times of our model, C4, and C5 on an AMD Ryzen Threadripper PRO 3975WX CPU and an NVIDIA RTX A6000 GPU. Our fast runtime performance is particularly well-suited for mobile camera ISPs, where limited computational latency is crucial due to the processing demands of other modules (e.g., denoising, local tone mapping) per frame. Our model runs in just 0.25 ms on the DSP and 0.80 ms on the CPU of the Samsung S24 Ultra.", + "bbox": [ + 89, + 176, + 482, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies: We conducted ablation studies to evaluate the impact of each input feature on the validation set of our dataset. Specifically, we assessed our method by excluding either the time-capture feature (c) or the histogram feature $(\\mathbf{H})$ . Additionally, we tested the method using only the time feature $(\\mathbf{p})$ and noise stats $(\\mathbf{n})$ , while excluding the histogram feature $(\\mathbf{H})$ . We also evaluated the method using the histogram feature along with the time feature and noise stats, excluding other capture information. Furthermore, we examined the accuracy of our method using all input features except for $\\mathbf{p}$ . Lastly, we present the results when all input features are used, but without the noise stats $(\\mathbf{n})$ and the SNR stats $(\\mathbf{r})$ . These results are presented in Table 4.", + "bbox": [ + 89, + 318, + 482, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 4, using only the time feature $(\\mathbf{p})$ and noise stats $(\\mathbf{n})$ yields a reasonable accuracy $(2.37^{\\circ}$ mean angular error), compared to FFCC [11], which achieves $2.19^{\\circ}$ mean angular error on the validation set. This validates the usefulness of time-of-day and noise information in providing contextual clues. As expected, incorporating the histogram feature $(\\mathbf{H})$ , which represents scene colors, along with the time-of-day and noise stats, significantly boosts accuracy, as demonstrated by the results in the fourth row. We further investigate the impact of using noise information in the last three rows: using SNR stats $(\\mathbf{r})$ , using noise stats $(\\mathbf{n})$ , or both. The combination of both noise features yields the lowest mean angular error. Additional ablation studies are provided in the supplementary material.", + "bbox": [ + 89, + 529, + 482, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and limitations", + "text_level": 1, + "bbox": [ + 89, + 753, + 339, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a method for in-camera AWB correction that leverages contextual information. Specifically, we proposed a lightweight model that leverages contextual metadata (notably time-of-day derived from timestamp and geolocation) to guide the illuminant estimation process. In addition to this contextual metadata, we incorporate image colors in the form of histogram features, as well as capture information such as ISO, shutter speed, and noise stats, to help", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Results of ablation studies on the validation set with neutral white-balance ground truth. We report the mean, median, best $25\\%$ , and worst $25\\%$ angular errors for our method with various configurations. Symbols $\\mathbf{H}$ , $\\mathbf{p}$ , $\\mathbf{n}$ , and $\\mathbf{r}$ denote the histogram feature, time feature, noise stats, and SNR stats, respectively. The symbol $\\mathbf{m}$ refers to the capture information feature, which includes ISO $(i)$ , shutter speed $(s)$ , and flash status $(f)$ , all of which are used in our input time-capture feature $\\mathbf{c}$ . The best results are highlighted.", + "bbox": [ + 511, + 89, + 903, + 214 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d7635400aa35b730e3020735d8ff1d7a4800b2ce4cec98a514dce17f89ecc031.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
HmpnrMeanMed.Best 25%Worst 25%Worst 5%
✓XXXX2.031.360.314.908.90
X✓✓✓✓2.291.810.504.958.58
XX✓✓X2.371.590.455.489.03
✓X✓✓X1.751.170.294.137.26
✓✓X✓✓1.891.210.314.609.41
✓✓✓XX1.851.320.354.267.53
✓✓✓X✓1.721.160.304.138.16
✓✓✓✓X1.671.070.294.047.90
✓✓✓✓✓1.661.200.333.776.95
", + "bbox": [ + 519, + 215, + 895, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the model distinguish between artificial and natural scenes and improve its final accuracy. Our method is fast and can achieve high frame rates on modern smartphone DSPs and CPUs, while maintaining accurate illuminant color estimation.", + "bbox": [ + 511, + 378, + 903, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A key contribution of this work is a large-scale dataset of raw images captured by a consumer smartphone camera, along with the necessary contextual metadata for training and evaluating our method. Beyond the traditional neutral white-balance ground truth extracted from a calibration color chart placed in each scene, we also include a user-preference ground truth that targets the observer's preference, validated through a user study. Results based on both ground truth types demonstrate that our method achieves comparable or superior performance to existing methods, which require larger models.", + "bbox": [ + 511, + 458, + 905, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our method represents a promising solution for mobile camera ISPs, its dependency on contextual metadata limits its optimal accuracy to devices that provide geolocation data, which may not be available on DSLR cameras. Additionally, while the contextual metadata is device-independent (i.e., the differences across devices are expected to be minimal), our method relies on additional capture information (i.e., ISO, shutter speed, noise, and SNR stats) and image colors, making it inherently camera-dependent in design. This dependency prevents our trained model from generalizing to new devices without fine-tuning or re-training. However, this issue can be mitigated through color calibration (e.g., [46]), which can be extended to calibrate capture information as well—by performing a preprocessing mapping from the new camera space (for both color and capture information) to the camera space used during training. Additional discussion on cross-camera generalization is provided in the supplementary material.", + "bbox": [ + 511, + 628, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5045", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdelrahman Abdelhamed, Abhijith Punnappurath, and Michael S Brown. Leveraging the availability of two cameras for illuminant estimation. In CVPR, 2021. 2", + "[2] Mahmoud Affifi and Michael S Brown. Sensor-independent illumination estimation for DNN models. In BMVC, 2019. 1, 6, 7", + "[3] Mahmoud Affifi and Michael S Brown. Deep white-balance editing. In CVPR, 2020. 1", + "[4] Mahmoud Affifi, Brian Price, Scott Cohen, and Michael S Brown. When color constancy goes wrong: Correcting improperly white-balanced images. In CVPR, 2019. 6, 7", + "[5] Mahmoud Affifi, Abhijith Punnappurath, Graham Finlayson, and Michael S Brown. As-projective-as-possible bias correction for illumination estimation algorithms. JOSA A, 36 (1):71-78, 2019. 6, 7", + "[6] Mahmoud Afifi, Jonathan T Barron, Chloe LeGendre, Yun-Ta Tsai, and Francois Bleibel. Cross-camera convolutional color constancy. In ICCV, 2021. 1, 2, 4, 6, 7, 8", + "[7] Mahmoud Affifi, Zhenhua Hu, and Liang Liang. Optimizing illuminant estimation in dual-exposure HDR imaging. In ECCV, 2024. 1, 2, 5, 6", + "[8] Vivek Agarwal, Besma R Abidi, Andreas Koschan, and Mongi A Abidi. An overview of color constancy algorithms. Journal of Pattern Recognition Research, 1(1):42-54, 2006. 1", + "[9] Relja Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In CVPR, 2012. 4", + "[10] Jonathan T Barron. Convolutional color constancy. In ICCV, 2015. 1, 2, 4", + "[11] Jonathan T Barron and Yun-Ta Tsai. Fast Fourier color constancy. In CVPR, 2017. 2, 4, 6, 7, 8", + "[12] Simone Bianco and Marco Buzzelli. Truncated edge-based color constancy. In ICCE, 2022. 6, 7", + "[13] Simone Bianco and Claudio Cusano. Quasi-unsupervised color constancy. In CVPR, 2019. 1, 6, 7", + "[14] David H Brainard and Brian A Wandell. Analysis of the Retinex theory of color vision. Journal of the Optical Society of America A, 3(10):1651-1661, 1986. 6, 7", + "[15] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin Institute, 310(1): 1-26, 1980. 1, 6, 7", + "[16] Marco Buzzelli and Simone Bianco. A convolutional framework for color constancy. IEEE Transactions on Neural Networks and Learning Systems, pages 1-15, 2024. 6, 7", + "[17] Cheng Cheng, Kai-Fu Yang, Xue-Mei Wan, Leanne Lai Hang Chan, and Yong-Jie Li. Nighttime color constancy using robust gray pixels. JOSA A, 41(3):476-488, 2024. 5", + "[18] Cheng Cheng, Kai Fu Yang, Xue Mei Wan, Leanne Lai Hang Chan, and Yonge Jie Li. Nighttime color constancy using robust gray pixels. JOsA A, 41(3):476-488, 2024. 6, 7", + "[19] Dongliang Cheng, Dilip K Prasad, and Michael S Brown. Illuminant estimation for color constancy: Why spatial-domain methods work and the role of the color distribution. JOSA A, 31(5):1049-1058, 2014. 2, 5, 6, 7" + ], + "bbox": [ + 93, + 114, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[20] Kyungah Choi and Hyeon-Jeong Suk. User-preferred color temperature adjustment for smartphone display under varying illuminants. Optical Engineering, 53(6):061708-061708, 2014. 5", + "[21] Djork-Arné Clevert. Fast and accurate deep network learning by exponential linear units (ELUs). arXiv preprint arXiv:1511.07289, 2015. 5", + "[22] Maurizio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Mobile computational photography: A tour. Annual Review of Vision Science, 7(1):571-604, 2021. 3", + "[23] Egor Ershov, Alexey Savchik, Illya Semenkov, Nikola Banić, Alexander Belokopytov, Daria Senshina, Karlo Koščević, Marko Subašić, and Sven Lončarić. The cube++ illumination estimation dataset. IEEE Access, 8:227511-227527, 2020. 2, 5, 6, 7", + "[24] Egor Ershov, Alex Savchik, Denis Shepelev, Nikola Banić, Michael S Brown, Radu Timofte, Karlo Koščević, Michael Freeman, Vasily Tesalin, Dmitry Bocharov, et al. NTIRE 2022 challenge on night photography rendering. In CVPRW, 2022. 5", + "[25] Graham D Finlayson and Steven D Hordley. Color constancy at a pixel. JOSA A, 18(2):253-264, 2001. 4", + "[26] Graham D Finlayson and Elisabetta Trezzi. Shades of gray and colour constancy. In CIC, 2004. 1, 6, 7", + "[27] Karl R Gegenfurtner, David Weiss, and Marina Bloj. Color constancy in real-world settings. Journal of Vision, 24(2): 12-12, 2024. 1", + "[28] Peter Vincent Gehler, Carsten Rother, Andrew Blake, Tom Minka, and Toby Sharp. Bayesian color constancy revisited. In CVPR, 2008. 2, 5", + "[29] Arjan Gijsenij and Theo Gevers. Color constancy using natural image statistics and scene semantics. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(4): 687-698, 2011. 6, 7", + "[30] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Generalized gamut mapping using image derivative structures for color constancy. International Journal of Computer Vision, 86(2):127-139, 2010. 2, 6, 7", + "[31] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Computational color constancy: Survey and experiments. IEEE Transactions on Image Processing, 20(9):2475-2489, 2011. 1", + "[32] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):918-929, 2011. 1, 6, 7", + "[33] Thorsten Hansen, Sebastian Walter, and Karl R Gegenfurtner. Effects of spatial and temporal context on color categories and color constancy. Journal of Vision, 7(4):2-2, 2007. 1", + "[34] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics, 35(6):1-12, 2016. 3" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5046", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Xuanhua He, Tao Hu, Guoli Wang, Zejin Wang, Run Wang, Qian Zhang, Keyu Yan, Ziyi Chen, Rui Li, Chengjun Xie, et al. Enhancing RAW-to-sRGB with decoupled style structure in Fourier domain. In AAAI, 2024. 6", + "[36] Steven D Hordley and Graham D Finlayson. Re-evaluating colour constancy algorithms. In ICPR, 2004. 5", + "[37] Yuanming Hu, Baoyuan Wang, and Stephen Lin. FC4: Fully convolutional color constancy with confidence-weighted pooling. In CVPR, 2017. 6, 7", + "[38] Andrey Ignatov, Anastasia Sycheva, Radu Timofte, Yu Tseng, Yu-Syuan Xu, Po-Hsiang Yu, Cheng-Ming Chiang, Hsien-Kai Kuo, Min-Hung Chen, Chia-Ming Cheng, et al. MicroISP: processing 32MP photos on mobile devices with deep learning. In ECCV, 2022. 6", + "[39] Woohyeok Kim, Geonu Kim, Junyong Lee, Seungyong Lee, Seung-Hwan Baek, and Sunghyun Cho. ParamISP: Learned forward and inverse ISPs using camera parameters. In CVPR, 2024. 6", + "[40] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[41] Firas Laakom, Nikolaos Passalis, Jenni Raitoharju, Jarno Nikkanen, Anastasios Tefas, Alexandros Iosifidis, and Moncef Gabbouj. Bag of color features for color constancy. IEEE Transactions on Image Processing, 29:7722-7734, 2020. 1, 6, 7", + "[42] Firas Laakom, Jenni Raitoharju, Jarno Nikkanen, Alexandros Iosifidis, and Moncef Gabbouj. Robust channel-wise illumination estimation. In BMVC, 2021. 2, 6, 7", + "[43] Firas Laakom, Jenni Raitoharju, Jarno Nikkanen, Alexandros Iosifidis, and Moncef Gabbouj. Intel-TAU: A color constancy dataset. IEEE Access, 9:39560-39567, 2021. 2, 5", + "[44] Shuwei Li and Robby T Tan. NightCC: Nighttime color constancy via adaptive channel masking. In CVPR, 2024. 5", + "[45] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM Transactions on Graphics, 38(6):164-1, 2019. 3", + "[46] Yi-Tun Lin, Bianjiang Yang, Hao Xie, Wenbin Wang, Honghong Peng, and Jun Hu. Color constancy: How to deal with camera bias? In BMVC, 2023. 8", + "[47] Ce Liu, William T Freeman, Richard Szeliski, and Sing Bing Kang. Noise estimation from a single image. In CVPR, 2006. 4", + "[48] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6", + "[49] Zhongyu Lou, Theo Gevers, Ninghang Hu, Marcel P Lucassen, et al. Color constancy by deep learning. In BMVC, 2015. 1", + "[50] Ming Luo. A review of chromatic adaptation transforms. Review of Progress in Coloration and Related Topics, 30:77 - 92, 2008. 5", + "[51] Laurence T Maloney. Illuminant estimation as cue combination. Journal of Vision, 2(6):6-6, 2002. 1" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Jean H Meeus. *Astronomical algorithms*. Willmann-Bell, Incorporated, 1991. 3", + "[53] Seoung Wug Oh and Seon Joo Kim. Approaching the computational color constancy as a classification problem through deep learning. Pattern Recognition, 61:405-416, 2017. 1, 6, 7", + "[54] Varad A Pimpalkhute, Rutvik Page, Ashwin Kothari, Kishor M Bhurchandi, and Vipin Milind Kamble. Digital image noise estimation using DWT coefficients. IEEE Transactions on Image Processing, 30:1962-1972, 2021. 4", + "[55] Stanislav Pyatykh, Jürgen Hesser, and Lei Zheng. Image noise level estimation by principal component analysis. IEEE Transactions on Image Processing, 22(2):687-699, 2012. 4", + "[56] Yanlin Qian, Joni-Kristian Kamarainen, Jarno Nikkanen, and Jiri Matas. On finding gray pixels. In CVPR, 2019. 1, 6, 7", + "[57] Yanlin Qian, Said Pertuz, Jarno Nikkanen, Joni-Kristian Kämäräinen, and Jiri Matas. Revisiting gray pixel for statistical illumination estimation. In VISSAP. 2019. 1, 6, 7", + "[58] Wu Shi, Chen Change Loy, and Xiaou Tang. Deep specialized network for illuminant estimation. In ECCV, pages 371-387, 2016. 1", + "[59] Yuxiang Tang, Xuejing Kang, Chunxiao Li, Zhaowen Lin, and Anlong Ming. Transfer learning for color constancy via statistic perspective. In AAAI, 2022. 2, 6, 7", + "[60] Shoji Tominaga, Takahiko Horiuchi, Shiori Nakajima, and Mariko Yano. Prediction of incomplete chromatic adaptation under illuminant A from images. In CIC, 2014. 2, 5", + "[61] Oguzhan Ulucan, Diclehan Ulucan, and Marc Ebner. Color constancy beyond standard illuminants. In ICIP, 2022. 2, 5", + "[62] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy. IEEE Transactions on Image Processing, 16(9):2207-2214, 2007. 1, 6, 7", + "[63] TC Van Flandern and KF Pulkkinen. Low-precision formulae for planetary positions. Astrophysical Journal Supplement Series, 41:391-411, 1979. 3", + "[64] Robert Walraven. Calculating the position of the sun. Solar energy, 20(5):393-397, 1978. 3", + "[65] Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible image signal processing. In CVPR, 2021. 6", + "[66] Huanglin Yu, Ke Chen, Kaiqi Wang, Yanlin Qian, Zhaoxiang Zhang, and Kui Jia. Cascading convolutional color constancy. In AAAI, 2020. 1, 6, 7, 8", + "[67] Shuwei Yue and Minchen Wei. Color constancy from a pure color view. Journal of the Optical Society of America A, 40 (3):602-610, 2023. 6, 7", + "[68] Yi Zhang, Hongwei Qin, Xiaogang Wang, and Hongsheng Li. Rethinking noise synthesis and modeling in raw denoising. In ICCV, 2021. 4" + ], + "bbox": [ + 516, + 90, + 903, + 789 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5047", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Time-Aware Auto White Balance in Mobile Photography/59296f3d-511f-4955-b6aa-60113b360f05_model.json b/2025/Time-Aware Auto White Balance in Mobile Photography/59296f3d-511f-4955-b6aa-60113b360f05_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4118c2820b87043369f11fd38527136f62de75da --- /dev/null +++ b/2025/Time-Aware Auto White Balance in Mobile Photography/59296f3d-511f-4955-b6aa-60113b360f05_model.json @@ -0,0 +1,2310 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.214, + 0.131, + 0.784, + 0.154 + ], + "angle": 0, + "content": "Time-Aware Auto White Balance in Mobile Photography" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.174, + 0.397, + 0.19 + ], + "angle": 0, + "content": "Mahmoud Affifi*" + }, + { + "type": "text", + "bbox": [ + 0.428, + 0.174, + 0.52, + 0.19 + ], + "angle": 0, + "content": "Luxi Zhao*" + }, + { + "type": "text", + "bbox": [ + 0.551, + 0.174, + 0.733, + 0.192 + ], + "angle": 0, + "content": "Abhijith Punnappurath" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.192, + 0.461, + 0.208 + ], + "angle": 0, + "content": "Mohammed A. Abdelsalam" + }, + { + "type": "text", + "bbox": [ + 0.492, + 0.192, + 0.758, + 0.21 + ], + "angle": 0, + "content": "Ran Zhang Michael S. Brown" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.21, + 0.663, + 0.227 + ], + "angle": 0, + "content": "AI Center-Toronto, Samsung Electronics" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.23, + 0.86, + 0.246 + ], + "angle": 0, + "content": "{m.afifil, lucy.zhao, abhijith.p, m.abdelsalam, ran.zhang, michael.b1}@samsung.com" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.257, + 0.297, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.463, + 0.147, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.165, + 0.463, + 0.195, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.463, + 0.227, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.258, + 0.497, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.463, + 0.347, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.463, + 0.428, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.258, + 0.7, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.463, + 0.548, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.463, + 0.628, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.259, + 0.902, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.463, + 0.748, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.463, + 0.829, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.498, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Figure 1. The time of day influences scene illumination, making it a valuable cue for improving illuminant estimation. Shown are white-balanced results (gamma-corrected for visualization) using our method with (1) colors only, (2) colors plus contextual (timestamp and geolocation) and capture data, and (3) ground truth (from a color chart). Angular errors show improvements when time-capture information is used." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.556, + 0.327, + 0.571 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.578, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Cameras rely on auto white balance (AWB) to correct undesirable color casts caused by scene illumination and the camera's spectral sensitivity. This is typically achieved using an illuminant estimator that determines the global color cast solely from the color information in the camera's raw sensor image. Mobile devices provide valuable additional metadata—such as capture timestamp and geolocation—that offers strong contextual clues to help narrow down the possible illumination solutions. This paper proposes a lightweight illuminant estimation method that incorporates such contextual metadata, along with additional capture information and image colors, into a lightweight model (\\(\\sim 5K\\) parameters), achieving promising results, matching or surpassing larger models. To validate our method, we introduce a dataset of 3,224 smartphone images with contextual metadata collected at various times of day and under diverse lighting conditions. The dataset includes ground-truth illuminant colors, determined using a color chart, and user-preferred illuminants validated through a user study, providing a comprehensive benchmark for AWB evaluation." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.22, + 0.9 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.603, + 0.793, + 0.619 + ], + "angle": 0, + "content": "1. Introduction and related work" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.628, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Color constancy refers to the ability of the human visual system to maintain stable object colors despite variations in lighting conditions by leveraging contextual cues within the scene [27, 33, 51]. Cameras approximate this effect using auto white balance (AWB) correction, which aims to partially neutralize color casts introduced by scene illumination and the camera's spectral sensitivity [3]. AWB first estimates the illumination color as an RGB vector in the camera's raw color space. The raw image is then corrected by scaling its color channels according to the estimated illumination, typically under the assumption of a single global light source [8, 31]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Conventional illuminant estimation methods primarily rely on image colors, either by directly processing the raw image (e.g., [13, 32, 41, 49, 53, 58, 62, 66]) or by analyzing color histograms (e.g., [2, 6, 7, 10]). These methods can be broadly categorized into two groups: (1) classical statistical-based methods (e.g., [15, 26, 32, 56, 57, 62])," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5038" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.475, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.465 + ], + "angle": 0, + "content": "Figure 2. The time of day at which an image is captured provides valuable information about the possible range of illuminant colors in outdoor scenes. The figure presents the correlated color temperature (CCT) of illuminant colors in our dataset (Sec. 3) for images captured at various times throughout the day and night. As shown, excluding images taken under artificial light, those captured at noon, for example, exhibit a different range of illuminant CCTs compared to images captured during sunset or sunrise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.547 + ], + "angle": 0, + "content": "which estimate the illuminant color based on image statistics, and (2) learning-based methods (e.g., [6, 10, 30, 42, 59]), which map image colors to their corresponding scene illuminant through data-driven models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.655 + ], + "angle": 0, + "content": "While image colors are a key cue for estimating the scene's illuminant, mobile devices provide an opportunity to integrate additional contextual information. For instance, the device's location, along with the date and time, offer valuable cues about outdoor lighting conditions (e.g., sunrise, noon, sunset), thereby improving illuminant estimation for outdoor scenes (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Intuitively, knowing the time of day when an outdoor scene is captured can help estimate the lighting conditions. Figure 2 further illustrates that the time of day, derived from contextual metadata (i.e., timestamp and geolocation) readily available on mobile devices, provides valuable insights into the likely range of illuminant correlated color temperature (CCT) in outdoor scenes. This information serves to narrow the range of possible illuminant colors. For instance, images taken at noon exhibit a different CCT range than those captured at sunrise or sunset. When combined with additional capture information to distinguish between environments (e.g., indoors vs. outdoors), such metadata can complement image colors to improve illuminant estimation accuracy." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Despite its potential, only a few attempts have explored leveraging additional information available to the camera's" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "image signal processor (ISP) for illuminant estimation, such as metadata-based model control [11] or data augmentation to enhance generalization [6]. However, to our knowledge, no previous work has investigated using contextual information—specifically, mobile device timestamps and geolocation—to refine illuminant estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.203, + 0.907, + 0.371 + ], + "angle": 0, + "content": "We introduce a method that leverages contextual metadata from mobile phones, along with additional capture information available in camera ISPs, to train a lightweight illuminant estimator model with \\(\\sim 5\\mathrm{K}\\) parameters. Our model delivers promising results, matching or surpassing that of larger models, while maintaining efficiency. Our model runs on a typical flagship mobile digital signal processor (DSP) and CPU in \\(0.25\\mathrm{ms}\\) and \\(0.80\\mathrm{ms}\\), respectively. This compact and efficient design is especially advantageous for mobile devices, where minimizing power consumption and memory usage is critical [1, 7, 11]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.908, + 0.662 + ], + "angle": 0, + "content": "Existing white-balance datasets (e.g., [19, 23, 28, 43]) lack the contextual information needed to validate our method. The absence of contextual information is because most available datasets (e.g., [19, 23]) used DSLR cameras, which typically lack built-in GPS functionality or accurate timestamps. To address this, we captured a new dataset of 3,224 images using a consumer smartphone camera, accompanied by contextual and capture information. The ground truth for the dataset was established using two approaches: (1) the conventional approach, selecting gray patches from a calibration color chart, and (2) a manually selected user-preference white-balance target, which enhances the image's aesthetic appeal and accounts for incomplete chromatic adaptation [60]. The user-preference white-balance \"ground truth\" was validated through a user study. The dataset spans a variety of lighting conditions, including non-standard artificial illuminants [61], and covers different times of day—sunrise, noon, sunset, and night." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Contribution: We propose an AWB method that utilizes smartphone contextual metadata (i.e., timestamp and geolocation) alongside capture information to enhance illuminant estimation. We demonstrate that integrating this additional data with conventional color information into a lightweight neural model leads to significant improvements in accuracy. Additionally, we introduce a large-scale dataset captured with a consumer smartphone at various times of day, with ground truth derived from both a calibration color chart and user-preference-based white balance. Benchmarking results show that our method achieves state-of-the-art results in most standard metrics (i.e., angular error statistics [19]), while maintaining computational efficiency on mobile devices." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5039" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.482, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.275, + 0.485, + 0.4 + ], + "angle": 0, + "content": "Figure 3. Our method includes a lightweight model consisting of a convolutional network that processes the histogram feature (derived from raw image colors and edge histograms concatenated with the \\(u / v\\) coordinates) to produce a latent feature \\(\\mathbf{v}_h\\in \\mathbb{R}^{16}\\). This is then concatenated with the latent feature of the processed time-capture feature, \\(\\mathbf{v}_t\\in \\mathbb{R}^{16}\\). The combined feature, \\(\\mathbf{v}\\in \\mathbb{R}^{32}\\), is passed through a lightweight MLP to produce the chromaticity of the scene illuminant, \\(\\ell_c\\in \\mathbb{R}^2\\), which is finally converted into normalized RGB illuminant color." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.419, + 0.182, + 0.433 + ], + "angle": 0, + "content": "2. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.484, + 0.687 + ], + "angle": 0, + "content": "An overview of our method is shown in Fig. 3. Our method employs a learnable model that processes two distinct inputs: (1) the time-capture feature (Sec. 2.1), which combines the contextual and capture information available on mobile devices and accessible by their camera ISPs, and (2) the histogram feature (Sec. 2.2), which represents the image's R/G and B/G chromaticity values. The time-capture feature is first processed to project it into a latent space, producing the time-capture latent feature vector, \\(\\mathbf{v}_t \\in \\mathbb{R}^{16}\\). The model processes the histogram feature to produce the histogram's latent feature vector, \\(\\mathbf{v}_h \\in \\mathbb{R}^{16}\\). Both the histogram and time-capture latent feature vectors are concatenated and processed by a set of learnable layers to output an R/G, B/G chromaticity vector, \\(\\ell_c \\in \\mathbb{R}^2\\), representing the scene illuminant. This 2D vector is then converted into the illuminant RGB color used to perform white balancing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.291, + 0.714 + ], + "angle": 0, + "content": "2.1. Time-capture feature" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our model incorporates contextual metadata, available on mobile devices, as one of its input features. Specifically, we use the geolocation and timestamp of image capture to compute the \"probability\" of the time of day (e.g., sunset, noon, etc.). Our approach converts conventional clock time (hour::minutes) to its corresponding time of day condition based on the date and geolocation of the scene. This approach allows the model to generalize across different time zones and prevents it from being influenced by the capturing location. The time probability vector represents the likelihood that the captured image corresponds to one of the solar event times (i.e., dawn, sunrise, noon, sunset, dusk, or mid" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.744, + 0.107 + ], + "angle": 0, + "content": "night) and is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.119, + 0.905, + 0.151 + ], + "angle": 0, + "content": "\\[\np _ {g} = 1 - \\frac {\\left| t _ {c} - t _ {g} \\right|}{t _ {s}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.157, + 0.907, + 0.262 + ], + "angle": 0, + "content": "where \\( t_c \\) is the capture time in seconds (adjusted to the local time zone based on geolocation information), and \\( t_g \\) is the local time in seconds of the solar event, \\( g \\in \\{\\text{dawn,sunrise,noon,sunset,dusk,midnight}\\} \\), computed using geolocation-based standard algorithms [52, 63, 64]. The scalar \\( t_s \\) represents the total number of seconds in a day (i.e., 86,400)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.264, + 0.907, + 0.445 + ], + "angle": 0, + "content": "We pre-process the probability of each solar event, \\( p_{g} \\), in our time probability vector by computing the square root to enhance feature representation—compressing high probabilities and amplifying lower ones, which is intended to help create a more balanced and smooth distribution for the model to leverage. We then augment this time probability vector with a one-hot vector, \\( \\mathbf{b} \\), that indicates whether the capture time \\( t_{c} \\) occurs before each solar event. This distinction helps the model account for expected variations in CCT, as illuminant colors can differ before and after certain solar events, such as sunset and sunrise. The value of this one-hot vector for a given solar event \\( g \\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.456, + 0.905, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathbf {b} _ {g} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} t _ {c} \\leq t _ {g} \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.508, + 0.905, + 0.554 + ], + "angle": 0, + "content": "where \\(\\mathbf{b}_g\\) corresponds to the entry in the one-hot vector for the solar event, \\(g\\). Both the time probability vector and the one-hot vector together form our time feature, \\(\\mathbf{p} \\in \\mathbb{R}^{12}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.554, + 0.906, + 0.627 + ], + "angle": 0, + "content": "To enrich our feature set with additional capture information available from the camera ISP and help distinguish the capturing environment (e.g., indoor vs. outdoor), we include the following features in our final time-capture feature, c:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.629, + 0.906, + 0.689 + ], + "angle": 0, + "content": "- ISO (i): The sensitivity of the camera's image sensor to light, where lower values indicate good lighting conditions (e.g., bright scenes) and higher values suggest low-light environments, such as poorly lit indoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.69, + 0.906, + 0.749 + ], + "angle": 0, + "content": "- Shutter speed \\((s)\\): The amount of time the camera's shutter remains open, allowing light to hit the image sensor. It provides an indication of lighting conditions, alongside the ISO value, \\(i\\)." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.75, + 0.905, + 0.78 + ], + "angle": 0, + "content": "- Flash status \\((f)\\): A binary value indicating whether flash light was used during capturing." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.781, + 0.907, + 0.901 + ], + "angle": 0, + "content": "- Noise information (optional): Since image denoising is typically applied before or in parallel with illuminant estimation in camera ISPs [22, 34, 45], we also explore the optional use of explicit noise information from the captured scene. More noise typically indicates low-light conditions, which can provide clues about the lighting color range of the scene. Unfortunately, while noise information is accessible within camera ISPs, obtaining accurate" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5040" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.483, + 0.166 + ], + "angle": 0, + "content": "noise information for public use is challenging, as the noise profiles in DNG files are not always reliable [68]. To address this, we simulate the noise information using two approaches: noise statistics (stats) and/or signal-to-noise ratio (SNR) stats, which are described below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.174, + 0.483, + 0.309 + ], + "angle": 0, + "content": "Noise stats (n): This represents the noise statistics in the captured raw image. We simulate this by denoising each raw image using Adobe Lighthroom and computing the noise stats as the mean and standard deviation of each color channel of the absolute difference between the denoised and noisy raw images. This approach provides a simplified method for estimating noise stats, as the denoised images are typically available within the camera ISPs, but difficult to extract from the DNG files." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.484, + 0.429 + ], + "angle": 0, + "content": "SNR stats (r): This alternative approach measures the noise information in the captured image without the need of a denoised reference. The SNR is computed by applying a \\(15 \\times 15\\) sliding window over the raw image and calculated as: \\(10\\log_{10}\\left(\\frac{\\mu}{\\sigma + \\epsilon}\\right)\\), where \\(\\mu\\) represents the mean RGB of the \\(15 \\times 15\\) patch, \\(\\sigma\\) is the standard deviation, and \\(\\epsilon\\) is a small value added for numerical stability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.49 + ], + "angle": 0, + "content": "Note that while these two approaches—namely, computing noise stats and SNR stats—yield satisfactory results, there is a wide body of research on noise estimation (e.g., [47, 54, 55]), which is beyond the scope of this paper." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.521 + ], + "angle": 0, + "content": "Our complete time-capture feature is the combination of these inputs and can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.533, + 0.483, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\mathbf {c} = \\left[ \\mathbf {p} ^ {T}, \\log (i), \\log (s), f, \\mathbf {w} ^ {T} \\right] ^ {T}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.626 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}\\) represents optional noise-related features, which can include either noise stats \\(\\mathbf{n}\\), SNR stats \\(\\mathbf{r}\\), or both. Alternatively, \\(\\mathbf{w}\\) can be omitted if noise information is not considered." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.688 + ], + "angle": 0, + "content": "Our time-capture feature is first normalized using min-max normalization (with the min and max values computed from the training data), and then processed through a learnable function, \\( f_{t} \\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.246, + 0.706, + 0.483, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {t} = f _ {t} (\\mathbf {c}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.777 + ], + "angle": 0, + "content": "where \\(f_{t}\\) is a learnable linear layer that transforms the time-capture feature, \\(\\mathbf{c}\\), into its latent representation, \\(\\mathbf{v}_t\\in \\mathbb{R}^{16}\\) (see Fig. 3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.268, + 0.803 + ], + "angle": 0, + "content": "2.2. Histogram feature" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In addition to the time-capture feature, c, we provide our neural model with a histogram feature, \\(\\mathbf{H}\\) which represents the colors of the raw image. Inspired by prior work [6, 10, 11], we use a 2D histogram to represent the R/G and B/G chromaticities of the input raw image, \\(\\mathbf{I} \\in \\mathbb{R}^{K \\times 3}\\), where \\(K\\) denotes the total number of pixels in the raw im" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "age. Specifically, we compute the 2D chromaticity histogram, \\(\\mathbf{H}_c\\in \\mathbb{R}^{h\\times h}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.134, + 0.905, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {c} ^ {(m, n)} = \\sum_ {k} \\| \\mathbf {I} ^ {(k)} \\| _ {2} \\cdot \\delta_ {m, n} ^ {(k)}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.192, + 0.905, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\delta_ {m, n} ^ {(k)} = \\left[ u _ {m} \\leq r g ^ {(k)} < u _ {m + 1} \\right] \\wedge \\left[ v _ {n} \\leq b g ^ {(k)} < v _ {n + 1} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.907, + 0.451 + ], + "angle": 0, + "content": "where \\( rg^{(k)} \\) and \\( bg^{(k)} \\) are the R/G and B/G chromaticity values of pixel \\( k \\) in I, and \\( \\| \\mathbf{I}^{(k)} \\|_2 \\) represents the intensity of pixel \\( k \\) (i.e., Euclidean norm of the pixel's RGB values). The notation \\( \\wedge \\) denotes the logical AND operator, and \\( [\\cdot] \\) is the Iverson bracket, which evaluates to 1 when the condition is true and 0 otherwise. The histogram bins are defined by the edges \\( \\{u_m\\} \\) and \\( \\{v_n\\} \\), where \\( u_m \\) and \\( v_n \\) are the lower edges of the bins, and \\( u_{m+1} \\) and \\( v_{n+1} \\) are the corresponding upper edges. The histogram accumulates the brightness values of pixels whose chromaticity values fall within the range \\( [u_m, u_{m+1}) \\) along the R/G axis (i.e., the \\( u \\)-axis) and \\( [v_n, v_{n+1}) \\) along the B/G axis (i.e., the \\( v \\)-axis). The number of histogram bins along each axis is denoted as \\( h \\). Following [10], we compute the square root of the histogram to enhance the utility of the histogram feature [9]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.906, + 0.541 + ], + "angle": 0, + "content": "Note that this histogram differs from the \\( \\log -uv \\) histogram used in prior work [6, 10, 11], which operates in the logarithmic space of G/R and G/B [25]. We found that the R/G, B/G chromaticity histogram performs better, as it aligns with our model's output space. See the supplemental material for an ablation study." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.906, + 0.616 + ], + "angle": 0, + "content": "In addition to the chromaticity histogram, \\(\\mathbf{H}_c\\), of the image colors, and building on prior work [6, 11], we augment it with the square-rooted chromaticity histogram of the image's edges, \\(\\mathbf{H}_e\\), where the image edges are computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.641, + 0.905, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} ^ {(x, y)} = \\frac {1}{8} \\sum_ {\\Delta x, \\Delta y} \\left| \\mathbf {I} ^ {\\prime (x, y)} - \\mathbf {I} ^ {\\prime (x + \\Delta x, y + \\Delta y)} \\right|, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.905, + 0.75 + ], + "angle": 0, + "content": "where \\(\\mathbf{E}\\) represents the image edges, \\(\\mathbf{I}'\\) refers to the raw image in 3D tensor form (height, width, channels), and \\(\\Delta x, \\Delta y \\in \\{-1,0,1\\}\\) with \\((\\Delta x, \\Delta y) \\neq (0,0)\\). The edge histogram, \\(\\mathbf{H}_e\\), is computed from \\(\\mathbf{E}\\) using Eqs. 5 and 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our histogram feature is constructed by concatenating these two histograms, \\(\\mathbf{H}_c\\) and \\(\\mathbf{H}_e\\). Since this histogram feature is first processed by convolutional (conv) layers, as shown in Fig. 3, we follow [6] by appending additional channels that encode the positional information of the \\(u / v\\) coordinates in histogram space, which helps capture spatial relationships within the histogram feature. Consequently, our final histogram feature, \\(\\mathbf{H} \\in \\mathbb{R}^{h \\times h \\times 4}\\), consists of 2 channels representing the chromaticity of the image and its edges, along with the additional \\(u / v\\) coordinate channels." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5041" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "This feature is processed through a series of conv layers with ELU activation [21]. The resulting latent representation undergoes adaptive average pooling before passing through a linear layer to produce the histogram's latent feature vector, \\(\\mathbf{v}_h\\in \\mathbb{R}^{16}\\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.183, + 0.483, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {h} = f _ {h} (\\mathbf {H}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.206, + 0.484, + 0.253 + ], + "angle": 0, + "content": "where \\( f_{h} \\) denotes the sub-model (i.e., conv layers, ELU activation, pooling, and linear layer) that maps the histogram feature into its latent space, as shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.262, + 0.295, + 0.277 + ], + "angle": 0, + "content": "2.3. Illuminant estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.284, + 0.484, + 0.33 + ], + "angle": 0, + "content": "Both feature vectors, \\(\\mathbf{v}_t\\) and \\(\\mathbf{v}_h\\), are concatenated to produce the latent vector \\(\\mathbf{v}\\) and processed by an illuminant estimation sub-model as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.345, + 0.483, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} = \\left[ \\mathbf {v} _ {t}; \\mathbf {v} _ {h} \\right], \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.246, + 0.384, + 0.483, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\ell_ {c} = f _ {\\ell} (\\mathbf {v}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.56 + ], + "angle": 0, + "content": "where \\( f_{\\ell} \\) consists of a set of linear layers, with batch normalization applied to the first layer, followed by activation functions—except for the final layer, which outputs \\( \\ell_{c} \\), the chromaticity vector of the scene illuminant. This vector is then transformed into an unnormalized RGB illuminant color by mapping \\( [\\mathrm{R / G},\\mathrm{B / G}]^T\\rightarrow [\\mathrm{R / G},1,\\mathrm{B / G}]^T \\), followed by normalization via division by its L2 norm to produce the final illuminant color. We optimize \\( f_{t} \\), \\( f_{h} \\), and \\( f_{\\ell} \\) to minimize the angular error [36] between the predicted RGB illuminant color and the ground-truth RGB illuminant color." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.574, + 0.182, + 0.589 + ], + "angle": 0, + "content": "3. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.84 + ], + "angle": 0, + "content": "To train and validate our method, we require a dataset that includes contextual information (i.e., timestamp and geolocation) h h image. Existing datasets (e.g., [17, 19, 23, 28, 43, 44]) lack this essential information, motivating us to collect a new dataset using a smartphone camera that provides contextual metadata for each image. Specifically, we captured 3,224 linear raw images with the Samsung S24 Ultra's main camera, covering a wide range of scenes both indoors and outdoors, at various times of day (e.g., sunset, sunrise, noon, night). Our dataset includes images captured under various light sources (e.g., sunlight, incandescent, LED), as well as non-standard illuminant colors (e.g., colored LED light), which are not present in existing datasets [61]; see Fig. 4. Additionally, our dataset captures scenes under different weather conditions (sunny, cloudy, rain, snow, etc.). Example scenes can be found in the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We follow prior work [7, 43] in collecting ground-truth illuminant colors for each scene. Specifically, for each scene, we first capture an image with a calibration color chart, which is used to extract the illuminant color from the" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.095, + 0.903, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.334, + 0.907, + 0.405 + ], + "angle": 0, + "content": "Figure 4. A plot of the \\( rg \\) chromaticity distribution of ground-truth illuminant colors for neutral white balance in our dataset. Example images in raw and sRGB spaces are shown (raw images are gamma-corrected for visualization). See supplemental material the \\( rg \\) chromaticity distribution of user-preference illuminants." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.906, + 0.532 + ], + "angle": 0, + "content": "gray patches. Next, we capture an image of the same scene without the color chart (see Fig. 5), resulting in \\(\\sim 6\\mathrm{K}\\) images. After obtaining the ground-truth illuminant color, we discard all color chart images, leaving us with 3,224 images in our dataset. This approach allows us to test with natural images that mimic real-world scenarios, without the need to mask out the color chart patch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.747 + ], + "angle": 0, + "content": "Since our dataset includes a wide variety of lighting conditions, such as sunset/sunrise, night scenes, and artificial light, we generate a \"user-preference\" ground truth in addition to the neutral ground truth obtained from the color chart for each scene in our dataset. This is to account for human incomplete chromatic adaptation in such scenes [50, 60] and user preference [20, 24]. Specifically, an expert photographer was asked to assign a ground truth illuminant to each scene to make it appear more natural, reflect real-world observations, and enhance the aesthetics of the image. Notably, the same person who captured the scene also performed the annotation, ensuring that the user-preference selection was based on real-world observations of how the scene should appear." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We validated the user-preference ground truth through a study with 20 participants, who selected the most natural image from pairs of white-balanced images corresponding to the user-preference ground truth and the neutral ground truth (obtained from the color chart). The user-preference ground truth was selected in \\(71.95\\%\\) of the trials, confirming its preference. See the supplemental material for additional details. In our experiments, both ground-truth types—neutral and user-preference—were used for training and evaluation (Sec. 4)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5042" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.905, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.204, + 0.908, + 0.261 + ], + "angle": 0, + "content": "Figure 5. For each scene in our dataset, an image with a color chart placed in the scene was captured. The gray patches on the color chart were used to measure the ground-truth illuminant. These color-chart images were discarded, and only the images of the scenes without the color chart were used for the training, validation, and testing sets. For each example, we show both raw and sRGB images, with the raw images gamma-corrected to enhance visualization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.279, + 0.482, + 0.4 + ], + "angle": 0, + "content": "Additionally, we created a mask for regions illuminated by non-dominant illuminants in each scene. These masks ensure that all scenes have only one dominant illuminant, matching the color of the neutral ground truth, without confounding effects from other illuminants. We applied these masks to the training images when training our method and others. To preserve privacy, sensitive information, such as car plates and faces, has been blurred." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.402, + 0.483, + 0.554 + ], + "angle": 0, + "content": "In addition to raw images, both ground-truth illuminant types, and contextual and capture information, we provide additional valuable auxiliary data to broaden the impact of our dataset. This includes locally tone-mapped sRGB images rendered by an expert, that can serve as ground-truth for applications beyond white-balance correction, such as neural ISPs [35, 38, 39, 65]. More information about the dataset can be found in the supplemental material. We organize our dataset into 2,619 raw images for training, 205 raw images for validation, and 400 raw images for testing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.573, + 0.224, + 0.59 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.606, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Implementation details: We train the model using the Adam optimizer [40], for 400 epochs with betas set to (0.9, 0.999) and a weight decay factor of \\(10^{-9}\\). A warm-up strategy is applied to gradually increase the learning rate from \\(10^{-6}\\) to \\(10^{-3}\\) over the first 5 epochs. After this, we use a cosine annealing schedule [48]. Following [6], we employ a batch-size increment strategy during training, starting with a batch size of 8 and doubling it every 100 epochs. Following prior work [6, 7], we use images of size \\(384 \\times 256\\) pixels in all experiments, which is a reasonable size for camera pipelines to reduce computational overhead. For our method, we use histograms with \\(h = 48\\) bins. The histogram boundaries are determined by computing the 10th and 95th percentiles of the chromaticity values along each chromaticity axis from the training set. We report the results of our method both with and without the inclusion of noise stats, n, and SNR stats, r." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Results: We report the results of our method alongside several others benchmarked on the proposed dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.905, + 0.385 + ], + "angle": 0, + "content": "Our method is compared with various statistical-based approaches [12, 14, 15, 18, 19, 26, 32, 56, 57, 62], camera-specific learning-based methods [4, 5, 11, 16, 29, 30, 37, 41, 42, 53, 59, 66, 67], and camera-independent techniques [2, 6, 13]. For the camera-specific learning-based approaches, including our method, all models were trained on our training set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.507 + ], + "angle": 0, + "content": "For the camera-independent methods [2, 6], we present results from models trained on the NUS [19] and Cube++ [23] datasets. Additionally, we report results from finetuned versions of these models [2, 6], incorporating our proposed dataset. We also evaluate camera-specific (CS) models of these methods [2, 6], which were trained exclusively on our training data, without any additional datasets. For further details, refer to the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Table 1 shows the results on the testing set of our dataset. We report the mean, median, best \\(25\\%\\), worst \\(25\\%\\), worst \\(5\\%\\), tri-mean, and maximum angular errors between the estimated illuminant colors and the ground truth colors for each method. Results are provided for both neutral and user-preference ground truth, with two models, one trained for each ground-truth type, except for the models SIIE [2], SIIE (tuned) [2], C5 [6], and C5 (tuned) [6], which were trained on images from the NUS [19] and Cube++ [23] datasets that do not include user-preference ground truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Additionally, we report the total number of parameters for methods that involve tunable or learnable parameters. The results in Table 1 are based on images without masking regions illuminated by a light source other than the dominant one used to obtain the ground truth, reflecting real-world scenarios where scenes with a single light source are not always guaranteed. For completeness, additional results are provided in the supplementary material, where regions illuminated by light sources different from the ground-truth illuminant color are masked out." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To further demonstrate the effectiveness of our method on existing DSLR datasets, where contextual metadata is typically unavailable, we report our results on the Simple Cube++ dataset [23]. Although geolocation metadata is absent, the dataset provides capture settings such as ISO and exposure time from the DSLR camera ISP." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5043" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.134 + ], + "angle": 0, + "content": "Table 1. Results on the testing set. We report the mean, median, best \\(25\\%\\), worst \\(25\\%\\), tri-mean, and maximum angular errors for each method on neutral and user-preference white-balance ground-truth illuminants, presented in the format (neutral / user-preference). Symbols n and r represent noise stats and SNR stats, respectively. The best and second-best results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.139, + 0.911, + 0.65 + ], + "angle": 0, + "content": "
MethodMeanMed.Best 25%Worst 25%Worst 5%Tri.Max#params (K)
GW [15]6.23 / 5.546.01 / 4.521.02 / 1.0012.43 / 11.9118.05 / 19.815.76 / 4.6528.09 / 31.70-
SoG [26]4.45 / 3.593.54 / 2.170.74 / 0.679.61 / 8.8114.94 / 15.613.77 / 2.6223.22 / 29.36-
GE-1st [62]4.21 / 3.333.29 / 2.120.71 / 0.559.34 / 8.2915.34 / 15.193.49 / 2.3627.81 / 28.80-
GE-2nd [62]4.11 / 3.183.17 / 1.890.70 / 0.589.09 / 7.7914.90 / 14.853.35 / 2.1825.09 / 29.42-
Max-RGB [14]4.01 / 2.612.92 / 1.881.06 / 0.978.57 / 5.6914.01 / 11.073.30 / 1.9434.22 / 23.63-
wGE [32]3.96 / 3.072.96 / 1.910.62 / 0.498.95 / 7.7914.94 / 15.083.21 / 2.1031.33 / 30.54-
PCA [19]4.42 / 3.633.54 / 1.950.70 / 0.619.59 / 9.1914.46 / 16.163.76 / 2.5222.87 / 31.58-
MSGP [57]6.39 / 5.695.72 / 4.380.94 / 0.9813.39 / 12.7320.97 / 22.335.64 / 4.6837.25 / 36.92-
GI [56]4.70 / 4.933.19 / 2.970.44 / 0.7811.63 / 12.2220.48 / 21.353.48 / 3.5336.34 / 36.02-
TECC [12]4.12 / 3.173.23 / 1.910.74 / 0.559.10 / 7.8314.73 / 14.483.46 / 2.1727.08 / 28.87-
Gamut (pixels) [30]3.77 / 2.402.81 / 1.490.77 / 0.638.31 / 5.8412.93 / 11.183.16 / 1.6321.53 / 23.310.636
Gamut (edges) [30]4.45 / 3.943.52 / 3.041.08 / 1.019.51 / 8.5015.00 / 15.253.70 / 3.1828.55 / 29.46324
Gamut (1st) [30]4.10 / 3.653.08 / 2.670.72 / 0.909.26 / 8.2514.75 / 14.693.28 / 2.9121.38 / 26.35279
NIS [29]4.58 / 3.903.76 / 2.570.77 / 0.759.81 / 9.1014.70 / 15.663.87 / 2.9720.76 / 31.800.078
Classification-CC [53]2.73 / 1.611.98 / 1.160.61 / 0.366.03 / 3.609.37 / 6.072.18 / 1.2719.34 / 9.5358,384
FFCC [11]2.62 / 1.501.46 / 0.810.37 / 0.246.89 / 3.9916.59 / 8.431.66 / 0.9548.97 / 18.6012
FFCC (capture info) [11]2.31 / 1.351.38 / 0.800.34 / 0.235.82 / 3.5112.44 / 7.361.60 / 0.8847.67 / 16.9636.9
FC4 [37]3.80 / 2.652.78 / 2.250.85 / 0.858.61 / 5.1415.06 / 7.522.86 / 2.3725.74 / 11.421,705
APAP (GW) [5]3.74 / 2.093.14 / 1.670.93 / 0.497.72 / 4.4411.09 / 6.893.26 / 1.7616.43 / 9.020.289
SIIE [2]4.09 / -3.25 / -0.91 / -8.97 / -15.96 / -3.37 / -43.24 / -1,008
SIIE (tuned) [2]3.15 / -2.22 / -0.51 / -7.28 / -12.14 / -2.46 / -34.52 / -1,008
SIIE (tuned-CS) [2]3.14 / 1.742.20 / 1.200.50 / 0.327.39 / 4.0613.61 / 6.732.41 / 1.3038.96 / 9.441,008
KNN (raw) [4]2.44 / 1.411.51 / 0.830.36 / 0.206.13 / 3.6611.64 / 7.011.66 / 0.9328.95 / 13.49757
Quasi-U-CC [13]3.85 / 3.262.97 / 1.810.55 / 0.588.47 / 8.2413.30 / 15.783.25 / 2.2124.74 / 32.1754,421
Quasi-U-CC (tuned) [13]3.11 / 2.542.27 / 1.430.49 / 0.447.14 / 6.5711.53 / 13.622.44 / 1.6322.67 / 33.7454,421
BoCF [41]3.54 / 2.142.68 / 1.590.96 / 0.487.31 / 4.7411.40 / 7.852.96 / 1.7222.31 / 19.6859
C4 [66]1.92 / 1.491.30 / 0.900.36 / 0.244.64 / 3.829.08 / 7.531.40 / 1.0321.55 / 18.095,116
CWCC [42]3.65 / 2.302.71 / 1.720.82 / 0.677.96 / 4.9512.52 / 9.652.99 / 1.8118.66 / 20.81101
C5 [6]3.22 / -2.51 / -0.78 / -6.97 / -10.54 / -2.68 / -16.38 / -412
C5 (tuned) [6]1.91 / -1.24 / -0.38 / -4.57 / -8.43 / -1.38 / -17.22 / -412
C5 (tuned-CS) [6]1.95 / 1.251.32 / 0.840.37 / 0.234.72 / 2.948.15 / 4.981.44 / 0.9316.78 / 9.23172
TLCC [59]2.71 / 2.742.06 / 1.830.66 / 0.695.89 / 6.3010.30 / 12.902.17 / 1.9921.44 / 33.3332,910
PCC [67]3.03 / 1.672.13 / 1.200.53 / 0.407.08 / 3.7911.20 / 6.842.34 / 1.2716.82 / 12.330.378
RGP [18]4.59 / 4.563.13 / 2.810.43 / 0.7011.13 / 11.2718.79 / 20.253.53 / 3.3132.11 / 33.98-
CFCC [16]3.07 / 1.542.20 / 1.050.73 / 0.396.87 / 3.5512.55 / 6.982.36 / 1.1323.34 / 14.400.283
Ours (w/o n, w/o r)1.93 / 1.261.35 / 0.770.38 / 0.234.56 / 3.139.48 / 5.881.43 / 0.8822.63 / 15.904.83
Ours (w/o n, w/ r)1.89 / 1.231.18 / 0.790.32 / 0.244.74 / 3.0110.10 / 5.071.30 / 0.9024.99 / 12.914.93
Ours (w/ n, w/o r)1.87 / 1.201.24 / 0.720.37 / 0.244.58 / 2.999.82 / 5.581.30 / 0.8229.46 / 15.124.93
Ours (w/ n, w/ r)1.84 / 1.201.24 / 0.770.35 / 0.194.41 / 2.959.17 / 5.121.32 / 0.8735.42 / 12.715.03
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.792 + ], + "angle": 0, + "content": "We trained our model using the histogram feature \\(\\mathbf{H}\\), along with capture features including ISO \\((i)\\), exposure time \\((e)\\), and SNR stats \\((\\mathbf{r})\\), deliberately excluding the contextual metadata that is not available in the Simple Cube++ DSLR dataset. The results of our method, both with and without the SNR stats, are compared to other methods in Table 2. As shown, our method outperforms competing methods across most evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Inference time: As shown in Table 1, our method performs comparably to or outperforms prior methods, while maintaining a compact model with only approximately 5K parameters. This lightweight design results in faster runtimes compared to other methods that achieve competitive results, namely C4 [66] and tuned C5 [6]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.67, + 0.907, + 0.756 + ], + "angle": 0, + "content": "Table 2. Results on the testing set of the 'Simple Cube++' dataset [23]. We report the mean, median, best \\(25\\%\\), worst \\(25\\%\\), tri-mean, and maximum angular errors for each method. Symbols c, i, e, and r refer to the time-capture feature, ISO, exposure time, and SNR stats, respectively. The best and second-best results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.519, + 0.763, + 0.902, + 0.862 + ], + "angle": 0, + "content": "
MethodMeanMed.Best 25%Worst 25%Worst 5%Tri.Max
FFCC [11]1.380.580.184.0510.300.6845.20
FC4 [37]3.722.020.479.9317.182.4029.92
C4 [66]1.160.650.243.006.720.7313.95
C5 (tuned-CS) [6]1.190.600.183.267.610.6815.62
TLCC [59]1.821.150.394.499.631.2019.02
Ours (c = [i, e]T)1.080.590.172.845.530.6911.03
Ours (c = [i, e, rT]T)1.010.600.172.685.710.6510.89
" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5044" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.072, + 0.483, + 0.099 + ], + "angle": 0, + "content": "Table 3. Processing time on an AMD Ryzen Threadripper PRO 3975WX CPU460 and an NVIDIA RTX A6000 GPU." + }, + { + "type": "table", + "bbox": [ + 0.163, + 0.107, + 0.418, + 0.16 + ], + "angle": 0, + "content": "
MethodCPU (ms)GPU (ms)FLOPs
C4 [66]49.0311.282.28G
C5 [6]8.284.50103.54M
Ours0.550.4516.78M
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.483, + 0.311 + ], + "angle": 0, + "content": "In Table 3, we present the processing times of our model, C4, and C5 on an AMD Ryzen Threadripper PRO 3975WX CPU and an NVIDIA RTX A6000 GPU. Our fast runtime performance is particularly well-suited for mobile camera ISPs, where limited computational latency is crucial due to the processing demands of other modules (e.g., denoising, local tone mapping) per frame. Our model runs in just 0.25 ms on the DSP and 0.80 ms on the CPU of the Samsung S24 Ultra." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.483, + 0.528 + ], + "angle": 0, + "content": "Ablation studies: We conducted ablation studies to evaluate the impact of each input feature on the validation set of our dataset. Specifically, we assessed our method by excluding either the time-capture feature (c) or the histogram feature \\((\\mathbf{H})\\). Additionally, we tested the method using only the time feature \\((\\mathbf{p})\\) and noise stats \\((\\mathbf{n})\\), while excluding the histogram feature \\((\\mathbf{H})\\). We also evaluated the method using the histogram feature along with the time feature and noise stats, excluding other capture information. Furthermore, we examined the accuracy of our method using all input features except for \\(\\mathbf{p}\\). Lastly, we present the results when all input features are used, but without the noise stats \\((\\mathbf{n})\\) and the SNR stats \\((\\mathbf{r})\\). These results are presented in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.483, + 0.742 + ], + "angle": 0, + "content": "As shown in Table 4, using only the time feature \\((\\mathbf{p})\\) and noise stats \\((\\mathbf{n})\\) yields a reasonable accuracy \\((2.37^{\\circ}\\) mean angular error), compared to FFCC [11], which achieves \\(2.19^{\\circ}\\) mean angular error on the validation set. This validates the usefulness of time-of-day and noise information in providing contextual clues. As expected, incorporating the histogram feature \\((\\mathbf{H})\\), which represents scene colors, along with the time-of-day and noise stats, significantly boosts accuracy, as demonstrated by the results in the fourth row. We further investigate the impact of using noise information in the last three rows: using SNR stats \\((\\mathbf{r})\\), using noise stats \\((\\mathbf{n})\\), or both. The combination of both noise features yields the lowest mean angular error. Additional ablation studies are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.341, + 0.77 + ], + "angle": 0, + "content": "5. Conclusion and limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We presented a method for in-camera AWB correction that leverages contextual information. Specifically, we proposed a lightweight model that leverages contextual metadata (notably time-of-day derived from timestamp and geolocation) to guide the illuminant estimation process. In addition to this contextual metadata, we incorporate image colors in the form of histogram features, as well as capture information such as ISO, shutter speed, and noise stats, to help" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.905, + 0.215 + ], + "angle": 0, + "content": "Table 4. Results of ablation studies on the validation set with neutral white-balance ground truth. We report the mean, median, best \\(25\\%\\), and worst \\(25\\%\\) angular errors for our method with various configurations. Symbols \\(\\mathbf{H}\\), \\(\\mathbf{p}\\), \\(\\mathbf{n}\\), and \\(\\mathbf{r}\\) denote the histogram feature, time feature, noise stats, and SNR stats, respectively. The symbol \\(\\mathbf{m}\\) refers to the capture information feature, which includes ISO \\((i)\\), shutter speed \\((s)\\), and flash status \\((f)\\), all of which are used in our input time-capture feature \\(\\mathbf{c}\\). The best results are highlighted." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.217, + 0.897, + 0.359 + ], + "angle": 0, + "content": "
HmpnrMeanMed.Best 25%Worst 25%Worst 5%
✓XXXX2.031.360.314.908.90
X✓✓✓✓2.291.810.504.958.58
XX✓✓X2.371.590.455.489.03
✓X✓✓X1.751.170.294.137.26
✓✓X✓✓1.891.210.314.609.41
✓✓✓XX1.851.320.354.267.53
✓✓✓X✓1.721.160.304.138.16
✓✓✓✓X1.671.070.294.047.90
✓✓✓✓✓1.661.200.333.776.95
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.905, + 0.454 + ], + "angle": 0, + "content": "the model distinguish between artificial and natural scenes and improve its final accuracy. Our method is fast and can achieve high frame rates on modern smartphone DSPs and CPUs, while maintaining accurate illuminant color estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.906, + 0.626 + ], + "angle": 0, + "content": "A key contribution of this work is a large-scale dataset of raw images captured by a consumer smartphone camera, along with the necessary contextual metadata for training and evaluating our method. Beyond the traditional neutral white-balance ground truth extracted from a calibration color chart placed in each scene, we also include a user-preference ground truth that targets the observer's preference, validated through a user study. Results based on both ground truth types demonstrate that our method achieves comparable or superior performance to existing methods, which require larger models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.901 + ], + "angle": 0, + "content": "While our method represents a promising solution for mobile camera ISPs, its dependency on contextual metadata limits its optimal accuracy to devices that provide geolocation data, which may not be available on DSLR cameras. Additionally, while the contextual metadata is device-independent (i.e., the differences across devices are expected to be minimal), our method relies on additional capture information (i.e., ISO, shutter speed, noise, and SNR stats) and image colors, making it inherently camera-dependent in design. This dependency prevents our trained model from generalizing to new devices without fine-tuning or re-training. However, this issue can be mitigated through color calibration (e.g., [46]), which can be extended to calibrate capture information as well—by performing a preprocessing mapping from the new camera space (for both color and capture information) to the camera space used during training. Additional discussion on cross-camera generalization is provided in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5045" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Abdelrahman Abdelhamed, Abhijith Punnappurath, and Michael S Brown. Leveraging the availability of two cameras for illuminant estimation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.199 + ], + "angle": 0, + "content": "[2] Mahmoud Affifi and Michael S Brown. Sensor-independent illumination estimation for DNN models. In BMVC, 2019. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.482, + 0.228 + ], + "angle": 0, + "content": "[3] Mahmoud Affifi and Michael S Brown. Deep white-balance editing. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[4] Mahmoud Affifi, Brian Price, Scott Cohen, and Michael S Brown. When color constancy goes wrong: Correcting improperly white-balanced images. In CVPR, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[5] Mahmoud Affifi, Abhijith Punnappurath, Graham Finlayson, and Michael S Brown. As-projective-as-possible bias correction for illumination estimation algorithms. JOSA A, 36 (1):71-78, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[6] Mahmoud Afifi, Jonathan T Barron, Chloe LeGendre, Yun-Ta Tsai, and Francois Bleibel. Cross-camera convolutional color constancy. In ICCV, 2021. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[7] Mahmoud Affifi, Zhenhua Hu, and Liang Liang. Optimizing illuminant estimation in dual-exposure HDR imaging. In ECCV, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[8] Vivek Agarwal, Besma R Abidi, Andreas Koschan, and Mongi A Abidi. An overview of color constancy algorithms. Journal of Pattern Recognition Research, 1(1):42-54, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.482, + 0.512 + ], + "angle": 0, + "content": "[9] Relja Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In CVPR, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.516, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[10] Jonathan T Barron. Convolutional color constancy. In ICCV, 2015. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.545, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[11] Jonathan T Barron and Yun-Ta Tsai. Fast Fourier color constancy. In CVPR, 2017. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.574, + 0.482, + 0.6 + ], + "angle": 0, + "content": "[12] Simone Bianco and Marco Buzzelli. Truncated edge-based color constancy. In ICCE, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[13] Simone Bianco and Claudio Cusano. Quasi-unsupervised color constancy. In CVPR, 2019. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.632, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[14] David H Brainard and Brian A Wandell. Analysis of the Retinex theory of color vision. Journal of the Optical Society of America A, 3(10):1651-1661, 1986. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.675, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[15] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin Institute, 310(1): 1-26, 1980. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[16] Marco Buzzelli and Simone Bianco. A convolutional framework for color constancy. IEEE Transactions on Neural Networks and Learning Systems, pages 1-15, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[17] Cheng Cheng, Kai-Fu Yang, Xue-Mei Wan, Leanne Lai Hang Chan, and Yong-Jie Li. Nighttime color constancy using robust gray pixels. JOSA A, 41(3):476-488, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[18] Cheng Cheng, Kai Fu Yang, Xue Mei Wan, Leanne Lai Hang Chan, and Yonge Jie Li. Nighttime color constancy using robust gray pixels. JOsA A, 41(3):476-488, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.847, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[19] Dongliang Cheng, Dilip K Prasad, and Michael S Brown. Illuminant estimation for color constancy: Why spatial-domain methods work and the role of the color distribution. JOSA A, 31(5):1049-1058, 2014. 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[20] Kyungah Choi and Hyeon-Jeong Suk. User-preferred color temperature adjustment for smartphone display under varying illuminants. Optical Engineering, 53(6):061708-061708, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.906, + 0.191 + ], + "angle": 0, + "content": "[21] Djork-Arné Clevert. Fast and accurate deep network learning by exponential linear units (ELUs). arXiv preprint arXiv:1511.07289, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.195, + 0.906, + 0.247 + ], + "angle": 0, + "content": "[22] Maurizio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Mobile computational photography: A tour. Annual Review of Vision Science, 7(1):571-604, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.252, + 0.906, + 0.32 + ], + "angle": 0, + "content": "[23] Egor Ershov, Alexey Savchik, Illya Semenkov, Nikola Banić, Alexander Belokopytov, Daria Senshina, Karlo Koščević, Marko Subašić, and Sven Lončarić. The cube++ illumination estimation dataset. IEEE Access, 8:227511-227527, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.324, + 0.906, + 0.391 + ], + "angle": 0, + "content": "[24] Egor Ershov, Alex Savchik, Denis Shepelev, Nikola Banić, Michael S Brown, Radu Timofte, Karlo Koščević, Michael Freeman, Vasily Tesalin, Dmitry Bocharov, et al. NTIRE 2022 challenge on night photography rendering. In CVPRW, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.395, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[25] Graham D Finlayson and Steven D Hordley. Color constancy at a pixel. JOSA A, 18(2):253-264, 2001. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.426, + 0.905, + 0.452 + ], + "angle": 0, + "content": "[26] Graham D Finlayson and Elisabetta Trezzi. Shades of gray and colour constancy. In CIC, 2004. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[27] Karl R Gegenfurtner, David Weiss, and Marina Bloj. Color constancy in real-world settings. Journal of Vision, 24(2): 12-12, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.499, + 0.905, + 0.54 + ], + "angle": 0, + "content": "[28] Peter Vincent Gehler, Carsten Rother, Andrew Blake, Tom Minka, and Toby Sharp. Bayesian color constancy revisited. In CVPR, 2008. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.543, + 0.905, + 0.597 + ], + "angle": 0, + "content": "[29] Arjan Gijsenij and Theo Gevers. Color constancy using natural image statistics and scene semantics. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(4): 687-698, 2011. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.601, + 0.905, + 0.656 + ], + "angle": 0, + "content": "[30] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Generalized gamut mapping using image derivative structures for color constancy. International Journal of Computer Vision, 86(2):127-139, 2010. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.659, + 0.905, + 0.712 + ], + "angle": 0, + "content": "[31] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Computational color constancy: Survey and experiments. IEEE Transactions on Image Processing, 20(9):2475-2489, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.716, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[32] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):918-929, 2011. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[33] Thorsten Hansen, Sebastian Walter, and Karl R Gegenfurtner. Effects of spatial and temporal context on color categories and color constancy. Journal of Vision, 7(4):2-2, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[34] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics, 35(6):1-12, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5046" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[35] Xuanhua He, Tao Hu, Guoli Wang, Zejin Wang, Run Wang, Qian Zhang, Keyu Yan, Ziyi Chen, Rui Li, Chengjun Xie, et al. Enhancing RAW-to-sRGB with decoupled style structure in Fourier domain. In AAAI, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.178 + ], + "angle": 0, + "content": "[36] Steven D Hordley and Graham D Finlayson. Re-evaluating colour constancy algorithms. In ICPR, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.482, + 0.221 + ], + "angle": 0, + "content": "[37] Yuanming Hu, Baoyuan Wang, and Stephen Lin. FC4: Fully convolutional color constancy with confidence-weighted pooling. In CVPR, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.224, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[38] Andrey Ignatov, Anastasia Sycheva, Radu Timofte, Yu Tseng, Yu-Syuan Xu, Po-Hsiang Yu, Cheng-Ming Chiang, Hsien-Kai Kuo, Min-Hung Chen, Chia-Ming Cheng, et al. MicroISP: processing 32MP photos on mobile devices with deep learning. In ECCV, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[39] Woohyeok Kim, Geonu Kim, Junyong Lee, Seungyong Lee, Seung-Hwan Baek, and Sunghyun Cho. ParamISP: Learned forward and inverse ISPs using camera parameters. In CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.352, + 0.482, + 0.393 + ], + "angle": 0, + "content": "[40] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.396, + 0.482, + 0.464 + ], + "angle": 0, + "content": "[41] Firas Laakom, Nikolaos Passalis, Jenni Raitoharju, Jarno Nikkanen, Anastasios Tefas, Alexandros Iosifidis, and Moncef Gabbouj. Bag of color features for color constancy. IEEE Transactions on Image Processing, 29:7722-7734, 2020. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.467, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[42] Firas Laakom, Jenni Raitoharju, Jarno Nikkanen, Alexandros Iosifidis, and Moncef Gabbouj. Robust channel-wise illumination estimation. In BMVC, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[43] Firas Laakom, Jenni Raitoharju, Jarno Nikkanen, Alexandros Iosifidis, and Moncef Gabbouj. Intel-TAU: A color constancy dataset. IEEE Access, 9:39560-39567, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[44] Shuwei Li and Robby T Tan. NightCC: Nighttime color constancy via adaptive channel masking. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.584, + 0.482, + 0.652 + ], + "angle": 0, + "content": "[45] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM Transactions on Graphics, 38(6):164-1, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.655, + 0.482, + 0.696 + ], + "angle": 0, + "content": "[46] Yi-Tun Lin, Bianjiang Yang, Hao Xie, Wenbin Wang, Honghong Peng, and Jun Hu. Color constancy: How to deal with camera bias? In BMVC, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.699, + 0.482, + 0.738 + ], + "angle": 0, + "content": "[47] Ce Liu, William T Freeman, Richard Szeliski, and Sing Bing Kang. Noise estimation from a single image. In CVPR, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.742, + 0.482, + 0.783 + ], + "angle": 0, + "content": "[48] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.786, + 0.482, + 0.826 + ], + "angle": 0, + "content": "[49] Zhongyu Lou, Theo Gevers, Ninghang Hu, Marcel P Lucassen, et al. Color constancy by deep learning. In BMVC, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[50] Ming Luo. A review of chromatic adaptation transforms. Review of Progress in Coloration and Related Topics, 30:77 - 92, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[51] Laurence T Maloney. Illuminant estimation as cue combination. Journal of Vision, 2(6):6-6, 2002. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[52] Jean H Meeus. *Astronomical algorithms*. Willmann-Bell, Incorporated, 1991. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[53] Seoung Wug Oh and Seon Joo Kim. Approaching the computational color constancy as a classification problem through deep learning. Pattern Recognition, 61:405-416, 2017. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[54] Varad A Pimpalkhute, Rutvik Page, Ashwin Kothari, Kishor M Bhurchandi, and Vipin Milind Kamble. Digital image noise estimation using DWT coefficients. IEEE Transactions on Image Processing, 30:1962-1972, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[55] Stanislav Pyatykh, Jürgen Hesser, and Lei Zheng. Image noise level estimation by principal component analysis. IEEE Transactions on Image Processing, 22(2):687-699, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[56] Yanlin Qian, Joni-Kristian Kamarainen, Jarno Nikkanen, and Jiri Matas. On finding gray pixels. In CVPR, 2019. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[57] Yanlin Qian, Said Pertuz, Jarno Nikkanen, Joni-Kristian Kämäräinen, and Jiri Matas. Revisiting gray pixel for statistical illumination estimation. In VISSAP. 2019. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[58] Wu Shi, Chen Change Loy, and Xiaou Tang. Deep specialized network for illuminant estimation. In ECCV, pages 371-387, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[59] Yuxiang Tang, Xuejing Kang, Chunxiao Li, Zhaowen Lin, and Anlong Ming. Transfer learning for color constancy via statistic perspective. In AAAI, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[60] Shoji Tominaga, Takahiko Horiuchi, Shiori Nakajima, and Mariko Yano. Prediction of incomplete chromatic adaptation under illuminant A from images. In CIC, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[61] Oguzhan Ulucan, Diclehan Ulucan, and Marc Ebner. Color constancy beyond standard illuminants. In ICIP, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.561 + ], + "angle": 0, + "content": "[62] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy. IEEE Transactions on Image Processing, 16(9):2207-2214, 2007. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.564, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[63] TC Van Flandern and KF Pulkkinen. Low-precision formulae for planetary positions. Astrophysical Journal Supplement Series, 41:391-411, 1979. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.607, + 0.905, + 0.633 + ], + "angle": 0, + "content": "[64] Robert Walraven. Calculating the position of the sun. Solar energy, 20(5):393-397, 1978. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.635, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[65] Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible image signal processing. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[66] Huanglin Yu, Ke Chen, Kaiqi Wang, Yanlin Qian, Zhaoxiang Zhang, and Kui Jia. Cascading convolutional color constancy. In AAAI, 2020. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[67] Shuwei Yue and Minchen Wei. Color constancy from a pure color view. Journal of the Optical Society of America A, 40 (3):602-610, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.79 + ], + "angle": 0, + "content": "[68] Yi Zhang, Hongwei Qin, Xiaogang Wang, and Hongsheng Li. Rethinking noise synthesis and modeling in raw denoising. In ICCV, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5047" + } + ] +] \ No newline at end of file diff --git a/2025/Time-Aware Auto White Balance in Mobile Photography/59296f3d-511f-4955-b6aa-60113b360f05_origin.pdf b/2025/Time-Aware Auto White Balance in Mobile Photography/59296f3d-511f-4955-b6aa-60113b360f05_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f0d8f6ab27d19f79e90daa894fc3f4670209e234 --- /dev/null +++ b/2025/Time-Aware Auto White Balance in Mobile Photography/59296f3d-511f-4955-b6aa-60113b360f05_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0825ab2b91ca914681d47ea659ce645f64f91eb826eb510eefc4250d013e9ee +size 2936607 diff --git a/2025/Time-Aware Auto White Balance in Mobile Photography/full.md b/2025/Time-Aware Auto White Balance in Mobile Photography/full.md new file mode 100644 index 0000000000000000000000000000000000000000..02cf43772657be01ccad0d382bc7275874f7f158 --- /dev/null +++ b/2025/Time-Aware Auto White Balance in Mobile Photography/full.md @@ -0,0 +1,327 @@ +# Time-Aware Auto White Balance in Mobile Photography + +Mahmoud Affifi* + +Luxi Zhao* + +Abhijith Punnappurath + +Mohammed A. Abdelsalam + +Ran Zhang Michael S. Brown + +AI Center-Toronto, Samsung Electronics + +{m.afifil, lucy.zhao, abhijith.p, m.abdelsalam, ran.zhang, michael.b1}@samsung.com + +![](images/7fc4c1df2b7ca5e35050456c912a700f35890770c4ff1ee1c02be792a3fb1784.jpg) + +![](images/20de2c620390972ff2c6eaea908e4ac94af663fb2c890494f3aef6ca9751b23c.jpg) +Figure 1. The time of day influences scene illumination, making it a valuable cue for improving illuminant estimation. Shown are white-balanced results (gamma-corrected for visualization) using our method with (1) colors only, (2) colors plus contextual (timestamp and geolocation) and capture data, and (3) ground truth (from a color chart). Angular errors show improvements when time-capture information is used. + +![](images/34933caa09610af20c0f945853961c95383be92bbb7d66865e91ff02329cd952.jpg) + +![](images/494610d5708b103912b8b3e91058b1c46fa8e0035839c2e65549f75d07e62175.jpg) + +![](images/0c0110915450ac8e4be1ea77794ccf5a754fce231ba0fdd1491162392c00be3b.jpg) + +![](images/804e0117b967e293d9741042ea051a11d7801395b6e4320f7e6ecfd0d683aa6f.jpg) + +![](images/f4234c86efb8c3db926f7d3948c6430d4f02764cf7391275e0dbe93abd58933c.jpg) + +![](images/890626b0b999577fb0a45a5a233fea3a624a3e705b69a7695a4c09f0cee8444f.jpg) + +![](images/7ab03fe008aea03315fc8bdcb39fa05d8d813d3261cbf357c35aab9c1ba117e5.jpg) + +![](images/4c8e35235cab070dd208c92e266b5e941db2ca4d6881db3c51cb5e93dabefb27.jpg) + +![](images/6b75febfbbb93d44c1e182766884ab821dc57151fc76fc0c04a60cc4f9a1100d.jpg) + +![](images/d2bbeef182343bfae2d9943f268bc9fc8e9c603a81f1c3e213ebe50da9ee833a.jpg) + +![](images/6e1502e8ef95c143df89633f3b5cf4f114bf3ff14e1af5a61e699f95bd8d5952.jpg) + +# Abstract + +Cameras rely on auto white balance (AWB) to correct undesirable color casts caused by scene illumination and the camera's spectral sensitivity. This is typically achieved using an illuminant estimator that determines the global color cast solely from the color information in the camera's raw sensor image. Mobile devices provide valuable additional metadata—such as capture timestamp and geolocation—that offers strong contextual clues to help narrow down the possible illumination solutions. This paper proposes a lightweight illuminant estimation method that incorporates such contextual metadata, along with additional capture information and image colors, into a lightweight model ( $\sim 5K$ parameters), achieving promising results, matching or surpassing larger models. To validate our method, we introduce a dataset of 3,224 smartphone images with contextual metadata collected at various times of day and under diverse lighting conditions. The dataset includes ground-truth illuminant colors, determined using a color chart, and user-preferred illuminants validated through a user study, providing a comprehensive benchmark for AWB evaluation. + +# 1. Introduction and related work + +Color constancy refers to the ability of the human visual system to maintain stable object colors despite variations in lighting conditions by leveraging contextual cues within the scene [27, 33, 51]. Cameras approximate this effect using auto white balance (AWB) correction, which aims to partially neutralize color casts introduced by scene illumination and the camera's spectral sensitivity [3]. AWB first estimates the illumination color as an RGB vector in the camera's raw color space. The raw image is then corrected by scaling its color channels according to the estimated illumination, typically under the assumption of a single global light source [8, 31]. + +Conventional illuminant estimation methods primarily rely on image colors, either by directly processing the raw image (e.g., [13, 32, 41, 49, 53, 58, 62, 66]) or by analyzing color histograms (e.g., [2, 6, 7, 10]). These methods can be broadly categorized into two groups: (1) classical statistical-based methods (e.g., [15, 26, 32, 56, 57, 62]), + +![](images/f233ecfc8dca891a11308e394d6b4a4823f290d025912e67d087ddffdabfae92.jpg) +Figure 2. The time of day at which an image is captured provides valuable information about the possible range of illuminant colors in outdoor scenes. The figure presents the correlated color temperature (CCT) of illuminant colors in our dataset (Sec. 3) for images captured at various times throughout the day and night. As shown, excluding images taken under artificial light, those captured at noon, for example, exhibit a different range of illuminant CCTs compared to images captured during sunset or sunrise. + +which estimate the illuminant color based on image statistics, and (2) learning-based methods (e.g., [6, 10, 30, 42, 59]), which map image colors to their corresponding scene illuminant through data-driven models. + +While image colors are a key cue for estimating the scene's illuminant, mobile devices provide an opportunity to integrate additional contextual information. For instance, the device's location, along with the date and time, offer valuable cues about outdoor lighting conditions (e.g., sunrise, noon, sunset), thereby improving illuminant estimation for outdoor scenes (see Fig. 1). + +Intuitively, knowing the time of day when an outdoor scene is captured can help estimate the lighting conditions. Figure 2 further illustrates that the time of day, derived from contextual metadata (i.e., timestamp and geolocation) readily available on mobile devices, provides valuable insights into the likely range of illuminant correlated color temperature (CCT) in outdoor scenes. This information serves to narrow the range of possible illuminant colors. For instance, images taken at noon exhibit a different CCT range than those captured at sunrise or sunset. When combined with additional capture information to distinguish between environments (e.g., indoors vs. outdoors), such metadata can complement image colors to improve illuminant estimation accuracy. + +Despite its potential, only a few attempts have explored leveraging additional information available to the camera's + +image signal processor (ISP) for illuminant estimation, such as metadata-based model control [11] or data augmentation to enhance generalization [6]. However, to our knowledge, no previous work has investigated using contextual information—specifically, mobile device timestamps and geolocation—to refine illuminant estimation. + +We introduce a method that leverages contextual metadata from mobile phones, along with additional capture information available in camera ISPs, to train a lightweight illuminant estimator model with $\sim 5\mathrm{K}$ parameters. Our model delivers promising results, matching or surpassing that of larger models, while maintaining efficiency. Our model runs on a typical flagship mobile digital signal processor (DSP) and CPU in $0.25\mathrm{ms}$ and $0.80\mathrm{ms}$ , respectively. This compact and efficient design is especially advantageous for mobile devices, where minimizing power consumption and memory usage is critical [1, 7, 11]. + +Existing white-balance datasets (e.g., [19, 23, 28, 43]) lack the contextual information needed to validate our method. The absence of contextual information is because most available datasets (e.g., [19, 23]) used DSLR cameras, which typically lack built-in GPS functionality or accurate timestamps. To address this, we captured a new dataset of 3,224 images using a consumer smartphone camera, accompanied by contextual and capture information. The ground truth for the dataset was established using two approaches: (1) the conventional approach, selecting gray patches from a calibration color chart, and (2) a manually selected user-preference white-balance target, which enhances the image's aesthetic appeal and accounts for incomplete chromatic adaptation [60]. The user-preference white-balance "ground truth" was validated through a user study. The dataset spans a variety of lighting conditions, including non-standard artificial illuminants [61], and covers different times of day—sunrise, noon, sunset, and night. + +Contribution: We propose an AWB method that utilizes smartphone contextual metadata (i.e., timestamp and geolocation) alongside capture information to enhance illuminant estimation. We demonstrate that integrating this additional data with conventional color information into a lightweight neural model leads to significant improvements in accuracy. Additionally, we introduce a large-scale dataset captured with a consumer smartphone at various times of day, with ground truth derived from both a calibration color chart and user-preference-based white balance. Benchmarking results show that our method achieves state-of-the-art results in most standard metrics (i.e., angular error statistics [19]), while maintaining computational efficiency on mobile devices. + +![](images/1591a830cb7446514d9e26f8a58a76bccbdae499c6a48f3d34c47ded2a49bdbd.jpg) +Figure 3. Our method includes a lightweight model consisting of a convolutional network that processes the histogram feature (derived from raw image colors and edge histograms concatenated with the $u / v$ coordinates) to produce a latent feature $\mathbf{v}_h\in \mathbb{R}^{16}$ . This is then concatenated with the latent feature of the processed time-capture feature, $\mathbf{v}_t\in \mathbb{R}^{16}$ . The combined feature, $\mathbf{v}\in \mathbb{R}^{32}$ , is passed through a lightweight MLP to produce the chromaticity of the scene illuminant, $\ell_c\in \mathbb{R}^2$ , which is finally converted into normalized RGB illuminant color. + +# 2. Method + +An overview of our method is shown in Fig. 3. Our method employs a learnable model that processes two distinct inputs: (1) the time-capture feature (Sec. 2.1), which combines the contextual and capture information available on mobile devices and accessible by their camera ISPs, and (2) the histogram feature (Sec. 2.2), which represents the image's R/G and B/G chromaticity values. The time-capture feature is first processed to project it into a latent space, producing the time-capture latent feature vector, $\mathbf{v}_t \in \mathbb{R}^{16}$ . The model processes the histogram feature to produce the histogram's latent feature vector, $\mathbf{v}_h \in \mathbb{R}^{16}$ . Both the histogram and time-capture latent feature vectors are concatenated and processed by a set of learnable layers to output an R/G, B/G chromaticity vector, $\ell_c \in \mathbb{R}^2$ , representing the scene illuminant. This 2D vector is then converted into the illuminant RGB color used to perform white balancing. + +# 2.1. Time-capture feature + +Our model incorporates contextual metadata, available on mobile devices, as one of its input features. Specifically, we use the geolocation and timestamp of image capture to compute the "probability" of the time of day (e.g., sunset, noon, etc.). Our approach converts conventional clock time (hour::minutes) to its corresponding time of day condition based on the date and geolocation of the scene. This approach allows the model to generalize across different time zones and prevents it from being influenced by the capturing location. The time probability vector represents the likelihood that the captured image corresponds to one of the solar event times (i.e., dawn, sunrise, noon, sunset, dusk, or mid + +night) and is computed as follows: + +$$ +p _ {g} = 1 - \frac {\left| t _ {c} - t _ {g} \right|}{t _ {s}}, \tag {1} +$$ + +where $t_c$ is the capture time in seconds (adjusted to the local time zone based on geolocation information), and $t_g$ is the local time in seconds of the solar event, $g \in \{\text{dawn,sunrise,noon,sunset,dusk,midnight}\}$ , computed using geolocation-based standard algorithms [52, 63, 64]. The scalar $t_s$ represents the total number of seconds in a day (i.e., 86,400). + +We pre-process the probability of each solar event, $p_{g}$ , in our time probability vector by computing the square root to enhance feature representation—compressing high probabilities and amplifying lower ones, which is intended to help create a more balanced and smooth distribution for the model to leverage. We then augment this time probability vector with a one-hot vector, $\mathbf{b}$ , that indicates whether the capture time $t_{c}$ occurs before each solar event. This distinction helps the model account for expected variations in CCT, as illuminant colors can differ before and after certain solar events, such as sunset and sunrise. The value of this one-hot vector for a given solar event $g$ is computed as: + +$$ +\mathbf {b} _ {g} = \left\{ \begin{array}{l l} 1, & \text {i f} t _ {c} \leq t _ {g} \\ 0, & \text {o t h e r w i s e}, \end{array} \right. \tag {2} +$$ + +where $\mathbf{b}_g$ corresponds to the entry in the one-hot vector for the solar event, $g$ . Both the time probability vector and the one-hot vector together form our time feature, $\mathbf{p} \in \mathbb{R}^{12}$ . + +To enrich our feature set with additional capture information available from the camera ISP and help distinguish the capturing environment (e.g., indoor vs. outdoor), we include the following features in our final time-capture feature, c: + +- ISO (i): The sensitivity of the camera's image sensor to light, where lower values indicate good lighting conditions (e.g., bright scenes) and higher values suggest low-light environments, such as poorly lit indoor scenes. +- Shutter speed $(s)$ : The amount of time the camera's shutter remains open, allowing light to hit the image sensor. It provides an indication of lighting conditions, alongside the ISO value, $i$ . +- Flash status $(f)$ : A binary value indicating whether flash light was used during capturing. +- Noise information (optional): Since image denoising is typically applied before or in parallel with illuminant estimation in camera ISPs [22, 34, 45], we also explore the optional use of explicit noise information from the captured scene. More noise typically indicates low-light conditions, which can provide clues about the lighting color range of the scene. Unfortunately, while noise information is accessible within camera ISPs, obtaining accurate + +noise information for public use is challenging, as the noise profiles in DNG files are not always reliable [68]. To address this, we simulate the noise information using two approaches: noise statistics (stats) and/or signal-to-noise ratio (SNR) stats, which are described below. + +Noise stats (n): This represents the noise statistics in the captured raw image. We simulate this by denoising each raw image using Adobe Lighthroom and computing the noise stats as the mean and standard deviation of each color channel of the absolute difference between the denoised and noisy raw images. This approach provides a simplified method for estimating noise stats, as the denoised images are typically available within the camera ISPs, but difficult to extract from the DNG files. + +SNR stats (r): This alternative approach measures the noise information in the captured image without the need of a denoised reference. The SNR is computed by applying a $15 \times 15$ sliding window over the raw image and calculated as: $10\log_{10}\left(\frac{\mu}{\sigma + \epsilon}\right)$ , where $\mu$ represents the mean RGB of the $15 \times 15$ patch, $\sigma$ is the standard deviation, and $\epsilon$ is a small value added for numerical stability. + +Note that while these two approaches—namely, computing noise stats and SNR stats—yield satisfactory results, there is a wide body of research on noise estimation (e.g., [47, 54, 55]), which is beyond the scope of this paper. + +Our complete time-capture feature is the combination of these inputs and can be expressed as follows: + +$$ +\mathbf {c} = \left[ \mathbf {p} ^ {T}, \log (i), \log (s), f, \mathbf {w} ^ {T} \right] ^ {T}, \tag {3} +$$ + +where $\mathbf{w}$ represents optional noise-related features, which can include either noise stats $\mathbf{n}$ , SNR stats $\mathbf{r}$ , or both. Alternatively, $\mathbf{w}$ can be omitted if noise information is not considered. + +Our time-capture feature is first normalized using min-max normalization (with the min and max values computed from the training data), and then processed through a learnable function, $f_{t}$ , as follows: + +$$ +\mathbf {v} _ {t} = f _ {t} (\mathbf {c}), \tag {4} +$$ + +where $f_{t}$ is a learnable linear layer that transforms the time-capture feature, $\mathbf{c}$ , into its latent representation, $\mathbf{v}_t\in \mathbb{R}^{16}$ (see Fig. 3). + +# 2.2. Histogram feature + +In addition to the time-capture feature, c, we provide our neural model with a histogram feature, $\mathbf{H}$ which represents the colors of the raw image. Inspired by prior work [6, 10, 11], we use a 2D histogram to represent the R/G and B/G chromaticities of the input raw image, $\mathbf{I} \in \mathbb{R}^{K \times 3}$ , where $K$ denotes the total number of pixels in the raw im + +age. Specifically, we compute the 2D chromaticity histogram, $\mathbf{H}_c\in \mathbb{R}^{h\times h}$ , as follows: + +$$ +\mathbf {H} _ {c} ^ {(m, n)} = \sum_ {k} \| \mathbf {I} ^ {(k)} \| _ {2} \cdot \delta_ {m, n} ^ {(k)}, \tag {5} +$$ + +$$ +\delta_ {m, n} ^ {(k)} = \left[ u _ {m} \leq r g ^ {(k)} < u _ {m + 1} \right] \wedge \left[ v _ {n} \leq b g ^ {(k)} < v _ {n + 1} \right], \tag {6} +$$ + +where $rg^{(k)}$ and $bg^{(k)}$ are the R/G and B/G chromaticity values of pixel $k$ in I, and $\| \mathbf{I}^{(k)} \|_2$ represents the intensity of pixel $k$ (i.e., Euclidean norm of the pixel's RGB values). The notation $\wedge$ denotes the logical AND operator, and $[\cdot]$ is the Iverson bracket, which evaluates to 1 when the condition is true and 0 otherwise. The histogram bins are defined by the edges $\{u_m\}$ and $\{v_n\}$ , where $u_m$ and $v_n$ are the lower edges of the bins, and $u_{m+1}$ and $v_{n+1}$ are the corresponding upper edges. The histogram accumulates the brightness values of pixels whose chromaticity values fall within the range $[u_m, u_{m+1})$ along the R/G axis (i.e., the $u$ -axis) and $[v_n, v_{n+1})$ along the B/G axis (i.e., the $v$ -axis). The number of histogram bins along each axis is denoted as $h$ . Following [10], we compute the square root of the histogram to enhance the utility of the histogram feature [9]. + +Note that this histogram differs from the $\log -uv$ histogram used in prior work [6, 10, 11], which operates in the logarithmic space of G/R and G/B [25]. We found that the R/G, B/G chromaticity histogram performs better, as it aligns with our model's output space. See the supplemental material for an ablation study. + +In addition to the chromaticity histogram, $\mathbf{H}_c$ , of the image colors, and building on prior work [6, 11], we augment it with the square-rooted chromaticity histogram of the image's edges, $\mathbf{H}_e$ , where the image edges are computed as follows: + +$$ +\mathbf {E} ^ {(x, y)} = \frac {1}{8} \sum_ {\Delta x, \Delta y} \left| \mathbf {I} ^ {\prime (x, y)} - \mathbf {I} ^ {\prime (x + \Delta x, y + \Delta y)} \right|, \tag {7} +$$ + +where $\mathbf{E}$ represents the image edges, $\mathbf{I}'$ refers to the raw image in 3D tensor form (height, width, channels), and $\Delta x, \Delta y \in \{-1,0,1\}$ with $(\Delta x, \Delta y) \neq (0,0)$ . The edge histogram, $\mathbf{H}_e$ , is computed from $\mathbf{E}$ using Eqs. 5 and 6. + +Our histogram feature is constructed by concatenating these two histograms, $\mathbf{H}_c$ and $\mathbf{H}_e$ . Since this histogram feature is first processed by convolutional (conv) layers, as shown in Fig. 3, we follow [6] by appending additional channels that encode the positional information of the $u / v$ coordinates in histogram space, which helps capture spatial relationships within the histogram feature. Consequently, our final histogram feature, $\mathbf{H} \in \mathbb{R}^{h \times h \times 4}$ , consists of 2 channels representing the chromaticity of the image and its edges, along with the additional $u / v$ coordinate channels. + +This feature is processed through a series of conv layers with ELU activation [21]. The resulting latent representation undergoes adaptive average pooling before passing through a linear layer to produce the histogram's latent feature vector, $\mathbf{v}_h\in \mathbb{R}^{16}$ , as follows: + +$$ +\mathbf {v} _ {h} = f _ {h} (\mathbf {H}), \tag {8} +$$ + +where $f_{h}$ denotes the sub-model (i.e., conv layers, ELU activation, pooling, and linear layer) that maps the histogram feature into its latent space, as shown in Fig. 3. + +# 2.3. Illuminant estimation + +Both feature vectors, $\mathbf{v}_t$ and $\mathbf{v}_h$ , are concatenated to produce the latent vector $\mathbf{v}$ and processed by an illuminant estimation sub-model as follows: + +$$ +\mathbf {v} = \left[ \mathbf {v} _ {t}; \mathbf {v} _ {h} \right], \tag {9} +$$ + +$$ +\ell_ {c} = f _ {\ell} (\mathbf {v}), \tag {10} +$$ + +where $f_{\ell}$ consists of a set of linear layers, with batch normalization applied to the first layer, followed by activation functions—except for the final layer, which outputs $\ell_{c}$ , the chromaticity vector of the scene illuminant. This vector is then transformed into an unnormalized RGB illuminant color by mapping $[\mathrm{R / G},\mathrm{B / G}]^T\rightarrow [\mathrm{R / G},1,\mathrm{B / G}]^T$ , followed by normalization via division by its L2 norm to produce the final illuminant color. We optimize $f_{t}$ , $f_{h}$ , and $f_{\ell}$ to minimize the angular error [36] between the predicted RGB illuminant color and the ground-truth RGB illuminant color. + +# 3. Dataset + +To train and validate our method, we require a dataset that includes contextual information (i.e., timestamp and geolocation) h h image. Existing datasets (e.g., [17, 19, 23, 28, 43, 44]) lack this essential information, motivating us to collect a new dataset using a smartphone camera that provides contextual metadata for each image. Specifically, we captured 3,224 linear raw images with the Samsung S24 Ultra's main camera, covering a wide range of scenes both indoors and outdoors, at various times of day (e.g., sunset, sunrise, noon, night). Our dataset includes images captured under various light sources (e.g., sunlight, incandescent, LED), as well as non-standard illuminant colors (e.g., colored LED light), which are not present in existing datasets [61]; see Fig. 4. Additionally, our dataset captures scenes under different weather conditions (sunny, cloudy, rain, snow, etc.). Example scenes can be found in the supplemental material. + +We follow prior work [7, 43] in collecting ground-truth illuminant colors for each scene. Specifically, for each scene, we first capture an image with a calibration color chart, which is used to extract the illuminant color from the + +![](images/37f97bafb85ba114e5ec6509bbbf0db7564b7396fd3d3680416eea20a931d984.jpg) +Figure 4. A plot of the $rg$ chromaticity distribution of ground-truth illuminant colors for neutral white balance in our dataset. Example images in raw and sRGB spaces are shown (raw images are gamma-corrected for visualization). See supplemental material the $rg$ chromaticity distribution of user-preference illuminants. + +gray patches. Next, we capture an image of the same scene without the color chart (see Fig. 5), resulting in $\sim 6\mathrm{K}$ images. After obtaining the ground-truth illuminant color, we discard all color chart images, leaving us with 3,224 images in our dataset. This approach allows us to test with natural images that mimic real-world scenarios, without the need to mask out the color chart patch. + +Since our dataset includes a wide variety of lighting conditions, such as sunset/sunrise, night scenes, and artificial light, we generate a "user-preference" ground truth in addition to the neutral ground truth obtained from the color chart for each scene in our dataset. This is to account for human incomplete chromatic adaptation in such scenes [50, 60] and user preference [20, 24]. Specifically, an expert photographer was asked to assign a ground truth illuminant to each scene to make it appear more natural, reflect real-world observations, and enhance the aesthetics of the image. Notably, the same person who captured the scene also performed the annotation, ensuring that the user-preference selection was based on real-world observations of how the scene should appear. + +We validated the user-preference ground truth through a study with 20 participants, who selected the most natural image from pairs of white-balanced images corresponding to the user-preference ground truth and the neutral ground truth (obtained from the color chart). The user-preference ground truth was selected in $71.95\%$ of the trials, confirming its preference. See the supplemental material for additional details. In our experiments, both ground-truth types—neutral and user-preference—were used for training and evaluation (Sec. 4). + +![](images/5cbfde5f3f2a93b14d0a501e0f559052dccf4bed8a822c976be0457aad762bfe.jpg) +Figure 5. For each scene in our dataset, an image with a color chart placed in the scene was captured. The gray patches on the color chart were used to measure the ground-truth illuminant. These color-chart images were discarded, and only the images of the scenes without the color chart were used for the training, validation, and testing sets. For each example, we show both raw and sRGB images, with the raw images gamma-corrected to enhance visualization. + +Additionally, we created a mask for regions illuminated by non-dominant illuminants in each scene. These masks ensure that all scenes have only one dominant illuminant, matching the color of the neutral ground truth, without confounding effects from other illuminants. We applied these masks to the training images when training our method and others. To preserve privacy, sensitive information, such as car plates and faces, has been blurred. + +In addition to raw images, both ground-truth illuminant types, and contextual and capture information, we provide additional valuable auxiliary data to broaden the impact of our dataset. This includes locally tone-mapped sRGB images rendered by an expert, that can serve as ground-truth for applications beyond white-balance correction, such as neural ISPs [35, 38, 39, 65]. More information about the dataset can be found in the supplemental material. We organize our dataset into 2,619 raw images for training, 205 raw images for validation, and 400 raw images for testing. + +# 4. Experiments + +Implementation details: We train the model using the Adam optimizer [40], for 400 epochs with betas set to (0.9, 0.999) and a weight decay factor of $10^{-9}$ . A warm-up strategy is applied to gradually increase the learning rate from $10^{-6}$ to $10^{-3}$ over the first 5 epochs. After this, we use a cosine annealing schedule [48]. Following [6], we employ a batch-size increment strategy during training, starting with a batch size of 8 and doubling it every 100 epochs. Following prior work [6, 7], we use images of size $384 \times 256$ pixels in all experiments, which is a reasonable size for camera pipelines to reduce computational overhead. For our method, we use histograms with $h = 48$ bins. The histogram boundaries are determined by computing the 10th and 95th percentiles of the chromaticity values along each chromaticity axis from the training set. We report the results of our method both with and without the inclusion of noise stats, n, and SNR stats, r. + +Results: We report the results of our method alongside several others benchmarked on the proposed dataset. + +Our method is compared with various statistical-based approaches [12, 14, 15, 18, 19, 26, 32, 56, 57, 62], camera-specific learning-based methods [4, 5, 11, 16, 29, 30, 37, 41, 42, 53, 59, 66, 67], and camera-independent techniques [2, 6, 13]. For the camera-specific learning-based approaches, including our method, all models were trained on our training set. + +For the camera-independent methods [2, 6], we present results from models trained on the NUS [19] and Cube++ [23] datasets. Additionally, we report results from finetuned versions of these models [2, 6], incorporating our proposed dataset. We also evaluate camera-specific (CS) models of these methods [2, 6], which were trained exclusively on our training data, without any additional datasets. For further details, refer to the supplemental material. + +Table 1 shows the results on the testing set of our dataset. We report the mean, median, best $25\%$ , worst $25\%$ , worst $5\%$ , tri-mean, and maximum angular errors between the estimated illuminant colors and the ground truth colors for each method. Results are provided for both neutral and user-preference ground truth, with two models, one trained for each ground-truth type, except for the models SIIE [2], SIIE (tuned) [2], C5 [6], and C5 (tuned) [6], which were trained on images from the NUS [19] and Cube++ [23] datasets that do not include user-preference ground truth. + +Additionally, we report the total number of parameters for methods that involve tunable or learnable parameters. The results in Table 1 are based on images without masking regions illuminated by a light source other than the dominant one used to obtain the ground truth, reflecting real-world scenarios where scenes with a single light source are not always guaranteed. For completeness, additional results are provided in the supplementary material, where regions illuminated by light sources different from the ground-truth illuminant color are masked out. + +To further demonstrate the effectiveness of our method on existing DSLR datasets, where contextual metadata is typically unavailable, we report our results on the Simple Cube++ dataset [23]. Although geolocation metadata is absent, the dataset provides capture settings such as ISO and exposure time from the DSLR camera ISP. + +Table 1. Results on the testing set. We report the mean, median, best $25\%$ , worst $25\%$ , tri-mean, and maximum angular errors for each method on neutral and user-preference white-balance ground-truth illuminants, presented in the format (neutral / user-preference). Symbols n and r represent noise stats and SNR stats, respectively. The best and second-best results are highlighted. + +
MethodMeanMed.Best 25%Worst 25%Worst 5%Tri.Max#params (K)
GW [15]6.23 / 5.546.01 / 4.521.02 / 1.0012.43 / 11.9118.05 / 19.815.76 / 4.6528.09 / 31.70-
SoG [26]4.45 / 3.593.54 / 2.170.74 / 0.679.61 / 8.8114.94 / 15.613.77 / 2.6223.22 / 29.36-
GE-1st [62]4.21 / 3.333.29 / 2.120.71 / 0.559.34 / 8.2915.34 / 15.193.49 / 2.3627.81 / 28.80-
GE-2nd [62]4.11 / 3.183.17 / 1.890.70 / 0.589.09 / 7.7914.90 / 14.853.35 / 2.1825.09 / 29.42-
Max-RGB [14]4.01 / 2.612.92 / 1.881.06 / 0.978.57 / 5.6914.01 / 11.073.30 / 1.9434.22 / 23.63-
wGE [32]3.96 / 3.072.96 / 1.910.62 / 0.498.95 / 7.7914.94 / 15.083.21 / 2.1031.33 / 30.54-
PCA [19]4.42 / 3.633.54 / 1.950.70 / 0.619.59 / 9.1914.46 / 16.163.76 / 2.5222.87 / 31.58-
MSGP [57]6.39 / 5.695.72 / 4.380.94 / 0.9813.39 / 12.7320.97 / 22.335.64 / 4.6837.25 / 36.92-
GI [56]4.70 / 4.933.19 / 2.970.44 / 0.7811.63 / 12.2220.48 / 21.353.48 / 3.5336.34 / 36.02-
TECC [12]4.12 / 3.173.23 / 1.910.74 / 0.559.10 / 7.8314.73 / 14.483.46 / 2.1727.08 / 28.87-
Gamut (pixels) [30]3.77 / 2.402.81 / 1.490.77 / 0.638.31 / 5.8412.93 / 11.183.16 / 1.6321.53 / 23.310.636
Gamut (edges) [30]4.45 / 3.943.52 / 3.041.08 / 1.019.51 / 8.5015.00 / 15.253.70 / 3.1828.55 / 29.46324
Gamut (1st) [30]4.10 / 3.653.08 / 2.670.72 / 0.909.26 / 8.2514.75 / 14.693.28 / 2.9121.38 / 26.35279
NIS [29]4.58 / 3.903.76 / 2.570.77 / 0.759.81 / 9.1014.70 / 15.663.87 / 2.9720.76 / 31.800.078
Classification-CC [53]2.73 / 1.611.98 / 1.160.61 / 0.366.03 / 3.609.37 / 6.072.18 / 1.2719.34 / 9.5358,384
FFCC [11]2.62 / 1.501.46 / 0.810.37 / 0.246.89 / 3.9916.59 / 8.431.66 / 0.9548.97 / 18.6012
FFCC (capture info) [11]2.31 / 1.351.38 / 0.800.34 / 0.235.82 / 3.5112.44 / 7.361.60 / 0.8847.67 / 16.9636.9
FC4 [37]3.80 / 2.652.78 / 2.250.85 / 0.858.61 / 5.1415.06 / 7.522.86 / 2.3725.74 / 11.421,705
APAP (GW) [5]3.74 / 2.093.14 / 1.670.93 / 0.497.72 / 4.4411.09 / 6.893.26 / 1.7616.43 / 9.020.289
SIIE [2]4.09 / -3.25 / -0.91 / -8.97 / -15.96 / -3.37 / -43.24 / -1,008
SIIE (tuned) [2]3.15 / -2.22 / -0.51 / -7.28 / -12.14 / -2.46 / -34.52 / -1,008
SIIE (tuned-CS) [2]3.14 / 1.742.20 / 1.200.50 / 0.327.39 / 4.0613.61 / 6.732.41 / 1.3038.96 / 9.441,008
KNN (raw) [4]2.44 / 1.411.51 / 0.830.36 / 0.206.13 / 3.6611.64 / 7.011.66 / 0.9328.95 / 13.49757
Quasi-U-CC [13]3.85 / 3.262.97 / 1.810.55 / 0.588.47 / 8.2413.30 / 15.783.25 / 2.2124.74 / 32.1754,421
Quasi-U-CC (tuned) [13]3.11 / 2.542.27 / 1.430.49 / 0.447.14 / 6.5711.53 / 13.622.44 / 1.6322.67 / 33.7454,421
BoCF [41]3.54 / 2.142.68 / 1.590.96 / 0.487.31 / 4.7411.40 / 7.852.96 / 1.7222.31 / 19.6859
C4 [66]1.92 / 1.491.30 / 0.900.36 / 0.244.64 / 3.829.08 / 7.531.40 / 1.0321.55 / 18.095,116
CWCC [42]3.65 / 2.302.71 / 1.720.82 / 0.677.96 / 4.9512.52 / 9.652.99 / 1.8118.66 / 20.81101
C5 [6]3.22 / -2.51 / -0.78 / -6.97 / -10.54 / -2.68 / -16.38 / -412
C5 (tuned) [6]1.91 / -1.24 / -0.38 / -4.57 / -8.43 / -1.38 / -17.22 / -412
C5 (tuned-CS) [6]1.95 / 1.251.32 / 0.840.37 / 0.234.72 / 2.948.15 / 4.981.44 / 0.9316.78 / 9.23172
TLCC [59]2.71 / 2.742.06 / 1.830.66 / 0.695.89 / 6.3010.30 / 12.902.17 / 1.9921.44 / 33.3332,910
PCC [67]3.03 / 1.672.13 / 1.200.53 / 0.407.08 / 3.7911.20 / 6.842.34 / 1.2716.82 / 12.330.378
RGP [18]4.59 / 4.563.13 / 2.810.43 / 0.7011.13 / 11.2718.79 / 20.253.53 / 3.3132.11 / 33.98-
CFCC [16]3.07 / 1.542.20 / 1.050.73 / 0.396.87 / 3.5512.55 / 6.982.36 / 1.1323.34 / 14.400.283
Ours (w/o n, w/o r)1.93 / 1.261.35 / 0.770.38 / 0.234.56 / 3.139.48 / 5.881.43 / 0.8822.63 / 15.904.83
Ours (w/o n, w/ r)1.89 / 1.231.18 / 0.790.32 / 0.244.74 / 3.0110.10 / 5.071.30 / 0.9024.99 / 12.914.93
Ours (w/ n, w/o r)1.87 / 1.201.24 / 0.720.37 / 0.244.58 / 2.999.82 / 5.581.30 / 0.8229.46 / 15.124.93
Ours (w/ n, w/ r)1.84 / 1.201.24 / 0.770.35 / 0.194.41 / 2.959.17 / 5.121.32 / 0.8735.42 / 12.715.03
+ +We trained our model using the histogram feature $\mathbf{H}$ , along with capture features including ISO $(i)$ , exposure time $(e)$ , and SNR stats $(\mathbf{r})$ , deliberately excluding the contextual metadata that is not available in the Simple Cube++ DSLR dataset. The results of our method, both with and without the SNR stats, are compared to other methods in Table 2. As shown, our method outperforms competing methods across most evaluation metrics. + +Inference time: As shown in Table 1, our method performs comparably to or outperforms prior methods, while maintaining a compact model with only approximately 5K parameters. This lightweight design results in faster runtimes compared to other methods that achieve competitive results, namely C4 [66] and tuned C5 [6]. + +Table 2. Results on the testing set of the 'Simple Cube++' dataset [23]. We report the mean, median, best $25\%$ , worst $25\%$ , tri-mean, and maximum angular errors for each method. Symbols c, i, e, and r refer to the time-capture feature, ISO, exposure time, and SNR stats, respectively. The best and second-best results are highlighted. + +
MethodMeanMed.Best 25%Worst 25%Worst 5%Tri.Max
FFCC [11]1.380.580.184.0510.300.6845.20
FC4 [37]3.722.020.479.9317.182.4029.92
C4 [66]1.160.650.243.006.720.7313.95
C5 (tuned-CS) [6]1.190.600.183.267.610.6815.62
TLCC [59]1.821.150.394.499.631.2019.02
Ours (c = [i, e]T)1.080.590.172.845.530.6911.03
Ours (c = [i, e, rT]T)1.010.600.172.685.710.6510.89
+ +Table 3. Processing time on an AMD Ryzen Threadripper PRO 3975WX CPU460 and an NVIDIA RTX A6000 GPU. + +
MethodCPU (ms)GPU (ms)FLOPs
C4 [66]49.0311.282.28G
C5 [6]8.284.50103.54M
Ours0.550.4516.78M
+ +In Table 3, we present the processing times of our model, C4, and C5 on an AMD Ryzen Threadripper PRO 3975WX CPU and an NVIDIA RTX A6000 GPU. Our fast runtime performance is particularly well-suited for mobile camera ISPs, where limited computational latency is crucial due to the processing demands of other modules (e.g., denoising, local tone mapping) per frame. Our model runs in just 0.25 ms on the DSP and 0.80 ms on the CPU of the Samsung S24 Ultra. + +Ablation studies: We conducted ablation studies to evaluate the impact of each input feature on the validation set of our dataset. Specifically, we assessed our method by excluding either the time-capture feature (c) or the histogram feature $(\mathbf{H})$ . Additionally, we tested the method using only the time feature $(\mathbf{p})$ and noise stats $(\mathbf{n})$ , while excluding the histogram feature $(\mathbf{H})$ . We also evaluated the method using the histogram feature along with the time feature and noise stats, excluding other capture information. Furthermore, we examined the accuracy of our method using all input features except for $\mathbf{p}$ . Lastly, we present the results when all input features are used, but without the noise stats $(\mathbf{n})$ and the SNR stats $(\mathbf{r})$ . These results are presented in Table 4. + +As shown in Table 4, using only the time feature $(\mathbf{p})$ and noise stats $(\mathbf{n})$ yields a reasonable accuracy $(2.37^{\circ}$ mean angular error), compared to FFCC [11], which achieves $2.19^{\circ}$ mean angular error on the validation set. This validates the usefulness of time-of-day and noise information in providing contextual clues. As expected, incorporating the histogram feature $(\mathbf{H})$ , which represents scene colors, along with the time-of-day and noise stats, significantly boosts accuracy, as demonstrated by the results in the fourth row. We further investigate the impact of using noise information in the last three rows: using SNR stats $(\mathbf{r})$ , using noise stats $(\mathbf{n})$ , or both. The combination of both noise features yields the lowest mean angular error. Additional ablation studies are provided in the supplementary material. + +# 5. Conclusion and limitations + +We presented a method for in-camera AWB correction that leverages contextual information. Specifically, we proposed a lightweight model that leverages contextual metadata (notably time-of-day derived from timestamp and geolocation) to guide the illuminant estimation process. In addition to this contextual metadata, we incorporate image colors in the form of histogram features, as well as capture information such as ISO, shutter speed, and noise stats, to help + +Table 4. Results of ablation studies on the validation set with neutral white-balance ground truth. We report the mean, median, best $25\%$ , and worst $25\%$ angular errors for our method with various configurations. Symbols $\mathbf{H}$ , $\mathbf{p}$ , $\mathbf{n}$ , and $\mathbf{r}$ denote the histogram feature, time feature, noise stats, and SNR stats, respectively. The symbol $\mathbf{m}$ refers to the capture information feature, which includes ISO $(i)$ , shutter speed $(s)$ , and flash status $(f)$ , all of which are used in our input time-capture feature $\mathbf{c}$ . The best results are highlighted. + +
HmpnrMeanMed.Best 25%Worst 25%Worst 5%
✓XXXX2.031.360.314.908.90
X✓✓✓✓2.291.810.504.958.58
XX✓✓X2.371.590.455.489.03
✓X✓✓X1.751.170.294.137.26
✓✓X✓✓1.891.210.314.609.41
✓✓✓XX1.851.320.354.267.53
✓✓✓X✓1.721.160.304.138.16
✓✓✓✓X1.671.070.294.047.90
✓✓✓✓✓1.661.200.333.776.95
+ +the model distinguish between artificial and natural scenes and improve its final accuracy. Our method is fast and can achieve high frame rates on modern smartphone DSPs and CPUs, while maintaining accurate illuminant color estimation. + +A key contribution of this work is a large-scale dataset of raw images captured by a consumer smartphone camera, along with the necessary contextual metadata for training and evaluating our method. Beyond the traditional neutral white-balance ground truth extracted from a calibration color chart placed in each scene, we also include a user-preference ground truth that targets the observer's preference, validated through a user study. Results based on both ground truth types demonstrate that our method achieves comparable or superior performance to existing methods, which require larger models. + +While our method represents a promising solution for mobile camera ISPs, its dependency on contextual metadata limits its optimal accuracy to devices that provide geolocation data, which may not be available on DSLR cameras. Additionally, while the contextual metadata is device-independent (i.e., the differences across devices are expected to be minimal), our method relies on additional capture information (i.e., ISO, shutter speed, noise, and SNR stats) and image colors, making it inherently camera-dependent in design. This dependency prevents our trained model from generalizing to new devices without fine-tuning or re-training. However, this issue can be mitigated through color calibration (e.g., [46]), which can be extended to calibrate capture information as well—by performing a preprocessing mapping from the new camera space (for both color and capture information) to the camera space used during training. Additional discussion on cross-camera generalization is provided in the supplementary material. + +# References + +[1] Abdelrahman Abdelhamed, Abhijith Punnappurath, and Michael S Brown. Leveraging the availability of two cameras for illuminant estimation. In CVPR, 2021. 2 +[2] Mahmoud Affifi and Michael S Brown. Sensor-independent illumination estimation for DNN models. In BMVC, 2019. 1, 6, 7 +[3] Mahmoud Affifi and Michael S Brown. Deep white-balance editing. In CVPR, 2020. 1 +[4] Mahmoud Affifi, Brian Price, Scott Cohen, and Michael S Brown. When color constancy goes wrong: Correcting improperly white-balanced images. In CVPR, 2019. 6, 7 +[5] Mahmoud Affifi, Abhijith Punnappurath, Graham Finlayson, and Michael S Brown. As-projective-as-possible bias correction for illumination estimation algorithms. JOSA A, 36 (1):71-78, 2019. 6, 7 +[6] Mahmoud Afifi, Jonathan T Barron, Chloe LeGendre, Yun-Ta Tsai, and Francois Bleibel. Cross-camera convolutional color constancy. In ICCV, 2021. 1, 2, 4, 6, 7, 8 +[7] Mahmoud Affifi, Zhenhua Hu, and Liang Liang. Optimizing illuminant estimation in dual-exposure HDR imaging. In ECCV, 2024. 1, 2, 5, 6 +[8] Vivek Agarwal, Besma R Abidi, Andreas Koschan, and Mongi A Abidi. An overview of color constancy algorithms. Journal of Pattern Recognition Research, 1(1):42-54, 2006. 1 +[9] Relja Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In CVPR, 2012. 4 +[10] Jonathan T Barron. Convolutional color constancy. In ICCV, 2015. 1, 2, 4 +[11] Jonathan T Barron and Yun-Ta Tsai. Fast Fourier color constancy. In CVPR, 2017. 2, 4, 6, 7, 8 +[12] Simone Bianco and Marco Buzzelli. Truncated edge-based color constancy. In ICCE, 2022. 6, 7 +[13] Simone Bianco and Claudio Cusano. Quasi-unsupervised color constancy. In CVPR, 2019. 1, 6, 7 +[14] David H Brainard and Brian A Wandell. Analysis of the Retinex theory of color vision. Journal of the Optical Society of America A, 3(10):1651-1661, 1986. 6, 7 +[15] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin Institute, 310(1): 1-26, 1980. 1, 6, 7 +[16] Marco Buzzelli and Simone Bianco. A convolutional framework for color constancy. IEEE Transactions on Neural Networks and Learning Systems, pages 1-15, 2024. 6, 7 +[17] Cheng Cheng, Kai-Fu Yang, Xue-Mei Wan, Leanne Lai Hang Chan, and Yong-Jie Li. Nighttime color constancy using robust gray pixels. JOSA A, 41(3):476-488, 2024. 5 +[18] Cheng Cheng, Kai Fu Yang, Xue Mei Wan, Leanne Lai Hang Chan, and Yonge Jie Li. Nighttime color constancy using robust gray pixels. JOsA A, 41(3):476-488, 2024. 6, 7 +[19] Dongliang Cheng, Dilip K Prasad, and Michael S Brown. Illuminant estimation for color constancy: Why spatial-domain methods work and the role of the color distribution. JOSA A, 31(5):1049-1058, 2014. 2, 5, 6, 7 + +[20] Kyungah Choi and Hyeon-Jeong Suk. User-preferred color temperature adjustment for smartphone display under varying illuminants. Optical Engineering, 53(6):061708-061708, 2014. 5 +[21] Djork-Arné Clevert. Fast and accurate deep network learning by exponential linear units (ELUs). arXiv preprint arXiv:1511.07289, 2015. 5 +[22] Maurizio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Mobile computational photography: A tour. Annual Review of Vision Science, 7(1):571-604, 2021. 3 +[23] Egor Ershov, Alexey Savchik, Illya Semenkov, Nikola Banić, Alexander Belokopytov, Daria Senshina, Karlo Koščević, Marko Subašić, and Sven Lončarić. The cube++ illumination estimation dataset. IEEE Access, 8:227511-227527, 2020. 2, 5, 6, 7 +[24] Egor Ershov, Alex Savchik, Denis Shepelev, Nikola Banić, Michael S Brown, Radu Timofte, Karlo Koščević, Michael Freeman, Vasily Tesalin, Dmitry Bocharov, et al. NTIRE 2022 challenge on night photography rendering. In CVPRW, 2022. 5 +[25] Graham D Finlayson and Steven D Hordley. Color constancy at a pixel. JOSA A, 18(2):253-264, 2001. 4 +[26] Graham D Finlayson and Elisabetta Trezzi. Shades of gray and colour constancy. In CIC, 2004. 1, 6, 7 +[27] Karl R Gegenfurtner, David Weiss, and Marina Bloj. Color constancy in real-world settings. Journal of Vision, 24(2): 12-12, 2024. 1 +[28] Peter Vincent Gehler, Carsten Rother, Andrew Blake, Tom Minka, and Toby Sharp. Bayesian color constancy revisited. In CVPR, 2008. 2, 5 +[29] Arjan Gijsenij and Theo Gevers. Color constancy using natural image statistics and scene semantics. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(4): 687-698, 2011. 6, 7 +[30] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Generalized gamut mapping using image derivative structures for color constancy. International Journal of Computer Vision, 86(2):127-139, 2010. 2, 6, 7 +[31] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Computational color constancy: Survey and experiments. IEEE Transactions on Image Processing, 20(9):2475-2489, 2011. 1 +[32] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):918-929, 2011. 1, 6, 7 +[33] Thorsten Hansen, Sebastian Walter, and Karl R Gegenfurtner. Effects of spatial and temporal context on color categories and color constancy. Journal of Vision, 7(4):2-2, 2007. 1 +[34] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics, 35(6):1-12, 2016. 3 + +[35] Xuanhua He, Tao Hu, Guoli Wang, Zejin Wang, Run Wang, Qian Zhang, Keyu Yan, Ziyi Chen, Rui Li, Chengjun Xie, et al. Enhancing RAW-to-sRGB with decoupled style structure in Fourier domain. In AAAI, 2024. 6 +[36] Steven D Hordley and Graham D Finlayson. Re-evaluating colour constancy algorithms. In ICPR, 2004. 5 +[37] Yuanming Hu, Baoyuan Wang, and Stephen Lin. FC4: Fully convolutional color constancy with confidence-weighted pooling. In CVPR, 2017. 6, 7 +[38] Andrey Ignatov, Anastasia Sycheva, Radu Timofte, Yu Tseng, Yu-Syuan Xu, Po-Hsiang Yu, Cheng-Ming Chiang, Hsien-Kai Kuo, Min-Hung Chen, Chia-Ming Cheng, et al. MicroISP: processing 32MP photos on mobile devices with deep learning. In ECCV, 2022. 6 +[39] Woohyeok Kim, Geonu Kim, Junyong Lee, Seungyong Lee, Seung-Hwan Baek, and Sunghyun Cho. ParamISP: Learned forward and inverse ISPs using camera parameters. In CVPR, 2024. 6 +[40] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[41] Firas Laakom, Nikolaos Passalis, Jenni Raitoharju, Jarno Nikkanen, Anastasios Tefas, Alexandros Iosifidis, and Moncef Gabbouj. Bag of color features for color constancy. IEEE Transactions on Image Processing, 29:7722-7734, 2020. 1, 6, 7 +[42] Firas Laakom, Jenni Raitoharju, Jarno Nikkanen, Alexandros Iosifidis, and Moncef Gabbouj. Robust channel-wise illumination estimation. In BMVC, 2021. 2, 6, 7 +[43] Firas Laakom, Jenni Raitoharju, Jarno Nikkanen, Alexandros Iosifidis, and Moncef Gabbouj. Intel-TAU: A color constancy dataset. IEEE Access, 9:39560-39567, 2021. 2, 5 +[44] Shuwei Li and Robby T Tan. NightCC: Nighttime color constancy via adaptive channel masking. In CVPR, 2024. 5 +[45] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM Transactions on Graphics, 38(6):164-1, 2019. 3 +[46] Yi-Tun Lin, Bianjiang Yang, Hao Xie, Wenbin Wang, Honghong Peng, and Jun Hu. Color constancy: How to deal with camera bias? In BMVC, 2023. 8 +[47] Ce Liu, William T Freeman, Richard Szeliski, and Sing Bing Kang. Noise estimation from a single image. In CVPR, 2006. 4 +[48] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6 +[49] Zhongyu Lou, Theo Gevers, Ninghang Hu, Marcel P Lucassen, et al. Color constancy by deep learning. In BMVC, 2015. 1 +[50] Ming Luo. A review of chromatic adaptation transforms. Review of Progress in Coloration and Related Topics, 30:77 - 92, 2008. 5 +[51] Laurence T Maloney. Illuminant estimation as cue combination. Journal of Vision, 2(6):6-6, 2002. 1 + +[52] Jean H Meeus. *Astronomical algorithms*. Willmann-Bell, Incorporated, 1991. 3 +[53] Seoung Wug Oh and Seon Joo Kim. Approaching the computational color constancy as a classification problem through deep learning. Pattern Recognition, 61:405-416, 2017. 1, 6, 7 +[54] Varad A Pimpalkhute, Rutvik Page, Ashwin Kothari, Kishor M Bhurchandi, and Vipin Milind Kamble. Digital image noise estimation using DWT coefficients. IEEE Transactions on Image Processing, 30:1962-1972, 2021. 4 +[55] Stanislav Pyatykh, Jürgen Hesser, and Lei Zheng. Image noise level estimation by principal component analysis. IEEE Transactions on Image Processing, 22(2):687-699, 2012. 4 +[56] Yanlin Qian, Joni-Kristian Kamarainen, Jarno Nikkanen, and Jiri Matas. On finding gray pixels. In CVPR, 2019. 1, 6, 7 +[57] Yanlin Qian, Said Pertuz, Jarno Nikkanen, Joni-Kristian Kämäräinen, and Jiri Matas. Revisiting gray pixel for statistical illumination estimation. In VISSAP. 2019. 1, 6, 7 +[58] Wu Shi, Chen Change Loy, and Xiaou Tang. Deep specialized network for illuminant estimation. In ECCV, pages 371-387, 2016. 1 +[59] Yuxiang Tang, Xuejing Kang, Chunxiao Li, Zhaowen Lin, and Anlong Ming. Transfer learning for color constancy via statistic perspective. In AAAI, 2022. 2, 6, 7 +[60] Shoji Tominaga, Takahiko Horiuchi, Shiori Nakajima, and Mariko Yano. Prediction of incomplete chromatic adaptation under illuminant A from images. In CIC, 2014. 2, 5 +[61] Oguzhan Ulucan, Diclehan Ulucan, and Marc Ebner. Color constancy beyond standard illuminants. In ICIP, 2022. 2, 5 +[62] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy. IEEE Transactions on Image Processing, 16(9):2207-2214, 2007. 1, 6, 7 +[63] TC Van Flandern and KF Pulkkinen. Low-precision formulae for planetary positions. Astrophysical Journal Supplement Series, 41:391-411, 1979. 3 +[64] Robert Walraven. Calculating the position of the sun. Solar energy, 20(5):393-397, 1978. 3 +[65] Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible image signal processing. In CVPR, 2021. 6 +[66] Huanglin Yu, Ke Chen, Kaiqi Wang, Yanlin Qian, Zhaoxiang Zhang, and Kui Jia. Cascading convolutional color constancy. In AAAI, 2020. 1, 6, 7, 8 +[67] Shuwei Yue and Minchen Wei. Color constancy from a pure color view. Journal of the Optical Society of America A, 40 (3):602-610, 2023. 6, 7 +[68] Yi Zhang, Hongwei Qin, Xiaogang Wang, and Hongsheng Li. Rethinking noise synthesis and modeling in raw denoising. In ICCV, 2021. 4 \ No newline at end of file diff --git a/2025/Time-Aware Auto White Balance in Mobile Photography/images.zip b/2025/Time-Aware Auto White Balance in Mobile Photography/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..36c4094dba3f979692a02e5c3b47267a83b97cb0 --- /dev/null +++ b/2025/Time-Aware Auto White Balance in Mobile Photography/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2888344632dea25ef8b4145d9ba737061b713ad415c408e0f7554421cc0be2f +size 742947 diff --git a/2025/Time-Aware Auto White Balance in Mobile Photography/layout.json b/2025/Time-Aware Auto White Balance in Mobile Photography/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c76b1499505cb4b4ba70937076c341ec8dbdb002 --- /dev/null +++ b/2025/Time-Aware Auto White Balance in Mobile Photography/layout.json @@ -0,0 +1,9791 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 479, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 479, + 121 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 479, + 121 + ], + "type": "text", + "content": "Time-Aware Auto White Balance in Mobile Photography" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 137, + 242, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 137, + 242, + 150 + ], + "spans": [ + { + "bbox": [ + 162, + 137, + 242, + 150 + ], + "type": "text", + "content": "Mahmoud Affifi*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 261, + 137, + 318, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 137, + 318, + 150 + ], + "spans": [ + { + "bbox": [ + 261, + 137, + 318, + 150 + ], + "type": "text", + "content": "Luxi Zhao*" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 337, + 137, + 448, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 137, + 448, + 152 + ], + "spans": [ + { + "bbox": [ + 337, + 137, + 448, + 152 + ], + "type": "text", + "content": "Abhijith Punnappurath" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 146, + 152, + 282, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 152, + 282, + 164 + ], + "spans": [ + { + "bbox": [ + 146, + 152, + 282, + 164 + ], + "type": "text", + "content": "Mohammed A. Abdelsalam" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 301, + 152, + 463, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 152, + 463, + 166 + ], + "spans": [ + { + "bbox": [ + 301, + 152, + 463, + 166 + ], + "type": "text", + "content": "Ran Zhang Michael S. Brown" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 205, + 166, + 405, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 166, + 405, + 179 + ], + "spans": [ + { + "bbox": [ + 205, + 166, + 405, + 179 + ], + "type": "text", + "content": "AI Center-Toronto, Samsung Electronics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 182, + 526, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 182, + 526, + 194 + ], + "spans": [ + { + "bbox": [ + 83, + 182, + 526, + 194 + ], + "type": "text", + "content": "{m.afifil, lucy.zhao, abhijith.p, m.abdelsalam, ran.zhang, michael.b1}@samsung.com" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 58, + 203, + 181, + 365 + ], + "blocks": [ + { + "bbox": [ + 58, + 203, + 181, + 365 + ], + "lines": [ + { + "bbox": [ + 58, + 203, + 181, + 365 + ], + "spans": [ + { + "bbox": [ + 58, + 203, + 181, + 365 + ], + "type": "image", + "image_path": "7fc4c1df2b7ca5e35050456c912a700f35890770c4ff1ee1c02be792a3fb1784.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 70, + 366, + 89, + 385 + ], + "blocks": [ + { + "bbox": [ + 70, + 366, + 89, + 385 + ], + "lines": [ + { + "bbox": [ + 70, + 366, + 89, + 385 + ], + "spans": [ + { + "bbox": [ + 70, + 366, + 89, + 385 + ], + "type": "image", + "image_path": "20de2c620390972ff2c6eaea908e4ac94af663fb2c890494f3aef6ca9751b23c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 394, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 54, + 394, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 394, + 555, + 437 + ], + "type": "text", + "content": "Figure 1. The time of day influences scene illumination, making it a valuable cue for improving illuminant estimation. Shown are white-balanced results (gamma-corrected for visualization) using our method with (1) colors only, (2) colors plus contextual (timestamp and geolocation) and capture data, and (3) ground truth (from a color chart). Angular errors show improvements when time-capture information is used." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 100, + 366, + 119, + 386 + ], + "blocks": [ + { + "bbox": [ + 100, + 366, + 119, + 386 + ], + "lines": [ + { + "bbox": [ + 100, + 366, + 119, + 386 + ], + "spans": [ + { + "bbox": [ + 100, + 366, + 119, + 386 + ], + "type": "image", + "image_path": "34933caa09610af20c0f945853961c95383be92bbb7d66865e91ff02329cd952.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 120, + 366, + 138, + 386 + ], + "blocks": [ + { + "bbox": [ + 120, + 366, + 138, + 386 + ], + "lines": [ + { + "bbox": [ + 120, + 366, + 138, + 386 + ], + "spans": [ + { + "bbox": [ + 120, + 366, + 138, + 386 + ], + "type": "image", + "image_path": "494610d5708b103912b8b3e91058b1c46fa8e0035839c2e65549f75d07e62175.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 182, + 204, + 304, + 365 + ], + "blocks": [ + { + "bbox": [ + 182, + 204, + 304, + 365 + ], + "lines": [ + { + "bbox": [ + 182, + 204, + 304, + 365 + ], + "spans": [ + { + "bbox": [ + 182, + 204, + 304, + 365 + ], + "type": "image", + "image_path": "0c0110915450ac8e4be1ea77794ccf5a754fce231ba0fdd1491162392c00be3b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 193, + 366, + 212, + 385 + ], + "blocks": [ + { + "bbox": [ + 193, + 366, + 212, + 385 + ], + "lines": [ + { + "bbox": [ + 193, + 366, + 212, + 385 + ], + "spans": [ + { + "bbox": [ + 193, + 366, + 212, + 385 + ], + "type": "image", + "image_path": "804e0117b967e293d9741042ea051a11d7801395b6e4320f7e6ecfd0d683aa6f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 224, + 366, + 261, + 386 + ], + "blocks": [ + { + "bbox": [ + 224, + 366, + 261, + 386 + ], + "lines": [ + { + "bbox": [ + 224, + 366, + 261, + 386 + ], + "spans": [ + { + "bbox": [ + 224, + 366, + 261, + 386 + ], + "type": "image", + "image_path": "f4234c86efb8c3db926f7d3948c6430d4f02764cf7391275e0dbe93abd58933c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 307, + 204, + 428, + 365 + ], + "blocks": [ + { + "bbox": [ + 307, + 204, + 428, + 365 + ], + "lines": [ + { + "bbox": [ + 307, + 204, + 428, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 204, + 428, + 365 + ], + "type": "image", + "image_path": "890626b0b999577fb0a45a5a233fea3a624a3e705b69a7695a4c09f0cee8444f.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 316, + 366, + 335, + 385 + ], + "blocks": [ + { + "bbox": [ + 316, + 366, + 335, + 385 + ], + "lines": [ + { + "bbox": [ + 316, + 366, + 335, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 335, + 385 + ], + "type": "image", + "image_path": "7ab03fe008aea03315fc8bdcb39fa05d8d813d3261cbf357c35aab9c1ba117e5.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 346, + 366, + 384, + 386 + ], + "blocks": [ + { + "bbox": [ + 346, + 366, + 384, + 386 + ], + "lines": [ + { + "bbox": [ + 346, + 366, + 384, + 386 + ], + "spans": [ + { + "bbox": [ + 346, + 366, + 384, + 386 + ], + "type": "image", + "image_path": "4c8e35235cab070dd208c92e266b5e941db2ca4d6881db3c51cb5e93dabefb27.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 429, + 205, + 552, + 365 + ], + "blocks": [ + { + "bbox": [ + 429, + 205, + 552, + 365 + ], + "lines": [ + { + "bbox": [ + 429, + 205, + 552, + 365 + ], + "spans": [ + { + "bbox": [ + 429, + 205, + 552, + 365 + ], + "type": "image", + "image_path": "6b75febfbbb93d44c1e182766884ab821dc57151fc76fc0c04a60cc4f9a1100d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 438, + 366, + 457, + 385 + ], + "blocks": [ + { + "bbox": [ + 438, + 366, + 457, + 385 + ], + "lines": [ + { + "bbox": [ + 438, + 366, + 457, + 385 + ], + "spans": [ + { + "bbox": [ + 438, + 366, + 457, + 385 + ], + "type": "image", + "image_path": "d2bbeef182343bfae2d9943f268bc9fc8e9c603a81f1c3e213ebe50da9ee833a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 470, + 366, + 507, + 386 + ], + "blocks": [ + { + "bbox": [ + 470, + 366, + 507, + 386 + ], + "lines": [ + { + "bbox": [ + 470, + 366, + 507, + 386 + ], + "spans": [ + { + "bbox": [ + 470, + 366, + 507, + 386 + ], + "type": "image", + "image_path": "6e1502e8ef95c143df89633f3b5cf4f114bf3ff14e1af5a61e699f95bd8d5952.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 151, + 440, + 200, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 440, + 200, + 452 + ], + "spans": [ + { + "bbox": [ + 151, + 440, + 200, + 452 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 54, + 457, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 457, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 457, + 295, + 696 + ], + "type": "text", + "content": "Cameras rely on auto white balance (AWB) to correct undesirable color casts caused by scene illumination and the camera's spectral sensitivity. This is typically achieved using an illuminant estimator that determines the global color cast solely from the color information in the camera's raw sensor image. Mobile devices provide valuable additional metadata—such as capture timestamp and geolocation—that offers strong contextual clues to help narrow down the possible illumination solutions. This paper proposes a lightweight illuminant estimation method that incorporates such contextual metadata, along with additional capture information and image colors, into a lightweight model (" + }, + { + "bbox": [ + 54, + 457, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\sim 5K" + }, + { + "bbox": [ + 54, + 457, + 295, + 696 + ], + "type": "text", + "content": " parameters), achieving promising results, matching or surpassing larger models. To validate our method, we introduce a dataset of 3,224 smartphone images with contextual metadata collected at various times of day and under diverse lighting conditions. The dataset includes ground-truth illuminant colors, determined using a color chart, and user-preferred illuminants validated through a user study, providing a comprehensive benchmark for AWB evaluation." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 477, + 485, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 477, + 485, + 490 + ], + "spans": [ + { + "bbox": [ + 314, + 477, + 485, + 490 + ], + "type": "text", + "content": "1. Introduction and related work" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 497, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 497, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 312, + 497, + 555, + 640 + ], + "type": "text", + "content": "Color constancy refers to the ability of the human visual system to maintain stable object colors despite variations in lighting conditions by leveraging contextual cues within the scene [27, 33, 51]. Cameras approximate this effect using auto white balance (AWB) correction, which aims to partially neutralize color casts introduced by scene illumination and the camera's spectral sensitivity [3]. AWB first estimates the illumination color as an RGB vector in the camera's raw color space. The raw image is then corrected by scaling its color channels according to the estimated illumination, typically under the assumption of a single global light source [8, 31]." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Conventional illuminant estimation methods primarily rely on image colors, either by directly processing the raw image (e.g., [13, 32, 41, 49, 53, 58, 62, 66]) or by analyzing color histograms (e.g., [2, 6, 7, 10]). These methods can be broadly categorized into two groups: (1) classical statistical-based methods (e.g., [15, 26, 32, 56, 57, 62])," + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 134, + 712 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5038" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 290, + 269 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 290, + 269 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 290, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 290, + 269 + ], + "type": "image", + "image_path": "f233ecfc8dca891a11308e394d6b4a4823f290d025912e67d087ddffdabfae92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 277, + 295, + 368 + ], + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 368 + ], + "type": "text", + "content": "Figure 2. The time of day at which an image is captured provides valuable information about the possible range of illuminant colors in outdoor scenes. The figure presents the correlated color temperature (CCT) of illuminant colors in our dataset (Sec. 3) for images captured at various times throughout the day and night. As shown, excluding images taken under artificial light, those captured at noon, for example, exhibit a different range of illuminant CCTs compared to images captured during sunset or sunrise." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 384, + 295, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 433 + ], + "type": "text", + "content": "which estimate the illuminant color based on image statistics, and (2) learning-based methods (e.g., [6, 10, 30, 42, 59]), which map image colors to their corresponding scene illuminant through data-driven models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 518 + ], + "type": "text", + "content": "While image colors are a key cue for estimating the scene's illuminant, mobile devices provide an opportunity to integrate additional contextual information. For instance, the device's location, along with the date and time, offer valuable cues about outdoor lighting conditions (e.g., sunrise, noon, sunset), thereby improving illuminant estimation for outdoor scenes (see Fig. 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 687 + ], + "type": "text", + "content": "Intuitively, knowing the time of day when an outdoor scene is captured can help estimate the lighting conditions. Figure 2 further illustrates that the time of day, derived from contextual metadata (i.e., timestamp and geolocation) readily available on mobile devices, provides valuable insights into the likely range of illuminant correlated color temperature (CCT) in outdoor scenes. This information serves to narrow the range of possible illuminant colors. For instance, images taken at noon exhibit a different CCT range than those captured at sunrise or sunset. When combined with additional capture information to distinguish between environments (e.g., indoors vs. outdoors), such metadata can complement image colors to improve illuminant estimation accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Despite its potential, only a few attempts have explored leveraging additional information available to the camera's" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "image signal processor (ISP) for illuminant estimation, such as metadata-based model control [11] or data augmentation to enhance generalization [6]. However, to our knowledge, no previous work has investigated using contextual information—specifically, mobile device timestamps and geolocation—to refine illuminant estimation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "text", + "content": "We introduce a method that leverages contextual metadata from mobile phones, along with additional capture information available in camera ISPs, to train a lightweight illuminant estimator model with " + }, + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\sim 5\\mathrm{K}" + }, + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "text", + "content": " parameters. Our model delivers promising results, matching or surpassing that of larger models, while maintaining efficiency. Our model runs on a typical flagship mobile digital signal processor (DSP) and CPU in " + }, + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "inline_equation", + "content": "0.25\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "inline_equation", + "content": "0.80\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 160, + 555, + 293 + ], + "type": "text", + "content": ", respectively. This compact and efficient design is especially advantageous for mobile devices, where minimizing power consumption and memory usage is critical [1, 7, 11]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 308, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 524 + ], + "type": "text", + "content": "Existing white-balance datasets (e.g., [19, 23, 28, 43]) lack the contextual information needed to validate our method. The absence of contextual information is because most available datasets (e.g., [19, 23]) used DSLR cameras, which typically lack built-in GPS functionality or accurate timestamps. To address this, we captured a new dataset of 3,224 images using a consumer smartphone camera, accompanied by contextual and capture information. The ground truth for the dataset was established using two approaches: (1) the conventional approach, selecting gray patches from a calibration color chart, and (2) a manually selected user-preference white-balance target, which enhances the image's aesthetic appeal and accounts for incomplete chromatic adaptation [60]. The user-preference white-balance \"ground truth\" was validated through a user study. The dataset spans a variety of lighting conditions, including non-standard artificial illuminants [61], and covers different times of day—sunrise, noon, sunset, and night." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "Contribution: We propose an AWB method that utilizes smartphone contextual metadata (i.e., timestamp and geolocation) alongside capture information to enhance illuminant estimation. We demonstrate that integrating this additional data with conventional color information into a lightweight neural model leads to significant improvements in accuracy. Additionally, we introduce a large-scale dataset captured with a consumer smartphone at various times of day, with ground truth derived from both a calibration color chart and user-preference-based white balance. Benchmarking results show that our method achieves state-of-the-art results in most standard metrics (i.e., angular error statistics [19]), while maintaining computational efficiency on mobile devices." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5039" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 294, + 206 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 294, + 206 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 294, + 206 + ], + "type": "image", + "image_path": "1591a830cb7446514d9e26f8a58a76bccbdae499c6a48f3d34c47ded2a49bdbd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "type": "text", + "content": "Figure 3. Our method includes a lightweight model consisting of a convolutional network that processes the histogram feature (derived from raw image colors and edge histograms concatenated with the " + }, + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "type": "inline_equation", + "content": "u / v" + }, + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "type": "text", + "content": " coordinates) to produce a latent feature " + }, + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_h\\in \\mathbb{R}^{16}" + }, + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "type": "text", + "content": ". This is then concatenated with the latent feature of the processed time-capture feature, " + }, + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t\\in \\mathbb{R}^{16}" + }, + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "type": "text", + "content": ". The combined feature, " + }, + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{v}\\in \\mathbb{R}^{32}" + }, + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "type": "text", + "content": ", is passed through a lightweight MLP to produce the chromaticity of the scene illuminant, " + }, + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "type": "inline_equation", + "content": "\\ell_c\\in \\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 217, + 296, + 316 + ], + "type": "text", + "content": ", which is finally converted into normalized RGB illuminant color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 331, + 111, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 111, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 111, + 342 + ], + "type": "text", + "content": "2. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 352, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 296, + 544 + ], + "type": "text", + "content": "An overview of our method is shown in Fig. 3. Our method employs a learnable model that processes two distinct inputs: (1) the time-capture feature (Sec. 2.1), which combines the contextual and capture information available on mobile devices and accessible by their camera ISPs, and (2) the histogram feature (Sec. 2.2), which represents the image's R/G and B/G chromaticity values. The time-capture feature is first processed to project it into a latent space, producing the time-capture latent feature vector, " + }, + { + "bbox": [ + 55, + 352, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t \\in \\mathbb{R}^{16}" + }, + { + "bbox": [ + 55, + 352, + 296, + 544 + ], + "type": "text", + "content": ". The model processes the histogram feature to produce the histogram's latent feature vector, " + }, + { + "bbox": [ + 55, + 352, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_h \\in \\mathbb{R}^{16}" + }, + { + "bbox": [ + 55, + 352, + 296, + 544 + ], + "type": "text", + "content": ". Both the histogram and time-capture latent feature vectors are concatenated and processed by a set of learnable layers to output an R/G, B/G chromaticity vector, " + }, + { + "bbox": [ + 55, + 352, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\ell_c \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 352, + 296, + 544 + ], + "type": "text", + "content": ", representing the scene illuminant. This 2D vector is then converted into the illuminant RGB color used to perform white balancing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 552, + 178, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 178, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 178, + 565 + ], + "type": "text", + "content": "2.1. Time-capture feature" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Our model incorporates contextual metadata, available on mobile devices, as one of its input features. Specifically, we use the geolocation and timestamp of image capture to compute the \"probability\" of the time of day (e.g., sunset, noon, etc.). Our approach converts conventional clock time (hour::minutes) to its corresponding time of day condition based on the date and geolocation of the scene. This approach allows the model to generalize across different time zones and prevents it from being influenced by the capturing location. The time probability vector represents the likelihood that the captured image corresponds to one of the solar event times (i.e., dawn, sunrise, noon, sunset, dusk, or mid" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 455, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 455, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 455, + 84 + ], + "type": "text", + "content": "night) and is computed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 392, + 94, + 553, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 94, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 392, + 94, + 553, + 119 + ], + "type": "interline_equation", + "content": "p _ {g} = 1 - \\frac {\\left| t _ {c} - t _ {g} \\right|}{t _ {s}}, \\tag {1}", + "image_path": "7bac2226fc0834fe2b8fbb1318ccec53ea1d0b74e0431b522812cba8dbaa6ff6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 124, + 555, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 124, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 124, + 555, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 124, + 555, + 207 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 313, + 124, + 555, + 207 + ], + "type": "text", + "content": " is the capture time in seconds (adjusted to the local time zone based on geolocation information), and " + }, + { + "bbox": [ + 313, + 124, + 555, + 207 + ], + "type": "inline_equation", + "content": "t_g" + }, + { + "bbox": [ + 313, + 124, + 555, + 207 + ], + "type": "text", + "content": " is the local time in seconds of the solar event, " + }, + { + "bbox": [ + 313, + 124, + 555, + 207 + ], + "type": "inline_equation", + "content": "g \\in \\{\\text{dawn,sunrise,noon,sunset,dusk,midnight}\\}" + }, + { + "bbox": [ + 313, + 124, + 555, + 207 + ], + "type": "text", + "content": ", computed using geolocation-based standard algorithms [52, 63, 64]. The scalar " + }, + { + "bbox": [ + 313, + 124, + 555, + 207 + ], + "type": "inline_equation", + "content": "t_s" + }, + { + "bbox": [ + 313, + 124, + 555, + 207 + ], + "type": "text", + "content": " represents the total number of seconds in a day (i.e., 86,400)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 209, + 555, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 352 + ], + "type": "text", + "content": "We pre-process the probability of each solar event, " + }, + { + "bbox": [ + 313, + 209, + 555, + 352 + ], + "type": "inline_equation", + "content": "p_{g}" + }, + { + "bbox": [ + 313, + 209, + 555, + 352 + ], + "type": "text", + "content": ", in our time probability vector by computing the square root to enhance feature representation—compressing high probabilities and amplifying lower ones, which is intended to help create a more balanced and smooth distribution for the model to leverage. We then augment this time probability vector with a one-hot vector, " + }, + { + "bbox": [ + 313, + 209, + 555, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 313, + 209, + 555, + 352 + ], + "type": "text", + "content": ", that indicates whether the capture time " + }, + { + "bbox": [ + 313, + 209, + 555, + 352 + ], + "type": "inline_equation", + "content": "t_{c}" + }, + { + "bbox": [ + 313, + 209, + 555, + 352 + ], + "type": "text", + "content": " occurs before each solar event. This distinction helps the model account for expected variations in CCT, as illuminant colors can differ before and after certain solar events, such as sunset and sunrise. The value of this one-hot vector for a given solar event " + }, + { + "bbox": [ + 313, + 209, + 555, + 352 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 209, + 555, + 352 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 386, + 361, + 553, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 361, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 386, + 361, + 553, + 393 + ], + "type": "interline_equation", + "content": "\\mathbf {b} _ {g} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} t _ {c} \\leq t _ {g} \\\\ 0, & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {2}", + "image_path": "6dfc06c74ab96156e16e9d0801671ef9f8a2a714955cc5eb06e26cb8af38cabc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 402, + 553, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 553, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 402, + 553, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_g" + }, + { + "bbox": [ + 313, + 402, + 553, + 438 + ], + "type": "text", + "content": " corresponds to the entry in the one-hot vector for the solar event, " + }, + { + "bbox": [ + 313, + 402, + 553, + 438 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 402, + 553, + 438 + ], + "type": "text", + "content": ". Both the time probability vector and the one-hot vector together form our time feature, " + }, + { + "bbox": [ + 313, + 402, + 553, + 438 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^{12}" + }, + { + "bbox": [ + 313, + 402, + 553, + 438 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 438, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 496 + ], + "type": "text", + "content": "To enrich our feature set with additional capture information available from the camera ISP and help distinguish the capturing environment (e.g., indoor vs. outdoor), we include the following features in our final time-capture feature, c:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 498, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 314, + 498, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 498, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 314, + 498, + 554, + 545 + ], + "type": "text", + "content": "- ISO (i): The sensitivity of the camera's image sensor to light, where lower values indicate good lighting conditions (e.g., bright scenes) and higher values suggest low-light environments, such as poorly lit indoor scenes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 546, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 546, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 315, + 546, + 554, + 593 + ], + "type": "text", + "content": "- Shutter speed " + }, + { + "bbox": [ + 315, + 546, + 554, + 593 + ], + "type": "inline_equation", + "content": "(s)" + }, + { + "bbox": [ + 315, + 546, + 554, + 593 + ], + "type": "text", + "content": ": The amount of time the camera's shutter remains open, allowing light to hit the image sensor. It provides an indication of lighting conditions, alongside the ISO value, " + }, + { + "bbox": [ + 315, + 546, + 554, + 593 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 315, + 546, + 554, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 594, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 594, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 315, + 594, + 553, + 617 + ], + "type": "text", + "content": "- Flash status " + }, + { + "bbox": [ + 315, + 594, + 553, + 617 + ], + "type": "inline_equation", + "content": "(f)" + }, + { + "bbox": [ + 315, + 594, + 553, + 617 + ], + "type": "text", + "content": ": A binary value indicating whether flash light was used during capturing." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 618, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 618, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 618, + 555, + 713 + ], + "type": "text", + "content": "- Noise information (optional): Since image denoising is typically applied before or in parallel with illuminant estimation in camera ISPs [22, 34, 45], we also explore the optional use of explicit noise information from the captured scene. More noise typically indicates low-light conditions, which can provide clues about the lighting color range of the scene. Unfortunately, while noise information is accessible within camera ISPs, obtaining accurate" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5040" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 131 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 131 + ], + "type": "text", + "content": "noise information for public use is challenging, as the noise profiles in DNG files are not always reliable [68]. To address this, we simulate the noise information using two approaches: noise statistics (stats) and/or signal-to-noise ratio (SNR) stats, which are described below." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 137, + 295, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 244 + ], + "type": "text", + "content": "Noise stats (n): This represents the noise statistics in the captured raw image. We simulate this by denoising each raw image using Adobe Lighthroom and computing the noise stats as the mean and standard deviation of each color channel of the absolute difference between the denoised and noisy raw images. This approach provides a simplified method for estimating noise stats, as the denoised images are typically available within the camera ISPs, but difficult to extract from the DNG files." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "text", + "content": "SNR stats (r): This alternative approach measures the noise information in the captured image without the need of a denoised reference. The SNR is computed by applying a " + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "inline_equation", + "content": "15 \\times 15" + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "text", + "content": " sliding window over the raw image and calculated as: " + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "inline_equation", + "content": "10\\log_{10}\\left(\\frac{\\mu}{\\sigma + \\epsilon}\\right)" + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "text", + "content": " represents the mean RGB of the " + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "inline_equation", + "content": "15 \\times 15" + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "text", + "content": " patch, " + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "text", + "content": " is the standard deviation, and " + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 251, + 296, + 339 + ], + "type": "text", + "content": " is a small value added for numerical stability." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 295, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 388 + ], + "type": "text", + "content": "Note that while these two approaches—namely, computing noise stats and SNR stats—yield satisfactory results, there is a wide body of research on noise estimation (e.g., [47, 54, 55]), which is beyond the scope of this paper." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 388, + 296, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 412 + ], + "type": "text", + "content": "Our complete time-capture feature is the combination of these inputs and can be expressed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 101, + 422, + 295, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 422, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 101, + 422, + 295, + 440 + ], + "type": "interline_equation", + "content": "\\mathbf {c} = \\left[ \\mathbf {p} ^ {T}, \\log (i), \\log (s), f, \\mathbf {w} ^ {T} \\right] ^ {T}, \\tag {3}", + "image_path": "d7bca9e1493e0a3bcbc932890b6088af8b3e861b12d231a432d474d3a97601e3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 449, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 449, + 295, + 495 + ], + "type": "text", + "content": " represents optional noise-related features, which can include either noise stats " + }, + { + "bbox": [ + 55, + 449, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 55, + 449, + 295, + 495 + ], + "type": "text", + "content": ", SNR stats " + }, + { + "bbox": [ + 55, + 449, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 55, + 449, + 295, + 495 + ], + "type": "text", + "content": ", or both. Alternatively, " + }, + { + "bbox": [ + 55, + 449, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 449, + 295, + 495 + ], + "type": "text", + "content": " can be omitted if noise information is not considered." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "text", + "content": "Our time-capture feature is first normalized using min-max normalization (with the min and max values computed from the training data), and then processed through a learnable function, " + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 497, + 296, + 544 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 150, + 559, + 295, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 559, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 150, + 559, + 295, + 571 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {t} = f _ {t} (\\mathbf {c}), \\tag {4}", + "image_path": "1665ec0612b94f1ed7499e3dc23852647ee4af76c1aab4a422281530003d0a6b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "text", + "content": " is a learnable linear layer that transforms the time-capture feature, " + }, + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "text", + "content": ", into its latent representation, " + }, + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t\\in \\mathbb{R}^{16}" + }, + { + "bbox": [ + 55, + 578, + 295, + 615 + ], + "type": "text", + "content": " (see Fig. 3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 624, + 164, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 164, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 164, + 635 + ], + "type": "text", + "content": "2.2. Histogram feature" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "In addition to the time-capture feature, c, we provide our neural model with a histogram feature, " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " which represents the colors of the raw image. Inspired by prior work [6, 10, 11], we use a 2D histogram to represent the R/G and B/G chromaticities of the input raw image, " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{K \\times 3}" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " denotes the total number of pixels in the raw im" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "age. Specifically, we compute the 2D chromaticity histogram, " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_c\\in \\mathbb{R}^{h\\times h}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 373, + 106, + 553, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 106, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 373, + 106, + 553, + 129 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {c} ^ {(m, n)} = \\sum_ {k} \\| \\mathbf {I} ^ {(k)} \\| _ {2} \\cdot \\delta_ {m, n} ^ {(k)}, \\tag {5}", + "image_path": "5f1da0bcad99bd4e776f28480d4ba1601b2e86690adfad9c6cc68a91c3e24850.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 152, + 553, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 553, + 168 + ], + "type": "interline_equation", + "content": "\\delta_ {m, n} ^ {(k)} = \\left[ u _ {m} \\leq r g ^ {(k)} < u _ {m + 1} \\right] \\wedge \\left[ v _ {n} \\leq b g ^ {(k)} < v _ {n + 1} \\right], \\tag {6}", + "image_path": "ecf2b9adf2a056e66e0d2c7ecfc81d984473e44ebf8c25e2f912a546be13f7df.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "rg^{(k)}" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "bg^{(k)}" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " are the R/G and B/G chromaticity values of pixel " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " in I, and " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{I}^{(k)} \\|_2" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " represents the intensity of pixel " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " (i.e., Euclidean norm of the pixel's RGB values). The notation " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\wedge" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " denotes the logical AND operator, and " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "[\\cdot]" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " is the Iverson bracket, which evaluates to 1 when the condition is true and 0 otherwise. The histogram bins are defined by the edges " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\{u_m\\}" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\{v_n\\}" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "u_m" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "v_n" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " are the lower edges of the bins, and " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "u_{m+1}" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "v_{n+1}" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " are the corresponding upper edges. The histogram accumulates the brightness values of pixels whose chromaticity values fall within the range " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "[u_m, u_{m+1})" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " along the R/G axis (i.e., the " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": "-axis) and " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "[v_n, v_{n+1})" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": " along the B/G axis (i.e., the " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": "-axis). The number of histogram bins along each axis is denoted as " + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 177, + 555, + 357 + ], + "type": "text", + "content": ". Following [10], we compute the square root of the histogram to enhance the utility of the histogram feature [9]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 357, + 554, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 428 + ], + "type": "text", + "content": "Note that this histogram differs from the " + }, + { + "bbox": [ + 313, + 357, + 554, + 428 + ], + "type": "inline_equation", + "content": "\\log -uv" + }, + { + "bbox": [ + 313, + 357, + 554, + 428 + ], + "type": "text", + "content": " histogram used in prior work [6, 10, 11], which operates in the logarithmic space of G/R and G/B [25]. We found that the R/G, B/G chromaticity histogram performs better, as it aligns with our model's output space. See the supplemental material for an ablation study." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 429, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 554, + 487 + ], + "type": "text", + "content": "In addition to the chromaticity histogram, " + }, + { + "bbox": [ + 313, + 429, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_c" + }, + { + "bbox": [ + 313, + 429, + 554, + 487 + ], + "type": "text", + "content": ", of the image colors, and building on prior work [6, 11], we augment it with the square-rooted chromaticity histogram of the image's edges, " + }, + { + "bbox": [ + 313, + 429, + 554, + 487 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_e" + }, + { + "bbox": [ + 313, + 429, + 554, + 487 + ], + "type": "text", + "content": ", where the image edges are computed as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 342, + 507, + 553, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 507, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 342, + 507, + 553, + 536 + ], + "type": "interline_equation", + "content": "\\mathbf {E} ^ {(x, y)} = \\frac {1}{8} \\sum_ {\\Delta x, \\Delta y} \\left| \\mathbf {I} ^ {\\prime (x, y)} - \\mathbf {I} ^ {\\prime (x + \\Delta x, y + \\Delta y)} \\right|, \\tag {7}", + "image_path": "59d9c9e8e26f4f1abae1ff20f6d60c0a2fd33013f94fb95c0d9912f0c01e6986.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "text", + "content": " represents the image edges, " + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{I}'" + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "text", + "content": " refers to the raw image in 3D tensor form (height, width, channels), and " + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "inline_equation", + "content": "\\Delta x, \\Delta y \\in \\{-1,0,1\\}" + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "inline_equation", + "content": "(\\Delta x, \\Delta y) \\neq (0,0)" + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "text", + "content": ". The edge histogram, " + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_e" + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "text", + "content": ", is computed from " + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 546, + 553, + 594 + ], + "type": "text", + "content": " using Eqs. 5 and 6." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Our histogram feature is constructed by concatenating these two histograms, " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_c" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_e" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ". Since this histogram feature is first processed by convolutional (conv) layers, as shown in Fig. 3, we follow [6] by appending additional channels that encode the positional information of the " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "u / v" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " coordinates in histogram space, which helps capture spatial relationships within the histogram feature. Consequently, our final histogram feature, " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{H} \\in \\mathbb{R}^{h \\times h \\times 4}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ", consists of 2 channels representing the chromaticity of the image and its edges, along with the additional " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "u / v" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " coordinate channels." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5041" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "This feature is processed through a series of conv layers with ELU activation [21]. The resulting latent representation undergoes adaptive average pooling before passing through a linear layer to produce the histogram's latent feature vector, " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_h\\in \\mathbb{R}^{16}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 144, + 295, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 144, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 146, + 144, + 295, + 158 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {h} = f _ {h} (\\mathbf {H}), \\tag {8}", + "image_path": "e56a07539203be90410ee5569029d2b5f9d8dcf71f6feade1550aed261458142.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 296, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 200 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 163, + 296, + 200 + ], + "type": "inline_equation", + "content": "f_{h}" + }, + { + "bbox": [ + 55, + 163, + 296, + 200 + ], + "type": "text", + "content": " denotes the sub-model (i.e., conv layers, ELU activation, pooling, and linear layer) that maps the histogram feature into its latent space, as shown in Fig. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 207, + 180, + 219 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 180, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 180, + 219 + ], + "type": "text", + "content": "2.3. Illuminant estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "text", + "content": "Both feature vectors, " + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_h" + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "text", + "content": ", are concatenated to produce the latent vector " + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "text", + "content": " and processed by an illuminant estimation sub-model as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 146, + 273, + 295, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 273, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 146, + 273, + 295, + 287 + ], + "type": "interline_equation", + "content": "\\mathbf {v} = \\left[ \\mathbf {v} _ {t}; \\mathbf {v} _ {h} \\right], \\tag {9}", + "image_path": "d57994e8570507c709853d2a96d2a86f8454ee59e20d59d4c1b0bd8a05341233.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 150, + 304, + 295, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 304, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 150, + 304, + 295, + 317 + ], + "type": "interline_equation", + "content": "\\ell_ {c} = f _ {\\ell} (\\mathbf {v}), \\tag {10}", + "image_path": "3d95eec02088618908c578f43e12b34640099f11c7b936388b97ee7c4c973f91.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "inline_equation", + "content": "f_{\\ell}" + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "text", + "content": " consists of a set of linear layers, with batch normalization applied to the first layer, followed by activation functions—except for the final layer, which outputs " + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\ell_{c}" + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "text", + "content": ", the chromaticity vector of the scene illuminant. This vector is then transformed into an unnormalized RGB illuminant color by mapping " + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "inline_equation", + "content": "[\\mathrm{R / G},\\mathrm{B / G}]^T\\rightarrow [\\mathrm{R / G},1,\\mathrm{B / G}]^T" + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "text", + "content": ", followed by normalization via division by its L2 norm to produce the final illuminant color. We optimize " + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "inline_equation", + "content": "f_{t}" + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "inline_equation", + "content": "f_{h}" + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "inline_equation", + "content": "f_{\\ell}" + }, + { + "bbox": [ + 55, + 323, + 296, + 443 + ], + "type": "text", + "content": " to minimize the angular error [36] between the predicted RGB illuminant color and the ground-truth RGB illuminant color." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 454, + 111, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 111, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 111, + 466 + ], + "type": "text", + "content": "3. Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 665 + ], + "type": "text", + "content": "To train and validate our method, we require a dataset that includes contextual information (i.e., timestamp and geolocation) h h image. Existing datasets (e.g., [17, 19, 23, 28, 43, 44]) lack this essential information, motivating us to collect a new dataset using a smartphone camera that provides contextual metadata for each image. Specifically, we captured 3,224 linear raw images with the Samsung S24 Ultra's main camera, covering a wide range of scenes both indoors and outdoors, at various times of day (e.g., sunset, sunrise, noon, night). Our dataset includes images captured under various light sources (e.g., sunlight, incandescent, LED), as well as non-standard illuminant colors (e.g., colored LED light), which are not present in existing datasets [61]; see Fig. 4. Additionally, our dataset captures scenes under different weather conditions (sunny, cloudy, rain, snow, etc.). Example scenes can be found in the supplemental material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "We follow prior work [7, 43] in collecting ground-truth illuminant colors for each scene. Specifically, for each scene, we first capture an image with a calibration color chart, which is used to extract the illuminant color from the" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 75, + 552, + 251 + ], + "blocks": [ + { + "bbox": [ + 318, + 75, + 552, + 251 + ], + "lines": [ + { + "bbox": [ + 318, + 75, + 552, + 251 + ], + "spans": [ + { + "bbox": [ + 318, + 75, + 552, + 251 + ], + "type": "image", + "image_path": "37f97bafb85ba114e5ec6509bbbf0db7564b7396fd3d3680416eea20a931d984.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 264, + 555, + 320 + ], + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 320 + ], + "type": "text", + "content": "Figure 4. A plot of the " + }, + { + "bbox": [ + 313, + 264, + 555, + 320 + ], + "type": "inline_equation", + "content": "rg" + }, + { + "bbox": [ + 313, + 264, + 555, + 320 + ], + "type": "text", + "content": " chromaticity distribution of ground-truth illuminant colors for neutral white balance in our dataset. Example images in raw and sRGB spaces are shown (raw images are gamma-corrected for visualization). See supplemental material the " + }, + { + "bbox": [ + 313, + 264, + 555, + 320 + ], + "type": "inline_equation", + "content": "rg" + }, + { + "bbox": [ + 313, + 264, + 555, + 320 + ], + "type": "text", + "content": " chromaticity distribution of user-preference illuminants." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 337, + 554, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 554, + 421 + ], + "type": "text", + "content": "gray patches. Next, we capture an image of the same scene without the color chart (see Fig. 5), resulting in " + }, + { + "bbox": [ + 313, + 337, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\sim 6\\mathrm{K}" + }, + { + "bbox": [ + 313, + 337, + 554, + 421 + ], + "type": "text", + "content": " images. After obtaining the ground-truth illuminant color, we discard all color chart images, leaving us with 3,224 images in our dataset. This approach allows us to test with natural images that mimic real-world scenarios, without the need to mask out the color chart patch." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 423, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 591 + ], + "type": "text", + "content": "Since our dataset includes a wide variety of lighting conditions, such as sunset/sunrise, night scenes, and artificial light, we generate a \"user-preference\" ground truth in addition to the neutral ground truth obtained from the color chart for each scene in our dataset. This is to account for human incomplete chromatic adaptation in such scenes [50, 60] and user preference [20, 24]. Specifically, an expert photographer was asked to assign a ground truth illuminant to each scene to make it appear more natural, reflect real-world observations, and enhance the aesthetics of the image. Notably, the same person who captured the scene also performed the annotation, ensuring that the user-preference selection was based on real-world observations of how the scene should appear." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "We validated the user-preference ground truth through a study with 20 participants, who selected the most natural image from pairs of white-balanced images corresponding to the user-preference ground truth and the neutral ground truth (obtained from the color chart). The user-preference ground truth was selected in " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "71.95\\%" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": " of the trials, confirming its preference. See the supplemental material for additional details. In our experiments, both ground-truth types—neutral and user-preference—were used for training and evaluation (Sec. 4)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5042" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 155 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 155 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 155 + ], + "type": "image", + "image_path": "5cbfde5f3f2a93b14d0a501e0f559052dccf4bed8a822c976be0457aad762bfe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 161, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 55, + 161, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 555, + 206 + ], + "type": "text", + "content": "Figure 5. For each scene in our dataset, an image with a color chart placed in the scene was captured. The gray patches on the color chart were used to measure the ground-truth illuminant. These color-chart images were discarded, and only the images of the scenes without the color chart were used for the training, validation, and testing sets. For each example, we show both raw and sRGB images, with the raw images gamma-corrected to enhance visualization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 220, + 294, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 220, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 294, + 316 + ], + "type": "text", + "content": "Additionally, we created a mask for regions illuminated by non-dominant illuminants in each scene. These masks ensure that all scenes have only one dominant illuminant, matching the color of the neutral ground truth, without confounding effects from other illuminants. We applied these masks to the training images when training our method and others. To preserve privacy, sensitive information, such as car plates and faces, has been blurred." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 318, + 295, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 318, + 295, + 438 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 295, + 438 + ], + "type": "text", + "content": "In addition to raw images, both ground-truth illuminant types, and contextual and capture information, we provide additional valuable auxiliary data to broaden the impact of our dataset. This includes locally tone-mapped sRGB images rendered by an expert, that can serve as ground-truth for applications beyond white-balance correction, such as neural ISPs [35, 38, 39, 65]. More information about the dataset can be found in the supplemental material. We organize our dataset into 2,619 raw images for training, 205 raw images for validation, and 400 raw images for testing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 137, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 137, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 137, + 467 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "type": "text", + "content": "Implementation details: We train the model using the Adam optimizer [40], for 400 epochs with betas set to (0.9, 0.999) and a weight decay factor of " + }, + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "type": "inline_equation", + "content": "10^{-9}" + }, + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "type": "text", + "content": ". A warm-up strategy is applied to gradually increase the learning rate from " + }, + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "type": "text", + "content": " over the first 5 epochs. After this, we use a cosine annealing schedule [48]. Following [6], we employ a batch-size increment strategy during training, starting with a batch size of 8 and doubling it every 100 epochs. Following prior work [6, 7], we use images of size " + }, + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "type": "inline_equation", + "content": "384 \\times 256" + }, + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "type": "text", + "content": " pixels in all experiments, which is a reasonable size for camera pipelines to reduce computational overhead. For our method, we use histograms with " + }, + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "type": "inline_equation", + "content": "h = 48" + }, + { + "bbox": [ + 56, + 479, + 295, + 684 + ], + "type": "text", + "content": " bins. The histogram boundaries are determined by computing the 10th and 95th percentiles of the chromaticity values along each chromaticity axis from the training set. We report the results of our method both with and without the inclusion of noise stats, n, and SNR stats, r." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Results: We report the results of our method alongside several others benchmarked on the proposed dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 220, + 553, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 304 + ], + "type": "text", + "content": "Our method is compared with various statistical-based approaches [12, 14, 15, 18, 19, 26, 32, 56, 57, 62], camera-specific learning-based methods [4, 5, 11, 16, 29, 30, 37, 41, 42, 53, 59, 66, 67], and camera-independent techniques [2, 6, 13]. For the camera-specific learning-based approaches, including our method, all models were trained on our training set." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 305, + 554, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 401 + ], + "type": "text", + "content": "For the camera-independent methods [2, 6], we present results from models trained on the NUS [19] and Cube++ [23] datasets. Additionally, we report results from finetuned versions of these models [2, 6], incorporating our proposed dataset. We also evaluate camera-specific (CS) models of these methods [2, 6], which were trained exclusively on our training data, without any additional datasets. For further details, refer to the supplemental material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "content": "Table 1 shows the results on the testing set of our dataset. We report the mean, median, best " + }, + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "content": ", worst " + }, + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "content": ", worst " + }, + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 402, + 554, + 521 + ], + "type": "text", + "content": ", tri-mean, and maximum angular errors between the estimated illuminant colors and the ground truth colors for each method. Results are provided for both neutral and user-preference ground truth, with two models, one trained for each ground-truth type, except for the models SIIE [2], SIIE (tuned) [2], C5 [6], and C5 (tuned) [6], which were trained on images from the NUS [19] and Cube++ [23] datasets that do not include user-preference ground truth." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 640 + ], + "type": "text", + "content": "Additionally, we report the total number of parameters for methods that involve tunable or learnable parameters. The results in Table 1 are based on images without masking regions illuminated by a light source other than the dominant one used to obtain the ground truth, reflecting real-world scenarios where scenes with a single light source are not always guaranteed. For completeness, additional results are provided in the supplementary material, where regions illuminated by light sources different from the ground-truth illuminant color are masked out." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "To further demonstrate the effectiveness of our method on existing DSLR datasets, where contextual metadata is typically unavailable, we report our results on the Simple Cube++ dataset [23]. Although geolocation metadata is absent, the dataset provides capture settings such as ISO and exposure time from the DSLR camera ISP." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5043" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 110, + 557, + 514 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 106 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 106 + ], + "type": "text", + "content": "Table 1. Results on the testing set. We report the mean, median, best " + }, + { + "bbox": [ + 55, + 70, + 555, + 106 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 70, + 555, + 106 + ], + "type": "text", + "content": ", worst " + }, + { + "bbox": [ + 55, + 70, + 555, + 106 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 70, + 555, + 106 + ], + "type": "text", + "content": ", tri-mean, and maximum angular errors for each method on neutral and user-preference white-balance ground-truth illuminants, presented in the format (neutral / user-preference). Symbols n and r represent noise stats and SNR stats, respectively. The best and second-best results are highlighted." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 110, + 557, + 514 + ], + "lines": [ + { + "bbox": [ + 57, + 110, + 557, + 514 + ], + "spans": [ + { + "bbox": [ + 57, + 110, + 557, + 514 + ], + "type": "table", + "html": "
MethodMeanMed.Best 25%Worst 25%Worst 5%Tri.Max#params (K)
GW [15]6.23 / 5.546.01 / 4.521.02 / 1.0012.43 / 11.9118.05 / 19.815.76 / 4.6528.09 / 31.70-
SoG [26]4.45 / 3.593.54 / 2.170.74 / 0.679.61 / 8.8114.94 / 15.613.77 / 2.6223.22 / 29.36-
GE-1st [62]4.21 / 3.333.29 / 2.120.71 / 0.559.34 / 8.2915.34 / 15.193.49 / 2.3627.81 / 28.80-
GE-2nd [62]4.11 / 3.183.17 / 1.890.70 / 0.589.09 / 7.7914.90 / 14.853.35 / 2.1825.09 / 29.42-
Max-RGB [14]4.01 / 2.612.92 / 1.881.06 / 0.978.57 / 5.6914.01 / 11.073.30 / 1.9434.22 / 23.63-
wGE [32]3.96 / 3.072.96 / 1.910.62 / 0.498.95 / 7.7914.94 / 15.083.21 / 2.1031.33 / 30.54-
PCA [19]4.42 / 3.633.54 / 1.950.70 / 0.619.59 / 9.1914.46 / 16.163.76 / 2.5222.87 / 31.58-
MSGP [57]6.39 / 5.695.72 / 4.380.94 / 0.9813.39 / 12.7320.97 / 22.335.64 / 4.6837.25 / 36.92-
GI [56]4.70 / 4.933.19 / 2.970.44 / 0.7811.63 / 12.2220.48 / 21.353.48 / 3.5336.34 / 36.02-
TECC [12]4.12 / 3.173.23 / 1.910.74 / 0.559.10 / 7.8314.73 / 14.483.46 / 2.1727.08 / 28.87-
Gamut (pixels) [30]3.77 / 2.402.81 / 1.490.77 / 0.638.31 / 5.8412.93 / 11.183.16 / 1.6321.53 / 23.310.636
Gamut (edges) [30]4.45 / 3.943.52 / 3.041.08 / 1.019.51 / 8.5015.00 / 15.253.70 / 3.1828.55 / 29.46324
Gamut (1st) [30]4.10 / 3.653.08 / 2.670.72 / 0.909.26 / 8.2514.75 / 14.693.28 / 2.9121.38 / 26.35279
NIS [29]4.58 / 3.903.76 / 2.570.77 / 0.759.81 / 9.1014.70 / 15.663.87 / 2.9720.76 / 31.800.078
Classification-CC [53]2.73 / 1.611.98 / 1.160.61 / 0.366.03 / 3.609.37 / 6.072.18 / 1.2719.34 / 9.5358,384
FFCC [11]2.62 / 1.501.46 / 0.810.37 / 0.246.89 / 3.9916.59 / 8.431.66 / 0.9548.97 / 18.6012
FFCC (capture info) [11]2.31 / 1.351.38 / 0.800.34 / 0.235.82 / 3.5112.44 / 7.361.60 / 0.8847.67 / 16.9636.9
FC4 [37]3.80 / 2.652.78 / 2.250.85 / 0.858.61 / 5.1415.06 / 7.522.86 / 2.3725.74 / 11.421,705
APAP (GW) [5]3.74 / 2.093.14 / 1.670.93 / 0.497.72 / 4.4411.09 / 6.893.26 / 1.7616.43 / 9.020.289
SIIE [2]4.09 / -3.25 / -0.91 / -8.97 / -15.96 / -3.37 / -43.24 / -1,008
SIIE (tuned) [2]3.15 / -2.22 / -0.51 / -7.28 / -12.14 / -2.46 / -34.52 / -1,008
SIIE (tuned-CS) [2]3.14 / 1.742.20 / 1.200.50 / 0.327.39 / 4.0613.61 / 6.732.41 / 1.3038.96 / 9.441,008
KNN (raw) [4]2.44 / 1.411.51 / 0.830.36 / 0.206.13 / 3.6611.64 / 7.011.66 / 0.9328.95 / 13.49757
Quasi-U-CC [13]3.85 / 3.262.97 / 1.810.55 / 0.588.47 / 8.2413.30 / 15.783.25 / 2.2124.74 / 32.1754,421
Quasi-U-CC (tuned) [13]3.11 / 2.542.27 / 1.430.49 / 0.447.14 / 6.5711.53 / 13.622.44 / 1.6322.67 / 33.7454,421
BoCF [41]3.54 / 2.142.68 / 1.590.96 / 0.487.31 / 4.7411.40 / 7.852.96 / 1.7222.31 / 19.6859
C4 [66]1.92 / 1.491.30 / 0.900.36 / 0.244.64 / 3.829.08 / 7.531.40 / 1.0321.55 / 18.095,116
CWCC [42]3.65 / 2.302.71 / 1.720.82 / 0.677.96 / 4.9512.52 / 9.652.99 / 1.8118.66 / 20.81101
C5 [6]3.22 / -2.51 / -0.78 / -6.97 / -10.54 / -2.68 / -16.38 / -412
C5 (tuned) [6]1.91 / -1.24 / -0.38 / -4.57 / -8.43 / -1.38 / -17.22 / -412
C5 (tuned-CS) [6]1.95 / 1.251.32 / 0.840.37 / 0.234.72 / 2.948.15 / 4.981.44 / 0.9316.78 / 9.23172
TLCC [59]2.71 / 2.742.06 / 1.830.66 / 0.695.89 / 6.3010.30 / 12.902.17 / 1.9921.44 / 33.3332,910
PCC [67]3.03 / 1.672.13 / 1.200.53 / 0.407.08 / 3.7911.20 / 6.842.34 / 1.2716.82 / 12.330.378
RGP [18]4.59 / 4.563.13 / 2.810.43 / 0.7011.13 / 11.2718.79 / 20.253.53 / 3.3132.11 / 33.98-
CFCC [16]3.07 / 1.542.20 / 1.050.73 / 0.396.87 / 3.5512.55 / 6.982.36 / 1.1323.34 / 14.400.283
Ours (w/o n, w/o r)1.93 / 1.261.35 / 0.770.38 / 0.234.56 / 3.139.48 / 5.881.43 / 0.8822.63 / 15.904.83
Ours (w/o n, w/ r)1.89 / 1.231.18 / 0.790.32 / 0.244.74 / 3.0110.10 / 5.071.30 / 0.9024.99 / 12.914.93
Ours (w/ n, w/o r)1.87 / 1.201.24 / 0.720.37 / 0.244.58 / 2.999.82 / 5.581.30 / 0.8229.46 / 15.124.93
Ours (w/ n, w/ r)1.84 / 1.201.24 / 0.770.35 / 0.194.41 / 2.959.17 / 5.121.32 / 0.8735.42 / 12.715.03
", + "image_path": "c3a0513e498cf13147d7f2739646f9ba91aefcf5fd5565513ae77bf3d767306e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "text", + "content": "We trained our model using the histogram feature " + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "text", + "content": ", along with capture features including ISO " + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "text", + "content": ", exposure time " + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "inline_equation", + "content": "(e)" + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "text", + "content": ", and SNR stats " + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "inline_equation", + "content": "(\\mathbf{r})" + }, + { + "bbox": [ + 55, + 532, + 295, + 627 + ], + "type": "text", + "content": ", deliberately excluding the contextual metadata that is not available in the Simple Cube++ DSLR dataset. The results of our method, both with and without the SNR stats, are compared to other methods in Table 2. As shown, our method outperforms competing methods across most evaluation metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "Inference time: As shown in Table 1, our method performs comparably to or outperforms prior methods, while maintaining a compact model with only approximately 5K parameters. This lightweight design results in faster runtimes compared to other methods that achieve competitive results, namely C4 [66] and tuned C5 [6]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 530, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 555, + 598 + ], + "type": "text", + "content": "Table 2. Results on the testing set of the 'Simple Cube++' dataset [23]. We report the mean, median, best " + }, + { + "bbox": [ + 313, + 530, + 555, + 598 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 530, + 555, + 598 + ], + "type": "text", + "content": ", worst " + }, + { + "bbox": [ + 313, + 530, + 555, + 598 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 530, + 555, + 598 + ], + "type": "text", + "content": ", tri-mean, and maximum angular errors for each method. Symbols c, i, e, and r refer to the time-capture feature, ISO, exposure time, and SNR stats, respectively. The best and second-best results are highlighted." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 604, + 552, + 682 + ], + "blocks": [ + { + "bbox": [ + 317, + 604, + 552, + 682 + ], + "lines": [ + { + "bbox": [ + 317, + 604, + 552, + 682 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 552, + 682 + ], + "type": "table", + "html": "
MethodMeanMed.Best 25%Worst 25%Worst 5%Tri.Max
FFCC [11]1.380.580.184.0510.300.6845.20
FC4 [37]3.722.020.479.9317.182.4029.92
C4 [66]1.160.650.243.006.720.7313.95
C5 (tuned-CS) [6]1.190.600.183.267.610.6815.62
TLCC [59]1.821.150.394.499.631.2019.02
Ours (c = [i, e]T)1.080.590.172.845.530.6911.03
Ours (c = [i, e, rT]T)1.010.600.172.685.710.6510.89
", + "image_path": "d67481ecb5d5df94a46120793116406a89df2975219ce6c792742523988d2219.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5044" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 84, + 255, + 126 + ], + "blocks": [ + { + "bbox": [ + 55, + 57, + 295, + 78 + ], + "lines": [ + { + "bbox": [ + 55, + 57, + 295, + 78 + ], + "spans": [ + { + "bbox": [ + 55, + 57, + 295, + 78 + ], + "type": "text", + "content": "Table 3. Processing time on an AMD Ryzen Threadripper PRO 3975WX CPU460 and an NVIDIA RTX A6000 GPU." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 99, + 84, + 255, + 126 + ], + "lines": [ + { + "bbox": [ + 99, + 84, + 255, + 126 + ], + "spans": [ + { + "bbox": [ + 99, + 84, + 255, + 126 + ], + "type": "table", + "html": "
MethodCPU (ms)GPU (ms)FLOPs
C4 [66]49.0311.282.28G
C5 [6]8.284.50103.54M
Ours0.550.4516.78M
", + "image_path": "8c50fac65325ec32da390d5d62047bb19702b1c3549669fed5e38994e8d777da.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 295, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 246 + ], + "type": "text", + "content": "In Table 3, we present the processing times of our model, C4, and C5 on an AMD Ryzen Threadripper PRO 3975WX CPU and an NVIDIA RTX A6000 GPU. Our fast runtime performance is particularly well-suited for mobile camera ISPs, where limited computational latency is crucial due to the processing demands of other modules (e.g., denoising, local tone mapping) per frame. Our model runs in just 0.25 ms on the DSP and 0.80 ms on the CPU of the Samsung S24 Ultra." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "text", + "content": "Ablation studies: We conducted ablation studies to evaluate the impact of each input feature on the validation set of our dataset. Specifically, we assessed our method by excluding either the time-capture feature (c) or the histogram feature " + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "inline_equation", + "content": "(\\mathbf{H})" + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "text", + "content": ". Additionally, we tested the method using only the time feature " + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "inline_equation", + "content": "(\\mathbf{p})" + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "text", + "content": " and noise stats " + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "inline_equation", + "content": "(\\mathbf{n})" + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "text", + "content": ", while excluding the histogram feature " + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "inline_equation", + "content": "(\\mathbf{H})" + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "text", + "content": ". We also evaluated the method using the histogram feature along with the time feature and noise stats, excluding other capture information. Furthermore, we examined the accuracy of our method using all input features except for " + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "text", + "content": ". Lastly, we present the results when all input features are used, but without the noise stats " + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "inline_equation", + "content": "(\\mathbf{n})" + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "text", + "content": " and the SNR stats " + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "inline_equation", + "content": "(\\mathbf{r})" + }, + { + "bbox": [ + 55, + 252, + 295, + 418 + ], + "type": "text", + "content": ". These results are presented in Table 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "text", + "content": "As shown in Table 4, using only the time feature " + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "inline_equation", + "content": "(\\mathbf{p})" + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "text", + "content": " and noise stats " + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "inline_equation", + "content": "(\\mathbf{n})" + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "text", + "content": " yields a reasonable accuracy " + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "inline_equation", + "content": "(2.37^{\\circ}" + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "text", + "content": " mean angular error), compared to FFCC [11], which achieves " + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "inline_equation", + "content": "2.19^{\\circ}" + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "text", + "content": " mean angular error on the validation set. This validates the usefulness of time-of-day and noise information in providing contextual clues. As expected, incorporating the histogram feature " + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "inline_equation", + "content": "(\\mathbf{H})" + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "text", + "content": ", which represents scene colors, along with the time-of-day and noise stats, significantly boosts accuracy, as demonstrated by the results in the fourth row. We further investigate the impact of using noise information in the last three rows: using SNR stats " + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "inline_equation", + "content": "(\\mathbf{r})" + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "text", + "content": ", using noise stats " + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "inline_equation", + "content": "(\\mathbf{n})" + }, + { + "bbox": [ + 55, + 419, + 295, + 587 + ], + "type": "text", + "content": ", or both. The combination of both noise features yields the lowest mean angular error. Additional ablation studies are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 597, + 208, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 208, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 208, + 609 + ], + "type": "text", + "content": "5. Conclusion and limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "We presented a method for in-camera AWB correction that leverages contextual information. Specifically, we proposed a lightweight model that leverages contextual metadata (notably time-of-day derived from timestamp and geolocation) to guide the illuminant estimation process. In addition to this contextual metadata, we incorporate image colors in the form of histogram features, as well as capture information such as ISO, shutter speed, and noise stats, to help" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "content": "Table 4. Results of ablation studies on the validation set with neutral white-balance ground truth. We report the mean, median, best " + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "content": ", and worst " + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "content": " angular errors for our method with various configurations. Symbols " + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "content": " denote the histogram feature, time feature, noise stats, and SNR stats, respectively. The symbol " + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "content": " refers to the capture information feature, which includes ISO " + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "content": ", shutter speed " + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "inline_equation", + "content": "(s)" + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "content": ", and flash status " + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "inline_equation", + "content": "(f)" + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "content": ", all of which are used in our input time-capture feature " + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 71, + 553, + 170 + ], + "type": "text", + "content": ". The best results are highlighted." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 171, + 548, + 284 + ], + "blocks": [ + { + "bbox": [ + 318, + 171, + 548, + 284 + ], + "lines": [ + { + "bbox": [ + 318, + 171, + 548, + 284 + ], + "spans": [ + { + "bbox": [ + 318, + 171, + 548, + 284 + ], + "type": "table", + "html": "
HmpnrMeanMed.Best 25%Worst 25%Worst 5%
✓XXXX2.031.360.314.908.90
X✓✓✓✓2.291.810.504.958.58
XX✓✓X2.371.590.455.489.03
✓X✓✓X1.751.170.294.137.26
✓✓X✓✓1.891.210.314.609.41
✓✓✓XX1.851.320.354.267.53
✓✓✓X✓1.721.160.304.138.16
✓✓✓✓X1.671.070.294.047.90
✓✓✓✓✓1.661.200.333.776.95
", + "image_path": "d7635400aa35b730e3020735d8ff1d7a4800b2ce4cec98a514dce17f89ecc031.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 553, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 359 + ], + "type": "text", + "content": "the model distinguish between artificial and natural scenes and improve its final accuracy. Our method is fast and can achieve high frame rates on modern smartphone DSPs and CPUs, while maintaining accurate illuminant color estimation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 363, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 554, + 495 + ], + "type": "text", + "content": "A key contribution of this work is a large-scale dataset of raw images captured by a consumer smartphone camera, along with the necessary contextual metadata for training and evaluating our method. Beyond the traditional neutral white-balance ground truth extracted from a calibration color chart placed in each scene, we also include a user-preference ground truth that targets the observer's preference, validated through a user study. Results based on both ground truth types demonstrate that our method achieves comparable or superior performance to existing methods, which require larger models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 713 + ], + "type": "text", + "content": "While our method represents a promising solution for mobile camera ISPs, its dependency on contextual metadata limits its optimal accuracy to devices that provide geolocation data, which may not be available on DSLR cameras. Additionally, while the contextual metadata is device-independent (i.e., the differences across devices are expected to be minimal), our method relies on additional capture information (i.e., ISO, shutter speed, noise, and SNR stats) and image colors, making it inherently camera-dependent in design. This dependency prevents our trained model from generalizing to new devices without fine-tuning or re-training. However, this issue can be mitigated through color calibration (e.g., [46]), which can be extended to calibrate capture information as well—by performing a preprocessing mapping from the new camera space (for both color and capture information) to the camera space used during training. Additional discussion on cross-camera generalization is provided in the supplementary material." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5045" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Abdelrahman Abdelhamed, Abhijith Punnappurath, and Michael S Brown. Leveraging the availability of two cameras for illuminant estimation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 157 + ], + "type": "text", + "content": "[2] Mahmoud Affifi and Michael S Brown. Sensor-independent illumination estimation for DNN models. In BMVC, 2019. 1, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 294, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 294, + 180 + ], + "type": "text", + "content": "[3] Mahmoud Affifi and Michael S Brown. Deep white-balance editing. In CVPR, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 182, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 294, + 215 + ], + "type": "text", + "content": "[4] Mahmoud Affifi, Brian Price, Scott Cohen, and Michael S Brown. When color constancy goes wrong: Correcting improperly white-balanced images. In CVPR, 2019. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 217, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 294, + 258 + ], + "type": "text", + "content": "[5] Mahmoud Affifi, Abhijith Punnappurath, Graham Finlayson, and Michael S Brown. As-projective-as-possible bias correction for illumination estimation algorithms. JOSA A, 36 (1):71-78, 2019. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 262, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 294, + 293 + ], + "type": "text", + "content": "[6] Mahmoud Afifi, Jonathan T Barron, Chloe LeGendre, Yun-Ta Tsai, and Francois Bleibel. Cross-camera convolutional color constancy. In ICCV, 2021. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 296, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 327 + ], + "type": "text", + "content": "[7] Mahmoud Affifi, Zhenhua Hu, and Liang Liang. Optimizing illuminant estimation in dual-exposure HDR imaging. In ECCV, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 329, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 294, + 371 + ], + "type": "text", + "content": "[8] Vivek Agarwal, Besma R Abidi, Andreas Koschan, and Mongi A Abidi. An overview of color constancy algorithms. Journal of Pattern Recognition Research, 1(1):42-54, 2006. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 374, + 294, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 294, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 294, + 405 + ], + "type": "text", + "content": "[9] Relja Arandjelović and Andrew Zisserman. Three things everyone should know to improve object retrieval. In CVPR, 2012. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 408, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 408, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 57, + 408, + 294, + 429 + ], + "type": "text", + "content": "[10] Jonathan T Barron. Convolutional color constancy. In ICCV, 2015. 1, 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 431, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 431, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 57, + 431, + 294, + 452 + ], + "type": "text", + "content": "[11] Jonathan T Barron and Yun-Ta Tsai. Fast Fourier color constancy. In CVPR, 2017. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 454, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 454, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 57, + 454, + 294, + 475 + ], + "type": "text", + "content": "[12] Simone Bianco and Marco Buzzelli. Truncated edge-based color constancy. In ICCE, 2022. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 477, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 294, + 498 + ], + "type": "text", + "content": "[13] Simone Bianco and Claudio Cusano. Quasi-unsupervised color constancy. In CVPR, 2019. 1, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 500, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 500, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 57, + 500, + 294, + 532 + ], + "type": "text", + "content": "[14] David H Brainard and Brian A Wandell. Analysis of the Retinex theory of color vision. Journal of the Optical Society of America A, 3(10):1651-1661, 1986. 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 534, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 534, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 57, + 534, + 294, + 565 + ], + "type": "text", + "content": "[15] Gershon Buchsbaum. A spatial processor model for object colour perception. Journal of the Franklin Institute, 310(1): 1-26, 1980. 1, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 568, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 294, + 601 + ], + "type": "text", + "content": "[16] Marco Buzzelli and Simone Bianco. A convolutional framework for color constancy. IEEE Transactions on Neural Networks and Learning Systems, pages 1-15, 2024. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 602, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 634 + ], + "type": "text", + "content": "[17] Cheng Cheng, Kai-Fu Yang, Xue-Mei Wan, Leanne Lai Hang Chan, and Yong-Jie Li. Nighttime color constancy using robust gray pixels. JOSA A, 41(3):476-488, 2024. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 57, + 636, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 668 + ], + "type": "text", + "content": "[18] Cheng Cheng, Kai Fu Yang, Xue Mei Wan, Leanne Lai Hang Chan, and Yonge Jie Li. Nighttime color constancy using robust gray pixels. JOsA A, 41(3):476-488, 2024. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[19] Dongliang Cheng, Dilip K Prasad, and Michael S Brown. Illuminant estimation for color constancy: Why spatial-domain methods work and the role of the color distribution. JOSA A, 31(5):1049-1058, 2014. 2, 5, 6, 7" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[20] Kyungah Choi and Hyeon-Jeong Suk. User-preferred color temperature adjustment for smartphone display under varying illuminants. Optical Engineering, 53(6):061708-061708, 2014. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 119, + 554, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 554, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 554, + 151 + ], + "type": "text", + "content": "[21] Djork-Arné Clevert. Fast and accurate deep network learning by exponential linear units (ELUs). arXiv preprint arXiv:1511.07289, 2015. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 154, + 554, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 154, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 154, + 554, + 195 + ], + "type": "text", + "content": "[22] Maurizio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Mobile computational photography: A tour. Annual Review of Vision Science, 7(1):571-604, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 199, + 554, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 199, + 554, + 253 + ], + "spans": [ + { + "bbox": [ + 317, + 199, + 554, + 253 + ], + "type": "text", + "content": "[23] Egor Ershov, Alexey Savchik, Illya Semenkov, Nikola Banić, Alexander Belokopytov, Daria Senshina, Karlo Koščević, Marko Subašić, and Sven Lončarić. The cube++ illumination estimation dataset. IEEE Access, 8:227511-227527, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 256, + 554, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 256, + 554, + 309 + ], + "spans": [ + { + "bbox": [ + 317, + 256, + 554, + 309 + ], + "type": "text", + "content": "[24] Egor Ershov, Alex Savchik, Denis Shepelev, Nikola Banić, Michael S Brown, Radu Timofte, Karlo Koščević, Michael Freeman, Vasily Tesalin, Dmitry Bocharov, et al. NTIRE 2022 challenge on night photography rendering. In CVPRW, 2022. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 312, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 312, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 553, + 334 + ], + "type": "text", + "content": "[25] Graham D Finlayson and Steven D Hordley. Color constancy at a pixel. JOSA A, 18(2):253-264, 2001. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 337, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 337, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 553, + 357 + ], + "type": "text", + "content": "[26] Graham D Finlayson and Elisabetta Trezzi. Shades of gray and colour constancy. In CIC, 2004. 1, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 361, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 392 + ], + "type": "text", + "content": "[27] Karl R Gegenfurtner, David Weiss, and Marina Bloj. Color constancy in real-world settings. Journal of Vision, 24(2): 12-12, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 395, + 553, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 395, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 317, + 395, + 553, + 427 + ], + "type": "text", + "content": "[28] Peter Vincent Gehler, Carsten Rother, Andrew Blake, Tom Minka, and Toby Sharp. Bayesian color constancy revisited. In CVPR, 2008. 2, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 430, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 430, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 317, + 430, + 553, + 472 + ], + "type": "text", + "content": "[29] Arjan Gijsenij and Theo Gevers. Color constancy using natural image statistics and scene semantics. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(4): 687-698, 2011. 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 475, + 553, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 475, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 317, + 475, + 553, + 519 + ], + "type": "text", + "content": "[30] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Generalized gamut mapping using image derivative structures for color constancy. International Journal of Computer Vision, 86(2):127-139, 2010. 2, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 521, + 553, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 521, + 553, + 563 + ], + "spans": [ + { + "bbox": [ + 317, + 521, + 553, + 563 + ], + "type": "text", + "content": "[31] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Computational color constancy: Survey and experiments. IEEE Transactions on Image Processing, 20(9):2475-2489, 2011. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "type": "text", + "content": "[32] Arjan Gijsenij, Theo Gevers, and Joost Van De Weijer. Improving color constancy by photometric edge weighting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):918-929, 2011. 1, 6, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "type": "text", + "content": "[33] Thorsten Hansen, Sebastian Walter, and Karl R Gegenfurtner. Effects of spatial and temporal context on color categories and color constancy. Journal of Vision, 7(4):2-2, 2007. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[34] Samuel W Hasinoff, Dillon Sharlet, Ryan Geiss, Andrew Adams, Jonathan T Barron, Florian Kainz, Jiawen Chen, and Marc Levoy. Burst photography for high dynamic range and low-light imaging on mobile cameras. ACM Transactions on Graphics, 35(6):1-12, 2016. 3" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5046" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[35] Xuanhua He, Tao Hu, Guoli Wang, Zejin Wang, Run Wang, Qian Zhang, Keyu Yan, Ziyi Chen, Rui Li, Chengjun Xie, et al. Enhancing RAW-to-sRGB with decoupled style structure in Fourier domain. In AAAI, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 140 + ], + "type": "text", + "content": "[36] Steven D Hordley and Graham D Finlayson. Re-evaluating colour constancy algorithms. In ICPR, 2004. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 294, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 175 + ], + "type": "text", + "content": "[37] Yuanming Hu, Baoyuan Wang, and Stephen Lin. FC4: Fully convolutional color constancy with confidence-weighted pooling. In CVPR, 2017. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 177, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 294, + 232 + ], + "type": "text", + "content": "[38] Andrey Ignatov, Anastasia Sycheva, Radu Timofte, Yu Tseng, Yu-Syuan Xu, Po-Hsiang Yu, Cheng-Ming Chiang, Hsien-Kai Kuo, Min-Hung Chen, Chia-Ming Cheng, et al. MicroISP: processing 32MP photos on mobile devices with deep learning. In ECCV, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 233, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 294, + 275 + ], + "type": "text", + "content": "[39] Woohyeok Kim, Geonu Kim, Junyong Lee, Seungyong Lee, Seung-Hwan Baek, and Sunghyun Cho. ParamISP: Learned forward and inverse ISPs using camera parameters. In CVPR, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 278, + 294, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 294, + 311 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 294, + 311 + ], + "type": "text", + "content": "[40] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 313, + 294, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 313, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 294, + 367 + ], + "type": "text", + "content": "[41] Firas Laakom, Nikolaos Passalis, Jenni Raitoharju, Jarno Nikkanen, Anastasios Tefas, Alexandros Iosifidis, and Moncef Gabbouj. Bag of color features for color constancy. IEEE Transactions on Image Processing, 29:7722-7734, 2020. 1, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 369, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 369, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 294, + 402 + ], + "type": "text", + "content": "[42] Firas Laakom, Jenni Raitoharju, Jarno Nikkanen, Alexandros Iosifidis, and Moncef Gabbouj. Robust channel-wise illumination estimation. In BMVC, 2021. 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "text", + "content": "[43] Firas Laakom, Jenni Raitoharju, Jarno Nikkanen, Alexandros Iosifidis, and Moncef Gabbouj. Intel-TAU: A color constancy dataset. IEEE Access, 9:39560-39567, 2021. 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 438, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 460 + ], + "type": "text", + "content": "[44] Shuwei Li and Robby T Tan. NightCC: Nighttime color constancy via adaptive channel masking. In CVPR, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 462, + 294, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 294, + 516 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 294, + 516 + ], + "type": "text", + "content": "[45] Orly Liba, Kiran Murthy, Yun-Ta Tsai, Tim Brooks, Tianfan Xue, Nikhil Karnad, Qiurui He, Jonathan T Barron, Dillon Sharlet, Ryan Geiss, et al. Handheld mobile photography in very low light. ACM Transactions on Graphics, 38(6):164-1, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 518, + 294, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 518, + 294, + 551 + ], + "spans": [ + { + "bbox": [ + 56, + 518, + 294, + 551 + ], + "type": "text", + "content": "[46] Yi-Tun Lin, Bianjiang Yang, Hao Xie, Wenbin Wang, Honghong Peng, and Jun Hu. Color constancy: How to deal with camera bias? In BMVC, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 553, + 294, + 584 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 553, + 294, + 584 + ], + "spans": [ + { + "bbox": [ + 56, + 553, + 294, + 584 + ], + "type": "text", + "content": "[47] Ce Liu, William T Freeman, Richard Szeliski, and Sing Bing Kang. Noise estimation from a single image. In CVPR, 2006. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 587, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 294, + 620 + ], + "type": "text", + "content": "[48] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 622, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 622, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 622, + 294, + 654 + ], + "type": "text", + "content": "[49] Zhongyu Lou, Theo Gevers, Ninghang Hu, Marcel P Lucassen, et al. Color constancy by deep learning. In BMVC, 2015. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "text", + "content": "[50] Ming Luo. A review of chromatic adaptation transforms. Review of Progress in Coloration and Related Topics, 30:77 - 92, 2008. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "text", + "content": "[51] Laurence T Maloney. Illuminant estimation as cue combination. Journal of Vision, 2(6):6-6, 2002. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 625 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "type": "text", + "content": "[52] Jean H Meeus. *Astronomical algorithms*. Willmann-Bell, Incorporated, 1991. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[53] Seoung Wug Oh and Seon Joo Kim. Approaching the computational color constancy as a classification problem through deep learning. Pattern Recognition, 61:405-416, 2017. 1, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[54] Varad A Pimpalkhute, Rutvik Page, Ashwin Kothari, Kishor M Bhurchandi, and Vipin Milind Kamble. Digital image noise estimation using DWT coefficients. IEEE Transactions on Image Processing, 30:1962-1972, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "text", + "content": "[55] Stanislav Pyatykh, Jürgen Hesser, and Lei Zheng. Image noise level estimation by principal component analysis. IEEE Transactions on Image Processing, 22(2):687-699, 2012. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 252 + ], + "type": "text", + "content": "[56] Yanlin Qian, Joni-Kristian Kamarainen, Jarno Nikkanen, and Jiri Matas. On finding gray pixels. In CVPR, 2019. 1, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[57] Yanlin Qian, Said Pertuz, Jarno Nikkanen, Joni-Kristian Kämäräinen, and Jiri Matas. Revisiting gray pixel for statistical illumination estimation. In VISSAP. 2019. 1, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[58] Wu Shi, Chen Change Loy, and Xiaou Tang. Deep specialized network for illuminant estimation. In ECCV, pages 371-387, 2016. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "text", + "content": "[59] Yuxiang Tang, Xuejing Kang, Chunxiao Li, Zhaowen Lin, and Anlong Ming. Transfer learning for color constancy via statistic perspective. In AAAI, 2022. 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "text", + "content": "[60] Shoji Tominaga, Takahiko Horiuchi, Shiori Nakajima, and Mariko Yano. Prediction of incomplete chromatic adaptation under illuminant A from images. In CIC, 2014. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 389, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 410 + ], + "type": "text", + "content": "[61] Oguzhan Ulucan, Diclehan Ulucan, and Marc Ebner. Color constancy beyond standard illuminants. In ICIP, 2022. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 412, + 553, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 444 + ], + "type": "text", + "content": "[62] Joost Van De Weijer, Theo Gevers, and Arjan Gijsenij. Edge-based color constancy. IEEE Transactions on Image Processing, 16(9):2207-2214, 2007. 1, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 446, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 446, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 446, + 553, + 478 + ], + "type": "text", + "content": "[63] TC Van Flandern and KF Pulkkinen. Low-precision formulae for planetary positions. Astrophysical Journal Supplement Series, 41:391-411, 1979. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 480, + 553, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 480, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 316, + 480, + 553, + 501 + ], + "type": "text", + "content": "[64] Robert Walraven. Calculating the position of the sun. Solar energy, 20(5):393-397, 1978. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 502, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 553, + 524 + ], + "type": "text", + "content": "[65] Yazhou Xing, Zian Qian, and Qifeng Chen. Invertible image signal processing. In CVPR, 2021. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 525, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 558 + ], + "type": "text", + "content": "[66] Huanglin Yu, Ke Chen, Kaiqi Wang, Yanlin Qian, Zhaoxiang Zhang, and Kui Jia. Cascading convolutional color constancy. In AAAI, 2020. 1, 6, 7, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 559, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 591 + ], + "type": "text", + "content": "[67] Shuwei Yue and Minchen Wei. Color constancy from a pure color view. Journal of the Optical Society of America A, 40 (3):602-610, 2023. 6, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 593, + 553, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 625 + ], + "type": "text", + "content": "[68] Yi Zhang, Hongwei Qin, Xiaogang Wang, and Hongsheng Li. Rethinking noise synthesis and modeling in raw denoising. In ICCV, 2021. 4" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5047" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/b7999e96-8abf-47f1-9b7e-e5c9222126c0_content_list.json b/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/b7999e96-8abf-47f1-9b7e-e5c9222126c0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2c7fc08ea76ce7538236e476d0237557066d4fda --- /dev/null +++ b/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/b7999e96-8abf-47f1-9b7e-e5c9222126c0_content_list.json @@ -0,0 +1,1657 @@ +[ + { + "type": "text", + "text": "TimeBooth: Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging", + "text_level": 1, + "bbox": [ + 125, + 128, + 872, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zepeng Su $^{1}$ , Zhulin Liu $^{1,2,*}$ , Zongyan Zhang $^{1,2}$ , Tong Zhang $^{1,2,3,4}$ , C. L. Philip Chen $^{1,2,3,4}$", + "bbox": [ + 153, + 200, + 841, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ South China University of Technology $^{2}$ Pazhou Lab", + "bbox": [ + 333, + 223, + 665, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Guangdong Provincial Key Laboratory of Computational AI Models and Cognitive Intelligence", + "bbox": [ + 210, + 239, + 785, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Engineering Research Center of the Ministry of Education on Health Intelligent Perception and Paralleled Digital-Human", + "bbox": [ + 133, + 258, + 857, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8637756c8197d5ae7f632050474a8555d9d09e5610dcaeccdfc6e19f47922052.jpg", + "image_caption": [ + "Nemo", + "(Source)" + ], + "image_footnote": [], + "bbox": [ + 94, + 300, + 171, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b571f28f5bed9f37e3aeeb3c5b74bea5a8db47ae8bb0b7e783684137c023373.jpg", + "image_caption": [ + "Film Mr. Nobody" + ], + "image_footnote": [], + "bbox": [ + 192, + 295, + 514, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c08c451133fe62cb2cbc55a076a6abbbf4c2c284c0401b38ad93a09096a17c71.jpg", + "image_caption": [ + "Figure 1. In Mr. Nobody, the protagonist Nemo traverses multiple parallel realities, showcasing distinct aged appearances. This illustrates how choices such as environmental factors and lifestyle influence the diversity of face aging. Inspired by this, we propose TimeBooth." + ], + "image_footnote": [], + "bbox": [ + 524, + 282, + 901, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 463, + 326, + 478 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face aging is a typical ill-posed problem influenced by various factors such as environment and genetics, leading to highly diverse outcomes. However, existing methods primarily rely on numerical age representations, making it difficult to accurately capture individual or group-level aging patterns. To address this, we introduce a novel disentangled face representation, where age features are modeled in the image modality—referred to as the Age Prompt—providing richer prior age information to constrain the generation results. To this end, we design an ID-age multi-task colearning framework and propose the Bidirectional Adversarial Disentanglement(BAD) strategy. This strategy maximizes the disentanglement of ID and age representation through bidirectional adversarial learning, extracting their attribute-invariant representations. Based on this representation, we propose TimeBooth, a personalized face aging model capable of generating diverse and individualized aging results. To optimize training, we construct a cross-age hybrid data pipeline and introduce various training strategies. Finally, we propose the R-AgeMAE metric and validate our method through extensive experiments, demonstrating that TimeBooth outperforms existing methods in both diversity and controllability. Our code will be publicly available at https://github.com/zpsuu/TimeBooth.", + "bbox": [ + 88, + 493, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 463, + 643, + 478 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, large-scale text-to-image generation models [27, 30, 32, 37, 46, 47] have demonstrated remarkable controllable generation capabilities, transforming imagination into reality. Face aging, also known as face age transformation, enables the synthesis of photorealistic face images at different ages by modifying age-related facial attributes while preserving identity-consistent features, such as facial structure, contour and so on. This technology has broad applications, including film post-production, forensic age progression for missing persons, and medical diagnostics. With the advancement of personalized generation techniques, face aging can achieve more precise and individualized simulations, further enhancing its real-world applicability.", + "bbox": [ + 511, + 489, + 906, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Face aging is a typical ill-posed problem, influenced by various factors [16, 31, 38] such as environment, genetics, lifestyle, and living conditions. For instance, individuals of the same age may exhibit distinct aging patterns—prolonged sun exposure can lead to increased freckles and skin laxity, while those carrying the MCR1 gene are more sensitive to light and prone to developing wrinkles. However, existing face aging methods [1, 13, 28, 41, 43] are predominantly based on GANs [15] and follow an image-to-image translation paradigm. These methods typically use age as the sole constraint—age, resulting in generated images that merely transfer the average aging characteristics learned from the dataset. Consequently, the existing", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Zhulin Liu (e-mail: liuzhl@scut.edu.cn).", + "bbox": [ + 107, + 886, + 452, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12147", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1c5756e0033b565dcc753b7919a3dbe7b6bff2063846509972980517a0916649.jpg", + "image_caption": [ + "Figure 2. Training pipelines of our proposed TimeBooth. It consists of two stages: (a) Through ID-age collaborative training, we learn ID and age feature representations and introduce the BAD strategy, where bidirectional adversarial disentanglement based on mutual GRL layers is applied to extract attribute-invariant features. $\\lambda_{1}$ and $\\lambda_{2}$ are positive numbers greater than 0, and (b) The disentangled encoder is frozen. Based on the learned disentangled representation, fusion is performed in the Age Adapter, followed by image cross-attention layers, which are used with a diffusion model for generation. Multiple training strategies are optimized concurrently during this process." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "paradigm can only provide a seemingly reasonable solution to this ill-posed problem but fails to accurately capture the diversity of individual or group-level aging patterns.", + "bbox": [ + 88, + 395, + 480, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fundamentally, the limited expressiveness of existing paradigms stems from the inadequate representation of age as a constraint. As a numerical modality, age is inherently an abstract and discrete label, making it insufficient to capture the diversity and complexity of facial features at different ages. Consequently, it fails to provide the model with adequate semantic guidance for diverse and controllable age transformation. Therefore, it is essential to introduce a constraint with a richer age representation that properly regularizes the ill-posed nature of face aging, leading to more precise and well-defined solutions for personalized and diverse face aging.", + "bbox": [ + 88, + 444, + 482, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As the saying goes, \"a picture is worth a thousand words.\" As a highly concretized modality, images naturally provide rich and intuitive information. Inspired by this, we propose a novel disentangled face representation, where the ID-invariant feature and age-related feature are disentangled, and age is modeled using an image modality, referred to as the Age Prompt. To this end, we construct a multi-task collaborative learning framework, leveraging ID recognition and age classification tasks for joint supervision. We propose a bidirectional adversarial disentanglement strategy, which introduces mutual gradient reversal layers to maximize feature disentanglement, thereby learning attribute-invariant representations. Based on this representation, we further design an Age Adapter, in which a lightweight Q-former is employed to fuse the disentangled features, while the image cross-attention layers inject facial information into the diffusion model for generation. To train the Age Adapter, we build a cross-age hybrid dataset", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pipeline and propose multiple training strategies to ensure the correct alignment of disentangled features. Finally, we introduce the R-AgeMAE as a metric to evaluate the accuracy of personalized face aging. Our contributions can be summarized as follows:", + "bbox": [ + 511, + 395, + 906, + 470 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel disentangled facial representation, where age constraints are modeled using an image modality instead of the traditional numerical modality. To the best of our knowledge, this is the first attempt to represent age using image prompts, providing a richer age prior for the inherently ill-posed problem of face aging and introducing a new generative paradigm.", + "- We construct an ID-age multi-task co-learning framework and propose a bidirectional adversarial disentanglement strategy to maximize the disentanglement of ID and age representations, enabling the learning of their attribute-invariant representations.", + "- We introduce TimeBooth, which enables users to generate more diverse and personalized aging face images. To enhance training, we construct a hybrid cross-age dataset and design multiple training strategies to optimize the Age Adapter.", + "- We introduce the R-AgeMAE metric to evaluate the accuracy of personalized face aging. Extensive experiments demonstrate the superiority of TimeBooth in terms of diversity, and controllability. Moreover, TimeBooth supports plug-and-play integration, seamlessly incorporating into existing community models to provide efficient personalized face aging generation in just a few seconds without test-time optimizing." + ], + "bbox": [ + 511, + 502, + 903, + 878 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12148", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e88a88ede282596d013a42efef7b51842bd9097fc820489b301a10370cf6034f.jpg", + "image_caption": [ + "Figure 3. Inference pipeline of TimeBooth." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 472, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 383, + 233, + 400 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Face Aging methods", + "text_level": 1, + "bbox": [ + 89, + 409, + 282, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, most face age transformation methods [1, 13, 28, 41, 43] have been GAN-based [15], following an image-to-image generation paradigm where age serves as a conditioning factor. Many approaches [2, 19, 41, 42] utilize pre-trained face recognition models and ID-preserving loss terms to maintain facial identity, while others achieve transformation via style transfer. CUSP [13] disentangles content and style encoding, using a content mask for customizable editing. With the rise of diffusion models [27, 32], recent methods have explored their potential for face age transformation. Inspired by [35], [3] introduces specific identifiers and an age regularization set to fine-tune diffusion models. FADING [7] applies Prompt-to-Prompt [18] local editing, modifying attention maps during generation. Additionally, general image editing approaches based on attention mechanisms [5, 18] can also perform face age transformation by leveraging pre-trained text-to-image priors, though with potentially lower accuracy.", + "bbox": [ + 89, + 431, + 483, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Personalization methods", + "text_level": 1, + "bbox": [ + 89, + 712, + 316, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although large-scale text-to-image diffusion models [27, 30, 32, 37] possess rich prior knowledge, they are limited to generating content from their pre-trained corpus and struggle to meet personalized user needs. Researchers have explored personalized generation methods. One method [36] involves test-time fine-tuning, where reference images are used to adapt diffusion models at inference, often with LoRA [20] for efficiency. Another method [40, 44, 45] leverages adapter-based architectures, learning feature representations to preserve identity without test-time training. Many approaches also incorporate recognition-based [26]", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/efe4e89776967d16ab3f7debebf29683e81dd772ee42c2134e2840957af43e72.jpg", + "image_caption": [ + "Figure 4. Visualization of disentangled representation." + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 885, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "embeddings into the generation process. IP-Adapter-Faceid [44] combines face embeddings with CLIP [29] image encoders and employs disentangled cross-attention for generation. PhotoMaker [25] stacks ID embeddings for higher fidelity but struggles with fewer reference images. InstantID [40], built on IP-Adapter [44], introduces a facial keypoint control branch to improve shape fidelity at the cost of editability. FaceChain-FACT [45] utilizes a pre-trained ViT face recognition model [10] and a gated cross-attention layer to enhance instruction-following without compromising identity preservation.", + "bbox": [ + 511, + 295, + 906, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 513, + 479, + 612, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 506, + 653, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent Diffusion Model. Latent diffusion models [32] are a class of generative models. During training, noise is progressively added to the latent representation of the original image, $z_0$ , to obtain $z_t$ through a forward process. This process can be formulated as follows:", + "bbox": [ + 511, + 530, + 906, + 604 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\varepsilon \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 611, + 903, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The denoising network needs to accept the time step $t$ and the condition $c$ to estimate the noise. After multiple steps of denoising, the latent representation of target image $z_0$ can be obtained, which can be expressed by the formula:", + "bbox": [ + 511, + 633, + 906, + 694 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t - 1} = \\frac {\\sqrt {\\alpha_ {t - 1}}}{\\sqrt {\\alpha_ {t}}} z _ {t} + \\sqrt {\\alpha_ {t - 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\varepsilon_ {\\theta} \\left(z _ {t}, t, c\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 699, + 906, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, we estimate the noise added in the forward process and use the denoising network for denoising in the reverse process. The objective function for training the denoising network is:", + "bbox": [ + 511, + 752, + 906, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D M} = E _ {z _ {0}, \\varepsilon \\sim \\mathcal {N} (0, \\mathcal {I}), c, t} \\| \\varepsilon - \\varepsilon_ {\\theta} (z _ {t}, c, t) \\| ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 816, + 906, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-Prompt Adapter. The image adapter [44] is designed to enable a pre-trained text-to-image diffusion model to generate images conditioned on image prompts. It consists of an image encoder and a cross-attention module,", + "bbox": [ + 511, + 840, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12149", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e26ee12c76d07491d3f24314d4df334d96866ba8f73235cdd922099501aadcd8.jpg", + "image_caption": [ + "Figure 5. TimeBooth is capable of capturing fine-grained aging patterns for both individuals (left) and groups (right). As the number of Age Prompts increases, the generated results gradually shift from individual diversity to a stable group-level aging pattern, enhancing the interpretability and diversity of face aging synthesis in a data-driven manner." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 903, + 310 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f8c2ae0e525c888b2a39699b561dc11eadab54b408b939b1ce35e95db605da22.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison with translation-based methods CUSP [14], IP2P [6], and FADING [8]. Compared to these methods, TimeBooth generates faces that are both more natural and diverse, effectively avoiding the issue of 'child faces with elderly textures.'" + ], + "image_footnote": [], + "bbox": [ + 112, + 376, + 888, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the encoded image prompts are fed into the cross-attention layers to guide image generation. The adapter is trained using the following loss function:", + "bbox": [ + 89, + 611, + 482, + 656 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I A} = E _ {z _ {0}, \\varepsilon \\sim \\mathcal {N} (0, \\mathcal {I}), c _ {t}, c _ {i}, t} \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, c _ {t}, c _ {i}, t\\right) \\| ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 660, + 482, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{t}, c_{i}$ are text and image conditions after encoding text prompt and image prompt.", + "bbox": [ + 89, + 684, + 483, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Disentangled Representation Encoder", + "text_level": 1, + "bbox": [ + 89, + 726, + 416, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ID-Age Representation Co-learning Framework. We divide facial representation into two components: the ID representation $c_{id}$ , which remain invariant across age transformations, and the age representation $c_{age}$ , which change with age. Based on previous work [45], using the output of a ViT-based face recognition model [10] as face representations has proven effective in achieving high-fidelity reconstructions in diffusion models. Following this idea, we adopt a similar method. However, we further differentiate the original output of the ViT model. Specifically, we append two", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "separate neck networks to refine and disentangle the original features. The specific model details can be found in the Appendix.", + "bbox": [ + 511, + 611, + 906, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One neck network is designed to extract ID age-invariant representation $c_{id}$ . For this, we use the CosFace loss [39] $\\mathcal{L}_{\\mathrm{COSFACE}}$ for supervision:", + "bbox": [ + 511, + 659, + 906, + 705 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {I D}} = \\mathcal {L} _ {\\text {C O S F A C E}} \\left(L \\left(X _ {\\mathrm {i d}}\\right), y _ {\\mathrm {i d}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 710, + 903, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $X_{\\mathrm{id}}$ represents the extracted ID features, $L(\\cdot)$ denotes the head network, $y_{\\mathrm{id}}$ is the corresponding ID label.", + "bbox": [ + 511, + 731, + 903, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meanwhile, the other neck network is used to extract age features $c_{age}$ . We use the mse loss as the supervision loss:", + "bbox": [ + 511, + 763, + 906, + 795 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {A g e}} = \\mathcal {L} _ {\\mathrm {M S E}} (\\mathrm {D E X} (A (X _ {\\mathrm {a g e}})), y _ {\\mathrm {a g e}}) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 800, + 903, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $X_{\\mathrm{age}}$ represents the extracted age features, $A(\\cdot)$ is the head network, and $\\mathrm{DEX}(\\cdot)$ is the deep expectation [33, 34], $y_{\\mathrm{age}}$ is the ground-truth age label.", + "bbox": [ + 511, + 821, + 903, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bidirectional Adversarial Disentanglement(BAD) Strategy for Disentangling Representation. Supervised co", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12150", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "learning will result in overlap between ID and age representation. To enable cross-age face synthesis by modifying only $c_{age}$ , we aim to maximize the disentanglement of these two representations. Typically, face recognition features encapsulate all identifiable attributes of an individual, including age-related characteristics. However, such entanglement compromises feature purity, thereby affecting controllability during generation. Therefore, inspired by [22], we seek to learn an age-invariant ID representation that remains stable across varying age, eliminating the influence of age on ID recognition. To achieve this, we introduce an adversarial learning strategy to enhance representation disentanglement. Specifically, we incorporate a Gradient Reversal Layer (GRL) [12] before the age classification head. During backpropagation, the GRL reverses the gradient, forcing the ID representation to become age-invariant while retaining identity-specific information. The adversarial constraint is imposed through the following loss function:", + "bbox": [ + 89, + 90, + 483, + 363 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {A d v I D}} = \\mathcal {L} _ {\\mathrm {C O S F A C E}} \\left(L \\left(\\operatorname {G R L} \\left(X _ {\\text {a g e}}\\right)\\right), y _ {\\mathrm {i d}}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 368, + 482, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{GRL}(\\cdot)$ is the Gradient Reversal Layer.", + "bbox": [ + 89, + 390, + 397, + 405 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, to prevent ID information from being encoded in age features—ensuring that interpolation at inference time does not compromise the fidelity of generated faces—we introduce an adversarial constraint on the age feature extraction branch. Specifically, we incorporate a Gradient Reversal Layer (GRL) before the ID classification head, forcing the extracted age representation to become ID-invariant. The loss function is:", + "bbox": [ + 89, + 406, + 483, + 526 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {A d v A g e}} = \\mathcal {L} _ {\\text {M S E}} (\\operatorname {D E X} (A (\\operatorname {G R L} (X _ {\\mathrm {i d}}))), y _ {\\text {a g e}}) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 532, + 482, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By leveraging this bidirectional adversarial learning strategy, we achieve more effective disentanglement of ID and age representations. The total loss of the disentangled encoder is as follows:", + "bbox": [ + 89, + 555, + 483, + 614 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D R E}} = \\mathcal {L} _ {\\mathrm {I D}} + \\lambda_ {\\mathrm {A g e}} \\mathcal {L} _ {\\mathrm {A g e}} + \\lambda_ {\\mathrm {A d v I D}} \\mathcal {L} _ {\\mathrm {A d v I D}} + \\lambda_ {\\mathrm {A d v A g e}} \\mathcal {L} _ {\\mathrm {A d v A g e}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 621, + 482, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{Age}}$ , $\\lambda_{\\mathrm{AdvID}}$ , and $\\lambda_{\\mathrm{AdvAge}}$ are weighting coefficients balancing each loss term.", + "bbox": [ + 89, + 652, + 482, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Face Reconstruction via Age Adapter", + "text_level": 1, + "bbox": [ + 89, + 695, + 415, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve personalized face aging, we first perform face reconstruction during training. We propose the Age Adapter, which integrates the Age-Conditioned ID Encoder and multiple image cross-attention modules.", + "bbox": [ + 89, + 718, + 482, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Age-conditioned Face Encoding. To fuse disentangled representations into a complete ID representation with age information, we propose the Age-Conditioned ID Encoder. It begins by applying dual cross attention modules to facilitate interaction between $c_{id}$ and $c_{age}$ , which are derived from the disentangled representation encoder. This step ensures that $c_{id}$ are enriched with necessary age-related representation $c_{age}$ , while the $c_{age}$ is simultaneously refined to", + "bbox": [ + 89, + 780, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "align with the ID representation $c_{id}$ . Once the age-enhanced ID representation and ID-conditioned age representation are obtained, we employ a lightweight Q-Former [24] to adaptively learn the final comprehensive facial representation. Specifically, we introduce 32 learnable tokens that interact with the processed feature representations, allowing the model to capture a holistic and semantically rich ID representation while incorporating the essential age attributes. These learnable tokens dynamically adjust their latent space embeddings, ensuring that the final identity feature faithfully preserves the original identity while seamlessly incorporating the desired age characteristics.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Face Reconstruction. To enable face reconstruction using the fused representations, we integrate multiple image cross attention layers within a text-to-image model to effectively extract and incorporate information from the Age-conditioned ID Encoder $\\mathcal{E}$ , as formulated below:", + "bbox": [ + 511, + 272, + 906, + 347 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ = \\text {A t t e n t i o n} _ {\\text {t e x t}} (Q, K, V) + \\lambda \\cdot \\text {A t t e n t i o n} _ {\\text {i m g}} (Q, K ^ {\\prime}, V ^ {\\prime}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 353, + 911, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $\\lambda$ is weight coefficient of image cross attention layers, $Q = ZW_{q}, K = c_{t}W_{k}, V = c_{t}W_{v}, K' = \\mathcal{E}(c_{id}, c_{age})W'_k, V' = \\mathcal{E}(c_{id}, c_{age})W'_v$ are the query, key, and values matrices from the text and face representation.", + "bbox": [ + 511, + 383, + 905, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we optimize the Age Adapter and the unfrozen components of the Age-conditioned ID Encoder via a reconstruction task, with the objective defined as follows:", + "bbox": [ + 511, + 445, + 905, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D M} = E _ {z _ {0}, \\varepsilon \\sim \\mathcal {N} (0, \\mathcal {I}), c _ {t}, c _ {i d}, c _ {a g e}, t} \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, c _ {t}, \\mathcal {E} \\left(c _ {i d}, c _ {a g e}\\right), t\\right) \\| ^ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 494, + 903, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Strategies for Age Adapter", + "text_level": 1, + "bbox": [ + 511, + 534, + 826, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Shuffle Strategy. To ensure that the age-invariant ID representation $\\mathbf{c}_{\\mathrm{id}}$ of the same identity maintains strong intraclass consistency in the feature space, we propose the Shuffle Strategy and design a cross-age data pipeline. During training, with probability $p_s$ , we dynamically replace the ID feature $\\mathbf{c}_{\\mathrm{id}}^i$ of a sample with that of the same identity at a different age, formulated as:", + "bbox": [ + 511, + 555, + 905, + 661 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {c} _ {\\mathrm {i d}} ^ {i} = \\mathbb {I} \\cdot \\mathbf {c} _ {\\mathrm {i d}} ^ {j} + (1 - \\mathbb {I}) \\cdot c _ {\\mathrm {i d}} ^ {i}, \\mathbb {I} \\sim \\operatorname {B e r n o u l l i} \\left(p _ {s}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 666, + 903, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathbb{I}$ follows a Bernoulli distribution with parameter $p_s$ , and $j$ represents another image with the same identity but a different age ( $j \\neq i, y_j = y_i$ ). This strategy not only enhances the learning effectiveness of the Age Adapter but also prevents overfitting due to direct reconstruction of the same image, thereby improving model generalization.", + "bbox": [ + 511, + 689, + 905, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Age Prototype Replacement(APR) Strategy. To prevent the Age Prompt from capturing spurious age-unrelated correlations and to leverage population-level statistical regularities of age attributes, we propose the APR strategy. We first extract age features $\\{\\mathbf{c}_{\\mathrm{age}}^i\\}_{i = 1}^N$ from all training samples and perform offline K-means clustering [17] (precomputed before training) to group them into $K$ cohorts. Each cohort's prototype vector $\\mu_{\\mathrm{age}}^k$ encapsulates the average age patterns", + "bbox": [ + 511, + 780, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12151", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of its members. During training, the original feature $\\mathbf{c}_{\\mathrm{age}}^i$ of each sample is randomly replaced with its cohort prototype $\\mu_{\\mathrm{age}}^{k(i)}$ with probability $p_a$ , effectively suppressing noisy individual-specific extremes:", + "bbox": [ + 89, + 90, + 480, + 154 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {c} _ {\\text {a g e}} ^ {i} = \\mathbb {I} \\cdot \\mu_ {\\text {a g e}} ^ {k (i)} + (1 - \\mathbb {I}) \\cdot c _ {\\text {a g e}} ^ {i}, \\quad \\mathbb {I} \\sim \\operatorname {B e r n o u l l i} \\left(p _ {a}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 157, + 482, + 179 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamic Training-set Sampling(DTS) Strategy. Current face image datasets often exhibit a long-tailed age distribution, where training on raw data may cause feature overfitting in well-represented age groups and suboptimal learning in underrepresented ones. To address this, we propose the DTS strategy. Before training, age labels are discretized into equal intervals. During training, adaptive sample selection is performed based on the statistical median: for age groups exceeding the median sample count, only a representative subset matching the median is retained per epoch, while all samples are used for underrepresented groups. This strategy improves face generation across age ranges and mitigates data imbalance.", + "bbox": [ + 89, + 196, + 483, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Drop Strategy. Similar to [44, 45], we randomly mask $c_{id}$ and $c_{age}$ with probability $p_d$ during training. During inference, the denoising process is formulated as follow:", + "bbox": [ + 89, + 393, + 483, + 439 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {t - 1} = \\left(1 + \\lambda_ {\\mathrm {C F G}}\\right) \\cdot \\mathcal {H} \\left(z _ {t}, t, c _ {\\text {t e x t}}, \\mathcal {E} \\left(c _ {i d}, c _ {a g e}\\right)\\right) \\\\ - \\lambda_ {\\text {C F G}} \\mathcal {H} \\left(z _ {t}, t, c _ {\\text {t e x t - n e g}}, \\mathcal {E} \\left(c _ {i d - d r o p}, c _ {a g e - d r o p}\\right)\\right) \\tag {14} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 445, + 480, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{H}$ represents the forward process for denoising U-Net and $\\lambda_{\\mathrm{CFG}}$ is the CFG scale.", + "bbox": [ + 89, + 492, + 482, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Personalized Face Aging during Inference", + "text_level": 1, + "bbox": [ + 89, + 534, + 449, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the inference phase (see Fig. 3), we employ the disentangled representation encoder to extract $c_{id}$ and $c_{age}^{src}$ from the source image. To achieve more controllable age transformation, we propose a fusion strategy based on Spherical Linear Interpolation. This method interpolates between the source age feature $c_{age}^{src}$ and the target Age Prompt $c_{age}^{tgt}$ on a unit hypersphere, generating a fused age feature $c_{age}^{*}$ with the hyperparameter aging guidance $\\alpha$ controlling the intervention intensity of the target attribute:", + "bbox": [ + 89, + 556, + 483, + 691 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nc _ {a g e} ^ {*} = \\frac {\\sin [ (1 - \\alpha) \\theta ]}{\\sin \\theta} c _ {a g e} ^ {s r c} + \\frac {\\sin [ \\alpha \\theta ]}{\\sin \\theta} c _ {a g e} ^ {t g t}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 696, + 482, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\theta = \\arccos \\left(c_{age}^{src} \\cdot c_{age}^{tgt}\\right)$ is the angular distance between feature vectors.", + "bbox": [ + 89, + 731, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, the cross-age face images are synthesized with identity consistency by combining $c_{id}$ and $c_{age}^*$ .", + "bbox": [ + 89, + 762, + 483, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 806, + 223, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiments Setup", + "text_level": 1, + "bbox": [ + 89, + 832, + 272, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset construction. In the first training stage, we utilize the MS1Mv2 [11] dataset for ID recognition and age estimation, with age annotations provided by [21]. In the", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "second stage, we construct a hybrid dataset containing approximately 500K high-quality facial images. This hybrid dataset consists of two parts: one is a general crossage facial dataset without identity overlap, composed of FFHQ [23] and SFHQ [4]; the other is a cross-age dataset with the same identity across different ages, derived from VGGFace2-HQ [9], based on which we establish a crossage identity pool. For more details on dataset construction, please refer to the Appendix.", + "bbox": [ + 511, + 90, + 905, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. Please refer to the Appendix.", + "text_level": 1, + "bbox": [ + 511, + 228, + 826, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline. To the best of our knowledge, there has been no method that directly utilizes age information as image prompts for face generation. Therefore, a direct comparison with existing personalized face aging methods is not feasible. To provide a reasonable baseline, we have constructed several baseline methods based on current common used face generation methods, including DreamBooth [36], FaceID [44], and FaceID+DreamBooth. Although these methods are not specifically designed for face aging, with appropriate configuration and adjustments, they can achieve a certain degree of control over face aging as image prompts. Detailed experimental configurations and parameter settings can be found in the appendix. Additionally, we selected several state-of-the-art methods [6, 8, 14] from the past in the field of face aging for comparison. These methods are based on image translation methods, which, while differing from our task, share a similar objective.", + "bbox": [ + 511, + 244, + 906, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. To evaluate personalized face aging performance, we employ seven quantitative metrics: Face Similarity (Face Sim.) for ID preservation, Kernel Inception Distance (KID) for distribution fidelity, Age Mean Absolute Error (AgeMAE) for age accuracy, and Relative AgeMAE (R-AgeMAE), our proposed metric, which specifically measures Age Prompt transfer accuracy. Additionally, we assess Face Quality, per-image inference time, and conduct a multidimensional user study for subjective perception. More details are provided in the Appendix.", + "bbox": [ + 511, + 501, + 905, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation of Age Prompt. Unless otherwise specified, the Age Prompts used are all sampled from FFHQ [23]. The sampling procedure follows three main rules: (1) The images are selected from the target age group, (2) the images must match the race and gender of the source face, and (3) dozens of target age images are sampled to ensure the use of prior knowledge from the population, which helps stabilize the results and ensure consistency.", + "bbox": [ + 511, + 652, + 906, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 787, + 696, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with translation-based methods. Time-Booth generates more diverse results compared to comparison methods, as illustrated in Fig. 6. Although translation-based approaches may appear to offer higher fidelity, the results generated by the comparison methods resemble 'children with elderly textures,' creating a strong sense of incon", + "bbox": [ + 511, + 809, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12152", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6c1d8cffc2e0a93ae1cb86c3e363a29ac687ead2b65841f28b0727e83cd539fd.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of lifespan synthesis. TimeBooth generates more natural images with high fidelity and consistency." + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 883, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0888239f85217be4045950af35edaa5c79b609a488f364b95808231ac0f5249f.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison. We selected in-the-wild celebrity photos and transformed them to the corresponding age in the GroundTruth. By transferring ID age-invariant features and leveraging the prior knowledge of Age Prompts, our results are superior." + ], + "image_footnote": [], + "bbox": [ + 116, + 324, + 883, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gruity, and do not resemble real human faces. In reality, both facial textures and bone structure undergo significant transformations from childhood to old age. By disentangling features and using Age Prompts as guidance, Time-Booth produces faces with semantics more aligned with the real aging process. Fig. 7 demonstrates that Time-Booth achieves a comparable level of cross-age consistency to the comparison methods, indicating accurate preservation of ID-invariant age features. Additionally, as shown in Fig. 9, Time-Booth enables more flexible and personalized generation, which translation-based methods cannot achieve.", + "bbox": [ + 88, + 526, + 482, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with baseline methods. As shown in Fig. 10, although the faces generated after fine-tuning with DreamBooth reflect the age features in the Age Prompt, their fidelity is relatively low. Furthermore, DreamBooth requires fine-tuning, which incurs higher computational and time costs. In contrast, our method is tuning-free, and the generated images can be produced in just a few seconds. For more qualitative experiments, please refer to the Appendix.", + "bbox": [ + 89, + 694, + 482, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 830, + 284, + 847 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted a quantitative evaluation to verify Time-Booth's effectiveness. In Table 1, while TimeBooth inherently differs from translation-based methods by generating", + "bbox": [ + 89, + 854, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "images from scratch, it achieves superior performance on several metrics. Furthermore, it outperforms baseline methods on most metrics, highlighting its advantages.", + "bbox": [ + 511, + 526, + 906, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. We conducted a user study to compare our method against baseline approaches, evaluating three key metrics: facial consistency, Age Prompt similarity, and visual quality. As shown in Table 1, our method outperforms the baselines across all three metrics.", + "bbox": [ + 511, + 573, + 906, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 662, + 666, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Training Strategy. In Table 2, we present an ablation study on various training strategies. The results indicate that removing the Shuffle Strategy or DTS leads to a significant decrease in R-AgeMAE, suggesting a weakened ability to extract age-related information from the Age Prompt. Additionally, removing PAR or DTS results in a drop in KID and Face Sim., implying a decline in fidelity.", + "bbox": [ + 511, + 686, + 906, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Aging Guidance $\\alpha$ . In Eq. 15, we introduce $\\alpha$ to control the influence of Age Prompt features on the target image. To evaluate its effect, we vary $\\alpha$ from 0 to 1, and the generated results are shown in Fig. 11. As $\\alpha$ increases, the facial features of the generated image gradually approach those of the Age Prompt, effectively simulating the aging process. Conversely, decreasing $\\alpha$ makes the generated im", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12153", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4529f45afefc8f603d3413bea29fa2531b4801758b7112bda5c5c8b60c2d9d1.jpg", + "image_caption": [ + "Figure 9. Qualitative comparison of face aging. TimeBooth demonstrates more flexible generation capabilities." + ], + "image_footnote": [], + "bbox": [ + 116, + 89, + 883, + 294 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/37f492601473b310c32cc0dc2065a7d4dbc9c21a4d5e5e14e958f891bf07c030.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodYoung-to-OldYoung-to-ChildUser StudyFace QualityInference Time
KID(X100) ↓Face-sim ↑AgeMAE ↓R-AgeMAE ↓KID(X100) ↓Face-sim ↑AgeMAE ↓R-AgeMAE ↓Face Sim.(%)Sim. to A.P.(%)Visual Quality(%)
CUSP5.300.7446.08-8.080.8208.99-28.6-32.968.05<10 s
FADING4.800.5946.96-4.290.8687.05-21.4-18.668.173 min
IP2P9.300.49127.92-7.170.78820.63-18.6-12.965.86<10 s
FaceID13.040.63021.5022.5914.020.7267.385.7127.120.045.752.48<10 s
DreamBooth12.960.6745.335.7615.820.7436.845.6510.045.725.762.9510 min
FaceID+DreamBooth9.750.65125.0424.607.730.7289.788.2031.454.358.654.366 min
Ours7.200.7314.523.539.760.7638.557.9778.677.175.872.20<10 s
", + "bbox": [ + 93, + 333, + 903, + 439 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2971bf21b6c7a8c4308221de961d16e549685e869af443ad788b0e1d38d43314.jpg", + "image_caption": [ + "Figure 10. Qualitative comparison with baseline methods." + ], + "image_footnote": [], + "bbox": [ + 94, + 500, + 480, + 664 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d95727e286b84c8263c99d186b9d1150b012f95f3685f34f5c02727792854760.jpg", + "table_caption": [ + "Table 1. Quantitative comparison across various aspects. Blue indicates the best result among translation-based methods, bold denotes the best among baseline methods, and bold blue represents the best overall." + ], + "table_footnote": [], + "table_body": "
MethodKID (×100) ↓Face Sim. ↑R-AgeMAE ↓
w/o Shuffle4.850.7288.62
w/o PAR6.800.7194.11
w/o DTS7.400.7248.37
Full4.250.7334.58
", + "bbox": [ + 140, + 710, + 431, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation study.", + "bbox": [ + 212, + 791, + 359, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "age more similar to the source image. Given the editability-fidelity trade-off, selecting an appropriate $\\alpha$ is crucial for achieving a balanced result. Empirical studies suggest that an $\\alpha$ value between 0.7 and 0.9 provides the best balance.", + "bbox": [ + 89, + 840, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5cec05d691f3a5854927beeb96c25d41f2bd3e4857c392d87610f314f2eacd9.jpg", + "image_caption": [ + "Figure 11. Visualization under different aging guidance $\\alpha$ : as $\\alpha$ increases, the age features from the Age Prompt are gradually injected into the source face." + ], + "image_footnote": [], + "bbox": [ + 526, + 502, + 893, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 691, + 633, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel disentangled face representation that, for the first time, models age using image prompts, providing richer priors for face aging. Based on this, we introduce TimeBooth, which surpasses existing methods by generating more personalized and diverse aged face images, establishing a new paradigm for face aging. Moreover, TimeBooth supports plug-and-play integration and requires no test-time fine-tuning, offering greater flexibility and higher computational efficiency. In the future, we will further explore invariant face representations and extend their applications to broader and more general face generation tasks.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12154", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was funded in part by the National Natural Science Foundation of China grant under number 62222603, in part by the STI2030-Major Projects grant from the Ministry of Science and Technology of the People's Republic of China under number 2021ZD0200700, in part by the Key-Area Research and Development Program of Guangdong Province under number 2023B0303030001, and in part by the Program for Guangdong Introducing Innovative and Entrepreneurial Teams (2019ZT08X214).", + "bbox": [ + 89, + 114, + 485, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 263, + 187, + 277 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Only a matter of style: age transformation using a style-based regression model. ACM Trans. Graph., 40(4):45:1-45:12, 2021. 1, 3", + "[2] Grigory Antipov, Moez Baccouche, and Jean-Luc Dugelay. Face aging with conditional generative adversarial networks. In 2017 IEEE International Conference on Image Processing, ICIP 2017, Beijing, China, September 17-20, 2017, pages 2089-2093. IEEE, 2017. 3", + "[3] Sudipta Banerjee, Govind Mittal, Ameya Joshi, Chinmay Hegde, and Nasir D. Memon. Identity-preserving aging of face images via latent diffusion models. In IEEE International Joint Conference on Biometrics, IJCB 2023, Ljubljana, Slovenia, September 25-28, 2023, pages 1-10. IEEE, 2023. 3", + "[4] David Beniaguev. Synthetic faces high quality (sfhq) dataset, 2022. 6", + "[5] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18392-18402. IEEE, 2023. 3", + "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 4, 6", + "[7] Xiangyi Chen and Stephane Lathuilière. Face aging via diffusion-based editing. In 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023, page 595. BMVA Press, 2023. 3", + "[8] Xiangyi Chen and Stéphane Lathuilière. Face aging via diffusion-based editing. arXiv preprint arXiv:2309.11321, 2023. 4, 6", + "[9] Xuanhong Chen, Bingbing Ni, Yutian Liu, Naiyuan Liu, Zhilin Zeng, and Hang Wang. Simswap++: Towards faster and high-quality identity swapping. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):576-592, 2023. 6", + "[10] Jun Dan, Yang Liu, Haoyu Xie, Jiankang Deng, Haoran Xie, Xuansong Xie, and Baigui Sun. Transface: Calibrating transformer training for face recognition from a data-centric perspective. In Proceedings of the IEEE/CVF international conference on computer vision, pages 20642-20653, 2023. 3, 4" + ], + "bbox": [ + 93, + 287, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6", + "[12] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 5", + "[13] Guillermo Gomez-Trenado, Stéphane Lathuilière, Pablo Mesejo, and Oscar Cordón. Custom structure preservation in face aging. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 565-580. Springer, 2022. 1, 3", + "[14] Guillermo Gomez-Trenado, Stéphane Lathuilière, Pablo Mesejo, and Oscar Cordón. Custom structure preservation in face aging. In European Conference on Computer Vision, pages 565-580. Springer, 2022. 4, 6", + "[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139–144, 2020. 1, 3", + "[16] Bahman Guyuron, David J Rowe, Adam Bryce Weinfeld, Yashar Eshraghi, Amir Fathi, and Seree Iamphongsai. Factors contributing to the facial aging of identical twins. Plastic and reconstructive surgery, 123(4):1321-1331, 2009. 1", + "[17] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics), 28(1):100-108, 1979. 5", + "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3", + "[19] Gee-Sern Hsu, Rui-Cang Xie, and Zhi-Ting Chen. Wasserstein divergence GAN with cross-age identity expert and attribute retainer for facial age transformation. IEEE Access, 9:39695-39706, 2021. 3", + "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3", + "[21] Zhizhong Huang, Junping Zhang, and Hongming Shan. When age-invariant face recognition meets face age synthesis: A multi-task learning framework. In CVPR, 2021. 6", + "[22] Zhizhong Huang, Junping Zhang, and Hongming Shan. When age-invariant face recognition meets face age synthesis: A multi-task learning framework. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7282-7291, 2021. 5", + "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6", + "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + ], + "bbox": [ + 516, + 92, + 906, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12155", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5", + "[25] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8640-8650, 2024. 3", + "[26] Haiqi Liu, C. L. Philip Chen, Xinrong Gong, and Tong Zhang. Robust saliency-aware distillation for few-shot fine-grained visual recognition. IEEE Transactions on Multimedia, 26:7529-7542, 2024. 3", + "[27] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, pages 16784-16804. PMLR, 2022. 1, 3", + "[28] Roy Or-El, Soumyadip Sengupta, Ohad Fried, Eli Shechtman, and Ira Kemelmacher-Shlizerman. Lifespan age transformation synthesis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI, pages 739-755. Springer, 2020. 1, 3", + "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3", + "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 1, 3", + "[31] Helle Rexbye, Inge Petersen, Mette Johansens, Louise Klitkou, Bernard Jeune, and Kaare Christensen. Influence of environmental factors on facial ageing. Age and Ageing, 35 (2):110-115, 2006. 1", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 1, 3", + "[33] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Dex: Deep expectation of apparent age from a single image. In Proceedings of the IEEE international conference on computer vision workshops, pages 10-15, 2015. 4", + "[34] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Deep expectation of real and apparent age from a single image without facial landmarks. International Journal of Computer Vision, 126(2):144-157, 2018. 4", + "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC," + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Canada, June 17-24, 2023, pages 22500-22510. IEEE, 2023. 3", + "[36] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3, 6", + "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. 1, 3", + "[38] Kestutis Sveikata, Irena Balciuniene, Janina Tutkuviene, et al. Factors influencing face aging. literature review. Stomatologija, 13(4):113-116, 2011. 1", + "[39] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 4", + "[40] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3", + "[41] Zongwei Wang, Xu Tang, Weixin Luo, and Shenghua Gao. Face aging with identity-preserved conditional generative adversarial networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 7939-7947. Computer Vision Foundation / IEEE Computer Society, 2018. 1, 3", + "[42] Hongyu Yang, Di Huang, Yunhong Wang, and Anil K. Jain. Learning continuous face age progression: A pyramid of gans. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):499-515, 2021. 3", + "[43] Xu Yao, Gilles Puy, Alasdair Newson, Yann Gousseau, and Pierre Hellier. High resolution face age editing. In 25th International Conference on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021, pages 8624-8631. IEEE, 2020. 1, 3", + "[44] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 6", + "[45] Cheng Yu, Haoyu Xie, Lei Shang, Yang Liu, Jun Dan, Liefeng Bo, and Baigui Sun. Facechain-fact: Face adapter with decoupled training for identity-preserved personalization. arXiv preprint arXiv:2410.12312, 2024. 3, 4, 6", + "[46] Zongyan Zhang, Haohan Weng, Tong Zhang, and C. L. Philip Chen. A broad generative network for two-stage image outpainting. IEEE Transactions on Neural Networks and Learning Systems, 35(9):12731-12745, 2024. 1" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12156", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[47] Zongyan Zhang, C. L. Philip Chen, Haohan Weng, and Tong Zhang. Self-prompt guided image outpainting model for captions absence in social scenes. IEEE Transactions on Computational Social Systems, pages 1-12, 2025. 1", + "bbox": [ + 91, + 90, + 486, + 150 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12157", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/b7999e96-8abf-47f1-9b7e-e5c9222126c0_model.json b/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/b7999e96-8abf-47f1-9b7e-e5c9222126c0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5199379f01d04684b4c93ac2d9c8bf01070e6c80 --- /dev/null +++ b/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/b7999e96-8abf-47f1-9b7e-e5c9222126c0_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.13, + 0.873, + 0.177 + ], + "angle": 0, + "content": "TimeBooth: Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.202, + 0.842, + 0.223 + ], + "angle": 0, + "content": "Zepeng Su\\(^{1}\\), Zhulin Liu\\(^{1,2,*}\\), Zongyan Zhang\\(^{1,2}\\), Tong Zhang\\(^{1,2,3,4}\\), C. L. Philip Chen\\(^{1,2,3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.224, + 0.666, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\)South China University of Technology \\(^{2}\\)Pazhou Lab" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.241, + 0.786, + 0.257 + ], + "angle": 0, + "content": "3Guangdong Provincial Key Laboratory of Computational AI Models and Cognitive Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.259, + 0.859, + 0.275 + ], + "angle": 0, + "content": "4Engineering Research Center of the Ministry of Education on Health Intelligent Perception and Paralleled Digital-Human" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.301, + 0.172, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.365, + 0.151, + 0.375 + ], + "angle": 0, + "content": "Nemo" + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.377, + 0.157, + 0.389 + ], + "angle": 0, + "content": "(Source)" + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.296, + 0.516, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.371, + 0.399, + 0.383 + ], + "angle": 0, + "content": "Film Mr. Nobody" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.284, + 0.902, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.401, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Figure 1. In Mr. Nobody, the protagonist Nemo traverses multiple parallel realities, showcasing distinct aged appearances. This illustrates how choices such as environmental factors and lifestyle influence the diversity of face aging. Inspired by this, we propose TimeBooth." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.464, + 0.327, + 0.479 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.494, + 0.486, + 0.871 + ], + "angle": 0, + "content": "Face aging is a typical ill-posed problem influenced by various factors such as environment and genetics, leading to highly diverse outcomes. However, existing methods primarily rely on numerical age representations, making it difficult to accurately capture individual or group-level aging patterns. To address this, we introduce a novel disentangled face representation, where age features are modeled in the image modality—referred to as the Age Prompt—providing richer prior age information to constrain the generation results. To this end, we design an ID-age multi-task colearning framework and propose the Bidirectional Adversarial Disentanglement(BAD) strategy. This strategy maximizes the disentanglement of ID and age representation through bidirectional adversarial learning, extracting their attribute-invariant representations. Based on this representation, we propose TimeBooth, a personalized face aging model capable of generating diverse and individualized aging results. To optimize training, we construct a cross-age hybrid data pipeline and introduce various training strategies. Finally, we propose the R-AgeMAE metric and validate our method through extensive experiments, demonstrating that TimeBooth outperforms existing methods in both diversity and controllability. Our code will be publicly available at https://github.com/zpsuu/TimeBooth." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.464, + 0.645, + 0.479 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.703 + ], + "angle": 0, + "content": "In recent years, large-scale text-to-image generation models [27, 30, 32, 37, 46, 47] have demonstrated remarkable controllable generation capabilities, transforming imagination into reality. Face aging, also known as face age transformation, enables the synthesis of photorealistic face images at different ages by modifying age-related facial attributes while preserving identity-consistent features, such as facial structure, contour and so on. This technology has broad applications, including film post-production, forensic age progression for missing persons, and medical diagnostics. With the advancement of personalized generation techniques, face aging can achieve more precise and individualized simulations, further enhancing its real-world applicability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Face aging is a typical ill-posed problem, influenced by various factors [16, 31, 38] such as environment, genetics, lifestyle, and living conditions. For instance, individuals of the same age may exhibit distinct aging patterns—prolonged sun exposure can lead to increased freckles and skin laxity, while those carrying the MCR1 gene are more sensitive to light and prone to developing wrinkles. However, existing face aging methods [1, 13, 28, 41, 43] are predominantly based on GANs [15] and follow an image-to-image translation paradigm. These methods typically use age as the sole constraint—age, resulting in generated images that merely transfer the average aging characteristics learned from the dataset. Consequently, the existing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.454, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: Zhulin Liu (e-mail: liuzhl@scut.edu.cn)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12147" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.299, + 0.908, + 0.373 + ], + "angle": 0, + "content": "Figure 2. Training pipelines of our proposed TimeBooth. It consists of two stages: (a) Through ID-age collaborative training, we learn ID and age feature representations and introduce the BAD strategy, where bidirectional adversarial disentanglement based on mutual GRL layers is applied to extract attribute-invariant features. \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are positive numbers greater than 0, and (b) The disentangled encoder is frozen. Based on the learned disentangled representation, fusion is performed in the Age Adapter, followed by image cross-attention layers, which are used with a diffusion model for generation. Multiple training strategies are optimized concurrently during this process." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.396, + 0.482, + 0.442 + ], + "angle": 0, + "content": "paradigm can only provide a seemingly reasonable solution to this ill-posed problem but fails to accurately capture the diversity of individual or group-level aging patterns." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.483, + 0.626 + ], + "angle": 0, + "content": "Fundamentally, the limited expressiveness of existing paradigms stems from the inadequate representation of age as a constraint. As a numerical modality, age is inherently an abstract and discrete label, making it insufficient to capture the diversity and complexity of facial features at different ages. Consequently, it fails to provide the model with adequate semantic guidance for diverse and controllable age transformation. Therefore, it is essential to introduce a constraint with a richer age representation that properly regularizes the ill-posed nature of face aging, leading to more precise and well-defined solutions for personalized and diverse face aging." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As the saying goes, \"a picture is worth a thousand words.\" As a highly concretized modality, images naturally provide rich and intuitive information. Inspired by this, we propose a novel disentangled face representation, where the ID-invariant feature and age-related feature are disentangled, and age is modeled using an image modality, referred to as the Age Prompt. To this end, we construct a multi-task collaborative learning framework, leveraging ID recognition and age classification tasks for joint supervision. We propose a bidirectional adversarial disentanglement strategy, which introduces mutual gradient reversal layers to maximize feature disentanglement, thereby learning attribute-invariant representations. Based on this representation, we further design an Age Adapter, in which a lightweight Q-former is employed to fuse the disentangled features, while the image cross-attention layers inject facial information into the diffusion model for generation. To train the Age Adapter, we build a cross-age hybrid dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.907, + 0.472 + ], + "angle": 0, + "content": "pipeline and propose multiple training strategies to ensure the correct alignment of disentangled features. Finally, we introduce the R-AgeMAE as a metric to evaluate the accuracy of personalized face aging. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.503, + 0.905, + 0.609 + ], + "angle": 0, + "content": "- We propose a novel disentangled facial representation, where age constraints are modeled using an image modality instead of the traditional numerical modality. To the best of our knowledge, this is the first attempt to represent age using image prompts, providing a richer age prior for the inherently ill-posed problem of face aging and introducing a new generative paradigm." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.61, + 0.905, + 0.683 + ], + "angle": 0, + "content": "- We construct an ID-age multi-task co-learning framework and propose a bidirectional adversarial disentanglement strategy to maximize the disentanglement of ID and age representations, enabling the learning of their attribute-invariant representations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.685, + 0.905, + 0.759 + ], + "angle": 0, + "content": "- We introduce TimeBooth, which enables users to generate more diverse and personalized aging face images. To enhance training, we construct a hybrid cross-age dataset and design multiple training strategies to optimize the Age Adapter." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.761, + 0.905, + 0.88 + ], + "angle": 0, + "content": "- We introduce the R-AgeMAE metric to evaluate the accuracy of personalized face aging. Extensive experiments demonstrate the superiority of TimeBooth in terms of diversity, and controllability. Moreover, TimeBooth supports plug-and-play integration, seamlessly incorporating into existing community models to provide efficient personalized face aging generation in just a few seconds without test-time optimizing." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.503, + 0.905, + 0.88 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12148" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.473, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.346, + 0.423, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Inference pipeline of TimeBooth." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.385, + 0.234, + 0.401 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.41, + 0.284, + 0.426 + ], + "angle": 0, + "content": "2.1. Face Aging methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.485, + 0.706 + ], + "angle": 0, + "content": "In recent years, most face age transformation methods [1, 13, 28, 41, 43] have been GAN-based [15], following an image-to-image generation paradigm where age serves as a conditioning factor. Many approaches [2, 19, 41, 42] utilize pre-trained face recognition models and ID-preserving loss terms to maintain facial identity, while others achieve transformation via style transfer. CUSP [13] disentangles content and style encoding, using a content mask for customizable editing. With the rise of diffusion models [27, 32], recent methods have explored their potential for face age transformation. Inspired by [35], [3] introduces specific identifiers and an age regularization set to fine-tune diffusion models. FADING [7] applies Prompt-to-Prompt [18] local editing, modifying attention maps during generation. Additionally, general image editing approaches based on attention mechanisms [5, 18] can also perform face age transformation by leveraging pre-trained text-to-image priors, though with potentially lower accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.317, + 0.728 + ], + "angle": 0, + "content": "2.2. Personalization methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Although large-scale text-to-image diffusion models [27, 30, 32, 37] possess rich prior knowledge, they are limited to generating content from their pre-trained corpus and struggle to meet personalized user needs. Researchers have explored personalized generation methods. One method [36] involves test-time fine-tuning, where reference images are used to adapt diffusion models at inference, often with LoRA [20] for efficiency. Another method [40, 44, 45] leverages adapter-based architectures, learning feature representations to preserve identity without test-time training. Many approaches also incorporate recognition-based [26]" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.091, + 0.887, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.251, + 0.874, + 0.266 + ], + "angle": 0, + "content": "Figure 4. Visualization of disentangled representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.907, + 0.463 + ], + "angle": 0, + "content": "embeddings into the generation process. IP-Adapter-Faceid [44] combines face embeddings with CLIP [29] image encoders and employs disentangled cross-attention for generation. PhotoMaker [25] stacks ID embeddings for higher fidelity but struggles with fewer reference images. InstantID [40], built on IP-Adapter [44], introduces a facial keypoint control branch to improve shape fidelity at the cost of editability. FaceChain-FACT [45] utilizes a pre-trained ViT face recognition model [10] and a gated cross-attention layer to enhance instruction-following without compromising identity preservation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.48, + 0.613, + 0.495 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.507, + 0.655, + 0.523 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.907, + 0.605 + ], + "angle": 0, + "content": "Latent Diffusion Model. Latent diffusion models [32] are a class of generative models. During training, noise is progressively added to the latent representation of the original image, \\( z_0 \\), to obtain \\( z_t \\) through a forward process. This process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.612, + 0.905, + 0.629 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\varepsilon \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.695 + ], + "angle": 0, + "content": "The denoising network needs to accept the time step \\( t \\) and the condition \\( c \\) to estimate the noise. After multiple steps of denoising, the latent representation of target image \\( z_0 \\) can be obtained, which can be expressed by the formula:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.7, + 0.907, + 0.752 + ], + "angle": 0, + "content": "\\[\nz _ {t - 1} = \\frac {\\sqrt {\\alpha_ {t - 1}}}{\\sqrt {\\alpha_ {t}}} z _ {t} + \\sqrt {\\alpha_ {t - 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\varepsilon_ {\\theta} \\left(z _ {t}, t, c\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.753, + 0.907, + 0.813 + ], + "angle": 0, + "content": "During training, we estimate the noise added in the forward process and use the denoising network for denoising in the reverse process. The objective function for training the denoising network is:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.817, + 0.907, + 0.836 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D M} = E _ {z _ {0}, \\varepsilon \\sim \\mathcal {N} (0, \\mathcal {I}), c, t} \\| \\varepsilon - \\varepsilon_ {\\theta} (z _ {t}, c, t) \\| ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Image-Prompt Adapter. The image adapter [44] is designed to enable a pre-trained text-to-image diffusion model to generate images conditioned on image prompts. It consists of an image encoder and a cross-attention module," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12149" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.905, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.32, + 0.908, + 0.364 + ], + "angle": 0, + "content": "Figure 5. TimeBooth is capable of capturing fine-grained aging patterns for both individuals (left) and groups (right). As the number of Age Prompts increases, the generated results gradually shift from individual diversity to a stable group-level aging pattern, enhancing the interpretability and diversity of face aging synthesis in a data-driven manner." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.377, + 0.889, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.557, + 0.908, + 0.586 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison with translation-based methods CUSP [14], IP2P [6], and FADING [8]. Compared to these methods, TimeBooth generates faces that are both more natural and diverse, effectively avoiding the issue of 'child faces with elderly textures.'" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.657 + ], + "angle": 0, + "content": "where the encoded image prompts are fed into the cross-attention layers to guide image generation. The adapter is trained using the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.661, + 0.483, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I A} = E _ {z _ {0}, \\varepsilon \\sim \\mathcal {N} (0, \\mathcal {I}), c _ {t}, c _ {i}, t} \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, c _ {t}, c _ {i}, t\\right) \\| ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.484, + 0.716 + ], + "angle": 0, + "content": "where \\( c_{t}, c_{i} \\) are text and image conditions after encoding text prompt and image prompt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.418, + 0.743 + ], + "angle": 0, + "content": "3.2. Disentangled Representation Encoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "ID-Age Representation Co-learning Framework. We divide facial representation into two components: the ID representation \\( c_{id} \\), which remain invariant across age transformations, and the age representation \\( c_{age} \\), which change with age. Based on previous work [45], using the output of a ViT-based face recognition model [10] as face representations has proven effective in achieving high-fidelity reconstructions in diffusion models. Following this idea, we adopt a similar method. However, we further differentiate the original output of the ViT model. Specifically, we append two" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.907, + 0.656 + ], + "angle": 0, + "content": "separate neck networks to refine and disentangle the original features. The specific model details can be found in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.907, + 0.706 + ], + "angle": 0, + "content": "One neck network is designed to extract ID age-invariant representation \\( c_{id} \\). For this, we use the CosFace loss [39] \\( \\mathcal{L}_{\\mathrm{COSFACE}} \\) for supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.711, + 0.905, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {I D}} = \\mathcal {L} _ {\\text {C O S F A C E}} \\left(L \\left(X _ {\\mathrm {i d}}\\right), y _ {\\mathrm {i d}}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.905, + 0.763 + ], + "angle": 0, + "content": "where \\(X_{\\mathrm{id}}\\) represents the extracted ID features, \\(L(\\cdot)\\) denotes the head network, \\(y_{\\mathrm{id}}\\) is the corresponding ID label." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.907, + 0.796 + ], + "angle": 0, + "content": "Meanwhile, the other neck network is used to extract age features \\( c_{age} \\). We use the mse loss as the supervision loss:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.801, + 0.905, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {A g e}} = \\mathcal {L} _ {\\mathrm {M S E}} (\\mathrm {D E X} (A (X _ {\\mathrm {a g e}})), y _ {\\mathrm {a g e}}) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.823, + 0.905, + 0.869 + ], + "angle": 0, + "content": "where \\( X_{\\mathrm{age}} \\) represents the extracted age features, \\( A(\\cdot) \\) is the head network, and \\( \\mathrm{DEX}(\\cdot) \\) is the deep expectation [33, 34], \\( y_{\\mathrm{age}} \\) is the ground-truth age label." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Bidirectional Adversarial Disentanglement(BAD) Strategy for Disentangling Representation. Supervised co" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12150" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.364 + ], + "angle": 0, + "content": "learning will result in overlap between ID and age representation. To enable cross-age face synthesis by modifying only \\( c_{age} \\), we aim to maximize the disentanglement of these two representations. Typically, face recognition features encapsulate all identifiable attributes of an individual, including age-related characteristics. However, such entanglement compromises feature purity, thereby affecting controllability during generation. Therefore, inspired by [22], we seek to learn an age-invariant ID representation that remains stable across varying age, eliminating the influence of age on ID recognition. To achieve this, we introduce an adversarial learning strategy to enhance representation disentanglement. Specifically, we incorporate a Gradient Reversal Layer (GRL) [12] before the age classification head. During backpropagation, the GRL reverses the gradient, forcing the ID representation to become age-invariant while retaining identity-specific information. The adversarial constraint is imposed through the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.369, + 0.483, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {A d v I D}} = \\mathcal {L} _ {\\mathrm {C O S F A C E}} \\left(L \\left(\\operatorname {G R L} \\left(X _ {\\text {a g e}}\\right)\\right), y _ {\\mathrm {i d}}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.391, + 0.398, + 0.406 + ], + "angle": 0, + "content": "where \\(\\mathrm{GRL}(\\cdot)\\) is the Gradient Reversal Layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.527 + ], + "angle": 0, + "content": "Similarly, to prevent ID information from being encoded in age features—ensuring that interpolation at inference time does not compromise the fidelity of generated faces—we introduce an adversarial constraint on the age feature extraction branch. Specifically, we incorporate a Gradient Reversal Layer (GRL) before the ID classification head, forcing the extracted age representation to become ID-invariant. The loss function is:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.534, + 0.483, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {A d v A g e}} = \\mathcal {L} _ {\\text {M S E}} (\\operatorname {D E X} (A (\\operatorname {G R L} (X _ {\\mathrm {i d}}))), y _ {\\text {a g e}}) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.484, + 0.615 + ], + "angle": 0, + "content": "By leveraging this bidirectional adversarial learning strategy, we achieve more effective disentanglement of ID and age representations. The total loss of the disentangled encoder is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.622, + 0.483, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D R E}} = \\mathcal {L} _ {\\mathrm {I D}} + \\lambda_ {\\mathrm {A g e}} \\mathcal {L} _ {\\mathrm {A g e}} + \\lambda_ {\\mathrm {A d v I D}} \\mathcal {L} _ {\\mathrm {A d v I D}} + \\lambda_ {\\mathrm {A d v A g e}} \\mathcal {L} _ {\\mathrm {A d v A g e}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.683 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{Age}}\\), \\(\\lambda_{\\mathrm{AdvID}}\\), and \\(\\lambda_{\\mathrm{AdvAge}}\\) are weighting coefficients balancing each loss term." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.416, + 0.712 + ], + "angle": 0, + "content": "3.3. Face Reconstruction via Age Adapter" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.779 + ], + "angle": 0, + "content": "To achieve personalized face aging, we first perform face reconstruction during training. We propose the Age Adapter, which integrates the Age-Conditioned ID Encoder and multiple image cross-attention modules." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Age-conditioned Face Encoding. To fuse disentangled representations into a complete ID representation with age information, we propose the Age-Conditioned ID Encoder. It begins by applying dual cross attention modules to facilitate interaction between \\( c_{id} \\) and \\( c_{age} \\), which are derived from the disentangled representation encoder. This step ensures that \\( c_{id} \\) are enriched with necessary age-related representation \\( c_{age} \\), while the \\( c_{age} \\) is simultaneously refined to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "align with the ID representation \\( c_{id} \\). Once the age-enhanced ID representation and ID-conditioned age representation are obtained, we employ a lightweight Q-Former [24] to adaptively learn the final comprehensive facial representation. Specifically, we introduce 32 learnable tokens that interact with the processed feature representations, allowing the model to capture a holistic and semantically rich ID representation while incorporating the essential age attributes. These learnable tokens dynamically adjust their latent space embeddings, ensuring that the final identity feature faithfully preserves the original identity while seamlessly incorporating the desired age characteristics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.907, + 0.348 + ], + "angle": 0, + "content": "Face Reconstruction. To enable face reconstruction using the fused representations, we integrate multiple image cross attention layers within a text-to-image model to effectively extract and incorporate information from the Age-conditioned ID Encoder \\(\\mathcal{E}\\), as formulated below:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.354, + 0.912, + 0.384 + ], + "angle": 0, + "content": "\\[\nZ = \\text {A t t e n t i o n} _ {\\text {t e x t}} (Q, K, V) + \\lambda \\cdot \\text {A t t e n t i o n} _ {\\text {i m g}} (Q, K ^ {\\prime}, V ^ {\\prime}) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.385, + 0.906, + 0.445 + ], + "angle": 0, + "content": "where, \\(\\lambda\\) is weight coefficient of image cross attention layers, \\(Q = ZW_{q}, K = c_{t}W_{k}, V = c_{t}W_{v}, K' = \\mathcal{E}(c_{id}, c_{age})W'_k, V' = \\mathcal{E}(c_{id}, c_{age})W'_v\\) are the query, key, and values matrices from the text and face representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.906, + 0.491 + ], + "angle": 0, + "content": "Finally, we optimize the Age Adapter and the unfrozen components of the Age-conditioned ID Encoder via a reconstruction task, with the objective defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.496, + 0.905, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D M} = E _ {z _ {0}, \\varepsilon \\sim \\mathcal {N} (0, \\mathcal {I}), c _ {t}, c _ {i d}, c _ {a g e}, t} \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, c _ {t}, \\mathcal {E} \\left(c _ {i d}, c _ {a g e}\\right), t\\right) \\| ^ {2} \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.535, + 0.827, + 0.551 + ], + "angle": 0, + "content": "3.4. Training Strategies for Age Adapter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.906, + 0.662 + ], + "angle": 0, + "content": "Shuffle Strategy. To ensure that the age-invariant ID representation \\(\\mathbf{c}_{\\mathrm{id}}\\) of the same identity maintains strong intraclass consistency in the feature space, we propose the Shuffle Strategy and design a cross-age data pipeline. During training, with probability \\(p_s\\), we dynamically replace the ID feature \\(\\mathbf{c}_{\\mathrm{id}}^i\\) of a sample with that of the same identity at a different age, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.667, + 0.905, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\tilde {c} _ {\\mathrm {i d}} ^ {i} = \\mathbb {I} \\cdot \\mathbf {c} _ {\\mathrm {i d}} ^ {j} + (1 - \\mathbb {I}) \\cdot c _ {\\mathrm {i d}} ^ {i}, \\mathbb {I} \\sim \\operatorname {B e r n o u l l i} \\left(p _ {s}\\right) \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Here, \\(\\mathbb{I}\\) follows a Bernoulli distribution with parameter \\(p_s\\), and \\(j\\) represents another image with the same identity but a different age (\\(j \\neq i, y_j = y_i\\)). This strategy not only enhances the learning effectiveness of the Age Adapter but also prevents overfitting due to direct reconstruction of the same image, thereby improving model generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Age Prototype Replacement(APR) Strategy. To prevent the Age Prompt from capturing spurious age-unrelated correlations and to leverage population-level statistical regularities of age attributes, we propose the APR strategy. We first extract age features \\(\\{\\mathbf{c}_{\\mathrm{age}}^i\\}_{i = 1}^N\\) from all training samples and perform offline K-means clustering [17] (precomputed before training) to group them into \\(K\\) cohorts. Each cohort's prototype vector \\(\\mu_{\\mathrm{age}}^k\\) encapsulates the average age patterns" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12151" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.155 + ], + "angle": 0, + "content": "of its members. During training, the original feature \\(\\mathbf{c}_{\\mathrm{age}}^i\\) of each sample is randomly replaced with its cohort prototype \\(\\mu_{\\mathrm{age}}^{k(i)}\\) with probability \\(p_a\\), effectively suppressing noisy individual-specific extremes:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.158, + 0.483, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\tilde {c} _ {\\text {a g e}} ^ {i} = \\mathbb {I} \\cdot \\mu_ {\\text {a g e}} ^ {k (i)} + (1 - \\mathbb {I}) \\cdot c _ {\\text {a g e}} ^ {i}, \\quad \\mathbb {I} \\sim \\operatorname {B e r n o u l l i} \\left(p _ {a}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.197, + 0.484, + 0.392 + ], + "angle": 0, + "content": "Dynamic Training-set Sampling(DTS) Strategy. Current face image datasets often exhibit a long-tailed age distribution, where training on raw data may cause feature overfitting in well-represented age groups and suboptimal learning in underrepresented ones. To address this, we propose the DTS strategy. Before training, age labels are discretized into equal intervals. During training, adaptive sample selection is performed based on the statistical median: for age groups exceeding the median sample count, only a representative subset matching the median is retained per epoch, while all samples are used for underrepresented groups. This strategy improves face generation across age ranges and mitigates data imbalance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.44 + ], + "angle": 0, + "content": "Drop Strategy. Similar to [44, 45], we randomly mask \\( c_{id} \\) and \\( c_{age} \\) with probability \\( p_d \\) during training. During inference, the denoising process is formulated as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.446, + 0.481, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {t - 1} = \\left(1 + \\lambda_ {\\mathrm {C F G}}\\right) \\cdot \\mathcal {H} \\left(z _ {t}, t, c _ {\\text {t e x t}}, \\mathcal {E} \\left(c _ {i d}, c _ {a g e}\\right)\\right) \\\\ - \\lambda_ {\\text {C F G}} \\mathcal {H} \\left(z _ {t}, t, c _ {\\text {t e x t - n e g}}, \\mathcal {E} \\left(c _ {i d - d r o p}, c _ {a g e - d r o p}\\right)\\right) \\tag {14} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.524 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}\\) represents the forward process for denoising U-Net and \\(\\lambda_{\\mathrm{CFG}}\\) is the CFG scale." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.45, + 0.551 + ], + "angle": 0, + "content": "3.5. Personalized Face Aging during Inference" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.693 + ], + "angle": 0, + "content": "During the inference phase (see Fig. 3), we employ the disentangled representation encoder to extract \\( c_{id} \\) and \\( c_{age}^{src} \\) from the source image. To achieve more controllable age transformation, we propose a fusion strategy based on Spherical Linear Interpolation. This method interpolates between the source age feature \\( c_{age}^{src} \\) and the target Age Prompt \\( c_{age}^{tgt} \\) on a unit hypersphere, generating a fused age feature \\( c_{age}^{*} \\) with the hyperparameter aging guidance \\( \\alpha \\) controlling the intervention intensity of the target attribute:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.697, + 0.483, + 0.728 + ], + "angle": 0, + "content": "\\[\nc _ {a g e} ^ {*} = \\frac {\\sin [ (1 - \\alpha) \\theta ]}{\\sin \\theta} c _ {a g e} ^ {s r c} + \\frac {\\sin [ \\alpha \\theta ]}{\\sin \\theta} c _ {a g e} ^ {t g t}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.483, + 0.761 + ], + "angle": 0, + "content": "where \\(\\theta = \\arccos \\left(c_{age}^{src} \\cdot c_{age}^{tgt}\\right)\\) is the angular distance between feature vectors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.484, + 0.796 + ], + "angle": 0, + "content": "Finally, the cross-age face images are synthesized with identity consistency by combining \\( c_{id} \\) and \\( c_{age}^* \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.808, + 0.224, + 0.825 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.274, + 0.85 + ], + "angle": 0, + "content": "4.1. Experiments Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Dataset construction. In the first training stage, we utilize the MS1Mv2 [11] dataset for ID recognition and age estimation, with age annotations provided by [21]. In the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.228 + ], + "angle": 0, + "content": "second stage, we construct a hybrid dataset containing approximately 500K high-quality facial images. This hybrid dataset consists of two parts: one is a general crossage facial dataset without identity overlap, composed of FFHQ [23] and SFHQ [4]; the other is a cross-age dataset with the same identity across different ages, derived from VGGFace2-HQ [9], based on which we establish a crossage identity pool. For more details on dataset construction, please refer to the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.229, + 0.828, + 0.244 + ], + "angle": 0, + "content": "Training details. Please refer to the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.245, + 0.907, + 0.501 + ], + "angle": 0, + "content": "Baseline. To the best of our knowledge, there has been no method that directly utilizes age information as image prompts for face generation. Therefore, a direct comparison with existing personalized face aging methods is not feasible. To provide a reasonable baseline, we have constructed several baseline methods based on current common used face generation methods, including DreamBooth [36], FaceID [44], and FaceID+DreamBooth. Although these methods are not specifically designed for face aging, with appropriate configuration and adjustments, they can achieve a certain degree of control over face aging as image prompts. Detailed experimental configurations and parameter settings can be found in the appendix. Additionally, we selected several state-of-the-art methods [6, 8, 14] from the past in the field of face aging for comparison. These methods are based on image translation methods, which, while differing from our task, share a similar objective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.906, + 0.653 + ], + "angle": 0, + "content": "Evaluation. To evaluate personalized face aging performance, we employ seven quantitative metrics: Face Similarity (Face Sim.) for ID preservation, Kernel Inception Distance (KID) for distribution fidelity, Age Mean Absolute Error (AgeMAE) for age accuracy, and Relative AgeMAE (R-AgeMAE), our proposed metric, which specifically measures Age Prompt transfer accuracy. Additionally, we assess Face Quality, per-image inference time, and conduct a multidimensional user study for subjective perception. More details are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.654, + 0.907, + 0.777 + ], + "angle": 0, + "content": "Implementation of Age Prompt. Unless otherwise specified, the Age Prompts used are all sampled from FFHQ [23]. The sampling procedure follows three main rules: (1) The images are selected from the target age group, (2) the images must match the race and gender of the source face, and (3) dozens of target age images are sampled to ensure the use of prior knowledge from the population, which helps stabilize the results and ensure consistency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.697, + 0.803 + ], + "angle": 0, + "content": "4.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Comparison with translation-based methods. Time-Booth generates more diverse results compared to comparison methods, as illustrated in Fig. 6. Although translation-based approaches may appear to offer higher fidelity, the results generated by the comparison methods resemble 'children with elderly textures,' creating a strong sense of incon" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12152" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.885, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.298, + 0.898, + 0.313 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of lifespan synthesis. TimeBooth generates more natural images with high fidelity and consistency." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.325, + 0.885, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.472, + 0.908, + 0.501 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison. We selected in-the-wild celebrity photos and transformed them to the corresponding age in the GroundTruth. By transferring ID age-invariant features and leveraging the prior knowledge of Age Prompts, our results are superior." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.527, + 0.483, + 0.692 + ], + "angle": 0, + "content": "gruity, and do not resemble real human faces. In reality, both facial textures and bone structure undergo significant transformations from childhood to old age. By disentangling features and using Age Prompts as guidance, Time-Booth produces faces with semantics more aligned with the real aging process. Fig. 7 demonstrates that Time-Booth achieves a comparable level of cross-age consistency to the comparison methods, indicating accurate preservation of ID-invariant age features. Additionally, as shown in Fig. 9, Time-Booth enables more flexible and personalized generation, which translation-based methods cannot achieve." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.483, + 0.817 + ], + "angle": 0, + "content": "Comparison with baseline methods. As shown in Fig. 10, although the faces generated after fine-tuning with DreamBooth reflect the age features in the Age Prompt, their fidelity is relatively low. Furthermore, DreamBooth requires fine-tuning, which incurs higher computational and time costs. In contrast, our method is tuning-free, and the generated images can be produced in just a few seconds. For more qualitative experiments, please refer to the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.285, + 0.848 + ], + "angle": 0, + "content": "4.3.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We conducted a quantitative evaluation to verify Time-Booth's effectiveness. In Table 1, while TimeBooth inherently differs from translation-based methods by generating" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.572 + ], + "angle": 0, + "content": "images from scratch, it achieves superior performance on several metrics. Furthermore, it outperforms baseline methods on most metrics, highlighting its advantages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.649 + ], + "angle": 0, + "content": "User Study. We conducted a user study to compare our method against baseline approaches, evaluating three key metrics: facial consistency, Age Prompt similarity, and visual quality. As shown in Table 1, our method outperforms the baselines across all three metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.667, + 0.681 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Effect of Training Strategy. In Table 2, we present an ablation study on various training strategies. The results indicate that removing the Shuffle Strategy or DTS leads to a significant decrease in R-AgeMAE, suggesting a weakened ability to extract age-related information from the Age Prompt. Additionally, removing PAR or DTS results in a drop in KID and Face Sim., implying a decline in fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effect of Aging Guidance \\(\\alpha\\). In Eq. 15, we introduce \\(\\alpha\\) to control the influence of Age Prompt features on the target image. To evaluate its effect, we vary \\(\\alpha\\) from 0 to 1, and the generated results are shown in Fig. 11. As \\(\\alpha\\) increases, the facial features of the generated image gradually approach those of the Age Prompt, effectively simulating the aging process. Conversely, decreasing \\(\\alpha\\) makes the generated im" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12153" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.09, + 0.885, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.164, + 0.306, + 0.832, + 0.322 + ], + "angle": 0, + "content": "Figure 9. Qualitative comparison of face aging. TimeBooth demonstrates more flexible generation capabilities." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.334, + 0.905, + 0.44 + ], + "angle": 0, + "content": "
MethodYoung-to-OldYoung-to-ChildUser StudyFace QualityInference Time
KID(X100) ↓Face-sim ↑AgeMAE ↓R-AgeMAE ↓KID(X100) ↓Face-sim ↑AgeMAE ↓R-AgeMAE ↓Face Sim.(%)Sim. to A.P.(%)Visual Quality(%)
CUSP5.300.7446.08-8.080.8208.99-28.6-32.968.05<10 s
FADING4.800.5946.96-4.290.8687.05-21.4-18.668.173 min
IP2P9.300.49127.92-7.170.78820.63-18.6-12.965.86<10 s
FaceID13.040.63021.5022.5914.020.7267.385.7127.120.045.752.48<10 s
DreamBooth12.960.6745.335.7615.820.7436.845.6510.045.725.762.9510 min
FaceID+DreamBooth9.750.65125.0424.607.730.7289.788.2031.454.358.654.366 min
Ours7.200.7314.523.539.760.7638.557.9778.677.175.872.20<10 s
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.45, + 0.907, + 0.478 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison across various aspects. Blue indicates the best result among translation-based methods, bold denotes the best among baseline methods, and bold blue represents the best overall." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.501, + 0.481, + 0.665 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.675, + 0.465, + 0.69 + ], + "angle": 0, + "content": "Figure 10. Qualitative comparison with baseline methods." + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.711, + 0.432, + 0.781 + ], + "angle": 0, + "content": "
MethodKID (×100) ↓Face Sim. ↑R-AgeMAE ↓
w/o Shuffle4.850.7288.62
w/o PAR6.800.7194.11
w/o DTS7.400.7248.37
Full4.250.7334.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.792, + 0.361, + 0.806 + ], + "angle": 0, + "content": "Table 2. Ablation study." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "age more similar to the source image. Given the editability-fidelity trade-off, selecting an appropriate \\(\\alpha\\) is crucial for achieving a balanced result. Empirical studies suggest that an \\(\\alpha\\) value between 0.7 and 0.9 provides the best balance." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.503, + 0.895, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.624, + 0.907, + 0.665 + ], + "angle": 0, + "content": "Figure 11. Visualization under different aging guidance \\(\\alpha\\): as \\(\\alpha\\) increases, the age features from the Age Prompt are gradually injected into the source face." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.693, + 0.634, + 0.708 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel disentangled face representation that, for the first time, models age using image prompts, providing richer priors for face aging. Based on this, we introduce TimeBooth, which surpasses existing methods by generating more personalized and diverse aged face images, establishing a new paradigm for face aging. Moreover, TimeBooth supports plug-and-play integration and requires no test-time fine-tuning, offering greater flexibility and higher computational efficiency. In the future, we will further explore invariant face representations and extend their applications to broader and more general face generation tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12154" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.253 + ], + "angle": 0, + "content": "This work was funded in part by the National Natural Science Foundation of China grant under number 62222603, in part by the STI2030-Major Projects grant from the Ministry of Science and Technology of the People's Republic of China under number 2021ZD0200700, in part by the Key-Area Research and Development Program of Guangdong Province under number 2023B0303030001, and in part by the Program for Guangdong Introducing Innovative and Entrepreneurial Teams (2019ZT08X214)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.264, + 0.188, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.288, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[1] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Only a matter of style: age transformation using a style-based regression model. ACM Trans. Graph., 40(4):45:1-45:12, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.344, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[2] Grigory Antipov, Moez Baccouche, and Jean-Luc Dugelay. Face aging with conditional generative adversarial networks. In 2017 IEEE International Conference on Image Processing, ICIP 2017, Beijing, China, September 17-20, 2017, pages 2089-2093. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[3] Sudipta Banerjee, Govind Mittal, Ameya Joshi, Chinmay Hegde, and Nasir D. Memon. Identity-preserving aging of face images via latent diffusion models. In IEEE International Joint Conference on Biometrics, IJCB 2023, Ljubljana, Slovenia, September 25-28, 2023, pages 1-10. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[4] David Beniaguev. Synthetic faces high quality (sfhq) dataset, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18392-18402. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.482, + 0.72 + ], + "angle": 0, + "content": "[7] Xiangyi Chen and Stephane Lathuilière. Face aging via diffusion-based editing. In 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023, page 595. BMVA Press, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.721, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[8] Xiangyi Chen and Stéphane Lathuilière. Face aging via diffusion-based editing. arXiv preprint arXiv:2309.11321, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[9] Xuanhong Chen, Bingbing Ni, Yutian Liu, Naiyuan Liu, Zhilin Zeng, and Hang Wang. Simswap++: Towards faster and high-quality identity swapping. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):576-592, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[10] Jun Dan, Yang Liu, Haoyu Xie, Jiankang Deng, Haoran Xie, Xuansong Xie, and Baigui Sun. Transface: Calibrating transformer training for face recognition from a data-centric perspective. In Proceedings of the IEEE/CVF international conference on computer vision, pages 20642-20653, 2023. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.288, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.162, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[12] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[13] Guillermo Gomez-Trenado, Stéphane Lathuilière, Pablo Mesejo, and Oscar Cordón. Custom structure preservation in face aging. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 565-580. Springer, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.904, + 0.357 + ], + "angle": 0, + "content": "[14] Guillermo Gomez-Trenado, Stéphane Lathuilière, Pablo Mesejo, and Oscar Cordón. Custom structure preservation in face aging. In European Conference on Computer Vision, pages 565-580. Springer, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.904, + 0.413 + ], + "angle": 0, + "content": "[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139–144, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.904, + 0.469 + ], + "angle": 0, + "content": "[16] Bahman Guyuron, David J Rowe, Adam Bryce Weinfeld, Yashar Eshraghi, Amir Fathi, and Seree Iamphongsai. Factors contributing to the facial aging of identical twins. Plastic and reconstructive surgery, 123(4):1321-1331, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.904, + 0.523 + ], + "angle": 0, + "content": "[17] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics), 28(1):100-108, 1979. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[19] Gee-Sern Hsu, Rui-Cang Xie, and Zhi-Ting Chen. Wasserstein divergence GAN with cross-age identity expert and attribute retainer for facial age transformation. IEEE Access, 9:39695-39706, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.65, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[21] Zhizhong Huang, Junping Zhang, and Hongming Shan. When age-invariant face recognition meets face age synthesis: A multi-task learning framework. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.907, + 0.817 + ], + "angle": 0, + "content": "[22] Zhizhong Huang, Junping Zhang, and Hongming Shan. When age-invariant face recognition meets face age synthesis: A multi-task learning framework. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7282-7291, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12155" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[25] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8640-8650, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[26] Haiqi Liu, C. L. Philip Chen, Xinrong Gong, and Tong Zhang. Robust saliency-aware distillation for few-shot fine-grained visual recognition. IEEE Transactions on Multimedia, 26:7529-7542, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.361 + ], + "angle": 0, + "content": "[27] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, pages 16784-16804. PMLR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.484, + 0.433 + ], + "angle": 0, + "content": "[28] Roy Or-El, Soumyadip Sengupta, Ohad Fried, Eli Shechtman, and Ira Kemelmacher-Shlizerman. Lifespan age transformation synthesis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI, pages 739-755. Springer, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.484, + 0.572 + ], + "angle": 0, + "content": "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[31] Helle Rexbye, Inge Petersen, Mette Johansens, Louise Klitkou, Bernard Jeune, and Kaare Christensen. Influence of environmental factors on facial ageing. Age and Ageing, 35 (2):110-115, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[33] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Dex: Deep expectation of apparent age from a single image. In Proceedings of the IEEE international conference on computer vision workshops, pages 10-15, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[34] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Deep expectation of real and apparent age from a single image without facial landmarks. International Journal of Computer Vision, 126(2):144-157, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Canada, June 17-24, 2023, pages 22500-22510. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[36] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.204, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[38] Kestutis Sveikata, Irena Balciuniene, Janina Tutkuviene, et al. Factors influencing face aging. literature review. Stomatologija, 13(4):113-116, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[39] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[40] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.606 + ], + "angle": 0, + "content": "[41] Zongwei Wang, Xu Tang, Weixin Luo, and Shenghua Gao. Face aging with identity-preserved conditional generative adversarial networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 7939-7947. Computer Vision Foundation / IEEE Computer Society, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.608, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[42] Hongyu Yang, Di Huang, Yunhong Wang, and Anil K. Jain. Learning continuous face age progression: A pyramid of gans. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):499-515, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[43] Xu Yao, Gilles Puy, Alasdair Newson, Yann Gousseau, and Pierre Hellier. High resolution face age editing. In 25th International Conference on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021, pages 8624-8631. IEEE, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[44] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[45] Cheng Yu, Haoyu Xie, Lei Shang, Yang Liu, Jun Dan, Liefeng Bo, and Baigui Sun. Facechain-fact: Face adapter with decoupled training for identity-preserved personalization. arXiv preprint arXiv:2410.12312, 2024. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Zongyan Zhang, Haohan Weng, Tong Zhang, and C. L. Philip Chen. A broad generative network for two-stage image outpainting. IEEE Transactions on Neural Networks and Learning Systems, 35(9):12731-12745, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12156" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.487, + 0.151 + ], + "angle": 0, + "content": "[47] Zongyan Zhang, C. L. Philip Chen, Haohan Weng, and Tong Zhang. Self-prompt guided image outpainting model for captions absence in social scenes. IEEE Transactions on Computational Social Systems, pages 1-12, 2025. 1" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12157" + } + ] +] \ No newline at end of file diff --git a/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/b7999e96-8abf-47f1-9b7e-e5c9222126c0_origin.pdf b/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/b7999e96-8abf-47f1-9b7e-e5c9222126c0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..63ad8a6248b22521c3784937348aec9735b111f3 --- /dev/null +++ b/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/b7999e96-8abf-47f1-9b7e-e5c9222126c0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:221a9e858545ef29ba5f03cf3677318b4c1ee7717e7cd982087dbe12a2c9c892 +size 4109035 diff --git a/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/full.md b/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4e857d4db8e70c410151470ef10dd6ac69de4c79 --- /dev/null +++ b/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/full.md @@ -0,0 +1,336 @@ +# TimeBooth: Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging + +Zepeng Su $^{1}$ , Zhulin Liu $^{1,2,*}$ , Zongyan Zhang $^{1,2}$ , Tong Zhang $^{1,2,3,4}$ , C. L. Philip Chen $^{1,2,3,4}$ + +$^{1}$ South China University of Technology $^{2}$ Pazhou Lab + +3Guangdong Provincial Key Laboratory of Computational AI Models and Cognitive Intelligence + +4Engineering Research Center of the Ministry of Education on Health Intelligent Perception and Paralleled Digital-Human + +![](images/8637756c8197d5ae7f632050474a8555d9d09e5610dcaeccdfc6e19f47922052.jpg) +Nemo +(Source) + +![](images/5b571f28f5bed9f37e3aeeb3c5b74bea5a8db47ae8bb0b7e783684137c023373.jpg) +Film Mr. Nobody + +![](images/c08c451133fe62cb2cbc55a076a6abbbf4c2c284c0401b38ad93a09096a17c71.jpg) +Figure 1. In Mr. Nobody, the protagonist Nemo traverses multiple parallel realities, showcasing distinct aged appearances. This illustrates how choices such as environmental factors and lifestyle influence the diversity of face aging. Inspired by this, we propose TimeBooth. + +# Abstract + +Face aging is a typical ill-posed problem influenced by various factors such as environment and genetics, leading to highly diverse outcomes. However, existing methods primarily rely on numerical age representations, making it difficult to accurately capture individual or group-level aging patterns. To address this, we introduce a novel disentangled face representation, where age features are modeled in the image modality—referred to as the Age Prompt—providing richer prior age information to constrain the generation results. To this end, we design an ID-age multi-task colearning framework and propose the Bidirectional Adversarial Disentanglement(BAD) strategy. This strategy maximizes the disentanglement of ID and age representation through bidirectional adversarial learning, extracting their attribute-invariant representations. Based on this representation, we propose TimeBooth, a personalized face aging model capable of generating diverse and individualized aging results. To optimize training, we construct a cross-age hybrid data pipeline and introduce various training strategies. Finally, we propose the R-AgeMAE metric and validate our method through extensive experiments, demonstrating that TimeBooth outperforms existing methods in both diversity and controllability. Our code will be publicly available at https://github.com/zpsuu/TimeBooth. + +# 1. Introduction + +In recent years, large-scale text-to-image generation models [27, 30, 32, 37, 46, 47] have demonstrated remarkable controllable generation capabilities, transforming imagination into reality. Face aging, also known as face age transformation, enables the synthesis of photorealistic face images at different ages by modifying age-related facial attributes while preserving identity-consistent features, such as facial structure, contour and so on. This technology has broad applications, including film post-production, forensic age progression for missing persons, and medical diagnostics. With the advancement of personalized generation techniques, face aging can achieve more precise and individualized simulations, further enhancing its real-world applicability. + +Face aging is a typical ill-posed problem, influenced by various factors [16, 31, 38] such as environment, genetics, lifestyle, and living conditions. For instance, individuals of the same age may exhibit distinct aging patterns—prolonged sun exposure can lead to increased freckles and skin laxity, while those carrying the MCR1 gene are more sensitive to light and prone to developing wrinkles. However, existing face aging methods [1, 13, 28, 41, 43] are predominantly based on GANs [15] and follow an image-to-image translation paradigm. These methods typically use age as the sole constraint—age, resulting in generated images that merely transfer the average aging characteristics learned from the dataset. Consequently, the existing + +![](images/1c5756e0033b565dcc753b7919a3dbe7b6bff2063846509972980517a0916649.jpg) +Figure 2. Training pipelines of our proposed TimeBooth. It consists of two stages: (a) Through ID-age collaborative training, we learn ID and age feature representations and introduce the BAD strategy, where bidirectional adversarial disentanglement based on mutual GRL layers is applied to extract attribute-invariant features. $\lambda_{1}$ and $\lambda_{2}$ are positive numbers greater than 0, and (b) The disentangled encoder is frozen. Based on the learned disentangled representation, fusion is performed in the Age Adapter, followed by image cross-attention layers, which are used with a diffusion model for generation. Multiple training strategies are optimized concurrently during this process. + +paradigm can only provide a seemingly reasonable solution to this ill-posed problem but fails to accurately capture the diversity of individual or group-level aging patterns. + +Fundamentally, the limited expressiveness of existing paradigms stems from the inadequate representation of age as a constraint. As a numerical modality, age is inherently an abstract and discrete label, making it insufficient to capture the diversity and complexity of facial features at different ages. Consequently, it fails to provide the model with adequate semantic guidance for diverse and controllable age transformation. Therefore, it is essential to introduce a constraint with a richer age representation that properly regularizes the ill-posed nature of face aging, leading to more precise and well-defined solutions for personalized and diverse face aging. + +As the saying goes, "a picture is worth a thousand words." As a highly concretized modality, images naturally provide rich and intuitive information. Inspired by this, we propose a novel disentangled face representation, where the ID-invariant feature and age-related feature are disentangled, and age is modeled using an image modality, referred to as the Age Prompt. To this end, we construct a multi-task collaborative learning framework, leveraging ID recognition and age classification tasks for joint supervision. We propose a bidirectional adversarial disentanglement strategy, which introduces mutual gradient reversal layers to maximize feature disentanglement, thereby learning attribute-invariant representations. Based on this representation, we further design an Age Adapter, in which a lightweight Q-former is employed to fuse the disentangled features, while the image cross-attention layers inject facial information into the diffusion model for generation. To train the Age Adapter, we build a cross-age hybrid dataset + +pipeline and propose multiple training strategies to ensure the correct alignment of disentangled features. Finally, we introduce the R-AgeMAE as a metric to evaluate the accuracy of personalized face aging. Our contributions can be summarized as follows: + +- We propose a novel disentangled facial representation, where age constraints are modeled using an image modality instead of the traditional numerical modality. To the best of our knowledge, this is the first attempt to represent age using image prompts, providing a richer age prior for the inherently ill-posed problem of face aging and introducing a new generative paradigm. +- We construct an ID-age multi-task co-learning framework and propose a bidirectional adversarial disentanglement strategy to maximize the disentanglement of ID and age representations, enabling the learning of their attribute-invariant representations. +- We introduce TimeBooth, which enables users to generate more diverse and personalized aging face images. To enhance training, we construct a hybrid cross-age dataset and design multiple training strategies to optimize the Age Adapter. +- We introduce the R-AgeMAE metric to evaluate the accuracy of personalized face aging. Extensive experiments demonstrate the superiority of TimeBooth in terms of diversity, and controllability. Moreover, TimeBooth supports plug-and-play integration, seamlessly incorporating into existing community models to provide efficient personalized face aging generation in just a few seconds without test-time optimizing. + +![](images/e88a88ede282596d013a42efef7b51842bd9097fc820489b301a10370cf6034f.jpg) +Figure 3. Inference pipeline of TimeBooth. + +# 2. Related Work + +# 2.1. Face Aging methods + +In recent years, most face age transformation methods [1, 13, 28, 41, 43] have been GAN-based [15], following an image-to-image generation paradigm where age serves as a conditioning factor. Many approaches [2, 19, 41, 42] utilize pre-trained face recognition models and ID-preserving loss terms to maintain facial identity, while others achieve transformation via style transfer. CUSP [13] disentangles content and style encoding, using a content mask for customizable editing. With the rise of diffusion models [27, 32], recent methods have explored their potential for face age transformation. Inspired by [35], [3] introduces specific identifiers and an age regularization set to fine-tune diffusion models. FADING [7] applies Prompt-to-Prompt [18] local editing, modifying attention maps during generation. Additionally, general image editing approaches based on attention mechanisms [5, 18] can also perform face age transformation by leveraging pre-trained text-to-image priors, though with potentially lower accuracy. + +# 2.2. Personalization methods + +Although large-scale text-to-image diffusion models [27, 30, 32, 37] possess rich prior knowledge, they are limited to generating content from their pre-trained corpus and struggle to meet personalized user needs. Researchers have explored personalized generation methods. One method [36] involves test-time fine-tuning, where reference images are used to adapt diffusion models at inference, often with LoRA [20] for efficiency. Another method [40, 44, 45] leverages adapter-based architectures, learning feature representations to preserve identity without test-time training. Many approaches also incorporate recognition-based [26] + +![](images/efe4e89776967d16ab3f7debebf29683e81dd772ee42c2134e2840957af43e72.jpg) +Figure 4. Visualization of disentangled representation. + +embeddings into the generation process. IP-Adapter-Faceid [44] combines face embeddings with CLIP [29] image encoders and employs disentangled cross-attention for generation. PhotoMaker [25] stacks ID embeddings for higher fidelity but struggles with fewer reference images. InstantID [40], built on IP-Adapter [44], introduces a facial keypoint control branch to improve shape fidelity at the cost of editability. FaceChain-FACT [45] utilizes a pre-trained ViT face recognition model [10] and a gated cross-attention layer to enhance instruction-following without compromising identity preservation. + +# 3. Methods + +# 3.1. Preliminaries + +Latent Diffusion Model. Latent diffusion models [32] are a class of generative models. During training, noise is progressively added to the latent representation of the original image, $z_0$ , to obtain $z_t$ through a forward process. This process can be formulated as follows: + +$$ +z _ {t} = \sqrt {\alpha_ {t}} z _ {0} + \sqrt {1 - \alpha_ {t}} \varepsilon \tag {1} +$$ + +The denoising network needs to accept the time step $t$ and the condition $c$ to estimate the noise. After multiple steps of denoising, the latent representation of target image $z_0$ can be obtained, which can be expressed by the formula: + +$$ +z _ {t - 1} = \frac {\sqrt {\alpha_ {t - 1}}}{\sqrt {\alpha_ {t}}} z _ {t} + \sqrt {\alpha_ {t - 1}} \left(\sqrt {\frac {1}{\alpha_ {t - 1}} - 1} - \sqrt {\frac {1}{\alpha_ {t}} - 1}\right) \varepsilon_ {\theta} \left(z _ {t}, t, c\right) \tag {2} +$$ + +During training, we estimate the noise added in the forward process and use the denoising network for denoising in the reverse process. The objective function for training the denoising network is: + +$$ +\mathcal {L} _ {D M} = E _ {z _ {0}, \varepsilon \sim \mathcal {N} (0, \mathcal {I}), c, t} \| \varepsilon - \varepsilon_ {\theta} (z _ {t}, c, t) \| ^ {2} \tag {3} +$$ + +Image-Prompt Adapter. The image adapter [44] is designed to enable a pre-trained text-to-image diffusion model to generate images conditioned on image prompts. It consists of an image encoder and a cross-attention module, + +![](images/e26ee12c76d07491d3f24314d4df334d96866ba8f73235cdd922099501aadcd8.jpg) +Figure 5. TimeBooth is capable of capturing fine-grained aging patterns for both individuals (left) and groups (right). As the number of Age Prompts increases, the generated results gradually shift from individual diversity to a stable group-level aging pattern, enhancing the interpretability and diversity of face aging synthesis in a data-driven manner. + +![](images/f8c2ae0e525c888b2a39699b561dc11eadab54b408b939b1ce35e95db605da22.jpg) +Figure 6. Qualitative comparison with translation-based methods CUSP [14], IP2P [6], and FADING [8]. Compared to these methods, TimeBooth generates faces that are both more natural and diverse, effectively avoiding the issue of 'child faces with elderly textures.' + +where the encoded image prompts are fed into the cross-attention layers to guide image generation. The adapter is trained using the following loss function: + +$$ +\mathcal {L} _ {I A} = E _ {z _ {0}, \varepsilon \sim \mathcal {N} (0, \mathcal {I}), c _ {t}, c _ {i}, t} \| \varepsilon - \varepsilon_ {\theta} \left(z _ {t}, c _ {t}, c _ {i}, t\right) \| ^ {2} \tag {4} +$$ + +where $c_{t}, c_{i}$ are text and image conditions after encoding text prompt and image prompt. + +# 3.2. Disentangled Representation Encoder + +ID-Age Representation Co-learning Framework. We divide facial representation into two components: the ID representation $c_{id}$ , which remain invariant across age transformations, and the age representation $c_{age}$ , which change with age. Based on previous work [45], using the output of a ViT-based face recognition model [10] as face representations has proven effective in achieving high-fidelity reconstructions in diffusion models. Following this idea, we adopt a similar method. However, we further differentiate the original output of the ViT model. Specifically, we append two + +separate neck networks to refine and disentangle the original features. The specific model details can be found in the Appendix. + +One neck network is designed to extract ID age-invariant representation $c_{id}$ . For this, we use the CosFace loss [39] $\mathcal{L}_{\mathrm{COSFACE}}$ for supervision: + +$$ +\mathcal {L} _ {\mathrm {I D}} = \mathcal {L} _ {\text {C O S F A C E}} \left(L \left(X _ {\mathrm {i d}}\right), y _ {\mathrm {i d}}\right) \tag {5} +$$ + +where $X_{\mathrm{id}}$ represents the extracted ID features, $L(\cdot)$ denotes the head network, $y_{\mathrm{id}}$ is the corresponding ID label. + +Meanwhile, the other neck network is used to extract age features $c_{age}$ . We use the mse loss as the supervision loss: + +$$ +\mathcal {L} _ {\mathrm {A g e}} = \mathcal {L} _ {\mathrm {M S E}} (\mathrm {D E X} (A (X _ {\mathrm {a g e}})), y _ {\mathrm {a g e}}) \tag {6} +$$ + +where $X_{\mathrm{age}}$ represents the extracted age features, $A(\cdot)$ is the head network, and $\mathrm{DEX}(\cdot)$ is the deep expectation [33, 34], $y_{\mathrm{age}}$ is the ground-truth age label. + +Bidirectional Adversarial Disentanglement(BAD) Strategy for Disentangling Representation. Supervised co + +learning will result in overlap between ID and age representation. To enable cross-age face synthesis by modifying only $c_{age}$ , we aim to maximize the disentanglement of these two representations. Typically, face recognition features encapsulate all identifiable attributes of an individual, including age-related characteristics. However, such entanglement compromises feature purity, thereby affecting controllability during generation. Therefore, inspired by [22], we seek to learn an age-invariant ID representation that remains stable across varying age, eliminating the influence of age on ID recognition. To achieve this, we introduce an adversarial learning strategy to enhance representation disentanglement. Specifically, we incorporate a Gradient Reversal Layer (GRL) [12] before the age classification head. During backpropagation, the GRL reverses the gradient, forcing the ID representation to become age-invariant while retaining identity-specific information. The adversarial constraint is imposed through the following loss function: + +$$ +\mathcal {L} _ {\mathrm {A d v I D}} = \mathcal {L} _ {\mathrm {C O S F A C E}} \left(L \left(\operatorname {G R L} \left(X _ {\text {a g e}}\right)\right), y _ {\mathrm {i d}}\right) \tag {7} +$$ + +where $\mathrm{GRL}(\cdot)$ is the Gradient Reversal Layer. + +Similarly, to prevent ID information from being encoded in age features—ensuring that interpolation at inference time does not compromise the fidelity of generated faces—we introduce an adversarial constraint on the age feature extraction branch. Specifically, we incorporate a Gradient Reversal Layer (GRL) before the ID classification head, forcing the extracted age representation to become ID-invariant. The loss function is: + +$$ +\mathcal {L} _ {\text {A d v A g e}} = \mathcal {L} _ {\text {M S E}} (\operatorname {D E X} (A (\operatorname {G R L} (X _ {\mathrm {i d}}))), y _ {\text {a g e}}) \tag {8} +$$ + +By leveraging this bidirectional adversarial learning strategy, we achieve more effective disentanglement of ID and age representations. The total loss of the disentangled encoder is as follows: + +$$ +\mathcal {L} _ {\mathrm {D R E}} = \mathcal {L} _ {\mathrm {I D}} + \lambda_ {\mathrm {A g e}} \mathcal {L} _ {\mathrm {A g e}} + \lambda_ {\mathrm {A d v I D}} \mathcal {L} _ {\mathrm {A d v I D}} + \lambda_ {\mathrm {A d v A g e}} \mathcal {L} _ {\mathrm {A d v A g e}} \tag {9} +$$ + +where $\lambda_{\mathrm{Age}}$ , $\lambda_{\mathrm{AdvID}}$ , and $\lambda_{\mathrm{AdvAge}}$ are weighting coefficients balancing each loss term. + +# 3.3. Face Reconstruction via Age Adapter + +To achieve personalized face aging, we first perform face reconstruction during training. We propose the Age Adapter, which integrates the Age-Conditioned ID Encoder and multiple image cross-attention modules. + +Age-conditioned Face Encoding. To fuse disentangled representations into a complete ID representation with age information, we propose the Age-Conditioned ID Encoder. It begins by applying dual cross attention modules to facilitate interaction between $c_{id}$ and $c_{age}$ , which are derived from the disentangled representation encoder. This step ensures that $c_{id}$ are enriched with necessary age-related representation $c_{age}$ , while the $c_{age}$ is simultaneously refined to + +align with the ID representation $c_{id}$ . Once the age-enhanced ID representation and ID-conditioned age representation are obtained, we employ a lightweight Q-Former [24] to adaptively learn the final comprehensive facial representation. Specifically, we introduce 32 learnable tokens that interact with the processed feature representations, allowing the model to capture a holistic and semantically rich ID representation while incorporating the essential age attributes. These learnable tokens dynamically adjust their latent space embeddings, ensuring that the final identity feature faithfully preserves the original identity while seamlessly incorporating the desired age characteristics. + +Face Reconstruction. To enable face reconstruction using the fused representations, we integrate multiple image cross attention layers within a text-to-image model to effectively extract and incorporate information from the Age-conditioned ID Encoder $\mathcal{E}$ , as formulated below: + +$$ +Z = \text {A t t e n t i o n} _ {\text {t e x t}} (Q, K, V) + \lambda \cdot \text {A t t e n t i o n} _ {\text {i m g}} (Q, K ^ {\prime}, V ^ {\prime}) \tag {10} +$$ + +where, $\lambda$ is weight coefficient of image cross attention layers, $Q = ZW_{q}, K = c_{t}W_{k}, V = c_{t}W_{v}, K' = \mathcal{E}(c_{id}, c_{age})W'_k, V' = \mathcal{E}(c_{id}, c_{age})W'_v$ are the query, key, and values matrices from the text and face representation. + +Finally, we optimize the Age Adapter and the unfrozen components of the Age-conditioned ID Encoder via a reconstruction task, with the objective defined as follows: + +$$ +\mathcal {L} _ {D M} = E _ {z _ {0}, \varepsilon \sim \mathcal {N} (0, \mathcal {I}), c _ {t}, c _ {i d}, c _ {a g e}, t} \| \varepsilon - \varepsilon_ {\theta} \left(z _ {t}, c _ {t}, \mathcal {E} \left(c _ {i d}, c _ {a g e}\right), t\right) \| ^ {2} \tag {11} +$$ + +# 3.4. Training Strategies for Age Adapter + +Shuffle Strategy. To ensure that the age-invariant ID representation $\mathbf{c}_{\mathrm{id}}$ of the same identity maintains strong intraclass consistency in the feature space, we propose the Shuffle Strategy and design a cross-age data pipeline. During training, with probability $p_s$ , we dynamically replace the ID feature $\mathbf{c}_{\mathrm{id}}^i$ of a sample with that of the same identity at a different age, formulated as: + +$$ +\tilde {c} _ {\mathrm {i d}} ^ {i} = \mathbb {I} \cdot \mathbf {c} _ {\mathrm {i d}} ^ {j} + (1 - \mathbb {I}) \cdot c _ {\mathrm {i d}} ^ {i}, \mathbb {I} \sim \operatorname {B e r n o u l l i} \left(p _ {s}\right) \tag {12} +$$ + +Here, $\mathbb{I}$ follows a Bernoulli distribution with parameter $p_s$ , and $j$ represents another image with the same identity but a different age ( $j \neq i, y_j = y_i$ ). This strategy not only enhances the learning effectiveness of the Age Adapter but also prevents overfitting due to direct reconstruction of the same image, thereby improving model generalization. + +Age Prototype Replacement(APR) Strategy. To prevent the Age Prompt from capturing spurious age-unrelated correlations and to leverage population-level statistical regularities of age attributes, we propose the APR strategy. We first extract age features $\{\mathbf{c}_{\mathrm{age}}^i\}_{i = 1}^N$ from all training samples and perform offline K-means clustering [17] (precomputed before training) to group them into $K$ cohorts. Each cohort's prototype vector $\mu_{\mathrm{age}}^k$ encapsulates the average age patterns + +of its members. During training, the original feature $\mathbf{c}_{\mathrm{age}}^i$ of each sample is randomly replaced with its cohort prototype $\mu_{\mathrm{age}}^{k(i)}$ with probability $p_a$ , effectively suppressing noisy individual-specific extremes: + +$$ +\tilde {c} _ {\text {a g e}} ^ {i} = \mathbb {I} \cdot \mu_ {\text {a g e}} ^ {k (i)} + (1 - \mathbb {I}) \cdot c _ {\text {a g e}} ^ {i}, \quad \mathbb {I} \sim \operatorname {B e r n o u l l i} \left(p _ {a}\right). \tag {13} +$$ + +Dynamic Training-set Sampling(DTS) Strategy. Current face image datasets often exhibit a long-tailed age distribution, where training on raw data may cause feature overfitting in well-represented age groups and suboptimal learning in underrepresented ones. To address this, we propose the DTS strategy. Before training, age labels are discretized into equal intervals. During training, adaptive sample selection is performed based on the statistical median: for age groups exceeding the median sample count, only a representative subset matching the median is retained per epoch, while all samples are used for underrepresented groups. This strategy improves face generation across age ranges and mitigates data imbalance. + +Drop Strategy. Similar to [44, 45], we randomly mask $c_{id}$ and $c_{age}$ with probability $p_d$ during training. During inference, the denoising process is formulated as follow: + +$$ +\begin{array}{l} z _ {t - 1} = \left(1 + \lambda_ {\mathrm {C F G}}\right) \cdot \mathcal {H} \left(z _ {t}, t, c _ {\text {t e x t}}, \mathcal {E} \left(c _ {i d}, c _ {a g e}\right)\right) \\ - \lambda_ {\text {C F G}} \mathcal {H} \left(z _ {t}, t, c _ {\text {t e x t - n e g}}, \mathcal {E} \left(c _ {i d - d r o p}, c _ {a g e - d r o p}\right)\right) \tag {14} \\ \end{array} +$$ + +where $\mathcal{H}$ represents the forward process for denoising U-Net and $\lambda_{\mathrm{CFG}}$ is the CFG scale. + +# 3.5. Personalized Face Aging during Inference + +During the inference phase (see Fig. 3), we employ the disentangled representation encoder to extract $c_{id}$ and $c_{age}^{src}$ from the source image. To achieve more controllable age transformation, we propose a fusion strategy based on Spherical Linear Interpolation. This method interpolates between the source age feature $c_{age}^{src}$ and the target Age Prompt $c_{age}^{tgt}$ on a unit hypersphere, generating a fused age feature $c_{age}^{*}$ with the hyperparameter aging guidance $\alpha$ controlling the intervention intensity of the target attribute: + +$$ +c _ {a g e} ^ {*} = \frac {\sin [ (1 - \alpha) \theta ]}{\sin \theta} c _ {a g e} ^ {s r c} + \frac {\sin [ \alpha \theta ]}{\sin \theta} c _ {a g e} ^ {t g t}, \tag {15} +$$ + +where $\theta = \arccos \left(c_{age}^{src} \cdot c_{age}^{tgt}\right)$ is the angular distance between feature vectors. + +Finally, the cross-age face images are synthesized with identity consistency by combining $c_{id}$ and $c_{age}^*$ . + +# 4. Experiments + +# 4.1. Experiments Setup + +Dataset construction. In the first training stage, we utilize the MS1Mv2 [11] dataset for ID recognition and age estimation, with age annotations provided by [21]. In the + +second stage, we construct a hybrid dataset containing approximately 500K high-quality facial images. This hybrid dataset consists of two parts: one is a general crossage facial dataset without identity overlap, composed of FFHQ [23] and SFHQ [4]; the other is a cross-age dataset with the same identity across different ages, derived from VGGFace2-HQ [9], based on which we establish a crossage identity pool. For more details on dataset construction, please refer to the Appendix. + +# Training details. Please refer to the Appendix. + +Baseline. To the best of our knowledge, there has been no method that directly utilizes age information as image prompts for face generation. Therefore, a direct comparison with existing personalized face aging methods is not feasible. To provide a reasonable baseline, we have constructed several baseline methods based on current common used face generation methods, including DreamBooth [36], FaceID [44], and FaceID+DreamBooth. Although these methods are not specifically designed for face aging, with appropriate configuration and adjustments, they can achieve a certain degree of control over face aging as image prompts. Detailed experimental configurations and parameter settings can be found in the appendix. Additionally, we selected several state-of-the-art methods [6, 8, 14] from the past in the field of face aging for comparison. These methods are based on image translation methods, which, while differing from our task, share a similar objective. + +Evaluation. To evaluate personalized face aging performance, we employ seven quantitative metrics: Face Similarity (Face Sim.) for ID preservation, Kernel Inception Distance (KID) for distribution fidelity, Age Mean Absolute Error (AgeMAE) for age accuracy, and Relative AgeMAE (R-AgeMAE), our proposed metric, which specifically measures Age Prompt transfer accuracy. Additionally, we assess Face Quality, per-image inference time, and conduct a multidimensional user study for subjective perception. More details are provided in the Appendix. + +Implementation of Age Prompt. Unless otherwise specified, the Age Prompts used are all sampled from FFHQ [23]. The sampling procedure follows three main rules: (1) The images are selected from the target age group, (2) the images must match the race and gender of the source face, and (3) dozens of target age images are sampled to ensure the use of prior knowledge from the population, which helps stabilize the results and ensure consistency. + +# 4.2. Qualitative Results + +Comparison with translation-based methods. Time-Booth generates more diverse results compared to comparison methods, as illustrated in Fig. 6. Although translation-based approaches may appear to offer higher fidelity, the results generated by the comparison methods resemble 'children with elderly textures,' creating a strong sense of incon + +![](images/6c1d8cffc2e0a93ae1cb86c3e363a29ac687ead2b65841f28b0727e83cd539fd.jpg) +Figure 7. Qualitative comparison of lifespan synthesis. TimeBooth generates more natural images with high fidelity and consistency. + +![](images/0888239f85217be4045950af35edaa5c79b609a488f364b95808231ac0f5249f.jpg) +Figure 8. Qualitative comparison. We selected in-the-wild celebrity photos and transformed them to the corresponding age in the GroundTruth. By transferring ID age-invariant features and leveraging the prior knowledge of Age Prompts, our results are superior. + +gruity, and do not resemble real human faces. In reality, both facial textures and bone structure undergo significant transformations from childhood to old age. By disentangling features and using Age Prompts as guidance, Time-Booth produces faces with semantics more aligned with the real aging process. Fig. 7 demonstrates that Time-Booth achieves a comparable level of cross-age consistency to the comparison methods, indicating accurate preservation of ID-invariant age features. Additionally, as shown in Fig. 9, Time-Booth enables more flexible and personalized generation, which translation-based methods cannot achieve. + +Comparison with baseline methods. As shown in Fig. 10, although the faces generated after fine-tuning with DreamBooth reflect the age features in the Age Prompt, their fidelity is relatively low. Furthermore, DreamBooth requires fine-tuning, which incurs higher computational and time costs. In contrast, our method is tuning-free, and the generated images can be produced in just a few seconds. For more qualitative experiments, please refer to the Appendix. + +# 4.3.Quantitative Results + +We conducted a quantitative evaluation to verify Time-Booth's effectiveness. In Table 1, while TimeBooth inherently differs from translation-based methods by generating + +images from scratch, it achieves superior performance on several metrics. Furthermore, it outperforms baseline methods on most metrics, highlighting its advantages. + +User Study. We conducted a user study to compare our method against baseline approaches, evaluating three key metrics: facial consistency, Age Prompt similarity, and visual quality. As shown in Table 1, our method outperforms the baselines across all three metrics. + +# 4.4. Ablation Study + +Effect of Training Strategy. In Table 2, we present an ablation study on various training strategies. The results indicate that removing the Shuffle Strategy or DTS leads to a significant decrease in R-AgeMAE, suggesting a weakened ability to extract age-related information from the Age Prompt. Additionally, removing PAR or DTS results in a drop in KID and Face Sim., implying a decline in fidelity. + +Effect of Aging Guidance $\alpha$ . In Eq. 15, we introduce $\alpha$ to control the influence of Age Prompt features on the target image. To evaluate its effect, we vary $\alpha$ from 0 to 1, and the generated results are shown in Fig. 11. As $\alpha$ increases, the facial features of the generated image gradually approach those of the Age Prompt, effectively simulating the aging process. Conversely, decreasing $\alpha$ makes the generated im + +![](images/f4529f45afefc8f603d3413bea29fa2531b4801758b7112bda5c5c8b60c2d9d1.jpg) +Figure 9. Qualitative comparison of face aging. TimeBooth demonstrates more flexible generation capabilities. + +
MethodYoung-to-OldYoung-to-ChildUser StudyFace QualityInference Time
KID(X100) ↓Face-sim ↑AgeMAE ↓R-AgeMAE ↓KID(X100) ↓Face-sim ↑AgeMAE ↓R-AgeMAE ↓Face Sim.(%)Sim. to A.P.(%)Visual Quality(%)
CUSP5.300.7446.08-8.080.8208.99-28.6-32.968.05<10 s
FADING4.800.5946.96-4.290.8687.05-21.4-18.668.173 min
IP2P9.300.49127.92-7.170.78820.63-18.6-12.965.86<10 s
FaceID13.040.63021.5022.5914.020.7267.385.7127.120.045.752.48<10 s
DreamBooth12.960.6745.335.7615.820.7436.845.6510.045.725.762.9510 min
FaceID+DreamBooth9.750.65125.0424.607.730.7289.788.2031.454.358.654.366 min
Ours7.200.7314.523.539.760.7638.557.9778.677.175.872.20<10 s
+ +![](images/2971bf21b6c7a8c4308221de961d16e549685e869af443ad788b0e1d38d43314.jpg) +Figure 10. Qualitative comparison with baseline methods. + +Table 1. Quantitative comparison across various aspects. Blue indicates the best result among translation-based methods, bold denotes the best among baseline methods, and bold blue represents the best overall. + +
MethodKID (×100) ↓Face Sim. ↑R-AgeMAE ↓
w/o Shuffle4.850.7288.62
w/o PAR6.800.7194.11
w/o DTS7.400.7248.37
Full4.250.7334.58
+ +Table 2. Ablation study. + +age more similar to the source image. Given the editability-fidelity trade-off, selecting an appropriate $\alpha$ is crucial for achieving a balanced result. Empirical studies suggest that an $\alpha$ value between 0.7 and 0.9 provides the best balance. + +![](images/b5cec05d691f3a5854927beeb96c25d41f2bd3e4857c392d87610f314f2eacd9.jpg) +Figure 11. Visualization under different aging guidance $\alpha$ : as $\alpha$ increases, the age features from the Age Prompt are gradually injected into the source face. + +# 5. Conclusion + +In this paper, we propose a novel disentangled face representation that, for the first time, models age using image prompts, providing richer priors for face aging. Based on this, we introduce TimeBooth, which surpasses existing methods by generating more personalized and diverse aged face images, establishing a new paradigm for face aging. Moreover, TimeBooth supports plug-and-play integration and requires no test-time fine-tuning, offering greater flexibility and higher computational efficiency. In the future, we will further explore invariant face representations and extend their applications to broader and more general face generation tasks. + +# Acknowledgments + +This work was funded in part by the National Natural Science Foundation of China grant under number 62222603, in part by the STI2030-Major Projects grant from the Ministry of Science and Technology of the People's Republic of China under number 2021ZD0200700, in part by the Key-Area Research and Development Program of Guangdong Province under number 2023B0303030001, and in part by the Program for Guangdong Introducing Innovative and Entrepreneurial Teams (2019ZT08X214). + +# References + +[1] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Only a matter of style: age transformation using a style-based regression model. ACM Trans. Graph., 40(4):45:1-45:12, 2021. 1, 3 +[2] Grigory Antipov, Moez Baccouche, and Jean-Luc Dugelay. Face aging with conditional generative adversarial networks. In 2017 IEEE International Conference on Image Processing, ICIP 2017, Beijing, China, September 17-20, 2017, pages 2089-2093. IEEE, 2017. 3 +[3] Sudipta Banerjee, Govind Mittal, Ameya Joshi, Chinmay Hegde, and Nasir D. Memon. Identity-preserving aging of face images via latent diffusion models. In IEEE International Joint Conference on Biometrics, IJCB 2023, Ljubljana, Slovenia, September 25-28, 2023, pages 1-10. IEEE, 2023. 3 +[4] David Beniaguev. Synthetic faces high quality (sfhq) dataset, 2022. 6 +[5] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18392-18402. IEEE, 2023. 3 +[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 4, 6 +[7] Xiangyi Chen and Stephane Lathuilière. Face aging via diffusion-based editing. In 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023, page 595. BMVA Press, 2023. 3 +[8] Xiangyi Chen and Stéphane Lathuilière. Face aging via diffusion-based editing. arXiv preprint arXiv:2309.11321, 2023. 4, 6 +[9] Xuanhong Chen, Bingbing Ni, Yutian Liu, Naiyuan Liu, Zhilin Zeng, and Hang Wang. Simswap++: Towards faster and high-quality identity swapping. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):576-592, 2023. 6 +[10] Jun Dan, Yang Liu, Haoyu Xie, Jiankang Deng, Haoran Xie, Xuansong Xie, and Baigui Sun. Transface: Calibrating transformer training for face recognition from a data-centric perspective. In Proceedings of the IEEE/CVF international conference on computer vision, pages 20642-20653, 2023. 3, 4 + +[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6 +[12] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 5 +[13] Guillermo Gomez-Trenado, Stéphane Lathuilière, Pablo Mesejo, and Oscar Cordón. Custom structure preservation in face aging. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 565-580. Springer, 2022. 1, 3 +[14] Guillermo Gomez-Trenado, Stéphane Lathuilière, Pablo Mesejo, and Oscar Cordón. Custom structure preservation in face aging. In European Conference on Computer Vision, pages 565-580. Springer, 2022. 4, 6 +[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139–144, 2020. 1, 3 +[16] Bahman Guyuron, David J Rowe, Adam Bryce Weinfeld, Yashar Eshraghi, Amir Fathi, and Seree Iamphongsai. Factors contributing to the facial aging of identical twins. Plastic and reconstructive surgery, 123(4):1321-1331, 2009. 1 +[17] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics), 28(1):100-108, 1979. 5 +[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3 +[19] Gee-Sern Hsu, Rui-Cang Xie, and Zhi-Ting Chen. Wasserstein divergence GAN with cross-age identity expert and attribute retainer for facial age transformation. IEEE Access, 9:39695-39706, 2021. 3 +[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3 +[21] Zhizhong Huang, Junping Zhang, and Hongming Shan. When age-invariant face recognition meets face age synthesis: A multi-task learning framework. In CVPR, 2021. 6 +[22] Zhizhong Huang, Junping Zhang, and Hongming Shan. When age-invariant face recognition meets face age synthesis: A multi-task learning framework. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7282-7291, 2021. 5 +[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6 +[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with + +frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5 +[25] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8640-8650, 2024. 3 +[26] Haiqi Liu, C. L. Philip Chen, Xinrong Gong, and Tong Zhang. Robust saliency-aware distillation for few-shot fine-grained visual recognition. IEEE Transactions on Multimedia, 26:7529-7542, 2024. 3 +[27] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, pages 16784-16804. PMLR, 2022. 1, 3 +[28] Roy Or-El, Soumyadip Sengupta, Ohad Fried, Eli Shechtman, and Ira Kemelmacher-Shlizerman. Lifespan age transformation synthesis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI, pages 739-755. Springer, 2020. 1, 3 +[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3 +[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 1, 3 +[31] Helle Rexbye, Inge Petersen, Mette Johansens, Louise Klitkou, Bernard Jeune, and Kaare Christensen. Influence of environmental factors on facial ageing. Age and Ageing, 35 (2):110-115, 2006. 1 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 1, 3 +[33] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Dex: Deep expectation of apparent age from a single image. In Proceedings of the IEEE international conference on computer vision workshops, pages 10-15, 2015. 4 +[34] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Deep expectation of real and apparent age from a single image without facial landmarks. International Journal of Computer Vision, 126(2):144-157, 2018. 4 +[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, + +Canada, June 17-24, 2023, pages 22500-22510. IEEE, 2023. 3 +[36] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3, 6 +[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. 1, 3 +[38] Kestutis Sveikata, Irena Balciuniene, Janina Tutkuviene, et al. Factors influencing face aging. literature review. Stomatologija, 13(4):113-116, 2011. 1 +[39] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 4 +[40] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3 +[41] Zongwei Wang, Xu Tang, Weixin Luo, and Shenghua Gao. Face aging with identity-preserved conditional generative adversarial networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 7939-7947. Computer Vision Foundation / IEEE Computer Society, 2018. 1, 3 +[42] Hongyu Yang, Di Huang, Yunhong Wang, and Anil K. Jain. Learning continuous face age progression: A pyramid of gans. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):499-515, 2021. 3 +[43] Xu Yao, Gilles Puy, Alasdair Newson, Yann Gousseau, and Pierre Hellier. High resolution face age editing. In 25th International Conference on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021, pages 8624-8631. IEEE, 2020. 1, 3 +[44] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 6 +[45] Cheng Yu, Haoyu Xie, Lei Shang, Yang Liu, Jun Dan, Liefeng Bo, and Baigui Sun. Facechain-fact: Face adapter with decoupled training for identity-preserved personalization. arXiv preprint arXiv:2410.12312, 2024. 3, 4, 6 +[46] Zongyan Zhang, Haohan Weng, Tong Zhang, and C. L. Philip Chen. A broad generative network for two-stage image outpainting. IEEE Transactions on Neural Networks and Learning Systems, 35(9):12731-12745, 2024. 1 + +[47] Zongyan Zhang, C. L. Philip Chen, Haohan Weng, and Tong Zhang. Self-prompt guided image outpainting model for captions absence in social scenes. IEEE Transactions on Computational Social Systems, pages 1-12, 2025. 1 \ No newline at end of file diff --git a/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/images.zip b/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1289c3549e1f42341e02ebdb9062d33b006313c7 --- /dev/null +++ b/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8904d9204f017b0259721dff8cbde5a5754207f9b4c2a51bd34d50063582f7d4 +size 1018416 diff --git a/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/layout.json b/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a88d3e79f3b6d46298d82fe527b16176e80ac191 --- /dev/null +++ b/2025/TimeBooth_ Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging/layout.json @@ -0,0 +1,8448 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 102, + 534, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 102, + 534, + 140 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 534, + 140 + ], + "type": "text", + "content": "TimeBooth: Disentangled Facial Invariant Representation for Diverse and Personalized Face Aging" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 159, + 515, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 159, + 515, + 176 + ], + "spans": [ + { + "bbox": [ + 94, + 159, + 515, + 176 + ], + "type": "text", + "content": "Zepeng Su" + }, + { + "bbox": [ + 94, + 159, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 176 + ], + "type": "text", + "content": ", Zhulin Liu" + }, + { + "bbox": [ + 94, + 159, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 94, + 159, + 515, + 176 + ], + "type": "text", + "content": ", Zongyan Zhang" + }, + { + "bbox": [ + 94, + 159, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 94, + 159, + 515, + 176 + ], + "type": "text", + "content": ", Tong Zhang" + }, + { + "bbox": [ + 94, + 159, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,3,4}" + }, + { + "bbox": [ + 94, + 159, + 515, + 176 + ], + "type": "text", + "content": ", C. L. Philip Chen" + }, + { + "bbox": [ + 94, + 159, + 515, + 176 + ], + "type": "inline_equation", + "content": "^{1,2,3,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 204, + 177, + 407, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 177, + 407, + 189 + ], + "spans": [ + { + "bbox": [ + 204, + 177, + 407, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 204, + 177, + 407, + 189 + ], + "type": "text", + "content": "South China University of Technology " + }, + { + "bbox": [ + 204, + 177, + 407, + 189 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 204, + 177, + 407, + 189 + ], + "type": "text", + "content": "Pazhou Lab" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 190, + 481, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 190, + 481, + 203 + ], + "spans": [ + { + "bbox": [ + 129, + 190, + 481, + 203 + ], + "type": "text", + "content": "3Guangdong Provincial Key Laboratory of Computational AI Models and Cognitive Intelligence" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 205, + 525, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 205, + 525, + 217 + ], + "spans": [ + { + "bbox": [ + 82, + 205, + 525, + 217 + ], + "type": "text", + "content": "4Engineering Research Center of the Ministry of Education on Health Intelligent Perception and Paralleled Digital-Human" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 238, + 105, + 285 + ], + "blocks": [ + { + "bbox": [ + 58, + 238, + 105, + 285 + ], + "lines": [ + { + "bbox": [ + 58, + 238, + 105, + 285 + ], + "spans": [ + { + "bbox": [ + 58, + 238, + 105, + 285 + ], + "type": "image", + "image_path": "8637756c8197d5ae7f632050474a8555d9d09e5610dcaeccdfc6e19f47922052.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 289, + 92, + 297 + ], + "lines": [ + { + "bbox": [ + 69, + 289, + 92, + 297 + ], + "spans": [ + { + "bbox": [ + 69, + 289, + 92, + 297 + ], + "type": "text", + "content": "Nemo" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 298, + 96, + 308 + ], + "lines": [ + { + "bbox": [ + 66, + 298, + 96, + 308 + ], + "spans": [ + { + "bbox": [ + 66, + 298, + 96, + 308 + ], + "type": "text", + "content": "(Source)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 118, + 234, + 315, + 291 + ], + "blocks": [ + { + "bbox": [ + 118, + 234, + 315, + 291 + ], + "lines": [ + { + "bbox": [ + 118, + 234, + 315, + 291 + ], + "spans": [ + { + "bbox": [ + 118, + 234, + 315, + 291 + ], + "type": "image", + "image_path": "5b571f28f5bed9f37e3aeeb3c5b74bea5a8db47ae8bb0b7e783684137c023373.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 293, + 244, + 303 + ], + "lines": [ + { + "bbox": [ + 184, + 293, + 244, + 303 + ], + "spans": [ + { + "bbox": [ + 184, + 293, + 244, + 303 + ], + "type": "text", + "content": "Film Mr. Nobody" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 224, + 552, + 308 + ], + "blocks": [ + { + "bbox": [ + 321, + 224, + 552, + 308 + ], + "lines": [ + { + "bbox": [ + 321, + 224, + 552, + 308 + ], + "spans": [ + { + "bbox": [ + 321, + 224, + 552, + 308 + ], + "type": "image", + "image_path": "c08c451133fe62cb2cbc55a076a6abbbf4c2c284c0401b38ad93a09096a17c71.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 555, + 340 + ], + "type": "text", + "content": "Figure 1. In Mr. Nobody, the protagonist Nemo traverses multiple parallel realities, showcasing distinct aged appearances. This illustrates how choices such as environmental factors and lifestyle influence the diversity of face aging. Inspired by this, we propose TimeBooth." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 151, + 367, + 200, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 367, + 200, + 379 + ], + "spans": [ + { + "bbox": [ + 151, + 367, + 200, + 379 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 391, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 391, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 391, + 297, + 689 + ], + "type": "text", + "content": "Face aging is a typical ill-posed problem influenced by various factors such as environment and genetics, leading to highly diverse outcomes. However, existing methods primarily rely on numerical age representations, making it difficult to accurately capture individual or group-level aging patterns. To address this, we introduce a novel disentangled face representation, where age features are modeled in the image modality—referred to as the Age Prompt—providing richer prior age information to constrain the generation results. To this end, we design an ID-age multi-task colearning framework and propose the Bidirectional Adversarial Disentanglement(BAD) strategy. This strategy maximizes the disentanglement of ID and age representation through bidirectional adversarial learning, extracting their attribute-invariant representations. Based on this representation, we propose TimeBooth, a personalized face aging model capable of generating diverse and individualized aging results. To optimize training, we construct a cross-age hybrid data pipeline and introduce various training strategies. Finally, we propose the R-AgeMAE metric and validate our method through extensive experiments, demonstrating that TimeBooth outperforms existing methods in both diversity and controllability. Our code will be publicly available at https://github.com/zpsuu/TimeBooth." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 367, + 394, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 367, + 394, + 379 + ], + "spans": [ + { + "bbox": [ + 315, + 367, + 394, + 379 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 556 + ], + "type": "text", + "content": "In recent years, large-scale text-to-image generation models [27, 30, 32, 37, 46, 47] have demonstrated remarkable controllable generation capabilities, transforming imagination into reality. Face aging, also known as face age transformation, enables the synthesis of photorealistic face images at different ages by modifying age-related facial attributes while preserving identity-consistent features, such as facial structure, contour and so on. This technology has broad applications, including film post-production, forensic age progression for missing persons, and medical diagnostics. With the advancement of personalized generation techniques, face aging can achieve more precise and individualized simulations, further enhancing its real-world applicability." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Face aging is a typical ill-posed problem, influenced by various factors [16, 31, 38] such as environment, genetics, lifestyle, and living conditions. For instance, individuals of the same age may exhibit distinct aging patterns—prolonged sun exposure can lead to increased freckles and skin laxity, while those carrying the MCR1 gene are more sensitive to light and prone to developing wrinkles. However, existing face aging methods [1, 13, 28, 41, 43] are predominantly based on GANs [15] and follow an image-to-image translation paradigm. These methods typically use age as the sole constraint—age, resulting in generated images that merely transfer the average aging characteristics learned from the dataset. Consequently, the existing" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 277, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 277, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 277, + 713 + ], + "type": "text", + "content": "*Corresponding author: Zhulin Liu (e-mail: liuzhl@scut.edu.cn)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12147" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 228 + ], + "type": "image", + "image_path": "1c5756e0033b565dcc753b7919a3dbe7b6bff2063846509972980517a0916649.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 236, + 555, + 295 + ], + "lines": [ + { + "bbox": [ + 54, + 236, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 555, + 295 + ], + "type": "text", + "content": "Figure 2. Training pipelines of our proposed TimeBooth. It consists of two stages: (a) Through ID-age collaborative training, we learn ID and age feature representations and introduce the BAD strategy, where bidirectional adversarial disentanglement based on mutual GRL layers is applied to extract attribute-invariant features. " + }, + { + "bbox": [ + 54, + 236, + 555, + 295 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 54, + 236, + 555, + 295 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 236, + 555, + 295 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 54, + 236, + 555, + 295 + ], + "type": "text", + "content": " are positive numbers greater than 0, and (b) The disentangled encoder is frozen. Based on the learned disentangled representation, fusion is performed in the Age Adapter, followed by image cross-attention layers, which are used with a diffusion model for generation. Multiple training strategies are optimized concurrently during this process." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 313, + 294, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 313, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 313, + 294, + 350 + ], + "type": "text", + "content": "paradigm can only provide a seemingly reasonable solution to this ill-posed problem but fails to accurately capture the diversity of individual or group-level aging patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 352, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 295, + 495 + ], + "type": "text", + "content": "Fundamentally, the limited expressiveness of existing paradigms stems from the inadequate representation of age as a constraint. As a numerical modality, age is inherently an abstract and discrete label, making it insufficient to capture the diversity and complexity of facial features at different ages. Consequently, it fails to provide the model with adequate semantic guidance for diverse and controllable age transformation. Therefore, it is essential to introduce a constraint with a richer age representation that properly regularizes the ill-posed nature of face aging, leading to more precise and well-defined solutions for personalized and diverse face aging." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "As the saying goes, \"a picture is worth a thousand words.\" As a highly concretized modality, images naturally provide rich and intuitive information. Inspired by this, we propose a novel disentangled face representation, where the ID-invariant feature and age-related feature are disentangled, and age is modeled using an image modality, referred to as the Age Prompt. To this end, we construct a multi-task collaborative learning framework, leveraging ID recognition and age classification tasks for joint supervision. We propose a bidirectional adversarial disentanglement strategy, which introduces mutual gradient reversal layers to maximize feature disentanglement, thereby learning attribute-invariant representations. Based on this representation, we further design an Age Adapter, in which a lightweight Q-former is employed to fuse the disentangled features, while the image cross-attention layers inject facial information into the diffusion model for generation. To train the Age Adapter, we build a cross-age hybrid dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 313, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 555, + 373 + ], + "type": "text", + "content": "pipeline and propose multiple training strategies to ensure the correct alignment of disentangled features. Finally, we introduce the R-AgeMAE as a metric to evaluate the accuracy of personalized face aging. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 398, + 553, + 696 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 398, + 553, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 553, + 482 + ], + "type": "text", + "content": "- We propose a novel disentangled facial representation, where age constraints are modeled using an image modality instead of the traditional numerical modality. To the best of our knowledge, this is the first attempt to represent age using image prompts, providing a richer age prior for the inherently ill-posed problem of face aging and introducing a new generative paradigm." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 483, + 553, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 553, + 540 + ], + "type": "text", + "content": "- We construct an ID-age multi-task co-learning framework and propose a bidirectional adversarial disentanglement strategy to maximize the disentanglement of ID and age representations, enabling the learning of their attribute-invariant representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 542, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 553, + 601 + ], + "type": "text", + "content": "- We introduce TimeBooth, which enables users to generate more diverse and personalized aging face images. To enhance training, we construct a hybrid cross-age dataset and design multiple training strategies to optimize the Age Adapter." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 602, + 553, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 553, + 696 + ], + "type": "text", + "content": "- We introduce the R-AgeMAE metric to evaluate the accuracy of personalized face aging. Extensive experiments demonstrate the superiority of TimeBooth in terms of diversity, and controllability. Moreover, TimeBooth supports plug-and-play integration, seamlessly incorporating into existing community models to provide efficient personalized face aging generation in just a few seconds without test-time optimizing." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12148" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 289, + 266 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 289, + 266 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 289, + 266 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 289, + 266 + ], + "type": "image", + "image_path": "e88a88ede282596d013a42efef7b51842bd9097fc820489b301a10370cf6034f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 274, + 258, + 285 + ], + "lines": [ + { + "bbox": [ + 93, + 274, + 258, + 285 + ], + "spans": [ + { + "bbox": [ + 93, + 274, + 258, + 285 + ], + "type": "text", + "content": "Figure 3. Inference pipeline of TimeBooth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 304, + 143, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 143, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 143, + 317 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 173, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 173, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 173, + 337 + ], + "type": "text", + "content": "2.1. Face Aging methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 559 + ], + "type": "text", + "content": "In recent years, most face age transformation methods [1, 13, 28, 41, 43] have been GAN-based [15], following an image-to-image generation paradigm where age serves as a conditioning factor. Many approaches [2, 19, 41, 42] utilize pre-trained face recognition models and ID-preserving loss terms to maintain facial identity, while others achieve transformation via style transfer. CUSP [13] disentangles content and style encoding, using a content mask for customizable editing. With the rise of diffusion models [27, 32], recent methods have explored their potential for face age transformation. Inspired by [35], [3] introduces specific identifiers and an age regularization set to fine-tune diffusion models. FADING [7] applies Prompt-to-Prompt [18] local editing, modifying attention maps during generation. Additionally, general image editing approaches based on attention mechanisms [5, 18] can also perform face age transformation by leveraging pre-trained text-to-image priors, though with potentially lower accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 564, + 194, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 194, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 194, + 576 + ], + "type": "text", + "content": "2.2. Personalization methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Although large-scale text-to-image diffusion models [27, 30, 32, 37] possess rich prior knowledge, they are limited to generating content from their pre-trained corpus and struggle to meet personalized user needs. Researchers have explored personalized generation methods. One method [36] involves test-time fine-tuning, where reference images are used to adapt diffusion models at inference, often with LoRA [20] for efficiency. Another method [40, 44, 45] leverages adapter-based architectures, learning feature representations to preserve identity without test-time training. Many approaches also incorporate recognition-based [26]" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 542, + 188 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 542, + 188 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 542, + 188 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 542, + 188 + ], + "type": "image", + "image_path": "efe4e89776967d16ab3f7debebf29683e81dd772ee42c2134e2840957af43e72.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 198, + 534, + 210 + ], + "lines": [ + { + "bbox": [ + 333, + 198, + 534, + 210 + ], + "spans": [ + { + "bbox": [ + 333, + 198, + 534, + 210 + ], + "type": "text", + "content": "Figure 4. Visualization of disentangled representation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 234, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 366 + ], + "type": "text", + "content": "embeddings into the generation process. IP-Adapter-Faceid [44] combines face embeddings with CLIP [29] image encoders and employs disentangled cross-attention for generation. PhotoMaker [25] stacks ID embeddings for higher fidelity but struggles with fewer reference images. InstantID [40], built on IP-Adapter [44], introduces a facial keypoint control branch to improve shape fidelity at the cost of editability. FaceChain-FACT [45] utilizes a pre-trained ViT face recognition model [10] and a gated cross-attention layer to enhance instruction-following without compromising identity preservation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 380, + 375, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 380, + 375, + 392 + ], + "spans": [ + { + "bbox": [ + 314, + 380, + 375, + 392 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 401, + 400, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 400, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 400, + 414 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 420, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 555, + 479 + ], + "type": "text", + "content": "Latent Diffusion Model. Latent diffusion models [32] are a class of generative models. During training, noise is progressively added to the latent representation of the original image, " + }, + { + "bbox": [ + 313, + 420, + 555, + 479 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 420, + 555, + 479 + ], + "type": "text", + "content": ", to obtain " + }, + { + "bbox": [ + 313, + 420, + 555, + 479 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 313, + 420, + 555, + 479 + ], + "type": "text", + "content": " through a forward process. This process can be formulated as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 382, + 484, + 553, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 484, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 382, + 484, + 553, + 498 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {\\alpha_ {t}} z _ {0} + \\sqrt {1 - \\alpha_ {t}} \\varepsilon \\tag {1}", + "image_path": "08e59065223e275193dd1517b83685686399de76ed558769a75f3d68f6cfed38.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "type": "text", + "content": "The denoising network needs to accept the time step " + }, + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "type": "text", + "content": " and the condition " + }, + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "type": "text", + "content": " to estimate the noise. After multiple steps of denoising, the latent representation of target image " + }, + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 502, + 555, + 550 + ], + "type": "text", + "content": " can be obtained, which can be expressed by the formula:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 554, + 555, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 554, + 555, + 595 + ], + "spans": [ + { + "bbox": [ + 314, + 554, + 555, + 595 + ], + "type": "interline_equation", + "content": "z _ {t - 1} = \\frac {\\sqrt {\\alpha_ {t - 1}}}{\\sqrt {\\alpha_ {t}}} z _ {t} + \\sqrt {\\alpha_ {t - 1}} \\left(\\sqrt {\\frac {1}{\\alpha_ {t - 1}} - 1} - \\sqrt {\\frac {1}{\\alpha_ {t}} - 1}\\right) \\varepsilon_ {\\theta} \\left(z _ {t}, t, c\\right) \\tag {2}", + "image_path": "a46a890ba52e1067cd6f32a489dfd47f1d9c6b7de131153044d0568dc9d10bf2.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 596, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 555, + 643 + ], + "type": "text", + "content": "During training, we estimate the noise added in the forward process and use the denoising network for denoising in the reverse process. The objective function for training the denoising network is:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 647, + 555, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 647, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 345, + 647, + 555, + 662 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D M} = E _ {z _ {0}, \\varepsilon \\sim \\mathcal {N} (0, \\mathcal {I}), c, t} \\| \\varepsilon - \\varepsilon_ {\\theta} (z _ {t}, c, t) \\| ^ {2} \\tag {3}", + "image_path": "ba52792a712e31b11b0ae4cb5ddc7d33d1d2cd2e332642e24e37ec126eb8b7eb.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 714 + ], + "type": "text", + "content": "Image-Prompt Adapter. The image adapter [44] is designed to enable a pre-trained text-to-image diffusion model to generate images conditioned on image prompts. It consists of an image encoder and a cross-attention module," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12149" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 553, + 246 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 246 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 246 + ], + "type": "image", + "image_path": "e26ee12c76d07491d3f24314d4df334d96866ba8f73235cdd922099501aadcd8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 253, + 555, + 288 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 555, + 288 + ], + "type": "text", + "content": "Figure 5. TimeBooth is capable of capturing fine-grained aging patterns for both individuals (left) and groups (right). As the number of Age Prompts increases, the generated results gradually shift from individual diversity to a stable group-level aging pattern, enhancing the interpretability and diversity of face aging synthesis in a data-driven manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 298, + 544, + 430 + ], + "blocks": [ + { + "bbox": [ + 69, + 298, + 544, + 430 + ], + "lines": [ + { + "bbox": [ + 69, + 298, + 544, + 430 + ], + "spans": [ + { + "bbox": [ + 69, + 298, + 544, + 430 + ], + "type": "image", + "image_path": "f8c2ae0e525c888b2a39699b561dc11eadab54b408b939b1ce35e95db605da22.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 441, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 555, + 464 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison with translation-based methods CUSP [14], IP2P [6], and FADING [8]. Compared to these methods, TimeBooth generates faces that are both more natural and diverse, effectively avoiding the issue of 'child faces with elderly textures.'" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 484, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 520 + ], + "type": "text", + "content": "where the encoded image prompts are fed into the cross-attention layers to guide image generation. The adapter is trained using the following loss function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 523, + 295, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 523, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 70, + 523, + 295, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I A} = E _ {z _ {0}, \\varepsilon \\sim \\mathcal {N} (0, \\mathcal {I}), c _ {t}, c _ {i}, t} \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, c _ {t}, c _ {i}, t\\right) \\| ^ {2} \\tag {4}", + "image_path": "cb873646cb38f55e06afc10ff5a61e3a0560634f242d17717e31448f90498be4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 542, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 542, + 296, + 567 + ], + "type": "inline_equation", + "content": "c_{t}, c_{i}" + }, + { + "bbox": [ + 55, + 542, + 296, + 567 + ], + "type": "text", + "content": " are text and image conditions after encoding text prompt and image prompt." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 575, + 255, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 255, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 255, + 588 + ], + "type": "text", + "content": "3.2. Disentangled Representation Encoder" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "ID-Age Representation Co-learning Framework. We divide facial representation into two components: the ID representation " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{id}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", which remain invariant across age transformations, and the age representation " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{age}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", which change with age. Based on previous work [45], using the output of a ViT-based face recognition model [10] as face representations has proven effective in achieving high-fidelity reconstructions in diffusion models. Following this idea, we adopt a similar method. However, we further differentiate the original output of the ViT model. Specifically, we append two" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 484, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 519 + ], + "type": "text", + "content": "separate neck networks to refine and disentangle the original features. The specific model details can be found in the Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 522, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 559 + ], + "type": "text", + "content": "One neck network is designed to extract ID age-invariant representation " + }, + { + "bbox": [ + 313, + 522, + 555, + 559 + ], + "type": "inline_equation", + "content": "c_{id}" + }, + { + "bbox": [ + 313, + 522, + 555, + 559 + ], + "type": "text", + "content": ". For this, we use the CosFace loss [39] " + }, + { + "bbox": [ + 313, + 522, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{COSFACE}}" + }, + { + "bbox": [ + 313, + 522, + 555, + 559 + ], + "type": "text", + "content": " for supervision:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 369, + 563, + 553, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 563, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 369, + 563, + 553, + 576 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {I D}} = \\mathcal {L} _ {\\text {C O S F A C E}} \\left(L \\left(X _ {\\mathrm {i d}}\\right), y _ {\\mathrm {i d}}\\right) \\tag {5}", + "image_path": "3c91ae6a4358e00dc895a04b7b87a9567e8ffd07b5ef54111ca4f4126fdac2f4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 579, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 553, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 579, + 553, + 604 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{id}}" + }, + { + "bbox": [ + 313, + 579, + 553, + 604 + ], + "type": "text", + "content": " represents the extracted ID features, " + }, + { + "bbox": [ + 313, + 579, + 553, + 604 + ], + "type": "inline_equation", + "content": "L(\\cdot)" + }, + { + "bbox": [ + 313, + 579, + 553, + 604 + ], + "type": "text", + "content": " denotes the head network, " + }, + { + "bbox": [ + 313, + 579, + 553, + 604 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{id}}" + }, + { + "bbox": [ + 313, + 579, + 553, + 604 + ], + "type": "text", + "content": " is the corresponding ID label." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 630 + ], + "type": "text", + "content": "Meanwhile, the other neck network is used to extract age features " + }, + { + "bbox": [ + 313, + 605, + 555, + 630 + ], + "type": "inline_equation", + "content": "c_{age}" + }, + { + "bbox": [ + 313, + 605, + 555, + 630 + ], + "type": "text", + "content": ". We use the mse loss as the supervision loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 634, + 553, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 634, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 361, + 634, + 553, + 647 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {A g e}} = \\mathcal {L} _ {\\mathrm {M S E}} (\\mathrm {D E X} (A (X _ {\\mathrm {a g e}})), y _ {\\mathrm {a g e}}) \\tag {6}", + "image_path": "b90388da3544be61cd93d492ad970e0422f03ddb9cb13b28abc1b79caea61f12.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 651, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 553, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 651, + 553, + 688 + ], + "type": "inline_equation", + "content": "X_{\\mathrm{age}}" + }, + { + "bbox": [ + 313, + 651, + 553, + 688 + ], + "type": "text", + "content": " represents the extracted age features, " + }, + { + "bbox": [ + 313, + 651, + 553, + 688 + ], + "type": "inline_equation", + "content": "A(\\cdot)" + }, + { + "bbox": [ + 313, + 651, + 553, + 688 + ], + "type": "text", + "content": " is the head network, and " + }, + { + "bbox": [ + 313, + 651, + 553, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{DEX}(\\cdot)" + }, + { + "bbox": [ + 313, + 651, + 553, + 688 + ], + "type": "text", + "content": " is the deep expectation [33, 34], " + }, + { + "bbox": [ + 313, + 651, + 553, + 688 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{age}}" + }, + { + "bbox": [ + 313, + 651, + 553, + 688 + ], + "type": "text", + "content": " is the ground-truth age label." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Bidirectional Adversarial Disentanglement(BAD) Strategy for Disentangling Representation. Supervised co" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12150" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "content": "learning will result in overlap between ID and age representation. To enable cross-age face synthesis by modifying only " + }, + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "inline_equation", + "content": "c_{age}" + }, + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "content": ", we aim to maximize the disentanglement of these two representations. Typically, face recognition features encapsulate all identifiable attributes of an individual, including age-related characteristics. However, such entanglement compromises feature purity, thereby affecting controllability during generation. Therefore, inspired by [22], we seek to learn an age-invariant ID representation that remains stable across varying age, eliminating the influence of age on ID recognition. To achieve this, we introduce an adversarial learning strategy to enhance representation disentanglement. Specifically, we incorporate a Gradient Reversal Layer (GRL) [12] before the age classification head. During backpropagation, the GRL reverses the gradient, forcing the ID representation to become age-invariant while retaining identity-specific information. The adversarial constraint is imposed through the following loss function:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 292, + 295, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 292, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 93, + 292, + 295, + 305 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {A d v I D}} = \\mathcal {L} _ {\\mathrm {C O S F A C E}} \\left(L \\left(\\operatorname {G R L} \\left(X _ {\\text {a g e}}\\right)\\right), y _ {\\mathrm {i d}}\\right) \\tag {7}", + "image_path": "68f4ddd823b3649fe4238da4850cf42f5ab167bd2b8ab02ea22e91888f388f04.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 309, + 243, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 243, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 243, + 321 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 309, + 243, + 321 + ], + "type": "inline_equation", + "content": "\\mathrm{GRL}(\\cdot)" + }, + { + "bbox": [ + 55, + 309, + 243, + 321 + ], + "type": "text", + "content": " is the Gradient Reversal Layer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 322, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 417 + ], + "type": "text", + "content": "Similarly, to prevent ID information from being encoded in age features—ensuring that interpolation at inference time does not compromise the fidelity of generated faces—we introduce an adversarial constraint on the age feature extraction branch. Specifically, we incorporate a Gradient Reversal Layer (GRL) before the ID classification head, forcing the extracted age representation to become ID-invariant. The loss function is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 422, + 295, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 422, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 84, + 422, + 295, + 436 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {A d v A g e}} = \\mathcal {L} _ {\\text {M S E}} (\\operatorname {D E X} (A (\\operatorname {G R L} (X _ {\\mathrm {i d}}))), y _ {\\text {a g e}}) \\tag {8}", + "image_path": "06006a3783d6a1512c580ed4819ac1958fee3b7d35652b933a8a01b9f9b835c8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 440, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 296, + 487 + ], + "type": "text", + "content": "By leveraging this bidirectional adversarial learning strategy, we achieve more effective disentanglement of ID and age representations. The total loss of the disentangled encoder is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 492, + 295, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 492, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 492, + 295, + 516 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D R E}} = \\mathcal {L} _ {\\mathrm {I D}} + \\lambda_ {\\mathrm {A g e}} \\mathcal {L} _ {\\mathrm {A g e}} + \\lambda_ {\\mathrm {A d v I D}} \\mathcal {L} _ {\\mathrm {A d v I D}} + \\lambda_ {\\mathrm {A d v A g e}} \\mathcal {L} _ {\\mathrm {A d v A g e}} \\tag {9}", + "image_path": "07c7dd67cb8ff8398edb9fda4ee23e141d39e207d015c60f3b2d361248df63bb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{Age}}" + }, + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{AdvID}}" + }, + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{AdvAge}}" + }, + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "type": "text", + "content": " are weighting coefficients balancing each loss term." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 551, + 254, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 254, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 254, + 563 + ], + "type": "text", + "content": "3.3. Face Reconstruction via Age Adapter" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 569, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 616 + ], + "type": "text", + "content": "To achieve personalized face aging, we first perform face reconstruction during training. We propose the Age Adapter, which integrates the Age-Conditioned ID Encoder and multiple image cross-attention modules." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": "Age-conditioned Face Encoding. To fuse disentangled representations into a complete ID representation with age information, we propose the Age-Conditioned ID Encoder. It begins by applying dual cross attention modules to facilitate interaction between " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "c_{id}" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "c_{age}" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": ", which are derived from the disentangled representation encoder. This step ensures that " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "c_{id}" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": " are enriched with necessary age-related representation " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "c_{age}" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": ", while the " + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "inline_equation", + "content": "c_{age}" + }, + { + "bbox": [ + 55, + 618, + 296, + 715 + ], + "type": "text", + "content": " is simultaneously refined to" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "align with the ID representation " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "c_{id}" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": ". Once the age-enhanced ID representation and ID-conditioned age representation are obtained, we employ a lightweight Q-Former [24] to adaptively learn the final comprehensive facial representation. Specifically, we introduce 32 learnable tokens that interact with the processed feature representations, allowing the model to capture a holistic and semantically rich ID representation while incorporating the essential age attributes. These learnable tokens dynamically adjust their latent space embeddings, ensuring that the final identity feature faithfully preserves the original identity while seamlessly incorporating the desired age characteristics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 216, + 555, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 275 + ], + "type": "text", + "content": "Face Reconstruction. To enable face reconstruction using the fused representations, we integrate multiple image cross attention layers within a text-to-image model to effectively extract and incorporate information from the Age-conditioned ID Encoder " + }, + { + "bbox": [ + 313, + 216, + 555, + 275 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 216, + 555, + 275 + ], + "type": "text", + "content": ", as formulated below:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 280, + 558, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 280, + 558, + 304 + ], + "spans": [ + { + "bbox": [ + 314, + 280, + 558, + 304 + ], + "type": "interline_equation", + "content": "Z = \\text {A t t e n t i o n} _ {\\text {t e x t}} (Q, K, V) + \\lambda \\cdot \\text {A t t e n t i o n} _ {\\text {i m g}} (Q, K ^ {\\prime}, V ^ {\\prime}) \\tag {10}", + "image_path": "fc7de46c30b96cd00bdf7c869a76949b2eb697e3eb7b17ba337e53afefd75729.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 304, + 554, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 352 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 313, + 304, + 554, + 352 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 304, + 554, + 352 + ], + "type": "text", + "content": " is weight coefficient of image cross attention layers, " + }, + { + "bbox": [ + 313, + 304, + 554, + 352 + ], + "type": "inline_equation", + "content": "Q = ZW_{q}, K = c_{t}W_{k}, V = c_{t}W_{v}, K' = \\mathcal{E}(c_{id}, c_{age})W'_k, V' = \\mathcal{E}(c_{id}, c_{age})W'_v" + }, + { + "bbox": [ + 313, + 304, + 554, + 352 + ], + "type": "text", + "content": " are the query, key, and values matrices from the text and face representation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 353, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 554, + 388 + ], + "type": "text", + "content": "Finally, we optimize the Age Adapter and the unfrozen components of the Age-conditioned ID Encoder via a reconstruction task, with the objective defined as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 392, + 553, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 392, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 319, + 392, + 553, + 417 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D M} = E _ {z _ {0}, \\varepsilon \\sim \\mathcal {N} (0, \\mathcal {I}), c _ {t}, c _ {i d}, c _ {a g e}, t} \\| \\varepsilon - \\varepsilon_ {\\theta} \\left(z _ {t}, c _ {t}, \\mathcal {E} \\left(c _ {i d}, c _ {a g e}\\right), t\\right) \\| ^ {2} \\tag {11}", + "image_path": "57561a269a9447e55625d20d6bd9219ad3c2c50e044b784673aac33d5769252f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 423, + 506, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 506, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 506, + 436 + ], + "type": "text", + "content": "3.4. Training Strategies for Age Adapter" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 440, + 554, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 554, + 524 + ], + "type": "text", + "content": "Shuffle Strategy. To ensure that the age-invariant ID representation " + }, + { + "bbox": [ + 313, + 440, + 554, + 524 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{id}}" + }, + { + "bbox": [ + 313, + 440, + 554, + 524 + ], + "type": "text", + "content": " of the same identity maintains strong intraclass consistency in the feature space, we propose the Shuffle Strategy and design a cross-age data pipeline. During training, with probability " + }, + { + "bbox": [ + 313, + 440, + 554, + 524 + ], + "type": "inline_equation", + "content": "p_s" + }, + { + "bbox": [ + 313, + 440, + 554, + 524 + ], + "type": "text", + "content": ", we dynamically replace the ID feature " + }, + { + "bbox": [ + 313, + 440, + 554, + 524 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{id}}^i" + }, + { + "bbox": [ + 313, + 440, + 554, + 524 + ], + "type": "text", + "content": " of a sample with that of the same identity at a different age, formulated as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 333, + 528, + 553, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 528, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 333, + 528, + 553, + 543 + ], + "type": "interline_equation", + "content": "\\tilde {c} _ {\\mathrm {i d}} ^ {i} = \\mathbb {I} \\cdot \\mathbf {c} _ {\\mathrm {i d}} ^ {j} + (1 - \\mathbb {I}) \\cdot c _ {\\mathrm {i d}} ^ {i}, \\mathbb {I} \\sim \\operatorname {B e r n o u l l i} \\left(p _ {s}\\right) \\tag {12}", + "image_path": "dc6a477bb9c242c224ce49c82fb95ea862a9bc89f6c7b736bfaa7ee74cf57a51.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "text", + "content": " follows a Bernoulli distribution with parameter " + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "inline_equation", + "content": "p_s" + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "text", + "content": " represents another image with the same identity but a different age (" + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "inline_equation", + "content": "j \\neq i, y_j = y_i" + }, + { + "bbox": [ + 313, + 546, + 554, + 617 + ], + "type": "text", + "content": "). This strategy not only enhances the learning effectiveness of the Age Adapter but also prevents overfitting due to direct reconstruction of the same image, thereby improving model generalization." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": "Age Prototype Replacement(APR) Strategy. To prevent the Age Prompt from capturing spurious age-unrelated correlations and to leverage population-level statistical regularities of age attributes, we propose the APR strategy. We first extract age features " + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{c}_{\\mathrm{age}}^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": " from all training samples and perform offline K-means clustering [17] (precomputed before training) to group them into " + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": " cohorts. Each cohort's prototype vector " + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{age}}^k" + }, + { + "bbox": [ + 313, + 618, + 555, + 715 + ], + "type": "text", + "content": " encapsulates the average age patterns" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12151" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": "of its members. During training, the original feature " + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{age}}^i" + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": " of each sample is randomly replaced with its cohort prototype " + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathrm{age}}^{k(i)}" + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": " with probability " + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "inline_equation", + "content": "p_a" + }, + { + "bbox": [ + 55, + 72, + 294, + 122 + ], + "type": "text", + "content": ", effectively suppressing noisy individual-specific extremes:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 125, + 295, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 142 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 142 + ], + "type": "interline_equation", + "content": "\\tilde {c} _ {\\text {a g e}} ^ {i} = \\mathbb {I} \\cdot \\mu_ {\\text {a g e}} ^ {k (i)} + (1 - \\mathbb {I}) \\cdot c _ {\\text {a g e}} ^ {i}, \\quad \\mathbb {I} \\sim \\operatorname {B e r n o u l l i} \\left(p _ {a}\\right). \\tag {13}", + "image_path": "5bbbbf2bd93c3c0eb1c8da34520d87c7b499d6f3eb601f5bd6f36179eb5e968f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 296, + 310 + ], + "type": "text", + "content": "Dynamic Training-set Sampling(DTS) Strategy. Current face image datasets often exhibit a long-tailed age distribution, where training on raw data may cause feature overfitting in well-represented age groups and suboptimal learning in underrepresented ones. To address this, we propose the DTS strategy. Before training, age labels are discretized into equal intervals. During training, adaptive sample selection is performed based on the statistical median: for age groups exceeding the median sample count, only a representative subset matching the median is retained per epoch, while all samples are used for underrepresented groups. This strategy improves face generation across age ranges and mitigates data imbalance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 348 + ], + "type": "text", + "content": "Drop Strategy. Similar to [44, 45], we randomly mask " + }, + { + "bbox": [ + 55, + 312, + 296, + 348 + ], + "type": "inline_equation", + "content": "c_{id}" + }, + { + "bbox": [ + 55, + 312, + 296, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 312, + 296, + 348 + ], + "type": "inline_equation", + "content": "c_{age}" + }, + { + "bbox": [ + 55, + 312, + 296, + 348 + ], + "type": "text", + "content": " with probability " + }, + { + "bbox": [ + 55, + 312, + 296, + 348 + ], + "type": "inline_equation", + "content": "p_d" + }, + { + "bbox": [ + 55, + 312, + 296, + 348 + ], + "type": "text", + "content": " during training. During inference, the denoising process is formulated as follow:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 353, + 294, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 353, + 294, + 389 + ], + "spans": [ + { + "bbox": [ + 66, + 353, + 294, + 389 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {t - 1} = \\left(1 + \\lambda_ {\\mathrm {C F G}}\\right) \\cdot \\mathcal {H} \\left(z _ {t}, t, c _ {\\text {t e x t}}, \\mathcal {E} \\left(c _ {i d}, c _ {a g e}\\right)\\right) \\\\ - \\lambda_ {\\text {C F G}} \\mathcal {H} \\left(z _ {t}, t, c _ {\\text {t e x t - n e g}}, \\mathcal {E} \\left(c _ {i d - d r o p}, c _ {a g e - d r o p}\\right)\\right) \\tag {14} \\\\ \\end{array}", + "image_path": "ea63202ddbc19bed7d4597f7eb9eabed7b5fc0557dff21f6618f3d56951a1818.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "type": "text", + "content": " represents the forward process for denoising U-Net and " + }, + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{CFG}}" + }, + { + "bbox": [ + 55, + 390, + 295, + 415 + ], + "type": "text", + "content": " is the CFG scale." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 423, + 275, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 275, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 275, + 436 + ], + "type": "text", + "content": "3.5. Personalized Face Aging during Inference" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "content": "During the inference phase (see Fig. 3), we employ the disentangled representation encoder to extract " + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "inline_equation", + "content": "c_{id}" + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "inline_equation", + "content": "c_{age}^{src}" + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "content": " from the source image. To achieve more controllable age transformation, we propose a fusion strategy based on Spherical Linear Interpolation. This method interpolates between the source age feature " + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "inline_equation", + "content": "c_{age}^{src}" + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "content": " and the target Age Prompt " + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "inline_equation", + "content": "c_{age}^{tgt}" + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "content": " on a unit hypersphere, generating a fused age feature " + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "inline_equation", + "content": "c_{age}^{*}" + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "content": " with the hyperparameter aging guidance " + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 441, + 296, + 548 + ], + "type": "text", + "content": " controlling the intervention intensity of the target attribute:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 552, + 295, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 552, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 82, + 552, + 295, + 576 + ], + "type": "interline_equation", + "content": "c _ {a g e} ^ {*} = \\frac {\\sin [ (1 - \\alpha) \\theta ]}{\\sin \\theta} c _ {a g e} ^ {s r c} + \\frac {\\sin [ \\alpha \\theta ]}{\\sin \\theta} c _ {a g e} ^ {t g t}, \\tag {15}", + "image_path": "096e5da40355f309788a8813027ede28bd05ba8a4494a5805f2460dd8ed7bef8.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 579, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 295, + 602 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 579, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\theta = \\arccos \\left(c_{age}^{src} \\cdot c_{age}^{tgt}\\right)" + }, + { + "bbox": [ + 55, + 579, + 295, + 602 + ], + "type": "text", + "content": " is the angular distance between feature vectors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 604, + 296, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 296, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 296, + 630 + ], + "type": "text", + "content": "Finally, the cross-age face images are synthesized with identity consistency by combining " + }, + { + "bbox": [ + 55, + 604, + 296, + 630 + ], + "type": "inline_equation", + "content": "c_{id}" + }, + { + "bbox": [ + 55, + 604, + 296, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 604, + 296, + 630 + ], + "type": "inline_equation", + "content": "c_{age}^*" + }, + { + "bbox": [ + 55, + 604, + 296, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 639, + 137, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 137, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 137, + 653 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 659, + 167, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 167, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 167, + 673 + ], + "type": "text", + "content": "4.1. Experiments Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Dataset construction. In the first training stage, we utilize the MS1Mv2 [11] dataset for ID recognition and age estimation, with age annotations provided by [21]. In the" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 554, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 180 + ], + "type": "text", + "content": "second stage, we construct a hybrid dataset containing approximately 500K high-quality facial images. This hybrid dataset consists of two parts: one is a general crossage facial dataset without identity overlap, composed of FFHQ [23] and SFHQ [4]; the other is a cross-age dataset with the same identity across different ages, derived from VGGFace2-HQ [9], based on which we establish a crossage identity pool. For more details on dataset construction, please refer to the Appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 181, + 506, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 506, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 506, + 193 + ], + "type": "text", + "content": "Training details. Please refer to the Appendix." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 194, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 396 + ], + "type": "text", + "content": "Baseline. To the best of our knowledge, there has been no method that directly utilizes age information as image prompts for face generation. Therefore, a direct comparison with existing personalized face aging methods is not feasible. To provide a reasonable baseline, we have constructed several baseline methods based on current common used face generation methods, including DreamBooth [36], FaceID [44], and FaceID+DreamBooth. Although these methods are not specifically designed for face aging, with appropriate configuration and adjustments, they can achieve a certain degree of control over face aging as image prompts. Detailed experimental configurations and parameter settings can be found in the appendix. Additionally, we selected several state-of-the-art methods [6, 8, 14] from the past in the field of face aging for comparison. These methods are based on image translation methods, which, while differing from our task, share a similar objective." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 397, + 554, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 554, + 517 + ], + "type": "text", + "content": "Evaluation. To evaluate personalized face aging performance, we employ seven quantitative metrics: Face Similarity (Face Sim.) for ID preservation, Kernel Inception Distance (KID) for distribution fidelity, Age Mean Absolute Error (AgeMAE) for age accuracy, and Relative AgeMAE (R-AgeMAE), our proposed metric, which specifically measures Age Prompt transfer accuracy. Additionally, we assess Face Quality, per-image inference time, and conduct a multidimensional user study for subjective perception. More details are provided in the Appendix." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 517, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 615 + ], + "type": "text", + "content": "Implementation of Age Prompt. Unless otherwise specified, the Age Prompts used are all sampled from FFHQ [23]. The sampling procedure follows three main rules: (1) The images are selected from the target age group, (2) the images must match the race and gender of the source face, and (3) dozens of target age images are sampled to ensure the use of prior knowledge from the population, which helps stabilize the results and ensure consistency." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 624, + 426, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 426, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 426, + 635 + ], + "type": "text", + "content": "4.2. Qualitative Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 714 + ], + "type": "text", + "content": "Comparison with translation-based methods. Time-Booth generates more diverse results compared to comparison methods, as illustrated in Fig. 6. Although translation-based approaches may appear to offer higher fidelity, the results generated by the comparison methods resemble 'children with elderly textures,' creating a strong sense of incon" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12152" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 541, + 225 + ], + "blocks": [ + { + "bbox": [ + 71, + 70, + 541, + 225 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 541, + 225 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 541, + 225 + ], + "type": "image", + "image_path": "6c1d8cffc2e0a93ae1cb86c3e363a29ac687ead2b65841f28b0727e83cd539fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 236, + 549, + 247 + ], + "lines": [ + { + "bbox": [ + 59, + 236, + 549, + 247 + ], + "spans": [ + { + "bbox": [ + 59, + 236, + 549, + 247 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of lifespan synthesis. TimeBooth generates more natural images with high fidelity and consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 71, + 257, + 541, + 365 + ], + "blocks": [ + { + "bbox": [ + 71, + 257, + 541, + 365 + ], + "lines": [ + { + "bbox": [ + 71, + 257, + 541, + 365 + ], + "spans": [ + { + "bbox": [ + 71, + 257, + 541, + 365 + ], + "type": "image", + "image_path": "0888239f85217be4045950af35edaa5c79b609a488f364b95808231ac0f5249f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 373, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 555, + 396 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison. We selected in-the-wild celebrity photos and transformed them to the corresponding age in the GroundTruth. By transferring ID age-invariant features and leveraging the prior knowledge of Age Prompts, our results are superior." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 417, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 417, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 54, + 417, + 295, + 548 + ], + "type": "text", + "content": "gruity, and do not resemble real human faces. In reality, both facial textures and bone structure undergo significant transformations from childhood to old age. By disentangling features and using Age Prompts as guidance, Time-Booth produces faces with semantics more aligned with the real aging process. Fig. 7 demonstrates that Time-Booth achieves a comparable level of cross-age consistency to the comparison methods, indicating accurate preservation of ID-invariant age features. Additionally, as shown in Fig. 9, Time-Booth enables more flexible and personalized generation, which translation-based methods cannot achieve." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 550, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 295, + 647 + ], + "type": "text", + "content": "Comparison with baseline methods. As shown in Fig. 10, although the faces generated after fine-tuning with DreamBooth reflect the age features in the Age Prompt, their fidelity is relatively low. Furthermore, DreamBooth requires fine-tuning, which incurs higher computational and time costs. In contrast, our method is tuning-free, and the generated images can be produced in just a few seconds. For more qualitative experiments, please refer to the Appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 658, + 174, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 174, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 174, + 671 + ], + "type": "text", + "content": "4.3.Quantitative Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 715 + ], + "type": "text", + "content": "We conducted a quantitative evaluation to verify Time-Booth's effectiveness. In Table 1, while TimeBooth inherently differs from translation-based methods by generating" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 417, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 453 + ], + "type": "text", + "content": "images from scratch, it achieves superior performance on several metrics. Furthermore, it outperforms baseline methods on most metrics, highlighting its advantages." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "text", + "content": "User Study. We conducted a user study to compare our method against baseline approaches, evaluating three key metrics: facial consistency, Age Prompt similarity, and visual quality. As shown in Table 1, our method outperforms the baselines across all three metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 525, + 408, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 408, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 408, + 539 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 544, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 628 + ], + "type": "text", + "content": "Effect of Training Strategy. In Table 2, we present an ablation study on various training strategies. The results indicate that removing the Shuffle Strategy or DTS leads to a significant decrease in R-AgeMAE, suggesting a weakened ability to extract age-related information from the Age Prompt. Additionally, removing PAR or DTS results in a drop in KID and Face Sim., implying a decline in fidelity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Effect of Aging Guidance " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": ". In Eq. 15, we introduce " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " to control the influence of Age Prompt features on the target image. To evaluate its effect, we vary " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " from 0 to 1, and the generated results are shown in Fig. 11. As " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " increases, the facial features of the generated image gradually approach those of the Age Prompt, effectively simulating the aging process. Conversely, decreasing " + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": " makes the generated im" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12153" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 71, + 541, + 233 + ], + "blocks": [ + { + "bbox": [ + 71, + 71, + 541, + 233 + ], + "lines": [ + { + "bbox": [ + 71, + 71, + 541, + 233 + ], + "spans": [ + { + "bbox": [ + 71, + 71, + 541, + 233 + ], + "type": "image", + "image_path": "f4529f45afefc8f603d3413bea29fa2531b4801758b7112bda5c5c8b60c2d9d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 242, + 509, + 255 + ], + "lines": [ + { + "bbox": [ + 100, + 242, + 509, + 255 + ], + "spans": [ + { + "bbox": [ + 100, + 242, + 509, + 255 + ], + "type": "text", + "content": "Figure 9. Qualitative comparison of face aging. TimeBooth demonstrates more flexible generation capabilities." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 264, + 553, + 348 + ], + "blocks": [ + { + "bbox": [ + 57, + 264, + 553, + 348 + ], + "lines": [ + { + "bbox": [ + 57, + 264, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 57, + 264, + 553, + 348 + ], + "type": "table", + "html": "
MethodYoung-to-OldYoung-to-ChildUser StudyFace QualityInference Time
KID(X100) ↓Face-sim ↑AgeMAE ↓R-AgeMAE ↓KID(X100) ↓Face-sim ↑AgeMAE ↓R-AgeMAE ↓Face Sim.(%)Sim. to A.P.(%)Visual Quality(%)
CUSP5.300.7446.08-8.080.8208.99-28.6-32.968.05<10 s
FADING4.800.5946.96-4.290.8687.05-21.4-18.668.173 min
IP2P9.300.49127.92-7.170.78820.63-18.6-12.965.86<10 s
FaceID13.040.63021.5022.5914.020.7267.385.7127.120.045.752.48<10 s
DreamBooth12.960.6745.335.7615.820.7436.845.6510.045.725.762.9510 min
FaceID+DreamBooth9.750.65125.0424.607.730.7289.788.2031.454.358.654.366 min
Ours7.200.7314.523.539.760.7638.557.9778.677.175.872.20<10 s
", + "image_path": "37f492601473b310c32cc0dc2065a7d4dbc9c21a4d5e5e14e958f891bf07c030.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 396, + 294, + 526 + ], + "blocks": [ + { + "bbox": [ + 58, + 396, + 294, + 526 + ], + "lines": [ + { + "bbox": [ + 58, + 396, + 294, + 526 + ], + "spans": [ + { + "bbox": [ + 58, + 396, + 294, + 526 + ], + "type": "image", + "image_path": "2971bf21b6c7a8c4308221de961d16e549685e869af443ad788b0e1d38d43314.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 534, + 284, + 546 + ], + "lines": [ + { + "bbox": [ + 67, + 534, + 284, + 546 + ], + "spans": [ + { + "bbox": [ + 67, + 534, + 284, + 546 + ], + "type": "text", + "content": "Figure 10. Qualitative comparison with baseline methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 86, + 563, + 264, + 618 + ], + "blocks": [ + { + "bbox": [ + 55, + 356, + 555, + 378 + ], + "lines": [ + { + "bbox": [ + 55, + 356, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 555, + 378 + ], + "type": "text", + "content": "Table 1. Quantitative comparison across various aspects. Blue indicates the best result among translation-based methods, bold denotes the best among baseline methods, and bold blue represents the best overall." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 563, + 264, + 618 + ], + "lines": [ + { + "bbox": [ + 86, + 563, + 264, + 618 + ], + "spans": [ + { + "bbox": [ + 86, + 563, + 264, + 618 + ], + "type": "table", + "html": "
MethodKID (×100) ↓Face Sim. ↑R-AgeMAE ↓
w/o Shuffle4.850.7288.62
w/o PAR6.800.7194.11
w/o DTS7.400.7248.37
Full4.250.7334.58
", + "image_path": "d95727e286b84c8263c99d186b9d1150b012f95f3685f34f5c02727792854760.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 627, + 220, + 638 + ], + "lines": [ + { + "bbox": [ + 130, + 627, + 220, + 638 + ], + "spans": [ + { + "bbox": [ + 130, + 627, + 220, + 638 + ], + "type": "text", + "content": "Table 2. Ablation study." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": "age more similar to the source image. Given the editability-fidelity trade-off, selecting an appropriate " + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": " is crucial for achieving a balanced result. Empirical studies suggest that an " + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": " value between 0.7 and 0.9 provides the best balance." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 398, + 547, + 485 + ], + "blocks": [ + { + "bbox": [ + 322, + 398, + 547, + 485 + ], + "lines": [ + { + "bbox": [ + 322, + 398, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 322, + 398, + 547, + 485 + ], + "type": "image", + "image_path": "b5cec05d691f3a5854927beeb96c25d41f2bd3e4857c392d87610f314f2eacd9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 494, + 555, + 526 + ], + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 526 + ], + "type": "text", + "content": "Figure 11. Visualization under different aging guidance " + }, + { + "bbox": [ + 313, + 494, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 494, + 555, + 526 + ], + "type": "text", + "content": ": as " + }, + { + "bbox": [ + 313, + 494, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 494, + 555, + 526 + ], + "type": "text", + "content": " increases, the age features from the Age Prompt are gradually injected into the source face." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 548, + 388, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 548, + 388, + 560 + ], + "spans": [ + { + "bbox": [ + 314, + 548, + 388, + 560 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel disentangled face representation that, for the first time, models age using image prompts, providing richer priors for face aging. Based on this, we introduce TimeBooth, which surpasses existing methods by generating more personalized and diverse aged face images, establishing a new paradigm for face aging. Moreover, TimeBooth supports plug-and-play integration and requires no test-time fine-tuning, offering greater flexibility and higher computational efficiency. In the future, we will further explore invariant face representations and extend their applications to broader and more general face generation tasks." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12154" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 200 + ], + "type": "text", + "content": "This work was funded in part by the National Natural Science Foundation of China grant under number 62222603, in part by the STI2030-Major Projects grant from the Ministry of Science and Technology of the People's Republic of China under number 2021ZD0200700, in part by the Key-Area Research and Development Program of Guangdong Province under number 2023B0303030001, and in part by the Program for Guangdong Introducing Innovative and Entrepreneurial Teams (2019ZT08X214)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 228, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 228, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 294, + 272 + ], + "type": "text", + "content": "[1] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Only a matter of style: age transformation using a style-based regression model. ACM Trans. Graph., 40(4):45:1-45:12, 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 272, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 272, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 61, + 272, + 296, + 327 + ], + "type": "text", + "content": "[2] Grigory Antipov, Moez Baccouche, and Jean-Luc Dugelay. Face aging with conditional generative adversarial networks. In 2017 IEEE International Conference on Image Processing, ICIP 2017, Beijing, China, September 17-20, 2017, pages 2089-2093. IEEE, 2017. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 327, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 296, + 392 + ], + "type": "text", + "content": "[3] Sudipta Banerjee, Govind Mittal, Ameya Joshi, Chinmay Hegde, and Nasir D. Memon. Identity-preserving aging of face images via latent diffusion models. In IEEE International Joint Conference on Biometrics, IJCB 2023, Ljubljana, Slovenia, September 25-28, 2023, pages 1-10. IEEE, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 393, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 294, + 414 + ], + "type": "text", + "content": "[4] David Beniaguev. Synthetic faces high quality (sfhq) dataset, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 294, + 470 + ], + "type": "text", + "content": "[5] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 18392-18402. IEEE, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 294, + 525 + ], + "type": "text", + "content": "[6] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18392-18402, 2023. 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 294, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 294, + 570 + ], + "type": "text", + "content": "[7] Xiangyi Chen and Stephane Lathuilière. Face aging via diffusion-based editing. In 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023, page 595. BMVA Press, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 571, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 294, + 602 + ], + "type": "text", + "content": "[8] Xiangyi Chen and Stéphane Lathuilière. Face aging via diffusion-based editing. arXiv preprint arXiv:2309.11321, 2023. 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 602, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 294, + 657 + ], + "type": "text", + "content": "[9] Xuanhong Chen, Bingbing Ni, Yutian Liu, Naiyuan Liu, Zhilin Zeng, and Hang Wang. Simswap++: Towards faster and high-quality identity swapping. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(1):576-592, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 714 + ], + "type": "text", + "content": "[10] Jun Dan, Yang Liu, Haoyu Xie, Jiankang Deng, Haoran Xie, Xuansong Xie, and Baigui Sun. Transface: Calibrating transformer training for face recognition from a data-centric perspective. In Proceedings of the IEEE/CVF international conference on computer vision, pages 20642-20653, 2023. 3, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[11] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 128, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 128, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 128, + 553, + 183 + ], + "type": "text", + "content": "[12] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1-35, 2016. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 184, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 239 + ], + "type": "text", + "content": "[13] Guillermo Gomez-Trenado, Stéphane Lathuilière, Pablo Mesejo, and Oscar Cordón. Custom structure preservation in face aging. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XVI, pages 565-580. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 282 + ], + "type": "text", + "content": "[14] Guillermo Gomez-Trenado, Stéphane Lathuilière, Pablo Mesejo, and Oscar Cordón. Custom structure preservation in face aging. In European Conference on Computer Vision, pages 565-580. Springer, 2022. 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 327 + ], + "type": "text", + "content": "[15] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Commun. ACM, 63(11):139–144, 2020. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "text", + "content": "[16] Bahman Guyuron, David J Rowe, Adam Bryce Weinfeld, Yashar Eshraghi, Amir Fathi, and Seree Iamphongsai. Factors contributing to the facial aging of identical twins. Plastic and reconstructive surgery, 123(4):1321-1331, 2009. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "text", + "content": "[17] John A Hartigan and Manchek A Wong. Algorithm as 136: A k-means clustering algorithm. Journal of the royal statistical society. series c (applied statistics), 28(1):100-108, 1979. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "text", + "content": "[18] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross-attention control. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 514 + ], + "type": "text", + "content": "[19] Gee-Sern Hsu, Rui-Cang Xie, and Zhi-Ting Chen. Wasserstein divergence GAN with cross-age identity expert and attribute retainer for facial age transformation. IEEE Access, 9:39695-39706, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 514, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 514, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 514, + 553, + 558 + ], + "type": "text", + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 592 + ], + "type": "text", + "content": "[21] Zhizhong Huang, Junping Zhang, and Hongming Shan. When age-invariant face recognition meets face age synthesis: A multi-task learning framework. In CVPR, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 592, + 555, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 555, + 647 + ], + "type": "text", + "content": "[22] Zhizhong Huang, Junping Zhang, and Hongming Shan. When age-invariant face recognition meets face age synthesis: A multi-task learning framework. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7282-7291, 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 715 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12155" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 296, + 106 + ], + "type": "text", + "content": "frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "text", + "content": "[25] Zhen Li, Mingdeng Cao, Xintao Wang, Zhongang Qi, MingMing Cheng, and Ying Shan. Photomaker: Customizing realistic human photos via stacked id embedding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8640-8650, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 296, + 207 + ], + "type": "text", + "content": "[26] Haiqi Liu, C. L. Philip Chen, Xinrong Gong, and Tong Zhang. Robust saliency-aware distillation for few-shot fine-grained visual recognition. IEEE Transactions on Multimedia, 26:7529-7542, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 285 + ], + "type": "text", + "content": "[27] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In International Conference on Machine Learning, ICML 2022, 17-23 July 2022, Baltimore, Maryland, USA, pages 16784-16804. PMLR, 2022. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 296, + 342 + ], + "type": "text", + "content": "[28] Roy Or-El, Soumyadip Sengupta, Ohad Fried, Eli Shechtman, and Ira Kemelmacher-Shlizerman. Lifespan age transformation synthesis. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI, pages 739-755. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 296, + 408 + ], + "type": "text", + "content": "[29] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 410, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 296, + 453 + ], + "type": "text", + "content": "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 456, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 499 + ], + "type": "text", + "content": "[31] Helle Rexbye, Inge Petersen, Mette Johansens, Louise Klitkou, Bernard Jeune, and Kaare Christensen. Influence of environmental factors on facial ageing. Age and Ageing, 35 (2):110-115, 2006. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 296, + 567 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 612 + ], + "type": "text", + "content": "[33] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Dex: Deep expectation of apparent age from a single image. In Proceedings of the IEEE international conference on computer vision workshops, pages 10-15, 2015. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 296, + 657 + ], + "type": "text", + "content": "[34] Rasmus Rothe, Radu Timofte, and Luc Van Gool. Deep expectation of real and apparent age from a single image without facial landmarks. International Journal of Computer Vision, 126(2):144-157, 2018. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 713 + ], + "type": "text", + "content": "[35] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "text", + "content": "Canada, June 17-24, 2023, pages 22500-22510. IEEE, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "text", + "content": "[36] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 161, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 161, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 553, + 270 + ], + "type": "text", + "content": "[37] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022, New Orleans, LA, USA, November 28 - December 9, 2022, 2022. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 271, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 553, + 304 + ], + "type": "text", + "content": "[38] Kestutis Sveikata, Irena Balciuniene, Janina Tutkuviene, et al. Factors influencing face aging. literature review. Stomatologija, 13(4):113-116, 2011. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "type": "text", + "content": "[39] Hao Wang, Yitong Wang, Zheng Zhou, Xing Ji, Dihong Gong, Jingchao Zhou, Zhifeng Li, and Wei Liu. Cosface: Large margin cosine loss for deep face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5265-5274, 2018. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "type": "text", + "content": "[40] Qixun Wang, Xu Bai, Haofan Wang, Zekui Qin, Anthony Chen, Huaxia Li, Xu Tang, and Yao Hu. Instantid: Zero-shot identity-preserving generation in seconds. arXiv preprint arXiv:2401.07519, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 404, + 553, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 479 + ], + "type": "text", + "content": "[41] Zongwei Wang, Xu Tang, Weixin Luo, and Shenghua Gao. Face aging with identity-preserved conditional generative adversarial networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 7939-7947. Computer Vision Foundation / IEEE Computer Society, 2018. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 481, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 553, + 525 + ], + "type": "text", + "content": "[42] Hongyu Yang, Di Huang, Yunhong Wang, and Anil K. Jain. Learning continuous face age progression: A pyramid of gans. IEEE Trans. Pattern Anal. Mach. Intell., 43(2):499-515, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 525, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 580 + ], + "type": "text", + "content": "[43] Xu Yao, Gilles Puy, Alasdair Newson, Yann Gousseau, and Pierre Hellier. High resolution face age editing. In 25th International Conference on Pattern Recognition, ICPR 2020, Virtual Event / Milan, Italy, January 10-15, 2021, pages 8624-8631. IEEE, 2020. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "type": "text", + "content": "[44] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "type": "text", + "content": "[45] Cheng Yu, Haoyu Xie, Lei Shang, Yang Liu, Jun Dan, Liefeng Bo, and Baigui Sun. Facechain-fact: Face adapter with decoupled training for identity-preserved personalization. arXiv preprint arXiv:2410.12312, 2024. 3, 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[46] Zongyan Zhang, Haohan Weng, Tong Zhang, and C. L. Philip Chen. A broad generative network for two-stage image outpainting. IEEE Transactions on Neural Networks and Learning Systems, 35(9):12731-12745, 2024. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12156" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 298, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 298, + 119 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 298, + 119 + ], + "type": "text", + "content": "[47] Zongyan Zhang, C. L. Philip Chen, Haohan Weng, and Tong Zhang. Self-prompt guided image outpainting model for captions absence in social scenes. IEEE Transactions on Computational Social Systems, pages 1-12, 2025. 1" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12157" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/a4a233d0-73c0-4327-bc43-78ef9be737e9_content_list.json b/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/a4a233d0-73c0-4327-bc43-78ef9be737e9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1ea4e87b0c8e61b23a2550cc54edcec542e70ebb --- /dev/null +++ b/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/a4a233d0-73c0-4327-bc43-78ef9be737e9_content_list.json @@ -0,0 +1,1511 @@ +[ + { + "type": "text", + "text": "TimeExpert: An Expert-Guided Video LLM for Video Temporal Grounding", + "text_level": 1, + "bbox": [ + 114, + 130, + 883, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zuhao Yang $^{1*}$ Yingchen Yu $^{2}$ Yunqing Zhao $^{2}$ Shijian Lu $^{1\\dagger}$ Song Bai $^{2}$", + "bbox": [ + 205, + 179, + 790, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Nanyang Technological University $^{2}$ ByteDance Inc.", + "bbox": [ + 272, + 205, + 723, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://mwxely.github.io/projects/yang2025time/index", + "bbox": [ + 281, + 231, + 709, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b0c95e601c0873448604778cced1d04c13252a9987dc544923b584b875bf68ad.jpg", + "image_caption": [ + "Figure 1. Left: Video Temporal Grounding (VTG) is a fine-grained video understanding task that aims to accurately locate content along with event timestamps based on natural language queries. In this work, we mainly consider three major types of VTG tasks: (1) Moment Retrieval (MR), (2) Video Highlight Detection (VHD), and (3) Dense Video Captioning (DVC). The outputs of VTG often contain textual captions, timestamps, and saliency scores. Right: Unlike existing methods (e.g., TimeChat [40]) that employ a single static model, motivated by expert specialization on different task tokens [16], we propose TimeExpert, an expert-guided Video LLM with dynamic token routing. Through task-aware expert allocation, TimeExpert demonstrates substantial improvements over state-of-the-art Video-LLMs on several VTG benchmarks. For example, here we visualize zero-shot F1 score for DVC on the YouCook2 dataset [64], $\\mathrm{R} @ 1_{\\mathrm{IoU}=0.7}$ for MR on the Charades-STA dataset [13], and HIT@1 for VHD on the QVHighlights dataset [24]. More results and analysis are in Section 4." + ], + "image_footnote": [], + "bbox": [ + 94, + 267, + 694, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/65e4838df396854c9b320b09a462517a187b3c3edd34b3b12e31370be77de8f5.jpg", + "image_caption": [ + "Performance on VTG Tasks" + ], + "image_footnote": [ + "MR: Moment Retrieval \nDVC: Dense Video Captioning \nVHD: VideoHighlight Detection" + ], + "bbox": [ + 707, + 311, + 906, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 638, + 326, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video Temporal Grounding (VTG) aims to precisely identify video event segments in response to textual queries. The outputs of VTG tasks manifest as sequences of events, each defined by precise timestamps, saliency scores, and textual descriptions. Despite recent advances, a fundamental limitation persists in existing Video Large Language Models (Video-LLMs): they process all task tokens through identical and static pathways, failing to recognize that temporal localization, saliency assessment, and textual generation represent fundamentally distinct tasks requiring specialized processing. To address this, we introduce TimeExpert, a Mixture-of-Experts (MoE)-based Video-LLM that effectively decomposes VTG tasks by dynamically routing task", + "bbox": [ + 89, + 669, + 483, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "specific tokens (e.g., timestamps, saliency scores) to specialized experts, with increased computational efficiency. Our design choices enable precise handling of each subtask, leading to improved event modeling across diverse VTG applications. Extensive experiments demonstrate that TimeExpert consistently achieves state-of-the-art performance on various VTG tasks such as Dense Video Captioning, Moment Retrieval, and Video Highlight Detection.", + "bbox": [ + 511, + 640, + 906, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 768, + 645, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consider watching a cooking tutorial video and attempting to pinpoint the exact moment when the chef fries the bacon (Figure 1-Left) or other highlighted visual segments. This seemingly simple task reveals a fundamental challenge: while humans can effortlessly recognize these actions visually, determining their precise temporal boundaries—down to the specific second—remains significantly challenging", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done while Zuhao Yang was interning at ByteDance.", + "bbox": [ + 107, + 875, + 473, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Shijian Lu is the corresponding author.", + "bbox": [ + 109, + 887, + 321, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24286", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/006e0b32fae927e7b0f24cd9df7232c869e4204106b850a7fd37a768b032e965.jpg", + "image_caption": [ + "(a) VTG-Specific Video-LLM" + ], + "image_footnote": [], + "bbox": [ + 94, + 97, + 321, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8cdc35db096fdd413da3c0fe6607220a59d1f82790d345aaa9c47fb3ce2f4bfe.jpg", + "image_caption": [ + "(b) Vanilla MoE" + ], + "image_footnote": [], + "bbox": [ + 375, + 95, + 588, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9878de8d107a84d2898c11809cabda9ccc5d5f27da7d645a0aa179b8e3ae5d4d.jpg", + "image_caption": [ + "removal new expert", + "(c) TimeExpert (Ours)" + ], + "image_footnote": [], + "bbox": [ + 619, + 95, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3c281f57eeeacc28a909a3388afc5e9152bb1f32410c4ef16693874e9b2bb086.jpg", + "image_caption": [ + "Figure 2. A Glimpse of Comparison across VTG Approaches. (a): VTG-specific Video-LLM [16, 40] relies on a single static model with shared parameters for all tasks, limiting their ability to specialize in diverse VTG subtasks. (b): Vanilla MoE improves upon this by activating a fixed set (e.g., $k = 2$ ) of experts, enabling some degree of task specialization. (c): Our TimeExpert takes a step further by implementing an adaptive routing mechanism that dynamically allocates new experts when tokens lack suitable matches and prunes unmatched experts when necessary. This dynamic approach significantly enhances computational efficiency while achieving superior task specialization, particularly for VTG subtasks that require distinct feature representations. Detailed design choices can be found at Section 3." + ], + "image_footnote": [], + "bbox": [ + 620, + 212, + 661, + 246 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "even for human observers. This problem forms the core motivation behind Video Temporal Grounding (VTG), which aims to develop intelligent systems capable of accurately locating relevant video segments based on natural language queries, enabling users to efficiently access desired content without manually searching through hours of footage.", + "bbox": [ + 88, + 368, + 480, + 459 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Research Gaps. While traditional video grounding models excel in individual VTG subtasks like Moment Retrieval [18, 24] and Video Highlight Detection [32], their practicality is limited as they cannot handle multiple VTG subtasks concurrently and require extensive task-specific fine-tuning to achieve adequate performance. More recently, Video Large Language Models (VideoLLMs) [5, 27, 29, 46, 55] have made remarkable advancements in coarse-grained video content understanding (e.g., VideoQA [11, 27]), yet they struggle to extend such capabilities to tasks requiring fine-grained temporal localization (e.g., VTG [13, 64]), due to the lack of dedicated mechanisms for explicitly modeling temporal boundaries.", + "bbox": [ + 88, + 460, + 482, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to generalist Video-LLMs, VTG-specific models that encode temporal tokens [15, 20, 45] or are trained on temporally-annotated instruction-tuning datasets [19, 40, 49] demonstrate marginal improvements, as they blend temporal information with text tokens without explicitly modeling timestamps and saliency scores, despite videos inherently possessing structured information beyond textual descriptions [16]. Furthermore, existing VTG-specific Video-LLMs rely exclusively on a single LLM as their reasoning backbone, processing all task tokens indiscriminately with shared parameters. This parameter-sharing paradigm inherently constrains the model's ability (Figure 1-Right) to specialize in the diverse subtasks of VTG—such as timestamp prediction, saliency estimation, and caption generation—each of which may demand distinct feature representations for optimal performance.", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Proposed Method. To address the misalignment between language modeling of existing Video-LLMs and inherent video temporal structure, we introduce TimeExpert, an event-centric modeling framework that decouples the prediction of timestamps, saliency scores, and textual descriptions into distinct specialized tasks, organized in an interleaved sequence. To precisely model the varying token importance across these decoupled tasks, we replace the single LLM backbone of existing Video-LLMs with a fine-grained Mixture-of-Experts (MoE) decoder, with our proposed token-aware dynamic gating and token-adaptive routing strategies, which achieves both computational efficiency and modeling precision by activating only task-relevant parameters during both training and inference. To harness each expert's potential, we develop a task-dependent auxiliary loss that encourages more frequently activated experts to process a greater proportion of task-relevant tokens.", + "bbox": [ + 511, + 368, + 903, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Contributions. (1): To our knowledge, TimeExpert is the first to explore the task token importance in VTG by introducing a MoE-based framework with dynamic expert routing. Our expert-guided architecture allows for adaptive learning across different VTG subtasks, effectively mitigating task interference while leveraging expert specialization. (2): We conduct extensive experiments across multiple VTG tasks on several benchmarks to demonstrate the effectiveness of TimeExpert. In particular, we demonstrate substantial improvements over state-of-the-art Video-LLMs through quantitative results and ablation studies, highlighting the advantages of expert-driven token specialization.", + "bbox": [ + 511, + 627, + 906, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 827, + 653, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Temporal Grounding. Video Temporal Grounding (VTG) aims to precisely localize relevant temporal boundaries in a video given textual queries [30]. It can be divided", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24287", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7530c8befb15395ae0e9f0a1e5586f316e0bc05cd32809cade8a001271f20f96.jpg", + "image_caption": [ + "Figure 3. Architecture Overview of TimeExpert. Our model leverages independent encoders and decoding heads to process time, score, and text inputs and outputs. The timestamps and saliency scores of sampled frames are encoded into special tokens and integrated into the corresponding visual tokens. During inference, the generated response follows a structured format, sequentially incorporating time tokens, score tokens, and text tokens." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "into subtasks like Moment Retrieval [13, 38], Dense Video Captioning [2, 43, 64], Video Summarization [17, 42], and Video Highlight Detection [24, 32]. Traditional methods primarily rely on large-scale video-text pretraining, using training objectives such as video-text contrastive learning [48, 54], video-text matching [4, 26], and timestamp regression [35, 36, 59, 60]. Despite their initial success, these methods are constrained by their generalization ability to handle only a single task per model and require substantial fine-tuning for each new downstream application.", + "bbox": [ + 88, + 415, + 482, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video-LLMs for VTG Tasks. With the growing reasoning capabilities of Video-LLMs [5, 27, 29, 46, 55], numerous studies have investigated their adaptation to VTG tasks (Figure 2-(a)). For example, TimeChat [40], VTimeLLM [19], and HawkEye [20] fine-tune the Video-LLMs on refined temporal-annotated instruction datasets. Following that, several studies further enhance the Video-LLM's understanding of timestamps by introducing special tokens [15, 20] or a temporal perception module [39]. However, these methods adopt a purely text-to-text framework, formulating VTG as a natural language generation problem, which lacks explicit temporal modeling. Recently, TRACE [16] proposes a causal event modeling framework to capture the inherent structure of videos. Nevertheless, it still fails to incorporate task-specific properties, as they process all task tokens within a single shared reasoning backbone, leading to task interference and limited specialization. In contrast, TimeExpert simultaneously structures Video-LLM outputs into event representations and leverages adaptive expert specialization, enhancing both efficiency and accuracy across diverse VTG tasks.", + "bbox": [ + 89, + 566, + 482, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLMs with Mixture-of-Experts. The Mixture-of-Experts", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(MoE) [8, 28, 41] architecture (Figure 2-(b)) has been widely adopted to improve the efficiency of large-scale foundation models during both training and inference by replacing feed-forward network layers in transformer models with multiple specialized sub-networks, known as experts, which enables sparse parameter activation. The activated experts are selected by a gating network, which contain a pre-defined router. Prior methods often implement static top- $k$ routing across all tokens, several studies [14, 21, 58, 61] explored on dynamic routing mechanism that allow tokens to activate varying numbers of experts. However, these approaches either introduce additional hyperparameter complexity or demonstrate limited validation confined to text-to-text tasks, constraining their flexibility and cross-task generalizability. In contrast, our proposed TimeExpert (Figure 2-(c)) introduces dynamic expert routing tailored for Video-LLMs through adaptively adjusting the number of activated experts based on task token importance, ensuring effective expert specialization while eliminating reliance on rigid hyperparameter tuning.", + "bbox": [ + 511, + 90, + 903, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. TimeExpert: An Expert-guided Video LLM", + "text_level": 1, + "bbox": [ + 511, + 405, + 903, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we firstly introduce some fundamental preliminaries in Section 3.1. Then, we propose TimeExpert (Figure 3), which integrates a dynamic expert selection mechanism with a structured event modeling framework. Our approach implements task-aware dynamic gating (Section 3.2) and token-adaptive routing (Section 3.3) to selectively activates specialized experts based on token relevance across diverse tasks. In Section 3.4, we discuss our proposed task-dependent auxiliary loss that optimizes expert utilization and thereby enhances the stability of MoE-based learning. In Section 3.5, we elaborate on our training recipe.", + "bbox": [ + 511, + 431, + 906, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 604, + 653, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Definition. Given a textual instruction $\\mathcal{I}$ and a sequence of video frames $\\mathcal{F}$ from a video, the objective of VTG task is to generate a structured event representation $\\mathcal{R}$ . The output sequence $\\mathcal{R}$ consists of discrete events $\\{\\epsilon_1,\\epsilon_2,\\dots ,\\epsilon_M\\}$ , where each event $\\epsilon_{m}$ encapsulates three key components: a timestamp $t_m$ , a saliency score $s_m$ , and a textual caption $c_{m}$ . Formally, we have:", + "bbox": [ + 511, + 626, + 903, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} = \\left\\{\\epsilon_ {1}, \\epsilon_ {2}, \\dots , \\epsilon_ {M} \\right\\} = \\left\\{\\left(t _ {m}, s _ {m}, c _ {m}\\right) \\mid 1 \\leq m \\leq M \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 732, + 903, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Causal Event Modeling. Following [16], we formulate the event prediction process in a causal manner, leveraging the sequential dependencies between events. Specifically, given the past event sequence $\\epsilon_{1:m-1}$ , the textual instruction $\\mathcal{I}$ , and the video frames $\\mathcal{F}$ , the probability distribution of the next event $\\epsilon_m$ is defined as:", + "bbox": [ + 511, + 753, + 906, + 839 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} \\left(\\epsilon_ {m} \\mid \\epsilon_ {1: m - 1}, \\mathcal {I}, \\mathcal {F}\\right) = \\mathcal {P} \\left(t _ {m}, s _ {m}, c _ {m} \\mid \\epsilon_ {1: m - 1}, \\mathcal {I}, \\mathcal {F}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 845, + 903, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This causal formulation ensures that event generation follows a structured temporal order, where each event is pro", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24288", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "gressively determined based on prior events, textual cues, and visual content.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Task-aware Dynamic Gating", + "text_level": 1, + "bbox": [ + 89, + 128, + 349, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vanilla MoEs. (Figure 2-(b)) employ the top- $k$ gating strategy, which activates a fixed number of experts for each input token, determined by computing gating scores via a learned network $g$ . Specifically, for a given token embedding $\\mathbf{x} \\in \\mathbb{R}^d$ , the gating function follows:", + "bbox": [ + 89, + 151, + 483, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng (\\mathbf {x}) \\in \\mathbb {R} ^ {K} := \\operatorname {s o f t m a x} \\left(\\mathbf {W} _ {g} ^ {T} \\mathbf {x}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 234, + 482, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_g\\in \\mathbb{R}^{d\\times K}$ represents the gating network parameters, and $K$ denotes the total number of experts. Then, the output of this MoE layer is then formulated as:", + "bbox": [ + 89, + 263, + 483, + 310 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\frac {1}{\\sum_ {e \\in \\operatorname {T o p} - k (g (\\mathbf {x}))} g (\\mathbf {x}) _ {e}} \\sum_ {e \\in \\operatorname {T o p} - k (g (\\mathbf {x}))} g (\\mathbf {x}) _ {e} \\mathbf {E} _ {e} (\\mathbf {x}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 316, + 482, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{E}_e(\\mathbf{x})\\in \\mathbb{R}^d$ is the output of the $e$ -th expert given input $\\mathbf{x}$ , and $g(\\mathbf{x})_e$ is the corresponding gating score.", + "bbox": [ + 89, + 364, + 482, + 396 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implicit Task Preference. Vanilla MoEs face some limitations, such as (1): the reliance on a predefined $k$ for expert selection leads to suboptimal flexibility [6, 9, 56], (as task complexity varies, different tokens may require a varying number of experts based on their importance); and (2): the standard gating function treats all tokens uniformly, neglecting the inherent differences in task-specific significance, which may lead to inefficient expert routing and task interference [21]. Surprisingly, we observe an implicit task preference in these models on VTG tasks, as visualized in Figure 4 of expert assignments. In particular, we measure the ratio of activated task-specific tokens to the number of processed text tokens in that layer. Our findings reveal that certain experts, such as expert #7, consistently activate in response to specific subtasks (e.g., score tokens), even though the model was not explicitly trained to establish such specialization, implying an inherent task-driven preference.", + "bbox": [ + 89, + 396, + 483, + 652 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/650963354869e4cee6fad7bb905e1f85f62ae301d731dd75b16a279bd9a5fc28.jpg", + "image_caption": [ + "Figure 4. Visualization of Expert Assignments on Various Task Tokens using Our Vanilla MoE Implementation. We take layer 4 as an example and only visualized the first 8 experts out of a total of 64, due to the space limitation." + ], + "image_footnote": [], + "bbox": [ + 94, + 664, + 480, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Explicit Task-Aware Gating. Based on this observation, we propose a task-aware dynamic gating mechanism that", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "jointly considers token-level probabilities and task-specific activation statistics. Instead of applying a static threshold, we introduce a task-weighted gating function that adjusts expert activation dynamically by incorporating task token activation rates. Formally, given an expert representation matrix $\\mathbf{W}_g\\in \\mathbb{R}^{d\\times K}$ and an input token $\\mathbf{x}\\in \\mathbb{R}^d$ , we define:", + "bbox": [ + 511, + 90, + 905, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns (\\mathbf {x}) = \\cos (\\mathbf {x}, \\mathbf {W} _ {g}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 196, + 906, + 213 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng (\\mathbf {x}) = \\operatorname {s i g n} \\left(\\sigma \\left(\\frac {s (\\mathbf {x}) + \\alpha A _ {t}}{1 + \\alpha}\\right) - \\sigma (\\mathbf {G})\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 218, + 906, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A_{t}$ represents the historical activation rate of the task token type associated with $\\mathbf{x}$ ; $\\alpha$ is a scaling coefficient that controls the influence of task importance; $\\sigma(\\cdot)$ denotes the sigmoid function for score normalization; $\\mathbf{G} \\in \\mathbb{R}^{K}$ is a learnable threshold where a token routes to expert $e$ only if its similarity score $s(\\mathbf{x}) \\in \\mathbb{R}^{K}$ exceeds $\\mathbf{G}$ . Following [14], we make the binary sign() differentiable by copying the upstream gradient of $g(\\mathbf{x})$ to the pre-activation term $\\sigma(s(\\mathbf{x}, A_{t})) - \\sigma(\\mathbf{G})$ . With $A_{t}$ , the model explicitly prioritizes experts that have demonstrated effectiveness for specific task tokens, encouraging specialization while ensuring under-utilized experts do not get excessively activated.", + "bbox": [ + 511, + 257, + 906, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Token-adaptive Routing", + "text_level": 1, + "bbox": [ + 511, + 448, + 736, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce a token-adaptive routing mechanism (Figure 2-(c)) that dynamically facilitates expert addition and removal based on recorded activation trends. The routing mechanism comprises three key components, including (1): Task-specific Routing Recording: We continuously monitor expert utilization across different task tokens and the routing results during training. (2): Adaptive Expert Addition: New experts are allocated when certain task tokens persistently fail to activate appropriate experts, and (3): Redundant Expert Removal: Experts with persistently low activation rates are pruned to maintain efficiency.", + "bbox": [ + 511, + 469, + 905, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-specific Routing Recording. Instead of solely tracking which experts get activated, we maintain a per-task token activation frequency metric. Specifically, (1) For each expert $e$ , we log its activation timestamps, denoted as $\\mathbf{R}_E \\in \\mathbb{R}^K$ ; (2) For input tokens that fail to activate any expert, we compute their aggregated embeddings as $\\mathbf{R}_S \\in \\mathbb{R}^d$ ; (3) We maintain an activation rate tracker $A_t$ per task token type, which influences expert selection in subsequent iterations.", + "bbox": [ + 511, + 636, + 905, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive Expert Addition. When a significant proportion of task tokens fail to activate any expert, we introduce new experts whose representations are initialized as:", + "bbox": [ + 511, + 757, + 905, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {g, K + 1} = \\frac {\\mathbf {R} _ {S}}{\\| \\mathbf {R} _ {S} \\|}, \\quad \\mathbf {G} _ {K + 1} = 0, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 811, + 906, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{R}_S$ captures the average embedding of underrepresented task tokens. This ensures that newly added experts are aligned with the missing task representations.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24289", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e8775313169e8bea0b1f58e66ac62ee987872d4949be55f1861da735b6460534.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Training StageDatasetsNo. of Samples
Stage 1: Task Module PretrainingValley [33], LLaVA-Image [31], TextVR [51], ShareGPT4Video [3], VTG-IT [15]1.9M
Stage 2: MoE Decoder PretrainingValley*, TextVR*, ShareGPT4Video*, VTG-IT*, ActivityNet Captions [2], VideoChatGPT [34], InternVid [48], Next-QA [53]0.9M
Stage 3: Supervised Fine-tuningFiltered and Re-annotated Data from: Previous Stages' Data, EgoQA [37], STAR [50], Moment-10M [39], LLaVA-Video-178K [63]2.3M
", + "bbox": [ + 176, + 88, + 823, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Training Data Recipe of TimeExpert. These data are categorized into three training stages. * denotes we utilize the filtered subset of the original dataset.", + "bbox": [ + 89, + 253, + 906, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Redundant Expert Removal. Experts are pruned if their activation rate $A_{e}$ remains below a threshold $\\tau_{\\mathrm{min}}$ :", + "bbox": [ + 89, + 309, + 483, + 339 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\text {r e m o v e}} = \\left\\{e \\mid A _ {e} < \\tau_ {\\min } \\right\\}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 345, + 482, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This prevents redundant experts from occupying parameter space while ensuring that experts specializing in low-frequency but important task tokens are retained.", + "bbox": [ + 89, + 367, + 483, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Task-dependent Auxiliary Loss", + "text_level": 1, + "bbox": [ + 89, + 419, + 366, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance the efficiency and specialization of expert activation in TimeExpert, we introduce a task-dependent auxiliary loss, which ensures that experts with higher activation rates receive more relevant task tokens while preventing excessive expert overlap.", + "bbox": [ + 89, + 441, + 483, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing auxiliary loss functions [10, 52] primarily focus on load balancing among experts, which may conflict with our goal of leveraging expert specialization for VTG tasks. Instead, our auxiliary loss promotes a dynamic allocation strategy by reinforcing the association between frequently activated experts and their corresponding task tokens while preventing under-utilization or excessive redundancy.", + "bbox": [ + 89, + 517, + 483, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our auxiliary loss consists of two key components: task-aware concentration term (Left) and activation regularization term (Right). The task-aware concentration loss ensures that experts specialize in processing specific types of task tokens (e.g., timestamps, saliency scores, captions) by amplifying the assignment of relevant tokens to frequently activated experts. The activation regularization loss prevents the over-activation of a single expert. The two loss terms work synergistically to ensure a diverse and efficient allocation of task tokens. In summary, we define our auxiliary loss as follows:", + "bbox": [ + 89, + 623, + 483, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a u x}} = \\lambda_ {1} \\sum_ {e = 1} ^ {K} \\left(\\frac {A _ {e}}{\\sum_ {j = 1} ^ {K} A _ {j}} - \\frac {N _ {e}}{\\sum_ {j = 1} ^ {K} N _ {j}}\\right) ^ {2} + \\lambda_ {2} \\sum_ {e = 1} ^ {K} \\| \\mathbf {w} _ {g, e} \\| _ {2} ^ {2} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 792, + 482, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A_{e}$ represents the activation count of expert $e$ over a batch, $N_{e}$ denotes the number of task tokens assigned to expert $e$ , $\\mathbf{w}_{g,e}$ is the representation vector of expert $e$ , and $\\lambda_{1}$ and $\\lambda_{2}$ are hyperparameters controlling the balance between task-aware specialization and numerical stability.", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training Recipe", + "text_level": 1, + "bbox": [ + 513, + 308, + 671, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Data. To ensure comprehensive learning across various video-language understanding tasks, we curate a diverse dataset spanning three progressive training stages, as summarized in Table 1. These utilized public datasets cover a wide range of temporal reasoning, video grounding, and captioning challenges. We conducted systematic processing, filtering, and re-annotating techniques that remove inaccurate temporal annotations, ensuring high-quality labels.", + "bbox": [ + 511, + 349, + 906, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Strategy. The training process of TimeExpert includes three stages: (1) Stage 1 - Task Module Pretraining. In the first stage, we pretrain the task-specific modules, including the vision compression layer, task encoder, and task heads, using 1.9M general multimodal video-text samples. This stage ensures that the model learns fundamental video representation capabilities before incorporating MoE decoder. (2) Stage 2 - MoE Decoder Pretraining. Before fine-tuning the full model, we introduce an intermediate MoE pretraining stage with 0.9M samples to ensure that expert routing aligns with different VTG-specific task tokens. This step mitigates the risk of expert collapse (i.e., all tokens being routed to the same expert) and enhances task-aware expert specialization. (3) Stage 3 - Supervised Fine-tuning. In the final stage, we fine-tune the entire Time-Expert framework on 2.3M samples, allowing both the task modules and the MoE decoder to be optimized jointly.", + "bbox": [ + 511, + 489, + 908, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Objectives. During the first stage, we adopt cross-entropy loss for each generated token to teach basic video-event formatting. During MoE Decoder Pretraining, besides the cross-entropy loss, we apply $z$ -loss [65] to stabilize training and proposed task-aware auxiliary loss to mitigate expert collapse and drive specialization. During Supervised Fine-tuning, we fine-tune the full model (except visual encoder) using same trio of objectives, ensuring stable routing while the model learns task-specific instructions.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24290", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/77cfdaac63444ff1214f8fca94fb9085ec8a9c94b3700a148adcae856c18affc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodNo. of \nActivated \nParametersDense Video Captioning \n(YouCook2 [64])Moment Retrieval \n(Charades-STA [13])Video Highlight Detection \n(QVHighlights [24])
SODAc(↑)CIDEr (↑)F1 Score (↑)R@IIoU=0.5 (↑)R@IIoU=0.7 (↑)mAP (↑)HIT@1 (↑)
TimeChat [40]7B1.23.412.632.213.414.523.9
VTimeLLM [19]7B---27.511.4--
Momentor [39]7B---26.611.67.6-
HawkEye [49]7B---31.414.5--
VTG-LLM [15]7B1.55.017.533.815.716.533.5
TRACE [16]7B2.28.122.440.319.426.842.7
TimeExpert (adaptive k)≈ 5.9B / 3.5B / 4.8B2.58.223.642.820.329.646.9
TimeExpert (TRACE's data)≈ 5.2B / 3.1B / 4.0B2.48.123.341.920.129.146.3
", + "bbox": [ + 91, + 88, + 906, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Zero-shot Performance Comparison of TimeExpert against several state-of-the-art VTG-specific Video-LLMs on Dense Video Captioning, Temporal Grounding, and Video Highlight Detection. Some results are sourced from [16]. The best results are in bold. We also underline the second-best results.", + "bbox": [ + 89, + 263, + 906, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 330, + 222, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Details", + "text_level": 1, + "bbox": [ + 89, + 357, + 290, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We adopt ARIA [25] as the base MoE Video-LLM. Specifically, we utilize a lightweight visual encoder with 438M parameters, consisting of a Vision Transformer (ViT) [7] and a projection module. The ViT weights are initialized from [62]. Each video frame is initially encoded into 128 (for medium resolution) or 256 (for high resolution) visual tokens, which are then compressed via slot-based token compression [15] down to 8 visual tokens per frame. The time encoder and score encoder share the same architecture, both initialized with a tokenizer containing 11 number tokens for representing timestamps/scores, a separator token marking the end of each timestamp/score, and a switching token indicating task transitions. Specifically, following [16], we insert the token $\\langle sep\\rangle$ between consecutive timestamps or scores and append $\\langle sync\\rangle$ at the end of the sequence.", + "bbox": [ + 89, + 381, + 482, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison Methods. We select TimeChat [40], VTimeLLM [19], Momentor [39], HawkEye [49], VTGLLM [15], and TRACE [16] as comparison methods.", + "bbox": [ + 89, + 625, + 482, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Benchmarks. To thoroughly evaluate TimeExpert in Video Temporal Grounding, we assess it across three VTG tasks: Dense Video Captioning, Moment Retrieval, and Video Highlight Detection, utilizing datasets such as YouCook2 [64], Charades-STA [13], QVHighlights [24], and ActivityNet Captions [23].", + "bbox": [ + 89, + 671, + 482, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For Dense Video Captioning, we use metrics including $\\mathrm{SODA_c}$ [12], which is specifically tailored for the video's storyline; CIDEr score [44], which measures consensus between generated and reference captions by computing the TF-IDF weighted n-gram similarity across multiple reference captions; F1 Score, which quantifies the model's capacity on precisely locating timestamps; and METEOR [1], which combines uni-gram precision and recall into a harmonic mean while incorporating stemming,", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "synonym matching, and a fragmentation penalty so that longer contiguous matches are rewarded. For Moment Retrieval, we compute the Intersection over Union (IoU) between the timestamps predicted by the model and ground truth, including Recall at IoU thresholds of $\\{0.5, 0.7\\}$ and mean IoU. For Video Highlight Detection, we calculate both the mean Average Precision (mAP) at IoU thresholds of $\\{0.5, 0.75\\}$ , and HIT@1, which represents the hit ratio of the highest scored clip.", + "bbox": [ + 511, + 332, + 906, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 482, + 705, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot Performance on VTG Tasks. In Table 2, we present the zero-shot performance of TimeExpert against several state-of-the-art baselines on three widely adopted VTG benchmarks. We report the number of activated parameters for a transparent comparison. Across the three benchmarks, the adaptive- $k$ variant activated an average of 14.5, 9.1, and 11.8 experts, whereas the TRACE-trained variant averaged 13.5, 8.3, and 10.5. Notably, our approach outperforms other VTG-specific Video-LLMs across all three benchmarks with fewer activated parameters. It achieves a $0.3\\%$ and $1.2\\%$ performance gain on YouCook2 [64] using the $\\mathrm{SODA_c}$ and F1 Score metrics, respectively; a $2.5\\%$ and $0.9\\%$ performance boost in Recall with IoU = {0.5, 0.7} thresholds on Charades-STA [13]; and a $2.8\\%$ and $4.2\\%$ performance gain for the mAP and HIT@1 metrics on QVHighlights [24]. This validates that dynamic gating and adaptive routing not only ensures efficient expert utilization but also fully leverages task specialization.", + "bbox": [ + 511, + 506, + 906, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fine-tuned Performance on VTG Tasks. In Table 3, we present the results of fine-tuning TimeExpert on YouCook2 and Charades-STA datasets for 2 epochs. The results indicate that: (1) TimeExpert surpasses both generalist and task-specific baselines. Unlike TimeChat and VTG-LLM, which struggle to fully benefit from fine-tuning, our MoE-based approach effectively adapts to downstream datasets, achieving new state-of-the-art results. Notably, it surpasses", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24291", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9d19316759034114927ed468b880ae21ae84a0b0d7ddf9b8ab0ba8e34c523ecb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDense Video Captioning\n(YouCook2 [64])Moment Retrieval\n(Charades-STA [13])
SODAc(↑)CIDEr(↑)F1 Score(↑)R@1IoU=0.5(↑)R@1IoU=0.7(↑)
Traditional Models
PDVC [47]4.422.7---
Vid2Seq [57]5.725.323.5--
Vid2Seq (Audio Input) [57]7.947.127.3--
CM² [22]5.331.728.4--
VTG-specific Video-LLMs
TimeChat [40]3.411.019.546.723.7
VTG-LLM [15]3.613.420.657.233.4
TRACE [16]6.735.531.861.741.4
TimeExpert (Ours)7.239.033.564.143.3
", + "bbox": [ + 163, + 88, + 834, + 343 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/69f6a8d62dfd11c2a1587077786169be7087561167b8926a27e274432f2d6b15.jpg", + "table_caption": [ + "Table 3. Fine-tuned Performance of TimeExpert. Some results are from [16]. We report the performance of each model after fine-tuning on YouCook2 and Charades-STA datasets. We bold the best scores and italicize results that benefit from additional audio inputs." + ], + "table_footnote": [], + "table_body": "
MethodDense Video CaptioningMoment Retrieval
METEOR (↑)SODAc (↑)CIDEr (↑)F1 Score (↑)R@1IoU=0.5 (↑)R@1IoU=0.7 (↑)mIoU (↑)
TimeChat [40]5.74.719.036.94.62.06.9
VTimeLLM [19]6.85.827.6-29.514.231.4
Momentor* [39]4.72.314.9-23.012.429.3
VTG-LLM [15]5.95.120.734.88.33.712.0
TRACE [16]6.46.025.939.337.724.039.0
TimeExpert (Ours)7.06.528.440.539.226.141.5
", + "bbox": [ + 101, + 393, + 897, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Performance Comparison on ActivityNet Captions [23]. Some results are sourced from [16]. * indicates zero-shot evaluation. The best results are in bold. TimeExpert achieves state-of-the-art results across both Dense Video Captioning and Moment Retrieval.", + "bbox": [ + 89, + 559, + 906, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TRACE in both Moment Retrieval and Dense Video Captioning tasks, highlighting the benefits of dynamic expert activation in video-language modeling. (2) TimeExpert enhances retrieval and captioning tasks simultaneously. Our model achieves a $+3.5$ improvement in CIDEr and a $+1.7$ boost in F1 Score on YouCook2, alongside an absolute gain of 2.4 in $\\mathrm{R}@\\mathrm{1}_{\\mathrm{IoU}} = 0.5$ and 1.9 in $\\mathrm{R}@\\mathrm{1}_{\\mathrm{IoU}} = 0.7$ on Charades-STA. These results demonstrate that TimeExpert's token-adaptive routing effectively captures event structures, enabling improved video reasoning. (3) Efficient expert utilization enables balanced trade-offs. Unlike static models, TimeExpert dynamically activates the most relevant experts per input, resulting in stronger generalization to unseen datasets while keeping computation manageable.", + "bbox": [ + 88, + 614, + 485, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance Comparison on ActivityNet Captions. In Table 4, we evaluate TimeExpert on the ActivityNet Captions [23] dataset. Notably, TimeExpert achieves the best performance across both Moment Retrieval and Dense", + "bbox": [ + 89, + 839, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Captioning tasks. It surpasses TRACE, which was previously the strongest baseline, by improving CIDER (+2.5) and F1 Score (+1.2) in Dense Video Captioning, while also significantly boosting R@ $1_{\\mathrm{IoU}} = 0.5$ (+1.5) and mIoU (+2.5) in Moment Retrieval. These results highlight TimeExpert's ability to effectively model long-range temporal dependencies while efficiently leveraging a MoE framework for adaptive token routing. Unlike previous methods, which rely on static expert activation, our model dynamically selects experts based on token-wise information, leading to improved alignment between textual queries and visual events.", + "bbox": [ + 511, + 614, + 906, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 815, + 679, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Design Choices. As shown in the upper part of Table 5, the removal of token-adaptive routing leads to a notable drop in all metrics, confirming its efficacy in efficiently allocating task tokens. Moreover, removing separate encoders", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24292", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3058beabc216fcaa7e62519dec58f05123fb06c4c28b6b612bf7bbc3d3b6d157.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDense Video Captioning\n(YouCook2 [64])Moment Retrieval\n(Charades-STA [13])Video Highlight Detection\n(QVHighlights [24])
SODAc(↑)CIDEr (↑)F1 Score (↑)R@1IoU=0.5 (↑)R@1IoU=0.7 (↑)mAP (↑)HIT@1 (↑)
Ablation Studies on Design Choices
w/o token-adaptive routing2.17.622.240.519.226.542.6
w/o separate encoders*-------
w/o task-dependent loss2.47.922.841.319.728.145.2
TimeExpert (Ours)2.58.223.642.820.329.646.9
Ablation Studies on No. of Input Frames
TimeExpert (8 frames)1.55.118.930.213.922.138.4
TimeExpert (64 frames)2.17.521.937.517.827.244.9
TimeExpert (128 frames)2.58.222.041.519.929.346.5
Ablation Studies on No. of Activated Experts
Vanilla MoE (k=2)2.38.122.842.119.829.245.8
Vanilla MoE (k=4)2.48.223.342.520.129.546.6
Vanilla MoE (k=6)2.58.223.542.820.229.646.9
Vanilla MoE (k=8)2.58.123.442.820.229.846.8
", + "bbox": [ + 89, + 88, + 906, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation Study Results of TimeExpert. We evaluate the impact of removing key components, varying the number of input frames, and adjusting the number of activated experts. Results under the default setting are highlighted in gray. * denotes that directly processing time tokens and score tokens with the text tokenizer causes the model to fail to follow instructions.", + "bbox": [ + 89, + 417, + 906, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and decoding heads for causal event modeling completely disrupts the model's ability to follow instructions, reinforcing the necessity of structured task processing, and removing the auxiliary loss results in a slight performance drop across all tasks, particularly in Video Highlight Detection and Moment Retrieval. This indicates our proposed task-dependent auxiliary loss improves expert specialization by forcing task-aware routing, leading to better alignment between task tokens and relevant experts.", + "bbox": [ + 88, + 487, + 480, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of Input Frames. As shown in the middle of Table 5, increasing the number of frames generally improves performance across all tasks, as expected. Notably, the transition from 64 to 128 frames boosts performance in Moment Retrieval and Video Highlight Detection, indicating that a larger temporal context helps refine event understanding.", + "bbox": [ + 89, + 625, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of Activated Experts. As shown in the bottom part of Table 5, there is a clear trend where the predefined number of activated experts directly influences performance. Activating more experts improves performance, particularly in Video Highlight Detection (mAP +0.4 from 2 to 6 experts). Nonetheless, beyond 6 experts, performance saturates, showing that excessive expert activation may introduce redundancy without additional benefits.", + "bbox": [ + 89, + 717, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Data. We also conduct a data variant study, training our model on the same data recipe as TRACE to enable a direct comparison. As shown in the bottom line of Table 2, TimeExpert still outperforms TRACE across all VTG", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tasks. Compared to our full-data model, this variant exhibits slight fluctuations, with some metrics even improving (e.g., CIDEr and $\\mathrm{R} @ 1_{\\mathrm{IoU}=0.7}$ ), suggesting that TimeExpert generalizes well even with limited data. These findings highlight TimeExpert's efficiency and adaptability, proving that our model design inherently enhances VTG performance, independent of dataset scale.", + "bbox": [ + 511, + 487, + 906, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 607, + 633, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we identify two critical limitations in existing VTG-specific Video-LLMs: inappropriate temporal event modeling and uniform task token processing. To address this, we propose TimeExpert, a novel framework that addresses these issues through a Mixture-of-Experts decoder that dynamically routes different task tokens (e.g., timestamps, saliency scores, and textual captions) to specialized experts, ensuring adaptive learning and mitigating task interference. Comprehensive experiments demonstrate Time-Expert's effectiveness across multiple VTG subtasks and general video understanding benchmarks, consistently outperforming existing VTG-specific Video-LLMs.", + "bbox": [ + 511, + 633, + 906, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 513, + 829, + 681, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This study is supported by the Ministry of Education Singapore, under the Tier-1 project scheme with the project number RT18/22.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24293", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6", + "[2] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 3, 5", + "[3] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. Sharegpt4video: Improving video understanding and generation with better captions. In Advances in Neural Information Processing Systems, pages 19472-19495, 2024. 5", + "[4] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. In Advances in Neural Information Processing Systems, pages 72842-72866, 2023. 3", + "[5] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2, 3", + "[6] Aidan Clark, Diego de Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. In International Conference on Machine Learning, pages 4057-4086. PMLR, 2022. 4", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6", + "[8] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. 3", + "[9] Dongyang Fan, Bettina Messmer, and Martin Jaggi. Towards an empirical understanding of moe design choices. arXiv preprint arXiv:2402.13089, 2024. 4", + "[10] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 5", + "[11] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24108-24118, 2025. 2" + ], + "bbox": [ + 93, + 114, + 483, + 897 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Soichiro Fujita, Tsutomu Hirao, Hidetakamigaito, Manabu Okumura, and Masaaki Nagata. Soda: Story oriented dense video captioning evaluation framework. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 517-531. Springer, 2020. 6", + "[13] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 3, 6, 7, 8", + "[14] Yongxin Guo, Zhenglin Cheng, Xiaoying Tang, Zhaopeng Tu, and Tao Lin. Dynamic mixture of experts: An autotuning approach for efficient transformer models. In International Conference on Learning Representations, 2024. 3, 4", + "[15] Yongxin Guo, Jingyu Liu, Mingda Li, Dingxin Cheng, Xiaoying Tang, Dianbo Sui, Qingbin Liu, Xi Chen, and Kevin Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. In Proceedings of the AAAI conference on artificial intelligence, 2025. 2, 3, 5, 6, 7", + "[16] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Qingbin Liu, and Xi Chen. Trace: Temporal grounding video llm via causal event modeling. In International Conference on Learning Representations, 2025. 1, 2, 3, 6, 7", + "[17] Michael Gygli, Helmut Grabner, Hayko Riemenschneider, and Luc Van Gool. Creating summaries from user videos. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13, pages 505-520. Springer, 2014. 3", + "[18] Donghoon Han, Seunghyeon Seo, Eunhwan Park, Seong-Uk Nam, and Nojun Kwak. Unleash the potential of clip for video highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8275-8279, 2024. 2", + "[19] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 2, 3, 6, 7", + "[20] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 2, 3", + "[21] Quzhe Huang, Zhenwei An, Nan Zhuang, Mingxu Tao, Chen Zhang, Yang Jin, Kun Xu, Liwei Chen, Songfang Huang, and Yansong Feng. Harder tasks need more experts: Dynamic routing in moe models. In Association for Computational Linguistics, 2024. 3, 4", + "[22] Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi, and Seong Tae Kim. Do you remember? dense video captioning with cross-modal memory retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894-13904, 2024. 7", + "[23] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-capitaling events in videos. In" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24294", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE international conference on computer vision, pages 706-715, 2017. 6, 7", + "[24] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In Advances in Neural Information Processing Systems, pages 11846-11858, 2021. 1, 2, 3, 6, 8", + "[25] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model. arXiv preprint arXiv:2410.05993, 2024. 6", + "[26] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19948-19960, 2023. 3", + "[27] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2, 3", + "[28] Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. Unimoe: Scaling unified multimodal llms with mixture of experts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3", + "[29] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Empirical Methods in Natural Language Processing, 2024. 2, 3", + "[30] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified videolanguage temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 2", + "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 5", + "[32] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3042-3051, 2022. 2, 3", + "[33] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.5", + "[34] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Association for Computational Linguistics, 2024. 5", + "[35] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration for video temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 3" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23023-23033, 2023. 3", + "[37] Thong Thanh Nguyen, Zhiyuan Hu, Xiaobao Wu, Cong-Duy T Nguyen, See-Kiong Ng, and Anh Tuan Luu. Encoding and controlling global semantics for long-form video question answering. arXiv preprint arXiv:2405.19723, 2024. 5", + "[38] Andreea-Maria Oncescu, Joao F Henriques, Yang Liu, Andrew Zisserman, and Samuel Albanie. Queryd: A video dataset with high-quality text and audio narrations. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2265-2269. IEEE, 2021. 3", + "[39] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In International Conference on Machine Learning, 2024. 3, 5, 6, 7", + "[40] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1, 2, 3, 6, 7", + "[41] N Shazeer, A Mirhoseini, K Maziarz, A Davis, Q Le, G Hinton, and J Dean. The sparsely-gated mixture-of-experts layer. Outrageously large neural networks, 2017. 3", + "[42] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. Tvsum: Summarizing web videos using titles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5179-5187, 2015. 3", + "[43] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 3", + "[44] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6", + "[45] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 2", + "[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3", + "[47] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6847-6857, 2021. 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24295", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 3, 5", + "[49] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 2, 6", + "[50] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. arXiv preprint arXiv:2405.09711, 2024. 5", + "[51] Weijia Wu, Yuzhong Zhao, Zhuang Li, Jiahong Li, Hong Zhou, Mike Zheng Shou, and Xiang Bai. A large cross-modal video retrieval dataset with reading comprehension. Pattern Recognition, 157:110818, 2025. 5", + "[52] Xun Wu, Shaohan Huang, Wenhui Wang, Shuming Ma, Li Dong, and Furu Wei. Multi-head mixture-of-experts. In Advances in Neural Information Processing Systems, pages 94073-94096, 2024. 5", + "[53] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 5", + "[54] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Empirical Methods in Natural Language Processing, 2021. 3", + "[55] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 2, 3", + "[56] An Yang, Junyang Lin, Rui Men, Chang Zhou, Le Jiang, Xianyan Jia, Ang Wang, Jie Zhang, Jiamang Wang, Yong Li, et al. M6-t: Exploring sparse expert models and beyond. arXiv preprint arXiv:2105.15082, 2021. 4", + "[57] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 7", + "[58] Yuanhang Yang, Shiyi Qi, Wenchao Gu, Chaozheng Wang, Cuiyun Gao, and Zenglin Xu. Xmoe: Sparse models with fine-grained and adaptive expert selection. In Findings of the Association for Computational Linguistics, 2024. 3", + "[59] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23056-23065, 2023. 3", + "[60] Yingsen Zeng, Yujie Zhong, Chengjian Feng, and Lin Ma. Unimd: Towards unifying moment retrieval and temporal action detection. In European Conference on Computer Vision, pages 286-304. Springer, 2024. 3" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Zihao Zeng, Yibo Miao, Hongcheng Gao, Hao Zhang, and Zhijie Deng. Adamoe: Token-adaptive routing with null experts for mixture-of-experts language models. In *Findings of the Association for Computational Linguistics*, 2024. 3", + "[62] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 6", + "[63] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Zwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 5", + "[64] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1, 2, 3, 6, 7, 8", + "[65] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 5" + ], + "bbox": [ + 514, + 90, + 906, + 359 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24296", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/a4a233d0-73c0-4327-bc43-78ef9be737e9_model.json b/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/a4a233d0-73c0-4327-bc43-78ef9be737e9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8959cd63dd93d7a22ff91e18c615baeb592245e3 --- /dev/null +++ b/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/a4a233d0-73c0-4327-bc43-78ef9be737e9_model.json @@ -0,0 +1,2235 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.131, + 0.885, + 0.155 + ], + "angle": 0, + "content": "TimeExpert: An Expert-Guided Video LLM for Video Temporal Grounding" + }, + { + "type": "text", + "bbox": [ + 0.206, + 0.18, + 0.792, + 0.201 + ], + "angle": 0, + "content": "Zuhao Yang\\(^{1*}\\) Yingchen Yu\\(^{2}\\) Yunqing Zhao\\(^{2}\\) Shijian Lu\\(^{1\\dagger}\\) Song Bai\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.206, + 0.725, + 0.225 + ], + "angle": 0, + "content": "\\(^{1}\\)Nanyang Technological University \\(^{2}\\)ByteDance Inc." + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.232, + 0.71, + 0.251 + ], + "angle": 0, + "content": "https://mwxely.github.io/projects/yang2025time/index" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.268, + 0.695, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.271, + 0.89, + 0.282 + ], + "angle": 0, + "content": "Performance on VTG Tasks" + }, + { + "type": "image_footnote", + "bbox": [ + 0.707, + 0.284, + 0.897, + 0.311 + ], + "angle": 0, + "content": "MR: Moment Retrieval \nDVC: Dense Video Captioning \nVHD: VideoHighlight Detection" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.313, + 0.907, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.515, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Figure 1. Left: Video Temporal Grounding (VTG) is a fine-grained video understanding task that aims to accurately locate content along with event timestamps based on natural language queries. In this work, we mainly consider three major types of VTG tasks: (1) Moment Retrieval (MR), (2) Video Highlight Detection (VHD), and (3) Dense Video Captioning (DVC). The outputs of VTG often contain textual captions, timestamps, and saliency scores. Right: Unlike existing methods (e.g., TimeChat [40]) that employ a single static model, motivated by expert specialization on different task tokens [16], we propose TimeExpert, an expert-guided Video LLM with dynamic token routing. Through task-aware expert allocation, TimeExpert demonstrates substantial improvements over state-of-the-art Video-LLMs on several VTG benchmarks. For example, here we visualize zero-shot F1 score for DVC on the YouCook2 dataset [64], \\( \\mathrm{R} @ 1_{\\mathrm{IoU}=0.7} \\) for MR on the Charades-STA dataset [13], and HIT@1 for VHD on the QVHighlights dataset [24]. More results and analysis are in Section 4." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.639, + 0.327, + 0.654 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.484, + 0.867 + ], + "angle": 0, + "content": "Video Temporal Grounding (VTG) aims to precisely identify video event segments in response to textual queries. The outputs of VTG tasks manifest as sequences of events, each defined by precise timestamps, saliency scores, and textual descriptions. Despite recent advances, a fundamental limitation persists in existing Video Large Language Models (Video-LLMs): they process all task tokens through identical and static pathways, failing to recognize that temporal localization, saliency assessment, and textual generation represent fundamentally distinct tasks requiring specialized processing. To address this, we introduce TimeExpert, a Mixture-of-Experts (MoE)-based Video-LLM that effectively decomposes VTG tasks by dynamically routing task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.763 + ], + "angle": 0, + "content": "specific tokens (e.g., timestamps, saliency scores) to specialized experts, with increased computational efficiency. Our design choices enable precise handling of each subtask, leading to improved event modeling across diverse VTG applications. Extensive experiments demonstrate that TimeExpert consistently achieves state-of-the-art performance on various VTG tasks such as Dense Video Captioning, Moment Retrieval, and Video Highlight Detection." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.77, + 0.646, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Consider watching a cooking tutorial video and attempting to pinpoint the exact moment when the chef fries the bacon (Figure 1-Left) or other highlighted visual segments. This seemingly simple task reveals a fundamental challenge: while humans can effortlessly recognize these actions visually, determining their precise temporal boundaries—down to the specific second—remains significantly challenging" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.474, + 0.888 + ], + "angle": 0, + "content": "*This work was done while Zuhao Yang was interning at ByteDance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.323, + 0.901 + ], + "angle": 0, + "content": "Shijian Lu is the corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.474, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24286" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.098, + 0.322, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.222, + 0.299, + 0.237 + ], + "angle": 0, + "content": "(a) VTG-Specific Video-LLM" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.096, + 0.589, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.222, + 0.534, + 0.236 + ], + "angle": 0, + "content": "(b) Vanilla MoE" + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.096, + 0.905, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.213, + 0.663, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.217, + 0.723, + 0.248 + ], + "angle": 0, + "content": "removal new expert" + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.222, + 0.874, + 0.238 + ], + "angle": 0, + "content": "(c) TimeExpert (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.258, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 2. A Glimpse of Comparison across VTG Approaches. (a): VTG-specific Video-LLM [16, 40] relies on a single static model with shared parameters for all tasks, limiting their ability to specialize in diverse VTG subtasks. (b): Vanilla MoE improves upon this by activating a fixed set (e.g., \\( k = 2 \\)) of experts, enabling some degree of task specialization. (c): Our TimeExpert takes a step further by implementing an adaptive routing mechanism that dynamically allocates new experts when tokens lack suitable matches and prunes unmatched experts when necessary. This dynamic approach significantly enhances computational efficiency while achieving superior task specialization, particularly for VTG subtasks that require distinct feature representations. Detailed design choices can be found at Section 3." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.369, + 0.482, + 0.46 + ], + "angle": 0, + "content": "even for human observers. This problem forms the core motivation behind Video Temporal Grounding (VTG), which aims to develop intelligent systems capable of accurately locating relevant video segments based on natural language queries, enabling users to efficiently access desired content without manually searching through hours of footage." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.657 + ], + "angle": 0, + "content": "Research Gaps. While traditional video grounding models excel in individual VTG subtasks like Moment Retrieval [18, 24] and Video Highlight Detection [32], their practicality is limited as they cannot handle multiple VTG subtasks concurrently and require extensive task-specific fine-tuning to achieve adequate performance. More recently, Video Large Language Models (VideoLLMs) [5, 27, 29, 46, 55] have made remarkable advancements in coarse-grained video content understanding (e.g., VideoQA [11, 27]), yet they struggle to extend such capabilities to tasks requiring fine-grained temporal localization (e.g., VTG [13, 64]), due to the lack of dedicated mechanisms for explicitly modeling temporal boundaries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Compared to generalist Video-LLMs, VTG-specific models that encode temporal tokens [15, 20, 45] or are trained on temporally-annotated instruction-tuning datasets [19, 40, 49] demonstrate marginal improvements, as they blend temporal information with text tokens without explicitly modeling timestamps and saliency scores, despite videos inherently possessing structured information beyond textual descriptions [16]. Furthermore, existing VTG-specific Video-LLMs rely exclusively on a single LLM as their reasoning backbone, processing all task tokens indiscriminately with shared parameters. This parameter-sharing paradigm inherently constrains the model's ability (Figure 1-Right) to specialize in the diverse subtasks of VTG—such as timestamp prediction, saliency estimation, and caption generation—each of which may demand distinct feature representations for optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.905, + 0.626 + ], + "angle": 0, + "content": "Proposed Method. To address the misalignment between language modeling of existing Video-LLMs and inherent video temporal structure, we introduce TimeExpert, an event-centric modeling framework that decouples the prediction of timestamps, saliency scores, and textual descriptions into distinct specialized tasks, organized in an interleaved sequence. To precisely model the varying token importance across these decoupled tasks, we replace the single LLM backbone of existing Video-LLMs with a fine-grained Mixture-of-Experts (MoE) decoder, with our proposed token-aware dynamic gating and token-adaptive routing strategies, which achieves both computational efficiency and modeling precision by activating only task-relevant parameters during both training and inference. To harness each expert's potential, we develop a task-dependent auxiliary loss that encourages more frequently activated experts to process a greater proportion of task-relevant tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Our Contributions. (1): To our knowledge, TimeExpert is the first to explore the task token importance in VTG by introducing a MoE-based framework with dynamic expert routing. Our expert-guided architecture allows for adaptive learning across different VTG subtasks, effectively mitigating task interference while leveraging expert specialization. (2): We conduct extensive experiments across multiple VTG tasks on several benchmarks to demonstrate the effectiveness of TimeExpert. In particular, we demonstrate substantial improvements over state-of-the-art Video-LLMs through quantitative results and ablation studies, highlighting the advantages of expert-driven token specialization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.828, + 0.655, + 0.844 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Video Temporal Grounding. Video Temporal Grounding (VTG) aims to precisely localize relevant temporal boundaries in a video given textual queries [30]. It can be divided" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24287" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.291, + 0.485, + 0.39 + ], + "angle": 0, + "content": "Figure 3. Architecture Overview of TimeExpert. Our model leverages independent encoders and decoding heads to process time, score, and text inputs and outputs. The timestamps and saliency scores of sampled frames are encoded into special tokens and integrated into the corresponding visual tokens. During inference, the generated response follows a structured format, sequentially incorporating time tokens, score tokens, and text tokens." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.483, + 0.568 + ], + "angle": 0, + "content": "into subtasks like Moment Retrieval [13, 38], Dense Video Captioning [2, 43, 64], Video Summarization [17, 42], and Video Highlight Detection [24, 32]. Traditional methods primarily rely on large-scale video-text pretraining, using training objectives such as video-text contrastive learning [48, 54], video-text matching [4, 26], and timestamp regression [35, 36, 59, 60]. Despite their initial success, these methods are constrained by their generalization ability to handle only a single task per model and require substantial fine-tuning for each new downstream application." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Video-LLMs for VTG Tasks. With the growing reasoning capabilities of Video-LLMs [5, 27, 29, 46, 55], numerous studies have investigated their adaptation to VTG tasks (Figure 2-(a)). For example, TimeChat [40], VTimeLLM [19], and HawkEye [20] fine-tune the Video-LLMs on refined temporal-annotated instruction datasets. Following that, several studies further enhance the Video-LLM's understanding of timestamps by introducing special tokens [15, 20] or a temporal perception module [39]. However, these methods adopt a purely text-to-text framework, formulating VTG as a natural language generation problem, which lacks explicit temporal modeling. Recently, TRACE [16] proposes a causal event modeling framework to capture the inherent structure of videos. Nevertheless, it still fails to incorporate task-specific properties, as they process all task tokens within a single shared reasoning backbone, leading to task interference and limited specialization. In contrast, TimeExpert simultaneously structures Video-LLM outputs into event representations and leverages adaptive expert specialization, enhancing both efficiency and accuracy across diverse VTG tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "LLMs with Mixture-of-Experts. The Mixture-of-Experts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.395 + ], + "angle": 0, + "content": "(MoE) [8, 28, 41] architecture (Figure 2-(b)) has been widely adopted to improve the efficiency of large-scale foundation models during both training and inference by replacing feed-forward network layers in transformer models with multiple specialized sub-networks, known as experts, which enables sparse parameter activation. The activated experts are selected by a gating network, which contain a pre-defined router. Prior methods often implement static top-\\(k\\) routing across all tokens, several studies [14, 21, 58, 61] explored on dynamic routing mechanism that allow tokens to activate varying numbers of experts. However, these approaches either introduce additional hyperparameter complexity or demonstrate limited validation confined to text-to-text tasks, constraining their flexibility and cross-task generalizability. In contrast, our proposed TimeExpert (Figure 2-(c)) introduces dynamic expert routing tailored for Video-LLMs through adaptively adjusting the number of activated experts based on task token importance, ensuring effective expert specialization while eliminating reliance on rigid hyperparameter tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.406, + 0.905, + 0.423 + ], + "angle": 0, + "content": "3. TimeExpert: An Expert-guided Video LLM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.599 + ], + "angle": 0, + "content": "In this section, we firstly introduce some fundamental preliminaries in Section 3.1. Then, we propose TimeExpert (Figure 3), which integrates a dynamic expert selection mechanism with a structured event modeling framework. Our approach implements task-aware dynamic gating (Section 3.2) and token-adaptive routing (Section 3.3) to selectively activates specialized experts based on token relevance across diverse tasks. In Section 3.4, we discuss our proposed task-dependent auxiliary loss that optimizes expert utilization and thereby enhances the stability of MoE-based learning. In Section 3.5, we elaborate on our training recipe." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.655, + 0.621 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.627, + 0.905, + 0.727 + ], + "angle": 0, + "content": "Problem Definition. Given a textual instruction \\(\\mathcal{I}\\) and a sequence of video frames \\(\\mathcal{F}\\) from a video, the objective of VTG task is to generate a structured event representation \\(\\mathcal{R}\\). The output sequence \\(\\mathcal{R}\\) consists of discrete events \\(\\{\\epsilon_1,\\epsilon_2,\\dots ,\\epsilon_M\\}\\), where each event \\(\\epsilon_{m}\\) encapsulates three key components: a timestamp \\(t_m\\), a saliency score \\(s_m\\), and a textual caption \\(c_{m}\\). Formally, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.733, + 0.905, + 0.75 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} = \\left\\{\\epsilon_ {1}, \\epsilon_ {2}, \\dots , \\epsilon_ {M} \\right\\} = \\left\\{\\left(t _ {m}, s _ {m}, c _ {m}\\right) \\mid 1 \\leq m \\leq M \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Causal Event Modeling. Following [16], we formulate the event prediction process in a causal manner, leveraging the sequential dependencies between events. Specifically, given the past event sequence \\(\\epsilon_{1:m-1}\\), the textual instruction \\(\\mathcal{I}\\), and the video frames \\(\\mathcal{F}\\), the probability distribution of the next event \\(\\epsilon_m\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.847, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} \\left(\\epsilon_ {m} \\mid \\epsilon_ {1: m - 1}, \\mathcal {I}, \\mathcal {F}\\right) = \\mathcal {P} \\left(t _ {m}, s _ {m}, c _ {m} \\mid \\epsilon_ {1: m - 1}, \\mathcal {I}, \\mathcal {F}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "This causal formulation ensures that event generation follows a structured temporal order, where each event is pro" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24288" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "gressively determined based on prior events, textual cues, and visual content." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.13, + 0.35, + 0.148 + ], + "angle": 0, + "content": "3.2. Task-aware Dynamic Gating" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.152, + 0.485, + 0.228 + ], + "angle": 0, + "content": "Vanilla MoEs. (Figure 2-(b)) employ the top-\\(k\\) gating strategy, which activates a fixed number of experts for each input token, determined by computing gating scores via a learned network \\(g\\). Specifically, for a given token embedding \\(\\mathbf{x} \\in \\mathbb{R}^d\\), the gating function follows:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.235, + 0.483, + 0.255 + ], + "angle": 0, + "content": "\\[\ng (\\mathbf {x}) \\in \\mathbb {R} ^ {K} := \\operatorname {s o f t m a x} \\left(\\mathbf {W} _ {g} ^ {T} \\mathbf {x}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.264, + 0.484, + 0.311 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_g\\in \\mathbb{R}^{d\\times K}\\) represents the gating network parameters, and \\(K\\) denotes the total number of experts. Then, the output of this MoE layer is then formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.318, + 0.483, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\frac {1}{\\sum_ {e \\in \\operatorname {T o p} - k (g (\\mathbf {x}))} g (\\mathbf {x}) _ {e}} \\sum_ {e \\in \\operatorname {T o p} - k (g (\\mathbf {x}))} g (\\mathbf {x}) _ {e} \\mathbf {E} _ {e} (\\mathbf {x}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.483, + 0.397 + ], + "angle": 0, + "content": "where \\(\\mathbf{E}_e(\\mathbf{x})\\in \\mathbb{R}^d\\) is the output of the \\(e\\)-th expert given input \\(\\mathbf{x}\\), and \\(g(\\mathbf{x})_e\\) is the corresponding gating score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.484, + 0.654 + ], + "angle": 0, + "content": "Implicit Task Preference. Vanilla MoEs face some limitations, such as (1): the reliance on a predefined \\( k \\) for expert selection leads to suboptimal flexibility [6, 9, 56], (as task complexity varies, different tokens may require a varying number of experts based on their importance); and (2): the standard gating function treats all tokens uniformly, neglecting the inherent differences in task-specific significance, which may lead to inefficient expert routing and task interference [21]. Surprisingly, we observe an implicit task preference in these models on VTG tasks, as visualized in Figure 4 of expert assignments. In particular, we measure the ratio of activated task-specific tokens to the number of processed text tokens in that layer. Our findings reveal that certain experts, such as expert #7, consistently activate in response to specific subtasks (e.g., score tokens), even though the model was not explicitly trained to establish such specialization, implying an inherent task-driven preference." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.665, + 0.481, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.8, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Figure 4. Visualization of Expert Assignments on Various Task Tokens using Our Vanilla MoE Implementation. We take layer 4 as an example and only visualized the first 8 experts out of a total of 64, due to the space limitation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Explicit Task-Aware Gating. Based on this observation, we propose a task-aware dynamic gating mechanism that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.184 + ], + "angle": 0, + "content": "jointly considers token-level probabilities and task-specific activation statistics. Instead of applying a static threshold, we introduce a task-weighted gating function that adjusts expert activation dynamically by incorporating task token activation rates. Formally, given an expert representation matrix \\(\\mathbf{W}_g\\in \\mathbb{R}^{d\\times K}\\) and an input token \\(\\mathbf{x}\\in \\mathbb{R}^d\\), we define:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.197, + 0.907, + 0.214 + ], + "angle": 0, + "content": "\\[\ns (\\mathbf {x}) = \\cos (\\mathbf {x}, \\mathbf {W} _ {g}), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.219, + 0.907, + 0.253 + ], + "angle": 0, + "content": "\\[\ng (\\mathbf {x}) = \\operatorname {s i g n} \\left(\\sigma \\left(\\frac {s (\\mathbf {x}) + \\alpha A _ {t}}{1 + \\alpha}\\right) - \\sigma (\\mathbf {G})\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.44 + ], + "angle": 0, + "content": "where \\(A_{t}\\) represents the historical activation rate of the task token type associated with \\(\\mathbf{x}\\); \\(\\alpha\\) is a scaling coefficient that controls the influence of task importance; \\(\\sigma(\\cdot)\\) denotes the sigmoid function for score normalization; \\(\\mathbf{G} \\in \\mathbb{R}^{K}\\) is a learnable threshold where a token routes to expert \\(e\\) only if its similarity score \\(s(\\mathbf{x}) \\in \\mathbb{R}^{K}\\) exceeds \\(\\mathbf{G}\\). Following [14], we make the binary sign() differentiable by copying the upstream gradient of \\(g(\\mathbf{x})\\) to the pre-activation term \\(\\sigma(s(\\mathbf{x}, A_{t})) - \\sigma(\\mathbf{G})\\). With \\(A_{t}\\), the model explicitly prioritizes experts that have demonstrated effectiveness for specific task tokens, encouraging specialization while ensuring under-utilized experts do not get excessively activated." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.449, + 0.738, + 0.466 + ], + "angle": 0, + "content": "3.3. Token-adaptive Routing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.636 + ], + "angle": 0, + "content": "In this section, we introduce a token-adaptive routing mechanism (Figure 2-(c)) that dynamically facilitates expert addition and removal based on recorded activation trends. The routing mechanism comprises three key components, including (1): Task-specific Routing Recording: We continuously monitor expert utilization across different task tokens and the routing results during training. (2): Adaptive Expert Addition: New experts are allocated when certain task tokens persistently fail to activate appropriate experts, and (3): Redundant Expert Removal: Experts with persistently low activation rates are pruned to maintain efficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.637, + 0.906, + 0.757 + ], + "angle": 0, + "content": "Task-specific Routing Recording. Instead of solely tracking which experts get activated, we maintain a per-task token activation frequency metric. Specifically, (1) For each expert \\( e \\), we log its activation timestamps, denoted as \\( \\mathbf{R}_E \\in \\mathbb{R}^K \\); (2) For input tokens that fail to activate any expert, we compute their aggregated embeddings as \\( \\mathbf{R}_S \\in \\mathbb{R}^d \\); (3) We maintain an activation rate tracker \\( A_t \\) per task token type, which influences expert selection in subsequent iterations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.758, + 0.906, + 0.803 + ], + "angle": 0, + "content": "Adaptive Expert Addition. When a significant proportion of task tokens fail to activate any expert, we introduce new experts whose representations are initialized as:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.813, + 0.907, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {g, K + 1} = \\frac {\\mathbf {R} _ {S}}{\\| \\mathbf {R} _ {S} \\|}, \\quad \\mathbf {G} _ {K + 1} = 0, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{R}_S\\) captures the average embedding of underrepresented task tokens. This ensures that newly added experts are aligned with the missing task representations." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24289" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.177, + 0.089, + 0.825, + 0.246 + ], + "angle": 0, + "content": "
Training StageDatasetsNo. of Samples
Stage 1: Task Module PretrainingValley [33], LLaVA-Image [31], TextVR [51], ShareGPT4Video [3], VTG-IT [15]1.9M
Stage 2: MoE Decoder PretrainingValley*, TextVR*, ShareGPT4Video*, VTG-IT*, ActivityNet Captions [2], VideoChatGPT [34], InternVid [48], Next-QA [53]0.9M
Stage 3: Supervised Fine-tuningFiltered and Re-annotated Data from: Previous Stages' Data, EgoQA [37], STAR [50], Moment-10M [39], LLaVA-Video-178K [63]2.3M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.255, + 0.908, + 0.283 + ], + "angle": 0, + "content": "Table 1. Training Data Recipe of TimeExpert. These data are categorized into three training stages. * denotes we utilize the filtered subset of the original dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.34 + ], + "angle": 0, + "content": "Redundant Expert Removal. Experts are pruned if their activation rate \\( A_{e} \\) remains below a threshold \\( \\tau_{\\mathrm{min}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.346, + 0.483, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\text {r e m o v e}} = \\left\\{e \\mid A _ {e} < \\tau_ {\\min } \\right\\}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.484, + 0.414 + ], + "angle": 0, + "content": "This prevents redundant experts from occupying parameter space while ensuring that experts specializing in low-frequency but important task tokens are retained." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.42, + 0.367, + 0.435 + ], + "angle": 0, + "content": "3.4. Task-dependent Auxiliary Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.442, + 0.484, + 0.517 + ], + "angle": 0, + "content": "To enhance the efficiency and specialization of expert activation in TimeExpert, we introduce a task-dependent auxiliary loss, which ensures that experts with higher activation rates receive more relevant task tokens while preventing excessive expert overlap." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.623 + ], + "angle": 0, + "content": "Existing auxiliary loss functions [10, 52] primarily focus on load balancing among experts, which may conflict with our goal of leveraging expert specialization for VTG tasks. Instead, our auxiliary loss promotes a dynamic allocation strategy by reinforcing the association between frequently activated experts and their corresponding task tokens while preventing under-utilization or excessive redundancy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.789 + ], + "angle": 0, + "content": "Our auxiliary loss consists of two key components: task-aware concentration term (Left) and activation regularization term (Right). The task-aware concentration loss ensures that experts specialize in processing specific types of task tokens (e.g., timestamps, saliency scores, captions) by amplifying the assignment of relevant tokens to frequently activated experts. The activation regularization loss prevents the over-activation of a single expert. The two loss terms work synergistically to ensure a diverse and efficient allocation of task tokens. In summary, we define our auxiliary loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.794, + 0.483, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a u x}} = \\lambda_ {1} \\sum_ {e = 1} ^ {K} \\left(\\frac {A _ {e}}{\\sum_ {j = 1} ^ {K} A _ {j}} - \\frac {N _ {e}}{\\sum_ {j = 1} ^ {K} N _ {j}}\\right) ^ {2} + \\lambda_ {2} \\sum_ {e = 1} ^ {K} \\| \\mathbf {w} _ {g, e} \\| _ {2} ^ {2} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(A_{e}\\) represents the activation count of expert \\(e\\) over a batch, \\(N_{e}\\) denotes the number of task tokens assigned to expert \\(e\\), \\(\\mathbf{w}_{g,e}\\) is the representation vector of expert \\(e\\), and \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyperparameters controlling the balance between task-aware specialization and numerical stability." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.309, + 0.673, + 0.326 + ], + "angle": 0, + "content": "3.5. Training Recipe" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.35, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Training Data. To ensure comprehensive learning across various video-language understanding tasks, we curate a diverse dataset spanning three progressive training stages, as summarized in Table 1. These utilized public datasets cover a wide range of temporal reasoning, video grounding, and captioning challenges. We conducted systematic processing, filtering, and re-annotating techniques that remove inaccurate temporal annotations, ensuring high-quality labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.909, + 0.747 + ], + "angle": 0, + "content": "Training Strategy. The training process of TimeExpert includes three stages: (1) Stage 1 - Task Module Pretraining. In the first stage, we pretrain the task-specific modules, including the vision compression layer, task encoder, and task heads, using 1.9M general multimodal video-text samples. This stage ensures that the model learns fundamental video representation capabilities before incorporating MoE decoder. (2) Stage 2 - MoE Decoder Pretraining. Before fine-tuning the full model, we introduce an intermediate MoE pretraining stage with 0.9M samples to ensure that expert routing aligns with different VTG-specific task tokens. This step mitigates the risk of expert collapse (i.e., all tokens being routed to the same expert) and enhances task-aware expert specialization. (3) Stage 3 - Supervised Fine-tuning. In the final stage, we fine-tune the entire Time-Expert framework on 2.3M samples, allowing both the task modules and the MoE decoder to be optimized jointly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Training Objectives. During the first stage, we adopt cross-entropy loss for each generated token to teach basic video-event formatting. During MoE Decoder Pretraining, besides the cross-entropy loss, we apply \\( z \\)-loss [65] to stabilize training and proposed task-aware auxiliary loss to mitigate expert collapse and drive specialization. During Supervised Fine-tuning, we fine-tune the full model (except visual encoder) using same trio of objectives, ensuring stable routing while the model learns task-specific instructions." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24290" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.907, + 0.254 + ], + "angle": 0, + "content": "
MethodNo. of \nActivated \nParametersDense Video Captioning \n(YouCook2 [64])Moment Retrieval \n(Charades-STA [13])Video Highlight Detection \n(QVHighlights [24])
SODAc(↑)CIDEr (↑)F1 Score (↑)R@IIoU=0.5 (↑)R@IIoU=0.7 (↑)mAP (↑)HIT@1 (↑)
TimeChat [40]7B1.23.412.632.213.414.523.9
VTimeLLM [19]7B---27.511.4--
Momentor [39]7B---26.611.67.6-
HawkEye [49]7B---31.414.5--
VTG-LLM [15]7B1.55.017.533.815.716.533.5
TRACE [16]7B2.28.122.440.319.426.842.7
TimeExpert (adaptive k)≈ 5.9B / 3.5B / 4.8B2.58.223.642.820.329.646.9
TimeExpert (TRACE's data)≈ 5.2B / 3.1B / 4.0B2.48.123.341.920.129.146.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.264, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Table 2. Zero-shot Performance Comparison of TimeExpert against several state-of-the-art VTG-specific Video-LLMs on Dense Video Captioning, Temporal Grounding, and Video Highlight Detection. Some results are sourced from [16]. The best results are in bold. We also underline the second-best results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.331, + 0.223, + 0.349 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.358, + 0.291, + 0.374 + ], + "angle": 0, + "content": "4.1. Experimental Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.483, + 0.623 + ], + "angle": 0, + "content": "Implementation Details. We adopt ARIA [25] as the base MoE Video-LLM. Specifically, we utilize a lightweight visual encoder with 438M parameters, consisting of a Vision Transformer (ViT) [7] and a projection module. The ViT weights are initialized from [62]. Each video frame is initially encoded into 128 (for medium resolution) or 256 (for high resolution) visual tokens, which are then compressed via slot-based token compression [15] down to 8 visual tokens per frame. The time encoder and score encoder share the same architecture, both initialized with a tokenizer containing 11 number tokens for representing timestamps/scores, a separator token marking the end of each timestamp/score, and a switching token indicating task transitions. Specifically, following [16], we insert the token \\(\\langle sep\\rangle\\) between consecutive timestamps or scores and append \\(\\langle sync\\rangle\\) at the end of the sequence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Comparison Methods. We select TimeChat [40], VTimeLLM [19], Momentor [39], HawkEye [49], VTGLLM [15], and TRACE [16] as comparison methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Evaluation Benchmarks. To thoroughly evaluate TimeExpert in Video Temporal Grounding, we assess it across three VTG tasks: Dense Video Captioning, Moment Retrieval, and Video Highlight Detection, utilizing datasets such as YouCook2 [64], Charades-STA [13], QVHighlights [24], and ActivityNet Captions [23]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. For Dense Video Captioning, we use metrics including \\(\\mathrm{SODA_c}\\) [12], which is specifically tailored for the video's storyline; CIDEr score [44], which measures consensus between generated and reference captions by computing the TF-IDF weighted n-gram similarity across multiple reference captions; F1 Score, which quantifies the model's capacity on precisely locating timestamps; and METEOR [1], which combines uni-gram precision and recall into a harmonic mean while incorporating stemming," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.907, + 0.47 + ], + "angle": 0, + "content": "synonym matching, and a fragmentation penalty so that longer contiguous matches are rewarded. For Moment Retrieval, we compute the Intersection over Union (IoU) between the timestamps predicted by the model and ground truth, including Recall at IoU thresholds of \\(\\{0.5, 0.7\\}\\) and mean IoU. For Video Highlight Detection, we calculate both the mean Average Precision (mAP) at IoU thresholds of \\(\\{0.5, 0.75\\}\\), and HIT@1, which represents the hit ratio of the highest scored clip." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.707, + 0.499 + ], + "angle": 0, + "content": "4.2.Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Zero-shot Performance on VTG Tasks. In Table 2, we present the zero-shot performance of TimeExpert against several state-of-the-art baselines on three widely adopted VTG benchmarks. We report the number of activated parameters for a transparent comparison. Across the three benchmarks, the adaptive-\\(k\\) variant activated an average of 14.5, 9.1, and 11.8 experts, whereas the TRACE-trained variant averaged 13.5, 8.3, and 10.5. Notably, our approach outperforms other VTG-specific Video-LLMs across all three benchmarks with fewer activated parameters. It achieves a \\(0.3\\%\\) and \\(1.2\\%\\) performance gain on YouCook2 [64] using the \\(\\mathrm{SODA_c}\\) and F1 Score metrics, respectively; a \\(2.5\\%\\) and \\(0.9\\%\\) performance boost in Recall with IoU = {0.5, 0.7} thresholds on Charades-STA [13]; and a \\(2.8\\%\\) and \\(4.2\\%\\) performance gain for the mAP and HIT@1 metrics on QVHighlights [24]. This validates that dynamic gating and adaptive routing not only ensures efficient expert utilization but also fully leverages task specialization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Fine-tuned Performance on VTG Tasks. In Table 3, we present the results of fine-tuning TimeExpert on YouCook2 and Charades-STA datasets for 2 epochs. The results indicate that: (1) TimeExpert surpasses both generalist and task-specific baselines. Unlike TimeChat and VTG-LLM, which struggle to fully benefit from fine-tuning, our MoE-based approach effectively adapts to downstream datasets, achieving new state-of-the-art results. Notably, it surpasses" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24291" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.164, + 0.089, + 0.836, + 0.344 + ], + "angle": 0, + "content": "
MethodDense Video Captioning\n(YouCook2 [64])Moment Retrieval\n(Charades-STA [13])
SODAc(↑)CIDEr(↑)F1 Score(↑)R@1IoU=0.5(↑)R@1IoU=0.7(↑)
Traditional Models
PDVC [47]4.422.7---
Vid2Seq [57]5.725.323.5--
Vid2Seq (Audio Input) [57]7.947.127.3--
CM² [22]5.331.728.4--
VTG-specific Video-LLMs
TimeChat [40]3.411.019.546.723.7
VTG-LLM [15]3.613.420.657.233.4
TRACE [16]6.735.531.861.741.4
TimeExpert (Ours)7.239.033.564.143.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.352, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Table 3. Fine-tuned Performance of TimeExpert. Some results are from [16]. We report the performance of each model after fine-tuning on YouCook2 and Charades-STA datasets. We bold the best scores and italicize results that benefit from additional audio inputs." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.395, + 0.898, + 0.553 + ], + "angle": 0, + "content": "
MethodDense Video CaptioningMoment Retrieval
METEOR (↑)SODAc (↑)CIDEr (↑)F1 Score (↑)R@1IoU=0.5 (↑)R@1IoU=0.7 (↑)mIoU (↑)
TimeChat [40]5.74.719.036.94.62.06.9
VTimeLLM [19]6.85.827.6-29.514.231.4
Momentor* [39]4.72.314.9-23.012.429.3
VTG-LLM [15]5.95.120.734.88.33.712.0
TRACE [16]6.46.025.939.337.724.039.0
TimeExpert (Ours)7.06.528.440.539.226.141.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.56, + 0.907, + 0.591 + ], + "angle": 0, + "content": "Table 4. Performance Comparison on ActivityNet Captions [23]. Some results are sourced from [16]. * indicates zero-shot evaluation. The best results are in bold. TimeExpert achieves state-of-the-art results across both Dense Video Captioning and Moment Retrieval." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.615, + 0.486, + 0.829 + ], + "angle": 0, + "content": "TRACE in both Moment Retrieval and Dense Video Captioning tasks, highlighting the benefits of dynamic expert activation in video-language modeling. (2) TimeExpert enhances retrieval and captioning tasks simultaneously. Our model achieves a \\(+3.5\\) improvement in CIDEr and a \\(+1.7\\) boost in F1 Score on YouCook2, alongside an absolute gain of 2.4 in \\(\\mathrm{R}@\\mathrm{1}_{\\mathrm{IoU}} = 0.5\\) and 1.9 in \\(\\mathrm{R}@\\mathrm{1}_{\\mathrm{IoU}} = 0.7\\) on Charades-STA. These results demonstrate that TimeExpert's token-adaptive routing effectively captures event structures, enabling improved video reasoning. (3) Efficient expert utilization enables balanced trade-offs. Unlike static models, TimeExpert dynamically activates the most relevant experts per input, resulting in stronger generalization to unseen datasets while keeping computation manageable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Performance Comparison on ActivityNet Captions. In Table 4, we evaluate TimeExpert on the ActivityNet Captions [23] dataset. Notably, TimeExpert achieves the best performance across both Moment Retrieval and Dense" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.907, + 0.797 + ], + "angle": 0, + "content": "Video Captioning tasks. It surpasses TRACE, which was previously the strongest baseline, by improving CIDER (+2.5) and F1 Score (+1.2) in Dense Video Captioning, while also significantly boosting R@ \\( 1_{\\mathrm{IoU}} = 0.5 \\) (+1.5) and mIoU (+2.5) in Moment Retrieval. These results highlight TimeExpert's ability to effectively model long-range temporal dependencies while efficiently leveraging a MoE framework for adaptive token routing. Unlike previous methods, which rely on static expert activation, our model dynamically selects experts based on token-wise information, leading to improved alignment between textual queries and visual events." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.681, + 0.831 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Design Choices. As shown in the upper part of Table 5, the removal of token-adaptive routing leads to a notable drop in all metrics, confirming its efficacy in efficiently allocating task tokens. Moreover, removing separate encoders" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24292" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.907, + 0.409 + ], + "angle": 0, + "content": "
MethodDense Video Captioning\n(YouCook2 [64])Moment Retrieval\n(Charades-STA [13])Video Highlight Detection\n(QVHighlights [24])
SODAc(↑)CIDEr (↑)F1 Score (↑)R@1IoU=0.5 (↑)R@1IoU=0.7 (↑)mAP (↑)HIT@1 (↑)
Ablation Studies on Design Choices
w/o token-adaptive routing2.17.622.240.519.226.542.6
w/o separate encoders*-------
w/o task-dependent loss2.47.922.841.319.728.145.2
TimeExpert (Ours)2.58.223.642.820.329.646.9
Ablation Studies on No. of Input Frames
TimeExpert (8 frames)1.55.118.930.213.922.138.4
TimeExpert (64 frames)2.17.521.937.517.827.244.9
TimeExpert (128 frames)2.58.222.041.519.929.346.5
Ablation Studies on No. of Activated Experts
Vanilla MoE (k=2)2.38.122.842.119.829.245.8
Vanilla MoE (k=4)2.48.223.342.520.129.546.6
Vanilla MoE (k=6)2.58.223.542.820.229.646.9
Vanilla MoE (k=8)2.58.123.442.820.229.846.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.418, + 0.907, + 0.463 + ], + "angle": 0, + "content": "Table 5. Ablation Study Results of TimeExpert. We evaluate the impact of removing key components, varying the number of input frames, and adjusting the number of activated experts. Results under the default setting are highlighted in gray. * denotes that directly processing time tokens and score tokens with the text tokenizer causes the model to fail to follow instructions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.488, + 0.482, + 0.625 + ], + "angle": 0, + "content": "and decoding heads for causal event modeling completely disrupts the model's ability to follow instructions, reinforcing the necessity of structured task processing, and removing the auxiliary loss results in a slight performance drop across all tasks, particularly in Video Highlight Detection and Moment Retrieval. This indicates our proposed task-dependent auxiliary loss improves expert specialization by forcing task-aware routing, leading to better alignment between task tokens and relevant experts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.717 + ], + "angle": 0, + "content": "Number of Input Frames. As shown in the middle of Table 5, increasing the number of frames generally improves performance across all tasks, as expected. Notably, the transition from 64 to 128 frames boosts performance in Moment Retrieval and Video Highlight Detection, indicating that a larger temporal context helps refine event understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Number of Activated Experts. As shown in the bottom part of Table 5, there is a clear trend where the predefined number of activated experts directly influences performance. Activating more experts improves performance, particularly in Video Highlight Detection (mAP +0.4 from 2 to 6 experts). Nonetheless, beyond 6 experts, performance saturates, showing that excessive expert activation may introduce redundancy without additional benefits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Training Data. We also conduct a data variant study, training our model on the same data recipe as TRACE to enable a direct comparison. As shown in the bottom line of Table 2, TimeExpert still outperforms TRACE across all VTG" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.595 + ], + "angle": 0, + "content": "tasks. Compared to our full-data model, this variant exhibits slight fluctuations, with some metrics even improving (e.g., CIDEr and \\( \\mathrm{R} @ 1_{\\mathrm{IoU}=0.7} \\)), suggesting that TimeExpert generalizes well even with limited data. These findings highlight TimeExpert's efficiency and adaptability, proving that our model design inherently enhances VTG performance, independent of dataset scale." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.608, + 0.634, + 0.625 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.907, + 0.817 + ], + "angle": 0, + "content": "In this work, we identify two critical limitations in existing VTG-specific Video-LLMs: inappropriate temporal event modeling and uniform task token processing. To address this, we propose TimeExpert, a novel framework that addresses these issues through a Mixture-of-Experts decoder that dynamically routes different task tokens (e.g., timestamps, saliency scores, and textual captions) to specialized experts, ensuring adaptive learning and mitigating task interference. Comprehensive experiments demonstrate Time-Expert's effectiveness across multiple VTG subtasks and general video understanding benchmarks, consistently outperforming existing VTG-specific Video-LLMs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.83, + 0.682, + 0.848 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This study is supported by the Ministry of Education Singapore, under the Tier-1 project scheme with the project number RT18/22." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24293" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[3] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. Sharegpt4video: Improving video understanding and generation with better captions. In Advances in Neural Information Processing Systems, pages 19472-19495, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.344, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[4] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. In Advances in Neural Information Processing Systems, pages 72842-72866, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[5] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.487, + 0.483, + 0.569 + ], + "angle": 0, + "content": "[6] Aidan Clark, Diego de Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. In International Conference on Machine Learning, pages 4057-4086. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.573, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.657, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[8] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.702, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[9] Dongyang Fan, Bettina Messmer, and Martin Jaggi. Towards an empirical understanding of moe design choices. arXiv preprint arXiv:2402.13089, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[10] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.898 + ], + "angle": 0, + "content": "[11] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24108-24118, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[12] Soichiro Fujita, Tsutomu Hirao, Hidetakamigaito, Manabu Okumura, and Masaaki Nagata. Soda: Story oriented dense video captioning evaluation framework. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 517-531. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[13] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[14] Yongxin Guo, Zhenglin Cheng, Xiaoying Tang, Zhaopeng Tu, and Tao Lin. Dynamic mixture of experts: An autotuning approach for efficient transformer models. In International Conference on Learning Representations, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[15] Yongxin Guo, Jingyu Liu, Mingda Li, Dingxin Cheng, Xiaoying Tang, Dianbo Sui, Qingbin Liu, Xi Chen, and Kevin Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. In Proceedings of the AAAI conference on artificial intelligence, 2025. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[16] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Qingbin Liu, and Xi Chen. Trace: Temporal grounding video llm via causal event modeling. In International Conference on Learning Representations, 2025. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[17] Michael Gygli, Helmut Grabner, Hayko Riemenschneider, and Luc Van Gool. Creating summaries from user videos. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13, pages 505-520. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[18] Donghoon Han, Seunghyeon Seo, Eunhwan Park, Seong-Uk Nam, and Nojun Kwak. Unleash the potential of clip for video highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8275-8279, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[19] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.66, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[20] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[21] Quzhe Huang, Zhenwei An, Nan Zhuang, Mingxu Tao, Chen Zhang, Yang Jin, Kun Xu, Liwei Chen, Songfang Huang, and Yansong Feng. Harder tasks need more experts: Dynamic routing in moe models. In Association for Computational Linguistics, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[22] Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi, and Seong Tae Kim. Do you remember? dense video captioning with cross-modal memory retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894-13904, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-capitaling events in videos. In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24294" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE international conference on computer vision, pages 706-715, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[24] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In Advances in Neural Information Processing Systems, pages 11846-11858, 2021. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[25] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model. arXiv preprint arXiv:2410.05993, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[26] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19948-19960, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[27] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[28] Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. Unimoe: Scaling unified multimodal llms with mixture of experts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[29] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Empirical Methods in Natural Language Processing, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[30] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified videolanguage temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[32] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3042-3051, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[33] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[34] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Association for Computational Linguistics, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[35] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration for video temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[36] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23023-23033, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.23 + ], + "angle": 0, + "content": "[37] Thong Thanh Nguyen, Zhiyuan Hu, Xiaobao Wu, Cong-Duy T Nguyen, See-Kiong Ng, and Anh Tuan Luu. Encoding and controlling global semantics for long-form video question answering. arXiv preprint arXiv:2405.19723, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[38] Andreea-Maria Oncescu, Joao F Henriques, Yang Liu, Andrew Zisserman, and Samuel Albanie. Queryd: A video dataset with high-quality text and audio narrations. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2265-2269. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[39] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In International Conference on Machine Learning, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[40] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[41] N Shazeer, A Mirhoseini, K Maziarz, A Davis, Q Le, G Hinton, and J Dean. The sparsely-gated mixture-of-experts layer. Outrageously large neural networks, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[42] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. Tvsum: Summarizing web videos using titles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5179-5187, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[43] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[44] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[45] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[47] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6847-6857, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24295" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[48] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[49] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[50] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. arXiv preprint arXiv:2405.09711, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.273, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[51] Weijia Wu, Yuzhong Zhao, Zhuang Li, Jiahong Li, Hong Zhou, Mike Zheng Shou, and Xiang Bai. A large cross-modal video retrieval dataset with reading comprehension. Pattern Recognition, 157:110818, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[52] Xun Wu, Shaohan Huang, Wenhui Wang, Shuming Ma, Li Dong, and Furu Wei. Multi-head mixture-of-experts. In Advances in Neural Information Processing Systems, pages 94073-94096, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[53] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[54] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Empirical Methods in Natural Language Processing, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[55] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[56] An Yang, Junyang Lin, Rui Men, Chang Zhou, Le Jiang, Xianyan Jia, Ang Wang, Jie Zhang, Jiamang Wang, Yong Li, et al. M6-t: Exploring sparse expert models and beyond. arXiv preprint arXiv:2105.15082, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[57] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[58] Yuanhang Yang, Shiyi Qi, Wenchao Gu, Chaozheng Wang, Cuiyun Gao, and Zenglin Xu. Xmoe: Sparse models with fine-grained and adaptive expert selection. In Findings of the Association for Computational Linguistics, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[59] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23056-23065, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[60] Yingsen Zeng, Yujie Zhong, Chengjian Feng, and Lin Ma. Unimd: Towards unifying moment retrieval and temporal action detection. In European Conference on Computer Vision, pages 286-304. Springer, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.148 + ], + "angle": 0, + "content": "[61] Zihao Zeng, Yibo Miao, Hongcheng Gao, Hao Zhang, and Zhijie Deng. Adamoe: Token-adaptive routing with null experts for mixture-of-experts language models. In *Findings of the Association for Computational Linguistics*, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[62] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.206, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[63] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Zwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.248, + 0.907, + 0.304 + ], + "angle": 0, + "content": "[64] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.305, + 0.907, + 0.361 + ], + "angle": 0, + "content": "[65] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24296" + } + ] +] \ No newline at end of file diff --git a/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/a4a233d0-73c0-4327-bc43-78ef9be737e9_origin.pdf b/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/a4a233d0-73c0-4327-bc43-78ef9be737e9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dd61cf840978dd0441ae574780babf0d633e2466 --- /dev/null +++ b/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/a4a233d0-73c0-4327-bc43-78ef9be737e9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:11d599ebe1f9bc2d71aea169aa0502d4930fa0c2965bc23d6a895b7c615bcf47 +size 987926 diff --git a/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/full.md b/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f29ef9547cd55aeaa261cfd99f9a9ed65e10a1e0 --- /dev/null +++ b/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/full.md @@ -0,0 +1,317 @@ +# TimeExpert: An Expert-Guided Video LLM for Video Temporal Grounding + +Zuhao Yang $^{1*}$ Yingchen Yu $^{2}$ Yunqing Zhao $^{2}$ Shijian Lu $^{1\dagger}$ Song Bai $^{2}$ + +$^{1}$ Nanyang Technological University $^{2}$ ByteDance Inc. + +https://mwxely.github.io/projects/yang2025time/index + +![](images/b0c95e601c0873448604778cced1d04c13252a9987dc544923b584b875bf68ad.jpg) +Figure 1. Left: Video Temporal Grounding (VTG) is a fine-grained video understanding task that aims to accurately locate content along with event timestamps based on natural language queries. In this work, we mainly consider three major types of VTG tasks: (1) Moment Retrieval (MR), (2) Video Highlight Detection (VHD), and (3) Dense Video Captioning (DVC). The outputs of VTG often contain textual captions, timestamps, and saliency scores. Right: Unlike existing methods (e.g., TimeChat [40]) that employ a single static model, motivated by expert specialization on different task tokens [16], we propose TimeExpert, an expert-guided Video LLM with dynamic token routing. Through task-aware expert allocation, TimeExpert demonstrates substantial improvements over state-of-the-art Video-LLMs on several VTG benchmarks. For example, here we visualize zero-shot F1 score for DVC on the YouCook2 dataset [64], $\mathrm{R} @ 1_{\mathrm{IoU}=0.7}$ for MR on the Charades-STA dataset [13], and HIT@1 for VHD on the QVHighlights dataset [24]. More results and analysis are in Section 4. + +Performance on VTG Tasks +![](images/65e4838df396854c9b320b09a462517a187b3c3edd34b3b12e31370be77de8f5.jpg) +MR: Moment Retrieval +DVC: Dense Video Captioning +VHD: VideoHighlight Detection + +# Abstract + +Video Temporal Grounding (VTG) aims to precisely identify video event segments in response to textual queries. The outputs of VTG tasks manifest as sequences of events, each defined by precise timestamps, saliency scores, and textual descriptions. Despite recent advances, a fundamental limitation persists in existing Video Large Language Models (Video-LLMs): they process all task tokens through identical and static pathways, failing to recognize that temporal localization, saliency assessment, and textual generation represent fundamentally distinct tasks requiring specialized processing. To address this, we introduce TimeExpert, a Mixture-of-Experts (MoE)-based Video-LLM that effectively decomposes VTG tasks by dynamically routing task + +specific tokens (e.g., timestamps, saliency scores) to specialized experts, with increased computational efficiency. Our design choices enable precise handling of each subtask, leading to improved event modeling across diverse VTG applications. Extensive experiments demonstrate that TimeExpert consistently achieves state-of-the-art performance on various VTG tasks such as Dense Video Captioning, Moment Retrieval, and Video Highlight Detection. + +# 1. Introduction + +Consider watching a cooking tutorial video and attempting to pinpoint the exact moment when the chef fries the bacon (Figure 1-Left) or other highlighted visual segments. This seemingly simple task reveals a fundamental challenge: while humans can effortlessly recognize these actions visually, determining their precise temporal boundaries—down to the specific second—remains significantly challenging + +![](images/006e0b32fae927e7b0f24cd9df7232c869e4204106b850a7fd37a768b032e965.jpg) +(a) VTG-Specific Video-LLM + +![](images/8cdc35db096fdd413da3c0fe6607220a59d1f82790d345aaa9c47fb3ce2f4bfe.jpg) +(b) Vanilla MoE + +![](images/9878de8d107a84d2898c11809cabda9ccc5d5f27da7d645a0aa179b8e3ae5d4d.jpg) +removal new expert +(c) TimeExpert (Ours) + +![](images/3c281f57eeeacc28a909a3388afc5e9152bb1f32410c4ef16693874e9b2bb086.jpg) +Figure 2. A Glimpse of Comparison across VTG Approaches. (a): VTG-specific Video-LLM [16, 40] relies on a single static model with shared parameters for all tasks, limiting their ability to specialize in diverse VTG subtasks. (b): Vanilla MoE improves upon this by activating a fixed set (e.g., $k = 2$ ) of experts, enabling some degree of task specialization. (c): Our TimeExpert takes a step further by implementing an adaptive routing mechanism that dynamically allocates new experts when tokens lack suitable matches and prunes unmatched experts when necessary. This dynamic approach significantly enhances computational efficiency while achieving superior task specialization, particularly for VTG subtasks that require distinct feature representations. Detailed design choices can be found at Section 3. + +even for human observers. This problem forms the core motivation behind Video Temporal Grounding (VTG), which aims to develop intelligent systems capable of accurately locating relevant video segments based on natural language queries, enabling users to efficiently access desired content without manually searching through hours of footage. + +Research Gaps. While traditional video grounding models excel in individual VTG subtasks like Moment Retrieval [18, 24] and Video Highlight Detection [32], their practicality is limited as they cannot handle multiple VTG subtasks concurrently and require extensive task-specific fine-tuning to achieve adequate performance. More recently, Video Large Language Models (VideoLLMs) [5, 27, 29, 46, 55] have made remarkable advancements in coarse-grained video content understanding (e.g., VideoQA [11, 27]), yet they struggle to extend such capabilities to tasks requiring fine-grained temporal localization (e.g., VTG [13, 64]), due to the lack of dedicated mechanisms for explicitly modeling temporal boundaries. + +Compared to generalist Video-LLMs, VTG-specific models that encode temporal tokens [15, 20, 45] or are trained on temporally-annotated instruction-tuning datasets [19, 40, 49] demonstrate marginal improvements, as they blend temporal information with text tokens without explicitly modeling timestamps and saliency scores, despite videos inherently possessing structured information beyond textual descriptions [16]. Furthermore, existing VTG-specific Video-LLMs rely exclusively on a single LLM as their reasoning backbone, processing all task tokens indiscriminately with shared parameters. This parameter-sharing paradigm inherently constrains the model's ability (Figure 1-Right) to specialize in the diverse subtasks of VTG—such as timestamp prediction, saliency estimation, and caption generation—each of which may demand distinct feature representations for optimal performance. + +Proposed Method. To address the misalignment between language modeling of existing Video-LLMs and inherent video temporal structure, we introduce TimeExpert, an event-centric modeling framework that decouples the prediction of timestamps, saliency scores, and textual descriptions into distinct specialized tasks, organized in an interleaved sequence. To precisely model the varying token importance across these decoupled tasks, we replace the single LLM backbone of existing Video-LLMs with a fine-grained Mixture-of-Experts (MoE) decoder, with our proposed token-aware dynamic gating and token-adaptive routing strategies, which achieves both computational efficiency and modeling precision by activating only task-relevant parameters during both training and inference. To harness each expert's potential, we develop a task-dependent auxiliary loss that encourages more frequently activated experts to process a greater proportion of task-relevant tokens. + +Our Contributions. (1): To our knowledge, TimeExpert is the first to explore the task token importance in VTG by introducing a MoE-based framework with dynamic expert routing. Our expert-guided architecture allows for adaptive learning across different VTG subtasks, effectively mitigating task interference while leveraging expert specialization. (2): We conduct extensive experiments across multiple VTG tasks on several benchmarks to demonstrate the effectiveness of TimeExpert. In particular, we demonstrate substantial improvements over state-of-the-art Video-LLMs through quantitative results and ablation studies, highlighting the advantages of expert-driven token specialization. + +# 2. Related Work + +Video Temporal Grounding. Video Temporal Grounding (VTG) aims to precisely localize relevant temporal boundaries in a video given textual queries [30]. It can be divided + +![](images/7530c8befb15395ae0e9f0a1e5586f316e0bc05cd32809cade8a001271f20f96.jpg) +Figure 3. Architecture Overview of TimeExpert. Our model leverages independent encoders and decoding heads to process time, score, and text inputs and outputs. The timestamps and saliency scores of sampled frames are encoded into special tokens and integrated into the corresponding visual tokens. During inference, the generated response follows a structured format, sequentially incorporating time tokens, score tokens, and text tokens. + +into subtasks like Moment Retrieval [13, 38], Dense Video Captioning [2, 43, 64], Video Summarization [17, 42], and Video Highlight Detection [24, 32]. Traditional methods primarily rely on large-scale video-text pretraining, using training objectives such as video-text contrastive learning [48, 54], video-text matching [4, 26], and timestamp regression [35, 36, 59, 60]. Despite their initial success, these methods are constrained by their generalization ability to handle only a single task per model and require substantial fine-tuning for each new downstream application. + +Video-LLMs for VTG Tasks. With the growing reasoning capabilities of Video-LLMs [5, 27, 29, 46, 55], numerous studies have investigated their adaptation to VTG tasks (Figure 2-(a)). For example, TimeChat [40], VTimeLLM [19], and HawkEye [20] fine-tune the Video-LLMs on refined temporal-annotated instruction datasets. Following that, several studies further enhance the Video-LLM's understanding of timestamps by introducing special tokens [15, 20] or a temporal perception module [39]. However, these methods adopt a purely text-to-text framework, formulating VTG as a natural language generation problem, which lacks explicit temporal modeling. Recently, TRACE [16] proposes a causal event modeling framework to capture the inherent structure of videos. Nevertheless, it still fails to incorporate task-specific properties, as they process all task tokens within a single shared reasoning backbone, leading to task interference and limited specialization. In contrast, TimeExpert simultaneously structures Video-LLM outputs into event representations and leverages adaptive expert specialization, enhancing both efficiency and accuracy across diverse VTG tasks. + +LLMs with Mixture-of-Experts. The Mixture-of-Experts + +(MoE) [8, 28, 41] architecture (Figure 2-(b)) has been widely adopted to improve the efficiency of large-scale foundation models during both training and inference by replacing feed-forward network layers in transformer models with multiple specialized sub-networks, known as experts, which enables sparse parameter activation. The activated experts are selected by a gating network, which contain a pre-defined router. Prior methods often implement static top- $k$ routing across all tokens, several studies [14, 21, 58, 61] explored on dynamic routing mechanism that allow tokens to activate varying numbers of experts. However, these approaches either introduce additional hyperparameter complexity or demonstrate limited validation confined to text-to-text tasks, constraining their flexibility and cross-task generalizability. In contrast, our proposed TimeExpert (Figure 2-(c)) introduces dynamic expert routing tailored for Video-LLMs through adaptively adjusting the number of activated experts based on task token importance, ensuring effective expert specialization while eliminating reliance on rigid hyperparameter tuning. + +# 3. TimeExpert: An Expert-guided Video LLM + +In this section, we firstly introduce some fundamental preliminaries in Section 3.1. Then, we propose TimeExpert (Figure 3), which integrates a dynamic expert selection mechanism with a structured event modeling framework. Our approach implements task-aware dynamic gating (Section 3.2) and token-adaptive routing (Section 3.3) to selectively activates specialized experts based on token relevance across diverse tasks. In Section 3.4, we discuss our proposed task-dependent auxiliary loss that optimizes expert utilization and thereby enhances the stability of MoE-based learning. In Section 3.5, we elaborate on our training recipe. + +# 3.1. Preliminaries + +Problem Definition. Given a textual instruction $\mathcal{I}$ and a sequence of video frames $\mathcal{F}$ from a video, the objective of VTG task is to generate a structured event representation $\mathcal{R}$ . The output sequence $\mathcal{R}$ consists of discrete events $\{\epsilon_1,\epsilon_2,\dots ,\epsilon_M\}$ , where each event $\epsilon_{m}$ encapsulates three key components: a timestamp $t_m$ , a saliency score $s_m$ , and a textual caption $c_{m}$ . Formally, we have: + +$$ +\mathcal {R} = \left\{\epsilon_ {1}, \epsilon_ {2}, \dots , \epsilon_ {M} \right\} = \left\{\left(t _ {m}, s _ {m}, c _ {m}\right) \mid 1 \leq m \leq M \right\}. \tag {1} +$$ + +Causal Event Modeling. Following [16], we formulate the event prediction process in a causal manner, leveraging the sequential dependencies between events. Specifically, given the past event sequence $\epsilon_{1:m-1}$ , the textual instruction $\mathcal{I}$ , and the video frames $\mathcal{F}$ , the probability distribution of the next event $\epsilon_m$ is defined as: + +$$ +\mathcal {P} \left(\epsilon_ {m} \mid \epsilon_ {1: m - 1}, \mathcal {I}, \mathcal {F}\right) = \mathcal {P} \left(t _ {m}, s _ {m}, c _ {m} \mid \epsilon_ {1: m - 1}, \mathcal {I}, \mathcal {F}\right). \tag {2} +$$ + +This causal formulation ensures that event generation follows a structured temporal order, where each event is pro + +gressively determined based on prior events, textual cues, and visual content. + +# 3.2. Task-aware Dynamic Gating + +Vanilla MoEs. (Figure 2-(b)) employ the top- $k$ gating strategy, which activates a fixed number of experts for each input token, determined by computing gating scores via a learned network $g$ . Specifically, for a given token embedding $\mathbf{x} \in \mathbb{R}^d$ , the gating function follows: + +$$ +g (\mathbf {x}) \in \mathbb {R} ^ {K} := \operatorname {s o f t m a x} \left(\mathbf {W} _ {g} ^ {T} \mathbf {x}\right), \tag {3} +$$ + +where $\mathbf{W}_g\in \mathbb{R}^{d\times K}$ represents the gating network parameters, and $K$ denotes the total number of experts. Then, the output of this MoE layer is then formulated as: + +$$ +\mathbf {y} = \frac {1}{\sum_ {e \in \operatorname {T o p} - k (g (\mathbf {x}))} g (\mathbf {x}) _ {e}} \sum_ {e \in \operatorname {T o p} - k (g (\mathbf {x}))} g (\mathbf {x}) _ {e} \mathbf {E} _ {e} (\mathbf {x}), \tag {4} +$$ + +where $\mathbf{E}_e(\mathbf{x})\in \mathbb{R}^d$ is the output of the $e$ -th expert given input $\mathbf{x}$ , and $g(\mathbf{x})_e$ is the corresponding gating score. + +Implicit Task Preference. Vanilla MoEs face some limitations, such as (1): the reliance on a predefined $k$ for expert selection leads to suboptimal flexibility [6, 9, 56], (as task complexity varies, different tokens may require a varying number of experts based on their importance); and (2): the standard gating function treats all tokens uniformly, neglecting the inherent differences in task-specific significance, which may lead to inefficient expert routing and task interference [21]. Surprisingly, we observe an implicit task preference in these models on VTG tasks, as visualized in Figure 4 of expert assignments. In particular, we measure the ratio of activated task-specific tokens to the number of processed text tokens in that layer. Our findings reveal that certain experts, such as expert #7, consistently activate in response to specific subtasks (e.g., score tokens), even though the model was not explicitly trained to establish such specialization, implying an inherent task-driven preference. + +![](images/650963354869e4cee6fad7bb905e1f85f62ae301d731dd75b16a279bd9a5fc28.jpg) +Figure 4. Visualization of Expert Assignments on Various Task Tokens using Our Vanilla MoE Implementation. We take layer 4 as an example and only visualized the first 8 experts out of a total of 64, due to the space limitation. + +Explicit Task-Aware Gating. Based on this observation, we propose a task-aware dynamic gating mechanism that + +jointly considers token-level probabilities and task-specific activation statistics. Instead of applying a static threshold, we introduce a task-weighted gating function that adjusts expert activation dynamically by incorporating task token activation rates. Formally, given an expert representation matrix $\mathbf{W}_g\in \mathbb{R}^{d\times K}$ and an input token $\mathbf{x}\in \mathbb{R}^d$ , we define: + +$$ +s (\mathbf {x}) = \cos (\mathbf {x}, \mathbf {W} _ {g}), \tag {5} +$$ + +$$ +g (\mathbf {x}) = \operatorname {s i g n} \left(\sigma \left(\frac {s (\mathbf {x}) + \alpha A _ {t}}{1 + \alpha}\right) - \sigma (\mathbf {G})\right), \tag {6} +$$ + +where $A_{t}$ represents the historical activation rate of the task token type associated with $\mathbf{x}$ ; $\alpha$ is a scaling coefficient that controls the influence of task importance; $\sigma(\cdot)$ denotes the sigmoid function for score normalization; $\mathbf{G} \in \mathbb{R}^{K}$ is a learnable threshold where a token routes to expert $e$ only if its similarity score $s(\mathbf{x}) \in \mathbb{R}^{K}$ exceeds $\mathbf{G}$ . Following [14], we make the binary sign() differentiable by copying the upstream gradient of $g(\mathbf{x})$ to the pre-activation term $\sigma(s(\mathbf{x}, A_{t})) - \sigma(\mathbf{G})$ . With $A_{t}$ , the model explicitly prioritizes experts that have demonstrated effectiveness for specific task tokens, encouraging specialization while ensuring under-utilized experts do not get excessively activated. + +# 3.3. Token-adaptive Routing + +In this section, we introduce a token-adaptive routing mechanism (Figure 2-(c)) that dynamically facilitates expert addition and removal based on recorded activation trends. The routing mechanism comprises three key components, including (1): Task-specific Routing Recording: We continuously monitor expert utilization across different task tokens and the routing results during training. (2): Adaptive Expert Addition: New experts are allocated when certain task tokens persistently fail to activate appropriate experts, and (3): Redundant Expert Removal: Experts with persistently low activation rates are pruned to maintain efficiency. + +Task-specific Routing Recording. Instead of solely tracking which experts get activated, we maintain a per-task token activation frequency metric. Specifically, (1) For each expert $e$ , we log its activation timestamps, denoted as $\mathbf{R}_E \in \mathbb{R}^K$ ; (2) For input tokens that fail to activate any expert, we compute their aggregated embeddings as $\mathbf{R}_S \in \mathbb{R}^d$ ; (3) We maintain an activation rate tracker $A_t$ per task token type, which influences expert selection in subsequent iterations. + +Adaptive Expert Addition. When a significant proportion of task tokens fail to activate any expert, we introduce new experts whose representations are initialized as: + +$$ +\mathbf {W} _ {g, K + 1} = \frac {\mathbf {R} _ {S}}{\| \mathbf {R} _ {S} \|}, \quad \mathbf {G} _ {K + 1} = 0, \tag {7} +$$ + +where $\mathbf{R}_S$ captures the average embedding of underrepresented task tokens. This ensures that newly added experts are aligned with the missing task representations. + +
Training StageDatasetsNo. of Samples
Stage 1: Task Module PretrainingValley [33], LLaVA-Image [31], TextVR [51], ShareGPT4Video [3], VTG-IT [15]1.9M
Stage 2: MoE Decoder PretrainingValley*, TextVR*, ShareGPT4Video*, VTG-IT*, ActivityNet Captions [2], VideoChatGPT [34], InternVid [48], Next-QA [53]0.9M
Stage 3: Supervised Fine-tuningFiltered and Re-annotated Data from: Previous Stages' Data, EgoQA [37], STAR [50], Moment-10M [39], LLaVA-Video-178K [63]2.3M
+ +Table 1. Training Data Recipe of TimeExpert. These data are categorized into three training stages. * denotes we utilize the filtered subset of the original dataset. + +Redundant Expert Removal. Experts are pruned if their activation rate $A_{e}$ remains below a threshold $\tau_{\mathrm{min}}$ : + +$$ +\mathcal {E} _ {\text {r e m o v e}} = \left\{e \mid A _ {e} < \tau_ {\min } \right\}. \tag {8} +$$ + +This prevents redundant experts from occupying parameter space while ensuring that experts specializing in low-frequency but important task tokens are retained. + +# 3.4. Task-dependent Auxiliary Loss + +To enhance the efficiency and specialization of expert activation in TimeExpert, we introduce a task-dependent auxiliary loss, which ensures that experts with higher activation rates receive more relevant task tokens while preventing excessive expert overlap. + +Existing auxiliary loss functions [10, 52] primarily focus on load balancing among experts, which may conflict with our goal of leveraging expert specialization for VTG tasks. Instead, our auxiliary loss promotes a dynamic allocation strategy by reinforcing the association between frequently activated experts and their corresponding task tokens while preventing under-utilization or excessive redundancy. + +Our auxiliary loss consists of two key components: task-aware concentration term (Left) and activation regularization term (Right). The task-aware concentration loss ensures that experts specialize in processing specific types of task tokens (e.g., timestamps, saliency scores, captions) by amplifying the assignment of relevant tokens to frequently activated experts. The activation regularization loss prevents the over-activation of a single expert. The two loss terms work synergistically to ensure a diverse and efficient allocation of task tokens. In summary, we define our auxiliary loss as follows: + +$$ +\mathcal {L} _ {\text {a u x}} = \lambda_ {1} \sum_ {e = 1} ^ {K} \left(\frac {A _ {e}}{\sum_ {j = 1} ^ {K} A _ {j}} - \frac {N _ {e}}{\sum_ {j = 1} ^ {K} N _ {j}}\right) ^ {2} + \lambda_ {2} \sum_ {e = 1} ^ {K} \| \mathbf {w} _ {g, e} \| _ {2} ^ {2} \tag {9} +$$ + +where $A_{e}$ represents the activation count of expert $e$ over a batch, $N_{e}$ denotes the number of task tokens assigned to expert $e$ , $\mathbf{w}_{g,e}$ is the representation vector of expert $e$ , and $\lambda_{1}$ and $\lambda_{2}$ are hyperparameters controlling the balance between task-aware specialization and numerical stability. + +# 3.5. Training Recipe + +Training Data. To ensure comprehensive learning across various video-language understanding tasks, we curate a diverse dataset spanning three progressive training stages, as summarized in Table 1. These utilized public datasets cover a wide range of temporal reasoning, video grounding, and captioning challenges. We conducted systematic processing, filtering, and re-annotating techniques that remove inaccurate temporal annotations, ensuring high-quality labels. + +Training Strategy. The training process of TimeExpert includes three stages: (1) Stage 1 - Task Module Pretraining. In the first stage, we pretrain the task-specific modules, including the vision compression layer, task encoder, and task heads, using 1.9M general multimodal video-text samples. This stage ensures that the model learns fundamental video representation capabilities before incorporating MoE decoder. (2) Stage 2 - MoE Decoder Pretraining. Before fine-tuning the full model, we introduce an intermediate MoE pretraining stage with 0.9M samples to ensure that expert routing aligns with different VTG-specific task tokens. This step mitigates the risk of expert collapse (i.e., all tokens being routed to the same expert) and enhances task-aware expert specialization. (3) Stage 3 - Supervised Fine-tuning. In the final stage, we fine-tune the entire Time-Expert framework on 2.3M samples, allowing both the task modules and the MoE decoder to be optimized jointly. + +Training Objectives. During the first stage, we adopt cross-entropy loss for each generated token to teach basic video-event formatting. During MoE Decoder Pretraining, besides the cross-entropy loss, we apply $z$ -loss [65] to stabilize training and proposed task-aware auxiliary loss to mitigate expert collapse and drive specialization. During Supervised Fine-tuning, we fine-tune the full model (except visual encoder) using same trio of objectives, ensuring stable routing while the model learns task-specific instructions. + +
MethodNo. of +Activated +ParametersDense Video Captioning +(YouCook2 [64])Moment Retrieval +(Charades-STA [13])Video Highlight Detection +(QVHighlights [24])
SODAc(↑)CIDEr (↑)F1 Score (↑)R@IIoU=0.5 (↑)R@IIoU=0.7 (↑)mAP (↑)HIT@1 (↑)
TimeChat [40]7B1.23.412.632.213.414.523.9
VTimeLLM [19]7B---27.511.4--
Momentor [39]7B---26.611.67.6-
HawkEye [49]7B---31.414.5--
VTG-LLM [15]7B1.55.017.533.815.716.533.5
TRACE [16]7B2.28.122.440.319.426.842.7
TimeExpert (adaptive k)≈ 5.9B / 3.5B / 4.8B2.58.223.642.820.329.646.9
TimeExpert (TRACE's data)≈ 5.2B / 3.1B / 4.0B2.48.123.341.920.129.146.3
+ +Table 2. Zero-shot Performance Comparison of TimeExpert against several state-of-the-art VTG-specific Video-LLMs on Dense Video Captioning, Temporal Grounding, and Video Highlight Detection. Some results are sourced from [16]. The best results are in bold. We also underline the second-best results. + +# 4. Experiments + +# 4.1. Experimental Details + +Implementation Details. We adopt ARIA [25] as the base MoE Video-LLM. Specifically, we utilize a lightweight visual encoder with 438M parameters, consisting of a Vision Transformer (ViT) [7] and a projection module. The ViT weights are initialized from [62]. Each video frame is initially encoded into 128 (for medium resolution) or 256 (for high resolution) visual tokens, which are then compressed via slot-based token compression [15] down to 8 visual tokens per frame. The time encoder and score encoder share the same architecture, both initialized with a tokenizer containing 11 number tokens for representing timestamps/scores, a separator token marking the end of each timestamp/score, and a switching token indicating task transitions. Specifically, following [16], we insert the token $\langle sep\rangle$ between consecutive timestamps or scores and append $\langle sync\rangle$ at the end of the sequence. + +Comparison Methods. We select TimeChat [40], VTimeLLM [19], Momentor [39], HawkEye [49], VTGLLM [15], and TRACE [16] as comparison methods. + +Evaluation Benchmarks. To thoroughly evaluate TimeExpert in Video Temporal Grounding, we assess it across three VTG tasks: Dense Video Captioning, Moment Retrieval, and Video Highlight Detection, utilizing datasets such as YouCook2 [64], Charades-STA [13], QVHighlights [24], and ActivityNet Captions [23]. + +Evaluation Metrics. For Dense Video Captioning, we use metrics including $\mathrm{SODA_c}$ [12], which is specifically tailored for the video's storyline; CIDEr score [44], which measures consensus between generated and reference captions by computing the TF-IDF weighted n-gram similarity across multiple reference captions; F1 Score, which quantifies the model's capacity on precisely locating timestamps; and METEOR [1], which combines uni-gram precision and recall into a harmonic mean while incorporating stemming, + +synonym matching, and a fragmentation penalty so that longer contiguous matches are rewarded. For Moment Retrieval, we compute the Intersection over Union (IoU) between the timestamps predicted by the model and ground truth, including Recall at IoU thresholds of $\{0.5, 0.7\}$ and mean IoU. For Video Highlight Detection, we calculate both the mean Average Precision (mAP) at IoU thresholds of $\{0.5, 0.75\}$ , and HIT@1, which represents the hit ratio of the highest scored clip. + +# 4.2.Quantitative Results + +Zero-shot Performance on VTG Tasks. In Table 2, we present the zero-shot performance of TimeExpert against several state-of-the-art baselines on three widely adopted VTG benchmarks. We report the number of activated parameters for a transparent comparison. Across the three benchmarks, the adaptive- $k$ variant activated an average of 14.5, 9.1, and 11.8 experts, whereas the TRACE-trained variant averaged 13.5, 8.3, and 10.5. Notably, our approach outperforms other VTG-specific Video-LLMs across all three benchmarks with fewer activated parameters. It achieves a $0.3\%$ and $1.2\%$ performance gain on YouCook2 [64] using the $\mathrm{SODA_c}$ and F1 Score metrics, respectively; a $2.5\%$ and $0.9\%$ performance boost in Recall with IoU = {0.5, 0.7} thresholds on Charades-STA [13]; and a $2.8\%$ and $4.2\%$ performance gain for the mAP and HIT@1 metrics on QVHighlights [24]. This validates that dynamic gating and adaptive routing not only ensures efficient expert utilization but also fully leverages task specialization. + +Fine-tuned Performance on VTG Tasks. In Table 3, we present the results of fine-tuning TimeExpert on YouCook2 and Charades-STA datasets for 2 epochs. The results indicate that: (1) TimeExpert surpasses both generalist and task-specific baselines. Unlike TimeChat and VTG-LLM, which struggle to fully benefit from fine-tuning, our MoE-based approach effectively adapts to downstream datasets, achieving new state-of-the-art results. Notably, it surpasses + +
MethodDense Video Captioning +(YouCook2 [64])Moment Retrieval +(Charades-STA [13])
SODAc(↑)CIDEr(↑)F1 Score(↑)R@1IoU=0.5(↑)R@1IoU=0.7(↑)
Traditional Models
PDVC [47]4.422.7---
Vid2Seq [57]5.725.323.5--
Vid2Seq (Audio Input) [57]7.947.127.3--
CM² [22]5.331.728.4--
VTG-specific Video-LLMs
TimeChat [40]3.411.019.546.723.7
VTG-LLM [15]3.613.420.657.233.4
TRACE [16]6.735.531.861.741.4
TimeExpert (Ours)7.239.033.564.143.3
+ +Table 3. Fine-tuned Performance of TimeExpert. Some results are from [16]. We report the performance of each model after fine-tuning on YouCook2 and Charades-STA datasets. We bold the best scores and italicize results that benefit from additional audio inputs. + +
MethodDense Video CaptioningMoment Retrieval
METEOR (↑)SODAc (↑)CIDEr (↑)F1 Score (↑)R@1IoU=0.5 (↑)R@1IoU=0.7 (↑)mIoU (↑)
TimeChat [40]5.74.719.036.94.62.06.9
VTimeLLM [19]6.85.827.6-29.514.231.4
Momentor* [39]4.72.314.9-23.012.429.3
VTG-LLM [15]5.95.120.734.88.33.712.0
TRACE [16]6.46.025.939.337.724.039.0
TimeExpert (Ours)7.06.528.440.539.226.141.5
+ +Table 4. Performance Comparison on ActivityNet Captions [23]. Some results are sourced from [16]. * indicates zero-shot evaluation. The best results are in bold. TimeExpert achieves state-of-the-art results across both Dense Video Captioning and Moment Retrieval. + +TRACE in both Moment Retrieval and Dense Video Captioning tasks, highlighting the benefits of dynamic expert activation in video-language modeling. (2) TimeExpert enhances retrieval and captioning tasks simultaneously. Our model achieves a $+3.5$ improvement in CIDEr and a $+1.7$ boost in F1 Score on YouCook2, alongside an absolute gain of 2.4 in $\mathrm{R}@\mathrm{1}_{\mathrm{IoU}} = 0.5$ and 1.9 in $\mathrm{R}@\mathrm{1}_{\mathrm{IoU}} = 0.7$ on Charades-STA. These results demonstrate that TimeExpert's token-adaptive routing effectively captures event structures, enabling improved video reasoning. (3) Efficient expert utilization enables balanced trade-offs. Unlike static models, TimeExpert dynamically activates the most relevant experts per input, resulting in stronger generalization to unseen datasets while keeping computation manageable. + +Performance Comparison on ActivityNet Captions. In Table 4, we evaluate TimeExpert on the ActivityNet Captions [23] dataset. Notably, TimeExpert achieves the best performance across both Moment Retrieval and Dense + +Video Captioning tasks. It surpasses TRACE, which was previously the strongest baseline, by improving CIDER (+2.5) and F1 Score (+1.2) in Dense Video Captioning, while also significantly boosting R@ $1_{\mathrm{IoU}} = 0.5$ (+1.5) and mIoU (+2.5) in Moment Retrieval. These results highlight TimeExpert's ability to effectively model long-range temporal dependencies while efficiently leveraging a MoE framework for adaptive token routing. Unlike previous methods, which rely on static expert activation, our model dynamically selects experts based on token-wise information, leading to improved alignment between textual queries and visual events. + +# 4.3. Ablation Studies + +Design Choices. As shown in the upper part of Table 5, the removal of token-adaptive routing leads to a notable drop in all metrics, confirming its efficacy in efficiently allocating task tokens. Moreover, removing separate encoders + +
MethodDense Video Captioning +(YouCook2 [64])Moment Retrieval +(Charades-STA [13])Video Highlight Detection +(QVHighlights [24])
SODAc(↑)CIDEr (↑)F1 Score (↑)R@1IoU=0.5 (↑)R@1IoU=0.7 (↑)mAP (↑)HIT@1 (↑)
Ablation Studies on Design Choices
w/o token-adaptive routing2.17.622.240.519.226.542.6
w/o separate encoders*-------
w/o task-dependent loss2.47.922.841.319.728.145.2
TimeExpert (Ours)2.58.223.642.820.329.646.9
Ablation Studies on No. of Input Frames
TimeExpert (8 frames)1.55.118.930.213.922.138.4
TimeExpert (64 frames)2.17.521.937.517.827.244.9
TimeExpert (128 frames)2.58.222.041.519.929.346.5
Ablation Studies on No. of Activated Experts
Vanilla MoE (k=2)2.38.122.842.119.829.245.8
Vanilla MoE (k=4)2.48.223.342.520.129.546.6
Vanilla MoE (k=6)2.58.223.542.820.229.646.9
Vanilla MoE (k=8)2.58.123.442.820.229.846.8
+ +Table 5. Ablation Study Results of TimeExpert. We evaluate the impact of removing key components, varying the number of input frames, and adjusting the number of activated experts. Results under the default setting are highlighted in gray. * denotes that directly processing time tokens and score tokens with the text tokenizer causes the model to fail to follow instructions. + +and decoding heads for causal event modeling completely disrupts the model's ability to follow instructions, reinforcing the necessity of structured task processing, and removing the auxiliary loss results in a slight performance drop across all tasks, particularly in Video Highlight Detection and Moment Retrieval. This indicates our proposed task-dependent auxiliary loss improves expert specialization by forcing task-aware routing, leading to better alignment between task tokens and relevant experts. + +Number of Input Frames. As shown in the middle of Table 5, increasing the number of frames generally improves performance across all tasks, as expected. Notably, the transition from 64 to 128 frames boosts performance in Moment Retrieval and Video Highlight Detection, indicating that a larger temporal context helps refine event understanding. + +Number of Activated Experts. As shown in the bottom part of Table 5, there is a clear trend where the predefined number of activated experts directly influences performance. Activating more experts improves performance, particularly in Video Highlight Detection (mAP +0.4 from 2 to 6 experts). Nonetheless, beyond 6 experts, performance saturates, showing that excessive expert activation may introduce redundancy without additional benefits. + +Training Data. We also conduct a data variant study, training our model on the same data recipe as TRACE to enable a direct comparison. As shown in the bottom line of Table 2, TimeExpert still outperforms TRACE across all VTG + +tasks. Compared to our full-data model, this variant exhibits slight fluctuations, with some metrics even improving (e.g., CIDEr and $\mathrm{R} @ 1_{\mathrm{IoU}=0.7}$ ), suggesting that TimeExpert generalizes well even with limited data. These findings highlight TimeExpert's efficiency and adaptability, proving that our model design inherently enhances VTG performance, independent of dataset scale. + +# 5. Conclusion + +In this work, we identify two critical limitations in existing VTG-specific Video-LLMs: inappropriate temporal event modeling and uniform task token processing. To address this, we propose TimeExpert, a novel framework that addresses these issues through a Mixture-of-Experts decoder that dynamically routes different task tokens (e.g., timestamps, saliency scores, and textual captions) to specialized experts, ensuring adaptive learning and mitigating task interference. Comprehensive experiments demonstrate Time-Expert's effectiveness across multiple VTG subtasks and general video understanding benchmarks, consistently outperforming existing VTG-specific Video-LLMs. + +# Acknowledgements + +This study is supported by the Ministry of Education Singapore, under the Tier-1 project scheme with the project number RT18/22. + +# References + +[1] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6 +[2] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 3, 5 +[3] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. Sharegpt4video: Improving video understanding and generation with better captions. In Advances in Neural Information Processing Systems, pages 19472-19495, 2024. 5 +[4] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. In Advances in Neural Information Processing Systems, pages 72842-72866, 2023. 3 +[5] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2, 3 +[6] Aidan Clark, Diego de Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. In International Conference on Machine Learning, pages 4057-4086. PMLR, 2022. 4 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6 +[8] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. 3 +[9] Dongyang Fan, Bettina Messmer, and Martin Jaggi. Towards an empirical understanding of moe design choices. arXiv preprint arXiv:2402.13089, 2024. 4 +[10] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 5 +[11] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24108-24118, 2025. 2 + +[12] Soichiro Fujita, Tsutomu Hirao, Hidetakamigaito, Manabu Okumura, and Masaaki Nagata. Soda: Story oriented dense video captioning evaluation framework. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 517-531. Springer, 2020. 6 +[13] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 3, 6, 7, 8 +[14] Yongxin Guo, Zhenglin Cheng, Xiaoying Tang, Zhaopeng Tu, and Tao Lin. Dynamic mixture of experts: An autotuning approach for efficient transformer models. In International Conference on Learning Representations, 2024. 3, 4 +[15] Yongxin Guo, Jingyu Liu, Mingda Li, Dingxin Cheng, Xiaoying Tang, Dianbo Sui, Qingbin Liu, Xi Chen, and Kevin Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. In Proceedings of the AAAI conference on artificial intelligence, 2025. 2, 3, 5, 6, 7 +[16] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Qingbin Liu, and Xi Chen. Trace: Temporal grounding video llm via causal event modeling. In International Conference on Learning Representations, 2025. 1, 2, 3, 6, 7 +[17] Michael Gygli, Helmut Grabner, Hayko Riemenschneider, and Luc Van Gool. Creating summaries from user videos. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13, pages 505-520. Springer, 2014. 3 +[18] Donghoon Han, Seunghyeon Seo, Eunhwan Park, Seong-Uk Nam, and Nojun Kwak. Unleash the potential of clip for video highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8275-8279, 2024. 2 +[19] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 2, 3, 6, 7 +[20] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 2, 3 +[21] Quzhe Huang, Zhenwei An, Nan Zhuang, Mingxu Tao, Chen Zhang, Yang Jin, Kun Xu, Liwei Chen, Songfang Huang, and Yansong Feng. Harder tasks need more experts: Dynamic routing in moe models. In Association for Computational Linguistics, 2024. 3, 4 +[22] Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi, and Seong Tae Kim. Do you remember? dense video captioning with cross-modal memory retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894-13904, 2024. 7 +[23] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-capitaling events in videos. In + +Proceedings of the IEEE international conference on computer vision, pages 706-715, 2017. 6, 7 +[24] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In Advances in Neural Information Processing Systems, pages 11846-11858, 2021. 1, 2, 3, 6, 8 +[25] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model. arXiv preprint arXiv:2410.05993, 2024. 6 +[26] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19948-19960, 2023. 3 +[27] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2, 3 +[28] Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. Unimoe: Scaling unified multimodal llms with mixture of experts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3 +[29] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Empirical Methods in Natural Language Processing, 2024. 2, 3 +[30] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified videolanguage temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 2 +[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 5 +[32] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3042-3051, 2022. 2, 3 +[33] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.5 +[34] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Association for Computational Linguistics, 2024. 5 +[35] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration for video temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 3 + +[36] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23023-23033, 2023. 3 +[37] Thong Thanh Nguyen, Zhiyuan Hu, Xiaobao Wu, Cong-Duy T Nguyen, See-Kiong Ng, and Anh Tuan Luu. Encoding and controlling global semantics for long-form video question answering. arXiv preprint arXiv:2405.19723, 2024. 5 +[38] Andreea-Maria Oncescu, Joao F Henriques, Yang Liu, Andrew Zisserman, and Samuel Albanie. Queryd: A video dataset with high-quality text and audio narrations. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2265-2269. IEEE, 2021. 3 +[39] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In International Conference on Machine Learning, 2024. 3, 5, 6, 7 +[40] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1, 2, 3, 6, 7 +[41] N Shazeer, A Mirhoseini, K Maziarz, A Davis, Q Le, G Hinton, and J Dean. The sparsely-gated mixture-of-experts layer. Outrageously large neural networks, 2017. 3 +[42] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. Tvsum: Summarizing web videos using titles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5179-5187, 2015. 3 +[43] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 3 +[44] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6 +[45] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 2 +[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3 +[47] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6847-6857, 2021. 7 + +[48] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 3, 5 +[49] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 2, 6 +[50] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. arXiv preprint arXiv:2405.09711, 2024. 5 +[51] Weijia Wu, Yuzhong Zhao, Zhuang Li, Jiahong Li, Hong Zhou, Mike Zheng Shou, and Xiang Bai. A large cross-modal video retrieval dataset with reading comprehension. Pattern Recognition, 157:110818, 2025. 5 +[52] Xun Wu, Shaohan Huang, Wenhui Wang, Shuming Ma, Li Dong, and Furu Wei. Multi-head mixture-of-experts. In Advances in Neural Information Processing Systems, pages 94073-94096, 2024. 5 +[53] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 5 +[54] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Empirical Methods in Natural Language Processing, 2021. 3 +[55] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 2, 3 +[56] An Yang, Junyang Lin, Rui Men, Chang Zhou, Le Jiang, Xianyan Jia, Ang Wang, Jie Zhang, Jiamang Wang, Yong Li, et al. M6-t: Exploring sparse expert models and beyond. arXiv preprint arXiv:2105.15082, 2021. 4 +[57] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 7 +[58] Yuanhang Yang, Shiyi Qi, Wenchao Gu, Chaozheng Wang, Cuiyun Gao, and Zenglin Xu. Xmoe: Sparse models with fine-grained and adaptive expert selection. In Findings of the Association for Computational Linguistics, 2024. 3 +[59] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23056-23065, 2023. 3 +[60] Yingsen Zeng, Yujie Zhong, Chengjian Feng, and Lin Ma. Unimd: Towards unifying moment retrieval and temporal action detection. In European Conference on Computer Vision, pages 286-304. Springer, 2024. 3 + +[61] Zihao Zeng, Yibo Miao, Hongcheng Gao, Hao Zhang, and Zhijie Deng. Adamoe: Token-adaptive routing with null experts for mixture-of-experts language models. In *Findings of the Association for Computational Linguistics*, 2024. 3 +[62] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 6 +[63] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Zwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 5 +[64] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1, 2, 3, 6, 7, 8 +[65] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 5 \ No newline at end of file diff --git a/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/images.zip b/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3ea84498607a2a0ef79da0c83c5aaedfc1913590 --- /dev/null +++ b/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c90337de8dd631e20ac64c073234bffd6bc65295d59bccd2108e091543453b09 +size 568998 diff --git a/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/layout.json b/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..079bdbca05b82dcee38a69de283c5e3793cefcce --- /dev/null +++ b/2025/TimeExpert_ An Expert-Guided Video LLM for Video Temporal Grounding/layout.json @@ -0,0 +1,8383 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 541, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 541, + 122 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 541, + 122 + ], + "type": "text", + "content": "TimeExpert: An Expert-Guided Video LLM for Video Temporal Grounding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 142, + 484, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 142, + 484, + 159 + ], + "spans": [ + { + "bbox": [ + 126, + 142, + 484, + 159 + ], + "type": "text", + "content": "Zuhao Yang" + }, + { + "bbox": [ + 126, + 142, + 484, + 159 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 126, + 142, + 484, + 159 + ], + "type": "text", + "content": " Yingchen Yu" + }, + { + "bbox": [ + 126, + 142, + 484, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 142, + 484, + 159 + ], + "type": "text", + "content": " Yunqing Zhao" + }, + { + "bbox": [ + 126, + 142, + 484, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 142, + 484, + 159 + ], + "type": "text", + "content": " Shijian Lu" + }, + { + "bbox": [ + 126, + 142, + 484, + 159 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 126, + 142, + 484, + 159 + ], + "type": "text", + "content": " Song Bai" + }, + { + "bbox": [ + 126, + 142, + 484, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 163, + 443, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 163, + 443, + 178 + ], + "spans": [ + { + "bbox": [ + 167, + 163, + 443, + 178 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 167, + 163, + 443, + 178 + ], + "type": "text", + "content": "Nanyang Technological University " + }, + { + "bbox": [ + 167, + 163, + 443, + 178 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 167, + 163, + 443, + 178 + ], + "type": "text", + "content": "ByteDance Inc." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 183, + 434, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 183, + 434, + 198 + ], + "spans": [ + { + "bbox": [ + 172, + 183, + 434, + 198 + ], + "type": "text", + "content": "https://mwxely.github.io/projects/yang2025time/index" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 212, + 425, + 398 + ], + "blocks": [ + { + "bbox": [ + 58, + 212, + 425, + 398 + ], + "lines": [ + { + "bbox": [ + 58, + 212, + 425, + 398 + ], + "spans": [ + { + "bbox": [ + 58, + 212, + 425, + 398 + ], + "type": "image", + "image_path": "b0c95e601c0873448604778cced1d04c13252a9987dc544923b584b875bf68ad.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 407, + 555, + 496 + ], + "lines": [ + { + "bbox": [ + 55, + 407, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 555, + 496 + ], + "type": "text", + "content": "Figure 1. Left: Video Temporal Grounding (VTG) is a fine-grained video understanding task that aims to accurately locate content along with event timestamps based on natural language queries. In this work, we mainly consider three major types of VTG tasks: (1) Moment Retrieval (MR), (2) Video Highlight Detection (VHD), and (3) Dense Video Captioning (DVC). The outputs of VTG often contain textual captions, timestamps, and saliency scores. Right: Unlike existing methods (e.g., TimeChat [40]) that employ a single static model, motivated by expert specialization on different task tokens [16], we propose TimeExpert, an expert-guided Video LLM with dynamic token routing. Through task-aware expert allocation, TimeExpert demonstrates substantial improvements over state-of-the-art Video-LLMs on several VTG benchmarks. For example, here we visualize zero-shot F1 score for DVC on the YouCook2 dataset [64], " + }, + { + "bbox": [ + 55, + 407, + 555, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{R} @ 1_{\\mathrm{IoU}=0.7}" + }, + { + "bbox": [ + 55, + 407, + 555, + 496 + ], + "type": "text", + "content": " for MR on the Charades-STA dataset [13], and HIT@1 for VHD on the QVHighlights dataset [24]. More results and analysis are in Section 4." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 433, + 247, + 555, + 396 + ], + "blocks": [ + { + "bbox": [ + 446, + 214, + 544, + 223 + ], + "lines": [ + { + "bbox": [ + 446, + 214, + 544, + 223 + ], + "spans": [ + { + "bbox": [ + 446, + 214, + 544, + 223 + ], + "type": "text", + "content": "Performance on VTG Tasks" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 224, + 548, + 246 + ], + "lines": [ + { + "bbox": [ + 432, + 224, + 548, + 246 + ], + "spans": [ + { + "bbox": [ + 432, + 224, + 548, + 246 + ], + "type": "text", + "content": "MR: Moment Retrieval \nDVC: Dense Video Captioning \nVHD: VideoHighlight Detection" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 433, + 247, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 433, + 247, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 433, + 247, + 555, + 396 + ], + "type": "image", + "image_path": "65e4838df396854c9b320b09a462517a187b3c3edd34b3b12e31370be77de8f5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 506, + 200, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 506, + 200, + 517 + ], + "spans": [ + { + "bbox": [ + 151, + 506, + 200, + 517 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 530, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 296, + 686 + ], + "type": "text", + "content": "Video Temporal Grounding (VTG) aims to precisely identify video event segments in response to textual queries. The outputs of VTG tasks manifest as sequences of events, each defined by precise timestamps, saliency scores, and textual descriptions. Despite recent advances, a fundamental limitation persists in existing Video Large Language Models (Video-LLMs): they process all task tokens through identical and static pathways, failing to recognize that temporal localization, saliency assessment, and textual generation represent fundamentally distinct tasks requiring specialized processing. To address this, we introduce TimeExpert, a Mixture-of-Experts (MoE)-based Video-LLM that effectively decomposes VTG tasks by dynamically routing task" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": "specific tokens (e.g., timestamps, saliency scores) to specialized experts, with increased computational efficiency. Our design choices enable precise handling of each subtask, leading to improved event modeling across diverse VTG applications. Extensive experiments demonstrate that TimeExpert consistently achieves state-of-the-art performance on various VTG tasks such as Dense Video Captioning, Moment Retrieval, and Video Highlight Detection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "Consider watching a cooking tutorial video and attempting to pinpoint the exact moment when the chef fries the bacon (Figure 1-Left) or other highlighted visual segments. This seemingly simple task reveals a fundamental challenge: while humans can effortlessly recognize these actions visually, determining their precise temporal boundaries—down to the specific second—remains significantly challenging" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 290, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 290, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 290, + 703 + ], + "type": "text", + "content": "*This work was done while Zuhao Yang was interning at ByteDance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 703, + 197, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 197, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 197, + 713 + ], + "type": "text", + "content": "Shijian Lu is the corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24286" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 77, + 197, + 171 + ], + "blocks": [ + { + "bbox": [ + 58, + 77, + 197, + 171 + ], + "lines": [ + { + "bbox": [ + 58, + 77, + 197, + 171 + ], + "spans": [ + { + "bbox": [ + 58, + 77, + 197, + 171 + ], + "type": "image", + "image_path": "006e0b32fae927e7b0f24cd9df7232c869e4204106b850a7fd37a768b032e965.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 175, + 182, + 187 + ], + "lines": [ + { + "bbox": [ + 70, + 175, + 182, + 187 + ], + "spans": [ + { + "bbox": [ + 70, + 175, + 182, + 187 + ], + "type": "text", + "content": "(a) VTG-Specific Video-LLM" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 230, + 76, + 360, + 167 + ], + "blocks": [ + { + "bbox": [ + 230, + 76, + 360, + 167 + ], + "lines": [ + { + "bbox": [ + 230, + 76, + 360, + 167 + ], + "spans": [ + { + "bbox": [ + 230, + 76, + 360, + 167 + ], + "type": "image", + "image_path": "8cdc35db096fdd413da3c0fe6607220a59d1f82790d345aaa9c47fb3ce2f4bfe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 175, + 326, + 186 + ], + "lines": [ + { + "bbox": [ + 262, + 175, + 326, + 186 + ], + "spans": [ + { + "bbox": [ + 262, + 175, + 326, + 186 + ], + "type": "text", + "content": "(b) Vanilla MoE" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 379, + 76, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 379, + 76, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 379, + 76, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 379, + 76, + 553, + 167 + ], + "type": "image", + "image_path": "9878de8d107a84d2898c11809cabda9ccc5d5f27da7d645a0aa179b8e3ae5d4d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 171, + 442, + 196 + ], + "lines": [ + { + "bbox": [ + 408, + 171, + 442, + 196 + ], + "spans": [ + { + "bbox": [ + 408, + 171, + 442, + 196 + ], + "type": "text", + "content": "removal new expert" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 446, + 175, + 534, + 188 + ], + "lines": [ + { + "bbox": [ + 446, + 175, + 534, + 188 + ], + "spans": [ + { + "bbox": [ + 446, + 175, + 534, + 188 + ], + "type": "text", + "content": "(c) TimeExpert (Ours)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 380, + 168, + 405, + 195 + ], + "blocks": [ + { + "bbox": [ + 380, + 168, + 405, + 195 + ], + "lines": [ + { + "bbox": [ + 380, + 168, + 405, + 195 + ], + "spans": [ + { + "bbox": [ + 380, + 168, + 405, + 195 + ], + "type": "image", + "image_path": "3c281f57eeeacc28a909a3388afc5e9152bb1f32410c4ef16693874e9b2bb086.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 204, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 54, + 204, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 204, + 555, + 270 + ], + "type": "text", + "content": "Figure 2. A Glimpse of Comparison across VTG Approaches. (a): VTG-specific Video-LLM [16, 40] relies on a single static model with shared parameters for all tasks, limiting their ability to specialize in diverse VTG subtasks. (b): Vanilla MoE improves upon this by activating a fixed set (e.g., " + }, + { + "bbox": [ + 54, + 204, + 555, + 270 + ], + "type": "inline_equation", + "content": "k = 2" + }, + { + "bbox": [ + 54, + 204, + 555, + 270 + ], + "type": "text", + "content": ") of experts, enabling some degree of task specialization. (c): Our TimeExpert takes a step further by implementing an adaptive routing mechanism that dynamically allocates new experts when tokens lack suitable matches and prunes unmatched experts when necessary. This dynamic approach significantly enhances computational efficiency while achieving superior task specialization, particularly for VTG subtasks that require distinct feature representations. Detailed design choices can be found at Section 3." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 292, + 294, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 292, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 54, + 292, + 294, + 364 + ], + "type": "text", + "content": "even for human observers. This problem forms the core motivation behind Video Temporal Grounding (VTG), which aims to develop intelligent systems capable of accurately locating relevant video segments based on natural language queries, enabling users to efficiently access desired content without manually searching through hours of footage." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 365, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 520 + ], + "type": "text", + "content": "Research Gaps. While traditional video grounding models excel in individual VTG subtasks like Moment Retrieval [18, 24] and Video Highlight Detection [32], their practicality is limited as they cannot handle multiple VTG subtasks concurrently and require extensive task-specific fine-tuning to achieve adequate performance. More recently, Video Large Language Models (VideoLLMs) [5, 27, 29, 46, 55] have made remarkable advancements in coarse-grained video content understanding (e.g., VideoQA [11, 27]), yet they struggle to extend such capabilities to tasks requiring fine-grained temporal localization (e.g., VTG [13, 64]), due to the lack of dedicated mechanisms for explicitly modeling temporal boundaries." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "Compared to generalist Video-LLMs, VTG-specific models that encode temporal tokens [15, 20, 45] or are trained on temporally-annotated instruction-tuning datasets [19, 40, 49] demonstrate marginal improvements, as they blend temporal information with text tokens without explicitly modeling timestamps and saliency scores, despite videos inherently possessing structured information beyond textual descriptions [16]. Furthermore, existing VTG-specific Video-LLMs rely exclusively on a single LLM as their reasoning backbone, processing all task tokens indiscriminately with shared parameters. This parameter-sharing paradigm inherently constrains the model's ability (Figure 1-Right) to specialize in the diverse subtasks of VTG—such as timestamp prediction, saliency estimation, and caption generation—each of which may demand distinct feature representations for optimal performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 292, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 553, + 495 + ], + "type": "text", + "content": "Proposed Method. To address the misalignment between language modeling of existing Video-LLMs and inherent video temporal structure, we introduce TimeExpert, an event-centric modeling framework that decouples the prediction of timestamps, saliency scores, and textual descriptions into distinct specialized tasks, organized in an interleaved sequence. To precisely model the varying token importance across these decoupled tasks, we replace the single LLM backbone of existing Video-LLMs with a fine-grained Mixture-of-Experts (MoE) decoder, with our proposed token-aware dynamic gating and token-adaptive routing strategies, which achieves both computational efficiency and modeling precision by activating only task-relevant parameters during both training and inference. To harness each expert's potential, we develop a task-dependent auxiliary loss that encourages more frequently activated experts to process a greater proportion of task-relevant tokens." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 641 + ], + "type": "text", + "content": "Our Contributions. (1): To our knowledge, TimeExpert is the first to explore the task token importance in VTG by introducing a MoE-based framework with dynamic expert routing. Our expert-guided architecture allows for adaptive learning across different VTG subtasks, effectively mitigating task interference while leveraging expert specialization. (2): We conduct extensive experiments across multiple VTG tasks on several benchmarks to demonstrate the effectiveness of TimeExpert. In particular, we demonstrate substantial improvements over state-of-the-art Video-LLMs through quantitative results and ablation studies, highlighting the advantages of expert-driven token specialization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 655, + 400, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 655, + 400, + 668 + ], + "spans": [ + { + "bbox": [ + 314, + 655, + 400, + 668 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Video Temporal Grounding. Video Temporal Grounding (VTG) aims to precisely localize relevant temporal boundaries in a video given textual queries [30]. It can be divided" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24287" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 222 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 222 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 222 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 222 + ], + "type": "image", + "image_path": "7530c8befb15395ae0e9f0a1e5586f316e0bc05cd32809cade8a001271f20f96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 230, + 296, + 308 + ], + "lines": [ + { + "bbox": [ + 55, + 230, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 296, + 308 + ], + "type": "text", + "content": "Figure 3. Architecture Overview of TimeExpert. Our model leverages independent encoders and decoding heads to process time, score, and text inputs and outputs. The timestamps and saliency scores of sampled frames are encoded into special tokens and integrated into the corresponding visual tokens. During inference, the generated response follows a structured format, sequentially incorporating time tokens, score tokens, and text tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 329, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 295, + 449 + ], + "type": "text", + "content": "into subtasks like Moment Retrieval [13, 38], Dense Video Captioning [2, 43, 64], Video Summarization [17, 42], and Video Highlight Detection [24, 32]. Traditional methods primarily rely on large-scale video-text pretraining, using training objectives such as video-text contrastive learning [48, 54], video-text matching [4, 26], and timestamp regression [35, 36, 59, 60]. Despite their initial success, these methods are constrained by their generalization ability to handle only a single task per model and require substantial fine-tuning for each new downstream application." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 700 + ], + "type": "text", + "content": "Video-LLMs for VTG Tasks. With the growing reasoning capabilities of Video-LLMs [5, 27, 29, 46, 55], numerous studies have investigated their adaptation to VTG tasks (Figure 2-(a)). For example, TimeChat [40], VTimeLLM [19], and HawkEye [20] fine-tune the Video-LLMs on refined temporal-annotated instruction datasets. Following that, several studies further enhance the Video-LLM's understanding of timestamps by introducing special tokens [15, 20] or a temporal perception module [39]. However, these methods adopt a purely text-to-text framework, formulating VTG as a natural language generation problem, which lacks explicit temporal modeling. Recently, TRACE [16] proposes a causal event modeling framework to capture the inherent structure of videos. Nevertheless, it still fails to incorporate task-specific properties, as they process all task tokens within a single shared reasoning backbone, leading to task interference and limited specialization. In contrast, TimeExpert simultaneously structures Video-LLM outputs into event representations and leverages adaptive expert specialization, enhancing both efficiency and accuracy across diverse VTG tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "LLMs with Mixture-of-Experts. The Mixture-of-Experts" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": "(MoE) [8, 28, 41] architecture (Figure 2-(b)) has been widely adopted to improve the efficiency of large-scale foundation models during both training and inference by replacing feed-forward network layers in transformer models with multiple specialized sub-networks, known as experts, which enables sparse parameter activation. The activated experts are selected by a gating network, which contain a pre-defined router. Prior methods often implement static top-" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 553, + 312 + ], + "type": "text", + "content": " routing across all tokens, several studies [14, 21, 58, 61] explored on dynamic routing mechanism that allow tokens to activate varying numbers of experts. However, these approaches either introduce additional hyperparameter complexity or demonstrate limited validation confined to text-to-text tasks, constraining their flexibility and cross-task generalizability. In contrast, our proposed TimeExpert (Figure 2-(c)) introduces dynamic expert routing tailored for Video-LLMs through adaptively adjusting the number of activated experts based on task token importance, ensuring effective expert specialization while eliminating reliance on rigid hyperparameter tuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 321, + 553, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 553, + 335 + ], + "type": "text", + "content": "3. TimeExpert: An Expert-guided Video LLM" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 342, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 474 + ], + "type": "text", + "content": "In this section, we firstly introduce some fundamental preliminaries in Section 3.1. Then, we propose TimeExpert (Figure 3), which integrates a dynamic expert selection mechanism with a structured event modeling framework. Our approach implements task-aware dynamic gating (Section 3.2) and token-adaptive routing (Section 3.3) to selectively activates specialized experts based on token relevance across diverse tasks. In Section 3.4, we discuss our proposed task-dependent auxiliary loss that optimizes expert utilization and thereby enhances the stability of MoE-based learning. In Section 3.5, we elaborate on our training recipe." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 479, + 400, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 400, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 400, + 491 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "text", + "content": "Problem Definition. Given a textual instruction " + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "text", + "content": " and a sequence of video frames " + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "text", + "content": " from a video, the objective of VTG task is to generate a structured event representation " + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "text", + "content": ". The output sequence " + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "text", + "content": " consists of discrete events " + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "inline_equation", + "content": "\\{\\epsilon_1,\\epsilon_2,\\dots ,\\epsilon_M\\}" + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "text", + "content": ", where each event " + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "inline_equation", + "content": "\\epsilon_{m}" + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "text", + "content": " encapsulates three key components: a timestamp " + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "inline_equation", + "content": "t_m" + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "text", + "content": ", a saliency score " + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "inline_equation", + "content": "s_m" + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "text", + "content": ", and a textual caption " + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "inline_equation", + "content": "c_{m}" + }, + { + "bbox": [ + 313, + 496, + 553, + 575 + ], + "type": "text", + "content": ". Formally, we have:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 322, + 580, + 553, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 580, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 322, + 580, + 553, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {R} = \\left\\{\\epsilon_ {1}, \\epsilon_ {2}, \\dots , \\epsilon_ {M} \\right\\} = \\left\\{\\left(t _ {m}, s _ {m}, c _ {m}\\right) \\mid 1 \\leq m \\leq M \\right\\}. \\tag {1}", + "image_path": "0f2d94a8b0cc5317b465a7fc5861d0a946e665977613dba9f167a627f839c2d0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 597, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 555, + 665 + ], + "type": "text", + "content": "Causal Event Modeling. Following [16], we formulate the event prediction process in a causal manner, leveraging the sequential dependencies between events. Specifically, given the past event sequence " + }, + { + "bbox": [ + 313, + 597, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_{1:m-1}" + }, + { + "bbox": [ + 313, + 597, + 555, + 665 + ], + "type": "text", + "content": ", the textual instruction " + }, + { + "bbox": [ + 313, + 597, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 597, + 555, + 665 + ], + "type": "text", + "content": ", and the video frames " + }, + { + "bbox": [ + 313, + 597, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 597, + 555, + 665 + ], + "type": "text", + "content": ", the probability distribution of the next event " + }, + { + "bbox": [ + 313, + 597, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\epsilon_m" + }, + { + "bbox": [ + 313, + 597, + 555, + 665 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 324, + 670, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 670, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 324, + 670, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {P} \\left(\\epsilon_ {m} \\mid \\epsilon_ {1: m - 1}, \\mathcal {I}, \\mathcal {F}\\right) = \\mathcal {P} \\left(t _ {m}, s _ {m}, c _ {m} \\mid \\epsilon_ {1: m - 1}, \\mathcal {I}, \\mathcal {F}\\right). \\tag {2}", + "image_path": "27ff356cb7491439c8ff49144791f9ab8868b5ca574606cefe813d728620b6f1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "This causal formulation ensures that event generation follows a structured temporal order, where each event is pro" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24288" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "gressively determined based on prior events, textual cues, and visual content." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 214, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 214, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 214, + 117 + ], + "type": "text", + "content": "3.2. Task-aware Dynamic Gating" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 120, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 296, + 180 + ], + "type": "text", + "content": "Vanilla MoEs. (Figure 2-(b)) employ the top-" + }, + { + "bbox": [ + 55, + 120, + 296, + 180 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 120, + 296, + 180 + ], + "type": "text", + "content": " gating strategy, which activates a fixed number of experts for each input token, determined by computing gating scores via a learned network " + }, + { + "bbox": [ + 55, + 120, + 296, + 180 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 120, + 296, + 180 + ], + "type": "text", + "content": ". Specifically, for a given token embedding " + }, + { + "bbox": [ + 55, + 120, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 120, + 296, + 180 + ], + "type": "text", + "content": ", the gating function follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 186, + 295, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 186, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 109, + 186, + 295, + 201 + ], + "type": "interline_equation", + "content": "g (\\mathbf {x}) \\in \\mathbb {R} ^ {K} := \\operatorname {s o f t m a x} \\left(\\mathbf {W} _ {g} ^ {T} \\mathbf {x}\\right), \\tag {3}", + "image_path": "36e2ddf68646ac51a3c7e0f5addb8d50ff39ae1dada6591ab9ec3cbb860278e1.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 209, + 296, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 246 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 209, + 296, + 246 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_g\\in \\mathbb{R}^{d\\times K}" + }, + { + "bbox": [ + 55, + 209, + 296, + 246 + ], + "type": "text", + "content": " represents the gating network parameters, and " + }, + { + "bbox": [ + 55, + 209, + 296, + 246 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 209, + 296, + 246 + ], + "type": "text", + "content": " denotes the total number of experts. Then, the output of this MoE layer is then formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 251, + 295, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 251, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 63, + 251, + 295, + 281 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\frac {1}{\\sum_ {e \\in \\operatorname {T o p} - k (g (\\mathbf {x}))} g (\\mathbf {x}) _ {e}} \\sum_ {e \\in \\operatorname {T o p} - k (g (\\mathbf {x}))} g (\\mathbf {x}) _ {e} \\mathbf {E} _ {e} (\\mathbf {x}), \\tag {4}", + "image_path": "00f1a4444ef98f71f9a87639fa9ec636af161861c1cc0f62bd13a00308b72446.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 289, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 295, + 314 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 289, + 295, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_e(\\mathbf{x})\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 289, + 295, + 314 + ], + "type": "text", + "content": " is the output of the " + }, + { + "bbox": [ + 55, + 289, + 295, + 314 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 289, + 295, + 314 + ], + "type": "text", + "content": "-th expert given input " + }, + { + "bbox": [ + 55, + 289, + 295, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 289, + 295, + 314 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 289, + 295, + 314 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x})_e" + }, + { + "bbox": [ + 55, + 289, + 295, + 314 + ], + "type": "text", + "content": " is the corresponding gating score." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 314, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 517 + ], + "type": "text", + "content": "Implicit Task Preference. Vanilla MoEs face some limitations, such as (1): the reliance on a predefined " + }, + { + "bbox": [ + 55, + 314, + 296, + 517 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 314, + 296, + 517 + ], + "type": "text", + "content": " for expert selection leads to suboptimal flexibility [6, 9, 56], (as task complexity varies, different tokens may require a varying number of experts based on their importance); and (2): the standard gating function treats all tokens uniformly, neglecting the inherent differences in task-specific significance, which may lead to inefficient expert routing and task interference [21]. Surprisingly, we observe an implicit task preference in these models on VTG tasks, as visualized in Figure 4 of expert assignments. In particular, we measure the ratio of activated task-specific tokens to the number of processed text tokens in that layer. Our findings reveal that certain experts, such as expert #7, consistently activate in response to specific subtasks (e.g., score tokens), even though the model was not explicitly trained to establish such specialization, implying an inherent task-driven preference." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 526, + 294, + 625 + ], + "blocks": [ + { + "bbox": [ + 58, + 526, + 294, + 625 + ], + "lines": [ + { + "bbox": [ + 58, + 526, + 294, + 625 + ], + "spans": [ + { + "bbox": [ + 58, + 526, + 294, + 625 + ], + "type": "image", + "image_path": "650963354869e4cee6fad7bb905e1f85f62ae301d731dd75b16a279bd9a5fc28.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 633, + 296, + 677 + ], + "lines": [ + { + "bbox": [ + 55, + 633, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 296, + 677 + ], + "type": "text", + "content": "Figure 4. Visualization of Expert Assignments on Various Task Tokens using Our Vanilla MoE Implementation. We take layer 4 as an example and only visualized the first 8 experts out of a total of 64, due to the space limitation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Explicit Task-Aware Gating. Based on this observation, we propose a task-aware dynamic gating mechanism that" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": "jointly considers token-level probabilities and task-specific activation statistics. Instead of applying a static threshold, we introduce a task-weighted gating function that adjusts expert activation dynamically by incorporating task token activation rates. Formally, given an expert representation matrix " + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_g\\in \\mathbb{R}^{d\\times K}" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": " and an input token " + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 72, + 554, + 145 + ], + "type": "text", + "content": ", we define:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 391, + 156, + 555, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 156, + 555, + 169 + ], + "spans": [ + { + "bbox": [ + 391, + 156, + 555, + 169 + ], + "type": "interline_equation", + "content": "s (\\mathbf {x}) = \\cos (\\mathbf {x}, \\mathbf {W} _ {g}), \\tag {5}", + "image_path": "14ac1ddf11067942d8e26dd2816d149a0cb1a2dd9a400ecb707ceb790708d722.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 343, + 173, + 555, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 173, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 343, + 173, + 555, + 200 + ], + "type": "interline_equation", + "content": "g (\\mathbf {x}) = \\operatorname {s i g n} \\left(\\sigma \\left(\\frac {s (\\mathbf {x}) + \\alpha A _ {t}}{1 + \\alpha}\\right) - \\sigma (\\mathbf {G})\\right), \\tag {6}", + "image_path": "ba6db8b67721cf2454a497f1cb45a0d2a10f86b80f98b7d1c65de3fdb29ff793.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " represents the historical activation rate of the task token type associated with " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " is a scaling coefficient that controls the influence of task importance; " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " denotes the sigmoid function for score normalization; " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{G} \\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " is a learnable threshold where a token routes to expert " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " only if its similarity score " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "s(\\mathbf{x}) \\in \\mathbb{R}^{K}" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " exceeds " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": ". Following [14], we make the binary sign() differentiable by copying the upstream gradient of " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": " to the pre-activation term " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\sigma(s(\\mathbf{x}, A_{t})) - \\sigma(\\mathbf{G})" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": ". With " + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "inline_equation", + "content": "A_{t}" + }, + { + "bbox": [ + 313, + 204, + 555, + 348 + ], + "type": "text", + "content": ", the model explicitly prioritizes experts that have demonstrated effectiveness for specific task tokens, encouraging specialization while ensuring under-utilized experts do not get excessively activated." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 355, + 451, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 451, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 451, + 369 + ], + "type": "text", + "content": "3.3. Token-adaptive Routing" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 372, + 554, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 503 + ], + "type": "text", + "content": "In this section, we introduce a token-adaptive routing mechanism (Figure 2-(c)) that dynamically facilitates expert addition and removal based on recorded activation trends. The routing mechanism comprises three key components, including (1): Task-specific Routing Recording: We continuously monitor expert utilization across different task tokens and the routing results during training. (2): Adaptive Expert Addition: New experts are allocated when certain task tokens persistently fail to activate appropriate experts, and (3): Redundant Expert Removal: Experts with persistently low activation rates are pruned to maintain efficiency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 504, + 554, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 554, + 599 + ], + "type": "text", + "content": "Task-specific Routing Recording. Instead of solely tracking which experts get activated, we maintain a per-task token activation frequency metric. Specifically, (1) For each expert " + }, + { + "bbox": [ + 313, + 504, + 554, + 599 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 313, + 504, + 554, + 599 + ], + "type": "text", + "content": ", we log its activation timestamps, denoted as " + }, + { + "bbox": [ + 313, + 504, + 554, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_E \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 504, + 554, + 599 + ], + "type": "text", + "content": "; (2) For input tokens that fail to activate any expert, we compute their aggregated embeddings as " + }, + { + "bbox": [ + 313, + 504, + 554, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_S \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 504, + 554, + 599 + ], + "type": "text", + "content": "; (3) We maintain an activation rate tracker " + }, + { + "bbox": [ + 313, + 504, + 554, + 599 + ], + "type": "inline_equation", + "content": "A_t" + }, + { + "bbox": [ + 313, + 504, + 554, + 599 + ], + "type": "text", + "content": " per task token type, which influences expert selection in subsequent iterations." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 600, + 554, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 554, + 635 + ], + "type": "text", + "content": "Adaptive Expert Addition. When a significant proportion of task tokens fail to activate any expert, we introduce new experts whose representations are initialized as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 362, + 643, + 555, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 643, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 362, + 643, + 555, + 669 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {g, K + 1} = \\frac {\\mathbf {R} _ {S}}{\\| \\mathbf {R} _ {S} \\|}, \\quad \\mathbf {G} _ {K + 1} = 0, \\tag {7}", + "image_path": "850f1d033b99510cb1bd3996668b29ceb0fa91574cbe6294a6257ad0b803f3ae.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_S" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " captures the average embedding of underrepresented task tokens. This ensures that newly added experts are aligned with the missing task representations." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24289" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 70, + 504, + 194 + ], + "blocks": [ + { + "bbox": [ + 108, + 70, + 504, + 194 + ], + "lines": [ + { + "bbox": [ + 108, + 70, + 504, + 194 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 504, + 194 + ], + "type": "table", + "html": "
Training StageDatasetsNo. of Samples
Stage 1: Task Module PretrainingValley [33], LLaVA-Image [31], TextVR [51], ShareGPT4Video [3], VTG-IT [15]1.9M
Stage 2: MoE Decoder PretrainingValley*, TextVR*, ShareGPT4Video*, VTG-IT*, ActivityNet Captions [2], VideoChatGPT [34], InternVid [48], Next-QA [53]0.9M
Stage 3: Supervised Fine-tuningFiltered and Re-annotated Data from: Previous Stages' Data, EgoQA [37], STAR [50], Moment-10M [39], LLaVA-Video-178K [63]2.3M
", + "image_path": "e8775313169e8bea0b1f58e66ac62ee987872d4949be55f1861da735b6460534.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "type": "text", + "content": "Table 1. Training Data Recipe of TimeExpert. These data are categorized into three training stages. * denotes we utilize the filtered subset of the original dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 245, + 296, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 269 + ], + "type": "text", + "content": "Redundant Expert Removal. Experts are pruned if their activation rate " + }, + { + "bbox": [ + 55, + 245, + 296, + 269 + ], + "type": "inline_equation", + "content": "A_{e}" + }, + { + "bbox": [ + 55, + 245, + 296, + 269 + ], + "type": "text", + "content": " remains below a threshold " + }, + { + "bbox": [ + 55, + 245, + 296, + 269 + ], + "type": "inline_equation", + "content": "\\tau_{\\mathrm{min}}" + }, + { + "bbox": [ + 55, + 245, + 296, + 269 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 274, + 295, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 274, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 120, + 274, + 295, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\text {r e m o v e}} = \\left\\{e \\mid A _ {e} < \\tau_ {\\min } \\right\\}. \\tag {8}", + "image_path": "4d6ca89168367254318c214678afc29223dddf8947ea8aabd78a1f346d2734b6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 291, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 327 + ], + "type": "text", + "content": "This prevents redundant experts from occupying parameter space while ensuring that experts specializing in low-frequency but important task tokens are retained." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 332, + 224, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 224, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 224, + 344 + ], + "type": "text", + "content": "3.4. Task-dependent Auxiliary Loss" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 350, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 409 + ], + "type": "text", + "content": "To enhance the efficiency and specialization of expert activation in TimeExpert, we introduce a task-dependent auxiliary loss, which ensures that experts with higher activation rates receive more relevant task tokens while preventing excessive expert overlap." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 410, + 296, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 493 + ], + "type": "text", + "content": "Existing auxiliary loss functions [10, 52] primarily focus on load balancing among experts, which may conflict with our goal of leveraging expert specialization for VTG tasks. Instead, our auxiliary loss promotes a dynamic allocation strategy by reinforcing the association between frequently activated experts and their corresponding task tokens while preventing under-utilization or excessive redundancy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 494, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 624 + ], + "type": "text", + "content": "Our auxiliary loss consists of two key components: task-aware concentration term (Left) and activation regularization term (Right). The task-aware concentration loss ensures that experts specialize in processing specific types of task tokens (e.g., timestamps, saliency scores, captions) by amplifying the assignment of relevant tokens to frequently activated experts. The activation regularization loss prevents the over-activation of a single expert. The two loss terms work synergistically to ensure a diverse and efficient allocation of task tokens. In summary, we define our auxiliary loss as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 71, + 628, + 295, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 628, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 71, + 628, + 295, + 651 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a u x}} = \\lambda_ {1} \\sum_ {e = 1} ^ {K} \\left(\\frac {A _ {e}}{\\sum_ {j = 1} ^ {K} A _ {j}} - \\frac {N _ {e}}{\\sum_ {j = 1} ^ {K} N _ {j}}\\right) ^ {2} + \\lambda_ {2} \\sum_ {e = 1} ^ {K} \\| \\mathbf {w} _ {g, e} \\| _ {2} ^ {2} \\tag {9}", + "image_path": "ba4cfb7914b7cfc71f4470105d3c0d4f3f7fe3eadb5c5f854fb077ff25996729.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "A_{e}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " represents the activation count of expert " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " over a batch, " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "N_{e}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " denotes the number of task tokens assigned to expert " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{g,e}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " is the representation vector of expert " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " are hyperparameters controlling the balance between task-aware specialization and numerical stability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 244, + 411, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 244, + 411, + 258 + ], + "spans": [ + { + "bbox": [ + 314, + 244, + 411, + 258 + ], + "type": "text", + "content": "3.5. Training Recipe" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 277, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 373 + ], + "type": "text", + "content": "Training Data. To ensure comprehensive learning across various video-language understanding tasks, we curate a diverse dataset spanning three progressive training stages, as summarized in Table 1. These utilized public datasets cover a wide range of temporal reasoning, video grounding, and captioning challenges. We conducted systematic processing, filtering, and re-annotating techniques that remove inaccurate temporal annotations, ensuring high-quality labels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 388, + 556, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 591 + ], + "type": "text", + "content": "Training Strategy. The training process of TimeExpert includes three stages: (1) Stage 1 - Task Module Pretraining. In the first stage, we pretrain the task-specific modules, including the vision compression layer, task encoder, and task heads, using 1.9M general multimodal video-text samples. This stage ensures that the model learns fundamental video representation capabilities before incorporating MoE decoder. (2) Stage 2 - MoE Decoder Pretraining. Before fine-tuning the full model, we introduce an intermediate MoE pretraining stage with 0.9M samples to ensure that expert routing aligns with different VTG-specific task tokens. This step mitigates the risk of expert collapse (i.e., all tokens being routed to the same expert) and enhances task-aware expert specialization. (3) Stage 3 - Supervised Fine-tuning. In the final stage, we fine-tune the entire Time-Expert framework on 2.3M samples, allowing both the task modules and the MoE decoder to be optimized jointly." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Training Objectives. During the first stage, we adopt cross-entropy loss for each generated token to teach basic video-event formatting. During MoE Decoder Pretraining, besides the cross-entropy loss, we apply " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "-loss [65] to stabilize training and proposed task-aware auxiliary loss to mitigate expert collapse and drive specialization. During Supervised Fine-tuning, we fine-tune the full model (except visual encoder) using same trio of objectives, ensuring stable routing while the model learns task-specific instructions." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24290" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 201 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 201 + ], + "type": "table", + "html": "
MethodNo. of \nActivated \nParametersDense Video Captioning \n(YouCook2 [64])Moment Retrieval \n(Charades-STA [13])Video Highlight Detection \n(QVHighlights [24])
SODAc(↑)CIDEr (↑)F1 Score (↑)R@IIoU=0.5 (↑)R@IIoU=0.7 (↑)mAP (↑)HIT@1 (↑)
TimeChat [40]7B1.23.412.632.213.414.523.9
VTimeLLM [19]7B---27.511.4--
Momentor [39]7B---26.611.67.6-
HawkEye [49]7B---31.414.5--
VTG-LLM [15]7B1.55.017.533.815.716.533.5
TRACE [16]7B2.28.122.440.319.426.842.7
TimeExpert (adaptive k)≈ 5.9B / 3.5B / 4.8B2.58.223.642.820.329.646.9
TimeExpert (TRACE's data)≈ 5.2B / 3.1B / 4.0B2.48.123.341.920.129.146.3
", + "image_path": "77cfdaac63444ff1214f8fca94fb9085ec8a9c94b3700a148adcae856c18affc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 555, + 243 + ], + "type": "text", + "content": "Table 2. Zero-shot Performance Comparison of TimeExpert against several state-of-the-art VTG-specific Video-LLMs on Dense Video Captioning, Temporal Grounding, and Video Highlight Detection. Some results are sourced from [16]. The best results are in bold. We also underline the second-best results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 262, + 136, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 136, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 136, + 276 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 178, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 178, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 178, + 296 + ], + "type": "text", + "content": "4.1. Experimental Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 302, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 295, + 493 + ], + "type": "text", + "content": "Implementation Details. We adopt ARIA [25] as the base MoE Video-LLM. Specifically, we utilize a lightweight visual encoder with 438M parameters, consisting of a Vision Transformer (ViT) [7] and a projection module. The ViT weights are initialized from [62]. Each video frame is initially encoded into 128 (for medium resolution) or 256 (for high resolution) visual tokens, which are then compressed via slot-based token compression [15] down to 8 visual tokens per frame. The time encoder and score encoder share the same architecture, both initialized with a tokenizer containing 11 number tokens for representing timestamps/scores, a separator token marking the end of each timestamp/score, and a switching token indicating task transitions. Specifically, following [16], we insert the token " + }, + { + "bbox": [ + 55, + 302, + 295, + 493 + ], + "type": "inline_equation", + "content": "\\langle sep\\rangle" + }, + { + "bbox": [ + 55, + 302, + 295, + 493 + ], + "type": "text", + "content": " between consecutive timestamps or scores and append " + }, + { + "bbox": [ + 55, + 302, + 295, + 493 + ], + "type": "inline_equation", + "content": "\\langle sync\\rangle" + }, + { + "bbox": [ + 55, + 302, + 295, + 493 + ], + "type": "text", + "content": " at the end of the sequence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 531 + ], + "type": "text", + "content": "Comparison Methods. We select TimeChat [40], VTimeLLM [19], Momentor [39], HawkEye [49], VTGLLM [15], and TRACE [16] as comparison methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 604 + ], + "type": "text", + "content": "Evaluation Benchmarks. To thoroughly evaluate TimeExpert in Video Temporal Grounding, we assess it across three VTG tasks: Dense Video Captioning, Moment Retrieval, and Video Highlight Detection, utilizing datasets such as YouCook2 [64], Charades-STA [13], QVHighlights [24], and ActivityNet Captions [23]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. For Dense Video Captioning, we use metrics including " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{SODA_c}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " [12], which is specifically tailored for the video's storyline; CIDEr score [44], which measures consensus between generated and reference captions by computing the TF-IDF weighted n-gram similarity across multiple reference captions; F1 Score, which quantifies the model's capacity on precisely locating timestamps; and METEOR [1], which combines uni-gram precision and recall into a harmonic mean while incorporating stemming," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 263, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 372 + ], + "type": "text", + "content": "synonym matching, and a fragmentation penalty so that longer contiguous matches are rewarded. For Moment Retrieval, we compute the Intersection over Union (IoU) between the timestamps predicted by the model and ground truth, including Recall at IoU thresholds of " + }, + { + "bbox": [ + 313, + 263, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\{0.5, 0.7\\}" + }, + { + "bbox": [ + 313, + 263, + 555, + 372 + ], + "type": "text", + "content": " and mean IoU. For Video Highlight Detection, we calculate both the mean Average Precision (mAP) at IoU thresholds of " + }, + { + "bbox": [ + 313, + 263, + 555, + 372 + ], + "type": "inline_equation", + "content": "\\{0.5, 0.75\\}" + }, + { + "bbox": [ + 313, + 263, + 555, + 372 + ], + "type": "text", + "content": ", and HIT@1, which represents the hit ratio of the highest scored clip." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 382, + 432, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 432, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 432, + 395 + ], + "type": "text", + "content": "4.2.Quantitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "text", + "content": "Zero-shot Performance on VTG Tasks. In Table 2, we present the zero-shot performance of TimeExpert against several state-of-the-art baselines on three widely adopted VTG benchmarks. We report the number of activated parameters for a transparent comparison. Across the three benchmarks, the adaptive-" + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "text", + "content": " variant activated an average of 14.5, 9.1, and 11.8 experts, whereas the TRACE-trained variant averaged 13.5, 8.3, and 10.5. Notably, our approach outperforms other VTG-specific Video-LLMs across all three benchmarks with fewer activated parameters. It achieves a " + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "text", + "content": " performance gain on YouCook2 [64] using the " + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "inline_equation", + "content": "\\mathrm{SODA_c}" + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "text", + "content": " and F1 Score metrics, respectively; a " + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "text", + "content": " performance boost in Recall with IoU = {0.5, 0.7} thresholds on Charades-STA [13]; and a " + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 313, + 401, + 555, + 616 + ], + "type": "text", + "content": " performance gain for the mAP and HIT@1 metrics on QVHighlights [24]. This validates that dynamic gating and adaptive routing not only ensures efficient expert utilization but also fully leverages task specialization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Fine-tuned Performance on VTG Tasks. In Table 3, we present the results of fine-tuning TimeExpert on YouCook2 and Charades-STA datasets for 2 epochs. The results indicate that: (1) TimeExpert surpasses both generalist and task-specific baselines. Unlike TimeChat and VTG-LLM, which struggle to fully benefit from fine-tuning, our MoE-based approach effectively adapts to downstream datasets, achieving new state-of-the-art results. Notably, it surpasses" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24291" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 70, + 511, + 272 + ], + "blocks": [ + { + "bbox": [ + 100, + 70, + 511, + 272 + ], + "lines": [ + { + "bbox": [ + 100, + 70, + 511, + 272 + ], + "spans": [ + { + "bbox": [ + 100, + 70, + 511, + 272 + ], + "type": "table", + "html": "
MethodDense Video Captioning\n(YouCook2 [64])Moment Retrieval\n(Charades-STA [13])
SODAc(↑)CIDEr(↑)F1 Score(↑)R@1IoU=0.5(↑)R@1IoU=0.7(↑)
Traditional Models
PDVC [47]4.422.7---
Vid2Seq [57]5.725.323.5--
Vid2Seq (Audio Input) [57]7.947.127.3--
CM² [22]5.331.728.4--
VTG-specific Video-LLMs
TimeChat [40]3.411.019.546.723.7
VTG-LLM [15]3.613.420.657.233.4
TRACE [16]6.735.531.861.741.4
TimeExpert (Ours)7.239.033.564.143.3
", + "image_path": "9d19316759034114927ed468b880ae21ae84a0b0d7ddf9b8ab0ba8e34c523ecb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 312, + 549, + 437 + ], + "blocks": [ + { + "bbox": [ + 55, + 278, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 278, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 555, + 301 + ], + "type": "text", + "content": "Table 3. Fine-tuned Performance of TimeExpert. Some results are from [16]. We report the performance of each model after fine-tuning on YouCook2 and Charades-STA datasets. We bold the best scores and italicize results that benefit from additional audio inputs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 312, + 549, + 437 + ], + "lines": [ + { + "bbox": [ + 62, + 312, + 549, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 312, + 549, + 437 + ], + "type": "table", + "html": "
MethodDense Video CaptioningMoment Retrieval
METEOR (↑)SODAc (↑)CIDEr (↑)F1 Score (↑)R@1IoU=0.5 (↑)R@1IoU=0.7 (↑)mIoU (↑)
TimeChat [40]5.74.719.036.94.62.06.9
VTimeLLM [19]6.85.827.6-29.514.231.4
Momentor* [39]4.72.314.9-23.012.429.3
VTG-LLM [15]5.95.120.734.88.33.712.0
TRACE [16]6.46.025.939.337.724.039.0
TimeExpert (Ours)7.06.528.440.539.226.141.5
", + "image_path": "69f6a8d62dfd11c2a1587077786169be7087561167b8926a27e274432f2d6b15.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 443, + 555, + 468 + ], + "lines": [ + { + "bbox": [ + 55, + 443, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 555, + 468 + ], + "type": "text", + "content": "Table 4. Performance Comparison on ActivityNet Captions [23]. Some results are sourced from [16]. * indicates zero-shot evaluation. The best results are in bold. TimeExpert achieves state-of-the-art results across both Dense Video Captioning and Moment Retrieval." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 487, + 297, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 297, + 656 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 297, + 656 + ], + "type": "text", + "content": "TRACE in both Moment Retrieval and Dense Video Captioning tasks, highlighting the benefits of dynamic expert activation in video-language modeling. (2) TimeExpert enhances retrieval and captioning tasks simultaneously. Our model achieves a " + }, + { + "bbox": [ + 54, + 487, + 297, + 656 + ], + "type": "inline_equation", + "content": "+3.5" + }, + { + "bbox": [ + 54, + 487, + 297, + 656 + ], + "type": "text", + "content": " improvement in CIDEr and a " + }, + { + "bbox": [ + 54, + 487, + 297, + 656 + ], + "type": "inline_equation", + "content": "+1.7" + }, + { + "bbox": [ + 54, + 487, + 297, + 656 + ], + "type": "text", + "content": " boost in F1 Score on YouCook2, alongside an absolute gain of 2.4 in " + }, + { + "bbox": [ + 54, + 487, + 297, + 656 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\mathrm{1}_{\\mathrm{IoU}} = 0.5" + }, + { + "bbox": [ + 54, + 487, + 297, + 656 + ], + "type": "text", + "content": " and 1.9 in " + }, + { + "bbox": [ + 54, + 487, + 297, + 656 + ], + "type": "inline_equation", + "content": "\\mathrm{R}@\\mathrm{1}_{\\mathrm{IoU}} = 0.7" + }, + { + "bbox": [ + 54, + 487, + 297, + 656 + ], + "type": "text", + "content": " on Charades-STA. These results demonstrate that TimeExpert's token-adaptive routing effectively captures event structures, enabling improved video reasoning. (3) Efficient expert utilization enables balanced trade-offs. Unlike static models, TimeExpert dynamically activates the most relevant experts per input, resulting in stronger generalization to unseen datasets while keeping computation manageable." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": "Performance Comparison on ActivityNet Captions. In Table 4, we evaluate TimeExpert on the ActivityNet Captions [23] dataset. Notably, TimeExpert achieves the best performance across both Moment Retrieval and Dense" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 487, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 631 + ], + "type": "text", + "content": "Video Captioning tasks. It surpasses TRACE, which was previously the strongest baseline, by improving CIDER (+2.5) and F1 Score (+1.2) in Dense Video Captioning, while also significantly boosting R@ " + }, + { + "bbox": [ + 313, + 487, + 555, + 631 + ], + "type": "inline_equation", + "content": "1_{\\mathrm{IoU}} = 0.5" + }, + { + "bbox": [ + 313, + 487, + 555, + 631 + ], + "type": "text", + "content": " (+1.5) and mIoU (+2.5) in Moment Retrieval. These results highlight TimeExpert's ability to effectively model long-range temporal dependencies while efficiently leveraging a MoE framework for adaptive token routing. Unlike previous methods, which rely on static expert activation, our model dynamically selects experts based on token-wise information, leading to improved alignment between textual queries and visual events." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 646, + 416, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 416, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 416, + 658 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Design Choices. As shown in the upper part of Table 5, the removal of token-adaptive routing leads to a notable drop in all metrics, confirming its efficacy in efficiently allocating task tokens. Moreover, removing separate encoders" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24292" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 555, + 323 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 323 + ], + "type": "table", + "html": "
MethodDense Video Captioning\n(YouCook2 [64])Moment Retrieval\n(Charades-STA [13])Video Highlight Detection\n(QVHighlights [24])
SODAc(↑)CIDEr (↑)F1 Score (↑)R@1IoU=0.5 (↑)R@1IoU=0.7 (↑)mAP (↑)HIT@1 (↑)
Ablation Studies on Design Choices
w/o token-adaptive routing2.17.622.240.519.226.542.6
w/o separate encoders*-------
w/o task-dependent loss2.47.922.841.319.728.145.2
TimeExpert (Ours)2.58.223.642.820.329.646.9
Ablation Studies on No. of Input Frames
TimeExpert (8 frames)1.55.118.930.213.922.138.4
TimeExpert (64 frames)2.17.521.937.517.827.244.9
TimeExpert (128 frames)2.58.222.041.519.929.346.5
Ablation Studies on No. of Activated Experts
Vanilla MoE (k=2)2.38.122.842.119.829.245.8
Vanilla MoE (k=4)2.48.223.342.520.129.546.6
Vanilla MoE (k=6)2.58.223.542.820.229.646.9
Vanilla MoE (k=8)2.58.123.442.820.229.846.8
", + "image_path": "3058beabc216fcaa7e62519dec58f05123fb06c4c28b6b612bf7bbc3d3b6d157.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 331, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 555, + 366 + ], + "type": "text", + "content": "Table 5. Ablation Study Results of TimeExpert. We evaluate the impact of removing key components, varying the number of input frames, and adjusting the number of activated experts. Results under the default setting are highlighted in gray. * denotes that directly processing time tokens and score tokens with the text tokenizer causes the model to fail to follow instructions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 386, + 294, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 386, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 386, + 294, + 495 + ], + "type": "text", + "content": "and decoding heads for causal event modeling completely disrupts the model's ability to follow instructions, reinforcing the necessity of structured task processing, and removing the auxiliary loss results in a slight performance drop across all tasks, particularly in Video Highlight Detection and Moment Retrieval. This indicates our proposed task-dependent auxiliary loss improves expert specialization by forcing task-aware routing, leading to better alignment between task tokens and relevant experts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": "Number of Input Frames. As shown in the middle of Table 5, increasing the number of frames generally improves performance across all tasks, as expected. Notably, the transition from 64 to 128 frames boosts performance in Moment Retrieval and Video Highlight Detection, indicating that a larger temporal context helps refine event understanding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 568, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 664 + ], + "type": "text", + "content": "Number of Activated Experts. As shown in the bottom part of Table 5, there is a clear trend where the predefined number of activated experts directly influences performance. Activating more experts improves performance, particularly in Video Highlight Detection (mAP +0.4 from 2 to 6 experts). Nonetheless, beyond 6 experts, performance saturates, showing that excessive expert activation may introduce redundancy without additional benefits." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Training Data. We also conduct a data variant study, training our model on the same data recipe as TRACE to enable a direct comparison. As shown in the bottom line of Table 2, TimeExpert still outperforms TRACE across all VTG" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 386, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 471 + ], + "type": "text", + "content": "tasks. Compared to our full-data model, this variant exhibits slight fluctuations, with some metrics even improving (e.g., CIDEr and " + }, + { + "bbox": [ + 313, + 386, + 555, + 471 + ], + "type": "inline_equation", + "content": "\\mathrm{R} @ 1_{\\mathrm{IoU}=0.7}" + }, + { + "bbox": [ + 313, + 386, + 555, + 471 + ], + "type": "text", + "content": "), suggesting that TimeExpert generalizes well even with limited data. These findings highlight TimeExpert's efficiency and adaptability, proving that our model design inherently enhances VTG performance, independent of dataset scale." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 481, + 388, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 481, + 388, + 495 + ], + "spans": [ + { + "bbox": [ + 314, + 481, + 388, + 495 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 647 + ], + "type": "text", + "content": "In this work, we identify two critical limitations in existing VTG-specific Video-LLMs: inappropriate temporal event modeling and uniform task token processing. To address this, we propose TimeExpert, a novel framework that addresses these issues through a Mixture-of-Experts decoder that dynamically routes different task tokens (e.g., timestamps, saliency scores, and textual captions) to specialized experts, ensuring adaptive learning and mitigating task interference. Comprehensive experiments demonstrate Time-Expert's effectiveness across multiple VTG subtasks and general video understanding benchmarks, consistently outperforming existing VTG-specific Video-LLMs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 657, + 417, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 657, + 417, + 671 + ], + "spans": [ + { + "bbox": [ + 314, + 657, + 417, + 671 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "This study is supported by the Ministry of Education Singapore, under the Tier-1 project scheme with the project number RT18/22." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24293" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 711 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 205, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 295, + 270 + ], + "type": "text", + "content": "[3] Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. Sharegpt4video: Improving video understanding and generation with better captions. In Advances in Neural Information Processing Systems, pages 19472-19495, 2024. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 272, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 272, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 61, + 272, + 295, + 327 + ], + "type": "text", + "content": "[4] Sihan Chen, Handong Li, Qunbo Wang, Zijia Zhao, Mingzhen Sun, Xinxin Zhu, and Jing Liu. Vast: A vision-audio-subtitle-text omni-modality foundation model and dataset. In Advances in Neural Information Processing Systems, pages 72842-72866, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 328, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 295, + 384 + ], + "type": "text", + "content": "[5] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 385, + 295, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 385, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 62, + 385, + 295, + 450 + ], + "type": "text", + "content": "[6] Aidan Clark, Diego de Las Casas, Aurelia Guy, Arthur Mensch, Michela Paganini, Jordan Hoffmann, Bogdan Damoc, Blake Hechtman, Trevor Cai, Sebastian Borgeaud, et al. Unified scaling laws for routed language models. In International Conference on Machine Learning, pages 4057-4086. PMLR, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 453, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 295, + 518 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 520, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 520, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 295, + 554 + ], + "type": "text", + "content": "[8] David Eigen, Marc'Aurelio Ranzato, and Ilya Sutskever. Learning factored representations in a deep mixture of experts. arXiv preprint arXiv:1312.4314, 2013. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 555, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 295, + 589 + ], + "type": "text", + "content": "[9] Dongyang Fan, Bettina Messmer, and Martin Jaggi. Towards an empirical understanding of moe design choices. arXiv preprint arXiv:2402.13089, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 590, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 634 + ], + "type": "text", + "content": "[10] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. Journal of Machine Learning Research, 23(120):1-39, 2022. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 636, + 295, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 711 + ], + "type": "text", + "content": "[11] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24108-24118, 2025. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "text", + "content": "[12] Soichiro Fujita, Tsutomu Hirao, Hidetakamigaito, Manabu Okumura, and Masaaki Nagata. Soda: Story oriented dense video captioning evaluation framework. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 517-531. Springer, 2020. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "text", + "content": "[13] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "text", + "content": "[14] Yongxin Guo, Zhenglin Cheng, Xiaoying Tang, Zhaopeng Tu, and Tao Lin. Dynamic mixture of experts: An autotuning approach for efficient transformer models. In International Conference on Learning Representations, 2024. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 242, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 306 + ], + "type": "text", + "content": "[15] Yongxin Guo, Jingyu Liu, Mingda Li, Dingxin Cheng, Xiaoying Tang, Dianbo Sui, Qingbin Liu, Xi Chen, and Kevin Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. In Proceedings of the AAAI conference on artificial intelligence, 2025. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 308, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 353 + ], + "type": "text", + "content": "[16] Yongxin Guo, Jingyu Liu, Mingda Li, Xiaoying Tang, Qingbin Liu, and Xi Chen. Trace: Temporal grounding video llm via causal event modeling. In International Conference on Learning Representations, 2025. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 354, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 409 + ], + "type": "text", + "content": "[17] Michael Gygli, Helmut Grabner, Hayko Riemenschneider, and Luc Van Gool. Creating summaries from user videos. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part VII 13, pages 505-520. Springer, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 464 + ], + "type": "text", + "content": "[18] Donghoon Han, Seunghyeon Seo, Eunhwan Park, Seong-Uk Nam, and Nojun Kwak. Unleash the potential of clip for video highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8275-8279, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 467, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 520 + ], + "type": "text", + "content": "[19] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14271-14280, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 522, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 553, + 577 + ], + "type": "text", + "content": "[20] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In European Conference on Computer Vision, pages 202-218. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 578, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 633 + ], + "type": "text", + "content": "[21] Quzhe Huang, Zhenwei An, Nan Zhuang, Mingxu Tao, Chen Zhang, Yang Jin, Kun Xu, Liwei Chen, Songfang Huang, and Yansong Feng. Harder tasks need more experts: Dynamic routing in moe models. In Association for Computational Linguistics, 2024. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[22] Minkuk Kim, Hyeon Bae Kim, Jinyoung Moon, Jinwoo Choi, and Seong Tae Kim. Do you remember? dense video captioning with cross-modal memory retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13894-13904, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[23] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-capitaling events in videos. In" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24294" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE international conference on computer vision, pages 706-715, 2017. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[24] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. In Advances in Neural Information Processing Systems, pages 11846-11858, 2021. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[25] Dongxu Li, Yudong Liu, Haoning Wu, Yue Wang, Zhiqi Shen, Bowen Qu, Xinyao Niu, Guoyin Wang, Bei Chen, and Junnan Li. Aria: An open multimodal native mixture-of-experts model. arXiv preprint arXiv:2410.05993, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 241 + ], + "type": "text", + "content": "[26] Kunchang Li, Yali Wang, Yizhuo Li, Yi Wang, Yinan He, Limin Wang, and Yu Qiao. Unmasked teacher: Towards training-efficient video foundation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19948-19960, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 308 + ], + "type": "text", + "content": "[27] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mybench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 364 + ], + "type": "text", + "content": "[28] Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. Unimoe: Scaling unified multimodal llms with mixture of experts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 409 + ], + "type": "text", + "content": "[29] Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Empirical Methods in Natural Language Processing, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 476 + ], + "type": "text", + "content": "[30] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified videolanguage temporal grounding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2794-2804, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 511 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36:34892-34916, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 567 + ], + "type": "text", + "content": "[32] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3042-3051, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 621 + ], + "type": "text", + "content": "[33] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "text", + "content": "[34] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In Association for Computational Linguistics, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[35] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration for video temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[36] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23023-23033, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[37] Thong Thanh Nguyen, Zhiyuan Hu, Xiaobao Wu, Cong-Duy T Nguyen, See-Kiong Ng, and Anh Tuan Luu. Encoding and controlling global semantics for long-form video question answering. arXiv preprint arXiv:2405.19723, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 183, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 553, + 249 + ], + "type": "text", + "content": "[38] Andreea-Maria Oncescu, Joao F Henriques, Yang Liu, Andrew Zisserman, and Samuel Albanie. Queryd: A video dataset with high-quality text and audio narrations. In ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 2265-2269. IEEE, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[39] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In International Conference on Machine Learning, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "type": "text", + "content": "[40] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313-14323, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 361, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 393 + ], + "type": "text", + "content": "[41] N Shazeer, A Mirhoseini, K Maziarz, A Davis, Q Le, G Hinton, and J Dean. The sparsely-gated mixture-of-experts layer. Outrageously large neural networks, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 437 + ], + "type": "text", + "content": "[42] Yale Song, Jordi Vallmitjana, Amanda Stent, and Alejandro Jaimes. Tvsum: Summarizing web videos using titles. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5179-5187, 2015. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 438, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 502 + ], + "type": "text", + "content": "[43] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1207-1216, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 548 + ], + "type": "text", + "content": "[44] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 603 + ], + "type": "text", + "content": "[45] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 657 + ], + "type": "text", + "content": "[46] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[47] Teng Wang, Ruimao Zhang, Zhichao Lu, Feng Zheng, Ran Cheng, and Ping Luo. End-to-end dense video captioning with parallel decoding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6847-6857, 2021. 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24295" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[48] Yi Wang, Kunchang Li, Yizhuo Li, Yinan He, Bingkun Huang, Zhiyu Zhao, Hongjie Zhang, Jilan Xu, Yi Liu, Zun Wang, et al. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191, 2022. 3, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "text", + "content": "[49] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 215 + ], + "type": "text", + "content": "[50] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. arXiv preprint arXiv:2405.09711, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 216, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 216, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 216, + 294, + 260 + ], + "type": "text", + "content": "[51] Weijia Wu, Yuzhong Zhao, Zhuang Li, Jiahong Li, Hong Zhou, Mike Zheng Shou, and Xiang Bai. A large cross-modal video retrieval dataset with reading comprehension. Pattern Recognition, 157:110818, 2025. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 304 + ], + "type": "text", + "content": "[52] Xun Wu, Shaohan Huang, Wenhui Wang, Shuming Ma, Li Dong, and Furu Wei. Multi-head mixture-of-experts. In Advances in Neural Information Processing Systems, pages 94073-94096, 2024. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "type": "text", + "content": "[53] Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9777–9786, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "text", + "content": "[54] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. In Empirical Methods in Natural Language Processing, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 459 + ], + "type": "text", + "content": "[55] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 459, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 294, + 503 + ], + "type": "text", + "content": "[56] An Yang, Junyang Lin, Rui Men, Chang Zhou, Le Jiang, Xianyan Jia, Ang Wang, Jie Zhang, Jiamang Wang, Yong Li, et al. M6-t: Exploring sparse expert models and beyond. arXiv preprint arXiv:2105.15082, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 569 + ], + "type": "text", + "content": "[57] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[58] Yuanhang Yang, Shiyi Qi, Wenchao Gu, Chaozheng Wang, Cuiyun Gao, and Zenglin Xu. Xmoe: Sparse models with fine-grained and adaptive expert selection. In Findings of the Association for Computational Linguistics, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 669 + ], + "type": "text", + "content": "[59] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23056-23065, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[60] Yingsen Zeng, Yujie Zhong, Chengjian Feng, and Lin Ma. Unimd: Towards unifying moment retrieval and temporal action detection. In European Conference on Computer Vision, pages 286-304. Springer, 2024. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 285 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 117 + ], + "type": "text", + "content": "[61] Zihao Zeng, Yibo Miao, Hongcheng Gao, Hao Zhang, and Zhijie Deng. Adamoe: Token-adaptive routing with null experts for mixture-of-experts language models. In *Findings of the Association for Computational Linguistics*, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 118, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 555, + 162 + ], + "type": "text", + "content": "[62] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11975-11986, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 163, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 163, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 555, + 196 + ], + "type": "text", + "content": "[63] Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Zwei Liu, and Chunyuan Li. Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713, 2024. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 196, + 555, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 196, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 315, + 196, + 555, + 240 + ], + "type": "text", + "content": "[64] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In Proceedings of the AAAI conference on artificial intelligence, 2018. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 241, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 241, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 315, + 241, + 555, + 285 + ], + "type": "text", + "content": "[65] Barret Zoph, Irwan Bello, Sameer Kumar, Nan Du, Yanping Huang, Jeff Dean, Noam Shazeer, and William Fedus. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906, 2022. 5" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24296" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/8cdc06bd-0ca5-4f6b-b012-50c8329288a4_content_list.json b/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/8cdc06bd-0ca5-4f6b-b012-50c8329288a4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fce2c8fa42a396ed5dc2014e86e0da03ea0feccc --- /dev/null +++ b/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/8cdc06bd-0ca5-4f6b-b012-50c8329288a4_content_list.json @@ -0,0 +1,2434 @@ +[ + { + "type": "text", + "text": "TimeFormer: Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction", + "text_level": 1, + "bbox": [ + 217, + 130, + 779, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dadong Jiang $^{1,3}$", + "bbox": [ + 184, + 202, + 316, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "patrickddj7@gmail.com", + "bbox": [ + 156, + 224, + 346, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xianghui Yang3", + "bbox": [ + 207, + 244, + 336, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yangxhui6@163.com", + "bbox": [ + 194, + 267, + 348, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhi Hou $^{2\\dagger}$", + "bbox": [ + 464, + 202, + 547, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhou9878@uni.sydney.edu.au", + "bbox": [ + 390, + 224, + 622, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaobo Zhou $^{1,4\\dagger}$", + "bbox": [ + 431, + 244, + 563, + 262 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xiaobo.zhou@tju.edu.cn", + "bbox": [ + 400, + 267, + 596, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhihui Ke", + "bbox": [ + 707, + 203, + 797, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kezhihui@tju.edu.cn", + "bbox": [ + 668, + 224, + 838, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tie Qiu $^{1,4}$", + "bbox": [ + 684, + 244, + 764, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "qiutie@tju.edu.cn", + "bbox": [ + 648, + 267, + 800, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tianjin University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ Tencent $^{4}$ Qinghai Minzu University", + "bbox": [ + 99, + 287, + 916, + 308 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2764ac663fe4ffba03aad8528a90d4a3b8562e60ca0658137ab81572ea2fec83.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 313, + 248, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/069a1eec7ac7d8dc1fb2244ca72e46900c11e67c65fedbbb6fdecc396a1eb5f7.jpg", + "image_caption": [ + "DeformGS PSNR:27.53" + ], + "image_footnote": [], + "bbox": [ + 94, + 503, + 169, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f757022999d07ed5ad3b24825151ebc029c4d1d53c41ab0783ceed1d888ed3fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 313, + 406, + 502 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8ad2ff566c29fd88f2f99299e3c47cbcf7c8094e14821d10e80efcfd52858d5c.jpg", + "image_caption": [ + "+TimeFormer PSNR:29.34" + ], + "image_footnote": [], + "bbox": [ + 251, + 503, + 330, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c8c76442a8f177dd0ef48b19245ad93ecaa72146f58c98faf504103471b9028a.jpg", + "image_caption": [ + "STGS PSNR: 27.28" + ], + "image_footnote": [], + "bbox": [ + 411, + 314, + 643, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a376af876f0b331f7d044f378acca83cf4c7f49be271faea76baad2b69927723.jpg", + "image_caption": [ + "+TimeFormer PSNR:27.98" + ], + "image_footnote": [], + "bbox": [ + 411, + 441, + 643, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1d41cf214f1312ac1c6c3efbd825eb5c769d445f030b0796e16d415089cf95b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 313, + 727, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d847fe0f9b12247ba1025d6a94de0001990026997c34b8c37dff12c15020136c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 375, + 725, + 435 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e26fe2e99b6de71585fa2f2cae4312fa65c814413e0920ea1fc2ee083618126e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 441, + 727, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1278efc39127dce26a8d1b72c57039e1bca6dd6bb13eef32ba6eebee05293336.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 503, + 725, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a530fc711584fb47472638fd78cea9077aebeefe035735f96b25c386ac02dfd.jpg", + "image_caption": [ + "Figure 1. We propose TimeFormer, a Transformer module that implicitly models the motion pattern via Temporal Attention from a learning perspective (right). TimeFormer is plug-and-play to existing deformable 3D Gaussian reconstruction methods [27, 60, 66] and enhances reconstruction results (left and middle) without affecting their original inference speed." + ], + "image_footnote": [], + "bbox": [ + 738, + 314, + 903, + 568 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 638, + 326, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dynamic scene reconstruction is a long-term challenge in 3D vision. Recent methods extend 3D Gaussian Splating to dynamic scenes via additional deformation fields and apply explicit constraints like motion flow to guide the deformation. However, they learn motion changes from individual timestamps independently, making it challenging to reconstruct complex scenes, particularly when dealing with violent movement, extreme-shaped geometries, or reflective surfaces. To address the above issue, we design a simple yet effective plug-and-play module called TimeFormer to enable existing deformable 3D Gaussians reconstruction methods with the ability to implicitly model motion patterns from a learning perspective. Specifically,", + "bbox": [ + 88, + 670, + 485, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "TimeFormer includes a Cross-Temporal Transformer Encoder, which adaptively learns the temporal relationships of deformable 3D Gaussians. Furthermore, we propose a two-stream optimization strategy that transfers the motion knowledge learned from TimeFormer to the base stream during the training phase. This allows us to remove TimeFormer during inference, thereby preserving the original rendering speed. Extensive experiments in the multi-view and monocular dynamic scenes validate qualitative and quantitative improvement brought by TimeFormer.", + "bbox": [ + 511, + 638, + 908, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 827, + 645, + 842 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-quality reconstruction of dynamic scenes is significantly challenging in computer vision and graphics, yet has a wide range of potential applications in movie production,", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author", + "bbox": [ + 112, + 875, + 246, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project Page: https://patrickddj.github.io/TimeFormer", + "bbox": [ + 112, + 887, + 400, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8721", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "virtual reality, and augmented reality. The difficulty stems from factors like occlusions, translucent materials, specular surfaces, and changing topology, all of which are prevalent in dynamic scenes.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the success of neural radiance field (NeRF) on static scenes [3, 36, 37], extending NeRF to dynamic scenes [15, 24, 26, 30, 38, 39, 50, 71] has been explored. However, these NeRF-based methods are limited by computationally intensive volume rendering [7], which makes real-time rendering almost impossible. Recently, 3D Gaussian Splatting (3DGS) [22] represents the scene with anisotropic 3D Gaussians and develops a rasterization-based rendering algorithm, which allows real-time rendering through directly projecting 3D Gaussians onto the image plane. Following the revolutionized 3DGS, dynamic Gaussian Splatting methods [9, 27, 48, 67] have been introduced for reconstructing dynamic scenes. These methods typically construct a canonical 3DGS and utilize a deformation field based on individual timestamps [27, 60, 66].", + "bbox": [ + 89, + 154, + 482, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, temporal relationships of 3D Gaussians have been poorly investigated. Previous studies have primarily modeled motion patterns using various types of the deformation field, for example, MLPs [2, 19, 33, 34, 43, 53, 66, 70], spatial-temporal planes [10, 32, 60], polynomial functions [27], Fourier series [21], and combinations of these methods [28]. These methods learn motion patterns from independent time input in a vanilla way, neglecting internal cross-time relationships. Some studies introduce the motion flow regularization [23] to explicitly learn motion patterns from neighboring frames [32, 34]. Although these methods promote similar motion between adjacent timestamps, they adopt a local perspective on time series during optimization. This constraint makes it challenging to reconstruct scenes with more complex motion patterns, such as sudden appearances, violent movements, or reflective surfaces.", + "bbox": [ + 89, + 383, + 482, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce TimeFormer, a transformer module designed to implicitly learn motion patterns across multiple timestamps. By modeling temporal relationships within a time batch, we aim to provide a global view of the entire time series, enabling the deformation field backbones themselves to capture motion patterns from a learning perspective, as shown in Fig. 1. Specifically, TimeFormer utilizes a Cross-Temporal Encoder to capture the implicit motion patterns of 3D Gaussians across multiple sampled timestamps using a self-attention mechanism. Moreover, to avoid additional computational costs of TimeFormer during inference, we present a two-stream optimization strategy. By sharing the weights of two deformation fields, we transfer the motion pattern learning from TimeFormer stream to the base stream during training. Therefore, we can eliminate TimeFormer during inference and maintain the original rendering speed. Notably, TimeFormer does not require any prior information and extracts motion patterns solely from", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/735ecea33472d60d7c3725f3768d2de64dc9e8f2c06dca090d6cab986072efaf.jpg", + "image_caption": [ + "Points: $154.5\\mathrm{k}$ FPS:31.3" + ], + "image_footnote": [], + "bbox": [ + 521, + 97, + 696, + 160 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f67580d67fc1e2454ab9360a894d963dcda2536e4378dfe2707028f203912bd5.jpg", + "image_caption": [ + "Points: $63.6\\mathrm{k}$ FPS: 79.7" + ], + "image_footnote": [], + "bbox": [ + 718, + 87, + 885, + 157 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/29581fe486f0f084c13f286aabcf26656abec65d2ae7a2ca2beaa611bb05b393.jpg", + "image_caption": [ + "DeformGS PSNR:27.17" + ], + "image_footnote": [], + "bbox": [ + 516, + 176, + 702, + 262 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d92e509b77958c36d5f22608a21db0bb64a51b1d6d8d9a7ef5ea5706895c0b9c.jpg", + "image_caption": [ + "+TimeFormer PSNR:28.70", + "Figure 2. TimeFormer guides towards more efficiently distributed canonical space, showing higher FPS and better quality. The results are from \"cut lemon\" in the HyperNeRF Dataset [39]." + ], + "image_footnote": [], + "bbox": [ + 707, + 175, + 903, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RGB supervision, which can be seamlessly adapted to previous deformable 3D Gaussian methods in a plug-and-play manner. Furthermore, TimeFormer promotes faster gradient descent and guides a more efficient canonical space, ultimately increasing FPS during inference, as in Fig. 2.", + "bbox": [ + 511, + 333, + 903, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, our main contributions are as follows:", + "bbox": [ + 532, + 410, + 861, + 424 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose TimeFormer, a transformer module that enhances current deformable 3D Gaussians reconstruction methods in a plug-and-play manner from an automatic learning perspective.", + "- The two-stream optimization strategy allows the exclusion of TimeFormer during inference while maintaining and even improving rendering speed.", + "- Extensive experiments on real-world datasets validate the effectiveness of TimeFormer, achieving state-of-the-art rendering quality." + ], + "bbox": [ + 513, + 426, + 903, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 594, + 661, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Dynamic Scene Reconstruction", + "text_level": 1, + "bbox": [ + 513, + 619, + 787, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic Scene Reconstruction has been extensively researched over many years, with a wealth of studies [20, 39, 65, 66, 73] contributing to current progress. The pioneering neural radiance field [3, 12, 36, 47, 68] has demonstrated photorealistic rendering for novel view synthesis from calibrated multiview images, inspiring extensive approaches [8, 11, 38, 39, 41, 51, 62] to extend NeRF to 4D space-time field for dynamic scene reconstruction. There are majorly two lines of NeRF-based methods for dynamic scene reconstruction: 1) deformation-based methods [11, 38, 39, 41, 51] model deformation changes by using a deformation field to map the queried positions in different timestamps to a canonical space; 2) compact 4D space-time fields [5, 8, 25, 62] take the timestamps, positions, and directions as input to predict color and density. However, NeRF-based approaches are limited to training and rendering speed. Thus, a large number of meth-", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8722", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ods [1, 4, 13, 24, 29, 31, 42, 46, 55-57, 61, 63] have been proposed to accelerate NeRF-based methods.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent 3DGS [22] achieves real-time rendering and dynamic 3DGS approaches [27, 66, 67] emerge. In details, deformation-based methods [2, 10, 19, 21, 27, 28, 32-34, 43, 53, 60, 66, 70] employ a deformation field to predict per-Gaussian offsets and deform a canonical 3DGS according to different timestamps. Meanwhile, 4DGS [67] and 4D-Rotor [9] add the time dimension to 3D Gaussian, forming a 4D Gaussian representation. TimeFormer focuses on modeling the temporal relationship and is plug-and-play to deformation-based methods [66] and 4D space-time representations [27, 67]. There are also a few attempts [28, 43, 48] in leveraging the nearby motion or flow for dynamic Gaussian reconstruction. For example, 3DGStream [48] proposes a per-frame optimization strategy, which predicts current Gaussian attributes based on previous Gaussian attributes. Meanwhile, Gaussian-Flow [28] presents explicitly model time-dependent residual of each attribute. To model a long video sequence, SWinGS [43] splits a video sequence into multiple sliding windows based on optical flow and uses a deformation-based method to model each sliding window.", + "bbox": [ + 91, + 123, + 483, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Motion Modeling", + "text_level": 1, + "bbox": [ + 89, + 452, + 261, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motion prediction in dynamic scene reconstruction from multi-view videos or monocular videos is an ill-posed problem. Early dynamic NeRF [15, 38, 39, 51] directly learns motion patterns from individual timestamps, and several later works propose motion flow regularization terms [15, 26, 30, 35, 40, 44, 45, 50, 54, 69] to promote the learning of cross-time motion patterns. These methods typically utilize 2D prior information (e.g., optical flow) from pre-trained networks [49] to supervise the scene flows within neighboring frames. PREF [45] uses motion predictor to infer current motion based on the previous four frames and propose a self-supervision strategy without optical flow prior. KFD-NeRF [69] models motion patterns by Kalman Filter based on the previous two frames. CT-NeRF [35] employs a cross-attention mechanism from transformer [52] to learn the correlation of conservative frames.", + "bbox": [ + 89, + 476, + 482, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic 3D Gaussian Splatting (3DGS) methods [16, 17, 58, 72] also utilize optical flow supervision to enhance reconstruction performance. Techniques such as MD-Splatting [10], D3DG [34], and ST-4DGS [23] focus on minimizing both forward and backward Gaussian flow to promote temporal smoothness in Gaussian motions. DN4DGS [32] incorporates information from previous, current, and next timestamps into the deformation field to capture cross-time motion patterns. However, these methods primarily establish temporal correlations with neighboring timestamps, which limits their ability to address long-term motion changes. Additionally, they introduce extra compu", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tational costs during inference, which can reduce rendering speed. In contrast, TimeFormer captures the temporal relationships of Gaussians from a global perspective across the entire time series. Notably, TimeFormer is employed only during the training phase, without any additional computational costs during inference.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary: Deformable 3D Gaussians", + "text_level": 1, + "bbox": [ + 511, + 194, + 870, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussians [22] represent the scene with a set of 3D Gaussians, each of which has unique opacity $o \\in [0,1]$ , center position $\\mu \\in \\mathbb{R}^{3 \\times 1}$ , and covariance matrix $\\Sigma \\in \\mathbb{R}^{3 \\times 3}$ . For a position $x \\in \\mathbb{R}^{3 \\times 1}$ in 3D space, the corresponding contribution of a 3D Gaussian on it can be formulated as:", + "bbox": [ + 511, + 219, + 903, + 294 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (x) = o \\cdot e ^ {- \\frac {1}{2} (x - \\mu) ^ {\\top} \\Sigma^ {- 1} (x - \\mu)}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 300, + 903, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The covariance matrix $\\Sigma$ can be decomposed into a scaling matrix $\\mathbf{S}$ and a rotation matrix $\\mathbf{R}$ : $\\boldsymbol{\\Sigma} = \\mathbf{RSS}^T\\mathbf{R}^T$ , where $\\mathbf{S} = \\mathrm{diag}([s_x, s_y, s_z])$ and $\\mathbf{R}$ can be transformed from a quaternion $[r_w, r_x, r_y, r_z]$ . Then the 3D Gaussians can be splatted to a 2D camera plane through differential Gaussian splatting. To model the appearance of 3D Gaussians, spherical harmonics (SH) are introduced to define the color $c$ . Finally, for each pixel, the rendering results of 3DGS can be derived by calculating the color contribution of all the related Gaussians. This process is known as $\\alpha$ -blending:", + "bbox": [ + 511, + 330, + 906, + 482 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i} ^ {N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 489, + 903, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c_{i},\\alpha_{i}$ represent the color and density computed from the $i$ -th 3D Gaussian.", + "bbox": [ + 511, + 540, + 903, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent methods [27, 60, 66] utilize deformation fields that extend 3DGS to 4D space, inspired from NeRF-based methods such as D-NeRF [41]. The structure of the deformation field $\\mathcal{D}$ can vary among MLP [66], K-Plane [60] and Polynomial [27], while the deformation process can be summarized as follows:", + "bbox": [ + 511, + 570, + 905, + 660 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n(\\Delta \\mu , \\Delta r, \\Delta s) = \\mathcal {D} (\\mu , t), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 670, + 903, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where timestamp $t \\in \\mathcal{T}$ , $\\mathcal{T} \\in \\mathbb{R}^{T \\times 1}$ contains $T$ linear time inputs, $\\mu, r, s$ are the center position, rotation quaternion, and scaling factors of 3D Gaussians, and $\\Delta \\mu, \\Delta r, \\Delta s$ are their residuals, respectively.", + "bbox": [ + 511, + 696, + 905, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 768, + 604, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first provide an overview of the enhanced deformable 3D Gaussians reconstruction with the proposed TimeFormer (Sec. 4.1). We then introduce the TimeFormer in detail (Sec. 4.2) which consists of a Cross-Temporal Encoder and shared deformation fields (Sec. 4.3). We also discuss the insights behind TimeFormer from the view of gradient flow.", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8723", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5b6bd7ab238578bd016fa8c5cb0ad9395d482c0ac9a297a2a556a4a391767be1.jpg", + "image_caption": [ + "Figure 3. The Framework of Deformable 3D Gaussians Reconstruction with TimeFormer. Existing deformable 3D Gaussians framework usually includes the canonical space and the deformation field (first row), we incorporate TimeFormer to capture cross-time relationships and explore motion patterns implicitly (second row). We share weights of two deformation fields to transfer the learned motion knowledge. This allows us to exclude this Auxiliary Training Module during inference." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 908, + 255 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/80eb44fe0b95b28ef390966ae250d0731cd90aeacf0c4463daa07356ce68b0fa.jpg", + "image_caption": [ + "Figure 4. The Structure of Cross-Temporal Encoder. We concatenate randomly sampled timestamps to position $\\mu$ , treating them as special tokens in a sequence. This module is designed to model multi-temporal relationships and produce distinct time-variant position offsets $\\Delta p_0, \\ldots, \\Delta p_{B-1}$ ." + ], + "image_footnote": [], + "bbox": [ + 161, + 320, + 413, + 489 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/28776ba2a191a383669d96123c17102aeb2d8a3fa26e7b7d2ed65afbb0d40170.jpg", + "image_caption": [ + "Figure 5. Data Flow Changes in the Deformation Field. Dashed lines represent new data flow among time samples $t_0, \\dots, t_{B-1}$ ." + ], + "image_footnote": [], + "bbox": [ + 148, + 578, + 428, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 681, + 202, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous methods model motion patterns by explicitly learning temporal relationships on individual or neighboring timestamps, failing on those complex scenes containing violent movement or dynamic reflective surfaces. In contrast, we present TimeFormer to enable the deformable 3D Gaussian backbones themselves to model cross-time relationships from an implicit learning perspective. The main framework with the proposed TimeFormer is shown in Fig. 3. Our approach retains standard reconstruction modules, which include (1) 3D Gaussians in the canonical space and (2) a deformation field that applies time-variant transformation. Additionally, TimeFormer is introduced before the deformation field to extract implicit cross-time motion", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "features for each Gaussian through a self-attention mechanism along the time dimension. Moreover, we share the weights of two deformation fields to transfer the motion knowledge from TimeFormer to mitigate the gap between the original branch and TimeFormer branch, which supports real-time rendering without TimeFormer during inference.", + "bbox": [ + 511, + 321, + 906, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. TimeFormer", + "text_level": 1, + "bbox": [ + 511, + 424, + 650, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-Temporal Encoder. BatchFormer [18] demonstrates that the attention mechanism helps learn sample relationships from batch dimension, rather than channel and spatial dimensions [6, 52], inspiring us that different timestamps can also be considered as a special time batch. Let $\\mathcal{T}_s\\subset \\mathcal{T},\\mathcal{T}_s = \\{t_i\\}_{i = 0}^{B - 1}$ denotes randomly sampled timestamps, and let $\\mathcal{G}\\in \\mathbb{R}^{N\\times (3 + C)}$ denotes Gaussians in the canonical space, where $B$ is the size of time batch, $N$ is the number of Gaussians, $3 + C$ means each Gaussian has 3 position channels and $C$ additional channels. As in Fig. 4, all Gaussians's positions are made into $B$ copies, expanded into $\\mathcal{G}_c\\in \\mathbb{R}^{B\\times N\\times 3}$ , and sampled timestamps are made into $N$ copies, expanded into $\\mathcal{T}'\\in \\mathbb{R}^{B\\times N\\times 1}$ . Then, we compose $\\mathcal{G}_c$ and $\\mathcal{T}'$ together and apply position encoding function $\\gamma$ to extract high frequency information, as in Eq. 4:", + "bbox": [ + 511, + 446, + 908, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\gamma (p) = \\left(\\sin \\left(2 ^ {0} \\pi p\\right), \\cos \\left(2 ^ {0} \\pi p\\right), \\dots \\right. \\\\ \\sin \\left(2 ^ {L - 1} \\pi p\\right), \\cos \\left(2 ^ {L - 1} \\pi p\\right)) \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 686, + 903, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our experiments, we set $L = 6$ for both positions $x$ and time $t$ . We treat $[\\gamma(\\mathcal{G}_c, \\gamma(\\mathcal{T}')] \\in \\mathbb{R}^{B \\times N \\times (3 \\times 2L + 2L)}$ as the original input $F_0$ to TimeFormer. It can be considered as $N$ sequences of the length $B$ , containing $8L$ feature channels. With $M$ transformer encoder layers, for $m^{th}$ layer, intermediate features are calculated through multi-head self-attention(MSA) and MLP blocks. In the final stage, we use a tiny MLP to transform the last encoded features $F_{M-1}$ into offset $\\mathcal{O} \\in \\mathbb{R}^{B \\times N \\times 3}$ in the linear space:", + "bbox": [ + 511, + 734, + 906, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {O} = M L P \\left(F _ {M - 1}\\right), \\quad \\mathcal {G} _ {t} = \\mathcal {G} _ {c} + \\mathcal {O} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 883, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8724", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider the output from the Cross-Temporal Encoder as a fixing residual term to the original positions to encourage a gradual, steady learning process on motion patterns. Implicit cross-time relationship learning in TimeFormer enables the automatic aggregation of Gaussians with similar variations during optimization, promoting more efficient spatial distribution and accelerating rendering speed.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gradient Analysis. Let $\\mathcal{D}$ and $\\mathcal{P}$ be the deformation field and TimeFormer respectively, the output of the deformation field $\\Delta \\mu_{i}$ and partial derivative for $\\mu$ in time $t_i\\in \\mathcal{T}_s$ are formulated as follows:", + "bbox": [ + 89, + 196, + 480, + 256 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mu_ {i} = \\mathcal {D} (\\mu , t _ {i}), \\quad \\frac {\\partial \\Delta \\mu_ {i}}{\\partial \\mu} = \\frac {\\partial \\mathcal {D} (\\mu , t _ {i})}{\\partial \\mu} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 265, + 482, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With TimeFormer applied on time batch $\\mathcal{T}_s$ , we reformulate as Eq. 7 and data flow changes as in Fig. 5.", + "bbox": [ + 89, + 308, + 480, + 339 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mu_ {i} = \\mathcal {D} (\\mu + \\mathcal {P} (\\mu , \\mathcal {T} _ {s}), t _ {i}) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 351, + 482, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To calculate the partial derivative for $\\mu$ , we apply the chain rule. Let $a = \\mu + \\mathcal{P}(\\mu, \\mathcal{T}_s)$ , then:", + "bbox": [ + 89, + 380, + 482, + 410 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\Delta \\mu_ {i}}{\\partial \\mu} = \\frac {\\partial \\mathcal {D} (a , t _ {i})}{\\partial a} \\cdot \\frac {\\partial a}{\\partial \\mu}, \\quad \\frac {\\partial a}{\\partial \\mu} = 1 + \\frac {\\partial \\mathcal {P} (\\mu , \\mathcal {T} _ {s})}{\\partial \\mu} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 420, + 482, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {\\partial \\Delta \\mu_ {i}}{\\partial \\mu} = \\frac {\\partial \\mathcal {D} (a , t _ {i})}{\\partial a} \\cdot \\left(1 + \\frac {\\partial \\mathcal {P} (\\mu , \\mathcal {T} _ {s})}{\\partial \\mu}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 468, + 482, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eq. 6 has a weaker dynamic nature, as it only considers the current timestamp $t_i$ . In contrast, Eq. 9 incorporates an additional gradient term, $\\frac{\\partial \\mathcal{P}(\\mu, \\mathcal{T}_s)}{\\partial \\mu}$ , enabling the current state to be influenced by any past or future states. In other words, $\\Delta \\mu_i$ also optimizes the models according to other timestamps $t_j \\in \\mathcal{T}_s (j \\neq i)$ , which is a significant difference compared to the backward process without Time-Former. Such design accounts for cross-time attention and allows the models to capture more challenging motion patterns from a global view of the entire time series.", + "bbox": [ + 89, + 508, + 482, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Shared Deformation Fields", + "text_level": 1, + "bbox": [ + 89, + 670, + 334, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Accounting for additional computation costs of Time-Former, which can significantly decrease rendering speed, we force the original deformation field and the auxiliary deformation field to share weights for knowledge transferring.", + "bbox": [ + 89, + 693, + 482, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\mathcal{V}$ be the camera viewpoints, we apply the shared deformation fields to predict deformed space from both Gaussians in the canonical space and Gaussians from Time-Former. Then, we apply the splatting algorithm to these two groups of deformed space. We calculate the losses between rendered images and ground truth $\\mathcal{T}_{gt}$ as follows:", + "bbox": [ + 89, + 753, + 482, + 844 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\left\\| S p l a t t i n g \\left(\\mathcal {D} \\left(\\mathcal {G} _ {c}, \\mathcal {T}\\right), \\mathcal {V}\\right) - \\mathcal {I} _ {g t} \\right\\| _ {1} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 856, + 482, + 873 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = \\| S p l a t t i n g (\\mathcal {D} (\\mathcal {G} _ {t}, \\mathcal {T}), \\mathcal {V}) - \\mathcal {I} _ {g t} \\| _ {1} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 883, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {t} \\mathcal {L} _ {t} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 90, + 903, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": ", where $\\mathcal{L}_c, \\mathcal{L}_t$ represent losses of original branch and Time-Former branch with $\\lambda_c > \\lambda_t$ . We use a relatively smaller $\\lambda_t$ because we find it easy to overfit on the second branch with TimeFormer, causing a degradation in inference quality.", + "bbox": [ + 511, + 114, + 903, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 513, + 186, + 638, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 212, + 730, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our implementation is tested on a single A100 GPU. We use $M = 4$ transformer encoder layers in TimeFormer and the number of time samples is $B = 4$ . We set $\\lambda_c = 1$ and $\\lambda_t = 0.8$ to prevent overfitting on TimeFormer. We assess our experimental results using image quality metrics, including peak-signal-to-noise ratio (PSNR) and structural similarity index (SSIM [59]).", + "bbox": [ + 511, + 234, + 903, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines & Datasets. We apply TimeFormer to 4DGS [60], STGS [27] on multi-view videos, e.g., N3DV Dataset [25], and evaluate TimeFormer on 4DGS [60] and DeformGS [66] on monocular dynamic videos, e.g., HyperNeRF Dataset [39], respectively. We also utilize NeRF-DS Dataset [64] to demonstrate the robustness of TimeFormer on moving objects with specular surfaces.", + "bbox": [ + 511, + 340, + 905, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Experimental Comparisons", + "text_level": 1, + "bbox": [ + 513, + 455, + 761, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-View Video Dataset. In Tab. 1, TimeFormer improves the reconstruction quality (e.g., PSNR) of original 4DGS [60] and STGS [27] by 0.74 and 0.61, outperforming all baselines methods. We extract the frames at an initial time for scene Cook Spanish in Fig. 6, and TimeFormer can reconstruct a clearer geometry and specular effects compared to the original results.", + "bbox": [ + 511, + 477, + 905, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To have a more comprehensive overview of Time-Former's capacity on different temporal stages, we collect PSNR for 300 frames of all six scenes from N3DV Dataset [25] and calculate the average per-frame PSNR. As in Fig. 7, the reconstruction quality increases more significantly at the beginning and end, and such balanced reconstruction results are attributed to TimeFormer's cross-temporal attention mechanism on the whole time series.", + "bbox": [ + 511, + 584, + 905, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Monocular Video Dataset. Tab. 3 shows the enhancements achieved by TimeFormer on the HyperNeRF Dataset. It's observed that TimeFormer increases the PSNR by 0.89 for 4DGS [60] and 0.94 for DeformGS [66], and the SSIM by 0.019 and 0.028, respectively. Fig. 9 additionally provides visualization of the improvement in the image quality: on the left, TimeFormer generates a clearer outline between fingers, while on the right the texture on the front of the broom is clearer. TimeFormer also eliminates artifacts and reveals a clear structure of the broom.", + "bbox": [ + 511, + 704, + 905, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dynamic Reflective Video Dataset. Fig. 10 additionally demonstrates that TimeFormer can work well on scenes with dynamic specular objects on NeRF-DS Dataset [64].", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8725", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/00d6885c985cbc500c23f25fdf81f5443ae2cfafc988b84f0d2f900bf79a8ab2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSear SteakFlame SSIMSalmon SSIMCut Beef SSIMFlame SSINRSteak SSIMCook Spinach SSINRSSIMCoffee SSINRMartini SSIMMean
PSNRSSIMPSNRSSIM
K-Plane [14]32.520.97130.440.94231.820.96532.380.97030.600.96829.990.94331.63
MixVovels [55]31.210.97129.920.94531.300.96531.430.97031.610.96529.360.94630.80
GS4D [67]32.920.95326.390.89733.080.95933.810.96732.770.95625.230.88430.07
4D-Rotor [9]32.860.95628.250.91333.140.95231.610.95332.560.94927.950.90831.06
4DGS [60]32.490.94928.920.91732.900.95632.510.95432.460.94827.340.90331.10
+TimeFormer33.380.95529.330.92433.110.95733.250.95333.030.94928.930.91031.84
STGS [27]33.710.96228.210.92133.520.95833.460.96333.130.95527.710.91531.62
+TimeFormer34.340.96529.130.92433.570.95834.040.96433.450.95628.830.91732.23
", + "bbox": [ + 91, + 87, + 911, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative Comparisons on N3DV Dataset [25]. We use bold font to indicate the improvement, statistics of K-Plane [14] and MixVoxels [55] are from their original paper, while we calculate metrics of all Gaussian-based methods by running their official codes.", + "bbox": [ + 89, + 250, + 906, + 280 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1953b579de1e33da08265ef263e1ef704b7e78269a1ec5fe933b2a0348e1968a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 292, + 254, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6ec2eafbdc5e22dcda778ea59587681b2e915465db2ffc82a126e0c3d7fa7398.jpg", + "image_caption": [ + "Figure 6. Visualization of Comparisons on N3DV Dataset [25]." + ], + "image_footnote": [], + "bbox": [ + 254, + 292, + 415, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/32026f7bef44af4dfcbc0d87200b9546234bbfdbdebc365982f538eccf83330e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 292, + 578, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62b48a06311c8803e6d29909eff5674607083d784b65be180f27b417fcab333f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 292, + 741, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da94bbadd807988c465daf84fd10986d831afb8edfd4c3e91b2abb3dce654e08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 292, + 905, + 460 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5385e5e9c2ca0018d6ff00bd8373c5c0ca03fa0c28a0024005b1f004ae5b4861.jpg", + "image_caption": [ + "Figure 7. Per-Frame PSNR on N3DV Dataset [25]. TimeFormer improves the reconstruction quality, especially at the beginning and end time series." + ], + "image_footnote": [], + "bbox": [ + 109, + 508, + 465, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TimeFormer can reconstruct a cleaner appearance on reflective material (e.g., a metallic cup, a glass bottle), and this improvement is attributed to the attention mechanism which enables the deformation field to detect changes on specular surfaces from a learning perspective automatically.", + "bbox": [ + 89, + 703, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of FPS & Canonical Space. We also find that TimeFormer achieves more efficient representation with much fewer points in the final canonical space, thereby achieving a higher FPS in Tab. 2. We attribute the phenomenon to the reason that TimeFormer guides faster gradient descent and promotes the spatial layout of Gaussian points during optimization. Interestingly, in Fig. 8, the L1 loss of both branches in our method decreases", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f29f1ee14da7316287f670665b2067da9c1e5a95d60286b3bb38e234d09c4327.jpg", + "image_caption": [ + "Figure 8. Comparisons of Convergence Speed. We use the same batch size for 4DGS [60] and calculate the loss of both branches in our method." + ], + "image_footnote": [], + "bbox": [ + 549, + 517, + 870, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "faster and shows almost the same convergence trend during training, exceeding the optimization speed of the original method. This eliminates many redundant points and guides the canonical space toward a more efficient distribution, as in Fig. 2. Moreover, Fig. 8 also proves that the cross-time relationship learned in TimeFormer has been successfully transferred to the base branch, as introduced in Sec. 4.3.", + "bbox": [ + 511, + 731, + 906, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of Motion Patterns. We argue that TimeFormer achieves more robust learning of motion patterns. To take a deeper insight, we visualize the motion changes as follows: for each Gaussian, we calculate its bias $(\\Delta_x, \\Delta_y, \\Delta_z)$", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8726", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/83bf71128457713c2e5b31099a96df3930484451c7af3415d374a0664849a128.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 87, + 171, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4a8920e947b25e399638c6cddcd2083c61301303318b285911de78fbfbda1ff.jpg", + "image_caption": [ + "DeformGS PSNR:27.63" + ], + "image_footnote": [], + "bbox": [ + 96, + 199, + 169, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e86b021e18a22956a95a60c93984286616ebd4f37e4485d3b9bb25fdd1cf39d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 174, + 88, + 253, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99b0a988d13804c01e166678c2cf8615ced9c2ef71aeab62ef0d045ee5a682e1.jpg", + "image_caption": [ + "+TimeFormer PSNR:28.59" + ], + "image_footnote": [], + "bbox": [ + 178, + 199, + 251, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e3eb461b7d0a5dcbcec9a624a9e0a7a94858a7b351c819a3bb8bf6456f0d1fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 88, + 333, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a3c4d553a845fd812df9bb676a241b350919c7b1461400f2d7257fe30591768.jpg", + "image_caption": [ + "4DGS PSNR:28.55" + ], + "image_footnote": [], + "bbox": [ + 256, + 199, + 331, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6e8358bc3ac0cb5eb4433d2cab300554d7f0e6decce986e78f7127abcbb0668.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 88, + 415, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25454d64dca6f24a9919a8941575cae6a4fe15683ba04ff7827d9fdf2b904181.jpg", + "image_caption": [ + "+TimeFormer PSNR:29.51" + ], + "image_footnote": [], + "bbox": [ + 339, + 199, + 413, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e227a24b7cc91f7884655ccd8e833f855b7bec2586c6dfe8545f3b77d22044b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 88, + 496, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/77a375c9d9e614eeed111ccf790af346748f75352babf27b1f8525c75c7aa034.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 419, + 199, + 493, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ad325931a0d50b4a3dccdcd456c7e75f946961dc92323367418cd44208e51f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 88, + 576, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/92b543abe69218a2214023fb87267060df7370ee96d2a62e3702aa297d262770.jpg", + "image_caption": [ + "DeformGS PSNR:20.63" + ], + "image_footnote": [], + "bbox": [ + 501, + 199, + 575, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3401f94a25153b73a4a7ad2ea72c93623d2422ef4eed5b6d405e8f527771966.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 88, + 660, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1e6434f642df0e02b72666f406f1108467a231e81ac71b77f6c699c765f0458.jpg", + "image_caption": [ + "+TimeFormer PSNR:21.13" + ], + "image_footnote": [], + "bbox": [ + 584, + 199, + 656, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd7da060915c87ee872b913a6556c2574c1d5bc20cbd9b6257db98183995073e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 88, + 741, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/486c3d8b531296b2be472980f5879879334757fa3df4c72b34afd231a9553f69.jpg", + "image_caption": [ + "4DGS PSNR:21.28" + ], + "image_footnote": [], + "bbox": [ + 666, + 199, + 738, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e1e26777659cdaeac54f4ede6be381ece76d8c4f18028757c1f1d78c2257f54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 88, + 821, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a002a9ba9de71dc98c829e62ea673044c8b3f5bfbb8e57a5358d9d7d69242cfa.jpg", + "image_caption": [ + "+TimeFormer PSNR:22.23" + ], + "image_footnote": [], + "bbox": [ + 746, + 199, + 820, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/479af13db44a726d37af437335040c70701c69b4e5f41cbec9c49e2f3a4ca293.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 826, + 88, + 903, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/caada894c47253a00648e58a31d9d65af5fe811f36b4283875cfc40b0119cd64.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 828, + 199, + 900, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/636636a55414d32f71812b1047e518df9603037018f16c81439f5345f79b4698.jpg", + "image_caption": [ + "Figure 9. Visualization of Comparisons on HyperNeRF Dataset [39]." + ], + "image_footnote": [], + "bbox": [ + 98, + 306, + 272, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/90c51d445920d79dee882b75a1abb57a77d629cfbe95365898d5c07a09fa0664.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 306, + 429, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a78033a76223bd4e7efbafc92a4b346b5f5de741ebf5c43344ed5bd8502112f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 306, + 586, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a3bd1bfdc5b4371a9d9156a2a2357c88be8698083d1320d632217fb3748c436.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 306, + 746, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b0e790bde867741af637cf11ad6acc1b33764edd6acdc8c61920bbf439455938.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 306, + 903, + 376 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96c9520128b8114043673795b0ae9145ebe6c7ed66e9b2c80dfd42f638debb7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 377, + 271, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/04b69a301ed48abb34f2757c26f02cb1a667868d84784aa32e07ef22e4bebaae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 377, + 429, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f963041b35a5c71889e748965829793d45d4be7befe542c168b493b0afd0dbe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 377, + 586, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8f8a2b84a32cc1a366788266b4ae56a56454eb9eda67ba85ff2fbb43a911bab3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 377, + 746, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9f643ecef67c785672df5638bad6cd3b0f16277582e2f3a033ef6fd270322d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 377, + 903, + 446 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81a2c7df2388b93e95ed341480865c92cd83fa396c1a95a9e4f6adf95396c13b.jpg", + "image_caption": [ + "Figure 10. Visualization of Comparisons on NeRF-DS Dataset [64]. Compared with original results from DeformGS [66], TimeFormer presents a clearer visual effect on dynamic objects with specular surfaces." + ], + "image_footnote": [], + "bbox": [ + 98, + 448, + 271, + 516 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4b0d92a467d6cfaf042031d33b1e9046acfeb3120d39d789ee4d6c72e1064188.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 448, + 429, + 516 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20b2a83316c6bc1f1f6e3383f72e99551995f66ce993daf1a9c02b6d58eae03a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 431, + 448, + 586, + 516 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/def1b06f93e0a05a0290988e7b38b2b2f5c3908591da6b0d508d4f127d49f738.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 448, + 746, + 516 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/460e8424bf472d5f76e830b56a3fafe67406231ed166ed629daac01cba9d6bcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 748, + 448, + 903, + 516 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9ef168bf17071bf0e36cfd832444b3078cd2fdfd7205c82d1531133f7af168fb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodN ↓Training ↓FPS ↑
STGS [27]172.1 k41 min82.1
+TimeFormer148.9 k78 min88.9
4DGS [60]145.9 k52 min31.7
+TimeFormer113.1 k94 min37.9
", + "bbox": [ + 116, + 579, + 457, + 660 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b0a6987311efd2a24052a6208c8f546c03d400ec68c33507391c12d5247c8d34.jpg", + "table_caption": [ + "(a) Quantitative comparisons on N3DV Dataset [25]." + ], + "table_footnote": [], + "table_body": "
MethodN↓Training ↓FPS ↑
DeformGS [66]169.6 k25 min30.1
+TimeFormer82.9 k35 min58.9
4DGS [60]172.3 k32 min35.7
+TimeFormer135.5 k48 min40.9
", + "bbox": [ + 117, + 680, + 457, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(b) Quantitative comparisons on HyperNeRF Dataset [39].", + "bbox": [ + 130, + 763, + 441, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparisons of Gaussian Number, Training Time & FPS.", + "bbox": [ + 89, + 787, + 482, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "towards canonical space in time $t$ and replace the color attributes $(r, g, b)$ with the absolute value of this bias: $(r, g, b) \\gets (|\\Delta_x|, |\\Delta_y|, |\\Delta_z|)$ . We use the same splatting algorithm to render the accumulated motion bias, transfer", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ring the rendered result as a heatmap, as in Fig. 11.", + "bbox": [ + 511, + 583, + 849, + 597 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 11a, TimeFormer identifies the blade's motion as a rigid deformation, showing a sharper outline and consistent motion, especially as it contacts the initially static lemon. This demonstrates TimeFormer's ability to distinguish between moving and static objects. In Fig. 11b, TimeFormer captures subtle motion on the spray gun and clear flame flickering on the beef. Fig. 11c shows that TimeFormer also detects the violent movement of the elongated broom, where 4DGS [60] fails, proving its robustness in handling extremely geometries.", + "bbox": [ + 511, + 598, + 906, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 757, + 678, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cross-Temporal Encoder. Tab. 5 demonstrates that the performance of TimeFormer is not sensitive to time batch $B$ and transformer encoder layer $M$ . In Tab. 5b, we explore two sampling methods: random sampling and continuous sampling of timestamps. Both strategies demonstrate improvements in PSNR and other metrics compared to the baseline. Notably, the random sampling method yields more significant enhancements. We hypothesize that this", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8727", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e029d70d85ad9696aa983d1634b971cdff71d3e90ee3675996837f6f4d215dab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBroomChickenCut LemonTorchcoPeel BananaHandMean
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DeformGS [66]20.740.32226.320.78627.940.71427.40.87726.380.83627.790.7826.090.719
+TimeFormer21.010.32426.550.79230.010.78027.550.88327.240.85229.790.84827.030.747
4DGS[60]21.530.35126.820.79729.720.76327.450.88327.820.84429.520.84127.140.747
+TimeFormer22.840.40127.060.80130.230.77829.480.90228.210.85330.380.86228.030.766
", + "bbox": [ + 93, + 87, + 903, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5242b14c867ca8f47a5437c2d54c643622f8fdabc89397e634a051f591122b08.jpg", + "table_caption": [ + "Table 3. Quantitative Comparisons on HyperNeRF Dataset [39]." + ], + "table_footnote": [], + "table_body": "
MethodPressPlateBasinSieveBellCupMean
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
HyperNeRF [39]25.40.87318.10.71420.20.82925.00.90924.00.88424.10.89622.80.851
NeRF-DS [64]26.40.91120.80.86720.30.86826.10.93523.30.87224.50.91623.570.895
DeformGS[66]25.680.86620.820.81219.870.80425.710.88124.920.85424.520.89723.590.852
+TimeFormer26.290.86720.900.81719.930.80526.260.89125.900.87325.160.90324.100.859
", + "bbox": [ + 91, + 220, + 911, + 320 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fab3f703efa079a01918d7752d39f32e42468806559975932e1922028cced503.jpg", + "image_caption": [ + "(a) Cut Lemon" + ], + "image_footnote": [], + "bbox": [ + 98, + 367, + 480, + 529 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/134b17a7009fe9a1932a98d991747c9bdeb08dc63b2b48883c63487cbb735450.jpg", + "image_caption": [ + "(b) Torchocolate", + "(c) Broom", + "Figure 11. Motion Visualization on HyperNeRF Dataset [39]." + ], + "image_footnote": [], + "bbox": [ + 98, + 541, + 480, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "phenomenon occurs because the random sampling method allows for more effective propagation of dynamic behavior modeling across the entire time series.", + "bbox": [ + 89, + 805, + 482, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Shared Deformation Fields. In Tab. 5b, the image quality during inference suffers a considerable decrease if we train two deformation fields without sharing their weights.", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/393688fcb3f0844c9a5e3ad9d5a62eb11463a667b468199b3f9d375b53d03c6f.jpg", + "table_caption": [ + "Table 4. Quantitative Comparisons on NeRF-DS Dataset [64]." + ], + "table_footnote": [], + "table_body": "
SettingRandom SamplingContinuous Sampling
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
B=225.890.8750.11725.730.8710.124
B=426.060.8710.11925.780.8740.117
B=625.990.870.12325.730.8650.125
", + "bbox": [ + 522, + 367, + 915, + 446 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7bf28136f7fb00cf41557f1f2d960b86d5b89703b91a734511a8ce80c9ae1764.jpg", + "table_caption": [ + "(a) Ablation Studies on Batch Size and Sampling Strategies." + ], + "table_footnote": [ + "(b) Ablation Results. $M$ is the number of encoder layers, \"w/o Shared\" means not using shared weights in two deformation fields.", + "Table 5. Ablation Results. The results are from three scenes press, sieve and bell on NeRF-DS Dataset [64]." + ], + "table_body": "
SettingPSNR↑SSIM↑LPIPS↓
Baseline25.440.8670.125
M=125.960.8740.117
M=225.900.8740.120
M=326.050.8760.112
M=426.040.8730.119
w/o Shared24.810.8520.132
", + "bbox": [ + 566, + 468, + 870, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 8 also demonstrates the effectiveness of the two-stream strategy and TimeFormer's ability of re-balancing.", + "bbox": [ + 511, + 657, + 906, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 707, + 633, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose TimeFormer, a Transformer module which enhances existing deformable 3D Gaussians methods by implicitly modeling motion patterns from a learning perspective. In addition, we design a two-stream optimization strategy to transfer the learned motion knowledge from TimeFormer to the original deformation branch. This allows us to remove TimeFormer during inference and thus maintain the same inference speed. Extensive experiments demonstrate that TimeFormer facilitates the reconstruction of three state-of-the-art methods among three datasets, and illustrate the improvement on reconstructing complex scenes.", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8728", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 236, + 187, + 253 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 3", + "[2] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. arXiv preprint arXiv:2404.03613, 2024. 2, 3", + "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855-5864, 2021. 2", + "[4] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 3", + "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2", + "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 4", + "[7] Robert A Drebin, Loren Carpenter, and Pat Hanrahan. Volume rendering. ACM Siggraph Computer Graphics, 22(4): 65-74, 1988. 2", + "[8] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2", + "[9] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. In Proc. SIGGRAPH, 2024. 2, 3, 6", + "[10] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Mike Zheng Shou, Shuran Song, and Jeffrey Ichnowski. Md-splatting: Learning metric deformation from" + ], + "bbox": [ + 93, + 261, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work is completed during the author's internship at the Tencent Hunyuan 3D Department. This work is supported in part by the National Science Fund for Distinguished Young Scholars (No. 62325208), in part by the National Natural Science Foundation of China (No. 62232002), and in part by the Natural Science Foundation of Tianjin (No. 23ZGZNGX00020 and 24JCZXJC00050).", + "4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583, 2023. 2, 3", + "[11] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2", + "[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2", + "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 3", + "[14] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 6", + "[15] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2, 3", + "[16] Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wenchao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splitting gaussian dynamics for 4d content creation. arXiv preprint arXiv:2403.12365, 2024. 3", + "[17] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. arXiv preprint arXiv:2403.11447, 2024. 3", + "[18] Zhi Hou, Baosheng Yu, and Dacheng Tao. Batchformer: Learning to explore sample relationships for robust representation learning. In CVPR, 2022. 4", + "[19] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4220-4230, 2024. 2, 3", + "[20] Takeo Kanade, Peter Rander, and PJ Narayanan. Virtualized reality: Constructing virtual worlds from real scenes. IEEE multimedia, 4(1):34-47, 1997. 2", + "[21] Kai Katsumata, Duc Minh Vo, and Hideki Nakayama. A compact dynamic 3d gaussian representation for real-time dynamic view synthesis. In Computer Vision - ECCV 2024, pages 394-412. Springer Nature Switzerland, 2025. 2, 3", + "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3", + "[23] Deqi Li, Shi-Sheng Huang, Zhiyuan Lu, Xinran Duan, and Hua Huang. St-4dgs: Spatial-temporally consistent 4d gaussian splatting for efficient dynamic scene rendering. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3" + ], + "bbox": [ + 93, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8729", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 2, 3", + "[25] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2, 5, 6, 7", + "[26] Zhengqi Li, Simon Niklaus, Noah Snively, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2, 3", + "[27] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8508-8520, 2024. 1, 2, 3, 5, 6, 7", + "[28] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2, 3", + "[29] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. Advances in Neural Information Processing Systems, 35:36762-36775, 2022. 3", + "[30] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13-23, 2023. 2, 3", + "[31] Ange Lou, Benjamin Planche, Zhongpai Gao, Yamin Li, Tianyu Luan, Hao Ding, Terrence Chen, Jack Noble, and Ziyan Wu. Darenerf: Direction-aware representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5031-5042, 2024. 3", + "[32] Jiahao Lu, Jiacheng Deng, Ruijie Zhu, Yanzhe Liang, Wenfei Yang, Tianzhu Zhang, and Xu Zhou. Dn-4dgs: Denoised deformable network with temporal-spatial aggregation for dynamic scene rendering. arXiv preprint arXiv:2410.13607, 2024. 2, 3", + "[33] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2", + "[34] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 2024 International Conference on 3D Vision (3DV), pages 800-809. IEEE, 2024. 2, 3" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Xingyu Miao, Yang Bai, Haoran Duan, Fan Wan, Yawen Huang, Yang Long, and Yefeng Zheng. Ctnerf: Cross-time transformer for dynamic neural radiance field from monocular video. Pattern Recognition, 156:110729, 2024. 3", + "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[37] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2", + "[38] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2, 3", + "[39] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 2, 3, 5, 7, 8", + "[40] Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahiup Kang. Temporal interpolation is all you need for dynamic neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4212-4221, 2023. 3", + "[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2, 3", + "[42] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boring Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 3", + "[43] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Pérez-Pellitero. Swings: sliding windows for dynamic 3d gaussian splatting. In European Conference on Computer Vision. Springer, 2024. 2, 3", + "[44] Nagabhushan Somraj, Kapil Choudhary, Sai Harsha Mupparaju, and Rajiv Soundararajan. Factorized motion fields for fast sparse input dynamic view synthesis. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3", + "[45] Liangchen Song, Xuan Gong, Benjamin Planche, Meng Zheng, David Doermann, Junsong Yuan, Terrence Chen, and Ziyan Wu. Pref: Predictability regularized neural motion fields. In European Conference on Computer Vision, pages 664-681. Springer, 2022. 3", + "[46] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with de" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8730", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "composed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 3", + "[47] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2", + "[48] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20675-20685, 2024. 2, 3", + "[49] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3", + "[50] Fengrui Tian, Shaoyi Du, and Yueqi Duan. Mononerf: Learning a generalizable dynamic radiance field from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17903-17913, 2023. 2, 3", + "[51] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2, 3", + "[52] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3, 4", + "[53] Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3", + "[54] Chaoyang Wang, Lachlan Ewen MacDonald, Laszlo A Jeni, and Simon Lucey. Flow supervision for deformable nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21128-21137, 2023. 3", + "[55] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, Yafei Song, and Huaping Liu. Mixed neural voxels for fast multiview video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19706-19716, 2023. 3, 6", + "[56] Feng Wang, Zilong Chen, Guokang Wang, Yafei Song, and Huaping Liu. Masked space-time hash encoding for efficient dynamic scene reconstruction. Advances in Neural Information Processing Systems, 36, 2024.", + "[57] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 76-87, 2023. 3", + "[58] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. arXiv preprint arXiv:2405.18426, 2024. 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5", + "[60] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 1, 2, 3, 5, 6, 7, 8", + "[61] Minye Wu, Zehao Wang, Georgios Kouros, and Tinne Tuytelaars. Tetrirf: Temporal tri-plane radiance fields for efficient free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6487-6496, 2024. 3", + "[62] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9421-9431, 2021. 2", + "[63] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20029-20040, 2024. 3", + "[64] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8285-8295, 2023. 5, 7, 8", + "[65] Jason C Yang, Matthew Everett, Chris Buehler, and Leonard McMillan. A real-time distributed light field camera. Rendering Techniques, 2002(77-86):2, 2002. 2", + "[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 1, 2, 3, 5, 7, 8", + "[67] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 6", + "[68] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2", + "[69] Yifan Zhan, Zhuoxiao Li, Muyao Niu, Zhihang Zhong, Shohei Nobuhara, Ko Nishino, and Yinqiang Zheng. Kfdnerf: Rethinking dynamic nerf with kalman filter. arXiv preprint arXiv:2407.13185, 2024. 3", + "[70] Boming Zhao, Yuan Li, Ziyu Sun, Lin Zeng, Yujun Shen, Rui Ma, Yinda Zhang, Hujun Bao, and Zhaopeng Cui. Gaussian prediction: Dynamic 3d gaussian prediction for motion extrapolation and free view synthesis. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2, 3", + "[71] Kaichen Zhou, Jia-Xing Zhong, Sangyun Shin, Kai Lu, Yiyuan Yang, Andrew Markham, and Niki Trigoni. Dynpoint: Dynamic neural point for view synthesis. Advances in Neural Information Processing Systems, 36, 2024. 2" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8731", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Ruijie Zhu, Yanzhe Liang, Hanzhi Chang, Jiacheng Deng, Jiahao Lu, Wenfei Yang, Tianzhu Zhang, and Yongdong Zhang. Motions: Exploring explicit motion guidance for deformable 3d gaussian splatting. arXiv preprint arXiv:2410.07707, 2024. 3", + "[73] C Lawrence Zitnick, Sing Bing Kang, Matthew Uytendaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM transactions on graphics (TOG), 23(3):600-608, 2004. 2" + ], + "bbox": [ + 91, + 90, + 480, + 218 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8732", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/8cdc06bd-0ca5-4f6b-b012-50c8329288a4_model.json b/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/8cdc06bd-0ca5-4f6b-b012-50c8329288a4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f364f419163f96beabea768fa5498fd0ed8cadb9 --- /dev/null +++ b/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/8cdc06bd-0ca5-4f6b-b012-50c8329288a4_model.json @@ -0,0 +1,3425 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.218, + 0.131, + 0.78, + 0.175 + ], + "angle": 0, + "content": "TimeFormer: Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.203, + 0.317, + 0.223 + ], + "angle": 0, + "content": "Dadong Jiang\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.226, + 0.348, + 0.24 + ], + "angle": 0, + "content": "patrickddj7@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.246, + 0.337, + 0.265 + ], + "angle": 0, + "content": "Xianghui Yang3" + }, + { + "type": "text", + "bbox": [ + 0.196, + 0.268, + 0.349, + 0.282 + ], + "angle": 0, + "content": "yangxhui6@163.com" + }, + { + "type": "text", + "bbox": [ + 0.465, + 0.203, + 0.549, + 0.22 + ], + "angle": 0, + "content": "Zhi Hou\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.226, + 0.624, + 0.239 + ], + "angle": 0, + "content": "zhou9878@uni.sydney.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.433, + 0.246, + 0.564, + 0.263 + ], + "angle": 0, + "content": "Xiaobo Zhou\\(^{1,4\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.401, + 0.268, + 0.597, + 0.282 + ], + "angle": 0, + "content": "xiaobo.zhou@tju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.708, + 0.204, + 0.798, + 0.22 + ], + "angle": 0, + "content": "Zhihui Ke" + }, + { + "type": "text", + "bbox": [ + 0.669, + 0.226, + 0.839, + 0.239 + ], + "angle": 0, + "content": "kezhihui@tju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.685, + 0.246, + 0.766, + 0.264 + ], + "angle": 0, + "content": "Tie Qiu\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.65, + 0.268, + 0.802, + 0.282 + ], + "angle": 0, + "content": "qiutie@tju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.289, + 0.918, + 0.309 + ], + "angle": 0, + "content": "\\(^{1}\\) Tianjin University \\(^{2}\\) Shanghai Artificial Intelligence Laboratory \\(^{3}\\) Tencent \\(^{4}\\) Qinghai Minzu University" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.314, + 0.249, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.504, + 0.171, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.564, + 0.238, + 0.575 + ], + "angle": 0, + "content": "DeformGS PSNR:27.53" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.314, + 0.408, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.504, + 0.331, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.564, + 0.401, + 0.575 + ], + "angle": 0, + "content": "+TimeFormer PSNR:29.34" + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.315, + 0.645, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.479, + 0.434, + 0.582, + 0.443 + ], + "angle": 0, + "content": "STGS PSNR: 27.28" + }, + { + "type": "image", + "bbox": [ + 0.413, + 0.443, + 0.645, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.564, + 0.609, + 0.574 + ], + "angle": 0, + "content": "+TimeFormer PSNR:27.98" + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.314, + 0.728, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.376, + 0.727, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.442, + 0.728, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.504, + 0.727, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.739, + 0.315, + 0.904, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.579, + 0.908, + 0.622 + ], + "angle": 0, + "content": "Figure 1. We propose TimeFormer, a Transformer module that implicitly models the motion pattern via Temporal Attention from a learning perspective (right). TimeFormer is plug-and-play to existing deformable 3D Gaussian reconstruction methods [27, 60, 66] and enhances reconstruction results (left and middle) without affecting their original inference speed." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.639, + 0.327, + 0.654 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.671, + 0.486, + 0.867 + ], + "angle": 0, + "content": "Dynamic scene reconstruction is a long-term challenge in 3D vision. Recent methods extend 3D Gaussian Splating to dynamic scenes via additional deformation fields and apply explicit constraints like motion flow to guide the deformation. However, they learn motion changes from individual timestamps independently, making it challenging to reconstruct complex scenes, particularly when dealing with violent movement, extreme-shaped geometries, or reflective surfaces. To address the above issue, we design a simple yet effective plug-and-play module called TimeFormer to enable existing deformable 3D Gaussians reconstruction methods with the ability to implicitly model motion patterns from a learning perspective. Specifically," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.909, + 0.792 + ], + "angle": 0, + "content": "TimeFormer includes a Cross-Temporal Transformer Encoder, which adaptively learns the temporal relationships of deformable 3D Gaussians. Furthermore, we propose a two-stream optimization strategy that transfers the motion knowledge learned from TimeFormer to the base stream during the training phase. This allows us to remove TimeFormer during inference, thereby preserving the original rendering speed. Extensive experiments in the multi-view and monocular dynamic scenes validate qualitative and quantitative improvement brought by TimeFormer." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.828, + 0.646, + 0.843 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "High-quality reconstruction of dynamic scenes is significantly challenging in computer vision and graphics, yet has a wide range of potential applications in movie production," + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.876, + 0.248, + 0.888 + ], + "angle": 0, + "content": "† Corresponding Author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.401, + 0.901 + ], + "angle": 0, + "content": "Project Page: https://patrickddj.github.io/TimeFormer" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.876, + 0.401, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8721" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "virtual reality, and augmented reality. The difficulty stems from factors like occlusions, translucent materials, specular surfaces, and changing topology, all of which are prevalent in dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.483, + 0.382 + ], + "angle": 0, + "content": "Inspired by the success of neural radiance field (NeRF) on static scenes [3, 36, 37], extending NeRF to dynamic scenes [15, 24, 26, 30, 38, 39, 50, 71] has been explored. However, these NeRF-based methods are limited by computationally intensive volume rendering [7], which makes real-time rendering almost impossible. Recently, 3D Gaussian Splatting (3DGS) [22] represents the scene with anisotropic 3D Gaussians and develops a rasterization-based rendering algorithm, which allows real-time rendering through directly projecting 3D Gaussians onto the image plane. Following the revolutionized 3DGS, dynamic Gaussian Splatting methods [9, 27, 48, 67] have been introduced for reconstructing dynamic scenes. These methods typically construct a canonical 3DGS and utilize a deformation field based on individual timestamps [27, 60, 66]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.626 + ], + "angle": 0, + "content": "However, temporal relationships of 3D Gaussians have been poorly investigated. Previous studies have primarily modeled motion patterns using various types of the deformation field, for example, MLPs [2, 19, 33, 34, 43, 53, 66, 70], spatial-temporal planes [10, 32, 60], polynomial functions [27], Fourier series [21], and combinations of these methods [28]. These methods learn motion patterns from independent time input in a vanilla way, neglecting internal cross-time relationships. Some studies introduce the motion flow regularization [23] to explicitly learn motion patterns from neighboring frames [32, 34]. Although these methods promote similar motion between adjacent timestamps, they adopt a local perspective on time series during optimization. This constraint makes it challenging to reconstruct scenes with more complex motion patterns, such as sudden appearances, violent movements, or reflective surfaces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce TimeFormer, a transformer module designed to implicitly learn motion patterns across multiple timestamps. By modeling temporal relationships within a time batch, we aim to provide a global view of the entire time series, enabling the deformation field backbones themselves to capture motion patterns from a learning perspective, as shown in Fig. 1. Specifically, TimeFormer utilizes a Cross-Temporal Encoder to capture the implicit motion patterns of 3D Gaussians across multiple sampled timestamps using a self-attention mechanism. Moreover, to avoid additional computational costs of TimeFormer during inference, we present a two-stream optimization strategy. By sharing the weights of two deformation fields, we transfer the motion pattern learning from TimeFormer stream to the base stream during training. Therefore, we can eliminate TimeFormer during inference and maintain the original rendering speed. Notably, TimeFormer does not require any prior information and extracts motion patterns solely from" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.098, + 0.697, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.162, + 0.678, + 0.172 + ], + "angle": 0, + "content": "Points: \\(154.5\\mathrm{k}\\) FPS:31.3" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.088, + 0.887, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.161, + 0.871, + 0.172 + ], + "angle": 0, + "content": "Points: \\(63.6\\mathrm{k}\\) FPS: 79.7" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.177, + 0.703, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.265, + 0.674, + 0.276 + ], + "angle": 0, + "content": "DeformGS PSNR:27.17" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.176, + 0.905, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.266, + 0.876, + 0.276 + ], + "angle": 0, + "content": "+TimeFormer PSNR:28.70" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.281, + 0.905, + 0.323 + ], + "angle": 0, + "content": "Figure 2. TimeFormer guides towards more efficiently distributed canonical space, showing higher FPS and better quality. The results are from \"cut lemon\" in the HyperNeRF Dataset [39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.905, + 0.409 + ], + "angle": 0, + "content": "RGB supervision, which can be seamlessly adapted to previous deformable 3D Gaussian methods in a plug-and-play manner. Furthermore, TimeFormer promotes faster gradient descent and guides a more efficient canonical space, ultimately increasing FPS during inference, as in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.411, + 0.862, + 0.425 + ], + "angle": 0, + "content": "To sum up, our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.427, + 0.905, + 0.487 + ], + "angle": 0, + "content": "- We propose TimeFormer, a transformer module that enhances current deformable 3D Gaussians reconstruction methods in a plug-and-play manner from an automatic learning perspective." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.488, + 0.905, + 0.533 + ], + "angle": 0, + "content": "- The two-stream optimization strategy allows the exclusion of TimeFormer during inference while maintaining and even improving rendering speed." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.533, + 0.905, + 0.579 + ], + "angle": 0, + "content": "- Extensive experiments on real-world datasets validate the effectiveness of TimeFormer, achieving state-of-the-art rendering quality." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.427, + 0.905, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.595, + 0.663, + 0.61 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.788, + 0.637 + ], + "angle": 0, + "content": "2.1. Dynamic Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Dynamic Scene Reconstruction has been extensively researched over many years, with a wealth of studies [20, 39, 65, 66, 73] contributing to current progress. The pioneering neural radiance field [3, 12, 36, 47, 68] has demonstrated photorealistic rendering for novel view synthesis from calibrated multiview images, inspiring extensive approaches [8, 11, 38, 39, 41, 51, 62] to extend NeRF to 4D space-time field for dynamic scene reconstruction. There are majorly two lines of NeRF-based methods for dynamic scene reconstruction: 1) deformation-based methods [11, 38, 39, 41, 51] model deformation changes by using a deformation field to map the queried positions in different timestamps to a canonical space; 2) compact 4D space-time fields [5, 8, 25, 62] take the timestamps, positions, and directions as input to predict color and density. However, NeRF-based approaches are limited to training and rendering speed. Thus, a large number of meth-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8722" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "ods [1, 4, 13, 24, 29, 31, 42, 46, 55-57, 61, 63] have been proposed to accelerate NeRF-based methods." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.124, + 0.484, + 0.441 + ], + "angle": 0, + "content": "Recent 3DGS [22] achieves real-time rendering and dynamic 3DGS approaches [27, 66, 67] emerge. In details, deformation-based methods [2, 10, 19, 21, 27, 28, 32-34, 43, 53, 60, 66, 70] employ a deformation field to predict per-Gaussian offsets and deform a canonical 3DGS according to different timestamps. Meanwhile, 4DGS [67] and 4D-Rotor [9] add the time dimension to 3D Gaussian, forming a 4D Gaussian representation. TimeFormer focuses on modeling the temporal relationship and is plug-and-play to deformation-based methods [66] and 4D space-time representations [27, 67]. There are also a few attempts [28, 43, 48] in leveraging the nearby motion or flow for dynamic Gaussian reconstruction. For example, 3DGStream [48] proposes a per-frame optimization strategy, which predicts current Gaussian attributes based on previous Gaussian attributes. Meanwhile, Gaussian-Flow [28] presents explicitly model time-dependent residual of each attribute. To model a long video sequence, SWinGS [43] splits a video sequence into multiple sliding windows based on optical flow and uses a deformation-based method to model each sliding window." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.262, + 0.47 + ], + "angle": 0, + "content": "2.2. Motion Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Motion prediction in dynamic scene reconstruction from multi-view videos or monocular videos is an ill-posed problem. Early dynamic NeRF [15, 38, 39, 51] directly learns motion patterns from individual timestamps, and several later works propose motion flow regularization terms [15, 26, 30, 35, 40, 44, 45, 50, 54, 69] to promote the learning of cross-time motion patterns. These methods typically utilize 2D prior information (e.g., optical flow) from pre-trained networks [49] to supervise the scene flows within neighboring frames. PREF [45] uses motion predictor to infer current motion based on the previous four frames and propose a self-supervision strategy without optical flow prior. KFD-NeRF [69] models motion patterns by Kalman Filter based on the previous two frames. CT-NeRF [35] employs a cross-attention mechanism from transformer [52] to learn the correlation of conservative frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Dynamic 3D Gaussian Splatting (3DGS) methods [16, 17, 58, 72] also utilize optical flow supervision to enhance reconstruction performance. Techniques such as MD-Splatting [10], D3DG [34], and ST-4DGS [23] focus on minimizing both forward and backward Gaussian flow to promote temporal smoothness in Gaussian motions. DN4DGS [32] incorporates information from previous, current, and next timestamps into the deformation field to capture cross-time motion patterns. However, these methods primarily establish temporal correlations with neighboring timestamps, which limits their ability to address long-term motion changes. Additionally, they introduce extra compu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "tational costs during inference, which can reduce rendering speed. In contrast, TimeFormer captures the temporal relationships of Gaussians from a global perspective across the entire time series. Notably, TimeFormer is employed only during the training phase, without any additional computational costs during inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.195, + 0.871, + 0.212 + ], + "angle": 0, + "content": "3. Preliminary: Deformable 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.22, + 0.905, + 0.295 + ], + "angle": 0, + "content": "3D Gaussians [22] represent the scene with a set of 3D Gaussians, each of which has unique opacity \\(o \\in [0,1]\\), center position \\(\\mu \\in \\mathbb{R}^{3 \\times 1}\\), and covariance matrix \\(\\Sigma \\in \\mathbb{R}^{3 \\times 3}\\). For a position \\(x \\in \\mathbb{R}^{3 \\times 1}\\) in 3D space, the corresponding contribution of a 3D Gaussian on it can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.301, + 0.905, + 0.322 + ], + "angle": 0, + "content": "\\[\nG (x) = o \\cdot e ^ {- \\frac {1}{2} (x - \\mu) ^ {\\top} \\Sigma^ {- 1} (x - \\mu)}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.907, + 0.483 + ], + "angle": 0, + "content": "The covariance matrix \\(\\Sigma\\) can be decomposed into a scaling matrix \\(\\mathbf{S}\\) and a rotation matrix \\(\\mathbf{R}\\): \\(\\boldsymbol{\\Sigma} = \\mathbf{RSS}^T\\mathbf{R}^T\\), where \\(\\mathbf{S} = \\mathrm{diag}([s_x, s_y, s_z])\\) and \\(\\mathbf{R}\\) can be transformed from a quaternion \\([r_w, r_x, r_y, r_z]\\). Then the 3D Gaussians can be splatted to a 2D camera plane through differential Gaussian splatting. To model the appearance of 3D Gaussians, spherical harmonics (SH) are introduced to define the color \\(c\\). Finally, for each pixel, the rendering results of 3DGS can be derived by calculating the color contribution of all the related Gaussians. This process is known as \\(\\alpha\\)-blending:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.49, + 0.905, + 0.533 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i} ^ {N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.905, + 0.57 + ], + "angle": 0, + "content": "where \\(c_{i},\\alpha_{i}\\) represent the color and density computed from the \\(i\\) -th 3D Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.906, + 0.661 + ], + "angle": 0, + "content": "Recent methods [27, 60, 66] utilize deformation fields that extend 3DGS to 4D space, inspired from NeRF-based methods such as D-NeRF [41]. The structure of the deformation field \\(\\mathcal{D}\\) can vary among MLP [66], K-Plane [60] and Polynomial [27], while the deformation process can be summarized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.671, + 0.905, + 0.689 + ], + "angle": 0, + "content": "\\[\n(\\Delta \\mu , \\Delta r, \\Delta s) = \\mathcal {D} (\\mu , t), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.906, + 0.758 + ], + "angle": 0, + "content": "where timestamp \\(t \\in \\mathcal{T}\\), \\(\\mathcal{T} \\in \\mathbb{R}^{T \\times 1}\\) contains \\(T\\) linear time inputs, \\(\\mu, r, s\\) are the center position, rotation quaternion, and scaling factors of 3D Gaussians, and \\(\\Delta \\mu, \\Delta r, \\Delta s\\) are their residuals, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.605, + 0.785 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this section, we first provide an overview of the enhanced deformable 3D Gaussians reconstruction with the proposed TimeFormer (Sec. 4.1). We then introduce the TimeFormer in detail (Sec. 4.2) which consists of a Cross-Temporal Encoder and shared deformation fields (Sec. 4.3). We also discuss the insights behind TimeFormer from the view of gradient flow." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8723" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.91, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.258, + 0.908, + 0.314 + ], + "angle": 0, + "content": "Figure 3. The Framework of Deformable 3D Gaussians Reconstruction with TimeFormer. Existing deformable 3D Gaussians framework usually includes the canonical space and the deformation field (first row), we incorporate TimeFormer to capture cross-time relationships and explore motion patterns implicitly (second row). We share weights of two deformation fields to transfer the learned motion knowledge. This allows us to exclude this Auxiliary Training Module during inference." + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.321, + 0.415, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.504, + 0.484, + 0.574 + ], + "angle": 0, + "content": "Figure 4. The Structure of Cross-Temporal Encoder. We concatenate randomly sampled timestamps to position \\(\\mu\\), treating them as special tokens in a sequence. This module is designed to model multi-temporal relationships and produce distinct time-variant position offsets \\(\\Delta p_0, \\ldots, \\Delta p_{B-1}\\)." + }, + { + "type": "image", + "bbox": [ + 0.15, + 0.579, + 0.429, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.672 + ], + "angle": 0, + "content": "Figure 5. Data Flow Changes in the Deformation Field. Dashed lines represent new data flow among time samples \\( t_0, \\dots, t_{B-1} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.203, + 0.696 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Previous methods model motion patterns by explicitly learning temporal relationships on individual or neighboring timestamps, failing on those complex scenes containing violent movement or dynamic reflective surfaces. In contrast, we present TimeFormer to enable the deformable 3D Gaussian backbones themselves to model cross-time relationships from an implicit learning perspective. The main framework with the proposed TimeFormer is shown in Fig. 3. Our approach retains standard reconstruction modules, which include (1) 3D Gaussians in the canonical space and (2) a deformation field that applies time-variant transformation. Additionally, TimeFormer is introduced before the deformation field to extract implicit cross-time motion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.907, + 0.414 + ], + "angle": 0, + "content": "features for each Gaussian through a self-attention mechanism along the time dimension. Moreover, we share the weights of two deformation fields to transfer the motion knowledge from TimeFormer to mitigate the gap between the original branch and TimeFormer branch, which supports real-time rendering without TimeFormer during inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.425, + 0.651, + 0.44 + ], + "angle": 0, + "content": "4.2. TimeFormer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.909, + 0.678 + ], + "angle": 0, + "content": "Cross-Temporal Encoder. BatchFormer [18] demonstrates that the attention mechanism helps learn sample relationships from batch dimension, rather than channel and spatial dimensions [6, 52], inspiring us that different timestamps can also be considered as a special time batch. Let \\(\\mathcal{T}_s\\subset \\mathcal{T},\\mathcal{T}_s = \\{t_i\\}_{i = 0}^{B - 1}\\) denotes randomly sampled timestamps, and let \\(\\mathcal{G}\\in \\mathbb{R}^{N\\times (3 + C)}\\) denotes Gaussians in the canonical space, where \\(B\\) is the size of time batch, \\(N\\) is the number of Gaussians, \\(3 + C\\) means each Gaussian has 3 position channels and \\(C\\) additional channels. As in Fig. 4, all Gaussians's positions are made into \\(B\\) copies, expanded into \\(\\mathcal{G}_c\\in \\mathbb{R}^{B\\times N\\times 3}\\), and sampled timestamps are made into \\(N\\) copies, expanded into \\(\\mathcal{T}'\\in \\mathbb{R}^{B\\times N\\times 1}\\). Then, we compose \\(\\mathcal{G}_c\\) and \\(\\mathcal{T}'\\) together and apply position encoding function \\(\\gamma\\) to extract high frequency information, as in Eq. 4:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.688, + 0.905, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\gamma (p) = \\left(\\sin \\left(2 ^ {0} \\pi p\\right), \\cos \\left(2 ^ {0} \\pi p\\right), \\dots \\right. \\\\ \\sin \\left(2 ^ {L - 1} \\pi p\\right), \\cos \\left(2 ^ {L - 1} \\pi p\\right)) \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.907, + 0.872 + ], + "angle": 0, + "content": "In our experiments, we set \\( L = 6 \\) for both positions \\( x \\) and time \\( t \\). We treat \\( [\\gamma(\\mathcal{G}_c, \\gamma(\\mathcal{T}')] \\in \\mathbb{R}^{B \\times N \\times (3 \\times 2L + 2L)} \\) as the original input \\( F_0 \\) to TimeFormer. It can be considered as \\( N \\) sequences of the length \\( B \\), containing \\( 8L \\) feature channels. With \\( M \\) transformer encoder layers, for \\( m^{th} \\) layer, intermediate features are calculated through multi-head self-attention(MSA) and MLP blocks. In the final stage, we use a tiny MLP to transform the last encoded features \\( F_{M-1} \\) into offset \\( \\mathcal{O} \\in \\mathbb{R}^{B \\times N \\times 3} \\) in the linear space:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.885, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {O} = M L P \\left(F _ {M - 1}\\right), \\quad \\mathcal {G} _ {t} = \\mathcal {G} _ {c} + \\mathcal {O} \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8724" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.197 + ], + "angle": 0, + "content": "We consider the output from the Cross-Temporal Encoder as a fixing residual term to the original positions to encourage a gradual, steady learning process on motion patterns. Implicit cross-time relationship learning in TimeFormer enables the automatic aggregation of Gaussians with similar variations during optimization, promoting more efficient spatial distribution and accelerating rendering speed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.482, + 0.257 + ], + "angle": 0, + "content": "Gradient Analysis. Let \\(\\mathcal{D}\\) and \\(\\mathcal{P}\\) be the deformation field and TimeFormer respectively, the output of the deformation field \\(\\Delta \\mu_{i}\\) and partial derivative for \\(\\mu\\) in time \\(t_i\\in \\mathcal{T}_s\\) are formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.266, + 0.483, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mu_ {i} = \\mathcal {D} (\\mu , t _ {i}), \\quad \\frac {\\partial \\Delta \\mu_ {i}}{\\partial \\mu} = \\frac {\\partial \\mathcal {D} (\\mu , t _ {i})}{\\partial \\mu} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.482, + 0.34 + ], + "angle": 0, + "content": "With TimeFormer applied on time batch \\(\\mathcal{T}_s\\), we reformulate as Eq. 7 and data flow changes as in Fig. 5." + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.352, + 0.483, + 0.369 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mu_ {i} = \\mathcal {D} (\\mu + \\mathcal {P} (\\mu , \\mathcal {T} _ {s}), t _ {i}) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.411 + ], + "angle": 0, + "content": "To calculate the partial derivative for \\(\\mu\\), we apply the chain rule. Let \\(a = \\mu + \\mathcal{P}(\\mu, \\mathcal{T}_s)\\), then:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.421, + 0.483, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\Delta \\mu_ {i}}{\\partial \\mu} = \\frac {\\partial \\mathcal {D} (a , t _ {i})}{\\partial a} \\cdot \\frac {\\partial a}{\\partial \\mu}, \\quad \\frac {\\partial a}{\\partial \\mu} = 1 + \\frac {\\partial \\mathcal {P} (\\mu , \\mathcal {T} _ {s})}{\\partial \\mu} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.469, + 0.483, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\frac {\\partial \\Delta \\mu_ {i}}{\\partial \\mu} = \\frac {\\partial \\mathcal {D} (a , t _ {i})}{\\partial a} \\cdot \\left(1 + \\frac {\\partial \\mathcal {P} (\\mu , \\mathcal {T} _ {s})}{\\partial \\mu}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.661 + ], + "angle": 0, + "content": "Eq. 6 has a weaker dynamic nature, as it only considers the current timestamp \\( t_i \\). In contrast, Eq. 9 incorporates an additional gradient term, \\( \\frac{\\partial \\mathcal{P}(\\mu, \\mathcal{T}_s)}{\\partial \\mu} \\), enabling the current state to be influenced by any past or future states. In other words, \\( \\Delta \\mu_i \\) also optimizes the models according to other timestamps \\( t_j \\in \\mathcal{T}_s (j \\neq i) \\), which is a significant difference compared to the backward process without Time-Former. Such design accounts for cross-time attention and allows the models to capture more challenging motion patterns from a global view of the entire time series." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.671, + 0.336, + 0.687 + ], + "angle": 0, + "content": "4.3. Shared Deformation Fields" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.694, + 0.483, + 0.754 + ], + "angle": 0, + "content": "Accounting for additional computation costs of Time-Former, which can significantly decrease rendering speed, we force the original deformation field and the auxiliary deformation field to share weights for knowledge transferring." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.483, + 0.845 + ], + "angle": 0, + "content": "Let \\(\\mathcal{V}\\) be the camera viewpoints, we apply the shared deformation fields to predict deformed space from both Gaussians in the canonical space and Gaussians from Time-Former. Then, we apply the splatting algorithm to these two groups of deformed space. We calculate the losses between rendered images and ground truth \\(\\mathcal{T}_{gt}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.857, + 0.483, + 0.874 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\left\\| S p l a t t i n g \\left(\\mathcal {D} \\left(\\mathcal {G} _ {c}, \\mathcal {T}\\right), \\mathcal {V}\\right) - \\mathcal {I} _ {g t} \\right\\| _ {1} \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = \\| S p l a t t i n g (\\mathcal {D} (\\mathcal {G} _ {t}, \\mathcal {T}), \\mathcal {V}) - \\mathcal {I} _ {g t} \\| _ {1} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.092, + 0.905, + 0.107 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {t} \\mathcal {L} _ {t} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.175 + ], + "angle": 0, + "content": ", where \\(\\mathcal{L}_c, \\mathcal{L}_t\\) represent losses of original branch and Time-Former branch with \\(\\lambda_c > \\lambda_t\\). We use a relatively smaller \\(\\lambda_t\\) because we find it easy to overfit on the second branch with TimeFormer, causing a degradation in inference quality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.188, + 0.64, + 0.206 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.213, + 0.731, + 0.229 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.235, + 0.905, + 0.341 + ], + "angle": 0, + "content": "Our implementation is tested on a single A100 GPU. We use \\( M = 4 \\) transformer encoder layers in TimeFormer and the number of time samples is \\( B = 4 \\). We set \\( \\lambda_c = 1 \\) and \\( \\lambda_t = 0.8 \\) to prevent overfitting on TimeFormer. We assess our experimental results using image quality metrics, including peak-signal-to-noise ratio (PSNR) and structural similarity index (SSIM [59])." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.906, + 0.447 + ], + "angle": 0, + "content": "Baselines & Datasets. We apply TimeFormer to 4DGS [60], STGS [27] on multi-view videos, e.g., N3DV Dataset [25], and evaluate TimeFormer on 4DGS [60] and DeformGS [66] on monocular dynamic videos, e.g., HyperNeRF Dataset [39], respectively. We also utilize NeRF-DS Dataset [64] to demonstrate the robustness of TimeFormer on moving objects with specular surfaces." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.456, + 0.762, + 0.472 + ], + "angle": 0, + "content": "5.2. Experimental Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.583 + ], + "angle": 0, + "content": "Multi-View Video Dataset. In Tab. 1, TimeFormer improves the reconstruction quality (e.g., PSNR) of original 4DGS [60] and STGS [27] by 0.74 and 0.61, outperforming all baselines methods. We extract the frames at an initial time for scene Cook Spanish in Fig. 6, and TimeFormer can reconstruct a clearer geometry and specular effects compared to the original results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.704 + ], + "angle": 0, + "content": "To have a more comprehensive overview of Time-Former's capacity on different temporal stages, we collect PSNR for 300 frames of all six scenes from N3DV Dataset [25] and calculate the average per-frame PSNR. As in Fig. 7, the reconstruction quality increases more significantly at the beginning and end, and such balanced reconstruction results are attributed to TimeFormer's cross-temporal attention mechanism on the whole time series." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.854 + ], + "angle": 0, + "content": "Monocular Video Dataset. Tab. 3 shows the enhancements achieved by TimeFormer on the HyperNeRF Dataset. It's observed that TimeFormer increases the PSNR by 0.89 for 4DGS [60] and 0.94 for DeformGS [66], and the SSIM by 0.019 and 0.028, respectively. Fig. 9 additionally provides visualization of the improvement in the image quality: on the left, TimeFormer generates a clearer outline between fingers, while on the right the texture on the front of the broom is clearer. TimeFormer also eliminates artifacts and reveals a clear structure of the broom." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Dynamic Reflective Video Dataset. Fig. 10 additionally demonstrates that TimeFormer can work well on scenes with dynamic specular objects on NeRF-DS Dataset [64]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8725" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.088, + 0.912, + 0.248 + ], + "angle": 0, + "content": "
MethodSear SteakFlame SSIMSalmon SSIMCut Beef SSIMFlame SSINRSteak SSIMCook Spinach SSINRSSIMCoffee SSINRMartini SSIMMean
PSNRSSIMPSNRSSIM
K-Plane [14]32.520.97130.440.94231.820.96532.380.97030.600.96829.990.94331.63
MixVovels [55]31.210.97129.920.94531.300.96531.430.97031.610.96529.360.94630.80
GS4D [67]32.920.95326.390.89733.080.95933.810.96732.770.95625.230.88430.07
4D-Rotor [9]32.860.95628.250.91333.140.95231.610.95332.560.94927.950.90831.06
4DGS [60]32.490.94928.920.91732.900.95632.510.95432.460.94827.340.90331.10
+TimeFormer33.380.95529.330.92433.110.95733.250.95333.030.94928.930.91031.84
STGS [27]33.710.96228.210.92133.520.95833.460.96333.130.95527.710.91531.62
+TimeFormer34.340.96529.130.92433.570.95834.040.96433.450.95628.830.91732.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.251, + 0.908, + 0.281 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparisons on N3DV Dataset [25]. We use bold font to indicate the improvement, statistics of K-Plane [14] and MixVoxels [55] are from their original paper, while we calculate metrics of all Gaussian-based methods by running their official codes." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.293, + 0.255, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.293, + 0.416, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.293, + 0.579, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.293, + 0.742, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.293, + 0.906, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.309, + 0.472, + 0.687, + 0.487 + ], + "angle": 0, + "content": "Figure 6. Visualization of Comparisons on N3DV Dataset [25]." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.509, + 0.466, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Figure 7. Per-Frame PSNR on N3DV Dataset [25]. TimeFormer improves the reconstruction quality, especially at the beginning and end time series." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.779 + ], + "angle": 0, + "content": "TimeFormer can reconstruct a cleaner appearance on reflective material (e.g., a metallic cup, a glass bottle), and this improvement is attributed to the attention mechanism which enables the deformation field to detect changes on specular surfaces from a learning perspective automatically." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Analysis of FPS & Canonical Space. We also find that TimeFormer achieves more efficient representation with much fewer points in the final canonical space, thereby achieving a higher FPS in Tab. 2. We attribute the phenomenon to the reason that TimeFormer guides faster gradient descent and promotes the spatial layout of Gaussian points during optimization. Interestingly, in Fig. 8, the L1 loss of both branches in our method decreases" + }, + { + "type": "image", + "bbox": [ + 0.55, + 0.518, + 0.871, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.658, + 0.908, + 0.699 + ], + "angle": 0, + "content": "Figure 8. Comparisons of Convergence Speed. We use the same batch size for 4DGS [60] and calculate the loss of both branches in our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.732, + 0.907, + 0.838 + ], + "angle": 0, + "content": "faster and shows almost the same convergence trend during training, exceeding the optimization speed of the original method. This eliminates many redundant points and guides the canonical space toward a more efficient distribution, as in Fig. 2. Moreover, Fig. 8 also proves that the cross-time relationship learned in TimeFormer has been successfully transferred to the base branch, as introduced in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Analysis of Motion Patterns. We argue that TimeFormer achieves more robust learning of motion patterns. To take a deeper insight, we visualize the motion changes as follows: for each Gaussian, we calculate its bias \\((\\Delta_x, \\Delta_y, \\Delta_z)\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8726" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.172, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.2, + 0.17, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.256, + 0.165, + 0.275 + ], + "angle": 0, + "content": "DeformGS PSNR:27.63" + }, + { + "type": "image", + "bbox": [ + 0.175, + 0.089, + 0.254, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.2, + 0.252, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.256, + 0.25, + 0.275 + ], + "angle": 0, + "content": "+TimeFormer PSNR:28.59" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.089, + 0.334, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.2, + 0.333, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.256, + 0.326, + 0.275 + ], + "angle": 0, + "content": "4DGS PSNR:28.55" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.089, + 0.416, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.2, + 0.414, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.256, + 0.41, + 0.275 + ], + "angle": 0, + "content": "+TimeFormer PSNR:29.51" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.089, + 0.498, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.2, + 0.495, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.256, + 0.49, + 0.275 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.578, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.2, + 0.576, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.256, + 0.57, + 0.275 + ], + "angle": 0, + "content": "DeformGS PSNR:20.63" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.089, + 0.661, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.2, + 0.658, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.256, + 0.655, + 0.275 + ], + "angle": 0, + "content": "+TimeFormer PSNR:21.13" + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.089, + 0.742, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.2, + 0.739, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.256, + 0.731, + 0.275 + ], + "angle": 0, + "content": "4DGS PSNR:21.28" + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.089, + 0.822, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.2, + 0.821, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.256, + 0.817, + 0.275 + ], + "angle": 0, + "content": "+TimeFormer PSNR:22.23" + }, + { + "type": "image", + "bbox": [ + 0.827, + 0.089, + 0.905, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.829, + 0.2, + 0.901, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.834, + 0.256, + 0.896, + 0.266 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.28, + 0.703, + 0.294 + ], + "angle": 0, + "content": "Figure 9. Visualization of Comparisons on HyperNeRF Dataset [39]." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.307, + 0.273, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.307, + 0.43, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.307, + 0.588, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.307, + 0.747, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.307, + 0.905, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.378, + 0.272, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.378, + 0.43, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.378, + 0.588, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.378, + 0.747, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.378, + 0.905, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.449, + 0.272, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.449, + 0.43, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.449, + 0.588, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.449, + 0.747, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.449, + 0.905, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.528, + 0.907, + 0.556 + ], + "angle": 0, + "content": "Figure 10. Visualization of Comparisons on NeRF-DS Dataset [64]. Compared with original results from DeformGS [66], TimeFormer presents a clearer visual effect on dynamic objects with specular surfaces." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.58, + 0.458, + 0.661 + ], + "angle": 0, + "content": "
MethodN ↓Training ↓FPS ↑
STGS [27]172.1 k41 min82.1
+TimeFormer148.9 k78 min88.9
4DGS [60]145.9 k52 min31.7
+TimeFormer113.1 k94 min37.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.148, + 0.662, + 0.427, + 0.675 + ], + "angle": 0, + "content": "(a) Quantitative comparisons on N3DV Dataset [25]." + }, + { + "type": "table", + "bbox": [ + 0.118, + 0.681, + 0.458, + 0.762 + ], + "angle": 0, + "content": "
MethodN↓Training ↓FPS ↑
DeformGS [66]169.6 k25 min30.1
+TimeFormer82.9 k35 min58.9
4DGS [60]172.3 k32 min35.7
+TimeFormer135.5 k48 min40.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.132, + 0.764, + 0.442, + 0.777 + ], + "angle": 0, + "content": "(b) Quantitative comparisons on HyperNeRF Dataset [39]." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.788, + 0.483, + 0.803 + ], + "angle": 0, + "content": "Table 2. Comparisons of Gaussian Number, Training Time & FPS." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "towards canonical space in time \\( t \\) and replace the color attributes \\( (r, g, b) \\) with the absolute value of this bias: \\( (r, g, b) \\gets (|\\Delta_x|, |\\Delta_y|, |\\Delta_z|) \\). We use the same splatting algorithm to render the accumulated motion bias, transfer" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.584, + 0.851, + 0.598 + ], + "angle": 0, + "content": "ring the rendered result as a heatmap, as in Fig. 11." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.75 + ], + "angle": 0, + "content": "In Fig. 11a, TimeFormer identifies the blade's motion as a rigid deformation, showing a sharper outline and consistent motion, especially as it contacts the initially static lemon. This demonstrates TimeFormer's ability to distinguish between moving and static objects. In Fig. 11b, TimeFormer captures subtle motion on the spray gun and clear flame flickering on the beef. Fig. 11c shows that TimeFormer also detects the violent movement of the elongated broom, where 4DGS [60] fails, proving its robustness in handling extremely geometries." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.679, + 0.772 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Cross-Temporal Encoder. Tab. 5 demonstrates that the performance of TimeFormer is not sensitive to time batch \\( B \\) and transformer encoder layer \\( M \\). In Tab. 5b, we explore two sampling methods: random sampling and continuous sampling of timestamps. Both strategies demonstrate improvements in PSNR and other metrics compared to the baseline. Notably, the random sampling method yields more significant enhancements. We hypothesize that this" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8727" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.088, + 0.905, + 0.187 + ], + "angle": 0, + "content": "
MethodBroomChickenCut LemonTorchcoPeel BananaHandMean
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DeformGS [66]20.740.32226.320.78627.940.71427.40.87726.380.83627.790.7826.090.719
+TimeFormer21.010.32426.550.79230.010.78027.550.88327.240.85229.790.84827.030.747
4DGS[60]21.530.35126.820.79729.720.76327.450.88327.820.84429.520.84127.140.747
+TimeFormer22.840.40127.060.80130.230.77829.480.90228.210.85330.380.86228.030.766
" + }, + { + "type": "table_caption", + "bbox": [ + 0.307, + 0.195, + 0.692, + 0.21 + ], + "angle": 0, + "content": "Table 3. Quantitative Comparisons on HyperNeRF Dataset [39]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.222, + 0.913, + 0.321 + ], + "angle": 0, + "content": "
MethodPressPlateBasinSieveBellCupMean
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
HyperNeRF [39]25.40.87318.10.71420.20.82925.00.90924.00.88424.10.89622.80.851
NeRF-DS [64]26.40.91120.80.86720.30.86826.10.93523.30.87224.50.91623.570.895
DeformGS[66]25.680.86620.820.81219.870.80425.710.88124.920.85424.520.89723.590.852
+TimeFormer26.290.86720.900.81719.930.80526.260.89125.900.87325.160.90324.100.859
" + }, + { + "type": "table_caption", + "bbox": [ + 0.314, + 0.33, + 0.684, + 0.344 + ], + "angle": 0, + "content": "Table 4. Quantitative Comparisons on NeRF-DS Dataset [64]." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.368, + 0.482, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.53, + 0.338, + 0.54 + ], + "angle": 0, + "content": "(a) Cut Lemon" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.542, + 0.482, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.161, + 0.745, + 0.245, + 0.756 + ], + "angle": 0, + "content": "(b) Torchocolate" + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.745, + 0.418, + 0.756 + ], + "angle": 0, + "content": "(c) Broom" + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.772, + 0.471, + 0.788 + ], + "angle": 0, + "content": "Figure 11. Motion Visualization on HyperNeRF Dataset [39]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.806, + 0.483, + 0.852 + ], + "angle": 0, + "content": "phenomenon occurs because the random sampling method allows for more effective propagation of dynamic behavior modeling across the entire time series." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Shared Deformation Fields. In Tab. 5b, the image quality during inference suffers a considerable decrease if we train two deformation fields without sharing their weights." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.368, + 0.916, + 0.447 + ], + "angle": 0, + "content": "
SettingRandom SamplingContinuous Sampling
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
B=225.890.8750.11725.730.8710.124
B=426.060.8710.11925.780.8740.117
B=625.990.870.12325.730.8650.125
" + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.449, + 0.878, + 0.462 + ], + "angle": 0, + "content": "(a) Ablation Studies on Batch Size and Sampling Strategies." + }, + { + "type": "table", + "bbox": [ + 0.568, + 0.469, + 0.871, + 0.579 + ], + "angle": 0, + "content": "
SettingPSNR↑SSIM↑LPIPS↓
Baseline25.440.8670.125
M=125.960.8740.117
M=225.900.8740.120
M=326.050.8760.112
M=426.040.8730.119
w/o Shared24.810.8520.132
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.514, + 0.581, + 0.924, + 0.607 + ], + "angle": 0, + "content": "(b) Ablation Results. \\( M \\) is the number of encoder layers, \"w/o Shared\" means not using shared weights in two deformation fields." + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.618, + 0.907, + 0.646 + ], + "angle": 0, + "content": "Table 5. Ablation Results. The results are from three scenes press, sieve and bell on NeRF-DS Dataset [64]." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.581, + 0.924, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.659, + 0.907, + 0.69 + ], + "angle": 0, + "content": "Fig. 8 also demonstrates the effectiveness of the two-stream strategy and TimeFormer's ability of re-balancing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.708, + 0.634, + 0.723 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We propose TimeFormer, a Transformer module which enhances existing deformable 3D Gaussians methods by implicitly modeling motion patterns from a learning perspective. In addition, we design a two-stream optimization strategy to transfer the learned motion knowledge from TimeFormer to the original deformation branch. This allows us to remove TimeFormer during inference and thus maintain the same inference speed. Extensive experiments demonstrate that TimeFormer facilitates the reconstruction of three state-of-the-art methods among three datasets, and illustrate the improvement on reconstructing complex scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8728" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This work is completed during the author's internship at the Tencent Hunyuan 3D Department. This work is supported in part by the National Science Fund for Distinguished Young Scholars (No. 62325208), in part by the National Natural Science Foundation of China (No. 62232002), and in part by the Natural Science Foundation of Tianjin (No. 23ZGZNGX00020 and 24JCZXJC00050)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.237, + 0.188, + 0.254 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.262, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.347, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[2] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. arXiv preprint arXiv:2404.03613, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.484, + 0.487 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855-5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.484, + 0.544 + ], + "angle": 0, + "content": "[4] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.484, + 0.686 + ], + "angle": 0, + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[7] Robert A Drebin, Loren Carpenter, and Pat Hanrahan. Volume rendering. ACM Siggraph Computer Graphics, 22(4): 65-74, 1988. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.732, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[8] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. In Proc. SIGGRAPH, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Mike Zheng Shou, Shuran Song, and Jeffrey Ichnowski. Md-splatting: Learning metric deformation from" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.262, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[11] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.906, + 0.333 + ], + "angle": 0, + "content": "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[14] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[15] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[16] Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wenchao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splitting gaussian dynamics for 4d content creation. arXiv preprint arXiv:2403.12365, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[17] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. arXiv preprint arXiv:2403.11447, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.906, + 0.603 + ], + "angle": 0, + "content": "[18] Zhi Hou, Baosheng Yu, and Dacheng Tao. Batchformer: Learning to explore sample relationships for robust representation learning. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.906, + 0.674 + ], + "angle": 0, + "content": "[19] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4220-4230, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[20] Takeo Kanade, Peter Rander, and PJ Narayanan. Virtualized reality: Constructing virtual worlds from real scenes. IEEE multimedia, 4(1):34-47, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[21] Kai Katsumata, Duc Minh Vo, and Hideki Nakayama. A compact dynamic 3d gaussian representation for real-time dynamic view synthesis. In Computer Vision - ECCV 2024, pages 394-412. Springer Nature Switzerland, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[23] Deqi Li, Shi-Sheng Huang, Zhiyuan Lu, Xinran Duan, and Hua Huang. St-4dgs: Spatial-temporally consistent 4d gaussian splatting for efficient dynamic scene rendering. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8729" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[24] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.244 + ], + "angle": 0, + "content": "[25] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[26] Zhengqi Li, Simon Niklaus, Noah Snively, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[27] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8508-8520, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[28] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[29] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. Advances in Neural Information Processing Systems, 35:36762-36775, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[30] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13-23, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[31] Ange Lou, Benjamin Planche, Zhongpai Gao, Yamin Li, Tianyu Luan, Hao Ding, Terrence Chen, Jack Noble, and Ziyan Wu. Darenerf: Direction-aware representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5031-5042, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[32] Jiahao Lu, Jiacheng Deng, Ruijie Zhu, Yanzhe Liang, Wenfei Yang, Tianzhu Zhang, and Xu Zhou. Dn-4dgs: Denoised deformable network with temporal-spatial aggregation for dynamic scene rendering. arXiv preprint arXiv:2410.13607, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[33] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[34] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 2024 International Conference on 3D Vision (3DV), pages 800-809. IEEE, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[35] Xingyu Miao, Yang Bai, Haoran Duan, Fan Wan, Yawen Huang, Yang Long, and Yefeng Zheng. Ctnerf: Cross-time transformer for dynamic neural radiance field from monocular video. Pattern Recognition, 156:110729, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[37] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[38] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[39] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[40] Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahiup Kang. Temporal interpolation is all you need for dynamic neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4212-4221, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[42] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boring Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[43] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Pérez-Pellitero. Swings: sliding windows for dynamic 3d gaussian splatting. In European Conference on Computer Vision. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[44] Nagabhushan Somraj, Kapil Choudhary, Sai Harsha Mupparaju, and Rajiv Soundararajan. Factorized motion fields for fast sparse input dynamic view synthesis. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[45] Liangchen Song, Xuan Gong, Benjamin Planche, Meng Zheng, David Doermann, Junsong Yuan, Terrence Chen, and Ziyan Wu. Pref: Predictability regularized neural motion fields. In European Conference on Computer Vision, pages 664-681. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with de" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8730" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "composed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[47] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.483, + 0.286 + ], + "angle": 0, + "content": "[48] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20675-20685, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.483, + 0.356 + ], + "angle": 0, + "content": "[49] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[50] Fengrui Tian, Shaoyi Du, and Yueqi Duan. Mononerf: Learning a generalizable dynamic radiance field from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17903-17913, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.483, + 0.51 + ], + "angle": 0, + "content": "[51] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.538 + ], + "angle": 0, + "content": "[52] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[53] Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[54] Chaoyang Wang, Lachlan Ewen MacDonald, Laszlo A Jeni, and Simon Lucey. Flow supervision for deformable nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21128-21137, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[55] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, Yafei Song, and Huaping Liu. Mixed neural voxels for fast multiview video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19706-19716, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[56] Feng Wang, Zilong Chen, Guokang Wang, Yafei Song, and Huaping Liu. Masked space-time hash encoding for efficient dynamic scene reconstruction. Advances in Neural Information Processing Systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[57] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 76-87, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[58] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. arXiv preprint arXiv:2405.18426, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[59] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[60] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[61] Minye Wu, Zehao Wang, Georgios Kouros, and Tinne Tuytelaars. Tetrirf: Temporal tri-plane radiance fields for efficient free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6487-6496, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.301, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[62] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9421-9431, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[63] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20029-20040, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.907, + 0.496 + ], + "angle": 0, + "content": "[64] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8285-8295, 2023. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[65] Jason C Yang, Matthew Everett, Chris Buehler, and Leonard McMillan. A real-time distributed light field camera. Rendering Techniques, 2002(77-86):2, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[67] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[68] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[69] Yifan Zhan, Zhuoxiao Li, Muyao Niu, Zhihang Zhong, Shohei Nobuhara, Ko Nishino, and Yinqiang Zheng. Kfdnerf: Rethinking dynamic nerf with kalman filter. arXiv preprint arXiv:2407.13185, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[70] Boming Zhao, Yuan Li, Ziyu Sun, Lin Zeng, Yujun Shen, Rui Ma, Yinda Zhang, Hujun Bao, and Zhaopeng Cui. Gaussian prediction: Dynamic 3d gaussian prediction for motion extrapolation and free view synthesis. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[71] Kaichen Zhou, Jia-Xing Zhong, Sangyun Shin, Kai Lu, Yiyuan Yang, Andrew Markham, and Niki Trigoni. Dynpoint: Dynamic neural point for view synthesis. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8731" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[72] Ruijie Zhu, Yanzhe Liang, Hanzhi Chang, Jiacheng Deng, Jiahao Lu, Wenfei Yang, Tianzhu Zhang, and Yongdong Zhang. Motions: Exploring explicit motion guidance for deformable 3d gaussian splatting. arXiv preprint arXiv:2410.07707, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[73] C Lawrence Zitnick, Sing Bing Kang, Matthew Uytendaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM transactions on graphics (TOG), 23(3):600-608, 2004. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8732" + } + ] +] \ No newline at end of file diff --git a/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/8cdc06bd-0ca5-4f6b-b012-50c8329288a4_origin.pdf b/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/8cdc06bd-0ca5-4f6b-b012-50c8329288a4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..60ec1ff69a85d8ad5493d9e0b4549ed711aa177f --- /dev/null +++ b/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/8cdc06bd-0ca5-4f6b-b012-50c8329288a4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a2a88d2899125960470243ebfd8c708ee6c7af044b63f82c53c6501c8d463704 +size 7376854 diff --git a/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/full.md b/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e8c8dcb15589c7fe09a0973131f56f32303cb23d --- /dev/null +++ b/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/full.md @@ -0,0 +1,485 @@ +# TimeFormer: Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction + +Dadong Jiang $^{1,3}$ + +patrickddj7@gmail.com + +Xianghui Yang3 + +yangxhui6@163.com + +Zhi Hou $^{2\dagger}$ + +zhou9878@uni.sydney.edu.au + +Xiaobo Zhou $^{1,4\dagger}$ + +xiaobo.zhou@tju.edu.cn + +Zhihui Ke + +kezhihui@tju.edu.cn + +Tie Qiu $^{1,4}$ + +qiutie@tju.edu.cn + +$^{1}$ Tianjin University $^{2}$ Shanghai Artificial Intelligence Laboratory $^{3}$ Tencent $^{4}$ Qinghai Minzu University + +![](images/2764ac663fe4ffba03aad8528a90d4a3b8562e60ca0658137ab81572ea2fec83.jpg) + +![](images/069a1eec7ac7d8dc1fb2244ca72e46900c11e67c65fedbbb6fdecc396a1eb5f7.jpg) +DeformGS PSNR:27.53 + +![](images/f757022999d07ed5ad3b24825151ebc029c4d1d53c41ab0783ceed1d888ed3fd.jpg) + +![](images/8ad2ff566c29fd88f2f99299e3c47cbcf7c8094e14821d10e80efcfd52858d5c.jpg) ++TimeFormer PSNR:29.34 + +![](images/c8c76442a8f177dd0ef48b19245ad93ecaa72146f58c98faf504103471b9028a.jpg) +STGS PSNR: 27.28 + +![](images/a376af876f0b331f7d044f378acca83cf4c7f49be271faea76baad2b69927723.jpg) ++TimeFormer PSNR:27.98 + +![](images/b1d41cf214f1312ac1c6c3efbd825eb5c769d445f030b0796e16d415089cf95b.jpg) + +![](images/d847fe0f9b12247ba1025d6a94de0001990026997c34b8c37dff12c15020136c.jpg) + +![](images/e26fe2e99b6de71585fa2f2cae4312fa65c814413e0920ea1fc2ee083618126e.jpg) + +![](images/1278efc39127dce26a8d1b72c57039e1bca6dd6bb13eef32ba6eebee05293336.jpg) + +![](images/3a530fc711584fb47472638fd78cea9077aebeefe035735f96b25c386ac02dfd.jpg) +Figure 1. We propose TimeFormer, a Transformer module that implicitly models the motion pattern via Temporal Attention from a learning perspective (right). TimeFormer is plug-and-play to existing deformable 3D Gaussian reconstruction methods [27, 60, 66] and enhances reconstruction results (left and middle) without affecting their original inference speed. + +# Abstract + +Dynamic scene reconstruction is a long-term challenge in 3D vision. Recent methods extend 3D Gaussian Splating to dynamic scenes via additional deformation fields and apply explicit constraints like motion flow to guide the deformation. However, they learn motion changes from individual timestamps independently, making it challenging to reconstruct complex scenes, particularly when dealing with violent movement, extreme-shaped geometries, or reflective surfaces. To address the above issue, we design a simple yet effective plug-and-play module called TimeFormer to enable existing deformable 3D Gaussians reconstruction methods with the ability to implicitly model motion patterns from a learning perspective. Specifically, + +TimeFormer includes a Cross-Temporal Transformer Encoder, which adaptively learns the temporal relationships of deformable 3D Gaussians. Furthermore, we propose a two-stream optimization strategy that transfers the motion knowledge learned from TimeFormer to the base stream during the training phase. This allows us to remove TimeFormer during inference, thereby preserving the original rendering speed. Extensive experiments in the multi-view and monocular dynamic scenes validate qualitative and quantitative improvement brought by TimeFormer. + +# 1. Introduction + +High-quality reconstruction of dynamic scenes is significantly challenging in computer vision and graphics, yet has a wide range of potential applications in movie production, + +virtual reality, and augmented reality. The difficulty stems from factors like occlusions, translucent materials, specular surfaces, and changing topology, all of which are prevalent in dynamic scenes. + +Inspired by the success of neural radiance field (NeRF) on static scenes [3, 36, 37], extending NeRF to dynamic scenes [15, 24, 26, 30, 38, 39, 50, 71] has been explored. However, these NeRF-based methods are limited by computationally intensive volume rendering [7], which makes real-time rendering almost impossible. Recently, 3D Gaussian Splatting (3DGS) [22] represents the scene with anisotropic 3D Gaussians and develops a rasterization-based rendering algorithm, which allows real-time rendering through directly projecting 3D Gaussians onto the image plane. Following the revolutionized 3DGS, dynamic Gaussian Splatting methods [9, 27, 48, 67] have been introduced for reconstructing dynamic scenes. These methods typically construct a canonical 3DGS and utilize a deformation field based on individual timestamps [27, 60, 66]. + +However, temporal relationships of 3D Gaussians have been poorly investigated. Previous studies have primarily modeled motion patterns using various types of the deformation field, for example, MLPs [2, 19, 33, 34, 43, 53, 66, 70], spatial-temporal planes [10, 32, 60], polynomial functions [27], Fourier series [21], and combinations of these methods [28]. These methods learn motion patterns from independent time input in a vanilla way, neglecting internal cross-time relationships. Some studies introduce the motion flow regularization [23] to explicitly learn motion patterns from neighboring frames [32, 34]. Although these methods promote similar motion between adjacent timestamps, they adopt a local perspective on time series during optimization. This constraint makes it challenging to reconstruct scenes with more complex motion patterns, such as sudden appearances, violent movements, or reflective surfaces. + +In this paper, we introduce TimeFormer, a transformer module designed to implicitly learn motion patterns across multiple timestamps. By modeling temporal relationships within a time batch, we aim to provide a global view of the entire time series, enabling the deformation field backbones themselves to capture motion patterns from a learning perspective, as shown in Fig. 1. Specifically, TimeFormer utilizes a Cross-Temporal Encoder to capture the implicit motion patterns of 3D Gaussians across multiple sampled timestamps using a self-attention mechanism. Moreover, to avoid additional computational costs of TimeFormer during inference, we present a two-stream optimization strategy. By sharing the weights of two deformation fields, we transfer the motion pattern learning from TimeFormer stream to the base stream during training. Therefore, we can eliminate TimeFormer during inference and maintain the original rendering speed. Notably, TimeFormer does not require any prior information and extracts motion patterns solely from + +![](images/735ecea33472d60d7c3725f3768d2de64dc9e8f2c06dca090d6cab986072efaf.jpg) +Points: $154.5\mathrm{k}$ FPS:31.3 + +![](images/f67580d67fc1e2454ab9360a894d963dcda2536e4378dfe2707028f203912bd5.jpg) +Points: $63.6\mathrm{k}$ FPS: 79.7 + +![](images/29581fe486f0f084c13f286aabcf26656abec65d2ae7a2ca2beaa611bb05b393.jpg) +DeformGS PSNR:27.17 + +![](images/d92e509b77958c36d5f22608a21db0bb64a51b1d6d8d9a7ef5ea5706895c0b9c.jpg) ++TimeFormer PSNR:28.70 +Figure 2. TimeFormer guides towards more efficiently distributed canonical space, showing higher FPS and better quality. The results are from "cut lemon" in the HyperNeRF Dataset [39]. + +RGB supervision, which can be seamlessly adapted to previous deformable 3D Gaussian methods in a plug-and-play manner. Furthermore, TimeFormer promotes faster gradient descent and guides a more efficient canonical space, ultimately increasing FPS during inference, as in Fig. 2. + +To sum up, our main contributions are as follows: + +- We propose TimeFormer, a transformer module that enhances current deformable 3D Gaussians reconstruction methods in a plug-and-play manner from an automatic learning perspective. +- The two-stream optimization strategy allows the exclusion of TimeFormer during inference while maintaining and even improving rendering speed. +- Extensive experiments on real-world datasets validate the effectiveness of TimeFormer, achieving state-of-the-art rendering quality. + +# 2. Related Works + +# 2.1. Dynamic Scene Reconstruction + +Dynamic Scene Reconstruction has been extensively researched over many years, with a wealth of studies [20, 39, 65, 66, 73] contributing to current progress. The pioneering neural radiance field [3, 12, 36, 47, 68] has demonstrated photorealistic rendering for novel view synthesis from calibrated multiview images, inspiring extensive approaches [8, 11, 38, 39, 41, 51, 62] to extend NeRF to 4D space-time field for dynamic scene reconstruction. There are majorly two lines of NeRF-based methods for dynamic scene reconstruction: 1) deformation-based methods [11, 38, 39, 41, 51] model deformation changes by using a deformation field to map the queried positions in different timestamps to a canonical space; 2) compact 4D space-time fields [5, 8, 25, 62] take the timestamps, positions, and directions as input to predict color and density. However, NeRF-based approaches are limited to training and rendering speed. Thus, a large number of meth- + +ods [1, 4, 13, 24, 29, 31, 42, 46, 55-57, 61, 63] have been proposed to accelerate NeRF-based methods. + +Recent 3DGS [22] achieves real-time rendering and dynamic 3DGS approaches [27, 66, 67] emerge. In details, deformation-based methods [2, 10, 19, 21, 27, 28, 32-34, 43, 53, 60, 66, 70] employ a deformation field to predict per-Gaussian offsets and deform a canonical 3DGS according to different timestamps. Meanwhile, 4DGS [67] and 4D-Rotor [9] add the time dimension to 3D Gaussian, forming a 4D Gaussian representation. TimeFormer focuses on modeling the temporal relationship and is plug-and-play to deformation-based methods [66] and 4D space-time representations [27, 67]. There are also a few attempts [28, 43, 48] in leveraging the nearby motion or flow for dynamic Gaussian reconstruction. For example, 3DGStream [48] proposes a per-frame optimization strategy, which predicts current Gaussian attributes based on previous Gaussian attributes. Meanwhile, Gaussian-Flow [28] presents explicitly model time-dependent residual of each attribute. To model a long video sequence, SWinGS [43] splits a video sequence into multiple sliding windows based on optical flow and uses a deformation-based method to model each sliding window. + +# 2.2. Motion Modeling + +Motion prediction in dynamic scene reconstruction from multi-view videos or monocular videos is an ill-posed problem. Early dynamic NeRF [15, 38, 39, 51] directly learns motion patterns from individual timestamps, and several later works propose motion flow regularization terms [15, 26, 30, 35, 40, 44, 45, 50, 54, 69] to promote the learning of cross-time motion patterns. These methods typically utilize 2D prior information (e.g., optical flow) from pre-trained networks [49] to supervise the scene flows within neighboring frames. PREF [45] uses motion predictor to infer current motion based on the previous four frames and propose a self-supervision strategy without optical flow prior. KFD-NeRF [69] models motion patterns by Kalman Filter based on the previous two frames. CT-NeRF [35] employs a cross-attention mechanism from transformer [52] to learn the correlation of conservative frames. + +Dynamic 3D Gaussian Splatting (3DGS) methods [16, 17, 58, 72] also utilize optical flow supervision to enhance reconstruction performance. Techniques such as MD-Splatting [10], D3DG [34], and ST-4DGS [23] focus on minimizing both forward and backward Gaussian flow to promote temporal smoothness in Gaussian motions. DN4DGS [32] incorporates information from previous, current, and next timestamps into the deformation field to capture cross-time motion patterns. However, these methods primarily establish temporal correlations with neighboring timestamps, which limits their ability to address long-term motion changes. Additionally, they introduce extra compu + +tational costs during inference, which can reduce rendering speed. In contrast, TimeFormer captures the temporal relationships of Gaussians from a global perspective across the entire time series. Notably, TimeFormer is employed only during the training phase, without any additional computational costs during inference. + +# 3. Preliminary: Deformable 3D Gaussians + +3D Gaussians [22] represent the scene with a set of 3D Gaussians, each of which has unique opacity $o \in [0,1]$ , center position $\mu \in \mathbb{R}^{3 \times 1}$ , and covariance matrix $\Sigma \in \mathbb{R}^{3 \times 3}$ . For a position $x \in \mathbb{R}^{3 \times 1}$ in 3D space, the corresponding contribution of a 3D Gaussian on it can be formulated as: + +$$ +G (x) = o \cdot e ^ {- \frac {1}{2} (x - \mu) ^ {\top} \Sigma^ {- 1} (x - \mu)}. \tag {1} +$$ + +The covariance matrix $\Sigma$ can be decomposed into a scaling matrix $\mathbf{S}$ and a rotation matrix $\mathbf{R}$ : $\boldsymbol{\Sigma} = \mathbf{RSS}^T\mathbf{R}^T$ , where $\mathbf{S} = \mathrm{diag}([s_x, s_y, s_z])$ and $\mathbf{R}$ can be transformed from a quaternion $[r_w, r_x, r_y, r_z]$ . Then the 3D Gaussians can be splatted to a 2D camera plane through differential Gaussian splatting. To model the appearance of 3D Gaussians, spherical harmonics (SH) are introduced to define the color $c$ . Finally, for each pixel, the rendering results of 3DGS can be derived by calculating the color contribution of all the related Gaussians. This process is known as $\alpha$ -blending: + +$$ +C = \sum_ {i} ^ {N} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {2} +$$ + +where $c_{i},\alpha_{i}$ represent the color and density computed from the $i$ -th 3D Gaussian. + +Recent methods [27, 60, 66] utilize deformation fields that extend 3DGS to 4D space, inspired from NeRF-based methods such as D-NeRF [41]. The structure of the deformation field $\mathcal{D}$ can vary among MLP [66], K-Plane [60] and Polynomial [27], while the deformation process can be summarized as follows: + +$$ +(\Delta \mu , \Delta r, \Delta s) = \mathcal {D} (\mu , t), \tag {3} +$$ + +where timestamp $t \in \mathcal{T}$ , $\mathcal{T} \in \mathbb{R}^{T \times 1}$ contains $T$ linear time inputs, $\mu, r, s$ are the center position, rotation quaternion, and scaling factors of 3D Gaussians, and $\Delta \mu, \Delta r, \Delta s$ are their residuals, respectively. + +# 4. Method + +In this section, we first provide an overview of the enhanced deformable 3D Gaussians reconstruction with the proposed TimeFormer (Sec. 4.1). We then introduce the TimeFormer in detail (Sec. 4.2) which consists of a Cross-Temporal Encoder and shared deformation fields (Sec. 4.3). We also discuss the insights behind TimeFormer from the view of gradient flow. + +![](images/5b6bd7ab238578bd016fa8c5cb0ad9395d482c0ac9a297a2a556a4a391767be1.jpg) +Figure 3. The Framework of Deformable 3D Gaussians Reconstruction with TimeFormer. Existing deformable 3D Gaussians framework usually includes the canonical space and the deformation field (first row), we incorporate TimeFormer to capture cross-time relationships and explore motion patterns implicitly (second row). We share weights of two deformation fields to transfer the learned motion knowledge. This allows us to exclude this Auxiliary Training Module during inference. + +![](images/80eb44fe0b95b28ef390966ae250d0731cd90aeacf0c4463daa07356ce68b0fa.jpg) +Figure 4. The Structure of Cross-Temporal Encoder. We concatenate randomly sampled timestamps to position $\mu$ , treating them as special tokens in a sequence. This module is designed to model multi-temporal relationships and produce distinct time-variant position offsets $\Delta p_0, \ldots, \Delta p_{B-1}$ . + +![](images/28776ba2a191a383669d96123c17102aeb2d8a3fa26e7b7d2ed65afbb0d40170.jpg) +Figure 5. Data Flow Changes in the Deformation Field. Dashed lines represent new data flow among time samples $t_0, \dots, t_{B-1}$ . + +# 4.1. Overview + +Previous methods model motion patterns by explicitly learning temporal relationships on individual or neighboring timestamps, failing on those complex scenes containing violent movement or dynamic reflective surfaces. In contrast, we present TimeFormer to enable the deformable 3D Gaussian backbones themselves to model cross-time relationships from an implicit learning perspective. The main framework with the proposed TimeFormer is shown in Fig. 3. Our approach retains standard reconstruction modules, which include (1) 3D Gaussians in the canonical space and (2) a deformation field that applies time-variant transformation. Additionally, TimeFormer is introduced before the deformation field to extract implicit cross-time motion + +features for each Gaussian through a self-attention mechanism along the time dimension. Moreover, we share the weights of two deformation fields to transfer the motion knowledge from TimeFormer to mitigate the gap between the original branch and TimeFormer branch, which supports real-time rendering without TimeFormer during inference. + +# 4.2. TimeFormer + +Cross-Temporal Encoder. BatchFormer [18] demonstrates that the attention mechanism helps learn sample relationships from batch dimension, rather than channel and spatial dimensions [6, 52], inspiring us that different timestamps can also be considered as a special time batch. Let $\mathcal{T}_s\subset \mathcal{T},\mathcal{T}_s = \{t_i\}_{i = 0}^{B - 1}$ denotes randomly sampled timestamps, and let $\mathcal{G}\in \mathbb{R}^{N\times (3 + C)}$ denotes Gaussians in the canonical space, where $B$ is the size of time batch, $N$ is the number of Gaussians, $3 + C$ means each Gaussian has 3 position channels and $C$ additional channels. As in Fig. 4, all Gaussians's positions are made into $B$ copies, expanded into $\mathcal{G}_c\in \mathbb{R}^{B\times N\times 3}$ , and sampled timestamps are made into $N$ copies, expanded into $\mathcal{T}'\in \mathbb{R}^{B\times N\times 1}$ . Then, we compose $\mathcal{G}_c$ and $\mathcal{T}'$ together and apply position encoding function $\gamma$ to extract high frequency information, as in Eq. 4: + +$$ +\begin{array}{l} \gamma (p) = \left(\sin \left(2 ^ {0} \pi p\right), \cos \left(2 ^ {0} \pi p\right), \dots \right. \\ \sin \left(2 ^ {L - 1} \pi p\right), \cos \left(2 ^ {L - 1} \pi p\right)) \tag {4} \\ \end{array} +$$ + +In our experiments, we set $L = 6$ for both positions $x$ and time $t$ . We treat $[\gamma(\mathcal{G}_c, \gamma(\mathcal{T}')] \in \mathbb{R}^{B \times N \times (3 \times 2L + 2L)}$ as the original input $F_0$ to TimeFormer. It can be considered as $N$ sequences of the length $B$ , containing $8L$ feature channels. With $M$ transformer encoder layers, for $m^{th}$ layer, intermediate features are calculated through multi-head self-attention(MSA) and MLP blocks. In the final stage, we use a tiny MLP to transform the last encoded features $F_{M-1}$ into offset $\mathcal{O} \in \mathbb{R}^{B \times N \times 3}$ in the linear space: + +$$ +\mathcal {O} = M L P \left(F _ {M - 1}\right), \quad \mathcal {G} _ {t} = \mathcal {G} _ {c} + \mathcal {O} \tag {5} +$$ + +We consider the output from the Cross-Temporal Encoder as a fixing residual term to the original positions to encourage a gradual, steady learning process on motion patterns. Implicit cross-time relationship learning in TimeFormer enables the automatic aggregation of Gaussians with similar variations during optimization, promoting more efficient spatial distribution and accelerating rendering speed. + +Gradient Analysis. Let $\mathcal{D}$ and $\mathcal{P}$ be the deformation field and TimeFormer respectively, the output of the deformation field $\Delta \mu_{i}$ and partial derivative for $\mu$ in time $t_i\in \mathcal{T}_s$ are formulated as follows: + +$$ +\Delta \mu_ {i} = \mathcal {D} (\mu , t _ {i}), \quad \frac {\partial \Delta \mu_ {i}}{\partial \mu} = \frac {\partial \mathcal {D} (\mu , t _ {i})}{\partial \mu} \tag {6} +$$ + +With TimeFormer applied on time batch $\mathcal{T}_s$ , we reformulate as Eq. 7 and data flow changes as in Fig. 5. + +$$ +\Delta \mu_ {i} = \mathcal {D} (\mu + \mathcal {P} (\mu , \mathcal {T} _ {s}), t _ {i}) \tag {7} +$$ + +To calculate the partial derivative for $\mu$ , we apply the chain rule. Let $a = \mu + \mathcal{P}(\mu, \mathcal{T}_s)$ , then: + +$$ +\frac {\partial \Delta \mu_ {i}}{\partial \mu} = \frac {\partial \mathcal {D} (a , t _ {i})}{\partial a} \cdot \frac {\partial a}{\partial \mu}, \quad \frac {\partial a}{\partial \mu} = 1 + \frac {\partial \mathcal {P} (\mu , \mathcal {T} _ {s})}{\partial \mu} \tag {8} +$$ + +$$ +\frac {\partial \Delta \mu_ {i}}{\partial \mu} = \frac {\partial \mathcal {D} (a , t _ {i})}{\partial a} \cdot \left(1 + \frac {\partial \mathcal {P} (\mu , \mathcal {T} _ {s})}{\partial \mu}\right) \tag {9} +$$ + +Eq. 6 has a weaker dynamic nature, as it only considers the current timestamp $t_i$ . In contrast, Eq. 9 incorporates an additional gradient term, $\frac{\partial \mathcal{P}(\mu, \mathcal{T}_s)}{\partial \mu}$ , enabling the current state to be influenced by any past or future states. In other words, $\Delta \mu_i$ also optimizes the models according to other timestamps $t_j \in \mathcal{T}_s (j \neq i)$ , which is a significant difference compared to the backward process without Time-Former. Such design accounts for cross-time attention and allows the models to capture more challenging motion patterns from a global view of the entire time series. + +# 4.3. Shared Deformation Fields + +Accounting for additional computation costs of Time-Former, which can significantly decrease rendering speed, we force the original deformation field and the auxiliary deformation field to share weights for knowledge transferring. + +Let $\mathcal{V}$ be the camera viewpoints, we apply the shared deformation fields to predict deformed space from both Gaussians in the canonical space and Gaussians from Time-Former. Then, we apply the splatting algorithm to these two groups of deformed space. We calculate the losses between rendered images and ground truth $\mathcal{T}_{gt}$ as follows: + +$$ +\mathcal {L} _ {c} = \left\| S p l a t t i n g \left(\mathcal {D} \left(\mathcal {G} _ {c}, \mathcal {T}\right), \mathcal {V}\right) - \mathcal {I} _ {g t} \right\| _ {1} \tag {10} +$$ + +$$ +\mathcal {L} _ {t} = \| S p l a t t i n g (\mathcal {D} (\mathcal {G} _ {t}, \mathcal {T}), \mathcal {V}) - \mathcal {I} _ {g t} \| _ {1} \tag {11} +$$ + +$$ +\mathcal {L} = \lambda_ {c} \mathcal {L} _ {c} + \lambda_ {t} \mathcal {L} _ {t} \tag {12} +$$ + +, where $\mathcal{L}_c, \mathcal{L}_t$ represent losses of original branch and Time-Former branch with $\lambda_c > \lambda_t$ . We use a relatively smaller $\lambda_t$ because we find it easy to overfit on the second branch with TimeFormer, causing a degradation in inference quality. + +# 5. Experiment + +# 5.1. Implementation Details + +Our implementation is tested on a single A100 GPU. We use $M = 4$ transformer encoder layers in TimeFormer and the number of time samples is $B = 4$ . We set $\lambda_c = 1$ and $\lambda_t = 0.8$ to prevent overfitting on TimeFormer. We assess our experimental results using image quality metrics, including peak-signal-to-noise ratio (PSNR) and structural similarity index (SSIM [59]). + +Baselines & Datasets. We apply TimeFormer to 4DGS [60], STGS [27] on multi-view videos, e.g., N3DV Dataset [25], and evaluate TimeFormer on 4DGS [60] and DeformGS [66] on monocular dynamic videos, e.g., HyperNeRF Dataset [39], respectively. We also utilize NeRF-DS Dataset [64] to demonstrate the robustness of TimeFormer on moving objects with specular surfaces. + +# 5.2. Experimental Comparisons + +Multi-View Video Dataset. In Tab. 1, TimeFormer improves the reconstruction quality (e.g., PSNR) of original 4DGS [60] and STGS [27] by 0.74 and 0.61, outperforming all baselines methods. We extract the frames at an initial time for scene Cook Spanish in Fig. 6, and TimeFormer can reconstruct a clearer geometry and specular effects compared to the original results. + +To have a more comprehensive overview of Time-Former's capacity on different temporal stages, we collect PSNR for 300 frames of all six scenes from N3DV Dataset [25] and calculate the average per-frame PSNR. As in Fig. 7, the reconstruction quality increases more significantly at the beginning and end, and such balanced reconstruction results are attributed to TimeFormer's cross-temporal attention mechanism on the whole time series. + +Monocular Video Dataset. Tab. 3 shows the enhancements achieved by TimeFormer on the HyperNeRF Dataset. It's observed that TimeFormer increases the PSNR by 0.89 for 4DGS [60] and 0.94 for DeformGS [66], and the SSIM by 0.019 and 0.028, respectively. Fig. 9 additionally provides visualization of the improvement in the image quality: on the left, TimeFormer generates a clearer outline between fingers, while on the right the texture on the front of the broom is clearer. TimeFormer also eliminates artifacts and reveals a clear structure of the broom. + +Dynamic Reflective Video Dataset. Fig. 10 additionally demonstrates that TimeFormer can work well on scenes with dynamic specular objects on NeRF-DS Dataset [64]. + +
MethodSear SteakFlame SSIMSalmon SSIMCut Beef SSIMFlame SSINRSteak SSIMCook Spinach SSINRSSIMCoffee SSINRMartini SSIMMean
PSNRSSIMPSNRSSIM
K-Plane [14]32.520.97130.440.94231.820.96532.380.97030.600.96829.990.94331.63
MixVovels [55]31.210.97129.920.94531.300.96531.430.97031.610.96529.360.94630.80
GS4D [67]32.920.95326.390.89733.080.95933.810.96732.770.95625.230.88430.07
4D-Rotor [9]32.860.95628.250.91333.140.95231.610.95332.560.94927.950.90831.06
4DGS [60]32.490.94928.920.91732.900.95632.510.95432.460.94827.340.90331.10
+TimeFormer33.380.95529.330.92433.110.95733.250.95333.030.94928.930.91031.84
STGS [27]33.710.96228.210.92133.520.95833.460.96333.130.95527.710.91531.62
+TimeFormer34.340.96529.130.92433.570.95834.040.96433.450.95628.830.91732.23
+ +Table 1. Quantitative Comparisons on N3DV Dataset [25]. We use bold font to indicate the improvement, statistics of K-Plane [14] and MixVoxels [55] are from their original paper, while we calculate metrics of all Gaussian-based methods by running their official codes. + +![](images/1953b579de1e33da08265ef263e1ef704b7e78269a1ec5fe933b2a0348e1968a.jpg) + +![](images/6ec2eafbdc5e22dcda778ea59587681b2e915465db2ffc82a126e0c3d7fa7398.jpg) +Figure 6. Visualization of Comparisons on N3DV Dataset [25]. + +![](images/32026f7bef44af4dfcbc0d87200b9546234bbfdbdebc365982f538eccf83330e.jpg) + +![](images/62b48a06311c8803e6d29909eff5674607083d784b65be180f27b417fcab333f.jpg) + +![](images/da94bbadd807988c465daf84fd10986d831afb8edfd4c3e91b2abb3dce654e08.jpg) + +![](images/5385e5e9c2ca0018d6ff00bd8373c5c0ca03fa0c28a0024005b1f004ae5b4861.jpg) +Figure 7. Per-Frame PSNR on N3DV Dataset [25]. TimeFormer improves the reconstruction quality, especially at the beginning and end time series. + +TimeFormer can reconstruct a cleaner appearance on reflective material (e.g., a metallic cup, a glass bottle), and this improvement is attributed to the attention mechanism which enables the deformation field to detect changes on specular surfaces from a learning perspective automatically. + +Analysis of FPS & Canonical Space. We also find that TimeFormer achieves more efficient representation with much fewer points in the final canonical space, thereby achieving a higher FPS in Tab. 2. We attribute the phenomenon to the reason that TimeFormer guides faster gradient descent and promotes the spatial layout of Gaussian points during optimization. Interestingly, in Fig. 8, the L1 loss of both branches in our method decreases + +![](images/f29f1ee14da7316287f670665b2067da9c1e5a95d60286b3bb38e234d09c4327.jpg) +Figure 8. Comparisons of Convergence Speed. We use the same batch size for 4DGS [60] and calculate the loss of both branches in our method. + +faster and shows almost the same convergence trend during training, exceeding the optimization speed of the original method. This eliminates many redundant points and guides the canonical space toward a more efficient distribution, as in Fig. 2. Moreover, Fig. 8 also proves that the cross-time relationship learned in TimeFormer has been successfully transferred to the base branch, as introduced in Sec. 4.3. + +Analysis of Motion Patterns. We argue that TimeFormer achieves more robust learning of motion patterns. To take a deeper insight, we visualize the motion changes as follows: for each Gaussian, we calculate its bias $(\Delta_x, \Delta_y, \Delta_z)$ + +![](images/83bf71128457713c2e5b31099a96df3930484451c7af3415d374a0664849a128.jpg) + +![](images/a4a8920e947b25e399638c6cddcd2083c61301303318b285911de78fbfbda1ff.jpg) +DeformGS PSNR:27.63 + +![](images/3e86b021e18a22956a95a60c93984286616ebd4f37e4485d3b9bb25fdd1cf39d.jpg) + +![](images/99b0a988d13804c01e166678c2cf8615ced9c2ef71aeab62ef0d045ee5a682e1.jpg) ++TimeFormer PSNR:28.59 + +![](images/4e3eb461b7d0a5dcbcec9a624a9e0a7a94858a7b351c819a3bb8bf6456f0d1fd.jpg) + +![](images/6a3c4d553a845fd812df9bb676a241b350919c7b1461400f2d7257fe30591768.jpg) +4DGS PSNR:28.55 + +![](images/e6e8358bc3ac0cb5eb4433d2cab300554d7f0e6decce986e78f7127abcbb0668.jpg) + +![](images/25454d64dca6f24a9919a8941575cae6a4fe15683ba04ff7827d9fdf2b904181.jpg) ++TimeFormer PSNR:29.51 + +![](images/e227a24b7cc91f7884655ccd8e833f855b7bec2586c6dfe8545f3b77d22044b1.jpg) + +![](images/77a375c9d9e614eeed111ccf790af346748f75352babf27b1f8525c75c7aa034.jpg) +Ground Truth + +![](images/2ad325931a0d50b4a3dccdcd456c7e75f946961dc92323367418cd44208e51f8.jpg) + +![](images/92b543abe69218a2214023fb87267060df7370ee96d2a62e3702aa297d262770.jpg) +DeformGS PSNR:20.63 + +![](images/a3401f94a25153b73a4a7ad2ea72c93623d2422ef4eed5b6d405e8f527771966.jpg) + +![](images/f1e6434f642df0e02b72666f406f1108467a231e81ac71b77f6c699c765f0458.jpg) ++TimeFormer PSNR:21.13 + +![](images/dd7da060915c87ee872b913a6556c2574c1d5bc20cbd9b6257db98183995073e.jpg) + +![](images/486c3d8b531296b2be472980f5879879334757fa3df4c72b34afd231a9553f69.jpg) +4DGS PSNR:21.28 + +![](images/2e1e26777659cdaeac54f4ede6be381ece76d8c4f18028757c1f1d78c2257f54.jpg) + +![](images/a002a9ba9de71dc98c829e62ea673044c8b3f5bfbb8e57a5358d9d7d69242cfa.jpg) ++TimeFormer PSNR:22.23 + +![](images/479af13db44a726d37af437335040c70701c69b4e5f41cbec9c49e2f3a4ca293.jpg) + +![](images/caada894c47253a00648e58a31d9d65af5fe811f36b4283875cfc40b0119cd64.jpg) +Ground Truth + +![](images/636636a55414d32f71812b1047e518df9603037018f16c81439f5345f79b4698.jpg) +Figure 9. Visualization of Comparisons on HyperNeRF Dataset [39]. + +![](images/90c51d445920d79dee882b75a1abb57a77d629cfbe95365898d5c07a09fa0664.jpg) + +![](images/9a78033a76223bd4e7efbafc92a4b346b5f5de741ebf5c43344ed5bd8502112f.jpg) + +![](images/4a3bd1bfdc5b4371a9d9156a2a2357c88be8698083d1320d632217fb3748c436.jpg) + +![](images/b0e790bde867741af637cf11ad6acc1b33764edd6acdc8c61920bbf439455938.jpg) + +![](images/96c9520128b8114043673795b0ae9145ebe6c7ed66e9b2c80dfd42f638debb7e.jpg) + +![](images/04b69a301ed48abb34f2757c26f02cb1a667868d84784aa32e07ef22e4bebaae.jpg) + +![](images/7f963041b35a5c71889e748965829793d45d4be7befe542c168b493b0afd0dbe.jpg) + +![](images/8f8a2b84a32cc1a366788266b4ae56a56454eb9eda67ba85ff2fbb43a911bab3.jpg) + +![](images/f9f643ecef67c785672df5638bad6cd3b0f16277582e2f3a033ef6fd270322d0.jpg) + +![](images/81a2c7df2388b93e95ed341480865c92cd83fa396c1a95a9e4f6adf95396c13b.jpg) +Figure 10. Visualization of Comparisons on NeRF-DS Dataset [64]. Compared with original results from DeformGS [66], TimeFormer presents a clearer visual effect on dynamic objects with specular surfaces. + +![](images/4b0d92a467d6cfaf042031d33b1e9046acfeb3120d39d789ee4d6c72e1064188.jpg) + +![](images/20b2a83316c6bc1f1f6e3383f72e99551995f66ce993daf1a9c02b6d58eae03a.jpg) + +![](images/def1b06f93e0a05a0290988e7b38b2b2f5c3908591da6b0d508d4f127d49f738.jpg) + +![](images/460e8424bf472d5f76e830b56a3fafe67406231ed166ed629daac01cba9d6bcb.jpg) + +
MethodN ↓Training ↓FPS ↑
STGS [27]172.1 k41 min82.1
+TimeFormer148.9 k78 min88.9
4DGS [60]145.9 k52 min31.7
+TimeFormer113.1 k94 min37.9
+ +(a) Quantitative comparisons on N3DV Dataset [25]. + +
MethodN↓Training ↓FPS ↑
DeformGS [66]169.6 k25 min30.1
+TimeFormer82.9 k35 min58.9
4DGS [60]172.3 k32 min35.7
+TimeFormer135.5 k48 min40.9
+ +(b) Quantitative comparisons on HyperNeRF Dataset [39]. + +Table 2. Comparisons of Gaussian Number, Training Time & FPS. + +towards canonical space in time $t$ and replace the color attributes $(r, g, b)$ with the absolute value of this bias: $(r, g, b) \gets (|\Delta_x|, |\Delta_y|, |\Delta_z|)$ . We use the same splatting algorithm to render the accumulated motion bias, transfer + +ring the rendered result as a heatmap, as in Fig. 11. + +In Fig. 11a, TimeFormer identifies the blade's motion as a rigid deformation, showing a sharper outline and consistent motion, especially as it contacts the initially static lemon. This demonstrates TimeFormer's ability to distinguish between moving and static objects. In Fig. 11b, TimeFormer captures subtle motion on the spray gun and clear flame flickering on the beef. Fig. 11c shows that TimeFormer also detects the violent movement of the elongated broom, where 4DGS [60] fails, proving its robustness in handling extremely geometries. + +# 5.3. Ablation Studies + +Cross-Temporal Encoder. Tab. 5 demonstrates that the performance of TimeFormer is not sensitive to time batch $B$ and transformer encoder layer $M$ . In Tab. 5b, we explore two sampling methods: random sampling and continuous sampling of timestamps. Both strategies demonstrate improvements in PSNR and other metrics compared to the baseline. Notably, the random sampling method yields more significant enhancements. We hypothesize that this + +
MethodBroomChickenCut LemonTorchcoPeel BananaHandMean
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DeformGS [66]20.740.32226.320.78627.940.71427.40.87726.380.83627.790.7826.090.719
+TimeFormer21.010.32426.550.79230.010.78027.550.88327.240.85229.790.84827.030.747
4DGS[60]21.530.35126.820.79729.720.76327.450.88327.820.84429.520.84127.140.747
+TimeFormer22.840.40127.060.80130.230.77829.480.90228.210.85330.380.86228.030.766
+ +Table 3. Quantitative Comparisons on HyperNeRF Dataset [39]. + +
MethodPressPlateBasinSieveBellCupMean
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
HyperNeRF [39]25.40.87318.10.71420.20.82925.00.90924.00.88424.10.89622.80.851
NeRF-DS [64]26.40.91120.80.86720.30.86826.10.93523.30.87224.50.91623.570.895
DeformGS[66]25.680.86620.820.81219.870.80425.710.88124.920.85424.520.89723.590.852
+TimeFormer26.290.86720.900.81719.930.80526.260.89125.900.87325.160.90324.100.859
+ +![](images/fab3f703efa079a01918d7752d39f32e42468806559975932e1922028cced503.jpg) +(a) Cut Lemon + +![](images/134b17a7009fe9a1932a98d991747c9bdeb08dc63b2b48883c63487cbb735450.jpg) +(b) Torchocolate +(c) Broom +Figure 11. Motion Visualization on HyperNeRF Dataset [39]. + +phenomenon occurs because the random sampling method allows for more effective propagation of dynamic behavior modeling across the entire time series. + +Shared Deformation Fields. In Tab. 5b, the image quality during inference suffers a considerable decrease if we train two deformation fields without sharing their weights. + +Table 4. Quantitative Comparisons on NeRF-DS Dataset [64]. + +
SettingRandom SamplingContinuous Sampling
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
B=225.890.8750.11725.730.8710.124
B=426.060.8710.11925.780.8740.117
B=625.990.870.12325.730.8650.125
+ +(a) Ablation Studies on Batch Size and Sampling Strategies. + +
SettingPSNR↑SSIM↑LPIPS↓
Baseline25.440.8670.125
M=125.960.8740.117
M=225.900.8740.120
M=326.050.8760.112
M=426.040.8730.119
w/o Shared24.810.8520.132
+ +(b) Ablation Results. $M$ is the number of encoder layers, "w/o Shared" means not using shared weights in two deformation fields. +Table 5. Ablation Results. The results are from three scenes press, sieve and bell on NeRF-DS Dataset [64]. + +Fig. 8 also demonstrates the effectiveness of the two-stream strategy and TimeFormer's ability of re-balancing. + +# 6. Conclusion + +We propose TimeFormer, a Transformer module which enhances existing deformable 3D Gaussians methods by implicitly modeling motion patterns from a learning perspective. In addition, we design a two-stream optimization strategy to transfer the learned motion knowledge from TimeFormer to the original deformation branch. This allows us to remove TimeFormer during inference and thus maintain the same inference speed. Extensive experiments demonstrate that TimeFormer facilitates the reconstruction of three state-of-the-art methods among three datasets, and illustrate the improvement on reconstructing complex scenes. + +# 7. Acknowledgement + +# References + +[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 3 +[2] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. arXiv preprint arXiv:2404.03613, 2024. 2, 3 +[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855-5864, 2021. 2 +[4] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 3 +[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2 +[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 4 +[7] Robert A Drebin, Loren Carpenter, and Pat Hanrahan. Volume rendering. ACM Siggraph Computer Graphics, 22(4): 65-74, 1988. 2 +[8] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2 +[9] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. In Proc. SIGGRAPH, 2024. 2, 3, 6 +[10] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Mike Zheng Shou, Shuran Song, and Jeffrey Ichnowski. Md-splatting: Learning metric deformation from + +This work is completed during the author's internship at the Tencent Hunyuan 3D Department. This work is supported in part by the National Science Fund for Distinguished Young Scholars (No. 62325208), in part by the National Natural Science Foundation of China (No. 62232002), and in part by the Natural Science Foundation of Tianjin (No. 23ZGZNGX00020 and 24JCZXJC00050). +4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583, 2023. 2, 3 +[11] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2 +[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2 +[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 3 +[14] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 6 +[15] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2, 3 +[16] Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wenchao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splitting gaussian dynamics for 4d content creation. arXiv preprint arXiv:2403.12365, 2024. 3 +[17] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. arXiv preprint arXiv:2403.11447, 2024. 3 +[18] Zhi Hou, Baosheng Yu, and Dacheng Tao. Batchformer: Learning to explore sample relationships for robust representation learning. In CVPR, 2022. 4 +[19] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4220-4230, 2024. 2, 3 +[20] Takeo Kanade, Peter Rander, and PJ Narayanan. Virtualized reality: Constructing virtual worlds from real scenes. IEEE multimedia, 4(1):34-47, 1997. 2 +[21] Kai Katsumata, Duc Minh Vo, and Hideki Nakayama. A compact dynamic 3d gaussian representation for real-time dynamic view synthesis. In Computer Vision - ECCV 2024, pages 394-412. Springer Nature Switzerland, 2025. 2, 3 +[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3 +[23] Deqi Li, Shi-Sheng Huang, Zhiyuan Lu, Xinran Duan, and Hua Huang. St-4dgs: Spatial-temporally consistent 4d gaussian splatting for efficient dynamic scene rendering. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3 + +[24] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 2, 3 +[25] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2, 5, 6, 7 +[26] Zhengqi Li, Simon Niklaus, Noah Snively, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2, 3 +[27] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8508-8520, 2024. 1, 2, 3, 5, 6, 7 +[28] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2, 3 +[29] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. Advances in Neural Information Processing Systems, 35:36762-36775, 2022. 3 +[30] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13-23, 2023. 2, 3 +[31] Ange Lou, Benjamin Planche, Zhongpai Gao, Yamin Li, Tianyu Luan, Hao Ding, Terrence Chen, Jack Noble, and Ziyan Wu. Darenerf: Direction-aware representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5031-5042, 2024. 3 +[32] Jiahao Lu, Jiacheng Deng, Ruijie Zhu, Yanzhe Liang, Wenfei Yang, Tianzhu Zhang, and Xu Zhou. Dn-4dgs: Denoised deformable network with temporal-spatial aggregation for dynamic scene rendering. arXiv preprint arXiv:2410.13607, 2024. 2, 3 +[33] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2 +[34] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 2024 International Conference on 3D Vision (3DV), pages 800-809. IEEE, 2024. 2, 3 + +[35] Xingyu Miao, Yang Bai, Haoran Duan, Fan Wan, Yawen Huang, Yang Long, and Yefeng Zheng. Ctnerf: Cross-time transformer for dynamic neural radiance field from monocular video. Pattern Recognition, 156:110729, 2024. 3 +[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[37] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2 +[38] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2, 3 +[39] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 2, 3, 5, 7, 8 +[40] Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahiup Kang. Temporal interpolation is all you need for dynamic neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4212-4221, 2023. 3 +[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2, 3 +[42] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boring Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 3 +[43] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Pérez-Pellitero. Swings: sliding windows for dynamic 3d gaussian splatting. In European Conference on Computer Vision. Springer, 2024. 2, 3 +[44] Nagabhushan Somraj, Kapil Choudhary, Sai Harsha Mupparaju, and Rajiv Soundararajan. Factorized motion fields for fast sparse input dynamic view synthesis. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3 +[45] Liangchen Song, Xuan Gong, Benjamin Planche, Meng Zheng, David Doermann, Junsong Yuan, Terrence Chen, and Ziyan Wu. Pref: Predictability regularized neural motion fields. In European Conference on Computer Vision, pages 664-681. Springer, 2022. 3 +[46] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with de + +composed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 3 +[47] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2 +[48] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20675-20685, 2024. 2, 3 +[49] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3 +[50] Fengrui Tian, Shaoyi Du, and Yueqi Duan. Mononerf: Learning a generalizable dynamic radiance field from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17903-17913, 2023. 2, 3 +[51] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2, 3 +[52] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3, 4 +[53] Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3 +[54] Chaoyang Wang, Lachlan Ewen MacDonald, Laszlo A Jeni, and Simon Lucey. Flow supervision for deformable nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21128-21137, 2023. 3 +[55] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, Yafei Song, and Huaping Liu. Mixed neural voxels for fast multiview video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19706-19716, 2023. 3, 6 +[56] Feng Wang, Zilong Chen, Guokang Wang, Yafei Song, and Huaping Liu. Masked space-time hash encoding for efficient dynamic scene reconstruction. Advances in Neural Information Processing Systems, 36, 2024. +[57] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 76-87, 2023. 3 +[58] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. arXiv preprint arXiv:2405.18426, 2024. 3 + +[59] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5 +[60] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 1, 2, 3, 5, 6, 7, 8 +[61] Minye Wu, Zehao Wang, Georgios Kouros, and Tinne Tuytelaars. Tetrirf: Temporal tri-plane radiance fields for efficient free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6487-6496, 2024. 3 +[62] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9421-9431, 2021. 2 +[63] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20029-20040, 2024. 3 +[64] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8285-8295, 2023. 5, 7, 8 +[65] Jason C Yang, Matthew Everett, Chris Buehler, and Leonard McMillan. A real-time distributed light field camera. Rendering Techniques, 2002(77-86):2, 2002. 2 +[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 1, 2, 3, 5, 7, 8 +[67] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 6 +[68] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2 +[69] Yifan Zhan, Zhuoxiao Li, Muyao Niu, Zhihang Zhong, Shohei Nobuhara, Ko Nishino, and Yinqiang Zheng. Kfdnerf: Rethinking dynamic nerf with kalman filter. arXiv preprint arXiv:2407.13185, 2024. 3 +[70] Boming Zhao, Yuan Li, Ziyu Sun, Lin Zeng, Yujun Shen, Rui Ma, Yinda Zhang, Hujun Bao, and Zhaopeng Cui. Gaussian prediction: Dynamic 3d gaussian prediction for motion extrapolation and free view synthesis. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2, 3 +[71] Kaichen Zhou, Jia-Xing Zhong, Sangyun Shin, Kai Lu, Yiyuan Yang, Andrew Markham, and Niki Trigoni. Dynpoint: Dynamic neural point for view synthesis. Advances in Neural Information Processing Systems, 36, 2024. 2 + +[72] Ruijie Zhu, Yanzhe Liang, Hanzhi Chang, Jiacheng Deng, Jiahao Lu, Wenfei Yang, Tianzhu Zhang, and Yongdong Zhang. Motions: Exploring explicit motion guidance for deformable 3d gaussian splatting. arXiv preprint arXiv:2410.07707, 2024. 3 +[73] C Lawrence Zitnick, Sing Bing Kang, Matthew Uytendaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM transactions on graphics (TOG), 23(3):600-608, 2004. 2 \ No newline at end of file diff --git a/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/images.zip b/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..818294698500e0cffedca31ecc44eedd9df14544 --- /dev/null +++ b/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:155ebb40749f2599ae88fb14e6394ad6c2a2708ae15fe11768b216197b30eb6e +size 1106666 diff --git a/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/layout.json b/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1548710f46655710fbad1b5481994eae09e82b48 --- /dev/null +++ b/2025/TimeFormer_ Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction/layout.json @@ -0,0 +1,12644 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 133, + 103, + 477, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 103, + 477, + 138 + ], + "spans": [ + { + "bbox": [ + 133, + 103, + 477, + 138 + ], + "type": "text", + "content": "TimeFormer: Capturing Temporal Relationships of Deformable 3D Gaussians for Robust Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 160, + 194, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 160, + 194, + 176 + ], + "spans": [ + { + "bbox": [ + 113, + 160, + 194, + 176 + ], + "type": "text", + "content": "Dadong Jiang" + }, + { + "bbox": [ + 113, + 160, + 194, + 176 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 178, + 212, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 178, + 212, + 190 + ], + "spans": [ + { + "bbox": [ + 96, + 178, + 212, + 190 + ], + "type": "text", + "content": "patrickddj7@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 194, + 206, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 194, + 206, + 209 + ], + "spans": [ + { + "bbox": [ + 127, + 194, + 206, + 209 + ], + "type": "text", + "content": "Xianghui Yang3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 119, + 212, + 213, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 212, + 213, + 223 + ], + "spans": [ + { + "bbox": [ + 119, + 212, + 213, + 223 + ], + "type": "text", + "content": "yangxhui6@163.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 284, + 160, + 335, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 284, + 160, + 335, + 174 + ], + "spans": [ + { + "bbox": [ + 284, + 160, + 335, + 174 + ], + "type": "text", + "content": "Zhi Hou" + }, + { + "bbox": [ + 284, + 160, + 335, + 174 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 239, + 178, + 381, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 178, + 381, + 189 + ], + "spans": [ + { + "bbox": [ + 239, + 178, + 381, + 189 + ], + "type": "text", + "content": "zhou9878@uni.sydney.edu.au" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 264, + 194, + 345, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 194, + 345, + 208 + ], + "spans": [ + { + "bbox": [ + 264, + 194, + 345, + 208 + ], + "type": "text", + "content": "Xiaobo Zhou" + }, + { + "bbox": [ + 264, + 194, + 345, + 208 + ], + "type": "inline_equation", + "content": "^{1,4\\dagger}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 245, + 212, + 365, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 212, + 365, + 223 + ], + "spans": [ + { + "bbox": [ + 245, + 212, + 365, + 223 + ], + "type": "text", + "content": "xiaobo.zhou@tju.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 433, + 161, + 488, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 433, + 161, + 488, + 174 + ], + "spans": [ + { + "bbox": [ + 433, + 161, + 488, + 174 + ], + "type": "text", + "content": "Zhihui Ke" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 409, + 178, + 513, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 409, + 178, + 513, + 189 + ], + "spans": [ + { + "bbox": [ + 409, + 178, + 513, + 189 + ], + "type": "text", + "content": "kezhihui@tju.edu.cn" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 419, + 194, + 468, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 194, + 468, + 209 + ], + "spans": [ + { + "bbox": [ + 419, + 194, + 468, + 209 + ], + "type": "text", + "content": "Tie Qiu" + }, + { + "bbox": [ + 419, + 194, + 468, + 209 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 397, + 212, + 490, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 212, + 490, + 223 + ], + "spans": [ + { + "bbox": [ + 397, + 212, + 490, + 223 + ], + "type": "text", + "content": "qiutie@tju.edu.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 61, + 228, + 561, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 561, + 244 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 561, + 244 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 228, + 561, + 244 + ], + "type": "text", + "content": " Tianjin University " + }, + { + "bbox": [ + 61, + 228, + 561, + 244 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 228, + 561, + 244 + ], + "type": "text", + "content": " Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 61, + 228, + 561, + 244 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 228, + 561, + 244 + ], + "type": "text", + "content": " Tencent " + }, + { + "bbox": [ + 61, + 228, + 561, + 244 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 228, + 561, + 244 + ], + "type": "text", + "content": " Qinghai Minzu University" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 58, + 248, + 152, + 398 + ], + "blocks": [ + { + "bbox": [ + 58, + 248, + 152, + 398 + ], + "lines": [ + { + "bbox": [ + 58, + 248, + 152, + 398 + ], + "spans": [ + { + "bbox": [ + 58, + 248, + 152, + 398 + ], + "type": "image", + "image_path": "2764ac663fe4ffba03aad8528a90d4a3b8562e60ca0658137ab81572ea2fec83.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 58, + 399, + 104, + 445 + ], + "blocks": [ + { + "bbox": [ + 58, + 399, + 104, + 445 + ], + "lines": [ + { + "bbox": [ + 58, + 399, + 104, + 445 + ], + "spans": [ + { + "bbox": [ + 58, + 399, + 104, + 445 + ], + "type": "image", + "image_path": "069a1eec7ac7d8dc1fb2244ca72e46900c11e67c65fedbbb6fdecc396a1eb5f7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 446, + 145, + 455 + ], + "lines": [ + { + "bbox": [ + 69, + 446, + 145, + 455 + ], + "spans": [ + { + "bbox": [ + 69, + 446, + 145, + 455 + ], + "type": "text", + "content": "DeformGS PSNR:27.53" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 154, + 248, + 249, + 398 + ], + "blocks": [ + { + "bbox": [ + 154, + 248, + 249, + 398 + ], + "lines": [ + { + "bbox": [ + 154, + 248, + 249, + 398 + ], + "spans": [ + { + "bbox": [ + 154, + 248, + 249, + 398 + ], + "type": "image", + "image_path": "f757022999d07ed5ad3b24825151ebc029c4d1d53c41ab0783ceed1d888ed3fd.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 154, + 399, + 202, + 445 + ], + "blocks": [ + { + "bbox": [ + 154, + 399, + 202, + 445 + ], + "lines": [ + { + "bbox": [ + 154, + 399, + 202, + 445 + ], + "spans": [ + { + "bbox": [ + 154, + 399, + 202, + 445 + ], + "type": "image", + "image_path": "8ad2ff566c29fd88f2f99299e3c47cbcf7c8094e14821d10e80efcfd52858d5c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 446, + 245, + 455 + ], + "lines": [ + { + "bbox": [ + 157, + 446, + 245, + 455 + ], + "spans": [ + { + "bbox": [ + 157, + 446, + 245, + 455 + ], + "type": "text", + "content": "+TimeFormer PSNR:29.34" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 252, + 249, + 394, + 344 + ], + "blocks": [ + { + "bbox": [ + 252, + 249, + 394, + 344 + ], + "lines": [ + { + "bbox": [ + 252, + 249, + 394, + 344 + ], + "spans": [ + { + "bbox": [ + 252, + 249, + 394, + 344 + ], + "type": "image", + "image_path": "c8c76442a8f177dd0ef48b19245ad93ecaa72146f58c98faf504103471b9028a.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 343, + 356, + 350 + ], + "lines": [ + { + "bbox": [ + 293, + 343, + 356, + 350 + ], + "spans": [ + { + "bbox": [ + 293, + 343, + 356, + 350 + ], + "type": "text", + "content": "STGS PSNR: 27.28" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 252, + 350, + 394, + 445 + ], + "blocks": [ + { + "bbox": [ + 252, + 350, + 394, + 445 + ], + "lines": [ + { + "bbox": [ + 252, + 350, + 394, + 445 + ], + "spans": [ + { + "bbox": [ + 252, + 350, + 394, + 445 + ], + "type": "image", + "image_path": "a376af876f0b331f7d044f378acca83cf4c7f49be271faea76baad2b69927723.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 446, + 372, + 454 + ], + "lines": [ + { + "bbox": [ + 285, + 446, + 372, + 454 + ], + "spans": [ + { + "bbox": [ + 285, + 446, + 372, + 454 + ], + "type": "text", + "content": "+TimeFormer PSNR:27.98" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 397, + 248, + 445, + 296 + ], + "blocks": [ + { + "bbox": [ + 397, + 248, + 445, + 296 + ], + "lines": [ + { + "bbox": [ + 397, + 248, + 445, + 296 + ], + "spans": [ + { + "bbox": [ + 397, + 248, + 445, + 296 + ], + "type": "image", + "image_path": "b1d41cf214f1312ac1c6c3efbd825eb5c769d445f030b0796e16d415089cf95b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 399, + 297, + 444, + 345 + ], + "blocks": [ + { + "bbox": [ + 399, + 297, + 444, + 345 + ], + "lines": [ + { + "bbox": [ + 399, + 297, + 444, + 345 + ], + "spans": [ + { + "bbox": [ + 399, + 297, + 444, + 345 + ], + "type": "image", + "image_path": "d847fe0f9b12247ba1025d6a94de0001990026997c34b8c37dff12c15020136c.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 397, + 350, + 445, + 397 + ], + "blocks": [ + { + "bbox": [ + 397, + 350, + 445, + 397 + ], + "lines": [ + { + "bbox": [ + 397, + 350, + 445, + 397 + ], + "spans": [ + { + "bbox": [ + 397, + 350, + 445, + 397 + ], + "type": "image", + "image_path": "e26fe2e99b6de71585fa2f2cae4312fa65c814413e0920ea1fc2ee083618126e.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 398, + 399, + 444, + 445 + ], + "blocks": [ + { + "bbox": [ + 398, + 399, + 444, + 445 + ], + "lines": [ + { + "bbox": [ + 398, + 399, + 444, + 445 + ], + "spans": [ + { + "bbox": [ + 398, + 399, + 444, + 445 + ], + "type": "image", + "image_path": "1278efc39127dce26a8d1b72c57039e1bca6dd6bb13eef32ba6eebee05293336.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 452, + 249, + 553, + 450 + ], + "blocks": [ + { + "bbox": [ + 452, + 249, + 553, + 450 + ], + "lines": [ + { + "bbox": [ + 452, + 249, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 452, + 249, + 553, + 450 + ], + "type": "image", + "image_path": "3a530fc711584fb47472638fd78cea9077aebeefe035735f96b25c386ac02dfd.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 458, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 55, + 458, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 555, + 492 + ], + "type": "text", + "content": "Figure 1. We propose TimeFormer, a Transformer module that implicitly models the motion pattern via Temporal Attention from a learning perspective (right). TimeFormer is plug-and-play to existing deformable 3D Gaussian reconstruction methods [27, 60, 66] and enhances reconstruction results (left and middle) without affecting their original inference speed." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 152, + 506, + 200, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 506, + 200, + 517 + ], + "spans": [ + { + "bbox": [ + 152, + 506, + 200, + 517 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 54, + 531, + 297, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 531, + 297, + 686 + ], + "spans": [ + { + "bbox": [ + 54, + 531, + 297, + 686 + ], + "type": "text", + "content": "Dynamic scene reconstruction is a long-term challenge in 3D vision. Recent methods extend 3D Gaussian Splating to dynamic scenes via additional deformation fields and apply explicit constraints like motion flow to guide the deformation. However, they learn motion changes from individual timestamps independently, making it challenging to reconstruct complex scenes, particularly when dealing with violent movement, extreme-shaped geometries, or reflective surfaces. To address the above issue, we design a simple yet effective plug-and-play module called TimeFormer to enable existing deformable 3D Gaussians reconstruction methods with the ability to implicitly model motion patterns from a learning perspective. Specifically," + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 506, + 556, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 627 + ], + "type": "text", + "content": "TimeFormer includes a Cross-Temporal Transformer Encoder, which adaptively learns the temporal relationships of deformable 3D Gaussians. Furthermore, we propose a two-stream optimization strategy that transfers the motion knowledge learned from TimeFormer to the base stream during the training phase. This allows us to remove TimeFormer during inference, thereby preserving the original rendering speed. Extensive experiments in the multi-view and monocular dynamic scenes validate qualitative and quantitative improvement brought by TimeFormer." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 314, + 655, + 395, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 655, + 395, + 667 + ], + "spans": [ + { + "bbox": [ + 314, + 655, + 395, + 667 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "High-quality reconstruction of dynamic scenes is significantly challenging in computer vision and graphics, yet has a wide range of potential applications in movie production," + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 151, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 151, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 151, + 703 + ], + "type": "text", + "content": "† Corresponding Author" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 69, + 703, + 245, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 245, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 245, + 713 + ], + "type": "text", + "content": "Project Page: https://patrickddj.github.io/TimeFormer" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "8721" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "virtual reality, and augmented reality. The difficulty stems from factors like occlusions, translucent materials, specular surfaces, and changing topology, all of which are prevalent in dynamic scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 295, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 295, + 302 + ], + "type": "text", + "content": "Inspired by the success of neural radiance field (NeRF) on static scenes [3, 36, 37], extending NeRF to dynamic scenes [15, 24, 26, 30, 38, 39, 50, 71] has been explored. However, these NeRF-based methods are limited by computationally intensive volume rendering [7], which makes real-time rendering almost impossible. Recently, 3D Gaussian Splatting (3DGS) [22] represents the scene with anisotropic 3D Gaussians and develops a rasterization-based rendering algorithm, which allows real-time rendering through directly projecting 3D Gaussians onto the image plane. Following the revolutionized 3DGS, dynamic Gaussian Splatting methods [9, 27, 48, 67] have been introduced for reconstructing dynamic scenes. These methods typically construct a canonical 3DGS and utilize a deformation field based on individual timestamps [27, 60, 66]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 304, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 495 + ], + "type": "text", + "content": "However, temporal relationships of 3D Gaussians have been poorly investigated. Previous studies have primarily modeled motion patterns using various types of the deformation field, for example, MLPs [2, 19, 33, 34, 43, 53, 66, 70], spatial-temporal planes [10, 32, 60], polynomial functions [27], Fourier series [21], and combinations of these methods [28]. These methods learn motion patterns from independent time input in a vanilla way, neglecting internal cross-time relationships. Some studies introduce the motion flow regularization [23] to explicitly learn motion patterns from neighboring frames [32, 34]. Although these methods promote similar motion between adjacent timestamps, they adopt a local perspective on time series during optimization. This constraint makes it challenging to reconstruct scenes with more complex motion patterns, such as sudden appearances, violent movements, or reflective surfaces." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we introduce TimeFormer, a transformer module designed to implicitly learn motion patterns across multiple timestamps. By modeling temporal relationships within a time batch, we aim to provide a global view of the entire time series, enabling the deformation field backbones themselves to capture motion patterns from a learning perspective, as shown in Fig. 1. Specifically, TimeFormer utilizes a Cross-Temporal Encoder to capture the implicit motion patterns of 3D Gaussians across multiple sampled timestamps using a self-attention mechanism. Moreover, to avoid additional computational costs of TimeFormer during inference, we present a two-stream optimization strategy. By sharing the weights of two deformation fields, we transfer the motion pattern learning from TimeFormer stream to the base stream during training. Therefore, we can eliminate TimeFormer during inference and maintain the original rendering speed. Notably, TimeFormer does not require any prior information and extracts motion patterns solely from" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 319, + 77, + 426, + 127 + ], + "blocks": [ + { + "bbox": [ + 319, + 77, + 426, + 127 + ], + "lines": [ + { + "bbox": [ + 319, + 77, + 426, + 127 + ], + "spans": [ + { + "bbox": [ + 319, + 77, + 426, + 127 + ], + "type": "image", + "image_path": "735ecea33472d60d7c3725f3768d2de64dc9e8f2c06dca090d6cab986072efaf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 128, + 414, + 136 + ], + "lines": [ + { + "bbox": [ + 334, + 128, + 414, + 136 + ], + "spans": [ + { + "bbox": [ + 334, + 128, + 414, + 136 + ], + "type": "text", + "content": "Points: " + }, + { + "bbox": [ + 334, + 128, + 414, + 136 + ], + "type": "inline_equation", + "content": "154.5\\mathrm{k}" + }, + { + "bbox": [ + 334, + 128, + 414, + 136 + ], + "type": "text", + "content": " FPS:31.3" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 440, + 69, + 542, + 125 + ], + "blocks": [ + { + "bbox": [ + 440, + 69, + 542, + 125 + ], + "lines": [ + { + "bbox": [ + 440, + 69, + 542, + 125 + ], + "spans": [ + { + "bbox": [ + 440, + 69, + 542, + 125 + ], + "type": "image", + "image_path": "f67580d67fc1e2454ab9360a894d963dcda2536e4378dfe2707028f203912bd5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 127, + 533, + 136 + ], + "lines": [ + { + "bbox": [ + 456, + 127, + 533, + 136 + ], + "spans": [ + { + "bbox": [ + 456, + 127, + 533, + 136 + ], + "type": "text", + "content": "Points: " + }, + { + "bbox": [ + 456, + 127, + 533, + 136 + ], + "type": "inline_equation", + "content": "63.6\\mathrm{k}" + }, + { + "bbox": [ + 456, + 127, + 533, + 136 + ], + "type": "text", + "content": " FPS: 79.7" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 140, + 430, + 208 + ], + "blocks": [ + { + "bbox": [ + 316, + 140, + 430, + 208 + ], + "lines": [ + { + "bbox": [ + 316, + 140, + 430, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 430, + 208 + ], + "type": "image", + "image_path": "29581fe486f0f084c13f286aabcf26656abec65d2ae7a2ca2beaa611bb05b393.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 209, + 412, + 218 + ], + "lines": [ + { + "bbox": [ + 337, + 209, + 412, + 218 + ], + "spans": [ + { + "bbox": [ + 337, + 209, + 412, + 218 + ], + "type": "text", + "content": "DeformGS PSNR:27.17" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 433, + 139, + 553, + 209 + ], + "blocks": [ + { + "bbox": [ + 433, + 139, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 433, + 139, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 433, + 139, + 553, + 209 + ], + "type": "image", + "image_path": "d92e509b77958c36d5f22608a21db0bb64a51b1d6d8d9a7ef5ea5706895c0b9c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 210, + 536, + 218 + ], + "lines": [ + { + "bbox": [ + 452, + 210, + 536, + 218 + ], + "spans": [ + { + "bbox": [ + 452, + 210, + 536, + 218 + ], + "type": "text", + "content": "+TimeFormer PSNR:28.70" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 222, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 313, + 222, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 553, + 255 + ], + "type": "text", + "content": "Figure 2. TimeFormer guides towards more efficiently distributed canonical space, showing higher FPS and better quality. The results are from \"cut lemon\" in the HyperNeRF Dataset [39]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 264, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 323 + ], + "type": "text", + "content": "RGB supervision, which can be seamlessly adapted to previous deformable 3D Gaussian methods in a plug-and-play manner. Furthermore, TimeFormer promotes faster gradient descent and guides a more efficient canonical space, ultimately increasing FPS during inference, as in Fig. 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 325, + 527, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 325, + 527, + 336 + ], + "spans": [ + { + "bbox": [ + 326, + 325, + 527, + 336 + ], + "type": "text", + "content": "To sum up, our main contributions are as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 338, + 553, + 458 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 314, + 338, + 553, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 553, + 385 + ], + "type": "text", + "content": "- We propose TimeFormer, a transformer module that enhances current deformable 3D Gaussians reconstruction methods in a plug-and-play manner from an automatic learning perspective." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 386, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 386, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 314, + 386, + 553, + 422 + ], + "type": "text", + "content": "- The two-stream optimization strategy allows the exclusion of TimeFormer during inference while maintaining and even improving rendering speed." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 422, + 553, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 553, + 458 + ], + "type": "text", + "content": "- Extensive experiments on real-world datasets validate the effectiveness of TimeFormer, achieving state-of-the-art rendering quality." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 471, + 405, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 471, + 405, + 483 + ], + "spans": [ + { + "bbox": [ + 314, + 471, + 405, + 483 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 491, + 482, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 482, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 482, + 504 + ], + "type": "text", + "content": "2.1. Dynamic Scene Reconstruction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "Dynamic Scene Reconstruction has been extensively researched over many years, with a wealth of studies [20, 39, 65, 66, 73] contributing to current progress. The pioneering neural radiance field [3, 12, 36, 47, 68] has demonstrated photorealistic rendering for novel view synthesis from calibrated multiview images, inspiring extensive approaches [8, 11, 38, 39, 41, 51, 62] to extend NeRF to 4D space-time field for dynamic scene reconstruction. There are majorly two lines of NeRF-based methods for dynamic scene reconstruction: 1) deformation-based methods [11, 38, 39, 41, 51] model deformation changes by using a deformation field to map the queried positions in different timestamps to a canonical space; 2) compact 4D space-time fields [5, 8, 25, 62] take the timestamps, positions, and directions as input to predict color and density. However, NeRF-based approaches are limited to training and rendering speed. Thus, a large number of meth-" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8722" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "ods [1, 4, 13, 24, 29, 31, 42, 46, 55-57, 61, 63] have been proposed to accelerate NeRF-based methods." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 98, + 296, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 98, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 98, + 296, + 349 + ], + "type": "text", + "content": "Recent 3DGS [22] achieves real-time rendering and dynamic 3DGS approaches [27, 66, 67] emerge. In details, deformation-based methods [2, 10, 19, 21, 27, 28, 32-34, 43, 53, 60, 66, 70] employ a deformation field to predict per-Gaussian offsets and deform a canonical 3DGS according to different timestamps. Meanwhile, 4DGS [67] and 4D-Rotor [9] add the time dimension to 3D Gaussian, forming a 4D Gaussian representation. TimeFormer focuses on modeling the temporal relationship and is plug-and-play to deformation-based methods [66] and 4D space-time representations [27, 67]. There are also a few attempts [28, 43, 48] in leveraging the nearby motion or flow for dynamic Gaussian reconstruction. For example, 3DGStream [48] proposes a per-frame optimization strategy, which predicts current Gaussian attributes based on previous Gaussian attributes. Meanwhile, Gaussian-Flow [28] presents explicitly model time-dependent residual of each attribute. To model a long video sequence, SWinGS [43] splits a video sequence into multiple sliding windows based on optical flow and uses a deformation-based method to model each sliding window." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 358, + 160, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 160, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 160, + 372 + ], + "type": "text", + "content": "2.2. Motion Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 377, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 568 + ], + "type": "text", + "content": "Motion prediction in dynamic scene reconstruction from multi-view videos or monocular videos is an ill-posed problem. Early dynamic NeRF [15, 38, 39, 51] directly learns motion patterns from individual timestamps, and several later works propose motion flow regularization terms [15, 26, 30, 35, 40, 44, 45, 50, 54, 69] to promote the learning of cross-time motion patterns. These methods typically utilize 2D prior information (e.g., optical flow) from pre-trained networks [49] to supervise the scene flows within neighboring frames. PREF [45] uses motion predictor to infer current motion based on the previous four frames and propose a self-supervision strategy without optical flow prior. KFD-NeRF [69] models motion patterns by Kalman Filter based on the previous two frames. CT-NeRF [35] employs a cross-attention mechanism from transformer [52] to learn the correlation of conservative frames." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Dynamic 3D Gaussian Splatting (3DGS) methods [16, 17, 58, 72] also utilize optical flow supervision to enhance reconstruction performance. Techniques such as MD-Splatting [10], D3DG [34], and ST-4DGS [23] focus on minimizing both forward and backward Gaussian flow to promote temporal smoothness in Gaussian motions. DN4DGS [32] incorporates information from previous, current, and next timestamps into the deformation field to capture cross-time motion patterns. However, these methods primarily establish temporal correlations with neighboring timestamps, which limits their ability to address long-term motion changes. Additionally, they introduce extra compu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "tational costs during inference, which can reduce rendering speed. In contrast, TimeFormer captures the temporal relationships of Gaussians from a global perspective across the entire time series. Notably, TimeFormer is employed only during the training phase, without any additional computational costs during inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 154, + 533, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 154, + 533, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 154, + 533, + 167 + ], + "type": "text", + "content": "3. Preliminary: Deformable 3D Gaussians" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 174, + 553, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 553, + 233 + ], + "type": "text", + "content": "3D Gaussians [22] represent the scene with a set of 3D Gaussians, each of which has unique opacity " + }, + { + "bbox": [ + 313, + 174, + 553, + 233 + ], + "type": "inline_equation", + "content": "o \\in [0,1]" + }, + { + "bbox": [ + 313, + 174, + 553, + 233 + ], + "type": "text", + "content": ", center position " + }, + { + "bbox": [ + 313, + 174, + 553, + 233 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^{3 \\times 1}" + }, + { + "bbox": [ + 313, + 174, + 553, + 233 + ], + "type": "text", + "content": ", and covariance matrix " + }, + { + "bbox": [ + 313, + 174, + 553, + 233 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 313, + 174, + 553, + 233 + ], + "type": "text", + "content": ". For a position " + }, + { + "bbox": [ + 313, + 174, + 553, + 233 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{3 \\times 1}" + }, + { + "bbox": [ + 313, + 174, + 553, + 233 + ], + "type": "text", + "content": " in 3D space, the corresponding contribution of a 3D Gaussian on it can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 367, + 238, + 553, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 238, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 367, + 238, + 553, + 255 + ], + "type": "interline_equation", + "content": "G (x) = o \\cdot e ^ {- \\frac {1}{2} (x - \\mu) ^ {\\top} \\Sigma^ {- 1} (x - \\mu)}. \\tag {1}", + "image_path": "c27194cda1e0416ba0d6ffbcc139fcb7fa2465570d03fbd6ad1ba0901ed81578.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "content": "The covariance matrix " + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "content": " can be decomposed into a scaling matrix " + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "content": " and a rotation matrix " + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma} = \\mathbf{RSS}^T\\mathbf{R}^T" + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\mathrm{diag}([s_x, s_y, s_z])" + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "content": " can be transformed from a quaternion " + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "inline_equation", + "content": "[r_w, r_x, r_y, r_z]" + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "content": ". Then the 3D Gaussians can be splatted to a 2D camera plane through differential Gaussian splatting. To model the appearance of 3D Gaussians, spherical harmonics (SH) are introduced to define the color " + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "content": ". Finally, for each pixel, the rendering results of 3DGS can be derived by calculating the color contribution of all the related Gaussians. This process is known as " + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 262, + 555, + 382 + ], + "type": "text", + "content": "-blending:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 378, + 388, + 553, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 388, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 378, + 388, + 553, + 422 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i} ^ {N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}", + "image_path": "9d35a1f21805715cd0115a15926456f99f72cc95d040349c02c18c17340aaa88.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 428, + 553, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 553, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 553, + 451 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 428, + 553, + 451 + ], + "type": "inline_equation", + "content": "c_{i},\\alpha_{i}" + }, + { + "bbox": [ + 313, + 428, + 553, + 451 + ], + "type": "text", + "content": " represent the color and density computed from the " + }, + { + "bbox": [ + 313, + 428, + 553, + 451 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 428, + 553, + 451 + ], + "type": "text", + "content": " -th 3D Gaussian." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 452, + 554, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 554, + 523 + ], + "type": "text", + "content": "Recent methods [27, 60, 66] utilize deformation fields that extend 3DGS to 4D space, inspired from NeRF-based methods such as D-NeRF [41]. The structure of the deformation field " + }, + { + "bbox": [ + 313, + 452, + 554, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 452, + 554, + 523 + ], + "type": "text", + "content": " can vary among MLP [66], K-Plane [60] and Polynomial [27], while the deformation process can be summarized as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 382, + 531, + 553, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 531, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 382, + 531, + 553, + 545 + ], + "type": "interline_equation", + "content": "(\\Delta \\mu , \\Delta r, \\Delta s) = \\mathcal {D} (\\mu , t), \\tag {3}", + "image_path": "53a6d47d117fd3f99a0b2475c9c12932eb4877f6a88adcef700b7b3ce509de94.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "type": "text", + "content": "where timestamp " + }, + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "type": "inline_equation", + "content": "t \\in \\mathcal{T}" + }, + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "type": "inline_equation", + "content": "\\mathcal{T} \\in \\mathbb{R}^{T \\times 1}" + }, + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "type": "text", + "content": " linear time inputs, " + }, + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "type": "inline_equation", + "content": "\\mu, r, s" + }, + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "type": "text", + "content": " are the center position, rotation quaternion, and scaling factors of 3D Gaussians, and " + }, + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "type": "inline_equation", + "content": "\\Delta \\mu, \\Delta r, \\Delta s" + }, + { + "bbox": [ + 313, + 552, + 554, + 600 + ], + "type": "text", + "content": " are their residuals, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 609, + 370, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 370, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 370, + 621 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "In this section, we first provide an overview of the enhanced deformable 3D Gaussians reconstruction with the proposed TimeFormer (Sec. 4.1). We then introduce the TimeFormer in detail (Sec. 4.2) which consists of a Cross-Temporal Encoder and shared deformation fields (Sec. 4.3). We also discuss the insights behind TimeFormer from the view of gradient flow." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8723" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 556, + 202 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 556, + 202 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 556, + 202 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 556, + 202 + ], + "type": "image", + "image_path": "5b6bd7ab238578bd016fa8c5cb0ad9395d482c0ac9a297a2a556a4a391767be1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 204, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 55, + 204, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 555, + 248 + ], + "type": "text", + "content": "Figure 3. The Framework of Deformable 3D Gaussians Reconstruction with TimeFormer. Existing deformable 3D Gaussians framework usually includes the canonical space and the deformation field (first row), we incorporate TimeFormer to capture cross-time relationships and explore motion patterns implicitly (second row). We share weights of two deformation fields to transfer the learned motion knowledge. This allows us to exclude this Auxiliary Training Module during inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 99, + 254, + 253, + 388 + ], + "blocks": [ + { + "bbox": [ + 99, + 254, + 253, + 388 + ], + "lines": [ + { + "bbox": [ + 99, + 254, + 253, + 388 + ], + "spans": [ + { + "bbox": [ + 99, + 254, + 253, + 388 + ], + "type": "image", + "image_path": "80eb44fe0b95b28ef390966ae250d0731cd90aeacf0c4463daa07356ce68b0fa.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 399, + 296, + 454 + ], + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 454 + ], + "type": "text", + "content": "Figure 4. The Structure of Cross-Temporal Encoder. We concatenate randomly sampled timestamps to position " + }, + { + "bbox": [ + 55, + 399, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 399, + 296, + 454 + ], + "type": "text", + "content": ", treating them as special tokens in a sequence. This module is designed to model multi-temporal relationships and produce distinct time-variant position offsets " + }, + { + "bbox": [ + 55, + 399, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\Delta p_0, \\ldots, \\Delta p_{B-1}" + }, + { + "bbox": [ + 55, + 399, + 296, + 454 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 91, + 458, + 262, + 506 + ], + "blocks": [ + { + "bbox": [ + 91, + 458, + 262, + 506 + ], + "lines": [ + { + "bbox": [ + 91, + 458, + 262, + 506 + ], + "spans": [ + { + "bbox": [ + 91, + 458, + 262, + 506 + ], + "type": "image", + "image_path": "28776ba2a191a383669d96123c17102aeb2d8a3fa26e7b7d2ed65afbb0d40170.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 509, + 295, + 532 + ], + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 532 + ], + "type": "text", + "content": "Figure 5. Data Flow Changes in the Deformation Field. Dashed lines represent new data flow among time samples " + }, + { + "bbox": [ + 55, + 509, + 295, + 532 + ], + "type": "inline_equation", + "content": "t_0, \\dots, t_{B-1}" + }, + { + "bbox": [ + 55, + 509, + 295, + 532 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 124, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 124, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 124, + 551 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Previous methods model motion patterns by explicitly learning temporal relationships on individual or neighboring timestamps, failing on those complex scenes containing violent movement or dynamic reflective surfaces. In contrast, we present TimeFormer to enable the deformable 3D Gaussian backbones themselves to model cross-time relationships from an implicit learning perspective. The main framework with the proposed TimeFormer is shown in Fig. 3. Our approach retains standard reconstruction modules, which include (1) 3D Gaussians in the canonical space and (2) a deformation field that applies time-variant transformation. Additionally, TimeFormer is introduced before the deformation field to extract implicit cross-time motion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 255, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 327 + ], + "type": "text", + "content": "features for each Gaussian through a self-attention mechanism along the time dimension. Moreover, we share the weights of two deformation fields to transfer the motion knowledge from TimeFormer to mitigate the gap between the original branch and TimeFormer branch, which supports real-time rendering without TimeFormer during inference." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 336, + 398, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 398, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 398, + 348 + ], + "type": "text", + "content": "4.2. TimeFormer" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": "Cross-Temporal Encoder. BatchFormer [18] demonstrates that the attention mechanism helps learn sample relationships from batch dimension, rather than channel and spatial dimensions [6, 52], inspiring us that different timestamps can also be considered as a special time batch. Let " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s\\subset \\mathcal{T},\\mathcal{T}_s = \\{t_i\\}_{i = 0}^{B - 1}" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": " denotes randomly sampled timestamps, and let " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{G}\\in \\mathbb{R}^{N\\times (3 + C)}" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": " denotes Gaussians in the canonical space, where " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": " is the size of time batch, " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": " is the number of Gaussians, " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "3 + C" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": " means each Gaussian has 3 position channels and " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": " additional channels. As in Fig. 4, all Gaussians's positions are made into " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": " copies, expanded into " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_c\\in \\mathbb{R}^{B\\times N\\times 3}" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": ", and sampled timestamps are made into " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": " copies, expanded into " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'\\in \\mathbb{R}^{B\\times N\\times 1}" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": ". Then, we compose " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_c" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{T}'" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": " together and apply position encoding function " + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 354, + 556, + 536 + ], + "type": "text", + "content": " to extract high frequency information, as in Eq. 4:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 544, + 553, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 544, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 354, + 544, + 553, + 577 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\gamma (p) = \\left(\\sin \\left(2 ^ {0} \\pi p\\right), \\cos \\left(2 ^ {0} \\pi p\\right), \\dots \\right. \\\\ \\sin \\left(2 ^ {L - 1} \\pi p\\right), \\cos \\left(2 ^ {L - 1} \\pi p\\right)) \\tag {4} \\\\ \\end{array}", + "image_path": "44273ede873def4a6cf15702625c77895f1f2f37c78bf417f38cdedae8b63a18.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": "In our experiments, we set " + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "inline_equation", + "content": "L = 6" + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": " for both positions " + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": ". We treat " + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "inline_equation", + "content": "[\\gamma(\\mathcal{G}_c, \\gamma(\\mathcal{T}')] \\in \\mathbb{R}^{B \\times N \\times (3 \\times 2L + 2L)}" + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": " as the original input " + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "inline_equation", + "content": "F_0" + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": " to TimeFormer. It can be considered as " + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": " sequences of the length " + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": ", containing " + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "inline_equation", + "content": "8L" + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": " feature channels. With " + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": " transformer encoder layers, for " + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "inline_equation", + "content": "m^{th}" + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": " layer, intermediate features are calculated through multi-head self-attention(MSA) and MLP blocks. In the final stage, we use a tiny MLP to transform the last encoded features " + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "inline_equation", + "content": "F_{M-1}" + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": " into offset " + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "inline_equation", + "content": "\\mathcal{O} \\in \\mathbb{R}^{B \\times N \\times 3}" + }, + { + "bbox": [ + 313, + 582, + 555, + 690 + ], + "type": "text", + "content": " in the linear space:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 700, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 700, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 358, + 700, + 555, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {O} = M L P \\left(F _ {M - 1}\\right), \\quad \\mathcal {G} _ {t} = \\mathcal {G} _ {c} + \\mathcal {O} \\tag {5}", + "image_path": "c9c16ac6f2e26d0fd2c3802d04403d6c2d0ec9cf82b99d249eec67117ccd8492.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8724" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "We consider the output from the Cross-Temporal Encoder as a fixing residual term to the original positions to encourage a gradual, steady learning process on motion patterns. Implicit cross-time relationship learning in TimeFormer enables the automatic aggregation of Gaussians with similar variations during optimization, promoting more efficient spatial distribution and accelerating rendering speed." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "type": "text", + "content": "Gradient Analysis. Let " + }, + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "type": "text", + "content": " be the deformation field and TimeFormer respectively, the output of the deformation field " + }, + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "type": "inline_equation", + "content": "\\Delta \\mu_{i}" + }, + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "type": "text", + "content": " and partial derivative for " + }, + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "type": "text", + "content": " in time " + }, + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "type": "inline_equation", + "content": "t_i\\in \\mathcal{T}_s" + }, + { + "bbox": [ + 55, + 156, + 294, + 203 + ], + "type": "text", + "content": " are formulated as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 210, + 295, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 210, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 94, + 210, + 295, + 236 + ], + "type": "interline_equation", + "content": "\\Delta \\mu_ {i} = \\mathcal {D} (\\mu , t _ {i}), \\quad \\frac {\\partial \\Delta \\mu_ {i}}{\\partial \\mu} = \\frac {\\partial \\mathcal {D} (\\mu , t _ {i})}{\\partial \\mu} \\tag {6}", + "image_path": "75c677bfc99beb69b7ecb833b5b4521d39f6bf3d653051e1356331dfc96e7307.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 244, + 294, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 294, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 294, + 269 + ], + "type": "text", + "content": "With TimeFormer applied on time batch " + }, + { + "bbox": [ + 55, + 244, + 294, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_s" + }, + { + "bbox": [ + 55, + 244, + 294, + 269 + ], + "type": "text", + "content": ", we reformulate as Eq. 7 and data flow changes as in Fig. 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 278, + 295, + 292 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 278, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 118, + 278, + 295, + 292 + ], + "type": "interline_equation", + "content": "\\Delta \\mu_ {i} = \\mathcal {D} (\\mu + \\mathcal {P} (\\mu , \\mathcal {T} _ {s}), t _ {i}) \\tag {7}", + "image_path": "a1c3719180def751c2765b2c2541492a97b6923789f9f76b30c26e73608221d6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 301, + 295, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 325 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 325 + ], + "type": "text", + "content": "To calculate the partial derivative for " + }, + { + "bbox": [ + 55, + 301, + 295, + 325 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 301, + 295, + 325 + ], + "type": "text", + "content": ", we apply the chain rule. Let " + }, + { + "bbox": [ + 55, + 301, + 295, + 325 + ], + "type": "inline_equation", + "content": "a = \\mu + \\mathcal{P}(\\mu, \\mathcal{T}_s)" + }, + { + "bbox": [ + 55, + 301, + 295, + 325 + ], + "type": "text", + "content": ", then:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 333, + 295, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 333, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 67, + 333, + 295, + 360 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\Delta \\mu_ {i}}{\\partial \\mu} = \\frac {\\partial \\mathcal {D} (a , t _ {i})}{\\partial a} \\cdot \\frac {\\partial a}{\\partial \\mu}, \\quad \\frac {\\partial a}{\\partial \\mu} = 1 + \\frac {\\partial \\mathcal {P} (\\mu , \\mathcal {T} _ {s})}{\\partial \\mu} \\tag {8}", + "image_path": "d1f2214b216339a0272137aa896575d2a94b7761c6e3b39a730909e42f302e69.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 371, + 295, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 371, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 93, + 371, + 295, + 399 + ], + "type": "interline_equation", + "content": "\\frac {\\partial \\Delta \\mu_ {i}}{\\partial \\mu} = \\frac {\\partial \\mathcal {D} (a , t _ {i})}{\\partial a} \\cdot \\left(1 + \\frac {\\partial \\mathcal {P} (\\mu , \\mathcal {T} _ {s})}{\\partial \\mu}\\right) \\tag {9}", + "image_path": "88469c75f45d173f3e436b9ad8d29e25b95bc6ef50e143e39948f77fe6b6d124.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 403, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 523 + ], + "type": "text", + "content": "Eq. 6 has a weaker dynamic nature, as it only considers the current timestamp " + }, + { + "bbox": [ + 55, + 403, + 295, + 523 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 403, + 295, + 523 + ], + "type": "text", + "content": ". In contrast, Eq. 9 incorporates an additional gradient term, " + }, + { + "bbox": [ + 55, + 403, + 295, + 523 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\mathcal{P}(\\mu, \\mathcal{T}_s)}{\\partial \\mu}" + }, + { + "bbox": [ + 55, + 403, + 295, + 523 + ], + "type": "text", + "content": ", enabling the current state to be influenced by any past or future states. In other words, " + }, + { + "bbox": [ + 55, + 403, + 295, + 523 + ], + "type": "inline_equation", + "content": "\\Delta \\mu_i" + }, + { + "bbox": [ + 55, + 403, + 295, + 523 + ], + "type": "text", + "content": " also optimizes the models according to other timestamps " + }, + { + "bbox": [ + 55, + 403, + 295, + 523 + ], + "type": "inline_equation", + "content": "t_j \\in \\mathcal{T}_s (j \\neq i)" + }, + { + "bbox": [ + 55, + 403, + 295, + 523 + ], + "type": "text", + "content": ", which is a significant difference compared to the backward process without Time-Former. Such design accounts for cross-time attention and allows the models to capture more challenging motion patterns from a global view of the entire time series." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 531, + 205, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 205, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 205, + 544 + ], + "type": "text", + "content": "4.3. Shared Deformation Fields" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 549, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 295, + 597 + ], + "type": "text", + "content": "Accounting for additional computation costs of Time-Former, which can significantly decrease rendering speed, we force the original deformation field and the auxiliary deformation field to share weights for knowledge transferring." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 597, + 295, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 669 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 597, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 597, + 295, + 669 + ], + "type": "text", + "content": " be the camera viewpoints, we apply the shared deformation fields to predict deformed space from both Gaussians in the canonical space and Gaussians from Time-Former. Then, we apply the splatting algorithm to these two groups of deformed space. We calculate the losses between rendered images and ground truth " + }, + { + "bbox": [ + 55, + 597, + 295, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{gt}" + }, + { + "bbox": [ + 55, + 597, + 295, + 669 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 92, + 678, + 295, + 692 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 678, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 92, + 678, + 295, + 692 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\left\\| S p l a t t i n g \\left(\\mathcal {D} \\left(\\mathcal {G} _ {c}, \\mathcal {T}\\right), \\mathcal {V}\\right) - \\mathcal {I} _ {g t} \\right\\| _ {1} \\tag {10}", + "image_path": "588e9d62e54174a0d2770b6d102789143b50905132d6e03533fae03068af5a51.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 93, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 93, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = \\| S p l a t t i n g (\\mathcal {D} (\\mathcal {G} _ {t}, \\mathcal {T}), \\mathcal {V}) - \\mathcal {I} _ {g t} \\| _ {1} \\tag {11}", + "image_path": "f74d8ae624377b72466913262106d32cdf8f2b4eedacd585735e6f5662538a82.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 396, + 72, + 553, + 84 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 72, + 553, + 84 + ], + "spans": [ + { + "bbox": [ + 396, + 72, + 553, + 84 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {c} \\mathcal {L} _ {c} + \\lambda_ {t} \\mathcal {L} _ {t} \\tag {12}", + "image_path": "febf61fee22d77d71a3c7b392bc3b29e3f1942617d624e4fa0b37c543bcbdf86.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 91, + 553, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 138 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 91, + 553, + 138 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c, \\mathcal{L}_t" + }, + { + "bbox": [ + 313, + 91, + 553, + 138 + ], + "type": "text", + "content": " represent losses of original branch and Time-Former branch with " + }, + { + "bbox": [ + 313, + 91, + 553, + 138 + ], + "type": "inline_equation", + "content": "\\lambda_c > \\lambda_t" + }, + { + "bbox": [ + 313, + 91, + 553, + 138 + ], + "type": "text", + "content": ". We use a relatively smaller " + }, + { + "bbox": [ + 313, + 91, + 553, + 138 + ], + "type": "inline_equation", + "content": "\\lambda_t" + }, + { + "bbox": [ + 313, + 91, + 553, + 138 + ], + "type": "text", + "content": " because we find it easy to overfit on the second branch with TimeFormer, causing a degradation in inference quality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 148, + 391, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 148, + 391, + 163 + ], + "spans": [ + { + "bbox": [ + 314, + 148, + 391, + 163 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 168, + 447, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 168, + 447, + 181 + ], + "spans": [ + { + "bbox": [ + 314, + 168, + 447, + 181 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "text", + "content": "Our implementation is tested on a single A100 GPU. We use " + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "text", + "content": " transformer encoder layers in TimeFormer and the number of time samples is " + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "inline_equation", + "content": "B = 4" + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "inline_equation", + "content": "\\lambda_c = 1" + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "inline_equation", + "content": "\\lambda_t = 0.8" + }, + { + "bbox": [ + 313, + 186, + 553, + 270 + ], + "type": "text", + "content": " to prevent overfitting on TimeFormer. We assess our experimental results using image quality metrics, including peak-signal-to-noise ratio (PSNR) and structural similarity index (SSIM [59])." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 270, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 554, + 354 + ], + "type": "text", + "content": "Baselines & Datasets. We apply TimeFormer to 4DGS [60], STGS [27] on multi-view videos, e.g., N3DV Dataset [25], and evaluate TimeFormer on 4DGS [60] and DeformGS [66] on monocular dynamic videos, e.g., HyperNeRF Dataset [39], respectively. We also utilize NeRF-DS Dataset [64] to demonstrate the robustness of TimeFormer on moving objects with specular surfaces." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 361, + 466, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 466, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 466, + 373 + ], + "type": "text", + "content": "5.2. Experimental Comparisons" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 378, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 461 + ], + "type": "text", + "content": "Multi-View Video Dataset. In Tab. 1, TimeFormer improves the reconstruction quality (e.g., PSNR) of original 4DGS [60] and STGS [27] by 0.74 and 0.61, outperforming all baselines methods. We extract the frames at an initial time for scene Cook Spanish in Fig. 6, and TimeFormer can reconstruct a clearer geometry and specular effects compared to the original results." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 463, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 557 + ], + "type": "text", + "content": "To have a more comprehensive overview of Time-Former's capacity on different temporal stages, we collect PSNR for 300 frames of all six scenes from N3DV Dataset [25] and calculate the average per-frame PSNR. As in Fig. 7, the reconstruction quality increases more significantly at the beginning and end, and such balanced reconstruction results are attributed to TimeFormer's cross-temporal attention mechanism on the whole time series." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 558, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 676 + ], + "type": "text", + "content": "Monocular Video Dataset. Tab. 3 shows the enhancements achieved by TimeFormer on the HyperNeRF Dataset. It's observed that TimeFormer increases the PSNR by 0.89 for 4DGS [60] and 0.94 for DeformGS [66], and the SSIM by 0.019 and 0.028, respectively. Fig. 9 additionally provides visualization of the improvement in the image quality: on the left, TimeFormer generates a clearer outline between fingers, while on the right the texture on the front of the broom is clearer. TimeFormer also eliminates artifacts and reveals a clear structure of the broom." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Dynamic Reflective Video Dataset. Fig. 10 additionally demonstrates that TimeFormer can work well on scenes with dynamic specular objects on NeRF-DS Dataset [64]." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8725" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 69, + 558, + 196 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 558, + 196 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 558, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 558, + 196 + ], + "type": "table", + "html": "
MethodSear SteakFlame SSIMSalmon SSIMCut Beef SSIMFlame SSINRSteak SSIMCook Spinach SSINRSSIMCoffee SSINRMartini SSIMMean
PSNRSSIMPSNRSSIM
K-Plane [14]32.520.97130.440.94231.820.96532.380.97030.600.96829.990.94331.63
MixVovels [55]31.210.97129.920.94531.300.96531.430.97031.610.96529.360.94630.80
GS4D [67]32.920.95326.390.89733.080.95933.810.96732.770.95625.230.88430.07
4D-Rotor [9]32.860.95628.250.91333.140.95231.610.95332.560.94927.950.90831.06
4DGS [60]32.490.94928.920.91732.900.95632.510.95432.460.94827.340.90331.10
+TimeFormer33.380.95529.330.92433.110.95733.250.95333.030.94928.930.91031.84
STGS [27]33.710.96228.210.92133.520.95833.460.96333.130.95527.710.91531.62
+TimeFormer34.340.96529.130.92433.570.95834.040.96433.450.95628.830.91732.23
", + "image_path": "00d6885c985cbc500c23f25fdf81f5443ae2cfafc988b84f0d2f900bf79a8ab2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 198, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 222 + ], + "type": "text", + "content": "Table 1. Quantitative Comparisons on N3DV Dataset [25]. We use bold font to indicate the improvement, statistics of K-Plane [14] and MixVoxels [55] are from their original paper, while we calculate metrics of all Gaussian-based methods by running their official codes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 56, + 232, + 156, + 365 + ], + "blocks": [ + { + "bbox": [ + 56, + 232, + 156, + 365 + ], + "lines": [ + { + "bbox": [ + 56, + 232, + 156, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 156, + 365 + ], + "type": "image", + "image_path": "1953b579de1e33da08265ef263e1ef704b7e78269a1ec5fe933b2a0348e1968a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 156, + 232, + 254, + 365 + ], + "blocks": [ + { + "bbox": [ + 156, + 232, + 254, + 365 + ], + "lines": [ + { + "bbox": [ + 156, + 232, + 254, + 365 + ], + "spans": [ + { + "bbox": [ + 156, + 232, + 254, + 365 + ], + "type": "image", + "image_path": "6ec2eafbdc5e22dcda778ea59587681b2e915465db2ffc82a126e0c3d7fa7398.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 373, + 420, + 385 + ], + "lines": [ + { + "bbox": [ + 189, + 373, + 420, + 385 + ], + "spans": [ + { + "bbox": [ + 189, + 373, + 420, + 385 + ], + "type": "text", + "content": "Figure 6. Visualization of Comparisons on N3DV Dataset [25]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 255, + 232, + 354, + 365 + ], + "blocks": [ + { + "bbox": [ + 255, + 232, + 354, + 365 + ], + "lines": [ + { + "bbox": [ + 255, + 232, + 354, + 365 + ], + "spans": [ + { + "bbox": [ + 255, + 232, + 354, + 365 + ], + "type": "image", + "image_path": "32026f7bef44af4dfcbc0d87200b9546234bbfdbdebc365982f538eccf83330e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 354, + 232, + 454, + 365 + ], + "blocks": [ + { + "bbox": [ + 354, + 232, + 454, + 365 + ], + "lines": [ + { + "bbox": [ + 354, + 232, + 454, + 365 + ], + "spans": [ + { + "bbox": [ + 354, + 232, + 454, + 365 + ], + "type": "image", + "image_path": "62b48a06311c8803e6d29909eff5674607083d784b65be180f27b417fcab333f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 455, + 232, + 554, + 365 + ], + "blocks": [ + { + "bbox": [ + 455, + 232, + 554, + 365 + ], + "lines": [ + { + "bbox": [ + 455, + 232, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 455, + 232, + 554, + 365 + ], + "type": "image", + "image_path": "da94bbadd807988c465daf84fd10986d831afb8edfd4c3e91b2abb3dce654e08.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 67, + 403, + 285, + 494 + ], + "blocks": [ + { + "bbox": [ + 67, + 403, + 285, + 494 + ], + "lines": [ + { + "bbox": [ + 67, + 403, + 285, + 494 + ], + "spans": [ + { + "bbox": [ + 67, + 403, + 285, + 494 + ], + "type": "image", + "image_path": "5385e5e9c2ca0018d6ff00bd8373c5c0ca03fa0c28a0024005b1f004ae5b4861.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 501, + 295, + 534 + ], + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 534 + ], + "type": "text", + "content": "Figure 7. Per-Frame PSNR on N3DV Dataset [25]. TimeFormer improves the reconstruction quality, especially at the beginning and end time series." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 557, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 616 + ], + "type": "text", + "content": "TimeFormer can reconstruct a cleaner appearance on reflective material (e.g., a metallic cup, a glass bottle), and this improvement is attributed to the attention mechanism which enables the deformation field to detect changes on specular surfaces from a learning perspective automatically." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Analysis of FPS & Canonical Space. We also find that TimeFormer achieves more efficient representation with much fewer points in the final canonical space, thereby achieving a higher FPS in Tab. 2. We attribute the phenomenon to the reason that TimeFormer guides faster gradient descent and promotes the spatial layout of Gaussian points during optimization. Interestingly, in Fig. 8, the L1 loss of both branches in our method decreases" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 336, + 410, + 533, + 509 + ], + "blocks": [ + { + "bbox": [ + 336, + 410, + 533, + 509 + ], + "lines": [ + { + "bbox": [ + 336, + 410, + 533, + 509 + ], + "spans": [ + { + "bbox": [ + 336, + 410, + 533, + 509 + ], + "type": "image", + "image_path": "f29f1ee14da7316287f670665b2067da9c1e5a95d60286b3bb38e234d09c4327.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 521, + 555, + 553 + ], + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 553 + ], + "type": "text", + "content": "Figure 8. Comparisons of Convergence Speed. We use the same batch size for 4DGS [60] and calculate the loss of both branches in our method." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 579, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 663 + ], + "type": "text", + "content": "faster and shows almost the same convergence trend during training, exceeding the optimization speed of the original method. This eliminates many redundant points and guides the canonical space toward a more efficient distribution, as in Fig. 2. Moreover, Fig. 8 also proves that the cross-time relationship learned in TimeFormer has been successfully transferred to the base branch, as introduced in Sec. 4.3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Analysis of Motion Patterns. We argue that TimeFormer achieves more robust learning of motion patterns. To take a deeper insight, we visualize the motion changes as follows: for each Gaussian, we calculate its bias " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "(\\Delta_x, \\Delta_y, \\Delta_z)" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8726" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 105, + 156 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 105, + 156 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 105, + 156 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 105, + 156 + ], + "type": "image", + "image_path": "83bf71128457713c2e5b31099a96df3930484451c7af3415d374a0664849a128.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 158, + 104, + 201 + ], + "blocks": [ + { + "bbox": [ + 59, + 158, + 104, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 158, + 104, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 158, + 104, + 201 + ], + "type": "image", + "image_path": "a4a8920e947b25e399638c6cddcd2083c61301303318b285911de78fbfbda1ff.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 202, + 100, + 217 + ], + "lines": [ + { + "bbox": [ + 66, + 202, + 100, + 217 + ], + "spans": [ + { + "bbox": [ + 66, + 202, + 100, + 217 + ], + "type": "text", + "content": "DeformGS PSNR:27.63" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 107, + 70, + 155, + 156 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 155, + 156 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 155, + 156 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 155, + 156 + ], + "type": "image", + "image_path": "3e86b021e18a22956a95a60c93984286616ebd4f37e4485d3b9bb25fdd1cf39d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 109, + 158, + 154, + 201 + ], + "blocks": [ + { + "bbox": [ + 109, + 158, + 154, + 201 + ], + "lines": [ + { + "bbox": [ + 109, + 158, + 154, + 201 + ], + "spans": [ + { + "bbox": [ + 109, + 158, + 154, + 201 + ], + "type": "image", + "image_path": "99b0a988d13804c01e166678c2cf8615ced9c2ef71aeab62ef0d045ee5a682e1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 202, + 153, + 217 + ], + "lines": [ + { + "bbox": [ + 112, + 202, + 153, + 217 + ], + "spans": [ + { + "bbox": [ + 112, + 202, + 153, + 217 + ], + "type": "text", + "content": "+TimeFormer PSNR:28.59" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 156, + 70, + 204, + 156 + ], + "blocks": [ + { + "bbox": [ + 156, + 70, + 204, + 156 + ], + "lines": [ + { + "bbox": [ + 156, + 70, + 204, + 156 + ], + "spans": [ + { + "bbox": [ + 156, + 70, + 204, + 156 + ], + "type": "image", + "image_path": "4e3eb461b7d0a5dcbcec9a624a9e0a7a94858a7b351c819a3bb8bf6456f0d1fd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 157, + 158, + 203, + 201 + ], + "blocks": [ + { + "bbox": [ + 157, + 158, + 203, + 201 + ], + "lines": [ + { + "bbox": [ + 157, + 158, + 203, + 201 + ], + "spans": [ + { + "bbox": [ + 157, + 158, + 203, + 201 + ], + "type": "image", + "image_path": "6a3c4d553a845fd812df9bb676a241b350919c7b1461400f2d7257fe30591768.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 202, + 199, + 217 + ], + "lines": [ + { + "bbox": [ + 164, + 202, + 199, + 217 + ], + "spans": [ + { + "bbox": [ + 164, + 202, + 199, + 217 + ], + "type": "text", + "content": "4DGS PSNR:28.55" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 206, + 70, + 254, + 156 + ], + "blocks": [ + { + "bbox": [ + 206, + 70, + 254, + 156 + ], + "lines": [ + { + "bbox": [ + 206, + 70, + 254, + 156 + ], + "spans": [ + { + "bbox": [ + 206, + 70, + 254, + 156 + ], + "type": "image", + "image_path": "e6e8358bc3ac0cb5eb4433d2cab300554d7f0e6decce986e78f7127abcbb0668.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 208, + 158, + 253, + 201 + ], + "blocks": [ + { + "bbox": [ + 208, + 158, + 253, + 201 + ], + "lines": [ + { + "bbox": [ + 208, + 158, + 253, + 201 + ], + "spans": [ + { + "bbox": [ + 208, + 158, + 253, + 201 + ], + "type": "image", + "image_path": "25454d64dca6f24a9919a8941575cae6a4fe15683ba04ff7827d9fdf2b904181.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 202, + 250, + 217 + ], + "lines": [ + { + "bbox": [ + 211, + 202, + 250, + 217 + ], + "spans": [ + { + "bbox": [ + 211, + 202, + 250, + 217 + ], + "type": "text", + "content": "+TimeFormer PSNR:29.51" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 255, + 70, + 304, + 156 + ], + "blocks": [ + { + "bbox": [ + 255, + 70, + 304, + 156 + ], + "lines": [ + { + "bbox": [ + 255, + 70, + 304, + 156 + ], + "spans": [ + { + "bbox": [ + 255, + 70, + 304, + 156 + ], + "type": "image", + "image_path": "e227a24b7cc91f7884655ccd8e833f855b7bec2586c6dfe8545f3b77d22044b1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 257, + 158, + 302, + 201 + ], + "blocks": [ + { + "bbox": [ + 257, + 158, + 302, + 201 + ], + "lines": [ + { + "bbox": [ + 257, + 158, + 302, + 201 + ], + "spans": [ + { + "bbox": [ + 257, + 158, + 302, + 201 + ], + "type": "image", + "image_path": "77a375c9d9e614eeed111ccf790af346748f75352babf27b1f8525c75c7aa034.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 202, + 299, + 217 + ], + "lines": [ + { + "bbox": [ + 262, + 202, + 299, + 217 + ], + "spans": [ + { + "bbox": [ + 262, + 202, + 299, + 217 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 353, + 156 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 353, + 156 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 353, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 353, + 156 + ], + "type": "image", + "image_path": "2ad325931a0d50b4a3dccdcd456c7e75f946961dc92323367418cd44208e51f8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 307, + 158, + 352, + 201 + ], + "blocks": [ + { + "bbox": [ + 307, + 158, + 352, + 201 + ], + "lines": [ + { + "bbox": [ + 307, + 158, + 352, + 201 + ], + "spans": [ + { + "bbox": [ + 307, + 158, + 352, + 201 + ], + "type": "image", + "image_path": "92b543abe69218a2214023fb87267060df7370ee96d2a62e3702aa297d262770.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 202, + 348, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 202, + 348, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 348, + 217 + ], + "type": "text", + "content": "DeformGS PSNR:20.63" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 356, + 70, + 404, + 156 + ], + "blocks": [ + { + "bbox": [ + 356, + 70, + 404, + 156 + ], + "lines": [ + { + "bbox": [ + 356, + 70, + 404, + 156 + ], + "spans": [ + { + "bbox": [ + 356, + 70, + 404, + 156 + ], + "type": "image", + "image_path": "a3401f94a25153b73a4a7ad2ea72c93623d2422ef4eed5b6d405e8f527771966.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 358, + 158, + 402, + 201 + ], + "blocks": [ + { + "bbox": [ + 358, + 158, + 402, + 201 + ], + "lines": [ + { + "bbox": [ + 358, + 158, + 402, + 201 + ], + "spans": [ + { + "bbox": [ + 358, + 158, + 402, + 201 + ], + "type": "image", + "image_path": "f1e6434f642df0e02b72666f406f1108467a231e81ac71b77f6c699c765f0458.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 202, + 400, + 217 + ], + "lines": [ + { + "bbox": [ + 361, + 202, + 400, + 217 + ], + "spans": [ + { + "bbox": [ + 361, + 202, + 400, + 217 + ], + "type": "text", + "content": "+TimeFormer PSNR:21.13" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 406, + 70, + 454, + 156 + ], + "blocks": [ + { + "bbox": [ + 406, + 70, + 454, + 156 + ], + "lines": [ + { + "bbox": [ + 406, + 70, + 454, + 156 + ], + "spans": [ + { + "bbox": [ + 406, + 70, + 454, + 156 + ], + "type": "image", + "image_path": "dd7da060915c87ee872b913a6556c2574c1d5bc20cbd9b6257db98183995073e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 408, + 158, + 452, + 201 + ], + "blocks": [ + { + "bbox": [ + 408, + 158, + 452, + 201 + ], + "lines": [ + { + "bbox": [ + 408, + 158, + 452, + 201 + ], + "spans": [ + { + "bbox": [ + 408, + 158, + 452, + 201 + ], + "type": "image", + "image_path": "486c3d8b531296b2be472980f5879879334757fa3df4c72b34afd231a9553f69.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 202, + 447, + 217 + ], + "lines": [ + { + "bbox": [ + 412, + 202, + 447, + 217 + ], + "spans": [ + { + "bbox": [ + 412, + 202, + 447, + 217 + ], + "type": "text", + "content": "4DGS PSNR:21.28" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 455, + 70, + 503, + 156 + ], + "blocks": [ + { + "bbox": [ + 455, + 70, + 503, + 156 + ], + "lines": [ + { + "bbox": [ + 455, + 70, + 503, + 156 + ], + "spans": [ + { + "bbox": [ + 455, + 70, + 503, + 156 + ], + "type": "image", + "image_path": "2e1e26777659cdaeac54f4ede6be381ece76d8c4f18028757c1f1d78c2257f54.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 457, + 158, + 502, + 201 + ], + "blocks": [ + { + "bbox": [ + 457, + 158, + 502, + 201 + ], + "lines": [ + { + "bbox": [ + 457, + 158, + 502, + 201 + ], + "spans": [ + { + "bbox": [ + 457, + 158, + 502, + 201 + ], + "type": "image", + "image_path": "a002a9ba9de71dc98c829e62ea673044c8b3f5bfbb8e57a5358d9d7d69242cfa.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 202, + 500, + 217 + ], + "lines": [ + { + "bbox": [ + 459, + 202, + 500, + 217 + ], + "spans": [ + { + "bbox": [ + 459, + 202, + 500, + 217 + ], + "type": "text", + "content": "+TimeFormer PSNR:22.23" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 506, + 70, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 506, + 70, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 506, + 70, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 506, + 70, + 553, + 156 + ], + "type": "image", + "image_path": "479af13db44a726d37af437335040c70701c69b4e5f41cbec9c49e2f3a4ca293.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 507, + 158, + 551, + 201 + ], + "blocks": [ + { + "bbox": [ + 507, + 158, + 551, + 201 + ], + "lines": [ + { + "bbox": [ + 507, + 158, + 551, + 201 + ], + "spans": [ + { + "bbox": [ + 507, + 158, + 551, + 201 + ], + "type": "image", + "image_path": "caada894c47253a00648e58a31d9d65af5fe811f36b4283875cfc40b0119cd64.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 510, + 202, + 548, + 210 + ], + "lines": [ + { + "bbox": [ + 510, + 202, + 548, + 210 + ], + "spans": [ + { + "bbox": [ + 510, + 202, + 548, + 210 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 60, + 243, + 167, + 297 + ], + "blocks": [ + { + "bbox": [ + 179, + 221, + 430, + 232 + ], + "lines": [ + { + "bbox": [ + 179, + 221, + 430, + 232 + ], + "spans": [ + { + "bbox": [ + 179, + 221, + 430, + 232 + ], + "type": "text", + "content": "Figure 9. Visualization of Comparisons on HyperNeRF Dataset [39]." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 243, + 167, + 297 + ], + "lines": [ + { + "bbox": [ + 60, + 243, + 167, + 297 + ], + "spans": [ + { + "bbox": [ + 60, + 243, + 167, + 297 + ], + "type": "image", + "image_path": "636636a55414d32f71812b1047e518df9603037018f16c81439f5345f79b4698.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 168, + 243, + 263, + 297 + ], + "blocks": [ + { + "bbox": [ + 168, + 243, + 263, + 297 + ], + "lines": [ + { + "bbox": [ + 168, + 243, + 263, + 297 + ], + "spans": [ + { + "bbox": [ + 168, + 243, + 263, + 297 + ], + "type": "image", + "image_path": "90c51d445920d79dee882b75a1abb57a77d629cfbe95365898d5c07a09fa0664.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 264, + 243, + 359, + 297 + ], + "blocks": [ + { + "bbox": [ + 264, + 243, + 359, + 297 + ], + "lines": [ + { + "bbox": [ + 264, + 243, + 359, + 297 + ], + "spans": [ + { + "bbox": [ + 264, + 243, + 359, + 297 + ], + "type": "image", + "image_path": "9a78033a76223bd4e7efbafc92a4b346b5f5de741ebf5c43344ed5bd8502112f.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 360, + 243, + 457, + 298 + ], + "blocks": [ + { + "bbox": [ + 360, + 243, + 457, + 298 + ], + "lines": [ + { + "bbox": [ + 360, + 243, + 457, + 298 + ], + "spans": [ + { + "bbox": [ + 360, + 243, + 457, + 298 + ], + "type": "image", + "image_path": "4a3bd1bfdc5b4371a9d9156a2a2357c88be8698083d1320d632217fb3748c436.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 458, + 243, + 553, + 298 + ], + "blocks": [ + { + "bbox": [ + 458, + 243, + 553, + 298 + ], + "lines": [ + { + "bbox": [ + 458, + 243, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 458, + 243, + 553, + 298 + ], + "type": "image", + "image_path": "b0e790bde867741af637cf11ad6acc1b33764edd6acdc8c61920bbf439455938.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 60, + 299, + 166, + 354 + ], + "blocks": [ + { + "bbox": [ + 60, + 299, + 166, + 354 + ], + "lines": [ + { + "bbox": [ + 60, + 299, + 166, + 354 + ], + "spans": [ + { + "bbox": [ + 60, + 299, + 166, + 354 + ], + "type": "image", + "image_path": "96c9520128b8114043673795b0ae9145ebe6c7ed66e9b2c80dfd42f638debb7e.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 168, + 299, + 263, + 354 + ], + "blocks": [ + { + "bbox": [ + 168, + 299, + 263, + 354 + ], + "lines": [ + { + "bbox": [ + 168, + 299, + 263, + 354 + ], + "spans": [ + { + "bbox": [ + 168, + 299, + 263, + 354 + ], + "type": "image", + "image_path": "04b69a301ed48abb34f2757c26f02cb1a667868d84784aa32e07ef22e4bebaae.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 264, + 299, + 359, + 354 + ], + "blocks": [ + { + "bbox": [ + 264, + 299, + 359, + 354 + ], + "lines": [ + { + "bbox": [ + 264, + 299, + 359, + 354 + ], + "spans": [ + { + "bbox": [ + 264, + 299, + 359, + 354 + ], + "type": "image", + "image_path": "7f963041b35a5c71889e748965829793d45d4be7befe542c168b493b0afd0dbe.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 360, + 299, + 457, + 354 + ], + "blocks": [ + { + "bbox": [ + 360, + 299, + 457, + 354 + ], + "lines": [ + { + "bbox": [ + 360, + 299, + 457, + 354 + ], + "spans": [ + { + "bbox": [ + 360, + 299, + 457, + 354 + ], + "type": "image", + "image_path": "8f8a2b84a32cc1a366788266b4ae56a56454eb9eda67ba85ff2fbb43a911bab3.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 458, + 299, + 553, + 354 + ], + "blocks": [ + { + "bbox": [ + 458, + 299, + 553, + 354 + ], + "lines": [ + { + "bbox": [ + 458, + 299, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 458, + 299, + 553, + 354 + ], + "type": "image", + "image_path": "f9f643ecef67c785672df5638bad6cd3b0f16277582e2f3a033ef6fd270322d0.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 60, + 355, + 166, + 409 + ], + "blocks": [ + { + "bbox": [ + 60, + 355, + 166, + 409 + ], + "lines": [ + { + "bbox": [ + 60, + 355, + 166, + 409 + ], + "spans": [ + { + "bbox": [ + 60, + 355, + 166, + 409 + ], + "type": "image", + "image_path": "81a2c7df2388b93e95ed341480865c92cd83fa396c1a95a9e4f6adf95396c13b.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 418, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 418, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 555, + 440 + ], + "type": "text", + "content": "Figure 10. Visualization of Comparisons on NeRF-DS Dataset [64]. Compared with original results from DeformGS [66], TimeFormer presents a clearer visual effect on dynamic objects with specular surfaces." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 168, + 355, + 263, + 409 + ], + "blocks": [ + { + "bbox": [ + 168, + 355, + 263, + 409 + ], + "lines": [ + { + "bbox": [ + 168, + 355, + 263, + 409 + ], + "spans": [ + { + "bbox": [ + 168, + 355, + 263, + 409 + ], + "type": "image", + "image_path": "4b0d92a467d6cfaf042031d33b1e9046acfeb3120d39d789ee4d6c72e1064188.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 264, + 355, + 359, + 409 + ], + "blocks": [ + { + "bbox": [ + 264, + 355, + 359, + 409 + ], + "lines": [ + { + "bbox": [ + 264, + 355, + 359, + 409 + ], + "spans": [ + { + "bbox": [ + 264, + 355, + 359, + 409 + ], + "type": "image", + "image_path": "20b2a83316c6bc1f1f6e3383f72e99551995f66ce993daf1a9c02b6d58eae03a.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 360, + 355, + 457, + 409 + ], + "blocks": [ + { + "bbox": [ + 360, + 355, + 457, + 409 + ], + "lines": [ + { + "bbox": [ + 360, + 355, + 457, + 409 + ], + "spans": [ + { + "bbox": [ + 360, + 355, + 457, + 409 + ], + "type": "image", + "image_path": "def1b06f93e0a05a0290988e7b38b2b2f5c3908591da6b0d508d4f127d49f738.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 458, + 355, + 553, + 409 + ], + "blocks": [ + { + "bbox": [ + 458, + 355, + 553, + 409 + ], + "lines": [ + { + "bbox": [ + 458, + 355, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 458, + 355, + 553, + 409 + ], + "type": "image", + "image_path": "460e8424bf472d5f76e830b56a3fafe67406231ed166ed629daac01cba9d6bcb.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "table", + "bbox": [ + 71, + 459, + 280, + 523 + ], + "blocks": [ + { + "bbox": [ + 71, + 459, + 280, + 523 + ], + "lines": [ + { + "bbox": [ + 71, + 459, + 280, + 523 + ], + "spans": [ + { + "bbox": [ + 71, + 459, + 280, + 523 + ], + "type": "table", + "html": "
MethodN ↓Training ↓FPS ↑
STGS [27]172.1 k41 min82.1
+TimeFormer148.9 k78 min88.9
4DGS [60]145.9 k52 min31.7
+TimeFormer113.1 k94 min37.9
", + "image_path": "9ef168bf17071bf0e36cfd832444b3078cd2fdfd7205c82d1531133f7af168fb.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "table_body" + } + ], + "index": 47 + }, + { + "type": "table", + "bbox": [ + 72, + 539, + 280, + 603 + ], + "blocks": [ + { + "bbox": [ + 90, + 524, + 261, + 534 + ], + "lines": [ + { + "bbox": [ + 90, + 524, + 261, + 534 + ], + "spans": [ + { + "bbox": [ + 90, + 524, + 261, + 534 + ], + "type": "text", + "content": "(a) Quantitative comparisons on N3DV Dataset [25]." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 72, + 539, + 280, + 603 + ], + "lines": [ + { + "bbox": [ + 72, + 539, + 280, + 603 + ], + "spans": [ + { + "bbox": [ + 72, + 539, + 280, + 603 + ], + "type": "table", + "html": "
MethodN↓Training ↓FPS ↑
DeformGS [66]169.6 k25 min30.1
+TimeFormer82.9 k35 min58.9
4DGS [60]172.3 k32 min35.7
+TimeFormer135.5 k48 min40.9
", + "image_path": "b0a6987311efd2a24052a6208c8f546c03d400ec68c33507391c12d5247c8d34.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "table_body" + } + ], + "index": 49 + }, + { + "bbox": [ + 80, + 605, + 270, + 615 + ], + "lines": [ + { + "bbox": [ + 80, + 605, + 270, + 615 + ], + "spans": [ + { + "bbox": [ + 80, + 605, + 270, + 615 + ], + "type": "text", + "content": "(b) Quantitative comparisons on HyperNeRF Dataset [39]." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 624, + 295, + 635 + ], + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 635 + ], + "type": "text", + "content": "Table 2. Comparisons of Gaussian Number, Training Time & FPS." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "towards canonical space in time " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " and replace the color attributes " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "(r, g, b)" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " with the absolute value of this bias: " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "(r, g, b) \\gets (|\\Delta_x|, |\\Delta_y|, |\\Delta_z|)" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": ". We use the same splatting algorithm to render the accumulated motion bias, transfer" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 313, + 462, + 520, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 520, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 520, + 473 + ], + "type": "text", + "content": "ring the rendered result as a heatmap, as in Fig. 11." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 313, + 474, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 594 + ], + "type": "text", + "content": "In Fig. 11a, TimeFormer identifies the blade's motion as a rigid deformation, showing a sharper outline and consistent motion, especially as it contacts the initially static lemon. This demonstrates TimeFormer's ability to distinguish between moving and static objects. In Fig. 11b, TimeFormer captures subtle motion on the spray gun and clear flame flickering on the beef. Fig. 11c shows that TimeFormer also detects the violent movement of the elongated broom, where 4DGS [60] fails, proving its robustness in handling extremely geometries." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 313, + 600, + 415, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 415, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 415, + 611 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Cross-Temporal Encoder. Tab. 5 demonstrates that the performance of TimeFormer is not sensitive to time batch " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " and transformer encoder layer " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": ". In Tab. 5b, we explore two sampling methods: random sampling and continuous sampling of timestamps. Both strategies demonstrate improvements in PSNR and other metrics compared to the baseline. Notably, the random sampling method yields more significant enhancements. We hypothesize that this" + } + ] + } + ], + "index": 56 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8727" + } + ] + } + ], + "index": 57 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 69, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 553, + 148 + ], + "type": "table", + "html": "
MethodBroomChickenCut LemonTorchcoPeel BananaHandMean
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
DeformGS [66]20.740.32226.320.78627.940.71427.40.87726.380.83627.790.7826.090.719
+TimeFormer21.010.32426.550.79230.010.78027.550.88327.240.85229.790.84827.030.747
4DGS[60]21.530.35126.820.79729.720.76327.450.88327.820.84429.520.84127.140.747
+TimeFormer22.840.40127.060.80130.230.77829.480.90228.210.85330.380.86228.030.766
", + "image_path": "e029d70d85ad9696aa983d1634b971cdff71d3e90ee3675996837f6f4d215dab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 175, + 558, + 254 + ], + "blocks": [ + { + "bbox": [ + 187, + 154, + 423, + 166 + ], + "lines": [ + { + "bbox": [ + 187, + 154, + 423, + 166 + ], + "spans": [ + { + "bbox": [ + 187, + 154, + 423, + 166 + ], + "type": "text", + "content": "Table 3. Quantitative Comparisons on HyperNeRF Dataset [39]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 175, + 558, + 254 + ], + "lines": [ + { + "bbox": [ + 56, + 175, + 558, + 254 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 558, + 254 + ], + "type": "table", + "html": "
MethodPressPlateBasinSieveBellCupMean
PSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIMPSNRSSIM
HyperNeRF [39]25.40.87318.10.71420.20.82925.00.90924.00.88424.10.89622.80.851
NeRF-DS [64]26.40.91120.80.86720.30.86826.10.93523.30.87224.50.91623.570.895
DeformGS[66]25.680.86620.820.81219.870.80425.710.88124.920.85424.520.89723.590.852
+TimeFormer26.290.86720.900.81719.930.80526.260.89125.900.87325.160.90324.100.859
", + "image_path": "5242b14c867ca8f47a5437c2d54c643622f8fdabc89397e634a051f591122b08.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 291, + 294, + 419 + ], + "blocks": [ + { + "bbox": [ + 60, + 291, + 294, + 419 + ], + "lines": [ + { + "bbox": [ + 60, + 291, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 60, + 291, + 294, + 419 + ], + "type": "image", + "image_path": "fab3f703efa079a01918d7752d39f32e42468806559975932e1922028cced503.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 419, + 206, + 427 + ], + "lines": [ + { + "bbox": [ + 161, + 419, + 206, + 427 + ], + "spans": [ + { + "bbox": [ + 161, + 419, + 206, + 427 + ], + "type": "text", + "content": "(a) Cut Lemon" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 60, + 429, + 294, + 588 + ], + "blocks": [ + { + "bbox": [ + 60, + 429, + 294, + 588 + ], + "lines": [ + { + "bbox": [ + 60, + 429, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 60, + 429, + 294, + 588 + ], + "type": "image", + "image_path": "134b17a7009fe9a1932a98d991747c9bdeb08dc63b2b48883c63487cbb735450.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 98, + 590, + 149, + 598 + ], + "lines": [ + { + "bbox": [ + 98, + 590, + 149, + 598 + ], + "spans": [ + { + "bbox": [ + 98, + 590, + 149, + 598 + ], + "type": "text", + "content": "(b) Torchocolate" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 590, + 255, + 598 + ], + "lines": [ + { + "bbox": [ + 223, + 590, + 255, + 598 + ], + "spans": [ + { + "bbox": [ + 223, + 590, + 255, + 598 + ], + "type": "text", + "content": "(c) Broom" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 611, + 288, + 624 + ], + "lines": [ + { + "bbox": [ + 63, + 611, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 63, + 611, + 288, + 624 + ], + "type": "text", + "content": "Figure 11. Motion Visualization on HyperNeRF Dataset [39]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 638, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 638, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 295, + 674 + ], + "type": "text", + "content": "phenomenon occurs because the random sampling method allows for more effective propagation of dynamic behavior modeling across the entire time series." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Shared Deformation Fields. In Tab. 5b, the image quality during inference suffers a considerable decrease if we train two deformation fields without sharing their weights." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 320, + 291, + 560, + 354 + ], + "blocks": [ + { + "bbox": [ + 192, + 261, + 418, + 272 + ], + "lines": [ + { + "bbox": [ + 192, + 261, + 418, + 272 + ], + "spans": [ + { + "bbox": [ + 192, + 261, + 418, + 272 + ], + "type": "text", + "content": "Table 4. Quantitative Comparisons on NeRF-DS Dataset [64]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 291, + 560, + 354 + ], + "lines": [ + { + "bbox": [ + 320, + 291, + 560, + 354 + ], + "spans": [ + { + "bbox": [ + 320, + 291, + 560, + 354 + ], + "type": "table", + "html": "
SettingRandom SamplingContinuous Sampling
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
B=225.890.8750.11725.730.8710.124
B=426.060.8710.11925.780.8740.117
B=625.990.870.12325.730.8650.125
", + "image_path": "393688fcb3f0844c9a5e3ad9d5a62eb11463a667b468199b3f9d375b53d03c6f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 347, + 371, + 533, + 458 + ], + "blocks": [ + { + "bbox": [ + 343, + 355, + 537, + 365 + ], + "lines": [ + { + "bbox": [ + 343, + 355, + 537, + 365 + ], + "spans": [ + { + "bbox": [ + 343, + 355, + 537, + 365 + ], + "type": "text", + "content": "(a) Ablation Studies on Batch Size and Sampling Strategies." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 371, + 533, + 458 + ], + "lines": [ + { + "bbox": [ + 347, + 371, + 533, + 458 + ], + "spans": [ + { + "bbox": [ + 347, + 371, + 533, + 458 + ], + "type": "table", + "html": "
SettingPSNR↑SSIM↑LPIPS↓
Baseline25.440.8670.125
M=125.960.8740.117
M=225.900.8740.120
M=326.050.8760.112
M=426.040.8730.119
w/o Shared24.810.8520.132
", + "image_path": "7bf28136f7fb00cf41557f1f2d960b86d5b89703b91a734511a8ce80c9ae1764.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 314, + 460, + 565, + 480 + ], + "lines": [ + { + "bbox": [ + 314, + 460, + 565, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 460, + 565, + 480 + ], + "type": "text", + "content": "(b) Ablation Results. " + }, + { + "bbox": [ + 314, + 460, + 565, + 480 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 314, + 460, + 565, + 480 + ], + "type": "text", + "content": " is the number of encoder layers, \"w/o Shared\" means not using shared weights in two deformation fields." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 313, + 489, + 555, + 511 + ], + "lines": [ + { + "bbox": [ + 313, + 489, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 555, + 511 + ], + "type": "text", + "content": "Table 5. Ablation Results. The results are from three scenes press, sieve and bell on NeRF-DS Dataset [64]." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 546 + ], + "type": "text", + "content": "Fig. 8 also demonstrates the effectiveness of the two-stream strategy and TimeFormer's ability of re-balancing." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 560, + 388, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 388, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 388, + 572 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "We propose TimeFormer, a Transformer module which enhances existing deformable 3D Gaussians methods by implicitly modeling motion patterns from a learning perspective. In addition, we design a two-stream optimization strategy to transfer the learned motion knowledge from TimeFormer to the original deformation branch. This allows us to remove TimeFormer during inference and thus maintain the same inference speed. Extensive experiments demonstrate that TimeFormer facilitates the reconstruction of three state-of-the-art methods among three datasets, and illustrate the improvement on reconstructing complex scenes." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8728" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 207, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 207, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 207, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 61, + 207, + 296, + 274 + ], + "type": "text", + "content": "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 274, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 319 + ], + "type": "text", + "content": "[2] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. arXiv preprint arXiv:2404.03613, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 319, + 296, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 385 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5855-5864, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "text", + "content": "[4] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 130-141, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "type": "text", + "content": "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European conference on computer vision, pages 333-350. Springer, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 477, + 296, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 296, + 543 + ], + "type": "text", + "content": "[6] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 544, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 578 + ], + "type": "text", + "content": "[7] Robert A Drebin, Loren Carpenter, and Pat Hanrahan. Volume rendering. ACM Siggraph Computer Graphics, 22(4): 65-74, 1988. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 579, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 579, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 579, + 296, + 634 + ], + "type": "text", + "content": "[8] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "type": "text", + "content": "[9] Yuanxing Duan, Fangyin Wei, Qiyu Dai, Yuhang He, Wenzheng Chen, and Baoquan Chen. 4d-rotor gaussian splatting: Towards efficient novel view synthesis for dynamic scenes. In Proc. SIGGRAPH, 2024. 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "text", + "content": "[10] Bardienus P Duisterhof, Zhao Mandi, Yunchao Yao, Jia-Wei Liu, Mike Zheng Shou, Shuran Song, and Jeffrey Ichnowski. Md-splatting: Learning metric deformation from" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work is completed during the author's internship at the Tencent Hunyuan 3D Department. This work is supported in part by the National Science Fund for Distinguished Young Scholars (No. 62325208), in part by the National Natural Science Foundation of China (No. 62232002), and in part by the Natural Science Foundation of Tianjin (No. 23ZGZNGX00020 and 24JCZXJC00050)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "4d gaussians in highly deformable scenes. arXiv preprint arXiv:2312.00583, 2023. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[11] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 207 + ], + "type": "text", + "content": "[12] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 554, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 554, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 554, + 263 + ], + "type": "text", + "content": "[13] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbek Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12479–12488, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "text", + "content": "[14] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "type": "text", + "content": "[15] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "text", + "content": "[16] Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wenchao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splitting gaussian dynamics for 4d content creation. arXiv preprint arXiv:2403.12365, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "text", + "content": "[17] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. arXiv preprint arXiv:2403.11447, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 444, + 554, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 554, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 554, + 477 + ], + "type": "text", + "content": "[18] Zhi Hou, Baosheng Yu, and Dacheng Tao. Batchformer: Learning to explore sample relationships for robust representation learning. In CVPR, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 554, + 533 + ], + "type": "text", + "content": "[19] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4220-4230, 2024. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 534, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 554, + 567 + ], + "type": "text", + "content": "[20] Takeo Kanade, Peter Rander, and PJ Narayanan. Virtualized reality: Constructing virtual worlds from real scenes. IEEE multimedia, 4(1):34-47, 1997. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 554, + 612 + ], + "type": "text", + "content": "[21] Kai Katsumata, Duc Minh Vo, and Hideki Nakayama. A compact dynamic 3d gaussian representation for real-time dynamic view synthesis. In Computer Vision - ECCV 2024, pages 394-412. Springer Nature Switzerland, 2025. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 657 + ], + "type": "text", + "content": "[22] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 714 + ], + "type": "text", + "content": "[23] Deqi Li, Shi-Sheng Huang, Zhiyuan Lu, Xinran Duan, and Hua Huang. St-4dgs: Spatial-temporally consistent 4d gaussian splatting for efficient dynamic scene rendering. In ACM SIGGRAPH 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8729" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[24] Lingzhi Li, Zhen Shen, Zhongshu Wang, Li Shen, and Ping Tan. Streaming radiance fields for 3d video synthesis. Advances in Neural Information Processing Systems, 35: 13485-13498, 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 193 + ], + "type": "text", + "content": "[25] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, et al. Neural 3d video synthesis from multi-view video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5521-5531, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 248 + ], + "type": "text", + "content": "[26] Zhengqi Li, Simon Niklaus, Noah Snively, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "text", + "content": "[27] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8508-8520, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 358 + ], + "type": "text", + "content": "[28] Youtian Lin, Zuozhuo Dai, Siyu Zhu, and Yao Yao. Gaussian-flow: 4d reconstruction with dynamic 3d gaussian particle. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21136-21145, 2024. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 360, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 414 + ], + "type": "text", + "content": "[29] Jia-Wei Liu, Yan-Pei Cao, Weijia Mao, Wenqiao Zhang, David Junhao Zhang, Jussi Keppo, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Devrf: Fast deformable voxel radiance fields for dynamic scenes. Advances in Neural Information Processing Systems, 35:36762-36775, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 415, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 480 + ], + "type": "text", + "content": "[30] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13-23, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 482, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 546 + ], + "type": "text", + "content": "[31] Ange Lou, Benjamin Planche, Zhongpai Gao, Yamin Li, Tianyu Luan, Hao Ding, Terrence Chen, Jack Noble, and Ziyan Wu. Darenerf: Direction-aware representation for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5031-5042, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[32] Jiahao Lu, Jiacheng Deng, Ruijie Zhu, Yanzhe Liang, Wenfei Yang, Tianzhu Zhang, and Xu Zhou. Dn-4dgs: Denoised deformable network with temporal-spatial aggregation for dynamic scene rendering. arXiv preprint arXiv:2410.13607, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "text", + "content": "[33] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[34] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 2024 International Conference on 3D Vision (3DV), pages 800-809. IEEE, 2024. 2, 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[35] Xingyu Miao, Yang Bai, Haoran Duan, Fan Wan, Yawen Huang, Yang Long, and Yefeng Zheng. Ctnerf: Cross-time transformer for dynamic neural radiance field from monocular video. Pattern Recognition, 156:110729, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 174, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 218 + ], + "type": "text", + "content": "[37] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "text", + "content": "[38] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 276, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 340 + ], + "type": "text", + "content": "[39] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), 2021. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 342, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 398 + ], + "type": "text", + "content": "[40] Sungheon Park, Minjung Son, Seokhwan Jang, Young Chun Ahn, Ji-Yeon Kim, and Nahiup Kang. Temporal interpolation is all you need for dynamic neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4212-4221, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[41] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 455, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 520 + ], + "type": "text", + "content": "[42] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boring Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16632-16642, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 522, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 577 + ], + "type": "text", + "content": "[43] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Pérez-Pellitero. Swings: sliding windows for dynamic 3d gaussian splatting. In European Conference on Computer Vision. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 578, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 623 + ], + "type": "text", + "content": "[44] Nagabhushan Somraj, Kapil Choudhary, Sai Harsha Mupparaju, and Rajiv Soundararajan. Factorized motion fields for fast sparse input dynamic view synthesis. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "text", + "content": "[45] Liangchen Song, Xuan Gong, Benjamin Planche, Meng Zheng, David Doermann, Junsong Yuan, Terrence Chen, and Ziyan Wu. Pref: Predictability regularized neural motion fields. In European Conference on Computer Vision, pages 664-681. Springer, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[46] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with de" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8730" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "composed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 29(5):2732-2742, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "text", + "content": "[47] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 295, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 295, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 295, + 226 + ], + "type": "text", + "content": "[48] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20675-20685, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 281 + ], + "type": "text", + "content": "[49] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 402-419. Springer, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 295, + 337 + ], + "type": "text", + "content": "[50] Fengrui Tian, Shaoyi Du, and Yueqi Duan. Mononerf: Learning a generalizable dynamic radiance field from monocular videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17903-17913, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 337, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 295, + 403 + ], + "type": "text", + "content": "[51] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 426 + ], + "type": "text", + "content": "[52] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "type": "text", + "content": "[53] Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 524 + ], + "type": "text", + "content": "[54] Chaoyang Wang, Lachlan Ewen MacDonald, Laszlo A Jeni, and Simon Lucey. Flow supervision for deformable nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21128-21137, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 295, + 581 + ], + "type": "text", + "content": "[55] Feng Wang, Sinan Tan, Xinghang Li, Zeyue Tian, Yafei Song, and Huaping Liu. Mixed neural voxels for fast multiview video synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19706-19716, 2023. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "type": "text", + "content": "[56] Feng Wang, Zilong Chen, Guokang Wang, Yafei Song, and Huaping Liu. Masked space-time hash encoding for efficient dynamic scene reconstruction. Advances in Neural Information Processing Systems, 36, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 680 + ], + "type": "text", + "content": "[57] Liao Wang, Qiang Hu, Qihan He, Ziyu Wang, Jingyi Yu, Tinne Tuytelaars, Lan Xu, and Minye Wu. Neural residual radiance fields for streamably free-viewpoint videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 76-87, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[58] Shizun Wang, Xingyi Yang, Qiuhong Shen, Zhenxiang Jiang, and Xinchao Wang. Gflow: Recovering 4d world from monocular video. arXiv preprint arXiv:2405.18426, 2024. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[59] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "type": "text", + "content": "[60] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20310-20320, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 183, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 553, + 237 + ], + "type": "text", + "content": "[61] Minye Wu, Zehao Wang, Georgios Kouros, and Tinne Tuytelaars. Tetrirf: Temporal tri-plane radiance fields for efficient free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6487-6496, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 238, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 238, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 553, + 293 + ], + "type": "text", + "content": "[62] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9421-9431, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 293, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 348 + ], + "type": "text", + "content": "[63] Zhen Xu, Sida Peng, Haotong Lin, Guangzhao He, Jiaming Sun, Yujun Shen, Hujun Bao, and Xiaowei Zhou. 4k4d: Real-time 4d view synthesis at 4k resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20029-20040, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 349, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 555, + 392 + ], + "type": "text", + "content": "[64] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8285-8295, 2023. 5, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 426 + ], + "type": "text", + "content": "[65] Jason C Yang, Matthew Everett, Chris Buehler, and Leonard McMillan. A real-time distributed light field camera. Rendering Techniques, 2002(77-86):2, 2002. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "text", + "content": "[66] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[67] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In International Conference on Learning Representations (ICLR), 2024. 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "type": "text", + "content": "[68] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoptrees for real-time rendering of neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5752-5761, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 613 + ], + "type": "text", + "content": "[69] Yifan Zhan, Zhuoxiao Li, Muyao Niu, Zhihang Zhong, Shohei Nobuhara, Ko Nishino, and Yinqiang Zheng. Kfdnerf: Rethinking dynamic nerf with kalman filter. arXiv preprint arXiv:2407.13185, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 669 + ], + "type": "text", + "content": "[70] Boming Zhao, Yuan Li, Ziyu Sun, Lin Zeng, Yujun Shen, Rui Ma, Yinda Zhang, Hujun Bao, and Zhaopeng Cui. Gaussian prediction: Dynamic 3d gaussian prediction for motion extrapolation and free view synthesis. In ACM SIGGRAPH 2024 Conference Papers, pages 1-12, 2024. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[71] Kaichen Zhou, Jia-Xing Zhong, Sangyun Shin, Kai Lu, Yiyuan Yang, Andrew Markham, and Niki Trigoni. Dynpoint: Dynamic neural point for view synthesis. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "8731" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 173 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[72] Ruijie Zhu, Yanzhe Liang, Hanzhi Chang, Jiacheng Deng, Jiahao Lu, Wenfei Yang, Tianzhu Zhang, and Yongdong Zhang. Motions: Exploring explicit motion guidance for deformable 3d gaussian splatting. arXiv preprint arXiv:2410.07707, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[73] C Lawrence Zitnick, Sing Bing Kang, Matthew Uytendaele, Simon Winder, and Richard Szeliski. High-quality video view interpolation using a layered representation. ACM transactions on graphics (TOG), 23(3):600-608, 2004. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8732" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/5b02e9a3-c733-4ab2-9aa6-549d5b3e9de2_content_list.json b/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/5b02e9a3-c733-4ab2-9aa6-549d5b3e9de2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e92a59e9beebaa990d15b886f6d051de21f575e3 --- /dev/null +++ b/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/5b02e9a3-c733-4ab2-9aa6-549d5b3e9de2_content_list.json @@ -0,0 +1,2036 @@ +[ + { + "type": "text", + "text": "Timestep-Aware Diffusion Model for Extreme Image Rescaling", + "text_level": 1, + "bbox": [ + 181, + 130, + 816, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ce Wang, Zhenyu Hu, Wanjie Sun* Zhenzhong Chen Wuhan University", + "bbox": [ + 285, + 180, + 705, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cewang, zhenyuhu, sunwanjie, zzchen}@whu.edu.cn", + "bbox": [ + 282, + 218, + 705, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image rescaling aims to learn the optimal low-resolution (LR) image that can be accurately reconstructed to its original high-resolution (HR) counterpart, providing an efficient image processing and storage method for ultra-high definition media. However, extreme downscaling factors pose significant challenges to the upscaling process due to its highly ill-posed nature, causing existing image rescaling methods to struggle in generating semantically correct structures and perceptual friendly textures. In this work, we propose a novel framework called Timestep-Aware Diffusion Model (TADM) for extreme image rescaling, which performs rescaling operations in the latent space of a pretrained autoencoder and effectively leverages powerful natural image priors learned by a pre-trained text-to-image diffusion model. Specifically, TADM adopts a pseudo-invertible module to establish the bidirectional mapping between the latent features of the HR image and the target-sized LR image. Then, the rescaled latent features are enhanced by a pre-trained diffusion model to generate more faithful details. Considering the spatially non-uniform degradation caused by the rescaling operation, we propose a novel time-step alignment strategy, which can adaptively allocate the generative capacity of the diffusion model based on the quality of the reconstructed latent features. Extensive experiments demonstrate the superiority of TADM over previous methods in both quantitative and qualitative evaluations. The code will be available at: https://github.com/wwangcece/TADM.", + "bbox": [ + 91, + 301, + 480, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 756, + 220, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the explosive growth of ultra high-resolution (HR) images, image rescaling has become essential for enhancing data storage and transmission efficiency [39], optimizing resource usage in real-time applications [22], and ensuring proper display on devices with varying resolutions [21]. During image rescaling, HR images are downscaled", + "bbox": [ + 89, + 782, + 480, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to low-resolution (LR) images for storage, transmission, or display. When HR images are needed, an upscaling process is employed to restore the downscaled images to their original resolution. Unlike image super-resolution (SR), image rescaling allows for joint optimization of both downscaling and upscaling processes, preserving detailed information crucial for effective upscaling.", + "bbox": [ + 511, + 268, + 903, + 377 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently proposed image rescaling methods can be classified into two main categories: encoder-decoder architecture [12, 25] and invertible neural networks [17, 37]. The encoder-decoder architecture employs two neural networks to perform downscaling and upscaling. While, invertible neural networks view image rescaling as a reversible process, using the forward and backward passes of an invertible neural network to simulate the downscaling and upscaling. Despite the significant improvements in reconstruction quality achieved by these methods, they are limited to small rescaling factors, such as $2 \\times$ or $4 \\times$ .", + "bbox": [ + 511, + 378, + 903, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, with the advancement of ultra HR imaging devices, there is an increasing need for extreme rescaling factors (e.g., $16 \\times$ or $32 \\times$ ) to achieve more efficient storage and transmission of ultra HR data [7, 47]. Existing image rescaling methods often struggle to generate sufficient detail under extreme rescaling factors. Therefore, some approaches propose using pre-trained GANs as priors to constrain the solution space of the reconstructed results. GRAIN [47] and BDFlow [14] use pre-trained StyleGAN [9, 10] to generate high-quality upscaled face images. VQIR [34] leverages the high-quality visual embeddings encapsulated in pre-trained VQGAN [3] to achieve extreme rescaling on natural images. Although these methods have made significant progress in perceptual metrics, their reconstruction results still suffer from incorrect semantics and structure.", + "bbox": [ + 511, + 549, + 903, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address the aforementioned issues, we propose a novel framework named Timestep-Aware Diffusion Model (TADM) for extreme image rescaling, which performs image rescaling in the latent space of a pre-trained autoencoder and leverages the natural image priors stored in large text-to-image diffusion models to restore HR images. Specifically, first, the HR images are encoded by the via the VAE encoder to obtain the latent features. Then, we use an au", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15594", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "toencoder to rescale the latent code of the HR image to the target size, and employ a set of invertible neural networks (INN) to perform the mapping from the feature space to the pixel space to obtain the LR image. Next, the rescaled features are fed into a pre-trained diffusion model to perform a single denoising step, thereby enhancing their perceptual quality. Finally, the enhanced features are decoded via the VAE decoder to obtain the rescaled image. Additionally, we find that the rescaled latent features exhibit non-uniform degradation, which can be interpreted as the addition of noise at varying levels. As a result, different images require different time-steps to perform the denoising process. Previous studies have shown that the time-step in diffusion models directly controls the generative capacity of the model [29]. Therefore, we propose a time-step alignment mechanism that predicts the appropriate timestep based on the reconstruction quality of the rescaled latent features, thereby adaptively tuning the generative capacity of the diffusion model.", + "bbox": [ + 89, + 90, + 480, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 378, + 434, + 393 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Timestep-Aware Diffusion Model that performs well in global semantic and structural reconstruction for the extreme image rescaling task.", + "- We propose a novel decoupled strategy for feature rescaling. It decouples the generation of LR images from the rescaling operation, thereby improving the reconstruction accuracy of the rescaling process.", + "- We design a time-step alignment strategy to address the spatially non-uniform reconstruction quality of the rescaled latent features, allowing for more effective utilization of the prior information from the text-to-image diffusion model." + ], + "bbox": [ + 91, + 396, + 482, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 594, + 232, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Rescaling", + "text_level": 1, + "bbox": [ + 89, + 619, + 253, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image rescaling aims to downscale HR images to visually coherent LR images and then plausibly reconstruct the original HR images. Unlike image SR, image rescaling involves joint modeling of both downscaling and upscaling processes, leading to higher reconstruction accuracy. Earlier researches [12, 16] utilize CNNs to build an encoder-decoder framework, treating image rescaling as a unified task by jointly training the downscaling and upscaling processes. Instead of directly generating LR images, CAR [25] proposes generating content-adaptive resampling kernels based on HR images to perform the downsampling operation, avoiding pixel-level constraints on the downscaled images. To fully leverage the inherent reversibility of image rescaling, there has been a surge in the use of INNs for this task [15, 38, 45, 48]. IRN [37] is the first invertible framework that models image rescaling as a bijective transformation, embedding residual high-frequency (HF) components", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "into a case-agnostic latent distribution for efficient reconstruction. HCFlow [17] assumes that the HF information depends on the LR image, thereby achieving better performance by incorporating LR conditions in HF information estimation. In extreme image rescaling, generative priors are usually employed to constrain the outcomes. GRAIN [47] and VQIR [34] leverage pre-trained StyleGAN [9, 10] and VQGAN [3], respectively, to achieve notable enhancements in perceptual quality. Nonetheless, these approaches are often limited in image contents, either confined to handling images from specific domains or encountering difficulties when reconstructing faces and textual elements.", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Diffusion Model for Image SR", + "text_level": 1, + "bbox": [ + 511, + 284, + 782, + 300 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Here, we briefly introduce the application of diffusion models in a related task, i.e., image SR. Diffusion models promote the development of natural image SR in two main ways: training diffusion models from scratch [4, 13, 41] and leveraging pre-trained SD model [19, 36, 40]. The former is represented by SR3 [24], which achieves perceptual quality comparable to GANs. The latter is represented by StableSR [28], which aims to build a model capable of handling image restoration under any complex degradation in real-world scenarios.", + "bbox": [ + 511, + 306, + 903, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the above methods typically require dozens or even hundreds of iterations during inference, which significantly increases computational costs. Researchers have begun exploring how to achieve image SR using one-step diffusion models. SinSR [31], based on ResShift [41], derives a deterministic one-step sampling process and employs a consistency preserving loss to distill the prior knowledge from a multi-step teacher network into a one-step student network. A representative work, OSEDiff [35], directly feeds the LQ image into the SD model and employs variational score distillation for regularization. S3Diff [43] proposes a degradation-aware LoRA and utilizes a pre-trained DINO [44] model as a discriminator for adversarial training. InvSR [42] designs a partial noise prediction strategy to provide a dynamic starting sampling point and supports inference with an arbitrary number of sampling steps ranging from one to five.", + "bbox": [ + 511, + 458, + 906, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 731, + 648, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Overview of TADM", + "text_level": 1, + "bbox": [ + 511, + 756, + 700, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Fig. 1, TADM consists of four parts: Latent Encoding, Feature Rescaling in the Latent Space, Denoising Guided Perception Enhancement and Latent Decoding. First, the input HR images $x$ are mapped to the latent features $z$ via the pre-trained VAE encoder $\\mathcal{E}$ . Subsequently, the proposed Decoupled Feature Rescaling Module (DFRM) rescales $z$ to the target size, outputting LR images $y$ and the rescaled latent features $\\hat{z}$ . Our aim is to make $\\hat{z}$", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15595", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/20f560aec677f0e2202ba95b14927a29a9381aa962307d0d48974ecb76432715.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dffab37f138dca29e215e6a3d9eb66e5ee2a8f067984774baa57e3286b203c6a.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed Timestep-Aware Diffusion Model (TADM). First, the HR images $x$ are encoded to obtain the latent features $z$ . The latent features are then rescaled to the target size using the decoupled feature rescaling module (DFRM), producing the LR images $y$ and outputting the rescaled latent features $\\hat{z}$ . Next, the rescaled latent features $\\hat{z}$ are passed into a pre-trained diffusion model to perform a single denoising step for perceptual enhancement to obtain $\\hat{z}_0$ . In this process, a time prediction module (TPM) is used to estimate the time step $t$ based on $\\hat{z}$ and the predicted $t$ is then fed into both the U-Net and the time scheduler. Finally, the perceptually enhanced latent features $\\hat{z}_0$ are decoded to obtain the rescaled image $\\hat{x}$ ." + ], + "image_footnote": [], + "bbox": [ + 133, + 271, + 867, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as close as possible to $z$ so that the rescaled results align with the latent space of the pre-trained VAE, while ensuring that $y$ and the original HR images $x$ share the same content. Then, the rescaled features $\\hat{z}$ are fed into a denoising U-Net network to perform a single denoising step to get the perceptually enhanced latent features $\\hat{z}_0$ . Finally, $\\hat{z}_0$ is decoded via the pre-trained VAE decoder $\\mathcal{D}$ to obtain the rescaled image $\\hat{x}$ . Overall, the latent rescaling process can be regarded as a noise-adding process, while the latent enhancement process can be viewed as a denoising process. To dynamically estimate the 'noise' density introduced by latent rescaling, we design a time alignment strategy, which consists of a time-step prediction module (TPM) and a time scheduler module. It can better address the non-uniform degradation caused by rescaling, thereby enabling the dynamic allocation of generative capabilities of pre-trained SD.", + "bbox": [ + 88, + 522, + 485, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Feature Rescaling in the Latent Space", + "text_level": 1, + "bbox": [ + 89, + 794, + 416, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Network Architecture.", + "text_level": 1, + "bbox": [ + 89, + 819, + 294, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this part, we perform rescaling of the HR latent features $z$ , obtaining LR images $y$ and rescaled features $\\hat{z}$ . A straightforward idea is to construct a transformation chain as $y = G_{\\mathrm{e}}(z)$ , $\\hat{z} = G_{\\mathrm{d}}(y)$ . Here, $G_{\\mathrm{e}}$ is the encoder used", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to downscale the input features to the target sized LR images, while $G_{\\mathrm{d}}$ represents the decoder used to reconstruct the original HR features. Guidance loss $\\mathcal{L}_{\\mathrm{gui}}$ and reconstruction loss $\\mathcal{L}_{\\mathrm{rec}}$ are applied to $y$ and $\\hat{z}$ respectively, to force $y$ to have the same content as $x$ , and $\\hat{z}$ to be as similar as possible to $z$ :", + "bbox": [ + 511, + 522, + 906, + 613 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {g u i .}} = \\| y - \\operatorname {B i c u b i c} (x) \\| _ {1} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 626, + 903, + 642 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c .}} = \\| \\hat {z} - z \\| _ {1} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 645, + 903, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, these two losses are contradictory to each other. When the guidance loss approaches zero, the rescaling problem degenerates into a classical SR problem [33], embedding little information about the downscaling process into the LR image. In contrast, when the reconstruction loss reaches its global optimum, the rescaling problem degenerates into an image compression problem [20], causing the LR image to lack meaningful visual content. Therefore, directly applying the guidance loss $\\mathcal{L}_{\\mathrm{gui}}$ on $y$ could adversely impact subsequent upscaling.", + "bbox": [ + 511, + 672, + 906, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering the above issues, we devise the DFRM as shown in Fig. 2, which divides the rescaling operation into two sub-tasks: feature rescaling and LR image generation. Inspired by FGRN [15], we use a set of CNNs to perform downscaling and upscaling, and a separate INN to", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15596", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1258a43e802c5300b8cc4c98aa94c29ee00708d7a8662f578ecd125fad9967a0.jpg", + "image_caption": [ + "Figure 2. Detailed architecture of Decoupled Feature Rescaling Module (DFRM)." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 473, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "perform the bidirectional mapping between the feature domain and the pixel domain. To achieve a balance between feature rescaling accuracy and visual quality of LR images, we introduce another pixel guidance module in the encoder to incorporate pixel-level information, which can improve the quality of the LR image. Specifically, DFRM perform feature rescaling through two independent transformation chains $((x,z)\\to z_{\\mathrm{lr}}\\to \\hat{z})$ . $(z_{\\mathrm{lr}}\\leftrightarrow y)$ as follows:", + "bbox": [ + 89, + 247, + 483, + 368 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\mathrm {l r}} = G _ {\\mathrm {e}} (x, z) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 376, + 480, + 392 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\mathrm {l r}} \\approx F ^ {- 1} (F (z _ {\\mathrm {l r}})) = F ^ {- 1} (y) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 395, + 480, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} = G _ {\\mathrm {d}} \\left(z _ {\\mathrm {l r}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 415, + 480, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $F$ and $F^{-1}$ represent the forward and backward processes of the INN, respectively. Given the latent variable $z \\in \\mathbb{R}^{\\frac{H}{8} \\times \\frac{W}{8} \\times 4}$ of the pre-trained SD VAE as input, the encoder $G_{\\mathrm{e}}$ first transforms it into a compact representation $z_{\\mathrm{lr}} \\in \\mathbb{R}^{\\frac{H}{N} \\times \\frac{W}{N} \\times 3}$ , whose size is the same as that of the target LR image and $N$ represents the target downscaling factor. Subsequently, a set of INNs perform the forward process to transform $z_{\\mathrm{lr}}$ from the feature domain to the pixel domain and obtain the LR image $y \\in \\mathbb{R}^{\\frac{H}{N} \\times \\frac{W}{N} \\times 3}$ by quantization. In the upscaling process, the inverse of the INN is used to process the LR image $y$ and reconstruct $z_{\\mathrm{lr}}$ . Finally, the decoder $G_{\\mathrm{d}}$ is utilized to obtain the rescaled feature $\\hat{z}$ .", + "bbox": [ + 89, + 439, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Loss Function.", + "text_level": 1, + "bbox": [ + 89, + 626, + 238, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training of this part involves two losses: reconstruction loss and guidance loss. Considering the information lost in quantization operations, the bidirectional mapping between $z_{\\mathrm{lr}}$ and $y$ is not exactly reversible. To ensure the robustness of the network to quantization operations, we consider two types of transformation chains in designing the reconstruction loss: $(x,z)\\to z_{\\mathrm{lr}}\\to \\hat{z}$ and $(x,z)\\to z_{\\mathrm{lr}}\\to y\\to \\hat{z}_{\\mathrm{lr}}\\to \\hat{z}$ , which can be expressed as the following formula:", + "bbox": [ + 89, + 645, + 483, + 765 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e c .}} = \\| G _ {\\mathrm {d}} \\left(G _ {\\mathrm {e}} (x, z)\\right) - z \\| _ {1} + \\tag {6} \\\\ \\left. \\left\\| G _ {\\mathrm {d}} \\left(F ^ {- 1} \\left(F \\left(G _ {\\mathrm {e}} (x, z)\\right)\\right)\\right) - z \\right\\| _ {1} \\right. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 772, + 482, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This combination can facilitate the network to reconstruct $\\hat{z}$ with high fidelity in both the existence and absence of quantization operations. Additionally, the guidance loss is designed as follows:", + "bbox": [ + 89, + 816, + 483, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {g u i .}} = \\left\\| F \\left(G _ {\\mathrm {e}} (x, z)\\right) - \\operatorname {B i c u b i c} (x) \\right\\| _ {1} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the loss function of this part is a combination of these two losses:", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e s .}} = \\lambda_ {\\text {r e c .}} \\mathcal {L} _ {\\text {r e c .}} + \\lambda_ {\\text {g u i .}} \\mathcal {L} _ {\\text {g u i .}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 130, + 903, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Denoising Guided Perception Enhancement", + "text_level": 1, + "bbox": [ + 511, + 154, + 883, + 169 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The latent variable $\\hat{z}$ obtained in Sec. 3.2 retains sufficient semantic information about the HR image, but lacks adequate texture. Therefore, we propose to leverage the strong priors stored in the pre-trained SD model to enhance the perceptual quality of the rescaling results in this part. Original diffusion model requires a multi-step sampling algorithm, significantly increases inference time and makes it difficult to scale to practical applications. Inspired by recent work on accelerating diffusion models [8], we adopt an one-step inference strategy. Specifically, we can directly feed the rescaled latent variable $\\hat{z}$ into the denoising U-Net $\\epsilon = \\epsilon_{\\theta}(\\hat{z},t)$ and perform a single denoising step as follows:", + "bbox": [ + 511, + 176, + 906, + 357 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {0} = \\frac {\\hat {z} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon}{\\sqrt {\\bar {\\alpha} _ {t}}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 377, + 903, + 411 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{z}_0$ represents the perceptually enhanced latent features, $\\bar{\\alpha}_{t}$ is the noise scheduler of the forward diffusion process of the SD model, and $t$ represents the time-step associated with the noise intensity. Finally, $\\hat{z}_0$ can be transformed to the pixel domain through the pre-trained VAE decoder $\\mathcal{D}$ to obtain the upscaled image $\\hat{x} = \\mathcal{D}(\\hat{z}_0)$ . Then, we can adopt the loss function between $\\hat{x}$ and $x$ to optimize the U-Net using LoRA layers.", + "bbox": [ + 511, + 417, + 905, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the key problem of the aforementioned approaches lies in the selection of time-step $t$ . We can regard the rescaling process as the noise-adding process of diffusion model, while Eq. (9) represents the denoising process. For different rescaling factors and different images, the 'noise' introduced by the rescaling operation varies in intensity. In Fig. 3, we present the global rescaling error of an image under different rescaling factors, as well as the local rescaling error of different patches within the image. Additionally, we plot the error curve of the forward noise-adding process of the SD model. It can be observed that as the rescaling factor and image content change, rescaling MSE also varies and corresponds to the diffusion MSE at different time-steps. Thus, we propose to dynamically predict the time-step $t$ based on the quality of the rescaled latent features $\\hat{z}$ . A simple yet effective network named timestep prediction module (TPM), whose structure is presented in Fig. 1, is used to predict $t$ based on the rescaled latent features: $t = \\mathrm{TPM}(\\hat{z})$ . During the training phase, we compute a single time-step for each input image patch. In the inference phase, we adopt a tiled inference approach (which will be elaborated in detail in the supplementary materials), allowing for predicting different time-steps for various regions of a high-resolution input image.", + "bbox": [ + 511, + 539, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15597", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1b8b3468943aab65fa6153d8e81a81c03b422dbd2e378353f4c35c2984780218.jpg", + "image_caption": [ + "Figure 3. We compare the MSE introduced by the rescaling operation with that caused by the forward diffusion process. The rescaling degradation severity is correlated with both the image content and the rescaling scale. Therefore, the model requires adaptive time-step to align the rescaling MSE with the diffusion MSE." + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 460, + 289 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/151175251986af6e47c306ecd577e79aeb9bce6dfba03ffba3970274e61c1362.jpg", + "image_caption": [ + "Figure 4. The structure of the time scheduler module. It combines both the fixed scheduler and learnable scheduler to perform the denoising process. The two schedulers are connected by a zero-convolution layer to stabilize the training." + ], + "image_footnote": [], + "bbox": [ + 114, + 381, + 460, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, the predicted time-step $t$ is used to perform a single denoising step. However, the vanilla SD time scheduler as shown in Eq. (9) is not differentiable with respect to $t$ , which prevents gradients from being backpropagated to the TPM. Considering that given $\\hat{z}$ , $\\epsilon$ and $t$ , Eq. (9) essentially represents a linear transformation, which can be learned by a neural network. Therefore, we propose a hybrid time scheduler module consisting of a fixed scheduler and a learnable scheduler, whose structure is presented in Fig. 4. The former uses Eq. (9) to perform the denoising step, while the latter employs a neural network to simulate the denoising scheduler. To ensure training stability, the final layer of the learnable scheduler is a zero-initialized convolution, allowing it to gradually correct the results of the fixed scheduler as training progresses. This process can be described as follows:", + "bbox": [ + 89, + 575, + 483, + 815 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {0} = \\mathcal {S} _ {\\text {f i x e d}} (\\hat {z}, \\epsilon , t _ {0}) + \\mathcal {S} _ {\\text {l e a r n e d}} (\\hat {z}, \\epsilon , t) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 828, + 482, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t_0$ is the fixed pre-set time-step and $S_{\\mathrm{fixed}}$ denotes the vanilla SD time scheduler as shown in Eq. (9).", + "bbox": [ + 89, + 854, + 482, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we fine-tune the denoising U-Net network and", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the VAE decoder using LoRA, while also jointly training the TPM and time scheduler module. To achieve a balance between fidelity and realism, we supervise the reconstruction results $\\hat{x}$ using pixel loss and perceptual loss as follows:", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {e n h .}} = \\| \\hat {x} - x \\| _ {1} + \\tag {11} \\\\ \\lambda_ {\\text {p e c .}} \\left(\\mathcal {L} _ {\\text {l p i p s}} (\\hat {x}, x) + \\mathcal {L} _ {\\text {d i s t s}} (\\hat {x}, x)\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 162, + 905, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 209, + 645, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 233, + 702, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1. Datasets and Evaluation Metrics.", + "text_level": 1, + "bbox": [ + 511, + 256, + 790, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train the proposed TADM using the DF2K dataset [1, 18]. Considering that traditional benchmarks such as Set5, Set14, and BSD100 often have lower quality and resolution [5], they are not suitable for evaluating extreme rescaling methods. Therefore, we utilize four high-resolution datasets: DIV2K, CLIC2020 [26], Urban100 [6], and DIV8K [5] to evaluate models been compared (results on the latter two datasets are presented in the supplementary materials). We adopt PSNR and SSIM [32] for quantitative evaluation. For perceptual quality evaluation, reference-based metrics LPIPS [46] and DISTS [2] are used. For non-reference image quality assessment, we use metrics including MUSIQ [11] and CLIPIQA [27].", + "bbox": [ + 511, + 273, + 906, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2. Training Details", + "text_level": 1, + "bbox": [ + 511, + 478, + 671, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TADM is built based on the SD 2.1-base model, and the training is divided into three stages. First, we train the DFRM using Eq. (8). Then we adopt Eq. (11) to jointly train the LoRA layers, the TPM and the time scheduler module. Finally, we use a smaller learning rate to jointly fine-tune DFRM, LoRA layers, TPM and scheduler module using a combination of Eq. (8) and (11). More training details are provided in the supplementary material.", + "bbox": [ + 511, + 496, + 905, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Results and Comparison with SOTA", + "text_level": 1, + "bbox": [ + 511, + 626, + 826, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare the proposed TADM with three kinds of methods: (1) bicubic downscaling and SR, i.e., ESRGAN [30], StableSR [28] (based on multi-step diffusion) and S3Diff (based on single-step diffusion)[43]; (2) flow-based rescaling models, i.e., IRN [37] and HCFlow [17]; (3) prior-based extreme rescaling method VQIR [34]. Some of them are proposed for $4 \\times$ rescaling factors and we extended them to $16 \\times$ or $32 \\times$ . To ensure a fair comparison, all methods are trained until convergence on the DF2K dataset following their original configurations.", + "bbox": [ + 511, + 648, + 906, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 806, + 702, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 presents the $16\\times$ and $32\\times$ quantitative image rescaling results on four datasets. The performance of SR methods are generally poor. Both gan-based (ESRGAN) and diffusion-based (StableSR and S3Diff) methods can only achieve relative good performance on no-reference metrics.", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15598", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b49bae9e01fbfa476ab8eb510004c6378cfe0478dec1e4a942d880b2db589c55.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons with different methods at ${16} \\times$ and ${32} \\times$ (results on Urban100 and DIV8K are presented in the supplementary materials). The symbols $\\uparrow$ and $\\downarrow$ respectively represent that higher or lower values indicate better performance. Bold represents the best and underline represents the second best." + ], + "table_footnote": [], + "table_body": "
DatasetMethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓MUSIQ ↑CLIPIQA ↑
16×32×16×32×16×32×16×32×16×32×16×32×
DIV2KESRGAN23.1519.940.59460.56890.44780.58920.23780.423059.8043.350.61610.3734
StableSR22.0219.920.59830.55880.49800.57360.20780.260351.2848.940.37800.3203
S3Diff20.2217.810.54910.46010.40330.48950.13090.181064.3767.920.62280.6991
IRN26.0422.840.70200.62120.51530.60950.30410.403643.7228.700.27990.2336
HCFlow26.6623.890.71760.64670.48850.58160.28660.385246.4337.250.27350.2792
VQIR23.9122.020.64980.58230.31740.45680.10240.266364.0458.210.63500.6293
Ours23.9822.180.67360.58690.29790.42210.08860.168466.5669.120.71890.7204
CLIC2020ESRGAN25.4322.470.70180.68040.42130.54290.22090.401157.8041.280.61070.3558
StableSR24.1122.240.70520.67260.46250.52880.18600.228252.6949.800.37790.3325
S3Diff22.0419.870.66150.57110.39680.48290.13360.181263.0766.400.60010.6838
IRN28.5524.930.79070.72280.46760.55280.27460.377647.6627.810.29500.2495
HCFlow28.8226.210.79490.74510.45420.53250.26350.361147.7237.150.27330.2884
VQIR25.6924.120.73630.69420.28540.41910.09010.244564.3654.830.61050.6068
Ours26.0424.260.75420.69030.27240.39110.07920.151365.3569.100.68150.7271
", + "bbox": [ + 91, + 141, + 906, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This is because they rely solely on the LR image to reconstruct the original image and cannot jointly optimize the downscaling and upscaling processes. While methods based on INNs (HCFlow and IRN) achieve the highest PSNR and SSIM due to its consideration of the correlation between downscaling and upscaling. In all generative methods, VQIR achieves relatively good perceptual metrics by leveraging the priors of the pre-trained VQGAN. In contrast, our method achieves optimal performance across all perceptual metrics on all datasets. For instance, in the $32 \\times$ rescaling on the DIV2K validation set, our method improves the DISTS metric by $36.76\\%$ compared to the second-best VQIR. Importantly, our approach is designed for applications like social media delivery, where semantic fidelity and perceptual realism are more critical. In such cases, metrics like LPIPS and DISTS better reflect image quality [23].", + "bbox": [ + 88, + 393, + 485, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 667, + 269, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 5, we present visual comparisons at $16 \\times$ and $32 \\times$ rescaling factors. Due to the ill-posed nature of extreme rescaling, pixel-oriented optimization methods such as IRN, and HCFlow produce overly-smoothed results. Although ESRGAN incorporates perceptual optimization it still cannot generate adequate details. VQIR leverages the pretrained VQGAN as a prior, allowing it to generate rich details and sharp edges. However, it struggles with faces and more complex structures. Although S3Diff can generate sharp edges and rich details, its reconstructed results have low fidelity and exhibit significant artifacts and color distortions. Thanks to the powerful prior provided by the pre-trained SD and the carefully designed architecture,, our method effectively addresses these issues. It not only gener", + "bbox": [ + 89, + 688, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ates sufficiently rich textures but also reconstructs accuracy semantics and regular structures, such as text, faces, vehicles, and leaves.", + "bbox": [ + 511, + 393, + 906, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 448, + 666, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct ablation experiments to validate the effectiveness of each component. All ablations are trained on the DF2K dataset and tested on the validation set of the DIV2K for $16 \\times$ rescaling task.", + "bbox": [ + 511, + 470, + 906, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1. Effectiveness of Feature Rescaling", + "text_level": 1, + "bbox": [ + 511, + 537, + 797, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To verify the effectiveness of the feature rescaling operation, we move the rescaling operation from the latent space to the pixel space. Specifically, we use HCFlow [17], the comparative method with the best PSNR performance, to perform the rescaling operation in the pixel domain, followed by perceptual enhancement using the same approach described in Sec. 3.3. As shown the 1st row in Table 2, the rescaling operation in the latent space can improve reconstruction performance, as evidenced by much lower LPIPS and DISTS. Additional details and results are presented in the supplementary materials.", + "bbox": [ + 511, + 556, + 908, + 724 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4c7d13a6aac895a6756f4c8294e6e9ae954b3340d894cffcc93a028822f35a03.jpg", + "table_caption": [ + "Table 2. Ablation study of rescaling space and SD prior." + ], + "table_footnote": [], + "table_body": "
Latent spacePixel spaceSD priorLPIPS ↓DISTS ↓
×✓✓0.36300.1154
✓××0.46750.3109
✓×✓0.29790.0886
", + "bbox": [ + 550, + 761, + 870, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2. Necessity of SD Prior", + "text_level": 1, + "bbox": [ + 511, + 835, + 709, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate that SD model provides rich priors about real-world HR images, we directly omit the perceptual enhancement part. As shown in Table 2 (2nd row), the introduction", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15599", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd4ddc584944e372d5db7496c087e1745d12376f4e726b92dcdbb857fdf82d0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 87, + 250, + 147 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ad96ea359717441ed6b97350ac3e3e7d239ed570350e582011de0a70f6c02f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 148, + 248, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/736ea63ecf9ac2858c6dbe394a4b539e83d3d845dc5d1f275e68e346e4756525.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 210, + 248, + 268 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a528fdf9259a50289bf98dc1ee7f8704c1dcc188e435c4975b945dbcaddf813a.jpg", + "image_caption": [ + "(a) StableSR" + ], + "image_footnote": [], + "bbox": [ + 112, + 271, + 248, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f21437c8c9194ba8527af1d16b2c1aad498338ffb8e7c457be5593ca691a4620.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 87, + 328, + 147 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f192fb0a0ce666244f33b1c476f3c57ef5db36b9869194dc90449b99c9bed680.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 148, + 328, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec99751125d639c4342048c63d38e32ecd36f0712f709c016003797bfb635bdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 210, + 328, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/40480862e59a7bc2154612944e23626525d056aa94f6688975f88bcf76b12d3e.jpg", + "image_caption": [ + "(b) S3Diff" + ], + "image_footnote": [], + "bbox": [ + 253, + 271, + 328, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce7e424750f685b1e9c7fdc635fbf278bba6abe447500e7e7e9bdef55c50ed61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 87, + 406, + 147 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0e870a5e34d9d4b162d2cc0a77664cfe0f80e682e46d5061a82fe9006d2003a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 148, + 406, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75fff6b0e603a820472cc8ab0c24b0b4511190d136b7d88cd1cd3daf86fa54ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 210, + 406, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2e8c024997cf5bcc4672a14ce78c1b6d5c3d4e50c668f37d615a4aecb85e0e5a.jpg", + "image_caption": [ + "(c)ESRGAN" + ], + "image_footnote": [], + "bbox": [ + 331, + 271, + 406, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff32ee53fec9d1c2fffa6c6e92789b67184aae4e33ed5f6e51c663c127cec4f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 87, + 486, + 147 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d7a76e49bfc44da7c92f83cf97ff6f3ae70bb234824947c3e841ddec9461a5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 148, + 486, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1555b512a06c3b435d04bde73b0f7450fadf6ec84e2c98f5e728359fe0a1adab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 210, + 486, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0224e89361a889d0b280f64e1e88a0f3750f236c8872ad13250582f078d8fac4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 271, + 486, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c722e92a09990f1d30c2e50ef1e560ba8338762fc8ee3849fb5fc43228aef23b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 87, + 566, + 147 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9c133f8e1dcfab98e86580b5fce12b436319dd8f1d90c57a9a0d398692c5a2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 148, + 566, + 208 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68e57e783fd83330b93f324cfe085b9130b9d13367d09244b26d4e727d7b6412.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 210, + 566, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1be1b4abf1fb7b7801e69b69c925ceaf6e86950729c8432674164e36d44647ae.jpg", + "image_caption": [ + "(d) IRN" + ], + "image_footnote": [], + "bbox": [ + 491, + 271, + 566, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d7b6aab8d3d937a2a1762e9951702b989d46df1a6dd7e0f0278b352910a8463.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 87, + 645, + 147 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc21f908646bd65b2d004886f51b93615c6cff9db8f55252296676e49d37da14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 148, + 645, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1005b8308246bc87729221cdd4025084f83e4121fda414d86ae1467fe7467ed0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 570, + 210, + 645, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05a06bd329b41e86ca5e52de11dd852a67b150c6f7839b592dce35404b1c8325.jpg", + "image_caption": [ + "(e) HCFlow" + ], + "image_footnote": [], + "bbox": [ + 570, + 271, + 645, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3009741a3727e4f37d7b124c161837c2907bc01464ce4fa0e7f24f3cc708bd30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 87, + 725, + 147 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d88d75630eb5fcdab0e3240428056c0451d037d8036f9bca63b6db0633f7e7e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 148, + 725, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aecbabf27b17e5d3b8ade4b49cb218b54f363c93e43b35dbf6de95d6c129ad96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 210, + 725, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ed0ceeabb474c54a0036586d0201ef594251d5fddea50ed4bcb8325072e0634.jpg", + "image_caption": [ + "(f) VQIR" + ], + "image_footnote": [], + "bbox": [ + 650, + 271, + 725, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5127534e08da84053f78292987e6df2a887351d853a1db216f01c8ce402a3898.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 87, + 803, + 147 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bdf4734a648115027bd2e98154f6476b394374c83d6073ea9b033d1b1888c0b9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 148, + 803, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/95cb22709bb6411cd8b818ed25e8e70d4609198834788e83093ede0df8580e94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 210, + 803, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a52781ffb42af446eabbe24d887b259d1a241569c777e7087e6e144bf1c34e62.jpg", + "image_caption": [ + "(g) Ours" + ], + "image_footnote": [], + "bbox": [ + 728, + 271, + 803, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f399334587a6bd0fe7c8bde7430e45ec43bcabe23f6c1fbe9609f1b79dac0765.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 807, + 87, + 885, + 147 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db04ff57706227217ee1d6f47cb7266816b0ba4f085fd9e6006bb2474e6da60a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 807, + 148, + 885, + 209 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d39d8aaad51264f7728292b9edff9f5e396d3edb9b2690a01e1d2a40339f9a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 807, + 210, + 885, + 270 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9744568492a477fd40883c68d35c53e14d0f492aa7d70b30af99eb9e734b291.jpg", + "image_caption": [ + "(h) HR" + ], + "image_footnote": [], + "bbox": [ + 807, + 271, + 885, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1fdf82630b05f0f11338afd1c89209e58becfdbed9239da1c0270373597491b6.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons of $16 \\times$ (the 1st and 2nd rows) and $32 \\times$ (the 3rd and 4th rows) rescaling results. Our method achieves higher semantic accuracy, such as more clearer text, more recognizable faces, and more realistic structures." + ], + "image_footnote": [], + "bbox": [ + 117, + 417, + 284, + 517 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/136107dc95f13832cf2213ed1751495fecff08233853d76d0167ef3061ecb0f6.jpg", + "image_caption": [ + "Figure 6. The validation loss curves when removing a specific component from the DFRM." + ], + "image_footnote": [], + "bbox": [ + 285, + 417, + 457, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of the SD priors can significantly improve performance of LPIPS and DISTS and greatly enhance the perceptual quality of the reconstruction results.", + "bbox": [ + 89, + 599, + 483, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.3. Effectiveness of DFRM Design", + "text_level": 1, + "bbox": [ + 89, + 676, + 348, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our DFRM includes two specific designs: the domain converter module and the pixel guidance module. To analyze their impact, we remove each component individually and plot the validation loss curves, as shown in Fig. 6. It can be observed that the introduction of the INN can reduce reconstruction error. Although the pixel guidance module does not improve reconstruction performance, our experiments show that it can enhance the fidelity and visual quality of the LR images, as demonstrated by the lower guidance loss in Fig. 6. Therefore, the design of our DFRM strikes a good balance between feature reconstruction and LR image quality. More experiments and the visual quality comparisons of LR images are presented in the supplementary materials.", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12d6e60aab704f14b0c01532d5e8675ad64b812228a3e09ef474ec91432817d4.jpg", + "image_caption": [ + "Figure 7. Ablation study about time alignment." + ], + "image_footnote": [], + "bbox": [ + 550, + 431, + 875, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.4. Effectiveness of Time Alignment", + "text_level": 1, + "bbox": [ + 511, + 636, + 785, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the effectiveness of our proposed time-step alignment strategy, we retrain our model using fixed timesteps. This involves removing the TPM and time scheduler modules and employing a manually selected time-step to perform a single denoising step based on Eq. (9). Following the setting of previous work [35], we select fixed time-steps 1 and 999 for comparison. As shown in Fig. 7, when the time-step is set to a smaller value, the model achieves higher fidelity but lower perceptual quality, and the situation is reversed when a larger time-step is chosen. In contrast, our time-step alignment mechanism can dynamically determine the level of generative capability based on the image content, thereby achieving optimal performance in both fidelity and perceptual quality.", + "bbox": [ + 511, + 656, + 906, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 8, we present a schematic illustration of the tiled inference process along with the predicted time-step map.", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15600", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7407938f94e679e96e92d8d91be4971cc70e9c2a24382da3d43da75011d6ed5e.jpg", + "image_caption": [ + "Figure 8. Illustration of the tiled inference process and the predicted time-step map." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 236 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It can be observed that the time alignment mechanism effectively captures more complex regions in high-resolution images and assigns larger time steps to these areas. Details regarding the selection of patch size and stride length are provided in the supplementary materials.", + "bbox": [ + 89, + 304, + 483, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Discussions", + "text_level": 1, + "bbox": [ + 89, + 392, + 217, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.1. Application to Image Compression", + "text_level": 1, + "bbox": [ + 89, + 415, + 377, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our model can be combined with existing lossless image compression methods to achieve extreme compression ratios. To validate this, we use PNG to compress LR images and compare it with JPEG at various bit-rates. As shown in Fig. 9, at the same bits per pixel (bpp), the image reconstruction quality of our model significantly surpasses that of the JPEG compression. These comparisons demonstrate the potential of our model for image compression.", + "bbox": [ + 89, + 434, + 483, + 555 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9fe5368702fcc26077d80f5fbcd88347ea06463541624be324166e00968d5859.jpg", + "image_caption": [ + "Figure 9. R-D curves of JPEG compression and our proposed model on the DIV2K validation set." + ], + "image_footnote": [], + "bbox": [ + 116, + 571, + 447, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2. Inference Efficiency Comparison", + "text_level": 1, + "bbox": [ + 89, + 775, + 366, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 3, we provide comparisons of inference time and memory consumption for representative methods. All tests are performed on an RTX 3090 using images with a resolution of $1024 \\times 1024$ . It can be observed that flow-based models (IRN and HCFlow) require less GPU memory consumption and enable fast inference. In contrast, StableSR, which is based on multi-step diffusion models, demands", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "significantly more inference time and larger GPU memory consumption. Although VQIR requires less memory, the nearest neighbor search process in VQGAN results in longer inference time. For one-step diffusion-based method S3Diff, due to its degradation-guided LoRA which needs to dynamically predict LoRA parameters during inference, it requires longer ingerence time and higher memory consumption. In contrast, our proposed TADM demonstrates moderate efficiency and scalability compared to other methods. In the future, we can utilize distillation techniques to further reduce the model size.", + "bbox": [ + 511, + 90, + 906, + 256 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/29094e4ed95f51eee30b83dd26c7e297716cf84a69fe007772bead61fd47b8c6.jpg", + "table_caption": [ + "Table 3. Comparison of inference time and memory consumption." + ], + "table_footnote": [], + "table_body": "
IRNHCFlowVQIRStableSRS3DiffOurs
Inference Time (s)0.3590.3630.897135.22.1860.711
Memory Consumption (G)0.4980.0980.3446.3805.0252.598
", + "bbox": [ + 519, + 296, + 901, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.3. Upper Bound of TADM", + "text_level": 1, + "bbox": [ + 511, + 354, + 723, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since our method performs the rescaling operation in the latent space of a pre-trained VAE, its performance upper bound is inherently constrained by the VAE. In Table 4, we present a performance comparison between VAE and two $4 \\times$ rescaling models. For VAEs, we directly feed the high-resolution images to the encoder and decoder without performing any rescaling operations. This performance represents the upper bound of our method's performance across all rescaling factors. It can be observed that VAE significantly lags behind the latter. In the future, we could explore integrating stronger VAEs, such as those from SDXL, SD3, thereby enabling the TADM to be more applicable to small-scale rescaling tasks.", + "bbox": [ + 511, + 375, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/69dd70094646137f2d06d6e7c6ed0b6e507b4d0f588963fe6d9daf5c5c784dcd.jpg", + "table_caption": [ + "Table 4. Comparison of VAE and $4 \\times$ rescaling methods." + ], + "table_footnote": [], + "table_body": "
IRNHCFlowVAE
PSNR35.0735.2329.01
SSIM0.93180.93460.8046
", + "bbox": [ + 593, + 613, + 826, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 691, + 633, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel framework called TADM to address issues such as insufficient texture and semantic inaccuracies in extreme image rescaling. Specifically, we design a DFRM to achieve bidirectional mapping between HR features and LR images. Then, we utilize a pre-trained SD model to perform a single denoising step on the rescaled features, enhancing its perceptual quality. Considering the non-uniform quality of the rescaled latent features, we propose a novel time-step alignment strategy to achieve a balance between fidelity and perceptual quality. Both qualitative and quantitative experiments demonstrate the superiority of our approach.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15601", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 217, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5", + "[2] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2567-2581, 2020. 5", + "[3] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021. 1, 2", + "[4] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10021-10030, 2023. 2", + "[5] Shuhang Gu, Andreas Lugmayr, Martin Danelljan, Manuel Fritsche, Julien Lamour, and Radu Timofte. Div8k: Diverse 8k resolution image dataset. In IEEE International Conference on Computer Vision Workshop, pages 3512-3516, 2019. 5", + "[6] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 5", + "[7] Zhaoyang Jia, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Generative latent coding for ultra-low bitrate image compression. In IEEE Conference on Computer Vision and Pattern Recognition, pages 26088-26098, 2024. 1", + "[8] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into conditional gans. arXiv preprint arXiv:2405.05967, 2024. 4", + "[9] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 1, 2", + "[10] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1, 2", + "[11] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In IEEE International Conference on Computer Vision, pages 5148-5157, 2021. 5" + ], + "bbox": [ + 93, + 243, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported in part by the National Natural Science Foundation of China under Grant 62201403, in part by Hubei Provincial Natural Science Foundation of China under Grant 2024AFB549. The numerical calculations in this paper have been done on the supercomputing system in the Supercomputing Center of Wuhan University.", + "[12] Heewon Kim, Myungsub Choi, Bee Lim, and Kyoung Mu Lee. Task-aware image downscaling. In European Conference on Computer Vision, pages 399-414, 2018. 1, 2", + "[13] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2", + "[14] Jinmin Li, Tao Dai, Jingyun Zhang, Kang Liu, Jun Wang, Shaoming Wang, Shu-Tao Xia, et al. Boundary-aware decoupled flow networks for realistic extreme rescaling. arXiv preprint arXiv:2405.02941, 2024. 1", + "[15] Shang Li, Guixuan Zhang, Zhengxiong Luo, Jie Liu, Zhi Zeng, and Shuwu Zhang. Approaching the limit of image rescaling via flow guidance. arXiv preprint arXiv:2111.05133, 2021. 2, 3", + "[16] Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing, 28(3): 1092-1107, 2018. 2", + "[17] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image superresolution and image rescaling. In IEEE International Conference on Computer Vision, pages 4076-4085, 2021. 1, 2, 5, 6", + "[18] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 5", + "[19] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diff-bir: Toward blind image restoration with generative diffusion prior. In European Conference on Computer Vision, pages 430-448, 2024. 2", + "[20] Siwei Ma, Xinfeng Zhang, Chuanmin Jia, Zhenghui Zhao, Shiqi Wang, and Shanshe Wang. Image and video compression with neural networks: A review. IEEE Transactions on Circuits and Systems for Video Technology, 30(6):1683-1698, 2019. 3", + "[21] Zhihong Pan, Baopu Li, Dongliang He, Mingde Yao, Henhao Wu, Tianwei Lin, Xin Li, and Errui Ding. Towards bidirectional arbitrary image rescaling: Joint optimization and cycle idempotence. In IEEE Conference on Computer Vision and Pattern Recognition, pages 17389-17398, 2022. 1", + "[22] Chenyang Qi, Xin Yang, Ka Leong Cheng, Ying-Cong Chen, and Qifeng Chen. Real-time 6k image rescaling with rate-distortion optimization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14092-14101, 2023. 1", + "[23] Lucas Relic, Roberto Azevedo, Markus Gross, and Christopher Schroers. Lossy image compression with foundation diffusion models. In European Conference on Computer Vision, pages 303-319, 2024. 6", + "[24] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions" + ], + "bbox": [ + 93, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15602", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2", + "[25] Wanjie Sun and Zhenzhong Chen. Learned image downscaling for upscaling using content adaptive resampler. IEEE Transactions on Image Processing, 29:4027-4040, 2020. 1, 2", + "[26] George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Ballé, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020. 5", + "[27] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5", + "[28] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 5", + "[29] Luozhou Wang, Shuai Yang, Shu Liu, and Ying-cong Chen. Not all steps are created equal: Selective diffusion distillation for image manipulation. In IEEE International Conference on Computer Vision, pages 7472-7481, 2023. 2", + "[30] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision Workshops, pages 0–0, 2018. 5", + "[31] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In IEEE Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 2", + "[32] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5", + "[33] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3365-3387, 2020. 3", + "[34] Hao Wei, Chenyang Ge, Zhiyuan Li, Xin Qiao, and Pengchao Deng. Towards extreme image rescaling with generative prior and invertible prior. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6181-6193, 2024. 1, 2, 5", + "[35] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. In Advances in Neural Information Processing Systems, pages 92529-92553, 2024. 2, 7", + "[36] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 25456-25467, 2024. 2", + "[37] Mingqing Xiao, Shuxin Zheng, Chang Liu, Zhouchen Lin, and Tie-Yan Liu. Invertible rescaling network and its ex" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tensions. International Journal of Computer Vision, 131(1): 134-159, 2023. 1, 2, 5", + "[38] Bingna Xu, Yong Guo, Luoqian Jiang, Mianjie Yu, and Jian Chen. Downscaled representation matters: Improving image rescaling with collaborative downscaled images. In IEEE International Conference on Computer Vision, pages 12237-12247, 2023. 2", + "[39] Jinhai Yang, Mengxi Guo, Shijie Zhao, Junlin Li, and Li Zhang. Self-asymmetric invertible network for compression-aware image rescaling. In AAAI Conference on Artificial Intelligence, pages 3155-3163, 2023. 1", + "[40] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In European Conference on Computer Vision, pages 74-91, 2024. 2", + "[41] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. In Advances in Neural Information Processing Systems, pages 13294-13307, 2023. 2", + "[42] Zongsheng Yue, Kang Liao, and Chen Change Loy. Arbitrary-steps image super-resolution via diffusion inversion. arXiv preprint arXiv:2412.09013, 2024. 2", + "[43] Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step image super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 2, 5", + "[44] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2", + "[45] Min Zhang, Zhihong Pan, Xin Zhou, and C-C Jay Kuo. Enhancing image rescaling using dual latent variables in invertible neural network. In ACM International Conference on Multimedia, pages 5602-5610, 2022. 2", + "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5", + "[47] Zhixuan Zhong, Liangyu Chai, Yang Zhou, Bailin Deng, Jia Pan, and Shengfeng He. Faithful extreme rescaling via generative prior reciprocated invertible representations. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5708-5717, 2022. 1, 2", + "[48] Yiming Zhu, Cairong Wang, Chenyu Dong, Ke Zhang, Hongyang Gao, and Chun Yuan. High-frequency normalizing flow for image rescaling. IEEE Transactions on Image Processing, 32:6223-6233, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 768 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15603", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/5b02e9a3-c733-4ab2-9aa6-549d5b3e9de2_model.json b/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/5b02e9a3-c733-4ab2-9aa6-549d5b3e9de2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..40cfe877286ba61b8a2285fa3f04a538efefa6e9 --- /dev/null +++ b/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/5b02e9a3-c733-4ab2-9aa6-549d5b3e9de2_model.json @@ -0,0 +1,2629 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.817, + 0.154 + ], + "angle": 0, + "content": "Timestep-Aware Diffusion Model for Extreme Image Rescaling" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.181, + 0.707, + 0.217 + ], + "angle": 0, + "content": "Ce Wang, Zhenyu Hu, Wanjie Sun* Zhenzhong Chen Wuhan University" + }, + { + "type": "text", + "bbox": [ + 0.283, + 0.219, + 0.706, + 0.234 + ], + "angle": 0, + "content": "{cewang, zhenyuhu, sunwanjie, zzchen}@whu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.302, + 0.482, + 0.725 + ], + "angle": 0, + "content": "Image rescaling aims to learn the optimal low-resolution (LR) image that can be accurately reconstructed to its original high-resolution (HR) counterpart, providing an efficient image processing and storage method for ultra-high definition media. However, extreme downscaling factors pose significant challenges to the upscaling process due to its highly ill-posed nature, causing existing image rescaling methods to struggle in generating semantically correct structures and perceptual friendly textures. In this work, we propose a novel framework called Timestep-Aware Diffusion Model (TADM) for extreme image rescaling, which performs rescaling operations in the latent space of a pretrained autoencoder and effectively leverages powerful natural image priors learned by a pre-trained text-to-image diffusion model. Specifically, TADM adopts a pseudo-invertible module to establish the bidirectional mapping between the latent features of the HR image and the target-sized LR image. Then, the rescaled latent features are enhanced by a pre-trained diffusion model to generate more faithful details. Considering the spatially non-uniform degradation caused by the rescaling operation, we propose a novel time-step alignment strategy, which can adaptively allocate the generative capacity of the diffusion model based on the quality of the reconstructed latent features. Extensive experiments demonstrate the superiority of TADM over previous methods in both quantitative and qualitative evaluations. The code will be available at: https://github.com/wwangcece/TADM." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.757, + 0.222, + 0.772 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.783, + 0.482, + 0.875 + ], + "angle": 0, + "content": "With the explosive growth of ultra high-resolution (HR) images, image rescaling has become essential for enhancing data storage and transmission efficiency [39], optimizing resource usage in real-time applications [22], and ensuring proper display on devices with varying resolutions [21]. During image rescaling, HR images are downscaled" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.905, + 0.378 + ], + "angle": 0, + "content": "to low-resolution (LR) images for storage, transmission, or display. When HR images are needed, an upscaling process is employed to restore the downscaled images to their original resolution. Unlike image super-resolution (SR), image rescaling allows for joint optimization of both downscaling and upscaling processes, preserving detailed information crucial for effective upscaling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.905, + 0.546 + ], + "angle": 0, + "content": "Recently proposed image rescaling methods can be classified into two main categories: encoder-decoder architecture [12, 25] and invertible neural networks [17, 37]. The encoder-decoder architecture employs two neural networks to perform downscaling and upscaling. While, invertible neural networks view image rescaling as a reversible process, using the forward and backward passes of an invertible neural network to simulate the downscaling and upscaling. Despite the significant improvements in reconstruction quality achieved by these methods, they are limited to small rescaling factors, such as \\(2 \\times\\) or \\(4 \\times\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.905, + 0.777 + ], + "angle": 0, + "content": "However, with the advancement of ultra HR imaging devices, there is an increasing need for extreme rescaling factors (e.g., \\(16 \\times\\) or \\(32 \\times\\)) to achieve more efficient storage and transmission of ultra HR data [7, 47]. Existing image rescaling methods often struggle to generate sufficient detail under extreme rescaling factors. Therefore, some approaches propose using pre-trained GANs as priors to constrain the solution space of the reconstructed results. GRAIN [47] and BDFlow [14] use pre-trained StyleGAN [9, 10] to generate high-quality upscaled face images. VQIR [34] leverages the high-quality visual embeddings encapsulated in pre-trained VQGAN [3] to achieve extreme rescaling on natural images. Although these methods have made significant progress in perceptual metrics, their reconstruction results still suffer from incorrect semantics and structure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To address the aforementioned issues, we propose a novel framework named Timestep-Aware Diffusion Model (TADM) for extreme image rescaling, which performs image rescaling in the latent space of a pre-trained autoencoder and leverages the natural image priors stored in large text-to-image diffusion models to restore HR images. Specifically, first, the HR images are encoded by the via the VAE encoder to obtain the latent features. Then, we use an au" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15594" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.378 + ], + "angle": 0, + "content": "toencoder to rescale the latent code of the HR image to the target size, and employ a set of invertible neural networks (INN) to perform the mapping from the feature space to the pixel space to obtain the LR image. Next, the rescaled features are fed into a pre-trained diffusion model to perform a single denoising step, thereby enhancing their perceptual quality. Finally, the enhanced features are decoded via the VAE decoder to obtain the rescaled image. Additionally, we find that the rescaled latent features exhibit non-uniform degradation, which can be interpreted as the addition of noise at varying levels. As a result, different images require different time-steps to perform the denoising process. Previous studies have shown that the time-step in diffusion models directly controls the generative capacity of the model [29]. Therefore, we propose a time-step alignment mechanism that predicts the appropriate timestep based on the reconstruction quality of the rescaled latent features, thereby adaptively tuning the generative capacity of the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.38, + 0.435, + 0.394 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.397, + 0.483, + 0.44 + ], + "angle": 0, + "content": "- We propose a Timestep-Aware Diffusion Model that performs well in global semantic and structural reconstruction for the extreme image rescaling task." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.442, + 0.483, + 0.5 + ], + "angle": 0, + "content": "- We propose a novel decoupled strategy for feature rescaling. It decouples the generation of LR images from the rescaling operation, thereby improving the reconstruction accuracy of the rescaling process." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.502, + 0.483, + 0.576 + ], + "angle": 0, + "content": "- We design a time-step alignment strategy to address the spatially non-uniform reconstruction quality of the rescaled latent features, allowing for more effective utilization of the prior information from the text-to-image diffusion model." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.397, + 0.483, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.595, + 0.233, + 0.61 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.254, + 0.637 + ], + "angle": 0, + "content": "2.1. Image Rescaling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Image rescaling aims to downscale HR images to visually coherent LR images and then plausibly reconstruct the original HR images. Unlike image SR, image rescaling involves joint modeling of both downscaling and upscaling processes, leading to higher reconstruction accuracy. Earlier researches [12, 16] utilize CNNs to build an encoder-decoder framework, treating image rescaling as a unified task by jointly training the downscaling and upscaling processes. Instead of directly generating LR images, CAR [25] proposes generating content-adaptive resampling kernels based on HR images to perform the downsampling operation, avoiding pixel-level constraints on the downscaled images. To fully leverage the inherent reversibility of image rescaling, there has been a surge in the use of INNs for this task [15, 38, 45, 48]. IRN [37] is the first invertible framework that models image rescaling as a bijective transformation, embedding residual high-frequency (HF) components" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "into a case-agnostic latent distribution for efficient reconstruction. HCFlow [17] assumes that the HF information depends on the LR image, thereby achieving better performance by incorporating LR conditions in HF information estimation. In extreme image rescaling, generative priors are usually employed to constrain the outcomes. GRAIN [47] and VQIR [34] leverage pre-trained StyleGAN [9, 10] and VQGAN [3], respectively, to achieve notable enhancements in perceptual quality. Nonetheless, these approaches are often limited in image contents, either confined to handling images from specific domains or encountering difficulties when reconstructing faces and textual elements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.285, + 0.783, + 0.301 + ], + "angle": 0, + "content": "2.2. Diffusion Model for Image SR" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.905, + 0.457 + ], + "angle": 0, + "content": "Here, we briefly introduce the application of diffusion models in a related task, i.e., image SR. Diffusion models promote the development of natural image SR in two main ways: training diffusion models from scratch [4, 13, 41] and leveraging pre-trained SD model [19, 36, 40]. The former is represented by SR3 [24], which achieves perceptual quality comparable to GANs. The latter is represented by StableSR [28], which aims to build a model capable of handling image restoration under any complex degradation in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.717 + ], + "angle": 0, + "content": "However, the above methods typically require dozens or even hundreds of iterations during inference, which significantly increases computational costs. Researchers have begun exploring how to achieve image SR using one-step diffusion models. SinSR [31], based on ResShift [41], derives a deterministic one-step sampling process and employs a consistency preserving loss to distill the prior knowledge from a multi-step teacher network into a one-step student network. A representative work, OSEDiff [35], directly feeds the LQ image into the SD model and employs variational score distillation for regularization. S3Diff [43] proposes a degradation-aware LoRA and utilizes a pre-trained DINO [44] model as a discriminator for adversarial training. InvSR [42] designs a partial noise prediction strategy to provide a dynamic starting sampling point and supports inference with an arbitrary number of sampling steps ranging from one to five." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.732, + 0.649, + 0.749 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.702, + 0.772 + ], + "angle": 0, + "content": "3.1. Overview of TADM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Fig. 1, TADM consists of four parts: Latent Encoding, Feature Rescaling in the Latent Space, Denoising Guided Perception Enhancement and Latent Decoding. First, the input HR images \\( x \\) are mapped to the latent features \\( z \\) via the pre-trained VAE encoder \\( \\mathcal{E} \\). Subsequently, the proposed Decoupled Feature Rescaling Module (DFRM) rescales \\( z \\) to the target size, outputting LR images \\( y \\) and the rescaled latent features \\( \\hat{z} \\). Our aim is to make \\( \\hat{z} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15595" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.272, + 0.868, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.413, + 0.908, + 0.497 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed Timestep-Aware Diffusion Model (TADM). First, the HR images \\( x \\) are encoded to obtain the latent features \\( z \\). The latent features are then rescaled to the target size using the decoupled feature rescaling module (DFRM), producing the LR images \\( y \\) and outputting the rescaled latent features \\( \\hat{z} \\). Next, the rescaled latent features \\( \\hat{z} \\) are passed into a pre-trained diffusion model to perform a single denoising step for perceptual enhancement to obtain \\( \\hat{z}_0 \\). In this process, a time prediction module (TPM) is used to estimate the time step \\( t \\) based on \\( \\hat{z} \\) and the predicted \\( t \\) is then fed into both the U-Net and the time scheduler. Finally, the perceptually enhanced latent features \\( \\hat{z}_0 \\) are decoded to obtain the rescaled image \\( \\hat{x} \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.523, + 0.486, + 0.779 + ], + "angle": 0, + "content": "as close as possible to \\( z \\) so that the rescaled results align with the latent space of the pre-trained VAE, while ensuring that \\( y \\) and the original HR images \\( x \\) share the same content. Then, the rescaled features \\( \\hat{z} \\) are fed into a denoising U-Net network to perform a single denoising step to get the perceptually enhanced latent features \\( \\hat{z}_0 \\). Finally, \\( \\hat{z}_0 \\) is decoded via the pre-trained VAE decoder \\( \\mathcal{D} \\) to obtain the rescaled image \\( \\hat{x} \\). Overall, the latent rescaling process can be regarded as a noise-adding process, while the latent enhancement process can be viewed as a denoising process. To dynamically estimate the 'noise' density introduced by latent rescaling, we design a time alignment strategy, which consists of a time-step prediction module (TPM) and a time scheduler module. It can better address the non-uniform degradation caused by rescaling, thereby enabling the dynamic allocation of generative capabilities of pre-trained SD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.795, + 0.418, + 0.813 + ], + "angle": 0, + "content": "3.2. Feature Rescaling in the Latent Space" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.82, + 0.295, + 0.833 + ], + "angle": 0, + "content": "3.2.1. Network Architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "In this part, we perform rescaling of the HR latent features \\( z \\), obtaining LR images \\( y \\) and rescaled features \\( \\hat{z} \\). A straightforward idea is to construct a transformation chain as \\( y = G_{\\mathrm{e}}(z) \\), \\( \\hat{z} = G_{\\mathrm{d}}(y) \\). Here, \\( G_{\\mathrm{e}} \\) is the encoder used" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.614 + ], + "angle": 0, + "content": "to downscale the input features to the target sized LR images, while \\( G_{\\mathrm{d}} \\) represents the decoder used to reconstruct the original HR features. Guidance loss \\( \\mathcal{L}_{\\mathrm{gui}} \\) and reconstruction loss \\( \\mathcal{L}_{\\mathrm{rec}} \\) are applied to \\( y \\) and \\( \\hat{z} \\) respectively, to force \\( y \\) to have the same content as \\( x \\), and \\( \\hat{z} \\) to be as similar as possible to \\( z \\):" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.627, + 0.905, + 0.643 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {g u i .}} = \\| y - \\operatorname {B i c u b i c} (x) \\| _ {1} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.646, + 0.905, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c .}} = \\| \\hat {z} - z \\| _ {1} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.825 + ], + "angle": 0, + "content": "However, these two losses are contradictory to each other. When the guidance loss approaches zero, the rescaling problem degenerates into a classical SR problem [33], embedding little information about the downscaling process into the LR image. In contrast, when the reconstruction loss reaches its global optimum, the rescaling problem degenerates into an image compression problem [20], causing the LR image to lack meaningful visual content. Therefore, directly applying the guidance loss \\(\\mathcal{L}_{\\mathrm{gui}}\\) on \\(y\\) could adversely impact subsequent upscaling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Considering the above issues, we devise the DFRM as shown in Fig. 2, which divides the rescaling operation into two sub-tasks: feature rescaling and LR image generation. Inspired by FGRN [15], we use a set of CNNs to perform downscaling and upscaling, and a separate INN to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15596" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.475, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.197, + 0.483, + 0.224 + ], + "angle": 0, + "content": "Figure 2. Detailed architecture of Decoupled Feature Rescaling Module (DFRM)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.484, + 0.369 + ], + "angle": 0, + "content": "perform the bidirectional mapping between the feature domain and the pixel domain. To achieve a balance between feature rescaling accuracy and visual quality of LR images, we introduce another pixel guidance module in the encoder to incorporate pixel-level information, which can improve the quality of the LR image. Specifically, DFRM perform feature rescaling through two independent transformation chains \\(((x,z)\\to z_{\\mathrm{lr}}\\to \\hat{z})\\) . \\((z_{\\mathrm{lr}}\\leftrightarrow y)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.377, + 0.482, + 0.393 + ], + "angle": 0, + "content": "\\[\nz _ {\\mathrm {l r}} = G _ {\\mathrm {e}} (x, z) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.396, + 0.482, + 0.414 + ], + "angle": 0, + "content": "\\[\nz _ {\\mathrm {l r}} \\approx F ^ {- 1} (F (z _ {\\mathrm {l r}})) = F ^ {- 1} (y) \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.416, + 0.482, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\hat {z} = G _ {\\mathrm {d}} \\left(z _ {\\mathrm {l r}}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Here, \\(F\\) and \\(F^{-1}\\) represent the forward and backward processes of the INN, respectively. Given the latent variable \\(z \\in \\mathbb{R}^{\\frac{H}{8} \\times \\frac{W}{8} \\times 4}\\) of the pre-trained SD VAE as input, the encoder \\(G_{\\mathrm{e}}\\) first transforms it into a compact representation \\(z_{\\mathrm{lr}} \\in \\mathbb{R}^{\\frac{H}{N} \\times \\frac{W}{N} \\times 3}\\), whose size is the same as that of the target LR image and \\(N\\) represents the target downscaling factor. Subsequently, a set of INNs perform the forward process to transform \\(z_{\\mathrm{lr}}\\) from the feature domain to the pixel domain and obtain the LR image \\(y \\in \\mathbb{R}^{\\frac{H}{N} \\times \\frac{W}{N} \\times 3}\\) by quantization. In the upscaling process, the inverse of the INN is used to process the LR image \\(y\\) and reconstruct \\(z_{\\mathrm{lr}}\\). Finally, the decoder \\(G_{\\mathrm{d}}\\) is utilized to obtain the rescaled feature \\(\\hat{z}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.627, + 0.24, + 0.64 + ], + "angle": 0, + "content": "3.2.2. Loss Function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.766 + ], + "angle": 0, + "content": "Training of this part involves two losses: reconstruction loss and guidance loss. Considering the information lost in quantization operations, the bidirectional mapping between \\( z_{\\mathrm{lr}} \\) and \\( y \\) is not exactly reversible. To ensure the robustness of the network to quantization operations, we consider two types of transformation chains in designing the reconstruction loss: \\( (x,z)\\to z_{\\mathrm{lr}}\\to \\hat{z} \\) and \\( (x,z)\\to z_{\\mathrm{lr}}\\to y\\to \\hat{z}_{\\mathrm{lr}}\\to \\hat{z} \\), which can be expressed as the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.773, + 0.483, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {r e c .}} = \\| G _ {\\mathrm {d}} \\left(G _ {\\mathrm {e}} (x, z)\\right) - z \\| _ {1} + \\tag {6} \\\\ \\left. \\left\\| G _ {\\mathrm {d}} \\left(F ^ {- 1} \\left(F \\left(G _ {\\mathrm {e}} (x, z)\\right)\\right)\\right) - z \\right\\| _ {1} \\right. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.484, + 0.877 + ], + "angle": 0, + "content": "This combination can facilitate the network to reconstruct \\(\\hat{z}\\) with high fidelity in both the existence and absence of quantization operations. Additionally, the guidance loss is designed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {g u i .}} = \\left\\| F \\left(G _ {\\mathrm {e}} (x, z)\\right) - \\operatorname {B i c u b i c} (x) \\right\\| _ {1} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.12 + ], + "angle": 0, + "content": "Finally, the loss function of this part is a combination of these two losses:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.131, + 0.905, + 0.148 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e s .}} = \\lambda_ {\\text {r e c .}} \\mathcal {L} _ {\\text {r e c .}} + \\lambda_ {\\text {g u i .}} \\mathcal {L} _ {\\text {g u i .}} \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.155, + 0.884, + 0.17 + ], + "angle": 0, + "content": "3.3. Denoising Guided Perception Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.177, + 0.907, + 0.358 + ], + "angle": 0, + "content": "The latent variable \\(\\hat{z}\\) obtained in Sec. 3.2 retains sufficient semantic information about the HR image, but lacks adequate texture. Therefore, we propose to leverage the strong priors stored in the pre-trained SD model to enhance the perceptual quality of the rescaling results in this part. Original diffusion model requires a multi-step sampling algorithm, significantly increases inference time and makes it difficult to scale to practical applications. Inspired by recent work on accelerating diffusion models [8], we adopt an one-step inference strategy. Specifically, we can directly feed the rescaled latent variable \\(\\hat{z}\\) into the denoising U-Net \\(\\epsilon = \\epsilon_{\\theta}(\\hat{z},t)\\) and perform a single denoising step as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.378, + 0.905, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {0} = \\frac {\\hat {z} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon}{\\sqrt {\\bar {\\alpha} _ {t}}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.906, + 0.538 + ], + "angle": 0, + "content": "where \\(\\hat{z}_0\\) represents the perceptually enhanced latent features, \\(\\bar{\\alpha}_{t}\\) is the noise scheduler of the forward diffusion process of the SD model, and \\(t\\) represents the time-step associated with the noise intensity. Finally, \\(\\hat{z}_0\\) can be transformed to the pixel domain through the pre-trained VAE decoder \\(\\mathcal{D}\\) to obtain the upscaled image \\(\\hat{x} = \\mathcal{D}(\\hat{z}_0)\\). Then, we can adopt the loss function between \\(\\hat{x}\\) and \\(x\\) to optimize the U-Net using LoRA layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.903 + ], + "angle": 0, + "content": "However, the key problem of the aforementioned approaches lies in the selection of time-step \\( t \\). We can regard the rescaling process as the noise-adding process of diffusion model, while Eq. (9) represents the denoising process. For different rescaling factors and different images, the 'noise' introduced by the rescaling operation varies in intensity. In Fig. 3, we present the global rescaling error of an image under different rescaling factors, as well as the local rescaling error of different patches within the image. Additionally, we plot the error curve of the forward noise-adding process of the SD model. It can be observed that as the rescaling factor and image content change, rescaling MSE also varies and corresponds to the diffusion MSE at different time-steps. Thus, we propose to dynamically predict the time-step \\( t \\) based on the quality of the rescaled latent features \\( \\hat{z} \\). A simple yet effective network named timestep prediction module (TPM), whose structure is presented in Fig. 1, is used to predict \\( t \\) based on the rescaled latent features: \\( t = \\mathrm{TPM}(\\hat{z}) \\). During the training phase, we compute a single time-step for each input image patch. In the inference phase, we adopt a tiled inference approach (which will be elaborated in detail in the supplementary materials), allowing for predicting different time-steps for various regions of a high-resolution input image." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15597" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.089, + 0.462, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.485, + 0.371 + ], + "angle": 0, + "content": "Figure 3. We compare the MSE introduced by the rescaling operation with that caused by the forward diffusion process. The rescaling degradation severity is correlated with both the image content and the rescaling scale. Therefore, the model requires adaptive time-step to align the rescaling MSE with the diffusion MSE." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.382, + 0.462, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.551 + ], + "angle": 0, + "content": "Figure 4. The structure of the time scheduler module. It combines both the fixed scheduler and learnable scheduler to perform the denoising process. The two schedulers are connected by a zero-convolution layer to stabilize the training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.484, + 0.816 + ], + "angle": 0, + "content": "Subsequently, the predicted time-step \\( t \\) is used to perform a single denoising step. However, the vanilla SD time scheduler as shown in Eq. (9) is not differentiable with respect to \\( t \\), which prevents gradients from being backpropagated to the TPM. Considering that given \\( \\hat{z} \\), \\( \\epsilon \\) and \\( t \\), Eq. (9) essentially represents a linear transformation, which can be learned by a neural network. Therefore, we propose a hybrid time scheduler module consisting of a fixed scheduler and a learnable scheduler, whose structure is presented in Fig. 4. The former uses Eq. (9) to perform the denoising step, while the latter employs a neural network to simulate the denoising scheduler. To ensure training stability, the final layer of the learnable scheduler is a zero-initialized convolution, allowing it to gradually correct the results of the fixed scheduler as training progresses. This process can be described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.829, + 0.483, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {0} = \\mathcal {S} _ {\\text {f i x e d}} (\\hat {z}, \\epsilon , t _ {0}) + \\mathcal {S} _ {\\text {l e a r n e d}} (\\hat {z}, \\epsilon , t) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.885 + ], + "angle": 0, + "content": "where \\( t_0 \\) is the fixed pre-set time-step and \\( S_{\\mathrm{fixed}} \\) denotes the vanilla SD time scheduler as shown in Eq. (9)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Finally, we fine-tune the denoising U-Net network and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "the VAE decoder using LoRA, while also jointly training the TPM and time scheduler module. To achieve a balance between fidelity and realism, we supervise the reconstruction results \\(\\hat{x}\\) using pixel loss and perceptual loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.163, + 0.906, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {e n h .}} = \\| \\hat {x} - x \\| _ {1} + \\tag {11} \\\\ \\lambda_ {\\text {p e c .}} \\left(\\mathcal {L} _ {\\text {l p i p s}} (\\hat {x}, x) + \\mathcal {L} _ {\\text {d i s t s}} (\\hat {x}, x)\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.21, + 0.646, + 0.226 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.234, + 0.704, + 0.251 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.257, + 0.791, + 0.27 + ], + "angle": 0, + "content": "4.1.1. Datasets and Evaluation Metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.472 + ], + "angle": 0, + "content": "We train the proposed TADM using the DF2K dataset [1, 18]. Considering that traditional benchmarks such as Set5, Set14, and BSD100 often have lower quality and resolution [5], they are not suitable for evaluating extreme rescaling methods. Therefore, we utilize four high-resolution datasets: DIV2K, CLIC2020 [26], Urban100 [6], and DIV8K [5] to evaluate models been compared (results on the latter two datasets are presented in the supplementary materials). We adopt PSNR and SSIM [32] for quantitative evaluation. For perceptual quality evaluation, reference-based metrics LPIPS [46] and DISTS [2] are used. For non-reference image quality assessment, we use metrics including MUSIQ [11] and CLIPIQA [27]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.479, + 0.673, + 0.493 + ], + "angle": 0, + "content": "4.1.2. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.618 + ], + "angle": 0, + "content": "TADM is built based on the SD 2.1-base model, and the training is divided into three stages. First, we train the DFRM using Eq. (8). Then we adopt Eq. (11) to jointly train the LoRA layers, the TPM and the time scheduler module. Finally, we use a smaller learning rate to jointly fine-tune DFRM, LoRA layers, TPM and scheduler module using a combination of Eq. (8) and (11). More training details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.627, + 0.828, + 0.643 + ], + "angle": 0, + "content": "4.2. Results and Comparison with SOTA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.907, + 0.8 + ], + "angle": 0, + "content": "We compare the proposed TADM with three kinds of methods: (1) bicubic downscaling and SR, i.e., ESRGAN [30], StableSR [28] (based on multi-step diffusion) and S3Diff (based on single-step diffusion)[43]; (2) flow-based rescaling models, i.e., IRN [37] and HCFlow [17]; (3) prior-based extreme rescaling method VQIR [34]. Some of them are proposed for \\(4 \\times\\) rescaling factors and we extended them to \\(16 \\times\\) or \\(32 \\times\\). To ensure a fair comparison, all methods are trained until convergence on the DF2K dataset following their original configurations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.807, + 0.704, + 0.821 + ], + "angle": 0, + "content": "4.2.1. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Table 1 presents the \\(16\\times\\) and \\(32\\times\\) quantitative image rescaling results on four datasets. The performance of SR methods are generally poor. Both gan-based (ESRGAN) and diffusion-based (StableSR and S3Diff) methods can only achieve relative good performance on no-reference metrics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15598" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with different methods at \\( {16} \\times \\) and \\( {32} \\times \\) (results on Urban100 and DIV8K are presented in the supplementary materials). The symbols \\( \\uparrow \\) and \\( \\downarrow \\) respectively represent that higher or lower values indicate better performance. Bold represents the best and underline represents the second best." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.142, + 0.907, + 0.371 + ], + "angle": 0, + "content": "
DatasetMethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓MUSIQ ↑CLIPIQA ↑
16×32×16×32×16×32×16×32×16×32×16×32×
DIV2KESRGAN23.1519.940.59460.56890.44780.58920.23780.423059.8043.350.61610.3734
StableSR22.0219.920.59830.55880.49800.57360.20780.260351.2848.940.37800.3203
S3Diff20.2217.810.54910.46010.40330.48950.13090.181064.3767.920.62280.6991
IRN26.0422.840.70200.62120.51530.60950.30410.403643.7228.700.27990.2336
HCFlow26.6623.890.71760.64670.48850.58160.28660.385246.4337.250.27350.2792
VQIR23.9122.020.64980.58230.31740.45680.10240.266364.0458.210.63500.6293
Ours23.9822.180.67360.58690.29790.42210.08860.168466.5669.120.71890.7204
CLIC2020ESRGAN25.4322.470.70180.68040.42130.54290.22090.401157.8041.280.61070.3558
StableSR24.1122.240.70520.67260.46250.52880.18600.228252.6949.800.37790.3325
S3Diff22.0419.870.66150.57110.39680.48290.13360.181263.0766.400.60010.6838
IRN28.5524.930.79070.72280.46760.55280.27460.377647.6627.810.29500.2495
HCFlow28.8226.210.79490.74510.45420.53250.26350.361147.7237.150.27330.2884
VQIR25.6924.120.73630.69420.28540.41910.09010.244564.3654.830.61050.6068
Ours26.0424.260.75420.69030.27240.39110.07920.151365.3569.100.68150.7271
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.394, + 0.486, + 0.653 + ], + "angle": 0, + "content": "This is because they rely solely on the LR image to reconstruct the original image and cannot jointly optimize the downscaling and upscaling processes. While methods based on INNs (HCFlow and IRN) achieve the highest PSNR and SSIM due to its consideration of the correlation between downscaling and upscaling. In all generative methods, VQIR achieves relatively good perceptual metrics by leveraging the priors of the pre-trained VQGAN. In contrast, our method achieves optimal performance across all perceptual metrics on all datasets. For instance, in the \\(32 \\times\\) rescaling on the DIV2K validation set, our method improves the DISTS metric by \\(36.76\\%\\) compared to the second-best VQIR. Importantly, our approach is designed for applications like social media delivery, where semantic fidelity and perceptual realism are more critical. In such cases, metrics like LPIPS and DISTS better reflect image quality [23]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.27, + 0.683 + ], + "angle": 0, + "content": "4.2.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.485, + 0.902 + ], + "angle": 0, + "content": "In Fig. 5, we present visual comparisons at \\(16 \\times\\) and \\(32 \\times\\) rescaling factors. Due to the ill-posed nature of extreme rescaling, pixel-oriented optimization methods such as IRN, and HCFlow produce overly-smoothed results. Although ESRGAN incorporates perceptual optimization it still cannot generate adequate details. VQIR leverages the pretrained VQGAN as a prior, allowing it to generate rich details and sharp edges. However, it struggles with faces and more complex structures. Although S3Diff can generate sharp edges and rich details, its reconstructed results have low fidelity and exhibit significant artifacts and color distortions. Thanks to the powerful prior provided by the pre-trained SD and the carefully designed architecture,, our method effectively addresses these issues. It not only gener" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.44 + ], + "angle": 0, + "content": "ates sufficiently rich textures but also reconstructs accuracy semantics and regular structures, such as text, faces, vehicles, and leaves." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.449, + 0.667, + 0.466 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.532 + ], + "angle": 0, + "content": "We conduct ablation experiments to validate the effectiveness of each component. All ablations are trained on the DF2K dataset and tested on the validation set of the DIV2K for \\(16 \\times\\) rescaling task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.539, + 0.798, + 0.555 + ], + "angle": 0, + "content": "4.3.1. Effectiveness of Feature Rescaling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.909, + 0.725 + ], + "angle": 0, + "content": "To verify the effectiveness of the feature rescaling operation, we move the rescaling operation from the latent space to the pixel space. Specifically, we use HCFlow [17], the comparative method with the best PSNR performance, to perform the rescaling operation in the pixel domain, followed by perceptual enhancement using the same approach described in Sec. 3.3. As shown the 1st row in Table 2, the rescaling operation in the latent space can improve reconstruction performance, as evidenced by much lower LPIPS and DISTS. Additional details and results are presented in the supplementary materials." + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.737, + 0.877, + 0.752 + ], + "angle": 0, + "content": "Table 2. Ablation study of rescaling space and SD prior." + }, + { + "type": "table", + "bbox": [ + 0.551, + 0.762, + 0.872, + 0.815 + ], + "angle": 0, + "content": "
Latent spacePixel spaceSD priorLPIPS ↓DISTS ↓
×✓✓0.36300.1154
✓××0.46750.3109
✓×✓0.29790.0886
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.837, + 0.71, + 0.852 + ], + "angle": 0, + "content": "4.3.2. Necessity of SD Prior" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To validate that SD model provides rich priors about real-world HR images, we directly omit the perceptual enhancement part. As shown in Table 2 (2nd row), the introduction" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15599" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.088, + 0.25, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.149, + 0.249, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.211, + 0.249, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.272, + 0.249, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.088, + 0.33, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.149, + 0.33, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.211, + 0.33, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.272, + 0.33, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.335, + 0.326, + 0.347 + ], + "angle": 0, + "content": "(a) StableSR" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.088, + 0.408, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.149, + 0.408, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.211, + 0.408, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.272, + 0.408, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.335, + 0.397, + 0.347 + ], + "angle": 0, + "content": "(b) S3Diff" + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.088, + 0.488, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.149, + 0.488, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.211, + 0.488, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.272, + 0.488, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.335, + 0.482, + 0.347 + ], + "angle": 0, + "content": "(c)ESRGAN" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.088, + 0.567, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.149, + 0.567, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.211, + 0.567, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.272, + 0.567, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.335, + 0.548, + 0.347 + ], + "angle": 0, + "content": "(d) IRN" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.088, + 0.646, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.149, + 0.646, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.211, + 0.646, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.272, + 0.646, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.335, + 0.639, + 0.347 + ], + "angle": 0, + "content": "(e) HCFlow" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.088, + 0.726, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.149, + 0.726, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.211, + 0.726, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.272, + 0.726, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.335, + 0.71, + 0.347 + ], + "angle": 0, + "content": "(f) VQIR" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.088, + 0.805, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.149, + 0.805, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.211, + 0.805, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.272, + 0.805, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.335, + 0.791, + 0.347 + ], + "angle": 0, + "content": "(g) Ours" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.088, + 0.886, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.149, + 0.886, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.211, + 0.886, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.272, + 0.886, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.825, + 0.335, + 0.863, + 0.347 + ], + "angle": 0, + "content": "(h) HR" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.365, + 0.908, + 0.394 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons of \\(16 \\times\\) (the 1st and 2nd rows) and \\(32 \\times\\) (the 3rd and 4th rows) rescaling results. Our method achieves higher semantic accuracy, such as more clearer text, more recognizable faces, and more realistic structures." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.419, + 0.285, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.419, + 0.458, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.529, + 0.483, + 0.559 + ], + "angle": 0, + "content": "Figure 6. The validation loss curves when removing a specific component from the DFRM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.484, + 0.645 + ], + "angle": 0, + "content": "of the SD priors can significantly improve performance of LPIPS and DISTS and greatly enhance the perceptual quality of the reconstruction results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.678, + 0.349, + 0.693 + ], + "angle": 0, + "content": "4.3.3. Effectiveness of DFRM Design" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our DFRM includes two specific designs: the domain converter module and the pixel guidance module. To analyze their impact, we remove each component individually and plot the validation loss curves, as shown in Fig. 6. It can be observed that the introduction of the INN can reduce reconstruction error. Although the pixel guidance module does not improve reconstruction performance, our experiments show that it can enhance the fidelity and visual quality of the LR images, as demonstrated by the lower guidance loss in Fig. 6. Therefore, the design of our DFRM strikes a good balance between feature reconstruction and LR image quality. More experiments and the visual quality comparisons of LR images are presented in the supplementary materials." + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.432, + 0.877, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.592, + 0.851, + 0.607 + ], + "angle": 0, + "content": "Figure 7. Ablation study about time alignment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.787, + 0.652 + ], + "angle": 0, + "content": "4.3.4. Effectiveness of Time Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.869 + ], + "angle": 0, + "content": "To validate the effectiveness of our proposed time-step alignment strategy, we retrain our model using fixed timesteps. This involves removing the TPM and time scheduler modules and employing a manually selected time-step to perform a single denoising step based on Eq. (9). Following the setting of previous work [35], we select fixed time-steps 1 and 999 for comparison. As shown in Fig. 7, when the time-step is set to a smaller value, the model achieves higher fidelity but lower perceptual quality, and the situation is reversed when a larger time-step is chosen. In contrast, our time-step alignment mechanism can dynamically determine the level of generative capability based on the image content, thereby achieving optimal performance in both fidelity and perceptual quality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In Fig. 8, we present a schematic illustration of the tiled inference process along with the predicted time-step map." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15600" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.249, + 0.483, + 0.278 + ], + "angle": 0, + "content": "Figure 8. Illustration of the tiled inference process and the predicted time-step map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.305, + 0.484, + 0.382 + ], + "angle": 0, + "content": "It can be observed that the time alignment mechanism effectively captures more complex regions in high-resolution images and assigns larger time steps to these areas. Details regarding the selection of patch size and stride length are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.393, + 0.218, + 0.407 + ], + "angle": 0, + "content": "4.4. Discussions" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.416, + 0.379, + 0.431 + ], + "angle": 0, + "content": "4.4.1. Application to Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.556 + ], + "angle": 0, + "content": "Our model can be combined with existing lossless image compression methods to achieve extreme compression ratios. To validate this, we use PNG to compress LR images and compare it with JPEG at various bit-rates. As shown in Fig. 9, at the same bits per pixel (bpp), the image reconstruction quality of our model significantly surpasses that of the JPEG compression. These comparisons demonstrate the potential of our model for image compression." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.573, + 0.449, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Figure 9. R-D curves of JPEG compression and our proposed model on the DIV2K validation set." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.367, + 0.791 + ], + "angle": 0, + "content": "4.4.2. Inference Efficiency Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In Table 3, we provide comparisons of inference time and memory consumption for representative methods. All tests are performed on an RTX 3090 using images with a resolution of \\(1024 \\times 1024\\). It can be observed that flow-based models (IRN and HCFlow) require less GPU memory consumption and enable fast inference. In contrast, StableSR, which is based on multi-step diffusion models, demands" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.257 + ], + "angle": 0, + "content": "significantly more inference time and larger GPU memory consumption. Although VQIR requires less memory, the nearest neighbor search process in VQGAN results in longer inference time. For one-step diffusion-based method S3Diff, due to its degradation-guided LoRA which needs to dynamically predict LoRA parameters during inference, it requires longer ingerence time and higher memory consumption. In contrast, our proposed TADM demonstrates moderate efficiency and scalability compared to other methods. In the future, we can utilize distillation techniques to further reduce the model size." + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.273, + 0.905, + 0.287 + ], + "angle": 0, + "content": "Table 3. Comparison of inference time and memory consumption." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.297, + 0.903, + 0.336 + ], + "angle": 0, + "content": "
IRNHCFlowVQIRStableSRS3DiffOurs
Inference Time (s)0.3590.3630.897135.22.1860.711
Memory Consumption (G)0.4980.0980.3446.3805.0252.598
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.355, + 0.725, + 0.37 + ], + "angle": 0, + "content": "4.4.3. Upper Bound of TADM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.572 + ], + "angle": 0, + "content": "Since our method performs the rescaling operation in the latent space of a pre-trained VAE, its performance upper bound is inherently constrained by the VAE. In Table 4, we present a performance comparison between VAE and two \\(4 \\times\\) rescaling models. For VAEs, we directly feed the high-resolution images to the encoder and decoder without performing any rescaling operations. This performance represents the upper bound of our method's performance across all rescaling factors. It can be observed that VAE significantly lags behind the latter. In the future, we could explore integrating stronger VAEs, such as those from SDXL, SD3, thereby enabling the TADM to be more applicable to small-scale rescaling tasks." + }, + { + "type": "table_caption", + "bbox": [ + 0.541, + 0.589, + 0.878, + 0.604 + ], + "angle": 0, + "content": "Table 4. Comparison of VAE and \\( 4 \\times \\) rescaling methods." + }, + { + "type": "table", + "bbox": [ + 0.594, + 0.614, + 0.828, + 0.656 + ], + "angle": 0, + "content": "
IRNHCFlowVAE
PSNR35.0735.2329.01
SSIM0.93180.93460.8046
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.693, + 0.634, + 0.708 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose a novel framework called TADM to address issues such as insufficient texture and semantic inaccuracies in extreme image rescaling. Specifically, we design a DFRM to achieve bidirectional mapping between HR features and LR images. Then, we utilize a pre-trained SD model to perform a single denoising step on the rescaled features, enhancing its perceptual quality. Considering the non-uniform quality of the rescaled latent features, we propose a novel time-step alignment strategy to achieve a balance between fidelity and perceptual quality. Both qualitative and quantitative experiments demonstrate the superiority of our approach." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15601" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62201403, in part by Hubei Provincial Natural Science Foundation of China under Grant 2024AFB549. The numerical calculations in this paper have been done on the supercomputing system in the Supercomputing Center of Wuhan University." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.218, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.484, + 0.299 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.3, + 0.483, + 0.356 + ], + "angle": 0, + "content": "[2] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2567-2581, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.357, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[3] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.413, + 0.483, + 0.481 + ], + "angle": 0, + "content": "[4] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10021-10030, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.482, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[5] Shuhang Gu, Andreas Lugmayr, Martin Danelljan, Manuel Fritsche, Julien Lamour, and Radu Timofte. Div8k: Diverse 8k resolution image dataset. In IEEE International Conference on Computer Vision Workshop, pages 3512-3516, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[6] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[7] Zhaoyang Jia, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Generative latent coding for ultra-low bitrate image compression. In IEEE Conference on Computer Vision and Pattern Recognition, pages 26088-26098, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[8] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into conditional gans. arXiv preprint arXiv:2405.05967, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[9] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[10] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In IEEE International Conference on Computer Vision, pages 5148-5157, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[12] Heewon Kim, Myungsub Choi, Bee Lim, and Kyoung Mu Lee. Task-aware image downscaling. In European Conference on Computer Vision, pages 399-414, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[13] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[14] Jinmin Li, Tao Dai, Jingyun Zhang, Kang Liu, Jun Wang, Shaoming Wang, Shu-Tao Xia, et al. Boundary-aware decoupled flow networks for realistic extreme rescaling. arXiv preprint arXiv:2405.02941, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[15] Shang Li, Guixuan Zhang, Zhengxiong Luo, Jie Liu, Zhi Zeng, and Shuwu Zhang. Approaching the limit of image rescaling via flow guidance. arXiv preprint arXiv:2111.05133, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[16] Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing, 28(3): 1092-1107, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[17] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image superresolution and image rescaling. In IEEE International Conference on Computer Vision, pages 4076-4085, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[18] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[19] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diff-bir: Toward blind image restoration with generative diffusion prior. In European Conference on Computer Vision, pages 430-448, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[20] Siwei Ma, Xinfeng Zhang, Chuanmin Jia, Zhenghui Zhao, Shiqi Wang, and Shanshe Wang. Image and video compression with neural networks: A review. IEEE Transactions on Circuits and Systems for Video Technology, 30(6):1683-1698, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[21] Zhihong Pan, Baopu Li, Dongliang He, Mingde Yao, Henhao Wu, Tianwei Lin, Xin Li, and Errui Ding. Towards bidirectional arbitrary image rescaling: Joint optimization and cycle idempotence. In IEEE Conference on Computer Vision and Pattern Recognition, pages 17389-17398, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[22] Chenyang Qi, Xin Yang, Ka Leong Cheng, Ying-Cong Chen, and Qifeng Chen. Real-time 6k image rescaling with rate-distortion optimization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14092-14101, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[23] Lucas Relic, Roberto Azevedo, Markus Gross, and Christopher Schroers. Lossy image compression with foundation diffusion models. In European Conference on Computer Vision, pages 303-319, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15602" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[25] Wanjie Sun and Zhenzhong Chen. Learned image downscaling for upscaling using content adaptive resampler. IEEE Transactions on Image Processing, 29:4027-4040, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[26] George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Ballé, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[27] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[28] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[29] Luozhou Wang, Shuai Yang, Shu Liu, and Ying-cong Chen. Not all steps are created equal: Selective diffusion distillation for image manipulation. In IEEE International Conference on Computer Vision, pages 7472-7481, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[30] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision Workshops, pages 0–0, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[31] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In IEEE Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[32] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[33] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3365-3387, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[34] Hao Wei, Chenyang Ge, Zhiyuan Li, Xin Qiao, and Pengchao Deng. Towards extreme image rescaling with generative prior and invertible prior. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6181-6193, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[35] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. In Advances in Neural Information Processing Systems, pages 92529-92553, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[36] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 25456-25467, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Mingqing Xiao, Shuxin Zheng, Chang Liu, Zhouchen Lin, and Tie-Yan Liu. Invertible rescaling network and its ex" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "tensions. International Journal of Computer Vision, 131(1): 134-159, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[38] Bingna Xu, Yong Guo, Luoqian Jiang, Mianjie Yu, and Jian Chen. Downscaled representation matters: Improving image rescaling with collaborative downscaled images. In IEEE International Conference on Computer Vision, pages 12237-12247, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[39] Jinhai Yang, Mengxi Guo, Shijie Zhao, Junlin Li, and Li Zhang. Self-asymmetric invertible network for compression-aware image rescaling. In AAAI Conference on Artificial Intelligence, pages 3155-3163, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[40] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In European Conference on Computer Vision, pages 74-91, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[41] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. In Advances in Neural Information Processing Systems, pages 13294-13307, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.904, + 0.403 + ], + "angle": 0, + "content": "[42] Zongsheng Yue, Kang Liao, and Chen Change Loy. Arbitrary-steps image super-resolution via diffusion inversion. arXiv preprint arXiv:2412.09013, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[43] Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step image super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[44] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[45] Min Zhang, Zhihong Pan, Xin Zhou, and C-C Jay Kuo. Enhancing image rescaling using dual latent variables in invertible neural network. In ACM International Conference on Multimedia, pages 5602-5610, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[47] Zhixuan Zhong, Liangyu Chai, Yang Zhou, Bailin Deng, Jia Pan, and Shengfeng He. Faithful extreme rescaling via generative prior reciprocated invertible representations. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5708-5717, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[48] Yiming Zhu, Cairong Wang, Chenyu Dong, Ke Zhang, Hongyang Gao, and Chun Yuan. High-frequency normalizing flow for image rescaling. IEEE Transactions on Image Processing, 32:6223-6233, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15603" + } + ] +] \ No newline at end of file diff --git a/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/5b02e9a3-c733-4ab2-9aa6-549d5b3e9de2_origin.pdf b/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/5b02e9a3-c733-4ab2-9aa6-549d5b3e9de2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ec1527ddd92b1f1ffaf52dcaecd1c88a820d8c09 --- /dev/null +++ b/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/5b02e9a3-c733-4ab2-9aa6-549d5b3e9de2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d936a203622968a7228bcc2279cfabef260d20f79014123f00a393d7b3fe6685 +size 3991253 diff --git a/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/full.md b/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a4e08b1f6132c87a6426d95be3414e62f441acd2 --- /dev/null +++ b/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/full.md @@ -0,0 +1,393 @@ +# Timestep-Aware Diffusion Model for Extreme Image Rescaling + +Ce Wang, Zhenyu Hu, Wanjie Sun* Zhenzhong Chen Wuhan University + +{cewang, zhenyuhu, sunwanjie, zzchen}@whu.edu.cn + +# Abstract + +Image rescaling aims to learn the optimal low-resolution (LR) image that can be accurately reconstructed to its original high-resolution (HR) counterpart, providing an efficient image processing and storage method for ultra-high definition media. However, extreme downscaling factors pose significant challenges to the upscaling process due to its highly ill-posed nature, causing existing image rescaling methods to struggle in generating semantically correct structures and perceptual friendly textures. In this work, we propose a novel framework called Timestep-Aware Diffusion Model (TADM) for extreme image rescaling, which performs rescaling operations in the latent space of a pretrained autoencoder and effectively leverages powerful natural image priors learned by a pre-trained text-to-image diffusion model. Specifically, TADM adopts a pseudo-invertible module to establish the bidirectional mapping between the latent features of the HR image and the target-sized LR image. Then, the rescaled latent features are enhanced by a pre-trained diffusion model to generate more faithful details. Considering the spatially non-uniform degradation caused by the rescaling operation, we propose a novel time-step alignment strategy, which can adaptively allocate the generative capacity of the diffusion model based on the quality of the reconstructed latent features. Extensive experiments demonstrate the superiority of TADM over previous methods in both quantitative and qualitative evaluations. The code will be available at: https://github.com/wwangcece/TADM. + +# 1. Introduction + +With the explosive growth of ultra high-resolution (HR) images, image rescaling has become essential for enhancing data storage and transmission efficiency [39], optimizing resource usage in real-time applications [22], and ensuring proper display on devices with varying resolutions [21]. During image rescaling, HR images are downscaled + +to low-resolution (LR) images for storage, transmission, or display. When HR images are needed, an upscaling process is employed to restore the downscaled images to their original resolution. Unlike image super-resolution (SR), image rescaling allows for joint optimization of both downscaling and upscaling processes, preserving detailed information crucial for effective upscaling. + +Recently proposed image rescaling methods can be classified into two main categories: encoder-decoder architecture [12, 25] and invertible neural networks [17, 37]. The encoder-decoder architecture employs two neural networks to perform downscaling and upscaling. While, invertible neural networks view image rescaling as a reversible process, using the forward and backward passes of an invertible neural network to simulate the downscaling and upscaling. Despite the significant improvements in reconstruction quality achieved by these methods, they are limited to small rescaling factors, such as $2 \times$ or $4 \times$ . + +However, with the advancement of ultra HR imaging devices, there is an increasing need for extreme rescaling factors (e.g., $16 \times$ or $32 \times$ ) to achieve more efficient storage and transmission of ultra HR data [7, 47]. Existing image rescaling methods often struggle to generate sufficient detail under extreme rescaling factors. Therefore, some approaches propose using pre-trained GANs as priors to constrain the solution space of the reconstructed results. GRAIN [47] and BDFlow [14] use pre-trained StyleGAN [9, 10] to generate high-quality upscaled face images. VQIR [34] leverages the high-quality visual embeddings encapsulated in pre-trained VQGAN [3] to achieve extreme rescaling on natural images. Although these methods have made significant progress in perceptual metrics, their reconstruction results still suffer from incorrect semantics and structure. + +To address the aforementioned issues, we propose a novel framework named Timestep-Aware Diffusion Model (TADM) for extreme image rescaling, which performs image rescaling in the latent space of a pre-trained autoencoder and leverages the natural image priors stored in large text-to-image diffusion models to restore HR images. Specifically, first, the HR images are encoded by the via the VAE encoder to obtain the latent features. Then, we use an au + +toencoder to rescale the latent code of the HR image to the target size, and employ a set of invertible neural networks (INN) to perform the mapping from the feature space to the pixel space to obtain the LR image. Next, the rescaled features are fed into a pre-trained diffusion model to perform a single denoising step, thereby enhancing their perceptual quality. Finally, the enhanced features are decoded via the VAE decoder to obtain the rescaled image. Additionally, we find that the rescaled latent features exhibit non-uniform degradation, which can be interpreted as the addition of noise at varying levels. As a result, different images require different time-steps to perform the denoising process. Previous studies have shown that the time-step in diffusion models directly controls the generative capacity of the model [29]. Therefore, we propose a time-step alignment mechanism that predicts the appropriate timestep based on the reconstruction quality of the rescaled latent features, thereby adaptively tuning the generative capacity of the diffusion model. + +Our contributions can be summarized as follows: + +- We propose a Timestep-Aware Diffusion Model that performs well in global semantic and structural reconstruction for the extreme image rescaling task. +- We propose a novel decoupled strategy for feature rescaling. It decouples the generation of LR images from the rescaling operation, thereby improving the reconstruction accuracy of the rescaling process. +- We design a time-step alignment strategy to address the spatially non-uniform reconstruction quality of the rescaled latent features, allowing for more effective utilization of the prior information from the text-to-image diffusion model. + +# 2. Related Work + +# 2.1. Image Rescaling + +Image rescaling aims to downscale HR images to visually coherent LR images and then plausibly reconstruct the original HR images. Unlike image SR, image rescaling involves joint modeling of both downscaling and upscaling processes, leading to higher reconstruction accuracy. Earlier researches [12, 16] utilize CNNs to build an encoder-decoder framework, treating image rescaling as a unified task by jointly training the downscaling and upscaling processes. Instead of directly generating LR images, CAR [25] proposes generating content-adaptive resampling kernels based on HR images to perform the downsampling operation, avoiding pixel-level constraints on the downscaled images. To fully leverage the inherent reversibility of image rescaling, there has been a surge in the use of INNs for this task [15, 38, 45, 48]. IRN [37] is the first invertible framework that models image rescaling as a bijective transformation, embedding residual high-frequency (HF) components + +into a case-agnostic latent distribution for efficient reconstruction. HCFlow [17] assumes that the HF information depends on the LR image, thereby achieving better performance by incorporating LR conditions in HF information estimation. In extreme image rescaling, generative priors are usually employed to constrain the outcomes. GRAIN [47] and VQIR [34] leverage pre-trained StyleGAN [9, 10] and VQGAN [3], respectively, to achieve notable enhancements in perceptual quality. Nonetheless, these approaches are often limited in image contents, either confined to handling images from specific domains or encountering difficulties when reconstructing faces and textual elements. + +# 2.2. Diffusion Model for Image SR + +Here, we briefly introduce the application of diffusion models in a related task, i.e., image SR. Diffusion models promote the development of natural image SR in two main ways: training diffusion models from scratch [4, 13, 41] and leveraging pre-trained SD model [19, 36, 40]. The former is represented by SR3 [24], which achieves perceptual quality comparable to GANs. The latter is represented by StableSR [28], which aims to build a model capable of handling image restoration under any complex degradation in real-world scenarios. + +However, the above methods typically require dozens or even hundreds of iterations during inference, which significantly increases computational costs. Researchers have begun exploring how to achieve image SR using one-step diffusion models. SinSR [31], based on ResShift [41], derives a deterministic one-step sampling process and employs a consistency preserving loss to distill the prior knowledge from a multi-step teacher network into a one-step student network. A representative work, OSEDiff [35], directly feeds the LQ image into the SD model and employs variational score distillation for regularization. S3Diff [43] proposes a degradation-aware LoRA and utilizes a pre-trained DINO [44] model as a discriminator for adversarial training. InvSR [42] designs a partial noise prediction strategy to provide a dynamic starting sampling point and supports inference with an arbitrary number of sampling steps ranging from one to five. + +# 3. Methodology + +# 3.1. Overview of TADM + +As illustrated in Fig. 1, TADM consists of four parts: Latent Encoding, Feature Rescaling in the Latent Space, Denoising Guided Perception Enhancement and Latent Decoding. First, the input HR images $x$ are mapped to the latent features $z$ via the pre-trained VAE encoder $\mathcal{E}$ . Subsequently, the proposed Decoupled Feature Rescaling Module (DFRM) rescales $z$ to the target size, outputting LR images $y$ and the rescaled latent features $\hat{z}$ . Our aim is to make $\hat{z}$ + +![](images/20f560aec677f0e2202ba95b14927a29a9381aa962307d0d48974ecb76432715.jpg) + +![](images/dffab37f138dca29e215e6a3d9eb66e5ee2a8f067984774baa57e3286b203c6a.jpg) +Figure 1. Overview of the proposed Timestep-Aware Diffusion Model (TADM). First, the HR images $x$ are encoded to obtain the latent features $z$ . The latent features are then rescaled to the target size using the decoupled feature rescaling module (DFRM), producing the LR images $y$ and outputting the rescaled latent features $\hat{z}$ . Next, the rescaled latent features $\hat{z}$ are passed into a pre-trained diffusion model to perform a single denoising step for perceptual enhancement to obtain $\hat{z}_0$ . In this process, a time prediction module (TPM) is used to estimate the time step $t$ based on $\hat{z}$ and the predicted $t$ is then fed into both the U-Net and the time scheduler. Finally, the perceptually enhanced latent features $\hat{z}_0$ are decoded to obtain the rescaled image $\hat{x}$ . + +as close as possible to $z$ so that the rescaled results align with the latent space of the pre-trained VAE, while ensuring that $y$ and the original HR images $x$ share the same content. Then, the rescaled features $\hat{z}$ are fed into a denoising U-Net network to perform a single denoising step to get the perceptually enhanced latent features $\hat{z}_0$ . Finally, $\hat{z}_0$ is decoded via the pre-trained VAE decoder $\mathcal{D}$ to obtain the rescaled image $\hat{x}$ . Overall, the latent rescaling process can be regarded as a noise-adding process, while the latent enhancement process can be viewed as a denoising process. To dynamically estimate the 'noise' density introduced by latent rescaling, we design a time alignment strategy, which consists of a time-step prediction module (TPM) and a time scheduler module. It can better address the non-uniform degradation caused by rescaling, thereby enabling the dynamic allocation of generative capabilities of pre-trained SD. + +# 3.2. Feature Rescaling in the Latent Space + +# 3.2.1. Network Architecture. + +In this part, we perform rescaling of the HR latent features $z$ , obtaining LR images $y$ and rescaled features $\hat{z}$ . A straightforward idea is to construct a transformation chain as $y = G_{\mathrm{e}}(z)$ , $\hat{z} = G_{\mathrm{d}}(y)$ . Here, $G_{\mathrm{e}}$ is the encoder used + +to downscale the input features to the target sized LR images, while $G_{\mathrm{d}}$ represents the decoder used to reconstruct the original HR features. Guidance loss $\mathcal{L}_{\mathrm{gui}}$ and reconstruction loss $\mathcal{L}_{\mathrm{rec}}$ are applied to $y$ and $\hat{z}$ respectively, to force $y$ to have the same content as $x$ , and $\hat{z}$ to be as similar as possible to $z$ : + +$$ +\mathcal {L} _ {\text {g u i .}} = \| y - \operatorname {B i c u b i c} (x) \| _ {1} \tag {1} +$$ + +$$ +\mathcal {L} _ {\text {r e c .}} = \| \hat {z} - z \| _ {1} \tag {2} +$$ + +However, these two losses are contradictory to each other. When the guidance loss approaches zero, the rescaling problem degenerates into a classical SR problem [33], embedding little information about the downscaling process into the LR image. In contrast, when the reconstruction loss reaches its global optimum, the rescaling problem degenerates into an image compression problem [20], causing the LR image to lack meaningful visual content. Therefore, directly applying the guidance loss $\mathcal{L}_{\mathrm{gui}}$ on $y$ could adversely impact subsequent upscaling. + +Considering the above issues, we devise the DFRM as shown in Fig. 2, which divides the rescaling operation into two sub-tasks: feature rescaling and LR image generation. Inspired by FGRN [15], we use a set of CNNs to perform downscaling and upscaling, and a separate INN to + +![](images/1258a43e802c5300b8cc4c98aa94c29ee00708d7a8662f578ecd125fad9967a0.jpg) +Figure 2. Detailed architecture of Decoupled Feature Rescaling Module (DFRM). + +perform the bidirectional mapping between the feature domain and the pixel domain. To achieve a balance between feature rescaling accuracy and visual quality of LR images, we introduce another pixel guidance module in the encoder to incorporate pixel-level information, which can improve the quality of the LR image. Specifically, DFRM perform feature rescaling through two independent transformation chains $((x,z)\to z_{\mathrm{lr}}\to \hat{z})$ . $(z_{\mathrm{lr}}\leftrightarrow y)$ as follows: + +$$ +z _ {\mathrm {l r}} = G _ {\mathrm {e}} (x, z) \tag {3} +$$ + +$$ +z _ {\mathrm {l r}} \approx F ^ {- 1} (F (z _ {\mathrm {l r}})) = F ^ {- 1} (y) \tag {4} +$$ + +$$ +\hat {z} = G _ {\mathrm {d}} \left(z _ {\mathrm {l r}}\right) \tag {5} +$$ + +Here, $F$ and $F^{-1}$ represent the forward and backward processes of the INN, respectively. Given the latent variable $z \in \mathbb{R}^{\frac{H}{8} \times \frac{W}{8} \times 4}$ of the pre-trained SD VAE as input, the encoder $G_{\mathrm{e}}$ first transforms it into a compact representation $z_{\mathrm{lr}} \in \mathbb{R}^{\frac{H}{N} \times \frac{W}{N} \times 3}$ , whose size is the same as that of the target LR image and $N$ represents the target downscaling factor. Subsequently, a set of INNs perform the forward process to transform $z_{\mathrm{lr}}$ from the feature domain to the pixel domain and obtain the LR image $y \in \mathbb{R}^{\frac{H}{N} \times \frac{W}{N} \times 3}$ by quantization. In the upscaling process, the inverse of the INN is used to process the LR image $y$ and reconstruct $z_{\mathrm{lr}}$ . Finally, the decoder $G_{\mathrm{d}}$ is utilized to obtain the rescaled feature $\hat{z}$ . + +# 3.2.2. Loss Function. + +Training of this part involves two losses: reconstruction loss and guidance loss. Considering the information lost in quantization operations, the bidirectional mapping between $z_{\mathrm{lr}}$ and $y$ is not exactly reversible. To ensure the robustness of the network to quantization operations, we consider two types of transformation chains in designing the reconstruction loss: $(x,z)\to z_{\mathrm{lr}}\to \hat{z}$ and $(x,z)\to z_{\mathrm{lr}}\to y\to \hat{z}_{\mathrm{lr}}\to \hat{z}$ , which can be expressed as the following formula: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {r e c .}} = \| G _ {\mathrm {d}} \left(G _ {\mathrm {e}} (x, z)\right) - z \| _ {1} + \tag {6} \\ \left. \left\| G _ {\mathrm {d}} \left(F ^ {- 1} \left(F \left(G _ {\mathrm {e}} (x, z)\right)\right)\right) - z \right\| _ {1} \right. \\ \end{array} +$$ + +This combination can facilitate the network to reconstruct $\hat{z}$ with high fidelity in both the existence and absence of quantization operations. Additionally, the guidance loss is designed as follows: + +$$ +\mathcal {L} _ {\text {g u i .}} = \left\| F \left(G _ {\mathrm {e}} (x, z)\right) - \operatorname {B i c u b i c} (x) \right\| _ {1} \tag {7} +$$ + +Finally, the loss function of this part is a combination of these two losses: + +$$ +\mathcal {L} _ {\text {r e s .}} = \lambda_ {\text {r e c .}} \mathcal {L} _ {\text {r e c .}} + \lambda_ {\text {g u i .}} \mathcal {L} _ {\text {g u i .}} \tag {8} +$$ + +# 3.3. Denoising Guided Perception Enhancement + +The latent variable $\hat{z}$ obtained in Sec. 3.2 retains sufficient semantic information about the HR image, but lacks adequate texture. Therefore, we propose to leverage the strong priors stored in the pre-trained SD model to enhance the perceptual quality of the rescaling results in this part. Original diffusion model requires a multi-step sampling algorithm, significantly increases inference time and makes it difficult to scale to practical applications. Inspired by recent work on accelerating diffusion models [8], we adopt an one-step inference strategy. Specifically, we can directly feed the rescaled latent variable $\hat{z}$ into the denoising U-Net $\epsilon = \epsilon_{\theta}(\hat{z},t)$ and perform a single denoising step as follows: + +$$ +\hat {z} _ {0} = \frac {\hat {z} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon}{\sqrt {\bar {\alpha} _ {t}}} \tag {9} +$$ + +where $\hat{z}_0$ represents the perceptually enhanced latent features, $\bar{\alpha}_{t}$ is the noise scheduler of the forward diffusion process of the SD model, and $t$ represents the time-step associated with the noise intensity. Finally, $\hat{z}_0$ can be transformed to the pixel domain through the pre-trained VAE decoder $\mathcal{D}$ to obtain the upscaled image $\hat{x} = \mathcal{D}(\hat{z}_0)$ . Then, we can adopt the loss function between $\hat{x}$ and $x$ to optimize the U-Net using LoRA layers. + +However, the key problem of the aforementioned approaches lies in the selection of time-step $t$ . We can regard the rescaling process as the noise-adding process of diffusion model, while Eq. (9) represents the denoising process. For different rescaling factors and different images, the 'noise' introduced by the rescaling operation varies in intensity. In Fig. 3, we present the global rescaling error of an image under different rescaling factors, as well as the local rescaling error of different patches within the image. Additionally, we plot the error curve of the forward noise-adding process of the SD model. It can be observed that as the rescaling factor and image content change, rescaling MSE also varies and corresponds to the diffusion MSE at different time-steps. Thus, we propose to dynamically predict the time-step $t$ based on the quality of the rescaled latent features $\hat{z}$ . A simple yet effective network named timestep prediction module (TPM), whose structure is presented in Fig. 1, is used to predict $t$ based on the rescaled latent features: $t = \mathrm{TPM}(\hat{z})$ . During the training phase, we compute a single time-step for each input image patch. In the inference phase, we adopt a tiled inference approach (which will be elaborated in detail in the supplementary materials), allowing for predicting different time-steps for various regions of a high-resolution input image. + +![](images/1b8b3468943aab65fa6153d8e81a81c03b422dbd2e378353f4c35c2984780218.jpg) +Figure 3. We compare the MSE introduced by the rescaling operation with that caused by the forward diffusion process. The rescaling degradation severity is correlated with both the image content and the rescaling scale. Therefore, the model requires adaptive time-step to align the rescaling MSE with the diffusion MSE. + +![](images/151175251986af6e47c306ecd577e79aeb9bce6dfba03ffba3970274e61c1362.jpg) +Figure 4. The structure of the time scheduler module. It combines both the fixed scheduler and learnable scheduler to perform the denoising process. The two schedulers are connected by a zero-convolution layer to stabilize the training. + +Subsequently, the predicted time-step $t$ is used to perform a single denoising step. However, the vanilla SD time scheduler as shown in Eq. (9) is not differentiable with respect to $t$ , which prevents gradients from being backpropagated to the TPM. Considering that given $\hat{z}$ , $\epsilon$ and $t$ , Eq. (9) essentially represents a linear transformation, which can be learned by a neural network. Therefore, we propose a hybrid time scheduler module consisting of a fixed scheduler and a learnable scheduler, whose structure is presented in Fig. 4. The former uses Eq. (9) to perform the denoising step, while the latter employs a neural network to simulate the denoising scheduler. To ensure training stability, the final layer of the learnable scheduler is a zero-initialized convolution, allowing it to gradually correct the results of the fixed scheduler as training progresses. This process can be described as follows: + +$$ +\hat {z} _ {0} = \mathcal {S} _ {\text {f i x e d}} (\hat {z}, \epsilon , t _ {0}) + \mathcal {S} _ {\text {l e a r n e d}} (\hat {z}, \epsilon , t) \tag {10} +$$ + +where $t_0$ is the fixed pre-set time-step and $S_{\mathrm{fixed}}$ denotes the vanilla SD time scheduler as shown in Eq. (9). + +Finally, we fine-tune the denoising U-Net network and + +the VAE decoder using LoRA, while also jointly training the TPM and time scheduler module. To achieve a balance between fidelity and realism, we supervise the reconstruction results $\hat{x}$ using pixel loss and perceptual loss as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {e n h .}} = \| \hat {x} - x \| _ {1} + \tag {11} \\ \lambda_ {\text {p e c .}} \left(\mathcal {L} _ {\text {l p i p s}} (\hat {x}, x) + \mathcal {L} _ {\text {d i s t s}} (\hat {x}, x)\right) \\ \end{array} +$$ + +# 4. Experiments + +# 4.1. Experimental Setup + +# 4.1.1. Datasets and Evaluation Metrics. + +We train the proposed TADM using the DF2K dataset [1, 18]. Considering that traditional benchmarks such as Set5, Set14, and BSD100 often have lower quality and resolution [5], they are not suitable for evaluating extreme rescaling methods. Therefore, we utilize four high-resolution datasets: DIV2K, CLIC2020 [26], Urban100 [6], and DIV8K [5] to evaluate models been compared (results on the latter two datasets are presented in the supplementary materials). We adopt PSNR and SSIM [32] for quantitative evaluation. For perceptual quality evaluation, reference-based metrics LPIPS [46] and DISTS [2] are used. For non-reference image quality assessment, we use metrics including MUSIQ [11] and CLIPIQA [27]. + +# 4.1.2. Training Details + +TADM is built based on the SD 2.1-base model, and the training is divided into three stages. First, we train the DFRM using Eq. (8). Then we adopt Eq. (11) to jointly train the LoRA layers, the TPM and the time scheduler module. Finally, we use a smaller learning rate to jointly fine-tune DFRM, LoRA layers, TPM and scheduler module using a combination of Eq. (8) and (11). More training details are provided in the supplementary material. + +# 4.2. Results and Comparison with SOTA + +We compare the proposed TADM with three kinds of methods: (1) bicubic downscaling and SR, i.e., ESRGAN [30], StableSR [28] (based on multi-step diffusion) and S3Diff (based on single-step diffusion)[43]; (2) flow-based rescaling models, i.e., IRN [37] and HCFlow [17]; (3) prior-based extreme rescaling method VQIR [34]. Some of them are proposed for $4 \times$ rescaling factors and we extended them to $16 \times$ or $32 \times$ . To ensure a fair comparison, all methods are trained until convergence on the DF2K dataset following their original configurations. + +# 4.2.1. Quantitative Results + +Table 1 presents the $16\times$ and $32\times$ quantitative image rescaling results on four datasets. The performance of SR methods are generally poor. Both gan-based (ESRGAN) and diffusion-based (StableSR and S3Diff) methods can only achieve relative good performance on no-reference metrics. + +Table 1. Quantitative comparisons with different methods at ${16} \times$ and ${32} \times$ (results on Urban100 and DIV8K are presented in the supplementary materials). The symbols $\uparrow$ and $\downarrow$ respectively represent that higher or lower values indicate better performance. Bold represents the best and underline represents the second best. + +
DatasetMethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓MUSIQ ↑CLIPIQA ↑
16×32×16×32×16×32×16×32×16×32×16×32×
DIV2KESRGAN23.1519.940.59460.56890.44780.58920.23780.423059.8043.350.61610.3734
StableSR22.0219.920.59830.55880.49800.57360.20780.260351.2848.940.37800.3203
S3Diff20.2217.810.54910.46010.40330.48950.13090.181064.3767.920.62280.6991
IRN26.0422.840.70200.62120.51530.60950.30410.403643.7228.700.27990.2336
HCFlow26.6623.890.71760.64670.48850.58160.28660.385246.4337.250.27350.2792
VQIR23.9122.020.64980.58230.31740.45680.10240.266364.0458.210.63500.6293
Ours23.9822.180.67360.58690.29790.42210.08860.168466.5669.120.71890.7204
CLIC2020ESRGAN25.4322.470.70180.68040.42130.54290.22090.401157.8041.280.61070.3558
StableSR24.1122.240.70520.67260.46250.52880.18600.228252.6949.800.37790.3325
S3Diff22.0419.870.66150.57110.39680.48290.13360.181263.0766.400.60010.6838
IRN28.5524.930.79070.72280.46760.55280.27460.377647.6627.810.29500.2495
HCFlow28.8226.210.79490.74510.45420.53250.26350.361147.7237.150.27330.2884
VQIR25.6924.120.73630.69420.28540.41910.09010.244564.3654.830.61050.6068
Ours26.0424.260.75420.69030.27240.39110.07920.151365.3569.100.68150.7271
+ +This is because they rely solely on the LR image to reconstruct the original image and cannot jointly optimize the downscaling and upscaling processes. While methods based on INNs (HCFlow and IRN) achieve the highest PSNR and SSIM due to its consideration of the correlation between downscaling and upscaling. In all generative methods, VQIR achieves relatively good perceptual metrics by leveraging the priors of the pre-trained VQGAN. In contrast, our method achieves optimal performance across all perceptual metrics on all datasets. For instance, in the $32 \times$ rescaling on the DIV2K validation set, our method improves the DISTS metric by $36.76\%$ compared to the second-best VQIR. Importantly, our approach is designed for applications like social media delivery, where semantic fidelity and perceptual realism are more critical. In such cases, metrics like LPIPS and DISTS better reflect image quality [23]. + +# 4.2.2. Qualitative Results + +In Fig. 5, we present visual comparisons at $16 \times$ and $32 \times$ rescaling factors. Due to the ill-posed nature of extreme rescaling, pixel-oriented optimization methods such as IRN, and HCFlow produce overly-smoothed results. Although ESRGAN incorporates perceptual optimization it still cannot generate adequate details. VQIR leverages the pretrained VQGAN as a prior, allowing it to generate rich details and sharp edges. However, it struggles with faces and more complex structures. Although S3Diff can generate sharp edges and rich details, its reconstructed results have low fidelity and exhibit significant artifacts and color distortions. Thanks to the powerful prior provided by the pre-trained SD and the carefully designed architecture,, our method effectively addresses these issues. It not only gener + +ates sufficiently rich textures but also reconstructs accuracy semantics and regular structures, such as text, faces, vehicles, and leaves. + +# 4.3. Ablation Study + +We conduct ablation experiments to validate the effectiveness of each component. All ablations are trained on the DF2K dataset and tested on the validation set of the DIV2K for $16 \times$ rescaling task. + +# 4.3.1. Effectiveness of Feature Rescaling + +To verify the effectiveness of the feature rescaling operation, we move the rescaling operation from the latent space to the pixel space. Specifically, we use HCFlow [17], the comparative method with the best PSNR performance, to perform the rescaling operation in the pixel domain, followed by perceptual enhancement using the same approach described in Sec. 3.3. As shown the 1st row in Table 2, the rescaling operation in the latent space can improve reconstruction performance, as evidenced by much lower LPIPS and DISTS. Additional details and results are presented in the supplementary materials. + +Table 2. Ablation study of rescaling space and SD prior. + +
Latent spacePixel spaceSD priorLPIPS ↓DISTS ↓
×✓✓0.36300.1154
✓××0.46750.3109
✓×✓0.29790.0886
+ +# 4.3.2. Necessity of SD Prior + +To validate that SD model provides rich priors about real-world HR images, we directly omit the perceptual enhancement part. As shown in Table 2 (2nd row), the introduction + +![](images/dd4ddc584944e372d5db7496c087e1745d12376f4e726b92dcdbb857fdf82d0e.jpg) + +![](images/3ad96ea359717441ed6b97350ac3e3e7d239ed570350e582011de0a70f6c02f6.jpg) + +![](images/736ea63ecf9ac2858c6dbe394a4b539e83d3d845dc5d1f275e68e346e4756525.jpg) + +![](images/a528fdf9259a50289bf98dc1ee7f8704c1dcc188e435c4975b945dbcaddf813a.jpg) +(a) StableSR + +![](images/f21437c8c9194ba8527af1d16b2c1aad498338ffb8e7c457be5593ca691a4620.jpg) + +![](images/f192fb0a0ce666244f33b1c476f3c57ef5db36b9869194dc90449b99c9bed680.jpg) + +![](images/ec99751125d639c4342048c63d38e32ecd36f0712f709c016003797bfb635bdf.jpg) + +![](images/40480862e59a7bc2154612944e23626525d056aa94f6688975f88bcf76b12d3e.jpg) +(b) S3Diff + +![](images/ce7e424750f685b1e9c7fdc635fbf278bba6abe447500e7e7e9bdef55c50ed61.jpg) + +![](images/c0e870a5e34d9d4b162d2cc0a77664cfe0f80e682e46d5061a82fe9006d2003a.jpg) + +![](images/75fff6b0e603a820472cc8ab0c24b0b4511190d136b7d88cd1cd3daf86fa54ba.jpg) + +![](images/2e8c024997cf5bcc4672a14ce78c1b6d5c3d4e50c668f37d615a4aecb85e0e5a.jpg) +(c)ESRGAN + +![](images/ff32ee53fec9d1c2fffa6c6e92789b67184aae4e33ed5f6e51c663c127cec4f1.jpg) + +![](images/1d7a76e49bfc44da7c92f83cf97ff6f3ae70bb234824947c3e841ddec9461a5b.jpg) + +![](images/1555b512a06c3b435d04bde73b0f7450fadf6ec84e2c98f5e728359fe0a1adab.jpg) + +![](images/0224e89361a889d0b280f64e1e88a0f3750f236c8872ad13250582f078d8fac4.jpg) + +![](images/c722e92a09990f1d30c2e50ef1e560ba8338762fc8ee3849fb5fc43228aef23b.jpg) + +![](images/b9c133f8e1dcfab98e86580b5fce12b436319dd8f1d90c57a9a0d398692c5a2f.jpg) + +![](images/68e57e783fd83330b93f324cfe085b9130b9d13367d09244b26d4e727d7b6412.jpg) + +![](images/1be1b4abf1fb7b7801e69b69c925ceaf6e86950729c8432674164e36d44647ae.jpg) +(d) IRN + +![](images/8d7b6aab8d3d937a2a1762e9951702b989d46df1a6dd7e0f0278b352910a8463.jpg) + +![](images/fc21f908646bd65b2d004886f51b93615c6cff9db8f55252296676e49d37da14.jpg) + +![](images/1005b8308246bc87729221cdd4025084f83e4121fda414d86ae1467fe7467ed0.jpg) + +![](images/05a06bd329b41e86ca5e52de11dd852a67b150c6f7839b592dce35404b1c8325.jpg) +(e) HCFlow + +![](images/3009741a3727e4f37d7b124c161837c2907bc01464ce4fa0e7f24f3cc708bd30.jpg) + +![](images/d88d75630eb5fcdab0e3240428056c0451d037d8036f9bca63b6db0633f7e7e6.jpg) + +![](images/aecbabf27b17e5d3b8ade4b49cb218b54f363c93e43b35dbf6de95d6c129ad96.jpg) + +![](images/2ed0ceeabb474c54a0036586d0201ef594251d5fddea50ed4bcb8325072e0634.jpg) +(f) VQIR + +![](images/5127534e08da84053f78292987e6df2a887351d853a1db216f01c8ce402a3898.jpg) + +![](images/bdf4734a648115027bd2e98154f6476b394374c83d6073ea9b033d1b1888c0b9.jpg) + +![](images/95cb22709bb6411cd8b818ed25e8e70d4609198834788e83093ede0df8580e94.jpg) + +![](images/a52781ffb42af446eabbe24d887b259d1a241569c777e7087e6e144bf1c34e62.jpg) +(g) Ours + +![](images/f399334587a6bd0fe7c8bde7430e45ec43bcabe23f6c1fbe9609f1b79dac0765.jpg) + +![](images/db04ff57706227217ee1d6f47cb7266816b0ba4f085fd9e6006bb2474e6da60a.jpg) + +![](images/6d39d8aaad51264f7728292b9edff9f5e396d3edb9b2690a01e1d2a40339f9a7.jpg) + +![](images/a9744568492a477fd40883c68d35c53e14d0f492aa7d70b30af99eb9e734b291.jpg) +(h) HR + +![](images/1fdf82630b05f0f11338afd1c89209e58becfdbed9239da1c0270373597491b6.jpg) +Figure 5. Qualitative comparisons of $16 \times$ (the 1st and 2nd rows) and $32 \times$ (the 3rd and 4th rows) rescaling results. Our method achieves higher semantic accuracy, such as more clearer text, more recognizable faces, and more realistic structures. + +![](images/136107dc95f13832cf2213ed1751495fecff08233853d76d0167ef3061ecb0f6.jpg) +Figure 6. The validation loss curves when removing a specific component from the DFRM. + +of the SD priors can significantly improve performance of LPIPS and DISTS and greatly enhance the perceptual quality of the reconstruction results. + +# 4.3.3. Effectiveness of DFRM Design + +Our DFRM includes two specific designs: the domain converter module and the pixel guidance module. To analyze their impact, we remove each component individually and plot the validation loss curves, as shown in Fig. 6. It can be observed that the introduction of the INN can reduce reconstruction error. Although the pixel guidance module does not improve reconstruction performance, our experiments show that it can enhance the fidelity and visual quality of the LR images, as demonstrated by the lower guidance loss in Fig. 6. Therefore, the design of our DFRM strikes a good balance between feature reconstruction and LR image quality. More experiments and the visual quality comparisons of LR images are presented in the supplementary materials. + +![](images/12d6e60aab704f14b0c01532d5e8675ad64b812228a3e09ef474ec91432817d4.jpg) +Figure 7. Ablation study about time alignment. + +# 4.3.4. Effectiveness of Time Alignment + +To validate the effectiveness of our proposed time-step alignment strategy, we retrain our model using fixed timesteps. This involves removing the TPM and time scheduler modules and employing a manually selected time-step to perform a single denoising step based on Eq. (9). Following the setting of previous work [35], we select fixed time-steps 1 and 999 for comparison. As shown in Fig. 7, when the time-step is set to a smaller value, the model achieves higher fidelity but lower perceptual quality, and the situation is reversed when a larger time-step is chosen. In contrast, our time-step alignment mechanism can dynamically determine the level of generative capability based on the image content, thereby achieving optimal performance in both fidelity and perceptual quality. + +In Fig. 8, we present a schematic illustration of the tiled inference process along with the predicted time-step map. + +![](images/7407938f94e679e96e92d8d91be4971cc70e9c2a24382da3d43da75011d6ed5e.jpg) +Figure 8. Illustration of the tiled inference process and the predicted time-step map. + +It can be observed that the time alignment mechanism effectively captures more complex regions in high-resolution images and assigns larger time steps to these areas. Details regarding the selection of patch size and stride length are provided in the supplementary materials. + +# 4.4. Discussions + +# 4.4.1. Application to Image Compression + +Our model can be combined with existing lossless image compression methods to achieve extreme compression ratios. To validate this, we use PNG to compress LR images and compare it with JPEG at various bit-rates. As shown in Fig. 9, at the same bits per pixel (bpp), the image reconstruction quality of our model significantly surpasses that of the JPEG compression. These comparisons demonstrate the potential of our model for image compression. + +![](images/9fe5368702fcc26077d80f5fbcd88347ea06463541624be324166e00968d5859.jpg) +Figure 9. R-D curves of JPEG compression and our proposed model on the DIV2K validation set. + +# 4.4.2. Inference Efficiency Comparison + +In Table 3, we provide comparisons of inference time and memory consumption for representative methods. All tests are performed on an RTX 3090 using images with a resolution of $1024 \times 1024$ . It can be observed that flow-based models (IRN and HCFlow) require less GPU memory consumption and enable fast inference. In contrast, StableSR, which is based on multi-step diffusion models, demands + +significantly more inference time and larger GPU memory consumption. Although VQIR requires less memory, the nearest neighbor search process in VQGAN results in longer inference time. For one-step diffusion-based method S3Diff, due to its degradation-guided LoRA which needs to dynamically predict LoRA parameters during inference, it requires longer ingerence time and higher memory consumption. In contrast, our proposed TADM demonstrates moderate efficiency and scalability compared to other methods. In the future, we can utilize distillation techniques to further reduce the model size. + +Table 3. Comparison of inference time and memory consumption. + +
IRNHCFlowVQIRStableSRS3DiffOurs
Inference Time (s)0.3590.3630.897135.22.1860.711
Memory Consumption (G)0.4980.0980.3446.3805.0252.598
+ +# 4.4.3. Upper Bound of TADM + +Since our method performs the rescaling operation in the latent space of a pre-trained VAE, its performance upper bound is inherently constrained by the VAE. In Table 4, we present a performance comparison between VAE and two $4 \times$ rescaling models. For VAEs, we directly feed the high-resolution images to the encoder and decoder without performing any rescaling operations. This performance represents the upper bound of our method's performance across all rescaling factors. It can be observed that VAE significantly lags behind the latter. In the future, we could explore integrating stronger VAEs, such as those from SDXL, SD3, thereby enabling the TADM to be more applicable to small-scale rescaling tasks. + +Table 4. Comparison of VAE and $4 \times$ rescaling methods. + +
IRNHCFlowVAE
PSNR35.0735.2329.01
SSIM0.93180.93460.8046
+ +# 5. Conclusion + +We propose a novel framework called TADM to address issues such as insufficient texture and semantic inaccuracies in extreme image rescaling. Specifically, we design a DFRM to achieve bidirectional mapping between HR features and LR images. Then, we utilize a pre-trained SD model to perform a single denoising step on the rescaled features, enhancing its perceptual quality. Considering the non-uniform quality of the rescaled latent features, we propose a novel time-step alignment strategy to achieve a balance between fidelity and perceptual quality. Both qualitative and quantitative experiments demonstrate the superiority of our approach. + +# Acknowledgments + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5 +[2] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2567-2581, 2020. 5 +[3] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021. 1, 2 +[4] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10021-10030, 2023. 2 +[5] Shuhang Gu, Andreas Lugmayr, Martin Danelljan, Manuel Fritsche, Julien Lamour, and Radu Timofte. Div8k: Diverse 8k resolution image dataset. In IEEE International Conference on Computer Vision Workshop, pages 3512-3516, 2019. 5 +[6] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 5 +[7] Zhaoyang Jia, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Generative latent coding for ultra-low bitrate image compression. In IEEE Conference on Computer Vision and Pattern Recognition, pages 26088-26098, 2024. 1 +[8] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into conditional gans. arXiv preprint arXiv:2405.05967, 2024. 4 +[9] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 1, 2 +[10] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1, 2 +[11] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In IEEE International Conference on Computer Vision, pages 5148-5157, 2021. 5 + +This work was supported in part by the National Natural Science Foundation of China under Grant 62201403, in part by Hubei Provincial Natural Science Foundation of China under Grant 2024AFB549. The numerical calculations in this paper have been done on the supercomputing system in the Supercomputing Center of Wuhan University. +[12] Heewon Kim, Myungsub Choi, Bee Lim, and Kyoung Mu Lee. Task-aware image downscaling. In European Conference on Computer Vision, pages 399-414, 2018. 1, 2 +[13] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2 +[14] Jinmin Li, Tao Dai, Jingyun Zhang, Kang Liu, Jun Wang, Shaoming Wang, Shu-Tao Xia, et al. Boundary-aware decoupled flow networks for realistic extreme rescaling. arXiv preprint arXiv:2405.02941, 2024. 1 +[15] Shang Li, Guixuan Zhang, Zhengxiong Luo, Jie Liu, Zhi Zeng, and Shuwu Zhang. Approaching the limit of image rescaling via flow guidance. arXiv preprint arXiv:2111.05133, 2021. 2, 3 +[16] Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing, 28(3): 1092-1107, 2018. 2 +[17] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image superresolution and image rescaling. In IEEE International Conference on Computer Vision, pages 4076-4085, 2021. 1, 2, 5, 6 +[18] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 5 +[19] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diff-bir: Toward blind image restoration with generative diffusion prior. In European Conference on Computer Vision, pages 430-448, 2024. 2 +[20] Siwei Ma, Xinfeng Zhang, Chuanmin Jia, Zhenghui Zhao, Shiqi Wang, and Shanshe Wang. Image and video compression with neural networks: A review. IEEE Transactions on Circuits and Systems for Video Technology, 30(6):1683-1698, 2019. 3 +[21] Zhihong Pan, Baopu Li, Dongliang He, Mingde Yao, Henhao Wu, Tianwei Lin, Xin Li, and Errui Ding. Towards bidirectional arbitrary image rescaling: Joint optimization and cycle idempotence. In IEEE Conference on Computer Vision and Pattern Recognition, pages 17389-17398, 2022. 1 +[22] Chenyang Qi, Xin Yang, Ka Leong Cheng, Ying-Cong Chen, and Qifeng Chen. Real-time 6k image rescaling with rate-distortion optimization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14092-14101, 2023. 1 +[23] Lucas Relic, Roberto Azevedo, Markus Gross, and Christopher Schroers. Lossy image compression with foundation diffusion models. In European Conference on Computer Vision, pages 303-319, 2024. 6 +[24] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions + +on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2 +[25] Wanjie Sun and Zhenzhong Chen. Learned image downscaling for upscaling using content adaptive resampler. IEEE Transactions on Image Processing, 29:4027-4040, 2020. 1, 2 +[26] George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Ballé, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020. 5 +[27] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5 +[28] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 5 +[29] Luozhou Wang, Shuai Yang, Shu Liu, and Ying-cong Chen. Not all steps are created equal: Selective diffusion distillation for image manipulation. In IEEE International Conference on Computer Vision, pages 7472-7481, 2023. 2 +[30] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision Workshops, pages 0–0, 2018. 5 +[31] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In IEEE Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 2 +[32] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5 +[33] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3365-3387, 2020. 3 +[34] Hao Wei, Chenyang Ge, Zhiyuan Li, Xin Qiao, and Pengchao Deng. Towards extreme image rescaling with generative prior and invertible prior. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6181-6193, 2024. 1, 2, 5 +[35] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. In Advances in Neural Information Processing Systems, pages 92529-92553, 2024. 2, 7 +[36] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 25456-25467, 2024. 2 +[37] Mingqing Xiao, Shuxin Zheng, Chang Liu, Zhouchen Lin, and Tie-Yan Liu. Invertible rescaling network and its ex + +tensions. International Journal of Computer Vision, 131(1): 134-159, 2023. 1, 2, 5 +[38] Bingna Xu, Yong Guo, Luoqian Jiang, Mianjie Yu, and Jian Chen. Downscaled representation matters: Improving image rescaling with collaborative downscaled images. In IEEE International Conference on Computer Vision, pages 12237-12247, 2023. 2 +[39] Jinhai Yang, Mengxi Guo, Shijie Zhao, Junlin Li, and Li Zhang. Self-asymmetric invertible network for compression-aware image rescaling. In AAAI Conference on Artificial Intelligence, pages 3155-3163, 2023. 1 +[40] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In European Conference on Computer Vision, pages 74-91, 2024. 2 +[41] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. In Advances in Neural Information Processing Systems, pages 13294-13307, 2023. 2 +[42] Zongsheng Yue, Kang Liao, and Chen Change Loy. Arbitrary-steps image super-resolution via diffusion inversion. arXiv preprint arXiv:2412.09013, 2024. 2 +[43] Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step image super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 2, 5 +[44] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2 +[45] Min Zhang, Zhihong Pan, Xin Zhou, and C-C Jay Kuo. Enhancing image rescaling using dual latent variables in invertible neural network. In ACM International Conference on Multimedia, pages 5602-5610, 2022. 2 +[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5 +[47] Zhixuan Zhong, Liangyu Chai, Yang Zhou, Bailin Deng, Jia Pan, and Shengfeng He. Faithful extreme rescaling via generative prior reciprocated invertible representations. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5708-5717, 2022. 1, 2 +[48] Yiming Zhu, Cairong Wang, Chenyu Dong, Ke Zhang, Hongyang Gao, and Chun Yuan. High-frequency normalizing flow for image rescaling. IEEE Transactions on Image Processing, 32:6223-6233, 2022. 2 \ No newline at end of file diff --git a/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/images.zip b/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..55c6762f17c54c07e22f1d1ea03ee7944a312140 --- /dev/null +++ b/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:115fd18fc1b554fa9cec0be66a388d7eea1816973ce2f07ac702cadca93103d9 +size 656251 diff --git a/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/layout.json b/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b060481dad74081afbf039779336648f253bf0cc --- /dev/null +++ b/2025/Timestep-Aware Diffusion Model for Extreme Image Rescaling/layout.json @@ -0,0 +1,10627 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 500, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 500, + 121 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 500, + 121 + ], + "type": "text", + "content": "Timestep-Aware Diffusion Model for Extreme Image Rescaling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 143, + 432, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 143, + 432, + 171 + ], + "spans": [ + { + "bbox": [ + 175, + 143, + 432, + 171 + ], + "type": "text", + "content": "Ce Wang, Zhenyu Hu, Wanjie Sun* Zhenzhong Chen Wuhan University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 173, + 173, + 432, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 173, + 432, + 185 + ], + "spans": [ + { + "bbox": [ + 173, + 173, + 432, + 185 + ], + "type": "text", + "content": "{cewang, zhenyuhu, sunwanjie, zzchen}@whu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 239, + 294, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 574 + ], + "type": "text", + "content": "Image rescaling aims to learn the optimal low-resolution (LR) image that can be accurately reconstructed to its original high-resolution (HR) counterpart, providing an efficient image processing and storage method for ultra-high definition media. However, extreme downscaling factors pose significant challenges to the upscaling process due to its highly ill-posed nature, causing existing image rescaling methods to struggle in generating semantically correct structures and perceptual friendly textures. In this work, we propose a novel framework called Timestep-Aware Diffusion Model (TADM) for extreme image rescaling, which performs rescaling operations in the latent space of a pretrained autoencoder and effectively leverages powerful natural image priors learned by a pre-trained text-to-image diffusion model. Specifically, TADM adopts a pseudo-invertible module to establish the bidirectional mapping between the latent features of the HR image and the target-sized LR image. Then, the rescaled latent features are enhanced by a pre-trained diffusion model to generate more faithful details. Considering the spatially non-uniform degradation caused by the rescaling operation, we propose a novel time-step alignment strategy, which can adaptively allocate the generative capacity of the diffusion model based on the quality of the reconstructed latent features. Extensive experiments demonstrate the superiority of TADM over previous methods in both quantitative and qualitative evaluations. The code will be available at: https://github.com/wwangcece/TADM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 599, + 135, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 599, + 135, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 135, + 611 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 620, + 294, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 294, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 294, + 693 + ], + "type": "text", + "content": "With the explosive growth of ultra high-resolution (HR) images, image rescaling has become essential for enhancing data storage and transmission efficiency [39], optimizing resource usage in real-time applications [22], and ensuring proper display on devices with varying resolutions [21]. During image rescaling, HR images are downscaled" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 299 + ], + "type": "text", + "content": "to low-resolution (LR) images for storage, transmission, or display. When HR images are needed, an upscaling process is employed to restore the downscaled images to their original resolution. Unlike image super-resolution (SR), image rescaling allows for joint optimization of both downscaling and upscaling processes, preserving detailed information crucial for effective upscaling." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 300, + 553, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 432 + ], + "type": "text", + "content": "Recently proposed image rescaling methods can be classified into two main categories: encoder-decoder architecture [12, 25] and invertible neural networks [17, 37]. The encoder-decoder architecture employs two neural networks to perform downscaling and upscaling. While, invertible neural networks view image rescaling as a reversible process, using the forward and backward passes of an invertible neural network to simulate the downscaling and upscaling. Despite the significant improvements in reconstruction quality achieved by these methods, they are limited to small rescaling factors, such as " + }, + { + "bbox": [ + 313, + 300, + 553, + 432 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 300, + 553, + 432 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 300, + 553, + 432 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 300, + 553, + 432 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 435, + 553, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 553, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 553, + 615 + ], + "type": "text", + "content": "However, with the advancement of ultra HR imaging devices, there is an increasing need for extreme rescaling factors (e.g., " + }, + { + "bbox": [ + 313, + 435, + 553, + 615 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 313, + 435, + 553, + 615 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 435, + 553, + 615 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 313, + 435, + 553, + 615 + ], + "type": "text", + "content": ") to achieve more efficient storage and transmission of ultra HR data [7, 47]. Existing image rescaling methods often struggle to generate sufficient detail under extreme rescaling factors. Therefore, some approaches propose using pre-trained GANs as priors to constrain the solution space of the reconstructed results. GRAIN [47] and BDFlow [14] use pre-trained StyleGAN [9, 10] to generate high-quality upscaled face images. VQIR [34] leverages the high-quality visual embeddings encapsulated in pre-trained VQGAN [3] to achieve extreme rescaling on natural images. Although these methods have made significant progress in perceptual metrics, their reconstruction results still suffer from incorrect semantics and structure." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "To address the aforementioned issues, we propose a novel framework named Timestep-Aware Diffusion Model (TADM) for extreme image rescaling, which performs image rescaling in the latent space of a pre-trained autoencoder and leverages the natural image priors stored in large text-to-image diffusion models to restore HR images. Specifically, first, the HR images are encoded by the via the VAE encoder to obtain the latent features. Then, we use an au" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15594" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 299 + ], + "type": "text", + "content": "toencoder to rescale the latent code of the HR image to the target size, and employ a set of invertible neural networks (INN) to perform the mapping from the feature space to the pixel space to obtain the LR image. Next, the rescaled features are fed into a pre-trained diffusion model to perform a single denoising step, thereby enhancing their perceptual quality. Finally, the enhanced features are decoded via the VAE decoder to obtain the rescaled image. Additionally, we find that the rescaled latent features exhibit non-uniform degradation, which can be interpreted as the addition of noise at varying levels. As a result, different images require different time-steps to perform the denoising process. Previous studies have shown that the time-step in diffusion models directly controls the generative capacity of the model [29]. Therefore, we propose a time-step alignment mechanism that predicts the appropriate timestep based on the reconstruction quality of the rescaled latent features, thereby adaptively tuning the generative capacity of the diffusion model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 300, + 266, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 300, + 266, + 312 + ], + "spans": [ + { + "bbox": [ + 67, + 300, + 266, + 312 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 314, + 295, + 456 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 314, + 295, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 295, + 348 + ], + "type": "text", + "content": "- We propose a Timestep-Aware Diffusion Model that performs well in global semantic and structural reconstruction for the extreme image rescaling task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 350, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 396 + ], + "type": "text", + "content": "- We propose a novel decoupled strategy for feature rescaling. It decouples the generation of LR images from the rescaling operation, thereby improving the reconstruction accuracy of the rescaling process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 397, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 397, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 56, + 397, + 295, + 456 + ], + "type": "text", + "content": "- We design a time-step alignment strategy to address the spatially non-uniform reconstruction quality of the rescaled latent features, allowing for more effective utilization of the prior information from the text-to-image diffusion model." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 471, + 142, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 142, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 142, + 483 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 491, + 155, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 155, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 155, + 504 + ], + "type": "text", + "content": "2.1. Image Rescaling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "Image rescaling aims to downscale HR images to visually coherent LR images and then plausibly reconstruct the original HR images. Unlike image SR, image rescaling involves joint modeling of both downscaling and upscaling processes, leading to higher reconstruction accuracy. Earlier researches [12, 16] utilize CNNs to build an encoder-decoder framework, treating image rescaling as a unified task by jointly training the downscaling and upscaling processes. Instead of directly generating LR images, CAR [25] proposes generating content-adaptive resampling kernels based on HR images to perform the downsampling operation, avoiding pixel-level constraints on the downscaled images. To fully leverage the inherent reversibility of image rescaling, there has been a surge in the use of INNs for this task [15, 38, 45, 48]. IRN [37] is the first invertible framework that models image rescaling as a bijective transformation, embedding residual high-frequency (HF) components" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "into a case-agnostic latent distribution for efficient reconstruction. HCFlow [17] assumes that the HF information depends on the LR image, thereby achieving better performance by incorporating LR conditions in HF information estimation. In extreme image rescaling, generative priors are usually employed to constrain the outcomes. GRAIN [47] and VQIR [34] leverage pre-trained StyleGAN [9, 10] and VQGAN [3], respectively, to achieve notable enhancements in perceptual quality. Nonetheless, these approaches are often limited in image contents, either confined to handling images from specific domains or encountering difficulties when reconstructing faces and textual elements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 225, + 479, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 479, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 479, + 238 + ], + "type": "text", + "content": "2.2. Diffusion Model for Image SR" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 243, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 361 + ], + "type": "text", + "content": "Here, we briefly introduce the application of diffusion models in a related task, i.e., image SR. Diffusion models promote the development of natural image SR in two main ways: training diffusion models from scratch [4, 13, 41] and leveraging pre-trained SD model [19, 36, 40]. The former is represented by SR3 [24], which achieves perceptual quality comparable to GANs. The latter is represented by StableSR [28], which aims to build a model capable of handling image restoration under any complex degradation in real-world scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 567 + ], + "type": "text", + "content": "However, the above methods typically require dozens or even hundreds of iterations during inference, which significantly increases computational costs. Researchers have begun exploring how to achieve image SR using one-step diffusion models. SinSR [31], based on ResShift [41], derives a deterministic one-step sampling process and employs a consistency preserving loss to distill the prior knowledge from a multi-step teacher network into a one-step student network. A representative work, OSEDiff [35], directly feeds the LQ image into the SD model and employs variational score distillation for regularization. S3Diff [43] proposes a degradation-aware LoRA and utilizes a pre-trained DINO [44] model as a discriminator for adversarial training. InvSR [42] designs a partial noise prediction strategy to provide a dynamic starting sampling point and supports inference with an arbitrary number of sampling steps ranging from one to five." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 579, + 397, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 397, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 397, + 593 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 599, + 429, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 429, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 429, + 611 + ], + "type": "text", + "content": "3.1. Overview of TADM" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "As illustrated in Fig. 1, TADM consists of four parts: Latent Encoding, Feature Rescaling in the Latent Space, Denoising Guided Perception Enhancement and Latent Decoding. First, the input HR images " + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": " are mapped to the latent features " + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": " via the pre-trained VAE encoder " + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": ". Subsequently, the proposed Decoupled Feature Rescaling Module (DFRM) rescales " + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": " to the target size, outputting LR images " + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": " and the rescaled latent features " + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": ". Our aim is to make " + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\hat{z}" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15595" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 213 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 213 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 213 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 213 + ], + "type": "image", + "image_path": "20f560aec677f0e2202ba95b14927a29a9381aa962307d0d48974ecb76432715.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 82, + 215, + 531, + 319 + ], + "blocks": [ + { + "bbox": [ + 82, + 215, + 531, + 319 + ], + "lines": [ + { + "bbox": [ + 82, + 215, + 531, + 319 + ], + "spans": [ + { + "bbox": [ + 82, + 215, + 531, + 319 + ], + "type": "image", + "image_path": "dffab37f138dca29e215e6a3d9eb66e5ee2a8f067984774baa57e3286b203c6a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "lines": [ + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed Timestep-Aware Diffusion Model (TADM). First, the HR images " + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "text", + "content": " are encoded to obtain the latent features " + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "text", + "content": ". The latent features are then rescaled to the target size using the decoupled feature rescaling module (DFRM), producing the LR images " + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "text", + "content": " and outputting the rescaled latent features " + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "text", + "content": ". Next, the rescaled latent features " + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "text", + "content": " are passed into a pre-trained diffusion model to perform a single denoising step for perceptual enhancement to obtain " + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "text", + "content": ". In this process, a time prediction module (TPM) is used to estimate the time step " + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "text", + "content": " and the predicted " + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "text", + "content": " is then fed into both the U-Net and the time scheduler. Finally, the perceptually enhanced latent features " + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "text", + "content": " are decoded to obtain the rescaled image " + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 54, + 327, + 555, + 393 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "text", + "content": "as close as possible to " + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "text", + "content": " so that the rescaled results align with the latent space of the pre-trained VAE, while ensuring that " + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "text", + "content": " and the original HR images " + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "text", + "content": " share the same content. Then, the rescaled features " + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "text", + "content": " are fed into a denoising U-Net network to perform a single denoising step to get the perceptually enhanced latent features " + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "text", + "content": " is decoded via the pre-trained VAE decoder " + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "text", + "content": " to obtain the rescaled image " + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 54, + 414, + 297, + 616 + ], + "type": "text", + "content": ". Overall, the latent rescaling process can be regarded as a noise-adding process, while the latent enhancement process can be viewed as a denoising process. To dynamically estimate the 'noise' density introduced by latent rescaling, we design a time alignment strategy, which consists of a time-step prediction module (TPM) and a time scheduler module. It can better address the non-uniform degradation caused by rescaling, thereby enabling the dynamic allocation of generative capabilities of pre-trained SD." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 255, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 255, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 255, + 643 + ], + "type": "text", + "content": "3.2. Feature Rescaling in the Latent Space" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 649, + 180, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 180, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 180, + 659 + ], + "type": "text", + "content": "3.2.1. Network Architecture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "In this part, we perform rescaling of the HR latent features " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", obtaining LR images " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " and rescaled features " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ". A straightforward idea is to construct a transformation chain as " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "y = G_{\\mathrm{e}}(z)" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{z} = G_{\\mathrm{d}}(y)" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{e}}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " is the encoder used" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "content": "to downscale the input features to the target sized LR images, while " + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "content": " represents the decoder used to reconstruct the original HR features. Guidance loss " + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gui}}" + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "content": " and reconstruction loss " + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "content": " are applied to " + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "content": " respectively, to force " + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "content": " to have the same content as " + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "content": " to be as similar as possible to " + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 414, + 555, + 486 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 375, + 496, + 553, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 496, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 375, + 496, + 553, + 509 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {g u i .}} = \\| y - \\operatorname {B i c u b i c} (x) \\| _ {1} \\tag {1}", + "image_path": "31f0698b544204cf0627c3ae2474025df7a13dce10f2397dd9a50316d520fcce.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 396, + 511, + 553, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 511, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 396, + 511, + 553, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c .}} = \\| \\hat {z} - z \\| _ {1} \\tag {2}", + "image_path": "ba44d834171c044f5122b535e0114af39f8675efa40e16d2d8bd783f41005db3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 533, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 653 + ], + "type": "text", + "content": "However, these two losses are contradictory to each other. When the guidance loss approaches zero, the rescaling problem degenerates into a classical SR problem [33], embedding little information about the downscaling process into the LR image. In contrast, when the reconstruction loss reaches its global optimum, the rescaling problem degenerates into an image compression problem [20], causing the LR image to lack meaningful visual content. Therefore, directly applying the guidance loss " + }, + { + "bbox": [ + 313, + 533, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gui}}" + }, + { + "bbox": [ + 313, + 533, + 555, + 653 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 533, + 555, + 653 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 533, + 555, + 653 + ], + "type": "text", + "content": " could adversely impact subsequent upscaling." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Considering the above issues, we devise the DFRM as shown in Fig. 2, which divides the rescaling operation into two sub-tasks: feature rescaling and LR image generation. Inspired by FGRN [15], we use a set of CNNs to perform downscaling and upscaling, and a separate INN to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15596" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 290, + 148 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 290, + 148 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 290, + 148 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 290, + 148 + ], + "type": "image", + "image_path": "1258a43e802c5300b8cc4c98aa94c29ee00708d7a8662f578ecd125fad9967a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 156, + 295, + 177 + ], + "lines": [ + { + "bbox": [ + 55, + 156, + 295, + 177 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 295, + 177 + ], + "type": "text", + "content": "Figure 2. Detailed architecture of Decoupled Feature Rescaling Module (DFRM)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "text", + "content": "perform the bidirectional mapping between the feature domain and the pixel domain. To achieve a balance between feature rescaling accuracy and visual quality of LR images, we introduce another pixel guidance module in the encoder to incorporate pixel-level information, which can improve the quality of the LR image. Specifically, DFRM perform feature rescaling through two independent transformation chains " + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "inline_equation", + "content": "((x,z)\\to z_{\\mathrm{lr}}\\to \\hat{z})" + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "text", + "content": " . " + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "inline_equation", + "content": "(z_{\\mathrm{lr}}\\leftrightarrow y)" + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 298, + 294, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 298, + 294, + 311 + ], + "spans": [ + { + "bbox": [ + 144, + 298, + 294, + 311 + ], + "type": "interline_equation", + "content": "z _ {\\mathrm {l r}} = G _ {\\mathrm {e}} (x, z) \\tag {3}", + "image_path": "b0e758023cbfbe1ca8b59c35eafe3b365ad278eccfaa2df0ac7056f3ba60a4a2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 313, + 294, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 313, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 113, + 313, + 294, + 327 + ], + "type": "interline_equation", + "content": "z _ {\\mathrm {l r}} \\approx F ^ {- 1} (F (z _ {\\mathrm {l r}})) = F ^ {- 1} (y) \\tag {4}", + "image_path": "23aa7864a8aebcc4863de166b8c8bf4de675f2b71c0670cb94cd4f43d9c6ce4b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 329, + 294, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 329, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 149, + 329, + 294, + 342 + ], + "type": "interline_equation", + "content": "\\hat {z} = G _ {\\mathrm {d}} \\left(z _ {\\mathrm {l r}}\\right) \\tag {5}", + "image_path": "e31ae4035f8af5195a662d6e41bc03f3e24c8dc03526ae7d4d7066fc159581e5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "inline_equation", + "content": "F^{-1}" + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": " represent the forward and backward processes of the INN, respectively. Given the latent variable " + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{\\frac{H}{8} \\times \\frac{W}{8} \\times 4}" + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": " of the pre-trained SD VAE as input, the encoder " + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{e}}" + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": " first transforms it into a compact representation " + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{lr}} \\in \\mathbb{R}^{\\frac{H}{N} \\times \\frac{W}{N} \\times 3}" + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": ", whose size is the same as that of the target LR image and " + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": " represents the target downscaling factor. Subsequently, a set of INNs perform the forward process to transform " + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{lr}}" + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": " from the feature domain to the pixel domain and obtain the LR image " + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}^{\\frac{H}{N} \\times \\frac{W}{N} \\times 3}" + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": " by quantization. In the upscaling process, the inverse of the INN is used to process the LR image " + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": " and reconstruct " + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{lr}}" + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": ". Finally, the decoder " + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": " is utilized to obtain the rescaled feature " + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 55, + 348, + 296, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 496, + 146, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 146, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 146, + 506 + ], + "type": "text", + "content": "3.2.2. Loss Function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 511, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 606 + ], + "type": "text", + "content": "Training of this part involves two losses: reconstruction loss and guidance loss. Considering the information lost in quantization operations, the bidirectional mapping between " + }, + { + "bbox": [ + 55, + 511, + 296, + 606 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{lr}}" + }, + { + "bbox": [ + 55, + 511, + 296, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 511, + 296, + 606 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 511, + 296, + 606 + ], + "type": "text", + "content": " is not exactly reversible. To ensure the robustness of the network to quantization operations, we consider two types of transformation chains in designing the reconstruction loss: " + }, + { + "bbox": [ + 55, + 511, + 296, + 606 + ], + "type": "inline_equation", + "content": "(x,z)\\to z_{\\mathrm{lr}}\\to \\hat{z}" + }, + { + "bbox": [ + 55, + 511, + 296, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 511, + 296, + 606 + ], + "type": "inline_equation", + "content": "(x,z)\\to z_{\\mathrm{lr}}\\to y\\to \\hat{z}_{\\mathrm{lr}}\\to \\hat{z}" + }, + { + "bbox": [ + 55, + 511, + 296, + 606 + ], + "type": "text", + "content": ", which can be expressed as the following formula:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 85, + 612, + 295, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 612, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 85, + 612, + 295, + 643 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {r e c .}} = \\| G _ {\\mathrm {d}} \\left(G _ {\\mathrm {e}} (x, z)\\right) - z \\| _ {1} + \\tag {6} \\\\ \\left. \\left\\| G _ {\\mathrm {d}} \\left(F ^ {- 1} \\left(F \\left(G _ {\\mathrm {e}} (x, z)\\right)\\right)\\right) - z \\right\\| _ {1} \\right. \\\\ \\end{array}", + "image_path": "a5c12a63851643a6d57623a39e3f4980954cb0a1cac804c7ce90a949eb5e7f4e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 694 + ], + "type": "text", + "content": "This combination can facilitate the network to reconstruct " + }, + { + "bbox": [ + 55, + 647, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 55, + 647, + 296, + 694 + ], + "type": "text", + "content": " with high fidelity in both the existence and absence of quantization operations. Additionally, the guidance loss is designed as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 91, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 91, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {g u i .}} = \\left\\| F \\left(G _ {\\mathrm {e}} (x, z)\\right) - \\operatorname {B i c u b i c} (x) \\right\\| _ {1} \\tag {7}", + "image_path": "b7e1a8405e43a0e59d8efc4c9eb19cca0d8e62d553236aee44c5756513290ad3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "Finally, the loss function of this part is a combination of these two losses:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 370, + 103, + 553, + 117 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 103, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 370, + 103, + 553, + 117 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e s .}} = \\lambda_ {\\text {r e c .}} \\mathcal {L} _ {\\text {r e c .}} + \\lambda_ {\\text {g u i .}} \\mathcal {L} _ {\\text {g u i .}} \\tag {8}", + "image_path": "870a4dfc76eec4879d3401b762187acdd7a2f94f9c4b9c0a4617571fe75ccb9c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 122, + 541, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 541, + 134 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 541, + 134 + ], + "type": "text", + "content": "3.3. Denoising Guided Perception Enhancement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 140, + 555, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 555, + 283 + ], + "type": "text", + "content": "The latent variable " + }, + { + "bbox": [ + 313, + 140, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 313, + 140, + 555, + 283 + ], + "type": "text", + "content": " obtained in Sec. 3.2 retains sufficient semantic information about the HR image, but lacks adequate texture. Therefore, we propose to leverage the strong priors stored in the pre-trained SD model to enhance the perceptual quality of the rescaling results in this part. Original diffusion model requires a multi-step sampling algorithm, significantly increases inference time and makes it difficult to scale to practical applications. Inspired by recent work on accelerating diffusion models [8], we adopt an one-step inference strategy. Specifically, we can directly feed the rescaled latent variable " + }, + { + "bbox": [ + 313, + 140, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 313, + 140, + 555, + 283 + ], + "type": "text", + "content": " into the denoising U-Net " + }, + { + "bbox": [ + 313, + 140, + 555, + 283 + ], + "type": "inline_equation", + "content": "\\epsilon = \\epsilon_{\\theta}(\\hat{z},t)" + }, + { + "bbox": [ + 313, + 140, + 555, + 283 + ], + "type": "text", + "content": " and perform a single denoising step as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 392, + 299, + 553, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 299, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 392, + 299, + 553, + 326 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {0} = \\frac {\\hat {z} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon}{\\sqrt {\\bar {\\alpha} _ {t}}} \\tag {9}", + "image_path": "b2bbf98a147c8d0e90d4bd9a1407d73c89f477b7484e8dd94adac1e7c89b8511.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": " represents the perceptually enhanced latent features, " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t}" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": " is the noise scheduler of the forward diffusion process of the SD model, and " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": " represents the time-step associated with the noise intensity. Finally, " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": " can be transformed to the pixel domain through the pre-trained VAE decoder " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": " to obtain the upscaled image " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\hat{x} = \\mathcal{D}(\\hat{z}_0)" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": ". Then, we can adopt the loss function between " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": " to optimize the U-Net using LoRA layers." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "type": "text", + "content": "However, the key problem of the aforementioned approaches lies in the selection of time-step " + }, + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "type": "text", + "content": ". We can regard the rescaling process as the noise-adding process of diffusion model, while Eq. (9) represents the denoising process. For different rescaling factors and different images, the 'noise' introduced by the rescaling operation varies in intensity. In Fig. 3, we present the global rescaling error of an image under different rescaling factors, as well as the local rescaling error of different patches within the image. Additionally, we plot the error curve of the forward noise-adding process of the SD model. It can be observed that as the rescaling factor and image content change, rescaling MSE also varies and corresponds to the diffusion MSE at different time-steps. Thus, we propose to dynamically predict the time-step " + }, + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "type": "text", + "content": " based on the quality of the rescaled latent features " + }, + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "type": "text", + "content": ". A simple yet effective network named timestep prediction module (TPM), whose structure is presented in Fig. 1, is used to predict " + }, + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "type": "text", + "content": " based on the rescaled latent features: " + }, + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "type": "inline_equation", + "content": "t = \\mathrm{TPM}(\\hat{z})" + }, + { + "bbox": [ + 313, + 427, + 555, + 715 + ], + "type": "text", + "content": ". During the training phase, we compute a single time-step for each input image patch. In the inference phase, we adopt a tiled inference approach (which will be elaborated in detail in the supplementary materials), allowing for predicting different time-steps for various regions of a high-resolution input image." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15597" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 282, + 229 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 282, + 229 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 282, + 229 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 282, + 229 + ], + "type": "image", + "image_path": "1b8b3468943aab65fa6153d8e81a81c03b422dbd2e378353f4c35c2984780218.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 237, + 296, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 293 + ], + "type": "text", + "content": "Figure 3. We compare the MSE introduced by the rescaling operation with that caused by the forward diffusion process. The rescaling degradation severity is correlated with both the image content and the rescaling scale. Therefore, the model requires adaptive time-step to align the rescaling MSE with the diffusion MSE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 302, + 282, + 383 + ], + "blocks": [ + { + "bbox": [ + 70, + 302, + 282, + 383 + ], + "lines": [ + { + "bbox": [ + 70, + 302, + 282, + 383 + ], + "spans": [ + { + "bbox": [ + 70, + 302, + 282, + 383 + ], + "type": "image", + "image_path": "151175251986af6e47c306ecd577e79aeb9bce6dfba03ffba3970274e61c1362.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 391, + 296, + 436 + ], + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 436 + ], + "type": "text", + "content": "Figure 4. The structure of the time scheduler module. It combines both the fixed scheduler and learnable scheduler to perform the denoising process. The two schedulers are connected by a zero-convolution layer to stabilize the training." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "type": "text", + "content": "Subsequently, the predicted time-step " + }, + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "type": "text", + "content": " is used to perform a single denoising step. However, the vanilla SD time scheduler as shown in Eq. (9) is not differentiable with respect to " + }, + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "type": "text", + "content": ", which prevents gradients from being backpropagated to the TPM. Considering that given " + }, + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 456, + 296, + 646 + ], + "type": "text", + "content": ", Eq. (9) essentially represents a linear transformation, which can be learned by a neural network. Therefore, we propose a hybrid time scheduler module consisting of a fixed scheduler and a learnable scheduler, whose structure is presented in Fig. 4. The former uses Eq. (9) to perform the denoising step, while the latter employs a neural network to simulate the denoising scheduler. To ensure training stability, the final layer of the learnable scheduler is a zero-initialized convolution, allowing it to gradually correct the results of the fixed scheduler as training progresses. This process can be described as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 656, + 295, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 656, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 96, + 656, + 295, + 669 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {0} = \\mathcal {S} _ {\\text {f i x e d}} (\\hat {z}, \\epsilon , t _ {0}) + \\mathcal {S} _ {\\text {l e a r n e d}} (\\hat {z}, \\epsilon , t) \\tag {10}", + "image_path": "455c69d8b6197b560fc7016ba1e63b90e048c3dea60a678b8cf6d6c552033130.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "content": " is the fixed pre-set time-step and " + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{fixed}}" + }, + { + "bbox": [ + 55, + 677, + 295, + 700 + ], + "type": "text", + "content": " denotes the vanilla SD time scheduler as shown in Eq. (9)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Finally, we fine-tune the denoising U-Net network and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "the VAE decoder using LoRA, while also jointly training the TPM and time scheduler module. To achieve a balance between fidelity and realism, we supervise the reconstruction results " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " using pixel loss and perceptual loss as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 337, + 129, + 554, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 129, + 554, + 158 + ], + "spans": [ + { + "bbox": [ + 337, + 129, + 554, + 158 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {e n h .}} = \\| \\hat {x} - x \\| _ {1} + \\tag {11} \\\\ \\lambda_ {\\text {p e c .}} \\left(\\mathcal {L} _ {\\text {l p i p s}} (\\hat {x}, x) + \\mathcal {L} _ {\\text {d i s t s}} (\\hat {x}, x)\\right) \\\\ \\end{array}", + "image_path": "ceb08773b457695c974998bdbe6a186b8dc61fd6a3565f5c890c226d8a3afeda.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 166, + 395, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 395, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 395, + 178 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 185, + 430, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 430, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 430, + 198 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 203, + 484, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 484, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 484, + 213 + ], + "type": "text", + "content": "4.1.1. Datasets and Evaluation Metrics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 217, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 373 + ], + "type": "text", + "content": "We train the proposed TADM using the DF2K dataset [1, 18]. Considering that traditional benchmarks such as Set5, Set14, and BSD100 often have lower quality and resolution [5], they are not suitable for evaluating extreme rescaling methods. Therefore, we utilize four high-resolution datasets: DIV2K, CLIC2020 [26], Urban100 [6], and DIV8K [5] to evaluate models been compared (results on the latter two datasets are presented in the supplementary materials). We adopt PSNR and SSIM [32] for quantitative evaluation. For perceptual quality evaluation, reference-based metrics LPIPS [46] and DISTS [2] are used. For non-reference image quality assessment, we use metrics including MUSIQ [11] and CLIPIQA [27]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 379, + 411, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 411, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 411, + 390 + ], + "type": "text", + "content": "4.1.2. Training Details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 393, + 554, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 489 + ], + "type": "text", + "content": "TADM is built based on the SD 2.1-base model, and the training is divided into three stages. First, we train the DFRM using Eq. (8). Then we adopt Eq. (11) to jointly train the LoRA layers, the TPM and the time scheduler module. Finally, we use a smaller learning rate to jointly fine-tune DFRM, LoRA layers, TPM and scheduler module using a combination of Eq. (8) and (11). More training details are provided in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 496, + 506, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 506, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 506, + 509 + ], + "type": "text", + "content": "4.2. Results and Comparison with SOTA" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 514, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 633 + ], + "type": "text", + "content": "We compare the proposed TADM with three kinds of methods: (1) bicubic downscaling and SR, i.e., ESRGAN [30], StableSR [28] (based on multi-step diffusion) and S3Diff (based on single-step diffusion)[43]; (2) flow-based rescaling models, i.e., IRN [37] and HCFlow [17]; (3) prior-based extreme rescaling method VQIR [34]. Some of them are proposed for " + }, + { + "bbox": [ + 313, + 514, + 555, + 633 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 514, + 555, + 633 + ], + "type": "text", + "content": " rescaling factors and we extended them to " + }, + { + "bbox": [ + 313, + 514, + 555, + 633 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 313, + 514, + 555, + 633 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 514, + 555, + 633 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 313, + 514, + 555, + 633 + ], + "type": "text", + "content": ". To ensure a fair comparison, all methods are trained until convergence on the DF2K dataset following their original configurations." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 639, + 430, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 430, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 430, + 650 + ], + "type": "text", + "content": "4.2.1. Quantitative Results" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Table 1 presents the " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "16\\times" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "32\\times" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " quantitative image rescaling results on four datasets. The performance of SR methods are generally poor. Both gan-based (ESRGAN) and diffusion-based (StableSR and S3Diff) methods can only achieve relative good performance on no-reference metrics." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15598" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 112, + 555, + 293 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with different methods at " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "{16} \\times" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "{32} \\times" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " (results on Urban100 and DIV8K are presented in the supplementary materials). The symbols " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " respectively represent that higher or lower values indicate better performance. Bold represents the best and underline represents the second best." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 112, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 56, + 112, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 112, + 555, + 293 + ], + "type": "table", + "html": "
DatasetMethodPSNR ↑SSIM ↑LPIPS ↓DISTS ↓MUSIQ ↑CLIPIQA ↑
16×32×16×32×16×32×16×32×16×32×16×32×
DIV2KESRGAN23.1519.940.59460.56890.44780.58920.23780.423059.8043.350.61610.3734
StableSR22.0219.920.59830.55880.49800.57360.20780.260351.2848.940.37800.3203
S3Diff20.2217.810.54910.46010.40330.48950.13090.181064.3767.920.62280.6991
IRN26.0422.840.70200.62120.51530.60950.30410.403643.7228.700.27990.2336
HCFlow26.6623.890.71760.64670.48850.58160.28660.385246.4337.250.27350.2792
VQIR23.9122.020.64980.58230.31740.45680.10240.266364.0458.210.63500.6293
Ours23.9822.180.67360.58690.29790.42210.08860.168466.5669.120.71890.7204
CLIC2020ESRGAN25.4322.470.70180.68040.42130.54290.22090.401157.8041.280.61070.3558
StableSR24.1122.240.70520.67260.46250.52880.18600.228252.6949.800.37790.3325
S3Diff22.0419.870.66150.57110.39680.48290.13360.181263.0766.400.60010.6838
IRN28.5524.930.79070.72280.46760.55280.27460.377647.6627.810.29500.2495
HCFlow28.8226.210.79490.74510.45420.53250.26350.361147.7237.150.27330.2884
VQIR25.6924.120.73630.69420.28540.41910.09010.244564.3654.830.61050.6068
Ours26.0424.260.75420.69030.27240.39110.07920.151365.3569.100.68150.7271
", + "image_path": "b49bae9e01fbfa476ab8eb510004c6378cfe0478dec1e4a942d880b2db589c55.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 312, + 297, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 297, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 297, + 517 + ], + "type": "text", + "content": "This is because they rely solely on the LR image to reconstruct the original image and cannot jointly optimize the downscaling and upscaling processes. While methods based on INNs (HCFlow and IRN) achieve the highest PSNR and SSIM due to its consideration of the correlation between downscaling and upscaling. In all generative methods, VQIR achieves relatively good perceptual metrics by leveraging the priors of the pre-trained VQGAN. In contrast, our method achieves optimal performance across all perceptual metrics on all datasets. For instance, in the " + }, + { + "bbox": [ + 54, + 312, + 297, + 517 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 54, + 312, + 297, + 517 + ], + "type": "text", + "content": " rescaling on the DIV2K validation set, our method improves the DISTS metric by " + }, + { + "bbox": [ + 54, + 312, + 297, + 517 + ], + "type": "inline_equation", + "content": "36.76\\%" + }, + { + "bbox": [ + 54, + 312, + 297, + 517 + ], + "type": "text", + "content": " compared to the second-best VQIR. Importantly, our approach is designed for applications like social media delivery, where semantic fidelity and perceptual realism are more critical. In such cases, metrics like LPIPS and DISTS better reflect image quality [23]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 529, + 165, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 165, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 165, + 540 + ], + "type": "text", + "content": "4.2.2. Qualitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": "In Fig. 5, we present visual comparisons at " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " rescaling factors. Due to the ill-posed nature of extreme rescaling, pixel-oriented optimization methods such as IRN, and HCFlow produce overly-smoothed results. Although ESRGAN incorporates perceptual optimization it still cannot generate adequate details. VQIR leverages the pretrained VQGAN as a prior, allowing it to generate rich details and sharp edges. However, it struggles with faces and more complex structures. Although S3Diff can generate sharp edges and rich details, its reconstructed results have low fidelity and exhibit significant artifacts and color distortions. Thanks to the powerful prior provided by the pre-trained SD and the carefully designed architecture,, our method effectively addresses these issues. It not only gener" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 312, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 348 + ], + "type": "text", + "content": "ates sufficiently rich textures but also reconstructs accuracy semantics and regular structures, such as text, faces, vehicles, and leaves." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 355, + 408, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 408, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 408, + 369 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": "We conduct ablation experiments to validate the effectiveness of each component. All ablations are trained on the DF2K dataset and tested on the validation set of the DIV2K for " + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": " rescaling task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 488, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 488, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 488, + 439 + ], + "type": "text", + "content": "4.3.1. Effectiveness of Feature Rescaling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 441, + 556, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 556, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 556, + 574 + ], + "type": "text", + "content": "To verify the effectiveness of the feature rescaling operation, we move the rescaling operation from the latent space to the pixel space. Specifically, we use HCFlow [17], the comparative method with the best PSNR performance, to perform the rescaling operation in the pixel domain, followed by perceptual enhancement using the same approach described in Sec. 3.3. As shown the 1st row in Table 2, the rescaling operation in the latent space can improve reconstruction performance, as evidenced by much lower LPIPS and DISTS. Additional details and results are presented in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 337, + 603, + 533, + 645 + ], + "blocks": [ + { + "bbox": [ + 331, + 583, + 536, + 595 + ], + "lines": [ + { + "bbox": [ + 331, + 583, + 536, + 595 + ], + "spans": [ + { + "bbox": [ + 331, + 583, + 536, + 595 + ], + "type": "text", + "content": "Table 2. Ablation study of rescaling space and SD prior." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 337, + 603, + 533, + 645 + ], + "lines": [ + { + "bbox": [ + 337, + 603, + 533, + 645 + ], + "spans": [ + { + "bbox": [ + 337, + 603, + 533, + 645 + ], + "type": "table", + "html": "
Latent spacePixel spaceSD priorLPIPS ↓DISTS ↓
×✓✓0.36300.1154
✓××0.46750.3109
✓×✓0.29790.0886
", + "image_path": "4c7d13a6aac895a6756f4c8294e6e9ae954b3340d894cffcc93a028822f35a03.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 662, + 434, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 434, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 434, + 674 + ], + "type": "text", + "content": "4.3.2. Necessity of SD Prior" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "To validate that SD model provides rich priors about real-world HR images, we directly omit the perceptual enhancement part. As shown in Table 2 (2nd row), the introduction" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15599" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 69, + 153, + 117 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 153, + 117 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 153, + 117 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 153, + 117 + ], + "type": "image", + "image_path": "dd4ddc584944e372d5db7496c087e1745d12376f4e726b92dcdbb857fdf82d0e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 118, + 152, + 164 + ], + "blocks": [ + { + "bbox": [ + 69, + 118, + 152, + 164 + ], + "lines": [ + { + "bbox": [ + 69, + 118, + 152, + 164 + ], + "spans": [ + { + "bbox": [ + 69, + 118, + 152, + 164 + ], + "type": "image", + "image_path": "3ad96ea359717441ed6b97350ac3e3e7d239ed570350e582011de0a70f6c02f6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 69, + 167, + 152, + 213 + ], + "blocks": [ + { + "bbox": [ + 69, + 167, + 152, + 213 + ], + "lines": [ + { + "bbox": [ + 69, + 167, + 152, + 213 + ], + "spans": [ + { + "bbox": [ + 69, + 167, + 152, + 213 + ], + "type": "image", + "image_path": "736ea63ecf9ac2858c6dbe394a4b539e83d3d845dc5d1f275e68e346e4756525.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 69, + 215, + 152, + 262 + ], + "blocks": [ + { + "bbox": [ + 69, + 215, + 152, + 262 + ], + "lines": [ + { + "bbox": [ + 69, + 215, + 152, + 262 + ], + "spans": [ + { + "bbox": [ + 69, + 215, + 152, + 262 + ], + "type": "image", + "image_path": "a528fdf9259a50289bf98dc1ee7f8704c1dcc188e435c4975b945dbcaddf813a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 265, + 199, + 274 + ], + "lines": [ + { + "bbox": [ + 155, + 265, + 199, + 274 + ], + "spans": [ + { + "bbox": [ + 155, + 265, + 199, + 274 + ], + "type": "text", + "content": "(a) StableSR" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 155, + 69, + 201, + 117 + ], + "blocks": [ + { + "bbox": [ + 155, + 69, + 201, + 117 + ], + "lines": [ + { + "bbox": [ + 155, + 69, + 201, + 117 + ], + "spans": [ + { + "bbox": [ + 155, + 69, + 201, + 117 + ], + "type": "image", + "image_path": "f21437c8c9194ba8527af1d16b2c1aad498338ffb8e7c457be5593ca691a4620.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 155, + 118, + 201, + 165 + ], + "blocks": [ + { + "bbox": [ + 155, + 118, + 201, + 165 + ], + "lines": [ + { + "bbox": [ + 155, + 118, + 201, + 165 + ], + "spans": [ + { + "bbox": [ + 155, + 118, + 201, + 165 + ], + "type": "image", + "image_path": "f192fb0a0ce666244f33b1c476f3c57ef5db36b9869194dc90449b99c9bed680.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 155, + 167, + 201, + 214 + ], + "blocks": [ + { + "bbox": [ + 155, + 167, + 201, + 214 + ], + "lines": [ + { + "bbox": [ + 155, + 167, + 201, + 214 + ], + "spans": [ + { + "bbox": [ + 155, + 167, + 201, + 214 + ], + "type": "image", + "image_path": "ec99751125d639c4342048c63d38e32ecd36f0712f709c016003797bfb635bdf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 155, + 215, + 201, + 262 + ], + "blocks": [ + { + "bbox": [ + 155, + 215, + 201, + 262 + ], + "lines": [ + { + "bbox": [ + 155, + 215, + 201, + 262 + ], + "spans": [ + { + "bbox": [ + 155, + 215, + 201, + 262 + ], + "type": "image", + "image_path": "40480862e59a7bc2154612944e23626525d056aa94f6688975f88bcf76b12d3e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 265, + 242, + 274 + ], + "lines": [ + { + "bbox": [ + 207, + 265, + 242, + 274 + ], + "spans": [ + { + "bbox": [ + 207, + 265, + 242, + 274 + ], + "type": "text", + "content": "(b) S3Diff" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 203, + 69, + 249, + 117 + ], + "blocks": [ + { + "bbox": [ + 203, + 69, + 249, + 117 + ], + "lines": [ + { + "bbox": [ + 203, + 69, + 249, + 117 + ], + "spans": [ + { + "bbox": [ + 203, + 69, + 249, + 117 + ], + "type": "image", + "image_path": "ce7e424750f685b1e9c7fdc635fbf278bba6abe447500e7e7e9bdef55c50ed61.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 203, + 118, + 249, + 165 + ], + "blocks": [ + { + "bbox": [ + 203, + 118, + 249, + 165 + ], + "lines": [ + { + "bbox": [ + 203, + 118, + 249, + 165 + ], + "spans": [ + { + "bbox": [ + 203, + 118, + 249, + 165 + ], + "type": "image", + "image_path": "c0e870a5e34d9d4b162d2cc0a77664cfe0f80e682e46d5061a82fe9006d2003a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 203, + 167, + 249, + 214 + ], + "blocks": [ + { + "bbox": [ + 203, + 167, + 249, + 214 + ], + "lines": [ + { + "bbox": [ + 203, + 167, + 249, + 214 + ], + "spans": [ + { + "bbox": [ + 203, + 167, + 249, + 214 + ], + "type": "image", + "image_path": "75fff6b0e603a820472cc8ab0c24b0b4511190d136b7d88cd1cd3daf86fa54ba.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 203, + 215, + 249, + 262 + ], + "blocks": [ + { + "bbox": [ + 203, + 215, + 249, + 262 + ], + "lines": [ + { + "bbox": [ + 203, + 215, + 249, + 262 + ], + "spans": [ + { + "bbox": [ + 203, + 215, + 249, + 262 + ], + "type": "image", + "image_path": "2e8c024997cf5bcc4672a14ce78c1b6d5c3d4e50c668f37d615a4aecb85e0e5a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 265, + 294, + 274 + ], + "lines": [ + { + "bbox": [ + 253, + 265, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 253, + 265, + 294, + 274 + ], + "type": "text", + "content": "(c)ESRGAN" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 252, + 69, + 298, + 117 + ], + "blocks": [ + { + "bbox": [ + 252, + 69, + 298, + 117 + ], + "lines": [ + { + "bbox": [ + 252, + 69, + 298, + 117 + ], + "spans": [ + { + "bbox": [ + 252, + 69, + 298, + 117 + ], + "type": "image", + "image_path": "ff32ee53fec9d1c2fffa6c6e92789b67184aae4e33ed5f6e51c663c127cec4f1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 252, + 118, + 298, + 166 + ], + "blocks": [ + { + "bbox": [ + 252, + 118, + 298, + 166 + ], + "lines": [ + { + "bbox": [ + 252, + 118, + 298, + 166 + ], + "spans": [ + { + "bbox": [ + 252, + 118, + 298, + 166 + ], + "type": "image", + "image_path": "1d7a76e49bfc44da7c92f83cf97ff6f3ae70bb234824947c3e841ddec9461a5b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 252, + 167, + 298, + 214 + ], + "blocks": [ + { + "bbox": [ + 252, + 167, + 298, + 214 + ], + "lines": [ + { + "bbox": [ + 252, + 167, + 298, + 214 + ], + "spans": [ + { + "bbox": [ + 252, + 167, + 298, + 214 + ], + "type": "image", + "image_path": "1555b512a06c3b435d04bde73b0f7450fadf6ec84e2c98f5e728359fe0a1adab.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 252, + 215, + 298, + 262 + ], + "blocks": [ + { + "bbox": [ + 252, + 215, + 298, + 262 + ], + "lines": [ + { + "bbox": [ + 252, + 215, + 298, + 262 + ], + "spans": [ + { + "bbox": [ + 252, + 215, + 298, + 262 + ], + "type": "image", + "image_path": "0224e89361a889d0b280f64e1e88a0f3750f236c8872ad13250582f078d8fac4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 301, + 69, + 347, + 117 + ], + "blocks": [ + { + "bbox": [ + 301, + 69, + 347, + 117 + ], + "lines": [ + { + "bbox": [ + 301, + 69, + 347, + 117 + ], + "spans": [ + { + "bbox": [ + 301, + 69, + 347, + 117 + ], + "type": "image", + "image_path": "c722e92a09990f1d30c2e50ef1e560ba8338762fc8ee3849fb5fc43228aef23b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 301, + 118, + 347, + 165 + ], + "blocks": [ + { + "bbox": [ + 301, + 118, + 347, + 165 + ], + "lines": [ + { + "bbox": [ + 301, + 118, + 347, + 165 + ], + "spans": [ + { + "bbox": [ + 301, + 118, + 347, + 165 + ], + "type": "image", + "image_path": "b9c133f8e1dcfab98e86580b5fce12b436319dd8f1d90c57a9a0d398692c5a2f.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 301, + 167, + 347, + 214 + ], + "blocks": [ + { + "bbox": [ + 301, + 167, + 347, + 214 + ], + "lines": [ + { + "bbox": [ + 301, + 167, + 347, + 214 + ], + "spans": [ + { + "bbox": [ + 301, + 167, + 347, + 214 + ], + "type": "image", + "image_path": "68e57e783fd83330b93f324cfe085b9130b9d13367d09244b26d4e727d7b6412.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 301, + 215, + 347, + 262 + ], + "blocks": [ + { + "bbox": [ + 301, + 215, + 347, + 262 + ], + "lines": [ + { + "bbox": [ + 301, + 215, + 347, + 262 + ], + "spans": [ + { + "bbox": [ + 301, + 215, + 347, + 262 + ], + "type": "image", + "image_path": "1be1b4abf1fb7b7801e69b69c925ceaf6e86950729c8432674164e36d44647ae.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 265, + 335, + 274 + ], + "lines": [ + { + "bbox": [ + 310, + 265, + 335, + 274 + ], + "spans": [ + { + "bbox": [ + 310, + 265, + 335, + 274 + ], + "type": "text", + "content": "(d) IRN" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 349, + 69, + 395, + 117 + ], + "blocks": [ + { + "bbox": [ + 349, + 69, + 395, + 117 + ], + "lines": [ + { + "bbox": [ + 349, + 69, + 395, + 117 + ], + "spans": [ + { + "bbox": [ + 349, + 69, + 395, + 117 + ], + "type": "image", + "image_path": "8d7b6aab8d3d937a2a1762e9951702b989d46df1a6dd7e0f0278b352910a8463.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 349, + 118, + 395, + 166 + ], + "blocks": [ + { + "bbox": [ + 349, + 118, + 395, + 166 + ], + "lines": [ + { + "bbox": [ + 349, + 118, + 395, + 166 + ], + "spans": [ + { + "bbox": [ + 349, + 118, + 395, + 166 + ], + "type": "image", + "image_path": "fc21f908646bd65b2d004886f51b93615c6cff9db8f55252296676e49d37da14.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 349, + 167, + 395, + 214 + ], + "blocks": [ + { + "bbox": [ + 349, + 167, + 395, + 214 + ], + "lines": [ + { + "bbox": [ + 349, + 167, + 395, + 214 + ], + "spans": [ + { + "bbox": [ + 349, + 167, + 395, + 214 + ], + "type": "image", + "image_path": "1005b8308246bc87729221cdd4025084f83e4121fda414d86ae1467fe7467ed0.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 349, + 215, + 395, + 262 + ], + "blocks": [ + { + "bbox": [ + 349, + 215, + 395, + 262 + ], + "lines": [ + { + "bbox": [ + 349, + 215, + 395, + 262 + ], + "spans": [ + { + "bbox": [ + 349, + 215, + 395, + 262 + ], + "type": "image", + "image_path": "05a06bd329b41e86ca5e52de11dd852a67b150c6f7839b592dce35404b1c8325.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 265, + 391, + 274 + ], + "lines": [ + { + "bbox": [ + 351, + 265, + 391, + 274 + ], + "spans": [ + { + "bbox": [ + 351, + 265, + 391, + 274 + ], + "type": "text", + "content": "(e) HCFlow" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 398, + 69, + 444, + 117 + ], + "blocks": [ + { + "bbox": [ + 398, + 69, + 444, + 117 + ], + "lines": [ + { + "bbox": [ + 398, + 69, + 444, + 117 + ], + "spans": [ + { + "bbox": [ + 398, + 69, + 444, + 117 + ], + "type": "image", + "image_path": "3009741a3727e4f37d7b124c161837c2907bc01464ce4fa0e7f24f3cc708bd30.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 398, + 118, + 444, + 166 + ], + "blocks": [ + { + "bbox": [ + 398, + 118, + 444, + 166 + ], + "lines": [ + { + "bbox": [ + 398, + 118, + 444, + 166 + ], + "spans": [ + { + "bbox": [ + 398, + 118, + 444, + 166 + ], + "type": "image", + "image_path": "d88d75630eb5fcdab0e3240428056c0451d037d8036f9bca63b6db0633f7e7e6.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 398, + 167, + 444, + 214 + ], + "blocks": [ + { + "bbox": [ + 398, + 167, + 444, + 214 + ], + "lines": [ + { + "bbox": [ + 398, + 167, + 444, + 214 + ], + "spans": [ + { + "bbox": [ + 398, + 167, + 444, + 214 + ], + "type": "image", + "image_path": "aecbabf27b17e5d3b8ade4b49cb218b54f363c93e43b35dbf6de95d6c129ad96.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 398, + 215, + 444, + 262 + ], + "blocks": [ + { + "bbox": [ + 398, + 215, + 444, + 262 + ], + "lines": [ + { + "bbox": [ + 398, + 215, + 444, + 262 + ], + "spans": [ + { + "bbox": [ + 398, + 215, + 444, + 262 + ], + "type": "image", + "image_path": "2ed0ceeabb474c54a0036586d0201ef594251d5fddea50ed4bcb8325072e0634.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 265, + 434, + 274 + ], + "lines": [ + { + "bbox": [ + 405, + 265, + 434, + 274 + ], + "spans": [ + { + "bbox": [ + 405, + 265, + 434, + 274 + ], + "type": "text", + "content": "(f) VQIR" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 446, + 69, + 492, + 117 + ], + "blocks": [ + { + "bbox": [ + 446, + 69, + 492, + 117 + ], + "lines": [ + { + "bbox": [ + 446, + 69, + 492, + 117 + ], + "spans": [ + { + "bbox": [ + 446, + 69, + 492, + 117 + ], + "type": "image", + "image_path": "5127534e08da84053f78292987e6df2a887351d853a1db216f01c8ce402a3898.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 446, + 118, + 492, + 166 + ], + "blocks": [ + { + "bbox": [ + 446, + 118, + 492, + 166 + ], + "lines": [ + { + "bbox": [ + 446, + 118, + 492, + 166 + ], + "spans": [ + { + "bbox": [ + 446, + 118, + 492, + 166 + ], + "type": "image", + "image_path": "bdf4734a648115027bd2e98154f6476b394374c83d6073ea9b033d1b1888c0b9.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 446, + 167, + 492, + 214 + ], + "blocks": [ + { + "bbox": [ + 446, + 167, + 492, + 214 + ], + "lines": [ + { + "bbox": [ + 446, + 167, + 492, + 214 + ], + "spans": [ + { + "bbox": [ + 446, + 167, + 492, + 214 + ], + "type": "image", + "image_path": "95cb22709bb6411cd8b818ed25e8e70d4609198834788e83093ede0df8580e94.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 446, + 215, + 492, + 262 + ], + "blocks": [ + { + "bbox": [ + 446, + 215, + 492, + 262 + ], + "lines": [ + { + "bbox": [ + 446, + 215, + 492, + 262 + ], + "spans": [ + { + "bbox": [ + 446, + 215, + 492, + 262 + ], + "type": "image", + "image_path": "a52781ffb42af446eabbe24d887b259d1a241569c777e7087e6e144bf1c34e62.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 265, + 484, + 274 + ], + "lines": [ + { + "bbox": [ + 453, + 265, + 484, + 274 + ], + "spans": [ + { + "bbox": [ + 453, + 265, + 484, + 274 + ], + "type": "text", + "content": "(g) Ours" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 494, + 69, + 542, + 117 + ], + "blocks": [ + { + "bbox": [ + 494, + 69, + 542, + 117 + ], + "lines": [ + { + "bbox": [ + 494, + 69, + 542, + 117 + ], + "spans": [ + { + "bbox": [ + 494, + 69, + 542, + 117 + ], + "type": "image", + "image_path": "f399334587a6bd0fe7c8bde7430e45ec43bcabe23f6c1fbe9609f1b79dac0765.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 494, + 118, + 542, + 166 + ], + "blocks": [ + { + "bbox": [ + 494, + 118, + 542, + 166 + ], + "lines": [ + { + "bbox": [ + 494, + 118, + 542, + 166 + ], + "spans": [ + { + "bbox": [ + 494, + 118, + 542, + 166 + ], + "type": "image", + "image_path": "db04ff57706227217ee1d6f47cb7266816b0ba4f085fd9e6006bb2474e6da60a.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 494, + 167, + 542, + 214 + ], + "blocks": [ + { + "bbox": [ + 494, + 167, + 542, + 214 + ], + "lines": [ + { + "bbox": [ + 494, + 167, + 542, + 214 + ], + "spans": [ + { + "bbox": [ + 494, + 167, + 542, + 214 + ], + "type": "image", + "image_path": "6d39d8aaad51264f7728292b9edff9f5e396d3edb9b2690a01e1d2a40339f9a7.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 494, + 215, + 542, + 262 + ], + "blocks": [ + { + "bbox": [ + 494, + 215, + 542, + 262 + ], + "lines": [ + { + "bbox": [ + 494, + 215, + 542, + 262 + ], + "spans": [ + { + "bbox": [ + 494, + 215, + 542, + 262 + ], + "type": "image", + "image_path": "a9744568492a477fd40883c68d35c53e14d0f492aa7d70b30af99eb9e734b291.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 504, + 265, + 528, + 274 + ], + "lines": [ + { + "bbox": [ + 504, + 265, + 528, + 274 + ], + "spans": [ + { + "bbox": [ + 504, + 265, + 528, + 274 + ], + "type": "text", + "content": "(h) HR" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 72, + 331, + 174, + 410 + ], + "blocks": [ + { + "bbox": [ + 55, + 289, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 555, + 312 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons of " + }, + { + "bbox": [ + 55, + 289, + 555, + 312 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 55, + 289, + 555, + 312 + ], + "type": "text", + "content": " (the 1st and 2nd rows) and " + }, + { + "bbox": [ + 55, + 289, + 555, + 312 + ], + "type": "inline_equation", + "content": "32 \\times" + }, + { + "bbox": [ + 55, + 289, + 555, + 312 + ], + "type": "text", + "content": " (the 3rd and 4th rows) rescaling results. Our method achieves higher semantic accuracy, such as more clearer text, more recognizable faces, and more realistic structures." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 331, + 174, + 410 + ], + "lines": [ + { + "bbox": [ + 72, + 331, + 174, + 410 + ], + "spans": [ + { + "bbox": [ + 72, + 331, + 174, + 410 + ], + "type": "image", + "image_path": "1fdf82630b05f0f11338afd1c89209e58becfdbed9239da1c0270373597491b6.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 175, + 331, + 280, + 410 + ], + "blocks": [ + { + "bbox": [ + 175, + 331, + 280, + 410 + ], + "lines": [ + { + "bbox": [ + 175, + 331, + 280, + 410 + ], + "spans": [ + { + "bbox": [ + 175, + 331, + 280, + 410 + ], + "type": "image", + "image_path": "136107dc95f13832cf2213ed1751495fecff08233853d76d0167ef3061ecb0f6.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 418, + 295, + 442 + ], + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 442 + ], + "type": "text", + "content": "Figure 6. The validation loss curves when removing a specific component from the DFRM." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "bbox": [ + 55, + 475, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 510 + ], + "type": "text", + "content": "of the SD priors can significantly improve performance of LPIPS and DISTS and greatly enhance the perceptual quality of the reconstruction results." + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 55, + 536, + 213, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 213, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 213, + 548 + ], + "type": "text", + "content": "4.3.3. Effectiveness of DFRM Design" + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Our DFRM includes two specific designs: the domain converter module and the pixel guidance module. To analyze their impact, we remove each component individually and plot the validation loss curves, as shown in Fig. 6. It can be observed that the introduction of the INN can reduce reconstruction error. Although the pixel guidance module does not improve reconstruction performance, our experiments show that it can enhance the fidelity and visual quality of the LR images, as demonstrated by the lower guidance loss in Fig. 6. Therefore, the design of our DFRM strikes a good balance between feature reconstruction and LR image quality. More experiments and the visual quality comparisons of LR images are presented in the supplementary materials." + } + ] + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 337, + 342, + 536, + 453 + ], + "blocks": [ + { + "bbox": [ + 337, + 342, + 536, + 453 + ], + "lines": [ + { + "bbox": [ + 337, + 342, + 536, + 453 + ], + "spans": [ + { + "bbox": [ + 337, + 342, + 536, + 453 + ], + "type": "image", + "image_path": "12d6e60aab704f14b0c01532d5e8675ad64b812228a3e09ef474ec91432817d4.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 468, + 520, + 480 + ], + "lines": [ + { + "bbox": [ + 348, + 468, + 520, + 480 + ], + "spans": [ + { + "bbox": [ + 348, + 468, + 520, + 480 + ], + "type": "text", + "content": "Figure 7. Ablation study about time alignment." + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 504, + 481, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 481, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 481, + 516 + ], + "type": "text", + "content": "4.3.4. Effectiveness of Time Alignment" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 313, + 520, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 688 + ], + "type": "text", + "content": "To validate the effectiveness of our proposed time-step alignment strategy, we retrain our model using fixed timesteps. This involves removing the TPM and time scheduler modules and employing a manually selected time-step to perform a single denoising step based on Eq. (9). Following the setting of previous work [35], we select fixed time-steps 1 and 999 for comparison. As shown in Fig. 7, when the time-step is set to a smaller value, the model achieves higher fidelity but lower perceptual quality, and the situation is reversed when a larger time-step is chosen. In contrast, our time-step alignment mechanism can dynamically determine the level of generative capability based on the image content, thereby achieving optimal performance in both fidelity and perceptual quality." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "In Fig. 8, we present a schematic illustration of the tiled inference process along with the predicted time-step map." + } + ] + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15600" + } + ] + } + ], + "index": 56 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 187 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 187 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 187 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 187 + ], + "type": "image", + "image_path": "7407938f94e679e96e92d8d91be4971cc70e9c2a24382da3d43da75011d6ed5e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 197, + 295, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 197, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 295, + 220 + ], + "type": "text", + "content": "Figure 8. Illustration of the tiled inference process and the predicted time-step map." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 241, + 296, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 296, + 302 + ], + "type": "text", + "content": "It can be observed that the time alignment mechanism effectively captures more complex regions in high-resolution images and assigns larger time steps to these areas. Details regarding the selection of patch size and stride length are provided in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 311, + 133, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 133, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 133, + 322 + ], + "type": "text", + "content": "4.4. Discussions" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 329, + 231, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 231, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 231, + 341 + ], + "type": "text", + "content": "4.4.1. Application to Image Compression" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 344, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 440 + ], + "type": "text", + "content": "Our model can be combined with existing lossless image compression methods to achieve extreme compression ratios. To validate this, we use PNG to compress LR images and compare it with JPEG at various bit-rates. As shown in Fig. 9, at the same bits per pixel (bpp), the image reconstruction quality of our model significantly surpasses that of the JPEG compression. These comparisons demonstrate the potential of our model for image compression." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 71, + 453, + 274, + 563 + ], + "blocks": [ + { + "bbox": [ + 71, + 453, + 274, + 563 + ], + "lines": [ + { + "bbox": [ + 71, + 453, + 274, + 563 + ], + "spans": [ + { + "bbox": [ + 71, + 453, + 274, + 563 + ], + "type": "image", + "image_path": "9fe5368702fcc26077d80f5fbcd88347ea06463541624be324166e00968d5859.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 570, + 295, + 592 + ], + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 592 + ], + "type": "text", + "content": "Figure 9. R-D curves of JPEG compression and our proposed model on the DIV2K validation set." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 614, + 224, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 224, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 224, + 626 + ], + "type": "text", + "content": "4.4.2. Inference Efficiency Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "In Table 3, we provide comparisons of inference time and memory consumption for representative methods. All tests are performed on an RTX 3090 using images with a resolution of " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ". It can be observed that flow-based models (IRN and HCFlow) require less GPU memory consumption and enable fast inference. In contrast, StableSR, which is based on multi-step diffusion models, demands" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": "significantly more inference time and larger GPU memory consumption. Although VQIR requires less memory, the nearest neighbor search process in VQGAN results in longer inference time. For one-step diffusion-based method S3Diff, due to its degradation-guided LoRA which needs to dynamically predict LoRA parameters during inference, it requires longer ingerence time and higher memory consumption. In contrast, our proposed TADM demonstrates moderate efficiency and scalability compared to other methods. In the future, we can utilize distillation techniques to further reduce the model size." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 235, + 552, + 266 + ], + "blocks": [ + { + "bbox": [ + 314, + 216, + 553, + 227 + ], + "lines": [ + { + "bbox": [ + 314, + 216, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 314, + 216, + 553, + 227 + ], + "type": "text", + "content": "Table 3. Comparison of inference time and memory consumption." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 235, + 552, + 266 + ], + "lines": [ + { + "bbox": [ + 318, + 235, + 552, + 266 + ], + "spans": [ + { + "bbox": [ + 318, + 235, + 552, + 266 + ], + "type": "table", + "html": "
IRNHCFlowVQIRStableSRS3DiffOurs
Inference Time (s)0.3590.3630.897135.22.1860.711
Memory Consumption (G)0.4980.0980.3446.3805.0252.598
", + "image_path": "29094e4ed95f51eee30b83dd26c7e297716cf84a69fe007772bead61fd47b8c6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 281, + 443, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 443, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 443, + 293 + ], + "type": "text", + "content": "4.4.3. Upper Bound of TADM" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 297, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 453 + ], + "type": "text", + "content": "Since our method performs the rescaling operation in the latent space of a pre-trained VAE, its performance upper bound is inherently constrained by the VAE. In Table 4, we present a performance comparison between VAE and two " + }, + { + "bbox": [ + 313, + 297, + 555, + 453 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 297, + 555, + 453 + ], + "type": "text", + "content": " rescaling models. For VAEs, we directly feed the high-resolution images to the encoder and decoder without performing any rescaling operations. This performance represents the upper bound of our method's performance across all rescaling factors. It can be observed that VAE significantly lags behind the latter. In the future, we could explore integrating stronger VAEs, such as those from SDXL, SD3, thereby enabling the TADM to be more applicable to small-scale rescaling tasks." + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 363, + 486, + 506, + 519 + ], + "blocks": [ + { + "bbox": [ + 331, + 466, + 537, + 478 + ], + "lines": [ + { + "bbox": [ + 331, + 466, + 537, + 478 + ], + "spans": [ + { + "bbox": [ + 331, + 466, + 537, + 478 + ], + "type": "text", + "content": "Table 4. Comparison of VAE and " + }, + { + "bbox": [ + 331, + 466, + 537, + 478 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 331, + 466, + 537, + 478 + ], + "type": "text", + "content": " rescaling methods." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 363, + 486, + 506, + 519 + ], + "lines": [ + { + "bbox": [ + 363, + 486, + 506, + 519 + ], + "spans": [ + { + "bbox": [ + 363, + 486, + 506, + 519 + ], + "type": "table", + "html": "
IRNHCFlowVAE
PSNR35.0735.2329.01
SSIM0.93180.93460.8046
", + "image_path": "69dd70094646137f2d06d6e7c6ed0b6e507b4d0f588963fe6d9daf5c5c784dcd.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "We propose a novel framework called TADM to address issues such as insufficient texture and semantic inaccuracies in extreme image rescaling. Specifically, we design a DFRM to achieve bidirectional mapping between HR features and LR images. Then, we utilize a pre-trained SD model to perform a single denoising step on the rescaled features, enhancing its perceptual quality. Considering the non-uniform quality of the rescaled latent features, we propose a novel time-step alignment strategy to achieve a balance between fidelity and perceptual quality. Both qualitative and quantitative experiments demonstrate the superiority of our approach." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15601" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 193, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 296, + 236 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 126-135, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 237, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 237, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 237, + 295, + 281 + ], + "type": "text", + "content": "[2] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2567-2581, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 282, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 282, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 61, + 282, + 295, + 326 + ], + "type": "text", + "content": "[3] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 295, + 380 + ], + "type": "text", + "content": "[4] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 10021-10030, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 381, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 381, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 381, + 296, + 435 + ], + "type": "text", + "content": "[5] Shuhang Gu, Andreas Lugmayr, Martin Danelljan, Manuel Fritsche, Julien Lamour, and Radu Timofte. Div8k: Diverse 8k resolution image dataset. In IEEE International Conference on Computer Vision Workshop, pages 3512-3516, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 480 + ], + "type": "text", + "content": "[6] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5197-5206, 2015. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 525 + ], + "type": "text", + "content": "[7] Zhaoyang Jia, Jiahao Li, Bin Li, Houqiang Li, and Yan Lu. Generative latent coding for ultra-low bitrate image compression. In IEEE Conference on Computer Vision and Pattern Recognition, pages 26088-26098, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 525, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 294, + 569 + ], + "type": "text", + "content": "[8] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling diffusion models into conditional gans. arXiv preprint arXiv:2405.05967, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 570, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 295, + 613 + ], + "type": "text", + "content": "[9] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In IEEE Conference on Computer Vision and Pattern Recognition, pages 4401-4410, 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 668 + ], + "type": "text", + "content": "[10] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In IEEE Conference on Computer Vision and Pattern Recognition, pages 8110-8119, 2020. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[11] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In IEEE International Conference on Computer Vision, pages 5148-5157, 2021. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62201403, in part by Hubei Provincial Natural Science Foundation of China under Grant 2024AFB549. The numerical calculations in this paper have been done on the supercomputing system in the Supercomputing Center of Wuhan University." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[12] Heewon Kim, Myungsub Choi, Bee Lim, and Kyoung Mu Lee. Task-aware image downscaling. In European Conference on Computer Vision, pages 399-414, 2018. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 152 + ], + "type": "text", + "content": "[13] Haoying Li, Yifan Yang, Meng Chang, Shiqi Chen, Huajun Feng, Zhihai Xu, Qi Li, and Yueting Chen. Srdiff: Single image super-resolution with diffusion probabilistic models. Neurocomputing, 479:47-59, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[14] Jinmin Li, Tao Dai, Jingyun Zhang, Kang Liu, Jun Wang, Shaoming Wang, Shu-Tao Xia, et al. Boundary-aware decoupled flow networks for realistic extreme rescaling. arXiv preprint arXiv:2405.02941, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[15] Shang Li, Guixuan Zhang, Zhengxiong Luo, Jie Liu, Zhi Zeng, and Shuwu Zhang. Approaching the limit of image rescaling via flow guidance. arXiv preprint arXiv:2111.05133, 2021. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[16] Yue Li, Dong Liu, Houqiang Li, Li Li, Zhu Li, and Feng Wu. Learning a convolutional neural network for image compact-resolution. IEEE Transactions on Image Processing, 28(3): 1092-1107, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 287, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 353 + ], + "type": "text", + "content": "[17] Jingyun Liang, Andreas Lugmayr, Kai Zhang, Martin Danelljan, Luc Van Gool, and Radu Timofte. Hierarchical conditional flow: A unified framework for image superresolution and image rescaling. In IEEE International Conference on Computer Vision, pages 4076-4085, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 408 + ], + "type": "text", + "content": "[18] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 136-144, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[19] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Yu Qiao, Wanli Ouyang, and Chao Dong. Diff-bir: Toward blind image restoration with generative diffusion prior. In European Conference on Computer Vision, pages 430-448, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 467, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 521 + ], + "type": "text", + "content": "[20] Siwei Ma, Xinfeng Zhang, Chuanmin Jia, Zhenghui Zhao, Shiqi Wang, and Shanshe Wang. Image and video compression with neural networks: A review. IEEE Transactions on Circuits and Systems for Video Technology, 30(6):1683-1698, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "type": "text", + "content": "[21] Zhihong Pan, Baopu Li, Dongliang He, Mingde Yao, Henhao Wu, Tianwei Lin, Xin Li, and Errui Ding. Towards bidirectional arbitrary image rescaling: Joint optimization and cycle idempotence. In IEEE Conference on Computer Vision and Pattern Recognition, pages 17389-17398, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "text", + "content": "[22] Chenyang Qi, Xin Yang, Ka Leong Cheng, Ying-Cong Chen, and Qifeng Chen. Real-time 6k image rescaling with rate-distortion optimization. In IEEE Conference on Computer Vision and Pattern Recognition, pages 14092-14101, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 679 + ], + "type": "text", + "content": "[23] Lucas Relic, Roberto Azevedo, Markus Gross, and Christopher Schroers. Lossy image compression with foundation diffusion models. In European Conference on Computer Vision, pages 303-319, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[24] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image super-resolution via iterative refinement. IEEE Transactions" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15602" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "text", + "content": "on Pattern Analysis and Machine Intelligence, 45(4):4713-4726, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 139 + ], + "type": "text", + "content": "[25] Wanjie Sun and Zhenzhong Chen. Learned image downscaling for upscaling using content adaptive resampler. IEEE Transactions on Image Processing, 29:4027-4040, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 185 + ], + "type": "text", + "content": "[26] George Toderici, Lucas Theis, Nick Johnston, Eirikur Agustsson, Fabian Mentzer, Johannes Ballé, Wenzhe Shi, and Radu Timofte. Clic 2020: Challenge on learned image compression, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "text", + "content": "[27] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 295, + 274 + ], + "type": "text", + "content": "[28] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. arXiv preprint arXiv:2305.07015, 2023. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 319 + ], + "type": "text", + "content": "[29] Luozhou Wang, Shuai Yang, Shu Liu, and Ying-cong Chen. Not all steps are created equal: Selective diffusion distillation for image manipulation. In IEEE International Conference on Computer Vision, pages 7472-7481, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 322, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 294, + 374 + ], + "type": "text", + "content": "[30] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In European Conference on Computer Vision Workshops, pages 0–0, 2018. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 441 + ], + "type": "text", + "content": "[31] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In IEEE Conference on Computer Vision and Pattern Recognition, pages 25796-25805, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 487 + ], + "type": "text", + "content": "[32] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 532 + ], + "type": "text", + "content": "[33] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(10):3365-3387, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 588 + ], + "type": "text", + "content": "[34] Hao Wei, Chenyang Ge, Zhiyuan Li, Xin Qiao, and Pengchao Deng. Towards extreme image rescaling with generative prior and invertible prior. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6181-6193, 2024. 1, 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "text", + "content": "[35] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. In Advances in Neural Information Processing Systems, pages 92529-92553, 2024. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[36] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In IEEE Conference on Computer Vision and Pattern Recognition, pages 25456-25467, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[37] Mingqing Xiao, Shuxin Zheng, Chang Liu, Zhouchen Lin, and Tie-Yan Liu. Invertible rescaling network and its ex" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 609 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "tensions. International Journal of Computer Vision, 131(1): 134-159, 2023. 1, 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[38] Bingna Xu, Yong Guo, Luoqian Jiang, Mianjie Yu, and Jian Chen. Downscaled representation matters: Improving image rescaling with collaborative downscaled images. In IEEE International Conference on Computer Vision, pages 12237-12247, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[39] Jinhai Yang, Mengxi Guo, Shijie Zhao, Junlin Li, and Li Zhang. Self-asymmetric invertible network for compression-aware image rescaling. In AAAI Conference on Artificial Intelligence, pages 3155-3163, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "text", + "content": "[40] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In European Conference on Computer Vision, pages 74-91, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[41] Zongsheng Yue, Jianyi Wang, and Chen Change Loy. Reshift: Efficient diffusion model for image superresolution by residual shifting. In Advances in Neural Information Processing Systems, pages 13294-13307, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[42] Zongsheng Yue, Kang Liao, and Chen Change Loy. Arbitrary-steps image super-resolution via diffusion inversion. arXiv preprint arXiv:2412.09013, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 321, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 363 + ], + "type": "text", + "content": "[43] Aiping Zhang, Zongsheng Yue, Renjing Pei, Wenqi Ren, and Xiaochun Cao. Degradation-guided one-step image super-resolution with diffusion priors. arXiv preprint arXiv:2409.17058, 2024. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "text", + "content": "[44] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv preprint arXiv:2203.03605, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[45] Min Zhang, Zhihong Pan, Xin Zhou, and C-C Jay Kuo. Enhancing image rescaling using dual latent variables in invertible neural network. In ACM International Conference on Multimedia, pages 5602-5610, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 456, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 508 + ], + "type": "text", + "content": "[46] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, pages 586-595, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 511, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 564 + ], + "type": "text", + "content": "[47] Zhixuan Zhong, Liangyu Chai, Yang Zhou, Bailin Deng, Jia Pan, and Shengfeng He. Faithful extreme rescaling via generative prior reciprocated invertible representations. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5708-5717, 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "type": "text", + "content": "[48] Yiming Zhu, Cairong Wang, Chenyu Dong, Ke Zhang, Hongyang Gao, and Chun Yuan. High-frequency normalizing flow for image rescaling. IEEE Transactions on Image Processing, 32:6223-6233, 2022. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15603" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/f0ba958e-31c7-4013-97d5-2552433cb070_content_list.json b/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/f0ba958e-31c7-4013-97d5-2552433cb070_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..40eec5d06bd2ea55220f78f0d976503368dd013a --- /dev/null +++ b/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/f0ba958e-31c7-4013-97d5-2552433cb070_content_list.json @@ -0,0 +1,1451 @@ +[ + { + "type": "text", + "text": "TinyViM: Frequency Decoupling for Tiny Hybrid Vision Mamba", + "text_level": 1, + "bbox": [ + 171, + 130, + 826, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaowen Ma * Zhenliang Ni * Xinghao Chen† Huawei Noah's Ark Lab", + "bbox": [ + 299, + 180, + 696, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xwma@zju.edu.cn, {nizhenliang2, xinghao.chen}@huawei.com", + "bbox": [ + 246, + 219, + 743, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mamba has shown great potential for computer vision due to its linear complexity in modeling the global context with respect to the input length. However, existing lightweight Mamba-based backbones cannot demonstrate performance that matches Convolution or Transformer-based methods. By observing, we find that simply modifying the scanning path in the image domain is not conducive to fully exploiting the potential of vision Mamba. In this paper, we first perform comprehensive spectral and quantitative analyses, and verify that the Mamba block mainly models low-frequency information under Convolution-Mamba hybrid architecture. Based on the analyses, we introduce a novel Laplace mixer to decouple the features in terms of frequency and input only the low-frequency components into the Mamba block. In addition, considering the redundancy of the features and the different requirements for high-frequency details and low-frequency global information at different stages, we introduce a frequency ramp inception, i.e., gradually reduce the input dimensions of the high-frequency branches, so as to efficiently trade-off the high-frequency and low-frequency components at different layers. By integrating mobile-friendly convolution and efficient Laplace mixer, we build a series of tiny hybrid vision Mamba called TinyViM. The proposed TinyViM achieves impressive performance on several downstream tasks including image classification, semantic segmentation, object detection and instance segmentation. In particular, TinyViM outperforms Convolution, Transformer and Mamba-based models with similar scales, and the throughput is about 2-3 times higher than that of other Mamba-based models. Code is available at https://github.com/xwmaxwma/TinyViM.", + "bbox": [ + 89, + 301, + 485, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 801, + 220, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, Mamba has gained widespread attention and application in the field of vision due to its advantages in long-sequence modeling [8]. Mamba is a selective struc", + "bbox": [ + 89, + 828, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a8656d77bb359f04e054d3a8191f64d6b442e0a356f64302bf4280292be95af9.jpg", + "image_caption": [ + "Figure 1. Comparison of GFLOPs, Throughput and accuracy between TinyViM and others. The top-1 accuracy is tested on ImageNet-1K and the throughput is measured on Nvidia V100 GPU with maximum power-of-two batch size that fits in memory. Compared to recent Mamba-based models such as EfficientV-Mamba, QuadMamba, and VMamba, TinyViM has higher top-1 accuracy with $2 \\times$ higher throughput and fewer GFLOPs." + ], + "image_footnote": [], + "bbox": [ + 517, + 268, + 718, + 479 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4ad6defaaaf37411d30ff5bd490200b612d065cb890789bfada110486644141d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 268, + 919, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tured state-space model that can overcome the limitations of local perception in Convolution neural networks [19] and the quadratic computational complexity of Transformers [3], thereby demonstrating strong capabilities in handling long-sequence data. Due to its advanced context extraction capabilities, Mamba has been widely applied in image classification tasks and has achieved advanced performance, such as Vim [56], VMamba [31], etc.", + "bbox": [ + 509, + 625, + 906, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vim [56] is one of the earliest works to apply Mamba to visual tasks, by introducing a bidirectional state-space model to process 2D image data. By introducing the Cross-Scan Module (CSM) and 2D Selective Scanning (SS2D) modules, VMamba [31] reduces the computational complexity from quadratic to linear while maintaining the global acceptance domain. MSVMamba [44] uses a multiscale 2D scanning method to apply Mamba to both original and down-sampled feature maps. This not only helps to learn long-distance dependencies, but also reduces compu", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, † Corresponding author", + "bbox": [ + 107, + 886, + 346, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23519", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5bd5024f51d23ddf4b352624596601f7d9fb9f70fb1daf4f2092893f4467bf9b.jpg", + "image_caption": [ + "Figure 2. Spectral analysis of Mamba. (a) Frequency magnitude $(14 \\times 14)$ from 8 output channels before- and after Mamba block. (b) Relative log amplitudes of Fourier transformed feature maps. The magnitude and amplitude are averaged over 384 samples. (a) and (b) show that Mamba block focuses on capturing low-frequency information under the Convolution-Mamba hybrid architecture." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 658, + 256 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fe08c43ef9675f8ddc3310f0a1780ac90a24720fce063b4ebef42fe5bdb16d68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 665, + 90, + 893, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tational costs. EfficientVMamba [39] is a lightweight Vision Mamba variant that reduces computational complexity through dilated-based selective scanning. Although these methods inherit the advantages of Mamba linear complexity and global receptive field, they do not show competitive performance with lightweight backbones based on Convolution [49, 53] and Transformer [14, 43, 48] in computationally limited and real-time deployment scenarios, as shown in Fig. 1. In fact, these methods that simply modify the scan path in the image domain do not utilize the full potential of visual Mamba [31, 56]. In this paper, we try to rethink the design of lightweight vision Mamba from the perspective of frequency decoupling, which is a reasonable and effective solution.", + "bbox": [ + 88, + 335, + 485, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through observation, we find that Mamba prefers to extract low-frequency features and ignores some high-frequency features. As shown in Fig. 2, we construct a baseline using mobile-friendly Convolution and vanilla Mamba, and perform spectral analysis of the features before and after the Mamba. Many previous works [26, 43, 50] have verified the necessity of using Convolution in lightweight backbones. After applying the Mamba block, the low-frequency component in the center of the two-dimensional spectrum map is highlighted, and the high-frequency component is suppressed. Therefore, to improve the efficiency of low-frequency modeling and keep the high-frequency features, we introduce an efficient Laplace pyramid to decompose the high- and low-frequency components of the features. Then, we input only the low-frequency components into the Mamba block to obtain a global receptive field, and for the high-frequency components, we use a reparameterized $3 \\times 3$ depth-wise Convolution to enhance high frequency information [5, 6]. It is worth noting that the size of the low frequency feature is much smaller than that of the original input feature, which can greatly reduce the computational complexity and improve the throughput of the Mamba. Therefore, the strategy of frequency decoupling and specif", + "bbox": [ + 91, + 553, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ically enhancement separately effectively improves the efficiency and the feature representation ability of the model.", + "bbox": [ + 511, + 335, + 903, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, previous models based on Convolution or Transformer have verified that: 1) deep neural networks have feature redundancy [15]; 2) deep neural networks tend to require more high-frequency information in the shallow layer and more global information in the deep layer [35, 45]. Therefore, we propose the frequency ramp inception to further improve the performance and efficiency of the model. Specifically, we split the feature map along the channels at each stage. Early stages prioritize high frequencies, reflected in a higher channel ratio for the high-frequency branch, while later stages concentrate on low frequencies, allowing for a greater channel allocation to low frequencies. This strategic distribution enables the network to learn features that are more suitable at each stage of processing effectively.", + "bbox": [ + 511, + 368, + 906, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By integrating frequency decoupling and frequency ramp inception, we propose a tiny and efficient Convolution-Mamba hybrid architecture TinyVim. We then conduct extensive experiments to validate the effectiveness of TinyViM, which achieves state-of-the-art image classification performance on ImageNet and advanced performance in downstream tasks such as detection and segmentation. Specifically, TinyViM has higher classification accuracy than other Convolution or Transformer-based backbones with a similar scale. In addition, compared to other Mamba-based backbones such as EfficientVMambaT, TinyViM has $2 \\times$ higher throughput and $2.7\\%$ higher accuracy, as shown in Fig. 1. Our contributions can be summarized as follows.", + "bbox": [ + 511, + 597, + 908, + 806 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We verify that Mamba mainly models low-frequency information under the Convolution-Mamba hybrid architecture based on qualitative analysis and quantitative comparison.", + "- We introduce an efficient Laplace mixer, which improves the modeling efficiency and performance of vi" + ], + "bbox": [ + 511, + 810, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23520", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sual Mamba through frequency decoupling and frequency ramp inception.", + "bbox": [ + 102, + 90, + 482, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We propose TinyViM by integrating the Laplace mixer and Convolution. Extensive experiments show that TinyViM outperforms other Convolutional, Transformer and Mamba-based models with a similar scale, and achieves a better trade-off between performance and efficiency.", + "bbox": [ + 91, + 121, + 482, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 227, + 232, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Efficient Generic Vision Backbones", + "text_level": 1, + "bbox": [ + 89, + 253, + 397, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Convolution has been widely employed for designing efficient vision backbones due to the inductive bias and efficient parallel processing on GPUs. Early works such as the MobileNet series [21, 22, 42] propose the depth-wise separable Convolution and the inverted residual blocks to further reduce the parameter and computational consumption of the plain Convolutional model [19, 46]. A series of subsequent works further improve the efficiency through depth-wise dilated Convolutions [38], network pruning [16], neural structure search [21], channel shuffling [33, 54] and structure reparameterization [48-50]. However, constrained by a limited receptive field, these efficient Convolutional networks lack global interactions between features, which degrades their performance. Although some works propose to increase the receptive field by increasing the size of the Convolutional kernel [2, 29], however, it imposes high memory access cost and computational cost, and significantly slows down the inference speed. Considering that Transformers are capable of achieving global attention but introduce quadratic complexity with respect to the resolution of the input image, a series of recent works have attempted to combine the advantages of CNNs and Transformers to develop more efficient models [26, 43, 48]. For example, and EfficientFormer only adds self-attention at more deep stages to capture global context, SwiftFormer adds effective additive attention at each stage of the model. The success of this hybrid design motivates us to explore whether a more advanced backbone can be constructed by integrating Convolution and Mamba.", + "bbox": [ + 89, + 276, + 483, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. State Space Models", + "text_level": 1, + "bbox": [ + 89, + 726, + 276, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a typical architectural paradigm for sequence-to-sequence transformations, state-space models (SSMs) play a pivotal role in handling long-range dependencies in sequence data [10, 11, 36, 47]. Early SSMs have challenges in capturing long-term dependency relationships and a unified understanding of memory [7]. To address these issues, Gu et al. enhance the ability of SSMs to capture extended dependencies using HiPPO initialization [9] and solve the continuous-time memory problem based on a linear state space layer (LSSL) [11]. However, these SSMs are still lim", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ited by computational requirements. Subsequent works propose a series of improvements including complex diagonal structures [12, 13], multiple-input multiple-output support [47], and diagonal-plus-low-rank arithmetic [17] to improve computational efficiency and generalization across tasks. In particular, these strategies have been extended to large representation models [4, 34, 37]. Recently Mamba [8] introduce a selection mechanism that outperforms Transformers on large-scale real data through a data-dependent SSM layer with linearly scaling sequence lengths.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The success of Mamba has caused the vision community to explore and apply it to various vision tasks including classification [31, 56], detection [51], segmentation [57], deraining [25] and multimodal learning [41]. In particular, ViM [56] and VMamba [31] propose bi-directional scanning and cross-scanning strategies to cope with the non-sequential structure of visual data, respectively. QuadMamba [51] proposes quad-tree scanning to capture local dependencies at different granularities. MSVMamba [44] proposes multiscale two-dimensional scanning, which efficiently mitigates the long-range forgetting problem and reduces the computational cost. A highly related work is EfficientVMamba [40], which improves the scanning efficiency of a model by an atrous-based selective scan approach. However, these works tend to show low throughput and fail to perform competitively with other advanced lightweight backbones based on Convolution or Transformers. This motivates us to design the TinyViM, which obtains satisfactory speed and performance based on Laplace frequency decoupling.", + "bbox": [ + 511, + 242, + 906, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 558, + 651, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "State Space Models (S4). Inspired by the continuous system, State space models (SSMs) are proposed in deep learning as common sequence models [10]. These models transform an input sequence $x(t) \\in \\mathbb{R}^{L \\times D}$ into an output sequence $y(t) \\in \\mathbb{R}^{L \\times D}$ by utilizing a learnable hidden state $h(t) \\in \\mathbb{R}^N$ . The process could be denoted as follows:", + "bbox": [ + 511, + 583, + 905, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} h ^ {\\prime} (t) = \\boldsymbol {A} h (t) + \\boldsymbol {B} x (t), \\tag {1} \\\\ y (t) = \\boldsymbol {C} h (t), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 685, + 905, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{A} \\in \\mathbb{R}^{N \\times N}$ is the evolution parameter, $\\mathbf{B}, \\mathbf{C} \\in \\mathbb{R}^{N \\times D}$ denote the learnable projection parameters, and $N$ is the state size.", + "bbox": [ + 511, + 733, + 905, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discretization. The above continuous-time SSMs are not well compatible with deep learning algorithms. Therefore, discretization is needed to align the model with the sampling frequency of the input signal to improve computational efficiency [11]. Following the previous work [13], given the sampling time scale parameter $\\Delta$ , the above continuous SSMs are discretized through zero-order hold rule, thus converting the continuous-time parameters $(A, B)$ to", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23521", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a2238a917e987242aee1d0cc9ccba8943687f34363faffc4d84de8965c94f748.jpg", + "image_caption": [ + "Figure 3. Overview of the proposed TinyViM, which has four stages and each stage consists of Local Blocks and TinyViM Blocks. The local block applies a reparameterized 3x3 Convolution to extract local features, and the TinyViM block is employed to capture the global context. The core component of the TinyViM block is the Laplace Mixer, which decouples the frequencies of the features based on an efficient Laplace pyramid and passes the state only for the low-frequencies. The enhanced low- and high-frequency components are then integrated based on a frequency ramp Inception structure. Thus, the high- and low-frequency components of different stages are appropriately trade-off and the efficiency is further improved. SS2D denotes 2D selective scanning, which is used in VMamba [31]." + ], + "image_footnote": [], + "bbox": [ + 130, + 85, + 885, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "their corresponding discrete counterparts $(\\overline{A},\\overline{B})$", + "bbox": [ + 89, + 428, + 419, + 444 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\overline {{\\boldsymbol {A}}} = e ^ {\\boldsymbol {\\Delta A}}, \\\\ - \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 448, + 480, + 473 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{B}} = (\\Delta A) ^ {- 1} (e ^ {\\Delta A} - I) \\cdot \\Delta B.\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 469, + 395, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, The discretized formulation of Eq. 1 is formulated as:", + "bbox": [ + 89, + 492, + 482, + 507 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {t} = \\bar {\\boldsymbol {A}} h _ {t - 1} + \\bar {\\boldsymbol {B}} x _ {t}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 518, + 482, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {t} = C h _ {t},\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 540, + 290, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\overline{A} \\in \\mathbb{R}^{N \\times N}$ , $\\overline{B} \\in \\mathbb{R}^{N \\times D}$ . In order to improve the computational efficiency, the iterative process described in Eq. 3 can be performed by the parallel computing mode of global convolution [8]:", + "bbox": [ + 89, + 558, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny = x * \\bar {K}, \\quad - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 626, + 482, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\overline{K} = (C\\overline{B}, C\\overline{A}\\overline{B}, \\dots, C\\overline{A}^{L-1}\\overline{B})$ (4)", + "bbox": [ + 138, + 643, + 483, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\ast$ denotes the convolution operation, and $\\overline{\\pmb{K}}\\in \\mathbb{R}^{L}$ serves as the kernel of the SSMs.", + "bbox": [ + 89, + 672, + 482, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Selective State Space Models (S6). Conventional SSMs (i.e., S4) have been implemented to capture sequence context under linear time complexity, despite the fact that they are constrained by static parameterization and cannot perform content-based reasoning. To address this problem, the selective state space model (i.e., Mamba [8]) has been proposed, which allows the model to selectively propagate or forget information based on the latitude of the current token along the length of the sequence by simply setting the parameters of the SSM as a function of the inputs. In S6, the parameters $B$ , $C$ , and $\\Delta$ are computed directly from the input sequence $x(t)$ , thus enabling sequence-aware parameterization.", + "bbox": [ + 89, + 704, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 426, + 604, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Analysis of Vanilla Mamba Block", + "text_level": 1, + "bbox": [ + 511, + 453, + 805, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As analyzed in Sec. 3, Mamba is able to model global context information based on linear complexity with respect to sequence length through pixel-by-pixel state transfer, which shows potential in lightweight model design. Moreover, considering the mobile-friendly and localized feature extraction capability of Convolution, a natural intuition arises: can we build a lightweight and efficient visual backbone based on Convolution and Mamba blocks? Although there have been some works [40] that attempt to do so, the performance and efficiency shown are not competitive with Convolution [49] or Transformer-based models [43, 48] so far.", + "bbox": [ + 511, + 476, + 906, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baseline Construction. In response to this intuition, we first build a baseline architecture using Convolution (the local block in Fig. 3) and a vanilla Mamba block (applying SS2D in VMamba [31] to adapt to the visual data), i.e., the Baseline in Table 1. Note that the baseline structure is identical to the final structure of TinyViM-S, which is designed with reference to [43], and specific structural configurations can be found in the Supplementary Material. The only difference between Baseline and TinyViM-S is that we replace the TinyViM blocks with SS2D-based vanilla Mamba blocks. We are surprised to find that this version achieves a Top-1 accuracy of $79.1\\%$ , which is significantly better than the purely Convolutional (i.e., Conv only) version. This can be explained by the fact that Mamba can improve the expressive ability of the model by capturing the global context at each stage. However, the low throughput limits the use of the model in real-time applications. We then attempt to", + "bbox": [ + 511, + 643, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23522", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3ae2cd0ef1e145e8edcc852a9bf91e8803e5c0d503f408670f19ac5699d8e689.jpg", + "table_caption": [ + "Table 1. Effect of frequency input Mamba block. Throughput is tested on a Nvidia V100 GPU with maximum power-of-two batch size that fits in memory. The Top-1 accuracy is the average of three experiments to exclude random fluctuations." + ], + "table_footnote": [], + "table_body": "
VariantsGMACsThroughputTop-1
Baseline0.96167379.1
Conv only0.88321277.5
Low only0.93257479.0
High only0.96137778.6
Low + High0.97150979.1
", + "bbox": [ + 130, + 156, + 441, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "improve the baseline based on qualitative and quantitative analysis.", + "bbox": [ + 89, + 263, + 482, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Analysis. We first perform a spectral analysis of the feature maps before and after inputting the Mamba block in the baseline, as shown in Fig. 2. It shows that the Mamba block has strong low-frequency capture ability under the Convolution-Mamba hybrid architecture, but loses high-frequency information, such as local edges and textures. This low-frequency preference impairs the performance and efficiency of vision Mamba because 1) Populating each stage with low-frequency information degrades the high-frequency component, weakening the fine-grained recognition ability of visual Mamba, and 2) Additional high-frequency component input to a Mamba block impairs the efficiency and parallel computation of the model. A feasible strategy for this phenomenon is that we decouple the high and low frequencies of the features and input only the low-frequency components of the features into the Mamba block for hidden state transfer. Therefore, we decouple the high and low frequencies of the feature maps based on the Laplace pyramid, which is a more efficient architecture compared to other methods such as the cascaded wavelet transform, to verify the effectiveness of our strategy.", + "bbox": [ + 91, + 295, + 483, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Analysis. We next set the inputs of the Mamba block for each stage to be low-frequency only, high-frequency only, high and low-frequency in parallel to explore the effect of the corresponding variants on the classification accuracy, respectively. As shown in Table 1, when only low frequencies are used as inputs, the model achieves $79.0\\%$ accuracy, reduces the computational consumption of the model and achieves 1.5 times higher throughput. This verifies that we can significantly improve the efficiency of the model without compromising the classification accuracy by using only low-frequency inputs to Mamba. Therefore, based on the above qualitative and quantitative analyses, we design the Laplace mixer to optimize Mamba in the construction of a lightweight vision backbone.", + "bbox": [ + 89, + 612, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Laplace Mixer", + "text_level": 1, + "bbox": [ + 89, + 832, + 241, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous work such as EfficientVMamba [40] uses dilated-based selective scanning to improve the efficiency of vanilla Mamba [8]. Despite reducing the number of input tokens,", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "EfficientVMamba performs unsatisfactorily on various visual tasks [40]. We argue that the potential of Mamba in lightweight backbones is not fully exploited simply by modifying the scanning path in the image domain. Based on the analysis in Sec. 4.1, we reconstruct the Mamba block based on the frequency decoupling and propose an elegant Laplace mixer. Furthermore, consider that in a general visual backbone, the bottom layer is more concerned with capturing high-frequency details, while the top layer is more concerned with modeling low-frequency global information [19]. As in humans, by capturing details in high-frequency components, the lower layers can capture the basic visual features and gradually collect local information to achieve a global understanding of the input. In particular, feature redundancy has been shown to exist in existing networks [15], and separate low-frequency and high-frequency processing for all channels is unnecessary. Therefore, we design a frequency ramp inception structure which reduces the channel size of the high-frequency branches as the depth of the network increases.", + "bbox": [ + 511, + 90, + 903, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, for the input feature $X \\in \\mathbb{R}^{H \\times W \\times D}$ , where H, W and D denote the height, width and channels of $X$ , we first split along the channel dimension and divide it into a low-frequency input $X_{l} \\in \\mathbb{R}^{H \\times W \\times \\alpha D}$ and a high-frequency input $X_{h} \\in \\mathbb{R}^{H \\times W \\times (1 - \\alpha)D}$ . $\\alpha$ denotes the partition coefficient, which determines the number of channels in the low-frequency branch. Then, for the low-frequency branch, we adopt a simple Laplace pyramid architecture. Specifically, we average pool $X_{l}$ to obtain the corresponding low-frequency components $X_{ll} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\alpha D}$ and shrink the number of input tokens. Next, we recover the resolution of $X_{ll}$ based on the nearest neighbor upsampling and perform pixel-by-pixel subtraction from $X_{l}$ to obtain the high-frequency component of $X_{l}$ , i.e., $X_{lh}$ . This process can be formulated as,", + "bbox": [ + 511, + 392, + 906, + 621 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {l l} = \\operatorname {P o o l} (X _ {l}), \\quad X _ {l h} = X _ {l} - \\operatorname {U p s a m p l e} (X _ {l l}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 628, + 906, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we concatenate $X_{lh}$ and $X_{h}$ along the channel to obtain the high-frequency input $X_{hh} \\in \\mathbb{R}^{H \\times W \\times D}$ , and apply a small depth-wise Convolution kernel, which usually tends to respond to the high frequencies in the input [5, 6], enhancing the high-frequency component of $X_{hh}$ . As for the low-frequency input $X_{ll}$ , we input it into the SS2D block for cross-scanning to capture the global context. The process of frequency separation processing can be described as,", + "bbox": [ + 511, + 652, + 908, + 776 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {l l} = \\operatorname {S S 2 D} \\left(X _ {l l}\\right), \\quad \\hat {X} _ {h h} = \\operatorname {R e p} _ {3} \\left(X _ {h h}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 787, + 906, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Rep}_3$ denotes the reparameterized $3\\times 3$ convolution. Finally, we sum up $\\hat{X}_{ll}$ and $\\hat{X}_{hh}$ of the corresponding channels at the element level, followed by a $1\\times 1$ convolution for fusion. Through the frequency decoupling process and frequency ramp inception, the Laplace mixer can effectively weigh and integrate the high and low-frequency", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23523", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0014ce45d9c7c1ac0438c98b9a135d13b619c2f625b61b4e19714deb04e40d3d.jpg", + "table_caption": [ + "Table 2. Classification performance on ImageNet-1K. Following [30], throughput is tested on Nvidia V100 GPU with maximum power-of-two batch size that fits in memory." + ], + "table_footnote": [], + "table_body": "
ModelConferenceTypeParams ↓ (M)GMACs ↓Throughput ↑ (im/s)Reso.Top-1 ↑ (%)
MobileOne-S1 [49]CVPR'23CNN4.80.8354522475.9
SwiftFormer-S [43]ICCV'23CNN+ViT6.11.0262622478.5
EfficientFormerV2-S0 [27]ICCV'23CNN+ViT3.50.498530075.7
EfficientVMamba-T [40]Arxiv'24Mamba6.11.0139622476.5
Vim-Ti [56]ICML'24Mamba7.01.561730076.1
QuadMamba-Li [51]NeurlIPS'24Mamba5.40.8119022474.2
MSVMamba-N [44]NeurlIPS'24Mamba7.00.9128322477.3
MobileMamba-S6 [18]CVPR'25CNN+Mamba15.00.7365022478.0
TinyViM-S-CNN+Mamba5.60.9256322479.2
PoolFormer-S12 [52]CVPR'22Pool12.02.0190222477.2
MobileOne-S3 [49]CVPR'23CNN10.11.9190022478.1
MobileOne-S4 [49]CVPR'23CNN14.83.0122322479.4
Agent-PVT-T [14]ECCV'24ViT11.62.0144722478.4
EfficientVMamba-S [40]Arxiv'24Mamba11.01.367422478.7
QuadMamba-T [51]NeurlIPS'24Mamba10.02.058522478.2
MSVMamba-M [44]NeurlIPS'24Mamba12.01.595722479.8
MobileMamba-B1 [18]CVPR'25CNN+Mamba17.11.1209725679.9
TinyViM-B-CNN+Mamba11.01.5185122481.2
PoolFormer-S36 [52]CVPR'22Pool31.05.266722481.4
InceptionNext-T [53]CVPR'24CNN28.04.298722482.3
Agent-PVT-S [14]ECCV'24ViT20.64.068622482.4
FastViT-SA24 [48]ICCV'23CNN+ViT20.63.886122482.6
EfficientVMamba-B [40]Arxiv'24Mamba33.04.058022481.8
LocalVMamba-T [23]Arxiv'24Mamba26.05.7-22482.7
ViM-S [56]ICML'24Mamba26.05.120122480.5
VMamba-T [31]NeurlIPS'24Mamba30.04.938322482.6
QuadMamba-S [51]NeurlIPS'24Mamba31.05.554122482.4
MobileMamba-B4 [18]CVPR'25CNN+Mamba17.14.351351282.5
TinyViM-L-CNN+Mamba31.74.784322483.3
", + "bbox": [ + 96, + 128, + 643, + 455 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e6733ab2d548048f5310c74df5dc89ac45d56b51195ae4d156e509f731c848f3.jpg", + "table_caption": [ + "Table 3. Ablation of frequency ramp inception design." + ], + "table_footnote": [], + "table_body": "
InceptionRampGMACsThroughputTop-1
XX0.94247879.2
✓X0.93257479.0
✓✓0.93256379.2
", + "bbox": [ + 663, + 121, + 908, + 176 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f658d47432438eeba0085e3f995d3ae8e06348cb4ca849e44dc5a67337fc52d6.jpg", + "table_caption": [ + "Table 4. Ablation of the partition coefficient $\\alpha$ of four stages." + ], + "table_footnote": [], + "table_body": "
α1α2α3α4GMACsThrou.Top-1
0.750.50.50.250.94259178.8
0.50.50.50.50.93257479.0
0.250.50.50.750.93256379.2
", + "bbox": [ + 665, + 212, + 906, + 263 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/34f3ffc9aa1fd7130e772436966c0e1d3033a4b06ed2179133093501b92058a8.jpg", + "table_caption": [ + "Table 5. Ablation of the convolution kernel size in Laplace Mixer." + ], + "table_footnote": [], + "table_body": "
kernelaxialGMACsThroughputTop-1
3X0.94251079.0
5✓0.93259879.0
7✓0.93256379.2
9✓0.94247979.2
", + "bbox": [ + 673, + 299, + 898, + 367 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3532c394d071fa8174b95e8804de0a61c0bb1b87f60dddf19fc3113e3fbb75ba.jpg", + "table_caption": [ + "Table 6. Ablation of re-parameterization." + ], + "table_footnote": [], + "table_body": "
VariantRepGMACsThrou.Top-1
TinyViM-S✓0.93256379.2
TinyViM-S✗0.93256379.2
TinyViM-L✓4.7284383.3
TinyViM-L✗4.7284383.1
", + "bbox": [ + 674, + 388, + 901, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "components of all layers. As a result, TinyViM achieves a better balance between accuracy and efficiency.", + "bbox": [ + 89, + 472, + 482, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Overall Architecture", + "text_level": 1, + "bbox": [ + 89, + 511, + 287, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TinyViM employs a similar multi-scale backbone design similar to previous works [26, 43]. As shown in Fig. 3, given an image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ , a simple stem is employed to output $X_{1} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times D_{1}}$ feature map. $H, W$ denote the height and width of the input image respectively, and $D_{1}$ is the channel number of $X_{1}$ . Note that the stem is implemented with two $3 \\times 3$ convolutions with a stride of 2. Then, the output feature maps are fed into the first stage, which begins with Local Blocks. The local block consists of a reparameterized $3 \\times 3$ convolution and a feed-forward network (FFN, implemented by two consecutive $1 \\times 1$ convolutions) that achieves local feature extraction and feature channel mixing, respectively. The local block can be formulated as,", + "bbox": [ + 88, + 532, + 482, + 744 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} _ {1} = \\operatorname {C o n v} _ {1} (\\operatorname {G e L U} (\\operatorname {C o n v} _ {1} (\\operatorname {R e p} _ {3} (X _ {1})))) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 755, + 482, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathrm{Conv}_1$ and $\\mathrm{Rep}_3$ denote the $1\\times 1$ and reparameterized $3\\times 3$ convolution, respectively. The features are then fed into the TinyViM block for global context-awareness. Similarly, the TinyViM block consists of a Laplace mixer and FFN for global feature extraction and channel mixing, respectively. The TinyViM Block can be described as,", + "bbox": [ + 89, + 782, + 482, + 875 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {X} ^ {\\prime} _ {1} = \\operatorname {L a p l a c e M i x e r} (\\hat {X} _ {1}) + \\hat {X} _ {1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 883, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {X} _ {1} = \\operatorname {C o n v} _ {1} \\left(\\operatorname {G e L U} \\left(\\operatorname {C o n v} _ {1} \\left(\\hat {X} ^ {\\prime} _ {1}\\right)\\right)\\right) + \\hat {X} ^ {\\prime} _ {1}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 486, + 906, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "There is a Patch Embedding layer between two consecutive stages for increasing the channel dimension and reducing the resolution of feature maps. Next, the resulting feature maps are subsequently fed into the second, third, and fourth stages, producing $\\frac{H}{8} \\times \\frac{W}{8} \\times D_2$ , $\\frac{H}{16} \\times \\frac{W}{16} \\times D_3$ and $\\frac{H}{32} \\times \\frac{W}{32} \\times D_4$ feature maps, respectively.", + "bbox": [ + 511, + 510, + 906, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 616, + 645, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the proposed TinyViM across four downstream tasks: classification on ImageNet-1K dataset [1], object detection and instance segmentation on MS-COCO 2017 dataset [28], and semantic segmentation on ADE20K dataset [55]. All experiments are implemented following the common training settings in [43]. Due to page limits, we provide a description of the dataset, the experimental setup, details of the model's architecture and more qualitative analysis in the Supplementary Material.", + "bbox": [ + 509, + 642, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Image Classification", + "text_level": 1, + "bbox": [ + 511, + 787, + 705, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate TinyViM's performance in image classification on ImageNet-1K [1] with some advanced Convolution, ViT and Mamba-based methods, and the comparison results are summarized in Table 2. With similar GMACs and throughput, TinyViM-S achieves a top-1 accuracy of $79.2\\%$ , outperforming SwiftFormer-S by $0.7\\%$ .", + "bbox": [ + 511, + 809, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23524", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In particular, TinyViM retains its performance advantages even at small and large scales. For example, TinyViMB achieves an accuracy of 81.2, which outperforms the convolution-based model MobileOne-S4 by $1.8\\%$ , the ViT-based model Agent-PVT-T by $2.8\\%$ , and the Mamba-based model MSVMamba-M by $1.4\\%$ , respectively. These comparison results show the excellent performance of TinyViM. In addition, we compare the visualisation of Effective Receptive Field (ERF) [32] with the classical lightweight backbone, as shown in Fig. 5. The comparison models include the pure Convolution-based MobileOne-S3 and the Convolution-ViT hybrid model SwiftFormer-S. It can be observed that TinyViM-S possesses a larger ERF at a similar size. This suggests that TinyViM is able to effectively model the global context through the Laplace Mixer, thus improving the semantic recognition ability of the model.", + "bbox": [ + 89, + 90, + 480, + 332 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In terms of computational efficiency, TinyViM-S achieves a throughput of 2574 images/s, which is significantly better than other visual Mamba models. For example, the throughput of EfficientVmamba is only 1396 images/s, which is only about half of the throughput of TinyViM. This huge efficiency advantage persists on TinyViM-B and TinyViM-L, where the throughput is 2.7 and 1.5 times higher than that of the similarly scaled EfficientVmamba, respectively. Compared with other convolutional and ViT-based models, TinyViM maintains its lead in performance while its throughput is highly competitive. It should be emphasized that this competitiveness is significantly improved on A100 GPUs. This is due to the fact that Mamba has a number of hardware-friendly improvements on the A100 such as triton, which are not yet supported on the V100 GPUs. In summary, the results show that TinyViM strikes a better balance between performance and efficiency than other state-of-the-art lightweight backbones, and offers high value in real-time applications.", + "bbox": [ + 91, + 333, + 482, + 621 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Object Detection and Instance Segmentation", + "text_level": 1, + "bbox": [ + 89, + 635, + 467, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the proposed TinyViM for object detection and instance segmentation using the Mask R-CNN [20] framework. As shown in Table 7, TinyViM outperforms competitive models such as the pooling-based model PoolFormer, the Transformer-based models FastVit and SwiftFormer, and the Mamba-based model EfficientVMamba in both $\\mathrm{AP}^{box}$ and $\\mathrm{AP}^{mask}$ metrics. Specifically, compared to recent models SwiftFormer-L1, TinyViM has $1.1\\%$ and $0.6\\%$ improvement on $\\mathrm{AP}^{box}$ and $\\mathrm{AP}^{mask}$ , respectively, and has a larger throughput. Compared to EfficientVMamba-S, TinyViM-B has more significant advantages, i.e., 3.0 and 2.0 improvement in $\\mathrm{AP}^{box}$ and $\\mathrm{AP}^{mask}$ , respectively, with a 1.7 times throughput. When comparing with larger models such as FastViT-SA24 and EfficientVMamba-b, TinyViM still has an advantage in throughput, and shows $2.5\\%$ and $0.8\\%$ improvements in $\\mathrm{AP}^{box}$ , respectively. These results", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "demonstrate the superiority of TinyViM when transferring to detection and instance segmentation tasks.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Semantic Segmentation", + "text_level": 1, + "bbox": [ + 511, + 132, + 730, + 148 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7 shows the semantic segmentation results of TinyViM as a backbone in comparison with other state-of-the-art backbones, where Semantic FPN [24] serves as the decode head. Specifically, compared with recent methods such as SwiftFormer, TinyVim obtains $0.8\\%$ and $0.3\\%$ improvement in base and large scale, respectively. Compared with the method FasterViT, TinyViM has more significant advantages, with $3.9\\%$ and $3.2\\%$ improvements, respectively. The experimental results verify the effectiveness of TinyViM in semantic segmentation tasks.", + "bbox": [ + 511, + 155, + 903, + 306 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 316, + 678, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of the Frequency Decoupling. In order to investigate the effectiveness of the frequency decoupling strategy, we perform spectral visualization of the feature maps of the low-frequency and high-frequency branches, respectively. As shown in Fig. 4, by applying Mamba, the feature maps have higher concentrations at low frequencies; by applying depth-wise convolution, the high frequencies of the model are effectively preserved. This proves that the Laplace mixer can effectively decouple the high and low frequency components of the features and specially enhance them respectively.", + "bbox": [ + 511, + 340, + 903, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of the Frequency Ramp Inception. We explore the effectiveness of the frequency ramp inception design, as shown in Table 3. We set up three experimental variants: variant 1, we implement Laplace frequency decoupling for all channels of the feature map, and input low and high frequencies into Mamba and depth-wise convolution, respectively; variant 2, we adopt a non-ramp structure, i.e., we set the segmentation coefficient $\\alpha$ of the four stages to be $1/2$ ; and variant 3, we gradually increase $\\alpha$ with the increase of the stage of the model. In the implementation of this paper, the $\\alpha$ of the four stages are 0.25, 0.5, 0.5 and 0.75, respectively. It can be observed that, even in the small size, the throughput of variant 1 is significantly reduced and more GMACs are required. Variant 2, although more efficient, has a decreased accuracy due to the lack of adjusting the weights of the low-frequency and high-frequency components in different stages. We also explore the inverse ramp structure, as shown in Table 4. The frequency ramp used in variant 3 effectively weighs and integrates the high and low frequency information, thus helping the model to achieve a better balance between performance and efficiency.", + "bbox": [ + 511, + 507, + 903, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of the convolution kernel size in Laplace Mixer. ViM and VMamba use a $3 \\times 3$ depthwise convolution before scanning. However, a larger convolution kernel is beneficial for parameters (i.e., $B$ , $C$ ) generation due to the large receptive fields, although it impairs the efficiency of", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23525", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0616b78a67bd2b407dd139d7eab16940aef7c4b3b7732106bc12257842a9ca92.jpg", + "table_caption": [ + "Table 7. Object detection & instance segmentation results on MS COCO 2017 with the Mask RCNN framework. Semantic segmentation results on ADE20K by integrating models into Semantic FPN." + ], + "table_footnote": [], + "table_body": "
BackboneThroughput ↑ (img/s)Object DetectionInstance SegmentationSemantic mIoU
APbox50APbox75APmaskAPmask50APmask75
EfficientVMamba-T [40]16835.657.738.033.254.435.1-
TinyViM-S26238.761.141.335.757.137.938.9
PoolFormer-S12 [52]17137.359.040.134.655.836.937.2
EfficientFormer-L1 [26]19737.960.341.035.457.337.338.9
FastViT-SA12 [48]16238.960.542.235.957.638.138.0
SwiftFormer-L1 [43]17441.263.244.838.160.240.741.1
EfficientVMamba-S [40]10439.361.842.636.758.939.2-
MobileMamba-B1 [18]-40.661.843.837.458.939.940.7
TinyViM-B18042.364.246.338.761.141.341.9
PoolFormer-S36 [52]8741.063.144.837.760.140.042.0
EfficientFormer-L3 [26]11741.463.944.738.161.040.443.5
FastViT-SA24 [48]11042.063.545.838.060.540.541.0
SwiftFormer-L3 [43]12042.764.446.739.161.741.843.9
EfficientVMamba-B [40]9043.766.247.940.263.342.9-
MobileMamba-B4 [18]-40.161.843.036.958.639.242.5
TinyViM-L11144.566.448.640.763.643.844.2
", + "bbox": [ + 133, + 121, + 861, + 383 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3c628bfb17906b90abdda1e5e6be0efa06b04bdcc51fe75ac10ea714f968e61.jpg", + "image_caption": [ + "Figure 4. Fourier spectrum of TinyViM for the Low frequency branch and High frequency branch. It can be observed that we achieve the decoupling of the high and low frequency components of the features and specifically enhance them separately." + ], + "image_footnote": [], + "bbox": [ + 101, + 400, + 480, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the model. Therefore, we introduce axial depthwise convolution, which increases the convolution's receptive field along the scan path without decreasing the model efficiency. We also explore $3 \\times 3$ depthwise convolution and axial convolution with kernel sizes 5 and 9, as shown in Table 5. The model achieves the best balance between performance and efficiency when the kernel size is 7.", + "bbox": [ + 89, + 632, + 482, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation of re-parameterization. As shown in Table 6, the reparameterization has no improvement on TinyViM-S, but brings an increase of 0.2 to TinyViM-L. Considering that the reparameterization technique does not harm the inference efficiency of the model, we apply it for TinyViM.", + "bbox": [ + 89, + 739, + 482, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 832, + 209, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motivated by Mamba's ability to achieve global context modeling based on linear complexity, we construct a tiny and efficient hybrid vision Mamba TinyViM by combin-", + "bbox": [ + 89, + 857, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac8bf6bfc4e3ab9a98b42cdf553e5b9a955acd83870c3f10800bdf0ada958cb6.jpg", + "image_caption": [ + "Figure 5. The ERF of MobileOne, SwiftFormer and TinyViM. Our TinyViM effectively obtains large ERFs with Laplace Mixer." + ], + "image_footnote": [], + "bbox": [ + 527, + 398, + 890, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing the mobile-friendly convolution and Mamba in this paper. In contrast to previous vision Mamba efforts that simply modify scan paths in the image domain, whose performance could not compete with Convolution or Transformer based models with similar scale, TinyViM attempts to build a lightweight backbone from a frequency decoupling perspective. Specifically, we verify that the Mamba block mainly models low-frequency information in a Convolution-Mamba hybrid architecture after spectral analysis and quantitative experiments. To this end, we design a frequency decoupling strategy to preserve the high-frequency information and improve the low-frequency modeling efficiency of Mamba. In particular, we introduce a frequency ramp inception to weigh the different strengths of the low-frequency and high-frequency components at different stages of the model, so as to help the model achieve a better balance between accuracy and efficiency. Extensive experiments have shown that TinyViM outperforms other Convolution, Transformer and Mamba-based models. In future work, we plan to apply it as a lightweight backbone to replace the image encoder in SAM to explore the potential of vision Mamba in multi-modal scenarios.", + "bbox": [ + 511, + 594, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23526", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 118, + 187, + 133 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6", + "[2] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11963-11975, 2022. 3", + "[3] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1", + "[4] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022.3", + "[5] Paul Gavrikov and Janis Keuper. Can biases in imagenet models explain generalization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22184-22194, 2024. 2, 5", + "[6] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, 2022. 2, 5", + "[7] Alex Graves. Long Short-Term Memory, pages 37-45. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. 3", + "[8] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 3, 4, 5", + "[9] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487, 2020. 3", + "[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3", + "[11] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3", + "[12] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 3", + "[13] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 3", + "[14] Dongchen Han, Tianzhu Ye, Yizeng Han, Zhuofan Xia, Siyuan Pan, Pengfei Wan, Shiji Song, and Gao Huang. Agent" + ], + "bbox": [ + 93, + 143, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "attention: On the integration of softmax and linear attention. In Computer Vision - ECCV 2024, pages 124-140, Cham, 2025. Springer Nature Switzerland. 2, 6", + "[15] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5", + "[16] Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding, 2016. 3", + "[17] Ramin Hasani, Mathias Lechner, Tsun-Hsuan Wang, Makram Chahine, Alexander Amini, and Daniela Rus. Liquid structural state-space models. arXiv preprint arXiv:2209.12951, 2022.3", + "[18] Haoyang He, Jiangning Zhang, Yuxuan Cai, Hongxu Chen, Xiaobin Hu, Zhenye Gan, Yabiao Wang, Chengjie Wang, Yunsheng Wu, and Lei Xie. Mobilemamba: Lightweight multi-receptive visual mamba network, 2024. 6, 8", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 3, 5", + "[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7", + "[21] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, and Hartwig Adam. Searching for mobilenetv3. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3", + "[22] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications, 2017. 3", + "[23] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan, 2024. 6", + "[24] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollar. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019. 7", + "[25] Dong Li, Yidi Liu, Xueyang Fu, Senyan Xu, and Zheng-Jun Zha. Fourier learning integration with state space models for image deraining, 2024. 3", + "[26] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35: 12934-12949, 2022. 2, 3, 6, 8", + "[27] Yanyu Li, Ju Hu, Yang Wen, Georgios Evangelidis, Kamyar Salahi, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Rethinking vision transformers for mobilenet size and speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16889-16900, 2023. 6", + "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23527", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6", + "[29] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Tommi Kärkkäinen, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity, 2023. 3", + "[30] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14430, 2023. 6", + "[31] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 1, 2, 3, 4, 6", + "[32] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 7", + "[33] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3", + "[34] Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer. Mega: moving average equipped gated attention. arXiv preprint arXiv:2209.10655, 2022. 3", + "[35] Xiaowen Ma, Rongrong Lian, Zhenkai Wu, Renxiang Guan, Tingfeng Hong, Mengjiao Zhao, Mengting Ma, Jiangtao Nie, Zhenhong Du, Siyang Song, and Wei Zhang. A novel scene coupling semantic mask network for remote sensing image segmentation. ISPRS Journal of Photogrammetry and Remote Sensing, 221:44-63, 2025. 2", + "[36] Xiaowen Ma, Zhen-Liang Ni, Shuai Xiao, and Xinghao Chen. Timepro: Efficient multivariate long-term time series forecasting with variable- and time-aware hyper-state. In Forty-second International Conference on Machine Learning, 2025. 3", + "[37] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 3", + "[38] Sachin Mehta, Mohammad Rastegari, Linda Shapiro, and Hannaneh Hajishirzi. Espnetv2: A light-weight, power efficient, and general purpose convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[39] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024. 2", + "[40] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba, 2024. 3, 4, 5, 6, 8", + "[41] Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba:" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Exploring state space models for multimodal learning, 2024. 3", + "[42] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition* (CVPR), 2018. 3", + "[43] Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Swiftformer: Efficient additive attention for transformer-based real-time mobile vision applications. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17425-17436, 2023. 2, 3, 4, 6, 8", + "[44] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. arXiv preprint arXiv:2405.14174, 2024. 1, 3, 6", + "[45] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xin chao Wang, and Shuicheng Yan. Inception transformer. In Advances in Neural Information Processing Systems, pages 23495-23509. Curran Associates, Inc., 2022. 2", + "[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition, 2015. 3", + "[47] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 3", + "[48] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Fastvit: A fast hybrid vision transformer using structural reparameterization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5785-5795, 2023. 2, 3, 4, 6, 8", + "[49] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7907-7917, 2023. 2, 4, 6", + "[50] Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Repvit: Revisiting mobile cnn from vit perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15909-15920, 2024. 2, 3", + "[51] Fei Xie, Weijia Zhang, Zhongdao Wang, and Chao Ma. Quadmamba: Learning quadtree-based selective scan for visual state space model, 2024. 3, 6", + "[52] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10819-10829, 2022. 6, 8", + "[53] Weihao Yu, Pan Zhou, Shuicheng Yan, and Xinchao Wang. Inceptionnext: When inception meets convnext. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5672-5683, 2024. 2, 6", + "[54] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23528", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6", + "[56] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In *Forty-first International Conference on Machine Learning*. 1, 2, 3, 6", + "[57] Qinfeng Zhu, Yuanzhi Cai, Yuan Fang, Yihan Yang, Cheng Chen, Lei Fan, and Anh Nguyen. Samba: Semantic segmentation of remotely sensed images with state space model, 2024. 3" + ], + "bbox": [ + 91, + 90, + 480, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23529", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/f0ba958e-31c7-4013-97d5-2552433cb070_model.json b/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/f0ba958e-31c7-4013-97d5-2552433cb070_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6e2c79259b3706fb1d94617737716a0ddfffb740 --- /dev/null +++ b/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/f0ba958e-31c7-4013-97d5-2552433cb070_model.json @@ -0,0 +1,2125 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.172, + 0.131, + 0.828, + 0.154 + ], + "angle": 0, + "content": "TinyViM: Frequency Decoupling for Tiny Hybrid Vision Mamba" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.181, + 0.697, + 0.216 + ], + "angle": 0, + "content": "Xiaowen Ma * Zhenliang Ni * Xinghao Chen† Huawei Noah's Ark Lab" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.22, + 0.745, + 0.234 + ], + "angle": 0, + "content": "xwma@zju.edu.cn, {nizhenliang2, xinghao.chen}@huawei.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.486, + 0.771 + ], + "angle": 0, + "content": "Mamba has shown great potential for computer vision due to its linear complexity in modeling the global context with respect to the input length. However, existing lightweight Mamba-based backbones cannot demonstrate performance that matches Convolution or Transformer-based methods. By observing, we find that simply modifying the scanning path in the image domain is not conducive to fully exploiting the potential of vision Mamba. In this paper, we first perform comprehensive spectral and quantitative analyses, and verify that the Mamba block mainly models low-frequency information under Convolution-Mamba hybrid architecture. Based on the analyses, we introduce a novel Laplace mixer to decouple the features in terms of frequency and input only the low-frequency components into the Mamba block. In addition, considering the redundancy of the features and the different requirements for high-frequency details and low-frequency global information at different stages, we introduce a frequency ramp inception, i.e., gradually reduce the input dimensions of the high-frequency branches, so as to efficiently trade-off the high-frequency and low-frequency components at different layers. By integrating mobile-friendly convolution and efficient Laplace mixer, we build a series of tiny hybrid vision Mamba called TinyViM. The proposed TinyViM achieves impressive performance on several downstream tasks including image classification, semantic segmentation, object detection and instance segmentation. In particular, TinyViM outperforms Convolution, Transformer and Mamba-based models with similar scales, and the throughput is about 2-3 times higher than that of other Mamba-based models. Code is available at https://github.com/xwmaxwma/TinyViM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.222, + 0.818 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.829, + 0.485, + 0.875 + ], + "angle": 0, + "content": "In recent years, Mamba has gained widespread attention and application in the field of vision due to its advantages in long-sequence modeling [8]. Mamba is a selective struc" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.27, + 0.72, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.27, + 0.92, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.496, + 0.908, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Comparison of GFLOPs, Throughput and accuracy between TinyViM and others. The top-1 accuracy is tested on ImageNet-1K and the throughput is measured on Nvidia V100 GPU with maximum power-of-two batch size that fits in memory. Compared to recent Mamba-based models such as EfficientV-Mamba, QuadMamba, and VMamba, TinyViM has higher top-1 accuracy with \\(2 \\times\\) higher throughput and fewer GFLOPs." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.626, + 0.907, + 0.748 + ], + "angle": 0, + "content": "tured state-space model that can overcome the limitations of local perception in Convolution neural networks [19] and the quadratic computational complexity of Transformers [3], thereby demonstrating strong capabilities in handling long-sequence data. Due to its advanced context extraction capabilities, Mamba has been widely applied in image classification tasks and has achieved advanced performance, such as Vim [56], VMamba [31], etc." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Vim [56] is one of the earliest works to apply Mamba to visual tasks, by introducing a bidirectional state-space model to process 2D image data. By introducing the Cross-Scan Module (CSM) and 2D Selective Scanning (SS2D) modules, VMamba [31] reduces the computational complexity from quadratic to linear while maintaining the global acceptance domain. MSVMamba [44] uses a multiscale 2D scanning method to apply Mamba to both original and down-sampled feature maps. This not only helps to learn long-distance dependencies, but also reduces compu" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.348, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution, † Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23519" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.092, + 0.66, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.092, + 0.895, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.274, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Spectral analysis of Mamba. (a) Frequency magnitude \\((14 \\times 14)\\) from 8 output channels before- and after Mamba block. (b) Relative log amplitudes of Fourier transformed feature maps. The magnitude and amplitude are averaged over 384 samples. (a) and (b) show that Mamba block focuses on capturing low-frequency information under the Convolution-Mamba hybrid architecture." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.486, + 0.547 + ], + "angle": 0, + "content": "tational costs. EfficientVMamba [39] is a lightweight Vision Mamba variant that reduces computational complexity through dilated-based selective scanning. Although these methods inherit the advantages of Mamba linear complexity and global receptive field, they do not show competitive performance with lightweight backbones based on Convolution [49, 53] and Transformer [14, 43, 48] in computationally limited and real-time deployment scenarios, as shown in Fig. 1. In fact, these methods that simply modify the scan path in the image domain do not utilize the full potential of visual Mamba [31, 56]. In this paper, we try to rethink the design of lightweight vision Mamba from the perspective of frequency decoupling, which is a reasonable and effective solution." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.554, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Through observation, we find that Mamba prefers to extract low-frequency features and ignores some high-frequency features. As shown in Fig. 2, we construct a baseline using mobile-friendly Convolution and vanilla Mamba, and perform spectral analysis of the features before and after the Mamba. Many previous works [26, 43, 50] have verified the necessity of using Convolution in lightweight backbones. After applying the Mamba block, the low-frequency component in the center of the two-dimensional spectrum map is highlighted, and the high-frequency component is suppressed. Therefore, to improve the efficiency of low-frequency modeling and keep the high-frequency features, we introduce an efficient Laplace pyramid to decompose the high- and low-frequency components of the features. Then, we input only the low-frequency components into the Mamba block to obtain a global receptive field, and for the high-frequency components, we use a reparameterized \\(3 \\times 3\\) depth-wise Convolution to enhance high frequency information [5, 6]. It is worth noting that the size of the low frequency feature is much smaller than that of the original input feature, which can greatly reduce the computational complexity and improve the throughput of the Mamba. Therefore, the strategy of frequency decoupling and specif" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.905, + 0.366 + ], + "angle": 0, + "content": "ically enhancement separately effectively improves the efficiency and the feature representation ability of the model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.907, + 0.595 + ], + "angle": 0, + "content": "In addition, previous models based on Convolution or Transformer have verified that: 1) deep neural networks have feature redundancy [15]; 2) deep neural networks tend to require more high-frequency information in the shallow layer and more global information in the deep layer [35, 45]. Therefore, we propose the frequency ramp inception to further improve the performance and efficiency of the model. Specifically, we split the feature map along the channels at each stage. Early stages prioritize high frequencies, reflected in a higher channel ratio for the high-frequency branch, while later stages concentrate on low frequencies, allowing for a greater channel allocation to low frequencies. This strategic distribution enables the network to learn features that are more suitable at each stage of processing effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.909, + 0.807 + ], + "angle": 0, + "content": "By integrating frequency decoupling and frequency ramp inception, we propose a tiny and efficient Convolution-Mamba hybrid architecture TinyVim. We then conduct extensive experiments to validate the effectiveness of TinyViM, which achieves state-of-the-art image classification performance on ImageNet and advanced performance in downstream tasks such as detection and segmentation. Specifically, TinyViM has higher classification accuracy than other Convolution or Transformer-based backbones with a similar scale. In addition, compared to other Mamba-based backbones such as EfficientVMambaT, TinyViM has \\(2 \\times\\) higher throughput and \\(2.7\\%\\) higher accuracy, as shown in Fig. 1. Our contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.87 + ], + "angle": 0, + "content": "- We verify that Mamba mainly models low-frequency information under the Convolution-Mamba hybrid architecture based on qualitative analysis and quantitative comparison." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- We introduce an efficient Laplace mixer, which improves the modeling efficiency and performance of vi" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23520" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "sual Mamba through frequency decoupling and frequency ramp inception." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.122, + 0.483, + 0.212 + ], + "angle": 0, + "content": "- We propose TinyViM by integrating the Laplace mixer and Convolution. Extensive experiments show that TinyViM outperforms other Convolutional, Transformer and Mamba-based models with a similar scale, and achieves a better trade-off between performance and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.228, + 0.233, + 0.244 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.254, + 0.398, + 0.269 + ], + "angle": 0, + "content": "2.1. Efficient Generic Vision Backbones" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.277, + 0.485, + 0.715 + ], + "angle": 0, + "content": "Convolution has been widely employed for designing efficient vision backbones due to the inductive bias and efficient parallel processing on GPUs. Early works such as the MobileNet series [21, 22, 42] propose the depth-wise separable Convolution and the inverted residual blocks to further reduce the parameter and computational consumption of the plain Convolutional model [19, 46]. A series of subsequent works further improve the efficiency through depth-wise dilated Convolutions [38], network pruning [16], neural structure search [21], channel shuffling [33, 54] and structure reparameterization [48-50]. However, constrained by a limited receptive field, these efficient Convolutional networks lack global interactions between features, which degrades their performance. Although some works propose to increase the receptive field by increasing the size of the Convolutional kernel [2, 29], however, it imposes high memory access cost and computational cost, and significantly slows down the inference speed. Considering that Transformers are capable of achieving global attention but introduce quadratic complexity with respect to the resolution of the input image, a series of recent works have attempted to combine the advantages of CNNs and Transformers to develop more efficient models [26, 43, 48]. For example, and EfficientFormer only adds self-attention at more deep stages to capture global context, SwiftFormer adds effective additive attention at each stage of the model. The success of this hybrid design motivates us to explore whether a more advanced backbone can be constructed by integrating Convolution and Mamba." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.277, + 0.743 + ], + "angle": 0, + "content": "2.2. State Space Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As a typical architectural paradigm for sequence-to-sequence transformations, state-space models (SSMs) play a pivotal role in handling long-range dependencies in sequence data [10, 11, 36, 47]. Early SSMs have challenges in capturing long-term dependency relationships and a unified understanding of memory [7]. To address these issues, Gu et al. enhance the ability of SSMs to capture extended dependencies using HiPPO initialization [9] and solve the continuous-time memory problem based on a linear state space layer (LSSL) [11]. However, these SSMs are still lim" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "ited by computational requirements. Subsequent works propose a series of improvements including complex diagonal structures [12, 13], multiple-input multiple-output support [47], and diagonal-plus-low-rank arithmetic [17] to improve computational efficiency and generalization across tasks. In particular, these strategies have been extended to large representation models [4, 34, 37]. Recently Mamba [8] introduce a selection mechanism that outperforms Transformers on large-scale real data through a data-dependent SSM layer with linearly scaling sequence lengths." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.907, + 0.545 + ], + "angle": 0, + "content": "The success of Mamba has caused the vision community to explore and apply it to various vision tasks including classification [31, 56], detection [51], segmentation [57], deraining [25] and multimodal learning [41]. In particular, ViM [56] and VMamba [31] propose bi-directional scanning and cross-scanning strategies to cope with the non-sequential structure of visual data, respectively. QuadMamba [51] proposes quad-tree scanning to capture local dependencies at different granularities. MSVMamba [44] proposes multiscale two-dimensional scanning, which efficiently mitigates the long-range forgetting problem and reduces the computational cost. A highly related work is EfficientVMamba [40], which improves the scanning efficiency of a model by an atrous-based selective scan approach. However, these works tend to show low throughput and fail to perform competitively with other advanced lightweight backbones based on Convolution or Transformers. This motivates us to design the TinyViM, which obtains satisfactory speed and performance based on Laplace frequency decoupling." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.653, + 0.573 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.675 + ], + "angle": 0, + "content": "State Space Models (S4). Inspired by the continuous system, State space models (SSMs) are proposed in deep learning as common sequence models [10]. These models transform an input sequence \\( x(t) \\in \\mathbb{R}^{L \\times D} \\) into an output sequence \\( y(t) \\in \\mathbb{R}^{L \\times D} \\) by utilizing a learnable hidden state \\( h(t) \\in \\mathbb{R}^N \\). The process could be denoted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.686, + 0.906, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} h ^ {\\prime} (t) = \\boldsymbol {A} h (t) + \\boldsymbol {B} x (t), \\tag {1} \\\\ y (t) = \\boldsymbol {C} h (t), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.734, + 0.906, + 0.779 + ], + "angle": 0, + "content": "where \\(\\mathbf{A} \\in \\mathbb{R}^{N \\times N}\\) is the evolution parameter, \\(\\mathbf{B}, \\mathbf{C} \\in \\mathbb{R}^{N \\times D}\\) denote the learnable projection parameters, and \\(N\\) is the state size." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Discretization. The above continuous-time SSMs are not well compatible with deep learning algorithms. Therefore, discretization is needed to align the model with the sampling frequency of the input signal to improve computational efficiency [11]. Following the previous work [13], given the sampling time scale parameter \\(\\Delta\\), the above continuous SSMs are discretized through zero-order hold rule, thus converting the continuous-time parameters \\((A, B)\\) to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23521" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.087, + 0.887, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.319, + 0.907, + 0.404 + ], + "angle": 0, + "content": "Figure 3. Overview of the proposed TinyViM, which has four stages and each stage consists of Local Blocks and TinyViM Blocks. The local block applies a reparameterized 3x3 Convolution to extract local features, and the TinyViM block is employed to capture the global context. The core component of the TinyViM block is the Laplace Mixer, which decouples the frequencies of the features based on an efficient Laplace pyramid and passes the state only for the low-frequencies. The enhanced low- and high-frequency components are then integrated based on a frequency ramp Inception structure. Thus, the high- and low-frequency components of different stages are appropriately trade-off and the efficiency is further improved. SS2D denotes 2D selective scanning, which is used in VMamba [31]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.429, + 0.42, + 0.445 + ], + "angle": 0, + "content": "their corresponding discrete counterparts \\((\\overline{A},\\overline{B})\\)" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.449, + 0.482, + 0.474 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\overline {{\\boldsymbol {A}}} = e ^ {\\boldsymbol {\\Delta A}}, \\\\ - \\end{array} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.47, + 0.397, + 0.487 + ], + "angle": 0, + "content": "\\[\n\\overline {{B}} = (\\Delta A) ^ {- 1} (e ^ {\\Delta A} - I) \\cdot \\Delta B.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.493, + 0.483, + 0.508 + ], + "angle": 0, + "content": "Then, The discretized formulation of Eq. 1 is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.519, + 0.483, + 0.544 + ], + "angle": 0, + "content": "\\[\nh _ {t} = \\bar {\\boldsymbol {A}} h _ {t - 1} + \\bar {\\boldsymbol {B}} x _ {t}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.541, + 0.292, + 0.555 + ], + "angle": 0, + "content": "\\[\ny _ {t} = C h _ {t},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.559, + 0.484, + 0.621 + ], + "angle": 0, + "content": "where \\(\\overline{A} \\in \\mathbb{R}^{N \\times N}\\), \\(\\overline{B} \\in \\mathbb{R}^{N \\times D}\\). In order to improve the computational efficiency, the iterative process described in Eq. 3 can be performed by the parallel computing mode of global convolution [8]:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.627, + 0.483, + 0.653 + ], + "angle": 0, + "content": "\\[\ny = x * \\bar {K}, \\quad - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.645, + 0.484, + 0.667 + ], + "angle": 0, + "content": "with \\(\\overline{K} = (C\\overline{B}, C\\overline{A}\\overline{B}, \\dots, C\\overline{A}^{L-1}\\overline{B})\\) (4)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.703 + ], + "angle": 0, + "content": "where \\(\\ast\\) denotes the convolution operation, and \\(\\overline{\\pmb{K}}\\in \\mathbb{R}^{L}\\) serves as the kernel of the SSMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Selective State Space Models (S6). Conventional SSMs (i.e., S4) have been implemented to capture sequence context under linear time complexity, despite the fact that they are constrained by static parameterization and cannot perform content-based reasoning. To address this problem, the selective state space model (i.e., Mamba [8]) has been proposed, which allows the model to selectively propagate or forget information based on the latitude of the current token along the length of the sequence by simply setting the parameters of the SSM as a function of the inputs. In S6, the parameters \\( B \\), \\( C \\), and \\( \\Delta \\) are computed directly from the input sequence \\( x(t) \\), thus enabling sequence-aware parameterization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.428, + 0.605, + 0.443 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.454, + 0.806, + 0.47 + ], + "angle": 0, + "content": "4.1. Analysis of Vanilla Mamba Block" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.643 + ], + "angle": 0, + "content": "As analyzed in Sec. 3, Mamba is able to model global context information based on linear complexity with respect to sequence length through pixel-by-pixel state transfer, which shows potential in lightweight model design. Moreover, considering the mobile-friendly and localized feature extraction capability of Convolution, a natural intuition arises: can we build a lightweight and efficient visual backbone based on Convolution and Mamba blocks? Although there have been some works [40] that attempt to do so, the performance and efficiency shown are not competitive with Convolution [49] or Transformer-based models [43, 48] so far." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Baseline Construction. In response to this intuition, we first build a baseline architecture using Convolution (the local block in Fig. 3) and a vanilla Mamba block (applying SS2D in VMamba [31] to adapt to the visual data), i.e., the Baseline in Table 1. Note that the baseline structure is identical to the final structure of TinyViM-S, which is designed with reference to [43], and specific structural configurations can be found in the Supplementary Material. The only difference between Baseline and TinyViM-S is that we replace the TinyViM blocks with SS2D-based vanilla Mamba blocks. We are surprised to find that this version achieves a Top-1 accuracy of \\(79.1\\%\\), which is significantly better than the purely Convolutional (i.e., Conv only) version. This can be explained by the fact that Mamba can improve the expressive ability of the model by capturing the global context at each stage. However, the low throughput limits the use of the model in real-time applications. We then attempt to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23522" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.147 + ], + "angle": 0, + "content": "Table 1. Effect of frequency input Mamba block. Throughput is tested on a Nvidia V100 GPU with maximum power-of-two batch size that fits in memory. The Top-1 accuracy is the average of three experiments to exclude random fluctuations." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.157, + 0.442, + 0.251 + ], + "angle": 0, + "content": "
VariantsGMACsThroughputTop-1
Baseline0.96167379.1
Conv only0.88321277.5
Low only0.93257479.0
High only0.96137778.6
Low + High0.97150979.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.265, + 0.483, + 0.295 + ], + "angle": 0, + "content": "improve the baseline based on qualitative and quantitative analysis." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.296, + 0.484, + 0.612 + ], + "angle": 0, + "content": "Qualitative Analysis. We first perform a spectral analysis of the feature maps before and after inputting the Mamba block in the baseline, as shown in Fig. 2. It shows that the Mamba block has strong low-frequency capture ability under the Convolution-Mamba hybrid architecture, but loses high-frequency information, such as local edges and textures. This low-frequency preference impairs the performance and efficiency of vision Mamba because 1) Populating each stage with low-frequency information degrades the high-frequency component, weakening the fine-grained recognition ability of visual Mamba, and 2) Additional high-frequency component input to a Mamba block impairs the efficiency and parallel computation of the model. A feasible strategy for this phenomenon is that we decouple the high and low frequencies of the features and input only the low-frequency components of the features into the Mamba block for hidden state transfer. Therefore, we decouple the high and low frequencies of the feature maps based on the Laplace pyramid, which is a more efficient architecture compared to other methods such as the cascaded wavelet transform, to verify the effectiveness of our strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Quantitative Analysis. We next set the inputs of the Mamba block for each stage to be low-frequency only, high-frequency only, high and low-frequency in parallel to explore the effect of the corresponding variants on the classification accuracy, respectively. As shown in Table 1, when only low frequencies are used as inputs, the model achieves \\(79.0\\%\\) accuracy, reduces the computational consumption of the model and achieves 1.5 times higher throughput. This verifies that we can significantly improve the efficiency of the model without compromising the classification accuracy by using only low-frequency inputs to Mamba. Therefore, based on the above qualitative and quantitative analyses, we design the Laplace mixer to optimize Mamba in the construction of a lightweight vision backbone." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.242, + 0.85 + ], + "angle": 0, + "content": "4.2. Laplace Mixer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Previous work such as EfficientVMamba [40] uses dilated-based selective scanning to improve the efficiency of vanilla Mamba [8]. Despite reducing the number of input tokens," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.392 + ], + "angle": 0, + "content": "EfficientVMamba performs unsatisfactorily on various visual tasks [40]. We argue that the potential of Mamba in lightweight backbones is not fully exploited simply by modifying the scanning path in the image domain. Based on the analysis in Sec. 4.1, we reconstruct the Mamba block based on the frequency decoupling and propose an elegant Laplace mixer. Furthermore, consider that in a general visual backbone, the bottom layer is more concerned with capturing high-frequency details, while the top layer is more concerned with modeling low-frequency global information [19]. As in humans, by capturing details in high-frequency components, the lower layers can capture the basic visual features and gradually collect local information to achieve a global understanding of the input. In particular, feature redundancy has been shown to exist in existing networks [15], and separate low-frequency and high-frequency processing for all channels is unnecessary. Therefore, we design a frequency ramp inception structure which reduces the channel size of the high-frequency branches as the depth of the network increases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.622 + ], + "angle": 0, + "content": "Specifically, for the input feature \\(X \\in \\mathbb{R}^{H \\times W \\times D}\\), where H, W and D denote the height, width and channels of \\(X\\), we first split along the channel dimension and divide it into a low-frequency input \\(X_{l} \\in \\mathbb{R}^{H \\times W \\times \\alpha D}\\) and a high-frequency input \\(X_{h} \\in \\mathbb{R}^{H \\times W \\times (1 - \\alpha)D}\\). \\(\\alpha\\) denotes the partition coefficient, which determines the number of channels in the low-frequency branch. Then, for the low-frequency branch, we adopt a simple Laplace pyramid architecture. Specifically, we average pool \\(X_{l}\\) to obtain the corresponding low-frequency components \\(X_{ll} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\alpha D}\\) and shrink the number of input tokens. Next, we recover the resolution of \\(X_{ll}\\) based on the nearest neighbor upsampling and perform pixel-by-pixel subtraction from \\(X_{l}\\) to obtain the high-frequency component of \\(X_{l}\\), i.e., \\(X_{lh}\\). This process can be formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.63, + 0.907, + 0.648 + ], + "angle": 0, + "content": "\\[\nX _ {l l} = \\operatorname {P o o l} (X _ {l}), \\quad X _ {l h} = X _ {l} - \\operatorname {U p s a m p l e} (X _ {l l}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.909, + 0.777 + ], + "angle": 0, + "content": "Then, we concatenate \\( X_{lh} \\) and \\( X_{h} \\) along the channel to obtain the high-frequency input \\( X_{hh} \\in \\mathbb{R}^{H \\times W \\times D} \\), and apply a small depth-wise Convolution kernel, which usually tends to respond to the high frequencies in the input [5, 6], enhancing the high-frequency component of \\( X_{hh} \\). As for the low-frequency input \\( X_{ll} \\), we input it into the SS2D block for cross-scanning to capture the global context. The process of frequency separation processing can be described as," + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.788, + 0.907, + 0.807 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {l l} = \\operatorname {S S 2 D} \\left(X _ {l l}\\right), \\quad \\hat {X} _ {h h} = \\operatorname {R e p} _ {3} \\left(X _ {h h}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathrm{Rep}_3\\) denotes the reparameterized \\(3\\times 3\\) convolution. Finally, we sum up \\(\\hat{X}_{ll}\\) and \\(\\hat{X}_{hh}\\) of the corresponding channels at the element level, followed by a \\(1\\times 1\\) convolution for fusion. Through the frequency decoupling process and frequency ramp inception, the Laplace mixer can effectively weigh and integrate the high and low-frequency" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23523" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.647, + 0.119 + ], + "angle": 0, + "content": "Table 2. Classification performance on ImageNet-1K. Following [30], throughput is tested on Nvidia V100 GPU with maximum power-of-two batch size that fits in memory." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.13, + 0.644, + 0.457 + ], + "angle": 0, + "content": "
ModelConferenceTypeParams ↓ (M)GMACs ↓Throughput ↑ (im/s)Reso.Top-1 ↑ (%)
MobileOne-S1 [49]CVPR'23CNN4.80.8354522475.9
SwiftFormer-S [43]ICCV'23CNN+ViT6.11.0262622478.5
EfficientFormerV2-S0 [27]ICCV'23CNN+ViT3.50.498530075.7
EfficientVMamba-T [40]Arxiv'24Mamba6.11.0139622476.5
Vim-Ti [56]ICML'24Mamba7.01.561730076.1
QuadMamba-Li [51]NeurlIPS'24Mamba5.40.8119022474.2
MSVMamba-N [44]NeurlIPS'24Mamba7.00.9128322477.3
MobileMamba-S6 [18]CVPR'25CNN+Mamba15.00.7365022478.0
TinyViM-S-CNN+Mamba5.60.9256322479.2
PoolFormer-S12 [52]CVPR'22Pool12.02.0190222477.2
MobileOne-S3 [49]CVPR'23CNN10.11.9190022478.1
MobileOne-S4 [49]CVPR'23CNN14.83.0122322479.4
Agent-PVT-T [14]ECCV'24ViT11.62.0144722478.4
EfficientVMamba-S [40]Arxiv'24Mamba11.01.367422478.7
QuadMamba-T [51]NeurlIPS'24Mamba10.02.058522478.2
MSVMamba-M [44]NeurlIPS'24Mamba12.01.595722479.8
MobileMamba-B1 [18]CVPR'25CNN+Mamba17.11.1209725679.9
TinyViM-B-CNN+Mamba11.01.5185122481.2
PoolFormer-S36 [52]CVPR'22Pool31.05.266722481.4
InceptionNext-T [53]CVPR'24CNN28.04.298722482.3
Agent-PVT-S [14]ECCV'24ViT20.64.068622482.4
FastViT-SA24 [48]ICCV'23CNN+ViT20.63.886122482.6
EfficientVMamba-B [40]Arxiv'24Mamba33.04.058022481.8
LocalVMamba-T [23]Arxiv'24Mamba26.05.7-22482.7
ViM-S [56]ICML'24Mamba26.05.120122480.5
VMamba-T [31]NeurlIPS'24Mamba30.04.938322482.6
QuadMamba-S [51]NeurlIPS'24Mamba31.05.554122482.4
MobileMamba-B4 [18]CVPR'25CNN+Mamba17.14.351351282.5
TinyViM-L-CNN+Mamba31.74.784322483.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.663, + 0.09, + 0.911, + 0.118 + ], + "angle": 0, + "content": "Table 3. Ablation of frequency ramp inception design." + }, + { + "type": "table", + "bbox": [ + 0.665, + 0.122, + 0.909, + 0.178 + ], + "angle": 0, + "content": "
InceptionRampGMACsThroughputTop-1
XX0.94247879.2
✓X0.93257479.0
✓✓0.93256379.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.664, + 0.181, + 0.911, + 0.209 + ], + "angle": 0, + "content": "Table 4. Ablation of the partition coefficient \\( \\alpha \\) of four stages." + }, + { + "type": "table", + "bbox": [ + 0.666, + 0.213, + 0.907, + 0.264 + ], + "angle": 0, + "content": "
α1α2α3α4GMACsThrou.Top-1
0.750.50.50.250.94259178.8
0.50.50.50.50.93257479.0
0.250.50.50.750.93256379.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.664, + 0.267, + 0.911, + 0.295 + ], + "angle": 0, + "content": "Table 5. Ablation of the convolution kernel size in Laplace Mixer." + }, + { + "type": "table", + "bbox": [ + 0.674, + 0.3, + 0.9, + 0.368 + ], + "angle": 0, + "content": "
kernelaxialGMACsThroughputTop-1
3X0.94251079.0
5✓0.93259879.0
7✓0.93256379.2
9✓0.94247979.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.666, + 0.371, + 0.91, + 0.385 + ], + "angle": 0, + "content": "Table 6. Ablation of re-parameterization." + }, + { + "type": "table", + "bbox": [ + 0.675, + 0.389, + 0.902, + 0.457 + ], + "angle": 0, + "content": "
VariantRepGMACsThrou.Top-1
TinyViM-S✓0.93256379.2
TinyViM-S✗0.93256379.2
TinyViM-L✓4.7284383.3
TinyViM-L✗4.7284383.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.504 + ], + "angle": 0, + "content": "components of all layers. As a result, TinyViM achieves a better balance between accuracy and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.512, + 0.288, + 0.527 + ], + "angle": 0, + "content": "4.3. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.534, + 0.483, + 0.746 + ], + "angle": 0, + "content": "TinyViM employs a similar multi-scale backbone design similar to previous works [26, 43]. As shown in Fig. 3, given an image \\( I \\in \\mathbb{R}^{H \\times W \\times 3} \\), a simple stem is employed to output \\( X_{1} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times D_{1}} \\) feature map. \\( H, W \\) denote the height and width of the input image respectively, and \\( D_{1} \\) is the channel number of \\( X_{1} \\). Note that the stem is implemented with two \\( 3 \\times 3 \\) convolutions with a stride of 2. Then, the output feature maps are fed into the first stage, which begins with Local Blocks. The local block consists of a reparameterized \\( 3 \\times 3 \\) convolution and a feed-forward network (FFN, implemented by two consecutive \\( 1 \\times 1 \\) convolutions) that achieves local feature extraction and feature channel mixing, respectively. The local block can be formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.756, + 0.483, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\hat {X} _ {1} = \\operatorname {C o n v} _ {1} (\\operatorname {G e L U} (\\operatorname {C o n v} _ {1} (\\operatorname {R e p} _ {3} (X _ {1})))) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.483, + 0.875 + ], + "angle": 0, + "content": "where \\(\\mathrm{Conv}_1\\) and \\(\\mathrm{Rep}_3\\) denote the \\(1\\times 1\\) and reparameterized \\(3\\times 3\\) convolution, respectively. The features are then fed into the TinyViM block for global context-awareness. Similarly, the TinyViM block consists of a Laplace mixer and FFN for global feature extraction and channel mixing, respectively. The TinyViM Block can be described as," + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\hat {X} ^ {\\prime} _ {1} = \\operatorname {L a p l a c e M i x e r} (\\hat {X} _ {1}) + \\hat {X} _ {1}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.487, + 0.907, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\tilde {X} _ {1} = \\operatorname {C o n v} _ {1} \\left(\\operatorname {G e L U} \\left(\\operatorname {C o n v} _ {1} \\left(\\hat {X} ^ {\\prime} _ {1}\\right)\\right)\\right) + \\hat {X} ^ {\\prime} _ {1}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.907, + 0.605 + ], + "angle": 0, + "content": "There is a Patch Embedding layer between two consecutive stages for increasing the channel dimension and reducing the resolution of feature maps. Next, the resulting feature maps are subsequently fed into the second, third, and fourth stages, producing \\(\\frac{H}{8} \\times \\frac{W}{8} \\times D_2\\), \\(\\frac{H}{16} \\times \\frac{W}{16} \\times D_3\\) and \\(\\frac{H}{32} \\times \\frac{W}{32} \\times D_4\\) feature maps, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.617, + 0.646, + 0.634 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.643, + 0.907, + 0.78 + ], + "angle": 0, + "content": "We evaluate the proposed TinyViM across four downstream tasks: classification on ImageNet-1K dataset [1], object detection and instance segmentation on MS-COCO 2017 dataset [28], and semantic segmentation on ADE20K dataset [55]. All experiments are implemented following the common training settings in [43]. Due to page limits, we provide a description of the dataset, the experimental setup, details of the model's architecture and more qualitative analysis in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.706, + 0.804 + ], + "angle": 0, + "content": "5.1. Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We evaluate TinyViM's performance in image classification on ImageNet-1K [1] with some advanced Convolution, ViT and Mamba-based methods, and the comparison results are summarized in Table 2. With similar GMACs and throughput, TinyViM-S achieves a top-1 accuracy of \\(79.2\\%\\), outperforming SwiftFormer-S by \\(0.7\\%\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23524" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.333 + ], + "angle": 0, + "content": "In particular, TinyViM retains its performance advantages even at small and large scales. For example, TinyViMB achieves an accuracy of 81.2, which outperforms the convolution-based model MobileOne-S4 by \\(1.8\\%\\), the ViT-based model Agent-PVT-T by \\(2.8\\%\\), and the Mamba-based model MSVMamba-M by \\(1.4\\%\\), respectively. These comparison results show the excellent performance of TinyViM. In addition, we compare the visualisation of Effective Receptive Field (ERF) [32] with the classical lightweight backbone, as shown in Fig. 5. The comparison models include the pure Convolution-based MobileOne-S3 and the Convolution-ViT hybrid model SwiftFormer-S. It can be observed that TinyViM-S possesses a larger ERF at a similar size. This suggests that TinyViM is able to effectively model the global context through the Laplace Mixer, thus improving the semantic recognition ability of the model." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.622 + ], + "angle": 0, + "content": "In terms of computational efficiency, TinyViM-S achieves a throughput of 2574 images/s, which is significantly better than other visual Mamba models. For example, the throughput of EfficientVmamba is only 1396 images/s, which is only about half of the throughput of TinyViM. This huge efficiency advantage persists on TinyViM-B and TinyViM-L, where the throughput is 2.7 and 1.5 times higher than that of the similarly scaled EfficientVmamba, respectively. Compared with other convolutional and ViT-based models, TinyViM maintains its lead in performance while its throughput is highly competitive. It should be emphasized that this competitiveness is significantly improved on A100 GPUs. This is due to the fact that Mamba has a number of hardware-friendly improvements on the A100 such as triton, which are not yet supported on the V100 GPUs. In summary, the results show that TinyViM strikes a better balance between performance and efficiency than other state-of-the-art lightweight backbones, and offers high value in real-time applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.468, + 0.652 + ], + "angle": 0, + "content": "5.2. Object Detection and Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate the proposed TinyViM for object detection and instance segmentation using the Mask R-CNN [20] framework. As shown in Table 7, TinyViM outperforms competitive models such as the pooling-based model PoolFormer, the Transformer-based models FastVit and SwiftFormer, and the Mamba-based model EfficientVMamba in both \\(\\mathrm{AP}^{box}\\) and \\(\\mathrm{AP}^{mask}\\) metrics. Specifically, compared to recent models SwiftFormer-L1, TinyViM has \\(1.1\\%\\) and \\(0.6\\%\\) improvement on \\(\\mathrm{AP}^{box}\\) and \\(\\mathrm{AP}^{mask}\\), respectively, and has a larger throughput. Compared to EfficientVMamba-S, TinyViM-B has more significant advantages, i.e., 3.0 and 2.0 improvement in \\(\\mathrm{AP}^{box}\\) and \\(\\mathrm{AP}^{mask}\\), respectively, with a 1.7 times throughput. When comparing with larger models such as FastViT-SA24 and EfficientVMamba-b, TinyViM still has an advantage in throughput, and shows \\(2.5\\%\\) and \\(0.8\\%\\) improvements in \\(\\mathrm{AP}^{box}\\), respectively. These results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "demonstrate the superiority of TinyViM when transferring to detection and instance segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.133, + 0.731, + 0.149 + ], + "angle": 0, + "content": "5.3. Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.156, + 0.905, + 0.307 + ], + "angle": 0, + "content": "Table 7 shows the semantic segmentation results of TinyViM as a backbone in comparison with other state-of-the-art backbones, where Semantic FPN [24] serves as the decode head. Specifically, compared with recent methods such as SwiftFormer, TinyVim obtains \\(0.8\\%\\) and \\(0.3\\%\\) improvement in base and large scale, respectively. Compared with the method FasterViT, TinyViM has more significant advantages, with \\(3.9\\%\\) and \\(3.2\\%\\) improvements, respectively. The experimental results verify the effectiveness of TinyViM in semantic segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.318, + 0.679, + 0.334 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.905, + 0.506 + ], + "angle": 0, + "content": "Effect of the Frequency Decoupling. In order to investigate the effectiveness of the frequency decoupling strategy, we perform spectral visualization of the feature maps of the low-frequency and high-frequency branches, respectively. As shown in Fig. 4, by applying Mamba, the feature maps have higher concentrations at low frequencies; by applying depth-wise convolution, the high frequencies of the model are effectively preserved. This proves that the Laplace mixer can effectively decouple the high and low frequency components of the features and specially enhance them respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.905, + 0.825 + ], + "angle": 0, + "content": "Ablation of the Frequency Ramp Inception. We explore the effectiveness of the frequency ramp inception design, as shown in Table 3. We set up three experimental variants: variant 1, we implement Laplace frequency decoupling for all channels of the feature map, and input low and high frequencies into Mamba and depth-wise convolution, respectively; variant 2, we adopt a non-ramp structure, i.e., we set the segmentation coefficient \\(\\alpha\\) of the four stages to be \\(1/2\\); and variant 3, we gradually increase \\(\\alpha\\) with the increase of the stage of the model. In the implementation of this paper, the \\(\\alpha\\) of the four stages are 0.25, 0.5, 0.5 and 0.75, respectively. It can be observed that, even in the small size, the throughput of variant 1 is significantly reduced and more GMACs are required. Variant 2, although more efficient, has a decreased accuracy due to the lack of adjusting the weights of the low-frequency and high-frequency components in different stages. We also explore the inverse ramp structure, as shown in Table 4. The frequency ramp used in variant 3 effectively weighs and integrates the high and low frequency information, thus helping the model to achieve a better balance between performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Ablation of the convolution kernel size in Laplace Mixer. ViM and VMamba use a \\(3 \\times 3\\) depthwise convolution before scanning. However, a larger convolution kernel is beneficial for parameters (i.e., \\(B\\), \\(C\\)) generation due to the large receptive fields, although it impairs the efficiency of" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23525" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 7. Object detection & instance segmentation results on MS COCO 2017 with the Mask RCNN framework. Semantic segmentation results on ADE20K by integrating models into Semantic FPN." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.122, + 0.862, + 0.384 + ], + "angle": 0, + "content": "
BackboneThroughput ↑ (img/s)Object DetectionInstance SegmentationSemantic mIoU
APbox50APbox75APmaskAPmask50APmask75
EfficientVMamba-T [40]16835.657.738.033.254.435.1-
TinyViM-S26238.761.141.335.757.137.938.9
PoolFormer-S12 [52]17137.359.040.134.655.836.937.2
EfficientFormer-L1 [26]19737.960.341.035.457.337.338.9
FastViT-SA12 [48]16238.960.542.235.957.638.138.0
SwiftFormer-L1 [43]17441.263.244.838.160.240.741.1
EfficientVMamba-S [40]10439.361.842.636.758.939.2-
MobileMamba-B1 [18]-40.661.843.837.458.939.940.7
TinyViM-B18042.364.246.338.761.141.341.9
PoolFormer-S36 [52]8741.063.144.837.760.140.042.0
EfficientFormer-L3 [26]11741.463.944.738.161.040.443.5
FastViT-SA24 [48]11042.063.545.838.060.540.541.0
SwiftFormer-L3 [43]12042.764.446.739.161.741.843.9
EfficientVMamba-B [40]9043.766.247.940.263.342.9-
MobileMamba-B4 [18]-40.161.843.036.958.639.242.5
TinyViM-L11144.566.448.640.763.643.844.2
" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.401, + 0.482, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.555, + 0.483, + 0.613 + ], + "angle": 0, + "content": "Figure 4. Fourier spectrum of TinyViM for the Low frequency branch and High frequency branch. It can be observed that we achieve the decoupling of the high and low frequency components of the features and specifically enhance them separately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.738 + ], + "angle": 0, + "content": "the model. Therefore, we introduce axial depthwise convolution, which increases the convolution's receptive field along the scan path without decreasing the model efficiency. We also explore \\(3 \\times 3\\) depthwise convolution and axial convolution with kernel sizes 5 and 9, as shown in Table 5. The model achieves the best balance between performance and efficiency when the kernel size is 7." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.483, + 0.815 + ], + "angle": 0, + "content": "Ablation of re-parameterization. As shown in Table 6, the reparameterization has no improvement on TinyViM-S, but brings an increase of 0.2 to TinyViM-L. Considering that the reparameterization technique does not harm the inference efficiency of the model, we apply it for TinyViM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.21, + 0.849 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.858, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Motivated by Mamba's ability to achieve global context modeling based on linear complexity, we construct a tiny and efficient hybrid vision Mamba TinyViM by combin-" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.399, + 0.892, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.531, + 0.907, + 0.561 + ], + "angle": 0, + "content": "Figure 5. The ERF of MobileOne, SwiftFormer and TinyViM. Our TinyViM effectively obtains large ERFs with Laplace Mixer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.907, + 0.901 + ], + "angle": 0, + "content": "ing the mobile-friendly convolution and Mamba in this paper. In contrast to previous vision Mamba efforts that simply modify scan paths in the image domain, whose performance could not compete with Convolution or Transformer based models with similar scale, TinyViM attempts to build a lightweight backbone from a frequency decoupling perspective. Specifically, we verify that the Mamba block mainly models low-frequency information in a Convolution-Mamba hybrid architecture after spectral analysis and quantitative experiments. To this end, we design a frequency decoupling strategy to preserve the high-frequency information and improve the low-frequency modeling efficiency of Mamba. In particular, we introduce a frequency ramp inception to weigh the different strengths of the low-frequency and high-frequency components at different stages of the model, so as to help the model achieve a better balance between accuracy and efficiency. Extensive experiments have shown that TinyViM outperforms other Convolution, Transformer and Mamba-based models. In future work, we plan to apply it as a lightweight backbone to replace the image encoder in SAM to explore the potential of vision Mamba in multi-modal scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23526" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.119, + 0.188, + 0.135 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.144, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[1] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.483, + 0.27 + ], + "angle": 0, + "content": "[2] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11963-11975, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[3] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.482, + 0.409 + ], + "angle": 0, + "content": "[4] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.482, + 0.465 + ], + "angle": 0, + "content": "[5] Paul Gavrikov and Janis Keuper. Can biases in imagenet models explain generalization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22184-22194, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[6] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[7] Alex Graves. Long Short-Term Memory, pages 37-45. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[8] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.482, + 0.647 + ], + "angle": 0, + "content": "[9] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[11] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[12] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[13] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Dongchen Han, Tianzhu Ye, Yizeng Han, Zhuofan Xia, Siyuan Pan, Pengfei Wan, Shiji Song, and Gao Huang. Agent" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.144, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "attention: On the integration of softmax and linear attention. In Computer Vision - ECCV 2024, pages 124-140, Cham, 2025. Springer Nature Switzerland. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.135, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[15] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[16] Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[17] Ramin Hasani, Mathias Lechner, Tsun-Hsuan Wang, Makram Chahine, Alexander Amini, and Daniela Rus. Liquid structural state-space models. arXiv preprint arXiv:2209.12951, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[18] Haoyang He, Jiangning Zhang, Yuxuan Cai, Hongxu Chen, Xiaobin Hu, Zhenye Gan, Yabiao Wang, Chengjie Wang, Yunsheng Wu, and Lei Xie. Mobilemamba: Lightweight multi-receptive visual mamba network, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[21] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, and Hartwig Adam. Searching for mobilenetv3. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.525, + 0.905, + 0.592 + ], + "angle": 0, + "content": "[22] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.594, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[23] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[24] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollar. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[25] Dong Li, Yidi Liu, Xueyang Fu, Senyan Xu, and Zheng-Jun Zha. Fourier learning integration with state space models for image deraining, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[26] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35: 12934-12949, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[27] Yanyu Li, Ju Hu, Yang Wen, Georgios Evangelidis, Kamyar Salahi, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Rethinking vision transformers for mobilenet size and speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16889-16900, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23527" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[29] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Tommi Kärkkäinen, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[30] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14430, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[31] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[32] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[33] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[34] Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer. Mega: moving average equipped gated attention. arXiv preprint arXiv:2209.10655, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[35] Xiaowen Ma, Rongrong Lian, Zhenkai Wu, Renxiang Guan, Tingfeng Hong, Mengjiao Zhao, Mengting Ma, Jiangtao Nie, Zhenhong Du, Siyang Song, and Wei Zhang. A novel scene coupling semantic mask network for remote sensing image segmentation. ISPRS Journal of Photogrammetry and Remote Sensing, 221:44-63, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[36] Xiaowen Ma, Zhen-Liang Ni, Shuai Xiao, and Xinghao Chen. Timepro: Efficient multivariate long-term time series forecasting with variable- and time-aware hyper-state. In Forty-second International Conference on Machine Learning, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[37] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[38] Sachin Mehta, Mohammad Rastegari, Linda Shapiro, and Hannaneh Hajishirzi. Espnetv2: A light-weight, power efficient, and general purpose convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[39] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[40] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba, 2024. 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[41] Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Exploring state space models for multimodal learning, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[42] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition* (CVPR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[43] Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Swiftformer: Efficient additive attention for transformer-based real-time mobile vision applications. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17425-17436, 2023. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[44] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. arXiv preprint arXiv:2405.14174, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[45] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xin chao Wang, and Shuicheng Yan. Inception transformer. In Advances in Neural Information Processing Systems, pages 23495-23509. Curran Associates, Inc., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[47] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[48] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Fastvit: A fast hybrid vision transformer using structural reparameterization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5785-5795, 2023. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.907, + 0.594 + ], + "angle": 0, + "content": "[49] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7907-7917, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[50] Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Repvit: Revisiting mobile cnn from vit perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15909-15920, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[51] Fei Xie, Weijia Zhang, Zhongdao Wang, and Chao Ma. Quadmamba: Learning quadtree-based selective scan for visual state space model, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.907, + 0.775 + ], + "angle": 0, + "content": "[52] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10819-10829, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[53] Weihao Yu, Pan Zhou, Shuicheng Yan, and Xinchao Wang. Inceptionnext: When inception meets convnext. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5672-5683, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[54] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23528" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[56] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In *Forty-first International Conference on Machine Learning*. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[57] Qinfeng Zhu, Yuanzhi Cai, Yuan Fang, Yihan Yang, Cheng Chen, Lei Fan, and Anh Nguyen. Samba: Semantic segmentation of remotely sensed images with state space model, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23529" + } + ] +] \ No newline at end of file diff --git a/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/f0ba958e-31c7-4013-97d5-2552433cb070_origin.pdf b/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/f0ba958e-31c7-4013-97d5-2552433cb070_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3c3a67e9cb5b010bf8efcdb68f68e08c568ac6e3 --- /dev/null +++ b/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/f0ba958e-31c7-4013-97d5-2552433cb070_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c06d841a83a1dd83dc25d9af60d5ab937f3e3bb2595579447cbaf31e270e5b31 +size 1794183 diff --git a/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/full.md b/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a8702fcdab4defdf300373c2ee9c64289de89f90 --- /dev/null +++ b/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/full.md @@ -0,0 +1,302 @@ +# TinyViM: Frequency Decoupling for Tiny Hybrid Vision Mamba + +Xiaowen Ma * Zhenliang Ni * Xinghao Chen† Huawei Noah's Ark Lab + +xwma@zju.edu.cn, {nizhenliang2, xinghao.chen}@huawei.com + +# Abstract + +Mamba has shown great potential for computer vision due to its linear complexity in modeling the global context with respect to the input length. However, existing lightweight Mamba-based backbones cannot demonstrate performance that matches Convolution or Transformer-based methods. By observing, we find that simply modifying the scanning path in the image domain is not conducive to fully exploiting the potential of vision Mamba. In this paper, we first perform comprehensive spectral and quantitative analyses, and verify that the Mamba block mainly models low-frequency information under Convolution-Mamba hybrid architecture. Based on the analyses, we introduce a novel Laplace mixer to decouple the features in terms of frequency and input only the low-frequency components into the Mamba block. In addition, considering the redundancy of the features and the different requirements for high-frequency details and low-frequency global information at different stages, we introduce a frequency ramp inception, i.e., gradually reduce the input dimensions of the high-frequency branches, so as to efficiently trade-off the high-frequency and low-frequency components at different layers. By integrating mobile-friendly convolution and efficient Laplace mixer, we build a series of tiny hybrid vision Mamba called TinyViM. The proposed TinyViM achieves impressive performance on several downstream tasks including image classification, semantic segmentation, object detection and instance segmentation. In particular, TinyViM outperforms Convolution, Transformer and Mamba-based models with similar scales, and the throughput is about 2-3 times higher than that of other Mamba-based models. Code is available at https://github.com/xwmaxwma/TinyViM. + +# 1. Introduction + +In recent years, Mamba has gained widespread attention and application in the field of vision due to its advantages in long-sequence modeling [8]. Mamba is a selective struc + +![](images/a8656d77bb359f04e054d3a8191f64d6b442e0a356f64302bf4280292be95af9.jpg) +Figure 1. Comparison of GFLOPs, Throughput and accuracy between TinyViM and others. The top-1 accuracy is tested on ImageNet-1K and the throughput is measured on Nvidia V100 GPU with maximum power-of-two batch size that fits in memory. Compared to recent Mamba-based models such as EfficientV-Mamba, QuadMamba, and VMamba, TinyViM has higher top-1 accuracy with $2 \times$ higher throughput and fewer GFLOPs. + +![](images/4ad6defaaaf37411d30ff5bd490200b612d065cb890789bfada110486644141d.jpg) + +tured state-space model that can overcome the limitations of local perception in Convolution neural networks [19] and the quadratic computational complexity of Transformers [3], thereby demonstrating strong capabilities in handling long-sequence data. Due to its advanced context extraction capabilities, Mamba has been widely applied in image classification tasks and has achieved advanced performance, such as Vim [56], VMamba [31], etc. + +Vim [56] is one of the earliest works to apply Mamba to visual tasks, by introducing a bidirectional state-space model to process 2D image data. By introducing the Cross-Scan Module (CSM) and 2D Selective Scanning (SS2D) modules, VMamba [31] reduces the computational complexity from quadratic to linear while maintaining the global acceptance domain. MSVMamba [44] uses a multiscale 2D scanning method to apply Mamba to both original and down-sampled feature maps. This not only helps to learn long-distance dependencies, but also reduces compu + +![](images/5bd5024f51d23ddf4b352624596601f7d9fb9f70fb1daf4f2092893f4467bf9b.jpg) +Figure 2. Spectral analysis of Mamba. (a) Frequency magnitude $(14 \times 14)$ from 8 output channels before- and after Mamba block. (b) Relative log amplitudes of Fourier transformed feature maps. The magnitude and amplitude are averaged over 384 samples. (a) and (b) show that Mamba block focuses on capturing low-frequency information under the Convolution-Mamba hybrid architecture. + +![](images/fe08c43ef9675f8ddc3310f0a1780ac90a24720fce063b4ebef42fe5bdb16d68.jpg) + +tational costs. EfficientVMamba [39] is a lightweight Vision Mamba variant that reduces computational complexity through dilated-based selective scanning. Although these methods inherit the advantages of Mamba linear complexity and global receptive field, they do not show competitive performance with lightweight backbones based on Convolution [49, 53] and Transformer [14, 43, 48] in computationally limited and real-time deployment scenarios, as shown in Fig. 1. In fact, these methods that simply modify the scan path in the image domain do not utilize the full potential of visual Mamba [31, 56]. In this paper, we try to rethink the design of lightweight vision Mamba from the perspective of frequency decoupling, which is a reasonable and effective solution. + +Through observation, we find that Mamba prefers to extract low-frequency features and ignores some high-frequency features. As shown in Fig. 2, we construct a baseline using mobile-friendly Convolution and vanilla Mamba, and perform spectral analysis of the features before and after the Mamba. Many previous works [26, 43, 50] have verified the necessity of using Convolution in lightweight backbones. After applying the Mamba block, the low-frequency component in the center of the two-dimensional spectrum map is highlighted, and the high-frequency component is suppressed. Therefore, to improve the efficiency of low-frequency modeling and keep the high-frequency features, we introduce an efficient Laplace pyramid to decompose the high- and low-frequency components of the features. Then, we input only the low-frequency components into the Mamba block to obtain a global receptive field, and for the high-frequency components, we use a reparameterized $3 \times 3$ depth-wise Convolution to enhance high frequency information [5, 6]. It is worth noting that the size of the low frequency feature is much smaller than that of the original input feature, which can greatly reduce the computational complexity and improve the throughput of the Mamba. Therefore, the strategy of frequency decoupling and specif + +ically enhancement separately effectively improves the efficiency and the feature representation ability of the model. + +In addition, previous models based on Convolution or Transformer have verified that: 1) deep neural networks have feature redundancy [15]; 2) deep neural networks tend to require more high-frequency information in the shallow layer and more global information in the deep layer [35, 45]. Therefore, we propose the frequency ramp inception to further improve the performance and efficiency of the model. Specifically, we split the feature map along the channels at each stage. Early stages prioritize high frequencies, reflected in a higher channel ratio for the high-frequency branch, while later stages concentrate on low frequencies, allowing for a greater channel allocation to low frequencies. This strategic distribution enables the network to learn features that are more suitable at each stage of processing effectively. + +By integrating frequency decoupling and frequency ramp inception, we propose a tiny and efficient Convolution-Mamba hybrid architecture TinyVim. We then conduct extensive experiments to validate the effectiveness of TinyViM, which achieves state-of-the-art image classification performance on ImageNet and advanced performance in downstream tasks such as detection and segmentation. Specifically, TinyViM has higher classification accuracy than other Convolution or Transformer-based backbones with a similar scale. In addition, compared to other Mamba-based backbones such as EfficientVMambaT, TinyViM has $2 \times$ higher throughput and $2.7\%$ higher accuracy, as shown in Fig. 1. Our contributions can be summarized as follows. + +- We verify that Mamba mainly models low-frequency information under the Convolution-Mamba hybrid architecture based on qualitative analysis and quantitative comparison. +- We introduce an efficient Laplace mixer, which improves the modeling efficiency and performance of vi + +sual Mamba through frequency decoupling and frequency ramp inception. + +- We propose TinyViM by integrating the Laplace mixer and Convolution. Extensive experiments show that TinyViM outperforms other Convolutional, Transformer and Mamba-based models with a similar scale, and achieves a better trade-off between performance and efficiency. + +# 2. Related Work + +# 2.1. Efficient Generic Vision Backbones + +Convolution has been widely employed for designing efficient vision backbones due to the inductive bias and efficient parallel processing on GPUs. Early works such as the MobileNet series [21, 22, 42] propose the depth-wise separable Convolution and the inverted residual blocks to further reduce the parameter and computational consumption of the plain Convolutional model [19, 46]. A series of subsequent works further improve the efficiency through depth-wise dilated Convolutions [38], network pruning [16], neural structure search [21], channel shuffling [33, 54] and structure reparameterization [48-50]. However, constrained by a limited receptive field, these efficient Convolutional networks lack global interactions between features, which degrades their performance. Although some works propose to increase the receptive field by increasing the size of the Convolutional kernel [2, 29], however, it imposes high memory access cost and computational cost, and significantly slows down the inference speed. Considering that Transformers are capable of achieving global attention but introduce quadratic complexity with respect to the resolution of the input image, a series of recent works have attempted to combine the advantages of CNNs and Transformers to develop more efficient models [26, 43, 48]. For example, and EfficientFormer only adds self-attention at more deep stages to capture global context, SwiftFormer adds effective additive attention at each stage of the model. The success of this hybrid design motivates us to explore whether a more advanced backbone can be constructed by integrating Convolution and Mamba. + +# 2.2. State Space Models + +As a typical architectural paradigm for sequence-to-sequence transformations, state-space models (SSMs) play a pivotal role in handling long-range dependencies in sequence data [10, 11, 36, 47]. Early SSMs have challenges in capturing long-term dependency relationships and a unified understanding of memory [7]. To address these issues, Gu et al. enhance the ability of SSMs to capture extended dependencies using HiPPO initialization [9] and solve the continuous-time memory problem based on a linear state space layer (LSSL) [11]. However, these SSMs are still lim + +ited by computational requirements. Subsequent works propose a series of improvements including complex diagonal structures [12, 13], multiple-input multiple-output support [47], and diagonal-plus-low-rank arithmetic [17] to improve computational efficiency and generalization across tasks. In particular, these strategies have been extended to large representation models [4, 34, 37]. Recently Mamba [8] introduce a selection mechanism that outperforms Transformers on large-scale real data through a data-dependent SSM layer with linearly scaling sequence lengths. + +The success of Mamba has caused the vision community to explore and apply it to various vision tasks including classification [31, 56], detection [51], segmentation [57], deraining [25] and multimodal learning [41]. In particular, ViM [56] and VMamba [31] propose bi-directional scanning and cross-scanning strategies to cope with the non-sequential structure of visual data, respectively. QuadMamba [51] proposes quad-tree scanning to capture local dependencies at different granularities. MSVMamba [44] proposes multiscale two-dimensional scanning, which efficiently mitigates the long-range forgetting problem and reduces the computational cost. A highly related work is EfficientVMamba [40], which improves the scanning efficiency of a model by an atrous-based selective scan approach. However, these works tend to show low throughput and fail to perform competitively with other advanced lightweight backbones based on Convolution or Transformers. This motivates us to design the TinyViM, which obtains satisfactory speed and performance based on Laplace frequency decoupling. + +# 3. Preliminaries + +State Space Models (S4). Inspired by the continuous system, State space models (SSMs) are proposed in deep learning as common sequence models [10]. These models transform an input sequence $x(t) \in \mathbb{R}^{L \times D}$ into an output sequence $y(t) \in \mathbb{R}^{L \times D}$ by utilizing a learnable hidden state $h(t) \in \mathbb{R}^N$ . The process could be denoted as follows: + +$$ +\begin{array}{l} h ^ {\prime} (t) = \boldsymbol {A} h (t) + \boldsymbol {B} x (t), \tag {1} \\ y (t) = \boldsymbol {C} h (t), \\ \end{array} +$$ + +where $\mathbf{A} \in \mathbb{R}^{N \times N}$ is the evolution parameter, $\mathbf{B}, \mathbf{C} \in \mathbb{R}^{N \times D}$ denote the learnable projection parameters, and $N$ is the state size. + +Discretization. The above continuous-time SSMs are not well compatible with deep learning algorithms. Therefore, discretization is needed to align the model with the sampling frequency of the input signal to improve computational efficiency [11]. Following the previous work [13], given the sampling time scale parameter $\Delta$ , the above continuous SSMs are discretized through zero-order hold rule, thus converting the continuous-time parameters $(A, B)$ to + +![](images/a2238a917e987242aee1d0cc9ccba8943687f34363faffc4d84de8965c94f748.jpg) +Figure 3. Overview of the proposed TinyViM, which has four stages and each stage consists of Local Blocks and TinyViM Blocks. The local block applies a reparameterized 3x3 Convolution to extract local features, and the TinyViM block is employed to capture the global context. The core component of the TinyViM block is the Laplace Mixer, which decouples the frequencies of the features based on an efficient Laplace pyramid and passes the state only for the low-frequencies. The enhanced low- and high-frequency components are then integrated based on a frequency ramp Inception structure. Thus, the high- and low-frequency components of different stages are appropriately trade-off and the efficiency is further improved. SS2D denotes 2D selective scanning, which is used in VMamba [31]. + +their corresponding discrete counterparts $(\overline{A},\overline{B})$ + +$$ +\begin{array}{l} \overline {{\boldsymbol {A}}} = e ^ {\boldsymbol {\Delta A}}, \\ - \end{array} \tag {2} +$$ + +$$ +\overline {{B}} = (\Delta A) ^ {- 1} (e ^ {\Delta A} - I) \cdot \Delta B. +$$ + +Then, The discretized formulation of Eq. 1 is formulated as: + +$$ +h _ {t} = \bar {\boldsymbol {A}} h _ {t - 1} + \bar {\boldsymbol {B}} x _ {t}, \tag {3} +$$ + +$$ +y _ {t} = C h _ {t}, +$$ + +where $\overline{A} \in \mathbb{R}^{N \times N}$ , $\overline{B} \in \mathbb{R}^{N \times D}$ . In order to improve the computational efficiency, the iterative process described in Eq. 3 can be performed by the parallel computing mode of global convolution [8]: + +$$ +y = x * \bar {K}, \quad - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - +$$ + +with $\overline{K} = (C\overline{B}, C\overline{A}\overline{B}, \dots, C\overline{A}^{L-1}\overline{B})$ (4) + +where $\ast$ denotes the convolution operation, and $\overline{\pmb{K}}\in \mathbb{R}^{L}$ serves as the kernel of the SSMs. + +Selective State Space Models (S6). Conventional SSMs (i.e., S4) have been implemented to capture sequence context under linear time complexity, despite the fact that they are constrained by static parameterization and cannot perform content-based reasoning. To address this problem, the selective state space model (i.e., Mamba [8]) has been proposed, which allows the model to selectively propagate or forget information based on the latitude of the current token along the length of the sequence by simply setting the parameters of the SSM as a function of the inputs. In S6, the parameters $B$ , $C$ , and $\Delta$ are computed directly from the input sequence $x(t)$ , thus enabling sequence-aware parameterization. + +# 4. Method + +# 4.1. Analysis of Vanilla Mamba Block + +As analyzed in Sec. 3, Mamba is able to model global context information based on linear complexity with respect to sequence length through pixel-by-pixel state transfer, which shows potential in lightweight model design. Moreover, considering the mobile-friendly and localized feature extraction capability of Convolution, a natural intuition arises: can we build a lightweight and efficient visual backbone based on Convolution and Mamba blocks? Although there have been some works [40] that attempt to do so, the performance and efficiency shown are not competitive with Convolution [49] or Transformer-based models [43, 48] so far. + +Baseline Construction. In response to this intuition, we first build a baseline architecture using Convolution (the local block in Fig. 3) and a vanilla Mamba block (applying SS2D in VMamba [31] to adapt to the visual data), i.e., the Baseline in Table 1. Note that the baseline structure is identical to the final structure of TinyViM-S, which is designed with reference to [43], and specific structural configurations can be found in the Supplementary Material. The only difference between Baseline and TinyViM-S is that we replace the TinyViM blocks with SS2D-based vanilla Mamba blocks. We are surprised to find that this version achieves a Top-1 accuracy of $79.1\%$ , which is significantly better than the purely Convolutional (i.e., Conv only) version. This can be explained by the fact that Mamba can improve the expressive ability of the model by capturing the global context at each stage. However, the low throughput limits the use of the model in real-time applications. We then attempt to + +Table 1. Effect of frequency input Mamba block. Throughput is tested on a Nvidia V100 GPU with maximum power-of-two batch size that fits in memory. The Top-1 accuracy is the average of three experiments to exclude random fluctuations. + +
VariantsGMACsThroughputTop-1
Baseline0.96167379.1
Conv only0.88321277.5
Low only0.93257479.0
High only0.96137778.6
Low + High0.97150979.1
+ +improve the baseline based on qualitative and quantitative analysis. + +Qualitative Analysis. We first perform a spectral analysis of the feature maps before and after inputting the Mamba block in the baseline, as shown in Fig. 2. It shows that the Mamba block has strong low-frequency capture ability under the Convolution-Mamba hybrid architecture, but loses high-frequency information, such as local edges and textures. This low-frequency preference impairs the performance and efficiency of vision Mamba because 1) Populating each stage with low-frequency information degrades the high-frequency component, weakening the fine-grained recognition ability of visual Mamba, and 2) Additional high-frequency component input to a Mamba block impairs the efficiency and parallel computation of the model. A feasible strategy for this phenomenon is that we decouple the high and low frequencies of the features and input only the low-frequency components of the features into the Mamba block for hidden state transfer. Therefore, we decouple the high and low frequencies of the feature maps based on the Laplace pyramid, which is a more efficient architecture compared to other methods such as the cascaded wavelet transform, to verify the effectiveness of our strategy. + +Quantitative Analysis. We next set the inputs of the Mamba block for each stage to be low-frequency only, high-frequency only, high and low-frequency in parallel to explore the effect of the corresponding variants on the classification accuracy, respectively. As shown in Table 1, when only low frequencies are used as inputs, the model achieves $79.0\%$ accuracy, reduces the computational consumption of the model and achieves 1.5 times higher throughput. This verifies that we can significantly improve the efficiency of the model without compromising the classification accuracy by using only low-frequency inputs to Mamba. Therefore, based on the above qualitative and quantitative analyses, we design the Laplace mixer to optimize Mamba in the construction of a lightweight vision backbone. + +# 4.2. Laplace Mixer + +Previous work such as EfficientVMamba [40] uses dilated-based selective scanning to improve the efficiency of vanilla Mamba [8]. Despite reducing the number of input tokens, + +EfficientVMamba performs unsatisfactorily on various visual tasks [40]. We argue that the potential of Mamba in lightweight backbones is not fully exploited simply by modifying the scanning path in the image domain. Based on the analysis in Sec. 4.1, we reconstruct the Mamba block based on the frequency decoupling and propose an elegant Laplace mixer. Furthermore, consider that in a general visual backbone, the bottom layer is more concerned with capturing high-frequency details, while the top layer is more concerned with modeling low-frequency global information [19]. As in humans, by capturing details in high-frequency components, the lower layers can capture the basic visual features and gradually collect local information to achieve a global understanding of the input. In particular, feature redundancy has been shown to exist in existing networks [15], and separate low-frequency and high-frequency processing for all channels is unnecessary. Therefore, we design a frequency ramp inception structure which reduces the channel size of the high-frequency branches as the depth of the network increases. + +Specifically, for the input feature $X \in \mathbb{R}^{H \times W \times D}$ , where H, W and D denote the height, width and channels of $X$ , we first split along the channel dimension and divide it into a low-frequency input $X_{l} \in \mathbb{R}^{H \times W \times \alpha D}$ and a high-frequency input $X_{h} \in \mathbb{R}^{H \times W \times (1 - \alpha)D}$ . $\alpha$ denotes the partition coefficient, which determines the number of channels in the low-frequency branch. Then, for the low-frequency branch, we adopt a simple Laplace pyramid architecture. Specifically, we average pool $X_{l}$ to obtain the corresponding low-frequency components $X_{ll} \in \mathbb{R}^{\hat{H} \times \hat{W} \times \alpha D}$ and shrink the number of input tokens. Next, we recover the resolution of $X_{ll}$ based on the nearest neighbor upsampling and perform pixel-by-pixel subtraction from $X_{l}$ to obtain the high-frequency component of $X_{l}$ , i.e., $X_{lh}$ . This process can be formulated as, + +$$ +X _ {l l} = \operatorname {P o o l} (X _ {l}), \quad X _ {l h} = X _ {l} - \operatorname {U p s a m p l e} (X _ {l l}). \tag {5} +$$ + +Then, we concatenate $X_{lh}$ and $X_{h}$ along the channel to obtain the high-frequency input $X_{hh} \in \mathbb{R}^{H \times W \times D}$ , and apply a small depth-wise Convolution kernel, which usually tends to respond to the high frequencies in the input [5, 6], enhancing the high-frequency component of $X_{hh}$ . As for the low-frequency input $X_{ll}$ , we input it into the SS2D block for cross-scanning to capture the global context. The process of frequency separation processing can be described as, + +$$ +\hat {X} _ {l l} = \operatorname {S S 2 D} \left(X _ {l l}\right), \quad \hat {X} _ {h h} = \operatorname {R e p} _ {3} \left(X _ {h h}\right). \tag {6} +$$ + +where $\mathrm{Rep}_3$ denotes the reparameterized $3\times 3$ convolution. Finally, we sum up $\hat{X}_{ll}$ and $\hat{X}_{hh}$ of the corresponding channels at the element level, followed by a $1\times 1$ convolution for fusion. Through the frequency decoupling process and frequency ramp inception, the Laplace mixer can effectively weigh and integrate the high and low-frequency + +Table 2. Classification performance on ImageNet-1K. Following [30], throughput is tested on Nvidia V100 GPU with maximum power-of-two batch size that fits in memory. + +
ModelConferenceTypeParams ↓ (M)GMACs ↓Throughput ↑ (im/s)Reso.Top-1 ↑ (%)
MobileOne-S1 [49]CVPR'23CNN4.80.8354522475.9
SwiftFormer-S [43]ICCV'23CNN+ViT6.11.0262622478.5
EfficientFormerV2-S0 [27]ICCV'23CNN+ViT3.50.498530075.7
EfficientVMamba-T [40]Arxiv'24Mamba6.11.0139622476.5
Vim-Ti [56]ICML'24Mamba7.01.561730076.1
QuadMamba-Li [51]NeurlIPS'24Mamba5.40.8119022474.2
MSVMamba-N [44]NeurlIPS'24Mamba7.00.9128322477.3
MobileMamba-S6 [18]CVPR'25CNN+Mamba15.00.7365022478.0
TinyViM-S-CNN+Mamba5.60.9256322479.2
PoolFormer-S12 [52]CVPR'22Pool12.02.0190222477.2
MobileOne-S3 [49]CVPR'23CNN10.11.9190022478.1
MobileOne-S4 [49]CVPR'23CNN14.83.0122322479.4
Agent-PVT-T [14]ECCV'24ViT11.62.0144722478.4
EfficientVMamba-S [40]Arxiv'24Mamba11.01.367422478.7
QuadMamba-T [51]NeurlIPS'24Mamba10.02.058522478.2
MSVMamba-M [44]NeurlIPS'24Mamba12.01.595722479.8
MobileMamba-B1 [18]CVPR'25CNN+Mamba17.11.1209725679.9
TinyViM-B-CNN+Mamba11.01.5185122481.2
PoolFormer-S36 [52]CVPR'22Pool31.05.266722481.4
InceptionNext-T [53]CVPR'24CNN28.04.298722482.3
Agent-PVT-S [14]ECCV'24ViT20.64.068622482.4
FastViT-SA24 [48]ICCV'23CNN+ViT20.63.886122482.6
EfficientVMamba-B [40]Arxiv'24Mamba33.04.058022481.8
LocalVMamba-T [23]Arxiv'24Mamba26.05.7-22482.7
ViM-S [56]ICML'24Mamba26.05.120122480.5
VMamba-T [31]NeurlIPS'24Mamba30.04.938322482.6
QuadMamba-S [51]NeurlIPS'24Mamba31.05.554122482.4
MobileMamba-B4 [18]CVPR'25CNN+Mamba17.14.351351282.5
TinyViM-L-CNN+Mamba31.74.784322483.3
+ +Table 3. Ablation of frequency ramp inception design. + +
InceptionRampGMACsThroughputTop-1
XX0.94247879.2
✓X0.93257479.0
✓✓0.93256379.2
+ +Table 4. Ablation of the partition coefficient $\alpha$ of four stages. + +
α1α2α3α4GMACsThrou.Top-1
0.750.50.50.250.94259178.8
0.50.50.50.50.93257479.0
0.250.50.50.750.93256379.2
+ +Table 5. Ablation of the convolution kernel size in Laplace Mixer. + +
kernelaxialGMACsThroughputTop-1
3X0.94251079.0
5✓0.93259879.0
7✓0.93256379.2
9✓0.94247979.2
+ +Table 6. Ablation of re-parameterization. + +
VariantRepGMACsThrou.Top-1
TinyViM-S✓0.93256379.2
TinyViM-S✗0.93256379.2
TinyViM-L✓4.7284383.3
TinyViM-L✗4.7284383.1
+ +components of all layers. As a result, TinyViM achieves a better balance between accuracy and efficiency. + +# 4.3. Overall Architecture + +TinyViM employs a similar multi-scale backbone design similar to previous works [26, 43]. As shown in Fig. 3, given an image $I \in \mathbb{R}^{H \times W \times 3}$ , a simple stem is employed to output $X_{1} \in \mathbb{R}^{\frac{H}{4} \times \frac{W}{4} \times D_{1}}$ feature map. $H, W$ denote the height and width of the input image respectively, and $D_{1}$ is the channel number of $X_{1}$ . Note that the stem is implemented with two $3 \times 3$ convolutions with a stride of 2. Then, the output feature maps are fed into the first stage, which begins with Local Blocks. The local block consists of a reparameterized $3 \times 3$ convolution and a feed-forward network (FFN, implemented by two consecutive $1 \times 1$ convolutions) that achieves local feature extraction and feature channel mixing, respectively. The local block can be formulated as, + +$$ +\hat {X} _ {1} = \operatorname {C o n v} _ {1} (\operatorname {G e L U} (\operatorname {C o n v} _ {1} (\operatorname {R e p} _ {3} (X _ {1})))) \tag {7} +$$ + +where $\mathrm{Conv}_1$ and $\mathrm{Rep}_3$ denote the $1\times 1$ and reparameterized $3\times 3$ convolution, respectively. The features are then fed into the TinyViM block for global context-awareness. Similarly, the TinyViM block consists of a Laplace mixer and FFN for global feature extraction and channel mixing, respectively. The TinyViM Block can be described as, + +$$ +\hat {X} ^ {\prime} _ {1} = \operatorname {L a p l a c e M i x e r} (\hat {X} _ {1}) + \hat {X} _ {1}, \tag {8} +$$ + +$$ +\tilde {X} _ {1} = \operatorname {C o n v} _ {1} \left(\operatorname {G e L U} \left(\operatorname {C o n v} _ {1} \left(\hat {X} ^ {\prime} _ {1}\right)\right)\right) + \hat {X} ^ {\prime} _ {1}. \tag {9} +$$ + +There is a Patch Embedding layer between two consecutive stages for increasing the channel dimension and reducing the resolution of feature maps. Next, the resulting feature maps are subsequently fed into the second, third, and fourth stages, producing $\frac{H}{8} \times \frac{W}{8} \times D_2$ , $\frac{H}{16} \times \frac{W}{16} \times D_3$ and $\frac{H}{32} \times \frac{W}{32} \times D_4$ feature maps, respectively. + +# 5. Experiments + +We evaluate the proposed TinyViM across four downstream tasks: classification on ImageNet-1K dataset [1], object detection and instance segmentation on MS-COCO 2017 dataset [28], and semantic segmentation on ADE20K dataset [55]. All experiments are implemented following the common training settings in [43]. Due to page limits, we provide a description of the dataset, the experimental setup, details of the model's architecture and more qualitative analysis in the Supplementary Material. + +# 5.1. Image Classification + +We evaluate TinyViM's performance in image classification on ImageNet-1K [1] with some advanced Convolution, ViT and Mamba-based methods, and the comparison results are summarized in Table 2. With similar GMACs and throughput, TinyViM-S achieves a top-1 accuracy of $79.2\%$ , outperforming SwiftFormer-S by $0.7\%$ . + +In particular, TinyViM retains its performance advantages even at small and large scales. For example, TinyViMB achieves an accuracy of 81.2, which outperforms the convolution-based model MobileOne-S4 by $1.8\%$ , the ViT-based model Agent-PVT-T by $2.8\%$ , and the Mamba-based model MSVMamba-M by $1.4\%$ , respectively. These comparison results show the excellent performance of TinyViM. In addition, we compare the visualisation of Effective Receptive Field (ERF) [32] with the classical lightweight backbone, as shown in Fig. 5. The comparison models include the pure Convolution-based MobileOne-S3 and the Convolution-ViT hybrid model SwiftFormer-S. It can be observed that TinyViM-S possesses a larger ERF at a similar size. This suggests that TinyViM is able to effectively model the global context through the Laplace Mixer, thus improving the semantic recognition ability of the model. + +In terms of computational efficiency, TinyViM-S achieves a throughput of 2574 images/s, which is significantly better than other visual Mamba models. For example, the throughput of EfficientVmamba is only 1396 images/s, which is only about half of the throughput of TinyViM. This huge efficiency advantage persists on TinyViM-B and TinyViM-L, where the throughput is 2.7 and 1.5 times higher than that of the similarly scaled EfficientVmamba, respectively. Compared with other convolutional and ViT-based models, TinyViM maintains its lead in performance while its throughput is highly competitive. It should be emphasized that this competitiveness is significantly improved on A100 GPUs. This is due to the fact that Mamba has a number of hardware-friendly improvements on the A100 such as triton, which are not yet supported on the V100 GPUs. In summary, the results show that TinyViM strikes a better balance between performance and efficiency than other state-of-the-art lightweight backbones, and offers high value in real-time applications. + +# 5.2. Object Detection and Instance Segmentation + +We evaluate the proposed TinyViM for object detection and instance segmentation using the Mask R-CNN [20] framework. As shown in Table 7, TinyViM outperforms competitive models such as the pooling-based model PoolFormer, the Transformer-based models FastVit and SwiftFormer, and the Mamba-based model EfficientVMamba in both $\mathrm{AP}^{box}$ and $\mathrm{AP}^{mask}$ metrics. Specifically, compared to recent models SwiftFormer-L1, TinyViM has $1.1\%$ and $0.6\%$ improvement on $\mathrm{AP}^{box}$ and $\mathrm{AP}^{mask}$ , respectively, and has a larger throughput. Compared to EfficientVMamba-S, TinyViM-B has more significant advantages, i.e., 3.0 and 2.0 improvement in $\mathrm{AP}^{box}$ and $\mathrm{AP}^{mask}$ , respectively, with a 1.7 times throughput. When comparing with larger models such as FastViT-SA24 and EfficientVMamba-b, TinyViM still has an advantage in throughput, and shows $2.5\%$ and $0.8\%$ improvements in $\mathrm{AP}^{box}$ , respectively. These results + +demonstrate the superiority of TinyViM when transferring to detection and instance segmentation tasks. + +# 5.3. Semantic Segmentation + +Table 7 shows the semantic segmentation results of TinyViM as a backbone in comparison with other state-of-the-art backbones, where Semantic FPN [24] serves as the decode head. Specifically, compared with recent methods such as SwiftFormer, TinyVim obtains $0.8\%$ and $0.3\%$ improvement in base and large scale, respectively. Compared with the method FasterViT, TinyViM has more significant advantages, with $3.9\%$ and $3.2\%$ improvements, respectively. The experimental results verify the effectiveness of TinyViM in semantic segmentation tasks. + +# 5.4. Ablation Studies + +Effect of the Frequency Decoupling. In order to investigate the effectiveness of the frequency decoupling strategy, we perform spectral visualization of the feature maps of the low-frequency and high-frequency branches, respectively. As shown in Fig. 4, by applying Mamba, the feature maps have higher concentrations at low frequencies; by applying depth-wise convolution, the high frequencies of the model are effectively preserved. This proves that the Laplace mixer can effectively decouple the high and low frequency components of the features and specially enhance them respectively. + +Ablation of the Frequency Ramp Inception. We explore the effectiveness of the frequency ramp inception design, as shown in Table 3. We set up three experimental variants: variant 1, we implement Laplace frequency decoupling for all channels of the feature map, and input low and high frequencies into Mamba and depth-wise convolution, respectively; variant 2, we adopt a non-ramp structure, i.e., we set the segmentation coefficient $\alpha$ of the four stages to be $1/2$ ; and variant 3, we gradually increase $\alpha$ with the increase of the stage of the model. In the implementation of this paper, the $\alpha$ of the four stages are 0.25, 0.5, 0.5 and 0.75, respectively. It can be observed that, even in the small size, the throughput of variant 1 is significantly reduced and more GMACs are required. Variant 2, although more efficient, has a decreased accuracy due to the lack of adjusting the weights of the low-frequency and high-frequency components in different stages. We also explore the inverse ramp structure, as shown in Table 4. The frequency ramp used in variant 3 effectively weighs and integrates the high and low frequency information, thus helping the model to achieve a better balance between performance and efficiency. + +Ablation of the convolution kernel size in Laplace Mixer. ViM and VMamba use a $3 \times 3$ depthwise convolution before scanning. However, a larger convolution kernel is beneficial for parameters (i.e., $B$ , $C$ ) generation due to the large receptive fields, although it impairs the efficiency of + +Table 7. Object detection & instance segmentation results on MS COCO 2017 with the Mask RCNN framework. Semantic segmentation results on ADE20K by integrating models into Semantic FPN. + +
BackboneThroughput ↑ (img/s)Object DetectionInstance SegmentationSemantic mIoU
APbox50APbox75APmaskAPmask50APmask75
EfficientVMamba-T [40]16835.657.738.033.254.435.1-
TinyViM-S26238.761.141.335.757.137.938.9
PoolFormer-S12 [52]17137.359.040.134.655.836.937.2
EfficientFormer-L1 [26]19737.960.341.035.457.337.338.9
FastViT-SA12 [48]16238.960.542.235.957.638.138.0
SwiftFormer-L1 [43]17441.263.244.838.160.240.741.1
EfficientVMamba-S [40]10439.361.842.636.758.939.2-
MobileMamba-B1 [18]-40.661.843.837.458.939.940.7
TinyViM-B18042.364.246.338.761.141.341.9
PoolFormer-S36 [52]8741.063.144.837.760.140.042.0
EfficientFormer-L3 [26]11741.463.944.738.161.040.443.5
FastViT-SA24 [48]11042.063.545.838.060.540.541.0
SwiftFormer-L3 [43]12042.764.446.739.161.741.843.9
EfficientVMamba-B [40]9043.766.247.940.263.342.9-
MobileMamba-B4 [18]-40.161.843.036.958.639.242.5
TinyViM-L11144.566.448.640.763.643.844.2
+ +![](images/c3c628bfb17906b90abdda1e5e6be0efa06b04bdcc51fe75ac10ea714f968e61.jpg) +Figure 4. Fourier spectrum of TinyViM for the Low frequency branch and High frequency branch. It can be observed that we achieve the decoupling of the high and low frequency components of the features and specifically enhance them separately. + +the model. Therefore, we introduce axial depthwise convolution, which increases the convolution's receptive field along the scan path without decreasing the model efficiency. We also explore $3 \times 3$ depthwise convolution and axial convolution with kernel sizes 5 and 9, as shown in Table 5. The model achieves the best balance between performance and efficiency when the kernel size is 7. + +Ablation of re-parameterization. As shown in Table 6, the reparameterization has no improvement on TinyViM-S, but brings an increase of 0.2 to TinyViM-L. Considering that the reparameterization technique does not harm the inference efficiency of the model, we apply it for TinyViM. + +# 6. Conclusion + +Motivated by Mamba's ability to achieve global context modeling based on linear complexity, we construct a tiny and efficient hybrid vision Mamba TinyViM by combin- + +![](images/ac8bf6bfc4e3ab9a98b42cdf553e5b9a955acd83870c3f10800bdf0ada958cb6.jpg) +Figure 5. The ERF of MobileOne, SwiftFormer and TinyViM. Our TinyViM effectively obtains large ERFs with Laplace Mixer. + +ing the mobile-friendly convolution and Mamba in this paper. In contrast to previous vision Mamba efforts that simply modify scan paths in the image domain, whose performance could not compete with Convolution or Transformer based models with similar scale, TinyViM attempts to build a lightweight backbone from a frequency decoupling perspective. Specifically, we verify that the Mamba block mainly models low-frequency information in a Convolution-Mamba hybrid architecture after spectral analysis and quantitative experiments. To this end, we design a frequency decoupling strategy to preserve the high-frequency information and improve the low-frequency modeling efficiency of Mamba. In particular, we introduce a frequency ramp inception to weigh the different strengths of the low-frequency and high-frequency components at different stages of the model, so as to help the model achieve a better balance between accuracy and efficiency. Extensive experiments have shown that TinyViM outperforms other Convolution, Transformer and Mamba-based models. In future work, we plan to apply it as a lightweight backbone to replace the image encoder in SAM to explore the potential of vision Mamba in multi-modal scenarios. + +# References + +[1] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6 +[2] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11963-11975, 2022. 3 +[3] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1 +[4] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022.3 +[5] Paul Gavrikov and Janis Keuper. Can biases in imagenet models explain generalization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22184-22194, 2024. 2, 5 +[6] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, 2022. 2, 5 +[7] Alex Graves. Long Short-Term Memory, pages 37-45. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. 3 +[8] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 3, 4, 5 +[9] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487, 2020. 3 +[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3 +[11] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3 +[12] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 3 +[13] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 3 +[14] Dongchen Han, Tianzhu Ye, Yizeng Han, Zhuofan Xia, Siyuan Pan, Pengfei Wan, Shiji Song, and Gao Huang. Agent + +attention: On the integration of softmax and linear attention. In Computer Vision - ECCV 2024, pages 124-140, Cham, 2025. Springer Nature Switzerland. 2, 6 +[15] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5 +[16] Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding, 2016. 3 +[17] Ramin Hasani, Mathias Lechner, Tsun-Hsuan Wang, Makram Chahine, Alexander Amini, and Daniela Rus. Liquid structural state-space models. arXiv preprint arXiv:2209.12951, 2022.3 +[18] Haoyang He, Jiangning Zhang, Yuxuan Cai, Hongxu Chen, Xiaobin Hu, Zhenye Gan, Yabiao Wang, Chengjie Wang, Yunsheng Wu, and Lei Xie. Mobilemamba: Lightweight multi-receptive visual mamba network, 2024. 6, 8 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 3, 5 +[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7 +[21] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, and Hartwig Adam. Searching for mobilenetv3. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3 +[22] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications, 2017. 3 +[23] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan, 2024. 6 +[24] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollar. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019. 7 +[25] Dong Li, Yidi Liu, Xueyang Fu, Senyan Xu, and Zheng-Jun Zha. Fourier learning integration with state space models for image deraining, 2024. 3 +[26] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35: 12934-12949, 2022. 2, 3, 6, 8 +[27] Yanyu Li, Ju Hu, Yang Wen, Georgios Evangelidis, Kamyar Salahi, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Rethinking vision transformers for mobilenet size and speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16889-16900, 2023. 6 +[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence + +Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 +[29] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Tommi Kärkkäinen, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity, 2023. 3 +[30] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14430, 2023. 6 +[31] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 1, 2, 3, 4, 6 +[32] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 7 +[33] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3 +[34] Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer. Mega: moving average equipped gated attention. arXiv preprint arXiv:2209.10655, 2022. 3 +[35] Xiaowen Ma, Rongrong Lian, Zhenkai Wu, Renxiang Guan, Tingfeng Hong, Mengjiao Zhao, Mengting Ma, Jiangtao Nie, Zhenhong Du, Siyang Song, and Wei Zhang. A novel scene coupling semantic mask network for remote sensing image segmentation. ISPRS Journal of Photogrammetry and Remote Sensing, 221:44-63, 2025. 2 +[36] Xiaowen Ma, Zhen-Liang Ni, Shuai Xiao, and Xinghao Chen. Timepro: Efficient multivariate long-term time series forecasting with variable- and time-aware hyper-state. In Forty-second International Conference on Machine Learning, 2025. 3 +[37] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 3 +[38] Sachin Mehta, Mohammad Rastegari, Linda Shapiro, and Hannaneh Hajishirzi. Espnetv2: A light-weight, power efficient, and general purpose convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[39] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024. 2 +[40] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba, 2024. 3, 4, 5, 6, 8 +[41] Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba: + +Exploring state space models for multimodal learning, 2024. 3 +[42] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition* (CVPR), 2018. 3 +[43] Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Swiftformer: Efficient additive attention for transformer-based real-time mobile vision applications. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17425-17436, 2023. 2, 3, 4, 6, 8 +[44] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. arXiv preprint arXiv:2405.14174, 2024. 1, 3, 6 +[45] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xin chao Wang, and Shuicheng Yan. Inception transformer. In Advances in Neural Information Processing Systems, pages 23495-23509. Curran Associates, Inc., 2022. 2 +[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition, 2015. 3 +[47] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 3 +[48] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Fastvit: A fast hybrid vision transformer using structural reparameterization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5785-5795, 2023. 2, 3, 4, 6, 8 +[49] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7907-7917, 2023. 2, 4, 6 +[50] Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Repvit: Revisiting mobile cnn from vit perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15909-15920, 2024. 2, 3 +[51] Fei Xie, Weijia Zhang, Zhongdao Wang, and Chao Ma. Quadmamba: Learning quadtree-based selective scan for visual state space model, 2024. 3, 6 +[52] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10819-10829, 2022. 6, 8 +[53] Weihao Yu, Pan Zhou, Shuicheng Yan, and Xinchao Wang. Inceptionnext: When inception meets convnext. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5672-5683, 2024. 2, 6 +[54] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3 + +[55] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6 +[56] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In *Forty-first International Conference on Machine Learning*. 1, 2, 3, 6 +[57] Qinfeng Zhu, Yuanzhi Cai, Yuan Fang, Yihan Yang, Cheng Chen, Lei Fan, and Anh Nguyen. Samba: Semantic segmentation of remotely sensed images with state space model, 2024. 3 \ No newline at end of file diff --git a/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/images.zip b/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..04820299002f2026f09e91cb904981aadb5252c2 --- /dev/null +++ b/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d760732910f55b589029ec39f09bd8de3c462fb96cd2f9cb04bc8b13600730b +size 605035 diff --git a/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/layout.json b/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e17e14442d57b091eb2272f6548aefcecf2e8326 --- /dev/null +++ b/2025/TinyViM_ Frequency Decoupling for Tiny Hybrid Vision Mamba/layout.json @@ -0,0 +1,8292 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 105, + 103, + 506, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 103, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 105, + 103, + 506, + 121 + ], + "type": "text", + "content": "TinyViM: Frequency Decoupling for Tiny Hybrid Vision Mamba" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 183, + 143, + 426, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 143, + 426, + 171 + ], + "spans": [ + { + "bbox": [ + 183, + 143, + 426, + 171 + ], + "type": "text", + "content": "Xiaowen Ma * Zhenliang Ni * Xinghao Chen† Huawei Noah's Ark Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 174, + 455, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 174, + 455, + 185 + ], + "spans": [ + { + "bbox": [ + 151, + 174, + 455, + 185 + ], + "type": "text", + "content": "xwma@zju.edu.cn, {nizhenliang2, xinghao.chen}@huawei.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 297, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 610 + ], + "type": "text", + "content": "Mamba has shown great potential for computer vision due to its linear complexity in modeling the global context with respect to the input length. However, existing lightweight Mamba-based backbones cannot demonstrate performance that matches Convolution or Transformer-based methods. By observing, we find that simply modifying the scanning path in the image domain is not conducive to fully exploiting the potential of vision Mamba. In this paper, we first perform comprehensive spectral and quantitative analyses, and verify that the Mamba block mainly models low-frequency information under Convolution-Mamba hybrid architecture. Based on the analyses, we introduce a novel Laplace mixer to decouple the features in terms of frequency and input only the low-frequency components into the Mamba block. In addition, considering the redundancy of the features and the different requirements for high-frequency details and low-frequency global information at different stages, we introduce a frequency ramp inception, i.e., gradually reduce the input dimensions of the high-frequency branches, so as to efficiently trade-off the high-frequency and low-frequency components at different layers. By integrating mobile-friendly convolution and efficient Laplace mixer, we build a series of tiny hybrid vision Mamba called TinyViM. The proposed TinyViM achieves impressive performance on several downstream tasks including image classification, semantic segmentation, object detection and instance segmentation. In particular, TinyViM outperforms Convolution, Transformer and Mamba-based models with similar scales, and the throughput is about 2-3 times higher than that of other Mamba-based models. Code is available at https://github.com/xwmaxwma/TinyViM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "content": "In recent years, Mamba has gained widespread attention and application in the field of vision due to its advantages in long-sequence modeling [8]. Mamba is a selective struc" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 213, + 440, + 380 + ], + "blocks": [ + { + "bbox": [ + 317, + 213, + 440, + 380 + ], + "lines": [ + { + "bbox": [ + 317, + 213, + 440, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 440, + 380 + ], + "type": "image", + "image_path": "a8656d77bb359f04e054d3a8191f64d6b442e0a356f64302bf4280292be95af9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 392, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. Comparison of GFLOPs, Throughput and accuracy between TinyViM and others. The top-1 accuracy is tested on ImageNet-1K and the throughput is measured on Nvidia V100 GPU with maximum power-of-two batch size that fits in memory. Compared to recent Mamba-based models such as EfficientV-Mamba, QuadMamba, and VMamba, TinyViM has higher top-1 accuracy with " + }, + { + "bbox": [ + 313, + 392, + 555, + 472 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 392, + 555, + 472 + ], + "type": "text", + "content": " higher throughput and fewer GFLOPs." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 441, + 213, + 563, + 381 + ], + "blocks": [ + { + "bbox": [ + 441, + 213, + 563, + 381 + ], + "lines": [ + { + "bbox": [ + 441, + 213, + 563, + 381 + ], + "spans": [ + { + "bbox": [ + 441, + 213, + 563, + 381 + ], + "type": "image", + "image_path": "4ad6defaaaf37411d30ff5bd490200b612d065cb890789bfada110486644141d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 495, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 495, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 312, + 495, + 555, + 592 + ], + "type": "text", + "content": "tured state-space model that can overcome the limitations of local perception in Convolution neural networks [19] and the quadratic computational complexity of Transformers [3], thereby demonstrating strong capabilities in handling long-sequence data. Due to its advanced context extraction capabilities, Mamba has been widely applied in image classification tasks and has achieved advanced performance, such as Vim [56], VMamba [31], etc." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Vim [56] is one of the earliest works to apply Mamba to visual tasks, by introducing a bidirectional state-space model to process 2D image data. By introducing the Cross-Scan Module (CSM) and 2D Selective Scanning (SS2D) modules, VMamba [31] reduces the computational complexity from quadratic to linear while maintaining the global acceptance domain. MSVMamba [44] uses a multiscale 2D scanning method to apply Mamba to both original and down-sampled feature maps. This not only helps to learn long-distance dependencies, but also reduces compu" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "type": "text", + "content": "*Equal contribution, † Corresponding author" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23519" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 403, + 203 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 403, + 203 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 403, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 403, + 203 + ], + "type": "image", + "image_path": "5bd5024f51d23ddf4b352624596601f7d9fb9f70fb1daf4f2092893f4467bf9b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "type": "text", + "content": "Figure 2. Spectral analysis of Mamba. (a) Frequency magnitude " + }, + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "type": "inline_equation", + "content": "(14 \\times 14)" + }, + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "type": "text", + "content": " from 8 output channels before- and after Mamba block. (b) Relative log amplitudes of Fourier transformed feature maps. The magnitude and amplitude are averaged over 384 samples. (a) and (b) show that Mamba block focuses on capturing low-frequency information under the Convolution-Mamba hybrid architecture." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 407, + 72, + 547, + 202 + ], + "blocks": [ + { + "bbox": [ + 407, + 72, + 547, + 202 + ], + "lines": [ + { + "bbox": [ + 407, + 72, + 547, + 202 + ], + "spans": [ + { + "bbox": [ + 407, + 72, + 547, + 202 + ], + "type": "image", + "image_path": "fe08c43ef9675f8ddc3310f0a1780ac90a24720fce063b4ebef42fe5bdb16d68.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 266, + 297, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 297, + 433 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 297, + 433 + ], + "type": "text", + "content": "tational costs. EfficientVMamba [39] is a lightweight Vision Mamba variant that reduces computational complexity through dilated-based selective scanning. Although these methods inherit the advantages of Mamba linear complexity and global receptive field, they do not show competitive performance with lightweight backbones based on Convolution [49, 53] and Transformer [14, 43, 48] in computationally limited and real-time deployment scenarios, as shown in Fig. 1. In fact, these methods that simply modify the scan path in the image domain do not utilize the full potential of visual Mamba [31, 56]. In this paper, we try to rethink the design of lightweight vision Mamba from the perspective of frequency decoupling, which is a reasonable and effective solution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "content": "Through observation, we find that Mamba prefers to extract low-frequency features and ignores some high-frequency features. As shown in Fig. 2, we construct a baseline using mobile-friendly Convolution and vanilla Mamba, and perform spectral analysis of the features before and after the Mamba. Many previous works [26, 43, 50] have verified the necessity of using Convolution in lightweight backbones. After applying the Mamba block, the low-frequency component in the center of the two-dimensional spectrum map is highlighted, and the high-frequency component is suppressed. Therefore, to improve the efficiency of low-frequency modeling and keep the high-frequency features, we introduce an efficient Laplace pyramid to decompose the high- and low-frequency components of the features. Then, we input only the low-frequency components into the Mamba block to obtain a global receptive field, and for the high-frequency components, we use a reparameterized " + }, + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "content": " depth-wise Convolution to enhance high frequency information [5, 6]. It is worth noting that the size of the low frequency feature is much smaller than that of the original input feature, which can greatly reduce the computational complexity and improve the throughput of the Mamba. Therefore, the strategy of frequency decoupling and specif" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 289 + ], + "type": "text", + "content": "ically enhancement separately effectively improves the efficiency and the feature representation ability of the model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 292, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 555, + 471 + ], + "type": "text", + "content": "In addition, previous models based on Convolution or Transformer have verified that: 1) deep neural networks have feature redundancy [15]; 2) deep neural networks tend to require more high-frequency information in the shallow layer and more global information in the deep layer [35, 45]. Therefore, we propose the frequency ramp inception to further improve the performance and efficiency of the model. Specifically, we split the feature map along the channels at each stage. Early stages prioritize high frequencies, reflected in a higher channel ratio for the high-frequency branch, while later stages concentrate on low frequencies, allowing for a greater channel allocation to low frequencies. This strategic distribution enables the network to learn features that are more suitable at each stage of processing effectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 473, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 556, + 639 + ], + "type": "text", + "content": "By integrating frequency decoupling and frequency ramp inception, we propose a tiny and efficient Convolution-Mamba hybrid architecture TinyVim. We then conduct extensive experiments to validate the effectiveness of TinyViM, which achieves state-of-the-art image classification performance on ImageNet and advanced performance in downstream tasks such as detection and segmentation. Specifically, TinyViM has higher classification accuracy than other Convolution or Transformer-based backbones with a similar scale. In addition, compared to other Mamba-based backbones such as EfficientVMambaT, TinyViM has " + }, + { + "bbox": [ + 313, + 473, + 556, + 639 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 473, + 556, + 639 + ], + "type": "text", + "content": " higher throughput and " + }, + { + "bbox": [ + 313, + 473, + 556, + 639 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 313, + 473, + 556, + 639 + ], + "type": "text", + "content": " higher accuracy, as shown in Fig. 1. Our contributions can be summarized as follows." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 642, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 642, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 689 + ], + "type": "text", + "content": "- We verify that Mamba mainly models low-frequency information under the Convolution-Mamba hybrid architecture based on qualitative analysis and quantitative comparison." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "- We introduce an efficient Laplace mixer, which improves the modeling efficiency and performance of vi" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23520" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "content": "sual Mamba through frequency decoupling and frequency ramp inception." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 167 + ], + "type": "text", + "content": "- We propose TinyViM by integrating the Laplace mixer and Convolution. Extensive experiments show that TinyViM outperforms other Convolutional, Transformer and Mamba-based models with a similar scale, and achieves a better trade-off between performance and efficiency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 142, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 142, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 142, + 193 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 201, + 243, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 243, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 243, + 213 + ], + "type": "text", + "content": "2.1. Efficient Generic Vision Backbones" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 219, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 296, + 566 + ], + "type": "text", + "content": "Convolution has been widely employed for designing efficient vision backbones due to the inductive bias and efficient parallel processing on GPUs. Early works such as the MobileNet series [21, 22, 42] propose the depth-wise separable Convolution and the inverted residual blocks to further reduce the parameter and computational consumption of the plain Convolutional model [19, 46]. A series of subsequent works further improve the efficiency through depth-wise dilated Convolutions [38], network pruning [16], neural structure search [21], channel shuffling [33, 54] and structure reparameterization [48-50]. However, constrained by a limited receptive field, these efficient Convolutional networks lack global interactions between features, which degrades their performance. Although some works propose to increase the receptive field by increasing the size of the Convolutional kernel [2, 29], however, it imposes high memory access cost and computational cost, and significantly slows down the inference speed. Considering that Transformers are capable of achieving global attention but introduce quadratic complexity with respect to the resolution of the input image, a series of recent works have attempted to combine the advantages of CNNs and Transformers to develop more efficient models [26, 43, 48]. For example, and EfficientFormer only adds self-attention at more deep stages to capture global context, SwiftFormer adds effective additive attention at each stage of the model. The success of this hybrid design motivates us to explore whether a more advanced backbone can be constructed by integrating Convolution and Mamba." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 575, + 169, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 169, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 169, + 588 + ], + "type": "text", + "content": "2.2. State Space Models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "As a typical architectural paradigm for sequence-to-sequence transformations, state-space models (SSMs) play a pivotal role in handling long-range dependencies in sequence data [10, 11, 36, 47]. Early SSMs have challenges in capturing long-term dependency relationships and a unified understanding of memory [7]. To address these issues, Gu et al. enhance the ability of SSMs to capture extended dependencies using HiPPO initialization [9] and solve the continuous-time memory problem based on a linear state space layer (LSSL) [11]. However, these SSMs are still lim" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "ited by computational requirements. Subsequent works propose a series of improvements including complex diagonal structures [12, 13], multiple-input multiple-output support [47], and diagonal-plus-low-rank arithmetic [17] to improve computational efficiency and generalization across tasks. In particular, these strategies have been extended to large representation models [4, 34, 37]. Recently Mamba [8] introduce a selection mechanism that outperforms Transformers on large-scale real data through a data-dependent SSM layer with linearly scaling sequence lengths." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 192, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 431 + ], + "type": "text", + "content": "The success of Mamba has caused the vision community to explore and apply it to various vision tasks including classification [31, 56], detection [51], segmentation [57], deraining [25] and multimodal learning [41]. In particular, ViM [56] and VMamba [31] propose bi-directional scanning and cross-scanning strategies to cope with the non-sequential structure of visual data, respectively. QuadMamba [51] proposes quad-tree scanning to capture local dependencies at different granularities. MSVMamba [44] proposes multiscale two-dimensional scanning, which efficiently mitigates the long-range forgetting problem and reduces the computational cost. A highly related work is EfficientVMamba [40], which improves the scanning efficiency of a model by an atrous-based selective scan approach. However, these works tend to show low throughput and fail to perform competitively with other advanced lightweight backbones based on Convolution or Transformers. This motivates us to design the TinyViM, which obtains satisfactory speed and performance based on Laplace frequency decoupling." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 442, + 399, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 399, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 399, + 453 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "content": "State Space Models (S4). Inspired by the continuous system, State space models (SSMs) are proposed in deep learning as common sequence models [10]. These models transform an input sequence " + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "inline_equation", + "content": "x(t) \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "content": " into an output sequence " + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "inline_equation", + "content": "y(t) \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "content": " by utilizing a learnable hidden state " + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "inline_equation", + "content": "h(t) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "content": ". The process could be denoted as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 381, + 543, + 554, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 543, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 381, + 543, + 554, + 571 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} h ^ {\\prime} (t) = \\boldsymbol {A} h (t) + \\boldsymbol {B} x (t), \\tag {1} \\\\ y (t) = \\boldsymbol {C} h (t), \\\\ \\end{array}", + "image_path": "0fa442cca0c613951010080522a0570e343bdf87d9357b5f58e72c38f2acd3db.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "text", + "content": " is the evolution parameter, " + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{B}, \\mathbf{C} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "text", + "content": " denote the learnable projection parameters, and " + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 581, + 554, + 616 + ], + "type": "text", + "content": " is the state size." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Discretization. The above continuous-time SSMs are not well compatible with deep learning algorithms. Therefore, discretization is needed to align the model with the sampling frequency of the input signal to improve computational efficiency [11]. Following the previous work [13], given the sampling time scale parameter " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": ", the above continuous SSMs are discretized through zero-order hold rule, thus converting the continuous-time parameters " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "(A, B)" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23521" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 68, + 542, + 246 + ], + "blocks": [ + { + "bbox": [ + 80, + 68, + 542, + 246 + ], + "lines": [ + { + "bbox": [ + 80, + 68, + 542, + 246 + ], + "spans": [ + { + "bbox": [ + 80, + 68, + 542, + 246 + ], + "type": "image", + "image_path": "a2238a917e987242aee1d0cc9ccba8943687f34363faffc4d84de8965c94f748.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 252, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 54, + 252, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 555, + 319 + ], + "type": "text", + "content": "Figure 3. Overview of the proposed TinyViM, which has four stages and each stage consists of Local Blocks and TinyViM Blocks. The local block applies a reparameterized 3x3 Convolution to extract local features, and the TinyViM block is employed to capture the global context. The core component of the TinyViM block is the Laplace Mixer, which decouples the frequencies of the features based on an efficient Laplace pyramid and passes the state only for the low-frequencies. The enhanced low- and high-frequency components are then integrated based on a frequency ramp Inception structure. Thus, the high- and low-frequency components of different stages are appropriately trade-off and the efficiency is further improved. SS2D denotes 2D selective scanning, which is used in VMamba [31]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 339, + 257, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 257, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 257, + 352 + ], + "type": "text", + "content": "their corresponding discrete counterparts " + }, + { + "bbox": [ + 55, + 339, + 257, + 352 + ], + "type": "inline_equation", + "content": "(\\overline{A},\\overline{B})" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 355, + 294, + 375 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 355, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 107, + 355, + 294, + 375 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\overline {{\\boldsymbol {A}}} = e ^ {\\boldsymbol {\\Delta A}}, \\\\ - \\end{array} \\tag {2}", + "image_path": "39f8c4ca1bd2b3a7043953e4d4730188d83b26e8e95ded6e19faf56599507082.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 372, + 242, + 385 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 372, + 242, + 385 + ], + "spans": [ + { + "bbox": [ + 108, + 372, + 242, + 385 + ], + "type": "interline_equation", + "content": "\\overline {{B}} = (\\Delta A) ^ {- 1} (e ^ {\\Delta A} - I) \\cdot \\Delta B.", + "image_path": "67abff8c1c964cfaf590f73e08f24ad183da376c3339592b307b3e61f3ff5467.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 390, + 295, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 402 + ], + "type": "text", + "content": "Then, The discretized formulation of Eq. 1 is formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 131, + 411, + 295, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 411, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 131, + 411, + 295, + 430 + ], + "type": "interline_equation", + "content": "h _ {t} = \\bar {\\boldsymbol {A}} h _ {t - 1} + \\bar {\\boldsymbol {B}} x _ {t}, \\tag {3}", + "image_path": "0b01544445bf6c8544b0a26a70fa3cc34b8aeb75eefe9da87c0a3b2cf236ea91.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 134, + 428, + 178, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 428, + 178, + 439 + ], + "spans": [ + { + "bbox": [ + 134, + 428, + 178, + 439 + ], + "type": "interline_equation", + "content": "y _ {t} = C h _ {t},", + "image_path": "bd2341e571b0870485de6af3cb0ba2e166a8cff9838cc43e875c01f2890f3793.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 442, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 296, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 442, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\overline{A} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 55, + 442, + 296, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 442, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\overline{B} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 55, + 442, + 296, + 491 + ], + "type": "text", + "content": ". In order to improve the computational efficiency, the iterative process described in Eq. 3 can be performed by the parallel computing mode of global convolution [8]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 117, + 496, + 295, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 496, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 117, + 496, + 295, + 517 + ], + "type": "interline_equation", + "content": "y = x * \\bar {K}, \\quad - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -", + "image_path": "e116a67c7f28652a2de092476debd186cc9a444ca200982bed426c6bb4c94ea5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 85, + 510, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 510, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 85, + 510, + 296, + 528 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 85, + 510, + 296, + 528 + ], + "type": "inline_equation", + "content": "\\overline{K} = (C\\overline{B}, C\\overline{A}\\overline{B}, \\dots, C\\overline{A}^{L-1}\\overline{B})" + }, + { + "bbox": [ + 85, + 510, + 296, + 528 + ], + "type": "text", + "content": " (4)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\ast" + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "content": " denotes the convolution operation, and " + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\overline{\\pmb{K}}\\in \\mathbb{R}^{L}" + }, + { + "bbox": [ + 55, + 533, + 295, + 556 + ], + "type": "text", + "content": " serves as the kernel of the SSMs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "text", + "content": "Selective State Space Models (S6). Conventional SSMs (i.e., S4) have been implemented to capture sequence context under linear time complexity, despite the fact that they are constrained by static parameterization and cannot perform content-based reasoning. To address this problem, the selective state space model (i.e., Mamba [8]) has been proposed, which allows the model to selectively propagate or forget information based on the latitude of the current token along the length of the sequence by simply setting the parameters of the SSM as a function of the inputs. In S6, the parameters " + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "text", + "content": " are computed directly from the input sequence " + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "inline_equation", + "content": "x(t)" + }, + { + "bbox": [ + 55, + 558, + 296, + 712 + ], + "type": "text", + "content": ", thus enabling sequence-aware parameterization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 338, + 370, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 370, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 370, + 350 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 359, + 493, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 493, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 493, + 372 + ], + "type": "text", + "content": "4.1. Analysis of Vanilla Mamba Block" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 377, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 509 + ], + "type": "text", + "content": "As analyzed in Sec. 3, Mamba is able to model global context information based on linear complexity with respect to sequence length through pixel-by-pixel state transfer, which shows potential in lightweight model design. Moreover, considering the mobile-friendly and localized feature extraction capability of Convolution, a natural intuition arises: can we build a lightweight and efficient visual backbone based on Convolution and Mamba blocks? Although there have been some works [40] that attempt to do so, the performance and efficiency shown are not competitive with Convolution [49] or Transformer-based models [43, 48] so far." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": "Baseline Construction. In response to this intuition, we first build a baseline architecture using Convolution (the local block in Fig. 3) and a vanilla Mamba block (applying SS2D in VMamba [31] to adapt to the visual data), i.e., the Baseline in Table 1. Note that the baseline structure is identical to the final structure of TinyViM-S, which is designed with reference to [43], and specific structural configurations can be found in the Supplementary Material. The only difference between Baseline and TinyViM-S is that we replace the TinyViM blocks with SS2D-based vanilla Mamba blocks. We are surprised to find that this version achieves a Top-1 accuracy of " + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "inline_equation", + "content": "79.1\\%" + }, + { + "bbox": [ + 313, + 510, + 556, + 714 + ], + "type": "text", + "content": ", which is significantly better than the purely Convolutional (i.e., Conv only) version. This can be explained by the fact that Mamba can improve the expressive ability of the model by capturing the global context at each stage. However, the low throughput limits the use of the model in real-time applications. We then attempt to" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23522" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 124, + 270, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "type": "text", + "content": "Table 1. Effect of frequency input Mamba block. Throughput is tested on a Nvidia V100 GPU with maximum power-of-two batch size that fits in memory. The Top-1 accuracy is the average of three experiments to exclude random fluctuations." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 124, + 270, + 198 + ], + "lines": [ + { + "bbox": [ + 80, + 124, + 270, + 198 + ], + "spans": [ + { + "bbox": [ + 80, + 124, + 270, + 198 + ], + "type": "table", + "html": "
VariantsGMACsThroughputTop-1
Baseline0.96167379.1
Conv only0.88321277.5
Low only0.93257479.0
High only0.96137778.6
Low + High0.97150979.1
", + "image_path": "3ae2cd0ef1e145e8edcc852a9bf91e8803e5c0d503f408670f19ac5699d8e689.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 209, + 295, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 233 + ], + "type": "text", + "content": "improve the baseline based on qualitative and quantitative analysis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 234, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 296, + 484 + ], + "type": "text", + "content": "Qualitative Analysis. We first perform a spectral analysis of the feature maps before and after inputting the Mamba block in the baseline, as shown in Fig. 2. It shows that the Mamba block has strong low-frequency capture ability under the Convolution-Mamba hybrid architecture, but loses high-frequency information, such as local edges and textures. This low-frequency preference impairs the performance and efficiency of vision Mamba because 1) Populating each stage with low-frequency information degrades the high-frequency component, weakening the fine-grained recognition ability of visual Mamba, and 2) Additional high-frequency component input to a Mamba block impairs the efficiency and parallel computation of the model. A feasible strategy for this phenomenon is that we decouple the high and low frequencies of the features and input only the low-frequency components of the features into the Mamba block for hidden state transfer. Therefore, we decouple the high and low frequencies of the feature maps based on the Laplace pyramid, which is a more efficient architecture compared to other methods such as the cascaded wavelet transform, to verify the effectiveness of our strategy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 653 + ], + "type": "text", + "content": "Quantitative Analysis. We next set the inputs of the Mamba block for each stage to be low-frequency only, high-frequency only, high and low-frequency in parallel to explore the effect of the corresponding variants on the classification accuracy, respectively. As shown in Table 1, when only low frequencies are used as inputs, the model achieves " + }, + { + "bbox": [ + 55, + 485, + 296, + 653 + ], + "type": "inline_equation", + "content": "79.0\\%" + }, + { + "bbox": [ + 55, + 485, + 296, + 653 + ], + "type": "text", + "content": " accuracy, reduces the computational consumption of the model and achieves 1.5 times higher throughput. This verifies that we can significantly improve the efficiency of the model without compromising the classification accuracy by using only low-frequency inputs to Mamba. Therefore, based on the above qualitative and quantitative analyses, we design the Laplace mixer to optimize Mamba in the construction of a lightweight vision backbone." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 659, + 148, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 148, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 148, + 673 + ], + "type": "text", + "content": "4.2. Laplace Mixer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Previous work such as EfficientVMamba [40] uses dilated-based selective scanning to improve the efficiency of vanilla Mamba [8]. Despite reducing the number of input tokens," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 310 + ], + "type": "text", + "content": "EfficientVMamba performs unsatisfactorily on various visual tasks [40]. We argue that the potential of Mamba in lightweight backbones is not fully exploited simply by modifying the scanning path in the image domain. Based on the analysis in Sec. 4.1, we reconstruct the Mamba block based on the frequency decoupling and propose an elegant Laplace mixer. Furthermore, consider that in a general visual backbone, the bottom layer is more concerned with capturing high-frequency details, while the top layer is more concerned with modeling low-frequency global information [19]. As in humans, by capturing details in high-frequency components, the lower layers can capture the basic visual features and gradually collect local information to achieve a global understanding of the input. In particular, feature redundancy has been shown to exist in existing networks [15], and separate low-frequency and high-frequency processing for all channels is unnecessary. Therefore, we design a frequency ramp inception structure which reduces the channel size of the high-frequency branches as the depth of the network increases." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "content": "Specifically, for the input feature " + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "content": ", where H, W and D denote the height, width and channels of " + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "content": ", we first split along the channel dimension and divide it into a low-frequency input " + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "inline_equation", + "content": "X_{l} \\in \\mathbb{R}^{H \\times W \\times \\alpha D}" + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "content": " and a high-frequency input " + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "inline_equation", + "content": "X_{h} \\in \\mathbb{R}^{H \\times W \\times (1 - \\alpha)D}" + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "content": " denotes the partition coefficient, which determines the number of channels in the low-frequency branch. Then, for the low-frequency branch, we adopt a simple Laplace pyramid architecture. Specifically, we average pool " + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "content": " to obtain the corresponding low-frequency components " + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "inline_equation", + "content": "X_{ll} \\in \\mathbb{R}^{\\hat{H} \\times \\hat{W} \\times \\alpha D}" + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "content": " and shrink the number of input tokens. Next, we recover the resolution of " + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "inline_equation", + "content": "X_{ll}" + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "content": " based on the nearest neighbor upsampling and perform pixel-by-pixel subtraction from " + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "content": " to obtain the high-frequency component of " + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "inline_equation", + "content": "X_{l}" + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "inline_equation", + "content": "X_{lh}" + }, + { + "bbox": [ + 313, + 311, + 555, + 492 + ], + "type": "text", + "content": ". This process can be formulated as," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 498, + 555, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 498, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 326, + 498, + 555, + 513 + ], + "type": "interline_equation", + "content": "X _ {l l} = \\operatorname {P o o l} (X _ {l}), \\quad X _ {l h} = X _ {l} - \\operatorname {U p s a m p l e} (X _ {l l}). \\tag {5}", + "image_path": "669eb373c9f8904a631d3eeab63648691bc74209046a482aa83bb8248796003e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "type": "text", + "content": "Then, we concatenate " + }, + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "type": "inline_equation", + "content": "X_{lh}" + }, + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "type": "inline_equation", + "content": "X_{h}" + }, + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "type": "text", + "content": " along the channel to obtain the high-frequency input " + }, + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "type": "inline_equation", + "content": "X_{hh} \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "type": "text", + "content": ", and apply a small depth-wise Convolution kernel, which usually tends to respond to the high frequencies in the input [5, 6], enhancing the high-frequency component of " + }, + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "type": "inline_equation", + "content": "X_{hh}" + }, + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "type": "text", + "content": ". As for the low-frequency input " + }, + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "type": "inline_equation", + "content": "X_{ll}" + }, + { + "bbox": [ + 313, + 517, + 556, + 615 + ], + "type": "text", + "content": ", we input it into the SS2D block for cross-scanning to capture the global context. The process of frequency separation processing can be described as," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 348, + 624, + 555, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 624, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 348, + 624, + 555, + 639 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {l l} = \\operatorname {S S 2 D} \\left(X _ {l l}\\right), \\quad \\hat {X} _ {h h} = \\operatorname {R e p} _ {3} \\left(X _ {h h}\\right). \\tag {6}", + "image_path": "7ce1e9b265b39f1458bdebed47436c061fc022bfe90f7655e33a37c149bfb1ee.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{Rep}_3" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " denotes the reparameterized " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " convolution. Finally, we sum up " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{X}_{ll}" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{X}_{hh}" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " of the corresponding channels at the element level, followed by a " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " convolution for fusion. Through the frequency decoupling process and frequency ramp inception, the Laplace mixer can effectively weigh and integrate the high and low-frequency" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23523" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 102, + 394, + 361 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 395, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 395, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 395, + 94 + ], + "type": "text", + "content": "Table 2. Classification performance on ImageNet-1K. Following [30], throughput is tested on Nvidia V100 GPU with maximum power-of-two batch size that fits in memory." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 102, + 394, + 361 + ], + "lines": [ + { + "bbox": [ + 59, + 102, + 394, + 361 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 394, + 361 + ], + "type": "table", + "html": "
ModelConferenceTypeParams ↓ (M)GMACs ↓Throughput ↑ (im/s)Reso.Top-1 ↑ (%)
MobileOne-S1 [49]CVPR'23CNN4.80.8354522475.9
SwiftFormer-S [43]ICCV'23CNN+ViT6.11.0262622478.5
EfficientFormerV2-S0 [27]ICCV'23CNN+ViT3.50.498530075.7
EfficientVMamba-T [40]Arxiv'24Mamba6.11.0139622476.5
Vim-Ti [56]ICML'24Mamba7.01.561730076.1
QuadMamba-Li [51]NeurlIPS'24Mamba5.40.8119022474.2
MSVMamba-N [44]NeurlIPS'24Mamba7.00.9128322477.3
MobileMamba-S6 [18]CVPR'25CNN+Mamba15.00.7365022478.0
TinyViM-S-CNN+Mamba5.60.9256322479.2
PoolFormer-S12 [52]CVPR'22Pool12.02.0190222477.2
MobileOne-S3 [49]CVPR'23CNN10.11.9190022478.1
MobileOne-S4 [49]CVPR'23CNN14.83.0122322479.4
Agent-PVT-T [14]ECCV'24ViT11.62.0144722478.4
EfficientVMamba-S [40]Arxiv'24Mamba11.01.367422478.7
QuadMamba-T [51]NeurlIPS'24Mamba10.02.058522478.2
MSVMamba-M [44]NeurlIPS'24Mamba12.01.595722479.8
MobileMamba-B1 [18]CVPR'25CNN+Mamba17.11.1209725679.9
TinyViM-B-CNN+Mamba11.01.5185122481.2
PoolFormer-S36 [52]CVPR'22Pool31.05.266722481.4
InceptionNext-T [53]CVPR'24CNN28.04.298722482.3
Agent-PVT-S [14]ECCV'24ViT20.64.068622482.4
FastViT-SA24 [48]ICCV'23CNN+ViT20.63.886122482.6
EfficientVMamba-B [40]Arxiv'24Mamba33.04.058022481.8
LocalVMamba-T [23]Arxiv'24Mamba26.05.7-22482.7
ViM-S [56]ICML'24Mamba26.05.120122480.5
VMamba-T [31]NeurlIPS'24Mamba30.04.938322482.6
QuadMamba-S [51]NeurlIPS'24Mamba31.05.554122482.4
MobileMamba-B4 [18]CVPR'25CNN+Mamba17.14.351351282.5
TinyViM-L-CNN+Mamba31.74.784322483.3
", + "image_path": "0014ce45d9c7c1ac0438c98b9a135d13b619c2f625b61b4e19714deb04e40d3d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 406, + 96, + 556, + 140 + ], + "blocks": [ + { + "bbox": [ + 405, + 71, + 557, + 93 + ], + "lines": [ + { + "bbox": [ + 405, + 71, + 557, + 93 + ], + "spans": [ + { + "bbox": [ + 405, + 71, + 557, + 93 + ], + "type": "text", + "content": "Table 3. Ablation of frequency ramp inception design." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 406, + 96, + 556, + 140 + ], + "lines": [ + { + "bbox": [ + 406, + 96, + 556, + 140 + ], + "spans": [ + { + "bbox": [ + 406, + 96, + 556, + 140 + ], + "type": "table", + "html": "
InceptionRampGMACsThroughputTop-1
XX0.94247879.2
✓X0.93257479.0
✓✓0.93256379.2
", + "image_path": "e6733ab2d548048f5310c74df5dc89ac45d56b51195ae4d156e509f731c848f3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 407, + 168, + 555, + 209 + ], + "blocks": [ + { + "bbox": [ + 406, + 143, + 557, + 165 + ], + "lines": [ + { + "bbox": [ + 406, + 143, + 557, + 165 + ], + "spans": [ + { + "bbox": [ + 406, + 143, + 557, + 165 + ], + "type": "text", + "content": "Table 4. Ablation of the partition coefficient " + }, + { + "bbox": [ + 406, + 143, + 557, + 165 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 406, + 143, + 557, + 165 + ], + "type": "text", + "content": " of four stages." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 407, + 168, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 407, + 168, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 407, + 168, + 555, + 209 + ], + "type": "table", + "html": "
α1α2α3α4GMACsThrou.Top-1
0.750.50.50.250.94259178.8
0.50.50.50.50.93257479.0
0.250.50.50.750.93256379.2
", + "image_path": "f658d47432438eeba0085e3f995d3ae8e06348cb4ca849e44dc5a67337fc52d6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 412, + 237, + 550, + 291 + ], + "blocks": [ + { + "bbox": [ + 406, + 211, + 557, + 233 + ], + "lines": [ + { + "bbox": [ + 406, + 211, + 557, + 233 + ], + "spans": [ + { + "bbox": [ + 406, + 211, + 557, + 233 + ], + "type": "text", + "content": "Table 5. Ablation of the convolution kernel size in Laplace Mixer." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 412, + 237, + 550, + 291 + ], + "lines": [ + { + "bbox": [ + 412, + 237, + 550, + 291 + ], + "spans": [ + { + "bbox": [ + 412, + 237, + 550, + 291 + ], + "type": "table", + "html": "
kernelaxialGMACsThroughputTop-1
3X0.94251079.0
5✓0.93259879.0
7✓0.93256379.2
9✓0.94247979.2
", + "image_path": "34f3ffc9aa1fd7130e772436966c0e1d3033a4b06ed2179133093501b92058a8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 413, + 308, + 552, + 361 + ], + "blocks": [ + { + "bbox": [ + 407, + 293, + 556, + 304 + ], + "lines": [ + { + "bbox": [ + 407, + 293, + 556, + 304 + ], + "spans": [ + { + "bbox": [ + 407, + 293, + 556, + 304 + ], + "type": "text", + "content": "Table 6. Ablation of re-parameterization." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 413, + 308, + 552, + 361 + ], + "lines": [ + { + "bbox": [ + 413, + 308, + 552, + 361 + ], + "spans": [ + { + "bbox": [ + 413, + 308, + 552, + 361 + ], + "type": "table", + "html": "
VariantRepGMACsThrou.Top-1
TinyViM-S✓0.93256379.2
TinyViM-S✗0.93256379.2
TinyViM-L✓4.7284383.3
TinyViM-L✗4.7284383.1
", + "image_path": "3532c394d071fa8174b95e8804de0a61c0bb1b87f60dddf19fc3113e3fbb75ba.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 374, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 399 + ], + "type": "text", + "content": "components of all layers. As a result, TinyViM achieves a better balance between accuracy and efficiency." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 405, + 176, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 176, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 176, + 417 + ], + "type": "text", + "content": "4.3. Overall Architecture" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "text", + "content": "TinyViM employs a similar multi-scale backbone design similar to previous works [26, 43]. As shown in Fig. 3, given an image " + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "text", + "content": ", a simple stem is employed to output " + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "inline_equation", + "content": "X_{1} \\in \\mathbb{R}^{\\frac{H}{4} \\times \\frac{W}{4} \\times D_{1}}" + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "text", + "content": " feature map. " + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "text", + "content": " denote the height and width of the input image respectively, and " + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "inline_equation", + "content": "D_{1}" + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "text", + "content": " is the channel number of " + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "inline_equation", + "content": "X_{1}" + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "text", + "content": ". Note that the stem is implemented with two " + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "text", + "content": " convolutions with a stride of 2. Then, the output feature maps are fed into the first stage, which begins with Local Blocks. The local block consists of a reparameterized " + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "text", + "content": " convolution and a feed-forward network (FFN, implemented by two consecutive " + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 54, + 422, + 295, + 590 + ], + "type": "text", + "content": " convolutions) that achieves local feature extraction and feature channel mixing, respectively. The local block can be formulated as," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 86, + 598, + 295, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 598, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 86, + 598, + 295, + 613 + ], + "type": "interline_equation", + "content": "\\hat {X} _ {1} = \\operatorname {C o n v} _ {1} (\\operatorname {G e L U} (\\operatorname {C o n v} _ {1} (\\operatorname {R e p} _ {3} (X _ {1})))) \\tag {7}", + "image_path": "4c04c5d454ba45823196b90b4976678a0cb49ca29f23ba1f55b1fd9d71f15308.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_1" + }, + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\mathrm{Rep}_3" + }, + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "type": "text", + "content": " denote the " + }, + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "type": "text", + "content": " and reparameterized " + }, + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 55, + 620, + 295, + 693 + ], + "type": "text", + "content": " convolution, respectively. The features are then fed into the TinyViM block for global context-awareness. Similarly, the TinyViM block consists of a Laplace mixer and FFN for global feature extraction and channel mixing, respectively. The TinyViM Block can be described as," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 105, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 105, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\hat {X} ^ {\\prime} _ {1} = \\operatorname {L a p l a c e M i x e r} (\\hat {X} _ {1}) + \\hat {X} _ {1}, \\tag {8}", + "image_path": "8c747d0f1b90f2bbd6663d66b60affc2bb3f233aa9730a34f6215813ffe70eac.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 385, + 555, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 385, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 343, + 385, + 555, + 399 + ], + "type": "interline_equation", + "content": "\\tilde {X} _ {1} = \\operatorname {C o n v} _ {1} \\left(\\operatorname {G e L U} \\left(\\operatorname {C o n v} _ {1} \\left(\\hat {X} ^ {\\prime} _ {1}\\right)\\right)\\right) + \\hat {X} ^ {\\prime} _ {1}. \\tag {9}", + "image_path": "d08e6f8f01a92e3b3c2f5ecedba2b449c0c111ab537215d8b6f7dd38bf1890a7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 404, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 555, + 479 + ], + "type": "text", + "content": "There is a Patch Embedding layer between two consecutive stages for increasing the channel dimension and reducing the resolution of feature maps. Next, the resulting feature maps are subsequently fed into the second, third, and fourth stages, producing " + }, + { + "bbox": [ + 313, + 404, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\frac{H}{8} \\times \\frac{W}{8} \\times D_2" + }, + { + "bbox": [ + 313, + 404, + 555, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 404, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\frac{H}{16} \\times \\frac{W}{16} \\times D_3" + }, + { + "bbox": [ + 313, + 404, + 555, + 479 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 404, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\frac{H}{32} \\times \\frac{W}{32} \\times D_4" + }, + { + "bbox": [ + 313, + 404, + 555, + 479 + ], + "type": "text", + "content": " feature maps, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 488, + 395, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 395, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 395, + 502 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 509, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 555, + 617 + ], + "type": "text", + "content": "We evaluate the proposed TinyViM across four downstream tasks: classification on ImageNet-1K dataset [1], object detection and instance segmentation on MS-COCO 2017 dataset [28], and semantic segmentation on ADE20K dataset [55]. All experiments are implemented following the common training settings in [43]. Due to page limits, we provide a description of the dataset, the experimental setup, details of the model's architecture and more qualitative analysis in the Supplementary Material." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 624, + 432, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 432, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 432, + 636 + ], + "type": "text", + "content": "5.1. Image Classification" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "We evaluate TinyViM's performance in image classification on ImageNet-1K [1] with some advanced Convolution, ViT and Mamba-based methods, and the comparison results are summarized in Table 2. With similar GMACs and throughput, TinyViM-S achieves a top-1 accuracy of " + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "79.2\\%" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": ", outperforming SwiftFormer-S by " + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 641, + 555, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23524" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "content": "In particular, TinyViM retains its performance advantages even at small and large scales. For example, TinyViMB achieves an accuracy of 81.2, which outperforms the convolution-based model MobileOne-S4 by " + }, + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "content": ", the ViT-based model Agent-PVT-T by " + }, + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "inline_equation", + "content": "2.8\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "content": ", and the Mamba-based model MSVMamba-M by " + }, + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 263 + ], + "type": "text", + "content": ", respectively. These comparison results show the excellent performance of TinyViM. In addition, we compare the visualisation of Effective Receptive Field (ERF) [32] with the classical lightweight backbone, as shown in Fig. 5. The comparison models include the pure Convolution-based MobileOne-S3 and the Convolution-ViT hybrid model SwiftFormer-S. It can be observed that TinyViM-S possesses a larger ERF at a similar size. This suggests that TinyViM is able to effectively model the global context through the Laplace Mixer, thus improving the semantic recognition ability of the model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 264, + 295, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 492 + ], + "type": "text", + "content": "In terms of computational efficiency, TinyViM-S achieves a throughput of 2574 images/s, which is significantly better than other visual Mamba models. For example, the throughput of EfficientVmamba is only 1396 images/s, which is only about half of the throughput of TinyViM. This huge efficiency advantage persists on TinyViM-B and TinyViM-L, where the throughput is 2.7 and 1.5 times higher than that of the similarly scaled EfficientVmamba, respectively. Compared with other convolutional and ViT-based models, TinyViM maintains its lead in performance while its throughput is highly competitive. It should be emphasized that this competitiveness is significantly improved on A100 GPUs. This is due to the fact that Mamba has a number of hardware-friendly improvements on the A100 such as triton, which are not yet supported on the V100 GPUs. In summary, the results show that TinyViM strikes a better balance between performance and efficiency than other state-of-the-art lightweight backbones, and offers high value in real-time applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 503, + 286, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 286, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 286, + 516 + ], + "type": "text", + "content": "5.2. Object Detection and Instance Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "We evaluate the proposed TinyViM for object detection and instance segmentation using the Mask R-CNN [20] framework. As shown in Table 7, TinyViM outperforms competitive models such as the pooling-based model PoolFormer, the Transformer-based models FastVit and SwiftFormer, and the Mamba-based model EfficientVMamba in both " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^{box}" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^{mask}" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " metrics. Specifically, compared to recent models SwiftFormer-L1, TinyViM has " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " improvement on " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^{box}" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^{mask}" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": ", respectively, and has a larger throughput. Compared to EfficientVMamba-S, TinyViM-B has more significant advantages, i.e., 3.0 and 2.0 improvement in " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^{box}" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^{mask}" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": ", respectively, with a 1.7 times throughput. When comparing with larger models such as FastViT-SA24 and EfficientVMamba-b, TinyViM still has an advantage in throughput, and shows " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": " improvements in " + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}^{box}" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": ", respectively. These results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "demonstrate the superiority of TinyViM when transferring to detection and instance segmentation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 105, + 447, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 105, + 447, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 105, + 447, + 118 + ], + "type": "text", + "content": "5.3. Semantic Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 123, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 123, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 123, + 553, + 243 + ], + "type": "text", + "content": "Table 7 shows the semantic segmentation results of TinyViM as a backbone in comparison with other state-of-the-art backbones, where Semantic FPN [24] serves as the decode head. Specifically, compared with recent methods such as SwiftFormer, TinyVim obtains " + }, + { + "bbox": [ + 313, + 123, + 553, + 243 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 313, + 123, + 553, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 123, + 553, + 243 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 313, + 123, + 553, + 243 + ], + "type": "text", + "content": " improvement in base and large scale, respectively. Compared with the method FasterViT, TinyViM has more significant advantages, with " + }, + { + "bbox": [ + 313, + 123, + 553, + 243 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 313, + 123, + 553, + 243 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 123, + 553, + 243 + ], + "type": "inline_equation", + "content": "3.2\\%" + }, + { + "bbox": [ + 313, + 123, + 553, + 243 + ], + "type": "text", + "content": " improvements, respectively. The experimental results verify the effectiveness of TinyViM in semantic segmentation tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 251, + 415, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 415, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 415, + 264 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 270, + 553, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 400 + ], + "type": "text", + "content": "Effect of the Frequency Decoupling. In order to investigate the effectiveness of the frequency decoupling strategy, we perform spectral visualization of the feature maps of the low-frequency and high-frequency branches, respectively. As shown in Fig. 4, by applying Mamba, the feature maps have higher concentrations at low frequencies; by applying depth-wise convolution, the high frequencies of the model are effectively preserved. This proves that the Laplace mixer can effectively decouple the high and low frequency components of the features and specially enhance them respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 402, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 553, + 653 + ], + "type": "text", + "content": "Ablation of the Frequency Ramp Inception. We explore the effectiveness of the frequency ramp inception design, as shown in Table 3. We set up three experimental variants: variant 1, we implement Laplace frequency decoupling for all channels of the feature map, and input low and high frequencies into Mamba and depth-wise convolution, respectively; variant 2, we adopt a non-ramp structure, i.e., we set the segmentation coefficient " + }, + { + "bbox": [ + 313, + 402, + 553, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 402, + 553, + 653 + ], + "type": "text", + "content": " of the four stages to be " + }, + { + "bbox": [ + 313, + 402, + 553, + 653 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 313, + 402, + 553, + 653 + ], + "type": "text", + "content": "; and variant 3, we gradually increase " + }, + { + "bbox": [ + 313, + 402, + 553, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 402, + 553, + 653 + ], + "type": "text", + "content": " with the increase of the stage of the model. In the implementation of this paper, the " + }, + { + "bbox": [ + 313, + 402, + 553, + 653 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 402, + 553, + 653 + ], + "type": "text", + "content": " of the four stages are 0.25, 0.5, 0.5 and 0.75, respectively. It can be observed that, even in the small size, the throughput of variant 1 is significantly reduced and more GMACs are required. Variant 2, although more efficient, has a decreased accuracy due to the lack of adjusting the weights of the low-frequency and high-frequency components in different stages. We also explore the inverse ramp structure, as shown in Table 4. The frequency ramp used in variant 3 effectively weighs and integrates the high and low frequency information, thus helping the model to achieve a better balance between performance and efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Ablation of the convolution kernel size in Laplace Mixer. ViM and VMamba use a " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": " depthwise convolution before scanning. However, a larger convolution kernel is beneficial for parameters (i.e., " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": ") generation due to the large receptive fields, although it impairs the efficiency of" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23525" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 96, + 527, + 304 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 7. Object detection & instance segmentation results on MS COCO 2017 with the Mask RCNN framework. Semantic segmentation results on ADE20K by integrating models into Semantic FPN." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 96, + 527, + 304 + ], + "lines": [ + { + "bbox": [ + 82, + 96, + 527, + 304 + ], + "spans": [ + { + "bbox": [ + 82, + 96, + 527, + 304 + ], + "type": "table", + "html": "
BackboneThroughput ↑ (img/s)Object DetectionInstance SegmentationSemantic mIoU
APbox50APbox75APmaskAPmask50APmask75
EfficientVMamba-T [40]16835.657.738.033.254.435.1-
TinyViM-S26238.761.141.335.757.137.938.9
PoolFormer-S12 [52]17137.359.040.134.655.836.937.2
EfficientFormer-L1 [26]19737.960.341.035.457.337.338.9
FastViT-SA12 [48]16238.960.542.235.957.638.138.0
SwiftFormer-L1 [43]17441.263.244.838.160.240.741.1
EfficientVMamba-S [40]10439.361.842.636.758.939.2-
MobileMamba-B1 [18]-40.661.843.837.458.939.940.7
TinyViM-B18042.364.246.338.761.141.341.9
PoolFormer-S36 [52]8741.063.144.837.760.140.042.0
EfficientFormer-L3 [26]11741.463.944.738.161.040.443.5
FastViT-SA24 [48]11042.063.545.838.060.540.541.0
SwiftFormer-L3 [43]12042.764.446.739.161.741.843.9
EfficientVMamba-B [40]9043.766.247.940.263.342.9-
MobileMamba-B4 [18]-40.161.843.036.958.639.242.5
TinyViM-L11144.566.448.640.763.643.844.2
", + "image_path": "0616b78a67bd2b407dd139d7eab16940aef7c4b3b7732106bc12257842a9ca92.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 317, + 294, + 430 + ], + "blocks": [ + { + "bbox": [ + 62, + 317, + 294, + 430 + ], + "lines": [ + { + "bbox": [ + 62, + 317, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 294, + 430 + ], + "type": "image", + "image_path": "c3c628bfb17906b90abdda1e5e6be0efa06b04bdcc51fe75ac10ea714f968e61.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 439, + 295, + 485 + ], + "lines": [ + { + "bbox": [ + 55, + 439, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 295, + 485 + ], + "type": "text", + "content": "Figure 4. Fourier spectrum of TinyViM for the Low frequency branch and High frequency branch. It can be observed that we achieve the decoupling of the high and low frequency components of the features and specifically enhance them separately." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 501, + 295, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 584 + ], + "type": "text", + "content": "the model. Therefore, we introduce axial depthwise convolution, which increases the convolution's receptive field along the scan path without decreasing the model efficiency. We also explore " + }, + { + "bbox": [ + 55, + 501, + 295, + 584 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 501, + 295, + 584 + ], + "type": "text", + "content": " depthwise convolution and axial convolution with kernel sizes 5 and 9, as shown in Table 5. The model achieves the best balance between performance and efficiency when the kernel size is 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 586, + 295, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 645 + ], + "type": "text", + "content": "Ablation of re-parameterization. As shown in Table 6, the reparameterization has no improvement on TinyViM-S, but brings an increase of 0.2 to TinyViM-L. Considering that the reparameterization technique does not harm the inference efficiency of the model, we apply it for TinyViM." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 659, + 128, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 128, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 128, + 672 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 679, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 679, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 679, + 295, + 713 + ], + "type": "text", + "content": "Motivated by Mamba's ability to achieve global context modeling based on linear complexity, we construct a tiny and efficient hybrid vision Mamba TinyViM by combin-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 323, + 316, + 545, + 407 + ], + "blocks": [ + { + "bbox": [ + 323, + 316, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 323, + 316, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 323, + 316, + 545, + 407 + ], + "type": "image", + "image_path": "ac8bf6bfc4e3ab9a98b42cdf553e5b9a955acd83870c3f10800bdf0ada958cb6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 420, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 313, + 420, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 555, + 444 + ], + "type": "text", + "content": "Figure 5. The ERF of MobileOne, SwiftFormer and TinyViM. Our TinyViM effectively obtains large ERFs with Laplace Mixer." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 471, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 555, + 713 + ], + "type": "text", + "content": "ing the mobile-friendly convolution and Mamba in this paper. In contrast to previous vision Mamba efforts that simply modify scan paths in the image domain, whose performance could not compete with Convolution or Transformer based models with similar scale, TinyViM attempts to build a lightweight backbone from a frequency decoupling perspective. Specifically, we verify that the Mamba block mainly models low-frequency information in a Convolution-Mamba hybrid architecture after spectral analysis and quantitative experiments. To this end, we design a frequency decoupling strategy to preserve the high-frequency information and improve the low-frequency modeling efficiency of Mamba. In particular, we introduce a frequency ramp inception to weigh the different strengths of the low-frequency and high-frequency components at different stages of the model, so as to help the model achieve a better balance between accuracy and efficiency. Extensive experiments have shown that TinyViM outperforms other Convolution, Transformer and Mamba-based models. In future work, we plan to apply it as a lightweight backbone to replace the image encoder in SAM to explore the potential of vision Mamba in multi-modal scenarios." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23526" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 94, + 115, + 106 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 114, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 114, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 295, + 158 + ], + "type": "text", + "content": "[1] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 159, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 295, + 213 + ], + "type": "text", + "content": "[2] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11963-11975, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 214, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 278 + ], + "type": "text", + "content": "[3] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 280, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 294, + 323 + ], + "type": "text", + "content": "[4] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 324, + 294, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 294, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 294, + 368 + ], + "type": "text", + "content": "[5] Paul Gavrikov and Janis Keuper. Can biases in imagenet models explain generalization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22184-22194, 2024. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 413 + ], + "type": "text", + "content": "[6] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, 2022. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 414, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 435 + ], + "type": "text", + "content": "[7] Alex Graves. Long Short-Term Memory, pages 37-45. Springer Berlin Heidelberg, Berlin, Heidelberg, 2012. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 294, + 468 + ], + "type": "text", + "content": "[8] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 3, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 294, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 294, + 512 + ], + "type": "text", + "content": "[9] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial projections. Advances in neural information processing systems, 33:1474-1487, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "type": "text", + "content": "[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 294, + 601 + ], + "type": "text", + "content": "[11] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "type": "text", + "content": "[12] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[13] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[14] Dongchen Han, Tianzhu Ye, Yizeng Han, Zhuofan Xia, Siyuan Pan, Pengfei Wan, Shiji Song, and Gao Huang. Agent" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "attention: On the integration of softmax and linear attention. In Computer Vision - ECCV 2024, pages 124-140, Cham, 2025. Springer Nature Switzerland. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 106, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 553, + 150 + ], + "type": "text", + "content": "[15] Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, and Chang Xu. Ghostnet: More features from cheap operations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 151, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 183 + ], + "type": "text", + "content": "[16] Song Han, Huizi Mao, and William J. Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding, 2016. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 184, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 227 + ], + "type": "text", + "content": "[17] Ramin Hasani, Mathias Lechner, Tsun-Hsuan Wang, Makram Chahine, Alexander Amini, and Daniela Rus. Liquid structural state-space models. arXiv preprint arXiv:2209.12951, 2022.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 271 + ], + "type": "text", + "content": "[18] Haoyang He, Jiangning Zhang, Yuxuan Cai, Hongxu Chen, Xiaobin Hu, Zhenye Gan, Yabiao Wang, Chengjie Wang, Yunsheng Wu, and Lei Xie. Mobilemamba: Lightweight multi-receptive visual mamba network, 2024. 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 316, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 349 + ], + "type": "text", + "content": "[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "type": "text", + "content": "[21] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, Quoc V. Le, and Hartwig Adam. Searching for mobilenetv3. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 415, + 553, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 415, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 317, + 415, + 553, + 468 + ], + "type": "text", + "content": "[22] Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 470, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 470, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 470, + 553, + 503 + ], + "type": "text", + "content": "[23] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan, 2024. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 504, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 548 + ], + "type": "text", + "content": "[24] Alexander Kirillov, Ross Girshick, Kaiming He, and Piotr Dollar. Panoptic feature pyramid networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6399-6408, 2019. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 548, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 581 + ], + "type": "text", + "content": "[25] Dong Li, Yidi Liu, Xueyang Fu, Senyan Xu, and Zheng-Jun Zha. Fourier learning integration with state space models for image deraining, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 553, + 635 + ], + "type": "text", + "content": "[26] Yanyu Li, Geng Yuan, Yang Wen, Ju Hu, Georgios Evangelidis, Sergey Tulyakov, Yanzhi Wang, and Jian Ren. Efficientformer: Vision transformers at mobilenet speed. Advances in Neural Information Processing Systems, 35: 12934-12949, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 635, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 691 + ], + "type": "text", + "content": "[27] Yanyu Li, Ju Hu, Yang Wen, Georgios Evangelidis, Kamyar Salahi, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Rethinking vision transformers for mobilenet size and speed. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16889-16900, 2023. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23527" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 117 + ], + "type": "text", + "content": "Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 172 + ], + "type": "text", + "content": "[29] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Tommi Kärkkäinen, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 295, + 229 + ], + "type": "text", + "content": "[30] Xinyu Liu, Houwen Peng, Ningxin Zheng, Yuqing Yang, Han Hu, and Yixuan Yuan. Efficientvit: Memory efficient vision transformer with cascaded group attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14430, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 274 + ], + "type": "text", + "content": "[31] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 295, + 319 + ], + "type": "text", + "content": "[32] Wenjie Luo, Yujia Li, Raquel Urtasun, and Richard Zemel. Understanding the effective receptive field in deep convolutional neural networks. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2016. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 295, + 364 + ], + "type": "text", + "content": "[33] Ningning Ma, Xiangyu Zhang, Hai-Tao Zheng, and Jian Sun. Shufflenet v2: Practical guidelines for efficient cnn architecture design. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "type": "text", + "content": "[34] Xuezhe Ma, Chunting Zhou, Xiang Kong, Junxian He, Liangke Gui, Graham Neubig, Jonathan May, and Luke Zettlemoyer. Mega: moving average equipped gated attention. arXiv preprint arXiv:2209.10655, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 410, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 295, + 476 + ], + "type": "text", + "content": "[35] Xiaowen Ma, Rongrong Lian, Zhenkai Wu, Renxiang Guan, Tingfeng Hong, Mengjiao Zhao, Mengting Ma, Jiangtao Nie, Zhenhong Du, Siyang Song, and Wei Zhang. A novel scene coupling semantic mask network for remote sensing image segmentation. ISPRS Journal of Photogrammetry and Remote Sensing, 221:44-63, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 477, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 295, + 532 + ], + "type": "text", + "content": "[36] Xiaowen Ma, Zhen-Liang Ni, Shuai Xiao, and Xinghao Chen. Timepro: Efficient multivariate long-term time series forecasting with variable- and time-aware hyper-state. In Forty-second International Conference on Machine Learning, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 533, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 295, + 566 + ], + "type": "text", + "content": "[37] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 295, + 622 + ], + "type": "text", + "content": "[38] Sachin Mehta, Mohammad Rastegari, Linda Shapiro, and Hannaneh Hajishirzi. Espnetv2: A light-weight, power efficient, and general purpose convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 656 + ], + "type": "text", + "content": "[39] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba. arXiv preprint arXiv:2403.09977, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 657, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 295, + 689 + ], + "type": "text", + "content": "[40] Xiaohuan Pei, Tao Huang, and Chang Xu. Efficientvmamba: Atrous selective scan for light weight visual mamba, 2024. 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 295, + 714 + ], + "type": "text", + "content": "[41] Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "type": "text", + "content": "Exploring state space models for multimodal learning, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "text", + "content": "[42] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition* (CVPR), 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 150, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 217 + ], + "type": "text", + "content": "[43] Abdelrahman Shaker, Muhammad Maaz, Hanoona Rasheed, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Swiftformer: Efficient additive attention for transformer-based real-time mobile vision applications. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17425-17436, 2023. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "type": "text", + "content": "[44] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. arXiv preprint arXiv:2405.14174, 2024. 1, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 294 + ], + "type": "text", + "content": "[45] Chenyang Si, Weihao Yu, Pan Zhou, Yichen Zhou, Xin chao Wang, and Shuicheng Yan. Inception transformer. In Advances in Neural Information Processing Systems, pages 23495-23509. Curran Associates, Inc., 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 316 + ], + "type": "text", + "content": "[46] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition, 2015. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 316, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 349 + ], + "type": "text", + "content": "[47] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 414 + ], + "type": "text", + "content": "[48] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Fastvit: A fast hybrid vision transformer using structural reparameterization. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5785-5795, 2023. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 415, + 555, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 555, + 470 + ], + "type": "text", + "content": "[49] Pavan Kumar Anasosalu Vasu, James Gabriel, Jeff Zhu, Oncel Tuzel, and Anurag Ranjan. Mobileone: An improved one millisecond mobile backbone. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7907-7917, 2023. 2, 4, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "text", + "content": "[50] Ao Wang, Hui Chen, Zijia Lin, Jungong Han, and Guiguang Ding. Repvit: Revisiting mobile cnn from vit perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15909-15920, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 525, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 558 + ], + "type": "text", + "content": "[51] Fei Xie, Weijia Zhang, Zhongdao Wang, and Chao Ma. Quadmamba: Learning quadtree-based selective scan for visual state space model, 2024. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 613 + ], + "type": "text", + "content": "[52] Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10819-10829, 2022. 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[53] Weihao Yu, Pan Zhou, Shuicheng Yan, and Xinchao Wang. Inceptionnext: When inception meets convnext. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5672-5683, 2024. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[54] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23528" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 228 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[55] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[56] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In *Forty-first International Conference on Machine Learning*. 1, 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[57] Qinfeng Zhu, Yuanzhi Cai, Yuan Fang, Yihan Yang, Cheng Chen, Lei Fan, and Anh Nguyen. Samba: Semantic segmentation of remotely sensed images with state space model, 2024. 3" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23529" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/8bac4e62-edf8-4833-a2b8-a238e4c4c68b_content_list.json b/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/8bac4e62-edf8-4833-a2b8-a238e4c4c68b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ac6c5c689ceaeb8016c8589c58ec6bb85b999b39 --- /dev/null +++ b/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/8bac4e62-edf8-4833-a2b8-a238e4c4c68b_content_list.json @@ -0,0 +1,1539 @@ +[ + { + "type": "text", + "text": "To Label or Not to Label: PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models", + "text_level": 1, + "bbox": [ + 194, + 130, + 803, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Julia Machnio \nPioneer Centre for AI \nUniversity of Copenhagen \njuma@di.ku.dk", + "bbox": [ + 140, + 203, + 352, + 275 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mads Nielsen \nPioneer Centre for AI \nUniversity of Copenhagen \nmadsn@di.ku.dk", + "bbox": [ + 393, + 203, + 604, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mostafa Mehdipour Ghazi \nPioneer Centre for AI \nUniversity of Copenhagen \nghazi@di.ku.dk", + "bbox": [ + 643, + 203, + 857, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Active learning (AL) seeks to reduce annotation costs by selecting the most informative samples for labeling, making it particularly valuable in resource-constrained settings. However, traditional evaluation methods, which focus solely on final accuracy, fail to capture the full dynamics of the learning process. To address this gap, we propose PALM (Performance Analysis of Active Learning Models), a unified and interpretable mathematical model that characterizes AL trajectories through four key parameters: achievable accuracy $(A_{\\mathrm{max}})$ , coverage efficiency $(\\delta)$ , early-stage performance $(\\alpha)$ , and scalability $(\\beta)$ . PALM provides a predictive description of AL behavior from partial observations, enabling the estimation of future performance and facilitating principled comparisons across different strategies. We validate PALM through extensive experiments on CIFAR-10/100 and ImageNet-50/100/200, covering a wide range of AL methods and self-supervised embeddings. Our results demonstrate that PALM generalizes effectively across datasets, budgets, and strategies, accurately predicting full learning curves from limited labeled data. Importantly, PALM reveals crucial insights into learning efficiency, data space coverage, and the scalability of AL methods. By enabling the selection of cost-effective strategies and predicting performance under tight budget constraints, PALM lays the basis for more systematic, reproducible, and data-efficient evaluation of AL in both research and real-world applications. The code is available at: https://github.com/juliamachnio/PALM.", + "bbox": [ + 89, + 343, + 483, + 765 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 797, + 220, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid growth of large-scale, domain-specific datasets in real-world applications, the cost and effort required for data labeling have become significant bottlenecks in machine learning pipelines. In many domains, obtaining labeled data necessitates substantial human expertise,", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf49bacbcccc097e2a3e9d321e30cd3052e580882a53c6e891e00bf97cc99e34.jpg", + "image_caption": [ + "Figure 1. Illustration of the PALM method for comparing AL strategies through interpretable parameters. Methods 1 and 2 differ in how labeled samples (green and orange) cover the data space, with $\\delta$ representing the average coverage per labeled sample. The corresponding PALM equations describe accuracy growth as a function of the cumulative annotation budget $(B)$ , with parameters controlling early-stage performance $(\\alpha)$ , scalability $(\\beta)$ , and maximum achievable accuracy $(\\mathrm{A}_{\\max})$ . The learning curves show how different parameter configurations affect AL dynamics: Method 1 achieves faster accuracy gains and higher asymptotic performance due to better coverage efficiency and scalability." + ], + "image_footnote": [], + "bbox": [ + 540, + 306, + 879, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "time, and resources, making it impractical to fully-associate large datasets. Active learning (AL) addresses this challenge by iteratively selecting the most informative samples for labeling, enabling models to achieve high performance with fewer labeled examples [32]. This makes AL particularly attractive in settings where labeled data is scarce or expensive, such as medical imaging [4, 35, 36], autonomous driving [12, 23], and anomaly detection in IoT systems [39].", + "bbox": [ + 511, + 718, + 906, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although AL provides a solution to reduce annotation costs, its effectiveness in deep learning largely depends on the sample selection strategy. Over the years, a variety of AL methods have been proposed [1, 34, 37, 40], typically", + "bbox": [ + 511, + 840, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4039", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "categorized into three main groups based on their selection criteria: uncertainty-based methods, diversity-based methods, and core-set approaches. Uncertainty-based methods prioritize samples where the model exhibits low confidence [11, 21, 28, 30]. Diversity-based methods aim to maximize data space coverage by selecting diverse samples [2, 16, 18, 19]. Core-set approaches focus on selecting representative subsets that best capture the global data distribution [22, 24, 31]. Recently, a fourth category, typicality-based methods, has emerged, particularly effective in scenarios with limited annotation budgets (the number of labeled samples). These methods select samples that are most representative of the underlying data distribution, aiming to enhance coverage and learning efficiency [14, 25, 38].", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Integrating self-supervised learning (SSL) into active learning pipelines has further enhanced typicality-based methods by providing richer feature representations that improve both sample selection and model initialization [14, 25, 38]. SSL techniques such as SimCLR [5], BYOL [13], or MoCov2+ and MoCov3 [6, 7] have become powerful tools for learning transferable representations from unlabeled data. The synergy between AL and SSL has proven especially beneficial under limited annotation budgets, where high-quality feature representations help identify the most informative and representative samples.", + "bbox": [ + 89, + 305, + 480, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advancements, the performance of AL methods remains highly sensitive to experimental settings, including model initialization, dataset properties, and task complexity [3]. Furthermore, current evaluation protocols typically focus on the final accuracy achieved after a fixed number of iterations, neglecting critical factors such as early learning efficiency, spatial coverage, and the scalability of learning gains over time. Recent work has highlighted the need for fair comparisons of AL methods in various scenarios but continues to rely on metrics such as accuracy and area under the ROC curve (AUC), which provide only a limited view of overall performance [26].", + "bbox": [ + 89, + 474, + 482, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These comparison limitations reveal a fundamental gap; there is no general predictive model for evaluating and comparing AL performance across diverse methods, datasets, and annotation budgets. In real-world scenarios, where annotation resources are finite, the objective is not only to achieve high accuracy but also to ensure robustness and generalization within limited budgets [32]. Moreover, deep learning models often retain predictive ability in uncovered regions of the data space, meaning that performance depends not only on labeled coverage but also on the model's inherent generalization capacity. Thus, evaluating active learning strategies requires considering both coverage expansion and the model's ability to learn from unlabeled regions. Without standardized, predictive evaluations incorporating these factors, selecting the most effective AL strategy remains a costly and inconsistent process.", + "bbox": [ + 89, + 659, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenge of fair comparison, we propose PALM, a mathematical model for the Performance Analysis of Active Learning Models, describing the dynamics of AL processes through a compact set of interpretable parameters as shown in Figure 1. PALM captures not only how accuracy grows with increasing labeled data but also how uncovered regions contribute to overall performance through generalization. It estimates the full learning curve from a limited number of iterations and predicts outcomes such as final accuracy, data coverage, early learning efficiency, and scalability. This formulation generalizes across different AL methods, datasets, and configurations and is compatible with or without SSL-based feature representations. By providing an interpretable and predictive description of AL behavior, PALM enables practitioners to estimate the number of labels required to achieve target performance and identify the most cost-effective strategies under resource constraints.", + "bbox": [ + 511, + 90, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows: (1) we propose PALM, a simple yet powerful mathematical model that describes the dynamics of AL and predicts key performance metrics from limited labeled data; (2) we empirically evaluate PALM's ability to compare different AL methods, datasets, and configurations without requiring full dataset annotation; (3) we demonstrate that PALM can estimate final accuracy, data space coverage, early learning efficiency, and scalability of learning gains across various AL settings; (4) we analyze the impact of incorporating SSL embeddings into AL pipelines, showing that PALM remains effective with or without SSL-based representations; and (5) we provide a practical approach for estimating the number of labeling samples needed to achieve target performance, facilitating real-world deployments under strict budget constraints.", + "bbox": [ + 511, + 348, + 903, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Methods", + "text_level": 1, + "bbox": [ + 513, + 588, + 611, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we present the underlying calculations of PALM and the properties that enable it to predict final accuracy and space coverage for active learning methods. Our formulation is grounded in fundamental probabilistic principles [29], considering the random distribution of objects in a space and their contribution to coverage estimation. This approach is inspired by the random covering problem [17], as well as concepts such as covering and effective sample size [8] and dynamic coverage and margin mix [25].", + "bbox": [ + 511, + 614, + 903, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Expected Coverage Fraction", + "text_level": 1, + "bbox": [ + 511, + 758, + 767, + 775 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 1 (Coverage Probability). Let $\\mathbb{X} \\subset \\mathbb{R}^d$ represent a $d$ -dimensional space, and let $\\{O_i\\}_{i=1}^s$ denote a collection of $s$ randomly placed objects within $\\mathbb{X}$ . Each object $O_i$ is a measurable subset of $\\mathbb{X}$ . For a given point $x \\in \\mathbb{X}$ , we define the coverage probability $P_C(x)$ as the probability that $x$ is covered by at least one object:", + "bbox": [ + 511, + 781, + 903, + 872 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nP _ {C} (x) = 1 - P _ {U C} (x), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 885, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4040", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $P_{UC}(x)$ represents the probability that $x$ is not covered by any object in $\\{O_i\\}_{i=1}^s$ .", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 2 (Coverage Probability with $s$ Independent Objects). Assume that each object $O_{i}$ is placed independently and covers a given point $x \\in \\mathbb{X}$ with the same probability $p$ . The probability that a single object does not cover $x$ is $1 - p$ . Since the objects are independent, the probability that no object covers $x$ is given by:", + "bbox": [ + 89, + 121, + 483, + 212 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {U C} (x) = (1 - p) ^ {s}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 218, + 482, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, the overall coverage probability of point $x$ is:", + "bbox": [ + 89, + 241, + 464, + 257 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP _ {C} (x) = 1 - (1 - p) ^ {s}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 263, + 482, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As $s \\to \\infty$ , $P_C(x)$ approaches 1, ensuring that $x$ is almost surely covered. When $s = 0$ , we have $P_C(x) = 0$ . Moreover, since $(1 - p)^s$ decreases with increasing $s$ for $0 < p < 1$ , the probability of non-coverage diminishes as the number of objects increases.", + "bbox": [ + 89, + 285, + 483, + 359 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Corollary 1 (Boundary Behavior of the Coverage Probability). The coverage probability $P_C(x)$ , as defined in Eq. (3), exhibits different boundary behaviors. As the number of objects grows infinitely, the probability that $x$ is covered approaches 1, ensuring full coverage of the space:", + "bbox": [ + 89, + 361, + 483, + 436 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {s \\rightarrow \\infty} P _ {C} (x) = 1 - \\lim _ {s \\rightarrow \\infty} (1 - p) ^ {s}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 443, + 482, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and since $0 < 1 - p < 1$ , we have:", + "bbox": [ + 89, + 469, + 323, + 484 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {s \\rightarrow \\infty} (1 - p) ^ {s} = 0. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 491, + 482, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, as the number of objects approaches infinity, the coverage probability becomes:", + "bbox": [ + 89, + 517, + 482, + 547 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {s \\rightarrow \\infty} P _ {C} (x) = 1. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 554, + 482, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conversely, as the number of objects approaches zero, the probability of coverage also approaches zero, indicating no coverage of the space:", + "bbox": [ + 89, + 580, + 483, + 626 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\lim _ {s \\rightarrow 0 ^ {+}} P _ {C} (x) = 1 - \\lim _ {s \\rightarrow 0 ^ {+}} (1 - p) ^ {s} = 0. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 632, + 482, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3 (Expected Coverage Fraction). Let $\\mathbb{X}$ be a space with finite volume $V$ , and let $\\delta = V_{O_i} / V$ denote the fraction of the total volume covered by a single randomly placed object $O_i$ , assuming all objects have the same volume and are placed independently and uniformly over $\\mathbb{X}$ . Under the assumption of independence, the expected fraction of $\\mathbb{X}$ covered by $s$ objects, denoted by $\\mathbb{E}_C$ , is given by:", + "bbox": [ + 89, + 659, + 483, + 766 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\mathrm {C}} = 1 - (1 - \\delta) ^ {s} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 772, + 482, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $0 \\leq \\delta \\leq 1$ . Note that $\\mathbb{E}_{\\mathbf{C}}$ describes the expected proportion of the entire space covered and is independent of any specific point in $\\mathbb{X}$ . This equation exhibits the same limiting behavior as outlined in the previous corollary. A larger $\\delta$ implies that fewer objects are needed to achieve significant coverage of $\\mathbb{X}$ since each object covers a greater proportion of the space.", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Active Learning Pools", + "text_level": 1, + "bbox": [ + 513, + 90, + 718, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To define the properties of AL methods, we adopt the notation introduced in [25], with a modification in the interpretation of $\\delta$ . Specifically, in this context, $\\delta$ represents the fraction of the space $\\mathbb{X}$ covered by a single labeled sample rather than the radius of a covering ball.", + "bbox": [ + 511, + 113, + 903, + 188 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define $\\mathbb{L} \\subseteq \\mathbb{X}$ as the labeled set of points and $\\mathbb{U} \\subseteq \\mathbb{X}$ as the unlabeled set, such that $\\mathbb{X} = \\mathbb{L} \\cup \\mathbb{U}$ , with the cardinality of the labeled set $|\\mathbb{L}| = B$ , representing the cumulative annotation budget collected iteratively from the AL pool. A labeling round refers to an iteration of the active learning process, in which a subset of points from $\\mathbb{U}$ is selected for labeling, and the labeled set $\\mathbb{L}$ is updated accordingly. Throughout the AL process, the cumulative budget $B \\leq |\\mathbb{X}|$ increases as more rounds are completed.", + "bbox": [ + 511, + 189, + 905, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 4 (Expected Coverage in Active Learning Pools). The expected fraction of $\\mathbb{X}$ covered by $B$ labeled samples, selected independently, is given by:", + "bbox": [ + 511, + 327, + 905, + 372 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {\\mathrm {C}} = 1 - (1 - \\delta) ^ {B}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 638, + 385, + 903, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where, $\\delta$ captures the average contribution of a single labeled sample to the overall coverage of $\\mathbb{X}$ , which inherently depends on the AL strategy used to select the samples. Specifically, more effective strategies prioritize samples that maximize coverage, leading to higher $\\delta$ values. Hence, comparing $\\delta$ across different strategies allows for assessing their relative effectiveness in covering the space.", + "bbox": [ + 511, + 416, + 905, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, the accuracy of a model trained on the labeled set $\\mathbb{L} \\subseteq \\mathbb{X}$ depends not only on the fraction of the space covered by labeled samples but also on the quality of the predictions within both the covered and uncovered regions [38]. Furthermore, the uncovered regions contribute to the model's generalization ability by influencing its behavior in areas with sparse or no labeled data.", + "bbox": [ + 511, + 523, + 905, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 5 (Accuracy as a Function of Coverage Probability). The overall test accuracy of a trained model on the labeled set can be expressed as:", + "bbox": [ + 511, + 630, + 905, + 675 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA = A _ {C} P _ {C} + A _ {U C} \\left(1 - P _ {C}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 690, + 903, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{A_C}$ is the model's accuracy in the covered regions of $\\mathbb{X}$ , $\\mathrm{A_{UC}}$ is its accuracy in the uncovered regions, and $P_{\\mathrm{C}}$ is the probability that a random test point lies within a covered region. By substituting $P_{\\mathrm{C}}$ from Eq. (9), we obtain:", + "bbox": [ + 511, + 719, + 905, + 781 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} = \\mathrm {A} _ {\\mathrm {C}} \\left(1 - (1 - \\delta) ^ {B}\\right) + \\mathrm {A} _ {\\mathrm {U C}} (1 - \\delta) ^ {B}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 792, + 903, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which establishes a simple yet powerful relationship between test accuracy and the proportion of covered space. When the labeled budget $B$ is small, the overall accuracy is dominated by $A_{\\mathrm{UC}}$ , as most of the space remains uncovered. As $B$ increases, coverage expands and accuracy improves,", + "bbox": [ + 511, + 824, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4041", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "eventually converging to $\\mathrm{A_C}$ as $B\\to \\infty$ . For small $\\delta$ values, the accuracy exhibits exponential growth behavior, approximately following $1 - e^{-B\\delta}$ , capturing the rate at which coverage and accuracy improve with increasing budget.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This formulation unifies the concepts of coverage and accuracy growth, showing how increasing labeled samples enhances accuracy while recognizing the importance of uncovered regions. In real-world scenarios, where models often retain some generalization capacity in uncovered areas, $\\mathrm{A_{UC}}$ becomes a crucial factor in performance evaluation. However, it does not fully describe how uncovered regions influence the rate of accuracy improvement during AL. To address this, we introduce additional parameters to better reflect the impact of uncovered areas on learning dynamics.", + "bbox": [ + 89, + 152, + 482, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 6 (Generalized Accuracy as a Function of Coverage with Exponential Adjustment). We define the generalized test accuracy function as:", + "bbox": [ + 89, + 304, + 483, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} = \\mathrm {A} _ {\\max } \\left(1 - (1 - \\delta) ^ {(B + \\alpha) ^ {\\beta}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 361, + 482, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{A}_{\\mathrm{max}}$ is the maximum achievable accuracy, $\\alpha > -B$ is a shifting parameter that adjusts the effective starting point of learning by capturing the contribution of uncovered regions, and $\\beta > 0$ is a scaling parameter that controls the rate at which accuracy increases with the number of labeled samples. These parameters introduce greater flexibility in modeling how space coverage shapes the learning outcome. An analysis of the asymptotic behavior of this function is provided in Supplementary Materials, showing how $\\alpha$ and $\\beta$ influence initial accuracy, growth rate, and the impact of uncovered regions on learning dynamics.", + "bbox": [ + 89, + 398, + 483, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The generalized accuracy function offers a robust model for approximating AL training curves across varying initial conditions, AL strategies, and datasets. Given accuracy observations at four or more distinct cumulative budgets $B$ , the parameters $A_{\\mathrm{max}}$ , $\\delta$ , $\\alpha$ , and $\\beta$ can be estimated by solving a system of equations using nonlinear regression methods. The computational complexity of this estimation is approximately $\\mathcal{O}(\\log (B))$ when employing optimized exponentiation techniques.", + "bbox": [ + 89, + 566, + 482, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More straightforward optimization approaches typically exhibit higher complexity, approaching $\\mathcal{O}(B^{\\beta})$ for large $B$ . Moreover, the rapid growth of the term $(B + \\alpha)^{\\beta}$ can lead to large numerical values, increasing the risk of floating-point precision errors. To mitigate these issues and enhance numerical stability, it is beneficial to normalize the cumulative budget $B$ by the mean annotation budget per iteration, denoted $b$ . This normalization reduces computational complexity to approximately $\\mathcal{O}((B / b)^{\\beta})$ and aligns the model with the iterative structure of AL, where labeled data is acquired in fixed increments of size $b$ per iteration.", + "bbox": [ + 89, + 704, + 482, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 7 (Normalized Accuracy Function for Active Learning Budget Scaling). The generalized accuracy func", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion with normalized budget $B$ is defined as:", + "bbox": [ + 511, + 90, + 808, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} = \\mathrm {A} _ {\\max } \\left(1 - (1 - \\delta) ^ {\\left(\\frac {B}{b} + \\alpha\\right) ^ {\\beta}}\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 116, + 903, + 142 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $b$ is the mean number of labeled samples per iteration. This normalization aligns the accuracy function with the iterative structure of AL and mitigates numerical instability associated with large budget values. Estimating the characteristic parameters $(\\delta, b, \\alpha, \\beta, \\mathrm{A}_{\\mathrm{max}})$ from observed learning curves enables direct and meaningful comparisons between different AL methods.", + "bbox": [ + 511, + 152, + 903, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 (Comparison of Active Learning Methods Using the Normalized Accuracy Function). Consider two active learning methods 1 and 2 described by their respective accuracy functions:", + "bbox": [ + 511, + 258, + 903, + 319 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} _ {1} = \\mathrm {A} _ {\\max , 1} \\left(1 - \\left(1 - \\delta_ {1}\\right) ^ {\\left(\\frac {B}{b _ {1}} + \\alpha_ {1}\\right) ^ {\\beta_ {1}}}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 329, + 903, + 362 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} _ {2} = \\mathrm {A} _ {\\max , 2} \\left(1 - \\left(1 - \\delta_ {2}\\right) ^ {\\left(\\frac {B}{b _ {2}} + \\alpha_ {2}\\right) ^ {\\beta_ {2}}}\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 364, + 903, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a given budget $B$ , Method 1 outperforms Method 2 if $\\mathrm{A}_1 > \\mathrm{A}_2$ . Importantly, $\\mathrm{A}_{\\max}$ determines the asymptotic performance as $B \\to \\infty$ , with the method having higher $\\mathrm{A}_{\\max}$ achieving superior long-term accuracy. The rate of accuracy improvement, influenced by $\\delta$ , $b$ , $\\alpha$ , and $\\beta$ , further captures the efficiency of each method in reaching its asymptotic potential. The full proof is presented in Supplementary Materials.", + "bbox": [ + 511, + 407, + 906, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lemma 1 (Comparison Criterion for Active Learning Methods). For any budget $B$ , two active learning methods, A and B, can be compared directly through:", + "bbox": [ + 511, + 529, + 903, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} _ {1} > \\mathrm {A} _ {2}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 673, + 585, + 903, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This provides a simple and effective criterion to identify the better-performing method for a given annotation budget, making it easier to evaluate performance across different stages of the AL process.", + "bbox": [ + 511, + 613, + 903, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Corollary 2 (Comparison of Learning Efficiency and Parameter Influence). Assuming both methods achieve the same maximum accuracy, their performance can be distinguished through the following considerations:", + "bbox": [ + 511, + 674, + 903, + 734 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Coverage Efficiency $(\\delta)$ : A higher $\\delta$ indicates that each labeled sample provides more effective coverage of the space, enhancing overall efficiency.", + "- Budget Efficiency $(b)$ : A smaller $b$ results in more frequent, smaller batches of labeled data, leading to faster and more incremental accuracy improvements.", + "- Initial Learning Efficiency $(\\alpha)$ : A lower $\\alpha$ boosts the starting accuracy, which is especially advantageous in low-budget scenarios.", + "- Scaling of Learning Gains $(\\beta)$ : A larger $\\beta$ increases the accuracy improvement rate as more samples are labeled." + ], + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4042", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hence, active learning methods can be quantitatively compared by evaluating these parameters and $\\mathrm{A}_{\\mathrm{max}}$ that defines the maximum achievable accuracy as the budget increases.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, this work presents a unified equation for modeling AL performance through a normalized accuracy function that captures the effects of coverage efficiency, budget scaling, and learning dynamics. By introducing coverage efficiency $(\\delta)$ , budget efficiency $(b)$ , initial accuracy shift $(\\alpha)$ , and learning gain scaling $(\\beta)$ , the model captures the early-stage and long-term behavior of AL methods. This formulation enables a direct, quantitative comparison of different AL strategies based on their estimated parameters, independent of the dataset or initial conditions.", + "bbox": [ + 88, + 136, + 480, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 300, + 282, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the proposed PALM model in fully supervised settings using ResNet-18 and ResNet-50 architectures [15] on the CIFAR-10 and CIFAR-100 datasets [20], as well as on subsets of ImageNet [9], including ImageNet-50, ImageNet-100, and ImageNet-200. In this context, fully supervised refers to training the models exclusively on the labeled subset $\\mathbb{L}$ of the dataset. At each iteration, $\\mathbb{L}$ is incrementally expanded with newly labeled samples in accordance with standard AL procedures. We adapted an existing AL framework to ensure reproducibility across various AL sampling strategies [14, 27]. In each experiment, we estimated the PALM parameters, which characterize the AL dynamics, based on the cumulative annotation budget $B$ at the corresponding iteration.", + "bbox": [ + 91, + 325, + 483, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Supervised Training on CIFARs", + "text_level": 1, + "bbox": [ + 89, + 545, + 372, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For CIFAR-10 and CIFAR-100, we trained a ResNet-18 model for at least 100 epochs. The training process employed stochastic gradient descent (SGD) with Nesterov momentum (0.9), weight decay (0.0003), and a cosine learning rate schedule, starting with an initial learning rate of 0.025. The models were trained using a batch size of 512. Standard data augmentation techniques, including random cropping and horizontal flipping, were applied. For further implementation details, please refer to [25, 27].", + "bbox": [ + 89, + 566, + 482, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Supervised Training on ImageNet Subsets", + "text_level": 1, + "bbox": [ + 89, + 712, + 447, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For ImageNet-50, ImageNet-100, and ImageNet-200, we adopted the same optimization settings and data augmentation protocol as used in the CIFAR experiments, but with a ResNet-50 backbone [15].", + "bbox": [ + 89, + 734, + 482, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Self-Supervised Feature Extraction", + "text_level": 1, + "bbox": [ + 89, + 803, + 397, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To obtain semantically meaningful feature representations, we trained SimCLR on each dataset using the official implementation from [5]. For CIFAR-10 and CIFAR-100, we employed a ResNet-18 encoder with a projection head mapping to a 128-dimensional vector, training for 500 epochs.", + "bbox": [ + 89, + 825, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training hyperparameters followed the settings outlined in SCAN [33]. After training, we extracted representations from the penultimate layer, resulting in 512-dimensional embeddings for CIFAR-10 and CIFAR-100 and 2048-dimensional embeddings for ImageNet.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the ImageNet subsets, we also extracted features using the official pretrained models of BYOL [13] (ResNet-50 trained for 300 epochs with a batch size of 512), MoCov2+ [6] (ResNet-50 trained for 800 epochs with a batch size of 256), and MoCov3 [7] (ResNet-50 trained for 1000 epochs with a batch size of 4096). We computed the mean and standard deviation of the features extracted from the training samples, applied normalization accordingly, and used the same values to normalize the features extracted from the validation and test samples.", + "bbox": [ + 511, + 167, + 906, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Active Learning Strategies", + "text_level": 1, + "bbox": [ + 511, + 325, + 754, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate PALM's ability to approximate active learning dynamics across diverse query strategies, we evaluated several commonly used AL methods: Random Sampling, TypiClust [14], Uncertainty Sampling, which selects samples with the lowest maximum softmax probability, Max Entropy, which selects samples with the highest predictive entropy, Min Margin, which selects samples with the smallest margin between the top two softmax probabilities, and DBAL [10]. For reproducibility and fair comparison, we followed the experimental settings outlined in [27].", + "bbox": [ + 511, + 348, + 906, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results and Discussion", + "text_level": 1, + "bbox": [ + 511, + 512, + 728, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AL Behavior: We first evaluate the ability of PALM to approximate AL dynamics across different sampling strategies on CIFAR-10 and CIFAR-100. For CIFAR-10, we used a mean budget size of $b = 20$ over 1000 iterations, while for CIFAR-100, the budget was $b = 100$ over 450 iterations. Each AL strategy was executed until the unlabeled pool was nearly exhausted. For each AL curve, we estimated the PALM parameters $(\\delta, \\alpha, \\beta, \\mathrm{A}_{\\mathrm{max}})$ using nonlinear least squares to fit the model to the observed empirical accuracy as a function of the cumulative annotation budget $B$ . To ensure stable and interpretable fits, we imposed constraints on the parameters. These constraints help reduce ambiguity in the parameter estimates. Figure 2 presents the AL curves together with their corresponding PALM fits. Across both datasets, two distinct behaviors emerge: (1) AL strategies that achieve gradual accuracy improvements over extended annotation budgets and (2) strategies leveraging pretrained embeddings, which exhibit rapid accuracy gains and early convergence toward $\\mathrm{A}_{\\mathrm{max}}$ . These differences in behavior are reflected in the fitted PALM parameters.", + "bbox": [ + 511, + 537, + 906, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The parameter $\\delta$ quantifies the coverage efficiency per labeled sample. Across both CIFAR-10 and CIFAR-100, methods that employ pretrained embeddings achieve significantly higher $\\delta$ values. For example, on CIFAR-10, Margin", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4043", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fb7d9db832b5a12b0aa53e49a2e7f4e54a710fc060f1f1296033aa2eee0fe385.jpg", + "image_caption": [ + "Figure 2. PALM approximation of AL curves on CIFAR-10 ( $b = 20$ ) and CIFAR-100 ( $b = 100$ ). Each plot displays the accuracy trajectories of multiple AL strategies fitted with the PALM model across the entire annotation budget. Two distinct behaviors emerge in both datasets: methods with gradual accuracy growth and those utilizing pretrained embeddings that quickly reach maximum accuracy. These differences are reflected in the fitted PALM parameters. Methods using pretrained embeddings consistently show higher $\\delta$ values, indicating greater sample efficiency, while non-embedding methods exhibit higher $\\alpha$ values, indicating delayed learning progress. Despite the varying complexity of the datasets, PALM successfully models the learning dynamics and highlights clear differences between AL strategies. Note that the highest test accuracy for CIFAR-100 is approximately $65\\%$ . All quantitative results are available in the Supplementary Materials." + ], + "image_footnote": [], + "bbox": [ + 176, + 94, + 480, + 255 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b63ce5efe75411643b4f96585d9e91ea9823cd9a5d2a8c33bb26b2e93d3beaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 95, + 805, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sampling without embeddings yields $\\delta = 0.094$ , whereas its embedding-based counterpart achieves $\\delta = 0.535$ . Similarly, on CIFAR-100, Random sampling improves from $\\delta = 0.048$ (without embeddings) to $\\delta = 0.318$ (with embeddings). These indicate that pretrained embeddings enhance the representational quality of selected samples, thereby improving space coverage and annotation efficiency.", + "bbox": [ + 88, + 393, + 482, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The parameter $\\alpha$ controls the initial offset of learning progress. Higher $\\alpha$ values indicate a delayed onset of learning, a trend that is especially pronounced in non-embedding methods. For example, on CIFAR-100, Margin sampling without embeddings exhibits $\\alpha = 10.643$ , whereas its embedding-based counterpart reduces $\\alpha$ to 0.068, indicating faster initial learning when using SSL-derived representations. Moreover, the parameter $\\beta$ captures the scaling of accuracy gains for the annotation budget. Larger $\\beta$ values correspond to steeper growth in accuracy. On CIFAR-100, Entropy sampling with embeddings demonstrates $\\beta = 0.651$ , the highest among the evaluated strategies, indicating rapid convergence.", + "bbox": [ + 88, + 503, + 482, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Despite similar final accuracies across methods, the PALM parameters offer a fine-grained view of sample efficiency and learning behavior. For example, while Random sampling with embeddings achieves a high $\\delta = 0.536$ , Entropy sampling without embeddings is limited to $\\delta = 0.066$ , despite achieving comparable asymptotic performance. This illustrates that PALM not only captures endpoint accuracy but also characterizes the trajectory of learning dynamics. These findings empirically validate Eq. (13), where $\\delta$ , $\\alpha$ , and $\\beta$ jointly describe the efficiency, onset, and scaling of active learning progress. The clear distinction in behaviors across AL strategies, as revealed by the fitted parameters, confirms that PALM provides a consistent and in", + "bbox": [ + 88, + 704, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "terpretable model of AL dynamics across different datasets, methods, and embedding strategies. Full quantitative results are provided in the Supplementary Materials.", + "bbox": [ + 511, + 393, + 903, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While PALM accurately fits the learning trajectories across AL strategies, we observe minor deviations in the later rounds of the CIFAR-100 experiments. This is likely due to the progressive exhaustion of informative examples in the unlabeled pool. As AL progresses, the remaining samples tend to be noisier or less representative, which increases prediction uncertainty and affects fit stability. Although PALM assumes an effectively infinite unlabeled pool, this simplification may lead to small estimation drift once the pool is nearly depleted. Nevertheless, the observed deviations remain within $2\\%$ of test accuracy across all budget levels. Importantly, PALM is designed to support early-stage decision-making, where accurate prediction from limited partial observations is most impactful.", + "bbox": [ + 511, + 441, + 905, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Limited Budgets: To further assess the predictive capability of PALM, we evaluated its performance when fitted to partial learning curves, using only a limited number of annotated samples across different AL methods and datasets. The results are presented in Figure 3. For CIFAR-10, an accurate approximation of the full learning dynamics was achieved with as few as 1,000 labeled samples, representing just $2\\%$ of the dataset. Notably, for some strategies, high-quality predictions emerged with only 200-300 labeled samples, corresponding to approximately $0.8 - 1.5\\%$ of the dataset. Methods leveraging pretrained embeddings required even fewer samples to achieve strong fits, further highlighting the efficiency gains provided by SSL-based representations. Additionally, for CIFAR-100, a more complex dataset with 100 classes, reliable approximations were obtained using 5,000-10,000 labeled samples. This", + "bbox": [ + 511, + 657, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4044", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5ca44b5c6d092850b4861acb11b47b121c94fd70a75afbb08a642b3429bc6954.jpg", + "image_caption": [ + "Figure 3. PALM predictions of AL training curves fitted using varying numbers of cumulative budget points, as indicated in the legend. Each plot compares the predicted trajectories with the actual performance across different AL strategies on CIFAR-10 and CIFAR-100. Note that the strategies in rows 2 and 4 utilize embeddings extracted using the SimCLR method. We find that accurate predictions of the full learning dynamics can be made with cumulative budgets of approximately 1,000 samples for CIFAR-10 (2% of the dataset) and 5,000 samples for CIFAR-100 (10% of the dataset). In certain cases on CIFAR-10, reasonable predictions can be made with as few as 400 samples (0.8% of the dataset), demonstrating PALM's ability to model learning behavior even with limited early-stage supervision." + ], + "image_footnote": [], + "bbox": [ + 163, + 85, + 834, + 334 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "translates to approximately 50-100 labeled examples per class, demonstrating that PALM can generalize effectively to higher-dimensional spaces with greater class diversity.", + "bbox": [ + 88, + 455, + 482, + 501 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TypiClust, a method specifically designed for small-budget scenarios and reliant on pretrained embeddings, demonstrated particularly strong approximation with extremely limited supervision. On CIFAR-100, PALM accurately modeled TypiClust's learning trajectory using just 1,000 labeled samples (roughly 10 examples per class), underscoring its effectiveness under resource-constrained conditions. The use of SSL-based embeddings significantly reduced the computational cost of AL iterations, offering practical advantages for large-scale or real-time annotation workflows. This finding suggests that embedding-based AL pipelines not only enhance model performance but also improve training efficiency, making them a compelling choice for the design of future AL systems.", + "bbox": [ + 88, + 503, + 482, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is worth noting that for CIFAR-10, PALM occasionally struggles to fit the sharp knee point of the learning curve due to the dataset's relatively simple structure and the rapid accuracy gains observed in early iterations. In contrast, this effect is less pronounced on CIFAR-100, where the increased complexity and class imbalance align more closely with real-world scenarios, enabling smoother approximations across the full budget range.", + "bbox": [ + 88, + 717, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Large-Scale Datasets: Furthermore, to assess the generalization capability of PALM on large-scale, natural datasets, we conducted experiments on ImageNet-50, ImageNet-100, and ImageNet-200 using a total annotation budget corre", + "bbox": [ + 88, + 839, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sponding to $4\\%$ of the training samples. Figure 4 presents the results of this experiment. As seen, PALM successfully approximated the learning dynamics across all tested active learning strategies and SSL-based embeddings, reinforcing its applicability to realistic, resource-constrained scenarios. Beyond capturing overall learning trajectories, PALM effectively models distinct learning behaviors, including methods that achieve rapid initial improvements and those with delayed performance onset. These differences are reflected in the fitted parameters, particularly $\\alpha$ and $\\delta$ , which control the starting point and coverage efficiency of the learning process.", + "bbox": [ + 511, + 455, + 903, + 637 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A particularly notable case arises with the BYOL embeddings. Due to BYOL's extremely slow and nearly linear accuracy progression within the small budget, PALM fits the curve as if the method has already plateaued. This behavior occurs because the early-stage dynamics are not sufficiently captured with the limited number of labeled samples. However, signs of acceleration appear in methods like TypiClust, where the final iterations show an upward trend, suggesting that with larger annotation budgets, BYOL would eventually follow the growth patterns predicted by our mathematical model. This observation underscores the need for more labeled samples to estimate future performance for slow-starting methods reliably.", + "bbox": [ + 511, + 640, + 905, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Differences in $\\beta$ across AL strategies also reveal important scalability trends, with higher $\\beta$ values indicating steeper accuracy gains following the initial labeling phase. Notably, strategies combined with MoCov3 consistently ex", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4045", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b625e2c3e82669574c7ed45ed7741f3273d7ef2316b5cc7947f2779f030dd61.jpg", + "image_caption": [ + "Figure 4. PALM predictions of AL curves on ImageNet subsets. Each plot compares the performance of different AL strategies combined with various SSL methods (MoCov2+, MoCov3, BYOL, and SimCLR). PALM accurately models learning dynamics under extreme annotation constraints, capturing distinct behaviors across methods and embeddings. MoCov3 consistently achieves the highest achievable accuracy $(\\mathrm{A}_{\\mathrm{max}})$ and coverage efficiency $(\\delta)$ , indicating superior representation quality and faster learning. In contrast, BYOL exhibits slow, nearly linear accuracy growth with minimal coverage, indicating delayed learning that requires larger annotation budgets. SimCLR and MoCov2+ show varying dynamics depending on the AL strategy, underscoring the sensitivity of embedding performance to sample selection methods. These findings highlight PALM's ability to model and compare AL behavior across diverse settings." + ], + "image_footnote": [], + "bbox": [ + 166, + 85, + 831, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "hibit higher $\\beta$ , reflecting their ability to capitalize on accumulated labeled data more effectively. A closer analysis of the ImageNet-200 results further illustrates these insights:", + "bbox": [ + 89, + 417, + 482, + 463 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- MoCov3 consistently achieves the highest $A_{\\max}$ and $\\delta$ values across all strategies, indicating superior representation quality and faster learning progression, making it highly effective under limited supervision.", + "- SimCLR shows competitive $\\mathrm{A}_{\\max}$ but lower $\\delta$ , suggesting slower per-sample gains and greater sensitivity to the choice of active learning strategy, as evidenced by the variability of $\\alpha$ across methods.", + "- BYOL displays consistently low $\\delta$ values and, in some cases, near-zero values (e.g., with TypiClust), reflecting minimal coverage per labeled sample and delayed learning. For TypiClust specifically, a high $\\alpha$ and steep $\\beta$ indicate a late but accelerating learning phase, aligning with the observed rise in the final iterations.", + "- MoCov2+ generally lags behind the other embeddings, with lower $A_{\\max}$ and $\\delta$ across strategies, further highlighting the varying effectiveness of SSL embeddings in active learning pipelines." + ], + "bbox": [ + 89, + 465, + 482, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The findings show that PALM not only captures overall performance trends but also uncovers nuanced interactions between SSL embeddings and AL strategies. It offers a comprehensive model for comparing methods, balancing trade-offs between early performance and long-term scalability, and guiding annotation planning in practical applications.", + "bbox": [ + 89, + 737, + 482, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 844, + 209, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented PALM, a unified mathematical model for the performance analysis of active learning models by captur", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing the dynamics of AL curves through a compact set of interpretable parameters. Unlike conventional evaluation protocols that compare methods based solely on accuracy at fixed budgets, PALM provides a principled approach to understanding how active learning strategies behave throughout the entire annotation process. By modeling coverage efficiency $(\\delta)$ , sampling behavior $(b)$ , early-stage performance $(\\alpha)$ , scalability $(\\beta)$ , and achievable accuracy $(A_{\\max})$ , PALM offers a comprehensive characterization of active learning trajectories from limited observations.", + "bbox": [ + 511, + 417, + 906, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our experiments on CIFAR and ImageNet benchmarks, covering a diverse range of AL strategies and SSL embeddings, demonstrate that PALM generalizes across datasets, methods, and budget regimes despite using only a small fraction of labeled data. It not only accurately predicts learning curves under extreme annotation constraints but also reveals key differences in how methods scale, where they accelerate, and how embedding quality interacts with sampling strategies. These insights expose the limitations of relying on final accuracy alone and establish PALM as a fair model for cost-aware comparison of AL methods.", + "bbox": [ + 511, + 574, + 906, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "PALM advances AL by shifting the focus from isolated performance measurements to predictive modeling of the learning process. This enables more informed, data-efficient decision-making in real-world settings, where annotation costs dominate and strategy selection is critical. By accounting for the dynamics and learning efficiency, PALM lays the basis for a new standard in AL evaluation, paving the way toward more systematic and scalable AL design.", + "bbox": [ + 511, + 744, + 908, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This project has received funding from the Pioneer Centre for AI, DNRF, grant number P1.", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4046", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amin Aghaee, Mehrdad Ghadiri, and Mahdieh Soleymani Baghshah. Active distance-based clustering using k-medoids. In Advances in Knowledge Discovery and Data Mining: 20th Pacific-Asia Conference, PAKDD 2016, Auckland, New Zealand, April 19-22, 2016, Proceedings, Part I 20, pages 253-264. Springer, 2016. 1, 11", + "[2] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019. 2, 11", + "[3] William H Beluch, Tim Genewein, Andreas Nurnberger, and Jan M Kohler. The power of ensembles for active learning in image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9368-9377, 2018. 2", + "[4] Jiayi Chen, Benteng Ma, Hengfei Cui, and Yong Xia. Think twice before selection: Federated evidential active learning for medical image analysis with domain shifts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11439-11449, 2024. 1", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 2, 5", + "[6] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2, 5", + "[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 2, 5", + "[8] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 2", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5", + "[10] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International conference on machine learning, pages 1183-1192. PMLR, 2017. 5, 11", + "[11] Daniel Gissin and Shai Shalev-Shwartz. Discriminative active learning. arXiv preprint arXiv:1907.06347, 2019. 2", + "[12] Ross Greer, Björk Antoniussen, Mathias V Andersen, Andreas Møgelmose, and Mohan M Trivedi. The why, when, and how to use active learning in large-data-driven 3d object detection for safe autonomous driving: An empirical exploration. arXiv preprint arXiv:2401.16634, 2024. 1", + "[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 5", + "[14] Guy Hacohen, Avihu Dekel, and Daphna Weinshall. Active learning on a budget: Opposite strategies suit high and low budgets. arXiv preprint arXiv:2202.02794, 2022. 2, 5, 11", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[16] Rong Hu, Brian Mac Namee, and Sarah Jane Delany. Off to a good start: Using clustering to select the initial training set in active learning. In FLAIRS, 2010. 2", + "[17] Svante Janson. Random coverings in several dimensions. 1986. 2", + "[18] Qiuye Jin, Mingzhi Yuan, Qin Qiao, and Zhijian Song. One-shot active learning for image segmentation via contrastive learning and diversity-based sampling. Knowledge-Based Systems, 241:108278, 2022. 2", + "[19] Seho Kee, Enrique Del Castillo, and George Runger. Query-by-committee improvement with diversity and density in batch active learning. Information Sciences, 454:401-418, 2018. 2", + "[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images, 2009. 5", + "[21] David D Lewis. A sequential algorithm for training text classifiers: Corrigendum and additional data. In Acm Sigir Forum, pages 13-19. ACM New York, NY, USA, 1995. 2, 11", + "[22] Hailun Liang, Sunyuan Qiang, Hui Ma, Jun Wan, and Yanyan Liang. Semantic segmentation active learning with scene coverage coreset. In Chinese Conference on Biometric Recognition, pages 238-247. Springer, 2024. 2", + "[23] Jinpeng Lin, Zhihao Liang, Shengheng Deng, Lile Cai, Tao Jiang, Tianrui Li, Kui Jia, and Xun Xu. Exploring diversity-based active learning for 3d object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2024. 1", + "[24] Alaa Maalouf, Gilad Eini, Ben Mussay, Dan Feldman, and Margarita Osadchy. A unified approach to coreset learning. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2", + "[25] Inbal Mishal and Daphna Weinshall. Dcom: Active learning for all learners. arXiv preprint arXiv:2407.01804, 2024. 2, 3, 5", + "[26] Sudhanshu Mittal, Joshua Niemeijer, Özgün Çiçek, Maxim Tatarchenko, Jan Ehrhardt, Jörg P Schäfer, Heinz Handels, and Thomas Brox. Realistic evaluation of deep active learning for image classification and semantic segmentation. International Journal of Computer Vision, pages 1-23, 2025. 2", + "[27] Prateek Munjal, Nasir Hayat, Munawar Hayat, Jamshid Sourati, and Shadab Khan. Towards robust and reproducible active learning using neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 223-232, 2022. 5", + "[28] Vu-Linh Nguyen, Mohammad Hossein Shaker, and Eyke Hüllermeier. How to measure uncertainty in uncertainty sampling for active learning. Machine Learning, 111(1):89-122, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4047", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Sheldon M Ross. Introduction to probability models. Academic press, 2014. 2", + "[30] Tobias Scheffer, Christian Decomain, and Stefan Wrobel. Active hidden markov models for information extraction. In International symposium on intelligent data analysis, pages 309-318. Springer, 2001. 2, 11", + "[31] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 2, 11", + "[32] Burr Settles. Active learning literature survey. 2009. 1, 2, 11", + "[33] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. 5", + "[34] Konstantin Voevodski, Maria-Florina Balcan, Heiko Röglin, Shang-Hua Teng, and Yu Xia. Active clustering of biological sequences. The Journal of Machine Learning Research, 13 (1):203–225, 2012. 1, 11", + "[35] Haoran Wang, Qiuye Jin, Shiman Li, Siyu Liu, Manning Wang, and Zhijian Song. A comprehensive survey on deep active learning in medical image analysis. Medical Image Analysis, page 103201, 2024. 1", + "[36] Ji Wu, Zhongfeng Kang, Sebastian Nørgaard Llambias, Mostafa Mehdipour Ghazi, and Mads Nielsen. Active transfer learning for 3d hippocampus segmentation. In Workshop on Medical Image Learning with Limited and Noisy Data, pages 224-234. Springer, 2023. 1", + "[37] Zhao Xu, Kai Yu, Volker Tresp, Xiaowei Xu, and Jizhi Wang. Representative sampling for text classification using support vector machines. In Advances in Information Retrieval: 25th European Conference on IR Research, ECIR 2003, Pisa, Italy, April 14-16, 2003. Proceedings 25, pages 393-407. Springer, 2003. 1, 11", + "[38] Ofer Yehuda, Avihu Dekel, Guy Hacohen, and Daphna Weinshall. Active learning through a covering lens. Advances in Neural Information Processing Systems, 35: 22354-22367, 2022. 2, 3", + "[39] Mohammed Zakariah and Abdulaziz S Almazyad. Anomaly detection for IoT systems using active learning. Applied Sciences, 13(21):12029, 2023. 1", + "[40] Jifan Zhang, Yifang Chen, Gregory Canal, Stephen Mussmann, Arnav M Das, Gantavya Bhatt, Yinglun Zhu, Jeffrey Bilmes, Simon Shaolei Du, Kevin Jamieson, et al. Labelbench: A comprehensive framework for benchmarking adaptive label-efficient learning. arXiv preprint arXiv:2306.09910, 2023. 1, 11" + ], + "bbox": [ + 89, + 92, + 482, + 753 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4048", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/8bac4e62-edf8-4833-a2b8-a238e4c4c68b_model.json b/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/8bac4e62-edf8-4833-a2b8-a238e4c4c68b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f493ec67b68370ff0986cba0c6d7fd5b4acfdf99 --- /dev/null +++ b/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/8bac4e62-edf8-4833-a2b8-a238e4c4c68b_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.131, + 0.804, + 0.177 + ], + "angle": 0, + "content": "To Label or Not to Label: PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.204, + 0.353, + 0.276 + ], + "angle": 0, + "content": "Julia Machnio \nPioneer Centre for AI \nUniversity of Copenhagen \njuma@di.ku.dk" + }, + { + "type": "text", + "bbox": [ + 0.394, + 0.204, + 0.605, + 0.275 + ], + "angle": 0, + "content": "Mads Nielsen \nPioneer Centre for AI \nUniversity of Copenhagen \nmadsn@di.ku.dk" + }, + { + "type": "text", + "bbox": [ + 0.645, + 0.204, + 0.858, + 0.275 + ], + "angle": 0, + "content": "Mostafa Mehdipour Ghazi \nPioneer Centre for AI \nUniversity of Copenhagen \nghazi@di.ku.dk" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.344, + 0.485, + 0.766 + ], + "angle": 0, + "content": "Active learning (AL) seeks to reduce annotation costs by selecting the most informative samples for labeling, making it particularly valuable in resource-constrained settings. However, traditional evaluation methods, which focus solely on final accuracy, fail to capture the full dynamics of the learning process. To address this gap, we propose PALM (Performance Analysis of Active Learning Models), a unified and interpretable mathematical model that characterizes AL trajectories through four key parameters: achievable accuracy \\((A_{\\mathrm{max}})\\), coverage efficiency \\((\\delta)\\), early-stage performance \\((\\alpha)\\), and scalability \\((\\beta)\\). PALM provides a predictive description of AL behavior from partial observations, enabling the estimation of future performance and facilitating principled comparisons across different strategies. We validate PALM through extensive experiments on CIFAR-10/100 and ImageNet-50/100/200, covering a wide range of AL methods and self-supervised embeddings. Our results demonstrate that PALM generalizes effectively across datasets, budgets, and strategies, accurately predicting full learning curves from limited labeled data. Importantly, PALM reveals crucial insights into learning efficiency, data space coverage, and the scalability of AL methods. By enabling the selection of cost-effective strategies and predicting performance under tight budget constraints, PALM lays the basis for more systematic, reproducible, and data-efficient evaluation of AL in both research and real-world applications. The code is available at: https://github.com/juliamachnio/PALM." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.799, + 0.222, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "With the rapid growth of large-scale, domain-specific datasets in real-world applications, the cost and effort required for data labeling have become significant bottlenecks in machine learning pipelines. In many domains, obtaining labeled data necessitates substantial human expertise," + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.308, + 0.88, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.538, + 0.908, + 0.692 + ], + "angle": 0, + "content": "Figure 1. Illustration of the PALM method for comparing AL strategies through interpretable parameters. Methods 1 and 2 differ in how labeled samples (green and orange) cover the data space, with \\(\\delta\\) representing the average coverage per labeled sample. The corresponding PALM equations describe accuracy growth as a function of the cumulative annotation budget \\((B)\\), with parameters controlling early-stage performance \\((\\alpha)\\), scalability \\((\\beta)\\), and maximum achievable accuracy \\((\\mathrm{A}_{\\max})\\). The learning curves show how different parameter configurations affect AL dynamics: Method 1 achieves faster accuracy gains and higher asymptotic performance due to better coverage efficiency and scalability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.907, + 0.84 + ], + "angle": 0, + "content": "time, and resources, making it impractical to fully-associate large datasets. Active learning (AL) addresses this challenge by iteratively selecting the most informative samples for labeling, enabling models to achieve high performance with fewer labeled examples [32]. This makes AL particularly attractive in settings where labeled data is scarce or expensive, such as medical imaging [4, 35, 36], autonomous driving [12, 23], and anomaly detection in IoT systems [39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Although AL provides a solution to reduce annotation costs, its effectiveness in deep learning largely depends on the sample selection strategy. Over the years, a variety of AL methods have been proposed [1, 34, 37, 40], typically" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4039" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "categorized into three main groups based on their selection criteria: uncertainty-based methods, diversity-based methods, and core-set approaches. Uncertainty-based methods prioritize samples where the model exhibits low confidence [11, 21, 28, 30]. Diversity-based methods aim to maximize data space coverage by selecting diverse samples [2, 16, 18, 19]. Core-set approaches focus on selecting representative subsets that best capture the global data distribution [22, 24, 31]. Recently, a fourth category, typicality-based methods, has emerged, particularly effective in scenarios with limited annotation budgets (the number of labeled samples). These methods select samples that are most representative of the underlying data distribution, aiming to enhance coverage and learning efficiency [14, 25, 38]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.482, + 0.473 + ], + "angle": 0, + "content": "Integrating self-supervised learning (SSL) into active learning pipelines has further enhanced typicality-based methods by providing richer feature representations that improve both sample selection and model initialization [14, 25, 38]. SSL techniques such as SimCLR [5], BYOL [13], or MoCov2+ and MoCov3 [6, 7] have become powerful tools for learning transferable representations from unlabeled data. The synergy between AL and SSL has proven especially beneficial under limited annotation budgets, where high-quality feature representations help identify the most informative and representative samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.656 + ], + "angle": 0, + "content": "Despite these advancements, the performance of AL methods remains highly sensitive to experimental settings, including model initialization, dataset properties, and task complexity [3]. Furthermore, current evaluation protocols typically focus on the final accuracy achieved after a fixed number of iterations, neglecting critical factors such as early learning efficiency, spatial coverage, and the scalability of learning gains over time. Recent work has highlighted the need for fair comparisons of AL methods in various scenarios but continues to rely on metrics such as accuracy and area under the ROC curve (AUC), which provide only a limited view of overall performance [26]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.902 + ], + "angle": 0, + "content": "These comparison limitations reveal a fundamental gap; there is no general predictive model for evaluating and comparing AL performance across diverse methods, datasets, and annotation budgets. In real-world scenarios, where annotation resources are finite, the objective is not only to achieve high accuracy but also to ensure robustness and generalization within limited budgets [32]. Moreover, deep learning models often retain predictive ability in uncovered regions of the data space, meaning that performance depends not only on labeled coverage but also on the model's inherent generalization capacity. Thus, evaluating active learning strategies requires considering both coverage expansion and the model's ability to learn from unlabeled regions. Without standardized, predictive evaluations incorporating these factors, selecting the most effective AL strategy remains a costly and inconsistent process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.348 + ], + "angle": 0, + "content": "To address the challenge of fair comparison, we propose PALM, a mathematical model for the Performance Analysis of Active Learning Models, describing the dynamics of AL processes through a compact set of interpretable parameters as shown in Figure 1. PALM captures not only how accuracy grows with increasing labeled data but also how uncovered regions contribute to overall performance through generalization. It estimates the full learning curve from a limited number of iterations and predicts outcomes such as final accuracy, data coverage, early learning efficiency, and scalability. This formulation generalizes across different AL methods, datasets, and configurations and is compatible with or without SSL-based feature representations. By providing an interpretable and predictive description of AL behavior, PALM enables practitioners to estimate the number of labels required to achieve target performance and identify the most cost-effective strategies under resource constraints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.905, + 0.575 + ], + "angle": 0, + "content": "Our main contributions are as follows: (1) we propose PALM, a simple yet powerful mathematical model that describes the dynamics of AL and predicts key performance metrics from limited labeled data; (2) we empirically evaluate PALM's ability to compare different AL methods, datasets, and configurations without requiring full dataset annotation; (3) we demonstrate that PALM can estimate final accuracy, data space coverage, early learning efficiency, and scalability of learning gains across various AL settings; (4) we analyze the impact of incorporating SSL embeddings into AL pipelines, showing that PALM remains effective with or without SSL-based representations; and (5) we provide a practical approach for estimating the number of labeling samples needed to achieve target performance, facilitating real-world deployments under strict budget constraints." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.589, + 0.612, + 0.605 + ], + "angle": 0, + "content": "2. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.905, + 0.751 + ], + "angle": 0, + "content": "In this section, we present the underlying calculations of PALM and the properties that enable it to predict final accuracy and space coverage for active learning methods. Our formulation is grounded in fundamental probabilistic principles [29], considering the random distribution of objects in a space and their contribution to coverage estimation. This approach is inspired by the random covering problem [17], as well as concepts such as covering and effective sample size [8] and dynamic coverage and margin mix [25]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.76, + 0.768, + 0.776 + ], + "angle": 0, + "content": "2.1. Expected Coverage Fraction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.782, + 0.905, + 0.873 + ], + "angle": 0, + "content": "Definition 1 (Coverage Probability). Let \\(\\mathbb{X} \\subset \\mathbb{R}^d\\) represent a \\(d\\)-dimensional space, and let \\(\\{O_i\\}_{i=1}^s\\) denote a collection of \\(s\\) randomly placed objects within \\(\\mathbb{X}\\). Each object \\(O_i\\) is a measurable subset of \\(\\mathbb{X}\\). For a given point \\(x \\in \\mathbb{X}\\), we define the coverage probability \\(P_C(x)\\) as the probability that \\(x\\) is covered by at least one object:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\nP _ {C} (x) = 1 - P _ {U C} (x), \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4040" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "where \\( P_{UC}(x) \\) represents the probability that \\( x \\) is not covered by any object in \\( \\{O_i\\}_{i=1}^s \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.213 + ], + "angle": 0, + "content": "Definition 2 (Coverage Probability with \\( s \\) Independent Objects). Assume that each object \\( O_{i} \\) is placed independently and covers a given point \\( x \\in \\mathbb{X} \\) with the same probability \\( p \\). The probability that a single object does not cover \\( x \\) is \\( 1 - p \\). Since the objects are independent, the probability that no object covers \\( x \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.219, + 0.483, + 0.236 + ], + "angle": 0, + "content": "\\[\nP _ {U C} (x) = (1 - p) ^ {s}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.242, + 0.465, + 0.258 + ], + "angle": 0, + "content": "Therefore, the overall coverage probability of point \\( x \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.264, + 0.483, + 0.281 + ], + "angle": 0, + "content": "\\[\nP _ {C} (x) = 1 - (1 - p) ^ {s}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.484, + 0.361 + ], + "angle": 0, + "content": "As \\( s \\to \\infty \\), \\( P_C(x) \\) approaches 1, ensuring that \\( x \\) is almost surely covered. When \\( s = 0 \\), we have \\( P_C(x) = 0 \\). Moreover, since \\( (1 - p)^s \\) decreases with increasing \\( s \\) for \\( 0 < p < 1 \\), the probability of non-coverage diminishes as the number of objects increases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.484, + 0.437 + ], + "angle": 0, + "content": "Corollary 1 (Boundary Behavior of the Coverage Probability). The coverage probability \\( P_C(x) \\), as defined in Eq. (3), exhibits different boundary behaviors. As the number of objects grows infinitely, the probability that \\( x \\) is covered approaches 1, ensuring full coverage of the space:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.444, + 0.483, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\lim _ {s \\rightarrow \\infty} P _ {C} (x) = 1 - \\lim _ {s \\rightarrow \\infty} (1 - p) ^ {s}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.47, + 0.325, + 0.485 + ], + "angle": 0, + "content": "and since \\(0 < 1 - p < 1\\), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.224, + 0.492, + 0.483, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\lim _ {s \\rightarrow \\infty} (1 - p) ^ {s} = 0. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.549 + ], + "angle": 0, + "content": "Thus, as the number of objects approaches infinity, the coverage probability becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.229, + 0.555, + 0.483, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\lim _ {s \\rightarrow \\infty} P _ {C} (x) = 1. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.484, + 0.627 + ], + "angle": 0, + "content": "Conversely, as the number of objects approaches zero, the probability of coverage also approaches zero, indicating no coverage of the space:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.633, + 0.483, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\lim _ {s \\rightarrow 0 ^ {+}} P _ {C} (x) = 1 - \\lim _ {s \\rightarrow 0 ^ {+}} (1 - p) ^ {s} = 0. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.767 + ], + "angle": 0, + "content": "Definition 3 (Expected Coverage Fraction). Let \\(\\mathbb{X}\\) be a space with finite volume \\(V\\), and let \\(\\delta = V_{O_i} / V\\) denote the fraction of the total volume covered by a single randomly placed object \\(O_i\\), assuming all objects have the same volume and are placed independently and uniformly over \\(\\mathbb{X}\\). Under the assumption of independence, the expected fraction of \\(\\mathbb{X}\\) covered by \\(s\\) objects, denoted by \\(\\mathbb{E}_C\\), is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.773, + 0.483, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\mathrm {C}} = 1 - (1 - \\delta) ^ {s} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(0 \\leq \\delta \\leq 1\\). Note that \\(\\mathbb{E}_{\\mathbf{C}}\\) describes the expected proportion of the entire space covered and is independent of any specific point in \\(\\mathbb{X}\\). This equation exhibits the same limiting behavior as outlined in the previous corollary. A larger \\(\\delta\\) implies that fewer objects are needed to achieve significant coverage of \\(\\mathbb{X}\\) since each object covers a greater proportion of the space." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.72, + 0.107 + ], + "angle": 0, + "content": "2.2. Active Learning Pools" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.114, + 0.905, + 0.189 + ], + "angle": 0, + "content": "To define the properties of AL methods, we adopt the notation introduced in [25], with a modification in the interpretation of \\(\\delta\\). Specifically, in this context, \\(\\delta\\) represents the fraction of the space \\(\\mathbb{X}\\) covered by a single labeled sample rather than the radius of a covering ball." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.19, + 0.906, + 0.326 + ], + "angle": 0, + "content": "We define \\(\\mathbb{L} \\subseteq \\mathbb{X}\\) as the labeled set of points and \\(\\mathbb{U} \\subseteq \\mathbb{X}\\) as the unlabeled set, such that \\(\\mathbb{X} = \\mathbb{L} \\cup \\mathbb{U}\\), with the cardinality of the labeled set \\(|\\mathbb{L}| = B\\), representing the cumulative annotation budget collected iteratively from the AL pool. A labeling round refers to an iteration of the active learning process, in which a subset of points from \\(\\mathbb{U}\\) is selected for labeling, and the labeled set \\(\\mathbb{L}\\) is updated accordingly. Throughout the AL process, the cumulative budget \\(B \\leq |\\mathbb{X}|\\) increases as more rounds are completed." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.328, + 0.906, + 0.373 + ], + "angle": 0, + "content": "Definition 4 (Expected Coverage in Active Learning Pools). The expected fraction of \\(\\mathbb{X}\\) covered by \\(B\\) labeled samples, selected independently, is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.64, + 0.386, + 0.905, + 0.405 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {\\mathrm {C}} = 1 - (1 - \\delta) ^ {B}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.906, + 0.522 + ], + "angle": 0, + "content": "where, \\(\\delta\\) captures the average contribution of a single labeled sample to the overall coverage of \\(\\mathbb{X}\\), which inherently depends on the AL strategy used to select the samples. Specifically, more effective strategies prioritize samples that maximize coverage, leading to higher \\(\\delta\\) values. Hence, comparing \\(\\delta\\) across different strategies allows for assessing their relative effectiveness in covering the space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.629 + ], + "angle": 0, + "content": "Additionally, the accuracy of a model trained on the labeled set \\(\\mathbb{L} \\subseteq \\mathbb{X}\\) depends not only on the fraction of the space covered by labeled samples but also on the quality of the predictions within both the covered and uncovered regions [38]. Furthermore, the uncovered regions contribute to the model's generalization ability by influencing its behavior in areas with sparse or no labeled data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.631, + 0.906, + 0.676 + ], + "angle": 0, + "content": "Definition 5 (Accuracy as a Function of Coverage Probability). The overall test accuracy of a trained model on the labeled set can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.691, + 0.905, + 0.707 + ], + "angle": 0, + "content": "\\[\nA = A _ {C} P _ {C} + A _ {U C} \\left(1 - P _ {C}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.906, + 0.782 + ], + "angle": 0, + "content": "where \\( \\mathrm{A_C} \\) is the model's accuracy in the covered regions of \\( \\mathbb{X} \\), \\( \\mathrm{A_{UC}} \\) is its accuracy in the uncovered regions, and \\( P_{\\mathrm{C}} \\) is the probability that a random test point lies within a covered region. By substituting \\( P_{\\mathrm{C}} \\) from Eq. (9), we obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.794, + 0.905, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} = \\mathrm {A} _ {\\mathrm {C}} \\left(1 - (1 - \\delta) ^ {B}\\right) + \\mathrm {A} _ {\\mathrm {U C}} (1 - \\delta) ^ {B}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.902 + ], + "angle": 0, + "content": "which establishes a simple yet powerful relationship between test accuracy and the proportion of covered space. When the labeled budget \\( B \\) is small, the overall accuracy is dominated by \\( A_{\\mathrm{UC}} \\), as most of the space remains uncovered. As \\( B \\) increases, coverage expands and accuracy improves," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4041" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "eventually converging to \\(\\mathrm{A_C}\\) as \\(B\\to \\infty\\) . For small \\(\\delta\\) values, the accuracy exhibits exponential growth behavior, approximately following \\(1 - e^{-B\\delta}\\) , capturing the rate at which coverage and accuracy improve with increasing budget." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.304 + ], + "angle": 0, + "content": "This formulation unifies the concepts of coverage and accuracy growth, showing how increasing labeled samples enhances accuracy while recognizing the importance of uncovered regions. In real-world scenarios, where models often retain some generalization capacity in uncovered areas, \\(\\mathrm{A_{UC}}\\) becomes a crucial factor in performance evaluation. However, it does not fully describe how uncovered regions influence the rate of accuracy improvement during AL. To address this, we introduce additional parameters to better reflect the impact of uncovered areas on learning dynamics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.305, + 0.484, + 0.351 + ], + "angle": 0, + "content": "Definition 6 (Generalized Accuracy as a Function of Coverage with Exponential Adjustment). We define the generalized test accuracy function as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.362, + 0.483, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} = \\mathrm {A} _ {\\max } \\left(1 - (1 - \\delta) ^ {(B + \\alpha) ^ {\\beta}}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.567 + ], + "angle": 0, + "content": "where \\( \\mathrm{A}_{\\mathrm{max}} \\) is the maximum achievable accuracy, \\( \\alpha > -B \\) is a shifting parameter that adjusts the effective starting point of learning by capturing the contribution of uncovered regions, and \\( \\beta > 0 \\) is a scaling parameter that controls the rate at which accuracy increases with the number of labeled samples. These parameters introduce greater flexibility in modeling how space coverage shapes the learning outcome. An analysis of the asymptotic behavior of this function is provided in Supplementary Materials, showing how \\( \\alpha \\) and \\( \\beta \\) influence initial accuracy, growth rate, and the impact of uncovered regions on learning dynamics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.703 + ], + "angle": 0, + "content": "The generalized accuracy function offers a robust model for approximating AL training curves across varying initial conditions, AL strategies, and datasets. Given accuracy observations at four or more distinct cumulative budgets \\( B \\), the parameters \\( A_{\\mathrm{max}} \\), \\( \\delta \\), \\( \\alpha \\), and \\( \\beta \\) can be estimated by solving a system of equations using nonlinear regression methods. The computational complexity of this estimation is approximately \\( \\mathcal{O}(\\log (B)) \\) when employing optimized exponentiation techniques." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.869 + ], + "angle": 0, + "content": "More straightforward optimization approaches typically exhibit higher complexity, approaching \\(\\mathcal{O}(B^{\\beta})\\) for large \\(B\\). Moreover, the rapid growth of the term \\((B + \\alpha)^{\\beta}\\) can lead to large numerical values, increasing the risk of floating-point precision errors. To mitigate these issues and enhance numerical stability, it is beneficial to normalize the cumulative budget \\(B\\) by the mean annotation budget per iteration, denoted \\(b\\). This normalization reduces computational complexity to approximately \\(\\mathcal{O}((B / b)^{\\beta})\\) and aligns the model with the iterative structure of AL, where labeled data is acquired in fixed increments of size \\(b\\) per iteration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Definition 7 (Normalized Accuracy Function for Active Learning Budget Scaling). The generalized accuracy func" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.81, + 0.107 + ], + "angle": 0, + "content": "tion with normalized budget \\(B\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.117, + 0.905, + 0.143 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} = \\mathrm {A} _ {\\max } \\left(1 - (1 - \\delta) ^ {\\left(\\frac {B}{b} + \\alpha\\right) ^ {\\beta}}\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.154, + 0.905, + 0.258 + ], + "angle": 0, + "content": "where \\(b\\) is the mean number of labeled samples per iteration. This normalization aligns the accuracy function with the iterative structure of AL and mitigates numerical instability associated with large budget values. Estimating the characteristic parameters \\((\\delta, b, \\alpha, \\beta, \\mathrm{A}_{\\mathrm{max}})\\) from observed learning curves enables direct and meaningful comparisons between different AL methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.259, + 0.905, + 0.32 + ], + "angle": 0, + "content": "Theorem 1 (Comparison of Active Learning Methods Using the Normalized Accuracy Function). Consider two active learning methods 1 and 2 described by their respective accuracy functions:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.33, + 0.905, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} _ {1} = \\mathrm {A} _ {\\max , 1} \\left(1 - \\left(1 - \\delta_ {1}\\right) ^ {\\left(\\frac {B}{b _ {1}} + \\alpha_ {1}\\right) ^ {\\beta_ {1}}}\\right), \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.366, + 0.905, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} _ {2} = \\mathrm {A} _ {\\max , 2} \\left(1 - \\left(1 - \\delta_ {2}\\right) ^ {\\left(\\frac {B}{b _ {2}} + \\alpha_ {2}\\right) ^ {\\beta_ {2}}}\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.409, + 0.907, + 0.528 + ], + "angle": 0, + "content": "For a given budget \\( B \\), Method 1 outperforms Method 2 if \\( \\mathrm{A}_1 > \\mathrm{A}_2 \\). Importantly, \\( \\mathrm{A}_{\\max} \\) determines the asymptotic performance as \\( B \\to \\infty \\), with the method having higher \\( \\mathrm{A}_{\\max} \\) achieving superior long-term accuracy. The rate of accuracy improvement, influenced by \\( \\delta \\), \\( b \\), \\( \\alpha \\), and \\( \\beta \\), further captures the efficiency of each method in reaching its asymptotic potential. The full proof is presented in Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.53, + 0.905, + 0.575 + ], + "angle": 0, + "content": "Lemma 1 (Comparison Criterion for Active Learning Methods). For any budget \\( B \\), two active learning methods, A and B, can be compared directly through:" + }, + { + "type": "equation", + "bbox": [ + 0.674, + 0.587, + 0.905, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} _ {1} > \\mathrm {A} _ {2}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.674 + ], + "angle": 0, + "content": "This provides a simple and effective criterion to identify the better-performing method for a given annotation budget, making it easier to evaluate performance across different stages of the AL process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.735 + ], + "angle": 0, + "content": "Corollary 2 (Comparison of Learning Efficiency and Parameter Influence). Assuming both methods achieve the same maximum accuracy, their performance can be distinguished through the following considerations:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.905, + 0.78 + ], + "angle": 0, + "content": "- Coverage Efficiency \\((\\delta)\\): A higher \\(\\delta\\) indicates that each labeled sample provides more effective coverage of the space, enhancing overall efficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.905, + 0.825 + ], + "angle": 0, + "content": "- Budget Efficiency \\((b)\\): A smaller \\(b\\) results in more frequent, smaller batches of labeled data, leading to faster and more incremental accuracy improvements." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.87 + ], + "angle": 0, + "content": "- Initial Learning Efficiency \\((\\alpha)\\): A lower \\(\\alpha\\) boosts the starting accuracy, which is especially advantageous in low-budget scenarios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Scaling of Learning Gains \\((\\beta)\\): A larger \\(\\beta\\) increases the accuracy improvement rate as more samples are labeled." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4042" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "Hence, active learning methods can be quantitatively compared by evaluating these parameters and \\( \\mathrm{A}_{\\mathrm{max}} \\) that defines the maximum achievable accuracy as the budget increases." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.137, + 0.482, + 0.288 + ], + "angle": 0, + "content": "Overall, this work presents a unified equation for modeling AL performance through a normalized accuracy function that captures the effects of coverage efficiency, budget scaling, and learning dynamics. By introducing coverage efficiency \\((\\delta)\\), budget efficiency \\((b)\\), initial accuracy shift \\((\\alpha)\\), and learning gain scaling \\((\\beta)\\), the model captures the early-stage and long-term behavior of AL methods. This formulation enables a direct, quantitative comparison of different AL strategies based on their estimated parameters, independent of the dataset or initial conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.301, + 0.283, + 0.318 + ], + "angle": 0, + "content": "3. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.326, + 0.485, + 0.538 + ], + "angle": 0, + "content": "We evaluate the proposed PALM model in fully supervised settings using ResNet-18 and ResNet-50 architectures [15] on the CIFAR-10 and CIFAR-100 datasets [20], as well as on subsets of ImageNet [9], including ImageNet-50, ImageNet-100, and ImageNet-200. In this context, fully supervised refers to training the models exclusively on the labeled subset \\(\\mathbb{L}\\) of the dataset. At each iteration, \\(\\mathbb{L}\\) is incrementally expanded with newly labeled samples in accordance with standard AL procedures. We adapted an existing AL framework to ensure reproducibility across various AL sampling strategies [14, 27]. In each experiment, we estimated the PALM parameters, which characterize the AL dynamics, based on the cumulative annotation budget \\(B\\) at the corresponding iteration." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.546, + 0.373, + 0.563 + ], + "angle": 0, + "content": "3.1. Supervised Training on CIFARs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.705 + ], + "angle": 0, + "content": "For CIFAR-10 and CIFAR-100, we trained a ResNet-18 model for at least 100 epochs. The training process employed stochastic gradient descent (SGD) with Nesterov momentum (0.9), weight decay (0.0003), and a cosine learning rate schedule, starting with an initial learning rate of 0.025. The models were trained using a batch size of 512. Standard data augmentation techniques, including random cropping and horizontal flipping, were applied. For further implementation details, please refer to [25, 27]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.449, + 0.729 + ], + "angle": 0, + "content": "3.2. Supervised Training on ImageNet Subsets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.796 + ], + "angle": 0, + "content": "For ImageNet-50, ImageNet-100, and ImageNet-200, we adopted the same optimization settings and data augmentation protocol as used in the CIFAR experiments, but with a ResNet-50 backbone [15]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.398, + 0.82 + ], + "angle": 0, + "content": "3.3. Self-Supervised Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.903 + ], + "angle": 0, + "content": "To obtain semantically meaningful feature representations, we trained SimCLR on each dataset using the official implementation from [5]. For CIFAR-10 and CIFAR-100, we employed a ResNet-18 encoder with a projection head mapping to a 128-dimensional vector, training for 500 epochs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "The training hyperparameters followed the settings outlined in SCAN [33]. After training, we extracted representations from the penultimate layer, resulting in 512-dimensional embeddings for CIFAR-10 and CIFAR-100 and 2048-dimensional embeddings for ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.319 + ], + "angle": 0, + "content": "For the ImageNet subsets, we also extracted features using the official pretrained models of BYOL [13] (ResNet-50 trained for 300 epochs with a batch size of 512), MoCov2+ [6] (ResNet-50 trained for 800 epochs with a batch size of 256), and MoCov3 [7] (ResNet-50 trained for 1000 epochs with a batch size of 4096). We computed the mean and standard deviation of the features extracted from the training samples, applied normalization accordingly, and used the same values to normalize the features extracted from the validation and test samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.327, + 0.755, + 0.344 + ], + "angle": 0, + "content": "3.4. Active Learning Strategies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.502 + ], + "angle": 0, + "content": "To demonstrate PALM's ability to approximate active learning dynamics across diverse query strategies, we evaluated several commonly used AL methods: Random Sampling, TypiClust [14], Uncertainty Sampling, which selects samples with the lowest maximum softmax probability, Max Entropy, which selects samples with the highest predictive entropy, Min Margin, which selects samples with the smallest margin between the top two softmax probabilities, and DBAL [10]. For reproducibility and fair comparison, we followed the experimental settings outlined in [27]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.513, + 0.729, + 0.529 + ], + "angle": 0, + "content": "4. Results and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.84 + ], + "angle": 0, + "content": "AL Behavior: We first evaluate the ability of PALM to approximate AL dynamics across different sampling strategies on CIFAR-10 and CIFAR-100. For CIFAR-10, we used a mean budget size of \\( b = 20 \\) over 1000 iterations, while for CIFAR-100, the budget was \\( b = 100 \\) over 450 iterations. Each AL strategy was executed until the unlabeled pool was nearly exhausted. For each AL curve, we estimated the PALM parameters \\( (\\delta, \\alpha, \\beta, \\mathrm{A}_{\\mathrm{max}}) \\) using nonlinear least squares to fit the model to the observed empirical accuracy as a function of the cumulative annotation budget \\( B \\). To ensure stable and interpretable fits, we imposed constraints on the parameters. These constraints help reduce ambiguity in the parameter estimates. Figure 2 presents the AL curves together with their corresponding PALM fits. Across both datasets, two distinct behaviors emerge: (1) AL strategies that achieve gradual accuracy improvements over extended annotation budgets and (2) strategies leveraging pretrained embeddings, which exhibit rapid accuracy gains and early convergence toward \\( \\mathrm{A}_{\\mathrm{max}} \\). These differences in behavior are reflected in the fitted PALM parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The parameter \\(\\delta\\) quantifies the coverage efficiency per labeled sample. Across both CIFAR-10 and CIFAR-100, methods that employ pretrained embeddings achieve significantly higher \\(\\delta\\) values. For example, on CIFAR-10, Margin" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4043" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.177, + 0.095, + 0.482, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.096, + 0.807, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.269, + 0.908, + 0.368 + ], + "angle": 0, + "content": "Figure 2. PALM approximation of AL curves on CIFAR-10 (\\( b = 20 \\)) and CIFAR-100 (\\( b = 100 \\)). Each plot displays the accuracy trajectories of multiple AL strategies fitted with the PALM model across the entire annotation budget. Two distinct behaviors emerge in both datasets: methods with gradual accuracy growth and those utilizing pretrained embeddings that quickly reach maximum accuracy. These differences are reflected in the fitted PALM parameters. Methods using pretrained embeddings consistently show higher \\( \\delta \\) values, indicating greater sample efficiency, while non-embedding methods exhibit higher \\( \\alpha \\) values, indicating delayed learning progress. Despite the varying complexity of the datasets, PALM successfully models the learning dynamics and highlights clear differences between AL strategies. Note that the highest test accuracy for CIFAR-100 is approximately \\( 65\\% \\). All quantitative results are available in the Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.394, + 0.483, + 0.501 + ], + "angle": 0, + "content": "sampling without embeddings yields \\(\\delta = 0.094\\), whereas its embedding-based counterpart achieves \\(\\delta = 0.535\\). Similarly, on CIFAR-100, Random sampling improves from \\(\\delta = 0.048\\) (without embeddings) to \\(\\delta = 0.318\\) (with embeddings). These indicate that pretrained embeddings enhance the representational quality of selected samples, thereby improving space coverage and annotation efficiency." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.504, + 0.483, + 0.701 + ], + "angle": 0, + "content": "The parameter \\(\\alpha\\) controls the initial offset of learning progress. Higher \\(\\alpha\\) values indicate a delayed onset of learning, a trend that is especially pronounced in non-embedding methods. For example, on CIFAR-100, Margin sampling without embeddings exhibits \\(\\alpha = 10.643\\), whereas its embedding-based counterpart reduces \\(\\alpha\\) to 0.068, indicating faster initial learning when using SSL-derived representations. Moreover, the parameter \\(\\beta\\) captures the scaling of accuracy gains for the annotation budget. Larger \\(\\beta\\) values correspond to steeper growth in accuracy. On CIFAR-100, Entropy sampling with embeddings demonstrates \\(\\beta = 0.651\\), the highest among the evaluated strategies, indicating rapid convergence." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Despite similar final accuracies across methods, the PALM parameters offer a fine-grained view of sample efficiency and learning behavior. For example, while Random sampling with embeddings achieves a high \\(\\delta = 0.536\\), Entropy sampling without embeddings is limited to \\(\\delta = 0.066\\), despite achieving comparable asymptotic performance. This illustrates that PALM not only captures endpoint accuracy but also characterizes the trajectory of learning dynamics. These findings empirically validate Eq. (13), where \\(\\delta\\), \\(\\alpha\\), and \\(\\beta\\) jointly describe the efficiency, onset, and scaling of active learning progress. The clear distinction in behaviors across AL strategies, as revealed by the fitted parameters, confirms that PALM provides a consistent and in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.905, + 0.44 + ], + "angle": 0, + "content": "terpretable model of AL dynamics across different datasets, methods, and embedding strategies. Full quantitative results are provided in the Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.906, + 0.655 + ], + "angle": 0, + "content": "While PALM accurately fits the learning trajectories across AL strategies, we observe minor deviations in the later rounds of the CIFAR-100 experiments. This is likely due to the progressive exhaustion of informative examples in the unlabeled pool. As AL progresses, the remaining samples tend to be noisier or less representative, which increases prediction uncertainty and affects fit stability. Although PALM assumes an effectively infinite unlabeled pool, this simplification may lead to small estimation drift once the pool is nearly depleted. Nevertheless, the observed deviations remain within \\(2\\%\\) of test accuracy across all budget levels. Importantly, PALM is designed to support early-stage decision-making, where accurate prediction from limited partial observations is most impactful." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Limited Budgets: To further assess the predictive capability of PALM, we evaluated its performance when fitted to partial learning curves, using only a limited number of annotated samples across different AL methods and datasets. The results are presented in Figure 3. For CIFAR-10, an accurate approximation of the full learning dynamics was achieved with as few as 1,000 labeled samples, representing just \\(2\\%\\) of the dataset. Notably, for some strategies, high-quality predictions emerged with only 200-300 labeled samples, corresponding to approximately \\(0.8 - 1.5\\%\\) of the dataset. Methods leveraging pretrained embeddings required even fewer samples to achieve strong fits, further highlighting the efficiency gains provided by SSL-based representations. Additionally, for CIFAR-100, a more complex dataset with 100 classes, reliable approximations were obtained using 5,000-10,000 labeled samples. This" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4044" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.165, + 0.087, + 0.835, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.346, + 0.908, + 0.431 + ], + "angle": 0, + "content": "Figure 3. PALM predictions of AL training curves fitted using varying numbers of cumulative budget points, as indicated in the legend. Each plot compares the predicted trajectories with the actual performance across different AL strategies on CIFAR-10 and CIFAR-100. Note that the strategies in rows 2 and 4 utilize embeddings extracted using the SimCLR method. We find that accurate predictions of the full learning dynamics can be made with cumulative budgets of approximately 1,000 samples for CIFAR-10 (2% of the dataset) and 5,000 samples for CIFAR-100 (10% of the dataset). In certain cases on CIFAR-10, reasonable predictions can be made with as few as 400 samples (0.8% of the dataset), demonstrating PALM's ability to model learning behavior even with limited early-stage supervision." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.456, + 0.483, + 0.502 + ], + "angle": 0, + "content": "translates to approximately 50-100 labeled examples per class, demonstrating that PALM can generalize effectively to higher-dimensional spaces with greater class diversity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.504, + 0.483, + 0.716 + ], + "angle": 0, + "content": "TypiClust, a method specifically designed for small-budget scenarios and reliant on pretrained embeddings, demonstrated particularly strong approximation with extremely limited supervision. On CIFAR-100, PALM accurately modeled TypiClust's learning trajectory using just 1,000 labeled samples (roughly 10 examples per class), underscoring its effectiveness under resource-constrained conditions. The use of SSL-based embeddings significantly reduced the computational cost of AL iterations, offering practical advantages for large-scale or real-time annotation workflows. This finding suggests that embedding-based AL pipelines not only enhance model performance but also improve training efficiency, making them a compelling choice for the design of future AL systems." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.718, + 0.483, + 0.839 + ], + "angle": 0, + "content": "It is worth noting that for CIFAR-10, PALM occasionally struggles to fit the sharp knee point of the learning curve due to the dataset's relatively simple structure and the rapid accuracy gains observed in early iterations. In contrast, this effect is less pronounced on CIFAR-100, where the increased complexity and class imbalance align more closely with real-world scenarios, enabling smoother approximations across the full budget range." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Large-Scale Datasets: Furthermore, to assess the generalization capability of PALM on large-scale, natural datasets, we conducted experiments on ImageNet-50, ImageNet-100, and ImageNet-200 using a total annotation budget corre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.905, + 0.638 + ], + "angle": 0, + "content": "sponding to \\(4\\%\\) of the training samples. Figure 4 presents the results of this experiment. As seen, PALM successfully approximated the learning dynamics across all tested active learning strategies and SSL-based embeddings, reinforcing its applicability to realistic, resource-constrained scenarios. Beyond capturing overall learning trajectories, PALM effectively models distinct learning behaviors, including methods that achieve rapid initial improvements and those with delayed performance onset. These differences are reflected in the fitted parameters, particularly \\(\\alpha\\) and \\(\\delta\\), which control the starting point and coverage efficiency of the learning process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.906, + 0.838 + ], + "angle": 0, + "content": "A particularly notable case arises with the BYOL embeddings. Due to BYOL's extremely slow and nearly linear accuracy progression within the small budget, PALM fits the curve as if the method has already plateaued. This behavior occurs because the early-stage dynamics are not sufficiently captured with the limited number of labeled samples. However, signs of acceleration appear in methods like TypiClust, where the final iterations show an upward trend, suggesting that with larger annotation budgets, BYOL would eventually follow the growth patterns predicted by our mathematical model. This observation underscores the need for more labeled samples to estimate future performance for slow-starting methods reliably." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Differences in \\(\\beta\\) across AL strategies also reveal important scalability trends, with higher \\(\\beta\\) values indicating steeper accuracy gains following the initial labeling phase. Notably, strategies combined with MoCov3 consistently ex" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4045" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.167, + 0.087, + 0.833, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.296, + 0.907, + 0.396 + ], + "angle": 0, + "content": "Figure 4. PALM predictions of AL curves on ImageNet subsets. Each plot compares the performance of different AL strategies combined with various SSL methods (MoCov2+, MoCov3, BYOL, and SimCLR). PALM accurately models learning dynamics under extreme annotation constraints, capturing distinct behaviors across methods and embeddings. MoCov3 consistently achieves the highest achievable accuracy \\((\\mathrm{A}_{\\mathrm{max}})\\) and coverage efficiency \\((\\delta)\\), indicating superior representation quality and faster learning. In contrast, BYOL exhibits slow, nearly linear accuracy growth with minimal coverage, indicating delayed learning that requires larger annotation budgets. SimCLR and MoCov2+ show varying dynamics depending on the AL strategy, underscoring the sensitivity of embedding performance to sample selection methods. These findings highlight PALM's ability to model and compare AL behavior across diverse settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.483, + 0.464 + ], + "angle": 0, + "content": "hibit higher \\(\\beta\\), reflecting their ability to capitalize on accumulated labeled data more effectively. A closer analysis of the ImageNet-200 results further illustrates these insights:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.466, + 0.483, + 0.525 + ], + "angle": 0, + "content": "- MoCov3 consistently achieves the highest \\( A_{\\max} \\) and \\( \\delta \\) values across all strategies, indicating superior representation quality and faster learning progression, making it highly effective under limited supervision." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.526, + 0.483, + 0.585 + ], + "angle": 0, + "content": "- SimCLR shows competitive \\( \\mathrm{A}_{\\max} \\) but lower \\( \\delta \\), suggesting slower per-sample gains and greater sensitivity to the choice of active learning strategy, as evidenced by the variability of \\( \\alpha \\) across methods." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.587, + 0.483, + 0.675 + ], + "angle": 0, + "content": "- BYOL displays consistently low \\(\\delta\\) values and, in some cases, near-zero values (e.g., with TypiClust), reflecting minimal coverage per labeled sample and delayed learning. For TypiClust specifically, a high \\(\\alpha\\) and steep \\(\\beta\\) indicate a late but accelerating learning phase, aligning with the observed rise in the final iterations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.677, + 0.483, + 0.736 + ], + "angle": 0, + "content": "- MoCov2+ generally lags behind the other embeddings, with lower \\( A_{\\max} \\) and \\( \\delta \\) across strategies, further highlighting the varying effectiveness of SSL embeddings in active learning pipelines." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.466, + 0.483, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.738, + 0.483, + 0.829 + ], + "angle": 0, + "content": "The findings show that PALM not only captures overall performance trends but also uncovers nuanced interactions between SSL embeddings and AL strategies. It offers a comprehensive model for comparing methods, balancing trade-offs between early performance and long-term scalability, and guiding annotation planning in practical applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.21, + 0.86 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We presented PALM, a unified mathematical model for the performance analysis of active learning models by captur" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.57 + ], + "angle": 0, + "content": "ing the dynamics of AL curves through a compact set of interpretable parameters. Unlike conventional evaluation protocols that compare methods based solely on accuracy at fixed budgets, PALM provides a principled approach to understanding how active learning strategies behave throughout the entire annotation process. By modeling coverage efficiency \\((\\delta)\\), sampling behavior \\((b)\\), early-stage performance \\((\\alpha)\\), scalability \\((\\beta)\\), and achievable accuracy \\((A_{\\max})\\), PALM offers a comprehensive characterization of active learning trajectories from limited observations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.907, + 0.741 + ], + "angle": 0, + "content": "Our experiments on CIFAR and ImageNet benchmarks, covering a diverse range of AL strategies and SSL embeddings, demonstrate that PALM generalizes across datasets, methods, and budget regimes despite using only a small fraction of labeled data. It not only accurately predicts learning curves under extreme annotation constraints but also reveals key differences in how methods scale, where they accelerate, and how embedding quality interacts with sampling strategies. These insights expose the limitations of relying on final accuracy alone and establish PALM as a fair model for cost-aware comparison of AL methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.909, + 0.867 + ], + "angle": 0, + "content": "PALM advances AL by shifting the focus from isolated performance measurements to predictive modeling of the learning process. This enables more informed, data-efficient decision-making in real-world settings, where annotation costs dominate and strategy selection is critical. By accounting for the dynamics and learning efficiency, PALM lays the basis for a new standard in AL evaluation, paving the way toward more systematic and scalable AL design." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This project has received funding from the Pioneer Centre for AI, DNRF, grant number P1." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4046" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.199 + ], + "angle": 0, + "content": "[1] Amin Aghaee, Mehrdad Ghadiri, and Mahdieh Soleymani Baghshah. Active distance-based clustering using k-medoids. In Advances in Knowledge Discovery and Data Mining: 20th Pacific-Asia Conference, PAKDD 2016, Auckland, New Zealand, April 19-22, 2016, Proceedings, Part I 20, pages 253-264. Springer, 2016. 1, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019. 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] William H Beluch, Tim Genewein, Andreas Nurnberger, and Jan M Kohler. The power of ensembles for active learning in image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9368-9377, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.329, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[4] Jiayi Chen, Benteng Ma, Hengfei Cui, and Yong Xia. Think twice before selection: Federated evidential active learning for medical image analysis with domain shifts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11439-11449, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.483, + 0.457 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[6] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.483, + 0.558 + ], + "angle": 0, + "content": "[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.483, + 0.627 + ], + "angle": 0, + "content": "[8] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[10] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International conference on machine learning, pages 1183-1192. PMLR, 2017. 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[11] Daniel Gissin and Shai Shalev-Shwartz. Discriminative active learning. arXiv preprint arXiv:1907.06347, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[12] Ross Greer, Björk Antoniussen, Mathias V Andersen, Andreas Møgelmose, and Mohan M Trivedi. The why, when, and how to use active learning in large-data-driven 3d object detection for safe autonomous driving: An empirical exploration. arXiv preprint arXiv:2401.16634, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[14] Guy Hacohen, Avihu Dekel, and Daphna Weinshall. Active learning on a budget: Opposite strategies suit high and low budgets. arXiv preprint arXiv:2202.02794, 2022. 2, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[16] Rong Hu, Brian Mac Namee, and Sarah Jane Delany. Off to a good start: Using clustering to select the initial training set in active learning. In FLAIRS, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[17] Svante Janson. Random coverings in several dimensions. 1986. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[18] Qiuye Jin, Mingzhi Yuan, Qin Qiao, and Zhijian Song. One-shot active learning for image segmentation via contrastive learning and diversity-based sampling. Knowledge-Based Systems, 241:108278, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[19] Seho Kee, Enrique Del Castillo, and George Runger. Query-by-committee improvement with diversity and density in batch active learning. Information Sciences, 454:401-418, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.428 + ], + "angle": 0, + "content": "[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.47 + ], + "angle": 0, + "content": "[21] David D Lewis. A sequential algorithm for training text classifiers: Corrigendum and additional data. In Acm Sigir Forum, pages 13-19. ACM New York, NY, USA, 1995. 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[22] Hailun Liang, Sunyuan Qiang, Hui Ma, Jun Wan, and Yanyan Liang. Semantic segmentation active learning with scene coverage coreset. In Chinese Conference on Biometric Recognition, pages 238-247. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[23] Jinpeng Lin, Zhihao Liang, Shengheng Deng, Lile Cai, Tao Jiang, Tianrui Li, Kui Jia, and Xun Xu. Exploring diversity-based active learning for 3d object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[24] Alaa Maalouf, Gilad Eini, Ben Mussay, Dan Feldman, and Margarita Osadchy. A unified approach to coreset learning. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[25] Inbal Mishal and Daphna Weinshall. Dcom: Active learning for all learners. arXiv preprint arXiv:2407.01804, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[26] Sudhanshu Mittal, Joshua Niemeijer, Özgün Çiçek, Maxim Tatarchenko, Jan Ehrhardt, Jörg P Schäfer, Heinz Handels, and Thomas Brox. Realistic evaluation of deep active learning for image classification and semantic segmentation. International Journal of Computer Vision, pages 1-23, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[27] Prateek Munjal, Nasir Hayat, Munawar Hayat, Jamshid Sourati, and Shadab Khan. Towards robust and reproducible active learning using neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 223-232, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[28] Vu-Linh Nguyen, Mohammad Hossein Shaker, and Eyke Hüllermeier. How to measure uncertainty in uncertainty sampling for active learning. Machine Learning, 111(1):89-122, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4047" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.119 + ], + "angle": 0, + "content": "[29] Sheldon M Ross. Introduction to probability models. Academic press, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.121, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[30] Tobias Scheffer, Christian Decomain, and Stefan Wrobel. Active hidden markov models for information extraction. In International symposium on intelligent data analysis, pages 309-318. Springer, 2001. 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.179, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[31] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.221, + 0.481, + 0.235 + ], + "angle": 0, + "content": "[32] Burr Settles. Active learning literature survey. 2009. 1, 2, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.237, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[33] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.307, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[34] Konstantin Voevodski, Maria-Florina Balcan, Heiko Röglin, Shang-Hua Teng, and Yu Xia. Active clustering of biological sequences. The Journal of Machine Learning Research, 13 (1):203–225, 2012. 1, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.364, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[35] Haoran Wang, Qiuye Jin, Shiman Li, Siyu Liu, Manning Wang, and Zhijian Song. A comprehensive survey on deep active learning in medical image analysis. Medical Image Analysis, page 103201, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.421, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[36] Ji Wu, Zhongfeng Kang, Sebastian Nørgaard Llambias, Mostafa Mehdipour Ghazi, and Mads Nielsen. Active transfer learning for 3d hippocampus segmentation. In Workshop on Medical Image Learning with Limited and Noisy Data, pages 224-234. Springer, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.491, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[37] Zhao Xu, Kai Yu, Volker Tresp, Xiaowei Xu, and Jizhi Wang. Representative sampling for text classification using support vector machines. In Advances in Information Retrieval: 25th European Conference on IR Research, ECIR 2003, Pisa, Italy, April 14-16, 2003. Proceedings 25, pages 393-407. Springer, 2003. 1, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.576, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[38] Ofer Yehuda, Avihu Dekel, Guy Hacohen, and Daphna Weinshall. Active learning through a covering lens. Advances in Neural Information Processing Systems, 35: 22354-22367, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.632, + 0.482, + 0.672 + ], + "angle": 0, + "content": "[39] Mohammed Zakariah and Abdulaziz S Almazyad. Anomaly detection for IoT systems using active learning. Applied Sciences, 13(21):12029, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.675, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[40] Jifan Zhang, Yifang Chen, Gregory Canal, Stephen Mussmann, Arnav M Das, Gantavya Bhatt, Yinglun Zhu, Jeffrey Bilmes, Simon Shaolei Du, Kevin Jamieson, et al. Labelbench: A comprehensive framework for benchmarking adaptive label-efficient learning. arXiv preprint arXiv:2306.09910, 2023. 1, 11" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4048" + } + ] +] \ No newline at end of file diff --git a/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/8bac4e62-edf8-4833-a2b8-a238e4c4c68b_origin.pdf b/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/8bac4e62-edf8-4833-a2b8-a238e4c4c68b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..093c55464b7cbbf671aa2fccda96d5f0e2d8a465 --- /dev/null +++ b/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/8bac4e62-edf8-4833-a2b8-a238e4c4c68b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:968719e3ca8e6a3418f07660c6d8324580cd1bec2fb4e356df799f69b59c480c +size 6463408 diff --git a/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/full.md b/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..71b51a8adfeb67b248fe26ccb62edc8af4a9746b --- /dev/null +++ b/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/full.md @@ -0,0 +1,324 @@ +# To Label or Not to Label: PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models + +Julia Machnio +Pioneer Centre for AI +University of Copenhagen +juma@di.ku.dk + +Mads Nielsen +Pioneer Centre for AI +University of Copenhagen +madsn@di.ku.dk + +Mostafa Mehdipour Ghazi +Pioneer Centre for AI +University of Copenhagen +ghazi@di.ku.dk + +# Abstract + +Active learning (AL) seeks to reduce annotation costs by selecting the most informative samples for labeling, making it particularly valuable in resource-constrained settings. However, traditional evaluation methods, which focus solely on final accuracy, fail to capture the full dynamics of the learning process. To address this gap, we propose PALM (Performance Analysis of Active Learning Models), a unified and interpretable mathematical model that characterizes AL trajectories through four key parameters: achievable accuracy $(A_{\mathrm{max}})$ , coverage efficiency $(\delta)$ , early-stage performance $(\alpha)$ , and scalability $(\beta)$ . PALM provides a predictive description of AL behavior from partial observations, enabling the estimation of future performance and facilitating principled comparisons across different strategies. We validate PALM through extensive experiments on CIFAR-10/100 and ImageNet-50/100/200, covering a wide range of AL methods and self-supervised embeddings. Our results demonstrate that PALM generalizes effectively across datasets, budgets, and strategies, accurately predicting full learning curves from limited labeled data. Importantly, PALM reveals crucial insights into learning efficiency, data space coverage, and the scalability of AL methods. By enabling the selection of cost-effective strategies and predicting performance under tight budget constraints, PALM lays the basis for more systematic, reproducible, and data-efficient evaluation of AL in both research and real-world applications. The code is available at: https://github.com/juliamachnio/PALM. + +# 1. Introduction + +With the rapid growth of large-scale, domain-specific datasets in real-world applications, the cost and effort required for data labeling have become significant bottlenecks in machine learning pipelines. In many domains, obtaining labeled data necessitates substantial human expertise, + +![](images/bf49bacbcccc097e2a3e9d321e30cd3052e580882a53c6e891e00bf97cc99e34.jpg) +Figure 1. Illustration of the PALM method for comparing AL strategies through interpretable parameters. Methods 1 and 2 differ in how labeled samples (green and orange) cover the data space, with $\delta$ representing the average coverage per labeled sample. The corresponding PALM equations describe accuracy growth as a function of the cumulative annotation budget $(B)$ , with parameters controlling early-stage performance $(\alpha)$ , scalability $(\beta)$ , and maximum achievable accuracy $(\mathrm{A}_{\max})$ . The learning curves show how different parameter configurations affect AL dynamics: Method 1 achieves faster accuracy gains and higher asymptotic performance due to better coverage efficiency and scalability. + +time, and resources, making it impractical to fully-associate large datasets. Active learning (AL) addresses this challenge by iteratively selecting the most informative samples for labeling, enabling models to achieve high performance with fewer labeled examples [32]. This makes AL particularly attractive in settings where labeled data is scarce or expensive, such as medical imaging [4, 35, 36], autonomous driving [12, 23], and anomaly detection in IoT systems [39]. + +Although AL provides a solution to reduce annotation costs, its effectiveness in deep learning largely depends on the sample selection strategy. Over the years, a variety of AL methods have been proposed [1, 34, 37, 40], typically + +categorized into three main groups based on their selection criteria: uncertainty-based methods, diversity-based methods, and core-set approaches. Uncertainty-based methods prioritize samples where the model exhibits low confidence [11, 21, 28, 30]. Diversity-based methods aim to maximize data space coverage by selecting diverse samples [2, 16, 18, 19]. Core-set approaches focus on selecting representative subsets that best capture the global data distribution [22, 24, 31]. Recently, a fourth category, typicality-based methods, has emerged, particularly effective in scenarios with limited annotation budgets (the number of labeled samples). These methods select samples that are most representative of the underlying data distribution, aiming to enhance coverage and learning efficiency [14, 25, 38]. + +Integrating self-supervised learning (SSL) into active learning pipelines has further enhanced typicality-based methods by providing richer feature representations that improve both sample selection and model initialization [14, 25, 38]. SSL techniques such as SimCLR [5], BYOL [13], or MoCov2+ and MoCov3 [6, 7] have become powerful tools for learning transferable representations from unlabeled data. The synergy between AL and SSL has proven especially beneficial under limited annotation budgets, where high-quality feature representations help identify the most informative and representative samples. + +Despite these advancements, the performance of AL methods remains highly sensitive to experimental settings, including model initialization, dataset properties, and task complexity [3]. Furthermore, current evaluation protocols typically focus on the final accuracy achieved after a fixed number of iterations, neglecting critical factors such as early learning efficiency, spatial coverage, and the scalability of learning gains over time. Recent work has highlighted the need for fair comparisons of AL methods in various scenarios but continues to rely on metrics such as accuracy and area under the ROC curve (AUC), which provide only a limited view of overall performance [26]. + +These comparison limitations reveal a fundamental gap; there is no general predictive model for evaluating and comparing AL performance across diverse methods, datasets, and annotation budgets. In real-world scenarios, where annotation resources are finite, the objective is not only to achieve high accuracy but also to ensure robustness and generalization within limited budgets [32]. Moreover, deep learning models often retain predictive ability in uncovered regions of the data space, meaning that performance depends not only on labeled coverage but also on the model's inherent generalization capacity. Thus, evaluating active learning strategies requires considering both coverage expansion and the model's ability to learn from unlabeled regions. Without standardized, predictive evaluations incorporating these factors, selecting the most effective AL strategy remains a costly and inconsistent process. + +To address the challenge of fair comparison, we propose PALM, a mathematical model for the Performance Analysis of Active Learning Models, describing the dynamics of AL processes through a compact set of interpretable parameters as shown in Figure 1. PALM captures not only how accuracy grows with increasing labeled data but also how uncovered regions contribute to overall performance through generalization. It estimates the full learning curve from a limited number of iterations and predicts outcomes such as final accuracy, data coverage, early learning efficiency, and scalability. This formulation generalizes across different AL methods, datasets, and configurations and is compatible with or without SSL-based feature representations. By providing an interpretable and predictive description of AL behavior, PALM enables practitioners to estimate the number of labels required to achieve target performance and identify the most cost-effective strategies under resource constraints. + +Our main contributions are as follows: (1) we propose PALM, a simple yet powerful mathematical model that describes the dynamics of AL and predicts key performance metrics from limited labeled data; (2) we empirically evaluate PALM's ability to compare different AL methods, datasets, and configurations without requiring full dataset annotation; (3) we demonstrate that PALM can estimate final accuracy, data space coverage, early learning efficiency, and scalability of learning gains across various AL settings; (4) we analyze the impact of incorporating SSL embeddings into AL pipelines, showing that PALM remains effective with or without SSL-based representations; and (5) we provide a practical approach for estimating the number of labeling samples needed to achieve target performance, facilitating real-world deployments under strict budget constraints. + +# 2. Methods + +In this section, we present the underlying calculations of PALM and the properties that enable it to predict final accuracy and space coverage for active learning methods. Our formulation is grounded in fundamental probabilistic principles [29], considering the random distribution of objects in a space and their contribution to coverage estimation. This approach is inspired by the random covering problem [17], as well as concepts such as covering and effective sample size [8] and dynamic coverage and margin mix [25]. + +# 2.1. Expected Coverage Fraction + +Definition 1 (Coverage Probability). Let $\mathbb{X} \subset \mathbb{R}^d$ represent a $d$ -dimensional space, and let $\{O_i\}_{i=1}^s$ denote a collection of $s$ randomly placed objects within $\mathbb{X}$ . Each object $O_i$ is a measurable subset of $\mathbb{X}$ . For a given point $x \in \mathbb{X}$ , we define the coverage probability $P_C(x)$ as the probability that $x$ is covered by at least one object: + +$$ +P _ {C} (x) = 1 - P _ {U C} (x), \tag {1} +$$ + +where $P_{UC}(x)$ represents the probability that $x$ is not covered by any object in $\{O_i\}_{i=1}^s$ . + +Definition 2 (Coverage Probability with $s$ Independent Objects). Assume that each object $O_{i}$ is placed independently and covers a given point $x \in \mathbb{X}$ with the same probability $p$ . The probability that a single object does not cover $x$ is $1 - p$ . Since the objects are independent, the probability that no object covers $x$ is given by: + +$$ +P _ {U C} (x) = (1 - p) ^ {s}. \tag {2} +$$ + +Therefore, the overall coverage probability of point $x$ is: + +$$ +P _ {C} (x) = 1 - (1 - p) ^ {s}. \tag {3} +$$ + +As $s \to \infty$ , $P_C(x)$ approaches 1, ensuring that $x$ is almost surely covered. When $s = 0$ , we have $P_C(x) = 0$ . Moreover, since $(1 - p)^s$ decreases with increasing $s$ for $0 < p < 1$ , the probability of non-coverage diminishes as the number of objects increases. + +Corollary 1 (Boundary Behavior of the Coverage Probability). The coverage probability $P_C(x)$ , as defined in Eq. (3), exhibits different boundary behaviors. As the number of objects grows infinitely, the probability that $x$ is covered approaches 1, ensuring full coverage of the space: + +$$ +\lim _ {s \rightarrow \infty} P _ {C} (x) = 1 - \lim _ {s \rightarrow \infty} (1 - p) ^ {s}, \tag {4} +$$ + +and since $0 < 1 - p < 1$ , we have: + +$$ +\lim _ {s \rightarrow \infty} (1 - p) ^ {s} = 0. \tag {5} +$$ + +Thus, as the number of objects approaches infinity, the coverage probability becomes: + +$$ +\lim _ {s \rightarrow \infty} P _ {C} (x) = 1. \tag {6} +$$ + +Conversely, as the number of objects approaches zero, the probability of coverage also approaches zero, indicating no coverage of the space: + +$$ +\lim _ {s \rightarrow 0 ^ {+}} P _ {C} (x) = 1 - \lim _ {s \rightarrow 0 ^ {+}} (1 - p) ^ {s} = 0. \tag {7} +$$ + +Definition 3 (Expected Coverage Fraction). Let $\mathbb{X}$ be a space with finite volume $V$ , and let $\delta = V_{O_i} / V$ denote the fraction of the total volume covered by a single randomly placed object $O_i$ , assuming all objects have the same volume and are placed independently and uniformly over $\mathbb{X}$ . Under the assumption of independence, the expected fraction of $\mathbb{X}$ covered by $s$ objects, denoted by $\mathbb{E}_C$ , is given by: + +$$ +\mathbb {E} _ {\mathrm {C}} = 1 - (1 - \delta) ^ {s} \tag {8} +$$ + +where $0 \leq \delta \leq 1$ . Note that $\mathbb{E}_{\mathbf{C}}$ describes the expected proportion of the entire space covered and is independent of any specific point in $\mathbb{X}$ . This equation exhibits the same limiting behavior as outlined in the previous corollary. A larger $\delta$ implies that fewer objects are needed to achieve significant coverage of $\mathbb{X}$ since each object covers a greater proportion of the space. + +# 2.2. Active Learning Pools + +To define the properties of AL methods, we adopt the notation introduced in [25], with a modification in the interpretation of $\delta$ . Specifically, in this context, $\delta$ represents the fraction of the space $\mathbb{X}$ covered by a single labeled sample rather than the radius of a covering ball. + +We define $\mathbb{L} \subseteq \mathbb{X}$ as the labeled set of points and $\mathbb{U} \subseteq \mathbb{X}$ as the unlabeled set, such that $\mathbb{X} = \mathbb{L} \cup \mathbb{U}$ , with the cardinality of the labeled set $|\mathbb{L}| = B$ , representing the cumulative annotation budget collected iteratively from the AL pool. A labeling round refers to an iteration of the active learning process, in which a subset of points from $\mathbb{U}$ is selected for labeling, and the labeled set $\mathbb{L}$ is updated accordingly. Throughout the AL process, the cumulative budget $B \leq |\mathbb{X}|$ increases as more rounds are completed. + +Definition 4 (Expected Coverage in Active Learning Pools). The expected fraction of $\mathbb{X}$ covered by $B$ labeled samples, selected independently, is given by: + +$$ +\mathbb {E} _ {\mathrm {C}} = 1 - (1 - \delta) ^ {B}, \tag {9} +$$ + +where, $\delta$ captures the average contribution of a single labeled sample to the overall coverage of $\mathbb{X}$ , which inherently depends on the AL strategy used to select the samples. Specifically, more effective strategies prioritize samples that maximize coverage, leading to higher $\delta$ values. Hence, comparing $\delta$ across different strategies allows for assessing their relative effectiveness in covering the space. + +Additionally, the accuracy of a model trained on the labeled set $\mathbb{L} \subseteq \mathbb{X}$ depends not only on the fraction of the space covered by labeled samples but also on the quality of the predictions within both the covered and uncovered regions [38]. Furthermore, the uncovered regions contribute to the model's generalization ability by influencing its behavior in areas with sparse or no labeled data. + +Definition 5 (Accuracy as a Function of Coverage Probability). The overall test accuracy of a trained model on the labeled set can be expressed as: + +$$ +A = A _ {C} P _ {C} + A _ {U C} \left(1 - P _ {C}\right), \tag {10} +$$ + +where $\mathrm{A_C}$ is the model's accuracy in the covered regions of $\mathbb{X}$ , $\mathrm{A_{UC}}$ is its accuracy in the uncovered regions, and $P_{\mathrm{C}}$ is the probability that a random test point lies within a covered region. By substituting $P_{\mathrm{C}}$ from Eq. (9), we obtain: + +$$ +\mathrm {A} = \mathrm {A} _ {\mathrm {C}} \left(1 - (1 - \delta) ^ {B}\right) + \mathrm {A} _ {\mathrm {U C}} (1 - \delta) ^ {B}, \tag {11} +$$ + +which establishes a simple yet powerful relationship between test accuracy and the proportion of covered space. When the labeled budget $B$ is small, the overall accuracy is dominated by $A_{\mathrm{UC}}$ , as most of the space remains uncovered. As $B$ increases, coverage expands and accuracy improves, + +eventually converging to $\mathrm{A_C}$ as $B\to \infty$ . For small $\delta$ values, the accuracy exhibits exponential growth behavior, approximately following $1 - e^{-B\delta}$ , capturing the rate at which coverage and accuracy improve with increasing budget. + +This formulation unifies the concepts of coverage and accuracy growth, showing how increasing labeled samples enhances accuracy while recognizing the importance of uncovered regions. In real-world scenarios, where models often retain some generalization capacity in uncovered areas, $\mathrm{A_{UC}}$ becomes a crucial factor in performance evaluation. However, it does not fully describe how uncovered regions influence the rate of accuracy improvement during AL. To address this, we introduce additional parameters to better reflect the impact of uncovered areas on learning dynamics. + +Definition 6 (Generalized Accuracy as a Function of Coverage with Exponential Adjustment). We define the generalized test accuracy function as: + +$$ +\mathrm {A} = \mathrm {A} _ {\max } \left(1 - (1 - \delta) ^ {(B + \alpha) ^ {\beta}}\right), \tag {12} +$$ + +where $\mathrm{A}_{\mathrm{max}}$ is the maximum achievable accuracy, $\alpha > -B$ is a shifting parameter that adjusts the effective starting point of learning by capturing the contribution of uncovered regions, and $\beta > 0$ is a scaling parameter that controls the rate at which accuracy increases with the number of labeled samples. These parameters introduce greater flexibility in modeling how space coverage shapes the learning outcome. An analysis of the asymptotic behavior of this function is provided in Supplementary Materials, showing how $\alpha$ and $\beta$ influence initial accuracy, growth rate, and the impact of uncovered regions on learning dynamics. + +The generalized accuracy function offers a robust model for approximating AL training curves across varying initial conditions, AL strategies, and datasets. Given accuracy observations at four or more distinct cumulative budgets $B$ , the parameters $A_{\mathrm{max}}$ , $\delta$ , $\alpha$ , and $\beta$ can be estimated by solving a system of equations using nonlinear regression methods. The computational complexity of this estimation is approximately $\mathcal{O}(\log (B))$ when employing optimized exponentiation techniques. + +More straightforward optimization approaches typically exhibit higher complexity, approaching $\mathcal{O}(B^{\beta})$ for large $B$ . Moreover, the rapid growth of the term $(B + \alpha)^{\beta}$ can lead to large numerical values, increasing the risk of floating-point precision errors. To mitigate these issues and enhance numerical stability, it is beneficial to normalize the cumulative budget $B$ by the mean annotation budget per iteration, denoted $b$ . This normalization reduces computational complexity to approximately $\mathcal{O}((B / b)^{\beta})$ and aligns the model with the iterative structure of AL, where labeled data is acquired in fixed increments of size $b$ per iteration. + +Definition 7 (Normalized Accuracy Function for Active Learning Budget Scaling). The generalized accuracy func + +tion with normalized budget $B$ is defined as: + +$$ +\mathrm {A} = \mathrm {A} _ {\max } \left(1 - (1 - \delta) ^ {\left(\frac {B}{b} + \alpha\right) ^ {\beta}}\right), \tag {13} +$$ + +where $b$ is the mean number of labeled samples per iteration. This normalization aligns the accuracy function with the iterative structure of AL and mitigates numerical instability associated with large budget values. Estimating the characteristic parameters $(\delta, b, \alpha, \beta, \mathrm{A}_{\mathrm{max}})$ from observed learning curves enables direct and meaningful comparisons between different AL methods. + +Theorem 1 (Comparison of Active Learning Methods Using the Normalized Accuracy Function). Consider two active learning methods 1 and 2 described by their respective accuracy functions: + +$$ +\mathrm {A} _ {1} = \mathrm {A} _ {\max , 1} \left(1 - \left(1 - \delta_ {1}\right) ^ {\left(\frac {B}{b _ {1}} + \alpha_ {1}\right) ^ {\beta_ {1}}}\right), \tag {14} +$$ + +$$ +\mathrm {A} _ {2} = \mathrm {A} _ {\max , 2} \left(1 - \left(1 - \delta_ {2}\right) ^ {\left(\frac {B}{b _ {2}} + \alpha_ {2}\right) ^ {\beta_ {2}}}\right). \tag {15} +$$ + +For a given budget $B$ , Method 1 outperforms Method 2 if $\mathrm{A}_1 > \mathrm{A}_2$ . Importantly, $\mathrm{A}_{\max}$ determines the asymptotic performance as $B \to \infty$ , with the method having higher $\mathrm{A}_{\max}$ achieving superior long-term accuracy. The rate of accuracy improvement, influenced by $\delta$ , $b$ , $\alpha$ , and $\beta$ , further captures the efficiency of each method in reaching its asymptotic potential. The full proof is presented in Supplementary Materials. + +Lemma 1 (Comparison Criterion for Active Learning Methods). For any budget $B$ , two active learning methods, A and B, can be compared directly through: + +$$ +\mathrm {A} _ {1} > \mathrm {A} _ {2}. \tag {16} +$$ + +This provides a simple and effective criterion to identify the better-performing method for a given annotation budget, making it easier to evaluate performance across different stages of the AL process. + +Corollary 2 (Comparison of Learning Efficiency and Parameter Influence). Assuming both methods achieve the same maximum accuracy, their performance can be distinguished through the following considerations: + +- Coverage Efficiency $(\delta)$ : A higher $\delta$ indicates that each labeled sample provides more effective coverage of the space, enhancing overall efficiency. +- Budget Efficiency $(b)$ : A smaller $b$ results in more frequent, smaller batches of labeled data, leading to faster and more incremental accuracy improvements. +- Initial Learning Efficiency $(\alpha)$ : A lower $\alpha$ boosts the starting accuracy, which is especially advantageous in low-budget scenarios. +- Scaling of Learning Gains $(\beta)$ : A larger $\beta$ increases the accuracy improvement rate as more samples are labeled. + +Hence, active learning methods can be quantitatively compared by evaluating these parameters and $\mathrm{A}_{\mathrm{max}}$ that defines the maximum achievable accuracy as the budget increases. + +Overall, this work presents a unified equation for modeling AL performance through a normalized accuracy function that captures the effects of coverage efficiency, budget scaling, and learning dynamics. By introducing coverage efficiency $(\delta)$ , budget efficiency $(b)$ , initial accuracy shift $(\alpha)$ , and learning gain scaling $(\beta)$ , the model captures the early-stage and long-term behavior of AL methods. This formulation enables a direct, quantitative comparison of different AL strategies based on their estimated parameters, independent of the dataset or initial conditions. + +# 3. Experimental Setup + +We evaluate the proposed PALM model in fully supervised settings using ResNet-18 and ResNet-50 architectures [15] on the CIFAR-10 and CIFAR-100 datasets [20], as well as on subsets of ImageNet [9], including ImageNet-50, ImageNet-100, and ImageNet-200. In this context, fully supervised refers to training the models exclusively on the labeled subset $\mathbb{L}$ of the dataset. At each iteration, $\mathbb{L}$ is incrementally expanded with newly labeled samples in accordance with standard AL procedures. We adapted an existing AL framework to ensure reproducibility across various AL sampling strategies [14, 27]. In each experiment, we estimated the PALM parameters, which characterize the AL dynamics, based on the cumulative annotation budget $B$ at the corresponding iteration. + +# 3.1. Supervised Training on CIFARs + +For CIFAR-10 and CIFAR-100, we trained a ResNet-18 model for at least 100 epochs. The training process employed stochastic gradient descent (SGD) with Nesterov momentum (0.9), weight decay (0.0003), and a cosine learning rate schedule, starting with an initial learning rate of 0.025. The models were trained using a batch size of 512. Standard data augmentation techniques, including random cropping and horizontal flipping, were applied. For further implementation details, please refer to [25, 27]. + +# 3.2. Supervised Training on ImageNet Subsets + +For ImageNet-50, ImageNet-100, and ImageNet-200, we adopted the same optimization settings and data augmentation protocol as used in the CIFAR experiments, but with a ResNet-50 backbone [15]. + +# 3.3. Self-Supervised Feature Extraction + +To obtain semantically meaningful feature representations, we trained SimCLR on each dataset using the official implementation from [5]. For CIFAR-10 and CIFAR-100, we employed a ResNet-18 encoder with a projection head mapping to a 128-dimensional vector, training for 500 epochs. + +The training hyperparameters followed the settings outlined in SCAN [33]. After training, we extracted representations from the penultimate layer, resulting in 512-dimensional embeddings for CIFAR-10 and CIFAR-100 and 2048-dimensional embeddings for ImageNet. + +For the ImageNet subsets, we also extracted features using the official pretrained models of BYOL [13] (ResNet-50 trained for 300 epochs with a batch size of 512), MoCov2+ [6] (ResNet-50 trained for 800 epochs with a batch size of 256), and MoCov3 [7] (ResNet-50 trained for 1000 epochs with a batch size of 4096). We computed the mean and standard deviation of the features extracted from the training samples, applied normalization accordingly, and used the same values to normalize the features extracted from the validation and test samples. + +# 3.4. Active Learning Strategies + +To demonstrate PALM's ability to approximate active learning dynamics across diverse query strategies, we evaluated several commonly used AL methods: Random Sampling, TypiClust [14], Uncertainty Sampling, which selects samples with the lowest maximum softmax probability, Max Entropy, which selects samples with the highest predictive entropy, Min Margin, which selects samples with the smallest margin between the top two softmax probabilities, and DBAL [10]. For reproducibility and fair comparison, we followed the experimental settings outlined in [27]. + +# 4. Results and Discussion + +AL Behavior: We first evaluate the ability of PALM to approximate AL dynamics across different sampling strategies on CIFAR-10 and CIFAR-100. For CIFAR-10, we used a mean budget size of $b = 20$ over 1000 iterations, while for CIFAR-100, the budget was $b = 100$ over 450 iterations. Each AL strategy was executed until the unlabeled pool was nearly exhausted. For each AL curve, we estimated the PALM parameters $(\delta, \alpha, \beta, \mathrm{A}_{\mathrm{max}})$ using nonlinear least squares to fit the model to the observed empirical accuracy as a function of the cumulative annotation budget $B$ . To ensure stable and interpretable fits, we imposed constraints on the parameters. These constraints help reduce ambiguity in the parameter estimates. Figure 2 presents the AL curves together with their corresponding PALM fits. Across both datasets, two distinct behaviors emerge: (1) AL strategies that achieve gradual accuracy improvements over extended annotation budgets and (2) strategies leveraging pretrained embeddings, which exhibit rapid accuracy gains and early convergence toward $\mathrm{A}_{\mathrm{max}}$ . These differences in behavior are reflected in the fitted PALM parameters. + +The parameter $\delta$ quantifies the coverage efficiency per labeled sample. Across both CIFAR-10 and CIFAR-100, methods that employ pretrained embeddings achieve significantly higher $\delta$ values. For example, on CIFAR-10, Margin + +![](images/fb7d9db832b5a12b0aa53e49a2e7f4e54a710fc060f1f1296033aa2eee0fe385.jpg) +Figure 2. PALM approximation of AL curves on CIFAR-10 ( $b = 20$ ) and CIFAR-100 ( $b = 100$ ). Each plot displays the accuracy trajectories of multiple AL strategies fitted with the PALM model across the entire annotation budget. Two distinct behaviors emerge in both datasets: methods with gradual accuracy growth and those utilizing pretrained embeddings that quickly reach maximum accuracy. These differences are reflected in the fitted PALM parameters. Methods using pretrained embeddings consistently show higher $\delta$ values, indicating greater sample efficiency, while non-embedding methods exhibit higher $\alpha$ values, indicating delayed learning progress. Despite the varying complexity of the datasets, PALM successfully models the learning dynamics and highlights clear differences between AL strategies. Note that the highest test accuracy for CIFAR-100 is approximately $65\%$ . All quantitative results are available in the Supplementary Materials. + +![](images/2b63ce5efe75411643b4f96585d9e91ea9823cd9a5d2a8c33bb26b2e93d3beaf.jpg) + +sampling without embeddings yields $\delta = 0.094$ , whereas its embedding-based counterpart achieves $\delta = 0.535$ . Similarly, on CIFAR-100, Random sampling improves from $\delta = 0.048$ (without embeddings) to $\delta = 0.318$ (with embeddings). These indicate that pretrained embeddings enhance the representational quality of selected samples, thereby improving space coverage and annotation efficiency. + +The parameter $\alpha$ controls the initial offset of learning progress. Higher $\alpha$ values indicate a delayed onset of learning, a trend that is especially pronounced in non-embedding methods. For example, on CIFAR-100, Margin sampling without embeddings exhibits $\alpha = 10.643$ , whereas its embedding-based counterpart reduces $\alpha$ to 0.068, indicating faster initial learning when using SSL-derived representations. Moreover, the parameter $\beta$ captures the scaling of accuracy gains for the annotation budget. Larger $\beta$ values correspond to steeper growth in accuracy. On CIFAR-100, Entropy sampling with embeddings demonstrates $\beta = 0.651$ , the highest among the evaluated strategies, indicating rapid convergence. + +Despite similar final accuracies across methods, the PALM parameters offer a fine-grained view of sample efficiency and learning behavior. For example, while Random sampling with embeddings achieves a high $\delta = 0.536$ , Entropy sampling without embeddings is limited to $\delta = 0.066$ , despite achieving comparable asymptotic performance. This illustrates that PALM not only captures endpoint accuracy but also characterizes the trajectory of learning dynamics. These findings empirically validate Eq. (13), where $\delta$ , $\alpha$ , and $\beta$ jointly describe the efficiency, onset, and scaling of active learning progress. The clear distinction in behaviors across AL strategies, as revealed by the fitted parameters, confirms that PALM provides a consistent and in + +terpretable model of AL dynamics across different datasets, methods, and embedding strategies. Full quantitative results are provided in the Supplementary Materials. + +While PALM accurately fits the learning trajectories across AL strategies, we observe minor deviations in the later rounds of the CIFAR-100 experiments. This is likely due to the progressive exhaustion of informative examples in the unlabeled pool. As AL progresses, the remaining samples tend to be noisier or less representative, which increases prediction uncertainty and affects fit stability. Although PALM assumes an effectively infinite unlabeled pool, this simplification may lead to small estimation drift once the pool is nearly depleted. Nevertheless, the observed deviations remain within $2\%$ of test accuracy across all budget levels. Importantly, PALM is designed to support early-stage decision-making, where accurate prediction from limited partial observations is most impactful. + +Limited Budgets: To further assess the predictive capability of PALM, we evaluated its performance when fitted to partial learning curves, using only a limited number of annotated samples across different AL methods and datasets. The results are presented in Figure 3. For CIFAR-10, an accurate approximation of the full learning dynamics was achieved with as few as 1,000 labeled samples, representing just $2\%$ of the dataset. Notably, for some strategies, high-quality predictions emerged with only 200-300 labeled samples, corresponding to approximately $0.8 - 1.5\%$ of the dataset. Methods leveraging pretrained embeddings required even fewer samples to achieve strong fits, further highlighting the efficiency gains provided by SSL-based representations. Additionally, for CIFAR-100, a more complex dataset with 100 classes, reliable approximations were obtained using 5,000-10,000 labeled samples. This + +![](images/5ca44b5c6d092850b4861acb11b47b121c94fd70a75afbb08a642b3429bc6954.jpg) +Figure 3. PALM predictions of AL training curves fitted using varying numbers of cumulative budget points, as indicated in the legend. Each plot compares the predicted trajectories with the actual performance across different AL strategies on CIFAR-10 and CIFAR-100. Note that the strategies in rows 2 and 4 utilize embeddings extracted using the SimCLR method. We find that accurate predictions of the full learning dynamics can be made with cumulative budgets of approximately 1,000 samples for CIFAR-10 (2% of the dataset) and 5,000 samples for CIFAR-100 (10% of the dataset). In certain cases on CIFAR-10, reasonable predictions can be made with as few as 400 samples (0.8% of the dataset), demonstrating PALM's ability to model learning behavior even with limited early-stage supervision. + +translates to approximately 50-100 labeled examples per class, demonstrating that PALM can generalize effectively to higher-dimensional spaces with greater class diversity. + +TypiClust, a method specifically designed for small-budget scenarios and reliant on pretrained embeddings, demonstrated particularly strong approximation with extremely limited supervision. On CIFAR-100, PALM accurately modeled TypiClust's learning trajectory using just 1,000 labeled samples (roughly 10 examples per class), underscoring its effectiveness under resource-constrained conditions. The use of SSL-based embeddings significantly reduced the computational cost of AL iterations, offering practical advantages for large-scale or real-time annotation workflows. This finding suggests that embedding-based AL pipelines not only enhance model performance but also improve training efficiency, making them a compelling choice for the design of future AL systems. + +It is worth noting that for CIFAR-10, PALM occasionally struggles to fit the sharp knee point of the learning curve due to the dataset's relatively simple structure and the rapid accuracy gains observed in early iterations. In contrast, this effect is less pronounced on CIFAR-100, where the increased complexity and class imbalance align more closely with real-world scenarios, enabling smoother approximations across the full budget range. + +Large-Scale Datasets: Furthermore, to assess the generalization capability of PALM on large-scale, natural datasets, we conducted experiments on ImageNet-50, ImageNet-100, and ImageNet-200 using a total annotation budget corre + +sponding to $4\%$ of the training samples. Figure 4 presents the results of this experiment. As seen, PALM successfully approximated the learning dynamics across all tested active learning strategies and SSL-based embeddings, reinforcing its applicability to realistic, resource-constrained scenarios. Beyond capturing overall learning trajectories, PALM effectively models distinct learning behaviors, including methods that achieve rapid initial improvements and those with delayed performance onset. These differences are reflected in the fitted parameters, particularly $\alpha$ and $\delta$ , which control the starting point and coverage efficiency of the learning process. + +A particularly notable case arises with the BYOL embeddings. Due to BYOL's extremely slow and nearly linear accuracy progression within the small budget, PALM fits the curve as if the method has already plateaued. This behavior occurs because the early-stage dynamics are not sufficiently captured with the limited number of labeled samples. However, signs of acceleration appear in methods like TypiClust, where the final iterations show an upward trend, suggesting that with larger annotation budgets, BYOL would eventually follow the growth patterns predicted by our mathematical model. This observation underscores the need for more labeled samples to estimate future performance for slow-starting methods reliably. + +Differences in $\beta$ across AL strategies also reveal important scalability trends, with higher $\beta$ values indicating steeper accuracy gains following the initial labeling phase. Notably, strategies combined with MoCov3 consistently ex + +![](images/7b625e2c3e82669574c7ed45ed7741f3273d7ef2316b5cc7947f2779f030dd61.jpg) +Figure 4. PALM predictions of AL curves on ImageNet subsets. Each plot compares the performance of different AL strategies combined with various SSL methods (MoCov2+, MoCov3, BYOL, and SimCLR). PALM accurately models learning dynamics under extreme annotation constraints, capturing distinct behaviors across methods and embeddings. MoCov3 consistently achieves the highest achievable accuracy $(\mathrm{A}_{\mathrm{max}})$ and coverage efficiency $(\delta)$ , indicating superior representation quality and faster learning. In contrast, BYOL exhibits slow, nearly linear accuracy growth with minimal coverage, indicating delayed learning that requires larger annotation budgets. SimCLR and MoCov2+ show varying dynamics depending on the AL strategy, underscoring the sensitivity of embedding performance to sample selection methods. These findings highlight PALM's ability to model and compare AL behavior across diverse settings. + +hibit higher $\beta$ , reflecting their ability to capitalize on accumulated labeled data more effectively. A closer analysis of the ImageNet-200 results further illustrates these insights: + +- MoCov3 consistently achieves the highest $A_{\max}$ and $\delta$ values across all strategies, indicating superior representation quality and faster learning progression, making it highly effective under limited supervision. +- SimCLR shows competitive $\mathrm{A}_{\max}$ but lower $\delta$ , suggesting slower per-sample gains and greater sensitivity to the choice of active learning strategy, as evidenced by the variability of $\alpha$ across methods. +- BYOL displays consistently low $\delta$ values and, in some cases, near-zero values (e.g., with TypiClust), reflecting minimal coverage per labeled sample and delayed learning. For TypiClust specifically, a high $\alpha$ and steep $\beta$ indicate a late but accelerating learning phase, aligning with the observed rise in the final iterations. +- MoCov2+ generally lags behind the other embeddings, with lower $A_{\max}$ and $\delta$ across strategies, further highlighting the varying effectiveness of SSL embeddings in active learning pipelines. + +The findings show that PALM not only captures overall performance trends but also uncovers nuanced interactions between SSL embeddings and AL strategies. It offers a comprehensive model for comparing methods, balancing trade-offs between early performance and long-term scalability, and guiding annotation planning in practical applications. + +# 5. Conclusion + +We presented PALM, a unified mathematical model for the performance analysis of active learning models by captur + +ing the dynamics of AL curves through a compact set of interpretable parameters. Unlike conventional evaluation protocols that compare methods based solely on accuracy at fixed budgets, PALM provides a principled approach to understanding how active learning strategies behave throughout the entire annotation process. By modeling coverage efficiency $(\delta)$ , sampling behavior $(b)$ , early-stage performance $(\alpha)$ , scalability $(\beta)$ , and achievable accuracy $(A_{\max})$ , PALM offers a comprehensive characterization of active learning trajectories from limited observations. + +Our experiments on CIFAR and ImageNet benchmarks, covering a diverse range of AL strategies and SSL embeddings, demonstrate that PALM generalizes across datasets, methods, and budget regimes despite using only a small fraction of labeled data. It not only accurately predicts learning curves under extreme annotation constraints but also reveals key differences in how methods scale, where they accelerate, and how embedding quality interacts with sampling strategies. These insights expose the limitations of relying on final accuracy alone and establish PALM as a fair model for cost-aware comparison of AL methods. + +PALM advances AL by shifting the focus from isolated performance measurements to predictive modeling of the learning process. This enables more informed, data-efficient decision-making in real-world settings, where annotation costs dominate and strategy selection is critical. By accounting for the dynamics and learning efficiency, PALM lays the basis for a new standard in AL evaluation, paving the way toward more systematic and scalable AL design. + +Acknowledgements. This project has received funding from the Pioneer Centre for AI, DNRF, grant number P1. + +# References + +[1] Amin Aghaee, Mehrdad Ghadiri, and Mahdieh Soleymani Baghshah. Active distance-based clustering using k-medoids. In Advances in Knowledge Discovery and Data Mining: 20th Pacific-Asia Conference, PAKDD 2016, Auckland, New Zealand, April 19-22, 2016, Proceedings, Part I 20, pages 253-264. Springer, 2016. 1, 11 +[2] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019. 2, 11 +[3] William H Beluch, Tim Genewein, Andreas Nurnberger, and Jan M Kohler. The power of ensembles for active learning in image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9368-9377, 2018. 2 +[4] Jiayi Chen, Benteng Ma, Hengfei Cui, and Yong Xia. Think twice before selection: Federated evidential active learning for medical image analysis with domain shifts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11439-11449, 2024. 1 +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 2, 5 +[6] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2, 5 +[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 2, 5 +[8] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 2 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 +[10] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International conference on machine learning, pages 1183-1192. PMLR, 2017. 5, 11 +[11] Daniel Gissin and Shai Shalev-Shwartz. Discriminative active learning. arXiv preprint arXiv:1907.06347, 2019. 2 +[12] Ross Greer, Björk Antoniussen, Mathias V Andersen, Andreas Møgelmose, and Mohan M Trivedi. The why, when, and how to use active learning in large-data-driven 3d object detection for safe autonomous driving: An empirical exploration. arXiv preprint arXiv:2401.16634, 2024. 1 +[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach + +to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 5 +[14] Guy Hacohen, Avihu Dekel, and Daphna Weinshall. Active learning on a budget: Opposite strategies suit high and low budgets. arXiv preprint arXiv:2202.02794, 2022. 2, 5, 11 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[16] Rong Hu, Brian Mac Namee, and Sarah Jane Delany. Off to a good start: Using clustering to select the initial training set in active learning. In FLAIRS, 2010. 2 +[17] Svante Janson. Random coverings in several dimensions. 1986. 2 +[18] Qiuye Jin, Mingzhi Yuan, Qin Qiao, and Zhijian Song. One-shot active learning for image segmentation via contrastive learning and diversity-based sampling. Knowledge-Based Systems, 241:108278, 2022. 2 +[19] Seho Kee, Enrique Del Castillo, and George Runger. Query-by-committee improvement with diversity and density in batch active learning. Information Sciences, 454:401-418, 2018. 2 +[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images, 2009. 5 +[21] David D Lewis. A sequential algorithm for training text classifiers: Corrigendum and additional data. In Acm Sigir Forum, pages 13-19. ACM New York, NY, USA, 1995. 2, 11 +[22] Hailun Liang, Sunyuan Qiang, Hui Ma, Jun Wan, and Yanyan Liang. Semantic segmentation active learning with scene coverage coreset. In Chinese Conference on Biometric Recognition, pages 238-247. Springer, 2024. 2 +[23] Jinpeng Lin, Zhihao Liang, Shengheng Deng, Lile Cai, Tao Jiang, Tianrui Li, Kui Jia, and Xun Xu. Exploring diversity-based active learning for 3d object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2024. 1 +[24] Alaa Maalouf, Gilad Eini, Ben Mussay, Dan Feldman, and Margarita Osadchy. A unified approach to coreset learning. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2 +[25] Inbal Mishal and Daphna Weinshall. Dcom: Active learning for all learners. arXiv preprint arXiv:2407.01804, 2024. 2, 3, 5 +[26] Sudhanshu Mittal, Joshua Niemeijer, Özgün Çiçek, Maxim Tatarchenko, Jan Ehrhardt, Jörg P Schäfer, Heinz Handels, and Thomas Brox. Realistic evaluation of deep active learning for image classification and semantic segmentation. International Journal of Computer Vision, pages 1-23, 2025. 2 +[27] Prateek Munjal, Nasir Hayat, Munawar Hayat, Jamshid Sourati, and Shadab Khan. Towards robust and reproducible active learning using neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 223-232, 2022. 5 +[28] Vu-Linh Nguyen, Mohammad Hossein Shaker, and Eyke Hüllermeier. How to measure uncertainty in uncertainty sampling for active learning. Machine Learning, 111(1):89-122, 2022. 2 + +[29] Sheldon M Ross. Introduction to probability models. Academic press, 2014. 2 +[30] Tobias Scheffer, Christian Decomain, and Stefan Wrobel. Active hidden markov models for information extraction. In International symposium on intelligent data analysis, pages 309-318. Springer, 2001. 2, 11 +[31] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 2, 11 +[32] Burr Settles. Active learning literature survey. 2009. 1, 2, 11 +[33] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. 5 +[34] Konstantin Voevodski, Maria-Florina Balcan, Heiko Röglin, Shang-Hua Teng, and Yu Xia. Active clustering of biological sequences. The Journal of Machine Learning Research, 13 (1):203–225, 2012. 1, 11 +[35] Haoran Wang, Qiuye Jin, Shiman Li, Siyu Liu, Manning Wang, and Zhijian Song. A comprehensive survey on deep active learning in medical image analysis. Medical Image Analysis, page 103201, 2024. 1 +[36] Ji Wu, Zhongfeng Kang, Sebastian Nørgaard Llambias, Mostafa Mehdipour Ghazi, and Mads Nielsen. Active transfer learning for 3d hippocampus segmentation. In Workshop on Medical Image Learning with Limited and Noisy Data, pages 224-234. Springer, 2023. 1 +[37] Zhao Xu, Kai Yu, Volker Tresp, Xiaowei Xu, and Jizhi Wang. Representative sampling for text classification using support vector machines. In Advances in Information Retrieval: 25th European Conference on IR Research, ECIR 2003, Pisa, Italy, April 14-16, 2003. Proceedings 25, pages 393-407. Springer, 2003. 1, 11 +[38] Ofer Yehuda, Avihu Dekel, Guy Hacohen, and Daphna Weinshall. Active learning through a covering lens. Advances in Neural Information Processing Systems, 35: 22354-22367, 2022. 2, 3 +[39] Mohammed Zakariah and Abdulaziz S Almazyad. Anomaly detection for IoT systems using active learning. Applied Sciences, 13(21):12029, 2023. 1 +[40] Jifan Zhang, Yifang Chen, Gregory Canal, Stephen Mussmann, Arnav M Das, Gantavya Bhatt, Yinglun Zhu, Jeffrey Bilmes, Simon Shaolei Du, Kevin Jamieson, et al. Labelbench: A comprehensive framework for benchmarking adaptive label-efficient learning. arXiv preprint arXiv:2306.09910, 2023. 1, 11 \ No newline at end of file diff --git a/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/images.zip b/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d084d1339fd6b2e7182cada2f94adc9b5d0ce881 --- /dev/null +++ b/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:057b0702e92d8200eb9222fad73d5c42ade582c2f4174482957303dac4661abf +size 371080 diff --git a/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/layout.json b/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c0f1fea2a7732f50f1b696f000ccab8128c6f0a5 --- /dev/null +++ b/2025/To Label or Not to Label_ PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models/layout.json @@ -0,0 +1,9836 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 492, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 492, + 140 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 492, + 140 + ], + "type": "text", + "content": "To Label or Not to Label: PALM - A Predictive Model for Evaluating Sample Efficiency in Active Learning Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 86, + 161, + 216, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 161, + 216, + 218 + ], + "spans": [ + { + "bbox": [ + 86, + 161, + 216, + 218 + ], + "type": "text", + "content": "Julia Machnio \nPioneer Centre for AI \nUniversity of Copenhagen \njuma@di.ku.dk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 241, + 161, + 370, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 161, + 370, + 217 + ], + "spans": [ + { + "bbox": [ + 241, + 161, + 370, + 217 + ], + "type": "text", + "content": "Mads Nielsen \nPioneer Centre for AI \nUniversity of Copenhagen \nmadsn@di.ku.dk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 394, + 161, + 525, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 161, + 525, + 217 + ], + "spans": [ + { + "bbox": [ + 394, + 161, + 525, + 217 + ], + "type": "text", + "content": "Mostafa Mehdipour Ghazi \nPioneer Centre for AI \nUniversity of Copenhagen \nghazi@di.ku.dk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 272, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 296, + 606 + ], + "type": "text", + "content": "Active learning (AL) seeks to reduce annotation costs by selecting the most informative samples for labeling, making it particularly valuable in resource-constrained settings. However, traditional evaluation methods, which focus solely on final accuracy, fail to capture the full dynamics of the learning process. To address this gap, we propose PALM (Performance Analysis of Active Learning Models), a unified and interpretable mathematical model that characterizes AL trajectories through four key parameters: achievable accuracy " + }, + { + "bbox": [ + 55, + 272, + 296, + 606 + ], + "type": "inline_equation", + "content": "(A_{\\mathrm{max}})" + }, + { + "bbox": [ + 55, + 272, + 296, + 606 + ], + "type": "text", + "content": ", coverage efficiency " + }, + { + "bbox": [ + 55, + 272, + 296, + 606 + ], + "type": "inline_equation", + "content": "(\\delta)" + }, + { + "bbox": [ + 55, + 272, + 296, + 606 + ], + "type": "text", + "content": ", early-stage performance " + }, + { + "bbox": [ + 55, + 272, + 296, + 606 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 55, + 272, + 296, + 606 + ], + "type": "text", + "content": ", and scalability " + }, + { + "bbox": [ + 55, + 272, + 296, + 606 + ], + "type": "inline_equation", + "content": "(\\beta)" + }, + { + "bbox": [ + 55, + 272, + 296, + 606 + ], + "type": "text", + "content": ". PALM provides a predictive description of AL behavior from partial observations, enabling the estimation of future performance and facilitating principled comparisons across different strategies. We validate PALM through extensive experiments on CIFAR-10/100 and ImageNet-50/100/200, covering a wide range of AL methods and self-supervised embeddings. Our results demonstrate that PALM generalizes effectively across datasets, budgets, and strategies, accurately predicting full learning curves from limited labeled data. Importantly, PALM reveals crucial insights into learning efficiency, data space coverage, and the scalability of AL methods. By enabling the selection of cost-effective strategies and predicting performance under tight budget constraints, PALM lays the basis for more systematic, reproducible, and data-efficient evaluation of AL in both research and real-world applications. The code is available at: https://github.com/juliamachnio/PALM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 632, + 135, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 632, + 135, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 632, + 135, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "With the rapid growth of large-scale, domain-specific datasets in real-world applications, the cost and effort required for data labeling have become significant bottlenecks in machine learning pipelines. In many domains, obtaining labeled data necessitates substantial human expertise," + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 243, + 538, + 417 + ], + "blocks": [ + { + "bbox": [ + 331, + 243, + 538, + 417 + ], + "lines": [ + { + "bbox": [ + 331, + 243, + 538, + 417 + ], + "spans": [ + { + "bbox": [ + 331, + 243, + 538, + 417 + ], + "type": "image", + "image_path": "bf49bacbcccc097e2a3e9d321e30cd3052e580882a53c6e891e00bf97cc99e34.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "type": "text", + "content": "Figure 1. Illustration of the PALM method for comparing AL strategies through interpretable parameters. Methods 1 and 2 differ in how labeled samples (green and orange) cover the data space, with " + }, + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "type": "text", + "content": " representing the average coverage per labeled sample. The corresponding PALM equations describe accuracy growth as a function of the cumulative annotation budget " + }, + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "type": "inline_equation", + "content": "(B)" + }, + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "type": "text", + "content": ", with parameters controlling early-stage performance " + }, + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "type": "text", + "content": ", scalability " + }, + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "type": "inline_equation", + "content": "(\\beta)" + }, + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "type": "text", + "content": ", and maximum achievable accuracy " + }, + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "type": "inline_equation", + "content": "(\\mathrm{A}_{\\max})" + }, + { + "bbox": [ + 313, + 426, + 555, + 548 + ], + "type": "text", + "content": ". The learning curves show how different parameter configurations affect AL dynamics: Method 1 achieves faster accuracy gains and higher asymptotic performance due to better coverage efficiency and scalability." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 569, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 555, + 665 + ], + "type": "text", + "content": "time, and resources, making it impractical to fully-associate large datasets. Active learning (AL) addresses this challenge by iteratively selecting the most informative samples for labeling, enabling models to achieve high performance with fewer labeled examples [32]. This makes AL particularly attractive in settings where labeled data is scarce or expensive, such as medical imaging [4, 35, 36], autonomous driving [12, 23], and anomaly detection in IoT systems [39]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": "Although AL provides a solution to reduce annotation costs, its effectiveness in deep learning largely depends on the sample selection strategy. Over the years, a variety of AL methods have been proposed [1, 34, 37, 40], typically" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4039" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "categorized into three main groups based on their selection criteria: uncertainty-based methods, diversity-based methods, and core-set approaches. Uncertainty-based methods prioritize samples where the model exhibits low confidence [11, 21, 28, 30]. Diversity-based methods aim to maximize data space coverage by selecting diverse samples [2, 16, 18, 19]. Core-set approaches focus on selecting representative subsets that best capture the global data distribution [22, 24, 31]. Recently, a fourth category, typicality-based methods, has emerged, particularly effective in scenarios with limited annotation budgets (the number of labeled samples). These methods select samples that are most representative of the underlying data distribution, aiming to enhance coverage and learning efficiency [14, 25, 38]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 242, + 294, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 294, + 374 + ], + "type": "text", + "content": "Integrating self-supervised learning (SSL) into active learning pipelines has further enhanced typicality-based methods by providing richer feature representations that improve both sample selection and model initialization [14, 25, 38]. SSL techniques such as SimCLR [5], BYOL [13], or MoCov2+ and MoCov3 [6, 7] have become powerful tools for learning transferable representations from unlabeled data. The synergy between AL and SSL has proven especially beneficial under limited annotation budgets, where high-quality feature representations help identify the most informative and representative samples." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 376, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 519 + ], + "type": "text", + "content": "Despite these advancements, the performance of AL methods remains highly sensitive to experimental settings, including model initialization, dataset properties, and task complexity [3]. Furthermore, current evaluation protocols typically focus on the final accuracy achieved after a fixed number of iterations, neglecting critical factors such as early learning efficiency, spatial coverage, and the scalability of learning gains over time. Recent work has highlighted the need for fair comparisons of AL methods in various scenarios but continues to rely on metrics such as accuracy and area under the ROC curve (AUC), which provide only a limited view of overall performance [26]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 714 + ], + "type": "text", + "content": "These comparison limitations reveal a fundamental gap; there is no general predictive model for evaluating and comparing AL performance across diverse methods, datasets, and annotation budgets. In real-world scenarios, where annotation resources are finite, the objective is not only to achieve high accuracy but also to ensure robustness and generalization within limited budgets [32]. Moreover, deep learning models often retain predictive ability in uncovered regions of the data space, meaning that performance depends not only on labeled coverage but also on the model's inherent generalization capacity. Thus, evaluating active learning strategies requires considering both coverage expansion and the model's ability to learn from unlabeled regions. Without standardized, predictive evaluations incorporating these factors, selecting the most effective AL strategy remains a costly and inconsistent process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": "To address the challenge of fair comparison, we propose PALM, a mathematical model for the Performance Analysis of Active Learning Models, describing the dynamics of AL processes through a compact set of interpretable parameters as shown in Figure 1. PALM captures not only how accuracy grows with increasing labeled data but also how uncovered regions contribute to overall performance through generalization. It estimates the full learning curve from a limited number of iterations and predicts outcomes such as final accuracy, data coverage, early learning efficiency, and scalability. This formulation generalizes across different AL methods, datasets, and configurations and is compatible with or without SSL-based feature representations. By providing an interpretable and predictive description of AL behavior, PALM enables practitioners to estimate the number of labels required to achieve target performance and identify the most cost-effective strategies under resource constraints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 276, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 455 + ], + "type": "text", + "content": "Our main contributions are as follows: (1) we propose PALM, a simple yet powerful mathematical model that describes the dynamics of AL and predicts key performance metrics from limited labeled data; (2) we empirically evaluate PALM's ability to compare different AL methods, datasets, and configurations without requiring full dataset annotation; (3) we demonstrate that PALM can estimate final accuracy, data space coverage, early learning efficiency, and scalability of learning gains across various AL settings; (4) we analyze the impact of incorporating SSL embeddings into AL pipelines, showing that PALM remains effective with or without SSL-based representations; and (5) we provide a practical approach for estimating the number of labeling samples needed to achieve target performance, facilitating real-world deployments under strict budget constraints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 466, + 374, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 466, + 374, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 466, + 374, + 479 + ], + "type": "text", + "content": "2. Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 487, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 553, + 594 + ], + "type": "text", + "content": "In this section, we present the underlying calculations of PALM and the properties that enable it to predict final accuracy and space coverage for active learning methods. Our formulation is grounded in fundamental probabilistic principles [29], considering the random distribution of objects in a space and their contribution to coverage estimation. This approach is inspired by the random covering problem [17], as well as concepts such as covering and effective sample size [8] and dynamic coverage and margin mix [25]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 601, + 470, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 470, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 470, + 614 + ], + "type": "text", + "content": "2.1. Expected Coverage Fraction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "text", + "content": "Definition 1 (Coverage Probability). Let " + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "inline_equation", + "content": "\\mathbb{X} \\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "text", + "content": " represent a " + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "text", + "content": "-dimensional space, and let " + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "inline_equation", + "content": "\\{O_i\\}_{i=1}^s" + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "text", + "content": " denote a collection of " + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "text", + "content": " randomly placed objects within " + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "text", + "content": ". Each object " + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "text", + "content": " is a measurable subset of " + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "text", + "content": ". For a given point " + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{X}" + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "text", + "content": ", we define the coverage probability " + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "inline_equation", + "content": "P_C(x)" + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "text", + "content": " as the probability that " + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 619, + 553, + 691 + ], + "type": "text", + "content": " is covered by at least one object:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 386, + 701, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 386, + 701, + 553, + 714 + ], + "type": "interline_equation", + "content": "P _ {C} (x) = 1 - P _ {U C} (x), \\tag {1}", + "image_path": "21b7a572e3591f075e7c89ed6fc54641fdcd3c8450e0b4a94bb577991e5bb356.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4040" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "P_{UC}(x)" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " represents the probability that " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " is not covered by any object in " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\{O_i\\}_{i=1}^s" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": "Definition 2 (Coverage Probability with " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": " Independent Objects). Assume that each object " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": " is placed independently and covers a given point " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{X}" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": " with the same probability " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": ". The probability that a single object does not cover " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "1 - p" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": ". Since the objects are independent, the probability that no object covers " + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 96, + 296, + 168 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 173, + 295, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 173, + 295, + 186 + ], + "spans": [ + { + "bbox": [ + 132, + 173, + 295, + 186 + ], + "type": "interline_equation", + "content": "P _ {U C} (x) = (1 - p) ^ {s}. \\tag {2}", + "image_path": "43e14bbe68057cf14783f1d13cb3ab0b2aa6324f104ce6b6010adf48f4cff11c.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 191, + 284, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 284, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 284, + 204 + ], + "type": "text", + "content": "Therefore, the overall coverage probability of point " + }, + { + "bbox": [ + 55, + 191, + 284, + 204 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 191, + 284, + 204 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 209, + 295, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 209, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 127, + 209, + 295, + 222 + ], + "type": "interline_equation", + "content": "P _ {C} (x) = 1 - (1 - p) ^ {s}. \\tag {3}", + "image_path": "cb4de70e0506504d71a6bf92b7d794ee5a1698b70219349b8545511cd5afa7bd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "text", + "content": "As " + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "inline_equation", + "content": "s \\to \\infty" + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "inline_equation", + "content": "P_C(x)" + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "text", + "content": " approaches 1, ensuring that " + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "text", + "content": " is almost surely covered. When " + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "inline_equation", + "content": "s = 0" + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "inline_equation", + "content": "P_C(x) = 0" + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "text", + "content": ". Moreover, since " + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "inline_equation", + "content": "(1 - p)^s" + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "text", + "content": " decreases with increasing " + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "inline_equation", + "content": "0 < p < 1" + }, + { + "bbox": [ + 55, + 226, + 296, + 285 + ], + "type": "text", + "content": ", the probability of non-coverage diminishes as the number of objects increases." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 286, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 296, + 346 + ], + "type": "text", + "content": "Corollary 1 (Boundary Behavior of the Coverage Probability). The coverage probability " + }, + { + "bbox": [ + 55, + 286, + 296, + 346 + ], + "type": "inline_equation", + "content": "P_C(x)" + }, + { + "bbox": [ + 55, + 286, + 296, + 346 + ], + "type": "text", + "content": ", as defined in Eq. (3), exhibits different boundary behaviors. As the number of objects grows infinitely, the probability that " + }, + { + "bbox": [ + 55, + 286, + 296, + 346 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 286, + 296, + 346 + ], + "type": "text", + "content": " is covered approaches 1, ensuring full coverage of the space:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 106, + 351, + 295, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 351, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 106, + 351, + 295, + 369 + ], + "type": "interline_equation", + "content": "\\lim _ {s \\rightarrow \\infty} P _ {C} (x) = 1 - \\lim _ {s \\rightarrow \\infty} (1 - p) ^ {s}, \\tag {4}", + "image_path": "1367958b36d1303871c035a96686be53eb1e1b0a69f64596fe5efb2d0f33afaf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 372, + 198, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 198, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 198, + 384 + ], + "type": "text", + "content": "and since " + }, + { + "bbox": [ + 55, + 372, + 198, + 384 + ], + "type": "inline_equation", + "content": "0 < 1 - p < 1" + }, + { + "bbox": [ + 55, + 372, + 198, + 384 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 137, + 389, + 295, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 389, + 295, + 406 + ], + "spans": [ + { + "bbox": [ + 137, + 389, + 295, + 406 + ], + "type": "interline_equation", + "content": "\\lim _ {s \\rightarrow \\infty} (1 - p) ^ {s} = 0. \\tag {5}", + "image_path": "adede1ebff7086126faa7be8a155d33da7e4ed99989c04f71c9cc645f08755e1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 434 + ], + "type": "text", + "content": "Thus, as the number of objects approaches infinity, the coverage probability becomes:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 140, + 439, + 295, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 439, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 140, + 439, + 295, + 456 + ], + "type": "interline_equation", + "content": "\\lim _ {s \\rightarrow \\infty} P _ {C} (x) = 1. \\tag {6}", + "image_path": "1e7a045c9c405ee4c84bee7cf14333ef97dd413a249d1ebb1e505c876e37158a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 460, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 496 + ], + "type": "text", + "content": "Conversely, as the number of objects approaches zero, the probability of coverage also approaches zero, indicating no coverage of the space:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 96, + 501, + 295, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 501, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 96, + 501, + 295, + 519 + ], + "type": "interline_equation", + "content": "\\lim _ {s \\rightarrow 0 ^ {+}} P _ {C} (x) = 1 - \\lim _ {s \\rightarrow 0 ^ {+}} (1 - p) ^ {s} = 0. \\tag {7}", + "image_path": "1006013e44a2ecee142b4cf510c6e6caa5d6b3e67b307906b0f058e251fff0e5.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "text", + "content": "Definition 3 (Expected Coverage Fraction). Let " + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "text", + "content": " be a space with finite volume " + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "text", + "content": ", and let " + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "inline_equation", + "content": "\\delta = V_{O_i} / V" + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "text", + "content": " denote the fraction of the total volume covered by a single randomly placed object " + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "text", + "content": ", assuming all objects have the same volume and are placed independently and uniformly over " + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "text", + "content": ". Under the assumption of independence, the expected fraction of " + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "text", + "content": " covered by " + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "text", + "content": " objects, denoted by " + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_C" + }, + { + "bbox": [ + 55, + 522, + 296, + 607 + ], + "type": "text", + "content": ", is given by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 135, + 612, + 295, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 612, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 135, + 612, + 295, + 624 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\mathrm {C}} = 1 - (1 - \\delta) ^ {s} \\tag {8}", + "image_path": "b9568a19e2dcc4490f21e195e789e6d74c6f198af546870c5be5e23eecc1cf5b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "0 \\leq \\delta \\leq 1" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\mathbf{C}}" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " describes the expected proportion of the entire space covered and is independent of any specific point in " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": ". This equation exhibits the same limiting behavior as outlined in the previous corollary. A larger " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " implies that fewer objects are needed to achieve significant coverage of " + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": " since each object covers a greater proportion of the space." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 72, + 440, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 440, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 440, + 84 + ], + "type": "text", + "content": "2.2. Active Learning Pools" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 90, + 553, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 553, + 149 + ], + "type": "text", + "content": "To define the properties of AL methods, we adopt the notation introduced in [25], with a modification in the interpretation of " + }, + { + "bbox": [ + 313, + 90, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 90, + 553, + 149 + ], + "type": "text", + "content": ". Specifically, in this context, " + }, + { + "bbox": [ + 313, + 90, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 90, + 553, + 149 + ], + "type": "text", + "content": " represents the fraction of the space " + }, + { + "bbox": [ + 313, + 90, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 313, + 90, + 553, + 149 + ], + "type": "text", + "content": " covered by a single labeled sample rather than the radius of a covering ball." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "text", + "content": "We define " + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "inline_equation", + "content": "\\mathbb{L} \\subseteq \\mathbb{X}" + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "text", + "content": " as the labeled set of points and " + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "inline_equation", + "content": "\\mathbb{U} \\subseteq \\mathbb{X}" + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "text", + "content": " as the unlabeled set, such that " + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "inline_equation", + "content": "\\mathbb{X} = \\mathbb{L} \\cup \\mathbb{U}" + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "text", + "content": ", with the cardinality of the labeled set " + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "inline_equation", + "content": "|\\mathbb{L}| = B" + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "text", + "content": ", representing the cumulative annotation budget collected iteratively from the AL pool. A labeling round refers to an iteration of the active learning process, in which a subset of points from " + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "inline_equation", + "content": "\\mathbb{U}" + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "text", + "content": " is selected for labeling, and the labeled set " + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "inline_equation", + "content": "\\mathbb{L}" + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "text", + "content": " is updated accordingly. Throughout the AL process, the cumulative budget " + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "inline_equation", + "content": "B \\leq |\\mathbb{X}|" + }, + { + "bbox": [ + 313, + 150, + 554, + 258 + ], + "type": "text", + "content": " increases as more rounds are completed." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 259, + 554, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 554, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 554, + 295 + ], + "type": "text", + "content": "Definition 4 (Expected Coverage in Active Learning Pools). The expected fraction of " + }, + { + "bbox": [ + 313, + 259, + 554, + 295 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 313, + 259, + 554, + 295 + ], + "type": "text", + "content": " covered by " + }, + { + "bbox": [ + 313, + 259, + 554, + 295 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 259, + 554, + 295 + ], + "type": "text", + "content": " labeled samples, selected independently, is given by:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 391, + 305, + 553, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 305, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 391, + 305, + 553, + 320 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {\\mathrm {C}} = 1 - (1 - \\delta) ^ {B}, \\tag {9}", + "image_path": "7942ae972feb7144ce5aef09779ebfc94b13869bac5cdc243645419ad4e2900f.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 330, + 554, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 413 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 313, + 330, + 554, + 413 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 330, + 554, + 413 + ], + "type": "text", + "content": " captures the average contribution of a single labeled sample to the overall coverage of " + }, + { + "bbox": [ + 313, + 330, + 554, + 413 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 313, + 330, + 554, + 413 + ], + "type": "text", + "content": ", which inherently depends on the AL strategy used to select the samples. Specifically, more effective strategies prioritize samples that maximize coverage, leading to higher " + }, + { + "bbox": [ + 313, + 330, + 554, + 413 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 330, + 554, + 413 + ], + "type": "text", + "content": " values. Hence, comparing " + }, + { + "bbox": [ + 313, + 330, + 554, + 413 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 330, + 554, + 413 + ], + "type": "text", + "content": " across different strategies allows for assessing their relative effectiveness in covering the space." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 415, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 498 + ], + "type": "text", + "content": "Additionally, the accuracy of a model trained on the labeled set " + }, + { + "bbox": [ + 313, + 415, + 554, + 498 + ], + "type": "inline_equation", + "content": "\\mathbb{L} \\subseteq \\mathbb{X}" + }, + { + "bbox": [ + 313, + 415, + 554, + 498 + ], + "type": "text", + "content": " depends not only on the fraction of the space covered by labeled samples but also on the quality of the predictions within both the covered and uncovered regions [38]. Furthermore, the uncovered regions contribute to the model's generalization ability by influencing its behavior in areas with sparse or no labeled data." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 499, + 554, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 554, + 535 + ], + "type": "text", + "content": "Definition 5 (Accuracy as a Function of Coverage Probability). The overall test accuracy of a trained model on the labeled set can be expressed as:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 374, + 547, + 553, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 547, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 374, + 547, + 553, + 559 + ], + "type": "interline_equation", + "content": "A = A _ {C} P _ {C} + A _ {U C} \\left(1 - P _ {C}\\right), \\tag {10}", + "image_path": "4f04acefdf9313fbcdde051ad276c09b96ef084a622e683dfcd83f3ef7d526d4.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "type": "inline_equation", + "content": "\\mathrm{A_C}" + }, + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "type": "text", + "content": " is the model's accuracy in the covered regions of " + }, + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "type": "inline_equation", + "content": "\\mathbb{X}" + }, + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "type": "inline_equation", + "content": "\\mathrm{A_{UC}}" + }, + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "type": "text", + "content": " is its accuracy in the uncovered regions, and " + }, + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{C}}" + }, + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "type": "text", + "content": " is the probability that a random test point lies within a covered region. By substituting " + }, + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{C}}" + }, + { + "bbox": [ + 313, + 570, + 554, + 619 + ], + "type": "text", + "content": " from Eq. (9), we obtain:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 347, + 628, + 553, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 628, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 347, + 628, + 553, + 643 + ], + "type": "interline_equation", + "content": "\\mathrm {A} = \\mathrm {A} _ {\\mathrm {C}} \\left(1 - (1 - \\delta) ^ {B}\\right) + \\mathrm {A} _ {\\mathrm {U C}} (1 - \\delta) ^ {B}, \\tag {11}", + "image_path": "bd31d48e9a209accf27417cc14ed5659edc0220888ff9ff4fa9491e7ea2e6668.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": "which establishes a simple yet powerful relationship between test accuracy and the proportion of covered space. When the labeled budget " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": " is small, the overall accuracy is dominated by " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{UC}}" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": ", as most of the space remains uncovered. As " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": " increases, coverage expands and accuracy improves," + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4041" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "eventually converging to " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\mathrm{A_C}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "B\\to \\infty" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " . For small " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " values, the accuracy exhibits exponential growth behavior, approximately following " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "1 - e^{-B\\delta}" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " , capturing the rate at which coverage and accuracy improve with increasing budget." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 240 + ], + "type": "text", + "content": "This formulation unifies the concepts of coverage and accuracy growth, showing how increasing labeled samples enhances accuracy while recognizing the importance of uncovered regions. In real-world scenarios, where models often retain some generalization capacity in uncovered areas, " + }, + { + "bbox": [ + 55, + 121, + 295, + 240 + ], + "type": "inline_equation", + "content": "\\mathrm{A_{UC}}" + }, + { + "bbox": [ + 55, + 121, + 295, + 240 + ], + "type": "text", + "content": " becomes a crucial factor in performance evaluation. However, it does not fully describe how uncovered regions influence the rate of accuracy improvement during AL. To address this, we introduce additional parameters to better reflect the impact of uncovered areas on learning dynamics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 241, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 296, + 277 + ], + "type": "text", + "content": "Definition 6 (Generalized Accuracy as a Function of Coverage with Exponential Adjustment). We define the generalized test accuracy function as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 286, + 295, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 286, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 105, + 286, + 295, + 308 + ], + "type": "interline_equation", + "content": "\\mathrm {A} = \\mathrm {A} _ {\\max } \\left(1 - (1 - \\delta) ^ {(B + \\alpha) ^ {\\beta}}\\right), \\tag {12}", + "image_path": "b2246c0b254d298c7b2656cc2a731109ec7b373e667d6464706d9e1e0b1095e8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "type": "text", + "content": " is the maximum achievable accuracy, " + }, + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\alpha > -B" + }, + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "type": "text", + "content": " is a shifting parameter that adjusts the effective starting point of learning by capturing the contribution of uncovered regions, and " + }, + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\beta > 0" + }, + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "type": "text", + "content": " is a scaling parameter that controls the rate at which accuracy increases with the number of labeled samples. These parameters introduce greater flexibility in modeling how space coverage shapes the learning outcome. An analysis of the asymptotic behavior of this function is provided in Supplementary Materials, showing how " + }, + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 316, + 296, + 449 + ], + "type": "text", + "content": " influence initial accuracy, growth rate, and the impact of uncovered regions on learning dynamics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "text", + "content": "The generalized accuracy function offers a robust model for approximating AL training curves across varying initial conditions, AL strategies, and datasets. Given accuracy observations at four or more distinct cumulative budgets " + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "text", + "content": ", the parameters " + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "text", + "content": " can be estimated by solving a system of equations using nonlinear regression methods. The computational complexity of this estimation is approximately " + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\log (B))" + }, + { + "bbox": [ + 55, + 449, + 295, + 556 + ], + "type": "text", + "content": " when employing optimized exponentiation techniques." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": "More straightforward optimization approaches typically exhibit higher complexity, approaching " + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(B^{\\beta})" + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": " for large " + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": ". Moreover, the rapid growth of the term " + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "inline_equation", + "content": "(B + \\alpha)^{\\beta}" + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": " can lead to large numerical values, increasing the risk of floating-point precision errors. To mitigate these issues and enhance numerical stability, it is beneficial to normalize the cumulative budget " + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": " by the mean annotation budget per iteration, denoted " + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": ". This normalization reduces computational complexity to approximately " + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathcal{O}((B / b)^{\\beta})" + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": " and aligns the model with the iterative structure of AL, where labeled data is acquired in fixed increments of size " + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": " per iteration." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Definition 7 (Normalized Accuracy Function for Active Learning Budget Scaling). The generalized accuracy func" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 495, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 495, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 495, + 84 + ], + "type": "text", + "content": "tion with normalized budget " + }, + { + "bbox": [ + 313, + 72, + 495, + 84 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 72, + 495, + 84 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 362, + 92, + 553, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 92, + 553, + 113 + ], + "spans": [ + { + "bbox": [ + 362, + 92, + 553, + 113 + ], + "type": "interline_equation", + "content": "\\mathrm {A} = \\mathrm {A} _ {\\max } \\left(1 - (1 - \\delta) ^ {\\left(\\frac {B}{b} + \\alpha\\right) ^ {\\beta}}\\right), \\tag {13}", + "image_path": "da7b04b7b529ca994ef7d658496736fe0d484550451de9a70261403b123624f9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 121, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 204 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 121, + 553, + 204 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 121, + 553, + 204 + ], + "type": "text", + "content": " is the mean number of labeled samples per iteration. This normalization aligns the accuracy function with the iterative structure of AL and mitigates numerical instability associated with large budget values. Estimating the characteristic parameters " + }, + { + "bbox": [ + 313, + 121, + 553, + 204 + ], + "type": "inline_equation", + "content": "(\\delta, b, \\alpha, \\beta, \\mathrm{A}_{\\mathrm{max}})" + }, + { + "bbox": [ + 313, + 121, + 553, + 204 + ], + "type": "text", + "content": " from observed learning curves enables direct and meaningful comparisons between different AL methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 205, + 553, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 253 + ], + "type": "text", + "content": "Theorem 1 (Comparison of Active Learning Methods Using the Normalized Accuracy Function). Consider two active learning methods 1 and 2 described by their respective accuracy functions:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 261, + 553, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 261, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 347, + 261, + 553, + 287 + ], + "type": "interline_equation", + "content": "\\mathrm {A} _ {1} = \\mathrm {A} _ {\\max , 1} \\left(1 - \\left(1 - \\delta_ {1}\\right) ^ {\\left(\\frac {B}{b _ {1}} + \\alpha_ {1}\\right) ^ {\\beta_ {1}}}\\right), \\tag {14}", + "image_path": "98a44e5cc486588f0b905177ac5cf4a329347413f0cd2bfad3c0a0069a9e0585.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 289, + 553, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 289, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 348, + 289, + 553, + 316 + ], + "type": "interline_equation", + "content": "\\mathrm {A} _ {2} = \\mathrm {A} _ {\\max , 2} \\left(1 - \\left(1 - \\delta_ {2}\\right) ^ {\\left(\\frac {B}{b _ {2}} + \\alpha_ {2}\\right) ^ {\\beta_ {2}}}\\right). \\tag {15}", + "image_path": "4d5fba4e220579bdfc1dc7f98adea0df316d05a6598ddbdaa0ca9f7ff80c9e48.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": "For a given budget " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": ", Method 1 outperforms Method 2 if " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_1 > \\mathrm{A}_2" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": ". Importantly, " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_{\\max}" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": " determines the asymptotic performance as " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "B \\to \\infty" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": ", with the method having higher " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_{\\max}" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": " achieving superior long-term accuracy. The rate of accuracy improvement, influenced by " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 323, + 555, + 418 + ], + "type": "text", + "content": ", further captures the efficiency of each method in reaching its asymptotic potential. The full proof is presented in Supplementary Materials." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "type": "text", + "content": "Lemma 1 (Comparison Criterion for Active Learning Methods). For any budget " + }, + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 419, + 553, + 455 + ], + "type": "text", + "content": ", two active learning methods, A and B, can be compared directly through:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 412, + 464, + 553, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 412, + 464, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 412, + 464, + 553, + 476 + ], + "type": "interline_equation", + "content": "\\mathrm {A} _ {1} > \\mathrm {A} _ {2}. \\tag {16}", + "image_path": "f8ae465ca3b44e486fc4ef540c91caf75eef991934668cf4fbbc7da068b5cdea.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 486, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 533 + ], + "type": "text", + "content": "This provides a simple and effective criterion to identify the better-performing method for a given annotation budget, making it easier to evaluate performance across different stages of the AL process." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 534, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 582 + ], + "type": "text", + "content": "Corollary 2 (Comparison of Learning Efficiency and Parameter Influence). Assuming both methods achieve the same maximum accuracy, their performance can be distinguished through the following considerations:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "content": "- Coverage Efficiency " + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "inline_equation", + "content": "(\\delta)" + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "content": ": A higher " + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 582, + 553, + 617 + ], + "type": "text", + "content": " indicates that each labeled sample provides more effective coverage of the space, enhancing overall efficiency." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 618, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 653 + ], + "type": "text", + "content": "- Budget Efficiency " + }, + { + "bbox": [ + 313, + 618, + 553, + 653 + ], + "type": "inline_equation", + "content": "(b)" + }, + { + "bbox": [ + 313, + 618, + 553, + 653 + ], + "type": "text", + "content": ": A smaller " + }, + { + "bbox": [ + 313, + 618, + 553, + 653 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 618, + 553, + 653 + ], + "type": "text", + "content": " results in more frequent, smaller batches of labeled data, leading to faster and more incremental accuracy improvements." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": "- Initial Learning Efficiency " + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": ": A lower " + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": " boosts the starting accuracy, which is especially advantageous in low-budget scenarios." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "- Scaling of Learning Gains " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "(\\beta)" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": ": A larger " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " increases the accuracy improvement rate as more samples are labeled." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4042" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "Hence, active learning methods can be quantitatively compared by evaluating these parameters and " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": " that defines the maximum achievable accuracy as the budget increases." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 108, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 108, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 54, + 108, + 294, + 228 + ], + "type": "text", + "content": "Overall, this work presents a unified equation for modeling AL performance through a normalized accuracy function that captures the effects of coverage efficiency, budget scaling, and learning dynamics. By introducing coverage efficiency " + }, + { + "bbox": [ + 54, + 108, + 294, + 228 + ], + "type": "inline_equation", + "content": "(\\delta)" + }, + { + "bbox": [ + 54, + 108, + 294, + 228 + ], + "type": "text", + "content": ", budget efficiency " + }, + { + "bbox": [ + 54, + 108, + 294, + 228 + ], + "type": "inline_equation", + "content": "(b)" + }, + { + "bbox": [ + 54, + 108, + 294, + 228 + ], + "type": "text", + "content": ", initial accuracy shift " + }, + { + "bbox": [ + 54, + 108, + 294, + 228 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 54, + 108, + 294, + 228 + ], + "type": "text", + "content": ", and learning gain scaling " + }, + { + "bbox": [ + 54, + 108, + 294, + 228 + ], + "type": "inline_equation", + "content": "(\\beta)" + }, + { + "bbox": [ + 54, + 108, + 294, + 228 + ], + "type": "text", + "content": ", the model captures the early-stage and long-term behavior of AL methods. This formulation enables a direct, quantitative comparison of different AL strategies based on their estimated parameters, independent of the dataset or initial conditions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 238, + 173, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 173, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 173, + 251 + ], + "type": "text", + "content": "3. Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 258, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 296, + 426 + ], + "type": "text", + "content": "We evaluate the proposed PALM model in fully supervised settings using ResNet-18 and ResNet-50 architectures [15] on the CIFAR-10 and CIFAR-100 datasets [20], as well as on subsets of ImageNet [9], including ImageNet-50, ImageNet-100, and ImageNet-200. In this context, fully supervised refers to training the models exclusively on the labeled subset " + }, + { + "bbox": [ + 56, + 258, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathbb{L}" + }, + { + "bbox": [ + 56, + 258, + 296, + 426 + ], + "type": "text", + "content": " of the dataset. At each iteration, " + }, + { + "bbox": [ + 56, + 258, + 296, + 426 + ], + "type": "inline_equation", + "content": "\\mathbb{L}" + }, + { + "bbox": [ + 56, + 258, + 296, + 426 + ], + "type": "text", + "content": " is incrementally expanded with newly labeled samples in accordance with standard AL procedures. We adapted an existing AL framework to ensure reproducibility across various AL sampling strategies [14, 27]. In each experiment, we estimated the PALM parameters, which characterize the AL dynamics, based on the cumulative annotation budget " + }, + { + "bbox": [ + 56, + 258, + 296, + 426 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 56, + 258, + 296, + 426 + ], + "type": "text", + "content": " at the corresponding iteration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 432, + 228, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 228, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 228, + 445 + ], + "type": "text", + "content": "3.1. Supervised Training on CIFARs" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 558 + ], + "type": "text", + "content": "For CIFAR-10 and CIFAR-100, we trained a ResNet-18 model for at least 100 epochs. The training process employed stochastic gradient descent (SGD) with Nesterov momentum (0.9), weight decay (0.0003), and a cosine learning rate schedule, starting with an initial learning rate of 0.025. The models were trained using a batch size of 512. Standard data augmentation techniques, including random cropping and horizontal flipping, were applied. For further implementation details, please refer to [25, 27]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 564, + 274, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 274, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 274, + 577 + ], + "type": "text", + "content": "3.2. Supervised Training on ImageNet Subsets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 630 + ], + "type": "text", + "content": "For ImageNet-50, ImageNet-100, and ImageNet-200, we adopted the same optimization settings and data augmentation protocol as used in the CIFAR experiments, but with a ResNet-50 backbone [15]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 636, + 243, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 243, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 243, + 649 + ], + "type": "text", + "content": "3.3. Self-Supervised Feature Extraction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "content": "To obtain semantically meaningful feature representations, we trained SimCLR on each dataset using the official implementation from [5]. For CIFAR-10 and CIFAR-100, we employed a ResNet-18 encoder with a projection head mapping to a 128-dimensional vector, training for 500 epochs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "The training hyperparameters followed the settings outlined in SCAN [33]. After training, we extracted representations from the penultimate layer, resulting in 512-dimensional embeddings for CIFAR-10 and CIFAR-100 and 2048-dimensional embeddings for ImageNet." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 133, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 252 + ], + "type": "text", + "content": "For the ImageNet subsets, we also extracted features using the official pretrained models of BYOL [13] (ResNet-50 trained for 300 epochs with a batch size of 512), MoCov2+ [6] (ResNet-50 trained for 800 epochs with a batch size of 256), and MoCov3 [7] (ResNet-50 trained for 1000 epochs with a batch size of 4096). We computed the mean and standard deviation of the features extracted from the training samples, applied normalization accordingly, and used the same values to normalize the features extracted from the validation and test samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 258, + 462, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 462, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 462, + 272 + ], + "type": "text", + "content": "3.4. Active Learning Strategies" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 276, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 397 + ], + "type": "text", + "content": "To demonstrate PALM's ability to approximate active learning dynamics across diverse query strategies, we evaluated several commonly used AL methods: Random Sampling, TypiClust [14], Uncertainty Sampling, which selects samples with the lowest maximum softmax probability, Max Entropy, which selects samples with the highest predictive entropy, Min Margin, which selects samples with the smallest margin between the top two softmax probabilities, and DBAL [10]. For reproducibility and fair comparison, we followed the experimental settings outlined in [27]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 406, + 446, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 446, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 446, + 418 + ], + "type": "text", + "content": "4. Results and Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "type": "text", + "content": "AL Behavior: We first evaluate the ability of PALM to approximate AL dynamics across different sampling strategies on CIFAR-10 and CIFAR-100. For CIFAR-10, we used a mean budget size of " + }, + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "type": "inline_equation", + "content": "b = 20" + }, + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "type": "text", + "content": " over 1000 iterations, while for CIFAR-100, the budget was " + }, + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "type": "inline_equation", + "content": "b = 100" + }, + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "type": "text", + "content": " over 450 iterations. Each AL strategy was executed until the unlabeled pool was nearly exhausted. For each AL curve, we estimated the PALM parameters " + }, + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "type": "inline_equation", + "content": "(\\delta, \\alpha, \\beta, \\mathrm{A}_{\\mathrm{max}})" + }, + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "type": "text", + "content": " using nonlinear least squares to fit the model to the observed empirical accuracy as a function of the cumulative annotation budget " + }, + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "type": "text", + "content": ". To ensure stable and interpretable fits, we imposed constraints on the parameters. These constraints help reduce ambiguity in the parameter estimates. Figure 2 presents the AL curves together with their corresponding PALM fits. Across both datasets, two distinct behaviors emerge: (1) AL strategies that achieve gradual accuracy improvements over extended annotation budgets and (2) strategies leveraging pretrained embeddings, which exhibit rapid accuracy gains and early convergence toward " + }, + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 426, + 555, + 665 + ], + "type": "text", + "content": ". These differences in behavior are reflected in the fitted PALM parameters." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "The parameter " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " quantifies the coverage efficiency per labeled sample. Across both CIFAR-10 and CIFAR-100, methods that employ pretrained embeddings achieve significantly higher " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " values. For example, on CIFAR-10, Margin" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4043" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 75, + 294, + 202 + ], + "blocks": [ + { + "bbox": [ + 108, + 75, + 294, + 202 + ], + "lines": [ + { + "bbox": [ + 108, + 75, + 294, + 202 + ], + "spans": [ + { + "bbox": [ + 108, + 75, + 294, + 202 + ], + "type": "image", + "image_path": "fb7d9db832b5a12b0aa53e49a2e7f4e54a710fc060f1f1296033aa2eee0fe385.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "type": "text", + "content": "Figure 2. PALM approximation of AL curves on CIFAR-10 (" + }, + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "type": "inline_equation", + "content": "b = 20" + }, + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "type": "text", + "content": ") and CIFAR-100 (" + }, + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "type": "inline_equation", + "content": "b = 100" + }, + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "type": "text", + "content": "). Each plot displays the accuracy trajectories of multiple AL strategies fitted with the PALM model across the entire annotation budget. Two distinct behaviors emerge in both datasets: methods with gradual accuracy growth and those utilizing pretrained embeddings that quickly reach maximum accuracy. These differences are reflected in the fitted PALM parameters. Methods using pretrained embeddings consistently show higher " + }, + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "type": "text", + "content": " values, indicating greater sample efficiency, while non-embedding methods exhibit higher " + }, + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "type": "text", + "content": " values, indicating delayed learning progress. Despite the varying complexity of the datasets, PALM successfully models the learning dynamics and highlights clear differences between AL strategies. Note that the highest test accuracy for CIFAR-100 is approximately " + }, + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 54, + 213, + 555, + 291 + ], + "type": "text", + "content": ". All quantitative results are available in the Supplementary Materials." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 310, + 76, + 493, + 204 + ], + "blocks": [ + { + "bbox": [ + 310, + 76, + 493, + 204 + ], + "lines": [ + { + "bbox": [ + 310, + 76, + 493, + 204 + ], + "spans": [ + { + "bbox": [ + 310, + 76, + 493, + 204 + ], + "type": "image", + "image_path": "2b63ce5efe75411643b4f96585d9e91ea9823cd9a5d2a8c33bb26b2e93d3beaf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 312, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 295, + 396 + ], + "type": "text", + "content": "sampling without embeddings yields " + }, + { + "bbox": [ + 54, + 312, + 295, + 396 + ], + "type": "inline_equation", + "content": "\\delta = 0.094" + }, + { + "bbox": [ + 54, + 312, + 295, + 396 + ], + "type": "text", + "content": ", whereas its embedding-based counterpart achieves " + }, + { + "bbox": [ + 54, + 312, + 295, + 396 + ], + "type": "inline_equation", + "content": "\\delta = 0.535" + }, + { + "bbox": [ + 54, + 312, + 295, + 396 + ], + "type": "text", + "content": ". Similarly, on CIFAR-100, Random sampling improves from " + }, + { + "bbox": [ + 54, + 312, + 295, + 396 + ], + "type": "inline_equation", + "content": "\\delta = 0.048" + }, + { + "bbox": [ + 54, + 312, + 295, + 396 + ], + "type": "text", + "content": " (without embeddings) to " + }, + { + "bbox": [ + 54, + 312, + 295, + 396 + ], + "type": "inline_equation", + "content": "\\delta = 0.318" + }, + { + "bbox": [ + 54, + 312, + 295, + 396 + ], + "type": "text", + "content": " (with embeddings). These indicate that pretrained embeddings enhance the representational quality of selected samples, thereby improving space coverage and annotation efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": "The parameter " + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": " controls the initial offset of learning progress. Higher " + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": " values indicate a delayed onset of learning, a trend that is especially pronounced in non-embedding methods. For example, on CIFAR-100, Margin sampling without embeddings exhibits " + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\alpha = 10.643" + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": ", whereas its embedding-based counterpart reduces " + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": " to 0.068, indicating faster initial learning when using SSL-derived representations. Moreover, the parameter " + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": " captures the scaling of accuracy gains for the annotation budget. Larger " + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": " values correspond to steeper growth in accuracy. On CIFAR-100, Entropy sampling with embeddings demonstrates " + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "inline_equation", + "content": "\\beta = 0.651" + }, + { + "bbox": [ + 54, + 399, + 295, + 555 + ], + "type": "text", + "content": ", the highest among the evaluated strategies, indicating rapid convergence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "content": "Despite similar final accuracies across methods, the PALM parameters offer a fine-grained view of sample efficiency and learning behavior. For example, while Random sampling with embeddings achieves a high " + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\delta = 0.536" + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "content": ", Entropy sampling without embeddings is limited to " + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\delta = 0.066" + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "content": ", despite achieving comparable asymptotic performance. This illustrates that PALM not only captures endpoint accuracy but also characterizes the trajectory of learning dynamics. These findings empirically validate Eq. (13), where " + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 54, + 558, + 295, + 713 + ], + "type": "text", + "content": " jointly describe the efficiency, onset, and scaling of active learning progress. The clear distinction in behaviors across AL strategies, as revealed by the fitted parameters, confirms that PALM provides a consistent and in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 312, + 553, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 348 + ], + "type": "text", + "content": "terpretable model of AL dynamics across different datasets, methods, and embedding strategies. Full quantitative results are provided in the Supplementary Materials." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 350, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 518 + ], + "type": "text", + "content": "While PALM accurately fits the learning trajectories across AL strategies, we observe minor deviations in the later rounds of the CIFAR-100 experiments. This is likely due to the progressive exhaustion of informative examples in the unlabeled pool. As AL progresses, the remaining samples tend to be noisier or less representative, which increases prediction uncertainty and affects fit stability. Although PALM assumes an effectively infinite unlabeled pool, this simplification may lead to small estimation drift once the pool is nearly depleted. Nevertheless, the observed deviations remain within " + }, + { + "bbox": [ + 313, + 350, + 554, + 518 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 350, + 554, + 518 + ], + "type": "text", + "content": " of test accuracy across all budget levels. Importantly, PALM is designed to support early-stage decision-making, where accurate prediction from limited partial observations is most impactful." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": "Limited Budgets: To further assess the predictive capability of PALM, we evaluated its performance when fitted to partial learning curves, using only a limited number of annotated samples across different AL methods and datasets. The results are presented in Figure 3. For CIFAR-10, an accurate approximation of the full learning dynamics was achieved with as few as 1,000 labeled samples, representing just " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": " of the dataset. Notably, for some strategies, high-quality predictions emerged with only 200-300 labeled samples, corresponding to approximately " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "0.8 - 1.5\\%" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": " of the dataset. Methods leveraging pretrained embeddings required even fewer samples to achieve strong fits, further highlighting the efficiency gains provided by SSL-based representations. Additionally, for CIFAR-100, a more complex dataset with 100 classes, reliable approximations were obtained using 5,000-10,000 labeled samples. This" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4044" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 68, + 511, + 265 + ], + "blocks": [ + { + "bbox": [ + 100, + 68, + 511, + 265 + ], + "lines": [ + { + "bbox": [ + 100, + 68, + 511, + 265 + ], + "spans": [ + { + "bbox": [ + 100, + 68, + 511, + 265 + ], + "type": "image", + "image_path": "5ca44b5c6d092850b4861acb11b47b121c94fd70a75afbb08a642b3429bc6954.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 274, + 555, + 341 + ], + "lines": [ + { + "bbox": [ + 54, + 274, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 555, + 341 + ], + "type": "text", + "content": "Figure 3. PALM predictions of AL training curves fitted using varying numbers of cumulative budget points, as indicated in the legend. Each plot compares the predicted trajectories with the actual performance across different AL strategies on CIFAR-10 and CIFAR-100. Note that the strategies in rows 2 and 4 utilize embeddings extracted using the SimCLR method. We find that accurate predictions of the full learning dynamics can be made with cumulative budgets of approximately 1,000 samples for CIFAR-10 (2% of the dataset) and 5,000 samples for CIFAR-100 (10% of the dataset). In certain cases on CIFAR-10, reasonable predictions can be made with as few as 400 samples (0.8% of the dataset), demonstrating PALM's ability to model learning behavior even with limited early-stage supervision." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 361, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 295, + 397 + ], + "type": "text", + "content": "translates to approximately 50-100 labeled examples per class, demonstrating that PALM can generalize effectively to higher-dimensional spaces with greater class diversity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 399, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 399, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 399, + 295, + 567 + ], + "type": "text", + "content": "TypiClust, a method specifically designed for small-budget scenarios and reliant on pretrained embeddings, demonstrated particularly strong approximation with extremely limited supervision. On CIFAR-100, PALM accurately modeled TypiClust's learning trajectory using just 1,000 labeled samples (roughly 10 examples per class), underscoring its effectiveness under resource-constrained conditions. The use of SSL-based embeddings significantly reduced the computational cost of AL iterations, offering practical advantages for large-scale or real-time annotation workflows. This finding suggests that embedding-based AL pipelines not only enhance model performance but also improve training efficiency, making them a compelling choice for the design of future AL systems." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 568, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 568, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 568, + 295, + 664 + ], + "type": "text", + "content": "It is worth noting that for CIFAR-10, PALM occasionally struggles to fit the sharp knee point of the learning curve due to the dataset's relatively simple structure and the rapid accuracy gains observed in early iterations. In contrast, this effect is less pronounced on CIFAR-100, where the increased complexity and class imbalance align more closely with real-world scenarios, enabling smoother approximations across the full budget range." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 295, + 713 + ], + "type": "text", + "content": "Large-Scale Datasets: Furthermore, to assess the generalization capability of PALM on large-scale, natural datasets, we conducted experiments on ImageNet-50, ImageNet-100, and ImageNet-200 using a total annotation budget corre" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 361, + 553, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 553, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 553, + 505 + ], + "type": "text", + "content": "sponding to " + }, + { + "bbox": [ + 313, + 361, + 553, + 505 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 361, + 553, + 505 + ], + "type": "text", + "content": " of the training samples. Figure 4 presents the results of this experiment. As seen, PALM successfully approximated the learning dynamics across all tested active learning strategies and SSL-based embeddings, reinforcing its applicability to realistic, resource-constrained scenarios. Beyond capturing overall learning trajectories, PALM effectively models distinct learning behaviors, including methods that achieve rapid initial improvements and those with delayed performance onset. These differences are reflected in the fitted parameters, particularly " + }, + { + "bbox": [ + 313, + 361, + 553, + 505 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 361, + 553, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 361, + 553, + 505 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 361, + 553, + 505 + ], + "type": "text", + "content": ", which control the starting point and coverage efficiency of the learning process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 507, + 554, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 554, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 554, + 663 + ], + "type": "text", + "content": "A particularly notable case arises with the BYOL embeddings. Due to BYOL's extremely slow and nearly linear accuracy progression within the small budget, PALM fits the curve as if the method has already plateaued. This behavior occurs because the early-stage dynamics are not sufficiently captured with the limited number of labeled samples. However, signs of acceleration appear in methods like TypiClust, where the final iterations show an upward trend, suggesting that with larger annotation budgets, BYOL would eventually follow the growth patterns predicted by our mathematical model. This observation underscores the need for more labeled samples to estimate future performance for slow-starting methods reliably." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Differences in " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " across AL strategies also reveal important scalability trends, with higher " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " values indicating steeper accuracy gains following the initial labeling phase. Notably, strategies combined with MoCov3 consistently ex" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4045" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 68, + 509, + 224 + ], + "blocks": [ + { + "bbox": [ + 102, + 68, + 509, + 224 + ], + "lines": [ + { + "bbox": [ + 102, + 68, + 509, + 224 + ], + "spans": [ + { + "bbox": [ + 102, + 68, + 509, + 224 + ], + "type": "image", + "image_path": "7b625e2c3e82669574c7ed45ed7741f3273d7ef2316b5cc7947f2779f030dd61.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 234, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 555, + 313 + ], + "type": "text", + "content": "Figure 4. PALM predictions of AL curves on ImageNet subsets. Each plot compares the performance of different AL strategies combined with various SSL methods (MoCov2+, MoCov3, BYOL, and SimCLR). PALM accurately models learning dynamics under extreme annotation constraints, capturing distinct behaviors across methods and embeddings. MoCov3 consistently achieves the highest achievable accuracy " + }, + { + "bbox": [ + 55, + 234, + 555, + 313 + ], + "type": "inline_equation", + "content": "(\\mathrm{A}_{\\mathrm{max}})" + }, + { + "bbox": [ + 55, + 234, + 555, + 313 + ], + "type": "text", + "content": " and coverage efficiency " + }, + { + "bbox": [ + 55, + 234, + 555, + 313 + ], + "type": "inline_equation", + "content": "(\\delta)" + }, + { + "bbox": [ + 55, + 234, + 555, + 313 + ], + "type": "text", + "content": ", indicating superior representation quality and faster learning. In contrast, BYOL exhibits slow, nearly linear accuracy growth with minimal coverage, indicating delayed learning that requires larger annotation budgets. SimCLR and MoCov2+ show varying dynamics depending on the AL strategy, underscoring the sensitivity of embedding performance to sample selection methods. These findings highlight PALM's ability to model and compare AL behavior across diverse settings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 331, + 295, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 367 + ], + "type": "text", + "content": "hibit higher " + }, + { + "bbox": [ + 55, + 331, + 295, + 367 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 331, + 295, + 367 + ], + "type": "text", + "content": ", reflecting their ability to capitalize on accumulated labeled data more effectively. A closer analysis of the ImageNet-200 results further illustrates these insights:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 295, + 582 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 369, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 415 + ], + "type": "text", + "content": "- MoCov3 consistently achieves the highest " + }, + { + "bbox": [ + 55, + 369, + 295, + 415 + ], + "type": "inline_equation", + "content": "A_{\\max}" + }, + { + "bbox": [ + 55, + 369, + 295, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 369, + 295, + 415 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 369, + 295, + 415 + ], + "type": "text", + "content": " values across all strategies, indicating superior representation quality and faster learning progression, making it highly effective under limited supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 416, + 295, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 295, + 463 + ], + "type": "text", + "content": "- SimCLR shows competitive " + }, + { + "bbox": [ + 55, + 416, + 295, + 463 + ], + "type": "inline_equation", + "content": "\\mathrm{A}_{\\max}" + }, + { + "bbox": [ + 55, + 416, + 295, + 463 + ], + "type": "text", + "content": " but lower " + }, + { + "bbox": [ + 55, + 416, + 295, + 463 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 416, + 295, + 463 + ], + "type": "text", + "content": ", suggesting slower per-sample gains and greater sensitivity to the choice of active learning strategy, as evidenced by the variability of " + }, + { + "bbox": [ + 55, + 416, + 295, + 463 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 416, + 295, + 463 + ], + "type": "text", + "content": " across methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 464, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 295, + 534 + ], + "type": "text", + "content": "- BYOL displays consistently low " + }, + { + "bbox": [ + 55, + 464, + 295, + 534 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 464, + 295, + 534 + ], + "type": "text", + "content": " values and, in some cases, near-zero values (e.g., with TypiClust), reflecting minimal coverage per labeled sample and delayed learning. For TypiClust specifically, a high " + }, + { + "bbox": [ + 55, + 464, + 295, + 534 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 464, + 295, + 534 + ], + "type": "text", + "content": " and steep " + }, + { + "bbox": [ + 55, + 464, + 295, + 534 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 464, + 295, + 534 + ], + "type": "text", + "content": " indicate a late but accelerating learning phase, aligning with the observed rise in the final iterations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 536, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 295, + 582 + ], + "type": "text", + "content": "- MoCov2+ generally lags behind the other embeddings, with lower " + }, + { + "bbox": [ + 55, + 536, + 295, + 582 + ], + "type": "inline_equation", + "content": "A_{\\max}" + }, + { + "bbox": [ + 55, + 536, + 295, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 536, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 536, + 295, + 582 + ], + "type": "text", + "content": " across strategies, further highlighting the varying effectiveness of SSL embeddings in active learning pipelines." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 584, + 295, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 295, + 656 + ], + "type": "text", + "content": "The findings show that PALM not only captures overall performance trends but also uncovers nuanced interactions between SSL embeddings and AL strategies. It offers a comprehensive model for comparing methods, balancing trade-offs between early performance and long-term scalability, and guiding annotation planning in practical applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 669, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 128, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "We presented PALM, a unified mathematical model for the performance analysis of active learning models by captur" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "type": "text", + "content": "ing the dynamics of AL curves through a compact set of interpretable parameters. Unlike conventional evaluation protocols that compare methods based solely on accuracy at fixed budgets, PALM provides a principled approach to understanding how active learning strategies behave throughout the entire annotation process. By modeling coverage efficiency " + }, + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "type": "inline_equation", + "content": "(\\delta)" + }, + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "type": "text", + "content": ", sampling behavior " + }, + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "type": "inline_equation", + "content": "(b)" + }, + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "type": "text", + "content": ", early-stage performance " + }, + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "type": "inline_equation", + "content": "(\\alpha)" + }, + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "type": "text", + "content": ", scalability " + }, + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "type": "inline_equation", + "content": "(\\beta)" + }, + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "type": "text", + "content": ", and achievable accuracy " + }, + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "type": "inline_equation", + "content": "(A_{\\max})" + }, + { + "bbox": [ + 313, + 331, + 555, + 451 + ], + "type": "text", + "content": ", PALM offers a comprehensive characterization of active learning trajectories from limited observations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 455, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 586 + ], + "type": "text", + "content": "Our experiments on CIFAR and ImageNet benchmarks, covering a diverse range of AL strategies and SSL embeddings, demonstrate that PALM generalizes across datasets, methods, and budget regimes despite using only a small fraction of labeled data. It not only accurately predicts learning curves under extreme annotation constraints but also reveals key differences in how methods scale, where they accelerate, and how embedding quality interacts with sampling strategies. These insights expose the limitations of relying on final accuracy alone and establish PALM as a fair model for cost-aware comparison of AL methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 590, + 556, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 556, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 556, + 686 + ], + "type": "text", + "content": "PALM advances AL by shifting the focus from isolated performance measurements to predictive modeling of the learning process. This enables more informed, data-efficient decision-making in real-world settings, where annotation costs dominate and strategy selection is critical. By accounting for the dynamics and learning efficiency, PALM lays the basis for a new standard in AL evaluation, paving the way toward more systematic and scalable AL design." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgements. This project has received funding from the Pioneer Centre for AI, DNRF, grant number P1." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4046" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 157 + ], + "type": "text", + "content": "[1] Amin Aghaee, Mehrdad Ghadiri, and Mahdieh Soleymani Baghshah. Active distance-based clustering using k-medoids. In Advances in Knowledge Discovery and Data Mining: 20th Pacific-Asia Conference, PAKDD 2016, Auckland, New Zealand, April 19-22, 2016, Proceedings, Part I 20, pages 253-264. Springer, 2016. 1, 11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 159, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 203 + ], + "type": "text", + "content": "[2] Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. Deep batch active learning by diverse, uncertain gradient lower bounds. arXiv preprint arXiv:1906.03671, 2019. 2, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "text", + "content": "[3] William H Beluch, Tim Genewein, Andreas Nurnberger, and Jan M Kohler. The power of ensembles for active learning in image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9368-9377, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 260, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 295, + 316 + ], + "type": "text", + "content": "[4] Jiayi Chen, Benteng Ma, Hengfei Cui, and Yong Xia. Think twice before selection: Federated evidential active learning for medical image analysis with domain shifts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11439-11449, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 317, + 295, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 295, + 361 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PmLR, 2020. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 362, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 294, + 396 + ], + "type": "text", + "content": "[6] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 396, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 295, + 441 + ], + "type": "text", + "content": "[7] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640-9649, 2021. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 442, + 295, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 295, + 496 + ], + "type": "text", + "content": "[8] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 498, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 295, + 543 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 588 + ], + "type": "text", + "content": "[10] Yarin Gal, Riashat Islam, and Zoubin Ghahramani. Deep bayesian active learning with image data. In International conference on machine learning, pages 1183-1192. PMLR, 2017. 5, 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 612 + ], + "type": "text", + "content": "[11] Daniel Gissin and Shai Shalev-Shwartz. Discriminative active learning. arXiv preprint arXiv:1907.06347, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 668 + ], + "type": "text", + "content": "[12] Ross Greer, Björk Antoniussen, Mathias V Andersen, Andreas Møgelmose, and Mohan M Trivedi. The why, when, and how to use active learning in large-data-driven 3d object detection for safe autonomous driving: An empirical exploration. arXiv preprint arXiv:2401.16634, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 714 + ], + "type": "text", + "content": "[13] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[14] Guy Hacohen, Avihu Dekel, and Daphna Weinshall. Active learning on a budget: Opposite strategies suit high and low budgets. arXiv preprint arXiv:2202.02794, 2022. 2, 5, 11" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 205 + ], + "type": "text", + "content": "[16] Rong Hu, Brian Mac Namee, and Sarah Jane Delany. Off to a good start: Using clustering to select the initial training set in active learning. In FLAIRS, 2010. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 227 + ], + "type": "text", + "content": "[17] Svante Janson. Random coverings in several dimensions. 1986. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[18] Qiuye Jin, Mingzhi Yuan, Qin Qiao, and Zhijian Song. One-shot active learning for image segmentation via contrastive learning and diversity-based sampling. Knowledge-Based Systems, 241:108278, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "text", + "content": "[19] Seho Kee, Enrique Del Castillo, and George Runger. Query-by-committee improvement with diversity and density in batch active learning. Information Sciences, 454:401-418, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 317, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 338 + ], + "type": "text", + "content": "[20] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images, 2009. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 339, + 553, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 372 + ], + "type": "text", + "content": "[21] David D Lewis. A sequential algorithm for training text classifiers: Corrigendum and additional data. In Acm Sigir Forum, pages 13-19. ACM New York, NY, USA, 1995. 2, 11" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 373, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 415 + ], + "type": "text", + "content": "[22] Hailun Liang, Sunyuan Qiang, Hui Ma, Jun Wan, and Yanyan Liang. Semantic segmentation active learning with scene coverage coreset. In Chinese Conference on Biometric Recognition, pages 238-247. Springer, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "type": "text", + "content": "[23] Jinpeng Lin, Zhihao Liang, Shengheng Deng, Lile Cai, Tao Jiang, Tianrui Li, Kui Jia, and Xun Xu. Exploring diversity-based active learning for 3d object detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[24] Alaa Maalouf, Gilad Eini, Ben Mussay, Dan Feldman, and Margarita Osadchy. A unified approach to coreset learning. IEEE Transactions on Neural Networks and Learning Systems, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 553, + 548 + ], + "type": "text", + "content": "[25] Inbal Mishal and Daphna Weinshall. Dcom: Active learning for all learners. arXiv preprint arXiv:2407.01804, 2024. 2, 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 548, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 612 + ], + "type": "text", + "content": "[26] Sudhanshu Mittal, Joshua Niemeijer, Özgün Çiçek, Maxim Tatarchenko, Jan Ehrhardt, Jörg P Schäfer, Heinz Handels, and Thomas Brox. Realistic evaluation of deep active learning for image classification and semantic segmentation. International Journal of Computer Vision, pages 1-23, 2025. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 614, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 669 + ], + "type": "text", + "content": "[27] Prateek Munjal, Nasir Hayat, Munawar Hayat, Jamshid Sourati, and Shadab Khan. Towards robust and reproducible active learning using neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 223-232, 2022. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[28] Vu-Linh Nguyen, Mohammad Hossein Shaker, and Eyke Hüllermeier. How to measure uncertainty in uncertainty sampling for active learning. Machine Learning, 111(1):89-122, 2022. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4047" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 597 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 94 + ], + "type": "text", + "content": "[29] Sheldon M Ross. Introduction to probability models. Academic press, 2014. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 95, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 95, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 95, + 295, + 140 + ], + "type": "text", + "content": "[30] Tobias Scheffer, Christian Decomain, and Stefan Wrobel. Active hidden markov models for information extraction. In International symposium on intelligent data analysis, pages 309-318. Springer, 2001. 2, 11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 141, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 141, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 57, + 141, + 295, + 173 + ], + "type": "text", + "content": "[31] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. arXiv preprint arXiv:1708.00489, 2017. 2, 11" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 294, + 186 + ], + "type": "text", + "content": "[32] Burr Settles. Active learning literature survey. 2009. 1, 2, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 187, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 187, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 57, + 187, + 295, + 239 + ], + "type": "text", + "content": "[33] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, and Luc Van Gool. Scan: Learning to classify images without labels. In European conference on computer vision, pages 268-285. Springer, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 243, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 243, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 57, + 243, + 294, + 285 + ], + "type": "text", + "content": "[34] Konstantin Voevodski, Maria-Florina Balcan, Heiko Röglin, Shang-Hua Teng, and Yu Xia. Active clustering of biological sequences. The Journal of Machine Learning Research, 13 (1):203–225, 2012. 1, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 288, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 288, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 57, + 288, + 294, + 331 + ], + "type": "text", + "content": "[35] Haoran Wang, Qiuye Jin, Shiman Li, Siyu Liu, Manning Wang, and Zhijian Song. A comprehensive survey on deep active learning in medical image analysis. Medical Image Analysis, page 103201, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 333, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 333, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 57, + 333, + 294, + 387 + ], + "type": "text", + "content": "[36] Ji Wu, Zhongfeng Kang, Sebastian Nørgaard Llambias, Mostafa Mehdipour Ghazi, and Mads Nielsen. Active transfer learning for 3d hippocampus segmentation. In Workshop on Medical Image Learning with Limited and Noisy Data, pages 224-234. Springer, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 388, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 388, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 57, + 388, + 294, + 453 + ], + "type": "text", + "content": "[37] Zhao Xu, Kai Yu, Volker Tresp, Xiaowei Xu, and Jizhi Wang. Representative sampling for text classification using support vector machines. In Advances in Information Retrieval: 25th European Conference on IR Research, ECIR 2003, Pisa, Italy, April 14-16, 2003. Proceedings 25, pages 393-407. Springer, 2003. 1, 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 456, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 456, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 57, + 456, + 294, + 497 + ], + "type": "text", + "content": "[38] Ofer Yehuda, Avihu Dekel, Guy Hacohen, and Daphna Weinshall. Active learning through a covering lens. Advances in Neural Information Processing Systems, 35: 22354-22367, 2022. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 500, + 294, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 500, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 57, + 500, + 294, + 532 + ], + "type": "text", + "content": "[39] Mohammed Zakariah and Abdulaziz S Almazyad. Anomaly detection for IoT systems using active learning. Applied Sciences, 13(21):12029, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 534, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 534, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 534, + 294, + 597 + ], + "type": "text", + "content": "[40] Jifan Zhang, Yifang Chen, Gregory Canal, Stephen Mussmann, Arnav M Das, Gantavya Bhatt, Yinglun Zhu, Jeffrey Bilmes, Simon Shaolei Du, Kevin Jamieson, et al. Labelbench: A comprehensive framework for benchmarking adaptive label-efficient learning. arXiv preprint arXiv:2306.09910, 2023. 1, 11" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4048" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/93b627e5-bb27-4dc6-aa8c-e29320f5f5eb_content_list.json b/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/93b627e5-bb27-4dc6-aa8c-e29320f5f5eb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b2188320c9ea9897adc3ad2501e73ae29c28b751 --- /dev/null +++ b/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/93b627e5-bb27-4dc6-aa8c-e29320f5f5eb_content_list.json @@ -0,0 +1,1542 @@ +[ + { + "type": "text", + "text": "ToF-Splatting: Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration", + "text_level": 1, + "bbox": [ + 176, + 128, + 821, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Andrea Conti $^{1,2*}$", + "bbox": [ + 222, + 202, + 362, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matteo Poggi", + "bbox": [ + 401, + 203, + 524, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Valerio Cambareri", + "bbox": [ + 571, + 203, + 730, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Martin R. Oswald", + "bbox": [ + 297, + 220, + 454, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stefano Mattoccia", + "bbox": [ + 501, + 220, + 656, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Bologna, Italy 2Sony DepthSensing Solutions, Belgium 3University of Amsterdam, Netherlands", + "bbox": [ + 161, + 239, + 828, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/154d8347800c509697f43c2698b4eac191b07da985beb92b68118ca4783670fb.jpg", + "image_caption": [ + "Figure 1. Overview of our ToF-Splatting method. Our method combines sparse ToF depth, multi-view geometry from a buffer of keyframes, and monocular cues (left) to perform into a unique end-to-end dense SLAM framework enabled by a Gaussian Splatting." + ], + "image_footnote": [], + "bbox": [ + 133, + 290, + 867, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 502, + 326, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Time-of-Flight (ToF) sensors provide efficient active depth sensing at relatively low power budgets; among such designs, only very sparse measurements from low-resolution sensors are considered to meet the increasingly limited power constraints of mobile and AR/VR devices. However, such extreme sparsity levels limit the seamless usage of ToF depth in SLAM. In this work, we propose ToF-Splatting, the first 3D Gaussian Splatting-based SLAM pipeline tailored for using effectively very sparse ToF input data. Our approach improves upon the state of the art by introducing a multi-frame integration module, which produces dense depth maps by merging cues from extremely sparse ToF depth, monocular color, and multi-view geometry. Extensive experiments on both real and synthetic sparse ToF datasets demonstrate the advantages of our approach, as it achieves state-of-the-art tracking and mapping performances on reference datasets.", + "bbox": [ + 88, + 534, + 485, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 820, + 220, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simultaneous Localization and Mapping (SLAM) [6, 8, 36, 43] entails the joint estimation of the current camera pose", + "bbox": [ + 89, + 845, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "against a 3D scene representation (i.e., the map) and the update of this representation with data from the current camera view. It is a fundamental task with applications in AR/VR, robotics, and autonomous navigation [14, 18]. In recent years, SLAM witnessed a revolution [33] following the success of Neural Radiance Fields (NeRF) [23] and 3D Gaussian Splatting (3DGS) [13] as scene representations from the adjacent field of novel view synthesis.", + "bbox": [ + 511, + 503, + 906, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper we focus on SLAM pipelines that leverage RGB-D video inputs, i.e., where color images and dense depth maps are collected by synchronized, accurate color and depth sensors. Beyond global scale, depth sensors provide accurate local geometry that grants better tracking and mapping quality. A common choice for active depth sensing are Time-of-Flight (ToF) cameras. Despite the increased accuracy provided by reliable depth measurements, the power consumption and cost of such sensors constrain their adoption to high-end industrial or automotive applications. A recent trend to favor ToF sensor integration in handheld consumer devices is to reduce the specifications of ToF sensors to provide low-resolution but more reliable depth measurements. Among other examples, LiDAR (an instance of ToF) appears in consumer mobile phones [20] and is known to feature up to at most 576 dots (interpolated to $256 \\times 192$ resolution via depth completion). Other recent ToF sensors provide only up to 64 depth points but consume a re", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* work started while visiting University of Amsterdam.", + "bbox": [ + 114, + 886, + 408, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28344", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ported $200\\mathrm{mW}$ [19]. In such settings a SLAM pipeline can expect to receive, e.g., $8 \\times 8$ very noisy, very sparse depth measurements per frame. These specifications prevent their straightforward integration in SLAM systems, as low-resolution noisy depth is known to harm the accuracy of dense SLAM systems built with NeRF and not originally designed to deal with this scenario [30, 36, 46]. This is even more evident with the more recent 3DGS-based SLAM systems [22], where accurate and dense depth is paramount to achieve satisfactory results.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, a straightforward strategy [38] would consist in recovering dense and accurate depth maps using a single-image depth completion framework [25]. However, models trained for this task are known to generalize poorly across domains and across input depth noise levels, potentially affecting the downstream SLAM pipeline accuracy and yielding both drifted camera trajectory and inaccurate mapping. Therefore, we explore 3DGS as an explicit representation that allows for injecting domain-specific knowledge and that is easily extended to, e.g., dynamic scenes [21] and accurate large mesh reconstruction [2, 9]. On the one hand, we argue that depth completion alone is insufficient to allow sparse ToF depth usage in SLAM systems. On the other, we acknowledge that even sparse and noisy depth measurements can bootstrap SLAM if properly assisted with geometry. Accordingly, inspired by frame-to-frame SLAM systems relying on pre-trained trackers for localization, we believe that multi-frame integration fusing the noisy depth measurements with multiview geometry across a small set of local frames could supply the dense SLAM system with the supervision necessary to fill this gap.", + "bbox": [ + 91, + 244, + 483, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose ToF-Splatting, a novel 3DGS-based SLAM system enabling dense reconstruction from sparse ToF sensors. ToF-Splatting harmoniously alternates tracking and mapping, with each one benefiting from the other. The former is carried out through backpropagation during the optimization of the 3DGS model on color images; the latter exploits a pre-trained network [3] as a multi-frame integration module, combining sparse depth measurements with multi-view geometry according to a buffer of keyframes. For each of the keyframes, we use the camera pose estimated through tracking and providing dense depth maps to improve the supervision given to the 3DGS model itself. Figure 1 shows an overview of the setting in which ToF-Splatting operates, as well as the quality of the 3D reconstructions it yields.", + "bbox": [ + 89, + 580, + 482, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ToF-Splatting is tested on the real-world ZJUL5 dataset [19], where we measure tracking and mapping quality and find the superiority of our framework with respect to existing solutions built over sparse ToF sensors [38]. Moreover, standard benchmarks such as Replica [28] are also provided and confirm the effectiveness of our mapping strat", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "egy. Our main contributions are as follows:", + "bbox": [ + 511, + 90, + 808, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose ToF-Splatting, the first 3DGS-based SLAM system suited for sparse ToF sensors.", + "- We introduce a novel multi-frame integration method combining sparse depth, multi-view geometry, and monocular cues for robust noise and outlier handling.", + "- We establish a new state-of-the-art for dense SLAM systems built on top of sparse ToF depth sensors." + ], + "bbox": [ + 511, + 106, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 224, + 653, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We briefly review the literature relevant to our work, referring to [33] for a detailed overview of the latest advances.", + "bbox": [ + 511, + 251, + 903, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional SLAM. Traditional visual SLAM pipelines, or the more recent ones introducing deep modules [15, 31, 32], usually perform tracking in frame-to-frame fashion, based on the visual features extracted from past frames either using handcrafted feature extractors [24] or deep networks [31, 32]. These systems are usually defined by four main steps: tracking, mapping, global bundle adjustment, and loop closure. Among the different representations used by traditional SLAM frameworks, depth points [1, 24], surfels [27], and volumetric representations [6] allows for achieving globally consistent 3D reconstruction.", + "bbox": [ + 511, + 281, + 903, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF-based SLAM. The advent of NeRF in the field of novel view synthesis also reached adjacent research areas, as well as SLAM a few years later. NeRF-based SLAM systems estimate camera poses and reconstruct dense meshes by modeling scenes via MLPs. iMAP [30] and NICE-SLAM [45] are the first SLAM frameworks following this approach, with the latter introducing dense feature grids supporting the MLPs. Point-SLAM [26] and Loopy-SLAM [17] switched feature grids with neural point embeddings, allowing for higher flexibility and correcting or adjusting local maps after global optimization, yet with slow processing prohibiting deployment in robotics. ESLAM [11] and Co-SLAM [34] deploy tri-planes and hash grid embeddings, improving both processing speed and reconstruction accuracy, with PLGSLAM [7] further improving the representation capacity in large indoor scenes. GO-SLAM [44] and KN-SLAM [37] implement frame-to-frame systems thanks to external trackers, yielding a good trade-off between tracking/mapping accuracy and frame rate.", + "bbox": [ + 511, + 446, + 905, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3DGS-based SLAM. More recent advances in novel view synthesis brought to the development of new SLAM systems using 3D Gaussian Splatting [13] to represent the mapped scene. Several concurrent frameworks emerged built over 3DGS, such as MonoGS [22], GaussianSLAM [42], SplaTAM [12], and GS-SLAM [39]. Different from NeRF-based SLAM systems, these frameworks allow for higher interpretability and higher rendering speed.", + "bbox": [ + 511, + 734, + 903, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sparse ToF-based SLAM. Only lately, the use of more compact and energy-efficient depth sensors has gained attention in the SLAM research community. ToF-SLAM [38]", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28345", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d20e17a8a638c546de8098bf09d21117f7d3ac261c3cba4a4d1b227262971f0f.jpg", + "image_caption": [ + "Figure 2. ToF-Splatting Pipeline. Our method involves three main modules: a Tracking frontend estimating camera poses, a Multi-Frame Integration module that predicts dense depth maps from sparse ToF measurements and multi-view geometry, and a Mapping backend modeling the 3D scene representation via 3D Gaussian Splatting." + ], + "image_footnote": [], + "bbox": [ + 122, + 90, + 867, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "represents the only attempt to exploit \"lightweight\", sparse depth measurements within a dense SLAM system, aided by a depth completion model [16]. However, depth completion alone fails to deal with noise in the sparse input depth, a shortcoming we aim to specifically overcome within our ToF-Splatting framework, the first one based on 3DGS and suited for sparse ToF sensors.", + "bbox": [ + 88, + 325, + 483, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Pipeline Architecture", + "text_level": 1, + "bbox": [ + 89, + 455, + 294, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present the core components of our framework. ToF-Splatting is a frame-to-model SLAM system conceptually divided into three modules: $(i)$ a mapping backend, $(ii)$ a tracking frontend, and $(iii)$ a peculiar multi-frame integration module yielding dense depth maps based on a subset of the keyframe graph, the current RGB frame, and the current sparse ToF depth map.", + "bbox": [ + 88, + 484, + 482, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 provides an overview of our pipeline. Our framework is characterized by three heterogeneous modules that are tightly related and influence each other. The tracking module stands as the most influencing element in the pipeline, as it selects the keyframes used by both the mapping and multi-frame integration parts, moreover, it computes the poses used for multi-frame depth estimation. Nonetheless, it relies on the quality of the mapping step to perform correct ego-motion estimation and exploits the depth cues to properly behave when photometric information is lacking or ambiguous.", + "bbox": [ + 88, + 595, + 482, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The depth perception part is mainly sustained by the tracking step and highly influences the mapping part since it is used to seed new Gaussians and as supervision. Finally, the mapping phase not only creates a unique smooth representation but also defines the rendered opacity that is used for keyframe selection. These connections lead to a smoothly integrated framework that not only exploits a multi-frame backbone but more importantly, defines an effective way to use such information in the SLAM scenario.", + "bbox": [ + 88, + 763, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background: 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 511, + 325, + 830, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splitting [13] (3DGS) represents one of the latest advances in novel view synthesis, fitting a dynamically defined set of multivariate Gaussian distributions over the scene to enable image rendering from arbitrary viewpoints. Each Gaussian $G(\\mathbf{x})$ is parametrized by its mean $\\boldsymbol{\\mu} \\in \\mathbb{R}^3$ and covariance matrix $\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}$ , reading", + "bbox": [ + 511, + 347, + 906, + 439 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (\\mathbf {x}) = \\exp \\left(- \\frac {1}{2} (\\mathbf {x} - \\boldsymbol {\\mu}) ^ {\\top} \\boldsymbol {\\Sigma} ^ {- 1} (\\mathbf {x} - \\boldsymbol {\\mu})\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 444, + 906, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, opacity $o \\in \\mathbb{R}$ and RGB color $\\mathbf{c} \\in \\mathbb{R}^3$ are assigned to each Gaussian.", + "bbox": [ + 511, + 483, + 905, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set of images from different viewpoints all around the scene, the Gaussian parameters are optimized using the splatting technique [35]. With such a method, 3D Gaussians are projected through rasterization over the 2D image plane, containing a 2D Gaussian distribution", + "bbox": [ + 511, + 513, + 906, + 589 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} ^ {2 D} = \\pi \\left(\\mathbf {T} _ {w c} \\bar {\\boldsymbol {\\mu}}\\right) \\quad \\boldsymbol {\\Sigma} ^ {2 D} = \\mathbf {J} \\mathbf {R} _ {w c} \\boldsymbol {\\Sigma} \\mathbf {R} _ {w c} ^ {\\top} \\mathbf {J} ^ {\\top} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 594, + 903, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\pmb{\\mu}}$ is $\\pmb{\\mu}$ in homogeneous coordinates, $\\mathbf{T}_{wc} \\in \\mathbb{R}^{4 \\times 4}$ , $\\mathbf{R}_{wc} \\in \\mathbb{R}^{3 \\times 3}$ are respectively the world-to-camera transformation and its rotational component, $\\pi$ is the projection matrix and $\\mathbf{J} \\in \\mathbb{R}^{2 \\times 3}$ is the Jacobian of the projective transformation [49]. The color $C$ of a pixel is then given by the combination of the $M$ weighted Gaussians, i.e.,", + "bbox": [ + 511, + 619, + 905, + 710 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} = \\sum_ {i \\in M} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 715, + 903, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $\\alpha_{i} = o_{i}G(\\mathbf{x}_{i})$ . Since $\\pmb{\\Sigma}$ is positive semi-definite, it is parametrized with a diagonal scaling matrix $\\mathbf{S}$ and a rotation matrix $\\mathbf{R}$ , so that $\\pmb{\\Sigma} = \\mathbf{RSS}^{\\top}\\mathbf{R}^{\\top}$ . $\\mathbf{R}$ is internally represented with quaternions.", + "bbox": [ + 511, + 765, + 906, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Multi-Frame Integration", + "text_level": 1, + "bbox": [ + 511, + 833, + 741, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dense depth cues are pivotal for proper initialization of the Gaussians position $\\pmb{\\mu} \\in \\mathbb{R}^3$ , as well as to provide supervision to both tracking and mapping threads - as already", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28346", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "known in the literature [22, 47]. However, sparse ToF sensors alone are insufficient for this purpose, because of the very sparse and noisy depth measurements they provide (e.g., only 64 points for a sensor such as that used in [19]). To overcome this limitation, ToF-SLAM [19] exploits an auxiliary completion model, DELTAR [16], to recover dense depth maps out of ToF measurements. However, depth completion is highly dependent on the quality of the sparse depth in input, rapidly degrading the accuracy of the densified depth maps in the presence of noise.", + "bbox": [ + 88, + 90, + 480, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we follow a different path: inspired by frame-to-frame SLAM systems deploying an external tracker, we believe that multi-frame integration can compensate for the noise in the ToF measurements by exploiting both single-view and multi-view geometry cues on a set of local frames, given the camera poses estimated by the tracking frontend. We choose to extend the Depth on Demand (DoD) framework [3] for this purpose.", + "bbox": [ + 89, + 244, + 480, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-Frame Integration. DoD [3] integrates sparse depth and two-view stereo cues for dense depth prediction. Specifically – given a frame $\\mathbf{F}^k$ – it processes the associated RGB image $I_k$ (the target view), another RGB view $I_j$ from a different frame $\\mathbf{F}^j$ (a source view), the relative pose $P^{jk}$ between the two, and the sparse depth points $H^k$ to predict a dense metric depth map $D^k$ . Differently from monocular depth prediction, DoD is aware of the scene scale either from $H^k$ or $P^{jk}$ , simplifying the problem of integrating scale-inconsistent geometries in the SLAM frontend. However, DoD is limited to two-view processing: this may lead to suboptimal results in the SLAM scenario, where an abundant amount of keyframes largely overlap and cover the same portion of the scene. Thus, we extend [3] to multiview processing. This is achieved by observing that the predicted depth map $D^k$ depends on the source view $I_j$ through its relative pose $P^{jk}$ only. DoD iteratively updates $D^k$ for multiple iterations, exploiting this intrinsic property we can integrate multiple views $(I_j, P^{jk})$ using a different one at each iteration [4]. This approach allows smooth integration of multi-view cues avoiding order dependency. To select the source views $I_j$ , at each prediction we order the keyframes selected by the SLAM pipeline according to their relative pose similarity and select the first $N$ frames with baseline distance close to $b = 15\\mathrm{cm}$ to ensure enough parallax. The model is trained from scratch on ScanNetv2 [5] and following the protocol of [3].", + "bbox": [ + 91, + 369, + 483, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Monocular Cues Integration. As shown in [3], DoD heavily relies on both geometry and depth measurements, thus struggling at generalizing in challenging scenarios where the two lack. This may happen in cases where the tight field of view, noisy sparse points, and textureless surfaces hamper multi-view matching. To mitigate such limitations, we integrate explicit monocular cues into DoD by feeding its monocular encoder with $I_{k}$ and the normalized depth", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "map $\\tilde{D}^k\\in [0,1]^{H\\times W}$ obtained through the single-image depth estimator Depth Anything v2 [40]. This approach allows the injection of a robust bias, avoiding issues related to monocular estimation, e.g., slanted surfaces or noisy scale estimates, and delegating to DoD their handling. These cues are included while retraining on ScanNetv2 [5].", + "bbox": [ + 511, + 89, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Outlier Handling. Furthermore, the noisy depth measurements in input to DoD can severely affect its accuracy. Therefore, we explicitly focus on dealing with outliers in the input sparse depth, that may be perceived by ToF devices over specific surfaces. Such errors are particularly evident in cheap sensors and may largely prejudice the information supplied by the few points they measure. Purposely, we deploy a simple yet effective method to identify such outliers by exploiting multi-view geometry again. Whenever depth needs to be predicted for a new frame, DoD predicts a depth map $\\hat{D}^k$ without processing the ToF measurements. Then, the $\\ell_1$ error between the latter and the depth measured by the ToF sensor $H^k$ is computed and finally sparse measures with error higher than a given quantile $q$ are discarded, and depth is predicted again by DoD by also processing the filtered sparse depth $\\tilde{H}^k$ this time. This way, we exploit multi-view cues to extract a clean depth prior to filtering out inconsistent sparse depth measures, and then we integrate the remaining points to improve the original prediction. Even without processing the sparse depth, DoD still predicts metric depth, as the metric scale is enforced by the relative camera poses.", + "bbox": [ + 511, + 181, + 906, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tracking Frontend", + "text_level": 1, + "bbox": [ + 513, + 521, + 694, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The tracking part of ToF-Splatting estimates the ego-motion of the camera for each new frame $\\mathbf{F}^k$ and builds a keyframe graph $\\mathbf{G} = \\{\\mathbf{KF}^k\\} \\subset \\{\\mathbf{F}^k\\}$ containing a set of meaningful frames for mapping and multi-frame depth estimation.", + "bbox": [ + 511, + 542, + 903, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pose Estimation. We initialize each new frame $\\mathbf{F}^k$ of our frame-to-model tracker with the pose of the previous frame $\\mathbf{F}^{k - 1}$ and then minimizing the tracking loss with respect to the relative pose between $\\mathbf{F}^k$ and $\\mathbf{F}^{k - 1}$ , following [22]. The tracking loss consists of two terms, $L_{\\mathrm{rgb}}$ and $L_{\\mathrm{depth}}$ , respectively accounting for photometric and geometric errors", + "bbox": [ + 511, + 604, + 903, + 695 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t r a c k}} = \\lambda_ {\\text {t r a c k}} L _ {\\text {r g b}} + (1 - \\lambda_ {\\text {t r a c k}}) L _ {\\text {d e p t h}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 717, + 903, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {r g b}} = \\left\\| I _ {k} - \\hat {I} \\left(\\mathbf {F} ^ {k}\\right) \\right\\| _ {1} \\quad L _ {\\mathrm {d e p t h}} = \\left\\| D ^ {k} - \\hat {D} \\left(\\mathbf {F} ^ {k}\\right) \\right\\| _ {1} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 736, + 903, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{k}$ is the RGB image from $\\mathbf{F}^k$ , $\\hat{I}(\\cdot)$ and $\\hat{D}(\\cdot)$ denote the rendered RGB and Depth and, $D^{k}$ is the depth prediction produced by the multi-frame integration module.", + "bbox": [ + 511, + 763, + 903, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike other methods, the tracking process happens in two stages. Initially, for a fixed number of steps $\\eta_{\\mathrm{rgb}}$ only the photometric loss $L_{\\mathrm{rgb}}^T$ is optimized. Then, such a pose is used for multi-frame integration, as described in Sec. 3.2, to generate $D^{k}$ . Finally, both losses in Eq. (5) are used for $\\eta_{\\mathrm{rgbd}}^T$ further steps.", + "bbox": [ + 511, + 810, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28347", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Keyframe Selection Policy. We deploy a simple yet effective keyframe selection policy that considers i) novel content to be mapped and ii) instabilities in the already mapped areas. Unlike other methods using complex approaches – e.g., frustum intersection, point clouds and pose analysis [22, 41] – we observe that the rendered opacity for a novel frame contains low opacity values where either Gaussians are not present or unstable. The first case highlights the presence of novel areas of the environment to be mapped, the second happens due to the pruning procedure involved in the mapping step (see Sec. 3.4). We set a frame $\\mathbf{F}^k$ as a keyframe $\\mathbf{K}\\mathbf{F}^k$ if its novelty factor $\\nu^k$ is higher than a threshold $\\nu^{th}$ , where we define $\\nu^k$ as", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nu^ {k} = \\frac {\\sum_ {i \\in \\mathbf {K F} ^ {k}} \\mathbf {M} \\left(\\mathbf {K F} _ {i} ^ {k}\\right)}{\\sum_ {i \\in \\mathbf {K F} ^ {k}} \\left(1 - \\mathbf {M} \\left(\\mathbf {K F} _ {i} ^ {k}\\right)\\right)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 299, + 482, + 337 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} \\left(\\mathbf {K F} _ {i} ^ {k}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\hat {O} \\left(\\mathbf {K F} _ {i} ^ {k}\\right) < \\sigma \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 339, + 482, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\hat{O} (\\mathbf{K}\\mathbf{F}_i^k)$ being the rendered opacity at pixel $i$ , and $\\mathbf{M}(\\mathbf{K}\\mathbf{F}_i^k)$ a binary uncertainty map.", + "bbox": [ + 89, + 392, + 482, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This approach allows for skipping several frames when the camera moves slowly or focusing deeply on areas where the ToF-Splatting struggles the most, by mapping it consecutively. Nonetheless, we enforce mapping after skipping a certain number of frames to avoid mapping too few frames.", + "bbox": [ + 89, + 425, + 483, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Mapping Backend", + "text_level": 1, + "bbox": [ + 89, + 511, + 267, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Whenever a new keyframe $\\mathbf{K}\\mathbf{F}^k$ is identified, mapping is triggered. This phase seeks to embed the new frame in the global 3DGS model, involving the following two steps.", + "bbox": [ + 89, + 532, + 482, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialization. In this step, new Gaussians are seeded in the 3DGS model. To limit the size of the model and reduce outliers, only areas where the rendered uncertainty $\\mathbf{M}(\\mathbf{KF}^k)$ is high are seeded with new points, with the new Gaussians also being randomly downsampled by a constant factor. Each Gaussian is initialized using depth $D^{k}$ predicted in the multi-frame integration step, with color from the RGB frame $I_{k}$ , and scale initialized with a constant value. Unlike other 3DGS pipelines [22, 41], we prove this approach effective and particularly efficient, replacing any complex 3D heuristics to a simple use of rendered opacity $O(\\cdot)$ , already a side-product of the rasterization process.", + "bbox": [ + 89, + 579, + 482, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimization. Firstly, ToF-Splating smoothly integrates the new information in the global model and secondly tunes the existing representation. This is achieved by optimizing the new keyframe $\\mathbf{KF}^k$ and a random subset of 5 keyframes among the last $N$ keyframes $\\{\\mathbf{KF}^j \\mid j \\geq \\max(k - N, 0)\\}$ . These are optimized for a fixed number of steps $\\eta_{\\mathrm{rgbd}}^M$ . Moreover, we enforce a sampling rate of $60\\%$ for $\\mathbf{KF}^k$ , with the remaining $40\\%$ for the remaining $\\mathbf{KF}^j$ to avoid forgetting. During this step, gradients are also back-propagated to the", + "bbox": [ + 89, + 761, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "camera poses of the sampled keyframes, thus acting as a global bundle adjustment. Finally, we prune Gaussians with opacities lower than 0.5 every $\\eta_{\\mathrm{rgbd}}^{M} / 2$ steps. In addition to the terms in Eq. (5), we optimize structural similarity as in Eq. (8) and, following [2], the normals as in Eq. (9), i.e.,", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {d s s i m}} = \\frac {1 - \\operatorname {S S I M} \\left(I _ {k} , \\hat {I} \\left(\\mathbf {K F} ^ {k}\\right)\\right)}{2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 175, + 905, + 209 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {n o r m a l s}} = \\| N ^ {k} - \\hat {N} \\| _ {1} + \\overline {{\\mathbf {1} - \\langle N ^ {k} , \\hat {N} \\rangle}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 212, + 903, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with $\\hat{N} \\coloneqq \\bar{\\nabla}\\hat{D}(\\mathbf{K}\\mathbf{F}^k)$ , $N^k \\coloneqq \\bar{\\nabla}D^k$ denoting the normals estimated via 3D gradients $\\bar{\\nabla}$ for the rendered and multiframe predicted depth respectively, $\\langle \\cdot, \\cdot \\rangle$ their inner product, and $\\overline{\\cdot}$ the average. $L_{\\mathrm{normals}}$ is masked on edges to preserve sharpness, as identified by running a Sobel filter.", + "bbox": [ + 511, + 241, + 905, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we introduce a regularization loss $L_{\\mathrm{iso}}$ to penalize elongated Gaussians (i.e., promote their isotropy) that usually leads to rendering artifacts as already observed in [22]. Eq. (10) defines $L_{\\mathrm{iso}}$ , where $\\frac{\\mathrm{diag}(\\cdot)}{\\mathrm{diag}(\\mathbf{S}_j)}$ extracts the diagonal values from its argument, and $\\overline{\\mathrm{diag}(\\mathbf{S}_j)}$ is the average of the resulting vector, i.e., of the elements of the $j$ -th diagonal scale matrix $\\mathbf{S}_j$ as defined in Sec. 3.1. Thus,", + "bbox": [ + 511, + 319, + 905, + 426 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {i s o}} = \\frac {1}{| \\mathcal {G} |} \\sum_ {j} ^ {| \\mathcal {G} |} \\| \\operatorname {d i a g} \\left(\\mathbf {S} _ {j}\\right) - \\overline {{\\operatorname {d i a g} \\left(\\mathbf {S} _ {j}\\right)}} \\cdot \\mathbf {1} _ {3 \\times 1} \\| _ {1} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 435, + 903, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final mapping loss aggregates the aforementioned loss terms with corresponding weights as follows:", + "bbox": [ + 511, + 488, + 905, + 520 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\text {m a p}} = \\lambda_ {\\text {m a p}} \\left(\\lambda_ {\\text {v i s u a l}} L _ {\\text {r g b}} + (1 - \\lambda_ {\\text {v i s u a l}}) L _ {\\text {d s s i m}}\\right) \\tag {11} \\\\ + \\left(1 - \\lambda_ {\\text {m a p}}\\right) L _ {\\text {d e p t h}} + \\lambda_ {\\text {n o r m a l s}} L _ {\\text {n o r m a l s}} + \\lambda_ {\\text {i s o}} L _ {\\text {i s o}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 527, + 903, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 588, + 645, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section assesses both the tracking and mapping performance of ToF-Splatting on the following datasets.", + "bbox": [ + 511, + 613, + 905, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ZJUL5. ZJUL5 [19] is the only existing public-domain real dataset providing sparse depth data from a low-resolution, low-cost ToF sensor. A VL53L5CX ToF sensor is assembled on a calibrated rig with an Intel RealSense 435i to provide dense depth ground truth. This benchmark comprises 7 diverse indoor scene recordings. Ground truth 3D meshes and camera poses are obtained following the ScanNet protocol [5]. This sensor provides $8 \\times 8$ depth points by counting the number of photons returned in each discretized time range. Of the at most 64 depth points provided, many are extremely noisy and introduce large outliers. This challenging real-world dataset provides real measurement noise that is not accurately represented in mainstream datasets.", + "bbox": [ + 511, + 643, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TUM RGB-D. This is a real dataset [29] providing indoor sequence captures and widely used as benchmark for RGB-D SLAM pipelines. The depth stream provides dense VGA depth maps from a high-resolution Kinect v1 depth sensor.", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28348", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/58ab1526af8198a569a54680c797c14f720c058b2f470e00f67f9d64b1e9298a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodKitchenSofaOfficeReceptionLiving RoomOffice2Sofa2Avg.
KinectFusion [10]X0.1460.2090.157X0.3210.1250.192
iMAP [30]X1.6580.3380.6480.6790.3440.2140.647
NICE-SLAM [46]0.7450.1440.1550.2510.2890.2280.4210.319
BundleFusion [6]0.1760.1020.1350.101X0.1630.1200.132
ElasticFusion [36]0.2530.1100.0700.1930.5300.1210.1460.203
MonoGS [22]0.2310.0320.0410.0440.1530.0350.0730.087
ToF-SLAM [19]0.1130.0810.0560.1140.2000.1010.0850.107
ToF-Splatting (ours)0.0880.0240.0220.0410.1220.0290.0300.051
", + "bbox": [ + 91, + 88, + 906, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. ZJUL5 Tracking Results. We show here the tracking performance on the ZJUL5 dataset [19] dataset with the ATE RMSE [m] (↓) metric on the 8 sequences available where $\\times$ indicates lost tracking. ToF-Splatting provides the better performance by a good margin on each sequence, demonstrating the optimal capability of our proposal to track the camera ego-motion accurately.", + "bbox": [ + 89, + 215, + 906, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Replica. Replica [28] is a synthetic dataset providing extremely realistic indoor sequences. It provides dense depth maps, which we use at different densities to perform experiments that prove the robustness and generalization capabilities of ToF-Splatting in challenging scenarios.", + "bbox": [ + 88, + 273, + 480, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare our method with both learning-based [19, 30, 46] and traditional pipelines [6, 10, 36]. Following ToF-SLAM [19], baselines requiring RGB-D frames use DELTAR [16] to densify the sparse ToF measurements. Moreover, we adapt MonoGS [22] densifying depth with the state-of-the-art depth completion framework OGNI-DC [48] for a Gaussian Splatting SLAM baseline.", + "bbox": [ + 88, + 349, + 482, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. In all our experiments, we use $N = 4$ source views $I_{j}$ (i.e., keyframe buffer elements) for the multi-view integration module and a baseline $b = 15\\mathrm{cm}$ for selecting keyframes. We use a quantile $q = 0.75$ to filter outliers that occur frequently in the sparse ToF depth of the ZJUL5 dataset [19]. Concerning tracking, we use a novelty factor threshold of $\\nu_{th} = 0.1$ , $\\sigma = 0.98$ . The total tracking loss uses $\\lambda_{\\mathrm{track}} = 0.9$ , a total number of $\\eta_{\\mathrm{rgb}}^T = 30$ and $\\eta_{\\mathrm{rgbd}}^T = 70$ steps are performed at each iteration. For mapping, we set the loss weights $\\lambda_{\\mathrm{map}} = 60$ , $\\lambda_{\\mathrm{visual}} = 0.20$ , $\\lambda_{\\mathrm{normals}} = 0.01$ , $\\lambda_{\\mathrm{iso}} = 1.0$ and the number of steps per iteration $\\eta_{\\mathrm{rgbd}}^M = 60$ . We test on a single RTX3090.", + "bbox": [ + 88, + 455, + 482, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Comparison with the State of the Art", + "text_level": 1, + "bbox": [ + 89, + 650, + 413, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare ToF-Splatting with existing methods in terms of tracking and mapping performances.", + "bbox": [ + 88, + 672, + 482, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tracking. In Table 1, we report the tracking performance by ToF-Splatting and various baselines on each sequence of the ZJUL5 dataset [19]. Following common practice we use the Absolute Trajectory Error metric [29], which directly compares the global consistency of a trajectory. While various frameworks [6, 10, 30] fail on the most challenging scenes, ToF-Splatting delivers accurate and consistent tracking. This is evident from the good margin it gains against the baselines. In Figure 3, we show the 3D trajectories on a subset of the scenes, with error below $3\\mathrm{cm}$ .", + "bbox": [ + 88, + 703, + 482, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mapping. In Table 2, we report the mapping performance by ToF-Splatting compared with the other baselines. For each scene, we collect the predicted pose of each frame", + "bbox": [ + 88, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/80030511516d6a512ed1b075e61b8fe373f7acf9c8973d7621a95480ea3c4880.jpg", + "image_caption": [ + "Figure 3. Qualitative results on the ZJUL5 dataset [19]. We show meshes obtained by fusing rendered depth maps with TSDF and marching cubes (left), and 3D trajectories (right) on 3 scenes selected from the ZJUL5 dataset [19]." + ], + "image_footnote": [], + "bbox": [ + 521, + 276, + 898, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and render depth and color from 3DGS. Then, we perform TSDF integration and extract the final mesh through marching cubes. Following [19], the meshes are then evaluated by computing Accuracy, Completion, and F-score versus the ground truth meshes. Accuracy and Completion respectively evaluate the mean distance between each predicted vertex and the nearest ground truth one and vice versa. The F-score takes into account both Accuracy and Completion with an aggregate metric. ToF-Splating delivers consistent results for mapping, always achieving the highest accuracy and significantly exceeding the baselines on average and specifically on the Sofa, Office, and Reception scenes which account for five over seven sequences. [19] is the second-best method and the main competitor in mapping.", + "bbox": [ + 511, + 688, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28349", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cdfd93b0e71c93afe0b7289433b526fc6b3c3cbdb92c15b9b16910c58e33ba07.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMetricKitchenSofaOfficeReceptionLiving RoomOffice2Sofa2Avg.
KinectFusion [10]Accuracy ↓X0.1900.2110.261X0.2670.135-
Completion ↓X0.0480.0460.064X0.0780.064-
F-score ↑X0.2780.2880.285X0.2740.381-
ElasticFusion [36]Accuracy ↓0.0920.1350.0840.2970.1510.0960.1220.140
Completion ↓0.0650.0480.0820.3050.2160.1470.0470.130
F-score ↑0.5530.4200.5290.2740.3820.4160.4810.436
BundleFusion [6]Accuracy ↓0.1700.1000.1030.122X0.1210.123-
Completion ↓0.0880.0300.0380.057X0.2140.034-
F-score ↑0.3730.5710.4740.470X0.4420.527-
iMAP [30]Accuracy ↓X0.1350.2290.3650.2250.2330.139-
Completion ↓X0.0540.1030.2450.2910.1390.069-
F-score ↑X0.4450.3150.2380.1700.2550.416-
NICE-SLAM [46]Accuracy ↓0.3030.1190.1160.2160.1030.1560.4640.211
Completion ↓0.4560.0420.0700.1990.0890.1630.0450.152
F-score ↑0.2210.5540.4110.4020.4000.2730.4010.380
MonoGS [22]Accuracy ↓0.0900.0860.0830.0710.0690.0810.0890.081
Completion ↓0.2460.1220.1040.2380.1930.1570.1700.175
F-score ↑0.2900.2990.3460.3670.2590.3740.2890.318
ToF-SLAM [19]Accuracy ↓0.0810.0680.0670.0790.0780.1130.1210.087
Completion ↓0.0710.0410.0450.0620.1220.0850.0330.066
F-score ↑0.5590.6610.6460.6430.4960.5570.6560.604
ToF-Splatting (ours)Accuracy ↓0.0640.0310.0320.0410.0590.0340.0430.043
Completion ↓0.0950.0380.0290.0560.1310.0540.0460.064
F-score ↑0.5270.7910.8400.7100.3590.7790.6420.664
", + "bbox": [ + 94, + 88, + 906, + 366 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6a1aa086b182d275b084baf0c8482cd8fb43afe64b5c8d2d9c7747423b83125b.jpg", + "table_caption": [ + "Table 2. ZJUL5 Mapping Results. We perform the mapping evaluation on the ZJUL5 dataset (7 indoor sequences) and report results of three metrics including accuracy (Acc.), completion (Comp.), and F-score. The failure cases are marked as $\\times$ ." + ], + "table_footnote": [], + "table_body": "
Monocular CuesMulti-View CuesMetricKitchenSofaOfficeReceptionLiving RoomOffice2Sofa2Avg.
✓XATE ↓0.1030.0540.0460.0500.1100.0410.0550.066
F-score ↑0.2160.6750.5640.6450.2410.6110.5570.501
X✓ATE ↓0.2340.0240.0250.0370.1090.0340.0280.072
F-score ↑0.2840.7810.8260.7030.2530.7660.6830.614
✓✓ATE ↓0.0880.0240.0220.0410.1220.0290.0300.051
F-score ↑0.5270.7910.8400.7100.3590.7790.6420.664
", + "bbox": [ + 91, + 406, + 906, + 501 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5971744a6c15c7ee6b05e581c3e9f311589b5fbfdadcde0925b65731a6a81d1c.jpg", + "table_caption": [ + "Table 3. Multi-Frame Integration Ablation Study. Analysis of the impact of monocular and multi-view cues in the multi-frame integration step. Monocular cues are less impactful than multi-view ones, boosting performance in the challenging scenes where multi-view cues struggle most. Nonetheless, their combination yields the best results." + ], + "table_footnote": [], + "table_body": "
MethodLCInputDensityfr1/deskfr2/xyzfr3/officeAvg.
DROID-VOXRGB0.00%0.0520.1070.0730.077
MonoGSXRGB0.00%0.0380.0460.0350.040
ToF-SplattingXRGB-D0.02%0.0300.0220.0300.027
0.04%0.0270.0190.0210.022
ESLAMXRGB-D100%0.0250.0110.0240.020
Co-SLAMXRGB-D100%0.0240.0170.0240.022
CG-SLAMXRGB-D100%0.0240.0120.0240.020
RTG-SLAMXRGB-D100%0.0170.0040.0110.011
MonoGSXRGB-D100%0.0150.0140.0150.015
ORB-SLAM2✓RGB0.00%0.0190.0060.0240.016
ORB-SLAM2✓RGB-D100%0.0160.0400.0100.010
", + "bbox": [ + 93, + 563, + 480, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Results on TUM RGB-D dataset. Top: competitors w/o loop-closure (LC); middle: most representative competitor with $100\\%$ density; bottom: competitors w/ loop-closure.", + "bbox": [ + 89, + 704, + 482, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, the other methods – like [6, 36] – may provide good performance on specific scenes but demonstrate unreliability on average usually failing in the particularly challenging Kitchen and Living Room scenes. In Figure 3, we show Office, Reception, and Sofa meshes on the left.", + "bbox": [ + 89, + 755, + 482, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 845, + 256, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform additional experiments aimed at studying the impact of different factors on ToF-Splatting.", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-Frame Integration. In Table 3, we ablate the main components of the multi-frame integration and assess their contribution to the whole SLAM framework. By retaining monocular cues alone at the expense of multi-view geometry, we observe severe drops in both tracking and mapping accuracy, confirming the paramount importance of the latter; monocular cues alone are effective only where multiview cues are ineffective – e.g., in the absence of texture, as occurs mostly in the Kitchen scene.", + "bbox": [ + 511, + 566, + 906, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TUM RGB-D. In Table 4, we simulate sparse ToF data by sampling $0.02\\%$ and $0.04\\%$ depth points, and compare ToF-Splatting with existing RGB and RGB-D methods. With as few as $0.02\\%$ of points, ToF-Splatting largely improves over RGB methods [22, 32], approaching the $100\\%$ density RGB-D approach of the most representative competitor [22] at only $0.04\\%$ density.", + "bbox": [ + 511, + 703, + 908, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Replica: Input Depth Sparsity. We conduct additional studies on the Replica dataset, for which Figure 4 provides a qualitative overview of our 3D reconstructions. ToF-Splatting enables performing SLAM with very few sparse points – the 64 sparse points provided by a $8 \\times 8$ sparse account only for $0.02\\%$ of a $640 \\times 480$ image. We here", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28350", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/beff949188f76e61ce25e2f0d0d661966e91962c7328414bac13b1b9985dd5c4.jpg", + "image_caption": [ + "Figure 4. Replica Qualitatives. We provide qualitative results on Replica [28] to demonstrate the generalization capabilities of our method. On the left from top to bottom, meshes obtained respectively on scenes Office2 and Room2. On the right, details from the scene Room0. ToF-Splating delivers accurate details and allows for nice photometric and depth rendering." + ], + "image_footnote": [], + "bbox": [ + 140, + 90, + 861, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5507bdfda1e2142e9434ff9414b17a3166fbf7544d3334aefe737a3d70836de0.jpg", + "image_caption": [ + "Figure 5. Impact of depth sparsity. We test on Replica [28] with different simulated depth sparsity levels to assess the capability to exploit higher input densities. MAE and ATE smoothly decrease, whereas rendering metrics appear to be less affected." + ], + "image_footnote": [], + "bbox": [ + 116, + 305, + 460, + 406 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/24c5def5c78c0a1803aae4182b7e0931978cdc3bbe579554c5c536f42e892d08.jpg", + "image_caption": [ + "Figure 6. Impact of depth noise. We test on Replica [28] injecting different amounts of noise $\\xi \\in [0.00, 0.01, 0.05, 0.10]$ . ToF-Splatting demonstrates to be effective at dealing with noise, with error increasing almost linearly with the injected amount of noise." + ], + "image_footnote": [], + "bbox": [ + 116, + 470, + 460, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "explore the impact of higher sparse points density on its overall performance. We simulate on Replica densities of $\\{0.02\\%, 0.04\\%, 0.08\\%, 0.16\\% \\}$ corresponding to about $\\{64, 128, 256, 512\\}$ points. Figure 5 shows the performance variation on both tracking and mapping metrics using the aforementioned sparsities.", + "bbox": [ + 89, + 642, + 482, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Replica: Input Depth Noise. We study various noise levels impact in the input depth to assess our pipeline robustness. We model noise as additive but dependent on depth, injecting heteroscedastic Gaussian noise $\\mathcal{N}(d,\\epsilon d)$ where $\\epsilon$ modulates the variance of the noise source on Replica. Figure 6 shows that MAE, PSNR, and ATE all expose linear degradation trends as $\\epsilon$ increases.", + "bbox": [ + 89, + 733, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Runtime. In Table 5 we measure the runtime of ToF-Splatting in comparison with the main baselines. The tracking time of ToF-Splatting already includes depth estimates with DoD, with average runtime $\\bar{t}_{\\mathrm{DoD}} \\approx 90 ms$ (compris", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/213d997618fb32462fd6c9205e2a5dce00b4604d8e15988c0038993546e0dcab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method NameTracking TimeMapping Time
iMAP [30]101 ms448 ms
NICE-SLAM [46]470 ms1300 ms
MonoGS [22]1169 ms650 ms
ToF-SLAM [19]116 ms380 ms
ToF-Splatting (ours)614 ms865 ms
", + "bbox": [ + 514, + 303, + 905, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Runtime Comparison. We compare the runtime of ToF-Splatting and other learned baselines measuring their runtime for tracking and mapping. MonoGS [22] tracking time contains also the depth completion inference time [48].", + "bbox": [ + 511, + 391, + 906, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing DepthAnything v2 for monocular cues inference [40]). To date, 3DGS-based mapping methods are not capable of real-time performances. Nevertheless, exploiting the fact that the mapping step accounts only for just a few frames (typically $4\\%$ of the total in sequences of the ZJUL5 dataset [19]) and can be parallelized, the pipeline achieves 1.5 fps without further optimization.", + "bbox": [ + 511, + 478, + 906, + 584 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Even with state-of-the-art quality, ToF-Splatting manifests some issues that will be addressed in future developments. The most concerning is the runtime, which is too slow for real-time applications and requires memory and compute-intensive backpropagation at deployment. Even though fast 3DGS frameworks exist, such methodologies have not been transferred yet into the SLAM community, that being outside the scope of this work.", + "bbox": [ + 511, + 590, + 908, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 747, + 633, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented the first 3DGS-based SLAM pipeline relying on sparse ToF depth sensing, as it provides accurate tracking and mapping from cheap and low-power cameras. Moreover, we showed how this result can be achieved through the integration of multi-view geometry, sparse depth data, and monocular cues, yielding an end-to-end 3DGS-based SLAM system. Finally, we assessed the effectiveness of our approach on the ZJUL5 and Replica datasets.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28351", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 2", + "[2] Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yunsong Wang, and Gim Hee Lee. Vcr-gaus: View consistent depth-normal regularizer for gaussian surface reconstruction. arXiv preprint arXiv:2406.05774, 2024. 2, 5", + "[3] Andrea Conti, Matteo Poggi, Valerio Cambareri, and Stefano Mattoccia. Depth on demand: Streaming dense depth from a low frame-rate active sensor. In European Conference on Computer Vision (ECCV), 2024. 2, 4", + "[4] Andrea Conti, Matteo Poggi, Valerio Cambareri, and S. Mattoccia. Range-agnostic multi-view depth estimation with keyframe selection. 2024 International Conference on 3D Vision (3DV), pages 1350–1359, 2024. 4", + "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 4, 5", + "[6] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. BundleFusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4): 1, 2017. 1, 2, 6, 7", + "[7] Tianchen Deng, Guole Shen, Tong Qin, Jianyu Wang, Wentao Zhao, Jingchuan Wang, Danwei Wang, and Weidong Chen. Plgslam: Progressive neural scene representation with local to global bundle adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19657-19666, 2024. 2", + "[8] Zilong Dong, Guofeng Zhang, Jiaya Jia, and Hujun Bao. Keyframe-based real-time camera tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1538-1545. IEEE, 2009. 1", + "[9] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR, 2024. 2", + "[10] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. KinectFusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User Interface Software and Technology, pages 559-568, 2011. 6, 7", + "[11] Mohammad Mahdi Johari, Camilla Carta, and François Fleuret. Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17408-17419, 2023. 2", + "[12] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for densergb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 2", + "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. 1, 2, 3", + "[14] Georg Klein and David Murray. Parallel tracking and mapping for small AR workspaces. In Proceedings of the IEEE and ACM International Symposium on Mixed and Augmented Reality, pages 225-234. IEEE, 2007. 1", + "[15] Yanyan Li, Nikolas Brasch, Yida Wang, Nassir Navab, and Federico Tombari. Structure-slam: Low-drift monocular slam in indoor environments. IEEE Robotics and Automation Letters, 5(4):6583-6590, 2020. 2", + "[16] Yijin Li, Xinyang Liu, Wenqian Dong, Han Zhou, Hujun Bao, Guofeng Zhang, Yinda Zhang, and Zhaopeng Cui. Deltar: Depth estimation from a light-weight tof sensor and rgb image. In European Conference on Computer Vision, 2022. 3, 4, 6", + "[17] Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R Oswald. Loopy-slam: Dense neural slam with loop closures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20363-20373, 2024. 2", + "[18] Haomin Liu, Guofeng Zhang, and Hujun Bao. Robust keyframe-based monocular SLAM for augmented reality. In Proceedings of the IEEE International Symposium on Mixed and Augmented Reality, pages 1-10. IEEE, 2016. 1", + "[19] Xinyang Liu, Yijin Li, Yanbin Teng, Hujun Bao, Guofeng Zhang, Yinda Zhang, and Zhaopeng Cui. Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1-11, 2023. 2, 4, 5, 6, 7, 8", + "[20] Gregor Luetzenburg, Aart Kroon, and Anders A. Bjørk. Evaluation of the Apple iPhone 12 Pro LiDAR for an Application in Geosciences. Scientific Reports, 11(1), 2021. 1", + "[21] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. 2024 International Conference on 3D Vision (3DV), pages 800-809, 2023. 2", + "[22] Hidenobu Matsuki, Riku Murai, Paul H.J. Kelly, and Andrew J. Davison. Gaussian splattering slam. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18039-18048, 2023. 2, 4, 5, 6, 7, 8", + "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision, pages 405-421. Springer, 2020. 1", + "[24] Raul Mur-Artal and Juan D. Tardós. ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo, and RGB-D Cameras. IEEE Transactions on Robotics, 33(5): 1255–1262, 2017. 2", + "[25] Xin Qiao, Matteo Poggi, Pengchao Deng, Hao Wei, Chenyang Ge, and Stefano Mattoccia. Rgb guided tof imag-" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28352", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing system: A survey of deep learning-based methods. International Journal of Computer Vision, pages 1-38, 2024. 2", + "[26] Erik Sandström, Yue Li, Luc Van Gool, and Martin R Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433-18444, 2023. 2", + "[27] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2", + "[28] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 8", + "[29] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 573-580. IEEE, 2012. 5, 6", + "[30] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2, 6, 7, 8", + "[31] Keisuke Tateno, Federico Tombari, Iro Laina, and Nassir Navab. Cnn-slam: Real-time dense monocular slam with learned depth prediction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6243-6252, 2017. 2", + "[32] Zachary Teed and Jia Deng. Droid-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. Advances in Neural Information Processing Systems, 34: 16558-16569, 2021. 2, 7", + "[33] Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R. Oswald, and Matteo Poggi. How nerfs and 3d gaussian splattering are reshaping slam: a survey, 2024. 1, 2", + "[34] Hengyi Wang, Jingwen Wang, and Lourdes Agapito. Coslam: Joint coordinate and sparse parametric encodings for neural real-time slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13293-13302, 2023. 2", + "[35] Yifan Wang, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38:1-14, 2019. 3", + "[36] Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, Andrew J Davison, and Stefan Leutenegger. ElasticFusion: Real-time dense SLAM and light source estimation. The International Journal of Robotics Research, 35(14):1697-1716, 2016. 1, 2, 6, 7" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Xingming Wu, Zimeng Liu, Yuxin Tian, Zhong Liu, and Weihai Chen. Kn-slam: Keypoints and neural implicit encoding slam. IEEE Transactions on Instrumentation and Measurement, 73:1-12, 2024. 2", + "[38] Liu Xinyang, Li Yijin, Teng Yanbin, Bao Hujun, Zhang Guofeng, Zhang Yinda, and Cui Zhaopeng. Multi-modal neural radiance field for monocular dense slam with a lightweight tof sensor. In International Conference on Computer Vision (ICCV), 2023. 2", + "[39] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595-19604, 2024. 2", + "[40] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. ArXiv, abs/2406.09414, 2024. 4, 8", + "[41] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. ArXiv, abs/2312.10070, 2023. 5", + "[42] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting, 2023. 2", + "[43] Guofeng Zhang, Jiaya Jia, Tien-Tsin Wong, and Hujun Bao. Recovering consistent video depth maps via bundle optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2008. 1", + "[44] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimization for consistent 3d instant reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3727-3737, 2023. 2", + "[45] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 2", + "[46] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. NICE-SLAM: Neural implicit scalable encoding for SLAM. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2, 6, 7, 8", + "[47] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R. Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. 2024 International Conference on 3D Vision (3DV), pages 42-52, 2023. 4", + "[48] Yiming Zuo and Jia Deng. Ogni-dc: Robust depth completion with optimization-guided neural iterations. arXiv preprint arXiv:2406.11711, 2024. 6, 8", + "[49] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. Surface splatting. Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001. 3" + ], + "bbox": [ + 516, + 90, + 905, + 882 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28353", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/93b627e5-bb27-4dc6-aa8c-e29320f5f5eb_model.json b/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/93b627e5-bb27-4dc6-aa8c-e29320f5f5eb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3ddfe900f9338d3595475ff72ebe55f892b85c76 --- /dev/null +++ b/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/93b627e5-bb27-4dc6-aa8c-e29320f5f5eb_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.178, + 0.13, + 0.823, + 0.177 + ], + "angle": 0, + "content": "ToF-Splatting: Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.203, + 0.364, + 0.221 + ], + "angle": 0, + "content": "Andrea Conti \\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.403, + 0.204, + 0.525, + 0.222 + ], + "angle": 0, + "content": "Matteo Poggi" + }, + { + "type": "text", + "bbox": [ + 0.572, + 0.204, + 0.732, + 0.221 + ], + "angle": 0, + "content": "Valerio Cambareri" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.222, + 0.455, + 0.238 + ], + "angle": 0, + "content": "Martin R. Oswald" + }, + { + "type": "text", + "bbox": [ + 0.502, + 0.222, + 0.658, + 0.238 + ], + "angle": 0, + "content": "Stefano Mattoccia" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.241, + 0.83, + 0.257 + ], + "angle": 0, + "content": "1University of Bologna, Italy 2Sony DepthSensing Solutions, Belgium 3University of Amsterdam, Netherlands" + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.291, + 0.868, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.462, + 0.908, + 0.491 + ], + "angle": 0, + "content": "Figure 1. Overview of our ToF-Splatting method. Our method combines sparse ToF depth, multi-view geometry from a buffer of keyframes, and monocular cues (left) to perform into a unique end-to-end dense SLAM framework enabled by a Gaussian Splatting." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.503, + 0.327, + 0.518 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.535, + 0.486, + 0.793 + ], + "angle": 0, + "content": "Time-of-Flight (ToF) sensors provide efficient active depth sensing at relatively low power budgets; among such designs, only very sparse measurements from low-resolution sensors are considered to meet the increasingly limited power constraints of mobile and AR/VR devices. However, such extreme sparsity levels limit the seamless usage of ToF depth in SLAM. In this work, we propose ToF-Splatting, the first 3D Gaussian Splatting-based SLAM pipeline tailored for using effectively very sparse ToF input data. Our approach improves upon the state of the art by introducing a multi-frame integration module, which produces dense depth maps by merging cues from extremely sparse ToF depth, monocular color, and multi-view geometry. Extensive experiments on both real and synthetic sparse ToF datasets demonstrate the advantages of our approach, as it achieves state-of-the-art tracking and mapping performances on reference datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.821, + 0.222, + 0.836 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.847, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Simultaneous Localization and Mapping (SLAM) [6, 8, 36, 43] entails the joint estimation of the current camera pose" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.907, + 0.626 + ], + "angle": 0, + "content": "against a 3D scene representation (i.e., the map) and the update of this representation with data from the current camera view. It is a fundamental task with applications in AR/VR, robotics, and autonomous navigation [14, 18]. In recent years, SLAM witnessed a revolution [33] following the success of Neural Radiance Fields (NeRF) [23] and 3D Gaussian Splatting (3DGS) [13] as scene representations from the adjacent field of novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this paper we focus on SLAM pipelines that leverage RGB-D video inputs, i.e., where color images and dense depth maps are collected by synchronized, accurate color and depth sensors. Beyond global scale, depth sensors provide accurate local geometry that grants better tracking and mapping quality. A common choice for active depth sensing are Time-of-Flight (ToF) cameras. Despite the increased accuracy provided by reliable depth measurements, the power consumption and cost of such sensors constrain their adoption to high-end industrial or automotive applications. A recent trend to favor ToF sensor integration in handheld consumer devices is to reduce the specifications of ToF sensors to provide low-resolution but more reliable depth measurements. Among other examples, LiDAR (an instance of ToF) appears in consumer mobile phones [20] and is known to feature up to at most 576 dots (interpolated to \\(256 \\times 192\\) resolution via depth completion). Other recent ToF sensors provide only up to 64 depth points but consume a re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.409, + 0.9 + ], + "angle": 0, + "content": "* work started while visiting University of Amsterdam." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28344" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "ported \\(200\\mathrm{mW}\\) [19]. In such settings a SLAM pipeline can expect to receive, e.g., \\(8 \\times 8\\) very noisy, very sparse depth measurements per frame. These specifications prevent their straightforward integration in SLAM systems, as low-resolution noisy depth is known to harm the accuracy of dense SLAM systems built with NeRF and not originally designed to deal with this scenario [30, 36, 46]. This is even more evident with the more recent 3DGS-based SLAM systems [22], where accurate and dense depth is paramount to achieve satisfactory results." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.246, + 0.485, + 0.578 + ], + "angle": 0, + "content": "To overcome these limitations, a straightforward strategy [38] would consist in recovering dense and accurate depth maps using a single-image depth completion framework [25]. However, models trained for this task are known to generalize poorly across domains and across input depth noise levels, potentially affecting the downstream SLAM pipeline accuracy and yielding both drifted camera trajectory and inaccurate mapping. Therefore, we explore 3DGS as an explicit representation that allows for injecting domain-specific knowledge and that is easily extended to, e.g., dynamic scenes [21] and accurate large mesh reconstruction [2, 9]. On the one hand, we argue that depth completion alone is insufficient to allow sparse ToF depth usage in SLAM systems. On the other, we acknowledge that even sparse and noisy depth measurements can bootstrap SLAM if properly assisted with geometry. Accordingly, inspired by frame-to-frame SLAM systems relying on pre-trained trackers for localization, we believe that multi-frame integration fusing the noisy depth measurements with multiview geometry across a small set of local frames could supply the dense SLAM system with the supervision necessary to fill this gap." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.807 + ], + "angle": 0, + "content": "In this paper, we propose ToF-Splatting, a novel 3DGS-based SLAM system enabling dense reconstruction from sparse ToF sensors. ToF-Splatting harmoniously alternates tracking and mapping, with each one benefiting from the other. The former is carried out through backpropagation during the optimization of the 3DGS model on color images; the latter exploits a pre-trained network [3] as a multi-frame integration module, combining sparse depth measurements with multi-view geometry according to a buffer of keyframes. For each of the keyframes, we use the camera pose estimated through tracking and providing dense depth maps to improve the supervision given to the 3DGS model itself. Figure 1 shows an overview of the setting in which ToF-Splatting operates, as well as the quality of the 3D reconstructions it yields." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "ToF-Splatting is tested on the real-world ZJUL5 dataset [19], where we measure tracking and mapping quality and find the superiority of our framework with respect to existing solutions built over sparse ToF sensors [38]. Moreover, standard benchmarks such as Replica [28] are also provided and confirm the effectiveness of our mapping strat" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.809, + 0.106 + ], + "angle": 0, + "content": "egy. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.107, + 0.904, + 0.137 + ], + "angle": 0, + "content": "- We propose ToF-Splatting, the first 3DGS-based SLAM system suited for sparse ToF sensors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- We introduce a novel multi-frame integration method combining sparse depth, multi-view geometry, and monocular cues for robust noise and outlier handling." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.182, + 0.905, + 0.213 + ], + "angle": 0, + "content": "- We establish a new state-of-the-art for dense SLAM systems built on top of sparse ToF depth sensors." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.226, + 0.655, + 0.241 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.252, + 0.905, + 0.281 + ], + "angle": 0, + "content": "We briefly review the literature relevant to our work, referring to [33] for a detailed overview of the latest advances." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.282, + 0.905, + 0.447 + ], + "angle": 0, + "content": "Traditional SLAM. Traditional visual SLAM pipelines, or the more recent ones introducing deep modules [15, 31, 32], usually perform tracking in frame-to-frame fashion, based on the visual features extracted from past frames either using handcrafted feature extractors [24] or deep networks [31, 32]. These systems are usually defined by four main steps: tracking, mapping, global bundle adjustment, and loop closure. Among the different representations used by traditional SLAM frameworks, depth points [1, 24], surfels [27], and volumetric representations [6] allows for achieving globally consistent 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.448, + 0.906, + 0.735 + ], + "angle": 0, + "content": "NeRF-based SLAM. The advent of NeRF in the field of novel view synthesis also reached adjacent research areas, as well as SLAM a few years later. NeRF-based SLAM systems estimate camera poses and reconstruct dense meshes by modeling scenes via MLPs. iMAP [30] and NICE-SLAM [45] are the first SLAM frameworks following this approach, with the latter introducing dense feature grids supporting the MLPs. Point-SLAM [26] and Loopy-SLAM [17] switched feature grids with neural point embeddings, allowing for higher flexibility and correcting or adjusting local maps after global optimization, yet with slow processing prohibiting deployment in robotics. ESLAM [11] and Co-SLAM [34] deploy tri-planes and hash grid embeddings, improving both processing speed and reconstruction accuracy, with PLGSLAM [7] further improving the representation capacity in large indoor scenes. GO-SLAM [44] and KN-SLAM [37] implement frame-to-frame systems thanks to external trackers, yielding a good trade-off between tracking/mapping accuracy and frame rate." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.905, + 0.856 + ], + "angle": 0, + "content": "3DGS-based SLAM. More recent advances in novel view synthesis brought to the development of new SLAM systems using 3D Gaussian Splatting [13] to represent the mapped scene. Several concurrent frameworks emerged built over 3DGS, such as MonoGS [22], GaussianSLAM [42], SplaTAM [12], and GS-SLAM [39]. Different from NeRF-based SLAM systems, these frameworks allow for higher interpretability and higher rendering speed." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Sparse ToF-based SLAM. Only lately, the use of more compact and energy-efficient depth sensors has gained attention in the SLAM research community. ToF-SLAM [38]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28345" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.124, + 0.092, + 0.868, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.269, + 0.908, + 0.312 + ], + "angle": 0, + "content": "Figure 2. ToF-Splatting Pipeline. Our method involves three main modules: a Tracking frontend estimating camera poses, a Multi-Frame Integration module that predicts dense depth maps from sparse ToF measurements and multi-view geometry, and a Mapping backend modeling the 3D scene representation via 3D Gaussian Splatting." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.327, + 0.485, + 0.435 + ], + "angle": 0, + "content": "represents the only attempt to exploit \"lightweight\", sparse depth measurements within a dense SLAM system, aided by a depth completion model [16]. However, depth completion alone fails to deal with noise in the sparse input depth, a shortcoming we aim to specifically overcome within our ToF-Splatting framework, the first one based on 3DGS and suited for sparse ToF sensors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.457, + 0.295, + 0.475 + ], + "angle": 0, + "content": "3. Pipeline Architecture" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.486, + 0.483, + 0.593 + ], + "angle": 0, + "content": "In this section, we present the core components of our framework. ToF-Splatting is a frame-to-model SLAM system conceptually divided into three modules: \\((i)\\) a mapping backend, \\((ii)\\) a tracking frontend, and \\((iii)\\) a peculiar multi-frame integration module yielding dense depth maps based on a subset of the keyframe graph, the current RGB frame, and the current sparse ToF depth map." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.596, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Figure 2 provides an overview of our pipeline. Our framework is characterized by three heterogeneous modules that are tightly related and influence each other. The tracking module stands as the most influencing element in the pipeline, as it selects the keyframes used by both the mapping and multi-frame integration parts, moreover, it computes the poses used for multi-frame depth estimation. Nonetheless, it relies on the quality of the mapping step to perform correct ego-motion estimation and exploits the depth cues to properly behave when photometric information is lacking or ambiguous." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The depth perception part is mainly sustained by the tracking step and highly influences the mapping part since it is used to seed new Gaussians and as supervision. Finally, the mapping phase not only creates a unique smooth representation but also defines the rendered opacity that is used for keyframe selection. These connections lead to a smoothly integrated framework that not only exploits a multi-frame backbone but more importantly, defines an effective way to use such information in the SLAM scenario." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.327, + 0.831, + 0.344 + ], + "angle": 0, + "content": "3.1. Background: 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.907, + 0.44 + ], + "angle": 0, + "content": "3D Gaussian Splitting [13] (3DGS) represents one of the latest advances in novel view synthesis, fitting a dynamically defined set of multivariate Gaussian distributions over the scene to enable image rendering from arbitrary viewpoints. Each Gaussian \\( G(\\mathbf{x}) \\) is parametrized by its mean \\( \\boldsymbol{\\mu} \\in \\mathbb{R}^3 \\) and covariance matrix \\( \\boldsymbol{\\Sigma} \\in \\mathbb{R}^{3 \\times 3} \\), reading" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.445, + 0.907, + 0.479 + ], + "angle": 0, + "content": "\\[\nG (\\mathbf {x}) = \\exp \\left(- \\frac {1}{2} (\\mathbf {x} - \\boldsymbol {\\mu}) ^ {\\top} \\boldsymbol {\\Sigma} ^ {- 1} (\\mathbf {x} - \\boldsymbol {\\mu})\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.906, + 0.514 + ], + "angle": 0, + "content": "Moreover, opacity \\(o \\in \\mathbb{R}\\) and RGB color \\(\\mathbf{c} \\in \\mathbb{R}^3\\) are assigned to each Gaussian." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.907, + 0.59 + ], + "angle": 0, + "content": "Given a set of images from different viewpoints all around the scene, the Gaussian parameters are optimized using the splatting technique [35]. With such a method, 3D Gaussians are projected through rasterization over the 2D image plane, containing a 2D Gaussian distribution" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.595, + 0.905, + 0.615 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} ^ {2 D} = \\pi \\left(\\mathbf {T} _ {w c} \\bar {\\boldsymbol {\\mu}}\\right) \\quad \\boldsymbol {\\Sigma} ^ {2 D} = \\mathbf {J} \\mathbf {R} _ {w c} \\boldsymbol {\\Sigma} \\mathbf {R} _ {w c} ^ {\\top} \\mathbf {J} ^ {\\top} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.906, + 0.711 + ], + "angle": 0, + "content": "where \\(\\bar{\\pmb{\\mu}}\\) is \\(\\pmb{\\mu}\\) in homogeneous coordinates, \\(\\mathbf{T}_{wc} \\in \\mathbb{R}^{4 \\times 4}\\), \\(\\mathbf{R}_{wc} \\in \\mathbb{R}^{3 \\times 3}\\) are respectively the world-to-camera transformation and its rotational component, \\(\\pi\\) is the projection matrix and \\(\\mathbf{J} \\in \\mathbb{R}^{2 \\times 3}\\) is the Jacobian of the projective transformation [49]. The color \\(C\\) of a pixel is then given by the combination of the \\(M\\) weighted Gaussians, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.717, + 0.905, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} = \\sum_ {i \\in M} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.908, + 0.827 + ], + "angle": 0, + "content": "with \\(\\alpha_{i} = o_{i}G(\\mathbf{x}_{i})\\). Since \\(\\pmb{\\Sigma}\\) is positive semi-definite, it is parametrized with a diagonal scaling matrix \\(\\mathbf{S}\\) and a rotation matrix \\(\\mathbf{R}\\), so that \\(\\pmb{\\Sigma} = \\mathbf{RSS}^{\\top}\\mathbf{R}^{\\top}\\). \\(\\mathbf{R}\\) is internally represented with quaternions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.742, + 0.85 + ], + "angle": 0, + "content": "3.2. Multi-Frame Integration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Dense depth cues are pivotal for proper initialization of the Gaussians position \\(\\pmb{\\mu} \\in \\mathbb{R}^3\\), as well as to provide supervision to both tracking and mapping threads - as already" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28346" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.091, + 0.482, + 0.243 + ], + "angle": 0, + "content": "known in the literature [22, 47]. However, sparse ToF sensors alone are insufficient for this purpose, because of the very sparse and noisy depth measurements they provide (e.g., only 64 points for a sensor such as that used in [19]). To overcome this limitation, ToF-SLAM [19] exploits an auxiliary completion model, DELTAR [16], to recover dense depth maps out of ToF measurements. However, depth completion is highly dependent on the quality of the sparse depth in input, rapidly degrading the accuracy of the densified depth maps in the presence of noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.482, + 0.367 + ], + "angle": 0, + "content": "In this paper, we follow a different path: inspired by frame-to-frame SLAM systems deploying an external tracker, we believe that multi-frame integration can compensate for the noise in the ToF measurements by exploiting both single-view and multi-view geometry cues on a set of local frames, given the camera poses estimated by the tracking frontend. We choose to extend the Depth on Demand (DoD) framework [3] for this purpose." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.37, + 0.485, + 0.779 + ], + "angle": 0, + "content": "Multi-Frame Integration. DoD [3] integrates sparse depth and two-view stereo cues for dense depth prediction. Specifically – given a frame \\(\\mathbf{F}^k\\) – it processes the associated RGB image \\(I_k\\) (the target view), another RGB view \\(I_j\\) from a different frame \\(\\mathbf{F}^j\\) (a source view), the relative pose \\(P^{jk}\\) between the two, and the sparse depth points \\(H^k\\) to predict a dense metric depth map \\(D^k\\). Differently from monocular depth prediction, DoD is aware of the scene scale either from \\(H^k\\) or \\(P^{jk}\\), simplifying the problem of integrating scale-inconsistent geometries in the SLAM frontend. However, DoD is limited to two-view processing: this may lead to suboptimal results in the SLAM scenario, where an abundant amount of keyframes largely overlap and cover the same portion of the scene. Thus, we extend [3] to multiview processing. This is achieved by observing that the predicted depth map \\(D^k\\) depends on the source view \\(I_j\\) through its relative pose \\(P^{jk}\\) only. DoD iteratively updates \\(D^k\\) for multiple iterations, exploiting this intrinsic property we can integrate multiple views \\((I_j, P^{jk})\\) using a different one at each iteration [4]. This approach allows smooth integration of multi-view cues avoiding order dependency. To select the source views \\(I_j\\), at each prediction we order the keyframes selected by the SLAM pipeline according to their relative pose similarity and select the first \\(N\\) frames with baseline distance close to \\(b = 15\\mathrm{cm}\\) to ensure enough parallax. The model is trained from scratch on ScanNetv2 [5] and following the protocol of [3]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Monocular Cues Integration. As shown in [3], DoD heavily relies on both geometry and depth measurements, thus struggling at generalizing in challenging scenarios where the two lack. This may happen in cases where the tight field of view, noisy sparse points, and textureless surfaces hamper multi-view matching. To mitigate such limitations, we integrate explicit monocular cues into DoD by feeding its monocular encoder with \\( I_{k} \\) and the normalized depth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.905, + 0.182 + ], + "angle": 0, + "content": "map \\(\\tilde{D}^k\\in [0,1]^{H\\times W}\\) obtained through the single-image depth estimator Depth Anything v2 [40]. This approach allows the injection of a robust bias, avoiding issues related to monocular estimation, e.g., slanted surfaces or noisy scale estimates, and delegating to DoD their handling. These cues are included while retraining on ScanNetv2 [5]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Outlier Handling. Furthermore, the noisy depth measurements in input to DoD can severely affect its accuracy. Therefore, we explicitly focus on dealing with outliers in the input sparse depth, that may be perceived by ToF devices over specific surfaces. Such errors are particularly evident in cheap sensors and may largely prejudice the information supplied by the few points they measure. Purposely, we deploy a simple yet effective method to identify such outliers by exploiting multi-view geometry again. Whenever depth needs to be predicted for a new frame, DoD predicts a depth map \\(\\hat{D}^k\\) without processing the ToF measurements. Then, the \\(\\ell_1\\) error between the latter and the depth measured by the ToF sensor \\(H^k\\) is computed and finally sparse measures with error higher than a given quantile \\(q\\) are discarded, and depth is predicted again by DoD by also processing the filtered sparse depth \\(\\tilde{H}^k\\) this time. This way, we exploit multi-view cues to extract a clean depth prior to filtering out inconsistent sparse depth measures, and then we integrate the remaining points to improve the original prediction. Even without processing the sparse depth, DoD still predicts metric depth, as the metric scale is enforced by the relative camera poses." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.522, + 0.696, + 0.538 + ], + "angle": 0, + "content": "3.3. Tracking Frontend" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.905, + 0.604 + ], + "angle": 0, + "content": "The tracking part of ToF-Splatting estimates the ego-motion of the camera for each new frame \\(\\mathbf{F}^k\\) and builds a keyframe graph \\(\\mathbf{G} = \\{\\mathbf{KF}^k\\} \\subset \\{\\mathbf{F}^k\\}\\) containing a set of meaningful frames for mapping and multi-frame depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.605, + 0.905, + 0.696 + ], + "angle": 0, + "content": "Pose Estimation. We initialize each new frame \\(\\mathbf{F}^k\\) of our frame-to-model tracker with the pose of the previous frame \\(\\mathbf{F}^{k - 1}\\) and then minimizing the tracking loss with respect to the relative pose between \\(\\mathbf{F}^k\\) and \\(\\mathbf{F}^{k - 1}\\), following [22]. The tracking loss consists of two terms, \\(L_{\\mathrm{rgb}}\\) and \\(L_{\\mathrm{depth}}\\), respectively accounting for photometric and geometric errors" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.718, + 0.905, + 0.735 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t r a c k}} = \\lambda_ {\\text {t r a c k}} L _ {\\text {r g b}} + (1 - \\lambda_ {\\text {t r a c k}}) L _ {\\text {d e p t h}} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.737, + 0.905, + 0.756 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {r g b}} = \\left\\| I _ {k} - \\hat {I} \\left(\\mathbf {F} ^ {k}\\right) \\right\\| _ {1} \\quad L _ {\\mathrm {d e p t h}} = \\left\\| D ^ {k} - \\hat {D} \\left(\\mathbf {F} ^ {k}\\right) \\right\\| _ {1} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.905, + 0.81 + ], + "angle": 0, + "content": "where \\(I_{k}\\) is the RGB image from \\(\\mathbf{F}^k\\), \\(\\hat{I}(\\cdot)\\) and \\(\\hat{D}(\\cdot)\\) denote the rendered RGB and Depth and, \\(D^{k}\\) is the depth prediction produced by the multi-frame integration module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Unlike other methods, the tracking process happens in two stages. Initially, for a fixed number of steps \\(\\eta_{\\mathrm{rgb}}\\) only the photometric loss \\(L_{\\mathrm{rgb}}^T\\) is optimized. Then, such a pose is used for multi-frame integration, as described in Sec. 3.2, to generate \\(D^{k}\\). Finally, both losses in Eq. (5) are used for \\(\\eta_{\\mathrm{rgbd}}^T\\) further steps." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28347" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.288 + ], + "angle": 0, + "content": "Keyframe Selection Policy. We deploy a simple yet effective keyframe selection policy that considers i) novel content to be mapped and ii) instabilities in the already mapped areas. Unlike other methods using complex approaches – e.g., frustum intersection, point clouds and pose analysis [22, 41] – we observe that the rendered opacity for a novel frame contains low opacity values where either Gaussians are not present or unstable. The first case highlights the presence of novel areas of the environment to be mapped, the second happens due to the pruning procedure involved in the mapping step (see Sec. 3.4). We set a frame \\(\\mathbf{F}^k\\) as a keyframe \\(\\mathbf{K}\\mathbf{F}^k\\) if its novelty factor \\(\\nu^k\\) is higher than a threshold \\(\\nu^{th}\\), where we define \\(\\nu^k\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.3, + 0.483, + 0.338 + ], + "angle": 0, + "content": "\\[\n\\nu^ {k} = \\frac {\\sum_ {i \\in \\mathbf {K F} ^ {k}} \\mathbf {M} \\left(\\mathbf {K F} _ {i} ^ {k}\\right)}{\\sum_ {i \\in \\mathbf {K F} ^ {k}} \\left(1 - \\mathbf {M} \\left(\\mathbf {K F} _ {i} ^ {k}\\right)\\right)} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.34, + 0.483, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} \\left(\\mathbf {K F} _ {i} ^ {k}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\hat {O} \\left(\\mathbf {K F} _ {i} ^ {k}\\right) < \\sigma \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.483, + 0.426 + ], + "angle": 0, + "content": "with \\(\\hat{O} (\\mathbf{K}\\mathbf{F}_i^k)\\) being the rendered opacity at pixel \\(i\\), and \\(\\mathbf{M}(\\mathbf{K}\\mathbf{F}_i^k)\\) a binary uncertainty map." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.502 + ], + "angle": 0, + "content": "This approach allows for skipping several frames when the camera moves slowly or focusing deeply on areas where the ToF-Splatting struggles the most, by mapping it consecutively. Nonetheless, we enforce mapping after skipping a certain number of frames to avoid mapping too few frames." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.512, + 0.268, + 0.529 + ], + "angle": 0, + "content": "3.4. Mapping Backend" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.483, + 0.58 + ], + "angle": 0, + "content": "Whenever a new keyframe \\(\\mathbf{K}\\mathbf{F}^k\\) is identified, mapping is triggered. This phase seeks to embed the new frame in the global 3DGS model, involving the following two steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Initialization. In this step, new Gaussians are seeded in the 3DGS model. To limit the size of the model and reduce outliers, only areas where the rendered uncertainty \\(\\mathbf{M}(\\mathbf{KF}^k)\\) is high are seeded with new points, with the new Gaussians also being randomly downsampled by a constant factor. Each Gaussian is initialized using depth \\(D^{k}\\) predicted in the multi-frame integration step, with color from the RGB frame \\(I_{k}\\), and scale initialized with a constant value. Unlike other 3DGS pipelines [22, 41], we prove this approach effective and particularly efficient, replacing any complex 3D heuristics to a simple use of rendered opacity \\(O(\\cdot)\\), already a side-product of the rasterization process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Optimization. Firstly, ToF-Splating smoothly integrates the new information in the global model and secondly tunes the existing representation. This is achieved by optimizing the new keyframe \\(\\mathbf{KF}^k\\) and a random subset of 5 keyframes among the last \\(N\\) keyframes \\(\\{\\mathbf{KF}^j \\mid j \\geq \\max(k - N, 0)\\}\\). These are optimized for a fixed number of steps \\(\\eta_{\\mathrm{rgbd}}^M\\). Moreover, we enforce a sampling rate of \\(60\\%\\) for \\(\\mathbf{KF}^k\\), with the remaining \\(40\\%\\) for the remaining \\(\\mathbf{KF}^j\\) to avoid forgetting. During this step, gradients are also back-propagated to the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.169 + ], + "angle": 0, + "content": "camera poses of the sampled keyframes, thus acting as a global bundle adjustment. Finally, we prune Gaussians with opacities lower than 0.5 every \\(\\eta_{\\mathrm{rgbd}}^{M} / 2\\) steps. In addition to the terms in Eq. (5), we optimize structural similarity as in Eq. (8) and, following [2], the normals as in Eq. (9), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.176, + 0.906, + 0.21 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {d s s i m}} = \\frac {1 - \\operatorname {S S I M} \\left(I _ {k} , \\hat {I} \\left(\\mathbf {K F} ^ {k}\\right)\\right)}{2} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.213, + 0.905, + 0.233 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {n o r m a l s}} = \\| N ^ {k} - \\hat {N} \\| _ {1} + \\overline {{\\mathbf {1} - \\langle N ^ {k} , \\hat {N} \\rangle}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.242, + 0.906, + 0.32 + ], + "angle": 0, + "content": "with \\(\\hat{N} \\coloneqq \\bar{\\nabla}\\hat{D}(\\mathbf{K}\\mathbf{F}^k)\\), \\(N^k \\coloneqq \\bar{\\nabla}D^k\\) denoting the normals estimated via 3D gradients \\(\\bar{\\nabla}\\) for the rendered and multiframe predicted depth respectively, \\(\\langle \\cdot, \\cdot \\rangle\\) their inner product, and \\(\\overline{\\cdot}\\) the average. \\(L_{\\mathrm{normals}}\\) is masked on edges to preserve sharpness, as identified by running a Sobel filter." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.906, + 0.427 + ], + "angle": 0, + "content": "Finally, we introduce a regularization loss \\( L_{\\mathrm{iso}} \\) to penalize elongated Gaussians (i.e., promote their isotropy) that usually leads to rendering artifacts as already observed in [22]. Eq. (10) defines \\( L_{\\mathrm{iso}} \\), where \\( \\frac{\\mathrm{diag}(\\cdot)}{\\mathrm{diag}(\\mathbf{S}_j)} \\) extracts the diagonal values from its argument, and \\( \\overline{\\mathrm{diag}(\\mathbf{S}_j)} \\) is the average of the resulting vector, i.e., of the elements of the \\( j \\)-th diagonal scale matrix \\( \\mathbf{S}_j \\) as defined in Sec. 3.1. Thus," + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.436, + 0.905, + 0.481 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {i s o}} = \\frac {1}{| \\mathcal {G} |} \\sum_ {j} ^ {| \\mathcal {G} |} \\| \\operatorname {d i a g} \\left(\\mathbf {S} _ {j}\\right) - \\overline {{\\operatorname {d i a g} \\left(\\mathbf {S} _ {j}\\right)}} \\cdot \\mathbf {1} _ {3 \\times 1} \\| _ {1} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.489, + 0.906, + 0.521 + ], + "angle": 0, + "content": "The final mapping loss aggregates the aforementioned loss terms with corresponding weights as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.529, + 0.905, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\text {m a p}} = \\lambda_ {\\text {m a p}} \\left(\\lambda_ {\\text {v i s u a l}} L _ {\\text {r g b}} + (1 - \\lambda_ {\\text {v i s u a l}}) L _ {\\text {d s s i m}}\\right) \\tag {11} \\\\ + \\left(1 - \\lambda_ {\\text {m a p}}\\right) L _ {\\text {d e p t h}} + \\lambda_ {\\text {n o r m a l s}} L _ {\\text {n o r m a l s}} + \\lambda_ {\\text {i s o}} L _ {\\text {i s o}} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.646, + 0.607 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.644 + ], + "angle": 0, + "content": "This section assesses both the tracking and mapping performance of ToF-Splatting on the following datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.84 + ], + "angle": 0, + "content": "ZJUL5. ZJUL5 [19] is the only existing public-domain real dataset providing sparse depth data from a low-resolution, low-cost ToF sensor. A VL53L5CX ToF sensor is assembled on a calibrated rig with an Intel RealSense 435i to provide dense depth ground truth. This benchmark comprises 7 diverse indoor scene recordings. Ground truth 3D meshes and camera poses are obtained following the ScanNet protocol [5]. This sensor provides \\(8 \\times 8\\) depth points by counting the number of photons returned in each discretized time range. Of the at most 64 depth points provided, many are extremely noisy and introduce large outliers. This challenging real-world dataset provides real measurement noise that is not accurately represented in mainstream datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "TUM RGB-D. This is a real dataset [29] providing indoor sequence captures and widely used as benchmark for RGB-D SLAM pipelines. The depth stream provides dense VGA depth maps from a high-resolution Kinect v1 depth sensor." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28348" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.213 + ], + "angle": 0, + "content": "
MethodKitchenSofaOfficeReceptionLiving RoomOffice2Sofa2Avg.
KinectFusion [10]X0.1460.2090.157X0.3210.1250.192
iMAP [30]X1.6580.3380.6480.6790.3440.2140.647
NICE-SLAM [46]0.7450.1440.1550.2510.2890.2280.4210.319
BundleFusion [6]0.1760.1020.1350.101X0.1630.1200.132
ElasticFusion [36]0.2530.1100.0700.1930.5300.1210.1460.203
MonoGS [22]0.2310.0320.0410.0440.1530.0350.0730.087
ToF-SLAM [19]0.1130.0810.0560.1140.2000.1010.0850.107
ToF-Splatting (ours)0.0880.0240.0220.0410.1220.0290.0300.051
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.217, + 0.907, + 0.26 + ], + "angle": 0, + "content": "Table 1. ZJUL5 Tracking Results. We show here the tracking performance on the ZJUL5 dataset [19] dataset with the ATE RMSE [m] (↓) metric on the 8 sequences available where \\( \\times \\) indicates lost tracking. ToF-Splatting provides the better performance by a good margin on each sequence, demonstrating the optimal capability of our proposal to track the camera ego-motion accurately." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.274, + 0.482, + 0.35 + ], + "angle": 0, + "content": "Replica. Replica [28] is a synthetic dataset providing extremely realistic indoor sequences. It provides dense depth maps, which we use at different densities to perform experiments that prove the robustness and generalization capabilities of ToF-Splatting in challenging scenarios." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.35, + 0.483, + 0.457 + ], + "angle": 0, + "content": "Baselines. We compare our method with both learning-based [19, 30, 46] and traditional pipelines [6, 10, 36]. Following ToF-SLAM [19], baselines requiring RGB-D frames use DELTAR [16] to densify the sparse ToF measurements. Moreover, we adapt MonoGS [22] densifying depth with the state-of-the-art depth completion framework OGNI-DC [48] for a Gaussian Splatting SLAM baseline." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.457, + 0.483, + 0.643 + ], + "angle": 0, + "content": "Implementation Details. In all our experiments, we use \\(N = 4\\) source views \\(I_{j}\\) (i.e., keyframe buffer elements) for the multi-view integration module and a baseline \\(b = 15\\mathrm{cm}\\) for selecting keyframes. We use a quantile \\(q = 0.75\\) to filter outliers that occur frequently in the sparse ToF depth of the ZJUL5 dataset [19]. Concerning tracking, we use a novelty factor threshold of \\(\\nu_{th} = 0.1\\), \\(\\sigma = 0.98\\). The total tracking loss uses \\(\\lambda_{\\mathrm{track}} = 0.9\\), a total number of \\(\\eta_{\\mathrm{rgb}}^T = 30\\) and \\(\\eta_{\\mathrm{rgbd}}^T = 70\\) steps are performed at each iteration. For mapping, we set the loss weights \\(\\lambda_{\\mathrm{map}} = 60\\), \\(\\lambda_{\\mathrm{visual}} = 0.20\\), \\(\\lambda_{\\mathrm{normals}} = 0.01\\), \\(\\lambda_{\\mathrm{iso}} = 1.0\\) and the number of steps per iteration \\(\\eta_{\\mathrm{rgbd}}^M = 60\\). We test on a single RTX3090." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.651, + 0.414, + 0.667 + ], + "angle": 0, + "content": "4.1. Comparison with the State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.674, + 0.483, + 0.704 + ], + "angle": 0, + "content": "In this section, we compare ToF-Splatting with existing methods in terms of tracking and mapping performances." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.704, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Tracking. In Table 1, we report the tracking performance by ToF-Splatting and various baselines on each sequence of the ZJUL5 dataset [19]. Following common practice we use the Absolute Trajectory Error metric [29], which directly compares the global consistency of a trajectory. While various frameworks [6, 10, 30] fail on the most challenging scenes, ToF-Splatting delivers accurate and consistent tracking. This is evident from the good margin it gains against the baselines. In Figure 3, we show the 3D trajectories on a subset of the scenes, with error below \\(3\\mathrm{cm}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Mapping. In Table 2, we report the mapping performance by ToF-Splatting compared with the other baselines. For each scene, we collect the predicted pose of each frame" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.277, + 0.9, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.604, + 0.907, + 0.66 + ], + "angle": 0, + "content": "Figure 3. Qualitative results on the ZJUL5 dataset [19]. We show meshes obtained by fusing rendered depth maps with TSDF and marching cubes (left), and 3D trajectories (right) on 3 scenes selected from the ZJUL5 dataset [19]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.903 + ], + "angle": 0, + "content": "and render depth and color from 3DGS. Then, we perform TSDF integration and extract the final mesh through marching cubes. Following [19], the meshes are then evaluated by computing Accuracy, Completion, and F-score versus the ground truth meshes. Accuracy and Completion respectively evaluate the mean distance between each predicted vertex and the nearest ground truth one and vice versa. The F-score takes into account both Accuracy and Completion with an aggregate metric. ToF-Splating delivers consistent results for mapping, always achieving the highest accuracy and significantly exceeding the baselines on average and specifically on the Sofa, Office, and Reception scenes which account for five over seven sequences. [19] is the second-best method and the main competitor in mapping." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28349" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.367 + ], + "angle": 0, + "content": "
MethodMetricKitchenSofaOfficeReceptionLiving RoomOffice2Sofa2Avg.
KinectFusion [10]Accuracy ↓X0.1900.2110.261X0.2670.135-
Completion ↓X0.0480.0460.064X0.0780.064-
F-score ↑X0.2780.2880.285X0.2740.381-
ElasticFusion [36]Accuracy ↓0.0920.1350.0840.2970.1510.0960.1220.140
Completion ↓0.0650.0480.0820.3050.2160.1470.0470.130
F-score ↑0.5530.4200.5290.2740.3820.4160.4810.436
BundleFusion [6]Accuracy ↓0.1700.1000.1030.122X0.1210.123-
Completion ↓0.0880.0300.0380.057X0.2140.034-
F-score ↑0.3730.5710.4740.470X0.4420.527-
iMAP [30]Accuracy ↓X0.1350.2290.3650.2250.2330.139-
Completion ↓X0.0540.1030.2450.2910.1390.069-
F-score ↑X0.4450.3150.2380.1700.2550.416-
NICE-SLAM [46]Accuracy ↓0.3030.1190.1160.2160.1030.1560.4640.211
Completion ↓0.4560.0420.0700.1990.0890.1630.0450.152
F-score ↑0.2210.5540.4110.4020.4000.2730.4010.380
MonoGS [22]Accuracy ↓0.0900.0860.0830.0710.0690.0810.0890.081
Completion ↓0.2460.1220.1040.2380.1930.1570.1700.175
F-score ↑0.2900.2990.3460.3670.2590.3740.2890.318
ToF-SLAM [19]Accuracy ↓0.0810.0680.0670.0790.0780.1130.1210.087
Completion ↓0.0710.0410.0450.0620.1220.0850.0330.066
F-score ↑0.5590.6610.6460.6430.4960.5570.6560.604
ToF-Splatting (ours)Accuracy ↓0.0640.0310.0320.0410.0590.0340.0430.043
Completion ↓0.0950.0380.0290.0560.1310.0540.0460.064
F-score ↑0.5270.7910.8400.7100.3590.7790.6420.664
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.372, + 0.907, + 0.401 + ], + "angle": 0, + "content": "Table 2. ZJUL5 Mapping Results. We perform the mapping evaluation on the ZJUL5 dataset (7 indoor sequences) and report results of three metrics including accuracy (Acc.), completion (Comp.), and F-score. The failure cases are marked as \\( \\times \\)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.407, + 0.907, + 0.502 + ], + "angle": 0, + "content": "
Monocular CuesMulti-View CuesMetricKitchenSofaOfficeReceptionLiving RoomOffice2Sofa2Avg.
✓XATE ↓0.1030.0540.0460.0500.1100.0410.0550.066
F-score ↑0.2160.6750.5640.6450.2410.6110.5570.501
X✓ATE ↓0.2340.0240.0250.0370.1090.0340.0280.072
F-score ↑0.2840.7810.8260.7030.2530.7660.6830.614
✓✓ATE ↓0.0880.0240.0220.0410.1220.0290.0300.051
F-score ↑0.5270.7910.8400.7100.3590.7790.6420.664
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.506, + 0.907, + 0.548 + ], + "angle": 0, + "content": "Table 3. Multi-Frame Integration Ablation Study. Analysis of the impact of monocular and multi-view cues in the multi-frame integration step. Monocular cues are less impactful than multi-view ones, boosting performance in the challenging scenes where multi-view cues struggle most. Nonetheless, their combination yields the best results." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.564, + 0.481, + 0.704 + ], + "angle": 0, + "content": "
MethodLCInputDensityfr1/deskfr2/xyzfr3/officeAvg.
DROID-VOXRGB0.00%0.0520.1070.0730.077
MonoGSXRGB0.00%0.0380.0460.0350.040
ToF-SplattingXRGB-D0.02%0.0300.0220.0300.027
0.04%0.0270.0190.0210.022
ESLAMXRGB-D100%0.0250.0110.0240.020
Co-SLAMXRGB-D100%0.0240.0170.0240.022
CG-SLAMXRGB-D100%0.0240.0120.0240.020
RTG-SLAMXRGB-D100%0.0170.0040.0110.011
MonoGSXRGB-D100%0.0150.0140.0150.015
ORB-SLAM2✓RGB0.00%0.0190.0060.0240.016
ORB-SLAM2✓RGB-D100%0.0160.0400.0100.010
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Table 4. Results on TUM RGB-D dataset. Top: competitors w/o loop-closure (LC); middle: most representative competitor with \\(100\\%\\) density; bottom: competitors w/ loop-closure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.483, + 0.832 + ], + "angle": 0, + "content": "Notably, the other methods – like [6, 36] – may provide good performance on specific scenes but demonstrate unreliability on average usually failing in the particularly challenging Kitchen and Living Room scenes. In Figure 3, we show Office, Reception, and Sofa meshes on the left." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.258, + 0.862 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In this section, we perform additional experiments aimed at studying the impact of different factors on ToF-Splatting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Multi-Frame Integration. In Table 3, we ablate the main components of the multi-frame integration and assess their contribution to the whole SLAM framework. By retaining monocular cues alone at the expense of multi-view geometry, we observe severe drops in both tracking and mapping accuracy, confirming the paramount importance of the latter; monocular cues alone are effective only where multiview cues are ineffective – e.g., in the absence of texture, as occurs mostly in the Kitchen scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.809 + ], + "angle": 0, + "content": "TUM RGB-D. In Table 4, we simulate sparse ToF data by sampling \\(0.02\\%\\) and \\(0.04\\%\\) depth points, and compare ToF-Splatting with existing RGB and RGB-D methods. With as few as \\(0.02\\%\\) of points, ToF-Splatting largely improves over RGB methods [22, 32], approaching the \\(100\\%\\) density RGB-D approach of the most representative competitor [22] at only \\(0.04\\%\\) density." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Replica: Input Depth Sparsity. We conduct additional studies on the Replica dataset, for which Figure 4 provides a qualitative overview of our 3D reconstructions. ToF-Splatting enables performing SLAM with very few sparse points – the 64 sparse points provided by a \\(8 \\times 8\\) sparse account only for \\(0.02\\%\\) of a \\(640 \\times 480\\) image. We here" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28350" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.091, + 0.862, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.908, + 0.294 + ], + "angle": 0, + "content": "Figure 4. Replica Qualitatives. We provide qualitative results on Replica [28] to demonstrate the generalization capabilities of our method. On the left from top to bottom, meshes obtained respectively on scenes Office2 and Room2. On the right, details from the scene Room0. ToF-Splating delivers accurate details and allows for nice photometric and depth rendering." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.306, + 0.462, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.41, + 0.486, + 0.466 + ], + "angle": 0, + "content": "Figure 5. Impact of depth sparsity. We test on Replica [28] with different simulated depth sparsity levels to assess the capability to exploit higher input densities. MAE and ATE smoothly decrease, whereas rendering metrics appear to be less affected." + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.472, + 0.462, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.632 + ], + "angle": 0, + "content": "Figure 6. Impact of depth noise. We test on Replica [28] injecting different amounts of noise \\(\\xi \\in [0.00, 0.01, 0.05, 0.10]\\). ToF-Splatting demonstrates to be effective at dealing with noise, with error increasing almost linearly with the injected amount of noise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.732 + ], + "angle": 0, + "content": "explore the impact of higher sparse points density on its overall performance. We simulate on Replica densities of \\(\\{0.02\\%, 0.04\\%, 0.08\\%, 0.16\\% \\}\\) corresponding to about \\(\\{64, 128, 256, 512\\}\\) points. Figure 5 shows the performance variation on both tracking and mapping metrics using the aforementioned sparsities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Replica: Input Depth Noise. We study various noise levels impact in the input depth to assess our pipeline robustness. We model noise as additive but dependent on depth, injecting heteroscedastic Gaussian noise \\(\\mathcal{N}(d,\\epsilon d)\\) where \\(\\epsilon\\) modulates the variance of the noise source on Replica. Figure 6 shows that MAE, PSNR, and ATE all expose linear degradation trends as \\(\\epsilon\\) increases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Runtime. In Table 5 we measure the runtime of ToF-Splatting in comparison with the main baselines. The tracking time of ToF-Splatting already includes depth estimates with DoD, with average runtime \\(\\bar{t}_{\\mathrm{DoD}} \\approx 90 ms\\) (compris" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.304, + 0.906, + 0.387 + ], + "angle": 0, + "content": "
Method NameTracking TimeMapping Time
iMAP [30]101 ms448 ms
NICE-SLAM [46]470 ms1300 ms
MonoGS [22]1169 ms650 ms
ToF-SLAM [19]116 ms380 ms
ToF-Splatting (ours)614 ms865 ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.392, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Table 5. Runtime Comparison. We compare the runtime of ToF-Splatting and other learned baselines measuring their runtime for tracking and mapping. MonoGS [22] tracking time contains also the depth completion inference time [48]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.585 + ], + "angle": 0, + "content": "ing DepthAnything v2 for monocular cues inference [40]). To date, 3DGS-based mapping methods are not capable of real-time performances. Nevertheless, exploiting the fact that the mapping step accounts only for just a few frames (typically \\(4\\%\\) of the total in sequences of the ZJUL5 dataset [19]) and can be parallelized, the pipeline achieves 1.5 fps without further optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.909, + 0.713 + ], + "angle": 0, + "content": "Limitations. Even with state-of-the-art quality, ToF-Splatting manifests some issues that will be addressed in future developments. The most concerning is the runtime, which is too slow for real-time applications and requires memory and compute-intensive backpropagation at deployment. Even though fast 3DGS frameworks exist, such methodologies have not been transferred yet into the SLAM community, that being outside the scope of this work." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.748, + 0.634, + 0.763 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We presented the first 3DGS-based SLAM pipeline relying on sparse ToF depth sensing, as it provides accurate tracking and mapping from cheap and low-power cameras. Moreover, we showed how this result can be achieved through the integration of multi-view geometry, sparse depth data, and monocular cues, yielding an end-to-end 3DGS-based SLAM system. Finally, we assessed the effectiveness of our approach on the ZJUL5 and Replica datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28351" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yunsong Wang, and Gim Hee Lee. Vcr-gaus: View consistent depth-normal regularizer for gaussian surface reconstruction. arXiv preprint arXiv:2406.05774, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Andrea Conti, Matteo Poggi, Valerio Cambareri, and Stefano Mattoccia. Depth on demand: Streaming dense depth from a low frame-rate active sensor. In European Conference on Computer Vision (ECCV), 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.483, + 0.354 + ], + "angle": 0, + "content": "[4] Andrea Conti, Matteo Poggi, Valerio Cambareri, and S. Mattoccia. Range-agnostic multi-view depth estimation with keyframe selection. 2024 International Conference on 3D Vision (3DV), pages 1350–1359, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[6] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. BundleFusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4): 1, 2017. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[7] Tianchen Deng, Guole Shen, Tong Qin, Jianyu Wang, Wentao Zhao, Jingchuan Wang, Danwei Wang, and Weidong Chen. Plgslam: Progressive neural scene representation with local to global bundle adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19657-19666, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.483, + 0.649 + ], + "angle": 0, + "content": "[8] Zilong Dong, Guofeng Zhang, Jiaya Jia, and Hujun Bao. Keyframe-based real-time camera tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1538-1545. IEEE, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[9] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[10] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. KinectFusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User Interface Software and Technology, pages 559-568, 2011. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[11] Mohammad Mahdi Johari, Camilla Carta, and François Fleuret. Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17408-17419, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "for densergb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[14] Georg Klein and David Murray. Parallel tracking and mapping for small AR workspaces. In Proceedings of the IEEE and ACM International Symposium on Mixed and Augmented Reality, pages 225-234. IEEE, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[15] Yanyan Li, Nikolas Brasch, Yida Wang, Nassir Navab, and Federico Tombari. Structure-slam: Low-drift monocular slam in indoor environments. IEEE Robotics and Automation Letters, 5(4):6583-6590, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[16] Yijin Li, Xinyang Liu, Wenqian Dong, Han Zhou, Hujun Bao, Guofeng Zhang, Yinda Zhang, and Zhaopeng Cui. Deltar: Depth estimation from a light-weight tof sensor and rgb image. In European Conference on Computer Vision, 2022. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[17] Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R Oswald. Loopy-slam: Dense neural slam with loop closures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20363-20373, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[18] Haomin Liu, Guofeng Zhang, and Hujun Bao. Robust keyframe-based monocular SLAM for augmented reality. In Proceedings of the IEEE International Symposium on Mixed and Augmented Reality, pages 1-10. IEEE, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.585 + ], + "angle": 0, + "content": "[19] Xinyang Liu, Yijin Li, Yanbin Teng, Hujun Bao, Guofeng Zhang, Yinda Zhang, and Zhaopeng Cui. Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1-11, 2023. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[20] Gregor Luetzenburg, Aart Kroon, and Anders A. Bjørk. Evaluation of the Apple iPhone 12 Pro LiDAR for an Application in Geosciences. Scientific Reports, 11(1), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[21] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. 2024 International Conference on 3D Vision (3DV), pages 800-809, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[22] Hidenobu Matsuki, Riku Murai, Paul H.J. Kelly, and Andrew J. Davison. Gaussian splattering slam. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18039-18048, 2023. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision, pages 405-421. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Raul Mur-Artal and Juan D. Tardós. ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo, and RGB-D Cameras. IEEE Transactions on Robotics, 33(5): 1255–1262, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Xin Qiao, Matteo Poggi, Pengchao Deng, Hao Wei, Chenyang Ge, and Stefano Mattoccia. Rgb guided tof imag-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28352" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "ing system: A survey of deep learning-based methods. International Journal of Computer Vision, pages 1-38, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[26] Erik Sandström, Yue Li, Luc Van Gool, and Martin R Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433-18444, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[27] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.388 + ], + "angle": 0, + "content": "[28] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[29] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 573-580. IEEE, 2012. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[30] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.483, + 0.586 + ], + "angle": 0, + "content": "[31] Keisuke Tateno, Federico Tombari, Iro Laina, and Nassir Navab. Cnn-slam: Real-time dense monocular slam with learned depth prediction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6243-6252, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[32] Zachary Teed and Jia Deng. Droid-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. Advances in Neural Information Processing Systems, 34: 16558-16569, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[33] Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R. Oswald, and Matteo Poggi. How nerfs and 3d gaussian splattering are reshaping slam: a survey, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[34] Hengyi Wang, Jingwen Wang, and Lourdes Agapito. Coslam: Joint coordinate and sparse parametric encodings for neural real-time slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13293-13302, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[35] Yifan Wang, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38:1-14, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[36] Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, Andrew J Davison, and Stefan Leutenegger. ElasticFusion: Real-time dense SLAM and light source estimation. The International Journal of Robotics Research, 35(14):1697-1716, 2016. 1, 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[37] Xingming Wu, Zimeng Liu, Yuxin Tian, Zhong Liu, and Weihai Chen. Kn-slam: Keypoints and neural implicit encoding slam. IEEE Transactions on Instrumentation and Measurement, 73:1-12, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[38] Liu Xinyang, Li Yijin, Teng Yanbin, Bao Hujun, Zhang Guofeng, Zhang Yinda, and Cui Zhaopeng. Multi-modal neural radiance field for monocular dense slam with a lightweight tof sensor. In International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[39] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595-19604, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.906, + 0.331 + ], + "angle": 0, + "content": "[40] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. ArXiv, abs/2406.09414, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.906, + 0.374 + ], + "angle": 0, + "content": "[41] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. ArXiv, abs/2312.10070, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[42] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.906, + 0.486 + ], + "angle": 0, + "content": "[43] Guofeng Zhang, Jiaya Jia, Tien-Tsin Wong, and Hujun Bao. Recovering consistent video depth maps via bundle optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.906, + 0.557 + ], + "angle": 0, + "content": "[44] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimization for consistent 3d instant reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3727-3737, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.559, + 0.906, + 0.628 + ], + "angle": 0, + "content": "[45] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.906, + 0.712 + ], + "angle": 0, + "content": "[46] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. NICE-SLAM: Neural implicit scalable encoding for SLAM. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.714, + 0.906, + 0.782 + ], + "angle": 0, + "content": "[47] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R. Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. 2024 International Conference on 3D Vision (3DV), pages 42-52, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.785, + 0.906, + 0.826 + ], + "angle": 0, + "content": "[48] Yiming Zuo and Jia Deng. Ogni-dc: Robust depth completion with optimization-guided neural iterations. arXiv preprint arXiv:2406.11711, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.828, + 0.906, + 0.883 + ], + "angle": 0, + "content": "[49] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. Surface splatting. Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28353" + } + ] +] \ No newline at end of file diff --git a/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/93b627e5-bb27-4dc6-aa8c-e29320f5f5eb_origin.pdf b/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/93b627e5-bb27-4dc6-aa8c-e29320f5f5eb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..551266fc34c8d324e33f1de8ca98de000824f693 --- /dev/null +++ b/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/93b627e5-bb27-4dc6-aa8c-e29320f5f5eb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10d62f8991392f20e695ef5292359aaa25529b0e3c52bcb48ec026935a315411 +size 4084894 diff --git a/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/full.md b/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5ef577f303bfe625464624c8586ca4b88b5607e6 --- /dev/null +++ b/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/full.md @@ -0,0 +1,310 @@ +# ToF-Splatting: Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration + +Andrea Conti $^{1,2*}$ + +Matteo Poggi + +Valerio Cambareri + +Martin R. Oswald + +Stefano Mattoccia + +1University of Bologna, Italy 2Sony DepthSensing Solutions, Belgium 3University of Amsterdam, Netherlands + +![](images/154d8347800c509697f43c2698b4eac191b07da985beb92b68118ca4783670fb.jpg) +Figure 1. Overview of our ToF-Splatting method. Our method combines sparse ToF depth, multi-view geometry from a buffer of keyframes, and monocular cues (left) to perform into a unique end-to-end dense SLAM framework enabled by a Gaussian Splatting. + +# Abstract + +Time-of-Flight (ToF) sensors provide efficient active depth sensing at relatively low power budgets; among such designs, only very sparse measurements from low-resolution sensors are considered to meet the increasingly limited power constraints of mobile and AR/VR devices. However, such extreme sparsity levels limit the seamless usage of ToF depth in SLAM. In this work, we propose ToF-Splatting, the first 3D Gaussian Splatting-based SLAM pipeline tailored for using effectively very sparse ToF input data. Our approach improves upon the state of the art by introducing a multi-frame integration module, which produces dense depth maps by merging cues from extremely sparse ToF depth, monocular color, and multi-view geometry. Extensive experiments on both real and synthetic sparse ToF datasets demonstrate the advantages of our approach, as it achieves state-of-the-art tracking and mapping performances on reference datasets. + +# 1. Introduction + +Simultaneous Localization and Mapping (SLAM) [6, 8, 36, 43] entails the joint estimation of the current camera pose + +against a 3D scene representation (i.e., the map) and the update of this representation with data from the current camera view. It is a fundamental task with applications in AR/VR, robotics, and autonomous navigation [14, 18]. In recent years, SLAM witnessed a revolution [33] following the success of Neural Radiance Fields (NeRF) [23] and 3D Gaussian Splatting (3DGS) [13] as scene representations from the adjacent field of novel view synthesis. + +In this paper we focus on SLAM pipelines that leverage RGB-D video inputs, i.e., where color images and dense depth maps are collected by synchronized, accurate color and depth sensors. Beyond global scale, depth sensors provide accurate local geometry that grants better tracking and mapping quality. A common choice for active depth sensing are Time-of-Flight (ToF) cameras. Despite the increased accuracy provided by reliable depth measurements, the power consumption and cost of such sensors constrain their adoption to high-end industrial or automotive applications. A recent trend to favor ToF sensor integration in handheld consumer devices is to reduce the specifications of ToF sensors to provide low-resolution but more reliable depth measurements. Among other examples, LiDAR (an instance of ToF) appears in consumer mobile phones [20] and is known to feature up to at most 576 dots (interpolated to $256 \times 192$ resolution via depth completion). Other recent ToF sensors provide only up to 64 depth points but consume a re + +ported $200\mathrm{mW}$ [19]. In such settings a SLAM pipeline can expect to receive, e.g., $8 \times 8$ very noisy, very sparse depth measurements per frame. These specifications prevent their straightforward integration in SLAM systems, as low-resolution noisy depth is known to harm the accuracy of dense SLAM systems built with NeRF and not originally designed to deal with this scenario [30, 36, 46]. This is even more evident with the more recent 3DGS-based SLAM systems [22], where accurate and dense depth is paramount to achieve satisfactory results. + +To overcome these limitations, a straightforward strategy [38] would consist in recovering dense and accurate depth maps using a single-image depth completion framework [25]. However, models trained for this task are known to generalize poorly across domains and across input depth noise levels, potentially affecting the downstream SLAM pipeline accuracy and yielding both drifted camera trajectory and inaccurate mapping. Therefore, we explore 3DGS as an explicit representation that allows for injecting domain-specific knowledge and that is easily extended to, e.g., dynamic scenes [21] and accurate large mesh reconstruction [2, 9]. On the one hand, we argue that depth completion alone is insufficient to allow sparse ToF depth usage in SLAM systems. On the other, we acknowledge that even sparse and noisy depth measurements can bootstrap SLAM if properly assisted with geometry. Accordingly, inspired by frame-to-frame SLAM systems relying on pre-trained trackers for localization, we believe that multi-frame integration fusing the noisy depth measurements with multiview geometry across a small set of local frames could supply the dense SLAM system with the supervision necessary to fill this gap. + +In this paper, we propose ToF-Splatting, a novel 3DGS-based SLAM system enabling dense reconstruction from sparse ToF sensors. ToF-Splatting harmoniously alternates tracking and mapping, with each one benefiting from the other. The former is carried out through backpropagation during the optimization of the 3DGS model on color images; the latter exploits a pre-trained network [3] as a multi-frame integration module, combining sparse depth measurements with multi-view geometry according to a buffer of keyframes. For each of the keyframes, we use the camera pose estimated through tracking and providing dense depth maps to improve the supervision given to the 3DGS model itself. Figure 1 shows an overview of the setting in which ToF-Splatting operates, as well as the quality of the 3D reconstructions it yields. + +ToF-Splatting is tested on the real-world ZJUL5 dataset [19], where we measure tracking and mapping quality and find the superiority of our framework with respect to existing solutions built over sparse ToF sensors [38]. Moreover, standard benchmarks such as Replica [28] are also provided and confirm the effectiveness of our mapping strat + +egy. Our main contributions are as follows: + +- We propose ToF-Splatting, the first 3DGS-based SLAM system suited for sparse ToF sensors. +- We introduce a novel multi-frame integration method combining sparse depth, multi-view geometry, and monocular cues for robust noise and outlier handling. +- We establish a new state-of-the-art for dense SLAM systems built on top of sparse ToF depth sensors. + +# 2. Related Work + +We briefly review the literature relevant to our work, referring to [33] for a detailed overview of the latest advances. + +Traditional SLAM. Traditional visual SLAM pipelines, or the more recent ones introducing deep modules [15, 31, 32], usually perform tracking in frame-to-frame fashion, based on the visual features extracted from past frames either using handcrafted feature extractors [24] or deep networks [31, 32]. These systems are usually defined by four main steps: tracking, mapping, global bundle adjustment, and loop closure. Among the different representations used by traditional SLAM frameworks, depth points [1, 24], surfels [27], and volumetric representations [6] allows for achieving globally consistent 3D reconstruction. + +NeRF-based SLAM. The advent of NeRF in the field of novel view synthesis also reached adjacent research areas, as well as SLAM a few years later. NeRF-based SLAM systems estimate camera poses and reconstruct dense meshes by modeling scenes via MLPs. iMAP [30] and NICE-SLAM [45] are the first SLAM frameworks following this approach, with the latter introducing dense feature grids supporting the MLPs. Point-SLAM [26] and Loopy-SLAM [17] switched feature grids with neural point embeddings, allowing for higher flexibility and correcting or adjusting local maps after global optimization, yet with slow processing prohibiting deployment in robotics. ESLAM [11] and Co-SLAM [34] deploy tri-planes and hash grid embeddings, improving both processing speed and reconstruction accuracy, with PLGSLAM [7] further improving the representation capacity in large indoor scenes. GO-SLAM [44] and KN-SLAM [37] implement frame-to-frame systems thanks to external trackers, yielding a good trade-off between tracking/mapping accuracy and frame rate. + +3DGS-based SLAM. More recent advances in novel view synthesis brought to the development of new SLAM systems using 3D Gaussian Splatting [13] to represent the mapped scene. Several concurrent frameworks emerged built over 3DGS, such as MonoGS [22], GaussianSLAM [42], SplaTAM [12], and GS-SLAM [39]. Different from NeRF-based SLAM systems, these frameworks allow for higher interpretability and higher rendering speed. + +Sparse ToF-based SLAM. Only lately, the use of more compact and energy-efficient depth sensors has gained attention in the SLAM research community. ToF-SLAM [38] + +![](images/d20e17a8a638c546de8098bf09d21117f7d3ac261c3cba4a4d1b227262971f0f.jpg) +Figure 2. ToF-Splatting Pipeline. Our method involves three main modules: a Tracking frontend estimating camera poses, a Multi-Frame Integration module that predicts dense depth maps from sparse ToF measurements and multi-view geometry, and a Mapping backend modeling the 3D scene representation via 3D Gaussian Splatting. + +represents the only attempt to exploit "lightweight", sparse depth measurements within a dense SLAM system, aided by a depth completion model [16]. However, depth completion alone fails to deal with noise in the sparse input depth, a shortcoming we aim to specifically overcome within our ToF-Splatting framework, the first one based on 3DGS and suited for sparse ToF sensors. + +# 3. Pipeline Architecture + +In this section, we present the core components of our framework. ToF-Splatting is a frame-to-model SLAM system conceptually divided into three modules: $(i)$ a mapping backend, $(ii)$ a tracking frontend, and $(iii)$ a peculiar multi-frame integration module yielding dense depth maps based on a subset of the keyframe graph, the current RGB frame, and the current sparse ToF depth map. + +Figure 2 provides an overview of our pipeline. Our framework is characterized by three heterogeneous modules that are tightly related and influence each other. The tracking module stands as the most influencing element in the pipeline, as it selects the keyframes used by both the mapping and multi-frame integration parts, moreover, it computes the poses used for multi-frame depth estimation. Nonetheless, it relies on the quality of the mapping step to perform correct ego-motion estimation and exploits the depth cues to properly behave when photometric information is lacking or ambiguous. + +The depth perception part is mainly sustained by the tracking step and highly influences the mapping part since it is used to seed new Gaussians and as supervision. Finally, the mapping phase not only creates a unique smooth representation but also defines the rendered opacity that is used for keyframe selection. These connections lead to a smoothly integrated framework that not only exploits a multi-frame backbone but more importantly, defines an effective way to use such information in the SLAM scenario. + +# 3.1. Background: 3D Gaussian Splatting + +3D Gaussian Splitting [13] (3DGS) represents one of the latest advances in novel view synthesis, fitting a dynamically defined set of multivariate Gaussian distributions over the scene to enable image rendering from arbitrary viewpoints. Each Gaussian $G(\mathbf{x})$ is parametrized by its mean $\boldsymbol{\mu} \in \mathbb{R}^3$ and covariance matrix $\boldsymbol{\Sigma} \in \mathbb{R}^{3 \times 3}$ , reading + +$$ +G (\mathbf {x}) = \exp \left(- \frac {1}{2} (\mathbf {x} - \boldsymbol {\mu}) ^ {\top} \boldsymbol {\Sigma} ^ {- 1} (\mathbf {x} - \boldsymbol {\mu})\right) \tag {1} +$$ + +Moreover, opacity $o \in \mathbb{R}$ and RGB color $\mathbf{c} \in \mathbb{R}^3$ are assigned to each Gaussian. + +Given a set of images from different viewpoints all around the scene, the Gaussian parameters are optimized using the splatting technique [35]. With such a method, 3D Gaussians are projected through rasterization over the 2D image plane, containing a 2D Gaussian distribution + +$$ +\boldsymbol {\mu} ^ {2 D} = \pi \left(\mathbf {T} _ {w c} \bar {\boldsymbol {\mu}}\right) \quad \boldsymbol {\Sigma} ^ {2 D} = \mathbf {J} \mathbf {R} _ {w c} \boldsymbol {\Sigma} \mathbf {R} _ {w c} ^ {\top} \mathbf {J} ^ {\top} \tag {2} +$$ + +where $\bar{\pmb{\mu}}$ is $\pmb{\mu}$ in homogeneous coordinates, $\mathbf{T}_{wc} \in \mathbb{R}^{4 \times 4}$ , $\mathbf{R}_{wc} \in \mathbb{R}^{3 \times 3}$ are respectively the world-to-camera transformation and its rotational component, $\pi$ is the projection matrix and $\mathbf{J} \in \mathbb{R}^{2 \times 3}$ is the Jacobian of the projective transformation [49]. The color $C$ of a pixel is then given by the combination of the $M$ weighted Gaussians, i.e., + +$$ +\mathbf {C} = \sum_ {i \in M} \mathbf {c} _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}) \tag {3} +$$ + +with $\alpha_{i} = o_{i}G(\mathbf{x}_{i})$ . Since $\pmb{\Sigma}$ is positive semi-definite, it is parametrized with a diagonal scaling matrix $\mathbf{S}$ and a rotation matrix $\mathbf{R}$ , so that $\pmb{\Sigma} = \mathbf{RSS}^{\top}\mathbf{R}^{\top}$ . $\mathbf{R}$ is internally represented with quaternions. + +# 3.2. Multi-Frame Integration + +Dense depth cues are pivotal for proper initialization of the Gaussians position $\pmb{\mu} \in \mathbb{R}^3$ , as well as to provide supervision to both tracking and mapping threads - as already + +known in the literature [22, 47]. However, sparse ToF sensors alone are insufficient for this purpose, because of the very sparse and noisy depth measurements they provide (e.g., only 64 points for a sensor such as that used in [19]). To overcome this limitation, ToF-SLAM [19] exploits an auxiliary completion model, DELTAR [16], to recover dense depth maps out of ToF measurements. However, depth completion is highly dependent on the quality of the sparse depth in input, rapidly degrading the accuracy of the densified depth maps in the presence of noise. + +In this paper, we follow a different path: inspired by frame-to-frame SLAM systems deploying an external tracker, we believe that multi-frame integration can compensate for the noise in the ToF measurements by exploiting both single-view and multi-view geometry cues on a set of local frames, given the camera poses estimated by the tracking frontend. We choose to extend the Depth on Demand (DoD) framework [3] for this purpose. + +Multi-Frame Integration. DoD [3] integrates sparse depth and two-view stereo cues for dense depth prediction. Specifically – given a frame $\mathbf{F}^k$ – it processes the associated RGB image $I_k$ (the target view), another RGB view $I_j$ from a different frame $\mathbf{F}^j$ (a source view), the relative pose $P^{jk}$ between the two, and the sparse depth points $H^k$ to predict a dense metric depth map $D^k$ . Differently from monocular depth prediction, DoD is aware of the scene scale either from $H^k$ or $P^{jk}$ , simplifying the problem of integrating scale-inconsistent geometries in the SLAM frontend. However, DoD is limited to two-view processing: this may lead to suboptimal results in the SLAM scenario, where an abundant amount of keyframes largely overlap and cover the same portion of the scene. Thus, we extend [3] to multiview processing. This is achieved by observing that the predicted depth map $D^k$ depends on the source view $I_j$ through its relative pose $P^{jk}$ only. DoD iteratively updates $D^k$ for multiple iterations, exploiting this intrinsic property we can integrate multiple views $(I_j, P^{jk})$ using a different one at each iteration [4]. This approach allows smooth integration of multi-view cues avoiding order dependency. To select the source views $I_j$ , at each prediction we order the keyframes selected by the SLAM pipeline according to their relative pose similarity and select the first $N$ frames with baseline distance close to $b = 15\mathrm{cm}$ to ensure enough parallax. The model is trained from scratch on ScanNetv2 [5] and following the protocol of [3]. + +Monocular Cues Integration. As shown in [3], DoD heavily relies on both geometry and depth measurements, thus struggling at generalizing in challenging scenarios where the two lack. This may happen in cases where the tight field of view, noisy sparse points, and textureless surfaces hamper multi-view matching. To mitigate such limitations, we integrate explicit monocular cues into DoD by feeding its monocular encoder with $I_{k}$ and the normalized depth + +map $\tilde{D}^k\in [0,1]^{H\times W}$ obtained through the single-image depth estimator Depth Anything v2 [40]. This approach allows the injection of a robust bias, avoiding issues related to monocular estimation, e.g., slanted surfaces or noisy scale estimates, and delegating to DoD their handling. These cues are included while retraining on ScanNetv2 [5]. + +Outlier Handling. Furthermore, the noisy depth measurements in input to DoD can severely affect its accuracy. Therefore, we explicitly focus on dealing with outliers in the input sparse depth, that may be perceived by ToF devices over specific surfaces. Such errors are particularly evident in cheap sensors and may largely prejudice the information supplied by the few points they measure. Purposely, we deploy a simple yet effective method to identify such outliers by exploiting multi-view geometry again. Whenever depth needs to be predicted for a new frame, DoD predicts a depth map $\hat{D}^k$ without processing the ToF measurements. Then, the $\ell_1$ error between the latter and the depth measured by the ToF sensor $H^k$ is computed and finally sparse measures with error higher than a given quantile $q$ are discarded, and depth is predicted again by DoD by also processing the filtered sparse depth $\tilde{H}^k$ this time. This way, we exploit multi-view cues to extract a clean depth prior to filtering out inconsistent sparse depth measures, and then we integrate the remaining points to improve the original prediction. Even without processing the sparse depth, DoD still predicts metric depth, as the metric scale is enforced by the relative camera poses. + +# 3.3. Tracking Frontend + +The tracking part of ToF-Splatting estimates the ego-motion of the camera for each new frame $\mathbf{F}^k$ and builds a keyframe graph $\mathbf{G} = \{\mathbf{KF}^k\} \subset \{\mathbf{F}^k\}$ containing a set of meaningful frames for mapping and multi-frame depth estimation. + +Pose Estimation. We initialize each new frame $\mathbf{F}^k$ of our frame-to-model tracker with the pose of the previous frame $\mathbf{F}^{k - 1}$ and then minimizing the tracking loss with respect to the relative pose between $\mathbf{F}^k$ and $\mathbf{F}^{k - 1}$ , following [22]. The tracking loss consists of two terms, $L_{\mathrm{rgb}}$ and $L_{\mathrm{depth}}$ , respectively accounting for photometric and geometric errors + +$$ +L _ {\text {t r a c k}} = \lambda_ {\text {t r a c k}} L _ {\text {r g b}} + (1 - \lambda_ {\text {t r a c k}}) L _ {\text {d e p t h}} \tag {4} +$$ + +$$ +L _ {\mathrm {r g b}} = \left\| I _ {k} - \hat {I} \left(\mathbf {F} ^ {k}\right) \right\| _ {1} \quad L _ {\mathrm {d e p t h}} = \left\| D ^ {k} - \hat {D} \left(\mathbf {F} ^ {k}\right) \right\| _ {1} \tag {5} +$$ + +where $I_{k}$ is the RGB image from $\mathbf{F}^k$ , $\hat{I}(\cdot)$ and $\hat{D}(\cdot)$ denote the rendered RGB and Depth and, $D^{k}$ is the depth prediction produced by the multi-frame integration module. + +Unlike other methods, the tracking process happens in two stages. Initially, for a fixed number of steps $\eta_{\mathrm{rgb}}$ only the photometric loss $L_{\mathrm{rgb}}^T$ is optimized. Then, such a pose is used for multi-frame integration, as described in Sec. 3.2, to generate $D^{k}$ . Finally, both losses in Eq. (5) are used for $\eta_{\mathrm{rgbd}}^T$ further steps. + +Keyframe Selection Policy. We deploy a simple yet effective keyframe selection policy that considers i) novel content to be mapped and ii) instabilities in the already mapped areas. Unlike other methods using complex approaches – e.g., frustum intersection, point clouds and pose analysis [22, 41] – we observe that the rendered opacity for a novel frame contains low opacity values where either Gaussians are not present or unstable. The first case highlights the presence of novel areas of the environment to be mapped, the second happens due to the pruning procedure involved in the mapping step (see Sec. 3.4). We set a frame $\mathbf{F}^k$ as a keyframe $\mathbf{K}\mathbf{F}^k$ if its novelty factor $\nu^k$ is higher than a threshold $\nu^{th}$ , where we define $\nu^k$ as + +$$ +\nu^ {k} = \frac {\sum_ {i \in \mathbf {K F} ^ {k}} \mathbf {M} \left(\mathbf {K F} _ {i} ^ {k}\right)}{\sum_ {i \in \mathbf {K F} ^ {k}} \left(1 - \mathbf {M} \left(\mathbf {K F} _ {i} ^ {k}\right)\right)} \tag {6} +$$ + +$$ +\mathbf {M} \left(\mathbf {K F} _ {i} ^ {k}\right) = \left\{ \begin{array}{l l} 1 & \text {i f} \hat {O} \left(\mathbf {K F} _ {i} ^ {k}\right) < \sigma \\ 0 & \text {o t h e r w i s e} \end{array} \right. \tag {7} +$$ + +with $\hat{O} (\mathbf{K}\mathbf{F}_i^k)$ being the rendered opacity at pixel $i$ , and $\mathbf{M}(\mathbf{K}\mathbf{F}_i^k)$ a binary uncertainty map. + +This approach allows for skipping several frames when the camera moves slowly or focusing deeply on areas where the ToF-Splatting struggles the most, by mapping it consecutively. Nonetheless, we enforce mapping after skipping a certain number of frames to avoid mapping too few frames. + +# 3.4. Mapping Backend + +Whenever a new keyframe $\mathbf{K}\mathbf{F}^k$ is identified, mapping is triggered. This phase seeks to embed the new frame in the global 3DGS model, involving the following two steps. + +Initialization. In this step, new Gaussians are seeded in the 3DGS model. To limit the size of the model and reduce outliers, only areas where the rendered uncertainty $\mathbf{M}(\mathbf{KF}^k)$ is high are seeded with new points, with the new Gaussians also being randomly downsampled by a constant factor. Each Gaussian is initialized using depth $D^{k}$ predicted in the multi-frame integration step, with color from the RGB frame $I_{k}$ , and scale initialized with a constant value. Unlike other 3DGS pipelines [22, 41], we prove this approach effective and particularly efficient, replacing any complex 3D heuristics to a simple use of rendered opacity $O(\cdot)$ , already a side-product of the rasterization process. + +Optimization. Firstly, ToF-Splating smoothly integrates the new information in the global model and secondly tunes the existing representation. This is achieved by optimizing the new keyframe $\mathbf{KF}^k$ and a random subset of 5 keyframes among the last $N$ keyframes $\{\mathbf{KF}^j \mid j \geq \max(k - N, 0)\}$ . These are optimized for a fixed number of steps $\eta_{\mathrm{rgbd}}^M$ . Moreover, we enforce a sampling rate of $60\%$ for $\mathbf{KF}^k$ , with the remaining $40\%$ for the remaining $\mathbf{KF}^j$ to avoid forgetting. During this step, gradients are also back-propagated to the + +camera poses of the sampled keyframes, thus acting as a global bundle adjustment. Finally, we prune Gaussians with opacities lower than 0.5 every $\eta_{\mathrm{rgbd}}^{M} / 2$ steps. In addition to the terms in Eq. (5), we optimize structural similarity as in Eq. (8) and, following [2], the normals as in Eq. (9), i.e., + +$$ +L _ {\mathrm {d s s i m}} = \frac {1 - \operatorname {S S I M} \left(I _ {k} , \hat {I} \left(\mathbf {K F} ^ {k}\right)\right)}{2} \tag {8} +$$ + +$$ +L _ {\text {n o r m a l s}} = \| N ^ {k} - \hat {N} \| _ {1} + \overline {{\mathbf {1} - \langle N ^ {k} , \hat {N} \rangle}} \tag {9} +$$ + +with $\hat{N} \coloneqq \bar{\nabla}\hat{D}(\mathbf{K}\mathbf{F}^k)$ , $N^k \coloneqq \bar{\nabla}D^k$ denoting the normals estimated via 3D gradients $\bar{\nabla}$ for the rendered and multiframe predicted depth respectively, $\langle \cdot, \cdot \rangle$ their inner product, and $\overline{\cdot}$ the average. $L_{\mathrm{normals}}$ is masked on edges to preserve sharpness, as identified by running a Sobel filter. + +Finally, we introduce a regularization loss $L_{\mathrm{iso}}$ to penalize elongated Gaussians (i.e., promote their isotropy) that usually leads to rendering artifacts as already observed in [22]. Eq. (10) defines $L_{\mathrm{iso}}$ , where $\frac{\mathrm{diag}(\cdot)}{\mathrm{diag}(\mathbf{S}_j)}$ extracts the diagonal values from its argument, and $\overline{\mathrm{diag}(\mathbf{S}_j)}$ is the average of the resulting vector, i.e., of the elements of the $j$ -th diagonal scale matrix $\mathbf{S}_j$ as defined in Sec. 3.1. Thus, + +$$ +L _ {\mathrm {i s o}} = \frac {1}{| \mathcal {G} |} \sum_ {j} ^ {| \mathcal {G} |} \| \operatorname {d i a g} \left(\mathbf {S} _ {j}\right) - \overline {{\operatorname {d i a g} \left(\mathbf {S} _ {j}\right)}} \cdot \mathbf {1} _ {3 \times 1} \| _ {1} \tag {10} +$$ + +The final mapping loss aggregates the aforementioned loss terms with corresponding weights as follows: + +$$ +\begin{array}{l} L _ {\text {m a p}} = \lambda_ {\text {m a p}} \left(\lambda_ {\text {v i s u a l}} L _ {\text {r g b}} + (1 - \lambda_ {\text {v i s u a l}}) L _ {\text {d s s i m}}\right) \tag {11} \\ + \left(1 - \lambda_ {\text {m a p}}\right) L _ {\text {d e p t h}} + \lambda_ {\text {n o r m a l s}} L _ {\text {n o r m a l s}} + \lambda_ {\text {i s o}} L _ {\text {i s o}} \\ \end{array} +$$ + +# 4. Experiments + +This section assesses both the tracking and mapping performance of ToF-Splatting on the following datasets. + +ZJUL5. ZJUL5 [19] is the only existing public-domain real dataset providing sparse depth data from a low-resolution, low-cost ToF sensor. A VL53L5CX ToF sensor is assembled on a calibrated rig with an Intel RealSense 435i to provide dense depth ground truth. This benchmark comprises 7 diverse indoor scene recordings. Ground truth 3D meshes and camera poses are obtained following the ScanNet protocol [5]. This sensor provides $8 \times 8$ depth points by counting the number of photons returned in each discretized time range. Of the at most 64 depth points provided, many are extremely noisy and introduce large outliers. This challenging real-world dataset provides real measurement noise that is not accurately represented in mainstream datasets. + +TUM RGB-D. This is a real dataset [29] providing indoor sequence captures and widely used as benchmark for RGB-D SLAM pipelines. The depth stream provides dense VGA depth maps from a high-resolution Kinect v1 depth sensor. + +
MethodKitchenSofaOfficeReceptionLiving RoomOffice2Sofa2Avg.
KinectFusion [10]X0.1460.2090.157X0.3210.1250.192
iMAP [30]X1.6580.3380.6480.6790.3440.2140.647
NICE-SLAM [46]0.7450.1440.1550.2510.2890.2280.4210.319
BundleFusion [6]0.1760.1020.1350.101X0.1630.1200.132
ElasticFusion [36]0.2530.1100.0700.1930.5300.1210.1460.203
MonoGS [22]0.2310.0320.0410.0440.1530.0350.0730.087
ToF-SLAM [19]0.1130.0810.0560.1140.2000.1010.0850.107
ToF-Splatting (ours)0.0880.0240.0220.0410.1220.0290.0300.051
+ +Table 1. ZJUL5 Tracking Results. We show here the tracking performance on the ZJUL5 dataset [19] dataset with the ATE RMSE [m] (↓) metric on the 8 sequences available where $\times$ indicates lost tracking. ToF-Splatting provides the better performance by a good margin on each sequence, demonstrating the optimal capability of our proposal to track the camera ego-motion accurately. + +Replica. Replica [28] is a synthetic dataset providing extremely realistic indoor sequences. It provides dense depth maps, which we use at different densities to perform experiments that prove the robustness and generalization capabilities of ToF-Splatting in challenging scenarios. + +Baselines. We compare our method with both learning-based [19, 30, 46] and traditional pipelines [6, 10, 36]. Following ToF-SLAM [19], baselines requiring RGB-D frames use DELTAR [16] to densify the sparse ToF measurements. Moreover, we adapt MonoGS [22] densifying depth with the state-of-the-art depth completion framework OGNI-DC [48] for a Gaussian Splatting SLAM baseline. + +Implementation Details. In all our experiments, we use $N = 4$ source views $I_{j}$ (i.e., keyframe buffer elements) for the multi-view integration module and a baseline $b = 15\mathrm{cm}$ for selecting keyframes. We use a quantile $q = 0.75$ to filter outliers that occur frequently in the sparse ToF depth of the ZJUL5 dataset [19]. Concerning tracking, we use a novelty factor threshold of $\nu_{th} = 0.1$ , $\sigma = 0.98$ . The total tracking loss uses $\lambda_{\mathrm{track}} = 0.9$ , a total number of $\eta_{\mathrm{rgb}}^T = 30$ and $\eta_{\mathrm{rgbd}}^T = 70$ steps are performed at each iteration. For mapping, we set the loss weights $\lambda_{\mathrm{map}} = 60$ , $\lambda_{\mathrm{visual}} = 0.20$ , $\lambda_{\mathrm{normals}} = 0.01$ , $\lambda_{\mathrm{iso}} = 1.0$ and the number of steps per iteration $\eta_{\mathrm{rgbd}}^M = 60$ . We test on a single RTX3090. + +# 4.1. Comparison with the State of the Art + +In this section, we compare ToF-Splatting with existing methods in terms of tracking and mapping performances. + +Tracking. In Table 1, we report the tracking performance by ToF-Splatting and various baselines on each sequence of the ZJUL5 dataset [19]. Following common practice we use the Absolute Trajectory Error metric [29], which directly compares the global consistency of a trajectory. While various frameworks [6, 10, 30] fail on the most challenging scenes, ToF-Splatting delivers accurate and consistent tracking. This is evident from the good margin it gains against the baselines. In Figure 3, we show the 3D trajectories on a subset of the scenes, with error below $3\mathrm{cm}$ . + +Mapping. In Table 2, we report the mapping performance by ToF-Splatting compared with the other baselines. For each scene, we collect the predicted pose of each frame + +![](images/80030511516d6a512ed1b075e61b8fe373f7acf9c8973d7621a95480ea3c4880.jpg) +Figure 3. Qualitative results on the ZJUL5 dataset [19]. We show meshes obtained by fusing rendered depth maps with TSDF and marching cubes (left), and 3D trajectories (right) on 3 scenes selected from the ZJUL5 dataset [19]. + +and render depth and color from 3DGS. Then, we perform TSDF integration and extract the final mesh through marching cubes. Following [19], the meshes are then evaluated by computing Accuracy, Completion, and F-score versus the ground truth meshes. Accuracy and Completion respectively evaluate the mean distance between each predicted vertex and the nearest ground truth one and vice versa. The F-score takes into account both Accuracy and Completion with an aggregate metric. ToF-Splating delivers consistent results for mapping, always achieving the highest accuracy and significantly exceeding the baselines on average and specifically on the Sofa, Office, and Reception scenes which account for five over seven sequences. [19] is the second-best method and the main competitor in mapping. + +
MethodMetricKitchenSofaOfficeReceptionLiving RoomOffice2Sofa2Avg.
KinectFusion [10]Accuracy ↓X0.1900.2110.261X0.2670.135-
Completion ↓X0.0480.0460.064X0.0780.064-
F-score ↑X0.2780.2880.285X0.2740.381-
ElasticFusion [36]Accuracy ↓0.0920.1350.0840.2970.1510.0960.1220.140
Completion ↓0.0650.0480.0820.3050.2160.1470.0470.130
F-score ↑0.5530.4200.5290.2740.3820.4160.4810.436
BundleFusion [6]Accuracy ↓0.1700.1000.1030.122X0.1210.123-
Completion ↓0.0880.0300.0380.057X0.2140.034-
F-score ↑0.3730.5710.4740.470X0.4420.527-
iMAP [30]Accuracy ↓X0.1350.2290.3650.2250.2330.139-
Completion ↓X0.0540.1030.2450.2910.1390.069-
F-score ↑X0.4450.3150.2380.1700.2550.416-
NICE-SLAM [46]Accuracy ↓0.3030.1190.1160.2160.1030.1560.4640.211
Completion ↓0.4560.0420.0700.1990.0890.1630.0450.152
F-score ↑0.2210.5540.4110.4020.4000.2730.4010.380
MonoGS [22]Accuracy ↓0.0900.0860.0830.0710.0690.0810.0890.081
Completion ↓0.2460.1220.1040.2380.1930.1570.1700.175
F-score ↑0.2900.2990.3460.3670.2590.3740.2890.318
ToF-SLAM [19]Accuracy ↓0.0810.0680.0670.0790.0780.1130.1210.087
Completion ↓0.0710.0410.0450.0620.1220.0850.0330.066
F-score ↑0.5590.6610.6460.6430.4960.5570.6560.604
ToF-Splatting (ours)Accuracy ↓0.0640.0310.0320.0410.0590.0340.0430.043
Completion ↓0.0950.0380.0290.0560.1310.0540.0460.064
F-score ↑0.5270.7910.8400.7100.3590.7790.6420.664
+ +Table 2. ZJUL5 Mapping Results. We perform the mapping evaluation on the ZJUL5 dataset (7 indoor sequences) and report results of three metrics including accuracy (Acc.), completion (Comp.), and F-score. The failure cases are marked as $\times$ . + +
Monocular CuesMulti-View CuesMetricKitchenSofaOfficeReceptionLiving RoomOffice2Sofa2Avg.
✓XATE ↓0.1030.0540.0460.0500.1100.0410.0550.066
F-score ↑0.2160.6750.5640.6450.2410.6110.5570.501
X✓ATE ↓0.2340.0240.0250.0370.1090.0340.0280.072
F-score ↑0.2840.7810.8260.7030.2530.7660.6830.614
✓✓ATE ↓0.0880.0240.0220.0410.1220.0290.0300.051
F-score ↑0.5270.7910.8400.7100.3590.7790.6420.664
+ +Table 3. Multi-Frame Integration Ablation Study. Analysis of the impact of monocular and multi-view cues in the multi-frame integration step. Monocular cues are less impactful than multi-view ones, boosting performance in the challenging scenes where multi-view cues struggle most. Nonetheless, their combination yields the best results. + +
MethodLCInputDensityfr1/deskfr2/xyzfr3/officeAvg.
DROID-VOXRGB0.00%0.0520.1070.0730.077
MonoGSXRGB0.00%0.0380.0460.0350.040
ToF-SplattingXRGB-D0.02%0.0300.0220.0300.027
0.04%0.0270.0190.0210.022
ESLAMXRGB-D100%0.0250.0110.0240.020
Co-SLAMXRGB-D100%0.0240.0170.0240.022
CG-SLAMXRGB-D100%0.0240.0120.0240.020
RTG-SLAMXRGB-D100%0.0170.0040.0110.011
MonoGSXRGB-D100%0.0150.0140.0150.015
ORB-SLAM2✓RGB0.00%0.0190.0060.0240.016
ORB-SLAM2✓RGB-D100%0.0160.0400.0100.010
+ +Table 4. Results on TUM RGB-D dataset. Top: competitors w/o loop-closure (LC); middle: most representative competitor with $100\%$ density; bottom: competitors w/ loop-closure. + +Notably, the other methods – like [6, 36] – may provide good performance on specific scenes but demonstrate unreliability on average usually failing in the particularly challenging Kitchen and Living Room scenes. In Figure 3, we show Office, Reception, and Sofa meshes on the left. + +# 4.2. Ablation Studies + +In this section, we perform additional experiments aimed at studying the impact of different factors on ToF-Splatting. + +Multi-Frame Integration. In Table 3, we ablate the main components of the multi-frame integration and assess their contribution to the whole SLAM framework. By retaining monocular cues alone at the expense of multi-view geometry, we observe severe drops in both tracking and mapping accuracy, confirming the paramount importance of the latter; monocular cues alone are effective only where multiview cues are ineffective – e.g., in the absence of texture, as occurs mostly in the Kitchen scene. + +TUM RGB-D. In Table 4, we simulate sparse ToF data by sampling $0.02\%$ and $0.04\%$ depth points, and compare ToF-Splatting with existing RGB and RGB-D methods. With as few as $0.02\%$ of points, ToF-Splatting largely improves over RGB methods [22, 32], approaching the $100\%$ density RGB-D approach of the most representative competitor [22] at only $0.04\%$ density. + +Replica: Input Depth Sparsity. We conduct additional studies on the Replica dataset, for which Figure 4 provides a qualitative overview of our 3D reconstructions. ToF-Splatting enables performing SLAM with very few sparse points – the 64 sparse points provided by a $8 \times 8$ sparse account only for $0.02\%$ of a $640 \times 480$ image. We here + +![](images/beff949188f76e61ce25e2f0d0d661966e91962c7328414bac13b1b9985dd5c4.jpg) +Figure 4. Replica Qualitatives. We provide qualitative results on Replica [28] to demonstrate the generalization capabilities of our method. On the left from top to bottom, meshes obtained respectively on scenes Office2 and Room2. On the right, details from the scene Room0. ToF-Splating delivers accurate details and allows for nice photometric and depth rendering. + +![](images/5507bdfda1e2142e9434ff9414b17a3166fbf7544d3334aefe737a3d70836de0.jpg) +Figure 5. Impact of depth sparsity. We test on Replica [28] with different simulated depth sparsity levels to assess the capability to exploit higher input densities. MAE and ATE smoothly decrease, whereas rendering metrics appear to be less affected. + +![](images/24c5def5c78c0a1803aae4182b7e0931978cdc3bbe579554c5c536f42e892d08.jpg) +Figure 6. Impact of depth noise. We test on Replica [28] injecting different amounts of noise $\xi \in [0.00, 0.01, 0.05, 0.10]$ . ToF-Splatting demonstrates to be effective at dealing with noise, with error increasing almost linearly with the injected amount of noise. + +explore the impact of higher sparse points density on its overall performance. We simulate on Replica densities of $\{0.02\%, 0.04\%, 0.08\%, 0.16\% \}$ corresponding to about $\{64, 128, 256, 512\}$ points. Figure 5 shows the performance variation on both tracking and mapping metrics using the aforementioned sparsities. + +Replica: Input Depth Noise. We study various noise levels impact in the input depth to assess our pipeline robustness. We model noise as additive but dependent on depth, injecting heteroscedastic Gaussian noise $\mathcal{N}(d,\epsilon d)$ where $\epsilon$ modulates the variance of the noise source on Replica. Figure 6 shows that MAE, PSNR, and ATE all expose linear degradation trends as $\epsilon$ increases. + +Runtime. In Table 5 we measure the runtime of ToF-Splatting in comparison with the main baselines. The tracking time of ToF-Splatting already includes depth estimates with DoD, with average runtime $\bar{t}_{\mathrm{DoD}} \approx 90 ms$ (compris + +
Method NameTracking TimeMapping Time
iMAP [30]101 ms448 ms
NICE-SLAM [46]470 ms1300 ms
MonoGS [22]1169 ms650 ms
ToF-SLAM [19]116 ms380 ms
ToF-Splatting (ours)614 ms865 ms
+ +Table 5. Runtime Comparison. We compare the runtime of ToF-Splatting and other learned baselines measuring their runtime for tracking and mapping. MonoGS [22] tracking time contains also the depth completion inference time [48]. + +ing DepthAnything v2 for monocular cues inference [40]). To date, 3DGS-based mapping methods are not capable of real-time performances. Nevertheless, exploiting the fact that the mapping step accounts only for just a few frames (typically $4\%$ of the total in sequences of the ZJUL5 dataset [19]) and can be parallelized, the pipeline achieves 1.5 fps without further optimization. + +Limitations. Even with state-of-the-art quality, ToF-Splatting manifests some issues that will be addressed in future developments. The most concerning is the runtime, which is too slow for real-time applications and requires memory and compute-intensive backpropagation at deployment. Even though fast 3DGS frameworks exist, such methodologies have not been transferred yet into the SLAM community, that being outside the scope of this work. + +# 5. Conclusion + +We presented the first 3DGS-based SLAM pipeline relying on sparse ToF depth sensing, as it provides accurate tracking and mapping from cheap and low-power cameras. Moreover, we showed how this result can be achieved through the integration of multi-view geometry, sparse depth data, and monocular cues, yielding an end-to-end 3DGS-based SLAM system. Finally, we assessed the effectiveness of our approach on the ZJUL5 and Replica datasets. + +# References + +[1] Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 2 +[2] Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yunsong Wang, and Gim Hee Lee. Vcr-gaus: View consistent depth-normal regularizer for gaussian surface reconstruction. arXiv preprint arXiv:2406.05774, 2024. 2, 5 +[3] Andrea Conti, Matteo Poggi, Valerio Cambareri, and Stefano Mattoccia. Depth on demand: Streaming dense depth from a low frame-rate active sensor. In European Conference on Computer Vision (ECCV), 2024. 2, 4 +[4] Andrea Conti, Matteo Poggi, Valerio Cambareri, and S. Mattoccia. Range-agnostic multi-view depth estimation with keyframe selection. 2024 International Conference on 3D Vision (3DV), pages 1350–1359, 2024. 4 +[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 4, 5 +[6] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. BundleFusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4): 1, 2017. 1, 2, 6, 7 +[7] Tianchen Deng, Guole Shen, Tong Qin, Jianyu Wang, Wentao Zhao, Jingchuan Wang, Danwei Wang, and Weidong Chen. Plgslam: Progressive neural scene representation with local to global bundle adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19657-19666, 2024. 2 +[8] Zilong Dong, Guofeng Zhang, Jiaya Jia, and Hujun Bao. Keyframe-based real-time camera tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1538-1545. IEEE, 2009. 1 +[9] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR, 2024. 2 +[10] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. KinectFusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User Interface Software and Technology, pages 559-568, 2011. 6, 7 +[11] Mohammad Mahdi Johari, Camilla Carta, and François Fleuret. Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17408-17419, 2023. 2 +[12] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians + +for densergb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 2 +[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. 1, 2, 3 +[14] Georg Klein and David Murray. Parallel tracking and mapping for small AR workspaces. In Proceedings of the IEEE and ACM International Symposium on Mixed and Augmented Reality, pages 225-234. IEEE, 2007. 1 +[15] Yanyan Li, Nikolas Brasch, Yida Wang, Nassir Navab, and Federico Tombari. Structure-slam: Low-drift monocular slam in indoor environments. IEEE Robotics and Automation Letters, 5(4):6583-6590, 2020. 2 +[16] Yijin Li, Xinyang Liu, Wenqian Dong, Han Zhou, Hujun Bao, Guofeng Zhang, Yinda Zhang, and Zhaopeng Cui. Deltar: Depth estimation from a light-weight tof sensor and rgb image. In European Conference on Computer Vision, 2022. 3, 4, 6 +[17] Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R Oswald. Loopy-slam: Dense neural slam with loop closures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20363-20373, 2024. 2 +[18] Haomin Liu, Guofeng Zhang, and Hujun Bao. Robust keyframe-based monocular SLAM for augmented reality. In Proceedings of the IEEE International Symposium on Mixed and Augmented Reality, pages 1-10. IEEE, 2016. 1 +[19] Xinyang Liu, Yijin Li, Yanbin Teng, Hujun Bao, Guofeng Zhang, Yinda Zhang, and Zhaopeng Cui. Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1-11, 2023. 2, 4, 5, 6, 7, 8 +[20] Gregor Luetzenburg, Aart Kroon, and Anders A. Bjørk. Evaluation of the Apple iPhone 12 Pro LiDAR for an Application in Geosciences. Scientific Reports, 11(1), 2021. 1 +[21] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. 2024 International Conference on 3D Vision (3DV), pages 800-809, 2023. 2 +[22] Hidenobu Matsuki, Riku Murai, Paul H.J. Kelly, and Andrew J. Davison. Gaussian splattering slam. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18039-18048, 2023. 2, 4, 5, 6, 7, 8 +[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision, pages 405-421. Springer, 2020. 1 +[24] Raul Mur-Artal and Juan D. Tardós. ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo, and RGB-D Cameras. IEEE Transactions on Robotics, 33(5): 1255–1262, 2017. 2 +[25] Xin Qiao, Matteo Poggi, Pengchao Deng, Hao Wei, Chenyang Ge, and Stefano Mattoccia. Rgb guided tof imag- + +ing system: A survey of deep learning-based methods. International Journal of Computer Vision, pages 1-38, 2024. 2 +[26] Erik Sandström, Yue Li, Luc Van Gool, and Martin R Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433-18444, 2023. 2 +[27] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2 +[28] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 8 +[29] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 573-580. IEEE, 2012. 5, 6 +[30] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2, 6, 7, 8 +[31] Keisuke Tateno, Federico Tombari, Iro Laina, and Nassir Navab. Cnn-slam: Real-time dense monocular slam with learned depth prediction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6243-6252, 2017. 2 +[32] Zachary Teed and Jia Deng. Droid-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. Advances in Neural Information Processing Systems, 34: 16558-16569, 2021. 2, 7 +[33] Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R. Oswald, and Matteo Poggi. How nerfs and 3d gaussian splattering are reshaping slam: a survey, 2024. 1, 2 +[34] Hengyi Wang, Jingwen Wang, and Lourdes Agapito. Coslam: Joint coordinate and sparse parametric encodings for neural real-time slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13293-13302, 2023. 2 +[35] Yifan Wang, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38:1-14, 2019. 3 +[36] Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, Andrew J Davison, and Stefan Leutenegger. ElasticFusion: Real-time dense SLAM and light source estimation. The International Journal of Robotics Research, 35(14):1697-1716, 2016. 1, 2, 6, 7 + +[37] Xingming Wu, Zimeng Liu, Yuxin Tian, Zhong Liu, and Weihai Chen. Kn-slam: Keypoints and neural implicit encoding slam. IEEE Transactions on Instrumentation and Measurement, 73:1-12, 2024. 2 +[38] Liu Xinyang, Li Yijin, Teng Yanbin, Bao Hujun, Zhang Guofeng, Zhang Yinda, and Cui Zhaopeng. Multi-modal neural radiance field for monocular dense slam with a lightweight tof sensor. In International Conference on Computer Vision (ICCV), 2023. 2 +[39] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595-19604, 2024. 2 +[40] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. ArXiv, abs/2406.09414, 2024. 4, 8 +[41] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. ArXiv, abs/2312.10070, 2023. 5 +[42] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting, 2023. 2 +[43] Guofeng Zhang, Jiaya Jia, Tien-Tsin Wong, and Hujun Bao. Recovering consistent video depth maps via bundle optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2008. 1 +[44] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimization for consistent 3d instant reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3727-3737, 2023. 2 +[45] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 2 +[46] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. NICE-SLAM: Neural implicit scalable encoding for SLAM. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2, 6, 7, 8 +[47] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R. Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. 2024 International Conference on 3D Vision (3DV), pages 42-52, 2023. 4 +[48] Yiming Zuo and Jia Deng. Ogni-dc: Robust depth completion with optimization-guided neural iterations. arXiv preprint arXiv:2406.11711, 2024. 6, 8 +[49] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. Surface splatting. Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001. 3 \ No newline at end of file diff --git a/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/images.zip b/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e5c033d95f177343eac767b91bc124f651e1429 --- /dev/null +++ b/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95dd7e660a14cb8a7fe96faf43f85a2eb9d8307505903a9fdfbcad9284622023 +size 624889 diff --git a/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/layout.json b/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4cb26547817f4af3fe41d6e6a5a170263efdf350 --- /dev/null +++ b/2025/ToF-Splatting_ Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration/layout.json @@ -0,0 +1,9169 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 102, + 503, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 102, + 503, + 140 + ], + "spans": [ + { + "bbox": [ + 108, + 102, + 503, + 140 + ], + "type": "text", + "content": "ToF-Splatting: Dense SLAM using Sparse Time-of-Flight Depth and Multi-Frame Integration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 136, + 160, + 222, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 160, + 222, + 175 + ], + "spans": [ + { + "bbox": [ + 136, + 160, + 222, + 175 + ], + "type": "text", + "content": "Andrea Conti " + }, + { + "bbox": [ + 136, + 160, + 222, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 246, + 161, + 321, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 161, + 321, + 175 + ], + "spans": [ + { + "bbox": [ + 246, + 161, + 321, + 175 + ], + "type": "text", + "content": "Matteo Poggi" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 350, + 161, + 447, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 161, + 447, + 175 + ], + "spans": [ + { + "bbox": [ + 350, + 161, + 447, + 175 + ], + "type": "text", + "content": "Valerio Cambareri" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 182, + 175, + 278, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 175, + 278, + 188 + ], + "spans": [ + { + "bbox": [ + 182, + 175, + 278, + 188 + ], + "type": "text", + "content": "Martin R. Oswald" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 175, + 402, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 402, + 188 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 402, + 188 + ], + "type": "text", + "content": "Stefano Mattoccia" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 190, + 507, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 190, + 507, + 203 + ], + "spans": [ + { + "bbox": [ + 99, + 190, + 507, + 203 + ], + "type": "text", + "content": "1University of Bologna, Italy 2Sony DepthSensing Solutions, Belgium 3University of Amsterdam, Netherlands" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 82, + 230, + 531, + 358 + ], + "blocks": [ + { + "bbox": [ + 82, + 230, + 531, + 358 + ], + "lines": [ + { + "bbox": [ + 82, + 230, + 531, + 358 + ], + "spans": [ + { + "bbox": [ + 82, + 230, + 531, + 358 + ], + "type": "image", + "image_path": "154d8347800c509697f43c2698b4eac191b07da985beb92b68118ca4783670fb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 365, + 555, + 388 + ], + "lines": [ + { + "bbox": [ + 54, + 365, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 555, + 388 + ], + "type": "text", + "content": "Figure 1. Overview of our ToF-Splatting method. Our method combines sparse ToF depth, multi-view geometry from a buffer of keyframes, and monocular cues (left) to perform into a unique end-to-end dense SLAM framework enabled by a Gaussian Splatting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 398, + 200, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 398, + 200, + 410 + ], + "spans": [ + { + "bbox": [ + 151, + 398, + 200, + 410 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 423, + 297, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 423, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 297, + 628 + ], + "type": "text", + "content": "Time-of-Flight (ToF) sensors provide efficient active depth sensing at relatively low power budgets; among such designs, only very sparse measurements from low-resolution sensors are considered to meet the increasingly limited power constraints of mobile and AR/VR devices. However, such extreme sparsity levels limit the seamless usage of ToF depth in SLAM. In this work, we propose ToF-Splatting, the first 3D Gaussian Splatting-based SLAM pipeline tailored for using effectively very sparse ToF input data. Our approach improves upon the state of the art by introducing a multi-frame integration module, which produces dense depth maps by merging cues from extremely sparse ToF depth, monocular color, and multi-view geometry. Extensive experiments on both real and synthetic sparse ToF datasets demonstrate the advantages of our approach, as it achieves state-of-the-art tracking and mapping performances on reference datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 650, + 135, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 135, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 135, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 670, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 295, + 696 + ], + "type": "text", + "content": "Simultaneous Localization and Mapping (SLAM) [6, 8, 36, 43] entails the joint estimation of the current camera pose" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 399, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 495 + ], + "type": "text", + "content": "against a 3D scene representation (i.e., the map) and the update of this representation with data from the current camera view. It is a fundamental task with applications in AR/VR, robotics, and autonomous navigation [14, 18]. In recent years, SLAM witnessed a revolution [33] following the success of Neural Radiance Fields (NeRF) [23] and 3D Gaussian Splatting (3DGS) [13] as scene representations from the adjacent field of novel view synthesis." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "In this paper we focus on SLAM pipelines that leverage RGB-D video inputs, i.e., where color images and dense depth maps are collected by synchronized, accurate color and depth sensors. Beyond global scale, depth sensors provide accurate local geometry that grants better tracking and mapping quality. A common choice for active depth sensing are Time-of-Flight (ToF) cameras. Despite the increased accuracy provided by reliable depth measurements, the power consumption and cost of such sensors constrain their adoption to high-end industrial or automotive applications. A recent trend to favor ToF sensor integration in handheld consumer devices is to reduce the specifications of ToF sensors to provide low-resolution but more reliable depth measurements. Among other examples, LiDAR (an instance of ToF) appears in consumer mobile phones [20] and is known to feature up to at most 576 dots (interpolated to " + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "inline_equation", + "content": "256 \\times 192" + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": " resolution via depth completion). Other recent ToF sensors provide only up to 64 depth points but consume a re" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 250, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 250, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 250, + 712 + ], + "type": "text", + "content": "* work started while visiting University of Amsterdam." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28344" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "ported " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "200\\mathrm{mW}" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " [19]. In such settings a SLAM pipeline can expect to receive, e.g., " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " very noisy, very sparse depth measurements per frame. These specifications prevent their straightforward integration in SLAM systems, as low-resolution noisy depth is known to harm the accuracy of dense SLAM systems built with NeRF and not originally designed to deal with this scenario [30, 36, 46]. This is even more evident with the more recent 3DGS-based SLAM systems [22], where accurate and dense depth is paramount to achieve satisfactory results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 194, + 296, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 457 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 457 + ], + "type": "text", + "content": "To overcome these limitations, a straightforward strategy [38] would consist in recovering dense and accurate depth maps using a single-image depth completion framework [25]. However, models trained for this task are known to generalize poorly across domains and across input depth noise levels, potentially affecting the downstream SLAM pipeline accuracy and yielding both drifted camera trajectory and inaccurate mapping. Therefore, we explore 3DGS as an explicit representation that allows for injecting domain-specific knowledge and that is easily extended to, e.g., dynamic scenes [21] and accurate large mesh reconstruction [2, 9]. On the one hand, we argue that depth completion alone is insufficient to allow sparse ToF depth usage in SLAM systems. On the other, we acknowledge that even sparse and noisy depth measurements can bootstrap SLAM if properly assisted with geometry. Accordingly, inspired by frame-to-frame SLAM systems relying on pre-trained trackers for localization, we believe that multi-frame integration fusing the noisy depth measurements with multiview geometry across a small set of local frames could supply the dense SLAM system with the supervision necessary to fill this gap." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 460, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 639 + ], + "type": "text", + "content": "In this paper, we propose ToF-Splatting, a novel 3DGS-based SLAM system enabling dense reconstruction from sparse ToF sensors. ToF-Splatting harmoniously alternates tracking and mapping, with each one benefiting from the other. The former is carried out through backpropagation during the optimization of the 3DGS model on color images; the latter exploits a pre-trained network [3] as a multi-frame integration module, combining sparse depth measurements with multi-view geometry according to a buffer of keyframes. For each of the keyframes, we use the camera pose estimated through tracking and providing dense depth maps to improve the supervision given to the 3DGS model itself. Figure 1 shows an overview of the setting in which ToF-Splatting operates, as well as the quality of the 3D reconstructions it yields." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "ToF-Splatting is tested on the real-world ZJUL5 dataset [19], where we measure tracking and mapping quality and find the superiority of our framework with respect to existing solutions built over sparse ToF sensors [38]. Moreover, standard benchmarks such as Replica [28] are also provided and confirm the effectiveness of our mapping strat" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 495, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 495, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 495, + 83 + ], + "type": "text", + "content": "egy. Our main contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 84, + 553, + 168 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 313, + 84, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 108 + ], + "type": "text", + "content": "- We propose ToF-Splatting, the first 3DGS-based SLAM system suited for sparse ToF sensors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 144 + ], + "type": "text", + "content": "- We introduce a novel multi-frame integration method combining sparse depth, multi-view geometry, and monocular cues for robust noise and outlier handling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 168 + ], + "type": "text", + "content": "- We establish a new state-of-the-art for dense SLAM systems built on top of sparse ToF depth sensors." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "spans": [ + { + "bbox": [ + 314, + 178, + 400, + 190 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 199, + 553, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 553, + 222 + ], + "type": "text", + "content": "We briefly review the literature relevant to our work, referring to [33] for a detailed overview of the latest advances." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 223, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 553, + 354 + ], + "type": "text", + "content": "Traditional SLAM. Traditional visual SLAM pipelines, or the more recent ones introducing deep modules [15, 31, 32], usually perform tracking in frame-to-frame fashion, based on the visual features extracted from past frames either using handcrafted feature extractors [24] or deep networks [31, 32]. These systems are usually defined by four main steps: tracking, mapping, global bundle adjustment, and loop closure. Among the different representations used by traditional SLAM frameworks, depth points [1, 24], surfels [27], and volumetric representations [6] allows for achieving globally consistent 3D reconstruction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 354, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 582 + ], + "type": "text", + "content": "NeRF-based SLAM. The advent of NeRF in the field of novel view synthesis also reached adjacent research areas, as well as SLAM a few years later. NeRF-based SLAM systems estimate camera poses and reconstruct dense meshes by modeling scenes via MLPs. iMAP [30] and NICE-SLAM [45] are the first SLAM frameworks following this approach, with the latter introducing dense feature grids supporting the MLPs. Point-SLAM [26] and Loopy-SLAM [17] switched feature grids with neural point embeddings, allowing for higher flexibility and correcting or adjusting local maps after global optimization, yet with slow processing prohibiting deployment in robotics. ESLAM [11] and Co-SLAM [34] deploy tri-planes and hash grid embeddings, improving both processing speed and reconstruction accuracy, with PLGSLAM [7] further improving the representation capacity in large indoor scenes. GO-SLAM [44] and KN-SLAM [37] implement frame-to-frame systems thanks to external trackers, yielding a good trade-off between tracking/mapping accuracy and frame rate." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 677 + ], + "type": "text", + "content": "3DGS-based SLAM. More recent advances in novel view synthesis brought to the development of new SLAM systems using 3D Gaussian Splatting [13] to represent the mapped scene. Several concurrent frameworks emerged built over 3DGS, such as MonoGS [22], GaussianSLAM [42], SplaTAM [12], and GS-SLAM [39]. Different from NeRF-based SLAM systems, these frameworks allow for higher interpretability and higher rendering speed." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Sparse ToF-based SLAM. Only lately, the use of more compact and energy-efficient depth sensors has gained attention in the SLAM research community. ToF-SLAM [38]" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28345" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 72, + 531, + 209 + ], + "blocks": [ + { + "bbox": [ + 75, + 72, + 531, + 209 + ], + "lines": [ + { + "bbox": [ + 75, + 72, + 531, + 209 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 531, + 209 + ], + "type": "image", + "image_path": "d20e17a8a638c546de8098bf09d21117f7d3ac261c3cba4a4d1b227262971f0f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 213, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 54, + 213, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 555, + 247 + ], + "type": "text", + "content": "Figure 2. ToF-Splatting Pipeline. Our method involves three main modules: a Tracking frontend estimating camera poses, a Multi-Frame Integration module that predicts dense depth maps from sparse ToF measurements and multi-view geometry, and a Mapping backend modeling the 3D scene representation via 3D Gaussian Splatting." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 258, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 296, + 344 + ], + "type": "text", + "content": "represents the only attempt to exploit \"lightweight\", sparse depth measurements within a dense SLAM system, aided by a depth completion model [16]. However, depth completion alone fails to deal with noise in the sparse input depth, a shortcoming we aim to specifically overcome within our ToF-Splatting framework, the first one based on 3DGS and suited for sparse ToF sensors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 180, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 180, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 180, + 376 + ], + "type": "text", + "content": "3. Pipeline Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 384, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 295, + 469 + ], + "type": "text", + "content": "In this section, we present the core components of our framework. ToF-Splatting is a frame-to-model SLAM system conceptually divided into three modules: " + }, + { + "bbox": [ + 54, + 384, + 295, + 469 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 54, + 384, + 295, + 469 + ], + "type": "text", + "content": " a mapping backend, " + }, + { + "bbox": [ + 54, + 384, + 295, + 469 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 54, + 384, + 295, + 469 + ], + "type": "text", + "content": " a tracking frontend, and " + }, + { + "bbox": [ + 54, + 384, + 295, + 469 + ], + "type": "inline_equation", + "content": "(iii)" + }, + { + "bbox": [ + 54, + 384, + 295, + 469 + ], + "type": "text", + "content": " a peculiar multi-frame integration module yielding dense depth maps based on a subset of the keyframe graph, the current RGB frame, and the current sparse ToF depth map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 472, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 472, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 54, + 472, + 295, + 604 + ], + "type": "text", + "content": "Figure 2 provides an overview of our pipeline. Our framework is characterized by three heterogeneous modules that are tightly related and influence each other. The tracking module stands as the most influencing element in the pipeline, as it selects the keyframes used by both the mapping and multi-frame integration parts, moreover, it computes the poses used for multi-frame depth estimation. Nonetheless, it relies on the quality of the mapping step to perform correct ego-motion estimation and exploits the depth cues to properly behave when photometric information is lacking or ambiguous." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": "The depth perception part is mainly sustained by the tracking step and highly influences the mapping part since it is used to seed new Gaussians and as supervision. Finally, the mapping phase not only creates a unique smooth representation but also defines the rendered opacity that is used for keyframe selection. These connections lead to a smoothly integrated framework that not only exploits a multi-frame backbone but more importantly, defines an effective way to use such information in the SLAM scenario." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 258, + 508, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 508, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 508, + 272 + ], + "type": "text", + "content": "3.1. Background: 3D Gaussian Splatting" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": "3D Gaussian Splitting [13] (3DGS) represents one of the latest advances in novel view synthesis, fitting a dynamically defined set of multivariate Gaussian distributions over the scene to enable image rendering from arbitrary viewpoints. Each Gaussian " + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "inline_equation", + "content": "G(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": " is parametrized by its mean " + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\mu} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": " and covariance matrix " + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 313, + 275, + 555, + 348 + ], + "type": "text", + "content": ", reading" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 342, + 352, + 555, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 352, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 342, + 352, + 555, + 379 + ], + "type": "interline_equation", + "content": "G (\\mathbf {x}) = \\exp \\left(- \\frac {1}{2} (\\mathbf {x} - \\boldsymbol {\\mu}) ^ {\\top} \\boldsymbol {\\Sigma} ^ {- 1} (\\mathbf {x} - \\boldsymbol {\\mu})\\right) \\tag {1}", + "image_path": "136186e0a4ff5378d36b93a792a0ea17b5aeca2afe952c51f9c6daa00d129724.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 383, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 554, + 407 + ], + "type": "text", + "content": "Moreover, opacity " + }, + { + "bbox": [ + 313, + 383, + 554, + 407 + ], + "type": "inline_equation", + "content": "o \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 383, + 554, + 407 + ], + "type": "text", + "content": " and RGB color " + }, + { + "bbox": [ + 313, + 383, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 383, + 554, + 407 + ], + "type": "text", + "content": " are assigned to each Gaussian." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 407, + 555, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 467 + ], + "type": "text", + "content": "Given a set of images from different viewpoints all around the scene, the Gaussian parameters are optimized using the splatting technique [35]. With such a method, 3D Gaussians are projected through rasterization over the 2D image plane, containing a 2D Gaussian distribution" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 338, + 471, + 553, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 471, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 338, + 471, + 553, + 487 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} ^ {2 D} = \\pi \\left(\\mathbf {T} _ {w c} \\bar {\\boldsymbol {\\mu}}\\right) \\quad \\boldsymbol {\\Sigma} ^ {2 D} = \\mathbf {J} \\mathbf {R} _ {w c} \\boldsymbol {\\Sigma} \\mathbf {R} _ {w c} ^ {\\top} \\mathbf {J} ^ {\\top} \\tag {2}", + "image_path": "85ff421beec51a6b35e81c2786c28102543adf1876afec9f262a16d49e93b7c2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{\\mu}}" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": " in homogeneous coordinates, " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{wc} \\in \\mathbb{R}^{4 \\times 4}" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{wc} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": " are respectively the world-to-camera transformation and its rotational component, " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": " is the projection matrix and " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{J} \\in \\mathbb{R}^{2 \\times 3}" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": " is the Jacobian of the projective transformation [49]. The color " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": " of a pixel is then given by the combination of the " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": " weighted Gaussians, i.e.," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 377, + 567, + 553, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 567, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 377, + 567, + 553, + 602 + ], + "type": "interline_equation", + "content": "\\mathbf {C} = \\sum_ {i \\in M} \\mathbf {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}) \\tag {3}", + "image_path": "433d042169534fb2aff51e5232bd335c9585ca78b9031230bb18cdb3f4b4c228.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = o_{i}G(\\mathbf{x}_{i})" + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "text", + "content": " is positive semi-definite, it is parametrized with a diagonal scaling matrix " + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "text", + "content": " and a rotation matrix " + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "text", + "content": ", so that " + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma} = \\mathbf{RSS}^{\\top}\\mathbf{R}^{\\top}" + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 606, + 555, + 654 + ], + "type": "text", + "content": " is internally represented with quaternions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 660, + 454, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 454, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 454, + 673 + ], + "type": "text", + "content": "3.2. Multi-Frame Integration" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Dense depth cues are pivotal for proper initialization of the Gaussians position " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": ", as well as to provide supervision to both tracking and mapping threads - as already" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28346" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 192 + ], + "type": "text", + "content": "known in the literature [22, 47]. However, sparse ToF sensors alone are insufficient for this purpose, because of the very sparse and noisy depth measurements they provide (e.g., only 64 points for a sensor such as that used in [19]). To overcome this limitation, ToF-SLAM [19] exploits an auxiliary completion model, DELTAR [16], to recover dense depth maps out of ToF measurements. However, depth completion is highly dependent on the quality of the sparse depth in input, rapidly degrading the accuracy of the densified depth maps in the presence of noise." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 294, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 294, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 294, + 290 + ], + "type": "text", + "content": "In this paper, we follow a different path: inspired by frame-to-frame SLAM systems deploying an external tracker, we believe that multi-frame integration can compensate for the noise in the ToF measurements by exploiting both single-view and multi-view geometry cues on a set of local frames, given the camera poses estimated by the tracking frontend. We choose to extend the Depth on Demand (DoD) framework [3] for this purpose." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": "Multi-Frame Integration. DoD [3] integrates sparse depth and two-view stereo cues for dense depth prediction. Specifically – given a frame " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^k" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " – it processes the associated RGB image " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "I_k" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " (the target view), another RGB view " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " from a different frame " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^j" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " (a source view), the relative pose " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "P^{jk}" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " between the two, and the sparse depth points " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "H^k" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " to predict a dense metric depth map " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "D^k" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": ". Differently from monocular depth prediction, DoD is aware of the scene scale either from " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "H^k" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "P^{jk}" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": ", simplifying the problem of integrating scale-inconsistent geometries in the SLAM frontend. However, DoD is limited to two-view processing: this may lead to suboptimal results in the SLAM scenario, where an abundant amount of keyframes largely overlap and cover the same portion of the scene. Thus, we extend [3] to multiview processing. This is achieved by observing that the predicted depth map " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "D^k" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " depends on the source view " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " through its relative pose " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "P^{jk}" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " only. DoD iteratively updates " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "D^k" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " for multiple iterations, exploiting this intrinsic property we can integrate multiple views " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "(I_j, P^{jk})" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " using a different one at each iteration [4]. This approach allows smooth integration of multi-view cues avoiding order dependency. To select the source views " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": ", at each prediction we order the keyframes selected by the SLAM pipeline according to their relative pose similarity and select the first " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " frames with baseline distance close to " + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "inline_equation", + "content": "b = 15\\mathrm{cm}" + }, + { + "bbox": [ + 56, + 293, + 296, + 616 + ], + "type": "text", + "content": " to ensure enough parallax. The model is trained from scratch on ScanNetv2 [5] and following the protocol of [3]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Monocular Cues Integration. As shown in [3], DoD heavily relies on both geometry and depth measurements, thus struggling at generalizing in challenging scenarios where the two lack. This may happen in cases where the tight field of view, noisy sparse points, and textureless surfaces hamper multi-view matching. To mitigate such limitations, we integrate explicit monocular cues into DoD by feeding its monocular encoder with " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "I_{k}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " and the normalized depth" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "content": "map " + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "inline_equation", + "content": "\\tilde{D}^k\\in [0,1]^{H\\times W}" + }, + { + "bbox": [ + 313, + 71, + 553, + 144 + ], + "type": "text", + "content": " obtained through the single-image depth estimator Depth Anything v2 [40]. This approach allows the injection of a robust bias, avoiding issues related to monocular estimation, e.g., slanted surfaces or noisy scale estimates, and delegating to DoD their handling. These cues are included while retraining on ScanNetv2 [5]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "type": "text", + "content": "Outlier Handling. Furthermore, the noisy depth measurements in input to DoD can severely affect its accuracy. Therefore, we explicitly focus on dealing with outliers in the input sparse depth, that may be perceived by ToF devices over specific surfaces. Such errors are particularly evident in cheap sensors and may largely prejudice the information supplied by the few points they measure. Purposely, we deploy a simple yet effective method to identify such outliers by exploiting multi-view geometry again. Whenever depth needs to be predicted for a new frame, DoD predicts a depth map " + }, + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\hat{D}^k" + }, + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "type": "text", + "content": " without processing the ToF measurements. Then, the " + }, + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "type": "text", + "content": " error between the latter and the depth measured by the ToF sensor " + }, + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "type": "inline_equation", + "content": "H^k" + }, + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "type": "text", + "content": " is computed and finally sparse measures with error higher than a given quantile " + }, + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "type": "text", + "content": " are discarded, and depth is predicted again by DoD by also processing the filtered sparse depth " + }, + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "type": "inline_equation", + "content": "\\tilde{H}^k" + }, + { + "bbox": [ + 313, + 144, + 555, + 407 + ], + "type": "text", + "content": " this time. This way, we exploit multi-view cues to extract a clean depth prior to filtering out inconsistent sparse depth measures, and then we integrate the remaining points to improve the original prediction. Even without processing the sparse depth, DoD still predicts metric depth, as the metric scale is enforced by the relative camera poses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 413, + 425, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 413, + 425, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 413, + 425, + 426 + ], + "type": "text", + "content": "3.3. Tracking Frontend" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 430, + 553, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 553, + 478 + ], + "type": "text", + "content": "The tracking part of ToF-Splatting estimates the ego-motion of the camera for each new frame " + }, + { + "bbox": [ + 313, + 430, + 553, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^k" + }, + { + "bbox": [ + 313, + 430, + 553, + 478 + ], + "type": "text", + "content": " and builds a keyframe graph " + }, + { + "bbox": [ + 313, + 430, + 553, + 478 + ], + "type": "inline_equation", + "content": "\\mathbf{G} = \\{\\mathbf{KF}^k\\} \\subset \\{\\mathbf{F}^k\\}" + }, + { + "bbox": [ + 313, + 430, + 553, + 478 + ], + "type": "text", + "content": " containing a set of meaningful frames for mapping and multi-frame depth estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "text", + "content": "Pose Estimation. We initialize each new frame " + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^k" + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "text", + "content": " of our frame-to-model tracker with the pose of the previous frame " + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{k - 1}" + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "text", + "content": " and then minimizing the tracking loss with respect to the relative pose between " + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^k" + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{k - 1}" + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "text", + "content": ", following [22]. The tracking loss consists of two terms, " + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{depth}}" + }, + { + "bbox": [ + 313, + 479, + 553, + 551 + ], + "type": "text", + "content": ", respectively accounting for photometric and geometric errors" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 321, + 568, + 553, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 568, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 321, + 568, + 553, + 582 + ], + "type": "interline_equation", + "content": "L _ {\\text {t r a c k}} = \\lambda_ {\\text {t r a c k}} L _ {\\text {r g b}} + (1 - \\lambda_ {\\text {t r a c k}}) L _ {\\text {d e p t h}} \\tag {4}", + "image_path": "32aab2b0b29f1d3d364d181a11184cce0f79bb8f3530113f9371eac7cb9f1d6c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 583, + 553, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 583, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 326, + 583, + 553, + 598 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {r g b}} = \\left\\| I _ {k} - \\hat {I} \\left(\\mathbf {F} ^ {k}\\right) \\right\\| _ {1} \\quad L _ {\\mathrm {d e p t h}} = \\left\\| D ^ {k} - \\hat {D} \\left(\\mathbf {F} ^ {k}\\right) \\right\\| _ {1} \\tag {5}", + "image_path": "3d585ce463fd97cae33c7b037d11a3fb49f8ed625d0d4f4cec1640e0a4d17361.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "inline_equation", + "content": "I_{k}" + }, + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "text", + "content": " is the RGB image from " + }, + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^k" + }, + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "inline_equation", + "content": "\\hat{I}(\\cdot)" + }, + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "inline_equation", + "content": "\\hat{D}(\\cdot)" + }, + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "text", + "content": " denote the rendered RGB and Depth and, " + }, + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "inline_equation", + "content": "D^{k}" + }, + { + "bbox": [ + 313, + 605, + 553, + 641 + ], + "type": "text", + "content": " is the depth prediction produced by the multi-frame integration module." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 715 + ], + "type": "text", + "content": "Unlike other methods, the tracking process happens in two stages. Initially, for a fixed number of steps " + }, + { + "bbox": [ + 313, + 642, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 313, + 642, + 553, + 715 + ], + "type": "text", + "content": " only the photometric loss " + }, + { + "bbox": [ + 313, + 642, + 553, + 715 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{rgb}}^T" + }, + { + "bbox": [ + 313, + 642, + 553, + 715 + ], + "type": "text", + "content": " is optimized. Then, such a pose is used for multi-frame integration, as described in Sec. 3.2, to generate " + }, + { + "bbox": [ + 313, + 642, + 553, + 715 + ], + "type": "inline_equation", + "content": "D^{k}" + }, + { + "bbox": [ + 313, + 642, + 553, + 715 + ], + "type": "text", + "content": ". Finally, both losses in Eq. (5) are used for " + }, + { + "bbox": [ + 313, + 642, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{rgbd}}^T" + }, + { + "bbox": [ + 313, + 642, + 553, + 715 + ], + "type": "text", + "content": " further steps." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28347" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "Keyframe Selection Policy. We deploy a simple yet effective keyframe selection policy that considers i) novel content to be mapped and ii) instabilities in the already mapped areas. Unlike other methods using complex approaches – e.g., frustum intersection, point clouds and pose analysis [22, 41] – we observe that the rendered opacity for a novel frame contains low opacity values where either Gaussians are not present or unstable. The first case highlights the presence of novel areas of the environment to be mapped, the second happens due to the pruning procedure involved in the mapping step (see Sec. 3.4). We set a frame " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^k" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " as a keyframe " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{K}\\mathbf{F}^k" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " if its novelty factor " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\nu^k" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " is higher than a threshold " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\nu^{th}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": ", where we define " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\nu^k" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 119, + 237, + 295, + 267 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 237, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 119, + 237, + 295, + 267 + ], + "type": "interline_equation", + "content": "\\nu^ {k} = \\frac {\\sum_ {i \\in \\mathbf {K F} ^ {k}} \\mathbf {M} \\left(\\mathbf {K F} _ {i} ^ {k}\\right)}{\\sum_ {i \\in \\mathbf {K F} ^ {k}} \\left(1 - \\mathbf {M} \\left(\\mathbf {K F} _ {i} ^ {k}\\right)\\right)} \\tag {6}", + "image_path": "3639ff58268ca68ee774f6815b04244081b2d4c3966529e43da3f71ebd5fc8af.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 91, + 269, + 295, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 269, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 91, + 269, + 295, + 301 + ], + "type": "interline_equation", + "content": "\\mathbf {M} \\left(\\mathbf {K F} _ {i} ^ {k}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\hat {O} \\left(\\mathbf {K F} _ {i} ^ {k}\\right) < \\sigma \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {7}", + "image_path": "466f2c22ff34d5856ad59bf847d4380c4afebb11317f32bb45fc63c7e6bfd254.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 311, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 295, + 337 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 55, + 311, + 295, + 337 + ], + "type": "inline_equation", + "content": "\\hat{O} (\\mathbf{K}\\mathbf{F}_i^k)" + }, + { + "bbox": [ + 55, + 311, + 295, + 337 + ], + "type": "text", + "content": " being the rendered opacity at pixel " + }, + { + "bbox": [ + 55, + 311, + 295, + 337 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 311, + 295, + 337 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 311, + 295, + 337 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(\\mathbf{K}\\mathbf{F}_i^k)" + }, + { + "bbox": [ + 55, + 311, + 295, + 337 + ], + "type": "text", + "content": " a binary uncertainty map." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 337, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 397 + ], + "type": "text", + "content": "This approach allows for skipping several frames when the camera moves slowly or focusing deeply on areas where the ToF-Splatting struggles the most, by mapping it consecutively. Nonetheless, we enforce mapping after skipping a certain number of frames to avoid mapping too few frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 405, + 164, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 164, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 164, + 418 + ], + "type": "text", + "content": "3.4. Mapping Backend" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 422, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 459 + ], + "type": "text", + "content": "Whenever a new keyframe " + }, + { + "bbox": [ + 55, + 422, + 295, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{K}\\mathbf{F}^k" + }, + { + "bbox": [ + 55, + 422, + 295, + 459 + ], + "type": "text", + "content": " is identified, mapping is triggered. This phase seeks to embed the new frame in the global 3DGS model, involving the following two steps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 459, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 602 + ], + "type": "text", + "content": "Initialization. In this step, new Gaussians are seeded in the 3DGS model. To limit the size of the model and reduce outliers, only areas where the rendered uncertainty " + }, + { + "bbox": [ + 55, + 459, + 295, + 602 + ], + "type": "inline_equation", + "content": "\\mathbf{M}(\\mathbf{KF}^k)" + }, + { + "bbox": [ + 55, + 459, + 295, + 602 + ], + "type": "text", + "content": " is high are seeded with new points, with the new Gaussians also being randomly downsampled by a constant factor. Each Gaussian is initialized using depth " + }, + { + "bbox": [ + 55, + 459, + 295, + 602 + ], + "type": "inline_equation", + "content": "D^{k}" + }, + { + "bbox": [ + 55, + 459, + 295, + 602 + ], + "type": "text", + "content": " predicted in the multi-frame integration step, with color from the RGB frame " + }, + { + "bbox": [ + 55, + 459, + 295, + 602 + ], + "type": "inline_equation", + "content": "I_{k}" + }, + { + "bbox": [ + 55, + 459, + 295, + 602 + ], + "type": "text", + "content": ", and scale initialized with a constant value. Unlike other 3DGS pipelines [22, 41], we prove this approach effective and particularly efficient, replacing any complex 3D heuristics to a simple use of rendered opacity " + }, + { + "bbox": [ + 55, + 459, + 295, + 602 + ], + "type": "inline_equation", + "content": "O(\\cdot)" + }, + { + "bbox": [ + 55, + 459, + 295, + 602 + ], + "type": "text", + "content": ", already a side-product of the rasterization process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "text", + "content": "Optimization. Firstly, ToF-Splating smoothly integrates the new information in the global model and secondly tunes the existing representation. This is achieved by optimizing the new keyframe " + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{KF}^k" + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "text", + "content": " and a random subset of 5 keyframes among the last " + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "text", + "content": " keyframes " + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{KF}^j \\mid j \\geq \\max(k - N, 0)\\}" + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "text", + "content": ". These are optimized for a fixed number of steps " + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{rgbd}}^M" + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "text", + "content": ". Moreover, we enforce a sampling rate of " + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{KF}^k" + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "text", + "content": ", with the remaining " + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "text", + "content": " for the remaining " + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{KF}^j" + }, + { + "bbox": [ + 55, + 603, + 295, + 714 + ], + "type": "text", + "content": " to avoid forgetting. During this step, gradients are also back-propagated to the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "camera poses of the sampled keyframes, thus acting as a global bundle adjustment. Finally, we prune Gaussians with opacities lower than 0.5 every " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{rgbd}}^{M} / 2" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " steps. In addition to the terms in Eq. (5), we optimize structural similarity as in Eq. (8) and, following [2], the normals as in Eq. (9), i.e.," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 359, + 139, + 554, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 139, + 554, + 166 + ], + "spans": [ + { + "bbox": [ + 359, + 139, + 554, + 166 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {d s s i m}} = \\frac {1 - \\operatorname {S S I M} \\left(I _ {k} , \\hat {I} \\left(\\mathbf {K F} ^ {k}\\right)\\right)}{2} \\tag {8}", + "image_path": "6b1de7f5099fb81efdc3d5d0ffaf4fc1694f3052b12939ee46f6f961bf24c64d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 168, + 553, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 168, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 354, + 168, + 553, + 184 + ], + "type": "interline_equation", + "content": "L _ {\\text {n o r m a l s}} = \\| N ^ {k} - \\hat {N} \\| _ {1} + \\overline {{\\mathbf {1} - \\langle N ^ {k} , \\hat {N} \\rangle}} \\tag {9}", + "image_path": "9791e9ed0fccf5c4b96abcef6862d78388e1ceb29d5eb640de20e175683bdb29.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "inline_equation", + "content": "\\hat{N} \\coloneqq \\bar{\\nabla}\\hat{D}(\\mathbf{K}\\mathbf{F}^k)" + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "inline_equation", + "content": "N^k \\coloneqq \\bar{\\nabla}D^k" + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "text", + "content": " denoting the normals estimated via 3D gradients " + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "inline_equation", + "content": "\\bar{\\nabla}" + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "text", + "content": " for the rendered and multiframe predicted depth respectively, " + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "inline_equation", + "content": "\\langle \\cdot, \\cdot \\rangle" + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "text", + "content": " their inner product, and " + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "inline_equation", + "content": "\\overline{\\cdot}" + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "text", + "content": " the average. " + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{normals}}" + }, + { + "bbox": [ + 313, + 191, + 554, + 253 + ], + "type": "text", + "content": " is masked on edges to preserve sharpness, as identified by running a Sobel filter." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "text", + "content": "Finally, we introduce a regularization loss " + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{iso}}" + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "text", + "content": " to penalize elongated Gaussians (i.e., promote their isotropy) that usually leads to rendering artifacts as already observed in [22]. Eq. (10) defines " + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{iso}}" + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "inline_equation", + "content": "\\frac{\\mathrm{diag}(\\cdot)}{\\mathrm{diag}(\\mathbf{S}_j)}" + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "text", + "content": " extracts the diagonal values from its argument, and " + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "inline_equation", + "content": "\\overline{\\mathrm{diag}(\\mathbf{S}_j)}" + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "text", + "content": " is the average of the resulting vector, i.e., of the elements of the " + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "text", + "content": "-th diagonal scale matrix " + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_j" + }, + { + "bbox": [ + 313, + 253, + 554, + 338 + ], + "type": "text", + "content": " as defined in Sec. 3.1. Thus," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 331, + 345, + 553, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 345, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 331, + 345, + 553, + 380 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {i s o}} = \\frac {1}{| \\mathcal {G} |} \\sum_ {j} ^ {| \\mathcal {G} |} \\| \\operatorname {d i a g} \\left(\\mathbf {S} _ {j}\\right) - \\overline {{\\operatorname {d i a g} \\left(\\mathbf {S} _ {j}\\right)}} \\cdot \\mathbf {1} _ {3 \\times 1} \\| _ {1} \\tag {10}", + "image_path": "fa67312933bfb1c464e8d90a3fa6d471e4eefa4c6ca624b9203cefa3c838f5e1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 387, + 554, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 554, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 554, + 412 + ], + "type": "text", + "content": "The final mapping loss aggregates the aforementioned loss terms with corresponding weights as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 418, + 553, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 418, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 321, + 418, + 553, + 449 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\text {m a p}} = \\lambda_ {\\text {m a p}} \\left(\\lambda_ {\\text {v i s u a l}} L _ {\\text {r g b}} + (1 - \\lambda_ {\\text {v i s u a l}}) L _ {\\text {d s s i m}}\\right) \\tag {11} \\\\ + \\left(1 - \\lambda_ {\\text {m a p}}\\right) L _ {\\text {d e p t h}} + \\lambda_ {\\text {n o r m a l s}} L _ {\\text {n o r m a l s}} + \\lambda_ {\\text {i s o}} L _ {\\text {i s o}} \\\\ \\end{array}", + "image_path": "db8ba1e63fe408ec49b06268ccb0e37cf531f0213c8e5635d2f21c9f405a1a83.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 466, + 395, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 395, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 395, + 480 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 486, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 510 + ], + "type": "text", + "content": "This section assesses both the tracking and mapping performance of ToF-Splatting on the following datasets." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "type": "text", + "content": "ZJUL5. ZJUL5 [19] is the only existing public-domain real dataset providing sparse depth data from a low-resolution, low-cost ToF sensor. A VL53L5CX ToF sensor is assembled on a calibrated rig with an Intel RealSense 435i to provide dense depth ground truth. This benchmark comprises 7 diverse indoor scene recordings. Ground truth 3D meshes and camera poses are obtained following the ScanNet protocol [5]. This sensor provides " + }, + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 510, + 554, + 665 + ], + "type": "text", + "content": " depth points by counting the number of photons returned in each discretized time range. Of the at most 64 depth points provided, many are extremely noisy and introduce large outliers. This challenging real-world dataset provides real measurement noise that is not accurately represented in mainstream datasets." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "TUM RGB-D. This is a real dataset [29] providing indoor sequence captures and widely used as benchmark for RGB-D SLAM pipelines. The depth stream provides dense VGA depth maps from a high-resolution Kinect v1 depth sensor." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28348" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 168 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 168 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 168 + ], + "type": "table", + "html": "
MethodKitchenSofaOfficeReceptionLiving RoomOffice2Sofa2Avg.
KinectFusion [10]X0.1460.2090.157X0.3210.1250.192
iMAP [30]X1.6580.3380.6480.6790.3440.2140.647
NICE-SLAM [46]0.7450.1440.1550.2510.2890.2280.4210.319
BundleFusion [6]0.1760.1020.1350.101X0.1630.1200.132
ElasticFusion [36]0.2530.1100.0700.1930.5300.1210.1460.203
MonoGS [22]0.2310.0320.0410.0440.1530.0350.0730.087
ToF-SLAM [19]0.1130.0810.0560.1140.2000.1010.0850.107
ToF-Splatting (ours)0.0880.0240.0220.0410.1220.0290.0300.051
", + "image_path": "58ab1526af8198a569a54680c797c14f720c058b2f470e00f67f9d64b1e9298a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "type": "text", + "content": "Table 1. ZJUL5 Tracking Results. We show here the tracking performance on the ZJUL5 dataset [19] dataset with the ATE RMSE [m] (↓) metric on the 8 sequences available where " + }, + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 171, + 555, + 205 + ], + "type": "text", + "content": " indicates lost tracking. ToF-Splatting provides the better performance by a good margin on each sequence, demonstrating the optimal capability of our proposal to track the camera ego-motion accurately." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 217, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 217, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 294, + 277 + ], + "type": "text", + "content": "Replica. Replica [28] is a synthetic dataset providing extremely realistic indoor sequences. It provides dense depth maps, which we use at different densities to perform experiments that prove the robustness and generalization capabilities of ToF-Splatting in challenging scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 277, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 277, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 295, + 361 + ], + "type": "text", + "content": "Baselines. We compare our method with both learning-based [19, 30, 46] and traditional pipelines [6, 10, 36]. Following ToF-SLAM [19], baselines requiring RGB-D frames use DELTAR [16] to densify the sparse ToF measurements. Moreover, we adapt MonoGS [22] densifying depth with the state-of-the-art depth completion framework OGNI-DC [48] for a Gaussian Splatting SLAM baseline." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": "Implementation Details. In all our experiments, we use " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "N = 4" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": " source views " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": " (i.e., keyframe buffer elements) for the multi-view integration module and a baseline " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "b = 15\\mathrm{cm}" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": " for selecting keyframes. We use a quantile " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "q = 0.75" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": " to filter outliers that occur frequently in the sparse ToF depth of the ZJUL5 dataset [19]. Concerning tracking, we use a novelty factor threshold of " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\nu_{th} = 0.1" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\sigma = 0.98" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": ". The total tracking loss uses " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{track}} = 0.9" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": ", a total number of " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{rgb}}^T = 30" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{rgbd}}^T = 70" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": " steps are performed at each iteration. For mapping, we set the loss weights " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{map}} = 60" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{visual}} = 0.20" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{normals}} = 0.01" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{iso}} = 1.0" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": " and the number of steps per iteration " + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "inline_equation", + "content": "\\eta_{\\mathrm{rgbd}}^M = 60" + }, + { + "bbox": [ + 54, + 361, + 295, + 509 + ], + "type": "text", + "content": ". We test on a single RTX3090." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 515, + 253, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 253, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 253, + 528 + ], + "type": "text", + "content": "4.1. Comparison with the State of the Art" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 533, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 557 + ], + "type": "text", + "content": "In this section, we compare ToF-Splatting with existing methods in terms of tracking and mapping performances." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 557, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 557, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 54, + 557, + 295, + 676 + ], + "type": "text", + "content": "Tracking. In Table 1, we report the tracking performance by ToF-Splatting and various baselines on each sequence of the ZJUL5 dataset [19]. Following common practice we use the Absolute Trajectory Error metric [29], which directly compares the global consistency of a trajectory. While various frameworks [6, 10, 30] fail on the most challenging scenes, ToF-Splatting delivers accurate and consistent tracking. This is evident from the good margin it gains against the baselines. In Figure 3, we show the 3D trajectories on a subset of the scenes, with error below " + }, + { + "bbox": [ + 54, + 557, + 295, + 676 + ], + "type": "inline_equation", + "content": "3\\mathrm{cm}" + }, + { + "bbox": [ + 54, + 557, + 295, + 676 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 677, + 296, + 713 + ], + "type": "text", + "content": "Mapping. In Table 2, we report the mapping performance by ToF-Splatting compared with the other baselines. For each scene, we collect the predicted pose of each frame" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 219, + 550, + 473 + ], + "blocks": [ + { + "bbox": [ + 319, + 219, + 550, + 473 + ], + "lines": [ + { + "bbox": [ + 319, + 219, + 550, + 473 + ], + "spans": [ + { + "bbox": [ + 319, + 219, + 550, + 473 + ], + "type": "image", + "image_path": "80030511516d6a512ed1b075e61b8fe373f7acf9c8973d7621a95480ea3c4880.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 478, + 555, + 522 + ], + "lines": [ + { + "bbox": [ + 313, + 478, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 555, + 522 + ], + "type": "text", + "content": "Figure 3. Qualitative results on the ZJUL5 dataset [19]. We show meshes obtained by fusing rendered depth maps with TSDF and marching cubes (left), and 3D trajectories (right) on 3 scenes selected from the ZJUL5 dataset [19]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": "and render depth and color from 3DGS. Then, we perform TSDF integration and extract the final mesh through marching cubes. Following [19], the meshes are then evaluated by computing Accuracy, Completion, and F-score versus the ground truth meshes. Accuracy and Completion respectively evaluate the mean distance between each predicted vertex and the nearest ground truth one and vice versa. The F-score takes into account both Accuracy and Completion with an aggregate metric. ToF-Splating delivers consistent results for mapping, always achieving the highest accuracy and significantly exceeding the baselines on average and specifically on the Sofa, Office, and Reception scenes which account for five over seven sequences. [19] is the second-best method and the main competitor in mapping." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28349" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 555, + 290 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 290 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 290 + ], + "type": "table", + "html": "
MethodMetricKitchenSofaOfficeReceptionLiving RoomOffice2Sofa2Avg.
KinectFusion [10]Accuracy ↓X0.1900.2110.261X0.2670.135-
Completion ↓X0.0480.0460.064X0.0780.064-
F-score ↑X0.2780.2880.285X0.2740.381-
ElasticFusion [36]Accuracy ↓0.0920.1350.0840.2970.1510.0960.1220.140
Completion ↓0.0650.0480.0820.3050.2160.1470.0470.130
F-score ↑0.5530.4200.5290.2740.3820.4160.4810.436
BundleFusion [6]Accuracy ↓0.1700.1000.1030.122X0.1210.123-
Completion ↓0.0880.0300.0380.057X0.2140.034-
F-score ↑0.3730.5710.4740.470X0.4420.527-
iMAP [30]Accuracy ↓X0.1350.2290.3650.2250.2330.139-
Completion ↓X0.0540.1030.2450.2910.1390.069-
F-score ↑X0.4450.3150.2380.1700.2550.416-
NICE-SLAM [46]Accuracy ↓0.3030.1190.1160.2160.1030.1560.4640.211
Completion ↓0.4560.0420.0700.1990.0890.1630.0450.152
F-score ↑0.2210.5540.4110.4020.4000.2730.4010.380
MonoGS [22]Accuracy ↓0.0900.0860.0830.0710.0690.0810.0890.081
Completion ↓0.2460.1220.1040.2380.1930.1570.1700.175
F-score ↑0.2900.2990.3460.3670.2590.3740.2890.318
ToF-SLAM [19]Accuracy ↓0.0810.0680.0670.0790.0780.1130.1210.087
Completion ↓0.0710.0410.0450.0620.1220.0850.0330.066
F-score ↑0.5590.6610.6460.6430.4960.5570.6560.604
ToF-Splatting (ours)Accuracy ↓0.0640.0310.0320.0410.0590.0340.0430.043
Completion ↓0.0950.0380.0290.0560.1310.0540.0460.064
F-score ↑0.5270.7910.8400.7100.3590.7790.6420.664
", + "image_path": "cdfd93b0e71c93afe0b7289433b526fc6b3c3cbdb92c15b9b16910c58e33ba07.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 322, + 555, + 397 + ], + "blocks": [ + { + "bbox": [ + 55, + 294, + 555, + 317 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 555, + 317 + ], + "type": "text", + "content": "Table 2. ZJUL5 Mapping Results. We perform the mapping evaluation on the ZJUL5 dataset (7 indoor sequences) and report results of three metrics including accuracy (Acc.), completion (Comp.), and F-score. The failure cases are marked as " + }, + { + "bbox": [ + 55, + 294, + 555, + 317 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 55, + 294, + 555, + 317 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 322, + 555, + 397 + ], + "lines": [ + { + "bbox": [ + 56, + 322, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 555, + 397 + ], + "type": "table", + "html": "
Monocular CuesMulti-View CuesMetricKitchenSofaOfficeReceptionLiving RoomOffice2Sofa2Avg.
✓XATE ↓0.1030.0540.0460.0500.1100.0410.0550.066
F-score ↑0.2160.6750.5640.6450.2410.6110.5570.501
X✓ATE ↓0.2340.0240.0250.0370.1090.0340.0280.072
F-score ↑0.2840.7810.8260.7030.2530.7660.6830.614
✓✓ATE ↓0.0880.0240.0220.0410.1220.0290.0300.051
F-score ↑0.5270.7910.8400.7100.3590.7790.6420.664
", + "image_path": "6a1aa086b182d275b084baf0c8482cd8fb43afe64b5c8d2d9c7747423b83125b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 446, + 294, + 557 + ], + "blocks": [ + { + "bbox": [ + 55, + 400, + 555, + 434 + ], + "lines": [ + { + "bbox": [ + 55, + 400, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 555, + 434 + ], + "type": "text", + "content": "Table 3. Multi-Frame Integration Ablation Study. Analysis of the impact of monocular and multi-view cues in the multi-frame integration step. Monocular cues are less impactful than multi-view ones, boosting performance in the challenging scenes where multi-view cues struggle most. Nonetheless, their combination yields the best results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 446, + 294, + 557 + ], + "lines": [ + { + "bbox": [ + 57, + 446, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 446, + 294, + 557 + ], + "type": "table", + "html": "
MethodLCInputDensityfr1/deskfr2/xyzfr3/officeAvg.
DROID-VOXRGB0.00%0.0520.1070.0730.077
MonoGSXRGB0.00%0.0380.0460.0350.040
ToF-SplattingXRGB-D0.02%0.0300.0220.0300.027
0.04%0.0270.0190.0210.022
ESLAMXRGB-D100%0.0250.0110.0240.020
Co-SLAMXRGB-D100%0.0240.0170.0240.022
CG-SLAMXRGB-D100%0.0240.0120.0240.020
RTG-SLAMXRGB-D100%0.0170.0040.0110.011
MonoGSXRGB-D100%0.0150.0140.0150.015
ORB-SLAM2✓RGB0.00%0.0190.0060.0240.016
ORB-SLAM2✓RGB-D100%0.0160.0400.0100.010
", + "image_path": "5971744a6c15c7ee6b05e581c3e9f311589b5fbfdadcde0925b65731a6a81d1c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 592 + ], + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 592 + ], + "type": "text", + "content": "Table 4. Results on TUM RGB-D dataset. Top: competitors w/o loop-closure (LC); middle: most representative competitor with " + }, + { + "bbox": [ + 55, + 558, + 295, + 592 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 592 + ], + "type": "text", + "content": " density; bottom: competitors w/ loop-closure." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 598, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 295, + 658 + ], + "type": "text", + "content": "Notably, the other methods – like [6, 36] – may provide good performance on specific scenes but demonstrate unreliability on average usually failing in the particularly challenging Kitchen and Living Room scenes. In Figure 3, we show Office, Reception, and Sofa meshes on the left." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 670, + 157, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 157, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 157, + 682 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "In this section, we perform additional experiments aimed at studying the impact of different factors on ToF-Splatting." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 556 + ], + "type": "text", + "content": "Multi-Frame Integration. In Table 3, we ablate the main components of the multi-frame integration and assess their contribution to the whole SLAM framework. By retaining monocular cues alone at the expense of multi-view geometry, we observe severe drops in both tracking and mapping accuracy, confirming the paramount importance of the latter; monocular cues alone are effective only where multiview cues are ineffective – e.g., in the absence of texture, as occurs mostly in the Kitchen scene." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "type": "text", + "content": "TUM RGB-D. In Table 4, we simulate sparse ToF data by sampling " + }, + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "type": "inline_equation", + "content": "0.04\\%" + }, + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "type": "text", + "content": " depth points, and compare ToF-Splatting with existing RGB and RGB-D methods. With as few as " + }, + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "type": "text", + "content": " of points, ToF-Splatting largely improves over RGB methods [22, 32], approaching the " + }, + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "type": "text", + "content": " density RGB-D approach of the most representative competitor [22] at only " + }, + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "type": "inline_equation", + "content": "0.04\\%" + }, + { + "bbox": [ + 313, + 557, + 556, + 640 + ], + "type": "text", + "content": " density." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Replica: Input Depth Sparsity. We conduct additional studies on the Replica dataset, for which Figure 4 provides a qualitative overview of our 3D reconstructions. ToF-Splatting enables performing SLAM with very few sparse points – the 64 sparse points provided by a " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " sparse account only for " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " of a " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "640 \\times 480" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " image. We here" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28350" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 72, + 527, + 194 + ], + "blocks": [ + { + "bbox": [ + 86, + 72, + 527, + 194 + ], + "lines": [ + { + "bbox": [ + 86, + 72, + 527, + 194 + ], + "spans": [ + { + "bbox": [ + 86, + 72, + 527, + 194 + ], + "type": "image", + "image_path": "beff949188f76e61ce25e2f0d0d661966e91962c7328414bac13b1b9985dd5c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 232 + ], + "type": "text", + "content": "Figure 4. Replica Qualitatives. We provide qualitative results on Replica [28] to demonstrate the generalization capabilities of our method. On the left from top to bottom, meshes obtained respectively on scenes Office2 and Room2. On the right, details from the scene Room0. ToF-Splating delivers accurate details and allows for nice photometric and depth rendering." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 71, + 242, + 282, + 322 + ], + "blocks": [ + { + "bbox": [ + 71, + 242, + 282, + 322 + ], + "lines": [ + { + "bbox": [ + 71, + 242, + 282, + 322 + ], + "spans": [ + { + "bbox": [ + 71, + 242, + 282, + 322 + ], + "type": "image", + "image_path": "5507bdfda1e2142e9434ff9414b17a3166fbf7544d3334aefe737a3d70836de0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 324, + 297, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 324, + 297, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 297, + 369 + ], + "type": "text", + "content": "Figure 5. Impact of depth sparsity. We test on Replica [28] with different simulated depth sparsity levels to assess the capability to exploit higher input densities. MAE and ATE smoothly decrease, whereas rendering metrics appear to be less affected." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 71, + 373, + 282, + 453 + ], + "blocks": [ + { + "bbox": [ + 71, + 373, + 282, + 453 + ], + "lines": [ + { + "bbox": [ + 71, + 373, + 282, + 453 + ], + "spans": [ + { + "bbox": [ + 71, + 373, + 282, + 453 + ], + "type": "image", + "image_path": "24c5def5c78c0a1803aae4182b7e0931978cdc3bbe579554c5c536f42e892d08.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 455, + 295, + 500 + ], + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 500 + ], + "type": "text", + "content": "Figure 6. Impact of depth noise. We test on Replica [28] injecting different amounts of noise " + }, + { + "bbox": [ + 55, + 455, + 295, + 500 + ], + "type": "inline_equation", + "content": "\\xi \\in [0.00, 0.01, 0.05, 0.10]" + }, + { + "bbox": [ + 55, + 455, + 295, + 500 + ], + "type": "text", + "content": ". ToF-Splatting demonstrates to be effective at dealing with noise, with error increasing almost linearly with the injected amount of noise." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 579 + ], + "type": "text", + "content": "explore the impact of higher sparse points density on its overall performance. We simulate on Replica densities of " + }, + { + "bbox": [ + 55, + 509, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\{0.02\\%, 0.04\\%, 0.08\\%, 0.16\\% \\}" + }, + { + "bbox": [ + 55, + 509, + 295, + 579 + ], + "type": "text", + "content": " corresponding to about " + }, + { + "bbox": [ + 55, + 509, + 295, + 579 + ], + "type": "inline_equation", + "content": "\\{64, 128, 256, 512\\}" + }, + { + "bbox": [ + 55, + 509, + 295, + 579 + ], + "type": "text", + "content": " points. Figure 5 shows the performance variation on both tracking and mapping metrics using the aforementioned sparsities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 581, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 664 + ], + "type": "text", + "content": "Replica: Input Depth Noise. We study various noise levels impact in the input depth to assess our pipeline robustness. We model noise as additive but dependent on depth, injecting heteroscedastic Gaussian noise " + }, + { + "bbox": [ + 55, + 581, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(d,\\epsilon d)" + }, + { + "bbox": [ + 55, + 581, + 295, + 664 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 581, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 581, + 295, + 664 + ], + "type": "text", + "content": " modulates the variance of the noise source on Replica. Figure 6 shows that MAE, PSNR, and ATE all expose linear degradation trends as " + }, + { + "bbox": [ + 55, + 581, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 581, + 295, + 664 + ], + "type": "text", + "content": " increases." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Runtime. In Table 5 we measure the runtime of ToF-Splatting in comparison with the main baselines. The tracking time of ToF-Splatting already includes depth estimates with DoD, with average runtime " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\bar{t}_{\\mathrm{DoD}} \\approx 90 ms" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " (compris" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 240, + 554, + 306 + ], + "blocks": [ + { + "bbox": [ + 315, + 240, + 554, + 306 + ], + "lines": [ + { + "bbox": [ + 315, + 240, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 315, + 240, + 554, + 306 + ], + "type": "table", + "html": "
Method NameTracking TimeMapping Time
iMAP [30]101 ms448 ms
NICE-SLAM [46]470 ms1300 ms
MonoGS [22]1169 ms650 ms
ToF-SLAM [19]116 ms380 ms
ToF-Splatting (ours)614 ms865 ms
", + "image_path": "213d997618fb32462fd6c9205e2a5dce00b4604d8e15988c0038993546e0dcab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 310, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 310, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 555, + 354 + ], + "type": "text", + "content": "Table 5. Runtime Comparison. We compare the runtime of ToF-Splatting and other learned baselines measuring their runtime for tracking and mapping. MonoGS [22] tracking time contains also the depth completion inference time [48]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 379, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 463 + ], + "type": "text", + "content": "ing DepthAnything v2 for monocular cues inference [40]). To date, 3DGS-based mapping methods are not capable of real-time performances. Nevertheless, exploiting the fact that the mapping step accounts only for just a few frames (typically " + }, + { + "bbox": [ + 313, + 379, + 555, + 463 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 379, + 555, + 463 + ], + "type": "text", + "content": " of the total in sequences of the ZJUL5 dataset [19]) and can be parallelized, the pipeline achieves 1.5 fps without further optimization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 468, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 564 + ], + "type": "text", + "content": "Limitations. Even with state-of-the-art quality, ToF-Splatting manifests some issues that will be addressed in future developments. The most concerning is the runtime, which is too slow for real-time applications and requires memory and compute-intensive backpropagation at deployment. Even though fast 3DGS frameworks exist, such methodologies have not been transferred yet into the SLAM community, that being outside the scope of this work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 592, + 388, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 592, + 388, + 604 + ], + "spans": [ + { + "bbox": [ + 314, + 592, + 388, + 604 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "We presented the first 3DGS-based SLAM pipeline relying on sparse ToF depth sensing, as it provides accurate tracking and mapping from cheap and low-power cameras. Moreover, we showed how this result can be achieved through the integration of multi-view geometry, sparse depth data, and monocular cues, yielding an end-to-end 3DGS-based SLAM system. Finally, we assessed the effectiveness of our approach on the ZJUL5 and Replica datasets." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28351" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Carlos Campos, Richard Elvira, Juan J Gómez Rodríguez, José MM Montiel, and Juan D Tardós. Orb-slam3: An accurate open-source library for visual, visual-inertial, and multimap slam. IEEE Transactions on Robotics, 37(6):1874-1890, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "text", + "content": "[2] Hanlin Chen, Fangyin Wei, Chen Li, Tianxin Huang, Yunsong Wang, and Gim Hee Lee. Vcr-gaus: View consistent depth-normal regularizer for gaussian surface reconstruction. arXiv preprint arXiv:2406.05774, 2024. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "text", + "content": "[3] Andrea Conti, Matteo Poggi, Valerio Cambareri, and Stefano Mattoccia. Depth on demand: Streaming dense depth from a low frame-rate active sensor. In European Conference on Computer Vision (ECCV), 2024. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 295, + 280 + ], + "type": "text", + "content": "[4] Andrea Conti, Matteo Poggi, Valerio Cambareri, and S. Mattoccia. Range-agnostic multi-view depth estimation with keyframe selection. 2024 International Conference on 3D Vision (3DV), pages 1350–1359, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 345 + ], + "type": "text", + "content": "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5828-5839, 2017. 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 295, + 401 + ], + "type": "text", + "content": "[6] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. BundleFusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4): 1, 2017. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 468 + ], + "type": "text", + "content": "[7] Tianchen Deng, Guole Shen, Tong Qin, Jianyu Wang, Wentao Zhao, Jingchuan Wang, Danwei Wang, and Weidong Chen. Plgslam: Progressive neural scene representation with local to global bundle adjustment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19657-19666, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 468, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 295, + 514 + ], + "type": "text", + "content": "[8] Zilong Dong, Guofeng Zhang, Jiaya Jia, and Hujun Bao. Keyframe-based real-time camera tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1538-1545. IEEE, 2009. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 547 + ], + "type": "text", + "content": "[9] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 548, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 295, + 624 + ], + "type": "text", + "content": "[10] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. KinectFusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User Interface Software and Technology, pages 559-568, 2011. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 680 + ], + "type": "text", + "content": "[11] Mohammad Mahdi Johari, Camilla Carta, and François Fleuret. Eslam: Efficient dense slam system based on hybrid representation of signed distance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17408-17419, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "text", + "content": "[12] Nikhil Keetha, Jay Karhade, Krishna Murthy Jatavallabhula, Gengshan Yang, Sebastian Scherer, Deva Ramanan, and Jonathon Luiten. Splatam: Splat track & map 3d gaussians" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "for densergb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21357-21366, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[13] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 2023. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[14] Georg Klein and David Murray. Parallel tracking and mapping for small AR workspaces. In Proceedings of the IEEE and ACM International Symposium on Mixed and Augmented Reality, pages 225-234. IEEE, 2007. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "text", + "content": "[15] Yanyan Li, Nikolas Brasch, Yida Wang, Nassir Navab, and Federico Tombari. Structure-slam: Low-drift monocular slam in indoor environments. IEEE Robotics and Automation Letters, 5(4):6583-6590, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 242, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 297 + ], + "type": "text", + "content": "[16] Yijin Li, Xinyang Liu, Wenqian Dong, Han Zhou, Hujun Bao, Guofeng Zhang, Yinda Zhang, and Zhaopeng Cui. Deltar: Depth estimation from a light-weight tof sensor and rgb image. In European Conference on Computer Vision, 2022. 3, 4, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 354 + ], + "type": "text", + "content": "[17] Lorenzo Liso, Erik Sandström, Vladimir Yugay, Luc Van Gool, and Martin R Oswald. Loopy-slam: Dense neural slam with loop closures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20363-20373, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 399 + ], + "type": "text", + "content": "[18] Haomin Liu, Guofeng Zhang, and Hujun Bao. Robust keyframe-based monocular SLAM for augmented reality. In Proceedings of the IEEE International Symposium on Mixed and Augmented Reality, pages 1-10. IEEE, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 463 + ], + "type": "text", + "content": "[19] Xinyang Liu, Yijin Li, Yanbin Teng, Hujun Bao, Guofeng Zhang, Yinda Zhang, and Zhaopeng Cui. Multi-modal neural radiance field for monocular dense slam with a light-weight tof sensor. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1-11, 2023. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 466, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 499 + ], + "type": "text", + "content": "[20] Gregor Luetzenburg, Aart Kroon, and Anders A. Bjørk. Evaluation of the Apple iPhone 12 Pro LiDAR for an Application in Geosciences. Scientific Reports, 11(1), 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "text", + "content": "[21] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. 2024 International Conference on 3D Vision (3DV), pages 800-809, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "text", + "content": "[22] Hidenobu Matsuki, Riku Murai, Paul H.J. Kelly, and Andrew J. Davison. Gaussian splattering slam. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18039-18048, 2023. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 590, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 646 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision, pages 405-421. Springer, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[24] Raul Mur-Artal and Juan D. Tardós. ORB-SLAM2: An Open-Source SLAM System for Monocular, Stereo, and RGB-D Cameras. IEEE Transactions on Robotics, 33(5): 1255–1262, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[25] Xin Qiao, Matteo Poggi, Pengchao Deng, Hao Wei, Chenyang Ge, and Stefano Mattoccia. Rgb guided tof imag-" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28352" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 105 + ], + "type": "text", + "content": "ing system: A survey of deep learning-based methods. International Journal of Computer Vision, pages 1-38, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 151 + ], + "type": "text", + "content": "[26] Erik Sandström, Yue Li, Luc Van Gool, and Martin R Oswald. Point-slam: Dense neural point cloud-based slam. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18433-18444, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 196 + ], + "type": "text", + "content": "[27] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 307 + ], + "type": "text", + "content": "[28] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 2, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 295, + 363 + ], + "type": "text", + "content": "[29] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of RGB-D SLAM systems. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 573-580. IEEE, 2012. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 409 + ], + "type": "text", + "content": "[30] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 410, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 295, + 464 + ], + "type": "text", + "content": "[31] Keisuke Tateno, Federico Tombari, Iro Laina, and Nassir Navab. Cnn-slam: Real-time dense monocular slam with learned depth prediction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6243-6252, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 466, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 295, + 510 + ], + "type": "text", + "content": "[32] Zachary Teed and Jia Deng. Droid-SLAM: Deep visual SLAM for monocular, stereo, and RGB-D cameras. Advances in Neural Information Processing Systems, 34: 16558-16569, 2021. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 555 + ], + "type": "text", + "content": "[33] Fabio Tosi, Youmin Zhang, Ziren Gong, Erik Sandström, Stefano Mattoccia, Martin R. Oswald, and Matteo Poggi. How nerfs and 3d gaussian splattering are reshaping slam: a survey, 2024. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 295, + 610 + ], + "type": "text", + "content": "[34] Hengyi Wang, Jingwen Wang, and Lourdes Agapito. Coslam: Joint coordinate and sparse parametric encodings for neural real-time slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13293-13302, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 656 + ], + "type": "text", + "content": "[35] Yifan Wang, Felice Serena, Shihao Wu, Cengiz Öztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38:1-14, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[36] Thomas Whelan, Renato F Salas-Moreno, Ben Glocker, Andrew J Davison, and Stefan Leutenegger. ElasticFusion: Real-time dense SLAM and light source estimation. The International Journal of Robotics Research, 35(14):1697-1716, 2016. 1, 2, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 699 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "text", + "content": "[37] Xingming Wu, Zimeng Liu, Yuxin Tian, Zhong Liu, and Weihai Chen. Kn-slam: Keypoints and neural implicit encoding slam. IEEE Transactions on Instrumentation and Measurement, 73:1-12, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "type": "text", + "content": "[38] Liu Xinyang, Li Yijin, Teng Yanbin, Bao Hujun, Zhang Guofeng, Zhang Yinda, and Cui Zhaopeng. Multi-modal neural radiance field for monocular dense slam with a lightweight tof sensor. In International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 554, + 228 + ], + "type": "text", + "content": "[39] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19595-19604, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 229, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 554, + 262 + ], + "type": "text", + "content": "[40] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. ArXiv, abs/2406.09414, 2024. 4, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 554, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 554, + 296 + ], + "type": "text", + "content": "[41] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. ArXiv, abs/2312.10070, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 330 + ], + "type": "text", + "content": "[42] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R. Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 554, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 384 + ], + "type": "text", + "content": "[43] Guofeng Zhang, Jiaya Jia, Tien-Tsin Wong, and Hujun Bao. Recovering consistent video depth maps via bundle optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1-8. IEEE, 2008. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 387, + 554, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 554, + 441 + ], + "type": "text", + "content": "[44] Youmin Zhang, Fabio Tosi, Stefano Mattoccia, and Matteo Poggi. Go-slam: Global optimization for consistent 3d instant reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3727-3737, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 442, + 554, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 554, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 554, + 497 + ], + "type": "text", + "content": "[45] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12786-12796, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 498, + 554, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 554, + 563 + ], + "type": "text", + "content": "[46] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. NICE-SLAM: Neural implicit scalable encoding for SLAM. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 565, + 554, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 565, + 554, + 619 + ], + "spans": [ + { + "bbox": [ + 316, + 565, + 554, + 619 + ], + "type": "text", + "content": "[47] Zihan Zhu, Songyou Peng, Viktor Larsson, Zhaopeng Cui, Martin R. Oswald, Andreas Geiger, and Marc Pollefeys. Nicer-slam: Neural implicit scene encoding for rgb slam. 2024 International Conference on 3D Vision (3DV), pages 42-52, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 621, + 554, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 621, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 621, + 554, + 654 + ], + "type": "text", + "content": "[48] Yiming Zuo and Jia Deng. Ogni-dc: Robust depth completion with optimization-guided neural iterations. arXiv preprint arXiv:2406.11711, 2024. 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 655, + 554, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 655, + 554, + 699 + ], + "spans": [ + { + "bbox": [ + 316, + 655, + 554, + 699 + ], + "type": "text", + "content": "[49] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. Surface splatting. Proceedings of the 28th annual conference on Computer graphics and interactive techniques, 2001. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28353" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Token Activation Map to Visually Explain Multimodal LLMs/b852d151-66d5-47a7-8861-5fbe9cff3adc_content_list.json b/2025/Token Activation Map to Visually Explain Multimodal LLMs/b852d151-66d5-47a7-8861-5fbe9cff3adc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1a2d82e9cc6b8b5e21c45d05726642778cbe88a7 --- /dev/null +++ b/2025/Token Activation Map to Visually Explain Multimodal LLMs/b852d151-66d5-47a7-8861-5fbe9cff3adc_content_list.json @@ -0,0 +1,1615 @@ +[ + { + "type": "text", + "text": "Token Activation Map to Visually Explain Multimodal LLMs", + "text_level": 1, + "bbox": [ + 189, + 130, + 808, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yi Li, Hualiang Wang, Xinpeng Ding, Haonan Wang, Xiaomeng Li*", + "bbox": [ + 233, + 180, + 772, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology, Hong Kong SAR, China.", + "bbox": [ + 99, + 200, + 898, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ylini,xmli} $@$ ust.hk", + "bbox": [ + 419, + 215, + 583, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal large language models (MLLMs) are broadly empowering various fields. Despite their advancements, the explainability of MLLMs remains less explored, hindering deeper understanding, model credibility, and effective visualization. Unlike conventional vision models (e.g., CNNs, ViTs, CLIP) that produce a single output, MLLMs generate sequences of tokens progressively, where each generated token depends on the previous context. Therefore, earlier context tokens can introduce redundant activations that interfere with the explanation of later tokens beyond their original information. Existing studies often overlook this issue, but our observations reveal that these redundant correlations can significantly hurt the reliability of explanations. To address this, we propose an estimated causal inference method to mitigate the interference of context to achieve high-quality MLLM explanation, with a novel rank Gaussian filter to further reduce activation noises. We term this method Token Activation Map (TAM) to highlight the consideration of interactions between tokens. TAM also indicates that it excels at explaining multiple tokens of MLLM, which is different from the Class Activation Map (CAM) for a single prediction. Our TAM method significantly outperforms existing SoTA methods, showcasing high-quality visualization results that can be utilized for various scenarios, such as object localization, failure case analysis, video visualization, MLLMs visual comparison, and model understanding (e.g., color, shape, action, location, visual reasoning, multi-turn conversation, etc). The code is available at github.com/xmed-lab/TAM.", + "bbox": [ + 88, + 301, + 485, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 771, + 220, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal large language models (MLLMs or Multimodal LLMs) increasingly empower wide applications [17, 29, 30, 34], enabling multimodal inputs (e.g., images, videos, text) and human-like conversations. Although extensive efforts have been devoted to enhancing the performance of", + "bbox": [ + 89, + 797, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2a3412272f6e2129275a35c47ac409a77f84afd6c81c3ea480de827eafa11480.jpg", + "image_caption": [ + "(a) Conventional models vs. MLLMs" + ], + "image_footnote": [], + "bbox": [ + 517, + 267, + 669, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6f335c94f06f1a55049345c3fa1d7115795ab805e3cbcb67385e68d60bbc04c.jpg", + "image_caption": [ + "(b) Information flow of MLLM" + ], + "image_footnote": [], + "bbox": [ + 669, + 268, + 802, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bccf1e88009a26b6bebdf77574a705286367ea6675bfa676d7123bb27ac6c61f.jpg", + "image_caption": [ + "(C) Correlated activations" + ], + "image_footnote": [], + "bbox": [ + 802, + 268, + 906, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3859da4497307882cd97f8f1d183b03fea829e0040d55fb37c4fbdcffb775218.jpg", + "image_caption": [ + "(d) Qualitative example of interference activations" + ], + "image_footnote": [], + "bbox": [ + 516, + 414, + 647, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/85dc2c305cb925a43f0289596bfe6c312c26f97e3bbbf8d8f9f906a48f352a7d.jpg", + "image_caption": [ + "(e) TAM (ours) for MLLM explanation" + ], + "image_footnote": [], + "bbox": [ + 645, + 414, + 772, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57d544af825f0f08bf5f11b60798d366f21eea240fe56dd630a5c65cc905e931.jpg", + "image_caption": [ + "Figure 1. Illustration of Motivation. (a) MLLMs generate multiple tokens beyond a single output. (b) The information flow indicates that MLLM generates tokens progressively, where one generated token (each row) is correlated with the context (prompt + earlier answer tokens). (c) We randomly pair CAMs and count their L1 distance against text correlation. Higher text correlation corresponds to lower distance, indicating concurrent interferences. (d) The example shows that the context token \"plate\" introduces interference activations (marked by white boxes) to the later explained token \"fork\". (e) Our TAM reveals the original information of the explained token \"fork\", eliminating context interferences and beyond CAM for a single output in (d). Results in (c, d, e) derive from Qwen2-VL-2B on the COCO Caption dataset." + ], + "image_footnote": [], + "bbox": [ + 772, + 414, + 906, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MLLMs [4, 16, 35, 47, 51], research on MLLM explainability remains less explored. It is crucial for user trust, model understanding, case analysis, and visualization. However, exploring the explainability of MLLMs is more challenging than conventional visual models (such as CNN [24], ViT [23], CLIP [41]), where MLLMs progressively generate multiple tokens beyond a single output as Fig. 1a. Thus, we aim to explain all tokens of MLLM, revealing the visual cues via token-level activation maps (e.g., Fig. 1e).", + "bbox": [ + 511, + 718, + 906, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current MLLM explanation works [7, 48, 55] are generally borrowed from techniques [38, 45] for conventional models. For example, LLaVA-CAM [55] uses Smooth", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "48", + "bbox": [ + 490, + 945, + 506, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Grad-CAM++ [38] to explain LLaVA [35]. The self-attention weights [49] are applied to understand math problem-solving [48]. Besides, LVLM-Interpret [7] integrates some conventional methods as a tool. Unfortunately, existing visual explainability methods are designed for conventional models [23, 24, 41] that produce a single output, such as class activation maps (CAM) series [9, 26, 45, 56], attention-based approaches [1, 28], attention with relevance [2, 11], model-agnostic methods [36, 43], or multimodal explanation [3, 19, 31, 32]. However, these methods overlook that MLLMs generate multiple tokens progressively, where earlier context tokens interfere with later tokens to be explained. As shown in Fig. 1d, the conventional CAM [56] for the token \"fork\" is confounded by numerous redundant activations concurrent with the token \"plate\" in context (context refers to prompt and earlier answer tokens than the explained token). Quantitative analysis in Fig. 1c reveals that this issue is widespread, where correlated tokens lead to close and interference activations. These interferences obscure the original information conveyed by the explained token, significantly degrading the explainability of MLLMs. Notably, this activation interference across tokens is unique to MLLMs and has been overlooked.", + "bbox": [ + 89, + 90, + 480, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We aim to reveal the original information for the explained token while minimizing interference from context, as discussed above. Herein, the original information reveals a causal relation [40] between the current token and the input, while the interferences represent the correlations that need to be excluded. Inspired by the potential outcome model (POM) of causal inference [27], we propose the estimated causal inference to explore this causal relation. Unlike raw causal inference by comparing ready-made results (yes vs. no), there is no single given output to represent the \"no\" result. Therefore, we estimated interferences from multiple context tokens that are highly related to the target, with a scale factor optimized by the least-squares method to generate the refined causal activation map. We also propose a novel rank Gaussian filter to reduce noises, thereby further enhancing the quality of activation maps. The above methods are collectively referred to as token activation map (TAM), emphasizing the consideration of interactions between tokens. TAM also indicates that it excels at explaining tokens in multiple rounds of MLLM, which differs from the Class Activation Map (CAM) for a single prediction.", + "bbox": [ + 89, + 441, + 482, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimentally, the proposed TAM demonstrates quantitative improvement over existing SoTA explainability methods by more than $8.96\\%$ on the COCO Caption dataset [13] and $8.54\\%$ on the OpenPSG dataset [57] when applied to visually explain the Qwen2-VL-2B [51]. Meanwhile, it is complementary to existing methods to boost their performance. We further verify the scalability and applicability on 7 advanced MLLMs from the series of Qwen2-VL [51], InternVL2_5 [15] and LLaVA1_5 [35], where the", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performance improvements range from $5.45\\%$ to $11.0\\%$ on COCO Caption dataset [13]. Besides quantitative results, we conduct qualitative experiments to understand MLLM attributes about color, shape, action, location, etc. Furthermore, we illustrate that TAM is capable of explaining failure cases on the QK-VQA dataset. [37], and it supports video visualization well on the STAR dataset [52]. In addition, it can be used as a tool to visually compare MLLMs qualitatively. Moreover, TAM shows wide applicability to explain multi-turn conversations, multi-image scenarios, and visual reasoning (see the catalog Table 5 in supplementary for extensive examples). Our work has three main contributions:", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce TAM, a novel approach to explain MLLMs, which produces high-quality activation maps for multiple generative tokens by incorporating the estimated causal inference to alleviate interference from context tokens, with the rank Gaussian filter to reduce activation noises.", + "TAM significantly outperforms current methods for MLLM explanation, while also complementing them.", + "- TAM demonstrates wide applicability and scalability, serving as a versatile tool for object localization, model understanding, failure case analysis, video visualization, MLLM comparison, and supporting diverse scenarios (e.g., multi-turn, multi-image, visual reasoning)." + ], + "bbox": [ + 511, + 273, + 903, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 467, + 653, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Explainability for Multimodal LLM. The emergence of MLLMs such as LLaVA [35], Qwen2-VL [51], InternVL [15], and closed-source GPT-4o [25] has significantly impacted various tasks involving multimodal inputs like text, images, and video. We mainly focus on the visual explanation of MLLMs via the activation maps for the input image or video frames. Regarding the modality perspective, the explainability of MLLMs is partially related to methods to explain multimodal models such as CLIP Surgery [32] and transformer register [19] designed for CLIP [41]; InterpreT [3] and Bi-Modal [10] for BERT [20]. From the technical aspect, conventional explanation methods are also related, including class activation map series (CAM [56], Grad-CAM [45], Grad-CAM++ [9], LayerCAM [26]), attention-based mechanisms (LRP [28], Rollout [1]), the combination (AttnLRP [2], Grad×AttnRoll [11]), or model-agnostic methods (LIME [43], SHAP [36]).", + "bbox": [ + 511, + 492, + 906, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although existing methods may be valid to explain MLLMs, they are not the optimal solution. Because they are usually designed for conventional models (e.g., CNN [24], ViT [23], CLIP [41]) that produce a single output rather than generating multiple tokens progressively in MLLMs. We have observed that these earlier context tokens interfere with later tokens by introducing redundant visual activations, as shown in Fig. 1. This phenomenon is first studied by us, solved by a novel estimated causal inference method, which is specially designed for MLLM. Notably,", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "49", + "bbox": [ + 490, + 945, + 506, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "some workshop papers (LLaVA-CAM [55] and LVLM-Interpret [7]) applied conventional methods [38, 44] to explain LLaVA [35], but they also overlooked the interference from context tokens and treated each token independently. Causal Model. Causal inference [54] seeks to establish causal relation [40] between variables, focusing on how changes in one variable influence another. In contrast to statistical correlation analysis, it emphasizes causal relations. Another related concept, causal intervention [27], involves manipulating variables using a do operator. Model-agnostic explainability methods, such as LIME [43] and SHAP [36], can be viewed as causal interventions that explore the relationship between inputs and structured outputs by masking inputs or selecting subsets [12]. This concept is also applied to transformers through token masking (CLEANN [44]) or hidden state replacement [39]. However, causal intervention for MLLMs is impractical. Because there are too many input tokens are outputs in MLLM, requiring unbearable inference times. Besides, modified input leads to a changed context every inference, which is hard to evaluate. Our method is inspired by the potential outcome model (POM) to achieve causal inference without additional model inferences. Our contribution lies in estimating the unprocessed output from multiple tokens rather than a single ready-made output, considering the uniqueness of MLLMs.", + "bbox": [ + 89, + 90, + 480, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer Denoising. Activation maps of transformers [49] usually present many noise activations, impeding high-quality visualization. Recent studies give their explanations about it, including attention sink [53], lack of register [19], and redundant features across classes [32]. In MLLM, these noises still exist obversely, even system tokens already play the role of registers, or deploying the feature surgery to mitigate redundant features (See Table 2). To solve this problem practically, we aim to introduce denoising filters to mitigate it as post-processing. Since these noises belong to salt-and-pepper noise morphologically, we apply the median filter, adaptive median filter [8], and the Gaussian filter. However, these filters are not the optimal solution, where the Gaussian filter keeps too much noise signal and the median filter overlooks smaller responses. Therefore, we propose the rank Gaussian filter, a novel, simple, and effective filter to denoise transformer activation. It merges ranked values within a sliding window, weighted summed by a 1-d Gaussian kernel with the center at the median rank. Besides, we improve the Gaussian kernel by the coefficient of variation for more robust results. Overall, our method addresses the activation noises in a practical aspect through a new filter.", + "bbox": [ + 89, + 468, + 482, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 814, + 181, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first provide an overview of the proposed Token Activation Map (TAM), as illustrated in Fig. 2. Next, we elaborate on two key modules within TAM: the estimated causal inference and the rank Gaussian filter. Finally,", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we introduce three metrics designed to enable fine-grained evaluations of the explanations for MLLMs.", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Token Activation Map", + "text_level": 1, + "bbox": [ + 511, + 128, + 720, + 145 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TAM aims to depict explainable activation maps for multiple tokens generated from MLLMs by leveraging causal inference to mitigate inter-token interference. To be more specific, given the visual data and prompt text as the input, the MLLM is employed to progressively generate the visual features $\\pmb{F}^{v} \\in \\mathbb{R}^{n_{v} \\times c}$ , prompt features $\\pmb{F}^{p} \\in \\mathbb{R}^{n_{p} \\times c}$ and answer features $\\pmb{F}^{a} \\in \\mathbb{R}^{n_{a} \\times c}$ . $n_{v}, n_{p}, n_{a}$ and $c$ indicate the number of three types of features, and the feature dimension, respectively. Next, a token classifier, instantiated as a fully connected layer, is used to generate answer tokens $\\pmb{t}^{a} = \\{t_{1}^{a} \\dots t_{n_{a}}^{a}\\}$ from these features. Prompt tokens are defined as $\\pmb{t}^{p} = \\{t_{1}^{p} \\dots t_{n_{p}}^{p}\\}$ similarly. Based on the variables described above, the visual activation maps for the prompt tokens and answer tokens are calculated as follows:", + "bbox": [ + 511, + 151, + 906, + 362 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} _ {i} ^ {p} = \\left\\lfloor \\boldsymbol {F} ^ {v} \\boldsymbol {w} _ {t _ {i} ^ {p}} \\right\\rfloor_ {+}, i \\in [ 1, n _ {p} ]\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 369, + 803, + 387 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} _ {i} ^ {a} = \\left\\lfloor \\boldsymbol {F} ^ {v} \\boldsymbol {w} _ {t _ {i} ^ {a}} \\right\\rfloor_ {+}, i \\in [ 1, n _ {a} ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 382, + 903, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{A}_i^p \\in \\mathbb{R}^{n_v \\times 1}$ and $\\mathbf{A}_i^a \\in \\mathbb{R}^{n_v \\times 1}$ present the activation map for the $i$ -th prompt token $t_i^p$ and answer token $t_i^a$ . $\\mathbf{w}_{t_i^p} \\in \\mathbb{R}^{c \\times 1}$ and $\\mathbf{w}_{t_i^a} \\in \\mathbb{R}^{c \\times 1}$ are the corresponding weight vectors within the token classifier. $\\lfloor \\rfloor_+$ is the function to keep positive activations.", + "bbox": [ + 511, + 416, + 905, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Then, the activation maps for prompt and answer tokens are concatenated $(\\bot)$ as $\\pmb{A}_{:n_p + n_a} = \\pmb{A}^p\\bot \\pmb{A}^a$ . For the $i$ -th answer token, its context token is formulated as $\\pmb{A}_{:n_p + i - 1}$ , including all earlier context maps (prompt tokens + answer tokens earlier than $i$ ). On top of it, the refined activation maps for answer tokens are formulated as follows:", + "bbox": [ + 511, + 492, + 905, + 582 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\boldsymbol {A}} _ {i} ^ {a} = \\left\\lfloor \\mathcal {D} \\left(\\boldsymbol {A} _ {i} ^ {a} - s \\mathcal {E} \\left(\\boldsymbol {A} _ {: n _ {p} + i - 1}\\right)\\right) \\right\\rfloor_ {+}, i \\in [ 1, n _ {a} ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 590, + 903, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{A}_i^a$ indicates the refined activation map for the $i$ -th answer token. The number of $i$ increases sequentially from 1 to $n_a$ . $A_{i}^{a} - s\\mathcal{E}(A_{:n_{p} + i - 1})$ is the proposed estimated causal inference that produces the causal activation map with the estimation function $\\mathcal{E}$ described in Sec. 3.2. $\\mathcal{D}$ is the proposed rank Gaussian filter module in Sec. 3.3.", + "bbox": [ + 511, + 619, + 905, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the last, we concatenate $(\\perp)$ the refined visual activation map $\\bar{A}_i^a$ of the $i$ -th answer token with the raw textual relevance $\\mathbf{r}_i \\in \\mathbb{R}^{n_p + i - 1}$ for a multimodal activation map (image or video with text) $M_i \\in \\mathbb{R}^{n_v + n_p + i - 1}$ as follows:", + "bbox": [ + 511, + 710, + 905, + 771 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} _ {i} = \\mathcal {N} \\left(\\bar {\\boldsymbol {A}} _ {i} ^ {a} \\perp \\mathbf {r} _ {i}\\right), i \\in [ 1, n _ {a} ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 780, + 903, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{N}()$ is a min-max normalization function for visualizing visual and textual activations at an aligned level.", + "bbox": [ + 511, + 809, + 905, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Estimated Causal Inference", + "text_level": 1, + "bbox": [ + 511, + 848, + 761, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall, Eq. 2 optimizes later visual activation maps based on visual activation maps of all earlier context tokens. To", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "50", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eeac58906753acd5b559578b2f97e86510bf04ec0dc649f6e62c85b13ec2f662.jpg", + "image_caption": [ + "Figure 2. Illustration of method. (a, Eq. 1 - Eq. 3) The overall framework of TAM. (b, Eq. 4 - Eq. 5) Details of the estimated casual inference module. (d, Eq. 6 - Eq. 7) Details of the rank Gaussian filter module. (d, Eq. 8 - Eq. 10) Fine-grained evaluation metrics." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 315, + 333 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/69b6e0b7bca0597e4ebf01775f955dde8c33175b395bc38a0322b3ced06b1c6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 88, + 529, + 333 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0611e6074c1428e876590a78e1a68a1f020fffed3d6fa478a4d73e48bec275e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 539, + 88, + 763, + 333 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/272d80884ac3d74fc7477ac3a6125281c6eea90bf7d5a6a512e62d6c8fcdd9c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 88, + 903, + 333 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "be more specific, causal inference is employed to investigate the correct causal relation between the current answer token and visual tokens, eliminating interferences from prompt tokens and preceding answer tokens. To this end, we quantify the interference map from the context tokens for the $i$ -th answer token as follows:", + "bbox": [ + 89, + 396, + 483, + 486 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} \\left(\\boldsymbol {A} _ {: n _ {p} + i - 1}\\right) = \\sum_ {k = 1} ^ {n _ {p} + i - 1} \\frac {r _ {i} ^ {k}}{\\sum \\mathbf {r} _ {i} + \\epsilon} \\boldsymbol {A} _ {k} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 494, + 482, + 537 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns. t. r _ {i} ^ {k} = 0, \\mathrm {i f} t _ {i} ^ {a} = t _ {k},\n$$\n", + "text_format": "latex", + "bbox": [ + 261, + 539, + 418, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $r_i^k$ is the textual relevance between the $k$ -th context token $t_k$ and the current answer token $t_i^a$ . The relevance is obtained from activations among textual tokens $\\lfloor (\\pmb{F}^p \\perp \\pmb{F}^a) \\pmb{w}_{t_i^a} \\rfloor_+$ following Eq. 1. If the current context token $t_k$ is the same as the answer token $t_i^a$ , we set their relevance as 0, i.e., $r_i^k = 0$ , to avoid eliminating visual activations for tokens identical to the current answer token.", + "bbox": [ + 89, + 568, + 483, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, Eq. 4 aims for each answer token to encapsulate more visual information, which implies minimizing similarity to the visual semantics of previously context tokens. Meanwhile, we use the scale factor $s$ in Eq. 2 to control the scale of the interference map, ensuring that the influence of the interference map is removed from the activation map at a comparable degree. Specifically, we employ the least squares method to minimize the residual between the two maps, for the optimal value of $s$ , as follows:", + "bbox": [ + 89, + 674, + 483, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns = \\arg \\min _ {s} \\sum_ {j = 1} ^ {n _ {v}} \\left(\\boldsymbol {A} _ {j, i} ^ {a} - s \\mathcal {E} \\left(\\boldsymbol {A} _ {j,: n _ {p} + i - 1}\\right)\\right) ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 820, + 482, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After processing by the estimated causal inference module, there are still many salt-and-pepper noises in activation", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "maps. To address this issue, we will elaborate on the solution in the following section.", + "bbox": [ + 511, + 396, + 906, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Rank Gaussian Filter", + "text_level": 1, + "bbox": [ + 511, + 435, + 714, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As described in Sec. 2 the activation noise exists obviously and belongs to the salt-and-pepper type. Different from existing methods [19, 32] with suboptimal performances on MLLM, we address this problem in an new aspect: image denoising. Compared with existing filters like the Gaussian filter that keeps too much noise signal and the median filter which overlooks smaller responses, our rank Gaussian filter is more robust beyond existing methods. The overall equation of the rank Gaussian filter for any activation map $A_{i,j} \\in \\mathbb{R}^{h \\times w}$ is written as:", + "bbox": [ + 511, + 457, + 906, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {D} \\left(\\boldsymbol {A} _ {i, j}, k\\right) = \\sum \\boldsymbol {A} _ {i, j} \\mathcal {G} \\left(\\boldsymbol {a} _ {i, j}, k\\right), \\forall i \\in H, \\forall j \\in W \\tag {6} \\\\ s. t. \\mathbf {a} _ {i, j} = r a n k (\\mathbf {A} _ {i: i + k, j: j + k}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 619, + 903, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Herein, $A_{i:i + k,j:j + k}$ indicates local values within a sliding window at kernel size $k$ . We sort them to get a ranked array to multiply with a custom 1-d Gaussian kernel $\\mathcal{G}(\\cdot)$ . This custom kernel allows the median to have the largest weight, meanwhile aggregating signals ranked close to the median. It can be regarded as a kind of smoothed median filter to enhance the robustness weighted by the Gaussian kernel.", + "bbox": [ + 511, + 671, + 905, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the custom 1-d Gaussian kernel, we formulate it as Eq. 7. Specifically, $(i - k^2 / / 2)^2$ indicates the distance from the median rank $(k^2 / / 2$ , size exact division 2), and $\\sigma (\\mathbf{a}) / \\mu (\\mathbf{a})$ is the coefficient of variation which is more stable than the standard division $\\sigma ()$ in general Gaussian kernel using the mean $\\mu ()$ only.", + "bbox": [ + 511, + 777, + 905, + 868 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {i} (\\mathbf {a}, k) = e ^ {- \\frac {(i - k ^ {2} / / 2) ^ {2}}{2 (\\sigma (\\mathbf {a}) / \\mu (\\mathbf {a})) ^ {2}}}, \\forall i \\in k ^ {2} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 877, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "51", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Evaluation for MLLM Explanation", + "text_level": 1, + "bbox": [ + 89, + 90, + 398, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike conventional models with a single output, MLLMs generate multiple tokens. To evaluate how well activation maps correspond to objects, we propose two fine-grained IoU-based metrics: Obj-IoU for objects with manual masks, and Func-IoU for backgrounds. These metrics focus on plausibility instead of faithfulness (see concept differences in Supp. E). Plausibility better suits MLLMs since perturbation-based faithfulness tests [2, 11] which cause inconsistent text generation, making evaluation invalid (see detailed reasons in Supp. E).", + "bbox": [ + 88, + 112, + 482, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To be specific, we formulate Obj-IoU in Eq. 8, counting the IoU between activation map $(A_{i})$ after binarization $(\\mathcal{B}()$ using opencv OTSU auto threshold) and the groundtruth mask $G_{i}$ . Since the object classes may be unfixed, we count the IoU for all tokens at size $o$ over the dataset without an average on classes. We obtain the mask by matching the token to mask names. For a word or phrase containing multiple tokens, we record their max IoU for a single object.", + "bbox": [ + 89, + 263, + 483, + 385 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {O b j} - \\operatorname {I o U} = \\frac {1}{o} \\sum_ {i = 1} ^ {o} \\frac {\\mathcal {B} (\\boldsymbol {A} _ {i}) \\cap \\boldsymbol {G} _ {i}}{\\mathcal {B} (\\boldsymbol {A} _ {i}) \\cup \\boldsymbol {G} _ {i}} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 395, + 482, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Func-IoU measures the activation degree of tokens that are unreadable from the image (e.g., \"is\", \"and\", \"so\", \"the\"). Higher Func-IoU indicates fewer false positives and unnecessary responses. Specifically, we record the mean binarization thresholds $b_{i}$ (opencv OTSU auto threshold) for all noun tokens in the same image to divide foreground and background on function words. Activation map lower than it $(A_{i} < b_{i})$ is the background prediction used to count IoU with the all-one matrix $J$ (ground-truth is all background class). It is applied to total $u$ tokens over the whole dataset.", + "bbox": [ + 89, + 429, + 483, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {F u n c - I o U} = \\frac {1}{u} \\sum_ {i = 1} ^ {u} \\frac {\\left(\\boldsymbol {A} _ {i} < b _ {i}\\right) \\cap \\boldsymbol {J}}{\\left(\\boldsymbol {A} _ {i} < b _ {i}\\right) \\cup \\boldsymbol {J}} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 590, + 482, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We try to merge the above IoUs as the overall major metrics. We find their average is insufficient to measure biased and low-quality activation maps. For example, if one method predicts almost all tokens to be background, the Obj-IoU is very low but Func-IoU is high, forming a certain inverse ratio (e.g., $5.74\\%$ vs. $96.5\\%$ of Attention-Rollout [1] in Table 3). One suitable solution is to use the F1 value to merge metrics with a certain inverse ratio. Thus, we have the F1-IoU to evaluate the overall plausibility stably:", + "bbox": [ + 89, + 626, + 483, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F 1 - I o U} = \\frac {2 * \\text {O b j - I o U} * \\text {F u n c - I o U}}{\\text {O b j - I o U} + \\text {F u n c - I o U}} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 768, + 482, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 808, + 223, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 833, + 171, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. For quantitative experiments, we use datasets with texts and pixel-level annotations. Our main dataset is the COCO Caption dataset [13], which draws images and", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "masks from COCO2014 [33]. Since explainability methods require testing only, we use its 5K-image minival split without training. Additional datasets include GranDf [42] (1K images) and OpenPSG [57] (3,176-image validation set). Masks in COCO Caption and GranDf are manually annotated, while OpenPSG masks are integrated by Rasheed et al. [42]. For datasets without masks, we perform qualitative visualization tests (Sec. 4.3), including attribute-only images and the QK-VQA [37] validation set (5,046 VQA samples with textual answers). We also qualitatively evaluate the STAR video dataset [52] (914 videos).", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementations. We use the SciPy minimize function with the BFGS method to optimize the scale factor $s$ in Eq. 5. The only hyperparameter, the kernel size $k$ in Eq. 7, is set to 3, as our observations indicate that salt-and-pepper noises typically cluster between 1 to 5 tokens. A kernel size of 3, consisting of 9 elements, is deemed suitable. Eq. 9 incorporates part-of-speech identification. For Func-IoU and Obj-IoU, we tag each word's part of speech using the pos_tag function from the NLTK Python package (version 3.8.1, see tag details in Supp. E). For word matching to mask names, we employ the lemmatize function from nltk.stem. Note that the thresholds in Eq. 8 and Eq. 9 are automatically derived from OpenCV's OTSU method, requiring no manual operation. Details of the baselines are provided in Supp. D.", + "bbox": [ + 511, + 257, + 906, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 478, + 707, + 494 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ff11e5f34b41fc4d2378868b9fa1d778a66ed22a9001d5238ed757addb6dff78.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ECIRGFObj-IoU (%)Func-IoU (%)F1-IoU (%)
XX21.2351.9330.14
✓X22.4169.0333.84+3.7
X✓24.8243.3431.57+1.33
✓✓27.3768.4439.1+8.96
", + "bbox": [ + 516, + 503, + 901, + 569 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ccce2867c4d653abd5ec0d974ba1addfb3c7dd0c220f1d65f53f00478f554b25.jpg", + "table_caption": [ + "Table 1. Ablation study on the COCO Caption [13] dataset using Qwen2-VL-2B [51]. These two modules are mutually beneficial, the combination exceeds the sum of their individual gains. ECI indicates the proposed estimated causal inference, and RGF is the rank Gaussian filter. Metrics are IoU for object words, IoU for function words, and their F1-score-like combination, respectively." + ], + "table_footnote": [], + "table_body": "
SettingObj-IoU (%)Func-IoU (%)F1-IoU (%)
Baseline21.2351.9330.14
Replace rank Gaussian filter
Adaptive Median [8]25.4868.237.1
Median Filter26.0168.2637.67
Gaussian Filter26.5667.7838.16
Replace estimated causal inference
Feature surgery [32]18.548.6626.81
ECI-mean27.8449.8535.72
ECI-attnWeights27.1154.6136.23
Ours27.3768.4439.1
", + "bbox": [ + 514, + 666, + 908, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Effectiveness study on the COCO Caption [13] dataset using Qwen2-VL-2B [51]. ECI-mean and ECI-attnWeights are candidate implementations of ECI described in Supp. D.", + "bbox": [ + 511, + 806, + 906, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ablation and Effectiveness. We present the ablation study results in Table 1, which indicates that each module of the TAM is valuable. Specifically, the estimated causal infer", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "52", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6c8e84c515b962f58f84113d103427477d9e007aee5284e84c659c0e79533e80.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTypeCOCO CaptionGranDfOpenPSG
Obj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoU
Grad-CAM [45]21.2351.9330.1417.8562.1527.7422.9348.5731.15
Grad-CAM++ [9]Gradient19.5262.8329.7817.373.4228.0122.2159.9532.41
Grad-Rollout [1]1.2799.512.511.499.612.771.5799.5813.08
Layer-CAM [26]11.4384.8820.1513.1182.0922.6214.1285.2924.22
AttentionAttention8.292.8715.079.693.5617.4210.5894.2819.03
Attention-Rollout [1]5.7496.510.837.2196.6513.427.9497.0414.68
CP-LRP [5]Combination9.953.9716.7312.6153.2420.3913.353.3621.3
Attn-LRP [2]9.9252.4116.6912.1552.1919.7212.7852.2620.54
CAM [56]Logit21.2351.9330.1417.8562.1527.7422.9348.5731.15
Archi.-Surgery [32]15.6963.8225.1916.5962.2826.219.8358.7729.65
TAM (ours)27.3768.4439.118.6588.9730.8326.2692.9940.95
Archi.-Surgery [32]+ Ours+4.13+10.68+6.12+1.41+26.48+3.74+1.59+31.92+5.0
Grad-CAM++ [9]+2.83+3.91+3.7+1.26+5.39+2.04+2.75+9.44+4.3
Layer-CAM [26]+5.04-7.95+6.98+2.79+0.52+4.05+6.62-2.49+8.95
CAM [56]+6.14+16.52+8.96+0.79+26.82+3.09+3.33+44.42+9.8
", + "bbox": [ + 96, + 88, + 898, + 295 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fe41abbb7d27820028c61f34410566c33cf83030cacbc71cdadd49446db4b647.jpg", + "table_caption": [ + "Table 3. Comparison with SoTA methods using Qwen2-VL-2B on diverse datasets. We adopt the \"Logit\" type that uses the classifier weights in Eq. 1, without back-propagation in \"Gradient\" and \"Combination\" (Gradient + Attention). Besides, we support FlashAttention [18] and SdpaAttention, which do not return attention weights, while \"Attention\" and \"Combination\" rely on it. Our TAM is also complementary to existing methods with gains marked by \"+\" Note, CAM and Grad-CAM are equivalent as discussed in Supp. D. The major metric is the F1-IoU $(\\%)$ to reflect the overall result, merged from IoUs for Object words (Obj-IoU) and function words (Func-IoU)." + ], + "table_footnote": [], + "table_body": "
CAM [56] +OursCOCO CaptionGranDfOpenPSG
Obj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoU
LLaVA1_5-7B [35]+4.47+18.27+7.97+0.65+11.67+2.47+3.47+9.51+5.16
LLaVA1_5-13B [35]+4.3+7.31+5.45+0.76+7.03+2.11+4.22+11.51+6.37
InternVL2_5-2B [16]+5.44+19.48+8.57+2.2+48.29+8.47+3.24+40.45+8.65
InternVL2_5-4B [16]+3.53+22.17+7.14+1.62+45.19+7.56+2.2+55.05+10.85
InternVL2_5-8B [16]+5.38+2.13+6.93+3.53+28.53+7.02+3.27+26.54+6.44
Qwen2-VL-2B [51]+6.14+16.52+8.96+0.79+26.82+3.09+3.33+44.49+9.8
Qwen2-VL-7B [51]+5.62+29.4+11.01+1.29+22.54+3.4+3.53+46.94+11.15
Mean Improvements+4.98+16.47+8.01+1.55+27.15+4.88+3.32+33.5+8.35
", + "bbox": [ + 91, + 388, + 903, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Explainability improvements of TAM on diverse MLLMs and datasets. Specific results are listed in Table 6 of Supp. K.", + "bbox": [ + 116, + 523, + 879, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ence significantly improves Func-IoU by reducing context interference on function words, while the rank Gaussian filter performs better for object words. Importantly, these modules are mutually beneficial; their combination leads to an F1-IoU improvement of $8.96\\%$ , exceeding the total of their individual contributions (3.7% and 1.33%).", + "bbox": [ + 88, + 565, + 482, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further assess the effectiveness of these modules by comparing them with other methods, as shown in Table 2. Replacing our proposed rank Gaussian filter with alternatives yields lower performance, such as the adaptive median filter [8], the Gaussian filter, and the median filter. Although the differences are less pronounced than those seen with the causal inference module, denoising remains a challenging area. Our method outperforms feature surgery [32] (which operates on the class dimension) by $12.29\\%$ and also surpasses other ECI implementations in Supp. D.", + "bbox": [ + 88, + 657, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with SoTA Methods. Table 3 presents a comparison between TAM and SoTA methods. Our TAM outperforms all others in the major F1-IoU and Obj-IoU metrics. While some methods, such as Grad-Rollout [1] and Attention, achieve high Func-IoU by producing a limited number of activations, their performance on other metrics re", + "bbox": [ + 88, + 810, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mains low. The second-best method is Grad-CAM++ [9] on the COCO Caption [13] and OpenPSG [57] datasets, while CAM [56] and Grad-CAM [45] rank second on GranDf [42]. Our method surpasses these approaches by $8.96\\%$ , $2.82\\%$ , and $8.54\\%$ on these datasets, respectively. Furthermore, TAM is complementary to existing methods and enhances them significantly, with maximum gains of $6.62\\%$ , $44.42\\%$ , and $9.8\\%$ across the three metrics, respectively.", + "bbox": [ + 511, + 565, + 903, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Applicability and Scalability. We evaluate the applicability on three prominent MLLMs, as shown in Table 4, and assess its scalability across various model sizes. The results in Table 3 indicate that TAM consistently outperforms the baseline across all MLLMs and model sizes by significant margins. This demonstrates that TAM can be effectively applied to a range of MLLMs while providing superior explainability compared to existing methods. Moreover, TAM supports MLLMs comparison, complementing quantitative results. It introduces a new evaluation perspective by highlighting visual alignment quality, as illustrated in Fig. 8.", + "bbox": [ + 511, + 686, + 903, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 862, + 694, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison of Explainability Methods.", + "bbox": [ + 511, + 885, + 792, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "53", + "bbox": [ + 490, + 945, + 506, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/793b00d420c7a03590beea94a3c4f37f08f790a2f15101a48c6c8fa90eabcef2.jpg", + "image_caption": [ + "Figure 3. Visual comparison between TAM and SoTA methods on the COCO Caption dataset [13] using the Qwen2-VL-2B [51] model. The \"is\" in the last row is a functional word with background ground-truth. More examples are shown in Supp. G including complex cases." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/133398b6e279c2171db9f3b72958c2f2decb373681c29d9dc9479c7df46a451c.jpg", + "image_caption": [ + "Figure 4. TAM presents high-quality localization results on diverse datasets [13, 42, 57] under the Qwen2-VL-7B [51] model." + ], + "image_footnote": [], + "bbox": [ + 94, + 273, + 478, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform comparisons between the proposed TAM and SoTA explanation methods, as illustrated in Fig. 3. The visualization results clearly show that TAM more accurately highlights target objects and renders fewer false activations on function words. In contrast, attention-based methods [1, 2, 5] produce only a limited number of activations, while gradient-based methods [9, 45] present too many correlated responses and noise, performing significantly worse than our approach. Additionally, we present more complex examples in Supp. G, where TAM notably outperforms existing SoTA methods. TAM excels at explaining multiple tokens of MLLMs. In Supp. C, results show how TAM focuses on important tokens, whereas the baseline methods generate many redundant activations.", + "bbox": [ + 88, + 551, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "High-quality Localization Results. We visualized the results using the multimodal activation map in Eq. 3 (see detailed examples in Supp. B). The results demonstrate that TAM effectively locates objects across diverse datasets, as shown in Fig. 4. These findings indicate that TAM enhances MLLMs with localization capabilities, which could be potentially beneficial for various downstream tasks, including segmentation [32], object counting [46], anomaly detection [14], image editing [6], autonomous driving [22], and ap", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9944da5bff3f04ff6c48ab4097e93cc9aabd7648e6c5599cd74f2dae88d5aec.jpg", + "image_caption": [ + "Figure 5. TAM supports to explain and analyze diverse attributes of MLLM (Qwen2-VL-7B). See extensive examples in Supp. M." + ], + "image_footnote": [], + "bbox": [ + 517, + 273, + 903, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "plications in the medical field [21, 50], which require pixellevel activations.", + "bbox": [ + 511, + 561, + 903, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization for Various Attributes. We apply TAM to explain various attributes of MLLMs, specifically focusing on fine-grained token types as Fig. 5. The results indicate that TAM effectively supports the explanation of attributes such as color, shape, action, number, location, and text. Notably, the representation of numbers is less pronounced compared to object localization, suggesting that some attributes may be encoded within related object tokens. In contrast, other attributes are more prominently featured.", + "bbox": [ + 511, + 595, + 905, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Explaining Failure Cases. We apply TAM to analyze failure cases in the QK-VQA dataset [37] as Fig. 7. Our findings indicate that while the model successfully locates objects, it struggles to align them with specific knowledge. For instance, the model can identify the desert and focus on the key prompt (\"year\") when outputting \"don't know.\" It suggests a misalignment between the desert and the year, rather than recognition or prompt issues. Besides, some responses consist of synonyms, hyphenms, or hyponyms of the correct answers, resulting in mismatches. For more case studies, please refer to the catalog in Supp. Table 5.", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "54", + "bbox": [ + 490, + 945, + 506, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d01dc3e24f8cd68c76f5ddebf2c7295eb8854fd87679b8cbce63c576eb3ca3c.jpg", + "image_caption": [ + "Figure 6. TAM significantly improves the quality of video visualization, with much fewer redundant activation and noises on STAR [52] dataset using Qwen2-VL-2B [51]. Please see extensive examples in Supp. 27 and failure cases analysis for videos in Fig. 25." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 496, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a41589abb588a35c283540f42a091f30434ad52b580b37a4735ba4fd9b9f6ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 90, + 903, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fceab8ba0355ed65903104963968dc5153bae6d2e3ce952854f51147e2d0fb26.jpg", + "image_caption": [ + "Figure 7. TAM supports failure cases analysis on QK-VQA [37]. See more cases in Supp. about VQA (Fig. 24), videos (Fig. 25), visual reasoning (Fig. 28), and multi-turn conversation (Fig. 31)." + ], + "image_footnote": [], + "bbox": [ + 96, + 300, + 223, + 410 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/587602b470b3cd0bd90a00cda47eaad67cf026e08726ddf9ea733c09f803aa53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 223, + 300, + 351, + 410 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8aa8054cb80577901dddab1e8628e4d104752e90d969668eed0cd7bced221f34.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 300, + 477, + 410 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dbe630a86dc5d791fac0cae47d5996d220b75f4366ae3655691154b0df768c24.jpg", + "image_caption": [ + "Figure 8. TAM enables qualitative comparison among MLLMs. Extensive examples are given in Fig. 17 and Fig. 18 in Supp. K." + ], + "image_footnote": [], + "bbox": [ + 94, + 481, + 478, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video Visualization. The proposed TAM excels at video visualization, where there are much fewer redundant activations and noises as shown in Fig. 6. Our method provides very clear results, while the conventional CAM makes it hard to see the video content without obvious highlights.", + "bbox": [ + 89, + 746, + 482, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A Tool to Visually Comparison MLLMs. TAM can be used as a tool to compare MLLMs qualitatively. If an MLLM exhibits superior activation maps, it can be inferred that its explainability and degree of vision alignment are also better, offering valuable pixel-level predictions at no", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "additional cost. As shown in Fig. 8, Qwen2-VL-7B [51] shows greater explainability than LLaVA1_5-7B [35] and InternVL2_5-8B [16]. Furthermore, we observe that InternVL tends to focus more on text, resulting in weaker activation levels; in some cases, this performance is even better than that of LLaVA (e.g., with the broccoli). More extensive examples can be found in Supp. K.", + "bbox": [ + 511, + 301, + 906, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 422, + 633, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work focuses on the unique nature of MLLMs in visual explainability, which generates multiple tokens progressively, complicating the explanation process. In response, we introduced the Token Activation Map (TAM), a novel method that utilizes estimated causal inference to mitigate interferences from context tokens, with a rank Gaussian filter to reduce activation noise, thereby providing clearer visual explanations. Our results demonstrate that TAM significantly outperforms existing SoTA methods, offering high-quality visualizations applicable to diverse scenarios.", + "bbox": [ + 511, + 446, + 906, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite the success, we focus on visual inputs, other modalities, such as audio, remain under-explored. Additionally, interpreting model decisions is an extensible further aspect. The potential applications of our TAM method are extensive, including open-vocabulary segmentation, detection, image grounding, anomaly detection, remote sensing, and medical fields that require pixel-level activations. Our work lays a foundation for advancing their explainability and practical utility across various domains.", + "bbox": [ + 511, + 598, + 908, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 511, + 748, + 666, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by a research grant from the Joint Research Scheme (JRS) under the National Natural Science Foundation of China (NSFC) and the Research Grants Council (RGC) of Hong Kong (Project No. N_HKUST654/24), as well as a grant from the RGC of the Hong Kong Special Administrative Region, China (Project No. R6005-24).", + "bbox": [ + 511, + 773, + 906, + 878 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "55", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4190-4197, 2020. 2, 5, 6, 7, 14, 15", + "[2] Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. Attnlrp: Attention-aware layer-wise relevance propagation for transformers. In International Conference on Machine Learning, pages 135-168. PMLR, 2024. 2, 5, 6, 7, 14, 15", + "[3] Estelle Aflalo, Meng Du, Shao-Yen Tseng, Yongfei Liu, Chenfei Wu, Nan Duan, and Vasudev Lal. Vl-interpret: An interactive visualization tool for interpreting vision-language transformers. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 21406–21415, 2022. 2", + "[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1", + "[5] Ameen Ali, Thomas Schnake, Oliver Eberle, Grégoire Montavon, Klaus-Robert Müller, and Lior Wolf. Xai for transformers: Better explanations through conservative propagation. In International Conference on Machine Learning, pages 435-451. PMLR, 2022. 6, 7, 14, 15", + "[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 7", + "[7] Gabriela Ben Melech Stan, Estelle Aflalo, Raanan Yehezkel Rohekar, Anahita Bhiwandiwalla, Shao-Yen Tseng, Matthew Lyle Olson, Yaniv Gurwicz, Chenfei Wu, Nan Duan, and Vasudev Lal. Lvlm-intrepret: An interpretability tool for large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8182-8187, 2024. 1, 2, 3", + "[8] Chin-Chen Chang, Ju-Yuan Hsiao, and Chih-Ping Hsieh. An adaptive median filter for image denoising. In 2008 Second international symposium on intelligent information technology application, pages 346-350. IEEE, 2008. 3, 5, 6, 16", + "[9] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE winter conference on applications of computer vision (WACV), pages 839-847. IEEE, 2018. 2, 6, 7, 14, 15", + "[10] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 2", + "[11] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF conference on computer vision and pattern recognition, pages 782-791, 2021. 2, 5, 15", + "[12] Ruoyu Chen, Hua Zhang, Siyuan Liang, Jingzhi Li, and Xiaochun Cao. Less is more: Fewer interpretable region via submodular subset selection. In The Twelfth International Conference on Learning Representations. 3", + "[13] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 2, 5, 6, 7, 13, 16, 17, 20, 21, 22, 23, 24, 26, 29, 30", + "[14] Xuhai Chen, Jiangning Zhang, Guanzhong Tian, Haoyang He, Wuhao Zhang, Yabiao Wang, Chengjie Wang, and Yong Liu. Clip-ad: A language-guided staged dual-path model for zero-shot anomaly detection. In International Joint Conference on Artificial Intelligence, pages 17-33. Springer, 2024. 7", + "[15] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 2, 13", + "[16] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 1, 6, 8, 16, 17, 19, 25, 35", + "[17] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1", + "[18] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 6", + "[19] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations. 2, 3, 4", + "[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 2", + "[21] Xinpeng Ding, Yongqiang Chu, Renjie Pi, Hualiang Wang, and Xiaomeng Li. Hia: Towards chinese multimodal llms for comparative high-resolution joint diagnosis. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 575-586. Springer, 2024. 7", + "[22] Xinpeng Ding, Jianhua Han, Hang Xu, Xiaodan Liang, Wei Zhang, and Xiaomeng Li. Holistic autonomous driving understanding by bird's view injected multi-modal large models." + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "56", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13668-13677, 2024. 7", + "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2", + "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2", + "[25] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2", + "[26] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875-5888, 2021. 2, 6, 14", + "[27] Licheng Jiao, Yuhan Wang, Xu Liu, Lingling Li, Fang Liu, Wenping Ma, Yuwei Guo, Puhua Chen, Shuyuan Yang, and Biao Hou. Causal inference meets deep learning: A comprehensive survey. Research, 7:0467, 2024. 2, 3", + "[28] Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Unmasking clever hans predictors and assessing what machines really learn. Nature communications, 10(1):1096, 2019. 2", + "[29] Unggi Lee, Minji Jeon, Yunseo Lee, Gyuri Byun, Yoorim Son, Jaeyoon Shin, Hongkyu Ko, and Hyeoncheol Kim. Llava-docent: Instruction tuning with multimodal large language model to support art appreciation education. Computers and Education: Artificial Intelligence, 7:100297, 2024. 1", + "[30] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18061-18070, 2024. 1", + "[31] Yi Li, Hualiang Wang, Yiqun Duan, Hang Xu, and Xiaomeng Li. Exploring visual interpretability for contrastive language-image pre-training. arXiv preprint arXiv:2209.07046, 2022. 2", + "[32] Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. A closer look at the explainability of contrastive language-image pre-training. Pattern Recognition, page 111409, 2025. 2, 3, 4, 5, 6, 7, 14, 15, 22", + "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[34] Fenglin Liu, Tingting Zhu, Xian Wu, Bang Yang, Chenyu You, Chenyang Wang, Lei Lu, Zhangdaihong Liu, Yefeng" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zheng, Xu Sun, et al. A medical multimodal large language model for future pandemics. NPJ Digital Medicine, 6(1): 226, 2023. 1", + "[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3, 6, 8, 13, 16, 17, 25", + "[36] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017. 2, 3", + "[37] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 5, 7, 8, 14, 18, 31, 33", + "[38] Daniel Omeiza, Skyler Speakman, Celia Cintas, and Komminist Weldermariam. Smooth grad-cam++: An enhanced inference level visualization technique for deep convolutional neural network models. arXiv preprint arXiv:1908.01224, 2019. 1, 2, 3", + "[39] Vedant Palit, Rohan Pandey, Aryaman Arora, and Paul Pu Liang. Towards vision-language mechanistic interpretability: A causal tracing tool for blip. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2856-2861, 2023. 3", + "[40] Judea Pearl. Causality. Cambridge university press, 2009. 2,3", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 22", + "[42] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 5, 6, 7, 13, 16, 17", + "[43] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144, 2016. 2, 3", + "[44] Raanan Y Rohekar, Yaniv Gurwicz, and Shami Nisimov. Causal interpretation of self-attention in pre-trained transformers. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[45] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1, 2, 6, 7, 14, 15, 16, 17, 18, 20, 24, 30, 34", + "[46] Zenglin Shi, Ying Sun, and Mengmi Zhang. Training-free object counting with prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 323-331, 2024. 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "57", + "bbox": [ + 488, + 945, + 506, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1", + "[48] Joseph Tey. Understanding how vision-language models reason when solving visual math problems. 1, 2", + "[49] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3", + "[50] Lehan Wang, Haonan Wang, Honglong Yang, Jiaji Mao, Zehong Yang, Jun Shen, and Xiaomeng Li. Interpretable bilingual multimodal large language model for diverse biomedical tasks. In The Thirteenth International Conference on Learning Representations, 2025. 7", + "[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2, 5, 6, 7, 8, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38", + "[52] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2). 2, 5, 8, 14, 18, 29, 32, 34", + "[53] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. In The Twelfth International Conference on Learning Representations. 3", + "[54] Liuyi Yao, Zhixuan Chu, Sheng Li, Yaliang Li, Jing Gao, and Aidong Zhang. A survey on causal inference. ACM Transactions on Knowledge Discovery from Data (TKDD), 15(5):1-46, 2021. 3", + "[55] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multimodal large language models. Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, 2025. 1, 3", + "[56] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 6, 16, 17, 18, 20, 24, 30, 34", + "[57] Zijian Zhou, Zheng Zhu, Holger Caesar, and Miaojing Shi. Openpsg: Open-set panoptic scene graph generation via large multimodal models. In European Conference on Computer Vision, pages 199-215. Springer, 2024. 2, 5, 6, 7, 14, 16, 17" + ], + "bbox": [ + 91, + 92, + 480, + 808 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "58", + "bbox": [ + 490, + 946, + 506, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Token Activation Map to Visually Explain Multimodal LLMs/b852d151-66d5-47a7-8861-5fbe9cff3adc_model.json b/2025/Token Activation Map to Visually Explain Multimodal LLMs/b852d151-66d5-47a7-8861-5fbe9cff3adc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..45724e9463a91fd88becbba4b7e638e938ac2d89 --- /dev/null +++ b/2025/Token Activation Map to Visually Explain Multimodal LLMs/b852d151-66d5-47a7-8861-5fbe9cff3adc_model.json @@ -0,0 +1,2312 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.191, + 0.131, + 0.809, + 0.154 + ], + "angle": 0, + "content": "Token Activation Map to Visually Explain Multimodal LLMs" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.181, + 0.774, + 0.2 + ], + "angle": 0, + "content": "Yi Li, Hualiang Wang, Xinpeng Ding, Haonan Wang, Xiaomeng Li*" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.202, + 0.9, + 0.217 + ], + "angle": 0, + "content": "Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology, Hong Kong SAR, China." + }, + { + "type": "text", + "bbox": [ + 0.421, + 0.217, + 0.584, + 0.235 + ], + "angle": 0, + "content": "{ylini,xmli} \\(@\\) ust.hk" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.741 + ], + "angle": 0, + "content": "Multimodal large language models (MLLMs) are broadly empowering various fields. Despite their advancements, the explainability of MLLMs remains less explored, hindering deeper understanding, model credibility, and effective visualization. Unlike conventional vision models (e.g., CNNs, ViTs, CLIP) that produce a single output, MLLMs generate sequences of tokens progressively, where each generated token depends on the previous context. Therefore, earlier context tokens can introduce redundant activations that interfere with the explanation of later tokens beyond their original information. Existing studies often overlook this issue, but our observations reveal that these redundant correlations can significantly hurt the reliability of explanations. To address this, we propose an estimated causal inference method to mitigate the interference of context to achieve high-quality MLLM explanation, with a novel rank Gaussian filter to further reduce activation noises. We term this method Token Activation Map (TAM) to highlight the consideration of interactions between tokens. TAM also indicates that it excels at explaining multiple tokens of MLLM, which is different from the Class Activation Map (CAM) for a single prediction. Our TAM method significantly outperforms existing SoTA methods, showcasing high-quality visualization results that can be utilized for various scenarios, such as object localization, failure case analysis, video visualization, MLLMs visual comparison, and model understanding (e.g., color, shape, action, location, visual reasoning, multi-turn conversation, etc). The code is available at github.com/xmed-lab/TAM." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.772, + 0.222, + 0.788 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.798, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Multimodal large language models (MLLMs or Multimodal LLMs) increasingly empower wide applications [17, 29, 30, 34], enabling multimodal inputs (e.g., images, videos, text) and human-like conversations. Although extensive efforts have been devoted to enhancing the performance of" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.268, + 0.67, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.401, + 0.657, + 0.41 + ], + "angle": 0, + "content": "(a) Conventional models vs. MLLMs" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.269, + 0.803, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.401, + 0.785, + 0.41 + ], + "angle": 0, + "content": "(b) Information flow of MLLM" + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.269, + 0.907, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.401, + 0.902, + 0.41 + ], + "angle": 0, + "content": "(C) Correlated activations" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.415, + 0.648, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.499, + 0.731, + 0.507 + ], + "angle": 0, + "content": "(d) Qualitative example of interference activations" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.415, + 0.773, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.415, + 0.907, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.499, + 0.907, + 0.508 + ], + "angle": 0, + "content": "(e) TAM (ours) for MLLM explanation" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Figure 1. Illustration of Motivation. (a) MLLMs generate multiple tokens beyond a single output. (b) The information flow indicates that MLLM generates tokens progressively, where one generated token (each row) is correlated with the context (prompt + earlier answer tokens). (c) We randomly pair CAMs and count their L1 distance against text correlation. Higher text correlation corresponds to lower distance, indicating concurrent interferences. (d) The example shows that the context token \"plate\" introduces interference activations (marked by white boxes) to the later explained token \"fork\". (e) Our TAM reveals the original information of the explained token \"fork\", eliminating context interferences and beyond CAM for a single output in (d). Results in (c, d, e) derive from Qwen2-VL-2B on the COCO Caption dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.907, + 0.856 + ], + "angle": 0, + "content": "MLLMs [4, 16, 35, 47, 51], research on MLLM explainability remains less explored. It is crucial for user trust, model understanding, case analysis, and visualization. However, exploring the explainability of MLLMs is more challenging than conventional visual models (such as CNN [24], ViT [23], CLIP [41]), where MLLMs progressively generate multiple tokens beyond a single output as Fig. 1a. Thus, we aim to explain all tokens of MLLM, revealing the visual cues via token-level activation maps (e.g., Fig. 1e)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Current MLLM explanation works [7, 48, 55] are generally borrowed from techniques [38, 45] for conventional models. For example, LLaVA-CAM [55] uses Smooth" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.946, + 0.508, + 0.957 + ], + "angle": 0, + "content": "48" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.438 + ], + "angle": 0, + "content": "Grad-CAM++ [38] to explain LLaVA [35]. The self-attention weights [49] are applied to understand math problem-solving [48]. Besides, LVLM-Interpret [7] integrates some conventional methods as a tool. Unfortunately, existing visual explainability methods are designed for conventional models [23, 24, 41] that produce a single output, such as class activation maps (CAM) series [9, 26, 45, 56], attention-based approaches [1, 28], attention with relevance [2, 11], model-agnostic methods [36, 43], or multimodal explanation [3, 19, 31, 32]. However, these methods overlook that MLLMs generate multiple tokens progressively, where earlier context tokens interfere with later tokens to be explained. As shown in Fig. 1d, the conventional CAM [56] for the token \"fork\" is confounded by numerous redundant activations concurrent with the token \"plate\" in context (context refers to prompt and earlier answer tokens than the explained token). Quantitative analysis in Fig. 1c reveals that this issue is widespread, where correlated tokens lead to close and interference activations. These interferences obscure the original information conveyed by the explained token, significantly degrading the explainability of MLLMs. Notably, this activation interference across tokens is unique to MLLMs and has been overlooked." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.483, + 0.761 + ], + "angle": 0, + "content": "We aim to reveal the original information for the explained token while minimizing interference from context, as discussed above. Herein, the original information reveals a causal relation [40] between the current token and the input, while the interferences represent the correlations that need to be excluded. Inspired by the potential outcome model (POM) of causal inference [27], we propose the estimated causal inference to explore this causal relation. Unlike raw causal inference by comparing ready-made results (yes vs. no), there is no single given output to represent the \"no\" result. Therefore, we estimated interferences from multiple context tokens that are highly related to the target, with a scale factor optimized by the least-squares method to generate the refined causal activation map. We also propose a novel rank Gaussian filter to reduce noises, thereby further enhancing the quality of activation maps. The above methods are collectively referred to as token activation map (TAM), emphasizing the consideration of interactions between tokens. TAM also indicates that it excels at explaining tokens in multiple rounds of MLLM, which differs from the Class Activation Map (CAM) for a single prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Experimentally, the proposed TAM demonstrates quantitative improvement over existing SoTA explainability methods by more than \\(8.96\\%\\) on the COCO Caption dataset [13] and \\(8.54\\%\\) on the OpenPSG dataset [57] when applied to visually explain the Qwen2-VL-2B [51]. Meanwhile, it is complementary to existing methods to boost their performance. We further verify the scalability and applicability on 7 advanced MLLMs from the series of Qwen2-VL [51], InternVL2_5 [15] and LLaVA1_5 [35], where the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "performance improvements range from \\(5.45\\%\\) to \\(11.0\\%\\) on COCO Caption dataset [13]. Besides quantitative results, we conduct qualitative experiments to understand MLLM attributes about color, shape, action, location, etc. Furthermore, we illustrate that TAM is capable of explaining failure cases on the QK-VQA dataset. [37], and it supports video visualization well on the STAR dataset [52]. In addition, it can be used as a tool to visually compare MLLMs qualitatively. Moreover, TAM shows wide applicability to explain multi-turn conversations, multi-image scenarios, and visual reasoning (see the catalog Table 5 in supplementary for extensive examples). Our work has three main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.274, + 0.905, + 0.349 + ], + "angle": 0, + "content": "- We introduce TAM, a novel approach to explain MLLMs, which produces high-quality activation maps for multiple generative tokens by incorporating the estimated causal inference to alleviate interference from context tokens, with the rank Gaussian filter to reduce activation noises." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.35, + 0.905, + 0.38 + ], + "angle": 0, + "content": "TAM significantly outperforms current methods for MLLM explanation, while also complementing them." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.38, + 0.905, + 0.456 + ], + "angle": 0, + "content": "- TAM demonstrates wide applicability and scalability, serving as a versatile tool for object localization, model understanding, failure case analysis, video visualization, MLLM comparison, and supporting diverse scenarios (e.g., multi-turn, multi-image, visual reasoning)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.274, + 0.905, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.468, + 0.655, + 0.484 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.75 + ], + "angle": 0, + "content": "Visual Explainability for Multimodal LLM. The emergence of MLLMs such as LLaVA [35], Qwen2-VL [51], InternVL [15], and closed-source GPT-4o [25] has significantly impacted various tasks involving multimodal inputs like text, images, and video. We mainly focus on the visual explanation of MLLMs via the activation maps for the input image or video frames. Regarding the modality perspective, the explainability of MLLMs is partially related to methods to explain multimodal models such as CLIP Surgery [32] and transformer register [19] designed for CLIP [41]; InterpreT [3] and Bi-Modal [10] for BERT [20]. From the technical aspect, conventional explanation methods are also related, including class activation map series (CAM [56], Grad-CAM [45], Grad-CAM++ [9], LayerCAM [26]), attention-based mechanisms (LRP [28], Rollout [1]), the combination (AttnLRP [2], Grad×AttnRoll [11]), or model-agnostic methods (LIME [43], SHAP [36])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Although existing methods may be valid to explain MLLMs, they are not the optimal solution. Because they are usually designed for conventional models (e.g., CNN [24], ViT [23], CLIP [41]) that produce a single output rather than generating multiple tokens progressively in MLLMs. We have observed that these earlier context tokens interfere with later tokens by introducing redundant visual activations, as shown in Fig. 1. This phenomenon is first studied by us, solved by a novel estimated causal inference method, which is specially designed for MLLM. Notably," + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "49" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.468 + ], + "angle": 0, + "content": "some workshop papers (LLaVA-CAM [55] and LVLM-Interpret [7]) applied conventional methods [38, 44] to explain LLaVA [35], but they also overlooked the interference from context tokens and treated each token independently. Causal Model. Causal inference [54] seeks to establish causal relation [40] between variables, focusing on how changes in one variable influence another. In contrast to statistical correlation analysis, it emphasizes causal relations. Another related concept, causal intervention [27], involves manipulating variables using a do operator. Model-agnostic explainability methods, such as LIME [43] and SHAP [36], can be viewed as causal interventions that explore the relationship between inputs and structured outputs by masking inputs or selecting subsets [12]. This concept is also applied to transformers through token masking (CLEANN [44]) or hidden state replacement [39]. However, causal intervention for MLLMs is impractical. Because there are too many input tokens are outputs in MLLM, requiring unbearable inference times. Besides, modified input leads to a changed context every inference, which is hard to evaluate. Our method is inspired by the potential outcome model (POM) to achieve causal inference without additional model inferences. Our contribution lies in estimating the unprocessed output from multiple tokens rather than a single ready-made output, considering the uniqueness of MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.802 + ], + "angle": 0, + "content": "Transformer Denoising. Activation maps of transformers [49] usually present many noise activations, impeding high-quality visualization. Recent studies give their explanations about it, including attention sink [53], lack of register [19], and redundant features across classes [32]. In MLLM, these noises still exist obversely, even system tokens already play the role of registers, or deploying the feature surgery to mitigate redundant features (See Table 2). To solve this problem practically, we aim to introduce denoising filters to mitigate it as post-processing. Since these noises belong to salt-and-pepper noise morphologically, we apply the median filter, adaptive median filter [8], and the Gaussian filter. However, these filters are not the optimal solution, where the Gaussian filter keeps too much noise signal and the median filter overlooks smaller responses. Therefore, we propose the rank Gaussian filter, a novel, simple, and effective filter to denoise transformer activation. It merges ranked values within a sliding window, weighted summed by a 1-d Gaussian kernel with the center at the median rank. Besides, we improve the Gaussian kernel by the coefficient of variation for more robust results. Overall, our method addresses the activation noises in a practical aspect through a new filter." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.182, + 0.831 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we first provide an overview of the proposed Token Activation Map (TAM), as illustrated in Fig. 2. Next, we elaborate on two key modules within TAM: the estimated causal inference and the rank Gaussian filter. Finally," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "we introduce three metrics designed to enable fine-grained evaluations of the explanations for MLLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.13, + 0.722, + 0.146 + ], + "angle": 0, + "content": "3.1. Token Activation Map" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.907, + 0.363 + ], + "angle": 0, + "content": "TAM aims to depict explainable activation maps for multiple tokens generated from MLLMs by leveraging causal inference to mitigate inter-token interference. To be more specific, given the visual data and prompt text as the input, the MLLM is employed to progressively generate the visual features \\( \\pmb{F}^{v} \\in \\mathbb{R}^{n_{v} \\times c} \\), prompt features \\( \\pmb{F}^{p} \\in \\mathbb{R}^{n_{p} \\times c} \\) and answer features \\( \\pmb{F}^{a} \\in \\mathbb{R}^{n_{a} \\times c} \\). \\( n_{v}, n_{p}, n_{a} \\) and \\( c \\) indicate the number of three types of features, and the feature dimension, respectively. Next, a token classifier, instantiated as a fully connected layer, is used to generate answer tokens \\( \\pmb{t}^{a} = \\{t_{1}^{a} \\dots t_{n_{a}}^{a}\\} \\) from these features. Prompt tokens are defined as \\( \\pmb{t}^{p} = \\{t_{1}^{p} \\dots t_{n_{p}}^{p}\\} \\) similarly. Based on the variables described above, the visual activation maps for the prompt tokens and answer tokens are calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.37, + 0.805, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} _ {i} ^ {p} = \\left\\lfloor \\boldsymbol {F} ^ {v} \\boldsymbol {w} _ {t _ {i} ^ {p}} \\right\\rfloor_ {+}, i \\in [ 1, n _ {p} ]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.383, + 0.904, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} _ {i} ^ {a} = \\left\\lfloor \\boldsymbol {F} ^ {v} \\boldsymbol {w} _ {t _ {i} ^ {a}} \\right\\rfloor_ {+}, i \\in [ 1, n _ {a} ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.417, + 0.906, + 0.492 + ], + "angle": 0, + "content": "where \\( \\mathbf{A}_i^p \\in \\mathbb{R}^{n_v \\times 1} \\) and \\( \\mathbf{A}_i^a \\in \\mathbb{R}^{n_v \\times 1} \\) present the activation map for the \\( i \\)-th prompt token \\( t_i^p \\) and answer token \\( t_i^a \\). \\( \\mathbf{w}_{t_i^p} \\in \\mathbb{R}^{c \\times 1} \\) and \\( \\mathbf{w}_{t_i^a} \\in \\mathbb{R}^{c \\times 1} \\) are the corresponding weight vectors within the token classifier. \\( \\lfloor \\rfloor_+ \\) is the function to keep positive activations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.583 + ], + "angle": 0, + "content": "Then, the activation maps for prompt and answer tokens are concatenated \\((\\bot)\\) as \\(\\pmb{A}_{:n_p + n_a} = \\pmb{A}^p\\bot \\pmb{A}^a\\). For the \\(i\\)-th answer token, its context token is formulated as \\(\\pmb{A}_{:n_p + i - 1}\\), including all earlier context maps (prompt tokens + answer tokens earlier than \\(i\\)). On top of it, the refined activation maps for answer tokens are formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.592, + 0.905, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\bar {\\boldsymbol {A}} _ {i} ^ {a} = \\left\\lfloor \\mathcal {D} \\left(\\boldsymbol {A} _ {i} ^ {a} - s \\mathcal {E} \\left(\\boldsymbol {A} _ {: n _ {p} + i - 1}\\right)\\right) \\right\\rfloor_ {+}, i \\in [ 1, n _ {a} ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.906, + 0.711 + ], + "angle": 0, + "content": "where \\(\\bar{A}_i^a\\) indicates the refined activation map for the \\(i\\)-th answer token. The number of \\(i\\) increases sequentially from 1 to \\(n_a\\). \\(A_{i}^{a} - s\\mathcal{E}(A_{:n_{p} + i - 1})\\) is the proposed estimated causal inference that produces the causal activation map with the estimation function \\(\\mathcal{E}\\) described in Sec. 3.2. \\(\\mathcal{D}\\) is the proposed rank Gaussian filter module in Sec. 3.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.906, + 0.772 + ], + "angle": 0, + "content": "In the last, we concatenate \\((\\perp)\\) the refined visual activation map \\(\\bar{A}_i^a\\) of the \\(i\\)-th answer token with the raw textual relevance \\(\\mathbf{r}_i \\in \\mathbb{R}^{n_p + i - 1}\\) for a multimodal activation map (image or video with text) \\(M_i \\in \\mathbb{R}^{n_v + n_p + i - 1}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.781, + 0.905, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} _ {i} = \\mathcal {N} \\left(\\bar {\\boldsymbol {A}} _ {i} ^ {a} \\perp \\mathbf {r} _ {i}\\right), i \\in [ 1, n _ {a} ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.81, + 0.906, + 0.84 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}()\\) is a min-max normalization function for visualizing visual and textual activations at an aligned level." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.763, + 0.863 + ], + "angle": 0, + "content": "3.2. Estimated Causal Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Overall, Eq. 2 optimizes later visual activation maps based on visual activation maps of all earlier context tokens. To" + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "50" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.316, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.322, + 0.089, + 0.531, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.089, + 0.764, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.089, + 0.905, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.342, + 0.908, + 0.371 + ], + "angle": 0, + "content": "Figure 2. Illustration of method. (a, Eq. 1 - Eq. 3) The overall framework of TAM. (b, Eq. 4 - Eq. 5) Details of the estimated casual inference module. (d, Eq. 6 - Eq. 7) Details of the rank Gaussian filter module. (d, Eq. 8 - Eq. 10) Fine-grained evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.484, + 0.487 + ], + "angle": 0, + "content": "be more specific, causal inference is employed to investigate the correct causal relation between the current answer token and visual tokens, eliminating interferences from prompt tokens and preceding answer tokens. To this end, we quantify the interference map from the context tokens for the \\(i\\)-th answer token as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.495, + 0.483, + 0.538 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} \\left(\\boldsymbol {A} _ {: n _ {p} + i - 1}\\right) = \\sum_ {k = 1} ^ {n _ {p} + i - 1} \\frac {r _ {i} ^ {k}}{\\sum \\mathbf {r} _ {i} + \\epsilon} \\boldsymbol {A} _ {k} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.262, + 0.54, + 0.419, + 0.559 + ], + "angle": 0, + "content": "\\[\ns. t. r _ {i} ^ {k} = 0, \\mathrm {i f} t _ {i} ^ {a} = t _ {k},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.674 + ], + "angle": 0, + "content": "where \\(r_i^k\\) is the textual relevance between the \\(k\\)-th context token \\(t_k\\) and the current answer token \\(t_i^a\\). The relevance is obtained from activations among textual tokens \\(\\lfloor (\\pmb{F}^p \\perp \\pmb{F}^a) \\pmb{w}_{t_i^a} \\rfloor_+\\) following Eq. 1. If the current context token \\(t_k\\) is the same as the answer token \\(t_i^a\\), we set their relevance as 0, i.e., \\(r_i^k = 0\\), to avoid eliminating visual activations for tokens identical to the current answer token." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.812 + ], + "angle": 0, + "content": "Overall, Eq. 4 aims for each answer token to encapsulate more visual information, which implies minimizing similarity to the visual semantics of previously context tokens. Meanwhile, we use the scale factor \\( s \\) in Eq. 2 to control the scale of the interference map, ensuring that the influence of the interference map is removed from the activation map at a comparable degree. Specifically, we employ the least squares method to minimize the residual between the two maps, for the optimal value of \\( s \\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.821, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\ns = \\arg \\min _ {s} \\sum_ {j = 1} ^ {n _ {v}} \\left(\\boldsymbol {A} _ {j, i} ^ {a} - s \\mathcal {E} \\left(\\boldsymbol {A} _ {j,: n _ {p} + i - 1}\\right)\\right) ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "After processing by the estimated causal inference module, there are still many salt-and-pepper noises in activation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.907, + 0.427 + ], + "angle": 0, + "content": "maps. To address this issue, we will elaborate on the solution in the following section." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.436, + 0.715, + 0.451 + ], + "angle": 0, + "content": "3.3. Rank Gaussian Filter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.61 + ], + "angle": 0, + "content": "As described in Sec. 2 the activation noise exists obviously and belongs to the salt-and-pepper type. Different from existing methods [19, 32] with suboptimal performances on MLLM, we address this problem in an new aspect: image denoising. Compared with existing filters like the Gaussian filter that keeps too much noise signal and the median filter which overlooks smaller responses, our rank Gaussian filter is more robust beyond existing methods. The overall equation of the rank Gaussian filter for any activation map \\(A_{i,j} \\in \\mathbb{R}^{h \\times w}\\) is written as:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.62, + 0.905, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {D} \\left(\\boldsymbol {A} _ {i, j}, k\\right) = \\sum \\boldsymbol {A} _ {i, j} \\mathcal {G} \\left(\\boldsymbol {a} _ {i, j}, k\\right), \\forall i \\in H, \\forall j \\in W \\tag {6} \\\\ s. t. \\mathbf {a} _ {i, j} = r a n k (\\mathbf {A} _ {i: i + k, j: j + k}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.777 + ], + "angle": 0, + "content": "Herein, \\(A_{i:i + k,j:j + k}\\) indicates local values within a sliding window at kernel size \\(k\\). We sort them to get a ranked array to multiply with a custom 1-d Gaussian kernel \\(\\mathcal{G}(\\cdot)\\). This custom kernel allows the median to have the largest weight, meanwhile aggregating signals ranked close to the median. It can be regarded as a kind of smoothed median filter to enhance the robustness weighted by the Gaussian kernel." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.778, + 0.906, + 0.869 + ], + "angle": 0, + "content": "For the custom 1-d Gaussian kernel, we formulate it as Eq. 7. Specifically, \\((i - k^2 / / 2)^2\\) indicates the distance from the median rank \\((k^2 / / 2\\), size exact division 2), and \\(\\sigma (\\mathbf{a}) / \\mu (\\mathbf{a})\\) is the coefficient of variation which is more stable than the standard division \\(\\sigma ()\\) in general Gaussian kernel using the mean \\(\\mu ()\\) only." + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.878, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {i} (\\mathbf {a}, k) = e ^ {- \\frac {(i - k ^ {2} / / 2) ^ {2}}{2 (\\sigma (\\mathbf {a}) / \\mu (\\mathbf {a})) ^ {2}}}, \\forall i \\in k ^ {2} \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.507, + 0.957 + ], + "angle": 0, + "content": "51" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.4, + 0.108 + ], + "angle": 0, + "content": "3.4. Evaluation for MLLM Explanation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.113, + 0.483, + 0.265 + ], + "angle": 0, + "content": "Unlike conventional models with a single output, MLLMs generate multiple tokens. To evaluate how well activation maps correspond to objects, we propose two fine-grained IoU-based metrics: Obj-IoU for objects with manual masks, and Func-IoU for backgrounds. These metrics focus on plausibility instead of faithfulness (see concept differences in Supp. E). Plausibility better suits MLLMs since perturbation-based faithfulness tests [2, 11] which cause inconsistent text generation, making evaluation invalid (see detailed reasons in Supp. E)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.265, + 0.484, + 0.386 + ], + "angle": 0, + "content": "To be specific, we formulate Obj-IoU in Eq. 8, counting the IoU between activation map \\((A_{i})\\) after binarization \\((\\mathcal{B}()\\) using opencv OTSU auto threshold) and the groundtruth mask \\(G_{i}\\). Since the object classes may be unfixed, we count the IoU for all tokens at size \\(o\\) over the dataset without an average on classes. We obtain the mask by matching the token to mask names. For a word or phrase containing multiple tokens, we record their max IoU for a single object." + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.396, + 0.483, + 0.429 + ], + "angle": 0, + "content": "\\[\n\\operatorname {O b j} - \\operatorname {I o U} = \\frac {1}{o} \\sum_ {i = 1} ^ {o} \\frac {\\mathcal {B} (\\boldsymbol {A} _ {i}) \\cap \\boldsymbol {G} _ {i}}{\\mathcal {B} (\\boldsymbol {A} _ {i}) \\cup \\boldsymbol {G} _ {i}} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Func-IoU measures the activation degree of tokens that are unreadable from the image (e.g., \"is\", \"and\", \"so\", \"the\"). Higher Func-IoU indicates fewer false positives and unnecessary responses. Specifically, we record the mean binarization thresholds \\( b_{i} \\) (opencv OTSU auto threshold) for all noun tokens in the same image to divide foreground and background on function words. Activation map lower than it \\( (A_{i} < b_{i}) \\) is the background prediction used to count IoU with the all-one matrix \\( J \\) (ground-truth is all background class). It is applied to total \\( u \\) tokens over the whole dataset." + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.592, + 0.483, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\operatorname {F u n c - I o U} = \\frac {1}{u} \\sum_ {i = 1} ^ {u} \\frac {\\left(\\boldsymbol {A} _ {i} < b _ {i}\\right) \\cap \\boldsymbol {J}}{\\left(\\boldsymbol {A} _ {i} < b _ {i}\\right) \\cup \\boldsymbol {J}} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.763 + ], + "angle": 0, + "content": "We try to merge the above IoUs as the overall major metrics. We find their average is insufficient to measure biased and low-quality activation maps. For example, if one method predicts almost all tokens to be background, the Obj-IoU is very low but Func-IoU is high, forming a certain inverse ratio (e.g., \\(5.74\\%\\) vs. \\(96.5\\%\\) of Attention-Rollout [1] in Table 3). One suitable solution is to use the F1 value to merge metrics with a certain inverse ratio. Thus, we have the F1-IoU to evaluate the overall plausibility stably:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.77, + 0.483, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F 1 - I o U} = \\frac {2 * \\text {O b j - I o U} * \\text {F u n c - I o U}}{\\text {O b j - I o U} + \\text {F u n c - I o U}} \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.809, + 0.224, + 0.827 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.173, + 0.851 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Datasets. For quantitative experiments, we use datasets with texts and pixel-level annotations. Our main dataset is the COCO Caption dataset [13], which draws images and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "masks from COCO2014 [33]. Since explainability methods require testing only, we use its 5K-image minival split without training. Additional datasets include GranDf [42] (1K images) and OpenPSG [57] (3,176-image validation set). Masks in COCO Caption and GranDf are manually annotated, while OpenPSG masks are integrated by Rasheed et al. [42]. For datasets without masks, we perform qualitative visualization tests (Sec. 4.3), including attribute-only images and the QK-VQA [37] validation set (5,046 VQA samples with textual answers). We also qualitatively evaluate the STAR video dataset [52] (914 videos)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Implementations. We use the SciPy minimize function with the BFGS method to optimize the scale factor \\( s \\) in Eq. 5. The only hyperparameter, the kernel size \\( k \\) in Eq. 7, is set to 3, as our observations indicate that salt-and-pepper noises typically cluster between 1 to 5 tokens. A kernel size of 3, consisting of 9 elements, is deemed suitable. Eq. 9 incorporates part-of-speech identification. For Func-IoU and Obj-IoU, we tag each word's part of speech using the pos_tag function from the NLTK Python package (version 3.8.1, see tag details in Supp. E). For word matching to mask names, we employ the lemmatize function from nltk.stem. Note that the thresholds in Eq. 8 and Eq. 9 are automatically derived from OpenCV's OTSU method, requiring no manual operation. Details of the baselines are provided in Supp. D." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.479, + 0.708, + 0.495 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.505, + 0.902, + 0.57 + ], + "angle": 0, + "content": "
ECIRGFObj-IoU (%)Func-IoU (%)F1-IoU (%)
XX21.2351.9330.14
✓X22.4169.0333.84+3.7
X✓24.8243.3431.57+1.33
✓✓27.3768.4439.1+8.96
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.573, + 0.907, + 0.657 + ], + "angle": 0, + "content": "Table 1. Ablation study on the COCO Caption [13] dataset using Qwen2-VL-2B [51]. These two modules are mutually beneficial, the combination exceeds the sum of their individual gains. ECI indicates the proposed estimated causal inference, and RGF is the rank Gaussian filter. Metrics are IoU for object words, IoU for function words, and their F1-score-like combination, respectively." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.667, + 0.91, + 0.804 + ], + "angle": 0, + "content": "
SettingObj-IoU (%)Func-IoU (%)F1-IoU (%)
Baseline21.2351.9330.14
Replace rank Gaussian filter
Adaptive Median [8]25.4868.237.1
Median Filter26.0168.2637.67
Gaussian Filter26.5667.7838.16
Replace estimated causal inference
Feature surgery [32]18.548.6626.81
ECI-mean27.8449.8535.72
ECI-attnWeights27.1154.6136.23
Ours27.3768.4439.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.807, + 0.907, + 0.85 + ], + "angle": 0, + "content": "Table 2. Effectiveness study on the COCO Caption [13] dataset using Qwen2-VL-2B [51]. ECI-mean and ECI-attnWeights are candidate implementations of ECI described in Supp. D." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Ablation and Effectiveness. We present the ablation study results in Table 1, which indicates that each module of the TAM is valuable. Specifically, the estimated causal infer" + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.957 + ], + "angle": 0, + "content": "52" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.9, + 0.296 + ], + "angle": 0, + "content": "
MethodTypeCOCO CaptionGranDfOpenPSG
Obj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoU
Grad-CAM [45]21.2351.9330.1417.8562.1527.7422.9348.5731.15
Grad-CAM++ [9]Gradient19.5262.8329.7817.373.4228.0122.2159.9532.41
Grad-Rollout [1]1.2799.512.511.499.612.771.5799.5813.08
Layer-CAM [26]11.4384.8820.1513.1182.0922.6214.1285.2924.22
AttentionAttention8.292.8715.079.693.5617.4210.5894.2819.03
Attention-Rollout [1]5.7496.510.837.2196.6513.427.9497.0414.68
CP-LRP [5]Combination9.953.9716.7312.6153.2420.3913.353.3621.3
Attn-LRP [2]9.9252.4116.6912.1552.1919.7212.7852.2620.54
CAM [56]Logit21.2351.9330.1417.8562.1527.7422.9348.5731.15
Archi.-Surgery [32]15.6963.8225.1916.5962.2826.219.8358.7729.65
TAM (ours)27.3768.4439.118.6588.9730.8326.2692.9940.95
Archi.-Surgery [32]+ Ours+4.13+10.68+6.12+1.41+26.48+3.74+1.59+31.92+5.0
Grad-CAM++ [9]+2.83+3.91+3.7+1.26+5.39+2.04+2.75+9.44+4.3
Layer-CAM [26]+5.04-7.95+6.98+2.79+0.52+4.05+6.62-2.49+8.95
CAM [56]+6.14+16.52+8.96+0.79+26.82+3.09+3.33+44.42+9.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.307, + 0.907, + 0.379 + ], + "angle": 0, + "content": "Table 3. Comparison with SoTA methods using Qwen2-VL-2B on diverse datasets. We adopt the \"Logit\" type that uses the classifier weights in Eq. 1, without back-propagation in \"Gradient\" and \"Combination\" (Gradient + Attention). Besides, we support FlashAttention [18] and SdpaAttention, which do not return attention weights, while \"Attention\" and \"Combination\" rely on it. Our TAM is also complementary to existing methods with gains marked by \"+\" Note, CAM and Grad-CAM are equivalent as discussed in Supp. D. The major metric is the F1-IoU \\((\\%)\\) to reflect the overall result, merged from IoUs for Object words (Obj-IoU) and function words (Func-IoU)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.39, + 0.905, + 0.515 + ], + "angle": 0, + "content": "
CAM [56] +OursCOCO CaptionGranDfOpenPSG
Obj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoU
LLaVA1_5-7B [35]+4.47+18.27+7.97+0.65+11.67+2.47+3.47+9.51+5.16
LLaVA1_5-13B [35]+4.3+7.31+5.45+0.76+7.03+2.11+4.22+11.51+6.37
InternVL2_5-2B [16]+5.44+19.48+8.57+2.2+48.29+8.47+3.24+40.45+8.65
InternVL2_5-4B [16]+3.53+22.17+7.14+1.62+45.19+7.56+2.2+55.05+10.85
InternVL2_5-8B [16]+5.38+2.13+6.93+3.53+28.53+7.02+3.27+26.54+6.44
Qwen2-VL-2B [51]+6.14+16.52+8.96+0.79+26.82+3.09+3.33+44.49+9.8
Qwen2-VL-7B [51]+5.62+29.4+11.01+1.29+22.54+3.4+3.53+46.94+11.15
Mean Improvements+4.98+16.47+8.01+1.55+27.15+4.88+3.32+33.5+8.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.117, + 0.525, + 0.88, + 0.54 + ], + "angle": 0, + "content": "Table 4. Explainability improvements of TAM on diverse MLLMs and datasets. Specific results are listed in Table 6 of Supp. K." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.566, + 0.483, + 0.656 + ], + "angle": 0, + "content": "ence significantly improves Func-IoU by reducing context interference on function words, while the rank Gaussian filter performs better for object words. Importantly, these modules are mutually beneficial; their combination leads to an F1-IoU improvement of \\(8.96\\%\\), exceeding the total of their individual contributions (3.7% and 1.33%)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.658, + 0.483, + 0.809 + ], + "angle": 0, + "content": "We further assess the effectiveness of these modules by comparing them with other methods, as shown in Table 2. Replacing our proposed rank Gaussian filter with alternatives yields lower performance, such as the adaptive median filter [8], the Gaussian filter, and the median filter. Although the differences are less pronounced than those seen with the causal inference module, denoising remains a challenging area. Our method outperforms feature surgery [32] (which operates on the class dimension) by \\(12.29\\%\\) and also surpasses other ECI implementations in Supp. D." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Comparison with SoTA Methods. Table 3 presents a comparison between TAM and SoTA methods. Our TAM outperforms all others in the major F1-IoU and Obj-IoU metrics. While some methods, such as Grad-Rollout [1] and Attention, achieve high Func-IoU by producing a limited number of activations, their performance on other metrics re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.905, + 0.686 + ], + "angle": 0, + "content": "mains low. The second-best method is Grad-CAM++ [9] on the COCO Caption [13] and OpenPSG [57] datasets, while CAM [56] and Grad-CAM [45] rank second on GranDf [42]. Our method surpasses these approaches by \\(8.96\\%\\), \\(2.82\\%\\), and \\(8.54\\%\\) on these datasets, respectively. Furthermore, TAM is complementary to existing methods and enhances them significantly, with maximum gains of \\(6.62\\%\\), \\(44.42\\%\\), and \\(9.8\\%\\) across the three metrics, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.905, + 0.853 + ], + "angle": 0, + "content": "Applicability and Scalability. We evaluate the applicability on three prominent MLLMs, as shown in Table 4, and assess its scalability across various model sizes. The results in Table 3 indicate that TAM consistently outperforms the baseline across all MLLMs and model sizes by significant margins. This demonstrates that TAM can be effectively applied to a range of MLLMs while providing superior explainability compared to existing methods. Moreover, TAM supports MLLMs comparison, complementing quantitative results. It introduces a new evaluation perspective by highlighting visual alignment quality, as illustrated in Fig. 8." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.863, + 0.696, + 0.879 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.794, + 0.901 + ], + "angle": 0, + "content": "Comparison of Explainability Methods." + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.946, + 0.507, + 0.956 + ], + "angle": 0, + "content": "53" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.905, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.222, + 0.905, + 0.251 + ], + "angle": 0, + "content": "Figure 3. Visual comparison between TAM and SoTA methods on the COCO Caption dataset [13] using the Qwen2-VL-2B [51] model. The \"is\" in the last row is a functional word with background ground-truth. More examples are shown in Supp. G including complex cases." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.275, + 0.48, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.495, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Figure 4. TAM presents high-quality localization results on diverse datasets [13, 42, 57] under the Qwen2-VL-7B [51] model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.764 + ], + "angle": 0, + "content": "We perform comparisons between the proposed TAM and SoTA explanation methods, as illustrated in Fig. 3. The visualization results clearly show that TAM more accurately highlights target objects and renders fewer false activations on function words. In contrast, attention-based methods [1, 2, 5] produce only a limited number of activations, while gradient-based methods [9, 45] present too many correlated responses and noise, performing significantly worse than our approach. Additionally, we present more complex examples in Supp. G, where TAM notably outperforms existing SoTA methods. TAM excels at explaining multiple tokens of MLLMs. In Supp. C, results show how TAM focuses on important tokens, whereas the baseline methods generate many redundant activations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "High-quality Localization Results. We visualized the results using the multimodal activation map in Eq. 3 (see detailed examples in Supp. B). The results demonstrate that TAM effectively locates objects across diverse datasets, as shown in Fig. 4. These findings indicate that TAM enhances MLLMs with localization capabilities, which could be potentially beneficial for various downstream tasks, including segmentation [32], object counting [46], anomaly detection [14], image editing [6], autonomous driving [22], and ap" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.274, + 0.905, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.5, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Figure 5. TAM supports to explain and analyze diverse attributes of MLLM (Qwen2-VL-7B). See extensive examples in Supp. M." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.905, + 0.592 + ], + "angle": 0, + "content": "plications in the medical field [21, 50], which require pixellevel activations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.732 + ], + "angle": 0, + "content": "Visualization for Various Attributes. We apply TAM to explain various attributes of MLLMs, specifically focusing on fine-grained token types as Fig. 5. The results indicate that TAM effectively supports the explanation of attributes such as color, shape, action, number, location, and text. Notably, the representation of numbers is less pronounced compared to object localization, suggesting that some attributes may be encoded within related object tokens. In contrast, other attributes are more prominently featured." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Explaining Failure Cases. We apply TAM to analyze failure cases in the QK-VQA dataset [37] as Fig. 7. Our findings indicate that while the model successfully locates objects, it struggles to align them with specific knowledge. For instance, the model can identify the desert and focus on the key prompt (\"year\") when outputting \"don't know.\" It suggests a misalignment between the desert and the year, rather than recognition or prompt issues. Besides, some responses consist of synonyms, hyphenms, or hyponyms of the correct answers, resulting in mismatches. For more case studies, please refer to the catalog in Supp. Table 5." + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "54" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.498, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.091, + 0.904, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.249, + 0.908, + 0.278 + ], + "angle": 0, + "content": "Figure 6. TAM significantly improves the quality of video visualization, with much fewer redundant activation and noises on STAR [52] dataset using Qwen2-VL-2B [51]. Please see extensive examples in Supp. 27 and failure cases analysis for videos in Fig. 25." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.301, + 0.225, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.301, + 0.352, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.301, + 0.478, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.421, + 0.483, + 0.464 + ], + "angle": 0, + "content": "Figure 7. TAM supports failure cases analysis on QK-VQA [37]. See more cases in Supp. about VQA (Fig. 24), videos (Fig. 25), visual reasoning (Fig. 28), and multi-turn conversation (Fig. 31)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.482, + 0.48, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Figure 8. TAM enables qualitative comparison among MLLMs. Extensive examples are given in Fig. 17 and Fig. 18 in Supp. K." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.747, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Video Visualization. The proposed TAM excels at video visualization, where there are much fewer redundant activations and noises as shown in Fig. 6. Our method provides very clear results, while the conventional CAM makes it hard to see the video content without obvious highlights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "A Tool to Visually Comparison MLLMs. TAM can be used as a tool to compare MLLMs qualitatively. If an MLLM exhibits superior activation maps, it can be inferred that its explainability and degree of vision alignment are also better, offering valuable pixel-level predictions at no" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.907, + 0.41 + ], + "angle": 0, + "content": "additional cost. As shown in Fig. 8, Qwen2-VL-7B [51] shows greater explainability than LLaVA1_5-7B [35] and InternVL2_5-8B [16]. Furthermore, we observe that InternVL tends to focus more on text, resulting in weaker activation levels; in some cases, this performance is even better than that of LLaVA (e.g., with the broccoli). More extensive examples can be found in Supp. K." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.423, + 0.634, + 0.438 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.599 + ], + "angle": 0, + "content": "This work focuses on the unique nature of MLLMs in visual explainability, which generates multiple tokens progressively, complicating the explanation process. In response, we introduced the Token Activation Map (TAM), a novel method that utilizes estimated causal inference to mitigate interferences from context tokens, with a rank Gaussian filter to reduce activation noise, thereby providing clearer visual explanations. Our results demonstrate that TAM significantly outperforms existing SoTA methods, offering high-quality visualizations applicable to diverse scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.736 + ], + "angle": 0, + "content": "Despite the success, we focus on visual inputs, other modalities, such as audio, remain under-explored. Additionally, interpreting model decisions is an extensible further aspect. The potential applications of our TAM method are extensive, including open-vocabulary segmentation, detection, image grounding, anomaly detection, remote sensing, and medical fields that require pixel-level activations. Our work lays a foundation for advancing their explainability and practical utility across various domains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.749, + 0.667, + 0.766 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.774, + 0.907, + 0.88 + ], + "angle": 0, + "content": "This work was supported by a research grant from the Joint Research Scheme (JRS) under the National Natural Science Foundation of China (NSFC) and the Research Grants Council (RGC) of Hong Kong (Project No. N_HKUST654/24), as well as a grant from the RGC of the Hong Kong Special Administrative Region, China (Project No. R6005-24)." + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.957 + ], + "angle": 0, + "content": "55" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4190-4197, 2020. 2, 5, 6, 7, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.256 + ], + "angle": 0, + "content": "[2] Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. Attnlrp: Attention-aware layer-wise relevance propagation for transformers. In International Conference on Machine Learning, pages 135-168. PMLR, 2024. 2, 5, 6, 7, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[3] Estelle Aflalo, Meng Du, Shao-Yen Tseng, Yongfei Liu, Chenfei Wu, Nan Duan, and Vasudev Lal. Vl-interpret: An interactive visualization tool for interpreting vision-language transformers. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 21406–21415, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.429, + 0.483, + 0.499 + ], + "angle": 0, + "content": "[5] Ameen Ali, Thomas Schnake, Oliver Eberle, Grégoire Montavon, Klaus-Robert Müller, and Lior Wolf. Xai for transformers: Better explanations through conservative propagation. In International Conference on Machine Learning, pages 435-451. PMLR, 2022. 6, 7, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.501, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.484, + 0.656 + ], + "angle": 0, + "content": "[7] Gabriela Ben Melech Stan, Estelle Aflalo, Raanan Yehezkel Rohekar, Anahita Bhiwandiwalla, Shao-Yen Tseng, Matthew Lyle Olson, Yaniv Gurwicz, Chenfei Wu, Nan Duan, and Vasudev Lal. Lvlm-intrepret: An interpretability tool for large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8182-8187, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.483, + 0.714 + ], + "angle": 0, + "content": "[8] Chin-Chen Chang, Ju-Yuan Hsiao, and Chih-Ping Hsieh. An adaptive median filter for image denoising. In 2008 Second international symposium on intelligent information technology application, pages 346-350. IEEE, 2008. 3, 5, 6, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.716, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[9] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE winter conference on applications of computer vision (WACV), pages 839-847. IEEE, 2018. 2, 6, 7, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[10] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 782-791, 2021. 2, 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[12] Ruoyu Chen, Hua Zhang, Siyuan Liang, Jingzhi Li, and Xiaochun Cao. Less is more: Fewer interpretable region via submodular subset selection. In The Twelfth International Conference on Learning Representations. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[13] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 2, 5, 6, 7, 13, 16, 17, 20, 21, 22, 23, 24, 26, 29, 30" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[14] Xuhai Chen, Jiangning Zhang, Guanzhong Tian, Haoyang He, Wuhao Zhang, Yabiao Wang, Chengjie Wang, and Yong Liu. Clip-ad: A language-guided staged dual-path model for zero-shot anomaly detection. In International Joint Conference on Artificial Intelligence, pages 17-33. Springer, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[15] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[16] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 1, 6, 8, 16, 17, 19, 25, 35" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.497, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[17] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[18] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[19] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[21] Xinpeng Ding, Yongqiang Chu, Renjie Pi, Hualiang Wang, and Xiaomeng Li. Hia: Towards chinese multimodal llms for comparative high-resolution joint diagnosis. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 575-586. Springer, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[22] Xinpeng Ding, Jianhua Han, Hang Xu, Xiaodan Liang, Wei Zhang, and Xiaomeng Li. Holistic autonomous driving understanding by bird's view injected multi-modal large models." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "56" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13668-13677, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[25] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[26] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875-5888, 2021. 2, 6, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[27] Licheng Jiao, Yuhan Wang, Xu Liu, Lingling Li, Fang Liu, Wenping Ma, Yuwei Guo, Puhua Chen, Shuyuan Yang, and Biao Hou. Causal inference meets deep learning: A comprehensive survey. Research, 7:0467, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[28] Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Unmasking clever hans predictors and assessing what machines really learn. Nature communications, 10(1):1096, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[29] Unggi Lee, Minji Jeon, Yunseo Lee, Gyuri Byun, Yoorim Son, Jaeyoon Shin, Hongkyu Ko, and Hyeoncheol Kim. Llava-docent: Instruction tuning with multimodal large language model to support art appreciation education. Computers and Education: Artificial Intelligence, 7:100297, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[30] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18061-18070, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[31] Yi Li, Hualiang Wang, Yiqun Duan, Hang Xu, and Xiaomeng Li. Exploring visual interpretability for contrastive language-image pre-training. arXiv preprint arXiv:2209.07046, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[32] Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. A closer look at the explainability of contrastive language-image pre-training. Pattern Recognition, page 111409, 2025. 2, 3, 4, 5, 6, 7, 14, 15, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[34] Fenglin Liu, Tingting Zhu, Xian Wu, Bang Yang, Chenyu You, Chenyang Wang, Lei Lu, Zhangdaihong Liu, Yefeng" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Zheng, Xu Sun, et al. A medical multimodal large language model for future pandemics. NPJ Digital Medicine, 6(1): 226, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3, 6, 8, 13, 16, 17, 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[36] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[37] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 5, 7, 8, 14, 18, 31, 33" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[38] Daniel Omeiza, Skyler Speakman, Celia Cintas, and Komminist Weldermariam. Smooth grad-cam++: An enhanced inference level visualization technique for deep convolutional neural network models. arXiv preprint arXiv:1908.01224, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[39] Vedant Palit, Rohan Pandey, Aryaman Arora, and Paul Pu Liang. Towards vision-language mechanistic interpretability: A causal tracing tool for blip. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2856-2861, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[40] Judea Pearl. Causality. Cambridge university press, 2009. 2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[42] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 5, 6, 7, 13, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[43] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[44] Raanan Y Rohekar, Yaniv Gurwicz, and Shami Nisimov. Causal interpretation of self-attention in pre-trained transformers. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[45] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1, 2, 6, 7, 14, 15, 16, 17, 18, 20, 24, 30, 34" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Zenglin Shi, Ying Sun, and Mengmi Zhang. Training-free object counting with prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 323-331, 2024. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.49, + 0.946, + 0.508, + 0.956 + ], + "angle": 0, + "content": "57" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[47] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[48] Joseph Tey. Understanding how vision-language models reason when solving visual math problems. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[49] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[50] Lehan Wang, Haonan Wang, Honglong Yang, Jiaji Mao, Zehong Yang, Jun Shen, and Xiaomeng Li. Interpretable bilingual multimodal large language model for diverse biomedical tasks. In The Thirteenth International Conference on Learning Representations, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2, 5, 6, 7, 8, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[52] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2). 2, 5, 8, 14, 18, 29, 32, 34" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[53] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. In The Twelfth International Conference on Learning Representations. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[54] Liuyi Yao, Zhixuan Chu, Sheng Li, Yaliang Li, Jing Gao, and Aidong Zhang. A survey on causal inference. ACM Transactions on Knowledge Discovery from Data (TKDD), 15(5):1-46, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[55] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multimodal large language models. Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.672, + 0.482, + 0.739 + ], + "angle": 0, + "content": "[56] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 6, 16, 17, 18, 20, 24, 30, 34" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.742, + 0.482, + 0.809 + ], + "angle": 0, + "content": "[57] Zijian Zhou, Zheng Zhu, Holger Caesar, and Miaojing Shi. Openpsg: Open-set panoptic scene graph generation via large multimodal models. In European Conference on Computer Vision, pages 199-215. Springer, 2024. 2, 5, 6, 7, 14, 16, 17" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.809 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.491, + 0.947, + 0.508, + 0.956 + ], + "angle": 0, + "content": "58" + } + ] +] \ No newline at end of file diff --git a/2025/Token Activation Map to Visually Explain Multimodal LLMs/b852d151-66d5-47a7-8861-5fbe9cff3adc_origin.pdf b/2025/Token Activation Map to Visually Explain Multimodal LLMs/b852d151-66d5-47a7-8861-5fbe9cff3adc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ce95e266ce0810223eae0aa47054ddff2e620f41 --- /dev/null +++ b/2025/Token Activation Map to Visually Explain Multimodal LLMs/b852d151-66d5-47a7-8861-5fbe9cff3adc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:03da8a0fac3a79ea047489c3aafc2869a05a8ddfcf6cda07667142eeba86c2ec +size 4190729 diff --git a/2025/Token Activation Map to Visually Explain Multimodal LLMs/full.md b/2025/Token Activation Map to Visually Explain Multimodal LLMs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a79e53d1c31430a9195d1d598772a330fd5c0584 --- /dev/null +++ b/2025/Token Activation Map to Visually Explain Multimodal LLMs/full.md @@ -0,0 +1,329 @@ +# Token Activation Map to Visually Explain Multimodal LLMs + +Yi Li, Hualiang Wang, Xinpeng Ding, Haonan Wang, Xiaomeng Li* + +Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology, Hong Kong SAR, China. + +{ylini,xmli} $@$ ust.hk + +# Abstract + +Multimodal large language models (MLLMs) are broadly empowering various fields. Despite their advancements, the explainability of MLLMs remains less explored, hindering deeper understanding, model credibility, and effective visualization. Unlike conventional vision models (e.g., CNNs, ViTs, CLIP) that produce a single output, MLLMs generate sequences of tokens progressively, where each generated token depends on the previous context. Therefore, earlier context tokens can introduce redundant activations that interfere with the explanation of later tokens beyond their original information. Existing studies often overlook this issue, but our observations reveal that these redundant correlations can significantly hurt the reliability of explanations. To address this, we propose an estimated causal inference method to mitigate the interference of context to achieve high-quality MLLM explanation, with a novel rank Gaussian filter to further reduce activation noises. We term this method Token Activation Map (TAM) to highlight the consideration of interactions between tokens. TAM also indicates that it excels at explaining multiple tokens of MLLM, which is different from the Class Activation Map (CAM) for a single prediction. Our TAM method significantly outperforms existing SoTA methods, showcasing high-quality visualization results that can be utilized for various scenarios, such as object localization, failure case analysis, video visualization, MLLMs visual comparison, and model understanding (e.g., color, shape, action, location, visual reasoning, multi-turn conversation, etc). The code is available at github.com/xmed-lab/TAM. + +# 1. Introduction + +Multimodal large language models (MLLMs or Multimodal LLMs) increasingly empower wide applications [17, 29, 30, 34], enabling multimodal inputs (e.g., images, videos, text) and human-like conversations. Although extensive efforts have been devoted to enhancing the performance of + +![](images/2a3412272f6e2129275a35c47ac409a77f84afd6c81c3ea480de827eafa11480.jpg) +(a) Conventional models vs. MLLMs + +![](images/f6f335c94f06f1a55049345c3fa1d7115795ab805e3cbcb67385e68d60bbc04c.jpg) +(b) Information flow of MLLM + +![](images/bccf1e88009a26b6bebdf77574a705286367ea6675bfa676d7123bb27ac6c61f.jpg) +(C) Correlated activations + +![](images/3859da4497307882cd97f8f1d183b03fea829e0040d55fb37c4fbdcffb775218.jpg) +(d) Qualitative example of interference activations + +![](images/85dc2c305cb925a43f0289596bfe6c312c26f97e3bbbf8d8f9f906a48f352a7d.jpg) +(e) TAM (ours) for MLLM explanation + +![](images/57d544af825f0f08bf5f11b60798d366f21eea240fe56dd630a5c65cc905e931.jpg) +Figure 1. Illustration of Motivation. (a) MLLMs generate multiple tokens beyond a single output. (b) The information flow indicates that MLLM generates tokens progressively, where one generated token (each row) is correlated with the context (prompt + earlier answer tokens). (c) We randomly pair CAMs and count their L1 distance against text correlation. Higher text correlation corresponds to lower distance, indicating concurrent interferences. (d) The example shows that the context token "plate" introduces interference activations (marked by white boxes) to the later explained token "fork". (e) Our TAM reveals the original information of the explained token "fork", eliminating context interferences and beyond CAM for a single output in (d). Results in (c, d, e) derive from Qwen2-VL-2B on the COCO Caption dataset. + +MLLMs [4, 16, 35, 47, 51], research on MLLM explainability remains less explored. It is crucial for user trust, model understanding, case analysis, and visualization. However, exploring the explainability of MLLMs is more challenging than conventional visual models (such as CNN [24], ViT [23], CLIP [41]), where MLLMs progressively generate multiple tokens beyond a single output as Fig. 1a. Thus, we aim to explain all tokens of MLLM, revealing the visual cues via token-level activation maps (e.g., Fig. 1e). + +Current MLLM explanation works [7, 48, 55] are generally borrowed from techniques [38, 45] for conventional models. For example, LLaVA-CAM [55] uses Smooth + +Grad-CAM++ [38] to explain LLaVA [35]. The self-attention weights [49] are applied to understand math problem-solving [48]. Besides, LVLM-Interpret [7] integrates some conventional methods as a tool. Unfortunately, existing visual explainability methods are designed for conventional models [23, 24, 41] that produce a single output, such as class activation maps (CAM) series [9, 26, 45, 56], attention-based approaches [1, 28], attention with relevance [2, 11], model-agnostic methods [36, 43], or multimodal explanation [3, 19, 31, 32]. However, these methods overlook that MLLMs generate multiple tokens progressively, where earlier context tokens interfere with later tokens to be explained. As shown in Fig. 1d, the conventional CAM [56] for the token "fork" is confounded by numerous redundant activations concurrent with the token "plate" in context (context refers to prompt and earlier answer tokens than the explained token). Quantitative analysis in Fig. 1c reveals that this issue is widespread, where correlated tokens lead to close and interference activations. These interferences obscure the original information conveyed by the explained token, significantly degrading the explainability of MLLMs. Notably, this activation interference across tokens is unique to MLLMs and has been overlooked. + +We aim to reveal the original information for the explained token while minimizing interference from context, as discussed above. Herein, the original information reveals a causal relation [40] between the current token and the input, while the interferences represent the correlations that need to be excluded. Inspired by the potential outcome model (POM) of causal inference [27], we propose the estimated causal inference to explore this causal relation. Unlike raw causal inference by comparing ready-made results (yes vs. no), there is no single given output to represent the "no" result. Therefore, we estimated interferences from multiple context tokens that are highly related to the target, with a scale factor optimized by the least-squares method to generate the refined causal activation map. We also propose a novel rank Gaussian filter to reduce noises, thereby further enhancing the quality of activation maps. The above methods are collectively referred to as token activation map (TAM), emphasizing the consideration of interactions between tokens. TAM also indicates that it excels at explaining tokens in multiple rounds of MLLM, which differs from the Class Activation Map (CAM) for a single prediction. + +Experimentally, the proposed TAM demonstrates quantitative improvement over existing SoTA explainability methods by more than $8.96\%$ on the COCO Caption dataset [13] and $8.54\%$ on the OpenPSG dataset [57] when applied to visually explain the Qwen2-VL-2B [51]. Meanwhile, it is complementary to existing methods to boost their performance. We further verify the scalability and applicability on 7 advanced MLLMs from the series of Qwen2-VL [51], InternVL2_5 [15] and LLaVA1_5 [35], where the + +performance improvements range from $5.45\%$ to $11.0\%$ on COCO Caption dataset [13]. Besides quantitative results, we conduct qualitative experiments to understand MLLM attributes about color, shape, action, location, etc. Furthermore, we illustrate that TAM is capable of explaining failure cases on the QK-VQA dataset. [37], and it supports video visualization well on the STAR dataset [52]. In addition, it can be used as a tool to visually compare MLLMs qualitatively. Moreover, TAM shows wide applicability to explain multi-turn conversations, multi-image scenarios, and visual reasoning (see the catalog Table 5 in supplementary for extensive examples). Our work has three main contributions: + +- We introduce TAM, a novel approach to explain MLLMs, which produces high-quality activation maps for multiple generative tokens by incorporating the estimated causal inference to alleviate interference from context tokens, with the rank Gaussian filter to reduce activation noises. +TAM significantly outperforms current methods for MLLM explanation, while also complementing them. +- TAM demonstrates wide applicability and scalability, serving as a versatile tool for object localization, model understanding, failure case analysis, video visualization, MLLM comparison, and supporting diverse scenarios (e.g., multi-turn, multi-image, visual reasoning). + +# 2. Related Work + +Visual Explainability for Multimodal LLM. The emergence of MLLMs such as LLaVA [35], Qwen2-VL [51], InternVL [15], and closed-source GPT-4o [25] has significantly impacted various tasks involving multimodal inputs like text, images, and video. We mainly focus on the visual explanation of MLLMs via the activation maps for the input image or video frames. Regarding the modality perspective, the explainability of MLLMs is partially related to methods to explain multimodal models such as CLIP Surgery [32] and transformer register [19] designed for CLIP [41]; InterpreT [3] and Bi-Modal [10] for BERT [20]. From the technical aspect, conventional explanation methods are also related, including class activation map series (CAM [56], Grad-CAM [45], Grad-CAM++ [9], LayerCAM [26]), attention-based mechanisms (LRP [28], Rollout [1]), the combination (AttnLRP [2], Grad×AttnRoll [11]), or model-agnostic methods (LIME [43], SHAP [36]). + +Although existing methods may be valid to explain MLLMs, they are not the optimal solution. Because they are usually designed for conventional models (e.g., CNN [24], ViT [23], CLIP [41]) that produce a single output rather than generating multiple tokens progressively in MLLMs. We have observed that these earlier context tokens interfere with later tokens by introducing redundant visual activations, as shown in Fig. 1. This phenomenon is first studied by us, solved by a novel estimated causal inference method, which is specially designed for MLLM. Notably, + +some workshop papers (LLaVA-CAM [55] and LVLM-Interpret [7]) applied conventional methods [38, 44] to explain LLaVA [35], but they also overlooked the interference from context tokens and treated each token independently. Causal Model. Causal inference [54] seeks to establish causal relation [40] between variables, focusing on how changes in one variable influence another. In contrast to statistical correlation analysis, it emphasizes causal relations. Another related concept, causal intervention [27], involves manipulating variables using a do operator. Model-agnostic explainability methods, such as LIME [43] and SHAP [36], can be viewed as causal interventions that explore the relationship between inputs and structured outputs by masking inputs or selecting subsets [12]. This concept is also applied to transformers through token masking (CLEANN [44]) or hidden state replacement [39]. However, causal intervention for MLLMs is impractical. Because there are too many input tokens are outputs in MLLM, requiring unbearable inference times. Besides, modified input leads to a changed context every inference, which is hard to evaluate. Our method is inspired by the potential outcome model (POM) to achieve causal inference without additional model inferences. Our contribution lies in estimating the unprocessed output from multiple tokens rather than a single ready-made output, considering the uniqueness of MLLMs. + +Transformer Denoising. Activation maps of transformers [49] usually present many noise activations, impeding high-quality visualization. Recent studies give their explanations about it, including attention sink [53], lack of register [19], and redundant features across classes [32]. In MLLM, these noises still exist obversely, even system tokens already play the role of registers, or deploying the feature surgery to mitigate redundant features (See Table 2). To solve this problem practically, we aim to introduce denoising filters to mitigate it as post-processing. Since these noises belong to salt-and-pepper noise morphologically, we apply the median filter, adaptive median filter [8], and the Gaussian filter. However, these filters are not the optimal solution, where the Gaussian filter keeps too much noise signal and the median filter overlooks smaller responses. Therefore, we propose the rank Gaussian filter, a novel, simple, and effective filter to denoise transformer activation. It merges ranked values within a sliding window, weighted summed by a 1-d Gaussian kernel with the center at the median rank. Besides, we improve the Gaussian kernel by the coefficient of variation for more robust results. Overall, our method addresses the activation noises in a practical aspect through a new filter. + +# 3. Method + +In this section, we first provide an overview of the proposed Token Activation Map (TAM), as illustrated in Fig. 2. Next, we elaborate on two key modules within TAM: the estimated causal inference and the rank Gaussian filter. Finally, + +we introduce three metrics designed to enable fine-grained evaluations of the explanations for MLLMs. + +# 3.1. Token Activation Map + +TAM aims to depict explainable activation maps for multiple tokens generated from MLLMs by leveraging causal inference to mitigate inter-token interference. To be more specific, given the visual data and prompt text as the input, the MLLM is employed to progressively generate the visual features $\pmb{F}^{v} \in \mathbb{R}^{n_{v} \times c}$ , prompt features $\pmb{F}^{p} \in \mathbb{R}^{n_{p} \times c}$ and answer features $\pmb{F}^{a} \in \mathbb{R}^{n_{a} \times c}$ . $n_{v}, n_{p}, n_{a}$ and $c$ indicate the number of three types of features, and the feature dimension, respectively. Next, a token classifier, instantiated as a fully connected layer, is used to generate answer tokens $\pmb{t}^{a} = \{t_{1}^{a} \dots t_{n_{a}}^{a}\}$ from these features. Prompt tokens are defined as $\pmb{t}^{p} = \{t_{1}^{p} \dots t_{n_{p}}^{p}\}$ similarly. Based on the variables described above, the visual activation maps for the prompt tokens and answer tokens are calculated as follows: + +$$ +\boldsymbol {A} _ {i} ^ {p} = \left\lfloor \boldsymbol {F} ^ {v} \boldsymbol {w} _ {t _ {i} ^ {p}} \right\rfloor_ {+}, i \in [ 1, n _ {p} ] +$$ + +$$ +\boldsymbol {A} _ {i} ^ {a} = \left\lfloor \boldsymbol {F} ^ {v} \boldsymbol {w} _ {t _ {i} ^ {a}} \right\rfloor_ {+}, i \in [ 1, n _ {a} ], \tag {1} +$$ + +where $\mathbf{A}_i^p \in \mathbb{R}^{n_v \times 1}$ and $\mathbf{A}_i^a \in \mathbb{R}^{n_v \times 1}$ present the activation map for the $i$ -th prompt token $t_i^p$ and answer token $t_i^a$ . $\mathbf{w}_{t_i^p} \in \mathbb{R}^{c \times 1}$ and $\mathbf{w}_{t_i^a} \in \mathbb{R}^{c \times 1}$ are the corresponding weight vectors within the token classifier. $\lfloor \rfloor_+$ is the function to keep positive activations. + +Then, the activation maps for prompt and answer tokens are concatenated $(\bot)$ as $\pmb{A}_{:n_p + n_a} = \pmb{A}^p\bot \pmb{A}^a$ . For the $i$ -th answer token, its context token is formulated as $\pmb{A}_{:n_p + i - 1}$ , including all earlier context maps (prompt tokens + answer tokens earlier than $i$ ). On top of it, the refined activation maps for answer tokens are formulated as follows: + +$$ +\bar {\boldsymbol {A}} _ {i} ^ {a} = \left\lfloor \mathcal {D} \left(\boldsymbol {A} _ {i} ^ {a} - s \mathcal {E} \left(\boldsymbol {A} _ {: n _ {p} + i - 1}\right)\right) \right\rfloor_ {+}, i \in [ 1, n _ {a} ], \tag {2} +$$ + +where $\bar{A}_i^a$ indicates the refined activation map for the $i$ -th answer token. The number of $i$ increases sequentially from 1 to $n_a$ . $A_{i}^{a} - s\mathcal{E}(A_{:n_{p} + i - 1})$ is the proposed estimated causal inference that produces the causal activation map with the estimation function $\mathcal{E}$ described in Sec. 3.2. $\mathcal{D}$ is the proposed rank Gaussian filter module in Sec. 3.3. + +In the last, we concatenate $(\perp)$ the refined visual activation map $\bar{A}_i^a$ of the $i$ -th answer token with the raw textual relevance $\mathbf{r}_i \in \mathbb{R}^{n_p + i - 1}$ for a multimodal activation map (image or video with text) $M_i \in \mathbb{R}^{n_v + n_p + i - 1}$ as follows: + +$$ +\boldsymbol {M} _ {i} = \mathcal {N} \left(\bar {\boldsymbol {A}} _ {i} ^ {a} \perp \mathbf {r} _ {i}\right), i \in [ 1, n _ {a} ], \tag {3} +$$ + +where $\mathcal{N}()$ is a min-max normalization function for visualizing visual and textual activations at an aligned level. + +# 3.2. Estimated Causal Inference + +Overall, Eq. 2 optimizes later visual activation maps based on visual activation maps of all earlier context tokens. To + +![](images/eeac58906753acd5b559578b2f97e86510bf04ec0dc649f6e62c85b13ec2f662.jpg) +Figure 2. Illustration of method. (a, Eq. 1 - Eq. 3) The overall framework of TAM. (b, Eq. 4 - Eq. 5) Details of the estimated casual inference module. (d, Eq. 6 - Eq. 7) Details of the rank Gaussian filter module. (d, Eq. 8 - Eq. 10) Fine-grained evaluation metrics. + +![](images/69b6e0b7bca0597e4ebf01775f955dde8c33175b395bc38a0322b3ced06b1c6a.jpg) + +![](images/0611e6074c1428e876590a78e1a68a1f020fffed3d6fa478a4d73e48bec275e4.jpg) + +![](images/272d80884ac3d74fc7477ac3a6125281c6eea90bf7d5a6a512e62d6c8fcdd9c8.jpg) + +be more specific, causal inference is employed to investigate the correct causal relation between the current answer token and visual tokens, eliminating interferences from prompt tokens and preceding answer tokens. To this end, we quantify the interference map from the context tokens for the $i$ -th answer token as follows: + +$$ +\mathcal {E} \left(\boldsymbol {A} _ {: n _ {p} + i - 1}\right) = \sum_ {k = 1} ^ {n _ {p} + i - 1} \frac {r _ {i} ^ {k}}{\sum \mathbf {r} _ {i} + \epsilon} \boldsymbol {A} _ {k} \tag {4} +$$ + +$$ +s. t. r _ {i} ^ {k} = 0, \mathrm {i f} t _ {i} ^ {a} = t _ {k}, +$$ + +where $r_i^k$ is the textual relevance between the $k$ -th context token $t_k$ and the current answer token $t_i^a$ . The relevance is obtained from activations among textual tokens $\lfloor (\pmb{F}^p \perp \pmb{F}^a) \pmb{w}_{t_i^a} \rfloor_+$ following Eq. 1. If the current context token $t_k$ is the same as the answer token $t_i^a$ , we set their relevance as 0, i.e., $r_i^k = 0$ , to avoid eliminating visual activations for tokens identical to the current answer token. + +Overall, Eq. 4 aims for each answer token to encapsulate more visual information, which implies minimizing similarity to the visual semantics of previously context tokens. Meanwhile, we use the scale factor $s$ in Eq. 2 to control the scale of the interference map, ensuring that the influence of the interference map is removed from the activation map at a comparable degree. Specifically, we employ the least squares method to minimize the residual between the two maps, for the optimal value of $s$ , as follows: + +$$ +s = \arg \min _ {s} \sum_ {j = 1} ^ {n _ {v}} \left(\boldsymbol {A} _ {j, i} ^ {a} - s \mathcal {E} \left(\boldsymbol {A} _ {j,: n _ {p} + i - 1}\right)\right) ^ {2}. \tag {5} +$$ + +After processing by the estimated causal inference module, there are still many salt-and-pepper noises in activation + +maps. To address this issue, we will elaborate on the solution in the following section. + +# 3.3. Rank Gaussian Filter + +As described in Sec. 2 the activation noise exists obviously and belongs to the salt-and-pepper type. Different from existing methods [19, 32] with suboptimal performances on MLLM, we address this problem in an new aspect: image denoising. Compared with existing filters like the Gaussian filter that keeps too much noise signal and the median filter which overlooks smaller responses, our rank Gaussian filter is more robust beyond existing methods. The overall equation of the rank Gaussian filter for any activation map $A_{i,j} \in \mathbb{R}^{h \times w}$ is written as: + +$$ +\begin{array}{l} \mathcal {D} \left(\boldsymbol {A} _ {i, j}, k\right) = \sum \boldsymbol {A} _ {i, j} \mathcal {G} \left(\boldsymbol {a} _ {i, j}, k\right), \forall i \in H, \forall j \in W \tag {6} \\ s. t. \mathbf {a} _ {i, j} = r a n k (\mathbf {A} _ {i: i + k, j: j + k}). \\ \end{array} +$$ + +Herein, $A_{i:i + k,j:j + k}$ indicates local values within a sliding window at kernel size $k$ . We sort them to get a ranked array to multiply with a custom 1-d Gaussian kernel $\mathcal{G}(\cdot)$ . This custom kernel allows the median to have the largest weight, meanwhile aggregating signals ranked close to the median. It can be regarded as a kind of smoothed median filter to enhance the robustness weighted by the Gaussian kernel. + +For the custom 1-d Gaussian kernel, we formulate it as Eq. 7. Specifically, $(i - k^2 / / 2)^2$ indicates the distance from the median rank $(k^2 / / 2$ , size exact division 2), and $\sigma (\mathbf{a}) / \mu (\mathbf{a})$ is the coefficient of variation which is more stable than the standard division $\sigma ()$ in general Gaussian kernel using the mean $\mu ()$ only. + +$$ +\mathcal {G} _ {i} (\mathbf {a}, k) = e ^ {- \frac {(i - k ^ {2} / / 2) ^ {2}}{2 (\sigma (\mathbf {a}) / \mu (\mathbf {a})) ^ {2}}}, \forall i \in k ^ {2} \tag {7} +$$ + +# 3.4. Evaluation for MLLM Explanation + +Unlike conventional models with a single output, MLLMs generate multiple tokens. To evaluate how well activation maps correspond to objects, we propose two fine-grained IoU-based metrics: Obj-IoU for objects with manual masks, and Func-IoU for backgrounds. These metrics focus on plausibility instead of faithfulness (see concept differences in Supp. E). Plausibility better suits MLLMs since perturbation-based faithfulness tests [2, 11] which cause inconsistent text generation, making evaluation invalid (see detailed reasons in Supp. E). + +To be specific, we formulate Obj-IoU in Eq. 8, counting the IoU between activation map $(A_{i})$ after binarization $(\mathcal{B}()$ using opencv OTSU auto threshold) and the groundtruth mask $G_{i}$ . Since the object classes may be unfixed, we count the IoU for all tokens at size $o$ over the dataset without an average on classes. We obtain the mask by matching the token to mask names. For a word or phrase containing multiple tokens, we record their max IoU for a single object. + +$$ +\operatorname {O b j} - \operatorname {I o U} = \frac {1}{o} \sum_ {i = 1} ^ {o} \frac {\mathcal {B} (\boldsymbol {A} _ {i}) \cap \boldsymbol {G} _ {i}}{\mathcal {B} (\boldsymbol {A} _ {i}) \cup \boldsymbol {G} _ {i}} \tag {8} +$$ + +Func-IoU measures the activation degree of tokens that are unreadable from the image (e.g., "is", "and", "so", "the"). Higher Func-IoU indicates fewer false positives and unnecessary responses. Specifically, we record the mean binarization thresholds $b_{i}$ (opencv OTSU auto threshold) for all noun tokens in the same image to divide foreground and background on function words. Activation map lower than it $(A_{i} < b_{i})$ is the background prediction used to count IoU with the all-one matrix $J$ (ground-truth is all background class). It is applied to total $u$ tokens over the whole dataset. + +$$ +\operatorname {F u n c - I o U} = \frac {1}{u} \sum_ {i = 1} ^ {u} \frac {\left(\boldsymbol {A} _ {i} < b _ {i}\right) \cap \boldsymbol {J}}{\left(\boldsymbol {A} _ {i} < b _ {i}\right) \cup \boldsymbol {J}} \tag {9} +$$ + +We try to merge the above IoUs as the overall major metrics. We find their average is insufficient to measure biased and low-quality activation maps. For example, if one method predicts almost all tokens to be background, the Obj-IoU is very low but Func-IoU is high, forming a certain inverse ratio (e.g., $5.74\%$ vs. $96.5\%$ of Attention-Rollout [1] in Table 3). One suitable solution is to use the F1 value to merge metrics with a certain inverse ratio. Thus, we have the F1-IoU to evaluate the overall plausibility stably: + +$$ +\mathrm {F 1 - I o U} = \frac {2 * \text {O b j - I o U} * \text {F u n c - I o U}}{\text {O b j - I o U} + \text {F u n c - I o U}} \tag {10} +$$ + +# 4. Experiments + +# 4.1. Setup + +Datasets. For quantitative experiments, we use datasets with texts and pixel-level annotations. Our main dataset is the COCO Caption dataset [13], which draws images and + +masks from COCO2014 [33]. Since explainability methods require testing only, we use its 5K-image minival split without training. Additional datasets include GranDf [42] (1K images) and OpenPSG [57] (3,176-image validation set). Masks in COCO Caption and GranDf are manually annotated, while OpenPSG masks are integrated by Rasheed et al. [42]. For datasets without masks, we perform qualitative visualization tests (Sec. 4.3), including attribute-only images and the QK-VQA [37] validation set (5,046 VQA samples with textual answers). We also qualitatively evaluate the STAR video dataset [52] (914 videos). + +Implementations. We use the SciPy minimize function with the BFGS method to optimize the scale factor $s$ in Eq. 5. The only hyperparameter, the kernel size $k$ in Eq. 7, is set to 3, as our observations indicate that salt-and-pepper noises typically cluster between 1 to 5 tokens. A kernel size of 3, consisting of 9 elements, is deemed suitable. Eq. 9 incorporates part-of-speech identification. For Func-IoU and Obj-IoU, we tag each word's part of speech using the pos_tag function from the NLTK Python package (version 3.8.1, see tag details in Supp. E). For word matching to mask names, we employ the lemmatize function from nltk.stem. Note that the thresholds in Eq. 8 and Eq. 9 are automatically derived from OpenCV's OTSU method, requiring no manual operation. Details of the baselines are provided in Supp. D. + +# 4.2. Quantitative Results + +
ECIRGFObj-IoU (%)Func-IoU (%)F1-IoU (%)
XX21.2351.9330.14
✓X22.4169.0333.84+3.7
X✓24.8243.3431.57+1.33
✓✓27.3768.4439.1+8.96
+ +Table 1. Ablation study on the COCO Caption [13] dataset using Qwen2-VL-2B [51]. These two modules are mutually beneficial, the combination exceeds the sum of their individual gains. ECI indicates the proposed estimated causal inference, and RGF is the rank Gaussian filter. Metrics are IoU for object words, IoU for function words, and their F1-score-like combination, respectively. + +
SettingObj-IoU (%)Func-IoU (%)F1-IoU (%)
Baseline21.2351.9330.14
Replace rank Gaussian filter
Adaptive Median [8]25.4868.237.1
Median Filter26.0168.2637.67
Gaussian Filter26.5667.7838.16
Replace estimated causal inference
Feature surgery [32]18.548.6626.81
ECI-mean27.8449.8535.72
ECI-attnWeights27.1154.6136.23
Ours27.3768.4439.1
+ +Table 2. Effectiveness study on the COCO Caption [13] dataset using Qwen2-VL-2B [51]. ECI-mean and ECI-attnWeights are candidate implementations of ECI described in Supp. D. + +Ablation and Effectiveness. We present the ablation study results in Table 1, which indicates that each module of the TAM is valuable. Specifically, the estimated causal infer + +
MethodTypeCOCO CaptionGranDfOpenPSG
Obj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoU
Grad-CAM [45]21.2351.9330.1417.8562.1527.7422.9348.5731.15
Grad-CAM++ [9]Gradient19.5262.8329.7817.373.4228.0122.2159.9532.41
Grad-Rollout [1]1.2799.512.511.499.612.771.5799.5813.08
Layer-CAM [26]11.4384.8820.1513.1182.0922.6214.1285.2924.22
AttentionAttention8.292.8715.079.693.5617.4210.5894.2819.03
Attention-Rollout [1]5.7496.510.837.2196.6513.427.9497.0414.68
CP-LRP [5]Combination9.953.9716.7312.6153.2420.3913.353.3621.3
Attn-LRP [2]9.9252.4116.6912.1552.1919.7212.7852.2620.54
CAM [56]Logit21.2351.9330.1417.8562.1527.7422.9348.5731.15
Archi.-Surgery [32]15.6963.8225.1916.5962.2826.219.8358.7729.65
TAM (ours)27.3768.4439.118.6588.9730.8326.2692.9940.95
Archi.-Surgery [32]+ Ours+4.13+10.68+6.12+1.41+26.48+3.74+1.59+31.92+5.0
Grad-CAM++ [9]+2.83+3.91+3.7+1.26+5.39+2.04+2.75+9.44+4.3
Layer-CAM [26]+5.04-7.95+6.98+2.79+0.52+4.05+6.62-2.49+8.95
CAM [56]+6.14+16.52+8.96+0.79+26.82+3.09+3.33+44.42+9.8
+ +Table 3. Comparison with SoTA methods using Qwen2-VL-2B on diverse datasets. We adopt the "Logit" type that uses the classifier weights in Eq. 1, without back-propagation in "Gradient" and "Combination" (Gradient + Attention). Besides, we support FlashAttention [18] and SdpaAttention, which do not return attention weights, while "Attention" and "Combination" rely on it. Our TAM is also complementary to existing methods with gains marked by "+" Note, CAM and Grad-CAM are equivalent as discussed in Supp. D. The major metric is the F1-IoU $(\%)$ to reflect the overall result, merged from IoUs for Object words (Obj-IoU) and function words (Func-IoU). + +
CAM [56] +OursCOCO CaptionGranDfOpenPSG
Obj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoU
LLaVA1_5-7B [35]+4.47+18.27+7.97+0.65+11.67+2.47+3.47+9.51+5.16
LLaVA1_5-13B [35]+4.3+7.31+5.45+0.76+7.03+2.11+4.22+11.51+6.37
InternVL2_5-2B [16]+5.44+19.48+8.57+2.2+48.29+8.47+3.24+40.45+8.65
InternVL2_5-4B [16]+3.53+22.17+7.14+1.62+45.19+7.56+2.2+55.05+10.85
InternVL2_5-8B [16]+5.38+2.13+6.93+3.53+28.53+7.02+3.27+26.54+6.44
Qwen2-VL-2B [51]+6.14+16.52+8.96+0.79+26.82+3.09+3.33+44.49+9.8
Qwen2-VL-7B [51]+5.62+29.4+11.01+1.29+22.54+3.4+3.53+46.94+11.15
Mean Improvements+4.98+16.47+8.01+1.55+27.15+4.88+3.32+33.5+8.35
+ +Table 4. Explainability improvements of TAM on diverse MLLMs and datasets. Specific results are listed in Table 6 of Supp. K. + +ence significantly improves Func-IoU by reducing context interference on function words, while the rank Gaussian filter performs better for object words. Importantly, these modules are mutually beneficial; their combination leads to an F1-IoU improvement of $8.96\%$ , exceeding the total of their individual contributions (3.7% and 1.33%). + +We further assess the effectiveness of these modules by comparing them with other methods, as shown in Table 2. Replacing our proposed rank Gaussian filter with alternatives yields lower performance, such as the adaptive median filter [8], the Gaussian filter, and the median filter. Although the differences are less pronounced than those seen with the causal inference module, denoising remains a challenging area. Our method outperforms feature surgery [32] (which operates on the class dimension) by $12.29\%$ and also surpasses other ECI implementations in Supp. D. + +Comparison with SoTA Methods. Table 3 presents a comparison between TAM and SoTA methods. Our TAM outperforms all others in the major F1-IoU and Obj-IoU metrics. While some methods, such as Grad-Rollout [1] and Attention, achieve high Func-IoU by producing a limited number of activations, their performance on other metrics re + +mains low. The second-best method is Grad-CAM++ [9] on the COCO Caption [13] and OpenPSG [57] datasets, while CAM [56] and Grad-CAM [45] rank second on GranDf [42]. Our method surpasses these approaches by $8.96\%$ , $2.82\%$ , and $8.54\%$ on these datasets, respectively. Furthermore, TAM is complementary to existing methods and enhances them significantly, with maximum gains of $6.62\%$ , $44.42\%$ , and $9.8\%$ across the three metrics, respectively. + +Applicability and Scalability. We evaluate the applicability on three prominent MLLMs, as shown in Table 4, and assess its scalability across various model sizes. The results in Table 3 indicate that TAM consistently outperforms the baseline across all MLLMs and model sizes by significant margins. This demonstrates that TAM can be effectively applied to a range of MLLMs while providing superior explainability compared to existing methods. Moreover, TAM supports MLLMs comparison, complementing quantitative results. It introduces a new evaluation perspective by highlighting visual alignment quality, as illustrated in Fig. 8. + +# 4.3. Qualitative Results + +Comparison of Explainability Methods. + +![](images/793b00d420c7a03590beea94a3c4f37f08f790a2f15101a48c6c8fa90eabcef2.jpg) +Figure 3. Visual comparison between TAM and SoTA methods on the COCO Caption dataset [13] using the Qwen2-VL-2B [51] model. The "is" in the last row is a functional word with background ground-truth. More examples are shown in Supp. G including complex cases. + +![](images/133398b6e279c2171db9f3b72958c2f2decb373681c29d9dc9479c7df46a451c.jpg) +Figure 4. TAM presents high-quality localization results on diverse datasets [13, 42, 57] under the Qwen2-VL-7B [51] model. + +We perform comparisons between the proposed TAM and SoTA explanation methods, as illustrated in Fig. 3. The visualization results clearly show that TAM more accurately highlights target objects and renders fewer false activations on function words. In contrast, attention-based methods [1, 2, 5] produce only a limited number of activations, while gradient-based methods [9, 45] present too many correlated responses and noise, performing significantly worse than our approach. Additionally, we present more complex examples in Supp. G, where TAM notably outperforms existing SoTA methods. TAM excels at explaining multiple tokens of MLLMs. In Supp. C, results show how TAM focuses on important tokens, whereas the baseline methods generate many redundant activations. + +High-quality Localization Results. We visualized the results using the multimodal activation map in Eq. 3 (see detailed examples in Supp. B). The results demonstrate that TAM effectively locates objects across diverse datasets, as shown in Fig. 4. These findings indicate that TAM enhances MLLMs with localization capabilities, which could be potentially beneficial for various downstream tasks, including segmentation [32], object counting [46], anomaly detection [14], image editing [6], autonomous driving [22], and ap + +![](images/d9944da5bff3f04ff6c48ab4097e93cc9aabd7648e6c5599cd74f2dae88d5aec.jpg) +Figure 5. TAM supports to explain and analyze diverse attributes of MLLM (Qwen2-VL-7B). See extensive examples in Supp. M. + +plications in the medical field [21, 50], which require pixellevel activations. + +Visualization for Various Attributes. We apply TAM to explain various attributes of MLLMs, specifically focusing on fine-grained token types as Fig. 5. The results indicate that TAM effectively supports the explanation of attributes such as color, shape, action, number, location, and text. Notably, the representation of numbers is less pronounced compared to object localization, suggesting that some attributes may be encoded within related object tokens. In contrast, other attributes are more prominently featured. + +Explaining Failure Cases. We apply TAM to analyze failure cases in the QK-VQA dataset [37] as Fig. 7. Our findings indicate that while the model successfully locates objects, it struggles to align them with specific knowledge. For instance, the model can identify the desert and focus on the key prompt ("year") when outputting "don't know." It suggests a misalignment between the desert and the year, rather than recognition or prompt issues. Besides, some responses consist of synonyms, hyphenms, or hyponyms of the correct answers, resulting in mismatches. For more case studies, please refer to the catalog in Supp. Table 5. + +![](images/3d01dc3e24f8cd68c76f5ddebf2c7295eb8854fd87679b8cbce63c576eb3ca3c.jpg) +Figure 6. TAM significantly improves the quality of video visualization, with much fewer redundant activation and noises on STAR [52] dataset using Qwen2-VL-2B [51]. Please see extensive examples in Supp. 27 and failure cases analysis for videos in Fig. 25. + +![](images/5a41589abb588a35c283540f42a091f30434ad52b580b37a4735ba4fd9b9f6ac.jpg) + +![](images/fceab8ba0355ed65903104963968dc5153bae6d2e3ce952854f51147e2d0fb26.jpg) +Figure 7. TAM supports failure cases analysis on QK-VQA [37]. See more cases in Supp. about VQA (Fig. 24), videos (Fig. 25), visual reasoning (Fig. 28), and multi-turn conversation (Fig. 31). + +![](images/587602b470b3cd0bd90a00cda47eaad67cf026e08726ddf9ea733c09f803aa53.jpg) + +![](images/8aa8054cb80577901dddab1e8628e4d104752e90d969668eed0cd7bced221f34.jpg) + +![](images/dbe630a86dc5d791fac0cae47d5996d220b75f4366ae3655691154b0df768c24.jpg) +Figure 8. TAM enables qualitative comparison among MLLMs. Extensive examples are given in Fig. 17 and Fig. 18 in Supp. K. + +Video Visualization. The proposed TAM excels at video visualization, where there are much fewer redundant activations and noises as shown in Fig. 6. Our method provides very clear results, while the conventional CAM makes it hard to see the video content without obvious highlights. + +A Tool to Visually Comparison MLLMs. TAM can be used as a tool to compare MLLMs qualitatively. If an MLLM exhibits superior activation maps, it can be inferred that its explainability and degree of vision alignment are also better, offering valuable pixel-level predictions at no + +additional cost. As shown in Fig. 8, Qwen2-VL-7B [51] shows greater explainability than LLaVA1_5-7B [35] and InternVL2_5-8B [16]. Furthermore, we observe that InternVL tends to focus more on text, resulting in weaker activation levels; in some cases, this performance is even better than that of LLaVA (e.g., with the broccoli). More extensive examples can be found in Supp. K. + +# 5. Conclusion + +This work focuses on the unique nature of MLLMs in visual explainability, which generates multiple tokens progressively, complicating the explanation process. In response, we introduced the Token Activation Map (TAM), a novel method that utilizes estimated causal inference to mitigate interferences from context tokens, with a rank Gaussian filter to reduce activation noise, thereby providing clearer visual explanations. Our results demonstrate that TAM significantly outperforms existing SoTA methods, offering high-quality visualizations applicable to diverse scenarios. + +Despite the success, we focus on visual inputs, other modalities, such as audio, remain under-explored. Additionally, interpreting model decisions is an extensible further aspect. The potential applications of our TAM method are extensive, including open-vocabulary segmentation, detection, image grounding, anomaly detection, remote sensing, and medical fields that require pixel-level activations. Our work lays a foundation for advancing their explainability and practical utility across various domains. + +# Acknowledgment + +This work was supported by a research grant from the Joint Research Scheme (JRS) under the National Natural Science Foundation of China (NSFC) and the Research Grants Council (RGC) of Hong Kong (Project No. N_HKUST654/24), as well as a grant from the RGC of the Hong Kong Special Administrative Region, China (Project No. R6005-24). + +# References + +[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4190-4197, 2020. 2, 5, 6, 7, 14, 15 +[2] Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. Attnlrp: Attention-aware layer-wise relevance propagation for transformers. In International Conference on Machine Learning, pages 135-168. PMLR, 2024. 2, 5, 6, 7, 14, 15 +[3] Estelle Aflalo, Meng Du, Shao-Yen Tseng, Yongfei Liu, Chenfei Wu, Nan Duan, and Vasudev Lal. Vl-interpret: An interactive visualization tool for interpreting vision-language transformers. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 21406–21415, 2022. 2 +[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1 +[5] Ameen Ali, Thomas Schnake, Oliver Eberle, Grégoire Montavon, Klaus-Robert Müller, and Lior Wolf. Xai for transformers: Better explanations through conservative propagation. In International Conference on Machine Learning, pages 435-451. PMLR, 2022. 6, 7, 14, 15 +[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 7 +[7] Gabriela Ben Melech Stan, Estelle Aflalo, Raanan Yehezkel Rohekar, Anahita Bhiwandiwalla, Shao-Yen Tseng, Matthew Lyle Olson, Yaniv Gurwicz, Chenfei Wu, Nan Duan, and Vasudev Lal. Lvlm-intrepret: An interpretability tool for large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8182-8187, 2024. 1, 2, 3 +[8] Chin-Chen Chang, Ju-Yuan Hsiao, and Chih-Ping Hsieh. An adaptive median filter for image denoising. In 2008 Second international symposium on intelligent information technology application, pages 346-350. IEEE, 2008. 3, 5, 6, 16 +[9] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE winter conference on applications of computer vision (WACV), pages 839-847. IEEE, 2018. 2, 6, 7, 14, 15 +[10] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 2 +[11] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of + +the IEEE/CVF conference on computer vision and pattern recognition, pages 782-791, 2021. 2, 5, 15 +[12] Ruoyu Chen, Hua Zhang, Siyuan Liang, Jingzhi Li, and Xiaochun Cao. Less is more: Fewer interpretable region via submodular subset selection. In The Twelfth International Conference on Learning Representations. 3 +[13] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 2, 5, 6, 7, 13, 16, 17, 20, 21, 22, 23, 24, 26, 29, 30 +[14] Xuhai Chen, Jiangning Zhang, Guanzhong Tian, Haoyang He, Wuhao Zhang, Yabiao Wang, Chengjie Wang, and Yong Liu. Clip-ad: A language-guided staged dual-path model for zero-shot anomaly detection. In International Joint Conference on Artificial Intelligence, pages 17-33. Springer, 2024. 7 +[15] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 2, 13 +[16] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 1, 6, 8, 16, 17, 19, 25, 35 +[17] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1 +[18] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 6 +[19] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations. 2, 3, 4 +[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 2 +[21] Xinpeng Ding, Yongqiang Chu, Renjie Pi, Hualiang Wang, and Xiaomeng Li. Hia: Towards chinese multimodal llms for comparative high-resolution joint diagnosis. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 575-586. Springer, 2024. 7 +[22] Xinpeng Ding, Jianhua Han, Hang Xu, Xiaodan Liang, Wei Zhang, and Xiaomeng Li. Holistic autonomous driving understanding by bird's view injected multi-modal large models. + +In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13668-13677, 2024. 7 +[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2 +[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2 +[25] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2 +[26] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875-5888, 2021. 2, 6, 14 +[27] Licheng Jiao, Yuhan Wang, Xu Liu, Lingling Li, Fang Liu, Wenping Ma, Yuwei Guo, Puhua Chen, Shuyuan Yang, and Biao Hou. Causal inference meets deep learning: A comprehensive survey. Research, 7:0467, 2024. 2, 3 +[28] Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Unmasking clever hans predictors and assessing what machines really learn. Nature communications, 10(1):1096, 2019. 2 +[29] Unggi Lee, Minji Jeon, Yunseo Lee, Gyuri Byun, Yoorim Son, Jaeyoon Shin, Hongkyu Ko, and Hyeoncheol Kim. Llava-docent: Instruction tuning with multimodal large language model to support art appreciation education. Computers and Education: Artificial Intelligence, 7:100297, 2024. 1 +[30] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18061-18070, 2024. 1 +[31] Yi Li, Hualiang Wang, Yiqun Duan, Hang Xu, and Xiaomeng Li. Exploring visual interpretability for contrastive language-image pre-training. arXiv preprint arXiv:2209.07046, 2022. 2 +[32] Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. A closer look at the explainability of contrastive language-image pre-training. Pattern Recognition, page 111409, 2025. 2, 3, 4, 5, 6, 7, 14, 15, 22 +[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[34] Fenglin Liu, Tingting Zhu, Xian Wu, Bang Yang, Chenyu You, Chenyang Wang, Lei Lu, Zhangdaihong Liu, Yefeng + +Zheng, Xu Sun, et al. A medical multimodal large language model for future pandemics. NPJ Digital Medicine, 6(1): 226, 2023. 1 +[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3, 6, 8, 13, 16, 17, 25 +[36] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017. 2, 3 +[37] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 5, 7, 8, 14, 18, 31, 33 +[38] Daniel Omeiza, Skyler Speakman, Celia Cintas, and Komminist Weldermariam. Smooth grad-cam++: An enhanced inference level visualization technique for deep convolutional neural network models. arXiv preprint arXiv:1908.01224, 2019. 1, 2, 3 +[39] Vedant Palit, Rohan Pandey, Aryaman Arora, and Paul Pu Liang. Towards vision-language mechanistic interpretability: A causal tracing tool for blip. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2856-2861, 2023. 3 +[40] Judea Pearl. Causality. Cambridge university press, 2009. 2,3 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 22 +[42] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 5, 6, 7, 13, 16, 17 +[43] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. "why should i trust you?" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144, 2016. 2, 3 +[44] Raanan Y Rohekar, Yaniv Gurwicz, and Shami Nisimov. Causal interpretation of self-attention in pre-trained transformers. Advances in Neural Information Processing Systems, 36, 2024. 3 +[45] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1, 2, 6, 7, 14, 15, 16, 17, 18, 20, 24, 30, 34 +[46] Zenglin Shi, Ying Sun, and Mengmi Zhang. Training-free object counting with prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 323-331, 2024. 7 + +[47] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1 +[48] Joseph Tey. Understanding how vision-language models reason when solving visual math problems. 1, 2 +[49] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3 +[50] Lehan Wang, Haonan Wang, Honglong Yang, Jiaji Mao, Zehong Yang, Jun Shen, and Xiaomeng Li. Interpretable bilingual multimodal large language model for diverse biomedical tasks. In The Thirteenth International Conference on Learning Representations, 2025. 7 +[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2, 5, 6, 7, 8, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38 +[52] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2). 2, 5, 8, 14, 18, 29, 32, 34 +[53] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. In The Twelfth International Conference on Learning Representations. 3 +[54] Liuyi Yao, Zhixuan Chu, Sheng Li, Yaliang Li, Jing Gao, and Aidong Zhang. A survey on causal inference. ACM Transactions on Knowledge Discovery from Data (TKDD), 15(5):1-46, 2021. 3 +[55] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multimodal large language models. Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, 2025. 1, 3 +[56] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 6, 16, 17, 18, 20, 24, 30, 34 +[57] Zijian Zhou, Zheng Zhu, Holger Caesar, and Miaojing Shi. Openpsg: Open-set panoptic scene graph generation via large multimodal models. In European Conference on Computer Vision, pages 199-215. Springer, 2024. 2, 5, 6, 7, 14, 16, 17 \ No newline at end of file diff --git a/2025/Token Activation Map to Visually Explain Multimodal LLMs/images.zip b/2025/Token Activation Map to Visually Explain Multimodal LLMs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8f2b5badc2c2b10ce56b19db8a9dbf5248ca874d --- /dev/null +++ b/2025/Token Activation Map to Visually Explain Multimodal LLMs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f52f4cc33c7a2555bd80f4e0981ccad577532d88545974fca907fe1213370cdc +size 928178 diff --git a/2025/Token Activation Map to Visually Explain Multimodal LLMs/layout.json b/2025/Token Activation Map to Visually Explain Multimodal LLMs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..642f7eaf6ea2eaca707d67d76693bf940537d20e --- /dev/null +++ b/2025/Token Activation Map to Visually Explain Multimodal LLMs/layout.json @@ -0,0 +1,8749 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 495, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 495, + 121 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 495, + 121 + ], + "type": "text", + "content": "Token Activation Map to Visually Explain Multimodal LLMs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 143, + 143, + 473, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 143, + 473, + 158 + ], + "spans": [ + { + "bbox": [ + 143, + 143, + 473, + 158 + ], + "type": "text", + "content": "Yi Li, Hualiang Wang, Xinpeng Ding, Haonan Wang, Xiaomeng Li*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 159, + 550, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 550, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 550, + 171 + ], + "type": "text", + "content": "Department of Electronic and Computer Engineering, The Hong Kong University of Science and Technology, Hong Kong SAR, China." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 257, + 171, + 357, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 257, + 171, + 357, + 186 + ], + "spans": [ + { + "bbox": [ + 257, + 171, + 357, + 186 + ], + "type": "text", + "content": "{ylini,xmli} " + }, + { + "bbox": [ + 257, + 171, + 357, + 186 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 257, + 171, + 357, + 186 + ], + "type": "text", + "content": " ust.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 239, + 297, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 586 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 586 + ], + "type": "text", + "content": "Multimodal large language models (MLLMs) are broadly empowering various fields. Despite their advancements, the explainability of MLLMs remains less explored, hindering deeper understanding, model credibility, and effective visualization. Unlike conventional vision models (e.g., CNNs, ViTs, CLIP) that produce a single output, MLLMs generate sequences of tokens progressively, where each generated token depends on the previous context. Therefore, earlier context tokens can introduce redundant activations that interfere with the explanation of later tokens beyond their original information. Existing studies often overlook this issue, but our observations reveal that these redundant correlations can significantly hurt the reliability of explanations. To address this, we propose an estimated causal inference method to mitigate the interference of context to achieve high-quality MLLM explanation, with a novel rank Gaussian filter to further reduce activation noises. We term this method Token Activation Map (TAM) to highlight the consideration of interactions between tokens. TAM also indicates that it excels at explaining multiple tokens of MLLM, which is different from the Class Activation Map (CAM) for a single prediction. Our TAM method significantly outperforms existing SoTA methods, showcasing high-quality visualization results that can be utilized for various scenarios, such as object localization, failure case analysis, video visualization, MLLMs visual comparison, and model understanding (e.g., color, shape, action, location, visual reasoning, multi-turn conversation, etc). The code is available at github.com/xmed-lab/TAM." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 611, + 135, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 611, + 135, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 611, + 135, + 624 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 296, + 693 + ], + "type": "text", + "content": "Multimodal large language models (MLLMs or Multimodal LLMs) increasingly empower wide applications [17, 29, 30, 34], enabling multimodal inputs (e.g., images, videos, text) and human-like conversations. Although extensive efforts have been devoted to enhancing the performance of" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 212, + 410, + 316 + ], + "blocks": [ + { + "bbox": [ + 317, + 212, + 410, + 316 + ], + "lines": [ + { + "bbox": [ + 317, + 212, + 410, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 410, + 316 + ], + "type": "image", + "image_path": "2a3412272f6e2129275a35c47ac409a77f84afd6c81c3ea480de827eafa11480.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 317, + 402, + 324 + ], + "lines": [ + { + "bbox": [ + 317, + 317, + 402, + 324 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 402, + 324 + ], + "type": "text", + "content": "(a) Conventional models vs. MLLMs" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 410, + 213, + 491, + 316 + ], + "blocks": [ + { + "bbox": [ + 410, + 213, + 491, + 316 + ], + "lines": [ + { + "bbox": [ + 410, + 213, + 491, + 316 + ], + "spans": [ + { + "bbox": [ + 410, + 213, + 491, + 316 + ], + "type": "image", + "image_path": "f6f335c94f06f1a55049345c3fa1d7115795ab805e3cbcb67385e68d60bbc04c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 317, + 480, + 324 + ], + "lines": [ + { + "bbox": [ + 410, + 317, + 480, + 324 + ], + "spans": [ + { + "bbox": [ + 410, + 317, + 480, + 324 + ], + "type": "text", + "content": "(b) Information flow of MLLM" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 491, + 213, + 555, + 316 + ], + "blocks": [ + { + "bbox": [ + 491, + 213, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 491, + 213, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 491, + 213, + 555, + 316 + ], + "type": "image", + "image_path": "bccf1e88009a26b6bebdf77574a705286367ea6675bfa676d7123bb27ac6c61f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 317, + 552, + 324 + ], + "lines": [ + { + "bbox": [ + 492, + 317, + 552, + 324 + ], + "spans": [ + { + "bbox": [ + 492, + 317, + 552, + 324 + ], + "type": "text", + "content": "(C) Correlated activations" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 328, + 396, + 393 + ], + "blocks": [ + { + "bbox": [ + 316, + 328, + 396, + 393 + ], + "lines": [ + { + "bbox": [ + 316, + 328, + 396, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 396, + 393 + ], + "type": "image", + "image_path": "3859da4497307882cd97f8f1d183b03fea829e0040d55fb37c4fbdcffb775218.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 395, + 447, + 401 + ], + "lines": [ + { + "bbox": [ + 315, + 395, + 447, + 401 + ], + "spans": [ + { + "bbox": [ + 315, + 395, + 447, + 401 + ], + "type": "text", + "content": "(d) Qualitative example of interference activations" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 395, + 328, + 473, + 393 + ], + "blocks": [ + { + "bbox": [ + 395, + 328, + 473, + 393 + ], + "lines": [ + { + "bbox": [ + 395, + 328, + 473, + 393 + ], + "spans": [ + { + "bbox": [ + 395, + 328, + 473, + 393 + ], + "type": "image", + "image_path": "85dc2c305cb925a43f0289596bfe6c312c26f97e3bbbf8d8f9f906a48f352a7d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 472, + 395, + 555, + 402 + ], + "lines": [ + { + "bbox": [ + 472, + 395, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 472, + 395, + 555, + 402 + ], + "type": "text", + "content": "(e) TAM (ours) for MLLM explanation" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 473, + 328, + 555, + 393 + ], + "blocks": [ + { + "bbox": [ + 473, + 328, + 555, + 393 + ], + "lines": [ + { + "bbox": [ + 473, + 328, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 473, + 328, + 555, + 393 + ], + "type": "image", + "image_path": "57d544af825f0f08bf5f11b60798d366f21eea240fe56dd630a5c65cc905e931.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 555 + ], + "type": "text", + "content": "Figure 1. Illustration of Motivation. (a) MLLMs generate multiple tokens beyond a single output. (b) The information flow indicates that MLLM generates tokens progressively, where one generated token (each row) is correlated with the context (prompt + earlier answer tokens). (c) We randomly pair CAMs and count their L1 distance against text correlation. Higher text correlation corresponds to lower distance, indicating concurrent interferences. (d) The example shows that the context token \"plate\" introduces interference activations (marked by white boxes) to the later explained token \"fork\". (e) Our TAM reveals the original information of the explained token \"fork\", eliminating context interferences and beyond CAM for a single output in (d). Results in (c, d, e) derive from Qwen2-VL-2B on the COCO Caption dataset." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 569, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 555, + 677 + ], + "type": "text", + "content": "MLLMs [4, 16, 35, 47, 51], research on MLLM explainability remains less explored. It is crucial for user trust, model understanding, case analysis, and visualization. However, exploring the explainability of MLLMs is more challenging than conventional visual models (such as CNN [24], ViT [23], CLIP [41]), where MLLMs progressively generate multiple tokens beyond a single output as Fig. 1a. Thus, we aim to explain all tokens of MLLM, revealing the visual cues via token-level activation maps (e.g., Fig. 1e)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Current MLLM explanation works [7, 48, 55] are generally borrowed from techniques [38, 45] for conventional models. For example, LLaVA-CAM [55] uses Smooth" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "text", + "content": "48" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 346 + ], + "type": "text", + "content": "Grad-CAM++ [38] to explain LLaVA [35]. The self-attention weights [49] are applied to understand math problem-solving [48]. Besides, LVLM-Interpret [7] integrates some conventional methods as a tool. Unfortunately, existing visual explainability methods are designed for conventional models [23, 24, 41] that produce a single output, such as class activation maps (CAM) series [9, 26, 45, 56], attention-based approaches [1, 28], attention with relevance [2, 11], model-agnostic methods [36, 43], or multimodal explanation [3, 19, 31, 32]. However, these methods overlook that MLLMs generate multiple tokens progressively, where earlier context tokens interfere with later tokens to be explained. As shown in Fig. 1d, the conventional CAM [56] for the token \"fork\" is confounded by numerous redundant activations concurrent with the token \"plate\" in context (context refers to prompt and earlier answer tokens than the explained token). Quantitative analysis in Fig. 1c reveals that this issue is widespread, where correlated tokens lead to close and interference activations. These interferences obscure the original information conveyed by the explained token, significantly degrading the explainability of MLLMs. Notably, this activation interference across tokens is unique to MLLMs and has been overlooked." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 350, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 602 + ], + "type": "text", + "content": "We aim to reveal the original information for the explained token while minimizing interference from context, as discussed above. Herein, the original information reveals a causal relation [40] between the current token and the input, while the interferences represent the correlations that need to be excluded. Inspired by the potential outcome model (POM) of causal inference [27], we propose the estimated causal inference to explore this causal relation. Unlike raw causal inference by comparing ready-made results (yes vs. no), there is no single given output to represent the \"no\" result. Therefore, we estimated interferences from multiple context tokens that are highly related to the target, with a scale factor optimized by the least-squares method to generate the refined causal activation map. We also propose a novel rank Gaussian filter to reduce noises, thereby further enhancing the quality of activation maps. The above methods are collectively referred to as token activation map (TAM), emphasizing the consideration of interactions between tokens. TAM also indicates that it excels at explaining tokens in multiple rounds of MLLM, which differs from the Class Activation Map (CAM) for a single prediction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "Experimentally, the proposed TAM demonstrates quantitative improvement over existing SoTA explainability methods by more than " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "8.96\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": " on the COCO Caption dataset [13] and " + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "inline_equation", + "content": "8.54\\%" + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": " on the OpenPSG dataset [57] when applied to visually explain the Qwen2-VL-2B [51]. Meanwhile, it is complementary to existing methods to boost their performance. We further verify the scalability and applicability on 7 advanced MLLMs from the series of Qwen2-VL [51], InternVL2_5 [15] and LLaVA1_5 [35], where the" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "performance improvements range from " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "5.45\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "11.0\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " on COCO Caption dataset [13]. Besides quantitative results, we conduct qualitative experiments to understand MLLM attributes about color, shape, action, location, etc. Furthermore, we illustrate that TAM is capable of explaining failure cases on the QK-VQA dataset. [37], and it supports video visualization well on the STAR dataset [52]. In addition, it can be used as a tool to visually compare MLLMs qualitatively. Moreover, TAM shows wide applicability to explain multi-turn conversations, multi-image scenarios, and visual reasoning (see the catalog Table 5 in supplementary for extensive examples). Our work has three main contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 217, + 553, + 361 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 217, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 217, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 314, + 217, + 553, + 276 + ], + "type": "text", + "content": "- We introduce TAM, a novel approach to explain MLLMs, which produces high-quality activation maps for multiple generative tokens by incorporating the estimated causal inference to alleviate interference from context tokens, with the rank Gaussian filter to reduce activation noises." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 277, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 300 + ], + "type": "text", + "content": "TAM significantly outperforms current methods for MLLM explanation, while also complementing them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "type": "text", + "content": "- TAM demonstrates wide applicability and scalability, serving as a versatile tool for object localization, model understanding, failure case analysis, video visualization, MLLM comparison, and supporting diverse scenarios (e.g., multi-turn, multi-image, visual reasoning)." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 370, + 400, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 370, + 400, + 383 + ], + "spans": [ + { + "bbox": [ + 314, + 370, + 400, + 383 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 390, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 594 + ], + "type": "text", + "content": "Visual Explainability for Multimodal LLM. The emergence of MLLMs such as LLaVA [35], Qwen2-VL [51], InternVL [15], and closed-source GPT-4o [25] has significantly impacted various tasks involving multimodal inputs like text, images, and video. We mainly focus on the visual explanation of MLLMs via the activation maps for the input image or video frames. Regarding the modality perspective, the explainability of MLLMs is partially related to methods to explain multimodal models such as CLIP Surgery [32] and transformer register [19] designed for CLIP [41]; InterpreT [3] and Bi-Modal [10] for BERT [20]. From the technical aspect, conventional explanation methods are also related, including class activation map series (CAM [56], Grad-CAM [45], Grad-CAM++ [9], LayerCAM [26]), attention-based mechanisms (LRP [28], Rollout [1]), the combination (AttnLRP [2], Grad×AttnRoll [11]), or model-agnostic methods (LIME [43], SHAP [36])." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Although existing methods may be valid to explain MLLMs, they are not the optimal solution. Because they are usually designed for conventional models (e.g., CNN [24], ViT [23], CLIP [41]) that produce a single output rather than generating multiple tokens progressively in MLLMs. We have observed that these earlier context tokens interfere with later tokens by introducing redundant visual activations, as shown in Fig. 1. This phenomenon is first studied by us, solved by a novel estimated causal inference method, which is specially designed for MLLM. Notably," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "text", + "content": "49" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 370 + ], + "type": "text", + "content": "some workshop papers (LLaVA-CAM [55] and LVLM-Interpret [7]) applied conventional methods [38, 44] to explain LLaVA [35], but they also overlooked the interference from context tokens and treated each token independently. Causal Model. Causal inference [54] seeks to establish causal relation [40] between variables, focusing on how changes in one variable influence another. In contrast to statistical correlation analysis, it emphasizes causal relations. Another related concept, causal intervention [27], involves manipulating variables using a do operator. Model-agnostic explainability methods, such as LIME [43] and SHAP [36], can be viewed as causal interventions that explore the relationship between inputs and structured outputs by masking inputs or selecting subsets [12]. This concept is also applied to transformers through token masking (CLEANN [44]) or hidden state replacement [39]. However, causal intervention for MLLMs is impractical. Because there are too many input tokens are outputs in MLLM, requiring unbearable inference times. Besides, modified input leads to a changed context every inference, which is hard to evaluate. Our method is inspired by the potential outcome model (POM) to achieve causal inference without additional model inferences. Our contribution lies in estimating the unprocessed output from multiple tokens rather than a single ready-made output, considering the uniqueness of MLLMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 371, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 635 + ], + "type": "text", + "content": "Transformer Denoising. Activation maps of transformers [49] usually present many noise activations, impeding high-quality visualization. Recent studies give their explanations about it, including attention sink [53], lack of register [19], and redundant features across classes [32]. In MLLM, these noises still exist obversely, even system tokens already play the role of registers, or deploying the feature surgery to mitigate redundant features (See Table 2). To solve this problem practically, we aim to introduce denoising filters to mitigate it as post-processing. Since these noises belong to salt-and-pepper noise morphologically, we apply the median filter, adaptive median filter [8], and the Gaussian filter. However, these filters are not the optimal solution, where the Gaussian filter keeps too much noise signal and the median filter overlooks smaller responses. Therefore, we propose the rank Gaussian filter, a novel, simple, and effective filter to denoise transformer activation. It merges ranked values within a sliding window, weighted summed by a 1-d Gaussian kernel with the center at the median rank. Besides, we improve the Gaussian kernel by the coefficient of variation for more robust results. Overall, our method addresses the activation noises in a practical aspect through a new filter." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 645, + 111, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 111, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 111, + 658 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "In this section, we first provide an overview of the proposed Token Activation Map (TAM), as illustrated in Fig. 2. Next, we elaborate on two key modules within TAM: the estimated causal inference and the rank Gaussian filter. Finally," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "we introduce three metrics designed to enable fine-grained evaluations of the explanations for MLLMs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 102, + 441, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 102, + 441, + 115 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 441, + 115 + ], + "type": "text", + "content": "3.1. Token Activation Map" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "text", + "content": "TAM aims to depict explainable activation maps for multiple tokens generated from MLLMs by leveraging causal inference to mitigate inter-token interference. To be more specific, given the visual data and prompt text as the input, the MLLM is employed to progressively generate the visual features " + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\pmb{F}^{v} \\in \\mathbb{R}^{n_{v} \\times c}" + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "text", + "content": ", prompt features " + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\pmb{F}^{p} \\in \\mathbb{R}^{n_{p} \\times c}" + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "text", + "content": " and answer features " + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\pmb{F}^{a} \\in \\mathbb{R}^{n_{a} \\times c}" + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "inline_equation", + "content": "n_{v}, n_{p}, n_{a}" + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "text", + "content": " indicate the number of three types of features, and the feature dimension, respectively. Next, a token classifier, instantiated as a fully connected layer, is used to generate answer tokens " + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\pmb{t}^{a} = \\{t_{1}^{a} \\dots t_{n_{a}}^{a}\\}" + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "text", + "content": " from these features. Prompt tokens are defined as " + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "inline_equation", + "content": "\\pmb{t}^{p} = \\{t_{1}^{p} \\dots t_{n_{p}}^{p}\\}" + }, + { + "bbox": [ + 313, + 120, + 555, + 287 + ], + "type": "text", + "content": " similarly. Based on the variables described above, the visual activation maps for the prompt tokens and answer tokens are calculated as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 375, + 293, + 492, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 293, + 492, + 307 + ], + "spans": [ + { + "bbox": [ + 375, + 293, + 492, + 307 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} _ {i} ^ {p} = \\left\\lfloor \\boldsymbol {F} ^ {v} \\boldsymbol {w} _ {t _ {i} ^ {p}} \\right\\rfloor_ {+}, i \\in [ 1, n _ {p} ]", + "image_path": "8ee77e686867cb3e520eb333618d7495da364c27626523b96bcc94c635da7b5f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 374, + 303, + 553, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 303, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 374, + 303, + 553, + 323 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} _ {i} ^ {a} = \\left\\lfloor \\boldsymbol {F} ^ {v} \\boldsymbol {w} _ {t _ {i} ^ {a}} \\right\\rfloor_ {+}, i \\in [ 1, n _ {a} ], \\tag {1}", + "image_path": "633e4108620662e7bbba4f941c87812aa2f008ec0e8e55df4c47d32af5fbb1e4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i^p \\in \\mathbb{R}^{n_v \\times 1}" + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i^a \\in \\mathbb{R}^{n_v \\times 1}" + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "text", + "content": " present the activation map for the " + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "text", + "content": "-th prompt token " + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "inline_equation", + "content": "t_i^p" + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "text", + "content": " and answer token " + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "inline_equation", + "content": "t_i^a" + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{t_i^p} \\in \\mathbb{R}^{c \\times 1}" + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{t_i^a} \\in \\mathbb{R}^{c \\times 1}" + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "text", + "content": " are the corresponding weight vectors within the token classifier. " + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\lfloor \\rfloor_+" + }, + { + "bbox": [ + 313, + 330, + 554, + 389 + ], + "type": "text", + "content": " is the function to keep positive activations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "type": "text", + "content": "Then, the activation maps for prompt and answer tokens are concatenated " + }, + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "type": "inline_equation", + "content": "(\\bot)" + }, + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{A}_{:n_p + n_a} = \\pmb{A}^p\\bot \\pmb{A}^a" + }, + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "type": "text", + "content": "-th answer token, its context token is formulated as " + }, + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "type": "inline_equation", + "content": "\\pmb{A}_{:n_p + i - 1}" + }, + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "type": "text", + "content": ", including all earlier context maps (prompt tokens + answer tokens earlier than " + }, + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 390, + 554, + 461 + ], + "type": "text", + "content": "). On top of it, the refined activation maps for answer tokens are formulated as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 468, + 553, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 468, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 331, + 468, + 553, + 483 + ], + "type": "interline_equation", + "content": "\\bar {\\boldsymbol {A}} _ {i} ^ {a} = \\left\\lfloor \\mathcal {D} \\left(\\boldsymbol {A} _ {i} ^ {a} - s \\mathcal {E} \\left(\\boldsymbol {A} _ {: n _ {p} + i - 1}\\right)\\right) \\right\\rfloor_ {+}, i \\in [ 1, n _ {a} ], \\tag {2}", + "image_path": "bc37a7b16b18e95eeac0fecad9c33974c4245a0e9409ba2e38a6c3948b130868.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\bar{A}_i^a" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": " indicates the refined activation map for the " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": "-th answer token. The number of " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": " increases sequentially from 1 to " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "n_a" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "A_{i}^{a} - s\\mathcal{E}(A_{:n_{p} + i - 1})" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": " is the proposed estimated causal inference that produces the causal activation map with the estimation function " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": " described in Sec. 3.2. " + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 491, + 554, + 563 + ], + "type": "text", + "content": " is the proposed rank Gaussian filter module in Sec. 3.3." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "content": "In the last, we concatenate " + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "inline_equation", + "content": "(\\perp)" + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "content": " the refined visual activation map " + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "inline_equation", + "content": "\\bar{A}_i^a" + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "content": "-th answer token with the raw textual relevance " + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i \\in \\mathbb{R}^{n_p + i - 1}" + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "content": " for a multimodal activation map (image or video with text) " + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "inline_equation", + "content": "M_i \\in \\mathbb{R}^{n_v + n_p + i - 1}" + }, + { + "bbox": [ + 313, + 563, + 554, + 611 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 618, + 553, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 618, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 370, + 618, + 553, + 633 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} _ {i} = \\mathcal {N} \\left(\\bar {\\boldsymbol {A}} _ {i} ^ {a} \\perp \\mathbf {r} _ {i}\\right), i \\in [ 1, n _ {a} ], \\tag {3}", + "image_path": "1453fb624a69b539f4490bd1d7d21cd0c38863a3e00f7185ed44c32b8312dc7b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{N}()" + }, + { + "bbox": [ + 313, + 641, + 554, + 665 + ], + "type": "text", + "content": " is a min-max normalization function for visualizing visual and textual activations at an aligned level." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 672, + 466, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 466, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 466, + 683 + ], + "type": "text", + "content": "3.2. Estimated Causal Inference" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Overall, Eq. 2 optimizes later visual activation maps based on visual activation maps of all earlier context tokens. To" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "50" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 193, + 264 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 193, + 264 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 193, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 193, + 264 + ], + "type": "image", + "image_path": "eeac58906753acd5b559578b2f97e86510bf04ec0dc649f6e62c85b13ec2f662.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 270, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 555, + 293 + ], + "type": "text", + "content": "Figure 2. Illustration of method. (a, Eq. 1 - Eq. 3) The overall framework of TAM. (b, Eq. 4 - Eq. 5) Details of the estimated casual inference module. (d, Eq. 6 - Eq. 7) Details of the rank Gaussian filter module. (d, Eq. 8 - Eq. 10) Fine-grained evaluation metrics." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 197, + 70, + 324, + 264 + ], + "blocks": [ + { + "bbox": [ + 197, + 70, + 324, + 264 + ], + "lines": [ + { + "bbox": [ + 197, + 70, + 324, + 264 + ], + "spans": [ + { + "bbox": [ + 197, + 70, + 324, + 264 + ], + "type": "image", + "image_path": "69b6e0b7bca0597e4ebf01775f955dde8c33175b395bc38a0322b3ced06b1c6a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 330, + 70, + 467, + 264 + ], + "blocks": [ + { + "bbox": [ + 330, + 70, + 467, + 264 + ], + "lines": [ + { + "bbox": [ + 330, + 70, + 467, + 264 + ], + "spans": [ + { + "bbox": [ + 330, + 70, + 467, + 264 + ], + "type": "image", + "image_path": "0611e6074c1428e876590a78e1a68a1f020fffed3d6fa478a4d73e48bec275e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 471, + 70, + 553, + 264 + ], + "blocks": [ + { + "bbox": [ + 471, + 70, + 553, + 264 + ], + "lines": [ + { + "bbox": [ + 471, + 70, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 471, + 70, + 553, + 264 + ], + "type": "image", + "image_path": "272d80884ac3d74fc7477ac3a6125281c6eea90bf7d5a6a512e62d6c8fcdd9c8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 314, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 296, + 385 + ], + "type": "text", + "content": "be more specific, causal inference is employed to investigate the correct causal relation between the current answer token and visual tokens, eliminating interferences from prompt tokens and preceding answer tokens. To this end, we quantify the interference map from the context tokens for the " + }, + { + "bbox": [ + 55, + 314, + 296, + 385 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 314, + 296, + 385 + ], + "type": "text", + "content": "-th answer token as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 392, + 295, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 392, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 94, + 392, + 295, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {E} \\left(\\boldsymbol {A} _ {: n _ {p} + i - 1}\\right) = \\sum_ {k = 1} ^ {n _ {p} + i - 1} \\frac {r _ {i} ^ {k}}{\\sum \\mathbf {r} _ {i} + \\epsilon} \\boldsymbol {A} _ {k} \\tag {4}", + "image_path": "22ccea5624c52e3207a41fedb4099a021b5cc25634f4d868dbc9e7094e82195a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 160, + 427, + 256, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 427, + 256, + 442 + ], + "spans": [ + { + "bbox": [ + 160, + 427, + 256, + 442 + ], + "type": "interline_equation", + "content": "s. t. r _ {i} ^ {k} = 0, \\mathrm {i f} t _ {i} ^ {a} = t _ {k},", + "image_path": "0f6c9897519fb6983775174f89b51e5815f393496f1228b670e6ffadbe690012.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "inline_equation", + "content": "r_i^k" + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "text", + "content": " is the textual relevance between the " + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "text", + "content": "-th context token " + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "text", + "content": " and the current answer token " + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "inline_equation", + "content": "t_i^a" + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "text", + "content": ". The relevance is obtained from activations among textual tokens " + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "inline_equation", + "content": "\\lfloor (\\pmb{F}^p \\perp \\pmb{F}^a) \\pmb{w}_{t_i^a} \\rfloor_+" + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "text", + "content": " following Eq. 1. If the current context token " + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "text", + "content": " is the same as the answer token " + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "inline_equation", + "content": "t_i^a" + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "text", + "content": ", we set their relevance as 0, i.e., " + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "inline_equation", + "content": "r_i^k = 0" + }, + { + "bbox": [ + 55, + 450, + 296, + 533 + ], + "type": "text", + "content": ", to avoid eliminating visual activations for tokens identical to the current answer token." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "type": "text", + "content": "Overall, Eq. 4 aims for each answer token to encapsulate more visual information, which implies minimizing similarity to the visual semantics of previously context tokens. Meanwhile, we use the scale factor " + }, + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "type": "text", + "content": " in Eq. 2 to control the scale of the interference map, ensuring that the influence of the interference map is removed from the activation map at a comparable degree. Specifically, we employ the least squares method to minimize the residual between the two maps, for the optimal value of " + }, + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 534, + 296, + 643 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 82, + 650, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 650, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 82, + 650, + 295, + 682 + ], + "type": "interline_equation", + "content": "s = \\arg \\min _ {s} \\sum_ {j = 1} ^ {n _ {v}} \\left(\\boldsymbol {A} _ {j, i} ^ {a} - s \\mathcal {E} \\left(\\boldsymbol {A} _ {j,: n _ {p} + i - 1}\\right)\\right) ^ {2}. \\tag {5}", + "image_path": "37b482e0ec261a2aed2288e890bec8bce028fd07a142d737d0ae100a7a704857.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "After processing by the estimated causal inference module, there are still many salt-and-pepper noises in activation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 314, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 555, + 338 + ], + "type": "text", + "content": "maps. To address this issue, we will elaborate on the solution in the following section." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 345, + 437, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 437, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 437, + 357 + ], + "type": "text", + "content": "3.3. Rank Gaussian Filter" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 362, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 483 + ], + "type": "text", + "content": "As described in Sec. 2 the activation noise exists obviously and belongs to the salt-and-pepper type. Different from existing methods [19, 32] with suboptimal performances on MLLM, we address this problem in an new aspect: image denoising. Compared with existing filters like the Gaussian filter that keeps too much noise signal and the median filter which overlooks smaller responses, our rank Gaussian filter is more robust beyond existing methods. The overall equation of the rank Gaussian filter for any activation map " + }, + { + "bbox": [ + 313, + 362, + 555, + 483 + ], + "type": "inline_equation", + "content": "A_{i,j} \\in \\mathbb{R}^{h \\times w}" + }, + { + "bbox": [ + 313, + 362, + 555, + 483 + ], + "type": "text", + "content": " is written as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 491, + 553, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 491, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 321, + 491, + 553, + 525 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {D} \\left(\\boldsymbol {A} _ {i, j}, k\\right) = \\sum \\boldsymbol {A} _ {i, j} \\mathcal {G} \\left(\\boldsymbol {a} _ {i, j}, k\\right), \\forall i \\in H, \\forall j \\in W \\tag {6} \\\\ s. t. \\mathbf {a} _ {i, j} = r a n k (\\mathbf {A} _ {i: i + k, j: j + k}). \\\\ \\end{array}", + "image_path": "196da3d418346b47300a7f325d7abeda024dda48045ca1cdcf75ef84a971bc1c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 532, + 554, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 615 + ], + "type": "text", + "content": "Herein, " + }, + { + "bbox": [ + 313, + 532, + 554, + 615 + ], + "type": "inline_equation", + "content": "A_{i:i + k,j:j + k}" + }, + { + "bbox": [ + 313, + 532, + 554, + 615 + ], + "type": "text", + "content": " indicates local values within a sliding window at kernel size " + }, + { + "bbox": [ + 313, + 532, + 554, + 615 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 532, + 554, + 615 + ], + "type": "text", + "content": ". We sort them to get a ranked array to multiply with a custom 1-d Gaussian kernel " + }, + { + "bbox": [ + 313, + 532, + 554, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\cdot)" + }, + { + "bbox": [ + 313, + 532, + 554, + 615 + ], + "type": "text", + "content": ". This custom kernel allows the median to have the largest weight, meanwhile aggregating signals ranked close to the median. It can be regarded as a kind of smoothed median filter to enhance the robustness weighted by the Gaussian kernel." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "text", + "content": "For the custom 1-d Gaussian kernel, we formulate it as Eq. 7. Specifically, " + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "inline_equation", + "content": "(i - k^2 / / 2)^2" + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "text", + "content": " indicates the distance from the median rank " + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "inline_equation", + "content": "(k^2 / / 2" + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "text", + "content": ", size exact division 2), and " + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\sigma (\\mathbf{a}) / \\mu (\\mathbf{a})" + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "text", + "content": " is the coefficient of variation which is more stable than the standard division " + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\sigma ()" + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "text", + "content": " in general Gaussian kernel using the mean " + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "inline_equation", + "content": "\\mu ()" + }, + { + "bbox": [ + 313, + 616, + 554, + 688 + ], + "type": "text", + "content": " only." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 695, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 695, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 361, + 695, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {i} (\\mathbf {a}, k) = e ^ {- \\frac {(i - k ^ {2} / / 2) ^ {2}}{2 (\\sigma (\\mathbf {a}) / \\mu (\\mathbf {a})) ^ {2}}}, \\forall i \\in k ^ {2} \\tag {7}", + "image_path": "d9aa01bcc3b3917b3331eb7dbeb8517b1639b07c8d77b9860d8c892ab676de63.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "51" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 244, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 244, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 244, + 85 + ], + "type": "text", + "content": "3.4. Evaluation for MLLM Explanation" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 89, + 295, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 89, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 54, + 89, + 295, + 209 + ], + "type": "text", + "content": "Unlike conventional models with a single output, MLLMs generate multiple tokens. To evaluate how well activation maps correspond to objects, we propose two fine-grained IoU-based metrics: Obj-IoU for objects with manual masks, and Func-IoU for backgrounds. These metrics focus on plausibility instead of faithfulness (see concept differences in Supp. E). Plausibility better suits MLLMs since perturbation-based faithfulness tests [2, 11] which cause inconsistent text generation, making evaluation invalid (see detailed reasons in Supp. E)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 209, + 296, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 305 + ], + "type": "text", + "content": "To be specific, we formulate Obj-IoU in Eq. 8, counting the IoU between activation map " + }, + { + "bbox": [ + 55, + 209, + 296, + 305 + ], + "type": "inline_equation", + "content": "(A_{i})" + }, + { + "bbox": [ + 55, + 209, + 296, + 305 + ], + "type": "text", + "content": " after binarization " + }, + { + "bbox": [ + 55, + 209, + 296, + 305 + ], + "type": "inline_equation", + "content": "(\\mathcal{B}()" + }, + { + "bbox": [ + 55, + 209, + 296, + 305 + ], + "type": "text", + "content": " using opencv OTSU auto threshold) and the groundtruth mask " + }, + { + "bbox": [ + 55, + 209, + 296, + 305 + ], + "type": "inline_equation", + "content": "G_{i}" + }, + { + "bbox": [ + 55, + 209, + 296, + 305 + ], + "type": "text", + "content": ". Since the object classes may be unfixed, we count the IoU for all tokens at size " + }, + { + "bbox": [ + 55, + 209, + 296, + 305 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 209, + 296, + 305 + ], + "type": "text", + "content": " over the dataset without an average on classes. We obtain the mask by matching the token to mask names. For a word or phrase containing multiple tokens, we record their max IoU for a single object." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 313, + 295, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 313, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 121, + 313, + 295, + 339 + ], + "type": "interline_equation", + "content": "\\operatorname {O b j} - \\operatorname {I o U} = \\frac {1}{o} \\sum_ {i = 1} ^ {o} \\frac {\\mathcal {B} (\\boldsymbol {A} _ {i}) \\cap \\boldsymbol {G} _ {i}}{\\mathcal {B} (\\boldsymbol {A} _ {i}) \\cup \\boldsymbol {G} _ {i}} \\tag {8}", + "image_path": "3b9aecc7ccb0964e1b7a0dc34fc20ca07410c53663fde671e2d75ff36378f7db.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "content": "Func-IoU measures the activation degree of tokens that are unreadable from the image (e.g., \"is\", \"and\", \"so\", \"the\"). Higher Func-IoU indicates fewer false positives and unnecessary responses. Specifically, we record the mean binarization thresholds " + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "content": " (opencv OTSU auto threshold) for all noun tokens in the same image to divide foreground and background on function words. Activation map lower than it " + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "inline_equation", + "content": "(A_{i} < b_{i})" + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "content": " is the background prediction used to count IoU with the all-one matrix " + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "content": " (ground-truth is all background class). It is applied to total " + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "content": " tokens over the whole dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 468, + 295, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 468, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 115, + 468, + 295, + 495 + ], + "type": "interline_equation", + "content": "\\operatorname {F u n c - I o U} = \\frac {1}{u} \\sum_ {i = 1} ^ {u} \\frac {\\left(\\boldsymbol {A} _ {i} < b _ {i}\\right) \\cap \\boldsymbol {J}}{\\left(\\boldsymbol {A} _ {i} < b _ {i}\\right) \\cup \\boldsymbol {J}} \\tag {9}", + "image_path": "952b5d0ef50f69764f8a14a828fde5c8247df0b12fdaefe080b93039c22ed230.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 496, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 604 + ], + "type": "text", + "content": "We try to merge the above IoUs as the overall major metrics. We find their average is insufficient to measure biased and low-quality activation maps. For example, if one method predicts almost all tokens to be background, the Obj-IoU is very low but Func-IoU is high, forming a certain inverse ratio (e.g., " + }, + { + "bbox": [ + 55, + 496, + 296, + 604 + ], + "type": "inline_equation", + "content": "5.74\\%" + }, + { + "bbox": [ + 55, + 496, + 296, + 604 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 55, + 496, + 296, + 604 + ], + "type": "inline_equation", + "content": "96.5\\%" + }, + { + "bbox": [ + 55, + 496, + 296, + 604 + ], + "type": "text", + "content": " of Attention-Rollout [1] in Table 3). One suitable solution is to use the F1 value to merge metrics with a certain inverse ratio. Thus, we have the F1-IoU to evaluate the overall plausibility stably:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 609, + 295, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 609, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 113, + 609, + 295, + 633 + ], + "type": "interline_equation", + "content": "\\mathrm {F 1 - I o U} = \\frac {2 * \\text {O b j - I o U} * \\text {F u n c - I o U}}{\\text {O b j - I o U} + \\text {F u n c - I o U}} \\tag {10}", + "image_path": "a260f9a64f753cd0bcc7511cfe702365efd7bd37945b8b6a4648a2e1ecb2ae1d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 640, + 137, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 137, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 137, + 654 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 660, + 105, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 105, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 105, + 673 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Datasets. For quantitative experiments, we use datasets with texts and pixel-level annotations. Our main dataset is the COCO Caption dataset [13], which draws images and" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "masks from COCO2014 [33]. Since explainability methods require testing only, we use its 5K-image minival split without training. Additional datasets include GranDf [42] (1K images) and OpenPSG [57] (3,176-image validation set). Masks in COCO Caption and GranDf are manually annotated, while OpenPSG masks are integrated by Rasheed et al. [42]. For datasets without masks, we perform qualitative visualization tests (Sec. 4.3), including attribute-only images and the QK-VQA [37] validation set (5,046 VQA samples with textual answers). We also qualitatively evaluate the STAR video dataset [52] (914 videos)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 204, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 373 + ], + "type": "text", + "content": "Implementations. We use the SciPy minimize function with the BFGS method to optimize the scale factor " + }, + { + "bbox": [ + 313, + 204, + 555, + 373 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 204, + 555, + 373 + ], + "type": "text", + "content": " in Eq. 5. The only hyperparameter, the kernel size " + }, + { + "bbox": [ + 313, + 204, + 555, + 373 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 204, + 555, + 373 + ], + "type": "text", + "content": " in Eq. 7, is set to 3, as our observations indicate that salt-and-pepper noises typically cluster between 1 to 5 tokens. A kernel size of 3, consisting of 9 elements, is deemed suitable. Eq. 9 incorporates part-of-speech identification. For Func-IoU and Obj-IoU, we tag each word's part of speech using the pos_tag function from the NLTK Python package (version 3.8.1, see tag details in Supp. E). For word matching to mask names, we employ the lemmatize function from nltk.stem. Note that the thresholds in Eq. 8 and Eq. 9 are automatically derived from OpenCV's OTSU method, requiring no manual operation. Details of the baselines are provided in Supp. D." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 379, + 433, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 433, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 433, + 392 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 316, + 399, + 552, + 451 + ], + "blocks": [ + { + "bbox": [ + 316, + 399, + 552, + 451 + ], + "lines": [ + { + "bbox": [ + 316, + 399, + 552, + 451 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 552, + 451 + ], + "type": "table", + "html": "
ECIRGFObj-IoU (%)Func-IoU (%)F1-IoU (%)
XX21.2351.9330.14
✓X22.4169.0333.84+3.7
X✓24.8243.3431.57+1.33
✓✓27.3768.4439.1+8.96
", + "image_path": "ff11e5f34b41fc4d2378868b9fa1d778a66ed22a9001d5238ed757addb6dff78.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 315, + 528, + 556, + 636 + ], + "blocks": [ + { + "bbox": [ + 313, + 453, + 555, + 520 + ], + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 520 + ], + "type": "text", + "content": "Table 1. Ablation study on the COCO Caption [13] dataset using Qwen2-VL-2B [51]. These two modules are mutually beneficial, the combination exceeds the sum of their individual gains. ECI indicates the proposed estimated causal inference, and RGF is the rank Gaussian filter. Metrics are IoU for object words, IoU for function words, and their F1-score-like combination, respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 528, + 556, + 636 + ], + "lines": [ + { + "bbox": [ + 315, + 528, + 556, + 636 + ], + "spans": [ + { + "bbox": [ + 315, + 528, + 556, + 636 + ], + "type": "table", + "html": "
SettingObj-IoU (%)Func-IoU (%)F1-IoU (%)
Baseline21.2351.9330.14
Replace rank Gaussian filter
Adaptive Median [8]25.4868.237.1
Median Filter26.0168.2637.67
Gaussian Filter26.5667.7838.16
Replace estimated causal inference
Feature surgery [32]18.548.6626.81
ECI-mean27.8449.8535.72
ECI-attnWeights27.1154.6136.23
Ours27.3768.4439.1
", + "image_path": "ccce2867c4d653abd5ec0d974ba1addfb3c7dd0c220f1d65f53f00478f554b25.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 639, + 555, + 673 + ], + "lines": [ + { + "bbox": [ + 313, + 639, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 555, + 673 + ], + "type": "text", + "content": "Table 2. Effectiveness study on the COCO Caption [13] dataset using Qwen2-VL-2B [51]. ECI-mean and ECI-attnWeights are candidate implementations of ECI described in Supp. D." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Ablation and Effectiveness. We present the ablation study results in Table 1, which indicates that each module of the TAM is valuable. Specifically, the estimated causal infer" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "52" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 550, + 234 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 550, + 234 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 550, + 234 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 550, + 234 + ], + "type": "table", + "html": "
MethodTypeCOCO CaptionGranDfOpenPSG
Obj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoU
Grad-CAM [45]21.2351.9330.1417.8562.1527.7422.9348.5731.15
Grad-CAM++ [9]Gradient19.5262.8329.7817.373.4228.0122.2159.9532.41
Grad-Rollout [1]1.2799.512.511.499.612.771.5799.5813.08
Layer-CAM [26]11.4384.8820.1513.1182.0922.6214.1285.2924.22
AttentionAttention8.292.8715.079.693.5617.4210.5894.2819.03
Attention-Rollout [1]5.7496.510.837.2196.6513.427.9497.0414.68
CP-LRP [5]Combination9.953.9716.7312.6153.2420.3913.353.3621.3
Attn-LRP [2]9.9252.4116.6912.1552.1919.7212.7852.2620.54
CAM [56]Logit21.2351.9330.1417.8562.1527.7422.9348.5731.15
Archi.-Surgery [32]15.6963.8225.1916.5962.2826.219.8358.7729.65
TAM (ours)27.3768.4439.118.6588.9730.8326.2692.9940.95
Archi.-Surgery [32]+ Ours+4.13+10.68+6.12+1.41+26.48+3.74+1.59+31.92+5.0
Grad-CAM++ [9]+2.83+3.91+3.7+1.26+5.39+2.04+2.75+9.44+4.3
Layer-CAM [26]+5.04-7.95+6.98+2.79+0.52+4.05+6.62-2.49+8.95
CAM [56]+6.14+16.52+8.96+0.79+26.82+3.09+3.33+44.42+9.8
", + "image_path": "6c8e84c515b962f58f84113d103427477d9e007aee5284e84c659c0e79533e80.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 308, + 553, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 243, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 300 + ], + "type": "text", + "content": "Table 3. Comparison with SoTA methods using Qwen2-VL-2B on diverse datasets. We adopt the \"Logit\" type that uses the classifier weights in Eq. 1, without back-propagation in \"Gradient\" and \"Combination\" (Gradient + Attention). Besides, we support FlashAttention [18] and SdpaAttention, which do not return attention weights, while \"Attention\" and \"Combination\" rely on it. Our TAM is also complementary to existing methods with gains marked by \"+\" Note, CAM and Grad-CAM are equivalent as discussed in Supp. D. The major metric is the F1-IoU " + }, + { + "bbox": [ + 55, + 243, + 555, + 300 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 243, + 555, + 300 + ], + "type": "text", + "content": " to reflect the overall result, merged from IoUs for Object words (Obj-IoU) and function words (Func-IoU)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 308, + 553, + 407 + ], + "lines": [ + { + "bbox": [ + 56, + 308, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 553, + 407 + ], + "type": "table", + "html": "
CAM [56] +OursCOCO CaptionGranDfOpenPSG
Obj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoUObj-IoUFunc-IoUF1-IoU
LLaVA1_5-7B [35]+4.47+18.27+7.97+0.65+11.67+2.47+3.47+9.51+5.16
LLaVA1_5-13B [35]+4.3+7.31+5.45+0.76+7.03+2.11+4.22+11.51+6.37
InternVL2_5-2B [16]+5.44+19.48+8.57+2.2+48.29+8.47+3.24+40.45+8.65
InternVL2_5-4B [16]+3.53+22.17+7.14+1.62+45.19+7.56+2.2+55.05+10.85
InternVL2_5-8B [16]+5.38+2.13+6.93+3.53+28.53+7.02+3.27+26.54+6.44
Qwen2-VL-2B [51]+6.14+16.52+8.96+0.79+26.82+3.09+3.33+44.49+9.8
Qwen2-VL-7B [51]+5.62+29.4+11.01+1.29+22.54+3.4+3.53+46.94+11.15
Mean Improvements+4.98+16.47+8.01+1.55+27.15+4.88+3.32+33.5+8.35
", + "image_path": "fe41abbb7d27820028c61f34410566c33cf83030cacbc71cdadd49446db4b647.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 415, + 538, + 427 + ], + "lines": [ + { + "bbox": [ + 71, + 415, + 538, + 427 + ], + "spans": [ + { + "bbox": [ + 71, + 415, + 538, + 427 + ], + "type": "text", + "content": "Table 4. Explainability improvements of TAM on diverse MLLMs and datasets. Specific results are listed in Table 6 of Supp. K." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 448, + 295, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 448, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 295, + 519 + ], + "type": "text", + "content": "ence significantly improves Func-IoU by reducing context interference on function words, while the rank Gaussian filter performs better for object words. Importantly, these modules are mutually beneficial; their combination leads to an F1-IoU improvement of " + }, + { + "bbox": [ + 54, + 448, + 295, + 519 + ], + "type": "inline_equation", + "content": "8.96\\%" + }, + { + "bbox": [ + 54, + 448, + 295, + 519 + ], + "type": "text", + "content": ", exceeding the total of their individual contributions (3.7% and 1.33%)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "text", + "content": "We further assess the effectiveness of these modules by comparing them with other methods, as shown in Table 2. Replacing our proposed rank Gaussian filter with alternatives yields lower performance, such as the adaptive median filter [8], the Gaussian filter, and the median filter. Although the differences are less pronounced than those seen with the causal inference module, denoising remains a challenging area. Our method outperforms feature surgery [32] (which operates on the class dimension) by " + }, + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "12.29\\%" + }, + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "text", + "content": " and also surpasses other ECI implementations in Supp. D." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 642, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 295, + 712 + ], + "type": "text", + "content": "Comparison with SoTA Methods. Table 3 presents a comparison between TAM and SoTA methods. Our TAM outperforms all others in the major F1-IoU and Obj-IoU metrics. While some methods, such as Grad-Rollout [1] and Attention, achieve high Func-IoU by producing a limited number of activations, their performance on other metrics re" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "text", + "content": "mains low. The second-best method is Grad-CAM++ [9] on the COCO Caption [13] and OpenPSG [57] datasets, while CAM [56] and Grad-CAM [45] rank second on GranDf [42]. Our method surpasses these approaches by " + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "inline_equation", + "content": "8.96\\%" + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "inline_equation", + "content": "2.82\\%" + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "inline_equation", + "content": "8.54\\%" + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "text", + "content": " on these datasets, respectively. Furthermore, TAM is complementary to existing methods and enhances them significantly, with maximum gains of " + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "inline_equation", + "content": "6.62\\%" + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "inline_equation", + "content": "44.42\\%" + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "inline_equation", + "content": "9.8\\%" + }, + { + "bbox": [ + 313, + 448, + 553, + 543 + ], + "type": "text", + "content": " across the three metrics, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 544, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 675 + ], + "type": "text", + "content": "Applicability and Scalability. We evaluate the applicability on three prominent MLLMs, as shown in Table 4, and assess its scalability across various model sizes. The results in Table 3 indicate that TAM consistently outperforms the baseline across all MLLMs and model sizes by significant margins. This demonstrates that TAM can be effectively applied to a range of MLLMs while providing superior explainability compared to existing methods. Moreover, TAM supports MLLMs comparison, complementing quantitative results. It introduces a new evaluation perspective by highlighting visual alignment quality, as illustrated in Fig. 8." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 683, + 425, + 696 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 425, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 425, + 696 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 701, + 485, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 485, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 485, + 713 + ], + "type": "text", + "content": "Comparison of Explainability Methods." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "text", + "content": "53" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 167 + ], + "type": "image", + "image_path": "793b00d420c7a03590beea94a3c4f37f08f790a2f15101a48c6c8fa90eabcef2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 175, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 553, + 198 + ], + "type": "text", + "content": "Figure 3. Visual comparison between TAM and SoTA methods on the COCO Caption dataset [13] using the Qwen2-VL-2B [51] model. The \"is\" in the last row is a functional word with background ground-truth. More examples are shown in Supp. G including complex cases." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 217, + 293, + 383 + ], + "blocks": [ + { + "bbox": [ + 58, + 217, + 293, + 383 + ], + "lines": [ + { + "bbox": [ + 58, + 217, + 293, + 383 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 293, + 383 + ], + "type": "image", + "image_path": "133398b6e279c2171db9f3b72958c2f2decb373681c29d9dc9479c7df46a451c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 392, + 295, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 415 + ], + "type": "text", + "content": "Figure 4. TAM presents high-quality localization results on diverse datasets [13, 42, 57] under the Qwen2-VL-7B [51] model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 437, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 605 + ], + "type": "text", + "content": "We perform comparisons between the proposed TAM and SoTA explanation methods, as illustrated in Fig. 3. The visualization results clearly show that TAM more accurately highlights target objects and renders fewer false activations on function words. In contrast, attention-based methods [1, 2, 5] produce only a limited number of activations, while gradient-based methods [9, 45] present too many correlated responses and noise, performing significantly worse than our approach. Additionally, we present more complex examples in Supp. G, where TAM notably outperforms existing SoTA methods. TAM excels at explaining multiple tokens of MLLMs. In Supp. C, results show how TAM focuses on important tokens, whereas the baseline methods generate many redundant activations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "High-quality Localization Results. We visualized the results using the multimodal activation map in Eq. 3 (see detailed examples in Supp. B). The results demonstrate that TAM effectively locates objects across diverse datasets, as shown in Fig. 4. These findings indicate that TAM enhances MLLMs with localization capabilities, which could be potentially beneficial for various downstream tasks, including segmentation [32], object counting [46], anomaly detection [14], image editing [6], autonomous driving [22], and ap" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 217, + 553, + 388 + ], + "blocks": [ + { + "bbox": [ + 317, + 217, + 553, + 388 + ], + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 388 + ], + "type": "image", + "image_path": "d9944da5bff3f04ff6c48ab4097e93cc9aabd7648e6c5599cd74f2dae88d5aec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 396, + 555, + 419 + ], + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 419 + ], + "type": "text", + "content": "Figure 5. TAM supports to explain and analyze diverse attributes of MLLM (Qwen2-VL-7B). See extensive examples in Supp. M." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 445, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 553, + 468 + ], + "type": "text", + "content": "plications in the medical field [21, 50], which require pixellevel activations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 579 + ], + "type": "text", + "content": "Visualization for Various Attributes. We apply TAM to explain various attributes of MLLMs, specifically focusing on fine-grained token types as Fig. 5. The results indicate that TAM effectively supports the explanation of attributes such as color, shape, action, number, location, and text. Notably, the representation of numbers is less pronounced compared to object localization, suggesting that some attributes may be encoded within related object tokens. In contrast, other attributes are more prominently featured." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "Explaining Failure Cases. We apply TAM to analyze failure cases in the QK-VQA dataset [37] as Fig. 7. Our findings indicate that while the model successfully locates objects, it struggles to align them with specific knowledge. For instance, the model can identify the desert and focus on the key prompt (\"year\") when outputting \"don't know.\" It suggests a misalignment between the desert and the year, rather than recognition or prompt issues. Besides, some responses consist of synonyms, hyphenms, or hyponyms of the correct answers, resulting in mismatches. For more case studies, please refer to the catalog in Supp. Table 5." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 749, + 310, + 757 + ], + "type": "text", + "content": "54" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 304, + 186 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 304, + 186 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 304, + 186 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 304, + 186 + ], + "type": "image", + "image_path": "3d01dc3e24f8cd68c76f5ddebf2c7295eb8854fd87679b8cbce63c576eb3ca3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 197, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 197, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 555, + 220 + ], + "type": "text", + "content": "Figure 6. TAM significantly improves the quality of video visualization, with much fewer redundant activation and noises on STAR [52] dataset using Qwen2-VL-2B [51]. Please see extensive examples in Supp. 27 and failure cases analysis for videos in Fig. 25." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 553, + 187 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 553, + 187 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 553, + 187 + ], + "type": "image", + "image_path": "5a41589abb588a35c283540f42a091f30434ad52b580b37a4735ba4fd9b9f6ac.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 238, + 137, + 325 + ], + "blocks": [ + { + "bbox": [ + 59, + 238, + 137, + 325 + ], + "lines": [ + { + "bbox": [ + 59, + 238, + 137, + 325 + ], + "spans": [ + { + "bbox": [ + 59, + 238, + 137, + 325 + ], + "type": "image", + "image_path": "fceab8ba0355ed65903104963968dc5153bae6d2e3ce952854f51147e2d0fb26.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 333, + 295, + 367 + ], + "lines": [ + { + "bbox": [ + 55, + 333, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 295, + 367 + ], + "type": "text", + "content": "Figure 7. TAM supports failure cases analysis on QK-VQA [37]. See more cases in Supp. about VQA (Fig. 24), videos (Fig. 25), visual reasoning (Fig. 28), and multi-turn conversation (Fig. 31)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 137, + 238, + 215, + 325 + ], + "blocks": [ + { + "bbox": [ + 137, + 238, + 215, + 325 + ], + "lines": [ + { + "bbox": [ + 137, + 238, + 215, + 325 + ], + "spans": [ + { + "bbox": [ + 137, + 238, + 215, + 325 + ], + "type": "image", + "image_path": "587602b470b3cd0bd90a00cda47eaad67cf026e08726ddf9ea733c09f803aa53.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 215, + 238, + 292, + 325 + ], + "blocks": [ + { + "bbox": [ + 215, + 238, + 292, + 325 + ], + "lines": [ + { + "bbox": [ + 215, + 238, + 292, + 325 + ], + "spans": [ + { + "bbox": [ + 215, + 238, + 292, + 325 + ], + "type": "image", + "image_path": "8aa8054cb80577901dddab1e8628e4d104752e90d969668eed0cd7bced221f34.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 381, + 293, + 536 + ], + "blocks": [ + { + "bbox": [ + 58, + 381, + 293, + 536 + ], + "lines": [ + { + "bbox": [ + 58, + 381, + 293, + 536 + ], + "spans": [ + { + "bbox": [ + 58, + 381, + 293, + 536 + ], + "type": "image", + "image_path": "dbe630a86dc5d791fac0cae47d5996d220b75f4366ae3655691154b0df768c24.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 544, + 295, + 569 + ], + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 569 + ], + "type": "text", + "content": "Figure 8. TAM enables qualitative comparison among MLLMs. Extensive examples are given in Fig. 17 and Fig. 18 in Supp. K." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 591, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 295, + 652 + ], + "type": "text", + "content": "Video Visualization. The proposed TAM excels at video visualization, where there are much fewer redundant activations and noises as shown in Fig. 6. Our method provides very clear results, while the conventional CAM makes it hard to see the video content without obvious highlights." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "A Tool to Visually Comparison MLLMs. TAM can be used as a tool to compare MLLMs qualitatively. If an MLLM exhibits superior activation maps, it can be inferred that its explainability and degree of vision alignment are also better, offering valuable pixel-level predictions at no" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 239, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 324 + ], + "type": "text", + "content": "additional cost. As shown in Fig. 8, Qwen2-VL-7B [51] shows greater explainability than LLaVA1_5-7B [35] and InternVL2_5-8B [16]. Furthermore, we observe that InternVL tends to focus more on text, resulting in weaker activation levels; in some cases, this performance is even better than that of LLaVA (e.g., with the broccoli). More extensive examples can be found in Supp. K." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 335, + 388, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 388, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 388, + 346 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 354, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 474 + ], + "type": "text", + "content": "This work focuses on the unique nature of MLLMs in visual explainability, which generates multiple tokens progressively, complicating the explanation process. In response, we introduced the Token Activation Map (TAM), a novel method that utilizes estimated causal inference to mitigate interferences from context tokens, with a rank Gaussian filter to reduce activation noise, thereby providing clearer visual explanations. Our results demonstrate that TAM significantly outperforms existing SoTA methods, offering high-quality visualizations applicable to diverse scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 474, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 582 + ], + "type": "text", + "content": "Despite the success, we focus on visual inputs, other modalities, such as audio, remain under-explored. Additionally, interpreting model decisions is an extensible further aspect. The potential applications of our TAM method are extensive, including open-vocabulary segmentation, detection, image grounding, anomaly detection, remote sensing, and medical fields that require pixel-level activations. Our work lays a foundation for advancing their explainability and practical utility across various domains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 593, + 408, + 606 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 408, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 408, + 606 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 613, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 555, + 696 + ], + "type": "text", + "content": "This work was supported by a research grant from the Joint Research Scheme (JRS) under the National Natural Science Foundation of China (NSFC) and the Research Grants Council (RGC) of Hong Kong (Project No. N_HKUST654/24), as well as a grant from the RGC of the Hong Kong Special Administrative Region, China (Project No. R6005-24)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "55" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4190-4197, 2020. 2, 5, 6, 7, 14, 15" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 202 + ], + "type": "text", + "content": "[2] Reduan Achtibat, Sayed Mohammad Vakilzadeh Hatefi, Maximilian Dreyer, Aakriti Jain, Thomas Wiegand, Sebastian Lapuschkin, and Wojciech Samek. Attnlrp: Attention-aware layer-wise relevance propagation for transformers. In International Conference on Machine Learning, pages 135-168. PMLR, 2024. 2, 5, 6, 7, 14, 15" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "type": "text", + "content": "[3] Estelle Aflalo, Meng Du, Shao-Yen Tseng, Yongfei Liu, Chenfei Wu, Nan Duan, and Vasudev Lal. Vl-interpret: An interactive visualization tool for interpreting vision-language transformers. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 21406–21415, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 272, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 295, + 337 + ], + "type": "text", + "content": "[4] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 339, + 295, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 339, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 339, + 295, + 395 + ], + "type": "text", + "content": "[5] Ameen Ali, Thomas Schnake, Oliver Eberle, Grégoire Montavon, Klaus-Robert Müller, and Lior Wolf. Xai for transformers: Better explanations through conservative propagation. In International Conference on Machine Learning, pages 435-451. PMLR, 2022. 6, 7, 14, 15" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 396, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 295, + 441 + ], + "type": "text", + "content": "[6] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 442, + 296, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 296, + 519 + ], + "type": "text", + "content": "[7] Gabriela Ben Melech Stan, Estelle Aflalo, Raanan Yehezkel Rohekar, Anahita Bhiwandiwalla, Shao-Yen Tseng, Matthew Lyle Olson, Yaniv Gurwicz, Chenfei Wu, Nan Duan, and Vasudev Lal. Lvlm-intrepret: An interpretability tool for large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8182-8187, 2024. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 521, + 295, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 295, + 565 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 295, + 565 + ], + "type": "text", + "content": "[8] Chin-Chen Chang, Ju-Yuan Hsiao, and Chih-Ping Hsieh. An adaptive median filter for image denoising. In 2008 Second international symposium on intelligent information technology application, pages 346-350. IEEE, 2008. 3, 5, 6, 16" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 567, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 295, + 632 + ], + "type": "text", + "content": "[9] Aditya Chattopadhyay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE winter conference on applications of computer vision (WACV), pages 839-847. IEEE, 2018. 2, 6, 7, 14, 15" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[10] Hila Chefer, Shir Gur, and Lior Wolf. Generic attention-model explainability for interpreting bi-modal and encoder-decoder transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 397-406, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[11] Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "the IEEE/CVF conference on computer vision and pattern recognition, pages 782-791, 2021. 2, 5, 15" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[12] Ruoyu Chen, Hua Zhang, Siyuan Liang, Jingzhi Li, and Xiaochun Cao. Less is more: Fewer interpretable region via submodular subset selection. In The Twelfth International Conference on Learning Representations. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "text", + "content": "[13] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco captions: Data collection and evaluation server. arXiv preprint arXiv:1504.00325, 2015. 2, 5, 6, 7, 13, 16, 17, 20, 21, 22, 23, 24, 26, 29, 30" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 195, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 258 + ], + "type": "text", + "content": "[14] Xuhai Chen, Jiangning Zhang, Guanzhong Tian, Haoyang He, Wuhao Zhang, Yabiao Wang, Chengjie Wang, and Yong Liu. Clip-ad: A language-guided staged dual-path model for zero-shot anomaly detection. In International Joint Conference on Artificial Intelligence, pages 17-33. Springer, 2024. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[15] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 2, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 317, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 553, + 392 + ], + "type": "text", + "content": "[16] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024. 1, 6, 8, 16, 17, 19, 25, 35" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 393, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 393, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 393, + 553, + 458 + ], + "type": "text", + "content": "[17] Can Cui, Yunsheng Ma, Xu Cao, Wenqian Ye, Yang Zhou, Kaizhao Liang, Jintai Chen, Juanwu Lu, Zichong Yang, Kuei-Da Liao, et al. A survey on multimodal large language models for autonomous driving. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 958-979, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "type": "text", + "content": "[18] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 504, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 547 + ], + "type": "text", + "content": "[19] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In The Twelfth International Conference on Learning Representations. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 548, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 624 + ], + "type": "text", + "content": "[20] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 625, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 680 + ], + "type": "text", + "content": "[21] Xinpeng Ding, Yongqiang Chu, Renjie Pi, Hualiang Wang, and Xiaomeng Li. Hia: Towards chinese multimodal llms for comparative high-resolution joint diagnosis. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 575-586. Springer, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[22] Xinpeng Ding, Jianhua Han, Hang Xu, Xiaodan Liang, Wei Zhang, and Xiaomeng Li. Holistic autonomous driving understanding by bird's view injected multi-modal large models." + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "56" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13668-13677, 2024. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[23] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "text", + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[25] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[26] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, Ming-Ming Cheng, and Yunchao Wei. Layercam: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875-5888, 2021. 2, 6, 14" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "text", + "content": "[27] Licheng Jiao, Yuhan Wang, Xu Liu, Lingling Li, Fang Liu, Wenping Ma, Yuwei Guo, Puhua Chen, Shuyuan Yang, and Biao Hou. Causal inference meets deep learning: A comprehensive survey. Research, 7:0467, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 398 + ], + "type": "text", + "content": "[28] Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Grégoire Montavon, Wojciech Samek, and Klaus-Robert Müller. Unmasking clever hans predictors and assessing what machines really learn. Nature communications, 10(1):1096, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 464 + ], + "type": "text", + "content": "[29] Unggi Lee, Minji Jeon, Yunseo Lee, Gyuri Byun, Yoorim Son, Jaeyoon Shin, Hongkyu Ko, and Hyeoncheol Kim. Llava-docent: Instruction tuning with multimodal large language model to support art appreciation education. Computers and Education: Artificial Intelligence, 7:100297, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 533 + ], + "type": "text", + "content": "[30] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Maniplm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18061-18070, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 577 + ], + "type": "text", + "content": "[31] Yi Li, Hualiang Wang, Yiqun Duan, Hang Xu, and Xiaomeng Li. Exploring visual interpretability for contrastive language-image pre-training. arXiv preprint arXiv:2209.07046, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "text", + "content": "[32] Yi Li, Hualiang Wang, Yiqun Duan, Jiheng Zhang, and Xiaomeng Li. A closer look at the explainability of contrastive language-image pre-training. Pattern Recognition, page 111409, 2025. 2, 3, 4, 5, 6, 7, 14, 15, 22" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[33] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[34] Fenglin Liu, Tingting Zhu, Xian Wu, Bang Yang, Chenyu You, Chenyang Wang, Lei Lu, Zhangdaihong Liu, Yefeng" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "Zheng, Xu Sun, et al. A medical multimodal large language model for future pandemics. NPJ Digital Medicine, 6(1): 226, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 106, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 139 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3, 6, 8, 13, 16, 17, 25" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "text", + "content": "[36] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. Advances in neural information processing systems, 30, 2017. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 227 + ], + "type": "text", + "content": "[37] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 5, 7, 8, 14, 18, 31, 33" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 281 + ], + "type": "text", + "content": "[38] Daniel Omeiza, Skyler Speakman, Celia Cintas, and Komminist Weldermariam. Smooth grad-cam++: An enhanced inference level visualization technique for deep convolutional neural network models. arXiv preprint arXiv:1908.01224, 2019. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 283, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 338 + ], + "type": "text", + "content": "[39] Vedant Palit, Rohan Pandey, Aryaman Arora, and Paul Pu Liang. Towards vision-language mechanistic interpretability: A causal tracing tool for blip. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2856-2861, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 358 + ], + "type": "text", + "content": "[40] Judea Pearl. Causality. Cambridge university press, 2009. 2,3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 425 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 22" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 426, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 502 + ], + "type": "text", + "content": "[42] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009-13018, 2024. 5, 6, 7, 13, 16, 17" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 504, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 558 + ], + "type": "text", + "content": "[43] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should i trust you?\" explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, pages 1135-1144, 2016. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "text", + "content": "[44] Raanan Y Rohekar, Yaniv Gurwicz, and Shami Nisimov. Causal interpretation of self-attention in pre-trained transformers. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 604, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 668 + ], + "type": "text", + "content": "[45] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 1, 2, 6, 7, 14, 15, 16, 17, 18, 20, 24, 30, 34" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[46] Zenglin Shi, Ying Sun, and Mengmi Zhang. Training-free object counting with prompts. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 323-331, 2024. 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 749, + 310, + 757 + ], + "type": "text", + "content": "57" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 640 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[47] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 150 + ], + "type": "text", + "content": "[48] Joseph Tey. Understanding how vision-language models reason when solving visual math problems. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 174 + ], + "type": "text", + "content": "[49] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "text", + "content": "[50] Lehan Wang, Haonan Wang, Honglong Yang, Jiaji Mao, Zehong Yang, Jun Shen, and Xiaomeng Li. Interpretable bilingual multimodal large language model for diverse biomedical tasks. In The Thirteenth International Conference on Learning Representations, 2025. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 306 + ], + "type": "text", + "content": "[51] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 1, 2, 5, 6, 7, 8, 13, 14, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "type": "text", + "content": "[52] Bo Wu, Shoubin Yu, Zhenfang Chen, Joshua B Tenenbaum, and Chuang Gan. Star: A benchmark for situated reasoning in real-world videos. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2). 2, 5, 8, 14, 18, 29, 32, 34" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 407 + ], + "type": "text", + "content": "[53] Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, and Mike Lewis. Efficient streaming language models with attention sinks. In The Twelfth International Conference on Learning Representations. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 453 + ], + "type": "text", + "content": "[54] Liuyi Yao, Zhixuan Chu, Sheng Li, Yaliang Li, Jing Gao, and Aidong Zhang. A survey on causal inference. ACM Transactions on Knowledge Discovery from Data (TKDD), 15(5):1-46, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 529 + ], + "type": "text", + "content": "[55] Xiaofeng Zhang, Yihao Quan, Chen Shen, Xiaosong Yuan, Shaotian Yan, Liang Xie, Wenxiao Wang, Chaochen Gu, Hao Tang, and Jieping Ye. From redundancy to relevance: Enhancing explainability in multimodal large language models. Annual Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics, 2025. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 532, + 294, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 532, + 294, + 585 + ], + "spans": [ + { + "bbox": [ + 56, + 532, + 294, + 585 + ], + "type": "text", + "content": "[56] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2921-2929, 2016. 2, 6, 16, 17, 18, 20, 24, 30, 34" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 587, + 294, + 640 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 294, + 640 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 294, + 640 + ], + "type": "text", + "content": "[57] Zijian Zhou, Zheng Zhu, Holger Caesar, and Miaojing Shi. Openpsg: Open-set panoptic scene graph generation via large multimodal models. In European Conference on Computer Vision, pages 199-215. Springer, 2024. 2, 5, 6, 7, 14, 16, 17" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 300, + 750, + 310, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 300, + 750, + 310, + 757 + ], + "spans": [ + { + "bbox": [ + 300, + 750, + 310, + 757 + ], + "type": "text", + "content": "58" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/b324faad-4fe9-44a7-a833-a82bac07835f_content_list.json b/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/b324faad-4fe9-44a7-a833-a82bac07835f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..33ff1345e7028eba4a03d4f98ec9391304f2198b --- /dev/null +++ b/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/b324faad-4fe9-44a7-a833-a82bac07835f_content_list.json @@ -0,0 +1,1900 @@ +[ + { + "type": "text", + "text": "Token-Efficient VLM: High-Resolution Image Understanding via Dynamic Region Proposal", + "text_level": 1, + "bbox": [ + 122, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yitong Jiang $^{1,2,\\dagger}$ Jinwei Gu $^{1,2}$ Tianfan Xue $^{2}$ Ka Chun Cheung $^{1}$ Pavlo Molchanov $^{1}$ Hongxu Yin $^{1\\dagger}$ Sifei Liu $^{1\\dagger}$ $^{1}$ NVIDIA ${}^{2}$ The Chinese University of Hong Kong", + "bbox": [ + 148, + 200, + 849, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b9f8094134eef96cdae7f92e7dc7b7351220f5e5b3d420f27fb60a273388fe58.jpg", + "image_caption": [ + "Figure 1. Illustration of the Token-Efficient Vision LLanguage Model (TEVA) for high-resolution scenarios. Given a user query about a specific detail $(20\\times 50\\mathrm{px}$ region) in a high-resolution image $(5760\\times 2880\\mathrm{px})$ , traditional VLMs (Left) often miss critical details due to uniform patch processing and computational limits. TEVA (Right) addresses this by using a Relevant Area Proposal module to identify key subjects (e.g., people) and Dynamic Patch Sampling to focus on high-resolution patches of interest while retaining global context patches. This approach enables accurate interpretation of details at arbitrary resolutions without increasing the token count." + ], + "image_footnote": [], + "bbox": [ + 91, + 279, + 372, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/76caf698d7b59308f0c8fd33c6975bf48e88287e18735ba1e06e37cd2681f745.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 287, + 563, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c451e37662a929f2eb525e829632f8fbaea2e883ee094318beda32abcd874f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 280, + 771, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ca2907859b78258310a8467c1209cce64ff9587719301615d11edaa65b2782c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 280, + 898, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 513, + 326, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) excel at visual understanding by leveraging pretrained image encoders to generate visual tokens. However, they struggle with high-resolution images and zoomed-in regions due to the computational burden and token redundancy of uniform patch-based processing, often leading to the loss of critical details. To address these challenges, we propose Token-Efficient Vision LLanguage Model (TEVA), a novel framework that detects key regions and applies dynamic patch sampling to efficiently capture fine-grained details while preserving global context. Our approach first identifies subject-oriented regions using an adaptive detection strategy. Then, a dynamic patch sampling mechanism selects and arranges patches at varying scales, ensuring efficient processing without increasing token count. Extensive experiments demonstrate that Token-Efficient Vision LLanguage Model (TEVA) significantly enhances VLM performance in handling visual details, seamlessly integrating with various decoders and LLMs.", + "bbox": [ + 88, + 547, + 485, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 513, + 643, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language Models (VLMs) [1, 36, 37, 80] have recently received significant attention due to their ability to integrate visual understanding into Large Language Models (LLMs) [50, 59, 77], enhancing performance in tasks like image captioning and visual question answering. However, supporting applications that require precise attention to specific details—such as video surveillance analysis, satellite imagery interpretation, and scene OCR [49]—remains challenging. These tasks often necessitate zooming into fine-grained regions of high-resolution images to capture essential information, but many existing VLMs [1, 28, 37, 80] rely on low-resolution pretrained image encoders (e.g., $224 \\times 224$ pixels), leading to the loss of important details [38, 52].", + "bbox": [ + 511, + 545, + 906, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this issue, previous methods have focused on either increasing global resolution or applying region-aware techniques. Enhancing resolution in pretrained image encoders typically involves two strategies: scaling resolution via curriculum learning [2, 5, 73] and partitioning images into fixed-size sub-images for independent processing [9, 33, 35, 36, 38, 42, 52]. These approaches have notable limitations. Scaling resolution increases computational complexity quadratically, raising self-attention", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Yitong Jiang was an intern at NVIDIA during the project. \n†Equal Advising.", + "bbox": [ + 107, + 873, + 418, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24147", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "costs. Partitioning images boosts the number of visual tokens (e.g., over 5000 tokens for an $896 \\times 896$ image [31]), leading to redundancy and degraded performance [35, 38]. Moreover, the lack of interaction among image tokens within the vision encoder limits the model's ability to capture contextual relationships, requiring specialized design interventions [31, 38].", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Alternatively, region-aware methods [6, 15, 66, 72, 73] focus on fine-grained details by leveraging manually defined Regions of Interest (ROIs). However, these methods either compress visual tokens within target regions, potentially losing critical information [6, 15, 66], or require multiple independent passes through the vision encoder, significantly increasing computational cost and token usage [72, 73].", + "bbox": [ + 89, + 199, + 480, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose Token-Efficient Vision LLanguage Model (TEVA), which leverages a dynamic patch sampling framework for the vision encoder that balances visual token count and image resolution. Our primary concept involves informative patch sampling within the vision encoder, which selectively focuses on key parts of the image while preserving the understanding of the global context, represented in low resolution. As shown in Figure 1, unlike conventional vision encoders that uniformly partition the entire image into patches, our method makes an early determination of which image regions are essential for answering a given question and applies selective patch sampling before inputting patches into the vision encoder. This approach not only prevents an increase in the number of patches when processing high-resolution images of any size, but also reduces self-attention computation and minimizes tokens the LLM needs to process. Additionally, this approach maintains coherence among sampled image regions, enhancing the model's ability to understand the global context and structure of the input image [63, 67].", + "bbox": [ + 91, + 321, + 482, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our TEVA consists of three main steps. First, the Relevant Area Proposal (RAP) module identifies key locations and scales of interest using a subject proposal agent powered by an LLM, and a subject detection agent via a vision foundation model. This enables efficient sampling of informative patches as vision tokens. Next, patches from localized regions are arranged with low-resolution global patches, integrating positional information through an affine transformation strategy. This allows for sampling patches of varying sizes and locations without additional tokens or complex architectures [8, 31, 38, 64]. Finally, the mixture of local and global patch embeddings is processed in the vision encoder (SigLIP [71] in our work) to extract visual tokens for the Large Language Models (LLMs).", + "bbox": [ + 89, + 626, + 482, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, to enhance dynamic patch sampling for VLMs, we introduce a data curation pipeline and curate a Subject-Oriented (SO) dataset containing 8 million subject-focused conversations. This is the first large-scale Subject", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Oriented VQA dataset for VLM pretraining, enabling a framework that focuses on specific image details. By integrating dynamic image tokens with early interactions between local and global visual elements, Token-Efficient Vision LLanguage Model (TEVA) achieves superior VLM performance across various benchmarks, particularly excelling in flexible-resolution scenarios, as shown in our comparative analysis in Section 4.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 224, + 653, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies [58, 70, 78] highlight a bottleneck in Vision-Language Models (VLMs), where image resolution constraints limit the model's ability to capture fine-grained visual details. Previous approaches of handling high-resolution images in Vision-Language Models (VLMs) can be grouped into two main strategies: (1) increasing the global image resolution, (2) applying region-aware techniques, as discussed in the following.", + "bbox": [ + 511, + 250, + 903, + 371 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the first strategy, methods like Qwen-VL [2] and Griffon v2 [73] employ curriculum learning and bilinear interpolation to gradually increase the resolution in the pretrained image encoder. However, a significant challenge with transformer-based image encoders [10] is their reliance on uniform patch sampling, which leads to a quadratic increase in computational complexity as image resolution grows. Other methods [30, 31, 33, 38, 64, 76] partition high-resolution images into smaller sub-images, treating each as an independent input to the vision encoder. While this approach enables the processing of higher-resolution images, it imposes a constraint on the input sequence length that the Large Language Model (LLM) can handle, thereby limiting its overall capacity. To mitigate the computational burden, learnable queries within the Perceiver architecture [31, 38] and pooling strategies [52, 73] have been introduced to reduce the number of visual tokens generated by the vision encoder from these sub-images. However, such compression techniques often lack selectivity, leading to the loss of fine-grained visual details that are crucial for optimal Multimodal Large Language Model (MLLM) performance.", + "bbox": [ + 511, + 372, + 903, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The second strategy involves explicitly identifying and processing Regions-of-Interest (ROIs). Some methods [6, 15, 66, 74] require manual specification of ROIs, followed by compression of low-resolution visual tokens. Others [72, 73] necessitate multiple passes through the vision encoder, increasing computational overhead. $V^{*}$ [61] further extends this idea by iteratively searching for ROIs through multiple passes through an additional Vision-Language Model (VLM). However, this iterative search is computationally expensive and also requires multiple forward passes through the vision encoder for each ROI candidate.", + "bbox": [ + 511, + 688, + 903, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to these methods, our approach automatically identifies and focuses on informative regions within high-resolution images. By applying dynamic patch sam", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24148", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/631ab1c885928cca65fbfb51c07bd31086577eeca79d6879ff5df3ffe108b651.jpg", + "image_caption": [ + "Figure 2. The TEVA architecture. Given a user query, the Subject Proposal Agent identifies relevant subjects (e.g., \"signs\"). The Subject Detection Agent then locates all candidate signs in the image. Informative patch sampling extracts high-resolution local patches at these locations, which are concatenated with low-resolution global patches and fed into the VLM for final interpretation." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 903, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pling before the vision encoder, we preserve essential details without resorting to compression, while also minimizing computational costs. This is achieved by skipping less informative patches and avoiding multiple passes through the vision encoder, resulting in a more efficient and detailed visual representation for VLMs.", + "bbox": [ + 89, + 321, + 483, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 422, + 225, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We show the overall flowchart of our proposed TokenEfficient Vision LLanguage Model (TEVA) in Figure 2. TEVA is a VLM framework designed to handle arbitrary high-resolution images by detecting and focusing on informative visual details within images. It incorporates a Relevant Area Proposal (RAP) module to identify and extract key regions from the input image. A targeted patch sampling strategy then partitions these selected areas into informative high-resolution local patches, while retaining low-resolution global patches for context. The concatenation of local and global patches is processed by the image encoder to generate visual tokens. Additionally, we introduce a Subject-Oriented (SO) dataset to enhance the VLM's capability to manage dynamic patch sampling and improve overall performance. Details are described below.", + "bbox": [ + 88, + 448, + 482, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Relevant Area Proposal Module", + "text_level": 1, + "bbox": [ + 89, + 681, + 372, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on our insight that image tokens are essential for Vision-Language Models (VLMs) to generate accurate content descriptions and answer questions, we address a key challenge: as shown in Figure 3 (left), when VLMs process uniformly partitioned image patches, they often focus on prominent elements, such as the path, trees, and signs on the trees, while overlooking finer details, like the presence of two individuals in the forest.", + "bbox": [ + 89, + 704, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome this, we propose an automatic Relevant Area Proposal (RAP) module to capture key local details that help the VLM focus on the most informative and relevant areas in response to user queries. Current open-vocabulary detection methods, such as Grounding", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a9073fbb70f73f59809c1fdb59add3299aebaba28c6e8ff54eab95d2e6e84d60.jpg", + "image_caption": [ + "What can you see in this image?", + "Assistant: A forest with tall trees and a trail in the middle.", + "Figure 3. Example of visual focus in Vision-Language Models (VLMs) with uniform image patches (Left) versus high-resolution, relevant patches (Right). In the uniform approach, the model emphasizes broad features like trees, trails, and visible signs, missing subtler details, such as the presence of people in the scene. By providing high-resolution relevant patches, the model can identify and interpret finer details." + ], + "image_footnote": [], + "bbox": [ + 516, + 335, + 705, + 438 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8111a0443b3cb485999879f5b82792f2b6ddf61921e925c2c9a09aa417fd0f46.jpg", + "image_caption": [ + "Assistant: A group of people are walking through a forest." + ], + "image_footnote": [], + "bbox": [ + 715, + 335, + 903, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dino [39], have shown strong performance in detecting regions of interest based on straightforward text cues. However, user queries often involve complex questions or instructions without an explicit subject, which limits the applicability of open-vocabulary detection.", + "bbox": [ + 511, + 597, + 905, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this, we use two specialized agents: a Subject-Promise Agent using LLM, e.g., Llama3.1 [59], and a Subject Detection Agent via Grounding Dino. Given a user query, we employ a structured prompt (e.g., \"You have limited vision of the image; tell me the subjects you need to observe to answer the question + User query\") to guide the LLM in proposing candidate subjects for detection. Once identified, the Subject Detection Agent then locates the regions of interest (ROI) based on these suggested subjects.", + "bbox": [ + 511, + 674, + 905, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While this process effectively isolates the ROIs, integrating these fine-grained local details into the VLM remains challenging. To this end, we introduce the Informative Patch Sampling module in the following, which efficiently incorporates high-resolution information without increasing the token count or computational burden.", + "bbox": [ + 511, + 810, + 905, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24149", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Informative Patch Sampling", + "text_level": 1, + "bbox": [ + 89, + 90, + 344, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous region-level VLMs [15, 66] handle regions of interest by relying on low-resolution features and apply pooling strategies to compress information within each region. Although effective for reducing computational load, these pooling methods often lead to a loss of critical details, especially for high-resolution images. Alternative approaches [61, 72, 73] treat each target region as a separate image that passes through a vision tower. However, this approach is computationally expensive and significantly increases the number of image tokens and inference time. To better leverage high-resolution image information without incurring excessive computational costs, we propose a novel approach that utilizes informative patch sampling within each high-resolution target region. This technique preserves fine-grained details while maintaining computational efficiency and the same token count comparable to a low-resolution image encoder, as demonstrated in Table 1.", + "bbox": [ + 89, + 112, + 483, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Patch allocation. As shown in Algorithm 1, our method begins by allocating informative patch numbers to regions of interest based on their relative sizes. Next, we align the assigned patches within each bounding box, ensuring even coverage across each target area. Finally, we remove overlapping bounding boxes to maintain a total of $n$ , yielding the positions $P \\in \\mathbb{R}^{n \\times 2}$ and scaling factors $S \\in \\mathbb{R}^{n \\times 2}$ .", + "bbox": [ + 89, + 387, + 483, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Patch sampling. Given the positions and the scaling factors, to link these positional clues to exact image patches, we embed $\\mathsf{P} = \\{(p_i^x,p_i^y)|i\\in (0,n - 1)\\}$ and $S = \\{(s_i^x,s_i^y)|i\\in (0,n - 1)\\}$ into an affine transformation matrix to project image pixels into patches:", + "bbox": [ + 89, + 510, + 483, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA = \\left[ \\begin{array}{c c c} s _ {i} ^ {x} & 0 & p _ {i} ^ {x} \\\\ 0 & s _ {i} ^ {y} & p _ {i} ^ {y} \\end{array} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 593, + 482, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With this setup, the correspondence between spatial coordinates of patches and input image can now be defined as:", + "bbox": [ + 89, + 633, + 482, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc = A \\left[ \\begin{array}{l l l} u & v & 1 \\end{array} \\right] ^ {T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 671, + 482, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c = [c_u, c_v]$ are the spatial coordinates in the original image, while $u$ and $v$ are coordinates in the patch space, sampled on a uniform grid. Through sampling patch spatial coordinates with Eq. 2 we can determine the corresponding coordinates in the original image. Since $c$ is not discrete, proper interpolation is necessary. Inspired by Spatial Transformer Networks [19], we employ bilinear interpolation to address this:", + "bbox": [ + 89, + 700, + 483, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} c _ {u, v} = \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} I _ {(h, w)} \\max (0, 1 - | c _ {v} - h |) \\tag {3} \\\\ \\times \\max (0, 1 - | c _ {u} - w |), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 840, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{(h,w)}$ represents the pixel value in the original image $I$ of size $H \\times W$ at the location $(h,w)$ . For a patch of size $\\hat{H} \\times \\hat{W}$ , the coordinates are sampled as $u = \\{1,2,\\dots,\\hat{H}\\}$ ; $v = \\{1,2,\\dots,\\hat{W}\\}$ . Unlike the original integration-based approach [19], which is time-consuming for high-resolution tasks, we use a look-up operation to accelerate by considering only the four neighboring pixels around each patch coordinate.", + "bbox": [ + 511, + 90, + 906, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To incorporate spatial information for each patch, we concatenate the affine transformation matrix embeddings with each patch's positional embedding. Experimental results in Table 4 demonstrate the effectiveness of this approach in enhancing the model's spatial awareness.", + "bbox": [ + 511, + 212, + 906, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 Simplified Informative Patch Sampling", + "text_level": 1, + "bbox": [ + 514, + 321, + 879, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input:", + "bbox": [ + 514, + 338, + 553, + 349 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "Bounding boxes $\\mathsf{B} = \\{b_1,\\dots ,b_{\\mathrm{box\\_num}}\\} \\in \\mathbb{R}^{\\mathrm{box\\_num}\\times 4},$ Each bounding box $b_{j} = [x_{min},y_{min},x_{max},y_{max}],$ Number of local patches $n$", + "guess_lang": "txt", + "bbox": [ + 522, + 349, + 867, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output:", + "bbox": [ + 514, + 386, + 562, + 397 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "Sampled patch locations $\\mathsf{P}\\in \\mathbb{R}^{n\\times 2}$ Sampled patch scaling factors $\\mathsf{S}\\in \\mathbb{R}^{n\\times 2}$", + "guess_lang": "latex", + "bbox": [ + 522, + 398, + 823, + 421 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "1. Allocate patches Based on Area area $(\\mathsf{b_j}) = (\\mathbf{x}_{\\mathrm{max}} - \\mathbf{x}_{\\mathrm{min}})(\\mathbf{y}_{\\mathrm{max}} - \\mathbf{y}_{\\mathrm{min}})$ $\\forall \\mathsf{b_j}\\in \\mathsf{B}$ n=j = num-patches(bj) = area(bj) n, $\\forall \\mathsf{b_j}\\in \\mathsf{B}$", + "guess_lang": "latex", + "bbox": [ + 514, + 426, + 836, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2. Generate patch locations and scales for $\\mathbf{b}_{\\mathrm{j}}\\in \\mathsf{B}$ do", + "bbox": [ + 514, + 472, + 838, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\cdot \\operatorname {w i d t h} (\\mathfrak {b} _ {\\mathrm {j}}) = \\mathbf {x} _ {\\max } - \\mathbf {x} _ {\\min }, \\operatorname {h e i g h t} (\\mathfrak {b} _ {\\mathrm {j}}) = \\mathbf {y} _ {\\max } - \\mathbf {y} _ {\\min } \\\\ \\cdot \\text {p a t c h e s} _ {\\mathrm {w}}. \\text {s i d e} = \\sqrt {\\frac {\\operatorname {n u m} . \\operatorname {p a t c h e s} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right) \\cdot \\operatorname {w i d t h} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right)}{\\operatorname {h e i g h t} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right)}} \\\\ \\cdot \\text {p a t c h e s} _ {\\mathrm {h}} \\text {s i d e} = \\sqrt {\\frac {\\operatorname {n u m} . \\operatorname {p a t c h e s} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right) \\cdot \\operatorname {h e i g h t} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right)}{\\operatorname {w i d t h} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right)}} \\\\ \\bullet \\mathrm {x} = \\operatorname {l i n s p a c e} \\left(\\mathbf {x} _ {\\min }, \\mathbf {x} _ {\\max }, \\text {p a t c h e s - w - s i d e}\\right) \\\\ \\cdot \\mathbf {y} = \\operatorname {l i n s p a c e} \\left(\\mathbf {y} _ {\\min }, \\mathbf {y} _ {\\max }, \\text {p a t c h e s} _ {\\mathrm {h}} \\text {s i d e}\\right) \\\\ \\cdot P _ {j} = \\text {m e s h g r i d} (x, y) \\\\ \\bullet \\mathbf {S} _ {\\mathrm {j}} = \\left(\\frac {\\operatorname {w i d t h} (\\mathbf {b} _ {\\mathrm {j}})}{\\text {p a t c h e s . w . s i d e}}, \\frac {\\operatorname {h e i g h t} (\\mathbf {b} _ {\\mathrm {j}})}{\\text {p a t c h e s . h . s i d e}}\\right). \\exp a n d (\\mathbf {n} _ {\\mathrm {j}}, 2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 498, + 880, + 617 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end", + "bbox": [ + 514, + 619, + 537, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Remove redundant patches", + "bbox": [ + 514, + 636, + 751, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "P,S $\\leftarrow$ Remove_overlapedpatches(Concat(Pj),Concat(Sj))", + "bbox": [ + 522, + 648, + 893, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "return P and S", + "bbox": [ + 514, + 661, + 599, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Subject-Oriented Data Curation", + "text_level": 1, + "bbox": [ + 511, + 696, + 797, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To fine-tune our TEVA model for dynamically sampled patches, we introduce a data curation pipeline that creates subject-oriented training data, emphasizing specific regions of interest within images. While large datasets [3, 51] are commonly used to enhance the generalization of VLMs, they typically consist of image-caption pairs providing broad, whole-image descriptions. Our approach curates question-answer pairs that emphasize zoomed-in regions, enabling the model to better understand and prioritize detailed image areas. To address this, we developed an efficient method that curates large-scale, subject-oriented conversations (8 million pairs) based on existing image-caption", + "bbox": [ + 511, + 719, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24150", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/42471df3e5794d0cd39d3db5b0ef00738f654aa5888678eb70205e736b071bda.jpg", + "image_caption": [ + "The image is only for illustration. The images for subject-oriented VCA generation." + ], + "image_footnote": [], + "bbox": [ + 102, + 95, + 194, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Caption: The image portrays a fairy with white wings and a purple dress, standing in a field of lavender flowers. She is holding a lavender plant in her hands. The background is a light blue color, and there are two butterflies visible in the scene. The fairy is positioned in the center of the image, with the lavender plants surrounding her. The butterflies are located in the top left and bottom right corners of the image. The fairy's wings are spread out, and her dress is flowing, giving the impression of movement. The lavender plants are tall and thin, with small purple flowers at the top. The butterflies are small and white, with black spots on their wings. The overall color scheme of the image is purple, white, and light blue.", + "bbox": [ + 202, + 95, + 473, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8fba0d29514340cd3abd3b56a4d7ffb1f2eac1acd306f445dfc941a3496d5d6e.jpg", + "image_caption": [ + "Figure 4. Example of constructing Subject-Oriented QA pairs. Using the image captions, we employ the Llama3.1 [59] to generate an average of three questions per image, resulting in 8 million QA pairs focused on different objects described in each caption." + ], + "image_footnote": [], + "bbox": [ + 102, + 204, + 127, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": " Generate three question-answer pairs based on this image description Each question should ask about one different object in the image without repeating objects.", + "bbox": [ + 135, + 205, + 470, + 231 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b2cafc2701ac137f928f5b6ccf1737aae6d577a336bc8dcbc4d8b04be533aa11.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 239, + 129, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Llama", + "bbox": [ + 102, + 257, + 130, + 265 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Question: what is the fairy wearing? Answer: The fairy is wearing a purple dress. \n2. Question: How many butterflies are there? Answer: There are two.", + "3. Question: What do the lavender plants look like? Answer: They are tall and thin, with small purple flowers at the top." + ], + "bbox": [ + 138, + 241, + 470, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "datasets [3, 4, 11]. As illustrated in Figure 4, we leverage Llama 3.1 [59] for its strong contextual language reasoning ability, using only structured text prompts as input to generate subject-focused conversations. After generating these conversations, we apply an open-vocabulary detection method to validate the presence of specified objects in the images, filtering out any noisy data and ensuring data quality. Experiments in Table 4 demonstrate the significance of this pretraining dataset, especially for challenging benchmarks that require an understanding of zoomed-in patches or complex reasoning tasks. The details of data curation are available in the supplementary material.", + "bbox": [ + 88, + 376, + 483, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 571, + 223, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we provide the training settings and evaluation benchmarks employed to assess our model's performance on high-resolution and general Vision-Language Model (VLM) tasks. Subsequently, we present comparative results, benchmarking our method against state-of-the-art approaches. Further details are available in the supplementary material.", + "bbox": [ + 89, + 595, + 483, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 712, + 281, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training stages The goal of TEVA is to enhance the model's ability to process high-resolution images by focusing on informative patches. Two key requirements are essential for achieving this: (1) enabling the vision tower to handle dynamic patch sampling, and (2) ensuring that dynamic vision tokens can be effectively interpreted by the LLM. To meet these goals, we structured our training pipeline into four stages:", + "bbox": [ + 89, + 734, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Alignment Stage: Following the process used in LLaVA [37], we begin by aligning visual features with the language space. In this stage, we freeze the large", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "language model (LLM) and optimize only the vision-language (VL) adapter with LLaVA alignment dataset (595k), ensuring that the visual features are properly mapped to the language domain.", + "bbox": [ + 524, + 90, + 906, + 151 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Vision Tower Adapting: Using 4M samples from our proposed Subject-Oriented Dataset, we fine-tune the vision tower to handle dynamic image patches. During this stage, both the vision tower and VL adapter are unfrozen for training.", + "- Integrated Interpretation Stage: To further enhance the LLM's ability to interpret dynamic vision tokens, we unfreeze the vision tower, VL adapter, and LLM, then fine-tune using 4M samples from the Subject-Oriented dataset and 1M samples from the ShareGPT4V [4] caption dataset. This step aligns all components for seamless interpretation of dynamic visual information.", + "- Instruction Fine-tuning: In the final stage, following the training strategy of VILA [37], we continue fine-tuning the vision tower, VL adapter, and LLM with a mix of Supervised Fine-Tuning (SFT) Dataset (4.17M) built on VILA [37]. This instruction fine-tuning stage enhances the model's instruction-following and dialogue capabilities on multiple tasks. Details of the training settings and datasets are provided in the supplementary materials." + ], + "bbox": [ + 513, + 152, + 906, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation benchmarks To assess our model's capability in handling complex high-resolution images, we report results on challenging high-resolution benchmarks, including $V^{*}$ [61] and MME-RealWorld [79]. Additionally, we evaluate our model on commonly used general VQA and MLLM benchmarks including MMBench [41], SQA (ScienceQA) [43], $\\mathrm{VQAT}$ (TextVQA) [54], POPE [29], AI2D [22], LLaVAw (LLaVA-Bench In-the-Wild) [37], $\\mathrm{MMMU}^{\\mathrm{t}}$ (MMMU-test) [69], and $\\mathrm{VQAV}^{\\mathrm{v2}}$ (VQA-v2) [14]. Evaluation of more benchmarks is available in the supplementary.", + "bbox": [ + 511, + 455, + 906, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 635, + 653, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with region-aware VLMs. We evaluate Token-Efficient Vision LLanguage Model (TEVA) on the high-resolution visual detail benchmark $V^{*}$ [61], comparing its performance and efficiency against state-of-the-art region-aware VLMs. As shown in Table 1, TEVA consistently outperforms existing methods in attribute recognition (AR), spatial relationship reasoning (SRR), and scene OCR tasks. Our dynamic patch sampling strategy significantly reduces computational overhead compared to SpatialGPT (mask pooling) [6], Griffon-G [72] and $V^{*}$ [61], which require multiple vision tower passes to process individual ROIs. Specifically, TEVA achieves a $7.53 \\times$ speedup over $V^{*}$ in ROI aggregation and improves overall inference efficiency by $12.31 \\times$ compared to $V^{*}$ [61], benefiting from both token efficiency and the lightweight RAP module, which eliminates the need for iterative region searching.", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24151", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/085235beebac11b4f33d69d6877492a23ce14f4570b380128d8a1f963079ef42.jpg", + "image_caption": [ + "How many people are there on the street?" + ], + "image_footnote": [ + "VILA-1.5", + "There is only one person on the street." + ], + "bbox": [ + 106, + 111, + 354, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f8b68f9c05c5bc06578ee1d47e0188bebaeb629b3f58091a4e5a4f62b091e58e.jpg", + "image_caption": [], + "image_footnote": [ + "Monkey", + "ChatGPT-40" + ], + "bbox": [ + 117, + 292, + 140, + 306 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1c03c33f34dd65596fba737f31f944c0d4911914a99ecc3cf3149604dbb26c5e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 319, + 142, + 340 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d14656949e1ee1f5aacdbb1ed0b61243d3bb2100ff5996ec3031812e72783e7.jpg", + "image_caption": [ + "What is the breed of the dog?" + ], + "image_footnote": [ + "VILA-1.5", + "The dog in the image is a white dog.", + "Pug." + ], + "bbox": [ + 379, + 111, + 627, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/60af72dfd219e8d855f0abfbec23ebe28574bb77f811b94b3f3e49117486cbd2.jpg", + "image_caption": [], + "image_footnote": [ + "There is one person visible on the street.", + "Monkey", + "difficult to identify the breed." + ], + "bbox": [ + 370, + 292, + 431, + 316 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/43cb6b63581acaebe012f234d233b642e9323d55430d6ff4c705cabc2204a7c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 319, + 411, + 340 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/91f5ccf34abeafccae6483240717f387b505bce7a61f9d6a27230b3755b76cf0.jpg", + "image_caption": [ + "What color is the bottle cap?" + ], + "image_footnote": [ + "VILA-1.5", + "The bottle has a blue cap.", + "Yellow." + ], + "bbox": [ + 651, + 111, + 815, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0a0b956b9bb5161918f9cc06637ccc05d35f17fdcbe11e51da90eda1f5859fa5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 111, + 897, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/553d44c330c57927bc9f3368901a6b50ef78cdd9ca48860753933d2268cb6515.jpg", + "image_caption": [], + "image_footnote": [ + "Monkey", + "ChatGPT-40", + "The color of the cap appears to be red." + ], + "bbox": [ + 658, + 292, + 681, + 308 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6f24cc55cee359e0aba86ad7b6cb7d82683414772db15359e36f7def45fb098e.jpg", + "image_caption": [ + "What is the speed limit?" + ], + "image_footnote": [], + "bbox": [ + 109, + 369, + 160, + 422 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0ba9204fef177b283b3250feb35545e346d371c02659dbb4cbbd736ffe99c3d.jpg", + "image_caption": [], + "image_footnote": [ + "There are 2.", + "VILA-1.5", + "It is $10\\mathrm{km / h}$", + "50." + ], + "bbox": [ + 161, + 369, + 354, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c43006d0fce0039a6d9d486f3787ef41bd340821e1b4bd761cfbf16c5ac1d097.jpg", + "image_caption": [], + "image_footnote": [ + "Monkey", + "ChatGPT-40" + ], + "bbox": [ + 119, + 551, + 140, + 566 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/953c4b79f9ba2434e1c42fc26b0130d3ccff034123a9ebf8a280b3fb2718630c.jpg", + "image_caption": [ + "Figure 5. Evaluation of our model against leading open-source VLMs (VILA-1.5 [32], Monkey [31]) and the advanced commercial VLM, GPT-4o [50], in high-resolution, real-world scenarios. Each image displays the proposed subject in the bottom left corner. The first row presents responses to general questions, while the second row addresses more challenging OCR-based questions. The proposed subject is in the left-bottom corner of each image. Despite using a smaller language model (Sheared-LLaMA-2.7B [62]), our model demonstrates robust capability in observing precise details for high-resolution images and can adapt to various image resolutions, providing flexibility and enhanced performance in demanding high-resolution scenarios. Please note that all sample images were processed at their original resolution $(2246\\times 1582$ on average) for TEVA-3B." + ], + "image_footnote": [ + "It is 40." + ], + "bbox": [ + 120, + 580, + 140, + 599 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8a7c6982364fe91658d78a93f969354fc3d4cc145684e03181d5c77733af1a15.jpg", + "image_caption": [ + "What is the written on the large boat?" + ], + "image_footnote": [ + "Golden Retriever.", + "VILA-1.5", + "\"Oranje\" is written on it.", + "Unanswerable." + ], + "bbox": [ + 380, + 369, + 627, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/15b6f7534fd3582d3f29008687d15573d545b17fd122636612f695d70f1fed32.jpg", + "image_caption": [], + "image_footnote": [ + "The speed limit appears to be 30.", + "Monkey", + "The text on the boat reads \"INTERCOMBI\".", + "Nitrico." + ], + "bbox": [ + 369, + 554, + 431, + 575 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b95613fedc10d8511381c4aebc4f39423c81bf8867134618bcacc85adec5511.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 580, + 413, + 599 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ec8214adc5331ac7380da365b2f65983cd5dfdda8eeb9d1cc8aab824a5de3a1d.jpg", + "image_caption": [ + "Tell me the yellow text on the airship." + ], + "image_footnote": [ + "It is red.", + "VILA-1.5", + "\"Airship\".", + "Unanswerable." + ], + "bbox": [ + 648, + 369, + 897, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/feb679bfac17eff5710fe9a143d10bc73df4d2b69d6835edc5a07f2ee83979b1.jpg", + "image_caption": [], + "image_footnote": [ + "Monkey", + "ChatGPT-40", + "The text appears to say \"GOODYEAR\"." + ], + "bbox": [ + 665, + 554, + 684, + 568 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/895625fa76c67137b73f26c3e736eeba01853e859ab89f60e1ef97bcd55a0e3f.jpg", + "image_caption": [], + "image_footnote": [ + "The yellow text reads \"GoodYear\"." + ], + "bbox": [ + 663, + 580, + 684, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlike SpatialGPT [6] and Griffon-G [72], which rely on manual region selection, TEVA automatically detects and proposes ROIs. For a fair comparison, we use TEVA's RAP module to predict ROIs for SpatialGPT and Griffon-G.", + "bbox": [ + 89, + 744, + 482, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "High resolution benchmark results. To demonstrate TEVA's capability in handling high-resolution images, we evaluated our method on high-resolution VLM benchmarks: $V^{*}$ [61], with an average image resolution of $2246 \\times 1582$ pixels and target subjects, and MME-RealWorld [79], with an average resolution of $2000 \\times 1500$ pixels.", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 2, TEVA excels in tasks that require precise detail recognition, such as attribute recognition (AR) and spatial relationship reasoning (SRR) in $V^{*}$ benchmark, outperforming both the commercial model GPT-4V [50] by 21.6 points and all the other open-source high resolution VLMs. In the complex MME-RealWorld benchmark, TEVA also surpasses open-source methods in both perception and reasoning ability. In applications that demand high-resolution fine visual detail perception, such as remote sensing, video monitoring, and autonomous driv", + "bbox": [ + 511, + 744, + 906, + 896 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24152", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3dd3fba3ce5262ce457a8915186bf32b6e2c436b4851f2856f88ec911646b3f6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodV* Bench
ARSRROCR
SpatialGPT [6]36.556.540.0
Griffon-G [72]44.771.056.6
V* [61]74.776.380.0
TEVA-7B76.577.683.3
MethodInference Time
AIPFPI
SpatialGPT [6]9.57×1.05×
Griffon-G [72]10.24×4.83×
V* [61]7.53×12.31×
TEVA-7B1.00×1.00×
", + "bbox": [ + 99, + 89, + 359, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1dfe1c0a1964722ddf371451fdbddfde63d70b05c0f1fbee9c569fa98ec1d5b4.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-art region-aware 7B VLMs on the high-resolution $V^{*}$ benchmark, evaluating attribute recognition (AR), spatial relationship reasoning (SRR), and scene OCR tasks. Inference times are reported for Additional Information Processing (AIP) and Full Pipeline Inference (FPI), demonstrating the efficiency and performance gains of dynamic patch sampling over SpatialGPT(pooling), Griffon-G and $V^{*}$ (multiple vision tower passes)." + ], + "table_footnote": [], + "table_body": "
MethodV* AverageMME-RWPRCMME-RWRSN
OCRRSMOADOCRMOAD
GPT-4V [50]55.9-------
GPT-4o [50]-77.728.933.922.461.436.526.4
Gemini Pro [56]50.5-------
Gemini-1.5-pro [56]-67.614.031.126.652.728.319.2
LLaVA-1.5 [35]50.538.722.119.126.026.025.924.2
LLaVA-1.5-S2[52]56.5-------
Text-Monkey [40]-37.311.716.114.330.44.420.0
Monkey [31]-54.625.028.029.727.227.333.0
LLaVA-Next [36]61.340.631.635.427.214.929.528.2
Cambrian-1 [57]60.832.735.827.635.527.531.231.2
MLLM-K [75]62.3-------
TEVA-3B61.643.132.332.633.041.833.529.9
TEVA-7B77.160.738.236.935.249.434.735.5
", + "bbox": [ + 397, + 89, + 887, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8c7437fd5645ffe4e60d57f9c65fd5e7ca6b6d214201a1e540dfd0bc2c31f982.jpg", + "table_caption": [ + "Table 2. Comparison with state-of-the-art open-source and commercial methods on the high-resolution benchmark $V^{*}$ [61] (averaging attribute recognition(AR) and spatial relationship reasoning(SRR)), as well as on MME-RealWorld [79], which includes perception benchmark MME-RW $_{PRC}$ and reasoning benchmark MME-RW $_{RSN}$ . Task categories are abbreviated due to space limits: OCR (Optical Character Recognition), RS (Remote Sensing), MO (Video Monitoring), AD (Autonomous Driving). Our model consistently outperforms other open-source methods on Reasoning tasks such as OCR, RS, MO, AR and SRR. It even surpasses leading commercial VLMs, GPT-4o and Gemini-1.5-pro, on RS, MO, and AD perception tasks that require precise visual grounding. The best results are bold. Notably, all samples are evaluated at their original image resolution for TEVA and at the largest default resolution for other methods." + ], + "table_footnote": [], + "table_body": "
MethodLLM#Visual TokenMMBSQAVQATPOPEAI2DLLaVAwMMMUVVQAV2
InstructBLIP [7]Vicuna-7B-36.063.150.7--60.9--
IDEFICS-9B [18]LLAMA-7B-48.2------50.9
LLaVA-1.5 [35]Vicuna-7B57664.366.866.885.9-63.435.778.5
MLLM-K [75]Vicuna-7B---56.187.3---76.5
LLaVA-1.5-S2[52]Vicuna-7B57666.2-----37.780.0
VILA-1.5 [32]Llama-2-7B72968.968.268.285.5-69.738.679.9
Mini-Gemini-HD [30]Vicuna-7B288072.775.168.4-73.5-37.3-
Monkey [31]Qwen-7B5120-69.467.6-62.6--80.3
LLaVA-Next [36]LLama3-8B288072.172.864.686.771.683.241.781.8
Cambrian-1 [57]Llama3-8B-Instruct230475.980.471.7-73.0-42.781.2
TEVA-3BSheared-llama-2.7B72964.275.766.887.064.985.032.480.3
TEVA-7BQwen2-7B72980.392.572.587.976.394.348.482.8
", + "bbox": [ + 91, + 459, + 903, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison with leading methods on 8 commonly used VLM benchmarks: MMBench [41], SQA: ScienceQA [43], VQATextVQA [54], POPE [29], AI2D [22], LLaVAw: LLaVA-Bench (In-the-Wild) [37], MMMUv: MMMU-validation [69], VQAV2: VQAV2 [14]. Our model consistently outperforms the other methods without image tiling and increasing visual tokens. The best results are bold. Please note that all the samples are evaluated on original image resolution for our model.", + "bbox": [ + 89, + 643, + 906, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing (e.g., identifying traffic signs), TEVA demonstrates robust performance and even surpassing the most powerful commercial models, GPT-4o [50] and Gemini-1.5-pro [56]", + "bbox": [ + 89, + 720, + 480, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "General resolution benchmark results. We also evaluate our model on eight commonly used VQA and MLLM benchmarks, comparing it with state-of-the-art VLMs, as shown in Table 3. The results demonstrate that TEVA consistently outperforms other leading methods across all the benchmarks and achieves this without requiring image tiling or visual token extending.", + "bbox": [ + 89, + 772, + 482, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Notably, on general resolution subject-oriented QA tasks", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "like MMBench [41], AI2D [22] and $\\mathrm{MMMU}^{\\mathrm{V}}$ [69], TEVA achieves superior results compared to all other methods. This highlights the generalizability of our approach in focusing on contextually important areas within an image.", + "bbox": [ + 511, + 719, + 906, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. In Figure 5, we present qualitative results that demonstrate the effectiveness of TEVA-3B in challenging high-resolution scenarios requiring precise detail recognition. By leveraging our informative patch sampling strategy, TEVA dynamically focuses on relevant, information-rich regions based on the user query, enabling inference on images at arbitrary resolutions without increas", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24153", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1dcfc085b1a86ab4f5056d8b296b5b3ad36c4b3c3299ad1ce664d2e58bf46e31.jpg", + "table_caption": [ + "Results with Different Dynamic and Static Token Numbers" + ], + "table_footnote": [], + "table_body": "
DynamicStaticV*MRWPOPEVQAT
025645.030.083.044.5
808150.431.585.654.2
47325661.634.087.066.8
", + "bbox": [ + 94, + 107, + 475, + 169 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3f892eecd0a7f7fb5445460ed1d8a788c0898db5a9a924714cd871bd3bbd78c3.jpg", + "table_caption": [ + "Comparison of w/o Subject-Orient (SO) Data" + ], + "table_footnote": [], + "table_body": "
SO DataMRWPRCMRWRSNPOPEVQAT
X33.128.586.558.7
✓33.834.387.066.8
", + "bbox": [ + 99, + 186, + 475, + 237 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bb5339fb463b57a54e52695d492689d04f9e268f79bbb892c3ce8ab9ee6327c8.jpg", + "table_caption": [ + "Comparison of w/o Matrix Embedding (ME)" + ], + "table_footnote": [], + "table_body": "
MEV*ARV*SRRMRWPOPE
X48.350.031.884.8
✓60.063.234.087.0
", + "bbox": [ + 99, + 253, + 475, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation Studies on Token Numbers, w/o Subject-Orient Pretraining and w/o Affine Transformation Matrix Embedding. Benchmark abbreviation MRW: MME-RealWorld [79].", + "bbox": [ + 89, + 311, + 482, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing computational cost. For example, processing a 5760 $\\times$ 2880 (5K-resolution) image takes only 2.5 seconds on a single NVIDIA A100 GPU with TEVA-7B. This efficiency enables TEVA to handle high-resolution images seamlessly, outperforming leading open-source VLMs, such as VILA1.5-7B [32], which uses the same number of visual tokens (729), and Monkey-7B [31], which relies on image tiling for high-resolution support. Furthermore, our approach surpasses the advanced commercial VLM, ChatGPT-4o [50], in precise detail observation for high-resolution images, despite the latter's use of a more powerful language model. The evaluation covers real-world, high-difficulty OCR tasks along with diverse general queries, such as counting, identification, and attribute recognition, providing a comprehensive assessment of TEVA's capabilities.", + "bbox": [ + 88, + 369, + 482, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 606, + 243, + 622 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamic vs. Static Token Numbers. We assess the impact of using dynamic, informative image patches compared to static, uniformly sampled patches. As shown in Table 4, adding 473 dynamic patches to the baseline of 256 static patches (729 tokens in total) leads to significant improvements in tasks requiring detailed structural understanding, such as scene OCR tasks in $\\mathrm{VQA}^{\\mathrm{T}}$ (TextVQA) [54], as well as in handling high-resolution images, as demonstrated on the $V^{*}$ [61] benchmark. Additionally, we conducted experiments using a reduced set of 144 patches (80 dynamic and 81 static). Despite the lower patch count, incorporating dynamic tokens consistently enhanced performance on challenging datasets that demand high-resolution or precise observation.", + "bbox": [ + 89, + 628, + 482, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Subject-Oriented Pretraining. To evaluate the benefits of pretraining on our Subject-Oriented (SO) Dataset, we compare results with and without the Subject-Oriented Pretraining phase. As illustrated in Table 4, pretrain", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing on a large-scale Subject-Oriented Dataset notably improves the model's complex reasoning abilities (MMERealWorld $_{RSN}$ : MME-RealWorld-Reasoning [79]) and scene understanding, especially in tasks requiring OCR, such as $\\mathrm{VQAT}$ (TextVQA) [54]. However, for tasks requiring more straightforward scene comprehension, such as basic perception (MME-RealWorld $_{PRC}$ : MME-RealWorld-Perception [79]) or hallucination assessments (POPE [29]), the improvements from Subject-Oriented Pretraining are less pronounced. These findings demonstrate the effectiveness of Subject-Oriented Pretraining in enhancing the model's capacity for complex visual understanding.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Affine Transformation Matrix Embeddings. We examine the effectiveness of embedding the affine transformation matrix to provide positional information for each patch. Results in Table 4 demonstrate that the Affine Transformation Matrix Embeddings (ME) consistently improve performance across various benchmarks, particularly enhancing the model's ability to comprehend the relative positions of objects in the $V_{SRR}^{*}$ ( $V^{*}$ Spatial Relationship reasoning) benchmark. This aligns well with our intuition, as the additional spatial context helps the model maintain awareness of object locations and relationships.", + "bbox": [ + 511, + 272, + 906, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Limitation", + "text_level": 1, + "bbox": [ + 511, + 448, + 633, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While TEVA can process images at arbitrary resolutions without increasing the token count, its effectiveness may be limited in scenarios involving densely packed high-resolution images, such as text-heavy documents. In these cases, where information is uniformly distributed across the image, there are no distinct regions to prioritize. Consequently, dynamic sampling must cover the entire image at a lower resolution, limiting the advantages of high-resolution focus on key areas. Additional experiments and discussion are provided in the supplementary material.", + "bbox": [ + 511, + 469, + 905, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced Token-Efficient Vision LLanguage Model (TEVA), a straightforward and token-efficient framework that enhances Vision-Language Models (VLMs) for high-resolution image processing. TEVA employs a subject-oriented strategy and Dynamic Patch Sampling to balance fine details with global context, supported by a specialized Subject-Oriented Dataset for improved learning. Comprehensive evaluations against state-of-the-art VLMs, including GPT-4o, show that TEVA excels in challenging benchmarks and holds strong potential for applications requiring high-resolution precision, such as remote sensing, autonomous driving, and video monitoring.", + "bbox": [ + 511, + 659, + 905, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24154", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. 2022. 1", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2", + "[3] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.4.5", + "[4] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 5, 7", + "[5] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 1", + "[6] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062-135093, 2025. 2, 5, 6, 7", + "[7] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 7", + "[8] Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuoling Yang, Zihan Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nvlm: Open frontier-class multimodal llms. arXiv preprint arXiv:2409.11402, 2024. 2", + "[9] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 1", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[11] Yunhao Fang, Ligeng Zhu, Yao Lu, Yan Wang, Pavlo Molchanov, Jang Hyun Cho, Marco Pavone, Song Han, and Hongxu Yin. vila2: Vila augmented vila. arXiv preprint arXiv:2407.17453, 2024. 5", + "[12] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "comprehensive evaluation benchmark for multimodal large language models. ArXiv, abs/2306.13394, 2023. 1, 3", + "[13] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023. 7", + "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 5, 7, 1, 3", + "[15] Qiushan Guo, Shalini De Mello, Hongxu Yin, Wonmin Byeon, Ka Chun Cheung, Yizhou Yu, Ping Luo, and Sifei Liu. Regionpt: Towards region understanding vision language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13796-13806, 2024. 2, 4", + "[16] Jack Hessel, Jena D Hwang, Jae Sung Park, Rowan Zellers, Chandra Bhagavatula, Anna Rohrbach, Kate Saenko, and Yejin Choi. The abduction of sherlock holmes: A dataset for visual abductive reasoning. In European Conference on Computer Vision, pages 558-575. Springer, 2022. 7", + "[17] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 3, 7", + "[18] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. 2023. 7", + "[19] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. 4", + "[20] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901-2910, 2017. 7", + "[21] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 7", + "[22] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 5, 7, 1, 3", + "[23] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeong Yeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24155", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Jonathan Krause, Justin Johnson, Ranjay Krishna, and Li Fei-Fei. A hierarchical approach for generating descriptive image paragraphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 317–325, 2017. 7", + "[25] Paul Lerner, Olivier Ferret, Camille Guinaudeau, Hervé Le Borgne, Romaric Besançon, José G Moreno, and Jesús Lovón Melgarejo. Viquae, a dataset for knowledge-based visual question answering about named entities. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 3108-3120, 2022. 7", + "[26] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 1, 3", + "[27] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 7", + "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1", + "[29] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 5, 7, 8, 1, 3", + "[30] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 2, 7", + "[31] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26763-26773, 2024. 2, 6, 7, 8", + "[32] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 6, 7, 8, 1, 2", + "[33] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1, 2", + "[34] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. Mmc: Advancing multimodal chart understanding with large-scale instruction tuning. arXiv preprint arXiv:2311.10774, 2023. 7", + "[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 7" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.1,7", + "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NIPS, 2024. 1, 5, 7, 3", + "[38] Haogeng Liu, Quanzeng You, Xiaotian Han, Yiqi Wang, Bohan Zhai, Yongfei Liu, Yunzhe Tao, Huaibo Huang, Ran He, and Hongxia Yang. Infimm-hd: A leap forward in high-resolution multimodal understanding. arXiv preprint arXiv:2403.01487, 2024. 1, 2", + "[39] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 2, 5", + "[40] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 7", + "[41] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 5, 7, 1, 3", + "[42] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, Xiuyu Li, Yunhao Fang, Yukang Chen, Cheng-Yu Hsieh, De-An Huang, An-Chieh Cheng, Vishwesh Nath, Jinyi Hu, Sifei Liu, Ranjay Krishna, Daguang Xu, Xiaolong Wang, Pavlo Molchanov, Jan Kautz, Hongxu Yin, Song Han, and Yao Lu. Nvila: Efficient frontier visual language models, 2024. 1", + "[43] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5, 7, 1, 3", + "[44] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 7", + "[45] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2, 7", + "[46] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2", + "[47] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 2", + "[48] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24156", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 7", + "[49] Ravina Mithe, Supriya Indalkar, and Nilam Divekar. Optical character recognition. International journal of recent technology and engineering (IJRTE), 2(1):72-75, 2013. 1", + "[50] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2(5), 2023. 1, 6, 7, 8, 2", + "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 4", + "[52] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? In European Conference on Computer Vision, pages 444-462. Springer, 2025. 1, 2, 7", + "[53] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 742–758, 2020. 7", + "[54] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 7, 8, 1, 3", + "[55] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. Wit: Wikipedia-based image text dataset for multimodal multilingual machine learning. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, pages 2443-2449, 2021. 7", + "[56] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 7", + "[57] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2024. 7", + "[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 5, 2", + "[60] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 7", + "[61] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084-13094, 2024. 2, 4, 5, 6, 7, 8, 1", + "[62] Mengzhou Xia, Tianyu Gao, Zhiyuan Zeng, and Danqi Chen. Sheared llama: Accelerating language model pre-training via structured pruning. arXiv preprint arXiv:2310.06694, 2023. 6", + "[63] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794-4803, 2022. 2", + "[64] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 2", + "[65] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1", + "[66] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 4", + "[67] Xiaoyu Yue, Shuyang Sun, Zhanghui Kuang, Meng Wei, Philip HS Torr, Wayne Zhang, and Dahua Lin. Vision transformer with progressive sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 387-396, 2021. 2", + "[68] Xiang Yue, Xingwei Qu, Ge Zhang, Yao Fu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mammoth: Building math generalist models through hybrid instruction tuning. arXiv preprint arXiv:2309.05653, 2023. 7", + "[69] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567, 2024. 5, 7, 1, 3", + "[70] Bohan Zhai, Shijia Yang, Chenfeng Xu, Sheng Shen, Kurt Keutzer, and Manling Li. Halle-switch: Controlling object hallucination in large vision language models. arXiv e-prints, 2023. 2", + "[71] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 1", + "[72] Yufei Zhan, Hongyin Zhao, Yousong Zhu, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon-g: Bridging vision-language and vision-centric tasks via large multimodal models. arXiv preprint arXiv:2410.16163, 2024. 2, 4, 5, 6, 7", + "[73] Yufei Zhan, Yousong Zhu, Hongyin Zhao, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon v2: Advancing multimodal" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24157", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "perception with high-resolution scaling and visual-language co-referring. arXiv preprint arXiv:2403.09333, 2024. 1, 2, 4", + "[74] Haotian Zhang, Haoxuan You, Philipp Duffer, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 2", + "[75] Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara, and Filip Ilievski. Mllms know where to look: Training-free perception of small visual details with multimodal llms. Proceedings of International Conference on Learning Representations (ICLR), 2025. 7", + "[76] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Haodong Duan, Songyang Zhang, Shuangrui Ding, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 2", + "[77] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1", + "[78] Yanzhe Zhang, Ruiyi Zhang, Jiuming Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 2, 7", + "[79] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, et al. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans? arXiv preprint arXiv:2408.13257, 2024. 5, 6, 7, 8, 1", + "[80] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1" + ], + "bbox": [ + 91, + 92, + 482, + 626 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "24158", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/b324faad-4fe9-44a7-a833-a82bac07835f_model.json b/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/b324faad-4fe9-44a7-a833-a82bac07835f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4e6775dd53b5b9d243c0faf5c801951b9ec4f5ab --- /dev/null +++ b/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/b324faad-4fe9-44a7-a833-a82bac07835f_model.json @@ -0,0 +1,3249 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.13, + 0.877, + 0.178 + ], + "angle": 0, + "content": "Token-Efficient VLM: High-Resolution Image Understanding via Dynamic Region Proposal" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.202, + 0.851, + 0.26 + ], + "angle": 0, + "content": "Yitong Jiang\\(^{1,2,\\dagger}\\) Jinwei Gu\\(^{1,2}\\) Tianfan Xue\\(^{2}\\) Ka Chun Cheung\\(^{1}\\) Pavlo Molchanov\\(^{1}\\) Hongxu Yin\\(^{1\\dagger}\\) Sifei Liu\\(^{1\\dagger}\\) \n\\(^{1}\\) NVIDIA \\({}^{2}\\) The Chinese University of Hong Kong" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.28, + 0.373, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.289, + 0.564, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.281, + 0.772, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.281, + 0.9, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.431, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Figure 1. Illustration of the Token-Efficient Vision LLanguage Model (TEVA) for high-resolution scenarios. Given a user query about a specific detail \\((20\\times 50\\mathrm{px}\\) region) in a high-resolution image \\((5760\\times 2880\\mathrm{px})\\) , traditional VLMs (Left) often miss critical details due to uniform patch processing and computational limits. TEVA (Right) addresses this by using a Relevant Area Proposal module to identify key subjects (e.g., people) and Dynamic Patch Sampling to focus on high-resolution patches of interest while retaining global context patches. This approach enables accurate interpretation of details at arbitrary resolutions without increasing the token count." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.514, + 0.328, + 0.53 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.549, + 0.486, + 0.837 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) excel at visual understanding by leveraging pretrained image encoders to generate visual tokens. However, they struggle with high-resolution images and zoomed-in regions due to the computational burden and token redundancy of uniform patch-based processing, often leading to the loss of critical details. To address these challenges, we propose Token-Efficient Vision LLanguage Model (TEVA), a novel framework that detects key regions and applies dynamic patch sampling to efficiently capture fine-grained details while preserving global context. Our approach first identifies subject-oriented regions using an adaptive detection strategy. Then, a dynamic patch sampling mechanism selects and arranges patches at varying scales, ensuring efficient processing without increasing token count. Extensive experiments demonstrate that Token-Efficient Vision LLanguage Model (TEVA) significantly enhances VLM performance in handling visual details, seamlessly integrating with various decoders and LLMs." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.514, + 0.645, + 0.53 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.759 + ], + "angle": 0, + "content": "Vision-Language Models (VLMs) [1, 36, 37, 80] have recently received significant attention due to their ability to integrate visual understanding into Large Language Models (LLMs) [50, 59, 77], enhancing performance in tasks like image captioning and visual question answering. However, supporting applications that require precise attention to specific details—such as video surveillance analysis, satellite imagery interpretation, and scene OCR [49]—remains challenging. These tasks often necessitate zooming into fine-grained regions of high-resolution images to capture essential information, but many existing VLMs [1, 28, 37, 80] rely on low-resolution pretrained image encoders (e.g., \\(224 \\times 224\\) pixels), leading to the loss of important details [38, 52]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address this issue, previous methods have focused on either increasing global resolution or applying region-aware techniques. Enhancing resolution in pretrained image encoders typically involves two strategies: scaling resolution via curriculum learning [2, 5, 73] and partitioning images into fixed-size sub-images for independent processing [9, 33, 35, 36, 38, 42, 52]. These approaches have notable limitations. Scaling resolution increases computational complexity quadratically, raising self-attention" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.874, + 0.419, + 0.901 + ], + "angle": 0, + "content": "†Yitong Jiang was an intern at NVIDIA during the project. \n†Equal Advising." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "24147" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "costs. Partitioning images boosts the number of visual tokens (e.g., over 5000 tokens for an \\(896 \\times 896\\) image [31]), leading to redundancy and degraded performance [35, 38]. Moreover, the lack of interaction among image tokens within the vision encoder limits the model's ability to capture contextual relationships, requiring specialized design interventions [31, 38]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.482, + 0.32 + ], + "angle": 0, + "content": "Alternatively, region-aware methods [6, 15, 66, 72, 73] focus on fine-grained details by leveraging manually defined Regions of Interest (ROIs). However, these methods either compress visual tokens within target regions, potentially losing critical information [6, 15, 66], or require multiple independent passes through the vision encoder, significantly increasing computational cost and token usage [72, 73]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.323, + 0.483, + 0.625 + ], + "angle": 0, + "content": "To address these limitations, we propose Token-Efficient Vision LLanguage Model (TEVA), which leverages a dynamic patch sampling framework for the vision encoder that balances visual token count and image resolution. Our primary concept involves informative patch sampling within the vision encoder, which selectively focuses on key parts of the image while preserving the understanding of the global context, represented in low resolution. As shown in Figure 1, unlike conventional vision encoders that uniformly partition the entire image into patches, our method makes an early determination of which image regions are essential for answering a given question and applies selective patch sampling before inputting patches into the vision encoder. This approach not only prevents an increase in the number of patches when processing high-resolution images of any size, but also reduces self-attention computation and minimizes tokens the LLM needs to process. Additionally, this approach maintains coherence among sampled image regions, enhancing the model's ability to understand the global context and structure of the input image [63, 67]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Our TEVA consists of three main steps. First, the Relevant Area Proposal (RAP) module identifies key locations and scales of interest using a subject proposal agent powered by an LLM, and a subject detection agent via a vision foundation model. This enables efficient sampling of informative patches as vision tokens. Next, patches from localized regions are arranged with low-resolution global patches, integrating positional information through an affine transformation strategy. This allows for sampling patches of varying sizes and locations without additional tokens or complex architectures [8, 31, 38, 64]. Finally, the mixture of local and global patch embeddings is processed in the vision encoder (SigLIP [71] in our work) to extract visual tokens for the Large Language Models (LLMs)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Furthermore, to enhance dynamic patch sampling for VLMs, we introduce a data curation pipeline and curate a Subject-Oriented (SO) dataset containing 8 million subject-focused conversations. This is the first large-scale Subject" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "Oriented VQA dataset for VLM pretraining, enabling a framework that focuses on specific image details. By integrating dynamic image tokens with early interactions between local and global visual elements, Token-Efficient Vision LLanguage Model (TEVA) achieves superior VLM performance across various benchmarks, particularly excelling in flexible-resolution scenarios, as shown in our comparative analysis in Section 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.226, + 0.655, + 0.241 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.905, + 0.372 + ], + "angle": 0, + "content": "Recent studies [58, 70, 78] highlight a bottleneck in Vision-Language Models (VLMs), where image resolution constraints limit the model's ability to capture fine-grained visual details. Previous approaches of handling high-resolution images in Vision-Language Models (VLMs) can be grouped into two main strategies: (1) increasing the global image resolution, (2) applying region-aware techniques, as discussed in the following." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.905, + 0.689 + ], + "angle": 0, + "content": "In the first strategy, methods like Qwen-VL [2] and Griffon v2 [73] employ curriculum learning and bilinear interpolation to gradually increase the resolution in the pretrained image encoder. However, a significant challenge with transformer-based image encoders [10] is their reliance on uniform patch sampling, which leads to a quadratic increase in computational complexity as image resolution grows. Other methods [30, 31, 33, 38, 64, 76] partition high-resolution images into smaller sub-images, treating each as an independent input to the vision encoder. While this approach enables the processing of higher-resolution images, it imposes a constraint on the input sequence length that the Large Language Model (LLM) can handle, thereby limiting its overall capacity. To mitigate the computational burden, learnable queries within the Perceiver architecture [31, 38] and pooling strategies [52, 73] have been introduced to reduce the number of visual tokens generated by the vision encoder from these sub-images. However, such compression techniques often lack selectivity, leading to the loss of fine-grained visual details that are crucial for optimal Multimodal Large Language Model (MLLM) performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.854 + ], + "angle": 0, + "content": "The second strategy involves explicitly identifying and processing Regions-of-Interest (ROIs). Some methods [6, 15, 66, 74] require manual specification of ROIs, followed by compression of low-resolution visual tokens. Others [72, 73] necessitate multiple passes through the vision encoder, increasing computational overhead. \\( V^{*} \\) [61] further extends this idea by iteratively searching for ROIs through multiple passes through an additional Vision-Language Model (VLM). However, this iterative search is computationally expensive and also requires multiple forward passes through the vision encoder for each ROI candidate." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In contrast to these methods, our approach automatically identifies and focuses on informative regions within high-resolution images. By applying dynamic patch sam" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24148" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.904, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.253, + 0.908, + 0.296 + ], + "angle": 0, + "content": "Figure 2. The TEVA architecture. Given a user query, the Subject Proposal Agent identifies relevant subjects (e.g., \"signs\"). The Subject Detection Agent then locates all candidate signs in the image. Informative patch sampling extracts high-resolution local patches at these locations, which are concatenated with low-resolution global patches and fed into the VLM for final interpretation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.412 + ], + "angle": 0, + "content": "pling before the vision encoder, we preserve essential details without resorting to compression, while also minimizing computational costs. This is achieved by skipping less informative patches and avoiding multiple passes through the vision encoder, resulting in a more efficient and detailed visual representation for VLMs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.424, + 0.226, + 0.441 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.449, + 0.483, + 0.676 + ], + "angle": 0, + "content": "We show the overall flowchart of our proposed TokenEfficient Vision LLanguage Model (TEVA) in Figure 2. TEVA is a VLM framework designed to handle arbitrary high-resolution images by detecting and focusing on informative visual details within images. It incorporates a Relevant Area Proposal (RAP) module to identify and extract key regions from the input image. A targeted patch sampling strategy then partitions these selected areas into informative high-resolution local patches, while retaining low-resolution global patches for context. The concatenation of local and global patches is processed by the image encoder to generate visual tokens. Additionally, we introduce a Subject-Oriented (SO) dataset to enhance the VLM's capability to manage dynamic patch sampling and improve overall performance. Details are described below." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.683, + 0.373, + 0.699 + ], + "angle": 0, + "content": "3.1. Relevant Area Proposal Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Building on our insight that image tokens are essential for Vision-Language Models (VLMs) to generate accurate content descriptions and answer questions, we address a key challenge: as shown in Figure 3 (left), when VLMs process uniformly partitioned image patches, they often focus on prominent elements, such as the path, trees, and signs on the trees, while overlooking finer details, like the presence of two individuals in the forest." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To overcome this, we propose an automatic Relevant Area Proposal (RAP) module to capture key local details that help the VLM focus on the most informative and relevant areas in response to user queries. Current open-vocabulary detection methods, such as Grounding" + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.32, + 0.72, + 0.334 + ], + "angle": 0, + "content": "What can you see in this image?" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.337, + 0.707, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.441, + 0.697, + 0.466 + ], + "angle": 0, + "content": "Assistant: A forest with tall trees and a trail in the middle." + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.337, + 0.905, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.717, + 0.441, + 0.884, + 0.467 + ], + "angle": 0, + "content": "Assistant: A group of people are walking through a forest." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Figure 3. Example of visual focus in Vision-Language Models (VLMs) with uniform image patches (Left) versus high-resolution, relevant patches (Right). In the uniform approach, the model emphasizes broad features like trees, trails, and visible signs, missing subtler details, such as the presence of people in the scene. By providing high-resolution relevant patches, the model can identify and interpret finer details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.674 + ], + "angle": 0, + "content": "Dino [39], have shown strong performance in detecting regions of interest based on straightforward text cues. However, user queries often involve complex questions or instructions without an explicit subject, which limits the applicability of open-vocabulary detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.81 + ], + "angle": 0, + "content": "To address this, we use two specialized agents: a Subject-Promise Agent using LLM, e.g., Llama3.1 [59], and a Subject Detection Agent via Grounding Dino. Given a user query, we employ a structured prompt (e.g., \"You have limited vision of the image; tell me the subjects you need to observe to answer the question + User query\") to guide the LLM in proposing candidate subjects for detection. Once identified, the Subject Detection Agent then locates the regions of interest (ROI) based on these suggested subjects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.9 + ], + "angle": 0, + "content": "While this process effectively isolates the ROIs, integrating these fine-grained local details into the VLM remains challenging. To this end, we introduce the Informative Patch Sampling module in the following, which efficiently incorporates high-resolution information without increasing the token count or computational burden." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24149" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.345, + 0.108 + ], + "angle": 0, + "content": "3.2. Informative Patch Sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.371 + ], + "angle": 0, + "content": "Previous region-level VLMs [15, 66] handle regions of interest by relying on low-resolution features and apply pooling strategies to compress information within each region. Although effective for reducing computational load, these pooling methods often lead to a loss of critical details, especially for high-resolution images. Alternative approaches [61, 72, 73] treat each target region as a separate image that passes through a vision tower. However, this approach is computationally expensive and significantly increases the number of image tokens and inference time. To better leverage high-resolution image information without incurring excessive computational costs, we propose a novel approach that utilizes informative patch sampling within each high-resolution target region. This technique preserves fine-grained details while maintaining computational efficiency and the same token count comparable to a low-resolution image encoder, as demonstrated in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.484, + 0.495 + ], + "angle": 0, + "content": "Patch allocation. As shown in Algorithm 1, our method begins by allocating informative patch numbers to regions of interest based on their relative sizes. Next, we align the assigned patches within each bounding box, ensuring even coverage across each target area. Finally, we remove overlapping bounding boxes to maintain a total of \\( n \\), yielding the positions \\( P \\in \\mathbb{R}^{n \\times 2} \\) and scaling factors \\( S \\in \\mathbb{R}^{n \\times 2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.484, + 0.587 + ], + "angle": 0, + "content": "Patch sampling. Given the positions and the scaling factors, to link these positional clues to exact image patches, we embed \\( \\mathsf{P} = \\{(p_i^x,p_i^y)|i\\in (0,n - 1)\\} \\) and \\( S = \\{(s_i^x,s_i^y)|i\\in (0,n - 1)\\} \\) into an affine transformation matrix to project image pixels into patches:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.594, + 0.483, + 0.629 + ], + "angle": 0, + "content": "\\[\nA = \\left[ \\begin{array}{c c c} s _ {i} ^ {x} & 0 & p _ {i} ^ {x} \\\\ 0 & s _ {i} ^ {y} & p _ {i} ^ {y} \\end{array} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.483, + 0.665 + ], + "angle": 0, + "content": "With this setup, the correspondence between spatial coordinates of patches and input image can now be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.672, + 0.483, + 0.695 + ], + "angle": 0, + "content": "\\[\nc = A \\left[ \\begin{array}{l l l} u & v & 1 \\end{array} \\right] ^ {T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.484, + 0.822 + ], + "angle": 0, + "content": "where \\( c = [c_u, c_v] \\) are the spatial coordinates in the original image, while \\( u \\) and \\( v \\) are coordinates in the patch space, sampled on a uniform grid. Through sampling patch spatial coordinates with Eq. 2 we can determine the corresponding coordinates in the original image. Since \\( c \\) is not discrete, proper interpolation is necessary. Inspired by Spatial Transformer Networks [19], we employ bilinear interpolation to address this:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.842, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} c _ {u, v} = \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} I _ {(h, w)} \\max (0, 1 - | c _ {v} - h |) \\tag {3} \\\\ \\times \\max (0, 1 - | c _ {u} - w |), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.212 + ], + "angle": 0, + "content": "where \\( I_{(h,w)} \\) represents the pixel value in the original image \\( I \\) of size \\( H \\times W \\) at the location \\( (h,w) \\). For a patch of size \\( \\hat{H} \\times \\hat{W} \\), the coordinates are sampled as \\( u = \\{1,2,\\dots,\\hat{H}\\} \\); \\( v = \\{1,2,\\dots,\\hat{W}\\} \\). Unlike the original integration-based approach [19], which is time-consuming for high-resolution tasks, we use a look-up operation to accelerate by considering only the four neighboring pixels around each patch coordinate." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.907, + 0.289 + ], + "angle": 0, + "content": "To incorporate spatial information for each patch, we concatenate the affine transformation matrix embeddings with each patch's positional embedding. Experimental results in Table 4 demonstrate the effectiveness of this approach in enhancing the model's spatial awareness." + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.322, + 0.88, + 0.337 + ], + "angle": 0, + "content": "Algorithm 1 Simplified Informative Patch Sampling" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.339, + 0.555, + 0.35 + ], + "angle": 0, + "content": "Input:" + }, + { + "type": "code", + "bbox": [ + 0.523, + 0.35, + 0.869, + 0.386 + ], + "angle": 0, + "content": "Bounding boxes \\(\\mathsf{B} = \\{b_1,\\dots ,b_{\\mathrm{box\\_num}}\\} \\in \\mathbb{R}^{\\mathrm{box\\_num}\\times 4},\\) Each bounding box \\(b_{j} = [x_{min},y_{min},x_{max},y_{max}],\\) Number of local patches \\(n\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.387, + 0.563, + 0.398 + ], + "angle": 0, + "content": "Output:" + }, + { + "type": "code", + "bbox": [ + 0.524, + 0.399, + 0.824, + 0.422 + ], + "angle": 0, + "content": "Sampled patch locations \\(\\mathsf{P}\\in \\mathbb{R}^{n\\times 2}\\) Sampled patch scaling factors \\(\\mathsf{S}\\in \\mathbb{R}^{n\\times 2}\\)" + }, + { + "type": "code", + "bbox": [ + 0.516, + 0.427, + 0.837, + 0.47 + ], + "angle": 0, + "content": "1. Allocate patches Based on Area area \\((\\mathsf{b_j}) = (\\mathbf{x}_{\\mathrm{max}} - \\mathbf{x}_{\\mathrm{min}})(\\mathbf{y}_{\\mathrm{max}} - \\mathbf{y}_{\\mathrm{min}})\\) \\(\\forall \\mathsf{b_j}\\in \\mathsf{B}\\) n=j = num-patches(bj) = area(bj) n, \\(\\forall \\mathsf{b_j}\\in \\mathsf{B}\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.473, + 0.839, + 0.496 + ], + "angle": 0, + "content": "2. Generate patch locations and scales for \\(\\mathbf{b}_{\\mathrm{j}}\\in \\mathsf{B}\\) do" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.499, + 0.881, + 0.618 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\cdot \\operatorname {w i d t h} (\\mathfrak {b} _ {\\mathrm {j}}) = \\mathbf {x} _ {\\max } - \\mathbf {x} _ {\\min }, \\operatorname {h e i g h t} (\\mathfrak {b} _ {\\mathrm {j}}) = \\mathbf {y} _ {\\max } - \\mathbf {y} _ {\\min } \\\\ \\cdot \\text {p a t c h e s} _ {\\mathrm {w}}. \\text {s i d e} = \\sqrt {\\frac {\\operatorname {n u m} . \\operatorname {p a t c h e s} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right) \\cdot \\operatorname {w i d t h} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right)}{\\operatorname {h e i g h t} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right)}} \\\\ \\cdot \\text {p a t c h e s} _ {\\mathrm {h}} \\text {s i d e} = \\sqrt {\\frac {\\operatorname {n u m} . \\operatorname {p a t c h e s} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right) \\cdot \\operatorname {h e i g h t} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right)}{\\operatorname {w i d t h} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right)}} \\\\ \\bullet \\mathrm {x} = \\operatorname {l i n s p a c e} \\left(\\mathbf {x} _ {\\min }, \\mathbf {x} _ {\\max }, \\text {p a t c h e s - w - s i d e}\\right) \\\\ \\cdot \\mathbf {y} = \\operatorname {l i n s p a c e} \\left(\\mathbf {y} _ {\\min }, \\mathbf {y} _ {\\max }, \\text {p a t c h e s} _ {\\mathrm {h}} \\text {s i d e}\\right) \\\\ \\cdot P _ {j} = \\text {m e s h g r i d} (x, y) \\\\ \\bullet \\mathbf {S} _ {\\mathrm {j}} = \\left(\\frac {\\operatorname {w i d t h} (\\mathbf {b} _ {\\mathrm {j}})}{\\text {p a t c h e s . w . s i d e}}, \\frac {\\operatorname {h e i g h t} (\\mathbf {b} _ {\\mathrm {j}})}{\\text {p a t c h e s . h . s i d e}}\\right). \\exp a n d (\\mathbf {n} _ {\\mathrm {j}}, 2) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.621, + 0.539, + 0.631 + ], + "angle": 0, + "content": "end" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.637, + 0.753, + 0.648 + ], + "angle": 0, + "content": "3. Remove redundant patches" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.649, + 0.895, + 0.661 + ], + "angle": 0, + "content": "P,S \\(\\leftarrow\\) Remove_overlapedpatches(Concat(Pj),Concat(Sj))" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.662, + 0.6, + 0.671 + ], + "angle": 0, + "content": "return P and S" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.799, + 0.713 + ], + "angle": 0, + "content": "3.3. Subject-Oriented Data Curation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To fine-tune our TEVA model for dynamically sampled patches, we introduce a data curation pipeline that creates subject-oriented training data, emphasizing specific regions of interest within images. While large datasets [3, 51] are commonly used to enhance the generalization of VLMs, they typically consist of image-caption pairs providing broad, whole-image descriptions. Our approach curates question-answer pairs that emphasize zoomed-in regions, enabling the model to better understand and prioritize detailed image areas. To address this, we developed an efficient method that curates large-scale, subject-oriented conversations (8 million pairs) based on existing image-caption" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24150" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.096, + 0.196, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.176, + 0.195, + 0.193 + ], + "angle": 0, + "content": "The image is only for illustration. The images for subject-oriented VCA generation." + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.097, + 0.474, + 0.198 + ], + "angle": 0, + "content": "Image Caption: The image portrays a fairy with white wings and a purple dress, standing in a field of lavender flowers. She is holding a lavender plant in her hands. The background is a light blue color, and there are two butterflies visible in the scene. The fairy is positioned in the center of the image, with the lavender plants surrounding her. The butterflies are located in the top left and bottom right corners of the image. The fairy's wings are spread out, and her dress is flowing, giving the impression of movement. The lavender plants are tall and thin, with small purple flowers at the top. The butterflies are small and white, with black spots on their wings. The overall color scheme of the image is purple, white, and light blue." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.205, + 0.128, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.206, + 0.471, + 0.232 + ], + "angle": 0, + "content": " Generate three question-answer pairs based on this image description Each question should ask about one different object in the image without repeating objects." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.241, + 0.13, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.258, + 0.131, + 0.266 + ], + "angle": 0, + "content": "Llama" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.242, + 0.471, + 0.259 + ], + "angle": 0, + "content": "1. Question: what is the fairy wearing? Answer: The fairy is wearing a purple dress. \n2. Question: How many butterflies are there? Answer: There are two." + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.259, + 0.47, + 0.276 + ], + "angle": 0, + "content": "3. Question: What do the lavender plants look like? Answer: They are tall and thin, with small purple flowers at the top." + }, + { + "type": "list", + "bbox": [ + 0.139, + 0.242, + 0.471, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.294, + 0.484, + 0.351 + ], + "angle": 0, + "content": "Figure 4. Example of constructing Subject-Oriented QA pairs. Using the image captions, we employ the Llama3.1 [59] to generate an average of three questions per image, resulting in 8 million QA pairs focused on different objects described in each caption." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.377, + 0.485, + 0.558 + ], + "angle": 0, + "content": "datasets [3, 4, 11]. As illustrated in Figure 4, we leverage Llama 3.1 [59] for its strong contextual language reasoning ability, using only structured text prompts as input to generate subject-focused conversations. After generating these conversations, we apply an open-vocabulary detection method to validate the presence of specified objects in the images, filtering out any noisy data and ensuring data quality. Experiments in Table 4 demonstrate the significance of this pretraining dataset, especially for challenging benchmarks that require an understanding of zoomed-in patches or complex reasoning tasks. The details of data curation are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.572, + 0.224, + 0.589 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.703 + ], + "angle": 0, + "content": "In this section, we provide the training settings and evaluation benchmarks employed to assess our model's performance on high-resolution and general Vision-Language Model (VLM) tasks. Subsequently, we present comparative results, benchmarking our method against state-of-the-art approaches. Further details are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.282, + 0.729 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Training stages The goal of TEVA is to enhance the model's ability to process high-resolution images by focusing on informative patches. Two key requirements are essential for achieving this: (1) enabling the vision tower to handle dynamic patch sampling, and (2) ensuring that dynamic vision tokens can be effectively interpreted by the LLM. To meet these goals, we structured our training pipeline into four stages:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- Alignment Stage: Following the process used in LLaVA [37], we begin by aligning visual features with the language space. In this stage, we freeze the large" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.907, + 0.152 + ], + "angle": 0, + "content": "language model (LLM) and optimize only the vision-language (VL) adapter with LLaVA alignment dataset (595k), ensuring that the visual features are properly mapped to the language domain." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.907, + 0.228 + ], + "angle": 0, + "content": "- Vision Tower Adapting: Using 4M samples from our proposed Subject-Oriented Dataset, we fine-tune the vision tower to handle dynamic image patches. During this stage, both the vision tower and VL adapter are unfrozen for training." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.228, + 0.906, + 0.332 + ], + "angle": 0, + "content": "- Integrated Interpretation Stage: To further enhance the LLM's ability to interpret dynamic vision tokens, we unfreeze the vision tower, VL adapter, and LLM, then fine-tune using 4M samples from the Subject-Oriented dataset and 1M samples from the ShareGPT4V [4] caption dataset. This step aligns all components for seamless interpretation of dynamic visual information." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.333, + 0.906, + 0.455 + ], + "angle": 0, + "content": "- Instruction Fine-tuning: In the final stage, following the training strategy of VILA [37], we continue fine-tuning the vision tower, VL adapter, and LLM with a mix of Supervised Fine-Tuning (SFT) Dataset (4.17M) built on VILA [37]. This instruction fine-tuning stage enhances the model's instruction-following and dialogue capabilities on multiple tasks. Details of the training settings and datasets are provided in the supplementary materials." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.153, + 0.907, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.907, + 0.622 + ], + "angle": 0, + "content": "Evaluation benchmarks To assess our model's capability in handling complex high-resolution images, we report results on challenging high-resolution benchmarks, including \\( V^{*} \\) [61] and MME-RealWorld [79]. Additionally, we evaluate our model on commonly used general VQA and MLLM benchmarks including MMBench [41], SQA (ScienceQA) [43], \\( \\mathrm{VQAT} \\) (TextVQA) [54], POPE [29], AI2D [22], LLaVAw (LLaVA-Bench In-the-Wild) [37], \\( \\mathrm{MMMU}^{\\mathrm{t}} \\) (MMMU-test) [69], and \\( \\mathrm{VQAV}^{\\mathrm{v2}} \\) (VQA-v2) [14]. Evaluation of more benchmarks is available in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.636, + 0.654, + 0.65 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Comparison with region-aware VLMs. We evaluate Token-Efficient Vision LLanguage Model (TEVA) on the high-resolution visual detail benchmark \\(V^{*}\\) [61], comparing its performance and efficiency against state-of-the-art region-aware VLMs. As shown in Table 1, TEVA consistently outperforms existing methods in attribute recognition (AR), spatial relationship reasoning (SRR), and scene OCR tasks. Our dynamic patch sampling strategy significantly reduces computational overhead compared to SpatialGPT (mask pooling) [6], Griffon-G [72] and \\(V^{*}\\) [61], which require multiple vision tower passes to process individual ROIs. Specifically, TEVA achieves a \\(7.53 \\times\\) speedup over \\(V^{*}\\) in ROI aggregation and improves overall inference efficiency by \\(12.31 \\times\\) compared to \\(V^{*}\\) [61], benefiting from both token efficiency and the lightweight RAP module, which eliminates the need for iterative region searching." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24151" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.096, + 0.354, + 0.11 + ], + "angle": 0, + "content": "How many people are there on the street?" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.112, + 0.356, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.108, + 0.248, + 0.153, + 0.257 + ], + "angle": 0, + "content": "VILA-1.5" + }, + { + "type": "image_footnote", + "bbox": [ + 0.182, + 0.242, + 0.335, + 0.264 + ], + "angle": 0, + "content": "There is only one person on the street." + }, + { + "type": "image_footnote", + "bbox": [ + 0.113, + 0.274, + 0.151, + 0.283 + ], + "angle": 0, + "content": "Monkey" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.293, + 0.141, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.102, + 0.309, + 0.16, + 0.317 + ], + "angle": 0, + "content": "ChatGPT-40" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.32, + 0.143, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.177, + 0.296, + 0.34, + 0.317 + ], + "angle": 0, + "content": "There is one person visible on the street." + }, + { + "type": "image_footnote", + "bbox": [ + 0.227, + 0.328, + 0.291, + 0.337 + ], + "angle": 0, + "content": "There are 2." + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.097, + 0.581, + 0.111 + ], + "angle": 0, + "content": "What is the breed of the dog?" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.112, + 0.628, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.379, + 0.248, + 0.424, + 0.257 + ], + "angle": 0, + "content": "VILA-1.5" + }, + { + "type": "image_footnote", + "bbox": [ + 0.453, + 0.243, + 0.605, + 0.266 + ], + "angle": 0, + "content": "The dog in the image is a white dog." + }, + { + "type": "image_footnote", + "bbox": [ + 0.381, + 0.274, + 0.42, + 0.284 + ], + "angle": 0, + "content": "Monkey" + }, + { + "type": "image_footnote", + "bbox": [ + 0.517, + 0.276, + 0.539, + 0.286 + ], + "angle": 0, + "content": "Pug." + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.294, + 0.432, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.466, + 0.302, + 0.609, + 0.313 + ], + "angle": 0, + "content": "difficult to identify the breed." + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.32, + 0.413, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.482, + 0.328, + 0.576, + 0.337 + ], + "angle": 0, + "content": "Golden Retriever." + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.097, + 0.85, + 0.11 + ], + "angle": 0, + "content": "What color is the bottle cap?" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.112, + 0.816, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.112, + 0.898, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.649, + 0.248, + 0.694, + 0.257 + ], + "angle": 0, + "content": "VILA-1.5" + }, + { + "type": "image_footnote", + "bbox": [ + 0.74, + 0.248, + 0.862, + 0.259 + ], + "angle": 0, + "content": "The bottle has a blue cap." + }, + { + "type": "image_footnote", + "bbox": [ + 0.653, + 0.275, + 0.69, + 0.283 + ], + "angle": 0, + "content": "Monkey" + }, + { + "type": "image_footnote", + "bbox": [ + 0.78, + 0.275, + 0.819, + 0.284 + ], + "angle": 0, + "content": "Yellow." + }, + { + "type": "image_footnote", + "bbox": [ + 0.643, + 0.309, + 0.701, + 0.318 + ], + "angle": 0, + "content": "ChatGPT-40" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.294, + 0.683, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.719, + 0.296, + 0.883, + 0.316 + ], + "angle": 0, + "content": "The color of the cap appears to be red." + }, + { + "type": "image_footnote", + "bbox": [ + 0.778, + 0.328, + 0.823, + 0.337 + ], + "angle": 0, + "content": "It is red." + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.359, + 0.299, + 0.37 + ], + "angle": 0, + "content": "What is the speed limit?" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.371, + 0.161, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.163, + 0.371, + 0.356, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.108, + 0.507, + 0.152, + 0.516 + ], + "angle": 0, + "content": "VILA-1.5" + }, + { + "type": "image_footnote", + "bbox": [ + 0.223, + 0.509, + 0.288, + 0.517 + ], + "angle": 0, + "content": "It is \\(10\\mathrm{km / h}\\)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.11, + 0.534, + 0.148, + 0.543 + ], + "angle": 0, + "content": "Monkey" + }, + { + "type": "image_footnote", + "bbox": [ + 0.248, + 0.535, + 0.264, + 0.544 + ], + "angle": 0, + "content": "50." + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.553, + 0.141, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.102, + 0.568, + 0.159, + 0.576 + ], + "angle": 0, + "content": "ChatGPT-40" + }, + { + "type": "image", + "bbox": [ + 0.121, + 0.581, + 0.141, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.177, + 0.561, + 0.335, + 0.572 + ], + "angle": 0, + "content": "The speed limit appears to be 30." + }, + { + "type": "image_footnote", + "bbox": [ + 0.235, + 0.587, + 0.276, + 0.596 + ], + "angle": 0, + "content": "It is 40." + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.359, + 0.61, + 0.37 + ], + "angle": 0, + "content": "What is the written on the large boat?" + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.371, + 0.628, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.381, + 0.508, + 0.426, + 0.517 + ], + "angle": 0, + "content": "VILA-1.5" + }, + { + "type": "image_footnote", + "bbox": [ + 0.47, + 0.509, + 0.59, + 0.518 + ], + "angle": 0, + "content": "\"Oranje\" is written on it." + }, + { + "type": "image_footnote", + "bbox": [ + 0.381, + 0.534, + 0.421, + 0.544 + ], + "angle": 0, + "content": "Monkey" + }, + { + "type": "image_footnote", + "bbox": [ + 0.494, + 0.535, + 0.565, + 0.544 + ], + "angle": 0, + "content": "Unanswerable." + }, + { + "type": "image", + "bbox": [ + 0.37, + 0.555, + 0.433, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.467, + 0.556, + 0.592, + 0.576 + ], + "angle": 0, + "content": "The text on the boat reads \"INTERCOMBI\"." + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.581, + 0.415, + 0.6 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.509, + 0.588, + 0.549, + 0.597 + ], + "angle": 0, + "content": "Nitrico." + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.359, + 0.881, + 0.37 + ], + "angle": 0, + "content": "Tell me the yellow text on the airship." + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.371, + 0.898, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.649, + 0.508, + 0.694, + 0.517 + ], + "angle": 0, + "content": "VILA-1.5" + }, + { + "type": "image_footnote", + "bbox": [ + 0.775, + 0.509, + 0.826, + 0.518 + ], + "angle": 0, + "content": "\"Airship\"." + }, + { + "type": "image_footnote", + "bbox": [ + 0.654, + 0.535, + 0.693, + 0.544 + ], + "angle": 0, + "content": "Monkey" + }, + { + "type": "image_footnote", + "bbox": [ + 0.765, + 0.535, + 0.836, + 0.544 + ], + "angle": 0, + "content": "Unanswerable." + }, + { + "type": "image", + "bbox": [ + 0.666, + 0.555, + 0.686, + 0.569 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.646, + 0.57, + 0.704, + 0.578 + ], + "angle": 0, + "content": "ChatGPT-40" + }, + { + "type": "image_footnote", + "bbox": [ + 0.745, + 0.556, + 0.856, + 0.576 + ], + "angle": 0, + "content": "The text appears to say \"GOODYEAR\"." + }, + { + "type": "image", + "bbox": [ + 0.664, + 0.581, + 0.686, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.746, + 0.582, + 0.855, + 0.602 + ], + "angle": 0, + "content": "The yellow text reads \"GoodYear\"." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.621, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Figure 5. Evaluation of our model against leading open-source VLMs (VILA-1.5 [32], Monkey [31]) and the advanced commercial VLM, GPT-4o [50], in high-resolution, real-world scenarios. Each image displays the proposed subject in the bottom left corner. The first row presents responses to general questions, while the second row addresses more challenging OCR-based questions. The proposed subject is in the left-bottom corner of each image. Despite using a smaller language model (Sheared-LLaMA-2.7B [62]), our model demonstrates robust capability in observing precise details for high-resolution images and can adapt to various image resolutions, providing flexibility and enhanced performance in demanding high-resolution scenarios. Please note that all sample images were processed at their original resolution \\((2246\\times 1582\\) on average) for TEVA-3B." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.483, + 0.806 + ], + "angle": 0, + "content": "Unlike SpatialGPT [6] and Griffon-G [72], which rely on manual region selection, TEVA automatically detects and proposes ROIs. For a fair comparison, we use TEVA's RAP module to predict ROIs for SpatialGPT and Griffon-G." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "High resolution benchmark results. To demonstrate TEVA's capability in handling high-resolution images, we evaluated our method on high-resolution VLM benchmarks: \\(V^{*}\\) [61], with an average image resolution of \\(2246 \\times 1582\\) pixels and target subjects, and MME-RealWorld [79], with an average resolution of \\(2000 \\times 1500\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.907, + 0.897 + ], + "angle": 0, + "content": "As shown in Table 2, TEVA excels in tasks that require precise detail recognition, such as attribute recognition (AR) and spatial relationship reasoning (SRR) in \\( V^{*} \\) benchmark, outperforming both the commercial model GPT-4V [50] by 21.6 points and all the other open-source high resolution VLMs. In the complex MME-RealWorld benchmark, TEVA also surpasses open-source methods in both perception and reasoning ability. In applications that demand high-resolution fine visual detail perception, such as remote sensing, video monitoring, and autonomous driv" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24152" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.09, + 0.361, + 0.288 + ], + "angle": 0, + "content": "
MethodV* Bench
ARSRROCR
SpatialGPT [6]36.556.540.0
Griffon-G [72]44.771.056.6
V* [61]74.776.380.0
TEVA-7B76.577.683.3
MethodInference Time
AIPFPI
SpatialGPT [6]9.57×1.05×
Griffon-G [72]10.24×4.83×
V* [61]7.53×12.31×
TEVA-7B1.00×1.00×
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.293, + 0.374, + 0.446 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art region-aware 7B VLMs on the high-resolution \\(V^{*}\\) benchmark, evaluating attribute recognition (AR), spatial relationship reasoning (SRR), and scene OCR tasks. Inference times are reported for Additional Information Processing (AIP) and Full Pipeline Inference (FPI), demonstrating the efficiency and performance gains of dynamic patch sampling over SpatialGPT(pooling), Griffon-G and \\(V^{*}\\) (multiple vision tower passes)." + }, + { + "type": "table", + "bbox": [ + 0.398, + 0.09, + 0.888, + 0.288 + ], + "angle": 0, + "content": "
MethodV* AverageMME-RWPRCMME-RWRSN
OCRRSMOADOCRMOAD
GPT-4V [50]55.9-------
GPT-4o [50]-77.728.933.922.461.436.526.4
Gemini Pro [56]50.5-------
Gemini-1.5-pro [56]-67.614.031.126.652.728.319.2
LLaVA-1.5 [35]50.538.722.119.126.026.025.924.2
LLaVA-1.5-S2[52]56.5-------
Text-Monkey [40]-37.311.716.114.330.44.420.0
Monkey [31]-54.625.028.029.727.227.333.0
LLaVA-Next [36]61.340.631.635.427.214.929.528.2
Cambrian-1 [57]60.832.735.827.635.527.531.231.2
MLLM-K [75]62.3-------
TEVA-3B61.643.132.332.633.041.833.529.9
TEVA-7B77.160.738.236.935.249.434.735.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.381, + 0.293, + 0.908, + 0.446 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art open-source and commercial methods on the high-resolution benchmark \\( V^{*} \\) [61] (averaging attribute recognition(AR) and spatial relationship reasoning(SRR)), as well as on MME-RealWorld [79], which includes perception benchmark MME-RW\\(_{PRC}\\) and reasoning benchmark MME-RW\\(_{RSN}\\). Task categories are abbreviated due to space limits: OCR (Optical Character Recognition), RS (Remote Sensing), MO (Video Monitoring), AD (Autonomous Driving). Our model consistently outperforms other open-source methods on Reasoning tasks such as OCR, RS, MO, AR and SRR. It even surpasses leading commercial VLMs, GPT-4o and Gemini-1.5-pro, on RS, MO, and AD perception tasks that require precise visual grounding. The best results are bold. Notably, all samples are evaluated at their original image resolution for TEVA and at the largest default resolution for other methods." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.46, + 0.905, + 0.639 + ], + "angle": 0, + "content": "
MethodLLM#Visual TokenMMBSQAVQATPOPEAI2DLLaVAwMMMUVVQAV2
InstructBLIP [7]Vicuna-7B-36.063.150.7--60.9--
IDEFICS-9B [18]LLAMA-7B-48.2------50.9
LLaVA-1.5 [35]Vicuna-7B57664.366.866.885.9-63.435.778.5
MLLM-K [75]Vicuna-7B---56.187.3---76.5
LLaVA-1.5-S2[52]Vicuna-7B57666.2-----37.780.0
VILA-1.5 [32]Llama-2-7B72968.968.268.285.5-69.738.679.9
Mini-Gemini-HD [30]Vicuna-7B288072.775.168.4-73.5-37.3-
Monkey [31]Qwen-7B5120-69.467.6-62.6--80.3
LLaVA-Next [36]LLama3-8B288072.172.864.686.771.683.241.781.8
Cambrian-1 [57]Llama3-8B-Instruct230475.980.471.7-73.0-42.781.2
TEVA-3BSheared-llama-2.7B72964.275.766.887.064.985.032.480.3
TEVA-7BQwen2-7B72980.392.572.587.976.394.348.482.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.644, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Table 3. Comparison with leading methods on 8 commonly used VLM benchmarks: MMBench [41], SQA: ScienceQA [43], VQATextVQA [54], POPE [29], AI2D [22], LLaVAw: LLaVA-Bench (In-the-Wild) [37], MMMUv: MMMU-validation [69], VQAV2: VQAV2 [14]. Our model consistently outperforms the other methods without image tiling and increasing visual tokens. The best results are bold. Please note that all the samples are evaluated on original image resolution for our model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.482, + 0.768 + ], + "angle": 0, + "content": "ing (e.g., identifying traffic signs), TEVA demonstrates robust performance and even surpassing the most powerful commercial models, GPT-4o [50] and Gemini-1.5-pro [56]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.483, + 0.88 + ], + "angle": 0, + "content": "General resolution benchmark results. We also evaluate our model on eight commonly used VQA and MLLM benchmarks, comparing it with state-of-the-art VLMs, as shown in Table 3. The results demonstrate that TEVA consistently outperforms other leading methods across all the benchmarks and achieves this without requiring image tiling or visual token extending." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Notably, on general resolution subject-oriented QA tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.782 + ], + "angle": 0, + "content": "like MMBench [41], AI2D [22] and \\(\\mathrm{MMMU}^{\\mathrm{V}}\\) [69], TEVA achieves superior results compared to all other methods. This highlights the generalizability of our approach in focusing on contextually important areas within an image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative Results. In Figure 5, we present qualitative results that demonstrate the effectiveness of TEVA-3B in challenging high-resolution scenarios requiring precise detail recognition. By leveraging our informative patch sampling strategy, TEVA dynamically focuses on relevant, information-rich regions based on the user query, enabling inference on images at arbitrary resolutions without increas" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24153" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.122, + 0.094, + 0.451, + 0.106 + ], + "angle": 0, + "content": "Results with Different Dynamic and Static Token Numbers" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.108, + 0.477, + 0.17 + ], + "angle": 0, + "content": "
DynamicStaticV*MRWPOPEVQAT
025645.030.083.044.5
808150.431.585.654.2
47325661.634.087.066.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.159, + 0.173, + 0.413, + 0.185 + ], + "angle": 0, + "content": "Comparison of w/o Subject-Orient (SO) Data" + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.187, + 0.476, + 0.238 + ], + "angle": 0, + "content": "
SO DataMRWPRCMRWRSNPOPEVQAT
X33.128.586.558.7
✓33.834.387.066.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.16, + 0.24, + 0.412, + 0.252 + ], + "angle": 0, + "content": "Comparison of w/o Matrix Embedding (ME)" + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.254, + 0.476, + 0.306 + ], + "angle": 0, + "content": "
MEV*ARV*SRRMRWPOPE
X48.350.031.884.8
✓60.063.234.087.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.312, + 0.483, + 0.356 + ], + "angle": 0, + "content": "Table 4. Ablation Studies on Token Numbers, w/o Subject-Orient Pretraining and w/o Affine Transformation Matrix Embedding. Benchmark abbreviation MRW: MME-RealWorld [79]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.371, + 0.483, + 0.598 + ], + "angle": 0, + "content": "ing computational cost. For example, processing a 5760 \\(\\times\\) 2880 (5K-resolution) image takes only 2.5 seconds on a single NVIDIA A100 GPU with TEVA-7B. This efficiency enables TEVA to handle high-resolution images seamlessly, outperforming leading open-source VLMs, such as VILA1.5-7B [32], which uses the same number of visual tokens (729), and Monkey-7B [31], which relies on image tiling for high-resolution support. Furthermore, our approach surpasses the advanced commercial VLM, ChatGPT-4o [50], in precise detail observation for high-resolution images, despite the latter's use of a more powerful language model. The evaluation covers real-world, high-difficulty OCR tasks along with diverse general queries, such as counting, identification, and attribute recognition, providing a comprehensive assessment of TEVA's capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.245, + 0.623 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Dynamic vs. Static Token Numbers. We assess the impact of using dynamic, informative image patches compared to static, uniformly sampled patches. As shown in Table 4, adding 473 dynamic patches to the baseline of 256 static patches (729 tokens in total) leads to significant improvements in tasks requiring detailed structural understanding, such as scene OCR tasks in \\(\\mathrm{VQA}^{\\mathrm{T}}\\) (TextVQA) [54], as well as in handling high-resolution images, as demonstrated on the \\(V^{*}\\) [61] benchmark. Additionally, we conducted experiments using a reduced set of 144 patches (80 dynamic and 81 static). Despite the lower patch count, incorporating dynamic tokens consistently enhanced performance on challenging datasets that demand high-resolution or precise observation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Subject-Oriented Pretraining. To evaluate the benefits of pretraining on our Subject-Oriented (SO) Dataset, we compare results with and without the Subject-Oriented Pretraining phase. As illustrated in Table 4, pretrain" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "ing on a large-scale Subject-Oriented Dataset notably improves the model's complex reasoning abilities (MMERealWorld\\(_{RSN}\\): MME-RealWorld-Reasoning [79]) and scene understanding, especially in tasks requiring OCR, such as \\(\\mathrm{VQAT}\\) (TextVQA) [54]. However, for tasks requiring more straightforward scene comprehension, such as basic perception (MME-RealWorld\\(_{PRC}\\): MME-RealWorld-Perception [79]) or hallucination assessments (POPE [29]), the improvements from Subject-Oriented Pretraining are less pronounced. These findings demonstrate the effectiveness of Subject-Oriented Pretraining in enhancing the model's capacity for complex visual understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Affine Transformation Matrix Embeddings. We examine the effectiveness of embedding the affine transformation matrix to provide positional information for each patch. Results in Table 4 demonstrate that the Affine Transformation Matrix Embeddings (ME) consistently improve performance across various benchmarks, particularly enhancing the model's ability to comprehend the relative positions of objects in the \\( V_{SRR}^{*} \\) (\\( V^{*} \\) Spatial Relationship reasoning) benchmark. This aligns well with our intuition, as the additional spatial context helps the model maintain awareness of object locations and relationships." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.449, + 0.634, + 0.463 + ], + "angle": 0, + "content": "4.4. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.622 + ], + "angle": 0, + "content": "While TEVA can process images at arbitrary resolutions without increasing the token count, its effectiveness may be limited in scenarios involving densely packed high-resolution images, such as text-heavy documents. In these cases, where information is uniformly distributed across the image, there are no distinct regions to prioritize. Consequently, dynamic sampling must cover the entire image at a lower resolution, limiting the advantages of high-resolution focus on key areas. Additional experiments and discussion are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.634, + 0.65 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.842 + ], + "angle": 0, + "content": "In this work, we introduced Token-Efficient Vision LLanguage Model (TEVA), a straightforward and token-efficient framework that enhances Vision-Language Models (VLMs) for high-resolution image processing. TEVA employs a subject-oriented strategy and Dynamic Patch Sampling to balance fine details with global context, supported by a specialized Subject-Oriented Dataset for improved learning. Comprehensive evaluations against state-of-the-art VLMs, including GPT-4o, show that TEVA excels in challenging benchmarks and holds strong potential for applications requiring high-resolution precision, such as remote sensing, autonomous driving, and video monitoring." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24154" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.4.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.482, + 0.354 + ], + "angle": 0, + "content": "[4] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[5] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[6] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062-135093, 2025. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.482, + 0.563 + ], + "angle": 0, + "content": "[7] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[8] Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuoling Yang, Zihan Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nvlm: Open frontier-class multimodal llms. arXiv preprint arXiv:2409.11402, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[9] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[11] Yunhao Fang, Ligeng Zhu, Yao Lu, Yan Wang, Pavlo Molchanov, Jang Hyun Cho, Marco Pavone, Song Han, and Hongxu Yin. vila2: Vila augmented vila. arXiv preprint arXiv:2407.17453, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "comprehensive evaluation benchmark for multimodal large language models. ArXiv, abs/2306.13394, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[13] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.19, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 5, 7, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[15] Qiushan Guo, Shalini De Mello, Hongxu Yin, Wonmin Byeon, Ka Chun Cheung, Yizhou Yu, Ping Luo, and Sifei Liu. Regionpt: Towards region understanding vision language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13796-13806, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[16] Jack Hessel, Jena D Hwang, Jae Sung Park, Rowan Zellers, Chandra Bhagavatula, Anna Rohrbach, Kate Saenko, and Yejin Choi. The abduction of sherlock holmes: A dataset for visual abductive reasoning. In European Conference on Computer Vision, pages 558-575. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[17] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[18] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[19] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[20] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901-2910, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[21] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.734, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[22] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 5, 7, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[23] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeong Yeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24155" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[24] Jonathan Krause, Justin Johnson, Ranjay Krishna, and Li Fei-Fei. A hierarchical approach for generating descriptive image paragraphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 317–325, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[25] Paul Lerner, Olivier Ferret, Camille Guinaudeau, Hervé Le Borgne, Romaric Besançon, José G Moreno, and Jesús Lovón Melgarejo. Viquae, a dataset for knowledge-based visual question answering about named entities. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 3108-3120, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[26] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[27] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[29] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 5, 7, 8, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.55 + ], + "angle": 0, + "content": "[30] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[31] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26763-26773, 2024. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[32] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 6, 7, 8, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[33] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[34] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. Mmc: Advancing multimodal chart understanding with large-scale instruction tuning. arXiv preprint arXiv:2311.10774, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[36] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.1,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NIPS, 2024. 1, 5, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[38] Haogeng Liu, Quanzeng You, Xiaotian Han, Yiqi Wang, Bohan Zhai, Yongfei Liu, Yunzhe Tao, Huaibo Huang, Ran He, and Hongxia Yang. Infimm-hd: A leap forward in high-resolution multimodal understanding. arXiv preprint arXiv:2403.01487, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[39] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[40] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[41] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 5, 7, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[42] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, Xiuyu Li, Yunhao Fang, Yukang Chen, Cheng-Yu Hsieh, De-An Huang, An-Chieh Cheng, Vishwesh Nath, Jinyi Hu, Sifei Liu, Ranjay Krishna, Daguang Xu, Xiaolong Wang, Pavlo Molchanov, Jan Kautz, Hongxu Yin, Song Han, and Yao Lu. Nvila: Efficient frontier visual language models, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[43] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5, 7, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[44] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[45] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.907, + 0.814 + ], + "angle": 0, + "content": "[46] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[47] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24156" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[49] Ravina Mithe, Supriya Indalkar, and Nilam Divekar. Optical character recognition. International journal of recent technology and engineering (IJRTE), 2(1):72-75, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[50] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2(5), 2023. 1, 6, 7, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.483, + 0.292 + ], + "angle": 0, + "content": "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[52] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? In European Conference on Computer Vision, pages 444-462. Springer, 2025. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[53] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 742–758, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.483, + 0.49 + ], + "angle": 0, + "content": "[54] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 7, 8, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.483, + 0.575 + ], + "angle": 0, + "content": "[55] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. Wit: Wikipedia-based image text dataset for multimodal multilingual machine learning. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, pages 2443-2449, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[56] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[57] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[60] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[61] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084-13094, 2024. 2, 4, 5, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[62] Mengzhou Xia, Tianyu Gao, Zhiyuan Zeng, and Danqi Chen. Sheared llama: Accelerating language model pre-training via structured pruning. arXiv preprint arXiv:2310.06694, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[63] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794-4803, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[64] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[65] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[66] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[67] Xiaoyu Yue, Shuyang Sun, Zhanghui Kuang, Meng Wei, Philip HS Torr, Wayne Zhang, and Dahua Lin. Vision transformer with progressive sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 387-396, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[68] Xiang Yue, Xingwei Qu, Ge Zhang, Yao Fu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mammoth: Building math generalist models through hybrid instruction tuning. arXiv preprint arXiv:2309.05653, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[69] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567, 2024. 5, 7, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[70] Bohan Zhai, Shijia Yang, Chenfeng Xu, Sheng Shen, Kurt Keutzer, and Manling Li. Halle-switch: Controlling object hallucination in large vision language models. arXiv e-prints, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[71] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[72] Yufei Zhan, Hongyin Zhao, Yousong Zhu, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon-g: Bridging vision-language and vision-centric tasks via large multimodal models. arXiv preprint arXiv:2410.16163, 2024. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[73] Yufei Zhan, Yousong Zhu, Hongyin Zhao, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon v2: Advancing multimodal" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24157" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "perception with high-resolution scaling and visual-language co-referring. arXiv preprint arXiv:2403.09333, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[74] Haotian Zhang, Haoxuan You, Philipp Duffer, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[75] Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara, and Filip Ilievski. Mllms know where to look: Training-free perception of small visual details with multimodal llms. Proceedings of International Conference on Learning Representations (ICLR), 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[76] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Haodong Duan, Songyang Zhang, Shuangrui Ding, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[77] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.431, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[78] Yanzhe Zhang, Ruiyi Zhang, Jiuming Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[79] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, et al. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans? arXiv preprint arXiv:2408.13257, 2024. 5, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[80] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24158" + } + ] +] \ No newline at end of file diff --git a/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/b324faad-4fe9-44a7-a833-a82bac07835f_origin.pdf b/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/b324faad-4fe9-44a7-a833-a82bac07835f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8a932e311668ea23c3a47fd022d381f775809c42 --- /dev/null +++ b/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/b324faad-4fe9-44a7-a833-a82bac07835f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78fdacfe743e0f58673a6ff4d0bcc26333830dce94324e4aca6281e19dfb0206 +size 35723720 diff --git a/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/full.md b/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f9ee04aa4ffc1ddf894fadd96aa6fa1f3ab96e0a --- /dev/null +++ b/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/full.md @@ -0,0 +1,418 @@ +# Token-Efficient VLM: High-Resolution Image Understanding via Dynamic Region Proposal + +Yitong Jiang $^{1,2,\dagger}$ Jinwei Gu $^{1,2}$ Tianfan Xue $^{2}$ Ka Chun Cheung $^{1}$ Pavlo Molchanov $^{1}$ Hongxu Yin $^{1\dagger}$ Sifei Liu $^{1\dagger}$ $^{1}$ NVIDIA ${}^{2}$ The Chinese University of Hong Kong + +![](images/b9f8094134eef96cdae7f92e7dc7b7351220f5e5b3d420f27fb60a273388fe58.jpg) +Figure 1. Illustration of the Token-Efficient Vision LLanguage Model (TEVA) for high-resolution scenarios. Given a user query about a specific detail $(20\times 50\mathrm{px}$ region) in a high-resolution image $(5760\times 2880\mathrm{px})$ , traditional VLMs (Left) often miss critical details due to uniform patch processing and computational limits. TEVA (Right) addresses this by using a Relevant Area Proposal module to identify key subjects (e.g., people) and Dynamic Patch Sampling to focus on high-resolution patches of interest while retaining global context patches. This approach enables accurate interpretation of details at arbitrary resolutions without increasing the token count. + +![](images/76caf698d7b59308f0c8fd33c6975bf48e88287e18735ba1e06e37cd2681f745.jpg) + +![](images/8c451e37662a929f2eb525e829632f8fbaea2e883ee094318beda32abcd874f6.jpg) + +![](images/9ca2907859b78258310a8467c1209cce64ff9587719301615d11edaa65b2782c.jpg) + +# Abstract + +Vision-Language Models (VLMs) excel at visual understanding by leveraging pretrained image encoders to generate visual tokens. However, they struggle with high-resolution images and zoomed-in regions due to the computational burden and token redundancy of uniform patch-based processing, often leading to the loss of critical details. To address these challenges, we propose Token-Efficient Vision LLanguage Model (TEVA), a novel framework that detects key regions and applies dynamic patch sampling to efficiently capture fine-grained details while preserving global context. Our approach first identifies subject-oriented regions using an adaptive detection strategy. Then, a dynamic patch sampling mechanism selects and arranges patches at varying scales, ensuring efficient processing without increasing token count. Extensive experiments demonstrate that Token-Efficient Vision LLanguage Model (TEVA) significantly enhances VLM performance in handling visual details, seamlessly integrating with various decoders and LLMs. + +# 1. Introduction + +Vision-Language Models (VLMs) [1, 36, 37, 80] have recently received significant attention due to their ability to integrate visual understanding into Large Language Models (LLMs) [50, 59, 77], enhancing performance in tasks like image captioning and visual question answering. However, supporting applications that require precise attention to specific details—such as video surveillance analysis, satellite imagery interpretation, and scene OCR [49]—remains challenging. These tasks often necessitate zooming into fine-grained regions of high-resolution images to capture essential information, but many existing VLMs [1, 28, 37, 80] rely on low-resolution pretrained image encoders (e.g., $224 \times 224$ pixels), leading to the loss of important details [38, 52]. + +To address this issue, previous methods have focused on either increasing global resolution or applying region-aware techniques. Enhancing resolution in pretrained image encoders typically involves two strategies: scaling resolution via curriculum learning [2, 5, 73] and partitioning images into fixed-size sub-images for independent processing [9, 33, 35, 36, 38, 42, 52]. These approaches have notable limitations. Scaling resolution increases computational complexity quadratically, raising self-attention + +costs. Partitioning images boosts the number of visual tokens (e.g., over 5000 tokens for an $896 \times 896$ image [31]), leading to redundancy and degraded performance [35, 38]. Moreover, the lack of interaction among image tokens within the vision encoder limits the model's ability to capture contextual relationships, requiring specialized design interventions [31, 38]. + +Alternatively, region-aware methods [6, 15, 66, 72, 73] focus on fine-grained details by leveraging manually defined Regions of Interest (ROIs). However, these methods either compress visual tokens within target regions, potentially losing critical information [6, 15, 66], or require multiple independent passes through the vision encoder, significantly increasing computational cost and token usage [72, 73]. + +To address these limitations, we propose Token-Efficient Vision LLanguage Model (TEVA), which leverages a dynamic patch sampling framework for the vision encoder that balances visual token count and image resolution. Our primary concept involves informative patch sampling within the vision encoder, which selectively focuses on key parts of the image while preserving the understanding of the global context, represented in low resolution. As shown in Figure 1, unlike conventional vision encoders that uniformly partition the entire image into patches, our method makes an early determination of which image regions are essential for answering a given question and applies selective patch sampling before inputting patches into the vision encoder. This approach not only prevents an increase in the number of patches when processing high-resolution images of any size, but also reduces self-attention computation and minimizes tokens the LLM needs to process. Additionally, this approach maintains coherence among sampled image regions, enhancing the model's ability to understand the global context and structure of the input image [63, 67]. + +Our TEVA consists of three main steps. First, the Relevant Area Proposal (RAP) module identifies key locations and scales of interest using a subject proposal agent powered by an LLM, and a subject detection agent via a vision foundation model. This enables efficient sampling of informative patches as vision tokens. Next, patches from localized regions are arranged with low-resolution global patches, integrating positional information through an affine transformation strategy. This allows for sampling patches of varying sizes and locations without additional tokens or complex architectures [8, 31, 38, 64]. Finally, the mixture of local and global patch embeddings is processed in the vision encoder (SigLIP [71] in our work) to extract visual tokens for the Large Language Models (LLMs). + +Furthermore, to enhance dynamic patch sampling for VLMs, we introduce a data curation pipeline and curate a Subject-Oriented (SO) dataset containing 8 million subject-focused conversations. This is the first large-scale Subject + +Oriented VQA dataset for VLM pretraining, enabling a framework that focuses on specific image details. By integrating dynamic image tokens with early interactions between local and global visual elements, Token-Efficient Vision LLanguage Model (TEVA) achieves superior VLM performance across various benchmarks, particularly excelling in flexible-resolution scenarios, as shown in our comparative analysis in Section 4. + +# 2. Related Work + +Recent studies [58, 70, 78] highlight a bottleneck in Vision-Language Models (VLMs), where image resolution constraints limit the model's ability to capture fine-grained visual details. Previous approaches of handling high-resolution images in Vision-Language Models (VLMs) can be grouped into two main strategies: (1) increasing the global image resolution, (2) applying region-aware techniques, as discussed in the following. + +In the first strategy, methods like Qwen-VL [2] and Griffon v2 [73] employ curriculum learning and bilinear interpolation to gradually increase the resolution in the pretrained image encoder. However, a significant challenge with transformer-based image encoders [10] is their reliance on uniform patch sampling, which leads to a quadratic increase in computational complexity as image resolution grows. Other methods [30, 31, 33, 38, 64, 76] partition high-resolution images into smaller sub-images, treating each as an independent input to the vision encoder. While this approach enables the processing of higher-resolution images, it imposes a constraint on the input sequence length that the Large Language Model (LLM) can handle, thereby limiting its overall capacity. To mitigate the computational burden, learnable queries within the Perceiver architecture [31, 38] and pooling strategies [52, 73] have been introduced to reduce the number of visual tokens generated by the vision encoder from these sub-images. However, such compression techniques often lack selectivity, leading to the loss of fine-grained visual details that are crucial for optimal Multimodal Large Language Model (MLLM) performance. + +The second strategy involves explicitly identifying and processing Regions-of-Interest (ROIs). Some methods [6, 15, 66, 74] require manual specification of ROIs, followed by compression of low-resolution visual tokens. Others [72, 73] necessitate multiple passes through the vision encoder, increasing computational overhead. $V^{*}$ [61] further extends this idea by iteratively searching for ROIs through multiple passes through an additional Vision-Language Model (VLM). However, this iterative search is computationally expensive and also requires multiple forward passes through the vision encoder for each ROI candidate. + +In contrast to these methods, our approach automatically identifies and focuses on informative regions within high-resolution images. By applying dynamic patch sam + +![](images/631ab1c885928cca65fbfb51c07bd31086577eeca79d6879ff5df3ffe108b651.jpg) +Figure 2. The TEVA architecture. Given a user query, the Subject Proposal Agent identifies relevant subjects (e.g., "signs"). The Subject Detection Agent then locates all candidate signs in the image. Informative patch sampling extracts high-resolution local patches at these locations, which are concatenated with low-resolution global patches and fed into the VLM for final interpretation. + +pling before the vision encoder, we preserve essential details without resorting to compression, while also minimizing computational costs. This is achieved by skipping less informative patches and avoiding multiple passes through the vision encoder, resulting in a more efficient and detailed visual representation for VLMs. + +# 3. Methodology + +We show the overall flowchart of our proposed TokenEfficient Vision LLanguage Model (TEVA) in Figure 2. TEVA is a VLM framework designed to handle arbitrary high-resolution images by detecting and focusing on informative visual details within images. It incorporates a Relevant Area Proposal (RAP) module to identify and extract key regions from the input image. A targeted patch sampling strategy then partitions these selected areas into informative high-resolution local patches, while retaining low-resolution global patches for context. The concatenation of local and global patches is processed by the image encoder to generate visual tokens. Additionally, we introduce a Subject-Oriented (SO) dataset to enhance the VLM's capability to manage dynamic patch sampling and improve overall performance. Details are described below. + +# 3.1. Relevant Area Proposal Module + +Building on our insight that image tokens are essential for Vision-Language Models (VLMs) to generate accurate content descriptions and answer questions, we address a key challenge: as shown in Figure 3 (left), when VLMs process uniformly partitioned image patches, they often focus on prominent elements, such as the path, trees, and signs on the trees, while overlooking finer details, like the presence of two individuals in the forest. + +To overcome this, we propose an automatic Relevant Area Proposal (RAP) module to capture key local details that help the VLM focus on the most informative and relevant areas in response to user queries. Current open-vocabulary detection methods, such as Grounding + +![](images/a9073fbb70f73f59809c1fdb59add3299aebaba28c6e8ff54eab95d2e6e84d60.jpg) +What can you see in this image? +Assistant: A forest with tall trees and a trail in the middle. +Figure 3. Example of visual focus in Vision-Language Models (VLMs) with uniform image patches (Left) versus high-resolution, relevant patches (Right). In the uniform approach, the model emphasizes broad features like trees, trails, and visible signs, missing subtler details, such as the presence of people in the scene. By providing high-resolution relevant patches, the model can identify and interpret finer details. + +![](images/8111a0443b3cb485999879f5b82792f2b6ddf61921e925c2c9a09aa417fd0f46.jpg) +Assistant: A group of people are walking through a forest. + +Dino [39], have shown strong performance in detecting regions of interest based on straightforward text cues. However, user queries often involve complex questions or instructions without an explicit subject, which limits the applicability of open-vocabulary detection. + +To address this, we use two specialized agents: a Subject-Promise Agent using LLM, e.g., Llama3.1 [59], and a Subject Detection Agent via Grounding Dino. Given a user query, we employ a structured prompt (e.g., "You have limited vision of the image; tell me the subjects you need to observe to answer the question + User query") to guide the LLM in proposing candidate subjects for detection. Once identified, the Subject Detection Agent then locates the regions of interest (ROI) based on these suggested subjects. + +While this process effectively isolates the ROIs, integrating these fine-grained local details into the VLM remains challenging. To this end, we introduce the Informative Patch Sampling module in the following, which efficiently incorporates high-resolution information without increasing the token count or computational burden. + +# 3.2. Informative Patch Sampling + +Previous region-level VLMs [15, 66] handle regions of interest by relying on low-resolution features and apply pooling strategies to compress information within each region. Although effective for reducing computational load, these pooling methods often lead to a loss of critical details, especially for high-resolution images. Alternative approaches [61, 72, 73] treat each target region as a separate image that passes through a vision tower. However, this approach is computationally expensive and significantly increases the number of image tokens and inference time. To better leverage high-resolution image information without incurring excessive computational costs, we propose a novel approach that utilizes informative patch sampling within each high-resolution target region. This technique preserves fine-grained details while maintaining computational efficiency and the same token count comparable to a low-resolution image encoder, as demonstrated in Table 1. + +Patch allocation. As shown in Algorithm 1, our method begins by allocating informative patch numbers to regions of interest based on their relative sizes. Next, we align the assigned patches within each bounding box, ensuring even coverage across each target area. Finally, we remove overlapping bounding boxes to maintain a total of $n$ , yielding the positions $P \in \mathbb{R}^{n \times 2}$ and scaling factors $S \in \mathbb{R}^{n \times 2}$ . + +Patch sampling. Given the positions and the scaling factors, to link these positional clues to exact image patches, we embed $\mathsf{P} = \{(p_i^x,p_i^y)|i\in (0,n - 1)\}$ and $S = \{(s_i^x,s_i^y)|i\in (0,n - 1)\}$ into an affine transformation matrix to project image pixels into patches: + +$$ +A = \left[ \begin{array}{c c c} s _ {i} ^ {x} & 0 & p _ {i} ^ {x} \\ 0 & s _ {i} ^ {y} & p _ {i} ^ {y} \end{array} \right]. \tag {1} +$$ + +With this setup, the correspondence between spatial coordinates of patches and input image can now be defined as: + +$$ +c = A \left[ \begin{array}{l l l} u & v & 1 \end{array} \right] ^ {T}, \tag {2} +$$ + +where $c = [c_u, c_v]$ are the spatial coordinates in the original image, while $u$ and $v$ are coordinates in the patch space, sampled on a uniform grid. Through sampling patch spatial coordinates with Eq. 2 we can determine the corresponding coordinates in the original image. Since $c$ is not discrete, proper interpolation is necessary. Inspired by Spatial Transformer Networks [19], we employ bilinear interpolation to address this: + +$$ +\begin{array}{l} c _ {u, v} = \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} I _ {(h, w)} \max (0, 1 - | c _ {v} - h |) \tag {3} \\ \times \max (0, 1 - | c _ {u} - w |), \\ \end{array} +$$ + +where $I_{(h,w)}$ represents the pixel value in the original image $I$ of size $H \times W$ at the location $(h,w)$ . For a patch of size $\hat{H} \times \hat{W}$ , the coordinates are sampled as $u = \{1,2,\dots,\hat{H}\}$ ; $v = \{1,2,\dots,\hat{W}\}$ . Unlike the original integration-based approach [19], which is time-consuming for high-resolution tasks, we use a look-up operation to accelerate by considering only the four neighboring pixels around each patch coordinate. + +To incorporate spatial information for each patch, we concatenate the affine transformation matrix embeddings with each patch's positional embedding. Experimental results in Table 4 demonstrate the effectiveness of this approach in enhancing the model's spatial awareness. + +# Algorithm 1 Simplified Informative Patch Sampling + +Input: + +```txt +Bounding boxes $\mathsf{B} = \{b_1,\dots ,b_{\mathrm{box\_num}}\} \in \mathbb{R}^{\mathrm{box\_num}\times 4},$ Each bounding box $b_{j} = [x_{min},y_{min},x_{max},y_{max}],$ Number of local patches $n$ +``` + +Output: + +```latex +Sampled patch locations $\mathsf{P}\in \mathbb{R}^{n\times 2}$ Sampled patch scaling factors $\mathsf{S}\in \mathbb{R}^{n\times 2}$ +``` + +```latex +1. Allocate patches Based on Area area $(\mathsf{b_j}) = (\mathbf{x}_{\mathrm{max}} - \mathbf{x}_{\mathrm{min}})(\mathbf{y}_{\mathrm{max}} - \mathbf{y}_{\mathrm{min}})$ $\forall \mathsf{b_j}\in \mathsf{B}$ n=j = num-patches(bj) = area(bj) n, $\forall \mathsf{b_j}\in \mathsf{B}$ +``` + +2. Generate patch locations and scales for $\mathbf{b}_{\mathrm{j}}\in \mathsf{B}$ do + +$$ +\begin{array}{l} \cdot \operatorname {w i d t h} (\mathfrak {b} _ {\mathrm {j}}) = \mathbf {x} _ {\max } - \mathbf {x} _ {\min }, \operatorname {h e i g h t} (\mathfrak {b} _ {\mathrm {j}}) = \mathbf {y} _ {\max } - \mathbf {y} _ {\min } \\ \cdot \text {p a t c h e s} _ {\mathrm {w}}. \text {s i d e} = \sqrt {\frac {\operatorname {n u m} . \operatorname {p a t c h e s} \left(\mathbf {b} _ {\mathrm {j}}\right) \cdot \operatorname {w i d t h} \left(\mathbf {b} _ {\mathrm {j}}\right)}{\operatorname {h e i g h t} \left(\mathbf {b} _ {\mathrm {j}}\right)}} \\ \cdot \text {p a t c h e s} _ {\mathrm {h}} \text {s i d e} = \sqrt {\frac {\operatorname {n u m} . \operatorname {p a t c h e s} \left(\mathbf {b} _ {\mathrm {j}}\right) \cdot \operatorname {h e i g h t} \left(\mathbf {b} _ {\mathrm {j}}\right)}{\operatorname {w i d t h} \left(\mathbf {b} _ {\mathrm {j}}\right)}} \\ \bullet \mathrm {x} = \operatorname {l i n s p a c e} \left(\mathbf {x} _ {\min }, \mathbf {x} _ {\max }, \text {p a t c h e s - w - s i d e}\right) \\ \cdot \mathbf {y} = \operatorname {l i n s p a c e} \left(\mathbf {y} _ {\min }, \mathbf {y} _ {\max }, \text {p a t c h e s} _ {\mathrm {h}} \text {s i d e}\right) \\ \cdot P _ {j} = \text {m e s h g r i d} (x, y) \\ \bullet \mathbf {S} _ {\mathrm {j}} = \left(\frac {\operatorname {w i d t h} (\mathbf {b} _ {\mathrm {j}})}{\text {p a t c h e s . w . s i d e}}, \frac {\operatorname {h e i g h t} (\mathbf {b} _ {\mathrm {j}})}{\text {p a t c h e s . h . s i d e}}\right). \exp a n d (\mathbf {n} _ {\mathrm {j}}, 2) \\ \end{array} +$$ + +end + +3. Remove redundant patches + +P,S $\leftarrow$ Remove_overlapedpatches(Concat(Pj),Concat(Sj)) + +return P and S + +# 3.3. Subject-Oriented Data Curation + +To fine-tune our TEVA model for dynamically sampled patches, we introduce a data curation pipeline that creates subject-oriented training data, emphasizing specific regions of interest within images. While large datasets [3, 51] are commonly used to enhance the generalization of VLMs, they typically consist of image-caption pairs providing broad, whole-image descriptions. Our approach curates question-answer pairs that emphasize zoomed-in regions, enabling the model to better understand and prioritize detailed image areas. To address this, we developed an efficient method that curates large-scale, subject-oriented conversations (8 million pairs) based on existing image-caption + +![](images/42471df3e5794d0cd39d3db5b0ef00738f654aa5888678eb70205e736b071bda.jpg) +The image is only for illustration. The images for subject-oriented VCA generation. + +Image Caption: The image portrays a fairy with white wings and a purple dress, standing in a field of lavender flowers. She is holding a lavender plant in her hands. The background is a light blue color, and there are two butterflies visible in the scene. The fairy is positioned in the center of the image, with the lavender plants surrounding her. The butterflies are located in the top left and bottom right corners of the image. The fairy's wings are spread out, and her dress is flowing, giving the impression of movement. The lavender plants are tall and thin, with small purple flowers at the top. The butterflies are small and white, with black spots on their wings. The overall color scheme of the image is purple, white, and light blue. + +![](images/8fba0d29514340cd3abd3b56a4d7ffb1f2eac1acd306f445dfc941a3496d5d6e.jpg) +Figure 4. Example of constructing Subject-Oriented QA pairs. Using the image captions, we employ the Llama3.1 [59] to generate an average of three questions per image, resulting in 8 million QA pairs focused on different objects described in each caption. + + Generate three question-answer pairs based on this image description Each question should ask about one different object in the image without repeating objects. + +![](images/b2cafc2701ac137f928f5b6ccf1737aae6d577a336bc8dcbc4d8b04be533aa11.jpg) + +Llama + +1. Question: what is the fairy wearing? Answer: The fairy is wearing a purple dress. +2. Question: How many butterflies are there? Answer: There are two. +3. Question: What do the lavender plants look like? Answer: They are tall and thin, with small purple flowers at the top. + +datasets [3, 4, 11]. As illustrated in Figure 4, we leverage Llama 3.1 [59] for its strong contextual language reasoning ability, using only structured text prompts as input to generate subject-focused conversations. After generating these conversations, we apply an open-vocabulary detection method to validate the presence of specified objects in the images, filtering out any noisy data and ensuring data quality. Experiments in Table 4 demonstrate the significance of this pretraining dataset, especially for challenging benchmarks that require an understanding of zoomed-in patches or complex reasoning tasks. The details of data curation are available in the supplementary material. + +# 4. Experiments + +In this section, we provide the training settings and evaluation benchmarks employed to assess our model's performance on high-resolution and general Vision-Language Model (VLM) tasks. Subsequently, we present comparative results, benchmarking our method against state-of-the-art approaches. Further details are available in the supplementary material. + +# 4.1. Experimental Setup + +Training stages The goal of TEVA is to enhance the model's ability to process high-resolution images by focusing on informative patches. Two key requirements are essential for achieving this: (1) enabling the vision tower to handle dynamic patch sampling, and (2) ensuring that dynamic vision tokens can be effectively interpreted by the LLM. To meet these goals, we structured our training pipeline into four stages: + +- Alignment Stage: Following the process used in LLaVA [37], we begin by aligning visual features with the language space. In this stage, we freeze the large + +language model (LLM) and optimize only the vision-language (VL) adapter with LLaVA alignment dataset (595k), ensuring that the visual features are properly mapped to the language domain. + +- Vision Tower Adapting: Using 4M samples from our proposed Subject-Oriented Dataset, we fine-tune the vision tower to handle dynamic image patches. During this stage, both the vision tower and VL adapter are unfrozen for training. +- Integrated Interpretation Stage: To further enhance the LLM's ability to interpret dynamic vision tokens, we unfreeze the vision tower, VL adapter, and LLM, then fine-tune using 4M samples from the Subject-Oriented dataset and 1M samples from the ShareGPT4V [4] caption dataset. This step aligns all components for seamless interpretation of dynamic visual information. +- Instruction Fine-tuning: In the final stage, following the training strategy of VILA [37], we continue fine-tuning the vision tower, VL adapter, and LLM with a mix of Supervised Fine-Tuning (SFT) Dataset (4.17M) built on VILA [37]. This instruction fine-tuning stage enhances the model's instruction-following and dialogue capabilities on multiple tasks. Details of the training settings and datasets are provided in the supplementary materials. + +Evaluation benchmarks To assess our model's capability in handling complex high-resolution images, we report results on challenging high-resolution benchmarks, including $V^{*}$ [61] and MME-RealWorld [79]. Additionally, we evaluate our model on commonly used general VQA and MLLM benchmarks including MMBench [41], SQA (ScienceQA) [43], $\mathrm{VQAT}$ (TextVQA) [54], POPE [29], AI2D [22], LLaVAw (LLaVA-Bench In-the-Wild) [37], $\mathrm{MMMU}^{\mathrm{t}}$ (MMMU-test) [69], and $\mathrm{VQAV}^{\mathrm{v2}}$ (VQA-v2) [14]. Evaluation of more benchmarks is available in the supplementary. + +# 4.2. Main Results + +Comparison with region-aware VLMs. We evaluate Token-Efficient Vision LLanguage Model (TEVA) on the high-resolution visual detail benchmark $V^{*}$ [61], comparing its performance and efficiency against state-of-the-art region-aware VLMs. As shown in Table 1, TEVA consistently outperforms existing methods in attribute recognition (AR), spatial relationship reasoning (SRR), and scene OCR tasks. Our dynamic patch sampling strategy significantly reduces computational overhead compared to SpatialGPT (mask pooling) [6], Griffon-G [72] and $V^{*}$ [61], which require multiple vision tower passes to process individual ROIs. Specifically, TEVA achieves a $7.53 \times$ speedup over $V^{*}$ in ROI aggregation and improves overall inference efficiency by $12.31 \times$ compared to $V^{*}$ [61], benefiting from both token efficiency and the lightweight RAP module, which eliminates the need for iterative region searching. + +How many people are there on the street? +![](images/085235beebac11b4f33d69d6877492a23ce14f4570b380128d8a1f963079ef42.jpg) +VILA-1.5 +There is only one person on the street. + +![](images/f8b68f9c05c5bc06578ee1d47e0188bebaeb629b3f58091a4e5a4f62b091e58e.jpg) +Monkey +ChatGPT-40 + +![](images/1c03c33f34dd65596fba737f31f944c0d4911914a99ecc3cf3149604dbb26c5e.jpg) + +What is the breed of the dog? +![](images/3d14656949e1ee1f5aacdbb1ed0b61243d3bb2100ff5996ec3031812e72783e7.jpg) +VILA-1.5 +The dog in the image is a white dog. +Pug. + +![](images/60af72dfd219e8d855f0abfbec23ebe28574bb77f811b94b3f3e49117486cbd2.jpg) +There is one person visible on the street. +Monkey +difficult to identify the breed. + +![](images/43cb6b63581acaebe012f234d233b642e9323d55430d6ff4c705cabc2204a7c1.jpg) + +What color is the bottle cap? +![](images/91f5ccf34abeafccae6483240717f387b505bce7a61f9d6a27230b3755b76cf0.jpg) +VILA-1.5 +The bottle has a blue cap. +Yellow. + +![](images/0a0b956b9bb5161918f9cc06637ccc05d35f17fdcbe11e51da90eda1f5859fa5.jpg) + +![](images/553d44c330c57927bc9f3368901a6b50ef78cdd9ca48860753933d2268cb6515.jpg) +Monkey +ChatGPT-40 +The color of the cap appears to be red. + +![](images/6f24cc55cee359e0aba86ad7b6cb7d82683414772db15359e36f7def45fb098e.jpg) +What is the speed limit? + +![](images/c0ba9204fef177b283b3250feb35545e346d371c02659dbb4cbbd736ffe99c3d.jpg) +There are 2. +VILA-1.5 +It is $10\mathrm{km / h}$ +50. + +![](images/c43006d0fce0039a6d9d486f3787ef41bd340821e1b4bd761cfbf16c5ac1d097.jpg) +Monkey +ChatGPT-40 + +Figure 5. Evaluation of our model against leading open-source VLMs (VILA-1.5 [32], Monkey [31]) and the advanced commercial VLM, GPT-4o [50], in high-resolution, real-world scenarios. Each image displays the proposed subject in the bottom left corner. The first row presents responses to general questions, while the second row addresses more challenging OCR-based questions. The proposed subject is in the left-bottom corner of each image. Despite using a smaller language model (Sheared-LLaMA-2.7B [62]), our model demonstrates robust capability in observing precise details for high-resolution images and can adapt to various image resolutions, providing flexibility and enhanced performance in demanding high-resolution scenarios. Please note that all sample images were processed at their original resolution $(2246\times 1582$ on average) for TEVA-3B. +![](images/953c4b79f9ba2434e1c42fc26b0130d3ccff034123a9ebf8a280b3fb2718630c.jpg) +It is 40. + +What is the written on the large boat? +![](images/8a7c6982364fe91658d78a93f969354fc3d4cc145684e03181d5c77733af1a15.jpg) +Golden Retriever. +VILA-1.5 +"Oranje" is written on it. +Unanswerable. + +![](images/15b6f7534fd3582d3f29008687d15573d545b17fd122636612f695d70f1fed32.jpg) +The speed limit appears to be 30. +Monkey +The text on the boat reads "INTERCOMBI". +Nitrico. + +![](images/2b95613fedc10d8511381c4aebc4f39423c81bf8867134618bcacc85adec5511.jpg) + +Tell me the yellow text on the airship. +![](images/ec8214adc5331ac7380da365b2f65983cd5dfdda8eeb9d1cc8aab824a5de3a1d.jpg) +It is red. +VILA-1.5 +"Airship". +Unanswerable. + +![](images/feb679bfac17eff5710fe9a143d10bc73df4d2b69d6835edc5a07f2ee83979b1.jpg) +Monkey +ChatGPT-40 +The text appears to say "GOODYEAR". + +![](images/895625fa76c67137b73f26c3e736eeba01853e859ab89f60e1ef97bcd55a0e3f.jpg) +The yellow text reads "GoodYear". + +Unlike SpatialGPT [6] and Griffon-G [72], which rely on manual region selection, TEVA automatically detects and proposes ROIs. For a fair comparison, we use TEVA's RAP module to predict ROIs for SpatialGPT and Griffon-G. + +High resolution benchmark results. To demonstrate TEVA's capability in handling high-resolution images, we evaluated our method on high-resolution VLM benchmarks: $V^{*}$ [61], with an average image resolution of $2246 \times 1582$ pixels and target subjects, and MME-RealWorld [79], with an average resolution of $2000 \times 1500$ pixels. + +As shown in Table 2, TEVA excels in tasks that require precise detail recognition, such as attribute recognition (AR) and spatial relationship reasoning (SRR) in $V^{*}$ benchmark, outperforming both the commercial model GPT-4V [50] by 21.6 points and all the other open-source high resolution VLMs. In the complex MME-RealWorld benchmark, TEVA also surpasses open-source methods in both perception and reasoning ability. In applications that demand high-resolution fine visual detail perception, such as remote sensing, video monitoring, and autonomous driv + +
MethodV* Bench
ARSRROCR
SpatialGPT [6]36.556.540.0
Griffon-G [72]44.771.056.6
V* [61]74.776.380.0
TEVA-7B76.577.683.3
MethodInference Time
AIPFPI
SpatialGPT [6]9.57×1.05×
Griffon-G [72]10.24×4.83×
V* [61]7.53×12.31×
TEVA-7B1.00×1.00×
+ +Table 1. Comparison with state-of-the-art region-aware 7B VLMs on the high-resolution $V^{*}$ benchmark, evaluating attribute recognition (AR), spatial relationship reasoning (SRR), and scene OCR tasks. Inference times are reported for Additional Information Processing (AIP) and Full Pipeline Inference (FPI), demonstrating the efficiency and performance gains of dynamic patch sampling over SpatialGPT(pooling), Griffon-G and $V^{*}$ (multiple vision tower passes). + +
MethodV* AverageMME-RWPRCMME-RWRSN
OCRRSMOADOCRMOAD
GPT-4V [50]55.9-------
GPT-4o [50]-77.728.933.922.461.436.526.4
Gemini Pro [56]50.5-------
Gemini-1.5-pro [56]-67.614.031.126.652.728.319.2
LLaVA-1.5 [35]50.538.722.119.126.026.025.924.2
LLaVA-1.5-S2[52]56.5-------
Text-Monkey [40]-37.311.716.114.330.44.420.0
Monkey [31]-54.625.028.029.727.227.333.0
LLaVA-Next [36]61.340.631.635.427.214.929.528.2
Cambrian-1 [57]60.832.735.827.635.527.531.231.2
MLLM-K [75]62.3-------
TEVA-3B61.643.132.332.633.041.833.529.9
TEVA-7B77.160.738.236.935.249.434.735.5
+ +Table 2. Comparison with state-of-the-art open-source and commercial methods on the high-resolution benchmark $V^{*}$ [61] (averaging attribute recognition(AR) and spatial relationship reasoning(SRR)), as well as on MME-RealWorld [79], which includes perception benchmark MME-RW $_{PRC}$ and reasoning benchmark MME-RW $_{RSN}$ . Task categories are abbreviated due to space limits: OCR (Optical Character Recognition), RS (Remote Sensing), MO (Video Monitoring), AD (Autonomous Driving). Our model consistently outperforms other open-source methods on Reasoning tasks such as OCR, RS, MO, AR and SRR. It even surpasses leading commercial VLMs, GPT-4o and Gemini-1.5-pro, on RS, MO, and AD perception tasks that require precise visual grounding. The best results are bold. Notably, all samples are evaluated at their original image resolution for TEVA and at the largest default resolution for other methods. + +
MethodLLM#Visual TokenMMBSQAVQATPOPEAI2DLLaVAwMMMUVVQAV2
InstructBLIP [7]Vicuna-7B-36.063.150.7--60.9--
IDEFICS-9B [18]LLAMA-7B-48.2------50.9
LLaVA-1.5 [35]Vicuna-7B57664.366.866.885.9-63.435.778.5
MLLM-K [75]Vicuna-7B---56.187.3---76.5
LLaVA-1.5-S2[52]Vicuna-7B57666.2-----37.780.0
VILA-1.5 [32]Llama-2-7B72968.968.268.285.5-69.738.679.9
Mini-Gemini-HD [30]Vicuna-7B288072.775.168.4-73.5-37.3-
Monkey [31]Qwen-7B5120-69.467.6-62.6--80.3
LLaVA-Next [36]LLama3-8B288072.172.864.686.771.683.241.781.8
Cambrian-1 [57]Llama3-8B-Instruct230475.980.471.7-73.0-42.781.2
TEVA-3BSheared-llama-2.7B72964.275.766.887.064.985.032.480.3
TEVA-7BQwen2-7B72980.392.572.587.976.394.348.482.8
+ +Table 3. Comparison with leading methods on 8 commonly used VLM benchmarks: MMBench [41], SQA: ScienceQA [43], VQATextVQA [54], POPE [29], AI2D [22], LLaVAw: LLaVA-Bench (In-the-Wild) [37], MMMUv: MMMU-validation [69], VQAV2: VQAV2 [14]. Our model consistently outperforms the other methods without image tiling and increasing visual tokens. The best results are bold. Please note that all the samples are evaluated on original image resolution for our model. + +ing (e.g., identifying traffic signs), TEVA demonstrates robust performance and even surpassing the most powerful commercial models, GPT-4o [50] and Gemini-1.5-pro [56] + +General resolution benchmark results. We also evaluate our model on eight commonly used VQA and MLLM benchmarks, comparing it with state-of-the-art VLMs, as shown in Table 3. The results demonstrate that TEVA consistently outperforms other leading methods across all the benchmarks and achieves this without requiring image tiling or visual token extending. + +Notably, on general resolution subject-oriented QA tasks + +like MMBench [41], AI2D [22] and $\mathrm{MMMU}^{\mathrm{V}}$ [69], TEVA achieves superior results compared to all other methods. This highlights the generalizability of our approach in focusing on contextually important areas within an image. + +Qualitative Results. In Figure 5, we present qualitative results that demonstrate the effectiveness of TEVA-3B in challenging high-resolution scenarios requiring precise detail recognition. By leveraging our informative patch sampling strategy, TEVA dynamically focuses on relevant, information-rich regions based on the user query, enabling inference on images at arbitrary resolutions without increas + +Results with Different Dynamic and Static Token Numbers + +
DynamicStaticV*MRWPOPEVQAT
025645.030.083.044.5
808150.431.585.654.2
47325661.634.087.066.8
+ +Comparison of w/o Subject-Orient (SO) Data + +
SO DataMRWPRCMRWRSNPOPEVQAT
X33.128.586.558.7
✓33.834.387.066.8
+ +Comparison of w/o Matrix Embedding (ME) + +
MEV*ARV*SRRMRWPOPE
X48.350.031.884.8
✓60.063.234.087.0
+ +Table 4. Ablation Studies on Token Numbers, w/o Subject-Orient Pretraining and w/o Affine Transformation Matrix Embedding. Benchmark abbreviation MRW: MME-RealWorld [79]. + +ing computational cost. For example, processing a 5760 $\times$ 2880 (5K-resolution) image takes only 2.5 seconds on a single NVIDIA A100 GPU with TEVA-7B. This efficiency enables TEVA to handle high-resolution images seamlessly, outperforming leading open-source VLMs, such as VILA1.5-7B [32], which uses the same number of visual tokens (729), and Monkey-7B [31], which relies on image tiling for high-resolution support. Furthermore, our approach surpasses the advanced commercial VLM, ChatGPT-4o [50], in precise detail observation for high-resolution images, despite the latter's use of a more powerful language model. The evaluation covers real-world, high-difficulty OCR tasks along with diverse general queries, such as counting, identification, and attribute recognition, providing a comprehensive assessment of TEVA's capabilities. + +# 4.3. Ablation Study + +Dynamic vs. Static Token Numbers. We assess the impact of using dynamic, informative image patches compared to static, uniformly sampled patches. As shown in Table 4, adding 473 dynamic patches to the baseline of 256 static patches (729 tokens in total) leads to significant improvements in tasks requiring detailed structural understanding, such as scene OCR tasks in $\mathrm{VQA}^{\mathrm{T}}$ (TextVQA) [54], as well as in handling high-resolution images, as demonstrated on the $V^{*}$ [61] benchmark. Additionally, we conducted experiments using a reduced set of 144 patches (80 dynamic and 81 static). Despite the lower patch count, incorporating dynamic tokens consistently enhanced performance on challenging datasets that demand high-resolution or precise observation. + +Subject-Oriented Pretraining. To evaluate the benefits of pretraining on our Subject-Oriented (SO) Dataset, we compare results with and without the Subject-Oriented Pretraining phase. As illustrated in Table 4, pretrain + +ing on a large-scale Subject-Oriented Dataset notably improves the model's complex reasoning abilities (MMERealWorld $_{RSN}$ : MME-RealWorld-Reasoning [79]) and scene understanding, especially in tasks requiring OCR, such as $\mathrm{VQAT}$ (TextVQA) [54]. However, for tasks requiring more straightforward scene comprehension, such as basic perception (MME-RealWorld $_{PRC}$ : MME-RealWorld-Perception [79]) or hallucination assessments (POPE [29]), the improvements from Subject-Oriented Pretraining are less pronounced. These findings demonstrate the effectiveness of Subject-Oriented Pretraining in enhancing the model's capacity for complex visual understanding. + +Affine Transformation Matrix Embeddings. We examine the effectiveness of embedding the affine transformation matrix to provide positional information for each patch. Results in Table 4 demonstrate that the Affine Transformation Matrix Embeddings (ME) consistently improve performance across various benchmarks, particularly enhancing the model's ability to comprehend the relative positions of objects in the $V_{SRR}^{*}$ ( $V^{*}$ Spatial Relationship reasoning) benchmark. This aligns well with our intuition, as the additional spatial context helps the model maintain awareness of object locations and relationships. + +# 4.4. Limitation + +While TEVA can process images at arbitrary resolutions without increasing the token count, its effectiveness may be limited in scenarios involving densely packed high-resolution images, such as text-heavy documents. In these cases, where information is uniformly distributed across the image, there are no distinct regions to prioritize. Consequently, dynamic sampling must cover the entire image at a lower resolution, limiting the advantages of high-resolution focus on key areas. Additional experiments and discussion are provided in the supplementary material. + +# 5. Conclusion + +In this work, we introduced Token-Efficient Vision LLanguage Model (TEVA), a straightforward and token-efficient framework that enhances Vision-Language Models (VLMs) for high-resolution image processing. TEVA employs a subject-oriented strategy and Dynamic Patch Sampling to balance fine details with global context, supported by a specialized Subject-Oriented Dataset for improved learning. Comprehensive evaluations against state-of-the-art VLMs, including GPT-4o, show that TEVA excels in challenging benchmarks and holds strong potential for applications requiring high-resolution precision, such as remote sensing, autonomous driving, and video monitoring. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. 2022. 1 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2 +[3] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.4.5 +[4] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 5, 7 +[5] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 1 +[6] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062-135093, 2025. 2, 5, 6, 7 +[7] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 7 +[8] Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuoling Yang, Zihan Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nvlm: Open frontier-class multimodal llms. arXiv preprint arXiv:2409.11402, 2024. 2 +[9] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 1 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[11] Yunhao Fang, Ligeng Zhu, Yao Lu, Yan Wang, Pavlo Molchanov, Jang Hyun Cho, Marco Pavone, Song Han, and Hongxu Yin. vila2: Vila augmented vila. arXiv preprint arXiv:2407.17453, 2024. 5 +[12] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A + +comprehensive evaluation benchmark for multimodal large language models. ArXiv, abs/2306.13394, 2023. 1, 3 +[13] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023. 7 +[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 5, 7, 1, 3 +[15] Qiushan Guo, Shalini De Mello, Hongxu Yin, Wonmin Byeon, Ka Chun Cheung, Yizhou Yu, Ping Luo, and Sifei Liu. Regionpt: Towards region understanding vision language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13796-13806, 2024. 2, 4 +[16] Jack Hessel, Jena D Hwang, Jae Sung Park, Rowan Zellers, Chandra Bhagavatula, Anna Rohrbach, Kate Saenko, and Yejin Choi. The abduction of sherlock holmes: A dataset for visual abductive reasoning. In European Conference on Computer Vision, pages 558-575. Springer, 2022. 7 +[17] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 3, 7 +[18] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. 2023. 7 +[19] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. 4 +[20] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901-2910, 2017. 7 +[21] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 7 +[22] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 5, 7, 1, 3 +[23] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeong Yeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 7 + +[24] Jonathan Krause, Justin Johnson, Ranjay Krishna, and Li Fei-Fei. A hierarchical approach for generating descriptive image paragraphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 317–325, 2017. 7 +[25] Paul Lerner, Olivier Ferret, Camille Guinaudeau, Hervé Le Borgne, Romaric Besançon, José G Moreno, and Jesús Lovón Melgarejo. Viquae, a dataset for knowledge-based visual question answering about named entities. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 3108-3120, 2022. 7 +[26] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 1, 3 +[27] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 7 +[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1 +[29] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 5, 7, 8, 1, 3 +[30] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 2, 7 +[31] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26763-26773, 2024. 2, 6, 7, 8 +[32] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 6, 7, 8, 1, 2 +[33] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1, 2 +[34] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. Mmc: Advancing multimodal chart understanding with large-scale instruction tuning. arXiv preprint arXiv:2311.10774, 2023. 7 +[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 7 + +[36] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.1,7 +[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NIPS, 2024. 1, 5, 7, 3 +[38] Haogeng Liu, Quanzeng You, Xiaotian Han, Yiqi Wang, Bohan Zhai, Yongfei Liu, Yunzhe Tao, Huaibo Huang, Ran He, and Hongxia Yang. Infimm-hd: A leap forward in high-resolution multimodal understanding. arXiv preprint arXiv:2403.01487, 2024. 1, 2 +[39] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 2, 5 +[40] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 7 +[41] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 5, 7, 1, 3 +[42] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, Xiuyu Li, Yunhao Fang, Yukang Chen, Cheng-Yu Hsieh, De-An Huang, An-Chieh Cheng, Vishwesh Nath, Jinyi Hu, Sifei Liu, Ranjay Krishna, Daguang Xu, Xiaolong Wang, Pavlo Molchanov, Jan Kautz, Hongxu Yin, Song Han, and Yao Lu. Nvila: Efficient frontier visual language models, 2024. 1 +[43] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5, 7, 1, 3 +[44] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 7 +[45] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2, 7 +[46] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2 +[47] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 2 +[48] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering + +by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 7 +[49] Ravina Mithe, Supriya Indalkar, and Nilam Divekar. Optical character recognition. International journal of recent technology and engineering (IJRTE), 2(1):72-75, 2013. 1 +[50] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2(5), 2023. 1, 6, 7, 8, 2 +[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 4 +[52] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? In European Conference on Computer Vision, pages 444-462. Springer, 2025. 1, 2, 7 +[53] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 742–758, 2020. 7 +[54] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 7, 8, 1, 3 +[55] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. Wit: Wikipedia-based image text dataset for multimodal multilingual machine learning. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, pages 2443-2449, 2021. 7 +[56] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 7 +[57] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2024. 7 +[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 5, 2 +[60] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and + +Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 7 +[61] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084-13094, 2024. 2, 4, 5, 6, 7, 8, 1 +[62] Mengzhou Xia, Tianyu Gao, Zhiyuan Zeng, and Danqi Chen. Sheared llama: Accelerating language model pre-training via structured pruning. arXiv preprint arXiv:2310.06694, 2023. 6 +[63] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794-4803, 2022. 2 +[64] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 2 +[65] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1 +[66] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 4 +[67] Xiaoyu Yue, Shuyang Sun, Zhanghui Kuang, Meng Wei, Philip HS Torr, Wayne Zhang, and Dahua Lin. Vision transformer with progressive sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 387-396, 2021. 2 +[68] Xiang Yue, Xingwei Qu, Ge Zhang, Yao Fu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mammoth: Building math generalist models through hybrid instruction tuning. arXiv preprint arXiv:2309.05653, 2023. 7 +[69] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567, 2024. 5, 7, 1, 3 +[70] Bohan Zhai, Shijia Yang, Chenfeng Xu, Sheng Shen, Kurt Keutzer, and Manling Li. Halle-switch: Controlling object hallucination in large vision language models. arXiv e-prints, 2023. 2 +[71] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 1 +[72] Yufei Zhan, Hongyin Zhao, Yousong Zhu, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon-g: Bridging vision-language and vision-centric tasks via large multimodal models. arXiv preprint arXiv:2410.16163, 2024. 2, 4, 5, 6, 7 +[73] Yufei Zhan, Yousong Zhu, Hongyin Zhao, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon v2: Advancing multimodal + +perception with high-resolution scaling and visual-language co-referring. arXiv preprint arXiv:2403.09333, 2024. 1, 2, 4 +[74] Haotian Zhang, Haoxuan You, Philipp Duffer, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 2 +[75] Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara, and Filip Ilievski. Mllms know where to look: Training-free perception of small visual details with multimodal llms. Proceedings of International Conference on Learning Representations (ICLR), 2025. 7 +[76] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Haodong Duan, Songyang Zhang, Shuangrui Ding, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 2 +[77] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1 +[78] Yanzhe Zhang, Ruiyi Zhang, Jiuming Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 2, 7 +[79] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, et al. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans? arXiv preprint arXiv:2408.13257, 2024. 5, 6, 7, 8, 1 +[80] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1 \ No newline at end of file diff --git a/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/images.zip b/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fa781a9f96e0a4c1f9666cbcd8fd3615adde3f26 --- /dev/null +++ b/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:87b22f7ba8a27855369514d3fce3c0a11a2ed4402d03d19afe29b54af244d843 +size 603570 diff --git a/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/layout.json b/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..264aa37b9e4d2ffa88c887e9c78ee8c3d8311743 --- /dev/null +++ b/2025/Token-Efficient VLM_ High-Resolution Image Understanding via Dynamic Region Proposal/layout.json @@ -0,0 +1,11448 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "type": "text", + "content": "Token-Efficient VLM: High-Resolution Image Understanding via Dynamic Region Proposal" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "spans": [ + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "text", + "content": "Yitong Jiang" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "text", + "content": " Jinwei Gu" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "text", + "content": " Tianfan Xue" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "text", + "content": " Ka Chun Cheung" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "text", + "content": " Pavlo Molchanov" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "text", + "content": " Hongxu Yin" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "text", + "content": " Sifei Liu" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "text", + "content": " NVIDIA " + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 91, + 159, + 520, + 205 + ], + "type": "text", + "content": " The Chinese University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 221, + 228, + 333 + ], + "blocks": [ + { + "bbox": [ + 56, + 221, + 228, + 333 + ], + "lines": [ + { + "bbox": [ + 56, + 221, + 228, + 333 + ], + "spans": [ + { + "bbox": [ + 56, + 221, + 228, + 333 + ], + "type": "image", + "image_path": "b9f8094134eef96cdae7f92e7dc7b7351220f5e5b3d420f27fb60a273388fe58.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 341, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 54, + 341, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 341, + 555, + 399 + ], + "type": "text", + "content": "Figure 1. Illustration of the Token-Efficient Vision LLanguage Model (TEVA) for high-resolution scenarios. Given a user query about a specific detail " + }, + { + "bbox": [ + 54, + 341, + 555, + 399 + ], + "type": "inline_equation", + "content": "(20\\times 50\\mathrm{px}" + }, + { + "bbox": [ + 54, + 341, + 555, + 399 + ], + "type": "text", + "content": " region) in a high-resolution image " + }, + { + "bbox": [ + 54, + 341, + 555, + 399 + ], + "type": "inline_equation", + "content": "(5760\\times 2880\\mathrm{px})" + }, + { + "bbox": [ + 54, + 341, + 555, + 399 + ], + "type": "text", + "content": " , traditional VLMs (Left) often miss critical details due to uniform patch processing and computational limits. TEVA (Right) addresses this by using a Relevant Area Proposal module to identify key subjects (e.g., people) and Dynamic Patch Sampling to focus on high-resolution patches of interest while retaining global context patches. This approach enables accurate interpretation of details at arbitrary resolutions without increasing the token count." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 239, + 228, + 345, + 330 + ], + "blocks": [ + { + "bbox": [ + 239, + 228, + 345, + 330 + ], + "lines": [ + { + "bbox": [ + 239, + 228, + 345, + 330 + ], + "spans": [ + { + "bbox": [ + 239, + 228, + 345, + 330 + ], + "type": "image", + "image_path": "76caf698d7b59308f0c8fd33c6975bf48e88287e18735ba1e06e37cd2681f745.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 356, + 222, + 472, + 335 + ], + "blocks": [ + { + "bbox": [ + 356, + 222, + 472, + 335 + ], + "lines": [ + { + "bbox": [ + 356, + 222, + 472, + 335 + ], + "spans": [ + { + "bbox": [ + 356, + 222, + 472, + 335 + ], + "type": "image", + "image_path": "8c451e37662a929f2eb525e829632f8fbaea2e883ee094318beda32abcd874f6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 475, + 222, + 550, + 334 + ], + "blocks": [ + { + "bbox": [ + 475, + 222, + 550, + 334 + ], + "lines": [ + { + "bbox": [ + 475, + 222, + 550, + 334 + ], + "spans": [ + { + "bbox": [ + 475, + 222, + 550, + 334 + ], + "type": "image", + "image_path": "9ca2907859b78258310a8467c1209cce64ff9587719301615d11edaa65b2782c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 407, + 200, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 407, + 200, + 419 + ], + "spans": [ + { + "bbox": [ + 152, + 407, + 200, + 419 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 434, + 297, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 297, + 662 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 297, + 662 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) excel at visual understanding by leveraging pretrained image encoders to generate visual tokens. However, they struggle with high-resolution images and zoomed-in regions due to the computational burden and token redundancy of uniform patch-based processing, often leading to the loss of critical details. To address these challenges, we propose Token-Efficient Vision LLanguage Model (TEVA), a novel framework that detects key regions and applies dynamic patch sampling to efficiently capture fine-grained details while preserving global context. Our approach first identifies subject-oriented regions using an adaptive detection strategy. Then, a dynamic patch sampling mechanism selects and arranges patches at varying scales, ensuring efficient processing without increasing token count. Extensive experiments demonstrate that Token-Efficient Vision LLanguage Model (TEVA) significantly enhances VLM performance in handling visual details, seamlessly integrating with various decoders and LLMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 407, + 394, + 419 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 407, + 394, + 419 + ], + "spans": [ + { + "bbox": [ + 315, + 407, + 394, + 419 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 432, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 601 + ], + "type": "text", + "content": "Vision-Language Models (VLMs) [1, 36, 37, 80] have recently received significant attention due to their ability to integrate visual understanding into Large Language Models (LLMs) [50, 59, 77], enhancing performance in tasks like image captioning and visual question answering. However, supporting applications that require precise attention to specific details—such as video surveillance analysis, satellite imagery interpretation, and scene OCR [49]—remains challenging. These tasks often necessitate zooming into fine-grained regions of high-resolution images to capture essential information, but many existing VLMs [1, 28, 37, 80] rely on low-resolution pretrained image encoders (e.g., " + }, + { + "bbox": [ + 313, + 432, + 555, + 601 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 432, + 555, + 601 + ], + "type": "text", + "content": " pixels), leading to the loss of important details [38, 52]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "To address this issue, previous methods have focused on either increasing global resolution or applying region-aware techniques. Enhancing resolution in pretrained image encoders typically involves two strategies: scaling resolution via curriculum learning [2, 5, 73] and partitioning images into fixed-size sub-images for independent processing [9, 33, 35, 36, 38, 42, 52]. These approaches have notable limitations. Scaling resolution increases computational complexity quadratically, raising self-attention" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 692, + 256, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 692, + 256, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 692, + 256, + 713 + ], + "type": "text", + "content": "†Yitong Jiang was an intern at NVIDIA during the project. \n†Equal Advising." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24147" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "costs. Partitioning images boosts the number of visual tokens (e.g., over 5000 tokens for an " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "896 \\times 896" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " image [31]), leading to redundancy and degraded performance [35, 38]. Moreover, the lack of interaction among image tokens within the vision encoder limits the model's ability to capture contextual relationships, requiring specialized design interventions [31, 38]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 294, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 294, + 253 + ], + "type": "text", + "content": "Alternatively, region-aware methods [6, 15, 66, 72, 73] focus on fine-grained details by leveraging manually defined Regions of Interest (ROIs). However, these methods either compress visual tokens within target regions, potentially losing critical information [6, 15, 66], or require multiple independent passes through the vision encoder, significantly increasing computational cost and token usage [72, 73]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 255, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 295, + 495 + ], + "type": "text", + "content": "To address these limitations, we propose Token-Efficient Vision LLanguage Model (TEVA), which leverages a dynamic patch sampling framework for the vision encoder that balances visual token count and image resolution. Our primary concept involves informative patch sampling within the vision encoder, which selectively focuses on key parts of the image while preserving the understanding of the global context, represented in low resolution. As shown in Figure 1, unlike conventional vision encoders that uniformly partition the entire image into patches, our method makes an early determination of which image regions are essential for answering a given question and applies selective patch sampling before inputting patches into the vision encoder. This approach not only prevents an increase in the number of patches when processing high-resolution images of any size, but also reduces self-attention computation and minimizes tokens the LLM needs to process. Additionally, this approach maintains coherence among sampled image regions, enhancing the model's ability to understand the global context and structure of the input image [63, 67]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 663 + ], + "type": "text", + "content": "Our TEVA consists of three main steps. First, the Relevant Area Proposal (RAP) module identifies key locations and scales of interest using a subject proposal agent powered by an LLM, and a subject detection agent via a vision foundation model. This enables efficient sampling of informative patches as vision tokens. Next, patches from localized regions are arranged with low-resolution global patches, integrating positional information through an affine transformation strategy. This allows for sampling patches of varying sizes and locations without additional tokens or complex architectures [8, 31, 38, 64]. Finally, the mixture of local and global patch embeddings is processed in the vision encoder (SigLIP [71] in our work) to extract visual tokens for the Large Language Models (LLMs)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "Furthermore, to enhance dynamic patch sampling for VLMs, we introduce a data curation pipeline and curate a Subject-Oriented (SO) dataset containing 8 million subject-focused conversations. This is the first large-scale Subject" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "Oriented VQA dataset for VLM pretraining, enabling a framework that focuses on specific image details. By integrating dynamic image tokens with early interactions between local and global visual elements, Token-Efficient Vision LLanguage Model (TEVA) achieves superior VLM performance across various benchmarks, particularly excelling in flexible-resolution scenarios, as shown in our comparative analysis in Section 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 178, + 400, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 400, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 400, + 190 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 198, + 553, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 294 + ], + "type": "text", + "content": "Recent studies [58, 70, 78] highlight a bottleneck in Vision-Language Models (VLMs), where image resolution constraints limit the model's ability to capture fine-grained visual details. Previous approaches of handling high-resolution images in Vision-Language Models (VLMs) can be grouped into two main strategies: (1) increasing the global image resolution, (2) applying region-aware techniques, as discussed in the following." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 295, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 553, + 545 + ], + "type": "text", + "content": "In the first strategy, methods like Qwen-VL [2] and Griffon v2 [73] employ curriculum learning and bilinear interpolation to gradually increase the resolution in the pretrained image encoder. However, a significant challenge with transformer-based image encoders [10] is their reliance on uniform patch sampling, which leads to a quadratic increase in computational complexity as image resolution grows. Other methods [30, 31, 33, 38, 64, 76] partition high-resolution images into smaller sub-images, treating each as an independent input to the vision encoder. While this approach enables the processing of higher-resolution images, it imposes a constraint on the input sequence length that the Large Language Model (LLM) can handle, thereby limiting its overall capacity. To mitigate the computational burden, learnable queries within the Perceiver architecture [31, 38] and pooling strategies [52, 73] have been introduced to reduce the number of visual tokens generated by the vision encoder from these sub-images. However, such compression techniques often lack selectivity, leading to the loss of fine-grained visual details that are crucial for optimal Multimodal Large Language Model (MLLM) performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 676 + ], + "type": "text", + "content": "The second strategy involves explicitly identifying and processing Regions-of-Interest (ROIs). Some methods [6, 15, 66, 74] require manual specification of ROIs, followed by compression of low-resolution visual tokens. Others [72, 73] necessitate multiple passes through the vision encoder, increasing computational overhead. " + }, + { + "bbox": [ + 313, + 545, + 553, + 676 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 545, + 553, + 676 + ], + "type": "text", + "content": " [61] further extends this idea by iteratively searching for ROIs through multiple passes through an additional Vision-Language Model (VLM). However, this iterative search is computationally expensive and also requires multiple forward passes through the vision encoder for each ROI candidate." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "In contrast to these methods, our approach automatically identifies and focuses on informative regions within high-resolution images. By applying dynamic patch sam" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24148" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 553, + 201 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 201 + ], + "type": "image", + "image_path": "631ab1c885928cca65fbfb51c07bd31086577eeca79d6879ff5df3ffe108b651.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 200, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 55, + 200, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 555, + 234 + ], + "type": "text", + "content": "Figure 2. The TEVA architecture. Given a user query, the Subject Proposal Agent identifies relevant subjects (e.g., \"signs\"). The Subject Detection Agent then locates all candidate signs in the image. Informative patch sampling extracts high-resolution local patches at these locations, which are concatenated with low-resolution global patches and fed into the VLM for final interpretation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 326 + ], + "type": "text", + "content": "pling before the vision encoder, we preserve essential details without resorting to compression, while also minimizing computational costs. This is achieved by skipping less informative patches and avoiding multiple passes through the vision encoder, resulting in a more efficient and detailed visual representation for VLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 138, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 138, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 138, + 349 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 355, + 295, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 355, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 54, + 355, + 295, + 535 + ], + "type": "text", + "content": "We show the overall flowchart of our proposed TokenEfficient Vision LLanguage Model (TEVA) in Figure 2. TEVA is a VLM framework designed to handle arbitrary high-resolution images by detecting and focusing on informative visual details within images. It incorporates a Relevant Area Proposal (RAP) module to identify and extract key regions from the input image. A targeted patch sampling strategy then partitions these selected areas into informative high-resolution local patches, while retaining low-resolution global patches for context. The concatenation of local and global patches is processed by the image encoder to generate visual tokens. Additionally, we introduce a Subject-Oriented (SO) dataset to enhance the VLM's capability to manage dynamic patch sampling and improve overall performance. Details are described below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 228, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 228, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 228, + 553 + ], + "type": "text", + "content": "3.1. Relevant Area Proposal Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 653 + ], + "type": "text", + "content": "Building on our insight that image tokens are essential for Vision-Language Models (VLMs) to generate accurate content descriptions and answer questions, we address a key challenge: as shown in Figure 3 (left), when VLMs process uniformly partitioned image patches, they often focus on prominent elements, such as the path, trees, and signs on the trees, while overlooking finer details, like the presence of two individuals in the forest." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "To overcome this, we propose an automatic Relevant Area Proposal (RAP) module to capture key local details that help the VLM focus on the most informative and relevant areas in response to user queries. Current open-vocabulary detection methods, such as Grounding" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 266, + 432, + 347 + ], + "blocks": [ + { + "bbox": [ + 315, + 253, + 440, + 264 + ], + "lines": [ + { + "bbox": [ + 315, + 253, + 440, + 264 + ], + "spans": [ + { + "bbox": [ + 315, + 253, + 440, + 264 + ], + "type": "text", + "content": "What can you see in this image?" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 266, + 432, + 347 + ], + "lines": [ + { + "bbox": [ + 316, + 266, + 432, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 432, + 347 + ], + "type": "image", + "image_path": "a9073fbb70f73f59809c1fdb59add3299aebaba28c6e8ff54eab95d2e6e84d60.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 349, + 426, + 369 + ], + "lines": [ + { + "bbox": [ + 318, + 349, + 426, + 369 + ], + "spans": [ + { + "bbox": [ + 318, + 349, + 426, + 369 + ], + "type": "text", + "content": "Assistant: A forest with tall trees and a trail in the middle." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 379, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 456 + ], + "type": "text", + "content": "Figure 3. Example of visual focus in Vision-Language Models (VLMs) with uniform image patches (Left) versus high-resolution, relevant patches (Right). In the uniform approach, the model emphasizes broad features like trees, trails, and visible signs, missing subtler details, such as the presence of people in the scene. By providing high-resolution relevant patches, the model can identify and interpret finer details." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 438, + 266, + 553, + 347 + ], + "blocks": [ + { + "bbox": [ + 438, + 266, + 553, + 347 + ], + "lines": [ + { + "bbox": [ + 438, + 266, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 438, + 266, + 553, + 347 + ], + "type": "image", + "image_path": "8111a0443b3cb485999879f5b82792f2b6ddf61921e925c2c9a09aa417fd0f46.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 438, + 349, + 541, + 369 + ], + "lines": [ + { + "bbox": [ + 438, + 349, + 541, + 369 + ], + "spans": [ + { + "bbox": [ + 438, + 349, + 541, + 369 + ], + "type": "text", + "content": "Assistant: A group of people are walking through a forest." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 473, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 533 + ], + "type": "text", + "content": "Dino [39], have shown strong performance in detecting regions of interest based on straightforward text cues. However, user queries often involve complex questions or instructions without an explicit subject, which limits the applicability of open-vocabulary detection." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 641 + ], + "type": "text", + "content": "To address this, we use two specialized agents: a Subject-Promise Agent using LLM, e.g., Llama3.1 [59], and a Subject Detection Agent via Grounding Dino. Given a user query, we employ a structured prompt (e.g., \"You have limited vision of the image; tell me the subjects you need to observe to answer the question + User query\") to guide the LLM in proposing candidate subjects for detection. Once identified, the Subject Detection Agent then locates the regions of interest (ROI) based on these suggested subjects." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": "While this process effectively isolates the ROIs, integrating these fine-grained local details into the VLM remains challenging. To this end, we introduce the Informative Patch Sampling module in the following, which efficiently incorporates high-resolution information without increasing the token count or computational burden." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24149" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 211, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 211, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 211, + 85 + ], + "type": "text", + "content": "3.2. Informative Patch Sampling" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 293 + ], + "type": "text", + "content": "Previous region-level VLMs [15, 66] handle regions of interest by relying on low-resolution features and apply pooling strategies to compress information within each region. Although effective for reducing computational load, these pooling methods often lead to a loss of critical details, especially for high-resolution images. Alternative approaches [61, 72, 73] treat each target region as a separate image that passes through a vision tower. However, this approach is computationally expensive and significantly increases the number of image tokens and inference time. To better leverage high-resolution image information without incurring excessive computational costs, we propose a novel approach that utilizes informative patch sampling within each high-resolution target region. This technique preserves fine-grained details while maintaining computational efficiency and the same token count comparable to a low-resolution image encoder, as demonstrated in Table 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 307, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 296, + 392 + ], + "type": "text", + "content": "Patch allocation. As shown in Algorithm 1, our method begins by allocating informative patch numbers to regions of interest based on their relative sizes. Next, we align the assigned patches within each bounding box, ensuring even coverage across each target area. Finally, we remove overlapping bounding boxes to maintain a total of " + }, + { + "bbox": [ + 55, + 307, + 296, + 392 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 307, + 296, + 392 + ], + "type": "text", + "content": ", yielding the positions " + }, + { + "bbox": [ + 55, + 307, + 296, + 392 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{n \\times 2}" + }, + { + "bbox": [ + 55, + 307, + 296, + 392 + ], + "type": "text", + "content": " and scaling factors " + }, + { + "bbox": [ + 55, + 307, + 296, + 392 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{n \\times 2}" + }, + { + "bbox": [ + 55, + 307, + 296, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 404, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 296, + 464 + ], + "type": "text", + "content": "Patch sampling. Given the positions and the scaling factors, to link these positional clues to exact image patches, we embed " + }, + { + "bbox": [ + 55, + 404, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathsf{P} = \\{(p_i^x,p_i^y)|i\\in (0,n - 1)\\}" + }, + { + "bbox": [ + 55, + 404, + 296, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 404, + 296, + 464 + ], + "type": "inline_equation", + "content": "S = \\{(s_i^x,s_i^y)|i\\in (0,n - 1)\\}" + }, + { + "bbox": [ + 55, + 404, + 296, + 464 + ], + "type": "text", + "content": " into an affine transformation matrix to project image pixels into patches:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 470, + 295, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 470, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 131, + 470, + 295, + 498 + ], + "type": "interline_equation", + "content": "A = \\left[ \\begin{array}{c c c} s _ {i} ^ {x} & 0 & p _ {i} ^ {x} \\\\ 0 & s _ {i} ^ {y} & p _ {i} ^ {y} \\end{array} \\right]. \\tag {1}", + "image_path": "d08762d9e7621e0ff7fc58673f075f45fff486c98b3d1531955c91a5595cabae.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 502, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 526 + ], + "type": "text", + "content": "With this setup, the correspondence between spatial coordinates of patches and input image can now be defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 532, + 295, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 532, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 132, + 532, + 295, + 550 + ], + "type": "interline_equation", + "content": "c = A \\left[ \\begin{array}{l l l} u & v & 1 \\end{array} \\right] ^ {T}, \\tag {2}", + "image_path": "e1929b37cc8d5e9f670341e957797d6fa4f030bcd911d83282b5770014014a45.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "inline_equation", + "content": "c = [c_u, c_v]" + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "text", + "content": " are the spatial coordinates in the original image, while " + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "text", + "content": " are coordinates in the patch space, sampled on a uniform grid. Through sampling patch spatial coordinates with Eq. 2 we can determine the corresponding coordinates in the original image. Since " + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 555, + 296, + 651 + ], + "type": "text", + "content": " is not discrete, proper interpolation is necessary. Inspired by Spatial Transformer Networks [19], we employ bilinear interpolation to address this:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 71, + 666, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 666, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 71, + 666, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} c _ {u, v} = \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} I _ {(h, w)} \\max (0, 1 - | c _ {v} - h |) \\tag {3} \\\\ \\times \\max (0, 1 - | c _ {u} - w |), \\\\ \\end{array}", + "image_path": "e3e14b24c07b18a26bcd62beb624e917173cfcd617e4819a62882d198cc2ae47.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "I_{(h,w)}" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " represents the pixel value in the original image " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": " at the location " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ". For a patch of size " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "\\hat{H} \\times \\hat{W}" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ", the coordinates are sampled as " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "u = \\{1,2,\\dots,\\hat{H}\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "inline_equation", + "content": "v = \\{1,2,\\dots,\\hat{W}\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": ". Unlike the original integration-based approach [19], which is time-consuming for high-resolution tasks, we use a look-up operation to accelerate by considering only the four neighboring pixels around each patch coordinate." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 168, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 555, + 228 + ], + "type": "text", + "content": "To incorporate spatial information for each patch, we concatenate the affine transformation matrix embeddings with each patch's positional embedding. Experimental results in Table 4 demonstrate the effectiveness of this approach in enhancing the model's spatial awareness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 255, + 538, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 255, + 538, + 266 + ], + "spans": [ + { + "bbox": [ + 315, + 255, + 538, + 266 + ], + "type": "text", + "content": "Algorithm 1 Simplified Informative Patch Sampling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 268, + 339, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 268, + 339, + 277 + ], + "spans": [ + { + "bbox": [ + 315, + 268, + 339, + 277 + ], + "type": "text", + "content": "Input:" + } + ] + } + ], + "index": 12 + }, + { + "type": "code", + "bbox": [ + 320, + 277, + 531, + 305 + ], + "blocks": [ + { + "bbox": [ + 320, + 277, + 531, + 305 + ], + "lines": [ + { + "bbox": [ + 320, + 277, + 531, + 305 + ], + "spans": [ + { + "bbox": [ + 320, + 277, + 531, + 305 + ], + "type": "text", + "content": "Bounding boxes " + }, + { + "bbox": [ + 320, + 277, + 531, + 305 + ], + "type": "inline_equation", + "content": "\\mathsf{B} = \\{b_1,\\dots ,b_{\\mathrm{box\\_num}}\\} \\in \\mathbb{R}^{\\mathrm{box\\_num}\\times 4}," + }, + { + "bbox": [ + 320, + 277, + 531, + 305 + ], + "type": "text", + "content": " Each bounding box " + }, + { + "bbox": [ + 320, + 277, + 531, + 305 + ], + "type": "inline_equation", + "content": "b_{j} = [x_{min},y_{min},x_{max},y_{max}]," + }, + { + "bbox": [ + 320, + 277, + 531, + 305 + ], + "type": "text", + "content": " Number of local patches " + }, + { + "bbox": [ + 320, + 277, + 531, + 305 + ], + "type": "inline_equation", + "content": "n" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + } + ], + "index": 13, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 315, + 306, + 344, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 306, + 344, + 315 + ], + "spans": [ + { + "bbox": [ + 315, + 306, + 344, + 315 + ], + "type": "text", + "content": "Output:" + } + ] + } + ], + "index": 14 + }, + { + "type": "code", + "bbox": [ + 320, + 316, + 504, + 334 + ], + "blocks": [ + { + "bbox": [ + 320, + 316, + 504, + 334 + ], + "lines": [ + { + "bbox": [ + 320, + 316, + 504, + 334 + ], + "spans": [ + { + "bbox": [ + 320, + 316, + 504, + 334 + ], + "type": "text", + "content": "Sampled patch locations " + }, + { + "bbox": [ + 320, + 316, + 504, + 334 + ], + "type": "inline_equation", + "content": "\\mathsf{P}\\in \\mathbb{R}^{n\\times 2}" + }, + { + "bbox": [ + 320, + 316, + 504, + 334 + ], + "type": "text", + "content": " Sampled patch scaling factors " + }, + { + "bbox": [ + 320, + 316, + 504, + 334 + ], + "type": "inline_equation", + "content": "\\mathsf{S}\\in \\mathbb{R}^{n\\times 2}" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_body" + } + ], + "index": 15, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 315, + 338, + 512, + 372 + ], + "blocks": [ + { + "bbox": [ + 315, + 338, + 512, + 372 + ], + "lines": [ + { + "bbox": [ + 315, + 338, + 512, + 372 + ], + "spans": [ + { + "bbox": [ + 315, + 338, + 512, + 372 + ], + "type": "text", + "content": "1. Allocate patches Based on Area area " + }, + { + "bbox": [ + 315, + 338, + 512, + 372 + ], + "type": "inline_equation", + "content": "(\\mathsf{b_j}) = (\\mathbf{x}_{\\mathrm{max}} - \\mathbf{x}_{\\mathrm{min}})(\\mathbf{y}_{\\mathrm{max}} - \\mathbf{y}_{\\mathrm{min}})" + }, + { + "bbox": [ + 315, + 338, + 512, + 372 + ], + "type": "inline_equation", + "content": "\\forall \\mathsf{b_j}\\in \\mathsf{B}" + }, + { + "bbox": [ + 315, + 338, + 512, + 372 + ], + "type": "text", + "content": " n=j = num-patches(bj) = area(bj) n, " + }, + { + "bbox": [ + 315, + 338, + 512, + 372 + ], + "type": "inline_equation", + "content": "\\forall \\mathsf{b_j}\\in \\mathsf{B}" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "code_body" + } + ], + "index": 16, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "bbox": [ + 315, + 374, + 513, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 374, + 513, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 374, + 513, + 392 + ], + "type": "text", + "content": "2. Generate patch locations and scales for " + }, + { + "bbox": [ + 315, + 374, + 513, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_{\\mathrm{j}}\\in \\mathsf{B}" + }, + { + "bbox": [ + 315, + 374, + 513, + 392 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 331, + 395, + 539, + 489 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 395, + 539, + 489 + ], + "spans": [ + { + "bbox": [ + 331, + 395, + 539, + 489 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\cdot \\operatorname {w i d t h} (\\mathfrak {b} _ {\\mathrm {j}}) = \\mathbf {x} _ {\\max } - \\mathbf {x} _ {\\min }, \\operatorname {h e i g h t} (\\mathfrak {b} _ {\\mathrm {j}}) = \\mathbf {y} _ {\\max } - \\mathbf {y} _ {\\min } \\\\ \\cdot \\text {p a t c h e s} _ {\\mathrm {w}}. \\text {s i d e} = \\sqrt {\\frac {\\operatorname {n u m} . \\operatorname {p a t c h e s} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right) \\cdot \\operatorname {w i d t h} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right)}{\\operatorname {h e i g h t} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right)}} \\\\ \\cdot \\text {p a t c h e s} _ {\\mathrm {h}} \\text {s i d e} = \\sqrt {\\frac {\\operatorname {n u m} . \\operatorname {p a t c h e s} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right) \\cdot \\operatorname {h e i g h t} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right)}{\\operatorname {w i d t h} \\left(\\mathbf {b} _ {\\mathrm {j}}\\right)}} \\\\ \\bullet \\mathrm {x} = \\operatorname {l i n s p a c e} \\left(\\mathbf {x} _ {\\min }, \\mathbf {x} _ {\\max }, \\text {p a t c h e s - w - s i d e}\\right) \\\\ \\cdot \\mathbf {y} = \\operatorname {l i n s p a c e} \\left(\\mathbf {y} _ {\\min }, \\mathbf {y} _ {\\max }, \\text {p a t c h e s} _ {\\mathrm {h}} \\text {s i d e}\\right) \\\\ \\cdot P _ {j} = \\text {m e s h g r i d} (x, y) \\\\ \\bullet \\mathbf {S} _ {\\mathrm {j}} = \\left(\\frac {\\operatorname {w i d t h} (\\mathbf {b} _ {\\mathrm {j}})}{\\text {p a t c h e s . w . s i d e}}, \\frac {\\operatorname {h e i g h t} (\\mathbf {b} _ {\\mathrm {j}})}{\\text {p a t c h e s . h . s i d e}}\\right). \\exp a n d (\\mathbf {n} _ {\\mathrm {j}}, 2) \\\\ \\end{array}", + "image_path": "39cfb45e9a018425e009563ba99eef4082f054dc0a827d0f0e5734a14f18a683.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 491, + 329, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 491, + 329, + 499 + ], + "spans": [ + { + "bbox": [ + 315, + 491, + 329, + 499 + ], + "type": "text", + "content": "end" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 504, + 460, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 504, + 460, + 513 + ], + "spans": [ + { + "bbox": [ + 315, + 504, + 460, + 513 + ], + "type": "text", + "content": "3. Remove redundant patches" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 514, + 547, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 514, + 547, + 523 + ], + "spans": [ + { + "bbox": [ + 320, + 514, + 547, + 523 + ], + "type": "text", + "content": "P,S " + }, + { + "bbox": [ + 320, + 514, + 547, + 523 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 320, + 514, + 547, + 523 + ], + "type": "text", + "content": " Remove_overlapedpatches(Concat(Pj),Concat(Sj))" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 524, + 367, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 524, + 367, + 531 + ], + "spans": [ + { + "bbox": [ + 315, + 524, + 367, + 531 + ], + "type": "text", + "content": "return P and S" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 552, + 488, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 488, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 488, + 564 + ], + "type": "text", + "content": "3.3. Subject-Oriented Data Curation" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": "To fine-tune our TEVA model for dynamically sampled patches, we introduce a data curation pipeline that creates subject-oriented training data, emphasizing specific regions of interest within images. While large datasets [3, 51] are commonly used to enhance the generalization of VLMs, they typically consist of image-caption pairs providing broad, whole-image descriptions. Our approach curates question-answer pairs that emphasize zoomed-in regions, enabling the model to better understand and prioritize detailed image areas. To address this, we developed an efficient method that curates large-scale, subject-oriented conversations (8 million pairs) based on existing image-caption" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24150" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 76, + 119, + 137 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 119, + 137 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 119, + 137 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 119, + 137 + ], + "type": "image", + "image_path": "42471df3e5794d0cd39d3db5b0ef00738f654aa5888678eb70205e736b071bda.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 139, + 119, + 152 + ], + "lines": [ + { + "bbox": [ + 63, + 139, + 119, + 152 + ], + "spans": [ + { + "bbox": [ + 63, + 139, + 119, + 152 + ], + "type": "text", + "content": "The image is only for illustration. The images for subject-oriented VCA generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 124, + 76, + 290, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 76, + 290, + 156 + ], + "spans": [ + { + "bbox": [ + 124, + 76, + 290, + 156 + ], + "type": "text", + "content": "Image Caption: The image portrays a fairy with white wings and a purple dress, standing in a field of lavender flowers. She is holding a lavender plant in her hands. The background is a light blue color, and there are two butterflies visible in the scene. The fairy is positioned in the center of the image, with the lavender plants surrounding her. The butterflies are located in the top left and bottom right corners of the image. The fairy's wings are spread out, and her dress is flowing, giving the impression of movement. The lavender plants are tall and thin, with small purple flowers at the top. The butterflies are small and white, with black spots on their wings. The overall color scheme of the image is purple, white, and light blue." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 63, + 162, + 78, + 180 + ], + "blocks": [ + { + "bbox": [ + 63, + 162, + 78, + 180 + ], + "lines": [ + { + "bbox": [ + 63, + 162, + 78, + 180 + ], + "spans": [ + { + "bbox": [ + 63, + 162, + 78, + 180 + ], + "type": "image", + "image_path": "8fba0d29514340cd3abd3b56a4d7ffb1f2eac1acd306f445dfc941a3496d5d6e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 232, + 296, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 277 + ], + "type": "text", + "content": "Figure 4. Example of constructing Subject-Oriented QA pairs. Using the image captions, we employ the Llama3.1 [59] to generate an average of three questions per image, resulting in 8 million QA pairs focused on different objects described in each caption." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 163, + 288, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 163, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 83, + 163, + 288, + 183 + ], + "type": "text", + "content": " Generate three question-answer pairs based on this image description Each question should ask about one different object in the image without repeating objects." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 190, + 79, + 204 + ], + "blocks": [ + { + "bbox": [ + 63, + 190, + 79, + 204 + ], + "lines": [ + { + "bbox": [ + 63, + 190, + 79, + 204 + ], + "spans": [ + { + "bbox": [ + 63, + 190, + 79, + 204 + ], + "type": "image", + "image_path": "b2cafc2701ac137f928f5b6ccf1737aae6d577a336bc8dcbc4d8b04be533aa11.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 204, + 80, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 204, + 80, + 210 + ], + "spans": [ + { + "bbox": [ + 63, + 204, + 80, + 210 + ], + "type": "text", + "content": "Llama" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 85, + 191, + 288, + 218 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 85, + 191, + 288, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 191, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 85, + 191, + 288, + 205 + ], + "type": "text", + "content": "1. Question: what is the fairy wearing? Answer: The fairy is wearing a purple dress. \n2. Question: How many butterflies are there? Answer: There are two." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 205, + 287, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 205, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 85, + 205, + 287, + 218 + ], + "type": "text", + "content": "3. Question: What do the lavender plants look like? Answer: They are tall and thin, with small purple flowers at the top." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 298, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 298, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 298, + 296, + 441 + ], + "type": "text", + "content": "datasets [3, 4, 11]. As illustrated in Figure 4, we leverage Llama 3.1 [59] for its strong contextual language reasoning ability, using only structured text prompts as input to generate subject-focused conversations. After generating these conversations, we apply an open-vocabulary detection method to validate the presence of specified objects in the images, filtering out any noisy data and ensuring data quality. Experiments in Table 4 demonstrate the significance of this pretraining dataset, especially for challenging benchmarks that require an understanding of zoomed-in patches or complex reasoning tasks. The details of data curation are available in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 453, + 137, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 137, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 137, + 466 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 472, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 556 + ], + "type": "text", + "content": "In this section, we provide the training settings and evaluation benchmarks employed to assess our model's performance on high-resolution and general Vision-Language Model (VLM) tasks. Subsequently, we present comparative results, benchmarking our method against state-of-the-art approaches. Further details are available in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 564, + 172, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 172, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 172, + 577 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 677 + ], + "type": "text", + "content": "Training stages The goal of TEVA is to enhance the model's ability to process high-resolution images by focusing on informative patches. Two key requirements are essential for achieving this: (1) enabling the vision tower to handle dynamic patch sampling, and (2) ensuring that dynamic vision tokens can be effectively interpreted by the LLM. To meet these goals, we structured our training pipeline into four stages:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "- Alignment Stage: Following the process used in LLaVA [37], we begin by aligning visual features with the language space. In this stage, we freeze the large" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 555, + 120 + ], + "type": "text", + "content": "language model (LLM) and optimize only the vision-language (VL) adapter with LLaVA alignment dataset (595k), ensuring that the visual features are properly mapped to the language domain." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 121, + 555, + 360 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 314, + 121, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 555, + 180 + ], + "type": "text", + "content": "- Vision Tower Adapting: Using 4M samples from our proposed Subject-Oriented Dataset, we fine-tune the vision tower to handle dynamic image patches. During this stage, both the vision tower and VL adapter are unfrozen for training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 180, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 180, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 180, + 554, + 262 + ], + "type": "text", + "content": "- Integrated Interpretation Stage: To further enhance the LLM's ability to interpret dynamic vision tokens, we unfreeze the vision tower, VL adapter, and LLM, then fine-tune using 4M samples from the Subject-Oriented dataset and 1M samples from the ShareGPT4V [4] caption dataset. This step aligns all components for seamless interpretation of dynamic visual information." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 263, + 554, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 263, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 314, + 263, + 554, + 360 + ], + "type": "text", + "content": "- Instruction Fine-tuning: In the final stage, following the training strategy of VILA [37], we continue fine-tuning the vision tower, VL adapter, and LLM with a mix of Supervised Fine-Tuning (SFT) Dataset (4.17M) built on VILA [37]. This instruction fine-tuning stage enhances the model's instruction-following and dialogue capabilities on multiple tasks. Details of the training settings and datasets are provided in the supplementary materials." + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "type": "text", + "content": "Evaluation benchmarks To assess our model's capability in handling complex high-resolution images, we report results on challenging high-resolution benchmarks, including " + }, + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "type": "text", + "content": " [61] and MME-RealWorld [79]. Additionally, we evaluate our model on commonly used general VQA and MLLM benchmarks including MMBench [41], SQA (ScienceQA) [43], " + }, + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathrm{VQAT}" + }, + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "type": "text", + "content": " (TextVQA) [54], POPE [29], AI2D [22], LLaVAw (LLaVA-Bench In-the-Wild) [37], " + }, + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathrm{MMMU}^{\\mathrm{t}}" + }, + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "type": "text", + "content": " (MMMU-test) [69], and " + }, + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathrm{VQAV}^{\\mathrm{v2}}" + }, + { + "bbox": [ + 313, + 361, + 555, + 492 + ], + "type": "text", + "content": " (VQA-v2) [14]. Evaluation of more benchmarks is available in the supplementary." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 503, + 400, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 400, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 400, + 514 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "Comparison with region-aware VLMs. We evaluate Token-Efficient Vision LLanguage Model (TEVA) on the high-resolution visual detail benchmark " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " [61], comparing its performance and efficiency against state-of-the-art region-aware VLMs. As shown in Table 1, TEVA consistently outperforms existing methods in attribute recognition (AR), spatial relationship reasoning (SRR), and scene OCR tasks. Our dynamic patch sampling strategy significantly reduces computational overhead compared to SpatialGPT (mask pooling) [6], Griffon-G [72] and " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " [61], which require multiple vision tower passes to process individual ROIs. Specifically, TEVA achieves a " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "7.53 \\times" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " speedup over " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " in ROI aggregation and improves overall inference efficiency by " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "12.31 \\times" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " [61], benefiting from both token efficiency and the lightweight RAP module, which eliminates the need for iterative region searching." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24151" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 88, + 217, + 187 + ], + "blocks": [ + { + "bbox": [ + 65, + 76, + 216, + 87 + ], + "lines": [ + { + "bbox": [ + 65, + 76, + 216, + 87 + ], + "spans": [ + { + "bbox": [ + 65, + 76, + 216, + 87 + ], + "type": "text", + "content": "How many people are there on the street?" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 88, + 217, + 187 + ], + "lines": [ + { + "bbox": [ + 65, + 88, + 217, + 187 + ], + "spans": [ + { + "bbox": [ + 65, + 88, + 217, + 187 + ], + "type": "image", + "image_path": "085235beebac11b4f33d69d6877492a23ce14f4570b380128d8a1f963079ef42.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 196, + 93, + 203 + ], + "lines": [ + { + "bbox": [ + 66, + 196, + 93, + 203 + ], + "spans": [ + { + "bbox": [ + 66, + 196, + 93, + 203 + ], + "type": "text", + "content": "VILA-1.5" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 111, + 191, + 205, + 209 + ], + "lines": [ + { + "bbox": [ + 111, + 191, + 205, + 209 + ], + "spans": [ + { + "bbox": [ + 111, + 191, + 205, + 209 + ], + "type": "text", + "content": "There is only one person on the street." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 72, + 232, + 86, + 243 + ], + "blocks": [ + { + "bbox": [ + 69, + 217, + 92, + 224 + ], + "lines": [ + { + "bbox": [ + 69, + 217, + 92, + 224 + ], + "spans": [ + { + "bbox": [ + 69, + 217, + 92, + 224 + ], + "type": "text", + "content": "Monkey" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 72, + 232, + 86, + 243 + ], + "lines": [ + { + "bbox": [ + 72, + 232, + 86, + 243 + ], + "spans": [ + { + "bbox": [ + 72, + 232, + 86, + 243 + ], + "type": "image", + "image_path": "f8b68f9c05c5bc06578ee1d47e0188bebaeb629b3f58091a4e5a4f62b091e58e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 244, + 97, + 251 + ], + "lines": [ + { + "bbox": [ + 62, + 244, + 97, + 251 + ], + "spans": [ + { + "bbox": [ + 62, + 244, + 97, + 251 + ], + "type": "text", + "content": "ChatGPT-40" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 72, + 253, + 87, + 270 + ], + "blocks": [ + { + "bbox": [ + 72, + 253, + 87, + 270 + ], + "lines": [ + { + "bbox": [ + 72, + 253, + 87, + 270 + ], + "spans": [ + { + "bbox": [ + 72, + 253, + 87, + 270 + ], + "type": "image", + "image_path": "1c03c33f34dd65596fba737f31f944c0d4911914a99ecc3cf3149604dbb26c5e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 232, + 88, + 384, + 187 + ], + "blocks": [ + { + "bbox": [ + 249, + 76, + 355, + 87 + ], + "lines": [ + { + "bbox": [ + 249, + 76, + 355, + 87 + ], + "spans": [ + { + "bbox": [ + 249, + 76, + 355, + 87 + ], + "type": "text", + "content": "What is the breed of the dog?" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 232, + 88, + 384, + 187 + ], + "lines": [ + { + "bbox": [ + 232, + 88, + 384, + 187 + ], + "spans": [ + { + "bbox": [ + 232, + 88, + 384, + 187 + ], + "type": "image", + "image_path": "3d14656949e1ee1f5aacdbb1ed0b61243d3bb2100ff5996ec3031812e72783e7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 196, + 259, + 203 + ], + "lines": [ + { + "bbox": [ + 231, + 196, + 259, + 203 + ], + "spans": [ + { + "bbox": [ + 231, + 196, + 259, + 203 + ], + "type": "text", + "content": "VILA-1.5" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 277, + 192, + 370, + 210 + ], + "lines": [ + { + "bbox": [ + 277, + 192, + 370, + 210 + ], + "spans": [ + { + "bbox": [ + 277, + 192, + 370, + 210 + ], + "type": "text", + "content": "The dog in the image is a white dog." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 316, + 218, + 329, + 226 + ], + "lines": [ + { + "bbox": [ + 316, + 218, + 329, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 218, + 329, + 226 + ], + "type": "text", + "content": "Pug." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 227, + 232, + 264, + 251 + ], + "blocks": [ + { + "bbox": [ + 108, + 234, + 208, + 251 + ], + "lines": [ + { + "bbox": [ + 108, + 234, + 208, + 251 + ], + "spans": [ + { + "bbox": [ + 108, + 234, + 208, + 251 + ], + "type": "text", + "content": "There is one person visible on the street." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 233, + 217, + 257, + 224 + ], + "lines": [ + { + "bbox": [ + 233, + 217, + 257, + 224 + ], + "spans": [ + { + "bbox": [ + 233, + 217, + 257, + 224 + ], + "type": "text", + "content": "Monkey" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 227, + 232, + 264, + 251 + ], + "lines": [ + { + "bbox": [ + 227, + 232, + 264, + 251 + ], + "spans": [ + { + "bbox": [ + 227, + 232, + 264, + 251 + ], + "type": "image", + "image_path": "60af72dfd219e8d855f0abfbec23ebe28574bb77f811b94b3f3e49117486cbd2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 239, + 372, + 247 + ], + "lines": [ + { + "bbox": [ + 285, + 239, + 372, + 247 + ], + "spans": [ + { + "bbox": [ + 285, + 239, + 372, + 247 + ], + "type": "text", + "content": "difficult to identify the breed." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 238, + 253, + 252, + 270 + ], + "blocks": [ + { + "bbox": [ + 238, + 253, + 252, + 270 + ], + "lines": [ + { + "bbox": [ + 238, + 253, + 252, + 270 + ], + "spans": [ + { + "bbox": [ + 238, + 253, + 252, + 270 + ], + "type": "image", + "image_path": "43cb6b63581acaebe012f234d233b642e9323d55430d6ff4c705cabc2204a7c1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 399, + 88, + 499, + 186 + ], + "blocks": [ + { + "bbox": [ + 416, + 76, + 520, + 87 + ], + "lines": [ + { + "bbox": [ + 416, + 76, + 520, + 87 + ], + "spans": [ + { + "bbox": [ + 416, + 76, + 520, + 87 + ], + "type": "text", + "content": "What color is the bottle cap?" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 88, + 499, + 186 + ], + "lines": [ + { + "bbox": [ + 399, + 88, + 499, + 186 + ], + "spans": [ + { + "bbox": [ + 399, + 88, + 499, + 186 + ], + "type": "image", + "image_path": "91f5ccf34abeafccae6483240717f387b505bce7a61f9d6a27230b3755b76cf0.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 196, + 424, + 203 + ], + "lines": [ + { + "bbox": [ + 397, + 196, + 424, + 203 + ], + "spans": [ + { + "bbox": [ + 397, + 196, + 424, + 203 + ], + "type": "text", + "content": "VILA-1.5" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 452, + 196, + 527, + 205 + ], + "lines": [ + { + "bbox": [ + 452, + 196, + 527, + 205 + ], + "spans": [ + { + "bbox": [ + 452, + 196, + 527, + 205 + ], + "type": "text", + "content": "The bottle has a blue cap." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 477, + 217, + 501, + 224 + ], + "lines": [ + { + "bbox": [ + 477, + 217, + 501, + 224 + ], + "spans": [ + { + "bbox": [ + 477, + 217, + 501, + 224 + ], + "type": "text", + "content": "Yellow." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 501, + 88, + 549, + 156 + ], + "blocks": [ + { + "bbox": [ + 501, + 88, + 549, + 156 + ], + "lines": [ + { + "bbox": [ + 501, + 88, + 549, + 156 + ], + "spans": [ + { + "bbox": [ + 501, + 88, + 549, + 156 + ], + "type": "image", + "image_path": "0a0b956b9bb5161918f9cc06637ccc05d35f17fdcbe11e51da90eda1f5859fa5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 403, + 232, + 417, + 244 + ], + "blocks": [ + { + "bbox": [ + 399, + 217, + 422, + 224 + ], + "lines": [ + { + "bbox": [ + 399, + 217, + 422, + 224 + ], + "spans": [ + { + "bbox": [ + 399, + 217, + 422, + 224 + ], + "type": "text", + "content": "Monkey" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 393, + 244, + 429, + 251 + ], + "lines": [ + { + "bbox": [ + 393, + 244, + 429, + 251 + ], + "spans": [ + { + "bbox": [ + 393, + 244, + 429, + 251 + ], + "type": "text", + "content": "ChatGPT-40" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 403, + 232, + 417, + 244 + ], + "lines": [ + { + "bbox": [ + 403, + 232, + 417, + 244 + ], + "spans": [ + { + "bbox": [ + 403, + 232, + 417, + 244 + ], + "type": "image", + "image_path": "553d44c330c57927bc9f3368901a6b50ef78cdd9ca48860753933d2268cb6515.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 234, + 540, + 250 + ], + "lines": [ + { + "bbox": [ + 440, + 234, + 540, + 250 + ], + "spans": [ + { + "bbox": [ + 440, + 234, + 540, + 250 + ], + "type": "text", + "content": "The color of the cap appears to be red." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 67, + 293, + 98, + 335 + ], + "blocks": [ + { + "bbox": [ + 96, + 284, + 182, + 293 + ], + "lines": [ + { + "bbox": [ + 96, + 284, + 182, + 293 + ], + "spans": [ + { + "bbox": [ + 96, + 284, + 182, + 293 + ], + "type": "text", + "content": "What is the speed limit?" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 293, + 98, + 335 + ], + "lines": [ + { + "bbox": [ + 67, + 293, + 98, + 335 + ], + "spans": [ + { + "bbox": [ + 67, + 293, + 98, + 335 + ], + "type": "image", + "image_path": "6f24cc55cee359e0aba86ad7b6cb7d82683414772db15359e36f7def45fb098e.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 99, + 293, + 217, + 392 + ], + "blocks": [ + { + "bbox": [ + 138, + 259, + 178, + 266 + ], + "lines": [ + { + "bbox": [ + 138, + 259, + 178, + 266 + ], + "spans": [ + { + "bbox": [ + 138, + 259, + 178, + 266 + ], + "type": "text", + "content": "There are 2." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 99, + 293, + 217, + 392 + ], + "lines": [ + { + "bbox": [ + 99, + 293, + 217, + 392 + ], + "spans": [ + { + "bbox": [ + 99, + 293, + 217, + 392 + ], + "type": "image", + "image_path": "c0ba9204fef177b283b3250feb35545e346d371c02659dbb4cbbd736ffe99c3d.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 401, + 93, + 408 + ], + "lines": [ + { + "bbox": [ + 66, + 401, + 93, + 408 + ], + "spans": [ + { + "bbox": [ + 66, + 401, + 93, + 408 + ], + "type": "text", + "content": "VILA-1.5" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 136, + 403, + 176, + 409 + ], + "lines": [ + { + "bbox": [ + 136, + 403, + 176, + 409 + ], + "spans": [ + { + "bbox": [ + 136, + 403, + 176, + 409 + ], + "type": "text", + "content": "It is " + }, + { + "bbox": [ + 136, + 403, + 176, + 409 + ], + "type": "inline_equation", + "content": "10\\mathrm{km / h}" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 151, + 423, + 161, + 430 + ], + "lines": [ + { + "bbox": [ + 151, + 423, + 161, + 430 + ], + "spans": [ + { + "bbox": [ + 151, + 423, + 161, + 430 + ], + "type": "text", + "content": "50." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 73, + 437, + 86, + 449 + ], + "blocks": [ + { + "bbox": [ + 67, + 422, + 90, + 430 + ], + "lines": [ + { + "bbox": [ + 67, + 422, + 90, + 430 + ], + "spans": [ + { + "bbox": [ + 67, + 422, + 90, + 430 + ], + "type": "text", + "content": "Monkey" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 73, + 437, + 86, + 449 + ], + "lines": [ + { + "bbox": [ + 73, + 437, + 86, + 449 + ], + "spans": [ + { + "bbox": [ + 73, + 437, + 86, + 449 + ], + "type": "image", + "image_path": "c43006d0fce0039a6d9d486f3787ef41bd340821e1b4bd761cfbf16c5ac1d097.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 449, + 97, + 456 + ], + "lines": [ + { + "bbox": [ + 62, + 449, + 97, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 97, + 456 + ], + "type": "text", + "content": "ChatGPT-40" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 74, + 460, + 86, + 475 + ], + "blocks": [ + { + "bbox": [ + 74, + 460, + 86, + 475 + ], + "lines": [ + { + "bbox": [ + 74, + 460, + 86, + 475 + ], + "spans": [ + { + "bbox": [ + 74, + 460, + 86, + 475 + ], + "type": "image", + "image_path": "953c4b79f9ba2434e1c42fc26b0130d3ccff034123a9ebf8a280b3fb2718630c.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 464, + 168, + 472 + ], + "lines": [ + { + "bbox": [ + 143, + 464, + 168, + 472 + ], + "spans": [ + { + "bbox": [ + 143, + 464, + 168, + 472 + ], + "type": "text", + "content": "It is 40." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 491, + 555, + 568 + ], + "lines": [ + { + "bbox": [ + 55, + 491, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 555, + 568 + ], + "type": "text", + "content": "Figure 5. Evaluation of our model against leading open-source VLMs (VILA-1.5 [32], Monkey [31]) and the advanced commercial VLM, GPT-4o [50], in high-resolution, real-world scenarios. Each image displays the proposed subject in the bottom left corner. The first row presents responses to general questions, while the second row addresses more challenging OCR-based questions. The proposed subject is in the left-bottom corner of each image. Despite using a smaller language model (Sheared-LLaMA-2.7B [62]), our model demonstrates robust capability in observing precise details for high-resolution images and can adapt to various image resolutions, providing flexibility and enhanced performance in demanding high-resolution scenarios. Please note that all sample images were processed at their original resolution " + }, + { + "bbox": [ + 55, + 491, + 555, + 568 + ], + "type": "inline_equation", + "content": "(2246\\times 1582" + }, + { + "bbox": [ + 55, + 491, + 555, + 568 + ], + "type": "text", + "content": " on average) for TEVA-3B." + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 233, + 293, + 384, + 392 + ], + "blocks": [ + { + "bbox": [ + 294, + 259, + 352, + 266 + ], + "lines": [ + { + "bbox": [ + 294, + 259, + 352, + 266 + ], + "spans": [ + { + "bbox": [ + 294, + 259, + 352, + 266 + ], + "type": "text", + "content": "Golden Retriever." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 237, + 284, + 373, + 293 + ], + "lines": [ + { + "bbox": [ + 237, + 284, + 373, + 293 + ], + "spans": [ + { + "bbox": [ + 237, + 284, + 373, + 293 + ], + "type": "text", + "content": "What is the written on the large boat?" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 293, + 384, + 392 + ], + "lines": [ + { + "bbox": [ + 233, + 293, + 384, + 392 + ], + "spans": [ + { + "bbox": [ + 233, + 293, + 384, + 392 + ], + "type": "image", + "image_path": "8a7c6982364fe91658d78a93f969354fc3d4cc145684e03181d5c77733af1a15.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 402, + 260, + 409 + ], + "lines": [ + { + "bbox": [ + 233, + 402, + 260, + 409 + ], + "spans": [ + { + "bbox": [ + 233, + 402, + 260, + 409 + ], + "type": "text", + "content": "VILA-1.5" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 287, + 403, + 361, + 410 + ], + "lines": [ + { + "bbox": [ + 287, + 403, + 361, + 410 + ], + "spans": [ + { + "bbox": [ + 287, + 403, + 361, + 410 + ], + "type": "text", + "content": "\"Oranje\" is written on it." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 302, + 423, + 345, + 430 + ], + "lines": [ + { + "bbox": [ + 302, + 423, + 345, + 430 + ], + "spans": [ + { + "bbox": [ + 302, + 423, + 345, + 430 + ], + "type": "text", + "content": "Unanswerable." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 226, + 439, + 264, + 456 + ], + "blocks": [ + { + "bbox": [ + 108, + 444, + 205, + 453 + ], + "lines": [ + { + "bbox": [ + 108, + 444, + 205, + 453 + ], + "spans": [ + { + "bbox": [ + 108, + 444, + 205, + 453 + ], + "type": "text", + "content": "The speed limit appears to be 30." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 233, + 422, + 257, + 430 + ], + "lines": [ + { + "bbox": [ + 233, + 422, + 257, + 430 + ], + "spans": [ + { + "bbox": [ + 233, + 422, + 257, + 430 + ], + "type": "text", + "content": "Monkey" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 226, + 439, + 264, + 456 + ], + "lines": [ + { + "bbox": [ + 226, + 439, + 264, + 456 + ], + "spans": [ + { + "bbox": [ + 226, + 439, + 264, + 456 + ], + "type": "image", + "image_path": "15b6f7534fd3582d3f29008687d15573d545b17fd122636612f695d70f1fed32.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 440, + 362, + 456 + ], + "lines": [ + { + "bbox": [ + 285, + 440, + 362, + 456 + ], + "spans": [ + { + "bbox": [ + 285, + 440, + 362, + 456 + ], + "type": "text", + "content": "The text on the boat reads \"INTERCOMBI\"." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 311, + 465, + 335, + 472 + ], + "lines": [ + { + "bbox": [ + 311, + 465, + 335, + 472 + ], + "spans": [ + { + "bbox": [ + 311, + 465, + 335, + 472 + ], + "type": "text", + "content": "Nitrico." + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 241, + 460, + 253, + 475 + ], + "blocks": [ + { + "bbox": [ + 241, + 460, + 253, + 475 + ], + "lines": [ + { + "bbox": [ + 241, + 460, + 253, + 475 + ], + "spans": [ + { + "bbox": [ + 241, + 460, + 253, + 475 + ], + "type": "image", + "image_path": "2b95613fedc10d8511381c4aebc4f39423c81bf8867134618bcacc85adec5511.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 397, + 293, + 549, + 392 + ], + "blocks": [ + { + "bbox": [ + 476, + 259, + 503, + 266 + ], + "lines": [ + { + "bbox": [ + 476, + 259, + 503, + 266 + ], + "spans": [ + { + "bbox": [ + 476, + 259, + 503, + 266 + ], + "type": "text", + "content": "It is red." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 405, + 284, + 539, + 293 + ], + "lines": [ + { + "bbox": [ + 405, + 284, + 539, + 293 + ], + "spans": [ + { + "bbox": [ + 405, + 284, + 539, + 293 + ], + "type": "text", + "content": "Tell me the yellow text on the airship." + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 397, + 293, + 549, + 392 + ], + "lines": [ + { + "bbox": [ + 397, + 293, + 549, + 392 + ], + "spans": [ + { + "bbox": [ + 397, + 293, + 549, + 392 + ], + "type": "image", + "image_path": "ec8214adc5331ac7380da365b2f65983cd5dfdda8eeb9d1cc8aab824a5de3a1d.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 402, + 424, + 409 + ], + "lines": [ + { + "bbox": [ + 397, + 402, + 424, + 409 + ], + "spans": [ + { + "bbox": [ + 397, + 402, + 424, + 409 + ], + "type": "text", + "content": "VILA-1.5" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 474, + 403, + 505, + 410 + ], + "lines": [ + { + "bbox": [ + 474, + 403, + 505, + 410 + ], + "spans": [ + { + "bbox": [ + 474, + 403, + 505, + 410 + ], + "type": "text", + "content": "\"Airship\"." + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 468, + 423, + 511, + 430 + ], + "lines": [ + { + "bbox": [ + 468, + 423, + 511, + 430 + ], + "spans": [ + { + "bbox": [ + 468, + 423, + 511, + 430 + ], + "type": "text", + "content": "Unanswerable." + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 407, + 439, + 419, + 450 + ], + "blocks": [ + { + "bbox": [ + 400, + 423, + 424, + 430 + ], + "lines": [ + { + "bbox": [ + 400, + 423, + 424, + 430 + ], + "spans": [ + { + "bbox": [ + 400, + 423, + 424, + 430 + ], + "type": "text", + "content": "Monkey" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 407, + 439, + 419, + 450 + ], + "lines": [ + { + "bbox": [ + 407, + 439, + 419, + 450 + ], + "spans": [ + { + "bbox": [ + 407, + 439, + 419, + 450 + ], + "type": "image", + "image_path": "feb679bfac17eff5710fe9a143d10bc73df4d2b69d6835edc5a07f2ee83979b1.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 451, + 430, + 457 + ], + "lines": [ + { + "bbox": [ + 395, + 451, + 430, + 457 + ], + "spans": [ + { + "bbox": [ + 395, + 451, + 430, + 457 + ], + "type": "text", + "content": "ChatGPT-40" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 455, + 440, + 523, + 456 + ], + "lines": [ + { + "bbox": [ + 455, + 440, + 523, + 456 + ], + "spans": [ + { + "bbox": [ + 455, + 440, + 523, + 456 + ], + "type": "text", + "content": "The text appears to say \"GOODYEAR\"." + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 406, + 460, + 419, + 474 + ], + "blocks": [ + { + "bbox": [ + 406, + 460, + 419, + 474 + ], + "lines": [ + { + "bbox": [ + 406, + 460, + 419, + 474 + ], + "spans": [ + { + "bbox": [ + 406, + 460, + 419, + 474 + ], + "type": "image", + "image_path": "895625fa76c67137b73f26c3e736eeba01853e859ab89f60e1ef97bcd55a0e3f.jpg" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 460, + 523, + 476 + ], + "lines": [ + { + "bbox": [ + 456, + 460, + 523, + 476 + ], + "spans": [ + { + "bbox": [ + 456, + 460, + 523, + 476 + ], + "type": "text", + "content": "The yellow text reads \"GoodYear\"." + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 62 + }, + { + "bbox": [ + 55, + 590, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 295, + 638 + ], + "type": "text", + "content": "Unlike SpatialGPT [6] and Griffon-G [72], which rely on manual region selection, TEVA automatically detects and proposes ROIs. For a fair comparison, we use TEVA's RAP module to predict ROIs for SpatialGPT and Griffon-G." + } + ] + } + ], + "index": 65 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "High resolution benchmark results. To demonstrate TEVA's capability in handling high-resolution images, we evaluated our method on high-resolution VLM benchmarks: " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " [61], with an average image resolution of " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "2246 \\times 1582" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " pixels and target subjects, and MME-RealWorld [79], with an average resolution of " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "2000 \\times 1500" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 66 + }, + { + "bbox": [ + 313, + 590, + 555, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 710 + ], + "type": "text", + "content": "As shown in Table 2, TEVA excels in tasks that require precise detail recognition, such as attribute recognition (AR) and spatial relationship reasoning (SRR) in " + }, + { + "bbox": [ + 313, + 590, + 555, + 710 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 590, + 555, + 710 + ], + "type": "text", + "content": " benchmark, outperforming both the commercial model GPT-4V [50] by 21.6 points and all the other open-source high resolution VLMs. In the complex MME-RealWorld benchmark, TEVA also surpasses open-source methods in both perception and reasoning ability. In applications that demand high-resolution fine visual detail perception, such as remote sensing, video monitoring, and autonomous driv" + } + ] + } + ], + "index": 67 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24152" + } + ] + } + ], + "index": 68 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 71, + 220, + 228 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 220, + 228 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 220, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 220, + 228 + ], + "type": "table", + "html": "
MethodV* Bench
ARSRROCR
SpatialGPT [6]36.556.540.0
Griffon-G [72]44.771.056.6
V* [61]74.776.380.0
TEVA-7B76.577.683.3
MethodInference Time
AIPFPI
SpatialGPT [6]9.57×1.05×
Griffon-G [72]10.24×4.83×
V* [61]7.53×12.31×
TEVA-7B1.00×1.00×
", + "image_path": "3dd3fba3ce5262ce457a8915186bf32b6e2c436b4851f2856f88ec911646b3f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 243, + 71, + 543, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 232, + 228, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 228, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 228, + 353 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art region-aware 7B VLMs on the high-resolution " + }, + { + "bbox": [ + 55, + 232, + 228, + 353 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 55, + 232, + 228, + 353 + ], + "type": "text", + "content": " benchmark, evaluating attribute recognition (AR), spatial relationship reasoning (SRR), and scene OCR tasks. Inference times are reported for Additional Information Processing (AIP) and Full Pipeline Inference (FPI), demonstrating the efficiency and performance gains of dynamic patch sampling over SpatialGPT(pooling), Griffon-G and " + }, + { + "bbox": [ + 55, + 232, + 228, + 353 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 55, + 232, + 228, + 353 + ], + "type": "text", + "content": " (multiple vision tower passes)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 243, + 71, + 543, + 228 + ], + "lines": [ + { + "bbox": [ + 243, + 71, + 543, + 228 + ], + "spans": [ + { + "bbox": [ + 243, + 71, + 543, + 228 + ], + "type": "table", + "html": "
MethodV* AverageMME-RWPRCMME-RWRSN
OCRRSMOADOCRMOAD
GPT-4V [50]55.9-------
GPT-4o [50]-77.728.933.922.461.436.526.4
Gemini Pro [56]50.5-------
Gemini-1.5-pro [56]-67.614.031.126.652.728.319.2
LLaVA-1.5 [35]50.538.722.119.126.026.025.924.2
LLaVA-1.5-S2[52]56.5-------
Text-Monkey [40]-37.311.716.114.330.44.420.0
Monkey [31]-54.625.028.029.727.227.333.0
LLaVA-Next [36]61.340.631.635.427.214.929.528.2
Cambrian-1 [57]60.832.735.827.635.527.531.231.2
MLLM-K [75]62.3-------
TEVA-3B61.643.132.332.633.041.833.529.9
TEVA-7B77.160.738.236.935.249.434.735.5
", + "image_path": "1dfe1c0a1964722ddf371451fdbddfde63d70b05c0f1fbee9c569fa98ec1d5b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 364, + 553, + 506 + ], + "blocks": [ + { + "bbox": [ + 233, + 232, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 233, + 232, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 233, + 232, + 555, + 353 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art open-source and commercial methods on the high-resolution benchmark " + }, + { + "bbox": [ + 233, + 232, + 555, + 353 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 233, + 232, + 555, + 353 + ], + "type": "text", + "content": " [61] (averaging attribute recognition(AR) and spatial relationship reasoning(SRR)), as well as on MME-RealWorld [79], which includes perception benchmark MME-RW" + }, + { + "bbox": [ + 233, + 232, + 555, + 353 + ], + "type": "inline_equation", + "content": "_{PRC}" + }, + { + "bbox": [ + 233, + 232, + 555, + 353 + ], + "type": "text", + "content": " and reasoning benchmark MME-RW" + }, + { + "bbox": [ + 233, + 232, + 555, + 353 + ], + "type": "inline_equation", + "content": "_{RSN}" + }, + { + "bbox": [ + 233, + 232, + 555, + 353 + ], + "type": "text", + "content": ". Task categories are abbreviated due to space limits: OCR (Optical Character Recognition), RS (Remote Sensing), MO (Video Monitoring), AD (Autonomous Driving). Our model consistently outperforms other open-source methods on Reasoning tasks such as OCR, RS, MO, AR and SRR. It even surpasses leading commercial VLMs, GPT-4o and Gemini-1.5-pro, on RS, MO, and AD perception tasks that require precise visual grounding. The best results are bold. Notably, all samples are evaluated at their original image resolution for TEVA and at the largest default resolution for other methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 364, + 553, + 506 + ], + "lines": [ + { + "bbox": [ + 56, + 364, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 553, + 506 + ], + "type": "table", + "html": "
MethodLLM#Visual TokenMMBSQAVQATPOPEAI2DLLaVAwMMMUVVQAV2
InstructBLIP [7]Vicuna-7B-36.063.150.7--60.9--
IDEFICS-9B [18]LLAMA-7B-48.2------50.9
LLaVA-1.5 [35]Vicuna-7B57664.366.866.885.9-63.435.778.5
MLLM-K [75]Vicuna-7B---56.187.3---76.5
LLaVA-1.5-S2[52]Vicuna-7B57666.2-----37.780.0
VILA-1.5 [32]Llama-2-7B72968.968.268.285.5-69.738.679.9
Mini-Gemini-HD [30]Vicuna-7B288072.775.168.4-73.5-37.3-
Monkey [31]Qwen-7B5120-69.467.6-62.6--80.3
LLaVA-Next [36]LLama3-8B288072.172.864.686.771.683.241.781.8
Cambrian-1 [57]Llama3-8B-Instruct230475.980.471.7-73.0-42.781.2
TEVA-3BSheared-llama-2.7B72964.275.766.887.064.985.032.480.3
TEVA-7BQwen2-7B72980.392.572.587.976.394.348.482.8
", + "image_path": "8c7437fd5645ffe4e60d57f9c65fd5e7ca6b6d214201a1e540dfd0bc2c31f982.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 555, + 555 + ], + "lines": [ + { + "bbox": [ + 55, + 510, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 555, + 555 + ], + "type": "text", + "content": "Table 3. Comparison with leading methods on 8 commonly used VLM benchmarks: MMBench [41], SQA: ScienceQA [43], VQATextVQA [54], POPE [29], AI2D [22], LLaVAw: LLaVA-Bench (In-the-Wild) [37], MMMUv: MMMU-validation [69], VQAV2: VQAV2 [14]. Our model consistently outperforms the other methods without image tiling and increasing visual tokens. The best results are bold. Please note that all the samples are evaluated on original image resolution for our model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 571, + 294, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 294, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 294, + 608 + ], + "type": "text", + "content": "ing (e.g., identifying traffic signs), TEVA demonstrates robust performance and even surpassing the most powerful commercial models, GPT-4o [50] and Gemini-1.5-pro [56]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 612, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 295, + 696 + ], + "type": "text", + "content": "General resolution benchmark results. We also evaluate our model on eight commonly used VQA and MLLM benchmarks, comparing it with state-of-the-art VLMs, as shown in Table 3. The results demonstrate that TEVA consistently outperforms other leading methods across all the benchmarks and achieves this without requiring image tiling or visual token extending." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "Notably, on general resolution subject-oriented QA tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 555, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 619 + ], + "type": "text", + "content": "like MMBench [41], AI2D [22] and " + }, + { + "bbox": [ + 313, + 570, + 555, + 619 + ], + "type": "inline_equation", + "content": "\\mathrm{MMMU}^{\\mathrm{V}}" + }, + { + "bbox": [ + 313, + 570, + 555, + 619 + ], + "type": "text", + "content": " [69], TEVA achieves superior results compared to all other methods. This highlights the generalizability of our approach in focusing on contextually important areas within an image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Qualitative Results. In Figure 5, we present qualitative results that demonstrate the effectiveness of TEVA-3B in challenging high-resolution scenarios requiring precise detail recognition. By leveraging our informative patch sampling strategy, TEVA dynamically focuses on relevant, information-rich regions based on the user query, enabling inference on images at arbitrary resolutions without increas" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24153" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 85, + 291, + 134 + ], + "blocks": [ + { + "bbox": [ + 74, + 74, + 276, + 83 + ], + "lines": [ + { + "bbox": [ + 74, + 74, + 276, + 83 + ], + "spans": [ + { + "bbox": [ + 74, + 74, + 276, + 83 + ], + "type": "text", + "content": "Results with Different Dynamic and Static Token Numbers" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 85, + 291, + 134 + ], + "lines": [ + { + "bbox": [ + 58, + 85, + 291, + 134 + ], + "spans": [ + { + "bbox": [ + 58, + 85, + 291, + 134 + ], + "type": "table", + "html": "
DynamicStaticV*MRWPOPEVQAT
025645.030.083.044.5
808150.431.585.654.2
47325661.634.087.066.8
", + "image_path": "1dcfc085b1a86ab4f5056d8b296b5b3ad36c4b3c3299ad1ce664d2e58bf46e31.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 148, + 291, + 188 + ], + "blocks": [ + { + "bbox": [ + 97, + 137, + 252, + 146 + ], + "lines": [ + { + "bbox": [ + 97, + 137, + 252, + 146 + ], + "spans": [ + { + "bbox": [ + 97, + 137, + 252, + 146 + ], + "type": "text", + "content": "Comparison of w/o Subject-Orient (SO) Data" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 148, + 291, + 188 + ], + "lines": [ + { + "bbox": [ + 61, + 148, + 291, + 188 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 291, + 188 + ], + "type": "table", + "html": "
SO DataMRWPRCMRWRSNPOPEVQAT
X33.128.586.558.7
✓33.834.387.066.8
", + "image_path": "3f892eecd0a7f7fb5445460ed1d8a788c0898db5a9a924714cd871bd3bbd78c3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 61, + 201, + 291, + 242 + ], + "blocks": [ + { + "bbox": [ + 97, + 190, + 252, + 199 + ], + "lines": [ + { + "bbox": [ + 97, + 190, + 252, + 199 + ], + "spans": [ + { + "bbox": [ + 97, + 190, + 252, + 199 + ], + "type": "text", + "content": "Comparison of w/o Matrix Embedding (ME)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 201, + 291, + 242 + ], + "lines": [ + { + "bbox": [ + 61, + 201, + 291, + 242 + ], + "spans": [ + { + "bbox": [ + 61, + 201, + 291, + 242 + ], + "type": "table", + "html": "
MEV*ARV*SRRMRWPOPE
X48.350.031.884.8
✓60.063.234.087.0
", + "image_path": "bb5339fb463b57a54e52695d492689d04f9e268f79bbb892c3ce8ab9ee6327c8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 247, + 295, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 281 + ], + "type": "text", + "content": "Table 4. Ablation Studies on Token Numbers, w/o Subject-Orient Pretraining and w/o Affine Transformation Matrix Embedding. Benchmark abbreviation MRW: MME-RealWorld [79]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 293, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 295, + 473 + ], + "type": "text", + "content": "ing computational cost. For example, processing a 5760 " + }, + { + "bbox": [ + 54, + 293, + 295, + 473 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 54, + 293, + 295, + 473 + ], + "type": "text", + "content": " 2880 (5K-resolution) image takes only 2.5 seconds on a single NVIDIA A100 GPU with TEVA-7B. This efficiency enables TEVA to handle high-resolution images seamlessly, outperforming leading open-source VLMs, such as VILA1.5-7B [32], which uses the same number of visual tokens (729), and Monkey-7B [31], which relies on image tiling for high-resolution support. Furthermore, our approach surpasses the advanced commercial VLM, ChatGPT-4o [50], in precise detail observation for high-resolution images, despite the latter's use of a more powerful language model. The evaluation covers real-world, high-difficulty OCR tasks along with diverse general queries, such as counting, identification, and attribute recognition, providing a comprehensive assessment of TEVA's capabilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 480, + 149, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 149, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 149, + 493 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "type": "text", + "content": "Dynamic vs. Static Token Numbers. We assess the impact of using dynamic, informative image patches compared to static, uniformly sampled patches. As shown in Table 4, adding 473 dynamic patches to the baseline of 256 static patches (729 tokens in total) leads to significant improvements in tasks requiring detailed structural understanding, such as scene OCR tasks in " + }, + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\mathrm{VQA}^{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "type": "text", + "content": " (TextVQA) [54], as well as in handling high-resolution images, as demonstrated on the " + }, + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "type": "text", + "content": " [61] benchmark. Additionally, we conducted experiments using a reduced set of 144 patches (80 dynamic and 81 static). Despite the lower patch count, incorporating dynamic tokens consistently enhanced performance on challenging datasets that demand high-resolution or precise observation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Subject-Oriented Pretraining. To evaluate the benefits of pretraining on our Subject-Oriented (SO) Dataset, we compare results with and without the Subject-Oriented Pretraining phase. As illustrated in Table 4, pretrain" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "ing on a large-scale Subject-Oriented Dataset notably improves the model's complex reasoning abilities (MMERealWorld" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "_{RSN}" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": ": MME-RealWorld-Reasoning [79]) and scene understanding, especially in tasks requiring OCR, such as " + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "\\mathrm{VQAT}" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": " (TextVQA) [54]. However, for tasks requiring more straightforward scene comprehension, such as basic perception (MME-RealWorld" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "inline_equation", + "content": "_{PRC}" + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": ": MME-RealWorld-Perception [79]) or hallucination assessments (POPE [29]), the improvements from Subject-Oriented Pretraining are less pronounced. These findings demonstrate the effectiveness of Subject-Oriented Pretraining in enhancing the model's capacity for complex visual understanding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "content": "Affine Transformation Matrix Embeddings. We examine the effectiveness of embedding the affine transformation matrix to provide positional information for each patch. Results in Table 4 demonstrate that the Affine Transformation Matrix Embeddings (ME) consistently improve performance across various benchmarks, particularly enhancing the model's ability to comprehend the relative positions of objects in the " + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "inline_equation", + "content": "V_{SRR}^{*}" + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "inline_equation", + "content": "V^{*}" + }, + { + "bbox": [ + 313, + 216, + 555, + 348 + ], + "type": "text", + "content": " Spatial Relationship reasoning) benchmark. This aligns well with our intuition, as the additional spatial context helps the model maintain awareness of object locations and relationships." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 355, + 388, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 388, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 388, + 366 + ], + "type": "text", + "content": "4.4. Limitation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 372, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 492 + ], + "type": "text", + "content": "While TEVA can process images at arbitrary resolutions without increasing the token count, its effectiveness may be limited in scenarios involving densely packed high-resolution images, such as text-heavy documents. In these cases, where information is uniformly distributed across the image, there are no distinct regions to prioritize. Consequently, dynamic sampling must cover the entire image at a lower resolution, limiting the advantages of high-resolution focus on key areas. Additional experiments and discussion are provided in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 522, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 666 + ], + "type": "text", + "content": "In this work, we introduced Token-Efficient Vision LLanguage Model (TEVA), a straightforward and token-efficient framework that enhances Vision-Language Models (VLMs) for high-resolution image processing. TEVA employs a subject-oriented strategy and Dynamic Patch Sampling to balance fine details with global context, supported by a specialized Subject-Oriented Dataset for improved learning. Comprehensive evaluations against state-of-the-art VLMs, including GPT-4o, show that TEVA excels in challenging benchmarks and holds strong potential for applications requiring high-resolution precision, such as remote sensing, autonomous driving, and video monitoring." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24154" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 235 + ], + "type": "text", + "content": "[3] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.4.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 294, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 280 + ], + "type": "text", + "content": "[4] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 281, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 294, + 335 + ], + "type": "text", + "content": "[5] Xi Chen, Josip Djolonga, Piotr Padlewski, Basil Mustafa, Soravit Changpinyo, Jialin Wu, Carlos Riquelme Ruiz, Sebastian Goodman, Xiao Wang, Yi Tay, et al. Pali-x: On scaling up a multilingual vision and language model. arXiv preprint arXiv:2305.18565, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 294, + 390 + ], + "type": "text", + "content": "[6] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision-language models. Advances in Neural Information Processing Systems, 37:135062-135093, 2025. 2, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 294, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 445 + ], + "type": "text", + "content": "[7] Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructlip: Towards general-purpose vision-language models with instruction tuning, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 501 + ], + "type": "text", + "content": "[8] Wenliang Dai, Nayeon Lee, Boxin Wang, Zhuoling Yang, Zihan Liu, Jon Barker, Tuomas Rintamaki, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. Nvlm: Open frontier-class multimodal llms. arXiv preprint arXiv:2409.11402, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 568 + ], + "type": "text", + "content": "[9] Xiaoyi Dong, Pan Zhang, Yuhang Zang, Yuhang Cao, Bin Wang, Linke Ouyang, Songyang Zhang, Haodong Duan, Wenwei Zhang, Yining Li, et al. Internlm-xcomposer2-4khd: A pioneering large vision-language model handling resolutions from 336 pixels to 4k hd. arXiv preprint arXiv:2404.06512, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 569, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 294, + 635 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 679 + ], + "type": "text", + "content": "[11] Yunhao Fang, Ligeng Zhu, Yao Lu, Yan Wang, Pavlo Molchanov, Jang Hyun Cho, Marco Pavone, Song Han, and Hongxu Yin. vila2: Vila augmented vila. arXiv preprint arXiv:2407.17453, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[12] Chaoyou Fu, Peixian Chen, Yunhang Shen, Yulei Qin, Mengdan Zhang, Xu Lin, Zhenyu Qiu, Wei Lin, Jinrui Yang, Xiawu Zheng, Ke Li, Xing Sun, and Rongrong Ji. Mme: A" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "comprehensive evaluation benchmark for multimodal large language models. ArXiv, abs/2306.13394, 2023. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 96, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 148 + ], + "type": "text", + "content": "[13] Jiahui Gao, Renjie Pi, Jipeng Zhang, Jiacheng Ye, Wanjun Zhong, Yufei Wang, Lanqing Hong, Jianhua Han, Hang Xu, Zhenguo Li, et al. G-llava: Solving geometric problem with multi-modal large language model. arXiv preprint arXiv:2312.11370, 2023. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 150, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 150, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 150, + 553, + 215 + ], + "type": "text", + "content": "[14] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 5, 7, 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "text", + "content": "[15] Qiushan Guo, Shalini De Mello, Hongxu Yin, Wonmin Byeon, Ka Chun Cheung, Yizhou Yu, Ping Luo, and Sifei Liu. Regionpt: Towards region understanding vision language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13796-13806, 2024. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 283, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 338 + ], + "type": "text", + "content": "[16] Jack Hessel, Jena D Hwang, Jae Sung Park, Rowan Zellers, Chandra Bhagavatula, Anna Rohrbach, Kate Saenko, and Yejin Choi. The abduction of sherlock holmes: A dataset for visual abductive reasoning. In European Conference on Computer Vision, pages 558-575. Springer, 2022. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "text", + "content": "[17] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 1, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 393, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 415 + ], + "type": "text", + "content": "[18] IDEFICS. Introducing idefics: An open reproduction of state-of-the-art visual language model. 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 416, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 553, + 458 + ], + "type": "text", + "content": "[19] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Proceedings of Advances in Neural Information Processing Systems (NeurIPS), 28, 2015. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "type": "text", + "content": "[20] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901-2910, 2017. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 580 + ], + "type": "text", + "content": "[21] Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5648-5656, 2018. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 581, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 581, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 581, + 553, + 647 + ], + "type": "text", + "content": "[22] Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 235-251. Springer, 2016. 5, 7, 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "text", + "content": "[23] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeong Yeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24155" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[24] Jonathan Krause, Justin Johnson, Ranjay Krishna, and Li Fei-Fei. A hierarchical approach for generating descriptive image paragraphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 317–325, 2017. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 204 + ], + "type": "text", + "content": "[25] Paul Lerner, Olivier Ferret, Camille Guinaudeau, Hervé Le Borgne, Romaric Besançon, José G Moreno, and Jesús Lovón Melgarejo. Viquae, a dataset for knowledge-based visual question answering about named entities. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 3108-3120, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 249 + ], + "type": "text", + "content": "[26] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. arXiv preprint arXiv:2307.16125, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[27] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 336 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[29] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv preprint arXiv:2305.10355, 2023. 5, 7, 8, 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 435 + ], + "type": "text", + "content": "[30] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "type": "text", + "content": "[31] Zhang Li, Biao Yang, Qiang Liu, Zhiyin Ma, Shuo Zhang, Jingxu Yang, Yabo Sun, Yuliang Liu, and Xiang Bai. Monkey: Image resolution and text label are important things for large multi-modal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26763-26773, 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "text", + "content": "[32] Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26689-26699, 2024. 6, 7, 8, 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "text", + "content": "[33] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 668 + ], + "type": "text", + "content": "[34] Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. Mmc: Advancing multimodal chart understanding with large-scale instruction tuning. arXiv preprint arXiv:2311.10774, 2023. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[35] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[36] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.1,7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 108, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 129 + ], + "type": "text", + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NIPS, 2024. 1, 5, 7, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 130, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 185 + ], + "type": "text", + "content": "[38] Haogeng Liu, Quanzeng You, Xiaotian Han, Yiqi Wang, Bohan Zhai, Yongfei Liu, Yunzhe Tao, Huaibo Huang, Ran He, and Hongxia Yang. Infimm-hd: A leap forward in high-resolution multimodal understanding. arXiv preprint arXiv:2403.01487, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 186, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 241 + ], + "type": "text", + "content": "[39] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 243, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 286 + ], + "type": "text", + "content": "[40] Yuliang Liu, Biao Yang, Qiang Liu, Zhang Li, Zhiyin Ma, Shuo Zhang, and Xiang Bai. Textmonkey: AnOCR-free large multimodal model for understanding document. arXiv preprint arXiv:2403.04473, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "type": "text", + "content": "[41] Yuan Liu, Haodong Duan, Yuanhan Zhang, Bo Li, Songyang Zhang, Wangbo Zhao, Yike Yuan, Jiaqi Wang, Conghui He, Ziwei Liu, et al. Mmbench: Is your multi-modal model an all-around player? In European Conference on Computer Vision, pages 216-233. Springer, 2025. 5, 7, 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 344, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 431 + ], + "type": "text", + "content": "[42] Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, Xiuyu Li, Yunhao Fang, Yukang Chen, Cheng-Yu Hsieh, De-An Huang, An-Chieh Cheng, Vishwesh Nath, Jinyi Hu, Sifei Liu, Ranjay Krishna, Daguang Xu, Xiaolong Wang, Pavlo Molchanov, Jan Kautz, Hongxu Yin, Song Han, and Yao Lu. Nvila: Efficient frontier visual language models, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 433, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 498 + ], + "type": "text", + "content": "[43] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems, 35:2507-2521, 2022. 5, 7, 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 499, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 555 + ], + "type": "text", + "content": "[44] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 600 + ], + "type": "text", + "content": "[45] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 601, + 555, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 555, + 644 + ], + "type": "text", + "content": "[46] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "text", + "content": "[47] Minesh Mathew, Viraj Bagal, Ruben Tito, Dimosthenis Karatzas, Ernest Valveny, and CV Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697-1706, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[48] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24156" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "by reading text in images. In 2019 international conference on document analysis and recognition (ICDAR), pages 947-952. IEEE, 2019. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 140 + ], + "type": "text", + "content": "[49] Ravina Mithe, Supriya Indalkar, and Nilam Divekar. Optical character recognition. International journal of recent technology and engineering (IJRTE), 2(1):72-75, 2013. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 163 + ], + "type": "text", + "content": "[50] R OpenAI. Gpt-4 technical report. arxiv 2303.08774. View in Article, 2(5), 2023. 1, 6, 7, 8, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 165, + 295, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 295, + 231 + ], + "type": "text", + "content": "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 295, + 275 + ], + "type": "text", + "content": "[52] Baifeng Shi, Ziyang Wu, Maolin Mao, Xin Wang, and Trevor Darrell. When do we not need larger vision models? In European Conference on Computer Vision, pages 444-462. Springer, 2025. 1, 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 295, + 331 + ], + "type": "text", + "content": "[53] Oleksii Sidorov, Ronghang Hu, Marcus Rohrbach, and Amanpreet Singh. Textcaps: a dataset for image captioning with reading comprehension. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16, pages 742–758, 2020. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 295, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 295, + 388 + ], + "type": "text", + "content": "[54] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 5, 7, 8, 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 389, + 295, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 295, + 455 + ], + "type": "text", + "content": "[55] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. Wit: Wikipedia-based image text dataset for multimodal multilingual machine learning. In Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval, pages 2443-2449, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 295, + 510 + ], + "type": "text", + "content": "[56] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 578 + ], + "type": "text", + "content": "[57] Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. Advances in Neural Information Processing Systems, 37:87310-87356, 2024. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 634 + ], + "type": "text", + "content": "[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 690 + ], + "type": "text", + "content": "[59] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1, 3, 5, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[60] Jason Wei, Maarten Bosma, Vincent Y Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M Dai, and" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Quoc V Le. Finetuned language models are zero-shot learners. arXiv preprint arXiv:2109.01652, 2021. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "text", + "content": "[61] Penghao Wu and Saining Xie. V*: Guided visual search as a core mechanism in multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13084-13094, 2024. 2, 4, 5, 6, 7, 8, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "text", + "content": "[62] Mengzhou Xia, Tianyu Gao, Zhiyuan Zeng, and Danqi Chen. Sheared llama: Accelerating language model pre-training via structured pruning. arXiv preprint arXiv:2310.06694, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 228 + ], + "type": "text", + "content": "[63] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794-4803, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "text", + "content": "[64] Ruyi Xu, Yuan Yao, Zonghao Guo, Junbo Cui, Zanlin Ni, Chunjiang Ge, Tat-Seng Chua, Zhiyuan Liu, Maosong Sun, and Gao Huang. Llava-uhd: an Imm perceiving any aspect ratio and high-resolution images. arXiv preprint arXiv:2403.11703, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[65] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[66] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "text", + "content": "[67] Xiaoyu Yue, Shuyang Sun, Zhanghui Kuang, Meng Wei, Philip HS Torr, Wayne Zhang, and Dahua Lin. Vision transformer with progressive sampling. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 387-396, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 481 + ], + "type": "text", + "content": "[68] Xiang Yue, Xingwei Qu, Ge Zhang, Yao Fu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mammoth: Building math generalist models through hybrid instruction tuning. arXiv preprint arXiv:2309.05653, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 482, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 558 + ], + "type": "text", + "content": "[69] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567, 2024. 5, 7, 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 602 + ], + "type": "text", + "content": "[70] Bohan Zhai, Shijia Yang, Chenfeng Xu, Sheng Shen, Kurt Keutzer, and Manling Li. Halle-switch: Controlling object hallucination in large vision language models. arXiv e-prints, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "type": "text", + "content": "[71] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 691 + ], + "type": "text", + "content": "[72] Yufei Zhan, Hongyin Zhao, Yousong Zhu, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon-g: Bridging vision-language and vision-centric tasks via large multimodal models. arXiv preprint arXiv:2410.16163, 2024. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[73] Yufei Zhan, Yousong Zhu, Hongyin Zhao, Fan Yang, Ming Tang, and Jinqiao Wang. Griffon v2: Advancing multimodal" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24157" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 496 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "text", + "content": "perception with high-resolution scaling and visual-language co-referring. arXiv preprint arXiv:2403.09333, 2024. 1, 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 160 + ], + "type": "text", + "content": "[74] Haotian Zhang, Haoxuan You, Philipp Duffer, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[75] Jiarui Zhang, Mahyar Khayatkhoei, Prateek Chhikara, and Filip Ilievski. Mllms know where to look: Training-free perception of small visual details with multimodal llms. Proceedings of International Conference on Learning Representations (ICLR), 2025. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 283 + ], + "type": "text", + "content": "[76] Pan Zhang, Xiaoyi Dong, Bin Wang, Yuhang Cao, Chao Xu, Linke Ouyang, Zhiyuan Zhao, Haodong Duan, Songyang Zhang, Shuangrui Ding, et al. Internlmxcomposer: A vision-language large model for advanced text-image comprehension and composition. arXiv preprint arXiv:2309.15112, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 339 + ], + "type": "text", + "content": "[77] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 341, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 294, + 384 + ], + "type": "text", + "content": "[78] Yanzhe Zhang, Ruiyi Zhang, Jiuming Gu, Yufan Zhou, Nedim Lipka, Diyi Yang, and Tong Sun. Llavar: Enhanced visual instruction tuning for text-rich image understanding. arXiv preprint arXiv:2306.17107, 2023. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 451 + ], + "type": "text", + "content": "[79] Yi-Fan Zhang, Huanyu Zhang, Haochen Tian, Chaoyou Fu, Shuangqing Zhang, Junfei Wu, Feng Li, Kun Wang, Qing-song Wen, Zhang Zhang, et al. Mme-realworld: Could your multimodal llm challenge high-resolution real-world scenarios that are difficult for humans? arXiv preprint arXiv:2408.13257, 2024. 5, 6, 7, 8, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 453, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 496 + ], + "type": "text", + "content": "[80] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24158" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/0bebb7e8-7094-4aab-a136-ac9635434593_content_list.json b/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/0bebb7e8-7094-4aab-a136-ac9635434593_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..147647654033d36b03c312acb686cb15cc8511b4 --- /dev/null +++ b/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/0bebb7e8-7094-4aab-a136-ac9635434593_content_list.json @@ -0,0 +1,1560 @@ +[ + { + "type": "text", + "text": "TokenUnify: Scaling Up Autoregressive Pretraining for Neuron Segmentation", + "text_level": 1, + "bbox": [ + 109, + 130, + 888, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yinda Chen $^{1,2*}$ Haoyuan Shi $^{1,2*}$ Xiaoyu Liu $^{1}$ Te Shi $^{2}$ Ruobing Zhang $^{3,2}$ Dong Liu $^{1}$ Zhiwei Xiong $^{1,2\\dagger}$ Feng Wu $^{1,2‡}$", + "bbox": [ + 251, + 179, + 750, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China \n $^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center \n $^{3}$ Institute for Brain and Intelligence, Fudan University \n{cyd0806, haoyuan.shi}@mail.ustc.edu.cn, {zwxiong, fengwu}@ustc.edu.cn", + "bbox": [ + 174, + 232, + 823, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 356, + 326, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neuron segmentation from electron microscopy (EM) volumes is crucial for understanding brain circuits, yet the complex neuronal structures in high-resolution EM images present significant challenges. EM data exhibits unique characteristics including high noise levels, anisotropic voxel dimensions, and ultra-long spatial dependencies that make traditional vision models inadequate. Inspired by autoregressive pretraining in language models, we propose TokenUnify, a hierarchical predictive coding framework that captures multi-scale dependencies through three complementary learning objectives. TokenUnify integrates random token prediction, next-token prediction, and next-all token prediction to create a comprehensive representational space with emergent properties. From an information-theoretic perspective, these three tasks are complementary and provide optimal coverage of visual data structure, with our approach reducing autoregressive error accumulation from $O(K)$ to $O(\\sqrt{K})$ for sequences of length $K$ . We also introduce a large-scale EM dataset with 1.2 billion annotated voxels, offering ideal long-sequence visual data with spatial continuity. Leveraging the Mamba architecture's linear-time sequence modeling capabilities, TokenUnify achieves a $44\\%$ performance improvement on downstream neuron segmentation and outperforms MAE by $25\\%$ . Our approach demonstrates superior scaling properties as model size increases, effectively bridging the gap between pretraining strategies for language and vision models. Code is available at https://github.com/ydchen0806/TokenUnify.", + "bbox": [ + 89, + 388, + 485, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 356, + 643, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding neural circuits is a central goal in neuroscience. Electron microscopy (EM) imaging enables reconstruction of neurons and synaptic connections at high resolution [31, 33]. However, EM neuron segmentation presents significant challenges due to three key properties: (1) high noise levels from electron beam interactions, (2) anisotropic voxel dimensions with coarser z-axis resolution, and (3) long-range spatial continuity spanning thousands of patches. Current methods struggle to capture global context and long-range dependencies [8, 21, 30, 36]. Despite success in natural language processing [28], similar advancements for long-sequence visual data remain largely unexplored.", + "bbox": [ + 511, + 381, + 906, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large language models (LLMs) show remarkable scaling capabilities [1, 39, 40] through effective autoregressive modeling. However, visual data has more complex structures than text. Contrastive methods like DINO v2 [34] and masked reconstruction like MAE [8, 20] demonstrate strong representation learning but lack favorable scaling laws [37]. Masked prediction faces theoretical limitations where estimation error scales as $O(\\sqrt{s \\log p / n})$ for sparsity $s$ , dimensionality $p$ , and sample size $n$ . Autoregressive approaches like AIM [14] and LVM [4] attempt to bridge this gap, but error accumulation $O(K)$ with sequence length $K$ and computational complexity hinder their application to high-resolution EM images [3].", + "bbox": [ + 511, + 578, + 908, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper tackles visual autoregressive learning challenges through hierarchical predictive coding. Let $X = \\{x_{1}, x_{2}, \\ldots, x_{K}\\}$ represent visual tokens from an EM image, where $\\mathcal{M} \\subset \\{1, \\ldots, K\\}$ denotes masked indices and $\\mathcal{M}^{c}$ denotes unmasked tokens. We recognize that three prediction tasks capture complementary visual structure aspects: random token prediction learns noise-robust spatial patterns, next-token prediction captures sequential de", + "bbox": [ + 511, + 773, + 908, + 896 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution.", + "bbox": [ + 107, + 849, + 240, + 861 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Project Leader", + "bbox": [ + 109, + 861, + 196, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding Author", + "bbox": [ + 109, + 875, + 236, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13604", + "bbox": [ + 478, + 941, + 519, + 954 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f011e02d13fd4c507b30d39ab6ab8bfed21cdff1d57f560c39f9caf037286de.jpg", + "image_caption": [ + "Figure 1. TokenUnify divides the 3D EM image into non-overlapping patches, which are tokenized into a sequence of K tokens. Three complementary tasks are performed for rich 3D image representations: (a) random token prediction for spatial pattern learning, (b) next token prediction for sequential dependencies, and (c) next-all token prediction for global structure modeling." + ], + "image_footnote": [], + "bbox": [ + 133, + 85, + 869, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pencies, and next-all token prediction models global context. Under regularity conditions $(\\mathrm{AR}(\\infty)$ process with summable coefficients), autoregressive models achieve asymptotically optimal prediction error $O(\\sigma^2)$ as capacity increases, where $\\sigma^2$ is irreducible noise.", + "bbox": [ + 88, + 309, + 480, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose TokenUnify, a unified framework that integrates these three predictive tasks within a coherent information-theoretic formulation (Fig. 1). This creates a comprehensive representational space where random token prediction provides local robustness, next-token prediction ensures structural continuity, and next-all token prediction captures global morphology, enabling simultaneous encoding of fine-grained details and global relationships unachievable with single objectives. Our approach reduces cumulative errors from $O(K)$ to $O(\\sqrt{K})$ by distributing prediction errors across multiple positions rather than sequential accumulation, analogous to central limit theorem scaling. To handle massive high-resolution EM data efficiently, we leverage the Mamba architecture, which reduces computational complexity from quadratic to linear while maintaining superior performance compared to Transformers. We validate our approach on a large-scale EM dataset collected from mouse brain slices, featuring thousands of continuous tokens that ensure spatial continuity crucial for understanding complex neuronal structures. Through comprehensive annotation of six functional brain regions totaling 1.2 billion voxels, we create one of the largest manually annotated EM neuron segmentation datasets to date.", + "bbox": [ + 93, + 385, + 482, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Comprehensive evaluation demonstrates that TokenUnify achieves a $44\\%$ improvement on downstream neuron segmentation and outperforms MAE [20] by $25\\%$ in pretraining performance. Furthermore, TokenUnify demonstrates superior scaling properties, offering a promising paradigm for pretraining large-scale visual models.", + "bbox": [ + 89, + 732, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 823, + 434, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We introduce a novel hierarchical predictive coding framework that unifies three distinct visual structure perspectives—random token prediction for spatial robustness, next-token prediction for sequential continuity, and", + "bbox": [ + 91, + 837, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "next-all token prediction for global context—within a coherent information-theoretic formulation. Our theoretical analysis (detailed in supplementary material) demonstrates why these prediction tasks are complementary and together provide optimal coverage of visual data structure, while reducing autoregressive error accumulation from $O(K)$ to $O(\\sqrt{K})$ .", + "bbox": [ + 531, + 309, + 903, + 414 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. We achieve a $44\\%$ performance improvement on downstream neuron segmentation tasks and realize the first billion-parameter Mamba network for visual autoregression, demonstrating both effectiveness and computational efficiency in processing long-sequence visual data. Our approach outperforms existing methods including MAE by $25\\%$ while exhibiting favorable scaling properties as model size increases.", + "3. We construct one of the largest EM neuron segmentation datasets with 1.2 billion finely annotated voxels across six functional brain regions, providing an ideal testbed for long-sequence visual modeling. This comprehensive dataset offers spatially continuous visual data with thousands of tokens per volume, enabling rigorous validation of autoregressive methods for biological image analysis." + ], + "bbox": [ + 511, + 415, + 906, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 654, + 653, + 669 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neuron Segmentation. In the field of EM image analysis, neuron instance segmentation has emerged as a vital task. Turaga et al. [41] pioneered the use of convolutional neural networks (CNNs) for affinity generation, followed by clustering of voxels in the affinity graph using post-processing techniques such as waterz [15] and LMC [5]. Since then, affinity-based methods have seen significant advancements. Funke et al. [15] introduced the MALIS loss [6] during training to ensure topologically accurate segmentation. Recent work by Huang et al. [21] proposed an embedding pyramid module to capture affinity information across different scales. Liu et al. [29] combined both affinity and embedding features with graph neural networks (GNNs) to enhance the segmentation of adjacent objects in the feature", + "bbox": [ + 511, + 680, + 906, + 891 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13605", + "bbox": [ + 480, + 941, + 517, + 953 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "space. Furthermore, Liu [30] introduced a knowledge distillation framework, transferring knowledge from 3D models to 2D models, which resulted in improved performance and efficiency in neuron segmentation. Despite these advances, mamba-based architectures [17], have yet to be fully explored for the anisotropic resolution characteristics of EM imaging. In this paper, we enhance the Segmamba decoder [43] to better process the anisotropic features of EM data using an affinity-based approach.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-supervised Pretraining. Self-supervised pretraining has been a cornerstone for enhancing model representation capabilities. Approaches based on contrastive learning for representation extraction [9, 19, 24, 44] and masked reconstruction methods [8, 10, 11, 13, 20, 27] have shown promise. However, current vision models have not exhibited the same sublinear scaling laws as language models. To address this issue, some tasks have adopted autoregressive pre-training paradigms similar to those used in language models [4, 7, 14], though the training costs remain a significant concern. In this paper, we explore the potential of long visual token autoregressive pre-training and introduce a collaborative training scheme for long token prediction. Our approach aims to balance the scaling laws and training costs, demonstrating improvements in fine-grained visual tasks.", + "bbox": [ + 91, + 227, + 483, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Theoretical Motivations", + "text_level": 1, + "bbox": [ + 89, + 482, + 316, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section provides the theoretical foundations that motivate our approach. Detailed derivations and proofs are provided in the supplementary material.", + "bbox": [ + 89, + 508, + 482, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notation. Let $\\mathbf{X} = \\{x_{1}, x_{2}, \\ldots, x_{K}\\}$ represent a sequence of visual tokens extracted from an EM image, where $\\mathcal{M} \\subset \\{1, \\ldots, K\\}$ denotes masked token indices, $\\mathcal{M}^c$ denotes unmasked tokens, and $x_{ 0$ is a small constant, and $c > 0$ is the magnitude of negative curvature. This geometric property enables efficient encoding of both fine-grained details (in near-flat regions) and complex hierarchical structures (in hyperbolic-like regions).", + "bbox": [ + 511, + 402, + 903, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 511, + 505, + 604, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 530, + 624, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The workflow of our method is shown in Fig. 2. Our approach comprises two stages: (1) hierarchical predictive pre-training, where we learn a generic visual representation $f_{\\theta_1}(\\cdot)$ from unlabeled data $\\mathbf{X}$ by integrating three complementary prediction tasks that capture both local and global dependencies; and (2) task-specific fine-tuning, where we adapt the pre-trained model to downstream tasks using both $\\mathbf{X}$ and labels $Y$ through a task-specific model $g_{\\theta_2}(\\cdot)$ initialized with $\\theta_1$ . For EM neuron segmentation, we process input volumes $\\mathbf{X} = \\{X^{(1)},\\dots,X^{(T)}\\}$ , where each $X^{(t)} \\in \\mathbb{R}^{D\\times H\\times W}$ represents a 3D image, by partitioning them into smaller patches $x \\in \\mathbb{R}^{D' \\times H' \\times W'}$ and leveraging Mamba's efficient sequence modeling capabilities to handle the high-dimensional, long-sequence nature of the data.", + "bbox": [ + 511, + 551, + 906, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Hierarchical Predictive Coding in Token Space", + "text_level": 1, + "bbox": [ + 511, + 772, + 903, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on our theoretical insights, we develop a hierarchical predictive coding framework that operates on the manifold of token embeddings to effectively capture the multiscale structure of high-dimensional neural image data. Our framework is motivated by the principle of maximum information preservation across complementary representational dimensions.", + "bbox": [ + 511, + 794, + 903, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13606", + "bbox": [ + 480, + 941, + 519, + 953 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8b709dcf7c3700340886284460baf6e5edc34fa59e74e831e19d0cf1dd86d173.jpg", + "image_caption": [ + "Figure 2. Illustration of the TokenUnify pretraining workflow. The image $\\mathbf{X}$ is fed into the Tokenizer, transforming it into a long sequence of tokens $x_{i}|_{i = 1}^{K}$ . The three prediction tasks (random token, next token, and next-all token) are performed with complementary objectives. The Perceiver Resampler is employed to convert varying-size large feature maps into a few visual tokens." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 908, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notation. Let $\\mathbf{x} = \\{x_{1}, x_{2}, \\ldots, x_{K}\\}$ represent a sequence of visual tokens extracted from a volumetric EM image, where each $x_{i} \\in \\mathbb{R}^{d}$ . For random token prediction, $\\mathcal{M} \\subset \\{1, \\ldots, K\\}$ denotes the set of masked token indices with $|\\mathcal{M}| = \\rho K$ for masking ratio $\\rho$ , and $\\mathcal{M}^{c} = \\{1, \\ldots, K\\} \\setminus \\mathcal{M}$ denotes unmasked tokens. For sequential prediction, $x_{ 0$ is the annealing temperature controlling transition sharpness.", + "bbox": [ + 89, + 205, + 482, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our unified objective function is formulated as:", + "bbox": [ + 109, + 253, + 423, + 266 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {T o k e n U n i f y}} = \\alpha (t) \\cdot \\mathcal {L} _ {\\text {r a n d o m}} + \\beta (t) \\cdot \\mathcal {L} _ {\\text {n e x t}} + \\gamma (t) \\cdot \\mathcal {L} _ {\\text {n e x t - a l l}}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 280, + 482, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The curriculum implements progressive transition from local to global representation learning through:", + "bbox": [ + 89, + 308, + 482, + 339 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} (t) = \\left\\{ \\begin{array}{l l} [ 2. 0, - 1. 0, - 2. 0 ] & \\text {i f} t < T _ {1} (\\text {r a n d o m - d o m i n a n t}) \\\\ [ - 1. 0, 2. 0, - 1. 0 ] & \\text {i f} T _ {1} \\leq t < T _ {2} (\\text {n e x t - d o m i n a n t}), \\\\ [ - 2. 0, - 1. 0, 2. 0 ] & \\text {i f} t \\geq T _ {2} (\\text {n e x t - a l l - d o m i n a n t}) \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 349, + 503, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T_{1} = 0.3 \\times T_{\\mathrm{total}}$ and $T_{2} = 0.7 \\times T_{\\mathrm{total}}$ are transition thresholds. With $\\tau(t) = 1.0$ , this yields dominant weights of $\\sim 0.73$ for the primary task while maintaining auxiliary contributions of $\\sim 0.18$ and $\\sim 0.09$ to preserve multi-task synergy.", + "bbox": [ + 89, + 419, + 483, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For computational efficiency in next-all token prediction, we employ cross-attention with latent queries to aggregate sequence-wide information, enabling scalability to long token sequences without prohibitive complexity.", + "bbox": [ + 89, + 494, + 483, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This hierarchical integration constructs a multi-resolution representational manifold capturing neural tissue organization across microscopic, mesoscopic, and macroscopic scales, providing robust foundations for volumetric electron microscopy circuit reconstruction.", + "bbox": [ + 89, + 555, + 483, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Finetuning and Segmentation", + "text_level": 1, + "bbox": [ + 89, + 638, + 354, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The segmentation network, denoted as $g_{\\mathrm{seg}}(\\cdot)$ , consists of an encoder $g_{e}(\\cdot)$ and a decoder $g_{d}(\\cdot)$ :", + "bbox": [ + 89, + 661, + 483, + 693 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {\\mathrm {s e g}} (x; \\theta_ {s}) = g _ {d} \\left(g _ {e} (x)\\right), \\quad \\theta_ {s} = \\left\\{\\theta_ {e}, \\theta_ {d} \\right\\}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 704, + 482, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{s}$ represents the parameters of the entire segmentation network, and $\\theta_{e}$ and $\\theta_{d}$ are the parameters of the encoder and decoder, respectively.", + "bbox": [ + 89, + 732, + 482, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The encoder $g_{e}(\\cdot)$ gradually downsamples the input volume and extracts high-level semantic features, while the decoder $g_{d}(\\cdot)$ upsamples the encoded features back to the original resolution. Meanwhile, the output of each downsampling layer in the encoder is connected to the corresponding layer in the decoder via skip connections to fuse local and global multi-scale information.", + "bbox": [ + 89, + 777, + 483, + 883 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/57a13d443f133743c56e008953f5650b6d532745e4b3e216ed9fd0f08fcf37f4.jpg", + "image_caption": [ + "(a) Sampling Area", + "(b) Annotation", + "Figure 3. Illustration of the MEC dataset annotation. (a) shows the sampling locations of six different neural regions from the MEC raw data. Each region contains $1250 \\times 1250 \\times 125$ voxels. (b) shows an example of the fine segmentation labels in one sampling region, which takes two experts one month to annotate." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt 3D ResUNet/ViT/Mamba as the backbone network, respectively. For the Mamba variant, we adapt SegMamba [43] for EM data, which we refer to as EMmamba, with enhanced decoder design for anisotropic EM features.", + "bbox": [ + 511, + 290, + 903, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The output of the segmentation network $\\hat{y} = g_{\\mathrm{seg}}(x)\\in$ $\\mathbb{R}^{C\\times D\\times H\\times W}$ represents the predicted affinity map [25, 26], corresponding to the connectivity probability of each voxel in three directions. During training, the loss function for labeled samples is the mean squared error between the predicted and ground-truth affinity maps:", + "bbox": [ + 511, + 351, + 905, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s e g}} = \\frac {1}{| \\mathcal {D} _ {l} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {l} |} \\| g _ {\\mathrm {s e g}} \\left(x _ {i} ^ {l}\\right) - y _ {i} \\| ^ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 450, + 903, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, for any new test sample $x_{t}$ , forward propagation through $g_{\\mathrm{seg}}(x_{t})$ yields its predicted affinity map. This predicted affinity map is then post-processed using a seeded watershed algorithm and a structure-aware region agglomeration algorithm [5, 15] to obtain the final neuron instance segmentation. Detailed information on our segmentation process and the segmentation pipeline are illustrated in the supplementary material.", + "bbox": [ + 511, + 500, + 905, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Datasets and Metrics", + "text_level": 1, + "bbox": [ + 513, + 633, + 714, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. For pretraining TokenUnify, we collect over 1 TB of publicly available unlabeled EM imaging data from four large-scale datasets: FAFB [35], MitoEM [42], FIB-25 [38], and Kasthuri15 [22]. These datasets span diverse organisms including Drosophila, mouse, rat, and human samples, ensuring comprehensive coverage of EM morphology variations. We sample from these datasets with equal probability to guarantee diversity in pretraining data. Additional details are provided in supplementary material.", + "bbox": [ + 511, + 657, + 905, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For downstream fine-tuning and evaluation, we employ two datasets: AC3/4 [22] and our large-scale MEC dataset. The AC3/4 dataset consists of 256/100 successive EM images with $1024 \\times 1024$ spatial resolution from mouse somatosensory cortex. We partition AC4 into training (first 80 images) and validation (last 20 images), while using AC3 (first 100 images) for testing.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13608", + "bbox": [ + 480, + 941, + 517, + 953 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1caedde039a0f42b734773c3e47dd7d721dd912acb37e39828aa8a912b758063.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Post.MethodWafer4Wafer36-2Param. (M)
VOIM↓VOIS↓VOI↓ARAND↓VOIM↓VOIS↓VOI↓ARAND↓
LMC [5]Supervised Methods
Superhuman [23]0.33281.12581.45870.17360.15060.45880.60940.08361.478
MALA [15]0.54381.50272.03750.11150.31791.06641.38430.157084.03
PEA [21]0.33810.92761.26580.06770.27870.42790.70660.11691.480
UNETR [18]0.45041.65812.10850.26580.44780.52170.96960.2913129.1
EMmamba0.49151.29241.78390.20520.24060.41890.65950.123128.30
Superhuman [23]0.19481.96972.16440.24530.07921.16181.24100.13191.478
MALA [15]0.34162.41292.75450.25670.14481.96032.10520.197784.03
PEA [21]0.17051.59931.76980.15270.47191.12261.59450.15881.480
UNETR [18]0.17913.17153.35060.63300.09491.38581.48070.1578129.1
EMmamba0.15962.05802.21770.19730.08471.03511.11980.125328.30
LMC [5]Self-Supervised Methods
Random0.49151.29241.78390.20520.24060.41890.65950.123128.30
MAE [20]0.23251.09231.32480.09780.16290.45320.61610.0835
BYOL [16]0.25840.94531.20370.08910.17480.43680.61160.0826
dbMIM [8]0.23420.87961.11380.07420.14760.39820.54580.0782
MS-Con-EM [9]0.24830.91021.15850.08130.15420.40370.55790.0794
TokenUnify0.19530.79980.99510.05090.12620.35850.48480.0650
Random0.15962.05802.21770.19730.08471.03511.11980.125328.30
MAE [20]0.14931.92162.07090.15830.08921.09271.18190.1243
BYOL [16]0.16291.83742.00030.14270.09531.07821.17350.1287
dbMIM [8]0.15731.74361.90090.09130.09121.06491.15610.1185
MS-Con-EM [9]0.16091.76921.93010.10260.09631.07031.16660.1207
TokenUnify0.14181.51031.65210.05910.08271.02761.11030.1074
", + "bbox": [ + 96, + 89, + 903, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison of electron microscopy segmentation results on Wafer4 and Wafer36-2 datasets. Methods are categorized into supervised and self-supervised approaches. All self-supervised methods use the same EMmamba backbone for fair comparison. \"Random\" refers to EMmamba without any pretraining. The best results are in bold and the second best results are underlined.", + "bbox": [ + 89, + 476, + 906, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate algorithm performance on large-scale EM data, we collected a 2TB MEC dataset by imaging mouse somatosensory cortex, medial entorhinal cortex, and cerebral cortex regions. We selected 6 representative volumes from different neural regions (wafer4/25/26/26-2/36/36-2), each with dimensions $1250 \\times 1250 \\times 125$ voxels. Dense manual annotation was performed on these regions, yielding over 1.2 billion annotated voxels through six months of work by two experts, as illustrated in Fig. 3. The dataset is split into training (wafer25/26/26-2/36), validation (wafer4), and testing (wafer36-2) sets.", + "bbox": [ + 88, + 544, + 480, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. To evaluate the performance of neuron segmentation [12, 45], we employ two widely-used metrics: Variation of Information (VOI) [32] and Adjusted Rand Index (ARAND) [2]. These metrics quantify the agreement between the predicted segmentation and the ground truth from different perspectives. Detailed descriptions of these metrics can be found in the supplementary material.", + "bbox": [ + 89, + 714, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiment", + "text_level": 1, + "bbox": [ + 89, + 834, + 215, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our proposed method on both MEC and AC3/4 datasets. Implementation details are provided in the sup", + "bbox": [ + 89, + 859, + 482, + 891 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "plementary material. We compare our method with both supervised and self-supervised approaches. For supervised methods, we include Superhuman [23], MALA [15], PEA [21], and UNETR [18]. For self-supervised methods, we evaluate two categories: (1) general-purpose visual representation learning methods including MAE [20] and BYOL [16], which were originally designed for natural images but adapted to EM data; and (2) EM-specific self-supervised methods including dbMIM [8] and MS-Con-EM [9], which were specifically designed for electron microscopy images. All self-supervised methods utilize the same EMmamba backbone for fair comparison.", + "bbox": [ + 511, + 544, + 906, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on MEC Dataset. Table 1 presents the quantitative results on the large-scale MEC dataset. Compared to models with similar parameter counts, on the Wafer4 volume under the waterz post-processing, our TokenUnify approach achieves approximately a $25\\%$ performance improvement over MAE and a $44\\%$ improvement compared to direct training without pretraining. Visualization results in Fig. 4 and Fig. 5 further demonstrate that our approach significantly outperforms other methods, particularly in challenging regions with complex neuronal structures.", + "bbox": [ + 511, + 729, + 908, + 881 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13609", + "bbox": [ + 480, + 941, + 519, + 954 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ca765afeba898f22e39affead4bf0485ec10b71e7272a27d207f57ccf5b1281.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 87, + 906, + 378 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/12dd48e2cc5a9a338d09f9f6b46beb755632fecd9e76de421ddf68d95efe5aec.jpg", + "image_caption": [ + "Figure 4. 2D visualization results for two slices from the MEC dataset: wafer 4 and wafer 36-2. The left side displays the EM raw images, while the right side presents the affinity and segmentation results. Red boxes indicate the regions of mis-segmentation." + ], + "image_footnote": [], + "bbox": [ + 96, + 425, + 906, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on AC3/4 Datasets. We also evaluate performance on the smaller AC3/4 dataset, which contains only about 1/10 of the labeled pixels compared to MEC. As shown in Table 2, even in this low-data scenario, the Mamba architecture combined with TokenUnify pretraining achieves performance comparable to state-of-the-art supervised methods like PEA. Additionally, it demonstrates a $11\\%$ performance improvement over the MAE pretraining approach. This highlights the effectiveness of TokenUnify in scenarios with limited labeled data.", + "bbox": [ + 89, + 672, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling Law Analysis. To investigate how our method scales with model size, we evaluated various initialization and training strategies by scaling model parameters from 100M to 1B. Results in Fig. 6 show that all methods ex", + "bbox": [ + 89, + 828, + 482, + 888 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0d5089d7f9b9e2100df2021ffa21be50419adf742f75f8fc5a586d02765894c.jpg", + "image_caption": [ + "Figure 5. 3D visualization results on two MEC datasets: wafer 4 and wafer 36-2. The red arrows indicate over-merged structures, and the orange arrows highlight over-split structures.", + "Figure 6. Performance evaluation of models with 100M, 200M, 500M, and 1B parameters. The models were trained for 100 epochs on the MEC dataset." + ], + "image_footnote": [], + "bbox": [ + 516, + 672, + 898, + 816 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13610", + "bbox": [ + 478, + 941, + 519, + 953 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f783d666fe097bee87ed62b11499099d5dd7dabbbb4af74dabe2a5cde8ddc943.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVOIM↓VOIS↓VOI↓ARAND↓Param.(M)
Supervised Methods
Superhuman [23]0.48820.65631.14450.17481.478
MALA [15]0.45710.67671.13380.166484.02
PEA [21]0.55220.49801.05020.20931.480
UNETR [18]0.77990.73991.51980.2411129.1
EMmamba0.93780.86291.80070.284028.30
Self-Supervised Methods
Random0.93780.86291.80070.2840
MAE [20]0.31670.79631.11310.1050
BYOL [16]0.38420.76211.14630.152728.30
dbMIM [8]0.46230.62791.09020.1432
MS-Con-EM [9]0.47850.65381.13230.1498
TokenUnify0.44790.54390.99180.1366
", + "bbox": [ + 91, + 89, + 482, + 301 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/59c940550e87c7a9ecd6922775c3c3938307095cdb2c9f902477163b4f78196c.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of segmentation results on AC3/4 datasets with waterz [15] post-processing algorithm. All self-supervised methods use the same EMmamba backbone. \"Random\" refers to EMmamba without any pretraining. The best results are in bold and the second best results are underlined." + ], + "table_footnote": [], + "table_body": "
Pretraining StrategyVOI ↓ARAND ↓
RandomNextNext-all
✓1.26800.0862
✓4.04180.4416
✓✓1.13000.0692
✓✓1.19070.1203
✓✓✓0.99510.0509
", + "bbox": [ + 151, + 383, + 423, + 494 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f33e291facc650128b06e8b403d998585db514c7827ba1297d39d68d6ee344ba.jpg", + "table_caption": [ + "Table 3. Ablation for different pretraining strategies on wafer4." + ], + "table_footnote": [], + "table_body": "
Finetuning ModuleVOI ↓ARAND ↓
MambaEncoderDecoder
✓1.13620.0782
✓1.55560.1370
✓1.52950.1212
✓✓1.10650.0629
✓✓✓0.99510.0509
", + "bbox": [ + 143, + 522, + 431, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation for the fine-tuning schemes on wafer4.", + "bbox": [ + 117, + 643, + 454, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "hibit performance improvements after pretraining. In contrast, TokenUnify demonstrates strong scaling properties, consistently outperforming other methods across both small and large model sizes. Notably, the Mamba architecture achieves competitive segmentation performance with fewer parameters compared to Transformer, confirming its efficiency for processing long sequence data in connectomics.", + "bbox": [ + 89, + 683, + 482, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.1. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 799, + 254, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation experiments to analyze the key components of our approach. As shown in Table 3 and Table 4, we examine two critical aspects:", + "bbox": [ + 89, + 820, + 482, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Token Prediction Strategies. We systematically evaluate different token prediction combinations dur", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ing pretraining on the wafer4 dataset (Table 3). The full combination (Random + Next + Next-all) achieves optimal performance $(\\mathrm{VOI} = 0.9951)$ , significantly outperforming individual strategies and demonstrating substantial improvements over baseline approaches. Next-token prediction alone performs poorly $(\\mathrm{VOI} = 4.0418)$ , confirming that conventional autoregressive pretraining is insufficient for visual tasks. Unlike sequential text, visual data requires global spatial understanding that pure autoregressive modeling cannot capture effectively. Random token prediction provides better spatial initialization by encouraging spatial consistency, while the strategic combination of multiple prediction objectives yields more robust and generalizable representations for segmentation. These results validate our hypothesis that diverse pretraining tasks are essential for effective visual representation learning.", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tuning Components. We also investigate which network components benefit most from fine-tuning (Table 4). Our architecture consists of three main parts: Mamba blocks (sequence modeling), encoder (downsampling), and decoder (upsampling). Fine-tuning only the Mamba blocks yields substantially better results $(\\mathrm{VOI} = 1.1362)$ than fine-tuning only the encoder $(\\mathrm{VOI} = 1.5556)$ or decoder $(\\mathrm{VOI} = 1.5295)$ , demonstrating the critical role of sequence modeling in connectomics segmentation. For resource-constrained scenarios, prioritizing the fine-tuning of Mamba blocks provides an efficient balance between performance and computational cost.", + "bbox": [ + 511, + 335, + 906, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 531, + 633, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose TokenUnify, a framework that unifies random token prediction, next-token prediction, and next-all token prediction to capture multi-scale visual structures. Our analysis shows this approach mitigates cumulative errors in autoregressive modeling. To validate TokenUnify, we introduce a large-scale EM neuron segmentation dataset with 1.2 billion labeled voxels across six mouse brain regions. Experiments demonstrate a $44\\%$ improvement over training from scratch and a $25\\%$ gain over MAE pretraining. Scaling analysis further highlights its efficiency with increasing model size and data volume. By releasing our dataset and implementation, we aim to advance biological imaging and inspire future research in visual representation learning.", + "bbox": [ + 511, + 556, + 906, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 513, + 766, + 671, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grant 624B2137. We thank the staff at the Institute of Artificial Intelligence, Hefei Comprehensive National Science Center for their assistance with biological specimen preparation, imaging, and annotation. We are grateful to the anonymous reviewers for their thorough reviews and constructive suggestions.", + "bbox": [ + 511, + 791, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13611", + "bbox": [ + 480, + 941, + 517, + 954 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] Ignacio Arganda-Carreras, Srinivas C Turaga, Daniel R Berger, Dan Ciresan, Alessandro Giusti, Luca M Gambardella, Jürgen Schmidhuber, Dmitry Laptev, Sarvesh Dwivedi, Joachim M Buhmann, et al. Crowdsourcing the creation of image segmentation algorithms for connectomics. Frontiers in Neuroanatomy, 9:152591, 2015. 6", + "[3] Gregor Bachmann and Vaishnavh Nagarajan. The pitfalls of next-token prediction. arXiv preprint arXiv:2403.06963, 2024. 1", + "[4] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785, 2023. 1, 3", + "[5] Thorsten Beier, Constantin Pape, Nasim Rahaman, Timo Prange, Stuart Berg, Davi D Bock, Albert Cardona, Graham W Knott, Stephen M Plaza, Louis K Scheffer, et al. Multicut brings automated neurite segmentation closer to human performance. Nature Methods, 14(2):101-102, 2017. 2, 5, 6", + "[6] Kevin Briggman, Winfried Denk, Sebastian Seung, Moritz Helmstaedter, and Srinivas C Turaga. Maximin affinity learning of image segmentation. In NeurIPS, 2009. 2", + "[7] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 3", + "[8] Yinda Chen, Wei Huang, Shenglong Zhou, Qi Chen, and Zhiwei Xiong. Self-supervised neuron segmentation with multi-agent reinforcement learning. In IJCAI, 2023. 1, 3, 6, 8", + "[9] Yinda Chen, Wei Huang, Xiaoyu Liu, Shiyu Deng, Qi Chen, and Zhiwei Xiong. Learning multiscale consistency for self-supervised electron microscopy instance segmentation. In ICASSP, 2024. 3, 6, 8", + "[10] Zhimin Chen, Yingwei Li, Longlong Jing, Liang Yang, and Bing Li. Point cloud self-supervised learning via 3d to multiview masked autoencoder. arXiv preprint arXiv:2311.10887, 2023. 3", + "[11] Zhimin Chen, Longlong Jing, Yingwei Li, and Bing Li. Bridging the domain gap: Self-supervised 3d scene understanding with foundation models. In NeurIPS, 2024. 3", + "[12] Bo Dang, Wenchao Zhao, Yufeng Li, Danqing Ma, Qixuan Yu, and Elly Yijun Zhu. Real-time pill identification for the visually impaired using deep learning. arXiv preprint arXiv:2405.05983, 2024. 6", + "[13] Zhiyuan Ding, Qi Dong, Haote Xu, Chenxin Li, Xinghao Ding, and Yue Huang. Unsupervised anomaly segmentation for brain lesions using dual semantic-manifold reconstruction. In ICONIP, 2022. 3", + "[14] Alaaeldin El-Nouby, Michal Klein, Shuangfei Zhai, Miguel Angel Bautista, Alexander Toshev, Vaishaal Shankar," + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Joshua M Susskind, and Armand Joulin. Scalable pretraining of large autoregressive image models. arXiv preprint arXiv:2401.08541, 2024. 1, 3", + "[15] Jan Funke, Fabian Tschopp, William Grisaitis, Arlo Sheridan, Chandan Singh, Stephan Saalfeld, and Srinivas C Turaga. Large scale image segmentation with structured loss based deep learning for connectome reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(7):1669-1680, 2018. 2, 5, 6, 8", + "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020. 6, 8", + "[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3", + "[18] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In WACV, 2022. 6, 8", + "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3", + "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 2, 3, 6, 8", + "[21] Wei Huang, Shiyu Deng, Chang Chen, Xueyang Fu, and Zhiwei Xiong. Learning to model pixel-embedded affinity for homogeneous instance segmentation. In AAAI, 2022. 1, 2, 6, 8", + "[22] Narayanan Kasthuri, Kenneth Jeffrey Hayworth, Daniel Raimund Berger, Richard Lee Schalek, José Angel Conchello, Seymour Knowles-Barley, Dongil Lee, Amelio Vázquez-Reina, Verena Kaynig, Thouis Raymond Jones, et al. Saturated reconstruction of a volume of neocortex. Cell, 162(3):648-661, 2015. 5", + "[23] Kisuk Lee, Jonathan Zung, Peter Li, Viren Jain, and H Sebastian Seung. Superhuman accuracy on the snemi3d connectomics challenge. arXiv preprint arXiv:1706.00120, 2017. 6, 8", + "[24] Chenxin Li, Yunlong Zhang, Zhehan Liang, Wenao Ma, Yue Huang, and Xinghao Ding. Consistent posterior distributions under vessel-mixing: A regularization for cross-domain retinal artery/vein classification. In ICIP, 2021. 3", + "[25] Panfeng Li, Youzuo Lin, and Emily Schultz-Fellenz. Contextual hourglass network for semantic segmentation of high resolution aerial imagery. arXiv preprint arXiv:1810.12813, 2018. 5", + "[26] Panfeng Li, Mohamed Abouelenien, and Rada Mihalcea. Deception detection from linguistic and physiological data streams using bimodal convolutional neural networks. arXiv preprint arXiv:2311.10944, 2023. 5", + "[27] Tianhong Li, Huiwen Chang, Shlok Mishra, Han Zhang, Dina Katabi, and Dilip Krishnan. Mage: Masked generative encoder to unify representation learning and image synthesis. In CVPR, 2023. 3" + ], + "bbox": [ + 516, + 92, + 905, + 896 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13612", + "bbox": [ + 480, + 941, + 519, + 954 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9): 1-35, 2023. 1", + "[29] Xiaoyu Liu, Wei Huang, Yueyi Zhang, and Zhiwei Xiong. Biological instance segmentation with a superpixel-guided graph. In IJCAI, 2022. 2", + "[30] Xiaoyu Liu, Miaomiao Cai, Yinda Chen, Yueyi Zhang, Te Shi, Ruobing Zhang, Xuejin Chen, and Zhiwei Xiong. Cross-dimension affinity distillation for 3d em neuron segmentation. In CVPR, 2024. 1, 3", + "[31] Christopher W Lynn and Danielle S Bassett. The physics of brain network structure, function and control. Nature Reviews Physics, 1(5):318-332, 2019. 1", + "[32] Juan Nunez-Iglesias, Ryan Kennedy, Toufiq Parag, Jianbo Shi, and Dmitri B Chklovskii. Machine learning of hierarchical clustering to segment 2d and 3d images. PLoS ONE, 8(8):e71715, 2013. 6", + "[33] Nobuhiko Ohno, Mitsuhiko Katoh, Yurika Saitoh, Sei Saitoh, and Shinichi Ohno. Three-dimensional volume imaging with electron microscopy toward connectome. Microscopy, 64(1):17-26, 2015. 1", + "[34] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2023. 1", + "[35] Philipp Schlegel, Alexander S Bates, Tejal Parag, Gregory SXE Jefferis, and Davi D Bock. Automatic detection of synaptic partners in a whole-brain drosophila em dataset. Nature Methods, 18(8):877-884, 2021. 5", + "[36] Arlo Sheridan, Tri M Nguyen, Diptodip Deb, Wei-Chung Allen Lee, Stephan Saalfeld, Srinivas C Turaga, Uri Manor, and Jan Funke. Local shape descriptors for neuron segmentation. Nature Methods, 20(2):295-303, 2023. 1", + "[37] Mannat Singh, Quentin Duval, Kalyan Vasudev Alwala, Haoqi Fan, Vaibhav Aggarwal, Aaron Adcock, Armand Joulin, Piotr Dollar, Christoph Feichtenhofer, Ross Girshick, et al. The effectiveness of mae pre-pretraining for billionscale pretraining. In ICCV, 2023. 1", + "[38] Shin-ya Takemura, Yoshinori Aso, Toshihide Hige, Allan Wong, Zhiyuan Lu, C Shan Xu, Patricia K Rivlin, Harald Hess, Ting Zhao, Toufiq Parag, et al. A connectome of a learning and memory center in the adult drosophila brain. eLife, 6:e26975, 2017. 5", + "[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[40] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1" + ], + "bbox": [ + 91, + 90, + 480, + 896 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Srinivas C Turaga, Joseph F Murray, Viren Jain, Fabian Roth, Moritz Helmstaedter, Kevin Briggman, Winfried Denk, and H Sebastian Seung. Convolutional networks can learn to generate affinity graphs for image segmentation. Neural Computation, 22(2):511-538, 2010. 2", + "[42] Donglai Wei, Zudi Lin, Daniel Franco-Barranco, Nils Wendt, Xingyu Liu, Wenjie Yin, Xin Huang, Aarush Gupta, Won-Dong Jang, Xueying Wang, et al. Mitoem dataset: Large-scale 3d mitochondria instance segmentation from em images. In MICCAI, 2020. 5", + "[43] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In MICCAI, 2024. 3, 5", + "[44] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021. 3", + "[45] Jingyu Zhang, Ao Xiang, Yu Cheng, Qin Yang, and Liyang Wang. Research on detection of floating objects in river and lake based on ai intelligent image recognition. arXiv preprint arXiv:2404.06883, 2024. 6" + ], + "bbox": [ + 514, + 90, + 903, + 372 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13613", + "bbox": [ + 480, + 941, + 517, + 953 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/0bebb7e8-7094-4aab-a136-ac9635434593_model.json b/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/0bebb7e8-7094-4aab-a136-ac9635434593_model.json new file mode 100644 index 0000000000000000000000000000000000000000..22292c1477764e748826fb5625a56af827fb9cb2 --- /dev/null +++ b/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/0bebb7e8-7094-4aab-a136-ac9635434593_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.89, + 0.154 + ], + "angle": 0, + "content": "TokenUnify: Scaling Up Autoregressive Pretraining for Neuron Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.18, + 0.751, + 0.218 + ], + "angle": 0, + "content": "Yinda Chen\\(^{1,2*}\\) Haoyuan Shi\\(^{1,2*}\\) Xiaoyu Liu\\(^{1}\\) Te Shi\\(^{2}\\) Ruobing Zhang\\(^{3,2}\\) Dong Liu\\(^{1}\\) Zhiwei Xiong\\(^{1,2\\dagger}\\) Feng Wu\\(^{1,2‡}\\)" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.233, + 0.824, + 0.325 + ], + "angle": 0, + "content": "\\(^{1}\\)MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China \n\\(^{2}\\)Institute of Artificial Intelligence, Hefei Comprehensive National Science Center \n\\(^{3}\\)Institute for Brain and Intelligence, Fudan University \n{cyd0806, haoyuan.shi}@mail.ustc.edu.cn, {zwxiong, fengwu}@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.357, + 0.327, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.486, + 0.829 + ], + "angle": 0, + "content": "Neuron segmentation from electron microscopy (EM) volumes is crucial for understanding brain circuits, yet the complex neuronal structures in high-resolution EM images present significant challenges. EM data exhibits unique characteristics including high noise levels, anisotropic voxel dimensions, and ultra-long spatial dependencies that make traditional vision models inadequate. Inspired by autoregressive pretraining in language models, we propose TokenUnify, a hierarchical predictive coding framework that captures multi-scale dependencies through three complementary learning objectives. TokenUnify integrates random token prediction, next-token prediction, and next-all token prediction to create a comprehensive representational space with emergent properties. From an information-theoretic perspective, these three tasks are complementary and provide optimal coverage of visual data structure, with our approach reducing autoregressive error accumulation from \\( O(K) \\) to \\( O(\\sqrt{K}) \\) for sequences of length \\( K \\). We also introduce a large-scale EM dataset with 1.2 billion annotated voxels, offering ideal long-sequence visual data with spatial continuity. Leveraging the Mamba architecture's linear-time sequence modeling capabilities, TokenUnify achieves a \\( 44\\% \\) performance improvement on downstream neuron segmentation and outperforms MAE by \\( 25\\% \\). Our approach demonstrates superior scaling properties as model size increases, effectively bridging the gap between pretraining strategies for language and vision models. Code is available at https://github.com/ydchen0806/TokenUnify." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.357, + 0.645, + 0.373 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.908, + 0.579 + ], + "angle": 0, + "content": "Understanding neural circuits is a central goal in neuroscience. Electron microscopy (EM) imaging enables reconstruction of neurons and synaptic connections at high resolution [31, 33]. However, EM neuron segmentation presents significant challenges due to three key properties: (1) high noise levels from electron beam interactions, (2) anisotropic voxel dimensions with coarser z-axis resolution, and (3) long-range spatial continuity spanning thousands of patches. Current methods struggle to capture global context and long-range dependencies [8, 21, 30, 36]. Despite success in natural language processing [28], similar advancements for long-sequence visual data remain largely unexplored." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.909, + 0.775 + ], + "angle": 0, + "content": "Large language models (LLMs) show remarkable scaling capabilities [1, 39, 40] through effective autoregressive modeling. However, visual data has more complex structures than text. Contrastive methods like DINO v2 [34] and masked reconstruction like MAE [8, 20] demonstrate strong representation learning but lack favorable scaling laws [37]. Masked prediction faces theoretical limitations where estimation error scales as \\( O(\\sqrt{s \\log p / n}) \\) for sparsity \\( s \\), dimensionality \\( p \\), and sample size \\( n \\). Autoregressive approaches like AIM [14] and LVM [4] attempt to bridge this gap, but error accumulation \\( O(K) \\) with sequence length \\( K \\) and computational complexity hinder their application to high-resolution EM images [3]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.775, + 0.909, + 0.897 + ], + "angle": 0, + "content": "This paper tackles visual autoregressive learning challenges through hierarchical predictive coding. Let \\( X = \\{x_{1}, x_{2}, \\ldots, x_{K}\\} \\) represent visual tokens from an EM image, where \\( \\mathcal{M} \\subset \\{1, \\ldots, K\\} \\) denotes masked indices and \\( \\mathcal{M}^{c} \\) denotes unmasked tokens. We recognize that three prediction tasks capture complementary visual structure aspects: random token prediction learns noise-robust spatial patterns, next-token prediction captures sequential de" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.85, + 0.241, + 0.862 + ], + "angle": 0, + "content": "*Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.862, + 0.197, + 0.875 + ], + "angle": 0, + "content": "†Project Leader" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.875, + 0.238, + 0.887 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding Author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.85, + 0.241, + 0.887 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.942, + 0.521, + 0.955 + ], + "angle": 0, + "content": "13604" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.086, + 0.87, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.24, + 0.908, + 0.285 + ], + "angle": 0, + "content": "Figure 1. TokenUnify divides the 3D EM image into non-overlapping patches, which are tokenized into a sequence of K tokens. Three complementary tasks are performed for rich 3D image representations: (a) random token prediction for spatial pattern learning, (b) next token prediction for sequential dependencies, and (c) next-all token prediction for global structure modeling." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.31, + 0.482, + 0.384 + ], + "angle": 0, + "content": "pencies, and next-all token prediction models global context. Under regularity conditions \\((\\mathrm{AR}(\\infty)\\) process with summable coefficients), autoregressive models achieve asymptotically optimal prediction error \\(O(\\sigma^2)\\) as capacity increases, where \\(\\sigma^2\\) is irreducible noise." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.386, + 0.483, + 0.732 + ], + "angle": 0, + "content": "We propose TokenUnify, a unified framework that integrates these three predictive tasks within a coherent information-theoretic formulation (Fig. 1). This creates a comprehensive representational space where random token prediction provides local robustness, next-token prediction ensures structural continuity, and next-all token prediction captures global morphology, enabling simultaneous encoding of fine-grained details and global relationships unachievable with single objectives. Our approach reduces cumulative errors from \\( O(K) \\) to \\( O(\\sqrt{K}) \\) by distributing prediction errors across multiple positions rather than sequential accumulation, analogous to central limit theorem scaling. To handle massive high-resolution EM data efficiently, we leverage the Mamba architecture, which reduces computational complexity from quadratic to linear while maintaining superior performance compared to Transformers. We validate our approach on a large-scale EM dataset collected from mouse brain slices, featuring thousands of continuous tokens that ensure spatial continuity crucial for understanding complex neuronal structures. Through comprehensive annotation of six functional brain regions totaling 1.2 billion voxels, we create one of the largest manually annotated EM neuron segmentation datasets to date." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Comprehensive evaluation demonstrates that TokenUnify achieves a \\(44\\%\\) improvement on downstream neuron segmentation and outperforms MAE [20] by \\(25\\%\\) in pretraining performance. Furthermore, TokenUnify demonstrates superior scaling properties, offering a promising paradigm for pretraining large-scale visual models." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.824, + 0.436, + 0.837 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.838, + 0.483, + 0.899 + ], + "angle": 0, + "content": "1. We introduce a novel hierarchical predictive coding framework that unifies three distinct visual structure perspectives—random token prediction for spatial robustness, next-token prediction for sequential continuity, and" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.31, + 0.905, + 0.415 + ], + "angle": 0, + "content": "next-all token prediction for global context—within a coherent information-theoretic formulation. Our theoretical analysis (detailed in supplementary material) demonstrates why these prediction tasks are complementary and together provide optimal coverage of visual data structure, while reducing autoregressive error accumulation from \\( O(K) \\) to \\( O(\\sqrt{K}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.416, + 0.905, + 0.534 + ], + "angle": 0, + "content": "2. We achieve a \\(44\\%\\) performance improvement on downstream neuron segmentation tasks and realize the first billion-parameter Mamba network for visual autoregression, demonstrating both effectiveness and computational efficiency in processing long-sequence visual data. Our approach outperforms existing methods including MAE by \\(25\\%\\) while exhibiting favorable scaling properties as model size increases." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.537, + 0.907, + 0.643 + ], + "angle": 0, + "content": "3. We construct one of the largest EM neuron segmentation datasets with 1.2 billion finely annotated voxels across six functional brain regions, providing an ideal testbed for long-sequence visual modeling. This comprehensive dataset offers spatially continuous visual data with thousands of tokens per volume, enabling rigorous validation of autoregressive methods for biological image analysis." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.416, + 0.907, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.655, + 0.655, + 0.67 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.907, + 0.892 + ], + "angle": 0, + "content": "Neuron Segmentation. In the field of EM image analysis, neuron instance segmentation has emerged as a vital task. Turaga et al. [41] pioneered the use of convolutional neural networks (CNNs) for affinity generation, followed by clustering of voxels in the affinity graph using post-processing techniques such as waterz [15] and LMC [5]. Since then, affinity-based methods have seen significant advancements. Funke et al. [15] introduced the MALIS loss [6] during training to ensure topologically accurate segmentation. Recent work by Huang et al. [21] proposed an embedding pyramid module to capture affinity information across different scales. Liu et al. [29] combined both affinity and embedding features with graph neural networks (GNNs) to enhance the segmentation of adjacent objects in the feature" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.942, + 0.519, + 0.954 + ], + "angle": 0, + "content": "13605" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "space. Furthermore, Liu [30] introduced a knowledge distillation framework, transferring knowledge from 3D models to 2D models, which resulted in improved performance and efficiency in neuron segmentation. Despite these advances, mamba-based architectures [17], have yet to be fully explored for the anisotropic resolution characteristics of EM imaging. In this paper, we enhance the Segmamba decoder [43] to better process the anisotropic features of EM data using an affinity-based approach." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.228, + 0.485, + 0.471 + ], + "angle": 0, + "content": "Self-supervised Pretraining. Self-supervised pretraining has been a cornerstone for enhancing model representation capabilities. Approaches based on contrastive learning for representation extraction [9, 19, 24, 44] and masked reconstruction methods [8, 10, 11, 13, 20, 27] have shown promise. However, current vision models have not exhibited the same sublinear scaling laws as language models. To address this issue, some tasks have adopted autoregressive pre-training paradigms similar to those used in language models [4, 7, 14], though the training costs remain a significant concern. In this paper, we explore the potential of long visual token autoregressive pre-training and introduce a collaborative training scheme for long token prediction. Our approach aims to balance the scaling laws and training costs, demonstrating improvements in fine-grained visual tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.317, + 0.499 + ], + "angle": 0, + "content": "3. Theoretical Motivations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.554 + ], + "angle": 0, + "content": "This section provides the theoretical foundations that motivate our approach. Detailed derivations and proofs are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Notation. Let \\(\\mathbf{X} = \\{x_{1}, x_{2}, \\ldots, x_{K}\\}\\) represent a sequence of visual tokens extracted from an EM image, where \\(\\mathcal{M} \\subset \\{1, \\ldots, K\\}\\) denotes masked token indices, \\(\\mathcal{M}^c\\) denotes unmasked tokens, and \\(x_{ 0\\) is a small constant, and \\(c > 0\\) is the magnitude of negative curvature. This geometric property enables efficient encoding of both fine-grained details (in near-flat regions) and complex hierarchical structures (in hyperbolic-like regions)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.506, + 0.605, + 0.521 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.625, + 0.545 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.766 + ], + "angle": 0, + "content": "The workflow of our method is shown in Fig. 2. Our approach comprises two stages: (1) hierarchical predictive pre-training, where we learn a generic visual representation \\( f_{\\theta_1}(\\cdot) \\) from unlabeled data \\( \\mathbf{X} \\) by integrating three complementary prediction tasks that capture both local and global dependencies; and (2) task-specific fine-tuning, where we adapt the pre-trained model to downstream tasks using both \\( \\mathbf{X} \\) and labels \\( Y \\) through a task-specific model \\( g_{\\theta_2}(\\cdot) \\) initialized with \\( \\theta_1 \\). For EM neuron segmentation, we process input volumes \\( \\mathbf{X} = \\{X^{(1)},\\dots,X^{(T)}\\} \\), where each \\( X^{(t)} \\in \\mathbb{R}^{D\\times H\\times W} \\) represents a 3D image, by partitioning them into smaller patches \\( x \\in \\mathbb{R}^{D' \\times H' \\times W'} \\) and leveraging Mamba's efficient sequence modeling capabilities to handle the high-dimensional, long-sequence nature of the data." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.773, + 0.905, + 0.79 + ], + "angle": 0, + "content": "4.2. Hierarchical Predictive Coding in Token Space" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Building on our theoretical insights, we develop a hierarchical predictive coding framework that operates on the manifold of token embeddings to effectively capture the multiscale structure of high-dimensional neural image data. Our framework is motivated by the principle of maximum information preservation across complementary representational dimensions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.943, + 0.52, + 0.954 + ], + "angle": 0, + "content": "13606" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.91, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.199, + 0.908, + 0.242 + ], + "angle": 0, + "content": "Figure 2. Illustration of the TokenUnify pretraining workflow. The image \\(\\mathbf{X}\\) is fed into the Tokenizer, transforming it into a long sequence of tokens \\(x_{i}|_{i = 1}^{K}\\). The three prediction tasks (random token, next token, and next-all token) are performed with complementary objectives. The Perceiver Resampler is employed to convert varying-size large feature maps into a few visual tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.484, + 0.388 + ], + "angle": 0, + "content": "Notation. Let \\(\\mathbf{x} = \\{x_{1}, x_{2}, \\ldots, x_{K}\\}\\) represent a sequence of visual tokens extracted from a volumetric EM image, where each \\(x_{i} \\in \\mathbb{R}^{d}\\). For random token prediction, \\(\\mathcal{M} \\subset \\{1, \\ldots, K\\}\\) denotes the set of masked token indices with \\(|\\mathcal{M}| = \\rho K\\) for masking ratio \\(\\rho\\), and \\(\\mathcal{M}^{c} = \\{1, \\ldots, K\\} \\setminus \\mathcal{M}\\) denotes unmasked tokens. For sequential prediction, \\(x_{ 0\\) is the annealing temperature controlling transition sharpness." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.254, + 0.424, + 0.267 + ], + "angle": 0, + "content": "Our unified objective function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.281, + 0.483, + 0.297 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {T o k e n U n i f y}} = \\alpha (t) \\cdot \\mathcal {L} _ {\\text {r a n d o m}} + \\beta (t) \\cdot \\mathcal {L} _ {\\text {n e x t}} + \\gamma (t) \\cdot \\mathcal {L} _ {\\text {n e x t - a l l}}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.483, + 0.34 + ], + "angle": 0, + "content": "The curriculum implements progressive transition from local to global representation learning through:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.351, + 0.504, + 0.419 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} (t) = \\left\\{ \\begin{array}{l l} [ 2. 0, - 1. 0, - 2. 0 ] & \\text {i f} t < T _ {1} (\\text {r a n d o m - d o m i n a n t}) \\\\ [ - 1. 0, 2. 0, - 1. 0 ] & \\text {i f} T _ {1} \\leq t < T _ {2} (\\text {n e x t - d o m i n a n t}), \\\\ [ - 2. 0, - 1. 0, 2. 0 ] & \\text {i f} t \\geq T _ {2} (\\text {n e x t - a l l - d o m i n a n t}) \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.42, + 0.484, + 0.495 + ], + "angle": 0, + "content": "where \\( T_{1} = 0.3 \\times T_{\\mathrm{total}} \\) and \\( T_{2} = 0.7 \\times T_{\\mathrm{total}} \\) are transition thresholds. With \\( \\tau(t) = 1.0 \\), this yields dominant weights of \\( \\sim 0.73 \\) for the primary task while maintaining auxiliary contributions of \\( \\sim 0.18 \\) and \\( \\sim 0.09 \\) to preserve multi-task synergy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.484, + 0.556 + ], + "angle": 0, + "content": "For computational efficiency in next-all token prediction, we employ cross-attention with latent queries to aggregate sequence-wide information, enabling scalability to long token sequences without prohibitive complexity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.484, + 0.631 + ], + "angle": 0, + "content": "This hierarchical integration constructs a multi-resolution representational manifold capturing neural tissue organization across microscopic, mesoscopic, and macroscopic scales, providing robust foundations for volumetric electron microscopy circuit reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.64, + 0.355, + 0.657 + ], + "angle": 0, + "content": "4.3. Finetuning and Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.662, + 0.484, + 0.694 + ], + "angle": 0, + "content": "The segmentation network, denoted as \\( g_{\\mathrm{seg}}(\\cdot) \\), consists of an encoder \\( g_{e}(\\cdot) \\) and a decoder \\( g_{d}(\\cdot) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.705, + 0.483, + 0.722 + ], + "angle": 0, + "content": "\\[\ng _ {\\mathrm {s e g}} (x; \\theta_ {s}) = g _ {d} \\left(g _ {e} (x)\\right), \\quad \\theta_ {s} = \\left\\{\\theta_ {e}, \\theta_ {d} \\right\\}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.778 + ], + "angle": 0, + "content": "where \\(\\theta_{s}\\) represents the parameters of the entire segmentation network, and \\(\\theta_{e}\\) and \\(\\theta_{d}\\) are the parameters of the encoder and decoder, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.884 + ], + "angle": 0, + "content": "The encoder \\( g_{e}(\\cdot) \\) gradually downsamples the input volume and extracts high-level semantic features, while the decoder \\( g_{d}(\\cdot) \\) upsamples the encoded features back to the original resolution. Meanwhile, the output of each downsampling layer in the encoder is connected to the corresponding layer in the decoder via skip connections to fuse local and global multi-scale information." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.184, + 0.686, + 0.197 + ], + "angle": 0, + "content": "(a) Sampling Area" + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.184, + 0.874, + 0.197 + ], + "angle": 0, + "content": "(b) Annotation" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.208, + 0.907, + 0.278 + ], + "angle": 0, + "content": "Figure 3. Illustration of the MEC dataset annotation. (a) shows the sampling locations of six different neural regions from the MEC raw data. Each region contains \\(1250 \\times 1250 \\times 125\\) voxels. (b) shows an example of the fine segmentation labels in one sampling region, which takes two experts one month to annotate." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.905, + 0.351 + ], + "angle": 0, + "content": "We adopt 3D ResUNet/ViT/Mamba as the backbone network, respectively. For the Mamba variant, we adapt SegMamba [43] for EM data, which we refer to as EMmamba, with enhanced decoder design for anisotropic EM features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.352, + 0.906, + 0.442 + ], + "angle": 0, + "content": "The output of the segmentation network \\(\\hat{y} = g_{\\mathrm{seg}}(x)\\in\\) \\(\\mathbb{R}^{C\\times D\\times H\\times W}\\) represents the predicted affinity map [25, 26], corresponding to the connectivity probability of each voxel in three directions. During training, the loss function for labeled samples is the mean squared error between the predicted and ground-truth affinity maps:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.451, + 0.905, + 0.493 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s e g}} = \\frac {1}{| \\mathcal {D} _ {l} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {l} |} \\| g _ {\\mathrm {s e g}} \\left(x _ {i} ^ {l}\\right) - y _ {i} \\| ^ {2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.622 + ], + "angle": 0, + "content": "During inference, for any new test sample \\( x_{t} \\), forward propagation through \\( g_{\\mathrm{seg}}(x_{t}) \\) yields its predicted affinity map. This predicted affinity map is then post-processed using a seeded watershed algorithm and a structure-aware region agglomeration algorithm [5, 15] to obtain the final neuron instance segmentation. Detailed information on our segmentation process and the segmentation pipeline are illustrated in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.715, + 0.65 + ], + "angle": 0, + "content": "5. Datasets and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Datasets. For pretraining TokenUnify, we collect over 1 TB of publicly available unlabeled EM imaging data from four large-scale datasets: FAFB [35], MitoEM [42], FIB-25 [38], and Kasthuri15 [22]. These datasets span diverse organisms including Drosophila, mouse, rat, and human samples, ensuring comprehensive coverage of EM morphology variations. We sample from these datasets with equal probability to guarantee diversity in pretraining data. Additional details are provided in supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For downstream fine-tuning and evaluation, we employ two datasets: AC3/4 [22] and our large-scale MEC dataset. The AC3/4 dataset consists of 256/100 successive EM images with \\(1024 \\times 1024\\) spatial resolution from mouse somatosensory cortex. We partition AC4 into training (first 80 images) and validation (last 20 images), while using AC3 (first 100 images) for testing." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.943, + 0.519, + 0.954 + ], + "angle": 0, + "content": "13608" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.09, + 0.904, + 0.466 + ], + "angle": 0, + "content": "
Post.MethodWafer4Wafer36-2Param. (M)
VOIM↓VOIS↓VOI↓ARAND↓VOIM↓VOIS↓VOI↓ARAND↓
LMC [5]Supervised Methods
Superhuman [23]0.33281.12581.45870.17360.15060.45880.60940.08361.478
MALA [15]0.54381.50272.03750.11150.31791.06641.38430.157084.03
PEA [21]0.33810.92761.26580.06770.27870.42790.70660.11691.480
UNETR [18]0.45041.65812.10850.26580.44780.52170.96960.2913129.1
EMmamba0.49151.29241.78390.20520.24060.41890.65950.123128.30
Superhuman [23]0.19481.96972.16440.24530.07921.16181.24100.13191.478
MALA [15]0.34162.41292.75450.25670.14481.96032.10520.197784.03
PEA [21]0.17051.59931.76980.15270.47191.12261.59450.15881.480
UNETR [18]0.17913.17153.35060.63300.09491.38581.48070.1578129.1
EMmamba0.15962.05802.21770.19730.08471.03511.11980.125328.30
LMC [5]Self-Supervised Methods
Random0.49151.29241.78390.20520.24060.41890.65950.123128.30
MAE [20]0.23251.09231.32480.09780.16290.45320.61610.0835
BYOL [16]0.25840.94531.20370.08910.17480.43680.61160.0826
dbMIM [8]0.23420.87961.11380.07420.14760.39820.54580.0782
MS-Con-EM [9]0.24830.91021.15850.08130.15420.40370.55790.0794
TokenUnify0.19530.79980.99510.05090.12620.35850.48480.0650
Random0.15962.05802.21770.19730.08471.03511.11980.125328.30
MAE [20]0.14931.92162.07090.15830.08921.09271.18190.1243
BYOL [16]0.16291.83742.00030.14270.09531.07821.17350.1287
dbMIM [8]0.15731.74361.90090.09130.09121.06491.15610.1185
MS-Con-EM [9]0.16091.76921.93010.10260.09631.07031.16660.1207
TokenUnify0.14181.51031.65210.05910.08271.02761.11030.1074
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.477, + 0.907, + 0.521 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of electron microscopy segmentation results on Wafer4 and Wafer36-2 datasets. Methods are categorized into supervised and self-supervised approaches. All self-supervised methods use the same EMmamba backbone for fair comparison. \"Random\" refers to EMmamba without any pretraining. The best results are in bold and the second best results are underlined." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.545, + 0.482, + 0.711 + ], + "angle": 0, + "content": "To validate algorithm performance on large-scale EM data, we collected a 2TB MEC dataset by imaging mouse somatosensory cortex, medial entorhinal cortex, and cerebral cortex regions. We selected 6 representative volumes from different neural regions (wafer4/25/26/26-2/36/36-2), each with dimensions \\(1250 \\times 1250 \\times 125\\) voxels. Dense manual annotation was performed on these regions, yielding over 1.2 billion annotated voxels through six months of work by two experts, as illustrated in Fig. 3. The dataset is split into training (wafer25/26/26-2/36), validation (wafer4), and testing (wafer36-2) sets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Metrics. To evaluate the performance of neuron segmentation [12, 45], we employ two widely-used metrics: Variation of Information (VOI) [32] and Adjusted Rand Index (ARAND) [2]. These metrics quantify the agreement between the predicted segmentation and the ground truth from different perspectives. Detailed descriptions of these metrics can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.835, + 0.216, + 0.853 + ], + "angle": 0, + "content": "6. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.86, + 0.483, + 0.892 + ], + "angle": 0, + "content": "We evaluate our proposed method on both MEC and AC3/4 datasets. Implementation details are provided in the sup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.727 + ], + "angle": 0, + "content": "plementary material. We compare our method with both supervised and self-supervised approaches. For supervised methods, we include Superhuman [23], MALA [15], PEA [21], and UNETR [18]. For self-supervised methods, we evaluate two categories: (1) general-purpose visual representation learning methods including MAE [20] and BYOL [16], which were originally designed for natural images but adapted to EM data; and (2) EM-specific self-supervised methods including dbMIM [8] and MS-Con-EM [9], which were specifically designed for electron microscopy images. All self-supervised methods utilize the same EMmamba backbone for fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.909, + 0.882 + ], + "angle": 0, + "content": "Results on MEC Dataset. Table 1 presents the quantitative results on the large-scale MEC dataset. Compared to models with similar parameter counts, on the Wafer4 volume under the waterz post-processing, our TokenUnify approach achieves approximately a \\(25\\%\\) performance improvement over MAE and a \\(44\\%\\) improvement compared to direct training without pretraining. Visualization results in Fig. 4 and Fig. 5 further demonstrate that our approach significantly outperforms other methods, particularly in challenging regions with complex neuronal structures." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.942, + 0.52, + 0.955 + ], + "angle": 0, + "content": "13609" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.088, + 0.907, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.397, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Figure 4. 2D visualization results for two slices from the MEC dataset: wafer 4 and wafer 36-2. The left side displays the EM raw images, while the right side presents the affinity and segmentation results. Red boxes indicate the regions of mis-segmentation." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.426, + 0.907, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.634, + 0.907, + 0.661 + ], + "angle": 0, + "content": "Figure 5. 3D visualization results on two MEC datasets: wafer 4 and wafer 36-2. The red arrows indicate over-merged structures, and the orange arrows highlight over-split structures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Results on AC3/4 Datasets. We also evaluate performance on the smaller AC3/4 dataset, which contains only about 1/10 of the labeled pixels compared to MEC. As shown in Table 2, even in this low-data scenario, the Mamba architecture combined with TokenUnify pretraining achieves performance comparable to state-of-the-art supervised methods like PEA. Additionally, it demonstrates a \\(11\\%\\) performance improvement over the MAE pretraining approach. This highlights the effectiveness of TokenUnify in scenarios with limited labeled data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.829, + 0.483, + 0.89 + ], + "angle": 0, + "content": "Scaling Law Analysis. To investigate how our method scales with model size, we evaluated various initialization and training strategies by scaling model parameters from 100M to 1B. Results in Fig. 6 show that all methods ex" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.673, + 0.9, + 0.818 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.835, + 0.907, + 0.877 + ], + "angle": 0, + "content": "Figure 6. Performance evaluation of models with 100M, 200M, 500M, and 1B parameters. The models were trained for 100 epochs on the MEC dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.942, + 0.52, + 0.954 + ], + "angle": 0, + "content": "13610" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.09, + 0.483, + 0.302 + ], + "angle": 0, + "content": "
MethodVOIM↓VOIS↓VOI↓ARAND↓Param.(M)
Supervised Methods
Superhuman [23]0.48820.65631.14450.17481.478
MALA [15]0.45710.67671.13380.166484.02
PEA [21]0.55220.49801.05020.20931.480
UNETR [18]0.77990.73991.51980.2411129.1
EMmamba0.93780.86291.80070.284028.30
Self-Supervised Methods
Random0.93780.86291.80070.2840
MAE [20]0.31670.79631.11310.1050
BYOL [16]0.38420.76211.14630.152728.30
dbMIM [8]0.46230.62791.09020.1432
MS-Con-EM [9]0.47850.65381.13230.1498
TokenUnify0.44790.54390.99180.1366
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.311, + 0.483, + 0.381 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of segmentation results on AC3/4 datasets with waterz [15] post-processing algorithm. All self-supervised methods use the same EMmamba backbone. \"Random\" refers to EMmamba without any pretraining. The best results are in bold and the second best results are underlined." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.384, + 0.424, + 0.496 + ], + "angle": 0, + "content": "
Pretraining StrategyVOI ↓ARAND ↓
RandomNextNext-all
✓1.26800.0862
✓4.04180.4416
✓✓1.13000.0692
✓✓1.19070.1203
✓✓✓0.99510.0509
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.505, + 0.474, + 0.52 + ], + "angle": 0, + "content": "Table 3. Ablation for different pretraining strategies on wafer4." + }, + { + "type": "table", + "bbox": [ + 0.145, + 0.523, + 0.432, + 0.634 + ], + "angle": 0, + "content": "
Finetuning ModuleVOI ↓ARAND ↓
MambaEncoderDecoder
✓1.13620.0782
✓1.55560.1370
✓1.52950.1212
✓✓1.10650.0629
✓✓✓0.99510.0509
" + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.644, + 0.455, + 0.658 + ], + "angle": 0, + "content": "Table 4. Ablation for the fine-tuning schemes on wafer4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.483, + 0.79 + ], + "angle": 0, + "content": "hibit performance improvements after pretraining. In contrast, TokenUnify demonstrates strong scaling properties, consistently outperforming other methods across both small and large model sizes. Notably, the Mamba architecture achieves competitive segmentation performance with fewer parameters compared to Transformer, confirming its efficiency for processing long sequence data in connectomics." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.255, + 0.814 + ], + "angle": 0, + "content": "6.1. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.821, + 0.483, + 0.867 + ], + "angle": 0, + "content": "We conduct ablation experiments to analyze the key components of our approach. As shown in Table 3 and Table 4, we examine two critical aspects:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Ablation on Token Prediction Strategies. We systematically evaluate different token prediction combinations dur" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.333 + ], + "angle": 0, + "content": "ing pretraining on the wafer4 dataset (Table 3). The full combination (Random + Next + Next-all) achieves optimal performance \\((\\mathrm{VOI} = 0.9951)\\), significantly outperforming individual strategies and demonstrating substantial improvements over baseline approaches. Next-token prediction alone performs poorly \\((\\mathrm{VOI} = 4.0418)\\), confirming that conventional autoregressive pretraining is insufficient for visual tasks. Unlike sequential text, visual data requires global spatial understanding that pure autoregressive modeling cannot capture effectively. Random token prediction provides better spatial initialization by encouraging spatial consistency, while the strategic combination of multiple prediction objectives yields more robust and generalizable representations for segmentation. These results validate our hypothesis that diverse pretraining tasks are essential for effective visual representation learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.907, + 0.519 + ], + "angle": 0, + "content": "Fine-tuning Components. We also investigate which network components benefit most from fine-tuning (Table 4). Our architecture consists of three main parts: Mamba blocks (sequence modeling), encoder (downsampling), and decoder (upsampling). Fine-tuning only the Mamba blocks yields substantially better results \\((\\mathrm{VOI} = 1.1362)\\) than fine-tuning only the encoder \\((\\mathrm{VOI} = 1.5556)\\) or decoder \\((\\mathrm{VOI} = 1.5295)\\), demonstrating the critical role of sequence modeling in connectomics segmentation. For resource-constrained scenarios, prioritizing the fine-tuning of Mamba blocks provides an efficient balance between performance and computational cost." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.532, + 0.634, + 0.547 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.907, + 0.755 + ], + "angle": 0, + "content": "We propose TokenUnify, a framework that unifies random token prediction, next-token prediction, and next-all token prediction to capture multi-scale visual structures. Our analysis shows this approach mitigates cumulative errors in autoregressive modeling. To validate TokenUnify, we introduce a large-scale EM neuron segmentation dataset with 1.2 billion labeled voxels across six mouse brain regions. Experiments demonstrate a \\(44\\%\\) improvement over training from scratch and a \\(25\\%\\) gain over MAE pretraining. Scaling analysis further highlights its efficiency with increasing model size and data volume. By releasing our dataset and implementation, we aim to advance biological imaging and inspire future research in visual representation learning." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.767, + 0.673, + 0.784 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.906, + 0.899 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant 624B2137. We thank the staff at the Institute of Artificial Intelligence, Hefei Comprehensive National Science Center for their assistance with biological specimen preparation, imaging, and annotation. We are grateful to the anonymous reviewers for their thorough reviews and constructive suggestions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.943, + 0.518, + 0.955 + ], + "angle": 0, + "content": "13611" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.269 + ], + "angle": 0, + "content": "[2] Ignacio Arganda-Carreras, Srinivas C Turaga, Daniel R Berger, Dan Ciresan, Alessandro Giusti, Luca M Gambardella, Jürgen Schmidhuber, Dmitry Laptev, Sarvesh Dwivedi, Joachim M Buhmann, et al. Crowdsourcing the creation of image segmentation algorithms for connectomics. Frontiers in Neuroanatomy, 9:152591, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.27, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[3] Gregor Bachmann and Vaishnavh Nagarajan. The pitfalls of next-token prediction. arXiv preprint arXiv:2403.06963, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[4] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.464 + ], + "angle": 0, + "content": "[5] Thorsten Beier, Constantin Pape, Nasim Rahaman, Timo Prange, Stuart Berg, Davi D Bock, Albert Cardona, Graham W Knott, Stephen M Plaza, Louis K Scheffer, et al. Multicut brings automated neurite segmentation closer to human performance. Nature Methods, 14(2):101-102, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[6] Kevin Briggman, Winfried Denk, Sebastian Seung, Moritz Helmstaedter, and Srinivas C Turaga. Maximin affinity learning of image segmentation. In NeurIPS, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[7] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[8] Yinda Chen, Wei Huang, Shenglong Zhou, Qi Chen, and Zhiwei Xiong. Self-supervised neuron segmentation with multi-agent reinforcement learning. In IJCAI, 2023. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[9] Yinda Chen, Wei Huang, Xiaoyu Liu, Shiyu Deng, Qi Chen, and Zhiwei Xiong. Learning multiscale consistency for self-supervised electron microscopy instance segmentation. In ICASSP, 2024. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[10] Zhimin Chen, Yingwei Li, Longlong Jing, Liang Yang, and Bing Li. Point cloud self-supervised learning via 3d to multiview masked autoencoder. arXiv preprint arXiv:2311.10887, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[11] Zhimin Chen, Longlong Jing, Yingwei Li, and Bing Li. Bridging the domain gap: Self-supervised 3d scene understanding with foundation models. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[12] Bo Dang, Wenchao Zhao, Yufeng Li, Danqing Ma, Qixuan Yu, and Elly Yijun Zhu. Real-time pill identification for the visually impaired using deep learning. arXiv preprint arXiv:2405.05983, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[13] Zhiyuan Ding, Qi Dong, Haote Xu, Chenxin Li, Xinghao Ding, and Yue Huang. Unsupervised anomaly segmentation for brain lesions using dual semantic-manifold reconstruction. In ICONIP, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Alaaeldin El-Nouby, Michal Klein, Shuangfei Zhai, Miguel Angel Bautista, Alexander Toshev, Vaishaal Shankar," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "Joshua M Susskind, and Armand Joulin. Scalable pretraining of large autoregressive image models. arXiv preprint arXiv:2401.08541, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[15] Jan Funke, Fabian Tschopp, William Grisaitis, Arlo Sheridan, Chandan Singh, Stephan Saalfeld, and Srinivas C Turaga. Large scale image segmentation with structured loss based deep learning for connectome reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(7):1669-1680, 2018. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[18] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In WACV, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[21] Wei Huang, Shiyu Deng, Chang Chen, Xueyang Fu, and Zhiwei Xiong. Learning to model pixel-embedded affinity for homogeneous instance segmentation. In AAAI, 2022. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[22] Narayanan Kasthuri, Kenneth Jeffrey Hayworth, Daniel Raimund Berger, Richard Lee Schalek, José Angel Conchello, Seymour Knowles-Barley, Dongil Lee, Amelio Vázquez-Reina, Verena Kaynig, Thouis Raymond Jones, et al. Saturated reconstruction of a volume of neocortex. Cell, 162(3):648-661, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.67 + ], + "angle": 0, + "content": "[23] Kisuk Lee, Jonathan Zung, Peter Li, Viren Jain, and H Sebastian Seung. Superhuman accuracy on the snemi3d connectomics challenge. arXiv preprint arXiv:1706.00120, 2017. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[24] Chenxin Li, Yunlong Zhang, Zhehan Liang, Wenao Ma, Yue Huang, and Xinghao Ding. Consistent posterior distributions under vessel-mixing: A regularization for cross-domain retinal artery/vein classification. In ICIP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.783 + ], + "angle": 0, + "content": "[25] Panfeng Li, Youzuo Lin, and Emily Schultz-Fellenz. Contextual hourglass network for semantic segmentation of high resolution aerial imagery. arXiv preprint arXiv:1810.12813, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.841 + ], + "angle": 0, + "content": "[26] Panfeng Li, Mohamed Abouelenien, and Rada Mihalcea. Deception detection from linguistic and physiological data streams using bimodal convolutional neural networks. arXiv preprint arXiv:2311.10944, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.843, + 0.905, + 0.897 + ], + "angle": 0, + "content": "[27] Tianhong Li, Huiwen Chang, Shlok Mishra, Han Zhang, Dina Katabi, and Dilip Krishnan. Mage: Masked generative encoder to unify representation learning and image synthesis. In CVPR, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.943, + 0.52, + 0.955 + ], + "angle": 0, + "content": "13612" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[28] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9): 1-35, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[29] Xiaoyu Liu, Wei Huang, Yueyi Zhang, and Zhiwei Xiong. Biological instance segmentation with a superpixel-guided graph. In IJCAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[30] Xiaoyu Liu, Miaomiao Cai, Yinda Chen, Yueyi Zhang, Te Shi, Ruobing Zhang, Xuejin Chen, and Zhiwei Xiong. Cross-dimension affinity distillation for 3d em neuron segmentation. In CVPR, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[31] Christopher W Lynn and Danielle S Bassett. The physics of brain network structure, function and control. Nature Reviews Physics, 1(5):318-332, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[32] Juan Nunez-Iglesias, Ryan Kennedy, Toufiq Parag, Jianbo Shi, and Dmitri B Chklovskii. Machine learning of hierarchical clustering to segment 2d and 3d images. PLoS ONE, 8(8):e71715, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[33] Nobuhiko Ohno, Mitsuhiko Katoh, Yurika Saitoh, Sei Saitoh, and Shinichi Ohno. Three-dimensional volume imaging with electron microscopy toward connectome. Microscopy, 64(1):17-26, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[34] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[35] Philipp Schlegel, Alexander S Bates, Tejal Parag, Gregory SXE Jefferis, and Davi D Bock. Automatic detection of synaptic partners in a whole-brain drosophila em dataset. Nature Methods, 18(8):877-884, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[36] Arlo Sheridan, Tri M Nguyen, Diptodip Deb, Wei-Chung Allen Lee, Stephan Saalfeld, Srinivas C Turaga, Uri Manor, and Jan Funke. Local shape descriptors for neuron segmentation. Nature Methods, 20(2):295-303, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[37] Mannat Singh, Quentin Duval, Kalyan Vasudev Alwala, Haoqi Fan, Vaibhav Aggarwal, Aaron Adcock, Armand Joulin, Piotr Dollar, Christoph Feichtenhofer, Ross Girshick, et al. The effectiveness of mae pre-pretraining for billionscale pretraining. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[38] Shin-ya Takemura, Yoshinori Aso, Toshihide Hige, Allan Wong, Zhiyuan Lu, C Shan Xu, Patricia K Rivlin, Harald Hess, Ting Zhao, Toufiq Parag, et al. A connectome of a learning and memory center in the adult drosophila brain. eLife, 6:e26975, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.757, + 0.482, + 0.826 + ], + "angle": 0, + "content": "[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.828, + 0.482, + 0.897 + ], + "angle": 0, + "content": "[40] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[41] Srinivas C Turaga, Joseph F Murray, Viren Jain, Fabian Roth, Moritz Helmstaedter, Kevin Briggman, Winfried Denk, and H Sebastian Seung. Convolutional networks can learn to generate affinity graphs for image segmentation. Neural Computation, 22(2):511-538, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[42] Donglai Wei, Zudi Lin, Daniel Franco-Barranco, Nils Wendt, Xingyu Liu, Wenjie Yin, Xin Huang, Aarush Gupta, Won-Dong Jang, Xueying Wang, et al. Mitoem dataset: Large-scale 3d mitochondria instance segmentation from em images. In MICCAI, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.234, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[43] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In MICCAI, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.277, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[44] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.32, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[45] Jingyu Zhang, Ao Xiang, Yu Cheng, Qin Yang, and Liyang Wang. Research on detection of floating objects in river and lake based on ai intelligent image recognition. arXiv preprint arXiv:2404.06883, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.943, + 0.519, + 0.954 + ], + "angle": 0, + "content": "13613" + } + ] +] \ No newline at end of file diff --git a/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/0bebb7e8-7094-4aab-a136-ac9635434593_origin.pdf b/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/0bebb7e8-7094-4aab-a136-ac9635434593_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e1474254671b20aa8c969889c682d9ea3bf345fd --- /dev/null +++ b/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/0bebb7e8-7094-4aab-a136-ac9635434593_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4cbc635fd687902e1212dedf5896d24db9f21bd06b3be276619c03a7f58bc3de +size 5992028 diff --git a/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/full.md b/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9ad7c0c779a81b26799a3718870157dcafeef7e3 --- /dev/null +++ b/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/full.md @@ -0,0 +1,307 @@ +# TokenUnify: Scaling Up Autoregressive Pretraining for Neuron Segmentation + +Yinda Chen $^{1,2*}$ Haoyuan Shi $^{1,2*}$ Xiaoyu Liu $^{1}$ Te Shi $^{2}$ Ruobing Zhang $^{3,2}$ Dong Liu $^{1}$ Zhiwei Xiong $^{1,2\dagger}$ Feng Wu $^{1,2‡}$ + +$^{1}$ MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China + $^{2}$ Institute of Artificial Intelligence, Hefei Comprehensive National Science Center + $^{3}$ Institute for Brain and Intelligence, Fudan University +{cyd0806, haoyuan.shi}@mail.ustc.edu.cn, {zwxiong, fengwu}@ustc.edu.cn + +# Abstract + +Neuron segmentation from electron microscopy (EM) volumes is crucial for understanding brain circuits, yet the complex neuronal structures in high-resolution EM images present significant challenges. EM data exhibits unique characteristics including high noise levels, anisotropic voxel dimensions, and ultra-long spatial dependencies that make traditional vision models inadequate. Inspired by autoregressive pretraining in language models, we propose TokenUnify, a hierarchical predictive coding framework that captures multi-scale dependencies through three complementary learning objectives. TokenUnify integrates random token prediction, next-token prediction, and next-all token prediction to create a comprehensive representational space with emergent properties. From an information-theoretic perspective, these three tasks are complementary and provide optimal coverage of visual data structure, with our approach reducing autoregressive error accumulation from $O(K)$ to $O(\sqrt{K})$ for sequences of length $K$ . We also introduce a large-scale EM dataset with 1.2 billion annotated voxels, offering ideal long-sequence visual data with spatial continuity. Leveraging the Mamba architecture's linear-time sequence modeling capabilities, TokenUnify achieves a $44\%$ performance improvement on downstream neuron segmentation and outperforms MAE by $25\%$ . Our approach demonstrates superior scaling properties as model size increases, effectively bridging the gap between pretraining strategies for language and vision models. Code is available at https://github.com/ydchen0806/TokenUnify. + +# 1. Introduction + +Understanding neural circuits is a central goal in neuroscience. Electron microscopy (EM) imaging enables reconstruction of neurons and synaptic connections at high resolution [31, 33]. However, EM neuron segmentation presents significant challenges due to three key properties: (1) high noise levels from electron beam interactions, (2) anisotropic voxel dimensions with coarser z-axis resolution, and (3) long-range spatial continuity spanning thousands of patches. Current methods struggle to capture global context and long-range dependencies [8, 21, 30, 36]. Despite success in natural language processing [28], similar advancements for long-sequence visual data remain largely unexplored. + +Large language models (LLMs) show remarkable scaling capabilities [1, 39, 40] through effective autoregressive modeling. However, visual data has more complex structures than text. Contrastive methods like DINO v2 [34] and masked reconstruction like MAE [8, 20] demonstrate strong representation learning but lack favorable scaling laws [37]. Masked prediction faces theoretical limitations where estimation error scales as $O(\sqrt{s \log p / n})$ for sparsity $s$ , dimensionality $p$ , and sample size $n$ . Autoregressive approaches like AIM [14] and LVM [4] attempt to bridge this gap, but error accumulation $O(K)$ with sequence length $K$ and computational complexity hinder their application to high-resolution EM images [3]. + +This paper tackles visual autoregressive learning challenges through hierarchical predictive coding. Let $X = \{x_{1}, x_{2}, \ldots, x_{K}\}$ represent visual tokens from an EM image, where $\mathcal{M} \subset \{1, \ldots, K\}$ denotes masked indices and $\mathcal{M}^{c}$ denotes unmasked tokens. We recognize that three prediction tasks capture complementary visual structure aspects: random token prediction learns noise-robust spatial patterns, next-token prediction captures sequential de + +![](images/0f011e02d13fd4c507b30d39ab6ab8bfed21cdff1d57f560c39f9caf037286de.jpg) +Figure 1. TokenUnify divides the 3D EM image into non-overlapping patches, which are tokenized into a sequence of K tokens. Three complementary tasks are performed for rich 3D image representations: (a) random token prediction for spatial pattern learning, (b) next token prediction for sequential dependencies, and (c) next-all token prediction for global structure modeling. + +pencies, and next-all token prediction models global context. Under regularity conditions $(\mathrm{AR}(\infty)$ process with summable coefficients), autoregressive models achieve asymptotically optimal prediction error $O(\sigma^2)$ as capacity increases, where $\sigma^2$ is irreducible noise. + +We propose TokenUnify, a unified framework that integrates these three predictive tasks within a coherent information-theoretic formulation (Fig. 1). This creates a comprehensive representational space where random token prediction provides local robustness, next-token prediction ensures structural continuity, and next-all token prediction captures global morphology, enabling simultaneous encoding of fine-grained details and global relationships unachievable with single objectives. Our approach reduces cumulative errors from $O(K)$ to $O(\sqrt{K})$ by distributing prediction errors across multiple positions rather than sequential accumulation, analogous to central limit theorem scaling. To handle massive high-resolution EM data efficiently, we leverage the Mamba architecture, which reduces computational complexity from quadratic to linear while maintaining superior performance compared to Transformers. We validate our approach on a large-scale EM dataset collected from mouse brain slices, featuring thousands of continuous tokens that ensure spatial continuity crucial for understanding complex neuronal structures. Through comprehensive annotation of six functional brain regions totaling 1.2 billion voxels, we create one of the largest manually annotated EM neuron segmentation datasets to date. + +Comprehensive evaluation demonstrates that TokenUnify achieves a $44\%$ improvement on downstream neuron segmentation and outperforms MAE [20] by $25\%$ in pretraining performance. Furthermore, TokenUnify demonstrates superior scaling properties, offering a promising paradigm for pretraining large-scale visual models. + +Our contributions can be summarized as follows: + +1. We introduce a novel hierarchical predictive coding framework that unifies three distinct visual structure perspectives—random token prediction for spatial robustness, next-token prediction for sequential continuity, and + +next-all token prediction for global context—within a coherent information-theoretic formulation. Our theoretical analysis (detailed in supplementary material) demonstrates why these prediction tasks are complementary and together provide optimal coverage of visual data structure, while reducing autoregressive error accumulation from $O(K)$ to $O(\sqrt{K})$ . + +2. We achieve a $44\%$ performance improvement on downstream neuron segmentation tasks and realize the first billion-parameter Mamba network for visual autoregression, demonstrating both effectiveness and computational efficiency in processing long-sequence visual data. Our approach outperforms existing methods including MAE by $25\%$ while exhibiting favorable scaling properties as model size increases. +3. We construct one of the largest EM neuron segmentation datasets with 1.2 billion finely annotated voxels across six functional brain regions, providing an ideal testbed for long-sequence visual modeling. This comprehensive dataset offers spatially continuous visual data with thousands of tokens per volume, enabling rigorous validation of autoregressive methods for biological image analysis. + +# 2. Related Work + +Neuron Segmentation. In the field of EM image analysis, neuron instance segmentation has emerged as a vital task. Turaga et al. [41] pioneered the use of convolutional neural networks (CNNs) for affinity generation, followed by clustering of voxels in the affinity graph using post-processing techniques such as waterz [15] and LMC [5]. Since then, affinity-based methods have seen significant advancements. Funke et al. [15] introduced the MALIS loss [6] during training to ensure topologically accurate segmentation. Recent work by Huang et al. [21] proposed an embedding pyramid module to capture affinity information across different scales. Liu et al. [29] combined both affinity and embedding features with graph neural networks (GNNs) to enhance the segmentation of adjacent objects in the feature + +space. Furthermore, Liu [30] introduced a knowledge distillation framework, transferring knowledge from 3D models to 2D models, which resulted in improved performance and efficiency in neuron segmentation. Despite these advances, mamba-based architectures [17], have yet to be fully explored for the anisotropic resolution characteristics of EM imaging. In this paper, we enhance the Segmamba decoder [43] to better process the anisotropic features of EM data using an affinity-based approach. + +Self-supervised Pretraining. Self-supervised pretraining has been a cornerstone for enhancing model representation capabilities. Approaches based on contrastive learning for representation extraction [9, 19, 24, 44] and masked reconstruction methods [8, 10, 11, 13, 20, 27] have shown promise. However, current vision models have not exhibited the same sublinear scaling laws as language models. To address this issue, some tasks have adopted autoregressive pre-training paradigms similar to those used in language models [4, 7, 14], though the training costs remain a significant concern. In this paper, we explore the potential of long visual token autoregressive pre-training and introduce a collaborative training scheme for long token prediction. Our approach aims to balance the scaling laws and training costs, demonstrating improvements in fine-grained visual tasks. + +# 3. Theoretical Motivations + +This section provides the theoretical foundations that motivate our approach. Detailed derivations and proofs are provided in the supplementary material. + +Notation. Let $\mathbf{X} = \{x_{1}, x_{2}, \ldots, x_{K}\}$ represent a sequence of visual tokens extracted from an EM image, where $\mathcal{M} \subset \{1, \ldots, K\}$ denotes masked token indices, $\mathcal{M}^c$ denotes unmasked tokens, and $x_{ 0$ is a small constant, and $c > 0$ is the magnitude of negative curvature. This geometric property enables efficient encoding of both fine-grained details (in near-flat regions) and complex hierarchical structures (in hyperbolic-like regions). + +# 4. Method + +# 4.1. Overview + +The workflow of our method is shown in Fig. 2. Our approach comprises two stages: (1) hierarchical predictive pre-training, where we learn a generic visual representation $f_{\theta_1}(\cdot)$ from unlabeled data $\mathbf{X}$ by integrating three complementary prediction tasks that capture both local and global dependencies; and (2) task-specific fine-tuning, where we adapt the pre-trained model to downstream tasks using both $\mathbf{X}$ and labels $Y$ through a task-specific model $g_{\theta_2}(\cdot)$ initialized with $\theta_1$ . For EM neuron segmentation, we process input volumes $\mathbf{X} = \{X^{(1)},\dots,X^{(T)}\}$ , where each $X^{(t)} \in \mathbb{R}^{D\times H\times W}$ represents a 3D image, by partitioning them into smaller patches $x \in \mathbb{R}^{D' \times H' \times W'}$ and leveraging Mamba's efficient sequence modeling capabilities to handle the high-dimensional, long-sequence nature of the data. + +# 4.2. Hierarchical Predictive Coding in Token Space + +Building on our theoretical insights, we develop a hierarchical predictive coding framework that operates on the manifold of token embeddings to effectively capture the multiscale structure of high-dimensional neural image data. Our framework is motivated by the principle of maximum information preservation across complementary representational dimensions. + +![](images/8b709dcf7c3700340886284460baf6e5edc34fa59e74e831e19d0cf1dd86d173.jpg) +Figure 2. Illustration of the TokenUnify pretraining workflow. The image $\mathbf{X}$ is fed into the Tokenizer, transforming it into a long sequence of tokens $x_{i}|_{i = 1}^{K}$ . The three prediction tasks (random token, next token, and next-all token) are performed with complementary objectives. The Perceiver Resampler is employed to convert varying-size large feature maps into a few visual tokens. + +Notation. Let $\mathbf{x} = \{x_{1}, x_{2}, \ldots, x_{K}\}$ represent a sequence of visual tokens extracted from a volumetric EM image, where each $x_{i} \in \mathbb{R}^{d}$ . For random token prediction, $\mathcal{M} \subset \{1, \ldots, K\}$ denotes the set of masked token indices with $|\mathcal{M}| = \rho K$ for masking ratio $\rho$ , and $\mathcal{M}^{c} = \{1, \ldots, K\} \setminus \mathcal{M}$ denotes unmasked tokens. For sequential prediction, $x_{ 0$ is the annealing temperature controlling transition sharpness. + +Our unified objective function is formulated as: + +$$ +\mathcal {L} _ {\text {T o k e n U n i f y}} = \alpha (t) \cdot \mathcal {L} _ {\text {r a n d o m}} + \beta (t) \cdot \mathcal {L} _ {\text {n e x t}} + \gamma (t) \cdot \mathcal {L} _ {\text {n e x t - a l l}}. \tag {7} +$$ + +The curriculum implements progressive transition from local to global representation learning through: + +$$ +\mathbf {v} (t) = \left\{ \begin{array}{l l} [ 2. 0, - 1. 0, - 2. 0 ] & \text {i f} t < T _ {1} (\text {r a n d o m - d o m i n a n t}) \\ [ - 1. 0, 2. 0, - 1. 0 ] & \text {i f} T _ {1} \leq t < T _ {2} (\text {n e x t - d o m i n a n t}), \\ [ - 2. 0, - 1. 0, 2. 0 ] & \text {i f} t \geq T _ {2} (\text {n e x t - a l l - d o m i n a n t}) \end{array} \right. \tag {8} +$$ + +where $T_{1} = 0.3 \times T_{\mathrm{total}}$ and $T_{2} = 0.7 \times T_{\mathrm{total}}$ are transition thresholds. With $\tau(t) = 1.0$ , this yields dominant weights of $\sim 0.73$ for the primary task while maintaining auxiliary contributions of $\sim 0.18$ and $\sim 0.09$ to preserve multi-task synergy. + +For computational efficiency in next-all token prediction, we employ cross-attention with latent queries to aggregate sequence-wide information, enabling scalability to long token sequences without prohibitive complexity. + +This hierarchical integration constructs a multi-resolution representational manifold capturing neural tissue organization across microscopic, mesoscopic, and macroscopic scales, providing robust foundations for volumetric electron microscopy circuit reconstruction. + +# 4.3. Finetuning and Segmentation + +The segmentation network, denoted as $g_{\mathrm{seg}}(\cdot)$ , consists of an encoder $g_{e}(\cdot)$ and a decoder $g_{d}(\cdot)$ : + +$$ +g _ {\mathrm {s e g}} (x; \theta_ {s}) = g _ {d} \left(g _ {e} (x)\right), \quad \theta_ {s} = \left\{\theta_ {e}, \theta_ {d} \right\}, \tag {9} +$$ + +where $\theta_{s}$ represents the parameters of the entire segmentation network, and $\theta_{e}$ and $\theta_{d}$ are the parameters of the encoder and decoder, respectively. + +The encoder $g_{e}(\cdot)$ gradually downsamples the input volume and extracts high-level semantic features, while the decoder $g_{d}(\cdot)$ upsamples the encoded features back to the original resolution. Meanwhile, the output of each downsampling layer in the encoder is connected to the corresponding layer in the decoder via skip connections to fuse local and global multi-scale information. + +![](images/57a13d443f133743c56e008953f5650b6d532745e4b3e216ed9fd0f08fcf37f4.jpg) +(a) Sampling Area +(b) Annotation +Figure 3. Illustration of the MEC dataset annotation. (a) shows the sampling locations of six different neural regions from the MEC raw data. Each region contains $1250 \times 1250 \times 125$ voxels. (b) shows an example of the fine segmentation labels in one sampling region, which takes two experts one month to annotate. + +We adopt 3D ResUNet/ViT/Mamba as the backbone network, respectively. For the Mamba variant, we adapt SegMamba [43] for EM data, which we refer to as EMmamba, with enhanced decoder design for anisotropic EM features. + +The output of the segmentation network $\hat{y} = g_{\mathrm{seg}}(x)\in$ $\mathbb{R}^{C\times D\times H\times W}$ represents the predicted affinity map [25, 26], corresponding to the connectivity probability of each voxel in three directions. During training, the loss function for labeled samples is the mean squared error between the predicted and ground-truth affinity maps: + +$$ +\mathcal {L} _ {\mathrm {s e g}} = \frac {1}{| \mathcal {D} _ {l} |} \sum_ {i = 1} ^ {| \mathcal {D} _ {l} |} \| g _ {\mathrm {s e g}} \left(x _ {i} ^ {l}\right) - y _ {i} \| ^ {2}. \tag {10} +$$ + +During inference, for any new test sample $x_{t}$ , forward propagation through $g_{\mathrm{seg}}(x_{t})$ yields its predicted affinity map. This predicted affinity map is then post-processed using a seeded watershed algorithm and a structure-aware region agglomeration algorithm [5, 15] to obtain the final neuron instance segmentation. Detailed information on our segmentation process and the segmentation pipeline are illustrated in the supplementary material. + +# 5. Datasets and Metrics + +Datasets. For pretraining TokenUnify, we collect over 1 TB of publicly available unlabeled EM imaging data from four large-scale datasets: FAFB [35], MitoEM [42], FIB-25 [38], and Kasthuri15 [22]. These datasets span diverse organisms including Drosophila, mouse, rat, and human samples, ensuring comprehensive coverage of EM morphology variations. We sample from these datasets with equal probability to guarantee diversity in pretraining data. Additional details are provided in supplementary material. + +For downstream fine-tuning and evaluation, we employ two datasets: AC3/4 [22] and our large-scale MEC dataset. The AC3/4 dataset consists of 256/100 successive EM images with $1024 \times 1024$ spatial resolution from mouse somatosensory cortex. We partition AC4 into training (first 80 images) and validation (last 20 images), while using AC3 (first 100 images) for testing. + +
Post.MethodWafer4Wafer36-2Param. (M)
VOIM↓VOIS↓VOI↓ARAND↓VOIM↓VOIS↓VOI↓ARAND↓
LMC [5]Supervised Methods
Superhuman [23]0.33281.12581.45870.17360.15060.45880.60940.08361.478
MALA [15]0.54381.50272.03750.11150.31791.06641.38430.157084.03
PEA [21]0.33810.92761.26580.06770.27870.42790.70660.11691.480
UNETR [18]0.45041.65812.10850.26580.44780.52170.96960.2913129.1
EMmamba0.49151.29241.78390.20520.24060.41890.65950.123128.30
Superhuman [23]0.19481.96972.16440.24530.07921.16181.24100.13191.478
MALA [15]0.34162.41292.75450.25670.14481.96032.10520.197784.03
PEA [21]0.17051.59931.76980.15270.47191.12261.59450.15881.480
UNETR [18]0.17913.17153.35060.63300.09491.38581.48070.1578129.1
EMmamba0.15962.05802.21770.19730.08471.03511.11980.125328.30
LMC [5]Self-Supervised Methods
Random0.49151.29241.78390.20520.24060.41890.65950.123128.30
MAE [20]0.23251.09231.32480.09780.16290.45320.61610.0835
BYOL [16]0.25840.94531.20370.08910.17480.43680.61160.0826
dbMIM [8]0.23420.87961.11380.07420.14760.39820.54580.0782
MS-Con-EM [9]0.24830.91021.15850.08130.15420.40370.55790.0794
TokenUnify0.19530.79980.99510.05090.12620.35850.48480.0650
Random0.15962.05802.21770.19730.08471.03511.11980.125328.30
MAE [20]0.14931.92162.07090.15830.08921.09271.18190.1243
BYOL [16]0.16291.83742.00030.14270.09531.07821.17350.1287
dbMIM [8]0.15731.74361.90090.09130.09121.06491.15610.1185
MS-Con-EM [9]0.16091.76921.93010.10260.09631.07031.16660.1207
TokenUnify0.14181.51031.65210.05910.08271.02761.11030.1074
+ +Table 1. Quantitative comparison of electron microscopy segmentation results on Wafer4 and Wafer36-2 datasets. Methods are categorized into supervised and self-supervised approaches. All self-supervised methods use the same EMmamba backbone for fair comparison. "Random" refers to EMmamba without any pretraining. The best results are in bold and the second best results are underlined. + +To validate algorithm performance on large-scale EM data, we collected a 2TB MEC dataset by imaging mouse somatosensory cortex, medial entorhinal cortex, and cerebral cortex regions. We selected 6 representative volumes from different neural regions (wafer4/25/26/26-2/36/36-2), each with dimensions $1250 \times 1250 \times 125$ voxels. Dense manual annotation was performed on these regions, yielding over 1.2 billion annotated voxels through six months of work by two experts, as illustrated in Fig. 3. The dataset is split into training (wafer25/26/26-2/36), validation (wafer4), and testing (wafer36-2) sets. + +Metrics. To evaluate the performance of neuron segmentation [12, 45], we employ two widely-used metrics: Variation of Information (VOI) [32] and Adjusted Rand Index (ARAND) [2]. These metrics quantify the agreement between the predicted segmentation and the ground truth from different perspectives. Detailed descriptions of these metrics can be found in the supplementary material. + +# 6. Experiment + +We evaluate our proposed method on both MEC and AC3/4 datasets. Implementation details are provided in the sup + +plementary material. We compare our method with both supervised and self-supervised approaches. For supervised methods, we include Superhuman [23], MALA [15], PEA [21], and UNETR [18]. For self-supervised methods, we evaluate two categories: (1) general-purpose visual representation learning methods including MAE [20] and BYOL [16], which were originally designed for natural images but adapted to EM data; and (2) EM-specific self-supervised methods including dbMIM [8] and MS-Con-EM [9], which were specifically designed for electron microscopy images. All self-supervised methods utilize the same EMmamba backbone for fair comparison. + +Results on MEC Dataset. Table 1 presents the quantitative results on the large-scale MEC dataset. Compared to models with similar parameter counts, on the Wafer4 volume under the waterz post-processing, our TokenUnify approach achieves approximately a $25\%$ performance improvement over MAE and a $44\%$ improvement compared to direct training without pretraining. Visualization results in Fig. 4 and Fig. 5 further demonstrate that our approach significantly outperforms other methods, particularly in challenging regions with complex neuronal structures. + +![](images/1ca765afeba898f22e39affead4bf0485ec10b71e7272a27d207f57ccf5b1281.jpg) + +![](images/12dd48e2cc5a9a338d09f9f6b46beb755632fecd9e76de421ddf68d95efe5aec.jpg) +Figure 4. 2D visualization results for two slices from the MEC dataset: wafer 4 and wafer 36-2. The left side displays the EM raw images, while the right side presents the affinity and segmentation results. Red boxes indicate the regions of mis-segmentation. + +Results on AC3/4 Datasets. We also evaluate performance on the smaller AC3/4 dataset, which contains only about 1/10 of the labeled pixels compared to MEC. As shown in Table 2, even in this low-data scenario, the Mamba architecture combined with TokenUnify pretraining achieves performance comparable to state-of-the-art supervised methods like PEA. Additionally, it demonstrates a $11\%$ performance improvement over the MAE pretraining approach. This highlights the effectiveness of TokenUnify in scenarios with limited labeled data. + +Scaling Law Analysis. To investigate how our method scales with model size, we evaluated various initialization and training strategies by scaling model parameters from 100M to 1B. Results in Fig. 6 show that all methods ex + +![](images/a0d5089d7f9b9e2100df2021ffa21be50419adf742f75f8fc5a586d02765894c.jpg) +Figure 5. 3D visualization results on two MEC datasets: wafer 4 and wafer 36-2. The red arrows indicate over-merged structures, and the orange arrows highlight over-split structures. +Figure 6. Performance evaluation of models with 100M, 200M, 500M, and 1B parameters. The models were trained for 100 epochs on the MEC dataset. + +
MethodVOIM↓VOIS↓VOI↓ARAND↓Param.(M)
Supervised Methods
Superhuman [23]0.48820.65631.14450.17481.478
MALA [15]0.45710.67671.13380.166484.02
PEA [21]0.55220.49801.05020.20931.480
UNETR [18]0.77990.73991.51980.2411129.1
EMmamba0.93780.86291.80070.284028.30
Self-Supervised Methods
Random0.93780.86291.80070.2840
MAE [20]0.31670.79631.11310.1050
BYOL [16]0.38420.76211.14630.152728.30
dbMIM [8]0.46230.62791.09020.1432
MS-Con-EM [9]0.47850.65381.13230.1498
TokenUnify0.44790.54390.99180.1366
+ +Table 2. Quantitative comparison of segmentation results on AC3/4 datasets with waterz [15] post-processing algorithm. All self-supervised methods use the same EMmamba backbone. "Random" refers to EMmamba without any pretraining. The best results are in bold and the second best results are underlined. + +
Pretraining StrategyVOI ↓ARAND ↓
RandomNextNext-all
✓1.26800.0862
✓4.04180.4416
✓✓1.13000.0692
✓✓1.19070.1203
✓✓✓0.99510.0509
+ +Table 3. Ablation for different pretraining strategies on wafer4. + +
Finetuning ModuleVOI ↓ARAND ↓
MambaEncoderDecoder
✓1.13620.0782
✓1.55560.1370
✓1.52950.1212
✓✓1.10650.0629
✓✓✓0.99510.0509
+ +Table 4. Ablation for the fine-tuning schemes on wafer4. + +hibit performance improvements after pretraining. In contrast, TokenUnify demonstrates strong scaling properties, consistently outperforming other methods across both small and large model sizes. Notably, the Mamba architecture achieves competitive segmentation performance with fewer parameters compared to Transformer, confirming its efficiency for processing long sequence data in connectomics. + +# 6.1. Ablation Studies + +We conduct ablation experiments to analyze the key components of our approach. As shown in Table 3 and Table 4, we examine two critical aspects: + +Ablation on Token Prediction Strategies. We systematically evaluate different token prediction combinations dur + +ing pretraining on the wafer4 dataset (Table 3). The full combination (Random + Next + Next-all) achieves optimal performance $(\mathrm{VOI} = 0.9951)$ , significantly outperforming individual strategies and demonstrating substantial improvements over baseline approaches. Next-token prediction alone performs poorly $(\mathrm{VOI} = 4.0418)$ , confirming that conventional autoregressive pretraining is insufficient for visual tasks. Unlike sequential text, visual data requires global spatial understanding that pure autoregressive modeling cannot capture effectively. Random token prediction provides better spatial initialization by encouraging spatial consistency, while the strategic combination of multiple prediction objectives yields more robust and generalizable representations for segmentation. These results validate our hypothesis that diverse pretraining tasks are essential for effective visual representation learning. + +Fine-tuning Components. We also investigate which network components benefit most from fine-tuning (Table 4). Our architecture consists of three main parts: Mamba blocks (sequence modeling), encoder (downsampling), and decoder (upsampling). Fine-tuning only the Mamba blocks yields substantially better results $(\mathrm{VOI} = 1.1362)$ than fine-tuning only the encoder $(\mathrm{VOI} = 1.5556)$ or decoder $(\mathrm{VOI} = 1.5295)$ , demonstrating the critical role of sequence modeling in connectomics segmentation. For resource-constrained scenarios, prioritizing the fine-tuning of Mamba blocks provides an efficient balance between performance and computational cost. + +# 7. Conclusion + +We propose TokenUnify, a framework that unifies random token prediction, next-token prediction, and next-all token prediction to capture multi-scale visual structures. Our analysis shows this approach mitigates cumulative errors in autoregressive modeling. To validate TokenUnify, we introduce a large-scale EM neuron segmentation dataset with 1.2 billion labeled voxels across six mouse brain regions. Experiments demonstrate a $44\%$ improvement over training from scratch and a $25\%$ gain over MAE pretraining. Scaling analysis further highlights its efficiency with increasing model size and data volume. By releasing our dataset and implementation, we aim to advance biological imaging and inspire future research in visual representation learning. + +# Acknowledgments + +This work was supported by the National Natural Science Foundation of China under Grant 624B2137. We thank the staff at the Institute of Artificial Intelligence, Hefei Comprehensive National Science Center for their assistance with biological specimen preparation, imaging, and annotation. We are grateful to the anonymous reviewers for their thorough reviews and constructive suggestions. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Ignacio Arganda-Carreras, Srinivas C Turaga, Daniel R Berger, Dan Ciresan, Alessandro Giusti, Luca M Gambardella, Jürgen Schmidhuber, Dmitry Laptev, Sarvesh Dwivedi, Joachim M Buhmann, et al. Crowdsourcing the creation of image segmentation algorithms for connectomics. Frontiers in Neuroanatomy, 9:152591, 2015. 6 +[3] Gregor Bachmann and Vaishnavh Nagarajan. The pitfalls of next-token prediction. arXiv preprint arXiv:2403.06963, 2024. 1 +[4] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785, 2023. 1, 3 +[5] Thorsten Beier, Constantin Pape, Nasim Rahaman, Timo Prange, Stuart Berg, Davi D Bock, Albert Cardona, Graham W Knott, Stephen M Plaza, Louis K Scheffer, et al. Multicut brings automated neurite segmentation closer to human performance. Nature Methods, 14(2):101-102, 2017. 2, 5, 6 +[6] Kevin Briggman, Winfried Denk, Sebastian Seung, Moritz Helmstaedter, and Srinivas C Turaga. Maximin affinity learning of image segmentation. In NeurIPS, 2009. 2 +[7] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 3 +[8] Yinda Chen, Wei Huang, Shenglong Zhou, Qi Chen, and Zhiwei Xiong. Self-supervised neuron segmentation with multi-agent reinforcement learning. In IJCAI, 2023. 1, 3, 6, 8 +[9] Yinda Chen, Wei Huang, Xiaoyu Liu, Shiyu Deng, Qi Chen, and Zhiwei Xiong. Learning multiscale consistency for self-supervised electron microscopy instance segmentation. In ICASSP, 2024. 3, 6, 8 +[10] Zhimin Chen, Yingwei Li, Longlong Jing, Liang Yang, and Bing Li. Point cloud self-supervised learning via 3d to multiview masked autoencoder. arXiv preprint arXiv:2311.10887, 2023. 3 +[11] Zhimin Chen, Longlong Jing, Yingwei Li, and Bing Li. Bridging the domain gap: Self-supervised 3d scene understanding with foundation models. In NeurIPS, 2024. 3 +[12] Bo Dang, Wenchao Zhao, Yufeng Li, Danqing Ma, Qixuan Yu, and Elly Yijun Zhu. Real-time pill identification for the visually impaired using deep learning. arXiv preprint arXiv:2405.05983, 2024. 6 +[13] Zhiyuan Ding, Qi Dong, Haote Xu, Chenxin Li, Xinghao Ding, and Yue Huang. Unsupervised anomaly segmentation for brain lesions using dual semantic-manifold reconstruction. In ICONIP, 2022. 3 +[14] Alaaeldin El-Nouby, Michal Klein, Shuangfei Zhai, Miguel Angel Bautista, Alexander Toshev, Vaishaal Shankar, + +Joshua M Susskind, and Armand Joulin. Scalable pretraining of large autoregressive image models. arXiv preprint arXiv:2401.08541, 2024. 1, 3 +[15] Jan Funke, Fabian Tschopp, William Grisaitis, Arlo Sheridan, Chandan Singh, Stephan Saalfeld, and Srinivas C Turaga. Large scale image segmentation with structured loss based deep learning for connectome reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(7):1669-1680, 2018. 2, 5, 6, 8 +[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020. 6, 8 +[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3 +[18] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In WACV, 2022. 6, 8 +[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3 +[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 2, 3, 6, 8 +[21] Wei Huang, Shiyu Deng, Chang Chen, Xueyang Fu, and Zhiwei Xiong. Learning to model pixel-embedded affinity for homogeneous instance segmentation. In AAAI, 2022. 1, 2, 6, 8 +[22] Narayanan Kasthuri, Kenneth Jeffrey Hayworth, Daniel Raimund Berger, Richard Lee Schalek, José Angel Conchello, Seymour Knowles-Barley, Dongil Lee, Amelio Vázquez-Reina, Verena Kaynig, Thouis Raymond Jones, et al. Saturated reconstruction of a volume of neocortex. Cell, 162(3):648-661, 2015. 5 +[23] Kisuk Lee, Jonathan Zung, Peter Li, Viren Jain, and H Sebastian Seung. Superhuman accuracy on the snemi3d connectomics challenge. arXiv preprint arXiv:1706.00120, 2017. 6, 8 +[24] Chenxin Li, Yunlong Zhang, Zhehan Liang, Wenao Ma, Yue Huang, and Xinghao Ding. Consistent posterior distributions under vessel-mixing: A regularization for cross-domain retinal artery/vein classification. In ICIP, 2021. 3 +[25] Panfeng Li, Youzuo Lin, and Emily Schultz-Fellenz. Contextual hourglass network for semantic segmentation of high resolution aerial imagery. arXiv preprint arXiv:1810.12813, 2018. 5 +[26] Panfeng Li, Mohamed Abouelenien, and Rada Mihalcea. Deception detection from linguistic and physiological data streams using bimodal convolutional neural networks. arXiv preprint arXiv:2311.10944, 2023. 5 +[27] Tianhong Li, Huiwen Chang, Shlok Mishra, Han Zhang, Dina Katabi, and Dilip Krishnan. Mage: Masked generative encoder to unify representation learning and image synthesis. In CVPR, 2023. 3 + +[28] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9): 1-35, 2023. 1 +[29] Xiaoyu Liu, Wei Huang, Yueyi Zhang, and Zhiwei Xiong. Biological instance segmentation with a superpixel-guided graph. In IJCAI, 2022. 2 +[30] Xiaoyu Liu, Miaomiao Cai, Yinda Chen, Yueyi Zhang, Te Shi, Ruobing Zhang, Xuejin Chen, and Zhiwei Xiong. Cross-dimension affinity distillation for 3d em neuron segmentation. In CVPR, 2024. 1, 3 +[31] Christopher W Lynn and Danielle S Bassett. The physics of brain network structure, function and control. Nature Reviews Physics, 1(5):318-332, 2019. 1 +[32] Juan Nunez-Iglesias, Ryan Kennedy, Toufiq Parag, Jianbo Shi, and Dmitri B Chklovskii. Machine learning of hierarchical clustering to segment 2d and 3d images. PLoS ONE, 8(8):e71715, 2013. 6 +[33] Nobuhiko Ohno, Mitsuhiko Katoh, Yurika Saitoh, Sei Saitoh, and Shinichi Ohno. Three-dimensional volume imaging with electron microscopy toward connectome. Microscopy, 64(1):17-26, 2015. 1 +[34] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2023. 1 +[35] Philipp Schlegel, Alexander S Bates, Tejal Parag, Gregory SXE Jefferis, and Davi D Bock. Automatic detection of synaptic partners in a whole-brain drosophila em dataset. Nature Methods, 18(8):877-884, 2021. 5 +[36] Arlo Sheridan, Tri M Nguyen, Diptodip Deb, Wei-Chung Allen Lee, Stephan Saalfeld, Srinivas C Turaga, Uri Manor, and Jan Funke. Local shape descriptors for neuron segmentation. Nature Methods, 20(2):295-303, 2023. 1 +[37] Mannat Singh, Quentin Duval, Kalyan Vasudev Alwala, Haoqi Fan, Vaibhav Aggarwal, Aaron Adcock, Armand Joulin, Piotr Dollar, Christoph Feichtenhofer, Ross Girshick, et al. The effectiveness of mae pre-pretraining for billionscale pretraining. In ICCV, 2023. 1 +[38] Shin-ya Takemura, Yoshinori Aso, Toshihide Hige, Allan Wong, Zhiyuan Lu, C Shan Xu, Patricia K Rivlin, Harald Hess, Ting Zhao, Toufiq Parag, et al. A connectome of a learning and memory center in the adult drosophila brain. eLife, 6:e26975, 2017. 5 +[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[40] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1 + +[41] Srinivas C Turaga, Joseph F Murray, Viren Jain, Fabian Roth, Moritz Helmstaedter, Kevin Briggman, Winfried Denk, and H Sebastian Seung. Convolutional networks can learn to generate affinity graphs for image segmentation. Neural Computation, 22(2):511-538, 2010. 2 +[42] Donglai Wei, Zudi Lin, Daniel Franco-Barranco, Nils Wendt, Xingyu Liu, Wenjie Yin, Xin Huang, Aarush Gupta, Won-Dong Jang, Xueying Wang, et al. Mitoem dataset: Large-scale 3d mitochondria instance segmentation from em images. In MICCAI, 2020. 5 +[43] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In MICCAI, 2024. 3, 5 +[44] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021. 3 +[45] Jingyu Zhang, Ao Xiang, Yu Cheng, Qin Yang, and Liyang Wang. Research on detection of floating objects in river and lake based on ai intelligent image recognition. arXiv preprint arXiv:2404.06883, 2024. 6 \ No newline at end of file diff --git a/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/images.zip b/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2fe69318aaa45b60a7bfa7b6c2246adc35d92b4c --- /dev/null +++ b/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f141254fa7dc4c1a275fc02498f0a3d8a273b4dc90f4d68a663f36d7ac7cbda2 +size 872096 diff --git a/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/layout.json b/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..84a19545d5402ab579458ac3bd84f0c5f478e89f --- /dev/null +++ b/2025/TokenUnify_ Scaling Up Autoregressive Pretraining for Neuron Segmentation/layout.json @@ -0,0 +1,8809 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 544, + 121 + ], + "type": "text", + "content": "TokenUnify: Scaling Up Autoregressive Pretraining for Neuron Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "spans": [ + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "text", + "content": "Yinda Chen" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "text", + "content": " Haoyuan Shi" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "text", + "content": " Xiaoyu Liu" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "text", + "content": " Te Shi" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "text", + "content": " Ruobing Zhang" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{3,2}" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "text", + "content": " Dong Liu" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "text", + "content": " Zhiwei Xiong" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "text", + "content": " Feng Wu" + }, + { + "bbox": [ + 154, + 142, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{1,2‡}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 184, + 504, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 184, + 504, + 257 + ], + "spans": [ + { + "bbox": [ + 107, + 184, + 504, + 257 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 184, + 504, + 257 + ], + "type": "text", + "content": "MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, University of Science and Technology of China \n" + }, + { + "bbox": [ + 107, + 184, + 504, + 257 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 184, + 504, + 257 + ], + "type": "text", + "content": "Institute of Artificial Intelligence, Hefei Comprehensive National Science Center \n" + }, + { + "bbox": [ + 107, + 184, + 504, + 257 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 184, + 504, + 257 + ], + "type": "text", + "content": "Institute for Brain and Intelligence, Fudan University \n{cyd0806, haoyuan.shi}@mail.ustc.edu.cn, {zwxiong, fengwu}@ustc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "spans": [ + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "type": "text", + "content": "Neuron segmentation from electron microscopy (EM) volumes is crucial for understanding brain circuits, yet the complex neuronal structures in high-resolution EM images present significant challenges. EM data exhibits unique characteristics including high noise levels, anisotropic voxel dimensions, and ultra-long spatial dependencies that make traditional vision models inadequate. Inspired by autoregressive pretraining in language models, we propose TokenUnify, a hierarchical predictive coding framework that captures multi-scale dependencies through three complementary learning objectives. TokenUnify integrates random token prediction, next-token prediction, and next-all token prediction to create a comprehensive representational space with emergent properties. From an information-theoretic perspective, these three tasks are complementary and provide optimal coverage of visual data structure, with our approach reducing autoregressive error accumulation from " + }, + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "type": "inline_equation", + "content": "O(K)" + }, + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "type": "inline_equation", + "content": "O(\\sqrt{K})" + }, + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "type": "text", + "content": " for sequences of length " + }, + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "type": "text", + "content": ". We also introduce a large-scale EM dataset with 1.2 billion annotated voxels, offering ideal long-sequence visual data with spatial continuity. Leveraging the Mamba architecture's linear-time sequence modeling capabilities, TokenUnify achieves a " + }, + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "type": "text", + "content": " performance improvement on downstream neuron segmentation and outperforms MAE by " + }, + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 308, + 297, + 656 + ], + "type": "text", + "content": ". Our approach demonstrates superior scaling properties as model size increases, effectively bridging the gap between pretraining strategies for language and vision models. Code is available at https://github.com/ydchen0806/TokenUnify." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 282, + 394, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 282, + 394, + 295 + ], + "spans": [ + { + "bbox": [ + 315, + 282, + 394, + 295 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 302, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 458 + ], + "type": "text", + "content": "Understanding neural circuits is a central goal in neuroscience. Electron microscopy (EM) imaging enables reconstruction of neurons and synaptic connections at high resolution [31, 33]. However, EM neuron segmentation presents significant challenges due to three key properties: (1) high noise levels from electron beam interactions, (2) anisotropic voxel dimensions with coarser z-axis resolution, and (3) long-range spatial continuity spanning thousands of patches. Current methods struggle to capture global context and long-range dependencies [8, 21, 30, 36]. Despite success in natural language processing [28], similar advancements for long-sequence visual data remain largely unexplored." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "text", + "content": "Large language models (LLMs) show remarkable scaling capabilities [1, 39, 40] through effective autoregressive modeling. However, visual data has more complex structures than text. Contrastive methods like DINO v2 [34] and masked reconstruction like MAE [8, 20] demonstrate strong representation learning but lack favorable scaling laws [37]. Masked prediction faces theoretical limitations where estimation error scales as " + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "inline_equation", + "content": "O(\\sqrt{s \\log p / n})" + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "text", + "content": " for sparsity " + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "text", + "content": ", dimensionality " + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "text", + "content": ", and sample size " + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "text", + "content": ". Autoregressive approaches like AIM [14] and LVM [4] attempt to bridge this gap, but error accumulation " + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "inline_equation", + "content": "O(K)" + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "text", + "content": " with sequence length " + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 458, + 556, + 613 + ], + "type": "text", + "content": " and computational complexity hinder their application to high-resolution EM images [3]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 613, + 556, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 556, + 710 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 556, + 710 + ], + "type": "text", + "content": "This paper tackles visual autoregressive learning challenges through hierarchical predictive coding. Let " + }, + { + "bbox": [ + 313, + 613, + 556, + 710 + ], + "type": "inline_equation", + "content": "X = \\{x_{1}, x_{2}, \\ldots, x_{K}\\}" + }, + { + "bbox": [ + 313, + 613, + 556, + 710 + ], + "type": "text", + "content": " represent visual tokens from an EM image, where " + }, + { + "bbox": [ + 313, + 613, + 556, + 710 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\subset \\{1, \\ldots, K\\}" + }, + { + "bbox": [ + 313, + 613, + 556, + 710 + ], + "type": "text", + "content": " denotes masked indices and " + }, + { + "bbox": [ + 313, + 613, + 556, + 710 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{c}" + }, + { + "bbox": [ + 313, + 613, + 556, + 710 + ], + "type": "text", + "content": " denotes unmasked tokens. We recognize that three prediction tasks capture complementary visual structure aspects: random token prediction learns noise-robust spatial patterns, next-token prediction captures sequential de" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 673, + 147, + 682 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 673, + 147, + 682 + ], + "spans": [ + { + "bbox": [ + 66, + 673, + 147, + 682 + ], + "type": "text", + "content": "*Equal Contribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 682, + 120, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 682, + 120, + 693 + ], + "spans": [ + { + "bbox": [ + 67, + 682, + 120, + 693 + ], + "type": "text", + "content": "†Project Leader" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 693, + 145, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 145, + 702 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 145, + 702 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 693, + 145, + 702 + ], + "type": "text", + "content": " Corresponding Author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 746, + 318, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 746, + 318, + 756 + ], + "spans": [ + { + "bbox": [ + 293, + 746, + 318, + 756 + ], + "type": "text", + "content": "13604" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 68, + 532, + 185 + ], + "blocks": [ + { + "bbox": [ + 82, + 68, + 532, + 185 + ], + "lines": [ + { + "bbox": [ + 82, + 68, + 532, + 185 + ], + "spans": [ + { + "bbox": [ + 82, + 68, + 532, + 185 + ], + "type": "image", + "image_path": "0f011e02d13fd4c507b30d39ab6ab8bfed21cdff1d57f560c39f9caf037286de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 190, + 555, + 225 + ], + "lines": [ + { + "bbox": [ + 54, + 190, + 555, + 225 + ], + "spans": [ + { + "bbox": [ + 54, + 190, + 555, + 225 + ], + "type": "text", + "content": "Figure 1. TokenUnify divides the 3D EM image into non-overlapping patches, which are tokenized into a sequence of K tokens. Three complementary tasks are performed for rich 3D image representations: (a) random token prediction for spatial pattern learning, (b) next token prediction for sequential dependencies, and (c) next-all token prediction for global structure modeling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 245, + 294, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 245, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 54, + 245, + 294, + 304 + ], + "type": "text", + "content": "pencies, and next-all token prediction models global context. Under regularity conditions " + }, + { + "bbox": [ + 54, + 245, + 294, + 304 + ], + "type": "inline_equation", + "content": "(\\mathrm{AR}(\\infty)" + }, + { + "bbox": [ + 54, + 245, + 294, + 304 + ], + "type": "text", + "content": " process with summable coefficients), autoregressive models achieve asymptotically optimal prediction error " + }, + { + "bbox": [ + 54, + 245, + 294, + 304 + ], + "type": "inline_equation", + "content": "O(\\sigma^2)" + }, + { + "bbox": [ + 54, + 245, + 294, + 304 + ], + "type": "text", + "content": " as capacity increases, where " + }, + { + "bbox": [ + 54, + 245, + 294, + 304 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 54, + 245, + 294, + 304 + ], + "type": "text", + "content": " is irreducible noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 305, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 305, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 305, + 295, + 579 + ], + "type": "text", + "content": "We propose TokenUnify, a unified framework that integrates these three predictive tasks within a coherent information-theoretic formulation (Fig. 1). This creates a comprehensive representational space where random token prediction provides local robustness, next-token prediction ensures structural continuity, and next-all token prediction captures global morphology, enabling simultaneous encoding of fine-grained details and global relationships unachievable with single objectives. Our approach reduces cumulative errors from " + }, + { + "bbox": [ + 57, + 305, + 295, + 579 + ], + "type": "inline_equation", + "content": "O(K)" + }, + { + "bbox": [ + 57, + 305, + 295, + 579 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 57, + 305, + 295, + 579 + ], + "type": "inline_equation", + "content": "O(\\sqrt{K})" + }, + { + "bbox": [ + 57, + 305, + 295, + 579 + ], + "type": "text", + "content": " by distributing prediction errors across multiple positions rather than sequential accumulation, analogous to central limit theorem scaling. To handle massive high-resolution EM data efficiently, we leverage the Mamba architecture, which reduces computational complexity from quadratic to linear while maintaining superior performance compared to Transformers. We validate our approach on a large-scale EM dataset collected from mouse brain slices, featuring thousands of continuous tokens that ensure spatial continuity crucial for understanding complex neuronal structures. Through comprehensive annotation of six functional brain regions totaling 1.2 billion voxels, we create one of the largest manually annotated EM neuron segmentation datasets to date." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 580, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 651 + ], + "type": "text", + "content": "Comprehensive evaluation demonstrates that TokenUnify achieves a " + }, + { + "bbox": [ + 55, + 580, + 295, + 651 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 55, + 580, + 295, + 651 + ], + "type": "text", + "content": " improvement on downstream neuron segmentation and outperforms MAE [20] by " + }, + { + "bbox": [ + 55, + 580, + 295, + 651 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 580, + 295, + 651 + ], + "type": "text", + "content": " in pretraining performance. Furthermore, TokenUnify demonstrates superior scaling properties, offering a promising paradigm for pretraining large-scale visual models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 652, + 266, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 652, + 266, + 662 + ], + "spans": [ + { + "bbox": [ + 67, + 652, + 266, + 662 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 663, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 663, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 663, + 295, + 712 + ], + "type": "text", + "content": "1. We introduce a novel hierarchical predictive coding framework that unifies three distinct visual structure perspectives—random token prediction for spatial robustness, next-token prediction for sequential continuity, and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 325, + 245, + 553, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 245, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 325, + 245, + 553, + 328 + ], + "type": "text", + "content": "next-all token prediction for global context—within a coherent information-theoretic formulation. Our theoretical analysis (detailed in supplementary material) demonstrates why these prediction tasks are complementary and together provide optimal coverage of visual data structure, while reducing autoregressive error accumulation from " + }, + { + "bbox": [ + 325, + 245, + 553, + 328 + ], + "type": "inline_equation", + "content": "O(K)" + }, + { + "bbox": [ + 325, + 245, + 553, + 328 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 325, + 245, + 553, + 328 + ], + "type": "inline_equation", + "content": "O(\\sqrt{K})" + }, + { + "bbox": [ + 325, + 245, + 553, + 328 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 329, + 555, + 509 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 329, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 329, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 314, + 329, + 553, + 422 + ], + "type": "text", + "content": "2. We achieve a " + }, + { + "bbox": [ + 314, + 329, + 553, + 422 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 314, + 329, + 553, + 422 + ], + "type": "text", + "content": " performance improvement on downstream neuron segmentation tasks and realize the first billion-parameter Mamba network for visual autoregression, demonstrating both effectiveness and computational efficiency in processing long-sequence visual data. Our approach outperforms existing methods including MAE by " + }, + { + "bbox": [ + 314, + 329, + 553, + 422 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 314, + 329, + 553, + 422 + ], + "type": "text", + "content": " while exhibiting favorable scaling properties as model size increases." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 425, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 509 + ], + "type": "text", + "content": "3. We construct one of the largest EM neuron segmentation datasets with 1.2 billion finely annotated voxels across six functional brain regions, providing an ideal testbed for long-sequence visual modeling. This comprehensive dataset offers spatially continuous visual data with thousands of tokens per volume, enabling rigorous validation of autoregressive methods for biological image analysis." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 518, + 400, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 400, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 400, + 530 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 539, + 555, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 706 + ], + "type": "text", + "content": "Neuron Segmentation. In the field of EM image analysis, neuron instance segmentation has emerged as a vital task. Turaga et al. [41] pioneered the use of convolutional neural networks (CNNs) for affinity generation, followed by clustering of voxels in the affinity graph using post-processing techniques such as waterz [15] and LMC [5]. Since then, affinity-based methods have seen significant advancements. Funke et al. [15] introduced the MALIS loss [6] during training to ensure topologically accurate segmentation. Recent work by Huang et al. [21] proposed an embedding pyramid module to capture affinity information across different scales. Liu et al. [29] combined both affinity and embedding features with graph neural networks (GNNs) to enhance the segmentation of adjacent objects in the feature" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 746, + 317, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 746, + 317, + 755 + ], + "spans": [ + { + "bbox": [ + 294, + 746, + 317, + 755 + ], + "type": "text", + "content": "13605" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "space. Furthermore, Liu [30] introduced a knowledge distillation framework, transferring knowledge from 3D models to 2D models, which resulted in improved performance and efficiency in neuron segmentation. Despite these advances, mamba-based architectures [17], have yet to be fully explored for the anisotropic resolution characteristics of EM imaging. In this paper, we enhance the Segmamba decoder [43] to better process the anisotropic features of EM data using an affinity-based approach." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 180, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 180, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 296, + 373 + ], + "type": "text", + "content": "Self-supervised Pretraining. Self-supervised pretraining has been a cornerstone for enhancing model representation capabilities. Approaches based on contrastive learning for representation extraction [9, 19, 24, 44] and masked reconstruction methods [8, 10, 11, 13, 20, 27] have shown promise. However, current vision models have not exhibited the same sublinear scaling laws as language models. To address this issue, some tasks have adopted autoregressive pre-training paradigms similar to those used in language models [4, 7, 14], though the training costs remain a significant concern. In this paper, we explore the potential of long visual token autoregressive pre-training and introduce a collaborative training scheme for long token prediction. Our approach aims to balance the scaling laws and training costs, demonstrating improvements in fine-grained visual tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 382, + 194, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 194, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 194, + 395 + ], + "type": "text", + "content": "3. Theoretical Motivations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 403, + 295, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 438 + ], + "type": "text", + "content": "This section provides the theoretical foundations that motivate our approach. Detailed derivations and proofs are provided in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 498 + ], + "type": "text", + "content": "Notation. Let " + }, + { + "bbox": [ + 55, + 438, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\{x_{1}, x_{2}, \\ldots, x_{K}\\}" + }, + { + "bbox": [ + 55, + 438, + 295, + 498 + ], + "type": "text", + "content": " represent a sequence of visual tokens extracted from an EM image, where " + }, + { + "bbox": [ + 55, + 438, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\subset \\{1, \\ldots, K\\}" + }, + { + "bbox": [ + 55, + 438, + 295, + 498 + ], + "type": "text", + "content": " denotes masked token indices, " + }, + { + "bbox": [ + 55, + 438, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^c" + }, + { + "bbox": [ + 55, + 438, + 295, + 498 + ], + "type": "text", + "content": " denotes unmasked tokens, and " + }, + { + "bbox": [ + 55, + 438, + 295, + 498 + ], + "type": "inline_equation", + "content": "x_{ 0" + }, + { + "bbox": [ + 313, + 319, + 553, + 390 + ], + "type": "text", + "content": " is a small constant, and " + }, + { + "bbox": [ + 313, + 319, + 553, + 390 + ], + "type": "inline_equation", + "content": "c > 0" + }, + { + "bbox": [ + 313, + 319, + 553, + 390 + ], + "type": "text", + "content": " is the magnitude of negative curvature. This geometric property enables efficient encoding of both fine-grained details (in near-flat regions) and complex hierarchical structures (in hyperbolic-like regions)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 400, + 370, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 370, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 370, + 412 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 420, + 382, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 382, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 382, + 431 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "text", + "content": "The workflow of our method is shown in Fig. 2. Our approach comprises two stages: (1) hierarchical predictive pre-training, where we learn a generic visual representation " + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "inline_equation", + "content": "f_{\\theta_1}(\\cdot)" + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "text", + "content": " from unlabeled data " + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "text", + "content": " by integrating three complementary prediction tasks that capture both local and global dependencies; and (2) task-specific fine-tuning, where we adapt the pre-trained model to downstream tasks using both " + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "text", + "content": " and labels " + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "text", + "content": " through a task-specific model " + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "inline_equation", + "content": "g_{\\theta_2}(\\cdot)" + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "text", + "content": " initialized with " + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "inline_equation", + "content": "\\theta_1" + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "text", + "content": ". For EM neuron segmentation, we process input volumes " + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\{X^{(1)},\\dots,X^{(T)}\\}" + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "inline_equation", + "content": "X^{(t)} \\in \\mathbb{R}^{D\\times H\\times W}" + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "text", + "content": " represents a 3D image, by partitioning them into smaller patches " + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{D' \\times H' \\times W'}" + }, + { + "bbox": [ + 313, + 437, + 555, + 606 + ], + "type": "text", + "content": " and leveraging Mamba's efficient sequence modeling capabilities to handle the high-dimensional, long-sequence nature of the data." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 612, + 553, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 553, + 625 + ], + "type": "text", + "content": "4.2. Hierarchical Predictive Coding in Token Space" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 629, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 712 + ], + "type": "text", + "content": "Building on our theoretical insights, we develop a hierarchical predictive coding framework that operates on the manifold of token embeddings to effectively capture the multiscale structure of high-dimensional neural image data. Our framework is motivated by the principle of maximum information preservation across complementary representational dimensions." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 746, + 318, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 746, + 318, + 755 + ], + "spans": [ + { + "bbox": [ + 294, + 746, + 318, + 755 + ], + "type": "text", + "content": "13606" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 556, + 150 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 556, + 150 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 556, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 556, + 150 + ], + "type": "image", + "image_path": "8b709dcf7c3700340886284460baf6e5edc34fa59e74e831e19d0cf1dd86d173.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "text", + "content": "Figure 2. Illustration of the TokenUnify pretraining workflow. The image " + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "text", + "content": " is fed into the Tokenizer, transforming it into a long sequence of tokens " + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "x_{i}|_{i = 1}^{K}" + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "text", + "content": ". The three prediction tasks (random token, next token, and next-all token) are performed with complementary objectives. The Perceiver Resampler is employed to convert varying-size large feature maps into a few visual tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "text", + "content": "Notation. Let " + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = \\{x_{1}, x_{2}, \\ldots, x_{K}\\}" + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "text", + "content": " represent a sequence of visual tokens extracted from a volumetric EM image, where each " + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "inline_equation", + "content": "x_{i} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "text", + "content": ". For random token prediction, " + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\subset \\{1, \\ldots, K\\}" + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "text", + "content": " denotes the set of masked token indices with " + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "inline_equation", + "content": "|\\mathcal{M}| = \\rho K" + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "text", + "content": " for masking ratio " + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{c} = \\{1, \\ldots, K\\} \\setminus \\mathcal{M}" + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "text", + "content": " denotes unmasked tokens. For sequential prediction, " + }, + { + "bbox": [ + 55, + 212, + 296, + 307 + ], + "type": "inline_equation", + "content": "x_{ 0" + }, + { + "bbox": [ + 55, + 163, + 295, + 200 + ], + "type": "text", + "content": " is the annealing temperature controlling transition sharpness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 201, + 259, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 201, + 259, + 211 + ], + "spans": [ + { + "bbox": [ + 67, + 201, + 259, + 211 + ], + "type": "text", + "content": "Our unified objective function is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 222, + 295, + 235 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 222, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 61, + 222, + 295, + 235 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {T o k e n U n i f y}} = \\alpha (t) \\cdot \\mathcal {L} _ {\\text {r a n d o m}} + \\beta (t) \\cdot \\mathcal {L} _ {\\text {n e x t}} + \\gamma (t) \\cdot \\mathcal {L} _ {\\text {n e x t - a l l}}. \\tag {7}", + "image_path": "a57b6a81e43e197fcd536b899c7db01591d41e76eb79b1e179ce66ca78525b00.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 244, + 295, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 295, + 269 + ], + "type": "text", + "content": "The curriculum implements progressive transition from local to global representation learning through:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 277, + 308, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 308, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 308, + 331 + ], + "type": "interline_equation", + "content": "\\mathbf {v} (t) = \\left\\{ \\begin{array}{l l} [ 2. 0, - 1. 0, - 2. 0 ] & \\text {i f} t < T _ {1} (\\text {r a n d o m - d o m i n a n t}) \\\\ [ - 1. 0, 2. 0, - 1. 0 ] & \\text {i f} T _ {1} \\leq t < T _ {2} (\\text {n e x t - d o m i n a n t}), \\\\ [ - 2. 0, - 1. 0, 2. 0 ] & \\text {i f} t \\geq T _ {2} (\\text {n e x t - a l l - d o m i n a n t}) \\end{array} \\right. \\tag {8}", + "image_path": "9fd680198b519a5a47be65a3bc3789abb4a8661859bdecb05a634b05fe7e8bdb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "inline_equation", + "content": "T_{1} = 0.3 \\times T_{\\mathrm{total}}" + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "inline_equation", + "content": "T_{2} = 0.7 \\times T_{\\mathrm{total}}" + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "content": " are transition thresholds. With " + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\tau(t) = 1.0" + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "content": ", this yields dominant weights of " + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\sim 0.73" + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "content": " for the primary task while maintaining auxiliary contributions of " + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\sim 0.18" + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\sim 0.09" + }, + { + "bbox": [ + 55, + 332, + 296, + 392 + ], + "type": "text", + "content": " to preserve multi-task synergy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 392, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 440 + ], + "type": "text", + "content": "For computational efficiency in next-all token prediction, we employ cross-attention with latent queries to aggregate sequence-wide information, enabling scalability to long token sequences without prohibitive complexity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 440, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 296, + 499 + ], + "type": "text", + "content": "This hierarchical integration constructs a multi-resolution representational manifold capturing neural tissue organization across microscopic, mesoscopic, and macroscopic scales, providing robust foundations for volumetric electron microscopy circuit reconstruction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 506, + 217, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 217, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 217, + 520 + ], + "type": "text", + "content": "4.3. Finetuning and Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 524, + 296, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 296, + 549 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 296, + 549 + ], + "type": "text", + "content": "The segmentation network, denoted as " + }, + { + "bbox": [ + 55, + 524, + 296, + 549 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{seg}}(\\cdot)" + }, + { + "bbox": [ + 55, + 524, + 296, + 549 + ], + "type": "text", + "content": ", consists of an encoder " + }, + { + "bbox": [ + 55, + 524, + 296, + 549 + ], + "type": "inline_equation", + "content": "g_{e}(\\cdot)" + }, + { + "bbox": [ + 55, + 524, + 296, + 549 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 55, + 524, + 296, + 549 + ], + "type": "inline_equation", + "content": "g_{d}(\\cdot)" + }, + { + "bbox": [ + 55, + 524, + 296, + 549 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 90, + 558, + 295, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 558, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 90, + 558, + 295, + 571 + ], + "type": "interline_equation", + "content": "g _ {\\mathrm {s e g}} (x; \\theta_ {s}) = g _ {d} \\left(g _ {e} (x)\\right), \\quad \\theta_ {s} = \\left\\{\\theta_ {e}, \\theta_ {d} \\right\\}, \\tag {9}", + "image_path": "67c98cc383499ec8aca617091557835c57f118c214122cb61c978a7772e3526b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "content": " represents the parameters of the entire segmentation network, and " + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\theta_{e}" + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\theta_{d}" + }, + { + "bbox": [ + 55, + 580, + 295, + 616 + ], + "type": "text", + "content": " are the parameters of the encoder and decoder, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 616, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 700 + ], + "type": "text", + "content": "The encoder " + }, + { + "bbox": [ + 55, + 616, + 296, + 700 + ], + "type": "inline_equation", + "content": "g_{e}(\\cdot)" + }, + { + "bbox": [ + 55, + 616, + 296, + 700 + ], + "type": "text", + "content": " gradually downsamples the input volume and extracts high-level semantic features, while the decoder " + }, + { + "bbox": [ + 55, + 616, + 296, + 700 + ], + "type": "inline_equation", + "content": "g_{d}(\\cdot)" + }, + { + "bbox": [ + 55, + 616, + 296, + 700 + ], + "type": "text", + "content": " upsamples the encoded features back to the original resolution. Meanwhile, the output of each downsampling layer in the encoder is connected to the corresponding layer in the decoder via skip connections to fuse local and global multi-scale information." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 144 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 144 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 144 + ], + "type": "image", + "image_path": "57a13d443f133743c56e008953f5650b6d532745e4b3e216ed9fd0f08fcf37f4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 145, + 419, + 156 + ], + "lines": [ + { + "bbox": [ + 343, + 145, + 419, + 156 + ], + "spans": [ + { + "bbox": [ + 343, + 145, + 419, + 156 + ], + "type": "text", + "content": "(a) Sampling Area" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 145, + 534, + 156 + ], + "lines": [ + { + "bbox": [ + 474, + 145, + 534, + 156 + ], + "spans": [ + { + "bbox": [ + 474, + 145, + 534, + 156 + ], + "type": "text", + "content": "(b) Annotation" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 164, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 164, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 555, + 220 + ], + "type": "text", + "content": "Figure 3. Illustration of the MEC dataset annotation. (a) shows the sampling locations of six different neural regions from the MEC raw data. Each region contains " + }, + { + "bbox": [ + 313, + 164, + 555, + 220 + ], + "type": "inline_equation", + "content": "1250 \\times 1250 \\times 125" + }, + { + "bbox": [ + 313, + 164, + 555, + 220 + ], + "type": "text", + "content": " voxels. (b) shows an example of the fine segmentation labels in one sampling region, which takes two experts one month to annotate." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 230, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 553, + 277 + ], + "type": "text", + "content": "We adopt 3D ResUNet/ViT/Mamba as the backbone network, respectively. For the Mamba variant, we adapt SegMamba [43] for EM data, which we refer to as EMmamba, with enhanced decoder design for anisotropic EM features." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 278, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 554, + 350 + ], + "type": "text", + "content": "The output of the segmentation network " + }, + { + "bbox": [ + 313, + 278, + 554, + 350 + ], + "type": "inline_equation", + "content": "\\hat{y} = g_{\\mathrm{seg}}(x)\\in" + }, + { + "bbox": [ + 313, + 278, + 554, + 350 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times D\\times H\\times W}" + }, + { + "bbox": [ + 313, + 278, + 554, + 350 + ], + "type": "text", + "content": " represents the predicted affinity map [25, 26], corresponding to the connectivity probability of each voxel in three directions. During training, the loss function for labeled samples is the mean squared error between the predicted and ground-truth affinity maps:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 365, + 357, + 553, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 357, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 365, + 357, + 553, + 390 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s e g}} = \\frac {1}{| \\mathcal {D} _ {l} |} \\sum_ {i = 1} ^ {| \\mathcal {D} _ {l} |} \\| g _ {\\mathrm {s e g}} \\left(x _ {i} ^ {l}\\right) - y _ {i} \\| ^ {2}. \\tag {10}", + "image_path": "6278dbb5dad11945e1afd59f5aa757106b192b322468a1a3bc86b8b1b8d75398.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 396, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 492 + ], + "type": "text", + "content": "During inference, for any new test sample " + }, + { + "bbox": [ + 313, + 396, + 554, + 492 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 396, + 554, + 492 + ], + "type": "text", + "content": ", forward propagation through " + }, + { + "bbox": [ + 313, + 396, + 554, + 492 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{seg}}(x_{t})" + }, + { + "bbox": [ + 313, + 396, + 554, + 492 + ], + "type": "text", + "content": " yields its predicted affinity map. This predicted affinity map is then post-processed using a seeded watershed algorithm and a structure-aware region agglomeration algorithm [5, 15] to obtain the final neuron instance segmentation. Detailed information on our segmentation process and the segmentation pipeline are illustrated in the supplementary material." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 502, + 437, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 437, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 437, + 514 + ], + "type": "text", + "content": "5. Datasets and Metrics" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 521, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 629 + ], + "type": "text", + "content": "Datasets. For pretraining TokenUnify, we collect over 1 TB of publicly available unlabeled EM imaging data from four large-scale datasets: FAFB [35], MitoEM [42], FIB-25 [38], and Kasthuri15 [22]. These datasets span diverse organisms including Drosophila, mouse, rat, and human samples, ensuring comprehensive coverage of EM morphology variations. We sample from these datasets with equal probability to guarantee diversity in pretraining data. Additional details are provided in supplementary material." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "For downstream fine-tuning and evaluation, we employ two datasets: AC3/4 [22] and our large-scale MEC dataset. The AC3/4 dataset consists of 256/100 successive EM images with " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " spatial resolution from mouse somatosensory cortex. We partition AC4 into training (first 80 images) and validation (last 20 images), while using AC3 (first 100 images) for testing." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 746, + 317, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 746, + 317, + 755 + ], + "spans": [ + { + "bbox": [ + 294, + 746, + 317, + 755 + ], + "type": "text", + "content": "13608" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 553, + 369 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 553, + 369 + ], + "type": "table", + "html": "
Post.MethodWafer4Wafer36-2Param. (M)
VOIM↓VOIS↓VOI↓ARAND↓VOIM↓VOIS↓VOI↓ARAND↓
LMC [5]Supervised Methods
Superhuman [23]0.33281.12581.45870.17360.15060.45880.60940.08361.478
MALA [15]0.54381.50272.03750.11150.31791.06641.38430.157084.03
PEA [21]0.33810.92761.26580.06770.27870.42790.70660.11691.480
UNETR [18]0.45041.65812.10850.26580.44780.52170.96960.2913129.1
EMmamba0.49151.29241.78390.20520.24060.41890.65950.123128.30
Superhuman [23]0.19481.96972.16440.24530.07921.16181.24100.13191.478
MALA [15]0.34162.41292.75450.25670.14481.96032.10520.197784.03
PEA [21]0.17051.59931.76980.15270.47191.12261.59450.15881.480
UNETR [18]0.17913.17153.35060.63300.09491.38581.48070.1578129.1
EMmamba0.15962.05802.21770.19730.08471.03511.11980.125328.30
LMC [5]Self-Supervised Methods
Random0.49151.29241.78390.20520.24060.41890.65950.123128.30
MAE [20]0.23251.09231.32480.09780.16290.45320.61610.0835
BYOL [16]0.25840.94531.20370.08910.17480.43680.61160.0826
dbMIM [8]0.23420.87961.11380.07420.14760.39820.54580.0782
MS-Con-EM [9]0.24830.91021.15850.08130.15420.40370.55790.0794
TokenUnify0.19530.79980.99510.05090.12620.35850.48480.0650
Random0.15962.05802.21770.19730.08471.03511.11980.125328.30
MAE [20]0.14931.92162.07090.15830.08921.09271.18190.1243
BYOL [16]0.16291.83742.00030.14270.09531.07821.17350.1287
dbMIM [8]0.15731.74361.90090.09130.09121.06491.15610.1185
MS-Con-EM [9]0.16091.76921.93010.10260.09631.07031.16660.1207
TokenUnify0.14181.51031.65210.05910.08271.02761.11030.1074
", + "image_path": "1caedde039a0f42b734773c3e47dd7d721dd912acb37e39828aa8a912b758063.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 377, + 555, + 412 + ], + "lines": [ + { + "bbox": [ + 55, + 377, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 555, + 412 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of electron microscopy segmentation results on Wafer4 and Wafer36-2 datasets. Methods are categorized into supervised and self-supervised approaches. All self-supervised methods use the same EMmamba backbone for fair comparison. \"Random\" refers to EMmamba without any pretraining. The best results are in bold and the second best results are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 431, + 294, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 431, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 431, + 294, + 563 + ], + "type": "text", + "content": "To validate algorithm performance on large-scale EM data, we collected a 2TB MEC dataset by imaging mouse somatosensory cortex, medial entorhinal cortex, and cerebral cortex regions. We selected 6 representative volumes from different neural regions (wafer4/25/26/26-2/36/36-2), each with dimensions " + }, + { + "bbox": [ + 54, + 431, + 294, + 563 + ], + "type": "inline_equation", + "content": "1250 \\times 1250 \\times 125" + }, + { + "bbox": [ + 54, + 431, + 294, + 563 + ], + "type": "text", + "content": " voxels. Dense manual annotation was performed on these regions, yielding over 1.2 billion annotated voxels through six months of work by two experts, as illustrated in Fig. 3. The dataset is split into training (wafer25/26/26-2/36), validation (wafer4), and testing (wafer36-2) sets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 566, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 651 + ], + "type": "text", + "content": "Metrics. To evaluate the performance of neuron segmentation [12, 45], we employ two widely-used metrics: Variation of Information (VOI) [32] and Adjusted Rand Index (ARAND) [2]. These metrics quantify the agreement between the predicted segmentation and the ground truth from different perspectives. Detailed descriptions of these metrics can be found in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 661, + 132, + 675 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 132, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 132, + 675 + ], + "type": "text", + "content": "6. Experiment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 681, + 295, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 681, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 681, + 295, + 706 + ], + "type": "text", + "content": "We evaluate our proposed method on both MEC and AC3/4 datasets. Implementation details are provided in the sup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 431, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 575 + ], + "type": "text", + "content": "plementary material. We compare our method with both supervised and self-supervised approaches. For supervised methods, we include Superhuman [23], MALA [15], PEA [21], and UNETR [18]. For self-supervised methods, we evaluate two categories: (1) general-purpose visual representation learning methods including MAE [20] and BYOL [16], which were originally designed for natural images but adapted to EM data; and (2) EM-specific self-supervised methods including dbMIM [8] and MS-Con-EM [9], which were specifically designed for electron microscopy images. All self-supervised methods utilize the same EMmamba backbone for fair comparison." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 578, + 556, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 556, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 556, + 698 + ], + "type": "text", + "content": "Results on MEC Dataset. Table 1 presents the quantitative results on the large-scale MEC dataset. Compared to models with similar parameter counts, on the Wafer4 volume under the waterz post-processing, our TokenUnify approach achieves approximately a " + }, + { + "bbox": [ + 313, + 578, + 556, + 698 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 578, + 556, + 698 + ], + "type": "text", + "content": " performance improvement over MAE and a " + }, + { + "bbox": [ + 313, + 578, + 556, + 698 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 313, + 578, + 556, + 698 + ], + "type": "text", + "content": " improvement compared to direct training without pretraining. Visualization results in Fig. 4 and Fig. 5 further demonstrate that our approach significantly outperforms other methods, particularly in challenging regions with complex neuronal structures." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 746, + 318, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 746, + 318, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 746, + 318, + 756 + ], + "type": "text", + "content": "13609" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 555, + 300 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 555, + 300 + ], + "type": "image", + "image_path": "1ca765afeba898f22e39affead4bf0485ec10b71e7272a27d207f57ccf5b1281.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 337, + 555, + 490 + ], + "blocks": [ + { + "bbox": [ + 55, + 314, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 55, + 314, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 555, + 335 + ], + "type": "text", + "content": "Figure 4. 2D visualization results for two slices from the MEC dataset: wafer 4 and wafer 36-2. The left side displays the EM raw images, while the right side presents the affinity and segmentation results. Red boxes indicate the regions of mis-segmentation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 337, + 555, + 490 + ], + "lines": [ + { + "bbox": [ + 59, + 337, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 59, + 337, + 555, + 490 + ], + "type": "image", + "image_path": "12dd48e2cc5a9a338d09f9f6b46beb755632fecd9e76de421ddf68d95efe5aec.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 533, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 653 + ], + "type": "text", + "content": "Results on AC3/4 Datasets. We also evaluate performance on the smaller AC3/4 dataset, which contains only about 1/10 of the labeled pixels compared to MEC. As shown in Table 2, even in this low-data scenario, the Mamba architecture combined with TokenUnify pretraining achieves performance comparable to state-of-the-art supervised methods like PEA. Additionally, it demonstrates a " + }, + { + "bbox": [ + 55, + 533, + 295, + 653 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 55, + 533, + 295, + 653 + ], + "type": "text", + "content": " performance improvement over the MAE pretraining approach. This highlights the effectiveness of TokenUnify in scenarios with limited labeled data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 656, + 295, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 295, + 704 + ], + "type": "text", + "content": "Scaling Law Analysis. To investigate how our method scales with model size, we evaluated various initialization and training strategies by scaling model parameters from 100M to 1B. Results in Fig. 6 show that all methods ex" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 533, + 550, + 647 + ], + "blocks": [ + { + "bbox": [ + 55, + 502, + 555, + 523 + ], + "lines": [ + { + "bbox": [ + 55, + 502, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 555, + 523 + ], + "type": "text", + "content": "Figure 5. 3D visualization results on two MEC datasets: wafer 4 and wafer 36-2. The red arrows indicate over-merged structures, and the orange arrows highlight over-split structures." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 533, + 550, + 647 + ], + "lines": [ + { + "bbox": [ + 316, + 533, + 550, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 550, + 647 + ], + "type": "image", + "image_path": "a0d5089d7f9b9e2100df2021ffa21be50419adf742f75f8fc5a586d02765894c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 661, + 555, + 694 + ], + "lines": [ + { + "bbox": [ + 313, + 661, + 555, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 661, + 555, + 694 + ], + "type": "text", + "content": "Figure 6. Performance evaluation of models with 100M, 200M, 500M, and 1B parameters. The models were trained for 100 epochs on the MEC dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 746, + 318, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 746, + 318, + 755 + ], + "spans": [ + { + "bbox": [ + 293, + 746, + 318, + 755 + ], + "type": "text", + "content": "13610" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 295, + 239 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 239 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 239 + ], + "type": "table", + "html": "
MethodVOIM↓VOIS↓VOI↓ARAND↓Param.(M)
Supervised Methods
Superhuman [23]0.48820.65631.14450.17481.478
MALA [15]0.45710.67671.13380.166484.02
PEA [21]0.55220.49801.05020.20931.480
UNETR [18]0.77990.73991.51980.2411129.1
EMmamba0.93780.86291.80070.284028.30
Self-Supervised Methods
Random0.93780.86291.80070.2840
MAE [20]0.31670.79631.11310.1050
BYOL [16]0.38420.76211.14630.152728.30
dbMIM [8]0.46230.62791.09020.1432
MS-Con-EM [9]0.47850.65381.13230.1498
TokenUnify0.44790.54390.99180.1366
", + "image_path": "f783d666fe097bee87ed62b11499099d5dd7dabbbb4af74dabe2a5cde8ddc943.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 93, + 304, + 259, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 246, + 295, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 246, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 295, + 301 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of segmentation results on AC3/4 datasets with waterz [15] post-processing algorithm. All self-supervised methods use the same EMmamba backbone. \"Random\" refers to EMmamba without any pretraining. The best results are in bold and the second best results are underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 304, + 259, + 392 + ], + "lines": [ + { + "bbox": [ + 93, + 304, + 259, + 392 + ], + "spans": [ + { + "bbox": [ + 93, + 304, + 259, + 392 + ], + "type": "table", + "html": "
Pretraining StrategyVOI ↓ARAND ↓
RandomNextNext-all
✓1.26800.0862
✓4.04180.4416
✓✓1.13000.0692
✓✓1.19070.1203
✓✓✓0.99510.0509
", + "image_path": "59c940550e87c7a9ecd6922775c3c3938307095cdb2c9f902477163b4f78196c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 88, + 414, + 264, + 502 + ], + "blocks": [ + { + "bbox": [ + 61, + 399, + 290, + 411 + ], + "lines": [ + { + "bbox": [ + 61, + 399, + 290, + 411 + ], + "spans": [ + { + "bbox": [ + 61, + 399, + 290, + 411 + ], + "type": "text", + "content": "Table 3. Ablation for different pretraining strategies on wafer4." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 414, + 264, + 502 + ], + "lines": [ + { + "bbox": [ + 88, + 414, + 264, + 502 + ], + "spans": [ + { + "bbox": [ + 88, + 414, + 264, + 502 + ], + "type": "table", + "html": "
Finetuning ModuleVOI ↓ARAND ↓
MambaEncoderDecoder
✓1.13620.0782
✓1.55560.1370
✓1.52950.1212
✓✓1.10650.0629
✓✓✓0.99510.0509
", + "image_path": "f33e291facc650128b06e8b403d998585db514c7827ba1297d39d68d6ee344ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 510, + 278, + 521 + ], + "lines": [ + { + "bbox": [ + 72, + 510, + 278, + 521 + ], + "spans": [ + { + "bbox": [ + 72, + 510, + 278, + 521 + ], + "type": "text", + "content": "Table 4. Ablation for the fine-tuning schemes on wafer4." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 541, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 295, + 625 + ], + "type": "text", + "content": "hibit performance improvements after pretraining. In contrast, TokenUnify demonstrates strong scaling properties, consistently outperforming other methods across both small and large model sizes. Notably, the Mamba architecture achieves competitive segmentation performance with fewer parameters compared to Transformer, confirming its efficiency for processing long sequence data in connectomics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 633, + 156, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 156, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 156, + 644 + ], + "type": "text", + "content": "6.1. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 650, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 295, + 686 + ], + "type": "text", + "content": "We conduct ablation experiments to analyze the key components of our approach. As shown in Table 3 and Table 4, we examine two critical aspects:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Ablation on Token Prediction Strategies. We systematically evaluate different token prediction combinations dur" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "ing pretraining on the wafer4 dataset (Table 3). The full combination (Random + Next + Next-all) achieves optimal performance " + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "inline_equation", + "content": "(\\mathrm{VOI} = 0.9951)" + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": ", significantly outperforming individual strategies and demonstrating substantial improvements over baseline approaches. Next-token prediction alone performs poorly " + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "inline_equation", + "content": "(\\mathrm{VOI} = 4.0418)" + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": ", confirming that conventional autoregressive pretraining is insufficient for visual tasks. Unlike sequential text, visual data requires global spatial understanding that pure autoregressive modeling cannot capture effectively. Random token prediction provides better spatial initialization by encouraging spatial consistency, while the strategic combination of multiple prediction objectives yields more robust and generalizable representations for segmentation. These results validate our hypothesis that diverse pretraining tasks are essential for effective visual representation learning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 266, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 411 + ], + "type": "text", + "content": "Fine-tuning Components. We also investigate which network components benefit most from fine-tuning (Table 4). Our architecture consists of three main parts: Mamba blocks (sequence modeling), encoder (downsampling), and decoder (upsampling). Fine-tuning only the Mamba blocks yields substantially better results " + }, + { + "bbox": [ + 313, + 266, + 555, + 411 + ], + "type": "inline_equation", + "content": "(\\mathrm{VOI} = 1.1362)" + }, + { + "bbox": [ + 313, + 266, + 555, + 411 + ], + "type": "text", + "content": " than fine-tuning only the encoder " + }, + { + "bbox": [ + 313, + 266, + 555, + 411 + ], + "type": "inline_equation", + "content": "(\\mathrm{VOI} = 1.5556)" + }, + { + "bbox": [ + 313, + 266, + 555, + 411 + ], + "type": "text", + "content": " or decoder " + }, + { + "bbox": [ + 313, + 266, + 555, + 411 + ], + "type": "inline_equation", + "content": "(\\mathrm{VOI} = 1.5295)" + }, + { + "bbox": [ + 313, + 266, + 555, + 411 + ], + "type": "text", + "content": ", demonstrating the critical role of sequence modeling in connectomics segmentation. For resource-constrained scenarios, prioritizing the fine-tuning of Mamba blocks provides an efficient balance between performance and computational cost." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 421, + 388, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 388, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 388, + 433 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 441, + 555, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 597 + ], + "type": "text", + "content": "We propose TokenUnify, a framework that unifies random token prediction, next-token prediction, and next-all token prediction to capture multi-scale visual structures. Our analysis shows this approach mitigates cumulative errors in autoregressive modeling. To validate TokenUnify, we introduce a large-scale EM neuron segmentation dataset with 1.2 billion labeled voxels across six mouse brain regions. Experiments demonstrate a " + }, + { + "bbox": [ + 313, + 441, + 555, + 597 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 313, + 441, + 555, + 597 + ], + "type": "text", + "content": " improvement over training from scratch and a " + }, + { + "bbox": [ + 313, + 441, + 555, + 597 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 313, + 441, + 555, + 597 + ], + "type": "text", + "content": " gain over MAE pretraining. Scaling analysis further highlights its efficiency with increasing model size and data volume. By releasing our dataset and implementation, we aim to advance biological imaging and inspire future research in visual representation learning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 607, + 411, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 607, + 411, + 620 + ], + "spans": [ + { + "bbox": [ + 314, + 607, + 411, + 620 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 627, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 554, + 712 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant 624B2137. We thank the staff at the Institute of Artificial Intelligence, Hefei Comprehensive National Science Center for their assistance with biological specimen preparation, imaging, and annotation. We are grateful to the anonymous reviewers for their thorough reviews and constructive suggestions." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 746, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 746, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 746, + 317, + 756 + ], + "type": "text", + "content": "13611" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 213 + ], + "type": "text", + "content": "[2] Ignacio Arganda-Carreras, Srinivas C Turaga, Daniel R Berger, Dan Ciresan, Alessandro Giusti, Luca M Gambardella, Jürgen Schmidhuber, Dmitry Laptev, Sarvesh Dwivedi, Joachim M Buhmann, et al. Crowdsourcing the creation of image segmentation algorithms for connectomics. Frontiers in Neuroanatomy, 9:152591, 2015. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 213, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 245 + ], + "type": "text", + "content": "[3] Gregor Bachmann and Vaishnavh Nagarajan. The pitfalls of next-token prediction. arXiv preprint arXiv:2403.06963, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "text", + "content": "[4] Yutong Bai, Xinyang Geng, Karttikeya Mangalam, Amir Bar, Alan Yuille, Trevor Darrell, Jitendra Malik, and Alexei A Efros. Sequential modeling enables scalable learning for large vision models. arXiv preprint arXiv:2312.00785, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 295, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 367 + ], + "type": "text", + "content": "[5] Thorsten Beier, Constantin Pape, Nasim Rahaman, Timo Prange, Stuart Berg, Davi D Bock, Albert Cardona, Graham W Knott, Stephen M Plaza, Louis K Scheffer, et al. Multicut brings automated neurite segmentation closer to human performance. Nature Methods, 14(2):101-102, 2017. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 401 + ], + "type": "text", + "content": "[6] Kevin Briggman, Winfried Denk, Sebastian Seung, Moritz Helmstaedter, and Srinivas C Turaga. Maximin affinity learning of image segmentation. In NeurIPS, 2009. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 402, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 295, + 435 + ], + "type": "text", + "content": "[7] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "text", + "content": "[8] Yinda Chen, Wei Huang, Shenglong Zhou, Qi Chen, and Zhiwei Xiong. Self-supervised neuron segmentation with multi-agent reinforcement learning. In IJCAI, 2023. 1, 3, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "type": "text", + "content": "[9] Yinda Chen, Wei Huang, Xiaoyu Liu, Shiyu Deng, Qi Chen, and Zhiwei Xiong. Learning multiscale consistency for self-supervised electron microscopy instance segmentation. In ICASSP, 2024. 3, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 567 + ], + "type": "text", + "content": "[10] Zhimin Chen, Yingwei Li, Longlong Jing, Liang Yang, and Bing Li. Point cloud self-supervised learning via 3d to multiview masked autoencoder. arXiv preprint arXiv:2311.10887, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 602 + ], + "type": "text", + "content": "[11] Zhimin Chen, Longlong Jing, Yingwei Li, and Bing Li. Bridging the domain gap: Self-supervised 3d scene understanding with foundation models. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 645 + ], + "type": "text", + "content": "[12] Bo Dang, Wenchao Zhao, Yufeng Li, Danqing Ma, Qixuan Yu, and Elly Yijun Zhu. Real-time pill identification for the visually impaired using deep learning. arXiv preprint arXiv:2405.05983, 2024. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[13] Zhiyuan Ding, Qi Dong, Haote Xu, Chenxin Li, Xinghao Ding, and Yue Huang. Unsupervised anomaly segmentation for brain lesions using dual semantic-manifold reconstruction. In ICONIP, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[14] Alaaeldin El-Nouby, Michal Klein, Shuangfei Zhai, Miguel Angel Bautista, Alexander Toshev, Vaishaal Shankar," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 710 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 106 + ], + "type": "text", + "content": "Joshua M Susskind, and Armand Joulin. Scalable pretraining of large autoregressive image models. arXiv preprint arXiv:2401.08541, 2024. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 173 + ], + "type": "text", + "content": "[15] Jan Funke, Fabian Tschopp, William Grisaitis, Arlo Sheridan, Chandan Singh, Stephan Saalfeld, and Srinivas C Turaga. Large scale image segmentation with structured loss based deep learning for connectome reconstruction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(7):1669-1680, 2018. 2, 5, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[16] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, 2020. 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 262 + ], + "type": "text", + "content": "[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 307 + ], + "type": "text", + "content": "[18] Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. In WACV, 2022. 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 308, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 342 + ], + "type": "text", + "content": "[19] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "text", + "content": "[20] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 418 + ], + "type": "text", + "content": "[21] Wei Huang, Shiyu Deng, Chang Chen, Xueyang Fu, and Zhiwei Xiong. Learning to model pixel-embedded affinity for homogeneous instance segmentation. In AAAI, 2022. 1, 2, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 421, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 486 + ], + "type": "text", + "content": "[22] Narayanan Kasthuri, Kenneth Jeffrey Hayworth, Daniel Raimund Berger, Richard Lee Schalek, José Angel Conchello, Seymour Knowles-Barley, Dongil Lee, Amelio Vázquez-Reina, Verena Kaynig, Thouis Raymond Jones, et al. Saturated reconstruction of a volume of neocortex. Cell, 162(3):648-661, 2015. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 487, + 553, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 530 + ], + "type": "text", + "content": "[23] Kisuk Lee, Jonathan Zung, Peter Li, Viren Jain, and H Sebastian Seung. Superhuman accuracy on the snemi3d connectomics challenge. arXiv preprint arXiv:1706.00120, 2017. 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 533, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 576 + ], + "type": "text", + "content": "[24] Chenxin Li, Yunlong Zhang, Zhehan Liang, Wenao Ma, Yue Huang, and Xinghao Ding. Consistent posterior distributions under vessel-mixing: A regularization for cross-domain retinal artery/vein classification. In ICIP, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 620 + ], + "type": "text", + "content": "[25] Panfeng Li, Youzuo Lin, and Emily Schultz-Fellenz. Contextual hourglass network for semantic segmentation of high resolution aerial imagery. arXiv preprint arXiv:1810.12813, 2018. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 622, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 666 + ], + "type": "text", + "content": "[26] Panfeng Li, Mohamed Abouelenien, and Rada Mihalcea. Deception detection from linguistic and physiological data streams using bimodal convolutional neural networks. arXiv preprint arXiv:2311.10944, 2023. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 667, + 553, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 667, + 553, + 710 + ], + "spans": [ + { + "bbox": [ + 316, + 667, + 553, + 710 + ], + "type": "text", + "content": "[27] Tianhong Li, Huiwen Chang, Shlok Mishra, Han Zhang, Dina Katabi, and Dilip Krishnan. Mage: Masked generative encoder to unify representation learning and image synthesis. In CVPR, 2023. 3" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 746, + 318, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 746, + 318, + 756 + ], + "spans": [ + { + "bbox": [ + 294, + 746, + 318, + 756 + ], + "type": "text", + "content": "13612" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 710 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[28] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9): 1-35, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 161 + ], + "type": "text", + "content": "[29] Xiaoyu Liu, Wei Huang, Yueyi Zhang, and Zhiwei Xiong. Biological instance segmentation with a superpixel-guided graph. In IJCAI, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[30] Xiaoyu Liu, Miaomiao Cai, Yinda Chen, Yueyi Zhang, Te Shi, Ruobing Zhang, Xuejin Chen, and Zhiwei Xiong. Cross-dimension affinity distillation for 3d em neuron segmentation. In CVPR, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 239 + ], + "type": "text", + "content": "[31] Christopher W Lynn and Danielle S Bassett. The physics of brain network structure, function and control. Nature Reviews Physics, 1(5):318-332, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 285 + ], + "type": "text", + "content": "[32] Juan Nunez-Iglesias, Ryan Kennedy, Toufiq Parag, Jianbo Shi, and Dmitri B Chklovskii. Machine learning of hierarchical clustering to segment 2d and 3d images. PLoS ONE, 8(8):e71715, 2013. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[33] Nobuhiko Ohno, Mitsuhiko Katoh, Yurika Saitoh, Sei Saitoh, and Shinichi Ohno. Three-dimensional volume imaging with electron microscopy toward connectome. Microscopy, 64(1):17-26, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 396 + ], + "type": "text", + "content": "[34] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "type": "text", + "content": "[35] Philipp Schlegel, Alexander S Bates, Tejal Parag, Gregory SXE Jefferis, and Davi D Bock. Automatic detection of synaptic partners in a whole-brain drosophila em dataset. Nature Methods, 18(8):877-884, 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 487 + ], + "type": "text", + "content": "[36] Arlo Sheridan, Tri M Nguyen, Diptodip Deb, Wei-Chung Allen Lee, Stephan Saalfeld, Srinivas C Turaga, Uri Manor, and Jan Funke. Local shape descriptors for neuron segmentation. Nature Methods, 20(2):295-303, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 488, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 542 + ], + "type": "text", + "content": "[37] Mannat Singh, Quentin Duval, Kalyan Vasudev Alwala, Haoqi Fan, Vaibhav Aggarwal, Aaron Adcock, Armand Joulin, Piotr Dollar, Christoph Feichtenhofer, Ross Girshick, et al. The effectiveness of mae pre-pretraining for billionscale pretraining. In ICCV, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "type": "text", + "content": "[38] Shin-ya Takemura, Yoshinori Aso, Toshihide Hige, Allan Wong, Zhiyuan Lu, C Shan Xu, Patricia K Rivlin, Harald Hess, Ting Zhao, Toufiq Parag, et al. A connectome of a learning and memory center in the adult drosophila brain. eLife, 6:e26975, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 599, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 599, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 294, + 654 + ], + "type": "text", + "content": "[39] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 655, + 294, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 655, + 294, + 710 + ], + "spans": [ + { + "bbox": [ + 56, + 655, + 294, + 710 + ], + "type": "text", + "content": "[40] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 295 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 127 + ], + "type": "text", + "content": "[41] Srinivas C Turaga, Joseph F Murray, Viren Jain, Fabian Roth, Moritz Helmstaedter, Kevin Briggman, Winfried Denk, and H Sebastian Seung. Convolutional networks can learn to generate affinity graphs for image segmentation. Neural Computation, 22(2):511-538, 2010. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 553, + 183 + ], + "type": "text", + "content": "[42] Donglai Wei, Zudi Lin, Daniel Franco-Barranco, Nils Wendt, Xingyu Liu, Wenjie Yin, Xin Huang, Aarush Gupta, Won-Dong Jang, Xueying Wang, et al. Mitoem dataset: Large-scale 3d mitochondria instance segmentation from em images. In MICCAI, 2020. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 185, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 185, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 315, + 185, + 553, + 217 + ], + "type": "text", + "content": "[43] Zhaohu Xing, Tian Ye, Yijun Yang, Guang Liu, and Lei Zhu. Segmamba: Long-range sequential modeling mamba for 3d medical image segmentation. In MICCAI, 2024. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 219, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 219, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 315, + 219, + 553, + 251 + ], + "type": "text", + "content": "[44] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 253, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 253, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 315, + 253, + 553, + 295 + ], + "type": "text", + "content": "[45] Jingyu Zhang, Ao Xiang, Yu Cheng, Qin Yang, and Liyang Wang. Research on detection of floating objects in river and lake based on ai intelligent image recognition. arXiv preprint arXiv:2404.06883, 2024. 6" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 746, + 317, + 755 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 746, + 317, + 755 + ], + "spans": [ + { + "bbox": [ + 294, + 746, + 317, + 755 + ], + "type": "text", + "content": "13613" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/439118b2-5f82-4c2e-9d89-cceba716e36e_content_list.json b/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/439118b2-5f82-4c2e-9d89-cceba716e36e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a26a04014cac366742efe8e9ff4f12f41b8c4e5 --- /dev/null +++ b/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/439118b2-5f82-4c2e-9d89-cceba716e36e_content_list.json @@ -0,0 +1,1345 @@ +[ + { + "type": "text", + "text": "TokensGen: Harnessing Condensed Tokens for Long Video Generation", + "text_level": 1, + "bbox": [ + 140, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenqi Ouyang1, Shuai Yang3,", + "bbox": [ + 215, + 164, + 349, + 202 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zeqi Xiao1, Lei Yang2,", + "bbox": [ + 388, + 165, + 491, + 202 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Danni Yang2, \nJianlou Si2,", + "bbox": [ + 524, + 165, + 635, + 202 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifan Zhou1, Xingang Pan1", + "bbox": [ + 665, + 165, + 779, + 202 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ S-Lab, Nanyang Technological University, $^{2}$ SenseTime Research, $^{3}$ Wangxuan Institute of Computer Technology, Peking University", + "bbox": [ + 215, + 208, + 779, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://vicky0522.github.io/tokensgen-webpage/", + "bbox": [ + 290, + 247, + 700, + 261 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb603e98ca6d897a1e47aac36b49d88bd29bdc1ebab90182f66d88ea280bfd63.jpg", + "image_caption": [ + "Figure 1. Given the text prompt, TokensGen generates long videos of up to 2 minutes, maintaining consistent motions and content. Moreover, TokensGen supports zero-shot prompt-guided video-to-video editing for long videos." + ], + "image_footnote": [], + "bbox": [ + 93, + 277, + 890, + 440 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 494, + 326, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating consistent long videos is a complex challenge: while diffusion-based generative models generate visually impressive short clips, extending them to longer durations often leads to memory bottlenecks and long-term inconsistency. In this paper, we propose TokensGen, a novel two-stage framework that leverages condensed tokens to address these issues. Our method decomposes long video generation into three core tasks: (1) inner-clip semantic control, (2) long-term consistency control, and (3) inter-clip smooth transition. First, we train To2V (Token-to-Video), a short video diffusion model guided by text and video tokens, with a Video Tokenizer that condenses short clips into semantically rich tokens. Second, we introduce T2To (Text-to-Token), a video token diffusion transformer that generates all tokens at once, ensuring global consistency across clips. Finally, during inference, an adaptive FIFO-Diffusion strategy seamlessly connects adjacent clips, reducing boundary artifacts and enhancing smooth transitions. Experimental results demonstrate that our approach significantly enhances long-term temporal and content coherence without incurring prohibitive computational overhead. By leveraging condensed tokens and pre-trained short video models, our method provides a scalable, modular solution for long video generation, opening new possibilities for storytelling, cinematic production, and immersive simulations.", + "bbox": [ + 88, + 518, + 485, + 898 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 494, + 643, + 510 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating consistent and visually pleasing long videos remains a formidable challenge in the field of video generation. While diffusion-based methods excel at short video generation [1-3, 9, 24, 36, 37, 41, 47], extending them to longer durations is constrained by computational resources, posing significant challenges.", + "bbox": [ + 511, + 516, + 906, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many approaches attempt to decompose long video generation into manageable sub-problems, using short video models without adding memory overhead. Tuning-free methods [7, 26, 28, 29, 32, 35] generate long videos using pre-trained short video models with hand-crafted techniques like noise re-scheduling, sliding window fusion, and attention manipulation, combined with multi-prompt sampling to enrich content. While ensuring high frame quality, these methods struggle with unnatural transitions due to missing long-range priors. Auto-regressive methods [1, 10, 18, 22, 30] and image-to-video approaches [5, 10, 40, 45] generate clips sequentially, achieving smooth transitions. However, they suffer from error accumulation, limited context windows, and unstable long-term controllability.", + "bbox": [ + 511, + 609, + 908, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hierarchical methods, such as MovieDreamer [46], adopt a multi-stage pipeline to address long-range challenges efficiently. They generate keyframes and then synthesize short clips guided by these keyframes, producing", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18197", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "high-quality results. However, they mainly focus on multiscene generation and lack strict consistency in motion and appearance across adjacent clips.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These limitations in long-term and short-term content control underscore the need for a unified, scalable solution that maintains both long-term and short-term consistency without excessive memory overhead. Therefore, we propose TokensGen, which leverages condensed video tokens to bridge short-clip generation with long-term consistency. Unlike hierarchical methods relying on keyframes generation and interpolation, or purely frame-level auto-regressive sampling, TokensGen jointly models spatial and temporal distributions for long videos through a two-stage framework, as detailed below:", + "bbox": [ + 88, + 138, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "a) To2V Model (Inner-clip content control): We employ a conditional short video generation model guided by text and video tokens to produce semantically rich yet concise video segments. Built on a powerful pre-trained backbone (CogVideoX [41]), our Video Tokenizer encodes short clips into a condensed set of high-level semantic tokens. This enables robust spatial layouts and motion cues in an efficient representation space, achieving stronger per-clip semantic control than text prompts alone.", + "b) T2To Model (Long-term content consistency): We train a video token diffusion transformer to generate the full set of tokens for a minute-long video from text prompts. These tokens are derived by encoding the long video clip-by-clip using the Video Tokenizer. Operating in this token space enables the T2To Model to maintain content continuity and logical coherence across clips while significantly reducing memory demands compared to raw frame modeling, preserving sufficient semantic detail for global consistency.", + "c) Adaptive FIFO-Diffusion (Inter-clip temporal smoothness): During inference, we sample long video tokens via the T2To Model and employ them to guide clip generation in the To2V Model. However, naively concatenating clips may cause boundary discontinuities, even with consistent semantic tokens. To overcome this, we propose an adaptive FIFO-Diffusion process for the To2V Model, enabling diagonal denoising of consecutive clips. This approach prevents distributional artifacts caused by naive padding or frame replication in FIFO-Diffusion [22], ensuring smoother transitions and improving the overall fidelity of the long video." + ], + "bbox": [ + 89, + 306, + 480, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared to prior methods for long video generation, TokensGen offers several key advantages. First, by leveraging pre-trained short video models, it inherits strong knowledge priors and architectural designs, enabling a smooth transition from short clips to minute-long sequences without extensive re-engineering. Second, encoding long videos into condensed token representations significantly reduces computational overhead for minute-level generation. Third, because each component (To2V Model, T2To Model, and", + "bbox": [ + 89, + 763, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the inter-clip scheduling) operates in a clearly defined subtask, our pipeline is highly flexible. It can seamlessly integrate with other short-term control strategies (e.g., Progressive Diffusion [39], Rolling Diffusion Models [30]) or multi-prompt composition frameworks [1, 7, 32].", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, TokensGen offers a scalable and resource-efficient framework for generating long videos with long-term consistency and smooth transitions, as shown in Fig. 1. By harnessing condensed tokens and powerful short video models, our approach significantly lowers the barrier to high-quality long video generation, opening new possibilities for storytelling, simulation, and beyond.", + "bbox": [ + 511, + 167, + 903, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 291, + 653, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video diffusion models. Video diffusion models generate videos from text or image prompts. Early methods [5, 9, 15, 16, 36, 37, 40] extend U-Net-based image diffusion to the temporal domain. However, they struggle with motion dynamics and content richness due to separate spatial-temporal attention and limited temporal windows. Recent works [2, 24, 27, 33, 41, 47] enhance fidelity and consistency by integrating diffusion transformers with 3D full-attention to jointly model spatial-temporal correlations and improved text encoders for complex prompts. While effective for short videos, extending these models to long videos remains computationally prohibitive.", + "bbox": [ + 511, + 316, + 903, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long video generation. Long video generation poses additional challenges for achieving content coherence, consistent dynamics, and efficient resource usage. We categorize long video generation methods into two groups: 1) those that optimize resource usage via engineering techniques or efficient model design, and 2) those that decompose long video generation into short video sub-tasks.", + "bbox": [ + 511, + 500, + 903, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Resource usage optimization. Recent transformer-based methods [2, 24, 27, 33, 41, 47] employ 3D-VAE to compress videos. However, as noted in CogVideoX [41], excessive compression hinders 3D-VAE convergence. ExVideo [12] extends SVD [5] to 128 frames via small learnable parameters with low memory overhead. Pyramidal Flow Matching [20] reformulates diffusion into pyramid stages, enabling efficient generation of videos up to 240 frames. While effective, these methods still face challenges in scaling to much longer durations.", + "bbox": [ + 511, + 607, + 903, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Problem decomposition via short video generation.", + "text_level": 1, + "bbox": [ + 511, + 758, + 870, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Multi-scene generation Multi-shot approaches [14, 25, 44, 49] divide long videos into segments and align them under a unified narrative, conditioning video diffusion on scene-level text or styles for coherence. MovieDreamer [46] employs a hierarchical pipeline to draft keyframes and refine shots. These methods emphasize storytelling and character consistency with relaxed demands on interclip coherence.", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18198", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Tuning-free methods. Tuning-free methods extend short video generation to longer durations via hand-crafted designs, such as co-denoising [28, 35], noise re-scheduling with sliding windows [29], and attention control mechanisms [7, 26, 32]. Often paired with multi-prompt sampling for content richness, these methods lack long-range priors, leading to unnatural motion and appearance transitions over extended durations.", + "- Auto-regressive methods. Auto-regressive methods generate long videos frame-by-frame (or clip-by-clip), ensuring temporal consistency at the cost of increased inference time. Image(clip)-to-video models [5, 18, 34, 40, 45] generate long videos iteratively by treating the last generated frame (clip) as the initial one in the next iteration, but suffer quality degradation due to error accumulation. Loong [38] adopts an auto-regressive language model-like strategy but is limited to low resolutions $(128 \\times 128)$ . Causal denoising [22, 30, 39, 43] gradually increases noise scales for later frames, ensuring smooth transitions by referencing clearer earlier frames. Commercial tools like Kling [1] offer clip-by-clip video extension but are constrained by context windows and unstable long-term controllability.", + "- Hierarchical methods. Hierarchical methods [6, 17, 42] use hierarchical frameworks to generate keyframes and perform interpolation or super-resolution. However, keyframe-only propagation increases information loss, degrading quality as stages grow. StoryDiffusion [48] adopts a multi-stage pipeline, generating keyframes, predicting motion, and synthesizing clips guided by both. Different from those methods, our method models correlated spatial-temporal distribution jointly for minute-long videos via condensed tokens, utilizing the fine-grained information to guide the short clip generation to achieve more natural and consistent content across the long range." + ], + "bbox": [ + 89, + 90, + 480, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 633, + 227, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CogVideoX. We use the pre-trained text-to-video diffusion model CogVideoX [41] as the foundation of our framework. CogVideoX employs a 3D causal VAE [23] to compress videos into a latent space, achieving an $8 \\times 8 \\times 4$ compression ratio along the spatial and temporal axes. During input processing, video latents are patched and concatenated with text embeddings, which are then passed through Expert Transformer blocks featuring Expert Adaptive LayerNorm (AdaLN) and 3D Full Attention, as shown in Figure 2. Text and Vision Expert AdaLN separately modulate text and video features, improving alignment. To handle large motions, CogVideoX integrates 3D Text-Video Hybrid Attention with 3D Rotary Position Embedding (RoPE) [31], effectively capturing spatial-temporal relationships.", + "bbox": [ + 89, + 657, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FIFO-Diffusion. FIFO-Diffusion [22] introduces a diagonal denoising strategy to extend a pre-trained text-to-video", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3f4ca73138b8d00e4f601a54a0cfcc682f78cca3364eaa946bc5e4ec6fa0d572.jpg", + "image_caption": [ + "Figure 2. CogVideoX architecture." + ], + "image_footnote": [], + "bbox": [ + 566, + 88, + 857, + 204 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "model from $f$ frames ( $f \\ll M$ ) to generate long videos with $M$ frames. This method progressively denoises consecutive frames with increasing noise levels. Given a time step schedule $0 = \\tau_0 < \\tau_1 < \\ldots < \\tau_f = T$ , each denoising step is defined as follows:", + "bbox": [ + 511, + 227, + 903, + 303 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ z _ {\\tau_ {0}} ^ {1}; \\dots ; z _ {\\tau_ {f - 1}} ^ {f} \\right] = \\Phi \\left(\\left[ z _ {\\tau_ {1}} ^ {1}; \\dots ; z _ {\\tau_ {f}} ^ {f} \\right], \\left[ \\tau_ {1}; \\dots ; \\tau_ {f} \\right], c; \\epsilon_ {\\theta}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 311, + 903, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The diagonal latents $\\{z_{\\tau_i}^i\\}_{i = 1}^f$ are stored in a queue. At each step, the foremost frame is dequeued once it reaches $\\tau_0 = 0$ , while a new latent at $\\tau_f$ is enqueued. This ensures that later frames with higher noise levels reference earlier frames with lower noise levels during denoising, maintaining temporal consistency and coherence throughout the long video generation process.", + "bbox": [ + 511, + 340, + 903, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. TokensGen for Long Video Generation", + "text_level": 1, + "bbox": [ + 511, + 460, + 861, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 484, + 622, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a text prompt, our framework generates a consistent minute-long video aligned with the prompt. It consists of two main components: To2V and T2To Models, as shown in Figure 3. During training, we first train To2V, a conditional short video generation model, to control spatial layout and motion based on text and video prompts. A video tokenizer extracts compact semantic tokens $z_{sem}$ from short clips, which are then fed into a diffusion transformer for guided generation. Since these tokens encode richer spatial and motion information than text prompts, they provide more accurate semantic control over individual clips. For long videos, we segment them into short clips, each tokenized to produce a sequence of semantic tokens $\\{z_{sem,i}\\}_{i=1}^{N}$ , forming a resource-efficient high-level representation of the entire video. We then train T2To, a video token transformer, to generate these long video tokens simultaneously from text prompts, ensuring long-term content consistency across clips. During inference, we first sample long video semantic tokens using T2To, then pass them to To2V to generate each clip. To ensure temporal consistency, we introduce an adaptive FIFO denoising strategy for diagonal denoising across clips.", + "bbox": [ + 511, + 507, + 906, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. To2V Model: Inner-clip content control", + "text_level": 1, + "bbox": [ + 511, + 848, + 852, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design a conditional short video generation model, To2V, guided by both text and video prompts for precise", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18199", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8a8640b752b68ab8ad69fec4c8483a4ced9b875724d0ffe7ebd8da9602701695.jpg", + "image_caption": [ + "Figure 3. Overview of the model. Left: Overall Framework for TokensGen. Right: Trainable Modules." + ], + "image_footnote": [], + "bbox": [ + 132, + 88, + 493, + 306 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2cee7ceff24f9aec0a3886634daeedb679c926f55f2f645d8aeb4212c055e2e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 89, + 836, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "content control in short video generation. To2V builds on the pre-trained text-guided video generation model CogVideoX [41] and consists of two key components: the Video Tokenizer that encodes the input video clip into compact semantic tokens, and the Cross-Attention Branch integrated with CogVideoX that enables cross-attention between semantic tokens and noisy latents.", + "bbox": [ + 88, + 344, + 482, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video Tokenizer. The Video Tokenizer consists of a 3D causal variational autoencoder (3D-VAE), a Patchify Module, and a Resampler, as illustrated on the right side of Figure 3. The 3D-VAE and the Patchify Module are inherited from CogVideoX with fixed weights. They process the input video into a set of tokens $z_{source}$ with the shape $f_s \\times h_s \\times w_s \\times c_s$ , where $f$ , $h$ , $w$ , and $c$ represent the number of frames, height, width, and channels, respectively. The Resampler compresses and resamples $z_{source}$ into a more compact representation space, as illustrated in Figure 4. It comprises a learnable latent $z_{latent}$ with the shape $f_r \\times h_r \\times w_r \\times c_s$ , four blocks of the 3D Cross-Attention Module that perform cross-attention between $z_{source}$ and $z_{latent}$ , and a Projector that transforms $z_{latent}$ into $z_{sem}$ with shape $f_r \\times h_r \\times w_r \\times c_r$ , where $f_r < f_s, h_r < h_s, w_r < w_s, c_r < c_s$ . The semantic tokens $z_{sem}$ encoded by the Video Tokenizer encapsulate high-level spatial layouts and motion information from the input video while maintaining a more compact size compared to the original video.", + "bbox": [ + 91, + 452, + 483, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-Attention Branch. To effectively incorporate these semantic tokens with CogVideoX, we add a separate Cross-Attention Branch to handle the newly added semantic conditions. This branch consists of a Semantic Token Adaptive LayerNorm (Sem AdaLN) and a 3D Cross-Attention Module, as depicted on the right side of Figure 3. The process is as follows:", + "bbox": [ + 89, + 757, + 482, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Back projection: The semantic tokens $z_{sem}$ from the Video Tokenizer are back-projected to match the num", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/28f37149a1de396f382a2932abdd2a4ffdc88b34560e6819f687799a58802b3a.jpg", + "image_caption": [ + "Figure 4. The architecture of the Resampler." + ], + "image_footnote": [], + "bbox": [ + 553, + 348, + 864, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ber of channels of the combined text-video embeddings $\\mathbf{Z}_{\\mathrm{tv}} = [\\mathbf{Z}_{\\mathrm{text}};\\mathbf{Z}_{\\mathrm{video}}]$", + "bbox": [ + 524, + 463, + 903, + 493 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Concatenation: These back-projected semantic tokens are concatenated with the text-video embeddings.", + "- Modulation: Similar to the Text and Vision AdaLN, the Sem AdaLN modulates the semantic condition embeddings to ensure better feature alignment.", + "- Attention: The modulated embeddings are passed to the 3D Text-Video Attention and the 3D Cross Attention Module to perform 3D full attention on the combined embeddings. Given the combined embeddings $\\left[\\mathbf{Z}_{\\mathrm{text}}; \\mathbf{Z}_{\\mathrm{video}}; \\mathbf{Z}_{\\mathrm{sem}}\\right]$ , the output attention results $\\left[\\mathbf{Z}_{\\mathrm{text}}^{\\prime}; \\mathbf{Z}_{\\mathrm{video}}^{\\prime}; \\mathbf{Z}_{\\mathrm{sem}}^{\\prime}\\right]$ are represented as follows:" + ], + "bbox": [ + 513, + 494, + 906, + 661 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {\\mathrm {t v}} = \\left[ \\mathbf {Z} _ {\\text {t e x t}}; \\mathbf {Z} _ {\\text {v i d e o}} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 672, + 781, + 688 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\mathbf {Z} _ {\\mathrm {t v}} \\mathbf {W} _ {\\mathbf {q}}, \\mathbf {K} = \\mathbf {Z} _ {\\mathrm {t v}} \\mathbf {W} _ {\\mathbf {k}}, \\mathbf {V} = \\mathbf {Z} _ {\\mathrm {t v}} \\mathbf {W} _ {\\mathbf {v}}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 691, + 854, + 707 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {\\mathbf {s}} = \\mathbf {Z} _ {\\mathbf {s e m}} \\mathbf {W} _ {\\mathbf {q}} ^ {\\mathrm {c}}, \\mathbf {K} _ {\\mathbf {s}} = \\mathbf {Z} _ {\\mathbf {s e m}} \\mathbf {W} _ {\\mathbf {k}} ^ {\\mathrm {c}}, \\mathbf {V} _ {\\mathbf {s}} = \\mathbf {Z} _ {\\mathbf {s e m}} \\mathbf {W} _ {\\mathbf {v}} ^ {\\mathrm {c}}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 710, + 890, + 727 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {\\mathbf {t v}} = \\mathbf {Z} _ {\\mathbf {t v}} \\mathbf {W} _ {\\mathbf {q}} ^ {\\mathbf {c}}, \\mathbf {K} _ {\\mathbf {t v}} = \\mathbf {Z} _ {\\mathbf {t v}} \\mathbf {W} _ {\\mathbf {k}} ^ {\\mathbf {c}}, \\mathbf {V} _ {\\mathbf {t v}} = \\mathbf {Z} _ {\\mathbf {t v}} \\mathbf {W} _ {\\mathbf {v}} ^ {\\mathbf {c}}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 729, + 885, + 747 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left[ \\mathbf {Z} _ {\\text {t e x t}} ^ {\\prime}, \\mathbf {Z} _ {\\text {v i d e o}} ^ {\\prime} \\right] = \\operatorname {A t t n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) + \\operatorname {A t t n} \\left(\\mathbf {Q} _ {\\mathrm {t v}}, \\mathbf {K} _ {\\mathrm {s}}, \\mathbf {V} _ {\\mathrm {s}}\\right) \\\\ = \\mathrm {s o f t m a x} (\\frac {\\mathbf {Q} (\\mathbf {K}) ^ {T}}{\\sqrt {d}}) \\mathbf {V} + \\mathrm {s o f t m a x} (\\frac {\\mathbf {Q} _ {\\mathbf {t v}} (\\mathbf {K} _ {\\mathbf {s}}) ^ {T}}{\\sqrt {d}}) \\mathbf {V} _ {\\mathbf {s}}, \\\\ \\mathbf {Z} _ {\\text {s e m}} ^ {\\prime} = \\operatorname {A t t n} \\left(\\mathbf {Q} _ {\\mathbf {s}}, \\left[ \\mathbf {K} _ {\\mathbf {t v}}; \\mathbf {K} _ {\\mathbf {s}} \\right], \\left[ \\mathbf {V} _ {\\mathbf {t v}}; \\mathbf {V} _ {\\mathbf {s}} \\right]\\right) \\\\ = \\mathrm {s o f t m a x} (\\frac {\\mathbf {Q _ {s}} ([ \\mathbf {K _ {t v}} ; \\mathbf {K _ {s}} ]) ^ {T}}{\\sqrt {d}}) [ \\mathbf {V _ {t v}}; \\mathbf {V _ {s}} ], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 750, + 908, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{\\mathbf{q}}$ , $\\mathbf{W}_{\\mathbf{k}}$ , $\\mathbf{W}_{\\mathbf{v}}$ are fixed parameters of the 3D Text-Video Attention Module inherited from CogVideoX,", + "bbox": [ + 526, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18200", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Q, K, V are the corresponding query, key, and value matrices. $\\mathbf{W}_{\\mathbf{q}}^{\\mathrm{c}}, \\mathbf{W}_{\\mathbf{k}}^{\\mathrm{c}}, \\mathbf{W}_{\\mathbf{v}}^{\\mathrm{c}}$ are trainable parameters of the 3D Cross-Attention Module, $Q_{s}, K_{s}, V_{s}$ are the query, key, and value matrices for $\\mathbf{Z}_{\\mathrm{sem}}$ , while $Q_{\\mathrm{tv}}, K_{\\mathrm{tv}}, V_{\\mathrm{tv}}$ are the query, key, and value matrices for the combined text and video embeddings $\\mathbf{Z}_{\\mathrm{tv}}$ .", + "bbox": [ + 102, + 90, + 482, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. T2To Model: Long-term content consistency", + "text_level": 1, + "bbox": [ + 89, + 196, + 467, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To learn long-term content and logic knowledge across the minute-long video, we design a video token transformer, the T2To Model, to generate the semantic tokens $\\{z_{sem,i}\\}_{i=1}^{N}$ representing the whole long video given the input text prompt. We adopt the same model structure and training strategy of CogVideoX [41] for the T2To Model, except for the following modifications:", + "bbox": [ + 89, + 220, + 483, + 325 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The model aims to generate $\\{z_{sem,i}\\}_{i=1}^{N}$ with the shape $(Nf_r) \\times h_r \\times w_r \\times c_r$ . The total number of tokens is $(Nf_r) \\times h_r \\times w_r$ .", + "- Since the number of tokens along the temporal dimension is much larger than the spatial dimensions, for 3D-RoPE, we reallocate the hidden state channel for height, width, and temporal dimension as about $10\\%$ , $10\\%$ , $80\\%$ ." + ], + "bbox": [ + 89, + 330, + 482, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Inter-clip temporal consistency", + "text_level": 1, + "bbox": [ + 89, + 449, + 364, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If each clip is denoised separately with a corresponding semantic token $z_{sem,i}$ , the model will generate a group of discontinuous clips. To achieve temporal continuity, we apply the FIFO-denoising strategy during the inference stage. Specifically, we adopt latent partitioning and lookahead denoising, like the original FIFO. However, to maintain a queue with sufficient frames at the start of denoising, FIFO pads the positions ahead of the first clip with noise-augmented first frame replications. We observe that this approach introduces artifacts in our settings, as the replicated frames deviate from the intrinsic distribution of the training domain for the video diffusion model. To address this, we propose an improved version of FIFO, named adaptive-FIFO, which incorporates an adaptive padding strategy at the beginning of the denoising process. For a latent partition containing fewer than $f_s$ frames, we denoise all frames together and update them simultaneously. For a partition with exactly $f_s$ frames, we employ lookahead denoising: the frames are denoised together, but only the noisier frames in the latter portion are updated. By better aligning the initial padding with the model's learned distribution and ensuring continuity in partially filled partitions, this approach yields smoother transitions across clips and better frame quality.", + "bbox": [ + 89, + 472, + 482, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Training strategy", + "text_level": 1, + "bbox": [ + 89, + 832, + 261, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For To2V Model, we fix the weights of the pre-trained modules of the base model, train the Resampler of the Video Tokenizer and the Cross-Attention Branch. For T2To Model,", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we initialize the model with the weights of the base model and train all the modules.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt similar training strategies with CogVideoX [41], including Multi-Resolution Frame Pack and Explicit Uniform Sampling. For T2To Model, we pack videos with different time duration into the same batch, and apply an attention mask indicating the valid frames, as well as the attention mask for loss calculation, to ensure attention module focus on the right area of the input noisy latents, an approach similar with Patch'n Pack [11]. For both To2V and T2To Model, we employ the explicit uniform sampling strategy for sampling timesteps.", + "bbox": [ + 511, + 122, + 906, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 290, + 718, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 316, + 730, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model Architecture. We employ CogVideoX-5B [41] as the base model for both To2V and T2To. In To2V, the input tokens $z_{source}$ have the shape $13 \\times 30 \\times 45 \\times 3072$ . We observed that T2To struggles to converge when $c_r$ is large (e.g., comparable to $c_s$ ), so we set the compressed semantic tokens $z_{sem}$ to have dimensions $4 \\times 8 \\times 12 \\times 16$ . For the Projector in the Resampler, we observed that a linear projection via PCA [13] effectively reduces the channel dimension without sacrificing information, as further analyzed in Sec. 5.3. Compared to the original latent shape $13 \\times 60 \\times 90 \\times 16$ , we achieve a compression ratio of approximately $3 \\times 8 \\times 8$ . Thus, we first train To2V without the channel projection and then apply PCA to the Resampler's output embeddings on 300 samples to obtain the transformation matrix. In T2To, we set the maximum number of chunks $N = 24$ . Each chunk contains 49 frames, allowing our model to process videos up to $24 \\times 49 = 1176$ frames. Dataset. We use the MiraData dataset [21], comprising long videos with structured captions. We first collect 56k videos, using their dense captions for training. For To2V Model, we randomly sample 49-frame video clips at 10 fps from these long videos as training targets. For T2To Model, we select a filtered subset of around 16k high-quality videos that are at least one minute long, primarily consisting of gameplay footage and natural landscape. We filter out videos with abrupt scene changes via PySceneDetect [8] and human evaluations. This subset ensures consistency within long videos for the training of T2To Model.", + "bbox": [ + 511, + 339, + 906, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. We adopt a progressive learning strategy for both T2To and To2V Model. For To2V Model, we first train on smaller token shapes $(4 \\times 5 \\times 7 \\times 3072)$ for 1,200 iterations, using a batch size of 72 and a learning rate of $1 \\times 10^{-3}$ . We then transition to the full token resolution $(4 \\times 8 \\times 12 \\times 3072)$ for 2,600 more iterations, initializing from the previously trained model. For T2To Model, we begin with shorter videos ( $N = 3$ chunks of 49 frames each) for 1,200 iterations, using a learning rate of $1 \\times 10^{-3}$ . Next,", + "bbox": [ + 511, + 765, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18201", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4787b9de97dc8090a3209eef164010897cd21efe9c8841ade882509e4b7d0cd6.jpg", + "image_caption": [ + "Figure 5. The qualitative comparison. We recommend readers refer to our webpage for video comparisons." + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 903, + 383 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/594dc63f6da9986c6e6f5573a5fb65590a14499c0b3798ffbe51756721c44163.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation of comparison study." + ], + "table_footnote": [], + "table_body": "
VBranchHuman Study
ModelsSubject ConsistencyBackground ConsistencyTemporal FlickeringMotion SmoothnessImaging QualityDynamic DegreeText-Visual AlignmentMotion & Content Consistency
Video-Infinity81.8090.5696.6697.6561.5831.00.0%0.69%
DiTCtrl76.7687.9695.5297.7859.2675.00.0%1.0%
AP-FIFO+CogVideoX86.2292.8994.7897.4864.1078.5724.31%22.57%
Ours84.5792.295.4198.0863.3178.9575.69%75.74%
", + "bbox": [ + 135, + 448, + 867, + 566 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c5caab729c2cbc6beb1e2d3128f63ecd7d16758f1b0d3deb5cbc56dd9e206384.jpg", + "table_caption": [ + "Table 2. Ablation study on ways of incorporating video conditions." + ], + "table_footnote": [], + "table_body": "
MethodsMSE↓SSIM↑temporal flickering↑
SR1.12e-20.6097.10
4x8x12 (w/o projection)1.24e-20.6297.58
4x5x72.84e-20.4997.57
13x5x71.54e-20.5897.54
4x8x121.24e-20.6297.58
", + "bbox": [ + 112, + 613, + 464, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we expand to long videos with up to $N = 24$ chunks, training for 5,000 iterations, with a learning rate of $3 \\times 10^{-4}$ and a batch size of 105. This progressive training helps the model converge faster for more complex, extended video generation.", + "bbox": [ + 89, + 727, + 482, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Baseline comparisons", + "text_level": 1, + "bbox": [ + 89, + 816, + 294, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparisons We evaluate our approach against several recent multi-prompt long video generation methods, including Video-Infinity [32], DiTCtrl [7], and Kling [1], as well as a baseline that adopts FIFO-Diffusion [22] on", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CogVideoX with our adaptive-padding strategy. For multiprompt methods, we use GPT-4o [4] to split the prompt into 24 segments, which are used for guiding each clip generation. FIFO and ours use the same text prompt, abbreviated as: \"a person riding a horse along a path leading to a serene river.\" The results are shown in Fig. 5. VideoInfinity produces transitions primarily through background changes while failing to capture meaningful foreground motion. The person and the horse remain essentially static within each clip, resulting in low engagement and narrative drift over longer durations. DiTCtrl demonstrates intermittently aligned keyframes but struggles to maintain smooth transitions between clips, leading to abrupt scene shifts and a disjointed storyline. Kling generates visually consistent frames but exhibits erratic motions (e.g., the subject abruptly reversing direction) and occasional discontinuities in scene composition. Such artifacts disrupt the viewing experience and deviate from the intended story arc. For FIFO (with adaptive padding on CogVideoX), we observe progressive over-saturation and abrupt changes in appear", + "bbox": [ + 511, + 589, + 906, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18202", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ance or color schemes as the video extends. These issues are especially pronounced when generating complex scenes over hundreds of frames. By contrast, our method delivers smoother motion transitions and subject representation, adherent to the prompt throughout the entire minute-long sequence. More comparison results are included in the supplementary.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative comparisons. We conduct a quantitative comparison study on 100 prompts randomly selected from the MiraData [21] test set. As reported in Tab. 1, our approach achieves the highest scores on VBench [19] for both Motion Smoothness and Dynamic Degree. We observe that certain metrics in VBench (e.g., Subject and Background Consistency, and Temporal Flickering) may assign higher scores to less dynamic videos, motivating us to conduct a user study for a more comprehensive evaluation. For the user study, we generate 12 video results for each method, with each video lasting between one and two minutes, covering categories such as humans, cars, and natural scenes. All the resulting videos are included on our webpage. To ensure unbiased feedback, videos are randomly shuffled and presented to 24 participants. They evaluate each video on two criteria: text-visual alignment and motion & content consistency. As shown on the right side of Tab. 1, our method consistently outperforms others in both dimensions, reflecting its strong long-term control capabilities. These results demonstrate that our approach effectively maintains semantic alignment with the textual prompt while preserving smooth motion and stable content over extended sequences. Additional results and details are provided in the supplementary.", + "bbox": [ + 91, + 199, + 486, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 566, + 251, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on video conditions. We investigate various strategies for incorporating video conditions into the To2V Model: (1) the condensed token shape, (2) with or without channel projection, and (3) a super-resolution-based approach. Specifically, we experiment with three token shapes $(4\\times 5\\times 7,13\\times 5\\times 7,4\\times 8\\times 12)$ , train models with and without the Projector module, and compare them against a super-resolution setting (where the low-resolution video directly serves as the condition). We also include a baseline using FIFO-Diffusion to illustrate its potential shortcomings without the conditioning process. As shown in Figure 6, the FIFO-based approach often produces inconsistent foreground and background visuals, underscoring the difficulty of preserving spatial-temporal coherence from purely latent-level guidance. Meanwhile, the super-resolution method tends to duplicate low-level color and texture cues from the source, failing to capture higher-level semantics, leading to less meaningful renderings. Comparing models with and without the Projector, we observe similar performances, demonstrating that our PCA-based projection provides a lightweight yet effective means of dimen", + "bbox": [ + 89, + 583, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sionality reduction without sacrificing image quality. Concerning the shape of the condensed tokens, the smallest variant $(4\\times 5\\times 7)$ struggles to preserve essential layout and motion patterns, resulting in less accurate re-creations of the source video. Increasing the token's temporal or spatial resolution $(13\\times 5\\times 7,4\\times 8\\times 12)$ significantly improves alignment and maintains better semantic fidelity. Among these, $(4\\times 8\\times 12)$ achieves the most favorable balance of fine-grained detail and computational efficiency, as quantitatively confirmed in Tab. 2. Overall, these ablation studies demonstrate that our token representation, combined with optional PCA-based projection, offers a robust and effective pathway to incorporate video conditions in To2V Model.", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation on FIFO. We further examine the influence of FIFO-Diffusion and our adaptive padding technique by comparing three variants: (1) omitting FIFO entirely, (2) using FIFO without adaptive padding, and (3) our full approach, incorporating both FIFO and adaptive padding. As illustrated in Fig. 7, disabling FIFO leads to abrupt scene changes between consecutive clips, producing visually inconsistent transitions where subjects may teleport or backgrounds shift abruptly. Meanwhile, removing adaptive padding induces severe artifacts in the initial frames of the video, as the model relies on duplicated frames that deviate from the training distribution. These artifacts propagate into subsequent frames, degrading overall quality. In contrast, our adaptive padding strategy aligns the padding frames with the model's distribution, preventing unnatural discontinuities at clip boundaries.", + "bbox": [ + 511, + 287, + 903, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Long Video Editing", + "text_level": 1, + "bbox": [ + 511, + 537, + 700, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Beyond generating entirely novel content, our method readily adapts to various long video editing scenarios. Specifically, the To2V Model's capacity to integrate textual prompts with source video data allows for transformations that preserve the essential structure of the original footage while injecting new semantics. We directly combine the target text prompt and the source video as input conditions to generate the edited long video. For detailed results, please refer to the supplementary and our webpage.", + "bbox": [ + 511, + 560, + 903, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 709, + 761, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We introduce TokensGen, a two-stage framework that addresses key challenges in long video generation, controlling per-clip semantics, ensuring long-term coherence, and achieving smooth transitions. The To2V Model generates short clips guided by text and video prompts, capturing fine-grained motion and content. The T2To Model transformer then leverages condensed semantic tokens to preserve long-term consistency across clips. Finally, our adaptive FIFO-Diffusion strategy overcomes boundary artifacts by maintaining temporal continuity. This pipeline efficiently scales pre-trained short video models to longer videos, enabling", + "bbox": [ + 511, + 734, + 903, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18203", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dfe808f1af597a587d775f8222e65d0e79e2bdd8e7399236058655673fb92c34.jpg", + "image_caption": [ + "Figure 6. Ablation study on methods of incorporating video conditions." + ], + "image_footnote": [], + "bbox": [ + 106, + 92, + 898, + 465 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d7e1ea812f78c751112fd7f2fb5bdd540000383a622ff5a271768b4ee308f37.jpg", + "image_caption": [ + "Figure 7. Ablation study FIFO." + ], + "image_footnote": [], + "bbox": [ + 106, + 511, + 898, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "a scalable, flexible, and resource-efficient approach to long video generation.", + "bbox": [ + 89, + 763, + 482, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Despite the effectiveness of TokensGen in maintaining long-range consistency, it does not preserve all fine-grained details. Focusing on high-level semantics, tokens may cause gradual variations in foreground objects over extended sequences (detailed in the supplementary). In complex scenes, their capacity to capture intricate spatial-temporal cues may be insufficient, requiring fine-grained", + "bbox": [ + 88, + 795, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tokenization and stronger short-term consistency strategies beyond tuning-free FIFO. Our framework is tested on a limited dataset of gameplay and landscape videos, but is scalable to larger datasets for broader applications. In future work, exploring multi-scale tokenization or hybrid representations could bolster fine-grained controllability, retaining subtle attributes while preserving the scalability and resource efficiency.", + "bbox": [ + 511, + 763, + 906, + 885 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18204", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research is supported by NTU SUG-NAP and is also supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAFICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 89, + 108, + 485, + 185 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 196, + 187, + 213 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kling. https://kling.kuaishou.com/, 2024. 1, 2, 3, 6", + "[2] Hunyuanvideo: A systematic framework for large video generative models, 2024. 2", + "[3] sora. https://openai.com/sora/, 2024.1", + "[4] Gpt-4o. chatgpt.com, 2025.6", + "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2, 3", + "[6] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. Advances in Neural Information Processing Systems, 35:31769-31781, 2022. 3", + "[7] Minghong Cai, Xiaodong Cun, Xiaoyu Li, Wenze Liu, Zhaoyang Zhang, Yong Zhang, Ying Shan, and Xiangyu Yue. Ditctrl: Exploring attention control in multi-modal diffusion transformer for tuning-free multi-prompt longer video generation. arXiv:2412.18597, 2024. 1, 2, 3, 6", + "[8] Brandon Castellano. Pyscenedetect. https://github.com/Breakthrough/PySceneDetect, 2024.5", + "[9] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models, 2024. 1, 2", + "[10] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In ICLR, 2023. 1", + "[11] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M. Alabdulmohsin, Avital Oliver, Piotr Padlewski, Alexey A. Gritsenko, Mario Lucic, and Neil Houlsby. Patch n' pack: Navit, a vision transformer for any aspect ratio and resolution. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 5", + "[12] Zhongjie Duan, Wenmeng Zhou, Cen Chen, Yaliang Li, and Weining Qian. Exvideo: Extending video diffusion models via parameter-efficient post-tuning. arXiv preprint arXiv:2406.14130, 2024. 2", + "[13] Karl Pearson F.R.S. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and" + ], + "bbox": [ + 93, + 222, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dublin Philosophical Magazine and Journal of Science, 2 (11):559-572, 1901. 5", + "[14] Ting Yao Fuchen Long, Zhaofan Qiu and Tao Mei. Videostudio: Generating consistent-content and multi-scene videos. In ECCV, 2024. 2", + "[15] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933, 2023. 2", + "[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 2", + "[17] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. 2022. 3", + "[18] Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Streamingt2v: Consistent, dynamic, and extendable long video generation from text. arXiv preprint arXiv:2403.14773, 2024. 1, 3", + "[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7", + "[20] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 2", + "[21] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions, 2024. 5, 7", + "[22] Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fico-diffusion: Generating infinite videos from text without training. arXiv preprint arXiv:2405.11473, 2024. 1, 2, 3, 6", + "[23] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3", + "[24] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024. 1, 2", + "[25] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. 2023. 2", + "[26] Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. Freelong: Training-free long video generation with spectralblend temporal attention. arXiv preprint arXiv:2407.19918, 2024. 1, 3", + "[27] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte:" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18205", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2", + "[28] Yongjia Ma, Junlin Chen, Donglin Di, Qi Xie, Lei Fan, Wei Chen, Xiaofei Gou, Na Zhao, and Xun Yang. Tuning-free long video generation via global-local collaborative diffusion. arXiv preprint arXiv:2501.05484, 2025. 1, 3", + "[29] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling, 2023. 1, 3", + "[30] David Ruhe, Jonathan Heek, Tim Salimans, and Emiel Hoogeboom. Rolling diffusion models. arXiv preprint arXiv:2402.09470, 2024. 1, 2, 3", + "[31] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 3", + "[32] Zhenxiong Tan, Xingyi Yang, Songhua Liu, and Xinchao Wang. Video-infinity: Distributed long video generation. arXiv preprint arXiv:2406.16260, 2024. 1, 2, 3, 6", + "[33] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024.2", + "[34] Ye Tian, Ling Yang, Haotian Yang, Yuan Gao, Yufan Deng, Jingmin Chen, Xintao Wang, Zhaochen Yu, Xin Tao, Pengfei Wan, Di Zhang, and Bin Cui. Videotetris: Towards compositional text-to-video generation. arXiv preprint arXiv:2406.04277, 2024. 3", + "[35] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-1-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 1, 3", + "[36] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2", + "[37] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 1, 2", + "[38] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 3", + "[39] Desai Xie, Zhan Xu, Yicong Hong, Hao Tan, Difan Liu, Feng Liu, Arie Kaufman, and Yang Zhou. Progressive autoregressive video diffusion models. arXiv preprint arXiv:2410.08151, 2024. 2, 3", + "[40] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccrafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 1, 2, 3", + "[41] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 4, 5" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023. 3", + "[43] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024. 3", + "[44] Zhengqing Yuan, Ruoxi Chen, Zhaoxu Li, Haolong Jia, Lifang He, Chi Wang, and Lichao Sun. Mora: Enabling generalist video generation via a multi-agent framework. arXiv preprint arXiv:2403.13248, 2024. 2", + "[45] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qing, Xiang Wang, Deli Zhao, and Jingren Zhou. 12vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. 2023. 1, 3", + "[46] Canyu Zhao, Mingyu Liu, Wen Wang, Jianlong Yuan, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence, 2024. 1, 2", + "[47] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2", + "[48] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Ji-ashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. NeurIPS 2024, 2024. 3", + "[49] Shaobin Zhuang, Kunchang Li, Xinyuan Chen, Yaohui Wang, Ziwei Liu, Yu Qiao, and Yali Wang. Vlogger: Make your dream a vlog. arXiv preprint arXiv:2401.09414, 2024. 2" + ], + "bbox": [ + 516, + 90, + 903, + 555 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18206", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/439118b2-5f82-4c2e-9d89-cceba716e36e_model.json b/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/439118b2-5f82-4c2e-9d89-cceba716e36e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aaebd2b351e22e7fa3b2cf992001d0c48de4be23 --- /dev/null +++ b/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/439118b2-5f82-4c2e-9d89-cceba716e36e_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.142, + 0.131, + 0.856, + 0.153 + ], + "angle": 0, + "content": "TokensGen: Harnessing Condensed Tokens for Long Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.165, + 0.351, + 0.203 + ], + "angle": 0, + "content": "Wenqi Ouyang1, Shuai Yang3," + }, + { + "type": "text", + "bbox": [ + 0.389, + 0.166, + 0.493, + 0.203 + ], + "angle": 0, + "content": "Zeqi Xiao1, Lei Yang2," + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.166, + 0.636, + 0.203 + ], + "angle": 0, + "content": "Danni Yang2, \nJianlou Si2," + }, + { + "type": "text", + "bbox": [ + 0.666, + 0.166, + 0.781, + 0.203 + ], + "angle": 0, + "content": "Yifan Zhou1, Xingang Pan1" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.209, + 0.781, + 0.246 + ], + "angle": 0, + "content": "\\(^{1}\\)S-Lab, Nanyang Technological University, \\(^{2}\\)SenseTime Research, \\(^{3}\\)Wangxuan Institute of Computer Technology, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.248, + 0.702, + 0.262 + ], + "angle": 0, + "content": "https://vicky0522.github.io/tokensgen-webpage/" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.279, + 0.892, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.453, + 0.905, + 0.482 + ], + "angle": 0, + "content": "Figure 1. Given the text prompt, TokensGen generates long videos of up to 2 minutes, maintaining consistent motions and content. Moreover, TokensGen supports zero-shot prompt-guided video-to-video editing for long videos." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.496, + 0.327, + 0.511 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.52, + 0.486, + 0.899 + ], + "angle": 0, + "content": "Generating consistent long videos is a complex challenge: while diffusion-based generative models generate visually impressive short clips, extending them to longer durations often leads to memory bottlenecks and long-term inconsistency. In this paper, we propose TokensGen, a novel two-stage framework that leverages condensed tokens to address these issues. Our method decomposes long video generation into three core tasks: (1) inner-clip semantic control, (2) long-term consistency control, and (3) inter-clip smooth transition. First, we train To2V (Token-to-Video), a short video diffusion model guided by text and video tokens, with a Video Tokenizer that condenses short clips into semantically rich tokens. Second, we introduce T2To (Text-to-Token), a video token diffusion transformer that generates all tokens at once, ensuring global consistency across clips. Finally, during inference, an adaptive FIFO-Diffusion strategy seamlessly connects adjacent clips, reducing boundary artifacts and enhancing smooth transitions. Experimental results demonstrate that our approach significantly enhances long-term temporal and content coherence without incurring prohibitive computational overhead. By leveraging condensed tokens and pre-trained short video models, our method provides a scalable, modular solution for long video generation, opening new possibilities for storytelling, cinematic production, and immersive simulations." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.496, + 0.645, + 0.511 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.907, + 0.608 + ], + "angle": 0, + "content": "Generating consistent and visually pleasing long videos remains a formidable challenge in the field of video generation. While diffusion-based methods excel at short video generation [1-3, 9, 24, 36, 37, 41, 47], extending them to longer durations is constrained by computational resources, posing significant challenges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Many approaches attempt to decompose long video generation into manageable sub-problems, using short video models without adding memory overhead. Tuning-free methods [7, 26, 28, 29, 32, 35] generate long videos using pre-trained short video models with hand-crafted techniques like noise re-scheduling, sliding window fusion, and attention manipulation, combined with multi-prompt sampling to enrich content. While ensuring high frame quality, these methods struggle with unnatural transitions due to missing long-range priors. Auto-regressive methods [1, 10, 18, 22, 30] and image-to-video approaches [5, 10, 40, 45] generate clips sequentially, achieving smooth transitions. However, they suffer from error accumulation, limited context windows, and unstable long-term controllability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Hierarchical methods, such as MovieDreamer [46], adopt a multi-stage pipeline to address long-range challenges efficiently. They generate keyframes and then synthesize short clips guided by these keyframes, producing" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18197" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "high-quality results. However, they mainly focus on multiscene generation and lack strict consistency in motion and appearance across adjacent clips." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.139, + 0.482, + 0.304 + ], + "angle": 0, + "content": "These limitations in long-term and short-term content control underscore the need for a unified, scalable solution that maintains both long-term and short-term consistency without excessive memory overhead. Therefore, we propose TokensGen, which leverages condensed video tokens to bridge short-clip generation with long-term consistency. Unlike hierarchical methods relying on keyframes generation and interpolation, or purely frame-level auto-regressive sampling, TokensGen jointly models spatial and temporal distributions for long videos through a two-stage framework, as detailed below:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.482, + 0.442 + ], + "angle": 0, + "content": "a) To2V Model (Inner-clip content control): We employ a conditional short video generation model guided by text and video tokens to produce semantically rich yet concise video segments. Built on a powerful pre-trained backbone (CogVideoX [41]), our Video Tokenizer encodes short clips into a condensed set of high-level semantic tokens. This enables robust spatial layouts and motion cues in an efficient representation space, achieving stronger per-clip semantic control than text prompts alone." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.482, + 0.581 + ], + "angle": 0, + "content": "b) T2To Model (Long-term content consistency): We train a video token diffusion transformer to generate the full set of tokens for a minute-long video from text prompts. These tokens are derived by encoding the long video clip-by-clip using the Video Tokenizer. Operating in this token space enables the T2To Model to maintain content continuity and logical coherence across clips while significantly reducing memory demands compared to raw frame modeling, preserving sufficient semantic detail for global consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.482, + 0.763 + ], + "angle": 0, + "content": "c) Adaptive FIFO-Diffusion (Inter-clip temporal smoothness): During inference, we sample long video tokens via the T2To Model and employ them to guide clip generation in the To2V Model. However, naively concatenating clips may cause boundary discontinuities, even with consistent semantic tokens. To overcome this, we propose an adaptive FIFO-Diffusion process for the To2V Model, enabling diagonal denoising of consecutive clips. This approach prevents distributional artifacts caused by naive padding or frame replication in FIFO-Diffusion [22], ensuring smoother transitions and improving the overall fidelity of the long video." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.307, + 0.482, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Compared to prior methods for long video generation, TokensGen offers several key advantages. First, by leveraging pre-trained short video models, it inherits strong knowledge priors and architectural designs, enabling a smooth transition from short clips to minute-long sequences without extensive re-engineering. Second, encoding long videos into condensed token representations significantly reduces computational overhead for minute-level generation. Third, because each component (To2V Model, T2To Model, and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "the inter-clip scheduling) operates in a clearly defined subtask, our pipeline is highly flexible. It can seamlessly integrate with other short-term control strategies (e.g., Progressive Diffusion [39], Rolling Diffusion Models [30]) or multi-prompt composition frameworks [1, 7, 32]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.905, + 0.275 + ], + "angle": 0, + "content": "In summary, TokensGen offers a scalable and resource-efficient framework for generating long videos with long-term consistency and smooth transitions, as shown in Fig. 1. By harnessing condensed tokens and powerful short video models, our approach significantly lowers the barrier to high-quality long video generation, opening new possibilities for storytelling, simulation, and beyond." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.292, + 0.655, + 0.308 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.5 + ], + "angle": 0, + "content": "Video diffusion models. Video diffusion models generate videos from text or image prompts. Early methods [5, 9, 15, 16, 36, 37, 40] extend U-Net-based image diffusion to the temporal domain. However, they struggle with motion dynamics and content richness due to separate spatial-temporal attention and limited temporal windows. Recent works [2, 24, 27, 33, 41, 47] enhance fidelity and consistency by integrating diffusion transformers with 3D full-attention to jointly model spatial-temporal correlations and improved text encoders for complex prompts. While effective for short videos, extending these models to long videos remains computationally prohibitive." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.905, + 0.606 + ], + "angle": 0, + "content": "Long video generation. Long video generation poses additional challenges for achieving content coherence, consistent dynamics, and efficient resource usage. We categorize long video generation methods into two groups: 1) those that optimize resource usage via engineering techniques or efficient model design, and 2) those that decompose long video generation into short video sub-tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.905, + 0.759 + ], + "angle": 0, + "content": "Resource usage optimization. Recent transformer-based methods [2, 24, 27, 33, 41, 47] employ 3D-VAE to compress videos. However, as noted in CogVideoX [41], excessive compression hinders 3D-VAE convergence. ExVideo [12] extends SVD [5] to 128 frames via small learnable parameters with low memory overhead. Pyramidal Flow Matching [20] reformulates diffusion into pyramid stages, enabling efficient generation of videos up to 240 frames. While effective, these methods still face challenges in scaling to much longer durations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.76, + 0.871, + 0.775 + ], + "angle": 0, + "content": "Problem decomposition via short video generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Multi-scene generation Multi-shot approaches [14, 25, 44, 49] divide long videos into segments and align them under a unified narrative, conditioning video diffusion on scene-level text or styles for coherence. MovieDreamer [46] employs a hierarchical pipeline to draft keyframes and refine shots. These methods emphasize storytelling and character consistency with relaxed demands on interclip coherence." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18198" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.211 + ], + "angle": 0, + "content": "- Tuning-free methods. Tuning-free methods extend short video generation to longer durations via hand-crafted designs, such as co-denoising [28, 35], noise re-scheduling with sliding windows [29], and attention control mechanisms [7, 26, 32]. Often paired with multi-prompt sampling for content richness, these methods lack long-range priors, leading to unnatural motion and appearance transitions over extended durations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.213, + 0.482, + 0.438 + ], + "angle": 0, + "content": "- Auto-regressive methods. Auto-regressive methods generate long videos frame-by-frame (or clip-by-clip), ensuring temporal consistency at the cost of increased inference time. Image(clip)-to-video models [5, 18, 34, 40, 45] generate long videos iteratively by treating the last generated frame (clip) as the initial one in the next iteration, but suffer quality degradation due to error accumulation. Loong [38] adopts an auto-regressive language model-like strategy but is limited to low resolutions \\((128 \\times 128)\\). Causal denoising [22, 30, 39, 43] gradually increases noise scales for later frames, ensuring smooth transitions by referencing clearer earlier frames. Commercial tools like Kling [1] offer clip-by-clip video extension but are constrained by context windows and unstable long-term controllability." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.439, + 0.482, + 0.621 + ], + "angle": 0, + "content": "- Hierarchical methods. Hierarchical methods [6, 17, 42] use hierarchical frameworks to generate keyframes and perform interpolation or super-resolution. However, keyframe-only propagation increases information loss, degrading quality as stages grow. StoryDiffusion [48] adopts a multi-stage pipeline, generating keyframes, predicting motion, and synthesizing clips guided by both. Different from those methods, our method models correlated spatial-temporal distribution jointly for minute-long videos via condensed tokens, utilizing the fine-grained information to guide the short clip generation to achieve more natural and consistent content across the long range." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.228, + 0.649 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.871 + ], + "angle": 0, + "content": "CogVideoX. We use the pre-trained text-to-video diffusion model CogVideoX [41] as the foundation of our framework. CogVideoX employs a 3D causal VAE [23] to compress videos into a latent space, achieving an \\(8 \\times 8 \\times 4\\) compression ratio along the spatial and temporal axes. During input processing, video latents are patched and concatenated with text embeddings, which are then passed through Expert Transformer blocks featuring Expert Adaptive LayerNorm (AdaLN) and 3D Full Attention, as shown in Figure 2. Text and Vision Expert AdaLN separately modulate text and video features, improving alignment. To handle large motions, CogVideoX integrates 3D Text-Video Hybrid Attention with 3D Rotary Position Embedding (RoPE) [31], effectively capturing spatial-temporal relationships." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "FIFO-Diffusion. FIFO-Diffusion [22] introduces a diagonal denoising strategy to extend a pre-trained text-to-video" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.089, + 0.858, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.207, + 0.813, + 0.221 + ], + "angle": 0, + "content": "Figure 2. CogVideoX architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.304 + ], + "angle": 0, + "content": "model from \\( f \\) frames (\\( f \\ll M \\)) to generate long videos with \\( M \\) frames. This method progressively denoises consecutive frames with increasing noise levels. Given a time step schedule \\( 0 = \\tau_0 < \\tau_1 < \\ldots < \\tau_f = T \\), each denoising step is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.312, + 0.905, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\left[ z _ {\\tau_ {0}} ^ {1}; \\dots ; z _ {\\tau_ {f - 1}} ^ {f} \\right] = \\Phi \\left(\\left[ z _ {\\tau_ {1}} ^ {1}; \\dots ; z _ {\\tau_ {f}} ^ {f} \\right], \\left[ \\tau_ {1}; \\dots ; \\tau_ {f} \\right], c; \\epsilon_ {\\theta}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.449 + ], + "angle": 0, + "content": "The diagonal latents \\(\\{z_{\\tau_i}^i\\}_{i = 1}^f\\) are stored in a queue. At each step, the foremost frame is dequeued once it reaches \\(\\tau_0 = 0\\), while a new latent at \\(\\tau_f\\) is enqueued. This ensures that later frames with higher noise levels reference earlier frames with lower noise levels during denoising, maintaining temporal consistency and coherence throughout the long video generation process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.461, + 0.862, + 0.478 + ], + "angle": 0, + "content": "4. TokensGen for Long Video Generation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.486, + 0.624, + 0.501 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.841 + ], + "angle": 0, + "content": "Given a text prompt, our framework generates a consistent minute-long video aligned with the prompt. It consists of two main components: To2V and T2To Models, as shown in Figure 3. During training, we first train To2V, a conditional short video generation model, to control spatial layout and motion based on text and video prompts. A video tokenizer extracts compact semantic tokens \\( z_{sem} \\) from short clips, which are then fed into a diffusion transformer for guided generation. Since these tokens encode richer spatial and motion information than text prompts, they provide more accurate semantic control over individual clips. For long videos, we segment them into short clips, each tokenized to produce a sequence of semantic tokens \\( \\{z_{sem,i}\\}_{i=1}^{N} \\), forming a resource-efficient high-level representation of the entire video. We then train T2To, a video token transformer, to generate these long video tokens simultaneously from text prompts, ensuring long-term content consistency across clips. During inference, we first sample long video semantic tokens using T2To, then pass them to To2V to generate each clip. To ensure temporal consistency, we introduce an adaptive FIFO denoising strategy for diagonal denoising across clips." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.854, + 0.865 + ], + "angle": 0, + "content": "4.2. To2V Model: Inner-clip content control" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We design a conditional short video generation model, To2V, guided by both text and video prompts for precise" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18199" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.089, + 0.495, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.09, + 0.838, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.193, + 0.315, + 0.802, + 0.33 + ], + "angle": 0, + "content": "Figure 3. Overview of the model. Left: Overall Framework for TokensGen. Right: Trainable Modules." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.345, + 0.483, + 0.451 + ], + "angle": 0, + "content": "content control in short video generation. To2V builds on the pre-trained text-guided video generation model CogVideoX [41] and consists of two key components: the Video Tokenizer that encodes the input video clip into compact semantic tokens, and the Cross-Attention Branch integrated with CogVideoX that enables cross-attention between semantic tokens and noisy latents." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.453, + 0.485, + 0.756 + ], + "angle": 0, + "content": "Video Tokenizer. The Video Tokenizer consists of a 3D causal variational autoencoder (3D-VAE), a Patchify Module, and a Resampler, as illustrated on the right side of Figure 3. The 3D-VAE and the Patchify Module are inherited from CogVideoX with fixed weights. They process the input video into a set of tokens \\( z_{source} \\) with the shape \\( f_s \\times h_s \\times w_s \\times c_s \\), where \\( f \\), \\( h \\), \\( w \\), and \\( c \\) represent the number of frames, height, width, and channels, respectively. The Resampler compresses and resamples \\( z_{source} \\) into a more compact representation space, as illustrated in Figure 4. It comprises a learnable latent \\( z_{latent} \\) with the shape \\( f_r \\times h_r \\times w_r \\times c_s \\), four blocks of the 3D Cross-Attention Module that perform cross-attention between \\( z_{source} \\) and \\( z_{latent} \\), and a Projector that transforms \\( z_{latent} \\) into \\( z_{sem} \\) with shape \\( f_r \\times h_r \\times w_r \\times c_r \\), where \\( f_r < f_s, h_r < h_s, w_r < w_s, c_r < c_s \\). The semantic tokens \\( z_{sem} \\) encoded by the Video Tokenizer encapsulate high-level spatial layouts and motion information from the input video while maintaining a more compact size compared to the original video." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Cross-Attention Branch. To effectively incorporate these semantic tokens with CogVideoX, we add a separate Cross-Attention Branch to handle the newly added semantic conditions. This branch consists of a Semantic Token Adaptive LayerNorm (Sem AdaLN) and a 3D Cross-Attention Module, as depicted on the right side of Figure 3. The process is as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Back projection: The semantic tokens \\( z_{sem} \\) from the Video Tokenizer are back-projected to match the num" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.349, + 0.866, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.445, + 0.842, + 0.459 + ], + "angle": 0, + "content": "Figure 4. The architecture of the Resampler." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.464, + 0.905, + 0.494 + ], + "angle": 0, + "content": "ber of channels of the combined text-video embeddings \\(\\mathbf{Z}_{\\mathrm{tv}} = [\\mathbf{Z}_{\\mathrm{text}};\\mathbf{Z}_{\\mathrm{video}}]\\)" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.495, + 0.905, + 0.524 + ], + "angle": 0, + "content": "- Concatenation: These back-projected semantic tokens are concatenated with the text-video embeddings." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.525, + 0.905, + 0.569 + ], + "angle": 0, + "content": "- Modulation: Similar to the Text and Vision AdaLN, the Sem AdaLN modulates the semantic condition embeddings to ensure better feature alignment." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.57, + 0.907, + 0.662 + ], + "angle": 0, + "content": "- Attention: The modulated embeddings are passed to the 3D Text-Video Attention and the 3D Cross Attention Module to perform 3D full attention on the combined embeddings. Given the combined embeddings \\(\\left[\\mathbf{Z}_{\\mathrm{text}}; \\mathbf{Z}_{\\mathrm{video}}; \\mathbf{Z}_{\\mathrm{sem}}\\right]\\), the output attention results \\(\\left[\\mathbf{Z}_{\\mathrm{text}}^{\\prime}; \\mathbf{Z}_{\\mathrm{video}}^{\\prime}; \\mathbf{Z}_{\\mathrm{sem}}^{\\prime}\\right]\\) are represented as follows:" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.495, + 0.907, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.673, + 0.782, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {\\mathrm {t v}} = \\left[ \\mathbf {Z} _ {\\text {t e x t}}; \\mathbf {Z} _ {\\text {v i d e o}} \\right]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.693, + 0.856, + 0.708 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\mathbf {Z} _ {\\mathrm {t v}} \\mathbf {W} _ {\\mathbf {q}}, \\mathbf {K} = \\mathbf {Z} _ {\\mathrm {t v}} \\mathbf {W} _ {\\mathbf {k}}, \\mathbf {V} = \\mathbf {Z} _ {\\mathrm {t v}} \\mathbf {W} _ {\\mathbf {v}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.711, + 0.892, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {\\mathbf {s}} = \\mathbf {Z} _ {\\mathbf {s e m}} \\mathbf {W} _ {\\mathbf {q}} ^ {\\mathrm {c}}, \\mathbf {K} _ {\\mathbf {s}} = \\mathbf {Z} _ {\\mathbf {s e m}} \\mathbf {W} _ {\\mathbf {k}} ^ {\\mathrm {c}}, \\mathbf {V} _ {\\mathbf {s}} = \\mathbf {Z} _ {\\mathbf {s e m}} \\mathbf {W} _ {\\mathbf {v}} ^ {\\mathrm {c}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.731, + 0.886, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {\\mathbf {t v}} = \\mathbf {Z} _ {\\mathbf {t v}} \\mathbf {W} _ {\\mathbf {q}} ^ {\\mathbf {c}}, \\mathbf {K} _ {\\mathbf {t v}} = \\mathbf {Z} _ {\\mathbf {t v}} \\mathbf {W} _ {\\mathbf {k}} ^ {\\mathbf {c}}, \\mathbf {V} _ {\\mathbf {t v}} = \\mathbf {Z} _ {\\mathbf {t v}} \\mathbf {W} _ {\\mathbf {v}} ^ {\\mathbf {c}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.751, + 0.91, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left[ \\mathbf {Z} _ {\\text {t e x t}} ^ {\\prime}, \\mathbf {Z} _ {\\text {v i d e o}} ^ {\\prime} \\right] = \\operatorname {A t t n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) + \\operatorname {A t t n} \\left(\\mathbf {Q} _ {\\mathrm {t v}}, \\mathbf {K} _ {\\mathrm {s}}, \\mathbf {V} _ {\\mathrm {s}}\\right) \\\\ = \\mathrm {s o f t m a x} (\\frac {\\mathbf {Q} (\\mathbf {K}) ^ {T}}{\\sqrt {d}}) \\mathbf {V} + \\mathrm {s o f t m a x} (\\frac {\\mathbf {Q} _ {\\mathbf {t v}} (\\mathbf {K} _ {\\mathbf {s}}) ^ {T}}{\\sqrt {d}}) \\mathbf {V} _ {\\mathbf {s}}, \\\\ \\mathbf {Z} _ {\\text {s e m}} ^ {\\prime} = \\operatorname {A t t n} \\left(\\mathbf {Q} _ {\\mathbf {s}}, \\left[ \\mathbf {K} _ {\\mathbf {t v}}; \\mathbf {K} _ {\\mathbf {s}} \\right], \\left[ \\mathbf {V} _ {\\mathbf {t v}}; \\mathbf {V} _ {\\mathbf {s}} \\right]\\right) \\\\ = \\mathrm {s o f t m a x} (\\frac {\\mathbf {Q _ {s}} ([ \\mathbf {K _ {t v}} ; \\mathbf {K _ {s}} ]) ^ {T}}{\\sqrt {d}}) [ \\mathbf {V _ {t v}}; \\mathbf {V _ {s}} ], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{\\mathbf{q}}\\), \\(\\mathbf{W}_{\\mathbf{k}}\\), \\(\\mathbf{W}_{\\mathbf{v}}\\) are fixed parameters of the 3D Text-Video Attention Module inherited from CogVideoX," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18200" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.091, + 0.483, + 0.184 + ], + "angle": 0, + "content": "Q, K, V are the corresponding query, key, and value matrices. \\(\\mathbf{W}_{\\mathbf{q}}^{\\mathrm{c}}, \\mathbf{W}_{\\mathbf{k}}^{\\mathrm{c}}, \\mathbf{W}_{\\mathbf{v}}^{\\mathrm{c}}\\) are trainable parameters of the 3D Cross-Attention Module, \\(Q_{s}, K_{s}, V_{s}\\) are the query, key, and value matrices for \\(\\mathbf{Z}_{\\mathrm{sem}}\\), while \\(Q_{\\mathrm{tv}}, K_{\\mathrm{tv}}, V_{\\mathrm{tv}}\\) are the query, key, and value matrices for the combined text and video embeddings \\(\\mathbf{Z}_{\\mathrm{tv}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.198, + 0.468, + 0.215 + ], + "angle": 0, + "content": "4.3. T2To Model: Long-term content consistency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.484, + 0.327 + ], + "angle": 0, + "content": "To learn long-term content and logic knowledge across the minute-long video, we design a video token transformer, the T2To Model, to generate the semantic tokens \\(\\{z_{sem,i}\\}_{i=1}^{N}\\) representing the whole long video given the input text prompt. We adopt the same model structure and training strategy of CogVideoX [41] for the T2To Model, except for the following modifications:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.332, + 0.483, + 0.375 + ], + "angle": 0, + "content": "- The model aims to generate \\(\\{z_{sem,i}\\}_{i=1}^{N}\\) with the shape \\((Nf_r) \\times h_r \\times w_r \\times c_r\\). The total number of tokens is \\((Nf_r) \\times h_r \\times w_r\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.377, + 0.482, + 0.436 + ], + "angle": 0, + "content": "- Since the number of tokens along the temporal dimension is much larger than the spatial dimensions, for 3D-RoPE, we reallocate the hidden state channel for height, width, and temporal dimension as about \\(10\\%\\), \\(10\\%\\), \\(80\\%\\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.332, + 0.483, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.45, + 0.365, + 0.466 + ], + "angle": 0, + "content": "4.4. Inter-clip temporal consistency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.82 + ], + "angle": 0, + "content": "If each clip is denoised separately with a corresponding semantic token \\( z_{sem,i} \\), the model will generate a group of discontinuous clips. To achieve temporal continuity, we apply the FIFO-denoising strategy during the inference stage. Specifically, we adopt latent partitioning and lookahead denoising, like the original FIFO. However, to maintain a queue with sufficient frames at the start of denoising, FIFO pads the positions ahead of the first clip with noise-augmented first frame replications. We observe that this approach introduces artifacts in our settings, as the replicated frames deviate from the intrinsic distribution of the training domain for the video diffusion model. To address this, we propose an improved version of FIFO, named adaptive-FIFO, which incorporates an adaptive padding strategy at the beginning of the denoising process. For a latent partition containing fewer than \\( f_s \\) frames, we denoise all frames together and update them simultaneously. For a partition with exactly \\( f_s \\) frames, we employ lookahead denoising: the frames are denoised together, but only the noisier frames in the latter portion are updated. By better aligning the initial padding with the model's learned distribution and ensuring continuity in partially filled partitions, this approach yields smoother transitions across clips and better frame quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.262, + 0.849 + ], + "angle": 0, + "content": "4.5. Training strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For To2V Model, we fix the weights of the pre-trained modules of the base model, train the Resampler of the Video Tokenizer and the Cross-Attention Branch. For T2To Model," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "we initialize the model with the weights of the base model and train all the modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.907, + 0.275 + ], + "angle": 0, + "content": "We adopt similar training strategies with CogVideoX [41], including Multi-Resolution Frame Pack and Explicit Uniform Sampling. For T2To Model, we pack videos with different time duration into the same batch, and apply an attention mask indicating the valid frames, as well as the attention mask for loss calculation, to ensure attention module focus on the right area of the input noisy latents, an approach similar with Patch'n Pack [11]. For both To2V and T2To Model, we employ the explicit uniform sampling strategy for sampling timesteps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.291, + 0.72, + 0.308 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.317, + 0.731, + 0.334 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Model Architecture. We employ CogVideoX-5B [41] as the base model for both To2V and T2To. In To2V, the input tokens \\( z_{source} \\) have the shape \\( 13 \\times 30 \\times 45 \\times 3072 \\). We observed that T2To struggles to converge when \\( c_r \\) is large (e.g., comparable to \\( c_s \\)), so we set the compressed semantic tokens \\( z_{sem} \\) to have dimensions \\( 4 \\times 8 \\times 12 \\times 16 \\). For the Projector in the Resampler, we observed that a linear projection via PCA [13] effectively reduces the channel dimension without sacrificing information, as further analyzed in Sec. 5.3. Compared to the original latent shape \\( 13 \\times 60 \\times 90 \\times 16 \\), we achieve a compression ratio of approximately \\( 3 \\times 8 \\times 8 \\). Thus, we first train To2V without the channel projection and then apply PCA to the Resampler's output embeddings on 300 samples to obtain the transformation matrix. In T2To, we set the maximum number of chunks \\( N = 24 \\). Each chunk contains 49 frames, allowing our model to process videos up to \\( 24 \\times 49 = 1176 \\) frames. Dataset. We use the MiraData dataset [21], comprising long videos with structured captions. We first collect 56k videos, using their dense captions for training. For To2V Model, we randomly sample 49-frame video clips at 10 fps from these long videos as training targets. For T2To Model, we select a filtered subset of around 16k high-quality videos that are at least one minute long, primarily consisting of gameplay footage and natural landscape. We filter out videos with abrupt scene changes via PySceneDetect [8] and human evaluations. This subset ensures consistency within long videos for the training of T2To Model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Training Details. We adopt a progressive learning strategy for both T2To and To2V Model. For To2V Model, we first train on smaller token shapes \\((4 \\times 5 \\times 7 \\times 3072)\\) for 1,200 iterations, using a batch size of 72 and a learning rate of \\(1 \\times 10^{-3}\\). We then transition to the full token resolution \\((4 \\times 8 \\times 12 \\times 3072)\\) for 2,600 more iterations, initializing from the previously trained model. For T2To Model, we begin with shorter videos (\\(N = 3\\) chunks of 49 frames each) for 1,200 iterations, using a learning rate of \\(1 \\times 10^{-3}\\). Next," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18201" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.093, + 0.904, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.181, + 0.395, + 0.816, + 0.41 + ], + "angle": 0, + "content": "Figure 5. The qualitative comparison. We recommend readers refer to our webpage for video comparisons." + }, + { + "type": "table_caption", + "bbox": [ + 0.339, + 0.424, + 0.658, + 0.438 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation of comparison study." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.449, + 0.868, + 0.567 + ], + "angle": 0, + "content": "
VBranchHuman Study
ModelsSubject ConsistencyBackground ConsistencyTemporal FlickeringMotion SmoothnessImaging QualityDynamic DegreeText-Visual AlignmentMotion & Content Consistency
Video-Infinity81.8090.5696.6697.6561.5831.00.0%0.69%
DiTCtrl76.7687.9695.5297.7859.2675.00.0%1.0%
AP-FIFO+CogVideoX86.2292.8994.7897.4864.1078.5724.31%22.57%
Ours84.5792.295.4198.0863.3178.9575.69%75.74%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.589, + 0.483, + 0.604 + ], + "angle": 0, + "content": "Table 2. Ablation study on ways of incorporating video conditions." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.614, + 0.465, + 0.724 + ], + "angle": 0, + "content": "
MethodsMSE↓SSIM↑temporal flickering↑
SR1.12e-20.6097.10
4x8x12 (w/o projection)1.24e-20.6297.58
4x5x72.84e-20.4997.57
13x5x71.54e-20.5897.54
4x8x121.24e-20.6297.58
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.805 + ], + "angle": 0, + "content": "we expand to long videos with up to \\( N = 24 \\) chunks, training for 5,000 iterations, with a learning rate of \\( 3 \\times 10^{-4} \\) and a batch size of 105. This progressive training helps the model converge faster for more complex, extended video generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.295, + 0.833 + ], + "angle": 0, + "content": "5.2. Baseline comparisons" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Qualitative comparisons We evaluate our approach against several recent multi-prompt long video generation methods, including Video-Infinity [32], DiTCtrl [7], and Kling [1], as well as a baseline that adopts FIFO-Diffusion [22] on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.895 + ], + "angle": 0, + "content": "CogVideoX with our adaptive-padding strategy. For multiprompt methods, we use GPT-4o [4] to split the prompt into 24 segments, which are used for guiding each clip generation. FIFO and ours use the same text prompt, abbreviated as: \"a person riding a horse along a path leading to a serene river.\" The results are shown in Fig. 5. VideoInfinity produces transitions primarily through background changes while failing to capture meaningful foreground motion. The person and the horse remain essentially static within each clip, resulting in low engagement and narrative drift over longer durations. DiTCtrl demonstrates intermittently aligned keyframes but struggles to maintain smooth transitions between clips, leading to abrupt scene shifts and a disjointed storyline. Kling generates visually consistent frames but exhibits erratic motions (e.g., the subject abruptly reversing direction) and occasional discontinuities in scene composition. Such artifacts disrupt the viewing experience and deviate from the intended story arc. For FIFO (with adaptive padding on CogVideoX), we observe progressive over-saturation and abrupt changes in appear" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18202" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "ance or color schemes as the video extends. These issues are especially pronounced when generating complex scenes over hundreds of frames. By contrast, our method delivers smoother motion transitions and subject representation, adherent to the prompt throughout the entire minute-long sequence. More comparison results are included in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.2, + 0.487, + 0.562 + ], + "angle": 0, + "content": "Quantitative comparisons. We conduct a quantitative comparison study on 100 prompts randomly selected from the MiraData [21] test set. As reported in Tab. 1, our approach achieves the highest scores on VBench [19] for both Motion Smoothness and Dynamic Degree. We observe that certain metrics in VBench (e.g., Subject and Background Consistency, and Temporal Flickering) may assign higher scores to less dynamic videos, motivating us to conduct a user study for a more comprehensive evaluation. For the user study, we generate 12 video results for each method, with each video lasting between one and two minutes, covering categories such as humans, cars, and natural scenes. All the resulting videos are included on our webpage. To ensure unbiased feedback, videos are randomly shuffled and presented to 24 participants. They evaluate each video on two criteria: text-visual alignment and motion & content consistency. As shown on the right side of Tab. 1, our method consistently outperforms others in both dimensions, reflecting its strong long-term control capabilities. These results demonstrate that our approach effectively maintains semantic alignment with the textual prompt while preserving smooth motion and stable content over extended sequences. Additional results and details are provided in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.567, + 0.253, + 0.582 + ], + "angle": 0, + "content": "5.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablation on video conditions. We investigate various strategies for incorporating video conditions into the To2V Model: (1) the condensed token shape, (2) with or without channel projection, and (3) a super-resolution-based approach. Specifically, we experiment with three token shapes \\((4\\times 5\\times 7,13\\times 5\\times 7,4\\times 8\\times 12)\\), train models with and without the Projector module, and compare them against a super-resolution setting (where the low-resolution video directly serves as the condition). We also include a baseline using FIFO-Diffusion to illustrate its potential shortcomings without the conditioning process. As shown in Figure 6, the FIFO-based approach often produces inconsistent foreground and background visuals, underscoring the difficulty of preserving spatial-temporal coherence from purely latent-level guidance. Meanwhile, the super-resolution method tends to duplicate low-level color and texture cues from the source, failing to capture higher-level semantics, leading to less meaningful renderings. Comparing models with and without the Projector, we observe similar performances, demonstrating that our PCA-based projection provides a lightweight yet effective means of dimen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "sionality reduction without sacrificing image quality. Concerning the shape of the condensed tokens, the smallest variant \\((4\\times 5\\times 7)\\) struggles to preserve essential layout and motion patterns, resulting in less accurate re-creations of the source video. Increasing the token's temporal or spatial resolution \\((13\\times 5\\times 7,4\\times 8\\times 12)\\) significantly improves alignment and maintains better semantic fidelity. Among these, \\((4\\times 8\\times 12)\\) achieves the most favorable balance of fine-grained detail and computational efficiency, as quantitatively confirmed in Tab. 2. Overall, these ablation studies demonstrate that our token representation, combined with optional PCA-based projection, offers a robust and effective pathway to incorporate video conditions in To2V Model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.529 + ], + "angle": 0, + "content": "Ablation on FIFO. We further examine the influence of FIFO-Diffusion and our adaptive padding technique by comparing three variants: (1) omitting FIFO entirely, (2) using FIFO without adaptive padding, and (3) our full approach, incorporating both FIFO and adaptive padding. As illustrated in Fig. 7, disabling FIFO leads to abrupt scene changes between consecutive clips, producing visually inconsistent transitions where subjects may teleport or backgrounds shift abruptly. Meanwhile, removing adaptive padding induces severe artifacts in the initial frames of the video, as the model relies on duplicated frames that deviate from the training distribution. These artifacts propagate into subsequent frames, degrading overall quality. In contrast, our adaptive padding strategy aligns the padding frames with the model's distribution, preventing unnatural discontinuities at clip boundaries." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.538, + 0.702, + 0.556 + ], + "angle": 0, + "content": "5.4. Long Video Editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.905, + 0.697 + ], + "angle": 0, + "content": "Beyond generating entirely novel content, our method readily adapts to various long video editing scenarios. Specifically, the To2V Model's capacity to integrate textual prompts with source video data allows for transformations that preserve the essential structure of the original footage while injecting new semantics. We directly combine the target text prompt and the source video as input conditions to generate the edited long video. For detailed results, please refer to the supplementary and our webpage." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.763, + 0.726 + ], + "angle": 0, + "content": "6. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.903 + ], + "angle": 0, + "content": "We introduce TokensGen, a two-stage framework that addresses key challenges in long video generation, controlling per-clip semantics, ensuring long-term coherence, and achieving smooth transitions. The To2V Model generates short clips guided by text and video prompts, capturing fine-grained motion and content. The T2To Model transformer then leverages condensed semantic tokens to preserve long-term consistency across clips. Finally, our adaptive FIFO-Diffusion strategy overcomes boundary artifacts by maintaining temporal continuity. This pipeline efficiently scales pre-trained short video models to longer videos, enabling" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18203" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.093, + 0.899, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.286, + 0.48, + 0.712, + 0.496 + ], + "angle": 0, + "content": "Figure 6. Ablation study on methods of incorporating video conditions." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.512, + 0.899, + 0.71 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.403, + 0.724, + 0.595, + 0.739 + ], + "angle": 0, + "content": "Figure 7. Ablation study FIFO." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.796 + ], + "angle": 0, + "content": "a scalable, flexible, and resource-efficient approach to long video generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Despite the effectiveness of TokensGen in maintaining long-range consistency, it does not preserve all fine-grained details. Focusing on high-level semantics, tokens may cause gradual variations in foreground objects over extended sequences (detailed in the supplementary). In complex scenes, their capacity to capture intricate spatial-temporal cues may be insufficient, requiring fine-grained" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.886 + ], + "angle": 0, + "content": "tokenization and stronger short-term consistency strategies beyond tuning-free FIFO. Our framework is tested on a limited dataset of gameplay and landscape videos, but is scalable to larger datasets for broader applications. In future work, exploring multi-scale tokenization or hybrid representations could bolster fine-grained controllability, retaining subtle attributes while preserving the scalability and resource efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18204" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.486, + 0.186 + ], + "angle": 0, + "content": "This research is supported by NTU SUG-NAP and is also supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAFICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.198, + 0.188, + 0.214 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.223, + 0.482, + 0.249 + ], + "angle": 0, + "content": "[1] Kling. https://kling.kuaishou.com/, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.252, + 0.483, + 0.279 + ], + "angle": 0, + "content": "[2] Hunyuanvideo: A systematic framework for large video generative models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.282, + 0.425, + 0.295 + ], + "angle": 0, + "content": "[3] sora. https://openai.com/sora/, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.296, + 0.329, + 0.31 + ], + "angle": 0, + "content": "[4] Gpt-4o. chatgpt.com, 2025.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.311, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.381, + 0.483, + 0.45 + ], + "angle": 0, + "content": "[6] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. Advances in Neural Information Processing Systems, 35:31769-31781, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.483, + 0.521 + ], + "angle": 0, + "content": "[7] Minghong Cai, Xiaodong Cun, Xiaoyu Li, Wenze Liu, Zhaoyang Zhang, Yong Zhang, Ying Shan, and Xiangyu Yue. Ditctrl: Exploring attention control in multi-modal diffusion transformer for tuning-free multi-prompt longer video generation. arXiv:2412.18597, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.48, + 0.549 + ], + "angle": 0, + "content": "[8] Brandon Castellano. Pyscenedetect. https://github.com/Breakthrough/PySceneDetect, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[9] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[10] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[11] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M. Alabdulmohsin, Avital Oliver, Piotr Padlewski, Alexey A. Gritsenko, Mario Lucic, and Neil Houlsby. Patch n' pack: Navit, a vision transformer for any aspect ratio and resolution. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[12] Zhongjie Duan, Wenmeng Zhou, Cen Chen, Yaliang Li, and Weining Qian. Exvideo: Extending video diffusion models via parameter-efficient post-tuning. arXiv preprint arXiv:2406.14130, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Karl Pearson F.R.S. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.223, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.906, + 0.12 + ], + "angle": 0, + "content": "Dublin Philosophical Magazine and Journal of Science, 2 (11):559-572, 1901. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[14] Ting Yao Fuchen Long, Zhaofan Qiu and Tao Mei. Videostudio: Generating consistent-content and multi-scene videos. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[15] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[17] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[18] Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Streamingt2v: Consistent, dynamic, and extendable long video generation from text. arXiv preprint arXiv:2403.14773, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[20] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[21] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.631, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[22] Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fico-diffusion: Generating infinite videos from text without training. arXiv preprint arXiv:2405.11473, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[23] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[24] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[25] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[26] Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. Freelong: Training-free long video generation with spectralblend temporal attention. arXiv preprint arXiv:2407.19918, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18205" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[28] Yongjia Ma, Junlin Chen, Donglin Di, Qi Xie, Lei Fan, Wei Chen, Xiaofei Gou, Na Zhao, and Xun Yang. Tuning-free long video generation via global-local collaborative diffusion. arXiv preprint arXiv:2501.05484, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.223 + ], + "angle": 0, + "content": "[29] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.224, + 0.482, + 0.266 + ], + "angle": 0, + "content": "[30] David Ruhe, Jonathan Heek, Tim Salimans, and Emiel Hoogeboom. Rolling diffusion models. arXiv preprint arXiv:2402.09470, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.268, + 0.482, + 0.322 + ], + "angle": 0, + "content": "[31] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.326, + 0.482, + 0.368 + ], + "angle": 0, + "content": "[32] Zhenxiong Tan, Xingyi Yang, Songhua Liu, and Xinchao Wang. Video-infinity: Distributed long video generation. arXiv preprint arXiv:2406.16260, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.37, + 0.482, + 0.397 + ], + "angle": 0, + "content": "[33] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[34] Ye Tian, Ling Yang, Haotian Yang, Yuan Gao, Yufan Deng, Jingmin Chen, Xintao Wang, Zhaochen Yu, Xin Tao, Pengfei Wan, Di Zhang, and Bin Cui. Videotetris: Towards compositional text-to-video generation. arXiv preprint arXiv:2406.04277, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.471, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[35] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-1-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.529, + 0.482, + 0.583 + ], + "angle": 0, + "content": "[36] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.587, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[37] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[38] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[39] Desai Xie, Zhan Xu, Yicong Hong, Hao Tan, Difan Liu, Feng Liu, Arie Kaufman, and Yang Zhou. Progressive autoregressive video diffusion models. arXiv preprint arXiv:2410.08151, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[40] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccrafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[41] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[42] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[43] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[44] Zhengqing Yuan, Ruoxi Chen, Zhaoxu Li, Haolong Jia, Lifang He, Chi Wang, and Lichao Sun. Mora: Enabling generalist video generation via a multi-agent framework. arXiv preprint arXiv:2403.13248, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[45] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qing, Xiang Wang, Deli Zhao, and Jingren Zhou. 12vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[46] Canyu Zhao, Mingyu Liu, Wen Wang, Jianlong Yuan, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[47] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[48] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Ji-ashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. NeurIPS 2024, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.556 + ], + "angle": 0, + "content": "[49] Shaobin Zhuang, Kunchang Li, Xinyuan Chen, Yaohui Wang, Ziwei Liu, Yu Qiao, and Yali Wang. Vlogger: Make your dream a vlog. arXiv preprint arXiv:2401.09414, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18206" + } + ] +] \ No newline at end of file diff --git a/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/439118b2-5f82-4c2e-9d89-cceba716e36e_origin.pdf b/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/439118b2-5f82-4c2e-9d89-cceba716e36e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..967d98781ac7da6902b2ba6c487717af2d10d5e1 --- /dev/null +++ b/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/439118b2-5f82-4c2e-9d89-cceba716e36e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:af95defe319e5050c76fd5efcbbdb6f9706542dc93590a2bd11ed977a8c0b88e +size 5518224 diff --git a/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/full.md b/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d7563eab369da0e9f0e8e6345357ce13378e6419 --- /dev/null +++ b/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/full.md @@ -0,0 +1,271 @@ +# TokensGen: Harnessing Condensed Tokens for Long Video Generation + +Wenqi Ouyang1, Shuai Yang3, + +Zeqi Xiao1, Lei Yang2, + +Danni Yang2, +Jianlou Si2, + +Yifan Zhou1, Xingang Pan1 + +$^{1}$ S-Lab, Nanyang Technological University, $^{2}$ SenseTime Research, $^{3}$ Wangxuan Institute of Computer Technology, Peking University + +https://vicky0522.github.io/tokensgen-webpage/ + +![](images/cb603e98ca6d897a1e47aac36b49d88bd29bdc1ebab90182f66d88ea280bfd63.jpg) +Figure 1. Given the text prompt, TokensGen generates long videos of up to 2 minutes, maintaining consistent motions and content. Moreover, TokensGen supports zero-shot prompt-guided video-to-video editing for long videos. + +# Abstract + +Generating consistent long videos is a complex challenge: while diffusion-based generative models generate visually impressive short clips, extending them to longer durations often leads to memory bottlenecks and long-term inconsistency. In this paper, we propose TokensGen, a novel two-stage framework that leverages condensed tokens to address these issues. Our method decomposes long video generation into three core tasks: (1) inner-clip semantic control, (2) long-term consistency control, and (3) inter-clip smooth transition. First, we train To2V (Token-to-Video), a short video diffusion model guided by text and video tokens, with a Video Tokenizer that condenses short clips into semantically rich tokens. Second, we introduce T2To (Text-to-Token), a video token diffusion transformer that generates all tokens at once, ensuring global consistency across clips. Finally, during inference, an adaptive FIFO-Diffusion strategy seamlessly connects adjacent clips, reducing boundary artifacts and enhancing smooth transitions. Experimental results demonstrate that our approach significantly enhances long-term temporal and content coherence without incurring prohibitive computational overhead. By leveraging condensed tokens and pre-trained short video models, our method provides a scalable, modular solution for long video generation, opening new possibilities for storytelling, cinematic production, and immersive simulations. + +# 1. Introduction + +Generating consistent and visually pleasing long videos remains a formidable challenge in the field of video generation. While diffusion-based methods excel at short video generation [1-3, 9, 24, 36, 37, 41, 47], extending them to longer durations is constrained by computational resources, posing significant challenges. + +Many approaches attempt to decompose long video generation into manageable sub-problems, using short video models without adding memory overhead. Tuning-free methods [7, 26, 28, 29, 32, 35] generate long videos using pre-trained short video models with hand-crafted techniques like noise re-scheduling, sliding window fusion, and attention manipulation, combined with multi-prompt sampling to enrich content. While ensuring high frame quality, these methods struggle with unnatural transitions due to missing long-range priors. Auto-regressive methods [1, 10, 18, 22, 30] and image-to-video approaches [5, 10, 40, 45] generate clips sequentially, achieving smooth transitions. However, they suffer from error accumulation, limited context windows, and unstable long-term controllability. + +Hierarchical methods, such as MovieDreamer [46], adopt a multi-stage pipeline to address long-range challenges efficiently. They generate keyframes and then synthesize short clips guided by these keyframes, producing + +high-quality results. However, they mainly focus on multiscene generation and lack strict consistency in motion and appearance across adjacent clips. + +These limitations in long-term and short-term content control underscore the need for a unified, scalable solution that maintains both long-term and short-term consistency without excessive memory overhead. Therefore, we propose TokensGen, which leverages condensed video tokens to bridge short-clip generation with long-term consistency. Unlike hierarchical methods relying on keyframes generation and interpolation, or purely frame-level auto-regressive sampling, TokensGen jointly models spatial and temporal distributions for long videos through a two-stage framework, as detailed below: + +a) To2V Model (Inner-clip content control): We employ a conditional short video generation model guided by text and video tokens to produce semantically rich yet concise video segments. Built on a powerful pre-trained backbone (CogVideoX [41]), our Video Tokenizer encodes short clips into a condensed set of high-level semantic tokens. This enables robust spatial layouts and motion cues in an efficient representation space, achieving stronger per-clip semantic control than text prompts alone. +b) T2To Model (Long-term content consistency): We train a video token diffusion transformer to generate the full set of tokens for a minute-long video from text prompts. These tokens are derived by encoding the long video clip-by-clip using the Video Tokenizer. Operating in this token space enables the T2To Model to maintain content continuity and logical coherence across clips while significantly reducing memory demands compared to raw frame modeling, preserving sufficient semantic detail for global consistency. +c) Adaptive FIFO-Diffusion (Inter-clip temporal smoothness): During inference, we sample long video tokens via the T2To Model and employ them to guide clip generation in the To2V Model. However, naively concatenating clips may cause boundary discontinuities, even with consistent semantic tokens. To overcome this, we propose an adaptive FIFO-Diffusion process for the To2V Model, enabling diagonal denoising of consecutive clips. This approach prevents distributional artifacts caused by naive padding or frame replication in FIFO-Diffusion [22], ensuring smoother transitions and improving the overall fidelity of the long video. + +Compared to prior methods for long video generation, TokensGen offers several key advantages. First, by leveraging pre-trained short video models, it inherits strong knowledge priors and architectural designs, enabling a smooth transition from short clips to minute-long sequences without extensive re-engineering. Second, encoding long videos into condensed token representations significantly reduces computational overhead for minute-level generation. Third, because each component (To2V Model, T2To Model, and + +the inter-clip scheduling) operates in a clearly defined subtask, our pipeline is highly flexible. It can seamlessly integrate with other short-term control strategies (e.g., Progressive Diffusion [39], Rolling Diffusion Models [30]) or multi-prompt composition frameworks [1, 7, 32]. + +In summary, TokensGen offers a scalable and resource-efficient framework for generating long videos with long-term consistency and smooth transitions, as shown in Fig. 1. By harnessing condensed tokens and powerful short video models, our approach significantly lowers the barrier to high-quality long video generation, opening new possibilities for storytelling, simulation, and beyond. + +# 2. Related Work + +Video diffusion models. Video diffusion models generate videos from text or image prompts. Early methods [5, 9, 15, 16, 36, 37, 40] extend U-Net-based image diffusion to the temporal domain. However, they struggle with motion dynamics and content richness due to separate spatial-temporal attention and limited temporal windows. Recent works [2, 24, 27, 33, 41, 47] enhance fidelity and consistency by integrating diffusion transformers with 3D full-attention to jointly model spatial-temporal correlations and improved text encoders for complex prompts. While effective for short videos, extending these models to long videos remains computationally prohibitive. + +Long video generation. Long video generation poses additional challenges for achieving content coherence, consistent dynamics, and efficient resource usage. We categorize long video generation methods into two groups: 1) those that optimize resource usage via engineering techniques or efficient model design, and 2) those that decompose long video generation into short video sub-tasks. + +Resource usage optimization. Recent transformer-based methods [2, 24, 27, 33, 41, 47] employ 3D-VAE to compress videos. However, as noted in CogVideoX [41], excessive compression hinders 3D-VAE convergence. ExVideo [12] extends SVD [5] to 128 frames via small learnable parameters with low memory overhead. Pyramidal Flow Matching [20] reformulates diffusion into pyramid stages, enabling efficient generation of videos up to 240 frames. While effective, these methods still face challenges in scaling to much longer durations. + +# Problem decomposition via short video generation. + +- Multi-scene generation Multi-shot approaches [14, 25, 44, 49] divide long videos into segments and align them under a unified narrative, conditioning video diffusion on scene-level text or styles for coherence. MovieDreamer [46] employs a hierarchical pipeline to draft keyframes and refine shots. These methods emphasize storytelling and character consistency with relaxed demands on interclip coherence. + +- Tuning-free methods. Tuning-free methods extend short video generation to longer durations via hand-crafted designs, such as co-denoising [28, 35], noise re-scheduling with sliding windows [29], and attention control mechanisms [7, 26, 32]. Often paired with multi-prompt sampling for content richness, these methods lack long-range priors, leading to unnatural motion and appearance transitions over extended durations. +- Auto-regressive methods. Auto-regressive methods generate long videos frame-by-frame (or clip-by-clip), ensuring temporal consistency at the cost of increased inference time. Image(clip)-to-video models [5, 18, 34, 40, 45] generate long videos iteratively by treating the last generated frame (clip) as the initial one in the next iteration, but suffer quality degradation due to error accumulation. Loong [38] adopts an auto-regressive language model-like strategy but is limited to low resolutions $(128 \times 128)$ . Causal denoising [22, 30, 39, 43] gradually increases noise scales for later frames, ensuring smooth transitions by referencing clearer earlier frames. Commercial tools like Kling [1] offer clip-by-clip video extension but are constrained by context windows and unstable long-term controllability. +- Hierarchical methods. Hierarchical methods [6, 17, 42] use hierarchical frameworks to generate keyframes and perform interpolation or super-resolution. However, keyframe-only propagation increases information loss, degrading quality as stages grow. StoryDiffusion [48] adopts a multi-stage pipeline, generating keyframes, predicting motion, and synthesizing clips guided by both. Different from those methods, our method models correlated spatial-temporal distribution jointly for minute-long videos via condensed tokens, utilizing the fine-grained information to guide the short clip generation to achieve more natural and consistent content across the long range. + +# 3. Preliminaries + +CogVideoX. We use the pre-trained text-to-video diffusion model CogVideoX [41] as the foundation of our framework. CogVideoX employs a 3D causal VAE [23] to compress videos into a latent space, achieving an $8 \times 8 \times 4$ compression ratio along the spatial and temporal axes. During input processing, video latents are patched and concatenated with text embeddings, which are then passed through Expert Transformer blocks featuring Expert Adaptive LayerNorm (AdaLN) and 3D Full Attention, as shown in Figure 2. Text and Vision Expert AdaLN separately modulate text and video features, improving alignment. To handle large motions, CogVideoX integrates 3D Text-Video Hybrid Attention with 3D Rotary Position Embedding (RoPE) [31], effectively capturing spatial-temporal relationships. + +FIFO-Diffusion. FIFO-Diffusion [22] introduces a diagonal denoising strategy to extend a pre-trained text-to-video + +![](images/3f4ca73138b8d00e4f601a54a0cfcc682f78cca3364eaa946bc5e4ec6fa0d572.jpg) +Figure 2. CogVideoX architecture. + +model from $f$ frames ( $f \ll M$ ) to generate long videos with $M$ frames. This method progressively denoises consecutive frames with increasing noise levels. Given a time step schedule $0 = \tau_0 < \tau_1 < \ldots < \tau_f = T$ , each denoising step is defined as follows: + +$$ +\left[ z _ {\tau_ {0}} ^ {1}; \dots ; z _ {\tau_ {f - 1}} ^ {f} \right] = \Phi \left(\left[ z _ {\tau_ {1}} ^ {1}; \dots ; z _ {\tau_ {f}} ^ {f} \right], \left[ \tau_ {1}; \dots ; \tau_ {f} \right], c; \epsilon_ {\theta}\right). \tag {1} +$$ + +The diagonal latents $\{z_{\tau_i}^i\}_{i = 1}^f$ are stored in a queue. At each step, the foremost frame is dequeued once it reaches $\tau_0 = 0$ , while a new latent at $\tau_f$ is enqueued. This ensures that later frames with higher noise levels reference earlier frames with lower noise levels during denoising, maintaining temporal consistency and coherence throughout the long video generation process. + +# 4. TokensGen for Long Video Generation + +# 4.1. Overview + +Given a text prompt, our framework generates a consistent minute-long video aligned with the prompt. It consists of two main components: To2V and T2To Models, as shown in Figure 3. During training, we first train To2V, a conditional short video generation model, to control spatial layout and motion based on text and video prompts. A video tokenizer extracts compact semantic tokens $z_{sem}$ from short clips, which are then fed into a diffusion transformer for guided generation. Since these tokens encode richer spatial and motion information than text prompts, they provide more accurate semantic control over individual clips. For long videos, we segment them into short clips, each tokenized to produce a sequence of semantic tokens $\{z_{sem,i}\}_{i=1}^{N}$ , forming a resource-efficient high-level representation of the entire video. We then train T2To, a video token transformer, to generate these long video tokens simultaneously from text prompts, ensuring long-term content consistency across clips. During inference, we first sample long video semantic tokens using T2To, then pass them to To2V to generate each clip. To ensure temporal consistency, we introduce an adaptive FIFO denoising strategy for diagonal denoising across clips. + +# 4.2. To2V Model: Inner-clip content control + +We design a conditional short video generation model, To2V, guided by both text and video prompts for precise + +![](images/8a8640b752b68ab8ad69fec4c8483a4ced9b875724d0ffe7ebd8da9602701695.jpg) +Figure 3. Overview of the model. Left: Overall Framework for TokensGen. Right: Trainable Modules. + +![](images/2cee7ceff24f9aec0a3886634daeedb679c926f55f2f645d8aeb4212c055e2e7.jpg) + +content control in short video generation. To2V builds on the pre-trained text-guided video generation model CogVideoX [41] and consists of two key components: the Video Tokenizer that encodes the input video clip into compact semantic tokens, and the Cross-Attention Branch integrated with CogVideoX that enables cross-attention between semantic tokens and noisy latents. + +Video Tokenizer. The Video Tokenizer consists of a 3D causal variational autoencoder (3D-VAE), a Patchify Module, and a Resampler, as illustrated on the right side of Figure 3. The 3D-VAE and the Patchify Module are inherited from CogVideoX with fixed weights. They process the input video into a set of tokens $z_{source}$ with the shape $f_s \times h_s \times w_s \times c_s$ , where $f$ , $h$ , $w$ , and $c$ represent the number of frames, height, width, and channels, respectively. The Resampler compresses and resamples $z_{source}$ into a more compact representation space, as illustrated in Figure 4. It comprises a learnable latent $z_{latent}$ with the shape $f_r \times h_r \times w_r \times c_s$ , four blocks of the 3D Cross-Attention Module that perform cross-attention between $z_{source}$ and $z_{latent}$ , and a Projector that transforms $z_{latent}$ into $z_{sem}$ with shape $f_r \times h_r \times w_r \times c_r$ , where $f_r < f_s, h_r < h_s, w_r < w_s, c_r < c_s$ . The semantic tokens $z_{sem}$ encoded by the Video Tokenizer encapsulate high-level spatial layouts and motion information from the input video while maintaining a more compact size compared to the original video. + +Cross-Attention Branch. To effectively incorporate these semantic tokens with CogVideoX, we add a separate Cross-Attention Branch to handle the newly added semantic conditions. This branch consists of a Semantic Token Adaptive LayerNorm (Sem AdaLN) and a 3D Cross-Attention Module, as depicted on the right side of Figure 3. The process is as follows: + +- Back projection: The semantic tokens $z_{sem}$ from the Video Tokenizer are back-projected to match the num + +![](images/28f37149a1de396f382a2932abdd2a4ffdc88b34560e6819f687799a58802b3a.jpg) +Figure 4. The architecture of the Resampler. + +ber of channels of the combined text-video embeddings $\mathbf{Z}_{\mathrm{tv}} = [\mathbf{Z}_{\mathrm{text}};\mathbf{Z}_{\mathrm{video}}]$ + +- Concatenation: These back-projected semantic tokens are concatenated with the text-video embeddings. +- Modulation: Similar to the Text and Vision AdaLN, the Sem AdaLN modulates the semantic condition embeddings to ensure better feature alignment. +- Attention: The modulated embeddings are passed to the 3D Text-Video Attention and the 3D Cross Attention Module to perform 3D full attention on the combined embeddings. Given the combined embeddings $\left[\mathbf{Z}_{\mathrm{text}}; \mathbf{Z}_{\mathrm{video}}; \mathbf{Z}_{\mathrm{sem}}\right]$ , the output attention results $\left[\mathbf{Z}_{\mathrm{text}}^{\prime}; \mathbf{Z}_{\mathrm{video}}^{\prime}; \mathbf{Z}_{\mathrm{sem}}^{\prime}\right]$ are represented as follows: + +$$ +\mathbf {Z} _ {\mathrm {t v}} = \left[ \mathbf {Z} _ {\text {t e x t}}; \mathbf {Z} _ {\text {v i d e o}} \right] +$$ + +$$ +\mathbf {Q} = \mathbf {Z} _ {\mathrm {t v}} \mathbf {W} _ {\mathbf {q}}, \mathbf {K} = \mathbf {Z} _ {\mathrm {t v}} \mathbf {W} _ {\mathbf {k}}, \mathbf {V} = \mathbf {Z} _ {\mathrm {t v}} \mathbf {W} _ {\mathbf {v}} +$$ + +$$ +\mathbf {Q} _ {\mathbf {s}} = \mathbf {Z} _ {\mathbf {s e m}} \mathbf {W} _ {\mathbf {q}} ^ {\mathrm {c}}, \mathbf {K} _ {\mathbf {s}} = \mathbf {Z} _ {\mathbf {s e m}} \mathbf {W} _ {\mathbf {k}} ^ {\mathrm {c}}, \mathbf {V} _ {\mathbf {s}} = \mathbf {Z} _ {\mathbf {s e m}} \mathbf {W} _ {\mathbf {v}} ^ {\mathrm {c}} +$$ + +$$ +\mathbf {Q} _ {\mathbf {t v}} = \mathbf {Z} _ {\mathbf {t v}} \mathbf {W} _ {\mathbf {q}} ^ {\mathbf {c}}, \mathbf {K} _ {\mathbf {t v}} = \mathbf {Z} _ {\mathbf {t v}} \mathbf {W} _ {\mathbf {k}} ^ {\mathbf {c}}, \mathbf {V} _ {\mathbf {t v}} = \mathbf {Z} _ {\mathbf {t v}} \mathbf {W} _ {\mathbf {v}} ^ {\mathbf {c}} +$$ + +$$ +\begin{array}{l} \left[ \mathbf {Z} _ {\text {t e x t}} ^ {\prime}, \mathbf {Z} _ {\text {v i d e o}} ^ {\prime} \right] = \operatorname {A t t n} (\mathbf {Q}, \mathbf {K}, \mathbf {V}) + \operatorname {A t t n} \left(\mathbf {Q} _ {\mathrm {t v}}, \mathbf {K} _ {\mathrm {s}}, \mathbf {V} _ {\mathrm {s}}\right) \\ = \mathrm {s o f t m a x} (\frac {\mathbf {Q} (\mathbf {K}) ^ {T}}{\sqrt {d}}) \mathbf {V} + \mathrm {s o f t m a x} (\frac {\mathbf {Q} _ {\mathbf {t v}} (\mathbf {K} _ {\mathbf {s}}) ^ {T}}{\sqrt {d}}) \mathbf {V} _ {\mathbf {s}}, \\ \mathbf {Z} _ {\text {s e m}} ^ {\prime} = \operatorname {A t t n} \left(\mathbf {Q} _ {\mathbf {s}}, \left[ \mathbf {K} _ {\mathbf {t v}}; \mathbf {K} _ {\mathbf {s}} \right], \left[ \mathbf {V} _ {\mathbf {t v}}; \mathbf {V} _ {\mathbf {s}} \right]\right) \\ = \mathrm {s o f t m a x} (\frac {\mathbf {Q _ {s}} ([ \mathbf {K _ {t v}} ; \mathbf {K _ {s}} ]) ^ {T}}{\sqrt {d}}) [ \mathbf {V _ {t v}}; \mathbf {V _ {s}} ], \\ \end{array} +$$ + +where $\mathbf{W}_{\mathbf{q}}$ , $\mathbf{W}_{\mathbf{k}}$ , $\mathbf{W}_{\mathbf{v}}$ are fixed parameters of the 3D Text-Video Attention Module inherited from CogVideoX, + +Q, K, V are the corresponding query, key, and value matrices. $\mathbf{W}_{\mathbf{q}}^{\mathrm{c}}, \mathbf{W}_{\mathbf{k}}^{\mathrm{c}}, \mathbf{W}_{\mathbf{v}}^{\mathrm{c}}$ are trainable parameters of the 3D Cross-Attention Module, $Q_{s}, K_{s}, V_{s}$ are the query, key, and value matrices for $\mathbf{Z}_{\mathrm{sem}}$ , while $Q_{\mathrm{tv}}, K_{\mathrm{tv}}, V_{\mathrm{tv}}$ are the query, key, and value matrices for the combined text and video embeddings $\mathbf{Z}_{\mathrm{tv}}$ . + +# 4.3. T2To Model: Long-term content consistency + +To learn long-term content and logic knowledge across the minute-long video, we design a video token transformer, the T2To Model, to generate the semantic tokens $\{z_{sem,i}\}_{i=1}^{N}$ representing the whole long video given the input text prompt. We adopt the same model structure and training strategy of CogVideoX [41] for the T2To Model, except for the following modifications: + +- The model aims to generate $\{z_{sem,i}\}_{i=1}^{N}$ with the shape $(Nf_r) \times h_r \times w_r \times c_r$ . The total number of tokens is $(Nf_r) \times h_r \times w_r$ . +- Since the number of tokens along the temporal dimension is much larger than the spatial dimensions, for 3D-RoPE, we reallocate the hidden state channel for height, width, and temporal dimension as about $10\%$ , $10\%$ , $80\%$ . + +# 4.4. Inter-clip temporal consistency + +If each clip is denoised separately with a corresponding semantic token $z_{sem,i}$ , the model will generate a group of discontinuous clips. To achieve temporal continuity, we apply the FIFO-denoising strategy during the inference stage. Specifically, we adopt latent partitioning and lookahead denoising, like the original FIFO. However, to maintain a queue with sufficient frames at the start of denoising, FIFO pads the positions ahead of the first clip with noise-augmented first frame replications. We observe that this approach introduces artifacts in our settings, as the replicated frames deviate from the intrinsic distribution of the training domain for the video diffusion model. To address this, we propose an improved version of FIFO, named adaptive-FIFO, which incorporates an adaptive padding strategy at the beginning of the denoising process. For a latent partition containing fewer than $f_s$ frames, we denoise all frames together and update them simultaneously. For a partition with exactly $f_s$ frames, we employ lookahead denoising: the frames are denoised together, but only the noisier frames in the latter portion are updated. By better aligning the initial padding with the model's learned distribution and ensuring continuity in partially filled partitions, this approach yields smoother transitions across clips and better frame quality. + +# 4.5. Training strategy + +For To2V Model, we fix the weights of the pre-trained modules of the base model, train the Resampler of the Video Tokenizer and the Cross-Attention Branch. For T2To Model, + +we initialize the model with the weights of the base model and train all the modules. + +We adopt similar training strategies with CogVideoX [41], including Multi-Resolution Frame Pack and Explicit Uniform Sampling. For T2To Model, we pack videos with different time duration into the same batch, and apply an attention mask indicating the valid frames, as well as the attention mask for loss calculation, to ensure attention module focus on the right area of the input noisy latents, an approach similar with Patch'n Pack [11]. For both To2V and T2To Model, we employ the explicit uniform sampling strategy for sampling timesteps. + +# 5. Experimental Results + +# 5.1. Implementation Details + +Model Architecture. We employ CogVideoX-5B [41] as the base model for both To2V and T2To. In To2V, the input tokens $z_{source}$ have the shape $13 \times 30 \times 45 \times 3072$ . We observed that T2To struggles to converge when $c_r$ is large (e.g., comparable to $c_s$ ), so we set the compressed semantic tokens $z_{sem}$ to have dimensions $4 \times 8 \times 12 \times 16$ . For the Projector in the Resampler, we observed that a linear projection via PCA [13] effectively reduces the channel dimension without sacrificing information, as further analyzed in Sec. 5.3. Compared to the original latent shape $13 \times 60 \times 90 \times 16$ , we achieve a compression ratio of approximately $3 \times 8 \times 8$ . Thus, we first train To2V without the channel projection and then apply PCA to the Resampler's output embeddings on 300 samples to obtain the transformation matrix. In T2To, we set the maximum number of chunks $N = 24$ . Each chunk contains 49 frames, allowing our model to process videos up to $24 \times 49 = 1176$ frames. Dataset. We use the MiraData dataset [21], comprising long videos with structured captions. We first collect 56k videos, using their dense captions for training. For To2V Model, we randomly sample 49-frame video clips at 10 fps from these long videos as training targets. For T2To Model, we select a filtered subset of around 16k high-quality videos that are at least one minute long, primarily consisting of gameplay footage and natural landscape. We filter out videos with abrupt scene changes via PySceneDetect [8] and human evaluations. This subset ensures consistency within long videos for the training of T2To Model. + +Training Details. We adopt a progressive learning strategy for both T2To and To2V Model. For To2V Model, we first train on smaller token shapes $(4 \times 5 \times 7 \times 3072)$ for 1,200 iterations, using a batch size of 72 and a learning rate of $1 \times 10^{-3}$ . We then transition to the full token resolution $(4 \times 8 \times 12 \times 3072)$ for 2,600 more iterations, initializing from the previously trained model. For T2To Model, we begin with shorter videos ( $N = 3$ chunks of 49 frames each) for 1,200 iterations, using a learning rate of $1 \times 10^{-3}$ . Next, + +![](images/4787b9de97dc8090a3209eef164010897cd21efe9c8841ade882509e4b7d0cd6.jpg) +Figure 5. The qualitative comparison. We recommend readers refer to our webpage for video comparisons. + +Table 1. Quantitative evaluation of comparison study. + +
VBranchHuman Study
ModelsSubject ConsistencyBackground ConsistencyTemporal FlickeringMotion SmoothnessImaging QualityDynamic DegreeText-Visual AlignmentMotion & Content Consistency
Video-Infinity81.8090.5696.6697.6561.5831.00.0%0.69%
DiTCtrl76.7687.9695.5297.7859.2675.00.0%1.0%
AP-FIFO+CogVideoX86.2292.8994.7897.4864.1078.5724.31%22.57%
Ours84.5792.295.4198.0863.3178.9575.69%75.74%
+ +Table 2. Ablation study on ways of incorporating video conditions. + +
MethodsMSE↓SSIM↑temporal flickering↑
SR1.12e-20.6097.10
4x8x12 (w/o projection)1.24e-20.6297.58
4x5x72.84e-20.4997.57
13x5x71.54e-20.5897.54
4x8x121.24e-20.6297.58
+ +we expand to long videos with up to $N = 24$ chunks, training for 5,000 iterations, with a learning rate of $3 \times 10^{-4}$ and a batch size of 105. This progressive training helps the model converge faster for more complex, extended video generation. + +# 5.2. Baseline comparisons + +Qualitative comparisons We evaluate our approach against several recent multi-prompt long video generation methods, including Video-Infinity [32], DiTCtrl [7], and Kling [1], as well as a baseline that adopts FIFO-Diffusion [22] on + +CogVideoX with our adaptive-padding strategy. For multiprompt methods, we use GPT-4o [4] to split the prompt into 24 segments, which are used for guiding each clip generation. FIFO and ours use the same text prompt, abbreviated as: "a person riding a horse along a path leading to a serene river." The results are shown in Fig. 5. VideoInfinity produces transitions primarily through background changes while failing to capture meaningful foreground motion. The person and the horse remain essentially static within each clip, resulting in low engagement and narrative drift over longer durations. DiTCtrl demonstrates intermittently aligned keyframes but struggles to maintain smooth transitions between clips, leading to abrupt scene shifts and a disjointed storyline. Kling generates visually consistent frames but exhibits erratic motions (e.g., the subject abruptly reversing direction) and occasional discontinuities in scene composition. Such artifacts disrupt the viewing experience and deviate from the intended story arc. For FIFO (with adaptive padding on CogVideoX), we observe progressive over-saturation and abrupt changes in appear + +ance or color schemes as the video extends. These issues are especially pronounced when generating complex scenes over hundreds of frames. By contrast, our method delivers smoother motion transitions and subject representation, adherent to the prompt throughout the entire minute-long sequence. More comparison results are included in the supplementary. + +Quantitative comparisons. We conduct a quantitative comparison study on 100 prompts randomly selected from the MiraData [21] test set. As reported in Tab. 1, our approach achieves the highest scores on VBench [19] for both Motion Smoothness and Dynamic Degree. We observe that certain metrics in VBench (e.g., Subject and Background Consistency, and Temporal Flickering) may assign higher scores to less dynamic videos, motivating us to conduct a user study for a more comprehensive evaluation. For the user study, we generate 12 video results for each method, with each video lasting between one and two minutes, covering categories such as humans, cars, and natural scenes. All the resulting videos are included on our webpage. To ensure unbiased feedback, videos are randomly shuffled and presented to 24 participants. They evaluate each video on two criteria: text-visual alignment and motion & content consistency. As shown on the right side of Tab. 1, our method consistently outperforms others in both dimensions, reflecting its strong long-term control capabilities. These results demonstrate that our approach effectively maintains semantic alignment with the textual prompt while preserving smooth motion and stable content over extended sequences. Additional results and details are provided in the supplementary. + +# 5.3. Ablation studies + +Ablation on video conditions. We investigate various strategies for incorporating video conditions into the To2V Model: (1) the condensed token shape, (2) with or without channel projection, and (3) a super-resolution-based approach. Specifically, we experiment with three token shapes $(4\times 5\times 7,13\times 5\times 7,4\times 8\times 12)$ , train models with and without the Projector module, and compare them against a super-resolution setting (where the low-resolution video directly serves as the condition). We also include a baseline using FIFO-Diffusion to illustrate its potential shortcomings without the conditioning process. As shown in Figure 6, the FIFO-based approach often produces inconsistent foreground and background visuals, underscoring the difficulty of preserving spatial-temporal coherence from purely latent-level guidance. Meanwhile, the super-resolution method tends to duplicate low-level color and texture cues from the source, failing to capture higher-level semantics, leading to less meaningful renderings. Comparing models with and without the Projector, we observe similar performances, demonstrating that our PCA-based projection provides a lightweight yet effective means of dimen + +sionality reduction without sacrificing image quality. Concerning the shape of the condensed tokens, the smallest variant $(4\times 5\times 7)$ struggles to preserve essential layout and motion patterns, resulting in less accurate re-creations of the source video. Increasing the token's temporal or spatial resolution $(13\times 5\times 7,4\times 8\times 12)$ significantly improves alignment and maintains better semantic fidelity. Among these, $(4\times 8\times 12)$ achieves the most favorable balance of fine-grained detail and computational efficiency, as quantitatively confirmed in Tab. 2. Overall, these ablation studies demonstrate that our token representation, combined with optional PCA-based projection, offers a robust and effective pathway to incorporate video conditions in To2V Model. + +Ablation on FIFO. We further examine the influence of FIFO-Diffusion and our adaptive padding technique by comparing three variants: (1) omitting FIFO entirely, (2) using FIFO without adaptive padding, and (3) our full approach, incorporating both FIFO and adaptive padding. As illustrated in Fig. 7, disabling FIFO leads to abrupt scene changes between consecutive clips, producing visually inconsistent transitions where subjects may teleport or backgrounds shift abruptly. Meanwhile, removing adaptive padding induces severe artifacts in the initial frames of the video, as the model relies on duplicated frames that deviate from the training distribution. These artifacts propagate into subsequent frames, degrading overall quality. In contrast, our adaptive padding strategy aligns the padding frames with the model's distribution, preventing unnatural discontinuities at clip boundaries. + +# 5.4. Long Video Editing + +Beyond generating entirely novel content, our method readily adapts to various long video editing scenarios. Specifically, the To2V Model's capacity to integrate textual prompts with source video data allows for transformations that preserve the essential structure of the original footage while injecting new semantics. We directly combine the target text prompt and the source video as input conditions to generate the edited long video. For detailed results, please refer to the supplementary and our webpage. + +# 6. Conclusion and Discussion + +We introduce TokensGen, a two-stage framework that addresses key challenges in long video generation, controlling per-clip semantics, ensuring long-term coherence, and achieving smooth transitions. The To2V Model generates short clips guided by text and video prompts, capturing fine-grained motion and content. The T2To Model transformer then leverages condensed semantic tokens to preserve long-term consistency across clips. Finally, our adaptive FIFO-Diffusion strategy overcomes boundary artifacts by maintaining temporal continuity. This pipeline efficiently scales pre-trained short video models to longer videos, enabling + +![](images/dfe808f1af597a587d775f8222e65d0e79e2bdd8e7399236058655673fb92c34.jpg) +Figure 6. Ablation study on methods of incorporating video conditions. + +![](images/8d7e1ea812f78c751112fd7f2fb5bdd540000383a622ff5a271768b4ee308f37.jpg) +Figure 7. Ablation study FIFO. + +a scalable, flexible, and resource-efficient approach to long video generation. + +Despite the effectiveness of TokensGen in maintaining long-range consistency, it does not preserve all fine-grained details. Focusing on high-level semantics, tokens may cause gradual variations in foreground objects over extended sequences (detailed in the supplementary). In complex scenes, their capacity to capture intricate spatial-temporal cues may be insufficient, requiring fine-grained + +tokenization and stronger short-term consistency strategies beyond tuning-free FIFO. Our framework is tested on a limited dataset of gameplay and landscape videos, but is scalable to larger datasets for broader applications. In future work, exploring multi-scale tokenization or hybrid representations could bolster fine-grained controllability, retaining subtle attributes while preserving the scalability and resource efficiency. + +# Acknowledgements + +This research is supported by NTU SUG-NAP and is also supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAFICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] Kling. https://kling.kuaishou.com/, 2024. 1, 2, 3, 6 +[2] Hunyuanvideo: A systematic framework for large video generative models, 2024. 2 +[3] sora. https://openai.com/sora/, 2024.1 +[4] Gpt-4o. chatgpt.com, 2025.6 +[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2, 3 +[6] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. Advances in Neural Information Processing Systems, 35:31769-31781, 2022. 3 +[7] Minghong Cai, Xiaodong Cun, Xiaoyu Li, Wenze Liu, Zhaoyang Zhang, Yong Zhang, Ying Shan, and Xiangyu Yue. Ditctrl: Exploring attention control in multi-modal diffusion transformer for tuning-free multi-prompt longer video generation. arXiv:2412.18597, 2024. 1, 2, 3, 6 +[8] Brandon Castellano. Pyscenedetect. https://github.com/Breakthrough/PySceneDetect, 2024.5 +[9] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models, 2024. 1, 2 +[10] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In ICLR, 2023. 1 +[11] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M. Alabdulmohsin, Avital Oliver, Piotr Padlewski, Alexey A. Gritsenko, Mario Lucic, and Neil Houlsby. Patch n' pack: Navit, a vision transformer for any aspect ratio and resolution. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 5 +[12] Zhongjie Duan, Wenmeng Zhou, Cen Chen, Yaliang Li, and Weining Qian. Exvideo: Extending video diffusion models via parameter-efficient post-tuning. arXiv preprint arXiv:2406.14130, 2024. 2 +[13] Karl Pearson F.R.S. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and + +Dublin Philosophical Magazine and Journal of Science, 2 (11):559-572, 1901. 5 +[14] Ting Yao Fuchen Long, Zhaofan Qiu and Tao Mei. Videostudio: Generating consistent-content and multi-scene videos. In ECCV, 2024. 2 +[15] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933, 2023. 2 +[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 2 +[17] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. 2022. 3 +[18] Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Streamingt2v: Consistent, dynamic, and extendable long video generation from text. arXiv preprint arXiv:2403.14773, 2024. 1, 3 +[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7 +[20] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 2 +[21] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions, 2024. 5, 7 +[22] Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fico-diffusion: Generating infinite videos from text without training. arXiv preprint arXiv:2405.11473, 2024. 1, 2, 3, 6 +[23] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3 +[24] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024. 1, 2 +[25] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. 2023. 2 +[26] Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. Freelong: Training-free long video generation with spectralblend temporal attention. arXiv preprint arXiv:2407.19918, 2024. 1, 3 +[27] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: + +Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2 +[28] Yongjia Ma, Junlin Chen, Donglin Di, Qi Xie, Lei Fan, Wei Chen, Xiaofei Gou, Na Zhao, and Xun Yang. Tuning-free long video generation via global-local collaborative diffusion. arXiv preprint arXiv:2501.05484, 2025. 1, 3 +[29] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling, 2023. 1, 3 +[30] David Ruhe, Jonathan Heek, Tim Salimans, and Emiel Hoogeboom. Rolling diffusion models. arXiv preprint arXiv:2402.09470, 2024. 1, 2, 3 +[31] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 3 +[32] Zhenxiong Tan, Xingyi Yang, Songhua Liu, and Xinchao Wang. Video-infinity: Distributed long video generation. arXiv preprint arXiv:2406.16260, 2024. 1, 2, 3, 6 +[33] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024.2 +[34] Ye Tian, Ling Yang, Haotian Yang, Yuan Gao, Yufan Deng, Jingmin Chen, Xintao Wang, Zhaochen Yu, Xin Tao, Pengfei Wan, Di Zhang, and Bin Cui. Videotetris: Towards compositional text-to-video generation. arXiv preprint arXiv:2406.04277, 2024. 3 +[35] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-1-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 1, 3 +[36] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2 +[37] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 1, 2 +[38] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 3 +[39] Desai Xie, Zhan Xu, Yicong Hong, Hao Tan, Difan Liu, Feng Liu, Arie Kaufman, and Yang Zhou. Progressive autoregressive video diffusion models. arXiv preprint arXiv:2410.08151, 2024. 2, 3 +[40] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccrafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 1, 2, 3 +[41] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 4, 5 + +[42] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023. 3 +[43] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024. 3 +[44] Zhengqing Yuan, Ruoxi Chen, Zhaoxu Li, Haolong Jia, Lifang He, Chi Wang, and Lichao Sun. Mora: Enabling generalist video generation via a multi-agent framework. arXiv preprint arXiv:2403.13248, 2024. 2 +[45] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qing, Xiang Wang, Deli Zhao, and Jingren Zhou. 12vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. 2023. 1, 3 +[46] Canyu Zhao, Mingyu Liu, Wen Wang, Jianlong Yuan, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence, 2024. 1, 2 +[47] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2 +[48] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Ji-ashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. NeurIPS 2024, 2024. 3 +[49] Shaobin Zhuang, Kunchang Li, Xinyuan Chen, Yaohui Wang, Ziwei Liu, Yu Qiao, and Yali Wang. Vlogger: Make your dream a vlog. arXiv preprint arXiv:2401.09414, 2024. 2 \ No newline at end of file diff --git a/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/images.zip b/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..08d6787e93da6973d9fd3c15e26354998fdf1ec8 --- /dev/null +++ b/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3d0513034768c03132a3f69d2425215fcfaac392aefbf1882ea149f27d0ca9d1 +size 850466 diff --git a/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/layout.json b/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3a4fe6db4f5d388704f570590f75b657e9cfe1bc --- /dev/null +++ b/2025/TokensGen_ Harnessing Condensed Tokens for Long Video Generation/layout.json @@ -0,0 +1,7472 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 523, + 121 + ], + "type": "text", + "content": "TokensGen: Harnessing Condensed Tokens for Long Video Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 130, + 214, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 130, + 214, + 160 + ], + "spans": [ + { + "bbox": [ + 132, + 130, + 214, + 160 + ], + "type": "text", + "content": "Wenqi Ouyang1, Shuai Yang3," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 238, + 131, + 301, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 238, + 131, + 301, + 160 + ], + "spans": [ + { + "bbox": [ + 238, + 131, + 301, + 160 + ], + "type": "text", + "content": "Zeqi Xiao1, Lei Yang2," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 321, + 131, + 389, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 131, + 389, + 160 + ], + "spans": [ + { + "bbox": [ + 321, + 131, + 389, + 160 + ], + "type": "text", + "content": "Danni Yang2, \nJianlou Si2," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 407, + 131, + 477, + 160 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 131, + 477, + 160 + ], + "spans": [ + { + "bbox": [ + 407, + 131, + 477, + 160 + ], + "type": "text", + "content": "Yifan Zhou1, Xingang Pan1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 132, + 165, + 477, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 165, + 477, + 194 + ], + "spans": [ + { + "bbox": [ + 132, + 165, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 165, + 477, + 194 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University, " + }, + { + "bbox": [ + 132, + 165, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 165, + 477, + 194 + ], + "type": "text", + "content": "SenseTime Research, " + }, + { + "bbox": [ + 132, + 165, + 477, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 165, + 477, + 194 + ], + "type": "text", + "content": "Wangxuan Institute of Computer Technology, Peking University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 178, + 196, + 429, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 196, + 429, + 207 + ], + "spans": [ + { + "bbox": [ + 178, + 196, + 429, + 207 + ], + "type": "text", + "content": "https://vicky0522.github.io/tokensgen-webpage/" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 57, + 220, + 545, + 349 + ], + "blocks": [ + { + "bbox": [ + 57, + 220, + 545, + 349 + ], + "lines": [ + { + "bbox": [ + 57, + 220, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 545, + 349 + ], + "type": "image", + "image_path": "cb603e98ca6d897a1e47aac36b49d88bd29bdc1ebab90182f66d88ea280bfd63.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 358, + 553, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 358, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 553, + 381 + ], + "type": "text", + "content": "Figure 1. Given the text prompt, TokensGen generates long videos of up to 2 minutes, maintaining consistent motions and content. Moreover, TokensGen supports zero-shot prompt-guided video-to-video editing for long videos." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 392, + 200, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 392, + 200, + 404 + ], + "spans": [ + { + "bbox": [ + 151, + 392, + 200, + 404 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 411, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 297, + 712 + ], + "type": "text", + "content": "Generating consistent long videos is a complex challenge: while diffusion-based generative models generate visually impressive short clips, extending them to longer durations often leads to memory bottlenecks and long-term inconsistency. In this paper, we propose TokensGen, a novel two-stage framework that leverages condensed tokens to address these issues. Our method decomposes long video generation into three core tasks: (1) inner-clip semantic control, (2) long-term consistency control, and (3) inter-clip smooth transition. First, we train To2V (Token-to-Video), a short video diffusion model guided by text and video tokens, with a Video Tokenizer that condenses short clips into semantically rich tokens. Second, we introduce T2To (Text-to-Token), a video token diffusion transformer that generates all tokens at once, ensuring global consistency across clips. Finally, during inference, an adaptive FIFO-Diffusion strategy seamlessly connects adjacent clips, reducing boundary artifacts and enhancing smooth transitions. Experimental results demonstrate that our approach significantly enhances long-term temporal and content coherence without incurring prohibitive computational overhead. By leveraging condensed tokens and pre-trained short video models, our method provides a scalable, modular solution for long video generation, opening new possibilities for storytelling, cinematic production, and immersive simulations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 392, + 394, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 392, + 394, + 404 + ], + "spans": [ + { + "bbox": [ + 315, + 392, + 394, + 404 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 409, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 481 + ], + "type": "text", + "content": "Generating consistent and visually pleasing long videos remains a formidable challenge in the field of video generation. While diffusion-based methods excel at short video generation [1-3, 9, 24, 36, 37, 41, 47], extending them to longer durations is constrained by computational resources, posing significant challenges." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 483, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 663 + ], + "type": "text", + "content": "Many approaches attempt to decompose long video generation into manageable sub-problems, using short video models without adding memory overhead. Tuning-free methods [7, 26, 28, 29, 32, 35] generate long videos using pre-trained short video models with hand-crafted techniques like noise re-scheduling, sliding window fusion, and attention manipulation, combined with multi-prompt sampling to enrich content. While ensuring high frame quality, these methods struggle with unnatural transitions due to missing long-range priors. Auto-regressive methods [1, 10, 18, 22, 30] and image-to-video approaches [5, 10, 40, 45] generate clips sequentially, achieving smooth transitions. However, they suffer from error accumulation, limited context windows, and unstable long-term controllability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Hierarchical methods, such as MovieDreamer [46], adopt a multi-stage pipeline to address long-range challenges efficiently. They generate keyframes and then synthesize short clips guided by these keyframes, producing" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18197" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "high-quality results. However, they mainly focus on multiscene generation and lack strict consistency in motion and appearance across adjacent clips." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 110, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 110, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 110, + 294, + 240 + ], + "type": "text", + "content": "These limitations in long-term and short-term content control underscore the need for a unified, scalable solution that maintains both long-term and short-term consistency without excessive memory overhead. Therefore, we propose TokensGen, which leverages condensed video tokens to bridge short-clip generation with long-term consistency. Unlike hierarchical methods relying on keyframes generation and interpolation, or purely frame-level auto-regressive sampling, TokensGen jointly models spatial and temporal distributions for long videos through a two-stage framework, as detailed below:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 294, + 604 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 243, + 294, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 294, + 350 + ], + "type": "text", + "content": "a) To2V Model (Inner-clip content control): We employ a conditional short video generation model guided by text and video tokens to produce semantically rich yet concise video segments. Built on a powerful pre-trained backbone (CogVideoX [41]), our Video Tokenizer encodes short clips into a condensed set of high-level semantic tokens. This enables robust spatial layouts and motion cues in an efficient representation space, achieving stronger per-clip semantic control than text prompts alone." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 352, + 294, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 294, + 460 + ], + "type": "text", + "content": "b) T2To Model (Long-term content consistency): We train a video token diffusion transformer to generate the full set of tokens for a minute-long video from text prompts. These tokens are derived by encoding the long video clip-by-clip using the Video Tokenizer. Operating in this token space enables the T2To Model to maintain content continuity and logical coherence across clips while significantly reducing memory demands compared to raw frame modeling, preserving sufficient semantic detail for global consistency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 461, + 294, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 294, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 294, + 604 + ], + "type": "text", + "content": "c) Adaptive FIFO-Diffusion (Inter-clip temporal smoothness): During inference, we sample long video tokens via the T2To Model and employ them to guide clip generation in the To2V Model. However, naively concatenating clips may cause boundary discontinuities, even with consistent semantic tokens. To overcome this, we propose an adaptive FIFO-Diffusion process for the To2V Model, enabling diagonal denoising of consecutive clips. This approach prevents distributional artifacts caused by naive padding or frame replication in FIFO-Diffusion [22], ensuring smoother transitions and improving the overall fidelity of the long video." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 294, + 713 + ], + "type": "text", + "content": "Compared to prior methods for long video generation, TokensGen offers several key advantages. First, by leveraging pre-trained short video models, it inherits strong knowledge priors and architectural designs, enabling a smooth transition from short clips to minute-long sequences without extensive re-engineering. Second, encoding long videos into condensed token representations significantly reduces computational overhead for minute-level generation. Third, because each component (To2V Model, T2To Model, and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "the inter-clip scheduling) operates in a clearly defined subtask, our pipeline is highly flexible. It can seamlessly integrate with other short-term control strategies (e.g., Progressive Diffusion [39], Rolling Diffusion Models [30]) or multi-prompt composition frameworks [1, 7, 32]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 217 + ], + "type": "text", + "content": "In summary, TokensGen offers a scalable and resource-efficient framework for generating long videos with long-term consistency and smooth transitions, as shown in Fig. 1. By harnessing condensed tokens and powerful short video models, our approach significantly lowers the barrier to high-quality long video generation, opening new possibilities for storytelling, simulation, and beyond." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 231, + 400, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 231, + 400, + 243 + ], + "spans": [ + { + "bbox": [ + 314, + 231, + 400, + 243 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 251, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 396 + ], + "type": "text", + "content": "Video diffusion models. Video diffusion models generate videos from text or image prompts. Early methods [5, 9, 15, 16, 36, 37, 40] extend U-Net-based image diffusion to the temporal domain. However, they struggle with motion dynamics and content richness due to separate spatial-temporal attention and limited temporal windows. Recent works [2, 24, 27, 33, 41, 47] enhance fidelity and consistency by integrating diffusion transformers with 3D full-attention to jointly model spatial-temporal correlations and improved text encoders for complex prompts. While effective for short videos, extending these models to long videos remains computationally prohibitive." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 396, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 553, + 479 + ], + "type": "text", + "content": "Long video generation. Long video generation poses additional challenges for achieving content coherence, consistent dynamics, and efficient resource usage. We categorize long video generation methods into two groups: 1) those that optimize resource usage via engineering techniques or efficient model design, and 2) those that decompose long video generation into short video sub-tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 481, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 553, + 601 + ], + "type": "text", + "content": "Resource usage optimization. Recent transformer-based methods [2, 24, 27, 33, 41, 47] employ 3D-VAE to compress videos. However, as noted in CogVideoX [41], excessive compression hinders 3D-VAE convergence. ExVideo [12] extends SVD [5] to 128 frames via small learnable parameters with low memory overhead. Pyramidal Flow Matching [20] reformulates diffusion into pyramid stages, enabling efficient generation of videos up to 240 frames. While effective, these methods still face challenges in scaling to much longer durations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 601, + 533, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 533, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 533, + 613 + ], + "type": "text", + "content": "Problem decomposition via short video generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "- Multi-scene generation Multi-shot approaches [14, 25, 44, 49] divide long videos into segments and align them under a unified narrative, conditioning video diffusion on scene-level text or styles for coherence. MovieDreamer [46] employs a hierarchical pipeline to draft keyframes and refine shots. These methods emphasize storytelling and character consistency with relaxed demands on interclip coherence." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18198" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 491 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "- Tuning-free methods. Tuning-free methods extend short video generation to longer durations via hand-crafted designs, such as co-denoising [28, 35], noise re-scheduling with sliding windows [29], and attention control mechanisms [7, 26, 32]. Often paired with multi-prompt sampling for content richness, these methods lack long-range priors, leading to unnatural motion and appearance transitions over extended durations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 294, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 294, + 346 + ], + "type": "text", + "content": "- Auto-regressive methods. Auto-regressive methods generate long videos frame-by-frame (or clip-by-clip), ensuring temporal consistency at the cost of increased inference time. Image(clip)-to-video models [5, 18, 34, 40, 45] generate long videos iteratively by treating the last generated frame (clip) as the initial one in the next iteration, but suffer quality degradation due to error accumulation. Loong [38] adopts an auto-regressive language model-like strategy but is limited to low resolutions " + }, + { + "bbox": [ + 55, + 168, + 294, + 346 + ], + "type": "inline_equation", + "content": "(128 \\times 128)" + }, + { + "bbox": [ + 55, + 168, + 294, + 346 + ], + "type": "text", + "content": ". Causal denoising [22, 30, 39, 43] gradually increases noise scales for later frames, ensuring smooth transitions by referencing clearer earlier frames. Commercial tools like Kling [1] offer clip-by-clip video extension but are constrained by context windows and unstable long-term controllability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 347, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 294, + 491 + ], + "type": "text", + "content": "- Hierarchical methods. Hierarchical methods [6, 17, 42] use hierarchical frameworks to generate keyframes and perform interpolation or super-resolution. However, keyframe-only propagation increases information loss, degrading quality as stages grow. StoryDiffusion [48] adopts a multi-stage pipeline, generating keyframes, predicting motion, and synthesizing clips guided by both. Different from those methods, our method models correlated spatial-temporal distribution jointly for minute-long videos via condensed tokens, utilizing the fine-grained information to guide the short clip generation to achieve more natural and consistent content across the long range." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 502, + 139, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 139, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 139, + 514 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 689 + ], + "type": "text", + "content": "CogVideoX. We use the pre-trained text-to-video diffusion model CogVideoX [41] as the foundation of our framework. CogVideoX employs a 3D causal VAE [23] to compress videos into a latent space, achieving an " + }, + { + "bbox": [ + 55, + 521, + 295, + 689 + ], + "type": "inline_equation", + "content": "8 \\times 8 \\times 4" + }, + { + "bbox": [ + 55, + 521, + 295, + 689 + ], + "type": "text", + "content": " compression ratio along the spatial and temporal axes. During input processing, video latents are patched and concatenated with text embeddings, which are then passed through Expert Transformer blocks featuring Expert Adaptive LayerNorm (AdaLN) and 3D Full Attention, as shown in Figure 2. Text and Vision Expert AdaLN separately modulate text and video features, improving alignment. To handle large motions, CogVideoX integrates 3D Text-Video Hybrid Attention with 3D Rotary Position Embedding (RoPE) [31], effectively capturing spatial-temporal relationships." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "FIFO-Diffusion. FIFO-Diffusion [22] introduces a diagonal denoising strategy to extend a pre-trained text-to-video" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 347, + 70, + 525, + 162 + ], + "blocks": [ + { + "bbox": [ + 347, + 70, + 525, + 162 + ], + "lines": [ + { + "bbox": [ + 347, + 70, + 525, + 162 + ], + "spans": [ + { + "bbox": [ + 347, + 70, + 525, + 162 + ], + "type": "image", + "image_path": "3f4ca73138b8d00e4f601a54a0cfcc682f78cca3364eaa946bc5e4ec6fa0d572.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 163, + 497, + 175 + ], + "lines": [ + { + "bbox": [ + 369, + 163, + 497, + 175 + ], + "spans": [ + { + "bbox": [ + 369, + 163, + 497, + 175 + ], + "type": "text", + "content": "Figure 2. CogVideoX architecture." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "type": "text", + "content": "model from " + }, + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "type": "text", + "content": " frames (" + }, + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "type": "inline_equation", + "content": "f \\ll M" + }, + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "type": "text", + "content": ") to generate long videos with " + }, + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "type": "text", + "content": " frames. This method progressively denoises consecutive frames with increasing noise levels. Given a time step schedule " + }, + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "type": "inline_equation", + "content": "0 = \\tau_0 < \\tau_1 < \\ldots < \\tau_f = T" + }, + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "type": "text", + "content": ", each denoising step is defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 324, + 247, + 553, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 247, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 324, + 247, + 553, + 262 + ], + "type": "interline_equation", + "content": "\\left[ z _ {\\tau_ {0}} ^ {1}; \\dots ; z _ {\\tau_ {f - 1}} ^ {f} \\right] = \\Phi \\left(\\left[ z _ {\\tau_ {1}} ^ {1}; \\dots ; z _ {\\tau_ {f}} ^ {f} \\right], \\left[ \\tau_ {1}; \\dots ; \\tau_ {f} \\right], c; \\epsilon_ {\\theta}\\right). \\tag {1}", + "image_path": "b734491518e4fb07eb97af6156d163685e1a195d3902796c869e8c56b09c8752.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 270, + 553, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 355 + ], + "type": "text", + "content": "The diagonal latents " + }, + { + "bbox": [ + 313, + 270, + 553, + 355 + ], + "type": "inline_equation", + "content": "\\{z_{\\tau_i}^i\\}_{i = 1}^f" + }, + { + "bbox": [ + 313, + 270, + 553, + 355 + ], + "type": "text", + "content": " are stored in a queue. At each step, the foremost frame is dequeued once it reaches " + }, + { + "bbox": [ + 313, + 270, + 553, + 355 + ], + "type": "inline_equation", + "content": "\\tau_0 = 0" + }, + { + "bbox": [ + 313, + 270, + 553, + 355 + ], + "type": "text", + "content": ", while a new latent at " + }, + { + "bbox": [ + 313, + 270, + 553, + 355 + ], + "type": "inline_equation", + "content": "\\tau_f" + }, + { + "bbox": [ + 313, + 270, + 553, + 355 + ], + "type": "text", + "content": " is enqueued. This ensures that later frames with higher noise levels reference earlier frames with lower noise levels during denoising, maintaining temporal consistency and coherence throughout the long video generation process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 365, + 527, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 527, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 527, + 378 + ], + "type": "text", + "content": "4. TokensGen for Long Video Generation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 384, + 381, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 381, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 381, + 396 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 402, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 666 + ], + "type": "text", + "content": "Given a text prompt, our framework generates a consistent minute-long video aligned with the prompt. It consists of two main components: To2V and T2To Models, as shown in Figure 3. During training, we first train To2V, a conditional short video generation model, to control spatial layout and motion based on text and video prompts. A video tokenizer extracts compact semantic tokens " + }, + { + "bbox": [ + 313, + 402, + 555, + 666 + ], + "type": "inline_equation", + "content": "z_{sem}" + }, + { + "bbox": [ + 313, + 402, + 555, + 666 + ], + "type": "text", + "content": " from short clips, which are then fed into a diffusion transformer for guided generation. Since these tokens encode richer spatial and motion information than text prompts, they provide more accurate semantic control over individual clips. For long videos, we segment them into short clips, each tokenized to produce a sequence of semantic tokens " + }, + { + "bbox": [ + 313, + 402, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\{z_{sem,i}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 313, + 402, + 555, + 666 + ], + "type": "text", + "content": ", forming a resource-efficient high-level representation of the entire video. We then train T2To, a video token transformer, to generate these long video tokens simultaneously from text prompts, ensuring long-term content consistency across clips. During inference, we first sample long video semantic tokens using T2To, then pass them to To2V to generate each clip. To ensure temporal consistency, we introduce an adaptive FIFO denoising strategy for diagonal denoising across clips." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 672, + 522, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 522, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 522, + 685 + ], + "type": "text", + "content": "4.2. To2V Model: Inner-clip content control" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "We design a conditional short video generation model, To2V, guided by both text and video prompts for precise" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18199" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 70, + 302, + 243 + ], + "blocks": [ + { + "bbox": [ + 81, + 70, + 302, + 243 + ], + "lines": [ + { + "bbox": [ + 81, + 70, + 302, + 243 + ], + "spans": [ + { + "bbox": [ + 81, + 70, + 302, + 243 + ], + "type": "image", + "image_path": "8a8640b752b68ab8ad69fec4c8483a4ced9b875724d0ffe7ebd8da9602701695.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 249, + 490, + 261 + ], + "lines": [ + { + "bbox": [ + 118, + 249, + 490, + 261 + ], + "spans": [ + { + "bbox": [ + 118, + 249, + 490, + 261 + ], + "type": "text", + "content": "Figure 3. Overview of the model. Left: Overall Framework for TokensGen. Right: Trainable Modules." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 71, + 512, + 243 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 512, + 243 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 512, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 512, + 243 + ], + "type": "image", + "image_path": "2cee7ceff24f9aec0a3886634daeedb679c926f55f2f645d8aeb4212c055e2e7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 273, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 273, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 273, + 295, + 357 + ], + "type": "text", + "content": "content control in short video generation. To2V builds on the pre-trained text-guided video generation model CogVideoX [41] and consists of two key components: the Video Tokenizer that encodes the input video clip into compact semantic tokens, and the Cross-Attention Branch integrated with CogVideoX that enables cross-attention between semantic tokens and noisy latents." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": "Video Tokenizer. The Video Tokenizer consists of a 3D causal variational autoencoder (3D-VAE), a Patchify Module, and a Resampler, as illustrated on the right side of Figure 3. The 3D-VAE and the Patchify Module are inherited from CogVideoX with fixed weights. They process the input video into a set of tokens " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "z_{source}" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": " with the shape " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "f_s \\times h_s \\times w_s \\times c_s" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": " represent the number of frames, height, width, and channels, respectively. The Resampler compresses and resamples " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "z_{source}" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": " into a more compact representation space, as illustrated in Figure 4. It comprises a learnable latent " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "z_{latent}" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": " with the shape " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "f_r \\times h_r \\times w_r \\times c_s" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": ", four blocks of the 3D Cross-Attention Module that perform cross-attention between " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "z_{source}" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "z_{latent}" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": ", and a Projector that transforms " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "z_{latent}" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "z_{sem}" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": " with shape " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "f_r \\times h_r \\times w_r \\times c_r" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "f_r < f_s, h_r < h_s, w_r < w_s, c_r < c_s" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": ". The semantic tokens " + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "inline_equation", + "content": "z_{sem}" + }, + { + "bbox": [ + 56, + 358, + 296, + 598 + ], + "type": "text", + "content": " encoded by the Video Tokenizer encapsulate high-level spatial layouts and motion information from the input video while maintaining a more compact size compared to the original video." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 295, + 684 + ], + "type": "text", + "content": "Cross-Attention Branch. To effectively incorporate these semantic tokens with CogVideoX, we add a separate Cross-Attention Branch to handle the newly added semantic conditions. This branch consists of a Semantic Token Adaptive LayerNorm (Sem AdaLN) and a 3D Cross-Attention Module, as depicted on the right side of Figure 3. The process is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- Back projection: The semantic tokens " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "z_{sem}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " from the Video Tokenizer are back-projected to match the num" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 339, + 276, + 529, + 346 + ], + "blocks": [ + { + "bbox": [ + 339, + 276, + 529, + 346 + ], + "lines": [ + { + "bbox": [ + 339, + 276, + 529, + 346 + ], + "spans": [ + { + "bbox": [ + 339, + 276, + 529, + 346 + ], + "type": "image", + "image_path": "28f37149a1de396f382a2932abdd2a4ffdc88b34560e6819f687799a58802b3a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 352, + 515, + 363 + ], + "lines": [ + { + "bbox": [ + 352, + 352, + 515, + 363 + ], + "spans": [ + { + "bbox": [ + 352, + 352, + 515, + 363 + ], + "type": "text", + "content": "Figure 4. The architecture of the Resampler." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 367, + 553, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 367, + 553, + 391 + ], + "spans": [ + { + "bbox": [ + 321, + 367, + 553, + 391 + ], + "type": "text", + "content": "ber of channels of the combined text-video embeddings " + }, + { + "bbox": [ + 321, + 367, + 553, + 391 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathrm{tv}} = [\\mathbf{Z}_{\\mathrm{text}};\\mathbf{Z}_{\\mathrm{video}}]" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 392, + 555, + 524 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 392, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 392, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 314, + 392, + 553, + 415 + ], + "type": "text", + "content": "- Concatenation: These back-projected semantic tokens are concatenated with the text-video embeddings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 415, + 553, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 415, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 553, + 450 + ], + "type": "text", + "content": "- Modulation: Similar to the Text and Vision AdaLN, the Sem AdaLN modulates the semantic condition embeddings to ensure better feature alignment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 451, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 451, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 314, + 451, + 555, + 524 + ], + "type": "text", + "content": "- Attention: The modulated embeddings are passed to the 3D Text-Video Attention and the 3D Cross Attention Module to perform 3D full attention on the combined embeddings. Given the combined embeddings " + }, + { + "bbox": [ + 314, + 451, + 555, + 524 + ], + "type": "inline_equation", + "content": "\\left[\\mathbf{Z}_{\\mathrm{text}}; \\mathbf{Z}_{\\mathrm{video}}; \\mathbf{Z}_{\\mathrm{sem}}\\right]" + }, + { + "bbox": [ + 314, + 451, + 555, + 524 + ], + "type": "text", + "content": ", the output attention results " + }, + { + "bbox": [ + 314, + 451, + 555, + 524 + ], + "type": "inline_equation", + "content": "\\left[\\mathbf{Z}_{\\mathrm{text}}^{\\prime}; \\mathbf{Z}_{\\mathrm{video}}^{\\prime}; \\mathbf{Z}_{\\mathrm{sem}}^{\\prime}\\right]" + }, + { + "bbox": [ + 314, + 451, + 555, + 524 + ], + "type": "text", + "content": " are represented as follows:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 384, + 533, + 478, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 533, + 478, + 545 + ], + "spans": [ + { + "bbox": [ + 384, + 533, + 478, + 545 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {\\mathrm {t v}} = \\left[ \\mathbf {Z} _ {\\text {t e x t}}; \\mathbf {Z} _ {\\text {v i d e o}} \\right]", + "image_path": "69c7bba4c91b5fc14a9b3c88a88f1f5372d926d7a7ff049021cbf83ebd7a62a3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 348, + 548, + 523, + 560 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 548, + 523, + 560 + ], + "spans": [ + { + "bbox": [ + 348, + 548, + 523, + 560 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\mathbf {Z} _ {\\mathrm {t v}} \\mathbf {W} _ {\\mathbf {q}}, \\mathbf {K} = \\mathbf {Z} _ {\\mathrm {t v}} \\mathbf {W} _ {\\mathbf {k}}, \\mathbf {V} = \\mathbf {Z} _ {\\mathrm {t v}} \\mathbf {W} _ {\\mathbf {v}}", + "image_path": "1a74362523c91d2fe309dfcc93d63f6e1dd894d207d344d2986d23a098b34430.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 337, + 563, + 545, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 563, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 337, + 563, + 545, + 576 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {\\mathbf {s}} = \\mathbf {Z} _ {\\mathbf {s e m}} \\mathbf {W} _ {\\mathbf {q}} ^ {\\mathrm {c}}, \\mathbf {K} _ {\\mathbf {s}} = \\mathbf {Z} _ {\\mathbf {s e m}} \\mathbf {W} _ {\\mathbf {k}} ^ {\\mathrm {c}}, \\mathbf {V} _ {\\mathbf {s}} = \\mathbf {Z} _ {\\mathbf {s e m}} \\mathbf {W} _ {\\mathbf {v}} ^ {\\mathrm {c}}", + "image_path": "e2d2afbc008af09046c3a7997e71a1b5a760c5b4fa1eeb5017de2ca7376aaef6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 339, + 578, + 542, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 578, + 542, + 592 + ], + "spans": [ + { + "bbox": [ + 339, + 578, + 542, + 592 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {\\mathbf {t v}} = \\mathbf {Z} _ {\\mathbf {t v}} \\mathbf {W} _ {\\mathbf {q}} ^ {\\mathbf {c}}, \\mathbf {K} _ {\\mathbf {t v}} = \\mathbf {Z} _ {\\mathbf {t v}} \\mathbf {W} _ {\\mathbf {k}} ^ {\\mathbf {c}}, \\mathbf {V} _ {\\mathbf {t v}} = \\mathbf {Z} _ {\\mathbf {t v}} \\mathbf {W} _ {\\mathbf {v}} ^ {\\mathbf {c}}", + "image_path": "12d2ab04f9b60347ed98011c90c8a8d6b1c4f4ab9b1737ca2dbffc06234ecc02.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 324, + 594, + 556, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 594, + 556, + 682 + ], + "spans": [ + { + "bbox": [ + 324, + 594, + 556, + 682 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left[ \\mathbf {Z} _ {\\text {t e x t}} ^ {\\prime}, \\mathbf {Z} _ {\\text {v i d e o}} ^ {\\prime} \\right] = \\operatorname {A t t n} (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) + \\operatorname {A t t n} \\left(\\mathbf {Q} _ {\\mathrm {t v}}, \\mathbf {K} _ {\\mathrm {s}}, \\mathbf {V} _ {\\mathrm {s}}\\right) \\\\ = \\mathrm {s o f t m a x} (\\frac {\\mathbf {Q} (\\mathbf {K}) ^ {T}}{\\sqrt {d}}) \\mathbf {V} + \\mathrm {s o f t m a x} (\\frac {\\mathbf {Q} _ {\\mathbf {t v}} (\\mathbf {K} _ {\\mathbf {s}}) ^ {T}}{\\sqrt {d}}) \\mathbf {V} _ {\\mathbf {s}}, \\\\ \\mathbf {Z} _ {\\text {s e m}} ^ {\\prime} = \\operatorname {A t t n} \\left(\\mathbf {Q} _ {\\mathbf {s}}, \\left[ \\mathbf {K} _ {\\mathbf {t v}}; \\mathbf {K} _ {\\mathbf {s}} \\right], \\left[ \\mathbf {V} _ {\\mathbf {t v}}; \\mathbf {V} _ {\\mathbf {s}} \\right]\\right) \\\\ = \\mathrm {s o f t m a x} (\\frac {\\mathbf {Q _ {s}} ([ \\mathbf {K _ {t v}} ; \\mathbf {K _ {s}} ]) ^ {T}}{\\sqrt {d}}) [ \\mathbf {V _ {t v}}; \\mathbf {V _ {s}} ], \\\\ \\end{array}", + "image_path": "228cf46063925b1859425b2dc6dbd2909663f55bb44baf65e63d16f1fc3db529.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 322, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 322, + 689, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 322, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{q}}" + }, + { + "bbox": [ + 322, + 689, + 553, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 322, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{k}}" + }, + { + "bbox": [ + 322, + 689, + 553, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 322, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{v}}" + }, + { + "bbox": [ + 322, + 689, + 553, + 713 + ], + "type": "text", + "content": " are fixed parameters of the 3D Text-Video Attention Module inherited from CogVideoX," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18200" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "type": "text", + "content": "Q, K, V are the corresponding query, key, and value matrices. " + }, + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{q}}^{\\mathrm{c}}, \\mathbf{W}_{\\mathbf{k}}^{\\mathrm{c}}, \\mathbf{W}_{\\mathbf{v}}^{\\mathrm{c}}" + }, + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "type": "text", + "content": " are trainable parameters of the 3D Cross-Attention Module, " + }, + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "type": "inline_equation", + "content": "Q_{s}, K_{s}, V_{s}" + }, + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "type": "text", + "content": " are the query, key, and value matrices for " + }, + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathrm{sem}}" + }, + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "type": "inline_equation", + "content": "Q_{\\mathrm{tv}}, K_{\\mathrm{tv}}, V_{\\mathrm{tv}}" + }, + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "type": "text", + "content": " are the query, key, and value matrices for the combined text and video embeddings " + }, + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{\\mathrm{tv}}" + }, + { + "bbox": [ + 63, + 72, + 295, + 145 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 286, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 286, + 170 + ], + "type": "text", + "content": "4.3. T2To Model: Long-term content consistency" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "text", + "content": "To learn long-term content and logic knowledge across the minute-long video, we design a video token transformer, the T2To Model, to generate the semantic tokens " + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\{z_{sem,i}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "text", + "content": " representing the whole long video given the input text prompt. We adopt the same model structure and training strategy of CogVideoX [41] for the T2To Model, except for the following modifications:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 262, + 295, + 345 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 262, + 295, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 295, + 297 + ], + "type": "text", + "content": "- The model aims to generate " + }, + { + "bbox": [ + 55, + 262, + 295, + 297 + ], + "type": "inline_equation", + "content": "\\{z_{sem,i}\\}_{i=1}^{N}" + }, + { + "bbox": [ + 55, + 262, + 295, + 297 + ], + "type": "text", + "content": " with the shape " + }, + { + "bbox": [ + 55, + 262, + 295, + 297 + ], + "type": "inline_equation", + "content": "(Nf_r) \\times h_r \\times w_r \\times c_r" + }, + { + "bbox": [ + 55, + 262, + 295, + 297 + ], + "type": "text", + "content": ". The total number of tokens is " + }, + { + "bbox": [ + 55, + 262, + 295, + 297 + ], + "type": "inline_equation", + "content": "(Nf_r) \\times h_r \\times w_r" + }, + { + "bbox": [ + 55, + 262, + 295, + 297 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 298, + 294, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 294, + 345 + ], + "type": "text", + "content": "- Since the number of tokens along the temporal dimension is much larger than the spatial dimensions, for 3D-RoPE, we reallocate the hidden state channel for height, width, and temporal dimension as about " + }, + { + "bbox": [ + 55, + 298, + 294, + 345 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 298, + 294, + 345 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 298, + 294, + 345 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 298, + 294, + 345 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 298, + 294, + 345 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 298, + 294, + 345 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 356, + 223, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 223, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 223, + 369 + ], + "type": "text", + "content": "4.4. Inter-clip temporal consistency" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 374, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 649 + ], + "type": "text", + "content": "If each clip is denoised separately with a corresponding semantic token " + }, + { + "bbox": [ + 55, + 374, + 295, + 649 + ], + "type": "inline_equation", + "content": "z_{sem,i}" + }, + { + "bbox": [ + 55, + 374, + 295, + 649 + ], + "type": "text", + "content": ", the model will generate a group of discontinuous clips. To achieve temporal continuity, we apply the FIFO-denoising strategy during the inference stage. Specifically, we adopt latent partitioning and lookahead denoising, like the original FIFO. However, to maintain a queue with sufficient frames at the start of denoising, FIFO pads the positions ahead of the first clip with noise-augmented first frame replications. We observe that this approach introduces artifacts in our settings, as the replicated frames deviate from the intrinsic distribution of the training domain for the video diffusion model. To address this, we propose an improved version of FIFO, named adaptive-FIFO, which incorporates an adaptive padding strategy at the beginning of the denoising process. For a latent partition containing fewer than " + }, + { + "bbox": [ + 55, + 374, + 295, + 649 + ], + "type": "inline_equation", + "content": "f_s" + }, + { + "bbox": [ + 55, + 374, + 295, + 649 + ], + "type": "text", + "content": " frames, we denoise all frames together and update them simultaneously. For a partition with exactly " + }, + { + "bbox": [ + 55, + 374, + 295, + 649 + ], + "type": "inline_equation", + "content": "f_s" + }, + { + "bbox": [ + 55, + 374, + 295, + 649 + ], + "type": "text", + "content": " frames, we employ lookahead denoising: the frames are denoised together, but only the noisier frames in the latter portion are updated. By better aligning the initial padding with the model's learned distribution and ensuring continuity in partially filled partitions, this approach yields smoother transitions across clips and better frame quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 160, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 160, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 160, + 672 + ], + "type": "text", + "content": "4.5. Training strategy" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "For To2V Model, we fix the weights of the pre-trained modules of the base model, train the Resampler of the Video Tokenizer and the Cross-Attention Branch. For T2To Model," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "we initialize the model with the weights of the base model and train all the modules." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 97, + 555, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 555, + 217 + ], + "type": "text", + "content": "We adopt similar training strategies with CogVideoX [41], including Multi-Resolution Frame Pack and Explicit Uniform Sampling. For T2To Model, we pack videos with different time duration into the same batch, and apply an attention mask indicating the valid frames, as well as the attention mask for loss calculation, to ensure attention module focus on the right area of the input noisy latents, an approach similar with Patch'n Pack [11]. For both To2V and T2To Model, we employ the explicit uniform sampling strategy for sampling timesteps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 230, + 440, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 440, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 440, + 243 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 251, + 447, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 447, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 447, + 264 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "text", + "content": "Model Architecture. We employ CogVideoX-5B [41] as the base model for both To2V and T2To. In To2V, the input tokens " + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "inline_equation", + "content": "z_{source}" + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "text", + "content": " have the shape " + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "inline_equation", + "content": "13 \\times 30 \\times 45 \\times 3072" + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "text", + "content": ". We observed that T2To struggles to converge when " + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "inline_equation", + "content": "c_r" + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "text", + "content": " is large (e.g., comparable to " + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "inline_equation", + "content": "c_s" + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "text", + "content": "), so we set the compressed semantic tokens " + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "inline_equation", + "content": "z_{sem}" + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "text", + "content": " to have dimensions " + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "inline_equation", + "content": "4 \\times 8 \\times 12 \\times 16" + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "text", + "content": ". For the Projector in the Resampler, we observed that a linear projection via PCA [13] effectively reduces the channel dimension without sacrificing information, as further analyzed in Sec. 5.3. Compared to the original latent shape " + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "inline_equation", + "content": "13 \\times 60 \\times 90 \\times 16" + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "text", + "content": ", we achieve a compression ratio of approximately " + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "inline_equation", + "content": "3 \\times 8 \\times 8" + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "text", + "content": ". Thus, we first train To2V without the channel projection and then apply PCA to the Resampler's output embeddings on 300 samples to obtain the transformation matrix. In T2To, we set the maximum number of chunks " + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "inline_equation", + "content": "N = 24" + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "text", + "content": ". Each chunk contains 49 frames, allowing our model to process videos up to " + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "inline_equation", + "content": "24 \\times 49 = 1176" + }, + { + "bbox": [ + 313, + 269, + 555, + 605 + ], + "type": "text", + "content": " frames. Dataset. We use the MiraData dataset [21], comprising long videos with structured captions. We first collect 56k videos, using their dense captions for training. For To2V Model, we randomly sample 49-frame video clips at 10 fps from these long videos as training targets. For T2To Model, we select a filtered subset of around 16k high-quality videos that are at least one minute long, primarily consisting of gameplay footage and natural landscape. We filter out videos with abrupt scene changes via PySceneDetect [8] and human evaluations. This subset ensures consistency within long videos for the training of T2To Model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": "Training Details. We adopt a progressive learning strategy for both T2To and To2V Model. For To2V Model, we first train on smaller token shapes " + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "inline_equation", + "content": "(4 \\times 5 \\times 7 \\times 3072)" + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": " for 1,200 iterations, using a batch size of 72 and a learning rate of " + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": ". We then transition to the full token resolution " + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "inline_equation", + "content": "(4 \\times 8 \\times 12 \\times 3072)" + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": " for 2,600 more iterations, initializing from the previously trained model. For T2To Model, we begin with shorter videos (" + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "inline_equation", + "content": "N = 3" + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": " chunks of 49 frames each) for 1,200 iterations, using a learning rate of " + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": ". Next," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18201" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 553, + 304 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 553, + 304 + ], + "type": "image", + "image_path": "4787b9de97dc8090a3209eef164010897cd21efe9c8841ade882509e4b7d0cd6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 312, + 499, + 324 + ], + "lines": [ + { + "bbox": [ + 110, + 312, + 499, + 324 + ], + "spans": [ + { + "bbox": [ + 110, + 312, + 499, + 324 + ], + "type": "text", + "content": "Figure 5. The qualitative comparison. We recommend readers refer to our webpage for video comparisons." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 83, + 355, + 531, + 449 + ], + "blocks": [ + { + "bbox": [ + 207, + 335, + 402, + 346 + ], + "lines": [ + { + "bbox": [ + 207, + 335, + 402, + 346 + ], + "spans": [ + { + "bbox": [ + 207, + 335, + 402, + 346 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation of comparison study." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 355, + 531, + 449 + ], + "lines": [ + { + "bbox": [ + 83, + 355, + 531, + 449 + ], + "spans": [ + { + "bbox": [ + 83, + 355, + 531, + 449 + ], + "type": "table", + "html": "
VBranchHuman Study
ModelsSubject ConsistencyBackground ConsistencyTemporal FlickeringMotion SmoothnessImaging QualityDynamic DegreeText-Visual AlignmentMotion & Content Consistency
Video-Infinity81.8090.5696.6697.6561.5831.00.0%0.69%
DiTCtrl76.7687.9695.5297.7859.2675.00.0%1.0%
AP-FIFO+CogVideoX86.2292.8994.7897.4864.1078.5724.31%22.57%
Ours84.5792.295.4198.0863.3178.9575.69%75.74%
", + "image_path": "594dc63f6da9986c6e6f5573a5fb65590a14499c0b3798ffbe51756721c44163.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 69, + 486, + 284, + 573 + ], + "blocks": [ + { + "bbox": [ + 55, + 466, + 295, + 478 + ], + "lines": [ + { + "bbox": [ + 55, + 466, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 295, + 478 + ], + "type": "text", + "content": "Table 2. Ablation study on ways of incorporating video conditions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 486, + 284, + 573 + ], + "lines": [ + { + "bbox": [ + 69, + 486, + 284, + 573 + ], + "spans": [ + { + "bbox": [ + 69, + 486, + 284, + 573 + ], + "type": "table", + "html": "
MethodsMSE↓SSIM↑temporal flickering↑
SR1.12e-20.6097.10
4x8x12 (w/o projection)1.24e-20.6297.58
4x5x72.84e-20.4997.57
13x5x71.54e-20.5897.54
4x8x121.24e-20.6297.58
", + "image_path": "c5caab729c2cbc6beb1e2d3128f63ecd7d16758f1b0d3deb5cbc56dd9e206384.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 576, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 637 + ], + "type": "text", + "content": "we expand to long videos with up to " + }, + { + "bbox": [ + 55, + 576, + 295, + 637 + ], + "type": "inline_equation", + "content": "N = 24" + }, + { + "bbox": [ + 55, + 576, + 295, + 637 + ], + "type": "text", + "content": " chunks, training for 5,000 iterations, with a learning rate of " + }, + { + "bbox": [ + 55, + 576, + 295, + 637 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 576, + 295, + 637 + ], + "type": "text", + "content": " and a batch size of 105. This progressive training helps the model converge faster for more complex, extended video generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 180, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 180, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 180, + 659 + ], + "type": "text", + "content": "5.2. Baseline comparisons" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Qualitative comparisons We evaluate our approach against several recent multi-prompt long video generation methods, including Video-Infinity [32], DiTCtrl [7], and Kling [1], as well as a baseline that adopts FIFO-Diffusion [22] on" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 467, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 708 + ], + "type": "text", + "content": "CogVideoX with our adaptive-padding strategy. For multiprompt methods, we use GPT-4o [4] to split the prompt into 24 segments, which are used for guiding each clip generation. FIFO and ours use the same text prompt, abbreviated as: \"a person riding a horse along a path leading to a serene river.\" The results are shown in Fig. 5. VideoInfinity produces transitions primarily through background changes while failing to capture meaningful foreground motion. The person and the horse remain essentially static within each clip, resulting in low engagement and narrative drift over longer durations. DiTCtrl demonstrates intermittently aligned keyframes but struggles to maintain smooth transitions between clips, leading to abrupt scene shifts and a disjointed storyline. Kling generates visually consistent frames but exhibits erratic motions (e.g., the subject abruptly reversing direction) and occasional discontinuities in scene composition. Such artifacts disrupt the viewing experience and deviate from the intended story arc. For FIFO (with adaptive padding on CogVideoX), we observe progressive over-saturation and abrupt changes in appear" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18202" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "ance or color schemes as the video extends. These issues are especially pronounced when generating complex scenes over hundreds of frames. By contrast, our method delivers smoother motion transitions and subject representation, adherent to the prompt throughout the entire minute-long sequence. More comparison results are included in the supplementary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 158, + 298, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 158, + 298, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 158, + 298, + 445 + ], + "type": "text", + "content": "Quantitative comparisons. We conduct a quantitative comparison study on 100 prompts randomly selected from the MiraData [21] test set. As reported in Tab. 1, our approach achieves the highest scores on VBench [19] for both Motion Smoothness and Dynamic Degree. We observe that certain metrics in VBench (e.g., Subject and Background Consistency, and Temporal Flickering) may assign higher scores to less dynamic videos, motivating us to conduct a user study for a more comprehensive evaluation. For the user study, we generate 12 video results for each method, with each video lasting between one and two minutes, covering categories such as humans, cars, and natural scenes. All the resulting videos are included on our webpage. To ensure unbiased feedback, videos are randomly shuffled and presented to 24 participants. They evaluate each video on two criteria: text-visual alignment and motion & content consistency. As shown on the right side of Tab. 1, our method consistently outperforms others in both dimensions, reflecting its strong long-term control capabilities. These results demonstrate that our approach effectively maintains semantic alignment with the textual prompt while preserving smooth motion and stable content over extended sequences. Additional results and details are provided in the supplementary." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 449, + 154, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 154, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 154, + 460 + ], + "type": "text", + "content": "5.3. Ablation studies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": "Ablation on video conditions. We investigate various strategies for incorporating video conditions into the To2V Model: (1) the condensed token shape, (2) with or without channel projection, and (3) a super-resolution-based approach. Specifically, we experiment with three token shapes " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "(4\\times 5\\times 7,13\\times 5\\times 7,4\\times 8\\times 12)" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": ", train models with and without the Projector module, and compare them against a super-resolution setting (where the low-resolution video directly serves as the condition). We also include a baseline using FIFO-Diffusion to illustrate its potential shortcomings without the conditioning process. As shown in Figure 6, the FIFO-based approach often produces inconsistent foreground and background visuals, underscoring the difficulty of preserving spatial-temporal coherence from purely latent-level guidance. Meanwhile, the super-resolution method tends to duplicate low-level color and texture cues from the source, failing to capture higher-level semantics, leading to less meaningful renderings. Comparing models with and without the Projector, we observe similar performances, demonstrating that our PCA-based projection provides a lightweight yet effective means of dimen" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "sionality reduction without sacrificing image quality. Concerning the shape of the condensed tokens, the smallest variant " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "(4\\times 5\\times 7)" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " struggles to preserve essential layout and motion patterns, resulting in less accurate re-creations of the source video. Increasing the token's temporal or spatial resolution " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "(13\\times 5\\times 7,4\\times 8\\times 12)" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " significantly improves alignment and maintains better semantic fidelity. Among these, " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "(4\\times 8\\times 12)" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " achieves the most favorable balance of fine-grained detail and computational efficiency, as quantitatively confirmed in Tab. 2. Overall, these ablation studies demonstrate that our token representation, combined with optional PCA-based projection, offers a robust and effective pathway to incorporate video conditions in To2V Model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 228, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 418 + ], + "type": "text", + "content": "Ablation on FIFO. We further examine the influence of FIFO-Diffusion and our adaptive padding technique by comparing three variants: (1) omitting FIFO entirely, (2) using FIFO without adaptive padding, and (3) our full approach, incorporating both FIFO and adaptive padding. As illustrated in Fig. 7, disabling FIFO leads to abrupt scene changes between consecutive clips, producing visually inconsistent transitions where subjects may teleport or backgrounds shift abruptly. Meanwhile, removing adaptive padding induces severe artifacts in the initial frames of the video, as the model relies on duplicated frames that deviate from the training distribution. These artifacts propagate into subsequent frames, degrading overall quality. In contrast, our adaptive padding strategy aligns the padding frames with the model's distribution, preventing unnatural discontinuities at clip boundaries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 426, + 429, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 429, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 429, + 440 + ], + "type": "text", + "content": "5.4. Long Video Editing" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 444, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 553, + 552 + ], + "type": "text", + "content": "Beyond generating entirely novel content, our method readily adapts to various long video editing scenarios. Specifically, the To2V Model's capacity to integrate textual prompts with source video data allows for transformations that preserve the essential structure of the original footage while injecting new semantics. We directly combine the target text prompt and the source video as input conditions to generate the edited long video. For detailed results, please refer to the supplementary and our webpage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 562, + 466, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 466, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 466, + 574 + ], + "type": "text", + "content": "6. Conclusion and Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 715 + ], + "type": "text", + "content": "We introduce TokensGen, a two-stage framework that addresses key challenges in long video generation, controlling per-clip semantics, ensuring long-term coherence, and achieving smooth transitions. The To2V Model generates short clips guided by text and video prompts, capturing fine-grained motion and content. The T2To Model transformer then leverages condensed semantic tokens to preserve long-term consistency across clips. Finally, our adaptive FIFO-Diffusion strategy overcomes boundary artifacts by maintaining temporal continuity. This pipeline efficiently scales pre-trained short video models to longer videos, enabling" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18203" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 73, + 550, + 369 + ], + "blocks": [ + { + "bbox": [ + 65, + 73, + 550, + 369 + ], + "lines": [ + { + "bbox": [ + 65, + 73, + 550, + 369 + ], + "spans": [ + { + "bbox": [ + 65, + 73, + 550, + 369 + ], + "type": "image", + "image_path": "dfe808f1af597a587d775f8222e65d0e79e2bdd8e7399236058655673fb92c34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 175, + 380, + 435, + 392 + ], + "lines": [ + { + "bbox": [ + 175, + 380, + 435, + 392 + ], + "spans": [ + { + "bbox": [ + 175, + 380, + 435, + 392 + ], + "type": "text", + "content": "Figure 6. Ablation study on methods of incorporating video conditions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 405, + 550, + 562 + ], + "blocks": [ + { + "bbox": [ + 65, + 405, + 550, + 562 + ], + "lines": [ + { + "bbox": [ + 65, + 405, + 550, + 562 + ], + "spans": [ + { + "bbox": [ + 65, + 405, + 550, + 562 + ], + "type": "image", + "image_path": "8d7e1ea812f78c751112fd7f2fb5bdd540000383a622ff5a271768b4ee308f37.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 573, + 364, + 585 + ], + "lines": [ + { + "bbox": [ + 246, + 573, + 364, + 585 + ], + "spans": [ + { + "bbox": [ + 246, + 573, + 364, + 585 + ], + "type": "text", + "content": "Figure 7. Ablation study FIFO." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 630 + ], + "type": "text", + "content": "a scalable, flexible, and resource-efficient approach to long video generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "content": "Despite the effectiveness of TokensGen in maintaining long-range consistency, it does not preserve all fine-grained details. Focusing on high-level semantics, tokens may cause gradual variations in foreground objects over extended sequences (detailed in the supplementary). In complex scenes, their capacity to capture intricate spatial-temporal cues may be insufficient, requiring fine-grained" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 701 + ], + "type": "text", + "content": "tokenization and stronger short-term consistency strategies beyond tuning-free FIFO. Our framework is tested on a limited dataset of gameplay and landscape videos, but is scalable to larger datasets for broader applications. In future work, exploring multi-scale tokenization or hybrid representations could bolster fine-grained controllability, retaining subtle attributes while preserving the scalability and resource efficiency." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18204" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 297, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 297, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 297, + 147 + ], + "type": "text", + "content": "This research is supported by NTU SUG-NAP and is also supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAFICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 115, + 169 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 176, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 176, + 294, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 176, + 294, + 197 + ], + "spans": [ + { + "bbox": [ + 61, + 176, + 294, + 197 + ], + "type": "text", + "content": "[1] Kling. https://kling.kuaishou.com/, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 199, + 295, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 199, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 62, + 199, + 295, + 220 + ], + "type": "text", + "content": "[2] Hunyuanvideo: A systematic framework for large video generative models, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 223, + 260, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 223, + 260, + 233 + ], + "spans": [ + { + "bbox": [ + 62, + 223, + 260, + 233 + ], + "type": "text", + "content": "[3] sora. https://openai.com/sora/, 2024.1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 234, + 201, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 234, + 201, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 234, + 201, + 245 + ], + "type": "text", + "content": "[4] Gpt-4o. chatgpt.com, 2025.6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 246, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 295, + 300 + ], + "type": "text", + "content": "[5] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 301, + 295, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 301, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 301, + 295, + 356 + ], + "type": "text", + "content": "[6] Tim Brooks, Janne Hellsten, Miika Aittala, Ting-Chun Wang, Timo Aila, Jaakko Lehtinen, Ming-Yu Liu, Alexei Efros, and Tero Karras. Generating long videos of dynamic scenes. Advances in Neural Information Processing Systems, 35:31769-31781, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "type": "text", + "content": "[7] Minghong Cai, Xiaodong Cun, Xiaoyu Li, Wenze Liu, Zhaoyang Zhang, Yong Zhang, Ying Shan, and Xiangyu Yue. Ditctrl: Exploring attention control in multi-modal diffusion transformer for tuning-free multi-prompt longer video generation. arXiv:2412.18597, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 413, + 293, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 293, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 293, + 434 + ], + "type": "text", + "content": "[8] Brandon Castellano. Pyscenedetect. https://github.com/Breakthrough/PySceneDetect, 2024.5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "text", + "content": "[9] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 480, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 295, + 534 + ], + "type": "text", + "content": "[10] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In ICLR, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 536, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 645 + ], + "type": "text", + "content": "[11] Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M. Alabdulmohsin, Avital Oliver, Piotr Padlewski, Alexey A. Gritsenko, Mario Lucic, and Neil Houlsby. Patch n' pack: Navit, a vision transformer for any aspect ratio and resolution. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[12] Zhongjie Duan, Wenmeng Zhou, Cen Chen, Yaliang Li, and Weining Qian. Exvideo: Extending video diffusion models via parameter-efficient post-tuning. arXiv preprint arXiv:2406.14130, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[13] Karl Pearson F.R.S. Liii. on lines and planes of closest fit to systems of points in space. The London, Edinburgh, and" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "text", + "content": "Dublin Philosophical Magazine and Journal of Science, 2 (11):559-572, 1901. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[14] Ting Yao Fuchen Long, Zhaofan Qiu and Tao Mei. Videostudio: Generating consistent-content and multi-scene videos. In ECCV, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 173 + ], + "type": "text", + "content": "[15] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. arXiv preprint arXiv:2311.16933, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 175, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 230 + ], + "type": "text", + "content": "[16] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 232, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 263 + ], + "type": "text", + "content": "[17] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 319 + ], + "type": "text", + "content": "[18] Roberto Henschel, Levon Khachatryan, Daniil Hayrapetyan, Hayk Poghosyan, Vahram Tadevosyan, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Streamingt2v: Consistent, dynamic, and extendable long video generation from text. arXiv preprint arXiv:2403.14773, 2024. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 321, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 397 + ], + "type": "text", + "content": "[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[20] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. arXiv preprint arXiv:2410.05954, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 498 + ], + "type": "text", + "content": "[21] Xuan Ju, Yiming Gao, Zhaoyang Zhang, Ziyang Yuan, Xintao Wang, Ailing Zeng, Yu Xiong, Qiang Xu, and Ying Shan. Miradata: A large-scale video dataset with long durations and structured captions, 2024. 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 499, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 499, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 499, + 553, + 533 + ], + "type": "text", + "content": "[22] Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fico-diffusion: Generating infinite videos from text without training. arXiv preprint arXiv:2405.11473, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 534, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 555 + ], + "type": "text", + "content": "[23] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 610 + ], + "type": "text", + "content": "[24] Bin Lin, Yunyang Ge, Xinhua Cheng, Zongjian Li, Bin Zhu, Shaodong Wang, Xianyi He, Yang Ye, Shenghai Yuan, Liuhan Chen, et al. Open-sora plan: Open-source large video generation model. arXiv preprint arXiv:2412.00131, 2024. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 645 + ], + "type": "text", + "content": "[25] Han Lin, Abhay Zala, Jaemin Cho, and Mohit Bansal. Videodirectorgpt: Consistent multi-scene video generation via llm-guided planning. 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[26] Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. Freelong: Training-free long video generation with spectralblend temporal attention. arXiv preprint arXiv:2407.19918, 2024. 1, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[27] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Zwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte:" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18205" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "text", + "content": "Latent diffusion transformer for video generation. arXiv preprint arXiv:2401.03048, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 140 + ], + "type": "text", + "content": "[28] Yongjia Ma, Junlin Chen, Donglin Di, Qi Xie, Lei Fan, Wei Chen, Xiaofei Gou, Na Zhao, and Xun Yang. Tuning-free long video generation via global-local collaborative diffusion. arXiv preprint arXiv:2501.05484, 2025. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 176 + ], + "type": "text", + "content": "[29] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 177, + 294, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 294, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 294, + 210 + ], + "type": "text", + "content": "[30] David Ruhe, Jonathan Heek, Tim Salimans, and Emiel Hoogeboom. Rolling diffusion models. arXiv preprint arXiv:2402.09470, 2024. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 212, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 294, + 255 + ], + "type": "text", + "content": "[31] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 258, + 294, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 294, + 291 + ], + "type": "text", + "content": "[32] Zhenxiong Tan, Xingyi Yang, Songhua Liu, and Xinchao Wang. Video-infinity: Distributed long video generation. arXiv preprint arXiv:2406.16260, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 293, + 294, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 314 + ], + "type": "text", + "content": "[33] Genmo Team. Mochi 1. https://github.com/genmoai/models, 2024.2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 370 + ], + "type": "text", + "content": "[34] Ye Tian, Ling Yang, Haotian Yang, Yuan Gao, Yufan Deng, Jingmin Chen, Xintao Wang, Zhaochen Yu, Xin Tao, Pengfei Wan, Di Zhang, and Bin Cui. Videotetris: Towards compositional text-to-video generation. arXiv preprint arXiv:2406.04277, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 373, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 294, + 416 + ], + "type": "text", + "content": "[35] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-1-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 418, + 294, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 294, + 461 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 294, + 461 + ], + "type": "text", + "content": "[36] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. arXiv preprint arXiv:2308.06571, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 464, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 508 + ], + "type": "text", + "content": "[37] Yaohui Wang, Xinyuan Chen, Xin Ma, Shangchen Zhou, Ziqi Huang, Yi Wang, Ceyuan Yang, Yinan He, Jiashuo Yu, Peiqing Yang, et al. Lavie: High-quality video generation with cascaded latent diffusion models. IJCV, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 510, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 554 + ], + "type": "text", + "content": "[38] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models. arXiv preprint arXiv:2410.02757, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 555, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 599 + ], + "type": "text", + "content": "[39] Desai Xie, Zhan Xu, Yicong Hong, Hao Tan, Difan Liu, Feng Liu, Arie Kaufman, and Yang Zhou. Progressive autoregressive video diffusion models. arXiv preprint arXiv:2410.08151, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "type": "text", + "content": "[40] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccrafter: Animating open-domain images with video diffusion priors. arXiv preprint arXiv:2310.12190, 2023. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[41] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 440 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[42] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[43] Tianwei Yin, Qiang Zhang, Richard Zhang, William T Freeman, Fredo Durand, Eli Shechtman, and Xun Huang. From slow bidirectional to fast causal video generators. arXiv preprint arXiv:2412.07772, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[44] Zhengqing Yuan, Ruoxi Chen, Zhaoxu Li, Haolong Jia, Lifang He, Chi Wang, and Lichao Sun. Mora: Enabling generalist video generation via a multi-agent framework. arXiv preprint arXiv:2403.13248, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[45] Shiwei Zhang, Jiayu Wang, Yingya Zhang, Kang Zhao, Hangjie Yuan, Zhiwu Qing, Xiang Wang, Deli Zhao, and Jingren Zhou. 12vgen-xl: High-quality image-to-video synthesis via cascaded diffusion models. 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 263, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 306 + ], + "type": "text", + "content": "[46] Canyu Zhao, Mingyu Liu, Wen Wang, Jianlong Yuan, Hao Chen, Bo Zhang, and Chunhua Shen. Moviedreamer: Hierarchical generation for coherent long visual sequence, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 308, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 351 + ], + "type": "text", + "content": "[47] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "text", + "content": "[48] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Ji-ashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. NeurIPS 2024, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 398, + 553, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 440 + ], + "type": "text", + "content": "[49] Shaobin Zhuang, Kunchang Li, Xinyuan Chen, Yaohui Wang, Ziwei Liu, Yu Qiao, and Yali Wang. Vlogger: Make your dream a vlog. arXiv preprint arXiv:2401.09414, 2024. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18206" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/dd0c7eba-b554-4e21-9d01-9fb4080c4a97_content_list.json b/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/dd0c7eba-b554-4e21-9d01-9fb4080c4a97_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d72ebec668d8d24de6fe7c566997292855afc623 --- /dev/null +++ b/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/dd0c7eba-b554-4e21-9d01-9fb4080c4a97_content_list.json @@ -0,0 +1,1666 @@ +[ + { + "type": "text", + "text": "ToolVQA: A Dataset for Multi-step Reasoning VQA with External Tools", + "text_level": 1, + "bbox": [ + 135, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shaofeng Yin Ting Lei Yang Liu* Wangxuan Institute of Computer Technology, Peking University", + "bbox": [ + 240, + 181, + 751, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yin_shaofeng@stu.pku.edu.cn {tinglei, yangliu}@pku.edu.cn", + "bbox": [ + 238, + 219, + 753, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Integrating external tools into Large Foundation Models (LFMs) has emerged as a promising approach to enhance their problem-solving capabilities. While existing studies have demonstrated strong performance in tool-augmented Visual Question Answering (VQA), recent benchmarks reveal significant gaps in real-world tool-use proficiency, particularly in functionally diverse multimodal settings requiring multi-step reasoning. In this work, we introduce ToolVQA, a large-scale multimodal dataset comprising 23K samples, designed to bridge this gap. Unlike previous datasets that rely on synthetic scenarios and simplified queries, ToolVQA features real-world visual contexts and challenging implicit multi-step reasoning tasks, better aligning with real user interactions. To construct this dataset, we propose ToolEngine, a novel data generation pipeline that employs image-guided Depth-First Search (DFS) with a Longest Common Subsequence (LCS)-based example matching mechanism to simulate human-like tool-use reasoning. ToolVQA encompasses 10 multimodal tools across 7 diverse domains, with an average inference length of 2.78 reasoning steps per sample. The LLaVA-7B model fine-tuned on ToolVQA not only achieves impressive performance on the ToolVQA test set, but also surpasses the large closed-source model GPT-3.5-turbo on five out-of-distribution (OOD) datasets, showing strong generalizability in real-world tool-use scenarios. Code is available at https://github.com/Fugtemypt123/ToolVQA-release.", + "bbox": [ + 88, + 301, + 485, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 741, + 220, + 757 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Integrating external tools into Large Foundation Models (LFMs) [31] is a promising way to build versatile AI assistants. By leveraging external tools, LFMs can break down complex tasks into smaller, more manageable subtasks, each handled by a specialized tool. Prior works [7, 10, 11, 33, 37] show strong performance on traditional Visual Question Answering (VQA) benchmarks using extern", + "bbox": [ + 89, + 767, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Non Real-World Scenarios", + "text_level": 1, + "bbox": [ + 545, + 282, + 674, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5fb20371181387a9671fca03aff1c85993942266b1e06d4fedd8f184bc448f26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 299, + 586, + 325 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n2 x ^ {2} - 1 2 = 0\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 301, + 676, + 319 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/83aea2f60d18d44e2e35805ed5a13d2ee244cbdd04f0df59ee00139b8fb1732f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 542, + 335, + 560, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "From:Shanghai To:Beijing", + "bbox": [ + 573, + 335, + 666, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Speed:", + "bbox": [ + 576, + 347, + 596, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "250 km/h", + "bbox": [ + 630, + 347, + 663, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(Overly simplistic synthetic scenarios)", + "bbox": [ + 540, + 367, + 678, + 376 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Non Real-World Queries", + "text_level": 1, + "bbox": [ + 549, + 400, + 669, + 409 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Based on the text 'James Hutton is often viewed as the first modern geologist...', could you help me find out who the first modern geologist is?", + "bbox": [ + 535, + 415, + 679, + 449 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(Simple retrieve, no tool use)", + "bbox": [ + 555, + 450, + 663, + 457 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can you help me find popular videos and trending gaming videos. I'm using the 'Cheap YouTube API' tool. (Explicit tool use, no reasoning)", + "bbox": [ + 534, + 465, + 681, + 497 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Find me some interesting news articles about the culinary world. Additionally, provide me with the current threshold securities list for NVIDIA's stock. (Just tool list, no multi-step reasoning)", + "bbox": [ + 534, + 505, + 676, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Real-World Setting", + "text_level": 1, + "bbox": [ + 754, + 282, + 846, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f0c4732c18d2011d9a0fb34ecdf24e773a6c943030d6e62c55ffa727c2ac4881.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 295, + 795, + 357 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c213c1444745538399e00d78ad01501a71c5f019d5a7b06ce0966362d12f5bb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 296, + 887, + 356 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c43bf96797156d8322e7dff3ed906168cf2d888a04bdfe0c64299600a2235337.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 372, + 733, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a bowl of salad, a sandwich, a sandwich of", + "bbox": [ + 736, + 373, + 790, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and a bottle of beer", + "bbox": [ + 736, + 388, + 790, + 393 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a2d1d008694823f1b0d6450fb3296c6e4e858cedea6b1629a3d60ffbbafcb30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 393, + 787, + 430 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eef1f2be73b208412e8577ed42655b778145cceb77262b0deb811d3683b2b583.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 807, + 372, + 883, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a22fbf87ed9ca6623669e86e370614171a9a99632620fc62893dceecd6c72c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 807, + 402, + 882, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/13a6a9027e41233b41a46d31e33001dcbabf30f21836ca84fa66f0b0f782336a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 431, + 733, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bridgeport Rewiring Company closed in 2019.", + "bbox": [ + 741, + 446, + 787, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2024-2019=5", + "bbox": [ + 741, + 468, + 782, + 476 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Q: How many years has it been since the brewery that produced this beer ceased operations? A: 5", + "bbox": [ + 715, + 500, + 787, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f0c1c23afc4d881dc16d83c41b674c15e69a29880981ad7c205abe89504c7580.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 459, + 883, + 507 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8717f26408b739818d5b3a8e9179a4ee1ad09a8e432378113083908b9c2eec85.jpg", + "image_caption": [ + "Figure 1. Our real-world setting includes (1) complex visual scenarios with real-world context; (2) challenging queries with an implicit multi-step reasoning process. Existing datasets (left) do not meet these requirements, while our ToolVQA (right) does." + ], + "image_footnote": [], + "bbox": [ + 807, + 510, + 883, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nal tools. Recent works [24, 27, 39] highlight the limitations of current LFMs in their tool-use capabilities within challenging contexts. These contexts require functionally diverse multimodal tools and multi-step reasoning processes associated with tool usage, posing major challenges to LFMs' tool-use proficiency.", + "bbox": [ + 509, + 659, + 906, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One effective approach to enhancing LFM's tool-use capability is fine-tuning on large-scale datasets [8, 10, 12, 33]. However, there remains a significant gap between existing large-scale datasets and real-world user needs. Through the study of real user logs, we find that real-world tool usage involves (1) real-world scenarios: complex visual scenarios with real-world context (e.g. real-taken photos rather than synthetic images), and (2) real-world queries: challenging queries with an implicit multi-step reasoning process. However, existing datasets often lack these two criti", + "bbox": [ + 509, + 750, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4424", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cal properties, as shown in Fig. 1. The scenarios are synthesized instead of real-taken [8, 39], resulting in simplified scenes that differ from real-world scenarios in complexity. The queries require only single-step reasoning or explicitly provide hints about the reasoning process [10, 30], such as \"using the Cheap YouTube API tool\". These setups simplify the task and may lead to a gap between synthetic and real-world tool usage, which limits our ability to analyze LFMs' performance on this task. Additionally, some datasets [27, 39] rely on costly human annotations, making them difficult to scale, which limits fine-tuning to enhance LFMs' tool-using capabilities.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge these gaps, we introduce ToolEngine, a data construction pipeline designed to generate multi-step reasoning VQA data for tool usage. ToolEngine employs an image-guided Depth-First Search (DFS) approach [30] to simulate the human-like multi-step reasoning process under the guidance of real user context examples. To facilitate multi-step reasoning in data, we introduce a Longest Common Subsequence (LCS)-based example matching mechanism. During each step in the iterative search process, this mechanism matches different real-world tool-use examples based on the previous search trajectory, guiding the controller in making the next selection. This mechanism enables: (1) Extracting rich visual information from the input image; (2) Integrating this information during the reasoning process to propose challenging queries. By freeing query generation from the constraints of fixed templates [8, 24, 30], ToolEngine enhances the utilization of visual information, consequently increasing the reasoning complexity of the generated queries.", + "bbox": [ + 91, + 275, + 482, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We leverage ToolEngine to construct ToolVQA, a large-scale multimodal tool-usage dataset comprising 23K samples. We assess data quality via a manual evaluation conducted on a randomly sampled 4k subset of the training set, which yields an accuracy of $90.8\\%$ . The test set is carefully filtered and fully human-annotated. As shown in Tab. 1, ToolVQA outperforms prior tool-use datasets across multiple perspectives, providing a more comprehensive representation of real-world scenarios. We evaluate state-of-the-art (SOTA) LFMs on ToolVQA and observe that: (1) Larger LFMs demonstrate superior reasoning capabilities. (2) LFMs consistently underperform relative to humans, primarily due to their inability to integrate new information introduced in multi-turn dialogues. We fine-tune LLaVA-7B [21] on ToolVQA, and the model demonstrated performance surpassing or approaching that of the large closed-source model GPT-3.5-Turbo on both our test set and five out-of-distribution (OOD) benchmarks—TextVQA [36], TallyQA [1], Infoseek [4], TEMPLAMA [6], and GTA [39], which include unseen tasks and tools. In summary, our contributions are as follows:", + "bbox": [ + 91, + 563, + 482, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. ToolEngine: We introduce ToolEngine, a VQA data", + "bbox": [ + 91, + 885, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "construction engine that generates reasonable multi-step tool-use trajectories from unannotated images, and constructs challenging question-answer pairs accordingly.", + "bbox": [ + 532, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. ToolVQA: We propose ToolVQA, a large-scale multimodal tool-use dataset that encompasses 10 multimodal tools across 7 diverse domains, providing a benchmark for evaluating and advancing LFMs' tool-use capabilities on real-world tasks.", + "3. Fine-tuning: We fine-tune LLaVA-7B on ToolVQA to obtain a new tool-use agent, which outperforms large closed-source models like GPT-3.5-Turbo on six benchmarks (ToolVQA and the other five OOD benchmarks)." + ], + "bbox": [ + 511, + 136, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 289, + 653, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Datasets and Benchmarks for Tool Agent", + "text_level": 1, + "bbox": [ + 511, + 314, + 864, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the rise of real-world tool-use applications, many works have attempted to build datasets and benchmarks to evaluate the tool-use capabilities of LFMs from different perspectives. Text-based datasets [12, 19, 29, 30, 32, 44, 47, 50] collect a large number of real-world APIs to build large-scale, diverse text-based tool-use examples. On the other hand, multimodal datasets [1, 4, 13, 24, 25, 27, 35, 36, 39, 40] integrate visual input into the query process, making queries challenging and more adaptable to real-world needs. However, they either fail to capture real-world scenarios or rely on expensive human annotation. MM-Traj [8] is a recent work which also constructs large-scale training data using an automated pipeline, but has the following limitations: (1) Synthesizes over-simplified PDF files as the multimodal contexts; (2) Their answers are long-form texts that have not been verified for correctness.", + "bbox": [ + 511, + 337, + 906, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. LFM-based Tool Agents", + "text_level": 1, + "bbox": [ + 511, + 589, + 735, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most recent LFMs have demonstrated impressive reasoning capabilities on numerous zero-shot tasks, which makes it possible to build LFM-based tool agents for real-world tasks. Mainstream approaches include in-context learning [3, 9, 22, 37, 42, 45, 46] and instruction fine-tuning [10, 28, 30, 33, 38]. However, these agents still face challenges in real-world applications. Additionally, recent advancements in web navigation agents [14, 16, 17, 23, 41, 43, 48] have primarily focused on GUI-based interactions, often designed as browser plugins, but their applicability is limited to web-based tasks. In contrast, our ToolEngine extends to a broader range of real-world scenarios.", + "bbox": [ + 511, + 611, + 906, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Other VQA Datasets", + "text_level": 1, + "bbox": [ + 511, + 801, + 710, + 818 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Question Answering (VQA) is one of the most fundamental tasks in multimodal learning. Early VQA datasets [2, 5, 13, 18, 20, 34, 49] primarily focus on simple commonsense question-answering, which typically consists of tasks that humans can easily solve. Recently, some", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4425", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/e1f21e94af63c72f601a9a2fbef8d17fbd0306cf5f71b127f1a7f2511f26136a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetNumber of SamplesNumber of ToolsMultrimodal Inputs?Real-World S/Q Setting?Evaluable Answers?Reasoning Complexity↑
ToolBench [30]126.5K3451X- / XX1.43
ToolQA [50]1.5K13X- / ✓✓1.35
Infoseek [4]1.3M1✓✓ / ✓✓1.13
VPD [10]*310.5K4✓✓ / ✓--
M&m's [24]1.6K33✓✓ / XX1.63
GAIA [27]†0.5K-✓✓ / ✓✓2.61
GTA [39]0.2K14✓X/ ✓✓1.86
MM-Traj [8]‡20K9✓X/ ✓X1.77
ToolVQA (Ours)23.7K10✓✓ / ✓✓2.38
", + "bbox": [ + 142, + 88, + 857, + 284 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison between ToolVQA and other datasets. Real-World S/Q Setting: whether each sample in the dataset involves: (1) Scenario—utilizes real rather than synthetic multimodal inputs; (2) Query—implicitly embeds tool usage and reasoning within the query. Reasoning Complexity: the average reasoning depth of humans for each task. VPD* constructs trajectories from existing QA pairs and is not publicly accessible. GAIA†'s each sample corresponds to a distinct set of human-annotated tools. ‡MM-Traj generates over-simplified PDF files as multimodal contexts, and their answers are long-form texts that have not been verified for correctness.", + "bbox": [ + 89, + 294, + 906, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "studies [1, 4, 10, 26, 36] begin exploring more challenging forms of VQA, which require various tools like Wikipedia, OCR, Object Detection, etc. However, these datasets typically only evaluate LFM's ability to use one specific tool instead of collaboration between multiple tools, limiting LFM's multi-step reasoning ability within a tool-agent framework. In contrast, ToolVQA contains high-quality and diverse multi-step tool usage traces. It encompasses 10 multimodal tools across 7 diverse domains, with an average inference length of 2.78 reasoning steps per sample.", + "bbox": [ + 88, + 390, + 480, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Data Collection", + "text_level": 1, + "bbox": [ + 89, + 569, + 245, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task Formulation", + "text_level": 1, + "bbox": [ + 89, + 597, + 264, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tool-use VQA Task. Given the image scenario $\\mathcal{I}$ , the toolset $\\mathcal{T}$ , and the user query $\\mathcal{Q}$ , a tool agent is required to provide an answer to the query $\\mathcal{A}$ and the tool-use trajectory $\\mathcal{P} = [(t_i, a_i)]_{i=1}^n$ , in which $t_i \\in \\mathcal{T}$ represents the tool used in step $i$ and $a_i$ represents the calling arguments for $t_i$ . Consequently, each sample is a five-tuple $(\\mathcal{I}, \\mathcal{T}, \\mathcal{Q}, \\mathcal{A}, \\mathcal{P})$ .", + "bbox": [ + 89, + 625, + 482, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Toolset and Image Source. We collect 10 tools with various functions as our toolset, including (1) Perception: ImageCaption, OCR, ObjectDetection, and RegionDescription. (2) Operation: DrawBox, GoogleSearch. (3) Logic: Calculator, Plot, ItemCount. (4) Creativity: TextToImage. These tools cover a wide range of real-world tasks and offer strong generalization capabilities (e.g. GoogleSearch can retrieve any external knowledge). The different sources of image $\\mathcal{I}$ and the corresponding toolset $\\mathcal{T}$ are shown in Fig. 2. These sources are diverse and complex, including real-world photography, e-commerce products, school exam charts, and other practical scenarios.", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/32e94094f293334f51ac2db55851378ee6b650d5dfeb49d8afaa8c352f42a3dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 390, + 862, + 470 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3ad9bd00f9213a45a40c701e168215cfe855292ecb258b1507e8ecd29af2e54d.jpg", + "image_caption": [ + "Figure 2. Image sources and corresponding tools of ToolVQA. We filter out overly simplistic tables and images from the data sources." + ], + "image_footnote": [], + "bbox": [ + 555, + 473, + 859, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. ToolEngine", + "text_level": 1, + "bbox": [ + 511, + 786, + 638, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We specify the definition of our real-world setting into the following criteria: (1) Align with human users: Each sample's scenario $\\mathcal{I}$ and query $\\mathcal{Q}$ should align with real-world user needs. We have filtered out uninformative scenarios $\\mathcal{I}$ , but it remains necessary to ensure that the tool-use queries $\\mathcal{Q}$ correspond to human demands. (2) Real-world", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4426", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ea19b08f67518f02dbd735eb50cde54c647096a479758998eac6756a23c6e66e.jpg", + "image_caption": [ + "Figure 3. The pipeline of ToolEngine, contains three core components: Real-world Example Construction, Image-guided DFS on Tool Graph, and LCS-based Example Matching. Given an input image, we perform DFS on the complete tool graph. At each step, an LFM controller generates the next tool's name and arguments, guided by the image, current tool-use trajectory, and matched examples. Once DFS is complete, the tool-use trajectory is determined, and then used to generate the query and answer." + ], + "image_footnote": [], + "bbox": [ + 130, + 97, + 872, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "deployed tools: Real-world tool results are complex distributions with noise, and the model must learn how to deal with the noise from real results. Some work [27, 28] uses simulated tool outputs to reduce noise, but we believe this reduces the realism of the training environment. (3) Multi-step reasoning: The tool-use trajectory $\\mathcal{P}$ should involve multiple steps. Each step in it must have an inherent logical connection, forming a complete problem-solving trajectory. It should not include ineffective steps (e.g. apply OCR to an image without text), or contain non-related tool usage in a single task (e.g. use \"culinary world\" and \"NVIDIA's stock\" sequentially, which should be split into two separate tasks). We propose ToolEngine, an automated data synthesis pipeline that ensures these conditions. Fig. 3 illustrates the three core components of ToolEngine:", + "bbox": [ + 88, + 508, + 485, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Real-World Example Construction. To better reflect real user needs, we incorporate human prior tool-use knowledge into the LFM controller through in-context examples. Human-constructed examples provide prior knowledge about typical tool-use scenarios (e.g. images containing specific entities may require a search tool, while tables with multi-digit numbers may necessitate a calculator). This information is crucial for synthesizing real-world queries. Therefore, we ask human users to construct a small set of real-world examples $\\mathcal{E} = \\{(\\mathcal{I},\\mathcal{T},\\mathcal{Q},\\mathcal{A},\\mathcal{P})\\}$ . Each", + "bbox": [ + 88, + 750, + 485, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "example in this set captures a logically valid trajectory, serving as prior knowledge for automatic construction. This approach not only aligns with how humans intuitively associate tools but also provides a structured, intuitive foundation for the construction process.", + "bbox": [ + 511, + 508, + 906, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-guided DFS on Tool Graph. To construct a multi-step tool-use trajectory, we employ image-guided Depth-First Search (DFS) on a tool graph to construct a reasonable tool-use trajectory for each image. While LFM can directly generate reasoning sequences, unrealistic tool invocations often lead to less reliable tool-use trajectories in practice. Our DFS trajectory involves real tool calls to extract detailed information from the input image, which is then used to generate challenging question-answer pairs. Specifically, we utilize the most advanced LFM (ChatGPT-4o-latest), denoted as $\\mathcal{M}$ , as the controller to select the tool at each step and generate the corresponding arguments. Formally:", + "bbox": [ + 511, + 584, + 908, + 767 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {i} = \\mathcal {M} (\\text {c h o i c e s} = \\mathcal {T}, \\text {i m a g e} = \\mathcal {I}, \\text {e x a m p l e s} = \\operatorname {R e t} (\\mathcal {E}, \\mathcal {P} _ {i - 1}))\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 771, + 908, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\na _ {i} = \\mathcal {M} (t o o l = t _ {i}, i m a g e = \\mathcal {I}, e x a m p l e s = \\operatorname {R e t} (\\mathcal {E}, \\mathcal {P} _ {i}))\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 789, + 867, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LCS-based Example-Matching. We combine the information obtained from different tools into a reasonable reasoning trajectory to perform multi-step reasoning. Previous DFS-based methods typically match a fixed group of examples, which limited their ability to combine sufficient information for multi-step reasoning. As illustrated in Fig. 4, we", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4427", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/da01cb65839426249d40d19d6191ec62278e3a9a1b4188a2d0366cd5fabc1748.jpg", + "image_caption": [ + "Figure 4. Comparison between different matching strategies. When a fixed example limits the diversity of generation, LCS matching can integrate multiple examples to enhance generation." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 478, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "propose an example-matching method based on the Longest Common Subsequence (LCS) algorithm, which enables dynamic matching during DFS. This approach better leverages the information extraction capabilities of diverse examples, facilitating more effective multi-step reasoning. Specifically, in step $i$ , the current solution process is denoted as $\\mathcal{P}_i = [(t_1,a_1),(t_2,a_2),\\dots ,(t_i,a_i)]$ and the set of example solution processes as $\\mathcal{P}_e\\in \\mathcal{E}$ . At each step, we perform the LCS matching between $\\mathcal{P}_i$ and each element of $\\mathcal{P}^e$ , retrieving the top-k examples in $\\mathcal{P}^e$ with the highest matching degree. Formally:", + "bbox": [ + 89, + 300, + 483, + 467 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e t} (\\mathcal {E}, \\mathcal {P} _ {i}) = \\operatorname {T o p K} _ {\\mathcal {P} ^ {e} \\in \\mathcal {E}} \\left\\{\\operatorname {L C S} (\\mathcal {P} ^ {e}, \\mathcal {P} _ {i}) \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 474, + 429, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This approach enables us to adaptively match each image with its most relevant example, while iteratively refining the match as new information is acquired at each step.", + "bbox": [ + 89, + 507, + 483, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human Annotations and Analysis. After the data synthesis process, we invite 10 human annotators to filter the generated data and re-associate the highest-quality subset, resulting in 2,550 test samples. The initial synthetic data is estimated to have a $90.8\\%$ accuracy in human validation on 4k randomly sampled data, which is higher than that reported in previous synthetic data methods [8, 24]. Notably, the ChatGPT-4o-latest model, which we use to generate questions, demonstrates a low accuracy rate (less than $40\\%$ ) in solving them. This highlights the effectiveness of our ToolEngine pipeline in decoupling complex reasoning into a series of simpler, single-step tasks. While ChatGPT-4o-latest excels in handling $\\mathcal{I} \\to \\mathcal{P} \\to \\mathcal{Q}, \\mathcal{A}$ process for straightforward, step-by-step reasoning, it faces challenges with the more challenging end-to-end $\\mathcal{I}, \\mathcal{Q} \\to \\mathcal{P} \\to \\mathcal{A}$ reasoning. This observation indicates a clear gap between LFM's single-step and multi-step reasoning capabilities.", + "bbox": [ + 89, + 553, + 483, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Dataset Statistics", + "text_level": 1, + "bbox": [ + 89, + 816, + 259, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tab. 2 presents the basic statistics of ToolVQA, which encompasses various question types (with both text and image as answers) and complex reasoning processes (with an average reasoning trajectory length of 2.78 steps). Notably,", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ToolVQA's questions and answers are concisely formulated (with an average token count of 15.74 for questions and 2.69 for answers). In contrast, previous synthetic datasets [8, 24] often fail to control answer length, leading to hallucinated content within the answers, which is detrimental to training. Fig. 5 demonstrates that our tool frequency aligns with real-world user needs, reinforcing the real-world setting of our dataset. Fig. 6 presents the distribution of reasoning trajectory lengths in ToolVQA.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Ablation on ToolEngine", + "text_level": 1, + "bbox": [ + 511, + 237, + 732, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To verify whether ToolEngine truly improves the quality of synthetic data, we conducted an ablation study on different synthetic methods. We invite 10 users to evaluate 1k randomly sampled data based on four key metrics: (1) answer correctness (Acc.); (2) correlation (Cor.): whether each sample's image content and Q-A pair are aligned; (3) tool redundancy (Nec.): whether each tool call is necessary for obtaining the answer; and (4) reasoning complexity (R.C.): the average reasoning depth for each task (same as defined in Tab. 1). The results, presented in Tab. 3, show that Real-World Example and LCS matching significantly improved Acc., Nec., and R.C., which highlights their role in the DFS-based simulation of reasoning chains: (1) Real-World Example and LCS matching introduce prior knowledge aligned with image content, enhancing both answer accuracy and tool necessity. (2) Real-World Example and LCS matching incorporate different types of knowledge at each step, increasing the reasoning complexity. Furthermore, the absence of LCS matching has a significant impact on data quality (e.g. Acc. dropped from 90.8 to 41.6), indicating that a fixed set of examples cannot adapt to different scenarios. Dynamically matching appropriate examples is crucial for significantly enhancing the generalization ability of the synthetic pipeline. However, both methods had a limited impact on Cor., likely because the existing controller already effectively captures image content, minimizing errors caused by misleading examples.", + "bbox": [ + 511, + 260, + 906, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 683, + 645, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Training and Test Settings", + "text_level": 1, + "bbox": [ + 511, + 708, + 751, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Objective. In our dataset $\\mathcal{D}$ , for each datapoint $\\mathcal{E} = (\\mathcal{I},\\mathcal{T},\\mathcal{Q},\\mathcal{A},\\mathcal{P})$ contains $n$ rounds of dialogue ( $n - 1$ rounds for tool usage and one round for final answer), we train our tool agent using cross-entropy loss:", + "bbox": [ + 511, + 729, + 906, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\mathcal {E} \\sim \\mathcal {D}} \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} - \\log p \\left(t _ {i}, a _ {i}, r _ {i} \\mid \\mathcal {I}, \\mathcal {T}, \\mathcal {Q}, \\mathcal {P} _ {i - 1}\\right) \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 801, + 890, + 844 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Setting. We divide ToolVQA into training and test sets: the training set consists of 21,105 automatically generated samples (with a human verification accu", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4428", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f0581da53ba6bddb739037969a1ab2b5331473d60be5e1f629e1e41f57b47206.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
StatisticsNumber
Samples23655
Tool Calls65785
Text answers15806
Image answers7849
Query length15.74
Answer length2.69
Trajectory length2.78
", + "bbox": [ + 130, + 102, + 321, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/189e4ff5e10624ab50756d59c0f851d89545143760aef97fe169224840ee0a87.jpg", + "image_caption": [ + "Figure 5. Tool frequency." + ], + "image_footnote": [], + "bbox": [ + 328, + 97, + 576, + 243 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56531ecff38782fea97298190c0d2befefb583f0c22918bd695d327e26b18ea7.jpg", + "image_caption": [ + "Figure 6. Trajectory length distribution." + ], + "image_footnote": [], + "bbox": [ + 581, + 97, + 874, + 234 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6dbb94790b3c2f0c17bbf43dadc1a1f67c09e5635948635e6a99e1743811bcb1.jpg", + "table_caption": [ + "Table 2. Basic statistics." + ], + "table_footnote": [], + "table_body": "
MethodAcc.↑Cor. ↑Nec. ↑R.C. ↑
ToolEngine90.885.287.512.38
-w/o Example + LCS27.377.621.041.1
-w/o LCS41.681.454.261.61
", + "bbox": [ + 98, + 292, + 475, + 372 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Ablation study on different synthetic methods.", + "bbox": [ + 120, + 381, + 450, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "racy of $90.8\\%$ ), while the test set contains 2,550 humanreannotated samples. We fine-tune LLaVA-7B [21] in the Lego Agent framework [15].", + "bbox": [ + 88, + 425, + 482, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Test Setting. We apply two test modes: end-to-end and step-by-step modes, to comprehensively evaluate LFMs' performance under different settings. The end-to-end mode only focuses on the end-to-end VQA ability of LFMs. Formally, we evaluate the $\\mathcal{I} + \\mathcal{Q}\\rightarrow \\mathcal{A}$ process. The step-by-step mode focuses on evaluating LM's tool-usage capabilities, formally the $\\mathcal{I} + \\mathcal{Q} + \\mathcal{P}_i\\to \\mathcal{P}_{i + 1}$ process. We use four metrics [39]: (1) InstAcc (Inst.): the percentage of steps executed without errors; (2) ToolAcc (Tool.): the accuracy of tool selection; (3) ArgAcc (Arg.): the accuracy of tool's argument name prediction, and (4) SummAcc (Summ.): the accuracy of final answer's summary. Previous work has focused on the research of large language model (LLM)-based tool agents while ignoring vision-language model (VLM)-based tool agents and the end-to-end VQA abilities of VLMs. To bridge these gaps, our test involves three settings: (1) VLM, directly answering questions end-to-end, formally $\\mathcal{Q} + \\mathcal{I}\\rightarrow \\mathcal{A}$ , (2) VLM+tool, calling the tool through VLM, formally $\\mathcal{Q} + \\mathcal{I} + \\mathcal{T}\\rightarrow \\mathcal{P} + \\mathcal{A}$ , (3) LLM+tool, calling the tool through LLM, formally $\\mathcal{Q} + \\mathcal{T}\\rightarrow \\mathcal{P} + \\mathcal{A}$ (Note that image information can still be obtained through tools such as caption under this setting).", + "bbox": [ + 91, + 472, + 483, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 816, + 230, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scale. As shown in Tab. 4, the tool-use capabilities of LFMs generally improve with the increase in model parameters. For example, Qwen2-7B-instruct surpasses Qwen2-2B-instruct by $7.03\\%$ , and closed-sourced LM (larger than", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "100B) is generally better than open-sourced (2B,7B). After fine-tuning, our 7B model's performance (last 3 lines) is close to the large scale LFM GPT-3.5-Turbo. This demonstrates that fine-tuning significantly enhances the model's real-world tool-use capabilities, even in smaller models.", + "bbox": [ + 511, + 295, + 903, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. By comparing the performance of the same model under different settings (VLM, VLM+tool, and LLM+tool) in Tab. 4, we identify several interesting observations. (1) In models such as GPT-4o, Qwen2-7B/2B, and LLaVA-7B, the VLM+tool setting performs worse than the LLM+tool setting, suggesting that the visual modules in current LFM architectures are still ineffective at guiding tool usage. (2) For models including GPT-4o, Claude-3.5, LLaVA-7B, and Qwen2-2B, we find that VLM+tool underperforms VLM alone. This indicates that the noise introduced by tool usage can outweigh the benefits brought by the tools. However, the tuned LLaVA model demonstrates a clear progression: VLM $<$ LLM+tool $<$ VLM+tool, suggesting that fine-tuning can indeed enhance the model's ability to make effective use of tools while suppressing tool-induced noise. Nevertheless, the similar performance of the LLM+tool and VLM+tool settings implies that the visual module's potential in facilitating tool use has not yet been fully explored.", + "bbox": [ + 511, + 371, + 906, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bottlenecks. The last two rows of Tab. 4 indicate that the performance of fine-tuned models is often constrained by their weakest capabilities, which are typically argument prediction and answer summarization. While fine-tuning significantly improves instruction formatting and tool selection, it makes little progress in weaker capabilities. Specifically, instruction formatting and tool selection can be accomplished through memorizing patterns, as the number of reasonable tool invocation types in real-world scenarios is limited. However, argument prediction and answer summarization require the model to genuinely understand the new information returned by tools and extract meaningful responses from it. This information varies across different scenarios and queries, and even within multi-turn dialogues of the same scenario. Consequently, a high level of generalization capability in vision-language joint reasoning is essential to handle these issues in tool-use tasks.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4429", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3463d219daf11aab206bcb08441f1991560f27ef681e452b977cc2ddf70f8403.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSettingEnd-to-End ModeStep-by-Step Mode
Acc. (%)Tool Call (Times)Tool Error (Times)Inst. (%)Tool. (%)Arg. (%)Summ. (%)
Closed-source LFMs
ChatGPT-4o-latestVLM38.29------
VLM + tool34.962003180436.514.688.9256.1
LLM + tool37.624687196273.0241.1225.2765.45
Claude-3-5-sonnetVLM30.33------
VLM + tool30.5258552445.7116.947.3646.06
LLM + tool24.48183467266.8339.2913.9875.52
GPT-3.5-TurboLLM + tool18.373178132173.2430.4620.0858.18
Open-source LFMs
Qwen2-VL-7B-InstructVLM3.64------
VLM + tool5.861373105135.129.021.0938.18
LLM + tool6.781366108939.4611.23.5544.85
Qwen2-7B-InstructLLM + tool11.53177370363.124.458.269.7
LLaVA-v1.5-7BVLM8.57------
VLM + tool1.179684968416.399.4300.01
LLM + tool1.6610421104219.2515.03011.52
LLaMA-3-8b-instructLLM + tool4.936607654251.6314.130.5341.99
Qwen2-VL-2B-InstructVLM7.71------
VLM + tool2.14067391547.1612.70.4121.82
LLM + tool2.226892665536.3410.520.6832.12
Qwen2-1.5B-InstructLLM + tool4.57933614150.9521.454.5144.24
Tuned LLaVA-7B (Ours)VLM7.21------
VLM + tool18.8431161786.6261.6139.3430.91
LLM + tool18.43446268087.5162.340.5726.67
", + "bbox": [ + 151, + 88, + 846, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Results on ToolVQA's test set.", + "bbox": [ + 380, + 507, + 614, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Out-of-Distribution Benchmarks", + "text_level": 1, + "bbox": [ + 89, + 547, + 377, + 563 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To test the generalization capability of our tuned model, we conduct experiments on out-of-distribution (OOD) benchmarks including (1) Seen tools, unseen queries: TextVQA [36], TallyQA [1], InfoSeek [4]; and (2) Unseen tools: GTA [39], TEMPLAMA [6]. The results are shown in the Tab. 5. Our fine-tuned model significantly outperforms the baseline model LLaVA-7B by $5.8\\%$ , $4.2\\%$ , $8.6\\%$ , $21.17\\%$ , and $18.37\\%$ accuracy on these datasets, showing excellent generalizability. Its performance also surpasses the large closed-source model GPT-3.5-Turbo on unseen queries. However, its performance on TEMPLAMA is inferior to GPT-3.5-Turbo, possibly because GPT encountered similar data during its training (due to its great coverage of extensive training data), whereas our model had no exposure to such data at all.", + "bbox": [ + 89, + 573, + 483, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Few-shot In-context Learning", + "text_level": 1, + "bbox": [ + 89, + 816, + 354, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Few-shot in-context learning (ICL) has become a common and important technique in tool-use tasks. However, prior model fine-tuning work often overlooks evaluating the ICL capabilities of fine-tuned models. We believe that fine-", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tuning and ICL can be complementary, offering significant potential for improvement. Tab. 6 shows few-shot ICL performance on ToolVQA. GPT-4o steadily improves with more shots, while GPT-3.5 and LLaVA both show limited gains and even decline slightly at 10-shot. This may be due to long contexts impairing the LLM's ability to interpret precise tool-use instructions. Fine-tuned LLaVA outperforms the baseline and continues to benefit from ICL. These results suggest that (1) compared to fine-tuning, ICL's improvement is limited and affected by context lengths, and (2) fine-tuned models can still benefit from ICL, highlighting the complementary value of these two approaches in enhancing tool-use performance.", + "bbox": [ + 511, + 549, + 906, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 757, + 666, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further evaluate the quality of the fine-tuning data from multiple perspectives, we conduct an ablation study of the different training strategies. As shown in Tab. 7, removing the ImageCaption tool significantly degrades both accuracy and parameter correctness. This highlights the model's reliance on high-quality captions for initial scene understanding, likely due to its limited inherent visual capability for fine-grained information extraction. When all tools", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4430", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d894eae8e6fb7369d8d684ebd1d9e0c2974510f893e741b0d0001e291cdce705.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTextVQA [36]TallyQA [1]InfoSeek [4]GTA [39]TEMPLAMA [6]
GPT-3.5-Turbo36.36111.323.6233.67
LLaVA-7B41.260.15.212.123.06
Tuned LLaVA-7B4764.313.833.2921.43
", + "bbox": [ + 161, + 88, + 836, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f650cf0d5a8ebf989e7ac41b90634206fe7c1992e73fa766f1b94c15acc6b25c.jpg", + "table_caption": [ + "Table 5. Results on OOD benchmarks." + ], + "table_footnote": [], + "table_body": "
ModelZero-shot1-shot5-shot10-shot
GPT-4o34.9637.2038.4138.63
GPT-3.5-Turbo18.3719.8320.4019.04
LLaVA-7B1.173.454.273.67
Tuned LLaVA-7B18.8019.4121.1320.69
", + "bbox": [ + 96, + 204, + 485, + 292 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0193ed63fdc784f5658d82a55dc108d2183f03a0ac1525b5e992f342297a3096.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSettingAcc.Tool CallTool Error
Tuned LLaVA-7BVLM + tool18.84311617
-w/o CaptionVLM + tool11.11861192
-w/o PVLM8.01--
-w/o P + ILLM0.18--
", + "bbox": [ + 500, + 204, + 903, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Few-shot performance on ToolVQA.", + "bbox": [ + 151, + 300, + 406, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation study on training methods.", + "bbox": [ + 581, + 300, + 816, + 313 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e17e61b0d7afdc57519a0e7fc829a3c3f1fd9106b1f61b08059026f3baf96e63.jpg", + "image_caption": [ + "Figure 7. Visualization of the main error types made by tuned LLaVA." + ], + "image_footnote": [], + "bbox": [ + 132, + 330, + 859, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are removed and ToolVQA is reduced to a single-hop VQA model, performance drops substantially, underscoring the necessity of multi-step tool usage. In the absence of both tools and images, performance approaches zero, indicating that the model is heavily dependent on visual input when answering questions in ToolVQA.", + "bbox": [ + 89, + 584, + 482, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Error Analysis", + "text_level": 1, + "bbox": [ + 89, + 681, + 241, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 7, we randomly select 100 errors made by tuned LLaVA, and manually count and classify the main error types. The two main error types are: (1) Argument prediction error, for example, missing essential keywords \"age\" during a search. (2) Answer summary error, for example, extracting incorrect information \"1,300\" from tool outputs. The common cause of these errors is the model's failure to recognize key terms (e.g. \"age\" and \"pound\") in the tool outputs, even though this should not be a challenging task for a language model. This suggests that the model still struggles with dynamically adapting to additional information provided by tools. Since multi-step reasoning often suffers from error accumulation (early errors propagate to sub", + "bbox": [ + 89, + 703, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sequent steps), errors in processing tool results at any step will affect the results. Fine-tuning has limited improvement on the capability for dynamically processing tool results, but ToolVQA still provides a valuable benchmark for evaluating this capability.", + "bbox": [ + 511, + 585, + 906, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present ToolVQA, a large multimodal dataset for real-world tool-augmented reasoning. Using our ToolEngine pipeline, we generate 23K real-world samples with implicit multi-step reasoning over 10 tools and 7 domains. This dataset challenges existing models beyond synthetic settings and better reflects real user needs. Experiments show that LLaVA-7B fine-tuned on ToolVQA not only excels on the in-domain test set but also outperforms GPT-3.5-turbo on five OOD benchmarks, establishing a new state-of-the-art for open-source models in complex tool-use VQA. ToolVQA serves as both a benchmark and a training ground for developing more capable, generalizable tool-using agents.", + "bbox": [ + 511, + 703, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4431", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the grants from the National Natural Science Foundation of China 62372014, Beijing Nova Program and Beijing Natural Science Foundation 4252040.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 167, + 187, + 184 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions, 2018. 2, 3, 7, 8", + "[2] Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C. Lawrence Zitnick, Dhruv Batra, and Devi Parikh. Vqa: Visual question answering, 2016. 2", + "[3] Lluis Castrejon, Thomas Mensink, Howard Zhou, Vittorio Ferrari, Andre Araujo, and Jasper Uijlings. Hammr: Hierarchical multimodal react agents for generic vqa, 2024. 2", + "[4] Yang Chen, Hexiang Hu, Yi Luan, Haitian Sun, Soravit Changpinyo, Alan Ritter, and Ming-Wei Chang. Can pre-trained vision and language models answer visual information-seeking questions?, 2023. 2, 3, 7, 8", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2", + "[6] Bhuwan Dhingra, Jeremy R. Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W. Cohen. Time-aware language models as temporal knowledge bases. Transactions of the Association for Computational Linguistics, 10:257-273, 2022. 2, 7, 8", + "[7] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn, 2023. 1", + "[8] Zhi Gao, Bofei Zhang, Pengxiang Li, Xiaojian Ma, Tao Yuan, Yue Fan, Yuwei Wu, Yunde Jia, Song-Chun Zhu, and Qing Li. Multi-modal agent tuning: Building a vlm-driven agent for efficient tool usage, 2025. 1, 2, 3, 5", + "[9] Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. Toolkengpt: Augmenting frozen language models with massive tools via tool embeddings, 2024. 2", + "[10] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models, 2024. 1, 2, 3", + "[11] Ziniu Hu, Ahmet Iscen, Chen Sun, Kai-Wei Chang, Yizhou Sun, David A Ross, Cordelia Schmid, and Alireza Fathi. Avis: Autonomous visual information seeking with large language model agent, 2023. 1", + "[12] Yue Huang, Jiawen Shi, Yuan Li, Chenrui Fan, Siyuan Wu, Qihui Zhang, Yixin Liu, Pan Zhou, Yao Wan, Neil Zhenqiang Gong, and Lichao Sun. Metatool benchmark for large language models: Deciding whether to use tools and which to use, 2024. 1, 2", + "[13] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional" + ], + "bbox": [ + 93, + 193, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2", + "[14] Faria Huq, Zora Zhiruo Wang, Frank F. Xu, Tianyue Ou, Shuyan Zhou, Jeffrey P. Bigham, and Graham Neubig. Cow-pilot: A framework for autonomous and human-agent collaborative web navigation, 2025. 2", + "[15] InternLM. Agentlego: Enhance llm agents with rich tool apis. https://github.com/InternLM/agentlego, 2024.6", + "[16] Iat Long Iong, Xiao Liu, Yuxuan Chen, Hanyu Lai, Shuntian Yao, Pengbo Shen, Hao Yu, Yuxiao Dong, and Jie Tang. OpenWebAgent: An open toolkit to enable web agents on large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 72-81, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2", + "[17] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks, 2024. 2", + "[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2", + "[19] Minghao Li, Feifan Song, Yu Bowen, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. Api-bank: A comprehensive benchmark for tool-augmented llms. In Conference on Empirical Methods in Natural Language Processing, 2023. 2", + "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2", + "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 6", + "[22] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In The 37th Conference on Neural Information Processing Systems (NeurIPS), 2023. 2", + "[23] Kaixin Ma, Hongming Zhang, Hongwei Wang, Xiaoman Pan, Wenhao Yu, and Dong Yu. Laser: LIm agent with statespace exploration for web navigation, 2024. 2", + "[24] Zixian Ma, Weikai Huang, Jieyu Zhang, Tanmay Gupta, and Ranjay Krishna. m&m's: A benchmark to evaluate tool-use for multi-step multi-modal tasks. EECV 2024, 2024. 1, 2, 3, 5", + "[25] Arjun Mani, Nobline Yoo, Will Hinthorn, and Olga Russakovsky. Point and ask: Incorporating pointing into visual question answering, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4432", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge, 2019. 3", + "[27] Grégoire Mialon, Clémentine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, and Thomas Scialom. Gaia: a benchmark for general ai assistants, 2023. 1, 2, 3, 4", + "[28] Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 2, 4", + "[29] Yun Peng, Shuqing Li, Wenwei Gu, Yichen Li, Wenxuan Wang, Cuiyun Gao, and Michael R. Lyu. Revisiting, benchmarking and exploring api recommendation: How far are we? IEEE Transactions on Software Engineering, 49:1876-1897, 2021. 2", + "[30] Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. Toollm: Facilitating large language models to master 16000+ real-world apis, 2023. 2, 3", + "[31] Yujia Qin, Shengding Hu, Yankai Lin, Weize Chen, Ning Ding, Ganqu Cui, Zheni Zeng, Yufei Huang, Chaojun Xiao, Chi Han, Yi Ren Fung, Yusheng Su, Huadong Wang, Cheng Qian, Runchu Tian, Kunlun Zhu, Shihao Liang, Xingyu Shen, Bokai Xu, Zhen Zhang, Yining Ye, Bowen Li, Ziwei Tang, Jing Yi, Yuzhang Zhu, Zhenning Dai, Lan Yan, Xin Cong, Yaxi Lu, Weilin Zhao, Yuxiang Huang, Junxi Yan, Xu Han, Xian Sun, Dahai Li, Jason Phang, Cheng Yang, Tongshuang Wu, Heng Ji, Zhiyuan Liu, and Maosong Sun. Tool learning with foundation models, 2024. 1", + "[32] Yangjun Ruan, Honghua Dong, Andrew Wang, Silviu Pitis, Yongchao Zhou, Jimmy Ba, Yann Dubois, Chris J. Maddison, and Tatsunori Hashimoto. Identifying the risks of lm agents with an lm-emulated sandbox, 2024. 2", + "[33] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools, 2023. 1, 2", + "[34] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge, 2022. 2", + "[35] Yongliang Shen, Kaitao Song, Xu Tan, Wenqi Zhang, Kan Ren, Siyu Yuan, Weiming Lu, Dongsheng Li, and Yueting Zhuang. Taskbench: Benchmarking large language models for task automation, 2024. 2", + "[36] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read, 2019. 2, 3, 7, 8", + "[37] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11888-11898, 2023. 1, 2", + "[38] Chenyu Wang, Weixin Luo, Qianyu Chen, Haonan Mai, Jindi Guo, Sixun Dong, Xiaohua (Michael) Xuan, Zhengxin Li, Lin Ma, and Shenghua Gao. Mllm-tool: A multimodal" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "large language model for tool agent learning. arXiv preprint arXiv:2401.10727, 2024. 2", + "[39] Jize Wang, Zerun Ma, Yining Li, Songyang Zhang, Cailian Chen, Kai Chen, and Xinyi Le. Gta: A benchmark for general tool agents, 2024. 1, 2, 3, 6, 7, 8", + "[40] Yan Wang, Yawen Zeng, Jingsheng Zheng, Xiaofen Xing, Jin Xu, and Xiangmin Xu. Videocot: A video chain-of-thought dataset with active annotation tool, 2024. 2", + "[41] Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. Agent workflow memory, 2024. 2", + "[42] Shirley Wu, Shiyu Zhao, Qian Huang, Kexin Huang, Michihiro Yasunaga, Kaidi Cao, Vassilis N. Ioannidis, Karthik Subbian, Jure Leskovec, and James Zou. Avatar: Optimizing llm agents for tool usage via contrastive reasoning, 2024. 2", + "[43] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments, 2024. 2", + "[44] Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering, 2018. 2", + "[45] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action, 2023. 2", + "[46] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent), 2024. 2", + "[47] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents, 2023. 2", + "[48] Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. Webarena: A realistic web environment for building autonomous agents, 2024. 2", + "[49] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images, 2016. 2", + "[50] Yuchen Zhuang, Yue Yu, Kuan Wang, Haotian Sun, and Chao Zhang. Toolqa: A dataset for llm question answering with external tools, 2023. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 729 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4433", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/dd0c7eba-b554-4e21-9d01-9fb4080c4a97_model.json b/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/dd0c7eba-b554-4e21-9d01-9fb4080c4a97_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b6970708bdd65a470e9e8080cf1b3bedc27cdf6f --- /dev/null +++ b/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/dd0c7eba-b554-4e21-9d01-9fb4080c4a97_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.865, + 0.154 + ], + "angle": 0, + "content": "ToolVQA: A Dataset for Multi-step Reasoning VQA with External Tools" + }, + { + "type": "text", + "bbox": [ + 0.241, + 0.182, + 0.753, + 0.22 + ], + "angle": 0, + "content": "Shaofeng Yin Ting Lei Yang Liu* Wangxuan Institute of Computer Technology, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.22, + 0.754, + 0.235 + ], + "angle": 0, + "content": "yin_shaofeng@stu.pku.edu.cn {tinglei, yangliu}@pku.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.71 + ], + "angle": 0, + "content": "Integrating external tools into Large Foundation Models (LFMs) has emerged as a promising approach to enhance their problem-solving capabilities. While existing studies have demonstrated strong performance in tool-augmented Visual Question Answering (VQA), recent benchmarks reveal significant gaps in real-world tool-use proficiency, particularly in functionally diverse multimodal settings requiring multi-step reasoning. In this work, we introduce ToolVQA, a large-scale multimodal dataset comprising 23K samples, designed to bridge this gap. Unlike previous datasets that rely on synthetic scenarios and simplified queries, ToolVQA features real-world visual contexts and challenging implicit multi-step reasoning tasks, better aligning with real user interactions. To construct this dataset, we propose ToolEngine, a novel data generation pipeline that employs image-guided Depth-First Search (DFS) with a Longest Common Subsequence (LCS)-based example matching mechanism to simulate human-like tool-use reasoning. ToolVQA encompasses 10 multimodal tools across 7 diverse domains, with an average inference length of 2.78 reasoning steps per sample. The LLaVA-7B model fine-tuned on ToolVQA not only achieves impressive performance on the ToolVQA test set, but also surpasses the large closed-source model GPT-3.5-turbo on five out-of-distribution (OOD) datasets, showing strong generalizability in real-world tool-use scenarios. Code is available at https://github.com/Fugtemypt123/ToolVQA-release." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.742, + 0.222, + 0.758 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Integrating external tools into Large Foundation Models (LFMs) [31] is a promising way to build versatile AI assistants. By leveraging external tools, LFMs can break down complex tasks into smaller, more manageable subtasks, each handled by a specialized tool. Prior works [7, 10, 11, 33, 37] show strong performance on traditional Visual Question Answering (VQA) benchmarks using extern" + }, + { + "type": "title", + "bbox": [ + 0.547, + 0.284, + 0.676, + 0.294 + ], + "angle": 0, + "content": "Non Real-World Scenarios" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.3, + 0.588, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.302, + 0.678, + 0.32 + ], + "angle": 0, + "content": "\\[\n2 x ^ {2} - 1 2 = 0\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.336, + 0.562, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.575, + 0.337, + 0.668, + 0.345 + ], + "angle": 0, + "content": "From:Shanghai To:Beijing" + }, + { + "type": "text", + "bbox": [ + 0.577, + 0.348, + 0.597, + 0.355 + ], + "angle": 0, + "content": "Speed:" + }, + { + "type": "text", + "bbox": [ + 0.632, + 0.348, + 0.664, + 0.355 + ], + "angle": 0, + "content": "250 km/h" + }, + { + "type": "text", + "bbox": [ + 0.541, + 0.368, + 0.679, + 0.377 + ], + "angle": 0, + "content": "(Overly simplistic synthetic scenarios)" + }, + { + "type": "title", + "bbox": [ + 0.55, + 0.401, + 0.67, + 0.41 + ], + "angle": 0, + "content": "Non Real-World Queries" + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.416, + 0.681, + 0.45 + ], + "angle": 0, + "content": "Based on the text 'James Hutton is often viewed as the first modern geologist...', could you help me find out who the first modern geologist is?" + }, + { + "type": "text", + "bbox": [ + 0.557, + 0.451, + 0.664, + 0.458 + ], + "angle": 0, + "content": "(Simple retrieve, no tool use)" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.466, + 0.682, + 0.498 + ], + "angle": 0, + "content": "Can you help me find popular videos and trending gaming videos. I'm using the 'Cheap YouTube API' tool. (Explicit tool use, no reasoning)" + }, + { + "type": "text", + "bbox": [ + 0.535, + 0.506, + 0.678, + 0.548 + ], + "angle": 0, + "content": "Find me some interesting news articles about the culinary world. Additionally, provide me with the current threshold securities list for NVIDIA's stock. (Just tool list, no multi-step reasoning)" + }, + { + "type": "title", + "bbox": [ + 0.755, + 0.284, + 0.848, + 0.294 + ], + "angle": 0, + "content": "Real-World Setting" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.296, + 0.796, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.297, + 0.888, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.373, + 0.734, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.737, + 0.374, + 0.791, + 0.389 + ], + "angle": 0, + "content": "a bowl of salad, a sandwich, a sandwich of" + }, + { + "type": "text", + "bbox": [ + 0.737, + 0.389, + 0.791, + 0.395 + ], + "angle": 0, + "content": "and a bottle of beer" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.395, + 0.788, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.373, + 0.884, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.404, + 0.883, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.433, + 0.735, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.743, + 0.448, + 0.788, + 0.468 + ], + "angle": 0, + "content": "Bridgeport Rewiring Company closed in 2019." + }, + { + "type": "text", + "bbox": [ + 0.743, + 0.469, + 0.784, + 0.477 + ], + "angle": 0, + "content": "2024-2019=5" + }, + { + "type": "text", + "bbox": [ + 0.716, + 0.501, + 0.788, + 0.544 + ], + "angle": 0, + "content": "Q: How many years has it been since the brewery that produced this beer ceased operations? A: 5" + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.46, + 0.884, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.511, + 0.884, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.577, + 0.907, + 0.633 + ], + "angle": 0, + "content": "Figure 1. Our real-world setting includes (1) complex visual scenarios with real-world context; (2) challenging queries with an implicit multi-step reasoning process. Existing datasets (left) do not meet these requirements, while our ToolVQA (right) does." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.66, + 0.907, + 0.75 + ], + "angle": 0, + "content": "nal tools. Recent works [24, 27, 39] highlight the limitations of current LFMs in their tool-use capabilities within challenging contexts. These contexts require functionally diverse multimodal tools and multi-step reasoning processes associated with tool usage, posing major challenges to LFMs' tool-use proficiency." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "One effective approach to enhancing LFM's tool-use capability is fine-tuning on large-scale datasets [8, 10, 12, 33]. However, there remains a significant gap between existing large-scale datasets and real-world user needs. Through the study of real user logs, we find that real-world tool usage involves (1) real-world scenarios: complex visual scenarios with real-world context (e.g. real-taken photos rather than synthetic images), and (2) real-world queries: challenging queries with an implicit multi-step reasoning process. However, existing datasets often lack these two criti" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4424" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "cal properties, as shown in Fig. 1. The scenarios are synthesized instead of real-taken [8, 39], resulting in simplified scenes that differ from real-world scenarios in complexity. The queries require only single-step reasoning or explicitly provide hints about the reasoning process [10, 30], such as \"using the Cheap YouTube API tool\". These setups simplify the task and may lead to a gap between synthetic and real-world tool usage, which limits our ability to analyze LFMs' performance on this task. Additionally, some datasets [27, 39] rely on costly human annotations, making them difficult to scale, which limits fine-tuning to enhance LFMs' tool-using capabilities." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.276, + 0.483, + 0.562 + ], + "angle": 0, + "content": "To bridge these gaps, we introduce ToolEngine, a data construction pipeline designed to generate multi-step reasoning VQA data for tool usage. ToolEngine employs an image-guided Depth-First Search (DFS) approach [30] to simulate the human-like multi-step reasoning process under the guidance of real user context examples. To facilitate multi-step reasoning in data, we introduce a Longest Common Subsequence (LCS)-based example matching mechanism. During each step in the iterative search process, this mechanism matches different real-world tool-use examples based on the previous search trajectory, guiding the controller in making the next selection. This mechanism enables: (1) Extracting rich visual information from the input image; (2) Integrating this information during the reasoning process to propose challenging queries. By freeing query generation from the constraints of fixed templates [8, 24, 30], ToolEngine enhances the utilization of visual information, consequently increasing the reasoning complexity of the generated queries." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.564, + 0.483, + 0.88 + ], + "angle": 0, + "content": "We leverage ToolEngine to construct ToolVQA, a large-scale multimodal tool-usage dataset comprising 23K samples. We assess data quality via a manual evaluation conducted on a randomly sampled 4k subset of the training set, which yields an accuracy of \\(90.8\\%\\). The test set is carefully filtered and fully human-annotated. As shown in Tab. 1, ToolVQA outperforms prior tool-use datasets across multiple perspectives, providing a more comprehensive representation of real-world scenarios. We evaluate state-of-the-art (SOTA) LFMs on ToolVQA and observe that: (1) Larger LFMs demonstrate superior reasoning capabilities. (2) LFMs consistently underperform relative to humans, primarily due to their inability to integrate new information introduced in multi-turn dialogues. We fine-tune LLaVA-7B [21] on ToolVQA, and the model demonstrated performance surpassing or approaching that of the large closed-source model GPT-3.5-Turbo on both our test set and five out-of-distribution (OOD) benchmarks—TextVQA [36], TallyQA [1], Infoseek [4], TEMPLAMA [6], and GTA [39], which include unseen tasks and tools. In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1. ToolEngine: We introduce ToolEngine, a VQA data" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "construction engine that generates reasonable multi-step tool-use trajectories from unannotated images, and constructs challenging question-answer pairs accordingly." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.211 + ], + "angle": 0, + "content": "2. ToolVQA: We propose ToolVQA, a large-scale multimodal tool-use dataset that encompasses 10 multimodal tools across 7 diverse domains, providing a benchmark for evaluating and advancing LFMs' tool-use capabilities on real-world tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.212, + 0.905, + 0.272 + ], + "angle": 0, + "content": "3. Fine-tuning: We fine-tune LLaVA-7B on ToolVQA to obtain a new tool-use agent, which outperforms large closed-source models like GPT-3.5-Turbo on six benchmarks (ToolVQA and the other five OOD benchmarks)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.29, + 0.655, + 0.307 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.315, + 0.866, + 0.332 + ], + "angle": 0, + "content": "2.1. Datasets and Benchmarks for Tool Agent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.907, + 0.579 + ], + "angle": 0, + "content": "With the rise of real-world tool-use applications, many works have attempted to build datasets and benchmarks to evaluate the tool-use capabilities of LFMs from different perspectives. Text-based datasets [12, 19, 29, 30, 32, 44, 47, 50] collect a large number of real-world APIs to build large-scale, diverse text-based tool-use examples. On the other hand, multimodal datasets [1, 4, 13, 24, 25, 27, 35, 36, 39, 40] integrate visual input into the query process, making queries challenging and more adaptable to real-world needs. However, they either fail to capture real-world scenarios or rely on expensive human annotation. MM-Traj [8] is a recent work which also constructs large-scale training data using an automated pipeline, but has the following limitations: (1) Synthesizes over-simplified PDF files as the multimodal contexts; (2) Their answers are long-form texts that have not been verified for correctness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.59, + 0.736, + 0.606 + ], + "angle": 0, + "content": "2.2. LFM-based Tool Agents" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Most recent LFMs have demonstrated impressive reasoning capabilities on numerous zero-shot tasks, which makes it possible to build LFM-based tool agents for real-world tasks. Mainstream approaches include in-context learning [3, 9, 22, 37, 42, 45, 46] and instruction fine-tuning [10, 28, 30, 33, 38]. However, these agents still face challenges in real-world applications. Additionally, recent advancements in web navigation agents [14, 16, 17, 23, 41, 43, 48] have primarily focused on GUI-based interactions, often designed as browser plugins, but their applicability is limited to web-based tasks. In contrast, our ToolEngine extends to a broader range of real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.712, + 0.819 + ], + "angle": 0, + "content": "2.3. Other VQA Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Visual Question Answering (VQA) is one of the most fundamental tasks in multimodal learning. Early VQA datasets [2, 5, 13, 18, 20, 34, 49] primarily focus on simple commonsense question-answering, which typically consists of tasks that humans can easily solve. Recently, some" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4425" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.143, + 0.089, + 0.858, + 0.285 + ], + "angle": 0, + "content": "
DatasetNumber of SamplesNumber of ToolsMultrimodal Inputs?Real-World S/Q Setting?Evaluable Answers?Reasoning Complexity↑
ToolBench [30]126.5K3451X- / XX1.43
ToolQA [50]1.5K13X- / ✓✓1.35
Infoseek [4]1.3M1✓✓ / ✓✓1.13
VPD [10]*310.5K4✓✓ / ✓--
M&m's [24]1.6K33✓✓ / XX1.63
GAIA [27]†0.5K-✓✓ / ✓✓2.61
GTA [39]0.2K14✓X/ ✓✓1.86
MM-Traj [8]‡20K9✓X/ ✓X1.77
ToolVQA (Ours)23.7K10✓✓ / ✓✓2.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.295, + 0.907, + 0.366 + ], + "angle": 0, + "content": "Table 1. Comparison between ToolVQA and other datasets. Real-World S/Q Setting: whether each sample in the dataset involves: (1) Scenario—utilizes real rather than synthetic multimodal inputs; (2) Query—implicitly embeds tool usage and reasoning within the query. Reasoning Complexity: the average reasoning depth of humans for each task. VPD* constructs trajectories from existing QA pairs and is not publicly accessible. GAIA†'s each sample corresponds to a distinct set of human-annotated tools. ‡MM-Traj generates over-simplified PDF files as multimodal contexts, and their answers are long-form texts that have not been verified for correctness." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.391, + 0.482, + 0.543 + ], + "angle": 0, + "content": "studies [1, 4, 10, 26, 36] begin exploring more challenging forms of VQA, which require various tools like Wikipedia, OCR, Object Detection, etc. However, these datasets typically only evaluate LFM's ability to use one specific tool instead of collaboration between multiple tools, limiting LFM's multi-step reasoning ability within a tool-agent framework. In contrast, ToolVQA contains high-quality and diverse multi-step tool usage traces. It encompasses 10 multimodal tools across 7 diverse domains, with an average inference length of 2.78 reasoning steps per sample." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.57, + 0.246, + 0.585 + ], + "angle": 0, + "content": "3. Data Collection" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.598, + 0.265, + 0.613 + ], + "angle": 0, + "content": "3.1. Task Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Tool-use VQA Task. Given the image scenario \\(\\mathcal{I}\\), the toolset \\(\\mathcal{T}\\), and the user query \\(\\mathcal{Q}\\), a tool agent is required to provide an answer to the query \\(\\mathcal{A}\\) and the tool-use trajectory \\(\\mathcal{P} = [(t_i, a_i)]_{i=1}^n\\), in which \\(t_i \\in \\mathcal{T}\\) represents the tool used in step \\(i\\) and \\(a_i\\) represents the calling arguments for \\(t_i\\). Consequently, each sample is a five-tuple \\((\\mathcal{I}, \\mathcal{T}, \\mathcal{Q}, \\mathcal{A}, \\mathcal{P})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Toolset and Image Source. We collect 10 tools with various functions as our toolset, including (1) Perception: ImageCaption, OCR, ObjectDetection, and RegionDescription. (2) Operation: DrawBox, GoogleSearch. (3) Logic: Calculator, Plot, ItemCount. (4) Creativity: TextToImage. These tools cover a wide range of real-world tasks and offer strong generalization capabilities (e.g. GoogleSearch can retrieve any external knowledge). The different sources of image \\(\\mathcal{I}\\) and the corresponding toolset \\(\\mathcal{T}\\) are shown in Fig. 2. These sources are diverse and complex, including real-world photography, e-commerce products, school exam charts, and other practical scenarios." + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.391, + 0.864, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.474, + 0.861, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.73, + 0.907, + 0.759 + ], + "angle": 0, + "content": "Figure 2. Image sources and corresponding tools of ToolVQA. We filter out overly simplistic tables and images from the data sources." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.639, + 0.803 + ], + "angle": 0, + "content": "3.2. ToolEngine" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We specify the definition of our real-world setting into the following criteria: (1) Align with human users: Each sample's scenario \\(\\mathcal{I}\\) and query \\(\\mathcal{Q}\\) should align with real-world user needs. We have filtered out uninformative scenarios \\(\\mathcal{I}\\), but it remains necessary to ensure that the tool-use queries \\(\\mathcal{Q}\\) correspond to human demands. (2) Real-world" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4426" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.098, + 0.873, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.427, + 0.908, + 0.486 + ], + "angle": 0, + "content": "Figure 3. The pipeline of ToolEngine, contains three core components: Real-world Example Construction, Image-guided DFS on Tool Graph, and LCS-based Example Matching. Given an input image, we perform DFS on the complete tool graph. At each step, an LFM controller generates the next tool's name and arguments, guided by the image, current tool-use trajectory, and matched examples. Once DFS is complete, the tool-use trajectory is determined, and then used to generate the query and answer." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.509, + 0.486, + 0.738 + ], + "angle": 0, + "content": "deployed tools: Real-world tool results are complex distributions with noise, and the model must learn how to deal with the noise from real results. Some work [27, 28] uses simulated tool outputs to reduce noise, but we believe this reduces the realism of the training environment. (3) Multi-step reasoning: The tool-use trajectory \\(\\mathcal{P}\\) should involve multiple steps. Each step in it must have an inherent logical connection, forming a complete problem-solving trajectory. It should not include ineffective steps (e.g. apply OCR to an image without text), or contain non-related tool usage in a single task (e.g. use \"culinary world\" and \"NVIDIA's stock\" sequentially, which should be split into two separate tasks). We propose ToolEngine, an automated data synthesis pipeline that ensures these conditions. Fig. 3 illustrates the three core components of ToolEngine:" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.486, + 0.904 + ], + "angle": 0, + "content": "Real-World Example Construction. To better reflect real user needs, we incorporate human prior tool-use knowledge into the LFM controller through in-context examples. Human-constructed examples provide prior knowledge about typical tool-use scenarios (e.g. images containing specific entities may require a search tool, while tables with multi-digit numbers may necessitate a calculator). This information is crucial for synthesizing real-world queries. Therefore, we ask human users to construct a small set of real-world examples \\(\\mathcal{E} = \\{(\\mathcal{I},\\mathcal{T},\\mathcal{Q},\\mathcal{A},\\mathcal{P})\\}\\). Each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.584 + ], + "angle": 0, + "content": "example in this set captures a logically valid trajectory, serving as prior knowledge for automatic construction. This approach not only aligns with how humans intuitively associate tools but also provides a structured, intuitive foundation for the construction process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.909, + 0.768 + ], + "angle": 0, + "content": "Image-guided DFS on Tool Graph. To construct a multi-step tool-use trajectory, we employ image-guided Depth-First Search (DFS) on a tool graph to construct a reasonable tool-use trajectory for each image. While LFM can directly generate reasoning sequences, unrealistic tool invocations often lead to less reliable tool-use trajectories in practice. Our DFS trajectory involves real tool calls to extract detailed information from the input image, which is then used to generate challenging question-answer pairs. Specifically, we utilize the most advanced LFM (ChatGPT-4o-latest), denoted as \\(\\mathcal{M}\\), as the controller to select the tool at each step and generate the corresponding arguments. Formally:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.772, + 0.909, + 0.788 + ], + "angle": 0, + "content": "\\[\nt _ {i} = \\mathcal {M} (\\text {c h o i c e s} = \\mathcal {T}, \\text {i m a g e} = \\mathcal {I}, \\text {e x a m p l e s} = \\operatorname {R e t} (\\mathcal {E}, \\mathcal {P} _ {i - 1}))\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.79, + 0.868, + 0.805 + ], + "angle": 0, + "content": "\\[\na _ {i} = \\mathcal {M} (t o o l = t _ {i}, i m a g e = \\mathcal {I}, e x a m p l e s = \\operatorname {R e t} (\\mathcal {E}, \\mathcal {P} _ {i}))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "LCS-based Example-Matching. We combine the information obtained from different tools into a reasonable reasoning trajectory to perform multi-step reasoning. Previous DFS-based methods typically match a fixed group of examples, which limited their ability to combine sufficient information for multi-step reasoning. As illustrated in Fig. 4, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4427" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.48, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.235, + 0.483, + 0.278 + ], + "angle": 0, + "content": "Figure 4. Comparison between different matching strategies. When a fixed example limits the diversity of generation, LCS matching can integrate multiple examples to enhance generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.484, + 0.468 + ], + "angle": 0, + "content": "propose an example-matching method based on the Longest Common Subsequence (LCS) algorithm, which enables dynamic matching during DFS. This approach better leverages the information extraction capabilities of diverse examples, facilitating more effective multi-step reasoning. Specifically, in step \\(i\\), the current solution process is denoted as \\(\\mathcal{P}_i = [(t_1,a_1),(t_2,a_2),\\dots ,(t_i,a_i)]\\) and the set of example solution processes as \\(\\mathcal{P}_e\\in \\mathcal{E}\\). At each step, we perform the LCS matching between \\(\\mathcal{P}_i\\) and each element of \\(\\mathcal{P}^e\\), retrieving the top-k examples in \\(\\mathcal{P}^e\\) with the highest matching degree. Formally:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.476, + 0.43, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e t} (\\mathcal {E}, \\mathcal {P} _ {i}) = \\operatorname {T o p K} _ {\\mathcal {P} ^ {e} \\in \\mathcal {E}} \\left\\{\\operatorname {L C S} (\\mathcal {P} ^ {e}, \\mathcal {P} _ {i}) \\right\\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.554 + ], + "angle": 0, + "content": "This approach enables us to adaptively match each image with its most relevant example, while iteratively refining the match as new information is acquired at each step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.812 + ], + "angle": 0, + "content": "Human Annotations and Analysis. After the data synthesis process, we invite 10 human annotators to filter the generated data and re-associate the highest-quality subset, resulting in 2,550 test samples. The initial synthetic data is estimated to have a \\(90.8\\%\\) accuracy in human validation on 4k randomly sampled data, which is higher than that reported in previous synthetic data methods [8, 24]. Notably, the ChatGPT-4o-latest model, which we use to generate questions, demonstrates a low accuracy rate (less than \\(40\\%\\)) in solving them. This highlights the effectiveness of our ToolEngine pipeline in decoupling complex reasoning into a series of simpler, single-step tasks. While ChatGPT-4o-latest excels in handling \\(\\mathcal{I} \\to \\mathcal{P} \\to \\mathcal{Q}, \\mathcal{A}\\) process for straightforward, step-by-step reasoning, it faces challenges with the more challenging end-to-end \\(\\mathcal{I}, \\mathcal{Q} \\to \\mathcal{P} \\to \\mathcal{A}\\) reasoning. This observation indicates a clear gap between LFM's single-step and multi-step reasoning capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.26, + 0.833 + ], + "angle": 0, + "content": "3.3. Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Tab. 2 presents the basic statistics of ToolVQA, which encompasses various question types (with both text and image as answers) and complex reasoning processes (with an average reasoning trajectory length of 2.78 steps). Notably," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "ToolVQA's questions and answers are concisely formulated (with an average token count of 15.74 for questions and 2.69 for answers). In contrast, previous synthetic datasets [8, 24] often fail to control answer length, leading to hallucinated content within the answers, which is detrimental to training. Fig. 5 demonstrates that our tool frequency aligns with real-world user needs, reinforcing the real-world setting of our dataset. Fig. 6 presents the distribution of reasoning trajectory lengths in ToolVQA." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.238, + 0.733, + 0.254 + ], + "angle": 0, + "content": "3.4. Ablation on ToolEngine" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.908, + 0.669 + ], + "angle": 0, + "content": "To verify whether ToolEngine truly improves the quality of synthetic data, we conducted an ablation study on different synthetic methods. We invite 10 users to evaluate 1k randomly sampled data based on four key metrics: (1) answer correctness (Acc.); (2) correlation (Cor.): whether each sample's image content and Q-A pair are aligned; (3) tool redundancy (Nec.): whether each tool call is necessary for obtaining the answer; and (4) reasoning complexity (R.C.): the average reasoning depth for each task (same as defined in Tab. 1). The results, presented in Tab. 3, show that Real-World Example and LCS matching significantly improved Acc., Nec., and R.C., which highlights their role in the DFS-based simulation of reasoning chains: (1) Real-World Example and LCS matching introduce prior knowledge aligned with image content, enhancing both answer accuracy and tool necessity. (2) Real-World Example and LCS matching incorporate different types of knowledge at each step, increasing the reasoning complexity. Furthermore, the absence of LCS matching has a significant impact on data quality (e.g. Acc. dropped from 90.8 to 41.6), indicating that a fixed set of examples cannot adapt to different scenarios. Dynamically matching appropriate examples is crucial for significantly enhancing the generalization ability of the synthetic pipeline. However, both methods had a limited impact on Cor., likely because the existing controller already effectively captures image content, minimizing errors caused by misleading examples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.684, + 0.646, + 0.7 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.753, + 0.725 + ], + "angle": 0, + "content": "4.1. Training and Test Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.907, + 0.793 + ], + "angle": 0, + "content": "Training Objective. In our dataset \\(\\mathcal{D}\\), for each datapoint \\(\\mathcal{E} = (\\mathcal{I},\\mathcal{T},\\mathcal{Q},\\mathcal{A},\\mathcal{P})\\) contains \\(n\\) rounds of dialogue (\\(n - 1\\) rounds for tool usage and one round for final answer), we train our tool agent using cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.803, + 0.892, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\mathcal {E} \\sim \\mathcal {D}} \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} - \\log p \\left(t _ {i}, a _ {i}, r _ {i} \\mid \\mathcal {I}, \\mathcal {T}, \\mathcal {Q}, \\mathcal {P} _ {i - 1}\\right) \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Training Setting. We divide ToolVQA into training and test sets: the training set consists of 21,105 automatically generated samples (with a human verification accu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4428" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.131, + 0.103, + 0.323, + 0.231 + ], + "angle": 0, + "content": "
StatisticsNumber
Samples23655
Tool Calls65785
Text answers15806
Image answers7849
Query length15.74
Answer length2.69
Trajectory length2.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.152, + 0.241, + 0.299, + 0.254 + ], + "angle": 0, + "content": "Table 2. Basic statistics." + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.098, + 0.577, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.255, + 0.528, + 0.27 + ], + "angle": 0, + "content": "Figure 5. Tool frequency." + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.098, + 0.875, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.247, + 0.846, + 0.262 + ], + "angle": 0, + "content": "Figure 6. Trajectory length distribution." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.293, + 0.477, + 0.373 + ], + "angle": 0, + "content": "
MethodAcc.↑Cor. ↑Nec. ↑R.C. ↑
ToolEngine90.885.287.512.38
-w/o Example + LCS27.377.621.041.1
-w/o LCS41.681.454.261.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.122, + 0.382, + 0.452, + 0.396 + ], + "angle": 0, + "content": "Table 3. Ablation study on different synthetic methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.426, + 0.483, + 0.47 + ], + "angle": 0, + "content": "racy of \\(90.8\\%\\) ), while the test set contains 2,550 humanreannotated samples. We fine-tune LLaVA-7B [21] in the Lego Agent framework [15]." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.473, + 0.484, + 0.805 + ], + "angle": 0, + "content": "Test Setting. We apply two test modes: end-to-end and step-by-step modes, to comprehensively evaluate LFMs' performance under different settings. The end-to-end mode only focuses on the end-to-end VQA ability of LFMs. Formally, we evaluate the \\(\\mathcal{I} + \\mathcal{Q}\\rightarrow \\mathcal{A}\\) process. The step-by-step mode focuses on evaluating LM's tool-usage capabilities, formally the \\(\\mathcal{I} + \\mathcal{Q} + \\mathcal{P}_i\\to \\mathcal{P}_{i + 1}\\) process. We use four metrics [39]: (1) InstAcc (Inst.): the percentage of steps executed without errors; (2) ToolAcc (Tool.): the accuracy of tool selection; (3) ArgAcc (Arg.): the accuracy of tool's argument name prediction, and (4) SummAcc (Summ.): the accuracy of final answer's summary. Previous work has focused on the research of large language model (LLM)-based tool agents while ignoring vision-language model (VLM)-based tool agents and the end-to-end VQA abilities of VLMs. To bridge these gaps, our test involves three settings: (1) VLM, directly answering questions end-to-end, formally \\(\\mathcal{Q} + \\mathcal{I}\\rightarrow \\mathcal{A}\\), (2) VLM+tool, calling the tool through VLM, formally \\(\\mathcal{Q} + \\mathcal{I} + \\mathcal{T}\\rightarrow \\mathcal{P} + \\mathcal{A}\\), (3) LLM+tool, calling the tool through LLM, formally \\(\\mathcal{Q} + \\mathcal{T}\\rightarrow \\mathcal{P} + \\mathcal{A}\\) (Note that image information can still be obtained through tools such as caption under this setting)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.231, + 0.833 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Scale. As shown in Tab. 4, the tool-use capabilities of LFMs generally improve with the increase in model parameters. For example, Qwen2-7B-instruct surpasses Qwen2-2B-instruct by \\(7.03\\%\\), and closed-sourced LM (larger than" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.905, + 0.371 + ], + "angle": 0, + "content": "100B) is generally better than open-sourced (2B,7B). After fine-tuning, our 7B model's performance (last 3 lines) is close to the large scale LFM GPT-3.5-Turbo. This demonstrates that fine-tuning significantly enhances the model's real-world tool-use capabilities, even in smaller models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.907, + 0.644 + ], + "angle": 0, + "content": "Settings. By comparing the performance of the same model under different settings (VLM, VLM+tool, and LLM+tool) in Tab. 4, we identify several interesting observations. (1) In models such as GPT-4o, Qwen2-7B/2B, and LLaVA-7B, the VLM+tool setting performs worse than the LLM+tool setting, suggesting that the visual modules in current LFM architectures are still ineffective at guiding tool usage. (2) For models including GPT-4o, Claude-3.5, LLaVA-7B, and Qwen2-2B, we find that VLM+tool underperforms VLM alone. This indicates that the noise introduced by tool usage can outweigh the benefits brought by the tools. However, the tuned LLaVA model demonstrates a clear progression: VLM \\(<\\) LLM+tool \\(<\\) VLM+tool, suggesting that fine-tuning can indeed enhance the model's ability to make effective use of tools while suppressing tool-induced noise. Nevertheless, the similar performance of the LLM+tool and VLM+tool settings implies that the visual module's potential in facilitating tool use has not yet been fully explored." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Bottlenecks. The last two rows of Tab. 4 indicate that the performance of fine-tuned models is often constrained by their weakest capabilities, which are typically argument prediction and answer summarization. While fine-tuning significantly improves instruction formatting and tool selection, it makes little progress in weaker capabilities. Specifically, instruction formatting and tool selection can be accomplished through memorizing patterns, as the number of reasonable tool invocation types in real-world scenarios is limited. However, argument prediction and answer summarization require the model to genuinely understand the new information returned by tools and extract meaningful responses from it. This information varies across different scenarios and queries, and even within multi-turn dialogues of the same scenario. Consequently, a high level of generalization capability in vision-language joint reasoning is essential to handle these issues in tool-use tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4429" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.153, + 0.089, + 0.847, + 0.499 + ], + "angle": 0, + "content": "
ModelSettingEnd-to-End ModeStep-by-Step Mode
Acc. (%)Tool Call (Times)Tool Error (Times)Inst. (%)Tool. (%)Arg. (%)Summ. (%)
Closed-source LFMs
ChatGPT-4o-latestVLM38.29------
VLM + tool34.962003180436.514.688.9256.1
LLM + tool37.624687196273.0241.1225.2765.45
Claude-3-5-sonnetVLM30.33------
VLM + tool30.5258552445.7116.947.3646.06
LLM + tool24.48183467266.8339.2913.9875.52
GPT-3.5-TurboLLM + tool18.373178132173.2430.4620.0858.18
Open-source LFMs
Qwen2-VL-7B-InstructVLM3.64------
VLM + tool5.861373105135.129.021.0938.18
LLM + tool6.781366108939.4611.23.5544.85
Qwen2-7B-InstructLLM + tool11.53177370363.124.458.269.7
LLaVA-v1.5-7BVLM8.57------
VLM + tool1.179684968416.399.4300.01
LLM + tool1.6610421104219.2515.03011.52
LLaMA-3-8b-instructLLM + tool4.936607654251.6314.130.5341.99
Qwen2-VL-2B-InstructVLM7.71------
VLM + tool2.14067391547.1612.70.4121.82
LLM + tool2.226892665536.3410.520.6832.12
Qwen2-1.5B-InstructLLM + tool4.57933614150.9521.454.5144.24
Tuned LLaVA-7B (Ours)VLM7.21------
VLM + tool18.8431161786.6261.6139.3430.91
LLM + tool18.43446268087.5162.340.5726.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.381, + 0.508, + 0.615, + 0.523 + ], + "angle": 0, + "content": "Table 4. Results on ToolVQA's test set." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.549, + 0.379, + 0.564 + ], + "angle": 0, + "content": "4.3. Out-of-Distribution Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.801 + ], + "angle": 0, + "content": "To test the generalization capability of our tuned model, we conduct experiments on out-of-distribution (OOD) benchmarks including (1) Seen tools, unseen queries: TextVQA [36], TallyQA [1], InfoSeek [4]; and (2) Unseen tools: GTA [39], TEMPLAMA [6]. The results are shown in the Tab. 5. Our fine-tuned model significantly outperforms the baseline model LLaVA-7B by \\(5.8\\%\\), \\(4.2\\%\\), \\(8.6\\%\\), \\(21.17\\%\\), and \\(18.37\\%\\) accuracy on these datasets, showing excellent generalizability. Its performance also surpasses the large closed-source model GPT-3.5-Turbo on unseen queries. However, its performance on TEMPLAMA is inferior to GPT-3.5-Turbo, possibly because GPT encountered similar data during its training (due to its great coverage of extensive training data), whereas our model had no exposure to such data at all." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.817, + 0.356, + 0.833 + ], + "angle": 0, + "content": "4.4. Few-shot In-context Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Few-shot in-context learning (ICL) has become a common and important technique in tool-use tasks. However, prior model fine-tuning work often overlooks evaluating the ICL capabilities of fine-tuned models. We believe that fine-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.747 + ], + "angle": 0, + "content": "tuning and ICL can be complementary, offering significant potential for improvement. Tab. 6 shows few-shot ICL performance on ToolVQA. GPT-4o steadily improves with more shots, while GPT-3.5 and LLaVA both show limited gains and even decline slightly at 10-shot. This may be due to long contexts impairing the LLM's ability to interpret precise tool-use instructions. Fine-tuned LLaVA outperforms the baseline and continues to benefit from ICL. These results suggest that (1) compared to fine-tuning, ICL's improvement is limited and affected by context lengths, and (2) fine-tuned models can still benefit from ICL, highlighting the complementary value of these two approaches in enhancing tool-use performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.667, + 0.774 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To further evaluate the quality of the fine-tuning data from multiple perspectives, we conduct an ablation study of the different training strategies. As shown in Tab. 7, removing the ImageCaption tool significantly degrades both accuracy and parameter correctness. This highlights the model's reliance on high-quality captions for initial scene understanding, likely due to its limited inherent visual capability for fine-grained information extraction. When all tools" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4430" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.162, + 0.089, + 0.837, + 0.168 + ], + "angle": 0, + "content": "
ModelTextVQA [36]TallyQA [1]InfoSeek [4]GTA [39]TEMPLAMA [6]
GPT-3.5-Turbo36.36111.323.6233.67
LLaVA-7B41.260.15.212.123.06
Tuned LLaVA-7B4764.313.833.2921.43
" + }, + { + "type": "table_caption", + "bbox": [ + 0.382, + 0.177, + 0.615, + 0.19 + ], + "angle": 0, + "content": "Table 5. Results on OOD benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.205, + 0.486, + 0.293 + ], + "angle": 0, + "content": "
ModelZero-shot1-shot5-shot10-shot
GPT-4o34.9637.2038.4138.63
GPT-3.5-Turbo18.3719.8320.4019.04
LLaVA-7B1.173.454.273.67
Tuned LLaVA-7B18.8019.4121.1320.69
" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.205, + 0.905, + 0.293 + ], + "angle": 0, + "content": "
ModelSettingAcc.Tool CallTool Error
Tuned LLaVA-7BVLM + tool18.84311617
-w/o CaptionVLM + tool11.11861192
-w/o PVLM8.01--
-w/o P + ILLM0.18--
" + }, + { + "type": "table_caption", + "bbox": [ + 0.152, + 0.301, + 0.408, + 0.315 + ], + "angle": 0, + "content": "Table 6. Few-shot performance on ToolVQA." + }, + { + "type": "table_caption", + "bbox": [ + 0.583, + 0.301, + 0.818, + 0.314 + ], + "angle": 0, + "content": "Table 7. Ablation study on training methods." + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.332, + 0.861, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.544, + 0.707, + 0.558 + ], + "angle": 0, + "content": "Figure 7. Visualization of the main error types made by tuned LLaVA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.676 + ], + "angle": 0, + "content": "are removed and ToolVQA is reduced to a single-hop VQA model, performance drops substantially, underscoring the necessity of multi-step tool usage. In the absence of both tools and images, performance approaches zero, indicating that the model is heavily dependent on visual input when answering questions in ToolVQA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.242, + 0.699 + ], + "angle": 0, + "content": "4.6. Error Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In Fig. 7, we randomly select 100 errors made by tuned LLaVA, and manually count and classify the main error types. The two main error types are: (1) Argument prediction error, for example, missing essential keywords \"age\" during a search. (2) Answer summary error, for example, extracting incorrect information \"1,300\" from tool outputs. The common cause of these errors is the model's failure to recognize key terms (e.g. \"age\" and \"pound\") in the tool outputs, even though this should not be a challenging task for a language model. This suggests that the model still struggles with dynamically adapting to additional information provided by tools. Since multi-step reasoning often suffers from error accumulation (early errors propagate to sub" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.586, + 0.907, + 0.661 + ], + "angle": 0, + "content": "sequent steps), errors in processing tool results at any step will affect the results. Fine-tuning has limited improvement on the capability for dynamically processing tool results, but ToolVQA still provides a valuable benchmark for evaluating this capability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.634, + 0.694 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We present ToolVQA, a large multimodal dataset for real-world tool-augmented reasoning. Using our ToolEngine pipeline, we generate 23K real-world samples with implicit multi-step reasoning over 10 tools and 7 domains. This dataset challenges existing models beyond synthetic settings and better reflects real user needs. Experiments show that LLaVA-7B fine-tuned on ToolVQA not only excels on the in-domain test set but also outperforms GPT-3.5-turbo on five OOD benchmarks, establishing a new state-of-the-art for open-source models in complex tool-use VQA. ToolVQA serves as both a benchmark and a training ground for developing more capable, generalizable tool-using agents." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4431" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the grants from the National Natural Science Foundation of China 62372014, Beijing Nova Program and Beijing Natural Science Foundation 4252040." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.169, + 0.188, + 0.185 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.194, + 0.484, + 0.235 + ], + "angle": 0, + "content": "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions, 2018. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.239, + 0.484, + 0.28 + ], + "angle": 0, + "content": "[2] Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C. Lawrence Zitnick, Dhruv Batra, and Devi Parikh. Vqa: Visual question answering, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.282, + 0.483, + 0.323 + ], + "angle": 0, + "content": "[3] Lluis Castrejon, Thomas Mensink, Howard Zhou, Vittorio Ferrari, Andre Araujo, and Jasper Uijlings. Hammr: Hierarchical multimodal react agents for generic vqa, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.326, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[4] Yang Chen, Hexiang Hu, Yi Luan, Haitian Sun, Soravit Changpinyo, Alan Ritter, and Ming-Wei Chang. Can pre-trained vision and language models answer visual information-seeking questions?, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.384, + 0.484, + 0.439 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[6] Bhuwan Dhingra, Jeremy R. Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W. Cohen. Time-aware language models as temporal knowledge bases. Transactions of the Association for Computational Linguistics, 10:257-273, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.568 + ], + "angle": 0, + "content": "[7] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.571, + 0.484, + 0.626 + ], + "angle": 0, + "content": "[8] Zhi Gao, Bofei Zhang, Pengxiang Li, Xiaojian Ma, Tao Yuan, Yue Fan, Yuwei Wu, Yunde Jia, Song-Chun Zhu, and Qing Li. Multi-modal agent tuning: Building a vlm-driven agent for efficient tool usage, 2025. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.629, + 0.484, + 0.67 + ], + "angle": 0, + "content": "[9] Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. Toolkengpt: Augmenting frozen language models with massive tools via tool embeddings, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[10] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.744, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[11] Ziniu Hu, Ahmet Iscen, Chen Sun, Kai-Wei Chang, Yizhou Sun, David A Ross, Cordelia Schmid, and Alireza Fathi. Avis: Autonomous visual information seeking with large language model agent, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Yue Huang, Jiawen Shi, Yuan Li, Chenrui Fan, Siyuan Wu, Qihui Zhang, Yixin Liu, Pan Zhou, Yao Wan, Neil Zhenqiang Gong, and Lichao Sun. Metatool benchmark for large language models: Deciding whether to use tools and which to use, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.194, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[14] Faria Huq, Zora Zhiruo Wang, Frank F. Xu, Tianyue Ou, Shuyan Zhou, Jeffrey P. Bigham, and Graham Neubig. Cow-pilot: A framework for autonomous and human-agent collaborative web navigation, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[15] InternLM. Agentlego: Enhance llm agents with rich tool apis. https://github.com/InternLM/agentlego, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.237, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[16] Iat Long Iong, Xiao Liu, Yuxuan Chen, Hanyu Lai, Shuntian Yao, Pengbo Shen, Hao Yu, Yuxiao Dong, and Jie Tang. OpenWebAgent: An open toolkit to enable web agents on large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 72-81, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[17] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[19] Minghao Li, Feifan Song, Yu Bowen, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. Api-bank: A comprehensive benchmark for tool-augmented llms. In Conference on Empirical Methods in Natural Language Processing, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.904, + 0.687 + ], + "angle": 0, + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[22] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In The 37th Conference on Neural Information Processing Systems (NeurIPS), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[23] Kaixin Ma, Hongming Zhang, Hongwei Wang, Xiaoman Pan, Wenhao Yu, and Dong Yu. Laser: LIm agent with statespace exploration for web navigation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[24] Zixian Ma, Weikai Huang, Jieyu Zhang, Tanmay Gupta, and Ranjay Krishna. m&m's: A benchmark to evaluate tool-use for multi-step multi-modal tasks. EECV 2024, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[25] Arjun Mani, Nobline Yoo, Will Hinthorn, and Olga Russakovsky. Point and ask: Incorporating pointing into visual question answering, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4432" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[26] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[27] Grégoire Mialon, Clémentine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, and Thomas Scialom. Gaia: a benchmark for general ai assistants, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.221 + ], + "angle": 0, + "content": "[28] Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.291 + ], + "angle": 0, + "content": "[29] Yun Peng, Shuqing Li, Wenwei Gu, Yichen Li, Wenxuan Wang, Cuiyun Gao, and Michael R. Lyu. Revisiting, benchmarking and exploring api recommendation: How far are we? IEEE Transactions on Software Engineering, 49:1876-1897, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[30] Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. Toollm: Facilitating large language models to master 16000+ real-world apis, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[31] Yujia Qin, Shengding Hu, Yankai Lin, Weize Chen, Ning Ding, Ganqu Cui, Zheni Zeng, Yufei Huang, Chaojun Xiao, Chi Han, Yi Ren Fung, Yusheng Su, Huadong Wang, Cheng Qian, Runchu Tian, Kunlun Zhu, Shihao Liang, Xingyu Shen, Bokai Xu, Zhen Zhang, Yining Ye, Bowen Li, Ziwei Tang, Jing Yi, Yuzhang Zhu, Zhenning Dai, Lan Yan, Xin Cong, Yaxi Lu, Weilin Zhao, Yuxiang Huang, Junxi Yan, Xu Han, Xian Sun, Dahai Li, Jason Phang, Cheng Yang, Tongshuang Wu, Heng Ji, Zhiyuan Liu, and Maosong Sun. Tool learning with foundation models, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[32] Yangjun Ruan, Honghua Dong, Andrew Wang, Silviu Pitis, Yongchao Zhou, Jimmy Ba, Yann Dubois, Chris J. Maddison, and Tatsunori Hashimoto. Identifying the risks of lm agents with an lm-emulated sandbox, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[33] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[34] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[35] Yongliang Shen, Kaitao Song, Xu Tan, Wenqi Zhang, Kan Ren, Siyu Yuan, Weiming Lu, Dongsheng Li, and Yueting Zhuang. Taskbench: Benchmarking large language models for task automation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[36] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read, 2019. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[37] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11888-11898, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Chenyu Wang, Weixin Luo, Qianyu Chen, Haonan Mai, Jindi Guo, Sixun Dong, Xiaohua (Michael) Xuan, Zhengxin Li, Lin Ma, and Shenghua Gao. Mllm-tool: A multimodal" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "large language model for tool agent learning. arXiv preprint arXiv:2401.10727, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[39] Jize Wang, Zerun Ma, Yining Li, Songyang Zhang, Cailian Chen, Kai Chen, and Xinyi Le. Gta: A benchmark for general tool agents, 2024. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[40] Yan Wang, Yawen Zeng, Jingsheng Zheng, Xiaofen Xing, Jin Xu, and Xiangmin Xu. Videocot: A video chain-of-thought dataset with active annotation tool, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[41] Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. Agent workflow memory, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[42] Shirley Wu, Shiyu Zhao, Qian Huang, Kexin Huang, Michihiro Yasunaga, Kaidi Cao, Vassilis N. Ioannidis, Karthik Subbian, Jure Leskovec, and James Zou. Avatar: Optimizing llm agents for tool usage via contrastive reasoning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.376 + ], + "angle": 0, + "content": "[43] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[44] Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[45] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[46] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[47] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[48] Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. Webarena: A realistic web environment for building autonomous agents, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[49] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[50] Yuchen Zhuang, Yue Yu, Kuan Wang, Haotian Sun, and Chao Zhang. Toolqa: A dataset for llm question answering with external tools, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4433" + } + ] +] \ No newline at end of file diff --git a/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/dd0c7eba-b554-4e21-9d01-9fb4080c4a97_origin.pdf b/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/dd0c7eba-b554-4e21-9d01-9fb4080c4a97_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d503342d7cd2f19773b6def88e6057d74980b727 --- /dev/null +++ b/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/dd0c7eba-b554-4e21-9d01-9fb4080c4a97_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:917d6ca72806c14afe7cbe1681c2b39f4dc081b2e3d3fd17f836bbb3d1035167 +size 1361791 diff --git a/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/full.md b/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e6ddd187cb8670319ec038caabcf3f3cfd07cfe8 --- /dev/null +++ b/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/full.md @@ -0,0 +1,315 @@ +# ToolVQA: A Dataset for Multi-step Reasoning VQA with External Tools + +Shaofeng Yin Ting Lei Yang Liu* Wangxuan Institute of Computer Technology, Peking University + +yin_shaofeng@stu.pku.edu.cn {tinglei, yangliu}@pku.edu.cn + +# Abstract + +Integrating external tools into Large Foundation Models (LFMs) has emerged as a promising approach to enhance their problem-solving capabilities. While existing studies have demonstrated strong performance in tool-augmented Visual Question Answering (VQA), recent benchmarks reveal significant gaps in real-world tool-use proficiency, particularly in functionally diverse multimodal settings requiring multi-step reasoning. In this work, we introduce ToolVQA, a large-scale multimodal dataset comprising 23K samples, designed to bridge this gap. Unlike previous datasets that rely on synthetic scenarios and simplified queries, ToolVQA features real-world visual contexts and challenging implicit multi-step reasoning tasks, better aligning with real user interactions. To construct this dataset, we propose ToolEngine, a novel data generation pipeline that employs image-guided Depth-First Search (DFS) with a Longest Common Subsequence (LCS)-based example matching mechanism to simulate human-like tool-use reasoning. ToolVQA encompasses 10 multimodal tools across 7 diverse domains, with an average inference length of 2.78 reasoning steps per sample. The LLaVA-7B model fine-tuned on ToolVQA not only achieves impressive performance on the ToolVQA test set, but also surpasses the large closed-source model GPT-3.5-turbo on five out-of-distribution (OOD) datasets, showing strong generalizability in real-world tool-use scenarios. Code is available at https://github.com/Fugtemypt123/ToolVQA-release. + +# 1. Introduction + +Integrating external tools into Large Foundation Models (LFMs) [31] is a promising way to build versatile AI assistants. By leveraging external tools, LFMs can break down complex tasks into smaller, more manageable subtasks, each handled by a specialized tool. Prior works [7, 10, 11, 33, 37] show strong performance on traditional Visual Question Answering (VQA) benchmarks using extern + +# Non Real-World Scenarios + +![](images/5fb20371181387a9671fca03aff1c85993942266b1e06d4fedd8f184bc448f26.jpg) + +$$ +2 x ^ {2} - 1 2 = 0 +$$ + +![](images/83aea2f60d18d44e2e35805ed5a13d2ee244cbdd04f0df59ee00139b8fb1732f.jpg) + +From:Shanghai To:Beijing + +Speed: + +250 km/h + +(Overly simplistic synthetic scenarios) + +# Non Real-World Queries + +Based on the text 'James Hutton is often viewed as the first modern geologist...', could you help me find out who the first modern geologist is? + +(Simple retrieve, no tool use) + +Can you help me find popular videos and trending gaming videos. I'm using the 'Cheap YouTube API' tool. (Explicit tool use, no reasoning) + +Find me some interesting news articles about the culinary world. Additionally, provide me with the current threshold securities list for NVIDIA's stock. (Just tool list, no multi-step reasoning) + +# Real-World Setting + +![](images/f0c4732c18d2011d9a0fb34ecdf24e773a6c943030d6e62c55ffa727c2ac4881.jpg) + +![](images/c213c1444745538399e00d78ad01501a71c5f019d5a7b06ce0966362d12f5bb2.jpg) + +![](images/c43bf96797156d8322e7dff3ed906168cf2d888a04bdfe0c64299600a2235337.jpg) + +a bowl of salad, a sandwich, a sandwich of + +and a bottle of beer + +![](images/0a2d1d008694823f1b0d6450fb3296c6e4e858cedea6b1629a3d60ffbbafcb30.jpg) + +![](images/eef1f2be73b208412e8577ed42655b778145cceb77262b0deb811d3683b2b583.jpg) + +![](images/0a22fbf87ed9ca6623669e86e370614171a9a99632620fc62893dceecd6c72c1.jpg) + +![](images/13a6a9027e41233b41a46d31e33001dcbabf30f21836ca84fa66f0b0f782336a.jpg) + +Bridgeport Rewiring Company closed in 2019. + +2024-2019=5 + +Q: How many years has it been since the brewery that produced this beer ceased operations? A: 5 + +![](images/f0c1c23afc4d881dc16d83c41b674c15e69a29880981ad7c205abe89504c7580.jpg) + +![](images/8717f26408b739818d5b3a8e9179a4ee1ad09a8e432378113083908b9c2eec85.jpg) +Figure 1. Our real-world setting includes (1) complex visual scenarios with real-world context; (2) challenging queries with an implicit multi-step reasoning process. Existing datasets (left) do not meet these requirements, while our ToolVQA (right) does. + +nal tools. Recent works [24, 27, 39] highlight the limitations of current LFMs in their tool-use capabilities within challenging contexts. These contexts require functionally diverse multimodal tools and multi-step reasoning processes associated with tool usage, posing major challenges to LFMs' tool-use proficiency. + +One effective approach to enhancing LFM's tool-use capability is fine-tuning on large-scale datasets [8, 10, 12, 33]. However, there remains a significant gap between existing large-scale datasets and real-world user needs. Through the study of real user logs, we find that real-world tool usage involves (1) real-world scenarios: complex visual scenarios with real-world context (e.g. real-taken photos rather than synthetic images), and (2) real-world queries: challenging queries with an implicit multi-step reasoning process. However, existing datasets often lack these two criti + +cal properties, as shown in Fig. 1. The scenarios are synthesized instead of real-taken [8, 39], resulting in simplified scenes that differ from real-world scenarios in complexity. The queries require only single-step reasoning or explicitly provide hints about the reasoning process [10, 30], such as "using the Cheap YouTube API tool". These setups simplify the task and may lead to a gap between synthetic and real-world tool usage, which limits our ability to analyze LFMs' performance on this task. Additionally, some datasets [27, 39] rely on costly human annotations, making them difficult to scale, which limits fine-tuning to enhance LFMs' tool-using capabilities. + +To bridge these gaps, we introduce ToolEngine, a data construction pipeline designed to generate multi-step reasoning VQA data for tool usage. ToolEngine employs an image-guided Depth-First Search (DFS) approach [30] to simulate the human-like multi-step reasoning process under the guidance of real user context examples. To facilitate multi-step reasoning in data, we introduce a Longest Common Subsequence (LCS)-based example matching mechanism. During each step in the iterative search process, this mechanism matches different real-world tool-use examples based on the previous search trajectory, guiding the controller in making the next selection. This mechanism enables: (1) Extracting rich visual information from the input image; (2) Integrating this information during the reasoning process to propose challenging queries. By freeing query generation from the constraints of fixed templates [8, 24, 30], ToolEngine enhances the utilization of visual information, consequently increasing the reasoning complexity of the generated queries. + +We leverage ToolEngine to construct ToolVQA, a large-scale multimodal tool-usage dataset comprising 23K samples. We assess data quality via a manual evaluation conducted on a randomly sampled 4k subset of the training set, which yields an accuracy of $90.8\%$ . The test set is carefully filtered and fully human-annotated. As shown in Tab. 1, ToolVQA outperforms prior tool-use datasets across multiple perspectives, providing a more comprehensive representation of real-world scenarios. We evaluate state-of-the-art (SOTA) LFMs on ToolVQA and observe that: (1) Larger LFMs demonstrate superior reasoning capabilities. (2) LFMs consistently underperform relative to humans, primarily due to their inability to integrate new information introduced in multi-turn dialogues. We fine-tune LLaVA-7B [21] on ToolVQA, and the model demonstrated performance surpassing or approaching that of the large closed-source model GPT-3.5-Turbo on both our test set and five out-of-distribution (OOD) benchmarks—TextVQA [36], TallyQA [1], Infoseek [4], TEMPLAMA [6], and GTA [39], which include unseen tasks and tools. In summary, our contributions are as follows: + +1. ToolEngine: We introduce ToolEngine, a VQA data + +construction engine that generates reasonable multi-step tool-use trajectories from unannotated images, and constructs challenging question-answer pairs accordingly. + +2. ToolVQA: We propose ToolVQA, a large-scale multimodal tool-use dataset that encompasses 10 multimodal tools across 7 diverse domains, providing a benchmark for evaluating and advancing LFMs' tool-use capabilities on real-world tasks. +3. Fine-tuning: We fine-tune LLaVA-7B on ToolVQA to obtain a new tool-use agent, which outperforms large closed-source models like GPT-3.5-Turbo on six benchmarks (ToolVQA and the other five OOD benchmarks). + +# 2. Related Work + +# 2.1. Datasets and Benchmarks for Tool Agent + +With the rise of real-world tool-use applications, many works have attempted to build datasets and benchmarks to evaluate the tool-use capabilities of LFMs from different perspectives. Text-based datasets [12, 19, 29, 30, 32, 44, 47, 50] collect a large number of real-world APIs to build large-scale, diverse text-based tool-use examples. On the other hand, multimodal datasets [1, 4, 13, 24, 25, 27, 35, 36, 39, 40] integrate visual input into the query process, making queries challenging and more adaptable to real-world needs. However, they either fail to capture real-world scenarios or rely on expensive human annotation. MM-Traj [8] is a recent work which also constructs large-scale training data using an automated pipeline, but has the following limitations: (1) Synthesizes over-simplified PDF files as the multimodal contexts; (2) Their answers are long-form texts that have not been verified for correctness. + +# 2.2. LFM-based Tool Agents + +Most recent LFMs have demonstrated impressive reasoning capabilities on numerous zero-shot tasks, which makes it possible to build LFM-based tool agents for real-world tasks. Mainstream approaches include in-context learning [3, 9, 22, 37, 42, 45, 46] and instruction fine-tuning [10, 28, 30, 33, 38]. However, these agents still face challenges in real-world applications. Additionally, recent advancements in web navigation agents [14, 16, 17, 23, 41, 43, 48] have primarily focused on GUI-based interactions, often designed as browser plugins, but their applicability is limited to web-based tasks. In contrast, our ToolEngine extends to a broader range of real-world scenarios. + +# 2.3. Other VQA Datasets + +Visual Question Answering (VQA) is one of the most fundamental tasks in multimodal learning. Early VQA datasets [2, 5, 13, 18, 20, 34, 49] primarily focus on simple commonsense question-answering, which typically consists of tasks that humans can easily solve. Recently, some + +
DatasetNumber of SamplesNumber of ToolsMultrimodal Inputs?Real-World S/Q Setting?Evaluable Answers?Reasoning Complexity↑
ToolBench [30]126.5K3451X- / XX1.43
ToolQA [50]1.5K13X- / ✓✓1.35
Infoseek [4]1.3M1✓✓ / ✓✓1.13
VPD [10]*310.5K4✓✓ / ✓--
M&m's [24]1.6K33✓✓ / XX1.63
GAIA [27]†0.5K-✓✓ / ✓✓2.61
GTA [39]0.2K14✓X/ ✓✓1.86
MM-Traj [8]‡20K9✓X/ ✓X1.77
ToolVQA (Ours)23.7K10✓✓ / ✓✓2.38
+ +Table 1. Comparison between ToolVQA and other datasets. Real-World S/Q Setting: whether each sample in the dataset involves: (1) Scenario—utilizes real rather than synthetic multimodal inputs; (2) Query—implicitly embeds tool usage and reasoning within the query. Reasoning Complexity: the average reasoning depth of humans for each task. VPD* constructs trajectories from existing QA pairs and is not publicly accessible. GAIA†'s each sample corresponds to a distinct set of human-annotated tools. ‡MM-Traj generates over-simplified PDF files as multimodal contexts, and their answers are long-form texts that have not been verified for correctness. + +studies [1, 4, 10, 26, 36] begin exploring more challenging forms of VQA, which require various tools like Wikipedia, OCR, Object Detection, etc. However, these datasets typically only evaluate LFM's ability to use one specific tool instead of collaboration between multiple tools, limiting LFM's multi-step reasoning ability within a tool-agent framework. In contrast, ToolVQA contains high-quality and diverse multi-step tool usage traces. It encompasses 10 multimodal tools across 7 diverse domains, with an average inference length of 2.78 reasoning steps per sample. + +# 3. Data Collection + +# 3.1. Task Formulation + +Tool-use VQA Task. Given the image scenario $\mathcal{I}$ , the toolset $\mathcal{T}$ , and the user query $\mathcal{Q}$ , a tool agent is required to provide an answer to the query $\mathcal{A}$ and the tool-use trajectory $\mathcal{P} = [(t_i, a_i)]_{i=1}^n$ , in which $t_i \in \mathcal{T}$ represents the tool used in step $i$ and $a_i$ represents the calling arguments for $t_i$ . Consequently, each sample is a five-tuple $(\mathcal{I}, \mathcal{T}, \mathcal{Q}, \mathcal{A}, \mathcal{P})$ . + +Toolset and Image Source. We collect 10 tools with various functions as our toolset, including (1) Perception: ImageCaption, OCR, ObjectDetection, and RegionDescription. (2) Operation: DrawBox, GoogleSearch. (3) Logic: Calculator, Plot, ItemCount. (4) Creativity: TextToImage. These tools cover a wide range of real-world tasks and offer strong generalization capabilities (e.g. GoogleSearch can retrieve any external knowledge). The different sources of image $\mathcal{I}$ and the corresponding toolset $\mathcal{T}$ are shown in Fig. 2. These sources are diverse and complex, including real-world photography, e-commerce products, school exam charts, and other practical scenarios. + +![](images/32e94094f293334f51ac2db55851378ee6b650d5dfeb49d8afaa8c352f42a3dc.jpg) + +![](images/3ad9bd00f9213a45a40c701e168215cfe855292ecb258b1507e8ecd29af2e54d.jpg) +Figure 2. Image sources and corresponding tools of ToolVQA. We filter out overly simplistic tables and images from the data sources. + +# 3.2. ToolEngine + +We specify the definition of our real-world setting into the following criteria: (1) Align with human users: Each sample's scenario $\mathcal{I}$ and query $\mathcal{Q}$ should align with real-world user needs. We have filtered out uninformative scenarios $\mathcal{I}$ , but it remains necessary to ensure that the tool-use queries $\mathcal{Q}$ correspond to human demands. (2) Real-world + +![](images/ea19b08f67518f02dbd735eb50cde54c647096a479758998eac6756a23c6e66e.jpg) +Figure 3. The pipeline of ToolEngine, contains three core components: Real-world Example Construction, Image-guided DFS on Tool Graph, and LCS-based Example Matching. Given an input image, we perform DFS on the complete tool graph. At each step, an LFM controller generates the next tool's name and arguments, guided by the image, current tool-use trajectory, and matched examples. Once DFS is complete, the tool-use trajectory is determined, and then used to generate the query and answer. + +deployed tools: Real-world tool results are complex distributions with noise, and the model must learn how to deal with the noise from real results. Some work [27, 28] uses simulated tool outputs to reduce noise, but we believe this reduces the realism of the training environment. (3) Multi-step reasoning: The tool-use trajectory $\mathcal{P}$ should involve multiple steps. Each step in it must have an inherent logical connection, forming a complete problem-solving trajectory. It should not include ineffective steps (e.g. apply OCR to an image without text), or contain non-related tool usage in a single task (e.g. use "culinary world" and "NVIDIA's stock" sequentially, which should be split into two separate tasks). We propose ToolEngine, an automated data synthesis pipeline that ensures these conditions. Fig. 3 illustrates the three core components of ToolEngine: + +Real-World Example Construction. To better reflect real user needs, we incorporate human prior tool-use knowledge into the LFM controller through in-context examples. Human-constructed examples provide prior knowledge about typical tool-use scenarios (e.g. images containing specific entities may require a search tool, while tables with multi-digit numbers may necessitate a calculator). This information is crucial for synthesizing real-world queries. Therefore, we ask human users to construct a small set of real-world examples $\mathcal{E} = \{(\mathcal{I},\mathcal{T},\mathcal{Q},\mathcal{A},\mathcal{P})\}$ . Each + +example in this set captures a logically valid trajectory, serving as prior knowledge for automatic construction. This approach not only aligns with how humans intuitively associate tools but also provides a structured, intuitive foundation for the construction process. + +Image-guided DFS on Tool Graph. To construct a multi-step tool-use trajectory, we employ image-guided Depth-First Search (DFS) on a tool graph to construct a reasonable tool-use trajectory for each image. While LFM can directly generate reasoning sequences, unrealistic tool invocations often lead to less reliable tool-use trajectories in practice. Our DFS trajectory involves real tool calls to extract detailed information from the input image, which is then used to generate challenging question-answer pairs. Specifically, we utilize the most advanced LFM (ChatGPT-4o-latest), denoted as $\mathcal{M}$ , as the controller to select the tool at each step and generate the corresponding arguments. Formally: + +$$ +t _ {i} = \mathcal {M} (\text {c h o i c e s} = \mathcal {T}, \text {i m a g e} = \mathcal {I}, \text {e x a m p l e s} = \operatorname {R e t} (\mathcal {E}, \mathcal {P} _ {i - 1})) +$$ + +$$ +a _ {i} = \mathcal {M} (t o o l = t _ {i}, i m a g e = \mathcal {I}, e x a m p l e s = \operatorname {R e t} (\mathcal {E}, \mathcal {P} _ {i})) +$$ + +LCS-based Example-Matching. We combine the information obtained from different tools into a reasonable reasoning trajectory to perform multi-step reasoning. Previous DFS-based methods typically match a fixed group of examples, which limited their ability to combine sufficient information for multi-step reasoning. As illustrated in Fig. 4, we + +![](images/da01cb65839426249d40d19d6191ec62278e3a9a1b4188a2d0366cd5fabc1748.jpg) +Figure 4. Comparison between different matching strategies. When a fixed example limits the diversity of generation, LCS matching can integrate multiple examples to enhance generation. + +propose an example-matching method based on the Longest Common Subsequence (LCS) algorithm, which enables dynamic matching during DFS. This approach better leverages the information extraction capabilities of diverse examples, facilitating more effective multi-step reasoning. Specifically, in step $i$ , the current solution process is denoted as $\mathcal{P}_i = [(t_1,a_1),(t_2,a_2),\dots ,(t_i,a_i)]$ and the set of example solution processes as $\mathcal{P}_e\in \mathcal{E}$ . At each step, we perform the LCS matching between $\mathcal{P}_i$ and each element of $\mathcal{P}^e$ , retrieving the top-k examples in $\mathcal{P}^e$ with the highest matching degree. Formally: + +$$ +\operatorname {R e t} (\mathcal {E}, \mathcal {P} _ {i}) = \operatorname {T o p K} _ {\mathcal {P} ^ {e} \in \mathcal {E}} \left\{\operatorname {L C S} (\mathcal {P} ^ {e}, \mathcal {P} _ {i}) \right\} +$$ + +This approach enables us to adaptively match each image with its most relevant example, while iteratively refining the match as new information is acquired at each step. + +Human Annotations and Analysis. After the data synthesis process, we invite 10 human annotators to filter the generated data and re-associate the highest-quality subset, resulting in 2,550 test samples. The initial synthetic data is estimated to have a $90.8\%$ accuracy in human validation on 4k randomly sampled data, which is higher than that reported in previous synthetic data methods [8, 24]. Notably, the ChatGPT-4o-latest model, which we use to generate questions, demonstrates a low accuracy rate (less than $40\%$ ) in solving them. This highlights the effectiveness of our ToolEngine pipeline in decoupling complex reasoning into a series of simpler, single-step tasks. While ChatGPT-4o-latest excels in handling $\mathcal{I} \to \mathcal{P} \to \mathcal{Q}, \mathcal{A}$ process for straightforward, step-by-step reasoning, it faces challenges with the more challenging end-to-end $\mathcal{I}, \mathcal{Q} \to \mathcal{P} \to \mathcal{A}$ reasoning. This observation indicates a clear gap between LFM's single-step and multi-step reasoning capabilities. + +# 3.3. Dataset Statistics + +Tab. 2 presents the basic statistics of ToolVQA, which encompasses various question types (with both text and image as answers) and complex reasoning processes (with an average reasoning trajectory length of 2.78 steps). Notably, + +ToolVQA's questions and answers are concisely formulated (with an average token count of 15.74 for questions and 2.69 for answers). In contrast, previous synthetic datasets [8, 24] often fail to control answer length, leading to hallucinated content within the answers, which is detrimental to training. Fig. 5 demonstrates that our tool frequency aligns with real-world user needs, reinforcing the real-world setting of our dataset. Fig. 6 presents the distribution of reasoning trajectory lengths in ToolVQA. + +# 3.4. Ablation on ToolEngine + +To verify whether ToolEngine truly improves the quality of synthetic data, we conducted an ablation study on different synthetic methods. We invite 10 users to evaluate 1k randomly sampled data based on four key metrics: (1) answer correctness (Acc.); (2) correlation (Cor.): whether each sample's image content and Q-A pair are aligned; (3) tool redundancy (Nec.): whether each tool call is necessary for obtaining the answer; and (4) reasoning complexity (R.C.): the average reasoning depth for each task (same as defined in Tab. 1). The results, presented in Tab. 3, show that Real-World Example and LCS matching significantly improved Acc., Nec., and R.C., which highlights their role in the DFS-based simulation of reasoning chains: (1) Real-World Example and LCS matching introduce prior knowledge aligned with image content, enhancing both answer accuracy and tool necessity. (2) Real-World Example and LCS matching incorporate different types of knowledge at each step, increasing the reasoning complexity. Furthermore, the absence of LCS matching has a significant impact on data quality (e.g. Acc. dropped from 90.8 to 41.6), indicating that a fixed set of examples cannot adapt to different scenarios. Dynamically matching appropriate examples is crucial for significantly enhancing the generalization ability of the synthetic pipeline. However, both methods had a limited impact on Cor., likely because the existing controller already effectively captures image content, minimizing errors caused by misleading examples. + +# 4. Experiments + +# 4.1. Training and Test Settings + +Training Objective. In our dataset $\mathcal{D}$ , for each datapoint $\mathcal{E} = (\mathcal{I},\mathcal{T},\mathcal{Q},\mathcal{A},\mathcal{P})$ contains $n$ rounds of dialogue ( $n - 1$ rounds for tool usage and one round for final answer), we train our tool agent using cross-entropy loss: + +$$ +\mathcal {L} = \mathbb {E} _ {\mathcal {E} \sim \mathcal {D}} \left[ \frac {1}{n} \sum_ {i = 1} ^ {n} - \log p \left(t _ {i}, a _ {i}, r _ {i} \mid \mathcal {I}, \mathcal {T}, \mathcal {Q}, \mathcal {P} _ {i - 1}\right) \right] +$$ + +Training Setting. We divide ToolVQA into training and test sets: the training set consists of 21,105 automatically generated samples (with a human verification accu + +
StatisticsNumber
Samples23655
Tool Calls65785
Text answers15806
Image answers7849
Query length15.74
Answer length2.69
Trajectory length2.78
+ +![](images/189e4ff5e10624ab50756d59c0f851d89545143760aef97fe169224840ee0a87.jpg) +Figure 5. Tool frequency. + +![](images/56531ecff38782fea97298190c0d2befefb583f0c22918bd695d327e26b18ea7.jpg) +Figure 6. Trajectory length distribution. + +Table 2. Basic statistics. + +
MethodAcc.↑Cor. ↑Nec. ↑R.C. ↑
ToolEngine90.885.287.512.38
-w/o Example + LCS27.377.621.041.1
-w/o LCS41.681.454.261.61
+ +Table 3. Ablation study on different synthetic methods. + +racy of $90.8\%$ ), while the test set contains 2,550 humanreannotated samples. We fine-tune LLaVA-7B [21] in the Lego Agent framework [15]. + +Test Setting. We apply two test modes: end-to-end and step-by-step modes, to comprehensively evaluate LFMs' performance under different settings. The end-to-end mode only focuses on the end-to-end VQA ability of LFMs. Formally, we evaluate the $\mathcal{I} + \mathcal{Q}\rightarrow \mathcal{A}$ process. The step-by-step mode focuses on evaluating LM's tool-usage capabilities, formally the $\mathcal{I} + \mathcal{Q} + \mathcal{P}_i\to \mathcal{P}_{i + 1}$ process. We use four metrics [39]: (1) InstAcc (Inst.): the percentage of steps executed without errors; (2) ToolAcc (Tool.): the accuracy of tool selection; (3) ArgAcc (Arg.): the accuracy of tool's argument name prediction, and (4) SummAcc (Summ.): the accuracy of final answer's summary. Previous work has focused on the research of large language model (LLM)-based tool agents while ignoring vision-language model (VLM)-based tool agents and the end-to-end VQA abilities of VLMs. To bridge these gaps, our test involves three settings: (1) VLM, directly answering questions end-to-end, formally $\mathcal{Q} + \mathcal{I}\rightarrow \mathcal{A}$ , (2) VLM+tool, calling the tool through VLM, formally $\mathcal{Q} + \mathcal{I} + \mathcal{T}\rightarrow \mathcal{P} + \mathcal{A}$ , (3) LLM+tool, calling the tool through LLM, formally $\mathcal{Q} + \mathcal{T}\rightarrow \mathcal{P} + \mathcal{A}$ (Note that image information can still be obtained through tools such as caption under this setting). + +# 4.2. Main Results + +Scale. As shown in Tab. 4, the tool-use capabilities of LFMs generally improve with the increase in model parameters. For example, Qwen2-7B-instruct surpasses Qwen2-2B-instruct by $7.03\%$ , and closed-sourced LM (larger than + +100B) is generally better than open-sourced (2B,7B). After fine-tuning, our 7B model's performance (last 3 lines) is close to the large scale LFM GPT-3.5-Turbo. This demonstrates that fine-tuning significantly enhances the model's real-world tool-use capabilities, even in smaller models. + +Settings. By comparing the performance of the same model under different settings (VLM, VLM+tool, and LLM+tool) in Tab. 4, we identify several interesting observations. (1) In models such as GPT-4o, Qwen2-7B/2B, and LLaVA-7B, the VLM+tool setting performs worse than the LLM+tool setting, suggesting that the visual modules in current LFM architectures are still ineffective at guiding tool usage. (2) For models including GPT-4o, Claude-3.5, LLaVA-7B, and Qwen2-2B, we find that VLM+tool underperforms VLM alone. This indicates that the noise introduced by tool usage can outweigh the benefits brought by the tools. However, the tuned LLaVA model demonstrates a clear progression: VLM $<$ LLM+tool $<$ VLM+tool, suggesting that fine-tuning can indeed enhance the model's ability to make effective use of tools while suppressing tool-induced noise. Nevertheless, the similar performance of the LLM+tool and VLM+tool settings implies that the visual module's potential in facilitating tool use has not yet been fully explored. + +Bottlenecks. The last two rows of Tab. 4 indicate that the performance of fine-tuned models is often constrained by their weakest capabilities, which are typically argument prediction and answer summarization. While fine-tuning significantly improves instruction formatting and tool selection, it makes little progress in weaker capabilities. Specifically, instruction formatting and tool selection can be accomplished through memorizing patterns, as the number of reasonable tool invocation types in real-world scenarios is limited. However, argument prediction and answer summarization require the model to genuinely understand the new information returned by tools and extract meaningful responses from it. This information varies across different scenarios and queries, and even within multi-turn dialogues of the same scenario. Consequently, a high level of generalization capability in vision-language joint reasoning is essential to handle these issues in tool-use tasks. + +
ModelSettingEnd-to-End ModeStep-by-Step Mode
Acc. (%)Tool Call (Times)Tool Error (Times)Inst. (%)Tool. (%)Arg. (%)Summ. (%)
Closed-source LFMs
ChatGPT-4o-latestVLM38.29------
VLM + tool34.962003180436.514.688.9256.1
LLM + tool37.624687196273.0241.1225.2765.45
Claude-3-5-sonnetVLM30.33------
VLM + tool30.5258552445.7116.947.3646.06
LLM + tool24.48183467266.8339.2913.9875.52
GPT-3.5-TurboLLM + tool18.373178132173.2430.4620.0858.18
Open-source LFMs
Qwen2-VL-7B-InstructVLM3.64------
VLM + tool5.861373105135.129.021.0938.18
LLM + tool6.781366108939.4611.23.5544.85
Qwen2-7B-InstructLLM + tool11.53177370363.124.458.269.7
LLaVA-v1.5-7BVLM8.57------
VLM + tool1.179684968416.399.4300.01
LLM + tool1.6610421104219.2515.03011.52
LLaMA-3-8b-instructLLM + tool4.936607654251.6314.130.5341.99
Qwen2-VL-2B-InstructVLM7.71------
VLM + tool2.14067391547.1612.70.4121.82
LLM + tool2.226892665536.3410.520.6832.12
Qwen2-1.5B-InstructLLM + tool4.57933614150.9521.454.5144.24
Tuned LLaVA-7B (Ours)VLM7.21------
VLM + tool18.8431161786.6261.6139.3430.91
LLM + tool18.43446268087.5162.340.5726.67
+ +Table 4. Results on ToolVQA's test set. + +# 4.3. Out-of-Distribution Benchmarks + +To test the generalization capability of our tuned model, we conduct experiments on out-of-distribution (OOD) benchmarks including (1) Seen tools, unseen queries: TextVQA [36], TallyQA [1], InfoSeek [4]; and (2) Unseen tools: GTA [39], TEMPLAMA [6]. The results are shown in the Tab. 5. Our fine-tuned model significantly outperforms the baseline model LLaVA-7B by $5.8\%$ , $4.2\%$ , $8.6\%$ , $21.17\%$ , and $18.37\%$ accuracy on these datasets, showing excellent generalizability. Its performance also surpasses the large closed-source model GPT-3.5-Turbo on unseen queries. However, its performance on TEMPLAMA is inferior to GPT-3.5-Turbo, possibly because GPT encountered similar data during its training (due to its great coverage of extensive training data), whereas our model had no exposure to such data at all. + +# 4.4. Few-shot In-context Learning + +Few-shot in-context learning (ICL) has become a common and important technique in tool-use tasks. However, prior model fine-tuning work often overlooks evaluating the ICL capabilities of fine-tuned models. We believe that fine- + +tuning and ICL can be complementary, offering significant potential for improvement. Tab. 6 shows few-shot ICL performance on ToolVQA. GPT-4o steadily improves with more shots, while GPT-3.5 and LLaVA both show limited gains and even decline slightly at 10-shot. This may be due to long contexts impairing the LLM's ability to interpret precise tool-use instructions. Fine-tuned LLaVA outperforms the baseline and continues to benefit from ICL. These results suggest that (1) compared to fine-tuning, ICL's improvement is limited and affected by context lengths, and (2) fine-tuned models can still benefit from ICL, highlighting the complementary value of these two approaches in enhancing tool-use performance. + +# 4.5. Ablation Study + +To further evaluate the quality of the fine-tuning data from multiple perspectives, we conduct an ablation study of the different training strategies. As shown in Tab. 7, removing the ImageCaption tool significantly degrades both accuracy and parameter correctness. This highlights the model's reliance on high-quality captions for initial scene understanding, likely due to its limited inherent visual capability for fine-grained information extraction. When all tools + +
ModelTextVQA [36]TallyQA [1]InfoSeek [4]GTA [39]TEMPLAMA [6]
GPT-3.5-Turbo36.36111.323.6233.67
LLaVA-7B41.260.15.212.123.06
Tuned LLaVA-7B4764.313.833.2921.43
+ +Table 5. Results on OOD benchmarks. + +
ModelZero-shot1-shot5-shot10-shot
GPT-4o34.9637.2038.4138.63
GPT-3.5-Turbo18.3719.8320.4019.04
LLaVA-7B1.173.454.273.67
Tuned LLaVA-7B18.8019.4121.1320.69
+ +
ModelSettingAcc.Tool CallTool Error
Tuned LLaVA-7BVLM + tool18.84311617
-w/o CaptionVLM + tool11.11861192
-w/o PVLM8.01--
-w/o P + ILLM0.18--
+ +Table 6. Few-shot performance on ToolVQA. + +Table 7. Ablation study on training methods. + +![](images/e17e61b0d7afdc57519a0e7fc829a3c3f1fd9106b1f61b08059026f3baf96e63.jpg) +Figure 7. Visualization of the main error types made by tuned LLaVA. + +are removed and ToolVQA is reduced to a single-hop VQA model, performance drops substantially, underscoring the necessity of multi-step tool usage. In the absence of both tools and images, performance approaches zero, indicating that the model is heavily dependent on visual input when answering questions in ToolVQA. + +# 4.6. Error Analysis + +In Fig. 7, we randomly select 100 errors made by tuned LLaVA, and manually count and classify the main error types. The two main error types are: (1) Argument prediction error, for example, missing essential keywords "age" during a search. (2) Answer summary error, for example, extracting incorrect information "1,300" from tool outputs. The common cause of these errors is the model's failure to recognize key terms (e.g. "age" and "pound") in the tool outputs, even though this should not be a challenging task for a language model. This suggests that the model still struggles with dynamically adapting to additional information provided by tools. Since multi-step reasoning often suffers from error accumulation (early errors propagate to sub + +sequent steps), errors in processing tool results at any step will affect the results. Fine-tuning has limited improvement on the capability for dynamically processing tool results, but ToolVQA still provides a valuable benchmark for evaluating this capability. + +# 5. Conclusion + +We present ToolVQA, a large multimodal dataset for real-world tool-augmented reasoning. Using our ToolEngine pipeline, we generate 23K real-world samples with implicit multi-step reasoning over 10 tools and 7 domains. This dataset challenges existing models beyond synthetic settings and better reflects real user needs. Experiments show that LLaVA-7B fine-tuned on ToolVQA not only excels on the in-domain test set but also outperforms GPT-3.5-turbo on five OOD benchmarks, establishing a new state-of-the-art for open-source models in complex tool-use VQA. ToolVQA serves as both a benchmark and a training ground for developing more capable, generalizable tool-using agents. + +Acknowledgements. This work was supported by the grants from the National Natural Science Foundation of China 62372014, Beijing Nova Program and Beijing Natural Science Foundation 4252040. + +# References + +[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions, 2018. 2, 3, 7, 8 +[2] Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C. Lawrence Zitnick, Dhruv Batra, and Devi Parikh. Vqa: Visual question answering, 2016. 2 +[3] Lluis Castrejon, Thomas Mensink, Howard Zhou, Vittorio Ferrari, Andre Araujo, and Jasper Uijlings. Hammr: Hierarchical multimodal react agents for generic vqa, 2024. 2 +[4] Yang Chen, Hexiang Hu, Yi Luan, Haitian Sun, Soravit Changpinyo, Alan Ritter, and Ming-Wei Chang. Can pre-trained vision and language models answer visual information-seeking questions?, 2023. 2, 3, 7, 8 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2 +[6] Bhuwan Dhingra, Jeremy R. Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W. Cohen. Time-aware language models as temporal knowledge bases. Transactions of the Association for Computational Linguistics, 10:257-273, 2022. 2, 7, 8 +[7] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn, 2023. 1 +[8] Zhi Gao, Bofei Zhang, Pengxiang Li, Xiaojian Ma, Tao Yuan, Yue Fan, Yuwei Wu, Yunde Jia, Song-Chun Zhu, and Qing Li. Multi-modal agent tuning: Building a vlm-driven agent for efficient tool usage, 2025. 1, 2, 3, 5 +[9] Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. Toolkengpt: Augmenting frozen language models with massive tools via tool embeddings, 2024. 2 +[10] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models, 2024. 1, 2, 3 +[11] Ziniu Hu, Ahmet Iscen, Chen Sun, Kai-Wei Chang, Yizhou Sun, David A Ross, Cordelia Schmid, and Alireza Fathi. Avis: Autonomous visual information seeking with large language model agent, 2023. 1 +[12] Yue Huang, Jiawen Shi, Yuan Li, Chenrui Fan, Siyuan Wu, Qihui Zhang, Yixin Liu, Pan Zhou, Yao Wan, Neil Zhenqiang Gong, and Lichao Sun. Metatool benchmark for large language models: Deciding whether to use tools and which to use, 2024. 1, 2 +[13] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional + +question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2 +[14] Faria Huq, Zora Zhiruo Wang, Frank F. Xu, Tianyue Ou, Shuyan Zhou, Jeffrey P. Bigham, and Graham Neubig. Cow-pilot: A framework for autonomous and human-agent collaborative web navigation, 2025. 2 +[15] InternLM. Agentlego: Enhance llm agents with rich tool apis. https://github.com/InternLM/agentlego, 2024.6 +[16] Iat Long Iong, Xiao Liu, Yuxuan Chen, Hanyu Lai, Shuntian Yao, Pengbo Shen, Hao Yu, Yuxiao Dong, and Jie Tang. OpenWebAgent: An open toolkit to enable web agents on large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 72-81, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2 +[17] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks, 2024. 2 +[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2 +[19] Minghao Li, Feifan Song, Yu Bowen, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. Api-bank: A comprehensive benchmark for tool-augmented llms. In Conference on Empirical Methods in Natural Language Processing, 2023. 2 +[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2 +[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 6 +[22] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In The 37th Conference on Neural Information Processing Systems (NeurIPS), 2023. 2 +[23] Kaixin Ma, Hongming Zhang, Hongwei Wang, Xiaoman Pan, Wenhao Yu, and Dong Yu. Laser: LIm agent with statespace exploration for web navigation, 2024. 2 +[24] Zixian Ma, Weikai Huang, Jieyu Zhang, Tanmay Gupta, and Ranjay Krishna. m&m's: A benchmark to evaluate tool-use for multi-step multi-modal tasks. EECV 2024, 2024. 1, 2, 3, 5 +[25] Arjun Mani, Nobline Yoo, Will Hinthorn, and Olga Russakovsky. Point and ask: Incorporating pointing into visual question answering, 2022. 2 + +[26] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge, 2019. 3 +[27] Grégoire Mialon, Clémentine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, and Thomas Scialom. Gaia: a benchmark for general ai assistants, 2023. 1, 2, 3, 4 +[28] Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 2, 4 +[29] Yun Peng, Shuqing Li, Wenwei Gu, Yichen Li, Wenxuan Wang, Cuiyun Gao, and Michael R. Lyu. Revisiting, benchmarking and exploring api recommendation: How far are we? IEEE Transactions on Software Engineering, 49:1876-1897, 2021. 2 +[30] Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. Toollm: Facilitating large language models to master 16000+ real-world apis, 2023. 2, 3 +[31] Yujia Qin, Shengding Hu, Yankai Lin, Weize Chen, Ning Ding, Ganqu Cui, Zheni Zeng, Yufei Huang, Chaojun Xiao, Chi Han, Yi Ren Fung, Yusheng Su, Huadong Wang, Cheng Qian, Runchu Tian, Kunlun Zhu, Shihao Liang, Xingyu Shen, Bokai Xu, Zhen Zhang, Yining Ye, Bowen Li, Ziwei Tang, Jing Yi, Yuzhang Zhu, Zhenning Dai, Lan Yan, Xin Cong, Yaxi Lu, Weilin Zhao, Yuxiang Huang, Junxi Yan, Xu Han, Xian Sun, Dahai Li, Jason Phang, Cheng Yang, Tongshuang Wu, Heng Ji, Zhiyuan Liu, and Maosong Sun. Tool learning with foundation models, 2024. 1 +[32] Yangjun Ruan, Honghua Dong, Andrew Wang, Silviu Pitis, Yongchao Zhou, Jimmy Ba, Yann Dubois, Chris J. Maddison, and Tatsunori Hashimoto. Identifying the risks of lm agents with an lm-emulated sandbox, 2024. 2 +[33] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools, 2023. 1, 2 +[34] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge, 2022. 2 +[35] Yongliang Shen, Kaitao Song, Xu Tan, Wenqi Zhang, Kan Ren, Siyu Yuan, Weiming Lu, Dongsheng Li, and Yueting Zhuang. Taskbench: Benchmarking large language models for task automation, 2024. 2 +[36] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read, 2019. 2, 3, 7, 8 +[37] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11888-11898, 2023. 1, 2 +[38] Chenyu Wang, Weixin Luo, Qianyu Chen, Haonan Mai, Jindi Guo, Sixun Dong, Xiaohua (Michael) Xuan, Zhengxin Li, Lin Ma, and Shenghua Gao. Mllm-tool: A multimodal + +large language model for tool agent learning. arXiv preprint arXiv:2401.10727, 2024. 2 +[39] Jize Wang, Zerun Ma, Yining Li, Songyang Zhang, Cailian Chen, Kai Chen, and Xinyi Le. Gta: A benchmark for general tool agents, 2024. 1, 2, 3, 6, 7, 8 +[40] Yan Wang, Yawen Zeng, Jingsheng Zheng, Xiaofen Xing, Jin Xu, and Xiangmin Xu. Videocot: A video chain-of-thought dataset with active annotation tool, 2024. 2 +[41] Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. Agent workflow memory, 2024. 2 +[42] Shirley Wu, Shiyu Zhao, Qian Huang, Kexin Huang, Michihiro Yasunaga, Kaidi Cao, Vassilis N. Ioannidis, Karthik Subbian, Jure Leskovec, and James Zou. Avatar: Optimizing llm agents for tool usage via contrastive reasoning, 2024. 2 +[43] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments, 2024. 2 +[44] Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering, 2018. 2 +[45] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action, 2023. 2 +[46] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent), 2024. 2 +[47] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents, 2023. 2 +[48] Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. Webarena: A realistic web environment for building autonomous agents, 2024. 2 +[49] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images, 2016. 2 +[50] Yuchen Zhuang, Yue Yu, Kuan Wang, Haotian Sun, and Chao Zhang. Toolqa: A dataset for llm question answering with external tools, 2023. 2, 3 \ No newline at end of file diff --git a/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/images.zip b/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..15460d94dd83d15dbc8de79686f03768b3e946f4 --- /dev/null +++ b/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:601903aa99d373fd5916fdc0bd3d2690d6bfbc20902a72ed803e52b634f35f74 +size 672688 diff --git a/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/layout.json b/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..50cc6e7cfc2d1ee20c0f02150b68bba489ea6be2 --- /dev/null +++ b/2025/ToolVQA_ A Dataset for Multi-step Reasoning VQA with External Tools/layout.json @@ -0,0 +1,7860 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "type": "text", + "content": "ToolVQA: A Dataset for Multi-step Reasoning VQA with External Tools" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 147, + 144, + 460, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 144, + 460, + 174 + ], + "spans": [ + { + "bbox": [ + 147, + 144, + 460, + 174 + ], + "type": "text", + "content": "Shaofeng Yin Ting Lei Yang Liu* Wangxuan Institute of Computer Technology, Peking University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 174, + 461, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 174, + 461, + 186 + ], + "spans": [ + { + "bbox": [ + 146, + 174, + 461, + 186 + ], + "type": "text", + "content": "yin_shaofeng@stu.pku.edu.cn {tinglei, yangliu}@pku.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 239, + 297, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 562 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 562 + ], + "type": "text", + "content": "Integrating external tools into Large Foundation Models (LFMs) has emerged as a promising approach to enhance their problem-solving capabilities. While existing studies have demonstrated strong performance in tool-augmented Visual Question Answering (VQA), recent benchmarks reveal significant gaps in real-world tool-use proficiency, particularly in functionally diverse multimodal settings requiring multi-step reasoning. In this work, we introduce ToolVQA, a large-scale multimodal dataset comprising 23K samples, designed to bridge this gap. Unlike previous datasets that rely on synthetic scenarios and simplified queries, ToolVQA features real-world visual contexts and challenging implicit multi-step reasoning tasks, better aligning with real user interactions. To construct this dataset, we propose ToolEngine, a novel data generation pipeline that employs image-guided Depth-First Search (DFS) with a Longest Common Subsequence (LCS)-based example matching mechanism to simulate human-like tool-use reasoning. ToolVQA encompasses 10 multimodal tools across 7 diverse domains, with an average inference length of 2.78 reasoning steps per sample. The LLaVA-7B model fine-tuned on ToolVQA not only achieves impressive performance on the ToolVQA test set, but also surpasses the large closed-source model GPT-3.5-turbo on five out-of-distribution (OOD) datasets, showing strong generalizability in real-world tool-use scenarios. Code is available at https://github.com/Fugtemypt123/ToolVQA-release." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 587, + 135, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 587, + 135, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 587, + 135, + 600 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 608, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 296, + 693 + ], + "type": "text", + "content": "Integrating external tools into Large Foundation Models (LFMs) [31] is a promising way to build versatile AI assistants. By leveraging external tools, LFMs can break down complex tasks into smaller, more manageable subtasks, each handled by a specialized tool. Prior works [7, 10, 11, 33, 37] show strong performance on traditional Visual Question Answering (VQA) benchmarks using extern" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 334, + 224, + 413, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 224, + 413, + 232 + ], + "spans": [ + { + "bbox": [ + 334, + 224, + 413, + 232 + ], + "type": "text", + "content": "Non Real-World Scenarios" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 329, + 237, + 359, + 258 + ], + "blocks": [ + { + "bbox": [ + 329, + 237, + 359, + 258 + ], + "lines": [ + { + "bbox": [ + 329, + 237, + 359, + 258 + ], + "spans": [ + { + "bbox": [ + 329, + 237, + 359, + 258 + ], + "type": "image", + "image_path": "5fb20371181387a9671fca03aff1c85993942266b1e06d4fedd8f184bc448f26.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 364, + 239, + 414, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 239, + 414, + 253 + ], + "spans": [ + { + "bbox": [ + 364, + 239, + 414, + 253 + ], + "type": "interline_equation", + "content": "2 x ^ {2} - 1 2 = 0", + "image_path": "682537afe0d563e13f7d1ad13378210415a49ae0ac7e30148f3950ef3a114777.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 332, + 266, + 343, + 281 + ], + "blocks": [ + { + "bbox": [ + 332, + 266, + 343, + 281 + ], + "lines": [ + { + "bbox": [ + 332, + 266, + 343, + 281 + ], + "spans": [ + { + "bbox": [ + 332, + 266, + 343, + 281 + ], + "type": "image", + "image_path": "83aea2f60d18d44e2e35805ed5a13d2ee244cbdd04f0df59ee00139b8fb1732f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 266, + 408, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 266, + 408, + 273 + ], + "spans": [ + { + "bbox": [ + 351, + 266, + 408, + 273 + ], + "type": "text", + "content": "From:Shanghai To:Beijing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 353, + 275, + 365, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 275, + 365, + 281 + ], + "spans": [ + { + "bbox": [ + 353, + 275, + 365, + 281 + ], + "type": "text", + "content": "Speed:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 386, + 275, + 406, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 275, + 406, + 281 + ], + "spans": [ + { + "bbox": [ + 386, + 275, + 406, + 281 + ], + "type": "text", + "content": "250 km/h" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 331, + 291, + 415, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 291, + 415, + 298 + ], + "spans": [ + { + "bbox": [ + 331, + 291, + 415, + 298 + ], + "type": "text", + "content": "(Overly simplistic synthetic scenarios)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 336, + 317, + 410, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 317, + 410, + 324 + ], + "spans": [ + { + "bbox": [ + 336, + 317, + 410, + 324 + ], + "type": "text", + "content": "Non Real-World Queries" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 328, + 329, + 416, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 329, + 416, + 356 + ], + "spans": [ + { + "bbox": [ + 328, + 329, + 416, + 356 + ], + "type": "text", + "content": "Based on the text 'James Hutton is often viewed as the first modern geologist...', could you help me find out who the first modern geologist is?" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 340, + 357, + 406, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 357, + 406, + 362 + ], + "spans": [ + { + "bbox": [ + 340, + 357, + 406, + 362 + ], + "type": "text", + "content": "(Simple retrieve, no tool use)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 327, + 369, + 417, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 369, + 417, + 394 + ], + "spans": [ + { + "bbox": [ + 327, + 369, + 417, + 394 + ], + "type": "text", + "content": "Can you help me find popular videos and trending gaming videos. I'm using the 'Cheap YouTube API' tool. (Explicit tool use, no reasoning)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 327, + 400, + 414, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 400, + 414, + 434 + ], + "spans": [ + { + "bbox": [ + 327, + 400, + 414, + 434 + ], + "type": "text", + "content": "Find me some interesting news articles about the culinary world. Additionally, provide me with the current threshold securities list for NVIDIA's stock. (Just tool list, no multi-step reasoning)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 462, + 224, + 518, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 462, + 224, + 518, + 232 + ], + "spans": [ + { + "bbox": [ + 462, + 224, + 518, + 232 + ], + "type": "text", + "content": "Real-World Setting" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 436, + 234, + 487, + 283 + ], + "blocks": [ + { + "bbox": [ + 436, + 234, + 487, + 283 + ], + "lines": [ + { + "bbox": [ + 436, + 234, + 487, + 283 + ], + "spans": [ + { + "bbox": [ + 436, + 234, + 487, + 283 + ], + "type": "image", + "image_path": "f0c4732c18d2011d9a0fb34ecdf24e773a6c943030d6e62c55ffa727c2ac4881.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 492, + 235, + 543, + 282 + ], + "blocks": [ + { + "bbox": [ + 492, + 235, + 543, + 282 + ], + "lines": [ + { + "bbox": [ + 492, + 235, + 543, + 282 + ], + "spans": [ + { + "bbox": [ + 492, + 235, + 543, + 282 + ], + "type": "image", + "image_path": "c213c1444745538399e00d78ad01501a71c5f019d5a7b06ce0966362d12f5bb2.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 437, + 295, + 449, + 304 + ], + "blocks": [ + { + "bbox": [ + 437, + 295, + 449, + 304 + ], + "lines": [ + { + "bbox": [ + 437, + 295, + 449, + 304 + ], + "spans": [ + { + "bbox": [ + 437, + 295, + 449, + 304 + ], + "type": "image", + "image_path": "c43bf96797156d8322e7dff3ed906168cf2d888a04bdfe0c64299600a2235337.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 451, + 296, + 484, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 296, + 484, + 308 + ], + "spans": [ + { + "bbox": [ + 451, + 296, + 484, + 308 + ], + "type": "text", + "content": "a bowl of salad, a sandwich, a sandwich of" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 451, + 308, + 484, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 308, + 484, + 312 + ], + "spans": [ + { + "bbox": [ + 451, + 308, + 484, + 312 + ], + "type": "text", + "content": "and a bottle of beer" + } + ] + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 438, + 312, + 482, + 341 + ], + "blocks": [ + { + "bbox": [ + 438, + 312, + 482, + 341 + ], + "lines": [ + { + "bbox": [ + 438, + 312, + 482, + 341 + ], + "spans": [ + { + "bbox": [ + 438, + 312, + 482, + 341 + ], + "type": "image", + "image_path": "0a2d1d008694823f1b0d6450fb3296c6e4e858cedea6b1629a3d60ffbbafcb30.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 494, + 295, + 541, + 316 + ], + "blocks": [ + { + "bbox": [ + 494, + 295, + 541, + 316 + ], + "lines": [ + { + "bbox": [ + 494, + 295, + 541, + 316 + ], + "spans": [ + { + "bbox": [ + 494, + 295, + 541, + 316 + ], + "type": "image", + "image_path": "eef1f2be73b208412e8577ed42655b778145cceb77262b0deb811d3683b2b583.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 494, + 319, + 540, + 352 + ], + "blocks": [ + { + "bbox": [ + 494, + 319, + 540, + 352 + ], + "lines": [ + { + "bbox": [ + 494, + 319, + 540, + 352 + ], + "spans": [ + { + "bbox": [ + 494, + 319, + 540, + 352 + ], + "type": "image", + "image_path": "0a22fbf87ed9ca6623669e86e370614171a9a99632620fc62893dceecd6c72c1.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 438, + 342, + 449, + 352 + ], + "blocks": [ + { + "bbox": [ + 438, + 342, + 449, + 352 + ], + "lines": [ + { + "bbox": [ + 438, + 342, + 449, + 352 + ], + "spans": [ + { + "bbox": [ + 438, + 342, + 449, + 352 + ], + "type": "image", + "image_path": "13a6a9027e41233b41a46d31e33001dcbabf30f21836ca84fa66f0b0f782336a.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 454, + 354, + 482, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 454, + 354, + 482, + 370 + ], + "spans": [ + { + "bbox": [ + 454, + 354, + 482, + 370 + ], + "type": "text", + "content": "Bridgeport Rewiring Company closed in 2019." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 454, + 371, + 479, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 454, + 371, + 479, + 377 + ], + "spans": [ + { + "bbox": [ + 454, + 371, + 479, + 377 + ], + "type": "text", + "content": "2024-2019=5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 438, + 396, + 482, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 438, + 396, + 482, + 430 + ], + "spans": [ + { + "bbox": [ + 438, + 396, + 482, + 430 + ], + "type": "text", + "content": "Q: How many years has it been since the brewery that produced this beer ceased operations? A: 5" + } + ] + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 493, + 364, + 541, + 402 + ], + "blocks": [ + { + "bbox": [ + 493, + 364, + 541, + 402 + ], + "lines": [ + { + "bbox": [ + 493, + 364, + 541, + 402 + ], + "spans": [ + { + "bbox": [ + 493, + 364, + 541, + 402 + ], + "type": "image", + "image_path": "f0c1c23afc4d881dc16d83c41b674c15e69a29880981ad7c205abe89504c7580.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 494, + 404, + 541, + 434 + ], + "blocks": [ + { + "bbox": [ + 494, + 404, + 541, + 434 + ], + "lines": [ + { + "bbox": [ + 494, + 404, + 541, + 434 + ], + "spans": [ + { + "bbox": [ + 494, + 404, + 541, + 434 + ], + "type": "image", + "image_path": "8717f26408b739818d5b3a8e9179a4ee1ad09a8e432378113083908b9c2eec85.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 456, + 555, + 501 + ], + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 501 + ], + "type": "text", + "content": "Figure 1. Our real-world setting includes (1) complex visual scenarios with real-world context; (2) challenging queries with an implicit multi-step reasoning process. Existing datasets (left) do not meet these requirements, while our ToolVQA (right) does." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "bbox": [ + 312, + 522, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 522, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 312, + 522, + 555, + 594 + ], + "type": "text", + "content": "nal tools. Recent works [24, 27, 39] highlight the limitations of current LFMs in their tool-use capabilities within challenging contexts. These contexts require functionally diverse multimodal tools and multi-step reasoning processes associated with tool usage, posing major challenges to LFMs' tool-use proficiency." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 312, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 556, + 714 + ], + "type": "text", + "content": "One effective approach to enhancing LFM's tool-use capability is fine-tuning on large-scale datasets [8, 10, 12, 33]. However, there remains a significant gap between existing large-scale datasets and real-world user needs. Through the study of real user logs, we find that real-world tool usage involves (1) real-world scenarios: complex visual scenarios with real-world context (e.g. real-taken photos rather than synthetic images), and (2) real-world queries: challenging queries with an implicit multi-step reasoning process. However, existing datasets often lack these two criti" + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4424" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "cal properties, as shown in Fig. 1. The scenarios are synthesized instead of real-taken [8, 39], resulting in simplified scenes that differ from real-world scenarios in complexity. The queries require only single-step reasoning or explicitly provide hints about the reasoning process [10, 30], such as \"using the Cheap YouTube API tool\". These setups simplify the task and may lead to a gap between synthetic and real-world tool usage, which limits our ability to analyze LFMs' performance on this task. Additionally, some datasets [27, 39] rely on costly human annotations, making them difficult to scale, which limits fine-tuning to enhance LFMs' tool-using capabilities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 218, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 295, + 445 + ], + "type": "text", + "content": "To bridge these gaps, we introduce ToolEngine, a data construction pipeline designed to generate multi-step reasoning VQA data for tool usage. ToolEngine employs an image-guided Depth-First Search (DFS) approach [30] to simulate the human-like multi-step reasoning process under the guidance of real user context examples. To facilitate multi-step reasoning in data, we introduce a Longest Common Subsequence (LCS)-based example matching mechanism. During each step in the iterative search process, this mechanism matches different real-world tool-use examples based on the previous search trajectory, guiding the controller in making the next selection. This mechanism enables: (1) Extracting rich visual information from the input image; (2) Integrating this information during the reasoning process to propose challenging queries. By freeing query generation from the constraints of fixed templates [8, 24, 30], ToolEngine enhances the utilization of visual information, consequently increasing the reasoning complexity of the generated queries." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 446, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 446, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 56, + 446, + 295, + 696 + ], + "type": "text", + "content": "We leverage ToolEngine to construct ToolVQA, a large-scale multimodal tool-usage dataset comprising 23K samples. We assess data quality via a manual evaluation conducted on a randomly sampled 4k subset of the training set, which yields an accuracy of " + }, + { + "bbox": [ + 56, + 446, + 295, + 696 + ], + "type": "inline_equation", + "content": "90.8\\%" + }, + { + "bbox": [ + 56, + 446, + 295, + 696 + ], + "type": "text", + "content": ". The test set is carefully filtered and fully human-annotated. As shown in Tab. 1, ToolVQA outperforms prior tool-use datasets across multiple perspectives, providing a more comprehensive representation of real-world scenarios. We evaluate state-of-the-art (SOTA) LFMs on ToolVQA and observe that: (1) Larger LFMs demonstrate superior reasoning capabilities. (2) LFMs consistently underperform relative to humans, primarily due to their inability to integrate new information introduced in multi-turn dialogues. We fine-tune LLaVA-7B [21] on ToolVQA, and the model demonstrated performance surpassing or approaching that of the large closed-source model GPT-3.5-Turbo on both our test set and five out-of-distribution (OOD) benchmarks—TextVQA [36], TallyQA [1], Infoseek [4], TEMPLAMA [6], and GTA [39], which include unseen tasks and tools. In summary, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 701, + 295, + 713 + ], + "type": "text", + "content": "1. ToolEngine: We introduce ToolEngine, a VQA data" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 326, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 553, + 108 + ], + "type": "text", + "content": "construction engine that generates reasonable multi-step tool-use trajectories from unannotated images, and constructs challenging question-answer pairs accordingly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 108, + 553, + 215 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 108, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 167 + ], + "type": "text", + "content": "2. ToolVQA: We propose ToolVQA, a large-scale multimodal tool-use dataset that encompasses 10 multimodal tools across 7 diverse domains, providing a benchmark for evaluating and advancing LFMs' tool-use capabilities on real-world tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 167, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 553, + 215 + ], + "type": "text", + "content": "3. Fine-tuning: We fine-tune LLaVA-7B on ToolVQA to obtain a new tool-use agent, which outperforms large closed-source models like GPT-3.5-Turbo on six benchmarks (ToolVQA and the other five OOD benchmarks)." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 229, + 400, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 400, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 400, + 243 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 249, + 529, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 529, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 529, + 262 + ], + "type": "text", + "content": "2.1. Datasets and Benchmarks for Tool Agent" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 267, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 555, + 458 + ], + "type": "text", + "content": "With the rise of real-world tool-use applications, many works have attempted to build datasets and benchmarks to evaluate the tool-use capabilities of LFMs from different perspectives. Text-based datasets [12, 19, 29, 30, 32, 44, 47, 50] collect a large number of real-world APIs to build large-scale, diverse text-based tool-use examples. On the other hand, multimodal datasets [1, 4, 13, 24, 25, 27, 35, 36, 39, 40] integrate visual input into the query process, making queries challenging and more adaptable to real-world needs. However, they either fail to capture real-world scenarios or rely on expensive human annotation. MM-Traj [8] is a recent work which also constructs large-scale training data using an automated pipeline, but has the following limitations: (1) Synthesizes over-simplified PDF files as the multimodal contexts; (2) Their answers are long-form texts that have not been verified for correctness." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 467, + 450, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 450, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 450, + 479 + ], + "type": "text", + "content": "2.2. LFM-based Tool Agents" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 628 + ], + "type": "text", + "content": "Most recent LFMs have demonstrated impressive reasoning capabilities on numerous zero-shot tasks, which makes it possible to build LFM-based tool agents for real-world tasks. Mainstream approaches include in-context learning [3, 9, 22, 37, 42, 45, 46] and instruction fine-tuning [10, 28, 30, 33, 38]. However, these agents still face challenges in real-world applications. Additionally, recent advancements in web navigation agents [14, 16, 17, 23, 41, 43, 48] have primarily focused on GUI-based interactions, often designed as browser plugins, but their applicability is limited to web-based tasks. In contrast, our ToolEngine extends to a broader range of real-world scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 635, + 435, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 435, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 435, + 648 + ], + "type": "text", + "content": "2.3. Other VQA Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Visual Question Answering (VQA) is one of the most fundamental tasks in multimodal learning. Early VQA datasets [2, 5, 13, 18, 20, 34, 49] primarily focus on simple commonsense question-answering, which typically consists of tasks that humans can easily solve. Recently, some" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4425" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 87, + 70, + 525, + 225 + ], + "blocks": [ + { + "bbox": [ + 87, + 70, + 525, + 225 + ], + "lines": [ + { + "bbox": [ + 87, + 70, + 525, + 225 + ], + "spans": [ + { + "bbox": [ + 87, + 70, + 525, + 225 + ], + "type": "table", + "html": "
DatasetNumber of SamplesNumber of ToolsMultrimodal Inputs?Real-World S/Q Setting?Evaluable Answers?Reasoning Complexity↑
ToolBench [30]126.5K3451X- / XX1.43
ToolQA [50]1.5K13X- / ✓✓1.35
Infoseek [4]1.3M1✓✓ / ✓✓1.13
VPD [10]*310.5K4✓✓ / ✓--
M&m's [24]1.6K33✓✓ / XX1.63
GAIA [27]†0.5K-✓✓ / ✓✓2.61
GTA [39]0.2K14✓X/ ✓✓1.86
MM-Traj [8]‡20K9✓X/ ✓X1.77
ToolVQA (Ours)23.7K10✓✓ / ✓✓2.38
", + "image_path": "e1f21e94af63c72f601a9a2fbef8d17fbd0306cf5f71b127f1a7f2511f26136a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 233, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 555, + 289 + ], + "type": "text", + "content": "Table 1. Comparison between ToolVQA and other datasets. Real-World S/Q Setting: whether each sample in the dataset involves: (1) Scenario—utilizes real rather than synthetic multimodal inputs; (2) Query—implicitly embeds tool usage and reasoning within the query. Reasoning Complexity: the average reasoning depth of humans for each task. VPD* constructs trajectories from existing QA pairs and is not publicly accessible. GAIA†'s each sample corresponds to a distinct set of human-annotated tools. ‡MM-Traj generates over-simplified PDF files as multimodal contexts, and their answers are long-form texts that have not been verified for correctness." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 309, + 294, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 309, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 54, + 309, + 294, + 430 + ], + "type": "text", + "content": "studies [1, 4, 10, 26, 36] begin exploring more challenging forms of VQA, which require various tools like Wikipedia, OCR, Object Detection, etc. However, these datasets typically only evaluate LFM's ability to use one specific tool instead of collaboration between multiple tools, limiting LFM's multi-step reasoning ability within a tool-agent framework. In contrast, ToolVQA contains high-quality and diverse multi-step tool usage traces. It encompasses 10 multimodal tools across 7 diverse domains, with an average inference length of 2.78 reasoning steps per sample." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 451, + 150, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 150, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 150, + 463 + ], + "type": "text", + "content": "3. Data Collection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 162, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 162, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 162, + 485 + ], + "type": "text", + "content": "3.1. Task Formulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": "Tool-use VQA Task. Given the image scenario " + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": ", the toolset " + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": ", and the user query " + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": ", a tool agent is required to provide an answer to the query " + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": " and the tool-use trajectory " + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = [(t_i, a_i)]_{i=1}^n" + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": ", in which " + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "inline_equation", + "content": "t_i \\in \\mathcal{T}" + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": " represents the tool used in step " + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": " represents the calling arguments for " + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": ". Consequently, each sample is a five-tuple " + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "inline_equation", + "content": "(\\mathcal{I}, \\mathcal{T}, \\mathcal{Q}, \\mathcal{A}, \\mathcal{P})" + }, + { + "bbox": [ + 55, + 495, + 295, + 567 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Toolset and Image Source. We collect 10 tools with various functions as our toolset, including (1) Perception: ImageCaption, OCR, ObjectDetection, and RegionDescription. (2) Operation: DrawBox, GoogleSearch. (3) Logic: Calculator, Plot, ItemCount. (4) Creativity: TextToImage. These tools cover a wide range of real-world tasks and offer strong generalization capabilities (e.g. GoogleSearch can retrieve any external knowledge). The different sources of image " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " and the corresponding toolset " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " are shown in Fig. 2. These sources are diverse and complex, including real-world photography, e-commerce products, school exam charts, and other practical scenarios." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 342, + 309, + 528, + 373 + ], + "blocks": [ + { + "bbox": [ + 342, + 309, + 528, + 373 + ], + "lines": [ + { + "bbox": [ + 342, + 309, + 528, + 373 + ], + "spans": [ + { + "bbox": [ + 342, + 309, + 528, + 373 + ], + "type": "image", + "image_path": "32e94094f293334f51ac2db55851378ee6b650d5dfeb49d8afaa8c352f42a3dc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 340, + 375, + 526, + 564 + ], + "blocks": [ + { + "bbox": [ + 340, + 375, + 526, + 564 + ], + "lines": [ + { + "bbox": [ + 340, + 375, + 526, + 564 + ], + "spans": [ + { + "bbox": [ + 340, + 375, + 526, + 564 + ], + "type": "image", + "image_path": "3ad9bd00f9213a45a40c701e168215cfe855292ecb258b1507e8ecd29af2e54d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 578, + 555, + 601 + ], + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 601 + ], + "type": "text", + "content": "Figure 2. Image sources and corresponding tools of ToolVQA. We filter out overly simplistic tables and images from the data sources." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 623, + 391, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 391, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 391, + 635 + ], + "type": "text", + "content": "3.2. ToolEngine" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "We specify the definition of our real-world setting into the following criteria: (1) Align with human users: Each sample's scenario " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": " and query " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": " should align with real-world user needs. We have filtered out uninformative scenarios " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": ", but it remains necessary to ensure that the tool-use queries " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": " correspond to human demands. (2) Real-world" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4426" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 77, + 534, + 323 + ], + "blocks": [ + { + "bbox": [ + 80, + 77, + 534, + 323 + ], + "lines": [ + { + "bbox": [ + 80, + 77, + 534, + 323 + ], + "spans": [ + { + "bbox": [ + 80, + 77, + 534, + 323 + ], + "type": "image", + "image_path": "ea19b08f67518f02dbd735eb50cde54c647096a479758998eac6756a23c6e66e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 338, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 54, + 338, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 54, + 338, + 555, + 384 + ], + "type": "text", + "content": "Figure 3. The pipeline of ToolEngine, contains three core components: Real-world Example Construction, Image-guided DFS on Tool Graph, and LCS-based Example Matching. Given an input image, we perform DFS on the complete tool graph. At each step, an LFM controller generates the next tool's name and arguments, guided by the image, current tool-use trajectory, and matched examples. Once DFS is complete, the tool-use trajectory is determined, and then used to generate the query and answer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 403, + 297, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 297, + 584 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 297, + 584 + ], + "type": "text", + "content": "deployed tools: Real-world tool results are complex distributions with noise, and the model must learn how to deal with the noise from real results. Some work [27, 28] uses simulated tool outputs to reduce noise, but we believe this reduces the realism of the training environment. (3) Multi-step reasoning: The tool-use trajectory " + }, + { + "bbox": [ + 54, + 403, + 297, + 584 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 54, + 403, + 297, + 584 + ], + "type": "text", + "content": " should involve multiple steps. Each step in it must have an inherent logical connection, forming a complete problem-solving trajectory. It should not include ineffective steps (e.g. apply OCR to an image without text), or contain non-related tool usage in a single task (e.g. use \"culinary world\" and \"NVIDIA's stock\" sequentially, which should be split into two separate tasks). We propose ToolEngine, an automated data synthesis pipeline that ensures these conditions. Fig. 3 illustrates the three core components of ToolEngine:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 594, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 297, + 715 + ], + "type": "text", + "content": "Real-World Example Construction. To better reflect real user needs, we incorporate human prior tool-use knowledge into the LFM controller through in-context examples. Human-constructed examples provide prior knowledge about typical tool-use scenarios (e.g. images containing specific entities may require a search tool, while tables with multi-digit numbers may necessitate a calculator). This information is crucial for synthesizing real-world queries. Therefore, we ask human users to construct a small set of real-world examples " + }, + { + "bbox": [ + 54, + 594, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{(\\mathcal{I},\\mathcal{T},\\mathcal{Q},\\mathcal{A},\\mathcal{P})\\}" + }, + { + "bbox": [ + 54, + 594, + 297, + 715 + ], + "type": "text", + "content": ". Each" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 403, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 462 + ], + "type": "text", + "content": "example in this set captures a logically valid trajectory, serving as prior knowledge for automatic construction. This approach not only aligns with how humans intuitively associate tools but also provides a structured, intuitive foundation for the construction process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 463, + 556, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 556, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 556, + 608 + ], + "type": "text", + "content": "Image-guided DFS on Tool Graph. To construct a multi-step tool-use trajectory, we employ image-guided Depth-First Search (DFS) on a tool graph to construct a reasonable tool-use trajectory for each image. While LFM can directly generate reasoning sequences, unrealistic tool invocations often lead to less reliable tool-use trajectories in practice. Our DFS trajectory involves real tool calls to extract detailed information from the input image, which is then used to generate challenging question-answer pairs. Specifically, we utilize the most advanced LFM (ChatGPT-4o-latest), denoted as " + }, + { + "bbox": [ + 313, + 463, + 556, + 608 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 463, + 556, + 608 + ], + "type": "text", + "content": ", as the controller to select the tool at each step and generate the corresponding arguments. Formally:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 611, + 556, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 611, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 314, + 611, + 556, + 624 + ], + "type": "interline_equation", + "content": "t _ {i} = \\mathcal {M} (\\text {c h o i c e s} = \\mathcal {T}, \\text {i m a g e} = \\mathcal {I}, \\text {e x a m p l e s} = \\operatorname {R e t} (\\mathcal {E}, \\mathcal {P} _ {i - 1}))", + "image_path": "af95426e06099057ccc6666a10e343e6911d24f774a18f07aaed31becb0de3f7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 625, + 531, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 625, + 531, + 637 + ], + "spans": [ + { + "bbox": [ + 315, + 625, + 531, + 637 + ], + "type": "interline_equation", + "content": "a _ {i} = \\mathcal {M} (t o o l = t _ {i}, i m a g e = \\mathcal {I}, e x a m p l e s = \\operatorname {R e t} (\\mathcal {E}, \\mathcal {P} _ {i}))", + "image_path": "025c1122fe232c572b40a7eac3bb6874673f28d8ae124a57b7a9e6b942437ee8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "LCS-based Example-Matching. We combine the information obtained from different tools into a reasonable reasoning trajectory to perform multi-step reasoning. Previous DFS-based methods typically match a fixed group of examples, which limited their ability to combine sufficient information for multi-step reasoning. As illustrated in Fig. 4, we" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4427" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 293, + 174 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 293, + 174 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 293, + 174 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 293, + 174 + ], + "type": "image", + "image_path": "da01cb65839426249d40d19d6191ec62278e3a9a1b4188a2d0366cd5fabc1748.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 295, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 295, + 220 + ], + "type": "text", + "content": "Figure 4. Comparison between different matching strategies. When a fixed example limits the diversity of generation, LCS matching can integrate multiple examples to enhance generation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "text", + "content": "propose an example-matching method based on the Longest Common Subsequence (LCS) algorithm, which enables dynamic matching during DFS. This approach better leverages the information extraction capabilities of diverse examples, facilitating more effective multi-step reasoning. Specifically, in step " + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "text", + "content": ", the current solution process is denoted as " + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = [(t_1,a_1),(t_2,a_2),\\dots ,(t_i,a_i)]" + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "text", + "content": " and the set of example solution processes as " + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_e\\in \\mathcal{E}" + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "text", + "content": ". At each step, we perform the LCS matching between " + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "text", + "content": " and each element of " + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^e" + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "text", + "content": ", retrieving the top-k examples in " + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^e" + }, + { + "bbox": [ + 55, + 238, + 296, + 370 + ], + "type": "text", + "content": " with the highest matching degree. Formally:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 376, + 263, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 376, + 263, + 397 + ], + "spans": [ + { + "bbox": [ + 88, + 376, + 263, + 397 + ], + "type": "interline_equation", + "content": "\\operatorname {R e t} (\\mathcal {E}, \\mathcal {P} _ {i}) = \\operatorname {T o p K} _ {\\mathcal {P} ^ {e} \\in \\mathcal {E}} \\left\\{\\operatorname {L C S} (\\mathcal {P} ^ {e}, \\mathcal {P} _ {i}) \\right\\}", + "image_path": "7e3ba5eece021baabaad4d15e1578af123e1825b3d2bb26a4edd3b3a7d177900.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 402, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 438 + ], + "type": "text", + "content": "This approach enables us to adaptively match each image with its most relevant example, while iteratively refining the match as new information is acquired at each step." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 438, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 643 + ], + "type": "text", + "content": "Human Annotations and Analysis. After the data synthesis process, we invite 10 human annotators to filter the generated data and re-associate the highest-quality subset, resulting in 2,550 test samples. The initial synthetic data is estimated to have a " + }, + { + "bbox": [ + 55, + 438, + 296, + 643 + ], + "type": "inline_equation", + "content": "90.8\\%" + }, + { + "bbox": [ + 55, + 438, + 296, + 643 + ], + "type": "text", + "content": " accuracy in human validation on 4k randomly sampled data, which is higher than that reported in previous synthetic data methods [8, 24]. Notably, the ChatGPT-4o-latest model, which we use to generate questions, demonstrates a low accuracy rate (less than " + }, + { + "bbox": [ + 55, + 438, + 296, + 643 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 438, + 296, + 643 + ], + "type": "text", + "content": ") in solving them. This highlights the effectiveness of our ToolEngine pipeline in decoupling complex reasoning into a series of simpler, single-step tasks. While ChatGPT-4o-latest excels in handling " + }, + { + "bbox": [ + 55, + 438, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\to \\mathcal{P} \\to \\mathcal{Q}, \\mathcal{A}" + }, + { + "bbox": [ + 55, + 438, + 296, + 643 + ], + "type": "text", + "content": " process for straightforward, step-by-step reasoning, it faces challenges with the more challenging end-to-end " + }, + { + "bbox": [ + 55, + 438, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\mathcal{I}, \\mathcal{Q} \\to \\mathcal{P} \\to \\mathcal{A}" + }, + { + "bbox": [ + 55, + 438, + 296, + 643 + ], + "type": "text", + "content": " reasoning. This observation indicates a clear gap between LFM's single-step and multi-step reasoning capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 159, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 159, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 159, + 659 + ], + "type": "text", + "content": "3.3. Dataset Statistics" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Tab. 2 presents the basic statistics of ToolVQA, which encompasses various question types (with both text and image as answers) and complex reasoning processes (with an average reasoning trajectory length of 2.78 steps). Notably," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "ToolVQA's questions and answers are concisely formulated (with an average token count of 15.74 for questions and 2.69 for answers). In contrast, previous synthetic datasets [8, 24] often fail to control answer length, leading to hallucinated content within the answers, which is detrimental to training. Fig. 5 demonstrates that our tool frequency aligns with real-world user needs, reinforcing the real-world setting of our dataset. Fig. 6 presents the distribution of reasoning trajectory lengths in ToolVQA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 188, + 448, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 448, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 448, + 201 + ], + "type": "text", + "content": "3.4. Ablation on ToolEngine" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 206, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 555, + 529 + ], + "type": "text", + "content": "To verify whether ToolEngine truly improves the quality of synthetic data, we conducted an ablation study on different synthetic methods. We invite 10 users to evaluate 1k randomly sampled data based on four key metrics: (1) answer correctness (Acc.); (2) correlation (Cor.): whether each sample's image content and Q-A pair are aligned; (3) tool redundancy (Nec.): whether each tool call is necessary for obtaining the answer; and (4) reasoning complexity (R.C.): the average reasoning depth for each task (same as defined in Tab. 1). The results, presented in Tab. 3, show that Real-World Example and LCS matching significantly improved Acc., Nec., and R.C., which highlights their role in the DFS-based simulation of reasoning chains: (1) Real-World Example and LCS matching introduce prior knowledge aligned with image content, enhancing both answer accuracy and tool necessity. (2) Real-World Example and LCS matching incorporate different types of knowledge at each step, increasing the reasoning complexity. Furthermore, the absence of LCS matching has a significant impact on data quality (e.g. Acc. dropped from 90.8 to 41.6), indicating that a fixed set of examples cannot adapt to different scenarios. Dynamically matching appropriate examples is crucial for significantly enhancing the generalization ability of the synthetic pipeline. However, both methods had a limited impact on Cor., likely because the existing controller already effectively captures image content, minimizing errors caused by misleading examples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 541, + 395, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 395, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 395, + 554 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 561, + 460, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 460, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 460, + 574 + ], + "type": "text", + "content": "4.1. Training and Test Settings" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 578, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 628 + ], + "type": "text", + "content": "Training Objective. In our dataset " + }, + { + "bbox": [ + 313, + 578, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 578, + 555, + 628 + ], + "type": "text", + "content": ", for each datapoint " + }, + { + "bbox": [ + 313, + 578, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = (\\mathcal{I},\\mathcal{T},\\mathcal{Q},\\mathcal{A},\\mathcal{P})" + }, + { + "bbox": [ + 313, + 578, + 555, + 628 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 313, + 578, + 555, + 628 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 578, + 555, + 628 + ], + "type": "text", + "content": " rounds of dialogue (" + }, + { + "bbox": [ + 313, + 578, + 555, + 628 + ], + "type": "inline_equation", + "content": "n - 1" + }, + { + "bbox": [ + 313, + 578, + 555, + 628 + ], + "type": "text", + "content": " rounds for tool usage and one round for final answer), we train our tool agent using cross-entropy loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 635, + 545, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 635, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 321, + 635, + 545, + 669 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\mathcal {E} \\sim \\mathcal {D}} \\left[ \\frac {1}{n} \\sum_ {i = 1} ^ {n} - \\log p \\left(t _ {i}, a _ {i}, r _ {i} \\mid \\mathcal {I}, \\mathcal {T}, \\mathcal {Q}, \\mathcal {P} _ {i - 1}\\right) \\right]", + "image_path": "123025adb8cb1fad84e1eafd338a5ad4a0119ac12a0b208d4753e3c4816f7b7b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Training Setting. We divide ToolVQA into training and test sets: the training set consists of 21,105 automatically generated samples (with a human verification accu" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4428" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 81, + 197, + 182 + ], + "blocks": [ + { + "bbox": [ + 80, + 81, + 197, + 182 + ], + "lines": [ + { + "bbox": [ + 80, + 81, + 197, + 182 + ], + "spans": [ + { + "bbox": [ + 80, + 81, + 197, + 182 + ], + "type": "table", + "html": "
StatisticsNumber
Samples23655
Tool Calls65785
Text answers15806
Image answers7849
Query length15.74
Answer length2.69
Trajectory length2.78
", + "image_path": "f0581da53ba6bddb739037969a1ab2b5331473d60be5e1f629e1e41f57b47206.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 201, + 77, + 353, + 193 + ], + "blocks": [ + { + "bbox": [ + 201, + 77, + 353, + 193 + ], + "lines": [ + { + "bbox": [ + 201, + 77, + 353, + 193 + ], + "spans": [ + { + "bbox": [ + 201, + 77, + 353, + 193 + ], + "type": "image", + "image_path": "189e4ff5e10624ab50756d59c0f851d89545143760aef97fe169224840ee0a87.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 201, + 323, + 213 + ], + "lines": [ + { + "bbox": [ + 229, + 201, + 323, + 213 + ], + "spans": [ + { + "bbox": [ + 229, + 201, + 323, + 213 + ], + "type": "text", + "content": "Figure 5. Tool frequency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 356, + 77, + 535, + 186 + ], + "blocks": [ + { + "bbox": [ + 356, + 77, + 535, + 186 + ], + "lines": [ + { + "bbox": [ + 356, + 77, + 535, + 186 + ], + "spans": [ + { + "bbox": [ + 356, + 77, + 535, + 186 + ], + "type": "image", + "image_path": "56531ecff38782fea97298190c0d2befefb583f0c22918bd695d327e26b18ea7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 195, + 517, + 207 + ], + "lines": [ + { + "bbox": [ + 371, + 195, + 517, + 207 + ], + "spans": [ + { + "bbox": [ + 371, + 195, + 517, + 207 + ], + "type": "text", + "content": "Figure 6. Trajectory length distribution." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 60, + 232, + 291, + 295 + ], + "blocks": [ + { + "bbox": [ + 93, + 190, + 182, + 201 + ], + "lines": [ + { + "bbox": [ + 93, + 190, + 182, + 201 + ], + "spans": [ + { + "bbox": [ + 93, + 190, + 182, + 201 + ], + "type": "text", + "content": "Table 2. Basic statistics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 232, + 291, + 295 + ], + "lines": [ + { + "bbox": [ + 60, + 232, + 291, + 295 + ], + "spans": [ + { + "bbox": [ + 60, + 232, + 291, + 295 + ], + "type": "table", + "html": "
MethodAcc.↑Cor. ↑Nec. ↑R.C. ↑
ToolEngine90.885.287.512.38
-w/o Example + LCS27.377.621.041.1
-w/o LCS41.681.454.261.61
", + "image_path": "6dbb94790b3c2f0c17bbf43dadc1a1f67c09e5635948635e6a99e1743811bcb1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 302, + 276, + 313 + ], + "lines": [ + { + "bbox": [ + 74, + 302, + 276, + 313 + ], + "spans": [ + { + "bbox": [ + 74, + 302, + 276, + 313 + ], + "type": "text", + "content": "Table 3. Ablation study on different synthetic methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 337, + 295, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 337, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 54, + 337, + 295, + 372 + ], + "type": "text", + "content": "racy of " + }, + { + "bbox": [ + 54, + 337, + 295, + 372 + ], + "type": "inline_equation", + "content": "90.8\\%" + }, + { + "bbox": [ + 54, + 337, + 295, + 372 + ], + "type": "text", + "content": " ), while the test set contains 2,550 humanreannotated samples. We fine-tune LLaVA-7B [21] in the Lego Agent framework [15]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "type": "text", + "content": "Test Setting. We apply two test modes: end-to-end and step-by-step modes, to comprehensively evaluate LFMs' performance under different settings. The end-to-end mode only focuses on the end-to-end VQA ability of LFMs. Formally, we evaluate the " + }, + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{I} + \\mathcal{Q}\\rightarrow \\mathcal{A}" + }, + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "type": "text", + "content": " process. The step-by-step mode focuses on evaluating LM's tool-usage capabilities, formally the " + }, + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{I} + \\mathcal{Q} + \\mathcal{P}_i\\to \\mathcal{P}_{i + 1}" + }, + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "type": "text", + "content": " process. We use four metrics [39]: (1) InstAcc (Inst.): the percentage of steps executed without errors; (2) ToolAcc (Tool.): the accuracy of tool selection; (3) ArgAcc (Arg.): the accuracy of tool's argument name prediction, and (4) SummAcc (Summ.): the accuracy of final answer's summary. Previous work has focused on the research of large language model (LLM)-based tool agents while ignoring vision-language model (VLM)-based tool agents and the end-to-end VQA abilities of VLMs. To bridge these gaps, our test involves three settings: (1) VLM, directly answering questions end-to-end, formally " + }, + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} + \\mathcal{I}\\rightarrow \\mathcal{A}" + }, + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "type": "text", + "content": ", (2) VLM+tool, calling the tool through VLM, formally " + }, + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} + \\mathcal{I} + \\mathcal{T}\\rightarrow \\mathcal{P} + \\mathcal{A}" + }, + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "type": "text", + "content": ", (3) LLM+tool, calling the tool through LLM, formally " + }, + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} + \\mathcal{T}\\rightarrow \\mathcal{P} + \\mathcal{A}" + }, + { + "bbox": [ + 56, + 374, + 296, + 637 + ], + "type": "text", + "content": " (Note that image information can still be obtained through tools such as caption under this setting)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 141, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 141, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 141, + 659 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Scale. As shown in Tab. 4, the tool-use capabilities of LFMs generally improve with the increase in model parameters. For example, Qwen2-7B-instruct surpasses Qwen2-2B-instruct by " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "7.03\\%" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", and closed-sourced LM (larger than" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 234, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 553, + 293 + ], + "type": "text", + "content": "100B) is generally better than open-sourced (2B,7B). After fine-tuning, our 7B model's performance (last 3 lines) is close to the large scale LFM GPT-3.5-Turbo. This demonstrates that fine-tuning significantly enhances the model's real-world tool-use capabilities, even in smaller models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 294, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 510 + ], + "type": "text", + "content": "Settings. By comparing the performance of the same model under different settings (VLM, VLM+tool, and LLM+tool) in Tab. 4, we identify several interesting observations. (1) In models such as GPT-4o, Qwen2-7B/2B, and LLaVA-7B, the VLM+tool setting performs worse than the LLM+tool setting, suggesting that the visual modules in current LFM architectures are still ineffective at guiding tool usage. (2) For models including GPT-4o, Claude-3.5, LLaVA-7B, and Qwen2-2B, we find that VLM+tool underperforms VLM alone. This indicates that the noise introduced by tool usage can outweigh the benefits brought by the tools. However, the tuned LLaVA model demonstrates a clear progression: VLM " + }, + { + "bbox": [ + 313, + 294, + 555, + 510 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 313, + 294, + 555, + 510 + ], + "type": "text", + "content": " LLM+tool " + }, + { + "bbox": [ + 313, + 294, + 555, + 510 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 313, + 294, + 555, + 510 + ], + "type": "text", + "content": " VLM+tool, suggesting that fine-tuning can indeed enhance the model's ability to make effective use of tools while suppressing tool-induced noise. Nevertheless, the similar performance of the LLM+tool and VLM+tool settings implies that the visual module's potential in facilitating tool use has not yet been fully explored." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Bottlenecks. The last two rows of Tab. 4 indicate that the performance of fine-tuned models is often constrained by their weakest capabilities, which are typically argument prediction and answer summarization. While fine-tuning significantly improves instruction formatting and tool selection, it makes little progress in weaker capabilities. Specifically, instruction formatting and tool selection can be accomplished through memorizing patterns, as the number of reasonable tool invocation types in real-world scenarios is limited. However, argument prediction and answer summarization require the model to genuinely understand the new information returned by tools and extract meaningful responses from it. This information varies across different scenarios and queries, and even within multi-turn dialogues of the same scenario. Consequently, a high level of generalization capability in vision-language joint reasoning is essential to handle these issues in tool-use tasks." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4429" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 70, + 518, + 395 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 518, + 395 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 518, + 395 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 518, + 395 + ], + "type": "table", + "html": "
ModelSettingEnd-to-End ModeStep-by-Step Mode
Acc. (%)Tool Call (Times)Tool Error (Times)Inst. (%)Tool. (%)Arg. (%)Summ. (%)
Closed-source LFMs
ChatGPT-4o-latestVLM38.29------
VLM + tool34.962003180436.514.688.9256.1
LLM + tool37.624687196273.0241.1225.2765.45
Claude-3-5-sonnetVLM30.33------
VLM + tool30.5258552445.7116.947.3646.06
LLM + tool24.48183467266.8339.2913.9875.52
GPT-3.5-TurboLLM + tool18.373178132173.2430.4620.0858.18
Open-source LFMs
Qwen2-VL-7B-InstructVLM3.64------
VLM + tool5.861373105135.129.021.0938.18
LLM + tool6.781366108939.4611.23.5544.85
Qwen2-7B-InstructLLM + tool11.53177370363.124.458.269.7
LLaVA-v1.5-7BVLM8.57------
VLM + tool1.179684968416.399.4300.01
LLM + tool1.6610421104219.2515.03011.52
LLaMA-3-8b-instructLLM + tool4.936607654251.6314.130.5341.99
Qwen2-VL-2B-InstructVLM7.71------
VLM + tool2.14067391547.1612.70.4121.82
LLM + tool2.226892665536.3410.520.6832.12
Qwen2-1.5B-InstructLLM + tool4.57933614150.9521.454.5144.24
Tuned LLaVA-7B (Ours)VLM7.21------
VLM + tool18.8431161786.6261.6139.3430.91
LLM + tool18.43446268087.5162.340.5726.67
", + "image_path": "3463d219daf11aab206bcb08441f1991560f27ef681e452b977cc2ddf70f8403.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 233, + 402, + 376, + 414 + ], + "lines": [ + { + "bbox": [ + 233, + 402, + 376, + 414 + ], + "spans": [ + { + "bbox": [ + 233, + 402, + 376, + 414 + ], + "type": "text", + "content": "Table 4. Results on ToolVQA's test set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 434, + 231, + 446 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 231, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 231, + 446 + ], + "type": "text", + "content": "4.3. Out-of-Distribution Benchmarks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "type": "text", + "content": "To test the generalization capability of our tuned model, we conduct experiments on out-of-distribution (OOD) benchmarks including (1) Seen tools, unseen queries: TextVQA [36], TallyQA [1], InfoSeek [4]; and (2) Unseen tools: GTA [39], TEMPLAMA [6]. The results are shown in the Tab. 5. Our fine-tuned model significantly outperforms the baseline model LLaVA-7B by " + }, + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "type": "inline_equation", + "content": "21.17\\%" + }, + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "type": "inline_equation", + "content": "18.37\\%" + }, + { + "bbox": [ + 55, + 454, + 296, + 634 + ], + "type": "text", + "content": " accuracy on these datasets, showing excellent generalizability. Its performance also surpasses the large closed-source model GPT-3.5-Turbo on unseen queries. However, its performance on TEMPLAMA is inferior to GPT-3.5-Turbo, possibly because GPT encountered similar data during its training (due to its great coverage of extensive training data), whereas our model had no exposure to such data at all." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 647, + 217, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 217, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 217, + 659 + ], + "type": "text", + "content": "4.4. Few-shot In-context Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Few-shot in-context learning (ICL) has become a common and important technique in tool-use tasks. However, prior model fine-tuning work often overlooks evaluating the ICL capabilities of fine-tuned models. We believe that fine-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 435, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 591 + ], + "type": "text", + "content": "tuning and ICL can be complementary, offering significant potential for improvement. Tab. 6 shows few-shot ICL performance on ToolVQA. GPT-4o steadily improves with more shots, while GPT-3.5 and LLaVA both show limited gains and even decline slightly at 10-shot. This may be due to long contexts impairing the LLM's ability to interpret precise tool-use instructions. Fine-tuned LLaVA outperforms the baseline and continues to benefit from ICL. These results suggest that (1) compared to fine-tuning, ICL's improvement is limited and affected by context lengths, and (2) fine-tuned models can still benefit from ICL, highlighting the complementary value of these two approaches in enhancing tool-use performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 600, + 408, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 408, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 408, + 613 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "To further evaluate the quality of the fine-tuning data from multiple perspectives, we conduct an ablation study of the different training strategies. As shown in Tab. 7, removing the ImageCaption tool significantly degrades both accuracy and parameter correctness. This highlights the model's reliance on high-quality captions for initial scene understanding, likely due to its limited inherent visual capability for fine-grained information extraction. When all tools" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4430" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 99, + 70, + 512, + 133 + ], + "blocks": [ + { + "bbox": [ + 99, + 70, + 512, + 133 + ], + "lines": [ + { + "bbox": [ + 99, + 70, + 512, + 133 + ], + "spans": [ + { + "bbox": [ + 99, + 70, + 512, + 133 + ], + "type": "table", + "html": "
ModelTextVQA [36]TallyQA [1]InfoSeek [4]GTA [39]TEMPLAMA [6]
GPT-3.5-Turbo36.36111.323.6233.67
LLaVA-7B41.260.15.212.123.06
Tuned LLaVA-7B4764.313.833.2921.43
", + "image_path": "d894eae8e6fb7369d8d684ebd1d9e0c2974510f893e741b0d0001e291cdce705.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 162, + 297, + 232 + ], + "blocks": [ + { + "bbox": [ + 233, + 140, + 376, + 150 + ], + "lines": [ + { + "bbox": [ + 233, + 140, + 376, + 150 + ], + "spans": [ + { + "bbox": [ + 233, + 140, + 376, + 150 + ], + "type": "text", + "content": "Table 5. Results on OOD benchmarks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 162, + 297, + 232 + ], + "lines": [ + { + "bbox": [ + 59, + 162, + 297, + 232 + ], + "spans": [ + { + "bbox": [ + 59, + 162, + 297, + 232 + ], + "type": "table", + "html": "
ModelZero-shot1-shot5-shot10-shot
GPT-4o34.9637.2038.4138.63
GPT-3.5-Turbo18.3719.8320.4019.04
LLaVA-7B1.173.454.273.67
Tuned LLaVA-7B18.8019.4121.1320.69
", + "image_path": "f650cf0d5a8ebf989e7ac41b90634206fe7c1992e73fa766f1b94c15acc6b25c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 306, + 162, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 306, + 162, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 306, + 162, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 306, + 162, + 553, + 232 + ], + "type": "table", + "html": "
ModelSettingAcc.Tool CallTool Error
Tuned LLaVA-7BVLM + tool18.84311617
-w/o CaptionVLM + tool11.11861192
-w/o PVLM8.01--
-w/o P + ILLM0.18--
", + "image_path": "0193ed63fdc784f5658d82a55dc108d2183f03a0ac1525b5e992f342297a3096.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 238, + 249, + 249 + ], + "lines": [ + { + "bbox": [ + 93, + 238, + 249, + 249 + ], + "spans": [ + { + "bbox": [ + 93, + 238, + 249, + 249 + ], + "type": "text", + "content": "Table 6. Few-shot performance on ToolVQA." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 356, + 238, + 500, + 248 + ], + "lines": [ + { + "bbox": [ + 356, + 238, + 500, + 248 + ], + "spans": [ + { + "bbox": [ + 356, + 238, + 500, + 248 + ], + "type": "text", + "content": "Table 7. Ablation study on training methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 81, + 262, + 526, + 417 + ], + "blocks": [ + { + "bbox": [ + 81, + 262, + 526, + 417 + ], + "lines": [ + { + "bbox": [ + 81, + 262, + 526, + 417 + ], + "spans": [ + { + "bbox": [ + 81, + 262, + 526, + 417 + ], + "type": "image", + "image_path": "e17e61b0d7afdc57519a0e7fc829a3c3f1fd9106b1f61b08059026f3baf96e63.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 430, + 432, + 441 + ], + "lines": [ + { + "bbox": [ + 176, + 430, + 432, + 441 + ], + "spans": [ + { + "bbox": [ + 176, + 430, + 432, + 441 + ], + "type": "text", + "content": "Figure 7. Visualization of the main error types made by tuned LLaVA." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 463, + 295, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 535 + ], + "type": "text", + "content": "are removed and ToolVQA is reduced to a single-hop VQA model, performance drops substantially, underscoring the necessity of multi-step tool usage. In the absence of both tools and images, performance approaches zero, indicating that the model is heavily dependent on visual input when answering questions in ToolVQA." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 540, + 148, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 148, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 148, + 553 + ], + "type": "text", + "content": "4.6. Error Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 296, + 714 + ], + "type": "text", + "content": "In Fig. 7, we randomly select 100 errors made by tuned LLaVA, and manually count and classify the main error types. The two main error types are: (1) Argument prediction error, for example, missing essential keywords \"age\" during a search. (2) Answer summary error, for example, extracting incorrect information \"1,300\" from tool outputs. The common cause of these errors is the model's failure to recognize key terms (e.g. \"age\" and \"pound\") in the tool outputs, even though this should not be a challenging task for a language model. This suggests that the model still struggles with dynamically adapting to additional information provided by tools. Since multi-step reasoning often suffers from error accumulation (early errors propagate to sub" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 464, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 555, + 523 + ], + "type": "text", + "content": "sequent steps), errors in processing tool results at any step will affect the results. Fine-tuning has limited improvement on the capability for dynamically processing tool results, but ToolVQA still provides a valuable benchmark for evaluating this capability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": "We present ToolVQA, a large multimodal dataset for real-world tool-augmented reasoning. Using our ToolEngine pipeline, we generate 23K real-world samples with implicit multi-step reasoning over 10 tools and 7 domains. This dataset challenges existing models beyond synthetic settings and better reflects real user needs. Experiments show that LLaVA-7B fine-tuned on ToolVQA not only excels on the in-domain test set but also outperforms GPT-3.5-turbo on five OOD benchmarks, establishing a new state-of-the-art for open-source models in complex tool-use VQA. ToolVQA serves as both a benchmark and a training ground for developing more capable, generalizable tool-using agents." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4431" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the grants from the National Natural Science Foundation of China 62372014, Beijing Nova Program and Beijing Natural Science Foundation 4252040." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "spans": [ + { + "bbox": [ + 56, + 133, + 115, + 146 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 153, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 153, + 296, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 153, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 153, + 296, + 186 + ], + "type": "text", + "content": "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions, 2018. 2, 3, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 189, + 296, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 189, + 296, + 221 + ], + "spans": [ + { + "bbox": [ + 61, + 189, + 296, + 221 + ], + "type": "text", + "content": "[2] Aishwarya Agrawal, Jiasen Lu, Stanislaw Antol, Margaret Mitchell, C. Lawrence Zitnick, Dhruv Batra, and Devi Parikh. Vqa: Visual question answering, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 223, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 223, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 61, + 223, + 295, + 255 + ], + "type": "text", + "content": "[3] Lluis Castrejon, Thomas Mensink, Howard Zhou, Vittorio Ferrari, Andre Araujo, and Jasper Uijlings. Hammr: Hierarchical multimodal react agents for generic vqa, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 258, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 258, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 61, + 258, + 295, + 302 + ], + "type": "text", + "content": "[4] Yang Chen, Hexiang Hu, Yi Luan, Haitian Sun, Soravit Changpinyo, Alan Ritter, and Ming-Wei Chang. Can pre-trained vision and language models answer visual information-seeking questions?, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 304, + 296, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 304, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 61, + 304, + 296, + 347 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 350, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 296, + 403 + ], + "type": "text", + "content": "[6] Bhuwan Dhingra, Jeremy R. Cole, Julian Martin Eisenschlos, Daniel Gillick, Jacob Eisenstein, and William W. Cohen. Time-aware language models as temporal knowledge bases. Transactions of the Association for Computational Linguistics, 10:257-273, 2022. 2, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 449 + ], + "type": "text", + "content": "[7] Difei Gao, Lei Ji, Luowei Zhou, Kevin Qinghong Lin, Joya Chen, Zihan Fan, and Mike Zheng Shou. Assistgpt: A general multi-modal assistant that can plan, execute, inspect, and learn, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 452, + 296, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 452, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 452, + 296, + 495 + ], + "type": "text", + "content": "[8] Zhi Gao, Bofei Zhang, Pengxiang Li, Xiaojian Ma, Tao Yuan, Yue Fan, Yuwei Wu, Yunde Jia, Song-Chun Zhu, and Qing Li. Multi-modal agent tuning: Building a vlm-driven agent for efficient tool usage, 2025. 1, 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 498, + 296, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 296, + 530 + ], + "type": "text", + "content": "[9] Shibo Hao, Tianyang Liu, Zhen Wang, and Zhiting Hu. Toolkengpt: Augmenting frozen language models with massive tools via tool embeddings, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 586 + ], + "type": "text", + "content": "[10] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models, 2024. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 589, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 589, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 589, + 295, + 633 + ], + "type": "text", + "content": "[11] Ziniu Hu, Ahmet Iscen, Chen Sun, Kai-Wei Chang, Yizhou Sun, David A Ross, Cordelia Schmid, and Alireza Fathi. Avis: Autonomous visual information seeking with large language model agent, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[12] Yue Huang, Jiawen Shi, Yuan Li, Chenrui Fan, Siyuan Wu, Qihui Zhang, Yixin Liu, Pan Zhou, Yao Wan, Neil Zhenqiang Gong, and Lichao Sun. Metatool benchmark for large language models: Deciding whether to use tools and which to use, 2024. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[13] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 152 + ], + "type": "text", + "content": "[14] Faria Huq, Zora Zhiruo Wang, Frank F. Xu, Tianyue Ou, Shuyan Zhou, Jeffrey P. Bigham, and Graham Neubig. Cow-pilot: A framework for autonomous and human-agent collaborative web navigation, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 153, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 185 + ], + "type": "text", + "content": "[15] InternLM. Agentlego: Enhance llm agents with rich tool apis. https://github.com/InternLM/agentlego, 2024.6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 187, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 187, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 187, + 553, + 273 + ], + "type": "text", + "content": "[16] Iat Long Iong, Xiao Liu, Yuxuan Chen, Hanyu Lai, Shuntian Yao, Pengbo Shen, Hao Yu, Yuxiao Dong, and Jie Tang. OpenWebAgent: An open toolkit to enable web agents on large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 72-81, Bangkok, Thailand, 2024. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "text", + "content": "[17] Jing Yu Koh, Robert Lo, Lawrence Jang, Vikram Duvvur, Ming Chong Lim, Po-Yu Huang, Graham Neubig, Shuyan Zhou, Ruslan Salakhutdinov, and Daniel Fried. Visualwebarena: Evaluating multimodal agents on realistic visual web tasks, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 332, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 396 + ], + "type": "text", + "content": "[18] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 453 + ], + "type": "text", + "content": "[19] Minghao Li, Feifan Song, Yu Bowen, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. Api-bank: A comprehensive benchmark for tool-augmented llms. In Conference on Empirical Methods in Natural Language Processing, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 455, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 553, + 521 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 522, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 544 + ], + "type": "text", + "content": "[21] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 545, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 599 + ], + "type": "text", + "content": "[22] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In The 37th Conference on Neural Information Processing Systems (NeurIPS), 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "text", + "content": "[23] Kaixin Ma, Hongming Zhang, Hongwei Wang, Xiaoman Pan, Wenhao Yu, and Dong Yu. Laser: LIm agent with statespace exploration for web navigation, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 677 + ], + "type": "text", + "content": "[24] Zixian Ma, Weikai Huang, Jieyu Zhang, Tanmay Gupta, and Ranjay Krishna. m&m's: A benchmark to evaluate tool-use for multi-step multi-modal tasks. EECV 2024, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "text", + "content": "[25] Arjun Mani, Nobline Yoo, Will Hinthorn, and Olga Russakovsky. Point and ask: Incorporating pointing into visual question answering, 2022. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4432" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[26] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[27] Grégoire Mialon, Clémentine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, and Thomas Scialom. Gaia: a benchmark for general ai assistants, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 175 + ], + "type": "text", + "content": "[28] Shishir G Patil, Tianjun Zhang, Xin Wang, and Joseph E Gonzalez. Gorilla: Large language model connected with massive apis. arXiv preprint arXiv:2305.15334, 2023. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 230 + ], + "type": "text", + "content": "[29] Yun Peng, Shuqing Li, Wenwei Gu, Yichen Li, Wenxuan Wang, Cuiyun Gao, and Michael R. Lyu. Revisiting, benchmarking and exploring api recommendation: How far are we? IEEE Transactions on Software Engineering, 49:1876-1897, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "type": "text", + "content": "[30] Yujia Qin, Shihao Liang, Yining Ye, Kunlun Zhu, Lan Yan, Yaxi Lu, Yankai Lin, Xin Cong, Xiangru Tang, Bill Qian, Sihan Zhao, Runchu Tian, Ruobing Xie, Jie Zhou, Mark Gerstein, Dahai Li, Zhiyuan Liu, and Maosong Sun. Toollm: Facilitating large language models to master 16000+ real-world apis, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 408 + ], + "type": "text", + "content": "[31] Yujia Qin, Shengding Hu, Yankai Lin, Weize Chen, Ning Ding, Ganqu Cui, Zheni Zeng, Yufei Huang, Chaojun Xiao, Chi Han, Yi Ren Fung, Yusheng Su, Huadong Wang, Cheng Qian, Runchu Tian, Kunlun Zhu, Shihao Liang, Xingyu Shen, Bokai Xu, Zhen Zhang, Yining Ye, Bowen Li, Ziwei Tang, Jing Yi, Yuzhang Zhu, Zhenning Dai, Lan Yan, Xin Cong, Yaxi Lu, Weilin Zhao, Yuxiang Huang, Junxi Yan, Xu Han, Xian Sun, Dahai Li, Jason Phang, Cheng Yang, Tongshuang Wu, Heng Ji, Zhiyuan Liu, and Maosong Sun. Tool learning with foundation models, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "text", + "content": "[32] Yangjun Ruan, Honghua Dong, Andrew Wang, Silviu Pitis, Yongchao Zhou, Jimmy Ba, Yann Dubois, Chris J. Maddison, and Tatsunori Hashimoto. Identifying the risks of lm agents with an lm-emulated sandbox, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "text", + "content": "[33] Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "text", + "content": "[34] Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 588 + ], + "type": "text", + "content": "[35] Yongliang Shen, Kaitao Song, Xu Tan, Wenqi Zhang, Kan Ren, Siyu Yuan, Weiming Lu, Dongsheng Li, and Yueting Zhuang. Taskbench: Benchmarking large language models for task automation, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[36] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read, 2019. 2, 3, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 679 + ], + "type": "text", + "content": "[37] Dídac Surís, Sachit Menon, and Carl Vondrick. Vipergpt: Visual inference via python execution for reasoning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11888-11898, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[38] Chenyu Wang, Weixin Luo, Qianyu Chen, Haonan Mai, Jindi Guo, Sixun Dong, Xiaohua (Michael) Xuan, Zhengxin Li, Lin Ma, and Shenghua Gao. Mllm-tool: A multimodal" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 578 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "large language model for tool agent learning. arXiv preprint arXiv:2401.10727, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[39] Jize Wang, Zerun Ma, Yining Li, Songyang Zhang, Cailian Chen, Kai Chen, and Xinyi Le. Gta: A benchmark for general tool agents, 2024. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "text", + "content": "[40] Yan Wang, Yawen Zeng, Jingsheng Zheng, Xiaofen Xing, Jin Xu, and Xiangmin Xu. Videocot: A video chain-of-thought dataset with active annotation tool, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 164, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 186 + ], + "type": "text", + "content": "[41] Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. Agent workflow memory, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 230 + ], + "type": "text", + "content": "[42] Shirley Wu, Shiyu Zhao, Qian Huang, Kexin Huang, Michihiro Yasunaga, Kaidi Cao, Vassilis N. Ioannidis, Karthik Subbian, Jure Leskovec, and James Zou. Avatar: Optimizing llm agents for tool usage via contrastive reasoning, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 232, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 297 + ], + "type": "text", + "content": "[43] Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu. Osworld: Benchmarking multimodal agents for open-ended tasks in real computer environments, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 342 + ], + "type": "text", + "content": "[44] Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 344, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 388 + ], + "type": "text", + "content": "[45] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[46] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent), 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 433, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 466 + ], + "type": "text", + "content": "[47] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "text", + "content": "[48] Shuyan Zhou, Frank F. Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, Uri Alon, and Graham Neubig. Webarena: A realistic web environment for building autonomous agents, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 544 + ], + "type": "text", + "content": "[49] Yuke Zhu, Oliver Groth, Michael Bernstein, and Li Fei-Fei. Visual7w: Grounded question answering in images, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "text", + "content": "[50] Yuchen Zhuang, Yue Yu, Kuan Wang, Haotian Sun, and Chao Zhang. Toolqa: A dataset for llm question answering with external tools, 2023. 2, 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4433" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/ddd3f9c4-d03d-4cd7-96eb-1ef2931311e2_content_list.json b/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/ddd3f9c4-d03d-4cd7-96eb-1ef2931311e2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b80b8f425ac470e34f26683d92fb72a69604043f --- /dev/null +++ b/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/ddd3f9c4-d03d-4cd7-96eb-1ef2931311e2_content_list.json @@ -0,0 +1,2269 @@ +[ + { + "type": "text", + "text": "Top2Pano: Learning to Generate Indoor Panoramas from Top-Down View", + "text_level": 1, + "bbox": [ + 122, + 130, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zitong Zhang Suranjan Gautam Rui Yu University of Louisville", + "bbox": [ + 303, + 181, + 692, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zitong.zhang, suranjan.gautam, rui.yu}@louisville.edu", + "bbox": [ + 259, + 219, + 736, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://top2pano.github.io/", + "bbox": [ + 375, + 237, + 616, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/87224dcd1f0f73f69582edfd7f31a0c55d8cd50f55ccf841fd2f52435aa16045.jpg", + "image_caption": [ + "Top-down view" + ], + "image_footnote": [], + "bbox": [ + 91, + 261, + 166, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/735e1e120e15d3c20da8305349706970f7009c55167331b92e15b05c5aac50f6.jpg", + "image_caption": [ + "Top2Pano" + ], + "image_footnote": [], + "bbox": [ + 163, + 289, + 269, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ec1b6120b05d374645c2eecb4d9b4651a0606aac0e8d229920a75de6a7588942.jpg", + "image_caption": [ + "Output 1" + ], + "image_footnote": [], + "bbox": [ + 271, + 261, + 480, + 343 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/725d35ef4e56b533ccc99ba1b5887ba31087e881a160f35a9ac35bd23c67f706.jpg", + "image_caption": [ + "Output 2" + ], + "image_footnote": [], + "bbox": [ + 482, + 262, + 691, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff0842f581694a85ce859498055eee298b328dabe8060f2a95f1f2a14f845357.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 694, + 262, + 903, + 344 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aeb079062defbd1466a8935594ae1bb340a8600929d40752fdf9aea4c059d5ba.jpg", + "image_caption": [ + "Floorplan", + "Figure 1. Top: We present Top2Pano, a method for synthesizing high-quality indoor panoramas from a top-down view. Given a camera position, Top2Pano generates panoramas that are both visually compelling and geometrically accurate. Bottom: Our model demonstrates strong generalization capabilities. When provided with schematic floor plans as input, Top2Pano produces photorealistic and structurally coherent panoramas. Additionally, our approach can be easily adapted for stylized synthesis, enabling diverse design variations. Note: The original dataset (Matterport3D [3]) contains blurry regions near the upper and lower edges of the panoramic image." + ], + "image_footnote": [], + "bbox": [ + 91, + 369, + 269, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92074427435354c982e115ec357e4ef9b8aeeb3fbd374e657c29893892d2ee9c.jpg", + "image_caption": [ + "Default" + ], + "image_footnote": [], + "bbox": [ + 269, + 359, + 478, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/165a518e082ea816a5ffe3c95a493f513fe530e02f5812876f6926f2b2e32f23.jpg", + "image_caption": [ + "\"Chinese Style\"" + ], + "image_footnote": [], + "bbox": [ + 480, + 359, + 691, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/205b4f0aa706d48aba89f5609ed437fe16f30b53af5bd772c92b61d0a7e623d1.jpg", + "image_caption": [ + "\"Stone Wall\"" + ], + "image_footnote": [], + "bbox": [ + 694, + 359, + 903, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 541, + 326, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating immersive $360^{\\circ}$ indoor panoramas from 2D top-down views has applications in virtual reality, interior design, real estate, and robotics. This task is challenging due to the lack of explicit 3D structure and the need for geometric consistency and photorealism. We propose Top2Pano, an end-to-end model for synthesizing realistic indoor panoramas from top-down views. Our method estimates volumetric occupancy to infer 3D structures, then uses volumetric rendering to generate coarse color and depth panoramas. These guide a diffusion-based refinement stage using ControlNet, enhancing realism and structural fidelity. Evaluations on two datasets show Top2Pano outperforms baselines, effectively reconstructing geometry, occlusions, and spatial arrangements. It also generalizes well, producing high-quality panoramas from schematic floorplans. Our results highlight Top2Pano's potential in bridging top-down views with immersive indoor synthesis.", + "bbox": [ + 88, + 571, + 485, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 844, + 220, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding and synthesizing immersive indoor scenes from minimal structural information is a fundamental chal", + "bbox": [ + 89, + 869, + 483, + 898 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lenge in computer vision and graphics [16, 20, 21, 32, 40, 44]. The ability to generate realistic indoor panorama images from a 2D top-down view holds immense potential for a wide range of applications, including virtual reality (VR) [23], interior design [27], real estate visualization [15], and robotics [11]. For instance, real estate platforms can leverage this technology to offer potential buyers photorealistic virtual walkthroughs generated directly from architectural floorplans, enhancing the property viewing experience. Similarly, VR applications can benefit from automatically synthesized environments that create more engaging and immersive user experiences. Additionally, robots operating in indoor environments can utilize synthesized panoramas to improve their spatial understanding and navigation capabilities, enabling more efficient and accurate movement in complex spaces. Despite its broad applicability, the task of generating high-quality indoor panoramas from top-down views remains surprisingly underexplored in the literature. Recent advancements in large multimodal models have enabled the synthesis of panoramas directly from text input [10, 39, 41]; however, these approaches often overlook critical geometric and textural constraints. Other studies have focused on generating 3D models from semantic layouts [2, 4, 8, 9, 30, 38];", + "bbox": [ + 511, + 542, + 908, + 891 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28493", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/800c677146bc589beee6fcb77f6e408d9d46964197f4e1dcf18574abd727b4a3.jpg", + "image_caption": [ + "(a) Volumetric Occupancy Estimation", + "(b) Coarse Panorama Rendering", + "(c) Photorealistic Panorama Synthesis", + "Figure 2. Overview of the proposed Top2Pano pipeline. The pipeline begins by segmenting the top-down view using SAM [19]. Both the segmented top-down image and the original top-down view are then processed by the OccRecon module to estimate the scene's 3D volumetric occupancy. Next, given the camera position, the system employs volumetric rendering to generate coarse depth and color panoramas. These coarse images are subsequently refined by the PanoGen module to produce the final photorealistic panorama. The PanoGen module also supports stylized panorama generation based on textual or visual conditions." + ], + "image_footnote": [], + "bbox": [ + 94, + 107, + 906, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "however, these methods are often limited by the quality of the resulting 3D meshes, making them unsuitable for rendering high-quality panoramas. Furthermore, semantic information is often unavailable in top-down views or floorplans, posing an additional challenge for existing approaches.", + "bbox": [ + 88, + 385, + 482, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Addressing this gap requires overcoming several significant technical challenges. First, a 2D top-down view provides only limited visual cues about the actual appearance and layout of the scene, making it difficult to infer occluded structures and fine texture details. Second, generating geometrically consistent indoor scenes demands accurate reasoning about 3D spatial occupancy from a 2D input, which is inherently ambiguous. Third, achieving photorealism while maintaining structural coherence necessitates a synthesis approach that effectively balances fidelity and realism, ensuring that the generated scenes are both visually appealing and functionally accurate.", + "bbox": [ + 88, + 462, + 482, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these challenges, we introduce Top2Pano, a novel framework for generating photorealistic indoor panorama images from 2D top-down views. Our approach consists of three main stages. First, we learn the volumetric occupancy of the indoor scene, enabling the model to infer plausible spatial structures and layout configurations. Next, we employ volumetric rendering to generate coarse depth and colored panorama images, providing an initial estimate of the scene's appearance and geometry. Finally, we refine the synthesized panoramas using a diffusion-based model [42] conditioned on the coarse representations, enhancing both realism and structural consistency. By incorporating learned occupancy priors and diffusion-based refinement, our model effectively bridges the gap between schematic top-down views and immersive indoor panoramas, producing results that are both visually compelling and geometrically accurate.", + "bbox": [ + 89, + 643, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate Top2Pano on two indoor datasets and demonstrate its effectiveness compared to baseline methods. Our model not only generates higher-quality images with improved geometric consistency but also exhibits strong generalization capabilities. Even when provided with schematic floorplans as input, Top2Pano can produce photorealistic and structurally coherent panoramas. Moreover, we show that our method can be easily adapted for stylized synthesis, allowing for diverse design variations and enabling users to explore different interior aesthetics with ease.", + "bbox": [ + 511, + 385, + 906, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key contributions are as follows:", + "bbox": [ + 531, + 537, + 779, + 551 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce Top2Pano, a novel framework for generating indoor panoramas from 2D top-down views, integrating volumetric occupancy learning, coarse synthesis, and diffusion-based refinement to achieve high-quality results.", + "- We conduct extensive experiments on two indoor datasets, demonstrating that our model surpasses baseline methods in both image quality and structural consistency, setting a new benchmark for this task.", + "- We show that Top2Pano generalizes well to schematic floorplans, producing high-quality, geometry-consistent panoramas. Furthermore, our approach supports stylish synthesis, enabling the generation of panoramas with diverse interior design aesthetics, making it a versatile tool for various applications." + ], + "bbox": [ + 511, + 551, + 906, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 777, + 653, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Panorama Generation", + "text_level": 1, + "bbox": [ + 513, + 803, + 720, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditionally, panoramas were generated using image stitching and feature matching methods. With the recent advancements in generative machine learning, text-driven panorama generation techniques have gained popularity. These methods have utilized GANs, VAEs or a combina-", + "bbox": [ + 511, + 825, + 908, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28494", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion of GANs and VAEs [5] and more recently, diffusion models [10, 39, 41] to synthesize panoramic images from textual descriptions. Another popular field of research is panorama synthesis from narrow-FOV images using image out-painting. Some methods rely solely on narrow-FOV images [1], while other incorporate textual descriptions alongside the images [6, 17, 34]. Cross-view panorama generation is another well-explored area, particularly challenging due to large shifts in camera perspective. Most techniques in this domain have focused on generating ground-view panoramas from aerial images. Some approaches directly use top-down images as input [35], while others [22, 28, 31, 37] extract geometric and segmentation information from top-down images to enhance quality. To the best of our knowledge, there is no prior work that explores the generation of indoor panoramic images from floor plans or top-down views of indoor spaces.", + "bbox": [ + 89, + 90, + 485, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Layout-Guided 3D Scene Generation", + "text_level": 1, + "bbox": [ + 89, + 357, + 410, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent approaches to 3D scene generation leverage layouts for semantic and physical plausibility. Plan2Scene [33] reconstructs 3D meshes from floor plans, while ATISS [26] employs Transformers conditioned on scene layout. CC3D [2] follows a 3D GAN-based approach using 2D semantic layouts. Diffusion-based methods such as SceneCraft [38], Layout2Scene [4], and Prim2Room [9] have further improved synthesis quality. ControlRoom3D [30] and Ctrl-Room [8] are closely related to our work, as both generate panoramas during reconstruction process. Unlike prior methods that rely on explicit semantic layouts detailing object classes, positions, orientations, and sizes, our approach only requires a top-down view – an easily obtainable and lightweight input format. Instead of generating full 3D scenes, we produce panoramas, offering a more efficient and realistic solution for AR/VR, and autonomous robotic navigation by enabling immersive experiences like virtual tours and supporting real-time robotic navigation.", + "bbox": [ + 89, + 378, + 483, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 662, + 181, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pipeline of our Top2Pano method is illustrated in Figure 2. Given an input top-down view $I_{\\mathrm{top}} \\in \\mathbb{R}^{H \\times W \\times 3}$ , we first generate its segmentation map $I_{\\mathrm{seg}} \\in \\mathbb{R}^{H \\times W \\times 3}$ using a pretrained model. Both the top-down view and its segmentation, along with the specified camera position, are then fed into an encoder-decoder occupancy estimation module, OccRecon, which reconstructs a 3D volumetric occupancy map $V_{\\mathrm{occ}} \\in \\mathbb{R}^{H \\times W \\times N}$ , where $N$ represents the number of vertical voxels. From this occupancy map, we render a coarse depth panorama $I_{\\mathrm{depth}}$ via volumetric rendering and project colors from the top-down view to obtain a coarse color panorama $I_{\\mathrm{color}}$ . To ensure geometric consistency, we enforce structural constraints on walls and floor, preserving occlusion relationships and realistic spatial structure. Finally,", + "bbox": [ + 89, + 688, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "both coarse depth and color panoramas serve as conditions for a diffusion-based synthesis module, PanoGen, which generates photorealistic panoramic images $I_{\\mathrm{pano}} \\in \\mathbb{R}^{H \\times W \\times 3}$ . These images faithfully capture the scene's spatial layout, furniture, and fine color details. Additionally, PanoGen module supports stylized synthesis with optional textual or imagery-based controls.", + "bbox": [ + 511, + 90, + 908, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Volumetric Occupancy Estimation", + "text_level": 1, + "bbox": [ + 511, + 207, + 815, + 223 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The 2D top-down views lack 3D structural information about objects and furniture. To render panoramas that accurately reflect the geometric spatial relationships of objects, we propose training an OccRecon module to estimate the scene's 3D occupancy or density.", + "bbox": [ + 511, + 229, + 906, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input Representations. Unlike the layout-guided 3D scene generation setting, our top-down input lacks semantic information. In our preliminary study, we found that current semantic segmentation models struggle to generalize to indoor top-down views, making it challenging to estimate the 3D structure without semantic guidance. To address this, we propose leveraging a pretrained segmentation model SAM [19] to extract the 2D structure of the scene. Both the top-down image and the segmentation view are then fed into the encoder of the OccRecon module. The segmentation provides valuable details, such as room boundaries, furniture positions, and shapes, which significantly enhance the OccRecon module's ability to learn the overall 3D structure of the rooms. Moreover, this semantic-free input design enables our model to generalize effectively to more abstract inputs, such as semantic floorplans.", + "bbox": [ + 511, + 310, + 908, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OccRecon Module. We propose a diffusion-based encoder-decoder framework that efficiently extracts and reconstructs the 3D spatial structure of a scene. Instead of using computationally intensive 3D diffusion models, our approach leverages a 2D diffusion model [42], significantly reducing resource demands. The OccRecon module processes 2D images and segmentations to extract spatial information and reconstruct room structures, including wall and furniture heights. In the final step, a 3D convolutional layer integrates the learned spatial features to generate a comprehensive 3D occupancy map. By relying on 2D inputs for most of the process, our method drastically lowers computational costs while still capturing the full scene layout. This balance of efficiency and accuracy makes it a practical solution for 3D structural modeling without semantic inputs. The OccRecon module outputs a 3D volumetric occupancy map $V_{\\mathrm{occ}}$ , which explicitly represents the scene's 3D geometry.", + "bbox": [ + 511, + 556, + 908, + 814 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\text {o c c}} = \\text {O c c R e c o n} \\left(I _ {\\text {t o p}}, I _ {\\text {s e g}}\\right) _ {\\text {c o n d i t i o n}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 825, + 906, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We normalize the occupancy values to the range [0, 1]. Given the scale information of the scene, we transform the real-world coordinates of a 3D point into the estimated volumetric", + "bbox": [ + 511, + 854, + 910, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28495", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "space. The point's density is then obtained by querying the learned occupancy map $V_{\\mathrm{occ}}$ using tri-linear interpolation.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Structural Reinforcement. Top-down views offer a comprehensive overview of the entire floor, but when observed from a first-person perspective, the scene typically reveals only the details of the current space, obscuring rooms beyond the walls. After the model learns the overall geometric structure of the rooms through the OccRecon module, we apply structural reinforcement to refine the wall geometry. This provides precise depth information, enhancing the reconstruction of the room's geometric layout. We solidify the wall voxels by applying the maximum value (1 after normalization). Additionally, top-down priors can constrain the floor's geometry and help infer the texture of the furniture, which is essential for generating accurate indoor panoramas with correctly placed and colored furniture. Since the OccNet modules are optimized end-to-end with the subsequent modules, the height of the furniture can be inferred from the 3D occupancy map. By encoding structural constraints from the walls and floor, we achieve a more accurate representation of the scene's geometry, including the positions, colors, and other attributes of the furniture.", + "bbox": [ + 89, + 126, + 486, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Coarse Panorama Rendering", + "text_level": 1, + "bbox": [ + 89, + 440, + 352, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the 3D occupancy map of the scene, we render coarse depth and color images based on the specified camera position. To ensure accurate mapping, we first compute the ratio between pixel resolution and physical dimensions. With the room height known, we apply this ratio along with the pixel coordinates in the top-down image to determine the corresponding position in the occupancy map $V_{\\mathrm{occ}}$ . To generate the panoramic image from the 3D occupancy map, we employ equirectangular projection along with a spherical coordinate system.", + "bbox": [ + 89, + 463, + 483, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The coarse depth panorama $I_{\\mathrm{depth}}$ is obtained through volumetric rendering [24] of the occupancy map. The depth of a projected ray at pixel $(u, v)$ is computed as follows:", + "bbox": [ + 89, + 614, + 483, + 660 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {d e p t h}} ^ {(u, v)} = \\sum_ {i = 1} ^ {S} T _ {i} \\alpha_ {i} d _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 672, + 483, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{i}$ represents the transparency level, and $d_{i}$ denotes the distance from the sampled position to the camera.", + "bbox": [ + 89, + 728, + 482, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The coarse color panorama $I_{\\mathrm{color}}$ is obtained by directly projecting the indoor top-down view along the camera rays [28], assigning colors based on the corresponding intersections and bilinear interpolation. Specifically, the color at pixel $(u,v)$ is determined by sampling the top-down image without learning a radiance field.", + "bbox": [ + 89, + 758, + 483, + 849 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {c o l o r}} ^ {(u, v)} = \\sum_ {i} ^ {S} T _ {i} \\alpha_ {i} c _ {i} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 863, + 483, + 904 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{i}$ is the color copied from the top-down image. This approach directly maps color from the top-down view, offering a more straightforward and computationally efficient solution compared to NeRF [25], which reconstructs scenes by learning a radiance field for view synthesis.", + "bbox": [ + 511, + 90, + 906, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employed a uniform voxel sampling strategy, where voxels were sampled along a fixed-length ray for both coarse color and coarse depth. However, this approach introduced banding artifacts, particularly noticeable on the floor directly beneath the camera in the coarse color image. Since the top-down view served as the floor texture, maintaining high-quality details was essential for achieving realistic rendering. To mitigate these artifacts, we reduced the ray length by half for coarse color sampling. This adjustment increased the density of sample points within the same spatial region, enhancing sampling accuracy and producing smoother color and texture transitions. As a result, banding artifacts were significantly diminished, improving overall rendering quality. For coarse depth, preserving scene structure was paramount, so we retained the original sampling strategy.", + "bbox": [ + 511, + 167, + 908, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Photorealistic Synthesis", + "text_level": 1, + "bbox": [ + 513, + 402, + 732, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Generating photorealistic panoramic images directly from top-down views is challenging. To address this, we propose a two-stage pipeline. In the second stage, the PanoGen module synthesizes photorealistic indoor panoramic images from coarse color and depth inputs. We implement PanoGen using a diffusion-based ControlNet [42], enabling the restoration of fine details. This two-stage approach not only reconstructs the house's structural layout, including precise wall positions, but also restores elements such as windows, lighting, and furniture. The final panoramic image $I_{\\mathrm{pano}}$ is generated based on the two coarse inputs from the previous stage:", + "bbox": [ + 511, + 424, + 908, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\text {p a n o}} = \\text {P a n o G e n} \\left(I _ {\\text {c o l o r}}, I _ {\\text {d e p t h}}\\right) _ {\\text {c o n d i t i o n}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 602, + 906, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The PanoGen module modifies the conditioning mechanism of ControlNet by treating coarse color and depth images as separate inputs before combining them. This design enables PanoGen to capture both the scene's geometric structure and the furniture's color and position, resulting in more accurate and high-quality panoramic images. To further enhance consistency, we incorporate an alignment loss to prevent structural distortions when the viewpoint changes and a color loss to ensure accurate color reproduction in the synthesized output.", + "bbox": [ + 511, + 630, + 908, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Training and Optimization", + "text_level": 1, + "bbox": [ + 511, + 790, + 756, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Top2Pano model employs denoising MSE loss, alignment loss, and color loss functions for optimization. The denoising MSE loss function [42] is defined as:", + "bbox": [ + 511, + 811, + 906, + 858 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f}} = \\mathbb {E} _ {z _ {0}, t, c _ {t}, c _ {c}, c _ {d}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, c _ {t}, c _ {c}, c _ {d}\\right) \\| _ {2} ^ {2} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 867, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28496", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ denotes the number of noise addition steps, $c_{\\mathrm{c}}$ represents the corresponding coarse color image, and $c_{\\mathrm{d}}$ represents the coarse depth image. The variables $t$ and $c_{t}$ correspond to the time step and simple text prompts, respectively. The diffusion algorithm trains a neural network $\\epsilon_{\\theta}$ to predict the noise added to the noisy image $z_{t}$ . The alignment loss is formulated as:", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a l i g n m e n t}} = \\left[ \\| I _ {D} - \\hat {I _ {D}} \\| _ {2} ^ {2} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 202, + 483, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I_{D}$ represents the depth images generated by the model and $\\hat{I_D}$ represents the ground truth depth images. This loss function is used to address distortions in the furniture.", + "bbox": [ + 89, + 237, + 482, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $I$ be the rendered image and $G$ the ground truth image, both with $C = 3$ color channels. For each channel $c \\in \\{1, \\ldots, C\\}$ , we define the normalized histogram as", + "bbox": [ + 89, + 282, + 483, + 329 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH ^ {c} (I) = \\left(h _ {1} ^ {c} (I), h _ {2} ^ {c} (I), \\dots , h _ {\\mathrm {b i n s}} ^ {c} (I)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 337, + 483, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where bins is the number of histogram bins (e.g., 256), and each $h_k^c (I)$ represents the normalized frequency (probability) of pixel intensities falling into bin $k$ . Likewise, for the ground truth image $G$ , we have", + "bbox": [ + 89, + 364, + 483, + 425 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH ^ {c} (G) = \\big (h _ {1} ^ {c} (G), h _ {2} ^ {c} (G), \\ldots , h _ {\\mathrm {b i n s}} ^ {c} (G) \\big). \\qquad (8)\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 435, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The color histogram loss, measured using the $L_{1}$ norm, is given by", + "bbox": [ + 89, + 462, + 483, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o l o r}} (I, G) = \\sum_ {c = 1} ^ {C} \\| H ^ {c} (I) - H ^ {c} (G) \\| _ {1} \\tag {9} \\\\ = \\sum_ {c = 1} ^ {C} \\sum_ {k = 1} ^ {\\text {b i n s}} \\left| h _ {k} ^ {c} (I) - h _ {k} ^ {c} (G) \\right|. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 502, + 483, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final loss function combines three loss terms as follows:", + "bbox": [ + 89, + 597, + 483, + 612 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {d i f f}} + \\mathcal {L} _ {\\text {a l g n m e n t}} + \\mathcal {L} _ {\\text {c o l o r}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 623, + 483, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Generalization and Stylized Synthesis", + "text_level": 1, + "bbox": [ + 89, + 648, + 416, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generalization to Floorplan. We aim to generate first-person view panoramas that faithfully represent a scene. While top-down views lack vertical details like walls and windows, this omission grants flexibility in generating panoramas. This flexibility is especially useful in interior design, where inputs are often simple floorplans. To improve generalization, particularly to schematic floorplans, we train on orthographic rather than perspective views, as they better match floorplans. Empirical results show that Top2Pano generalizes well to schematic and even hand-drawn floorplans while maintaining photorealism. Additionally, our model enables stylized synthesis guided by text or images, supporting diverse design needs.", + "bbox": [ + 89, + 670, + 483, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Text-Guided Stylization. The PanoGen module, built upon the text-driven Stable Diffusion model, inherently supports", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "text-conditioned image generation. As illustrated in Figure 2, PanoGen synthesizes panoramas using three conditions: coarse depth, coarse colored image, and stylized textual guidance. When the input is a textureless floorplan, the stylized textual condition effectively guides the style of the synthesized result, as demonstrated in Figure 1. However, when a colored top-down view is provided, the influence of text-guided stylization becomes less pronounced. This occurs because the rendered coarse colored panorama constrains the final output to closely align with the input view, thereby diminishing the impact of the textual stylization.", + "bbox": [ + 511, + 90, + 906, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance text-guided stylization, the weight of the coarse colored panorama in the PanoGen conditions can be reduced. However, this introduces a tradeoff: prioritizing stylization may come at the expense of fidelity to the top-down view. Regardless of this tradeoff, the coarse depth condition consistently ensures that the synthesized panorama adheres to the underlying scene geometry. Notably, this tradeoff does not apply to the text-to-panorama generation task, where textual guidance plays a more dominant role.", + "bbox": [ + 511, + 260, + 908, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image-Guided Stylization. Given several scene images (not necessarily panoramas), we can fine-tune the PanoGen module using low-rank adaptation (LoRA) [14] to generate panoramas that align with the visual styles present in the provided images. To guide this process, we introduce structured textural prompts augmented with style tags (e.g., [Japanese]) at the beginning of the input prompts. These tags act as conditional modifiers, steering the model toward synthesizing images that follow specific aesthetic themes, such as regional design styles. The framework is applied solely to the PanoGen module, ensuring both computational and parameter efficiency. Notably, our method requires fewer than five in-the-wild images per target style, substantially reducing data demands. This approach achieves its efficiency by decomposing weight updates into low-rank matrices. For a pretrained weight matrix $\\mathbf{W}_0 \\in \\mathbb{R}^{d \\times d}$ , the update $\\Delta \\mathbf{W}$ is constrained as follows:", + "bbox": [ + 511, + 405, + 908, + 661 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {W} = \\mathbf {B} \\mathbf {A}, \\quad \\text {w h e r e} \\mathbf {A} \\in \\mathbb {R} ^ {d \\times r}, \\mathbf {B} \\in \\mathbb {R} ^ {r \\times d} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 678, + 906, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $r \\ll d$ represents the intrinsic rank (we use $r = 8$ ). During fine-tuning, only the matrices $\\mathbf{A}$ and $\\mathbf{B}$ are updated, while the original weights $\\mathbf{W}_0$ remain fixed. The forward pass is modified as follows:", + "bbox": [ + 511, + 713, + 906, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {\\text {o u t}} = \\mathbf {W} _ {0} \\mathbf {h} _ {\\text {i n}} + \\alpha \\cdot \\mathbf {B A h} _ {\\text {i n}} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 791, + 906, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is a scaling coefficient. This lightweight adaptation (0.8% new parameters) mitigates catastrophic forgetting and maintains the model's baseline generation quality for generic prompts while enabling precise style control through our [style] textual conditioning.", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28497", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/65f305cfb4c13676df540f952b0be67aaa8aade81f5e88d7ecbe01bed1e7b612.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TrainingTesting
ScenesFloorsPanoramasScenesFloorsPanoramas
Matterport3D61127617714291405
Gibson152203537939761672
", + "bbox": [ + 94, + 88, + 478, + 147 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4c9c883bd5bbeaa6bba8545b72ae1d3179f249a1a6f2ed88d3d293da7bbaa0cc.jpg", + "table_caption": [ + "Table 1. The numbers of scenes, floors, and panorama images in the training and testing sets of the two datasets" + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑FID ↓LPIPS ↓
Materport3DSat2Density[28]+LDM[29]11.270.4221100.060.6237
Sat2Density[28]+ControlNet[42]11.420.422285.780.6163
PanFusion[41]11.450.437285.740.6153
Top2Pano (Ours)11.720.440930.840.6029
GibsonSat2Density[28]+LDM[29]10.540.448084.330.6462
Sat2Density[28]+ControlNet[42]10.970.458285.210.6645
PanFusion[41]11.360.474479.530.6634
Top2Pano (Ours)11.580.485128.680.6282
", + "bbox": [ + 94, + 195, + 475, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 347, + 223, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Data Preparation", + "text_level": 1, + "bbox": [ + 89, + 372, + 261, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For evaluation, we use the Matterport3D [3] and Gibson [36] datasets. Since no existing dataset provides both top-down views and high-quality panoramic images, we generate top-down views from 3D models in these datasets using Blender. Specifically, we import textured 3D meshes into Blender and render top-down views with an orthographic camera. The top-down view we render closely resembles a floorplan, unlike the perspective-rendered views used in embodied dialog localization [12]. This similarity enhances our model's ability to generalize to floorplan inputs. To determine the number of floors in each scene, we apply DBSCAN [7] clustering to the camera positions within the datasets. We exclude certain scenes, such as airports and large supermarkets, as well as panoramic images depicting outdoor environments to ensure alignment with our task. After processing, the final dataset sizes are summarized in Table 1.", + "bbox": [ + 89, + 395, + 483, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 648, + 274, + 664 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the quality of the generated panoramas, we employ both pixel-based and perceptual evaluation metrics. For pixel-level assessment, we utilize peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) to quantify image fidelity. Additionally, we incorporate perceptual metrics such as Fréchet Inception Distance (FID) [13] and Learned Perceptual Image Patch Similarity (LPIPS) [43] to capture higher-level visual realism.", + "bbox": [ + 89, + 669, + 483, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implement Details", + "text_level": 1, + "bbox": [ + 89, + 801, + 269, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our code runs on an NVIDIA RTX A6000 GPU with 48GB of memory. The model has 3.3 billion parameters and is trained with a batch size of 21 for 100 epochs. On average, each experiment takes approximately two days to complete on both the Matterport3D and Gibson datasets. We optimize", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/52b46dd0201b9cb851f318d47b311aca5baf4cc0e74b8767044495b48103cab5.jpg", + "table_caption": [ + "Table 2. Quantitative comparison with existing methods on the Matterport3D [3] and Gibson [36] datasets." + ], + "table_footnote": [], + "table_body": "
ModulesMatterport3DGibson
seg floor wall depth colorPSNR↑SSIM↑FID↓LPIPS↓PSNR↑SSIM↑FID↓LPIPS↓
✓✓11.080.4122104.030.646311.070.4551117.080.6574
✓✓✓11.260.4206102.840.647111.100.4564115.880.6568
✓✓✓10.780.409155.730.616111.190.466832.280.6310
✓✓✓11.220.420754.440.619311.280.461383.310.6429
✓✓✓11.570.437843.990.606711.160.473381.870.6549
✓✓✓10.980.419653.340.611511.330.460133.890.6307
✓✓11.210.419935.080.613111.380.464136.290.6384
✓✓✓11.260.429934.610.606011.490.467334.360.6336
✓✓✓✓11.590.438167.020.619211.380.476165.660.6557
✓✓✓✓11.720.440930.840.602911.580.485128.680.6282
", + "bbox": [ + 517, + 88, + 898, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Ablation study on five designs in our Top2Pano model (top-down view segmentation, floor reinforcement, wall reinforcement, coarse depth panorama, and coarse colored panorama).", + "bbox": [ + 511, + 217, + 906, + 260 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "our model using the Adam optimizer [18] with default parameters $(\\beta_{1} = 0.9, \\beta_{2} = 0.999, \\epsilon = 10^{-8})$ and a learning rate of $10^{-5}$ .", + "bbox": [ + 511, + 287, + 906, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparison with Previous Methods", + "text_level": 1, + "bbox": [ + 511, + 344, + 825, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 2, we compare our method against three baseline approaches across four evaluation metrics. Sat2Density [28] is a satellite-to-ground panorama synthesis method, which we adapt for indoor panorama generation using latent diffusion model (LDM) [29] and ControlNet [42]. PanFusion [41] is a text-to-panorama generation framework that also incorporates layout-conditioned generation via ControlNet. Table 2 shows that our method outperforms all baselines across all four metrics on both datasets, demonstrating its effectiveness. Furthermore, qualitative comparisons in Figures 3 and 4 highlight that our approach generates more realistic and structurally accurate house reconstructions, including furniture placement.", + "bbox": [ + 511, + 367, + 908, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 575, + 666, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conducted comprehensive ablation studies to analyze and validate the contribution of each component in our model. Specifically, we performed experiments comparing several model variants against the original. These experiments involved removing key elements such as the structural reinforcement of the floor and walls, the segmentation input to the OccRecon module, and the coarse depth and colored panoramicas as conditional inputs to the PanoGen module.", + "bbox": [ + 511, + 597, + 906, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 3, our original model achieves the highest overall scores across all four metrics, with any modification leading to some degree of performance degradation. Removing coarse colored panoramas or the embedded floor significantly disrupts furniture placement and color accuracy. While the room structure remains mostly intact, furniture positions become unreliable, and color representations appear distorted. Conversely, excluding coarse depth panoramas or embedded walls maintains color and furniture accuracy but compromises the spatial understanding and overall quality of room structure reconstruction. These effects are further illustrated in the qualitative results in the supp. materials.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28498", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d87bfacbbc6d25388816b311b302213cb3d0bb46625e8a1b8a330748ebf26ffd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 88, + 187, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3ca2b08a577bc0255d88476a19b5b742486e8bc1aabfbb723f06961184c25bd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 88, + 364, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e220fba2817c759044c7626b257b0c9b4adfc237980f430092901eb295208e35.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 88, + 542, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dab818e600dee898acfc2dc454e591b49ce9d5b58d62c1aff061ef0fa156a63d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 88, + 717, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6454c2d8d9f2c5252d908f6876fd693422c00b507e8997bb017a8d2b70043d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 88, + 893, + 157 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5f7da3d5ae4cdde7b4064799ae46cde9ec32987c8729f96f8f682d3354539fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 159, + 187, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75ad0ed1229b98c0dbca869cc726476e855f3964e596058b116451dfbddc4b47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 159, + 364, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2348aef428cb6c447d27ce7ae2b820d3afed552c37d78c221a5543c0157a9d1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 159, + 542, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/41173e3b0c831caa22b0d9be55a8dd33d3b34beef106cb034f5b38dcd1cd8388.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 159, + 717, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb192194dd9117f977de4e20e85edd5efc1511e495e37f2bfaa00d955589b7db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 159, + 893, + 226 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe2fbd79e74a5ad1b0a976300bc0e99b339fcd993ff76e81b541f306f1d64780.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 228, + 187, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3426376a90bee262dc255136afccf5f52ffc8fe24712aefff2cfc93a82dbfbea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 228, + 364, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10b3cc6e7819e1040a40a49e6805c63f35f2e99e97adfbc7d12e7b9532764ffb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 228, + 542, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/67e10f29dc3dac2670427d6102daaab6639597534db638ac9ae29272a2172d32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 228, + 717, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/11ab96e5e6346a65369b30f3e2581fca1700ce78d394df5d3e1b3654120bbdc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 228, + 893, + 297 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d261615fde255af288974800b7dd9e7c9836e1b410e134e97253837a45fb78c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 299, + 187, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0c73e5110e422f43205eee42ed5486b5d3b29226158e5fc75f728f4a3909f48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 299, + 364, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee5b0a2f7f93b90f13aa4246689e5f90868d3abe4ad49b80c40986e1e9ac86de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 297, + 542, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4a9af37419317d636564f7f7cff1e21c0c46d045673fe8b898d89b2408d2ba0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 297, + 717, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b40e50d7b666915888905233a3b749b7a8e10ce11572bf9a2be83ad0dc765b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 299, + 893, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6edcdca92b1bdc686e11fec9330d3bdb8bfd32397fd5076cffbfa7777824660a.jpg", + "image_caption": [ + "(a) Top-down" + ], + "image_footnote": [], + "bbox": [ + 99, + 369, + 187, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8e49994ea7e6a2599dd1cb16ec253e349de08635308ebce6ccd6983484f79d5.jpg", + "image_caption": [ + "(b) Sat2Density [28]" + ], + "image_footnote": [], + "bbox": [ + 191, + 368, + 364, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea2eca47b313d457ee0e09d8e10eb968faa0f6789babfd373153daf2b6e0e402.jpg", + "image_caption": [ + "(c) PanFusion [41]" + ], + "image_footnote": [], + "bbox": [ + 367, + 368, + 542, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d629db6494fc760ba7f900d97a389c87fe76ff2726a5e4151385b61d555da4c5.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 544, + 368, + 717, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e35433d217b9c84193629a8f88df49aeeb454c2a7f1a711a1344bbf68d9e5a9.jpg", + "image_caption": [ + "(e) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 720, + 368, + 893, + 436 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c14d0188b5709c4fcbef83d223b3baff16efbd9fcead728a823820e16b301518.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 484, + 187, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc66637b156586e359c5d763c911da2cd6e54ce72334141b7a32be5fce994a7b.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons on the Matterport3D dataset." + ], + "image_footnote": [], + "bbox": [ + 191, + 484, + 364, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/223edc01e4497d93b35884ac4c903647637fe4fb48b10291873c1bd5ad10a030.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 484, + 542, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f05c4ef89b20a0e2e1237535624cd2f5b6cbb051067fa6eab9fc5f027f22caee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 484, + 717, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/284404ffe703471b9303da5234fb7ef40c9d5e000682f300df7d45a0ec3d2f9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 484, + 893, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0db089f8e1927247b9f64db373875313f8eb6fc7bad8ca1644dd5cd83cde8c6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 554, + 187, + 622 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb8385ff7fb9c9656d5b7030844879f209718570148899aba6e1e8286ac99e1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 554, + 364, + 622 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/af0af0dff358f985d9eabdf9f30854ae5126e1ef205533739985f1e90181fcc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 554, + 542, + 622 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b72e69f0ce8dc37538fc9b372ec4f33b91a85d401b8efa37d1e69aafb350f367.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 554, + 717, + 622 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/22b649c8498fa3f2069c391825e9625df562d97c8777744372d84556772beb96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 554, + 893, + 622 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3105be3aacb1178bd379a3a4559793fb7a1e91a71f067604035014c37e3e5c92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 625, + 187, + 690 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25c43bb4407057ca1401bad007444e19a6bab19777f98842606b27b490c9b705.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 625, + 364, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3cccea65905067192fb436a816d70d258ce8907d519b0fc53233e6cbe43da0e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 625, + 542, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cbe4164e2c0a9d6548b9b168044fea952d010f4764a162c0876e163a32c49a1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 625, + 717, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aff2a99c7786f45ad56c41a9428a718b2db168ccad78cae9595139915c3cb531.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 625, + 893, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6516874589b8658a39e1599fa4a02d532010c2f7145f7bf1aee50ed6297019a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 693, + 187, + 762 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9eb348b30e81c6d5ce07f54c50d188fa9f80800e7dfdb75956eceb1602c9f2b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 693, + 364, + 762 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9651671a11b4ba16871f82afc7c6627c3e00276eac76b7ee7f7e70b655ead32c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 693, + 542, + 762 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ce4e1b5238ee97ddc15fe9d29318d483bc0ad3f77fa5989fa6e6fe5e1d5c6c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 693, + 717, + 762 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f48ef503b889b0948444e0128a995a2d1ddc10ed6b726a4b911bdc15927c3e43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 693, + 893, + 762 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1531d71541dcd9280f6684c6862637804b1ead2ddf6762d121623b8c07cfac53.jpg", + "image_caption": [ + "(a) Top-down" + ], + "image_footnote": [], + "bbox": [ + 99, + 763, + 187, + 832 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/746aacbb27b068d1b2a05b0f8494a04c5f0afb44fc3a1462cac537e24b9d597e.jpg", + "image_caption": [ + "(b) Sat2Density [28]" + ], + "image_footnote": [], + "bbox": [ + 191, + 763, + 364, + 832 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b8b42fa12bebc8a610f4b66ac65710f1fc61c3d641929e8e3e62452604cfcc1.jpg", + "image_caption": [ + "(c) PanFusion [41]" + ], + "image_footnote": [], + "bbox": [ + 367, + 763, + 542, + 832 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fe67f44e29f29015cdf3ba2dcc34507800e3ea435d7b3c635134d10929fcd704.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 544, + 763, + 717, + 832 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/560c00334ce9b6bfc96c237b6c9333336e4f7cc21d64e2c2230b4a3bc8c80e8f.jpg", + "image_caption": [ + "(e) Ground Truth", + "Figure 4. Qualitative comparisons on the Gibson dataset." + ], + "image_footnote": [], + "bbox": [ + 720, + 763, + 893, + 832 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28499", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7311c6f1b39e342a05b4e70eebbd812954f58f5c573d95902b26f2482de4c617.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑FID ↓LPIPS ↓
G → M11.080.439740.740.6285
M → G11.030.436645.870.6353
", + "bbox": [ + 132, + 88, + 439, + 140 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Cross-dataset evaluation. \"G\" represents Gibson; \"M\" represents Matterport3D.", + "bbox": [ + 89, + 143, + 485, + 171 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb27ff5d3b186bf25393e6e2934fbc1426ed1b363a55116bcf741168233b5655.jpg", + "image_caption": [ + "Figure 5. Our Top2Pano model generalizes to both textured floorplans (first row) and hand-drawn sketched floorplans (second row), incorporating stylized [Japanese] control." + ], + "image_footnote": [], + "bbox": [ + 127, + 175, + 442, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Generalization, Stylization, Manipulation", + "text_level": 1, + "bbox": [ + 89, + 378, + 444, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We employed cross-dataset evaluation to assess our model's generalization capability. As shown in Table 4, although there is a slight decline in metric scores, our model maintains strong performance. Additionally, the trained parameters tend to generate photorealistic walls that reflect characteristics of the training dataset.", + "bbox": [ + 89, + 401, + 483, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our model also generalizes well to floorplans. When given a specific floorplan and camera positions, the model assists users in generating indoor panoramic images. By using different text prompts, users can create various interior design styles, explore the house from a first-person perspective, and modify its style according to their preferences. We tested our model with three types of floorplans. The first type is a colored floorplan (Figure 5, first row), which provides detailed information about room colors and furniture hues, making it highly informative. The second type is a plain floorplan (Figure 1, bottom), which lacks color information and shows only the room structure and furniture layout. The third type is a hand-drawn floorplan sketch (Figure 5, first row), which offers a rough visual representation of the room. Our model successfully generates accurate panoramic images from these floorplans and adapts the visual style based on the provided textual descriptions. Moreover, our model enables panorama manipulation by editing objects in the floorplan, such as adding new items, as illustrated in Figure 6.", + "bbox": [ + 89, + 491, + 483, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Limitations", + "text_level": 1, + "bbox": [ + 89, + 803, + 215, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Failure cases. Figure 7 shows representative failure cases. We annotate different types of failures with numbered labels:", + "bbox": [ + 89, + 825, + 483, + 854 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Ceiling: ① missing fan, ② vaulted ceiling, ③ false light;", + "- Wall: 4 height error, 5 false or missing decorations;", + "- Window: 6 false window;" + ], + "bbox": [ + 89, + 854, + 475, + 898 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45012e95f40adb80c8263859436dab7eb4b3f3599ce770cc79ad7ff8b743e267.jpg", + "image_caption": [ + "Figure 6. Top2Pano enables panorama manipulation via compositional floorplan editing. In the second row, adding a rectangular object to the floorplan (compared to the first row) leads the model to generate a washstand with a mirror in the panorama." + ], + "image_footnote": [], + "bbox": [ + 552, + 88, + 866, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fc250a1e614b2c9d2c81f0f66d5713977d5d4e5d91f9c52eec3769209a8769e.jpg", + "image_caption": [ + "Figure 7. Failure cases (zoom in to view error types)." + ], + "image_footnote": [], + "bbox": [ + 517, + 301, + 911, + 503 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Furniture: height error;", + "- Thin object: 8 missing flat-screen TV;", + "Stairs: false direction." + ], + "bbox": [ + 513, + 529, + 782, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The failure cases stem from the ambiguity of the 2D input, leading to hallucinated objects that are not observable from the top-down view. The limitations in handling vertical structural details are largely due to the inherent ambiguity of the task.", + "bbox": [ + 511, + 575, + 905, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limited Vertical FoV. The generated panoramas exhibit a limited vertical field of view (FoV), reflecting the constraints of the training data. We expect improved performance with future datasets that include full vertical FoV panoramas.", + "bbox": [ + 511, + 654, + 905, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 727, + 638, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present Top2Pano, a novel method for generating high-quality $360^{\\circ}$ indoor panoramas from 2D top-down views. The model first estimates volumetric occupancy to infer 3D structure, then applies volumetric rendering to produce coarse color and depth panoramas. These guide a diffusion-based refinement stage via ControlNet. To our knowledge, this is the first approach to generate panoramas from top-down views. Experiments on two datasets show that Top2Pano outperforms baselines in reconstructing room layouts and realistic furniture.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28500", + "bbox": [ + 478, + 943, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. Diverse plausible 360-degree image outpainting for efficient 3dcg background creation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 11431-11440, 2022. 3", + "[2] Sherwin Bahmani, Jeong Joon Park, Despoina Paschalidou, Xingguang Yan, Gordon Wetzstein, Leonidas J. Guibas, and Andrea Tagliasacchi. CC3D: layout-conditioned generation of compositional 3d scenes. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 7137-7147, 2023. 1, 3", + "[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1, 6", + "[4] Minglin Chen, Longguang Wang, Sheng Ao, Ye Zhang, Kai Xu, and Yulan Guo. Layout2scene: 3d semantic layout guided scene generation via geometry and appearance diffusion priors. arXiv preprint arXiv:2501.02519, 2025. 1, 3", + "[5] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 3", + "[6] Mohammad Reza Karimi Dastjerdi, Yannick Hold-Geoffroy, Jonathan Eisenmann, Siavash Khodadadeh, and Jean-François Lalonde. Guided co-modulated gan for $360^{\\circ}$ field of view extrapolation. In 2022 International Conference on 3D Vision (3DV), pages 475–485, 2022. 3", + "[7] Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD), pages 226-231, 1996. 6", + "[8] Chuan Fang, Xiaotao Hu, Kunming Luo, and Ping Tan. Ctrlroom: Controllable text-to-3d room meshes generation with layout constraints. arXiv preprint arXiv:2310.03602, 2023. 1, 3", + "[9] Chengzeng Feng, Jiacheng Wei, Cheng Chen, Yang Li, Pan Ji, Fayao Liu, Hongdong Li, and Guosheng Lin. Prim2room: Layout-controllable room mesh generation from primitives. arXiv preprint arXiv:2409.05380, 2024. 1, 3", + "[10] Mengyang Feng, Jinlin Liu, Miaomiao Cui, and Xuansong Xie. Diffusion360: Seamless 360 degree panoramic image generation based on diffusion models. arXiv preprint arXiv:2311.13141, 2023. 1, 3", + "[11] Julia Guerrero-Viu, Clara Fernandez-Labrador, Cédric Demonceaux, and José Jesús Guerrero. What's in my room? object recognition on indoor panoramic images. In IEEE International Conference on Robotics and Automation, ICRA, pages 567-573, 2020. 1", + "[12] Meera Hahn, Jacob Krantz, Dhruv Batra, Devi Parikh, James M Rehg, Stefan Lee, and Peter Anderson. Where are you? localization from embodied dialog. arXiv preprint arXiv:2011.08277, 2020. 6", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "time-scale update rule converge to a local nash equilibrium. In Annual Conference on Neural Information Processing Systems, pages 6626-6637, 2017. 6", + "[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, ICLR, 2022. 5", + "[15] Zilong Huang, Jun He, Junyan Ye, Lihan Jiang, Weijia Li, Yiping Chen, and Ting Han. Scene4u: Hierarchical layered 3d scene reconstruction from single panoramic image for your immerse exploration. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 26723-26733, 2025. 1", + "[16] Zhe Huang, Yizhe Zhao, Hao Xiao, Chenyan Wu, and Lingting Ge. Duospacenet: Leveraging both bird's-eye-view and perspective view representations for 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops, pages 2560-2570, 2025. 1", + "[17] Nikolai Kalischek, Michael Oechsle, Fabian Manhardt, Philipp Henzler, Konrad Schindler, and Federico Tombari. Cubediff: Repurposing diffusion-based image models for panorama generation. In International Conference on Learning Representations, ICLR, 2025. 3", + "[18] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, ICLR, 2015. 6", + "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 3992-4003, 2023. 2, 3", + "[20] Jiachen Liu, Yuan Xue, Haomiao Ni, Rui Yu, Zihan Zhou, and Sharon X Huang. Computer-aided layout generation for building design: A review. arXiv preprint arXiv:2504.09694, 2025.1", + "[21] Jiachen Liu, Rui Yu, Sili Chen, Sharon X. Huang, and Hengkai Guo. Towards in-the-wild 3d plane reconstruction from a single image. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 27027-27037, 2025. 1", + "[22] Xiaohu Lu, Zuoyue Li, Zhaopeng Cui, Martin R. Oswald, Marc Pollefeys, and Rongjun Qin. Geometry-aware satellite-to-ground image synthesis for urban areas. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 856-864, 2020. 3", + "[23] Wentao Lyu, Peng Ding, Yingliang Zhang, Anpei Chen, Minye Wu, Shu Yin, and Jingyi Yu. Refocusable gigapixel panoramas for immersive VR experiences. IEEE Trans. Vis. Comput. Graph., 27(3):2028-2040, 2021. 1", + "[24] Nelson L. Max. Optical models for direct volume rendering. IEEE Trans. Vis. Comput. Graph., 1(2):99-108, 1995. 4", + "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28501", + "bbox": [ + 478, + 943, + 517, + 954 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "thesis. In European Conference on Computer Vision, ECCV, pages 405-421, 2020. 4", + "[26] Despoina Paschalidou, Amlan Kar, Maria Shugrina, Karsten Kreis, Andreas Geiger, and Sanja Fidler. Atiss: Autoregressive transformers for indoor scene synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[27] Giovanni Pintore, Fabio Bettio, Marco Agus, and Enrico Gobbetti. Deep scene synthesis of atlanta-world interiors from a single omnidirectional image. IEEE Trans. Vis. Comput. Graph., 29(11):4708-4718, 2023. 1", + "[28] Ming Qian, Jincheng Xiong, Gui-Song Xia, and Nan Xue. Sat2density: Faithful density learning from satellite-ground image pairs. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 3660-3669, 2023. 3, 4, 6, 7", + "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 10674-10685, 2022. 6", + "[30] Jonas Schult, Sam Tsai, Lukas Hollein, Bichen Wu, Jialiang Wang, Chih-Yao Ma, Kunpeng Li, Xiaofang Wang, Felix Wimbauer, Zijian He, Peizhao Zhang, Bastian Leibe, Peter Vajda, and Ji Hou. Controlroom3d: Room generation using semantic proxy rooms. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 3", + "[31] Yujiao Shi, Dylan Campbell, Xin Yu, and Hongdong Li. Geometry-guided street-view panorama synthesis from satellite imagery. IEEE Trans. Pattern Anal. Mach. Intell., 44(12): 10009-10022, 2022. 3", + "[32] Bin Tan, Rui Yu, Yujun Shen, and Nan Xue. Planarsplatting: Accurate planar surface reconstruction in 3 minutes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 1190-1199, 2025. 1", + "[33] Madhawa Vidanapathirana, Qirui Wu, Yasutaka Furukawa, Angel X. Chang, and Manolis Savva. Plan2scene: Converting floorplans to 3d scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 10733-10742, 2021. 3", + "[34] Jionghao Wang, Ziyu Chen, Jun Ling, Rong Xie, and Li Song. 360-degree panorama generation from few unregistered nfov images. In Proceedings of the 31st ACM International Conference on Multimedia, MM, pages 6811-6821, 2023. 3", + "[35] Songsong Wu, Hao Tang, Xiao-Yuan Jing, Haifeng Zhao, Jianjun Qian, Nicu Sebe, and Yan Yan. Cross-view panorama image synthesis. IEEE Trans. Multim., 25:3546-3559, 2023. 3", + "[36] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: real-world perception for embodied agents. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, 2018. 6", + "[37] Ningli Xu and Rongjun Qin. Geospecific view generation geometry-context aware high-resolution ground view inference from satellite views. In European Conference on Computer Vision, ECCV, pages 349-366, 2024. 3", + "[38] Xiuyu Yang, Yunze Man, Jun-Kun Chen, and Yu-Xiong Wang. Scenecraft: Layout-guided 3d scene generation. In Advances in Neural Information Processing Systems, 2024. 1, 3" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Weicai Ye, Chenhao Ji, Zheng Chen, Junyao Gao, Xiaoshui Huang, Song-Hai Zhang, Wanli Ouyang, Tong He, Cairong Zhao, and Guofeng Zhang. Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion. In Annual Conference on Neural Information Processing Systems, NeurIPS, 2024. 1, 3", + "[40] Rui Yu, Jiachen Liu, Zihan Zhou, and Sharon X. Huang. Nerf-enhanced outpainting for faithful field-of-view extrapolation. In IEEE International Conference on Robotics and Automation, ICRA, pages 16826-16833, 2024. 1", + "[41] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to $360^{\\circ}$ panorama image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 3, 6, 7", + "[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 3813-3824, 2023. 2, 3, 4, 6", + "[43] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pages 586-595, 2018. 6", + "[44] Dian Zheng, Cheng Zhang, Xiao-Ming Wu, Cao Li, Chengfei Lv, Jian-Fang Hu, and Wei-Shi Zheng. Panorama generation from nfov image done right. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 21610-21619, 2025. 1" + ], + "bbox": [ + 516, + 90, + 906, + 500 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28502", + "bbox": [ + 478, + 943, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/ddd3f9c4-d03d-4cd7-96eb-1ef2931311e2_model.json b/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/ddd3f9c4-d03d-4cd7-96eb-1ef2931311e2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9f6d97081c01ffefddb4b5d73284e4d97d6c5df1 --- /dev/null +++ b/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/ddd3f9c4-d03d-4cd7-96eb-1ef2931311e2_model.json @@ -0,0 +1,2937 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.124, + 0.131, + 0.877, + 0.155 + ], + "angle": 0, + "content": "Top2Pano: Learning to Generate Indoor Panoramas from Top-Down View" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.182, + 0.694, + 0.217 + ], + "angle": 0, + "content": "Zitong Zhang Suranjan Gautam Rui Yu University of Louisville" + }, + { + "type": "text", + "bbox": [ + 0.261, + 0.22, + 0.737, + 0.235 + ], + "angle": 0, + "content": "{zitong.zhang, suranjan.gautam, rui.yu}@louisville.edu" + }, + { + "type": "text", + "bbox": [ + 0.376, + 0.238, + 0.617, + 0.251 + ], + "angle": 0, + "content": "https://top2pano.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.262, + 0.167, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.346, + 0.18, + 0.357 + ], + "angle": 0, + "content": "Top-down view" + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.275, + 0.255, + 0.289 + ], + "angle": 0, + "content": "Top2Pano" + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.29, + 0.271, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.272, + 0.262, + 0.481, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.344, + 0.4, + 0.356 + ], + "angle": 0, + "content": "Output 1" + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.263, + 0.692, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.345, + 0.612, + 0.356 + ], + "angle": 0, + "content": "Output 2" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.263, + 0.905, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.345, + 0.838, + 0.356 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.37, + 0.27, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.442, + 0.194, + 0.454 + ], + "angle": 0, + "content": "Floorplan" + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.36, + 0.48, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.442, + 0.398, + 0.453 + ], + "angle": 0, + "content": "Default" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.36, + 0.692, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.442, + 0.63, + 0.453 + ], + "angle": 0, + "content": "\"Chinese Style\"" + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.36, + 0.904, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.442, + 0.835, + 0.453 + ], + "angle": 0, + "content": "\"Stone Wall\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.46, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Figure 1. Top: We present Top2Pano, a method for synthesizing high-quality indoor panoramas from a top-down view. Given a camera position, Top2Pano generates panoramas that are both visually compelling and geometrically accurate. Bottom: Our model demonstrates strong generalization capabilities. When provided with schematic floor plans as input, Top2Pano produces photorealistic and structurally coherent panoramas. Additionally, our approach can be easily adapted for stylized synthesis, enabling diverse design variations. Note: The original dataset (Matterport3D [3]) contains blurry regions near the upper and lower edges of the panoramic image." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.542, + 0.327, + 0.558 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.573, + 0.486, + 0.831 + ], + "angle": 0, + "content": "Generating immersive \\(360^{\\circ}\\) indoor panoramas from 2D top-down views has applications in virtual reality, interior design, real estate, and robotics. This task is challenging due to the lack of explicit 3D structure and the need for geometric consistency and photorealism. We propose Top2Pano, an end-to-end model for synthesizing realistic indoor panoramas from top-down views. Our method estimates volumetric occupancy to infer 3D structures, then uses volumetric rendering to generate coarse color and depth panoramas. These guide a diffusion-based refinement stage using ControlNet, enhancing realism and structural fidelity. Evaluations on two datasets show Top2Pano outperforms baselines, effectively reconstructing geometry, occlusions, and spatial arrangements. It also generalizes well, producing high-quality panoramas from schematic floorplans. Our results highlight Top2Pano's potential in bridging top-down views with immersive indoor synthesis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.222, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.9 + ], + "angle": 0, + "content": "Understanding and synthesizing immersive indoor scenes from minimal structural information is a fundamental chal" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.909, + 0.892 + ], + "angle": 0, + "content": "lenge in computer vision and graphics [16, 20, 21, 32, 40, 44]. The ability to generate realistic indoor panorama images from a 2D top-down view holds immense potential for a wide range of applications, including virtual reality (VR) [23], interior design [27], real estate visualization [15], and robotics [11]. For instance, real estate platforms can leverage this technology to offer potential buyers photorealistic virtual walkthroughs generated directly from architectural floorplans, enhancing the property viewing experience. Similarly, VR applications can benefit from automatically synthesized environments that create more engaging and immersive user experiences. Additionally, robots operating in indoor environments can utilize synthesized panoramas to improve their spatial understanding and navigation capabilities, enabling more efficient and accurate movement in complex spaces. Despite its broad applicability, the task of generating high-quality indoor panoramas from top-down views remains surprisingly underexplored in the literature. Recent advancements in large multimodal models have enabled the synthesis of panoramas directly from text input [10, 39, 41]; however, these approaches often overlook critical geometric and textural constraints. Other studies have focused on generating 3D models from semantic layouts [2, 4, 8, 9, 30, 38];" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "28493" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.093, + 0.35, + 0.106 + ], + "angle": 0, + "content": "(a) Volumetric Occupancy Estimation" + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.093, + 0.576, + 0.106 + ], + "angle": 0, + "content": "(b) Coarse Panorama Rendering" + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.093, + 0.867, + 0.106 + ], + "angle": 0, + "content": "(c) Photorealistic Panorama Synthesis" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.108, + 0.907, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.29, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed Top2Pano pipeline. The pipeline begins by segmenting the top-down view using SAM [19]. Both the segmented top-down image and the original top-down view are then processed by the OccRecon module to estimate the scene's 3D volumetric occupancy. Next, given the camera position, the system employs volumetric rendering to generate coarse depth and color panoramas. These coarse images are subsequently refined by the PanoGen module to produce the final photorealistic panorama. The PanoGen module also supports stylized panorama generation based on textual or visual conditions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.386, + 0.483, + 0.463 + ], + "angle": 0, + "content": "however, these methods are often limited by the quality of the resulting 3D meshes, making them unsuitable for rendering high-quality panoramas. Furthermore, semantic information is often unavailable in top-down views or floorplans, posing an additional challenge for existing approaches." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.463, + 0.483, + 0.644 + ], + "angle": 0, + "content": "Addressing this gap requires overcoming several significant technical challenges. First, a 2D top-down view provides only limited visual cues about the actual appearance and layout of the scene, making it difficult to infer occluded structures and fine texture details. Second, generating geometrically consistent indoor scenes demands accurate reasoning about 3D spatial occupancy from a 2D input, which is inherently ambiguous. Third, achieving photorealism while maintaining structural coherence necessitates a synthesis approach that effectively balances fidelity and realism, ensuring that the generated scenes are both visually appealing and functionally accurate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.486, + 0.902 + ], + "angle": 0, + "content": "To tackle these challenges, we introduce Top2Pano, a novel framework for generating photorealistic indoor panorama images from 2D top-down views. Our approach consists of three main stages. First, we learn the volumetric occupancy of the indoor scene, enabling the model to infer plausible spatial structures and layout configurations. Next, we employ volumetric rendering to generate coarse depth and colored panorama images, providing an initial estimate of the scene's appearance and geometry. Finally, we refine the synthesized panoramas using a diffusion-based model [42] conditioned on the coarse representations, enhancing both realism and structural consistency. By incorporating learned occupancy priors and diffusion-based refinement, our model effectively bridges the gap between schematic top-down views and immersive indoor panoramas, producing results that are both visually compelling and geometrically accurate." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.908, + 0.537 + ], + "angle": 0, + "content": "We evaluate Top2Pano on two indoor datasets and demonstrate its effectiveness compared to baseline methods. Our model not only generates higher-quality images with improved geometric consistency but also exhibits strong generalization capabilities. Even when provided with schematic floorplans as input, Top2Pano can produce photorealistic and structurally coherent panoramas. Moreover, we show that our method can be easily adapted for stylized synthesis, allowing for diverse design variations and enabling users to explore different interior aesthetics with ease." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.538, + 0.78, + 0.552 + ], + "angle": 0, + "content": "Our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.553, + 0.907, + 0.613 + ], + "angle": 0, + "content": "- We introduce Top2Pano, a novel framework for generating indoor panoramas from 2D top-down views, integrating volumetric occupancy learning, coarse synthesis, and diffusion-based refinement to achieve high-quality results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.614, + 0.907, + 0.673 + ], + "angle": 0, + "content": "- We conduct extensive experiments on two indoor datasets, demonstrating that our model surpasses baseline methods in both image quality and structural consistency, setting a new benchmark for this task." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.673, + 0.907, + 0.765 + ], + "angle": 0, + "content": "- We show that Top2Pano generalizes well to schematic floorplans, producing high-quality, geometry-consistent panoramas. Furthermore, our approach supports stylish synthesis, enabling the generation of panoramas with diverse interior design aesthetics, making it a versatile tool for various applications." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.553, + 0.907, + 0.765 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.779, + 0.655, + 0.794 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.804, + 0.722, + 0.818 + ], + "angle": 0, + "content": "2.1. Panorama Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Traditionally, panoramas were generated using image stitching and feature matching methods. With the recent advancements in generative machine learning, text-driven panorama generation techniques have gained popularity. These methods have utilized GANs, VAEs or a combina-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "28494" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.35 + ], + "angle": 0, + "content": "tion of GANs and VAEs [5] and more recently, diffusion models [10, 39, 41] to synthesize panoramic images from textual descriptions. Another popular field of research is panorama synthesis from narrow-FOV images using image out-painting. Some methods rely solely on narrow-FOV images [1], while other incorporate textual descriptions alongside the images [6, 17, 34]. Cross-view panorama generation is another well-explored area, particularly challenging due to large shifts in camera perspective. Most techniques in this domain have focused on generating ground-view panoramas from aerial images. Some approaches directly use top-down images as input [35], while others [22, 28, 31, 37] extract geometric and segmentation information from top-down images to enhance quality. To the best of our knowledge, there is no prior work that explores the generation of indoor panoramic images from floor plans or top-down views of indoor spaces." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.358, + 0.411, + 0.373 + ], + "angle": 0, + "content": "2.2. Layout-Guided 3D Scene Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.485, + 0.651 + ], + "angle": 0, + "content": "Recent approaches to 3D scene generation leverage layouts for semantic and physical plausibility. Plan2Scene [33] reconstructs 3D meshes from floor plans, while ATISS [26] employs Transformers conditioned on scene layout. CC3D [2] follows a 3D GAN-based approach using 2D semantic layouts. Diffusion-based methods such as SceneCraft [38], Layout2Scene [4], and Prim2Room [9] have further improved synthesis quality. ControlRoom3D [30] and Ctrl-Room [8] are closely related to our work, as both generate panoramas during reconstruction process. Unlike prior methods that rely on explicit semantic layouts detailing object classes, positions, orientations, and sizes, our approach only requires a top-down view – an easily obtainable and lightweight input format. Instead of generating full 3D scenes, we produce panoramas, offering a more efficient and realistic solution for AR/VR, and autonomous robotic navigation by enabling immersive experiences like virtual tours and supporting real-time robotic navigation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.182, + 0.68 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The pipeline of our Top2Pano method is illustrated in Figure 2. Given an input top-down view \\( I_{\\mathrm{top}} \\in \\mathbb{R}^{H \\times W \\times 3} \\), we first generate its segmentation map \\( I_{\\mathrm{seg}} \\in \\mathbb{R}^{H \\times W \\times 3} \\) using a pretrained model. Both the top-down view and its segmentation, along with the specified camera position, are then fed into an encoder-decoder occupancy estimation module, OccRecon, which reconstructs a 3D volumetric occupancy map \\( V_{\\mathrm{occ}} \\in \\mathbb{R}^{H \\times W \\times N} \\), where \\( N \\) represents the number of vertical voxels. From this occupancy map, we render a coarse depth panorama \\( I_{\\mathrm{depth}} \\) via volumetric rendering and project colors from the top-down view to obtain a coarse color panorama \\( I_{\\mathrm{color}} \\). To ensure geometric consistency, we enforce structural constraints on walls and floor, preserving occlusion relationships and realistic spatial structure. Finally," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.198 + ], + "angle": 0, + "content": "both coarse depth and color panoramas serve as conditions for a diffusion-based synthesis module, PanoGen, which generates photorealistic panoramic images \\( I_{\\mathrm{pano}} \\in \\mathbb{R}^{H \\times W \\times 3} \\). These images faithfully capture the scene's spatial layout, furniture, and fine color details. Additionally, PanoGen module supports stylized synthesis with optional textual or imagery-based controls." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.208, + 0.816, + 0.224 + ], + "angle": 0, + "content": "3.1. Volumetric Occupancy Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.23, + 0.907, + 0.306 + ], + "angle": 0, + "content": "The 2D top-down views lack 3D structural information about objects and furniture. To render panoramas that accurately reflect the geometric spatial relationships of objects, we propose training an OccRecon module to estimate the scene's 3D occupancy or density." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.909, + 0.553 + ], + "angle": 0, + "content": "Input Representations. Unlike the layout-guided 3D scene generation setting, our top-down input lacks semantic information. In our preliminary study, we found that current semantic segmentation models struggle to generalize to indoor top-down views, making it challenging to estimate the 3D structure without semantic guidance. To address this, we propose leveraging a pretrained segmentation model SAM [19] to extract the 2D structure of the scene. Both the top-down image and the segmentation view are then fed into the encoder of the OccRecon module. The segmentation provides valuable details, such as room boundaries, furniture positions, and shapes, which significantly enhance the OccRecon module's ability to learn the overall 3D structure of the rooms. Moreover, this semantic-free input design enables our model to generalize effectively to more abstract inputs, such as semantic floorplans." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.909, + 0.815 + ], + "angle": 0, + "content": "OccRecon Module. We propose a diffusion-based encoder-decoder framework that efficiently extracts and reconstructs the 3D spatial structure of a scene. Instead of using computationally intensive 3D diffusion models, our approach leverages a 2D diffusion model [42], significantly reducing resource demands. The OccRecon module processes 2D images and segmentations to extract spatial information and reconstruct room structures, including wall and furniture heights. In the final step, a 3D convolutional layer integrates the learned spatial features to generate a comprehensive 3D occupancy map. By relying on 2D inputs for most of the process, our method drastically lowers computational costs while still capturing the full scene layout. This balance of efficiency and accuracy makes it a practical solution for 3D structural modeling without semantic inputs. The OccRecon module outputs a 3D volumetric occupancy map \\( V_{\\mathrm{occ}} \\), which explicitly represents the scene's 3D geometry." + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.827, + 0.907, + 0.844 + ], + "angle": 0, + "content": "\\[\nV _ {\\text {o c c}} = \\text {O c c R e c o n} \\left(I _ {\\text {t o p}}, I _ {\\text {s e g}}\\right) _ {\\text {c o n d i t i o n}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.911, + 0.901 + ], + "angle": 0, + "content": "We normalize the occupancy values to the range [0, 1]. Given the scale information of the scene, we transform the real-world coordinates of a 3D point into the estimated volumetric" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.518, + 0.956 + ], + "angle": 0, + "content": "28495" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "space. The point's density is then obtained by querying the learned occupancy map \\( V_{\\mathrm{occ}} \\) using tri-linear interpolation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.127, + 0.487, + 0.43 + ], + "angle": 0, + "content": "Structural Reinforcement. Top-down views offer a comprehensive overview of the entire floor, but when observed from a first-person perspective, the scene typically reveals only the details of the current space, obscuring rooms beyond the walls. After the model learns the overall geometric structure of the rooms through the OccRecon module, we apply structural reinforcement to refine the wall geometry. This provides precise depth information, enhancing the reconstruction of the room's geometric layout. We solidify the wall voxels by applying the maximum value (1 after normalization). Additionally, top-down priors can constrain the floor's geometry and help infer the texture of the furniture, which is essential for generating accurate indoor panoramas with correctly placed and colored furniture. Since the OccNet modules are optimized end-to-end with the subsequent modules, the height of the furniture can be inferred from the 3D occupancy map. By encoding structural constraints from the walls and floor, we achieve a more accurate representation of the scene's geometry, including the positions, colors, and other attributes of the furniture." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.441, + 0.353, + 0.457 + ], + "angle": 0, + "content": "3.2. Coarse Panorama Rendering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.484, + 0.614 + ], + "angle": 0, + "content": "Given the 3D occupancy map of the scene, we render coarse depth and color images based on the specified camera position. To ensure accurate mapping, we first compute the ratio between pixel resolution and physical dimensions. With the room height known, we apply this ratio along with the pixel coordinates in the top-down image to determine the corresponding position in the occupancy map \\( V_{\\mathrm{occ}} \\). To generate the panoramic image from the 3D occupancy map, we employ equirectangular projection along with a spherical coordinate system." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.661 + ], + "angle": 0, + "content": "The coarse depth panorama \\( I_{\\mathrm{depth}} \\) is obtained through volumetric rendering [24] of the occupancy map. The depth of a projected ray at pixel \\( (u, v) \\) is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.673, + 0.484, + 0.719 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {d e p t h}} ^ {(u, v)} = \\sum_ {i = 1} ^ {S} T _ {i} \\alpha_ {i} d _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.483, + 0.759 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\) represents the transparency level, and \\(d_{i}\\) denotes the distance from the sampled position to the camera." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.484, + 0.851 + ], + "angle": 0, + "content": "The coarse color panorama \\( I_{\\mathrm{color}} \\) is obtained by directly projecting the indoor top-down view along the camera rays [28], assigning colors based on the corresponding intersections and bilinear interpolation. Specifically, the color at pixel \\( (u,v) \\) is determined by sampling the top-down image without learning a radiance field." + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.864, + 0.484, + 0.905 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {c o l o r}} ^ {(u, v)} = \\sum_ {i} ^ {S} T _ {i} \\alpha_ {i} c _ {i} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.168 + ], + "angle": 0, + "content": "where \\( c_{i} \\) is the color copied from the top-down image. This approach directly maps color from the top-down view, offering a more straightforward and computationally efficient solution compared to NeRF [25], which reconstructs scenes by learning a radiance field for view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.909, + 0.395 + ], + "angle": 0, + "content": "We employed a uniform voxel sampling strategy, where voxels were sampled along a fixed-length ray for both coarse color and coarse depth. However, this approach introduced banding artifacts, particularly noticeable on the floor directly beneath the camera in the coarse color image. Since the top-down view served as the floor texture, maintaining high-quality details was essential for achieving realistic rendering. To mitigate these artifacts, we reduced the ray length by half for coarse color sampling. This adjustment increased the density of sample points within the same spatial region, enhancing sampling accuracy and producing smoother color and texture transitions. As a result, banding artifacts were significantly diminished, improving overall rendering quality. For coarse depth, preserving scene structure was paramount, so we retained the original sampling strategy." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.403, + 0.733, + 0.419 + ], + "angle": 0, + "content": "3.3. Photorealistic Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.909, + 0.593 + ], + "angle": 0, + "content": "Generating photorealistic panoramic images directly from top-down views is challenging. To address this, we propose a two-stage pipeline. In the second stage, the PanoGen module synthesizes photorealistic indoor panoramic images from coarse color and depth inputs. We implement PanoGen using a diffusion-based ControlNet [42], enabling the restoration of fine details. This two-stage approach not only reconstructs the house's structural layout, including precise wall positions, but also restores elements such as windows, lighting, and furniture. The final panoramic image \\( I_{\\mathrm{pano}} \\) is generated based on the two coarse inputs from the previous stage:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.603, + 0.907, + 0.621 + ], + "angle": 0, + "content": "\\[\nI _ {\\text {p a n o}} = \\text {P a n o G e n} \\left(I _ {\\text {c o l o r}}, I _ {\\text {d e p t h}}\\right) _ {\\text {c o n d i t i o n}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.631, + 0.909, + 0.783 + ], + "angle": 0, + "content": "The PanoGen module modifies the conditioning mechanism of ControlNet by treating coarse color and depth images as separate inputs before combining them. This design enables PanoGen to capture both the scene's geometric structure and the furniture's color and position, resulting in more accurate and high-quality panoramic images. To further enhance consistency, we incorporate an alignment loss to prevent structural distortions when the viewpoint changes and a color loss to ensure accurate color reproduction in the synthesized output." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.791, + 0.757, + 0.807 + ], + "angle": 0, + "content": "3.4. Training and Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.813, + 0.908, + 0.859 + ], + "angle": 0, + "content": "The Top2Pano model employs denoising MSE loss, alignment loss, and color loss functions for optimization. The denoising MSE loss function [42] is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.868, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f}} = \\mathbb {E} _ {z _ {0}, t, c _ {t}, c _ {c}, c _ {d}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, c _ {t}, c _ {c}, c _ {d}\\right) \\| _ {2} ^ {2} \\right], \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.944, + 0.521, + 0.956 + ], + "angle": 0, + "content": "28496" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "where \\( t \\) denotes the number of noise addition steps, \\( c_{\\mathrm{c}} \\) represents the corresponding coarse color image, and \\( c_{\\mathrm{d}} \\) represents the coarse depth image. The variables \\( t \\) and \\( c_{t} \\) correspond to the time step and simple text prompts, respectively. The diffusion algorithm trains a neural network \\( \\epsilon_{\\theta} \\) to predict the noise added to the noisy image \\( z_{t} \\). The alignment loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.203, + 0.484, + 0.232 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a l i g n m e n t}} = \\left[ \\| I _ {D} - \\hat {I _ {D}} \\| _ {2} ^ {2} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.238, + 0.483, + 0.282 + ], + "angle": 0, + "content": "where \\( I_{D} \\) represents the depth images generated by the model and \\( \\hat{I_D} \\) represents the ground truth depth images. This loss function is used to address distortions in the furniture." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.283, + 0.484, + 0.33 + ], + "angle": 0, + "content": "Let \\(I\\) be the rendered image and \\(G\\) the ground truth image, both with \\(C = 3\\) color channels. For each channel \\(c \\in \\{1, \\ldots, C\\}\\), we define the normalized histogram as" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.338, + 0.484, + 0.358 + ], + "angle": 0, + "content": "\\[\nH ^ {c} (I) = \\left(h _ {1} ^ {c} (I), h _ {2} ^ {c} (I), \\dots , h _ {\\mathrm {b i n s}} ^ {c} (I)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.426 + ], + "angle": 0, + "content": "where bins is the number of histogram bins (e.g., 256), and each \\( h_k^c (I) \\) represents the normalized frequency (probability) of pixel intensities falling into bin \\( k \\). Likewise, for the ground truth image \\( G \\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.436, + 0.484, + 0.457 + ], + "angle": 0, + "content": "\\[\nH ^ {c} (G) = \\big (h _ {1} ^ {c} (G), h _ {2} ^ {c} (G), \\ldots , h _ {\\mathrm {b i n s}} ^ {c} (G) \\big). \\qquad (8)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.495 + ], + "angle": 0, + "content": "The color histogram loss, measured using the \\(L_{1}\\) norm, is given by" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.503, + 0.484, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {c o l o r}} (I, G) = \\sum_ {c = 1} ^ {C} \\| H ^ {c} (I) - H ^ {c} (G) \\| _ {1} \\tag {9} \\\\ = \\sum_ {c = 1} ^ {C} \\sum_ {k = 1} ^ {\\text {b i n s}} \\left| h _ {k} ^ {c} (I) - h _ {k} ^ {c} (G) \\right|. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.598, + 0.484, + 0.613 + ], + "angle": 0, + "content": "The final loss function combines three loss terms as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.624, + 0.484, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {d i f f}} + \\mathcal {L} _ {\\text {a l g n m e n t}} + \\mathcal {L} _ {\\text {c o l o r}}. \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.418, + 0.666 + ], + "angle": 0, + "content": "3.5. Generalization and Stylized Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.485, + 0.867 + ], + "angle": 0, + "content": "Generalization to Floorplan. We aim to generate first-person view panoramas that faithfully represent a scene. While top-down views lack vertical details like walls and windows, this omission grants flexibility in generating panoramas. This flexibility is especially useful in interior design, where inputs are often simple floorplans. To improve generalization, particularly to schematic floorplans, we train on orthographic rather than perspective views, as they better match floorplans. Empirical results show that Top2Pano generalizes well to schematic and even hand-drawn floorplans while maintaining photorealism. Additionally, our model enables stylized synthesis guided by text or images, supporting diverse design needs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Text-Guided Stylization. The PanoGen module, built upon the text-driven Stable Diffusion model, inherently supports" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.259 + ], + "angle": 0, + "content": "text-conditioned image generation. As illustrated in Figure 2, PanoGen synthesizes panoramas using three conditions: coarse depth, coarse colored image, and stylized textual guidance. When the input is a textureless floorplan, the stylized textual condition effectively guides the style of the synthesized result, as demonstrated in Figure 1. However, when a colored top-down view is provided, the influence of text-guided stylization becomes less pronounced. This occurs because the rendered coarse colored panorama constrains the final output to closely align with the input view, thereby diminishing the impact of the textual stylization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.909, + 0.398 + ], + "angle": 0, + "content": "To enhance text-guided stylization, the weight of the coarse colored panorama in the PanoGen conditions can be reduced. However, this introduces a tradeoff: prioritizing stylization may come at the expense of fidelity to the top-down view. Regardless of this tradeoff, the coarse depth condition consistently ensures that the synthesized panorama adheres to the underlying scene geometry. Notably, this tradeoff does not apply to the text-to-panorama generation task, where textual guidance plays a more dominant role." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.406, + 0.909, + 0.662 + ], + "angle": 0, + "content": "Image-Guided Stylization. Given several scene images (not necessarily panoramas), we can fine-tune the PanoGen module using low-rank adaptation (LoRA) [14] to generate panoramas that align with the visual styles present in the provided images. To guide this process, we introduce structured textural prompts augmented with style tags (e.g., [Japanese]) at the beginning of the input prompts. These tags act as conditional modifiers, steering the model toward synthesizing images that follow specific aesthetic themes, such as regional design styles. The framework is applied solely to the PanoGen module, ensuring both computational and parameter efficiency. Notably, our method requires fewer than five in-the-wild images per target style, substantially reducing data demands. This approach achieves its efficiency by decomposing weight updates into low-rank matrices. For a pretrained weight matrix \\(\\mathbf{W}_0 \\in \\mathbb{R}^{d \\times d}\\), the update \\(\\Delta \\mathbf{W}\\) is constrained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.679, + 0.907, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {W} = \\mathbf {B} \\mathbf {A}, \\quad \\text {w h e r e} \\mathbf {A} \\in \\mathbb {R} ^ {d \\times r}, \\mathbf {B} \\in \\mathbb {R} ^ {r \\times d} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.714, + 0.908, + 0.775 + ], + "angle": 0, + "content": "Here, \\( r \\ll d \\) represents the intrinsic rank (we use \\( r = 8 \\)). During fine-tuning, only the matrices \\( \\mathbf{A} \\) and \\( \\mathbf{B} \\) are updated, while the original weights \\( \\mathbf{W}_0 \\) remain fixed. The forward pass is modified as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.792, + 0.907, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {\\text {o u t}} = \\mathbf {W} _ {0} \\mathbf {h} _ {\\text {i n}} + \\alpha \\cdot \\mathbf {B A h} _ {\\text {i n}} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a scaling coefficient. This lightweight adaptation (0.8% new parameters) mitigates catastrophic forgetting and maintains the model's baseline generation quality for generic prompts while enabling precise style control through our [style] textual conditioning." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "28497" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.148 + ], + "angle": 0, + "content": "
TrainingTesting
ScenesFloorsPanoramasScenesFloorsPanoramas
Matterport3D61127617714291405
Gibson152203537939761672
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.153, + 0.483, + 0.181 + ], + "angle": 0, + "content": "Table 1. The numbers of scenes, floors, and panorama images in the training and testing sets of the two datasets" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.196, + 0.476, + 0.294 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑FID ↓LPIPS ↓
Materport3DSat2Density[28]+LDM[29]11.270.4221100.060.6237
Sat2Density[28]+ControlNet[42]11.420.422285.780.6163
PanFusion[41]11.450.437285.740.6153
Top2Pano (Ours)11.720.440930.840.6029
GibsonSat2Density[28]+LDM[29]10.540.448084.330.6462
Sat2Density[28]+ControlNet[42]10.970.458285.210.6645
PanFusion[41]11.360.474479.530.6634
Top2Pano (Ours)11.580.485128.680.6282
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.298, + 0.483, + 0.327 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with existing methods on the Matterport3D [3] and Gibson [36] datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.348, + 0.224, + 0.365 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.373, + 0.262, + 0.39 + ], + "angle": 0, + "content": "4.1. Data Preparation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.484, + 0.637 + ], + "angle": 0, + "content": "For evaluation, we use the Matterport3D [3] and Gibson [36] datasets. Since no existing dataset provides both top-down views and high-quality panoramic images, we generate top-down views from 3D models in these datasets using Blender. Specifically, we import textured 3D meshes into Blender and render top-down views with an orthographic camera. The top-down view we render closely resembles a floorplan, unlike the perspective-rendered views used in embodied dialog localization [12]. This similarity enhances our model's ability to generalize to floorplan inputs. To determine the number of floors in each scene, we apply DBSCAN [7] clustering to the camera positions within the datasets. We exclude certain scenes, such as airports and large supermarkets, as well as panoramic images depicting outdoor environments to ensure alignment with our task. After processing, the final dataset sizes are summarized in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.276, + 0.665 + ], + "angle": 0, + "content": "4.2. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.484, + 0.791 + ], + "angle": 0, + "content": "To assess the quality of the generated panoramas, we employ both pixel-based and perceptual evaluation metrics. For pixel-level assessment, we utilize peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) to quantify image fidelity. Additionally, we incorporate perceptual metrics such as Fréchet Inception Distance (FID) [13] and Learned Perceptual Image Patch Similarity (LPIPS) [43] to capture higher-level visual realism." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.27, + 0.819 + ], + "angle": 0, + "content": "4.3. Implement Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Our code runs on an NVIDIA RTX A6000 GPU with 48GB of memory. The model has 3.3 billion parameters and is trained with a batch size of 21 for 100 epochs. On average, each experiment takes approximately two days to complete on both the Matterport3D and Gibson datasets. We optimize" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.9, + 0.214 + ], + "angle": 0, + "content": "
ModulesMatterport3DGibson
seg floor wall depth colorPSNR↑SSIM↑FID↓LPIPS↓PSNR↑SSIM↑FID↓LPIPS↓
✓✓11.080.4122104.030.646311.070.4551117.080.6574
✓✓✓11.260.4206102.840.647111.100.4564115.880.6568
✓✓✓10.780.409155.730.616111.190.466832.280.6310
✓✓✓11.220.420754.440.619311.280.461383.310.6429
✓✓✓11.570.437843.990.606711.160.473381.870.6549
✓✓✓10.980.419653.340.611511.330.460133.890.6307
✓✓11.210.419935.080.613111.380.464136.290.6384
✓✓✓11.260.429934.610.606011.490.467334.360.6336
✓✓✓✓11.590.438167.020.619211.380.476165.660.6557
✓✓✓✓11.720.440930.840.602911.580.485128.680.6282
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.218, + 0.908, + 0.261 + ], + "angle": 0, + "content": "Table 3. Ablation study on five designs in our Top2Pano model (top-down view segmentation, floor reinforcement, wall reinforcement, coarse depth panorama, and coarse colored panorama)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.908, + 0.334 + ], + "angle": 0, + "content": "our model using the Adam optimizer [18] with default parameters \\((\\beta_{1} = 0.9, \\beta_{2} = 0.999, \\epsilon = 10^{-8})\\) and a learning rate of \\(10^{-5}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.345, + 0.826, + 0.362 + ], + "angle": 0, + "content": "4.4. Comparison with Previous Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.909, + 0.565 + ], + "angle": 0, + "content": "As shown in Table 2, we compare our method against three baseline approaches across four evaluation metrics. Sat2Density [28] is a satellite-to-ground panorama synthesis method, which we adapt for indoor panorama generation using latent diffusion model (LDM) [29] and ControlNet [42]. PanFusion [41] is a text-to-panorama generation framework that also incorporates layout-conditioned generation via ControlNet. Table 2 shows that our method outperforms all baselines across all four metrics on both datasets, demonstrating its effectiveness. Furthermore, qualitative comparisons in Figures 3 and 4 highlight that our approach generates more realistic and structurally accurate house reconstructions, including furniture placement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.667, + 0.592 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.908, + 0.719 + ], + "angle": 0, + "content": "We conducted comprehensive ablation studies to analyze and validate the contribution of each component in our model. Specifically, we performed experiments comparing several model variants against the original. These experiments involved removing key elements such as the structural reinforcement of the floor and walls, the segmentation input to the OccRecon module, and the coarse depth and colored panoramicas as conditional inputs to the PanoGen module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 3, our original model achieves the highest overall scores across all four metrics, with any modification leading to some degree of performance degradation. Removing coarse colored panoramas or the embedded floor significantly disrupts furniture placement and color accuracy. While the room structure remains mostly intact, furniture positions become unreliable, and color representations appear distorted. Conversely, excluding coarse depth panoramas or embedded walls maintains color and furniture accuracy but compromises the spatial understanding and overall quality of room structure reconstruction. These effects are further illustrated in the qualitative results in the supp. materials." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "28498" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.189, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.089, + 0.366, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.089, + 0.543, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.089, + 0.718, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.089, + 0.895, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.16, + 0.188, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.16, + 0.366, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.16, + 0.543, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.16, + 0.718, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.16, + 0.895, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.229, + 0.188, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.229, + 0.366, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.229, + 0.543, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.229, + 0.718, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.229, + 0.895, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.3, + 0.188, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.3, + 0.366, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.299, + 0.543, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.299, + 0.718, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.3, + 0.895, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.37, + 0.188, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.369, + 0.366, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.369, + 0.543, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.369, + 0.718, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.369, + 0.895, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.439, + 0.182, + 0.452 + ], + "angle": 0, + "content": "(a) Top-down" + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.439, + 0.333, + 0.452 + ], + "angle": 0, + "content": "(b) Sat2Density [28]" + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.439, + 0.505, + 0.452 + ], + "angle": 0, + "content": "(c) PanFusion [41]" + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.439, + 0.655, + 0.451 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.439, + 0.853, + 0.451 + ], + "angle": 0, + "content": "(e) Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.457, + 0.687, + 0.471 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons on the Matterport3D dataset." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.485, + 0.189, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.485, + 0.366, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.485, + 0.543, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.485, + 0.718, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.485, + 0.895, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.555, + 0.189, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.555, + 0.366, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.555, + 0.543, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.555, + 0.718, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.555, + 0.895, + 0.623 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.625, + 0.189, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.625, + 0.366, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.625, + 0.543, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.625, + 0.718, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.625, + 0.895, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.694, + 0.189, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.694, + 0.366, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.694, + 0.543, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.694, + 0.718, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.694, + 0.895, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.764, + 0.189, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.835, + 0.182, + 0.847 + ], + "angle": 0, + "content": "(a) Top-down" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.764, + 0.366, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.223, + 0.835, + 0.333, + 0.847 + ], + "angle": 0, + "content": "(b) Sat2Density [28]" + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.764, + 0.543, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.835, + 0.505, + 0.847 + ], + "angle": 0, + "content": "(c) PanFusion [41]" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.764, + 0.718, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.835, + 0.655, + 0.847 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.764, + 0.895, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.835, + 0.853, + 0.847 + ], + "angle": 0, + "content": "(e) Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.852, + 0.667, + 0.866 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons on the Gibson dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "28499" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.133, + 0.089, + 0.44, + 0.141 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑FID ↓LPIPS ↓
G → M11.080.439740.740.6285
M → G11.030.436645.870.6353
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.145, + 0.486, + 0.172 + ], + "angle": 0, + "content": "Table 4. Cross-dataset evaluation. \"G\" represents Gibson; \"M\" represents Matterport3D." + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.176, + 0.443, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.331, + 0.485, + 0.372 + ], + "angle": 0, + "content": "Figure 5. Our Top2Pano model generalizes to both textured floorplans (first row) and hand-drawn sketched floorplans (second row), incorporating stylized [Japanese] control." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.38, + 0.445, + 0.396 + ], + "angle": 0, + "content": "4.6. Generalization, Stylization, Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.492 + ], + "angle": 0, + "content": "We employed cross-dataset evaluation to assess our model's generalization capability. As shown in Table 4, although there is a slight decline in metric scores, our model maintains strong performance. Additionally, the trained parameters tend to generate photorealistic walls that reflect characteristics of the training dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Our model also generalizes well to floorplans. When given a specific floorplan and camera positions, the model assists users in generating indoor panoramic images. By using different text prompts, users can create various interior design styles, explore the house from a first-person perspective, and modify its style according to their preferences. We tested our model with three types of floorplans. The first type is a colored floorplan (Figure 5, first row), which provides detailed information about room colors and furniture hues, making it highly informative. The second type is a plain floorplan (Figure 1, bottom), which lacks color information and shows only the room structure and furniture layout. The third type is a hand-drawn floorplan sketch (Figure 5, first row), which offers a rough visual representation of the room. Our model successfully generates accurate panoramic images from these floorplans and adapts the visual style based on the provided textual descriptions. Moreover, our model enables panorama manipulation by editing objects in the floorplan, such as adding new items, as illustrated in Figure 6." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.217, + 0.818 + ], + "angle": 0, + "content": "4.7. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.855 + ], + "angle": 0, + "content": "Failure cases. Figure 7 shows representative failure cases. We annotate different types of failures with numbered labels:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.476, + 0.871 + ], + "angle": 0, + "content": "- Ceiling: ① missing fan, ② vaulted ceiling, ③ false light;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.453, + 0.885 + ], + "angle": 0, + "content": "- Wall: 4 height error, 5 false or missing decorations;" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.279, + 0.9 + ], + "angle": 0, + "content": "- Window: 6 false window;" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.856, + 0.476, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.553, + 0.089, + 0.867, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.244, + 0.908, + 0.3 + ], + "angle": 0, + "content": "Figure 6. Top2Pano enables panorama manipulation via compositional floorplan editing. In the second row, adding a rectangular object to the floorplan (compared to the first row) leads the model to generate a washstand with a mirror in the panorama." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.303, + 0.913, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.508, + 0.868, + 0.523 + ], + "angle": 0, + "content": "Figure 7. Failure cases (zoom in to view error types)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.53, + 0.703, + 0.545 + ], + "angle": 0, + "content": "- Furniture: height error;" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.546, + 0.784, + 0.56 + ], + "angle": 0, + "content": "- Thin object: 8 missing flat-screen TV;" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.56, + 0.692, + 0.574 + ], + "angle": 0, + "content": "Stairs: false direction." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.53, + 0.784, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.906, + 0.65 + ], + "angle": 0, + "content": "The failure cases stem from the ambiguity of the 2D input, leading to hallucinated objects that are not observable from the top-down view. The limitations in handling vertical structural details are largely due to the inherent ambiguity of the task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.906, + 0.715 + ], + "angle": 0, + "content": "Limited Vertical FoV. The generated panoramas exhibit a limited vertical field of view (FoV), reflecting the constraints of the training data. We expect improved performance with future datasets that include full vertical FoV panoramas." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.64, + 0.743 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We present Top2Pano, a novel method for generating high-quality \\(360^{\\circ}\\) indoor panoramas from 2D top-down views. The model first estimates volumetric occupancy to infer 3D structure, then applies volumetric rendering to produce coarse color and depth panoramas. These guide a diffusion-based refinement stage via ControlNet. To our knowledge, this is the first approach to generate panoramas from top-down views. Experiments on two datasets show that Top2Pano outperforms baselines in reconstructing room layouts and realistic furniture." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.519, + 0.956 + ], + "angle": 0, + "content": "28500" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.184 + ], + "angle": 0, + "content": "[1] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. Diverse plausible 360-degree image outpainting for efficient 3dcg background creation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 11431-11440, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.485, + 0.268 + ], + "angle": 0, + "content": "[2] Sherwin Bahmani, Jeong Joon Park, Despoina Paschalidou, Xingguang Yan, Gordon Wetzstein, Leonidas J. Guibas, and Andrea Tagliasacchi. CC3D: layout-conditioned generation of compositional 3d scenes. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 7137-7147, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.269, + 0.484, + 0.338 + ], + "angle": 0, + "content": "[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.34, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[4] Minglin Chen, Longguang Wang, Sheng Ao, Ye Zhang, Kai Xu, and Yulan Guo. Layout2scene: 3d semantic layout guided scene generation via geometry and appearance diffusion priors. arXiv preprint arXiv:2501.02519, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.484, + 0.437 + ], + "angle": 0, + "content": "[5] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.484, + 0.507 + ], + "angle": 0, + "content": "[6] Mohammad Reza Karimi Dastjerdi, Yannick Hold-Geoffroy, Jonathan Eisenmann, Siavash Khodadadeh, and Jean-François Lalonde. Guided co-modulated gan for \\(360^{\\circ}\\) field of view extrapolation. In 2022 International Conference on 3D Vision (3DV), pages 475–485, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.484, + 0.577 + ], + "angle": 0, + "content": "[7] Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD), pages 226-231, 1996. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.484, + 0.632 + ], + "angle": 0, + "content": "[8] Chuan Fang, Xiaotao Hu, Kunming Luo, and Ping Tan. Ctrlroom: Controllable text-to-3d room meshes generation with layout constraints. arXiv preprint arXiv:2310.03602, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.69 + ], + "angle": 0, + "content": "[9] Chengzeng Feng, Jiacheng Wei, Cheng Chen, Yang Li, Pan Ji, Fayao Liu, Hongdong Li, and Guosheng Lin. Prim2room: Layout-controllable room mesh generation from primitives. arXiv preprint arXiv:2409.05380, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[10] Mengyang Feng, Jinlin Liu, Miaomiao Cui, and Xuansong Xie. Diffusion360: Seamless 360 degree panoramic image generation based on diffusion models. arXiv preprint arXiv:2311.13141, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[11] Julia Guerrero-Viu, Clara Fernandez-Labrador, Cédric Demonceaux, and José Jesús Guerrero. What's in my room? object recognition on indoor panoramic images. In IEEE International Conference on Robotics and Automation, ICRA, pages 567-573, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[12] Meera Hahn, Jacob Krantz, Dhruv Batra, Devi Parikh, James M Rehg, Stefan Lee, and Peter Anderson. Where are you? localization from embodied dialog. arXiv preprint arXiv:2011.08277, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.908, + 0.135 + ], + "angle": 0, + "content": "time-scale update rule converge to a local nash equilibrium. In Annual Conference on Neural Information Processing Systems, pages 6626-6637, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.908, + 0.204 + ], + "angle": 0, + "content": "[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, ICLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.907, + 0.289 + ], + "angle": 0, + "content": "[15] Zilong Huang, Jun He, Junyan Ye, Lihan Jiang, Weijia Li, Yiping Chen, and Ting Han. Scene4u: Hierarchical layered 3d scene reconstruction from single panoramic image for your immerse exploration. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 26723-26733, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.908, + 0.373 + ], + "angle": 0, + "content": "[16] Zhe Huang, Yizhe Zhao, Hao Xiao, Chenyan Wu, and Lingting Ge. Duospacenet: Leveraging both bird's-eye-view and perspective view representations for 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops, pages 2560-2570, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.376, + 0.907, + 0.445 + ], + "angle": 0, + "content": "[17] Nikolai Kalischek, Michael Oechsle, Fabian Manhardt, Philipp Henzler, Konrad Schindler, and Federico Tombari. Cubediff: Repurposing diffusion-based image models for panorama generation. In International Conference on Learning Representations, ICLR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.906, + 0.488 + ], + "angle": 0, + "content": "[18] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, ICLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.908, + 0.572 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 3992-4003, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.575, + 0.907, + 0.628 + ], + "angle": 0, + "content": "[20] Jiachen Liu, Yuan Xue, Haomiao Ni, Rui Yu, Zihan Zhou, and Sharon X Huang. Computer-aided layout generation for building design: A review. arXiv preprint arXiv:2504.09694, 2025.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.632, + 0.907, + 0.699 + ], + "angle": 0, + "content": "[21] Jiachen Liu, Rui Yu, Sili Chen, Sharon X. Huang, and Hengkai Guo. Towards in-the-wild 3d plane reconstruction from a single image. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 27027-27037, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.908, + 0.771 + ], + "angle": 0, + "content": "[22] Xiaohu Lu, Zuoyue Li, Zhaopeng Cui, Martin R. Oswald, Marc Pollefeys, and Rongjun Qin. Geometry-aware satellite-to-ground image synthesis for urban areas. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 856-864, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.907, + 0.829 + ], + "angle": 0, + "content": "[23] Wentao Lyu, Peng Ding, Yingliang Zhang, Anpei Chen, Minye Wu, Shu Yin, and Jingyi Yu. Refocusable gigapixel panoramas for immersive VR experiences. IEEE Trans. Vis. Comput. Graph., 27(3):2028-2040, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[24] Nelson L. Max. Optical models for direct volume rendering. IEEE Trans. Vis. Comput. Graph., 1(2):99-108, 1995. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.518, + 0.955 + ], + "angle": 0, + "content": "28501" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "thesis. In European Conference on Computer Vision, ECCV, pages 405-421, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[26] Despoina Paschalidou, Amlan Kar, Maria Shugrina, Karsten Kreis, Andreas Geiger, and Sanja Fidler. Atiss: Autoregressive transformers for indoor scene synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.234 + ], + "angle": 0, + "content": "[27] Giovanni Pintore, Fabio Bettio, Marco Agus, and Enrico Gobbetti. Deep scene synthesis of atlanta-world interiors from a single omnidirectional image. IEEE Trans. Vis. Comput. Graph., 29(11):4708-4718, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.484, + 0.291 + ], + "angle": 0, + "content": "[28] Ming Qian, Jincheng Xiong, Gui-Song Xia, and Nan Xue. Sat2density: Faithful density learning from satellite-ground image pairs. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 3660-3669, 2023. 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.484, + 0.361 + ], + "angle": 0, + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 10674-10685, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.484, + 0.446 + ], + "angle": 0, + "content": "[30] Jonas Schult, Sam Tsai, Lukas Hollein, Bichen Wu, Jialiang Wang, Chih-Yao Ma, Kunpeng Li, Xiaofang Wang, Felix Wimbauer, Zijian He, Peizhao Zhang, Bastian Leibe, Peter Vajda, and Ji Hou. Controlroom3d: Room generation using semantic proxy rooms. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[31] Yujiao Shi, Dylan Campbell, Xin Yu, and Hongdong Li. Geometry-guided street-view panorama synthesis from satellite imagery. IEEE Trans. Pattern Anal. Mach. Intell., 44(12): 10009-10022, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[32] Bin Tan, Rui Yu, Yujun Shen, and Nan Xue. Planarsplatting: Accurate planar surface reconstruction in 3 minutes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 1190-1199, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[33] Madhawa Vidanapathirana, Qirui Wu, Yasutaka Furukawa, Angel X. Chang, and Manolis Savva. Plan2scene: Converting floorplans to 3d scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 10733-10742, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[34] Jionghao Wang, Ziyu Chen, Jun Ling, Rong Xie, and Li Song. 360-degree panorama generation from few unregistered nfov images. In Proceedings of the 31st ACM International Conference on Multimedia, MM, pages 6811-6821, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.484, + 0.742 + ], + "angle": 0, + "content": "[35] Songsong Wu, Hao Tang, Xiao-Yuan Jing, Haifeng Zhao, Jianjun Qian, Nicu Sebe, and Yan Yan. Cross-view panorama image synthesis. IEEE Trans. Multim., 25:3546-3559, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[36] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: real-world perception for embodied agents. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[37] Ningli Xu and Rongjun Qin. Geospecific view generation geometry-context aware high-resolution ground view inference from satellite views. In European Conference on Computer Vision, ECCV, pages 349-366, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[38] Xiuyu Yang, Yunze Man, Jun-Kun Chen, and Yu-Xiong Wang. Scenecraft: Layout-guided 3d scene generation. In Advances in Neural Information Processing Systems, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[39] Weicai Ye, Chenhao Ji, Zheng Chen, Junyao Gao, Xiaoshui Huang, Song-Hai Zhang, Wanli Ouyang, Tong He, Cairong Zhao, and Guofeng Zhang. Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion. In Annual Conference on Neural Information Processing Systems, NeurIPS, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.908, + 0.232 + ], + "angle": 0, + "content": "[40] Rui Yu, Jiachen Liu, Zihan Zhou, and Sharon X. Huang. Nerf-enhanced outpainting for faithful field-of-view extrapolation. In IEEE International Conference on Robotics and Automation, ICRA, pages 16826-16833, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.908, + 0.302 + ], + "angle": 0, + "content": "[41] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to \\(360^{\\circ}\\) panorama image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.908, + 0.359 + ], + "angle": 0, + "content": "[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 3813-3824, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.908, + 0.429 + ], + "angle": 0, + "content": "[43] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.431, + 0.908, + 0.5 + ], + "angle": 0, + "content": "[44] Dian Zheng, Cheng Zhang, Xiao-Ming Wu, Cao Li, Chengfei Lv, Jian-Fang Hu, and Wei-Shi Zheng. Panorama generation from nfov image done right. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 21610-21619, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.944, + 0.52, + 0.956 + ], + "angle": 0, + "content": "28502" + } + ] +] \ No newline at end of file diff --git a/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/ddd3f9c4-d03d-4cd7-96eb-1ef2931311e2_origin.pdf b/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/ddd3f9c4-d03d-4cd7-96eb-1ef2931311e2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8123793e40b0ff40ad1ac38865266bd9c09e73ad --- /dev/null +++ b/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/ddd3f9c4-d03d-4cd7-96eb-1ef2931311e2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bb4614a90f89a2da5a1fef962328394d13703987ade231d6c93ad44324b3cb6d +size 12268179 diff --git a/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/full.md b/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f2ee3976fae2d26b56eff854a20cc51cf8c2b43f --- /dev/null +++ b/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/full.md @@ -0,0 +1,438 @@ +# Top2Pano: Learning to Generate Indoor Panoramas from Top-Down View + +Zitong Zhang Suranjan Gautam Rui Yu University of Louisville + +{zitong.zhang, suranjan.gautam, rui.yu}@louisville.edu + +https://top2pano.github.io/ + +![](images/87224dcd1f0f73f69582edfd7f31a0c55d8cd50f55ccf841fd2f52435aa16045.jpg) +Top-down view + +![](images/735e1e120e15d3c20da8305349706970f7009c55167331b92e15b05c5aac50f6.jpg) +Top2Pano + +![](images/ec1b6120b05d374645c2eecb4d9b4651a0606aac0e8d229920a75de6a7588942.jpg) +Output 1 + +![](images/725d35ef4e56b533ccc99ba1b5887ba31087e881a160f35a9ac35bd23c67f706.jpg) +Output 2 + +![](images/ff0842f581694a85ce859498055eee298b328dabe8060f2a95f1f2a14f845357.jpg) +Ground Truth + +![](images/aeb079062defbd1466a8935594ae1bb340a8600929d40752fdf9aea4c059d5ba.jpg) +Floorplan +Figure 1. Top: We present Top2Pano, a method for synthesizing high-quality indoor panoramas from a top-down view. Given a camera position, Top2Pano generates panoramas that are both visually compelling and geometrically accurate. Bottom: Our model demonstrates strong generalization capabilities. When provided with schematic floor plans as input, Top2Pano produces photorealistic and structurally coherent panoramas. Additionally, our approach can be easily adapted for stylized synthesis, enabling diverse design variations. Note: The original dataset (Matterport3D [3]) contains blurry regions near the upper and lower edges of the panoramic image. + +![](images/92074427435354c982e115ec357e4ef9b8aeeb3fbd374e657c29893892d2ee9c.jpg) +Default + +![](images/165a518e082ea816a5ffe3c95a493f513fe530e02f5812876f6926f2b2e32f23.jpg) +"Chinese Style" + +![](images/205b4f0aa706d48aba89f5609ed437fe16f30b53af5bd772c92b61d0a7e623d1.jpg) +"Stone Wall" + +# Abstract + +Generating immersive $360^{\circ}$ indoor panoramas from 2D top-down views has applications in virtual reality, interior design, real estate, and robotics. This task is challenging due to the lack of explicit 3D structure and the need for geometric consistency and photorealism. We propose Top2Pano, an end-to-end model for synthesizing realistic indoor panoramas from top-down views. Our method estimates volumetric occupancy to infer 3D structures, then uses volumetric rendering to generate coarse color and depth panoramas. These guide a diffusion-based refinement stage using ControlNet, enhancing realism and structural fidelity. Evaluations on two datasets show Top2Pano outperforms baselines, effectively reconstructing geometry, occlusions, and spatial arrangements. It also generalizes well, producing high-quality panoramas from schematic floorplans. Our results highlight Top2Pano's potential in bridging top-down views with immersive indoor synthesis. + +# 1. Introduction + +Understanding and synthesizing immersive indoor scenes from minimal structural information is a fundamental chal + +lenge in computer vision and graphics [16, 20, 21, 32, 40, 44]. The ability to generate realistic indoor panorama images from a 2D top-down view holds immense potential for a wide range of applications, including virtual reality (VR) [23], interior design [27], real estate visualization [15], and robotics [11]. For instance, real estate platforms can leverage this technology to offer potential buyers photorealistic virtual walkthroughs generated directly from architectural floorplans, enhancing the property viewing experience. Similarly, VR applications can benefit from automatically synthesized environments that create more engaging and immersive user experiences. Additionally, robots operating in indoor environments can utilize synthesized panoramas to improve their spatial understanding and navigation capabilities, enabling more efficient and accurate movement in complex spaces. Despite its broad applicability, the task of generating high-quality indoor panoramas from top-down views remains surprisingly underexplored in the literature. Recent advancements in large multimodal models have enabled the synthesis of panoramas directly from text input [10, 39, 41]; however, these approaches often overlook critical geometric and textural constraints. Other studies have focused on generating 3D models from semantic layouts [2, 4, 8, 9, 30, 38]; + +![](images/800c677146bc589beee6fcb77f6e408d9d46964197f4e1dcf18574abd727b4a3.jpg) +(a) Volumetric Occupancy Estimation +(b) Coarse Panorama Rendering +(c) Photorealistic Panorama Synthesis +Figure 2. Overview of the proposed Top2Pano pipeline. The pipeline begins by segmenting the top-down view using SAM [19]. Both the segmented top-down image and the original top-down view are then processed by the OccRecon module to estimate the scene's 3D volumetric occupancy. Next, given the camera position, the system employs volumetric rendering to generate coarse depth and color panoramas. These coarse images are subsequently refined by the PanoGen module to produce the final photorealistic panorama. The PanoGen module also supports stylized panorama generation based on textual or visual conditions. + +however, these methods are often limited by the quality of the resulting 3D meshes, making them unsuitable for rendering high-quality panoramas. Furthermore, semantic information is often unavailable in top-down views or floorplans, posing an additional challenge for existing approaches. + +Addressing this gap requires overcoming several significant technical challenges. First, a 2D top-down view provides only limited visual cues about the actual appearance and layout of the scene, making it difficult to infer occluded structures and fine texture details. Second, generating geometrically consistent indoor scenes demands accurate reasoning about 3D spatial occupancy from a 2D input, which is inherently ambiguous. Third, achieving photorealism while maintaining structural coherence necessitates a synthesis approach that effectively balances fidelity and realism, ensuring that the generated scenes are both visually appealing and functionally accurate. + +To tackle these challenges, we introduce Top2Pano, a novel framework for generating photorealistic indoor panorama images from 2D top-down views. Our approach consists of three main stages. First, we learn the volumetric occupancy of the indoor scene, enabling the model to infer plausible spatial structures and layout configurations. Next, we employ volumetric rendering to generate coarse depth and colored panorama images, providing an initial estimate of the scene's appearance and geometry. Finally, we refine the synthesized panoramas using a diffusion-based model [42] conditioned on the coarse representations, enhancing both realism and structural consistency. By incorporating learned occupancy priors and diffusion-based refinement, our model effectively bridges the gap between schematic top-down views and immersive indoor panoramas, producing results that are both visually compelling and geometrically accurate. + +We evaluate Top2Pano on two indoor datasets and demonstrate its effectiveness compared to baseline methods. Our model not only generates higher-quality images with improved geometric consistency but also exhibits strong generalization capabilities. Even when provided with schematic floorplans as input, Top2Pano can produce photorealistic and structurally coherent panoramas. Moreover, we show that our method can be easily adapted for stylized synthesis, allowing for diverse design variations and enabling users to explore different interior aesthetics with ease. + +Our key contributions are as follows: + +- We introduce Top2Pano, a novel framework for generating indoor panoramas from 2D top-down views, integrating volumetric occupancy learning, coarse synthesis, and diffusion-based refinement to achieve high-quality results. +- We conduct extensive experiments on two indoor datasets, demonstrating that our model surpasses baseline methods in both image quality and structural consistency, setting a new benchmark for this task. +- We show that Top2Pano generalizes well to schematic floorplans, producing high-quality, geometry-consistent panoramas. Furthermore, our approach supports stylish synthesis, enabling the generation of panoramas with diverse interior design aesthetics, making it a versatile tool for various applications. + +# 2. Related Work + +# 2.1. Panorama Generation + +Traditionally, panoramas were generated using image stitching and feature matching methods. With the recent advancements in generative machine learning, text-driven panorama generation techniques have gained popularity. These methods have utilized GANs, VAEs or a combina- + +tion of GANs and VAEs [5] and more recently, diffusion models [10, 39, 41] to synthesize panoramic images from textual descriptions. Another popular field of research is panorama synthesis from narrow-FOV images using image out-painting. Some methods rely solely on narrow-FOV images [1], while other incorporate textual descriptions alongside the images [6, 17, 34]. Cross-view panorama generation is another well-explored area, particularly challenging due to large shifts in camera perspective. Most techniques in this domain have focused on generating ground-view panoramas from aerial images. Some approaches directly use top-down images as input [35], while others [22, 28, 31, 37] extract geometric and segmentation information from top-down images to enhance quality. To the best of our knowledge, there is no prior work that explores the generation of indoor panoramic images from floor plans or top-down views of indoor spaces. + +# 2.2. Layout-Guided 3D Scene Generation + +Recent approaches to 3D scene generation leverage layouts for semantic and physical plausibility. Plan2Scene [33] reconstructs 3D meshes from floor plans, while ATISS [26] employs Transformers conditioned on scene layout. CC3D [2] follows a 3D GAN-based approach using 2D semantic layouts. Diffusion-based methods such as SceneCraft [38], Layout2Scene [4], and Prim2Room [9] have further improved synthesis quality. ControlRoom3D [30] and Ctrl-Room [8] are closely related to our work, as both generate panoramas during reconstruction process. Unlike prior methods that rely on explicit semantic layouts detailing object classes, positions, orientations, and sizes, our approach only requires a top-down view – an easily obtainable and lightweight input format. Instead of generating full 3D scenes, we produce panoramas, offering a more efficient and realistic solution for AR/VR, and autonomous robotic navigation by enabling immersive experiences like virtual tours and supporting real-time robotic navigation. + +# 3. Method + +The pipeline of our Top2Pano method is illustrated in Figure 2. Given an input top-down view $I_{\mathrm{top}} \in \mathbb{R}^{H \times W \times 3}$ , we first generate its segmentation map $I_{\mathrm{seg}} \in \mathbb{R}^{H \times W \times 3}$ using a pretrained model. Both the top-down view and its segmentation, along with the specified camera position, are then fed into an encoder-decoder occupancy estimation module, OccRecon, which reconstructs a 3D volumetric occupancy map $V_{\mathrm{occ}} \in \mathbb{R}^{H \times W \times N}$ , where $N$ represents the number of vertical voxels. From this occupancy map, we render a coarse depth panorama $I_{\mathrm{depth}}$ via volumetric rendering and project colors from the top-down view to obtain a coarse color panorama $I_{\mathrm{color}}$ . To ensure geometric consistency, we enforce structural constraints on walls and floor, preserving occlusion relationships and realistic spatial structure. Finally, + +both coarse depth and color panoramas serve as conditions for a diffusion-based synthesis module, PanoGen, which generates photorealistic panoramic images $I_{\mathrm{pano}} \in \mathbb{R}^{H \times W \times 3}$ . These images faithfully capture the scene's spatial layout, furniture, and fine color details. Additionally, PanoGen module supports stylized synthesis with optional textual or imagery-based controls. + +# 3.1. Volumetric Occupancy Estimation + +The 2D top-down views lack 3D structural information about objects and furniture. To render panoramas that accurately reflect the geometric spatial relationships of objects, we propose training an OccRecon module to estimate the scene's 3D occupancy or density. + +Input Representations. Unlike the layout-guided 3D scene generation setting, our top-down input lacks semantic information. In our preliminary study, we found that current semantic segmentation models struggle to generalize to indoor top-down views, making it challenging to estimate the 3D structure without semantic guidance. To address this, we propose leveraging a pretrained segmentation model SAM [19] to extract the 2D structure of the scene. Both the top-down image and the segmentation view are then fed into the encoder of the OccRecon module. The segmentation provides valuable details, such as room boundaries, furniture positions, and shapes, which significantly enhance the OccRecon module's ability to learn the overall 3D structure of the rooms. Moreover, this semantic-free input design enables our model to generalize effectively to more abstract inputs, such as semantic floorplans. + +OccRecon Module. We propose a diffusion-based encoder-decoder framework that efficiently extracts and reconstructs the 3D spatial structure of a scene. Instead of using computationally intensive 3D diffusion models, our approach leverages a 2D diffusion model [42], significantly reducing resource demands. The OccRecon module processes 2D images and segmentations to extract spatial information and reconstruct room structures, including wall and furniture heights. In the final step, a 3D convolutional layer integrates the learned spatial features to generate a comprehensive 3D occupancy map. By relying on 2D inputs for most of the process, our method drastically lowers computational costs while still capturing the full scene layout. This balance of efficiency and accuracy makes it a practical solution for 3D structural modeling without semantic inputs. The OccRecon module outputs a 3D volumetric occupancy map $V_{\mathrm{occ}}$ , which explicitly represents the scene's 3D geometry. + +$$ +V _ {\text {o c c}} = \text {O c c R e c o n} \left(I _ {\text {t o p}}, I _ {\text {s e g}}\right) _ {\text {c o n d i t i o n}} \tag {1} +$$ + +We normalize the occupancy values to the range [0, 1]. Given the scale information of the scene, we transform the real-world coordinates of a 3D point into the estimated volumetric + +space. The point's density is then obtained by querying the learned occupancy map $V_{\mathrm{occ}}$ using tri-linear interpolation. + +Structural Reinforcement. Top-down views offer a comprehensive overview of the entire floor, but when observed from a first-person perspective, the scene typically reveals only the details of the current space, obscuring rooms beyond the walls. After the model learns the overall geometric structure of the rooms through the OccRecon module, we apply structural reinforcement to refine the wall geometry. This provides precise depth information, enhancing the reconstruction of the room's geometric layout. We solidify the wall voxels by applying the maximum value (1 after normalization). Additionally, top-down priors can constrain the floor's geometry and help infer the texture of the furniture, which is essential for generating accurate indoor panoramas with correctly placed and colored furniture. Since the OccNet modules are optimized end-to-end with the subsequent modules, the height of the furniture can be inferred from the 3D occupancy map. By encoding structural constraints from the walls and floor, we achieve a more accurate representation of the scene's geometry, including the positions, colors, and other attributes of the furniture. + +# 3.2. Coarse Panorama Rendering + +Given the 3D occupancy map of the scene, we render coarse depth and color images based on the specified camera position. To ensure accurate mapping, we first compute the ratio between pixel resolution and physical dimensions. With the room height known, we apply this ratio along with the pixel coordinates in the top-down image to determine the corresponding position in the occupancy map $V_{\mathrm{occ}}$ . To generate the panoramic image from the 3D occupancy map, we employ equirectangular projection along with a spherical coordinate system. + +The coarse depth panorama $I_{\mathrm{depth}}$ is obtained through volumetric rendering [24] of the occupancy map. The depth of a projected ray at pixel $(u, v)$ is computed as follows: + +$$ +I _ {\mathrm {d e p t h}} ^ {(u, v)} = \sum_ {i = 1} ^ {S} T _ {i} \alpha_ {i} d _ {i}, \quad T _ {i} = \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \qquad (2) +$$ + +where $\alpha_{i}$ represents the transparency level, and $d_{i}$ denotes the distance from the sampled position to the camera. + +The coarse color panorama $I_{\mathrm{color}}$ is obtained by directly projecting the indoor top-down view along the camera rays [28], assigning colors based on the corresponding intersections and bilinear interpolation. Specifically, the color at pixel $(u,v)$ is determined by sampling the top-down image without learning a radiance field. + +$$ +I _ {\mathrm {c o l o r}} ^ {(u, v)} = \sum_ {i} ^ {S} T _ {i} \alpha_ {i} c _ {i} \tag {3} +$$ + +where $c_{i}$ is the color copied from the top-down image. This approach directly maps color from the top-down view, offering a more straightforward and computationally efficient solution compared to NeRF [25], which reconstructs scenes by learning a radiance field for view synthesis. + +We employed a uniform voxel sampling strategy, where voxels were sampled along a fixed-length ray for both coarse color and coarse depth. However, this approach introduced banding artifacts, particularly noticeable on the floor directly beneath the camera in the coarse color image. Since the top-down view served as the floor texture, maintaining high-quality details was essential for achieving realistic rendering. To mitigate these artifacts, we reduced the ray length by half for coarse color sampling. This adjustment increased the density of sample points within the same spatial region, enhancing sampling accuracy and producing smoother color and texture transitions. As a result, banding artifacts were significantly diminished, improving overall rendering quality. For coarse depth, preserving scene structure was paramount, so we retained the original sampling strategy. + +# 3.3. Photorealistic Synthesis + +Generating photorealistic panoramic images directly from top-down views is challenging. To address this, we propose a two-stage pipeline. In the second stage, the PanoGen module synthesizes photorealistic indoor panoramic images from coarse color and depth inputs. We implement PanoGen using a diffusion-based ControlNet [42], enabling the restoration of fine details. This two-stage approach not only reconstructs the house's structural layout, including precise wall positions, but also restores elements such as windows, lighting, and furniture. The final panoramic image $I_{\mathrm{pano}}$ is generated based on the two coarse inputs from the previous stage: + +$$ +I _ {\text {p a n o}} = \text {P a n o G e n} \left(I _ {\text {c o l o r}}, I _ {\text {d e p t h}}\right) _ {\text {c o n d i t i o n}} \tag {4} +$$ + +The PanoGen module modifies the conditioning mechanism of ControlNet by treating coarse color and depth images as separate inputs before combining them. This design enables PanoGen to capture both the scene's geometric structure and the furniture's color and position, resulting in more accurate and high-quality panoramic images. To further enhance consistency, we incorporate an alignment loss to prevent structural distortions when the viewpoint changes and a color loss to ensure accurate color reproduction in the synthesized output. + +# 3.4. Training and Optimization + +The Top2Pano model employs denoising MSE loss, alignment loss, and color loss functions for optimization. The denoising MSE loss function [42] is defined as: + +$$ +\mathcal {L} _ {\text {d i f f}} = \mathbb {E} _ {z _ {0}, t, c _ {t}, c _ {c}, c _ {d}, \epsilon \sim \mathcal {N} (0, 1)} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, c _ {t}, c _ {c}, c _ {d}\right) \| _ {2} ^ {2} \right], \tag {5} +$$ + +where $t$ denotes the number of noise addition steps, $c_{\mathrm{c}}$ represents the corresponding coarse color image, and $c_{\mathrm{d}}$ represents the coarse depth image. The variables $t$ and $c_{t}$ correspond to the time step and simple text prompts, respectively. The diffusion algorithm trains a neural network $\epsilon_{\theta}$ to predict the noise added to the noisy image $z_{t}$ . The alignment loss is formulated as: + +$$ +\mathcal {L} _ {\mathrm {a l i g n m e n t}} = \left[ \| I _ {D} - \hat {I _ {D}} \| _ {2} ^ {2} \right], \tag {6} +$$ + +where $I_{D}$ represents the depth images generated by the model and $\hat{I_D}$ represents the ground truth depth images. This loss function is used to address distortions in the furniture. + +Let $I$ be the rendered image and $G$ the ground truth image, both with $C = 3$ color channels. For each channel $c \in \{1, \ldots, C\}$ , we define the normalized histogram as + +$$ +H ^ {c} (I) = \left(h _ {1} ^ {c} (I), h _ {2} ^ {c} (I), \dots , h _ {\mathrm {b i n s}} ^ {c} (I)\right), \tag {7} +$$ + +where bins is the number of histogram bins (e.g., 256), and each $h_k^c (I)$ represents the normalized frequency (probability) of pixel intensities falling into bin $k$ . Likewise, for the ground truth image $G$ , we have + +$$ +H ^ {c} (G) = \big (h _ {1} ^ {c} (G), h _ {2} ^ {c} (G), \ldots , h _ {\mathrm {b i n s}} ^ {c} (G) \big). \qquad (8) +$$ + +The color histogram loss, measured using the $L_{1}$ norm, is given by + +$$ +\begin{array}{l} \mathcal {L} _ {\text {c o l o r}} (I, G) = \sum_ {c = 1} ^ {C} \| H ^ {c} (I) - H ^ {c} (G) \| _ {1} \tag {9} \\ = \sum_ {c = 1} ^ {C} \sum_ {k = 1} ^ {\text {b i n s}} \left| h _ {k} ^ {c} (I) - h _ {k} ^ {c} (G) \right|. \\ \end{array} +$$ + +The final loss function combines three loss terms as follows: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {d i f f}} + \mathcal {L} _ {\text {a l g n m e n t}} + \mathcal {L} _ {\text {c o l o r}}. \tag {10} +$$ + +# 3.5. Generalization and Stylized Synthesis + +Generalization to Floorplan. We aim to generate first-person view panoramas that faithfully represent a scene. While top-down views lack vertical details like walls and windows, this omission grants flexibility in generating panoramas. This flexibility is especially useful in interior design, where inputs are often simple floorplans. To improve generalization, particularly to schematic floorplans, we train on orthographic rather than perspective views, as they better match floorplans. Empirical results show that Top2Pano generalizes well to schematic and even hand-drawn floorplans while maintaining photorealism. Additionally, our model enables stylized synthesis guided by text or images, supporting diverse design needs. + +Text-Guided Stylization. The PanoGen module, built upon the text-driven Stable Diffusion model, inherently supports + +text-conditioned image generation. As illustrated in Figure 2, PanoGen synthesizes panoramas using three conditions: coarse depth, coarse colored image, and stylized textual guidance. When the input is a textureless floorplan, the stylized textual condition effectively guides the style of the synthesized result, as demonstrated in Figure 1. However, when a colored top-down view is provided, the influence of text-guided stylization becomes less pronounced. This occurs because the rendered coarse colored panorama constrains the final output to closely align with the input view, thereby diminishing the impact of the textual stylization. + +To enhance text-guided stylization, the weight of the coarse colored panorama in the PanoGen conditions can be reduced. However, this introduces a tradeoff: prioritizing stylization may come at the expense of fidelity to the top-down view. Regardless of this tradeoff, the coarse depth condition consistently ensures that the synthesized panorama adheres to the underlying scene geometry. Notably, this tradeoff does not apply to the text-to-panorama generation task, where textual guidance plays a more dominant role. + +Image-Guided Stylization. Given several scene images (not necessarily panoramas), we can fine-tune the PanoGen module using low-rank adaptation (LoRA) [14] to generate panoramas that align with the visual styles present in the provided images. To guide this process, we introduce structured textural prompts augmented with style tags (e.g., [Japanese]) at the beginning of the input prompts. These tags act as conditional modifiers, steering the model toward synthesizing images that follow specific aesthetic themes, such as regional design styles. The framework is applied solely to the PanoGen module, ensuring both computational and parameter efficiency. Notably, our method requires fewer than five in-the-wild images per target style, substantially reducing data demands. This approach achieves its efficiency by decomposing weight updates into low-rank matrices. For a pretrained weight matrix $\mathbf{W}_0 \in \mathbb{R}^{d \times d}$ , the update $\Delta \mathbf{W}$ is constrained as follows: + +$$ +\Delta \mathbf {W} = \mathbf {B} \mathbf {A}, \quad \text {w h e r e} \mathbf {A} \in \mathbb {R} ^ {d \times r}, \mathbf {B} \in \mathbb {R} ^ {r \times d} \tag {11} +$$ + +Here, $r \ll d$ represents the intrinsic rank (we use $r = 8$ ). During fine-tuning, only the matrices $\mathbf{A}$ and $\mathbf{B}$ are updated, while the original weights $\mathbf{W}_0$ remain fixed. The forward pass is modified as follows: + +$$ +\mathbf {h} _ {\text {o u t}} = \mathbf {W} _ {0} \mathbf {h} _ {\text {i n}} + \alpha \cdot \mathbf {B A h} _ {\text {i n}} \tag {12} +$$ + +where $\alpha$ is a scaling coefficient. This lightweight adaptation (0.8% new parameters) mitigates catastrophic forgetting and maintains the model's baseline generation quality for generic prompts while enabling precise style control through our [style] textual conditioning. + +
TrainingTesting
ScenesFloorsPanoramasScenesFloorsPanoramas
Matterport3D61127617714291405
Gibson152203537939761672
+ +Table 1. The numbers of scenes, floors, and panorama images in the training and testing sets of the two datasets + +
MethodPSNR ↑SSIM ↑FID ↓LPIPS ↓
Materport3DSat2Density[28]+LDM[29]11.270.4221100.060.6237
Sat2Density[28]+ControlNet[42]11.420.422285.780.6163
PanFusion[41]11.450.437285.740.6153
Top2Pano (Ours)11.720.440930.840.6029
GibsonSat2Density[28]+LDM[29]10.540.448084.330.6462
Sat2Density[28]+ControlNet[42]10.970.458285.210.6645
PanFusion[41]11.360.474479.530.6634
Top2Pano (Ours)11.580.485128.680.6282
+ +# 4. Experiments + +# 4.1. Data Preparation + +For evaluation, we use the Matterport3D [3] and Gibson [36] datasets. Since no existing dataset provides both top-down views and high-quality panoramic images, we generate top-down views from 3D models in these datasets using Blender. Specifically, we import textured 3D meshes into Blender and render top-down views with an orthographic camera. The top-down view we render closely resembles a floorplan, unlike the perspective-rendered views used in embodied dialog localization [12]. This similarity enhances our model's ability to generalize to floorplan inputs. To determine the number of floors in each scene, we apply DBSCAN [7] clustering to the camera positions within the datasets. We exclude certain scenes, such as airports and large supermarkets, as well as panoramic images depicting outdoor environments to ensure alignment with our task. After processing, the final dataset sizes are summarized in Table 1. + +# 4.2. Evaluation Metrics + +To assess the quality of the generated panoramas, we employ both pixel-based and perceptual evaluation metrics. For pixel-level assessment, we utilize peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) to quantify image fidelity. Additionally, we incorporate perceptual metrics such as Fréchet Inception Distance (FID) [13] and Learned Perceptual Image Patch Similarity (LPIPS) [43] to capture higher-level visual realism. + +# 4.3. Implement Details + +Our code runs on an NVIDIA RTX A6000 GPU with 48GB of memory. The model has 3.3 billion parameters and is trained with a batch size of 21 for 100 epochs. On average, each experiment takes approximately two days to complete on both the Matterport3D and Gibson datasets. We optimize + +Table 2. Quantitative comparison with existing methods on the Matterport3D [3] and Gibson [36] datasets. + +
ModulesMatterport3DGibson
seg floor wall depth colorPSNR↑SSIM↑FID↓LPIPS↓PSNR↑SSIM↑FID↓LPIPS↓
✓✓11.080.4122104.030.646311.070.4551117.080.6574
✓✓✓11.260.4206102.840.647111.100.4564115.880.6568
✓✓✓10.780.409155.730.616111.190.466832.280.6310
✓✓✓11.220.420754.440.619311.280.461383.310.6429
✓✓✓11.570.437843.990.606711.160.473381.870.6549
✓✓✓10.980.419653.340.611511.330.460133.890.6307
✓✓11.210.419935.080.613111.380.464136.290.6384
✓✓✓11.260.429934.610.606011.490.467334.360.6336
✓✓✓✓11.590.438167.020.619211.380.476165.660.6557
✓✓✓✓11.720.440930.840.602911.580.485128.680.6282
+ +Table 3. Ablation study on five designs in our Top2Pano model (top-down view segmentation, floor reinforcement, wall reinforcement, coarse depth panorama, and coarse colored panorama). + +our model using the Adam optimizer [18] with default parameters $(\beta_{1} = 0.9, \beta_{2} = 0.999, \epsilon = 10^{-8})$ and a learning rate of $10^{-5}$ . + +# 4.4. Comparison with Previous Methods + +As shown in Table 2, we compare our method against three baseline approaches across four evaluation metrics. Sat2Density [28] is a satellite-to-ground panorama synthesis method, which we adapt for indoor panorama generation using latent diffusion model (LDM) [29] and ControlNet [42]. PanFusion [41] is a text-to-panorama generation framework that also incorporates layout-conditioned generation via ControlNet. Table 2 shows that our method outperforms all baselines across all four metrics on both datasets, demonstrating its effectiveness. Furthermore, qualitative comparisons in Figures 3 and 4 highlight that our approach generates more realistic and structurally accurate house reconstructions, including furniture placement. + +# 4.5. Ablation Study + +We conducted comprehensive ablation studies to analyze and validate the contribution of each component in our model. Specifically, we performed experiments comparing several model variants against the original. These experiments involved removing key elements such as the structural reinforcement of the floor and walls, the segmentation input to the OccRecon module, and the coarse depth and colored panoramicas as conditional inputs to the PanoGen module. + +As shown in Table 3, our original model achieves the highest overall scores across all four metrics, with any modification leading to some degree of performance degradation. Removing coarse colored panoramas or the embedded floor significantly disrupts furniture placement and color accuracy. While the room structure remains mostly intact, furniture positions become unreliable, and color representations appear distorted. Conversely, excluding coarse depth panoramas or embedded walls maintains color and furniture accuracy but compromises the spatial understanding and overall quality of room structure reconstruction. These effects are further illustrated in the qualitative results in the supp. materials. + +![](images/d87bfacbbc6d25388816b311b302213cb3d0bb46625e8a1b8a330748ebf26ffd.jpg) + +![](images/3ca2b08a577bc0255d88476a19b5b742486e8bc1aabfbb723f06961184c25bd7.jpg) + +![](images/e220fba2817c759044c7626b257b0c9b4adfc237980f430092901eb295208e35.jpg) + +![](images/dab818e600dee898acfc2dc454e591b49ce9d5b58d62c1aff061ef0fa156a63d.jpg) + +![](images/d6454c2d8d9f2c5252d908f6876fd693422c00b507e8997bb017a8d2b70043d3.jpg) + +![](images/f5f7da3d5ae4cdde7b4064799ae46cde9ec32987c8729f96f8f682d3354539fc.jpg) + +![](images/75ad0ed1229b98c0dbca869cc726476e855f3964e596058b116451dfbddc4b47.jpg) + +![](images/2348aef428cb6c447d27ce7ae2b820d3afed552c37d78c221a5543c0157a9d1f.jpg) + +![](images/41173e3b0c831caa22b0d9be55a8dd33d3b34beef106cb034f5b38dcd1cd8388.jpg) + +![](images/fb192194dd9117f977de4e20e85edd5efc1511e495e37f2bfaa00d955589b7db.jpg) + +![](images/fe2fbd79e74a5ad1b0a976300bc0e99b339fcd993ff76e81b541f306f1d64780.jpg) + +![](images/3426376a90bee262dc255136afccf5f52ffc8fe24712aefff2cfc93a82dbfbea.jpg) + +![](images/10b3cc6e7819e1040a40a49e6805c63f35f2e99e97adfbc7d12e7b9532764ffb.jpg) + +![](images/67e10f29dc3dac2670427d6102daaab6639597534db638ac9ae29272a2172d32.jpg) + +![](images/11ab96e5e6346a65369b30f3e2581fca1700ce78d394df5d3e1b3654120bbdc6.jpg) + +![](images/d261615fde255af288974800b7dd9e7c9836e1b410e134e97253837a45fb78c3.jpg) + +![](images/e0c73e5110e422f43205eee42ed5486b5d3b29226158e5fc75f728f4a3909f48.jpg) + +![](images/ee5b0a2f7f93b90f13aa4246689e5f90868d3abe4ad49b80c40986e1e9ac86de.jpg) + +![](images/f4a9af37419317d636564f7f7cff1e21c0c46d045673fe8b898d89b2408d2ba0.jpg) + +![](images/1b40e50d7b666915888905233a3b749b7a8e10ce11572bf9a2be83ad0dc765b5.jpg) + +![](images/6edcdca92b1bdc686e11fec9330d3bdb8bfd32397fd5076cffbfa7777824660a.jpg) +(a) Top-down + +![](images/f8e49994ea7e6a2599dd1cb16ec253e349de08635308ebce6ccd6983484f79d5.jpg) +(b) Sat2Density [28] + +![](images/ea2eca47b313d457ee0e09d8e10eb968faa0f6789babfd373153daf2b6e0e402.jpg) +(c) PanFusion [41] + +![](images/d629db6494fc760ba7f900d97a389c87fe76ff2726a5e4151385b61d555da4c5.jpg) +(d) Ours + +![](images/3e35433d217b9c84193629a8f88df49aeeb454c2a7f1a711a1344bbf68d9e5a9.jpg) +(e) Ground Truth + +![](images/c14d0188b5709c4fcbef83d223b3baff16efbd9fcead728a823820e16b301518.jpg) + +![](images/bc66637b156586e359c5d763c911da2cd6e54ce72334141b7a32be5fce994a7b.jpg) +Figure 3. Qualitative comparisons on the Matterport3D dataset. + +![](images/223edc01e4497d93b35884ac4c903647637fe4fb48b10291873c1bd5ad10a030.jpg) + +![](images/f05c4ef89b20a0e2e1237535624cd2f5b6cbb051067fa6eab9fc5f027f22caee.jpg) + +![](images/284404ffe703471b9303da5234fb7ef40c9d5e000682f300df7d45a0ec3d2f9f.jpg) + +![](images/0db089f8e1927247b9f64db373875313f8eb6fc7bad8ca1644dd5cd83cde8c6a.jpg) + +![](images/fb8385ff7fb9c9656d5b7030844879f209718570148899aba6e1e8286ac99e1e.jpg) + +![](images/af0af0dff358f985d9eabdf9f30854ae5126e1ef205533739985f1e90181fcc0.jpg) + +![](images/b72e69f0ce8dc37538fc9b372ec4f33b91a85d401b8efa37d1e69aafb350f367.jpg) + +![](images/22b649c8498fa3f2069c391825e9625df562d97c8777744372d84556772beb96.jpg) + +![](images/3105be3aacb1178bd379a3a4559793fb7a1e91a71f067604035014c37e3e5c92.jpg) + +![](images/25c43bb4407057ca1401bad007444e19a6bab19777f98842606b27b490c9b705.jpg) + +![](images/3cccea65905067192fb436a816d70d258ce8907d519b0fc53233e6cbe43da0e5.jpg) + +![](images/cbe4164e2c0a9d6548b9b168044fea952d010f4764a162c0876e163a32c49a1b.jpg) + +![](images/aff2a99c7786f45ad56c41a9428a718b2db168ccad78cae9595139915c3cb531.jpg) + +![](images/a6516874589b8658a39e1599fa4a02d532010c2f7145f7bf1aee50ed6297019a.jpg) + +![](images/9eb348b30e81c6d5ce07f54c50d188fa9f80800e7dfdb75956eceb1602c9f2b7.jpg) + +![](images/9651671a11b4ba16871f82afc7c6627c3e00276eac76b7ee7f7e70b655ead32c.jpg) + +![](images/2ce4e1b5238ee97ddc15fe9d29318d483bc0ad3f77fa5989fa6e6fe5e1d5c6c3.jpg) + +![](images/f48ef503b889b0948444e0128a995a2d1ddc10ed6b726a4b911bdc15927c3e43.jpg) + +![](images/1531d71541dcd9280f6684c6862637804b1ead2ddf6762d121623b8c07cfac53.jpg) +(a) Top-down + +![](images/746aacbb27b068d1b2a05b0f8494a04c5f0afb44fc3a1462cac537e24b9d597e.jpg) +(b) Sat2Density [28] + +![](images/6b8b42fa12bebc8a610f4b66ac65710f1fc61c3d641929e8e3e62452604cfcc1.jpg) +(c) PanFusion [41] + +![](images/fe67f44e29f29015cdf3ba2dcc34507800e3ea435d7b3c635134d10929fcd704.jpg) +(d) Ours + +![](images/560c00334ce9b6bfc96c237b6c9333336e4f7cc21d64e2c2230b4a3bc8c80e8f.jpg) +(e) Ground Truth +Figure 4. Qualitative comparisons on the Gibson dataset. + +
PSNR ↑SSIM ↑FID ↓LPIPS ↓
G → M11.080.439740.740.6285
M → G11.030.436645.870.6353
+ +Table 4. Cross-dataset evaluation. "G" represents Gibson; "M" represents Matterport3D. + +![](images/eb27ff5d3b186bf25393e6e2934fbc1426ed1b363a55116bcf741168233b5655.jpg) +Figure 5. Our Top2Pano model generalizes to both textured floorplans (first row) and hand-drawn sketched floorplans (second row), incorporating stylized [Japanese] control. + +# 4.6. Generalization, Stylization, Manipulation + +We employed cross-dataset evaluation to assess our model's generalization capability. As shown in Table 4, although there is a slight decline in metric scores, our model maintains strong performance. Additionally, the trained parameters tend to generate photorealistic walls that reflect characteristics of the training dataset. + +Our model also generalizes well to floorplans. When given a specific floorplan and camera positions, the model assists users in generating indoor panoramic images. By using different text prompts, users can create various interior design styles, explore the house from a first-person perspective, and modify its style according to their preferences. We tested our model with three types of floorplans. The first type is a colored floorplan (Figure 5, first row), which provides detailed information about room colors and furniture hues, making it highly informative. The second type is a plain floorplan (Figure 1, bottom), which lacks color information and shows only the room structure and furniture layout. The third type is a hand-drawn floorplan sketch (Figure 5, first row), which offers a rough visual representation of the room. Our model successfully generates accurate panoramic images from these floorplans and adapts the visual style based on the provided textual descriptions. Moreover, our model enables panorama manipulation by editing objects in the floorplan, such as adding new items, as illustrated in Figure 6. + +# 4.7. Limitations + +Failure cases. Figure 7 shows representative failure cases. We annotate different types of failures with numbered labels: + +- Ceiling: ① missing fan, ② vaulted ceiling, ③ false light; +- Wall: 4 height error, 5 false or missing decorations; +- Window: 6 false window; + +![](images/45012e95f40adb80c8263859436dab7eb4b3f3599ce770cc79ad7ff8b743e267.jpg) +Figure 6. Top2Pano enables panorama manipulation via compositional floorplan editing. In the second row, adding a rectangular object to the floorplan (compared to the first row) leads the model to generate a washstand with a mirror in the panorama. + +![](images/1fc250a1e614b2c9d2c81f0f66d5713977d5d4e5d91f9c52eec3769209a8769e.jpg) +Figure 7. Failure cases (zoom in to view error types). + +- Furniture: height error; +- Thin object: 8 missing flat-screen TV; +Stairs: false direction. + +The failure cases stem from the ambiguity of the 2D input, leading to hallucinated objects that are not observable from the top-down view. The limitations in handling vertical structural details are largely due to the inherent ambiguity of the task. + +Limited Vertical FoV. The generated panoramas exhibit a limited vertical field of view (FoV), reflecting the constraints of the training data. We expect improved performance with future datasets that include full vertical FoV panoramas. + +# 5. Conclusions + +We present Top2Pano, a novel method for generating high-quality $360^{\circ}$ indoor panoramas from 2D top-down views. The model first estimates volumetric occupancy to infer 3D structure, then applies volumetric rendering to produce coarse color and depth panoramas. These guide a diffusion-based refinement stage via ControlNet. To our knowledge, this is the first approach to generate panoramas from top-down views. Experiments on two datasets show that Top2Pano outperforms baselines in reconstructing room layouts and realistic furniture. + +# References + +[1] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. Diverse plausible 360-degree image outpainting for efficient 3dcg background creation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 11431-11440, 2022. 3 +[2] Sherwin Bahmani, Jeong Joon Park, Despoina Paschalidou, Xingguang Yan, Gordon Wetzstein, Leonidas J. Guibas, and Andrea Tagliasacchi. CC3D: layout-conditioned generation of compositional 3d scenes. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 7137-7147, 2023. 1, 3 +[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1, 6 +[4] Minglin Chen, Longguang Wang, Sheng Ao, Ye Zhang, Kai Xu, and Yulan Guo. Layout2scene: 3d semantic layout guided scene generation via geometry and appearance diffusion priors. arXiv preprint arXiv:2501.02519, 2025. 1, 3 +[5] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 3 +[6] Mohammad Reza Karimi Dastjerdi, Yannick Hold-Geoffroy, Jonathan Eisenmann, Siavash Khodadadeh, and Jean-François Lalonde. Guided co-modulated gan for $360^{\circ}$ field of view extrapolation. In 2022 International Conference on 3D Vision (3DV), pages 475–485, 2022. 3 +[7] Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD), pages 226-231, 1996. 6 +[8] Chuan Fang, Xiaotao Hu, Kunming Luo, and Ping Tan. Ctrlroom: Controllable text-to-3d room meshes generation with layout constraints. arXiv preprint arXiv:2310.03602, 2023. 1, 3 +[9] Chengzeng Feng, Jiacheng Wei, Cheng Chen, Yang Li, Pan Ji, Fayao Liu, Hongdong Li, and Guosheng Lin. Prim2room: Layout-controllable room mesh generation from primitives. arXiv preprint arXiv:2409.05380, 2024. 1, 3 +[10] Mengyang Feng, Jinlin Liu, Miaomiao Cui, and Xuansong Xie. Diffusion360: Seamless 360 degree panoramic image generation based on diffusion models. arXiv preprint arXiv:2311.13141, 2023. 1, 3 +[11] Julia Guerrero-Viu, Clara Fernandez-Labrador, Cédric Demonceaux, and José Jesús Guerrero. What's in my room? object recognition on indoor panoramic images. In IEEE International Conference on Robotics and Automation, ICRA, pages 567-573, 2020. 1 +[12] Meera Hahn, Jacob Krantz, Dhruv Batra, Devi Parikh, James M Rehg, Stefan Lee, and Peter Anderson. Where are you? localization from embodied dialog. arXiv preprint arXiv:2011.08277, 2020. 6 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two + +time-scale update rule converge to a local nash equilibrium. In Annual Conference on Neural Information Processing Systems, pages 6626-6637, 2017. 6 +[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, ICLR, 2022. 5 +[15] Zilong Huang, Jun He, Junyan Ye, Lihan Jiang, Weijia Li, Yiping Chen, and Ting Han. Scene4u: Hierarchical layered 3d scene reconstruction from single panoramic image for your immerse exploration. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 26723-26733, 2025. 1 +[16] Zhe Huang, Yizhe Zhao, Hao Xiao, Chenyan Wu, and Lingting Ge. Duospacenet: Leveraging both bird's-eye-view and perspective view representations for 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops, pages 2560-2570, 2025. 1 +[17] Nikolai Kalischek, Michael Oechsle, Fabian Manhardt, Philipp Henzler, Konrad Schindler, and Federico Tombari. Cubediff: Repurposing diffusion-based image models for panorama generation. In International Conference on Learning Representations, ICLR, 2025. 3 +[18] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, ICLR, 2015. 6 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 3992-4003, 2023. 2, 3 +[20] Jiachen Liu, Yuan Xue, Haomiao Ni, Rui Yu, Zihan Zhou, and Sharon X Huang. Computer-aided layout generation for building design: A review. arXiv preprint arXiv:2504.09694, 2025.1 +[21] Jiachen Liu, Rui Yu, Sili Chen, Sharon X. Huang, and Hengkai Guo. Towards in-the-wild 3d plane reconstruction from a single image. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 27027-27037, 2025. 1 +[22] Xiaohu Lu, Zuoyue Li, Zhaopeng Cui, Martin R. Oswald, Marc Pollefeys, and Rongjun Qin. Geometry-aware satellite-to-ground image synthesis for urban areas. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 856-864, 2020. 3 +[23] Wentao Lyu, Peng Ding, Yingliang Zhang, Anpei Chen, Minye Wu, Shu Yin, and Jingyi Yu. Refocusable gigapixel panoramas for immersive VR experiences. IEEE Trans. Vis. Comput. Graph., 27(3):2028-2040, 2021. 1 +[24] Nelson L. Max. Optical models for direct volume rendering. IEEE Trans. Vis. Comput. Graph., 1(2):99-108, 1995. 4 +[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn + +thesis. In European Conference on Computer Vision, ECCV, pages 405-421, 2020. 4 +[26] Despoina Paschalidou, Amlan Kar, Maria Shugrina, Karsten Kreis, Andreas Geiger, and Sanja Fidler. Atiss: Autoregressive transformers for indoor scene synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[27] Giovanni Pintore, Fabio Bettio, Marco Agus, and Enrico Gobbetti. Deep scene synthesis of atlanta-world interiors from a single omnidirectional image. IEEE Trans. Vis. Comput. Graph., 29(11):4708-4718, 2023. 1 +[28] Ming Qian, Jincheng Xiong, Gui-Song Xia, and Nan Xue. Sat2density: Faithful density learning from satellite-ground image pairs. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 3660-3669, 2023. 3, 4, 6, 7 +[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 10674-10685, 2022. 6 +[30] Jonas Schult, Sam Tsai, Lukas Hollein, Bichen Wu, Jialiang Wang, Chih-Yao Ma, Kunpeng Li, Xiaofang Wang, Felix Wimbauer, Zijian He, Peizhao Zhang, Bastian Leibe, Peter Vajda, and Ji Hou. Controlroom3d: Room generation using semantic proxy rooms. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 3 +[31] Yujiao Shi, Dylan Campbell, Xin Yu, and Hongdong Li. Geometry-guided street-view panorama synthesis from satellite imagery. IEEE Trans. Pattern Anal. Mach. Intell., 44(12): 10009-10022, 2022. 3 +[32] Bin Tan, Rui Yu, Yujun Shen, and Nan Xue. Planarsplatting: Accurate planar surface reconstruction in 3 minutes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 1190-1199, 2025. 1 +[33] Madhawa Vidanapathirana, Qirui Wu, Yasutaka Furukawa, Angel X. Chang, and Manolis Savva. Plan2scene: Converting floorplans to 3d scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 10733-10742, 2021. 3 +[34] Jionghao Wang, Ziyu Chen, Jun Ling, Rong Xie, and Li Song. 360-degree panorama generation from few unregistered nfov images. In Proceedings of the 31st ACM International Conference on Multimedia, MM, pages 6811-6821, 2023. 3 +[35] Songsong Wu, Hao Tang, Xiao-Yuan Jing, Haifeng Zhao, Jianjun Qian, Nicu Sebe, and Yan Yan. Cross-view panorama image synthesis. IEEE Trans. Multim., 25:3546-3559, 2023. 3 +[36] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: real-world perception for embodied agents. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, 2018. 6 +[37] Ningli Xu and Rongjun Qin. Geospecific view generation geometry-context aware high-resolution ground view inference from satellite views. In European Conference on Computer Vision, ECCV, pages 349-366, 2024. 3 +[38] Xiuyu Yang, Yunze Man, Jun-Kun Chen, and Yu-Xiong Wang. Scenecraft: Layout-guided 3d scene generation. In Advances in Neural Information Processing Systems, 2024. 1, 3 + +[39] Weicai Ye, Chenhao Ji, Zheng Chen, Junyao Gao, Xiaoshui Huang, Song-Hai Zhang, Wanli Ouyang, Tong He, Cairong Zhao, and Guofeng Zhang. Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion. In Annual Conference on Neural Information Processing Systems, NeurIPS, 2024. 1, 3 +[40] Rui Yu, Jiachen Liu, Zihan Zhou, and Sharon X. Huang. Nerf-enhanced outpainting for faithful field-of-view extrapolation. In IEEE International Conference on Robotics and Automation, ICRA, pages 16826-16833, 2024. 1 +[41] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to $360^{\circ}$ panorama image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 3, 6, 7 +[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 3813-3824, 2023. 2, 3, 4, 6 +[43] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pages 586-595, 2018. 6 +[44] Dian Zheng, Cheng Zhang, Xiao-Ming Wu, Cao Li, Chengfei Lv, Jian-Fang Hu, and Wei-Shi Zheng. Panorama generation from nfov image done right. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 21610-21619, 2025. 1 \ No newline at end of file diff --git a/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/images.zip b/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..83d63c8fe15d123bcdaea32c68b2f53562eb094d --- /dev/null +++ b/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bbd22e3dba5ecb8fde892317eff4d9af061fb93d025ec71fc7f315625806f6ba +size 771802 diff --git a/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/layout.json b/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bfdd366f3ef95cddee1bd4298de9a5ffd2a6e093 --- /dev/null +++ b/2025/Top2Pano_ Learning to Generate Indoor Panoramas from Top-Down View/layout.json @@ -0,0 +1,10528 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 103, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 536, + 122 + ], + "type": "text", + "content": "Top2Pano: Learning to Generate Indoor Panoramas from Top-Down View" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 186, + 144, + 424, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 144, + 424, + 171 + ], + "spans": [ + { + "bbox": [ + 186, + 144, + 424, + 171 + ], + "type": "text", + "content": "Zitong Zhang Suranjan Gautam Rui Yu University of Louisville" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 174, + 451, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 174, + 451, + 186 + ], + "spans": [ + { + "bbox": [ + 159, + 174, + 451, + 186 + ], + "type": "text", + "content": "{zitong.zhang, suranjan.gautam, rui.yu}@louisville.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 230, + 188, + 377, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 230, + 188, + 377, + 198 + ], + "spans": [ + { + "bbox": [ + 230, + 188, + 377, + 198 + ], + "type": "text", + "content": "https://top2pano.github.io/" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 207, + 102, + 273 + ], + "blocks": [ + { + "bbox": [ + 56, + 207, + 102, + 273 + ], + "lines": [ + { + "bbox": [ + 56, + 207, + 102, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 102, + 273 + ], + "type": "image", + "image_path": "87224dcd1f0f73f69582edfd7f31a0c55d8cd50f55ccf841fd2f52435aa16045.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 274, + 110, + 282 + ], + "lines": [ + { + "bbox": [ + 56, + 274, + 110, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 110, + 282 + ], + "type": "text", + "content": "Top-down view" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 100, + 229, + 165, + 258 + ], + "blocks": [ + { + "bbox": [ + 111, + 217, + 156, + 228 + ], + "lines": [ + { + "bbox": [ + 111, + 217, + 156, + 228 + ], + "spans": [ + { + "bbox": [ + 111, + 217, + 156, + 228 + ], + "type": "text", + "content": "Top2Pano" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 229, + 165, + 258 + ], + "lines": [ + { + "bbox": [ + 100, + 229, + 165, + 258 + ], + "spans": [ + { + "bbox": [ + 100, + 229, + 165, + 258 + ], + "type": "image", + "image_path": "735e1e120e15d3c20da8305349706970f7009c55167331b92e15b05c5aac50f6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 166, + 207, + 294, + 272 + ], + "blocks": [ + { + "bbox": [ + 166, + 207, + 294, + 272 + ], + "lines": [ + { + "bbox": [ + 166, + 207, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 166, + 207, + 294, + 272 + ], + "type": "image", + "image_path": "ec1b6120b05d374645c2eecb4d9b4651a0606aac0e8d229920a75de6a7588942.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 272, + 244, + 281 + ], + "lines": [ + { + "bbox": [ + 214, + 272, + 244, + 281 + ], + "spans": [ + { + "bbox": [ + 214, + 272, + 244, + 281 + ], + "type": "text", + "content": "Output 1" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 295, + 208, + 423, + 273 + ], + "blocks": [ + { + "bbox": [ + 295, + 208, + 423, + 273 + ], + "lines": [ + { + "bbox": [ + 295, + 208, + 423, + 273 + ], + "spans": [ + { + "bbox": [ + 295, + 208, + 423, + 273 + ], + "type": "image", + "image_path": "725d35ef4e56b533ccc99ba1b5887ba31087e881a160f35a9ac35bd23c67f706.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 273, + 374, + 281 + ], + "lines": [ + { + "bbox": [ + 343, + 273, + 374, + 281 + ], + "spans": [ + { + "bbox": [ + 343, + 273, + 374, + 281 + ], + "type": "text", + "content": "Output 2" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 425, + 208, + 553, + 273 + ], + "blocks": [ + { + "bbox": [ + 425, + 208, + 553, + 273 + ], + "lines": [ + { + "bbox": [ + 425, + 208, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 425, + 208, + 553, + 273 + ], + "type": "image", + "image_path": "ff0842f581694a85ce859498055eee298b328dabe8060f2a95f1f2a14f845357.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 273, + 512, + 281 + ], + "lines": [ + { + "bbox": [ + 465, + 273, + 512, + 281 + ], + "spans": [ + { + "bbox": [ + 465, + 273, + 512, + 281 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 56, + 293, + 165, + 350 + ], + "blocks": [ + { + "bbox": [ + 56, + 293, + 165, + 350 + ], + "lines": [ + { + "bbox": [ + 56, + 293, + 165, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 165, + 350 + ], + "type": "image", + "image_path": "aeb079062defbd1466a8935594ae1bb340a8600929d40752fdf9aea4c059d5ba.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 350, + 118, + 359 + ], + "lines": [ + { + "bbox": [ + 85, + 350, + 118, + 359 + ], + "spans": [ + { + "bbox": [ + 85, + 350, + 118, + 359 + ], + "type": "text", + "content": "Floorplan" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 364, + 555, + 419 + ], + "lines": [ + { + "bbox": [ + 55, + 364, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 555, + 419 + ], + "type": "text", + "content": "Figure 1. Top: We present Top2Pano, a method for synthesizing high-quality indoor panoramas from a top-down view. Given a camera position, Top2Pano generates panoramas that are both visually compelling and geometrically accurate. Bottom: Our model demonstrates strong generalization capabilities. When provided with schematic floor plans as input, Top2Pano produces photorealistic and structurally coherent panoramas. Additionally, our approach can be easily adapted for stylized synthesis, enabling diverse design variations. Note: The original dataset (Matterport3D [3]) contains blurry regions near the upper and lower edges of the panoramic image." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 165, + 285, + 293, + 350 + ], + "blocks": [ + { + "bbox": [ + 165, + 285, + 293, + 350 + ], + "lines": [ + { + "bbox": [ + 165, + 285, + 293, + 350 + ], + "spans": [ + { + "bbox": [ + 165, + 285, + 293, + 350 + ], + "type": "image", + "image_path": "92074427435354c982e115ec357e4ef9b8aeeb3fbd374e657c29893892d2ee9c.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 350, + 243, + 358 + ], + "lines": [ + { + "bbox": [ + 217, + 350, + 243, + 358 + ], + "spans": [ + { + "bbox": [ + 217, + 350, + 243, + 358 + ], + "type": "text", + "content": "Default" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 294, + 285, + 423, + 350 + ], + "blocks": [ + { + "bbox": [ + 294, + 285, + 423, + 350 + ], + "lines": [ + { + "bbox": [ + 294, + 285, + 423, + 350 + ], + "spans": [ + { + "bbox": [ + 294, + 285, + 423, + 350 + ], + "type": "image", + "image_path": "165a518e082ea816a5ffe3c95a493f513fe530e02f5812876f6926f2b2e32f23.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 350, + 385, + 358 + ], + "lines": [ + { + "bbox": [ + 334, + 350, + 385, + 358 + ], + "spans": [ + { + "bbox": [ + 334, + 350, + 385, + 358 + ], + "type": "text", + "content": "\"Chinese Style\"" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 425, + 285, + 553, + 350 + ], + "blocks": [ + { + "bbox": [ + 425, + 285, + 553, + 350 + ], + "lines": [ + { + "bbox": [ + 425, + 285, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 425, + 285, + 553, + 350 + ], + "type": "image", + "image_path": "205b4f0aa706d48aba89f5609ed437fe16f30b53af5bd772c92b61d0a7e623d1.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 350, + 511, + 358 + ], + "lines": [ + { + "bbox": [ + 468, + 350, + 511, + 358 + ], + "spans": [ + { + "bbox": [ + 468, + 350, + 511, + 358 + ], + "type": "text", + "content": "\"Stone Wall\"" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 151, + 429, + 200, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 429, + 200, + 441 + ], + "spans": [ + { + "bbox": [ + 151, + 429, + 200, + 441 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 54, + 453, + 297, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 453, + 297, + 658 + ], + "spans": [ + { + "bbox": [ + 54, + 453, + 297, + 658 + ], + "type": "text", + "content": "Generating immersive " + }, + { + "bbox": [ + 54, + 453, + 297, + 658 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 54, + 453, + 297, + 658 + ], + "type": "text", + "content": " indoor panoramas from 2D top-down views has applications in virtual reality, interior design, real estate, and robotics. This task is challenging due to the lack of explicit 3D structure and the need for geometric consistency and photorealism. We propose Top2Pano, an end-to-end model for synthesizing realistic indoor panoramas from top-down views. Our method estimates volumetric occupancy to infer 3D structures, then uses volumetric rendering to generate coarse color and depth panoramas. These guide a diffusion-based refinement stage using ControlNet, enhancing realism and structural fidelity. Evaluations on two datasets show Top2Pano outperforms baselines, effectively reconstructing geometry, occlusions, and spatial arrangements. It also generalizes well, producing high-quality panoramas from schematic floorplans. Our results highlight Top2Pano's potential in bridging top-down views with immersive indoor synthesis." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 669, + 135, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 135, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 135, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 712 + ], + "type": "text", + "content": "Understanding and synthesizing immersive indoor scenes from minimal structural information is a fundamental chal" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 430, + 556, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 706 + ], + "type": "text", + "content": "lenge in computer vision and graphics [16, 20, 21, 32, 40, 44]. The ability to generate realistic indoor panorama images from a 2D top-down view holds immense potential for a wide range of applications, including virtual reality (VR) [23], interior design [27], real estate visualization [15], and robotics [11]. For instance, real estate platforms can leverage this technology to offer potential buyers photorealistic virtual walkthroughs generated directly from architectural floorplans, enhancing the property viewing experience. Similarly, VR applications can benefit from automatically synthesized environments that create more engaging and immersive user experiences. Additionally, robots operating in indoor environments can utilize synthesized panoramas to improve their spatial understanding and navigation capabilities, enabling more efficient and accurate movement in complex spaces. Despite its broad applicability, the task of generating high-quality indoor panoramas from top-down views remains surprisingly underexplored in the literature. Recent advancements in large multimodal models have enabled the synthesis of panoramas directly from text input [10, 39, 41]; however, these approaches often overlook critical geometric and textural constraints. Other studies have focused on generating 3D models from semantic layouts [2, 4, 8, 9, 30, 38];" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "28493" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 85, + 555, + 220 + ], + "blocks": [ + { + "bbox": [ + 67, + 73, + 214, + 83 + ], + "lines": [ + { + "bbox": [ + 67, + 73, + 214, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 214, + 83 + ], + "type": "text", + "content": "(a) Volumetric Occupancy Estimation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 225, + 73, + 352, + 83 + ], + "lines": [ + { + "bbox": [ + 225, + 73, + 352, + 83 + ], + "spans": [ + { + "bbox": [ + 225, + 73, + 352, + 83 + ], + "type": "text", + "content": "(b) Coarse Panorama Rendering" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 73, + 530, + 83 + ], + "lines": [ + { + "bbox": [ + 383, + 73, + 530, + 83 + ], + "spans": [ + { + "bbox": [ + 383, + 73, + 530, + 83 + ], + "type": "text", + "content": "(c) Photorealistic Panorama Synthesis" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 85, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 58, + 85, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 85, + 555, + 220 + ], + "type": "image", + "image_path": "800c677146bc589beee6fcb77f6e408d9d46964197f4e1dcf18574abd727b4a3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 229, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 555, + 285 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed Top2Pano pipeline. The pipeline begins by segmenting the top-down view using SAM [19]. Both the segmented top-down image and the original top-down view are then processed by the OccRecon module to estimate the scene's 3D volumetric occupancy. Next, given the camera position, the system employs volumetric rendering to generate coarse depth and color panoramas. These coarse images are subsequently refined by the PanoGen module to produce the final photorealistic panorama. The PanoGen module also supports stylized panorama generation based on textual or visual conditions." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 305, + 295, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 305, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 295, + 366 + ], + "type": "text", + "content": "however, these methods are often limited by the quality of the resulting 3D meshes, making them unsuitable for rendering high-quality panoramas. Furthermore, semantic information is often unavailable in top-down views or floorplans, posing an additional challenge for existing approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 366, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 366, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 295, + 510 + ], + "type": "text", + "content": "Addressing this gap requires overcoming several significant technical challenges. First, a 2D top-down view provides only limited visual cues about the actual appearance and layout of the scene, making it difficult to infer occluded structures and fine texture details. Second, generating geometrically consistent indoor scenes demands accurate reasoning about 3D spatial occupancy from a 2D input, which is inherently ambiguous. Third, achieving photorealism while maintaining structural coherence necessitates a synthesis approach that effectively balances fidelity and realism, ensuring that the generated scenes are both visually appealing and functionally accurate." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 297, + 714 + ], + "type": "text", + "content": "To tackle these challenges, we introduce Top2Pano, a novel framework for generating photorealistic indoor panorama images from 2D top-down views. Our approach consists of three main stages. First, we learn the volumetric occupancy of the indoor scene, enabling the model to infer plausible spatial structures and layout configurations. Next, we employ volumetric rendering to generate coarse depth and colored panorama images, providing an initial estimate of the scene's appearance and geometry. Finally, we refine the synthesized panoramas using a diffusion-based model [42] conditioned on the coarse representations, enhancing both realism and structural consistency. By incorporating learned occupancy priors and diffusion-based refinement, our model effectively bridges the gap between schematic top-down views and immersive indoor panoramas, producing results that are both visually compelling and geometrically accurate." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 305, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 555, + 425 + ], + "type": "text", + "content": "We evaluate Top2Pano on two indoor datasets and demonstrate its effectiveness compared to baseline methods. Our model not only generates higher-quality images with improved geometric consistency but also exhibits strong generalization capabilities. Even when provided with schematic floorplans as input, Top2Pano can produce photorealistic and structurally coherent panoramas. Moreover, we show that our method can be easily adapted for stylized synthesis, allowing for diverse design variations and enabling users to explore different interior aesthetics with ease." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 426, + 477, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 426, + 477, + 437 + ], + "spans": [ + { + "bbox": [ + 325, + 426, + 477, + 437 + ], + "type": "text", + "content": "Our key contributions are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 437, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 437, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 314, + 437, + 555, + 485 + ], + "type": "text", + "content": "- We introduce Top2Pano, a novel framework for generating indoor panoramas from 2D top-down views, integrating volumetric occupancy learning, coarse synthesis, and diffusion-based refinement to achieve high-quality results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 533 + ], + "type": "text", + "content": "- We conduct extensive experiments on two indoor datasets, demonstrating that our model surpasses baseline methods in both image quality and structural consistency, setting a new benchmark for this task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 533, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 605 + ], + "type": "text", + "content": "- We show that Top2Pano generalizes well to schematic floorplans, producing high-quality, geometry-consistent panoramas. Furthermore, our approach supports stylish synthesis, enabling the generation of panoramas with diverse interior design aesthetics, making it a versatile tool for various applications." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 616, + 400, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 616, + 400, + 628 + ], + "spans": [ + { + "bbox": [ + 314, + 616, + 400, + 628 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 636, + 441, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 636, + 441, + 647 + ], + "spans": [ + { + "bbox": [ + 314, + 636, + 441, + 647 + ], + "type": "text", + "content": "2.1. Panorama Generation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 712 + ], + "type": "text", + "content": "Traditionally, panoramas were generated using image stitching and feature matching methods. With the recent advancements in generative machine learning, text-driven panorama generation techniques have gained popularity. These methods have utilized GANs, VAEs or a combina-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "28494" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 277 + ], + "type": "text", + "content": "tion of GANs and VAEs [5] and more recently, diffusion models [10, 39, 41] to synthesize panoramic images from textual descriptions. Another popular field of research is panorama synthesis from narrow-FOV images using image out-painting. Some methods rely solely on narrow-FOV images [1], while other incorporate textual descriptions alongside the images [6, 17, 34]. Cross-view panorama generation is another well-explored area, particularly challenging due to large shifts in camera perspective. Most techniques in this domain have focused on generating ground-view panoramas from aerial images. Some approaches directly use top-down images as input [35], while others [22, 28, 31, 37] extract geometric and segmentation information from top-down images to enhance quality. To the best of our knowledge, there is no prior work that explores the generation of indoor panoramic images from floor plans or top-down views of indoor spaces." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 283, + 251, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 251, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 251, + 295 + ], + "type": "text", + "content": "2.2. Layout-Guided 3D Scene Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 515 + ], + "type": "text", + "content": "Recent approaches to 3D scene generation leverage layouts for semantic and physical plausibility. Plan2Scene [33] reconstructs 3D meshes from floor plans, while ATISS [26] employs Transformers conditioned on scene layout. CC3D [2] follows a 3D GAN-based approach using 2D semantic layouts. Diffusion-based methods such as SceneCraft [38], Layout2Scene [4], and Prim2Room [9] have further improved synthesis quality. ControlRoom3D [30] and Ctrl-Room [8] are closely related to our work, as both generate panoramas during reconstruction process. Unlike prior methods that rely on explicit semantic layouts detailing object classes, positions, orientations, and sizes, our approach only requires a top-down view – an easily obtainable and lightweight input format. Instead of generating full 3D scenes, we produce panoramas, offering a more efficient and realistic solution for AR/VR, and autonomous robotic navigation by enabling immersive experiences like virtual tours and supporting real-time robotic navigation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 525, + 111, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 111, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 111, + 538 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": "The pipeline of our Top2Pano method is illustrated in Figure 2. Given an input top-down view " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{top}} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": ", we first generate its segmentation map " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{seg}} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": " using a pretrained model. Both the top-down view and its segmentation, along with the specified camera position, are then fed into an encoder-decoder occupancy estimation module, OccRecon, which reconstructs a 3D volumetric occupancy map " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{occ}} \\in \\mathbb{R}^{H \\times W \\times N}" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": " represents the number of vertical voxels. From this occupancy map, we render a coarse depth panorama " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{depth}}" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": " via volumetric rendering and project colors from the top-down view to obtain a coarse color panorama " + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{color}}" + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": ". To ensure geometric consistency, we enforce structural constraints on walls and floor, preserving occlusion relationships and realistic spatial structure. Finally," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 556, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 156 + ], + "type": "text", + "content": "both coarse depth and color panoramas serve as conditions for a diffusion-based synthesis module, PanoGen, which generates photorealistic panoramic images " + }, + { + "bbox": [ + 313, + 72, + 556, + 156 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{pano}} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 72, + 556, + 156 + ], + "type": "text", + "content": ". These images faithfully capture the scene's spatial layout, furniture, and fine color details. Additionally, PanoGen module supports stylized synthesis with optional textual or imagery-based controls." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 164, + 499, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 164, + 499, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 499, + 177 + ], + "type": "text", + "content": "3.1. Volumetric Occupancy Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 182, + 555, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 242 + ], + "type": "text", + "content": "The 2D top-down views lack 3D structural information about objects and furniture. To render panoramas that accurately reflect the geometric spatial relationships of objects, we propose training an OccRecon module to estimate the scene's 3D occupancy or density." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 246, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 556, + 437 + ], + "type": "text", + "content": "Input Representations. Unlike the layout-guided 3D scene generation setting, our top-down input lacks semantic information. In our preliminary study, we found that current semantic segmentation models struggle to generalize to indoor top-down views, making it challenging to estimate the 3D structure without semantic guidance. To address this, we propose leveraging a pretrained segmentation model SAM [19] to extract the 2D structure of the scene. Both the top-down image and the segmentation view are then fed into the encoder of the OccRecon module. The segmentation provides valuable details, such as room boundaries, furniture positions, and shapes, which significantly enhance the OccRecon module's ability to learn the overall 3D structure of the rooms. Moreover, this semantic-free input design enables our model to generalize effectively to more abstract inputs, such as semantic floorplans." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 441, + 556, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 556, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 556, + 645 + ], + "type": "text", + "content": "OccRecon Module. We propose a diffusion-based encoder-decoder framework that efficiently extracts and reconstructs the 3D spatial structure of a scene. Instead of using computationally intensive 3D diffusion models, our approach leverages a 2D diffusion model [42], significantly reducing resource demands. The OccRecon module processes 2D images and segmentations to extract spatial information and reconstruct room structures, including wall and furniture heights. In the final step, a 3D convolutional layer integrates the learned spatial features to generate a comprehensive 3D occupancy map. By relying on 2D inputs for most of the process, our method drastically lowers computational costs while still capturing the full scene layout. This balance of efficiency and accuracy makes it a practical solution for 3D structural modeling without semantic inputs. The OccRecon module outputs a 3D volumetric occupancy map " + }, + { + "bbox": [ + 313, + 441, + 556, + 645 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{occ}}" + }, + { + "bbox": [ + 313, + 441, + 556, + 645 + ], + "type": "text", + "content": ", which explicitly represents the scene's 3D geometry." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 353, + 654, + 555, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 654, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 353, + 654, + 555, + 668 + ], + "type": "interline_equation", + "content": "V _ {\\text {o c c}} = \\text {O c c R e c o n} \\left(I _ {\\text {t o p}}, I _ {\\text {s e g}}\\right) _ {\\text {c o n d i t i o n}} \\tag {1}", + "image_path": "c2fe3aa890f0e0ebf726abba30ab91c35e3e5d1f287cbc3d4a32ce7adc7e8cec.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 557, + 713 + ], + "type": "text", + "content": "We normalize the occupancy values to the range [0, 1]. Given the scale information of the scene, we transform the real-world coordinates of a 3D point into the estimated volumetric" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "28495" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "space. The point's density is then obtained by querying the learned occupancy map " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{occ}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " using tri-linear interpolation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 100, + 298, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 100, + 298, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 100, + 298, + 340 + ], + "type": "text", + "content": "Structural Reinforcement. Top-down views offer a comprehensive overview of the entire floor, but when observed from a first-person perspective, the scene typically reveals only the details of the current space, obscuring rooms beyond the walls. After the model learns the overall geometric structure of the rooms through the OccRecon module, we apply structural reinforcement to refine the wall geometry. This provides precise depth information, enhancing the reconstruction of the room's geometric layout. We solidify the wall voxels by applying the maximum value (1 after normalization). Additionally, top-down priors can constrain the floor's geometry and help infer the texture of the furniture, which is essential for generating accurate indoor panoramas with correctly placed and colored furniture. Since the OccNet modules are optimized end-to-end with the subsequent modules, the height of the furniture can be inferred from the 3D occupancy map. By encoding structural constraints from the walls and floor, we achieve a more accurate representation of the scene's geometry, including the positions, colors, and other attributes of the furniture." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 349, + 216, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 216, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 216, + 361 + ], + "type": "text", + "content": "3.2. Coarse Panorama Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 367, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 296, + 486 + ], + "type": "text", + "content": "Given the 3D occupancy map of the scene, we render coarse depth and color images based on the specified camera position. To ensure accurate mapping, we first compute the ratio between pixel resolution and physical dimensions. With the room height known, we apply this ratio along with the pixel coordinates in the top-down image to determine the corresponding position in the occupancy map " + }, + { + "bbox": [ + 55, + 367, + 296, + 486 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{occ}}" + }, + { + "bbox": [ + 55, + 367, + 296, + 486 + ], + "type": "text", + "content": ". To generate the panoramic image from the 3D occupancy map, we employ equirectangular projection along with a spherical coordinate system." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 487, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 523 + ], + "type": "text", + "content": "The coarse depth panorama " + }, + { + "bbox": [ + 55, + 487, + 296, + 523 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{depth}}" + }, + { + "bbox": [ + 55, + 487, + 296, + 523 + ], + "type": "text", + "content": " is obtained through volumetric rendering [24] of the occupancy map. The depth of a projected ray at pixel " + }, + { + "bbox": [ + 55, + 487, + 296, + 523 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 55, + 487, + 296, + 523 + ], + "type": "text", + "content": " is computed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 88, + 533, + 296, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 533, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 88, + 533, + 296, + 569 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {d e p t h}} ^ {(u, v)} = \\sum_ {i = 1} ^ {S} T _ {i} \\alpha_ {i} d _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\qquad (2)", + "image_path": "d5aca13f391a93a82cb72370e9901b5e7b8648db96ba483a9ddce26eefc09f8c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 577, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 295, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 577, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 577, + 295, + 601 + ], + "type": "text", + "content": " represents the transparency level, and " + }, + { + "bbox": [ + 55, + 577, + 295, + 601 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 55, + 577, + 295, + 601 + ], + "type": "text", + "content": " denotes the distance from the sampled position to the camera." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 601, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 673 + ], + "type": "text", + "content": "The coarse color panorama " + }, + { + "bbox": [ + 55, + 601, + 296, + 673 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{color}}" + }, + { + "bbox": [ + 55, + 601, + 296, + 673 + ], + "type": "text", + "content": " is obtained by directly projecting the indoor top-down view along the camera rays [28], assigning colors based on the corresponding intersections and bilinear interpolation. Specifically, the color at pixel " + }, + { + "bbox": [ + 55, + 601, + 296, + 673 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 601, + 296, + 673 + ], + "type": "text", + "content": " is determined by sampling the top-down image without learning a radiance field." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 684, + 296, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 684, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 134, + 684, + 296, + 716 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {c o l o r}} ^ {(u, v)} = \\sum_ {i} ^ {S} T _ {i} \\alpha_ {i} c _ {i} \\tag {3}", + "image_path": "ecfd744e8322907a6d33977754b23ee1a81acc71974f32ad7cdae44cebb23d05.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": " is the color copied from the top-down image. This approach directly maps color from the top-down view, offering a more straightforward and computationally efficient solution compared to NeRF [25], which reconstructs scenes by learning a radiance field for view synthesis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 133, + 556, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 556, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 556, + 312 + ], + "type": "text", + "content": "We employed a uniform voxel sampling strategy, where voxels were sampled along a fixed-length ray for both coarse color and coarse depth. However, this approach introduced banding artifacts, particularly noticeable on the floor directly beneath the camera in the coarse color image. Since the top-down view served as the floor texture, maintaining high-quality details was essential for achieving realistic rendering. To mitigate these artifacts, we reduced the ray length by half for coarse color sampling. This adjustment increased the density of sample points within the same spatial region, enhancing sampling accuracy and producing smoother color and texture transitions. As a result, banding artifacts were significantly diminished, improving overall rendering quality. For coarse depth, preserving scene structure was paramount, so we retained the original sampling strategy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 319, + 448, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 319, + 448, + 331 + ], + "spans": [ + { + "bbox": [ + 314, + 319, + 448, + 331 + ], + "type": "text", + "content": "3.3. Photorealistic Synthesis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 336, + 556, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 556, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 556, + 469 + ], + "type": "text", + "content": "Generating photorealistic panoramic images directly from top-down views is challenging. To address this, we propose a two-stage pipeline. In the second stage, the PanoGen module synthesizes photorealistic indoor panoramic images from coarse color and depth inputs. We implement PanoGen using a diffusion-based ControlNet [42], enabling the restoration of fine details. This two-stage approach not only reconstructs the house's structural layout, including precise wall positions, but also restores elements such as windows, lighting, and furniture. The final panoramic image " + }, + { + "bbox": [ + 313, + 336, + 556, + 469 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{pano}}" + }, + { + "bbox": [ + 313, + 336, + 556, + 469 + ], + "type": "text", + "content": " is generated based on the two coarse inputs from the previous stage:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 477, + 555, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 477, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 347, + 477, + 555, + 491 + ], + "type": "interline_equation", + "content": "I _ {\\text {p a n o}} = \\text {P a n o G e n} \\left(I _ {\\text {c o l o r}}, I _ {\\text {d e p t h}}\\right) _ {\\text {c o n d i t i o n}} \\tag {4}", + "image_path": "29774c62fb9f6b086a712d2b36df6d74ce6f9a7c20a1b60582294f07c3c63c52.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 499, + 556, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 556, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 556, + 620 + ], + "type": "text", + "content": "The PanoGen module modifies the conditioning mechanism of ControlNet by treating coarse color and depth images as separate inputs before combining them. This design enables PanoGen to capture both the scene's geometric structure and the furniture's color and position, resulting in more accurate and high-quality panoramic images. To further enhance consistency, we incorporate an alignment loss to prevent structural distortions when the viewpoint changes and a color loss to ensure accurate color reproduction in the synthesized output." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 626, + 463, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 463, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 463, + 639 + ], + "type": "text", + "content": "3.4. Training and Optimization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 643, + 555, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 643, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 643, + 555, + 680 + ], + "type": "text", + "content": "The Top2Pano model employs denoising MSE loss, alignment loss, and color loss functions for optimization. The denoising MSE loss function [42] is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 687, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 687, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 315, + 687, + 555, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f}} = \\mathbb {E} _ {z _ {0}, t, c _ {t}, c _ {c}, c _ {d}, \\epsilon \\sim \\mathcal {N} (0, 1)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, c _ {t}, c _ {c}, c _ {d}\\right) \\| _ {2} ^ {2} \\right], \\tag {5}", + "image_path": "4b7bdc9ddf8906a78b4eab21c8febd5a66b49cc6e8f73c7e03e1f6c8fa048852.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "28496" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " denotes the number of noise addition steps, " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " represents the corresponding coarse color image, and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " represents the coarse depth image. The variables " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "c_{t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " correspond to the time step and simple text prompts, respectively. The diffusion algorithm trains a neural network " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " to predict the noise added to the noisy image " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". The alignment loss is formulated as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 160, + 296, + 183 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 160, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 115, + 160, + 296, + 183 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a l i g n m e n t}} = \\left[ \\| I _ {D} - \\hat {I _ {D}} \\| _ {2} ^ {2} \\right], \\tag {6}", + "image_path": "a6e4eaf899bf05548b18a591fa44fe09c6f25e807973872d387123c564989109.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 188, + 295, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 295, + 223 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 188, + 295, + 223 + ], + "type": "inline_equation", + "content": "I_{D}" + }, + { + "bbox": [ + 55, + 188, + 295, + 223 + ], + "type": "text", + "content": " represents the depth images generated by the model and " + }, + { + "bbox": [ + 55, + 188, + 295, + 223 + ], + "type": "inline_equation", + "content": "\\hat{I_D}" + }, + { + "bbox": [ + 55, + 188, + 295, + 223 + ], + "type": "text", + "content": " represents the ground truth depth images. This loss function is used to address distortions in the furniture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "text", + "content": " be the rendered image and " + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "text", + "content": " the ground truth image, both with " + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "inline_equation", + "content": "C = 3" + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "text", + "content": " color channels. For each channel " + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "inline_equation", + "content": "c \\in \\{1, \\ldots, C\\}" + }, + { + "bbox": [ + 55, + 224, + 296, + 261 + ], + "type": "text", + "content": ", we define the normalized histogram as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 267, + 296, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 267, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 96, + 267, + 296, + 283 + ], + "type": "interline_equation", + "content": "H ^ {c} (I) = \\left(h _ {1} ^ {c} (I), h _ {2} ^ {c} (I), \\dots , h _ {\\mathrm {b i n s}} ^ {c} (I)\\right), \\tag {7}", + "image_path": "0e221c11e395bc548e3891eb054e3f93975d022fd51f949b3b245ba756f9cf92.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 289, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 337 + ], + "type": "text", + "content": "where bins is the number of histogram bins (e.g., 256), and each " + }, + { + "bbox": [ + 55, + 289, + 296, + 337 + ], + "type": "inline_equation", + "content": "h_k^c (I)" + }, + { + "bbox": [ + 55, + 289, + 296, + 337 + ], + "type": "text", + "content": " represents the normalized frequency (probability) of pixel intensities falling into bin " + }, + { + "bbox": [ + 55, + 289, + 296, + 337 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 289, + 296, + 337 + ], + "type": "text", + "content": ". Likewise, for the ground truth image " + }, + { + "bbox": [ + 55, + 289, + 296, + 337 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 289, + 296, + 337 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 345, + 296, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 345, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 91, + 345, + 296, + 361 + ], + "type": "interline_equation", + "content": "H ^ {c} (G) = \\big (h _ {1} ^ {c} (G), h _ {2} ^ {c} (G), \\ldots , h _ {\\mathrm {b i n s}} ^ {c} (G) \\big). \\qquad (8)", + "image_path": "ba8b1f9488a48b57a9da249b1dfccfbf3d9945e161a1f4cbb0f2b2040acd3991.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 366, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 392 + ], + "type": "text", + "content": "The color histogram loss, measured using the " + }, + { + "bbox": [ + 55, + 366, + 296, + 392 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 366, + 296, + 392 + ], + "type": "text", + "content": " norm, is given by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 91, + 398, + 296, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 398, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 91, + 398, + 296, + 466 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {c o l o r}} (I, G) = \\sum_ {c = 1} ^ {C} \\| H ^ {c} (I) - H ^ {c} (G) \\| _ {1} \\tag {9} \\\\ = \\sum_ {c = 1} ^ {C} \\sum_ {k = 1} ^ {\\text {b i n s}} \\left| h _ {k} ^ {c} (I) - h _ {k} ^ {c} (G) \\right|. \\\\ \\end{array}", + "image_path": "56d4b5ee468f9a4de922dbee95cde4ada762fbfca218c7d631168b19ca9b257b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 473, + 296, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 485 + ], + "type": "text", + "content": "The final loss function combines three loss terms as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 108, + 494, + 296, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 494, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 108, + 494, + 296, + 507 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {d i f f}} + \\mathcal {L} _ {\\text {a l g n m e n t}} + \\mathcal {L} _ {\\text {c o l o r}}. \\tag {10}", + "image_path": "5913e81493849929bfb360b1f2e8ac2e585570208da87bd1aaafca2d66a3631c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 514, + 255, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 255, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 255, + 527 + ], + "type": "text", + "content": "3.5. Generalization and Stylized Synthesis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 531, + 296, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 296, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 296, + 686 + ], + "type": "text", + "content": "Generalization to Floorplan. We aim to generate first-person view panoramas that faithfully represent a scene. While top-down views lack vertical details like walls and windows, this omission grants flexibility in generating panoramas. This flexibility is especially useful in interior design, where inputs are often simple floorplans. To improve generalization, particularly to schematic floorplans, we train on orthographic rather than perspective views, as they better match floorplans. Empirical results show that Top2Pano generalizes well to schematic and even hand-drawn floorplans while maintaining photorealism. Additionally, our model enables stylized synthesis guided by text or images, supporting diverse design needs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Text-Guided Stylization. The PanoGen module, built upon the text-driven Stable Diffusion model, inherently supports" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 205 + ], + "type": "text", + "content": "text-conditioned image generation. As illustrated in Figure 2, PanoGen synthesizes panoramas using three conditions: coarse depth, coarse colored image, and stylized textual guidance. When the input is a textureless floorplan, the stylized textual condition effectively guides the style of the synthesized result, as demonstrated in Figure 1. However, when a colored top-down view is provided, the influence of text-guided stylization becomes less pronounced. This occurs because the rendered coarse colored panorama constrains the final output to closely align with the input view, thereby diminishing the impact of the textual stylization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 206, + 556, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 556, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 556, + 315 + ], + "type": "text", + "content": "To enhance text-guided stylization, the weight of the coarse colored panorama in the PanoGen conditions can be reduced. However, this introduces a tradeoff: prioritizing stylization may come at the expense of fidelity to the top-down view. Regardless of this tradeoff, the coarse depth condition consistently ensures that the synthesized panorama adheres to the underlying scene geometry. Notably, this tradeoff does not apply to the text-to-panorama generation task, where textual guidance plays a more dominant role." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 321, + 556, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 556, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 556, + 524 + ], + "type": "text", + "content": "Image-Guided Stylization. Given several scene images (not necessarily panoramas), we can fine-tune the PanoGen module using low-rank adaptation (LoRA) [14] to generate panoramas that align with the visual styles present in the provided images. To guide this process, we introduce structured textural prompts augmented with style tags (e.g., [Japanese]) at the beginning of the input prompts. These tags act as conditional modifiers, steering the model toward synthesizing images that follow specific aesthetic themes, such as regional design styles. The framework is applied solely to the PanoGen module, ensuring both computational and parameter efficiency. Notably, our method requires fewer than five in-the-wild images per target style, substantially reducing data demands. This approach achieves its efficiency by decomposing weight updates into low-rank matrices. For a pretrained weight matrix " + }, + { + "bbox": [ + 313, + 321, + 556, + 524 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_0 \\in \\mathbb{R}^{d \\times d}" + }, + { + "bbox": [ + 313, + 321, + 556, + 524 + ], + "type": "text", + "content": ", the update " + }, + { + "bbox": [ + 313, + 321, + 556, + 524 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}" + }, + { + "bbox": [ + 313, + 321, + 556, + 524 + ], + "type": "text", + "content": " is constrained as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 334, + 537, + 555, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 537, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 334, + 537, + 555, + 552 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {W} = \\mathbf {B} \\mathbf {A}, \\quad \\text {w h e r e} \\mathbf {A} \\in \\mathbb {R} ^ {d \\times r}, \\mathbf {B} \\in \\mathbb {R} ^ {r \\times d} \\tag {11}", + "image_path": "0726aaeef57406a6dad8b610d16d0485bd61db65138c6422e136f51af38c1df7.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "type": "inline_equation", + "content": "r \\ll d" + }, + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "type": "text", + "content": " represents the intrinsic rank (we use " + }, + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "type": "inline_equation", + "content": "r = 8" + }, + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "type": "text", + "content": "). During fine-tuning, only the matrices " + }, + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "type": "text", + "content": " are updated, while the original weights " + }, + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_0" + }, + { + "bbox": [ + 313, + 565, + 555, + 613 + ], + "type": "text", + "content": " remain fixed. The forward pass is modified as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 373, + 627, + 555, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 627, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 373, + 627, + 555, + 640 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {\\text {o u t}} = \\mathbf {W} _ {0} \\mathbf {h} _ {\\text {i n}} + \\alpha \\cdot \\mathbf {B A h} _ {\\text {i n}} \\tag {12}", + "image_path": "de5791c040a4916c7b1b73fbd98b23668cad17ff9b899b63b64d52d50594db6b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": " is a scaling coefficient. This lightweight adaptation (0.8% new parameters) mitigates catastrophic forgetting and maintains the model's baseline generation quality for generic prompts while enabling precise style control through our [style] textual conditioning." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "28497" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 117 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 117 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 117 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 117 + ], + "type": "table", + "html": "
TrainingTesting
ScenesFloorsPanoramasScenesFloorsPanoramas
Matterport3D61127617714291405
Gibson152203537939761672
", + "image_path": "65f305cfb4c13676df540f952b0be67aaa8aade81f5e88d7ecbe01bed1e7b612.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 155, + 291, + 232 + ], + "blocks": [ + { + "bbox": [ + 56, + 121, + 295, + 143 + ], + "lines": [ + { + "bbox": [ + 56, + 121, + 295, + 143 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 295, + 143 + ], + "type": "text", + "content": "Table 1. The numbers of scenes, floors, and panorama images in the training and testing sets of the two datasets" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 155, + 291, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 155, + 291, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 155, + 291, + 232 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑FID ↓LPIPS ↓
Materport3DSat2Density[28]+LDM[29]11.270.4221100.060.6237
Sat2Density[28]+ControlNet[42]11.420.422285.780.6163
PanFusion[41]11.450.437285.740.6153
Top2Pano (Ours)11.720.440930.840.6029
GibsonSat2Density[28]+LDM[29]10.540.448084.330.6462
Sat2Density[28]+ControlNet[42]10.970.458285.210.6645
PanFusion[41]11.360.474479.530.6634
Top2Pano (Ours)11.580.485128.680.6282
", + "image_path": "4c9c883bd5bbeaa6bba8545b72ae1d3179f249a1a6f2ed88d3d293da7bbaa0cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 275, + 137, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 137, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 137, + 289 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 295, + 160, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 160, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 160, + 308 + ], + "type": "text", + "content": "4.1. Data Preparation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 313, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 296, + 504 + ], + "type": "text", + "content": "For evaluation, we use the Matterport3D [3] and Gibson [36] datasets. Since no existing dataset provides both top-down views and high-quality panoramic images, we generate top-down views from 3D models in these datasets using Blender. Specifically, we import textured 3D meshes into Blender and render top-down views with an orthographic camera. The top-down view we render closely resembles a floorplan, unlike the perspective-rendered views used in embodied dialog localization [12]. This similarity enhances our model's ability to generalize to floorplan inputs. To determine the number of floors in each scene, we apply DBSCAN [7] clustering to the camera positions within the datasets. We exclude certain scenes, such as airports and large supermarkets, as well as panoramic images depicting outdoor environments to ensure alignment with our task. After processing, the final dataset sizes are summarized in Table 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 168, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 168, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 168, + 526 + ], + "type": "text", + "content": "4.2. Evaluation Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 530, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 296, + 626 + ], + "type": "text", + "content": "To assess the quality of the generated panoramas, we employ both pixel-based and perceptual evaluation metrics. For pixel-level assessment, we utilize peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) to quantify image fidelity. Additionally, we incorporate perceptual metrics such as Fréchet Inception Distance (FID) [13] and Learned Perceptual Image Patch Similarity (LPIPS) [43] to capture higher-level visual realism." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 165, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 165, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 165, + 648 + ], + "type": "text", + "content": "4.3. Implement Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Our code runs on an NVIDIA RTX A6000 GPU with 48GB of memory. The model has 3.3 billion parameters and is trained with a batch size of 21 for 100 epochs. On average, each experiment takes approximately two days to complete on both the Matterport3D and Gibson datasets. We optimize" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 550, + 169 + ], + "blocks": [ + { + "bbox": [ + 55, + 236, + 295, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 295, + 258 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with existing methods on the Matterport3D [3] and Gibson [36] datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 550, + 169 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 550, + 169 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 550, + 169 + ], + "type": "table", + "html": "
ModulesMatterport3DGibson
seg floor wall depth colorPSNR↑SSIM↑FID↓LPIPS↓PSNR↑SSIM↑FID↓LPIPS↓
✓✓11.080.4122104.030.646311.070.4551117.080.6574
✓✓✓11.260.4206102.840.647111.100.4564115.880.6568
✓✓✓10.780.409155.730.616111.190.466832.280.6310
✓✓✓11.220.420754.440.619311.280.461383.310.6429
✓✓✓11.570.437843.990.606711.160.473381.870.6549
✓✓✓10.980.419653.340.611511.330.460133.890.6307
✓✓11.210.419935.080.613111.380.464136.290.6384
✓✓✓11.260.429934.610.606011.490.467334.360.6336
✓✓✓✓11.590.438167.020.619211.380.476165.660.6557
✓✓✓✓11.720.440930.840.602911.580.485128.680.6282
", + "image_path": "52b46dd0201b9cb851f318d47b311aca5baf4cc0e74b8767044495b48103cab5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 172, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 206 + ], + "type": "text", + "content": "Table 3. Ablation study on five designs in our Top2Pano model (top-down view segmentation, floor reinforcement, wall reinforcement, coarse depth panorama, and coarse colored panorama)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 228, + 555, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 264 + ], + "type": "text", + "content": "our model using the Adam optimizer [18] with default parameters " + }, + { + "bbox": [ + 313, + 228, + 555, + 264 + ], + "type": "inline_equation", + "content": "(\\beta_{1} = 0.9, \\beta_{2} = 0.999, \\epsilon = 10^{-8})" + }, + { + "bbox": [ + 313, + 228, + 555, + 264 + ], + "type": "text", + "content": " and a learning rate of " + }, + { + "bbox": [ + 313, + 228, + 555, + 264 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 313, + 228, + 555, + 264 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 273, + 505, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 505, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 505, + 286 + ], + "type": "text", + "content": "4.4. Comparison with Previous Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 291, + 556, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 556, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 556, + 447 + ], + "type": "text", + "content": "As shown in Table 2, we compare our method against three baseline approaches across four evaluation metrics. Sat2Density [28] is a satellite-to-ground panorama synthesis method, which we adapt for indoor panorama generation using latent diffusion model (LDM) [29] and ControlNet [42]. PanFusion [41] is a text-to-panorama generation framework that also incorporates layout-conditioned generation via ControlNet. Table 2 shows that our method outperforms all baselines across all four metrics on both datasets, demonstrating its effectiveness. Furthermore, qualitative comparisons in Figures 3 and 4 highlight that our approach generates more realistic and structurally accurate house reconstructions, including furniture placement." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 456, + 408, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 408, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 408, + 468 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 473, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 569 + ], + "type": "text", + "content": "We conducted comprehensive ablation studies to analyze and validate the contribution of each component in our model. Specifically, we performed experiments comparing several model variants against the original. These experiments involved removing key elements such as the structural reinforcement of the floor and walls, the segmentation input to the OccRecon module, and the coarse depth and colored panoramicas as conditional inputs to the PanoGen module." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "As shown in Table 3, our original model achieves the highest overall scores across all four metrics, with any modification leading to some degree of performance degradation. Removing coarse colored panoramas or the embedded floor significantly disrupts furniture placement and color accuracy. While the room structure remains mostly intact, furniture positions become unreliable, and color representations appear distorted. Conversely, excluding coarse depth panoramas or embedded walls maintains color and furniture accuracy but compromises the spatial understanding and overall quality of room structure reconstruction. These effects are further illustrated in the qualitative results in the supp. materials." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "28498" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 115, + 125 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 115, + 125 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 115, + 125 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 115, + 125 + ], + "type": "image", + "image_path": "d87bfacbbc6d25388816b311b302213cb3d0bb46625e8a1b8a330748ebf26ffd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 117, + 70, + 223, + 125 + ], + "blocks": [ + { + "bbox": [ + 117, + 70, + 223, + 125 + ], + "lines": [ + { + "bbox": [ + 117, + 70, + 223, + 125 + ], + "spans": [ + { + "bbox": [ + 117, + 70, + 223, + 125 + ], + "type": "image", + "image_path": "3ca2b08a577bc0255d88476a19b5b742486e8bc1aabfbb723f06961184c25bd7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 332, + 125 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 332, + 125 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 332, + 125 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 332, + 125 + ], + "type": "image", + "image_path": "e220fba2817c759044c7626b257b0c9b4adfc237980f430092901eb295208e35.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 333, + 70, + 439, + 125 + ], + "blocks": [ + { + "bbox": [ + 333, + 70, + 439, + 125 + ], + "lines": [ + { + "bbox": [ + 333, + 70, + 439, + 125 + ], + "spans": [ + { + "bbox": [ + 333, + 70, + 439, + 125 + ], + "type": "image", + "image_path": "dab818e600dee898acfc2dc454e591b49ce9d5b58d62c1aff061ef0fa156a63d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 441, + 70, + 547, + 125 + ], + "blocks": [ + { + "bbox": [ + 441, + 70, + 547, + 125 + ], + "lines": [ + { + "bbox": [ + 441, + 70, + 547, + 125 + ], + "spans": [ + { + "bbox": [ + 441, + 70, + 547, + 125 + ], + "type": "image", + "image_path": "d6454c2d8d9f2c5252d908f6876fd693422c00b507e8997bb017a8d2b70043d3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 126, + 115, + 179 + ], + "blocks": [ + { + "bbox": [ + 61, + 126, + 115, + 179 + ], + "lines": [ + { + "bbox": [ + 61, + 126, + 115, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 126, + 115, + 179 + ], + "type": "image", + "image_path": "f5f7da3d5ae4cdde7b4064799ae46cde9ec32987c8729f96f8f682d3354539fc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 117, + 126, + 223, + 179 + ], + "blocks": [ + { + "bbox": [ + 117, + 126, + 223, + 179 + ], + "lines": [ + { + "bbox": [ + 117, + 126, + 223, + 179 + ], + "spans": [ + { + "bbox": [ + 117, + 126, + 223, + 179 + ], + "type": "image", + "image_path": "75ad0ed1229b98c0dbca869cc726476e855f3964e596058b116451dfbddc4b47.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 225, + 126, + 332, + 179 + ], + "blocks": [ + { + "bbox": [ + 225, + 126, + 332, + 179 + ], + "lines": [ + { + "bbox": [ + 225, + 126, + 332, + 179 + ], + "spans": [ + { + "bbox": [ + 225, + 126, + 332, + 179 + ], + "type": "image", + "image_path": "2348aef428cb6c447d27ce7ae2b820d3afed552c37d78c221a5543c0157a9d1f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 333, + 126, + 439, + 179 + ], + "blocks": [ + { + "bbox": [ + 333, + 126, + 439, + 179 + ], + "lines": [ + { + "bbox": [ + 333, + 126, + 439, + 179 + ], + "spans": [ + { + "bbox": [ + 333, + 126, + 439, + 179 + ], + "type": "image", + "image_path": "41173e3b0c831caa22b0d9be55a8dd33d3b34beef106cb034f5b38dcd1cd8388.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 441, + 126, + 547, + 179 + ], + "blocks": [ + { + "bbox": [ + 441, + 126, + 547, + 179 + ], + "lines": [ + { + "bbox": [ + 441, + 126, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 441, + 126, + 547, + 179 + ], + "type": "image", + "image_path": "fb192194dd9117f977de4e20e85edd5efc1511e495e37f2bfaa00d955589b7db.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 61, + 181, + 115, + 236 + ], + "blocks": [ + { + "bbox": [ + 61, + 181, + 115, + 236 + ], + "lines": [ + { + "bbox": [ + 61, + 181, + 115, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 181, + 115, + 236 + ], + "type": "image", + "image_path": "fe2fbd79e74a5ad1b0a976300bc0e99b339fcd993ff76e81b541f306f1d64780.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 117, + 181, + 223, + 236 + ], + "blocks": [ + { + "bbox": [ + 117, + 181, + 223, + 236 + ], + "lines": [ + { + "bbox": [ + 117, + 181, + 223, + 236 + ], + "spans": [ + { + "bbox": [ + 117, + 181, + 223, + 236 + ], + "type": "image", + "image_path": "3426376a90bee262dc255136afccf5f52ffc8fe24712aefff2cfc93a82dbfbea.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 225, + 181, + 332, + 236 + ], + "blocks": [ + { + "bbox": [ + 225, + 181, + 332, + 236 + ], + "lines": [ + { + "bbox": [ + 225, + 181, + 332, + 236 + ], + "spans": [ + { + "bbox": [ + 225, + 181, + 332, + 236 + ], + "type": "image", + "image_path": "10b3cc6e7819e1040a40a49e6805c63f35f2e99e97adfbc7d12e7b9532764ffb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 333, + 181, + 439, + 236 + ], + "blocks": [ + { + "bbox": [ + 333, + 181, + 439, + 236 + ], + "lines": [ + { + "bbox": [ + 333, + 181, + 439, + 236 + ], + "spans": [ + { + "bbox": [ + 333, + 181, + 439, + 236 + ], + "type": "image", + "image_path": "67e10f29dc3dac2670427d6102daaab6639597534db638ac9ae29272a2172d32.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 441, + 181, + 547, + 236 + ], + "blocks": [ + { + "bbox": [ + 441, + 181, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 441, + 181, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 441, + 181, + 547, + 236 + ], + "type": "image", + "image_path": "11ab96e5e6346a65369b30f3e2581fca1700ce78d394df5d3e1b3654120bbdc6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 61, + 237, + 115, + 290 + ], + "blocks": [ + { + "bbox": [ + 61, + 237, + 115, + 290 + ], + "lines": [ + { + "bbox": [ + 61, + 237, + 115, + 290 + ], + "spans": [ + { + "bbox": [ + 61, + 237, + 115, + 290 + ], + "type": "image", + "image_path": "d261615fde255af288974800b7dd9e7c9836e1b410e134e97253837a45fb78c3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 117, + 237, + 223, + 290 + ], + "blocks": [ + { + "bbox": [ + 117, + 237, + 223, + 290 + ], + "lines": [ + { + "bbox": [ + 117, + 237, + 223, + 290 + ], + "spans": [ + { + "bbox": [ + 117, + 237, + 223, + 290 + ], + "type": "image", + "image_path": "e0c73e5110e422f43205eee42ed5486b5d3b29226158e5fc75f728f4a3909f48.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 225, + 236, + 332, + 290 + ], + "blocks": [ + { + "bbox": [ + 225, + 236, + 332, + 290 + ], + "lines": [ + { + "bbox": [ + 225, + 236, + 332, + 290 + ], + "spans": [ + { + "bbox": [ + 225, + 236, + 332, + 290 + ], + "type": "image", + "image_path": "ee5b0a2f7f93b90f13aa4246689e5f90868d3abe4ad49b80c40986e1e9ac86de.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 333, + 236, + 439, + 290 + ], + "blocks": [ + { + "bbox": [ + 333, + 236, + 439, + 290 + ], + "lines": [ + { + "bbox": [ + 333, + 236, + 439, + 290 + ], + "spans": [ + { + "bbox": [ + 333, + 236, + 439, + 290 + ], + "type": "image", + "image_path": "f4a9af37419317d636564f7f7cff1e21c0c46d045673fe8b898d89b2408d2ba0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 441, + 237, + 547, + 290 + ], + "blocks": [ + { + "bbox": [ + 441, + 237, + 547, + 290 + ], + "lines": [ + { + "bbox": [ + 441, + 237, + 547, + 290 + ], + "spans": [ + { + "bbox": [ + 441, + 237, + 547, + 290 + ], + "type": "image", + "image_path": "1b40e50d7b666915888905233a3b749b7a8e10ce11572bf9a2be83ad0dc765b5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 61, + 293, + 115, + 346 + ], + "blocks": [ + { + "bbox": [ + 61, + 293, + 115, + 346 + ], + "lines": [ + { + "bbox": [ + 61, + 293, + 115, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 293, + 115, + 346 + ], + "type": "image", + "image_path": "6edcdca92b1bdc686e11fec9330d3bdb8bfd32397fd5076cffbfa7777824660a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 347, + 111, + 357 + ], + "lines": [ + { + "bbox": [ + 66, + 347, + 111, + 357 + ], + "spans": [ + { + "bbox": [ + 66, + 347, + 111, + 357 + ], + "type": "text", + "content": "(a) Top-down" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 117, + 292, + 223, + 346 + ], + "blocks": [ + { + "bbox": [ + 117, + 292, + 223, + 346 + ], + "lines": [ + { + "bbox": [ + 117, + 292, + 223, + 346 + ], + "spans": [ + { + "bbox": [ + 117, + 292, + 223, + 346 + ], + "type": "image", + "image_path": "f8e49994ea7e6a2599dd1cb16ec253e349de08635308ebce6ccd6983484f79d5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 347, + 203, + 357 + ], + "lines": [ + { + "bbox": [ + 136, + 347, + 203, + 357 + ], + "spans": [ + { + "bbox": [ + 136, + 347, + 203, + 357 + ], + "type": "text", + "content": "(b) Sat2Density [28]" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 225, + 292, + 332, + 346 + ], + "blocks": [ + { + "bbox": [ + 225, + 292, + 332, + 346 + ], + "lines": [ + { + "bbox": [ + 225, + 292, + 332, + 346 + ], + "spans": [ + { + "bbox": [ + 225, + 292, + 332, + 346 + ], + "type": "image", + "image_path": "ea2eca47b313d457ee0e09d8e10eb968faa0f6789babfd373153daf2b6e0e402.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 347, + 309, + 357 + ], + "lines": [ + { + "bbox": [ + 247, + 347, + 309, + 357 + ], + "spans": [ + { + "bbox": [ + 247, + 347, + 309, + 357 + ], + "type": "text", + "content": "(c) PanFusion [41]" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 333, + 292, + 439, + 346 + ], + "blocks": [ + { + "bbox": [ + 333, + 292, + 439, + 346 + ], + "lines": [ + { + "bbox": [ + 333, + 292, + 439, + 346 + ], + "spans": [ + { + "bbox": [ + 333, + 292, + 439, + 346 + ], + "type": "image", + "image_path": "d629db6494fc760ba7f900d97a389c87fe76ff2726a5e4151385b61d555da4c5.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 347, + 400, + 357 + ], + "lines": [ + { + "bbox": [ + 372, + 347, + 400, + 357 + ], + "spans": [ + { + "bbox": [ + 372, + 347, + 400, + 357 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 441, + 292, + 547, + 346 + ], + "blocks": [ + { + "bbox": [ + 441, + 292, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 441, + 292, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 441, + 292, + 547, + 346 + ], + "type": "image", + "image_path": "3e35433d217b9c84193629a8f88df49aeeb454c2a7f1a711a1344bbf68d9e5a9.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 347, + 522, + 357 + ], + "lines": [ + { + "bbox": [ + 466, + 347, + 522, + 357 + ], + "spans": [ + { + "bbox": [ + 466, + 347, + 522, + 357 + ], + "type": "text", + "content": "(e) Ground Truth" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 61, + 384, + 115, + 437 + ], + "blocks": [ + { + "bbox": [ + 61, + 384, + 115, + 437 + ], + "lines": [ + { + "bbox": [ + 61, + 384, + 115, + 437 + ], + "spans": [ + { + "bbox": [ + 61, + 384, + 115, + 437 + ], + "type": "image", + "image_path": "c14d0188b5709c4fcbef83d223b3baff16efbd9fcead728a823820e16b301518.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 117, + 384, + 223, + 437 + ], + "blocks": [ + { + "bbox": [ + 189, + 361, + 420, + 373 + ], + "lines": [ + { + "bbox": [ + 189, + 361, + 420, + 373 + ], + "spans": [ + { + "bbox": [ + 189, + 361, + 420, + 373 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons on the Matterport3D dataset." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 117, + 384, + 223, + 437 + ], + "lines": [ + { + "bbox": [ + 117, + 384, + 223, + 437 + ], + "spans": [ + { + "bbox": [ + 117, + 384, + 223, + 437 + ], + "type": "image", + "image_path": "bc66637b156586e359c5d763c911da2cd6e54ce72334141b7a32be5fce994a7b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 225, + 384, + 332, + 437 + ], + "blocks": [ + { + "bbox": [ + 225, + 384, + 332, + 437 + ], + "lines": [ + { + "bbox": [ + 225, + 384, + 332, + 437 + ], + "spans": [ + { + "bbox": [ + 225, + 384, + 332, + 437 + ], + "type": "image", + "image_path": "223edc01e4497d93b35884ac4c903647637fe4fb48b10291873c1bd5ad10a030.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 333, + 384, + 439, + 437 + ], + "blocks": [ + { + "bbox": [ + 333, + 384, + 439, + 437 + ], + "lines": [ + { + "bbox": [ + 333, + 384, + 439, + 437 + ], + "spans": [ + { + "bbox": [ + 333, + 384, + 439, + 437 + ], + "type": "image", + "image_path": "f05c4ef89b20a0e2e1237535624cd2f5b6cbb051067fa6eab9fc5f027f22caee.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 441, + 384, + 547, + 437 + ], + "blocks": [ + { + "bbox": [ + 441, + 384, + 547, + 437 + ], + "lines": [ + { + "bbox": [ + 441, + 384, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 441, + 384, + 547, + 437 + ], + "type": "image", + "image_path": "284404ffe703471b9303da5234fb7ef40c9d5e000682f300df7d45a0ec3d2f9f.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 61, + 439, + 115, + 493 + ], + "blocks": [ + { + "bbox": [ + 61, + 439, + 115, + 493 + ], + "lines": [ + { + "bbox": [ + 61, + 439, + 115, + 493 + ], + "spans": [ + { + "bbox": [ + 61, + 439, + 115, + 493 + ], + "type": "image", + "image_path": "0db089f8e1927247b9f64db373875313f8eb6fc7bad8ca1644dd5cd83cde8c6a.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 117, + 439, + 223, + 493 + ], + "blocks": [ + { + "bbox": [ + 117, + 439, + 223, + 493 + ], + "lines": [ + { + "bbox": [ + 117, + 439, + 223, + 493 + ], + "spans": [ + { + "bbox": [ + 117, + 439, + 223, + 493 + ], + "type": "image", + "image_path": "fb8385ff7fb9c9656d5b7030844879f209718570148899aba6e1e8286ac99e1e.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 225, + 439, + 332, + 493 + ], + "blocks": [ + { + "bbox": [ + 225, + 439, + 332, + 493 + ], + "lines": [ + { + "bbox": [ + 225, + 439, + 332, + 493 + ], + "spans": [ + { + "bbox": [ + 225, + 439, + 332, + 493 + ], + "type": "image", + "image_path": "af0af0dff358f985d9eabdf9f30854ae5126e1ef205533739985f1e90181fcc0.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 333, + 439, + 439, + 493 + ], + "blocks": [ + { + "bbox": [ + 333, + 439, + 439, + 493 + ], + "lines": [ + { + "bbox": [ + 333, + 439, + 439, + 493 + ], + "spans": [ + { + "bbox": [ + 333, + 439, + 439, + 493 + ], + "type": "image", + "image_path": "b72e69f0ce8dc37538fc9b372ec4f33b91a85d401b8efa37d1e69aafb350f367.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 441, + 439, + 547, + 493 + ], + "blocks": [ + { + "bbox": [ + 441, + 439, + 547, + 493 + ], + "lines": [ + { + "bbox": [ + 441, + 439, + 547, + 493 + ], + "spans": [ + { + "bbox": [ + 441, + 439, + 547, + 493 + ], + "type": "image", + "image_path": "22b649c8498fa3f2069c391825e9625df562d97c8777744372d84556772beb96.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 61, + 495, + 115, + 547 + ], + "blocks": [ + { + "bbox": [ + 61, + 495, + 115, + 547 + ], + "lines": [ + { + "bbox": [ + 61, + 495, + 115, + 547 + ], + "spans": [ + { + "bbox": [ + 61, + 495, + 115, + 547 + ], + "type": "image", + "image_path": "3105be3aacb1178bd379a3a4559793fb7a1e91a71f067604035014c37e3e5c92.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 117, + 495, + 223, + 548 + ], + "blocks": [ + { + "bbox": [ + 117, + 495, + 223, + 548 + ], + "lines": [ + { + "bbox": [ + 117, + 495, + 223, + 548 + ], + "spans": [ + { + "bbox": [ + 117, + 495, + 223, + 548 + ], + "type": "image", + "image_path": "25c43bb4407057ca1401bad007444e19a6bab19777f98842606b27b490c9b705.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 225, + 495, + 332, + 548 + ], + "blocks": [ + { + "bbox": [ + 225, + 495, + 332, + 548 + ], + "lines": [ + { + "bbox": [ + 225, + 495, + 332, + 548 + ], + "spans": [ + { + "bbox": [ + 225, + 495, + 332, + 548 + ], + "type": "image", + "image_path": "3cccea65905067192fb436a816d70d258ce8907d519b0fc53233e6cbe43da0e5.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 333, + 495, + 439, + 548 + ], + "blocks": [ + { + "bbox": [ + 333, + 495, + 439, + 548 + ], + "lines": [ + { + "bbox": [ + 333, + 495, + 439, + 548 + ], + "spans": [ + { + "bbox": [ + 333, + 495, + 439, + 548 + ], + "type": "image", + "image_path": "cbe4164e2c0a9d6548b9b168044fea952d010f4764a162c0876e163a32c49a1b.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 441, + 495, + 547, + 548 + ], + "blocks": [ + { + "bbox": [ + 441, + 495, + 547, + 548 + ], + "lines": [ + { + "bbox": [ + 441, + 495, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 441, + 495, + 547, + 548 + ], + "type": "image", + "image_path": "aff2a99c7786f45ad56c41a9428a718b2db168ccad78cae9595139915c3cb531.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 61, + 549, + 115, + 604 + ], + "blocks": [ + { + "bbox": [ + 61, + 549, + 115, + 604 + ], + "lines": [ + { + "bbox": [ + 61, + 549, + 115, + 604 + ], + "spans": [ + { + "bbox": [ + 61, + 549, + 115, + 604 + ], + "type": "image", + "image_path": "a6516874589b8658a39e1599fa4a02d532010c2f7145f7bf1aee50ed6297019a.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 117, + 549, + 223, + 604 + ], + "blocks": [ + { + "bbox": [ + 117, + 549, + 223, + 604 + ], + "lines": [ + { + "bbox": [ + 117, + 549, + 223, + 604 + ], + "spans": [ + { + "bbox": [ + 117, + 549, + 223, + 604 + ], + "type": "image", + "image_path": "9eb348b30e81c6d5ce07f54c50d188fa9f80800e7dfdb75956eceb1602c9f2b7.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 225, + 549, + 332, + 604 + ], + "blocks": [ + { + "bbox": [ + 225, + 549, + 332, + 604 + ], + "lines": [ + { + "bbox": [ + 225, + 549, + 332, + 604 + ], + "spans": [ + { + "bbox": [ + 225, + 549, + 332, + 604 + ], + "type": "image", + "image_path": "9651671a11b4ba16871f82afc7c6627c3e00276eac76b7ee7f7e70b655ead32c.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 333, + 549, + 439, + 604 + ], + "blocks": [ + { + "bbox": [ + 333, + 549, + 439, + 604 + ], + "lines": [ + { + "bbox": [ + 333, + 549, + 439, + 604 + ], + "spans": [ + { + "bbox": [ + 333, + 549, + 439, + 604 + ], + "type": "image", + "image_path": "2ce4e1b5238ee97ddc15fe9d29318d483bc0ad3f77fa5989fa6e6fe5e1d5c6c3.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 441, + 549, + 547, + 604 + ], + "blocks": [ + { + "bbox": [ + 441, + 549, + 547, + 604 + ], + "lines": [ + { + "bbox": [ + 441, + 549, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 441, + 549, + 547, + 604 + ], + "type": "image", + "image_path": "f48ef503b889b0948444e0128a995a2d1ddc10ed6b726a4b911bdc15927c3e43.jpg" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_body" + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 61, + 605, + 115, + 659 + ], + "blocks": [ + { + "bbox": [ + 61, + 605, + 115, + 659 + ], + "lines": [ + { + "bbox": [ + 61, + 605, + 115, + 659 + ], + "spans": [ + { + "bbox": [ + 61, + 605, + 115, + 659 + ], + "type": "image", + "image_path": "1531d71541dcd9280f6684c6862637804b1ead2ddf6762d121623b8c07cfac53.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 661, + 111, + 670 + ], + "lines": [ + { + "bbox": [ + 66, + 661, + 111, + 670 + ], + "spans": [ + { + "bbox": [ + 66, + 661, + 111, + 670 + ], + "type": "text", + "content": "(a) Top-down" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 117, + 605, + 223, + 659 + ], + "blocks": [ + { + "bbox": [ + 117, + 605, + 223, + 659 + ], + "lines": [ + { + "bbox": [ + 117, + 605, + 223, + 659 + ], + "spans": [ + { + "bbox": [ + 117, + 605, + 223, + 659 + ], + "type": "image", + "image_path": "746aacbb27b068d1b2a05b0f8494a04c5f0afb44fc3a1462cac537e24b9d597e.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 136, + 661, + 203, + 670 + ], + "lines": [ + { + "bbox": [ + 136, + 661, + 203, + 670 + ], + "spans": [ + { + "bbox": [ + 136, + 661, + 203, + 670 + ], + "type": "text", + "content": "(b) Sat2Density [28]" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 225, + 605, + 332, + 659 + ], + "blocks": [ + { + "bbox": [ + 225, + 605, + 332, + 659 + ], + "lines": [ + { + "bbox": [ + 225, + 605, + 332, + 659 + ], + "spans": [ + { + "bbox": [ + 225, + 605, + 332, + 659 + ], + "type": "image", + "image_path": "6b8b42fa12bebc8a610f4b66ac65710f1fc61c3d641929e8e3e62452604cfcc1.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 661, + 309, + 670 + ], + "lines": [ + { + "bbox": [ + 247, + 661, + 309, + 670 + ], + "spans": [ + { + "bbox": [ + 247, + 661, + 309, + 670 + ], + "type": "text", + "content": "(c) PanFusion [41]" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 333, + 605, + 439, + 659 + ], + "blocks": [ + { + "bbox": [ + 333, + 605, + 439, + 659 + ], + "lines": [ + { + "bbox": [ + 333, + 605, + 439, + 659 + ], + "spans": [ + { + "bbox": [ + 333, + 605, + 439, + 659 + ], + "type": "image", + "image_path": "fe67f44e29f29015cdf3ba2dcc34507800e3ea435d7b3c635134d10929fcd704.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 661, + 400, + 670 + ], + "lines": [ + { + "bbox": [ + 372, + 661, + 400, + 670 + ], + "spans": [ + { + "bbox": [ + 372, + 661, + 400, + 670 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 441, + 605, + 547, + 659 + ], + "blocks": [ + { + "bbox": [ + 441, + 605, + 547, + 659 + ], + "lines": [ + { + "bbox": [ + 441, + 605, + 547, + 659 + ], + "spans": [ + { + "bbox": [ + 441, + 605, + 547, + 659 + ], + "type": "image", + "image_path": "560c00334ce9b6bfc96c237b6c9333336e4f7cc21d64e2c2230b4a3bc8c80e8f.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 661, + 522, + 670 + ], + "lines": [ + { + "bbox": [ + 465, + 661, + 522, + 670 + ], + "spans": [ + { + "bbox": [ + 465, + 661, + 522, + 670 + ], + "type": "text", + "content": "(e) Ground Truth" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 201, + 674, + 408, + 685 + ], + "lines": [ + { + "bbox": [ + 201, + 674, + 408, + 685 + ], + "spans": [ + { + "bbox": [ + 201, + 674, + 408, + 685 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons on the Gibson dataset." + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "28499" + } + ] + } + ], + "index": 62 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 81, + 70, + 269, + 111 + ], + "blocks": [ + { + "bbox": [ + 81, + 70, + 269, + 111 + ], + "lines": [ + { + "bbox": [ + 81, + 70, + 269, + 111 + ], + "spans": [ + { + "bbox": [ + 81, + 70, + 269, + 111 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑FID ↓LPIPS ↓
G → M11.080.439740.740.6285
M → G11.030.436645.870.6353
", + "image_path": "7311c6f1b39e342a05b4e70eebbd812954f58f5c573d95902b26f2482de4c617.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 114, + 297, + 136 + ], + "lines": [ + { + "bbox": [ + 55, + 114, + 297, + 136 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 297, + 136 + ], + "type": "text", + "content": "Table 4. Cross-dataset evaluation. \"G\" represents Gibson; \"M\" represents Matterport3D." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 78, + 139, + 271, + 259 + ], + "blocks": [ + { + "bbox": [ + 78, + 139, + 271, + 259 + ], + "lines": [ + { + "bbox": [ + 78, + 139, + 271, + 259 + ], + "spans": [ + { + "bbox": [ + 78, + 139, + 271, + 259 + ], + "type": "image", + "image_path": "eb27ff5d3b186bf25393e6e2934fbc1426ed1b363a55116bcf741168233b5655.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 296, + 294 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 294 + ], + "type": "text", + "content": "Figure 5. Our Top2Pano model generalizes to both textured floorplans (first row) and hand-drawn sketched floorplans (second row), incorporating stylized [Japanese] control." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 272, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 272, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 272, + 313 + ], + "type": "text", + "content": "4.6. Generalization, Stylization, Manipulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 318, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 389 + ], + "type": "text", + "content": "We employed cross-dataset evaluation to assess our model's generalization capability. As shown in Table 4, although there is a slight decline in metric scores, our model maintains strong performance. Additionally, the trained parameters tend to generate photorealistic walls that reflect characteristics of the training dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 628 + ], + "type": "text", + "content": "Our model also generalizes well to floorplans. When given a specific floorplan and camera positions, the model assists users in generating indoor panoramic images. By using different text prompts, users can create various interior design styles, explore the house from a first-person perspective, and modify its style according to their preferences. We tested our model with three types of floorplans. The first type is a colored floorplan (Figure 5, first row), which provides detailed information about room colors and furniture hues, making it highly informative. The second type is a plain floorplan (Figure 1, bottom), which lacks color information and shows only the room structure and furniture layout. The third type is a hand-drawn floorplan sketch (Figure 5, first row), which offers a rough visual representation of the room. Our model successfully generates accurate panoramic images from these floorplans and adapts the visual style based on the provided textual descriptions. Moreover, our model enables panorama manipulation by editing objects in the floorplan, such as adding new items, as illustrated in Figure 6." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 636, + 132, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 132, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 132, + 647 + ], + "type": "text", + "content": "4.7. Limitations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 677 + ], + "type": "text", + "content": "Failure cases. Figure 7 shows representative failure cases. We annotate different types of failures with numbered labels:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 291, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 55, + 677, + 291, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 291, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 291, + 689 + ], + "type": "text", + "content": "- Ceiling: ① missing fan, ② vaulted ceiling, ③ false light;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 277, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 277, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 277, + 700 + ], + "type": "text", + "content": "- Wall: 4 height error, 5 false or missing decorations;" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 701, + 170, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 170, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 170, + 712 + ], + "type": "text", + "content": "- Window: 6 false window;" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 338, + 70, + 530, + 190 + ], + "blocks": [ + { + "bbox": [ + 338, + 70, + 530, + 190 + ], + "lines": [ + { + "bbox": [ + 338, + 70, + 530, + 190 + ], + "spans": [ + { + "bbox": [ + 338, + 70, + 530, + 190 + ], + "type": "image", + "image_path": "45012e95f40adb80c8263859436dab7eb4b3f3599ce770cc79ad7ff8b743e267.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 193, + 555, + 237 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 555, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 555, + 237 + ], + "type": "text", + "content": "Figure 6. Top2Pano enables panorama manipulation via compositional floorplan editing. In the second row, adding a rectangular object to the floorplan (compared to the first row) leads the model to generate a washstand with a mirror in the panorama." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 239, + 558, + 399 + ], + "blocks": [ + { + "bbox": [ + 317, + 239, + 558, + 399 + ], + "lines": [ + { + "bbox": [ + 317, + 239, + 558, + 399 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 558, + 399 + ], + "type": "image", + "image_path": "1fc250a1e614b2c9d2c81f0f66d5713977d5d4e5d91f9c52eec3769209a8769e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 402, + 531, + 414 + ], + "lines": [ + { + "bbox": [ + 336, + 402, + 531, + 414 + ], + "spans": [ + { + "bbox": [ + 336, + 402, + 531, + 414 + ], + "type": "text", + "content": "Figure 7. Failure cases (zoom in to view error types)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 419, + 479, + 454 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 419, + 430, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 419, + 430, + 431 + ], + "spans": [ + { + "bbox": [ + 314, + 419, + 430, + 431 + ], + "type": "text", + "content": "- Furniture: height error;" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 432, + 479, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 432, + 479, + 443 + ], + "spans": [ + { + "bbox": [ + 314, + 432, + 479, + 443 + ], + "type": "text", + "content": "- Thin object: 8 missing flat-screen TV;" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 443, + 423, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 443, + 423, + 454 + ], + "spans": [ + { + "bbox": [ + 314, + 443, + 423, + 454 + ], + "type": "text", + "content": "Stairs: false direction." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 456, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 514 + ], + "type": "text", + "content": "The failure cases stem from the ambiguity of the 2D input, leading to hallucinated objects that are not observable from the top-down view. The limitations in handling vertical structural details are largely due to the inherent ambiguity of the task." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 518, + 554, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 554, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 554, + 566 + ], + "type": "text", + "content": "Limited Vertical FoV. The generated panoramas exhibit a limited vertical field of view (FoV), reflecting the constraints of the training data. We expect improved performance with future datasets that include full vertical FoV panoramas." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 576, + 391, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 391, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 391, + 588 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "We present Top2Pano, a novel method for generating high-quality " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " indoor panoramas from 2D top-down views. The model first estimates volumetric occupancy to infer 3D structure, then applies volumetric rendering to produce coarse color and depth panoramas. These guide a diffusion-based refinement stage via ControlNet. To our knowledge, this is the first approach to generate panoramas from top-down views. Experiments on two datasets show that Top2Pano outperforms baselines in reconstructing room layouts and realistic furniture." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 757 + ], + "type": "text", + "content": "28500" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "text", + "content": "[1] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. Diverse plausible 360-degree image outpainting for efficient 3dcg background creation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 11431-11440, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 212 + ], + "type": "text", + "content": "[2] Sherwin Bahmani, Jeong Joon Park, Despoina Paschalidou, Xingguang Yan, Gordon Wetzstein, Leonidas J. Guibas, and Andrea Tagliasacchi. CC3D: layout-conditioned generation of compositional 3d scenes. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 7137-7147, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 213, + 296, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 296, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 296, + 267 + ], + "type": "text", + "content": "[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 269, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 269, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 269, + 296, + 312 + ], + "type": "text", + "content": "[4] Minglin Chen, Longguang Wang, Sheng Ao, Ye Zhang, Kai Xu, and Yulan Guo. Layout2scene: 3d semantic layout guided scene generation via geometry and appearance diffusion priors. arXiv preprint arXiv:2501.02519, 2025. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 314, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 296, + 346 + ], + "type": "text", + "content": "[5] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 296, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 296, + 401 + ], + "type": "text", + "content": "[6] Mohammad Reza Karimi Dastjerdi, Yannick Hold-Geoffroy, Jonathan Eisenmann, Siavash Khodadadeh, and Jean-François Lalonde. Guided co-modulated gan for " + }, + { + "bbox": [ + 62, + 347, + 296, + 401 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 62, + 347, + 296, + 401 + ], + "type": "text", + "content": " field of view extrapolation. In 2022 International Conference on 3D Vision (3DV), pages 475–485, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 456 + ], + "type": "text", + "content": "[7] Martin Ester, Hans-Peter Kriegel, Jörg Sander, and Xiaowei Xu. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD), pages 226-231, 1996. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 296, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 296, + 500 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 296, + 500 + ], + "type": "text", + "content": "[8] Chuan Fang, Xiaotao Hu, Kunming Luo, and Ping Tan. Ctrlroom: Controllable text-to-3d room meshes generation with layout constraints. arXiv preprint arXiv:2310.03602, 2023. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 296, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 546 + ], + "type": "text", + "content": "[9] Chengzeng Feng, Jiacheng Wei, Cheng Chen, Yang Li, Pan Ji, Fayao Liu, Hongdong Li, and Guosheng Lin. Prim2room: Layout-controllable room mesh generation from primitives. arXiv preprint arXiv:2409.05380, 2024. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 296, + 590 + ], + "type": "text", + "content": "[10] Mengyang Feng, Jinlin Liu, Miaomiao Cui, and Xuansong Xie. Diffusion360: Seamless 360 degree panoramic image generation based on diffusion models. arXiv preprint arXiv:2311.13141, 2023. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 592, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 296, + 646 + ], + "type": "text", + "content": "[11] Julia Guerrero-Viu, Clara Fernandez-Labrador, Cédric Demonceaux, and José Jesús Guerrero. What's in my room? object recognition on indoor panoramic images. In IEEE International Conference on Robotics and Automation, ICRA, pages 567-573, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "text", + "content": "[12] Meera Hahn, Jacob Krantz, Dhruv Batra, Devi Parikh, James M Rehg, Stefan Lee, and Peter Anderson. Where are you? localization from embodied dialog. arXiv preprint arXiv:2011.08277, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "type": "text", + "content": "time-scale update rule converge to a local nash equilibrium. In Annual Conference on Neural Information Processing Systems, pages 6626-6637, 2017. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 108, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 555, + 161 + ], + "type": "text", + "content": "[14] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, ICLR, 2022. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 163, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 555, + 228 + ], + "type": "text", + "content": "[15] Zilong Huang, Jun He, Junyan Ye, Lihan Jiang, Weijia Li, Yiping Chen, and Ting Han. Scene4u: Hierarchical layered 3d scene reconstruction from single panoramic image for your immerse exploration. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 26723-26733, 2025. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 230, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 555, + 295 + ], + "type": "text", + "content": "[16] Zhe Huang, Yizhe Zhao, Hao Xiao, Chenyan Wu, and Lingting Ge. Duospacenet: Leveraging both bird's-eye-view and perspective view representations for 3d object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPR Workshops, pages 2560-2570, 2025. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 297, + 555, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 297, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 297, + 555, + 352 + ], + "type": "text", + "content": "[17] Nikolai Kalischek, Michael Oechsle, Fabian Manhardt, Philipp Henzler, Konrad Schindler, and Federico Tombari. Cubediff: Repurposing diffusion-based image models for panorama generation. In International Conference on Learning Representations, ICLR, 2025. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 554, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 386 + ], + "type": "text", + "content": "[18] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, ICLR, 2015. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 388, + 555, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 555, + 453 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross B. Girshick. Segment anything. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 3992-4003, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 455, + 555, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 455, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 317, + 455, + 555, + 497 + ], + "type": "text", + "content": "[20] Jiachen Liu, Yuan Xue, Haomiao Ni, Rui Yu, Zihan Zhou, and Sharon X Huang. Computer-aided layout generation for building design: A review. arXiv preprint arXiv:2504.09694, 2025.1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 500, + 555, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 500, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 317, + 500, + 555, + 553 + ], + "type": "text", + "content": "[21] Jiachen Liu, Rui Yu, Sili Chen, Sharon X. Huang, and Hengkai Guo. Towards in-the-wild 3d plane reconstruction from a single image. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 27027-27037, 2025. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 556, + 555, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 555, + 610 + ], + "type": "text", + "content": "[22] Xiaohu Lu, Zuoyue Li, Zhaopeng Cui, Martin R. Oswald, Marc Pollefeys, and Rongjun Qin. Geometry-aware satellite-to-ground image synthesis for urban areas. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 856-864, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 613, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 555, + 656 + ], + "type": "text", + "content": "[23] Wentao Lyu, Peng Ding, Yingliang Zhang, Anpei Chen, Minye Wu, Shu Yin, and Jingyi Yu. Refocusable gigapixel panoramas for immersive VR experiences. IEEE Trans. Vis. Comput. Graph., 27(3):2028-2040, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 679 + ], + "type": "text", + "content": "[24] Nelson L. Max. Optical models for direct volume rendering. IEEE Trans. Vis. Comput. Graph., 1(2):99-108, 1995. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 555, + 713 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 317, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 317, + 756 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 317, + 756 + ], + "type": "text", + "content": "28501" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 95 + ], + "type": "text", + "content": "thesis. In European Conference on Computer Vision, ECCV, pages 405-421, 2020. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[26] Despoina Paschalidou, Amlan Kar, Maria Shugrina, Karsten Kreis, Andreas Geiger, and Sanja Fidler. Atiss: Autoregressive transformers for indoor scene synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 185 + ], + "type": "text", + "content": "[27] Giovanni Pintore, Fabio Bettio, Marco Agus, and Enrico Gobbetti. Deep scene synthesis of atlanta-world interiors from a single omnidirectional image. IEEE Trans. Vis. Comput. Graph., 29(11):4708-4718, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 230 + ], + "type": "text", + "content": "[28] Ming Qian, Jincheng Xiong, Gui-Song Xia, and Nan Xue. Sat2density: Faithful density learning from satellite-ground image pairs. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 3660-3669, 2023. 3, 4, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 296, + 285 + ], + "type": "text", + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 10674-10685, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 287, + 296, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 296, + 353 + ], + "type": "text", + "content": "[30] Jonas Schult, Sam Tsai, Lukas Hollein, Bichen Wu, Jialiang Wang, Chih-Yao Ma, Kunpeng Li, Xiaofang Wang, Felix Wimbauer, Zijian He, Peizhao Zhang, Bastian Leibe, Peter Vajda, and Ji Hou. Controlroom3d: Room generation using semantic proxy rooms. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 397 + ], + "type": "text", + "content": "[31] Yujiao Shi, Dylan Campbell, Xin Yu, and Hongdong Li. Geometry-guided street-view panorama synthesis from satellite imagery. IEEE Trans. Pattern Anal. Mach. Intell., 44(12): 10009-10022, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 443 + ], + "type": "text", + "content": "[32] Bin Tan, Rui Yu, Yujun Shen, and Nan Xue. Planarsplatting: Accurate planar surface reconstruction in 3 minutes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 1190-1199, 2025. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 498 + ], + "type": "text", + "content": "[33] Madhawa Vidanapathirana, Qirui Wu, Yasutaka Furukawa, Angel X. Chang, and Manolis Savva. Plan2scene: Converting floorplans to 3d scenes. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 10733-10742, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 296, + 544 + ], + "type": "text", + "content": "[34] Jionghao Wang, Ziyu Chen, Jun Ling, Rong Xie, and Li Song. 360-degree panorama generation from few unregistered nfov images. In Proceedings of the 31st ACM International Conference on Multimedia, MM, pages 6811-6821, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 296, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 296, + 587 + ], + "type": "text", + "content": "[35] Songsong Wu, Hao Tang, Xiao-Yuan Jing, Haifeng Zhao, Jianjun Qian, Nicu Sebe, and Yan Yan. Cross-view panorama image synthesis. IEEE Trans. Multim., 25:3546-3559, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "text", + "content": "[36] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: real-world perception for embodied agents. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "text", + "content": "[37] Ningli Xu and Rongjun Qin. Geospecific view generation geometry-context aware high-resolution ground view inference from satellite views. In European Conference on Computer Vision, ECCV, pages 349-366, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "text", + "content": "[38] Xiuyu Yang, Yunze Man, Jun-Kun Chen, and Yu-Xiong Wang. Scenecraft: Layout-guided 3d scene generation. In Advances in Neural Information Processing Systems, 2024. 1, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 396 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 139 + ], + "type": "text", + "content": "[39] Weicai Ye, Chenhao Ji, Zheng Chen, Junyao Gao, Xiaoshui Huang, Song-Hai Zhang, Wanli Ouyang, Tong He, Cairong Zhao, and Guofeng Zhang. Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion. In Annual Conference on Neural Information Processing Systems, NeurIPS, 2024. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 183 + ], + "type": "text", + "content": "[40] Rui Yu, Jiachen Liu, Zihan Zhou, and Sharon X. Huang. Nerf-enhanced outpainting for faithful field-of-view extrapolation. In IEEE International Conference on Robotics and Automation, ICRA, pages 16826-16833, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "type": "text", + "content": "[41] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to " + }, + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 316, + 185, + 555, + 239 + ], + "type": "text", + "content": " panorama image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 240, + 555, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 555, + 284 + ], + "type": "text", + "content": "[42] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision, ICCV, pages 3813-3824, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 285, + 555, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 555, + 339 + ], + "type": "text", + "content": "[43] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 341, + 555, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 341, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 341, + 555, + 396 + ], + "type": "text", + "content": "[44] Dian Zheng, Cheng Zhang, Xiao-Ming Wu, Cao Li, Chengfei Lv, Jian-Fang Hu, and Wei-Shi Zheng. Panorama generation from nfov image done right. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pages 21610-21619, 2025. 1" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 747, + 318, + 757 + ], + "type": "text", + "content": "28502" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/43f123eb-2c45-4de4-8a4a-4958a152775f_content_list.json b/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/43f123eb-2c45-4de4-8a4a-4958a152775f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..259bc88d0f742cbe47d57cc545756996de2f72f7 --- /dev/null +++ b/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/43f123eb-2c45-4de4-8a4a-4958a152775f_content_list.json @@ -0,0 +1,1522 @@ +[ + { + "type": "text", + "text": "TopicGeo: An Efficient Unified Framework for Geolocation", + "text_level": 1, + "bbox": [ + 197, + 130, + 799, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xin Wang Xinlin Wang* Shuiping Gou", + "bbox": [ + 318, + 179, + 674, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xidian University, Xi'an, China", + "bbox": [ + 372, + 199, + 625, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xinwangai@stu., wangxinlin@, shpgou@mail.}xidian.edu.cn", + "bbox": [ + 250, + 219, + 741, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/adbfe4d900636fe8e1d4f8c38d08d587188756ab66a5d1b971a461df590c7459.jpg", + "image_caption": [ + "(a) Low-texture area(980 5000×5000 base images)" + ], + "image_footnote": [], + "bbox": [ + 98, + 273, + 294, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c608631d97709b16dee6ee44c8426b5c480dd7f4ea07db3ac59ae324409703f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 273, + 500, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e65a6d5df9f7ddbd40976f928198a49181734ea6e376e95f75b088baf1d082c7.jpg", + "image_caption": [ + "Figure 1. The visual geolocation results of our TopicGeo, and typical retrieval and matching approaches: Earthloc [4], SelaVPR [29], and LoFTR [47]. \"Recall\" represents retrieval results. The color-coded lines show matching correctness: red for higher-error matches, green for higher-precision ones. In various scenes, our method accurately retrieves the target base image containing the query image from the database, while establishing pixel-level correspondence between the smaller query and large-scale retrieved image." + ], + "image_footnote": [], + "bbox": [ + 501, + 273, + 602, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eb80c29623c87844f81db9ee512708cbf4591304e737610ea841957514f5c5e4.jpg", + "image_caption": [ + "(b) Multi-temporal change area (40 5120×5120 base images)" + ], + "image_footnote": [], + "bbox": [ + 602, + 273, + 702, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a0d42e5c664ef25bed7c7422f3aa395f48da4d8c7b84cce1accbf976d46112b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 273, + 800, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c91051cdae460b609a9f69a7e434c113ee6d5fe5755a263ebcbd889a3f4d4151.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 802, + 273, + 900, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 566, + 326, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-based geolocation techniques that establish spatial correspondences between smaller query images and larger georeferenced images have gained significant attention. Existing approaches typically employ a separate \"retrieve-then-match\" paradigm, whereas such paradigms suffer from computational inefficiency or precision limitations. To this end, we propose TopicGeo, a unified framework for direct and precise query-to-reference image matching via three key innovations. The textual object semantics, called topics, distilled from CLIP prompt learning are embedded into the geolocation framework to eliminate intra-class and inter-class distribution discrepancies while also enhancing processing efficiency. Center-based adaptive label assignment and outlier rejection mechanisms as a joint retrieval-matching optimization strategy ensure task-coherent feature learning and precise spatial correspondences. A multi-level fine matching pipeline is introduced to refine matching from quality and quantity. Evaluations on large-scale synthetic", + "bbox": [ + 88, + 599, + 485, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and real-world datasets illustrate that TopicGeo achieves state-of-the-art performance in retrieval recall and matching accuracy while maintaining a balance in computational efficiency.", + "bbox": [ + 511, + 568, + 906, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 660, + 643, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image matching through establishing spatial correspondence between localized query images and extensive georeferenced databases has emerged as a critical technique for vision-based geolocation, playing a vital role in environmental monitoring [13, 53], human activities[26, 54]. However, this task faces inherent challenges. Sensor variations and temporal differences induce substantial discrepancies [31] and nonlinear geometric deformations, complicating cross-image accurate matching. Moreover, repetitive low-saliency regions, such as forests, deserts, and water bodies in georeferenced images, pose matching ambiguities.", + "bbox": [ + 509, + 686, + 906, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in satellite image retrieval have focused on mitigating these challenges through global retrieval approaches [4, 24, 48]. These methods encode entire images", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 114, + 887, + 243, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8241", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as high-dimensional vectors for rapid retrieval of similar images from a database. However, such methods exhibit certain critical limitations. Global feature aggregation dilutes spatially fine-grained discriminative patterns, causing failures in low-saliency regions, such as uniform forests or barren land. In addition, when pairing large-size reference images, images are often cropped into smaller and partially overlapped pieces [4], which causes redundant computation. Local re-ranking methods [18, 29, 43] address these challenges by re-ranking candidates based on local matches. However, false positive candidates that are merely visually similar do not contribute to performance improvements. Retrieval-based methods exhibit limited geolocation accuracy, posing significant challenges for researchers in performing time-consuming and laborious manual corrections on multi-source imagery. A natural approach would involve first extracting candidate regions via a retrieval method, then refining for pixel-level matching with a specialized model [3, 31, 45]. However, a critical limitation of this approach lies in the decoupled retrieval and matching pipelines, which inherently prioritize different features in different tasks. In addition, the matching remains heavily dependent on the retrieval model's effectiveness.", + "bbox": [ + 89, + 90, + 480, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies [45, 49] apply a single end-to-end matching model for joint retrieval and matching. While enhanced by fine-grained features to boost accuracy, these methods face a huge time overhead [45], preventing real-time geolocation. The above geolocation methods rely on visual features for location estimation. However, remote sensing images exhibit a pronounced imbalance of areal distribution of terrestrial features, inducing inherent bias in visual information-dependent models toward overrepresented categories. Moreover, geometric distortions between multisource images and cross-temporal visual discrepancies lead to variations in the distribution of geographic features at the same location. Some methods [43, 55] learn more robust representations using natural images by incorporating semantic segmentation. In remote sensing, datasets [8, 39] and segmentation models [32, 33, 50] offer semantic information but face generalization limitations. Recent studies [15, 16] leverage self-supervised learning of image topics to enhance visual features. However, the resulting topic distributions often lack explicit semantic information.", + "bbox": [ + 89, + 441, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we propose a model that jointly optimizes retrieval and matching tasks for geographic positioning. Specifically, we employ asymmetric-input architecture to minimize redundant computation in large-displacement and large-scale scenarios. The core innovation lies in hierarchical textual embeddings for prompt learning and a topic-aware encoder trained via knowledge distillation and spatial alignment, yielding efficient topic features that are beneficial for both image retrieval and coarse matching. Building on this foundation, a multi-level refinement pipeline it-", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eratively resolves positional cues across scales, improving quality and quantity of matching points. Our unified framework is jointly optimized through effective label supervision and quality-aware rejection mechanisms. The contributions of the study are as follows:", + "bbox": [ + 511, + 90, + 903, + 165 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a unified retrieval and matching framework with asymmetric large-scale input, boosted by embedded topics for efficiency and robustness to intra-class discrepancies.", + "- The center-based label assignment is designed to optimize the contrastive loss, which is further weighted through topic-based category perception to balance inter-class discrepancies.", + "- The local spatial consensus filter and the global geometric verification module are co-designed to eliminate mismatched correspondences, thereby enhancing retrieval recall and matching precision in geolocation tasks." + ], + "bbox": [ + 511, + 167, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 361, + 653, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual Language Model. The pioneering visual-language model (VLM) CLIP [38] employs vision and text encoders to project images and texts into a shared semantic space, which achieves joint embedding of images and texts, making cross-modal retrieval and zero-shot learning possible. Since then, VLMs have rapidly advanced and been applied across fields like open-vocabulary detection [6], image-text retrieval [30], and action recognition [52]. Menon et al. [34] apply large language models (LLMs) to represent object classes as semantic attribute sets for enhanced classification. Recently, Liu et al. present the first remote sensing visual-language foundation model, RemoteCLIP[25], learning semantically rich visual features aligned with text embeddings for seamless downstream applications. Our method leverages the textual semantics information of objects to capture invariant matching features across multi-source and multi-temporal satellite images.", + "bbox": [ + 511, + 387, + 903, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Retrieval. The image retrieval task focuses on identifying images similar to a query image within a database, encompassing two primary approaches: global and local feature-based retrieval. In global retrieval, the fundamental challenge lies in effectively aggregating image features from diverse regions into a cohesive high-dimensional vector. Traditional methods accomplish this by aggregating handcrafted local features [2, 5, 28] into global representations through methods such as Fisher Vectors [35], and then performing efficient retrieval through nearest neighbor search algorithms. Recent advancements leverage deep learning for feature aggregation [1, 37], employing strategies such as margin-based classification loss functions [9, 27] and clustering training samples [4]. However, global retrieval often results in a low recall rate during the initial retrieval phase. To address the issue, local matching is often integrated to re-rank retrieval results [18, 29, 43]. Geomet-", + "bbox": [ + 511, + 643, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8242", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4605b39d3a4324b481c3988c80be0dd0c08ee6670da32e6037edc0533f20b3bd.jpg", + "image_caption": [ + "Figure 2. Overview of TopicGeo. (a) Multi-level features are extracted from input images. (b) CLIP-generated prompt embeddings and visual features are injected into cross-attention layers and then distilled to infer the topic distribution. The extracted topic embeddings are integrated with global context features from self-attention layers. Preliminary matches are obtained using dual-softmax and local mutual k-nearest neighbor (MKNN) filter, followed by global geometric verification(GGV) for retrieval. (c) Adaptive matching is performed on cropped patches at a secondary resolution, with further coordinate refinement through fine-grained feature patch cropping." + ], + "image_footnote": [], + "bbox": [ + 96, + 87, + 903, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ric verification methods like RANSAC [14] are employed to identify reliable candidates, though they frequently suffer from computational inefficiency. Fast verification approaches based on matching pair counts [29] face additional challenges in remote sensing applications, where weakly textured regions often lead to significant false positives. Unlike these retrieval-specific pipelines, this paper establishes a unified framework that seamlessly integrates retrieval and matching processes and introduces an efficient geometric verification mechanism.", + "bbox": [ + 88, + 468, + 480, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Matching Image matching establishes accurate spatial correspondences between two images, essential for tasks like visual localization [21, 41] and 3D reconstruction [23]. The core of image matching is keypoint detection and description. Classical methods [2, 5, 28] employ hand-crafted features to detect and describe keypoints. Recent advances leverage deep learning for improved sparse keypoint detection[10, 40, 44], and self-supervised descriptor learning [17]. SuperGlue [42] harnesses the graph neural network with attentional mechanism to address significant occlusion or viewpoint changes, but relying on the quality of keypoints. RoMa [12] significantly enhances accuracy via Markov chain-based dense matching but incurs high computational cost. Detector-free approaches like LoFTR [47] use position embedding and cross-attention for semi-dense matching, which has gained traction in remote sensing[3, 45, 49], yet struggle with large-scale or point-of-view transformation. AdaMatcher [20] proposes an adap", + "bbox": [ + 89, + 628, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tive label assignment strategy to solve this problem, but introduces inconsistent learning-driven global mismatch filtering. TopicFM [15] sets up learnable topic embeddings for topic perception, while the unknown topics diverge from human perception and fail in complex scenes. Most existing methods rely on cross-attention for feature interaction, which limits computational efficiency. In contrast, our approach leverages distilled topics for feature enhancement and a learning-free local filter for efficient correspondence refinement.", + "bbox": [ + 511, + 468, + 906, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 511, + 667, + 705, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to retrieve a reliable base image from a non-overlapping image database using a small query image and then match the query image to the retrieved larger-scale base image. Let $R = \\{B_i | i \\in [1, m]\\}$ be a retrieval database containing $m$ high resolution base images, and $A$ be a query image. We establish a unified retrieval and matching framework, which encodes $R$ into features $F_R = \\{F_{B_i} | i \\in [1, m]\\}$ storing as a feature database, and $A$ into $F_A$ . Then $A$ is retrieved, and the corresponding target base image $B$ is obtained. Finally, the fine-grained features of both $A$ and $B$ are extracted to refine the matching and compute the homography $H$ . Figure 2 gives an overview of our pipeline.", + "bbox": [ + 511, + 703, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8243", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Topic-aware Feature Extraction", + "text_level": 1, + "bbox": [ + 89, + 90, + 372, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This paper tackles distribution discrepancies of land objects by establishing an effective topic perception process that integrates primitive visual features with high-level text representations independent of vision. Firstly, we extract visual features of the query image $A$ and the base image $B$ . To be efficient and preserve sufficient information, both of them undergo processing through patch embedding [11]. Then the downsampled images are fed into a shared ResNet [19] to extract coarse-grained visual features, $(F_A^{1/64}, F_B^{1/64})$ , and fine-grained visual features $(F_A^{1/16}, F_B^{1/16})$ and $(F_A^{1/4}, F_B^{1/4})$ .", + "bbox": [ + 89, + 112, + 483, + 284 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hierarchical topic prompts. To accelerate convergence and integrate richer information, we design dual-level prompts for land objects. Specifically, generic and stable categories are extracted from established land cover datasets [50]. These categories are enriched through object attribute prompts generated by the GPT-4 language model. Subsequently, the embeddings of the augmented object prompts are encoded using the RemoteCLIP[25] text encoder. This process generates category embeddings $\\mathbf{F}_{\\mathrm{category}} \\in \\mathbb{R}^{C \\times d}$ and attribute embeddings $\\mathbf{F}_{\\mathrm{attribute}} \\in \\mathbb{R}^{C' \\times d}$ .", + "bbox": [ + 89, + 285, + 483, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object topic distillation. The obtained attribute and category prompt embeddings are embedded into visual features in sequence via cross-attention layers, where visual features serve as queries while these prompt embeddings act as keys and values. This process synthesizes the fused textual and visual representation to generate topic embeddings $\\mathbf{F}_{\\mathrm{topic}}^{1/64}$ . Notably, the sparsity of textual semantic tokens drastically lowers the computational cost of cross-attention.", + "bbox": [ + 89, + 438, + 483, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, such embeddings fail to align category-specific and attribute-related information with their corresponding visual semantics. Leveraging the inherent alignment between visual and textual features in RemoteCLIP, we employ its frozen vision encoder to further extract the visual embeddings $\\mathbf{F}_{\\mathrm{vision}} \\in \\mathbb{R}^{N \\times d}$ of $N$ patches from $A$ . Specifically, the patch-level category perception is established by computing the category probability distribution $p \\in \\mathbb{R}^{N \\times C}$ :", + "bbox": [ + 89, + 561, + 483, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {F} _ {\\text {v i s i o n}} \\mathbf {F} _ {\\text {c a t e g o r y}} ^ {T}}{t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 709, + 483, + 751 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ is the temperature parameter.", + "bbox": [ + 89, + 762, + 339, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To transfer the category-specific information from $\\mathbf{F}_{\\mathrm{vision}}$ to topic embedding $\\mathbf{F}_{\\mathrm{topic}}^{1/64}$ , the latter are spatially aligned with $\\mathbf{F}_{\\mathrm{vision}}$ via $2 \\times 2$ pooling, resulting in resized topic embedding $\\mathbf{F}_{\\mathrm{topic}}^{\\mathrm{p}} \\in \\mathbb{R}^{N \\times d}$ . The corresponding category probability distribution for the pooled topic embedding is computed analogously to $p$ , denoted as $q \\in \\mathbb{R}^{N \\times C}$ . Thus, the weighted Kullback-Leibler (KL) divergence loss $Loss_{\\mathrm{category}}$ is employed to transfer the category-specific in-", + "bbox": [ + 89, + 777, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "formation in $\\mathbf{F}_{\\mathrm{vision}}$ to topic embeddings, expressed as:", + "bbox": [ + 511, + 90, + 877, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL o s s _ {\\text {c a t e g o r y}} = \\sum_ {i = 0} ^ {N} \\left(\\sum_ {j = 0} ^ {C} \\left(p _ {i j} \\cdot \\log \\left(\\frac {p _ {i j}}{q _ {i j}}\\right)\\right) \\cdot w _ {i}\\right) \\cdot t ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 114, + 906, + 141 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{i}$ , the normalized class frequency weight of the $i$ -th image patch, is derived from $p$ . Analogously, attribute-related information is transferred using $Loss_{attribute}$ . Then the topic embeddings are enriched by minimizing the combined loss:", + "bbox": [ + 511, + 148, + 906, + 223 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d} = \\text {L o s s} _ {\\text {c a t e g o r y}} + \\text {L o s s} _ {\\text {a t t r i b u t e .}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 232, + 906, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that distillation operates exclusively on query images due to the fully shared parameters and training efficiency.", + "bbox": [ + 511, + 258, + 906, + 289 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Topic feature integration. To capture global context absent from $\\mathbf{F}_{\\mathrm{topic}}$ , the linear self-attention[51] with rotary position embedding[46] is further employed. Thus, the outputs from parallel encoding branches are integrated via a multilayer perceptron (MLP), with features from one branch stochastically masked during optimization to mitigate MLP-induced training bias. The final topic features, denoted as $\\hat{F}_A^{1/64}$ and $\\hat{F}_B^{1/64}$ , are utilized for downstream retrieval and matching.", + "bbox": [ + 511, + 289, + 906, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Retrieval & Coarse Matching", + "text_level": 1, + "bbox": [ + 511, + 435, + 777, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The multimodal topic features of query and base images, $\\hat{F}_A^{1/64}$ and $\\hat{F}_B^{1/64}$ , enable efficient retrieval and coarse matching. Inspired by [20], the coarse-level matching probability matrix $S$ and the bidirectional matching probability matrixes $\\mathcal{P}$ are calculated as:", + "bbox": [ + 511, + 457, + 906, + 534 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} (i, j) = \\left\\langle \\hat {F} _ {1 / 6 4} ^ {A} (i), \\hat {F} _ {1 / 6 4} ^ {B} (j) \\right\\rangle , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 542, + 906, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} ^ {A B} (i, j) = \\operatorname {S o f t m a x} \\left(\\mathcal {S} (i, \\cdot)\\right) _ {j}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 569, + 906, + 589 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} ^ {B A} (i, j) = \\operatorname {S o f t m a x} \\left(\\mathcal {S} (\\cdot , j)\\right) _ {i}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 590, + 906, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\langle .,.\\rangle$ is the inner product. Based on $\\mathcal{P}^{AB}$ , an index pair $(\\tilde{i},\\tilde{j})$ is selected as the match if its corresponding confidence value exceeds the threshold $\\theta$ and is the maximum in its row. Thus, matches $\\mathcal{M}_{AB}$ is obtained as:", + "bbox": [ + 511, + 619, + 905, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {A B} = \\left\\{\\left(\\tilde {i}, \\tilde {j}\\right) \\mid \\mathcal {P} ^ {A B} (\\tilde {i}, \\tilde {j}) = \\max _ {k} \\mathcal {P} ^ {A B} (\\tilde {i}, k), \\mathcal {P} ^ {A B} \\geq \\theta \\right\\}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 691, + 906, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analogously, $\\mathcal{M}_{BA}$ is derived by applying identical thresholding and column-wise maximization operations to $\\mathcal{P}^{BA}$ , completing the bidirectional coarse matching process.", + "bbox": [ + 511, + 719, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Local MKNN filter. While LoFTR [47] employs mutual nearest neighbor (MNN) filtering to suppress local mismatches, this strategy exhibits inherent limitations in one-to-many mismatches prevalent in resolution-scaled discrepancies. Instead, we propose mutual $k$ -nearest neighbor (MKNN) filtering, which relaxes mismatches constraints by retaining matches that satisfy $k$ -neighbor spatial consistency across bidirectional correspondence sets $\\mathcal{M}_{AB}$ and $\\mathcal{M}_{BA}$ . Specifically, a candidate match pair $(i,j) \\in \\mathcal{M}_{AB}$", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8244", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is preserved if there exists a reverse-matching pair $(i', j) \\in \\mathcal{M}_{BA}$ , and the 2D coordinate pair $(x, y)$ and $(x', y')$ corresponding to points $i$ and $i'$ are $k$ -nearest neighbors. Analogously, $\\mathcal{M}_{BA}$ is handled in the same manner. Finally, the refined coarse matches are combined using the OR operation to obtain $\\mathcal{M}_c$ , and the overall process is formulated as follows:", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {c} = \\operatorname {M K N N} \\left(\\mathcal {M} _ {A B}, \\mathcal {M} _ {B A}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 203, + 482, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Global geometric verification. It is known that correct correspondences have geometric spatial consistency. Therefore, we utilize the rule to filter false positive samples in the database efficiently. Inspired by [36], we independently analyze distance and angle compatibility for each correspondence pair $\\mathcal{M}_c(n)$ among the $N_{c}$ total pairs. For distance compatibility, we calculate pairwise Euclidean distances between each point and all other $N_{c} - 1$ points in both the query and base image coordinate systems, calculating cross-coordinate distance ratios and normalizing it. For angle compatibility, we evaluate the local geometric consistency by computing relative angles among each point's $k$ -nearest neighbors in both coordinate systems, and measuring their cross-coordinate absolute angular difference $c_{n}$ . Finally, the distance compatibility $\\alpha_{n}$ and angle compatibility $\\beta_{n}$ of each matching pair $\\mathcal{M}_c(n)$ are yielded to jointly quantify spatial coherence, defined as:", + "bbox": [ + 89, + 224, + 483, + 482 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {n} = \\max \\left(0, \\min \\left(Z - \\frac {\\operatorname {V a r} \\left(d _ {n}\\right)}{\\delta_ {d} {} ^ {2}}, Z - 1\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 487, + 482, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {n} = \\max \\left(0, \\min \\left(Z - \\frac {c _ {n} {} ^ {2}}{\\delta_ {c} {} ^ {2}}, Z - 1\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 536, + 482, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta_{d}$ and $\\delta_{c}$ are the maximum acceptable values, and $Z$ is a scalar, used to truncate outliers. The spatial coherence confidences weighted by $\\lambda$ is utilized to re-weight the per-pair bidirectional matching probabilities $E_{n}$ , expressed as:", + "bbox": [ + 89, + 574, + 483, + 636 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E _ {n} = \\left(\\lambda \\beta_ {n} + (1 - \\lambda) \\alpha_ {n}\\right) \\\\ \\times \\left(P ^ {A B} \\left(\\tilde {i} _ {n}, \\tilde {j} _ {n}\\right) + P ^ {B A} \\left(\\tilde {i} _ {n}, \\tilde {j} _ {n}\\right)\\right). \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 652, + 480, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The global confidence $E$ for the base image is then obtained by aggregating all individual matching probabilities. Finally, the base image only with the highest overall confidence score is selected as the target retrieval image.", + "bbox": [ + 89, + 695, + 483, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The joint retrieval-matching loss $\\mathrm{L}_c$ is defined as:", + "bbox": [ + 109, + 756, + 437, + 771 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} _ {c} = \\left(\\operatorname {F L} \\left(\\mathcal {P} ^ {A B}, \\hat {\\mathcal {P}} ^ {A B}\\right) + \\operatorname {F L} \\left(\\mathcal {P} ^ {B A}, \\hat {\\mathcal {P}} ^ {B A}\\right)\\right) \\times w _ {\\text {t o k e n}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 776, + 482, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{P}}$ is the ground truth matching matrix. We allow one-to-many and many-to-one label assignments [20] for $\\mathcal{P}^{AB}$ and $\\mathcal{P}^{BA}$ . FL refers to the Focal Loss [22]. $w_{token}$ is derived from the copy of $w$ in the early training period, and the weight calculated by the category perception at the $\\mathbf{F}_{\\mathrm{topic}}^{1/64}$ is used in the subsequent training.", + "bbox": [ + 89, + 806, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Adaptive Center Assignment", + "text_level": 1, + "bbox": [ + 513, + 90, + 771, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Downsampling operations inherently yield sparse matching pairs, exacerbated by resolution-scaled discrepancies and asymmetric input. The adaptive label assignment [20] and MKNN filter alleviate the issue, while its reliance on the upper-left coordinates of the warped grid [42, 47] for label assignment often induces suboptimal matching patch overlaps, especially under large rotational misalignments in remote sensing images. This paper proposes a center-based label adaptive assignment strategy. Labels are assigned by distorting grid center coordinates instead of upper-left points, which are later reconverted to the upper-left coordinates for feature patch indexing, thereby enhancing robustness in challenging conditions.", + "bbox": [ + 511, + 112, + 906, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Multi-level Fine Matching", + "text_level": 1, + "bbox": [ + 513, + 316, + 751, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The retrieval-oriented low-resolution encoding fundamentally limits matching precision. To achieve enhanced correspondence density through operationally consistent processing with the coarse matching stage, we employ the same adaptive local window matching on the finer-grained feature patches $(\\hat{F}_A^{1/16}, \\hat{F}_B^{1/16})$ corresponding to $\\mathcal{M}_c$ , thereby obtaining $\\mathcal{M}_a$ . Subsequently, following LoFTR[47], higher-resolution feature patches $(\\hat{F}_A^{1/4}, \\hat{F}_B^{1/4})$ extracted from $(F_A^{1/4}, F_B^{1/4})$ guided by $\\mathcal{M}_a$ are utilized for final expectation-based fine-grained matching, obtaining $\\mathcal{M}_f$ , achieving sub-pixel matching accuracy. For adaptive local window matching, the loss is formulated as:", + "bbox": [ + 511, + 339, + 906, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\operatorname {L o s s} _ {a} = \\frac {1}{| \\mathcal {M} _ {a} |} \\left(\\operatorname {F L} \\left(\\mathcal {P} _ {a} ^ {A B}, \\hat {\\mathcal {P}} _ {a} ^ {A B}\\right) + \\operatorname {F L} \\left(\\mathcal {P} _ {a} ^ {B A}, \\hat {\\mathcal {P}} _ {a} ^ {B A}\\right)\\right), \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 539, + 903, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{P}}_a$ represents the ground truth matching matrix. The expectation-based loss $\\mathrm{Loss}_f$ refers to [47]. Thus, the final fine-grained loss is the sum of both:", + "bbox": [ + 511, + 574, + 905, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} _ {f} = \\operatorname {L o s s} _ {a} + \\operatorname {L o s s} _ {f}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 632, + 903, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training", + "text_level": 1, + "bbox": [ + 513, + 657, + 617, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The model is trained by optimizing the topic distillation loss $\\mathrm{L}_d$ , the joint retrieval-coarse matching loss $\\mathrm{L}_c$ , and fine-grained matching loss $\\mathrm{L}_f$ , defined as:", + "bbox": [ + 511, + 680, + 905, + 727 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {L} = \\mathrm {L} _ {d} + \\mathrm {L} _ {c} + \\mathrm {L} _ {f}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 737, + 903, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 777, + 645, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 801, + 833, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TZC dataset [7] consists of 980 base maps with a spatial resolution of $1\\mathrm{m}$ , each with a high resolution of $5000\\times 5000$ pixels. This dataset captures various land cover types, including dense forests, barren land, water bodies, agricultural fields, and human settlements.", + "bbox": [ + 511, + 824, + 905, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8245", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/894ecb32145b41d3f4597755e43ef975db65681b3d9093b99d3cd09e7a92f129.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of retrieval approaches. Top 1, 5, 10 recall are shown on the TZC dataset, and top 1, 2, 3 recall are shown on the MTGL40-5 dataset. Re-ranking thresholds are set at 100 for TZC and 50 for MTGL40-5." + ], + "table_footnote": [], + "table_body": "
Retrieval modeMethodTZCMTGL40-5
R@1R@5R@10Time(s)R@1R@2R@3Time(s)
Global retrievalPatch-NetVLAD63.163.163.40.39158.260.762.30.352
SelaVPR68.969.269.50.04666.766.967.40.029
EarthLoc70.471.672.40.08267.267.568.20.064
Local Re-rankingPatch-NetVLAD65.666.266.421.562.364.264.810.7
SelaVPR72.673.173.40.27470.171.572.60.133
SP-SuperGlue71.472.873.512.169.370.773.65.87
Local retrievalSelaVPR74.475.675.93.2 × 10273.373.675.511.0
LoFTR_Coarse81.683.284.48.7 × 10376.478.480.53.9 × 102
TopicFM_Coarse83.783.884.59.5 × 10381.483.184.54.3 × 102
AdaMatcher_Coarse87.289.191.59.7 × 10384.686.287.84.5 × 102
Asymmetric local retrievalOur(LoFTR)86.188.490.134.282.383.183.71.81
Our(TopicFM)88.690.292.142.486.488.390.02.17
Our90.693.293.71.07191.693.193.90.165
", + "bbox": [ + 94, + 128, + 903, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MTGL40-5 dataset [31] focuses on 20 carefully selected scenes of ports and airports. Each scene has been standardized to $5120 \\times 5120$ pixels, with a spatial resolution of $0.5\\mathrm{m}$ , and comprises five temporally distinct images. The appendix comprehensively overviews both datasets, including the generated prompt texts.", + "bbox": [ + 88, + 393, + 480, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Parameter setting. We employ the Adam optimizer for model training, set the learning rate as $5.0 \\times 10^{-4}$ and the batch size as 2. The threshold $\\theta$ is fixed at 0.2. The window size for Multi-level Fine Matching is 4. The spatial compatibility parameter $Z$ is 10. The focal loss parameters $\\alpha$ and $\\gamma$ are configured as 0.25 and 2, respectively. For the MTGL40-5 dataset, $\\gamma$ is reduced to 1 to account for temporal variations. Additionally, during the category and attribute distillation, temperature parameters are respectively set as $1/5$ and 1 to facilitate effective knowledge transfer.", + "bbox": [ + 88, + 487, + 482, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Retrieval Performance", + "text_level": 1, + "bbox": [ + 89, + 650, + 302, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison. We evaluate retrieval performance of various state-of-the-art methods on TZC and MTGL40-5 datasets, as shown in Table 1. The global retrieval approaches, such as Patch-NetVLAD[18], SelaVPR[29], and EarthLoc[4], exhibit minimal latency but suffer from significantly low recall rates on both datasets.", + "bbox": [ + 89, + 672, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Local re-ranking strategies built upon global retrieval improve performance, where SP-SuperGlue re-ranks EarthLoc candidates by matching SuperPoint [10] features through SuperGlue [42] matcher. In the case, SelaVPR has significant gains, achieving $3.7\\%$ and $3.4\\%$ R@1 gains on TZC and MTGL40-5 datasets, respectively. Compared to similar methods, SelaVPR reduces computational overhead associated with geometric verification by leveraging matching-pair counts as similarity scores. However, these", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "enhancements remain constrained by the limited positive sample coverage in global retrieval candidate sets.", + "bbox": [ + 511, + 393, + 906, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, we compare methods that rely solely on local features for retrieval. It is obvious that the local feature retrieval-based SelaVPR incurs a two-order-of-magnitude computational cost increase, primarily due to redundant symmetric processing, which expands 980 base images into approximately $10^{5}$ database images. In contrast, our asymmetric-input architecture directly processes the 980 original images, significantly improving efficiency. Coarse matcher-based local retrieval approaches like LoFTR_Coarse [47], TopicFM_Coarse [15], and AdaMatcher_Coarse [20], outperform SelaVPR through fine-grained supervision but incur three-order-of-magnitude higher online retrieval latency. Furthermore, the coarse matching encoders of LoFTR and TopicFM are integrated into our framework to illustrate the effectiveness of the asymmetric-input architecture on enhancing performance and efficiency.", + "bbox": [ + 509, + 426, + 906, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, compared to the performance on both datasets, most methods underperform on the MTGL40-5 dataset. Since multi-source modality gaps in the MTGL40-5 dataset challenge robust features extraction, even though it contains significantly fewer base images than the TZC dataset. Conversely, our paradigm outperforms other methods on retrieval performance and efficiency on both datasets, which illustrates the effect of topic features on learning consistent and invariant representations.", + "bbox": [ + 509, + 684, + 908, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Matching Performance", + "text_level": 1, + "bbox": [ + 511, + 832, + 728, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison. Table 2 shows the matching performance and time of state-of-the-art methods. To avoid retrieval preference, positive samples from SelaVPR's lo", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8246", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/91ce6e5cb2abc04e2f423eaf14bc3f47b8e5c59d903079362052e88324b1b909.jpg", + "image_caption": [ + "Figure 3. A visualized comparison of the state-of-the-art approaches on TZC dataset. Our unified framework generates numerous accurate correspondences in both texture-scarce barren areas and texture-dense urban areas. For images with large resolution-scaled discrepancies, our approach matches regions that are easier to match although they are sparse." + ], + "image_footnote": [], + "bbox": [ + 135, + 94, + 864, + 477 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/325ee7946fb417d0f99a16284f224169898c59624c2c85a6217805388700e908.jpg", + "table_caption": [ + "Table 2. Comparison of matching approaches. $n = 1$ represents the original matching, while $n = 4$ indicates four iterations of the EarthMatch protocol. IoU is used as an effective evaluation metric." + ], + "table_footnote": [], + "table_body": "
CategoryMethodTZC\nn=1 n=4MTGL\nn=1 n=4Time(s)\nn=1 n=4
Detector-basedSIFT-SuperGlue82.4 89.570.6 78.60.24 2.81
SP-SuperGlue84.4 91.277.4 82.40.14 2.37
Dense matcherRoMa92.2 96.186.7 90.20.79 5.17
Detector\nfreeLoFTR85.7 90.377.3 84.10.12 2.18
AdaMatcher91.4 93.783.3 88.40.15 2.51
TopicFM89.2 92.478.5 84.50.13 2.32
Ours98.9 99.191.1 92.70.09 2.02
", + "bbox": [ + 91, + 643, + 478, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cal re-ranking are used for matching, while the EarthMatch [3] protocol is applied to iteratively optimize matching and compensate for the limited receptive field of other methods. It is obvious that our method achieves state-of-the-art performance on both protocols while significantly improving", + "bbox": [ + 89, + 825, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "efficiency. The proposed asymmetric-input direct matching framework reduces computational time by 1/20 of the EarthMatch protocol with merely $1.6\\%$ accuracy loss on TZC dataset. Notably, while recent dense matcher RoMa [12] delivers competitive accuracy ( $90.2\\%$ on the MTGL dataset), it requires $2.5 \\times$ longer for per match than our solution. The multi-temporal discrepancies in MTGL dataset pose greater challenges than TZC's synthetic data, where our matching method effectively handles heterogeneous differences through consistent semantics and adaptive matching mechanisms.", + "bbox": [ + 511, + 578, + 906, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual comparison. Figure 3 displays the matching results of the crucial first iteration under the EarthMatch protocol. Notably, our method searches over a broader range compared to others. It is seen that our method excels in texture-scarce (such as barren areas) and texture-dense (like urban regions) scenarios through adaptive refinement and relaxed matching constraints in MKNN. Under extreme scale variations, existing methods like LoFTR distribute numerous matching pairs on non-discriminative natural terrains, which leads to low matching rates and disruption of iter-", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8247", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0e00eea25299bf0919a2d249bd1c43ae8d1a07c8bff2d0b04ae58b5fb1cd723f.jpg", + "table_caption": [ + "Table 3. Ablation of topic-related cross-attention layers and weighted loss on TZC dataset. The baseline, C, and T are the self-attention encoder, the cross-attention layers in LoFTR, and the topic-related layers of TopicFM, respectively." + ], + "table_footnote": [], + "table_body": "
BaseC/TCategoryAttributeWLossR@1Time(s)
✓82.40.87
✓86.1/88.634/42
✓85.80.94
✓✓88.11.07
✓✓✓90.61.07
", + "bbox": [ + 91, + 156, + 488, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/eb6540f2d045956f938c51a99da860992d3f10a53991338c0c70d1899f95c1c1.jpg", + "table_caption": [ + "Table 4. Ablation of topic-independent components on TZC dataset. A baseline is used that incorporates the topic encoder, original adaptive label assignment [20], and expectation-based fine-grained matcher [47]." + ], + "table_footnote": [], + "table_body": "
BaseACAMKNNGGVMFMR@1IoUPairsTime(s)
✓(0,0)84.972.31550.94/0.07
✓(0,0)86.575.41710.94/0.07
✓(0,1)86.580.7171/1920.94/0.08
✓(0,1)✓90.685.1171/1921.07/0.08
✓(1,1)✓87.181.21921.07/0.08
✓(0,1)✓✓90.690.5171/8141.07/0.09
", + "bbox": [ + 91, + 325, + 483, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ative process. Whereas our topic-aware weighted loss encourages the model to match regions that are easier to match although they are sparse, thereby enhancing reliability.", + "bbox": [ + 89, + 460, + 483, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 516, + 243, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of topics. Our method uses distilled topics to improve efficiency and performance. Therefore, we separately evaluate the effect of the category and attribute, and the derived category weighted loss on the TZC dataset, as illustrated in Table 3. The cross-image perception of LoFTR improves $R@1$ by $3.7\\%$ , while TopicFM's self-supervised topic achieves $6.2\\%$ gains. Both encoders that require crossattention for information interaction cause $40\\times$ slower online retrieval than the baseline (0.87s). Our category embeddings brings $3.4\\%$ improvement, attribute embeddings adds $2.7\\%$ , and weight loss contributes $2.5\\%$ . Moreover, the lightweight parallel architecture avoids significant time consumption. Figure 4 further visually compares topic semantics of remote sensing land cover generated by our method and the self-supervised approach TopicFM. It is seen that our method's deterministic land-cover topic modeling better distinguishes terrain features than self-supervised approach, enhancing geolocation precision.", + "bbox": [ + 88, + 537, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of other modules. Our evaluation first examines adaptive center assignment (ACA) and MKNN modules, followed by global geometric verification (GGV) and multi-level fine matching (MFM). During testing, the retrieval and matching is in sequence: failed retrieval yields zero matching IoU. When GGV is disabled, match count", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28a710a3017d7830dcd9ad3e57b454b1f18bf524d8e2bb876acc9d89b3a3f540.jpg", + "image_caption": [ + "Figure 4. Visualization of category topics in remote images. Our human-perception-aligned topic distributions encourage multisource feature alignment for geospatial retrieval and matching, whereas self-supervised structural modeling struggles to adapt to complex remote sensing scenarios." + ], + "image_footnote": [], + "bbox": [ + 526, + 90, + 885, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "directly determines retrieval scores. The two values of MKNN correspond to the $k$ value in retrieval and coarse matching. As Table 4 shows, ACA increases initial matching pairs by $10.3\\%$ over the original adaptive top-left assignment. MKNN ( $k = 1$ ) further elevates this by $12.3\\%$ . The GGV enhances retrieval accuracy by $4.1\\%$ with only an additional time consumption (0.13s). One-to-many with $k = 1$ during retrieval makes the GGV score unstable and the recall rate decreases. MFM amplifies matching pairs by $5 \\times$ with $5.4\\%$ IoU improvement, while maintaining low time overhead. Crucially, correctly retrieved samples achieve $99\\%$ (90.5(matching)/90.6(retrieval)) average matching IoU, indicating the task-coherent feature learning of retrieval and matching tasks.", + "bbox": [ + 511, + 508, + 908, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 738, + 633, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced a unified asymmetric-input retrieval-matching framework for geolocation. By combining object semantic modeling with image retrieval and matching, our method learns a robust set of primary features that can be stored offline. Moreover, we design a center-based label adaptive assignment and a multi-level fine-grained local matching pipeline, optimizing the matching precision. Compared to state-of-the-art methods, the proposed approach is both efficient and interpretable.", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8248", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 236, + 187, + 253 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Artem Babenko and Victor Lempitsky. Aggregating local deep features for image retrieval. In Proceedings of the IEEE international conference on computer vision, pages 1269-1277, 2015. 2", + "[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer vision and image understanding, 110(3):346-359, 2008. 2, 3", + "[3] Gabriele Berton, Gabriele Goletto, Gabriele Trivigno, Alex Stoken, Barbara Caputo, and Carlo Masone. Earthmatch: Iterative coregistration for fine-grained localization of astronaut photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4264-4274, 2024. 2, 3, 7", + "[4] Gabriele Berton, Alex Stoken, Barbara Caputo, and Carlo Masone. Earthloc: Astronaut photography localization by indexing earth from space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12754-12764, 2024. 1, 2, 6", + "[5] Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 778-792. Springer, 2010. 2, 3", + "[6] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 2", + "[7] \"Tianzhi Cup\" Artificial Intelligence Challenge Organizing Committee. zhihuidiqiu2024, 2024. 5", + "[8] Rodrigo Caye Daudt, Bertrand Le Saux, Alexandre Boulch, and Yann Gousseau. Multitask learning for large-scale semantic change detection. Computer Vision and Image Understanding, 187:102783, 2019. 2", + "[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2", + "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 3, 6" + ], + "bbox": [ + 93, + 261, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported by the National Natural Science Foundation of China under Grant No.62372358 and 62302355, the Fundamental Research Funds for the Central Universities under Grant No.XJSJ24071 and No.XJSJ24072, and the Xidian University Specially Funded Project for Interdisciplinary Exploration under Grant No. TZJH2024026.", + "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4", + "[12] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 3, 7", + "[13] JR Elliott. Earth observation for the assessment of earthquake hazard, risk and disaster management. Surveys in geophysics, 41(6):1323-1354, 2020. 1", + "[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3", + "[15] Khang Truong Giang, Soohwan Song, and Sungho Jo. Top-icfm: Robust and interpretable topic-assisted feature matching. In Proceedings of the AAAI conference on artificial intelligence, pages 2447-2455, 2023. 2, 3, 6", + "[16] Khang Truong Giang, Soohwan Song, and Sungho Jo. Top-icfm+: Boosting accuracy and efficiency of topic-assisted feature matching. IEEE Transactions on Image Processing, 2024. 2", + "[17] Pierre Gleize, Weiyao Wang, and Matt Feiszli. Silk: Simple learned keypoints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22499-22508, 2023. 3", + "[18] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14141-14152, 2021. 2, 6", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[20] Dihe Huang, Ying Chen, Yong Liu, Jianlin Liu, Shang Xu, Wenlong Wu, Yikang Ding, Fan Tang, and Chengjie Wang. Adaptive assignment for geometry aware local feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5425-5434, 2023. 3, 4, 5, 6, 8", + "[21] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 3", + "[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5", + "[23] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 3", + "[24] Chao Liu, Jingjing Ma, Xu Tang, Fang Liu, Xiangrong Zhang, and Licheng Jiao. Deep hash learning for remote" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8249", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sensing image retrieval. IEEE Transactions on Geoscience and Remote Sensing, 59(4):3420-3443, 2020. 1", + "[25] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 4", + "[26] Ganchao Liu, Chao Li, Sihang Zhang, and Yuan Yuan. Vlmlf: Uav visual localization based on multi-source image feature learning. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1", + "[27] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 2", + "[28] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 2, 3", + "[29] Feng Lu, Lijun Zhang, Xiangyuan Lan, Shuting Dong, Yaowei Wang, and Chun Yuan. Towards seamless adaptation of pre-trained models for visual place recognition. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6", + "[30] Christian Lülf, Denis Mayr Lima Martins, Marcos Antonio Vaz Salles, Yongluan Zhou, and Fabian Gieseke. Clipbranches: Interactive fine-tuning for text-image retrieval. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2719-2723, 2024. 2", + "[31] Jingjing Ma, Shiji Pei, Yuqun Yang, Xu Tang, and Xiangrong Zhang. Mtgl40-5: A multi-temporal dataset for remote sensing image geo-localization. Remote Sensing, 15(17):4229, 2023. 1, 2, 6", + "[32] Xianping Ma, Xiaokang Zhang, and Man-On Pun. Rs 3 mamba: Visual state space model for remote sensing image semantic segmentation. IEEE Geoscience and Remote Sensing Letters, 2024. 2", + "[33] Xianping Ma, Xiaokang Zhang, Man-On Pun, and Ming Liu. A multilevel multimodal fusion transformer for remote sensing semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2", + "[34] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 2", + "[35] Florent Perronnin, Yan Liu, Jorge Sánchez, and Hervé Poirier. Large-scale image retrieval with compressed fisher vectors. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3384-3391. IEEE, 2010. 2", + "[36] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143-11152, 2022. 5", + "[37] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning cnn image retrieval with no human annotation. IEEE" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "transactions on pattern analysis and machine intelligence, 41(7):1655-1668, 2018. 2", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[39] Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and Robin Robertson Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. IEEE Access, 9:89644-89654, 2021. 2", + "[40] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in neural information processing systems, 32, 2019. 3", + "[41] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12716–12725, 2019. 3", + "[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3, 5, 6", + "[43] Yanqing Shen, Sanping Zhou, Jingwen Fu, Ruotong Wang, Shitao Chen, and Nanning Zheng. Structvpr: Distill structural knowledge with weighting samples for visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11217-11226, 2023. 2", + "[44] Edgar Simo-Serra, Eduard Trulls, Luis Ferraz, Iasonas Kokkinos, Pascal Fua, and Francesc Moreno-Noguer. Discriminative learning of deep convolutional feature point descriptors. In Proceedings of the IEEE international conference on computer vision, pages 118-126, 2015. 3", + "[45] Alex Stoken and Kenton Fisher. Find my astronaut photo: Automated localization and georectification of astronaut photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6196-6205, 2023. 2, 3", + "[46] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4", + "[47] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8918-8927, 2021. 1, 3, 4, 5, 6, 8", + "[48] Xu Tang, Qiushuo Ma, Xiangrong Zhang, Fang Liu, Jingjing Ma, and Licheng Jiao. Attention consistent network for remote sensing scene classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:2030-2045, 2021. 1" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8250", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Haoyang Wang, Fuhui Zhou, and Qihui Wu. Accurate vision-enabled uav location using feature-enhanced transformer-driven image matching. IEEE Transactions on Instrumentation and Measurement, 2024. 2, 3", + "[50] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733, 2021. 2, 4", + "[51] Shuhang Wang, Jin Zheng, Hai-Miao Hu, and Bo Li. Naturalness preserved enhancement algorithm for non-uniform illumination images. IEEE transactions on image processing, 22(9):3538-3548, 2013. 4", + "[52] Xiang Wang, Shiwei Zhang, Jun Cen, Changxin Gao, Yingya Zhang, Deli Zhao, and Nong Sang. Clip-guided prototype modulating for few-shot action recognition. International Journal of Computer Vision, 132(6):1899-1912, 2024. 2", + "[53] Qinghua Ye, Yuzhe Wang, Lin Liu, Linan Guo, Xueqin Zhang, Liyun Dai, Limin Zhai, Yafan Hu, Nauman Ali, Xinhui Ji, et al. Remote sensing and modeling of the cryosphere in high mountain asia: A multidisciplinary review. Remote Sensing, 16(10):1709, 2024. 1", + "[54] Hongsheng Zhang and Ru Xu. Exploring the optimal integration levels between sar and optical data for better urban land cover mapping in the pearl river delta. International journal of applied earth observation and geoinformation, 64:87-95, 2018. 1", + "[55] Yesheng Zhang and Xu Zhao. Mesa: Matching everything by segmenting anything. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20217-20226, 2024. 2" + ], + "bbox": [ + 89, + 92, + 482, + 513 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8251", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/43f123eb-2c45-4de4-8a4a-4958a152775f_model.json b/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/43f123eb-2c45-4de4-8a4a-4958a152775f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8af19105990511eea629b123b305f84acd5bdd98 --- /dev/null +++ b/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/43f123eb-2c45-4de4-8a4a-4958a152775f_model.json @@ -0,0 +1,2169 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.199, + 0.131, + 0.8, + 0.154 + ], + "angle": 0, + "content": "TopicGeo: An Efficient Unified Framework for Geolocation" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.18, + 0.676, + 0.2 + ], + "angle": 0, + "content": "Xin Wang Xinlin Wang* Shuiping Gou" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.2, + 0.626, + 0.218 + ], + "angle": 0, + "content": "Xidian University, Xi'an, China" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.22, + 0.743, + 0.235 + ], + "angle": 0, + "content": "{xinwangai@stu., wangxinlin@, shpgou@mail.}xidian.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.274, + 0.295, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.47, + 0.442, + 0.484 + ], + "angle": 0, + "content": "(a) Low-texture area(980 5000×5000 base images)" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.274, + 0.501, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.275, + 0.603, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.275, + 0.703, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.275, + 0.801, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.275, + 0.901, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.471, + 0.886, + 0.485 + ], + "angle": 0, + "content": "(b) Multi-temporal change area (40 5120×5120 base images)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.498, + 0.907, + 0.555 + ], + "angle": 0, + "content": "Figure 1. The visual geolocation results of our TopicGeo, and typical retrieval and matching approaches: Earthloc [4], SelaVPR [29], and LoFTR [47]. \"Recall\" represents retrieval results. The color-coded lines show matching correctness: red for higher-error matches, green for higher-precision ones. In various scenes, our method accurately retrieves the target base image containing the query image from the database, while establishing pixel-level correspondence between the smaller query and large-scale retrieved image." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.567, + 0.327, + 0.583 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.6, + 0.486, + 0.873 + ], + "angle": 0, + "content": "Vision-based geolocation techniques that establish spatial correspondences between smaller query images and larger georeferenced images have gained significant attention. Existing approaches typically employ a separate \"retrieve-then-match\" paradigm, whereas such paradigms suffer from computational inefficiency or precision limitations. To this end, we propose TopicGeo, a unified framework for direct and precise query-to-reference image matching via three key innovations. The textual object semantics, called topics, distilled from CLIP prompt learning are embedded into the geolocation framework to eliminate intra-class and inter-class distribution discrepancies while also enhancing processing efficiency. Center-based adaptive label assignment and outlier rejection mechanisms as a joint retrieval-matching optimization strategy ensure task-coherent feature learning and precise spatial correspondences. A multi-level fine matching pipeline is introduced to refine matching from quality and quantity. Evaluations on large-scale synthetic" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.569, + 0.907, + 0.63 + ], + "angle": 0, + "content": "and real-world datasets illustrate that TopicGeo achieves state-of-the-art performance in retrieval recall and matching accuracy while maintaining a balance in computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.661, + 0.645, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.688, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Image matching through establishing spatial correspondence between localized query images and extensive georeferenced databases has emerged as a critical technique for vision-based geolocation, playing a vital role in environmental monitoring [13, 53], human activities[26, 54]. However, this task faces inherent challenges. Sensor variations and temporal differences induce substantial discrepancies [31] and nonlinear geometric deformations, complicating cross-image accurate matching. Moreover, repetitive low-saliency regions, such as forests, deserts, and water bodies in georeferenced images, pose matching ambiguities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Recent advances in satellite image retrieval have focused on mitigating these challenges through global retrieval approaches [4, 24, 48]. These methods encode entire images" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.245, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8241" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.438 + ], + "angle": 0, + "content": "as high-dimensional vectors for rapid retrieval of similar images from a database. However, such methods exhibit certain critical limitations. Global feature aggregation dilutes spatially fine-grained discriminative patterns, causing failures in low-saliency regions, such as uniform forests or barren land. In addition, when pairing large-size reference images, images are often cropped into smaller and partially overlapped pieces [4], which causes redundant computation. Local re-ranking methods [18, 29, 43] address these challenges by re-ranking candidates based on local matches. However, false positive candidates that are merely visually similar do not contribute to performance improvements. Retrieval-based methods exhibit limited geolocation accuracy, posing significant challenges for researchers in performing time-consuming and laborious manual corrections on multi-source imagery. A natural approach would involve first extracting candidate regions via a retrieval method, then refining for pixel-level matching with a specialized model [3, 31, 45]. However, a critical limitation of this approach lies in the decoupled retrieval and matching pipelines, which inherently prioritize different features in different tasks. In addition, the matching remains heavily dependent on the retrieval model's effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.483, + 0.745 + ], + "angle": 0, + "content": "Recent studies [45, 49] apply a single end-to-end matching model for joint retrieval and matching. While enhanced by fine-grained features to boost accuracy, these methods face a huge time overhead [45], preventing real-time geolocation. The above geolocation methods rely on visual features for location estimation. However, remote sensing images exhibit a pronounced imbalance of areal distribution of terrestrial features, inducing inherent bias in visual information-dependent models toward overrepresented categories. Moreover, geometric distortions between multisource images and cross-temporal visual discrepancies lead to variations in the distribution of geographic features at the same location. Some methods [43, 55] learn more robust representations using natural images by incorporating semantic segmentation. In remote sensing, datasets [8, 39] and segmentation models [32, 33, 50] offer semantic information but face generalization limitations. Recent studies [15, 16] leverage self-supervised learning of image topics to enhance visual features. However, the resulting topic distributions often lack explicit semantic information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this study, we propose a model that jointly optimizes retrieval and matching tasks for geographic positioning. Specifically, we employ asymmetric-input architecture to minimize redundant computation in large-displacement and large-scale scenarios. The core innovation lies in hierarchical textual embeddings for prompt learning and a topic-aware encoder trained via knowledge distillation and spatial alignment, yielding efficient topic features that are beneficial for both image retrieval and coarse matching. Building on this foundation, a multi-level refinement pipeline it-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.166 + ], + "angle": 0, + "content": "eratively resolves positional cues across scales, improving quality and quantity of matching points. Our unified framework is jointly optimized through effective label supervision and quality-aware rejection mechanisms. The contributions of the study are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.168, + 0.905, + 0.227 + ], + "angle": 0, + "content": "- We propose a unified retrieval and matching framework with asymmetric large-scale input, boosted by embedded topics for efficiency and robustness to intra-class discrepancies." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.287 + ], + "angle": 0, + "content": "- The center-based label assignment is designed to optimize the contrastive loss, which is further weighted through topic-based category perception to balance inter-class discrepancies." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.905, + 0.348 + ], + "angle": 0, + "content": "- The local spatial consensus filter and the global geometric verification module are co-designed to eliminate mismatched correspondences, thereby enhancing retrieval recall and matching precision in geolocation tasks." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.168, + 0.905, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.655, + 0.377 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.388, + 0.905, + 0.644 + ], + "angle": 0, + "content": "Visual Language Model. The pioneering visual-language model (VLM) CLIP [38] employs vision and text encoders to project images and texts into a shared semantic space, which achieves joint embedding of images and texts, making cross-modal retrieval and zero-shot learning possible. Since then, VLMs have rapidly advanced and been applied across fields like open-vocabulary detection [6], image-text retrieval [30], and action recognition [52]. Menon et al. [34] apply large language models (LLMs) to represent object classes as semantic attribute sets for enhanced classification. Recently, Liu et al. present the first remote sensing visual-language foundation model, RemoteCLIP[25], learning semantically rich visual features aligned with text embeddings for seamless downstream applications. Our method leverages the textual semantics information of objects to capture invariant matching features across multi-source and multi-temporal satellite images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.645, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Image Retrieval. The image retrieval task focuses on identifying images similar to a query image within a database, encompassing two primary approaches: global and local feature-based retrieval. In global retrieval, the fundamental challenge lies in effectively aggregating image features from diverse regions into a cohesive high-dimensional vector. Traditional methods accomplish this by aggregating handcrafted local features [2, 5, 28] into global representations through methods such as Fisher Vectors [35], and then performing efficient retrieval through nearest neighbor search algorithms. Recent advancements leverage deep learning for feature aggregation [1, 37], employing strategies such as margin-based classification loss functions [9, 27] and clustering training samples [4]. However, global retrieval often results in a low recall rate during the initial retrieval phase. To address the issue, local matching is often integrated to re-rank retrieval results [18, 29, 43]. Geomet-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8242" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.088, + 0.904, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.373, + 0.908, + 0.445 + ], + "angle": 0, + "content": "Figure 2. Overview of TopicGeo. (a) Multi-level features are extracted from input images. (b) CLIP-generated prompt embeddings and visual features are injected into cross-attention layers and then distilled to infer the topic distribution. The extracted topic embeddings are integrated with global context features from self-attention layers. Preliminary matches are obtained using dual-softmax and local mutual k-nearest neighbor (MKNN) filter, followed by global geometric verification(GGV) for retrieval. (c) Adaptive matching is performed on cropped patches at a secondary resolution, with further coordinate refinement through fine-grained feature patch cropping." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.469, + 0.482, + 0.62 + ], + "angle": 0, + "content": "ric verification methods like RANSAC [14] are employed to identify reliable candidates, though they frequently suffer from computational inefficiency. Fast verification approaches based on matching pair counts [29] face additional challenges in remote sensing applications, where weakly textured regions often lead to significant false positives. Unlike these retrieval-specific pipelines, this paper establishes a unified framework that seamlessly integrates retrieval and matching processes and introduces an efficient geometric verification mechanism." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Image Matching Image matching establishes accurate spatial correspondences between two images, essential for tasks like visual localization [21, 41] and 3D reconstruction [23]. The core of image matching is keypoint detection and description. Classical methods [2, 5, 28] employ hand-crafted features to detect and describe keypoints. Recent advances leverage deep learning for improved sparse keypoint detection[10, 40, 44], and self-supervised descriptor learning [17]. SuperGlue [42] harnesses the graph neural network with attentional mechanism to address significant occlusion or viewpoint changes, but relying on the quality of keypoints. RoMa [12] significantly enhances accuracy via Markov chain-based dense matching but incurs high computational cost. Detector-free approaches like LoFTR [47] use position embedding and cross-attention for semi-dense matching, which has gained traction in remote sensing[3, 45, 49], yet struggle with large-scale or point-of-view transformation. AdaMatcher [20] proposes an adap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.621 + ], + "angle": 0, + "content": "tive label assignment strategy to solve this problem, but introduces inconsistent learning-driven global mismatch filtering. TopicFM [15] sets up learnable topic embeddings for topic perception, while the unknown topics diverge from human perception and fail in complex scenes. Most existing methods rely on cross-attention for feature interaction, which limits computational efficiency. In contrast, our approach leverages distilled topics for feature enhancement and a learning-free local filter for efficient correspondence refinement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.707, + 0.687 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Our goal is to retrieve a reliable base image from a non-overlapping image database using a small query image and then match the query image to the retrieved larger-scale base image. Let \\( R = \\{B_i | i \\in [1, m]\\} \\) be a retrieval database containing \\( m \\) high resolution base images, and \\( A \\) be a query image. We establish a unified retrieval and matching framework, which encodes \\( R \\) into features \\( F_R = \\{F_{B_i} | i \\in [1, m]\\} \\) storing as a feature database, and \\( A \\) into \\( F_A \\). Then \\( A \\) is retrieved, and the corresponding target base image \\( B \\) is obtained. Finally, the fine-grained features of both \\( A \\) and \\( B \\) are extracted to refine the matching and compute the homography \\( H \\). Figure 2 gives an overview of our pipeline." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8243" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.373, + 0.108 + ], + "angle": 0, + "content": "3.1. Topic-aware Feature Extraction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.285 + ], + "angle": 0, + "content": "This paper tackles distribution discrepancies of land objects by establishing an effective topic perception process that integrates primitive visual features with high-level text representations independent of vision. Firstly, we extract visual features of the query image \\( A \\) and the base image \\( B \\). To be efficient and preserve sufficient information, both of them undergo processing through patch embedding [11]. Then the downsampled images are fed into a shared ResNet [19] to extract coarse-grained visual features, \\( (F_A^{1/64}, F_B^{1/64}) \\), and fine-grained visual features \\( (F_A^{1/16}, F_B^{1/16}) \\) and \\( (F_A^{1/4}, F_B^{1/4}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.484, + 0.438 + ], + "angle": 0, + "content": "Hierarchical topic prompts. To accelerate convergence and integrate richer information, we design dual-level prompts for land objects. Specifically, generic and stable categories are extracted from established land cover datasets [50]. These categories are enriched through object attribute prompts generated by the GPT-4 language model. Subsequently, the embeddings of the augmented object prompts are encoded using the RemoteCLIP[25] text encoder. This process generates category embeddings \\(\\mathbf{F}_{\\mathrm{category}} \\in \\mathbb{R}^{C \\times d}\\) and attribute embeddings \\(\\mathbf{F}_{\\mathrm{attribute}} \\in \\mathbb{R}^{C' \\times d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.484, + 0.562 + ], + "angle": 0, + "content": "Object topic distillation. The obtained attribute and category prompt embeddings are embedded into visual features in sequence via cross-attention layers, where visual features serve as queries while these prompt embeddings act as keys and values. This process synthesizes the fused textual and visual representation to generate topic embeddings \\(\\mathbf{F}_{\\mathrm{topic}}^{1/64}\\). Notably, the sparsity of textual semantic tokens drastically lowers the computational cost of cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.699 + ], + "angle": 0, + "content": "However, such embeddings fail to align category-specific and attribute-related information with their corresponding visual semantics. Leveraging the inherent alignment between visual and textual features in RemoteCLIP, we employ its frozen vision encoder to further extract the visual embeddings \\(\\mathbf{F}_{\\mathrm{vision}} \\in \\mathbb{R}^{N \\times d}\\) of \\(N\\) patches from \\(A\\). Specifically, the patch-level category perception is established by computing the category probability distribution \\(p \\in \\mathbb{R}^{N \\times C}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.71, + 0.484, + 0.752 + ], + "angle": 0, + "content": "\\[\np = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {F} _ {\\text {v i s i o n}} \\mathbf {F} _ {\\text {c a t e g o r y}} ^ {T}}{t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.763, + 0.341, + 0.777 + ], + "angle": 0, + "content": "where \\(t\\) is the temperature parameter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To transfer the category-specific information from \\(\\mathbf{F}_{\\mathrm{vision}}\\) to topic embedding \\(\\mathbf{F}_{\\mathrm{topic}}^{1/64}\\), the latter are spatially aligned with \\(\\mathbf{F}_{\\mathrm{vision}}\\) via \\(2 \\times 2\\) pooling, resulting in resized topic embedding \\(\\mathbf{F}_{\\mathrm{topic}}^{\\mathrm{p}} \\in \\mathbb{R}^{N \\times d}\\). The corresponding category probability distribution for the pooled topic embedding is computed analogously to \\(p\\), denoted as \\(q \\in \\mathbb{R}^{N \\times C}\\). Thus, the weighted Kullback-Leibler (KL) divergence loss \\(Loss_{\\mathrm{category}}\\) is employed to transfer the category-specific in-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.878, + 0.108 + ], + "angle": 0, + "content": "formation in \\(\\mathbf{F}_{\\mathrm{vision}}\\) to topic embeddings, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.116, + 0.907, + 0.142 + ], + "angle": 0, + "content": "\\[\nL o s s _ {\\text {c a t e g o r y}} = \\sum_ {i = 0} ^ {N} \\left(\\sum_ {j = 0} ^ {C} \\left(p _ {i j} \\cdot \\log \\left(\\frac {p _ {i j}}{q _ {i j}}\\right)\\right) \\cdot w _ {i}\\right) \\cdot t ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.149, + 0.907, + 0.224 + ], + "angle": 0, + "content": "where \\(w_{i}\\), the normalized class frequency weight of the \\(i\\)-th image patch, is derived from \\(p\\). Analogously, attribute-related information is transferred using \\(Loss_{attribute}\\). Then the topic embeddings are enriched by minimizing the combined loss:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.233, + 0.907, + 0.25 + ], + "angle": 0, + "content": "\\[\nL _ {d} = \\text {L o s s} _ {\\text {c a t e g o r y}} + \\text {L o s s} _ {\\text {a t t r i b u t e .}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.907, + 0.29 + ], + "angle": 0, + "content": "Note that distillation operates exclusively on query images due to the fully shared parameters and training efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.908, + 0.429 + ], + "angle": 0, + "content": "Topic feature integration. To capture global context absent from \\(\\mathbf{F}_{\\mathrm{topic}}\\), the linear self-attention[51] with rotary position embedding[46] is further employed. Thus, the outputs from parallel encoding branches are integrated via a multilayer perceptron (MLP), with features from one branch stochastically masked during optimization to mitigate MLP-induced training bias. The final topic features, denoted as \\(\\hat{F}_A^{1/64}\\) and \\(\\hat{F}_B^{1/64}\\), are utilized for downstream retrieval and matching." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.436, + 0.779, + 0.453 + ], + "angle": 0, + "content": "3.2. Retrieval & Coarse Matching" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.458, + 0.907, + 0.535 + ], + "angle": 0, + "content": "The multimodal topic features of query and base images, \\(\\hat{F}_A^{1/64}\\) and \\(\\hat{F}_B^{1/64}\\), enable efficient retrieval and coarse matching. Inspired by [20], the coarse-level matching probability matrix \\(S\\) and the bidirectional matching probability matrixes \\(\\mathcal{P}\\) are calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.543, + 0.907, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} (i, j) = \\left\\langle \\hat {F} _ {1 / 6 4} ^ {A} (i), \\hat {F} _ {1 / 6 4} ^ {B} (j) \\right\\rangle , \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.57, + 0.907, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} ^ {A B} (i, j) = \\operatorname {S o f t m a x} \\left(\\mathcal {S} (i, \\cdot)\\right) _ {j}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.591, + 0.907, + 0.611 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} ^ {B A} (i, j) = \\operatorname {S o f t m a x} \\left(\\mathcal {S} (\\cdot , j)\\right) _ {i}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.906, + 0.682 + ], + "angle": 0, + "content": "where \\(\\langle .,.\\rangle\\) is the inner product. Based on \\(\\mathcal{P}^{AB}\\), an index pair \\((\\tilde{i},\\tilde{j})\\) is selected as the match if its corresponding confidence value exceeds the threshold \\(\\theta\\) and is the maximum in its row. Thus, matches \\(\\mathcal{M}_{AB}\\) is obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.693, + 0.907, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {A B} = \\left\\{\\left(\\tilde {i}, \\tilde {j}\\right) \\mid \\mathcal {P} ^ {A B} (\\tilde {i}, \\tilde {j}) = \\max _ {k} \\mathcal {P} ^ {A B} (\\tilde {i}, k), \\mathcal {P} ^ {A B} \\geq \\theta \\right\\}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.765 + ], + "angle": 0, + "content": "Analogously, \\(\\mathcal{M}_{BA}\\) is derived by applying identical thresholding and column-wise maximization operations to \\(\\mathcal{P}^{BA}\\), completing the bidirectional coarse matching process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Local MKNN filter. While LoFTR [47] employs mutual nearest neighbor (MNN) filtering to suppress local mismatches, this strategy exhibits inherent limitations in one-to-many mismatches prevalent in resolution-scaled discrepancies. Instead, we propose mutual \\( k \\)-nearest neighbor (MKNN) filtering, which relaxes mismatches constraints by retaining matches that satisfy \\( k \\)-neighbor spatial consistency across bidirectional correspondence sets \\( \\mathcal{M}_{AB} \\) and \\( \\mathcal{M}_{BA} \\). Specifically, a candidate match pair \\( (i,j) \\in \\mathcal{M}_{AB} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8244" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.198 + ], + "angle": 0, + "content": "is preserved if there exists a reverse-matching pair \\((i', j) \\in \\mathcal{M}_{BA}\\), and the 2D coordinate pair \\((x, y)\\) and \\((x', y')\\) corresponding to points \\(i\\) and \\(i'\\) are \\(k\\)-nearest neighbors. Analogously, \\(\\mathcal{M}_{BA}\\) is handled in the same manner. Finally, the refined coarse matches are combined using the OR operation to obtain \\(\\mathcal{M}_c\\), and the overall process is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.204, + 0.483, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {c} = \\operatorname {M K N N} \\left(\\mathcal {M} _ {A B}, \\mathcal {M} _ {B A}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.226, + 0.484, + 0.483 + ], + "angle": 0, + "content": "Global geometric verification. It is known that correct correspondences have geometric spatial consistency. Therefore, we utilize the rule to filter false positive samples in the database efficiently. Inspired by [36], we independently analyze distance and angle compatibility for each correspondence pair \\(\\mathcal{M}_c(n)\\) among the \\(N_{c}\\) total pairs. For distance compatibility, we calculate pairwise Euclidean distances between each point and all other \\(N_{c} - 1\\) points in both the query and base image coordinate systems, calculating cross-coordinate distance ratios and normalizing it. For angle compatibility, we evaluate the local geometric consistency by computing relative angles among each point's \\(k\\)-nearest neighbors in both coordinate systems, and measuring their cross-coordinate absolute angular difference \\(c_{n}\\). Finally, the distance compatibility \\(\\alpha_{n}\\) and angle compatibility \\(\\beta_{n}\\) of each matching pair \\(\\mathcal{M}_c(n)\\) are yielded to jointly quantify spatial coherence, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.488, + 0.483, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {n} = \\max \\left(0, \\min \\left(Z - \\frac {\\operatorname {V a r} \\left(d _ {n}\\right)}{\\delta_ {d} {} ^ {2}}, Z - 1\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.537, + 0.483, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\beta_ {n} = \\max \\left(0, \\min \\left(Z - \\frac {c _ {n} {} ^ {2}}{\\delta_ {c} {} ^ {2}}, Z - 1\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.637 + ], + "angle": 0, + "content": "where \\(\\delta_{d}\\) and \\(\\delta_{c}\\) are the maximum acceptable values, and \\(Z\\) is a scalar, used to truncate outliers. The spatial coherence confidences weighted by \\(\\lambda\\) is utilized to re-weight the per-pair bidirectional matching probabilities \\(E_{n}\\), expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.654, + 0.481, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E _ {n} = \\left(\\lambda \\beta_ {n} + (1 - \\lambda) \\alpha_ {n}\\right) \\\\ \\times \\left(P ^ {A B} \\left(\\tilde {i} _ {n}, \\tilde {j} _ {n}\\right) + P ^ {B A} \\left(\\tilde {i} _ {n}, \\tilde {j} _ {n}\\right)\\right). \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.484, + 0.757 + ], + "angle": 0, + "content": "The global confidence \\( E \\) for the base image is then obtained by aggregating all individual matching probabilities. Finally, the base image only with the highest overall confidence score is selected as the target retrieval image." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.757, + 0.438, + 0.772 + ], + "angle": 0, + "content": "The joint retrieval-matching loss \\( \\mathrm{L}_c \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.777, + 0.483, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} _ {c} = \\left(\\operatorname {F L} \\left(\\mathcal {P} ^ {A B}, \\hat {\\mathcal {P}} ^ {A B}\\right) + \\operatorname {F L} \\left(\\mathcal {P} ^ {B A}, \\hat {\\mathcal {P}} ^ {B A}\\right)\\right) \\times w _ {\\text {t o k e n}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.808, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{P}}\\) is the ground truth matching matrix. We allow one-to-many and many-to-one label assignments [20] for \\(\\mathcal{P}^{AB}\\) and \\(\\mathcal{P}^{BA}\\). FL refers to the Focal Loss [22]. \\(w_{token}\\) is derived from the copy of \\(w\\) in the early training period, and the weight calculated by the category perception at the \\(\\mathbf{F}_{\\mathrm{topic}}^{1/64}\\) is used in the subsequent training." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.772, + 0.108 + ], + "angle": 0, + "content": "3.3. Adaptive Center Assignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.31 + ], + "angle": 0, + "content": "Downsampling operations inherently yield sparse matching pairs, exacerbated by resolution-scaled discrepancies and asymmetric input. The adaptive label assignment [20] and MKNN filter alleviate the issue, while its reliance on the upper-left coordinates of the warped grid [42, 47] for label assignment often induces suboptimal matching patch overlaps, especially under large rotational misalignments in remote sensing images. This paper proposes a center-based label adaptive assignment strategy. Labels are assigned by distorting grid center coordinates instead of upper-left points, which are later reconverted to the upper-left coordinates for feature patch indexing, thereby enhancing robustness in challenging conditions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.318, + 0.753, + 0.336 + ], + "angle": 0, + "content": "3.4. Multi-level Fine Matching" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.907, + 0.53 + ], + "angle": 0, + "content": "The retrieval-oriented low-resolution encoding fundamentally limits matching precision. To achieve enhanced correspondence density through operationally consistent processing with the coarse matching stage, we employ the same adaptive local window matching on the finer-grained feature patches \\((\\hat{F}_A^{1/16}, \\hat{F}_B^{1/16})\\) corresponding to \\(\\mathcal{M}_c\\), thereby obtaining \\(\\mathcal{M}_a\\). Subsequently, following LoFTR[47], higher-resolution feature patches \\((\\hat{F}_A^{1/4}, \\hat{F}_B^{1/4})\\) extracted from \\((F_A^{1/4}, F_B^{1/4})\\) guided by \\(\\mathcal{M}_a\\) are utilized for final expectation-based fine-grained matching, obtaining \\(\\mathcal{M}_f\\), achieving sub-pixel matching accuracy. For adaptive local window matching, the loss is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.54, + 0.905, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\left. \\operatorname {L o s s} _ {a} = \\frac {1}{| \\mathcal {M} _ {a} |} \\left(\\operatorname {F L} \\left(\\mathcal {P} _ {a} ^ {A B}, \\hat {\\mathcal {P}} _ {a} ^ {A B}\\right) + \\operatorname {F L} \\left(\\mathcal {P} _ {a} ^ {B A}, \\hat {\\mathcal {P}} _ {a} ^ {B A}\\right)\\right), \\right. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.575, + 0.906, + 0.621 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{P}}_a\\) represents the ground truth matching matrix. The expectation-based loss \\(\\mathrm{Loss}_f\\) refers to [47]. Thus, the final fine-grained loss is the sum of both:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.633, + 0.905, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} _ {f} = \\operatorname {L o s s} _ {a} + \\operatorname {L o s s} _ {f}. \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.659, + 0.619, + 0.675 + ], + "angle": 0, + "content": "3.5. Training" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.681, + 0.906, + 0.728 + ], + "angle": 0, + "content": "The model is trained by optimizing the topic distillation loss \\( \\mathrm{L}_d \\), the joint retrieval-coarse matching loss \\( \\mathrm{L}_c \\), and fine-grained matching loss \\( \\mathrm{L}_f \\), defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.738, + 0.905, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\mathrm {L} = \\mathrm {L} _ {d} + \\mathrm {L} _ {c} + \\mathrm {L} _ {f}. \\tag {15}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.779, + 0.646, + 0.796 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.834, + 0.82 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.9 + ], + "angle": 0, + "content": "TZC dataset [7] consists of 980 base maps with a spatial resolution of \\(1\\mathrm{m}\\), each with a high resolution of \\(5000\\times 5000\\) pixels. This dataset captures various land cover types, including dense forests, barren land, water bodies, agricultural fields, and human settlements." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8245" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of retrieval approaches. Top 1, 5, 10 recall are shown on the TZC dataset, and top 1, 2, 3 recall are shown on the MTGL40-5 dataset. Re-ranking thresholds are set at 100 for TZC and 50 for MTGL40-5." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.13, + 0.905, + 0.37 + ], + "angle": 0, + "content": "
Retrieval modeMethodTZCMTGL40-5
R@1R@5R@10Time(s)R@1R@2R@3Time(s)
Global retrievalPatch-NetVLAD63.163.163.40.39158.260.762.30.352
SelaVPR68.969.269.50.04666.766.967.40.029
EarthLoc70.471.672.40.08267.267.568.20.064
Local Re-rankingPatch-NetVLAD65.666.266.421.562.364.264.810.7
SelaVPR72.673.173.40.27470.171.572.60.133
SP-SuperGlue71.472.873.512.169.370.773.65.87
Local retrievalSelaVPR74.475.675.93.2 × 10273.373.675.511.0
LoFTR_Coarse81.683.284.48.7 × 10376.478.480.53.9 × 102
TopicFM_Coarse83.783.884.59.5 × 10381.483.184.54.3 × 102
AdaMatcher_Coarse87.289.191.59.7 × 10384.686.287.84.5 × 102
Asymmetric local retrievalOur(LoFTR)86.188.490.134.282.383.183.71.81
Our(TopicFM)88.690.292.142.486.488.390.02.17
Our90.693.293.71.07191.693.193.90.165
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.395, + 0.482, + 0.488 + ], + "angle": 0, + "content": "MTGL40-5 dataset [31] focuses on 20 carefully selected scenes of ports and airports. Each scene has been standardized to \\(5120 \\times 5120\\) pixels, with a spatial resolution of \\(0.5\\mathrm{m}\\), and comprises five temporally distinct images. The appendix comprehensively overviews both datasets, including the generated prompt texts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.488, + 0.483, + 0.639 + ], + "angle": 0, + "content": "Parameter setting. We employ the Adam optimizer for model training, set the learning rate as \\(5.0 \\times 10^{-4}\\) and the batch size as 2. The threshold \\(\\theta\\) is fixed at 0.2. The window size for Multi-level Fine Matching is 4. The spatial compatibility parameter \\(Z\\) is 10. The focal loss parameters \\(\\alpha\\) and \\(\\gamma\\) are configured as 0.25 and 2, respectively. For the MTGL40-5 dataset, \\(\\gamma\\) is reduced to 1 to account for temporal variations. Additionally, during the category and attribute distillation, temperature parameters are respectively set as \\(1/5\\) and 1 to facilitate effective knowledge transfer." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.651, + 0.303, + 0.667 + ], + "angle": 0, + "content": "4.2. Retrieval Performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Quantitative comparison. We evaluate retrieval performance of various state-of-the-art methods on TZC and MTGL40-5 datasets, as shown in Table 1. The global retrieval approaches, such as Patch-NetVLAD[18], SelaVPR[29], and EarthLoc[4], exhibit minimal latency but suffer from significantly low recall rates on both datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Local re-ranking strategies built upon global retrieval improve performance, where SP-SuperGlue re-ranks EarthLoc candidates by matching SuperPoint [10] features through SuperGlue [42] matcher. In the case, SelaVPR has significant gains, achieving \\(3.7\\%\\) and \\(3.4\\%\\) R@1 gains on TZC and MTGL40-5 datasets, respectively. Compared to similar methods, SelaVPR reduces computational overhead associated with geometric verification by leveraging matching-pair counts as similarity scores. However, these" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.395, + 0.907, + 0.426 + ], + "angle": 0, + "content": "enhancements remain constrained by the limited positive sample coverage in global retrieval candidate sets." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.427, + 0.907, + 0.683 + ], + "angle": 0, + "content": "Moreover, we compare methods that rely solely on local features for retrieval. It is obvious that the local feature retrieval-based SelaVPR incurs a two-order-of-magnitude computational cost increase, primarily due to redundant symmetric processing, which expands 980 base images into approximately \\(10^{5}\\) database images. In contrast, our asymmetric-input architecture directly processes the 980 original images, significantly improving efficiency. Coarse matcher-based local retrieval approaches like LoFTR_Coarse [47], TopicFM_Coarse [15], and AdaMatcher_Coarse [20], outperform SelaVPR through fine-grained supervision but incur three-order-of-magnitude higher online retrieval latency. Furthermore, the coarse matching encoders of LoFTR and TopicFM are integrated into our framework to illustrate the effectiveness of the asymmetric-input architecture on enhancing performance and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.685, + 0.909, + 0.822 + ], + "angle": 0, + "content": "In addition, compared to the performance on both datasets, most methods underperform on the MTGL40-5 dataset. Since multi-source modality gaps in the MTGL40-5 dataset challenge robust features extraction, even though it contains significantly fewer base images than the TZC dataset. Conversely, our paradigm outperforms other methods on retrieval performance and efficiency on both datasets, which illustrates the effect of topic features on learning consistent and invariant representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.729, + 0.849 + ], + "angle": 0, + "content": "4.3. Matching Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Quantitative comparison. Table 2 shows the matching performance and time of state-of-the-art methods. To avoid retrieval preference, positive samples from SelaVPR's lo" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8246" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.095, + 0.865, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.51, + 0.907, + 0.555 + ], + "angle": 0, + "content": "Figure 3. A visualized comparison of the state-of-the-art approaches on TZC dataset. Our unified framework generates numerous accurate correspondences in both texture-scarce barren areas and texture-dense urban areas. For images with large resolution-scaled discrepancies, our approach matches regions that are easier to match although they are sparse." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.577, + 0.483, + 0.633 + ], + "angle": 0, + "content": "Table 2. Comparison of matching approaches. \\( n = 1 \\) represents the original matching, while \\( n = 4 \\) indicates four iterations of the EarthMatch protocol. IoU is used as an effective evaluation metric." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.644, + 0.48, + 0.786 + ], + "angle": 0, + "content": "
CategoryMethodTZC\nn=1 n=4MTGL\nn=1 n=4Time(s)\nn=1 n=4
Detector-basedSIFT-SuperGlue82.4 89.570.6 78.60.24 2.81
SP-SuperGlue84.4 91.277.4 82.40.14 2.37
Dense matcherRoMa92.2 96.186.7 90.20.79 5.17
Detector\nfreeLoFTR85.7 90.377.3 84.10.12 2.18
AdaMatcher91.4 93.783.3 88.40.15 2.51
TopicFM89.2 92.478.5 84.50.13 2.32
Ours98.9 99.191.1 92.70.09 2.02
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.903 + ], + "angle": 0, + "content": "cal re-ranking are used for matching, while the EarthMatch [3] protocol is applied to iteratively optimize matching and compensate for the limited receptive field of other methods. It is obvious that our method achieves state-of-the-art performance on both protocols while significantly improving" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.907, + 0.744 + ], + "angle": 0, + "content": "efficiency. The proposed asymmetric-input direct matching framework reduces computational time by 1/20 of the EarthMatch protocol with merely \\(1.6\\%\\) accuracy loss on TZC dataset. Notably, while recent dense matcher RoMa [12] delivers competitive accuracy (\\(90.2\\%\\) on the MTGL dataset), it requires \\(2.5 \\times\\) longer for per match than our solution. The multi-temporal discrepancies in MTGL dataset pose greater challenges than TZC's synthetic data, where our matching method effectively handles heterogeneous differences through consistent semantics and adaptive matching mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Visual comparison. Figure 3 displays the matching results of the crucial first iteration under the EarthMatch protocol. Notably, our method searches over a broader range compared to others. It is seen that our method excels in texture-scarce (such as barren areas) and texture-dense (like urban regions) scenarios through adaptive refinement and relaxed matching constraints in MKNN. Under extreme scale variations, existing methods like LoFTR distribute numerous matching pairs on non-discriminative natural terrains, which leads to low matching rates and disruption of iter-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8247" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 3. Ablation of topic-related cross-attention layers and weighted loss on TZC dataset. The baseline, C, and T are the self-attention encoder, the cross-attention layers in LoFTR, and the topic-related layers of TopicFM, respectively." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.157, + 0.49, + 0.246 + ], + "angle": 0, + "content": "
BaseC/TCategoryAttributeWLossR@1Time(s)
✓82.40.87
✓86.1/88.634/42
✓85.80.94
✓✓88.11.07
✓✓✓90.61.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.259, + 0.485, + 0.318 + ], + "angle": 0, + "content": "Table 4. Ablation of topic-independent components on TZC dataset. A baseline is used that incorporates the topic encoder, original adaptive label assignment [20], and expectation-based fine-grained matcher [47]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.327, + 0.485, + 0.438 + ], + "angle": 0, + "content": "
BaseACAMKNNGGVMFMR@1IoUPairsTime(s)
✓(0,0)84.972.31550.94/0.07
✓(0,0)86.575.41710.94/0.07
✓(0,1)86.580.7171/1920.94/0.08
✓(0,1)✓90.685.1171/1921.07/0.08
✓(1,1)✓87.181.21921.07/0.08
✓(0,1)✓✓90.690.5171/8141.07/0.09
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.508 + ], + "angle": 0, + "content": "ative process. Whereas our topic-aware weighted loss encourages the model to match regions that are easier to match although they are sparse, thereby enhancing reliability." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.517, + 0.245, + 0.533 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.538, + 0.483, + 0.81 + ], + "angle": 0, + "content": "The effect of topics. Our method uses distilled topics to improve efficiency and performance. Therefore, we separately evaluate the effect of the category and attribute, and the derived category weighted loss on the TZC dataset, as illustrated in Table 3. The cross-image perception of LoFTR improves \\( R@1 \\) by \\( 3.7\\% \\), while TopicFM's self-supervised topic achieves \\( 6.2\\% \\) gains. Both encoders that require crossattention for information interaction cause \\( 40\\times \\) slower online retrieval than the baseline (0.87s). Our category embeddings brings \\( 3.4\\% \\) improvement, attribute embeddings adds \\( 2.7\\% \\), and weight loss contributes \\( 2.5\\% \\). Moreover, the lightweight parallel architecture avoids significant time consumption. Figure 4 further visually compares topic semantics of remote sensing land cover generated by our method and the self-supervised approach TopicFM. It is seen that our method's deterministic land-cover topic modeling better distinguishes terrain features than self-supervised approach, enhancing geolocation precision." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The effect of other modules. Our evaluation first examines adaptive center assignment (ACA) and MKNN modules, followed by global geometric verification (GGV) and multi-level fine matching (MFM). During testing, the retrieval and matching is in sequence: failed retrieval yields zero matching IoU. When GGV is disabled, match count" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.092, + 0.887, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.411, + 0.908, + 0.482 + ], + "angle": 0, + "content": "Figure 4. Visualization of category topics in remote images. Our human-perception-aligned topic distributions encourage multisource feature alignment for geospatial retrieval and matching, whereas self-supervised structural modeling struggles to adapt to complex remote sensing scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.909, + 0.723 + ], + "angle": 0, + "content": "directly determines retrieval scores. The two values of MKNN correspond to the \\( k \\) value in retrieval and coarse matching. As Table 4 shows, ACA increases initial matching pairs by \\( 10.3\\% \\) over the original adaptive top-left assignment. MKNN (\\( k = 1 \\)) further elevates this by \\( 12.3\\% \\). The GGV enhances retrieval accuracy by \\( 4.1\\% \\) with only an additional time consumption (0.13s). One-to-many with \\( k = 1 \\) during retrieval makes the GGV score unstable and the recall rate decreases. MFM amplifies matching pairs by \\( 5 \\times \\) with \\( 5.4\\% \\) IoU improvement, while maintaining low time overhead. Crucially, correctly retrieved samples achieve \\( 99\\% \\) (90.5(matching)/90.6(retrieval)) average matching IoU, indicating the task-coherent feature learning of retrieval and matching tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.634, + 0.755 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We introduced a unified asymmetric-input retrieval-matching framework for geolocation. By combining object semantic modeling with image retrieval and matching, our method learns a robust set of primary features that can be stored offline. Moreover, we design a center-based label adaptive assignment and a multi-level fine-grained local matching pipeline, optimizing the matching precision. Compared to state-of-the-art methods, the proposed approach is both efficient and interpretable." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8248" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant No.62372358 and 62302355, the Fundamental Research Funds for the Central Universities under Grant No.XJSJ24071 and No.XJSJ24072, and the Xidian University Specially Funded Project for Interdisciplinary Exploration under Grant No. TZJH2024026." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.237, + 0.188, + 0.254 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.262, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[1] Artem Babenko and Victor Lempitsky. Aggregating local deep features for image retrieval. In Proceedings of the IEEE international conference on computer vision, pages 1269-1277, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.32, + 0.484, + 0.361 + ], + "angle": 0, + "content": "[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer vision and image understanding, 110(3):346-359, 2008. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.363, + 0.483, + 0.445 + ], + "angle": 0, + "content": "[3] Gabriele Berton, Gabriele Goletto, Gabriele Trivigno, Alex Stoken, Barbara Caputo, and Carlo Masone. Earthmatch: Iterative coregistration for fine-grained localization of astronaut photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4264-4274, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.448, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[4] Gabriele Berton, Alex Stoken, Barbara Caputo, and Carlo Masone. Earthloc: Astronaut photography localization by indexing earth from space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12754-12764, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.519, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[5] Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 778-792. Springer, 2010. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.484, + 0.673 + ], + "angle": 0, + "content": "[6] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[7] \"Tianzhi Cup\" Artificial Intelligence Challenge Organizing Committee. zhihuidiqiu2024, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[8] Rodrigo Caye Daudt, Bertrand Le Saux, Alexandre Boulch, and Yann Gousseau. Multitask learning for large-scale semantic change detection. Computer Vision and Image Understanding, 187:102783, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.262, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.133 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[12] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[13] JR Elliott. Earth observation for the assessment of earthquake hazard, risk and disaster management. Surveys in geophysics, 41(6):1323-1354, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.906, + 0.302 + ], + "angle": 0, + "content": "[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.906, + 0.357 + ], + "angle": 0, + "content": "[15] Khang Truong Giang, Soohwan Song, and Sungho Jo. Top-icfm: Robust and interpretable topic-assisted feature matching. In Proceedings of the AAAI conference on artificial intelligence, pages 2447-2455, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.906, + 0.412 + ], + "angle": 0, + "content": "[16] Khang Truong Giang, Soohwan Song, and Sungho Jo. Top-icfm+: Boosting accuracy and efficiency of topic-assisted feature matching. IEEE Transactions on Image Processing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.906, + 0.467 + ], + "angle": 0, + "content": "[17] Pierre Gleize, Weiyao Wang, and Matt Feiszli. Silk: Simple learned keypoints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22499-22508, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[18] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14141-14152, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.597, + 0.906, + 0.677 + ], + "angle": 0, + "content": "[20] Dihe Huang, Ying Chen, Yong Liu, Jianlin Liu, Shang Xu, Wenlong Wu, Yikang Ding, Fan Tang, and Chengjie Wang. Adaptive assignment for geometry aware local feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5425-5434, 2023. 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.746 + ], + "angle": 0, + "content": "[21] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.906, + 0.803 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.805, + 0.908, + 0.872 + ], + "angle": 0, + "content": "[23] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Chao Liu, Jingjing Ma, Xu Tang, Fang Liu, Xiangrong Zhang, and Licheng Jiao. Deep hash learning for remote" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8249" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "sensing image retrieval. IEEE Transactions on Geoscience and Remote Sensing, 59(4):3420-3443, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[25] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[26] Ganchao Liu, Chao Li, Sihang Zhang, and Yuan Yuan. Vlmlf: Uav visual localization based on multi-source image feature learning. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[27] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[28] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[29] Feng Lu, Lijun Zhang, Xiangyuan Lan, Shuting Dong, Yaowei Wang, and Chun Yuan. Towards seamless adaptation of pre-trained models for visual place recognition. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[30] Christian Lülf, Denis Mayr Lima Martins, Marcos Antonio Vaz Salles, Yongluan Zhou, and Fabian Gieseke. Clipbranches: Interactive fine-tuning for text-image retrieval. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2719-2723, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[31] Jingjing Ma, Shiji Pei, Yuqun Yang, Xu Tang, and Xiangrong Zhang. Mtgl40-5: A multi-temporal dataset for remote sensing image geo-localization. Remote Sensing, 15(17):4229, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[32] Xianping Ma, Xiaokang Zhang, and Man-On Pun. Rs 3 mamba: Visual state space model for remote sensing image semantic segmentation. IEEE Geoscience and Remote Sensing Letters, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[33] Xianping Ma, Xiaokang Zhang, Man-On Pun, and Ming Liu. A multilevel multimodal fusion transformer for remote sensing semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.73 + ], + "angle": 0, + "content": "[34] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[35] Florent Perronnin, Yan Liu, Jorge Sánchez, and Hervé Poirier. Large-scale image retrieval with compressed fisher vectors. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3384-3391. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[36] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143-11152, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning cnn image retrieval with no human annotation. IEEE" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "transactions on pattern analysis and machine intelligence, 41(7):1655-1668, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[39] Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and Robin Robertson Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. IEEE Access, 9:89644-89654, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[40] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[41] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12716–12725, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.907, + 0.475 + ], + "angle": 0, + "content": "[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[43] Yanqing Shen, Sanping Zhou, Jingwen Fu, Ruotong Wang, Shitao Chen, and Nanning Zheng. Structvpr: Distill structural knowledge with weighting samples for visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11217-11226, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[44] Edgar Simo-Serra, Eduard Trulls, Luis Ferraz, Iasonas Kokkinos, Pascal Fua, and Francesc Moreno-Noguer. Discriminative learning of deep convolutional feature point descriptors. In Proceedings of the IEEE international conference on computer vision, pages 118-126, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[45] Alex Stoken and Kenton Fisher. Find my astronaut photo: Automated localization and georectification of astronaut photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6196-6205, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[46] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[47] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8918-8927, 2021. 1, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[48] Xu Tang, Qiushuo Ma, Xiangrong Zhang, Fang Liu, Jingjing Ma, and Licheng Jiao. Attention consistent network for remote sensing scene classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:2030-2045, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8250" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[49] Haoyang Wang, Fuhui Zhou, and Qihui Wu. Accurate vision-enabled uav location using feature-enhanced transformer-driven image matching. IEEE Transactions on Instrumentation and Measurement, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[50] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.207, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[51] Shuhang Wang, Jin Zheng, Hai-Miao Hu, and Bo Li. Naturalness preserved enhancement algorithm for non-uniform illumination images. IEEE transactions on image processing, 22(9):3538-3548, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[52] Xiang Wang, Shiwei Zhang, Jun Cen, Changxin Gao, Yingya Zhang, Deli Zhao, and Nong Sang. Clip-guided prototype modulating for few-shot action recognition. International Journal of Computer Vision, 132(6):1899-1912, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.321, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[53] Qinghua Ye, Yuzhe Wang, Lin Liu, Linan Guo, Xueqin Zhang, Liyun Dai, Limin Zhai, Yafan Hu, Nauman Ali, Xinhui Ji, et al. Remote sensing and modeling of the cryosphere in high mountain asia: A multidisciplinary review. Remote Sensing, 16(10):1709, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.39, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[54] Hongsheng Zhang and Ru Xu. Exploring the optimal integration levels between sar and optical data for better urban land cover mapping in the pearl river delta. International journal of applied earth observation and geoinformation, 64:87-95, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.46, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[55] Yesheng Zhang and Xu Zhao. Mesa: Matching everything by segmenting anything. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20217-20226, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "8251" + } + ] +] \ No newline at end of file diff --git a/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/43f123eb-2c45-4de4-8a4a-4958a152775f_origin.pdf b/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/43f123eb-2c45-4de4-8a4a-4958a152775f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..334c36335d1426ad3adc0ec57e579673101cabb2 --- /dev/null +++ b/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/43f123eb-2c45-4de4-8a4a-4958a152775f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5fb861cebfc7bacf65b9254c53203690e5af285ec8b8a0e378844fed99cc8b8c +size 3634754 diff --git a/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/full.md b/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..25869510b1d8b9eeb6bfe9bf9b201b39037a2131 --- /dev/null +++ b/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/full.md @@ -0,0 +1,322 @@ +# TopicGeo: An Efficient Unified Framework for Geolocation + +Xin Wang Xinlin Wang* Shuiping Gou + +Xidian University, Xi'an, China + +{xinwangai@stu., wangxinlin@, shpgou@mail.}xidian.edu.cn + +![](images/adbfe4d900636fe8e1d4f8c38d08d587188756ab66a5d1b971a461df590c7459.jpg) +(a) Low-texture area(980 5000×5000 base images) + +![](images/4c608631d97709b16dee6ee44c8426b5c480dd7f4ea07db3ac59ae324409703f.jpg) + +![](images/e65a6d5df9f7ddbd40976f928198a49181734ea6e376e95f75b088baf1d082c7.jpg) +Figure 1. The visual geolocation results of our TopicGeo, and typical retrieval and matching approaches: Earthloc [4], SelaVPR [29], and LoFTR [47]. "Recall" represents retrieval results. The color-coded lines show matching correctness: red for higher-error matches, green for higher-precision ones. In various scenes, our method accurately retrieves the target base image containing the query image from the database, while establishing pixel-level correspondence between the smaller query and large-scale retrieved image. + +![](images/eb80c29623c87844f81db9ee512708cbf4591304e737610ea841957514f5c5e4.jpg) +(b) Multi-temporal change area (40 5120×5120 base images) + +![](images/7a0d42e5c664ef25bed7c7422f3aa395f48da4d8c7b84cce1accbf976d46112b.jpg) + +![](images/c91051cdae460b609a9f69a7e434c113ee6d5fe5755a263ebcbd889a3f4d4151.jpg) + +# Abstract + +Vision-based geolocation techniques that establish spatial correspondences between smaller query images and larger georeferenced images have gained significant attention. Existing approaches typically employ a separate "retrieve-then-match" paradigm, whereas such paradigms suffer from computational inefficiency or precision limitations. To this end, we propose TopicGeo, a unified framework for direct and precise query-to-reference image matching via three key innovations. The textual object semantics, called topics, distilled from CLIP prompt learning are embedded into the geolocation framework to eliminate intra-class and inter-class distribution discrepancies while also enhancing processing efficiency. Center-based adaptive label assignment and outlier rejection mechanisms as a joint retrieval-matching optimization strategy ensure task-coherent feature learning and precise spatial correspondences. A multi-level fine matching pipeline is introduced to refine matching from quality and quantity. Evaluations on large-scale synthetic + +and real-world datasets illustrate that TopicGeo achieves state-of-the-art performance in retrieval recall and matching accuracy while maintaining a balance in computational efficiency. + +# 1. Introduction + +Image matching through establishing spatial correspondence between localized query images and extensive georeferenced databases has emerged as a critical technique for vision-based geolocation, playing a vital role in environmental monitoring [13, 53], human activities[26, 54]. However, this task faces inherent challenges. Sensor variations and temporal differences induce substantial discrepancies [31] and nonlinear geometric deformations, complicating cross-image accurate matching. Moreover, repetitive low-saliency regions, such as forests, deserts, and water bodies in georeferenced images, pose matching ambiguities. + +Recent advances in satellite image retrieval have focused on mitigating these challenges through global retrieval approaches [4, 24, 48]. These methods encode entire images + +as high-dimensional vectors for rapid retrieval of similar images from a database. However, such methods exhibit certain critical limitations. Global feature aggregation dilutes spatially fine-grained discriminative patterns, causing failures in low-saliency regions, such as uniform forests or barren land. In addition, when pairing large-size reference images, images are often cropped into smaller and partially overlapped pieces [4], which causes redundant computation. Local re-ranking methods [18, 29, 43] address these challenges by re-ranking candidates based on local matches. However, false positive candidates that are merely visually similar do not contribute to performance improvements. Retrieval-based methods exhibit limited geolocation accuracy, posing significant challenges for researchers in performing time-consuming and laborious manual corrections on multi-source imagery. A natural approach would involve first extracting candidate regions via a retrieval method, then refining for pixel-level matching with a specialized model [3, 31, 45]. However, a critical limitation of this approach lies in the decoupled retrieval and matching pipelines, which inherently prioritize different features in different tasks. In addition, the matching remains heavily dependent on the retrieval model's effectiveness. + +Recent studies [45, 49] apply a single end-to-end matching model for joint retrieval and matching. While enhanced by fine-grained features to boost accuracy, these methods face a huge time overhead [45], preventing real-time geolocation. The above geolocation methods rely on visual features for location estimation. However, remote sensing images exhibit a pronounced imbalance of areal distribution of terrestrial features, inducing inherent bias in visual information-dependent models toward overrepresented categories. Moreover, geometric distortions between multisource images and cross-temporal visual discrepancies lead to variations in the distribution of geographic features at the same location. Some methods [43, 55] learn more robust representations using natural images by incorporating semantic segmentation. In remote sensing, datasets [8, 39] and segmentation models [32, 33, 50] offer semantic information but face generalization limitations. Recent studies [15, 16] leverage self-supervised learning of image topics to enhance visual features. However, the resulting topic distributions often lack explicit semantic information. + +In this study, we propose a model that jointly optimizes retrieval and matching tasks for geographic positioning. Specifically, we employ asymmetric-input architecture to minimize redundant computation in large-displacement and large-scale scenarios. The core innovation lies in hierarchical textual embeddings for prompt learning and a topic-aware encoder trained via knowledge distillation and spatial alignment, yielding efficient topic features that are beneficial for both image retrieval and coarse matching. Building on this foundation, a multi-level refinement pipeline it- + +eratively resolves positional cues across scales, improving quality and quantity of matching points. Our unified framework is jointly optimized through effective label supervision and quality-aware rejection mechanisms. The contributions of the study are as follows: + +- We propose a unified retrieval and matching framework with asymmetric large-scale input, boosted by embedded topics for efficiency and robustness to intra-class discrepancies. +- The center-based label assignment is designed to optimize the contrastive loss, which is further weighted through topic-based category perception to balance inter-class discrepancies. +- The local spatial consensus filter and the global geometric verification module are co-designed to eliminate mismatched correspondences, thereby enhancing retrieval recall and matching precision in geolocation tasks. + +# 2. Related Work + +Visual Language Model. The pioneering visual-language model (VLM) CLIP [38] employs vision and text encoders to project images and texts into a shared semantic space, which achieves joint embedding of images and texts, making cross-modal retrieval and zero-shot learning possible. Since then, VLMs have rapidly advanced and been applied across fields like open-vocabulary detection [6], image-text retrieval [30], and action recognition [52]. Menon et al. [34] apply large language models (LLMs) to represent object classes as semantic attribute sets for enhanced classification. Recently, Liu et al. present the first remote sensing visual-language foundation model, RemoteCLIP[25], learning semantically rich visual features aligned with text embeddings for seamless downstream applications. Our method leverages the textual semantics information of objects to capture invariant matching features across multi-source and multi-temporal satellite images. + +Image Retrieval. The image retrieval task focuses on identifying images similar to a query image within a database, encompassing two primary approaches: global and local feature-based retrieval. In global retrieval, the fundamental challenge lies in effectively aggregating image features from diverse regions into a cohesive high-dimensional vector. Traditional methods accomplish this by aggregating handcrafted local features [2, 5, 28] into global representations through methods such as Fisher Vectors [35], and then performing efficient retrieval through nearest neighbor search algorithms. Recent advancements leverage deep learning for feature aggregation [1, 37], employing strategies such as margin-based classification loss functions [9, 27] and clustering training samples [4]. However, global retrieval often results in a low recall rate during the initial retrieval phase. To address the issue, local matching is often integrated to re-rank retrieval results [18, 29, 43]. Geomet- + +![](images/4605b39d3a4324b481c3988c80be0dd0c08ee6670da32e6037edc0533f20b3bd.jpg) +Figure 2. Overview of TopicGeo. (a) Multi-level features are extracted from input images. (b) CLIP-generated prompt embeddings and visual features are injected into cross-attention layers and then distilled to infer the topic distribution. The extracted topic embeddings are integrated with global context features from self-attention layers. Preliminary matches are obtained using dual-softmax and local mutual k-nearest neighbor (MKNN) filter, followed by global geometric verification(GGV) for retrieval. (c) Adaptive matching is performed on cropped patches at a secondary resolution, with further coordinate refinement through fine-grained feature patch cropping. + +ric verification methods like RANSAC [14] are employed to identify reliable candidates, though they frequently suffer from computational inefficiency. Fast verification approaches based on matching pair counts [29] face additional challenges in remote sensing applications, where weakly textured regions often lead to significant false positives. Unlike these retrieval-specific pipelines, this paper establishes a unified framework that seamlessly integrates retrieval and matching processes and introduces an efficient geometric verification mechanism. + +Image Matching Image matching establishes accurate spatial correspondences between two images, essential for tasks like visual localization [21, 41] and 3D reconstruction [23]. The core of image matching is keypoint detection and description. Classical methods [2, 5, 28] employ hand-crafted features to detect and describe keypoints. Recent advances leverage deep learning for improved sparse keypoint detection[10, 40, 44], and self-supervised descriptor learning [17]. SuperGlue [42] harnesses the graph neural network with attentional mechanism to address significant occlusion or viewpoint changes, but relying on the quality of keypoints. RoMa [12] significantly enhances accuracy via Markov chain-based dense matching but incurs high computational cost. Detector-free approaches like LoFTR [47] use position embedding and cross-attention for semi-dense matching, which has gained traction in remote sensing[3, 45, 49], yet struggle with large-scale or point-of-view transformation. AdaMatcher [20] proposes an adap + +tive label assignment strategy to solve this problem, but introduces inconsistent learning-driven global mismatch filtering. TopicFM [15] sets up learnable topic embeddings for topic perception, while the unknown topics diverge from human perception and fail in complex scenes. Most existing methods rely on cross-attention for feature interaction, which limits computational efficiency. In contrast, our approach leverages distilled topics for feature enhancement and a learning-free local filter for efficient correspondence refinement. + +# 3. Proposed Approach + +Our goal is to retrieve a reliable base image from a non-overlapping image database using a small query image and then match the query image to the retrieved larger-scale base image. Let $R = \{B_i | i \in [1, m]\}$ be a retrieval database containing $m$ high resolution base images, and $A$ be a query image. We establish a unified retrieval and matching framework, which encodes $R$ into features $F_R = \{F_{B_i} | i \in [1, m]\}$ storing as a feature database, and $A$ into $F_A$ . Then $A$ is retrieved, and the corresponding target base image $B$ is obtained. Finally, the fine-grained features of both $A$ and $B$ are extracted to refine the matching and compute the homography $H$ . Figure 2 gives an overview of our pipeline. + +# 3.1. Topic-aware Feature Extraction + +This paper tackles distribution discrepancies of land objects by establishing an effective topic perception process that integrates primitive visual features with high-level text representations independent of vision. Firstly, we extract visual features of the query image $A$ and the base image $B$ . To be efficient and preserve sufficient information, both of them undergo processing through patch embedding [11]. Then the downsampled images are fed into a shared ResNet [19] to extract coarse-grained visual features, $(F_A^{1/64}, F_B^{1/64})$ , and fine-grained visual features $(F_A^{1/16}, F_B^{1/16})$ and $(F_A^{1/4}, F_B^{1/4})$ . + +Hierarchical topic prompts. To accelerate convergence and integrate richer information, we design dual-level prompts for land objects. Specifically, generic and stable categories are extracted from established land cover datasets [50]. These categories are enriched through object attribute prompts generated by the GPT-4 language model. Subsequently, the embeddings of the augmented object prompts are encoded using the RemoteCLIP[25] text encoder. This process generates category embeddings $\mathbf{F}_{\mathrm{category}} \in \mathbb{R}^{C \times d}$ and attribute embeddings $\mathbf{F}_{\mathrm{attribute}} \in \mathbb{R}^{C' \times d}$ . + +Object topic distillation. The obtained attribute and category prompt embeddings are embedded into visual features in sequence via cross-attention layers, where visual features serve as queries while these prompt embeddings act as keys and values. This process synthesizes the fused textual and visual representation to generate topic embeddings $\mathbf{F}_{\mathrm{topic}}^{1/64}$ . Notably, the sparsity of textual semantic tokens drastically lowers the computational cost of cross-attention. + +However, such embeddings fail to align category-specific and attribute-related information with their corresponding visual semantics. Leveraging the inherent alignment between visual and textual features in RemoteCLIP, we employ its frozen vision encoder to further extract the visual embeddings $\mathbf{F}_{\mathrm{vision}} \in \mathbb{R}^{N \times d}$ of $N$ patches from $A$ . Specifically, the patch-level category perception is established by computing the category probability distribution $p \in \mathbb{R}^{N \times C}$ : + +$$ +p = \operatorname {s o f t m a x} \left(\frac {\mathbf {F} _ {\text {v i s i o n}} \mathbf {F} _ {\text {c a t e g o r y}} ^ {T}}{t}\right), \tag {1} +$$ + +where $t$ is the temperature parameter. + +To transfer the category-specific information from $\mathbf{F}_{\mathrm{vision}}$ to topic embedding $\mathbf{F}_{\mathrm{topic}}^{1/64}$ , the latter are spatially aligned with $\mathbf{F}_{\mathrm{vision}}$ via $2 \times 2$ pooling, resulting in resized topic embedding $\mathbf{F}_{\mathrm{topic}}^{\mathrm{p}} \in \mathbb{R}^{N \times d}$ . The corresponding category probability distribution for the pooled topic embedding is computed analogously to $p$ , denoted as $q \in \mathbb{R}^{N \times C}$ . Thus, the weighted Kullback-Leibler (KL) divergence loss $Loss_{\mathrm{category}}$ is employed to transfer the category-specific in- + +formation in $\mathbf{F}_{\mathrm{vision}}$ to topic embeddings, expressed as: + +$$ +L o s s _ {\text {c a t e g o r y}} = \sum_ {i = 0} ^ {N} \left(\sum_ {j = 0} ^ {C} \left(p _ {i j} \cdot \log \left(\frac {p _ {i j}}{q _ {i j}}\right)\right) \cdot w _ {i}\right) \cdot t ^ {2}, \tag {2} +$$ + +where $w_{i}$ , the normalized class frequency weight of the $i$ -th image patch, is derived from $p$ . Analogously, attribute-related information is transferred using $Loss_{attribute}$ . Then the topic embeddings are enriched by minimizing the combined loss: + +$$ +L _ {d} = \text {L o s s} _ {\text {c a t e g o r y}} + \text {L o s s} _ {\text {a t t r i b u t e .}} \tag {3} +$$ + +Note that distillation operates exclusively on query images due to the fully shared parameters and training efficiency. + +Topic feature integration. To capture global context absent from $\mathbf{F}_{\mathrm{topic}}$ , the linear self-attention[51] with rotary position embedding[46] is further employed. Thus, the outputs from parallel encoding branches are integrated via a multilayer perceptron (MLP), with features from one branch stochastically masked during optimization to mitigate MLP-induced training bias. The final topic features, denoted as $\hat{F}_A^{1/64}$ and $\hat{F}_B^{1/64}$ , are utilized for downstream retrieval and matching. + +# 3.2. Retrieval & Coarse Matching + +The multimodal topic features of query and base images, $\hat{F}_A^{1/64}$ and $\hat{F}_B^{1/64}$ , enable efficient retrieval and coarse matching. Inspired by [20], the coarse-level matching probability matrix $S$ and the bidirectional matching probability matrixes $\mathcal{P}$ are calculated as: + +$$ +\mathcal {S} (i, j) = \left\langle \hat {F} _ {1 / 6 4} ^ {A} (i), \hat {F} _ {1 / 6 4} ^ {B} (j) \right\rangle , \tag {4} +$$ + +$$ +\mathcal {P} ^ {A B} (i, j) = \operatorname {S o f t m a x} \left(\mathcal {S} (i, \cdot)\right) _ {j}, \tag {5} +$$ + +$$ +\mathcal {P} ^ {B A} (i, j) = \operatorname {S o f t m a x} \left(\mathcal {S} (\cdot , j)\right) _ {i}, \tag {6} +$$ + +where $\langle .,.\rangle$ is the inner product. Based on $\mathcal{P}^{AB}$ , an index pair $(\tilde{i},\tilde{j})$ is selected as the match if its corresponding confidence value exceeds the threshold $\theta$ and is the maximum in its row. Thus, matches $\mathcal{M}_{AB}$ is obtained as: + +$$ +\mathcal {M} _ {A B} = \left\{\left(\tilde {i}, \tilde {j}\right) \mid \mathcal {P} ^ {A B} (\tilde {i}, \tilde {j}) = \max _ {k} \mathcal {P} ^ {A B} (\tilde {i}, k), \mathcal {P} ^ {A B} \geq \theta \right\}. \tag {7} +$$ + +Analogously, $\mathcal{M}_{BA}$ is derived by applying identical thresholding and column-wise maximization operations to $\mathcal{P}^{BA}$ , completing the bidirectional coarse matching process. + +Local MKNN filter. While LoFTR [47] employs mutual nearest neighbor (MNN) filtering to suppress local mismatches, this strategy exhibits inherent limitations in one-to-many mismatches prevalent in resolution-scaled discrepancies. Instead, we propose mutual $k$ -nearest neighbor (MKNN) filtering, which relaxes mismatches constraints by retaining matches that satisfy $k$ -neighbor spatial consistency across bidirectional correspondence sets $\mathcal{M}_{AB}$ and $\mathcal{M}_{BA}$ . Specifically, a candidate match pair $(i,j) \in \mathcal{M}_{AB}$ + +is preserved if there exists a reverse-matching pair $(i', j) \in \mathcal{M}_{BA}$ , and the 2D coordinate pair $(x, y)$ and $(x', y')$ corresponding to points $i$ and $i'$ are $k$ -nearest neighbors. Analogously, $\mathcal{M}_{BA}$ is handled in the same manner. Finally, the refined coarse matches are combined using the OR operation to obtain $\mathcal{M}_c$ , and the overall process is formulated as follows: + +$$ +\mathcal {M} _ {c} = \operatorname {M K N N} \left(\mathcal {M} _ {A B}, \mathcal {M} _ {B A}\right). \tag {8} +$$ + +Global geometric verification. It is known that correct correspondences have geometric spatial consistency. Therefore, we utilize the rule to filter false positive samples in the database efficiently. Inspired by [36], we independently analyze distance and angle compatibility for each correspondence pair $\mathcal{M}_c(n)$ among the $N_{c}$ total pairs. For distance compatibility, we calculate pairwise Euclidean distances between each point and all other $N_{c} - 1$ points in both the query and base image coordinate systems, calculating cross-coordinate distance ratios and normalizing it. For angle compatibility, we evaluate the local geometric consistency by computing relative angles among each point's $k$ -nearest neighbors in both coordinate systems, and measuring their cross-coordinate absolute angular difference $c_{n}$ . Finally, the distance compatibility $\alpha_{n}$ and angle compatibility $\beta_{n}$ of each matching pair $\mathcal{M}_c(n)$ are yielded to jointly quantify spatial coherence, defined as: + +$$ +\alpha_ {n} = \max \left(0, \min \left(Z - \frac {\operatorname {V a r} \left(d _ {n}\right)}{\delta_ {d} {} ^ {2}}, Z - 1\right)\right), \tag {9} +$$ + +$$ +\beta_ {n} = \max \left(0, \min \left(Z - \frac {c _ {n} {} ^ {2}}{\delta_ {c} {} ^ {2}}, Z - 1\right)\right), \tag {10} +$$ + +where $\delta_{d}$ and $\delta_{c}$ are the maximum acceptable values, and $Z$ is a scalar, used to truncate outliers. The spatial coherence confidences weighted by $\lambda$ is utilized to re-weight the per-pair bidirectional matching probabilities $E_{n}$ , expressed as: + +$$ +\begin{array}{l} E _ {n} = \left(\lambda \beta_ {n} + (1 - \lambda) \alpha_ {n}\right) \\ \times \left(P ^ {A B} \left(\tilde {i} _ {n}, \tilde {j} _ {n}\right) + P ^ {B A} \left(\tilde {i} _ {n}, \tilde {j} _ {n}\right)\right). \tag {11} \\ \end{array} +$$ + +The global confidence $E$ for the base image is then obtained by aggregating all individual matching probabilities. Finally, the base image only with the highest overall confidence score is selected as the target retrieval image. + +The joint retrieval-matching loss $\mathrm{L}_c$ is defined as: + +$$ +\mathrm {L} _ {c} = \left(\operatorname {F L} \left(\mathcal {P} ^ {A B}, \hat {\mathcal {P}} ^ {A B}\right) + \operatorname {F L} \left(\mathcal {P} ^ {B A}, \hat {\mathcal {P}} ^ {B A}\right)\right) \times w _ {\text {t o k e n}}, \tag {12} +$$ + +where $\hat{\mathcal{P}}$ is the ground truth matching matrix. We allow one-to-many and many-to-one label assignments [20] for $\mathcal{P}^{AB}$ and $\mathcal{P}^{BA}$ . FL refers to the Focal Loss [22]. $w_{token}$ is derived from the copy of $w$ in the early training period, and the weight calculated by the category perception at the $\mathbf{F}_{\mathrm{topic}}^{1/64}$ is used in the subsequent training. + +# 3.3. Adaptive Center Assignment + +Downsampling operations inherently yield sparse matching pairs, exacerbated by resolution-scaled discrepancies and asymmetric input. The adaptive label assignment [20] and MKNN filter alleviate the issue, while its reliance on the upper-left coordinates of the warped grid [42, 47] for label assignment often induces suboptimal matching patch overlaps, especially under large rotational misalignments in remote sensing images. This paper proposes a center-based label adaptive assignment strategy. Labels are assigned by distorting grid center coordinates instead of upper-left points, which are later reconverted to the upper-left coordinates for feature patch indexing, thereby enhancing robustness in challenging conditions. + +# 3.4. Multi-level Fine Matching + +The retrieval-oriented low-resolution encoding fundamentally limits matching precision. To achieve enhanced correspondence density through operationally consistent processing with the coarse matching stage, we employ the same adaptive local window matching on the finer-grained feature patches $(\hat{F}_A^{1/16}, \hat{F}_B^{1/16})$ corresponding to $\mathcal{M}_c$ , thereby obtaining $\mathcal{M}_a$ . Subsequently, following LoFTR[47], higher-resolution feature patches $(\hat{F}_A^{1/4}, \hat{F}_B^{1/4})$ extracted from $(F_A^{1/4}, F_B^{1/4})$ guided by $\mathcal{M}_a$ are utilized for final expectation-based fine-grained matching, obtaining $\mathcal{M}_f$ , achieving sub-pixel matching accuracy. For adaptive local window matching, the loss is formulated as: + +$$ +\left. \operatorname {L o s s} _ {a} = \frac {1}{| \mathcal {M} _ {a} |} \left(\operatorname {F L} \left(\mathcal {P} _ {a} ^ {A B}, \hat {\mathcal {P}} _ {a} ^ {A B}\right) + \operatorname {F L} \left(\mathcal {P} _ {a} ^ {B A}, \hat {\mathcal {P}} _ {a} ^ {B A}\right)\right), \right. \tag {13} +$$ + +where $\hat{\mathcal{P}}_a$ represents the ground truth matching matrix. The expectation-based loss $\mathrm{Loss}_f$ refers to [47]. Thus, the final fine-grained loss is the sum of both: + +$$ +\mathrm {L} _ {f} = \operatorname {L o s s} _ {a} + \operatorname {L o s s} _ {f}. \tag {14} +$$ + +# 3.5. Training + +The model is trained by optimizing the topic distillation loss $\mathrm{L}_d$ , the joint retrieval-coarse matching loss $\mathrm{L}_c$ , and fine-grained matching loss $\mathrm{L}_f$ , defined as: + +$$ +\mathrm {L} = \mathrm {L} _ {d} + \mathrm {L} _ {c} + \mathrm {L} _ {f}. \tag {15} +$$ + +# 4. Experiments + +# 4.1. Datasets and Implementation Details + +TZC dataset [7] consists of 980 base maps with a spatial resolution of $1\mathrm{m}$ , each with a high resolution of $5000\times 5000$ pixels. This dataset captures various land cover types, including dense forests, barren land, water bodies, agricultural fields, and human settlements. + +Table 1. Quantitative comparison of retrieval approaches. Top 1, 5, 10 recall are shown on the TZC dataset, and top 1, 2, 3 recall are shown on the MTGL40-5 dataset. Re-ranking thresholds are set at 100 for TZC and 50 for MTGL40-5. + +
Retrieval modeMethodTZCMTGL40-5
R@1R@5R@10Time(s)R@1R@2R@3Time(s)
Global retrievalPatch-NetVLAD63.163.163.40.39158.260.762.30.352
SelaVPR68.969.269.50.04666.766.967.40.029
EarthLoc70.471.672.40.08267.267.568.20.064
Local Re-rankingPatch-NetVLAD65.666.266.421.562.364.264.810.7
SelaVPR72.673.173.40.27470.171.572.60.133
SP-SuperGlue71.472.873.512.169.370.773.65.87
Local retrievalSelaVPR74.475.675.93.2 × 10273.373.675.511.0
LoFTR_Coarse81.683.284.48.7 × 10376.478.480.53.9 × 102
TopicFM_Coarse83.783.884.59.5 × 10381.483.184.54.3 × 102
AdaMatcher_Coarse87.289.191.59.7 × 10384.686.287.84.5 × 102
Asymmetric local retrievalOur(LoFTR)86.188.490.134.282.383.183.71.81
Our(TopicFM)88.690.292.142.486.488.390.02.17
Our90.693.293.71.07191.693.193.90.165
+ +MTGL40-5 dataset [31] focuses on 20 carefully selected scenes of ports and airports. Each scene has been standardized to $5120 \times 5120$ pixels, with a spatial resolution of $0.5\mathrm{m}$ , and comprises five temporally distinct images. The appendix comprehensively overviews both datasets, including the generated prompt texts. + +Parameter setting. We employ the Adam optimizer for model training, set the learning rate as $5.0 \times 10^{-4}$ and the batch size as 2. The threshold $\theta$ is fixed at 0.2. The window size for Multi-level Fine Matching is 4. The spatial compatibility parameter $Z$ is 10. The focal loss parameters $\alpha$ and $\gamma$ are configured as 0.25 and 2, respectively. For the MTGL40-5 dataset, $\gamma$ is reduced to 1 to account for temporal variations. Additionally, during the category and attribute distillation, temperature parameters are respectively set as $1/5$ and 1 to facilitate effective knowledge transfer. + +# 4.2. Retrieval Performance + +Quantitative comparison. We evaluate retrieval performance of various state-of-the-art methods on TZC and MTGL40-5 datasets, as shown in Table 1. The global retrieval approaches, such as Patch-NetVLAD[18], SelaVPR[29], and EarthLoc[4], exhibit minimal latency but suffer from significantly low recall rates on both datasets. + +Local re-ranking strategies built upon global retrieval improve performance, where SP-SuperGlue re-ranks EarthLoc candidates by matching SuperPoint [10] features through SuperGlue [42] matcher. In the case, SelaVPR has significant gains, achieving $3.7\%$ and $3.4\%$ R@1 gains on TZC and MTGL40-5 datasets, respectively. Compared to similar methods, SelaVPR reduces computational overhead associated with geometric verification by leveraging matching-pair counts as similarity scores. However, these + +enhancements remain constrained by the limited positive sample coverage in global retrieval candidate sets. + +Moreover, we compare methods that rely solely on local features for retrieval. It is obvious that the local feature retrieval-based SelaVPR incurs a two-order-of-magnitude computational cost increase, primarily due to redundant symmetric processing, which expands 980 base images into approximately $10^{5}$ database images. In contrast, our asymmetric-input architecture directly processes the 980 original images, significantly improving efficiency. Coarse matcher-based local retrieval approaches like LoFTR_Coarse [47], TopicFM_Coarse [15], and AdaMatcher_Coarse [20], outperform SelaVPR through fine-grained supervision but incur three-order-of-magnitude higher online retrieval latency. Furthermore, the coarse matching encoders of LoFTR and TopicFM are integrated into our framework to illustrate the effectiveness of the asymmetric-input architecture on enhancing performance and efficiency. + +In addition, compared to the performance on both datasets, most methods underperform on the MTGL40-5 dataset. Since multi-source modality gaps in the MTGL40-5 dataset challenge robust features extraction, even though it contains significantly fewer base images than the TZC dataset. Conversely, our paradigm outperforms other methods on retrieval performance and efficiency on both datasets, which illustrates the effect of topic features on learning consistent and invariant representations. + +# 4.3. Matching Performance + +Quantitative comparison. Table 2 shows the matching performance and time of state-of-the-art methods. To avoid retrieval preference, positive samples from SelaVPR's lo + +![](images/91ce6e5cb2abc04e2f423eaf14bc3f47b8e5c59d903079362052e88324b1b909.jpg) +Figure 3. A visualized comparison of the state-of-the-art approaches on TZC dataset. Our unified framework generates numerous accurate correspondences in both texture-scarce barren areas and texture-dense urban areas. For images with large resolution-scaled discrepancies, our approach matches regions that are easier to match although they are sparse. + +Table 2. Comparison of matching approaches. $n = 1$ represents the original matching, while $n = 4$ indicates four iterations of the EarthMatch protocol. IoU is used as an effective evaluation metric. + +
CategoryMethodTZC +n=1 n=4MTGL +n=1 n=4Time(s) +n=1 n=4
Detector-basedSIFT-SuperGlue82.4 89.570.6 78.60.24 2.81
SP-SuperGlue84.4 91.277.4 82.40.14 2.37
Dense matcherRoMa92.2 96.186.7 90.20.79 5.17
Detector +freeLoFTR85.7 90.377.3 84.10.12 2.18
AdaMatcher91.4 93.783.3 88.40.15 2.51
TopicFM89.2 92.478.5 84.50.13 2.32
Ours98.9 99.191.1 92.70.09 2.02
+ +cal re-ranking are used for matching, while the EarthMatch [3] protocol is applied to iteratively optimize matching and compensate for the limited receptive field of other methods. It is obvious that our method achieves state-of-the-art performance on both protocols while significantly improving + +efficiency. The proposed asymmetric-input direct matching framework reduces computational time by 1/20 of the EarthMatch protocol with merely $1.6\%$ accuracy loss on TZC dataset. Notably, while recent dense matcher RoMa [12] delivers competitive accuracy ( $90.2\%$ on the MTGL dataset), it requires $2.5 \times$ longer for per match than our solution. The multi-temporal discrepancies in MTGL dataset pose greater challenges than TZC's synthetic data, where our matching method effectively handles heterogeneous differences through consistent semantics and adaptive matching mechanisms. + +Visual comparison. Figure 3 displays the matching results of the crucial first iteration under the EarthMatch protocol. Notably, our method searches over a broader range compared to others. It is seen that our method excels in texture-scarce (such as barren areas) and texture-dense (like urban regions) scenarios through adaptive refinement and relaxed matching constraints in MKNN. Under extreme scale variations, existing methods like LoFTR distribute numerous matching pairs on non-discriminative natural terrains, which leads to low matching rates and disruption of iter- + +Table 3. Ablation of topic-related cross-attention layers and weighted loss on TZC dataset. The baseline, C, and T are the self-attention encoder, the cross-attention layers in LoFTR, and the topic-related layers of TopicFM, respectively. + +
BaseC/TCategoryAttributeWLossR@1Time(s)
✓82.40.87
✓86.1/88.634/42
✓85.80.94
✓✓88.11.07
✓✓✓90.61.07
+ +Table 4. Ablation of topic-independent components on TZC dataset. A baseline is used that incorporates the topic encoder, original adaptive label assignment [20], and expectation-based fine-grained matcher [47]. + +
BaseACAMKNNGGVMFMR@1IoUPairsTime(s)
✓(0,0)84.972.31550.94/0.07
✓(0,0)86.575.41710.94/0.07
✓(0,1)86.580.7171/1920.94/0.08
✓(0,1)✓90.685.1171/1921.07/0.08
✓(1,1)✓87.181.21921.07/0.08
✓(0,1)✓✓90.690.5171/8141.07/0.09
+ +ative process. Whereas our topic-aware weighted loss encourages the model to match regions that are easier to match although they are sparse, thereby enhancing reliability. + +# 4.4. Ablation Study + +The effect of topics. Our method uses distilled topics to improve efficiency and performance. Therefore, we separately evaluate the effect of the category and attribute, and the derived category weighted loss on the TZC dataset, as illustrated in Table 3. The cross-image perception of LoFTR improves $R@1$ by $3.7\%$ , while TopicFM's self-supervised topic achieves $6.2\%$ gains. Both encoders that require crossattention for information interaction cause $40\times$ slower online retrieval than the baseline (0.87s). Our category embeddings brings $3.4\%$ improvement, attribute embeddings adds $2.7\%$ , and weight loss contributes $2.5\%$ . Moreover, the lightweight parallel architecture avoids significant time consumption. Figure 4 further visually compares topic semantics of remote sensing land cover generated by our method and the self-supervised approach TopicFM. It is seen that our method's deterministic land-cover topic modeling better distinguishes terrain features than self-supervised approach, enhancing geolocation precision. + +The effect of other modules. Our evaluation first examines adaptive center assignment (ACA) and MKNN modules, followed by global geometric verification (GGV) and multi-level fine matching (MFM). During testing, the retrieval and matching is in sequence: failed retrieval yields zero matching IoU. When GGV is disabled, match count + +![](images/28a710a3017d7830dcd9ad3e57b454b1f18bf524d8e2bb876acc9d89b3a3f540.jpg) +Figure 4. Visualization of category topics in remote images. Our human-perception-aligned topic distributions encourage multisource feature alignment for geospatial retrieval and matching, whereas self-supervised structural modeling struggles to adapt to complex remote sensing scenarios. + +directly determines retrieval scores. The two values of MKNN correspond to the $k$ value in retrieval and coarse matching. As Table 4 shows, ACA increases initial matching pairs by $10.3\%$ over the original adaptive top-left assignment. MKNN ( $k = 1$ ) further elevates this by $12.3\%$ . The GGV enhances retrieval accuracy by $4.1\%$ with only an additional time consumption (0.13s). One-to-many with $k = 1$ during retrieval makes the GGV score unstable and the recall rate decreases. MFM amplifies matching pairs by $5 \times$ with $5.4\%$ IoU improvement, while maintaining low time overhead. Crucially, correctly retrieved samples achieve $99\%$ (90.5(matching)/90.6(retrieval)) average matching IoU, indicating the task-coherent feature learning of retrieval and matching tasks. + +# 5. Conclusion + +We introduced a unified asymmetric-input retrieval-matching framework for geolocation. By combining object semantic modeling with image retrieval and matching, our method learns a robust set of primary features that can be stored offline. Moreover, we design a center-based label adaptive assignment and a multi-level fine-grained local matching pipeline, optimizing the matching precision. Compared to state-of-the-art methods, the proposed approach is both efficient and interpretable. + +# 6. Acknowledgements + +# References + +[1] Artem Babenko and Victor Lempitsky. Aggregating local deep features for image retrieval. In Proceedings of the IEEE international conference on computer vision, pages 1269-1277, 2015. 2 +[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer vision and image understanding, 110(3):346-359, 2008. 2, 3 +[3] Gabriele Berton, Gabriele Goletto, Gabriele Trivigno, Alex Stoken, Barbara Caputo, and Carlo Masone. Earthmatch: Iterative coregistration for fine-grained localization of astronaut photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4264-4274, 2024. 2, 3, 7 +[4] Gabriele Berton, Alex Stoken, Barbara Caputo, and Carlo Masone. Earthloc: Astronaut photography localization by indexing earth from space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12754-12764, 2024. 1, 2, 6 +[5] Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 778-792. Springer, 2010. 2, 3 +[6] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 2 +[7] "Tianzhi Cup" Artificial Intelligence Challenge Organizing Committee. zhihuidiqiu2024, 2024. 5 +[8] Rodrigo Caye Daudt, Bertrand Le Saux, Alexandre Boulch, and Yann Gousseau. Multitask learning for large-scale semantic change detection. Computer Vision and Image Understanding, 187:102783, 2019. 2 +[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2 +[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 3, 6 + +This work was supported by the National Natural Science Foundation of China under Grant No.62372358 and 62302355, the Fundamental Research Funds for the Central Universities under Grant No.XJSJ24071 and No.XJSJ24072, and the Xidian University Specially Funded Project for Interdisciplinary Exploration under Grant No. TZJH2024026. +[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4 +[12] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 3, 7 +[13] JR Elliott. Earth observation for the assessment of earthquake hazard, risk and disaster management. Surveys in geophysics, 41(6):1323-1354, 2020. 1 +[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3 +[15] Khang Truong Giang, Soohwan Song, and Sungho Jo. Top-icfm: Robust and interpretable topic-assisted feature matching. In Proceedings of the AAAI conference on artificial intelligence, pages 2447-2455, 2023. 2, 3, 6 +[16] Khang Truong Giang, Soohwan Song, and Sungho Jo. Top-icfm+: Boosting accuracy and efficiency of topic-assisted feature matching. IEEE Transactions on Image Processing, 2024. 2 +[17] Pierre Gleize, Weiyao Wang, and Matt Feiszli. Silk: Simple learned keypoints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22499-22508, 2023. 3 +[18] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14141-14152, 2021. 2, 6 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[20] Dihe Huang, Ying Chen, Yong Liu, Jianlin Liu, Shang Xu, Wenlong Wu, Yikang Ding, Fan Tang, and Chengjie Wang. Adaptive assignment for geometry aware local feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5425-5434, 2023. 3, 4, 5, 6, 8 +[21] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 3 +[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5 +[23] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 3 +[24] Chao Liu, Jingjing Ma, Xu Tang, Fang Liu, Xiangrong Zhang, and Licheng Jiao. Deep hash learning for remote + +sensing image retrieval. IEEE Transactions on Geoscience and Remote Sensing, 59(4):3420-3443, 2020. 1 +[25] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 4 +[26] Ganchao Liu, Chao Li, Sihang Zhang, and Yuan Yuan. Vlmlf: Uav visual localization based on multi-source image feature learning. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1 +[27] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 2 +[28] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 2, 3 +[29] Feng Lu, Lijun Zhang, Xiangyuan Lan, Shuting Dong, Yaowei Wang, and Chun Yuan. Towards seamless adaptation of pre-trained models for visual place recognition. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6 +[30] Christian Lülf, Denis Mayr Lima Martins, Marcos Antonio Vaz Salles, Yongluan Zhou, and Fabian Gieseke. Clipbranches: Interactive fine-tuning for text-image retrieval. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2719-2723, 2024. 2 +[31] Jingjing Ma, Shiji Pei, Yuqun Yang, Xu Tang, and Xiangrong Zhang. Mtgl40-5: A multi-temporal dataset for remote sensing image geo-localization. Remote Sensing, 15(17):4229, 2023. 1, 2, 6 +[32] Xianping Ma, Xiaokang Zhang, and Man-On Pun. Rs 3 mamba: Visual state space model for remote sensing image semantic segmentation. IEEE Geoscience and Remote Sensing Letters, 2024. 2 +[33] Xianping Ma, Xiaokang Zhang, Man-On Pun, and Ming Liu. A multilevel multimodal fusion transformer for remote sensing semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2 +[34] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 2 +[35] Florent Perronnin, Yan Liu, Jorge Sánchez, and Hervé Poirier. Large-scale image retrieval with compressed fisher vectors. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3384-3391. IEEE, 2010. 2 +[36] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143-11152, 2022. 5 +[37] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning cnn image retrieval with no human annotation. IEEE + +transactions on pattern analysis and machine intelligence, 41(7):1655-1668, 2018. 2 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[39] Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and Robin Robertson Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. IEEE Access, 9:89644-89654, 2021. 2 +[40] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in neural information processing systems, 32, 2019. 3 +[41] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12716–12725, 2019. 3 +[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3, 5, 6 +[43] Yanqing Shen, Sanping Zhou, Jingwen Fu, Ruotong Wang, Shitao Chen, and Nanning Zheng. Structvpr: Distill structural knowledge with weighting samples for visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11217-11226, 2023. 2 +[44] Edgar Simo-Serra, Eduard Trulls, Luis Ferraz, Iasonas Kokkinos, Pascal Fua, and Francesc Moreno-Noguer. Discriminative learning of deep convolutional feature point descriptors. In Proceedings of the IEEE international conference on computer vision, pages 118-126, 2015. 3 +[45] Alex Stoken and Kenton Fisher. Find my astronaut photo: Automated localization and georectification of astronaut photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6196-6205, 2023. 2, 3 +[46] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4 +[47] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8918-8927, 2021. 1, 3, 4, 5, 6, 8 +[48] Xu Tang, Qiushuo Ma, Xiangrong Zhang, Fang Liu, Jingjing Ma, and Licheng Jiao. Attention consistent network for remote sensing scene classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:2030-2045, 2021. 1 + +[49] Haoyang Wang, Fuhui Zhou, and Qihui Wu. Accurate vision-enabled uav location using feature-enhanced transformer-driven image matching. IEEE Transactions on Instrumentation and Measurement, 2024. 2, 3 +[50] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733, 2021. 2, 4 +[51] Shuhang Wang, Jin Zheng, Hai-Miao Hu, and Bo Li. Naturalness preserved enhancement algorithm for non-uniform illumination images. IEEE transactions on image processing, 22(9):3538-3548, 2013. 4 +[52] Xiang Wang, Shiwei Zhang, Jun Cen, Changxin Gao, Yingya Zhang, Deli Zhao, and Nong Sang. Clip-guided prototype modulating for few-shot action recognition. International Journal of Computer Vision, 132(6):1899-1912, 2024. 2 +[53] Qinghua Ye, Yuzhe Wang, Lin Liu, Linan Guo, Xueqin Zhang, Liyun Dai, Limin Zhai, Yafan Hu, Nauman Ali, Xinhui Ji, et al. Remote sensing and modeling of the cryosphere in high mountain asia: A multidisciplinary review. Remote Sensing, 16(10):1709, 2024. 1 +[54] Hongsheng Zhang and Ru Xu. Exploring the optimal integration levels between sar and optical data for better urban land cover mapping in the pearl river delta. International journal of applied earth observation and geoinformation, 64:87-95, 2018. 1 +[55] Yesheng Zhang and Xu Zhao. Mesa: Matching everything by segmenting anything. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20217-20226, 2024. 2 \ No newline at end of file diff --git a/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/images.zip b/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..16b8ce5ae098d03f74ae6673211a303793332e98 --- /dev/null +++ b/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7dafd7c2e340d6868c0b76cce4481034446bef3c725b10d1bcd3b22aa53e0d67 +size 918609 diff --git a/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/layout.json b/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..79bec622b479e0ccd0d28c38c991c03ba6a63449 --- /dev/null +++ b/2025/TopicGeo_ An Efficient Unified Framework for Geolocation/layout.json @@ -0,0 +1,9273 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 103, + 489, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 103, + 489, + 121 + ], + "spans": [ + { + "bbox": [ + 121, + 103, + 489, + 121 + ], + "type": "text", + "content": "TopicGeo: An Efficient Unified Framework for Geolocation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 142, + 413, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 142, + 413, + 158 + ], + "spans": [ + { + "bbox": [ + 195, + 142, + 413, + 158 + ], + "type": "text", + "content": "Xin Wang Xinlin Wang* Shuiping Gou" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 158, + 383, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 158, + 383, + 172 + ], + "spans": [ + { + "bbox": [ + 228, + 158, + 383, + 172 + ], + "type": "text", + "content": "Xidian University, Xi'an, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 174, + 454, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 174, + 454, + 186 + ], + "spans": [ + { + "bbox": [ + 153, + 174, + 454, + 186 + ], + "type": "text", + "content": "{xinwangai@stu., wangxinlin@, shpgou@mail.}xidian.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 60, + 217, + 180, + 369 + ], + "blocks": [ + { + "bbox": [ + 60, + 217, + 180, + 369 + ], + "lines": [ + { + "bbox": [ + 60, + 217, + 180, + 369 + ], + "spans": [ + { + "bbox": [ + 60, + 217, + 180, + 369 + ], + "type": "image", + "image_path": "adbfe4d900636fe8e1d4f8c38d08d587188756ab66a5d1b971a461df590c7459.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 372, + 270, + 383 + ], + "lines": [ + { + "bbox": [ + 89, + 372, + 270, + 383 + ], + "spans": [ + { + "bbox": [ + 89, + 372, + 270, + 383 + ], + "type": "text", + "content": "(a) Low-texture area(980 5000×5000 base images)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 181, + 217, + 306, + 369 + ], + "blocks": [ + { + "bbox": [ + 181, + 217, + 306, + 369 + ], + "lines": [ + { + "bbox": [ + 181, + 217, + 306, + 369 + ], + "spans": [ + { + "bbox": [ + 181, + 217, + 306, + 369 + ], + "type": "image", + "image_path": "4c608631d97709b16dee6ee44c8426b5c480dd7f4ea07db3ac59ae324409703f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 217, + 369, + 369 + ], + "blocks": [ + { + "bbox": [ + 307, + 217, + 369, + 369 + ], + "lines": [ + { + "bbox": [ + 307, + 217, + 369, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 369, + 369 + ], + "type": "image", + "image_path": "e65a6d5df9f7ddbd40976f928198a49181734ea6e376e95f75b088baf1d082c7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 394, + 555, + 439 + ], + "lines": [ + { + "bbox": [ + 54, + 394, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 54, + 394, + 555, + 439 + ], + "type": "text", + "content": "Figure 1. The visual geolocation results of our TopicGeo, and typical retrieval and matching approaches: Earthloc [4], SelaVPR [29], and LoFTR [47]. \"Recall\" represents retrieval results. The color-coded lines show matching correctness: red for higher-error matches, green for higher-precision ones. In various scenes, our method accurately retrieves the target base image containing the query image from the database, while establishing pixel-level correspondence between the smaller query and large-scale retrieved image." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 369, + 217, + 430, + 369 + ], + "blocks": [ + { + "bbox": [ + 369, + 217, + 430, + 369 + ], + "lines": [ + { + "bbox": [ + 369, + 217, + 430, + 369 + ], + "spans": [ + { + "bbox": [ + 369, + 217, + 430, + 369 + ], + "type": "image", + "image_path": "eb80c29623c87844f81db9ee512708cbf4591304e737610ea841957514f5c5e4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 373, + 542, + 384 + ], + "lines": [ + { + "bbox": [ + 320, + 373, + 542, + 384 + ], + "spans": [ + { + "bbox": [ + 320, + 373, + 542, + 384 + ], + "type": "text", + "content": "(b) Multi-temporal change area (40 5120×5120 base images)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 430, + 217, + 490, + 369 + ], + "blocks": [ + { + "bbox": [ + 430, + 217, + 490, + 369 + ], + "lines": [ + { + "bbox": [ + 430, + 217, + 490, + 369 + ], + "spans": [ + { + "bbox": [ + 430, + 217, + 490, + 369 + ], + "type": "image", + "image_path": "7a0d42e5c664ef25bed7c7422f3aa395f48da4d8c7b84cce1accbf976d46112b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 491, + 217, + 551, + 369 + ], + "blocks": [ + { + "bbox": [ + 491, + 217, + 551, + 369 + ], + "lines": [ + { + "bbox": [ + 491, + 217, + 551, + 369 + ], + "spans": [ + { + "bbox": [ + 491, + 217, + 551, + 369 + ], + "type": "image", + "image_path": "c91051cdae460b609a9f69a7e434c113ee6d5fe5755a263ebcbd889a3f4d4151.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 152, + 449, + 200, + 461 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 449, + 200, + 461 + ], + "spans": [ + { + "bbox": [ + 152, + 449, + 200, + 461 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 475, + 297, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 475, + 297, + 691 + ], + "spans": [ + { + "bbox": [ + 54, + 475, + 297, + 691 + ], + "type": "text", + "content": "Vision-based geolocation techniques that establish spatial correspondences between smaller query images and larger georeferenced images have gained significant attention. Existing approaches typically employ a separate \"retrieve-then-match\" paradigm, whereas such paradigms suffer from computational inefficiency or precision limitations. To this end, we propose TopicGeo, a unified framework for direct and precise query-to-reference image matching via three key innovations. The textual object semantics, called topics, distilled from CLIP prompt learning are embedded into the geolocation framework to eliminate intra-class and inter-class distribution discrepancies while also enhancing processing efficiency. Center-based adaptive label assignment and outlier rejection mechanisms as a joint retrieval-matching optimization strategy ensure task-coherent feature learning and precise spatial correspondences. A multi-level fine matching pipeline is introduced to refine matching from quality and quantity. Evaluations on large-scale synthetic" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 450, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 498 + ], + "type": "text", + "content": "and real-world datasets illustrate that TopicGeo achieves state-of-the-art performance in retrieval recall and matching accuracy while maintaining a balance in computational efficiency." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 523, + 394, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 523, + 394, + 536 + ], + "spans": [ + { + "bbox": [ + 314, + 523, + 394, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 544, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 544, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 544, + 555, + 676 + ], + "type": "text", + "content": "Image matching through establishing spatial correspondence between localized query images and extensive georeferenced databases has emerged as a critical technique for vision-based geolocation, playing a vital role in environmental monitoring [13, 53], human activities[26, 54]. However, this task faces inherent challenges. Sensor variations and temporal differences induce substantial discrepancies [31] and nonlinear geometric deformations, complicating cross-image accurate matching. Moreover, repetitive low-saliency regions, such as forests, deserts, and water bodies in georeferenced images, pose matching ambiguities." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Recent advances in satellite image retrieval have focused on mitigating these challenges through global retrieval approaches [4, 24, 48]. These methods encode entire images" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 149, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 149, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 149, + 712 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8241" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 346 + ], + "type": "text", + "content": "as high-dimensional vectors for rapid retrieval of similar images from a database. However, such methods exhibit certain critical limitations. Global feature aggregation dilutes spatially fine-grained discriminative patterns, causing failures in low-saliency regions, such as uniform forests or barren land. In addition, when pairing large-size reference images, images are often cropped into smaller and partially overlapped pieces [4], which causes redundant computation. Local re-ranking methods [18, 29, 43] address these challenges by re-ranking candidates based on local matches. However, false positive candidates that are merely visually similar do not contribute to performance improvements. Retrieval-based methods exhibit limited geolocation accuracy, posing significant challenges for researchers in performing time-consuming and laborious manual corrections on multi-source imagery. A natural approach would involve first extracting candidate regions via a retrieval method, then refining for pixel-level matching with a specialized model [3, 31, 45]. However, a critical limitation of this approach lies in the decoupled retrieval and matching pipelines, which inherently prioritize different features in different tasks. In addition, the matching remains heavily dependent on the retrieval model's effectiveness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 350, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 295, + 590 + ], + "type": "text", + "content": "Recent studies [45, 49] apply a single end-to-end matching model for joint retrieval and matching. While enhanced by fine-grained features to boost accuracy, these methods face a huge time overhead [45], preventing real-time geolocation. The above geolocation methods rely on visual features for location estimation. However, remote sensing images exhibit a pronounced imbalance of areal distribution of terrestrial features, inducing inherent bias in visual information-dependent models toward overrepresented categories. Moreover, geometric distortions between multisource images and cross-temporal visual discrepancies lead to variations in the distribution of geographic features at the same location. Some methods [43, 55] learn more robust representations using natural images by incorporating semantic segmentation. In remote sensing, datasets [8, 39] and segmentation models [32, 33, 50] offer semantic information but face generalization limitations. Recent studies [15, 16] leverage self-supervised learning of image topics to enhance visual features. However, the resulting topic distributions often lack explicit semantic information." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "In this study, we propose a model that jointly optimizes retrieval and matching tasks for geographic positioning. Specifically, we employ asymmetric-input architecture to minimize redundant computation in large-displacement and large-scale scenarios. The core innovation lies in hierarchical textual embeddings for prompt learning and a topic-aware encoder trained via knowledge distillation and spatial alignment, yielding efficient topic features that are beneficial for both image retrieval and coarse matching. Building on this foundation, a multi-level refinement pipeline it-" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": "eratively resolves positional cues across scales, improving quality and quantity of matching points. Our unified framework is jointly optimized through effective label supervision and quality-aware rejection mechanisms. The contributions of the study are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 133, + 553, + 275 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 553, + 179 + ], + "type": "text", + "content": "- We propose a unified retrieval and matching framework with asymmetric large-scale input, boosted by embedded topics for efficiency and robustness to intra-class discrepancies." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 180, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 227 + ], + "type": "text", + "content": "- The center-based label assignment is designed to optimize the contrastive loss, which is further weighted through topic-based category perception to balance inter-class discrepancies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 228, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 275 + ], + "type": "text", + "content": "- The local spatial consensus filter and the global geometric verification module are co-designed to eliminate mismatched correspondences, thereby enhancing retrieval recall and matching precision in geolocation tasks." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 286, + 400, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 400, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 400, + 298 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 307, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 510 + ], + "type": "text", + "content": "Visual Language Model. The pioneering visual-language model (VLM) CLIP [38] employs vision and text encoders to project images and texts into a shared semantic space, which achieves joint embedding of images and texts, making cross-modal retrieval and zero-shot learning possible. Since then, VLMs have rapidly advanced and been applied across fields like open-vocabulary detection [6], image-text retrieval [30], and action recognition [52]. Menon et al. [34] apply large language models (LLMs) to represent object classes as semantic attribute sets for enhanced classification. Recently, Liu et al. present the first remote sensing visual-language foundation model, RemoteCLIP[25], learning semantically rich visual features aligned with text embeddings for seamless downstream applications. Our method leverages the textual semantics information of objects to capture invariant matching features across multi-source and multi-temporal satellite images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 713 + ], + "type": "text", + "content": "Image Retrieval. The image retrieval task focuses on identifying images similar to a query image within a database, encompassing two primary approaches: global and local feature-based retrieval. In global retrieval, the fundamental challenge lies in effectively aggregating image features from diverse regions into a cohesive high-dimensional vector. Traditional methods accomplish this by aggregating handcrafted local features [2, 5, 28] into global representations through methods such as Fisher Vectors [35], and then performing efficient retrieval through nearest neighbor search algorithms. Recent advancements leverage deep learning for feature aggregation [1, 37], employing strategies such as margin-based classification loss functions [9, 27] and clustering training samples [4]. However, global retrieval often results in a low recall rate during the initial retrieval phase. To address the issue, local matching is often integrated to re-rank retrieval results [18, 29, 43]. Geomet-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8242" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 69, + 553, + 277 + ], + "blocks": [ + { + "bbox": [ + 59, + 69, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 59, + 69, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 59, + 69, + 553, + 277 + ], + "type": "image", + "image_path": "4605b39d3a4324b481c3988c80be0dd0c08ee6670da32e6037edc0533f20b3bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 295, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 54, + 295, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 555, + 352 + ], + "type": "text", + "content": "Figure 2. Overview of TopicGeo. (a) Multi-level features are extracted from input images. (b) CLIP-generated prompt embeddings and visual features are injected into cross-attention layers and then distilled to infer the topic distribution. The extracted topic embeddings are integrated with global context features from self-attention layers. Preliminary matches are obtained using dual-softmax and local mutual k-nearest neighbor (MKNN) filter, followed by global geometric verification(GGV) for retrieval. (c) Adaptive matching is performed on cropped patches at a secondary resolution, with further coordinate refinement through fine-grained feature patch cropping." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 371, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 371, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 54, + 371, + 294, + 491 + ], + "type": "text", + "content": "ric verification methods like RANSAC [14] are employed to identify reliable candidates, though they frequently suffer from computational inefficiency. Fast verification approaches based on matching pair counts [29] face additional challenges in remote sensing applications, where weakly textured regions often lead to significant false positives. Unlike these retrieval-specific pipelines, this paper establishes a unified framework that seamlessly integrates retrieval and matching processes and introduces an efficient geometric verification mechanism." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": "Image Matching Image matching establishes accurate spatial correspondences between two images, essential for tasks like visual localization [21, 41] and 3D reconstruction [23]. The core of image matching is keypoint detection and description. Classical methods [2, 5, 28] employ hand-crafted features to detect and describe keypoints. Recent advances leverage deep learning for improved sparse keypoint detection[10, 40, 44], and self-supervised descriptor learning [17]. SuperGlue [42] harnesses the graph neural network with attentional mechanism to address significant occlusion or viewpoint changes, but relying on the quality of keypoints. RoMa [12] significantly enhances accuracy via Markov chain-based dense matching but incurs high computational cost. Detector-free approaches like LoFTR [47] use position embedding and cross-attention for semi-dense matching, which has gained traction in remote sensing[3, 45, 49], yet struggle with large-scale or point-of-view transformation. AdaMatcher [20] proposes an adap" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 371, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 491 + ], + "type": "text", + "content": "tive label assignment strategy to solve this problem, but introduces inconsistent learning-driven global mismatch filtering. TopicFM [15] sets up learnable topic embeddings for topic perception, while the unknown topics diverge from human perception and fail in complex scenes. Most existing methods rely on cross-attention for feature interaction, which limits computational efficiency. In contrast, our approach leverages distilled topics for feature enhancement and a learning-free local filter for efficient correspondence refinement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 529, + 432, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 432, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 432, + 544 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": "Our goal is to retrieve a reliable base image from a non-overlapping image database using a small query image and then match the query image to the retrieved larger-scale base image. Let " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "R = \\{B_i | i \\in [1, m]\\}" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " be a retrieval database containing " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " high resolution base images, and " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " be a query image. We establish a unified retrieval and matching framework, which encodes " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " into features " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "F_R = \\{F_{B_i} | i \\in [1, m]\\}" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " storing as a feature database, and " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "F_A" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " is retrieved, and the corresponding target base image " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " is obtained. Finally, the fine-grained features of both " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": " are extracted to refine the matching and compute the homography " + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": ". Figure 2 gives an overview of our pipeline." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8243" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 228, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 228, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 228, + 85 + ], + "type": "text", + "content": "3.1. Topic-aware Feature Extraction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "type": "text", + "content": "This paper tackles distribution discrepancies of land objects by establishing an effective topic perception process that integrates primitive visual features with high-level text representations independent of vision. Firstly, we extract visual features of the query image " + }, + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "type": "text", + "content": " and the base image " + }, + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "type": "text", + "content": ". To be efficient and preserve sufficient information, both of them undergo processing through patch embedding [11]. Then the downsampled images are fed into a shared ResNet [19] to extract coarse-grained visual features, " + }, + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "type": "inline_equation", + "content": "(F_A^{1/64}, F_B^{1/64})" + }, + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "type": "text", + "content": ", and fine-grained visual features " + }, + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "type": "inline_equation", + "content": "(F_A^{1/16}, F_B^{1/16})" + }, + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "type": "inline_equation", + "content": "(F_A^{1/4}, F_B^{1/4})" + }, + { + "bbox": [ + 55, + 89, + 296, + 225 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 226, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 346 + ], + "type": "text", + "content": "Hierarchical topic prompts. To accelerate convergence and integrate richer information, we design dual-level prompts for land objects. Specifically, generic and stable categories are extracted from established land cover datasets [50]. These categories are enriched through object attribute prompts generated by the GPT-4 language model. Subsequently, the embeddings of the augmented object prompts are encoded using the RemoteCLIP[25] text encoder. This process generates category embeddings " + }, + { + "bbox": [ + 55, + 226, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{category}} \\in \\mathbb{R}^{C \\times d}" + }, + { + "bbox": [ + 55, + 226, + 296, + 346 + ], + "type": "text", + "content": " and attribute embeddings " + }, + { + "bbox": [ + 55, + 226, + 296, + 346 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{attribute}} \\in \\mathbb{R}^{C' \\times d}" + }, + { + "bbox": [ + 55, + 226, + 296, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 347, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 445 + ], + "type": "text", + "content": "Object topic distillation. The obtained attribute and category prompt embeddings are embedded into visual features in sequence via cross-attention layers, where visual features serve as queries while these prompt embeddings act as keys and values. This process synthesizes the fused textual and visual representation to generate topic embeddings " + }, + { + "bbox": [ + 55, + 347, + 296, + 445 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{topic}}^{1/64}" + }, + { + "bbox": [ + 55, + 347, + 296, + 445 + ], + "type": "text", + "content": ". Notably, the sparsity of textual semantic tokens drastically lowers the computational cost of cross-attention." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 553 + ], + "type": "text", + "content": "However, such embeddings fail to align category-specific and attribute-related information with their corresponding visual semantics. Leveraging the inherent alignment between visual and textual features in RemoteCLIP, we employ its frozen vision encoder to further extract the visual embeddings " + }, + { + "bbox": [ + 55, + 445, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{vision}} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 55, + 445, + 296, + 553 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 445, + 296, + 553 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 445, + 296, + 553 + ], + "type": "text", + "content": " patches from " + }, + { + "bbox": [ + 55, + 445, + 296, + 553 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 445, + 296, + 553 + ], + "type": "text", + "content": ". Specifically, the patch-level category perception is established by computing the category probability distribution " + }, + { + "bbox": [ + 55, + 445, + 296, + 553 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 55, + 445, + 296, + 553 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 562, + 296, + 595 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 562, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 104, + 562, + 296, + 595 + ], + "type": "interline_equation", + "content": "p = \\operatorname {s o f t m a x} \\left(\\frac {\\mathbf {F} _ {\\text {v i s i o n}} \\mathbf {F} _ {\\text {c a t e g o r y}} ^ {T}}{t}\\right), \\tag {1}", + "image_path": "683d40cb40b5327dda6bf03072edba21258121a1a7f9ef48abe700697dd1dc21.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 604, + 208, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 208, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 208, + 615 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 604, + 208, + 615 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 604, + 208, + 615 + ], + "type": "text", + "content": " is the temperature parameter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": "To transfer the category-specific information from " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{vision}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " to topic embedding " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{topic}}^{1/64}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", the latter are spatially aligned with " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{vision}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " pooling, resulting in resized topic embedding " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{topic}}^{\\mathrm{p}} \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ". The corresponding category probability distribution for the pooled topic embedding is computed analogously to " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": ". Thus, the weighted Kullback-Leibler (KL) divergence loss " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "Loss_{\\mathrm{category}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " is employed to transfer the category-specific in-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "type": "text", + "content": "formation in " + }, + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{vision}}" + }, + { + "bbox": [ + 313, + 72, + 537, + 85 + ], + "type": "text", + "content": " to topic embeddings, expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 91, + 555, + 112 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 91, + 555, + 112 + ], + "spans": [ + { + "bbox": [ + 320, + 91, + 555, + 112 + ], + "type": "interline_equation", + "content": "L o s s _ {\\text {c a t e g o r y}} = \\sum_ {i = 0} ^ {N} \\left(\\sum_ {j = 0} ^ {C} \\left(p _ {i j} \\cdot \\log \\left(\\frac {p _ {i j}}{q _ {i j}}\\right)\\right) \\cdot w _ {i}\\right) \\cdot t ^ {2}, \\tag {2}", + "image_path": "60c066ef5b4da3951e2862cb18baf3f9dc34f0bd1ba1199a8c31797bb22d13ba.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 118, + 555, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 555, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 555, + 177 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 118, + 555, + 177 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 313, + 118, + 555, + 177 + ], + "type": "text", + "content": ", the normalized class frequency weight of the " + }, + { + "bbox": [ + 313, + 118, + 555, + 177 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 118, + 555, + 177 + ], + "type": "text", + "content": "-th image patch, is derived from " + }, + { + "bbox": [ + 313, + 118, + 555, + 177 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 118, + 555, + 177 + ], + "type": "text", + "content": ". Analogously, attribute-related information is transferred using " + }, + { + "bbox": [ + 313, + 118, + 555, + 177 + ], + "type": "inline_equation", + "content": "Loss_{attribute}" + }, + { + "bbox": [ + 313, + 118, + 555, + 177 + ], + "type": "text", + "content": ". Then the topic embeddings are enriched by minimizing the combined loss:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 367, + 184, + 555, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 184, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 367, + 184, + 555, + 198 + ], + "type": "interline_equation", + "content": "L _ {d} = \\text {L o s s} _ {\\text {c a t e g o r y}} + \\text {L o s s} _ {\\text {a t t r i b u t e .}} \\tag {3}", + "image_path": "01a20d2baab2de5cadade86575e2cb98a639fcd6b613fe1bb289c7ec982923e9.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 205, + 555, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 229 + ], + "type": "text", + "content": "Note that distillation operates exclusively on query images due to the fully shared parameters and training efficiency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 229, + 555, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 339 + ], + "type": "text", + "content": "Topic feature integration. To capture global context absent from " + }, + { + "bbox": [ + 313, + 229, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{topic}}" + }, + { + "bbox": [ + 313, + 229, + 555, + 339 + ], + "type": "text", + "content": ", the linear self-attention[51] with rotary position embedding[46] is further employed. Thus, the outputs from parallel encoding branches are integrated via a multilayer perceptron (MLP), with features from one branch stochastically masked during optimization to mitigate MLP-induced training bias. The final topic features, denoted as " + }, + { + "bbox": [ + 313, + 229, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\hat{F}_A^{1/64}" + }, + { + "bbox": [ + 313, + 229, + 555, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 229, + 555, + 339 + ], + "type": "inline_equation", + "content": "\\hat{F}_B^{1/64}" + }, + { + "bbox": [ + 313, + 229, + 555, + 339 + ], + "type": "text", + "content": ", are utilized for downstream retrieval and matching." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 345, + 476, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 476, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 476, + 358 + ], + "type": "text", + "content": "3.2. Retrieval & Coarse Matching" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 362, + 555, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 423 + ], + "type": "text", + "content": "The multimodal topic features of query and base images, " + }, + { + "bbox": [ + 313, + 362, + 555, + 423 + ], + "type": "inline_equation", + "content": "\\hat{F}_A^{1/64}" + }, + { + "bbox": [ + 313, + 362, + 555, + 423 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 362, + 555, + 423 + ], + "type": "inline_equation", + "content": "\\hat{F}_B^{1/64}" + }, + { + "bbox": [ + 313, + 362, + 555, + 423 + ], + "type": "text", + "content": ", enable efficient retrieval and coarse matching. Inspired by [20], the coarse-level matching probability matrix " + }, + { + "bbox": [ + 313, + 362, + 555, + 423 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 362, + 555, + 423 + ], + "type": "text", + "content": " and the bidirectional matching probability matrixes " + }, + { + "bbox": [ + 313, + 362, + 555, + 423 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 362, + 555, + 423 + ], + "type": "text", + "content": " are calculated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 373, + 430, + 555, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 430, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 373, + 430, + 555, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {S} (i, j) = \\left\\langle \\hat {F} _ {1 / 6 4} ^ {A} (i), \\hat {F} _ {1 / 6 4} ^ {B} (j) \\right\\rangle , \\tag {4}", + "image_path": "df9241abeb2dfa94dba948ea620e79584263f973e0fbf487630ef611170b836f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 451, + 555, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 451, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 359, + 451, + 555, + 467 + ], + "type": "interline_equation", + "content": "\\mathcal {P} ^ {A B} (i, j) = \\operatorname {S o f t m a x} \\left(\\mathcal {S} (i, \\cdot)\\right) _ {j}, \\tag {5}", + "image_path": "eac53cdcdb80ff04756fa72bf54c16de410b8347e8e766316e70fe71395cfdd5.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 360, + 468, + 555, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 468, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 360, + 468, + 555, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {P} ^ {B A} (i, j) = \\operatorname {S o f t m a x} \\left(\\mathcal {S} (\\cdot , j)\\right) _ {i}, \\tag {6}", + "image_path": "9a33ec0db63ddd56ef7bab5666e99879854139d38699dba70c1f922d1f995206.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\langle .,.\\rangle" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": " is the inner product. Based on " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{AB}" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": ", an index pair " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "(\\tilde{i},\\tilde{j})" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": " is selected as the match if its corresponding confidence value exceeds the threshold " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": " and is the maximum in its row. Thus, matches " + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{AB}" + }, + { + "bbox": [ + 313, + 491, + 554, + 540 + ], + "type": "text", + "content": " is obtained as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 548, + 555, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 548, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 320, + 548, + 555, + 562 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {A B} = \\left\\{\\left(\\tilde {i}, \\tilde {j}\\right) \\mid \\mathcal {P} ^ {A B} (\\tilde {i}, \\tilde {j}) = \\max _ {k} \\mathcal {P} ^ {A B} (\\tilde {i}, k), \\mathcal {P} ^ {A B} \\geq \\theta \\right\\}. \\tag {7}", + "image_path": "06a1728ff5b17540bced6af0b0b248cf966c9e593bfbfdceb99fe0ff016a8974.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": "Analogously, " + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{BA}" + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": " is derived by applying identical thresholding and column-wise maximization operations to " + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{BA}" + }, + { + "bbox": [ + 313, + 570, + 554, + 605 + ], + "type": "text", + "content": ", completing the bidirectional coarse matching process." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Local MKNN filter. While LoFTR [47] employs mutual nearest neighbor (MNN) filtering to suppress local mismatches, this strategy exhibits inherent limitations in one-to-many mismatches prevalent in resolution-scaled discrepancies. Instead, we propose mutual " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "-nearest neighbor (MKNN) filtering, which relaxes mismatches constraints by retaining matches that satisfy " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "-neighbor spatial consistency across bidirectional correspondence sets " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{AB}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{BA}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": ". Specifically, a candidate match pair " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "(i,j) \\in \\mathcal{M}_{AB}" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8244" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "is preserved if there exists a reverse-matching pair " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "(i', j) \\in \\mathcal{M}_{BA}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", and the 2D coordinate pair " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "(x', y')" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " corresponding to points " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "i'" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "-nearest neighbors. Analogously, " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{BA}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " is handled in the same manner. Finally, the refined coarse matches are combined using the OR operation to obtain " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", and the overall process is formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 161, + 295, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 161, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 110, + 161, + 295, + 174 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {c} = \\operatorname {M K N N} \\left(\\mathcal {M} _ {A B}, \\mathcal {M} _ {B A}\\right). \\tag {8}", + "image_path": "e0e2e4bde2b790f0ef6442e062b70f839f0b3d39960dc3cf75b5a0c3a27da3c9.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "text", + "content": "Global geometric verification. It is known that correct correspondences have geometric spatial consistency. Therefore, we utilize the rule to filter false positive samples in the database efficiently. Inspired by [36], we independently analyze distance and angle compatibility for each correspondence pair " + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c(n)" + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "text", + "content": " among the " + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "text", + "content": " total pairs. For distance compatibility, we calculate pairwise Euclidean distances between each point and all other " + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "inline_equation", + "content": "N_{c} - 1" + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "text", + "content": " points in both the query and base image coordinate systems, calculating cross-coordinate distance ratios and normalizing it. For angle compatibility, we evaluate the local geometric consistency by computing relative angles among each point's " + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "text", + "content": "-nearest neighbors in both coordinate systems, and measuring their cross-coordinate absolute angular difference " + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "inline_equation", + "content": "c_{n}" + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "text", + "content": ". Finally, the distance compatibility " + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "inline_equation", + "content": "\\alpha_{n}" + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "text", + "content": " and angle compatibility " + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "inline_equation", + "content": "\\beta_{n}" + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "text", + "content": " of each matching pair " + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c(n)" + }, + { + "bbox": [ + 55, + 178, + 296, + 382 + ], + "type": "text", + "content": " are yielded to jointly quantify spatial coherence, defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 386, + 295, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 386, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 72, + 386, + 295, + 413 + ], + "type": "interline_equation", + "content": "\\alpha_ {n} = \\max \\left(0, \\min \\left(Z - \\frac {\\operatorname {V a r} \\left(d _ {n}\\right)}{\\delta_ {d} {} ^ {2}}, Z - 1\\right)\\right), \\tag {9}", + "image_path": "61ace4cc42228cf0a3d20af363acbc419704772288096719f6b8593456d471e2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 425, + 295, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 425, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 79, + 425, + 295, + 452 + ], + "type": "interline_equation", + "content": "\\beta_ {n} = \\max \\left(0, \\min \\left(Z - \\frac {c _ {n} {} ^ {2}}{\\delta_ {c} {} ^ {2}}, Z - 1\\right)\\right), \\tag {10}", + "image_path": "1f7243d491b67511ef36309e5c5f8d2e74a1726903aedbd60176d544b93a22a8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "inline_equation", + "content": "\\delta_{d}" + }, + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "inline_equation", + "content": "\\delta_{c}" + }, + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "text", + "content": " are the maximum acceptable values, and " + }, + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "text", + "content": " is a scalar, used to truncate outliers. The spatial coherence confidences weighted by " + }, + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "text", + "content": " is utilized to re-weight the per-pair bidirectional matching probabilities " + }, + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "inline_equation", + "content": "E_{n}" + }, + { + "bbox": [ + 55, + 455, + 296, + 504 + ], + "type": "text", + "content": ", expressed as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 517, + 294, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 517, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 89, + 517, + 294, + 548 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E _ {n} = \\left(\\lambda \\beta_ {n} + (1 - \\lambda) \\alpha_ {n}\\right) \\\\ \\times \\left(P ^ {A B} \\left(\\tilde {i} _ {n}, \\tilde {j} _ {n}\\right) + P ^ {B A} \\left(\\tilde {i} _ {n}, \\tilde {j} _ {n}\\right)\\right). \\tag {11} \\\\ \\end{array}", + "image_path": "3e27e2c32a458d40d081d8626dfcf98b5a7bcc1337dfb86b12d9f40179b22bdd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 551, + 296, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 296, + 599 + ], + "type": "text", + "content": "The global confidence " + }, + { + "bbox": [ + 55, + 551, + 296, + 599 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 551, + 296, + 599 + ], + "type": "text", + "content": " for the base image is then obtained by aggregating all individual matching probabilities. Finally, the base image only with the highest overall confidence score is selected as the target retrieval image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 599, + 268, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 599, + 268, + 611 + ], + "spans": [ + { + "bbox": [ + 67, + 599, + 268, + 611 + ], + "type": "text", + "content": "The joint retrieval-matching loss " + }, + { + "bbox": [ + 67, + 599, + 268, + 611 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_c" + }, + { + "bbox": [ + 67, + 599, + 268, + 611 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 615, + 295, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 615, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 64, + 615, + 295, + 634 + ], + "type": "interline_equation", + "content": "\\mathrm {L} _ {c} = \\left(\\operatorname {F L} \\left(\\mathcal {P} ^ {A B}, \\hat {\\mathcal {P}} ^ {A B}\\right) + \\operatorname {F L} \\left(\\mathcal {P} ^ {B A}, \\hat {\\mathcal {P}} ^ {B A}\\right)\\right) \\times w _ {\\text {t o k e n}}, \\tag {12}", + "image_path": "65744da5b4844ebc6b251d01608a4662baf190191883152ab85d5b2f23087fb6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}" + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "text", + "content": " is the ground truth matching matrix. We allow one-to-many and many-to-one label assignments [20] for " + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{AB}" + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{BA}" + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "text", + "content": ". FL refers to the Focal Loss [22]. " + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "inline_equation", + "content": "w_{token}" + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "text", + "content": " is derived from the copy of " + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "text", + "content": " in the early training period, and the weight calculated by the category perception at the " + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{topic}}^{1/64}" + }, + { + "bbox": [ + 55, + 639, + 296, + 714 + ], + "type": "text", + "content": " is used in the subsequent training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 472, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 472, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 472, + 85 + ], + "type": "text", + "content": "3.3. Adaptive Center Assignment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 89, + 555, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 245 + ], + "type": "text", + "content": "Downsampling operations inherently yield sparse matching pairs, exacerbated by resolution-scaled discrepancies and asymmetric input. The adaptive label assignment [20] and MKNN filter alleviate the issue, while its reliance on the upper-left coordinates of the warped grid [42, 47] for label assignment often induces suboptimal matching patch overlaps, especially under large rotational misalignments in remote sensing images. This paper proposes a center-based label adaptive assignment strategy. Labels are assigned by distorting grid center coordinates instead of upper-left points, which are later reconverted to the upper-left coordinates for feature patch indexing, thereby enhancing robustness in challenging conditions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 251, + 460, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 460, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 460, + 266 + ], + "type": "text", + "content": "3.4. Multi-level Fine Matching" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "text", + "content": "The retrieval-oriented low-resolution encoding fundamentally limits matching precision. To achieve enhanced correspondence density through operationally consistent processing with the coarse matching stage, we employ the same adaptive local window matching on the finer-grained feature patches " + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "inline_equation", + "content": "(\\hat{F}_A^{1/16}, \\hat{F}_B^{1/16})" + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_c" + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "text", + "content": ", thereby obtaining " + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_a" + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "text", + "content": ". Subsequently, following LoFTR[47], higher-resolution feature patches " + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "inline_equation", + "content": "(\\hat{F}_A^{1/4}, \\hat{F}_B^{1/4})" + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "text", + "content": " extracted from " + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "inline_equation", + "content": "(F_A^{1/4}, F_B^{1/4})" + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "text", + "content": " guided by " + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_a" + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "text", + "content": " are utilized for final expectation-based fine-grained matching, obtaining " + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_f" + }, + { + "bbox": [ + 313, + 269, + 555, + 419 + ], + "type": "text", + "content": ", achieving sub-pixel matching accuracy. For adaptive local window matching, the loss is formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 323, + 427, + 553, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 427, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 323, + 427, + 553, + 445 + ], + "type": "interline_equation", + "content": "\\left. \\operatorname {L o s s} _ {a} = \\frac {1}{| \\mathcal {M} _ {a} |} \\left(\\operatorname {F L} \\left(\\mathcal {P} _ {a} ^ {A B}, \\hat {\\mathcal {P}} _ {a} ^ {A B}\\right) + \\operatorname {F L} \\left(\\mathcal {P} _ {a} ^ {B A}, \\hat {\\mathcal {P}} _ {a} ^ {B A}\\right)\\right), \\right. \\tag {13}", + "image_path": "0dc1ec36d5663683f68c14649979dbfcea06b7f05a7c8f43545f64d4dfd0fd41.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{P}}_a" + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": " represents the ground truth matching matrix. The expectation-based loss " + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "inline_equation", + "content": "\\mathrm{Loss}_f" + }, + { + "bbox": [ + 313, + 455, + 554, + 491 + ], + "type": "text", + "content": " refers to [47]. Thus, the final fine-grained loss is the sum of both:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 388, + 501, + 553, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 501, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 388, + 501, + 553, + 514 + ], + "type": "interline_equation", + "content": "\\mathrm {L} _ {f} = \\operatorname {L o s s} _ {a} + \\operatorname {L o s s} _ {f}. \\tag {14}", + "image_path": "8e43e42b38452f7cba35846978807d3db06dc75e07ff8a06e25a215cc4d63bb8.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 521, + 378, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 521, + 378, + 534 + ], + "spans": [ + { + "bbox": [ + 314, + 521, + 378, + 534 + ], + "type": "text", + "content": "3.5. Training" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 539, + 554, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 554, + 576 + ], + "type": "text", + "content": "The model is trained by optimizing the topic distillation loss " + }, + { + "bbox": [ + 313, + 539, + 554, + 576 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_d" + }, + { + "bbox": [ + 313, + 539, + 554, + 576 + ], + "type": "text", + "content": ", the joint retrieval-coarse matching loss " + }, + { + "bbox": [ + 313, + 539, + 554, + 576 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_c" + }, + { + "bbox": [ + 313, + 539, + 554, + 576 + ], + "type": "text", + "content": ", and fine-grained matching loss " + }, + { + "bbox": [ + 313, + 539, + 554, + 576 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_f" + }, + { + "bbox": [ + 313, + 539, + 554, + 576 + ], + "type": "text", + "content": ", defined as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 392, + 584, + 553, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 584, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 392, + 584, + 553, + 597 + ], + "type": "interline_equation", + "content": "\\mathrm {L} = \\mathrm {L} _ {d} + \\mathrm {L} _ {c} + \\mathrm {L} _ {f}. \\tag {15}", + "image_path": "16e2225b14ed640e6839c9b4eb9fed6bb90abd60a80857d9dc7b58b47d6f3324.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 616, + 395, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 616, + 395, + 630 + ], + "spans": [ + { + "bbox": [ + 314, + 616, + 395, + 630 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 635, + 510, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 510, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 510, + 649 + ], + "type": "text", + "content": "4.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "content": "TZC dataset [7] consists of 980 base maps with a spatial resolution of " + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "content": ", each with a high resolution of " + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "inline_equation", + "content": "5000\\times 5000" + }, + { + "bbox": [ + 313, + 653, + 554, + 712 + ], + "type": "text", + "content": " pixels. This dataset captures various land cover types, including dense forests, barren land, water bodies, agricultural fields, and human settlements." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8245" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 102, + 553, + 293 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of retrieval approaches. Top 1, 5, 10 recall are shown on the TZC dataset, and top 1, 2, 3 recall are shown on the MTGL40-5 dataset. Re-ranking thresholds are set at 100 for TZC and 50 for MTGL40-5." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 102, + 553, + 293 + ], + "lines": [ + { + "bbox": [ + 58, + 102, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 553, + 293 + ], + "type": "table", + "html": "
Retrieval modeMethodTZCMTGL40-5
R@1R@5R@10Time(s)R@1R@2R@3Time(s)
Global retrievalPatch-NetVLAD63.163.163.40.39158.260.762.30.352
SelaVPR68.969.269.50.04666.766.967.40.029
EarthLoc70.471.672.40.08267.267.568.20.064
Local Re-rankingPatch-NetVLAD65.666.266.421.562.364.264.810.7
SelaVPR72.673.173.40.27470.171.572.60.133
SP-SuperGlue71.472.873.512.169.370.773.65.87
Local retrievalSelaVPR74.475.675.93.2 × 10273.373.675.511.0
LoFTR_Coarse81.683.284.48.7 × 10376.478.480.53.9 × 102
TopicFM_Coarse83.783.884.59.5 × 10381.483.184.54.3 × 102
AdaMatcher_Coarse87.289.191.59.7 × 10384.686.287.84.5 × 102
Asymmetric local retrievalOur(LoFTR)86.188.490.134.282.383.183.71.81
Our(TopicFM)88.690.292.142.486.488.390.02.17
Our90.693.293.71.07191.693.193.90.165
", + "image_path": "894ecb32145b41d3f4597755e43ef975db65681b3d9093b99d3cd09e7a92f129.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 312, + 294, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 312, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 54, + 312, + 294, + 386 + ], + "type": "text", + "content": "MTGL40-5 dataset [31] focuses on 20 carefully selected scenes of ports and airports. Each scene has been standardized to " + }, + { + "bbox": [ + 54, + 312, + 294, + 386 + ], + "type": "inline_equation", + "content": "5120 \\times 5120" + }, + { + "bbox": [ + 54, + 312, + 294, + 386 + ], + "type": "text", + "content": " pixels, with a spatial resolution of " + }, + { + "bbox": [ + 54, + 312, + 294, + 386 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 54, + 312, + 294, + 386 + ], + "type": "text", + "content": ", and comprises five temporally distinct images. The appendix comprehensively overviews both datasets, including the generated prompt texts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "text", + "content": "Parameter setting. We employ the Adam optimizer for model training, set the learning rate as " + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "inline_equation", + "content": "5.0 \\times 10^{-4}" + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "text", + "content": " and the batch size as 2. The threshold " + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "text", + "content": " is fixed at 0.2. The window size for Multi-level Fine Matching is 4. The spatial compatibility parameter " + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "text", + "content": " is 10. The focal loss parameters " + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "text", + "content": " are configured as 0.25 and 2, respectively. For the MTGL40-5 dataset, " + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "text", + "content": " is reduced to 1 to account for temporal variations. Additionally, during the category and attribute distillation, temperature parameters are respectively set as " + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "inline_equation", + "content": "1/5" + }, + { + "bbox": [ + 54, + 386, + 295, + 506 + ], + "type": "text", + "content": " and 1 to facilitate effective knowledge transfer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 515, + 185, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 185, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 185, + 528 + ], + "type": "text", + "content": "4.2. Retrieval Performance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": "Quantitative comparison. We evaluate retrieval performance of various state-of-the-art methods on TZC and MTGL40-5 datasets, as shown in Table 1. The global retrieval approaches, such as Patch-NetVLAD[18], SelaVPR[29], and EarthLoc[4], exhibit minimal latency but suffer from significantly low recall rates on both datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "Local re-ranking strategies built upon global retrieval improve performance, where SP-SuperGlue re-ranks EarthLoc candidates by matching SuperPoint [10] features through SuperGlue [42] matcher. In the case, SelaVPR has significant gains, achieving " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": " R@1 gains on TZC and MTGL40-5 datasets, respectively. Compared to similar methods, SelaVPR reduces computational overhead associated with geometric verification by leveraging matching-pair counts as similarity scores. However, these" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 312, + 555, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 337 + ], + "type": "text", + "content": "enhancements remain constrained by the limited positive sample coverage in global retrieval candidate sets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 338, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 338, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 312, + 338, + 555, + 540 + ], + "type": "text", + "content": "Moreover, we compare methods that rely solely on local features for retrieval. It is obvious that the local feature retrieval-based SelaVPR incurs a two-order-of-magnitude computational cost increase, primarily due to redundant symmetric processing, which expands 980 base images into approximately " + }, + { + "bbox": [ + 312, + 338, + 555, + 540 + ], + "type": "inline_equation", + "content": "10^{5}" + }, + { + "bbox": [ + 312, + 338, + 555, + 540 + ], + "type": "text", + "content": " database images. In contrast, our asymmetric-input architecture directly processes the 980 original images, significantly improving efficiency. Coarse matcher-based local retrieval approaches like LoFTR_Coarse [47], TopicFM_Coarse [15], and AdaMatcher_Coarse [20], outperform SelaVPR through fine-grained supervision but incur three-order-of-magnitude higher online retrieval latency. Furthermore, the coarse matching encoders of LoFTR and TopicFM are integrated into our framework to illustrate the effectiveness of the asymmetric-input architecture on enhancing performance and efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 542, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 542, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 312, + 542, + 556, + 651 + ], + "type": "text", + "content": "In addition, compared to the performance on both datasets, most methods underperform on the MTGL40-5 dataset. Since multi-source modality gaps in the MTGL40-5 dataset challenge robust features extraction, even though it contains significantly fewer base images than the TZC dataset. Conversely, our paradigm outperforms other methods on retrieval performance and efficiency on both datasets, which illustrates the effect of topic features on learning consistent and invariant representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 659, + 446, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 446, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 446, + 672 + ], + "type": "text", + "content": "4.3. Matching Performance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Quantitative comparison. Table 2 shows the matching performance and time of state-of-the-art methods. To avoid retrieval preference, positive samples from SelaVPR's lo" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8246" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 75, + 529, + 378 + ], + "blocks": [ + { + "bbox": [ + 83, + 75, + 529, + 378 + ], + "lines": [ + { + "bbox": [ + 83, + 75, + 529, + 378 + ], + "spans": [ + { + "bbox": [ + 83, + 75, + 529, + 378 + ], + "type": "image", + "image_path": "91ce6e5cb2abc04e2f423eaf14bc3f47b8e5c59d903079362052e88324b1b909.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 403, + 555, + 439 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 555, + 439 + ], + "type": "text", + "content": "Figure 3. A visualized comparison of the state-of-the-art approaches on TZC dataset. Our unified framework generates numerous accurate correspondences in both texture-scarce barren areas and texture-dense urban areas. For images with large resolution-scaled discrepancies, our approach matches regions that are easier to match although they are sparse." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 510, + 293, + 622 + ], + "blocks": [ + { + "bbox": [ + 55, + 456, + 295, + 501 + ], + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 501 + ], + "type": "text", + "content": "Table 2. Comparison of matching approaches. " + }, + { + "bbox": [ + 55, + 456, + 295, + 501 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 55, + 456, + 295, + 501 + ], + "type": "text", + "content": " represents the original matching, while " + }, + { + "bbox": [ + 55, + 456, + 295, + 501 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 55, + 456, + 295, + 501 + ], + "type": "text", + "content": " indicates four iterations of the EarthMatch protocol. IoU is used as an effective evaluation metric." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 510, + 293, + 622 + ], + "lines": [ + { + "bbox": [ + 56, + 510, + 293, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 293, + 622 + ], + "type": "table", + "html": "
CategoryMethodTZC\nn=1 n=4MTGL\nn=1 n=4Time(s)\nn=1 n=4
Detector-basedSIFT-SuperGlue82.4 89.570.6 78.60.24 2.81
SP-SuperGlue84.4 91.277.4 82.40.14 2.37
Dense matcherRoMa92.2 96.186.7 90.20.79 5.17
Detector\nfreeLoFTR85.7 90.377.3 84.10.12 2.18
AdaMatcher91.4 93.783.3 88.40.15 2.51
TopicFM89.2 92.478.5 84.50.13 2.32
Ours98.9 99.191.1 92.70.09 2.02
", + "image_path": "325ee7946fb417d0f99a16284f224169898c59624c2c85a6217805388700e908.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "content": "cal re-ranking are used for matching, while the EarthMatch [3] protocol is applied to iteratively optimize matching and compensate for the limited receptive field of other methods. It is obvious that our method achieves state-of-the-art performance on both protocols while significantly improving" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 458, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 589 + ], + "type": "text", + "content": "efficiency. The proposed asymmetric-input direct matching framework reduces computational time by 1/20 of the EarthMatch protocol with merely " + }, + { + "bbox": [ + 313, + 458, + 555, + 589 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 458, + 555, + 589 + ], + "type": "text", + "content": " accuracy loss on TZC dataset. Notably, while recent dense matcher RoMa [12] delivers competitive accuracy (" + }, + { + "bbox": [ + 313, + 458, + 555, + 589 + ], + "type": "inline_equation", + "content": "90.2\\%" + }, + { + "bbox": [ + 313, + 458, + 555, + 589 + ], + "type": "text", + "content": " on the MTGL dataset), it requires " + }, + { + "bbox": [ + 313, + 458, + 555, + 589 + ], + "type": "inline_equation", + "content": "2.5 \\times" + }, + { + "bbox": [ + 313, + 458, + 555, + 589 + ], + "type": "text", + "content": " longer for per match than our solution. The multi-temporal discrepancies in MTGL dataset pose greater challenges than TZC's synthetic data, where our matching method effectively handles heterogeneous differences through consistent semantics and adaptive matching mechanisms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Visual comparison. Figure 3 displays the matching results of the crucial first iteration under the EarthMatch protocol. Notably, our method searches over a broader range compared to others. It is seen that our method excels in texture-scarce (such as barren areas) and texture-dense (like urban regions) scenarios through adaptive refinement and relaxed matching constraints in MKNN. Under extreme scale variations, existing methods like LoFTR distribute numerous matching pairs on non-discriminative natural terrains, which leads to low matching rates and disruption of iter-" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8247" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 124, + 299, + 194 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 297, + 117 + ], + "type": "text", + "content": "Table 3. Ablation of topic-related cross-attention layers and weighted loss on TZC dataset. The baseline, C, and T are the self-attention encoder, the cross-attention layers in LoFTR, and the topic-related layers of TopicFM, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 124, + 299, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 124, + 299, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 124, + 299, + 194 + ], + "type": "table", + "html": "
BaseC/TCategoryAttributeWLossR@1Time(s)
✓82.40.87
✓86.1/88.634/42
✓85.80.94
✓✓88.11.07
✓✓✓90.61.07
", + "image_path": "0e00eea25299bf0919a2d249bd1c43ae8d1a07c8bff2d0b04ae58b5fb1cd723f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 258, + 296, + 346 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 296, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 251 + ], + "type": "text", + "content": "Table 4. Ablation of topic-independent components on TZC dataset. A baseline is used that incorporates the topic encoder, original adaptive label assignment [20], and expectation-based fine-grained matcher [47]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 258, + 296, + 346 + ], + "lines": [ + { + "bbox": [ + 56, + 258, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 296, + 346 + ], + "type": "table", + "html": "
BaseACAMKNNGGVMFMR@1IoUPairsTime(s)
✓(0,0)84.972.31550.94/0.07
✓(0,0)86.575.41710.94/0.07
✓(0,1)86.580.7171/1920.94/0.08
✓(0,1)✓90.685.1171/1921.07/0.08
✓(1,1)✓87.181.21921.07/0.08
✓(0,1)✓✓90.690.5171/8141.07/0.09
", + "image_path": "eb6540f2d045956f938c51a99da860992d3f10a53991338c0c70d1899f95c1c1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 402 + ], + "type": "text", + "content": "ative process. Whereas our topic-aware weighted loss encourages the model to match regions that are easier to match although they are sparse, thereby enhancing reliability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 409, + 149, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 149, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 149, + 422 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "text", + "content": "The effect of topics. Our method uses distilled topics to improve efficiency and performance. Therefore, we separately evaluate the effect of the category and attribute, and the derived category weighted loss on the TZC dataset, as illustrated in Table 3. The cross-image perception of LoFTR improves " + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "inline_equation", + "content": "R@1" + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "text", + "content": ", while TopicFM's self-supervised topic achieves " + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "text", + "content": " gains. Both encoders that require crossattention for information interaction cause " + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "inline_equation", + "content": "40\\times" + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "text", + "content": " slower online retrieval than the baseline (0.87s). Our category embeddings brings " + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "inline_equation", + "content": "3.4\\%" + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "text", + "content": " improvement, attribute embeddings adds " + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "text", + "content": ", and weight loss contributes " + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 54, + 426, + 295, + 641 + ], + "type": "text", + "content": ". Moreover, the lightweight parallel architecture avoids significant time consumption. Figure 4 further visually compares topic semantics of remote sensing land cover generated by our method and the self-supervised approach TopicFM. It is seen that our method's deterministic land-cover topic modeling better distinguishes terrain features than self-supervised approach, enhancing geolocation precision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "The effect of other modules. Our evaluation first examines adaptive center assignment (ACA) and MKNN modules, followed by global geometric verification (GGV) and multi-level fine matching (MFM). During testing, the retrieval and matching is in sequence: failed retrieval yields zero matching IoU. When GGV is disabled, match count" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 542, + 306 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 542, + 306 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 542, + 306 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 542, + 306 + ], + "type": "image", + "image_path": "28a710a3017d7830dcd9ad3e57b454b1f18bf524d8e2bb876acc9d89b3a3f540.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 325, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 381 + ], + "type": "text", + "content": "Figure 4. Visualization of category topics in remote images. Our human-perception-aligned topic distributions encourage multisource feature alignment for geospatial retrieval and matching, whereas self-supervised structural modeling struggles to adapt to complex remote sensing scenarios." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "text", + "content": "directly determines retrieval scores. The two values of MKNN correspond to the " + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "text", + "content": " value in retrieval and coarse matching. As Table 4 shows, ACA increases initial matching pairs by " + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "inline_equation", + "content": "10.3\\%" + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "text", + "content": " over the original adaptive top-left assignment. MKNN (" + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "text", + "content": ") further elevates this by " + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "inline_equation", + "content": "12.3\\%" + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "text", + "content": ". The GGV enhances retrieval accuracy by " + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "text", + "content": " with only an additional time consumption (0.13s). One-to-many with " + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "text", + "content": " during retrieval makes the GGV score unstable and the recall rate decreases. MFM amplifies matching pairs by " + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "text", + "content": " IoU improvement, while maintaining low time overhead. Crucially, correctly retrieved samples achieve " + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 313, + 403, + 556, + 572 + ], + "type": "text", + "content": " (90.5(matching)/90.6(retrieval)) average matching IoU, indicating the task-coherent feature learning of retrieval and matching tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 585, + 388, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 388, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 388, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "We introduced a unified asymmetric-input retrieval-matching framework for geolocation. By combining object semantic modeling with image retrieval and matching, our method learns a robust set of primary features that can be stored offline. Moreover, we design a center-based label adaptive assignment and a multi-level fine-grained local matching pipeline, optimizing the matching precision. Compared to state-of-the-art methods, the proposed approach is both efficient and interpretable." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8248" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 207, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 207, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 207, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 207, + 296, + 251 + ], + "type": "text", + "content": "[1] Artem Babenko and Victor Lempitsky. Aggregating local deep features for image retrieval. In Proceedings of the IEEE international conference on computer vision, pages 1269-1277, 2015. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 253, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 253, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 253, + 296, + 285 + ], + "type": "text", + "content": "[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer vision and image understanding, 110(3):346-359, 2008. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 287, + 295, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 287, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 62, + 287, + 295, + 352 + ], + "type": "text", + "content": "[3] Gabriele Berton, Gabriele Goletto, Gabriele Trivigno, Alex Stoken, Barbara Caputo, and Carlo Masone. Earthmatch: Iterative coregistration for fine-grained localization of astronaut photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4264-4274, 2024. 2, 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 296, + 409 + ], + "type": "text", + "content": "[4] Gabriele Berton, Alex Stoken, Barbara Caputo, and Carlo Masone. Earthloc: Astronaut photography localization by indexing earth from space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12754-12764, 2024. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 411, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 411, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 411, + 296, + 476 + ], + "type": "text", + "content": "[5] Michael Calonder, Vincent Lepetit, Christoph Strecha, and Pascal Fua. Brief: Binary robust independent elementary features. In Computer Vision-ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part IV 11, pages 778-792. Springer, 2010. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 478, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 296, + 533 + ], + "type": "text", + "content": "[6] Tianheng Cheng, Lin Song, Yixiao Ge, Wenyu Liu, Xinggang Wang, and Ying Shan. Yolo-world: Real-time open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16901-16911, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 534, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 295, + 555 + ], + "type": "text", + "content": "[7] \"Tianzhi Cup\" Artificial Intelligence Challenge Organizing Committee. zhihuidiqiu2024, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 557, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 601 + ], + "type": "text", + "content": "[8] Rodrigo Caye Daudt, Bertrand Le Saux, Alexandre Boulch, and Yann Gousseau. Multitask learning for large-scale semantic change detection. Computer Vision and Image Understanding, 187:102783, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "type": "text", + "content": "[9] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 712 + ], + "type": "text", + "content": "[10] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 3, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant No.62372358 and 62302355, the Fundamental Research Funds for the Central Universities under Grant No.XJSJ24071 and No.XJSJ24072, and the Xidian University Specially Funded Project for Interdisciplinary Exploration under Grant No. TZJH2024026." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 105 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 107, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 160 + ], + "type": "text", + "content": "[12] Johan Edstedt, Qiyu Sun, Georg Bokman, Mårten Wadenback, and Michael Felsberg. Roma: Robust dense feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19790-19800, 2024. 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 554, + 194 + ], + "type": "text", + "content": "[13] JR Elliott. Earth observation for the assessment of earthquake hazard, risk and disaster management. Surveys in geophysics, 41(6):1323-1354, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 195, + 554, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 554, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 554, + 239 + ], + "type": "text", + "content": "[14] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 240, + 554, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 554, + 282 + ], + "type": "text", + "content": "[15] Khang Truong Giang, Soohwan Song, and Sungho Jo. Top-icfm: Robust and interpretable topic-assisted feature matching. In Proceedings of the AAAI conference on artificial intelligence, pages 2447-2455, 2023. 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 284, + 554, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 554, + 326 + ], + "type": "text", + "content": "[16] Khang Truong Giang, Soohwan Song, and Sungho Jo. Top-icfm+: Boosting accuracy and efficiency of topic-assisted feature matching. IEEE Transactions on Image Processing, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 554, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 369 + ], + "type": "text", + "content": "[17] Pierre Gleize, Weiyao Wang, and Matt Feiszli. Silk: Simple learned keypoints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22499-22508, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 426 + ], + "type": "text", + "content": "[18] Stephen Hausler, Sourav Garg, Ming Xu, Michael Milford, and Tobias Fischer. Patch-netvlad: Multi-scale fusion of locally-global descriptors for place recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14141-14152, 2021. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 427, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 554, + 470 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 472, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 554, + 536 + ], + "type": "text", + "content": "[20] Dihe Huang, Ying Chen, Yong Liu, Jianlin Liu, Shang Xu, Wenlong Wu, Yikang Ding, Fan Tang, and Chengjie Wang. Adaptive assignment for geometry aware local feature matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5425-5434, 2023. 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 537, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 590 + ], + "type": "text", + "content": "[21] Wei Jiang, Eduard Trulls, Jan Hosang, Andrea Tagliasacchi, and Kwang Moo Yi. Cotr: Correspondence transformer for matching across images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6207-6217, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 593, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 554, + 635 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 637, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 637, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 637, + 555, + 690 + ], + "type": "text", + "content": "[23] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5987-5997, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[24] Chao Liu, Jingjing Ma, Xu Tang, Fang Liu, Xiangrong Zhang, and Licheng Jiao. Deep hash learning for remote" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8249" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 95 + ], + "type": "text", + "content": "sensing image retrieval. IEEE Transactions on Geoscience and Remote Sensing, 59(4):3420-3443, 2020. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[25] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[26] Ganchao Liu, Chao Li, Sihang Zhang, and Yuan Yuan. Vlmlf: Uav visual localization based on multi-source image feature learning. IEEE Transactions on Geoscience and Remote Sensing, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[27] Weiyang Liu, Yandong Wen, Zhiding Yu, Ming Li, Bhiksha Raj, and Le Song. Spherface: Deep hypersphere embedding for face recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 212-220, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "text", + "content": "[28] David G Lowe. Distinctive image features from scale-invariant keypoints. International journal of computer vision, 60:91-110, 2004. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "text", + "content": "[29] Feng Lu, Lijun Zhang, Xiangyuan Lan, Shuting Dong, Yaowei Wang, and Chun Yuan. Towards seamless adaptation of pre-trained models for visual place recognition. In The Twelfth International Conference on Learning Representations, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 409 + ], + "type": "text", + "content": "[30] Christian Lülf, Denis Mayr Lima Martins, Marcos Antonio Vaz Salles, Yongluan Zhou, and Fabian Gieseke. Clipbranches: Interactive fine-tuning for text-image retrieval. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2719-2723, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "text", + "content": "[31] Jingjing Ma, Shiji Pei, Yuqun Yang, Xu Tang, and Xiangrong Zhang. Mtgl40-5: A multi-temporal dataset for remote sensing image geo-localization. Remote Sensing, 15(17):4229, 2023. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "text", + "content": "[32] Xianping Ma, Xiaokang Zhang, and Man-On Pun. Rs 3 mamba: Visual state space model for remote sensing image semantic segmentation. IEEE Geoscience and Remote Sensing Letters, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "text", + "content": "[33] Xianping Ma, Xiaokang Zhang, Man-On Pun, and Ming Liu. A multilevel multimodal fusion transformer for remote sensing semantic segmentation. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "text", + "content": "[34] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. arXiv preprint arXiv:2210.07183, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[35] Florent Perronnin, Yan Liu, Jorge Sánchez, and Hervé Poirier. Large-scale image retrieval with compressed fisher vectors. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3384-3391. IEEE, 2010. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 689 + ], + "type": "text", + "content": "[36] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11143-11152, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[37] Filip Radenovic, Giorgos Tolias, and Ondrej Chum. Finetuning cnn image retrieval with no human annotation. IEEE" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "transactions on pattern analysis and machine intelligence, 41(7):1655-1668, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[39] Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and Robin Robertson Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. IEEE Access, 9:89644-89654, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[40] Jerome Revaud, Cesar De Souza, Martin Humenberger, and Philippe Weinzaepfel. R2d2: Reliable and repeatable detector and descriptor. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[41] Paul-Edouard Sarlin, Cesar Cadena, Roland Siegwart, and Marcin Dymczyk. From coarse to fine: Robust hierarchical localization at large scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12716–12725, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 321, + 555, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 555, + 376 + ], + "type": "text", + "content": "[42] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 377, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 442 + ], + "type": "text", + "content": "[43] Yanqing Shen, Sanping Zhou, Jingwen Fu, Ruotong Wang, Shitao Chen, and Nanning Zheng. Structvpr: Distill structural knowledge with weighting samples for visual place recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11217-11226, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 499 + ], + "type": "text", + "content": "[44] Edgar Simo-Serra, Eduard Trulls, Luis Ferraz, Iasonas Kokkinos, Pascal Fua, and Francesc Moreno-Noguer. Discriminative learning of deep convolutional feature point descriptors. In Proceedings of the IEEE international conference on computer vision, pages 118-126, 2015. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "type": "text", + "content": "[45] Alex Stoken and Kenton Fisher. Find my astronaut photo: Automated localization and georectification of astronaut photography. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6196-6205, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "text", + "content": "[46] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "type": "text", + "content": "[47] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8918-8927, 2021. 1, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[48] Xu Tang, Qiushuo Ma, Xiangrong Zhang, Fang Liu, Jingjing Ma, and Licheng Jiao. Attention consistent network for remote sensing scene classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14:2030-2045, 2021. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8250" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 407 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[49] Haoyang Wang, Fuhui Zhou, and Qihui Wu. Accurate vision-enabled uav location using feature-enhanced transformer-driven image matching. IEEE Transactions on Instrumentation and Measurement, 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 295, + 161 + ], + "type": "text", + "content": "[50] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. Loveda: A remote sensing land-cover dataset for domain adaptive semantic segmentation. arXiv preprint arXiv:2110.08733, 2021. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 57, + 163, + 295, + 206 + ], + "type": "text", + "content": "[51] Shuhang Wang, Jin Zheng, Hai-Miao Hu, and Bo Li. Naturalness preserved enhancement algorithm for non-uniform illumination images. IEEE transactions on image processing, 22(9):3538-3548, 2013. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 208, + 294, + 251 + ], + "type": "text", + "content": "[52] Xiang Wang, Shiwei Zhang, Jun Cen, Changxin Gao, Yingya Zhang, Deli Zhao, and Nong Sang. Clip-guided prototype modulating for few-shot action recognition. International Journal of Computer Vision, 132(6):1899-1912, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 254, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 254, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 57, + 254, + 294, + 306 + ], + "type": "text", + "content": "[53] Qinghua Ye, Yuzhe Wang, Lin Liu, Linan Guo, Xueqin Zhang, Liyun Dai, Limin Zhai, Yafan Hu, Nauman Ali, Xinhui Ji, et al. Remote sensing and modeling of the cryosphere in high mountain asia: A multidisciplinary review. Remote Sensing, 16(10):1709, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 308, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 308, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 57, + 308, + 294, + 361 + ], + "type": "text", + "content": "[54] Hongsheng Zhang and Ru Xu. Exploring the optimal integration levels between sar and optical data for better urban land cover mapping in the pearl river delta. International journal of applied earth observation and geoinformation, 64:87-95, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 58, + 364, + 294, + 407 + ], + "type": "text", + "content": "[55] Yesheng Zhang and Xu Zhao. Mesa: Matching everything by segmenting anything. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20217-20226, 2024. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "8251" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/a88fe374-66af-4512-b87e-ddbd06fc2c78_content_list.json b/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/a88fe374-66af-4512-b87e-ddbd06fc2c78_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..62fd11343028ab8dc4df8de77a33e39b41648024 --- /dev/null +++ b/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/a88fe374-66af-4512-b87e-ddbd06fc2c78_content_list.json @@ -0,0 +1,1591 @@ +[ + { + "type": "text", + "text": "TopoTTA: Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation", + "text_level": 1, + "bbox": [ + 210, + 130, + 785, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiale Zhou $^{1,2}$ , Wenhan Wang $^{3}$ , Shikun Li $^{2}$ , Xiaolei Qu $^{3}$ , Xin Guo $^{3}$ , Yizhong Liu $^{3}$ , Wenzhong Tang $^{3}$ , Xun Lin $^{2,3,*}$ , Yefeng Zheng $^{2,*}$ $^{1}$ Zhejiang University $^{2}$ Westlake University $^{3}$ Beihang University \n{zhoujiale, zhengyefeng}@westlake.edu.cn linxun@buaa.edu.cn", + "bbox": [ + 187, + 202, + 810, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 309, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tubular structure segmentation (TSS) is important for various applications, such as hemodynamic analysis and route navigation. Despite significant progress in TSS, domain shifts remain a major challenge, leading to performance degradation in unseen target domains. Unlike other segmentation tasks, TSS is more sensitive to domain shifts, as changes in topological structures can compromise segmentation integrity, and variations in local features distinguishing foreground from background (e.g., texture and contrast) may further disrupt topological continuity. To address these challenges, we propose Topology-enhanced Test-Time Adaptation (TopoTTA), the first test-time adaptation framework designed specifically for TSS. TopoTTA consists of two stages: Stage 1 adapts models to cross-domain topological discrepancies using the proposed Topological Meta Difference Convolutions (TopoMDCs), which enhance topological representation without altering pre-trained parameters; Stage 2 improves topological continuity by a novel Topology Hard sample Generation (TopoHG) strategy and prediction alignment on hard samples with pseudo-labels in the generated pseudobreaks. Extensive experiments across four scenarios and ten datasets demonstrate TopoTTA's effectiveness in handling topological distribution shifts, achieving an average improvement of $31.81\\%$ in clDice. TopoTTA also serves as a plug-and-play TTA solution for CNN-based TSS models.", + "bbox": [ + 86, + 340, + 483, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tubular structure segmentation (TSS), also known as curvilinear object segmentation [45], plays a critical role in various applications (e.g., hemodynamic analysis [24, 55] and route planning [64, 73]). Although many TSS methods have been proposed and achieved inspiring performance [15, 21], domain shifts remain a common challenge", + "bbox": [ + 89, + 786, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9a7bccbfc9c7800d0f0a7cd0e86281b4f7acb1540d22a575ef1fc4b99911dd4a.jpg", + "image_caption": [ + "Figure 1. Comparison between TopoTTA and existing TTA solutions. (a) represents the model during the source domain training phase, while (b) denotes the model in the target domain testing phase. (b1-b4) show the results of directly testing the source-trained model, the TENT [58] method with updated BN layers, the CoTTA [60] method using a teacher-student scheme, and our topology-enhanced approach, respectively." + ], + "image_footnote": [], + "bbox": [ + 535, + 308, + 880, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in real-world scenarios [42] due to discrepancies in imaging devices [1, 13] and sample heterogeneity [11].", + "bbox": [ + 511, + 657, + 906, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domain shifts undermine the effectiveness of TSS models trained on the source domain when applied to an unseen target domain [36, 50]. Unlike segmentation models trained for other tasks, e.g., semantic [35, 67] or tumor segmentation [9, 30, 31, 47, 77], TSS models are more sensitive to domain shifts [50]. Beyond the common variations such as contrast and noise distribution [62, 63], TSS faces additional challenges related to unseen topological structural characteristics, including variations in trajectory, curvature, branching patterns, and thickness [6, 46]. These factors degrade segmentation performance and even disrupt the topological continuity of the structures [65] (see Fig. 1 (b1)), which may mislead critical downstream applications [37, 55, 64].", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 109, + 887, + 241, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24123", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, a special domain adaptation (DA) paradigm, namely test-time adaptation (TTA) [27, 40, 43], has been proposed and successfully applied to various image segmentation tasks, including semantic [22, 60, 61] and medical image segmentation [3, 74]. Unlike other DA paradigms, TTA performs continuous online adaptation during the test stage (after model deployment) [71]. It uses extremely small amounts of data (e.g., a single sample or batch), aligning better with real-world scenarios where data must be processed sequentially [58].", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most TTA methods update pre-trained models using self-supervised methods [27], which can typically be categorized into normalization-based methods [58, 72] and teacher-student schemes [10, 60]. Normalization-based methods update normalization layer statistics [53] (e.g., batch/layer normalization (BN/LN)) or leverage the difference in normalization statistics before and after adaptation to update external model parameters [3, 49]. These methods are initially designed for classification tasks, and their performance may degrade when applied to segmentation tasks with severe pixel-level class imbalance [5]. The second category, teacher-student schemes (based on pseudolabels), updates all model parameters and achieves performance gains in general segmentation tasks [69]. However, as shown in Fig. 1, these methods perform poorly in TSS due to two unique challenges in cross-domain scenarios:", + "bbox": [ + 89, + 244, + 482, + 484 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Challenge 1: Discrepant topological structures. Tubular structures in the source and target domains can exhibit substantial topological differences (e.g., thickness, curvature, and branching patterns). Prior TTA methods' general-purpose model adaptation strategies struggle to effectively capture these cross-domain topological variations when adapting to individual samples.", + "- Challenge 2: Fragile topological continuity. In cross-domain settings, the local characteristics (e.g., color, texture, and contrast) distinguishing foreground from background may change significantly [50]. Existing TTA methods lack targeted adaptation to these local variations, resulting in catastrophic disruptions to topological continuity (see Fig. 1) by misclassifying foreground pixels." + ], + "bbox": [ + 91, + 488, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we propose Topology-Enhanced Test-Time Adaptation (TopoTTA) in this paper. It includes two adaptation stages: topological structure adaptation (Stage 1) and topological continuity refinement (Stage 2).", + "bbox": [ + 89, + 700, + 482, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Stage 1 aims to enhance the model's capability to adapt to discrepant topological structures by utilizing a topology-perceiving model adaptation strategy. Inspired by the Central Difference Convolution (CDC) [70], which captures differences between the central pixel and its neighbors to enhance local representations, we extend this concept to accommodate the elongated and diverse trajectories of tubular structures. Specifically, we introduce Topological Meta", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Difference Convolutions (TopoMDCs), which perform a dual-pixel difference operation by extending the central pixel's interaction across eight directional neighbors. As shown in Fig. 2, we learn weighted combinations of different directional TopoMDCs (instead of the vanilla convolution) for different regions to enhance the representations of hard-to-detect tubular branches, improving adaptability to discrepant topological structures.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Within Stage 2, we design a Topology Hard sample Generation (TopoHG) strategy to construct samples with pseudo-breaks. These pseudo-breaks mislead the models to make weakened predictions with poor topological continuity. We then emphasize the local consistency between the predictions and original pseudo-labels in these hard regions. This stage encourages the models to adapt the characteristics distinguishing foreground from background and helps address continuity destructions in the target domain.", + "bbox": [ + 511, + 212, + 903, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "bbox": [ + 531, + 349, + 751, + 363 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the first TTA framework for tubular structure segmentation, named TopoTTA, which effectively adapts to unseen domains by enhancing topological information.", + "- We design eight directional Topological Meta Difference Convolutions (TopoMDCs) within TopoTTA, and learn how to adaptively combine different TopoMDCs based on the topological features of the test sample. TopoMDCs enhance adaptability to unseen topological structures.", + "- We propose a Topology Hard sample Generation strategy for TopoTTA, namely TopoHG, which generates pseudobreaks at confidently predicted regions. We then align the predictions around these pseudo breaks to the original pseudo labels, improving the topological continuity in the unseen domain.", + "- Extensive experiments on four scenarios across ten datasets validate the effectiveness of TopoTTA in addressing topological distribution shifts, achieving an average performance improvement of $31.81\\%$ in cI Dice." + ], + "bbox": [ + 513, + 366, + 903, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 655, + 663, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Tubular Structure Segmentation", + "text_level": 1, + "bbox": [ + 511, + 681, + 799, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Tubular structures, characterized by their thin and long topological structures as well as complex layouts, usually cannot be accurately and continuously localized by general segmentation methods [59]. To tackle these challenges, researchers have incorporated topological priors into model designs, including specialized deformable convolution kernels [46], attention mechanisms [14, 66], and tailored model architectures [33]. Complementary to these module innovations, PointScatter [59] was introduced to represent tubular structures as point sets, addressing the flexibility limitations of masks constrained by fixed grids. DconnNet [68] employed a unique design to separate directional subspaces from the latent space, enhancing feature representation for", + "bbox": [ + 511, + 703, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24124", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "connectivity. Meanwhile, a growing body of works focused on novel loss functions [2, 17, 19, 20, 23, 52, 57] that explicitly optimize the topological continuity of segmentation results. More recently, some research has started to explore the use of synthetic data [25, 28, 51] and the integration of additional topological information [21, 63] to further improve performance. However, these TSS methods ignore the existence of domain shifts, resulting in suboptimal performance when applied to unseen domains.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 89, + 234, + 294, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "TTA adaptively adjusts the model at test time based on a small number of unlabeled samples from a different distribution. Early TTA methods primarily modified model representations by adjusting BN's statistics [38, 44, 61, 75]. Mirza et al. [38] adapted target domain features by updating BN statistics continuously. Most TTA methods used self-supervised learning to adjust pre-trained parameters, including normalization-based methods [3, 43, 53, 58, 72] and teacher-student (TS) schemes [10, 41, 56, 60]. Wang et al. [58] proposed TENT, an entropy minimization method that updates BN affine transformation parameters to improve confidence in the target domain. Motivated by TENT, some recent works achieved better performance [3, 72] by designing various strategies to fine-tune BN. Zhang et al. [72] incorporated source domain statistics and extended the entropy minimization loss to utilize target domain information better. More recently, Chen et al. [3] reduced generalization errors by aligning source and target statistics to update external model parameters. However, many TTA methods are designed for classification tasks, ignoring pixel-wise class imbalance. In segmentation tasks, where sample imbalance is more severe, simple statistical adjustments and minor parameter updates are easily dominated by the majority of background pixels, reducing segmentation performance [69]. The TS scheme, which applies consistency regularization to update all model parameters, has shown effectiveness in general segmentation tasks. CoTTA [60] used weight-averaged and augmentation-averaged predictions, further reducing error accumulation in continuous inference. However, these methods lack a focus on tubular structure and topological continuity, leading to performance decline and continuity destruction in TSS.", + "bbox": [ + 89, + 255, + 483, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed TopoTTA", + "text_level": 1, + "bbox": [ + 89, + 750, + 277, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 775, + 220, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathcal{D}^s = \\{(\\pmb{x}_i^s,\\pmb{y}_i^s)\\}_{i = 1}^{N^s}$ denote the labeled source domain dataset, where $\\pmb{x}_i^s\\in \\mathbb{R}^{H\\times W\\times C}$ is an image with height $H$ width $W$ , and channel number $C$ , and $\\pmb{y}_i^s\\in \\{0,1\\}^{H\\times W\\times C}$ denotes the corresponding ground-truth label. Given a TSS model $\\mathcal{F}(\\cdot ;\\theta^s)$ trained on $\\mathcal{D}^s$ :", + "bbox": [ + 89, + 796, + 482, + 873 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta^ {s} = \\underset {\\theta^ {s}} {\\arg \\min } \\mathbb {E} _ {\\left(\\boldsymbol {x} ^ {s}, \\boldsymbol {y} ^ {s}\\right) \\sim \\mathcal {D} ^ {s}} \\left[ \\mathcal {L} _ {s e g} \\left(\\mathcal {F} \\left(\\boldsymbol {x} ^ {s}; \\theta^ {s}\\right), \\boldsymbol {y} ^ {s}\\right) \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 877, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{seg}(\\cdot ,\\cdot)$ can be any TSS loss function. Let $\\mathcal{D}^t = \\{\\pmb {x}_i^t\\}_{i = 1}^{N^t}$ represent the unlabeled target domain dataset. During testing, TTA methods aim to update the model parameters for each input image $\\pmb{x}_i^t\\in \\mathcal{D}^t$ and produce segmentation outputs $\\hat{\\pmb{y}}_i^t = \\mathcal{F}(\\pmb{x}_i^t;\\theta_{i - 1}^t)$ as follows:", + "bbox": [ + 511, + 90, + 903, + 170 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {i} ^ {t} = \\underset {\\theta_ {i - 1} ^ {t}} {\\arg \\min } \\mathcal {L} _ {t t a} \\left(\\mathcal {F} \\left(\\cdot ; \\theta_ {i - 1} ^ {t}\\right), \\boldsymbol {x} _ {i} ^ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 176, + 903, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_i^t$ are the parameters of the TSS model $\\mathcal{F}(\\cdot ;\\cdot)$ at the $i$ -th iteration, and $\\mathcal{L}_{tta}(\\cdot ,\\cdot)$ denotes the self-supervised loss function for TTA. Before performing TTA, we set $\\theta_0^t = \\theta^s$ .", + "bbox": [ + 511, + 214, + 903, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To improve readability, we use the simplified notation $\\pmb{x}$ instead of $\\pmb{x}_i^t$ in the following sections and figures.", + "bbox": [ + 511, + 260, + 903, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview of TopoTTA", + "text_level": 1, + "bbox": [ + 511, + 297, + 720, + 314 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As discussed in Section 1, existing TTA methods cannot effectively segment tubular structures due to two major challenges: discrepant topological structures and fragile topological continuity. To solve these issues, we propose TopoTTA, which consists of two stages: Stage 1 performs topological structure adaptation to overcome the insufficient cross-domain topological structure adaptation; Stage 2 conducts topological continuity refinement to enhance the topological continuity of the final predictions. An overview of our TopoTTA is illustrated in Fig. 2. The details of Stage 1 and Stage 2 are described in Section 3.3 and Section 3.4, respectively.", + "bbox": [ + 511, + 320, + 906, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Stage 1: Topological Structure Adaptation", + "text_level": 1, + "bbox": [ + 511, + 510, + 874, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This stage aims to adapt the pre-trained TSS model to discrepant topological structures caused by domain shifts. To achieve this, we propose TopoMDCs based on diverse fundamental topological patterns and adaptively reweight and combine TopoMDCs according to the topological features of different regions. We introduce the details of TopoMDCs and the model parameter updating process below.", + "bbox": [ + 511, + 532, + 905, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Topology-Meta Difference Convolutions (TopoMDCs). TopoMDCs draw inspiration from central difference convolution (CDC) [70], which captures the difference between the central pixel of a convolution kernel and its neighbors to detect local gradient changes. However, the local topological shapes of tubular structures often exhibit directionality and continuity. CDC, initially designed for other fine-grained vision tasks (e.g., face anti-spoofing [29, 32, 70]), lacks the ability to perceive these two features of tubular structures, making them ineffective in representing the complex topologies and geometric constraints within local regions. Let $x_{\\mathrm{in}}$ denote the input feature map. The vanilla convolution $\\mathcal{C}_0$ and CDC $\\mathcal{C}_c$ can be formulated as:", + "bbox": [ + 511, + 641, + 906, + 838 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {C} _ {0} (r _ {x}, r _ {y}) = \\sum_ {(\\Delta r _ {x}, \\Delta r _ {y}) \\in \\mathcal {R}} w (\\Delta r _ {x}, \\Delta r _ {y}) \\cdot \\pmb {x} _ {\\mathrm {i n}} (r _ {x} - \\Delta r _ {x}, r _ {y} - \\Delta r _ {y}), \\\\ \\mathcal{C}_{c}(r_{x},r_{y}) = \\sum_{(\\Delta r_{x},\\Delta r_{y})\\in \\mathcal{R}}w(\\Delta r_{x},\\Delta r_{y})\\cdot \\boldsymbol{x}_{\\mathrm{in}}(r_{x},r_{y}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 842, + 879, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24125", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ae5dbb41cbc112e71d559124140fdb6b81e4f67bbf450942b769c405f6928c10.jpg", + "image_caption": [ + "(a) Stage 1: Topological Structure Adaptation" + ], + "image_footnote": [], + "bbox": [ + 98, + 104, + 645, + 243 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a3edc90b2f535911d2f66a84eb8fc64e8bc35d1903870c4d19d40b66a0f9b5d5.jpg", + "image_caption": [ + "(c) TopoHG (in Stage 2)" + ], + "image_footnote": [], + "bbox": [ + 661, + 106, + 895, + 375 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2a5c0669467da68bedad559d3d8434723bd693c1527fc979a4c1e64ae7a3d341.jpg", + "image_caption": [ + "(b) Stage 2: Topological Continuity Refinement" + ], + "image_footnote": [], + "bbox": [ + 101, + 263, + 434, + 367 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a73c430afdc39bcaa88c982d220176019963c7ab7b7548228fd78da11e7acd7d.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed TopoTTA. TopoTTA consists of two stages: (a) Stage 1: Topology structure adaptation with Topological Meta Difference Convolutions (TopoMDCs), and (b) Stage 2: Topological continuity refinement. (c) Topology Hard sample Generation (TopoHG) constructs challenging samples with local pseudo-breaks. CDC is short for central difference convolution." + ], + "image_footnote": [], + "bbox": [ + 441, + 250, + 895, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{R} = \\{(1,1),(0,1),\\dots ,(-1,0),(-1, - 1)\\}$ represents the local receptive field of a $3\\times 3$ convolution kernel, $w$ denotes the parameters of the vanilla convolution layer, and $(r_x,r_y)$ represents the relative position within the convolution operation centered on the current pixel of $x_{\\mathrm{in}}$", + "bbox": [ + 89, + 425, + 482, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To overcome this limitation of CDC, the proposed TopoMDCs perform a two-pixel differential operation by extending their reach to the eight neighboring directions beyond the center pixel, as shown in Fig. 2 (a). This design simulates various basic topological patterns, effectively accommodating the elongated and diverse trajectories of tubular structures and enhancing the model's capability to recognize various tubular objects. Let $\\mathbf{x}_{\\mathrm{in}} \\in \\mathbb{R}^{H_{\\mathrm{in}} \\times W_{\\mathrm{in}} \\times C_{\\mathrm{in}}}$ denote the input feature map, the pre-trained vanilla convolution be $\\mathcal{C}_0$ , and the set of TopoMDCs be $\\{\\mathcal{C}_1, \\mathcal{C}_2, \\dots, \\mathcal{C}_8\\}$ .", + "bbox": [ + 89, + 501, + 482, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taking $\\mathcal{C}_1$ , which extends toward the top-left pixel (see Fig. 2 (a)), as an example, it is defined as:", + "bbox": [ + 89, + 652, + 482, + 681 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {C} _ {1} \\left(r _ {x}, r _ {y}\\right) = \\mathcal {C} _ {c} \\left(r _ {x}, r _ {y}\\right) - \\sum w \\left(\\Delta r _ {x}, \\Delta r _ {y}\\right) \\cdot \\boldsymbol {x} _ {\\text {i n}} \\left(r _ {x}, r _ {y}\\right) \\\\ (\\Delta r _ {x}, \\Delta r _ {y}) \\in \\mathcal {R} _ {1} \\\\ + \\sum_ {(\\Delta r _ {x}, \\Delta r _ {y}) \\in \\mathcal {R} _ {1}} w (\\Delta r _ {x}, \\Delta r _ {y}) \\cdot \\boldsymbol {x} _ {\\text {i n}} (r _ {x} + 1, r _ {y} + 1), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 684, + 480, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{R}_1 = \\{(-1, - 1),(-1,0),(0, - 1)\\}$ is the local receptive field of $C_1$ . Due to space limitations, the formal definitions of other TopoMDCs are provided in Appendix.", + "bbox": [ + 89, + 750, + 482, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TopoMDCs do not introduce any additional parameters for any convolution layer but instead, adaptively transform the vanilla convolution in $\\mathcal{F}(\\cdot ;\\theta)$ . Specifically, we replace all $3\\times 3$ convolution layers in the encoder of $\\mathcal{F}(\\cdot ;\\theta)$ with TopoMDCs. All TopoMDCs inherit the parameters of the vanilla convolutions. To adapt to a variety of complex (even unseen) topologies, a dynamic combination of TopoMDCs", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is required. Given that the topologies of local regions tend to be quite similar, while the topologies of distant regions can differ significantly, we introduce different learnable parameter sets $\\delta$ with eight learnable router parameters for different image patches. To adapt to the respective topological structures in different regions, we patchify the input image into $n\\times n$ non-overlap patches, each with the size of $H_{\\mathrm{in}} / n\\times W_{\\mathrm{in}} / n$ . Taking the $j$ -th image patches $\\pmb{p}_j$ from $\\pmb{x}_{\\mathrm{in}}$ , the corresponding router parameter set $\\delta$ assigns external parameters to different TopoMDCs, enabling adaptive enhancement of the perception of topological structures within the patch. The output feature $\\hat{\\pmb{x}}_{\\mathrm{out}}$ is obtained as:", + "bbox": [ + 511, + 425, + 906, + 608 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {x}} _ {\\text {o u t}} = \\mathcal {C} _ {0} \\left(\\boldsymbol {x} _ {\\text {i n}}\\right) - \\sum_ {j = 1} ^ {n \\times n} \\delta_ {j} \\sum_ {i = 1} ^ {8} \\mathcal {C} _ {i} \\left(\\boldsymbol {x} _ {\\text {i n}} ^ {j}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 618, + 906, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{x}_{\\mathrm{in}}^{j}$ denotes the $j$ -th patch of $\\pmb{x}_{\\mathrm{in}}$ . The patchification operations $\\mathcal{C}_0$ and $\\mathcal{C}_i$ differ slightly in their inputs in Eq. (4) and Eq. (5). In Eq. (5), $\\mathcal{C}_0$ and $\\mathcal{C}_i$ denote the same computation applied uniformly to every pixel in $\\pmb{x}_{\\mathrm{in}}^{j}$ .", + "bbox": [ + 511, + 669, + 905, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Router Parameter $(\\delta)$ Updating in Stage 1. As mentioned above, we do not update the original parameters of the convolution kernel or any other parameters in the network but only update a small number of introduced router parameters $\\delta$ . This is not just a consideration for inference speed. Simultaneously adjusting both $\\delta$ and the model parameters increases the search-space complexity, which may lead to conflicts or interference among parameters. This, in turn, could even disrupt the original feature representation. To this end, we only adjust $\\delta$ to adapt the topological response within an external feature space, thereby reducing the risk", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24126", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of instability. For each incoming test sample $\\pmb{x}$ , all router parameters are reset to zero, so that the model can independently adapt to the topological structures of the new sample. We use the self-supervised entropy minimization (EM) loss [58] to update $\\delta$ :", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E M}} = - \\sum_ {j = 1} ^ {2} \\hat {\\mathbf {y}} _ {j} \\log (\\hat {\\mathbf {y}} _ {j}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 176, + 482, + 214 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\delta} = \\underset {\\theta^ {t}} {\\arg \\min } \\mathcal {L} _ {\\mathrm {E M}} (\\mathcal {F} (\\cdot ; \\theta^ {t}; \\boldsymbol {\\delta}), \\boldsymbol {x}).\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 217, + 388, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Stage 2: Topological Continuity Refinement", + "text_level": 1, + "bbox": [ + 89, + 248, + 464, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As previously discussed in Sec. 3.3, Stage 1 enhances the representation of topological structures in $\\mathcal{D}^t$ . In Stage 2, we guide the model to learn local characteristics in $\\mathcal{D}^t$ for foreground-background differentiation, thereby improving the topological continuity of segmentation. To achieve this, we generate a series of pseudo-breaks at key topological structures on the test image $\\pmb{x}$ , producing the edited image $\\pmb{x}'$ through the proposed Topology Hard sample Generation (TopoHG) strategy, as shown in Fig. 2 (c).", + "bbox": [ + 89, + 270, + 483, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Topology Hard Sample Generation (TopoHG). The TopoHG process for generating local hard samples can be roughly divided into the following three steps:", + "bbox": [ + 89, + 410, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 1: Key Point Selection. Since we apply consistency supervision on the predictions of local hard samples, the selected regions must have highly reliable pseudo-labels. To this end, we first select a set of points with confidence above the threshold $\\tau = 0.95$ from the teacher's predictions $\\hat{\\pmb{y}}'$ , denoted as $\\mathcal{P} = \\{(u,v) \\mid \\hat{\\pmb{y}}'(u,v) > \\tau\\}$ . Then, we randomly select $N_p$ points from $\\mathcal{P}$ as key points. Since prediction confidence varies across different test samples, using a fixed $N_p$ for all input samples is suboptimal. Therefore, we allow samples with higher confidence to select more key points and create more challenging regions to contribute more to the weight update: $N_p = k \\cdot |\\mathcal{P}|$ , where $k$ is a hyperparameter.", + "bbox": [ + 89, + 457, + 483, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 2: Sliding to Search. For each selected key point $(u_c, v_c)$ , let $x_p^{\\mathrm{fg}}$ denote the fixed foreground window centered at this point, with a size of $s \\times s$ . Then, a background sliding window $x_p^{\\mathrm{bg}}$ , of the same size as the foreground window, slides over adjacent non-overlapping regions, identifying window $x_p^{\\mathrm{bg},*}$ with the lowest pseudo-label confidence of all pixels. Meanwhile, to prevent $x_p^{\\mathrm{bg},*}$ from carrying too much foreground information, we set a lower threshold $\\tau^{\\mathrm{bg}}$ . If the sum of foreground confidence of all pixels in $x_p^{\\mathrm{bg},*}$ exceeds $\\tau^{\\mathrm{bg}}$ , the key point $(u_c, v_c)$ is discarded.", + "bbox": [ + 89, + 655, + 483, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 3: Frequency-based Pseudo-break Generation. TopoHG aims to create local pseudo-break topological structures within the critical regions. This requires TopoHG to simulate the characteristics of structures that seem to be broken while retaining essential foreground features. Inspired by [34], we adopt a low-frequency swap method", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "between background and foreground to preserve high-frequency information that carries critical foreground features. Let FFT and iFFT denote the Fast Fourier Transform (FFT) and inverse Fast Fourier Transform (iFFT) operations [7], respectively, with $m_{\\mathrm{low}}$ as the mask for the low-frequency regions that need to be swapped. Applying FFT to $x_p^{\\mathrm{fg}}$ and $x_p^{\\mathrm{bg},*}$ yields the frequency domain representations $f_p^{\\mathrm{fg}} = \\mathrm{FFT}(x_p^{\\mathrm{fg}})$ and $f_p^{\\mathrm{bg}} = \\mathrm{FFT}(x_p^{\\mathrm{bg},*})$ . After performing low-frequency swap on $f_p^{\\mathrm{fg}}$ and $f_p^{\\mathrm{bg}}$ and applying iFFT, we obtain the modified image $x_p^{\\mathrm{swap}}$ , as follows:", + "bbox": [ + 511, + 90, + 906, + 247 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {p} ^ {\\text {s w a p}} = \\mathrm {i F F T} \\left(\\boldsymbol {f} _ {p} ^ {\\mathrm {f g}} \\cdot \\left(1 - \\boldsymbol {m} _ {\\text {l o w}}\\right) + \\boldsymbol {f} _ {p} ^ {\\mathrm {b g}} \\cdot \\boldsymbol {m} _ {\\text {l o w}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 253, + 906, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To avoid additional risks associated with unnecessary background changes when emphasizing consistency, we update only update pixels in $\\pmb{x}_p^{\\mathrm{fg}}$ that correspond to the foreground in the pseudo-label $\\hat{\\pmb{y}}_p'$ . The pseudo-break patch $\\pmb{x}_p'$ can be defined as follows:", + "bbox": [ + 511, + 281, + 906, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {p} ^ {\\prime} = \\boldsymbol {x} _ {p} ^ {\\text {s w a p}} \\cdot \\hat {\\boldsymbol {y}} _ {p} ^ {\\prime} + \\boldsymbol {x} _ {p} ^ {\\mathrm {f g}} \\cdot (1 - \\hat {\\boldsymbol {y}} _ {p} ^ {\\prime}). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 363, + 906, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After completing the above four steps, we obtain a hard sample $\\pmb{x}^{\\prime}$ that contains multiple local pseudo-breaks, resulting in weakened predictions to assist in model updates.", + "bbox": [ + 511, + 387, + 905, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model Updating in Stage 2. To align the predictions $\\hat{\\pmb{y}}''$ of these local hard samples $\\pmb{x}'$ with the original pseudo-label $\\hat{\\pmb{y}}'$ , we apply consistency regularization as supervision, implemented through the TS scheme as shown in Fig. 2 (c). The cross-entropy loss $\\mathcal{L}_{\\mathrm{CE}}$ is used to enhance the consistency between the student's (i.e., prediction of the hard sample $\\hat{\\pmb{y}}''$ ) and teacher's (i.e., pseudo-label $\\hat{\\pmb{y}}'$ of original samples) predictions, updating $\\theta_i^t$ as follows:", + "bbox": [ + 511, + 435, + 906, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} = - \\sum_ {(u, v)} \\mathcal {W} (u, v) \\cdot \\left(\\hat {\\boldsymbol {y}} ^ {\\prime} \\log \\left(\\hat {\\boldsymbol {y}} ^ {\\prime \\prime}\\right) + \\hat {\\boldsymbol {y}} ^ {\\prime \\prime} \\log \\left(\\hat {\\boldsymbol {y}} ^ {\\prime}\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 563, + 905, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_{i}^{t} = \\operatorname *{arg min}_{\\substack{\\theta_{i - 1}^{t}\\\\ }}\\mathcal{L}_{\\mathrm{CE}}(\\mathcal{F}(:,\\theta_{i - 1}^{t};\\delta),\\boldsymbol {x}),\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 597, + 764, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{W}(u,v)$ is a weight map. The local pseudo-breaks should receive higher attention in consistency regularization; therefore, $\\mathcal{W}(u,v)$ is defined as follows:", + "bbox": [ + 511, + 632, + 905, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} (u, v) = \\left\\{ \\begin{array}{l l} 1 0 & \\text {i f} (u, v) \\in \\boldsymbol {x} _ {p} ^ {\\prime} \\cdot \\hat {\\boldsymbol {y}} _ {p} ^ {\\prime} \\\\ 1 & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 685, + 905, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The gradient is used solely to update the student, while the teacher's parameters are updated by an exponential moving average (EMA) of the student's parameters. Finally, after several steps of updates, the unmodified input $\\pmb{x}$ is fed into the student model to obtain the final prediction.", + "bbox": [ + 511, + 727, + 905, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 815, + 645, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct comparison experiments across four scenarios, i.e., retinal vessel segmentation, road extraction, microscopic neuronal segmentation, and retinal OCT-angiography vessel segmentation covering ten datasets.", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24127", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8432e023444e835ae796e4d8ac16947948b96e4494a222fd9375dc56bfe7fb40.jpg", + "table_caption": [ + "Table 1. Cross-domain testing results in four different scenarios, i.e., retinal vessel segmentation, road extraction, microscopic neuronal segmentation, and retinal OCT-angiography vessel segmentation. Source Only: Trained on the source domain and tested on the target domain directly. The best and second-best results in each column are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
MethodDRIVE → CHASEDRIVE → STARECHASE → DRIVECHASE → STAREDeepGlobe → MR
Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓
Source Only22.3818.2141.2548.05/106.3062.9558.5989.9548.77/105.2044.8954.9776.83
TENT [58]66.9769.4337.7567.1460.7498.4064.7460.2588.3061.9155.39106.0042.6152.1877.71
CoTBA [60]68.6071.5336.3867.0259.8697.4067.6464.8081.2063.7257.6499.8550.8360.5874.82
SAR [43]66.9769.4937.1367.2960.9097.9065.0660.6988.3562.1755.68105.6043.8854.3076.70
DIGA [61]66.9170.3736.5059.2261.7782.0066.5462.4385.8563.8057.32103.7540.6551.0878.82
DomainAdaptor [72]64.9966.9941.5062.9855.51102.7564.4359.8988.5059.4152.05105.4545.1755.8676.98
MedBN [44]49.7751.0975.8857.8351.99102.5559.6354.1489.9554.0548.27109.0038.92/76.89
VPTTA [3]66.4868.9338.5067.1660.5498.9565.0260.7088.3561.7355.01105.5544.5155.9977.43
TopoTTA (Ours)70.7377.0525.3867.3662.7485.7572.9670.2679.1568.4361.2097.6050.7466.0869.23
MethodDeepGlobe → CNDSNeub1 → Neub2Neub2 → Neub1ROSE → OCTA500OCTA500 → ROSE
Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓
Source Only81.7492.227.0914.34/8.2260.9765.6260.5650.1057.6555.3671.3974.6419.55
TENT [58]78.6990.656.6551.5560.589.0663.2672.329.3768.9975.6943.0672.2575.6417.89
CoTBA [60]77.8391.185.9952.2162.7810.0664.4073.587.9670.0575.6541.3472.5875.7818.00
SAR [43]79.1990.726.8651.8560.809.1663.2172.319.6267.6574.1444.4272.4575.8517.44
DIGA [61]75.4583.2214.1059.1466.4419.7761.3273.408.1966.1772.9842.1871.5175.4017.99
DomainAdaptor [72]80.7191.726.8053.2460.967.9464.0472.388.7567.3974.6044.1471.7775.3717.00
MedBN [44]76.9687.949.2255.5963.9210.5661.3470.8211.5740.6447.53298.7471.7773.2916.11
VPTTA [3]77.0188.568.6454.2562.888.8363.4672.249.3467.5474.0144.9272.6175.9317.11
TopoTTA (Ours)89.1596.355.9866.8874.957.2261.8375.386.5070.7078.2431.8875.5977.7316.22
", + "bbox": [ + 91, + 130, + 903, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 330, + 266, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We adopt 10 widely-used TSS datasets for the following experiments. The retinal vessel segmentation datasets are collected from three different sources: DRIVE [54], STARE [16], and CHASEDB1 (CHASE) [12], which contain 40, 20, and 28 samples, respectively. Three road extraction datasets are DeepGlobe [8], Massachusetts road (MR) [39], and CNDS [4], providing 8,570, 1,171, and 224 samples, respectively. The microscopic neuronal segmentation datasets are Neub1 and Neub2 [76], with 112 and 98 samples, respectively. Two retinal OCT-angiography vessel segmentation datasets are from different sources, ROSE1 [37] and OCTA500 [26], containing 117 and 300 samples, respectively. All datasets follow their original training and testing splits, except for STARE, which is officially used as a test set due to its limited sample size [16].", + "bbox": [ + 88, + 361, + 482, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following prior works [46, 59, 65, 68], we use Dice as the primary metric for TSS performance evaluation, with clDice [52] and Betti errors $(\\beta)$ [18] as metrics of topological continuity. In Table 1, the diagonal marks “/” in clDice for some TTA methods indicate cases where no skeleton map can be extracted from their predictions, and thus clDice cannot be calculated. Further investigation reveals that these are failure cases.", + "bbox": [ + 89, + 607, + 482, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. For each scenario, we use a model pre-trained on a single dataset (source domain) and test it on other datasets (target domains) to calculate metrics. UNet [48] and CS2Net [24] (designed for TSS) are adopted as baseline networks for validating the performance of TopoTTA. To ensure consistent field-of-view and cross-domain accuracy, all datasets are resized to $384 \\times 384$ pixels, except for the OCTA dataset as ROSE and OCTA500 already share similar resolutions. During testing, TopoTTA and the comparison methods undergo six iterations of adaptation for each image, with TopoTTA performing three iter", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/35cebd7d0e70cdb1cc00546aea88eabd307804d74ac2464b510da0d9ca848ea2.jpg", + "table_caption": [ + "Table 2. Average cross-dataset testing results across ten datasets using UNet and CS2Net as baseline networks. The best and second-best results in each column are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
MethodUNetCS2Net
Dice (%)↑cI Dice (%)↑β↓Dice (%)↑cI Dice (%)↑β↓
Source Only50.5642.1951.9450.3151.1754.35
TENT [58]63.8167.2949.4263.7668.0951.87
CoTTA [60]65.4969.3447.3065.1269.8448.19
SAR [43]63.9767.4949.3263.6368.1151.84
DIGA [61]63.0767.4448.9263.9268.2051.97
DomainAdaptor [72]63.4166.5349.9860.9565.6554.68
MedBN [44]56.6554.9080.0562.6365.9754.63
VPTTA [3]63.9867.4849.7663.3067.7752.25
TopoTTA (Ours)69.4474.0043.0167.6872.7043.86
", + "bbox": [ + 516, + 387, + 903, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ations at each stage to ensure experimental consistency. For TopoTTA, we use the Adam optimizer with learning rates of 0.01 and $1 \\times 10^{-4}$ for the first and second stages, respectively. We use 0.5 as the threshold for mask binarization. The hyperparameters $n \\times n$ (number of TopoMDC regions), $s$ (modification window size), and $\\tau^{\\mathrm{bg}}$ (upper limit of the foreground pixel), respectively, are set to $4 \\times 4$ , 30, 0.05 for all tasks. We present more details in Appendix.", + "bbox": [ + 511, + 488, + 906, + 608 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison TTA Methods. In the following experiments, we compare our method with three categories of TTA methods: (1) methods designed for semantic segmentation, such as CoTTA [60] and DIGA [61], (2) a method designed for medical image segmentation, i.e., VPTTA [3], and (3) methods designed for classification, including TENT [58], SAR [43], DomainAdaptor [72] and MedBN [44], which can be easily adapted to segmentation tasks [3, 61] by using pixel-level self-supervision or BN statistics adjustment. We implement these TTA methods on the same baseline networks as TopoTTA, using hyperparameters as specified in the corresponding reference or assigning optimal ones.", + "bbox": [ + 511, + 611, + 908, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison Results", + "text_level": 1, + "bbox": [ + 511, + 803, + 705, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents the cross-dataset testing results across four different scenarios. Table 2 shows the average results for each baseline across all scenarios. Fig. 3 visualizes the segmentation results. From these results, we have the following key observations:", + "bbox": [ + 511, + 825, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24128", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1701b7d9a080c922f9d9c65f9c08f2b0d32279c255c29315c88a72da40b289d6.jpg", + "image_caption": [ + "Figure 3. Visualization of segmentation results for TopoTTA and seven comparison methods across four scenarios. \"Source Image\" is provided solely to illustrate domain differences across datasets. \"Source Only\" denotes results without any TTA methods applied, and GT is short for ground-truth labels. More visual examples are presented in Appendix." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key Observation 1: Improvement on topological continuity across four scenarios. As shown in Table 1, TopoTTA consistently achieves the highest topological continuity metrics. This improvement is not attained at the expense of segmentation accuracy (e.g., DIGA in the DRIVE $\\rightarrow$ STARE, and VPTTA in the DeepGlobe $\\rightarrow$ MR), but rather, TopoTTA enhances Dice score on most datasets. Notably, for the DeepGlobe $\\rightarrow$ CNDS experiment, where the source domain model has already demonstrated strong test time performance, the results of all comparison methods decline. However, TopoTTA substantially improves segmentation accuracy and topological continuity metrics in this experiment. In particular, in the challenging Neub1 $\\rightarrow$ Neub2 experiment, TopoTTA surpasses the second-best method, achieving gains of 7.74 and 8.51 in Dice and cIDice scores, respectively, underscoring the method's exceptional capacity for topological adaptation and continuity enhancement. Additionally, in the Neub1 $\\rightarrow$ Neub2 experiment, although TopoTTA's Dice score is lower than most comparison methods, visual analysis indicates this is likely due to label granularity rather than a true performance drop. As space is limited, details are provided in Appendix.", + "bbox": [ + 88, + 460, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key Observation 2: Consistent superiority across different baselines. TopoTTA can function as a plug-and-play module within standard convolutional neural networks. From Table 2, we observe that TopoTTA achieves significant improvements across ten experiments using both UNet and CS2Net as baseline networks. CoTTA attains the second-best performance, indicating that the TS scheme", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "may be well-suited for segmentation tasks. With UNet as the baseline network, TopoTTA outperforms CoTTA by $3.95\\%$ in Dice, $4.66\\%$ in clDice, and reduces Betti errors by $4.29\\%$ . Similar gains are observed using CS2Net. We also implement a TopoTTA variant on DSCNet, which delivers notable performance improvements, demonstrating TopoTTA's adaptability across different baseline networks. Detailed results on other baseline networks (i.e., CS2Net and DSCNet) can be found in Appendix.", + "bbox": [ + 511, + 460, + 906, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Key Observation 3: Clear visual proof of enhanced continuity and topological adaptability. In Fig. 3, we magnify specific regions of selected images for detailed comparison. In rows 1, 2, and 5, all comparison methods show apparent discontinuities in the red-boxed areas, while TopoTTA effectively connects these branches, underscoring its superior continuity enhancement. For the road image in row 3, TopoTTA achieves an almost perfect match with the ground truth—a level of accuracy beyond all comparison methods' reach. In the microscopic neuronal image in row 4, most comparison methods struggle to segment even the main branches, while TopoTTA closely matches the ground truth. These results showcase TopoTTA's adaptability to various topological features. Moreover, in row 6, comparison methods exhibit noise due to over-segmentation, which TopoTTA effectively suppresses.", + "bbox": [ + 511, + 598, + 908, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 848, + 668, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform ablation analyses of TopoMDCs and TopoHG. For all ablation experiments, we report the", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24129", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/57f8ec91fabe5297e73a296e38eefd2b9fad3dee8bd75dcfd6fe95427589f83c.jpg", + "table_caption": [ + "Table 3. Ablation results of the proposed TopoTTA. Baseline* is the model using the teacher-student scheme." + ], + "table_footnote": [], + "table_body": "
MethodDice (%) ↑cI Dice (%) ↑β ↓
Baseline65.3761.6982.24
Baseline + TopoMDCs (Stage 1)68.7065.1476.28
Baseline*65.9562.2077.90
Baseline* + TopoHG (Stage 2)68.8266.6173.63
TopoTTA69.8767.8173.27
", + "bbox": [ + 91, + 116, + 480, + 189 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2df67c0e8ce61a67c0f09da16b45fdc9f30bcfc18b4b4ac1b0ae6a2376c3627f.jpg", + "table_caption": [ + "Table 4. Ablation results of the proposed TopoMDCs using each TopoMDC type individually." + ], + "table_footnote": [], + "table_body": "
VariantMDC Operation SetDice (%) ↑clDice (%) ↑β ↓
CDC{Cc}68.9967.3572.91
Orthogonal MDCs{C1,C2,C3,C4}69.2667.5472.93
Diagonal MDCs{C5,C6,C7,C8}69.2967.4873.86
TopoMDCs (Full){C1,C2,···,C8}69.8767.8173.27
", + "bbox": [ + 91, + 220, + 480, + 279 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/99798830808ab7710ed1bbe4cf3008a3644ea98c0afa9a1e462a11b068eac611.jpg", + "table_caption": [ + "Table 5. Ablation results of the proposed TopoHG. We compare with three variants, i.e., Gaussian blur, random Gaussian noise, and image swap in the spatial domain." + ], + "table_footnote": [], + "table_body": "
MethodDice (%) ↑cIDice (%) ↑β ↓
Stage 168.7065.1476.28
Stage 1 + Blur69.1666.2676.19
Stage 1 + Noise68.4865.0077.76
Stage 1 + Image swap67.6566.3669.63
TopoTTA (Ours)69.8767.8173.27
", + "bbox": [ + 94, + 324, + 480, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "average results on the retinal vessel segmentation scenario. Due to the limited space, the ablation analysis on our hyperparameters is provided in Appendix.", + "bbox": [ + 89, + 395, + 482, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of TopoMDCs. To validate the effectiveness of the proposed TopoMDCs, we compare the network performance before and after applying TopoMDCs to the baseline or removing them from the TopoTTA. As shown in Table 3, TopoMDCs result in significant improvements across all metrics compared to the baseline that only updates statistical parameters, validating the effectiveness of TopoMDCs in adapting to various topological structures. Additionally, Table 4 provides a detailed analysis of the contribution of each type of TopoMDCs to performance enhancement. We categorize TopoMDCs into three types: central $(\\mathcal{C}_c)$ , orthogonal $(\\mathcal{C}_{1-4})$ , and diagonal $(\\mathcal{C}_{5-8})$ , and evaluate their impact individually. Results indicate that orthogonal and diagonal types relative to central lead to more significant performance improvements, with optimal topological awareness achieved when orthogonal and diagonal types are used together. We also visualize feature maps before and after incorporating TopoMDCs. Fig. 4 illustrates that in all four scenarios, TopoMDCs effectively enhance the topological structures on the target domain.", + "bbox": [ + 89, + 443, + 482, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of TopoHG. Here, we aim to answer two research questions: (1) Can TopoHG truly create pseudobreak predictions? (2) Is the low-frequency swap method in TopoHG more effective for enhancing topological continuity compared to other data augmentation methods?", + "bbox": [ + 89, + 750, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 3, using only Stage 2 with TopoHG still leads to improvements in both segmentation performance and topological continuity. Compared to Stage 1 alone, adding TopoHG notably enhances continuity. To answer the first question, we provide visualizations of selected", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f1f61249b925a28412e30172a9a87c80725479ea92ff297ae007b5f904cf3615.jpg", + "image_caption": [ + "Figure 4. Feature maps before and after TopoMDCs across four scenarios. VC represents the feature maps of vanilla convolutions." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 915, + 205 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56c7784fb1c2d14b4fd348e00e3f61e4c846b472ea66deb5c03f83c081fb03fc.jpg", + "image_caption": [ + "Figure 5. Visualizations of original patches $\\pmb{x}_{p}^{\\mathrm{fg}}$ , pseudo-labels $\\hat{\\pmb{y}}_p'$ , pseudo-break patches $\\pmb{x}_p'$ , and weakened prediction masks $\\pmb{y}_p'$ in selected regions across four scenarios." + ], + "image_footnote": [], + "bbox": [ + 521, + 237, + 906, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "regions' predictions before and after introducing pseudo-breaks in Fig. 5. The images generated by TopoHG significantly disrupt the model's previously confident predictions, effectively creating local pseudo-breaks. For the second question, Table 5 compares various data augmentation methods, including Gaussian blur, random Gaussian noise, and foreground swap with the background in the spatial domain (image swap). The blur method creates local hard samples by removing high-frequency details, which offers some improvement. The image swap method seems to yield superior topological performance, but a sharp drop in Dice indicates that lower Betti errors may stem from overly aggressive predictions connecting many regions incorrectly. This is likely due to consistent regularization on the background without high-frequency details, which can cause incorrect segmentation of the background as foreground. Due to the limited space, visual comparisons of different data augmentation methods are provided in Appendix.", + "bbox": [ + 511, + 420, + 906, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 708, + 633, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we proposed the first TTA framework, i.e., TopoTTA, for TSS to enhance cross-domain performance by addressing discrepant topological structures and fragile topological continuity. To improve the adaptability to topological structure, we proposed TopoMDCs to learn how to adaptively integrate various orientation information according to the topological feature of each test sample. Besides, we developed TopoHG to generate pseudo-breaks to enhance the predictions' topological continuity. Extensive experiments across four scenarios verified the superiority of our TopoTTA compared to other methods.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24130", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This research was supported by the National Natural Science Foundation of China (Grant No. 62202027).", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 152, + 187, + 167 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Cheng Chen, Quande Liu, Yueming Jin, Qi Dou, and Pheng-Ann Heng. Source-Free Domain Adaptive Fundus Image Segmentation with Denoised Pseudo-Labeling. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Interventions, pages 225–235, 2021. 1", + "[2] Zhuangzhuang Chen, Zhuonan Lai, Jie Chen, and Jianqiang Li. Mind Marginal Non-Crack Regions: Clustering-Inspired Representation Learning for Crack Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12698-12708, 2024. 3", + "[3] Ziyang Chen, Yongsheng Pan, Yiwen Ye, Mengkang Lu, and Yong Xia. Each Test Image Deserves a Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11184-11193, 2024. 2, 3, 6, 4, 5", + "[4] Guangliang Cheng, Ying Wang, Shibiao Xu, Hongzhen Wang, Shiming Xiang, and Chunhong Pan. Automatic Road Detection and Centerline Extraction via Cascaded End-to-End Convolutional Neural Network. IEEE Transactions on Geoscience and Remote Sensing, 55(6):3322-3337, 2017. 6", + "[5] Lechao Cheng, Chaowei Fang, Dingwen Zhang, Guanbin Li, and Gang Huang. Compound Batch Normalization for Long-Tailed Image Classification. In Proceedings of the ACM International Conference on Multimedia, pages 1925-1934, 2022. 2", + "[6] Mingfei Cheng, Kaili Zhao, Xuhong Guo, Yajing Xu, and Jun Guo. Joint Topology-Preserving and Feature-Refinement Network for Curvilinear Structure Segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 7147–7156, 2021. 1", + "[7] James W. Cooley and John W. Tukey. An Algorithm for the Machine Calculation of Complex Fourier Series. Mathematics of Computation, 19(90):297-301, 1965. 5", + "[8] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. DeepGlobe 2018: A Challenge to Parse the Earth Through Satellite Images. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 172-181, 2018. 6", + "[9] Jiahua Dong, Yang Cong, Gan Sun, Bineng Zhong, and Xiaowei Xu. What Can Be Transferred: Unsupervised Domain Adaptation for Endoscopic Lesions Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4023-4032, 2020. 1", + "[10] Mario Döbler, Robert A Marsden, and Bin Yang. Robust Mean Teacher for Continual and Gradual Test-Time Adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7704-7714, 2023. 2, 3" + ], + "bbox": [ + 93, + 176, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Zhen Fang, Jie Lu, Feng Liu, and Guangquan Zhang. Semi-Supervised Heterogeneous Domain Adaptation: Theory and Algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1087-1105, 2022. 1", + "[12] Muhammad Moazam Fraz, Paolo Remagnino, Andreas Hoppe, Bunyarit Uyyanonvara, Alicja R Rudnicka, Christopher G Owen, and Sarah A Barman. An Ensemble Classification-based Approach Applied to Retinal Blood Vessel Segmentation. IEEE Transactions on Biomedical Engineering, 59(9):2538-2548, 2012. 6", + "[13] Mohsen Ghafoorian, Alireza Mehrtash, Tina Kapur, Nico Karssemeijer, Elena Marchiori, Mehran Pesteie, Charles R. G. Guttmann, Frank-Erik de Leeuw, Clare M. Tempany, Bram van Ginneken, Andriy Fedorov, Purang Abolmaesumi, Bram Platel, and William M. Wells. Transfer Learning for Domain Adaptation in MRI: Application in Brain Lesion Segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Interventions, pages 516-524, 2017. 1", + "[14] Xin Guo, Wenzhong Tang, Haoran Wang, Jiale Wang, Shuai Wang, Xiaolei Qu, and Xun Lin. MorFormer: morphology-aware transformer for generalized pavement crack segmentation. IEEE Transactions on Intelligent Transportation Systems, 26(6):8219-8232, 2025. 2", + "[15] Saumya Gupta, Yikai Zhang, Xiaoling Hu, Prateek Prasanna, and Chao Chen. Topology-Aware Uncertainty for Image Segmentation. In Advances in the Neural Information Processing Systems, page 36, 2024. 1", + "[16] AD Hoover, Valentina Kouznetsova, and Michael Goldbaum. Locating Blood Vessels in Retinal Images by Piecewise Threshold Probing of a Matched Filter Response. IEEE Transactions on Medical Imaging, 19(3):203-210, 2000. 6", + "[17] Xiaoling Hu. Structure-Aware Image Segmentation with Homotopy Warping. In Advances in the Neural Information Processing Systems, pages 24046-24059, 2022. 3", + "[18] Xiaoling Hu, Fuxin Li, Dimitris Samaras, and Chao Chen. Topology-Preserving Deep Image Segmentation. In Advances in the Neural Information Processing Systems, 2019. 6", + "[19] Xiaoling Hu, Yusu Wang, Li Fuxin, Dimitris Samaras, and Chao Chen. Topology-Aware Segmentation Using Discrete Morse Theory. In Proceedings of the International Conference on Learning Representations, 2021. 3", + "[20] Xiaoling Hu, Dimitris Samaras, and Chao Chen. Learning Probabilistic Topological Representations Using Discrete Morse Theory. In Proceedings of the International Conference on Learning Representations, 2022. 3", + "[21] Jiaxing Huang, Yanfeng Zhou, Yaoru Luo, Guole Liu, Heng Guo, and Ge Yang. Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures. In Proceedings of the European Conference on Computer Vision, pages 143-160, 2025. 1, 3", + "[22] Klara Janouskova, Tamir Shor, Chaim Baskin, and Jiri Matas. Single Image Test-Time Adaptation for Segmentation. Transactions on Machine Learning Research, 2024. 2", + "[23] Yannick Kirchhoff, Maximilian R Rokuss, Saikat Roy, Balint Kovacs, Constantin Ulrich, Tassilo Wald, Maximilian Zenk, Philipp Vollmuth, Jens Kleesiek, Fabian Isensee," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24131", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Klaus Maier-Hein. Skeleton Recall Loss for Connectivity Conserving and Resource Efficient Segmentation of Thin Tubular Structures. In Proceedings of the European Conference on Computer Vision, pages 218-234, 2024. 3", + "[24] Mou Lei, Yitian Zhao, Huazhu Fu, Yonghuai Liu, Jun Cheng, Yalin Zheng, Pan Su, Jianlong Yang, Li Chen, Alejandro F Frangi, Masahiro Akiba, and Jiang Liu. CS2-Net: Deep learning Segmentation of Curvilinear Structures in Medical Imaging. Medical Image Analysis, 67:101874, 2021. 1, 6", + "[25] Qin Lei, Jiang Zhong, and Qizhu Dai. Enriching Information and Preserving Semantic Consistency in Expanding Curvilinear Object Segmentation Datasets. In Proceedings of the European Conference on Computer Vision, pages 233-250, 2025. 3", + "[26] Mingchao Li, Kun Huang, Qiuzhuo Xu, Jiadong Yang, Yuhan Zhang, Zexuan Ji, Keren Xie, Songtao Yuan, Qinghuai Liu, and Qiang Chen. OCTA-500: A Retinal Dataset for Optical Coherence Tomography Angiography Study. Medical Image Analysis, 93:103092, 2024. 6", + "[27] Jian Liang, Ran He, and Tieniu Tan. A Comprehensive Survey on Test-Time Adaptation Under Distribution Shifts. International Journal of Computer Vision, pages 1-34, 2024. 2", + "[28] Li Lin, Linkai Peng, Huaqing He, Pujin Cheng, Jiewei Wu, Kenneth K.Y. Wong, and Xiaoying Tang. YoloCurvSeg: You Only Label One Noisy Skeleton for Vessel-Style Curvilinear Structure Segmentation. Medical Image Analysis, 90: 102937, 2023. 3", + "[29] Xun Lin, Shuai Wang, Rizhao Cai, Yizhong Liu, Ying Fu, Wenzhong Tang, Zitong Yu, and Alex C. Kot. Suppress and rebalance: Towards generalized multi-modal face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 211-221, 2024. 3", + "[30] Xun Lin, Yi Yu, Song Xia, Jue Jiang, Haoran Wang, Zitong Yu, Yizhong Liu, Ying Fu, Shuai Wang, Wenzhong Tang, and Alex C. Kot. Safeguarding medical image segmentation datasets against unauthorized training via contour- and texture-aware perturbations. arXiv preprint arXiv:2403.14250, 2024. 1", + "[31] Xun Lin, Yi Yu, Zitong Yu, Ruohan Meng, Jiale Zhou, Ajian Liu, Yizhong Liu, Shuai Wang, Wenzhong Tang, Zhen Lei, and Alex C. Kot. *Hidemia: Hidden wavelet mining for privacy-enhancing medical image analysis*. In *Proceedings of the ACM International Conference on Multimedia*, pages 8110–8119, 2024. 1", + "[32] Xun Lin, Ajian Liu, Zitong Yu, Rizhao Cai, Shuai Wang, Yi Yu, Jun Wan, Zhen Lei, Xiaochun Cao, and Alex Kot. Reliable and balanced transfer learning for generalized multimodal face anti-spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025. 3", + "[33] Huajun Liu, Xiangyu Miao, Christoph Mertz, Chengzhong Xu, and Hui Kong. Crackformer: Transformer Network for Fine-Grained Crack Detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 3783-3792, 2021. 2", + "[34] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. FedDG: Federated Domain Generalization on" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Medical Image Segmentation via Episodic Learning in Continuous Frequency Space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1013-1023, 2021. 5", + "[35] Yuang Liu, Wei Zhang, and Jun Wang. Source-Free Domain Adaptation for Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1215-1224, 2021. 1", + "[36] Junyan Lyu, Yiqi Zhang, Yijin Huang, Li Lin, Pujin Cheng, and Xiaoying Tang. AADG: Automatic Augmentation for Domain Generalization on Retinal Image Segmentation. IEEE Transactions on Medical Imaging, 41(12):3699-3711, 2022. 1", + "[37] Yuhui Ma, Huaying Hao, Jianyang Xie, Huazhu Fu, Jiong Zhang, Jianlong Yang, Zhen Wang, Jiang Liu, Yalin Zheng, and Yitian Zhao. ROSE: A Retinal OCT-Angiography Vessel Segmentation Dataset and New Model. IEEE Transactions on Medical Imaging, 40(3):928-939, 2020. 1, 6", + "[38] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The Norm Must Go On: Dynamic Unsupervised Domain Adaptation by Normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 3", + "[39] Volodymyr Mnih. Machine Learning for Aerial Image Labeling. University of Toronto, Canada, 2013. 6", + "[40] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip H.S. Torr. Tipi: Test Time Adaptation with Transformation Invariance. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 24162-24171, 2023. 2", + "[41] Jiayi Ni, Senqiao Yang, Ran Xu, Jiaming Liu, Xiaoqi Li, Wenyu Jiao, Zehui Chen, Yi Liu, and Shanghang Zhang. Distribution-Aware Continual Test-Time Adaptation for Semantic Segmentation. In Proceedings of International Conference on Robotics and Automation, pages 3044-3050, 2024. 3", + "[42] Munan Ning, Donghuan Lu, Yujia Xie, Dongdong Chen, Dong Wei, Yefeng Zheng, Yonghong Tian, Shuicheng Yan, and Li Yuan. MADAv2: Advanced Multi-Anchors based Active Domain Adaptation Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11): 13553-13566, 2023. 1", + "[43] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards Stable Test-Time Adaptation in Dynamic Wild World. In Proceedings of the International Conference on Learning Representations, 2023. 2, 3, 6, 4, 5", + "[44] Hyejin Park, Jeongyeon Hwang, Sunung Mun, Sangdon Park, and Jungseul Ok. MedBN: Robust Test-Time Adaptation Against Malicious Test Samples. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5997-6007, 2024. 3, 6, 4, 5", + "[45] Bibiloni Pedro, González-Hidalgo M, and Massanet Sebastià. A Survey on Curvilinear Object Segmentation in Multiple Applications. Pattern Recognition, 60:949-970, 2016. 1", + "[46] Yaolei Qi, Yuting He, Xiaoming Qi, Yuan Zhang, and Guanyu Yang. Dynamic Snake Convolution Based on Topo" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24132", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "logical Geometric Constraints for Tubular Structure Segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 6070-6079, 2023. 1, 2, 6", + "[47] Xiaolei Qu, Jiale Zhou, Jue Jiang, Wenhan Wang, Haoran Wang, Shuai Wang, Wenzhong Tang, and Xun Lin. Ehformer: Regional easy-hard-aware transformer for breast lesion segmentation in ultrasound images. Information Fusion, 109:102430, 2024. 1", + "[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Interventions, pages 234–241, 2015. 6", + "[49] Jameel Hassan Abdul Samadh, Hanan Gani, Noor Hazim Hussein, Muhammad Uzair Khattak, Muzammal Naseer, Fahad Khan, and Salman Khan. Align Your Prompts: Test-Time Prompting with Distribution Alignment for Zero-Shot Generalization. In Advances in the Neural Information Processing Systems, 2023. 2", + "[50] Tianyi Shi, Nicolas Boutry, Yongchao Xu, and Thierry Géraud. Local Intensity Order Transformation for Robust Curvilinear Object Segmentation. IEEE Transactions on Image Processing, 31:2557-2569, 2022. 1, 2", + "[51] Tianyi Shi, Xiaohuan Ding, Liang Zhang, and Xin Yang. FreeCOS: Self-Supervised Learning from Fractals and Unlabeled Images for Curvilinear Object Segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 876-886, 2023. 3", + "[52] Suprosanna Shit, Johannes C Paetzold, Anjany Sekuboyina, Ivan Ezhov, Alexander Unger, Andrey Zhylka, Josien P.W. Plum, Ulrich Bauer, and Bjoern H Menze. cIDice: A Novel Topology-Preserving Loss Function for Tubular Structure Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16560-16569, 2021. 3, 6", + "[53] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. ECoTTA: Memory-Efficient Continual Test-Time Adaptation via Self-Distilled Regularization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11920-11929, 2023. 2, 3", + "[54] Joes Staal, Michael D Abramoff, Meindert Niemeijer, Max A Viergever, and Bram Van Ginneken. Ridge-based Vessel Segmentation in Color Images of the Retina. IEEE Transactions on Medical Imaging, 23(4):501-509, 2004. 6", + "[55] Yubo Tan, Kaifu Yang, Shixuan Zhao, and Yongjie Li. Retinal Vessel Segmentation with Skeletal Prior and Contrastive Loss. IEEE Transactions on Medical Imaging, 41(9):2238-2251, 2022. 1", + "[56] Devavrat Tomar, Guillaume Vray, Behzad Bozorgtabar, and Jean-Philippe Thiran. TeSLA: Test-Time Self-Learning with Automatic Adversarial Augmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20341–20350, 2023. 3", + "[57] Chenglong Wang, Yuichiro Hayashi, Masahiro Oda, Hayato Itoh, Takayuki Kitasaka, Alejandro F Frangi, and Kensaku Mori. Tubular Structure Segmentation Using Spatial Fully Connected Network with Radial Distance Loss for 3D Medical Images. In Proceedings of the International Conference" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Medical Image Computing and Computer Assisted Interventions, pages 348-356, 2019. 3", + "[58] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. TENT: Fully Test-Time Adaptation by Entropy Minimization. Proceedings of the International Conference on Learning Representations, 2021. 1, 2, 3, 5, 6, 4", + "[59] Dong Wang, Zhao Zhang, Ziwei Zhao, Yuhang Liu, Yihong Chen, and Liwei Wang. PointScatter: Point Set Representation for Tubular Structure Extraction. In Proceedings of the European Conference on Computer Vision, pages 366-383, 2022. 2, 6", + "[60] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual Test-Time Domain Adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 1, 2, 3, 6, 4, 5", + "[61] Wei Wang, Zhun Zhong, Weijie Wang, Xi Chen, Charles Ling, Boyu Wang, and Nicu Sebe. Dynamically Instance-Guided Adaptation: A Backward-Free Approach for Test-Time Domain Adaptive Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 24090-24099, 2023. 2, 3, 6, 4", + "[62] Wenhan Wang, Jiale Zhou, Jin Zhao, Xun Lin, Yan Zhang, Shan Lu, Wanchen Zhao, Shuai Wang, Wenzhong Tang, and Xiaolei Qu. Interactively fusing global and local features for benign and malignant classification of breast ultrasound images. Ultrasound in Medicine Biology, 51(3):525-534, 2025. 1", + "[63] Yan Wang, Xu Wei, Fengze Liu, Jieneng Chen, Yuyin Zhou, Wei Shen, Elliot K. Fishman, and Alan L. Yuille. Deep Distance Transform for Tubular Structure Segmentation in CT Scans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 3", + "[64] Ying Wang, Yuexing Peng, Wei Li, George C Alexandropoulos, Junchuan Yu, Daqing Ge, and Wei Xiang. DDU-Net: Dual-Decoder-U-Net for Road Extraction Using High-Resolution Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-12, 2022. 1", + "[65] Qian Wu, Yufei Chen, Wei Liu, Xiaodong Yue, and Xiahai Zhuang. Deep Closing: Enhancing Topological Connectivity in Medical Tubular Segmentation. IEEE Transactions on Medical Imaging, 43(11):3990-4003, 2024. 1, 6", + "[66] Jiaqi Yang, Xiaoling Hu, Chao Chen, and Chialing Tsai. A Topological-Attention ConvLSTM Network and Its Application to EM Images. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Interventions, pages 217–228, 2021. 2", + "[67] Yanchao Yang and Stefano Soatto. FDA: Fourier Domain Adaptation for Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4085-4095, 2020. 1", + "[68] Ziyun Yang and Sina Farsiu. Directional Connectivity-based Segmentation of Medical Images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11525-11535, 2023. 2, 6", + "[69] Chang'an Yi, Haotian Chen, Yifan Zhang, Yonghui Xu, Yan Zhou, and Lizhen Cui. From Question to Exploration:" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24133", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Can Classic Test-Time Adaptation Strategies Be Effectively Applied in Semantic Segmentation? In Proceedings of the ACM International Conference on Multimedia, pages 10085-10094, 2024. 2, 3", + "[70] Zitong Yu, Chenxu Zhao, Zezheng Wang, Yunxiao Qin, Zhuo Su, Xiaobai Li, Feng Zhou, and Guoying Zhao. Searching Central Difference Convolutional Networks for Face Anti-Spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5295-5305, 2020. 2, 3", + "[71] Longhui Yuan, Binhui Xie, and Shuang Li. Robust Test-Time Adaptation in Dynamic Scenarios. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 15922-15932, 2023. 2", + "[72] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. DomainAdaptor: A Novel Approach to Test-Time Adaptation. In Proceedings of the IEEE International Conference on Computer Vision, pages 18971-18981, 2023. 2, 3, 6, 4, 5", + "[73] Lefei Zhang, Meng Lan, Jing Zhang, and Dacheng Tao. Stagewise Unsupervised Domain Adaptation with Adversarial Self-Training for Road Segmentation of Remote-Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2021. 1", + "[74] Yizhe Zhang, Tao Zhou, Yuhui Tao, Shuo Wang, Ye Wu, Benyuan Liu, Pengfei Gu, Qiang Chen, and Danny Z Chen. TestFit: A Plug-and-Play One-Pass Test Time Method for Medical Image Segmentation. Medical Image Analysis, 92: 103069, 2024. 2", + "[75] Bowen Zhao, Chen Chen, and Shu-Tao Xia. DELTA: Degradation-Free Fully Test-Time Adaptation. In Proceedings of the International Conference on Learning Representations, 2023. 3", + "[76] He Zhao, Huiqi Li, Sebastian Maurer-Stroh, and Li Cheng. Synthesizing Retinal and Neuronal Images with Generative Adversarial Nets. Medical Image Analysis, 49:14-26, 2018. 6", + "[77] Jiale Zhou, Zuoxun Hou, Hongyan Lu, Wenhan Wang, Wanchen Zhao, Zenan Wang, Dezhi Zheng, Shuai Wang, Wenzhong Tang, and Xiaolei Qu. A deep supervised transformer u-shaped full-resolution residual network for the segmentation of breast ultrasound image. Medical Physics, 50 (12):7513-7524, 2023. 1" + ], + "bbox": [ + 91, + 90, + 483, + 683 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "24134", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/a88fe374-66af-4512-b87e-ddbd06fc2c78_model.json b/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/a88fe374-66af-4512-b87e-ddbd06fc2c78_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2261190f09d5d51950e393bfa603870b8be86139 --- /dev/null +++ b/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/a88fe374-66af-4512-b87e-ddbd06fc2c78_model.json @@ -0,0 +1,2534 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.212, + 0.131, + 0.787, + 0.176 + ], + "angle": 0, + "content": "TopoTTA: Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.203, + 0.812, + 0.275 + ], + "angle": 0, + "content": "Jiale Zhou\\(^{1,2}\\), Wenhan Wang\\(^{3}\\), Shikun Li\\(^{2}\\), Xiaolei Qu\\(^{3}\\), Xin Guo\\(^{3}\\), Yizhong Liu\\(^{3}\\), Wenzhong Tang\\(^{3}\\), Xun Lin\\(^{2,3,*}\\), Yefeng Zheng\\(^{2,*}\\) \n\\(^{1}\\)Zhejiang University \\(^{2}\\)Westlake University \\(^{3}\\)Beihang University \n{zhoujiale, zhengyefeng}@westlake.edu.cn linxun@buaa.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.31, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.342, + 0.485, + 0.735 + ], + "angle": 0, + "content": "Tubular structure segmentation (TSS) is important for various applications, such as hemodynamic analysis and route navigation. Despite significant progress in TSS, domain shifts remain a major challenge, leading to performance degradation in unseen target domains. Unlike other segmentation tasks, TSS is more sensitive to domain shifts, as changes in topological structures can compromise segmentation integrity, and variations in local features distinguishing foreground from background (e.g., texture and contrast) may further disrupt topological continuity. To address these challenges, we propose Topology-enhanced Test-Time Adaptation (TopoTTA), the first test-time adaptation framework designed specifically for TSS. TopoTTA consists of two stages: Stage 1 adapts models to cross-domain topological discrepancies using the proposed Topological Meta Difference Convolutions (TopoMDCs), which enhance topological representation without altering pre-trained parameters; Stage 2 improves topological continuity by a novel Topology Hard sample Generation (TopoHG) strategy and prediction alignment on hard samples with pseudo-labels in the generated pseudobreaks. Extensive experiments across four scenarios and ten datasets demonstrate TopoTTA's effectiveness in handling topological distribution shifts, achieving an average improvement of \\(31.81\\%\\) in clDice. TopoTTA also serves as a plug-and-play TTA solution for CNN-based TSS models." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Tubular structure segmentation (TSS), also known as curvilinear object segmentation [45], plays a critical role in various applications (e.g., hemodynamic analysis [24, 55] and route planning [64, 73]). Although many TSS methods have been proposed and achieved inspiring performance [15, 21], domain shifts remain a common challenge" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.309, + 0.882, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.651 + ], + "angle": 0, + "content": "Figure 1. Comparison between TopoTTA and existing TTA solutions. (a) represents the model during the source domain training phase, while (b) denotes the model in the target domain testing phase. (b1-b4) show the results of directly testing the source-trained model, the TENT [58] method with updated BN layers, the CoTTA [60] method using a teacher-student scheme, and our topology-enhanced approach, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.689 + ], + "angle": 0, + "content": "in real-world scenarios [42] due to discrepancies in imaging devices [1, 13] and sample heterogeneity [11]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Domain shifts undermine the effectiveness of TSS models trained on the source domain when applied to an unseen target domain [36, 50]. Unlike segmentation models trained for other tasks, e.g., semantic [35, 67] or tumor segmentation [9, 30, 31, 47, 77], TSS models are more sensitive to domain shifts [50]. Beyond the common variations such as contrast and noise distribution [62, 63], TSS faces additional challenges related to unseen topological structural characteristics, including variations in trajectory, curvature, branching patterns, and thickness [6, 46]. These factors degrade segmentation performance and even disrupt the topological continuity of the structures [65] (see Fig. 1 (b1)), which may mislead critical downstream applications [37, 55, 64]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.242, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24123" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "Recently, a special domain adaptation (DA) paradigm, namely test-time adaptation (TTA) [27, 40, 43], has been proposed and successfully applied to various image segmentation tasks, including semantic [22, 60, 61] and medical image segmentation [3, 74]. Unlike other DA paradigms, TTA performs continuous online adaptation during the test stage (after model deployment) [71]. It uses extremely small amounts of data (e.g., a single sample or batch), aligning better with real-world scenarios where data must be processed sequentially [58]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.245, + 0.483, + 0.486 + ], + "angle": 0, + "content": "Most TTA methods update pre-trained models using self-supervised methods [27], which can typically be categorized into normalization-based methods [58, 72] and teacher-student schemes [10, 60]. Normalization-based methods update normalization layer statistics [53] (e.g., batch/layer normalization (BN/LN)) or leverage the difference in normalization statistics before and after adaptation to update external model parameters [3, 49]. These methods are initially designed for classification tasks, and their performance may degrade when applied to segmentation tasks with severe pixel-level class imbalance [5]. The second category, teacher-student schemes (based on pseudolabels), updates all model parameters and achieves performance gains in general segmentation tasks [69]. However, as shown in Fig. 1, these methods perform poorly in TSS due to two unique challenges in cross-domain scenarios:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.489, + 0.483, + 0.594 + ], + "angle": 0, + "content": "- Challenge 1: Discrepant topological structures. Tubular structures in the source and target domains can exhibit substantial topological differences (e.g., thickness, curvature, and branching patterns). Prior TTA methods' general-purpose model adaptation strategies struggle to effectively capture these cross-domain topological variations when adapting to individual samples." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.595, + 0.483, + 0.701 + ], + "angle": 0, + "content": "- Challenge 2: Fragile topological continuity. In cross-domain settings, the local characteristics (e.g., color, texture, and contrast) distinguishing foreground from background may change significantly [50]. Existing TTA methods lack targeted adaptation to these local variations, resulting in catastrophic disruptions to topological continuity (see Fig. 1) by misclassifying foreground pixels." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.489, + 0.483, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.778 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we propose Topology-Enhanced Test-Time Adaptation (TopoTTA) in this paper. It includes two adaptation stages: topological structure adaptation (Stage 1) and topological continuity refinement (Stage 2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Stage 1 aims to enhance the model's capability to adapt to discrepant topological structures by utilizing a topology-perceiving model adaptation strategy. Inspired by the Central Difference Convolution (CDC) [70], which captures differences between the central pixel and its neighbors to enhance local representations, we extend this concept to accommodate the elongated and diverse trajectories of tubular structures. Specifically, we introduce Topological Meta" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "Difference Convolutions (TopoMDCs), which perform a dual-pixel difference operation by extending the central pixel's interaction across eight directional neighbors. As shown in Fig. 2, we learn weighted combinations of different directional TopoMDCs (instead of the vanilla convolution) for different regions to enhance the representations of hard-to-detect tubular branches, improving adaptability to discrepant topological structures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Within Stage 2, we design a Topology Hard sample Generation (TopoHG) strategy to construct samples with pseudo-breaks. These pseudo-breaks mislead the models to make weakened predictions with poor topological continuity. We then emphasize the local consistency between the predictions and original pseudo-labels in these hard regions. This stage encourages the models to adapt the characteristics distinguishing foreground from background and helps address continuity destructions in the target domain." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.351, + 0.753, + 0.364 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.367, + 0.905, + 0.412 + ], + "angle": 0, + "content": "- We propose the first TTA framework for tubular structure segmentation, named TopoTTA, which effectively adapts to unseen domains by enhancing topological information." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.413, + 0.905, + 0.488 + ], + "angle": 0, + "content": "- We design eight directional Topological Meta Difference Convolutions (TopoMDCs) within TopoTTA, and learn how to adaptively combine different TopoMDCs based on the topological features of the test sample. TopoMDCs enhance adaptability to unseen topological structures." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.488, + 0.905, + 0.577 + ], + "angle": 0, + "content": "- We propose a Topology Hard sample Generation strategy for TopoTTA, namely TopoHG, which generates pseudobreaks at confidently predicted regions. We then align the predictions around these pseudo breaks to the original pseudo labels, improving the topological continuity in the unseen domain." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.579, + 0.905, + 0.639 + ], + "angle": 0, + "content": "- Extensive experiments on four scenarios across ten datasets validate the effectiveness of TopoTTA in addressing topological distribution shifts, achieving an average performance improvement of \\(31.81\\%\\) in cI Dice." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.367, + 0.905, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.656, + 0.664, + 0.671 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.8, + 0.697 + ], + "angle": 0, + "content": "2.1. Tubular Structure Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Tubular structures, characterized by their thin and long topological structures as well as complex layouts, usually cannot be accurately and continuously localized by general segmentation methods [59]. To tackle these challenges, researchers have incorporated topological priors into model designs, including specialized deformable convolution kernels [46], attention mechanisms [14, 66], and tailored model architectures [33]. Complementary to these module innovations, PointScatter [59] was introduced to represent tubular structures as point sets, addressing the flexibility limitations of masks constrained by fixed grids. DconnNet [68] employed a unique design to separate directional subspaces from the latent space, enhancing feature representation for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24124" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "connectivity. Meanwhile, a growing body of works focused on novel loss functions [2, 17, 19, 20, 23, 52, 57] that explicitly optimize the topological continuity of segmentation results. More recently, some research has started to explore the use of synthetic data [25, 28, 51] and the integration of additional topological information [21, 63] to further improve performance. However, these TSS methods ignore the existence of domain shifts, resulting in suboptimal performance when applied to unseen domains." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.235, + 0.295, + 0.251 + ], + "angle": 0, + "content": "2.2. Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.485, + 0.741 + ], + "angle": 0, + "content": "TTA adaptively adjusts the model at test time based on a small number of unlabeled samples from a different distribution. Early TTA methods primarily modified model representations by adjusting BN's statistics [38, 44, 61, 75]. Mirza et al. [38] adapted target domain features by updating BN statistics continuously. Most TTA methods used self-supervised learning to adjust pre-trained parameters, including normalization-based methods [3, 43, 53, 58, 72] and teacher-student (TS) schemes [10, 41, 56, 60]. Wang et al. [58] proposed TENT, an entropy minimization method that updates BN affine transformation parameters to improve confidence in the target domain. Motivated by TENT, some recent works achieved better performance [3, 72] by designing various strategies to fine-tune BN. Zhang et al. [72] incorporated source domain statistics and extended the entropy minimization loss to utilize target domain information better. More recently, Chen et al. [3] reduced generalization errors by aligning source and target statistics to update external model parameters. However, many TTA methods are designed for classification tasks, ignoring pixel-wise class imbalance. In segmentation tasks, where sample imbalance is more severe, simple statistical adjustments and minor parameter updates are easily dominated by the majority of background pixels, reducing segmentation performance [69]. The TS scheme, which applies consistency regularization to update all model parameters, has shown effectiveness in general segmentation tasks. CoTTA [60] used weight-averaged and augmentation-averaged predictions, further reducing error accumulation in continuous inference. However, these methods lack a focus on tubular structure and topological continuity, leading to performance decline and continuity destruction in TSS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.751, + 0.279, + 0.768 + ], + "angle": 0, + "content": "3. Proposed TopoTTA" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.221, + 0.791 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.797, + 0.483, + 0.874 + ], + "angle": 0, + "content": "Let \\(\\mathcal{D}^s = \\{(\\pmb{x}_i^s,\\pmb{y}_i^s)\\}_{i = 1}^{N^s}\\) denote the labeled source domain dataset, where \\(\\pmb{x}_i^s\\in \\mathbb{R}^{H\\times W\\times C}\\) is an image with height \\(H\\) width \\(W\\), and channel number \\(C\\), and \\(\\pmb{y}_i^s\\in \\{0,1\\}^{H\\times W\\times C}\\) denotes the corresponding ground-truth label. Given a TSS model \\(\\mathcal{F}(\\cdot ;\\theta^s)\\) trained on \\(\\mathcal{D}^s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.878, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\theta^ {s} = \\underset {\\theta^ {s}} {\\arg \\min } \\mathbb {E} _ {\\left(\\boldsymbol {x} ^ {s}, \\boldsymbol {y} ^ {s}\\right) \\sim \\mathcal {D} ^ {s}} \\left[ \\mathcal {L} _ {s e g} \\left(\\mathcal {F} \\left(\\boldsymbol {x} ^ {s}; \\theta^ {s}\\right), \\boldsymbol {y} ^ {s}\\right) \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.171 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{seg}(\\cdot ,\\cdot)\\) can be any TSS loss function. Let \\(\\mathcal{D}^t = \\{\\pmb {x}_i^t\\}_{i = 1}^{N^t}\\) represent the unlabeled target domain dataset. During testing, TTA methods aim to update the model parameters for each input image \\(\\pmb{x}_i^t\\in \\mathcal{D}^t\\) and produce segmentation outputs \\(\\hat{\\pmb{y}}_i^t = \\mathcal{F}(\\pmb{x}_i^t;\\theta_{i - 1}^t)\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.178, + 0.905, + 0.207 + ], + "angle": 0, + "content": "\\[\n\\theta_ {i} ^ {t} = \\underset {\\theta_ {i - 1} ^ {t}} {\\arg \\min } \\mathcal {L} _ {t t a} \\left(\\mathcal {F} \\left(\\cdot ; \\theta_ {i - 1} ^ {t}\\right), \\boldsymbol {x} _ {i} ^ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.215, + 0.905, + 0.261 + ], + "angle": 0, + "content": "where \\(\\theta_i^t\\) are the parameters of the TSS model \\(\\mathcal{F}(\\cdot ;\\cdot)\\) at the \\(i\\)-th iteration, and \\(\\mathcal{L}_{tta}(\\cdot ,\\cdot)\\) denotes the self-supervised loss function for TTA. Before performing TTA, we set \\(\\theta_0^t = \\theta^s\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.261, + 0.905, + 0.291 + ], + "angle": 0, + "content": "To improve readability, we use the simplified notation \\( \\pmb{x} \\) instead of \\( \\pmb{x}_i^t \\) in the following sections and figures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.299, + 0.722, + 0.315 + ], + "angle": 0, + "content": "3.2. Overview of TopoTTA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.907, + 0.503 + ], + "angle": 0, + "content": "As discussed in Section 1, existing TTA methods cannot effectively segment tubular structures due to two major challenges: discrepant topological structures and fragile topological continuity. To solve these issues, we propose TopoTTA, which consists of two stages: Stage 1 performs topological structure adaptation to overcome the insufficient cross-domain topological structure adaptation; Stage 2 conducts topological continuity refinement to enhance the topological continuity of the final predictions. An overview of our TopoTTA is illustrated in Fig. 2. The details of Stage 1 and Stage 2 are described in Section 3.3 and Section 3.4, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.511, + 0.875, + 0.527 + ], + "angle": 0, + "content": "3.3. Stage 1: Topological Structure Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.639 + ], + "angle": 0, + "content": "This stage aims to adapt the pre-trained TSS model to discrepant topological structures caused by domain shifts. To achieve this, we propose TopoMDCs based on diverse fundamental topological patterns and adaptively reweight and combine TopoMDCs according to the topological features of different regions. We introduce the details of TopoMDCs and the model parameter updating process below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Topology-Meta Difference Convolutions (TopoMDCs). TopoMDCs draw inspiration from central difference convolution (CDC) [70], which captures the difference between the central pixel of a convolution kernel and its neighbors to detect local gradient changes. However, the local topological shapes of tubular structures often exhibit directionality and continuity. CDC, initially designed for other fine-grained vision tasks (e.g., face anti-spoofing [29, 32, 70]), lacks the ability to perceive these two features of tubular structures, making them ineffective in representing the complex topologies and geometric constraints within local regions. Let \\( x_{\\mathrm{in}} \\) denote the input feature map. The vanilla convolution \\( \\mathcal{C}_0 \\) and CDC \\( \\mathcal{C}_c \\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.843, + 0.88, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {C} _ {0} (r _ {x}, r _ {y}) = \\sum_ {(\\Delta r _ {x}, \\Delta r _ {y}) \\in \\mathcal {R}} w (\\Delta r _ {x}, \\Delta r _ {y}) \\cdot \\pmb {x} _ {\\mathrm {i n}} (r _ {x} - \\Delta r _ {x}, r _ {y} - \\Delta r _ {y}), \\\\ \\mathcal{C}_{c}(r_{x},r_{y}) = \\sum_{(\\Delta r_{x},\\Delta r_{y})\\in \\mathcal{R}}w(\\Delta r_{x},\\Delta r_{y})\\cdot \\boldsymbol{x}_{\\mathrm{in}}(r_{x},r_{y}), \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24125" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.093, + 0.351, + 0.106 + ], + "angle": 0, + "content": "(a) Stage 1: Topological Structure Adaptation" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.106, + 0.647, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.093, + 0.8, + 0.107 + ], + "angle": 0, + "content": "(c) TopoHG (in Stage 2)" + }, + { + "type": "image", + "bbox": [ + 0.662, + 0.107, + 0.896, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.251, + 0.361, + 0.264 + ], + "angle": 0, + "content": "(b) Stage 2: Topological Continuity Refinement" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.264, + 0.435, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.442, + 0.251, + 0.896, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.377, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed TopoTTA. TopoTTA consists of two stages: (a) Stage 1: Topology structure adaptation with Topological Meta Difference Convolutions (TopoMDCs), and (b) Stage 2: Topological continuity refinement. (c) Topology Hard sample Generation (TopoHG) constructs challenging samples with local pseudo-breaks. CDC is short for central difference convolution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.483, + 0.501 + ], + "angle": 0, + "content": "where \\(\\mathcal{R} = \\{(1,1),(0,1),\\dots ,(-1,0),(-1, - 1)\\}\\) represents the local receptive field of a \\(3\\times 3\\) convolution kernel, \\(w\\) denotes the parameters of the vanilla convolution layer, and \\((r_x,r_y)\\) represents the relative position within the convolution operation centered on the current pixel of \\(x_{\\mathrm{in}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.483, + 0.653 + ], + "angle": 0, + "content": "To overcome this limitation of CDC, the proposed TopoMDCs perform a two-pixel differential operation by extending their reach to the eight neighboring directions beyond the center pixel, as shown in Fig. 2 (a). This design simulates various basic topological patterns, effectively accommodating the elongated and diverse trajectories of tubular structures and enhancing the model's capability to recognize various tubular objects. Let \\( \\mathbf{x}_{\\mathrm{in}} \\in \\mathbb{R}^{H_{\\mathrm{in}} \\times W_{\\mathrm{in}} \\times C_{\\mathrm{in}}} \\) denote the input feature map, the pre-trained vanilla convolution be \\( \\mathcal{C}_0 \\), and the set of TopoMDCs be \\( \\{\\mathcal{C}_1, \\mathcal{C}_2, \\dots, \\mathcal{C}_8\\} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.682 + ], + "angle": 0, + "content": "Taking \\(\\mathcal{C}_1\\), which extends toward the top-left pixel (see Fig. 2 (a)), as an example, it is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.685, + 0.482, + 0.747 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {C} _ {1} \\left(r _ {x}, r _ {y}\\right) = \\mathcal {C} _ {c} \\left(r _ {x}, r _ {y}\\right) - \\sum w \\left(\\Delta r _ {x}, \\Delta r _ {y}\\right) \\cdot \\boldsymbol {x} _ {\\text {i n}} \\left(r _ {x}, r _ {y}\\right) \\\\ (\\Delta r _ {x}, \\Delta r _ {y}) \\in \\mathcal {R} _ {1} \\\\ + \\sum_ {(\\Delta r _ {x}, \\Delta r _ {y}) \\in \\mathcal {R} _ {1}} w (\\Delta r _ {x}, \\Delta r _ {y}) \\cdot \\boldsymbol {x} _ {\\text {i n}} (r _ {x} + 1, r _ {y} + 1), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.796 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}_1 = \\{(-1, - 1),(-1,0),(0, - 1)\\}\\) is the local receptive field of \\(C_1\\). Due to space limitations, the formal definitions of other TopoMDCs are provided in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "TopoMDCs do not introduce any additional parameters for any convolution layer but instead, adaptively transform the vanilla convolution in \\(\\mathcal{F}(\\cdot ;\\theta)\\). Specifically, we replace all \\(3\\times 3\\) convolution layers in the encoder of \\(\\mathcal{F}(\\cdot ;\\theta)\\) with TopoMDCs. All TopoMDCs inherit the parameters of the vanilla convolutions. To adapt to a variety of complex (even unseen) topologies, a dynamic combination of TopoMDCs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.907, + 0.609 + ], + "angle": 0, + "content": "is required. Given that the topologies of local regions tend to be quite similar, while the topologies of distant regions can differ significantly, we introduce different learnable parameter sets \\(\\delta\\) with eight learnable router parameters for different image patches. To adapt to the respective topological structures in different regions, we patchify the input image into \\(n\\times n\\) non-overlap patches, each with the size of \\(H_{\\mathrm{in}} / n\\times W_{\\mathrm{in}} / n\\). Taking the \\(j\\)-th image patches \\(\\pmb{p}_j\\) from \\(\\pmb{x}_{\\mathrm{in}}\\), the corresponding router parameter set \\(\\delta\\) assigns external parameters to different TopoMDCs, enabling adaptive enhancement of the perception of topological structures within the patch. The output feature \\(\\hat{\\pmb{x}}_{\\mathrm{out}}\\) is obtained as:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.619, + 0.907, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {x}} _ {\\text {o u t}} = \\mathcal {C} _ {0} \\left(\\boldsymbol {x} _ {\\text {i n}}\\right) - \\sum_ {j = 1} ^ {n \\times n} \\delta_ {j} \\sum_ {i = 1} ^ {8} \\mathcal {C} _ {i} \\left(\\boldsymbol {x} _ {\\text {i n}} ^ {j}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.906, + 0.734 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_{\\mathrm{in}}^{j}\\) denotes the \\(j\\)-th patch of \\(\\pmb{x}_{\\mathrm{in}}\\). The patchification operations \\(\\mathcal{C}_0\\) and \\(\\mathcal{C}_i\\) differ slightly in their inputs in Eq. (4) and Eq. (5). In Eq. (5), \\(\\mathcal{C}_0\\) and \\(\\mathcal{C}_i\\) denote the same computation applied uniformly to every pixel in \\(\\pmb{x}_{\\mathrm{in}}^{j}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Router Parameter \\((\\delta)\\) Updating in Stage 1. As mentioned above, we do not update the original parameters of the convolution kernel or any other parameters in the network but only update a small number of introduced router parameters \\(\\delta\\). This is not just a consideration for inference speed. Simultaneously adjusting both \\(\\delta\\) and the model parameters increases the search-space complexity, which may lead to conflicts or interference among parameters. This, in turn, could even disrupt the original feature representation. To this end, we only adjust \\(\\delta\\) to adapt the topological response within an external feature space, thereby reducing the risk" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24126" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "of instability. For each incoming test sample \\(\\pmb{x}\\), all router parameters are reset to zero, so that the model can independently adapt to the topological structures of the new sample. We use the self-supervised entropy minimization (EM) loss [58] to update \\(\\delta\\):" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.177, + 0.483, + 0.215 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E M}} = - \\sum_ {j = 1} ^ {2} \\hat {\\mathbf {y}} _ {j} \\log (\\hat {\\mathbf {y}} _ {j}), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.218, + 0.39, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\delta} = \\underset {\\theta^ {t}} {\\arg \\min } \\mathcal {L} _ {\\mathrm {E M}} (\\mathcal {F} (\\cdot ; \\theta^ {t}; \\boldsymbol {\\delta}), \\boldsymbol {x}).\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.249, + 0.465, + 0.266 + ], + "angle": 0, + "content": "3.4. Stage 2: Topological Continuity Refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.271, + 0.484, + 0.408 + ], + "angle": 0, + "content": "As previously discussed in Sec. 3.3, Stage 1 enhances the representation of topological structures in \\(\\mathcal{D}^t\\). In Stage 2, we guide the model to learn local characteristics in \\(\\mathcal{D}^t\\) for foreground-background differentiation, thereby improving the topological continuity of segmentation. To achieve this, we generate a series of pseudo-breaks at key topological structures on the test image \\(\\pmb{x}\\), producing the edited image \\(\\pmb{x}'\\) through the proposed Topology Hard sample Generation (TopoHG) strategy, as shown in Fig. 2 (c)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.484, + 0.457 + ], + "angle": 0, + "content": "Topology Hard Sample Generation (TopoHG). The TopoHG process for generating local hard samples can be roughly divided into the following three steps:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Step 1: Key Point Selection. Since we apply consistency supervision on the predictions of local hard samples, the selected regions must have highly reliable pseudo-labels. To this end, we first select a set of points with confidence above the threshold \\(\\tau = 0.95\\) from the teacher's predictions \\(\\hat{\\pmb{y}}'\\), denoted as \\(\\mathcal{P} = \\{(u,v) \\mid \\hat{\\pmb{y}}'(u,v) > \\tau\\}\\). Then, we randomly select \\(N_p\\) points from \\(\\mathcal{P}\\) as key points. Since prediction confidence varies across different test samples, using a fixed \\(N_p\\) for all input samples is suboptimal. Therefore, we allow samples with higher confidence to select more key points and create more challenging regions to contribute more to the weight update: \\(N_p = k \\cdot |\\mathcal{P}|\\), where \\(k\\) is a hyperparameter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.809 + ], + "angle": 0, + "content": "Step 2: Sliding to Search. For each selected key point \\((u_c, v_c)\\), let \\(x_p^{\\mathrm{fg}}\\) denote the fixed foreground window centered at this point, with a size of \\(s \\times s\\). Then, a background sliding window \\(x_p^{\\mathrm{bg}}\\), of the same size as the foreground window, slides over adjacent non-overlapping regions, identifying window \\(x_p^{\\mathrm{bg},*}\\) with the lowest pseudo-label confidence of all pixels. Meanwhile, to prevent \\(x_p^{\\mathrm{bg},*}\\) from carrying too much foreground information, we set a lower threshold \\(\\tau^{\\mathrm{bg}}\\). If the sum of foreground confidence of all pixels in \\(x_p^{\\mathrm{bg},*}\\) exceeds \\(\\tau^{\\mathrm{bg}}\\), the key point \\((u_c, v_c)\\) is discarded." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Step 3: Frequency-based Pseudo-break Generation. TopoHG aims to create local pseudo-break topological structures within the critical regions. This requires TopoHG to simulate the characteristics of structures that seem to be broken while retaining essential foreground features. Inspired by [34], we adopt a low-frequency swap method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.248 + ], + "angle": 0, + "content": "between background and foreground to preserve high-frequency information that carries critical foreground features. Let FFT and iFFT denote the Fast Fourier Transform (FFT) and inverse Fast Fourier Transform (iFFT) operations [7], respectively, with \\( m_{\\mathrm{low}} \\) as the mask for the low-frequency regions that need to be swapped. Applying FFT to \\( x_p^{\\mathrm{fg}} \\) and \\( x_p^{\\mathrm{bg},*} \\) yields the frequency domain representations \\( f_p^{\\mathrm{fg}} = \\mathrm{FFT}(x_p^{\\mathrm{fg}}) \\) and \\( f_p^{\\mathrm{bg}} = \\mathrm{FFT}(x_p^{\\mathrm{bg},*}) \\). After performing low-frequency swap on \\( f_p^{\\mathrm{fg}} \\) and \\( f_p^{\\mathrm{bg}} \\) and applying iFFT, we obtain the modified image \\( x_p^{\\mathrm{swap}} \\), as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.254, + 0.907, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {p} ^ {\\text {s w a p}} = \\mathrm {i F F T} \\left(\\boldsymbol {f} _ {p} ^ {\\mathrm {f g}} \\cdot \\left(1 - \\boldsymbol {m} _ {\\text {l o w}}\\right) + \\boldsymbol {f} _ {p} ^ {\\mathrm {b g}} \\cdot \\boldsymbol {m} _ {\\text {l o w}}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.907, + 0.356 + ], + "angle": 0, + "content": "To avoid additional risks associated with unnecessary background changes when emphasizing consistency, we update only update pixels in \\(\\pmb{x}_p^{\\mathrm{fg}}\\) that correspond to the foreground in the pseudo-label \\(\\hat{\\pmb{y}}_p'\\). The pseudo-break patch \\(\\pmb{x}_p'\\) can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.364, + 0.907, + 0.382 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {p} ^ {\\prime} = \\boldsymbol {x} _ {p} ^ {\\text {s w a p}} \\cdot \\hat {\\boldsymbol {y}} _ {p} ^ {\\prime} + \\boldsymbol {x} _ {p} ^ {\\mathrm {f g}} \\cdot (1 - \\hat {\\boldsymbol {y}} _ {p} ^ {\\prime}). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.906, + 0.433 + ], + "angle": 0, + "content": "After completing the above four steps, we obtain a hard sample \\( \\pmb{x}^{\\prime} \\) that contains multiple local pseudo-breaks, resulting in weakened predictions to assist in model updates." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.436, + 0.907, + 0.558 + ], + "angle": 0, + "content": "Model Updating in Stage 2. To align the predictions \\(\\hat{\\pmb{y}}''\\) of these local hard samples \\(\\pmb{x}'\\) with the original pseudo-label \\(\\hat{\\pmb{y}}'\\), we apply consistency regularization as supervision, implemented through the TS scheme as shown in Fig. 2 (c). The cross-entropy loss \\(\\mathcal{L}_{\\mathrm{CE}}\\) is used to enhance the consistency between the student's (i.e., prediction of the hard sample \\(\\hat{\\pmb{y}}''\\)) and teacher's (i.e., pseudo-label \\(\\hat{\\pmb{y}}'\\) of original samples) predictions, updating \\(\\theta_i^t\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.564, + 0.906, + 0.6 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} = - \\sum_ {(u, v)} \\mathcal {W} (u, v) \\cdot \\left(\\hat {\\boldsymbol {y}} ^ {\\prime} \\log \\left(\\hat {\\boldsymbol {y}} ^ {\\prime \\prime}\\right) + \\hat {\\boldsymbol {y}} ^ {\\prime \\prime} \\log \\left(\\hat {\\boldsymbol {y}} ^ {\\prime}\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.598, + 0.766, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\theta_{i}^{t} = \\operatorname *{arg min}_{\\substack{\\theta_{i - 1}^{t}\\\\ }}\\mathcal{L}_{\\mathrm{CE}}(\\mathcal{F}(:,\\theta_{i - 1}^{t};\\delta),\\boldsymbol {x}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.906, + 0.679 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}(u,v)\\) is a weight map. The local pseudo-breaks should receive higher attention in consistency regularization; therefore, \\(\\mathcal{W}(u,v)\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.686, + 0.906, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} (u, v) = \\left\\{ \\begin{array}{l l} 1 0 & \\text {i f} (u, v) \\in \\boldsymbol {x} _ {p} ^ {\\prime} \\cdot \\hat {\\boldsymbol {y}} _ {p} ^ {\\prime} \\\\ 1 & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.906, + 0.803 + ], + "angle": 0, + "content": "The gradient is used solely to update the student, while the teacher's parameters are updated by an exponential moving average (EMA) of the student's parameters. Finally, after several steps of updates, the unmodified input \\( \\pmb{x} \\) is fed into the student model to obtain the final prediction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.646, + 0.832 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We conduct comparison experiments across four scenarios, i.e., retinal vessel segmentation, road extraction, microscopic neuronal segmentation, and retinal OCT-angiography vessel segmentation covering ten datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24127" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 1. Cross-domain testing results in four different scenarios, i.e., retinal vessel segmentation, road extraction, microscopic neuronal segmentation, and retinal OCT-angiography vessel segmentation. Source Only: Trained on the source domain and tested on the target domain directly. The best and second-best results in each column are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.131, + 0.905, + 0.322 + ], + "angle": 0, + "content": "
MethodDRIVE → CHASEDRIVE → STARECHASE → DRIVECHASE → STAREDeepGlobe → MR
Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓
Source Only22.3818.2141.2548.05/106.3062.9558.5989.9548.77/105.2044.8954.9776.83
TENT [58]66.9769.4337.7567.1460.7498.4064.7460.2588.3061.9155.39106.0042.6152.1877.71
CoTBA [60]68.6071.5336.3867.0259.8697.4067.6464.8081.2063.7257.6499.8550.8360.5874.82
SAR [43]66.9769.4937.1367.2960.9097.9065.0660.6988.3562.1755.68105.6043.8854.3076.70
DIGA [61]66.9170.3736.5059.2261.7782.0066.5462.4385.8563.8057.32103.7540.6551.0878.82
DomainAdaptor [72]64.9966.9941.5062.9855.51102.7564.4359.8988.5059.4152.05105.4545.1755.8676.98
MedBN [44]49.7751.0975.8857.8351.99102.5559.6354.1489.9554.0548.27109.0038.92/76.89
VPTTA [3]66.4868.9338.5067.1660.5498.9565.0260.7088.3561.7355.01105.5544.5155.9977.43
TopoTTA (Ours)70.7377.0525.3867.3662.7485.7572.9670.2679.1568.4361.2097.6050.7466.0869.23
MethodDeepGlobe → CNDSNeub1 → Neub2Neub2 → Neub1ROSE → OCTA500OCTA500 → ROSE
Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓
Source Only81.7492.227.0914.34/8.2260.9765.6260.5650.1057.6555.3671.3974.6419.55
TENT [58]78.6990.656.6551.5560.589.0663.2672.329.3768.9975.6943.0672.2575.6417.89
CoTBA [60]77.8391.185.9952.2162.7810.0664.4073.587.9670.0575.6541.3472.5875.7818.00
SAR [43]79.1990.726.8651.8560.809.1663.2172.319.6267.6574.1444.4272.4575.8517.44
DIGA [61]75.4583.2214.1059.1466.4419.7761.3273.408.1966.1772.9842.1871.5175.4017.99
DomainAdaptor [72]80.7191.726.8053.2460.967.9464.0472.388.7567.3974.6044.1471.7775.3717.00
MedBN [44]76.9687.949.2255.5963.9210.5661.3470.8211.5740.6447.53298.7471.7773.2916.11
VPTTA [3]77.0188.568.6454.2562.888.8363.4672.249.3467.5474.0144.9272.6175.9317.11
TopoTTA (Ours)89.1596.355.9866.8874.957.2261.8375.386.5070.7078.2431.8875.5977.7316.22
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.332, + 0.267, + 0.349 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.362, + 0.483, + 0.603 + ], + "angle": 0, + "content": "Datasets. We adopt 10 widely-used TSS datasets for the following experiments. The retinal vessel segmentation datasets are collected from three different sources: DRIVE [54], STARE [16], and CHASEDB1 (CHASE) [12], which contain 40, 20, and 28 samples, respectively. Three road extraction datasets are DeepGlobe [8], Massachusetts road (MR) [39], and CNDS [4], providing 8,570, 1,171, and 224 samples, respectively. The microscopic neuronal segmentation datasets are Neub1 and Neub2 [76], with 112 and 98 samples, respectively. Two retinal OCT-angiography vessel segmentation datasets are from different sources, ROSE1 [37] and OCTA500 [26], containing 117 and 300 samples, respectively. All datasets follow their original training and testing splits, except for STARE, which is officially used as a test set due to its limited sample size [16]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.483, + 0.73 + ], + "angle": 0, + "content": "Evaluation Metrics. Following prior works [46, 59, 65, 68], we use Dice as the primary metric for TSS performance evaluation, with clDice [52] and Betti errors \\((\\beta)\\) [18] as metrics of topological continuity. In Table 1, the diagonal marks “/” in clDice for some TTA methods indicate cases where no skeleton map can be extracted from their predictions, and thus clDice cannot be calculated. Further investigation reveals that these are failure cases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. For each scenario, we use a model pre-trained on a single dataset (source domain) and test it on other datasets (target domains) to calculate metrics. UNet [48] and CS2Net [24] (designed for TSS) are adopted as baseline networks for validating the performance of TopoTTA. To ensure consistent field-of-view and cross-domain accuracy, all datasets are resized to \\(384 \\times 384\\) pixels, except for the OCTA dataset as ROSE and OCTA500 already share similar resolutions. During testing, TopoTTA and the comparison methods undergo six iterations of adaptation for each image, with TopoTTA performing three iter" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.331, + 0.907, + 0.386 + ], + "angle": 0, + "content": "Table 2. Average cross-dataset testing results across ten datasets using UNet and CS2Net as baseline networks. The best and second-best results in each column are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.388, + 0.905, + 0.48 + ], + "angle": 0, + "content": "
MethodUNetCS2Net
Dice (%)↑cI Dice (%)↑β↓Dice (%)↑cI Dice (%)↑β↓
Source Only50.5642.1951.9450.3151.1754.35
TENT [58]63.8167.2949.4263.7668.0951.87
CoTTA [60]65.4969.3447.3065.1269.8448.19
SAR [43]63.9767.4949.3263.6368.1151.84
DIGA [61]63.0767.4448.9263.9268.2051.97
DomainAdaptor [72]63.4166.5349.9860.9565.6554.68
MedBN [44]56.6554.9080.0562.6365.9754.63
VPTTA [3]63.9867.4849.7663.3067.7752.25
TopoTTA (Ours)69.4474.0043.0167.6872.7043.86
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.609 + ], + "angle": 0, + "content": "ations at each stage to ensure experimental consistency. For TopoTTA, we use the Adam optimizer with learning rates of 0.01 and \\(1 \\times 10^{-4}\\) for the first and second stages, respectively. We use 0.5 as the threshold for mask binarization. The hyperparameters \\(n \\times n\\) (number of TopoMDC regions), \\(s\\) (modification window size), and \\(\\tau^{\\mathrm{bg}}\\) (upper limit of the foreground pixel), respectively, are set to \\(4 \\times 4\\), 30, 0.05 for all tasks. We present more details in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Comparison TTA Methods. In the following experiments, we compare our method with three categories of TTA methods: (1) methods designed for semantic segmentation, such as CoTTA [60] and DIGA [61], (2) a method designed for medical image segmentation, i.e., VPTTA [3], and (3) methods designed for classification, including TENT [58], SAR [43], DomainAdaptor [72] and MedBN [44], which can be easily adapted to segmentation tasks [3, 61] by using pixel-level self-supervision or BN statistics adjustment. We implement these TTA methods on the same baseline networks as TopoTTA, using hyperparameters as specified in the corresponding reference or assigning optimal ones." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.707, + 0.82 + ], + "angle": 0, + "content": "4.2. Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Table 1 presents the cross-dataset testing results across four different scenarios. Table 2 shows the average results for each baseline across all scenarios. Fig. 3 visualizes the segmentation results. From these results, we have the following key observations:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24128" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.412, + 0.907, + 0.455 + ], + "angle": 0, + "content": "Figure 3. Visualization of segmentation results for TopoTTA and seven comparison methods across four scenarios. \"Source Image\" is provided solely to illustrate domain differences across datasets. \"Source Only\" denotes results without any TTA methods applied, and GT is short for ground-truth labels. More visual examples are presented in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.461, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Key Observation 1: Improvement on topological continuity across four scenarios. As shown in Table 1, TopoTTA consistently achieves the highest topological continuity metrics. This improvement is not attained at the expense of segmentation accuracy (e.g., DIGA in the DRIVE \\(\\rightarrow\\) STARE, and VPTTA in the DeepGlobe \\(\\rightarrow\\) MR), but rather, TopoTTA enhances Dice score on most datasets. Notably, for the DeepGlobe \\(\\rightarrow\\) CNDS experiment, where the source domain model has already demonstrated strong test time performance, the results of all comparison methods decline. However, TopoTTA substantially improves segmentation accuracy and topological continuity metrics in this experiment. In particular, in the challenging Neub1 \\(\\rightarrow\\) Neub2 experiment, TopoTTA surpasses the second-best method, achieving gains of 7.74 and 8.51 in Dice and cIDice scores, respectively, underscoring the method's exceptional capacity for topological adaptation and continuity enhancement. Additionally, in the Neub1 \\(\\rightarrow\\) Neub2 experiment, although TopoTTA's Dice score is lower than most comparison methods, visual analysis indicates this is likely due to label granularity rather than a true performance drop. As space is limited, details are provided in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Key Observation 2: Consistent superiority across different baselines. TopoTTA can function as a plug-and-play module within standard convolutional neural networks. From Table 2, we observe that TopoTTA achieves significant improvements across ten experiments using both UNet and CS2Net as baseline networks. CoTTA attains the second-best performance, indicating that the TS scheme" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.597 + ], + "angle": 0, + "content": "may be well-suited for segmentation tasks. With UNet as the baseline network, TopoTTA outperforms CoTTA by \\(3.95\\%\\) in Dice, \\(4.66\\%\\) in clDice, and reduces Betti errors by \\(4.29\\%\\). Similar gains are observed using CS2Net. We also implement a TopoTTA variant on DSCNet, which delivers notable performance improvements, demonstrating TopoTTA's adaptability across different baseline networks. Detailed results on other baseline networks (i.e., CS2Net and DSCNet) can be found in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.841 + ], + "angle": 0, + "content": "Key Observation 3: Clear visual proof of enhanced continuity and topological adaptability. In Fig. 3, we magnify specific regions of selected images for detailed comparison. In rows 1, 2, and 5, all comparison methods show apparent discontinuities in the red-boxed areas, while TopoTTA effectively connects these branches, underscoring its superior continuity enhancement. For the road image in row 3, TopoTTA achieves an almost perfect match with the ground truth—a level of accuracy beyond all comparison methods' reach. In the microscopic neuronal image in row 4, most comparison methods struggle to segment even the main branches, while TopoTTA closely matches the ground truth. These results showcase TopoTTA's adaptability to various topological features. Moreover, in row 6, comparison methods exhibit noise due to over-segmentation, which TopoTTA effectively suppresses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.669, + 0.866 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this section, we perform ablation analyses of TopoMDCs and TopoHG. For all ablation experiments, we report the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24129" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 3. Ablation results of the proposed TopoTTA. Baseline* is the model using the teacher-student scheme." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.117, + 0.482, + 0.19 + ], + "angle": 0, + "content": "
MethodDice (%) ↑cI Dice (%) ↑β ↓
Baseline65.3761.6982.24
Baseline + TopoMDCs (Stage 1)68.7065.1476.28
Baseline*65.9562.2077.90
Baseline* + TopoHG (Stage 2)68.8266.6173.63
TopoTTA69.8767.8173.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.193, + 0.482, + 0.221 + ], + "angle": 0, + "content": "Table 4. Ablation results of the proposed TopoMDCs using each TopoMDC type individually." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.28 + ], + "angle": 0, + "content": "
VariantMDC Operation SetDice (%) ↑clDice (%) ↑β ↓
CDC{Cc}68.9967.3572.91
Orthogonal MDCs{C1,C2,C3,C4}69.2667.5472.93
Diagonal MDCs{C5,C6,C7,C8}69.2967.4873.86
TopoMDCs (Full){C1,C2,···,C8}69.8767.8173.27
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.283, + 0.483, + 0.325 + ], + "angle": 0, + "content": "Table 5. Ablation results of the proposed TopoHG. We compare with three variants, i.e., Gaussian blur, random Gaussian noise, and image swap in the spatial domain." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.325, + 0.481, + 0.388 + ], + "angle": 0, + "content": "
MethodDice (%) ↑cIDice (%) ↑β ↓
Stage 168.7065.1476.28
Stage 1 + Blur69.1666.2676.19
Stage 1 + Noise68.4865.0077.76
Stage 1 + Image swap67.6566.3669.63
TopoTTA (Ours)69.8767.8173.27
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.483, + 0.442 + ], + "angle": 0, + "content": "average results on the retinal vessel segmentation scenario. Due to the limited space, the ablation analysis on our hyperparameters is provided in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.444, + 0.483, + 0.747 + ], + "angle": 0, + "content": "Effectiveness of TopoMDCs. To validate the effectiveness of the proposed TopoMDCs, we compare the network performance before and after applying TopoMDCs to the baseline or removing them from the TopoTTA. As shown in Table 3, TopoMDCs result in significant improvements across all metrics compared to the baseline that only updates statistical parameters, validating the effectiveness of TopoMDCs in adapting to various topological structures. Additionally, Table 4 provides a detailed analysis of the contribution of each type of TopoMDCs to performance enhancement. We categorize TopoMDCs into three types: central \\((\\mathcal{C}_c)\\), orthogonal \\((\\mathcal{C}_{1-4})\\), and diagonal \\((\\mathcal{C}_{5-8})\\), and evaluate their impact individually. Results indicate that orthogonal and diagonal types relative to central lead to more significant performance improvements, with optimal topological awareness achieved when orthogonal and diagonal types are used together. We also visualize feature maps before and after incorporating TopoMDCs. Fig. 4 illustrates that in all four scenarios, TopoMDCs effectively enhance the topological structures on the target domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Effectiveness of TopoHG. Here, we aim to answer two research questions: (1) Can TopoHG truly create pseudobreak predictions? (2) Is the low-frequency swap method in TopoHG more effective for enhancing topological continuity compared to other data augmentation methods?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 3, using only Stage 2 with TopoHG still leads to improvements in both segmentation performance and topological continuity. Compared to Stage 1 alone, adding TopoHG notably enhances continuity. To answer the first question, we provide visualizations of selected" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.089, + 0.916, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.207, + 0.907, + 0.235 + ], + "angle": 0, + "content": "Figure 4. Feature maps before and after TopoMDCs across four scenarios. VC represents the feature maps of vanilla convolutions." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.238, + 0.908, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.369, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Figure 5. Visualizations of original patches \\( \\pmb{x}_{p}^{\\mathrm{fg}} \\), pseudo-labels \\( \\hat{\\pmb{y}}_p' \\), pseudo-break patches \\( \\pmb{x}_p' \\), and weakened prediction masks \\( \\pmb{y}_p' \\) in selected regions across four scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.907, + 0.693 + ], + "angle": 0, + "content": "regions' predictions before and after introducing pseudo-breaks in Fig. 5. The images generated by TopoHG significantly disrupt the model's previously confident predictions, effectively creating local pseudo-breaks. For the second question, Table 5 compares various data augmentation methods, including Gaussian blur, random Gaussian noise, and foreground swap with the background in the spatial domain (image swap). The blur method creates local hard samples by removing high-frequency details, which offers some improvement. The image swap method seems to yield superior topological performance, but a sharp drop in Dice indicates that lower Betti errors may stem from overly aggressive predictions connecting many regions incorrectly. This is likely due to consistent regularization on the background without high-frequency details, which can cause incorrect segmentation of the background as foreground. Due to the limited space, visual comparisons of different data augmentation methods are provided in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.709, + 0.634, + 0.724 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we proposed the first TTA framework, i.e., TopoTTA, for TSS to enhance cross-domain performance by addressing discrepant topological structures and fragile topological continuity. To improve the adaptability to topological structure, we proposed TopoMDCs to learn how to adaptively integrate various orientation information according to the topological feature of each test sample. Besides, we developed TopoHG to generate pseudo-breaks to enhance the predictions' topological continuity. Extensive experiments across four scenarios verified the superiority of our TopoTTA compared to other methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24130" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "Acknowledgement. This research was supported by the National Natural Science Foundation of China (Grant No. 62202027)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.153, + 0.188, + 0.168 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.178, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[1] Cheng Chen, Quande Liu, Yueming Jin, Qi Dou, and Pheng-Ann Heng. Source-Free Domain Adaptive Fundus Image Segmentation with Denoised Pseudo-Labeling. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Interventions, pages 225–235, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.263, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[2] Zhuangzhuang Chen, Zhuonan Lai, Jie Chen, and Jianqiang Li. Mind Marginal Non-Crack Regions: Clustering-Inspired Representation Learning for Crack Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12698-12708, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.334, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[3] Ziyang Chen, Yongsheng Pan, Yiwen Ye, Mengkang Lu, and Yong Xia. Each Test Image Deserves a Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11184-11193, 2024. 2, 3, 6, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.419, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[4] Guangliang Cheng, Ying Wang, Shibiao Xu, Hongzhen Wang, Shiming Xiang, and Chunhong Pan. Automatic Road Detection and Centerline Extraction via Cascaded End-to-End Convolutional Neural Network. IEEE Transactions on Geoscience and Remote Sensing, 55(6):3322-3337, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[5] Lechao Cheng, Chaowei Fang, Dingwen Zhang, Guanbin Li, and Gang Huang. Compound Batch Normalization for Long-Tailed Image Classification. In Proceedings of the ACM International Conference on Multimedia, pages 1925-1934, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[6] Mingfei Cheng, Kaili Zhao, Xuhong Guo, Yajing Xu, and Jun Guo. Joint Topology-Preserving and Feature-Refinement Network for Curvilinear Structure Segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 7147–7156, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[7] James W. Cooley and John W. Tukey. An Algorithm for the Machine Calculation of Complex Fourier Series. Mathematics of Computation, 19(90):297-301, 1965. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[8] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. DeepGlobe 2018: A Challenge to Parse the Earth Through Satellite Images. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 172-181, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[9] Jiahua Dong, Yang Cong, Gan Sun, Bineng Zhong, and Xiaowei Xu. What Can Be Transferred: Unsupervised Domain Adaptation for Endoscopic Lesions Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4023-4032, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[10] Mario Döbler, Robert A Marsden, and Bin Yang. Robust Mean Teacher for Continual and Gradual Test-Time Adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7704-7714, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.178, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[11] Zhen Fang, Jie Lu, Feng Liu, and Guangquan Zhang. Semi-Supervised Heterogeneous Domain Adaptation: Theory and Algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1087-1105, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[12] Muhammad Moazam Fraz, Paolo Remagnino, Andreas Hoppe, Bunyarit Uyyanonvara, Alicja R Rudnicka, Christopher G Owen, and Sarah A Barman. An Ensemble Classification-based Approach Applied to Retinal Blood Vessel Segmentation. IEEE Transactions on Biomedical Engineering, 59(9):2538-2548, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[13] Mohsen Ghafoorian, Alireza Mehrtash, Tina Kapur, Nico Karssemeijer, Elena Marchiori, Mehran Pesteie, Charles R. G. Guttmann, Frank-Erik de Leeuw, Clare M. Tempany, Bram van Ginneken, Andriy Fedorov, Purang Abolmaesumi, Bram Platel, and William M. Wells. Transfer Learning for Domain Adaptation in MRI: Application in Brain Lesion Segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Interventions, pages 516-524, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.904, + 0.425 + ], + "angle": 0, + "content": "[14] Xin Guo, Wenzhong Tang, Haoran Wang, Jiale Wang, Shuai Wang, Xiaolei Qu, and Xun Lin. MorFormer: morphology-aware transformer for generalized pavement crack segmentation. IEEE Transactions on Intelligent Transportation Systems, 26(6):8219-8232, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.904, + 0.482 + ], + "angle": 0, + "content": "[15] Saumya Gupta, Yikai Zhang, Xiaoling Hu, Prateek Prasanna, and Chao Chen. Topology-Aware Uncertainty for Image Segmentation. In Advances in the Neural Information Processing Systems, page 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[16] AD Hoover, Valentina Kouznetsova, and Michael Goldbaum. Locating Blood Vessels in Retinal Images by Piecewise Threshold Probing of a Matched Filter Response. IEEE Transactions on Medical Imaging, 19(3):203-210, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[17] Xiaoling Hu. Structure-Aware Image Segmentation with Homotopy Warping. In Advances in the Neural Information Processing Systems, pages 24046-24059, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.905, + 0.633 + ], + "angle": 0, + "content": "[18] Xiaoling Hu, Fuxin Li, Dimitris Samaras, and Chao Chen. Topology-Preserving Deep Image Segmentation. In Advances in the Neural Information Processing Systems, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.636, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[19] Xiaoling Hu, Yusu Wang, Li Fuxin, Dimitris Samaras, and Chao Chen. Topology-Aware Segmentation Using Discrete Morse Theory. In Proceedings of the International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[20] Xiaoling Hu, Dimitris Samaras, and Chao Chen. Learning Probabilistic Topological Representations Using Discrete Morse Theory. In Proceedings of the International Conference on Learning Representations, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[21] Jiaxing Huang, Yanfeng Zhou, Yaoru Luo, Guole Liu, Heng Guo, and Ge Yang. Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures. In Proceedings of the European Conference on Computer Vision, pages 143-160, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[22] Klara Janouskova, Tamir Shor, Chaim Baskin, and Jiri Matas. Single Image Test-Time Adaptation for Segmentation. Transactions on Machine Learning Research, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Yannick Kirchhoff, Maximilian R Rokuss, Saikat Roy, Balint Kovacs, Constantin Ulrich, Tassilo Wald, Maximilian Zenk, Philipp Vollmuth, Jens Kleesiek, Fabian Isensee," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24131" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.148 + ], + "angle": 0, + "content": "and Klaus Maier-Hein. Skeleton Recall Loss for Connectivity Conserving and Resource Efficient Segmentation of Thin Tubular Structures. In Proceedings of the European Conference on Computer Vision, pages 218-234, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.218 + ], + "angle": 0, + "content": "[24] Mou Lei, Yitian Zhao, Huazhu Fu, Yonghuai Liu, Jun Cheng, Yalin Zheng, Pan Su, Jianlong Yang, Li Chen, Alejandro F Frangi, Masahiro Akiba, and Jiang Liu. CS2-Net: Deep learning Segmentation of Curvilinear Structures in Medical Imaging. Medical Image Analysis, 67:101874, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.483, + 0.286 + ], + "angle": 0, + "content": "[25] Qin Lei, Jiang Zhong, and Qizhu Dai. Enriching Information and Preserving Semantic Consistency in Expanding Curvilinear Object Segmentation Datasets. In Proceedings of the European Conference on Computer Vision, pages 233-250, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[26] Mingchao Li, Kun Huang, Qiuzhuo Xu, Jiadong Yang, Yuhan Zhang, Zexuan Ji, Keren Xie, Songtao Yuan, Qinghuai Liu, and Qiang Chen. OCTA-500: A Retinal Dataset for Optical Coherence Tomography Angiography Study. Medical Image Analysis, 93:103092, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.483, + 0.411 + ], + "angle": 0, + "content": "[27] Jian Liang, Ran He, and Tieniu Tan. A Comprehensive Survey on Test-Time Adaptation Under Distribution Shifts. International Journal of Computer Vision, pages 1-34, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[28] Li Lin, Linkai Peng, Huaqing He, Pujin Cheng, Jiewei Wu, Kenneth K.Y. Wong, and Xiaoying Tang. YoloCurvSeg: You Only Label One Noisy Skeleton for Vessel-Style Curvilinear Structure Segmentation. Medical Image Analysis, 90: 102937, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[29] Xun Lin, Shuai Wang, Rizhao Cai, Yizhong Liu, Ying Fu, Wenzhong Tang, Zitong Yu, and Alex C. Kot. Suppress and rebalance: Towards generalized multi-modal face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 211-221, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.483, + 0.649 + ], + "angle": 0, + "content": "[30] Xun Lin, Yi Yu, Song Xia, Jue Jiang, Haoran Wang, Zitong Yu, Yizhong Liu, Ying Fu, Shuai Wang, Wenzhong Tang, and Alex C. Kot. Safeguarding medical image segmentation datasets against unauthorized training via contour- and texture-aware perturbations. arXiv preprint arXiv:2403.14250, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[31] Xun Lin, Yi Yu, Zitong Yu, Ruohan Meng, Jiale Zhou, Ajian Liu, Yizhong Liu, Shuai Wang, Wenzhong Tang, Zhen Lei, and Alex C. Kot. *Hidemia: Hidden wavelet mining for privacy-enhancing medical image analysis*. In *Proceedings of the ACM International Conference on Multimedia*, pages 8110–8119, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[32] Xun Lin, Ajian Liu, Zitong Yu, Rizhao Cai, Shuai Wang, Yi Yu, Jun Wan, Zhen Lei, Xiaochun Cao, and Alex Kot. Reliable and balanced transfer learning for generalized multimodal face anti-spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[33] Huajun Liu, Xiangyu Miao, Christoph Mertz, Chengzhong Xu, and Hui Kong. Crackformer: Transformer Network for Fine-Grained Crack Detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 3783-3792, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[34] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. FedDG: Federated Domain Generalization on" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "Medical Image Segmentation via Episodic Learning in Continuous Frequency Space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1013-1023, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[35] Yuang Liu, Wei Zhang, and Jun Wang. Source-Free Domain Adaptation for Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1215-1224, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[36] Junyan Lyu, Yiqi Zhang, Yijin Huang, Li Lin, Pujin Cheng, and Xiaoying Tang. AADG: Automatic Augmentation for Domain Generalization on Retinal Image Segmentation. IEEE Transactions on Medical Imaging, 41(12):3699-3711, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[37] Yuhui Ma, Huaying Hao, Jianyang Xie, Huazhu Fu, Jiong Zhang, Jianlong Yang, Zhen Wang, Jiang Liu, Yalin Zheng, and Yitian Zhao. ROSE: A Retinal OCT-Angiography Vessel Segmentation Dataset and New Model. IEEE Transactions on Medical Imaging, 40(3):928-939, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[38] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The Norm Must Go On: Dynamic Unsupervised Domain Adaptation by Normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[39] Volodymyr Mnih. Machine Learning for Aerial Image Labeling. University of Toronto, Canada, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[40] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip H.S. Torr. Tipi: Test Time Adaptation with Transformation Invariance. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 24162-24171, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[41] Jiayi Ni, Senqiao Yang, Ran Xu, Jiaming Liu, Xiaoqi Li, Wenyu Jiao, Zehui Chen, Yi Liu, and Shanghang Zhang. Distribution-Aware Continual Test-Time Adaptation for Semantic Segmentation. In Proceedings of International Conference on Robotics and Automation, pages 3044-3050, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[42] Munan Ning, Donghuan Lu, Yujia Xie, Dongdong Chen, Dong Wei, Yefeng Zheng, Yonghong Tian, Shuicheng Yan, and Li Yuan. MADAv2: Advanced Multi-Anchors based Active Domain Adaptation Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11): 13553-13566, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[43] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards Stable Test-Time Adaptation in Dynamic Wild World. In Proceedings of the International Conference on Learning Representations, 2023. 2, 3, 6, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[44] Hyejin Park, Jeongyeon Hwang, Sunung Mun, Sangdon Park, and Jungseul Ok. MedBN: Robust Test-Time Adaptation Against Malicious Test Samples. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5997-6007, 2024. 3, 6, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[45] Bibiloni Pedro, González-Hidalgo M, and Massanet Sebastià. A Survey on Curvilinear Object Segmentation in Multiple Applications. Pattern Recognition, 60:949-970, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Yaolei Qi, Yuting He, Xiaoming Qi, Yuan Zhang, and Guanyu Yang. Dynamic Snake Convolution Based on Topo" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24132" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "logical Geometric Constraints for Tubular Structure Segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 6070-6079, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[47] Xiaolei Qu, Jiale Zhou, Jue Jiang, Wenhan Wang, Haoran Wang, Shuai Wang, Wenzhong Tang, and Xun Lin. Ehformer: Regional easy-hard-aware transformer for breast lesion segmentation in ultrasound images. Information Fusion, 109:102430, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.483, + 0.273 + ], + "angle": 0, + "content": "[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Interventions, pages 234–241, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[49] Jameel Hassan Abdul Samadh, Hanan Gani, Noor Hazim Hussein, Muhammad Uzair Khattak, Muzammal Naseer, Fahad Khan, and Salman Khan. Align Your Prompts: Test-Time Prompting with Distribution Alignment for Zero-Shot Generalization. In Advances in the Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[50] Tianyi Shi, Nicolas Boutry, Yongchao Xu, and Thierry Géraud. Local Intensity Order Transformation for Robust Curvilinear Object Segmentation. IEEE Transactions on Image Processing, 31:2557-2569, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[51] Tianyi Shi, Xiaohuan Ding, Liang Zhang, and Xin Yang. FreeCOS: Self-Supervised Learning from Fractals and Unlabeled Images for Curvilinear Object Segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 876-886, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[52] Suprosanna Shit, Johannes C Paetzold, Anjany Sekuboyina, Ivan Ezhov, Alexander Unger, Andrey Zhylka, Josien P.W. Plum, Ulrich Bauer, and Bjoern H Menze. cIDice: A Novel Topology-Preserving Loss Function for Tubular Structure Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16560-16569, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[53] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. ECoTTA: Memory-Efficient Continual Test-Time Adaptation via Self-Distilled Regularization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11920-11929, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[54] Joes Staal, Michael D Abramoff, Meindert Niemeijer, Max A Viergever, and Bram Van Ginneken. Ridge-based Vessel Segmentation in Color Images of the Retina. IEEE Transactions on Medical Imaging, 23(4):501-509, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[55] Yubo Tan, Kaifu Yang, Shixuan Zhao, and Yongjie Li. Retinal Vessel Segmentation with Skeletal Prior and Contrastive Loss. IEEE Transactions on Medical Imaging, 41(9):2238-2251, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[56] Devavrat Tomar, Guillaume Vray, Behzad Bozorgtabar, and Jean-Philippe Thiran. TeSLA: Test-Time Self-Learning with Automatic Adversarial Augmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20341–20350, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[57] Chenglong Wang, Yuichiro Hayashi, Masahiro Oda, Hayato Itoh, Takayuki Kitasaka, Alejandro F Frangi, and Kensaku Mori. Tubular Structure Segmentation Using Spatial Fully Connected Network with Radial Distance Loss for 3D Medical Images. In Proceedings of the International Conference" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "on Medical Image Computing and Computer Assisted Interventions, pages 348-356, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[58] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. TENT: Fully Test-Time Adaptation by Entropy Minimization. Proceedings of the International Conference on Learning Representations, 2021. 1, 2, 3, 5, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[59] Dong Wang, Zhao Zhang, Ziwei Zhao, Yuhang Liu, Yihong Chen, and Liwei Wang. PointScatter: Point Set Representation for Tubular Structure Extraction. In Proceedings of the European Conference on Computer Vision, pages 366-383, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[60] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual Test-Time Domain Adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 1, 2, 3, 6, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[61] Wei Wang, Zhun Zhong, Weijie Wang, Xi Chen, Charles Ling, Boyu Wang, and Nicu Sebe. Dynamically Instance-Guided Adaptation: A Backward-Free Approach for Test-Time Domain Adaptive Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 24090-24099, 2023. 2, 3, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[62] Wenhan Wang, Jiale Zhou, Jin Zhao, Xun Lin, Yan Zhang, Shan Lu, Wanchen Zhao, Shuai Wang, Wenzhong Tang, and Xiaolei Qu. Interactively fusing global and local features for benign and malignant classification of breast ultrasound images. Ultrasound in Medicine Biology, 51(3):525-534, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[63] Yan Wang, Xu Wei, Fengze Liu, Jieneng Chen, Yuyin Zhou, Wei Shen, Elliot K. Fishman, and Alan L. Yuille. Deep Distance Transform for Tubular Structure Segmentation in CT Scans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[64] Ying Wang, Yuexing Peng, Wei Li, George C Alexandropoulos, Junchuan Yu, Daqing Ge, and Wei Xiang. DDU-Net: Dual-Decoder-U-Net for Road Extraction Using High-Resolution Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-12, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[65] Qian Wu, Yufei Chen, Wei Liu, Xiaodong Yue, and Xiahai Zhuang. Deep Closing: Enhancing Topological Connectivity in Medical Tubular Segmentation. IEEE Transactions on Medical Imaging, 43(11):3990-4003, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[66] Jiaqi Yang, Xiaoling Hu, Chao Chen, and Chialing Tsai. A Topological-Attention ConvLSTM Network and Its Application to EM Images. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Interventions, pages 217–228, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[67] Yanchao Yang and Stefano Soatto. FDA: Fourier Domain Adaptation for Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4085-4095, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[68] Ziyun Yang and Sina Farsiu. Directional Connectivity-based Segmentation of Medical Images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11525-11535, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[69] Chang'an Yi, Haotian Chen, Yifan Zhang, Yonghui Xu, Yan Zhou, and Lizhen Cui. From Question to Exploration:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24133" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Can Classic Test-Time Adaptation Strategies Be Effectively Applied in Semantic Segmentation? In Proceedings of the ACM International Conference on Multimedia, pages 10085-10094, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[70] Zitong Yu, Chenxu Zhao, Zezheng Wang, Yunxiao Qin, Zhuo Su, Xiaobai Li, Feng Zhou, and Guoying Zhao. Searching Central Difference Convolutional Networks for Face Anti-Spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5295-5305, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[71] Longhui Yuan, Binhui Xie, and Shuang Li. Robust Test-Time Adaptation in Dynamic Scenarios. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 15922-15932, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[72] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. DomainAdaptor: A Novel Approach to Test-Time Adaptation. In Proceedings of the IEEE International Conference on Computer Vision, pages 18971-18981, 2023. 2, 3, 6, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[73] Lefei Zhang, Meng Lan, Jing Zhang, and Dacheng Tao. Stagewise Unsupervised Domain Adaptation with Adversarial Self-Training for Road Segmentation of Remote-Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.485 + ], + "angle": 0, + "content": "[74] Yizhe Zhang, Tao Zhou, Yuhui Tao, Shuo Wang, Ye Wu, Benyuan Liu, Pengfei Gu, Qiang Chen, and Danny Z Chen. TestFit: A Plug-and-Play One-Pass Test Time Method for Medical Image Segmentation. Medical Image Analysis, 92: 103069, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[75] Bowen Zhao, Chen Chen, and Shu-Tao Xia. DELTA: Degradation-Free Fully Test-Time Adaptation. In Proceedings of the International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[76] He Zhao, Huiqi Li, Sebastian Maurer-Stroh, and Li Cheng. Synthesizing Retinal and Neuronal Images with Generative Adversarial Nets. Medical Image Analysis, 49:14-26, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[77] Jiale Zhou, Zuoxun Hou, Hongyan Lu, Wenhan Wang, Wanchen Zhao, Zenan Wang, Dezhi Zheng, Shuai Wang, Wenzhong Tang, and Xiaolei Qu. A deep supervised transformer u-shaped full-resolution residual network for the segmentation of breast ultrasound image. Medical Physics, 50 (12):7513-7524, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24134" + } + ] +] \ No newline at end of file diff --git a/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/a88fe374-66af-4512-b87e-ddbd06fc2c78_origin.pdf b/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/a88fe374-66af-4512-b87e-ddbd06fc2c78_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2b9e6ff6080a671eaef90f305eb5f06d047a8f7a --- /dev/null +++ b/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/a88fe374-66af-4512-b87e-ddbd06fc2c78_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de8081ca9c4daa84190439bd898b56bd2eef4759c58b35acb2be5233d14cff6f +size 6323508 diff --git a/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/full.md b/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9cd562b3dd8ae35d8fff166e1f0846e06f17dd4e --- /dev/null +++ b/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/full.md @@ -0,0 +1,345 @@ +# TopoTTA: Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation + +Jiale Zhou $^{1,2}$ , Wenhan Wang $^{3}$ , Shikun Li $^{2}$ , Xiaolei Qu $^{3}$ , Xin Guo $^{3}$ , Yizhong Liu $^{3}$ , Wenzhong Tang $^{3}$ , Xun Lin $^{2,3,*}$ , Yefeng Zheng $^{2,*}$ $^{1}$ Zhejiang University $^{2}$ Westlake University $^{3}$ Beihang University +{zhoujiale, zhengyefeng}@westlake.edu.cn linxun@buaa.edu.cn + +# Abstract + +Tubular structure segmentation (TSS) is important for various applications, such as hemodynamic analysis and route navigation. Despite significant progress in TSS, domain shifts remain a major challenge, leading to performance degradation in unseen target domains. Unlike other segmentation tasks, TSS is more sensitive to domain shifts, as changes in topological structures can compromise segmentation integrity, and variations in local features distinguishing foreground from background (e.g., texture and contrast) may further disrupt topological continuity. To address these challenges, we propose Topology-enhanced Test-Time Adaptation (TopoTTA), the first test-time adaptation framework designed specifically for TSS. TopoTTA consists of two stages: Stage 1 adapts models to cross-domain topological discrepancies using the proposed Topological Meta Difference Convolutions (TopoMDCs), which enhance topological representation without altering pre-trained parameters; Stage 2 improves topological continuity by a novel Topology Hard sample Generation (TopoHG) strategy and prediction alignment on hard samples with pseudo-labels in the generated pseudobreaks. Extensive experiments across four scenarios and ten datasets demonstrate TopoTTA's effectiveness in handling topological distribution shifts, achieving an average improvement of $31.81\%$ in clDice. TopoTTA also serves as a plug-and-play TTA solution for CNN-based TSS models. + +# 1. Introduction + +Tubular structure segmentation (TSS), also known as curvilinear object segmentation [45], plays a critical role in various applications (e.g., hemodynamic analysis [24, 55] and route planning [64, 73]). Although many TSS methods have been proposed and achieved inspiring performance [15, 21], domain shifts remain a common challenge + +![](images/9a7bccbfc9c7800d0f0a7cd0e86281b4f7acb1540d22a575ef1fc4b99911dd4a.jpg) +Figure 1. Comparison between TopoTTA and existing TTA solutions. (a) represents the model during the source domain training phase, while (b) denotes the model in the target domain testing phase. (b1-b4) show the results of directly testing the source-trained model, the TENT [58] method with updated BN layers, the CoTTA [60] method using a teacher-student scheme, and our topology-enhanced approach, respectively. + +in real-world scenarios [42] due to discrepancies in imaging devices [1, 13] and sample heterogeneity [11]. + +Domain shifts undermine the effectiveness of TSS models trained on the source domain when applied to an unseen target domain [36, 50]. Unlike segmentation models trained for other tasks, e.g., semantic [35, 67] or tumor segmentation [9, 30, 31, 47, 77], TSS models are more sensitive to domain shifts [50]. Beyond the common variations such as contrast and noise distribution [62, 63], TSS faces additional challenges related to unseen topological structural characteristics, including variations in trajectory, curvature, branching patterns, and thickness [6, 46]. These factors degrade segmentation performance and even disrupt the topological continuity of the structures [65] (see Fig. 1 (b1)), which may mislead critical downstream applications [37, 55, 64]. + +Recently, a special domain adaptation (DA) paradigm, namely test-time adaptation (TTA) [27, 40, 43], has been proposed and successfully applied to various image segmentation tasks, including semantic [22, 60, 61] and medical image segmentation [3, 74]. Unlike other DA paradigms, TTA performs continuous online adaptation during the test stage (after model deployment) [71]. It uses extremely small amounts of data (e.g., a single sample or batch), aligning better with real-world scenarios where data must be processed sequentially [58]. + +Most TTA methods update pre-trained models using self-supervised methods [27], which can typically be categorized into normalization-based methods [58, 72] and teacher-student schemes [10, 60]. Normalization-based methods update normalization layer statistics [53] (e.g., batch/layer normalization (BN/LN)) or leverage the difference in normalization statistics before and after adaptation to update external model parameters [3, 49]. These methods are initially designed for classification tasks, and their performance may degrade when applied to segmentation tasks with severe pixel-level class imbalance [5]. The second category, teacher-student schemes (based on pseudolabels), updates all model parameters and achieves performance gains in general segmentation tasks [69]. However, as shown in Fig. 1, these methods perform poorly in TSS due to two unique challenges in cross-domain scenarios: + +- Challenge 1: Discrepant topological structures. Tubular structures in the source and target domains can exhibit substantial topological differences (e.g., thickness, curvature, and branching patterns). Prior TTA methods' general-purpose model adaptation strategies struggle to effectively capture these cross-domain topological variations when adapting to individual samples. +- Challenge 2: Fragile topological continuity. In cross-domain settings, the local characteristics (e.g., color, texture, and contrast) distinguishing foreground from background may change significantly [50]. Existing TTA methods lack targeted adaptation to these local variations, resulting in catastrophic disruptions to topological continuity (see Fig. 1) by misclassifying foreground pixels. + +To address the aforementioned challenges, we propose Topology-Enhanced Test-Time Adaptation (TopoTTA) in this paper. It includes two adaptation stages: topological structure adaptation (Stage 1) and topological continuity refinement (Stage 2). + +Stage 1 aims to enhance the model's capability to adapt to discrepant topological structures by utilizing a topology-perceiving model adaptation strategy. Inspired by the Central Difference Convolution (CDC) [70], which captures differences between the central pixel and its neighbors to enhance local representations, we extend this concept to accommodate the elongated and diverse trajectories of tubular structures. Specifically, we introduce Topological Meta + +Difference Convolutions (TopoMDCs), which perform a dual-pixel difference operation by extending the central pixel's interaction across eight directional neighbors. As shown in Fig. 2, we learn weighted combinations of different directional TopoMDCs (instead of the vanilla convolution) for different regions to enhance the representations of hard-to-detect tubular branches, improving adaptability to discrepant topological structures. + +Within Stage 2, we design a Topology Hard sample Generation (TopoHG) strategy to construct samples with pseudo-breaks. These pseudo-breaks mislead the models to make weakened predictions with poor topological continuity. We then emphasize the local consistency between the predictions and original pseudo-labels in these hard regions. This stage encourages the models to adapt the characteristics distinguishing foreground from background and helps address continuity destructions in the target domain. + +Our contributions are as follows: + +- We propose the first TTA framework for tubular structure segmentation, named TopoTTA, which effectively adapts to unseen domains by enhancing topological information. +- We design eight directional Topological Meta Difference Convolutions (TopoMDCs) within TopoTTA, and learn how to adaptively combine different TopoMDCs based on the topological features of the test sample. TopoMDCs enhance adaptability to unseen topological structures. +- We propose a Topology Hard sample Generation strategy for TopoTTA, namely TopoHG, which generates pseudobreaks at confidently predicted regions. We then align the predictions around these pseudo breaks to the original pseudo labels, improving the topological continuity in the unseen domain. +- Extensive experiments on four scenarios across ten datasets validate the effectiveness of TopoTTA in addressing topological distribution shifts, achieving an average performance improvement of $31.81\%$ in cI Dice. + +# 2. Related Works + +# 2.1. Tubular Structure Segmentation + +Tubular structures, characterized by their thin and long topological structures as well as complex layouts, usually cannot be accurately and continuously localized by general segmentation methods [59]. To tackle these challenges, researchers have incorporated topological priors into model designs, including specialized deformable convolution kernels [46], attention mechanisms [14, 66], and tailored model architectures [33]. Complementary to these module innovations, PointScatter [59] was introduced to represent tubular structures as point sets, addressing the flexibility limitations of masks constrained by fixed grids. DconnNet [68] employed a unique design to separate directional subspaces from the latent space, enhancing feature representation for + +connectivity. Meanwhile, a growing body of works focused on novel loss functions [2, 17, 19, 20, 23, 52, 57] that explicitly optimize the topological continuity of segmentation results. More recently, some research has started to explore the use of synthetic data [25, 28, 51] and the integration of additional topological information [21, 63] to further improve performance. However, these TSS methods ignore the existence of domain shifts, resulting in suboptimal performance when applied to unseen domains. + +# 2.2. Test-Time Adaptation + +TTA adaptively adjusts the model at test time based on a small number of unlabeled samples from a different distribution. Early TTA methods primarily modified model representations by adjusting BN's statistics [38, 44, 61, 75]. Mirza et al. [38] adapted target domain features by updating BN statistics continuously. Most TTA methods used self-supervised learning to adjust pre-trained parameters, including normalization-based methods [3, 43, 53, 58, 72] and teacher-student (TS) schemes [10, 41, 56, 60]. Wang et al. [58] proposed TENT, an entropy minimization method that updates BN affine transformation parameters to improve confidence in the target domain. Motivated by TENT, some recent works achieved better performance [3, 72] by designing various strategies to fine-tune BN. Zhang et al. [72] incorporated source domain statistics and extended the entropy minimization loss to utilize target domain information better. More recently, Chen et al. [3] reduced generalization errors by aligning source and target statistics to update external model parameters. However, many TTA methods are designed for classification tasks, ignoring pixel-wise class imbalance. In segmentation tasks, where sample imbalance is more severe, simple statistical adjustments and minor parameter updates are easily dominated by the majority of background pixels, reducing segmentation performance [69]. The TS scheme, which applies consistency regularization to update all model parameters, has shown effectiveness in general segmentation tasks. CoTTA [60] used weight-averaged and augmentation-averaged predictions, further reducing error accumulation in continuous inference. However, these methods lack a focus on tubular structure and topological continuity, leading to performance decline and continuity destruction in TSS. + +# 3. Proposed TopoTTA + +# 3.1. Preliminary + +Let $\mathcal{D}^s = \{(\pmb{x}_i^s,\pmb{y}_i^s)\}_{i = 1}^{N^s}$ denote the labeled source domain dataset, where $\pmb{x}_i^s\in \mathbb{R}^{H\times W\times C}$ is an image with height $H$ width $W$ , and channel number $C$ , and $\pmb{y}_i^s\in \{0,1\}^{H\times W\times C}$ denotes the corresponding ground-truth label. Given a TSS model $\mathcal{F}(\cdot ;\theta^s)$ trained on $\mathcal{D}^s$ : + +$$ +\theta^ {s} = \underset {\theta^ {s}} {\arg \min } \mathbb {E} _ {\left(\boldsymbol {x} ^ {s}, \boldsymbol {y} ^ {s}\right) \sim \mathcal {D} ^ {s}} \left[ \mathcal {L} _ {s e g} \left(\mathcal {F} \left(\boldsymbol {x} ^ {s}; \theta^ {s}\right), \boldsymbol {y} ^ {s}\right) \right], \tag {1} +$$ + +where $\mathcal{L}_{seg}(\cdot ,\cdot)$ can be any TSS loss function. Let $\mathcal{D}^t = \{\pmb {x}_i^t\}_{i = 1}^{N^t}$ represent the unlabeled target domain dataset. During testing, TTA methods aim to update the model parameters for each input image $\pmb{x}_i^t\in \mathcal{D}^t$ and produce segmentation outputs $\hat{\pmb{y}}_i^t = \mathcal{F}(\pmb{x}_i^t;\theta_{i - 1}^t)$ as follows: + +$$ +\theta_ {i} ^ {t} = \underset {\theta_ {i - 1} ^ {t}} {\arg \min } \mathcal {L} _ {t t a} \left(\mathcal {F} \left(\cdot ; \theta_ {i - 1} ^ {t}\right), \boldsymbol {x} _ {i} ^ {t}\right), \tag {2} +$$ + +where $\theta_i^t$ are the parameters of the TSS model $\mathcal{F}(\cdot ;\cdot)$ at the $i$ -th iteration, and $\mathcal{L}_{tta}(\cdot ,\cdot)$ denotes the self-supervised loss function for TTA. Before performing TTA, we set $\theta_0^t = \theta^s$ . + +To improve readability, we use the simplified notation $\pmb{x}$ instead of $\pmb{x}_i^t$ in the following sections and figures. + +# 3.2. Overview of TopoTTA + +As discussed in Section 1, existing TTA methods cannot effectively segment tubular structures due to two major challenges: discrepant topological structures and fragile topological continuity. To solve these issues, we propose TopoTTA, which consists of two stages: Stage 1 performs topological structure adaptation to overcome the insufficient cross-domain topological structure adaptation; Stage 2 conducts topological continuity refinement to enhance the topological continuity of the final predictions. An overview of our TopoTTA is illustrated in Fig. 2. The details of Stage 1 and Stage 2 are described in Section 3.3 and Section 3.4, respectively. + +# 3.3. Stage 1: Topological Structure Adaptation + +This stage aims to adapt the pre-trained TSS model to discrepant topological structures caused by domain shifts. To achieve this, we propose TopoMDCs based on diverse fundamental topological patterns and adaptively reweight and combine TopoMDCs according to the topological features of different regions. We introduce the details of TopoMDCs and the model parameter updating process below. + +Topology-Meta Difference Convolutions (TopoMDCs). TopoMDCs draw inspiration from central difference convolution (CDC) [70], which captures the difference between the central pixel of a convolution kernel and its neighbors to detect local gradient changes. However, the local topological shapes of tubular structures often exhibit directionality and continuity. CDC, initially designed for other fine-grained vision tasks (e.g., face anti-spoofing [29, 32, 70]), lacks the ability to perceive these two features of tubular structures, making them ineffective in representing the complex topologies and geometric constraints within local regions. Let $x_{\mathrm{in}}$ denote the input feature map. The vanilla convolution $\mathcal{C}_0$ and CDC $\mathcal{C}_c$ can be formulated as: + +$$ +\begin{array}{l} \mathcal {C} _ {0} (r _ {x}, r _ {y}) = \sum_ {(\Delta r _ {x}, \Delta r _ {y}) \in \mathcal {R}} w (\Delta r _ {x}, \Delta r _ {y}) \cdot \pmb {x} _ {\mathrm {i n}} (r _ {x} - \Delta r _ {x}, r _ {y} - \Delta r _ {y}), \\ \mathcal{C}_{c}(r_{x},r_{y}) = \sum_{(\Delta r_{x},\Delta r_{y})\in \mathcal{R}}w(\Delta r_{x},\Delta r_{y})\cdot \boldsymbol{x}_{\mathrm{in}}(r_{x},r_{y}), \\ \end{array} +$$ + +![](images/ae5dbb41cbc112e71d559124140fdb6b81e4f67bbf450942b769c405f6928c10.jpg) +(a) Stage 1: Topological Structure Adaptation + +![](images/a3edc90b2f535911d2f66a84eb8fc64e8bc35d1903870c4d19d40b66a0f9b5d5.jpg) +(c) TopoHG (in Stage 2) + +![](images/2a5c0669467da68bedad559d3d8434723bd693c1527fc979a4c1e64ae7a3d341.jpg) +(b) Stage 2: Topological Continuity Refinement + +![](images/a73c430afdc39bcaa88c982d220176019963c7ab7b7548228fd78da11e7acd7d.jpg) +Figure 2. Overview of the proposed TopoTTA. TopoTTA consists of two stages: (a) Stage 1: Topology structure adaptation with Topological Meta Difference Convolutions (TopoMDCs), and (b) Stage 2: Topological continuity refinement. (c) Topology Hard sample Generation (TopoHG) constructs challenging samples with local pseudo-breaks. CDC is short for central difference convolution. + +where $\mathcal{R} = \{(1,1),(0,1),\dots ,(-1,0),(-1, - 1)\}$ represents the local receptive field of a $3\times 3$ convolution kernel, $w$ denotes the parameters of the vanilla convolution layer, and $(r_x,r_y)$ represents the relative position within the convolution operation centered on the current pixel of $x_{\mathrm{in}}$ + +To overcome this limitation of CDC, the proposed TopoMDCs perform a two-pixel differential operation by extending their reach to the eight neighboring directions beyond the center pixel, as shown in Fig. 2 (a). This design simulates various basic topological patterns, effectively accommodating the elongated and diverse trajectories of tubular structures and enhancing the model's capability to recognize various tubular objects. Let $\mathbf{x}_{\mathrm{in}} \in \mathbb{R}^{H_{\mathrm{in}} \times W_{\mathrm{in}} \times C_{\mathrm{in}}}$ denote the input feature map, the pre-trained vanilla convolution be $\mathcal{C}_0$ , and the set of TopoMDCs be $\{\mathcal{C}_1, \mathcal{C}_2, \dots, \mathcal{C}_8\}$ . + +Taking $\mathcal{C}_1$ , which extends toward the top-left pixel (see Fig. 2 (a)), as an example, it is defined as: + +$$ +\begin{array}{l} \mathcal {C} _ {1} \left(r _ {x}, r _ {y}\right) = \mathcal {C} _ {c} \left(r _ {x}, r _ {y}\right) - \sum w \left(\Delta r _ {x}, \Delta r _ {y}\right) \cdot \boldsymbol {x} _ {\text {i n}} \left(r _ {x}, r _ {y}\right) \\ (\Delta r _ {x}, \Delta r _ {y}) \in \mathcal {R} _ {1} \\ + \sum_ {(\Delta r _ {x}, \Delta r _ {y}) \in \mathcal {R} _ {1}} w (\Delta r _ {x}, \Delta r _ {y}) \cdot \boldsymbol {x} _ {\text {i n}} (r _ {x} + 1, r _ {y} + 1), \tag {4} \\ \end{array} +$$ + +where $\mathcal{R}_1 = \{(-1, - 1),(-1,0),(0, - 1)\}$ is the local receptive field of $C_1$ . Due to space limitations, the formal definitions of other TopoMDCs are provided in Appendix. + +TopoMDCs do not introduce any additional parameters for any convolution layer but instead, adaptively transform the vanilla convolution in $\mathcal{F}(\cdot ;\theta)$ . Specifically, we replace all $3\times 3$ convolution layers in the encoder of $\mathcal{F}(\cdot ;\theta)$ with TopoMDCs. All TopoMDCs inherit the parameters of the vanilla convolutions. To adapt to a variety of complex (even unseen) topologies, a dynamic combination of TopoMDCs + +is required. Given that the topologies of local regions tend to be quite similar, while the topologies of distant regions can differ significantly, we introduce different learnable parameter sets $\delta$ with eight learnable router parameters for different image patches. To adapt to the respective topological structures in different regions, we patchify the input image into $n\times n$ non-overlap patches, each with the size of $H_{\mathrm{in}} / n\times W_{\mathrm{in}} / n$ . Taking the $j$ -th image patches $\pmb{p}_j$ from $\pmb{x}_{\mathrm{in}}$ , the corresponding router parameter set $\delta$ assigns external parameters to different TopoMDCs, enabling adaptive enhancement of the perception of topological structures within the patch. The output feature $\hat{\pmb{x}}_{\mathrm{out}}$ is obtained as: + +$$ +\hat {\boldsymbol {x}} _ {\text {o u t}} = \mathcal {C} _ {0} \left(\boldsymbol {x} _ {\text {i n}}\right) - \sum_ {j = 1} ^ {n \times n} \delta_ {j} \sum_ {i = 1} ^ {8} \mathcal {C} _ {i} \left(\boldsymbol {x} _ {\text {i n}} ^ {j}\right), \tag {5} +$$ + +where $\pmb{x}_{\mathrm{in}}^{j}$ denotes the $j$ -th patch of $\pmb{x}_{\mathrm{in}}$ . The patchification operations $\mathcal{C}_0$ and $\mathcal{C}_i$ differ slightly in their inputs in Eq. (4) and Eq. (5). In Eq. (5), $\mathcal{C}_0$ and $\mathcal{C}_i$ denote the same computation applied uniformly to every pixel in $\pmb{x}_{\mathrm{in}}^{j}$ . + +Router Parameter $(\delta)$ Updating in Stage 1. As mentioned above, we do not update the original parameters of the convolution kernel or any other parameters in the network but only update a small number of introduced router parameters $\delta$ . This is not just a consideration for inference speed. Simultaneously adjusting both $\delta$ and the model parameters increases the search-space complexity, which may lead to conflicts or interference among parameters. This, in turn, could even disrupt the original feature representation. To this end, we only adjust $\delta$ to adapt the topological response within an external feature space, thereby reducing the risk + +of instability. For each incoming test sample $\pmb{x}$ , all router parameters are reset to zero, so that the model can independently adapt to the topological structures of the new sample. We use the self-supervised entropy minimization (EM) loss [58] to update $\delta$ : + +$$ +\mathcal {L} _ {\mathrm {E M}} = - \sum_ {j = 1} ^ {2} \hat {\mathbf {y}} _ {j} \log (\hat {\mathbf {y}} _ {j}), \tag {6} +$$ + +$$ +\boldsymbol {\delta} = \underset {\theta^ {t}} {\arg \min } \mathcal {L} _ {\mathrm {E M}} (\mathcal {F} (\cdot ; \theta^ {t}; \boldsymbol {\delta}), \boldsymbol {x}). +$$ + +# 3.4. Stage 2: Topological Continuity Refinement + +As previously discussed in Sec. 3.3, Stage 1 enhances the representation of topological structures in $\mathcal{D}^t$ . In Stage 2, we guide the model to learn local characteristics in $\mathcal{D}^t$ for foreground-background differentiation, thereby improving the topological continuity of segmentation. To achieve this, we generate a series of pseudo-breaks at key topological structures on the test image $\pmb{x}$ , producing the edited image $\pmb{x}'$ through the proposed Topology Hard sample Generation (TopoHG) strategy, as shown in Fig. 2 (c). + +Topology Hard Sample Generation (TopoHG). The TopoHG process for generating local hard samples can be roughly divided into the following three steps: + +Step 1: Key Point Selection. Since we apply consistency supervision on the predictions of local hard samples, the selected regions must have highly reliable pseudo-labels. To this end, we first select a set of points with confidence above the threshold $\tau = 0.95$ from the teacher's predictions $\hat{\pmb{y}}'$ , denoted as $\mathcal{P} = \{(u,v) \mid \hat{\pmb{y}}'(u,v) > \tau\}$ . Then, we randomly select $N_p$ points from $\mathcal{P}$ as key points. Since prediction confidence varies across different test samples, using a fixed $N_p$ for all input samples is suboptimal. Therefore, we allow samples with higher confidence to select more key points and create more challenging regions to contribute more to the weight update: $N_p = k \cdot |\mathcal{P}|$ , where $k$ is a hyperparameter. + +Step 2: Sliding to Search. For each selected key point $(u_c, v_c)$ , let $x_p^{\mathrm{fg}}$ denote the fixed foreground window centered at this point, with a size of $s \times s$ . Then, a background sliding window $x_p^{\mathrm{bg}}$ , of the same size as the foreground window, slides over adjacent non-overlapping regions, identifying window $x_p^{\mathrm{bg},*}$ with the lowest pseudo-label confidence of all pixels. Meanwhile, to prevent $x_p^{\mathrm{bg},*}$ from carrying too much foreground information, we set a lower threshold $\tau^{\mathrm{bg}}$ . If the sum of foreground confidence of all pixels in $x_p^{\mathrm{bg},*}$ exceeds $\tau^{\mathrm{bg}}$ , the key point $(u_c, v_c)$ is discarded. + +Step 3: Frequency-based Pseudo-break Generation. TopoHG aims to create local pseudo-break topological structures within the critical regions. This requires TopoHG to simulate the characteristics of structures that seem to be broken while retaining essential foreground features. Inspired by [34], we adopt a low-frequency swap method + +between background and foreground to preserve high-frequency information that carries critical foreground features. Let FFT and iFFT denote the Fast Fourier Transform (FFT) and inverse Fast Fourier Transform (iFFT) operations [7], respectively, with $m_{\mathrm{low}}$ as the mask for the low-frequency regions that need to be swapped. Applying FFT to $x_p^{\mathrm{fg}}$ and $x_p^{\mathrm{bg},*}$ yields the frequency domain representations $f_p^{\mathrm{fg}} = \mathrm{FFT}(x_p^{\mathrm{fg}})$ and $f_p^{\mathrm{bg}} = \mathrm{FFT}(x_p^{\mathrm{bg},*})$ . After performing low-frequency swap on $f_p^{\mathrm{fg}}$ and $f_p^{\mathrm{bg}}$ and applying iFFT, we obtain the modified image $x_p^{\mathrm{swap}}$ , as follows: + +$$ +\boldsymbol {x} _ {p} ^ {\text {s w a p}} = \mathrm {i F F T} \left(\boldsymbol {f} _ {p} ^ {\mathrm {f g}} \cdot \left(1 - \boldsymbol {m} _ {\text {l o w}}\right) + \boldsymbol {f} _ {p} ^ {\mathrm {b g}} \cdot \boldsymbol {m} _ {\text {l o w}}\right). \tag {7} +$$ + +To avoid additional risks associated with unnecessary background changes when emphasizing consistency, we update only update pixels in $\pmb{x}_p^{\mathrm{fg}}$ that correspond to the foreground in the pseudo-label $\hat{\pmb{y}}_p'$ . The pseudo-break patch $\pmb{x}_p'$ can be defined as follows: + +$$ +\boldsymbol {x} _ {p} ^ {\prime} = \boldsymbol {x} _ {p} ^ {\text {s w a p}} \cdot \hat {\boldsymbol {y}} _ {p} ^ {\prime} + \boldsymbol {x} _ {p} ^ {\mathrm {f g}} \cdot (1 - \hat {\boldsymbol {y}} _ {p} ^ {\prime}). \tag {8} +$$ + +After completing the above four steps, we obtain a hard sample $\pmb{x}^{\prime}$ that contains multiple local pseudo-breaks, resulting in weakened predictions to assist in model updates. + +Model Updating in Stage 2. To align the predictions $\hat{\pmb{y}}''$ of these local hard samples $\pmb{x}'$ with the original pseudo-label $\hat{\pmb{y}}'$ , we apply consistency regularization as supervision, implemented through the TS scheme as shown in Fig. 2 (c). The cross-entropy loss $\mathcal{L}_{\mathrm{CE}}$ is used to enhance the consistency between the student's (i.e., prediction of the hard sample $\hat{\pmb{y}}''$ ) and teacher's (i.e., pseudo-label $\hat{\pmb{y}}'$ of original samples) predictions, updating $\theta_i^t$ as follows: + +$$ +\mathcal {L} _ {\mathrm {C E}} = - \sum_ {(u, v)} \mathcal {W} (u, v) \cdot \left(\hat {\boldsymbol {y}} ^ {\prime} \log \left(\hat {\boldsymbol {y}} ^ {\prime \prime}\right) + \hat {\boldsymbol {y}} ^ {\prime \prime} \log \left(\hat {\boldsymbol {y}} ^ {\prime}\right)\right), \tag {9} +$$ + +$$ +\theta_{i}^{t} = \operatorname *{arg min}_{\substack{\theta_{i - 1}^{t}\\ }}\mathcal{L}_{\mathrm{CE}}(\mathcal{F}(:,\theta_{i - 1}^{t};\delta),\boldsymbol {x}), +$$ + +where $\mathcal{W}(u,v)$ is a weight map. The local pseudo-breaks should receive higher attention in consistency regularization; therefore, $\mathcal{W}(u,v)$ is defined as follows: + +$$ +\mathcal {W} (u, v) = \left\{ \begin{array}{l l} 1 0 & \text {i f} (u, v) \in \boldsymbol {x} _ {p} ^ {\prime} \cdot \hat {\boldsymbol {y}} _ {p} ^ {\prime} \\ 1 & \text {o t h e r w i s e} \end{array} . \right. \tag {10} +$$ + +The gradient is used solely to update the student, while the teacher's parameters are updated by an exponential moving average (EMA) of the student's parameters. Finally, after several steps of updates, the unmodified input $\pmb{x}$ is fed into the student model to obtain the final prediction. + +# 4. Experiments + +We conduct comparison experiments across four scenarios, i.e., retinal vessel segmentation, road extraction, microscopic neuronal segmentation, and retinal OCT-angiography vessel segmentation covering ten datasets. + +Table 1. Cross-domain testing results in four different scenarios, i.e., retinal vessel segmentation, road extraction, microscopic neuronal segmentation, and retinal OCT-angiography vessel segmentation. Source Only: Trained on the source domain and tested on the target domain directly. The best and second-best results in each column are highlighted in bold and underline, respectively. + +
MethodDRIVE → CHASEDRIVE → STARECHASE → DRIVECHASE → STAREDeepGlobe → MR
Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓
Source Only22.3818.2141.2548.05/106.3062.9558.5989.9548.77/105.2044.8954.9776.83
TENT [58]66.9769.4337.7567.1460.7498.4064.7460.2588.3061.9155.39106.0042.6152.1877.71
CoTBA [60]68.6071.5336.3867.0259.8697.4067.6464.8081.2063.7257.6499.8550.8360.5874.82
SAR [43]66.9769.4937.1367.2960.9097.9065.0660.6988.3562.1755.68105.6043.8854.3076.70
DIGA [61]66.9170.3736.5059.2261.7782.0066.5462.4385.8563.8057.32103.7540.6551.0878.82
DomainAdaptor [72]64.9966.9941.5062.9855.51102.7564.4359.8988.5059.4152.05105.4545.1755.8676.98
MedBN [44]49.7751.0975.8857.8351.99102.5559.6354.1489.9554.0548.27109.0038.92/76.89
VPTTA [3]66.4868.9338.5067.1660.5498.9565.0260.7088.3561.7355.01105.5544.5155.9977.43
TopoTTA (Ours)70.7377.0525.3867.3662.7485.7572.9670.2679.1568.4361.2097.6050.7466.0869.23
MethodDeepGlobe → CNDSNeub1 → Neub2Neub2 → Neub1ROSE → OCTA500OCTA500 → ROSE
Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓
Source Only81.7492.227.0914.34/8.2260.9765.6260.5650.1057.6555.3671.3974.6419.55
TENT [58]78.6990.656.6551.5560.589.0663.2672.329.3768.9975.6943.0672.2575.6417.89
CoTBA [60]77.8391.185.9952.2162.7810.0664.4073.587.9670.0575.6541.3472.5875.7818.00
SAR [43]79.1990.726.8651.8560.809.1663.2172.319.6267.6574.1444.4272.4575.8517.44
DIGA [61]75.4583.2214.1059.1466.4419.7761.3273.408.1966.1772.9842.1871.5175.4017.99
DomainAdaptor [72]80.7191.726.8053.2460.967.9464.0472.388.7567.3974.6044.1471.7775.3717.00
MedBN [44]76.9687.949.2255.5963.9210.5661.3470.8211.5740.6447.53298.7471.7773.2916.11
VPTTA [3]77.0188.568.6454.2562.888.8363.4672.249.3467.5474.0144.9272.6175.9317.11
TopoTTA (Ours)89.1596.355.9866.8874.957.2261.8375.386.5070.7078.2431.8875.5977.7316.22
+ +# 4.1. Experiment Setup + +Datasets. We adopt 10 widely-used TSS datasets for the following experiments. The retinal vessel segmentation datasets are collected from three different sources: DRIVE [54], STARE [16], and CHASEDB1 (CHASE) [12], which contain 40, 20, and 28 samples, respectively. Three road extraction datasets are DeepGlobe [8], Massachusetts road (MR) [39], and CNDS [4], providing 8,570, 1,171, and 224 samples, respectively. The microscopic neuronal segmentation datasets are Neub1 and Neub2 [76], with 112 and 98 samples, respectively. Two retinal OCT-angiography vessel segmentation datasets are from different sources, ROSE1 [37] and OCTA500 [26], containing 117 and 300 samples, respectively. All datasets follow their original training and testing splits, except for STARE, which is officially used as a test set due to its limited sample size [16]. + +Evaluation Metrics. Following prior works [46, 59, 65, 68], we use Dice as the primary metric for TSS performance evaluation, with clDice [52] and Betti errors $(\beta)$ [18] as metrics of topological continuity. In Table 1, the diagonal marks “/” in clDice for some TTA methods indicate cases where no skeleton map can be extracted from their predictions, and thus clDice cannot be calculated. Further investigation reveals that these are failure cases. + +Implementation Details. For each scenario, we use a model pre-trained on a single dataset (source domain) and test it on other datasets (target domains) to calculate metrics. UNet [48] and CS2Net [24] (designed for TSS) are adopted as baseline networks for validating the performance of TopoTTA. To ensure consistent field-of-view and cross-domain accuracy, all datasets are resized to $384 \times 384$ pixels, except for the OCTA dataset as ROSE and OCTA500 already share similar resolutions. During testing, TopoTTA and the comparison methods undergo six iterations of adaptation for each image, with TopoTTA performing three iter + +Table 2. Average cross-dataset testing results across ten datasets using UNet and CS2Net as baseline networks. The best and second-best results in each column are highlighted in bold and underline, respectively. + +
MethodUNetCS2Net
Dice (%)↑cI Dice (%)↑β↓Dice (%)↑cI Dice (%)↑β↓
Source Only50.5642.1951.9450.3151.1754.35
TENT [58]63.8167.2949.4263.7668.0951.87
CoTTA [60]65.4969.3447.3065.1269.8448.19
SAR [43]63.9767.4949.3263.6368.1151.84
DIGA [61]63.0767.4448.9263.9268.2051.97
DomainAdaptor [72]63.4166.5349.9860.9565.6554.68
MedBN [44]56.6554.9080.0562.6365.9754.63
VPTTA [3]63.9867.4849.7663.3067.7752.25
TopoTTA (Ours)69.4474.0043.0167.6872.7043.86
+ +ations at each stage to ensure experimental consistency. For TopoTTA, we use the Adam optimizer with learning rates of 0.01 and $1 \times 10^{-4}$ for the first and second stages, respectively. We use 0.5 as the threshold for mask binarization. The hyperparameters $n \times n$ (number of TopoMDC regions), $s$ (modification window size), and $\tau^{\mathrm{bg}}$ (upper limit of the foreground pixel), respectively, are set to $4 \times 4$ , 30, 0.05 for all tasks. We present more details in Appendix. + +Comparison TTA Methods. In the following experiments, we compare our method with three categories of TTA methods: (1) methods designed for semantic segmentation, such as CoTTA [60] and DIGA [61], (2) a method designed for medical image segmentation, i.e., VPTTA [3], and (3) methods designed for classification, including TENT [58], SAR [43], DomainAdaptor [72] and MedBN [44], which can be easily adapted to segmentation tasks [3, 61] by using pixel-level self-supervision or BN statistics adjustment. We implement these TTA methods on the same baseline networks as TopoTTA, using hyperparameters as specified in the corresponding reference or assigning optimal ones. + +# 4.2. Comparison Results + +Table 1 presents the cross-dataset testing results across four different scenarios. Table 2 shows the average results for each baseline across all scenarios. Fig. 3 visualizes the segmentation results. From these results, we have the following key observations: + +![](images/1701b7d9a080c922f9d9c65f9c08f2b0d32279c255c29315c88a72da40b289d6.jpg) +Figure 3. Visualization of segmentation results for TopoTTA and seven comparison methods across four scenarios. "Source Image" is provided solely to illustrate domain differences across datasets. "Source Only" denotes results without any TTA methods applied, and GT is short for ground-truth labels. More visual examples are presented in Appendix. + +Key Observation 1: Improvement on topological continuity across four scenarios. As shown in Table 1, TopoTTA consistently achieves the highest topological continuity metrics. This improvement is not attained at the expense of segmentation accuracy (e.g., DIGA in the DRIVE $\rightarrow$ STARE, and VPTTA in the DeepGlobe $\rightarrow$ MR), but rather, TopoTTA enhances Dice score on most datasets. Notably, for the DeepGlobe $\rightarrow$ CNDS experiment, where the source domain model has already demonstrated strong test time performance, the results of all comparison methods decline. However, TopoTTA substantially improves segmentation accuracy and topological continuity metrics in this experiment. In particular, in the challenging Neub1 $\rightarrow$ Neub2 experiment, TopoTTA surpasses the second-best method, achieving gains of 7.74 and 8.51 in Dice and cIDice scores, respectively, underscoring the method's exceptional capacity for topological adaptation and continuity enhancement. Additionally, in the Neub1 $\rightarrow$ Neub2 experiment, although TopoTTA's Dice score is lower than most comparison methods, visual analysis indicates this is likely due to label granularity rather than a true performance drop. As space is limited, details are provided in Appendix. + +Key Observation 2: Consistent superiority across different baselines. TopoTTA can function as a plug-and-play module within standard convolutional neural networks. From Table 2, we observe that TopoTTA achieves significant improvements across ten experiments using both UNet and CS2Net as baseline networks. CoTTA attains the second-best performance, indicating that the TS scheme + +may be well-suited for segmentation tasks. With UNet as the baseline network, TopoTTA outperforms CoTTA by $3.95\%$ in Dice, $4.66\%$ in clDice, and reduces Betti errors by $4.29\%$ . Similar gains are observed using CS2Net. We also implement a TopoTTA variant on DSCNet, which delivers notable performance improvements, demonstrating TopoTTA's adaptability across different baseline networks. Detailed results on other baseline networks (i.e., CS2Net and DSCNet) can be found in Appendix. + +Key Observation 3: Clear visual proof of enhanced continuity and topological adaptability. In Fig. 3, we magnify specific regions of selected images for detailed comparison. In rows 1, 2, and 5, all comparison methods show apparent discontinuities in the red-boxed areas, while TopoTTA effectively connects these branches, underscoring its superior continuity enhancement. For the road image in row 3, TopoTTA achieves an almost perfect match with the ground truth—a level of accuracy beyond all comparison methods' reach. In the microscopic neuronal image in row 4, most comparison methods struggle to segment even the main branches, while TopoTTA closely matches the ground truth. These results showcase TopoTTA's adaptability to various topological features. Moreover, in row 6, comparison methods exhibit noise due to over-segmentation, which TopoTTA effectively suppresses. + +# 4.3. Ablation Study + +In this section, we perform ablation analyses of TopoMDCs and TopoHG. For all ablation experiments, we report the + +Table 3. Ablation results of the proposed TopoTTA. Baseline* is the model using the teacher-student scheme. + +
MethodDice (%) ↑cI Dice (%) ↑β ↓
Baseline65.3761.6982.24
Baseline + TopoMDCs (Stage 1)68.7065.1476.28
Baseline*65.9562.2077.90
Baseline* + TopoHG (Stage 2)68.8266.6173.63
TopoTTA69.8767.8173.27
+ +Table 4. Ablation results of the proposed TopoMDCs using each TopoMDC type individually. + +
VariantMDC Operation SetDice (%) ↑clDice (%) ↑β ↓
CDC{Cc}68.9967.3572.91
Orthogonal MDCs{C1,C2,C3,C4}69.2667.5472.93
Diagonal MDCs{C5,C6,C7,C8}69.2967.4873.86
TopoMDCs (Full){C1,C2,···,C8}69.8767.8173.27
+ +Table 5. Ablation results of the proposed TopoHG. We compare with three variants, i.e., Gaussian blur, random Gaussian noise, and image swap in the spatial domain. + +
MethodDice (%) ↑cIDice (%) ↑β ↓
Stage 168.7065.1476.28
Stage 1 + Blur69.1666.2676.19
Stage 1 + Noise68.4865.0077.76
Stage 1 + Image swap67.6566.3669.63
TopoTTA (Ours)69.8767.8173.27
+ +average results on the retinal vessel segmentation scenario. Due to the limited space, the ablation analysis on our hyperparameters is provided in Appendix. + +Effectiveness of TopoMDCs. To validate the effectiveness of the proposed TopoMDCs, we compare the network performance before and after applying TopoMDCs to the baseline or removing them from the TopoTTA. As shown in Table 3, TopoMDCs result in significant improvements across all metrics compared to the baseline that only updates statistical parameters, validating the effectiveness of TopoMDCs in adapting to various topological structures. Additionally, Table 4 provides a detailed analysis of the contribution of each type of TopoMDCs to performance enhancement. We categorize TopoMDCs into three types: central $(\mathcal{C}_c)$ , orthogonal $(\mathcal{C}_{1-4})$ , and diagonal $(\mathcal{C}_{5-8})$ , and evaluate their impact individually. Results indicate that orthogonal and diagonal types relative to central lead to more significant performance improvements, with optimal topological awareness achieved when orthogonal and diagonal types are used together. We also visualize feature maps before and after incorporating TopoMDCs. Fig. 4 illustrates that in all four scenarios, TopoMDCs effectively enhance the topological structures on the target domain. + +Effectiveness of TopoHG. Here, we aim to answer two research questions: (1) Can TopoHG truly create pseudobreak predictions? (2) Is the low-frequency swap method in TopoHG more effective for enhancing topological continuity compared to other data augmentation methods? + +As shown in Table 3, using only Stage 2 with TopoHG still leads to improvements in both segmentation performance and topological continuity. Compared to Stage 1 alone, adding TopoHG notably enhances continuity. To answer the first question, we provide visualizations of selected + +![](images/f1f61249b925a28412e30172a9a87c80725479ea92ff297ae007b5f904cf3615.jpg) +Figure 4. Feature maps before and after TopoMDCs across four scenarios. VC represents the feature maps of vanilla convolutions. + +![](images/56c7784fb1c2d14b4fd348e00e3f61e4c846b472ea66deb5c03f83c081fb03fc.jpg) +Figure 5. Visualizations of original patches $\pmb{x}_{p}^{\mathrm{fg}}$ , pseudo-labels $\hat{\pmb{y}}_p'$ , pseudo-break patches $\pmb{x}_p'$ , and weakened prediction masks $\pmb{y}_p'$ in selected regions across four scenarios. + +regions' predictions before and after introducing pseudo-breaks in Fig. 5. The images generated by TopoHG significantly disrupt the model's previously confident predictions, effectively creating local pseudo-breaks. For the second question, Table 5 compares various data augmentation methods, including Gaussian blur, random Gaussian noise, and foreground swap with the background in the spatial domain (image swap). The blur method creates local hard samples by removing high-frequency details, which offers some improvement. The image swap method seems to yield superior topological performance, but a sharp drop in Dice indicates that lower Betti errors may stem from overly aggressive predictions connecting many regions incorrectly. This is likely due to consistent regularization on the background without high-frequency details, which can cause incorrect segmentation of the background as foreground. Due to the limited space, visual comparisons of different data augmentation methods are provided in Appendix. + +# 5. Conclusion + +In this work, we proposed the first TTA framework, i.e., TopoTTA, for TSS to enhance cross-domain performance by addressing discrepant topological structures and fragile topological continuity. To improve the adaptability to topological structure, we proposed TopoMDCs to learn how to adaptively integrate various orientation information according to the topological feature of each test sample. Besides, we developed TopoHG to generate pseudo-breaks to enhance the predictions' topological continuity. Extensive experiments across four scenarios verified the superiority of our TopoTTA compared to other methods. + +Acknowledgement. This research was supported by the National Natural Science Foundation of China (Grant No. 62202027). + +# References + +[1] Cheng Chen, Quande Liu, Yueming Jin, Qi Dou, and Pheng-Ann Heng. Source-Free Domain Adaptive Fundus Image Segmentation with Denoised Pseudo-Labeling. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Interventions, pages 225–235, 2021. 1 +[2] Zhuangzhuang Chen, Zhuonan Lai, Jie Chen, and Jianqiang Li. Mind Marginal Non-Crack Regions: Clustering-Inspired Representation Learning for Crack Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12698-12708, 2024. 3 +[3] Ziyang Chen, Yongsheng Pan, Yiwen Ye, Mengkang Lu, and Yong Xia. Each Test Image Deserves a Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11184-11193, 2024. 2, 3, 6, 4, 5 +[4] Guangliang Cheng, Ying Wang, Shibiao Xu, Hongzhen Wang, Shiming Xiang, and Chunhong Pan. Automatic Road Detection and Centerline Extraction via Cascaded End-to-End Convolutional Neural Network. IEEE Transactions on Geoscience and Remote Sensing, 55(6):3322-3337, 2017. 6 +[5] Lechao Cheng, Chaowei Fang, Dingwen Zhang, Guanbin Li, and Gang Huang. Compound Batch Normalization for Long-Tailed Image Classification. In Proceedings of the ACM International Conference on Multimedia, pages 1925-1934, 2022. 2 +[6] Mingfei Cheng, Kaili Zhao, Xuhong Guo, Yajing Xu, and Jun Guo. Joint Topology-Preserving and Feature-Refinement Network for Curvilinear Structure Segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 7147–7156, 2021. 1 +[7] James W. Cooley and John W. Tukey. An Algorithm for the Machine Calculation of Complex Fourier Series. Mathematics of Computation, 19(90):297-301, 1965. 5 +[8] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. DeepGlobe 2018: A Challenge to Parse the Earth Through Satellite Images. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 172-181, 2018. 6 +[9] Jiahua Dong, Yang Cong, Gan Sun, Bineng Zhong, and Xiaowei Xu. What Can Be Transferred: Unsupervised Domain Adaptation for Endoscopic Lesions Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4023-4032, 2020. 1 +[10] Mario Döbler, Robert A Marsden, and Bin Yang. Robust Mean Teacher for Continual and Gradual Test-Time Adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7704-7714, 2023. 2, 3 + +[11] Zhen Fang, Jie Lu, Feng Liu, and Guangquan Zhang. Semi-Supervised Heterogeneous Domain Adaptation: Theory and Algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1087-1105, 2022. 1 +[12] Muhammad Moazam Fraz, Paolo Remagnino, Andreas Hoppe, Bunyarit Uyyanonvara, Alicja R Rudnicka, Christopher G Owen, and Sarah A Barman. An Ensemble Classification-based Approach Applied to Retinal Blood Vessel Segmentation. IEEE Transactions on Biomedical Engineering, 59(9):2538-2548, 2012. 6 +[13] Mohsen Ghafoorian, Alireza Mehrtash, Tina Kapur, Nico Karssemeijer, Elena Marchiori, Mehran Pesteie, Charles R. G. Guttmann, Frank-Erik de Leeuw, Clare M. Tempany, Bram van Ginneken, Andriy Fedorov, Purang Abolmaesumi, Bram Platel, and William M. Wells. Transfer Learning for Domain Adaptation in MRI: Application in Brain Lesion Segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Interventions, pages 516-524, 2017. 1 +[14] Xin Guo, Wenzhong Tang, Haoran Wang, Jiale Wang, Shuai Wang, Xiaolei Qu, and Xun Lin. MorFormer: morphology-aware transformer for generalized pavement crack segmentation. IEEE Transactions on Intelligent Transportation Systems, 26(6):8219-8232, 2025. 2 +[15] Saumya Gupta, Yikai Zhang, Xiaoling Hu, Prateek Prasanna, and Chao Chen. Topology-Aware Uncertainty for Image Segmentation. In Advances in the Neural Information Processing Systems, page 36, 2024. 1 +[16] AD Hoover, Valentina Kouznetsova, and Michael Goldbaum. Locating Blood Vessels in Retinal Images by Piecewise Threshold Probing of a Matched Filter Response. IEEE Transactions on Medical Imaging, 19(3):203-210, 2000. 6 +[17] Xiaoling Hu. Structure-Aware Image Segmentation with Homotopy Warping. In Advances in the Neural Information Processing Systems, pages 24046-24059, 2022. 3 +[18] Xiaoling Hu, Fuxin Li, Dimitris Samaras, and Chao Chen. Topology-Preserving Deep Image Segmentation. In Advances in the Neural Information Processing Systems, 2019. 6 +[19] Xiaoling Hu, Yusu Wang, Li Fuxin, Dimitris Samaras, and Chao Chen. Topology-Aware Segmentation Using Discrete Morse Theory. In Proceedings of the International Conference on Learning Representations, 2021. 3 +[20] Xiaoling Hu, Dimitris Samaras, and Chao Chen. Learning Probabilistic Topological Representations Using Discrete Morse Theory. In Proceedings of the International Conference on Learning Representations, 2022. 3 +[21] Jiaxing Huang, Yanfeng Zhou, Yaoru Luo, Guole Liu, Heng Guo, and Ge Yang. Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures. In Proceedings of the European Conference on Computer Vision, pages 143-160, 2025. 1, 3 +[22] Klara Janouskova, Tamir Shor, Chaim Baskin, and Jiri Matas. Single Image Test-Time Adaptation for Segmentation. Transactions on Machine Learning Research, 2024. 2 +[23] Yannick Kirchhoff, Maximilian R Rokuss, Saikat Roy, Balint Kovacs, Constantin Ulrich, Tassilo Wald, Maximilian Zenk, Philipp Vollmuth, Jens Kleesiek, Fabian Isensee, + +and Klaus Maier-Hein. Skeleton Recall Loss for Connectivity Conserving and Resource Efficient Segmentation of Thin Tubular Structures. In Proceedings of the European Conference on Computer Vision, pages 218-234, 2024. 3 +[24] Mou Lei, Yitian Zhao, Huazhu Fu, Yonghuai Liu, Jun Cheng, Yalin Zheng, Pan Su, Jianlong Yang, Li Chen, Alejandro F Frangi, Masahiro Akiba, and Jiang Liu. CS2-Net: Deep learning Segmentation of Curvilinear Structures in Medical Imaging. Medical Image Analysis, 67:101874, 2021. 1, 6 +[25] Qin Lei, Jiang Zhong, and Qizhu Dai. Enriching Information and Preserving Semantic Consistency in Expanding Curvilinear Object Segmentation Datasets. In Proceedings of the European Conference on Computer Vision, pages 233-250, 2025. 3 +[26] Mingchao Li, Kun Huang, Qiuzhuo Xu, Jiadong Yang, Yuhan Zhang, Zexuan Ji, Keren Xie, Songtao Yuan, Qinghuai Liu, and Qiang Chen. OCTA-500: A Retinal Dataset for Optical Coherence Tomography Angiography Study. Medical Image Analysis, 93:103092, 2024. 6 +[27] Jian Liang, Ran He, and Tieniu Tan. A Comprehensive Survey on Test-Time Adaptation Under Distribution Shifts. International Journal of Computer Vision, pages 1-34, 2024. 2 +[28] Li Lin, Linkai Peng, Huaqing He, Pujin Cheng, Jiewei Wu, Kenneth K.Y. Wong, and Xiaoying Tang. YoloCurvSeg: You Only Label One Noisy Skeleton for Vessel-Style Curvilinear Structure Segmentation. Medical Image Analysis, 90: 102937, 2023. 3 +[29] Xun Lin, Shuai Wang, Rizhao Cai, Yizhong Liu, Ying Fu, Wenzhong Tang, Zitong Yu, and Alex C. Kot. Suppress and rebalance: Towards generalized multi-modal face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 211-221, 2024. 3 +[30] Xun Lin, Yi Yu, Song Xia, Jue Jiang, Haoran Wang, Zitong Yu, Yizhong Liu, Ying Fu, Shuai Wang, Wenzhong Tang, and Alex C. Kot. Safeguarding medical image segmentation datasets against unauthorized training via contour- and texture-aware perturbations. arXiv preprint arXiv:2403.14250, 2024. 1 +[31] Xun Lin, Yi Yu, Zitong Yu, Ruohan Meng, Jiale Zhou, Ajian Liu, Yizhong Liu, Shuai Wang, Wenzhong Tang, Zhen Lei, and Alex C. Kot. *Hidemia: Hidden wavelet mining for privacy-enhancing medical image analysis*. In *Proceedings of the ACM International Conference on Multimedia*, pages 8110–8119, 2024. 1 +[32] Xun Lin, Ajian Liu, Zitong Yu, Rizhao Cai, Shuai Wang, Yi Yu, Jun Wan, Zhen Lei, Xiaochun Cao, and Alex Kot. Reliable and balanced transfer learning for generalized multimodal face anti-spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025. 3 +[33] Huajun Liu, Xiangyu Miao, Christoph Mertz, Chengzhong Xu, and Hui Kong. Crackformer: Transformer Network for Fine-Grained Crack Detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 3783-3792, 2021. 2 +[34] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. FedDG: Federated Domain Generalization on + +Medical Image Segmentation via Episodic Learning in Continuous Frequency Space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1013-1023, 2021. 5 +[35] Yuang Liu, Wei Zhang, and Jun Wang. Source-Free Domain Adaptation for Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1215-1224, 2021. 1 +[36] Junyan Lyu, Yiqi Zhang, Yijin Huang, Li Lin, Pujin Cheng, and Xiaoying Tang. AADG: Automatic Augmentation for Domain Generalization on Retinal Image Segmentation. IEEE Transactions on Medical Imaging, 41(12):3699-3711, 2022. 1 +[37] Yuhui Ma, Huaying Hao, Jianyang Xie, Huazhu Fu, Jiong Zhang, Jianlong Yang, Zhen Wang, Jiang Liu, Yalin Zheng, and Yitian Zhao. ROSE: A Retinal OCT-Angiography Vessel Segmentation Dataset and New Model. IEEE Transactions on Medical Imaging, 40(3):928-939, 2020. 1, 6 +[38] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The Norm Must Go On: Dynamic Unsupervised Domain Adaptation by Normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 3 +[39] Volodymyr Mnih. Machine Learning for Aerial Image Labeling. University of Toronto, Canada, 2013. 6 +[40] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip H.S. Torr. Tipi: Test Time Adaptation with Transformation Invariance. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 24162-24171, 2023. 2 +[41] Jiayi Ni, Senqiao Yang, Ran Xu, Jiaming Liu, Xiaoqi Li, Wenyu Jiao, Zehui Chen, Yi Liu, and Shanghang Zhang. Distribution-Aware Continual Test-Time Adaptation for Semantic Segmentation. In Proceedings of International Conference on Robotics and Automation, pages 3044-3050, 2024. 3 +[42] Munan Ning, Donghuan Lu, Yujia Xie, Dongdong Chen, Dong Wei, Yefeng Zheng, Yonghong Tian, Shuicheng Yan, and Li Yuan. MADAv2: Advanced Multi-Anchors based Active Domain Adaptation Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11): 13553-13566, 2023. 1 +[43] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards Stable Test-Time Adaptation in Dynamic Wild World. In Proceedings of the International Conference on Learning Representations, 2023. 2, 3, 6, 4, 5 +[44] Hyejin Park, Jeongyeon Hwang, Sunung Mun, Sangdon Park, and Jungseul Ok. MedBN: Robust Test-Time Adaptation Against Malicious Test Samples. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5997-6007, 2024. 3, 6, 4, 5 +[45] Bibiloni Pedro, González-Hidalgo M, and Massanet Sebastià. A Survey on Curvilinear Object Segmentation in Multiple Applications. Pattern Recognition, 60:949-970, 2016. 1 +[46] Yaolei Qi, Yuting He, Xiaoming Qi, Yuan Zhang, and Guanyu Yang. Dynamic Snake Convolution Based on Topo + +logical Geometric Constraints for Tubular Structure Segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 6070-6079, 2023. 1, 2, 6 +[47] Xiaolei Qu, Jiale Zhou, Jue Jiang, Wenhan Wang, Haoran Wang, Shuai Wang, Wenzhong Tang, and Xun Lin. Ehformer: Regional easy-hard-aware transformer for breast lesion segmentation in ultrasound images. Information Fusion, 109:102430, 2024. 1 +[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Interventions, pages 234–241, 2015. 6 +[49] Jameel Hassan Abdul Samadh, Hanan Gani, Noor Hazim Hussein, Muhammad Uzair Khattak, Muzammal Naseer, Fahad Khan, and Salman Khan. Align Your Prompts: Test-Time Prompting with Distribution Alignment for Zero-Shot Generalization. In Advances in the Neural Information Processing Systems, 2023. 2 +[50] Tianyi Shi, Nicolas Boutry, Yongchao Xu, and Thierry Géraud. Local Intensity Order Transformation for Robust Curvilinear Object Segmentation. IEEE Transactions on Image Processing, 31:2557-2569, 2022. 1, 2 +[51] Tianyi Shi, Xiaohuan Ding, Liang Zhang, and Xin Yang. FreeCOS: Self-Supervised Learning from Fractals and Unlabeled Images for Curvilinear Object Segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 876-886, 2023. 3 +[52] Suprosanna Shit, Johannes C Paetzold, Anjany Sekuboyina, Ivan Ezhov, Alexander Unger, Andrey Zhylka, Josien P.W. Plum, Ulrich Bauer, and Bjoern H Menze. cIDice: A Novel Topology-Preserving Loss Function for Tubular Structure Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16560-16569, 2021. 3, 6 +[53] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. ECoTTA: Memory-Efficient Continual Test-Time Adaptation via Self-Distilled Regularization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11920-11929, 2023. 2, 3 +[54] Joes Staal, Michael D Abramoff, Meindert Niemeijer, Max A Viergever, and Bram Van Ginneken. Ridge-based Vessel Segmentation in Color Images of the Retina. IEEE Transactions on Medical Imaging, 23(4):501-509, 2004. 6 +[55] Yubo Tan, Kaifu Yang, Shixuan Zhao, and Yongjie Li. Retinal Vessel Segmentation with Skeletal Prior and Contrastive Loss. IEEE Transactions on Medical Imaging, 41(9):2238-2251, 2022. 1 +[56] Devavrat Tomar, Guillaume Vray, Behzad Bozorgtabar, and Jean-Philippe Thiran. TeSLA: Test-Time Self-Learning with Automatic Adversarial Augmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20341–20350, 2023. 3 +[57] Chenglong Wang, Yuichiro Hayashi, Masahiro Oda, Hayato Itoh, Takayuki Kitasaka, Alejandro F Frangi, and Kensaku Mori. Tubular Structure Segmentation Using Spatial Fully Connected Network with Radial Distance Loss for 3D Medical Images. In Proceedings of the International Conference + +on Medical Image Computing and Computer Assisted Interventions, pages 348-356, 2019. 3 +[58] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. TENT: Fully Test-Time Adaptation by Entropy Minimization. Proceedings of the International Conference on Learning Representations, 2021. 1, 2, 3, 5, 6, 4 +[59] Dong Wang, Zhao Zhang, Ziwei Zhao, Yuhang Liu, Yihong Chen, and Liwei Wang. PointScatter: Point Set Representation for Tubular Structure Extraction. In Proceedings of the European Conference on Computer Vision, pages 366-383, 2022. 2, 6 +[60] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual Test-Time Domain Adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 1, 2, 3, 6, 4, 5 +[61] Wei Wang, Zhun Zhong, Weijie Wang, Xi Chen, Charles Ling, Boyu Wang, and Nicu Sebe. Dynamically Instance-Guided Adaptation: A Backward-Free Approach for Test-Time Domain Adaptive Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 24090-24099, 2023. 2, 3, 6, 4 +[62] Wenhan Wang, Jiale Zhou, Jin Zhao, Xun Lin, Yan Zhang, Shan Lu, Wanchen Zhao, Shuai Wang, Wenzhong Tang, and Xiaolei Qu. Interactively fusing global and local features for benign and malignant classification of breast ultrasound images. Ultrasound in Medicine Biology, 51(3):525-534, 2025. 1 +[63] Yan Wang, Xu Wei, Fengze Liu, Jieneng Chen, Yuyin Zhou, Wei Shen, Elliot K. Fishman, and Alan L. Yuille. Deep Distance Transform for Tubular Structure Segmentation in CT Scans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 3 +[64] Ying Wang, Yuexing Peng, Wei Li, George C Alexandropoulos, Junchuan Yu, Daqing Ge, and Wei Xiang. DDU-Net: Dual-Decoder-U-Net for Road Extraction Using High-Resolution Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-12, 2022. 1 +[65] Qian Wu, Yufei Chen, Wei Liu, Xiaodong Yue, and Xiahai Zhuang. Deep Closing: Enhancing Topological Connectivity in Medical Tubular Segmentation. IEEE Transactions on Medical Imaging, 43(11):3990-4003, 2024. 1, 6 +[66] Jiaqi Yang, Xiaoling Hu, Chao Chen, and Chialing Tsai. A Topological-Attention ConvLSTM Network and Its Application to EM Images. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Interventions, pages 217–228, 2021. 2 +[67] Yanchao Yang and Stefano Soatto. FDA: Fourier Domain Adaptation for Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4085-4095, 2020. 1 +[68] Ziyun Yang and Sina Farsiu. Directional Connectivity-based Segmentation of Medical Images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11525-11535, 2023. 2, 6 +[69] Chang'an Yi, Haotian Chen, Yifan Zhang, Yonghui Xu, Yan Zhou, and Lizhen Cui. From Question to Exploration: + +Can Classic Test-Time Adaptation Strategies Be Effectively Applied in Semantic Segmentation? In Proceedings of the ACM International Conference on Multimedia, pages 10085-10094, 2024. 2, 3 +[70] Zitong Yu, Chenxu Zhao, Zezheng Wang, Yunxiao Qin, Zhuo Su, Xiaobai Li, Feng Zhou, and Guoying Zhao. Searching Central Difference Convolutional Networks for Face Anti-Spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5295-5305, 2020. 2, 3 +[71] Longhui Yuan, Binhui Xie, and Shuang Li. Robust Test-Time Adaptation in Dynamic Scenarios. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 15922-15932, 2023. 2 +[72] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. DomainAdaptor: A Novel Approach to Test-Time Adaptation. In Proceedings of the IEEE International Conference on Computer Vision, pages 18971-18981, 2023. 2, 3, 6, 4, 5 +[73] Lefei Zhang, Meng Lan, Jing Zhang, and Dacheng Tao. Stagewise Unsupervised Domain Adaptation with Adversarial Self-Training for Road Segmentation of Remote-Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2021. 1 +[74] Yizhe Zhang, Tao Zhou, Yuhui Tao, Shuo Wang, Ye Wu, Benyuan Liu, Pengfei Gu, Qiang Chen, and Danny Z Chen. TestFit: A Plug-and-Play One-Pass Test Time Method for Medical Image Segmentation. Medical Image Analysis, 92: 103069, 2024. 2 +[75] Bowen Zhao, Chen Chen, and Shu-Tao Xia. DELTA: Degradation-Free Fully Test-Time Adaptation. In Proceedings of the International Conference on Learning Representations, 2023. 3 +[76] He Zhao, Huiqi Li, Sebastian Maurer-Stroh, and Li Cheng. Synthesizing Retinal and Neuronal Images with Generative Adversarial Nets. Medical Image Analysis, 49:14-26, 2018. 6 +[77] Jiale Zhou, Zuoxun Hou, Hongyan Lu, Wenhan Wang, Wanchen Zhao, Zenan Wang, Dezhi Zheng, Shuai Wang, Wenzhong Tang, and Xiaolei Qu. A deep supervised transformer u-shaped full-resolution residual network for the segmentation of breast ultrasound image. Medical Physics, 50 (12):7513-7524, 2023. 1 \ No newline at end of file diff --git a/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/images.zip b/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f79c0da7a43f6a653d5f70ab837a8e97cdb6894f --- /dev/null +++ b/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fecd170f37cee64667433df82f0b54a5322d1b7294b16bdfa973a471c268fbcc +size 813488 diff --git a/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/layout.json b/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6e1de25ad978a03675efb0b59d1f90d32f213019 --- /dev/null +++ b/2025/TopoTTA_ Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation/layout.json @@ -0,0 +1,10404 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 481, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 481, + 139 + ], + "type": "text", + "content": "TopoTTA: Topology-Enhanced Test-Time Adaptation for Tubular Structure Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "spans": [ + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "content": "Jiale Zhou" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "content": ", Wenhan Wang" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "content": ", Shikun Li" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "content": ", Xiaolei Qu" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "content": ", Xin Guo" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "content": ", Yizhong Liu" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "content": ", Wenzhong Tang" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "content": ", Xun Lin" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "inline_equation", + "content": "^{2,3,*}" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "content": ", Yefeng Zheng" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "content": "Zhejiang University " + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "content": "Westlake University " + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 160, + 496, + 217 + ], + "type": "text", + "content": "Beihang University \n{zhoujiale, zhengyefeng}@westlake.edu.cn linxun@buaa.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 245, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 245, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 245, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 296, + 582 + ], + "type": "text", + "content": "Tubular structure segmentation (TSS) is important for various applications, such as hemodynamic analysis and route navigation. Despite significant progress in TSS, domain shifts remain a major challenge, leading to performance degradation in unseen target domains. Unlike other segmentation tasks, TSS is more sensitive to domain shifts, as changes in topological structures can compromise segmentation integrity, and variations in local features distinguishing foreground from background (e.g., texture and contrast) may further disrupt topological continuity. To address these challenges, we propose Topology-enhanced Test-Time Adaptation (TopoTTA), the first test-time adaptation framework designed specifically for TSS. TopoTTA consists of two stages: Stage 1 adapts models to cross-domain topological discrepancies using the proposed Topological Meta Difference Convolutions (TopoMDCs), which enhance topological representation without altering pre-trained parameters; Stage 2 improves topological continuity by a novel Topology Hard sample Generation (TopoHG) strategy and prediction alignment on hard samples with pseudo-labels in the generated pseudobreaks. Extensive experiments across four scenarios and ten datasets demonstrate TopoTTA's effectiveness in handling topological distribution shifts, achieving an average improvement of " + }, + { + "bbox": [ + 53, + 270, + 296, + 582 + ], + "type": "inline_equation", + "content": "31.81\\%" + }, + { + "bbox": [ + 53, + 270, + 296, + 582 + ], + "type": "text", + "content": " in clDice. TopoTTA also serves as a plug-and-play TTA solution for CNN-based TSS models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "type": "text", + "content": "Tubular structure segmentation (TSS), also known as curvilinear object segmentation [45], plays a critical role in various applications (e.g., hemodynamic analysis [24, 55] and route planning [64, 73]). Although many TSS methods have been proposed and achieved inspiring performance [15, 21], domain shifts remain a common challenge" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 328, + 244, + 539, + 434 + ], + "blocks": [ + { + "bbox": [ + 328, + 244, + 539, + 434 + ], + "lines": [ + { + "bbox": [ + 328, + 244, + 539, + 434 + ], + "spans": [ + { + "bbox": [ + 328, + 244, + 539, + 434 + ], + "type": "image", + "image_path": "9a7bccbfc9c7800d0f0a7cd0e86281b4f7acb1540d22a575ef1fc4b99911dd4a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 437, + 555, + 515 + ], + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 515 + ], + "type": "text", + "content": "Figure 1. Comparison between TopoTTA and existing TTA solutions. (a) represents the model during the source domain training phase, while (b) denotes the model in the target domain testing phase. (b1-b4) show the results of directly testing the source-trained model, the TENT [58] method with updated BN layers, the CoTTA [60] method using a teacher-student scheme, and our topology-enhanced approach, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 545 + ], + "type": "text", + "content": "in real-world scenarios [42] due to discrepancies in imaging devices [1, 13] and sample heterogeneity [11]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "Domain shifts undermine the effectiveness of TSS models trained on the source domain when applied to an unseen target domain [36, 50]. Unlike segmentation models trained for other tasks, e.g., semantic [35, 67] or tumor segmentation [9, 30, 31, 47, 77], TSS models are more sensitive to domain shifts [50]. Beyond the common variations such as contrast and noise distribution [62, 63], TSS faces additional challenges related to unseen topological structural characteristics, including variations in trajectory, curvature, branching patterns, and thickness [6, 46]. These factors degrade segmentation performance and even disrupt the topological continuity of the structures [65] (see Fig. 1 (b1)), which may mislead critical downstream applications [37, 55, 64]." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 148, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24123" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "Recently, a special domain adaptation (DA) paradigm, namely test-time adaptation (TTA) [27, 40, 43], has been proposed and successfully applied to various image segmentation tasks, including semantic [22, 60, 61] and medical image segmentation [3, 74]. Unlike other DA paradigms, TTA performs continuous online adaptation during the test stage (after model deployment) [71]. It uses extremely small amounts of data (e.g., a single sample or batch), aligning better with real-world scenarios where data must be processed sequentially [58]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 295, + 384 + ], + "type": "text", + "content": "Most TTA methods update pre-trained models using self-supervised methods [27], which can typically be categorized into normalization-based methods [58, 72] and teacher-student schemes [10, 60]. Normalization-based methods update normalization layer statistics [53] (e.g., batch/layer normalization (BN/LN)) or leverage the difference in normalization statistics before and after adaptation to update external model parameters [3, 49]. These methods are initially designed for classification tasks, and their performance may degrade when applied to segmentation tasks with severe pixel-level class imbalance [5]. The second category, teacher-student schemes (based on pseudolabels), updates all model parameters and achieves performance gains in general segmentation tasks [69]. However, as shown in Fig. 1, these methods perform poorly in TSS due to two unique challenges in cross-domain scenarios:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 387, + 295, + 555 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 387, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 295, + 470 + ], + "type": "text", + "content": "- Challenge 1: Discrepant topological structures. Tubular structures in the source and target domains can exhibit substantial topological differences (e.g., thickness, curvature, and branching patterns). Prior TTA methods' general-purpose model adaptation strategies struggle to effectively capture these cross-domain topological variations when adapting to individual samples." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 471, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 555 + ], + "type": "text", + "content": "- Challenge 2: Fragile topological continuity. In cross-domain settings, the local characteristics (e.g., color, texture, and contrast) distinguishing foreground from background may change significantly [50]. Existing TTA methods lack targeted adaptation to these local variations, resulting in catastrophic disruptions to topological continuity (see Fig. 1) by misclassifying foreground pixels." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 555, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 616 + ], + "type": "text", + "content": "To address the aforementioned challenges, we propose Topology-Enhanced Test-Time Adaptation (TopoTTA) in this paper. It includes two adaptation stages: topological structure adaptation (Stage 1) and topological continuity refinement (Stage 2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Stage 1 aims to enhance the model's capability to adapt to discrepant topological structures by utilizing a topology-perceiving model adaptation strategy. Inspired by the Central Difference Convolution (CDC) [70], which captures differences between the central pixel and its neighbors to enhance local representations, we extend this concept to accommodate the elongated and diverse trajectories of tubular structures. Specifically, we introduce Topological Meta" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "Difference Convolutions (TopoMDCs), which perform a dual-pixel difference operation by extending the central pixel's interaction across eight directional neighbors. As shown in Fig. 2, we learn weighted combinations of different directional TopoMDCs (instead of the vanilla convolution) for different regions to enhance the representations of hard-to-detect tubular branches, improving adaptability to discrepant topological structures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 168, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 276 + ], + "type": "text", + "content": "Within Stage 2, we design a Topology Hard sample Generation (TopoHG) strategy to construct samples with pseudo-breaks. These pseudo-breaks mislead the models to make weakened predictions with poor topological continuity. We then emphasize the local consistency between the predictions and original pseudo-labels in these hard regions. This stage encourages the models to adapt the characteristics distinguishing foreground from background and helps address continuity destructions in the target domain." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 277, + 460, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 277, + 460, + 288 + ], + "spans": [ + { + "bbox": [ + 325, + 277, + 460, + 288 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 290, + 553, + 506 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 314, + 290, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 290, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 314, + 290, + 553, + 326 + ], + "type": "text", + "content": "- We propose the first TTA framework for tubular structure segmentation, named TopoTTA, which effectively adapts to unseen domains by enhancing topological information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 327, + 553, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 327, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 315, + 327, + 553, + 386 + ], + "type": "text", + "content": "- We design eight directional Topological Meta Difference Convolutions (TopoMDCs) within TopoTTA, and learn how to adaptively combine different TopoMDCs based on the topological features of the test sample. TopoMDCs enhance adaptability to unseen topological structures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 386, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 386, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 315, + 386, + 553, + 456 + ], + "type": "text", + "content": "- We propose a Topology Hard sample Generation strategy for TopoTTA, namely TopoHG, which generates pseudobreaks at confidently predicted regions. We then align the predictions around these pseudo breaks to the original pseudo labels, improving the topological continuity in the unseen domain." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 458, + 553, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 458, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 315, + 458, + 553, + 506 + ], + "type": "text", + "content": "- Extensive experiments on four scenarios across ten datasets validate the effectiveness of TopoTTA in addressing topological distribution shifts, achieving an average performance improvement of " + }, + { + "bbox": [ + 315, + 458, + 553, + 506 + ], + "type": "inline_equation", + "content": "31.81\\%" + }, + { + "bbox": [ + 315, + 458, + 553, + 506 + ], + "type": "text", + "content": " in cI Dice." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 519, + 406, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 519, + 406, + 531 + ], + "spans": [ + { + "bbox": [ + 314, + 519, + 406, + 531 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 540, + 489, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 489, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 489, + 552 + ], + "type": "text", + "content": "2.1. Tubular Structure Segmentation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 557, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 553, + 713 + ], + "type": "text", + "content": "Tubular structures, characterized by their thin and long topological structures as well as complex layouts, usually cannot be accurately and continuously localized by general segmentation methods [59]. To tackle these challenges, researchers have incorporated topological priors into model designs, including specialized deformable convolution kernels [46], attention mechanisms [14, 66], and tailored model architectures [33]. Complementary to these module innovations, PointScatter [59] was introduced to represent tubular structures as point sets, addressing the flexibility limitations of masks constrained by fixed grids. DconnNet [68] employed a unique design to separate directional subspaces from the latent space, enhancing feature representation for" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24124" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "connectivity. Meanwhile, a growing body of works focused on novel loss functions [2, 17, 19, 20, 23, 52, 57] that explicitly optimize the topological continuity of segmentation results. More recently, some research has started to explore the use of synthetic data [25, 28, 51] and the integration of additional topological information [21, 63] to further improve performance. However, these TSS methods ignore the existence of domain shifts, resulting in suboptimal performance when applied to unseen domains." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 186, + 180, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 180, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 180, + 198 + ], + "type": "text", + "content": "2.2. Test-Time Adaptation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 202, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 296, + 586 + ], + "type": "text", + "content": "TTA adaptively adjusts the model at test time based on a small number of unlabeled samples from a different distribution. Early TTA methods primarily modified model representations by adjusting BN's statistics [38, 44, 61, 75]. Mirza et al. [38] adapted target domain features by updating BN statistics continuously. Most TTA methods used self-supervised learning to adjust pre-trained parameters, including normalization-based methods [3, 43, 53, 58, 72] and teacher-student (TS) schemes [10, 41, 56, 60]. Wang et al. [58] proposed TENT, an entropy minimization method that updates BN affine transformation parameters to improve confidence in the target domain. Motivated by TENT, some recent works achieved better performance [3, 72] by designing various strategies to fine-tune BN. Zhang et al. [72] incorporated source domain statistics and extended the entropy minimization loss to utilize target domain information better. More recently, Chen et al. [3] reduced generalization errors by aligning source and target statistics to update external model parameters. However, many TTA methods are designed for classification tasks, ignoring pixel-wise class imbalance. In segmentation tasks, where sample imbalance is more severe, simple statistical adjustments and minor parameter updates are easily dominated by the majority of background pixels, reducing segmentation performance [69]. The TS scheme, which applies consistency regularization to update all model parameters, has shown effectiveness in general segmentation tasks. CoTTA [60] used weight-averaged and augmentation-averaged predictions, further reducing error accumulation in continuous inference. However, these methods lack a focus on tubular structure and topological continuity, leading to performance decline and continuity destruction in TSS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 170, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 170, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 170, + 608 + ], + "type": "text", + "content": "3. Proposed TopoTTA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 614, + 135, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 135, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 135, + 626 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^s = \\{(\\pmb{x}_i^s,\\pmb{y}_i^s)\\}_{i = 1}^{N^s}" + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "text", + "content": " denote the labeled source domain dataset, where " + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^s\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "text", + "content": " is an image with height " + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "text", + "content": " width " + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "text", + "content": ", and channel number " + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "inline_equation", + "content": "\\pmb{y}_i^s\\in \\{0,1\\}^{H\\times W\\times C}" + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "text", + "content": " denotes the corresponding ground-truth label. Given a TSS model " + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot ;\\theta^s)" + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "text", + "content": " trained on " + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^s" + }, + { + "bbox": [ + 55, + 631, + 295, + 692 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 695, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 695, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 80, + 695, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\theta^ {s} = \\underset {\\theta^ {s}} {\\arg \\min } \\mathbb {E} _ {\\left(\\boldsymbol {x} ^ {s}, \\boldsymbol {y} ^ {s}\\right) \\sim \\mathcal {D} ^ {s}} \\left[ \\mathcal {L} _ {s e g} \\left(\\mathcal {F} \\left(\\boldsymbol {x} ^ {s}; \\theta^ {s}\\right), \\boldsymbol {y} ^ {s}\\right) \\right], \\tag {1}", + "image_path": "42a87c93db405b2b14b13624f81a1c7380eff64f8f341663fd1ae920e9e1fc1f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 135 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 135 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 135 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{seg}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 135 + ], + "type": "text", + "content": " can be any TSS loss function. Let " + }, + { + "bbox": [ + 313, + 72, + 553, + 135 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t = \\{\\pmb {x}_i^t\\}_{i = 1}^{N^t}" + }, + { + "bbox": [ + 313, + 72, + 553, + 135 + ], + "type": "text", + "content": " represent the unlabeled target domain dataset. During testing, TTA methods aim to update the model parameters for each input image " + }, + { + "bbox": [ + 313, + 72, + 553, + 135 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^t\\in \\mathcal{D}^t" + }, + { + "bbox": [ + 313, + 72, + 553, + 135 + ], + "type": "text", + "content": " and produce segmentation outputs " + }, + { + "bbox": [ + 313, + 72, + 553, + 135 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_i^t = \\mathcal{F}(\\pmb{x}_i^t;\\theta_{i - 1}^t)" + }, + { + "bbox": [ + 313, + 72, + 553, + 135 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 367, + 140, + 553, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 140, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 367, + 140, + 553, + 163 + ], + "type": "interline_equation", + "content": "\\theta_ {i} ^ {t} = \\underset {\\theta_ {i - 1} ^ {t}} {\\arg \\min } \\mathcal {L} _ {t t a} \\left(\\mathcal {F} \\left(\\cdot ; \\theta_ {i - 1} ^ {t}\\right), \\boldsymbol {x} _ {i} ^ {t}\\right), \\tag {2}", + "image_path": "e817e997dffa8de7b687690b8ecbdc2ec3548745461e264383b356399c20bfbb.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "type": "inline_equation", + "content": "\\theta_i^t" + }, + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "type": "text", + "content": " are the parameters of the TSS model " + }, + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot ;\\cdot)" + }, + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "type": "text", + "content": "-th iteration, and " + }, + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{tta}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "type": "text", + "content": " denotes the self-supervised loss function for TTA. Before performing TTA, we set " + }, + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "type": "inline_equation", + "content": "\\theta_0^t = \\theta^s" + }, + { + "bbox": [ + 313, + 170, + 553, + 206 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 206, + 553, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 553, + 230 + ], + "type": "text", + "content": "To improve readability, we use the simplified notation " + }, + { + "bbox": [ + 313, + 206, + 553, + 230 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 206, + 553, + 230 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 313, + 206, + 553, + 230 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i^t" + }, + { + "bbox": [ + 313, + 206, + 553, + 230 + ], + "type": "text", + "content": " in the following sections and figures." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 236, + 441, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 441, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 441, + 249 + ], + "type": "text", + "content": "3.2. Overview of TopoTTA" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 254, + 555, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 398 + ], + "type": "text", + "content": "As discussed in Section 1, existing TTA methods cannot effectively segment tubular structures due to two major challenges: discrepant topological structures and fragile topological continuity. To solve these issues, we propose TopoTTA, which consists of two stages: Stage 1 performs topological structure adaptation to overcome the insufficient cross-domain topological structure adaptation; Stage 2 conducts topological continuity refinement to enhance the topological continuity of the final predictions. An overview of our TopoTTA is illustrated in Fig. 2. The details of Stage 1 and Stage 2 are described in Section 3.3 and Section 3.4, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 404, + 535, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 535, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 535, + 417 + ], + "type": "text", + "content": "3.3. Stage 1: Topological Structure Adaptation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 422, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 506 + ], + "type": "text", + "content": "This stage aims to adapt the pre-trained TSS model to discrepant topological structures caused by domain shifts. To achieve this, we propose TopoMDCs based on diverse fundamental topological patterns and adaptively reweight and combine TopoMDCs according to the topological features of different regions. We introduce the details of TopoMDCs and the model parameter updating process below." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 508, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 664 + ], + "type": "text", + "content": "Topology-Meta Difference Convolutions (TopoMDCs). TopoMDCs draw inspiration from central difference convolution (CDC) [70], which captures the difference between the central pixel of a convolution kernel and its neighbors to detect local gradient changes. However, the local topological shapes of tubular structures often exhibit directionality and continuity. CDC, initially designed for other fine-grained vision tasks (e.g., face anti-spoofing [29, 32, 70]), lacks the ability to perceive these two features of tubular structures, making them ineffective in representing the complex topologies and geometric constraints within local regions. Let " + }, + { + "bbox": [ + 313, + 508, + 555, + 664 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{in}}" + }, + { + "bbox": [ + 313, + 508, + 555, + 664 + ], + "type": "text", + "content": " denote the input feature map. The vanilla convolution " + }, + { + "bbox": [ + 313, + 508, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_0" + }, + { + "bbox": [ + 313, + 508, + 555, + 664 + ], + "type": "text", + "content": " and CDC " + }, + { + "bbox": [ + 313, + 508, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_c" + }, + { + "bbox": [ + 313, + 508, + 555, + 664 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 667, + 538, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 667, + 538, + 715 + ], + "spans": [ + { + "bbox": [ + 320, + 667, + 538, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {C} _ {0} (r _ {x}, r _ {y}) = \\sum_ {(\\Delta r _ {x}, \\Delta r _ {y}) \\in \\mathcal {R}} w (\\Delta r _ {x}, \\Delta r _ {y}) \\cdot \\pmb {x} _ {\\mathrm {i n}} (r _ {x} - \\Delta r _ {x}, r _ {y} - \\Delta r _ {y}), \\\\ \\mathcal{C}_{c}(r_{x},r_{y}) = \\sum_{(\\Delta r_{x},\\Delta r_{y})\\in \\mathcal{R}}w(\\Delta r_{x},\\Delta r_{y})\\cdot \\boldsymbol{x}_{\\mathrm{in}}(r_{x},r_{y}), \\\\ \\end{array}", + "image_path": "81375690653415443cc69bbd02610c6c772146c8b299fd6303a0535f43b14b76.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24125" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 83, + 395, + 193 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 214, + 83 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 214, + 83 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 214, + 83 + ], + "type": "text", + "content": "(a) Stage 1: Topological Structure Adaptation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 83, + 395, + 193 + ], + "lines": [ + { + "bbox": [ + 60, + 83, + 395, + 193 + ], + "spans": [ + { + "bbox": [ + 60, + 83, + 395, + 193 + ], + "type": "image", + "image_path": "ae5dbb41cbc112e71d559124140fdb6b81e4f67bbf450942b769c405f6928c10.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 405, + 84, + 548, + 297 + ], + "blocks": [ + { + "bbox": [ + 405, + 73, + 489, + 84 + ], + "lines": [ + { + "bbox": [ + 405, + 73, + 489, + 84 + ], + "spans": [ + { + "bbox": [ + 405, + 73, + 489, + 84 + ], + "type": "text", + "content": "(c) TopoHG (in Stage 2)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 405, + 84, + 548, + 297 + ], + "lines": [ + { + "bbox": [ + 405, + 84, + 548, + 297 + ], + "spans": [ + { + "bbox": [ + 405, + 84, + 548, + 297 + ], + "type": "image", + "image_path": "a3edc90b2f535911d2f66a84eb8fc64e8bc35d1903870c4d19d40b66a0f9b5d5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 209, + 266, + 291 + ], + "blocks": [ + { + "bbox": [ + 61, + 198, + 220, + 209 + ], + "lines": [ + { + "bbox": [ + 61, + 198, + 220, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 220, + 209 + ], + "type": "text", + "content": "(b) Stage 2: Topological Continuity Refinement" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 209, + 266, + 291 + ], + "lines": [ + { + "bbox": [ + 62, + 209, + 266, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 209, + 266, + 291 + ], + "type": "image", + "image_path": "2a5c0669467da68bedad559d3d8434723bd693c1527fc979a4c1e64ae7a3d341.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 270, + 198, + 548, + 297 + ], + "blocks": [ + { + "bbox": [ + 270, + 198, + 548, + 297 + ], + "lines": [ + { + "bbox": [ + 270, + 198, + 548, + 297 + ], + "spans": [ + { + "bbox": [ + 270, + 198, + 548, + 297 + ], + "type": "image", + "image_path": "a73c430afdc39bcaa88c982d220176019963c7ab7b7548228fd78da11e7acd7d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 298, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 298, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 555, + 332 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed TopoTTA. TopoTTA consists of two stages: (a) Stage 1: Topology structure adaptation with Topological Meta Difference Convolutions (TopoMDCs), and (b) Stage 2: Topological continuity refinement. (c) Topology Hard sample Generation (TopoHG) constructs challenging samples with local pseudo-breaks. CDC is short for central difference convolution." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 337, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 295, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 337, + 295, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{R} = \\{(1,1),(0,1),\\dots ,(-1,0),(-1, - 1)\\}" + }, + { + "bbox": [ + 55, + 337, + 295, + 396 + ], + "type": "text", + "content": " represents the local receptive field of a " + }, + { + "bbox": [ + 55, + 337, + 295, + 396 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 55, + 337, + 295, + 396 + ], + "type": "text", + "content": " convolution kernel, " + }, + { + "bbox": [ + 55, + 337, + 295, + 396 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 337, + 295, + 396 + ], + "type": "text", + "content": " denotes the parameters of the vanilla convolution layer, and " + }, + { + "bbox": [ + 55, + 337, + 295, + 396 + ], + "type": "inline_equation", + "content": "(r_x,r_y)" + }, + { + "bbox": [ + 55, + 337, + 295, + 396 + ], + "type": "text", + "content": " represents the relative position within the convolution operation centered on the current pixel of " + }, + { + "bbox": [ + 55, + 337, + 295, + 396 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{in}}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 397, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 295, + 517 + ], + "type": "text", + "content": "To overcome this limitation of CDC, the proposed TopoMDCs perform a two-pixel differential operation by extending their reach to the eight neighboring directions beyond the center pixel, as shown in Fig. 2 (a). This design simulates various basic topological patterns, effectively accommodating the elongated and diverse trajectories of tubular structures and enhancing the model's capability to recognize various tubular objects. Let " + }, + { + "bbox": [ + 55, + 397, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{\\mathrm{in}} \\in \\mathbb{R}^{H_{\\mathrm{in}} \\times W_{\\mathrm{in}} \\times C_{\\mathrm{in}}}" + }, + { + "bbox": [ + 55, + 397, + 295, + 517 + ], + "type": "text", + "content": " denote the input feature map, the pre-trained vanilla convolution be " + }, + { + "bbox": [ + 55, + 397, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_0" + }, + { + "bbox": [ + 55, + 397, + 295, + 517 + ], + "type": "text", + "content": ", and the set of TopoMDCs be " + }, + { + "bbox": [ + 55, + 397, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{C}_1, \\mathcal{C}_2, \\dots, \\mathcal{C}_8\\}" + }, + { + "bbox": [ + 55, + 397, + 295, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "type": "text", + "content": "Taking " + }, + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1" + }, + { + "bbox": [ + 55, + 517, + 295, + 540 + ], + "type": "text", + "content": ", which extends toward the top-left pixel (see Fig. 2 (a)), as an example, it is defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 64, + 542, + 294, + 591 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 542, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 64, + 542, + 294, + 591 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {C} _ {1} \\left(r _ {x}, r _ {y}\\right) = \\mathcal {C} _ {c} \\left(r _ {x}, r _ {y}\\right) - \\sum w \\left(\\Delta r _ {x}, \\Delta r _ {y}\\right) \\cdot \\boldsymbol {x} _ {\\text {i n}} \\left(r _ {x}, r _ {y}\\right) \\\\ (\\Delta r _ {x}, \\Delta r _ {y}) \\in \\mathcal {R} _ {1} \\\\ + \\sum_ {(\\Delta r _ {x}, \\Delta r _ {y}) \\in \\mathcal {R} _ {1}} w (\\Delta r _ {x}, \\Delta r _ {y}) \\cdot \\boldsymbol {x} _ {\\text {i n}} (r _ {x} + 1, r _ {y} + 1), \\tag {4} \\\\ \\end{array}", + "image_path": "2257ff9921ed0cd3d45a23e0fa7220f798ee90f086182dc0b09c6fdd462b5b98.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_1 = \\{(-1, - 1),(-1,0),(0, - 1)\\}" + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "content": " is the local receptive field of " + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "content": ". Due to space limitations, the formal definitions of other TopoMDCs are provided in Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "TopoMDCs do not introduce any additional parameters for any convolution layer but instead, adaptively transform the vanilla convolution in " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": ". Specifically, we replace all " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " convolution layers in the encoder of " + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot ;\\theta)" + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": " with TopoMDCs. All TopoMDCs inherit the parameters of the vanilla convolutions. To adapt to a variety of complex (even unseen) topologies, a dynamic combination of TopoMDCs" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "text", + "content": "is required. Given that the topologies of local regions tend to be quite similar, while the topologies of distant regions can differ significantly, we introduce different learnable parameter sets " + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "text", + "content": " with eight learnable router parameters for different image patches. To adapt to the respective topological structures in different regions, we patchify the input image into " + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "inline_equation", + "content": "n\\times n" + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "text", + "content": " non-overlap patches, each with the size of " + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{in}} / n\\times W_{\\mathrm{in}} / n" + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "text", + "content": ". Taking the " + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "text", + "content": "-th image patches " + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "inline_equation", + "content": "\\pmb{p}_j" + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{in}}" + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "text", + "content": ", the corresponding router parameter set " + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "text", + "content": " assigns external parameters to different TopoMDCs, enabling adaptive enhancement of the perception of topological structures within the patch. The output feature " + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}_{\\mathrm{out}}" + }, + { + "bbox": [ + 313, + 337, + 555, + 482 + ], + "type": "text", + "content": " is obtained as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 490, + 555, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 490, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 364, + 490, + 555, + 520 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {x}} _ {\\text {o u t}} = \\mathcal {C} _ {0} \\left(\\boldsymbol {x} _ {\\text {i n}}\\right) - \\sum_ {j = 1} ^ {n \\times n} \\delta_ {j} \\sum_ {i = 1} ^ {8} \\mathcal {C} _ {i} \\left(\\boldsymbol {x} _ {\\text {i n}} ^ {j}\\right), \\tag {5}", + "image_path": "7bd7802d84e7fbccdb3af755b6fb9b1237e1c5d7dffe8889407613a383539da5.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{in}}^{j}" + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "text", + "content": "-th patch of " + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{in}}" + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "text", + "content": ". The patchification operations " + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_0" + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "text", + "content": " differ slightly in their inputs in Eq. (4) and Eq. (5). In Eq. (5), " + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_0" + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_i" + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "text", + "content": " denote the same computation applied uniformly to every pixel in " + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{in}}^{j}" + }, + { + "bbox": [ + 313, + 530, + 554, + 581 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Router Parameter " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "(\\delta)" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " Updating in Stage 1. As mentioned above, we do not update the original parameters of the convolution kernel or any other parameters in the network but only update a small number of introduced router parameters " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". This is not just a consideration for inference speed. Simultaneously adjusting both " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and the model parameters increases the search-space complexity, which may lead to conflicts or interference among parameters. This, in turn, could even disrupt the original feature representation. To this end, we only adjust " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " to adapt the topological response within an external feature space, thereby reducing the risk" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24126" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "of instability. For each incoming test sample " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ", all router parameters are reset to zero, so that the model can independently adapt to the topological structures of the new sample. We use the self-supervised entropy minimization (EM) loss [58] to update " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 140, + 295, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 140, + 295, + 170 + ], + "spans": [ + { + "bbox": [ + 111, + 140, + 295, + 170 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E M}} = - \\sum_ {j = 1} ^ {2} \\hat {\\mathbf {y}} _ {j} \\log (\\hat {\\mathbf {y}} _ {j}), \\tag {6}", + "image_path": "8615c20d2aa8e601b0c22b9988025ea20ef25ea56ec3a3ff3fcaff0fa5577f05.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 112, + 172, + 238, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 172, + 238, + 190 + ], + "spans": [ + { + "bbox": [ + 112, + 172, + 238, + 190 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\delta} = \\underset {\\theta^ {t}} {\\arg \\min } \\mathcal {L} _ {\\mathrm {E M}} (\\mathcal {F} (\\cdot ; \\theta^ {t}; \\boldsymbol {\\delta}), \\boldsymbol {x}).", + "image_path": "a97671c4ad2abe3c1796dfd37de4e1db4a76c1f09e6934870121cecf194e5ef6.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 197, + 284, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 197, + 284, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 284, + 210 + ], + "type": "text", + "content": "3.4. Stage 2: Topological Continuity Refinement" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 214, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 214, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 214, + 296, + 323 + ], + "type": "text", + "content": "As previously discussed in Sec. 3.3, Stage 1 enhances the representation of topological structures in " + }, + { + "bbox": [ + 55, + 214, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t" + }, + { + "bbox": [ + 55, + 214, + 296, + 323 + ], + "type": "text", + "content": ". In Stage 2, we guide the model to learn local characteristics in " + }, + { + "bbox": [ + 55, + 214, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^t" + }, + { + "bbox": [ + 55, + 214, + 296, + 323 + ], + "type": "text", + "content": " for foreground-background differentiation, thereby improving the topological continuity of segmentation. To achieve this, we generate a series of pseudo-breaks at key topological structures on the test image " + }, + { + "bbox": [ + 55, + 214, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 214, + 296, + 323 + ], + "type": "text", + "content": ", producing the edited image " + }, + { + "bbox": [ + 55, + 214, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\pmb{x}'" + }, + { + "bbox": [ + 55, + 214, + 296, + 323 + ], + "type": "text", + "content": " through the proposed Topology Hard sample Generation (TopoHG) strategy, as shown in Fig. 2 (c)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 325, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 296, + 361 + ], + "type": "text", + "content": "Topology Hard Sample Generation (TopoHG). The TopoHG process for generating local hard samples can be roughly divided into the following three steps:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "text", + "content": "Step 1: Key Point Selection. Since we apply consistency supervision on the predictions of local hard samples, the selected regions must have highly reliable pseudo-labels. To this end, we first select a set of points with confidence above the threshold " + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\tau = 0.95" + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "text", + "content": " from the teacher's predictions " + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}'" + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{(u,v) \\mid \\hat{\\pmb{y}}'(u,v) > \\tau\\}" + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "text", + "content": ". Then, we randomly select " + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "text", + "content": " points from " + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "text", + "content": " as key points. Since prediction confidence varies across different test samples, using a fixed " + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "text", + "content": " for all input samples is suboptimal. Therefore, we allow samples with higher confidence to select more key points and create more challenging regions to contribute more to the weight update: " + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "inline_equation", + "content": "N_p = k \\cdot |\\mathcal{P}|" + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 362, + 296, + 517 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "content": "Step 2: Sliding to Search. For each selected key point " + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "inline_equation", + "content": "(u_c, v_c)" + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "content": ", let " + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "inline_equation", + "content": "x_p^{\\mathrm{fg}}" + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "content": " denote the fixed foreground window centered at this point, with a size of " + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "inline_equation", + "content": "s \\times s" + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "content": ". Then, a background sliding window " + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "inline_equation", + "content": "x_p^{\\mathrm{bg}}" + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "content": ", of the same size as the foreground window, slides over adjacent non-overlapping regions, identifying window " + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "inline_equation", + "content": "x_p^{\\mathrm{bg},*}" + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "content": " with the lowest pseudo-label confidence of all pixels. Meanwhile, to prevent " + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "inline_equation", + "content": "x_p^{\\mathrm{bg},*}" + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "content": " from carrying too much foreground information, we set a lower threshold " + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\tau^{\\mathrm{bg}}" + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "content": ". If the sum of foreground confidence of all pixels in " + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "inline_equation", + "content": "x_p^{\\mathrm{bg},*}" + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "content": " exceeds " + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\tau^{\\mathrm{bg}}" + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "content": ", the key point " + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "inline_equation", + "content": "(u_c, v_c)" + }, + { + "bbox": [ + 55, + 519, + 296, + 640 + ], + "type": "text", + "content": " is discarded." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Step 3: Frequency-based Pseudo-break Generation. TopoHG aims to create local pseudo-break topological structures within the critical regions. This requires TopoHG to simulate the characteristics of structures that seem to be broken while retaining essential foreground features. Inspired by [34], we adopt a low-frequency swap method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": "between background and foreground to preserve high-frequency information that carries critical foreground features. Let FFT and iFFT denote the Fast Fourier Transform (FFT) and inverse Fast Fourier Transform (iFFT) operations [7], respectively, with " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "m_{\\mathrm{low}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " as the mask for the low-frequency regions that need to be swapped. Applying FFT to " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "x_p^{\\mathrm{fg}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "x_p^{\\mathrm{bg},*}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " yields the frequency domain representations " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "f_p^{\\mathrm{fg}} = \\mathrm{FFT}(x_p^{\\mathrm{fg}})" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "f_p^{\\mathrm{bg}} = \\mathrm{FFT}(x_p^{\\mathrm{bg},*})" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": ". After performing low-frequency swap on " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "f_p^{\\mathrm{fg}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "f_p^{\\mathrm{bg}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " and applying iFFT, we obtain the modified image " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "x_p^{\\mathrm{swap}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": ", as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 201, + 555, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 201, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 344, + 201, + 555, + 220 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {p} ^ {\\text {s w a p}} = \\mathrm {i F F T} \\left(\\boldsymbol {f} _ {p} ^ {\\mathrm {f g}} \\cdot \\left(1 - \\boldsymbol {m} _ {\\text {l o w}}\\right) + \\boldsymbol {f} _ {p} ^ {\\mathrm {b g}} \\cdot \\boldsymbol {m} _ {\\text {l o w}}\\right). \\tag {7}", + "image_path": "6bd9d432e63a498072d62c2de68ff300800b2c3453a01c89d4ab6b1c78c3c4f5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 223, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 281 + ], + "type": "text", + "content": "To avoid additional risks associated with unnecessary background changes when emphasizing consistency, we update only update pixels in " + }, + { + "bbox": [ + 313, + 223, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\pmb{x}_p^{\\mathrm{fg}}" + }, + { + "bbox": [ + 313, + 223, + 555, + 281 + ], + "type": "text", + "content": " that correspond to the foreground in the pseudo-label " + }, + { + "bbox": [ + 313, + 223, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_p'" + }, + { + "bbox": [ + 313, + 223, + 555, + 281 + ], + "type": "text", + "content": ". The pseudo-break patch " + }, + { + "bbox": [ + 313, + 223, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\pmb{x}_p'" + }, + { + "bbox": [ + 313, + 223, + 555, + 281 + ], + "type": "text", + "content": " can be defined as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 370, + 288, + 555, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 288, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 370, + 288, + 555, + 302 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {p} ^ {\\prime} = \\boldsymbol {x} _ {p} ^ {\\text {s w a p}} \\cdot \\hat {\\boldsymbol {y}} _ {p} ^ {\\prime} + \\boldsymbol {x} _ {p} ^ {\\mathrm {f g}} \\cdot (1 - \\hat {\\boldsymbol {y}} _ {p} ^ {\\prime}). \\tag {8}", + "image_path": "cd59a2a1bc95f4403e640d57b8d58294aaaa94dec049291e5651a6888891355d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 307, + 554, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 342 + ], + "type": "text", + "content": "After completing the above four steps, we obtain a hard sample " + }, + { + "bbox": [ + 313, + 307, + 554, + 342 + ], + "type": "inline_equation", + "content": "\\pmb{x}^{\\prime}" + }, + { + "bbox": [ + 313, + 307, + 554, + 342 + ], + "type": "text", + "content": " that contains multiple local pseudo-breaks, resulting in weakened predictions to assist in model updates." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "text", + "content": "Model Updating in Stage 2. To align the predictions " + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}''" + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "text", + "content": " of these local hard samples " + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\pmb{x}'" + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "text", + "content": " with the original pseudo-label " + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}'" + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "text", + "content": ", we apply consistency regularization as supervision, implemented through the TS scheme as shown in Fig. 2 (c). The cross-entropy loss " + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "text", + "content": " is used to enhance the consistency between the student's (i.e., prediction of the hard sample " + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}''" + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "text", + "content": ") and teacher's (i.e., pseudo-label " + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}'" + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "text", + "content": " of original samples) predictions, updating " + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\theta_i^t" + }, + { + "bbox": [ + 313, + 345, + 555, + 441 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 446, + 554, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 446, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 326, + 446, + 554, + 475 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} = - \\sum_ {(u, v)} \\mathcal {W} (u, v) \\cdot \\left(\\hat {\\boldsymbol {y}} ^ {\\prime} \\log \\left(\\hat {\\boldsymbol {y}} ^ {\\prime \\prime}\\right) + \\hat {\\boldsymbol {y}} ^ {\\prime \\prime} \\log \\left(\\hat {\\boldsymbol {y}} ^ {\\prime}\\right)\\right), \\tag {9}", + "image_path": "0a5e190993049521e6812d828fe68a99642169991d02f9c56f1fa02b128a403d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 333, + 473, + 468, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 473, + 468, + 495 + ], + "spans": [ + { + "bbox": [ + 333, + 473, + 468, + 495 + ], + "type": "interline_equation", + "content": "\\theta_{i}^{t} = \\operatorname *{arg min}_{\\substack{\\theta_{i - 1}^{t}\\\\ }}\\mathcal{L}_{\\mathrm{CE}}(\\mathcal{F}(:,\\theta_{i - 1}^{t};\\delta),\\boldsymbol {x}),", + "image_path": "071ff8147ec2140fa1b6fd66e7277a5f92eee9d52f9e0d64785a034820d90f30.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(u,v)" + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": " is a weight map. The local pseudo-breaks should receive higher attention in consistency regularization; therefore, " + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "inline_equation", + "content": "\\mathcal{W}(u,v)" + }, + { + "bbox": [ + 313, + 501, + 554, + 537 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 543, + 554, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 543, + 554, + 572 + ], + "spans": [ + { + "bbox": [ + 361, + 543, + 554, + 572 + ], + "type": "interline_equation", + "content": "\\mathcal {W} (u, v) = \\left\\{ \\begin{array}{l l} 1 0 & \\text {i f} (u, v) \\in \\boldsymbol {x} _ {p} ^ {\\prime} \\cdot \\hat {\\boldsymbol {y}} _ {p} ^ {\\prime} \\\\ 1 & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {10}", + "image_path": "f2c98df9f42a77b1e24b77f53c779080cdf55f46714f9418883857f33fc6b0c4.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 576, + 554, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 554, + 635 + ], + "type": "text", + "content": "The gradient is used solely to update the student, while the teacher's parameters are updated by an exponential moving average (EMA) of the student's parameters. Finally, after several steps of updates, the unmodified input " + }, + { + "bbox": [ + 313, + 576, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 576, + 554, + 635 + ], + "type": "text", + "content": " is fed into the student model to obtain the final prediction." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 646, + 395, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 395, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 395, + 658 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "We conduct comparison experiments across four scenarios, i.e., retinal vessel segmentation, road extraction, microscopic neuronal segmentation, and retinal OCT-angiography vessel segmentation covering ten datasets." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24127" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 103, + 553, + 255 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Cross-domain testing results in four different scenarios, i.e., retinal vessel segmentation, road extraction, microscopic neuronal segmentation, and retinal OCT-angiography vessel segmentation. Source Only: Trained on the source domain and tested on the target domain directly. The best and second-best results in each column are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 103, + 553, + 255 + ], + "lines": [ + { + "bbox": [ + 56, + 103, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 553, + 255 + ], + "type": "table", + "html": "
MethodDRIVE → CHASEDRIVE → STARECHASE → DRIVECHASE → STAREDeepGlobe → MR
Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓
Source Only22.3818.2141.2548.05/106.3062.9558.5989.9548.77/105.2044.8954.9776.83
TENT [58]66.9769.4337.7567.1460.7498.4064.7460.2588.3061.9155.39106.0042.6152.1877.71
CoTBA [60]68.6071.5336.3867.0259.8697.4067.6464.8081.2063.7257.6499.8550.8360.5874.82
SAR [43]66.9769.4937.1367.2960.9097.9065.0660.6988.3562.1755.68105.6043.8854.3076.70
DIGA [61]66.9170.3736.5059.2261.7782.0066.5462.4385.8563.8057.32103.7540.6551.0878.82
DomainAdaptor [72]64.9966.9941.5062.9855.51102.7564.4359.8988.5059.4152.05105.4545.1755.8676.98
MedBN [44]49.7751.0975.8857.8351.99102.5559.6354.1489.9554.0548.27109.0038.92/76.89
VPTTA [3]66.4868.9338.5067.1660.5498.9565.0260.7088.3561.7355.01105.5544.5155.9977.43
TopoTTA (Ours)70.7377.0525.3867.3662.7485.7572.9670.2679.1568.4361.2097.6050.7466.0869.23
MethodDeepGlobe → CNDSNeub1 → Neub2Neub2 → Neub1ROSE → OCTA500OCTA500 → ROSE
Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓Dice (%) ↑cDice (%) ↑β ↓
Source Only81.7492.227.0914.34/8.2260.9765.6260.5650.1057.6555.3671.3974.6419.55
TENT [58]78.6990.656.6551.5560.589.0663.2672.329.3768.9975.6943.0672.2575.6417.89
CoTBA [60]77.8391.185.9952.2162.7810.0664.4073.587.9670.0575.6541.3472.5875.7818.00
SAR [43]79.1990.726.8651.8560.809.1663.2172.319.6267.6574.1444.4272.4575.8517.44
DIGA [61]75.4583.2214.1059.1466.4419.7761.3273.408.1966.1772.9842.1871.5175.4017.99
DomainAdaptor [72]80.7191.726.8053.2460.967.9464.0472.388.7567.3974.6044.1471.7775.3717.00
MedBN [44]76.9687.949.2255.5963.9210.5661.3470.8211.5740.6447.53298.7471.7773.2916.11
VPTTA [3]77.0188.568.6454.2562.888.8363.4672.249.3467.5474.0144.9272.6175.9317.11
TopoTTA (Ours)89.1596.355.9866.8874.957.2261.8375.386.5070.7078.2431.8875.5977.7316.22
", + "image_path": "8432e023444e835ae796e4d8ac16947948b96e4494a222fd9375dc56bfe7fb40.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 262, + 163, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 163, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 163, + 276 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 286, + 295, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 286, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 54, + 286, + 295, + 477 + ], + "type": "text", + "content": "Datasets. We adopt 10 widely-used TSS datasets for the following experiments. The retinal vessel segmentation datasets are collected from three different sources: DRIVE [54], STARE [16], and CHASEDB1 (CHASE) [12], which contain 40, 20, and 28 samples, respectively. Three road extraction datasets are DeepGlobe [8], Massachusetts road (MR) [39], and CNDS [4], providing 8,570, 1,171, and 224 samples, respectively. The microscopic neuronal segmentation datasets are Neub1 and Neub2 [76], with 112 and 98 samples, respectively. Two retinal OCT-angiography vessel segmentation datasets are from different sources, ROSE1 [37] and OCTA500 [26], containing 117 and 300 samples, respectively. All datasets follow their original training and testing splits, except for STARE, which is officially used as a test set due to its limited sample size [16]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 481, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 578 + ], + "type": "text", + "content": "Evaluation Metrics. Following prior works [46, 59, 65, 68], we use Dice as the primary metric for TSS performance evaluation, with clDice [52] and Betti errors " + }, + { + "bbox": [ + 55, + 481, + 295, + 578 + ], + "type": "inline_equation", + "content": "(\\beta)" + }, + { + "bbox": [ + 55, + 481, + 295, + 578 + ], + "type": "text", + "content": " [18] as metrics of topological continuity. In Table 1, the diagonal marks “/” in clDice for some TTA methods indicate cases where no skeleton map can be extracted from their predictions, and thus clDice cannot be calculated. Further investigation reveals that these are failure cases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Implementation Details. For each scenario, we use a model pre-trained on a single dataset (source domain) and test it on other datasets (target domains) to calculate metrics. UNet [48] and CS2Net [24] (designed for TSS) are adopted as baseline networks for validating the performance of TopoTTA. To ensure consistent field-of-view and cross-domain accuracy, all datasets are resized to " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " pixels, except for the OCTA dataset as ROSE and OCTA500 already share similar resolutions. During testing, TopoTTA and the comparison methods undergo six iterations of adaptation for each image, with TopoTTA performing three iter" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 307, + 553, + 380 + ], + "blocks": [ + { + "bbox": [ + 313, + 262, + 555, + 305 + ], + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 305 + ], + "type": "text", + "content": "Table 2. Average cross-dataset testing results across ten datasets using UNet and CS2Net as baseline networks. The best and second-best results in each column are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 307, + 553, + 380 + ], + "lines": [ + { + "bbox": [ + 316, + 307, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 307, + 553, + 380 + ], + "type": "table", + "html": "
MethodUNetCS2Net
Dice (%)↑cI Dice (%)↑β↓Dice (%)↑cI Dice (%)↑β↓
Source Only50.5642.1951.9450.3151.1754.35
TENT [58]63.8167.2949.4263.7668.0951.87
CoTTA [60]65.4969.3447.3065.1269.8448.19
SAR [43]63.9767.4949.3263.6368.1151.84
DIGA [61]63.0767.4448.9263.9268.2051.97
DomainAdaptor [72]63.4166.5349.9860.9565.6554.68
MedBN [44]56.6554.9080.0562.6365.9754.63
VPTTA [3]63.9867.4849.7663.3067.7752.25
TopoTTA (Ours)69.4474.0043.0167.6872.7043.86
", + "image_path": "35cebd7d0e70cdb1cc00546aea88eabd307804d74ac2464b510da0d9ca848ea2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "type": "text", + "content": "ations at each stage to ensure experimental consistency. For TopoTTA, we use the Adam optimizer with learning rates of 0.01 and " + }, + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "type": "text", + "content": " for the first and second stages, respectively. We use 0.5 as the threshold for mask binarization. The hyperparameters " + }, + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "type": "inline_equation", + "content": "n \\times n" + }, + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "type": "text", + "content": " (number of TopoMDC regions), " + }, + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "type": "text", + "content": " (modification window size), and " + }, + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "type": "inline_equation", + "content": "\\tau^{\\mathrm{bg}}" + }, + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "type": "text", + "content": " (upper limit of the foreground pixel), respectively, are set to " + }, + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 313, + 387, + 555, + 482 + ], + "type": "text", + "content": ", 30, 0.05 for all tasks. We present more details in Appendix." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 484, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 628 + ], + "type": "text", + "content": "Comparison TTA Methods. In the following experiments, we compare our method with three categories of TTA methods: (1) methods designed for semantic segmentation, such as CoTTA [60] and DIGA [61], (2) a method designed for medical image segmentation, i.e., VPTTA [3], and (3) methods designed for classification, including TENT [58], SAR [43], DomainAdaptor [72] and MedBN [44], which can be easily adapted to segmentation tasks [3, 61] by using pixel-level self-supervision or BN statistics adjustment. We implement these TTA methods on the same baseline networks as TopoTTA, using hyperparameters as specified in the corresponding reference or assigning optimal ones." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 636, + 432, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 432, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 432, + 649 + ], + "type": "text", + "content": "4.2. Comparison Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 713 + ], + "type": "text", + "content": "Table 1 presents the cross-dataset testing results across four different scenarios. Table 2 shows the average results for each baseline across all scenarios. Fig. 3 visualizes the segmentation results. From these results, we have the following key observations:" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24128" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 326 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 326 + ], + "type": "image", + "image_path": "1701b7d9a080c922f9d9c65f9c08f2b0d32279c255c29315c88a72da40b289d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 326, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 326, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 555, + 360 + ], + "type": "text", + "content": "Figure 3. Visualization of segmentation results for TopoTTA and seven comparison methods across four scenarios. \"Source Image\" is provided solely to illustrate domain differences across datasets. \"Source Only\" denotes results without any TTA methods applied, and GT is short for ground-truth labels. More visual examples are presented in Appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "type": "text", + "content": "Key Observation 1: Improvement on topological continuity across four scenarios. As shown in Table 1, TopoTTA consistently achieves the highest topological continuity metrics. This improvement is not attained at the expense of segmentation accuracy (e.g., DIGA in the DRIVE " + }, + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "type": "text", + "content": " STARE, and VPTTA in the DeepGlobe " + }, + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "type": "text", + "content": " MR), but rather, TopoTTA enhances Dice score on most datasets. Notably, for the DeepGlobe " + }, + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "type": "text", + "content": " CNDS experiment, where the source domain model has already demonstrated strong test time performance, the results of all comparison methods decline. However, TopoTTA substantially improves segmentation accuracy and topological continuity metrics in this experiment. In particular, in the challenging Neub1 " + }, + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "type": "text", + "content": " Neub2 experiment, TopoTTA surpasses the second-best method, achieving gains of 7.74 and 8.51 in Dice and cIDice scores, respectively, underscoring the method's exceptional capacity for topological adaptation and continuity enhancement. Additionally, in the Neub1 " + }, + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 54, + 365, + 295, + 628 + ], + "type": "text", + "content": " Neub2 experiment, although TopoTTA's Dice score is lower than most comparison methods, visual analysis indicates this is likely due to label granularity rather than a true performance drop. As space is limited, details are provided in Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "Key Observation 2: Consistent superiority across different baselines. TopoTTA can function as a plug-and-play module within standard convolutional neural networks. From Table 2, we observe that TopoTTA achieves significant improvements across ten experiments using both UNet and CS2Net as baseline networks. CoTTA attains the second-best performance, indicating that the TS scheme" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "type": "text", + "content": "may be well-suited for segmentation tasks. With UNet as the baseline network, TopoTTA outperforms CoTTA by " + }, + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "type": "inline_equation", + "content": "3.95\\%" + }, + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "type": "text", + "content": " in Dice, " + }, + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "type": "inline_equation", + "content": "4.66\\%" + }, + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "type": "text", + "content": " in clDice, and reduces Betti errors by " + }, + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "type": "inline_equation", + "content": "4.29\\%" + }, + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "type": "text", + "content": ". Similar gains are observed using CS2Net. We also implement a TopoTTA variant on DSCNet, which delivers notable performance improvements, demonstrating TopoTTA's adaptability across different baseline networks. Detailed results on other baseline networks (i.e., CS2Net and DSCNet) can be found in Appendix." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 474, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 666 + ], + "type": "text", + "content": "Key Observation 3: Clear visual proof of enhanced continuity and topological adaptability. In Fig. 3, we magnify specific regions of selected images for detailed comparison. In rows 1, 2, and 5, all comparison methods show apparent discontinuities in the red-boxed areas, while TopoTTA effectively connects these branches, underscoring its superior continuity enhancement. For the road image in row 3, TopoTTA achieves an almost perfect match with the ground truth—a level of accuracy beyond all comparison methods' reach. In the microscopic neuronal image in row 4, most comparison methods struggle to segment even the main branches, while TopoTTA closely matches the ground truth. These results showcase TopoTTA's adaptability to various topological features. Moreover, in row 6, comparison methods exhibit noise due to over-segmentation, which TopoTTA effectively suppresses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 672, + 409, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 409, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 409, + 685 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "In this section, we perform ablation analyses of TopoMDCs and TopoHG. For all ablation experiments, we report the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24129" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 92, + 294, + 150 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 3. Ablation results of the proposed TopoTTA. Baseline* is the model using the teacher-student scheme." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 92, + 294, + 150 + ], + "lines": [ + { + "bbox": [ + 56, + 92, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 92, + 294, + 150 + ], + "type": "table", + "html": "
MethodDice (%) ↑cI Dice (%) ↑β ↓
Baseline65.3761.6982.24
Baseline + TopoMDCs (Stage 1)68.7065.1476.28
Baseline*65.9562.2077.90
Baseline* + TopoHG (Stage 2)68.8266.6173.63
TopoTTA69.8767.8173.27
", + "image_path": "57f8ec91fabe5297e73a296e38eefd2b9fad3dee8bd75dcfd6fe95427589f83c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 175, + 294, + 221 + ], + "blocks": [ + { + "bbox": [ + 55, + 152, + 294, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 294, + 175 + ], + "type": "text", + "content": "Table 4. Ablation results of the proposed TopoMDCs using each TopoMDC type individually." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 175, + 294, + 221 + ], + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 221 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 221 + ], + "type": "table", + "html": "
VariantMDC Operation SetDice (%) ↑clDice (%) ↑β ↓
CDC{Cc}68.9967.3572.91
Orthogonal MDCs{C1,C2,C3,C4}69.2667.5472.93
Diagonal MDCs{C5,C6,C7,C8}69.2967.4873.86
TopoMDCs (Full){C1,C2,···,C8}69.8767.8173.27
", + "image_path": "2df67c0e8ce61a67c0f09da16b45fdc9f30bcfc18b4b4ac1b0ae6a2376c3627f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 257, + 294, + 307 + ], + "blocks": [ + { + "bbox": [ + 55, + 224, + 295, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 257 + ], + "type": "text", + "content": "Table 5. Ablation results of the proposed TopoHG. We compare with three variants, i.e., Gaussian blur, random Gaussian noise, and image swap in the spatial domain." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 257, + 294, + 307 + ], + "lines": [ + { + "bbox": [ + 58, + 257, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 58, + 257, + 294, + 307 + ], + "type": "table", + "html": "
MethodDice (%) ↑cIDice (%) ↑β ↓
Stage 168.7065.1476.28
Stage 1 + Blur69.1666.2676.19
Stage 1 + Noise68.4865.0077.76
Stage 1 + Image swap67.6566.3669.63
TopoTTA (Ours)69.8767.8173.27
", + "image_path": "99798830808ab7710ed1bbe4cf3008a3644ea98c0afa9a1e462a11b068eac611.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 313, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 295, + 350 + ], + "type": "text", + "content": "average results on the retinal vessel segmentation scenario. Due to the limited space, the ablation analysis on our hyperparameters is provided in Appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 351, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 295, + 591 + ], + "type": "text", + "content": "Effectiveness of TopoMDCs. To validate the effectiveness of the proposed TopoMDCs, we compare the network performance before and after applying TopoMDCs to the baseline or removing them from the TopoTTA. As shown in Table 3, TopoMDCs result in significant improvements across all metrics compared to the baseline that only updates statistical parameters, validating the effectiveness of TopoMDCs in adapting to various topological structures. Additionally, Table 4 provides a detailed analysis of the contribution of each type of TopoMDCs to performance enhancement. We categorize TopoMDCs into three types: central " + }, + { + "bbox": [ + 55, + 351, + 295, + 591 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}_c)" + }, + { + "bbox": [ + 55, + 351, + 295, + 591 + ], + "type": "text", + "content": ", orthogonal " + }, + { + "bbox": [ + 55, + 351, + 295, + 591 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}_{1-4})" + }, + { + "bbox": [ + 55, + 351, + 295, + 591 + ], + "type": "text", + "content": ", and diagonal " + }, + { + "bbox": [ + 55, + 351, + 295, + 591 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}_{5-8})" + }, + { + "bbox": [ + 55, + 351, + 295, + 591 + ], + "type": "text", + "content": ", and evaluate their impact individually. Results indicate that orthogonal and diagonal types relative to central lead to more significant performance improvements, with optimal topological awareness achieved when orthogonal and diagonal types are used together. We also visualize feature maps before and after incorporating TopoMDCs. Fig. 4 illustrates that in all four scenarios, TopoMDCs effectively enhance the topological structures on the target domain." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": "Effectiveness of TopoHG. Here, we aim to answer two research questions: (1) Can TopoHG truly create pseudobreak predictions? (2) Is the low-frequency swap method in TopoHG more effective for enhancing topological continuity compared to other data augmentation methods?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "As shown in Table 3, using only Stage 2 with TopoHG still leads to improvements in both segmentation performance and topological continuity. Compared to Stage 1 alone, adding TopoHG notably enhances continuity. To answer the first question, we provide visualizations of selected" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 560, + 163 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 560, + 163 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 560, + 163 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 560, + 163 + ], + "type": "image", + "image_path": "f1f61249b925a28412e30172a9a87c80725479ea92ff297ae007b5f904cf3615.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 163, + 555, + 186 + ], + "lines": [ + { + "bbox": [ + 314, + 163, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 314, + 163, + 555, + 186 + ], + "type": "text", + "content": "Figure 4. Feature maps before and after TopoMDCs across four scenarios. VC represents the feature maps of vanilla convolutions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 188, + 555, + 287 + ], + "blocks": [ + { + "bbox": [ + 319, + 188, + 555, + 287 + ], + "lines": [ + { + "bbox": [ + 319, + 188, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 319, + 188, + 555, + 287 + ], + "type": "image", + "image_path": "56c7784fb1c2d14b4fd348e00e3f61e4c846b472ea66deb5c03f83c081fb03fc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 292, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 313, + 292, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 555, + 326 + ], + "type": "text", + "content": "Figure 5. Visualizations of original patches " + }, + { + "bbox": [ + 313, + 292, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{p}^{\\mathrm{fg}}" + }, + { + "bbox": [ + 313, + 292, + 555, + 326 + ], + "type": "text", + "content": ", pseudo-labels " + }, + { + "bbox": [ + 313, + 292, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{y}}_p'" + }, + { + "bbox": [ + 313, + 292, + 555, + 326 + ], + "type": "text", + "content": ", pseudo-break patches " + }, + { + "bbox": [ + 313, + 292, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\pmb{x}_p'" + }, + { + "bbox": [ + 313, + 292, + 555, + 326 + ], + "type": "text", + "content": ", and weakened prediction masks " + }, + { + "bbox": [ + 313, + 292, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\pmb{y}_p'" + }, + { + "bbox": [ + 313, + 292, + 555, + 326 + ], + "type": "text", + "content": " in selected regions across four scenarios." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 333, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 548 + ], + "type": "text", + "content": "regions' predictions before and after introducing pseudo-breaks in Fig. 5. The images generated by TopoHG significantly disrupt the model's previously confident predictions, effectively creating local pseudo-breaks. For the second question, Table 5 compares various data augmentation methods, including Gaussian blur, random Gaussian noise, and foreground swap with the background in the spatial domain (image swap). The blur method creates local hard samples by removing high-frequency details, which offers some improvement. The image swap method seems to yield superior topological performance, but a sharp drop in Dice indicates that lower Betti errors may stem from overly aggressive predictions connecting many regions incorrectly. This is likely due to consistent regularization on the background without high-frequency details, which can cause incorrect segmentation of the background as foreground. Due to the limited space, visual comparisons of different data augmentation methods are provided in Appendix." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 561, + 388, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 561, + 388, + 573 + ], + "spans": [ + { + "bbox": [ + 314, + 561, + 388, + 573 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "In this work, we proposed the first TTA framework, i.e., TopoTTA, for TSS to enhance cross-domain performance by addressing discrepant topological structures and fragile topological continuity. To improve the adaptability to topological structure, we proposed TopoMDCs to learn how to adaptively integrate various orientation information according to the topological feature of each test sample. Besides, we developed TopoHG to generate pseudo-breaks to enhance the predictions' topological continuity. Extensive experiments across four scenarios verified the superiority of our TopoTTA compared to other methods." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24130" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "Acknowledgement. This research was supported by the National Natural Science Foundation of China (Grant No. 62202027)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 121, + 115, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 115, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 115, + 133 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 140, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 140, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 294, + 205 + ], + "type": "text", + "content": "[1] Cheng Chen, Quande Liu, Yueming Jin, Qi Dou, and Pheng-Ann Heng. Source-Free Domain Adaptive Fundus Image Segmentation with Denoised Pseudo-Labeling. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Interventions, pages 225–235, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 208, + 294, + 262 + ], + "type": "text", + "content": "[2] Zhuangzhuang Chen, Zhuonan Lai, Jie Chen, and Jianqiang Li. Mind Marginal Non-Crack Regions: Clustering-Inspired Representation Learning for Crack Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 12698-12708, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 264, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 264, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 62, + 264, + 294, + 329 + ], + "type": "text", + "content": "[3] Ziyang Chen, Yongsheng Pan, Yiwen Ye, Mengkang Lu, and Yong Xia. Each Test Image Deserves a Specific Prompt: Continual Test-Time Adaptation for 2D Medical Image Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11184-11193, 2024. 2, 3, 6, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 331, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 294, + 386 + ], + "type": "text", + "content": "[4] Guangliang Cheng, Ying Wang, Shibiao Xu, Hongzhen Wang, Shiming Xiang, and Chunhong Pan. Automatic Road Detection and Centerline Extraction via Cascaded End-to-End Convolutional Neural Network. IEEE Transactions on Geoscience and Remote Sensing, 55(6):3322-3337, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 388, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 294, + 441 + ], + "type": "text", + "content": "[5] Lechao Cheng, Chaowei Fang, Dingwen Zhang, Guanbin Li, and Gang Huang. Compound Batch Normalization for Long-Tailed Image Classification. In Proceedings of the ACM International Conference on Multimedia, pages 1925-1934, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 294, + 498 + ], + "type": "text", + "content": "[6] Mingfei Cheng, Kaili Zhao, Xuhong Guo, Yajing Xu, and Jun Guo. Joint Topology-Preserving and Feature-Refinement Network for Curvilinear Structure Segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 7147–7156, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 501, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 294, + 533 + ], + "type": "text", + "content": "[7] James W. Cooley and John W. Tukey. An Algorithm for the Machine Calculation of Complex Fourier Series. Mathematics of Computation, 19(90):297-301, 1965. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 535, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 294, + 600 + ], + "type": "text", + "content": "[8] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. DeepGlobe 2018: A Challenge to Parse the Earth Through Satellite Images. In IEEE Conference on Computer Vision and Pattern Recognition Workshop, pages 172-181, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 602, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 294, + 657 + ], + "type": "text", + "content": "[9] Jiahua Dong, Yang Cong, Gan Sun, Bineng Zhong, and Xiaowei Xu. What Can Be Transferred: Unsupervised Domain Adaptation for Endoscopic Lesions Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4023-4032, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[10] Mario Döbler, Robert A Marsden, and Bin Yang. Robust Mean Teacher for Continual and Gradual Test-Time Adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7704-7714, 2023. 2, 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[11] Zhen Fang, Jie Lu, Feng Liu, and Guangquan Zhang. Semi-Supervised Heterogeneous Domain Adaptation: Theory and Algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1087-1105, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "text", + "content": "[12] Muhammad Moazam Fraz, Paolo Remagnino, Andreas Hoppe, Bunyarit Uyyanonvara, Alicja R Rudnicka, Christopher G Owen, and Sarah A Barman. An Ensemble Classification-based Approach Applied to Retinal Blood Vessel Segmentation. IEEE Transactions on Biomedical Engineering, 59(9):2538-2548, 2012. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 184, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 281 + ], + "type": "text", + "content": "[13] Mohsen Ghafoorian, Alireza Mehrtash, Tina Kapur, Nico Karssemeijer, Elena Marchiori, Mehran Pesteie, Charles R. G. Guttmann, Frank-Erik de Leeuw, Clare M. Tempany, Bram van Ginneken, Andriy Fedorov, Purang Abolmaesumi, Bram Platel, and William M. Wells. Transfer Learning for Domain Adaptation in MRI: Application in Brain Lesion Segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Interventions, pages 516-524, 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 283, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 336 + ], + "type": "text", + "content": "[14] Xin Guo, Wenzhong Tang, Haoran Wang, Jiale Wang, Shuai Wang, Xiaolei Qu, and Xun Lin. MorFormer: morphology-aware transformer for generalized pavement crack segmentation. IEEE Transactions on Intelligent Transportation Systems, 26(6):8219-8232, 2025. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[15] Saumya Gupta, Yikai Zhang, Xiaoling Hu, Prateek Prasanna, and Chao Chen. Topology-Aware Uncertainty for Image Segmentation. In Advances in the Neural Information Processing Systems, page 36, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 425 + ], + "type": "text", + "content": "[16] AD Hoover, Valentina Kouznetsova, and Michael Goldbaum. Locating Blood Vessels in Retinal Images by Piecewise Threshold Probing of a Matched Filter Response. IEEE Transactions on Medical Imaging, 19(3):203-210, 2000. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 426, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 459 + ], + "type": "text", + "content": "[17] Xiaoling Hu. Structure-Aware Image Segmentation with Homotopy Warping. In Advances in the Neural Information Processing Systems, pages 24046-24059, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 460, + 553, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 501 + ], + "type": "text", + "content": "[18] Xiaoling Hu, Fuxin Li, Dimitris Samaras, and Chao Chen. Topology-Preserving Deep Image Segmentation. In Advances in the Neural Information Processing Systems, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 503, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 503, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 503, + 553, + 547 + ], + "type": "text", + "content": "[19] Xiaoling Hu, Yusu Wang, Li Fuxin, Dimitris Samaras, and Chao Chen. Topology-Aware Segmentation Using Discrete Morse Theory. In Proceedings of the International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "text", + "content": "[20] Xiaoling Hu, Dimitris Samaras, and Chao Chen. Learning Probabilistic Topological Representations Using Discrete Morse Theory. In Proceedings of the International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 593, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 646 + ], + "type": "text", + "content": "[21] Jiaxing Huang, Yanfeng Zhou, Yaoru Luo, Guole Liu, Heng Guo, and Ge Yang. Representing Topological Self-Similarity Using Fractal Feature Maps for Accurate Segmentation of Tubular Structures. In Proceedings of the European Conference on Computer Vision, pages 143-160, 2025. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "type": "text", + "content": "[22] Klara Janouskova, Tamir Shor, Chaim Baskin, and Jiri Matas. Single Image Test-Time Adaptation for Segmentation. Transactions on Machine Learning Research, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[23] Yannick Kirchhoff, Maximilian R Rokuss, Saikat Roy, Balint Kovacs, Constantin Ulrich, Tassilo Wald, Maximilian Zenk, Philipp Vollmuth, Jens Kleesiek, Fabian Isensee," + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24131" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 117 + ], + "type": "text", + "content": "and Klaus Maier-Hein. Skeleton Recall Loss for Connectivity Conserving and Resource Efficient Segmentation of Thin Tubular Structures. In Proceedings of the European Conference on Computer Vision, pages 218-234, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 172 + ], + "type": "text", + "content": "[24] Mou Lei, Yitian Zhao, Huazhu Fu, Yonghuai Liu, Jun Cheng, Yalin Zheng, Pan Su, Jianlong Yang, Li Chen, Alejandro F Frangi, Masahiro Akiba, and Jiang Liu. CS2-Net: Deep learning Segmentation of Curvilinear Structures in Medical Imaging. Medical Image Analysis, 67:101874, 2021. 1, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 295, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 295, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 295, + 226 + ], + "type": "text", + "content": "[25] Qin Lei, Jiang Zhong, and Qizhu Dai. Enriching Information and Preserving Semantic Consistency in Expanding Curvilinear Object Segmentation Datasets. In Proceedings of the European Conference on Computer Vision, pages 233-250, 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 282 + ], + "type": "text", + "content": "[26] Mingchao Li, Kun Huang, Qiuzhuo Xu, Jiadong Yang, Yuhan Zhang, Zexuan Ji, Keren Xie, Songtao Yuan, Qinghuai Liu, and Qiang Chen. OCTA-500: A Retinal Dataset for Optical Coherence Tomography Angiography Study. Medical Image Analysis, 93:103092, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 295, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 295, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 295, + 325 + ], + "type": "text", + "content": "[27] Jian Liang, Ran He, and Tieniu Tan. A Comprehensive Survey on Test-Time Adaptation Under Distribution Shifts. International Journal of Computer Vision, pages 1-34, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "text", + "content": "[28] Li Lin, Linkai Peng, Huaqing He, Pujin Cheng, Jiewei Wu, Kenneth K.Y. Wong, and Xiaoying Tang. YoloCurvSeg: You Only Label One Noisy Skeleton for Vessel-Style Curvilinear Structure Segmentation. Medical Image Analysis, 90: 102937, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 295, + 446 + ], + "type": "text", + "content": "[29] Xun Lin, Shuai Wang, Rizhao Cai, Yizhong Liu, Ying Fu, Wenzhong Tang, Zitong Yu, and Alex C. Kot. Suppress and rebalance: Towards generalized multi-modal face anti-spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 211-221, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 514 + ], + "type": "text", + "content": "[30] Xun Lin, Yi Yu, Song Xia, Jue Jiang, Haoran Wang, Zitong Yu, Yizhong Liu, Ying Fu, Shuai Wang, Wenzhong Tang, and Alex C. Kot. Safeguarding medical image segmentation datasets against unauthorized training via contour- and texture-aware perturbations. arXiv preprint arXiv:2403.14250, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 514, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 295, + 579 + ], + "type": "text", + "content": "[31] Xun Lin, Yi Yu, Zitong Yu, Ruohan Meng, Jiale Zhou, Ajian Liu, Yizhong Liu, Shuai Wang, Wenzhong Tang, Zhen Lei, and Alex C. Kot. *Hidemia: Hidden wavelet mining for privacy-enhancing medical image analysis*. In *Proceedings of the ACM International Conference on Multimedia*, pages 8110–8119, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 295, + 635 + ], + "type": "text", + "content": "[32] Xun Lin, Ajian Liu, Zitong Yu, Rizhao Cai, Shuai Wang, Yi Yu, Jun Wan, Zhen Lei, Xiaochun Cao, and Alex Kot. Reliable and balanced transfer learning for generalized multimodal face anti-spoofing. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "type": "text", + "content": "[33] Huajun Liu, Xiangyu Miao, Christoph Mertz, Chengzhong Xu, and Hui Kong. Crackformer: Transformer Network for Fine-Grained Crack Detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 3783-3792, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[34] Quande Liu, Cheng Chen, Jing Qin, Qi Dou, and Pheng-Ann Heng. FedDG: Federated Domain Generalization on" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "Medical Image Segmentation via Episodic Learning in Continuous Frequency Space. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1013-1023, 2021. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[35] Yuang Liu, Wei Zhang, and Jun Wang. Source-Free Domain Adaptation for Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1215-1224, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 215 + ], + "type": "text", + "content": "[36] Junyan Lyu, Yiqi Zhang, Yijin Huang, Li Lin, Pujin Cheng, and Xiaoying Tang. AADG: Automatic Augmentation for Domain Generalization on Retinal Image Segmentation. IEEE Transactions on Medical Imaging, 41(12):3699-3711, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[37] Yuhui Ma, Huaying Hao, Jianyang Xie, Huazhu Fu, Jiong Zhang, Jianlong Yang, Zhen Wang, Jiang Liu, Yalin Zheng, and Yitian Zhao. ROSE: A Retinal OCT-Angiography Vessel Segmentation Dataset and New Model. IEEE Transactions on Medical Imaging, 40(3):928-939, 2020. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 272, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 327 + ], + "type": "text", + "content": "[38] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The Norm Must Go On: Dynamic Unsupervised Domain Adaptation by Normalization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 14765-14775, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 349 + ], + "type": "text", + "content": "[39] Volodymyr Mnih. Machine Learning for Aerial Image Labeling. University of Toronto, Canada, 2013. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[40] A Tuan Nguyen, Thanh Nguyen-Tang, Ser-Nam Lim, and Philip H.S. Torr. Tipi: Test Time Adaptation with Transformation Invariance. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 24162-24171, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 404, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 469 + ], + "type": "text", + "content": "[41] Jiayi Ni, Senqiao Yang, Ran Xu, Jiaming Liu, Xiaoqi Li, Wenyu Jiao, Zehui Chen, Yi Liu, and Shanghang Zhang. Distribution-Aware Continual Test-Time Adaptation for Semantic Segmentation. In Proceedings of International Conference on Robotics and Automation, pages 3044-3050, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 471, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 535 + ], + "type": "text", + "content": "[42] Munan Ning, Donghuan Lu, Yujia Xie, Dongdong Chen, Dong Wei, Yefeng Zheng, Yonghong Tian, Shuicheng Yan, and Li Yuan. MADAv2: Advanced Multi-Anchors based Active Domain Adaptation Segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11): 13553-13566, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 536, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 591 + ], + "type": "text", + "content": "[43] Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards Stable Test-Time Adaptation in Dynamic Wild World. In Proceedings of the International Conference on Learning Representations, 2023. 2, 3, 6, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 592, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 646 + ], + "type": "text", + "content": "[44] Hyejin Park, Jeongyeon Hwang, Sunung Mun, Sangdon Park, and Jungseul Ok. MedBN: Robust Test-Time Adaptation Against Malicious Test Samples. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5997-6007, 2024. 3, 6, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[45] Bibiloni Pedro, González-Hidalgo M, and Massanet Sebastià. A Survey on Curvilinear Object Segmentation in Multiple Applications. Pattern Recognition, 60:949-970, 2016. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[46] Yaolei Qi, Yuting He, Xiaoming Qi, Yuan Zhang, and Guanyu Yang. Dynamic Snake Convolution Based on Topo" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24132" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "logical Geometric Constraints for Tubular Structure Segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 6070-6079, 2023. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 160 + ], + "type": "text", + "content": "[47] Xiaolei Qu, Jiale Zhou, Jue Jiang, Wenhan Wang, Haoran Wang, Shuai Wang, Wenzhong Tang, and Xun Lin. Ehformer: Regional easy-hard-aware transformer for breast lesion segmentation in ultrasound images. Information Fusion, 109:102430, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 295, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 295, + 216 + ], + "type": "text", + "content": "[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Proceedings of the International Conference on Medical Image Computing and Computer Assisted Interventions, pages 234–241, 2015. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "type": "text", + "content": "[49] Jameel Hassan Abdul Samadh, Hanan Gani, Noor Hazim Hussein, Muhammad Uzair Khattak, Muzammal Naseer, Fahad Khan, and Salman Khan. Align Your Prompts: Test-Time Prompting with Distribution Alignment for Zero-Shot Generalization. In Advances in the Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 326 + ], + "type": "text", + "content": "[50] Tianyi Shi, Nicolas Boutry, Yongchao Xu, and Thierry Géraud. Local Intensity Order Transformation for Robust Curvilinear Object Segmentation. IEEE Transactions on Image Processing, 31:2557-2569, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "text", + "content": "[51] Tianyi Shi, Xiaohuan Ding, Liang Zhang, and Xin Yang. FreeCOS: Self-Supervised Learning from Fractals and Unlabeled Images for Curvilinear Object Segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 876-886, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 458 + ], + "type": "text", + "content": "[52] Suprosanna Shit, Johannes C Paetzold, Anjany Sekuboyina, Ivan Ezhov, Alexander Unger, Andrey Zhylka, Josien P.W. Plum, Ulrich Bauer, and Bjoern H Menze. cIDice: A Novel Topology-Preserving Loss Function for Tubular Structure Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16560-16569, 2021. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[53] Junha Song, Jungsoo Lee, In So Kweon, and Sungha Choi. ECoTTA: Memory-Efficient Continual Test-Time Adaptation via Self-Distilled Regularization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11920-11929, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "type": "text", + "content": "[54] Joes Staal, Michael D Abramoff, Meindert Niemeijer, Max A Viergever, and Bram Van Ginneken. Ridge-based Vessel Segmentation in Color Images of the Retina. IEEE Transactions on Medical Imaging, 23(4):501-509, 2004. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 601 + ], + "type": "text", + "content": "[55] Yubo Tan, Kaifu Yang, Shixuan Zhao, and Yongjie Li. Retinal Vessel Segmentation with Skeletal Prior and Contrastive Loss. IEEE Transactions on Medical Imaging, 41(9):2238-2251, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "text", + "content": "[56] Devavrat Tomar, Guillaume Vray, Behzad Bozorgtabar, and Jean-Philippe Thiran. TeSLA: Test-Time Self-Learning with Automatic Adversarial Augmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 20341–20350, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[57] Chenglong Wang, Yuichiro Hayashi, Masahiro Oda, Hayato Itoh, Takayuki Kitasaka, Alejandro F Frangi, and Kensaku Mori. Tubular Structure Segmentation Using Spatial Fully Connected Network with Radial Distance Loss for 3D Medical Images. In Proceedings of the International Conference" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "on Medical Image Computing and Computer Assisted Interventions, pages 348-356, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[58] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. TENT: Fully Test-Time Adaptation by Entropy Minimization. Proceedings of the International Conference on Learning Representations, 2021. 1, 2, 3, 5, 6, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[59] Dong Wang, Zhao Zhang, Ziwei Zhao, Yuhang Liu, Yihong Chen, and Liwei Wang. PointScatter: Point Set Representation for Tubular Structure Extraction. In Proceedings of the European Conference on Computer Vision, pages 366-383, 2022. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 252 + ], + "type": "text", + "content": "[60] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual Test-Time Domain Adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 1, 2, 3, 6, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 254, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 319 + ], + "type": "text", + "content": "[61] Wei Wang, Zhun Zhong, Weijie Wang, Xi Chen, Charles Ling, Boyu Wang, and Nicu Sebe. Dynamically Instance-Guided Adaptation: A Backward-Free Approach for Test-Time Domain Adaptive Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 24090-24099, 2023. 2, 3, 6, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 320, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 385 + ], + "type": "text", + "content": "[62] Wenhan Wang, Jiale Zhou, Jin Zhao, Xun Lin, Yan Zhang, Shan Lu, Wanchen Zhao, Shuai Wang, Wenzhong Tang, and Xiaolei Qu. Interactively fusing global and local features for benign and malignant classification of breast ultrasound images. Ultrasound in Medicine Biology, 51(3):525-534, 2025. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[63] Yan Wang, Xu Wei, Fengze Liu, Jieneng Chen, Yuyin Zhou, Wei Shen, Elliot K. Fishman, and Alan L. Yuille. Deep Distance Transform for Tubular Structure Segmentation in CT Scans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "text", + "content": "[64] Ying Wang, Yuexing Peng, Wei Li, George C Alexandropoulos, Junchuan Yu, Daqing Ge, and Wei Xiang. DDU-Net: Dual-Decoder-U-Net for Road Extraction Using High-Resolution Remote Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-12, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "text", + "content": "[65] Qian Wu, Yufei Chen, Wei Liu, Xiaodong Yue, and Xiahai Zhuang. Deep Closing: Enhancing Topological Connectivity in Medical Tubular Segmentation. IEEE Transactions on Medical Imaging, 43(11):3990-4003, 2024. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "type": "text", + "content": "[66] Jiaqi Yang, Xiaoling Hu, Chao Chen, and Chialing Tsai. A Topological-Attention ConvLSTM Network and Its Application to EM Images. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Interventions, pages 217–228, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "text", + "content": "[67] Yanchao Yang and Stefano Soatto. FDA: Fourier Domain Adaptation for Semantic Segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4085-4095, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[68] Ziyun Yang and Sina Farsiu. Directional Connectivity-based Segmentation of Medical Images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 11525-11535, 2023. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[69] Chang'an Yi, Haotian Chen, Yifan Zhang, Yonghui Xu, Yan Zhou, and Lizhen Cui. From Question to Exploration:" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24133" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 541 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Can Classic Test-Time Adaptation Strategies Be Effectively Applied in Semantic Segmentation? In Proceedings of the ACM International Conference on Multimedia, pages 10085-10094, 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 183 + ], + "type": "text", + "content": "[70] Zitong Yu, Chenxu Zhao, Zezheng Wang, Yunxiao Qin, Zhuo Su, Xiaobai Li, Feng Zhou, and Guoying Zhao. Searching Central Difference Convolutional Networks for Face Anti-Spoofing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5295-5305, 2020. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 295, + 228 + ], + "type": "text", + "content": "[71] Longhui Yuan, Binhui Xie, and Shuang Li. Robust Test-Time Adaptation in Dynamic Scenarios. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 15922-15932, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[72] Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. DomainAdaptor: A Novel Approach to Test-Time Adaptation. In Proceedings of the IEEE International Conference on Computer Vision, pages 18971-18981, 2023. 2, 3, 6, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 328 + ], + "type": "text", + "content": "[73] Lefei Zhang, Meng Lan, Jing Zhang, and Dacheng Tao. Stagewise Unsupervised Domain Adaptation with Adversarial Self-Training for Road Segmentation of Remote-Sensing Images. IEEE Transactions on Geoscience and Remote Sensing, 60:1-13, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 384 + ], + "type": "text", + "content": "[74] Yizhe Zhang, Tao Zhou, Yuhui Tao, Shuo Wang, Ye Wu, Benyuan Liu, Pengfei Gu, Qiang Chen, and Danny Z Chen. TestFit: A Plug-and-Play One-Pass Test Time Method for Medical Image Segmentation. Medical Image Analysis, 92: 103069, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 429 + ], + "type": "text", + "content": "[75] Bowen Zhao, Chen Chen, and Shu-Tao Xia. DELTA: Degradation-Free Fully Test-Time Adaptation. In Proceedings of the International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 474 + ], + "type": "text", + "content": "[76] He Zhao, Huiqi Li, Sebastian Maurer-Stroh, and Li Cheng. Synthesizing Retinal and Neuronal Images with Generative Adversarial Nets. Medical Image Analysis, 49:14-26, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 541 + ], + "type": "text", + "content": "[77] Jiale Zhou, Zuoxun Hou, Hongyan Lu, Wenhan Wang, Wanchen Zhao, Zenan Wang, Dezhi Zheng, Shuai Wang, Wenzhong Tang, and Xiaolei Qu. A deep supervised transformer u-shaped full-resolution residual network for the segmentation of breast ultrasound image. Medical Physics, 50 (12):7513-7524, 2023. 1" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24134" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/60f819c6-a0ae-4fb0-be50-050c7068618b_content_list.json b/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/60f819c6-a0ae-4fb0-be50-050c7068618b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2ffb70c2c81a1c509e8bf8832465c41b259bc29c --- /dev/null +++ b/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/60f819c6-a0ae-4fb0-be50-050c7068618b_content_list.json @@ -0,0 +1,1836 @@ +[ + { + "type": "text", + "text": "TorchAdapt: Towards Light-Agnostic Real-Time Visual Perception", + "text_level": 1, + "bbox": [ + 161, + 130, + 836, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Khurram Azeem Hashmi $^{1,2}$", + "bbox": [ + 161, + 179, + 380, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Karthik Palyakere Suresh1,2", + "bbox": [ + 428, + 180, + 648, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Muhammad Zeshan Afzal1,2", + "bbox": [ + 385, + 198, + 611, + 214 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ DFKI", + "bbox": [ + 297, + 215, + 352, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ RPTU Kaiserslautern-Landau", + "bbox": [ + 457, + 215, + 700, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "khurram_azeem Hashmi@dfki.de", + "bbox": [ + 372, + 237, + 619, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Didier Stricker1,2", + "bbox": [ + 694, + 180, + 833, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Low-light conditions significantly degrade the performance of high-level vision tasks. Existing approaches either enhance low-light images without considering normal illumination scenarios, leading to poor generalization, or are tailored to specific tasks. We propose TorchAdapt, a real-time adaptive feature enhancement framework that generalizes robustly across varying illumination conditions without degrading performance in well-lit scenarios. TorchAdapt consists of two complementary modules: the Torch module enhances semantic features beneficial for downstream tasks, while the Adapt module dynamically modulates these enhancements based on input content. Leveraging a novel light-agnostic learning strategy, TorchAdapt aligns feature representations of enhanced and well-lit images to produce powerful illumination-invariant features. Extensive experiments on multiple high-level vision tasks, including object detection, face detection, instance segmentation, semantic segmentation, and video object detection, demonstrate that TorchAdapt consistently outperforms state-of-the-art low-light enhancement and task-specific methods in both low-light and light-agnostic settings. TorchAdapt thus provides a unified, flexible solution for robust visual perception across diverse lighting conditions.", + "bbox": [ + 89, + 316, + 485, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 694, + 220, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in visual perception tasks [13, 31, 77, 82] are highly effective at extracting features for complex tasks when provided with high-quality inputs. However, replicating this performance is challenging when data is scarce, noisy, or expensive to obtain. A common issue is varying illumination, particularly low-light conditions, which hinders the robustness and reliability of safety-critical visual perception systems [1, 3, 7, 63].", + "bbox": [ + 89, + 719, + 483, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A common strategy to improve low-light performance involves pre-processing images with Low-Light Image Enhancement (LLIE) methods [17, 24, 33, 44, 50, 51, 71, 84, 85] before feeding them into high-level vision mod", + "bbox": [ + 89, + 840, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fdef7c1f1dd5cc82fa6f742b32d83cd361a3634712105dbd6ce9d14c33371195.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 284, + 624, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7501f044036cdd8bc76406a59833e2f05e74cd09ccee5d765a4f21363bea0a95.jpg", + "image_caption": [ + "(a) Low-Light Vision" + ], + "image_footnote": [], + "bbox": [ + 625, + 285, + 736, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3158cf49653bf7978afeaf8384958c3197e9d7d59141b96ad775fa98e5129ce4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 285, + 903, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/393d6a987f61d20407ff9cce26d641509d477f9781212f06f5c3bc4f50e7bef8.jpg", + "image_caption": [ + "Figure 1. Prior downstream methods in low-light vision assume that all samples belong to low-light conditions as shown in (a), leading to underperformance in varying illumination environments, depicted in (b). TorchAdapt alleviates this bottleneck via learning powerful illumination-invariant features and achieves consistent gains across low-light and light-agnostic settings on Object Detection (OD), Face Detection (FD), Instance Segmentation (IS), Semantic Segmentation (SS), and Video Object Detection (VOD). Best viewed on the screen." + ], + "image_footnote": [], + "bbox": [ + 516, + 412, + 620, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8938b69d45be7c366a1c6a63ef8e45639c0e7021cb67de8f76069e68c29cb623.jpg", + "image_caption": [ + "(b) Light-Agnostic Vision" + ], + "image_footnote": [], + "bbox": [ + 620, + 414, + 733, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0895c393afe8577eb301659288cba0e1da35a2003dc2cb84ff8c7c41ce47946d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 414, + 903, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "els. However, this simple integration often leads to weaker gains [16, 26], as these arts conform to human visual perception (optimized on pixel-level objectives like L1 or MSE losses [4]) instead of machine visual perception. Moreover, applying these enhancement methods indiscriminately can degrade the performance on images/videos captured under standard lighting conditions, as alterations intended for low-light scenarios [47, 60, 80, 81] may be unnecessary or even detrimental in well-lit environments [15, 40, 58, 78].", + "bbox": [ + 511, + 700, + 908, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address low-light vision tasks, recent methods employ transfer learning [16, 59, 72] and domain adaptation techniques [20, 32, 48]. Despite improvements, these efforts are carefully designed for a single high-level vision task, such", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5645", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as object detection [16] or semantic segmentation [59, 72], and require cumbersome multi-stage training to adapt to each dense task [20, 32, 48]. Moreover, they often assume consistently low-light data, leading to performance degradation on well-lit images and videos, sometimes worse than baselines (see Table 2). In practical applications with unpredictable lighting—such as autonomous driving [7, 37] at different times of the day, surveillance systems [63] monitoring both indoor and outdoor environments or robots [3] operating between well-lit and low-lit areas—these limitations become critical. This highlights the need for high-level vision models to be light-agnostic—capable of maintaining high performance across varying lighting conditions, as visualized in Fig. 1(b).", + "bbox": [ + 89, + 90, + 485, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose TorchAdapt, a real-time, light-agnostic adaptive feature enhancement framework designed to improve the performance of any high-level vision task under varying illumination. TorchAdapt comprises two primary modules: Torch and Adapt, illustrated in Fig. 2. The Torch module is a lightweight feature enhancement network that enriches semantic representations favorable for downstream vision tasks. The Adapt module learns to modulate this feature enhancement dynamically based on input data, amplifying or reducing the effect of the Torch module.", + "bbox": [ + 89, + 305, + 485, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "TorchAdapt builds on the consensus [26, 38] that low-to-normal light adaptation lacks a physical model, leading to unexpected artifacts during optimization when transforming images in pixel space. Therefore, aligning representations in the feature space is a more practical approach for high-level vision tasks [48, 65]. We optimize TorchAdapt using a non-contrastive self-supervised learning objective [12, 23] to align representations from low-light and well-lit images. However, unlike prior works operating in a similar direction [16, 48, 65], TorchAdapt also learns not to adversely affect representations of well-lit images. As illustrated in Fig. 3, we take a well-lit image $V_{1}$ and its synthetically generated low-light counterpart $V_{2}$ , pass them through TorchAdapt to obtain enhanced images $\\hat{V}_{1}$ and $\\hat{V}_{2}$ , and feed them along with $V_{1}$ into a pre-trained frozen visual encoder [19, 27, 55] to generate representations. Since the encoder is frozen, only TorchAdapt is optimized to help the encoder produce well-lit representations (like $V_{1}$ ) for $\\hat{V}_{1}$ and $\\hat{V}_{2}$ .", + "bbox": [ + 89, + 459, + 485, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thanks to the conceptual simplicity, TorchAdapt enjoys two benefits. ① It can be trained with a simple self-supervised loss on a general-purpose dataset like ImageNet [18], circumventing large-scale real-world low-light datasets. ② The training of TorchAdapt is independent of the high-level vision task and needs to be done only once to align with the backbone networks [6, 19, 27]. Later, TorchAdapt can be plugged into any downstream tasks in a light-agnostic manner. We demonstrate the effectiveness of TorchAdapt on five representative visual perception tasks, including object detection [21, 47], face detection [78, 80], semantic", + "bbox": [ + 88, + 734, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "segmentation [15, 60], instance segmentation [10, 40], and video object detection [58, 81] under both low-light and light-agnostic settings. For all tasks, TorchAdapt brings consistent and significant gains over baselines under both low-light and light-agnostic settings, outperforming previous state-of-the-art task-specific and LLIE methods (see Fig. 1). Moreover, TorchAdapt accomplishes all this with negligible computational overhead (see Table 6).", + "bbox": [ + 511, + 90, + 908, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. Method: We propose TorchAdapt, a real-time, light-agnostic adaptive feature enhancement framework that improves the performance of any high-level vision task under varying illumination without compromising accuracy in well-lit conditions. Simplicity: Thanks to its architectural innovations and self-supervised objective, TorchAdapt significantly simplifies the training strategy for any high-level vision task pipeline, circumventing corresponding paired images/videos, as shown in Fig 3. Flexibility: TorchAdapt is independent of vision encoders and downstream task architectures, demonstrating model-agnostic generality (see § 5). Any zero-reference LLIE method [24, 33, 50, 67] can serve as the Torch module in our framework to achieve light-agnostic performance (see Fig. 5). Results: We conduct extensive experiments on five visual perception tasks involving images and videos under both low-light and light-agnostic settings using curated datasets. TorchAdapt outperforms state-of-the-art task-specific and LLIE methods across all settings when integrated with the same baselines (see Fig. 1).", + "bbox": [ + 511, + 213, + 910, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 532, + 658, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Enhancing Low-Light Images", + "text_level": 1, + "bbox": [ + 511, + 559, + 779, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Low-light Image Enhancement (LLIE) aims to improve the visual quality of images captured under insufficient lighting conditions. Modern learning-based approaches have become prevalent, thanks to the emergence of low-light datasets [5, 8, 25, 43, 70]. Supervised methods [70, 73, 74, 84, 86] leverage these datasets to design effective networks trained to predict normal-light images from low-light ones. Unsupervised LLIE approaches [30, 34, 79] employ adversarial learning from unpaired supervision to relax the requirement of paired normal-low light data. However, all of these works still rely on specific training data, which limits their generalizability to unseen scenarios. Another line of work known as zero-reference methods [24, 33, 44, 50, 67] get rid of both paired and unpaired data to enhance low-light images. These arts design enhancement networks that optimize on a set of non-reference loss functions. Since these works are mainly designed to conform to human visual perception, they bring sub-optimal gains for machine visual perception tasks. Furthermore, they are not light-agnostic (capable of handling both normal and low-light images). This work bridges this gap by complementing these zero-reference methods,", + "bbox": [ + 511, + 583, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5646", + "bbox": [ + 482, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6602a677e24ffe63a76d0f3c9d0e5d66f941f01870e4ff0cd1d6c44b1dca1f92.jpg", + "image_caption": [ + "Figure 2. TorchAdapt architectural overview. The Torch module enhances feature representation by applying $3 \\times 3$ convolutions and element-wise operations (Eq. 1). The Adapt module performs adaptive modulation, creating a content-aware response based on the input $I$ (Eq. 2). Both modules are fused via Adaptive Fusion, integrating feature enhancements from Torch and contextual adaptations from Adapt. This adaptive fusion produces illumination-invariant features, improving robustness across varying lighting conditions. Adaptive Fusion is explained in Eq. 4." + ], + "image_footnote": [], + "bbox": [ + 174, + 88, + 831, + 218 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where they can be employed as a Torch in our TorchAdapt framework to improve light-agnostic visual perception.", + "bbox": [ + 89, + 314, + 485, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Low-Light Visual Perception", + "text_level": 1, + "bbox": [ + 89, + 359, + 349, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This direction of work considers machine perception as the criteria for success while enhancing low-light images.", + "bbox": [ + 89, + 383, + 482, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Low-light Object Detection (LLOD). Object detection, as a fundamental problem in computer vision, has also been well-explored in low-light vision [16, 20, 39, 51, 61, 62, 64-66]. Based on the dataset, objects vary and can be generic [47, 52] or specific like pedestrians [53], and human faces [80]. One category of LLOD methods [51, 56, 62, 65] jointly learns the enhancement and detection task in an end-to-end fashion to improve the overall detection performance. Other approaches involve learning a low-light detector through fusing multiple models [61] or resorting to domain adaptation frameworks [16, 20, 64], minimizing the distribution variance between normal and low-light images. Most of these works are specifically designed to tackle the dark object detection task. Furthermore, when these methods are optimized for LLOD, they underperform on normal-light images. In contrast, TorchAdapt is fully task-agnostic for any high-level vision task, and it does not worsen the performance of the detector on normal-light data.", + "bbox": [ + 89, + 415, + 483, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Low-light Tasks-Agnostic Methods. Recently, various low-light task-agnostic methods [20, 26, 32, 48, 65] have emerged in the literature. Unlike prior approaches, they are more flexible and can be integrated to improve any downstream task under low-light vision. Our work follows the same spirit and offers similar flexibility. FeatEnHancer [26], as one of the representative works, learns to enhance the hierarchical features suitable for downstream tasks. Other seminal attempts [20, 32, 48] treat low-light perception tasks as domain adaption (DA) learning. They exploit well-lit source domain data [40, 78] and learn to adapt to low-light target domain data [47, 80]. Although flexible, the DA methods rely on multi-stage training frameworks for each task [48]. In contrast, our TorchAdapt is pre-trained with the frozen back", + "bbox": [ + 89, + 688, + 486, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bone on the domain-independent data like ImageNet [18]. Moreover, it is important to recognize the efforts of recent real low-light datasets [10, 35, 60, 68, 75, 81] that enable the increasing research on several downstream tasks in low-light vision.", + "bbox": [ + 511, + 314, + 906, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Learning Equivariant Representations in LowLight Vision with Self-Supervised Objectives", + "text_level": 1, + "bbox": [ + 511, + 406, + 906, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Annotating low-light vision datasets is a much more difficult task than annotating well-lit datasets. Hence, it is a common practice to apply self-supervised learning to learn illumination-invariant features [16, 65] or aligning well-lit and low-light data distributions [38, 48, 66]. MAET [16] adopts the autoencoding transformation [83] to decode illumination-degrading parameters and detection predictions. SACC [65] merges the pretext tasks of rotation prediction [22] and jigsaw puzzling [54] to learn illumination enhancement from low-light images. Most related to our work is SimMinMax [48], which adopts the BYOL [23] framework to learn model-level adaptation by maximizing the similarity of features between images at night and daytime. However, since SimMinMax [48] learns to adapt from daytime to nighttime images, it needs to be trained for each high-level vision task, requiring the corresponding daytime labeled dataset. In contrast, TorchAdapt leverages self-supervised objectives to align backbone features during the pretext task. Furthermore, it can be trained on any general-purpose dataset. This significantly simplifies the training procedure of TorchAdapt, as compared to SimMinMax, on any high-level vision task. TorchAdapt is also unique to prior works in its ability to handle light-agnostic data.", + "bbox": [ + 511, + 445, + 908, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. TorchAdapt", + "text_level": 1, + "bbox": [ + 511, + 827, + 640, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section formally introduces TorchAdapt, explaining the proposed Torch in § 3.1 and Adapt in § 3.2 modules, followed by the light-agnostic learning in § 3.3.", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5647", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ed71c641c4623764435fb269b0b8cb1c6f6bfd6a39089703dc0d05e1aa566bb1.jpg", + "image_caption": [ + "Figure 3. Light-Agnostic learning pipeline. $D(.)$ is a darkening function applied to the well-lit image $V_{1}$ to synthesize the low-light view $V_{2}$ . Both $V_{1}$ and $V_{2}$ are processed by TorchAdapt and then fed into a frozen pre-trained vision encoder $f$ . A cosine similarity loss optimizes TorchAdapt to produce well-lit representations for both $V_{1}$ and $V_{2}$ , enabling it to learn illumination-invariant features. The TorchAdapt(.) function integrates Eq. 1 and Eq. 4. Refer to § 3.3 for a complete explanation of each symbol." + ], + "image_footnote": [], + "bbox": [ + 119, + 95, + 467, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.Torch", + "text_level": 1, + "bbox": [ + 89, + 391, + 173, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Torch module in our TorchAdapt framework serves as an enhancement network as in Low-light Image Enhancement (LLIE) approaches [24, 33, 50]. However, unlike these methods—which focus on improving visual quality for human perception—our Torch module aims to boost semantic representations favorable for downstream vision tasks under varying lighting conditions. Given an input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ , the Torch module computes an enhancement map $\\mathcal{H}_{\\theta}(\\mathbf{I})$ , parameterized by network weights $\\theta$ . The transformed image $\\mathbf{T}$ is obtained by:", + "bbox": [ + 89, + 414, + 483, + 565 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} = \\mathbf {I} \\odot (1 + \\mathcal {H} _ {\\theta} (\\mathbf {I})) \\text {, w h e r e} \\mathcal {H} _ {\\theta} (\\mathbf {I}) = \\tanh (\\mathcal {F} (\\mathbf {I}; \\theta)) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 580, + 483, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise multiplication. The $\\tanh(\\cdot)$ activation ensures that the adjustment values are in the range $[-1, 1]$ , allowing the model to both amplify (positive values) and attenuate (negative values) features. This formulation enables the model to adapt to the image's content effectively. Torch network. The Torch network $\\mathcal{F}(\\mathbf{I};\\theta)$ is a lightweight convolutional network composed of an initial feature extraction layer followed by a residual block, as illustrated in Fig. 2. The residual learning enables the Torch module to focus on learning the necessary adjustments to improve the input image rather than reconstructing it entirely. Enhancing semantically important features while preserving original content makes the Torch module crucial in maintaining and improving performance in downstream vision tasks. Notably, thanks to the formulation explained in Eq 1, our Torch module only has 11,075 trainable parameters and requires 0.712 GFLOPs for $\\mathbf{I}$ of size $256 \\times 256 \\times 3$ , making it suitable for real-time high-level vision tasks under varying lighting conditions.", + "bbox": [ + 89, + 612, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Adapt", + "text_level": 1, + "bbox": [ + 513, + 90, + 599, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the Torch module enhances the input image $\\mathbf{I}$ to produce $\\mathbf{T}$ , applying this enhancement uniformly may not be optimal. Different regions within an image require varying degrees of enhancement based on content (low/well-lit images) and context (downstream task). To address this, we introduce the Adapt module, which learns a content-adaptive modulation matrix in the feature space. This allows the model to selectively amplify or suppress the enhancements produced by the Torch module, improving their effectiveness for downstream vision tasks in varying illumination scenarios, effectively making it light-agnostic.", + "bbox": [ + 511, + 112, + 906, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the Adapt module computes an Adaptive Matrix $\\mathbf{M} \\in \\mathbb{R}^{H \\times W}$ , where $H$ and $W$ are the spatial dimensions of $\\mathbf{I}$ . The purpose of the Adaptive Matrix $\\mathbf{M}$ is to modulate the feature enhancements from Torch $\\mathbf{T}$ before they are applied to $\\mathbf{I}$ . $\\mathbf{M}$ is computed as:", + "bbox": [ + 511, + 279, + 906, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\sigma \\left(\\mathcal {G} _ {\\phi} (\\mathbf {I})\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 361, + 906, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{G}_{\\phi}$ is a lightweight convolutional network parameterized by weights $\\phi$ , and $\\sigma(\\cdot)$ denotes the sigmoid activation function ensuring that the values of $\\mathbf{M}$ are in the range $[0,1]$ . We feed the input image $\\mathbf{I}$ into the Adapt module to compute $\\mathbf{M}$ , ensuring that the image content directly influences the modulation.", + "bbox": [ + 511, + 385, + 905, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adapt network. As shown in Fig. 2, the convolutional network $\\mathcal{G}_{\\phi}$ in the Adapt module is designed to extract hierarchical features, which are proven to be beneficial for high-level vision tasks such as object detection and semantic segmentation [26, 45, 46]. To capture multi-scale contextual information, the network consists of three $3\\times 3$ convolutional layers with increasing channel dimensions $C_1$ , $C_2$ , and $C_3$ , respectively. Each layer is followed by a batch normalization [29] and a ReLU activation function. This architecture allows the network to learn a hierarchy of features from low-level edges and textures to high-level semantic information. Finally, an additional $3\\times 3$ convolutional layer maps the features from $C_3$ to a single channel and applies a sigmoid activation function to generate the Adaptive Matrix $\\mathbf{M}\\in [0,1]^{H\\times W}$ . This design balances expressiveness and computational efficiency, enabling the Adapt module to adjust enhancements based on both local and global features, which is crucial for content-adaptive modulation.", + "bbox": [ + 511, + 476, + 908, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Illumination-aware scaling. We introduce a scale factor $\\gamma$ to ensure that the enhancement adapts appropriately to different illumination levels. This scale factor modulates the M globally based on the image's overall brightness, allowing the model to apply stronger enhancements to darker images and milder adjustments to well-lit images, maintaining visual consistency and preventing artifacts. We compute $\\gamma$ as a function of the average luminance $L(\\mathbf{I})$ of the input image:", + "bbox": [ + 511, + 747, + 908, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma = \\frac {\\kappa}{1 + \\exp (k (L (\\mathbf {I}) - L _ {0}))}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 875, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5648", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\kappa$ is the maximum scale factor (empirically set to 3 unless stated otherwise). $k$ controls the slope of the sigmoid function, and $L_{0}$ is the luminance value at the sigmoid's midpoint. The average luminance $L(\\mathbf{I})$ is computed following the standard luminance calculation in color science [2, 42, 87].", + "bbox": [ + 89, + 90, + 483, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive Fusion. The final enhanced image $\\mathbf{E}$ is then obtained by modulating the enhancement from the Torch module with the illumination-aware scaled Adaptive Matrix:", + "bbox": [ + 89, + 181, + 485, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} = \\mathbf {I} + ((\\gamma \\cdot \\mathbf {M}) \\odot \\mathbf {T}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 236, + 483, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise multiplication, and $\\mathbf{T}$ is obtained from Eq. (1). This formulation in Eq. 4 integrates both content-based modulation (through $\\mathbf{M}$ ) and illumination-based scaling (through $\\gamma$ ), enabling the model to adaptively enhance images under varying lighting conditions.", + "bbox": [ + 89, + 261, + 483, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Light-Agnostic Learning", + "text_level": 1, + "bbox": [ + 89, + 345, + 316, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Achieving robust performance across varying lighting conditions requires our model to learn representations invariant to illumination changes. Collecting labeled data under all possible lighting conditions is impractical. Hence, we adopt a self-supervised learning (SSL) strategy inspired by BYOL [23] to learn illumination-invariant features. Unlike prior works that design cumbersome pipelines to incorporate SSL into each high-level vision task [16, 48, 65], we leverage SSL to pre-train our TorchAdapt module with a pretext task that aligns backbone representations [19, 27] regardless of illumination. This approach eliminates the need for task-specific labeled data or low-light data, allowing training on general-purpose datasets like ImageNet [18]. Figure 3 illustrates the pre-training pipeline of TorchAdapt.", + "bbox": [ + 89, + 367, + 483, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Objective Function. Given a well-lit image $V_{1}$ , we generate its low-light counterpart $V_{2} = D(V_{1})$ using a darkening function $D(\\cdot)$ , which can be a gamma transformation, an illumination-degradation pipeline [16], or a darkening module from [48]. Let $f$ denote the frozen pre-trained backbone network (e.g., ResNet [27], ViT [19]). We pass $V_{1}$ and $V_{2}$ through the TorchAdapt module to obtain enhanced images $\\hat{V}_{1}$ and $\\hat{V}_{2}$ , respectively. The enhanced images are processed by the backbone $f$ and a projector network $g$ to obtain feature representations: $z_{1}^{\\prime} = g(f(\\hat{V}_{1}))$ and $z_{2} = g(f(\\hat{V}_{2}))$ . The target representation is obtained by passing the original well-lit image $V_{1}$ through $f^{\\prime}$ and $g^{\\prime}$ : $y_{1} = g^{\\prime}(f^{\\prime}(V_{1}))$ . Our training objective minimizes the cosine similarity loss $\\mathcal{D}_{\\mathrm{cos}}$ between the enhanced image representations and the well-lit target representation:", + "bbox": [ + 89, + 579, + 485, + 805 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left(1 - \\frac {z _ {1} ^ {\\prime}}{\\| z _ {1} ^ {\\prime} \\| _ {2} \\| y _ {1} \\| _ {2}}\\right) + \\left(1 - \\frac {z _ {2} ^ {\\top} y _ {1}}{\\| z _ {2} \\| _ {2} \\| y _ {1} \\| _ {2}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 811, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\| _2$ denotes the $\\ell_2$ norm of a vector. Since we employ identical, frozen backbone networks $f$ , the loss function guides TorchAdapt to enhance the low-light image $\\hat{V}_2$", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c75950f2f37899bdc86d43e1a36906c387eaf13d8446bfa7f78a9f8280345623.jpg", + "image_caption": [ + "Figure 4. TorchAdapt's downstream integration pipeline. TorchAdapt works as a plug-and-play, backbone-agnostic feature enhancement module, seamlessly integrating into various downstream vision tasks. Given any image or video input across diverse lighting conditions, TorchAdapt produces illumination-invariant representations, significantly boosting performance without requiring task-specific modifications." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 901, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "so its representation aligns with that of the well-lit image $V_{1}$ . Simultaneously, it discourages over-enhancement of the already well-lit image $\\hat{V}_{1}$ , as its representation should remain consistent with $V_{1}$ . This training strategy enables TorchAdapt to effectively enhance low-light images while preserving well-lit ones, promoting illumination invariance in the learned representations.", + "bbox": [ + 511, + 290, + 906, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. The training of TorchAdapt is fast and efficient due to its lightweight architecture and the absence of gradient propagation through the frozen backbone network. We train TorchAdapt for 5 epochs on the ImageNet [18] dataset using a single GPU, with a learning rate of 0.0001 and a batch size of 256. Since the backbone networks are pre-trained and we aim for consistent views, we apply only color jitter and random horizontal flip as data augmentations. Pre-training TorchAdapt with a ResNet-50 backbone takes $\\sim 6$ hours on a single A100 GPU. Once trained, TorchAdapt can be seamlessly integrated into any downstream vision task, as illustrated in Fig. 4.", + "bbox": [ + 511, + 396, + 908, + 578 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8171dd167c61ef425d4fe270ed9c336d6c9f63b6f6c55d6c1798078e2c4b4f06.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetTask#Cls#Train#Val
Low-light (LL)
ExDark [47]Object detection1258911472
DARK FACE [80]Face detection15400600
ACDC [60]Semantic segmentation19400106
LIS [10]Instance segmentation81561669
DarkVision [81]Video object detection426*6*
Light-Agnostic (LA)
COCO [21] ∩ ExDark [47]Object detection12923575156
WIDER FACE [78] ∩ DARK FACE [80]Face detection1182803822
CityScapes [15] ∩ ACDC [60]Semantic segmentation193375606
COCO [40] ∩ LIS [10]Instance segmentation8443212550
ImageNet VID [58] ∩ DarkVision [81]Video object detection2401*51*
", + "bbox": [ + 514, + 588, + 906, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Datasets Statistics.*: video samples. Light-Agnostic datasets are obtained through Eq. 6", + "bbox": [ + 511, + 727, + 906, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 768, + 643, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct extensive experiments on several representative visual tasks to demonstrate the superiority of TorchAdapt in both low-light and light-agnostic settings. These tasks include object detection [40, 47, 52], face detection [78, 80], instance segmentation [10, 40], semantic segmentation [15, 60], and video object detection [58, 81]. This section first presents the experimental settings, comparing TorchAdapt", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5649", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/66321bd3257498f3523c28c810e497bf99d469b9c9e78f725013b620322d402e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Tasks → Light Settings → MethodsObject Det.Face Det.Instance Seg.Semantic Seg.Video Object Det.
LLLALLLALLLALLLALLLA
mAP50mAPmAP50mAPAP50APAP50APmAPSegAPmAPSegAPmIoUmIoUmAPmAP
Baseline78.844.262.433.147.319.952.628.054.046.041.533.949.675.532.851.3
Enhancement
Zero-DCE [24]76.242.757.629.547.420.147.622.951.543.539.232.142.559.47.8328.3
SCI [84]75.542.257.929.745.119.345.421.752.344.235.128.342.157.46.7125.7
URetinexNet [71]74.642.057.329.244.718.346.622.359.543.137.429.741.258.86.6825.1
IAT [17]74.141.052.430.542.115.937.117.449.742.336.430.141.756.26.2424.5
Task-specific
Zero-DCE [24]‡79.144.462.933.250.421.150.123.559.152.744.435.744.761.034.151.5
SCI [50]‡79.244.263.033.451.622.254.729.958.451.043.734.943.559.932.950.7
MAET [16]‡79.244.357.327.344.318.747.923.1--------
Xue et al. [76]----46.117.744.221.0----42.169.7--
FeatEnHancer [26]‡79.244.862.933.347.219.951.624.559.652.444.936.150.176.134.652.7
TorchAdapt80.145.464.034.153.723.159.032.561.153.645.236.650.876.339.159.1
vs. prev. SoTA+0.9+0.6+1.0+0.7+2.1+0.9+4.3+2.6+1.5+0.9+0.3+0.5+0.7+0.2+4.5+6.4
", + "bbox": [ + 93, + 87, + 906, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison across five tasks under both Low-Light (LL) and Light-Agnostic (LA) settings. ‡: all these methods are reproduced by us with their official codebase and trained end-to-end for each task. The rest of the results are obtained from their official checkpoints. - indicates that the method is not applicable. TorchAdapt outperforms prior SOTA across all five tasks in LL and LA settings.", + "bbox": [ + 89, + 316, + 906, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with baselines, existing low-light image enhancement (LLIE) methods, and state-of-the-art task-specific approaches. We then examine critical design choices for TorchAdapt through ablation studies. Finally, we validate the generalizability of TorchAdapt by comparing it with recent state-of-the-art domain adaptation methods [20, 32, 48].", + "bbox": [ + 88, + 371, + 480, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Settings. We perform experiments under two settings for each task:", + "bbox": [ + 89, + 463, + 482, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1)- Low-Light (LL) Setting: Following prior works [16, 26, 76], we utilize publicly available low-light benchmarks specific to each task. These benchmarks consist exclusively of images captured under low-light conditions.", + "bbox": [ + 89, + 494, + 483, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2)- Light-Agnostic (LA) Setting: To introduce a more challenging and realistic evaluation, we curate datasets that encompass both well-lit and low-light images. We achieve this by selecting classes common to both well-lit and low-light datasets. Mathematically, for a given task, let $\\mathcal{C}_{\\mathrm{WL}}$ and $\\mathcal{C}_{\\mathrm{LL}}$ denote the sets of classes in the well-lit and low-light datasets, respectively. We define the set of common classes as $\\mathcal{C}_{\\mathrm{com}} = \\mathcal{C}_{\\mathrm{WL}} \\cap \\mathcal{C}_{\\mathrm{LL}}$ . The LA dataset is then formed by combining all images from both datasets that belong to classes in $\\mathcal{C}_{\\mathrm{com}}$ :", + "bbox": [ + 88, + 555, + 482, + 705 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\mathrm {L A}} = \\left\\{\\left(x, y\\right) \\mid (x, y) \\in \\mathcal {D} _ {\\mathrm {W L}} \\cup \\mathcal {D} _ {\\mathrm {L L}}, y \\in \\mathcal {C} _ {\\mathrm {c o m}} \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 718, + 482, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Where $\\mathcal{D}_{\\mathrm{WL}}$ and $\\mathcal{D}_{\\mathrm{LL}}$ are the well-lit and low-light datasets, respectively, and $(x,y)$ represents an image and its label. We summarize the key statistics of the employed benchmarks for both settings in Table 1. We provide complete implementation details for each experiment in Appendix A.", + "bbox": [ + 88, + 750, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons with State-of-the-Arts. We directly compare our approach with several LLIE methods, including Zero-DCE [24], IAT [17], SCI [50], and URetinexNet [71]. Following common practice [16, 26, 76], we use their released checkpoints to enhance all images before feeding them into", + "bbox": [ + 88, + 824, + 483, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the detector for performance evaluation. Notably, due to their zero-reference nature requiring no paired images, some of these LLIE methods [24, 50] can be trained end-to-end on downstream vision tasks. Therefore, besides evaluating their conventional enhancement results, we integrate them into each task and report their performance, similar to low-light task-specific works like MAET [16] and FeatEnHancer [26] (indicated with $\\ddagger$ in Table 2). Our findings reveal that end-to-end training of such LLIE works $[24, 50]^1$ significantly improves their performance on downstream vision tasks in both LL and LA settings. Concurrent to our work, we notice that two new methods [28, 38] appear in the literature on improving low-light object detection. However, since their code and checkpoints were not released by the time of the submission, we could not draw direct comparisons with them. Next, we discuss experiments on each task.", + "bbox": [ + 511, + 371, + 906, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Object Detection", + "text_level": 1, + "bbox": [ + 511, + 626, + 679, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Settings. Following common practice [16] in LL object detection, we adopt YOLOv3 [57] as the baseline detector. For the LL setting, COCO [40] pre-trained weights are utilized to fine-tune YOLOv3 on the ExDark dataset [47]. We train the model from scratch in the LA setting, as COCO images are also included in the training data (see Table. 1). Results. As shown in Table 2, TorchAdapt achieves a superior $\\mathrm{mAP}_{50}$ of 80.1 in LL and 64.1 in LA, compared to the baseline's 78.8 and 62.4. End-to-end training boosts the performance of enhancement methods like Zero-DCE [33] and SCI [50]; however, TorchAdapt still surpasses them, and the low-light object detection method FeatEnHancer [26] by $+0.9$ and $+1.1$ in LL and LA settings, respectively. These results affirm the efficacy of TorchAdapt for object detec", + "bbox": [ + 511, + 648, + 908, + 861 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1We select these works due to their efficiency and reproducibility. Experiments with more methods are in Appendix C.", + "bbox": [ + 511, + 875, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5650", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion across varying illumination conditions. We provide a qualitative comparison in Appendix B.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Face Detection", + "text_level": 1, + "bbox": [ + 89, + 130, + 241, + 146 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings. For face detection, we choose a different detector and adopt RetinaNet [41] to report results on the DARK FACE dataset [69, 80] in the LL setting and a combination of the DARK FACE and WIDER FACE [78] datasets in the LA setting. Following prior works [26], we resize images to a resolution of $1500 \\times 1000$ pixels and follow the $1 \\times$ schedule in MMDetection [9].", + "bbox": [ + 89, + 152, + 483, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Results in Table 2 reveal that TorchAdapt demonstrates superior performance on face detection under both LL and LA settings, outperforming the previous best method $(\\mathrm{SCI} \\ddagger [50])$ by $+2.1$ and $+4.3$ in mAP_50, respectively. The significant gains, especially in the challenging LA setting, highlight TorchAdapt's capacity to handle diverse illumination scenarios by adaptively modulating features based on content and illumination, thereby improving the robustness and accuracy of the baseline. We also compare TorchAdapt with recent domain adaptation methods [20, 48] using DSFD [36] in Sec. 4.7.", + "bbox": [ + 89, + 258, + 485, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Instance Segmentation", + "text_level": 1, + "bbox": [ + 89, + 434, + 302, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings. For the instance segmentation task, we adopt RTMDet-Ins-tiny $[49]^2$ as our baseline model and resize images to $640\\times 640$ pixels for training. Following Section 4.1, we use the COCO pre-trained weights of RTMDet-Ins-tiny for the LL setting and train the model from scratch for the LA setting (see Table 1).", + "bbox": [ + 89, + 455, + 483, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. As summarized in Table 2), TorchAdapt achieves SOTA SegAP performance in both low-light (LL) and light-agnostic (LA) settings. Under LL conditions, TorchAdapt attains a SegAP of 53.6, surpassing the previous best (FeatEnHancer†) by +1.2. In the LA setting, it achieves a SegAP of 36.6, slightly improving overFeatEnHancer† by +0.5. These results demonstrate TorchAdapt's effectiveness in enhancing instance segmentation across varying illumination by producing illumination-invariant features.", + "bbox": [ + 88, + 547, + 485, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Semantic Segmentation", + "text_level": 1, + "bbox": [ + 89, + 691, + 307, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings. Consistent with prior works [26, 76], we adopt DeepLabV3+[11] as our baseline model and follow [26, 76] for the LL setting. For the LA evaluation (Table 1), we train DeepLabV3+ from scratch using MMSegmentation [14]. Results. As shown in Table 2, TorchAdapt achieves an mIoU of 50.8 in the LL setting, surpassing the previous best (FeatEnHancer†) by +0.7, and an mIoU of 76.3 in the LA setting, improving overFeatEnHancer† by +0.2. These results highlight TorchAdapt's effectiveness for semantic segmentation under varying illumination conditions.", + "bbox": [ + 89, + 714, + 485, + 866 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9a1eb1374af3d822c6420a6151822a8cdffaf927d10f75265a404123489e4177.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
#Torch(§ 3.1)Adapt(§ 3.2)Scaling(§ 3.2)Object Det.Instance Seg.
LLLALLLA
1×××78.862.446.033.9
2✓××79.162.951.134.8
3✓✓×79.963.553.336.2
4✓×✓79.562.751.434.1
5✓✓✓80.164.053.636.6
", + "bbox": [ + 514, + 89, + 905, + 167 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8d0c465469dbe2148f7b50855a22233c597314fe8d77166ca7e99973c2babbe5.jpg", + "table_caption": [ + "Table 3. Contribution of each component. Highlighted settings are set as default." + ], + "table_footnote": [ + "(a) Scale factor $(\\kappa)$ in Eq. 3." + ], + "table_body": "
Scale factor (κ)ODSS
LLLALLLA
279.963.753.436.4
380.164.053.636.6
479.963.353.236.2
579.561.153.334.5
", + "bbox": [ + 521, + 210, + 696, + 261 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/88b41eb27354e01f4b82e718a576112b3aed18ddafa26394279f903437d53583.jpg", + "table_caption": [], + "table_footnote": [ + "(b) $D(\\mathbf{I})$ in $\\S 3.3$ and Fig. 3." + ], + "table_body": "
D(I)ODSS
LLLALLLA
ISP [16]79.461.747.629.9
D(I) in [48]79.663.153.236.3
Gamma curve80.164.053.636.6
", + "bbox": [ + 699, + 210, + 893, + 261 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablating design choices in TorchAdapt. Highlighted settings are set as default. OD and SS are object detection and semantic segmentation, respectively.", + "bbox": [ + 511, + 282, + 905, + 327 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Video Object Detection", + "text_level": 1, + "bbox": [ + 513, + 334, + 728, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Settings. In addition to image-based tasks, we evaluate the generalizability of TorchAdapt on video data. We present the dataset details in Table 1. Other experimental settings follow [26], with additional implementation details provided in Appendix A.", + "bbox": [ + 511, + 354, + 906, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We report the results of the challenging video object detection task in Table 2. While all LLIE methods perform poorly on this task, end-to-end training significantly boosts the performance of zero-reference methods [24, 50]. TorchAdapt, however, outperforms all methods under both LL and LA settings, achieving substantial gains of $+4.5$ and $+6.4$ , respectively. These improvements affirm TorchAdapt's robustness and effectiveness for video object detection across both low-light and light-agnostic settings.", + "bbox": [ + 511, + 431, + 908, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 575, + 679, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We ablate the components and design choices of TorchAdapt in object detection (OD) and instance segmentation (IS) tasks under (Low-light) LL and (Light-Agnostic) LA settings, following the same implementation details in § 4.1 and § 4.3, respectively. We provide more ablations in Appendix C.", + "bbox": [ + 511, + 598, + 908, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contribution of Each Component. Table 3 evaluates the impact of the Torch module, the Adapt module, and Illumination Scaling. Adding the Torch module alone (row 2) improves performance over the baseline with $+0.3\\%$ in LL OD and $+5.1\\%$ in LL IS, highlighting the benefit of feature enhancement in low-light vision. Introducing the Adapt module alongside Torch (row 3) yields significant improvements $(+0.8\\% / +0.6\\%)$ in LL/LA OD; $+2.2\\% / +1.4\\%$ in LL/LA IS), underscoring the importance of content-adaptive modulation. Employing scaling without the Adapt module (Row 4) offers minor gains in LL settings. However, it is detrimental in LA settings $(-0.2\\% / -0.7\\%)$ in OD/IS), indicating that content-based adaptation is more impactful than global scaling alone. The full TorchAdapt configuration, combining all three components (row 5), achieves the best results with", + "bbox": [ + 511, + 674, + 910, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "2https://github.com/open-mmlab/mmdetection/blob/main/configs/rtmdet/", + "bbox": [ + 89, + 875, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5651", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e603b3cb4e53545917c1c2f92488686b35b8e2f107fa9d9a38ce8ccdbfad4390.jpg", + "image_caption": [ + "Figure 5. Visualizing backbone features with SCI [50] on LA object detection using YOLOv3. (a) SCI's official checkpoint as an enhancer. (b) $\\mathrm{SCI}^{\\dagger}$ trained end-to-end. (c) SCI integrated into our TorchAdapt framework as a Torch. SCI with TorchAdapt extracts more representative features in both low and well-lit images." + ], + "image_footnote": [], + "bbox": [ + 93, + 92, + 480, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/39f0dd06dee8c9f16120ea873df2e77f9ba491e769d087cca921c04576e2393e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodmAP (%)
DSFD [36]16.1
CICony [32]18.4
SimMinMax [48]25.7
DAI-Net [20]‡28.0
DAI-Net+TorchAdapt31.9+3.9
", + "bbox": [ + 178, + 282, + 397, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "+1.3%/+1.6% in LL/LA OD and +7.6%/+2.5% in LL/LA IS over the baseline, confirming that these components complement each other and synergistically enhance performance in both LL and LA settings.", + "bbox": [ + 88, + 465, + 483, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Design Choices in TorchAdapt. Table 4 ablates important design choices in TorchAdapt. As shown in Table 4a, we empirically establish that when the scale factor $\\gamma$ in Eq. 3 is computed with $\\kappa = 3$ , it produces optimal performance. Results in Table 4b reveal that employing the Gamma Curve in $D(I)$ to generate $V_{2}$ in § 3.3 brings the biggest boost in performance across both LL and LA settings.", + "bbox": [ + 89, + 525, + 483, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Comparison with Domain-Adaptation Methods", + "text_level": 1, + "bbox": [ + 89, + 642, + 482, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate TorchAdapt's effectiveness for zero-shot domain adaptive face detection on the DARK FACE dataset [80]. Following settings from [20], we train the detector only on the WIDER FACE dataset [78]. As shown in Table 5, TorchAdapt integrated with DAI-Net [20], the current best-performing method, achieves a significant improvement of $+3.9\\%$ mAP, demonstrating its effectiveness in enabling robust zero-shot adaptation under diverse illumination conditions.", + "bbox": [ + 88, + 662, + 483, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Exploring TorchAdapt's Properties", + "text_level": 1, + "bbox": [ + 89, + 814, + 410, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model-Agnostic Generality. Experiments in § 4 confirm that TorchAdapt is a general-purpose module that can be integrated into any high-level vision task. Moreover, TorchAdapt as a framework is highly flexible: any zero-reference (unsu", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ee41f6e42f54df25a83aa563fd4d172a8a4f66b3f9ca9353912d2893c5abe8df.jpg", + "image_caption": [ + "I" + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 612, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b2a7897bd66a6e77f556447b7190a5df314b46bf3f6696eaef639d16f113a87.jpg", + "image_caption": [ + "TorchAdapt(I)" + ], + "image_footnote": [], + "bbox": [ + 614, + 89, + 707, + 142 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25b5ed0c93f79064680e3cc7139d3d8c17adda2977e3ec18b15eff4f99e37679.jpg", + "image_caption": [ + "I", + "Figure 6. Illustrating low-light image Enhancement as an emergent property of TorchAdapt on the DARK FACE validation set. More results can be found in Appendix B." + ], + "image_footnote": [], + "bbox": [ + 709, + 89, + 803, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d76615a5888f9954e5e83faa48066095eaedc649ff390d22ed8066855e23d58.jpg", + "image_caption": [ + "TorchAdapt(I)" + ], + "image_footnote": [], + "bbox": [ + 807, + 89, + 903, + 143 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/17783a468fc666ac2b016ad53ab0ddb1916d0c0074fda3e5ea9c5ddff04764ce.jpg", + "table_caption": [ + "Table 5. Zero-shot performance comparison of TorchAdapt with state-of-the-art domain adaptation methods on DARK FACE [80] using DSFD [36]. Results with $\\ddagger$ are reproduced using the original codebase. TorchAdapt enhances prior SOTA [20] by $+3.9\\mathrm{mAP}$ , demonstrating its effectiveness in zero-shot domain adaptation." + ], + "table_footnote": [], + "table_body": "
MethodmAP50(LL)mAP50(LA)#Params.(M)Latency (ms)
Baseline [57]78.862.465.221.9
+TorchAdapt80.1 (+1.3)64.0 (+1.6)65.3 (+0.1)23.7 (+1.8)
", + "bbox": [ + 514, + 226, + 903, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Computational analysis. With YOLOv3 [57], TorchAdapt introduces minimal computational overhead while bringing stronger gains across both LL and LA settings.", + "bbox": [ + 511, + 273, + 906, + 315 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pervised) LLIE methods [33, 50, 85] can serve as the Torch component to improve performance under varying illuminations. To verify this, we adopt SCI [50] as the Torch module in TorchAdapt and visualize the learned backbone features in Fig.5(c). Using SCI within TorchAdapt yields highly representative backbone features for both low-light and well-lit images. Appendix C provides additional examples and quantitative analyses.", + "bbox": [ + 511, + 334, + 906, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Enhancement as an Emergent Property. Although TorchAdapt is designed to produce illumination-invariant features for high-level vision tasks without employing any explicit enhancement loss functions during training, we observe that it inherently enhances low-light images. We illustrate this effect in Fig. 6. This emergent behavior likely results from the pre-training objective of aligning representations between well-lit and low-light images (see § 3.3), which encourages the network to implicitly adjust low-light images toward well-lit representations.", + "bbox": [ + 511, + 457, + 908, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational Analysis. We assess the computational efficiency of TorchAdapt using YOLOv3 [57] on the object detection task. As shown in Table 6, integrating TorchAdapt incurs minimal additional computational cost while delivering significant performance improvements. This demonstrates that TorchAdapt is compatible with lightweight models and suitable for real-time applications.", + "bbox": [ + 511, + 609, + 908, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 737, + 638, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce TorchAdapt, a real-time, general-purpose, light-agnostic adaptive feature enhancement framework that boosts the performance of high-level vision tasks under varying illumination conditions. By integrating its modules with light-agnostic learning, TorchAdapt produces powerful semantic representations for both low-light and well-lit data. Extensive experiments on five representative visual perception tasks involving images and videos confirm its effectiveness.", + "bbox": [ + 511, + 763, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5652", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially funded by the European Union under Horizon Europe, grant number 101135724 (LUMINOUS).", + "bbox": [ + 89, + 114, + 483, + 146 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 161, + 187, + 176 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Muhamamd Zeshan Afzal, SK Aziz Ali, Didier Stricker, Peter Eisert, Anna Hilsmann, Daniel Perez-Marcos, Marco Bianchi, Sonia Crottaz-Herbette, Roberto De Ioris, Eleni Mangina, et al. Next generation xr systems-large language models meet augmented and virtual reality. IEEE computer graphics and applications, 2025. 1", + "[2] Tadahiro Azetsu and Noriaki Suetake. Hue-preserving image enhancement in cielab color space considering color gamut. Optical Review, 26:283-294, 2019. 5", + "[3] Ruzena Bajcsy, Yiannis Aloimonos, and John K Tsotsos. Revisiting active perception. Autonomous Robots, 42:177-196, 2018. 1, 2", + "[4] Peter J Bickel and Kjell A Doksum. Mathematical statistics: basic ideas and selected topics, volumes I-II package. Chapman and Hall/CRC, 2015. 1", + "[5] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Fredo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. In CVPR 2011, pages 97-104. IEEE, 2011. 2", + "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2", + "[7] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE international conference on computer vision, pages 2722-2730, 2015. 1, 2", + "[8] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3291-3300, 2018. 2", + "[9] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7", + "[10] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. International Journal of Computer Vision, 131(8):2198-2218, 2023. 2, 3, 5", + "[11] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 7", + "[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 93, + 186, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2", + "[13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1", + "[14] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.7", + "[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 2, 5", + "[16] Ziteng Cui, Guo-Jun Qi, Lin Gu, Shaodi You, Zenghui Zhang, and Tatsuya Harada. Multitask aet with orthogonal tangent regularity for dark object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2553-2562, 2021. 1, 2, 3, 5, 6, 7", + "[17] Ziteng Cui, Kunchang Li, Lin Gu, Shenghan Su, Peng Gao, Zhengkai Jiang, Yu Qiao, and Tatsuya Harada. You only need 90k parameters to adapt light: a light weight transformer for image enhancement and exposure correction. arXiv preprint arXiv:2205.14871, 2022. 1, 6", + "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 3, 5", + "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2, 5", + "[20] Zhipeng Du, Miaojing Shi, and Jiankang Deng. Boosting object detection with zero-shot day-night domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12666-12676, 2024. 1, 2, 3, 6, 7, 8", + "[21] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 2, 5", + "[22] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 3", + "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 3, 5" + ], + "bbox": [ + 517, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5653", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. CoRR, abs/2001.06826, 2020. 1, 2, 4, 6, 7", + "[25] Jiang Hai, Zhu Xuan, Ren Yang, Yutong Hao, Fengzhu Zou, Fang Lin, and Songchen Han. R2rnet: Low-light image enhancement via real-low to real-normal network. Journal of Visual Communication and Image Representation, 90:103712, 2023. 2", + "[26] Khurram Azeem Hashmi, Goutham Kallempudi, Didier Stricker, and Muhammad Zeshan Afzal. Featenhancer: Enhancing hierarchical features for object detection and beyond under low-light vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6725-6735, 2023. 1, 2, 3, 4, 6, 7", + "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 2, 5", + "[28] Mingbo Hong, Shen Cheng, Haibin Huang, Haoqiang Fan, and Shuaicheng Liu. You only look around: Learning illumination invariant feature for low-light object detection. arXiv preprint arXiv:2410.18398, 2024. 6", + "[29] Sergey Ioffe. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. 4", + "[30] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhancement without paired supervision. IEEE Transactions on Image Processing, 30: 2340-2349, 2021. 2", + "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1", + "[32] Attila Lengyel, Sourav Garg, Michael Milford, and Jan C. van Gemert. Zero-shot day-night domain adaptation with a physics prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4399-4409, 2021. 1, 2, 3, 6, 8", + "[33] Chongyi Li, Chunle Guo Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 4, 6, 8", + "[34] Chongyi Li, Chunle Guo, Shangchen Zhou, Qiming Ai, Ruicheng Feng, and Chen Change Loy. Flexicurve: Flexible piecewise curves estimation for photo retouching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1092-1101, 2023. 2", + "[35] Hebei Li, Jin Wang, Jiahui Yuan, Yue Li, Wenming Weng, Yansong Peng, Yueyi Zhang, Zhiwei Xiong, and Xiaoyan Sun. Event-assisted low-light video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3250-3259, 2024. 3", + "[36] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd:" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 7, 8", + "[37] Jinlong Li, Baolu Li, Zhengzhong Tu, Xinyu Liu, Qing Guo, Felix Juefei-Xu, Runsheng Xu, and Hongkai Yu. Light the night: A multi-condition diffusion framework for unpaired low-light enhancement in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15205-15215, 2024. 2", + "[38] Yunan Li, Yihao Zhang, Shoude Li, Long Tian, Dou Quan, Chaoneng Li, and Qiguang Miao. Watching it in dark: A target-aware representation learning framework for high-level vision tasks in low illumination. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 3, 6", + "[39] Jinxiu Liang, Jingwen Wang, Yuhui Quan, Tianyi Chen, Jiaying Liu, Haibin Ling, and Yong Xu. Recurrent exposure generation for low-light face detection. IEEE Transactions on Multimedia, 24:1609-1621, 2022. 3", + "[40] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312, 2014. 1, 2, 3, 5, 6", + "[41] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. CoRR, abs/1708.02002, 2017. 7", + "[42] Hewei Liu, Renwei Yang, Shuyuan Zhu, Xing Wen, and Bing Zeng. Luminance-guided chrominance image enhancement for hevc intra coding. In 2022 IEEE International Symposium on Circuits and Systems (ISCAS), pages 3180-3184. IEEE, 2022. 5", + "[43] Jiaying Liu, Dejia Xu, Wenhan Yang, Minhao Fan, and Haofeng Huang. Benchmarking low-light image enhancement and beyond. International Journal of Computer Vision, 129:1153-1184, 2021. 2", + "[44] Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10561-10570, 2021. 1, 2", + "[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 4", + "[46] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12009-12019, 2022. 4", + "[47] Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. CoRR, abs/1805.11227, 2018. 1, 2, 3, 5, 6", + "[48] Rundong Luo, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Similarity min-max: Zero-shot day-night domain adaptation." + ], + "bbox": [ + 516, + 92, + 908, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5654", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8104-8114, 2023. 1, 2, 3, 5, 6, 7, 8", + "[49] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 7", + "[50] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5637-5646, 2022. 1, 2, 4, 6, 7, 8", + "[51] Tengyu Ma, Long Ma, Xin Fan, Zhongxuan Luo, and Risheng Liu. PIA: parallel architecture with illumination accelerator for joint enhancement and detection in low-light. In MM '22: The 30th ACM International Conference on Multimedia, Lisboa, Portugal, October 10 - 14, 2022, pages 2070-2078. ACM, 2022. 1, 3", + "[52] Igor Morawski, Yu-An Chen, Yu-Sheng Lin, and Winston H. Hsu. NOD: taking a closer look at detection under extreme low-light conditions with night object detection dataset. CoRR, abs/2110.10364, 2021. 3, 5", + "[53] Lukáš Neumann, Michelle Karg, Shanshan Zhang, Christian Scharfenberger, Eric Piegert, Sarah Mistr, Olga Prokofyeva, Robert Thiel, Andrea Vedaldi, Andrew Zisserman, et al. Nightowls: A pedestrians at night dataset. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part I 14, pages 691–705. Springer, 2019. 3", + "[54] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 3", + "[55] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2", + "[56] Qingpao Qin, Kan Chang, Mengyuan Huang, and Guiqing Li. Denet: Detection-driven enhancement network for object detection under adverse weather conditions. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 2813-2829, 2022. 3", + "[57] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. CoRR, abs/1804.02767, 2018. 6, 8", + "[58] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 1, 2, 5", + "[59] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Guided curriculum model adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7374-7383, 2019. 1, 2", + "[60] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "IEEE/CVF International Conference on Computer Vision (ICCV), pages 10765-10775, 2021. 1, 2, 3, 5", + "[61] Yukihiro Sasagawa and Hajime Nagahara. Yolo in the dark - domain adaptation method for merging multiple models. In Computer Vision – ECCV 2020, pages 345–359, Cham, 2020. Springer International Publishing. 3", + "[62] Shangquan Sun, Wenqi Ren, Tao Wang, and Xiaochun Cao. Rethinking image restoration for object detection. Advances in Neural Information Processing Systems, 35:4461-4474, 2022. 3", + "[63] Maria Valera and Sergio A Velastin. Intelligent distributed surveillance systems: a review. IEE Proceedings-Vision, Image and Signal Processing, 152(2):192-204, 2005. 1, 2", + "[64] Wenjing Wang, Wenhan Yang, and Jiaying Liu. Hla-face: Joint high-low adaptation for low light face detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16195-16204, 2021. 3", + "[65] Wenjing Wang, Zhengbo Xu, Haofeng Huang, and Jiaying Liu. Self-aligned concave curve: Illumination enhancement for unsupervised adaptation. In Proceedings of the 30th ACM international conference on multimedia, pages 2617-2626, 2022. 2, 3, 5", + "[66] Wenjing Wang, Xinhao Wang, Wenhan Yang, and Jiaying Liu. Unsupervised face detection in the dark. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1250-1266, 2023. 3", + "[67] Wenjing Wang, Huan Yang, Jianlong Fu, and Jiaying Liu. Zero-reference low-light enhancement via physical quadruple priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26057-26066, 2024. 2", + "[68] Xinzhe Wang, Kang Ma, Qiankun Liu, Yunhao Zou, and Ying Fu. Multi-object tracking in the dark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 382-392, 2024. 3", + "[69] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. CoRR, abs/1808.04560, 2018. 7", + "[70] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2", + "[71] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wenhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5901-5910, 2022. 1, 6", + "[72] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15769-15778, 2021. 1, 2", + "[73] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17714-17724, 2022. 2" + ], + "bbox": [ + 516, + 90, + 908, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5655", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Xiaogang Xu, Ruixing Wang, and Jiangbo Lu. Low-light image enhancement via structure modeling and guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9893-9903, 2023. 2", + "[75] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Arid: A new dataset for recognizing action in the dark. In Deep Learning for Human Activity Recognition: Second International Workshop, DL-HAR 2020, Held in Conjunction with IJCAI-PRICAI 2020, Kyoto, Japan, January 8, 2021, Proceedings 2, pages 70–84. Springer, 2021. 3", + "[76] Xinwei Xue, Jia He, Long Ma, Yi Wang, Xin Fan, and Risheng Liu. Best of both worlds: See and understand clearly in the dark. In Proceedings of the 30th ACM International Conference on Multimedia, page 2154–2162, New York, NY, USA, 2022. Association for Computing Machinery. 6, 7", + "[77] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1", + "[78] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5525-5533, 2016. 1, 2, 3, 5, 7, 8", + "[79] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12918-12927, 2023. 2", + "[80] Wenhan Yang, Ye Yuan, Wenqi Ren, Jiaying Liu, Walter J. Scheirer, Zhangyang Wang, Taiheng Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, Yuqiang Zheng, Yanyun Qu, Yuhong Xie, Liang Chen, Zhonghao Li, Chen Hong, Hao Jiang, Siyuan Yang, Yan Liu, Xiaochao Qu, Pengfei Wan, Shuai Zheng, Minhui Zhong, Taiyi Su, Lingzhi He, Yandong Guo, Yao Zhao, Zhenfeng Zhu, Jinxiu Liang, Jingwen Wang, Tianyi Chen, Yuhui Quan, Yong Xu, Bo Liu, Xin Liu, Qi Sun, Tingyu Lin, Xiaochuan Li, Feng Lu, Lin Gu, Shengdi Zhou, Cong Cao, Shifeng Zhang, Cheng Chi, Chubing Zhuang, Zhen Lei, Stan Z. Li, Shizheng Wang, Ruizhe Liu, Dong Yi, Zheming Zuo, Jianning Chi, Huan Wang, Kai Wang, Yixiu Liu, Xingyu Gao, Zhenyu Chen, Chang Guo, Yongzhou Li, Huicai Zhong, Jing Huang, Heng Guo, Jianfei Yang, Wenjuan Liao, Jiangang Yang, Liguo Zhou, Mingyue Feng, and Likun Qin. Advancing image understanding in poor visibility environments: A collective benchmark study. IEEE Transactions on Image Processing, 29:5737-5752, 2020. 1, 2, 3, 5, 7, 8", + "[81] Bo Zhang, Yuchen Guo, Runzhao Yang, Zhihong Zhang, Jiayi Xie, Jinli Suo, and Qionghai Dai. Darkvision: A benchmark for low-light image/video perception, 2023. 1, 2, 3, 5", + "[82] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on Learning Representations, 2023. 1", + "[83] Liheng Zhang, Guo-Jun Qi, Liqiang Wang, and Jiebo Luo. Aet vs. aed: Unsupervised representation learning by auto" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "encoding transformations rather than data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2547-2555, 2019. 3", + "[84] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. CoRR, abs/1905.04161, 2019. 1, 2, 6", + "[85] Zhaoyang Zhang, Yitong Jiang, Jun Jiang, Xiaogang Wang, Ping Luo, and Jinwei Gu. Star: A structure-aware lightweight transformer for real-time image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4106-4115, 2021. 1, 8", + "[86] Zhao Zhang, Huan Zheng, Richang Hong, Mingliang Xu, Shuicheng Yan, and Meng Wang. Deep color consistent network for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1899-1908, 2022. 2", + "[87] Hegui Zhu, Kai Wang, Ziwei Zhang, Yuelin Liu, and Wuming Jiang. Low-light image enhancement network with decomposition and adaptive information fusion. Neural Computing and Applications, 34(10):7733-7748, 2022. 5" + ], + "bbox": [ + 516, + 92, + 906, + 373 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "5656", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/60f819c6-a0ae-4fb0-be50-050c7068618b_model.json b/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/60f819c6-a0ae-4fb0-be50-050c7068618b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0e2b4ad92bf8fb87deeb2bb89b037fda868578bc --- /dev/null +++ b/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/60f819c6-a0ae-4fb0-be50-050c7068618b_model.json @@ -0,0 +1,2831 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.162, + 0.131, + 0.838, + 0.154 + ], + "angle": 0, + "content": "TorchAdapt: Towards Light-Agnostic Real-Time Visual Perception" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.18, + 0.382, + 0.198 + ], + "angle": 0, + "content": "Khurram Azeem Hashmi\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.181, + 0.65, + 0.198 + ], + "angle": 0, + "content": "Karthik Palyakere Suresh1,2" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.199, + 0.612, + 0.215 + ], + "angle": 0, + "content": "Muhammad Zeshan Afzal1,2" + }, + { + "type": "text", + "bbox": [ + 0.298, + 0.216, + 0.354, + 0.232 + ], + "angle": 0, + "content": "\\(^{1}\\)DFKI" + }, + { + "type": "text", + "bbox": [ + 0.458, + 0.217, + 0.701, + 0.233 + ], + "angle": 0, + "content": "\\(^{2}\\)RPTU Kaiserslautern-Landau" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.238, + 0.62, + 0.251 + ], + "angle": 0, + "content": "khurram_azeem Hashmi@dfki.de" + }, + { + "type": "text", + "bbox": [ + 0.696, + 0.181, + 0.834, + 0.198 + ], + "angle": 0, + "content": "Didier Stricker1,2" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.486, + 0.665 + ], + "angle": 0, + "content": "Low-light conditions significantly degrade the performance of high-level vision tasks. Existing approaches either enhance low-light images without considering normal illumination scenarios, leading to poor generalization, or are tailored to specific tasks. We propose TorchAdapt, a real-time adaptive feature enhancement framework that generalizes robustly across varying illumination conditions without degrading performance in well-lit scenarios. TorchAdapt consists of two complementary modules: the Torch module enhances semantic features beneficial for downstream tasks, while the Adapt module dynamically modulates these enhancements based on input content. Leveraging a novel light-agnostic learning strategy, TorchAdapt aligns feature representations of enhanced and well-lit images to produce powerful illumination-invariant features. Extensive experiments on multiple high-level vision tasks, including object detection, face detection, instance segmentation, semantic segmentation, and video object detection, demonstrate that TorchAdapt consistently outperforms state-of-the-art low-light enhancement and task-specific methods in both low-light and light-agnostic settings. TorchAdapt thus provides a unified, flexible solution for robust visual perception across diverse lighting conditions." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.695, + 0.222, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.84 + ], + "angle": 0, + "content": "Recent advances in visual perception tasks [13, 31, 77, 82] are highly effective at extracting features for complex tasks when provided with high-quality inputs. However, replicating this performance is challenging when data is scarce, noisy, or expensive to obtain. A common issue is varying illumination, particularly low-light conditions, which hinders the robustness and reliability of safety-critical visual perception systems [1, 3, 7, 63]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.486, + 0.902 + ], + "angle": 0, + "content": "A common strategy to improve low-light performance involves pre-processing images with Low-Light Image Enhancement (LLIE) methods [17, 24, 33, 44, 50, 51, 71, 84, 85] before feeding them into high-level vision mod" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.285, + 0.625, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.286, + 0.737, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.286, + 0.904, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.403, + 0.763, + 0.414 + ], + "angle": 0, + "content": "(a) Low-Light Vision" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.414, + 0.622, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.415, + 0.734, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.744, + 0.415, + 0.904, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.532, + 0.768, + 0.543 + ], + "angle": 0, + "content": "(b) Light-Agnostic Vision" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.678 + ], + "angle": 0, + "content": "Figure 1. Prior downstream methods in low-light vision assume that all samples belong to low-light conditions as shown in (a), leading to underperformance in varying illumination environments, depicted in (b). TorchAdapt alleviates this bottleneck via learning powerful illumination-invariant features and achieves consistent gains across low-light and light-agnostic settings on Object Detection (OD), Face Detection (FD), Instance Segmentation (IS), Semantic Segmentation (SS), and Video Object Detection (VOD). Best viewed on the screen." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.909, + 0.837 + ], + "angle": 0, + "content": "els. However, this simple integration often leads to weaker gains [16, 26], as these arts conform to human visual perception (optimized on pixel-level objectives like L1 or MSE losses [4]) instead of machine visual perception. Moreover, applying these enhancement methods indiscriminately can degrade the performance on images/videos captured under standard lighting conditions, as alterations intended for low-light scenarios [47, 60, 80, 81] may be unnecessary or even detrimental in well-lit environments [15, 40, 58, 78]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To address low-light vision tasks, recent methods employ transfer learning [16, 59, 72] and domain adaptation techniques [20, 32, 48]. Despite improvements, these efforts are carefully designed for a single high-level vision task, such" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5645" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.304 + ], + "angle": 0, + "content": "as object detection [16] or semantic segmentation [59, 72], and require cumbersome multi-stage training to adapt to each dense task [20, 32, 48]. Moreover, they often assume consistently low-light data, leading to performance degradation on well-lit images and videos, sometimes worse than baselines (see Table 2). In practical applications with unpredictable lighting—such as autonomous driving [7, 37] at different times of the day, surveillance systems [63] monitoring both indoor and outdoor environments or robots [3] operating between well-lit and low-lit areas—these limitations become critical. This highlights the need for high-level vision models to be light-agnostic—capable of maintaining high performance across varying lighting conditions, as visualized in Fig. 1(b)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.486, + 0.457 + ], + "angle": 0, + "content": "In this work, we propose TorchAdapt, a real-time, light-agnostic adaptive feature enhancement framework designed to improve the performance of any high-level vision task under varying illumination. TorchAdapt comprises two primary modules: Torch and Adapt, illustrated in Fig. 2. The Torch module is a lightweight feature enhancement network that enriches semantic representations favorable for downstream vision tasks. The Adapt module learns to modulate this feature enhancement dynamically based on input data, amplifying or reducing the effect of the Torch module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.486, + 0.732 + ], + "angle": 0, + "content": "TorchAdapt builds on the consensus [26, 38] that low-to-normal light adaptation lacks a physical model, leading to unexpected artifacts during optimization when transforming images in pixel space. Therefore, aligning representations in the feature space is a more practical approach for high-level vision tasks [48, 65]. We optimize TorchAdapt using a non-contrastive self-supervised learning objective [12, 23] to align representations from low-light and well-lit images. However, unlike prior works operating in a similar direction [16, 48, 65], TorchAdapt also learns not to adversely affect representations of well-lit images. As illustrated in Fig. 3, we take a well-lit image \\( V_{1} \\) and its synthetically generated low-light counterpart \\( V_{2} \\), pass them through TorchAdapt to obtain enhanced images \\( \\hat{V}_{1} \\) and \\( \\hat{V}_{2} \\), and feed them along with \\( V_{1} \\) into a pre-trained frozen visual encoder [19, 27, 55] to generate representations. Since the encoder is frozen, only TorchAdapt is optimized to help the encoder produce well-lit representations (like \\( V_{1} \\)) for \\( \\hat{V}_{1} \\) and \\( \\hat{V}_{2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Thanks to the conceptual simplicity, TorchAdapt enjoys two benefits. ① It can be trained with a simple self-supervised loss on a general-purpose dataset like ImageNet [18], circumventing large-scale real-world low-light datasets. ② The training of TorchAdapt is independent of the high-level vision task and needs to be done only once to align with the backbone networks [6, 19, 27]. Later, TorchAdapt can be plugged into any downstream tasks in a light-agnostic manner. We demonstrate the effectiveness of TorchAdapt on five representative visual perception tasks, including object detection [21, 47], face detection [78, 80], semantic" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.212 + ], + "angle": 0, + "content": "segmentation [15, 60], instance segmentation [10, 40], and video object detection [58, 81] under both low-light and light-agnostic settings. For all tasks, TorchAdapt brings consistent and significant gains over baselines under both low-light and light-agnostic settings, outperforming previous state-of-the-art task-specific and LLIE methods (see Fig. 1). Moreover, TorchAdapt accomplishes all this with negligible computational overhead (see Table 6)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.911, + 0.518 + ], + "angle": 0, + "content": "Contributions. Method: We propose TorchAdapt, a real-time, light-agnostic adaptive feature enhancement framework that improves the performance of any high-level vision task under varying illumination without compromising accuracy in well-lit conditions. Simplicity: Thanks to its architectural innovations and self-supervised objective, TorchAdapt significantly simplifies the training strategy for any high-level vision task pipeline, circumventing corresponding paired images/videos, as shown in Fig 3. Flexibility: TorchAdapt is independent of vision encoders and downstream task architectures, demonstrating model-agnostic generality (see § 5). Any zero-reference LLIE method [24, 33, 50, 67] can serve as the Torch module in our framework to achieve light-agnostic performance (see Fig. 5). Results: We conduct extensive experiments on five visual perception tasks involving images and videos under both low-light and light-agnostic settings using curated datasets. TorchAdapt outperforms state-of-the-art task-specific and LLIE methods across all settings when integrated with the same baselines (see Fig. 1)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.534, + 0.659, + 0.55 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.78, + 0.577 + ], + "angle": 0, + "content": "2.1. Enhancing Low-Light Images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.91, + 0.903 + ], + "angle": 0, + "content": "Low-light Image Enhancement (LLIE) aims to improve the visual quality of images captured under insufficient lighting conditions. Modern learning-based approaches have become prevalent, thanks to the emergence of low-light datasets [5, 8, 25, 43, 70]. Supervised methods [70, 73, 74, 84, 86] leverage these datasets to design effective networks trained to predict normal-light images from low-light ones. Unsupervised LLIE approaches [30, 34, 79] employ adversarial learning from unpaired supervision to relax the requirement of paired normal-low light data. However, all of these works still rely on specific training data, which limits their generalizability to unseen scenarios. Another line of work known as zero-reference methods [24, 33, 44, 50, 67] get rid of both paired and unpaired data to enhance low-light images. These arts design enhancement networks that optimize on a set of non-reference loss functions. Since these works are mainly designed to conform to human visual perception, they bring sub-optimal gains for machine visual perception tasks. Furthermore, they are not light-agnostic (capable of handling both normal and low-light images). This work bridges this gap by complementing these zero-reference methods," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "5646" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.089, + 0.833, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.231, + 0.908, + 0.304 + ], + "angle": 0, + "content": "Figure 2. TorchAdapt architectural overview. The Torch module enhances feature representation by applying \\(3 \\times 3\\) convolutions and element-wise operations (Eq. 1). The Adapt module performs adaptive modulation, creating a content-aware response based on the input \\(I\\) (Eq. 2). Both modules are fused via Adaptive Fusion, integrating feature enhancements from Torch and contextual adaptations from Adapt. This adaptive fusion produces illumination-invariant features, improving robustness across varying lighting conditions. Adaptive Fusion is explained in Eq. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.315, + 0.486, + 0.348 + ], + "angle": 0, + "content": "where they can be employed as a Torch in our TorchAdapt framework to improve light-agnostic visual perception." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.36, + 0.35, + 0.377 + ], + "angle": 0, + "content": "2.2. Low-Light Visual Perception" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.483, + 0.414 + ], + "angle": 0, + "content": "This direction of work considers machine perception as the criteria for success while enhancing low-light images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.688 + ], + "angle": 0, + "content": "Low-light Object Detection (LLOD). Object detection, as a fundamental problem in computer vision, has also been well-explored in low-light vision [16, 20, 39, 51, 61, 62, 64-66]. Based on the dataset, objects vary and can be generic [47, 52] or specific like pedestrians [53], and human faces [80]. One category of LLOD methods [51, 56, 62, 65] jointly learns the enhancement and detection task in an end-to-end fashion to improve the overall detection performance. Other approaches involve learning a low-light detector through fusing multiple models [61] or resorting to domain adaptation frameworks [16, 20, 64], minimizing the distribution variance between normal and low-light images. Most of these works are specifically designed to tackle the dark object detection task. Furthermore, when these methods are optimized for LLOD, they underperform on normal-light images. In contrast, TorchAdapt is fully task-agnostic for any high-level vision task, and it does not worsen the performance of the detector on normal-light data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.487, + 0.901 + ], + "angle": 0, + "content": "Low-light Tasks-Agnostic Methods. Recently, various low-light task-agnostic methods [20, 26, 32, 48, 65] have emerged in the literature. Unlike prior approaches, they are more flexible and can be integrated to improve any downstream task under low-light vision. Our work follows the same spirit and offers similar flexibility. FeatEnHancer [26], as one of the representative works, learns to enhance the hierarchical features suitable for downstream tasks. Other seminal attempts [20, 32, 48] treat low-light perception tasks as domain adaption (DA) learning. They exploit well-lit source domain data [40, 78] and learn to adapt to low-light target domain data [47, 80]. Although flexible, the DA methods rely on multi-stage training frameworks for each task [48]. In contrast, our TorchAdapt is pre-trained with the frozen back" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.908, + 0.39 + ], + "angle": 0, + "content": "bone on the domain-independent data like ImageNet [18]. Moreover, it is important to recognize the efforts of recent real low-light datasets [10, 35, 60, 68, 75, 81] that enable the increasing research on several downstream tasks in low-light vision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.407, + 0.907, + 0.439 + ], + "angle": 0, + "content": "2.3. Learning Equivariant Representations in LowLight Vision with Self-Supervised Objectives" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Annotating low-light vision datasets is a much more difficult task than annotating well-lit datasets. Hence, it is a common practice to apply self-supervised learning to learn illumination-invariant features [16, 65] or aligning well-lit and low-light data distributions [38, 48, 66]. MAET [16] adopts the autoencoding transformation [83] to decode illumination-degrading parameters and detection predictions. SACC [65] merges the pretext tasks of rotation prediction [22] and jigsaw puzzling [54] to learn illumination enhancement from low-light images. Most related to our work is SimMinMax [48], which adopts the BYOL [23] framework to learn model-level adaptation by maximizing the similarity of features between images at night and daytime. However, since SimMinMax [48] learns to adapt from daytime to nighttime images, it needs to be trained for each high-level vision task, requiring the corresponding daytime labeled dataset. In contrast, TorchAdapt leverages self-supervised objectives to align backbone features during the pretext task. Furthermore, it can be trained on any general-purpose dataset. This significantly simplifies the training procedure of TorchAdapt, as compared to SimMinMax, on any high-level vision task. TorchAdapt is also unique to prior works in its ability to handle light-agnostic data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.828, + 0.642, + 0.846 + ], + "angle": 0, + "content": "3. TorchAdapt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "This section formally introduces TorchAdapt, explaining the proposed Torch in § 3.1 and Adapt in § 3.2 modules, followed by the light-agnostic learning in § 3.3." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5647" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.096, + 0.468, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.27, + 0.485, + 0.382 + ], + "angle": 0, + "content": "Figure 3. Light-Agnostic learning pipeline. \\( D(.) \\) is a darkening function applied to the well-lit image \\( V_{1} \\) to synthesize the low-light view \\( V_{2} \\). Both \\( V_{1} \\) and \\( V_{2} \\) are processed by TorchAdapt and then fed into a frozen pre-trained vision encoder \\( f \\). A cosine similarity loss optimizes TorchAdapt to produce well-lit representations for both \\( V_{1} \\) and \\( V_{2} \\), enabling it to learn illumination-invariant features. The TorchAdapt(.) function integrates Eq. 1 and Eq. 4. Refer to § 3.3 for a complete explanation of each symbol." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.392, + 0.174, + 0.406 + ], + "angle": 0, + "content": "3.1.Torch" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.484, + 0.566 + ], + "angle": 0, + "content": "The Torch module in our TorchAdapt framework serves as an enhancement network as in Low-light Image Enhancement (LLIE) approaches [24, 33, 50]. However, unlike these methods—which focus on improving visual quality for human perception—our Torch module aims to boost semantic representations favorable for downstream vision tasks under varying lighting conditions. Given an input image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\), the Torch module computes an enhancement map \\(\\mathcal{H}_{\\theta}(\\mathbf{I})\\), parameterized by network weights \\(\\theta\\). The transformed image \\(\\mathbf{T}\\) is obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.581, + 0.484, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} = \\mathbf {I} \\odot (1 + \\mathcal {H} _ {\\theta} (\\mathbf {I})) \\text {, w h e r e} \\mathcal {H} _ {\\theta} (\\mathbf {I}) = \\tanh (\\mathcal {F} (\\mathbf {I}; \\theta)) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise multiplication. The \\(\\tanh(\\cdot)\\) activation ensures that the adjustment values are in the range \\([-1, 1]\\), allowing the model to both amplify (positive values) and attenuate (negative values) features. This formulation enables the model to adapt to the image's content effectively. Torch network. The Torch network \\(\\mathcal{F}(\\mathbf{I};\\theta)\\) is a lightweight convolutional network composed of an initial feature extraction layer followed by a residual block, as illustrated in Fig. 2. The residual learning enables the Torch module to focus on learning the necessary adjustments to improve the input image rather than reconstructing it entirely. Enhancing semantically important features while preserving original content makes the Torch module crucial in maintaining and improving performance in downstream vision tasks. Notably, thanks to the formulation explained in Eq 1, our Torch module only has 11,075 trainable parameters and requires 0.712 GFLOPs for \\(\\mathbf{I}\\) of size \\(256 \\times 256 \\times 3\\), making it suitable for real-time high-level vision tasks under varying lighting conditions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.6, + 0.107 + ], + "angle": 0, + "content": "3.2. Adapt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.279 + ], + "angle": 0, + "content": "While the Torch module enhances the input image \\(\\mathbf{I}\\) to produce \\(\\mathbf{T}\\), applying this enhancement uniformly may not be optimal. Different regions within an image require varying degrees of enhancement based on content (low/well-lit images) and context (downstream task). To address this, we introduce the Adapt module, which learns a content-adaptive modulation matrix in the feature space. This allows the model to selectively amplify or suppress the enhancements produced by the Torch module, improving their effectiveness for downstream vision tasks in varying illumination scenarios, effectively making it light-agnostic." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Specifically, the Adapt module computes an Adaptive Matrix \\(\\mathbf{M} \\in \\mathbb{R}^{H \\times W}\\), where \\(H\\) and \\(W\\) are the spatial dimensions of \\(\\mathbf{I}\\). The purpose of the Adaptive Matrix \\(\\mathbf{M}\\) is to modulate the feature enhancements from Torch \\(\\mathbf{T}\\) before they are applied to \\(\\mathbf{I}\\). \\(\\mathbf{M}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.362, + 0.907, + 0.379 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\sigma \\left(\\mathcal {G} _ {\\phi} (\\mathbf {I})\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.906, + 0.475 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}_{\\phi}\\) is a lightweight convolutional network parameterized by weights \\(\\phi\\), and \\(\\sigma(\\cdot)\\) denotes the sigmoid activation function ensuring that the values of \\(\\mathbf{M}\\) are in the range \\([0,1]\\). We feed the input image \\(\\mathbf{I}\\) into the Adapt module to compute \\(\\mathbf{M}\\), ensuring that the image content directly influences the modulation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.909, + 0.748 + ], + "angle": 0, + "content": "Adapt network. As shown in Fig. 2, the convolutional network \\(\\mathcal{G}_{\\phi}\\) in the Adapt module is designed to extract hierarchical features, which are proven to be beneficial for high-level vision tasks such as object detection and semantic segmentation [26, 45, 46]. To capture multi-scale contextual information, the network consists of three \\(3\\times 3\\) convolutional layers with increasing channel dimensions \\(C_1\\), \\(C_2\\), and \\(C_3\\), respectively. Each layer is followed by a batch normalization [29] and a ReLU activation function. This architecture allows the network to learn a hierarchy of features from low-level edges and textures to high-level semantic information. Finally, an additional \\(3\\times 3\\) convolutional layer maps the features from \\(C_3\\) to a single channel and applies a sigmoid activation function to generate the Adaptive Matrix \\(\\mathbf{M}\\in [0,1]^{H\\times W}\\). This design balances expressiveness and computational efficiency, enabling the Adapt module to adjust enhancements based on both local and global features, which is crucial for content-adaptive modulation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.748, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Illumination-aware scaling. We introduce a scale factor \\(\\gamma\\) to ensure that the enhancement adapts appropriately to different illumination levels. This scale factor modulates the M globally based on the image's overall brightness, allowing the model to apply stronger enhancements to darker images and milder adjustments to well-lit images, maintaining visual consistency and preventing artifacts. We compute \\(\\gamma\\) as a function of the average luminance \\(L(\\mathbf{I})\\) of the input image:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.876, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\gamma = \\frac {\\kappa}{1 + \\exp (k (L (\\mathbf {I}) - L _ {0}))}, \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5648" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.182 + ], + "angle": 0, + "content": "where \\(\\kappa\\) is the maximum scale factor (empirically set to 3 unless stated otherwise). \\(k\\) controls the slope of the sigmoid function, and \\(L_{0}\\) is the luminance value at the sigmoid's midpoint. The average luminance \\(L(\\mathbf{I})\\) is computed following the standard luminance calculation in color science [2, 42, 87]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.182, + 0.486, + 0.228 + ], + "angle": 0, + "content": "Adaptive Fusion. The final enhanced image \\(\\mathbf{E}\\) is then obtained by modulating the enhancement from the Torch module with the illumination-aware scaled Adaptive Matrix:" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.237, + 0.484, + 0.254 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} = \\mathbf {I} + ((\\gamma \\cdot \\mathbf {M}) \\odot \\mathbf {T}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.485, + 0.339 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise multiplication, and \\(\\mathbf{T}\\) is obtained from Eq. (1). This formulation in Eq. 4 integrates both content-based modulation (through \\(\\mathbf{M}\\)) and illumination-based scaling (through \\(\\gamma\\)), enabling the model to adaptively enhance images under varying lighting conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.346, + 0.318, + 0.362 + ], + "angle": 0, + "content": "3.3. Light-Agnostic Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.485, + 0.579 + ], + "angle": 0, + "content": "Achieving robust performance across varying lighting conditions requires our model to learn representations invariant to illumination changes. Collecting labeled data under all possible lighting conditions is impractical. Hence, we adopt a self-supervised learning (SSL) strategy inspired by BYOL [23] to learn illumination-invariant features. Unlike prior works that design cumbersome pipelines to incorporate SSL into each high-level vision task [16, 48, 65], we leverage SSL to pre-train our TorchAdapt module with a pretext task that aligns backbone representations [19, 27] regardless of illumination. This approach eliminates the need for task-specific labeled data or low-light data, allowing training on general-purpose datasets like ImageNet [18]. Figure 3 illustrates the pre-training pipeline of TorchAdapt." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.486, + 0.806 + ], + "angle": 0, + "content": "Objective Function. Given a well-lit image \\( V_{1} \\), we generate its low-light counterpart \\( V_{2} = D(V_{1}) \\) using a darkening function \\( D(\\cdot) \\), which can be a gamma transformation, an illumination-degradation pipeline [16], or a darkening module from [48]. Let \\( f \\) denote the frozen pre-trained backbone network (e.g., ResNet [27], ViT [19]). We pass \\( V_{1} \\) and \\( V_{2} \\) through the TorchAdapt module to obtain enhanced images \\( \\hat{V}_{1} \\) and \\( \\hat{V}_{2} \\), respectively. The enhanced images are processed by the backbone \\( f \\) and a projector network \\( g \\) to obtain feature representations: \\( z_{1}^{\\prime} = g(f(\\hat{V}_{1})) \\) and \\( z_{2} = g(f(\\hat{V}_{2})) \\). The target representation is obtained by passing the original well-lit image \\( V_{1} \\) through \\( f^{\\prime} \\) and \\( g^{\\prime} \\): \\( y_{1} = g^{\\prime}(f^{\\prime}(V_{1})) \\). Our training objective minimizes the cosine similarity loss \\( \\mathcal{D}_{\\mathrm{cos}} \\) between the enhanced image representations and the well-lit target representation:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.813, + 0.484, + 0.855 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left(1 - \\frac {z _ {1} ^ {\\prime}}{\\| z _ {1} ^ {\\prime} \\| _ {2} \\| y _ {1} \\| _ {2}}\\right) + \\left(1 - \\frac {z _ {2} ^ {\\top} y _ {1}}{\\| z _ {2} \\| _ {2} \\| y _ {1} \\| _ {2}}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.902 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\| _2\\) denotes the \\(\\ell_2\\) norm of a vector. Since we employ identical, frozen backbone networks \\(f\\), the loss function guides TorchAdapt to enhance the low-light image \\(\\hat{V}_2\\)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.903, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.187, + 0.908, + 0.286 + ], + "angle": 0, + "content": "Figure 4. TorchAdapt's downstream integration pipeline. TorchAdapt works as a plug-and-play, backbone-agnostic feature enhancement module, seamlessly integrating into various downstream vision tasks. Given any image or video input across diverse lighting conditions, TorchAdapt produces illumination-invariant representations, significantly boosting performance without requiring task-specific modifications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.907, + 0.397 + ], + "angle": 0, + "content": "so its representation aligns with that of the well-lit image \\( V_{1} \\). Simultaneously, it discourages over-enhancement of the already well-lit image \\( \\hat{V}_{1} \\), as its representation should remain consistent with \\( V_{1} \\). This training strategy enables TorchAdapt to effectively enhance low-light images while preserving well-lit ones, promoting illumination invariance in the learned representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.909, + 0.579 + ], + "angle": 0, + "content": "Training Details. The training of TorchAdapt is fast and efficient due to its lightweight architecture and the absence of gradient propagation through the frozen backbone network. We train TorchAdapt for 5 epochs on the ImageNet [18] dataset using a single GPU, with a learning rate of 0.0001 and a batch size of 256. Since the backbone networks are pre-trained and we aim for consistent views, we apply only color jitter and random horizontal flip as data augmentations. Pre-training TorchAdapt with a ResNet-50 backbone takes \\(\\sim 6\\) hours on a single A100 GPU. Once trained, TorchAdapt can be seamlessly integrated into any downstream vision task, as illustrated in Fig. 4." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.589, + 0.907, + 0.719 + ], + "angle": 0, + "content": "
DatasetTask#Cls#Train#Val
Low-light (LL)
ExDark [47]Object detection1258911472
DARK FACE [80]Face detection15400600
ACDC [60]Semantic segmentation19400106
LIS [10]Instance segmentation81561669
DarkVision [81]Video object detection426*6*
Light-Agnostic (LA)
COCO [21] ∩ ExDark [47]Object detection12923575156
WIDER FACE [78] ∩ DARK FACE [80]Face detection1182803822
CityScapes [15] ∩ ACDC [60]Semantic segmentation193375606
COCO [40] ∩ LIS [10]Instance segmentation8443212550
ImageNet VID [58] ∩ DarkVision [81]Video object detection2401*51*
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.728, + 0.907, + 0.756 + ], + "angle": 0, + "content": "Table 1. Datasets Statistics.*: video samples. Light-Agnostic datasets are obtained through Eq. 6" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.77, + 0.645, + 0.788 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We conduct extensive experiments on several representative visual tasks to demonstrate the superiority of TorchAdapt in both low-light and light-agnostic settings. These tasks include object detection [40, 47, 52], face detection [78, 80], instance segmentation [10, 40], semantic segmentation [15, 60], and video object detection [58, 81]. This section first presents the experimental settings, comparing TorchAdapt" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5649" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.088, + 0.907, + 0.306 + ], + "angle": 0, + "content": "
Tasks → Light Settings → MethodsObject Det.Face Det.Instance Seg.Semantic Seg.Video Object Det.
LLLALLLALLLALLLALLLA
mAP50mAPmAP50mAPAP50APAP50APmAPSegAPmAPSegAPmIoUmIoUmAPmAP
Baseline78.844.262.433.147.319.952.628.054.046.041.533.949.675.532.851.3
Enhancement
Zero-DCE [24]76.242.757.629.547.420.147.622.951.543.539.232.142.559.47.8328.3
SCI [84]75.542.257.929.745.119.345.421.752.344.235.128.342.157.46.7125.7
URetinexNet [71]74.642.057.329.244.718.346.622.359.543.137.429.741.258.86.6825.1
IAT [17]74.141.052.430.542.115.937.117.449.742.336.430.141.756.26.2424.5
Task-specific
Zero-DCE [24]‡79.144.462.933.250.421.150.123.559.152.744.435.744.761.034.151.5
SCI [50]‡79.244.263.033.451.622.254.729.958.451.043.734.943.559.932.950.7
MAET [16]‡79.244.357.327.344.318.747.923.1--------
Xue et al. [76]----46.117.744.221.0----42.169.7--
FeatEnHancer [26]‡79.244.862.933.347.219.951.624.559.652.444.936.150.176.134.652.7
TorchAdapt80.145.464.034.153.723.159.032.561.153.645.236.650.876.339.159.1
vs. prev. SoTA+0.9+0.6+1.0+0.7+2.1+0.9+4.3+2.6+1.5+0.9+0.3+0.5+0.7+0.2+4.5+6.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.317, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison across five tasks under both Low-Light (LL) and Light-Agnostic (LA) settings. ‡: all these methods are reproduced by us with their official codebase and trained end-to-end for each task. The rest of the results are obtained from their official checkpoints. - indicates that the method is not applicable. TorchAdapt outperforms prior SOTA across all five tasks in LL and LA settings." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.372, + 0.482, + 0.463 + ], + "angle": 0, + "content": "with baselines, existing low-light image enhancement (LLIE) methods, and state-of-the-art task-specific approaches. We then examine critical design choices for TorchAdapt through ablation studies. Finally, we validate the generalizability of TorchAdapt by comparing it with recent state-of-the-art domain adaptation methods [20, 32, 48]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.493 + ], + "angle": 0, + "content": "Experimental Settings. We perform experiments under two settings for each task:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.495, + 0.484, + 0.555 + ], + "angle": 0, + "content": "1)- Low-Light (LL) Setting: Following prior works [16, 26, 76], we utilize publicly available low-light benchmarks specific to each task. These benchmarks consist exclusively of images captured under low-light conditions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.556, + 0.483, + 0.706 + ], + "angle": 0, + "content": "2)- Light-Agnostic (LA) Setting: To introduce a more challenging and realistic evaluation, we curate datasets that encompass both well-lit and low-light images. We achieve this by selecting classes common to both well-lit and low-light datasets. Mathematically, for a given task, let \\(\\mathcal{C}_{\\mathrm{WL}}\\) and \\(\\mathcal{C}_{\\mathrm{LL}}\\) denote the sets of classes in the well-lit and low-light datasets, respectively. We define the set of common classes as \\(\\mathcal{C}_{\\mathrm{com}} = \\mathcal{C}_{\\mathrm{WL}} \\cap \\mathcal{C}_{\\mathrm{LL}}\\). The LA dataset is then formed by combining all images from both datasets that belong to classes in \\(\\mathcal{C}_{\\mathrm{com}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.719, + 0.483, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\mathrm {L A}} = \\left\\{\\left(x, y\\right) \\mid (x, y) \\in \\mathcal {D} _ {\\mathrm {W L}} \\cup \\mathcal {D} _ {\\mathrm {L L}}, y \\in \\mathcal {C} _ {\\mathrm {c o m}} \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Where \\(\\mathcal{D}_{\\mathrm{WL}}\\) and \\(\\mathcal{D}_{\\mathrm{LL}}\\) are the well-lit and low-light datasets, respectively, and \\((x,y)\\) represents an image and its label. We summarize the key statistics of the employed benchmarks for both settings in Table 1. We provide complete implementation details for each experiment in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Comparisons with State-of-the-Arts. We directly compare our approach with several LLIE methods, including Zero-DCE [24], IAT [17], SCI [50], and URetinexNet [71]. Following common practice [16, 26, 76], we use their released checkpoints to enhance all images before feeding them into" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.908, + 0.614 + ], + "angle": 0, + "content": "the detector for performance evaluation. Notably, due to their zero-reference nature requiring no paired images, some of these LLIE methods [24, 50] can be trained end-to-end on downstream vision tasks. Therefore, besides evaluating their conventional enhancement results, we integrate them into each task and report their performance, similar to low-light task-specific works like MAET [16] and FeatEnHancer [26] (indicated with \\(\\ddagger\\) in Table 2). Our findings reveal that end-to-end training of such LLIE works \\([24, 50]^1\\) significantly improves their performance on downstream vision tasks in both LL and LA settings. Concurrent to our work, we notice that two new methods [28, 38] appear in the literature on improving low-light object detection. However, since their code and checkpoints were not released by the time of the submission, we could not draw direct comparisons with them. Next, we discuss experiments on each task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.627, + 0.681, + 0.643 + ], + "angle": 0, + "content": "4.1. Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.909, + 0.862 + ], + "angle": 0, + "content": "Settings. Following common practice [16] in LL object detection, we adopt YOLOv3 [57] as the baseline detector. For the LL setting, COCO [40] pre-trained weights are utilized to fine-tune YOLOv3 on the ExDark dataset [47]. We train the model from scratch in the LA setting, as COCO images are also included in the training data (see Table. 1). Results. As shown in Table 2, TorchAdapt achieves a superior \\(\\mathrm{mAP}_{50}\\) of 80.1 in LL and 64.1 in LA, compared to the baseline's 78.8 and 62.4. End-to-end training boosts the performance of enhancement methods like Zero-DCE [33] and SCI [50]; however, TorchAdapt still surpasses them, and the low-light object detection method FeatEnHancer [26] by \\(+0.9\\) and \\(+1.1\\) in LL and LA settings, respectively. These results affirm the efficacy of TorchAdapt for object detec" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.908, + 0.901 + ], + "angle": 0, + "content": "1We select these works due to their efficiency and reproducibility. Experiments with more methods are in Appendix C." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5650" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "tion across varying illumination conditions. We provide a qualitative comparison in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.242, + 0.147 + ], + "angle": 0, + "content": "4.2. Face Detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.154, + 0.485, + 0.259 + ], + "angle": 0, + "content": "Settings. For face detection, we choose a different detector and adopt RetinaNet [41] to report results on the DARK FACE dataset [69, 80] in the LL setting and a combination of the DARK FACE and WIDER FACE [78] datasets in the LA setting. Following prior works [26], we resize images to a resolution of \\(1500 \\times 1000\\) pixels and follow the \\(1 \\times\\) schedule in MMDetection [9]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.486, + 0.426 + ], + "angle": 0, + "content": "Results. Results in Table 2 reveal that TorchAdapt demonstrates superior performance on face detection under both LL and LA settings, outperforming the previous best method \\((\\mathrm{SCI} \\ddagger [50])\\) by \\(+2.1\\) and \\(+4.3\\) in mAP_50, respectively. The significant gains, especially in the challenging LA setting, highlight TorchAdapt's capacity to handle diverse illumination scenarios by adaptively modulating features based on content and illumination, thereby improving the robustness and accuracy of the baseline. We also compare TorchAdapt with recent domain adaptation methods [20, 48] using DSFD [36] in Sec. 4.7." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.435, + 0.303, + 0.451 + ], + "angle": 0, + "content": "4.3. Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.485, + 0.547 + ], + "angle": 0, + "content": "Settings. For the instance segmentation task, we adopt RTMDet-Ins-tiny \\([49]^2\\) as our baseline model and resize images to \\(640\\times 640\\) pixels for training. Following Section 4.1, we use the COCO pre-trained weights of RTMDet-Ins-tiny for the LL setting and train the model from scratch for the LA setting (see Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.548, + 0.486, + 0.684 + ], + "angle": 0, + "content": "Results. As summarized in Table 2), TorchAdapt achieves SOTA SegAP performance in both low-light (LL) and light-agnostic (LA) settings. Under LL conditions, TorchAdapt attains a SegAP of 53.6, surpassing the previous best (FeatEnHancer†) by +1.2. In the LA setting, it achieves a SegAP of 36.6, slightly improving overFeatEnHancer† by +0.5. These results demonstrate TorchAdapt's effectiveness in enhancing instance segmentation across varying illumination by producing illumination-invariant features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.693, + 0.308, + 0.709 + ], + "angle": 0, + "content": "4.4. Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.486, + 0.867 + ], + "angle": 0, + "content": "Settings. Consistent with prior works [26, 76], we adopt DeepLabV3+[11] as our baseline model and follow [26, 76] for the LL setting. For the LA evaluation (Table 1), we train DeepLabV3+ from scratch using MMSegmentation [14]. Results. As shown in Table 2, TorchAdapt achieves an mIoU of 50.8 in the LL setting, surpassing the previous best (FeatEnHancer†) by +0.7, and an mIoU of 76.3 in the LA setting, improving overFeatEnHancer† by +0.2. These results highlight TorchAdapt's effectiveness for semantic segmentation under varying illumination conditions." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.09, + 0.906, + 0.168 + ], + "angle": 0, + "content": "
#Torch(§ 3.1)Adapt(§ 3.2)Scaling(§ 3.2)Object Det.Instance Seg.
LLLALLLA
1×××78.862.446.033.9
2✓××79.162.951.134.8
3✓✓×79.963.553.336.2
4✓×✓79.562.751.434.1
5✓✓✓80.164.053.636.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.179, + 0.907, + 0.207 + ], + "angle": 0, + "content": "Table 3. Contribution of each component. Highlighted settings are set as default." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.211, + 0.697, + 0.262 + ], + "angle": 0, + "content": "
Scale factor (κ)ODSS
LLLALLLA
279.963.753.436.4
380.164.053.636.6
479.963.353.236.2
579.561.153.334.5
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.532, + 0.264, + 0.688, + 0.277 + ], + "angle": 0, + "content": "(a) Scale factor \\((\\kappa)\\) in Eq. 3." + }, + { + "type": "table", + "bbox": [ + 0.7, + 0.211, + 0.895, + 0.262 + ], + "angle": 0, + "content": "
D(I)ODSS
LLLALLLA
ISP [16]79.461.747.629.9
D(I) in [48]79.663.153.236.3
Gamma curve80.164.053.636.6
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.719, + 0.264, + 0.873, + 0.276 + ], + "angle": 0, + "content": "(b) \\(D(\\mathbf{I})\\) in \\(\\S 3.3\\) and Fig. 3." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.284, + 0.906, + 0.328 + ], + "angle": 0, + "content": "Table 4. Ablating design choices in TorchAdapt. Highlighted settings are set as default. OD and SS are object detection and semantic segmentation, respectively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.335, + 0.729, + 0.351 + ], + "angle": 0, + "content": "4.5. Video Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.431 + ], + "angle": 0, + "content": "Settings. In addition to image-based tasks, we evaluate the generalizability of TorchAdapt on video data. We present the dataset details in Table 1. Other experimental settings follow [26], with additional implementation details provided in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.909, + 0.569 + ], + "angle": 0, + "content": "Results. We report the results of the challenging video object detection task in Table 2. While all LLIE methods perform poorly on this task, end-to-end training significantly boosts the performance of zero-reference methods [24, 50]. TorchAdapt, however, outperforms all methods under both LL and LA settings, achieving substantial gains of \\(+4.5\\) and \\(+6.4\\), respectively. These improvements affirm TorchAdapt's robustness and effectiveness for video object detection across both low-light and light-agnostic settings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.68, + 0.592 + ], + "angle": 0, + "content": "4.6. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.675 + ], + "angle": 0, + "content": "We ablate the components and design choices of TorchAdapt in object detection (OD) and instance segmentation (IS) tasks under (Low-light) LL and (Light-Agnostic) LA settings, following the same implementation details in § 4.1 and § 4.3, respectively. We provide more ablations in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Contribution of Each Component. Table 3 evaluates the impact of the Torch module, the Adapt module, and Illumination Scaling. Adding the Torch module alone (row 2) improves performance over the baseline with \\(+0.3\\%\\) in LL OD and \\(+5.1\\%\\) in LL IS, highlighting the benefit of feature enhancement in low-light vision. Introducing the Adapt module alongside Torch (row 3) yields significant improvements \\((+0.8\\% / +0.6\\%)\\) in LL/LA OD; \\(+2.2\\% / +1.4\\%\\) in LL/LA IS), underscoring the importance of content-adaptive modulation. Employing scaling without the Adapt module (Row 4) offers minor gains in LL settings. However, it is detrimental in LA settings \\((-0.2\\% / -0.7\\%)\\) in OD/IS), indicating that content-based adaptation is more impactful than global scaling alone. The full TorchAdapt configuration, combining all three components (row 5), achieves the best results with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.486, + 0.901 + ], + "angle": 0, + "content": "2https://github.com/open-mmlab/mmdetection/blob/main/configs/rtmdet/" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5651" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.093, + 0.482, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.212, + 0.484, + 0.282 + ], + "angle": 0, + "content": "Figure 5. Visualizing backbone features with SCI [50] on LA object detection using YOLOv3. (a) SCI's official checkpoint as an enhancer. (b) \\(\\mathrm{SCI}^{\\dagger}\\) trained end-to-end. (c) SCI integrated into our TorchAdapt framework as a Torch. SCI with TorchAdapt extracts more representative features in both low and well-lit images." + }, + { + "type": "table", + "bbox": [ + 0.179, + 0.284, + 0.398, + 0.364 + ], + "angle": 0, + "content": "
MethodmAP (%)
DSFD [36]16.1
CICony [32]18.4
SimMinMax [48]25.7
DAI-Net [20]‡28.0
DAI-Net+TorchAdapt31.9+3.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.374, + 0.483, + 0.457 + ], + "angle": 0, + "content": "Table 5. Zero-shot performance comparison of TorchAdapt with state-of-the-art domain adaptation methods on DARK FACE [80] using DSFD [36]. Results with \\(\\ddagger\\) are reproduced using the original codebase. TorchAdapt enhances prior SOTA [20] by \\(+3.9\\mathrm{mAP}\\), demonstrating its effectiveness in zero-shot domain adaptation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.466, + 0.484, + 0.526 + ], + "angle": 0, + "content": "+1.3%/+1.6% in LL/LA OD and +7.6%/+2.5% in LL/LA IS over the baseline, confirming that these components complement each other and synergistically enhance performance in both LL and LA settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.633 + ], + "angle": 0, + "content": "Design Choices in TorchAdapt. Table 4 ablates important design choices in TorchAdapt. As shown in Table 4a, we empirically establish that when the scale factor \\(\\gamma\\) in Eq. 3 is computed with \\(\\kappa = 3\\), it produces optimal performance. Results in Table 4b reveal that employing the Gamma Curve in \\(D(I)\\) to generate \\(V_{2}\\) in § 3.3 brings the biggest boost in performance across both LL and LA settings." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.659 + ], + "angle": 0, + "content": "4.7. Comparison with Domain-Adaptation Methods" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.664, + 0.484, + 0.8 + ], + "angle": 0, + "content": "We evaluate TorchAdapt's effectiveness for zero-shot domain adaptive face detection on the DARK FACE dataset [80]. Following settings from [20], we train the detector only on the WIDER FACE dataset [78]. As shown in Table 5, TorchAdapt integrated with DAI-Net [20], the current best-performing method, achieves a significant improvement of \\(+3.9\\%\\) mAP, demonstrating its effectiveness in enabling robust zero-shot adaptation under diverse illumination conditions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.815, + 0.411, + 0.833 + ], + "angle": 0, + "content": "5. Exploring TorchAdapt's Properties" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Model-Agnostic Generality. Experiments in § 4 confirm that TorchAdapt is a general-purpose module that can be integrated into any high-level vision task. Moreover, TorchAdapt as a framework is highly flexible: any zero-reference (unsu" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.613, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.144, + 0.569, + 0.153 + ], + "angle": 0, + "content": "I" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.09, + 0.709, + 0.143 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.144, + 0.702, + 0.154 + ], + "angle": 0, + "content": "TorchAdapt(I)" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.09, + 0.805, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.144, + 0.763, + 0.153 + ], + "angle": 0, + "content": "I" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.09, + 0.904, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.144, + 0.896, + 0.154 + ], + "angle": 0, + "content": "TorchAdapt(I)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.167, + 0.908, + 0.209 + ], + "angle": 0, + "content": "Figure 6. Illustrating low-light image Enhancement as an emergent property of TorchAdapt on the DARK FACE validation set. More results can be found in Appendix B." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.227, + 0.905, + 0.264 + ], + "angle": 0, + "content": "
MethodmAP50(LL)mAP50(LA)#Params.(M)Latency (ms)
Baseline [57]78.862.465.221.9
+TorchAdapt80.1 (+1.3)64.0 (+1.6)65.3 (+0.1)23.7 (+1.8)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.274, + 0.908, + 0.316 + ], + "angle": 0, + "content": "Table 6. Computational analysis. With YOLOv3 [57], TorchAdapt introduces minimal computational overhead while bringing stronger gains across both LL and LA settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.456 + ], + "angle": 0, + "content": "pervised) LLIE methods [33, 50, 85] can serve as the Torch component to improve performance under varying illuminations. To verify this, we adopt SCI [50] as the Torch module in TorchAdapt and visualize the learned backbone features in Fig.5(c). Using SCI within TorchAdapt yields highly representative backbone features for both low-light and well-lit images. Appendix C provides additional examples and quantitative analyses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.909, + 0.609 + ], + "angle": 0, + "content": "Enhancement as an Emergent Property. Although TorchAdapt is designed to produce illumination-invariant features for high-level vision tasks without employing any explicit enhancement loss functions during training, we observe that it inherently enhances low-light images. We illustrate this effect in Fig. 6. This emergent behavior likely results from the pre-training objective of aligning representations between well-lit and low-light images (see § 3.3), which encourages the network to implicitly adjust low-light images toward well-lit representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.909, + 0.718 + ], + "angle": 0, + "content": "Computational Analysis. We assess the computational efficiency of TorchAdapt using YOLOv3 [57] on the object detection task. As shown in Table 6, integrating TorchAdapt incurs minimal additional computational cost while delivering significant performance improvements. This demonstrates that TorchAdapt is compatible with lightweight models and suitable for real-time applications." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.738, + 0.64, + 0.753 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.9 + ], + "angle": 0, + "content": "In this work, we introduce TorchAdapt, a real-time, general-purpose, light-agnostic adaptive feature enhancement framework that boosts the performance of high-level vision tasks under varying illumination conditions. By integrating its modules with light-agnostic learning, TorchAdapt produces powerful semantic representations for both low-light and well-lit data. Extensive experiments on five representative visual perception tasks involving images and videos confirm its effectiveness." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5652" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.147 + ], + "angle": 0, + "content": "This work was partially funded by the European Union under Horizon Europe, grant number 101135724 (LUMINOUS)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.162, + 0.188, + 0.178 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.187, + 0.484, + 0.27 + ], + "angle": 0, + "content": "[1] Muhamamd Zeshan Afzal, SK Aziz Ali, Didier Stricker, Peter Eisert, Anna Hilsmann, Daniel Perez-Marcos, Marco Bianchi, Sonia Crottaz-Herbette, Roberto De Ioris, Eleni Mangina, et al. Next generation xr systems-large language models meet augmented and virtual reality. IEEE computer graphics and applications, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.273, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[2] Tadahiro Azetsu and Noriaki Suetake. Hue-preserving image enhancement in cielab color space considering color gamut. Optical Review, 26:283-294, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.485, + 0.356 + ], + "angle": 0, + "content": "[3] Ruzena Bajcsy, Yiannis Aloimonos, and John K Tsotsos. Revisiting active perception. Autonomous Robots, 42:177-196, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[4] Peter J Bickel and Kjell A Doksum. Mathematical statistics: basic ideas and selected topics, volumes I-II package. Chapman and Hall/CRC, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[5] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Fredo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. In CVPR 2011, pages 97-104. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.484, + 0.53 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.484, + 0.587 + ], + "angle": 0, + "content": "[7] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE international conference on computer vision, pages 2722-2730, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[8] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3291-3300, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.756 + ], + "angle": 0, + "content": "[9] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[10] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. International Journal of Computer Vision, 131(8):2198-2218, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.187, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.192, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[14] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.248, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.331, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[16] Ziteng Cui, Guo-Jun Qi, Lin Gu, Shaodi You, Zenghui Zhang, and Tatsuya Harada. Multitask aet with orthogonal tangent regularity for dark object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2553-2562, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.401, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[17] Ziteng Cui, Kunchang Li, Lin Gu, Shenghan Su, Peng Gao, Zhengkai Jiang, Yu Qiao, and Tatsuya Harada. You only need 90k parameters to adapt light: a light weight transformer for image enhancement and exposure correction. arXiv preprint arXiv:2205.14871, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.471, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.527, + 0.907, + 0.635 + ], + "angle": 0, + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.638, + 0.907, + 0.704 + ], + "angle": 0, + "content": "[20] Zhipeng Du, Miaojing Shi, and Jiankang Deng. Boosting object detection with zero-shot day-night domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12666-12676, 2024. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.907, + 0.773 + ], + "angle": 0, + "content": "[21] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.907, + 0.816 + ], + "angle": 0, + "content": "[22] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5653" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[24] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. CoRR, abs/2001.06826, 2020. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.486, + 0.216 + ], + "angle": 0, + "content": "[25] Jiang Hai, Zhu Xuan, Ren Yang, Yutong Hao, Fengzhu Zou, Fang Lin, and Songchen Han. R2rnet: Low-light image enhancement via real-low to real-normal network. Journal of Visual Communication and Image Representation, 90:103712, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.486, + 0.3 + ], + "angle": 0, + "content": "[26] Khurram Azeem Hashmi, Goutham Kallempudi, Didier Stricker, and Muhammad Zeshan Afzal. Featenhancer: Enhancing hierarchical features for object detection and beyond under low-light vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6725-6735, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.486, + 0.37 + ], + "angle": 0, + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.486, + 0.426 + ], + "angle": 0, + "content": "[28] Mingbo Hong, Shen Cheng, Haibin Huang, Haoqiang Fan, and Shuaicheng Liu. You only look around: Learning illumination invariant feature for low-light object detection. arXiv preprint arXiv:2410.18398, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.486, + 0.468 + ], + "angle": 0, + "content": "[29] Sergey Ioffe. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.486, + 0.537 + ], + "angle": 0, + "content": "[30] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhancement without paired supervision. IEEE Transactions on Image Processing, 30: 2340-2349, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.486, + 0.607 + ], + "angle": 0, + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.486, + 0.676 + ], + "angle": 0, + "content": "[32] Attila Lengyel, Sourav Garg, Michael Milford, and Jan C. van Gemert. Zero-shot day-night domain adaptation with a physics prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4399-4409, 2021. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.486, + 0.734 + ], + "angle": 0, + "content": "[33] Chongyi Li, Chunle Guo Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.486, + 0.803 + ], + "angle": 0, + "content": "[34] Chongyi Li, Chunle Guo, Shangchen Zhou, Qiming Ai, Ruicheng Feng, and Chen Change Loy. Flexicurve: Flexible piecewise curves estimation for photo retouching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1092-1101, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.486, + 0.872 + ], + "angle": 0, + "content": "[35] Hebei Li, Jin Wang, Jiahui Yuan, Yue Li, Wenming Weng, Yansong Peng, Yueyi Zhang, Zhiwei Xiong, and Xiaoyan Sun. Event-assisted low-light video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3250-3259, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.486, + 0.901 + ], + "angle": 0, + "content": "[36] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.909, + 0.134 + ], + "angle": 0, + "content": "dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.909, + 0.219 + ], + "angle": 0, + "content": "[37] Jinlong Li, Baolu Li, Zhengzhong Tu, Xinyu Liu, Qing Guo, Felix Juefei-Xu, Runsheng Xu, and Hongkai Yu. Light the night: A multi-condition diffusion framework for unpaired low-light enhancement in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15205-15215, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.909, + 0.302 + ], + "angle": 0, + "content": "[38] Yunan Li, Yihao Zhang, Shoude Li, Long Tian, Dou Quan, Chaoneng Li, and Qiguang Miao. Watching it in dark: A target-aware representation learning framework for high-level vision tasks in low illumination. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.909, + 0.359 + ], + "angle": 0, + "content": "[39] Jinxiu Liang, Jingwen Wang, Yuhui Quan, Tianyi Chen, Jiaying Liu, Haibin Ling, and Yong Xu. Recurrent exposure generation for low-light face detection. IEEE Transactions on Multimedia, 24:1609-1621, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.909, + 0.431 + ], + "angle": 0, + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312, 2014. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.909, + 0.473 + ], + "angle": 0, + "content": "[41] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. CoRR, abs/1708.02002, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.909, + 0.544 + ], + "angle": 0, + "content": "[42] Hewei Liu, Renwei Yang, Shuyuan Zhu, Xing Wen, and Bing Zeng. Luminance-guided chrominance image enhancement for hevc intra coding. In 2022 IEEE International Symposium on Circuits and Systems (ISCAS), pages 3180-3184. IEEE, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.909, + 0.601 + ], + "angle": 0, + "content": "[43] Jiaying Liu, Dejia Xu, Wenhan Yang, Minhao Fan, and Haofeng Huang. Benchmarking low-light image enhancement and beyond. International Journal of Computer Vision, 129:1153-1184, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.909, + 0.673 + ], + "angle": 0, + "content": "[44] Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10561-10570, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.909, + 0.744 + ], + "angle": 0, + "content": "[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.909, + 0.828 + ], + "angle": 0, + "content": "[46] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12009-12019, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.909, + 0.871 + ], + "angle": 0, + "content": "[47] Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. CoRR, abs/1805.11227, 2018. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.909, + 0.901 + ], + "angle": 0, + "content": "[48] Rundong Luo, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Similarity min-max: Zero-shot day-night domain adaptation." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5654" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8104-8114, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[49] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[50] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5637-5646, 2022. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[51] Tengyu Ma, Long Ma, Xin Fan, Zhongxuan Luo, and Risheng Liu. PIA: parallel architecture with illumination accelerator for joint enhancement and detection in low-light. In MM '22: The 30th ACM International Conference on Multimedia, Lisboa, Portugal, October 10 - 14, 2022, pages 2070-2078. ACM, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[52] Igor Morawski, Yu-An Chen, Yu-Sheng Lin, and Winston H. Hsu. NOD: taking a closer look at detection under extreme low-light conditions with night object detection dataset. CoRR, abs/2110.10364, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.484, + 0.489 + ], + "angle": 0, + "content": "[53] Lukáš Neumann, Michelle Karg, Shanshan Zhang, Christian Scharfenberger, Eric Piegert, Sarah Mistr, Olga Prokofyeva, Robert Thiel, Andrea Vedaldi, Andrew Zisserman, et al. Nightowls: A pedestrians at night dataset. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part I 14, pages 691–705. Springer, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.544 + ], + "angle": 0, + "content": "[54] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[55] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.484, + 0.686 + ], + "angle": 0, + "content": "[56] Qingpao Qin, Kan Chang, Mengyuan Huang, and Guiqing Li. Denet: Detection-driven enhancement network for object detection under adverse weather conditions. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 2813-2829, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[57] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. CoRR, abs/1804.02767, 2018. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[58] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[59] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Guided curriculum model adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7374-7383, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[60] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "IEEE/CVF International Conference on Computer Vision (ICCV), pages 10765-10775, 2021. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.909, + 0.177 + ], + "angle": 0, + "content": "[61] Yukihiro Sasagawa and Hajime Nagahara. Yolo in the dark - domain adaptation method for merging multiple models. In Computer Vision – ECCV 2020, pages 345–359, Cham, 2020. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.908, + 0.233 + ], + "angle": 0, + "content": "[62] Shangquan Sun, Wenqi Ren, Tao Wang, and Xiaochun Cao. Rethinking image restoration for object detection. Advances in Neural Information Processing Systems, 35:4461-4474, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.908, + 0.277 + ], + "angle": 0, + "content": "[63] Maria Valera and Sergio A Velastin. Intelligent distributed surveillance systems: a review. IEE Proceedings-Vision, Image and Signal Processing, 152(2):192-204, 2005. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.908, + 0.345 + ], + "angle": 0, + "content": "[64] Wenjing Wang, Wenhan Yang, and Jiaying Liu. Hla-face: Joint high-low adaptation for low light face detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16195-16204, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[65] Wenjing Wang, Zhengbo Xu, Haofeng Huang, and Jiaying Liu. Self-aligned concave curve: Illumination enhancement for unsupervised adaptation. In Proceedings of the 30th ACM international conference on multimedia, pages 2617-2626, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.908, + 0.474 + ], + "angle": 0, + "content": "[66] Wenjing Wang, Xinhao Wang, Wenhan Yang, and Jiaying Liu. Unsupervised face detection in the dark. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1250-1266, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.908, + 0.544 + ], + "angle": 0, + "content": "[67] Wenjing Wang, Huan Yang, Jianlong Fu, and Jiaying Liu. Zero-reference low-light enhancement via physical quadruple priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26057-26066, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.908, + 0.603 + ], + "angle": 0, + "content": "[68] Xinzhe Wang, Kang Ma, Qiankun Liu, Yunhao Zou, and Ying Fu. Multi-object tracking in the dark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 382-392, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.908, + 0.645 + ], + "angle": 0, + "content": "[69] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. CoRR, abs/1808.04560, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.908, + 0.689 + ], + "angle": 0, + "content": "[70] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.908, + 0.772 + ], + "angle": 0, + "content": "[71] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wenhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5901-5910, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.908, + 0.844 + ], + "angle": 0, + "content": "[72] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15769-15778, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[73] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17714-17724, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5655" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.147 + ], + "angle": 0, + "content": "[74] Xiaogang Xu, Ruixing Wang, and Jiangbo Lu. Low-light image enhancement via structure modeling and guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9893-9903, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.244 + ], + "angle": 0, + "content": "[75] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Arid: A new dataset for recognizing action in the dark. In Deep Learning for Human Activity Recognition: Second International Workshop, DL-HAR 2020, Held in Conjunction with IJCAI-PRICAI 2020, Kyoto, Japan, January 8, 2021, Proceedings 2, pages 70–84. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[76] Xinwei Xue, Jia He, Long Ma, Yi Wang, Xin Fan, and Risheng Liu. Best of both worlds: See and understand clearly in the dark. In Proceedings of the 30th ACM International Conference on Multimedia, page 2154–2162, New York, NY, USA, 2022. Association for Computing Machinery. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[77] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.484, + 0.441 + ], + "angle": 0, + "content": "[78] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5525-5533, 2016. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.509 + ], + "angle": 0, + "content": "[79] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12918-12927, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[80] Wenhan Yang, Ye Yuan, Wenqi Ren, Jiaying Liu, Walter J. Scheirer, Zhangyang Wang, Taiheng Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, Yuqiang Zheng, Yanyun Qu, Yuhong Xie, Liang Chen, Zhonghao Li, Chen Hong, Hao Jiang, Siyuan Yang, Yan Liu, Xiaochao Qu, Pengfei Wan, Shuai Zheng, Minhui Zhong, Taiyi Su, Lingzhi He, Yandong Guo, Yao Zhao, Zhenfeng Zhu, Jinxiu Liang, Jingwen Wang, Tianyi Chen, Yuhui Quan, Yong Xu, Bo Liu, Xin Liu, Qi Sun, Tingyu Lin, Xiaochuan Li, Feng Lu, Lin Gu, Shengdi Zhou, Cong Cao, Shifeng Zhang, Cheng Chi, Chubing Zhuang, Zhen Lei, Stan Z. Li, Shizheng Wang, Ruizhe Liu, Dong Yi, Zheming Zuo, Jianning Chi, Huan Wang, Kai Wang, Yixiu Liu, Xingyu Gao, Zhenyu Chen, Chang Guo, Yongzhou Li, Huicai Zhong, Jing Huang, Heng Guo, Jianfei Yang, Wenjuan Liao, Jiangang Yang, Liguo Zhou, Mingyue Feng, and Likun Qin. Advancing image understanding in poor visibility environments: A collective benchmark study. IEEE Transactions on Image Processing, 29:5737-5752, 2020. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[81] Bo Zhang, Yuchen Guo, Runzhao Yang, Zhihong Zhang, Jiayi Xie, Jinli Suo, and Qionghai Dai. Darkvision: A benchmark for low-light image/video perception, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[82] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[83] Liheng Zhang, Guo-Jun Qi, Liqiang Wang, and Jiebo Luo. Aet vs. aed: Unsupervised representation learning by auto" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.134 + ], + "angle": 0, + "content": "encoding transformations rather than data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2547-2555, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[84] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. CoRR, abs/1905.04161, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[85] Zhaoyang Zhang, Yitong Jiang, Jun Jiang, Xiaogang Wang, Ping Luo, and Jinwei Gu. Star: A structure-aware lightweight transformer for real-time image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4106-4115, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.317 + ], + "angle": 0, + "content": "[86] Zhao Zhang, Huan Zheng, Richang Hong, Mingliang Xu, Shuicheng Yan, and Meng Wang. Deep color consistent network for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1899-1908, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.907, + 0.374 + ], + "angle": 0, + "content": "[87] Hegui Zhu, Kai Wang, Ziwei Zhang, Yuelin Liu, and Wuming Jiang. Low-light image enhancement network with decomposition and adaptive information fusion. Neural Computing and Applications, 34(10):7733-7748, 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "5656" + } + ] +] \ No newline at end of file diff --git a/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/60f819c6-a0ae-4fb0-be50-050c7068618b_origin.pdf b/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/60f819c6-a0ae-4fb0-be50-050c7068618b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8ab977f02d22a9f39852df53a33785eae2ff1128 --- /dev/null +++ b/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/60f819c6-a0ae-4fb0-be50-050c7068618b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:13d8d78d87036e267d44894eac86f3c1e1244c0a93873685ce1506a1d1c1577e +size 6831243 diff --git a/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/full.md b/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/full.md new file mode 100644 index 0000000000000000000000000000000000000000..00cb818674fb70c1a3b5cbad95f4572bf2706877 --- /dev/null +++ b/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/full.md @@ -0,0 +1,377 @@ +# TorchAdapt: Towards Light-Agnostic Real-Time Visual Perception + +Khurram Azeem Hashmi $^{1,2}$ + +Karthik Palyakere Suresh1,2 + +Muhammad Zeshan Afzal1,2 + +$^{1}$ DFKI + +$^{2}$ RPTU Kaiserslautern-Landau + +khurram_azeem Hashmi@dfki.de + +Didier Stricker1,2 + +# Abstract + +Low-light conditions significantly degrade the performance of high-level vision tasks. Existing approaches either enhance low-light images without considering normal illumination scenarios, leading to poor generalization, or are tailored to specific tasks. We propose TorchAdapt, a real-time adaptive feature enhancement framework that generalizes robustly across varying illumination conditions without degrading performance in well-lit scenarios. TorchAdapt consists of two complementary modules: the Torch module enhances semantic features beneficial for downstream tasks, while the Adapt module dynamically modulates these enhancements based on input content. Leveraging a novel light-agnostic learning strategy, TorchAdapt aligns feature representations of enhanced and well-lit images to produce powerful illumination-invariant features. Extensive experiments on multiple high-level vision tasks, including object detection, face detection, instance segmentation, semantic segmentation, and video object detection, demonstrate that TorchAdapt consistently outperforms state-of-the-art low-light enhancement and task-specific methods in both low-light and light-agnostic settings. TorchAdapt thus provides a unified, flexible solution for robust visual perception across diverse lighting conditions. + +# 1. Introduction + +Recent advances in visual perception tasks [13, 31, 77, 82] are highly effective at extracting features for complex tasks when provided with high-quality inputs. However, replicating this performance is challenging when data is scarce, noisy, or expensive to obtain. A common issue is varying illumination, particularly low-light conditions, which hinders the robustness and reliability of safety-critical visual perception systems [1, 3, 7, 63]. + +A common strategy to improve low-light performance involves pre-processing images with Low-Light Image Enhancement (LLIE) methods [17, 24, 33, 44, 50, 51, 71, 84, 85] before feeding them into high-level vision mod + +![](images/fdef7c1f1dd5cc82fa6f742b32d83cd361a3634712105dbd6ce9d14c33371195.jpg) + +![](images/7501f044036cdd8bc76406a59833e2f05e74cd09ccee5d765a4f21363bea0a95.jpg) +(a) Low-Light Vision + +![](images/3158cf49653bf7978afeaf8384958c3197e9d7d59141b96ad775fa98e5129ce4.jpg) + +![](images/393d6a987f61d20407ff9cce26d641509d477f9781212f06f5c3bc4f50e7bef8.jpg) +Figure 1. Prior downstream methods in low-light vision assume that all samples belong to low-light conditions as shown in (a), leading to underperformance in varying illumination environments, depicted in (b). TorchAdapt alleviates this bottleneck via learning powerful illumination-invariant features and achieves consistent gains across low-light and light-agnostic settings on Object Detection (OD), Face Detection (FD), Instance Segmentation (IS), Semantic Segmentation (SS), and Video Object Detection (VOD). Best viewed on the screen. + +![](images/8938b69d45be7c366a1c6a63ef8e45639c0e7021cb67de8f76069e68c29cb623.jpg) +(b) Light-Agnostic Vision + +![](images/0895c393afe8577eb301659288cba0e1da35a2003dc2cb84ff8c7c41ce47946d.jpg) + +els. However, this simple integration often leads to weaker gains [16, 26], as these arts conform to human visual perception (optimized on pixel-level objectives like L1 or MSE losses [4]) instead of machine visual perception. Moreover, applying these enhancement methods indiscriminately can degrade the performance on images/videos captured under standard lighting conditions, as alterations intended for low-light scenarios [47, 60, 80, 81] may be unnecessary or even detrimental in well-lit environments [15, 40, 58, 78]. + +To address low-light vision tasks, recent methods employ transfer learning [16, 59, 72] and domain adaptation techniques [20, 32, 48]. Despite improvements, these efforts are carefully designed for a single high-level vision task, such + +as object detection [16] or semantic segmentation [59, 72], and require cumbersome multi-stage training to adapt to each dense task [20, 32, 48]. Moreover, they often assume consistently low-light data, leading to performance degradation on well-lit images and videos, sometimes worse than baselines (see Table 2). In practical applications with unpredictable lighting—such as autonomous driving [7, 37] at different times of the day, surveillance systems [63] monitoring both indoor and outdoor environments or robots [3] operating between well-lit and low-lit areas—these limitations become critical. This highlights the need for high-level vision models to be light-agnostic—capable of maintaining high performance across varying lighting conditions, as visualized in Fig. 1(b). + +In this work, we propose TorchAdapt, a real-time, light-agnostic adaptive feature enhancement framework designed to improve the performance of any high-level vision task under varying illumination. TorchAdapt comprises two primary modules: Torch and Adapt, illustrated in Fig. 2. The Torch module is a lightweight feature enhancement network that enriches semantic representations favorable for downstream vision tasks. The Adapt module learns to modulate this feature enhancement dynamically based on input data, amplifying or reducing the effect of the Torch module. + +TorchAdapt builds on the consensus [26, 38] that low-to-normal light adaptation lacks a physical model, leading to unexpected artifacts during optimization when transforming images in pixel space. Therefore, aligning representations in the feature space is a more practical approach for high-level vision tasks [48, 65]. We optimize TorchAdapt using a non-contrastive self-supervised learning objective [12, 23] to align representations from low-light and well-lit images. However, unlike prior works operating in a similar direction [16, 48, 65], TorchAdapt also learns not to adversely affect representations of well-lit images. As illustrated in Fig. 3, we take a well-lit image $V_{1}$ and its synthetically generated low-light counterpart $V_{2}$ , pass them through TorchAdapt to obtain enhanced images $\hat{V}_{1}$ and $\hat{V}_{2}$ , and feed them along with $V_{1}$ into a pre-trained frozen visual encoder [19, 27, 55] to generate representations. Since the encoder is frozen, only TorchAdapt is optimized to help the encoder produce well-lit representations (like $V_{1}$ ) for $\hat{V}_{1}$ and $\hat{V}_{2}$ . + +Thanks to the conceptual simplicity, TorchAdapt enjoys two benefits. ① It can be trained with a simple self-supervised loss on a general-purpose dataset like ImageNet [18], circumventing large-scale real-world low-light datasets. ② The training of TorchAdapt is independent of the high-level vision task and needs to be done only once to align with the backbone networks [6, 19, 27]. Later, TorchAdapt can be plugged into any downstream tasks in a light-agnostic manner. We demonstrate the effectiveness of TorchAdapt on five representative visual perception tasks, including object detection [21, 47], face detection [78, 80], semantic + +segmentation [15, 60], instance segmentation [10, 40], and video object detection [58, 81] under both low-light and light-agnostic settings. For all tasks, TorchAdapt brings consistent and significant gains over baselines under both low-light and light-agnostic settings, outperforming previous state-of-the-art task-specific and LLIE methods (see Fig. 1). Moreover, TorchAdapt accomplishes all this with negligible computational overhead (see Table 6). + +Contributions. Method: We propose TorchAdapt, a real-time, light-agnostic adaptive feature enhancement framework that improves the performance of any high-level vision task under varying illumination without compromising accuracy in well-lit conditions. Simplicity: Thanks to its architectural innovations and self-supervised objective, TorchAdapt significantly simplifies the training strategy for any high-level vision task pipeline, circumventing corresponding paired images/videos, as shown in Fig 3. Flexibility: TorchAdapt is independent of vision encoders and downstream task architectures, demonstrating model-agnostic generality (see § 5). Any zero-reference LLIE method [24, 33, 50, 67] can serve as the Torch module in our framework to achieve light-agnostic performance (see Fig. 5). Results: We conduct extensive experiments on five visual perception tasks involving images and videos under both low-light and light-agnostic settings using curated datasets. TorchAdapt outperforms state-of-the-art task-specific and LLIE methods across all settings when integrated with the same baselines (see Fig. 1). + +# 2. Related Work + +# 2.1. Enhancing Low-Light Images + +Low-light Image Enhancement (LLIE) aims to improve the visual quality of images captured under insufficient lighting conditions. Modern learning-based approaches have become prevalent, thanks to the emergence of low-light datasets [5, 8, 25, 43, 70]. Supervised methods [70, 73, 74, 84, 86] leverage these datasets to design effective networks trained to predict normal-light images from low-light ones. Unsupervised LLIE approaches [30, 34, 79] employ adversarial learning from unpaired supervision to relax the requirement of paired normal-low light data. However, all of these works still rely on specific training data, which limits their generalizability to unseen scenarios. Another line of work known as zero-reference methods [24, 33, 44, 50, 67] get rid of both paired and unpaired data to enhance low-light images. These arts design enhancement networks that optimize on a set of non-reference loss functions. Since these works are mainly designed to conform to human visual perception, they bring sub-optimal gains for machine visual perception tasks. Furthermore, they are not light-agnostic (capable of handling both normal and low-light images). This work bridges this gap by complementing these zero-reference methods, + +![](images/6602a677e24ffe63a76d0f3c9d0e5d66f941f01870e4ff0cd1d6c44b1dca1f92.jpg) +Figure 2. TorchAdapt architectural overview. The Torch module enhances feature representation by applying $3 \times 3$ convolutions and element-wise operations (Eq. 1). The Adapt module performs adaptive modulation, creating a content-aware response based on the input $I$ (Eq. 2). Both modules are fused via Adaptive Fusion, integrating feature enhancements from Torch and contextual adaptations from Adapt. This adaptive fusion produces illumination-invariant features, improving robustness across varying lighting conditions. Adaptive Fusion is explained in Eq. 4. + +where they can be employed as a Torch in our TorchAdapt framework to improve light-agnostic visual perception. + +# 2.2. Low-Light Visual Perception + +This direction of work considers machine perception as the criteria for success while enhancing low-light images. + +Low-light Object Detection (LLOD). Object detection, as a fundamental problem in computer vision, has also been well-explored in low-light vision [16, 20, 39, 51, 61, 62, 64-66]. Based on the dataset, objects vary and can be generic [47, 52] or specific like pedestrians [53], and human faces [80]. One category of LLOD methods [51, 56, 62, 65] jointly learns the enhancement and detection task in an end-to-end fashion to improve the overall detection performance. Other approaches involve learning a low-light detector through fusing multiple models [61] or resorting to domain adaptation frameworks [16, 20, 64], minimizing the distribution variance between normal and low-light images. Most of these works are specifically designed to tackle the dark object detection task. Furthermore, when these methods are optimized for LLOD, they underperform on normal-light images. In contrast, TorchAdapt is fully task-agnostic for any high-level vision task, and it does not worsen the performance of the detector on normal-light data. + +Low-light Tasks-Agnostic Methods. Recently, various low-light task-agnostic methods [20, 26, 32, 48, 65] have emerged in the literature. Unlike prior approaches, they are more flexible and can be integrated to improve any downstream task under low-light vision. Our work follows the same spirit and offers similar flexibility. FeatEnHancer [26], as one of the representative works, learns to enhance the hierarchical features suitable for downstream tasks. Other seminal attempts [20, 32, 48] treat low-light perception tasks as domain adaption (DA) learning. They exploit well-lit source domain data [40, 78] and learn to adapt to low-light target domain data [47, 80]. Although flexible, the DA methods rely on multi-stage training frameworks for each task [48]. In contrast, our TorchAdapt is pre-trained with the frozen back + +bone on the domain-independent data like ImageNet [18]. Moreover, it is important to recognize the efforts of recent real low-light datasets [10, 35, 60, 68, 75, 81] that enable the increasing research on several downstream tasks in low-light vision. + +# 2.3. Learning Equivariant Representations in LowLight Vision with Self-Supervised Objectives + +Annotating low-light vision datasets is a much more difficult task than annotating well-lit datasets. Hence, it is a common practice to apply self-supervised learning to learn illumination-invariant features [16, 65] or aligning well-lit and low-light data distributions [38, 48, 66]. MAET [16] adopts the autoencoding transformation [83] to decode illumination-degrading parameters and detection predictions. SACC [65] merges the pretext tasks of rotation prediction [22] and jigsaw puzzling [54] to learn illumination enhancement from low-light images. Most related to our work is SimMinMax [48], which adopts the BYOL [23] framework to learn model-level adaptation by maximizing the similarity of features between images at night and daytime. However, since SimMinMax [48] learns to adapt from daytime to nighttime images, it needs to be trained for each high-level vision task, requiring the corresponding daytime labeled dataset. In contrast, TorchAdapt leverages self-supervised objectives to align backbone features during the pretext task. Furthermore, it can be trained on any general-purpose dataset. This significantly simplifies the training procedure of TorchAdapt, as compared to SimMinMax, on any high-level vision task. TorchAdapt is also unique to prior works in its ability to handle light-agnostic data. + +# 3. TorchAdapt + +This section formally introduces TorchAdapt, explaining the proposed Torch in § 3.1 and Adapt in § 3.2 modules, followed by the light-agnostic learning in § 3.3. + +![](images/ed71c641c4623764435fb269b0b8cb1c6f6bfd6a39089703dc0d05e1aa566bb1.jpg) +Figure 3. Light-Agnostic learning pipeline. $D(.)$ is a darkening function applied to the well-lit image $V_{1}$ to synthesize the low-light view $V_{2}$ . Both $V_{1}$ and $V_{2}$ are processed by TorchAdapt and then fed into a frozen pre-trained vision encoder $f$ . A cosine similarity loss optimizes TorchAdapt to produce well-lit representations for both $V_{1}$ and $V_{2}$ , enabling it to learn illumination-invariant features. The TorchAdapt(.) function integrates Eq. 1 and Eq. 4. Refer to § 3.3 for a complete explanation of each symbol. + +# 3.1.Torch + +The Torch module in our TorchAdapt framework serves as an enhancement network as in Low-light Image Enhancement (LLIE) approaches [24, 33, 50]. However, unlike these methods—which focus on improving visual quality for human perception—our Torch module aims to boost semantic representations favorable for downstream vision tasks under varying lighting conditions. Given an input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ , the Torch module computes an enhancement map $\mathcal{H}_{\theta}(\mathbf{I})$ , parameterized by network weights $\theta$ . The transformed image $\mathbf{T}$ is obtained by: + +$$ +\mathbf {T} = \mathbf {I} \odot (1 + \mathcal {H} _ {\theta} (\mathbf {I})) \text {, w h e r e} \mathcal {H} _ {\theta} (\mathbf {I}) = \tanh (\mathcal {F} (\mathbf {I}; \theta)) \tag {1} +$$ + +where $\odot$ denotes element-wise multiplication. The $\tanh(\cdot)$ activation ensures that the adjustment values are in the range $[-1, 1]$ , allowing the model to both amplify (positive values) and attenuate (negative values) features. This formulation enables the model to adapt to the image's content effectively. Torch network. The Torch network $\mathcal{F}(\mathbf{I};\theta)$ is a lightweight convolutional network composed of an initial feature extraction layer followed by a residual block, as illustrated in Fig. 2. The residual learning enables the Torch module to focus on learning the necessary adjustments to improve the input image rather than reconstructing it entirely. Enhancing semantically important features while preserving original content makes the Torch module crucial in maintaining and improving performance in downstream vision tasks. Notably, thanks to the formulation explained in Eq 1, our Torch module only has 11,075 trainable parameters and requires 0.712 GFLOPs for $\mathbf{I}$ of size $256 \times 256 \times 3$ , making it suitable for real-time high-level vision tasks under varying lighting conditions. + +# 3.2. Adapt + +While the Torch module enhances the input image $\mathbf{I}$ to produce $\mathbf{T}$ , applying this enhancement uniformly may not be optimal. Different regions within an image require varying degrees of enhancement based on content (low/well-lit images) and context (downstream task). To address this, we introduce the Adapt module, which learns a content-adaptive modulation matrix in the feature space. This allows the model to selectively amplify or suppress the enhancements produced by the Torch module, improving their effectiveness for downstream vision tasks in varying illumination scenarios, effectively making it light-agnostic. + +Specifically, the Adapt module computes an Adaptive Matrix $\mathbf{M} \in \mathbb{R}^{H \times W}$ , where $H$ and $W$ are the spatial dimensions of $\mathbf{I}$ . The purpose of the Adaptive Matrix $\mathbf{M}$ is to modulate the feature enhancements from Torch $\mathbf{T}$ before they are applied to $\mathbf{I}$ . $\mathbf{M}$ is computed as: + +$$ +\mathbf {M} = \sigma \left(\mathcal {G} _ {\phi} (\mathbf {I})\right), \tag {2} +$$ + +where $\mathcal{G}_{\phi}$ is a lightweight convolutional network parameterized by weights $\phi$ , and $\sigma(\cdot)$ denotes the sigmoid activation function ensuring that the values of $\mathbf{M}$ are in the range $[0,1]$ . We feed the input image $\mathbf{I}$ into the Adapt module to compute $\mathbf{M}$ , ensuring that the image content directly influences the modulation. + +Adapt network. As shown in Fig. 2, the convolutional network $\mathcal{G}_{\phi}$ in the Adapt module is designed to extract hierarchical features, which are proven to be beneficial for high-level vision tasks such as object detection and semantic segmentation [26, 45, 46]. To capture multi-scale contextual information, the network consists of three $3\times 3$ convolutional layers with increasing channel dimensions $C_1$ , $C_2$ , and $C_3$ , respectively. Each layer is followed by a batch normalization [29] and a ReLU activation function. This architecture allows the network to learn a hierarchy of features from low-level edges and textures to high-level semantic information. Finally, an additional $3\times 3$ convolutional layer maps the features from $C_3$ to a single channel and applies a sigmoid activation function to generate the Adaptive Matrix $\mathbf{M}\in [0,1]^{H\times W}$ . This design balances expressiveness and computational efficiency, enabling the Adapt module to adjust enhancements based on both local and global features, which is crucial for content-adaptive modulation. + +Illumination-aware scaling. We introduce a scale factor $\gamma$ to ensure that the enhancement adapts appropriately to different illumination levels. This scale factor modulates the M globally based on the image's overall brightness, allowing the model to apply stronger enhancements to darker images and milder adjustments to well-lit images, maintaining visual consistency and preventing artifacts. We compute $\gamma$ as a function of the average luminance $L(\mathbf{I})$ of the input image: + +$$ +\gamma = \frac {\kappa}{1 + \exp (k (L (\mathbf {I}) - L _ {0}))}, \tag {3} +$$ + +where $\kappa$ is the maximum scale factor (empirically set to 3 unless stated otherwise). $k$ controls the slope of the sigmoid function, and $L_{0}$ is the luminance value at the sigmoid's midpoint. The average luminance $L(\mathbf{I})$ is computed following the standard luminance calculation in color science [2, 42, 87]. + +Adaptive Fusion. The final enhanced image $\mathbf{E}$ is then obtained by modulating the enhancement from the Torch module with the illumination-aware scaled Adaptive Matrix: + +$$ +\mathbf {E} = \mathbf {I} + ((\gamma \cdot \mathbf {M}) \odot \mathbf {T}). \tag {4} +$$ + +where $\odot$ denotes element-wise multiplication, and $\mathbf{T}$ is obtained from Eq. (1). This formulation in Eq. 4 integrates both content-based modulation (through $\mathbf{M}$ ) and illumination-based scaling (through $\gamma$ ), enabling the model to adaptively enhance images under varying lighting conditions. + +# 3.3. Light-Agnostic Learning + +Achieving robust performance across varying lighting conditions requires our model to learn representations invariant to illumination changes. Collecting labeled data under all possible lighting conditions is impractical. Hence, we adopt a self-supervised learning (SSL) strategy inspired by BYOL [23] to learn illumination-invariant features. Unlike prior works that design cumbersome pipelines to incorporate SSL into each high-level vision task [16, 48, 65], we leverage SSL to pre-train our TorchAdapt module with a pretext task that aligns backbone representations [19, 27] regardless of illumination. This approach eliminates the need for task-specific labeled data or low-light data, allowing training on general-purpose datasets like ImageNet [18]. Figure 3 illustrates the pre-training pipeline of TorchAdapt. + +Objective Function. Given a well-lit image $V_{1}$ , we generate its low-light counterpart $V_{2} = D(V_{1})$ using a darkening function $D(\cdot)$ , which can be a gamma transformation, an illumination-degradation pipeline [16], or a darkening module from [48]. Let $f$ denote the frozen pre-trained backbone network (e.g., ResNet [27], ViT [19]). We pass $V_{1}$ and $V_{2}$ through the TorchAdapt module to obtain enhanced images $\hat{V}_{1}$ and $\hat{V}_{2}$ , respectively. The enhanced images are processed by the backbone $f$ and a projector network $g$ to obtain feature representations: $z_{1}^{\prime} = g(f(\hat{V}_{1}))$ and $z_{2} = g(f(\hat{V}_{2}))$ . The target representation is obtained by passing the original well-lit image $V_{1}$ through $f^{\prime}$ and $g^{\prime}$ : $y_{1} = g^{\prime}(f^{\prime}(V_{1}))$ . Our training objective minimizes the cosine similarity loss $\mathcal{D}_{\mathrm{cos}}$ between the enhanced image representations and the well-lit target representation: + +$$ +\mathcal {L} = \left(1 - \frac {z _ {1} ^ {\prime}}{\| z _ {1} ^ {\prime} \| _ {2} \| y _ {1} \| _ {2}}\right) + \left(1 - \frac {z _ {2} ^ {\top} y _ {1}}{\| z _ {2} \| _ {2} \| y _ {1} \| _ {2}}\right), \tag {5} +$$ + +where $\| \cdot \| _2$ denotes the $\ell_2$ norm of a vector. Since we employ identical, frozen backbone networks $f$ , the loss function guides TorchAdapt to enhance the low-light image $\hat{V}_2$ + +![](images/c75950f2f37899bdc86d43e1a36906c387eaf13d8446bfa7f78a9f8280345623.jpg) +Figure 4. TorchAdapt's downstream integration pipeline. TorchAdapt works as a plug-and-play, backbone-agnostic feature enhancement module, seamlessly integrating into various downstream vision tasks. Given any image or video input across diverse lighting conditions, TorchAdapt produces illumination-invariant representations, significantly boosting performance without requiring task-specific modifications. + +so its representation aligns with that of the well-lit image $V_{1}$ . Simultaneously, it discourages over-enhancement of the already well-lit image $\hat{V}_{1}$ , as its representation should remain consistent with $V_{1}$ . This training strategy enables TorchAdapt to effectively enhance low-light images while preserving well-lit ones, promoting illumination invariance in the learned representations. + +Training Details. The training of TorchAdapt is fast and efficient due to its lightweight architecture and the absence of gradient propagation through the frozen backbone network. We train TorchAdapt for 5 epochs on the ImageNet [18] dataset using a single GPU, with a learning rate of 0.0001 and a batch size of 256. Since the backbone networks are pre-trained and we aim for consistent views, we apply only color jitter and random horizontal flip as data augmentations. Pre-training TorchAdapt with a ResNet-50 backbone takes $\sim 6$ hours on a single A100 GPU. Once trained, TorchAdapt can be seamlessly integrated into any downstream vision task, as illustrated in Fig. 4. + +
DatasetTask#Cls#Train#Val
Low-light (LL)
ExDark [47]Object detection1258911472
DARK FACE [80]Face detection15400600
ACDC [60]Semantic segmentation19400106
LIS [10]Instance segmentation81561669
DarkVision [81]Video object detection426*6*
Light-Agnostic (LA)
COCO [21] ∩ ExDark [47]Object detection12923575156
WIDER FACE [78] ∩ DARK FACE [80]Face detection1182803822
CityScapes [15] ∩ ACDC [60]Semantic segmentation193375606
COCO [40] ∩ LIS [10]Instance segmentation8443212550
ImageNet VID [58] ∩ DarkVision [81]Video object detection2401*51*
+ +Table 1. Datasets Statistics.*: video samples. Light-Agnostic datasets are obtained through Eq. 6 + +# 4. Experiments + +We conduct extensive experiments on several representative visual tasks to demonstrate the superiority of TorchAdapt in both low-light and light-agnostic settings. These tasks include object detection [40, 47, 52], face detection [78, 80], instance segmentation [10, 40], semantic segmentation [15, 60], and video object detection [58, 81]. This section first presents the experimental settings, comparing TorchAdapt + +
Tasks → Light Settings → MethodsObject Det.Face Det.Instance Seg.Semantic Seg.Video Object Det.
LLLALLLALLLALLLALLLA
mAP50mAPmAP50mAPAP50APAP50APmAPSegAPmAPSegAPmIoUmIoUmAPmAP
Baseline78.844.262.433.147.319.952.628.054.046.041.533.949.675.532.851.3
Enhancement
Zero-DCE [24]76.242.757.629.547.420.147.622.951.543.539.232.142.559.47.8328.3
SCI [84]75.542.257.929.745.119.345.421.752.344.235.128.342.157.46.7125.7
URetinexNet [71]74.642.057.329.244.718.346.622.359.543.137.429.741.258.86.6825.1
IAT [17]74.141.052.430.542.115.937.117.449.742.336.430.141.756.26.2424.5
Task-specific
Zero-DCE [24]‡79.144.462.933.250.421.150.123.559.152.744.435.744.761.034.151.5
SCI [50]‡79.244.263.033.451.622.254.729.958.451.043.734.943.559.932.950.7
MAET [16]‡79.244.357.327.344.318.747.923.1--------
Xue et al. [76]----46.117.744.221.0----42.169.7--
FeatEnHancer [26]‡79.244.862.933.347.219.951.624.559.652.444.936.150.176.134.652.7
TorchAdapt80.145.464.034.153.723.159.032.561.153.645.236.650.876.339.159.1
vs. prev. SoTA+0.9+0.6+1.0+0.7+2.1+0.9+4.3+2.6+1.5+0.9+0.3+0.5+0.7+0.2+4.5+6.4
+ +Table 2. Quantitative comparison across five tasks under both Low-Light (LL) and Light-Agnostic (LA) settings. ‡: all these methods are reproduced by us with their official codebase and trained end-to-end for each task. The rest of the results are obtained from their official checkpoints. - indicates that the method is not applicable. TorchAdapt outperforms prior SOTA across all five tasks in LL and LA settings. + +with baselines, existing low-light image enhancement (LLIE) methods, and state-of-the-art task-specific approaches. We then examine critical design choices for TorchAdapt through ablation studies. Finally, we validate the generalizability of TorchAdapt by comparing it with recent state-of-the-art domain adaptation methods [20, 32, 48]. + +Experimental Settings. We perform experiments under two settings for each task: + +1)- Low-Light (LL) Setting: Following prior works [16, 26, 76], we utilize publicly available low-light benchmarks specific to each task. These benchmarks consist exclusively of images captured under low-light conditions. + +2)- Light-Agnostic (LA) Setting: To introduce a more challenging and realistic evaluation, we curate datasets that encompass both well-lit and low-light images. We achieve this by selecting classes common to both well-lit and low-light datasets. Mathematically, for a given task, let $\mathcal{C}_{\mathrm{WL}}$ and $\mathcal{C}_{\mathrm{LL}}$ denote the sets of classes in the well-lit and low-light datasets, respectively. We define the set of common classes as $\mathcal{C}_{\mathrm{com}} = \mathcal{C}_{\mathrm{WL}} \cap \mathcal{C}_{\mathrm{LL}}$ . The LA dataset is then formed by combining all images from both datasets that belong to classes in $\mathcal{C}_{\mathrm{com}}$ : + +$$ +\mathcal {D} _ {\mathrm {L A}} = \left\{\left(x, y\right) \mid (x, y) \in \mathcal {D} _ {\mathrm {W L}} \cup \mathcal {D} _ {\mathrm {L L}}, y \in \mathcal {C} _ {\mathrm {c o m}} \right\}, \tag {6} +$$ + +Where $\mathcal{D}_{\mathrm{WL}}$ and $\mathcal{D}_{\mathrm{LL}}$ are the well-lit and low-light datasets, respectively, and $(x,y)$ represents an image and its label. We summarize the key statistics of the employed benchmarks for both settings in Table 1. We provide complete implementation details for each experiment in Appendix A. + +Comparisons with State-of-the-Arts. We directly compare our approach with several LLIE methods, including Zero-DCE [24], IAT [17], SCI [50], and URetinexNet [71]. Following common practice [16, 26, 76], we use their released checkpoints to enhance all images before feeding them into + +the detector for performance evaluation. Notably, due to their zero-reference nature requiring no paired images, some of these LLIE methods [24, 50] can be trained end-to-end on downstream vision tasks. Therefore, besides evaluating their conventional enhancement results, we integrate them into each task and report their performance, similar to low-light task-specific works like MAET [16] and FeatEnHancer [26] (indicated with $\ddagger$ in Table 2). Our findings reveal that end-to-end training of such LLIE works $[24, 50]^1$ significantly improves their performance on downstream vision tasks in both LL and LA settings. Concurrent to our work, we notice that two new methods [28, 38] appear in the literature on improving low-light object detection. However, since their code and checkpoints were not released by the time of the submission, we could not draw direct comparisons with them. Next, we discuss experiments on each task. + +# 4.1. Object Detection + +Settings. Following common practice [16] in LL object detection, we adopt YOLOv3 [57] as the baseline detector. For the LL setting, COCO [40] pre-trained weights are utilized to fine-tune YOLOv3 on the ExDark dataset [47]. We train the model from scratch in the LA setting, as COCO images are also included in the training data (see Table. 1). Results. As shown in Table 2, TorchAdapt achieves a superior $\mathrm{mAP}_{50}$ of 80.1 in LL and 64.1 in LA, compared to the baseline's 78.8 and 62.4. End-to-end training boosts the performance of enhancement methods like Zero-DCE [33] and SCI [50]; however, TorchAdapt still surpasses them, and the low-light object detection method FeatEnHancer [26] by $+0.9$ and $+1.1$ in LL and LA settings, respectively. These results affirm the efficacy of TorchAdapt for object detec + +tion across varying illumination conditions. We provide a qualitative comparison in Appendix B. + +# 4.2. Face Detection + +Settings. For face detection, we choose a different detector and adopt RetinaNet [41] to report results on the DARK FACE dataset [69, 80] in the LL setting and a combination of the DARK FACE and WIDER FACE [78] datasets in the LA setting. Following prior works [26], we resize images to a resolution of $1500 \times 1000$ pixels and follow the $1 \times$ schedule in MMDetection [9]. + +Results. Results in Table 2 reveal that TorchAdapt demonstrates superior performance on face detection under both LL and LA settings, outperforming the previous best method $(\mathrm{SCI} \ddagger [50])$ by $+2.1$ and $+4.3$ in mAP_50, respectively. The significant gains, especially in the challenging LA setting, highlight TorchAdapt's capacity to handle diverse illumination scenarios by adaptively modulating features based on content and illumination, thereby improving the robustness and accuracy of the baseline. We also compare TorchAdapt with recent domain adaptation methods [20, 48] using DSFD [36] in Sec. 4.7. + +# 4.3. Instance Segmentation + +Settings. For the instance segmentation task, we adopt RTMDet-Ins-tiny $[49]^2$ as our baseline model and resize images to $640\times 640$ pixels for training. Following Section 4.1, we use the COCO pre-trained weights of RTMDet-Ins-tiny for the LL setting and train the model from scratch for the LA setting (see Table 1). + +Results. As summarized in Table 2), TorchAdapt achieves SOTA SegAP performance in both low-light (LL) and light-agnostic (LA) settings. Under LL conditions, TorchAdapt attains a SegAP of 53.6, surpassing the previous best (FeatEnHancer†) by +1.2. In the LA setting, it achieves a SegAP of 36.6, slightly improving overFeatEnHancer† by +0.5. These results demonstrate TorchAdapt's effectiveness in enhancing instance segmentation across varying illumination by producing illumination-invariant features. + +# 4.4. Semantic Segmentation + +Settings. Consistent with prior works [26, 76], we adopt DeepLabV3+[11] as our baseline model and follow [26, 76] for the LL setting. For the LA evaluation (Table 1), we train DeepLabV3+ from scratch using MMSegmentation [14]. Results. As shown in Table 2, TorchAdapt achieves an mIoU of 50.8 in the LL setting, surpassing the previous best (FeatEnHancer†) by +0.7, and an mIoU of 76.3 in the LA setting, improving overFeatEnHancer† by +0.2. These results highlight TorchAdapt's effectiveness for semantic segmentation under varying illumination conditions. + +
#Torch(§ 3.1)Adapt(§ 3.2)Scaling(§ 3.2)Object Det.Instance Seg.
LLLALLLA
1×××78.862.446.033.9
2✓××79.162.951.134.8
3✓✓×79.963.553.336.2
4✓×✓79.562.751.434.1
5✓✓✓80.164.053.636.6
+ +Table 3. Contribution of each component. Highlighted settings are set as default. + +
Scale factor (κ)ODSS
LLLALLLA
279.963.753.436.4
380.164.053.636.6
479.963.353.236.2
579.561.153.334.5
+ +(a) Scale factor $(\kappa)$ in Eq. 3. + +
D(I)ODSS
LLLALLLA
ISP [16]79.461.747.629.9
D(I) in [48]79.663.153.236.3
Gamma curve80.164.053.636.6
+ +(b) $D(\mathbf{I})$ in $\S 3.3$ and Fig. 3. + +Table 4. Ablating design choices in TorchAdapt. Highlighted settings are set as default. OD and SS are object detection and semantic segmentation, respectively. + +# 4.5. Video Object Detection + +Settings. In addition to image-based tasks, we evaluate the generalizability of TorchAdapt on video data. We present the dataset details in Table 1. Other experimental settings follow [26], with additional implementation details provided in Appendix A. + +Results. We report the results of the challenging video object detection task in Table 2. While all LLIE methods perform poorly on this task, end-to-end training significantly boosts the performance of zero-reference methods [24, 50]. TorchAdapt, however, outperforms all methods under both LL and LA settings, achieving substantial gains of $+4.5$ and $+6.4$ , respectively. These improvements affirm TorchAdapt's robustness and effectiveness for video object detection across both low-light and light-agnostic settings. + +# 4.6. Ablation Studies + +We ablate the components and design choices of TorchAdapt in object detection (OD) and instance segmentation (IS) tasks under (Low-light) LL and (Light-Agnostic) LA settings, following the same implementation details in § 4.1 and § 4.3, respectively. We provide more ablations in Appendix C. + +Contribution of Each Component. Table 3 evaluates the impact of the Torch module, the Adapt module, and Illumination Scaling. Adding the Torch module alone (row 2) improves performance over the baseline with $+0.3\%$ in LL OD and $+5.1\%$ in LL IS, highlighting the benefit of feature enhancement in low-light vision. Introducing the Adapt module alongside Torch (row 3) yields significant improvements $(+0.8\% / +0.6\%)$ in LL/LA OD; $+2.2\% / +1.4\%$ in LL/LA IS), underscoring the importance of content-adaptive modulation. Employing scaling without the Adapt module (Row 4) offers minor gains in LL settings. However, it is detrimental in LA settings $(-0.2\% / -0.7\%)$ in OD/IS), indicating that content-based adaptation is more impactful than global scaling alone. The full TorchAdapt configuration, combining all three components (row 5), achieves the best results with + +![](images/e603b3cb4e53545917c1c2f92488686b35b8e2f107fa9d9a38ce8ccdbfad4390.jpg) +Figure 5. Visualizing backbone features with SCI [50] on LA object detection using YOLOv3. (a) SCI's official checkpoint as an enhancer. (b) $\mathrm{SCI}^{\dagger}$ trained end-to-end. (c) SCI integrated into our TorchAdapt framework as a Torch. SCI with TorchAdapt extracts more representative features in both low and well-lit images. + +
MethodmAP (%)
DSFD [36]16.1
CICony [32]18.4
SimMinMax [48]25.7
DAI-Net [20]‡28.0
DAI-Net+TorchAdapt31.9+3.9
+ ++1.3%/+1.6% in LL/LA OD and +7.6%/+2.5% in LL/LA IS over the baseline, confirming that these components complement each other and synergistically enhance performance in both LL and LA settings. + +Design Choices in TorchAdapt. Table 4 ablates important design choices in TorchAdapt. As shown in Table 4a, we empirically establish that when the scale factor $\gamma$ in Eq. 3 is computed with $\kappa = 3$ , it produces optimal performance. Results in Table 4b reveal that employing the Gamma Curve in $D(I)$ to generate $V_{2}$ in § 3.3 brings the biggest boost in performance across both LL and LA settings. + +# 4.7. Comparison with Domain-Adaptation Methods + +We evaluate TorchAdapt's effectiveness for zero-shot domain adaptive face detection on the DARK FACE dataset [80]. Following settings from [20], we train the detector only on the WIDER FACE dataset [78]. As shown in Table 5, TorchAdapt integrated with DAI-Net [20], the current best-performing method, achieves a significant improvement of $+3.9\%$ mAP, demonstrating its effectiveness in enabling robust zero-shot adaptation under diverse illumination conditions. + +# 5. Exploring TorchAdapt's Properties + +Model-Agnostic Generality. Experiments in § 4 confirm that TorchAdapt is a general-purpose module that can be integrated into any high-level vision task. Moreover, TorchAdapt as a framework is highly flexible: any zero-reference (unsu + +![](images/ee41f6e42f54df25a83aa563fd4d172a8a4f66b3f9ca9353912d2893c5abe8df.jpg) +I + +![](images/9b2a7897bd66a6e77f556447b7190a5df314b46bf3f6696eaef639d16f113a87.jpg) +TorchAdapt(I) + +![](images/25b5ed0c93f79064680e3cc7139d3d8c17adda2977e3ec18b15eff4f99e37679.jpg) +I +Figure 6. Illustrating low-light image Enhancement as an emergent property of TorchAdapt on the DARK FACE validation set. More results can be found in Appendix B. + +![](images/6d76615a5888f9954e5e83faa48066095eaedc649ff390d22ed8066855e23d58.jpg) +TorchAdapt(I) + +Table 5. Zero-shot performance comparison of TorchAdapt with state-of-the-art domain adaptation methods on DARK FACE [80] using DSFD [36]. Results with $\ddagger$ are reproduced using the original codebase. TorchAdapt enhances prior SOTA [20] by $+3.9\mathrm{mAP}$ , demonstrating its effectiveness in zero-shot domain adaptation. + +
MethodmAP50(LL)mAP50(LA)#Params.(M)Latency (ms)
Baseline [57]78.862.465.221.9
+TorchAdapt80.1 (+1.3)64.0 (+1.6)65.3 (+0.1)23.7 (+1.8)
+ +Table 6. Computational analysis. With YOLOv3 [57], TorchAdapt introduces minimal computational overhead while bringing stronger gains across both LL and LA settings. + +pervised) LLIE methods [33, 50, 85] can serve as the Torch component to improve performance under varying illuminations. To verify this, we adopt SCI [50] as the Torch module in TorchAdapt and visualize the learned backbone features in Fig.5(c). Using SCI within TorchAdapt yields highly representative backbone features for both low-light and well-lit images. Appendix C provides additional examples and quantitative analyses. + +Enhancement as an Emergent Property. Although TorchAdapt is designed to produce illumination-invariant features for high-level vision tasks without employing any explicit enhancement loss functions during training, we observe that it inherently enhances low-light images. We illustrate this effect in Fig. 6. This emergent behavior likely results from the pre-training objective of aligning representations between well-lit and low-light images (see § 3.3), which encourages the network to implicitly adjust low-light images toward well-lit representations. + +Computational Analysis. We assess the computational efficiency of TorchAdapt using YOLOv3 [57] on the object detection task. As shown in Table 6, integrating TorchAdapt incurs minimal additional computational cost while delivering significant performance improvements. This demonstrates that TorchAdapt is compatible with lightweight models and suitable for real-time applications. + +# 6. Conclusions + +In this work, we introduce TorchAdapt, a real-time, general-purpose, light-agnostic adaptive feature enhancement framework that boosts the performance of high-level vision tasks under varying illumination conditions. By integrating its modules with light-agnostic learning, TorchAdapt produces powerful semantic representations for both low-light and well-lit data. Extensive experiments on five representative visual perception tasks involving images and videos confirm its effectiveness. + +# 7. Acknowledgement + +This work was partially funded by the European Union under Horizon Europe, grant number 101135724 (LUMINOUS). + +# References + +[1] Muhamamd Zeshan Afzal, SK Aziz Ali, Didier Stricker, Peter Eisert, Anna Hilsmann, Daniel Perez-Marcos, Marco Bianchi, Sonia Crottaz-Herbette, Roberto De Ioris, Eleni Mangina, et al. Next generation xr systems-large language models meet augmented and virtual reality. IEEE computer graphics and applications, 2025. 1 +[2] Tadahiro Azetsu and Noriaki Suetake. Hue-preserving image enhancement in cielab color space considering color gamut. Optical Review, 26:283-294, 2019. 5 +[3] Ruzena Bajcsy, Yiannis Aloimonos, and John K Tsotsos. Revisiting active perception. Autonomous Robots, 42:177-196, 2018. 1, 2 +[4] Peter J Bickel and Kjell A Doksum. Mathematical statistics: basic ideas and selected topics, volumes I-II package. Chapman and Hall/CRC, 2015. 1 +[5] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Fredo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. In CVPR 2011, pages 97-104. IEEE, 2011. 2 +[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2 +[7] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE international conference on computer vision, pages 2722-2730, 2015. 1, 2 +[8] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3291-3300, 2018. 2 +[9] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7 +[10] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. International Journal of Computer Vision, 131(8):2198-2218, 2023. 2, 3, 5 +[11] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 7 +[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF + +conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2 +[13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1 +[14] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.7 +[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 2, 5 +[16] Ziteng Cui, Guo-Jun Qi, Lin Gu, Shaodi You, Zenghui Zhang, and Tatsuya Harada. Multitask aet with orthogonal tangent regularity for dark object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2553-2562, 2021. 1, 2, 3, 5, 6, 7 +[17] Ziteng Cui, Kunchang Li, Lin Gu, Shenghan Su, Peng Gao, Zhengkai Jiang, Yu Qiao, and Tatsuya Harada. You only need 90k parameters to adapt light: a light weight transformer for image enhancement and exposure correction. arXiv preprint arXiv:2205.14871, 2022. 1, 6 +[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 3, 5 +[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2, 5 +[20] Zhipeng Du, Miaojing Shi, and Jiankang Deng. Boosting object detection with zero-shot day-night domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12666-12676, 2024. 1, 2, 3, 6, 7, 8 +[21] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 2, 5 +[22] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 3 +[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 3, 5 + +[24] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. CoRR, abs/2001.06826, 2020. 1, 2, 4, 6, 7 +[25] Jiang Hai, Zhu Xuan, Ren Yang, Yutong Hao, Fengzhu Zou, Fang Lin, and Songchen Han. R2rnet: Low-light image enhancement via real-low to real-normal network. Journal of Visual Communication and Image Representation, 90:103712, 2023. 2 +[26] Khurram Azeem Hashmi, Goutham Kallempudi, Didier Stricker, and Muhammad Zeshan Afzal. Featenhancer: Enhancing hierarchical features for object detection and beyond under low-light vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6725-6735, 2023. 1, 2, 3, 4, 6, 7 +[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 2, 5 +[28] Mingbo Hong, Shen Cheng, Haibin Huang, Haoqiang Fan, and Shuaicheng Liu. You only look around: Learning illumination invariant feature for low-light object detection. arXiv preprint arXiv:2410.18398, 2024. 6 +[29] Sergey Ioffe. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. 4 +[30] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhancement without paired supervision. IEEE Transactions on Image Processing, 30: 2340-2349, 2021. 2 +[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1 +[32] Attila Lengyel, Sourav Garg, Michael Milford, and Jan C. van Gemert. Zero-shot day-night domain adaptation with a physics prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4399-4409, 2021. 1, 2, 3, 6, 8 +[33] Chongyi Li, Chunle Guo Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 4, 6, 8 +[34] Chongyi Li, Chunle Guo, Shangchen Zhou, Qiming Ai, Ruicheng Feng, and Chen Change Loy. Flexicurve: Flexible piecewise curves estimation for photo retouching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1092-1101, 2023. 2 +[35] Hebei Li, Jin Wang, Jiahui Yuan, Yue Li, Wenming Weng, Yansong Peng, Yueyi Zhang, Zhiwei Xiong, and Xiaoyan Sun. Event-assisted low-light video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3250-3259, 2024. 3 +[36] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd: + +dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 7, 8 +[37] Jinlong Li, Baolu Li, Zhengzhong Tu, Xinyu Liu, Qing Guo, Felix Juefei-Xu, Runsheng Xu, and Hongkai Yu. Light the night: A multi-condition diffusion framework for unpaired low-light enhancement in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15205-15215, 2024. 2 +[38] Yunan Li, Yihao Zhang, Shoude Li, Long Tian, Dou Quan, Chaoneng Li, and Qiguang Miao. Watching it in dark: A target-aware representation learning framework for high-level vision tasks in low illumination. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 3, 6 +[39] Jinxiu Liang, Jingwen Wang, Yuhui Quan, Tianyi Chen, Jiaying Liu, Haibin Ling, and Yong Xu. Recurrent exposure generation for low-light face detection. IEEE Transactions on Multimedia, 24:1609-1621, 2022. 3 +[40] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312, 2014. 1, 2, 3, 5, 6 +[41] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. CoRR, abs/1708.02002, 2017. 7 +[42] Hewei Liu, Renwei Yang, Shuyuan Zhu, Xing Wen, and Bing Zeng. Luminance-guided chrominance image enhancement for hevc intra coding. In 2022 IEEE International Symposium on Circuits and Systems (ISCAS), pages 3180-3184. IEEE, 2022. 5 +[43] Jiaying Liu, Dejia Xu, Wenhan Yang, Minhao Fan, and Haofeng Huang. Benchmarking low-light image enhancement and beyond. International Journal of Computer Vision, 129:1153-1184, 2021. 2 +[44] Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10561-10570, 2021. 1, 2 +[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 4 +[46] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12009-12019, 2022. 4 +[47] Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. CoRR, abs/1805.11227, 2018. 1, 2, 3, 5, 6 +[48] Rundong Luo, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Similarity min-max: Zero-shot day-night domain adaptation. + +In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8104-8114, 2023. 1, 2, 3, 5, 6, 7, 8 +[49] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 7 +[50] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5637-5646, 2022. 1, 2, 4, 6, 7, 8 +[51] Tengyu Ma, Long Ma, Xin Fan, Zhongxuan Luo, and Risheng Liu. PIA: parallel architecture with illumination accelerator for joint enhancement and detection in low-light. In MM '22: The 30th ACM International Conference on Multimedia, Lisboa, Portugal, October 10 - 14, 2022, pages 2070-2078. ACM, 2022. 1, 3 +[52] Igor Morawski, Yu-An Chen, Yu-Sheng Lin, and Winston H. Hsu. NOD: taking a closer look at detection under extreme low-light conditions with night object detection dataset. CoRR, abs/2110.10364, 2021. 3, 5 +[53] Lukáš Neumann, Michelle Karg, Shanshan Zhang, Christian Scharfenberger, Eric Piegert, Sarah Mistr, Olga Prokofyeva, Robert Thiel, Andrea Vedaldi, Andrew Zisserman, et al. Nightowls: A pedestrians at night dataset. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part I 14, pages 691–705. Springer, 2019. 3 +[54] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 3 +[55] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2 +[56] Qingpao Qin, Kan Chang, Mengyuan Huang, and Guiqing Li. Denet: Detection-driven enhancement network for object detection under adverse weather conditions. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 2813-2829, 2022. 3 +[57] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. CoRR, abs/1804.02767, 2018. 6, 8 +[58] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 1, 2, 5 +[59] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Guided curriculum model adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7374-7383, 2019. 1, 2 +[60] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the + +IEEE/CVF International Conference on Computer Vision (ICCV), pages 10765-10775, 2021. 1, 2, 3, 5 +[61] Yukihiro Sasagawa and Hajime Nagahara. Yolo in the dark - domain adaptation method for merging multiple models. In Computer Vision – ECCV 2020, pages 345–359, Cham, 2020. Springer International Publishing. 3 +[62] Shangquan Sun, Wenqi Ren, Tao Wang, and Xiaochun Cao. Rethinking image restoration for object detection. Advances in Neural Information Processing Systems, 35:4461-4474, 2022. 3 +[63] Maria Valera and Sergio A Velastin. Intelligent distributed surveillance systems: a review. IEE Proceedings-Vision, Image and Signal Processing, 152(2):192-204, 2005. 1, 2 +[64] Wenjing Wang, Wenhan Yang, and Jiaying Liu. Hla-face: Joint high-low adaptation for low light face detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16195-16204, 2021. 3 +[65] Wenjing Wang, Zhengbo Xu, Haofeng Huang, and Jiaying Liu. Self-aligned concave curve: Illumination enhancement for unsupervised adaptation. In Proceedings of the 30th ACM international conference on multimedia, pages 2617-2626, 2022. 2, 3, 5 +[66] Wenjing Wang, Xinhao Wang, Wenhan Yang, and Jiaying Liu. Unsupervised face detection in the dark. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1250-1266, 2023. 3 +[67] Wenjing Wang, Huan Yang, Jianlong Fu, and Jiaying Liu. Zero-reference low-light enhancement via physical quadruple priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26057-26066, 2024. 2 +[68] Xinzhe Wang, Kang Ma, Qiankun Liu, Yunhao Zou, and Ying Fu. Multi-object tracking in the dark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 382-392, 2024. 3 +[69] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. CoRR, abs/1808.04560, 2018. 7 +[70] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2 +[71] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wenhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5901-5910, 2022. 1, 6 +[72] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15769-15778, 2021. 1, 2 +[73] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17714-17724, 2022. 2 + +[74] Xiaogang Xu, Ruixing Wang, and Jiangbo Lu. Low-light image enhancement via structure modeling and guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9893-9903, 2023. 2 +[75] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Arid: A new dataset for recognizing action in the dark. In Deep Learning for Human Activity Recognition: Second International Workshop, DL-HAR 2020, Held in Conjunction with IJCAI-PRICAI 2020, Kyoto, Japan, January 8, 2021, Proceedings 2, pages 70–84. Springer, 2021. 3 +[76] Xinwei Xue, Jia He, Long Ma, Yi Wang, Xin Fan, and Risheng Liu. Best of both worlds: See and understand clearly in the dark. In Proceedings of the 30th ACM International Conference on Multimedia, page 2154–2162, New York, NY, USA, 2022. Association for Computing Machinery. 6, 7 +[77] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1 +[78] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5525-5533, 2016. 1, 2, 3, 5, 7, 8 +[79] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12918-12927, 2023. 2 +[80] Wenhan Yang, Ye Yuan, Wenqi Ren, Jiaying Liu, Walter J. Scheirer, Zhangyang Wang, Taiheng Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, Yuqiang Zheng, Yanyun Qu, Yuhong Xie, Liang Chen, Zhonghao Li, Chen Hong, Hao Jiang, Siyuan Yang, Yan Liu, Xiaochao Qu, Pengfei Wan, Shuai Zheng, Minhui Zhong, Taiyi Su, Lingzhi He, Yandong Guo, Yao Zhao, Zhenfeng Zhu, Jinxiu Liang, Jingwen Wang, Tianyi Chen, Yuhui Quan, Yong Xu, Bo Liu, Xin Liu, Qi Sun, Tingyu Lin, Xiaochuan Li, Feng Lu, Lin Gu, Shengdi Zhou, Cong Cao, Shifeng Zhang, Cheng Chi, Chubing Zhuang, Zhen Lei, Stan Z. Li, Shizheng Wang, Ruizhe Liu, Dong Yi, Zheming Zuo, Jianning Chi, Huan Wang, Kai Wang, Yixiu Liu, Xingyu Gao, Zhenyu Chen, Chang Guo, Yongzhou Li, Huicai Zhong, Jing Huang, Heng Guo, Jianfei Yang, Wenjuan Liao, Jiangang Yang, Liguo Zhou, Mingyue Feng, and Likun Qin. Advancing image understanding in poor visibility environments: A collective benchmark study. IEEE Transactions on Image Processing, 29:5737-5752, 2020. 1, 2, 3, 5, 7, 8 +[81] Bo Zhang, Yuchen Guo, Runzhao Yang, Zhihong Zhang, Jiayi Xie, Jinli Suo, and Qionghai Dai. Darkvision: A benchmark for low-light image/video perception, 2023. 1, 2, 3, 5 +[82] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on Learning Representations, 2023. 1 +[83] Liheng Zhang, Guo-Jun Qi, Liqiang Wang, and Jiebo Luo. Aet vs. aed: Unsupervised representation learning by auto + +encoding transformations rather than data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2547-2555, 2019. 3 +[84] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. CoRR, abs/1905.04161, 2019. 1, 2, 6 +[85] Zhaoyang Zhang, Yitong Jiang, Jun Jiang, Xiaogang Wang, Ping Luo, and Jinwei Gu. Star: A structure-aware lightweight transformer for real-time image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4106-4115, 2021. 1, 8 +[86] Zhao Zhang, Huan Zheng, Richang Hong, Mingliang Xu, Shuicheng Yan, and Meng Wang. Deep color consistent network for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1899-1908, 2022. 2 +[87] Hegui Zhu, Kai Wang, Ziwei Zhang, Yuelin Liu, and Wuming Jiang. Low-light image enhancement network with decomposition and adaptive information fusion. Neural Computing and Applications, 34(10):7733-7748, 2022. 5 \ No newline at end of file diff --git a/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/images.zip b/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b3ec27f95164f7e173dc532d99783bd92be0cb9f --- /dev/null +++ b/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7a8ba8459981fb8fffb902bebf0543d6415dacb88822fa4d1d50919757d8cce +size 443439 diff --git a/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/layout.json b/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e3c43acaac1bdc6d4ac917ed2cb876354e023e22 --- /dev/null +++ b/2025/TorchAdapt_ Towards Light-Agnostic Real-Time Visual Perception/layout.json @@ -0,0 +1,11226 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 512, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 512, + 121 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 512, + 121 + ], + "type": "text", + "content": "TorchAdapt: Towards Light-Agnostic Real-Time Visual Perception" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 142, + 233, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 142, + 233, + 156 + ], + "spans": [ + { + "bbox": [ + 99, + 142, + 233, + 156 + ], + "type": "text", + "content": "Khurram Azeem Hashmi" + }, + { + "bbox": [ + 99, + 142, + 233, + 156 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 262, + 143, + 397, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 143, + 397, + 156 + ], + "spans": [ + { + "bbox": [ + 262, + 143, + 397, + 156 + ], + "type": "text", + "content": "Karthik Palyakere Suresh1,2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 236, + 157, + 374, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 157, + 374, + 170 + ], + "spans": [ + { + "bbox": [ + 236, + 157, + 374, + 170 + ], + "type": "text", + "content": "Muhammad Zeshan Afzal1,2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 182, + 171, + 216, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 171, + 216, + 183 + ], + "spans": [ + { + "bbox": [ + 182, + 171, + 216, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 182, + 171, + 216, + 183 + ], + "type": "text", + "content": "DFKI" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 280, + 171, + 429, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 280, + 171, + 429, + 184 + ], + "spans": [ + { + "bbox": [ + 280, + 171, + 429, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 280, + 171, + 429, + 184 + ], + "type": "text", + "content": "RPTU Kaiserslautern-Landau" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 228, + 188, + 379, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 188, + 379, + 198 + ], + "spans": [ + { + "bbox": [ + 228, + 188, + 379, + 198 + ], + "type": "text", + "content": "khurram_azeem Hashmi@dfki.de" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 425, + 143, + 510, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 143, + 510, + 156 + ], + "spans": [ + { + "bbox": [ + 425, + 143, + 510, + 156 + ], + "type": "text", + "content": "Didier Stricker1,2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 251, + 297, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 297, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 297, + 526 + ], + "type": "text", + "content": "Low-light conditions significantly degrade the performance of high-level vision tasks. Existing approaches either enhance low-light images without considering normal illumination scenarios, leading to poor generalization, or are tailored to specific tasks. We propose TorchAdapt, a real-time adaptive feature enhancement framework that generalizes robustly across varying illumination conditions without degrading performance in well-lit scenarios. TorchAdapt consists of two complementary modules: the Torch module enhances semantic features beneficial for downstream tasks, while the Adapt module dynamically modulates these enhancements based on input content. Leveraging a novel light-agnostic learning strategy, TorchAdapt aligns feature representations of enhanced and well-lit images to produce powerful illumination-invariant features. Extensive experiments on multiple high-level vision tasks, including object detection, face detection, instance segmentation, semantic segmentation, and video object detection, demonstrate that TorchAdapt consistently outperforms state-of-the-art low-light enhancement and task-specific methods in both low-light and light-agnostic settings. TorchAdapt thus provides a unified, flexible solution for robust visual perception across diverse lighting conditions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 550, + 135, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 550, + 135, + 562 + ], + "spans": [ + { + "bbox": [ + 56, + 550, + 135, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 570, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 665 + ], + "type": "text", + "content": "Recent advances in visual perception tasks [13, 31, 77, 82] are highly effective at extracting features for complex tasks when provided with high-quality inputs. However, replicating this performance is challenging when data is scarce, noisy, or expensive to obtain. A common issue is varying illumination, particularly low-light conditions, which hinders the robustness and reliability of safety-critical visual perception systems [1, 3, 7, 63]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 666, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 297, + 714 + ], + "type": "text", + "content": "A common strategy to improve low-light performance involves pre-processing images with Low-Light Image Enhancement (LLIE) methods [17, 24, 33, 44, 50, 51, 71, 84, 85] before feeding them into high-level vision mod" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 315, + 225, + 382, + 319 + ], + "blocks": [ + { + "bbox": [ + 315, + 225, + 382, + 319 + ], + "lines": [ + { + "bbox": [ + 315, + 225, + 382, + 319 + ], + "spans": [ + { + "bbox": [ + 315, + 225, + 382, + 319 + ], + "type": "image", + "image_path": "fdef7c1f1dd5cc82fa6f742b32d83cd361a3634712105dbd6ce9d14c33371195.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 383, + 226, + 451, + 319 + ], + "blocks": [ + { + "bbox": [ + 383, + 226, + 451, + 319 + ], + "lines": [ + { + "bbox": [ + 383, + 226, + 451, + 319 + ], + "spans": [ + { + "bbox": [ + 383, + 226, + 451, + 319 + ], + "type": "image", + "image_path": "7501f044036cdd8bc76406a59833e2f05e74cd09ccee5d765a4f21363bea0a95.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 319, + 466, + 327 + ], + "lines": [ + { + "bbox": [ + 410, + 319, + 466, + 327 + ], + "spans": [ + { + "bbox": [ + 410, + 319, + 466, + 327 + ], + "type": "text", + "content": "(a) Low-Light Vision" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 455, + 226, + 553, + 317 + ], + "blocks": [ + { + "bbox": [ + 455, + 226, + 553, + 317 + ], + "lines": [ + { + "bbox": [ + 455, + 226, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 455, + 226, + 553, + 317 + ], + "type": "image", + "image_path": "3158cf49653bf7978afeaf8384958c3197e9d7d59141b96ad775fa98e5129ce4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 316, + 327, + 380, + 420 + ], + "blocks": [ + { + "bbox": [ + 316, + 327, + 380, + 420 + ], + "lines": [ + { + "bbox": [ + 316, + 327, + 380, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 380, + 420 + ], + "type": "image", + "image_path": "393d6a987f61d20407ff9cce26d641509d477f9781212f06f5c3bc4f50e7bef8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 438, + 556, + 536 + ], + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 536 + ], + "type": "text", + "content": "Figure 1. Prior downstream methods in low-light vision assume that all samples belong to low-light conditions as shown in (a), leading to underperformance in varying illumination environments, depicted in (b). TorchAdapt alleviates this bottleneck via learning powerful illumination-invariant features and achieves consistent gains across low-light and light-agnostic settings on Object Detection (OD), Face Detection (FD), Instance Segmentation (IS), Semantic Segmentation (SS), and Video Object Detection (VOD). Best viewed on the screen." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 380, + 328, + 449, + 420 + ], + "blocks": [ + { + "bbox": [ + 380, + 328, + 449, + 420 + ], + "lines": [ + { + "bbox": [ + 380, + 328, + 449, + 420 + ], + "spans": [ + { + "bbox": [ + 380, + 328, + 449, + 420 + ], + "type": "image", + "image_path": "8938b69d45be7c366a1c6a63ef8e45639c0e7021cb67de8f76069e68c29cb623.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 421, + 470, + 430 + ], + "lines": [ + { + "bbox": [ + 402, + 421, + 470, + 430 + ], + "spans": [ + { + "bbox": [ + 402, + 421, + 470, + 430 + ], + "type": "text", + "content": "(b) Light-Agnostic Vision" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 455, + 328, + 553, + 418 + ], + "blocks": [ + { + "bbox": [ + 455, + 328, + 553, + 418 + ], + "lines": [ + { + "bbox": [ + 455, + 328, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 455, + 328, + 553, + 418 + ], + "type": "image", + "image_path": "0895c393afe8577eb301659288cba0e1da35a2003dc2cb84ff8c7c41ce47946d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 555, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 662 + ], + "type": "text", + "content": "els. However, this simple integration often leads to weaker gains [16, 26], as these arts conform to human visual perception (optimized on pixel-level objectives like L1 or MSE losses [4]) instead of machine visual perception. Moreover, applying these enhancement methods indiscriminately can degrade the performance on images/videos captured under standard lighting conditions, as alterations intended for low-light scenarios [47, 60, 80, 81] may be unnecessary or even detrimental in well-lit environments [15, 40, 58, 78]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "To address low-light vision tasks, recent methods employ transfer learning [16, 59, 72] and domain adaptation techniques [20, 32, 48]. Despite improvements, these efforts are carefully designed for a single high-level vision task, such" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5645" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 240 + ], + "type": "text", + "content": "as object detection [16] or semantic segmentation [59, 72], and require cumbersome multi-stage training to adapt to each dense task [20, 32, 48]. Moreover, they often assume consistently low-light data, leading to performance degradation on well-lit images and videos, sometimes worse than baselines (see Table 2). In practical applications with unpredictable lighting—such as autonomous driving [7, 37] at different times of the day, surveillance systems [63] monitoring both indoor and outdoor environments or robots [3] operating between well-lit and low-lit areas—these limitations become critical. This highlights the need for high-level vision models to be light-agnostic—capable of maintaining high performance across varying lighting conditions, as visualized in Fig. 1(b)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 242, + 297, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 297, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 297, + 361 + ], + "type": "text", + "content": "In this work, we propose TorchAdapt, a real-time, light-agnostic adaptive feature enhancement framework designed to improve the performance of any high-level vision task under varying illumination. TorchAdapt comprises two primary modules: Torch and Adapt, illustrated in Fig. 2. The Torch module is a lightweight feature enhancement network that enriches semantic representations favorable for downstream vision tasks. The Adapt module learns to modulate this feature enhancement dynamically based on input data, amplifying or reducing the effect of the Torch module." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "text", + "content": "TorchAdapt builds on the consensus [26, 38] that low-to-normal light adaptation lacks a physical model, leading to unexpected artifacts during optimization when transforming images in pixel space. Therefore, aligning representations in the feature space is a more practical approach for high-level vision tasks [48, 65]. We optimize TorchAdapt using a non-contrastive self-supervised learning objective [12, 23] to align representations from low-light and well-lit images. However, unlike prior works operating in a similar direction [16, 48, 65], TorchAdapt also learns not to adversely affect representations of well-lit images. As illustrated in Fig. 3, we take a well-lit image " + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "text", + "content": " and its synthetically generated low-light counterpart " + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "inline_equation", + "content": "V_{2}" + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "text", + "content": ", pass them through TorchAdapt to obtain enhanced images " + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "inline_equation", + "content": "\\hat{V}_{1}" + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "inline_equation", + "content": "\\hat{V}_{2}" + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "text", + "content": ", and feed them along with " + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "text", + "content": " into a pre-trained frozen visual encoder [19, 27, 55] to generate representations. Since the encoder is frozen, only TorchAdapt is optimized to help the encoder produce well-lit representations (like " + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "text", + "content": ") for " + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "inline_equation", + "content": "\\hat{V}_{1}" + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "inline_equation", + "content": "\\hat{V}_{2}" + }, + { + "bbox": [ + 55, + 364, + 297, + 579 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 297, + 715 + ], + "type": "text", + "content": "Thanks to the conceptual simplicity, TorchAdapt enjoys two benefits. ① It can be trained with a simple self-supervised loss on a general-purpose dataset like ImageNet [18], circumventing large-scale real-world low-light datasets. ② The training of TorchAdapt is independent of the high-level vision task and needs to be done only once to align with the backbone networks [6, 19, 27]. Later, TorchAdapt can be plugged into any downstream tasks in a light-agnostic manner. We demonstrate the effectiveness of TorchAdapt on five representative visual perception tasks, including object detection [21, 47], face detection [78, 80], semantic" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 556, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 167 + ], + "type": "text", + "content": "segmentation [15, 60], instance segmentation [10, 40], and video object detection [58, 81] under both low-light and light-agnostic settings. For all tasks, TorchAdapt brings consistent and significant gains over baselines under both low-light and light-agnostic settings, outperforming previous state-of-the-art task-specific and LLIE methods (see Fig. 1). Moreover, TorchAdapt accomplishes all this with negligible computational overhead (see Table 6)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 169, + 557, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 557, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 557, + 410 + ], + "type": "text", + "content": "Contributions. Method: We propose TorchAdapt, a real-time, light-agnostic adaptive feature enhancement framework that improves the performance of any high-level vision task under varying illumination without compromising accuracy in well-lit conditions. Simplicity: Thanks to its architectural innovations and self-supervised objective, TorchAdapt significantly simplifies the training strategy for any high-level vision task pipeline, circumventing corresponding paired images/videos, as shown in Fig 3. Flexibility: TorchAdapt is independent of vision encoders and downstream task architectures, demonstrating model-agnostic generality (see § 5). Any zero-reference LLIE method [24, 33, 50, 67] can serve as the Torch module in our framework to achieve light-agnostic performance (see Fig. 5). Results: We conduct extensive experiments on five visual perception tasks involving images and videos under both low-light and light-agnostic settings using curated datasets. TorchAdapt outperforms state-of-the-art task-specific and LLIE methods across all settings when integrated with the same baselines (see Fig. 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 422, + 403, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 403, + 435 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 403, + 435 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 443, + 477, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 477, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 477, + 456 + ], + "type": "text", + "content": "2.1. Enhancing Low-Light Images" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 715 + ], + "type": "text", + "content": "Low-light Image Enhancement (LLIE) aims to improve the visual quality of images captured under insufficient lighting conditions. Modern learning-based approaches have become prevalent, thanks to the emergence of low-light datasets [5, 8, 25, 43, 70]. Supervised methods [70, 73, 74, 84, 86] leverage these datasets to design effective networks trained to predict normal-light images from low-light ones. Unsupervised LLIE approaches [30, 34, 79] employ adversarial learning from unpaired supervision to relax the requirement of paired normal-low light data. However, all of these works still rely on specific training data, which limits their generalizability to unseen scenarios. Another line of work known as zero-reference methods [24, 33, 44, 50, 67] get rid of both paired and unpaired data to enhance low-light images. These arts design enhancement networks that optimize on a set of non-reference loss functions. Since these works are mainly designed to conform to human visual perception, they bring sub-optimal gains for machine visual perception tasks. Furthermore, they are not light-agnostic (capable of handling both normal and low-light images). This work bridges this gap by complementing these zero-reference methods," + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 317, + 757 + ], + "type": "text", + "content": "5646" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 70, + 509, + 173 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 509, + 173 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 509, + 173 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 509, + 173 + ], + "type": "image", + "image_path": "6602a677e24ffe63a76d0f3c9d0e5d66f941f01870e4ff0cd1d6c44b1dca1f92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 182, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 54, + 182, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 54, + 182, + 555, + 240 + ], + "type": "text", + "content": "Figure 2. TorchAdapt architectural overview. The Torch module enhances feature representation by applying " + }, + { + "bbox": [ + 54, + 182, + 555, + 240 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 54, + 182, + 555, + 240 + ], + "type": "text", + "content": " convolutions and element-wise operations (Eq. 1). The Adapt module performs adaptive modulation, creating a content-aware response based on the input " + }, + { + "bbox": [ + 54, + 182, + 555, + 240 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 182, + 555, + 240 + ], + "type": "text", + "content": " (Eq. 2). Both modules are fused via Adaptive Fusion, integrating feature enhancements from Torch and contextual adaptations from Adapt. This adaptive fusion produces illumination-invariant features, improving robustness across varying lighting conditions. Adaptive Fusion is explained in Eq. 4." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 249, + 297, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 297, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 297, + 275 + ], + "type": "text", + "content": "where they can be employed as a Torch in our TorchAdapt framework to improve light-agnostic visual perception." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 214, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 214, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 214, + 298 + ], + "type": "text", + "content": "2.2. Low-Light Visual Perception" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 304, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 327 + ], + "type": "text", + "content": "This direction of work considers machine perception as the criteria for success while enhancing low-light images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 329, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 544 + ], + "type": "text", + "content": "Low-light Object Detection (LLOD). Object detection, as a fundamental problem in computer vision, has also been well-explored in low-light vision [16, 20, 39, 51, 61, 62, 64-66]. Based on the dataset, objects vary and can be generic [47, 52] or specific like pedestrians [53], and human faces [80]. One category of LLOD methods [51, 56, 62, 65] jointly learns the enhancement and detection task in an end-to-end fashion to improve the overall detection performance. Other approaches involve learning a low-light detector through fusing multiple models [61] or resorting to domain adaptation frameworks [16, 20, 64], minimizing the distribution variance between normal and low-light images. Most of these works are specifically designed to tackle the dark object detection task. Furthermore, when these methods are optimized for LLOD, they underperform on normal-light images. In contrast, TorchAdapt is fully task-agnostic for any high-level vision task, and it does not worsen the performance of the detector on normal-light data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 298, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 298, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 298, + 713 + ], + "type": "text", + "content": "Low-light Tasks-Agnostic Methods. Recently, various low-light task-agnostic methods [20, 26, 32, 48, 65] have emerged in the literature. Unlike prior approaches, they are more flexible and can be integrated to improve any downstream task under low-light vision. Our work follows the same spirit and offers similar flexibility. FeatEnHancer [26], as one of the representative works, learns to enhance the hierarchical features suitable for downstream tasks. Other seminal attempts [20, 32, 48] treat low-light perception tasks as domain adaption (DA) learning. They exploit well-lit source domain data [40, 78] and learn to adapt to low-light target domain data [47, 80]. Although flexible, the DA methods rely on multi-stage training frameworks for each task [48]. In contrast, our TorchAdapt is pre-trained with the frozen back" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 249, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 308 + ], + "type": "text", + "content": "bone on the domain-independent data like ImageNet [18]. Moreover, it is important to recognize the efforts of recent real low-light datasets [10, 35, 60, 68, 75, 81] that enable the increasing research on several downstream tasks in low-light vision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 322, + 555, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 347 + ], + "type": "text", + "content": "2.3. Learning Equivariant Representations in LowLight Vision with Self-Supervised Objectives" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 353, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 556, + 640 + ], + "type": "text", + "content": "Annotating low-light vision datasets is a much more difficult task than annotating well-lit datasets. Hence, it is a common practice to apply self-supervised learning to learn illumination-invariant features [16, 65] or aligning well-lit and low-light data distributions [38, 48, 66]. MAET [16] adopts the autoencoding transformation [83] to decode illumination-degrading parameters and detection predictions. SACC [65] merges the pretext tasks of rotation prediction [22] and jigsaw puzzling [54] to learn illumination enhancement from low-light images. Most related to our work is SimMinMax [48], which adopts the BYOL [23] framework to learn model-level adaptation by maximizing the similarity of features between images at night and daytime. However, since SimMinMax [48] learns to adapt from daytime to nighttime images, it needs to be trained for each high-level vision task, requiring the corresponding daytime labeled dataset. In contrast, TorchAdapt leverages self-supervised objectives to align backbone features during the pretext task. Furthermore, it can be trained on any general-purpose dataset. This significantly simplifies the training procedure of TorchAdapt, as compared to SimMinMax, on any high-level vision task. TorchAdapt is also unique to prior works in its ability to handle light-agnostic data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 655, + 392, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 655, + 392, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 655, + 392, + 670 + ], + "type": "text", + "content": "3. TorchAdapt" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "This section formally introduces TorchAdapt, explaining the proposed Torch in § 3.1 and Adapt in § 3.2 modules, followed by the light-agnostic learning in § 3.3." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5647" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 76, + 286, + 199 + ], + "blocks": [ + { + "bbox": [ + 73, + 76, + 286, + 199 + ], + "lines": [ + { + "bbox": [ + 73, + 76, + 286, + 199 + ], + "spans": [ + { + "bbox": [ + 73, + 76, + 286, + 199 + ], + "type": "image", + "image_path": "ed71c641c4623764435fb269b0b8cb1c6f6bfd6a39089703dc0d05e1aa566bb1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "text", + "content": "Figure 3. Light-Agnostic learning pipeline. " + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "inline_equation", + "content": "D(.)" + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "text", + "content": " is a darkening function applied to the well-lit image " + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "text", + "content": " to synthesize the low-light view " + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "inline_equation", + "content": "V_{2}" + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "text", + "content": ". Both " + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "inline_equation", + "content": "V_{2}" + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "text", + "content": " are processed by TorchAdapt and then fed into a frozen pre-trained vision encoder " + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "text", + "content": ". A cosine similarity loss optimizes TorchAdapt to produce well-lit representations for both " + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "inline_equation", + "content": "V_{2}" + }, + { + "bbox": [ + 55, + 213, + 296, + 302 + ], + "type": "text", + "content": ", enabling it to learn illumination-invariant features. The TorchAdapt(.) function integrates Eq. 1 and Eq. 4. Refer to § 3.3 for a complete explanation of each symbol." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 310, + 106, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 106, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 106, + 321 + ], + "type": "text", + "content": "3.1.Torch" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 328, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 296, + 448 + ], + "type": "text", + "content": "The Torch module in our TorchAdapt framework serves as an enhancement network as in Low-light Image Enhancement (LLIE) approaches [24, 33, 50]. However, unlike these methods—which focus on improving visual quality for human perception—our Torch module aims to boost semantic representations favorable for downstream vision tasks under varying lighting conditions. Given an input image " + }, + { + "bbox": [ + 55, + 328, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 328, + 296, + 448 + ], + "type": "text", + "content": ", the Torch module computes an enhancement map " + }, + { + "bbox": [ + 55, + 328, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\theta}(\\mathbf{I})" + }, + { + "bbox": [ + 55, + 328, + 296, + 448 + ], + "type": "text", + "content": ", parameterized by network weights " + }, + { + "bbox": [ + 55, + 328, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 328, + 296, + 448 + ], + "type": "text", + "content": ". The transformed image " + }, + { + "bbox": [ + 55, + 328, + 296, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 328, + 296, + 448 + ], + "type": "text", + "content": " is obtained by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 460, + 296, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 460, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 60, + 460, + 296, + 473 + ], + "type": "interline_equation", + "content": "\\mathbf {T} = \\mathbf {I} \\odot (1 + \\mathcal {H} _ {\\theta} (\\mathbf {I})) \\text {, w h e r e} \\mathcal {H} _ {\\theta} (\\mathbf {I}) = \\tanh (\\mathcal {F} (\\mathbf {I}; \\theta)) \\tag {1}", + "image_path": "13f2998343bc96fa2cfe2e67068a59a9fd81a2d1ae70780f7de4217d0a81f86e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "text", + "content": " denotes element-wise multiplication. The " + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tanh(\\cdot)" + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "text", + "content": " activation ensures that the adjustment values are in the range " + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "text", + "content": ", allowing the model to both amplify (positive values) and attenuate (negative values) features. This formulation enables the model to adapt to the image's content effectively. Torch network. The Torch network " + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathbf{I};\\theta)" + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "text", + "content": " is a lightweight convolutional network composed of an initial feature extraction layer followed by a residual block, as illustrated in Fig. 2. The residual learning enables the Torch module to focus on learning the necessary adjustments to improve the input image rather than reconstructing it entirely. Enhancing semantically important features while preserving original content makes the Torch module crucial in maintaining and improving performance in downstream vision tasks. Notably, thanks to the formulation explained in Eq 1, our Torch module only has 11,075 trainable parameters and requires 0.712 GFLOPs for " + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 3" + }, + { + "bbox": [ + 55, + 485, + 296, + 713 + ], + "type": "text", + "content": ", making it suitable for real-time high-level vision tasks under varying lighting conditions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 367, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 367, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 367, + 84 + ], + "type": "text", + "content": "3.2. Adapt" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 89, + 555, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 220 + ], + "type": "text", + "content": "While the Torch module enhances the input image " + }, + { + "bbox": [ + 313, + 89, + 555, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 89, + 555, + 220 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 313, + 89, + 555, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 89, + 555, + 220 + ], + "type": "text", + "content": ", applying this enhancement uniformly may not be optimal. Different regions within an image require varying degrees of enhancement based on content (low/well-lit images) and context (downstream task). To address this, we introduce the Adapt module, which learns a content-adaptive modulation matrix in the feature space. This allows the model to selectively amplify or suppress the enhancements produced by the Torch module, improving their effectiveness for downstream vision tasks in varying illumination scenarios, effectively making it light-agnostic." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "text", + "content": "Specifically, the Adapt module computes an Adaptive Matrix " + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "text", + "content": " are the spatial dimensions of " + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "text", + "content": ". The purpose of the Adaptive Matrix " + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "text", + "content": " is to modulate the feature enhancements from Torch " + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "text", + "content": " before they are applied to " + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 221, + 555, + 281 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 399, + 286, + 555, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 286, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 399, + 286, + 555, + 300 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\sigma \\left(\\mathcal {G} _ {\\phi} (\\mathbf {I})\\right), \\tag {2}", + "image_path": "24deb3615c6fc0b672f4740ff25cc6c24805938eccdaebcc0cc4a6fcad4036d8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi}" + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "text", + "content": " is a lightweight convolutional network parameterized by weights " + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "text", + "content": " denotes the sigmoid activation function ensuring that the values of " + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "text", + "content": " are in the range " + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "text", + "content": ". We feed the input image " + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "text", + "content": " into the Adapt module to compute " + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 313, + 305, + 554, + 376 + ], + "type": "text", + "content": ", ensuring that the image content directly influences the modulation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "text", + "content": "Adapt network. As shown in Fig. 2, the convolutional network " + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\phi}" + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "text", + "content": " in the Adapt module is designed to extract hierarchical features, which are proven to be beneficial for high-level vision tasks such as object detection and semantic segmentation [26, 45, 46]. To capture multi-scale contextual information, the network consists of three " + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "text", + "content": " convolutional layers with increasing channel dimensions " + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "inline_equation", + "content": "C_1" + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "inline_equation", + "content": "C_2" + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "inline_equation", + "content": "C_3" + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "text", + "content": ", respectively. Each layer is followed by a batch normalization [29] and a ReLU activation function. This architecture allows the network to learn a hierarchy of features from low-level edges and textures to high-level semantic information. Finally, an additional " + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "text", + "content": " convolutional layer maps the features from " + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "inline_equation", + "content": "C_3" + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "text", + "content": " to a single channel and applies a sigmoid activation function to generate the Adaptive Matrix " + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\in [0,1]^{H\\times W}" + }, + { + "bbox": [ + 313, + 377, + 556, + 592 + ], + "type": "text", + "content": ". This design balances expressiveness and computational efficiency, enabling the Adapt module to adjust enhancements based on both local and global features, which is crucial for content-adaptive modulation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 592, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 556, + 689 + ], + "type": "text", + "content": "Illumination-aware scaling. We introduce a scale factor " + }, + { + "bbox": [ + 313, + 592, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 592, + 556, + 689 + ], + "type": "text", + "content": " to ensure that the enhancement adapts appropriately to different illumination levels. This scale factor modulates the M globally based on the image's overall brightness, allowing the model to apply stronger enhancements to darker images and milder adjustments to well-lit images, maintaining visual consistency and preventing artifacts. We compute " + }, + { + "bbox": [ + 313, + 592, + 556, + 689 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 592, + 556, + 689 + ], + "type": "text", + "content": " as a function of the average luminance " + }, + { + "bbox": [ + 313, + 592, + 556, + 689 + ], + "type": "inline_equation", + "content": "L(\\mathbf{I})" + }, + { + "bbox": [ + 313, + 592, + 556, + 689 + ], + "type": "text", + "content": " of the input image:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 370, + 693, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 693, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 370, + 693, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\gamma = \\frac {\\kappa}{1 + \\exp (k (L (\\mathbf {I}) - L _ {0}))}, \\tag {3}", + "image_path": "0acabf00d32219a9706308c5f0e7fd78601c65957953baa5c913199ed1b74bec.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5648" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " is the maximum scale factor (empirically set to 3 unless stated otherwise). " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " controls the slope of the sigmoid function, and " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " is the luminance value at the sigmoid's midpoint. The average luminance " + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "inline_equation", + "content": "L(\\mathbf{I})" + }, + { + "bbox": [ + 55, + 72, + 296, + 144 + ], + "type": "text", + "content": " is computed following the standard luminance calculation in color science [2, 42, 87]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 297, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 297, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 297, + 180 + ], + "type": "text", + "content": "Adaptive Fusion. The final enhanced image " + }, + { + "bbox": [ + 55, + 144, + 297, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 144, + 297, + 180 + ], + "type": "text", + "content": " is then obtained by modulating the enhancement from the Torch module with the illumination-aware scaled Adaptive Matrix:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 123, + 187, + 296, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 187, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 123, + 187, + 296, + 201 + ], + "type": "interline_equation", + "content": "\\mathbf {E} = \\mathbf {I} + ((\\gamma \\cdot \\mathbf {M}) \\odot \\mathbf {T}). \\tag {4}", + "image_path": "83430ea0b79df605a4724efafae3d624eafe0f30f14aa72690e9dc214445bed5.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 207, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 268 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 207, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 207, + 296, + 268 + ], + "type": "text", + "content": " denotes element-wise multiplication, and " + }, + { + "bbox": [ + 55, + 207, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 207, + 296, + 268 + ], + "type": "text", + "content": " is obtained from Eq. (1). This formulation in Eq. 4 integrates both content-based modulation (through " + }, + { + "bbox": [ + 55, + 207, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 207, + 296, + 268 + ], + "type": "text", + "content": ") and illumination-based scaling (through " + }, + { + "bbox": [ + 55, + 207, + 296, + 268 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 207, + 296, + 268 + ], + "type": "text", + "content": "), enabling the model to adaptively enhance images under varying lighting conditions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 274, + 194, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 194, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 194, + 286 + ], + "type": "text", + "content": "3.3. Light-Agnostic Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 291, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 458 + ], + "type": "text", + "content": "Achieving robust performance across varying lighting conditions requires our model to learn representations invariant to illumination changes. Collecting labeled data under all possible lighting conditions is impractical. Hence, we adopt a self-supervised learning (SSL) strategy inspired by BYOL [23] to learn illumination-invariant features. Unlike prior works that design cumbersome pipelines to incorporate SSL into each high-level vision task [16, 48, 65], we leverage SSL to pre-train our TorchAdapt module with a pretext task that aligns backbone representations [19, 27] regardless of illumination. This approach eliminates the need for task-specific labeled data or low-light data, allowing training on general-purpose datasets like ImageNet [18]. Figure 3 illustrates the pre-training pipeline of TorchAdapt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": "Objective Function. Given a well-lit image " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": ", we generate its low-light counterpart " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "V_{2} = D(V_{1})" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": " using a darkening function " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": ", which can be a gamma transformation, an illumination-degradation pipeline [16], or a darkening module from [48]. Let " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": " denote the frozen pre-trained backbone network (e.g., ResNet [27], ViT [19]). We pass " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "V_{2}" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": " through the TorchAdapt module to obtain enhanced images " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "\\hat{V}_{1}" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "\\hat{V}_{2}" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": ", respectively. The enhanced images are processed by the backbone " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": " and a projector network " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": " to obtain feature representations: " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "z_{1}^{\\prime} = g(f(\\hat{V}_{1}))" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "z_{2} = g(f(\\hat{V}_{2}))" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": ". The target representation is obtained by passing the original well-lit image " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": " through " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "f^{\\prime}" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "g^{\\prime}" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "y_{1} = g^{\\prime}(f^{\\prime}(V_{1}))" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": ". Our training objective minimizes the cosine similarity loss " + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{cos}}" + }, + { + "bbox": [ + 55, + 459, + 297, + 638 + ], + "type": "text", + "content": " between the enhanced image representations and the well-lit target representation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 643, + 296, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 643, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 67, + 643, + 296, + 677 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left(1 - \\frac {z _ {1} ^ {\\prime}}{\\| z _ {1} ^ {\\prime} \\| _ {2} \\| y _ {1} \\| _ {2}}\\right) + \\left(1 - \\frac {z _ {2} ^ {\\top} y _ {1}}{\\| z _ {2} \\| _ {2} \\| y _ {1} \\| _ {2}}\\right), \\tag {5}", + "image_path": "a87de7ee87b381cf5e2add95950ce5919f47fd4c225d69e1a2f1a6178acd32b9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " norm of a vector. Since we employ identical, frozen backbone networks " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", the loss function guides TorchAdapt to enhance the low-light image " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{V}_2" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 552, + 140 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 552, + 140 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 552, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 552, + 140 + ], + "type": "image", + "image_path": "c75950f2f37899bdc86d43e1a36906c387eaf13d8446bfa7f78a9f8280345623.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 148, + 555, + 226 + ], + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 226 + ], + "type": "text", + "content": "Figure 4. TorchAdapt's downstream integration pipeline. TorchAdapt works as a plug-and-play, backbone-agnostic feature enhancement module, seamlessly integrating into various downstream vision tasks. Given any image or video input across diverse lighting conditions, TorchAdapt produces illumination-invariant representations, significantly boosting performance without requiring task-specific modifications." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 230, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 314 + ], + "type": "text", + "content": "so its representation aligns with that of the well-lit image " + }, + { + "bbox": [ + 313, + 230, + 555, + 314 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 313, + 230, + 555, + 314 + ], + "type": "text", + "content": ". Simultaneously, it discourages over-enhancement of the already well-lit image " + }, + { + "bbox": [ + 313, + 230, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\hat{V}_{1}" + }, + { + "bbox": [ + 313, + 230, + 555, + 314 + ], + "type": "text", + "content": ", as its representation should remain consistent with " + }, + { + "bbox": [ + 313, + 230, + 555, + 314 + ], + "type": "inline_equation", + "content": "V_{1}" + }, + { + "bbox": [ + 313, + 230, + 555, + 314 + ], + "type": "text", + "content": ". This training strategy enables TorchAdapt to effectively enhance low-light images while preserving well-lit ones, promoting illumination invariance in the learned representations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 314, + 556, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 556, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 556, + 458 + ], + "type": "text", + "content": "Training Details. The training of TorchAdapt is fast and efficient due to its lightweight architecture and the absence of gradient propagation through the frozen backbone network. We train TorchAdapt for 5 epochs on the ImageNet [18] dataset using a single GPU, with a learning rate of 0.0001 and a batch size of 256. Since the backbone networks are pre-trained and we aim for consistent views, we apply only color jitter and random horizontal flip as data augmentations. Pre-training TorchAdapt with a ResNet-50 backbone takes " + }, + { + "bbox": [ + 313, + 314, + 556, + 458 + ], + "type": "inline_equation", + "content": "\\sim 6" + }, + { + "bbox": [ + 313, + 314, + 556, + 458 + ], + "type": "text", + "content": " hours on a single A100 GPU. Once trained, TorchAdapt can be seamlessly integrated into any downstream vision task, as illustrated in Fig. 4." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 315, + 466, + 555, + 569 + ], + "blocks": [ + { + "bbox": [ + 315, + 466, + 555, + 569 + ], + "lines": [ + { + "bbox": [ + 315, + 466, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 315, + 466, + 555, + 569 + ], + "type": "table", + "html": "
DatasetTask#Cls#Train#Val
Low-light (LL)
ExDark [47]Object detection1258911472
DARK FACE [80]Face detection15400600
ACDC [60]Semantic segmentation19400106
LIS [10]Instance segmentation81561669
DarkVision [81]Video object detection426*6*
Light-Agnostic (LA)
COCO [21] ∩ ExDark [47]Object detection12923575156
WIDER FACE [78] ∩ DARK FACE [80]Face detection1182803822
CityScapes [15] ∩ ACDC [60]Semantic segmentation193375606
COCO [40] ∩ LIS [10]Instance segmentation8443212550
ImageNet VID [58] ∩ DarkVision [81]Video object detection2401*51*
", + "image_path": "8171dd167c61ef425d4fe270ed9c336d6c9f63b6f6c55d6c1798078e2c4b4f06.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 576, + 555, + 598 + ], + "lines": [ + { + "bbox": [ + 313, + 576, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 555, + 598 + ], + "type": "text", + "content": "Table 1. Datasets Statistics.*: video samples. Light-Agnostic datasets are obtained through Eq. 6" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 609, + 394, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 394, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 394, + 624 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "We conduct extensive experiments on several representative visual tasks to demonstrate the superiority of TorchAdapt in both low-light and light-agnostic settings. These tasks include object detection [40, 47, 52], face detection [78, 80], instance segmentation [10, 40], semantic segmentation [15, 60], and video object detection [58, 81]. This section first presents the experimental settings, comparing TorchAdapt" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5649" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 69, + 555, + 242 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 555, + 242 + ], + "type": "table", + "html": "
Tasks → Light Settings → MethodsObject Det.Face Det.Instance Seg.Semantic Seg.Video Object Det.
LLLALLLALLLALLLALLLA
mAP50mAPmAP50mAPAP50APAP50APmAPSegAPmAPSegAPmIoUmIoUmAPmAP
Baseline78.844.262.433.147.319.952.628.054.046.041.533.949.675.532.851.3
Enhancement
Zero-DCE [24]76.242.757.629.547.420.147.622.951.543.539.232.142.559.47.8328.3
SCI [84]75.542.257.929.745.119.345.421.752.344.235.128.342.157.46.7125.7
URetinexNet [71]74.642.057.329.244.718.346.622.359.543.137.429.741.258.86.6825.1
IAT [17]74.141.052.430.542.115.937.117.449.742.336.430.141.756.26.2424.5
Task-specific
Zero-DCE [24]‡79.144.462.933.250.421.150.123.559.152.744.435.744.761.034.151.5
SCI [50]‡79.244.263.033.451.622.254.729.958.451.043.734.943.559.932.950.7
MAET [16]‡79.244.357.327.344.318.747.923.1--------
Xue et al. [76]----46.117.744.221.0----42.169.7--
FeatEnHancer [26]‡79.244.862.933.347.219.951.624.559.652.444.936.150.176.134.652.7
TorchAdapt80.145.464.034.153.723.159.032.561.153.645.236.650.876.339.159.1
vs. prev. SoTA+0.9+0.6+1.0+0.7+2.1+0.9+4.3+2.6+1.5+0.9+0.3+0.5+0.7+0.2+4.5+6.4
", + "image_path": "66321bd3257498f3523c28c810e497bf99d469b9c9e78f725013b620322d402e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "type": "text", + "content": "Table 2. Quantitative comparison across five tasks under both Low-Light (LL) and Light-Agnostic (LA) settings. ‡: all these methods are reproduced by us with their official codebase and trained end-to-end for each task. The rest of the results are obtained from their official checkpoints. - indicates that the method is not applicable. TorchAdapt outperforms prior SOTA across all five tasks in LL and LA settings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 294, + 294, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 294, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 294, + 366 + ], + "type": "text", + "content": "with baselines, existing low-light image enhancement (LLIE) methods, and state-of-the-art task-specific approaches. We then examine critical design choices for TorchAdapt through ablation studies. Finally, we validate the generalizability of TorchAdapt by comparing it with recent state-of-the-art domain adaptation methods [20, 32, 48]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 367, + 295, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 390 + ], + "type": "text", + "content": "Experimental Settings. We perform experiments under two settings for each task:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 392, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 439 + ], + "type": "text", + "content": "1)- Low-Light (LL) Setting: Following prior works [16, 26, 76], we utilize publicly available low-light benchmarks specific to each task. These benchmarks consist exclusively of images captured under low-light conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "type": "text", + "content": "2)- Light-Agnostic (LA) Setting: To introduce a more challenging and realistic evaluation, we curate datasets that encompass both well-lit and low-light images. We achieve this by selecting classes common to both well-lit and low-light datasets. Mathematically, for a given task, let " + }, + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{WL}}" + }, + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{LL}}" + }, + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "type": "text", + "content": " denote the sets of classes in the well-lit and low-light datasets, respectively. We define the set of common classes as " + }, + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{com}} = \\mathcal{C}_{\\mathrm{WL}} \\cap \\mathcal{C}_{\\mathrm{LL}}" + }, + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "type": "text", + "content": ". The LA dataset is then formed by combining all images from both datasets that belong to classes in " + }, + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{com}}" + }, + { + "bbox": [ + 54, + 440, + 295, + 559 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 569, + 295, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 569, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 67, + 569, + 295, + 585 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\mathrm {L A}} = \\left\\{\\left(x, y\\right) \\mid (x, y) \\in \\mathcal {D} _ {\\mathrm {W L}} \\cup \\mathcal {D} _ {\\mathrm {L L}}, y \\in \\mathcal {C} _ {\\mathrm {c o m}} \\right\\}, \\tag {6}", + "image_path": "27a8842ce39bee3bb6b348111d18dd0d598c26413652d8cf9254689db95b7bd5.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 594, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 653 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 54, + 594, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{WL}}" + }, + { + "bbox": [ + 54, + 594, + 295, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 594, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{LL}}" + }, + { + "bbox": [ + 54, + 594, + 295, + 653 + ], + "type": "text", + "content": " are the well-lit and low-light datasets, respectively, and " + }, + { + "bbox": [ + 54, + 594, + 295, + 653 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 54, + 594, + 295, + 653 + ], + "type": "text", + "content": " represents an image and its label. We summarize the key statistics of the employed benchmarks for both settings in Table 1. We provide complete implementation details for each experiment in Appendix A." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 653, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 296, + 712 + ], + "type": "text", + "content": "Comparisons with State-of-the-Arts. We directly compare our approach with several LLIE methods, including Zero-DCE [24], IAT [17], SCI [50], and URetinexNet [71]. Following common practice [16, 26, 76], we use their released checkpoints to enhance all images before feeding them into" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 294, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 486 + ], + "type": "text", + "content": "the detector for performance evaluation. Notably, due to their zero-reference nature requiring no paired images, some of these LLIE methods [24, 50] can be trained end-to-end on downstream vision tasks. Therefore, besides evaluating their conventional enhancement results, we integrate them into each task and report their performance, similar to low-light task-specific works like MAET [16] and FeatEnHancer [26] (indicated with " + }, + { + "bbox": [ + 313, + 294, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 313, + 294, + 555, + 486 + ], + "type": "text", + "content": " in Table 2). Our findings reveal that end-to-end training of such LLIE works " + }, + { + "bbox": [ + 313, + 294, + 555, + 486 + ], + "type": "inline_equation", + "content": "[24, 50]^1" + }, + { + "bbox": [ + 313, + 294, + 555, + 486 + ], + "type": "text", + "content": " significantly improves their performance on downstream vision tasks in both LL and LA settings. Concurrent to our work, we notice that two new methods [28, 38] appear in the literature on improving low-light object detection. However, since their code and checkpoints were not released by the time of the submission, we could not draw direct comparisons with them. Next, we discuss experiments on each task." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 496, + 416, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 416, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 416, + 509 + ], + "type": "text", + "content": "4.1. Object Detection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 514, + 556, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 556, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 556, + 682 + ], + "type": "text", + "content": "Settings. Following common practice [16] in LL object detection, we adopt YOLOv3 [57] as the baseline detector. For the LL setting, COCO [40] pre-trained weights are utilized to fine-tune YOLOv3 on the ExDark dataset [47]. We train the model from scratch in the LA setting, as COCO images are also included in the training data (see Table. 1). Results. As shown in Table 2, TorchAdapt achieves a superior " + }, + { + "bbox": [ + 313, + 514, + 556, + 682 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 313, + 514, + 556, + 682 + ], + "type": "text", + "content": " of 80.1 in LL and 64.1 in LA, compared to the baseline's 78.8 and 62.4. End-to-end training boosts the performance of enhancement methods like Zero-DCE [33] and SCI [50]; however, TorchAdapt still surpasses them, and the low-light object detection method FeatEnHancer [26] by " + }, + { + "bbox": [ + 313, + 514, + 556, + 682 + ], + "type": "inline_equation", + "content": "+0.9" + }, + { + "bbox": [ + 313, + 514, + 556, + 682 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 514, + 556, + 682 + ], + "type": "inline_equation", + "content": "+1.1" + }, + { + "bbox": [ + 313, + 514, + 556, + 682 + ], + "type": "text", + "content": " in LL and LA settings, respectively. These results affirm the efficacy of TorchAdapt for object detec" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 713 + ], + "type": "text", + "content": "1We select these works due to their efficiency and reproducibility. Experiments with more methods are in Appendix C." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5650" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "tion across varying illumination conditions. We provide a qualitative comparison in Appendix B." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 103, + 148, + 116 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 148, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 148, + 116 + ], + "type": "text", + "content": "4.2. Face Detection" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": "Settings. For face detection, we choose a different detector and adopt RetinaNet [41] to report results on the DARK FACE dataset [69, 80] in the LL setting and a combination of the DARK FACE and WIDER FACE [78] datasets in the LA setting. Following prior works [26], we resize images to a resolution of " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "1500 \\times 1000" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": " pixels and follow the " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": " schedule in MMDetection [9]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 205, + 297, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 297, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 297, + 337 + ], + "type": "text", + "content": "Results. Results in Table 2 reveal that TorchAdapt demonstrates superior performance on face detection under both LL and LA settings, outperforming the previous best method " + }, + { + "bbox": [ + 55, + 205, + 297, + 337 + ], + "type": "inline_equation", + "content": "(\\mathrm{SCI} \\ddagger [50])" + }, + { + "bbox": [ + 55, + 205, + 297, + 337 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 205, + 297, + 337 + ], + "type": "inline_equation", + "content": "+2.1" + }, + { + "bbox": [ + 55, + 205, + 297, + 337 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 205, + 297, + 337 + ], + "type": "inline_equation", + "content": "+4.3" + }, + { + "bbox": [ + 55, + 205, + 297, + 337 + ], + "type": "text", + "content": " in mAP_50, respectively. The significant gains, especially in the challenging LA setting, highlight TorchAdapt's capacity to handle diverse illumination scenarios by adaptively modulating features based on content and illumination, thereby improving the robustness and accuracy of the baseline. We also compare TorchAdapt with recent domain adaptation methods [20, 48] using DSFD [36] in Sec. 4.7." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 344, + 185, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 185, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 185, + 357 + ], + "type": "text", + "content": "4.3. Instance Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": "Settings. For the instance segmentation task, we adopt RTMDet-Ins-tiny " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "[49]^2" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": " as our baseline model and resize images to " + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "inline_equation", + "content": "640\\times 640" + }, + { + "bbox": [ + 55, + 361, + 296, + 433 + ], + "type": "text", + "content": " pixels for training. Following Section 4.1, we use the COCO pre-trained weights of RTMDet-Ins-tiny for the LL setting and train the model from scratch for the LA setting (see Table 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 434, + 297, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 297, + 541 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 297, + 541 + ], + "type": "text", + "content": "Results. As summarized in Table 2), TorchAdapt achieves SOTA SegAP performance in both low-light (LL) and light-agnostic (LA) settings. Under LL conditions, TorchAdapt attains a SegAP of 53.6, surpassing the previous best (FeatEnHancer†) by +1.2. In the LA setting, it achieves a SegAP of 36.6, slightly improving overFeatEnHancer† by +0.5. These results demonstrate TorchAdapt's effectiveness in enhancing instance segmentation across varying illumination by producing illumination-invariant features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 548, + 188, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 188, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 188, + 561 + ], + "type": "text", + "content": "4.4. Semantic Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 566, + 297, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 297, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 297, + 686 + ], + "type": "text", + "content": "Settings. Consistent with prior works [26, 76], we adopt DeepLabV3+[11] as our baseline model and follow [26, 76] for the LL setting. For the LA evaluation (Table 1), we train DeepLabV3+ from scratch using MMSegmentation [14]. Results. As shown in Table 2, TorchAdapt achieves an mIoU of 50.8 in the LL setting, surpassing the previous best (FeatEnHancer†) by +0.7, and an mIoU of 76.3 in the LA setting, improving overFeatEnHancer† by +0.2. These results highlight TorchAdapt's effectiveness for semantic segmentation under varying illumination conditions." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 554, + 133 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 554, + 133 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 554, + 133 + ], + "type": "table", + "html": "
#Torch(§ 3.1)Adapt(§ 3.2)Scaling(§ 3.2)Object Det.Instance Seg.
LLLALLLA
1×××78.862.446.033.9
2✓××79.162.951.134.8
3✓✓×79.963.553.336.2
4✓×✓79.562.751.434.1
5✓✓✓80.164.053.636.6
", + "image_path": "9a1eb1374af3d822c6420a6151822a8cdffaf927d10f75265a404123489e4177.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 319, + 167, + 426, + 207 + ], + "blocks": [ + { + "bbox": [ + 313, + 141, + 555, + 163 + ], + "lines": [ + { + "bbox": [ + 313, + 141, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 141, + 555, + 163 + ], + "type": "text", + "content": "Table 3. Contribution of each component. Highlighted settings are set as default." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 167, + 426, + 207 + ], + "lines": [ + { + "bbox": [ + 319, + 167, + 426, + 207 + ], + "spans": [ + { + "bbox": [ + 319, + 167, + 426, + 207 + ], + "type": "table", + "html": "
Scale factor (κ)ODSS
LLLALLLA
279.963.753.436.4
380.164.053.636.6
479.963.353.236.2
579.561.153.334.5
", + "image_path": "8d0c465469dbe2148f7b50855a22233c597314fe8d77166ca7e99973c2babbe5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 325, + 209, + 421, + 219 + ], + "lines": [ + { + "bbox": [ + 325, + 209, + 421, + 219 + ], + "spans": [ + { + "bbox": [ + 325, + 209, + 421, + 219 + ], + "type": "text", + "content": "(a) Scale factor " + }, + { + "bbox": [ + 325, + 209, + 421, + 219 + ], + "type": "inline_equation", + "content": "(\\kappa)" + }, + { + "bbox": [ + 325, + 209, + 421, + 219 + ], + "type": "text", + "content": " in Eq. 3." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 428, + 167, + 547, + 207 + ], + "blocks": [ + { + "bbox": [ + 428, + 167, + 547, + 207 + ], + "lines": [ + { + "bbox": [ + 428, + 167, + 547, + 207 + ], + "spans": [ + { + "bbox": [ + 428, + 167, + 547, + 207 + ], + "type": "table", + "html": "
D(I)ODSS
LLLALLLA
ISP [16]79.461.747.629.9
D(I) in [48]79.663.153.236.3
Gamma curve80.164.053.636.6
", + "image_path": "88b41eb27354e01f4b82e718a576112b3aed18ddafa26394279f903437d53583.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 440, + 209, + 534, + 218 + ], + "lines": [ + { + "bbox": [ + 440, + 209, + 534, + 218 + ], + "spans": [ + { + "bbox": [ + 440, + 209, + 534, + 218 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 440, + 209, + 534, + 218 + ], + "type": "inline_equation", + "content": "D(\\mathbf{I})" + }, + { + "bbox": [ + 440, + 209, + 534, + 218 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 440, + 209, + 534, + 218 + ], + "type": "inline_equation", + "content": "\\S 3.3" + }, + { + "bbox": [ + 440, + 209, + 534, + 218 + ], + "type": "text", + "content": " and Fig. 3." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 224, + 554, + 259 + ], + "lines": [ + { + "bbox": [ + 313, + 224, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 554, + 259 + ], + "type": "text", + "content": "Table 4. Ablating design choices in TorchAdapt. Highlighted settings are set as default. OD and SS are object detection and semantic segmentation, respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 265, + 446, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 265, + 446, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 265, + 446, + 277 + ], + "type": "text", + "content": "4.5. Video Object Detection" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 281, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 341 + ], + "type": "text", + "content": "Settings. In addition to image-based tasks, we evaluate the generalizability of TorchAdapt on video data. We present the dataset details in Table 1. Other experimental settings follow [26], with additional implementation details provided in Appendix A." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 342, + 556, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 450 + ], + "type": "text", + "content": "Results. We report the results of the challenging video object detection task in Table 2. While all LLIE methods perform poorly on this task, end-to-end training significantly boosts the performance of zero-reference methods [24, 50]. TorchAdapt, however, outperforms all methods under both LL and LA settings, achieving substantial gains of " + }, + { + "bbox": [ + 313, + 342, + 556, + 450 + ], + "type": "inline_equation", + "content": "+4.5" + }, + { + "bbox": [ + 313, + 342, + 556, + 450 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 342, + 556, + 450 + ], + "type": "inline_equation", + "content": "+6.4" + }, + { + "bbox": [ + 313, + 342, + 556, + 450 + ], + "type": "text", + "content": ", respectively. These improvements affirm TorchAdapt's robustness and effectiveness for video object detection across both low-light and light-agnostic settings." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 456, + 416, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 416, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 416, + 468 + ], + "type": "text", + "content": "4.6. Ablation Studies" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 474, + 556, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 534 + ], + "type": "text", + "content": "We ablate the components and design choices of TorchAdapt in object detection (OD) and instance segmentation (IS) tasks under (Low-light) LL and (Light-Agnostic) LA settings, following the same implementation details in § 4.1 and § 4.3, respectively. We provide more ablations in Appendix C." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "content": "Contribution of Each Component. Table 3 evaluates the impact of the Torch module, the Adapt module, and Illumination Scaling. Adding the Torch module alone (row 2) improves performance over the baseline with " + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "inline_equation", + "content": "+0.3\\%" + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "content": " in LL OD and " + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "inline_equation", + "content": "+5.1\\%" + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "content": " in LL IS, highlighting the benefit of feature enhancement in low-light vision. Introducing the Adapt module alongside Torch (row 3) yields significant improvements " + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "inline_equation", + "content": "(+0.8\\% / +0.6\\%)" + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "content": " in LL/LA OD; " + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "inline_equation", + "content": "+2.2\\% / +1.4\\%" + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "content": " in LL/LA IS), underscoring the importance of content-adaptive modulation. Employing scaling without the Adapt module (Row 4) offers minor gains in LL settings. However, it is detrimental in LA settings " + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "inline_equation", + "content": "(-0.2\\% / -0.7\\%)" + }, + { + "bbox": [ + 313, + 534, + 557, + 713 + ], + "type": "text", + "content": " in OD/IS), indicating that content-based adaptation is more impactful than global scaling alone. The full TorchAdapt configuration, combining all three components (row 5), achieves the best results with" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 297, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 297, + 713 + ], + "type": "text", + "content": "2https://github.com/open-mmlab/mmdetection/blob/main/configs/rtmdet/" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5651" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 73, + 294, + 159 + ], + "blocks": [ + { + "bbox": [ + 57, + 73, + 294, + 159 + ], + "lines": [ + { + "bbox": [ + 57, + 73, + 294, + 159 + ], + "spans": [ + { + "bbox": [ + 57, + 73, + 294, + 159 + ], + "type": "image", + "image_path": "e603b3cb4e53545917c1c2f92488686b35b8e2f107fa9d9a38ce8ccdbfad4390.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 167, + 296, + 223 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 223 + ], + "type": "text", + "content": "Figure 5. Visualizing backbone features with SCI [50] on LA object detection using YOLOv3. (a) SCI's official checkpoint as an enhancer. (b) " + }, + { + "bbox": [ + 55, + 167, + 296, + 223 + ], + "type": "inline_equation", + "content": "\\mathrm{SCI}^{\\dagger}" + }, + { + "bbox": [ + 55, + 167, + 296, + 223 + ], + "type": "text", + "content": " trained end-to-end. (c) SCI integrated into our TorchAdapt framework as a Torch. SCI with TorchAdapt extracts more representative features in both low and well-lit images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 109, + 224, + 243, + 288 + ], + "blocks": [ + { + "bbox": [ + 109, + 224, + 243, + 288 + ], + "lines": [ + { + "bbox": [ + 109, + 224, + 243, + 288 + ], + "spans": [ + { + "bbox": [ + 109, + 224, + 243, + 288 + ], + "type": "table", + "html": "
MethodmAP (%)
DSFD [36]16.1
CICony [32]18.4
SimMinMax [48]25.7
DAI-Net [20]‡28.0
DAI-Net+TorchAdapt31.9+3.9
", + "image_path": "39f0dd06dee8c9f16120ea873df2e77f9ba491e769d087cca921c04576e2393e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 369, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 369, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 296, + 416 + ], + "type": "text", + "content": "+1.3%/+1.6% in LL/LA OD and +7.6%/+2.5% in LL/LA IS over the baseline, confirming that these components complement each other and synergistically enhance performance in both LL and LA settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 416, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 501 + ], + "type": "text", + "content": "Design Choices in TorchAdapt. Table 4 ablates important design choices in TorchAdapt. As shown in Table 4a, we empirically establish that when the scale factor " + }, + { + "bbox": [ + 55, + 416, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 416, + 296, + 501 + ], + "type": "text", + "content": " in Eq. 3 is computed with " + }, + { + "bbox": [ + 55, + 416, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\kappa = 3" + }, + { + "bbox": [ + 55, + 416, + 296, + 501 + ], + "type": "text", + "content": ", it produces optimal performance. Results in Table 4b reveal that employing the Gamma Curve in " + }, + { + "bbox": [ + 55, + 416, + 296, + 501 + ], + "type": "inline_equation", + "content": "D(I)" + }, + { + "bbox": [ + 55, + 416, + 296, + 501 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 55, + 416, + 296, + 501 + ], + "type": "inline_equation", + "content": "V_{2}" + }, + { + "bbox": [ + 55, + 416, + 296, + 501 + ], + "type": "text", + "content": " in § 3.3 brings the biggest boost in performance across both LL and LA settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 509, + 295, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 521 + ], + "type": "text", + "content": "4.7. Comparison with Domain-Adaptation Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 525, + 296, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 525, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 54, + 525, + 296, + 633 + ], + "type": "text", + "content": "We evaluate TorchAdapt's effectiveness for zero-shot domain adaptive face detection on the DARK FACE dataset [80]. Following settings from [20], we train the detector only on the WIDER FACE dataset [78]. As shown in Table 5, TorchAdapt integrated with DAI-Net [20], the current best-performing method, achieves a significant improvement of " + }, + { + "bbox": [ + 54, + 525, + 296, + 633 + ], + "type": "inline_equation", + "content": "+3.9\\%" + }, + { + "bbox": [ + 54, + 525, + 296, + 633 + ], + "type": "text", + "content": " mAP, demonstrating its effectiveness in enabling robust zero-shot adaptation under diverse illumination conditions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 645, + 251, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 251, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 251, + 659 + ], + "type": "text", + "content": "5. Exploring TorchAdapt's Properties" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Model-Agnostic Generality. Experiments in § 4 confirm that TorchAdapt is a general-purpose module that can be integrated into any high-level vision task. Moreover, TorchAdapt as a framework is highly flexible: any zero-reference (unsu" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 375, + 113 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 375, + 113 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 375, + 113 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 375, + 113 + ], + "type": "image", + "image_path": "ee41f6e42f54df25a83aa563fd4d172a8a4f66b3f9ca9353912d2893c5abe8df.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 114, + 348, + 121 + ], + "lines": [ + { + "bbox": [ + 343, + 114, + 348, + 121 + ], + "spans": [ + { + "bbox": [ + 343, + 114, + 348, + 121 + ], + "type": "text", + "content": "I" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 376, + 71, + 433, + 113 + ], + "blocks": [ + { + "bbox": [ + 376, + 71, + 433, + 113 + ], + "lines": [ + { + "bbox": [ + 376, + 71, + 433, + 113 + ], + "spans": [ + { + "bbox": [ + 376, + 71, + 433, + 113 + ], + "type": "image", + "image_path": "9b2a7897bd66a6e77f556447b7190a5df314b46bf3f6696eaef639d16f113a87.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 114, + 429, + 121 + ], + "lines": [ + { + "bbox": [ + 380, + 114, + 429, + 121 + ], + "spans": [ + { + "bbox": [ + 380, + 114, + 429, + 121 + ], + "type": "text", + "content": "TorchAdapt(I)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 434, + 71, + 492, + 114 + ], + "blocks": [ + { + "bbox": [ + 434, + 71, + 492, + 114 + ], + "lines": [ + { + "bbox": [ + 434, + 71, + 492, + 114 + ], + "spans": [ + { + "bbox": [ + 434, + 71, + 492, + 114 + ], + "type": "image", + "image_path": "25b5ed0c93f79064680e3cc7139d3d8c17adda2977e3ec18b15eff4f99e37679.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 114, + 466, + 121 + ], + "lines": [ + { + "bbox": [ + 462, + 114, + 466, + 121 + ], + "spans": [ + { + "bbox": [ + 462, + 114, + 466, + 121 + ], + "type": "text", + "content": "I" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 132, + 555, + 165 + ], + "lines": [ + { + "bbox": [ + 313, + 132, + 555, + 165 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 555, + 165 + ], + "type": "text", + "content": "Figure 6. Illustrating low-light image Enhancement as an emergent property of TorchAdapt on the DARK FACE validation set. More results can be found in Appendix B." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 494, + 71, + 553, + 114 + ], + "blocks": [ + { + "bbox": [ + 494, + 71, + 553, + 114 + ], + "lines": [ + { + "bbox": [ + 494, + 71, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 494, + 71, + 553, + 114 + ], + "type": "image", + "image_path": "6d76615a5888f9954e5e83faa48066095eaedc649ff390d22ed8066855e23d58.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 497, + 114, + 548, + 121 + ], + "lines": [ + { + "bbox": [ + 497, + 114, + 548, + 121 + ], + "spans": [ + { + "bbox": [ + 497, + 114, + 548, + 121 + ], + "type": "text", + "content": "TorchAdapt(I)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 315, + 179, + 553, + 209 + ], + "blocks": [ + { + "bbox": [ + 54, + 296, + 295, + 361 + ], + "lines": [ + { + "bbox": [ + 54, + 296, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 296, + 295, + 361 + ], + "type": "text", + "content": "Table 5. Zero-shot performance comparison of TorchAdapt with state-of-the-art domain adaptation methods on DARK FACE [80] using DSFD [36]. Results with " + }, + { + "bbox": [ + 54, + 296, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 54, + 296, + 295, + 361 + ], + "type": "text", + "content": " are reproduced using the original codebase. TorchAdapt enhances prior SOTA [20] by " + }, + { + "bbox": [ + 54, + 296, + 295, + 361 + ], + "type": "inline_equation", + "content": "+3.9\\mathrm{mAP}" + }, + { + "bbox": [ + 54, + 296, + 295, + 361 + ], + "type": "text", + "content": ", demonstrating its effectiveness in zero-shot domain adaptation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 179, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 315, + 179, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 315, + 179, + 553, + 209 + ], + "type": "table", + "html": "
MethodmAP50(LL)mAP50(LA)#Params.(M)Latency (ms)
Baseline [57]78.862.465.221.9
+TorchAdapt80.1 (+1.3)64.0 (+1.6)65.3 (+0.1)23.7 (+1.8)
", + "image_path": "17783a468fc666ac2b016ad53ab0ddb1916d0c0074fda3e5ea9c5ddff04764ce.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 217, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 250 + ], + "type": "text", + "content": "Table 6. Computational analysis. With YOLOv3 [57], TorchAdapt introduces minimal computational overhead while bringing stronger gains across both LL and LA settings." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 265, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 361 + ], + "type": "text", + "content": "pervised) LLIE methods [33, 50, 85] can serve as the Torch component to improve performance under varying illuminations. To verify this, we adopt SCI [50] as the Torch module in TorchAdapt and visualize the learned backbone features in Fig.5(c). Using SCI within TorchAdapt yields highly representative backbone features for both low-light and well-lit images. Appendix C provides additional examples and quantitative analyses." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 362, + 556, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 556, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 556, + 482 + ], + "type": "text", + "content": "Enhancement as an Emergent Property. Although TorchAdapt is designed to produce illumination-invariant features for high-level vision tasks without employing any explicit enhancement loss functions during training, we observe that it inherently enhances low-light images. We illustrate this effect in Fig. 6. This emergent behavior likely results from the pre-training objective of aligning representations between well-lit and low-light images (see § 3.3), which encourages the network to implicitly adjust low-light images toward well-lit representations." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 483, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 568 + ], + "type": "text", + "content": "Computational Analysis. We assess the computational efficiency of TorchAdapt using YOLOv3 [57] on the object detection task. As shown in Table 6, integrating TorchAdapt incurs minimal additional computational cost while delivering significant performance improvements. This demonstrates that TorchAdapt is compatible with lightweight models and suitable for real-time applications." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 584, + 391, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 391, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 391, + 596 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 605, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 712 + ], + "type": "text", + "content": "In this work, we introduce TorchAdapt, a real-time, general-purpose, light-agnostic adaptive feature enhancement framework that boosts the performance of high-level vision tasks under varying illumination conditions. By integrating its modules with light-agnostic learning, TorchAdapt produces powerful semantic representations for both low-light and well-lit data. Extensive experiments on five representative visual perception tasks involving images and videos confirm its effectiveness." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5652" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 116 + ], + "type": "text", + "content": "This work was partially funded by the European Union under Horizon Europe, grant number 101135724 (LUMINOUS)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 148, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 62, + 148, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 296, + 213 + ], + "type": "text", + "content": "[1] Muhamamd Zeshan Afzal, SK Aziz Ali, Didier Stricker, Peter Eisert, Anna Hilsmann, Daniel Perez-Marcos, Marco Bianchi, Sonia Crottaz-Herbette, Roberto De Ioris, Eleni Mangina, et al. Next generation xr systems-large language models meet augmented and virtual reality. IEEE computer graphics and applications, 2025. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 216, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 216, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 296, + 248 + ], + "type": "text", + "content": "[2] Tadahiro Azetsu and Noriaki Suetake. Hue-preserving image enhancement in cielab color space considering color gamut. Optical Review, 26:283-294, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 296, + 281 + ], + "type": "text", + "content": "[3] Ruzena Bajcsy, Yiannis Aloimonos, and John K Tsotsos. Revisiting active perception. Autonomous Robots, 42:177-196, 2018. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 296, + 316 + ], + "type": "text", + "content": "[4] Peter J Bickel and Kjell A Doksum. Mathematical statistics: basic ideas and selected topics, volumes I-II package. Chapman and Hall/CRC, 2015. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "type": "text", + "content": "[5] Vladimir Bychkovsky, Sylvain Paris, Eric Chan, and Fredo Durand. Learning photographic global tonal adjustment with a database of input/output image pairs. In CVPR 2011, pages 97-104. IEEE, 2011. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 365, + 296, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 296, + 419 + ], + "type": "text", + "content": "[6] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 421, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 296, + 464 + ], + "type": "text", + "content": "[7] Chenyi Chen, Ari Seff, Alain Kornhauser, and Jianxiong Xiao. Deepdriving: Learning affordance for direct perception in autonomous driving. In Proceedings of the IEEE international conference on computer vision, pages 2722-2730, 2015. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "type": "text", + "content": "[8] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3291-3300, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 512, + 296, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 598 + ], + "type": "text", + "content": "[9] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 633 + ], + "type": "text", + "content": "[10] Linwei Chen, Ying Fu, Kaixuan Wei, Dezhi Zheng, and Felix Heide. Instance segmentation in the dark. International Journal of Computer Vision, 131(8):2198-2218, 2023. 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[11] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), 2018. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[12] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "text", + "content": "conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 555, + 149 + ], + "type": "text", + "content": "[13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 555, + 194 + ], + "type": "text", + "content": "[14] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 196, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 196, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 196, + 554, + 259 + ], + "type": "text", + "content": "[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 1, 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 262, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 262, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 262, + 554, + 316 + ], + "type": "text", + "content": "[16] Ziteng Cui, Guo-Jun Qi, Lin Gu, Shaodi You, Zenghui Zhang, and Tatsuya Harada. Multitask aet with orthogonal tangent regularity for dark object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2553-2562, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 317, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 554, + 370 + ], + "type": "text", + "content": "[17] Ziteng Cui, Kunchang Li, Lin Gu, Shenghan Su, Peng Gao, Zhengkai Jiang, Yu Qiao, and Tatsuya Harada. You only need 90k parameters to adapt light: a light weight transformer for image enhancement and exposure correction. arXiv preprint arXiv:2205.14871, 2022. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 373, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 373, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 373, + 554, + 415 + ], + "type": "text", + "content": "[18] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 2, 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 417, + 555, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 417, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 417, + 555, + 502 + ], + "type": "text", + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 505, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 555, + 557 + ], + "type": "text", + "content": "[20] Zhipeng Du, Miaojing Shi, and Jiankang Deng. Boosting object detection with zero-shot day-night domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12666-12676, 2024. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 559, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 555, + 612 + ], + "type": "text", + "content": "[21] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 615, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 555, + 646 + ], + "type": "text", + "content": "[22] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. arXiv preprint arXiv:1803.07728, 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 647, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 555, + 713 + ], + "type": "text", + "content": "[23] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2, 3, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5653" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "text", + "content": "[24] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. CoRR, abs/2001.06826, 2020. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 297, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 297, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 297, + 171 + ], + "type": "text", + "content": "[25] Jiang Hai, Zhu Xuan, Ren Yang, Yutong Hao, Fengzhu Zou, Fang Lin, and Songchen Han. R2rnet: Low-light image enhancement via real-low to real-normal network. Journal of Visual Communication and Image Representation, 90:103712, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 297, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 297, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 297, + 237 + ], + "type": "text", + "content": "[26] Khurram Azeem Hashmi, Goutham Kallempudi, Didier Stricker, and Muhammad Zeshan Afzal. Featenhancer: Enhancing hierarchical features for object detection and beyond under low-light vision. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6725-6735, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 297, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 297, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 297, + 293 + ], + "type": "text", + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 297, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 297, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 297, + 337 + ], + "type": "text", + "content": "[28] Mingbo Hong, Shen Cheng, Haibin Huang, Haoqiang Fan, and Shuaicheng Liu. You only look around: Learning illumination invariant feature for low-light object detection. arXiv preprint arXiv:2410.18398, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 297, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 297, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 297, + 370 + ], + "type": "text", + "content": "[29] Sergey Ioffe. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 372, + 297, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 297, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 297, + 425 + ], + "type": "text", + "content": "[30] Yifan Jiang, Xinyu Gong, Ding Liu, Yu Cheng, Chen Fang, Xiaohui Shen, Jianchao Yang, Pan Zhou, and Zhangyang Wang. Enlightengan: Deep light enhancement without paired supervision. IEEE Transactions on Image Processing, 30: 2340-2349, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 297, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 297, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 297, + 480 + ], + "type": "text", + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 297, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 297, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 297, + 535 + ], + "type": "text", + "content": "[32] Attila Lengyel, Sourav Garg, Michael Milford, and Jan C. van Gemert. Zero-shot day-night domain adaptation with a physics prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4399-4409, 2021. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 297, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 297, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 297, + 581 + ], + "type": "text", + "content": "[33] Chongyi Li, Chunle Guo Guo, and Chen Change Loy. Learning to enhance low-light image via zero-reference deep curve estimation. In IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 297, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 297, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 297, + 635 + ], + "type": "text", + "content": "[34] Chongyi Li, Chunle Guo, Shangchen Zhou, Qiming Ai, Ruicheng Feng, and Chen Change Loy. Flexicurve: Flexible piecewise curves estimation for photo retouching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1092-1101, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 637, + 297, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 297, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 297, + 690 + ], + "type": "text", + "content": "[35] Hebei Li, Jin Wang, Jiahui Yuan, Yue Li, Wenming Weng, Yansong Peng, Yueyi Zhang, Zhiwei Xiong, and Xiaoyan Sun. Event-assisted low-light video object segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3250-3259, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 297, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 297, + 713 + ], + "type": "text", + "content": "[36] Jian Li, Yabiao Wang, Changan Wang, Ying Tai, Jianjun Qian, Jian Yang, Chengjie Wang, Jilin Li, and Feiyue Huang. Dsfd:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 556, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 556, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 556, + 106 + ], + "type": "text", + "content": "dual shot face detector. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5060-5069, 2019. 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 556, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 556, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 556, + 173 + ], + "type": "text", + "content": "[37] Jinlong Li, Baolu Li, Zhengzhong Tu, Xinyu Liu, Qing Guo, Felix Juefei-Xu, Runsheng Xu, and Hongkai Yu. Light the night: A multi-condition diffusion framework for unpaired low-light enhancement in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15205-15215, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 556, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 556, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 556, + 239 + ], + "type": "text", + "content": "[38] Yunan Li, Yihao Zhang, Shoude Li, Long Tian, Dou Quan, Chaoneng Li, and Qiguang Miao. Watching it in dark: A target-aware representation learning framework for high-level vision tasks in low illumination. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 240, + 556, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 556, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 556, + 284 + ], + "type": "text", + "content": "[39] Jinxiu Liang, Jingwen Wang, Yuhui Quan, Tianyi Chen, Jiaying Liu, Haibin Ling, and Yong Xu. Recurrent exposure generation for low-light face detection. IEEE Transactions on Multimedia, 24:1609-1621, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 556, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 556, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 556, + 341 + ], + "type": "text", + "content": "[40] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. CoRR, abs/1405.0312, 2014. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 556, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 556, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 556, + 374 + ], + "type": "text", + "content": "[41] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. CoRR, abs/1708.02002, 2017. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 556, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 556, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 556, + 430 + ], + "type": "text", + "content": "[42] Hewei Liu, Renwei Yang, Shuyuan Zhu, Xing Wen, and Bing Zeng. Luminance-guided chrominance image enhancement for hevc intra coding. In 2022 IEEE International Symposium on Circuits and Systems (ISCAS), pages 3180-3184. IEEE, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 433, + 556, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 556, + 475 + ], + "type": "text", + "content": "[43] Jiaying Liu, Dejia Xu, Wenhan Yang, Minhao Fan, and Haofeng Huang. Benchmarking low-light image enhancement and beyond. International Journal of Computer Vision, 129:1153-1184, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 478, + 556, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 556, + 533 + ], + "type": "text", + "content": "[44] Risheng Liu, Long Ma, Jiaao Zhang, Xin Fan, and Zhongxuan Luo. Retinex-inspired unrolling with cooperative prior architecture search for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10561-10570, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 534, + 556, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 556, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 556, + 589 + ], + "type": "text", + "content": "[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, 2021. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 556, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 556, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 556, + 655 + ], + "type": "text", + "content": "[46] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12009-12019, 2022. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 556, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 556, + 689 + ], + "type": "text", + "content": "[47] Yuen Peng Loh and Chee Seng Chan. Getting to know low-light images with the exclusively dark dataset. CoRR, abs/1805.11227, 2018. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 556, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 556, + 713 + ], + "type": "text", + "content": "[48] Rundong Luo, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Similarity min-max: Zero-shot day-night domain adaptation." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5654" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8104-8114, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[49] Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Yanyi Liu, Shilong Zhang, and Kai Chen. Rtmdet: An empirical study of designing real-time object detectors. arXiv preprint arXiv:2212.07784, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 196 + ], + "type": "text", + "content": "[50] Long Ma, Tengyu Ma, Risheng Liu, Xin Fan, and Zhongxuan Luo. Toward fast, flexible, and robust low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5637-5646, 2022. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 296, + 262 + ], + "type": "text", + "content": "[51] Tengyu Ma, Long Ma, Xin Fan, Zhongxuan Luo, and Risheng Liu. PIA: parallel architecture with illumination accelerator for joint enhancement and detection in low-light. In MM '22: The 30th ACM International Conference on Multimedia, Lisboa, Portugal, October 10 - 14, 2022, pages 2070-2078. ACM, 2022. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 308 + ], + "type": "text", + "content": "[52] Igor Morawski, Yu-An Chen, Yu-Sheng Lin, and Winston H. Hsu. NOD: taking a closer look at detection under extreme low-light conditions with night object detection dataset. CoRR, abs/2110.10364, 2021. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 296, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 296, + 387 + ], + "type": "text", + "content": "[53] Lukáš Neumann, Michelle Karg, Shanshan Zhang, Christian Scharfenberger, Eric Piegert, Sarah Mistr, Olga Prokofyeva, Robert Thiel, Andrea Vedaldi, Andrew Zisserman, et al. Nightowls: A pedestrians at night dataset. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part I 14, pages 691–705. Springer, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 430 + ], + "type": "text", + "content": "[54] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 296, + 487 + ], + "type": "text", + "content": "[55] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 296, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 296, + 543 + ], + "type": "text", + "content": "[56] Qingpao Qin, Kan Chang, Mengyuan Huang, and Guiqing Li. Denet: Detection-driven enhancement network for object detection under adverse weather conditions. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 2813-2829, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 544, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 296, + 567 + ], + "type": "text", + "content": "[57] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. CoRR, abs/1804.02767, 2018. 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 296, + 622 + ], + "type": "text", + "content": "[58] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 1, 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 678 + ], + "type": "text", + "content": "[59] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Guided curriculum model adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7374-7383, 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "text", + "content": "[60] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 95 + ], + "type": "text", + "content": "IEEE/CVF International Conference on Computer Vision (ICCV), pages 10765-10775, 2021. 1, 2, 3, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 556, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 556, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 556, + 140 + ], + "type": "text", + "content": "[61] Yukihiro Sasagawa and Hajime Nagahara. Yolo in the dark - domain adaptation method for merging multiple models. In Computer Vision – ECCV 2020, pages 345–359, Cham, 2020. Springer International Publishing. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "type": "text", + "content": "[62] Shangquan Sun, Wenqi Ren, Tao Wang, and Xiaochun Cao. Rethinking image restoration for object detection. Advances in Neural Information Processing Systems, 35:4461-4474, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 555, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 219 + ], + "type": "text", + "content": "[63] Maria Valera and Sergio A Velastin. Intelligent distributed surveillance systems: a review. IEE Proceedings-Vision, Image and Signal Processing, 152(2):192-204, 2005. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 555, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 273 + ], + "type": "text", + "content": "[64] Wenjing Wang, Wenhan Yang, and Jiaying Liu. Hla-face: Joint high-low adaptation for low light face detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16195-16204, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 330 + ], + "type": "text", + "content": "[65] Wenjing Wang, Zhengbo Xu, Haofeng Huang, and Jiaying Liu. Self-aligned concave curve: Illumination enhancement for unsupervised adaptation. In Proceedings of the 30th ACM international conference on multimedia, pages 2617-2626, 2022. 2, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 555, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 375 + ], + "type": "text", + "content": "[66] Wenjing Wang, Xinhao Wang, Wenhan Yang, and Jiaying Liu. Unsupervised face detection in the dark. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1250-1266, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 377, + 555, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 555, + 430 + ], + "type": "text", + "content": "[67] Wenjing Wang, Huan Yang, Jianlong Fu, and Jiaying Liu. Zero-reference low-light enhancement via physical quadruple priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26057-26066, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 555, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 555, + 477 + ], + "type": "text", + "content": "[68] Xinzhe Wang, Kang Ma, Qiankun Liu, Yunhao Zou, and Ying Fu. Multi-object tracking in the dark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 382-392, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 479, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 555, + 510 + ], + "type": "text", + "content": "[69] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. CoRR, abs/1808.04560, 2018. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 513, + 555, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 555, + 545 + ], + "type": "text", + "content": "[70] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 546, + 555, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 555, + 611 + ], + "type": "text", + "content": "[71] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wenhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5901-5910, 2022. 1, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 668 + ], + "type": "text", + "content": "[72] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15769-15778, 2021. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[73] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17714-17724, 2022. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5655" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "text", + "content": "[74] Xiaogang Xu, Ruixing Wang, and Jiangbo Lu. Low-light image enhancement via structure modeling and guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9893-9903, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 193 + ], + "type": "text", + "content": "[75] Yuecong Xu, Jianfei Yang, Haozhi Cao, Kezhi Mao, Jianxiong Yin, and Simon See. Arid: A new dataset for recognizing action in the dark. In Deep Learning for Human Activity Recognition: Second International Workshop, DL-HAR 2020, Held in Conjunction with IJCAI-PRICAI 2020, Kyoto, Japan, January 8, 2021, Proceedings 2, pages 70–84. Springer, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 295, + 250 + ], + "type": "text", + "content": "[76] Xinwei Xue, Jia He, Long Ma, Yi Wang, Xin Fan, and Risheng Liu. Best of both worlds: See and understand clearly in the dark. In Proceedings of the 30th ACM International Conference on Multimedia, page 2154–2162, New York, NY, USA, 2022. Association for Computing Machinery. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 296, + 304 + ], + "type": "text", + "content": "[77] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 305, + 296, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 296, + 349 + ], + "type": "text", + "content": "[78] Shuo Yang, Ping Luo, Chen-Change Loy, and Xiaou Tang. Wider face: A face detection benchmark. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5525-5533, 2016. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "text", + "content": "[79] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12918-12927, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 405, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 296, + 602 + ], + "type": "text", + "content": "[80] Wenhan Yang, Ye Yuan, Wenqi Ren, Jiaying Liu, Walter J. Scheirer, Zhangyang Wang, Taiheng Zhang, Qiaoyong Zhong, Di Xie, Shiliang Pu, Yuqiang Zheng, Yanyun Qu, Yuhong Xie, Liang Chen, Zhonghao Li, Chen Hong, Hao Jiang, Siyuan Yang, Yan Liu, Xiaochao Qu, Pengfei Wan, Shuai Zheng, Minhui Zhong, Taiyi Su, Lingzhi He, Yandong Guo, Yao Zhao, Zhenfeng Zhu, Jinxiu Liang, Jingwen Wang, Tianyi Chen, Yuhui Quan, Yong Xu, Bo Liu, Xin Liu, Qi Sun, Tingyu Lin, Xiaochuan Li, Feng Lu, Lin Gu, Shengdi Zhou, Cong Cao, Shifeng Zhang, Cheng Chi, Chubing Zhuang, Zhen Lei, Stan Z. Li, Shizheng Wang, Ruizhe Liu, Dong Yi, Zheming Zuo, Jianning Chi, Huan Wang, Kai Wang, Yixiu Liu, Xingyu Gao, Zhenyu Chen, Chang Guo, Yongzhou Li, Huicai Zhong, Jing Huang, Heng Guo, Jianfei Yang, Wenjuan Liao, Jiangang Yang, Liguo Zhou, Mingyue Feng, and Likun Qin. Advancing image understanding in poor visibility environments: A collective benchmark study. IEEE Transactions on Image Processing, 29:5737-5752, 2020. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 603, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 635 + ], + "type": "text", + "content": "[81] Bo Zhang, Yuchen Guo, Runzhao Yang, Zhihong Zhang, Jiayi Xie, Jinli Suo, and Qionghai Dai. Darkvision: A benchmark for low-light image/video perception, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 636, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 690 + ], + "type": "text", + "content": "[82] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Heung-Yeung Shum. DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[83] Liheng Zhang, Guo-Jun Qi, Liqiang Wang, and Jiebo Luo. Aet vs. aed: Unsupervised representation learning by auto" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 296 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "encoding transformations rather than data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2547-2555, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "type": "text", + "content": "[84] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. CoRR, abs/1905.04161, 2019. 1, 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "text", + "content": "[85] Zhaoyang Zhang, Yitong Jiang, Jun Jiang, Xiaogang Wang, Ping Luo, and Jinwei Gu. Star: A structure-aware lightweight transformer for real-time image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4106-4115, 2021. 1, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 197, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 251 + ], + "type": "text", + "content": "[86] Zhao Zhang, Huan Zheng, Richang Hong, Mingliang Xu, Shuicheng Yan, and Meng Wang. Deep color consistent network for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1899-1908, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 252, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 555, + 296 + ], + "type": "text", + "content": "[87] Hegui Zhu, Kai Wang, Ziwei Zhang, Yuelin Liu, and Wuming Jiang. Low-light image enhancement network with decomposition and adaptive information fusion. Neural Computing and Applications, 34(10):7733-7748, 2022. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5656" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/2c55e2cf-c2bf-4623-a4a2-6a1a33a77533_content_list.json b/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/2c55e2cf-c2bf-4623-a4a2-6a1a33a77533_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8188c899ffe035359657a3da6770b1f5bcb7773e --- /dev/null +++ b/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/2c55e2cf-c2bf-4623-a4a2-6a1a33a77533_content_list.json @@ -0,0 +1,1588 @@ +[ + { + "type": "text", + "text": "Toward Better Out-painting: Improving the Image Composition with Initialization Policy Model", + "text_level": 1, + "bbox": [ + 151, + 128, + 848, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xuan Han $^{1}$ Yihao Zhao $^{1}$ Yanhao Ge $^{1,2}$ Mingyu You $^{1*}$ $^{1}$ Tongji University, Shanghai, China \n $^{2}$ Future Imaging Area \n{hanxuan, zhaoyihao, myyou}@tongji.edu.cn {halege}@vivo.com", + "bbox": [ + 225, + 202, + 776, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With its extensive applications, Foreground Conditioned Out-painting (FCO) has attracted considerable attention in the research field. Through the utilization of text-driven FCO, users are enabled to generate diverse backgrounds for a given foreground by adjusting the text prompt, which considerably enhances the efficiency in fields like e-commerce. Since the foreground is fixed in FCO, a key concern is whether the generated background can match the foreground well to achieve a coherent composition. However, most existing methods are lacking in this regard. Artifacts and incorrect interactions are common defects in synthesized images. This issue is linked to the influence of the initial noise in the sampling process. As the initial noise is sampled independently, it's highly likely that the implied image composition will conflict with the given foreground. In this paper, a novel Initialization Policy Model (IPM) is proposed to address this problem. Its function is to replace the early denoising steps and directly predict the intermediate state that is conducive to the reasonable image composition. Since the IPM is designed to take only the foreground image and the text prompt as inputs, it isolates the impact of the initial noise. The subsequently proposed training paradigm that combines inversion-derived label supervision and diffusion reward supervision further ensures the efficient training of the IPM. The evaluation is conducted using the task-specific OpenImage-FCO dataset developed by us. The results verify that the introduction of the IPM can significantly improve the composition of the synthesized images and achieve advanced performance in the FCO task.", + "bbox": [ + 88, + 323, + 485, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 792, + 220, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nowadays, diffusion-based image editing has emerged as a trending research topic. Its progress in numerous tech-", + "bbox": [ + 89, + 819, + 482, + 851 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd8241776b463b7c8b42d48a7f02b84badb2ea3d5c6ee15389caaef35349f84e.jpg", + "image_caption": [ + "Figure 1. The illustrations concerning the flawed image composition in foreground conditioned out-painting task." + ], + "image_footnote": [], + "bbox": [ + 517, + 295, + 905, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "niques is continuously revolutionizing the practices in related fields such as commercial art [7, 11] and fashion design [6, 33]. Among the various techniques available, Foreground Conditioned Out-painting (FCO) [2, 14, 26] has attracted significant attention. FCO focuses on a special use case: The foreground of an image, usually containing the main object, is fixed, and image generators like Stable Diffusion [26] are used to fill the background based on the user's prompt. This technology allows users to generate diverse backgrounds for the given foreground simply by changing the text prompt, which significantly enhances efficiency compared to standard workflows.", + "bbox": [ + 511, + 491, + 906, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this task, a crucial concern is the quality of image composition. Take Fig. 1 as an example: the user wants to display the mug in a fresh scene. The expected result is that the synthesized background can be harmoniously integrated with the mug. However, existing methods demonstrate insufficient performance in this regard. In many results, background objects like the table fail to interact appropriately with the given mug, alongside artifacts generated in the process. Overall, the synthesized background often exists independently of the given foreground rather than forming a unified composition.", + "bbox": [ + 511, + 672, + 908, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To identify the source of the problem, we revisit the specificity of the FCO task. In diffusion-based FCO, images are gradually denoised from initial noise $x_{T}$ . According to existing literature [27, 35], variations in $x_{T}$ compre", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author. M. You is with the College of Electronic and Information Engineering, Tongji University, State Key Laboratory of Autonomous Intelligent Unmanned Systems", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16938", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hensively influence the position, scale, and pose of synthesized objects. However, in FCO, independently sampled $x_{T}$ frequently implies compositions conflicting with the fixed foreground. As shown in Fig. 1, different $x_{T}$ values lead to diverse table poses, many of which clash with the mug; artifacts are also generated to match these inconsistent backdrops. Evidently, addressing this issue requires mitigating the random impact of $x_{T}$ on image composition.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, a novel Initialization Policy Model (IPM) is introduced to address the problem. The method exploits the time-frequency coupling of the diffusion process, by which low-frequency signals determining image composition are primarily generated during early denoising steps [13]. In the proposed approach, the denoising process is truncated at timestep $t^s$ , with the IPM directly predicting the intermediate state $x_{t^s}$ . By conditioning solely on the foreground $c$ and text $y$ , IPM isolates image low-frequency signals from random $x_T$ influence. Subsequent normal denoising processes synthesize rich high-frequency details thereon. To facilitate the optimization of IPM, a comprehensive training paradigm integrating label supervision and reward supervision is proposed. Label supervision uses inversion-derived low-frequency signal ground truth to guide IPM output. Reward supervision subjects predicted $x_{t^s}$ to multiple denoising rounds, with the error between the final image and target image serving as the training reward criterion. We validate the method on our task-specific OpenImage-FCO dataset. The results confirm that it can notably improve the image composition in FCO task. The contributions of this work are as follows:", + "bbox": [ + 91, + 212, + 483, + 542 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel Initialization Policy Model (IPM) to enhance image composition in the FCO task. By using the IPM to directly predict the low-frequency signals that match the given foreground, it effectively suppresses the random influence of the initial noise, thus alleviating the issue of flawed composition.", + "- We devise a comprehensive training paradigm for the IPM. In addition to label supervision, we introduce novel reward supervision based on a pre-trained generative model, which has been proven to effectively increase the quality of the signals predicted by the IPM, resulting in the final image with reduced artifacts and improved detail accuracy.", + "- We construct the OpenImage-FCO dataset based on OpenImage [3] and an automated annotation pipeline. It consists of approximately 580K images, 800K filtered high-quality masks, and their associated tags, providing substantial support for research in the FCO field." + ], + "bbox": [ + 89, + 546, + 482, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 829, + 238, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging the abundant priors of large pre-trained models [8, 25, 26], diffusion-based approaches have achieved remarkable progress in many image editing tasks such as", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "object editing [4, 9, 29, 32], attribute editing [9, 19], and structure control [5, 30, 44]. This paper focuses on the Foreground Conditioned Out-painting (FCO) task. Related methods can be categorized into two types: learning-free and learning-based approaches. In this section, we introduce representative works from both categories and analyze their application characteristics.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning-Free Method In learning-free methods, the most prominent one is Blended latent Diffusion [1, 2]. In this method, the unknown region (background) is sampled through the standard denoising process. After each step, the known region (foreground) is pasted into the current result. This straightforward pipeline allows it to be plug-and-play in any pre-trained model. However, since the two regions are not acquired from the unified process, it is challenging to achieve a seamless combination between them. Subsequent research [10, 20] has strived to refine the pipeline. For example, DiffEdit [10] employs DDIM Inversion [23] to obtain the features of the known region, thereby reducing the distribution disparity between the features of the two regions. Although the related research is still advancing, at present, the mainstream learning-free methods still adopt the above spatially separated sampling paradigm. A common concern is that it hinders the feature interaction between the two regions, resulting in the inability to achieve mask-aware image completion [14]. With respect to the FCO task, the background synthesized by this type of method often struggles to exhibit a reasonable interaction with the given foreground.", + "bbox": [ + 511, + 200, + 906, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning-based Method Many studies [14, 37, 44, 46] to improve the synthesis effect through task-specific learning. A representative one is SD-Inpainting [26]. It fine-tunes the pre-trained model with a large-scale in-painting dataset and can be directly used for various image completion applications. Since the fine-tuning process involves most of the parameters of Stable Diffusion, the original diffusion prior in it is also affected to a certain extent. Subsequent research [40, 42] has developed a more refined framework. For example, PowerPoint [46] trains an additional task prompt to enhance the model's ability to fill images. After that, BrushNet [14] introduces a twin denoising network to process the features of the known region and optimizes the entire framework end-to-end. One of the most concerned solutions among them is ControlNet [44]. This method trains additional input branches to achieve various conditional generation tasks. When dealing with image completion, the new branch takes the RGB image and mask of the known area as the input. The feature of parameter invariance in the original network enables its direct combination with other methods.", + "bbox": [ + 511, + 534, + 908, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It must be pointed out that many studies mentioned above take general image completion rather than FCO as the major task. This is also reflected in the data they use. For example, the masks in the dataset of SD-Inpainting", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16939", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/04692cb158d94516b67758199e09daad01b183bb08f8a61843da6564cbbb3f89.jpg", + "image_caption": [ + "Figure 2. (a) The illustration about the differences between the proposed truncated diffusion process and the original diffusion process. (b) The principle and training paradigm of Initialization Policy Model (IPM)." + ], + "image_footnote": [], + "bbox": [ + 94, + 95, + 433, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7e90ca50fb8a810f2a0c70266bfb7b3f037f4429982d9a86f4b7612fcfe04186.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 447, + 95, + 903, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[26] are mostly randomly generated, which implies that the known region would contain some clues about the unknown part. But in the FCO task, the known region is the well-segmented foreground object. These differences will undoubtedly cause a series of issues when these methods are applied to FCO. The flawed image composition is one of the most significant problems. In this work, we conduct in-depth research on the FCO task and propose targeted solutions to improve the composition of synthesized images.", + "bbox": [ + 88, + 357, + 483, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Truncated Diffusion Model", + "text_level": 1, + "bbox": [ + 89, + 506, + 344, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 531, + 200, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper centers on the FCO task, with the aim of enhancing the image composition of generated images. The research primarily focuses on diffusion-based methods, where the image generation process is modeled as a series of sequential denoising steps commencing from random noise $x_{T}$ . Here, $T$ denotes the number of timesteps. Specifically for text-driven FCO, this denoising process is guided by both the text prompt $y$ and the condition image $c$ . The text prompt $y$ describes the semantics of the target image, while the condition image $c$ is typically the RGB image of the foreground we intend to present. The diffusion-based FCO model can be formally expressed as $p_{\\theta}(x_{0:T-1}|x_{T},y,c)$ , where $\\theta$ represents the denoising model.", + "bbox": [ + 88, + 551, + 482, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method originates from rethinking the influence of $x_{T}$ on image composition. According to existing research, even when using the same prompt, the results sampled from different $x_{T}$ values will differ. The pose, scale, and position of each synthesized object will change. Given that $x_{T}$ is completely independent, the image composition it implies is likely to conflict with the given foreground. As depicted in Fig. 2(a), we propose truncating the denoising process at a specific time step $t^{s}$ and substituting the early steps with an Initialization Policy Model (IPM) $\\phi$ . This method exploits", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the time-frequency coupling in the diffusion process, where the low-frequency signals that determine the image composition are mainly formed in the early stage. $\\phi$ is defined to be solely conditioned on the foreground image $c$ and the text prompt $y$ , thereby isolating the random impact of $x_{T}$ on the image composition. As a result, the original model is disassembled into $p_{\\phi}(x_{ts}|y,c)$ and $p_{\\theta}(x_{0:t^s +1}|x_{ts},y,c)$ . In this section, we first introduce the principle and network design of the IPM, followed by the detailed presentation of each component in the training paradigm.", + "bbox": [ + 511, + 357, + 906, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Initialization Policy Model", + "text_level": 1, + "bbox": [ + 511, + 516, + 754, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed method introduces a new IPM to obtain the intermediate state $x_{t^s}$ . According to the principle of the diffusion model, each intermediate state $x_{t}$ follows the distribution [31]:", + "bbox": [ + 511, + 537, + 906, + 599 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} \\sim \\mathcal {N} \\left(x _ {t}; \\sqrt {\\alpha_ {t}} \\tilde {x} _ {0}, (1 - \\alpha_ {t}) \\mathbf {I}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 614, + 906, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\alpha_{t}$ the parameter of the diffusion schedule. $\\tilde{x}_0$ represents the prediction of the target image $x_0$ at the current timestep, essentially serving as the deterministic signal after denoising steps from $T$ to $t - 1$ . To better clarify the meaning of each parameter, we rewrite Eq. (1) as follows:", + "bbox": [ + 511, + 635, + 905, + 712 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\underbrace {\\sqrt {\\alpha_ {t}} \\tilde {x} _ {0}} _ {\\text {S i g n a l}} + \\underbrace {\\sqrt {1 - \\alpha_ {t}} \\epsilon} _ {\\text {N o i s e}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 724, + 906, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When $t$ is large, $\\tilde{x}_0$ primarily contains low-frequency signals [12], which serve as the key determinants of image composition. $\\epsilon$ represents random noise, and different $\\epsilon$ lead to distinct high-frequency details in generated images. In this method, $\\phi$ is tasked with predicting the signal component $\\tilde{x}_0$ . During training, the IPM is required to learn the low-frequency signals corresponding to real target images. During sampling, its output is combined with random noise to generate diverse images with coherent compositions.", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16940", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notably, theoretically, there exists a broad range of potential compositions compatible with given $c$ and $y$ , and the ideal $x_{t^s}$ is expected to follow a complex distribution. However, in practice, it is challenging to obtain multiple real target images for a specific pair of $c$ and $y$ . Thus, the role of the IPM is to provide a specific cluster within the ideal distribution of $x_{t^s}$ . In the experiments section, we further discuss its impact on the original image generation model and how it helps the FCO model achieve a balance between diversity and coherence.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regarding the network architecture, the IPM $\\phi$ is directly initialized from the U-Net of Stable Diffusion [26]. This strategy aims to inherit the rich text-image priors embedded in its parameters. The text prompt $y$ (encoded by CLIP [24]) is integrated through cross-attention layers. The network's input layer is expanded with four additional channels to accommodate the condition image $c$ (encoded by the Latent VAE [26]). Furthermore, considering the U-Net was originally trained as a denoising model, a fixed noise strategy is introduced. Specifically, a static noise $x_{T}^{st}$ is fed into the original channels. The final prediction of the low-frequency signal is computed from the network output and $x_{T}^{st}$ according to the diffusion schedule: $\\sqrt{1 / \\alpha_{t^s}} x_T^{st} - \\sqrt{1 / \\alpha_{t^s} - 1}\\phi (y,c)$ . Since the schedule parameters and noise in this formula remain constant across training and sampling, we hereinafter use $\\phi (y,c)$ to denote the predicted low-frequency signal for simplicity.", + "bbox": [ + 89, + 244, + 486, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Inversion-derived Label Supervision", + "text_level": 1, + "bbox": [ + 89, + 523, + 406, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Supervision with reliable ground truth is one of the most effective training paradigms. For the prediction objectives of the IPM, the ground truth can be derived through the inversion of the diffusion process [15, 31, 38]. The inversion operation proceeds in the reverse direction of the sampling process, involving stepwise denoising from the target image $x_0$ to the corresponding initial noise $x_{T}$ . By adhering to the identical noise schedule, the trajectory of intermediate state $x_{t}$ obtained through inversion demonstrates high congruence with that of normal sampling operations, thereby qualifying it as valid ground truth. In this work, the widely adopted DDIM Inversion [31] is implemented, which operates within the range $t\\in [0,t^s]$ to compute the ground truth of the intermediate state at truncation timestep $x_{ts}^{gt}$ .", + "bbox": [ + 89, + 549, + 482, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the results of the inversion process, the ground truths of the signal and random noise can be further obtained: $\\tilde{x}_0^{gt} = \\sqrt{1 / \\alpha_{t^s}} x_{t^s} - \\sqrt{1 / \\alpha_{t^s} - 1}\\theta (x_t,y)$ , $\\epsilon^{gt} = \\theta (x_t,y)$ . Here, $\\theta$ represents the pretrained diffusion model. $\\tilde{x}_0^{gt}$ will be used as a label to supervise the training of the IPM, and the loss function is as follows:", + "bbox": [ + 89, + 765, + 482, + 854 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {l b l} = \\left\\| \\phi (y, c) - \\tilde {x} _ {0} ^ {g t} \\right\\| _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 881, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Diffusion Reward Supervision", + "text_level": 1, + "bbox": [ + 513, + 90, + 781, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our approach, the IPM is tasked with predicting the low-frequency signals of images. Solely using the MSE loss as $\\mathcal{L}_{lbl}$ may cause it to produce over-smoothed results [17]. Not to mention that results obtained via inversion operation inherently introduce deviations [15]. Therefore, we propose Diffusion Reward Supervision to assist in the training of the IPM. Its principle is illustrated in Fig. 2(b). The $x_{t^s}$ sampled by $\\phi$ will be input into the pre-trained diffusion model (reward model) $\\theta$ to perform multiple rounds of denoising, and the error between the resulting image and the target image $x_0$ will serve as the reward to supervise the training.", + "bbox": [ + 511, + 112, + 903, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The diffusion reward adopts the same fast denoising schedule as DDIM [31]. Assuming the number of denoising steps is $K$ , the sampling step size would be $\\lfloor (T - t^s) / K \\rfloor$ . To enhance the diversity of the sampling schedules, the step size of the first step is set to a random number within $[1, \\lfloor (T - t^s) / K \\rfloor]$ . Each denoising step can be formulated as follow:", + "bbox": [ + 511, + 279, + 903, + 382 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = \\sqrt {\\alpha_ {t - 1} / \\alpha_ {t}} x _ {t} - \\sqrt {\\alpha_ {t - 1} / \\alpha_ {t} - 1} \\theta (x _ {t}, y) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 409, + 903, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In principle, the gradients generated by the final reward can be back-propagated according to the chain rule to guide the training of $\\phi$ . The gradients between neighboring timesteps can be expressed as:", + "bbox": [ + 511, + 439, + 903, + 498 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\partial x _ {t - 1} = a _ {t} \\partial x _ {t} + b _ {t} \\partial \\theta (x _ {t}, y) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 513, + 903, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$a_{t}$ and $b_{t}$ are used to simplify the schedule parameters, where $a_{t} = \\sqrt{\\alpha_{t - 1} / \\alpha_{t}}$ and $b_{t} = \\sqrt{\\alpha_{t - 1} / \\alpha_{t} - 1}$ . Eq. (5) reveals a drawback of the diffusion reward: the propagation of the gradient requires repeated passage through the reward model $\\theta$ , which will occupy a large amount of video memory. To solve this problem, we refer to the work of Wu et al. [39] and use the technique of stop gradients here. The modified gradients between neighboring timesteps become:", + "bbox": [ + 511, + 537, + 906, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\partial x _ {t - 1} = a _ {t} \\partial x _ {t} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 672, + 903, + 688 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This operation reduces the accuracy of the gradients to some extent, but it also makes it possible to conduct the training with limited video memory. We use $msd(*)$ to represent the multi-step denoising operation, and the diffusion reward can be expressed as:", + "bbox": [ + 511, + 695, + 903, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r w d} = \\left\\| m s d \\left(\\sqrt {\\alpha_ {t ^ {s}}} \\phi (y, c) + \\sqrt {1 - \\alpha_ {t ^ {s}}} \\epsilon^ {g t}\\right) - x _ {0} \\right\\| _ {2} ^ {2} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 794, + 903, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\epsilon^{gt}$ is the noise corresponding to the target image obtained during the inversion process (see Sec. 3.3). The meaning of $\\mathcal{L}_{rwd}$ is that we expect the predicted low-frequency signal plus the ground truth of the noise to restore the corresponding target image. Experiments have", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16941", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a387335729bae23cef4e3dbb228a1c2edeb5560a39d03ca2c95dba8e64ad46c2.jpg", + "image_caption": [ + "Figure 3. The samples of generated images of the comparison methods. All compared methods are implemented with the recommended settings. $\\dagger$ means that the method is trained / finetuned with our dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "confirmed that $\\mathcal{L}_{rwd}$ can effectively enhance the quality of the signals predicted by the IPM, resulting in the final image with reduced artifacts and improved detail accuracy. The final loss function is as follows, where $\\lambda_{rwd}$ is the weight of the reward supervision term:", + "bbox": [ + 89, + 420, + 483, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {l b l} + \\lambda_ {r w d} \\mathcal {L} _ {r w d} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 513, + 480, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As described in Sec. 3.2, during sampling, the output of $\\phi$ is combined with random noise to obtain the intermediate state: $x_{ts} = \\sqrt{\\alpha_{ts}}\\phi (y,c) + \\sqrt{1 - \\alpha_t}\\epsilon$ , where $\\epsilon \\sim \\mathcal{N}(0,1)$ . Subsequent denoising steps are completed by conventional diffusion-based FCO. In this paper, a ControlNet trained on the same dataset is employed in this phase.", + "bbox": [ + 89, + 537, + 483, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 643, + 223, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Numerical experiments are conducted to evaluate the proposed method. This section begins with an introduction to the dataset and experimental setups. Afterward, the comparison between IPM and state-of-the-art methods is reported. Then, the impact of IPM on the generative model is discussed. Finally, ablation studies and hyperparameter analyses for the proposed method are presented.", + "bbox": [ + 89, + 670, + 483, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 89, + 787, + 186, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The advancement of the FCO technique demands high-quality task-specific datasets. Existing datasets, despite their large scale, often lack sufficient diversity in masked objects and their saliency in images. For instance, PowerPaint [46] utilized the OpenImage [3] segmentation subset (1.74M images, 2.8M masks) but focused on only 600", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "object classes, many of which were not the primary subjects in their images. To address this limitation, we introduce OpenImage-FCO, a novel task-specific dataset built on OpenImage-v7 [3].", + "bbox": [ + 511, + 420, + 906, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "OpenImage-FCO incorporates multi-modal filtering and semantic reasoning to enhance annotation quality. We leverage BLIP2 [18] and RAM [45] for image captioning and tagging, followed by semantic filtering via Llama3 [34] to retain only focused objects. GroundingDINO [21] and HQSAM [16] are used to generate precise object masks. The final dataset includes 580K images with 800K masks, 6K unique object tags, and image captions. For evaluation, we curated a test set of 250 cases, each equipped with manual foreground mask annotations and a text prompt. Its scale is aligned with EditBench [36], a widely adopted benchmark in the image editing domain, which encompasses 240 cases.", + "bbox": [ + 511, + 483, + 908, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 680, + 720, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details Our method is built upon Stable Diffusion V2.1 with an image size of $512 \\times 512$ . When training the IPM, the truncation timestep $t^s$ is set to 800, the diffusion reward parameter $K$ is set to 10, and $\\lambda_{rwd}$ is set to 1. The AdamW optimizer [22] is employed with a constant learning rate of $1 \\times 10^{-4}$ . Training is conducted on four Nvidia L40S GPUs with a batch size of 8 and a total of 400K iterations. The FCO ControlNet used for subsequent inference steps adopts the same structure as described in the original paper [44]. The AdamW optimizer is configured with a learning rate of $1 \\times 10^{-5}$ , a batch size of 8, and 200K iterations. Notably, all backbone parameters, including Cross-Attn layers, were frozen during Control-", + "bbox": [ + 511, + 703, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16942", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6168034b513bbedc4b15d86f6427bcdc455e03cb7cd4b4f505774fa0bed69d63.jpg", + "image_caption": [ + "Cond. 1", + "- A Siamese cat stands on a stone, surrounded by purple flowers. in a sunlit garden." + ], + "image_footnote": [], + "bbox": [ + 94, + 104, + 189, + 178 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/78daf4903e3afb1e64d63a6a030738fd9c163dd5c45f587687f4c20afacd2d66.jpg", + "image_caption": [ + "Cond. 2" + ], + "image_footnote": [], + "bbox": [ + 191, + 104, + 285, + 178 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/04a8937f0fa46ab3d8a2c619fb6a994f9478e80809622ce63812b10d932c1344.jpg", + "image_caption": [ + "Cond. 3" + ], + "image_footnote": [], + "bbox": [ + 287, + 104, + 383, + 178 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dceff543d3b90d98f36dd6aeeffe7235b01c8fa884596f6fc6c0a3d31d054a34.jpg", + "image_caption": [ + "Cond. 4" + ], + "image_footnote": [], + "bbox": [ + 385, + 104, + 480, + 178 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3b5e1e78d62cbdc1d0fa65405254b0f9655941cf398df6b4f27f8e4ebbea6dbb.jpg", + "image_caption": [ + "- A train speeds along the tracks, surrounded by lush green trees and hills." + ], + "image_footnote": [], + "bbox": [ + 94, + 190, + 187, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a3dac370d202ef8074b28f5afd65ce67379a24deef89afa313bd942f5522055e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 189, + 190, + 285, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aa0c5094d83ec7d107242c021e660927b11040e77382ee5643bc82391afd2b9b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 190, + 383, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f38ee668fff1e360d2735fd9e8101bd45d8356174494af4b05091ef2000f1d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 190, + 480, + 263 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7028898ee2daf61a550b04bb9f1b721183a1c3a1f0e274da309710e78ddb6010.jpg", + "image_caption": [ + "- A comfortable chair sits in the corner of a charmingly cozy living room.", + "Figure 4. The generated images of foreground objects with variable identities, scales, and positions. The low-frequency signals predicted by IPM are shown in the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 94, + 273, + 187, + 348 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/504b14b743ebcfc57333c03d30cbdd69678868dc9725fc9da8d3494382eb7370.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 189, + 273, + 285, + 348 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/765eb4c89557dd2446d8af5e11b75cb78881018e7641c3018ae4cbda99154461.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 273, + 383, + 348 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/90d228738ab98fb726ffa3b173a65fccb74f2d8c22b1fd819ea4c12b86b82891.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 273, + 480, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Net training to maximize the preservation of backbone text-image priors. All synthesized images presented in this article are sampled using the following strategy: After obtaining $x_{t^s}$ from the IPM, the DDIM sampler [31] completes the remaining sampling steps with a step size of 10 and a classifier-free guidance scale of 7.", + "bbox": [ + 88, + 440, + 482, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics We take into account 5 metrics from three aspects: image composition, text-image alignment and image quality. 1) Image Composition: The defective image composition is mainly manifested as artifacts and incorrect interactions. Specifically, GroundingDINO [21] and HQSAM [16] are used to detect all the masks that have the same semantics as the foreground objects. The size ratio between the overflow area and the ground-truth foreground is reported as the first metric, which is called SAM Score (SAM). The quality of the interaction is mainly measured by a user study (INT). Thirty individuals participate in the survey and assess the generated images regarding the object interaction. The scale ranges from 1 to 5, with a higher score indicating better results. 2) Text-Image Alignment: CLIP Similarity (CLIP) is reported to reflect the text-image alignment between the text prompt and the generated image. 3) Image Quality: Aesthetic Score (AS) [28] and Image Reward (IR) [41] are used to evaluate the image quality. These evaluation models are trained on human preference datasets and are considered to align well with the human perception.", + "bbox": [ + 89, + 532, + 482, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 89, + 847, + 398, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct comparisons with six representative related methods, including Blended Latent Diffusion (BLD) [2],", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2cf73e082c563d1b6d4669375ff5f4a6a5f206ab930c8a5103211dc077bbd11f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSAM↓INT↑CLIP↑IR↑AS↑
BLD [2]0.9861.7526.060.185.04
SD-Inpaint.[26]0.3173.3526.260.355.19
PowerPaint [46]0.8202.2326.130.345.05
BrushNet [14]0.1904.0626.620.605.31
Layer Diff. [43]0.0463.5026.150.225.27
ControlNet† [44]0.2353.8826.460.525.27
PowerPaint† [46]0.4333.2726.150.465.20
BrushNet† [14]0.1203.9426.560.585.29
Ours0.0704.6226.500.675.34
", + "bbox": [ + 517, + 88, + 903, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. The quantitative comparison of different methods. The items in bold and underline are ranked first and second. $\\dagger$ means that the method is trained / finetuned with our dataset.", + "bbox": [ + 511, + 287, + 906, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SD-Inpainting [26], PowerPoint [46], ControlNet [44], BrushNet [14] and Layer Diffusion [43]. Among them, all except BLD are learning-based methods. ControlNet is trained using the OpenImage-FCO dataset. For a fair comparison, we also fine-tuned BrushNet and PowerPaint using the proposed dataset. The intuitive and quantitative comparison results are shown in Fig. 3 and Tab. 1 respectively.", + "bbox": [ + 511, + 356, + 903, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the objective metric SAM, our method outperforms most comparison methods by at least 0.050. The only method that scores higher than our proposed method, Layer Diffusion, lacks competitiveness in other metrics due to poor coordination between the foreground and background. On the user-study metric INT, the average score given by users to our method was 0.56 higher than that of the second-best method. These results demonstrate the effectiveness of our method in improving image composition. The samples in Fig. 3 also intuitively illustrate this advantage. Among all compared methods, our approach produces artifact-free results across four cases. In terms of text-image alignment, our method achieves the second-best result. The first place is taken by BrushNet, partially due to its dataset. BrushNet was trained on LAION-Aesthetic [28] containing 120 million images, where abundant image-text pairs ensure advanced performance in text prompt controllability.", + "bbox": [ + 511, + 463, + 905, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The advantages in image composition also drive better performance in image quality metrics, with our method outperforming the second-best by 0.07 and 0.03 on IR and AS respectively. Notably, while suppressing artifacts and incorrect interactions, our method produces less cluttered backgrounds compared to others. This can be attributed to the distribution discrepancy between $x_{t^s}$ inferred by IPM and natural sampling. Fortunately, this characteristic also enhances visual hierarchy in synthesized images, making them more suitable for object presentation.", + "bbox": [ + 511, + 719, + 905, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A noteworthy result is that after fine-tuning on the proposed dataset, PowerPaint and BrushNet have shown signif", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16943", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d501d868eaf0710e5d86d6cb2364076ddb099898e1e5f3d73a809f9797c70f4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 87, + 480, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d13bbb518902ad94acab5b848d5c49d70cd10926e99d2b1eabf4363e164228b1.jpg", + "image_caption": [ + "Figure 5. (a) The generation results of the baseline model with or w/o IPM under different text prompts. (b) The multi generation results of the baseline model with or w/o IPM in same case." + ], + "image_footnote": [], + "bbox": [ + 89, + 261, + 480, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ed545122eb3927a872cf965dd86202118b181f8545af126dc6e84d257c3a25b4.jpg", + "image_caption": [ + "Figure 6. The samples of predicted low-frequency signals and final generated images with or w/o the diffusion reward loss $\\mathcal{L}_{rwd}$ ." + ], + "image_footnote": [], + "bbox": [ + 93, + 496, + 480, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "icant improvements in the SAM metric, which once again confirms the value of task-specific datasets. To further demonstrate the IPM's capabilities, we select foregrounds with diverse identities, scales, and positions as condition images and generate results under the same text prompt. The results in Fig. 4 show that our method adapts well to various scenarios and produces reasonable interactions between given objects and backdrops.", + "bbox": [ + 89, + 717, + 482, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. The Influence of IPM on Generation Model", + "text_level": 1, + "bbox": [ + 89, + 848, + 460, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given that the integration of IPM fundamentally alters the original diffusion process, its implications for the under", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3c1ceb69f2402a0b7b040973a4455dd77003ce33cd7cf52b342a2335a1e5b4c1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingSAM↓CLIP↑IR↑AS↑
Baseline0.23526.460.525.27
w/o L_rwd0.10326.440.645.26
w/o U-Net Init.0.07426.290.635.28
Ours0.07026.500.675.34
", + "bbox": [ + 547, + 88, + 874, + 188 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. The quantitative evaluation results of ablation study.", + "bbox": [ + 527, + 198, + 888, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lying generation model warrant detailed discussion. For image generation tasks, the primary concerns are semantic controllability and output diversity, both of which will be systematically addressed in this section.", + "bbox": [ + 511, + 242, + 903, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The generation results of the baseline model (ControlNet) with and without IPM under different text prompts $y$ are presented in Fig. 5(a). It can be observed that IPM does not significantly affect the model's controllability, as the model still synthesizes results with the required semantics. This finding is also supported by Tab. 1, where our method and ControlNet exhibit the closest CLIP scores among all approaches. The subsequent ablation studies will reveal that this benefit primarily stems from initializing IPM parameters with the pretrained U-Net. Turning to Fig. 5(b), the generation results of the baseline model with and without IPM for the same pair of $c$ and $y$ are shown. In this comparison, IPM results show more similar composition due to the shared low-frequency signals, while the baseline model's diverse outputs come at the cost of prevalent compositional defects. This reinforces that IPM trades inefficient diversity for compositional rationality.", + "bbox": [ + 511, + 303, + 906, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 573, + 666, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct ablation studies on critical components of the IPM training paradigm and network design, including the diffusion reward supervision $\\mathcal{L}_{rwd}$ and the strategy of initializing IPM parameters with a pretrained U-Net. Note that since removing $\\mathcal{L}_{lbl}$ would destabilize training, we exclude it from ablation analysis. The experimental results are presented in Tab. 2.", + "bbox": [ + 511, + 597, + 905, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study for $\\mathcal{L}_{rwd}$ Quantitative validation shows that diffusion reward supervision provides the most significant benefit in artifact suppression. When $\\mathcal{L}_{rwd}$ is removed, the SAM score increases by 0.033. The introduction of new artifacts also degrades image quality metrics, with AS dropping by 0.08. These artifacts correlate with the quality of low-frequency signals predicted by IPM. To visually corroborate this, we visualize these low-frequency signals (decoded via the Latent VAE) in Fig. 6. It can be observed that without $\\mathcal{L}_{rwd}$ , IPM outputs smoother signals. A direct consequence is that edges of some foreground objects become indistinct, which may develop into artifacts and imprecise details after subsequent denoising steps. Conversely, when", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16944", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ff068861785720bf7260705db46e0f920b9a7275b06f459da6d08f8e2fd394a9.jpg", + "image_caption": [ + "- A brand-new football is placed on the grass in the football field." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 480, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/66ec616b3cd2c81fda55ea260feb5eab73f4cd2c26a87f52bac0465a526e7b72.jpg", + "image_caption": [ + "Figure 7. The samples of generated images under different $t^s$ .", + "Figure 8. The quantitative evaluation results of the truncating timestep selection." + ], + "image_footnote": [], + "bbox": [ + 94, + 224, + 478, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "diffusion reward supervision is introduced, IPM is encouraged to output intermediate states that genuinely lead to target images, effectively mitigating this issue.", + "bbox": [ + 89, + 446, + 482, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study for U-Net Initialization As shown in Tab. 2, initializing IPM parameters with a pretrained U-Net primarily strengthens the model's text controllability. Removing this strategy causes the CLIP score to drop by 0.21, even below the baseline performance. Although IPM is only responsible for predicting low-frequency signals, training it from scratch limits its performance on complex text prompts due to the dataset's relatively limited corpus. This phenomenon substantiates the necessity of leveraging the rich text-image priors of pretrained models.", + "bbox": [ + 88, + 492, + 482, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. The Selection of Truncating Timestep", + "text_level": 1, + "bbox": [ + 89, + 652, + 415, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The primary hyperparameter when applying IPM is determining the appropriate timestep $t^s$ to truncate the diffusion process. We test values within the set {900, 850, 800, 750, 700}. The intuitive and quantitative comparison results are shown in Fig. 7 and Fig. 8.", + "bbox": [ + 89, + 674, + 482, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When the truncation timestep $t^s$ is set too late, the method's effectiveness declines significantly. In the $t^s = 700$ experiment group, although artifacts are further suppressed, this comes at the cost of overly simplified backgrounds and significantly reduced image quality. Its IR and AS metrics are 0.35 and 0.60 lower than those of $t^s = 800$ . This phenomenon is associated with the function of IPM. IPM directly predicts the image's low-frequency signal. When $t^s$ is set to a smaller value, IPM's task becomes more complex. Meanwhile, the number of subsequent normal de", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c401878804ca700c5644dd8c35a527b499d87851291814286201d76878db880.jpg", + "image_caption": [ + "Figure 9. The samples of failure case about the objects with extreme colors or lighting conditions." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 903, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "noising steps is reduced, requiring IPM to predict image signals across broader frequency bands. $t^s = 700$ is clearly an excessively challenging value. As illustrated in Fig. 7, the produced image exhibits obvious detail deficiencies due to the interplay of these factors. When $t^s = 900$ , the frequency band managed by IPM is too narrow, weakening its influence on final image. Its SAM score increases by 0.033 compared to $t^s = 800$ , indicating that subsequent denoising steps introduce more suboptimal compositions. In this paper, $t^s = 800$ is adopted across all experiments. g", + "bbox": [ + 511, + 253, + 906, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Failure case and Future Work", + "text_level": 1, + "bbox": [ + 511, + 414, + 779, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During testing, a notable failure case arises when the given foreground object exhibits extreme colors or lighting conditions, significantly degrading background controllability. Fig. 9 illustrates results for such cases. Despite prompts' requirements, synthesized images still adopt styles consistent with the provided foreground (case 1) or fail to include other backdrops (case 2). Root causes include dataset limitations (artistic photography underrepresented in OpenImage) and framework design. That is, features of objects with extreme colors tend to dominate background features during synthesis, compelling background features to always reflect similar appearances. This motivates exploring methods to decouple foreground-background color distributions and enable foreground relighting in FCO tasks.", + "bbox": [ + 511, + 436, + 906, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper focuses on the problem of flawed image composition in diffusion-based FCO methods. We believe that this problem can be partially attributed to the random impact of the initial noise. Inspired by this, we propose the Initialization Policy Model (IPM). It uses a deterministic model that only depends on the given foreground and text prompt to replace the early denoising steps, thus isolating the impact of the initial noise. The subsequently proposed comprehensive training paradigm ensures that the IPM can be effectively optimized. Experiments confirm that the proposed method can significantly improve the composition problems such as artifacts and incorrect interactions, and the comprehensive framework built upon IPM also achieves advanced performance in FCO task.", + "bbox": [ + 511, + 689, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16945", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant No. 62473290, 62088101, National Key Research and Development Program of China under Grant No. 2024YFB3311801, and the Shanghai Municipal Commission of Economy and Informatization under Grant No. 2024-GZL-RGZN-01008.", + "bbox": [ + 89, + 90, + 485, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 208, + 187, + 224 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18187-18197, 2022. 2", + "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM TOG, 42(4):149:1-149:11, 2023. 1, 2, 6", + "[3] Rodrigo Benenson and Vittorio Ferrari. From colouring-in to pointillism: revisiting semantic segmentation supervision. CoRR, abs/2210.14142, 2022. 2, 5", + "[4] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 2", + "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In CVPR, pages 22503-22513, 2023. 2", + "[6] Shidong Cao, Wenhao Chai, Shengyu Hao, Yanting Zhang, Hangyue Chen, and Gaoang Wang. Diffashion: Reference-based fashion design with structure-aware transfer by diffusion models. TMM, 26:3962-3975, 2024. 1", + "[7] Tingfeng Cao, Junsheng Kong, Xue Zhao, Wenqing Yao, Junwei Ding, Jinhui Zhu, and Jiandong Zhang. Product2img: Prompt-free e-commerce product background generation with diffusion model and self-improved lmm. In ACMMM, pages 10774-10783, 2024. 1", + "[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, pages 1–31, 2024. 2", + "[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, pages 6593-6602, 2024. 2", + "[10] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. CoRR, abs/2210.11427, 2022. 2", + "[11] Amir Erfan Eshratifar, Joao VB Soares, Kapil Thadani, Shaunak Mishra, Mikhail Kuznetsov, Yueh-Ning Ku, and Paloma De Juan. Salient object-aware background generation using text-guided diffusion models. In CVPR, pages 7489-7499, 2024. 1", + "[12] Xuan Han, Yihao Zhao, and Mingyu You. Scene diffusion: Text-driven scene image synthesis conditioning on a single 3d model. In ACMMM, pages 7862-7870, 2024. 3", + "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2" + ], + "bbox": [ + 93, + 233, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. CoRR, abs/2403.06976, 2024. 1, 2, 6", + "[15] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. In ICLR, pages 1-28, 2024. 4", + "[16] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In NeurIPS, pages 29914–29934, 2024. 5, 6", + "[17] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew P. Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, pages 105-114, 2017. 4", + "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742, 2023. 5", + "[19] Pengzhi Li, Qinxuan Huang, Yikang Ding, and Zhiheng Li. Layerdiffusion: Layered controlled image editing with diffusion models. In ACM SIGGRAPH Asia, pages 1-4, 2023. 2", + "[20] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, and Jianzhuang Liu. Zone: Zero-shot instruction-guided local editing. In CVPR, pages 6254-6263, 2024. 2", + "[21] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: marrying DINO with grounded pre-training for open-set object detection. CoRR, abs/2303.05499, 2023. 5, 6", + "[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, pages 1-11, 2019. 5", + "[23] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 2", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 4", + "[25] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. 2", + "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685, 2022. 1, 2, 3, 4, 6", + "[27] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In AAAI, pages 4695-4703, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16946", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, pages 1–21, 2022. 6", + "[29] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In CVPR, pages 8871-8879, 2024. 2", + "[30] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent Y. F. Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In CVPR, pages 8839-8849, 2024. 2", + "[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, pages 1-12, 2021. 3, 4, 6", + "[32] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In CVPR, pages 18310-18319, 2023. 2", + "[33] Zhengwentai Sun, Yanghong Zhou, Honghong He, and PY Mok. Sgdiff: A style guided diffusion model for fashion synthesis. In ACMMM, pages 8433-8442, 2023. 1", + "[34] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, and Faisal Azhar. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971, 2023. 5", + "[35] Ruoyu Wang, Yongqi Yang, Zhihao Qian, Ye Zhu, and Yu Wu. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, pages 1-21, 2024. 1", + "[36] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, pages 18359-18369, 2023. 5", + "[37] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, pages 18359-18369, 2023. 2", + "[38] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In ICCV, pages 7378-7387, 2023. 4", + "[39] Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, and Hongsheng Li. Deep reward supervisions for tuning text-to-image diffusion models. CoRR, abs/2405.00760, 2024. 4" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437, 2023. 2", + "[41] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In NeurIPS, pages 1-14, 2023. 6", + "[42] Shiyuan Yang, Xiaodong Chen, and Jing Liao. Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In ACM MM, page 3190-3199, 2023. 2", + "[43] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. ACM Transactions on Graphics (TOG), 43(4):1-15, 2024. 6", + "[44] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3813-3824, 2023. 2, 5, 6", + "[45] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize anything: A strong image tagging model. In CVPR, pages 1724-1732, 2024. 5", + "[46] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. CoRR, abs/2312.03594, 2023. 2, 5, 6" + ], + "bbox": [ + 516, + 90, + 903, + 472 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16947", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/2c55e2cf-c2bf-4623-a4a2-6a1a33a77533_model.json b/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/2c55e2cf-c2bf-4623-a4a2-6a1a33a77533_model.json new file mode 100644 index 0000000000000000000000000000000000000000..edc4511735b2b4accf4c209b29a692888db6ad16 --- /dev/null +++ b/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/2c55e2cf-c2bf-4623-a4a2-6a1a33a77533_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.13, + 0.849, + 0.177 + ], + "angle": 0, + "content": "Toward Better Out-painting: Improving the Image Composition with Initialization Policy Model" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.203, + 0.777, + 0.258 + ], + "angle": 0, + "content": "Xuan Han\\(^{1}\\) Yihao Zhao\\(^{1}\\) Yanhao Ge\\(^{1,2}\\) Mingyu You\\(^{1*}\\) \n\\(^{1}\\) Tongji University, Shanghai, China \n\\(^{2}\\) Future Imaging Area \n{hanxuan, zhaoyihao, myyou}@tongji.edu.cn {halege}@vivo.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.324, + 0.486, + 0.765 + ], + "angle": 0, + "content": "With its extensive applications, Foreground Conditioned Out-painting (FCO) has attracted considerable attention in the research field. Through the utilization of text-driven FCO, users are enabled to generate diverse backgrounds for a given foreground by adjusting the text prompt, which considerably enhances the efficiency in fields like e-commerce. Since the foreground is fixed in FCO, a key concern is whether the generated background can match the foreground well to achieve a coherent composition. However, most existing methods are lacking in this regard. Artifacts and incorrect interactions are common defects in synthesized images. This issue is linked to the influence of the initial noise in the sampling process. As the initial noise is sampled independently, it's highly likely that the implied image composition will conflict with the given foreground. In this paper, a novel Initialization Policy Model (IPM) is proposed to address this problem. Its function is to replace the early denoising steps and directly predict the intermediate state that is conducive to the reasonable image composition. Since the IPM is designed to take only the foreground image and the text prompt as inputs, it isolates the impact of the initial noise. The subsequently proposed training paradigm that combines inversion-derived label supervision and diffusion reward supervision further ensures the efficient training of the IPM. The evaluation is conducted using the task-specific OpenImage-FCO dataset developed by us. The results verify that the introduction of the IPM can significantly improve the composition of the synthesized images and achieve advanced performance in the FCO task." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.794, + 0.222, + 0.809 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.82, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Nowadays, diffusion-based image editing has emerged as a trending research topic. Its progress in numerous tech-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.296, + 0.906, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.436, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Figure 1. The illustrations concerning the flawed image composition in foreground conditioned out-painting task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.674 + ], + "angle": 0, + "content": "niques is continuously revolutionizing the practices in related fields such as commercial art [7, 11] and fashion design [6, 33]. Among the various techniques available, Foreground Conditioned Out-painting (FCO) [2, 14, 26] has attracted significant attention. FCO focuses on a special use case: The foreground of an image, usually containing the main object, is fixed, and image generators like Stable Diffusion [26] are used to fill the background based on the user's prompt. This technology allows users to generate diverse backgrounds for the given foreground simply by changing the text prompt, which significantly enhances efficiency compared to standard workflows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.84 + ], + "angle": 0, + "content": "In this task, a crucial concern is the quality of image composition. Take Fig. 1 as an example: the user wants to display the mug in a fresh scene. The expected result is that the synthesized background can be harmoniously integrated with the mug. However, existing methods demonstrate insufficient performance in this regard. In many results, background objects like the table fail to interact appropriately with the given mug, alongside artifacts generated in the process. Overall, the synthesized background often exists independently of the given foreground rather than forming a unified composition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To identify the source of the problem, we revisit the specificity of the FCO task. In diffusion-based FCO, images are gradually denoised from initial noise \\( x_{T} \\). According to existing literature [27, 35], variations in \\( x_{T} \\) compre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Corresponding Author. M. You is with the College of Electronic and Information Engineering, Tongji University, State Key Laboratory of Autonomous Intelligent Unmanned Systems" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16938" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "hensively influence the position, scale, and pose of synthesized objects. However, in FCO, independently sampled \\( x_{T} \\) frequently implies compositions conflicting with the fixed foreground. As shown in Fig. 1, different \\( x_{T} \\) values lead to diverse table poses, many of which clash with the mug; artifacts are also generated to match these inconsistent backdrops. Evidently, addressing this issue requires mitigating the random impact of \\( x_{T} \\) on image composition." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.213, + 0.485, + 0.543 + ], + "angle": 0, + "content": "In this paper, a novel Initialization Policy Model (IPM) is introduced to address the problem. The method exploits the time-frequency coupling of the diffusion process, by which low-frequency signals determining image composition are primarily generated during early denoising steps [13]. In the proposed approach, the denoising process is truncated at timestep \\( t^s \\), with the IPM directly predicting the intermediate state \\( x_{t^s} \\). By conditioning solely on the foreground \\( c \\) and text \\( y \\), IPM isolates image low-frequency signals from random \\( x_T \\) influence. Subsequent normal denoising processes synthesize rich high-frequency details thereon. To facilitate the optimization of IPM, a comprehensive training paradigm integrating label supervision and reward supervision is proposed. Label supervision uses inversion-derived low-frequency signal ground truth to guide IPM output. Reward supervision subjects predicted \\( x_{t^s} \\) to multiple denoising rounds, with the error between the final image and target image serving as the training reward criterion. We validate the method on our task-specific OpenImage-FCO dataset. The results confirm that it can notably improve the image composition in FCO task. The contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.547, + 0.483, + 0.636 + ], + "angle": 0, + "content": "- We propose a novel Initialization Policy Model (IPM) to enhance image composition in the FCO task. By using the IPM to directly predict the low-frequency signals that match the given foreground, it effectively suppresses the random influence of the initial noise, thus alleviating the issue of flawed composition." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.638, + 0.483, + 0.741 + ], + "angle": 0, + "content": "- We devise a comprehensive training paradigm for the IPM. In addition to label supervision, we introduce novel reward supervision based on a pre-trained generative model, which has been proven to effectively increase the quality of the signals predicted by the IPM, resulting in the final image with reduced artifacts and improved detail accuracy." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.743, + 0.483, + 0.817 + ], + "angle": 0, + "content": "- We construct the OpenImage-FCO dataset based on OpenImage [3] and an automated annotation pipeline. It consists of approximately 580K images, 800K filtered high-quality masks, and their associated tags, providing substantial support for research in the FCO field." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.547, + 0.483, + 0.817 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.24, + 0.846 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Leveraging the abundant priors of large pre-trained models [8, 25, 26], diffusion-based approaches have achieved remarkable progress in many image editing tasks such as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "object editing [4, 9, 29, 32], attribute editing [9, 19], and structure control [5, 30, 44]. This paper focuses on the Foreground Conditioned Out-painting (FCO) task. Related methods can be categorized into two types: learning-free and learning-based approaches. In this section, we introduce representative works from both categories and analyze their application characteristics." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.201, + 0.907, + 0.532 + ], + "angle": 0, + "content": "Learning-Free Method In learning-free methods, the most prominent one is Blended latent Diffusion [1, 2]. In this method, the unknown region (background) is sampled through the standard denoising process. After each step, the known region (foreground) is pasted into the current result. This straightforward pipeline allows it to be plug-and-play in any pre-trained model. However, since the two regions are not acquired from the unified process, it is challenging to achieve a seamless combination between them. Subsequent research [10, 20] has strived to refine the pipeline. For example, DiffEdit [10] employs DDIM Inversion [23] to obtain the features of the known region, thereby reducing the distribution disparity between the features of the two regions. Although the related research is still advancing, at present, the mainstream learning-free methods still adopt the above spatially separated sampling paradigm. A common concern is that it hinders the feature interaction between the two regions, resulting in the inability to achieve mask-aware image completion [14]. With respect to the FCO task, the background synthesized by this type of method often struggles to exhibit a reasonable interaction with the given foreground." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.535, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Learning-based Method Many studies [14, 37, 44, 46] to improve the synthesis effect through task-specific learning. A representative one is SD-Inpainting [26]. It fine-tunes the pre-trained model with a large-scale in-painting dataset and can be directly used for various image completion applications. Since the fine-tuning process involves most of the parameters of Stable Diffusion, the original diffusion prior in it is also affected to a certain extent. Subsequent research [40, 42] has developed a more refined framework. For example, PowerPoint [46] trains an additional task prompt to enhance the model's ability to fill images. After that, BrushNet [14] introduces a twin denoising network to process the features of the known region and optimizes the entire framework end-to-end. One of the most concerned solutions among them is ControlNet [44]. This method trains additional input branches to achieve various conditional generation tasks. When dealing with image completion, the new branch takes the RGB image and mask of the known area as the input. The feature of parameter invariance in the original network enables its direct combination with other methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "It must be pointed out that many studies mentioned above take general image completion rather than FCO as the major task. This is also reflected in the data they use. For example, the masks in the dataset of SD-Inpainting" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16939" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.096, + 0.434, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.449, + 0.096, + 0.904, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.303, + 0.908, + 0.334 + ], + "angle": 0, + "content": "Figure 2. (a) The illustration about the differences between the proposed truncated diffusion process and the original diffusion process. (b) The principle and training paradigm of Initialization Policy Model (IPM)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.358, + 0.485, + 0.496 + ], + "angle": 0, + "content": "[26] are mostly randomly generated, which implies that the known region would contain some clues about the unknown part. But in the FCO task, the known region is the well-segmented foreground object. These differences will undoubtedly cause a series of issues when these methods are applied to FCO. The flawed image composition is one of the most significant problems. In this work, we conduct in-depth research on the FCO task and propose targeted solutions to improve the composition of synthesized images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.507, + 0.345, + 0.523 + ], + "angle": 0, + "content": "3. Truncated Diffusion Model" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.532, + 0.202, + 0.546 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.75 + ], + "angle": 0, + "content": "This paper centers on the FCO task, with the aim of enhancing the image composition of generated images. The research primarily focuses on diffusion-based methods, where the image generation process is modeled as a series of sequential denoising steps commencing from random noise \\( x_{T} \\). Here, \\( T \\) denotes the number of timesteps. Specifically for text-driven FCO, this denoising process is guided by both the text prompt \\( y \\) and the condition image \\( c \\). The text prompt \\( y \\) describes the semantics of the target image, while the condition image \\( c \\) is typically the RGB image of the foreground we intend to present. The diffusion-based FCO model can be formally expressed as \\( p_{\\theta}(x_{0:T-1}|x_{T},y,c) \\), where \\( \\theta \\) represents the denoising model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our method originates from rethinking the influence of \\( x_{T} \\) on image composition. According to existing research, even when using the same prompt, the results sampled from different \\( x_{T} \\) values will differ. The pose, scale, and position of each synthesized object will change. Given that \\( x_{T} \\) is completely independent, the image composition it implies is likely to conflict with the given foreground. As depicted in Fig. 2(a), we propose truncating the denoising process at a specific time step \\( t^{s} \\) and substituting the early steps with an Initialization Policy Model (IPM) \\( \\phi \\). This method exploits" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.51 + ], + "angle": 0, + "content": "the time-frequency coupling in the diffusion process, where the low-frequency signals that determine the image composition are mainly formed in the early stage. \\(\\phi\\) is defined to be solely conditioned on the foreground image \\(c\\) and the text prompt \\(y\\), thereby isolating the random impact of \\(x_{T}\\) on the image composition. As a result, the original model is disassembled into \\(p_{\\phi}(x_{ts}|y,c)\\) and \\(p_{\\theta}(x_{0:t^s +1}|x_{ts},y,c)\\). In this section, we first introduce the principle and network design of the IPM, followed by the detailed presentation of each component in the training paradigm." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.756, + 0.533 + ], + "angle": 0, + "content": "3.2. Initialization Policy Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.6 + ], + "angle": 0, + "content": "The proposed method introduces a new IPM to obtain the intermediate state \\( x_{t^s} \\). According to the principle of the diffusion model, each intermediate state \\( x_{t} \\) follows the distribution [31]:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.615, + 0.907, + 0.631 + ], + "angle": 0, + "content": "\\[\nx _ {t} \\sim \\mathcal {N} \\left(x _ {t}; \\sqrt {\\alpha_ {t}} \\tilde {x} _ {0}, (1 - \\alpha_ {t}) \\mathbf {I}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.906, + 0.713 + ], + "angle": 0, + "content": "\\(\\alpha_{t}\\) the parameter of the diffusion schedule. \\(\\tilde{x}_0\\) represents the prediction of the target image \\(x_0\\) at the current timestep, essentially serving as the deterministic signal after denoising steps from \\(T\\) to \\(t - 1\\). To better clarify the meaning of each parameter, we rewrite Eq. (1) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.725, + 0.907, + 0.761 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\underbrace {\\sqrt {\\alpha_ {t}} \\tilde {x} _ {0}} _ {\\text {S i g n a l}} + \\underbrace {\\sqrt {1 - \\alpha_ {t}} \\epsilon} _ {\\text {N o i s e}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "When \\( t \\) is large, \\( \\tilde{x}_0 \\) primarily contains low-frequency signals [12], which serve as the key determinants of image composition. \\( \\epsilon \\) represents random noise, and different \\( \\epsilon \\) lead to distinct high-frequency details in generated images. In this method, \\( \\phi \\) is tasked with predicting the signal component \\( \\tilde{x}_0 \\). During training, the IPM is required to learn the low-frequency signals corresponding to real target images. During sampling, its output is combined with random noise to generate diverse images with coherent compositions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16940" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "Notably, theoretically, there exists a broad range of potential compositions compatible with given \\( c \\) and \\( y \\), and the ideal \\( x_{t^s} \\) is expected to follow a complex distribution. However, in practice, it is challenging to obtain multiple real target images for a specific pair of \\( c \\) and \\( y \\). Thus, the role of the IPM is to provide a specific cluster within the ideal distribution of \\( x_{t^s} \\). In the experiments section, we further discuss its impact on the original image generation model and how it helps the FCO model achieve a balance between diversity and coherence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.487, + 0.504 + ], + "angle": 0, + "content": "Regarding the network architecture, the IPM \\(\\phi\\) is directly initialized from the U-Net of Stable Diffusion [26]. This strategy aims to inherit the rich text-image priors embedded in its parameters. The text prompt \\(y\\) (encoded by CLIP [24]) is integrated through cross-attention layers. The network's input layer is expanded with four additional channels to accommodate the condition image \\(c\\) (encoded by the Latent VAE [26]). Furthermore, considering the U-Net was originally trained as a denoising model, a fixed noise strategy is introduced. Specifically, a static noise \\(x_{T}^{st}\\) is fed into the original channels. The final prediction of the low-frequency signal is computed from the network output and \\(x_{T}^{st}\\) according to the diffusion schedule: \\(\\sqrt{1 / \\alpha_{t^s}} x_T^{st} - \\sqrt{1 / \\alpha_{t^s} - 1}\\phi (y,c)\\). Since the schedule parameters and noise in this formula remain constant across training and sampling, we hereinafter use \\(\\phi (y,c)\\) to denote the predicted low-frequency signal for simplicity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.525, + 0.407, + 0.541 + ], + "angle": 0, + "content": "3.3. Inversion-derived Label Supervision" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Supervision with reliable ground truth is one of the most effective training paradigms. For the prediction objectives of the IPM, the ground truth can be derived through the inversion of the diffusion process [15, 31, 38]. The inversion operation proceeds in the reverse direction of the sampling process, involving stepwise denoising from the target image \\( x_0 \\) to the corresponding initial noise \\( x_{T} \\). By adhering to the identical noise schedule, the trajectory of intermediate state \\( x_{t} \\) obtained through inversion demonstrates high congruence with that of normal sampling operations, thereby qualifying it as valid ground truth. In this work, the widely adopted DDIM Inversion [31] is implemented, which operates within the range \\( t\\in [0,t^s] \\) to compute the ground truth of the intermediate state at truncation timestep \\( x_{ts}^{gt} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Based on the results of the inversion process, the ground truths of the signal and random noise can be further obtained: \\(\\tilde{x}_0^{gt} = \\sqrt{1 / \\alpha_{t^s}} x_{t^s} - \\sqrt{1 / \\alpha_{t^s} - 1}\\theta (x_t,y)\\), \\(\\epsilon^{gt} = \\theta (x_t,y)\\). Here, \\(\\theta\\) represents the pretrained diffusion model. \\(\\tilde{x}_0^{gt}\\) will be used as a label to supervise the training of the IPM, and the loss function is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.882, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {l b l} = \\left\\| \\phi (y, c) - \\tilde {x} _ {0} ^ {g t} \\right\\| _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.782, + 0.107 + ], + "angle": 0, + "content": "3.4. Diffusion Reward Supervision" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.905, + 0.279 + ], + "angle": 0, + "content": "In our approach, the IPM is tasked with predicting the low-frequency signals of images. Solely using the MSE loss as \\(\\mathcal{L}_{lbl}\\) may cause it to produce over-smoothed results [17]. Not to mention that results obtained via inversion operation inherently introduce deviations [15]. Therefore, we propose Diffusion Reward Supervision to assist in the training of the IPM. Its principle is illustrated in Fig. 2(b). The \\(x_{t^s}\\) sampled by \\(\\phi\\) will be input into the pre-trained diffusion model (reward model) \\(\\theta\\) to perform multiple rounds of denoising, and the error between the resulting image and the target image \\(x_0\\) will serve as the reward to supervise the training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.28, + 0.905, + 0.383 + ], + "angle": 0, + "content": "The diffusion reward adopts the same fast denoising schedule as DDIM [31]. Assuming the number of denoising steps is \\(K\\), the sampling step size would be \\(\\lfloor (T - t^s) / K \\rfloor\\). To enhance the diversity of the sampling schedules, the step size of the first step is set to a random number within \\([1, \\lfloor (T - t^s) / K \\rfloor]\\). Each denoising step can be formulated as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.41, + 0.905, + 0.429 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = \\sqrt {\\alpha_ {t - 1} / \\alpha_ {t}} x _ {t} - \\sqrt {\\alpha_ {t - 1} / \\alpha_ {t} - 1} \\theta (x _ {t}, y) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.905, + 0.499 + ], + "angle": 0, + "content": "In principle, the gradients generated by the final reward can be back-propagated according to the chain rule to guide the training of \\(\\phi\\). The gradients between neighboring timesteps can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.515, + 0.905, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\partial x _ {t - 1} = a _ {t} \\partial x _ {t} + b _ {t} \\partial \\theta (x _ {t}, y) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.658 + ], + "angle": 0, + "content": "\\(a_{t}\\) and \\(b_{t}\\) are used to simplify the schedule parameters, where \\(a_{t} = \\sqrt{\\alpha_{t - 1} / \\alpha_{t}}\\) and \\(b_{t} = \\sqrt{\\alpha_{t - 1} / \\alpha_{t} - 1}\\). Eq. (5) reveals a drawback of the diffusion reward: the propagation of the gradient requires repeated passage through the reward model \\(\\theta\\), which will occupy a large amount of video memory. To solve this problem, we refer to the work of Wu et al. [39] and use the technique of stop gradients here. The modified gradients between neighboring timesteps become:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.673, + 0.905, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\partial x _ {t - 1} = a _ {t} \\partial x _ {t} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.696, + 0.905, + 0.772 + ], + "angle": 0, + "content": "This operation reduces the accuracy of the gradients to some extent, but it also makes it possible to conduct the training with limited video memory. We use \\( msd(*) \\) to represent the multi-step denoising operation, and the diffusion reward can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.795, + 0.905, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r w d} = \\left\\| m s d \\left(\\sqrt {\\alpha_ {t ^ {s}}} \\phi (y, c) + \\sqrt {1 - \\alpha_ {t ^ {s}}} \\epsilon^ {g t}\\right) - x _ {0} \\right\\| _ {2} ^ {2} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\(\\epsilon^{gt}\\) is the noise corresponding to the target image obtained during the inversion process (see Sec. 3.3). The meaning of \\(\\mathcal{L}_{rwd}\\) is that we expect the predicted low-frequency signal plus the ground truth of the noise to restore the corresponding target image. Experiments have" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16941" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.366, + 0.908, + 0.396 + ], + "angle": 0, + "content": "Figure 3. The samples of generated images of the comparison methods. All compared methods are implemented with the recommended settings. \\(\\dagger\\) means that the method is trained / finetuned with our dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.484, + 0.496 + ], + "angle": 0, + "content": "confirmed that \\(\\mathcal{L}_{rwd}\\) can effectively enhance the quality of the signals predicted by the IPM, resulting in the final image with reduced artifacts and improved detail accuracy. The final loss function is as follows, where \\(\\lambda_{rwd}\\) is the weight of the reward supervision term:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.514, + 0.482, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {l b l} + \\lambda_ {r w d} \\mathcal {L} _ {r w d} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.631 + ], + "angle": 0, + "content": "As described in Sec. 3.2, during sampling, the output of \\(\\phi\\) is combined with random noise to obtain the intermediate state: \\(x_{ts} = \\sqrt{\\alpha_{ts}}\\phi (y,c) + \\sqrt{1 - \\alpha_t}\\epsilon\\), where \\(\\epsilon \\sim \\mathcal{N}(0,1)\\). Subsequent denoising steps are completed by conventional diffusion-based FCO. In this paper, a ControlNet trained on the same dataset is employed in this phase." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.645, + 0.224, + 0.661 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.484, + 0.777 + ], + "angle": 0, + "content": "Numerical experiments are conducted to evaluate the proposed method. This section begins with an introduction to the dataset and experimental setups. Afterward, the comparison between IPM and state-of-the-art methods is reported. Then, the impact of IPM on the generative model is discussed. Finally, ablation studies and hyperparameter analyses for the proposed method are presented." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.187, + 0.803 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The advancement of the FCO technique demands high-quality task-specific datasets. Existing datasets, despite their large scale, often lack sufficient diversity in masked objects and their saliency in images. For instance, PowerPaint [46] utilized the OpenImage [3] segmentation subset (1.74M images, 2.8M masks) but focused on only 600" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.907, + 0.481 + ], + "angle": 0, + "content": "object classes, many of which were not the primary subjects in their images. To address this limitation, we introduce OpenImage-FCO, a novel task-specific dataset built on OpenImage-v7 [3]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.909, + 0.665 + ], + "angle": 0, + "content": "OpenImage-FCO incorporates multi-modal filtering and semantic reasoning to enhance annotation quality. We leverage BLIP2 [18] and RAM [45] for image captioning and tagging, followed by semantic filtering via Llama3 [34] to retain only focused objects. GroundingDINO [21] and HQSAM [16] are used to generate precise object masks. The final dataset includes 580K images with 800K masks, 6K unique object tags, and image captions. For evaluation, we curated a test set of 250 cases, each equipped with manual foreground mask annotations and a text prompt. Its scale is aligned with EditBench [36], a widely adopted benchmark in the image editing domain, which encompasses 240 cases." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.722, + 0.697 + ], + "angle": 0, + "content": "4.2. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Implementation Details Our method is built upon Stable Diffusion V2.1 with an image size of \\(512 \\times 512\\). When training the IPM, the truncation timestep \\(t^s\\) is set to 800, the diffusion reward parameter \\(K\\) is set to 10, and \\(\\lambda_{rwd}\\) is set to 1. The AdamW optimizer [22] is employed with a constant learning rate of \\(1 \\times 10^{-4}\\). Training is conducted on four Nvidia L40S GPUs with a batch size of 8 and a total of 400K iterations. The FCO ControlNet used for subsequent inference steps adopts the same structure as described in the original paper [44]. The AdamW optimizer is configured with a learning rate of \\(1 \\times 10^{-5}\\), a batch size of 8, and 200K iterations. Notably, all backbone parameters, including Cross-Attn layers, were frozen during Control-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16942" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.092, + 0.166, + 0.103 + ], + "angle": 0, + "content": "Cond. 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.092, + 0.264, + 0.102 + ], + "angle": 0, + "content": "Cond. 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.092, + 0.36, + 0.102 + ], + "angle": 0, + "content": "Cond. 3" + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.092, + 0.458, + 0.102 + ], + "angle": 0, + "content": "Cond. 4" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.105, + 0.19, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.106, + 0.287, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.105, + 0.384, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.105, + 0.481, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.18, + 0.435, + 0.19 + ], + "angle": 0, + "content": "- A Siamese cat stands on a stone, surrounded by purple flowers. in a sunlit garden." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.191, + 0.189, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.191, + 0.286, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.191, + 0.385, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.191, + 0.481, + 0.264 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.265, + 0.402, + 0.274 + ], + "angle": 0, + "content": "- A train speeds along the tracks, surrounded by lush green trees and hills." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.275, + 0.189, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.191, + 0.275, + 0.287, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.275, + 0.385, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.275, + 0.481, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.35, + 0.395, + 0.359 + ], + "angle": 0, + "content": "- A comfortable chair sits in the corner of a charmingly cozy living room." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.412 + ], + "angle": 0, + "content": "Figure 4. The generated images of foreground objects with variable identities, scales, and positions. The low-frequency signals predicted by IPM are shown in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.441, + 0.483, + 0.531 + ], + "angle": 0, + "content": "Net training to maximize the preservation of backbone text-image priors. All synthesized images presented in this article are sampled using the following strategy: After obtaining \\( x_{t^s} \\) from the IPM, the DDIM sampler [31] completes the remaining sampling steps with a step size of 10 and a classifier-free guidance scale of 7." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.483, + 0.835 + ], + "angle": 0, + "content": "Metrics We take into account 5 metrics from three aspects: image composition, text-image alignment and image quality. 1) Image Composition: The defective image composition is mainly manifested as artifacts and incorrect interactions. Specifically, GroundingDINO [21] and HQSAM [16] are used to detect all the masks that have the same semantics as the foreground objects. The size ratio between the overflow area and the ground-truth foreground is reported as the first metric, which is called SAM Score (SAM). The quality of the interaction is mainly measured by a user study (INT). Thirty individuals participate in the survey and assess the generated images regarding the object interaction. The scale ranges from 1 to 5, with a higher score indicating better results. 2) Text-Image Alignment: CLIP Similarity (CLIP) is reported to reflect the text-image alignment between the text prompt and the generated image. 3) Image Quality: Aesthetic Score (AS) [28] and Image Reward (IR) [41] are used to evaluate the image quality. These evaluation models are trained on human preference datasets and are considered to align well with the human perception." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.848, + 0.4, + 0.865 + ], + "angle": 0, + "content": "4.3. Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We conduct comparisons with six representative related methods, including Blended Latent Diffusion (BLD) [2]," + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.278 + ], + "angle": 0, + "content": "
MethodSAM↓INT↑CLIP↑IR↑AS↑
BLD [2]0.9861.7526.060.185.04
SD-Inpaint.[26]0.3173.3526.260.355.19
PowerPaint [46]0.8202.2326.130.345.05
BrushNet [14]0.1904.0626.620.605.31
Layer Diff. [43]0.0463.5026.150.225.27
ControlNet† [44]0.2353.8826.460.525.27
PowerPaint† [46]0.4333.2726.150.465.20
BrushNet† [14]0.1203.9426.560.585.29
Ours0.0704.6226.500.675.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.289, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Table 1. The quantitative comparison of different methods. The items in bold and underline are ranked first and second. \\(\\dagger\\) means that the method is trained / finetuned with our dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.905, + 0.463 + ], + "angle": 0, + "content": "SD-Inpainting [26], PowerPoint [46], ControlNet [44], BrushNet [14] and Layer Diffusion [43]. Among them, all except BLD are learning-based methods. ControlNet is trained using the OpenImage-FCO dataset. For a fair comparison, we also fine-tuned BrushNet and PowerPaint using the proposed dataset. The intuitive and quantitative comparison results are shown in Fig. 3 and Tab. 1 respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.906, + 0.719 + ], + "angle": 0, + "content": "On the objective metric SAM, our method outperforms most comparison methods by at least 0.050. The only method that scores higher than our proposed method, Layer Diffusion, lacks competitiveness in other metrics due to poor coordination between the foreground and background. On the user-study metric INT, the average score given by users to our method was 0.56 higher than that of the second-best method. These results demonstrate the effectiveness of our method in improving image composition. The samples in Fig. 3 also intuitively illustrate this advantage. Among all compared methods, our approach produces artifact-free results across four cases. In terms of text-image alignment, our method achieves the second-best result. The first place is taken by BrushNet, partially due to its dataset. BrushNet was trained on LAION-Aesthetic [28] containing 120 million images, where abundant image-text pairs ensure advanced performance in text prompt controllability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.87 + ], + "angle": 0, + "content": "The advantages in image composition also drive better performance in image quality metrics, with our method outperforming the second-best by 0.07 and 0.03 on IR and AS respectively. Notably, while suppressing artifacts and incorrect interactions, our method produces less cluttered backgrounds compared to others. This can be attributed to the distribution discrepancy between \\( x_{t^s} \\) inferred by IPM and natural sampling. Fortunately, this characteristic also enhances visual hierarchy in synthesized images, making them more suitable for object presentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "A noteworthy result is that after fine-tuning on the proposed dataset, PowerPaint and BrushNet have shown signif" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16943" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.482, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.262, + 0.482, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.441, + 0.483, + 0.482 + ], + "angle": 0, + "content": "Figure 5. (a) The generation results of the baseline model with or w/o IPM under different text prompts. (b) The multi generation results of the baseline model with or w/o IPM in same case." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.497, + 0.481, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.662, + 0.483, + 0.691 + ], + "angle": 0, + "content": "Figure 6. The samples of predicted low-frequency signals and final generated images with or w/o the diffusion reward loss \\(\\mathcal{L}_{rwd}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.839 + ], + "angle": 0, + "content": "icant improvements in the SAM metric, which once again confirms the value of task-specific datasets. To further demonstrate the IPM's capabilities, we select foregrounds with diverse identities, scales, and positions as condition images and generate results under the same text prompt. The results in Fig. 4 show that our method adapts well to various scenarios and produces reasonable interactions between given objects and backdrops." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.461, + 0.863 + ], + "angle": 0, + "content": "4.4. The Influence of IPM on Generation Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Given that the integration of IPM fundamentally alters the original diffusion process, its implications for the under" + }, + { + "type": "table", + "bbox": [ + 0.548, + 0.089, + 0.875, + 0.189 + ], + "angle": 0, + "content": "
SettingSAM↓CLIP↑IR↑AS↑
Baseline0.23526.460.525.27
w/o L_rwd0.10326.440.645.26
w/o U-Net Init.0.07426.290.635.28
Ours0.07026.500.675.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.528, + 0.199, + 0.89, + 0.213 + ], + "angle": 0, + "content": "Table 2. The quantitative evaluation results of ablation study." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.905, + 0.303 + ], + "angle": 0, + "content": "lying generation model warrant detailed discussion. For image generation tasks, the primary concerns are semantic controllability and output diversity, both of which will be systematically addressed in this section." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.907, + 0.562 + ], + "angle": 0, + "content": "The generation results of the baseline model (ControlNet) with and without IPM under different text prompts \\( y \\) are presented in Fig. 5(a). It can be observed that IPM does not significantly affect the model's controllability, as the model still synthesizes results with the required semantics. This finding is also supported by Tab. 1, where our method and ControlNet exhibit the closest CLIP scores among all approaches. The subsequent ablation studies will reveal that this benefit primarily stems from initializing IPM parameters with the pretrained U-Net. Turning to Fig. 5(b), the generation results of the baseline model with and without IPM for the same pair of \\( c \\) and \\( y \\) are shown. In this comparison, IPM results show more similar composition due to the shared low-frequency signals, while the baseline model's diverse outputs come at the cost of prevalent compositional defects. This reinforces that IPM trades inefficient diversity for compositional rationality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.574, + 0.667, + 0.59 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.703 + ], + "angle": 0, + "content": "In this section, we conduct ablation studies on critical components of the IPM training paradigm and network design, including the diffusion reward supervision \\(\\mathcal{L}_{rwd}\\) and the strategy of initializing IPM parameters with a pretrained U-Net. Note that since removing \\(\\mathcal{L}_{lbl}\\) would destabilize training, we exclude it from ablation analysis. The experimental results are presented in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Ablation Study for \\(\\mathcal{L}_{rwd}\\) Quantitative validation shows that diffusion reward supervision provides the most significant benefit in artifact suppression. When \\(\\mathcal{L}_{rwd}\\) is removed, the SAM score increases by 0.033. The introduction of new artifacts also degrades image quality metrics, with AS dropping by 0.08. These artifacts correlate with the quality of low-frequency signals predicted by IPM. To visually corroborate this, we visualize these low-frequency signals (decoded via the Latent VAE) in Fig. 6. It can be observed that without \\(\\mathcal{L}_{rwd}\\), IPM outputs smoother signals. A direct consequence is that edges of some foreground objects become indistinct, which may develop into artifacts and imprecise details after subsequent denoising steps. Conversely, when" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16944" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.482, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.177, + 0.368, + 0.187 + ], + "angle": 0, + "content": "- A brand-new football is placed on the grass in the football field." + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.197, + 0.47, + 0.212 + ], + "angle": 0, + "content": "Figure 7. The samples of generated images under different \\(t^s\\)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.226, + 0.48, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.393, + 0.483, + 0.421 + ], + "angle": 0, + "content": "Figure 8. The quantitative evaluation results of the truncating timestep selection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.483, + 0.492 + ], + "angle": 0, + "content": "diffusion reward supervision is introduced, IPM is encouraged to output intermediate states that genuinely lead to target images, effectively mitigating this issue." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.493, + 0.483, + 0.645 + ], + "angle": 0, + "content": "Ablation Study for U-Net Initialization As shown in Tab. 2, initializing IPM parameters with a pretrained U-Net primarily strengthens the model's text controllability. Removing this strategy causes the CLIP score to drop by 0.21, even below the baseline performance. Although IPM is only responsible for predicting low-frequency signals, training it from scratch limits its performance on complex text prompts due to the dataset's relatively limited corpus. This phenomenon substantiates the necessity of leveraging the rich text-image priors of pretrained models." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.653, + 0.416, + 0.669 + ], + "angle": 0, + "content": "4.6. The Selection of Truncating Timestep" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.75 + ], + "angle": 0, + "content": "The primary hyperparameter when applying IPM is determining the appropriate timestep \\( t^s \\) to truncate the diffusion process. We test values within the set {900, 850, 800, 750, 700}. The intuitive and quantitative comparison results are shown in Fig. 7 and Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "When the truncation timestep \\( t^s \\) is set too late, the method's effectiveness declines significantly. In the \\( t^s = 700 \\) experiment group, although artifacts are further suppressed, this comes at the cost of overly simplified backgrounds and significantly reduced image quality. Its IR and AS metrics are 0.35 and 0.60 lower than those of \\( t^s = 800 \\). This phenomenon is associated with the function of IPM. IPM directly predicts the image's low-frequency signal. When \\( t^s \\) is set to a smaller value, IPM's task becomes more complex. Meanwhile, the number of subsequent normal de" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.905, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.198, + 0.905, + 0.226 + ], + "angle": 0, + "content": "Figure 9. The samples of failure case about the objects with extreme colors or lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.907, + 0.406 + ], + "angle": 0, + "content": "noising steps is reduced, requiring IPM to predict image signals across broader frequency bands. \\( t^s = 700 \\) is clearly an excessively challenging value. As illustrated in Fig. 7, the produced image exhibits obvious detail deficiencies due to the interplay of these factors. When \\( t^s = 900 \\), the frequency band managed by IPM is too narrow, weakening its influence on final image. Its SAM score increases by 0.033 compared to \\( t^s = 800 \\), indicating that subsequent denoising steps introduce more suboptimal compositions. In this paper, \\( t^s = 800 \\) is adopted across all experiments. g" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.415, + 0.78, + 0.43 + ], + "angle": 0, + "content": "4.7. Failure case and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.65 + ], + "angle": 0, + "content": "During testing, a notable failure case arises when the given foreground object exhibits extreme colors or lighting conditions, significantly degrading background controllability. Fig. 9 illustrates results for such cases. Despite prompts' requirements, synthesized images still adopt styles consistent with the provided foreground (case 1) or fail to include other backdrops (case 2). Root causes include dataset limitations (artistic photography underrepresented in OpenImage) and framework design. That is, features of objects with extreme colors tend to dominate background features during synthesis, compelling background features to always reflect similar appearances. This motivates exploring methods to decouple foreground-background color distributions and enable foreground relighting in FCO tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.9 + ], + "angle": 0, + "content": "This paper focuses on the problem of flawed image composition in diffusion-based FCO methods. We believe that this problem can be partially attributed to the random impact of the initial noise. Inspired by this, we propose the Initialization Policy Model (IPM). It uses a deterministic model that only depends on the given foreground and text prompt to replace the early denoising steps, thus isolating the impact of the initial noise. The subsequently proposed comprehensive training paradigm ensures that the IPM can be effectively optimized. Experiments confirm that the proposed method can significantly improve the composition problems such as artifacts and incorrect interactions, and the comprehensive framework built upon IPM also achieves advanced performance in FCO task." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16945" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.198 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant No. 62473290, 62088101, National Key Research and Development Program of China under Grant No. 2024YFB3311801, and the Shanghai Municipal Commission of Economy and Informatization under Grant No. 2024-GZL-RGZN-01008." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.209, + 0.188, + 0.225 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.234, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18187-18197, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.277, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM TOG, 42(4):149:1-149:11, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.318, + 0.482, + 0.358 + ], + "angle": 0, + "content": "[3] Rodrigo Benenson and Vittorio Ferrari. From colouring-in to pointillism: revisiting semantic segmentation supervision. CoRR, abs/2210.14142, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In CVPR, pages 22503-22513, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.482, + 0.512 + ], + "angle": 0, + "content": "[6] Shidong Cao, Wenhao Chai, Shengyu Hao, Yanting Zhang, Hangyue Chen, and Gaoang Wang. Diffashion: Reference-based fashion design with structure-aware transfer by diffusion models. TMM, 26:3962-3975, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.482, + 0.581 + ], + "angle": 0, + "content": "[7] Tingfeng Cao, Junsheng Kong, Xue Zhao, Wenqing Yao, Junwei Ding, Jinhui Zhu, and Jiandong Zhang. Product2img: Prompt-free e-commerce product background generation with diffusion model and self-improved lmm. In ACMMM, pages 10774-10783, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.482, + 0.651 + ], + "angle": 0, + "content": "[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, pages 1–31, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.652, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, pages 6593-6602, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.693, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[10] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. CoRR, abs/2210.11427, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.749, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[11] Amir Erfan Eshratifar, Joao VB Soares, Kapil Thadani, Shaunak Mishra, Mikhail Kuznetsov, Yueh-Ning Ku, and Paloma De Juan. Salient object-aware background generation using text-guided diffusion models. In CVPR, pages 7489-7499, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[12] Xuan Han, Yihao Zhao, and Mingyu You. Scene diffusion: Text-driven scene image synthesis conditioning on a single 3d model. In ACMMM, pages 7862-7870, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.234, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[14] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. CoRR, abs/2403.06976, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[15] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. In ICLR, pages 1-28, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[16] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In NeurIPS, pages 29914–29934, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[17] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew P. Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, pages 105-114, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[19] Pengzhi Li, Qinxuan Huang, Yikang Ding, and Zhiheng Li. Layerdiffusion: Layered controlled image editing with diffusion models. In ACM SIGGRAPH Asia, pages 1-4, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[20] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, and Jianzhuang Liu. Zone: Zero-shot instruction-guided local editing. In CVPR, pages 6254-6263, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[21] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: marrying DINO with grounded pre-training for open-set object detection. CoRR, abs/2303.05499, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.904, + 0.589 + ], + "angle": 0, + "content": "[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, pages 1-11, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[23] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[25] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685, 2022. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[27] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In AAAI, pages 4695-4703, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16946" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[28] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, pages 1–21, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[29] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In CVPR, pages 8871-8879, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[30] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent Y. F. Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In CVPR, pages 8839-8849, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, pages 1-12, 2021. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[32] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In CVPR, pages 18310-18319, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.476 + ], + "angle": 0, + "content": "[33] Zhengwentai Sun, Yanghong Zhou, Honghong He, and PY Mok. Sgdiff: A style guided diffusion model for fashion synthesis. In ACMMM, pages 8433-8442, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[34] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, and Faisal Azhar. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[35] Ruoyu Wang, Yongqi Yang, Zhihao Qian, Ye Zhu, and Yu Wu. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, pages 1-21, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[36] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, pages 18359-18369, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[37] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, pages 18359-18369, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[38] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In ICCV, pages 7378-7387, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[39] Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, and Hongsheng Li. Deep reward supervisions for tuning text-to-image diffusion models. CoRR, abs/2405.00760, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.146 + ], + "angle": 0, + "content": "[40] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[41] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In NeurIPS, pages 1-14, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[42] Shiyuan Yang, Xiaodong Chen, and Jing Liao. Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In ACM MM, page 3190-3199, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[43] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. ACM Transactions on Graphics (TOG), 43(4):1-15, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[44] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3813-3824, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[45] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize anything: A strong image tagging model. In CVPR, pages 1724-1732, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[46] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. CoRR, abs/2312.03594, 2023. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16947" + } + ] +] \ No newline at end of file diff --git a/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/2c55e2cf-c2bf-4623-a4a2-6a1a33a77533_origin.pdf b/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/2c55e2cf-c2bf-4623-a4a2-6a1a33a77533_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8dd95b12f8484f0ada147cad08e9e4464b67cb48 --- /dev/null +++ b/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/2c55e2cf-c2bf-4623-a4a2-6a1a33a77533_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3546cb0e70d574e11f0c30f59d58cf5de11650089e0d4105641085c57933bbd +size 25815429 diff --git a/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/full.md b/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c9a0f8a78b35c7ac73146906ec23e7d201ca29b3 --- /dev/null +++ b/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/full.md @@ -0,0 +1,310 @@ +# Toward Better Out-painting: Improving the Image Composition with Initialization Policy Model + +Xuan Han $^{1}$ Yihao Zhao $^{1}$ Yanhao Ge $^{1,2}$ Mingyu You $^{1*}$ $^{1}$ Tongji University, Shanghai, China + $^{2}$ Future Imaging Area +{hanxuan, zhaoyihao, myyou}@tongji.edu.cn {halege}@vivo.com + +# Abstract + +With its extensive applications, Foreground Conditioned Out-painting (FCO) has attracted considerable attention in the research field. Through the utilization of text-driven FCO, users are enabled to generate diverse backgrounds for a given foreground by adjusting the text prompt, which considerably enhances the efficiency in fields like e-commerce. Since the foreground is fixed in FCO, a key concern is whether the generated background can match the foreground well to achieve a coherent composition. However, most existing methods are lacking in this regard. Artifacts and incorrect interactions are common defects in synthesized images. This issue is linked to the influence of the initial noise in the sampling process. As the initial noise is sampled independently, it's highly likely that the implied image composition will conflict with the given foreground. In this paper, a novel Initialization Policy Model (IPM) is proposed to address this problem. Its function is to replace the early denoising steps and directly predict the intermediate state that is conducive to the reasonable image composition. Since the IPM is designed to take only the foreground image and the text prompt as inputs, it isolates the impact of the initial noise. The subsequently proposed training paradigm that combines inversion-derived label supervision and diffusion reward supervision further ensures the efficient training of the IPM. The evaluation is conducted using the task-specific OpenImage-FCO dataset developed by us. The results verify that the introduction of the IPM can significantly improve the composition of the synthesized images and achieve advanced performance in the FCO task. + +# 1. Introduction + +Nowadays, diffusion-based image editing has emerged as a trending research topic. Its progress in numerous tech- + +![](images/fd8241776b463b7c8b42d48a7f02b84badb2ea3d5c6ee15389caaef35349f84e.jpg) +Figure 1. The illustrations concerning the flawed image composition in foreground conditioned out-painting task. + +niques is continuously revolutionizing the practices in related fields such as commercial art [7, 11] and fashion design [6, 33]. Among the various techniques available, Foreground Conditioned Out-painting (FCO) [2, 14, 26] has attracted significant attention. FCO focuses on a special use case: The foreground of an image, usually containing the main object, is fixed, and image generators like Stable Diffusion [26] are used to fill the background based on the user's prompt. This technology allows users to generate diverse backgrounds for the given foreground simply by changing the text prompt, which significantly enhances efficiency compared to standard workflows. + +In this task, a crucial concern is the quality of image composition. Take Fig. 1 as an example: the user wants to display the mug in a fresh scene. The expected result is that the synthesized background can be harmoniously integrated with the mug. However, existing methods demonstrate insufficient performance in this regard. In many results, background objects like the table fail to interact appropriately with the given mug, alongside artifacts generated in the process. Overall, the synthesized background often exists independently of the given foreground rather than forming a unified composition. + +To identify the source of the problem, we revisit the specificity of the FCO task. In diffusion-based FCO, images are gradually denoised from initial noise $x_{T}$ . According to existing literature [27, 35], variations in $x_{T}$ compre + +hensively influence the position, scale, and pose of synthesized objects. However, in FCO, independently sampled $x_{T}$ frequently implies compositions conflicting with the fixed foreground. As shown in Fig. 1, different $x_{T}$ values lead to diverse table poses, many of which clash with the mug; artifacts are also generated to match these inconsistent backdrops. Evidently, addressing this issue requires mitigating the random impact of $x_{T}$ on image composition. + +In this paper, a novel Initialization Policy Model (IPM) is introduced to address the problem. The method exploits the time-frequency coupling of the diffusion process, by which low-frequency signals determining image composition are primarily generated during early denoising steps [13]. In the proposed approach, the denoising process is truncated at timestep $t^s$ , with the IPM directly predicting the intermediate state $x_{t^s}$ . By conditioning solely on the foreground $c$ and text $y$ , IPM isolates image low-frequency signals from random $x_T$ influence. Subsequent normal denoising processes synthesize rich high-frequency details thereon. To facilitate the optimization of IPM, a comprehensive training paradigm integrating label supervision and reward supervision is proposed. Label supervision uses inversion-derived low-frequency signal ground truth to guide IPM output. Reward supervision subjects predicted $x_{t^s}$ to multiple denoising rounds, with the error between the final image and target image serving as the training reward criterion. We validate the method on our task-specific OpenImage-FCO dataset. The results confirm that it can notably improve the image composition in FCO task. The contributions of this work are as follows: + +- We propose a novel Initialization Policy Model (IPM) to enhance image composition in the FCO task. By using the IPM to directly predict the low-frequency signals that match the given foreground, it effectively suppresses the random influence of the initial noise, thus alleviating the issue of flawed composition. +- We devise a comprehensive training paradigm for the IPM. In addition to label supervision, we introduce novel reward supervision based on a pre-trained generative model, which has been proven to effectively increase the quality of the signals predicted by the IPM, resulting in the final image with reduced artifacts and improved detail accuracy. +- We construct the OpenImage-FCO dataset based on OpenImage [3] and an automated annotation pipeline. It consists of approximately 580K images, 800K filtered high-quality masks, and their associated tags, providing substantial support for research in the FCO field. + +# 2. Related Works + +Leveraging the abundant priors of large pre-trained models [8, 25, 26], diffusion-based approaches have achieved remarkable progress in many image editing tasks such as + +object editing [4, 9, 29, 32], attribute editing [9, 19], and structure control [5, 30, 44]. This paper focuses on the Foreground Conditioned Out-painting (FCO) task. Related methods can be categorized into two types: learning-free and learning-based approaches. In this section, we introduce representative works from both categories and analyze their application characteristics. + +Learning-Free Method In learning-free methods, the most prominent one is Blended latent Diffusion [1, 2]. In this method, the unknown region (background) is sampled through the standard denoising process. After each step, the known region (foreground) is pasted into the current result. This straightforward pipeline allows it to be plug-and-play in any pre-trained model. However, since the two regions are not acquired from the unified process, it is challenging to achieve a seamless combination between them. Subsequent research [10, 20] has strived to refine the pipeline. For example, DiffEdit [10] employs DDIM Inversion [23] to obtain the features of the known region, thereby reducing the distribution disparity between the features of the two regions. Although the related research is still advancing, at present, the mainstream learning-free methods still adopt the above spatially separated sampling paradigm. A common concern is that it hinders the feature interaction between the two regions, resulting in the inability to achieve mask-aware image completion [14]. With respect to the FCO task, the background synthesized by this type of method often struggles to exhibit a reasonable interaction with the given foreground. + +Learning-based Method Many studies [14, 37, 44, 46] to improve the synthesis effect through task-specific learning. A representative one is SD-Inpainting [26]. It fine-tunes the pre-trained model with a large-scale in-painting dataset and can be directly used for various image completion applications. Since the fine-tuning process involves most of the parameters of Stable Diffusion, the original diffusion prior in it is also affected to a certain extent. Subsequent research [40, 42] has developed a more refined framework. For example, PowerPoint [46] trains an additional task prompt to enhance the model's ability to fill images. After that, BrushNet [14] introduces a twin denoising network to process the features of the known region and optimizes the entire framework end-to-end. One of the most concerned solutions among them is ControlNet [44]. This method trains additional input branches to achieve various conditional generation tasks. When dealing with image completion, the new branch takes the RGB image and mask of the known area as the input. The feature of parameter invariance in the original network enables its direct combination with other methods. + +It must be pointed out that many studies mentioned above take general image completion rather than FCO as the major task. This is also reflected in the data they use. For example, the masks in the dataset of SD-Inpainting + +![](images/04692cb158d94516b67758199e09daad01b183bb08f8a61843da6564cbbb3f89.jpg) +Figure 2. (a) The illustration about the differences between the proposed truncated diffusion process and the original diffusion process. (b) The principle and training paradigm of Initialization Policy Model (IPM). + +![](images/7e90ca50fb8a810f2a0c70266bfb7b3f037f4429982d9a86f4b7612fcfe04186.jpg) + +[26] are mostly randomly generated, which implies that the known region would contain some clues about the unknown part. But in the FCO task, the known region is the well-segmented foreground object. These differences will undoubtedly cause a series of issues when these methods are applied to FCO. The flawed image composition is one of the most significant problems. In this work, we conduct in-depth research on the FCO task and propose targeted solutions to improve the composition of synthesized images. + +# 3. Truncated Diffusion Model + +# 3.1. Overview + +This paper centers on the FCO task, with the aim of enhancing the image composition of generated images. The research primarily focuses on diffusion-based methods, where the image generation process is modeled as a series of sequential denoising steps commencing from random noise $x_{T}$ . Here, $T$ denotes the number of timesteps. Specifically for text-driven FCO, this denoising process is guided by both the text prompt $y$ and the condition image $c$ . The text prompt $y$ describes the semantics of the target image, while the condition image $c$ is typically the RGB image of the foreground we intend to present. The diffusion-based FCO model can be formally expressed as $p_{\theta}(x_{0:T-1}|x_{T},y,c)$ , where $\theta$ represents the denoising model. + +Our method originates from rethinking the influence of $x_{T}$ on image composition. According to existing research, even when using the same prompt, the results sampled from different $x_{T}$ values will differ. The pose, scale, and position of each synthesized object will change. Given that $x_{T}$ is completely independent, the image composition it implies is likely to conflict with the given foreground. As depicted in Fig. 2(a), we propose truncating the denoising process at a specific time step $t^{s}$ and substituting the early steps with an Initialization Policy Model (IPM) $\phi$ . This method exploits + +the time-frequency coupling in the diffusion process, where the low-frequency signals that determine the image composition are mainly formed in the early stage. $\phi$ is defined to be solely conditioned on the foreground image $c$ and the text prompt $y$ , thereby isolating the random impact of $x_{T}$ on the image composition. As a result, the original model is disassembled into $p_{\phi}(x_{ts}|y,c)$ and $p_{\theta}(x_{0:t^s +1}|x_{ts},y,c)$ . In this section, we first introduce the principle and network design of the IPM, followed by the detailed presentation of each component in the training paradigm. + +# 3.2. Initialization Policy Model + +The proposed method introduces a new IPM to obtain the intermediate state $x_{t^s}$ . According to the principle of the diffusion model, each intermediate state $x_{t}$ follows the distribution [31]: + +$$ +x _ {t} \sim \mathcal {N} \left(x _ {t}; \sqrt {\alpha_ {t}} \tilde {x} _ {0}, (1 - \alpha_ {t}) \mathbf {I}\right) \tag {1} +$$ + +$\alpha_{t}$ the parameter of the diffusion schedule. $\tilde{x}_0$ represents the prediction of the target image $x_0$ at the current timestep, essentially serving as the deterministic signal after denoising steps from $T$ to $t - 1$ . To better clarify the meaning of each parameter, we rewrite Eq. (1) as follows: + +$$ +x _ {t} = \underbrace {\sqrt {\alpha_ {t}} \tilde {x} _ {0}} _ {\text {S i g n a l}} + \underbrace {\sqrt {1 - \alpha_ {t}} \epsilon} _ {\text {N o i s e}} \tag {2} +$$ + +When $t$ is large, $\tilde{x}_0$ primarily contains low-frequency signals [12], which serve as the key determinants of image composition. $\epsilon$ represents random noise, and different $\epsilon$ lead to distinct high-frequency details in generated images. In this method, $\phi$ is tasked with predicting the signal component $\tilde{x}_0$ . During training, the IPM is required to learn the low-frequency signals corresponding to real target images. During sampling, its output is combined with random noise to generate diverse images with coherent compositions. + +Notably, theoretically, there exists a broad range of potential compositions compatible with given $c$ and $y$ , and the ideal $x_{t^s}$ is expected to follow a complex distribution. However, in practice, it is challenging to obtain multiple real target images for a specific pair of $c$ and $y$ . Thus, the role of the IPM is to provide a specific cluster within the ideal distribution of $x_{t^s}$ . In the experiments section, we further discuss its impact on the original image generation model and how it helps the FCO model achieve a balance between diversity and coherence. + +Regarding the network architecture, the IPM $\phi$ is directly initialized from the U-Net of Stable Diffusion [26]. This strategy aims to inherit the rich text-image priors embedded in its parameters. The text prompt $y$ (encoded by CLIP [24]) is integrated through cross-attention layers. The network's input layer is expanded with four additional channels to accommodate the condition image $c$ (encoded by the Latent VAE [26]). Furthermore, considering the U-Net was originally trained as a denoising model, a fixed noise strategy is introduced. Specifically, a static noise $x_{T}^{st}$ is fed into the original channels. The final prediction of the low-frequency signal is computed from the network output and $x_{T}^{st}$ according to the diffusion schedule: $\sqrt{1 / \alpha_{t^s}} x_T^{st} - \sqrt{1 / \alpha_{t^s} - 1}\phi (y,c)$ . Since the schedule parameters and noise in this formula remain constant across training and sampling, we hereinafter use $\phi (y,c)$ to denote the predicted low-frequency signal for simplicity. + +# 3.3. Inversion-derived Label Supervision + +Supervision with reliable ground truth is one of the most effective training paradigms. For the prediction objectives of the IPM, the ground truth can be derived through the inversion of the diffusion process [15, 31, 38]. The inversion operation proceeds in the reverse direction of the sampling process, involving stepwise denoising from the target image $x_0$ to the corresponding initial noise $x_{T}$ . By adhering to the identical noise schedule, the trajectory of intermediate state $x_{t}$ obtained through inversion demonstrates high congruence with that of normal sampling operations, thereby qualifying it as valid ground truth. In this work, the widely adopted DDIM Inversion [31] is implemented, which operates within the range $t\in [0,t^s]$ to compute the ground truth of the intermediate state at truncation timestep $x_{ts}^{gt}$ . + +Based on the results of the inversion process, the ground truths of the signal and random noise can be further obtained: $\tilde{x}_0^{gt} = \sqrt{1 / \alpha_{t^s}} x_{t^s} - \sqrt{1 / \alpha_{t^s} - 1}\theta (x_t,y)$ , $\epsilon^{gt} = \theta (x_t,y)$ . Here, $\theta$ represents the pretrained diffusion model. $\tilde{x}_0^{gt}$ will be used as a label to supervise the training of the IPM, and the loss function is as follows: + +$$ +\mathcal {L} _ {l b l} = \left\| \phi (y, c) - \tilde {x} _ {0} ^ {g t} \right\| _ {2} ^ {2} \tag {3} +$$ + +# 3.4. Diffusion Reward Supervision + +In our approach, the IPM is tasked with predicting the low-frequency signals of images. Solely using the MSE loss as $\mathcal{L}_{lbl}$ may cause it to produce over-smoothed results [17]. Not to mention that results obtained via inversion operation inherently introduce deviations [15]. Therefore, we propose Diffusion Reward Supervision to assist in the training of the IPM. Its principle is illustrated in Fig. 2(b). The $x_{t^s}$ sampled by $\phi$ will be input into the pre-trained diffusion model (reward model) $\theta$ to perform multiple rounds of denoising, and the error between the resulting image and the target image $x_0$ will serve as the reward to supervise the training. + +The diffusion reward adopts the same fast denoising schedule as DDIM [31]. Assuming the number of denoising steps is $K$ , the sampling step size would be $\lfloor (T - t^s) / K \rfloor$ . To enhance the diversity of the sampling schedules, the step size of the first step is set to a random number within $[1, \lfloor (T - t^s) / K \rfloor]$ . Each denoising step can be formulated as follow: + +$$ +x _ {t - 1} = \sqrt {\alpha_ {t - 1} / \alpha_ {t}} x _ {t} - \sqrt {\alpha_ {t - 1} / \alpha_ {t} - 1} \theta (x _ {t}, y) \tag {4} +$$ + +In principle, the gradients generated by the final reward can be back-propagated according to the chain rule to guide the training of $\phi$ . The gradients between neighboring timesteps can be expressed as: + +$$ +\partial x _ {t - 1} = a _ {t} \partial x _ {t} + b _ {t} \partial \theta (x _ {t}, y) \tag {5} +$$ + +$a_{t}$ and $b_{t}$ are used to simplify the schedule parameters, where $a_{t} = \sqrt{\alpha_{t - 1} / \alpha_{t}}$ and $b_{t} = \sqrt{\alpha_{t - 1} / \alpha_{t} - 1}$ . Eq. (5) reveals a drawback of the diffusion reward: the propagation of the gradient requires repeated passage through the reward model $\theta$ , which will occupy a large amount of video memory. To solve this problem, we refer to the work of Wu et al. [39] and use the technique of stop gradients here. The modified gradients between neighboring timesteps become: + +$$ +\partial x _ {t - 1} = a _ {t} \partial x _ {t} \tag {6} +$$ + +This operation reduces the accuracy of the gradients to some extent, but it also makes it possible to conduct the training with limited video memory. We use $msd(*)$ to represent the multi-step denoising operation, and the diffusion reward can be expressed as: + +$$ +\mathcal {L} _ {r w d} = \left\| m s d \left(\sqrt {\alpha_ {t ^ {s}}} \phi (y, c) + \sqrt {1 - \alpha_ {t ^ {s}}} \epsilon^ {g t}\right) - x _ {0} \right\| _ {2} ^ {2} \tag {7} +$$ + +$\epsilon^{gt}$ is the noise corresponding to the target image obtained during the inversion process (see Sec. 3.3). The meaning of $\mathcal{L}_{rwd}$ is that we expect the predicted low-frequency signal plus the ground truth of the noise to restore the corresponding target image. Experiments have + +![](images/a387335729bae23cef4e3dbb228a1c2edeb5560a39d03ca2c95dba8e64ad46c2.jpg) +Figure 3. The samples of generated images of the comparison methods. All compared methods are implemented with the recommended settings. $\dagger$ means that the method is trained / finetuned with our dataset. + +confirmed that $\mathcal{L}_{rwd}$ can effectively enhance the quality of the signals predicted by the IPM, resulting in the final image with reduced artifacts and improved detail accuracy. The final loss function is as follows, where $\lambda_{rwd}$ is the weight of the reward supervision term: + +$$ +\mathcal {L} = \mathcal {L} _ {l b l} + \lambda_ {r w d} \mathcal {L} _ {r w d} \tag {8} +$$ + +As described in Sec. 3.2, during sampling, the output of $\phi$ is combined with random noise to obtain the intermediate state: $x_{ts} = \sqrt{\alpha_{ts}}\phi (y,c) + \sqrt{1 - \alpha_t}\epsilon$ , where $\epsilon \sim \mathcal{N}(0,1)$ . Subsequent denoising steps are completed by conventional diffusion-based FCO. In this paper, a ControlNet trained on the same dataset is employed in this phase. + +# 4. Experiments + +Numerical experiments are conducted to evaluate the proposed method. This section begins with an introduction to the dataset and experimental setups. Afterward, the comparison between IPM and state-of-the-art methods is reported. Then, the impact of IPM on the generative model is discussed. Finally, ablation studies and hyperparameter analyses for the proposed method are presented. + +# 4.1. Dataset + +The advancement of the FCO technique demands high-quality task-specific datasets. Existing datasets, despite their large scale, often lack sufficient diversity in masked objects and their saliency in images. For instance, PowerPaint [46] utilized the OpenImage [3] segmentation subset (1.74M images, 2.8M masks) but focused on only 600 + +object classes, many of which were not the primary subjects in their images. To address this limitation, we introduce OpenImage-FCO, a novel task-specific dataset built on OpenImage-v7 [3]. + +OpenImage-FCO incorporates multi-modal filtering and semantic reasoning to enhance annotation quality. We leverage BLIP2 [18] and RAM [45] for image captioning and tagging, followed by semantic filtering via Llama3 [34] to retain only focused objects. GroundingDINO [21] and HQSAM [16] are used to generate precise object masks. The final dataset includes 580K images with 800K masks, 6K unique object tags, and image captions. For evaluation, we curated a test set of 250 cases, each equipped with manual foreground mask annotations and a text prompt. Its scale is aligned with EditBench [36], a widely adopted benchmark in the image editing domain, which encompasses 240 cases. + +# 4.2. Experimental Settings + +Implementation Details Our method is built upon Stable Diffusion V2.1 with an image size of $512 \times 512$ . When training the IPM, the truncation timestep $t^s$ is set to 800, the diffusion reward parameter $K$ is set to 10, and $\lambda_{rwd}$ is set to 1. The AdamW optimizer [22] is employed with a constant learning rate of $1 \times 10^{-4}$ . Training is conducted on four Nvidia L40S GPUs with a batch size of 8 and a total of 400K iterations. The FCO ControlNet used for subsequent inference steps adopts the same structure as described in the original paper [44]. The AdamW optimizer is configured with a learning rate of $1 \times 10^{-5}$ , a batch size of 8, and 200K iterations. Notably, all backbone parameters, including Cross-Attn layers, were frozen during Control- + +![](images/6168034b513bbedc4b15d86f6427bcdc455e03cb7cd4b4f505774fa0bed69d63.jpg) +Cond. 1 +- A Siamese cat stands on a stone, surrounded by purple flowers. in a sunlit garden. + +![](images/78daf4903e3afb1e64d63a6a030738fd9c163dd5c45f587687f4c20afacd2d66.jpg) +Cond. 2 + +![](images/04a8937f0fa46ab3d8a2c619fb6a994f9478e80809622ce63812b10d932c1344.jpg) +Cond. 3 + +![](images/dceff543d3b90d98f36dd6aeeffe7235b01c8fa884596f6fc6c0a3d31d054a34.jpg) +Cond. 4 + +![](images/3b5e1e78d62cbdc1d0fa65405254b0f9655941cf398df6b4f27f8e4ebbea6dbb.jpg) +- A train speeds along the tracks, surrounded by lush green trees and hills. + +![](images/a3dac370d202ef8074b28f5afd65ce67379a24deef89afa313bd942f5522055e.jpg) + +![](images/aa0c5094d83ec7d107242c021e660927b11040e77382ee5643bc82391afd2b9b.jpg) + +![](images/1f38ee668fff1e360d2735fd9e8101bd45d8356174494af4b05091ef2000f1d1.jpg) + +![](images/7028898ee2daf61a550b04bb9f1b721183a1c3a1f0e274da309710e78ddb6010.jpg) +- A comfortable chair sits in the corner of a charmingly cozy living room. +Figure 4. The generated images of foreground objects with variable identities, scales, and positions. The low-frequency signals predicted by IPM are shown in the supplementary materials. + +![](images/504b14b743ebcfc57333c03d30cbdd69678868dc9725fc9da8d3494382eb7370.jpg) + +![](images/765eb4c89557dd2446d8af5e11b75cb78881018e7641c3018ae4cbda99154461.jpg) + +![](images/90d228738ab98fb726ffa3b173a65fccb74f2d8c22b1fd819ea4c12b86b82891.jpg) + +Net training to maximize the preservation of backbone text-image priors. All synthesized images presented in this article are sampled using the following strategy: After obtaining $x_{t^s}$ from the IPM, the DDIM sampler [31] completes the remaining sampling steps with a step size of 10 and a classifier-free guidance scale of 7. + +Metrics We take into account 5 metrics from three aspects: image composition, text-image alignment and image quality. 1) Image Composition: The defective image composition is mainly manifested as artifacts and incorrect interactions. Specifically, GroundingDINO [21] and HQSAM [16] are used to detect all the masks that have the same semantics as the foreground objects. The size ratio between the overflow area and the ground-truth foreground is reported as the first metric, which is called SAM Score (SAM). The quality of the interaction is mainly measured by a user study (INT). Thirty individuals participate in the survey and assess the generated images regarding the object interaction. The scale ranges from 1 to 5, with a higher score indicating better results. 2) Text-Image Alignment: CLIP Similarity (CLIP) is reported to reflect the text-image alignment between the text prompt and the generated image. 3) Image Quality: Aesthetic Score (AS) [28] and Image Reward (IR) [41] are used to evaluate the image quality. These evaluation models are trained on human preference datasets and are considered to align well with the human perception. + +# 4.3. Comparison with Existing Methods + +We conduct comparisons with six representative related methods, including Blended Latent Diffusion (BLD) [2], + +
MethodSAM↓INT↑CLIP↑IR↑AS↑
BLD [2]0.9861.7526.060.185.04
SD-Inpaint.[26]0.3173.3526.260.355.19
PowerPaint [46]0.8202.2326.130.345.05
BrushNet [14]0.1904.0626.620.605.31
Layer Diff. [43]0.0463.5026.150.225.27
ControlNet† [44]0.2353.8826.460.525.27
PowerPaint† [46]0.4333.2726.150.465.20
BrushNet† [14]0.1203.9426.560.585.29
Ours0.0704.6226.500.675.34
+ +Table 1. The quantitative comparison of different methods. The items in bold and underline are ranked first and second. $\dagger$ means that the method is trained / finetuned with our dataset. + +SD-Inpainting [26], PowerPoint [46], ControlNet [44], BrushNet [14] and Layer Diffusion [43]. Among them, all except BLD are learning-based methods. ControlNet is trained using the OpenImage-FCO dataset. For a fair comparison, we also fine-tuned BrushNet and PowerPaint using the proposed dataset. The intuitive and quantitative comparison results are shown in Fig. 3 and Tab. 1 respectively. + +On the objective metric SAM, our method outperforms most comparison methods by at least 0.050. The only method that scores higher than our proposed method, Layer Diffusion, lacks competitiveness in other metrics due to poor coordination between the foreground and background. On the user-study metric INT, the average score given by users to our method was 0.56 higher than that of the second-best method. These results demonstrate the effectiveness of our method in improving image composition. The samples in Fig. 3 also intuitively illustrate this advantage. Among all compared methods, our approach produces artifact-free results across four cases. In terms of text-image alignment, our method achieves the second-best result. The first place is taken by BrushNet, partially due to its dataset. BrushNet was trained on LAION-Aesthetic [28] containing 120 million images, where abundant image-text pairs ensure advanced performance in text prompt controllability. + +The advantages in image composition also drive better performance in image quality metrics, with our method outperforming the second-best by 0.07 and 0.03 on IR and AS respectively. Notably, while suppressing artifacts and incorrect interactions, our method produces less cluttered backgrounds compared to others. This can be attributed to the distribution discrepancy between $x_{t^s}$ inferred by IPM and natural sampling. Fortunately, this characteristic also enhances visual hierarchy in synthesized images, making them more suitable for object presentation. + +A noteworthy result is that after fine-tuning on the proposed dataset, PowerPaint and BrushNet have shown signif + +![](images/d501d868eaf0710e5d86d6cb2364076ddb099898e1e5f3d73a809f9797c70f4d.jpg) + +![](images/d13bbb518902ad94acab5b848d5c49d70cd10926e99d2b1eabf4363e164228b1.jpg) +Figure 5. (a) The generation results of the baseline model with or w/o IPM under different text prompts. (b) The multi generation results of the baseline model with or w/o IPM in same case. + +![](images/ed545122eb3927a872cf965dd86202118b181f8545af126dc6e84d257c3a25b4.jpg) +Figure 6. The samples of predicted low-frequency signals and final generated images with or w/o the diffusion reward loss $\mathcal{L}_{rwd}$ . + +icant improvements in the SAM metric, which once again confirms the value of task-specific datasets. To further demonstrate the IPM's capabilities, we select foregrounds with diverse identities, scales, and positions as condition images and generate results under the same text prompt. The results in Fig. 4 show that our method adapts well to various scenarios and produces reasonable interactions between given objects and backdrops. + +# 4.4. The Influence of IPM on Generation Model + +Given that the integration of IPM fundamentally alters the original diffusion process, its implications for the under + +
SettingSAM↓CLIP↑IR↑AS↑
Baseline0.23526.460.525.27
w/o L_rwd0.10326.440.645.26
w/o U-Net Init.0.07426.290.635.28
Ours0.07026.500.675.34
+ +Table 2. The quantitative evaluation results of ablation study. + +lying generation model warrant detailed discussion. For image generation tasks, the primary concerns are semantic controllability and output diversity, both of which will be systematically addressed in this section. + +The generation results of the baseline model (ControlNet) with and without IPM under different text prompts $y$ are presented in Fig. 5(a). It can be observed that IPM does not significantly affect the model's controllability, as the model still synthesizes results with the required semantics. This finding is also supported by Tab. 1, where our method and ControlNet exhibit the closest CLIP scores among all approaches. The subsequent ablation studies will reveal that this benefit primarily stems from initializing IPM parameters with the pretrained U-Net. Turning to Fig. 5(b), the generation results of the baseline model with and without IPM for the same pair of $c$ and $y$ are shown. In this comparison, IPM results show more similar composition due to the shared low-frequency signals, while the baseline model's diverse outputs come at the cost of prevalent compositional defects. This reinforces that IPM trades inefficient diversity for compositional rationality. + +# 4.5. Ablation Study + +In this section, we conduct ablation studies on critical components of the IPM training paradigm and network design, including the diffusion reward supervision $\mathcal{L}_{rwd}$ and the strategy of initializing IPM parameters with a pretrained U-Net. Note that since removing $\mathcal{L}_{lbl}$ would destabilize training, we exclude it from ablation analysis. The experimental results are presented in Tab. 2. + +Ablation Study for $\mathcal{L}_{rwd}$ Quantitative validation shows that diffusion reward supervision provides the most significant benefit in artifact suppression. When $\mathcal{L}_{rwd}$ is removed, the SAM score increases by 0.033. The introduction of new artifacts also degrades image quality metrics, with AS dropping by 0.08. These artifacts correlate with the quality of low-frequency signals predicted by IPM. To visually corroborate this, we visualize these low-frequency signals (decoded via the Latent VAE) in Fig. 6. It can be observed that without $\mathcal{L}_{rwd}$ , IPM outputs smoother signals. A direct consequence is that edges of some foreground objects become indistinct, which may develop into artifacts and imprecise details after subsequent denoising steps. Conversely, when + +![](images/ff068861785720bf7260705db46e0f920b9a7275b06f459da6d08f8e2fd394a9.jpg) +- A brand-new football is placed on the grass in the football field. + +![](images/66ec616b3cd2c81fda55ea260feb5eab73f4cd2c26a87f52bac0465a526e7b72.jpg) +Figure 7. The samples of generated images under different $t^s$ . +Figure 8. The quantitative evaluation results of the truncating timestep selection. + +diffusion reward supervision is introduced, IPM is encouraged to output intermediate states that genuinely lead to target images, effectively mitigating this issue. + +Ablation Study for U-Net Initialization As shown in Tab. 2, initializing IPM parameters with a pretrained U-Net primarily strengthens the model's text controllability. Removing this strategy causes the CLIP score to drop by 0.21, even below the baseline performance. Although IPM is only responsible for predicting low-frequency signals, training it from scratch limits its performance on complex text prompts due to the dataset's relatively limited corpus. This phenomenon substantiates the necessity of leveraging the rich text-image priors of pretrained models. + +# 4.6. The Selection of Truncating Timestep + +The primary hyperparameter when applying IPM is determining the appropriate timestep $t^s$ to truncate the diffusion process. We test values within the set {900, 850, 800, 750, 700}. The intuitive and quantitative comparison results are shown in Fig. 7 and Fig. 8. + +When the truncation timestep $t^s$ is set too late, the method's effectiveness declines significantly. In the $t^s = 700$ experiment group, although artifacts are further suppressed, this comes at the cost of overly simplified backgrounds and significantly reduced image quality. Its IR and AS metrics are 0.35 and 0.60 lower than those of $t^s = 800$ . This phenomenon is associated with the function of IPM. IPM directly predicts the image's low-frequency signal. When $t^s$ is set to a smaller value, IPM's task becomes more complex. Meanwhile, the number of subsequent normal de + +![](images/3c401878804ca700c5644dd8c35a527b499d87851291814286201d76878db880.jpg) +Figure 9. The samples of failure case about the objects with extreme colors or lighting conditions. + +noising steps is reduced, requiring IPM to predict image signals across broader frequency bands. $t^s = 700$ is clearly an excessively challenging value. As illustrated in Fig. 7, the produced image exhibits obvious detail deficiencies due to the interplay of these factors. When $t^s = 900$ , the frequency band managed by IPM is too narrow, weakening its influence on final image. Its SAM score increases by 0.033 compared to $t^s = 800$ , indicating that subsequent denoising steps introduce more suboptimal compositions. In this paper, $t^s = 800$ is adopted across all experiments. g + +# 4.7. Failure case and Future Work + +During testing, a notable failure case arises when the given foreground object exhibits extreme colors or lighting conditions, significantly degrading background controllability. Fig. 9 illustrates results for such cases. Despite prompts' requirements, synthesized images still adopt styles consistent with the provided foreground (case 1) or fail to include other backdrops (case 2). Root causes include dataset limitations (artistic photography underrepresented in OpenImage) and framework design. That is, features of objects with extreme colors tend to dominate background features during synthesis, compelling background features to always reflect similar appearances. This motivates exploring methods to decouple foreground-background color distributions and enable foreground relighting in FCO tasks. + +# 5. Conclusion + +This paper focuses on the problem of flawed image composition in diffusion-based FCO methods. We believe that this problem can be partially attributed to the random impact of the initial noise. Inspired by this, we propose the Initialization Policy Model (IPM). It uses a deterministic model that only depends on the given foreground and text prompt to replace the early denoising steps, thus isolating the impact of the initial noise. The subsequently proposed comprehensive training paradigm ensures that the IPM can be effectively optimized. Experiments confirm that the proposed method can significantly improve the composition problems such as artifacts and incorrect interactions, and the comprehensive framework built upon IPM also achieves advanced performance in FCO task. + +Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant No. 62473290, 62088101, National Key Research and Development Program of China under Grant No. 2024YFB3311801, and the Shanghai Municipal Commission of Economy and Informatization under Grant No. 2024-GZL-RGZN-01008. + +# References + +[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18187-18197, 2022. 2 +[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM TOG, 42(4):149:1-149:11, 2023. 1, 2, 6 +[3] Rodrigo Benenson and Vittorio Ferrari. From colouring-in to pointillism: revisiting semantic segmentation supervision. CoRR, abs/2210.14142, 2022. 2, 5 +[4] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 2 +[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In CVPR, pages 22503-22513, 2023. 2 +[6] Shidong Cao, Wenhao Chai, Shengyu Hao, Yanting Zhang, Hangyue Chen, and Gaoang Wang. Diffashion: Reference-based fashion design with structure-aware transfer by diffusion models. TMM, 26:3962-3975, 2024. 1 +[7] Tingfeng Cao, Junsheng Kong, Xue Zhao, Wenqing Yao, Junwei Ding, Jinhui Zhu, and Jiandong Zhang. Product2img: Prompt-free e-commerce product background generation with diffusion model and self-improved lmm. In ACMMM, pages 10774-10783, 2024. 1 +[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, pages 1–31, 2024. 2 +[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, pages 6593-6602, 2024. 2 +[10] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. CoRR, abs/2210.11427, 2022. 2 +[11] Amir Erfan Eshratifar, Joao VB Soares, Kapil Thadani, Shaunak Mishra, Mikhail Kuznetsov, Yueh-Ning Ku, and Paloma De Juan. Salient object-aware background generation using text-guided diffusion models. In CVPR, pages 7489-7499, 2024. 1 +[12] Xuan Han, Yihao Zhao, and Mingyu You. Scene diffusion: Text-driven scene image synthesis conditioning on a single 3d model. In ACMMM, pages 7862-7870, 2024. 3 +[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2 + +[14] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. CoRR, abs/2403.06976, 2024. 1, 2, 6 +[15] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. In ICLR, pages 1-28, 2024. 4 +[16] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In NeurIPS, pages 29914–29934, 2024. 5, 6 +[17] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew P. Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, pages 105-114, 2017. 4 +[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742, 2023. 5 +[19] Pengzhi Li, Qinxuan Huang, Yikang Ding, and Zhiheng Li. Layerdiffusion: Layered controlled image editing with diffusion models. In ACM SIGGRAPH Asia, pages 1-4, 2023. 2 +[20] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, and Jianzhuang Liu. Zone: Zero-shot instruction-guided local editing. In CVPR, pages 6254-6263, 2024. 2 +[21] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: marrying DINO with grounded pre-training for open-set object detection. CoRR, abs/2303.05499, 2023. 5, 6 +[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, pages 1-11, 2019. 5 +[23] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 2 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 4 +[25] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. 2 +[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685, 2022. 1, 2, 3, 4, 6 +[27] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In AAAI, pages 4695-4703, 2024. 1 + +[28] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, pages 1–21, 2022. 6 +[29] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In CVPR, pages 8871-8879, 2024. 2 +[30] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent Y. F. Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In CVPR, pages 8839-8849, 2024. 2 +[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, pages 1-12, 2021. 3, 4, 6 +[32] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In CVPR, pages 18310-18319, 2023. 2 +[33] Zhengwentai Sun, Yanghong Zhou, Honghong He, and PY Mok. Sgdiff: A style guided diffusion model for fashion synthesis. In ACMMM, pages 8433-8442, 2023. 1 +[34] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, and Faisal Azhar. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971, 2023. 5 +[35] Ruoyu Wang, Yongqi Yang, Zhihao Qian, Ye Zhu, and Yu Wu. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, pages 1-21, 2024. 1 +[36] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, pages 18359-18369, 2023. 5 +[37] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, pages 18359-18369, 2023. 2 +[38] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In ICCV, pages 7378-7387, 2023. 4 +[39] Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, and Hongsheng Li. Deep reward supervisions for tuning text-to-image diffusion models. CoRR, abs/2405.00760, 2024. 4 + +[40] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437, 2023. 2 +[41] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In NeurIPS, pages 1-14, 2023. 6 +[42] Shiyuan Yang, Xiaodong Chen, and Jing Liao. Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In ACM MM, page 3190-3199, 2023. 2 +[43] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. ACM Transactions on Graphics (TOG), 43(4):1-15, 2024. 6 +[44] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3813-3824, 2023. 2, 5, 6 +[45] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize anything: A strong image tagging model. In CVPR, pages 1724-1732, 2024. 5 +[46] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. CoRR, abs/2312.03594, 2023. 2, 5, 6 \ No newline at end of file diff --git a/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/images.zip b/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..47ae31a3d1b0f3e3cf3b6d1da44c705201da0e00 --- /dev/null +++ b/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f972fe2148f1586836644f90ce5d18a438766ea0257e96d9b3985384b2820b27 +size 708794 diff --git a/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/layout.json b/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c7c201981664e9a1c6f05448dbd1df5441bcfd20 --- /dev/null +++ b/2025/Toward Better Out-painting_ Improving the Image Composition with Initialization Policy Model/layout.json @@ -0,0 +1,9154 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 93, + 102, + 519, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 102, + 519, + 140 + ], + "spans": [ + { + "bbox": [ + 93, + 102, + 519, + 140 + ], + "type": "text", + "content": "Toward Better Out-painting: Improving the Image Composition with Initialization Policy Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "spans": [ + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "text", + "content": "Xuan Han" + }, + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "text", + "content": " Yihao Zhao" + }, + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "text", + "content": " Yanhao Ge" + }, + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "text", + "content": " Mingyu You" + }, + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "text", + "content": " Tongji University, Shanghai, China \n" + }, + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 160, + 475, + 204 + ], + "type": "text", + "content": " Future Imaging Area \n{hanxuan, zhaoyihao, myyou}@tongji.edu.cn {halege}@vivo.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 256, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 256, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 297, + 605 + ], + "type": "text", + "content": "With its extensive applications, Foreground Conditioned Out-painting (FCO) has attracted considerable attention in the research field. Through the utilization of text-driven FCO, users are enabled to generate diverse backgrounds for a given foreground by adjusting the text prompt, which considerably enhances the efficiency in fields like e-commerce. Since the foreground is fixed in FCO, a key concern is whether the generated background can match the foreground well to achieve a coherent composition. However, most existing methods are lacking in this regard. Artifacts and incorrect interactions are common defects in synthesized images. This issue is linked to the influence of the initial noise in the sampling process. As the initial noise is sampled independently, it's highly likely that the implied image composition will conflict with the given foreground. In this paper, a novel Initialization Policy Model (IPM) is proposed to address this problem. Its function is to replace the early denoising steps and directly predict the intermediate state that is conducive to the reasonable image composition. Since the IPM is designed to take only the foreground image and the text prompt as inputs, it isolates the impact of the initial noise. The subsequently proposed training paradigm that combines inversion-derived label supervision and diffusion reward supervision further ensures the efficient training of the IPM. The evaluation is conducted using the task-specific OpenImage-FCO dataset developed by us. The results verify that the introduction of the IPM can significantly improve the composition of the synthesized images and achieve advanced performance in the FCO task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 628, + 135, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 628, + 135, + 640 + ], + "spans": [ + { + "bbox": [ + 56, + 628, + 135, + 640 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 649, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 295, + 674 + ], + "type": "text", + "content": "Nowadays, diffusion-based image editing has emerged as a trending research topic. Its progress in numerous tech-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 234, + 554, + 336 + ], + "blocks": [ + { + "bbox": [ + 317, + 234, + 554, + 336 + ], + "lines": [ + { + "bbox": [ + 317, + 234, + 554, + 336 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 554, + 336 + ], + "type": "image", + "image_path": "fd8241776b463b7c8b42d48a7f02b84badb2ea3d5c6ee15389caaef35349f84e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 345, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 313, + 345, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 555, + 368 + ], + "type": "text", + "content": "Figure 1. The illustrations concerning the flawed image composition in foreground conditioned out-painting task." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 389, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 533 + ], + "type": "text", + "content": "niques is continuously revolutionizing the practices in related fields such as commercial art [7, 11] and fashion design [6, 33]. Among the various techniques available, Foreground Conditioned Out-painting (FCO) [2, 14, 26] has attracted significant attention. FCO focuses on a special use case: The foreground of an image, usually containing the main object, is fixed, and image generators like Stable Diffusion [26] are used to fill the background based on the user's prompt. This technology allows users to generate diverse backgrounds for the given foreground simply by changing the text prompt, which significantly enhances efficiency compared to standard workflows." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 665 + ], + "type": "text", + "content": "In this task, a crucial concern is the quality of image composition. Take Fig. 1 as an example: the user wants to display the mug in a fresh scene. The expected result is that the synthesized background can be harmoniously integrated with the mug. However, existing methods demonstrate insufficient performance in this regard. In many results, background objects like the table fail to interact appropriately with the given mug, alongside artifacts generated in the process. Overall, the synthesized background often exists independently of the given foreground rather than forming a unified composition." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "To identify the source of the problem, we revisit the specificity of the FCO task. In diffusion-based FCO, images are gradually denoised from initial noise " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": ". According to existing literature [27, 35], variations in " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": " compre" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "*Corresponding Author. M. You is with the College of Electronic and Information Engineering, Tongji University, State Key Laboratory of Autonomous Intelligent Unmanned Systems" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16938" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "hensively influence the position, scale, and pose of synthesized objects. However, in FCO, independently sampled " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " frequently implies compositions conflicting with the fixed foreground. As shown in Fig. 1, different " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " values lead to diverse table poses, many of which clash with the mug; artifacts are also generated to match these inconsistent backdrops. Evidently, addressing this issue requires mitigating the random impact of " + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": " on image composition." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "text", + "content": "In this paper, a novel Initialization Policy Model (IPM) is introduced to address the problem. The method exploits the time-frequency coupling of the diffusion process, by which low-frequency signals determining image composition are primarily generated during early denoising steps [13]. In the proposed approach, the denoising process is truncated at timestep " + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "inline_equation", + "content": "t^s" + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "text", + "content": ", with the IPM directly predicting the intermediate state " + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "inline_equation", + "content": "x_{t^s}" + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "text", + "content": ". By conditioning solely on the foreground " + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "text", + "content": ", IPM isolates image low-frequency signals from random " + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "inline_equation", + "content": "x_T" + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "text", + "content": " influence. Subsequent normal denoising processes synthesize rich high-frequency details thereon. To facilitate the optimization of IPM, a comprehensive training paradigm integrating label supervision and reward supervision is proposed. Label supervision uses inversion-derived low-frequency signal ground truth to guide IPM output. Reward supervision subjects predicted " + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "inline_equation", + "content": "x_{t^s}" + }, + { + "bbox": [ + 56, + 168, + 296, + 430 + ], + "type": "text", + "content": " to multiple denoising rounds, with the error between the final image and target image serving as the training reward criterion. We validate the method on our task-specific OpenImage-FCO dataset. The results confirm that it can notably improve the image composition in FCO task. The contributions of this work are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 433, + 295, + 647 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 433, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 503 + ], + "type": "text", + "content": "- We propose a novel Initialization Policy Model (IPM) to enhance image composition in the FCO task. By using the IPM to directly predict the low-frequency signals that match the given foreground, it effectively suppresses the random influence of the initial noise, thus alleviating the issue of flawed composition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 505, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 505, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 505, + 295, + 586 + ], + "type": "text", + "content": "- We devise a comprehensive training paradigm for the IPM. In addition to label supervision, we introduce novel reward supervision based on a pre-trained generative model, which has been proven to effectively increase the quality of the signals predicted by the IPM, resulting in the final image with reduced artifacts and improved detail accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 588, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 295, + 647 + ], + "type": "text", + "content": "- We construct the OpenImage-FCO dataset based on OpenImage [3] and an automated annotation pipeline. It consists of approximately 580K images, 800K filtered high-quality masks, and their associated tags, providing substantial support for research in the FCO field." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 657, + 146, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 146, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 146, + 670 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Leveraging the abundant priors of large pre-trained models [8, 25, 26], diffusion-based approaches have achieved remarkable progress in many image editing tasks such as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "object editing [4, 9, 29, 32], attribute editing [9, 19], and structure control [5, 30, 44]. This paper focuses on the Foreground Conditioned Out-painting (FCO) task. Related methods can be categorized into two types: learning-free and learning-based approaches. In this section, we introduce representative works from both categories and analyze their application characteristics." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 159, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 555, + 421 + ], + "type": "text", + "content": "Learning-Free Method In learning-free methods, the most prominent one is Blended latent Diffusion [1, 2]. In this method, the unknown region (background) is sampled through the standard denoising process. After each step, the known region (foreground) is pasted into the current result. This straightforward pipeline allows it to be plug-and-play in any pre-trained model. However, since the two regions are not acquired from the unified process, it is challenging to achieve a seamless combination between them. Subsequent research [10, 20] has strived to refine the pipeline. For example, DiffEdit [10] employs DDIM Inversion [23] to obtain the features of the known region, thereby reducing the distribution disparity between the features of the two regions. Although the related research is still advancing, at present, the mainstream learning-free methods still adopt the above spatially separated sampling paradigm. A common concern is that it hinders the feature interaction between the two regions, resulting in the inability to achieve mask-aware image completion [14]. With respect to the FCO task, the background synthesized by this type of method often struggles to exhibit a reasonable interaction with the given foreground." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 423, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 556, + 663 + ], + "type": "text", + "content": "Learning-based Method Many studies [14, 37, 44, 46] to improve the synthesis effect through task-specific learning. A representative one is SD-Inpainting [26]. It fine-tunes the pre-trained model with a large-scale in-painting dataset and can be directly used for various image completion applications. Since the fine-tuning process involves most of the parameters of Stable Diffusion, the original diffusion prior in it is also affected to a certain extent. Subsequent research [40, 42] has developed a more refined framework. For example, PowerPoint [46] trains an additional task prompt to enhance the model's ability to fill images. After that, BrushNet [14] introduces a twin denoising network to process the features of the known region and optimizes the entire framework end-to-end. One of the most concerned solutions among them is ControlNet [44]. This method trains additional input branches to achieve various conditional generation tasks. When dealing with image completion, the new branch takes the RGB image and mask of the known area as the input. The feature of parameter invariance in the original network enables its direct combination with other methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "It must be pointed out that many studies mentioned above take general image completion rather than FCO as the major task. This is also reflected in the data they use. For example, the masks in the dataset of SD-Inpainting" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16939" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 76, + 265, + 232 + ], + "blocks": [ + { + "bbox": [ + 58, + 76, + 265, + 232 + ], + "lines": [ + { + "bbox": [ + 58, + 76, + 265, + 232 + ], + "spans": [ + { + "bbox": [ + 58, + 76, + 265, + 232 + ], + "type": "image", + "image_path": "04692cb158d94516b67758199e09daad01b183bb08f8a61843da6564cbbb3f89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 555, + 264 + ], + "type": "text", + "content": "Figure 2. (a) The illustration about the differences between the proposed truncated diffusion process and the original diffusion process. (b) The principle and training paradigm of Initialization Policy Model (IPM)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 274, + 76, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 274, + 76, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 274, + 76, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 274, + 76, + 553, + 232 + ], + "type": "image", + "image_path": "7e90ca50fb8a810f2a0c70266bfb7b3f037f4429982d9a86f4b7612fcfe04186.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 283, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 283, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 296, + 392 + ], + "type": "text", + "content": "[26] are mostly randomly generated, which implies that the known region would contain some clues about the unknown part. But in the FCO task, the known region is the well-segmented foreground object. These differences will undoubtedly cause a series of issues when these methods are applied to FCO. The flawed image composition is one of the most significant problems. In this work, we conduct in-depth research on the FCO task and propose targeted solutions to improve the composition of synthesized images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 401, + 211, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 211, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 211, + 414 + ], + "type": "text", + "content": "3. Truncated Diffusion Model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 421, + 123, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 123, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 123, + 432 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "text", + "content": "This paper centers on the FCO task, with the aim of enhancing the image composition of generated images. The research primarily focuses on diffusion-based methods, where the image generation process is modeled as a series of sequential denoising steps commencing from random noise " + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "text", + "content": " denotes the number of timesteps. Specifically for text-driven FCO, this denoising process is guided by both the text prompt " + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "text", + "content": " and the condition image " + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "text", + "content": ". The text prompt " + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "text", + "content": " describes the semantics of the target image, while the condition image " + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "text", + "content": " is typically the RGB image of the foreground we intend to present. The diffusion-based FCO model can be formally expressed as " + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{0:T-1}|x_{T},y,c)" + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 54, + 437, + 295, + 594 + ], + "type": "text", + "content": " represents the denoising model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Our method originates from rethinking the influence of " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " on image composition. According to existing research, even when using the same prompt, the results sampled from different " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " values will differ. The pose, scale, and position of each synthesized object will change. Given that " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " is completely independent, the image composition it implies is likely to conflict with the given foreground. As depicted in Fig. 2(a), we propose truncating the denoising process at a specific time step " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "t^{s}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " and substituting the early steps with an Initialization Policy Model (IPM) " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": ". This method exploits" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "text", + "content": "the time-frequency coupling in the diffusion process, where the low-frequency signals that determine the image composition are mainly formed in the early stage. " + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "text", + "content": " is defined to be solely conditioned on the foreground image " + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "text", + "content": " and the text prompt " + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "text", + "content": ", thereby isolating the random impact of " + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "text", + "content": " on the image composition. As a result, the original model is disassembled into " + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "inline_equation", + "content": "p_{\\phi}(x_{ts}|y,c)" + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{0:t^s +1}|x_{ts},y,c)" + }, + { + "bbox": [ + 313, + 283, + 555, + 403 + ], + "type": "text", + "content": ". In this section, we first introduce the principle and network design of the IPM, followed by the detailed presentation of each component in the training paradigm." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 409, + 462, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 462, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 462, + 422 + ], + "type": "text", + "content": "3.2. Initialization Policy Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "type": "text", + "content": "The proposed method introduces a new IPM to obtain the intermediate state " + }, + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "type": "inline_equation", + "content": "x_{t^s}" + }, + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "type": "text", + "content": ". According to the principle of the diffusion model, each intermediate state " + }, + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "type": "text", + "content": " follows the distribution [31]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 368, + 487, + 555, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 487, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 368, + 487, + 555, + 499 + ], + "type": "interline_equation", + "content": "x _ {t} \\sim \\mathcal {N} \\left(x _ {t}; \\sqrt {\\alpha_ {t}} \\tilde {x} _ {0}, (1 - \\alpha_ {t}) \\mathbf {I}\\right) \\tag {1}", + "image_path": "f4dcd2ad391d4c7a07c6d39c0d613f88042e45df61b86d7de908295111266eeb.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 503, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 554, + 564 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 503, + 554, + 564 + ], + "type": "text", + "content": " the parameter of the diffusion schedule. " + }, + { + "bbox": [ + 313, + 503, + 554, + 564 + ], + "type": "inline_equation", + "content": "\\tilde{x}_0" + }, + { + "bbox": [ + 313, + 503, + 554, + 564 + ], + "type": "text", + "content": " represents the prediction of the target image " + }, + { + "bbox": [ + 313, + 503, + 554, + 564 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 503, + 554, + 564 + ], + "type": "text", + "content": " at the current timestep, essentially serving as the deterministic signal after denoising steps from " + }, + { + "bbox": [ + 313, + 503, + 554, + 564 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 503, + 554, + 564 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 503, + 554, + 564 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 313, + 503, + 554, + 564 + ], + "type": "text", + "content": ". To better clarify the meaning of each parameter, we rewrite Eq. (1) as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 379, + 574, + 555, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 574, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 379, + 574, + 555, + 602 + ], + "type": "interline_equation", + "content": "x _ {t} = \\underbrace {\\sqrt {\\alpha_ {t}} \\tilde {x} _ {0}} _ {\\text {S i g n a l}} + \\underbrace {\\sqrt {1 - \\alpha_ {t}} \\epsilon} _ {\\text {N o i s e}} \\tag {2}", + "image_path": "93cbb22b5a4fb9da758c44d43edd9517f0c53cdec65e1cb22e3db5e4ae0e7cc2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "When " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " is large, " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{x}_0" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " primarily contains low-frequency signals [12], which serve as the key determinants of image composition. " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " represents random noise, and different " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " lead to distinct high-frequency details in generated images. In this method, " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " is tasked with predicting the signal component " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{x}_0" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ". During training, the IPM is required to learn the low-frequency signals corresponding to real target images. During sampling, its output is combined with random noise to generate diverse images with coherent compositions." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16940" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "Notably, theoretically, there exists a broad range of potential compositions compatible with given " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": ", and the ideal " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "x_{t^s}" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " is expected to follow a complex distribution. However, in practice, it is challenging to obtain multiple real target images for a specific pair of " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": ". Thus, the role of the IPM is to provide a specific cluster within the ideal distribution of " + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "inline_equation", + "content": "x_{t^s}" + }, + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": ". In the experiments section, we further discuss its impact on the original image generation model and how it helps the FCO model achieve a balance between diversity and coherence." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "text", + "content": "Regarding the network architecture, the IPM " + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "text", + "content": " is directly initialized from the U-Net of Stable Diffusion [26]. This strategy aims to inherit the rich text-image priors embedded in its parameters. The text prompt " + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "text", + "content": " (encoded by CLIP [24]) is integrated through cross-attention layers. The network's input layer is expanded with four additional channels to accommodate the condition image " + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "text", + "content": " (encoded by the Latent VAE [26]). Furthermore, considering the U-Net was originally trained as a denoising model, a fixed noise strategy is introduced. Specifically, a static noise " + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "inline_equation", + "content": "x_{T}^{st}" + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "text", + "content": " is fed into the original channels. The final prediction of the low-frequency signal is computed from the network output and " + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "inline_equation", + "content": "x_{T}^{st}" + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "text", + "content": " according to the diffusion schedule: " + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "inline_equation", + "content": "\\sqrt{1 / \\alpha_{t^s}} x_T^{st} - \\sqrt{1 / \\alpha_{t^s} - 1}\\phi (y,c)" + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "text", + "content": ". Since the schedule parameters and noise in this formula remain constant across training and sampling, we hereinafter use " + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "inline_equation", + "content": "\\phi (y,c)" + }, + { + "bbox": [ + 55, + 194, + 298, + 399 + ], + "type": "text", + "content": " to denote the predicted low-frequency signal for simplicity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 415, + 249, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 249, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 249, + 428 + ], + "type": "text", + "content": "3.3. Inversion-derived Label Supervision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "type": "text", + "content": "Supervision with reliable ground truth is one of the most effective training paradigms. For the prediction objectives of the IPM, the ground truth can be derived through the inversion of the diffusion process [15, 31, 38]. The inversion operation proceeds in the reverse direction of the sampling process, involving stepwise denoising from the target image " + }, + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "type": "text", + "content": " to the corresponding initial noise " + }, + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "type": "text", + "content": ". By adhering to the identical noise schedule, the trajectory of intermediate state " + }, + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "type": "text", + "content": " obtained through inversion demonstrates high congruence with that of normal sampling operations, thereby qualifying it as valid ground truth. In this work, the widely adopted DDIM Inversion [31] is implemented, which operates within the range " + }, + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "type": "inline_equation", + "content": "t\\in [0,t^s]" + }, + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "type": "text", + "content": " to compute the ground truth of the intermediate state at truncation timestep " + }, + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "type": "inline_equation", + "content": "x_{ts}^{gt}" + }, + { + "bbox": [ + 55, + 435, + 295, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 606, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 295, + 677 + ], + "type": "text", + "content": "Based on the results of the inversion process, the ground truths of the signal and random noise can be further obtained: " + }, + { + "bbox": [ + 55, + 606, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{x}_0^{gt} = \\sqrt{1 / \\alpha_{t^s}} x_{t^s} - \\sqrt{1 / \\alpha_{t^s} - 1}\\theta (x_t,y)" + }, + { + "bbox": [ + 55, + 606, + 295, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 606, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\epsilon^{gt} = \\theta (x_t,y)" + }, + { + "bbox": [ + 55, + 606, + 295, + 677 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 606, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 606, + 295, + 677 + ], + "type": "text", + "content": " represents the pretrained diffusion model. " + }, + { + "bbox": [ + 55, + 606, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\tilde{x}_0^{gt}" + }, + { + "bbox": [ + 55, + 606, + 295, + 677 + ], + "type": "text", + "content": " will be used as a label to supervise the training of the IPM, and the loss function is as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 698, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 698, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 124, + 698, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {l b l} = \\left\\| \\phi (y, c) - \\tilde {x} _ {0} ^ {g t} \\right\\| _ {2} ^ {2} \\tag {3}", + "image_path": "bacdda93e870d49f8ca83d2904af687f59e8a205662472a99d1e112a2576250a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 478, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 478, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 478, + 84 + ], + "type": "text", + "content": "3.4. Diffusion Reward Supervision" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "type": "text", + "content": "In our approach, the IPM is tasked with predicting the low-frequency signals of images. Solely using the MSE loss as " + }, + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lbl}" + }, + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "type": "text", + "content": " may cause it to produce over-smoothed results [17]. Not to mention that results obtained via inversion operation inherently introduce deviations [15]. Therefore, we propose Diffusion Reward Supervision to assist in the training of the IPM. Its principle is illustrated in Fig. 2(b). The " + }, + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "type": "inline_equation", + "content": "x_{t^s}" + }, + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "type": "text", + "content": " sampled by " + }, + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "type": "text", + "content": " will be input into the pre-trained diffusion model (reward model) " + }, + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "type": "text", + "content": " to perform multiple rounds of denoising, and the error between the resulting image and the target image " + }, + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 89, + 553, + 220 + ], + "type": "text", + "content": " will serve as the reward to supervise the training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 221, + 553, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 553, + 303 + ], + "type": "text", + "content": "The diffusion reward adopts the same fast denoising schedule as DDIM [31]. Assuming the number of denoising steps is " + }, + { + "bbox": [ + 313, + 221, + 553, + 303 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 221, + 553, + 303 + ], + "type": "text", + "content": ", the sampling step size would be " + }, + { + "bbox": [ + 313, + 221, + 553, + 303 + ], + "type": "inline_equation", + "content": "\\lfloor (T - t^s) / K \\rfloor" + }, + { + "bbox": [ + 313, + 221, + 553, + 303 + ], + "type": "text", + "content": ". To enhance the diversity of the sampling schedules, the step size of the first step is set to a random number within " + }, + { + "bbox": [ + 313, + 221, + 553, + 303 + ], + "type": "inline_equation", + "content": "[1, \\lfloor (T - t^s) / K \\rfloor]" + }, + { + "bbox": [ + 313, + 221, + 553, + 303 + ], + "type": "text", + "content": ". Each denoising step can be formulated as follow:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 327, + 324, + 553, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 324, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 327, + 324, + 553, + 339 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = \\sqrt {\\alpha_ {t - 1} / \\alpha_ {t}} x _ {t} - \\sqrt {\\alpha_ {t - 1} / \\alpha_ {t} - 1} \\theta (x _ {t}, y) \\tag {4}", + "image_path": "cdffe6d358aaef2a25f152bfb6d1e9596d1e793113fa9cc16bede40acd11ab40.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 348, + 553, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 553, + 395 + ], + "type": "text", + "content": "In principle, the gradients generated by the final reward can be back-propagated according to the chain rule to guide the training of " + }, + { + "bbox": [ + 313, + 348, + 553, + 395 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 348, + 553, + 395 + ], + "type": "text", + "content": ". The gradients between neighboring timesteps can be expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 367, + 407, + 553, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 407, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 367, + 407, + 553, + 420 + ], + "type": "interline_equation", + "content": "\\partial x _ {t - 1} = a _ {t} \\partial x _ {t} + b _ {t} \\partial \\theta (x _ {t}, y) \\tag {5}", + "image_path": "59517f6a555517899f5fb5dfa7d74580a0bf61993b3d5f7cc7e80799bed2a370.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "b_{t}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " are used to simplify the schedule parameters, where " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "a_{t} = \\sqrt{\\alpha_{t - 1} / \\alpha_{t}}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "b_{t} = \\sqrt{\\alpha_{t - 1} / \\alpha_{t} - 1}" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": ". Eq. (5) reveals a drawback of the diffusion reward: the propagation of the gradient requires repeated passage through the reward model " + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 426, + 555, + 521 + ], + "type": "text", + "content": ", which will occupy a large amount of video memory. To solve this problem, we refer to the work of Wu et al. [39] and use the technique of stop gradients here. The modified gradients between neighboring timesteps become:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 399, + 533, + 553, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 533, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 399, + 533, + 553, + 545 + ], + "type": "interline_equation", + "content": "\\partial x _ {t - 1} = a _ {t} \\partial x _ {t} \\tag {6}", + "image_path": "23576dcc678d5bada15565c4ce6fb3989880c728dd6107a8d9d2d634c983a866.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 551, + 553, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 553, + 611 + ], + "type": "text", + "content": "This operation reduces the accuracy of the gradients to some extent, but it also makes it possible to conduct the training with limited video memory. We use " + }, + { + "bbox": [ + 313, + 551, + 553, + 611 + ], + "type": "inline_equation", + "content": "msd(*)" + }, + { + "bbox": [ + 313, + 551, + 553, + 611 + ], + "type": "text", + "content": " to represent the multi-step denoising operation, and the diffusion reward can be expressed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 629, + 553, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 629, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 319, + 629, + 553, + 646 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r w d} = \\left\\| m s d \\left(\\sqrt {\\alpha_ {t ^ {s}}} \\phi (y, c) + \\sqrt {1 - \\alpha_ {t ^ {s}}} \\epsilon^ {g t}\\right) - x _ {0} \\right\\| _ {2} ^ {2} \\tag {7}", + "image_path": "8d2bb98e8b078bfe75918e0e4689de9fb3fe63369d9a0366208f69576692c2f3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon^{gt}" + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": " is the noise corresponding to the target image obtained during the inversion process (see Sec. 3.3). The meaning of " + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rwd}" + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": " is that we expect the predicted low-frequency signal plus the ground truth of the noise to restore the corresponding target image. Experiments have" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16941" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 281 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 281 + ], + "type": "image", + "image_path": "a387335729bae23cef4e3dbb228a1c2edeb5560a39d03ca2c95dba8e64ad46c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 289, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 555, + 313 + ], + "type": "text", + "content": "Figure 3. The samples of generated images of the comparison methods. All compared methods are implemented with the recommended settings. " + }, + { + "bbox": [ + 55, + 289, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 289, + 555, + 313 + ], + "type": "text", + "content": " means that the method is trained / finetuned with our dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 333, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 296, + 392 + ], + "type": "text", + "content": "confirmed that " + }, + { + "bbox": [ + 55, + 333, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rwd}" + }, + { + "bbox": [ + 55, + 333, + 296, + 392 + ], + "type": "text", + "content": " can effectively enhance the quality of the signals predicted by the IPM, resulting in the final image with reduced artifacts and improved detail accuracy. The final loss function is as follows, where " + }, + { + "bbox": [ + 55, + 333, + 296, + 392 + ], + "type": "inline_equation", + "content": "\\lambda_{rwd}" + }, + { + "bbox": [ + 55, + 333, + 296, + 392 + ], + "type": "text", + "content": " is the weight of the reward supervision term:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 129, + 407, + 294, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 407, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 129, + 407, + 294, + 419 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {l b l} + \\lambda_ {r w d} \\mathcal {L} _ {r w d} \\tag {8}", + "image_path": "707c281d084649a02a24513318469dcd9f4142787b2af47a3b9ba0aa52a78f8a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 499 + ], + "type": "text", + "content": "As described in Sec. 3.2, during sampling, the output of " + }, + { + "bbox": [ + 55, + 426, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 426, + 296, + 499 + ], + "type": "text", + "content": " is combined with random noise to obtain the intermediate state: " + }, + { + "bbox": [ + 55, + 426, + 296, + 499 + ], + "type": "inline_equation", + "content": "x_{ts} = \\sqrt{\\alpha_{ts}}\\phi (y,c) + \\sqrt{1 - \\alpha_t}\\epsilon" + }, + { + "bbox": [ + 55, + 426, + 296, + 499 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 426, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 55, + 426, + 296, + 499 + ], + "type": "text", + "content": ". Subsequent denoising steps are completed by conventional diffusion-based FCO. In this paper, a ControlNet trained on the same dataset is employed in this phase." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 137, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 137, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 137, + 523 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 531, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 296, + 615 + ], + "type": "text", + "content": "Numerical experiments are conducted to evaluate the proposed method. This section begins with an introduction to the dataset and experimental setups. Afterward, the comparison between IPM and state-of-the-art methods is reported. Then, the impact of IPM on the generative model is discussed. Finally, ablation studies and hyperparameter analyses for the proposed method are presented." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 114, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 114, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 114, + 635 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "The advancement of the FCO technique demands high-quality task-specific datasets. Existing datasets, despite their large scale, often lack sufficient diversity in masked objects and their saliency in images. For instance, PowerPaint [46] utilized the OpenImage [3] segmentation subset (1.74M images, 2.8M masks) but focused on only 600" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 333, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 380 + ], + "type": "text", + "content": "object classes, many of which were not the primary subjects in their images. To address this limitation, we introduce OpenImage-FCO, a novel task-specific dataset built on OpenImage-v7 [3]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 383, + 556, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 556, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 556, + 526 + ], + "type": "text", + "content": "OpenImage-FCO incorporates multi-modal filtering and semantic reasoning to enhance annotation quality. We leverage BLIP2 [18] and RAM [45] for image captioning and tagging, followed by semantic filtering via Llama3 [34] to retain only focused objects. GroundingDINO [21] and HQSAM [16] are used to generate precise object masks. The final dataset includes 580K images with 800K masks, 6K unique object tags, and image captions. For evaluation, we curated a test set of 250 cases, each equipped with manual foreground mask annotations and a text prompt. Its scale is aligned with EditBench [36], a widely adopted benchmark in the image editing domain, which encompasses 240 cases." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 539, + 441, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 441, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 441, + 552 + ], + "type": "text", + "content": "4.2. Experimental Settings" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": "Implementation Details Our method is built upon Stable Diffusion V2.1 with an image size of " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": ". When training the IPM, the truncation timestep " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "t^s" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": " is set to 800, the diffusion reward parameter " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": " is set to 10, and " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{rwd}" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": " is set to 1. The AdamW optimizer [22] is employed with a constant learning rate of " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": ". Training is conducted on four Nvidia L40S GPUs with a batch size of 8 and a total of 400K iterations. The FCO ControlNet used for subsequent inference steps adopts the same structure as described in the original paper [44]. The AdamW optimizer is configured with a learning rate of " + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": ", a batch size of 8, and 200K iterations. Notably, all backbone parameters, including Cross-Attn layers, were frozen during Control-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16942" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 83, + 116, + 141 + ], + "blocks": [ + { + "bbox": [ + 71, + 72, + 101, + 81 + ], + "lines": [ + { + "bbox": [ + 71, + 72, + 101, + 81 + ], + "spans": [ + { + "bbox": [ + 71, + 72, + 101, + 81 + ], + "type": "text", + "content": "Cond. 1" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 83, + 116, + 141 + ], + "lines": [ + { + "bbox": [ + 58, + 83, + 116, + 141 + ], + "spans": [ + { + "bbox": [ + 58, + 83, + 116, + 141 + ], + "type": "image", + "image_path": "6168034b513bbedc4b15d86f6427bcdc455e03cb7cd4b4f505774fa0bed69d63.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 142, + 266, + 150 + ], + "lines": [ + { + "bbox": [ + 58, + 142, + 266, + 150 + ], + "spans": [ + { + "bbox": [ + 58, + 142, + 266, + 150 + ], + "type": "text", + "content": "- A Siamese cat stands on a stone, surrounded by purple flowers. in a sunlit garden." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 117, + 83, + 175, + 141 + ], + "blocks": [ + { + "bbox": [ + 131, + 72, + 161, + 80 + ], + "lines": [ + { + "bbox": [ + 131, + 72, + 161, + 80 + ], + "spans": [ + { + "bbox": [ + 131, + 72, + 161, + 80 + ], + "type": "text", + "content": "Cond. 2" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 117, + 83, + 175, + 141 + ], + "lines": [ + { + "bbox": [ + 117, + 83, + 175, + 141 + ], + "spans": [ + { + "bbox": [ + 117, + 83, + 175, + 141 + ], + "type": "image", + "image_path": "78daf4903e3afb1e64d63a6a030738fd9c163dd5c45f587687f4c20afacd2d66.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 176, + 83, + 235, + 141 + ], + "blocks": [ + { + "bbox": [ + 190, + 72, + 220, + 80 + ], + "lines": [ + { + "bbox": [ + 190, + 72, + 220, + 80 + ], + "spans": [ + { + "bbox": [ + 190, + 72, + 220, + 80 + ], + "type": "text", + "content": "Cond. 3" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 176, + 83, + 235, + 141 + ], + "lines": [ + { + "bbox": [ + 176, + 83, + 235, + 141 + ], + "spans": [ + { + "bbox": [ + 176, + 83, + 235, + 141 + ], + "type": "image", + "image_path": "04a8937f0fa46ab3d8a2c619fb6a994f9478e80809622ce63812b10d932c1344.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 236, + 83, + 294, + 141 + ], + "blocks": [ + { + "bbox": [ + 249, + 72, + 280, + 80 + ], + "lines": [ + { + "bbox": [ + 249, + 72, + 280, + 80 + ], + "spans": [ + { + "bbox": [ + 249, + 72, + 280, + 80 + ], + "type": "text", + "content": "Cond. 4" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 236, + 83, + 294, + 141 + ], + "lines": [ + { + "bbox": [ + 236, + 83, + 294, + 141 + ], + "spans": [ + { + "bbox": [ + 236, + 83, + 294, + 141 + ], + "type": "image", + "image_path": "dceff543d3b90d98f36dd6aeeffe7235b01c8fa884596f6fc6c0a3d31d054a34.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 151, + 115, + 209 + ], + "blocks": [ + { + "bbox": [ + 58, + 151, + 115, + 209 + ], + "lines": [ + { + "bbox": [ + 58, + 151, + 115, + 209 + ], + "spans": [ + { + "bbox": [ + 58, + 151, + 115, + 209 + ], + "type": "image", + "image_path": "3b5e1e78d62cbdc1d0fa65405254b0f9655941cf398df6b4f27f8e4ebbea6dbb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 209, + 246, + 217 + ], + "lines": [ + { + "bbox": [ + 58, + 209, + 246, + 217 + ], + "spans": [ + { + "bbox": [ + 58, + 209, + 246, + 217 + ], + "type": "text", + "content": "- A train speeds along the tracks, surrounded by lush green trees and hills." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 116, + 151, + 175, + 209 + ], + "blocks": [ + { + "bbox": [ + 116, + 151, + 175, + 209 + ], + "lines": [ + { + "bbox": [ + 116, + 151, + 175, + 209 + ], + "spans": [ + { + "bbox": [ + 116, + 151, + 175, + 209 + ], + "type": "image", + "image_path": "a3dac370d202ef8074b28f5afd65ce67379a24deef89afa313bd942f5522055e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 176, + 151, + 235, + 209 + ], + "blocks": [ + { + "bbox": [ + 176, + 151, + 235, + 209 + ], + "lines": [ + { + "bbox": [ + 176, + 151, + 235, + 209 + ], + "spans": [ + { + "bbox": [ + 176, + 151, + 235, + 209 + ], + "type": "image", + "image_path": "aa0c5094d83ec7d107242c021e660927b11040e77382ee5643bc82391afd2b9b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 236, + 151, + 294, + 209 + ], + "blocks": [ + { + "bbox": [ + 236, + 151, + 294, + 209 + ], + "lines": [ + { + "bbox": [ + 236, + 151, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 236, + 151, + 294, + 209 + ], + "type": "image", + "image_path": "1f38ee668fff1e360d2735fd9e8101bd45d8356174494af4b05091ef2000f1d1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 58, + 217, + 115, + 276 + ], + "blocks": [ + { + "bbox": [ + 58, + 217, + 115, + 276 + ], + "lines": [ + { + "bbox": [ + 58, + 217, + 115, + 276 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 115, + 276 + ], + "type": "image", + "image_path": "7028898ee2daf61a550b04bb9f1b721183a1c3a1f0e274da309710e78ddb6010.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 277, + 241, + 284 + ], + "lines": [ + { + "bbox": [ + 59, + 277, + 241, + 284 + ], + "spans": [ + { + "bbox": [ + 59, + 277, + 241, + 284 + ], + "type": "text", + "content": "- A comfortable chair sits in the corner of a charmingly cozy living room." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 293, + 295, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 326 + ], + "type": "text", + "content": "Figure 4. The generated images of foreground objects with variable identities, scales, and positions. The low-frequency signals predicted by IPM are shown in the supplementary materials." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 116, + 217, + 175, + 276 + ], + "blocks": [ + { + "bbox": [ + 116, + 217, + 175, + 276 + ], + "lines": [ + { + "bbox": [ + 116, + 217, + 175, + 276 + ], + "spans": [ + { + "bbox": [ + 116, + 217, + 175, + 276 + ], + "type": "image", + "image_path": "504b14b743ebcfc57333c03d30cbdd69678868dc9725fc9da8d3494382eb7370.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 176, + 217, + 235, + 276 + ], + "blocks": [ + { + "bbox": [ + 176, + 217, + 235, + 276 + ], + "lines": [ + { + "bbox": [ + 176, + 217, + 235, + 276 + ], + "spans": [ + { + "bbox": [ + 176, + 217, + 235, + 276 + ], + "type": "image", + "image_path": "765eb4c89557dd2446d8af5e11b75cb78881018e7641c3018ae4cbda99154461.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 236, + 217, + 294, + 276 + ], + "blocks": [ + { + "bbox": [ + 236, + 217, + 294, + 276 + ], + "lines": [ + { + "bbox": [ + 236, + 217, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 236, + 217, + 294, + 276 + ], + "type": "image", + "image_path": "90d228738ab98fb726ffa3b173a65fccb74f2d8c22b1fd819ea4c12b86b82891.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 54, + 349, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 349, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 349, + 295, + 420 + ], + "type": "text", + "content": "Net training to maximize the preservation of backbone text-image priors. All synthesized images presented in this article are sampled using the following strategy: After obtaining " + }, + { + "bbox": [ + 54, + 349, + 295, + 420 + ], + "type": "inline_equation", + "content": "x_{t^s}" + }, + { + "bbox": [ + 54, + 349, + 295, + 420 + ], + "type": "text", + "content": " from the IPM, the DDIM sampler [31] completes the remaining sampling steps with a step size of 10 and a classifier-free guidance scale of 7." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 422, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 661 + ], + "type": "text", + "content": "Metrics We take into account 5 metrics from three aspects: image composition, text-image alignment and image quality. 1) Image Composition: The defective image composition is mainly manifested as artifacts and incorrect interactions. Specifically, GroundingDINO [21] and HQSAM [16] are used to detect all the masks that have the same semantics as the foreground objects. The size ratio between the overflow area and the ground-truth foreground is reported as the first metric, which is called SAM Score (SAM). The quality of the interaction is mainly measured by a user study (INT). Thirty individuals participate in the survey and assess the generated images regarding the object interaction. The scale ranges from 1 to 5, with a higher score indicating better results. 2) Text-Image Alignment: CLIP Similarity (CLIP) is reported to reflect the text-image alignment between the text prompt and the generated image. 3) Image Quality: Aesthetic Score (AS) [28] and Image Reward (IR) [41] are used to evaluate the image quality. These evaluation models are trained on human preference datasets and are considered to align well with the human perception." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 671, + 244, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 244, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 244, + 685 + ], + "type": "text", + "content": "4.3. Comparison with Existing Methods" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "We conduct comparisons with six representative related methods, including Blended Latent Diffusion (BLD) [2]," + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 220 + ], + "type": "table", + "html": "
MethodSAM↓INT↑CLIP↑IR↑AS↑
BLD [2]0.9861.7526.060.185.04
SD-Inpaint.[26]0.3173.3526.260.355.19
PowerPaint [46]0.8202.2326.130.345.05
BrushNet [14]0.1904.0626.620.605.31
Layer Diff. [43]0.0463.5026.150.225.27
ControlNet† [44]0.2353.8826.460.525.27
PowerPaint† [46]0.4333.2726.150.465.20
BrushNet† [14]0.1203.9426.560.585.29
Ours0.0704.6226.500.675.34
", + "image_path": "2cf73e082c563d1b6d4669375ff5f4a6a5f206ab930c8a5103211dc077bbd11f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 228, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 262 + ], + "type": "text", + "content": "Table 1. The quantitative comparison of different methods. The items in bold and underline are ranked first and second. " + }, + { + "bbox": [ + 313, + 228, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 228, + 555, + 262 + ], + "type": "text", + "content": " means that the method is trained / finetuned with our dataset." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 282, + 553, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 553, + 366 + ], + "type": "text", + "content": "SD-Inpainting [26], PowerPoint [46], ControlNet [44], BrushNet [14] and Layer Diffusion [43]. Among them, all except BLD are learning-based methods. ControlNet is trained using the OpenImage-FCO dataset. For a fair comparison, we also fine-tuned BrushNet and PowerPaint using the proposed dataset. The intuitive and quantitative comparison results are shown in Fig. 3 and Tab. 1 respectively." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 367, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 554, + 569 + ], + "type": "text", + "content": "On the objective metric SAM, our method outperforms most comparison methods by at least 0.050. The only method that scores higher than our proposed method, Layer Diffusion, lacks competitiveness in other metrics due to poor coordination between the foreground and background. On the user-study metric INT, the average score given by users to our method was 0.56 higher than that of the second-best method. These results demonstrate the effectiveness of our method in improving image composition. The samples in Fig. 3 also intuitively illustrate this advantage. Among all compared methods, our approach produces artifact-free results across four cases. In terms of text-image alignment, our method achieves the second-best result. The first place is taken by BrushNet, partially due to its dataset. BrushNet was trained on LAION-Aesthetic [28] containing 120 million images, where abundant image-text pairs ensure advanced performance in text prompt controllability." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": "The advantages in image composition also drive better performance in image quality metrics, with our method outperforming the second-best by 0.07 and 0.03 on IR and AS respectively. Notably, while suppressing artifacts and incorrect interactions, our method produces less cluttered backgrounds compared to others. This can be attributed to the distribution discrepancy between " + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "inline_equation", + "content": "x_{t^s}" + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": " inferred by IPM and natural sampling. Fortunately, this characteristic also enhances visual hierarchy in synthesized images, making them more suitable for object presentation." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "A noteworthy result is that after fine-tuning on the proposed dataset, PowerPaint and BrushNet have shown signif" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16943" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 294, + 204 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 294, + 204 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 294, + 204 + ], + "type": "image", + "image_path": "d501d868eaf0710e5d86d6cb2364076ddb099898e1e5f3d73a809f9797c70f4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 207, + 294, + 339 + ], + "blocks": [ + { + "bbox": [ + 55, + 207, + 294, + 339 + ], + "lines": [ + { + "bbox": [ + 55, + 207, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 294, + 339 + ], + "type": "image", + "image_path": "d13bbb518902ad94acab5b848d5c49d70cd10926e99d2b1eabf4363e164228b1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 349, + 295, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 349, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 295, + 381 + ], + "type": "text", + "content": "Figure 5. (a) The generation results of the baseline model with or w/o IPM under different text prompts. (b) The multi generation results of the baseline model with or w/o IPM in same case." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 393, + 294, + 517 + ], + "blocks": [ + { + "bbox": [ + 57, + 393, + 294, + 517 + ], + "lines": [ + { + "bbox": [ + 57, + 393, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 57, + 393, + 294, + 517 + ], + "type": "image", + "image_path": "ed545122eb3927a872cf965dd86202118b181f8545af126dc6e84d257c3a25b4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 524, + 295, + 547 + ], + "lines": [ + { + "bbox": [ + 55, + 524, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 295, + 547 + ], + "type": "text", + "content": "Figure 6. The samples of predicted low-frequency signals and final generated images with or w/o the diffusion reward loss " + }, + { + "bbox": [ + 55, + 524, + 295, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rwd}" + }, + { + "bbox": [ + 55, + 524, + 295, + 547 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 568, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 664 + ], + "type": "text", + "content": "icant improvements in the SAM metric, which once again confirms the value of task-specific datasets. To further demonstrate the IPM's capabilities, we select foregrounds with diverse identities, scales, and positions as condition images and generate results under the same text prompt. The results in Fig. 4 show that our method adapts well to various scenarios and produces reasonable interactions between given objects and backdrops." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 282, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 282, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 282, + 683 + ], + "type": "text", + "content": "4.4. The Influence of IPM on Generation Model" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Given that the integration of IPM fundamentally alters the original diffusion process, its implications for the under" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 335, + 70, + 535, + 149 + ], + "blocks": [ + { + "bbox": [ + 335, + 70, + 535, + 149 + ], + "lines": [ + { + "bbox": [ + 335, + 70, + 535, + 149 + ], + "spans": [ + { + "bbox": [ + 335, + 70, + 535, + 149 + ], + "type": "table", + "html": "
SettingSAM↓CLIP↑IR↑AS↑
Baseline0.23526.460.525.27
w/o L_rwd0.10326.440.645.26
w/o U-Net Init.0.07426.290.635.28
Ours0.07026.500.675.34
", + "image_path": "3c1ceb69f2402a0b7b040973a4455dd77003ce33cd7cf52b342a2335a1e5b4c1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 323, + 157, + 544, + 168 + ], + "lines": [ + { + "bbox": [ + 323, + 157, + 544, + 168 + ], + "spans": [ + { + "bbox": [ + 323, + 157, + 544, + 168 + ], + "type": "text", + "content": "Table 2. The quantitative evaluation results of ablation study." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 192, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 239 + ], + "type": "text", + "content": "lying generation model warrant detailed discussion. For image generation tasks, the primary concerns are semantic controllability and output diversity, both of which will be systematically addressed in this section." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 240, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 445 + ], + "type": "text", + "content": "The generation results of the baseline model (ControlNet) with and without IPM under different text prompts " + }, + { + "bbox": [ + 313, + 240, + 555, + 445 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 240, + 555, + 445 + ], + "type": "text", + "content": " are presented in Fig. 5(a). It can be observed that IPM does not significantly affect the model's controllability, as the model still synthesizes results with the required semantics. This finding is also supported by Tab. 1, where our method and ControlNet exhibit the closest CLIP scores among all approaches. The subsequent ablation studies will reveal that this benefit primarily stems from initializing IPM parameters with the pretrained U-Net. Turning to Fig. 5(b), the generation results of the baseline model with and without IPM for the same pair of " + }, + { + "bbox": [ + 313, + 240, + 555, + 445 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 240, + 555, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 240, + 555, + 445 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 240, + 555, + 445 + ], + "type": "text", + "content": " are shown. In this comparison, IPM results show more similar composition due to the shared low-frequency signals, while the baseline model's diverse outputs come at the cost of prevalent compositional defects. This reinforces that IPM trades inefficient diversity for compositional rationality." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 454, + 408, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 408, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 408, + 467 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 473, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 556 + ], + "type": "text", + "content": "In this section, we conduct ablation studies on critical components of the IPM training paradigm and network design, including the diffusion reward supervision " + }, + { + "bbox": [ + 313, + 473, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rwd}" + }, + { + "bbox": [ + 313, + 473, + 554, + 556 + ], + "type": "text", + "content": " and the strategy of initializing IPM parameters with a pretrained U-Net. Note that since removing " + }, + { + "bbox": [ + 313, + 473, + 554, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{lbl}" + }, + { + "bbox": [ + 313, + 473, + 554, + 556 + ], + "type": "text", + "content": " would destabilize training, we exclude it from ablation analysis. The experimental results are presented in Tab. 2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Ablation Study for " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rwd}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " Quantitative validation shows that diffusion reward supervision provides the most significant benefit in artifact suppression. When " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rwd}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": " is removed, the SAM score increases by 0.033. The introduction of new artifacts also degrades image quality metrics, with AS dropping by 0.08. These artifacts correlate with the quality of low-frequency signals predicted by IPM. To visually corroborate this, we visualize these low-frequency signals (decoded via the Latent VAE) in Fig. 6. It can be observed that without " + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rwd}" + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": ", IPM outputs smoother signals. A direct consequence is that edges of some foreground objects become indistinct, which may develop into artifacts and imprecise details after subsequent denoising steps. Conversely, when" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16944" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 294, + 139 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 294, + 139 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 294, + 139 + ], + "type": "image", + "image_path": "ff068861785720bf7260705db46e0f920b9a7275b06f459da6d08f8e2fd394a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 140, + 225, + 148 + ], + "lines": [ + { + "bbox": [ + 61, + 140, + 225, + 148 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 225, + 148 + ], + "type": "text", + "content": "- A brand-new football is placed on the grass in the football field." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 178, + 293, + 301 + ], + "blocks": [ + { + "bbox": [ + 62, + 156, + 287, + 167 + ], + "lines": [ + { + "bbox": [ + 62, + 156, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 62, + 156, + 287, + 167 + ], + "type": "text", + "content": "Figure 7. The samples of generated images under different " + }, + { + "bbox": [ + 62, + 156, + 287, + 167 + ], + "type": "inline_equation", + "content": "t^s" + }, + { + "bbox": [ + 62, + 156, + 287, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 178, + 293, + 301 + ], + "lines": [ + { + "bbox": [ + 58, + 178, + 293, + 301 + ], + "spans": [ + { + "bbox": [ + 58, + 178, + 293, + 301 + ], + "type": "image", + "image_path": "66ec616b3cd2c81fda55ea260feb5eab73f4cd2c26a87f52bac0465a526e7b72.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 311, + 295, + 333 + ], + "lines": [ + { + "bbox": [ + 55, + 311, + 295, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 295, + 333 + ], + "type": "text", + "content": "Figure 8. The quantitative evaluation results of the truncating timestep selection." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 389 + ], + "type": "text", + "content": "diffusion reward supervision is introduced, IPM is encouraged to output intermediate states that genuinely lead to target images, effectively mitigating this issue." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 390, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 390, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 295, + 510 + ], + "type": "text", + "content": "Ablation Study for U-Net Initialization As shown in Tab. 2, initializing IPM parameters with a pretrained U-Net primarily strengthens the model's text controllability. Removing this strategy causes the CLIP score to drop by 0.21, even below the baseline performance. Although IPM is only responsible for predicting low-frequency signals, training it from scratch limits its performance on complex text prompts due to the dataset's relatively limited corpus. This phenomenon substantiates the necessity of leveraging the rich text-image priors of pretrained models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 254, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 254, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 254, + 529 + ], + "type": "text", + "content": "4.6. The Selection of Truncating Timestep" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "content": "The primary hyperparameter when applying IPM is determining the appropriate timestep " + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "inline_equation", + "content": "t^s" + }, + { + "bbox": [ + 55, + 534, + 295, + 594 + ], + "type": "text", + "content": " to truncate the diffusion process. We test values within the set {900, 850, 800, 750, 700}. The intuitive and quantitative comparison results are shown in Fig. 7 and Fig. 8." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "When the truncation timestep " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "t^s" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " is set too late, the method's effectiveness declines significantly. In the " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "t^s = 700" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " experiment group, although artifacts are further suppressed, this comes at the cost of overly simplified backgrounds and significantly reduced image quality. Its IR and AS metrics are 0.35 and 0.60 lower than those of " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "t^s = 800" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": ". This phenomenon is associated with the function of IPM. IPM directly predicts the image's low-frequency signal. When " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "t^s" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " is set to a smaller value, IPM's task becomes more complex. Meanwhile, the number of subsequent normal de" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 553, + 148 + ], + "type": "image", + "image_path": "3c401878804ca700c5644dd8c35a527b499d87851291814286201d76878db880.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 156, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 553, + 178 + ], + "type": "text", + "content": "Figure 9. The samples of failure case about the objects with extreme colors or lighting conditions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 201, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 321 + ], + "type": "text", + "content": "noising steps is reduced, requiring IPM to predict image signals across broader frequency bands. " + }, + { + "bbox": [ + 313, + 201, + 555, + 321 + ], + "type": "inline_equation", + "content": "t^s = 700" + }, + { + "bbox": [ + 313, + 201, + 555, + 321 + ], + "type": "text", + "content": " is clearly an excessively challenging value. As illustrated in Fig. 7, the produced image exhibits obvious detail deficiencies due to the interplay of these factors. When " + }, + { + "bbox": [ + 313, + 201, + 555, + 321 + ], + "type": "inline_equation", + "content": "t^s = 900" + }, + { + "bbox": [ + 313, + 201, + 555, + 321 + ], + "type": "text", + "content": ", the frequency band managed by IPM is too narrow, weakening its influence on final image. Its SAM score increases by 0.033 compared to " + }, + { + "bbox": [ + 313, + 201, + 555, + 321 + ], + "type": "inline_equation", + "content": "t^s = 800" + }, + { + "bbox": [ + 313, + 201, + 555, + 321 + ], + "type": "text", + "content": ", indicating that subsequent denoising steps introduce more suboptimal compositions. In this paper, " + }, + { + "bbox": [ + 313, + 201, + 555, + 321 + ], + "type": "inline_equation", + "content": "t^s = 800" + }, + { + "bbox": [ + 313, + 201, + 555, + 321 + ], + "type": "text", + "content": " is adopted across all experiments. g" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 328, + 477, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 477, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 477, + 340 + ], + "type": "text", + "content": "4.7. Failure case and Future Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 346, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 514 + ], + "type": "text", + "content": "During testing, a notable failure case arises when the given foreground object exhibits extreme colors or lighting conditions, significantly degrading background controllability. Fig. 9 illustrates results for such cases. Despite prompts' requirements, synthesized images still adopt styles consistent with the provided foreground (case 1) or fail to include other backdrops (case 2). Root causes include dataset limitations (artistic photography underrepresented in OpenImage) and framework design. That is, features of objects with extreme colors tend to dominate background features during synthesis, compelling background features to always reflect similar appearances. This motivates exploring methods to decouple foreground-background color distributions and enable foreground relighting in FCO tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 546, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 712 + ], + "type": "text", + "content": "This paper focuses on the problem of flawed image composition in diffusion-based FCO methods. We believe that this problem can be partially attributed to the random impact of the initial noise. Inspired by this, we propose the Initialization Policy Model (IPM). It uses a deterministic model that only depends on the given foreground and text prompt to replace the early denoising steps, thus isolating the impact of the initial noise. The subsequently proposed comprehensive training paradigm ensures that the IPM can be effectively optimized. Experiments confirm that the proposed method can significantly improve the composition problems such as artifacts and incorrect interactions, and the comprehensive framework built upon IPM also achieves advanced performance in FCO task." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16945" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant No. 62473290, 62088101, National Key Research and Development Program of China under Grant No. 2024YFB3311801, and the Shanghai Municipal Commission of Economy and Informatization under Grant No. 2024-GZL-RGZN-01008." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 165, + 115, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 115, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 115, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 185, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 185, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 185, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 185, + 294, + 217 + ], + "type": "text", + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In CVPR, pages 18187-18197, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 219, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 219, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 219, + 294, + 250 + ], + "type": "text", + "content": "[2] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM TOG, 42(4):149:1-149:11, 2023. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 251, + 294, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 294, + 283 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 294, + 283 + ], + "type": "text", + "content": "[3] Rodrigo Benenson and Vittorio Ferrari. From colouring-in to pointillism: revisiting semantic segmentation supervision. CoRR, abs/2210.14142, 2022. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 294, + 317 + ], + "type": "text", + "content": "[4] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. Instructpix2pix: Learning to follow image editing instructions. In CVPR, pages 18392-18402, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 318, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 294, + 361 + ], + "type": "text", + "content": "[5] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In CVPR, pages 22503-22513, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 362, + 294, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 294, + 405 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 294, + 405 + ], + "type": "text", + "content": "[6] Shidong Cao, Wenhao Chai, Shengyu Hao, Yanting Zhang, Hangyue Chen, and Gaoang Wang. Diffashion: Reference-based fashion design with structure-aware transfer by diffusion models. TMM, 26:3962-3975, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 406, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 294, + 460 + ], + "type": "text", + "content": "[7] Tingfeng Cao, Junsheng Kong, Xue Zhao, Wenqing Yao, Junwei Ding, Jinhui Zhu, and Jiandong Zhang. Product2img: Prompt-free e-commerce product background generation with diffusion model and self-improved lmm. In ACMMM, pages 10774-10783, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 460, + 294, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 294, + 515 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 294, + 515 + ], + "type": "text", + "content": "[8] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In ICLR, pages 1–31, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 516, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 516, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 516, + 294, + 548 + ], + "type": "text", + "content": "[9] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, pages 6593-6602, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 548, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 294, + 592 + ], + "type": "text", + "content": "[10] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based semantic image editing with mask guidance. CoRR, abs/2210.11427, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 593, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 593, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 593, + 294, + 647 + ], + "type": "text", + "content": "[11] Amir Erfan Eshratifar, Joao VB Soares, Kapil Thadani, Shaunak Mishra, Mikhail Kuznetsov, Yueh-Ning Ku, and Paloma De Juan. Salient object-aware background generation using text-guided diffusion models. In CVPR, pages 7489-7499, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "type": "text", + "content": "[12] Xuan Han, Yihao Zhao, and Mingyu You. Scene diffusion: Text-driven scene image synthesis conditioning on a single 3d model. In ACMMM, pages 7862-7870, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[13] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[14] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. CoRR, abs/2403.06976, 2024. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 152 + ], + "type": "text", + "content": "[15] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. Pnp inversion: Boosting diffusion-based editing with 3 lines of code. In ICLR, pages 1-28, 2024. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 553, + 185 + ], + "type": "text", + "content": "[16] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Segment anything in high quality. In NeurIPS, pages 29914–29934, 2024. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 251 + ], + "type": "text", + "content": "[17] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew P. Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, pages 105-114, 2017. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "text", + "content": "[18] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730-19742, 2023. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 299, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 341 + ], + "type": "text", + "content": "[19] Pengzhi Li, Qinxuan Huang, Yikang Ding, and Zhiheng Li. Layerdiffusion: Layered controlled image editing with diffusion models. In ACM SIGGRAPH Asia, pages 1-4, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[20] Shanglin Li, Bohan Zeng, Yutang Feng, Sicheng Gao, Xiuhui Liu, Jiaming Liu, Lin Li, Xu Tang, Yao Hu, and Jianzhuang Liu. Zone: Zero-shot instruction-guided local editing. In CVPR, pages 6254-6263, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[21] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, and Lei Zhang. Grounding DINO: marrying DINO with grounded pre-training for open-set object detection. CoRR, abs/2303.05499, 2023. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 445, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 466 + ], + "type": "text", + "content": "[22] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, pages 1-11, 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "text", + "content": "[23] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 513, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 578 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 579, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 622 + ], + "type": "text", + "content": "[25] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, pages 8821-8831, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "type": "text", + "content": "[26] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10674-10685, 2022. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[27] Dvir Samuel, Rami Ben-Ari, Simon Raviv, Nir Darshan, and Gal Chechik. Generating images of rare concepts using pretrained diffusion models. In AAAI, pages 4695-4703, 2024. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16946" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 160 + ], + "type": "text", + "content": "[28] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In NeurIPS, pages 1–21, 2022. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[29] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In CVPR, pages 8871-8879, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 207, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 294, + 261 + ], + "type": "text", + "content": "[30] Yujun Shi, Chuhui Xue, Jun Hao Liew, Jiachun Pan, Hanshu Yan, Wenqing Zhang, Vincent Y. F. Tan, and Song Bai. Dragdiffusion: Harnessing diffusion models for interactive point-based image editing. In CVPR, pages 8839-8849, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 296 + ], + "type": "text", + "content": "[31] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, pages 1-12, 2021. 3, 4, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "text", + "content": "[32] Yizhi Song, Zhifei Zhang, Zhe Lin, Scott Cohen, Brian Price, Jianming Zhang, Soo Ye Kim, and Daniel Aliaga. Object-stitch: Object compositing with diffusion model. In CVPR, pages 18310-18319, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 376 + ], + "type": "text", + "content": "[33] Zhengwentai Sun, Yanghong Zhou, Honghong He, and PY Mok. Sgdiff: A style guided diffusion model for fashion synthesis. In ACMMM, pages 8433-8442, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "type": "text", + "content": "[34] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, and Faisal Azhar. Llama: Open and efficient foundation language models. CoRR, abs/2302.13971, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "text", + "content": "[35] Ruoyu Wang, Yongqi Yang, Zhihao Qian, Ye Zhu, and Yu Wu. Diffusion in diffusion: Cyclic one-way diffusion for text-vision-conditioned generation. In ICLR, pages 1-21, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 479, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 554 + ], + "type": "text", + "content": "[36] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, pages 18359-18369, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 556, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 632 + ], + "type": "text", + "content": "[37] Su Wang, Chitwan Sahara, Ceslee Montgomery, Jordi Pont-Tuset, Shai Noy, Stefano Pellegrini, Yasumasa Onoe, Sarah Laszlo, David J. Fleet, Radu Soricut, Jason Baldridge, Mohammad Norouzi, Peter Anderson, and William Chan. Image editor and editbench: Advancing and evaluating text-guided image inpainting. In CVPR, pages 18359-18369, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "type": "text", + "content": "[38] Chen Henry Wu and Fernando De la Torre. A latent space of stochastic diffusion models for zero-shot image editing and guidance. In ICCV, pages 7378-7387, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[39] Xiaoshi Wu, Yiming Hao, Manyuan Zhang, Keqiang Sun, Zhaoyang Huang, Guanglu Song, Yu Liu, and Hongsheng Li. Deep reward supervisions for tuning text-to-image diffusion models. CoRR, abs/2405.00760, 2024. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 374 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 115 + ], + "type": "text", + "content": "[40] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. In CVPR, pages 22428-22437, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[41] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. In NeurIPS, pages 1-14, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "type": "text", + "content": "[42] Shiyuan Yang, Xiaodong Chen, and Jing Liao. Uni-paint: A unified framework for multimodal image inpainting with pretrained diffusion model. In ACM MM, page 3190-3199, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 207, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 240 + ], + "type": "text", + "content": "[43] Lvmin Zhang and Maneesh Agrawala. Transparent image layer diffusion using latent transparency. ACM Transactions on Graphics (TOG), 43(4):1-15, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[44] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3813-3824, 2023. 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 330 + ], + "type": "text", + "content": "[45] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, Yandong Guo, and Lei Zhang. Recognize anything: A strong image tagging model. In CVPR, pages 1724-1732, 2024. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[46] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting. CoRR, abs/2312.03594, 2023. 2, 5, 6" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16947" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/0cab628c-59eb-4cba-b70b-2c418218511f_content_list.json b/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/0cab628c-59eb-4cba-b70b-2c418218511f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..822f9623edb6c5addde47a9ad4608dd90f5fc032 --- /dev/null +++ b/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/0cab628c-59eb-4cba-b70b-2c418218511f_content_list.json @@ -0,0 +1,1528 @@ +[ + { + "type": "image", + "img_path": "images/166718c4d156e9768c5b6b9751f77486acf30a91676fed38e1ffe982abe83859.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 111, + 126, + 158, + 157 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Toward Fair and Accurate Cross-Domain Medical Image Segmentation: A VLM-Driven Active Domain Adaptation Paradigm", + "text_level": 1, + "bbox": [ + 161, + 135, + 885, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongqiu Wang $^{1}$ Wu Chen $^{2}$ Xiangde Luo $^{3}$ Zhaohu Xing $^{1}$ Lihao Liu $^{4}$ Jing Qin $^{5}$ Shaozhi Wu $^{2*}$ Lei Zhu $^{1,6*}$", + "bbox": [ + 135, + 207, + 808, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Hong Kong University of Science and Technology (Guangzhou) $^{2}$ University of Electronic Science and Technology of China $^{3}$ West China Hospital $^{4}$ Amazon $^{5}$ Hong Kong Polytechnic University $^{6}$ Hong Kong University of Science and Technology", + "bbox": [ + 101, + 244, + 859, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fairness in AI-assisted medical image analysis is crucial for equitable healthcare, but is often neglected, especially in prevalent cross-domain scenarios (diverse demographics and imaging protocols). Effective and equitable deployment of AI models in these scenarios is critical, yet traditional Unsupervised Domain Adaptation (UDA) methods exhibit limited improvements. Emerging Active Domain Adaptation (ADA) approaches offer more effective enhancements, but all ignore fairness issues. Therefore, in this work, we propose the first fairness-aware ADA paradigm that simultaneously achieves both enhanced fairness and superior overall performance. Our method leverages the multimodal alignment capability of Vision-Language Models (VLMs): By performing medical images (vision) and sensitive attributes (language) learning, VLM inherently captures semantic correlations between visual features and protected attributes, enabling explicit attributes representation. Building on this foundation, we further devise an attribute-aware strategy (FairAP), which dynamically adapts to diverse patient demographics to promote equitable and high-quality outcomes by considering both Attribute and Polysemy. Extensive experiments on the FairDomain benchmark demonstrate that our method significantly reduces bias and maintains leading performance, outperforming existing UDA and ADA methods. This work pioneers a VLM-driven ADA paradigm for fair cross-domain medical segmentation, offering a blueprint for effective and equitable AI deployment in clinical practice. Code will be on (https://github.com/whq-xxh/Fair-AP).", + "bbox": [ + 86, + 340, + 483, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 821, + 220, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"Fairness is not an attitude. It's a professional skill that must be developed and exercised.\" — Brit Hume.", + "bbox": [ + 89, + 847, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3bf325cc1380e29fe850a634f2f1ebad8c401afb6b0a29e792f318a49298ac2.jpg", + "image_caption": [ + "Figure 1. Comparison of previous cross-domain fairness methods with our VLM-Driven fairness-aware ADA paradigm. Our method combines VLM and ADA to achieve both leading model performance and enhanced fairness." + ], + "image_footnote": [], + "bbox": [ + 517, + 308, + 906, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The advent of AI technologies has significantly transformed medical image analysis (MedIA), providing sophisticated tools for analyzing complex biological patterns within imaging data [37, 46, 49-51]. These advanced algorithms enhance diagnostic precision, support informed clinical decisions, and improve patient outcomes. However, as these models become increasingly integrated into clinical practice, critical challenges have emerged: inherent biases within these models [24, 32, 57] and domain shifts caused by diverse clinical scenarios [23, 41, 54].", + "bbox": [ + 509, + 537, + 906, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Numerous studies have reported that biases in MedIA tasks [13, 16, 19, 29, 31, 33, 35, 39, 60], stemming from training datasets failing to represent diverse populations adequately, raise profound ethical and fairness concerns [10, 34]. These biases cause severe health inequities, impacting marginalized groups in clinical applications. For example, Black individuals face nearly twice the risk of glaucoma than others, highlighting the need for accurately segmenting the optic disc and cup for early diagnosis, yet AI models often perform worst in this subgroup. This issue is particularly critical in cross-domain scenarios, where models deployed across institutions encounter heterogeneous data distributions that amplify performance disparities [36, 40, 41]. Such domain-specific biases not only undermine the per", + "bbox": [ + 511, + 689, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*corresponding authors.", + "bbox": [ + 114, + 888, + 228, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24102", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "formance of AI systems but also violate fairness. Thus, ensuring both efficient and fair adaptation in cross-domain real-world medical AI environments is crucial to translate advanced technology to equitable healthcare for all.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, significant efforts have been made to tackle fairness across various fields [1, 3, 4, 9, 11, 15, 20, 24, 29, 38-40, 58, 62, 64, 65]. These studies explore various attributes across different modalities or body parts. However, real-world cross-domain challenges remain underexplored. While Tian et al. [41] established the multiattribute benchmark dataset to evaluate cross-domain fairness, the limited performance of its Unsupervised Domain Adaptation (UDA) framework (e.g., IoU values $< 40\\%$ in Tables 1 & 3) severely restricts clinical applicability despite its fairness-aware design. Although emerging Active Domain Adaptation (ADA) has demonstrated significant performance gains over UDA by actively selecting valuable target-domain samples [45], its fairness implications remain largely unexplored. This creates a critical gap in achieving both equitable outcomes and clinically viable performance across demographic subgroups, as current methods force an undesirable trade-off between the fairness guarantees and the model effectiveness.", + "bbox": [ + 91, + 155, + 483, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, in this work, we propose the first fairness-aware ADA paradigm to simultaneously enhance fairness across all subgroups and boost overall performance for cross-domain medical image segmentation scenario applications. Specifically, according to the unique characteristics of medical images and corresponding sensitive attribute information (such as race: Black, White, Asian), we present a novel perspective that integrates the Vision-Language Model (VLM) to align image features with sensitive attributes. This alignment endows our model with the capability to perceive the sensitive attribute representations within images accurately. This VLM-attribute learning is essential as it establishes a solid foundation for subsequent representative attributes and polysemy sample selection. Then, we devise an attribute-aware FairAP sample selection strategy, where we select a quite small (only $5\\%$ here) but critical set of samples and annotate them for target fair model fine-tuning. The FairAP strategy allocates equal quotas to each subgroup to ensure balanced attention and leverages VLM attribute-latent space representations for attribute-representative and polysemic sample selection. Our process and its key distinctions from previous methods are visually demonstrated in Fig. 1. It is worth noting that the FairAP strategy is highly practical as it significantly reduces the workload of doctors, while simultaneously achieving both fairness and high performance (even improving IoU by around $30\\%$ than FairDA method [41] for rim segmentation as shown in Table 1). Comprehensive experiments on the FairDomain benchmark [41] confirm that our approach markedly diminishes bias and retains cutting", + "bbox": [ + 91, + 446, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "edge performance in segmentation tasks, surpassing traditional DA and ADA methods.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 532, + 122, + 893, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To the best of our knowledge, we are the first to tackle fairness challenges in cross-domain medical image segmentation with ADA paradigm, dynamically adapting to diverse patient subgroups for equitable outcomes.", + "- We propose to leverage the Vision-Language Model to seamlessly integrate demographic attributes like race or gender into the learning process, endowing the model with the ability to perceive these sensitive attributes.", + "- We devise the attribute-aware FairAP strategy to ensure both fairness and effectiveness. FairAP optimizes representation by balancing samples from underrepresented demographic groups while maintaining diversity to enhance overall performance comprehensively.", + "- Extensive experiments demonstrate that our method surpasses previous DA and ADA approaches by ensuring superior fairness across multiple protected attributes and maintaining top performance. Our method pioneers a novel solution paradigm for fairness and effectiveness in cross-domain medical AI model deployment." + ], + "bbox": [ + 513, + 138, + 903, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 440, + 663, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Fairness in Medical Imaging", + "text_level": 1, + "bbox": [ + 511, + 465, + 767, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In medical image processing, fairness aims to reduce model biases to ensure equitable outcomes across patient subgroups. Existing methods fall into three main categories: pre-processing, in-processing, and post-processing [6, 26, 57]. Pre-processing techniques, such as those proposed by Bissoto et al. [5] and others [24, 29, 59], aim to correct training data imbalances, but may affect computational efficiency. In-processing methods mitigate bias during training by incorporating fairness-aware loss functions [40] or fairness-specific modules [20, 25, 41], which can potentially trade off some accuracy. On the other hand, post-processing approaches, like those introduced by Wang et al. [53], Kim et al. [17], and others [22], adjust the model outputs to improve fairness, which may struggle to generalize well during testing. However, few studies focus on fairness in cross-domain scenarios within real-world deployment environments.", + "bbox": [ + 511, + 488, + 906, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Domain Adaptation", + "text_level": 1, + "bbox": [ + 511, + 757, + 702, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Domain Adaptation (UDA): UDA trains on the labeled source and unlabeled target data to bridge domain gaps, typically using adversarial training and minimizing inter-domain discrepancies [54]. In UDA methods, adversarial learning minimizes distributional differences between source and target domains, enhancing performance in the target domain [42, 44]. Additionally, self-training in UDA uses pseudo-labels from target domain data to en", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24103", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/08f6d3ad504d8ea729a41efa9bb0fef3fff8a084b9552ce8ed6366d6073d7e8e.jpg", + "image_caption": [ + "Figure 2. Overview of the Fair VLM-Driven ADA method (taking race attribute for illustration). Given the target domain images $x^T$ and their corresponding sensitive attribute information $\\text{attrs}$ , we first align them through VLM-attribute learning to obtain the vision-based sensitive attribute features $F_{attrR}$ . Then, we devise the attribute-aware FairAP Select strategy to select $\\alpha \\%$ of the samples that are representations of Attribute and Polysemy. These selected samples $S$ are then annotated by experts to obtain the annotations $y^{TL}$ . Finally, active domain adaptation fine-tuning is performed to obtain the target fair model $h_{\\phi}'$ for top performance and enhanced fairness." + ], + "image_footnote": [], + "bbox": [ + 106, + 87, + 883, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "hance performance in the target domain [12, 27, 66]. Building on the work by Hoyer et al. [12], Tian et al. [41] further explored fairness in cross-domain settings by integrating the fairness module. This new setting can more accurately reflect real-world clinical scenarios and provide a clear quantification of the performance and fairness of cross-domain methods [41]. However, very limited performance gains brought by UDA significantly diminish its clinical utility.", + "bbox": [ + 88, + 539, + 485, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Active Domain Adaptation (ADA): ADA combines active learning and domain adaptation to reduce annotation costs and sustain high performance by focusing on annotating the most valuable samples through strategies like uncertainty-based selection and diversity-based sampling [8, 14, 21, 48, 56]. For example, Wang et al. [45], and others [47, 52, 55, 61] enhanced cross-domain performance by strategically selecting target domain samples for model development. However, these current ADA methods have predominantly concentrated on optimizing overall model performance, neglecting the crucial aspect of fairness. Consequently, we are the pioneers in introducing a novel fairness-aware ADA paradigm, effectively minimizing annotation costs while simultaneously achieving both enhanced fairness and leading performance.", + "bbox": [ + 89, + 674, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 537, + 650, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setting", + "text_level": 1, + "bbox": [ + 511, + 568, + 674, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-domain MedIA often exhibits not only conventional performance degradation for domain shifts but also issues of unfairness related to sensitive attributes. Fair ADA is an effective and novel approach to address this challenge. Our work specifically targets fairness challenges in domain adaptation, focusing on fairness issues across multiple sensitive attributes in diverse subgroups. The problem is defined as follows: In medical segmentation tasks, consider the source domain data $D_{S} = \\{(x_{i}^{S},y_{i}^{S},\\text{attr}_{i})\\}_{i = 1}^{N_{S}}$ , where $x_{i}^{S}\\in \\mathbb{R}^{H\\times W}$ represents the $i_{th}$ image, the $y_{i}^{S}\\in \\mathbb{R}^{C\\times H\\times W}$ denotes the corresponding label with $C$ classes, and the $\\text{attr}_{i}\\in \\mathbb{R}^{n}$ means the sensitive attributes related to the sample, such as gender, race, or ethnicity. Meanwhile, there is an unlabeled dataset $D_{T} = \\{(x_{i}^{T},\\text{attr}_{i})\\}_{i = 1}^{N_{T}}$ whose domain vary, leading to different image features from $D_{S}$ . In the fair segmentation setting, note that the target attributes $\\text{attr}_{i}$ are available and accessible, as outlined in [40, 41]. The aim is to design strategies for selecting a small number of samples $S = \\alpha \\% \\cdot D_{T}\\ll D_{T}$ for labeling and active learning, thereby maximizing the fairness", + "bbox": [ + 511, + 595, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24104", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and performance of model $h_{\\phi}$ as much as possible while maintaining $\\alpha \\%$ low.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overview of the Fair-ADA", + "text_level": 1, + "bbox": [ + 89, + 128, + 328, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall structure of our proposed VLM-driven fairness-aware ADA framework is illustrated in Fig. 2. The entire methodology section uses race attribute $attrR$ as an example to illustrate. In the VLM-attribute Learning stage (a), we utilize a Vision-Language model $g_{\\theta}$ to align image samples with sensitive attribute information within the latent space. The model $g_{\\theta}$ performs vision (unlabeled medical images $x^{T}$ ) language (corresponding attribute information $attrR$ ) learning to achieve alignment. Then, we can extract vision-based sensitive attribute features $F_{attrR}$ . These feature representations consist of multiple subgroups corresponding to various attributes. For example, sensitive racial features $F_{attrR}$ include three subgroup features such as Asian features $F_{a}$ , Black features $F_{b}$ , and White features $F_{w}$ . The entire process can be expressed as follows:", + "bbox": [ + 89, + 150, + 483, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {a t t r R} = g _ {\\theta} \\left(x ^ {T}, a t t r R\\right) = \\left\\{F _ {a}, F _ {b}, F _ {w} \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 383, + 482, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the vision-based sensitive attribute features $F_{attrR}$ , the process moves into the FairAP Select stage (b), where our approach is organized into three main steps. In the first step, we ensure and amplify fairness within our active learning strategy by allocating an equal quota of samples from each sensitive attribute subcategory. In the second step, we employ clustering methods for each subgroup based on various sensitive attribute representations, followed by selecting attribute representative samples. In the third step, we further optimize the selected samples by expanding the selection to include polysemy representatives. This overall process can be summarized as:", + "bbox": [ + 89, + 407, + 483, + 589 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS = \\text {F a i r} A P \\left(F _ {\\text {a t t r} R}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 597, + 480, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the Active Domain Adaptation stage (c), we utilize the model $h_{\\phi}$ that has been previously trained on the source domain data $D_{S}$ . We then fine-tune this model using the sample set $S$ selected by our FairAP strategy to produce the optimized target fair model $h_{\\phi}^{\\prime}$ . The loss function for this model is defined by the following formula:", + "bbox": [ + 89, + 619, + 483, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {A D A} \\left(h _ {\\phi} ^ {\\prime} (x ^ {T L}), y ^ {T L}\\right) = - \\sum_ {c = 1} ^ {C} y _ {c} \\log \\left(h _ {\\phi} ^ {\\prime} (x)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 719, + 482, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C$ represents the number of classes, and selected list $S = \\{x^{TL},y^{TL}\\}$ , whose size corresponds to $\\alpha \\%$ of the total dataset.", + "bbox": [ + 89, + 765, + 483, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. VLM-attribute Learning", + "text_level": 1, + "bbox": [ + 89, + 816, + 318, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Traditional fairness-aware segmentation methods [40, 41] using the simplistic categorical encoding (e.g., 0/1 for gender male or female) suffer from semantic sparsity and contextual blindness, which fails to capture attribute-visual", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c646274ca138d0181a38b3873a00f629cd9bb42ab340cce4a970902c44a0c6d2.jpg", + "image_caption": [ + "VLM-attribute learning", + "Figure 3. Visualization of racial feature distribution before and after feature alignment via VLM-attribute learning. Our alignment process creates a denser distribution of features within racial subgroups, improving distinguishability across different races." + ], + "image_footnote": [], + "bbox": [ + 526, + 102, + 895, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "correlations. These disjoint representations limit ADA paradigms' ability to perceive subgroup distributions during cross-domain adaptation. To address this, we propose the VLM-based attribute-visual alignment to enable context-aware perception and provide continuous semantic gradients for precise subgroup-aware adaptation.", + "bbox": [ + 511, + 291, + 905, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 (a) provides a detailed illustration of our vision language learning process. Specifically, we utilize a pretrained vision-language model (CLIP [30, 63] here) denoted as $g_{\\theta}$ to process the samples $x^{T}$ along with their associated sensitive attribute information $\\text{attr } R$ . The model's vision encoder, $g_{\\theta_v}^v$ , extracts image tokens $z_{v}$ from the input $x^{T}$ with parameters $\\theta_{v}$ . Meanwhile, the language encoder, $g_{\\theta_l}^l$ , derives attribute tokens $z_{l}$ from the input $\\text{attr } R$ with parameters $\\theta_{l}$ . Therefore, we can express the model as $g_{\\theta} = \\{g_{\\theta_v}^v,g_\\theta_l^l\\}$ with the parameters as $\\theta = \\{\\theta_v,\\theta_l\\}$ . When training the model $g_{\\theta}$ using a batch of $N$ image-attribute token pairs, we generate a similarity matrix $A\\in \\mathbb{R}^{N\\times N}$ . Each entry of the matrix, $A_{ij} = z_{vi}^{T}z_{lj}$ , represents the cosine similarity between the $i_{th}$ image token $z_{vi}$ and $j_{th}$ attribute token $z_{lj}$ within the batch. The model $g_{\\theta}$ employs a contrastive learning approach that aims to maximize the cosine similarity for positive image-attribute token pairs while minimizing it for negative token pairs. The model $g_{\\theta}$ loss function can be expressed as follows:", + "bbox": [ + 511, + 382, + 906, + 671 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {V L M} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {1}{2} \\log \\left(\\frac {\\exp (2 A _ {i i})}{S _ {i} T _ {i}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 681, + 905, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S_{i} = \\sum_{j=1}^{N} \\exp(A_{ij})$ and $T_{i} = \\sum_{j=1}^{N} \\exp(A_{ji})$ . Since our focus is on enhancing the vision-based attribute perception capabilities, we choose to freeze the language encoder $g_{\\theta}^{l}$ to better train $g_{\\theta}$ as follows:", + "bbox": [ + 511, + 734, + 905, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA ^ {\\prime} \\leftarrow A,\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 809, + 661, + 825 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {v} ^ {\\prime} \\leftarrow \\theta_ {v} - \\eta \\frac {\\partial L _ {V L M}}{\\partial \\theta_ {v}}, \\quad \\theta_ {l} ^ {\\prime} = \\theta_ {l}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 828, + 903, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta$ is the learning rate. We employ a consistent template for sensitive attribute information as:", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24105", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A photo of a fundus and the race is $\\langle \\mathbf{RACE}\\rangle$", + "bbox": [ + 133, + 90, + 434, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The $\\langle \\mathbf{RACE}\\rangle$ represents specific racial information. After completing the training, we utilize a fixed vision encoder $g_{\\theta_v'}^v$ to process all target domain images $x^T$ , allowing extraction of the vision-based sensitive attribute features $F_{attrR}$ as following:", + "bbox": [ + 89, + 114, + 483, + 189 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {a t t r R} = g _ {\\theta_ {v} ^ {\\prime}} ^ {v} \\left(x ^ {T}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 189, + 480, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The visualization of racial characteristic distribution before and after our VLM-attribute learning is shown in Fig. 3.", + "bbox": [ + 89, + 210, + 483, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. FairAP Strategy", + "text_level": 1, + "bbox": [ + 89, + 250, + 253, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Building on the VLM-attribute learning described earlier, sensitive vision-based attribute features $F_{attrR}$ are successfully extracted. For clarity, we continue using sensitive racial features $F_{attrR}$ as an example to illustrate the Fair Attribute and Polysemy representation (FairAP) Select strategy. The sensitive racial features $F_{attrR}$ encompass three types of racial characteristics: Asian features $F_{a}$ , Black features $F_{b}$ , and White features $F_{w}$ . Thus, we define $F_{attrR} = \\{F_{a}, F_{b}, F_{w}\\}$ . Our FairAP Select strategy encompasses three critical components: Fair Quota Select, Attribute Representation Select, and Polysemy Representation Select. We will introduce each component in sequence.", + "bbox": [ + 89, + 272, + 483, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fair Quota Selection. We recognize the data-driven nature of AI. Therefore, to achieve our fairness objective ('treat all our subgroups with fairness'), we ensure a consistent sample quota for each racial subgroup. We define the sample quota for each subgroup as follows:", + "bbox": [ + 89, + 454, + 483, + 531 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {a} = \\Delta_ {b} = \\Delta_ {w} = \\Delta_ {R}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 540, + 480, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta_{a}$ , $\\Delta_{b}$ , and $\\Delta_{w}$ represent the quotas given to Asians, Blacks, and Whites respectively. The $\\Delta_{R}$ is calculated by dividing the overall quota $(\\alpha \\% \\cdot D_T)$ by the number of total racial subgroups, which is three here. Upon completing the evaluations of the subsequent components (Attribute and Polysemy), which assist in determining the specific samples, we compile the final selection list for each racial subgroup, represented as $S_{R} = \\{S_{\\hat{R}A}, S_{\\hat{R}P}\\}$ . Finally, we combine the selection lists $S_{R}$ of all racial subgroups and submitted them to doctors for annotations.", + "bbox": [ + 89, + 566, + 483, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attribute Representation Selection. To optimally select representative samples for each racial subgroup, we undertake clustering operation (k-means++ algorithm [2] here) to process the feature set $F_{R}$ , resulting in the clustered features $F_{R}^{\\prime}$ and the corresponding cluster centers $C_{R}^{N_{R}}$ , where $R$ represents a specific racial group (like Asian, Black, or White). This process is represented as follows:", + "bbox": [ + 89, + 718, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{F _ {R} ^ {\\prime}, C _ {R} ^ {N _ {R}} \\right\\} = \\operatorname {C l u s t e r} \\left(F _ {R}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 832, + 483, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this equation, $C_R^{N_R}$ denotes the cluster centers, and $N_R$ indicates the number of cluster centers for the racial sub", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: POLYSEMY-REP SELECT" + ], + "code_body": "Input: The $i_{th}$ cluster center of subgroup $R$ : $C_R^i$ , distance set: $D_R^i$ , selected distance set: $D_{\\hat{R}A}^i$ , features: $F_{R,m}''$ , number of $F_{R,m}': M_i$ , other racial subgroups' cluster centers: $C_{\\bar{R}}^k$ , quota: $\\Delta_R$ , subgroup cluster number: $N_R$ . Output: Polysemy-Rep selective racial list: $S_{\\hat{R}P}$ \n1 for $i = 1$ to $N_R$ do \n2 ${D_R'} \\gets {D_R} \\backslash {D_{\\hat{R}A}}^i // \\text{Remaining distances}$ \n3 $F_{R,m}'' \\gets \\text{Index} \\left( D_R^{i'} \\right) // \\text{Indexing features}$ \n4 $N_i \\gets M_i - \\frac{\\Delta_R}{2N_R} // \\text{Adjust selection range}$ \n5 Initialize empty list $E_{temp}$ ; \n6 for $m = 1$ to $N_i$ do \n7 if $F_{R,m}'' \\in C_R^i \\land F_{R,m}'' \\notin C_{\\bar{R}}^k$ then \n8 $\\begin{array}{c} e_i^m \\gets \\min \\left( \\| F_{R,m}' - C_{\\bar{R}}^k \\| \\right) // \\text{Min distance to other racial clusters} \\\\ \\text{Append } e_i^m \\text{ to } E_{temp}; \\end{array}$ \n9 end \n10 end \n11 Sort $E_{temp}$ in rising order; \n12 $E_{\\hat{R}P}^i \\gets \\text{top min distances from } E_{temp}$ ; \n14 end \n15 $F_{\\hat{R}P}' \\gets \\text{Index} \\left( \\{E_{\\hat{R}P}', ..., E_{\\hat{R}P}', ..., E_{\\hat{R}P}^{N_R}\\} \\right)$ \n16 $S_{\\hat{R}P} \\gets F_{\\hat{R}P}' // \\text{Get the final Polysemy-Rep racial subgroup list}$", + "bbox": [ + 516, + 112, + 903, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "group. For each racial subgroup, the number of cluster centers $C_R^{N_R}$ is calculated based on their overall distribution in the dataset to comprehensively represent the target domain. This is illustrated by the following formula:", + "bbox": [ + 511, + 553, + 906, + 614 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nN _ {a}: N _ {b}: N _ {w} \\approx \\operatorname {S u m} _ {a}: \\operatorname {S u m} _ {b}: \\operatorname {S u m} _ {w}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 626, + 903, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Sum}_a$ , $\\mathrm{Sum}_b$ , and $\\mathrm{Sum}_w$ denote the proportions of Asians, Blacks, and Whites, respectively. For the $i_{th}$ cluster center $C_R^i$ of race $R$ , we define a set of Euclidean distances $D_R^i$ , where each element $d_i^m$ in the set is the distance from sample $F_{R,m}^{\\prime}$ ( $m_{th}$ sample belongs to the $i_{th}$ cluster) to the cluster center $C_R^i$ . These distances are then sorted in ascending order. This can be represented as:", + "bbox": [ + 511, + 652, + 906, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D _ {R} ^ {i} = \\operatorname {s o r t} \\left(\\left\\{d _ {i} ^ {m} = \\| F _ {R, m} ^ {\\prime} - C _ {R} ^ {i} \\| \\mid F _ {R, m} ^ {\\prime} \\in C _ {R} ^ {i} \\right\\}\\right) \\\\ = \\left\\{d _ {i} ^ {1}, d _ {i} ^ {2}, \\dots , d _ {i} ^ {m}, \\dots , d _ {i} ^ {M _ {i}} \\right\\}, \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 771, + 903, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d_i^1 \\leq d_i^2 \\leq \\dots \\leq d_i^m \\leq \\dots \\leq d_i^{M_i}$ , and $M_{i}$ is the number of samples $F_{R}^{\\prime}$ that belong to the cluster center $C_R^i$ . To ensure that the selected attribute representatives $F_{\\hat{R}A}$ can comprehensively cover the entire distribution of the dataset,", + "bbox": [ + 511, + 837, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24106", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we establish an equal selection quota of $\\frac{\\Delta_R}{2N_R}$ for each cluster. Specifically, we select the top $\\frac{\\Delta_R}{2N_R}$ samples, denoted as the subset $D_{\\hat{R}A}^i$ , corresponding to the smallest Euclidean distances from the sorted set $D_R^i$ . The set $D_{\\hat{R}A}$ represents the distances selected from the $N_R$ clusters of the current racial subgroup. This can be expressed as:", + "bbox": [ + 89, + 89, + 483, + 189 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD _ {R \\hat {A}} ^ {i} = \\left\\{d _ {R} ^ {1}, d _ {R} ^ {2}, \\dots , d _ {R} ^ {\\Delta_ {R} / 2 N _ {R}} \\right\\} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 198, + 482, + 224 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\hat {R A}} = \\left\\{D _ {\\hat {R A}} ^ {1}, D _ {\\hat {R A}} ^ {2}, \\dots , D _ {\\hat {R A}} ^ {N _ {R}} \\right\\}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 232, + 482, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using the indices from the set $D_{\\hat{R}A}$ that correspond to the smallest distances, we can derive the feature set $F_{\\hat{R}A}^{\\prime}$ for attribute representative selection and obtain the corresponding sample set $S_{\\hat{R}A}$ .", + "bbox": [ + 89, + 262, + 482, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Polysemy Representation Selection. Given that representative samples for each racial subgroup have been selected in the previous phase, we focus on choosing diverse samples in the Polysemy phase to explore a wider range of expressions within a limited quota, thus enhancing performance. For example, as seen in the right (post-VLM learning) of Fig. 3, some samples within the Black clusters are actually from White individuals, which may embody characteristics of both White and similar attributes of Black.", + "bbox": [ + 89, + 324, + 483, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In detail, We first filter out the samples selected in the previous stage, as indicated by $D_{R}^{i'} = D_{R}^{i} \\setminus D_{\\hat{R}A}^{i}$ . Subsequently, we calculate the nearest distance $e_i^m$ from the remaining samples belonging to the $i_{th}$ cluster center $C_R^i$ of race $R$ to the other races' cluster centers. These distances are then sorted in rising order to obtain a new Euclidean distance set $E_R^i$ . After this, we will follow Algorithm 1 to identify the final selected set of polysemy representative features, $F_{\\hat{R}P}'$ , and samples list, $S_{\\hat{R}P}$ . Through the aforementioned selection strategy, the chosen samples not only exhibit characteristics of the race to which they belong but also possess features similar to other races. This demonstrates the polysemy of the selected samples. As illustrated in Fig. 2 (b), we use Asian samples as an example to demonstrate our selection process.", + "bbox": [ + 89, + 460, + 483, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 702, + 223, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 727, + 281, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We use NVIDIA RTX 3090Ti GPUs with 24GB of memory for all experiments. We utilize the official code repository of the CLIP model [30]. We employ the pre-trained ViT-B/16 model with a batch size of 32 and use the Adam optimizer [18]. The initial learning rate is set to 1e-7, and the total number of iterations is 20,000. For segmentation model in the ADA process, following previous works [40, 41], we use the TransUNet [7] as the backbone and set the batch size to 2. The total number of iterations: 40,000 for training on the source domain and", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a6532aa695cdeba009e1aff484cb50abec73ad11a36104a210a20af7243db4d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 923, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9cc126768d95ab511ea64d42ebca6c2784b6dd866963ccfda36deab6701a9414.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 154, + 921, + 217 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fc0a8e1fd698224b24fb0ece4235844e4de92df43211d5bb8d182cc92538eace.jpg", + "image_caption": [ + "GT Baseline", + "STDR", + "Ours", + "Figure 4. Visual comparisons of our method and representative methods. Baseline: cross-domain results without fine-tuning; FIA [41]: fairness-aware DA approach; STDR [47]: ADA approach. Samples from top to bottom: Asian, Black, White subgroups, with numerical results on each image. Please zoom in for the best view." + ], + "image_footnote": [], + "bbox": [ + 517, + 217, + 921, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "20,000 for fine-tuning on the target domain. The selection percentage of $\\alpha \\%$ is set to $5\\%$ .", + "bbox": [ + 511, + 378, + 903, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark Dataset. We employ the cross-domain medical segmentation benchmark of the Harvard-FairDomain dataset [41], comprising 10,000 multi-modal samples with sensitive attributes (eg., gender, race, ethnicity). Its rich demographic attributes and paired modalities make it suitable for assessing and quantifying cross-domain model fairness.", + "bbox": [ + 511, + 409, + 905, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We quantify fairness in medical image segmentation using metrics from previous studies [40, 41]: Dice, IoU, and the key composite metrics, ES-Dice and ES-IoU, to assess fairness alongside performance. For more details, please refer to the supplementary materials.", + "bbox": [ + 511, + 501, + 905, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 511, + 589, + 805, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following common practices in prior ADA works [45, 47], we comprehensively compare our approach with established DA methods and the latest ADA techniques. Our analysis primarily focused on fairness and performance comparisons across three sensitive attributes: Race, Ethnicity, and Gender (presented in Table 1, 3 and supplementary materials). Visualization examples of our method's results on racial attributes, compared to other representative methods (Baseline, FIA [41] and STDR [47]), are shown in Fig. 4.", + "bbox": [ + 511, + 612, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With SOTA DA Methods. We perform a comprehensive comparison with previous state-of-the-art DA methods (including PixMatch [28], DAFormer [12], and DAFormerFIA [41]) across three demographic attributes. Tables 1, & 3 prove that our method significantly outperforms these approaches across nearly all metrics. Taking the Race attribute (in Table 1) as an example, FairAP significantly outperforms the DA methods for all metrics. This reflects a stable improvement compared to the previous state-of-the", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24107", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f0eaa3b3a342d688e188a5593b7b34b23d9d07197dab702c96ad8af50780fb95.jpg", + "table_caption": [ + "Table 1. Cup and rim segmentation performance on the FairDomain-Segmentation benchmark using different DA and ADA methods with Race as the demographic attribute. The * denotes $p$ -value $< 0.05$ in all paired $t$ -test, indicating statistically significant differences." + ], + "table_footnote": [], + "table_body": "
MethodVenueOverall ES-Dice↑Overall Dice↑Overall ES-IoU↑Overall IoU↑Asian Dice↑Black Dice↑White Dice↑Asian IoU↑Black IoU↑White IoU↑
DACupBaseline (Source)-0.8680.8880.7840.8080.8810.9010.8860.8020.8300.805
Baseline (Target)-0.6730.7000.5350.5640.7030.7310.6940.5670.6060.555
RimBaseline (Source)-0.8200.8610.7160.7620.8410.8380.8680.7370.7310.771
Baseline (Target)-0.4810.4950.3340.3420.4960.4720.4990.3470.3260.345
DACupPixMatch [28]CVPR'210.7390.7750.6140.6600.7520.7930.7830.6330.693
DAFormer [12]CVPR'220.7850.8040.6690.6910.7940.8160.8020.6820.7120.688
DAFormer-FIA [41]ECCV'240.7960.8100.6820.7000.8040.8210.8090.6940.7170.697
RimPixMatch [28]CVPR'210.6270.6730.4930.5230.6530.6400.6930.5220.4980.559
DAFormer [12]CVPR'220.3260.3320.1990.2020.3340.3450.3290.2040.2120.199
DAFormer-FIA [41]ECCV'240.5210.5310.3630.3690.5220.5390.5300.3640.3800.368
CupRandom-0.8130.8340.7160.7340.8140.8310.8370.7120.7350.737
Entropy [44]CVPR'190.8140.8230.7080.7210.8140.8250.8240.7110.7280.720
MHPL [45]CVPR'230.8090.8340.7080.7330.8100.8300.8380.7020.7330.737
DML-Core [43]NeurIPS'240.8110.8390.7170.7400.8120.8350.8420.7120.7420.743
Detective [61]CVPR'240.8140.8370.7170.7400.8140.8410.8390.7180.7480.741
STDR [47]IEEE TMI'240.8130.8370.7190.7380.8140.8330.8400.7150.7370.741
Our FairAP-0.828*0.843*0.731*0.747*0.827*0.8430.845*0.730*0.7510.748*
RimRandom-0.6870.7800.5670.6540.7330.7120.8000.6000.5770.676
Entropy [44]CVPR'190.6860.7720.5650.6430.7270.7090.7900.5950.5730.662
MHPL [45]CVPR'230.6880.7820.5670.6560.7290.7180.8010.5960.5810.678
DML-Core [43]NeurIPS'240.6970.7830.5780.6580.7380.7220.8010.6090.5890.677
Detective [61]CVPR'240.7000.7840.5820.6590.7400.7250.8010.6120.5930.678
STDR [47]IEEE TMI'240.6880.7810.5700.6560.7330.7130.8000.6040.5790.677
Our FairAP-0.711*0.791*0.592*0.667*0.750*0.735*0.807*0.624*0.602*0.686*
", + "bbox": [ + 125, + 117, + 870, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "art DA models. For the rim segmentation, previous DA methods bring very limited performance improvements, and some (DAFormer [12]) even decline. For instance, as indicated in Table 1, the Baseline ES-IoU and IoU scores are 0.334 and 0.342 respectively, whereas DAFormer-FIA [41] achieved only 0.363 and 0.369. In contrast, our approach significantly outperforms them, achieving 0.592 and 0.667, respectively, which is approximately $20\\%$ and $30\\%$ higher. Overall, the previous fairness-aware DA method achieves fairness but with limited accuracy gains, while our FairAP ensures fairness and substantially boosts performance, underscoring its superiority.", + "bbox": [ + 88, + 446, + 480, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With SOTA ADA Methods. We further perform extensive comparison of these attributes against the ADA methods: Random (randomly select samples), Entropy [44], MHPL [45], DML-Core [43], Detective [61], and STDR [47], with the same labeling quota for a fair comparison. The experimental results indicate that our approach surpasses existing ADA methods in segmentation performance and demonstrates significant improvements in fairness. For instance, when analyzing the Race attribute (in Table 1), our method's overall Dice score of 0.791 outperforms the STDR method's 0.781 by 1 percentage with statistically significant differences. However, when considering both performance and fairness, our ES-Dice score of 0.711 exceeds STDR's 0.688 by 2.3 percentages, demonstrating FairAP's ability to significantly enhance fairness while maintaining performance. These findings are also reflected in Table 1 and Table 3, which assess Race and Ethnicity attributes. Compared to other ADA methods, our approach not only excels in seg", + "bbox": [ + 91, + 628, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6db440a0f7045d5d83211b4b210f9c769fc2db1f0323a283f9782f0163f1db9c.jpg", + "table_caption": [ + "Table 2. Ablation studies for VLM-learning on sensitive attributes. The * denotes $p$ -value $< 0.05$ in all paired $t$ -test." + ], + "table_footnote": [], + "table_body": "
AttributesMethodCUPRim
Fair-APVLMES-Dice ↑ES-IoU ↑ES-Dice ↑ES-IoU ↑
Race✓0.8110.7110.6940.575
✓✓0.828*0.731*0.711*0.592*
Ethnicity✓0.8210.7190.7810.660
✓✓0.836*0.730*0.784*0.663*
", + "bbox": [ + 516, + 472, + 898, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mentation performance but also demonstrates effective improvements in fairness.", + "bbox": [ + 511, + 574, + 903, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 618, + 679, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study on VLM-Learning. We conducted ablation studies to evaluate the impact of VLM-learning on model performance. In Table 2, we compared Fair-AP outcomes with and without VLM-learning under two sensitive attributes. The results demonstrate that VLM-learning significantly improves segmentation performance ( $p$ -value $< 0.05$ ). This statistically significant enhancement validates the effectiveness of VLM-learning in representing demographic attributes through cross-modal vision-language alignment, thereby enabling equitable performance across diverse subgroups.", + "bbox": [ + 509, + 642, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study on Fair-AP. We conducted ablation experiments to assess the impact of Fair-AP's design, focusing on four specific configurations: 1) Fair-Base, which evenly allocates quotas across subgroups (like White, Black, and Asian) and selects samples randomly. 2) Fair-Attr, which selects samples based on Attribute representativeness. 3)", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24108", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/812d3b1322ab5f65dd7b8ce84be786a61a451438e1a923a7f0e09b1f23601b71.jpg", + "table_caption": [ + "Table 3. Cup and rim segmentation performance on the FairDomain-Segmentation benchmark using different DA and ADA methods with Ethnicity as the demographic attribute. The * denotes $p$ -value $< {0.05}$ in all paired $t$ -test,indicating statistically significant differences." + ], + "table_footnote": [], + "table_body": "
MethodVenueOverall ES-Dice↑Overall Dice↑Overall ES-IoU↑Overall IoU↑Hispanic Dice↑Non-Hispanic Dice↑Hispanic IoU↑Non-Hispanic IoU↑
DACupBaseline (Source)-0.8710.8880.7840.8080.8870.9060.8070.839
Baseline (Target)-0.6840.7000.5510.5550.6990.7220.5630.587
RimBaseline (Source)-0.8450.8610.7430.7620.8600.8790.7610.786
Baseline (Target)-0.4890.4950.3400.3420.4950.5060.3420.348
DAPixMatch [28]CVPR'210.7400.7750.6280.6600.7630.8110.6490.699
DAFormer [12]CVPR'220.7730.7960.6540.6800.7940.8240.6770.717
DAFormer-FIA [41]ECCV'240.7900.8100.6740.7000.8080.8340.6970.734
RimPixMatch [28]CVPR'210.6430.6730.4990.5230.6730.7190.5130.562
DAFormer [12]CVPR'220.4050.4180.2640.2700.4160.4480.2690.295
DAFormer-FIA [41]ECCV'240.5160.5310.3590.3690.5290.5580.3680.395
CupRandom-0.8160.8340.7110.7340.8530.8310.7640.731
Entropy [44]CVPR'190.8010.8230.6930.7210.8480.8210.7580.717
MHPL [45]CVPR'230.8230.8340.7180.7330.8460.8320.7520.731
DML-Core [43]NeurIPS'240.8240.8390.7200.7400.8550.8360.7660.737
Detective [61]CVPR'240.8220.8370.7220.7400.8540.8360.7630.738
STDR [47]IEEE TMI'240.8120.8370.7100.7380.8650.8350.7740.735
ADAOur FairAP-0.836*0.8400.730*0.743*0.8440.8390.7590.741*
RimRandom-0.7750.7800.6510.6540.7860.7790.6570.653
Entropy [44]CVPR'190.7610.7720.6350.6430.7850.7700.6540.641
MHPL [45]CVPR'230.7740.7820.6500.6560.7900.7800.6620.654
DML-Core [43]NeurIPS'240.7750.7830.6530.6580.7910.7810.6620.655
Detective [61]CVPR'240.7770.7840.6560.6590.7920.7820.6620.658
STDR [47]IEEE TMI'240.7750.7810.6520.6560.7880.7790.6600.654
Our FairAP-0.784*0.790*0.663*0.668*0.796*0.788*0.673*0.666*
", + "bbox": [ + 133, + 119, + 862, + 425 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4a6e528e7aa9a6b94e0bd9cadec8eb436fbd6d7488ca5899858ecb8d9fe42336.jpg", + "image_caption": [ + "Figure 5. Top: bar charts showing Dice-related (left) and IoU-related (right) scores across four ablation models with race attribute. Bottom: radar charts showing Dice (left) and IoU (right) values for Cup and Rim regions, divided by racial subgroups." + ], + "image_footnote": [], + "bbox": [ + 99, + 435, + 473, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fair-Poly, which chooses samples for Polysemy representativeness. 4) Our Fair-AP, which considers the average division of Fair-Base and prioritizes sample selection with both Attribute and Polysemy. Fig. 5 demonstrates the ablation study results. Standalone Fair-Attr or Fair-Poly strategies may be one-sided and show unstable improvements, while our FairAP—jointly modeling Attribute and Polysemy achieves significant gains in all metrics. The radar chart further confirms this across specific sub-groups, demonstrating our design's effectiveness.", + "bbox": [ + 89, + 699, + 482, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Exploration on Labeling Quota $\\alpha\\%$ . We investigate the impact of varying labeling quotas $\\alpha\\%$ on Fair-AP's performance compared to some representative ADA methods. As", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c866706a14acbd9849e5d095d4eeae3e49c18305200092a06c2e727a642e1b8c.jpg", + "image_caption": [ + "Figure 6. Ablation studies of different labeling quota $\\alpha \\%$ for Race attribute on the key composite metric ES-Dice." + ], + "image_footnote": [], + "bbox": [ + 522, + 435, + 707, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c0ec4b432d0d1e94021d5fd0d55a46ddddad311130ef71b2c8c9656e595d294.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 434, + 897, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shown in Fig. 6, increasing $\\alpha\\%$ improves model accuracy with diminishing returns: significant gains emerge from $0\\%, \\rightarrow, 2\\%, \\rightarrow, 5\\%$ , followed by marginal improvements beyond $5\\%$ . Crucially, our method consistently achieves superior performance across all quotas, justifying $\\alpha = 5\\%$ as the optimal accuracy-cost balance.", + "bbox": [ + 511, + 601, + 906, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 719, + 633, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we pioneer a novel fairness-aware ADA paradigm to address fairness in real-world cross-domain medical scenarios. By utilizing the VLM to perceive demographic attributes and devising the attribute-aware FairAP strategy, we significantly enhance both fairness and segmentation performance. Extensive experiments confirm that our approach surpasses existing DA and ADA methods, demonstrating superior performance and enhanced fairness, providing a fresh perspective on addressing fairness challenges toward more equitable AI-assisted healthcare.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24109", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the Guangdong Science and Technology Department (2024ZDZX2004), a General Research Fund of Hong Kong Research Grants Council (project no. 15218521), and a grant under the scheme of Collaborative Research with World-leading Research Groups in The Hong Kong Polytechnic University (project no. G-SACF).", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 236, + 187, + 253 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ehsan Adeli, Qingyu Zhao, Adolf Pfefferbaum, Edith V Sullivan, Li Fei-Fei, Juan Carlos Niebles, and Kilian M Pohl. Representation learning with statistical independence to mitigate bias. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2513–2523, 2021. 2", + "[2] David Arthur and Sergei Vassilvitskii. k-means++: The advantages of careful seeding. Technical report, Stanford, 2006. 5", + "[3] Jonathan T Barron. A generalization of otsu's method and minimum error thresholding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 455-470. Springer, 2020. 2", + "[4] Peter J Bevan and Amir Atapour-Abarghoueiet. Detecting melanoma fairly: Skin tone detection and debiasing for skin lesion classification. In MICCAI Workshop on Domain Adaptation and Representation Transfer, pages 1-11. Springer, 2022. 2", + "[5] Alceu Bissoto, Michel Fornaciali, Eduardo Valle, and Sandra Avila. (de) constructing bias on skin lesion datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 2", + "[6] Simon Caton and Christian Haas. Fairness in machine learning: A survey. ACM Computing Surveys, 56(7):1-38, 2024. 2", + "[7] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 6", + "[8] Jiayi Chen, Benteng Ma, Hengfei Cui, and Yong Xia. Think twice before selection: Federated evidential active learning for medical image analysis with domain shifts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11439-11449, 2024. 3", + "[9] Julia Dressel and Hany Farid. The accuracy, fairness, and limits of predicting recidivism. Science advances, 4(1): eaao5580, 2018. 2", + "[10] Judy Wawira Gichoya, Imon Banerjee, Ananth Reddy Bhimireddy, John L Burns, Leo Anthony Celi, Li-Ching Chen, Ramon Correa, Natalie Dullerud, Marzyeh Ghassemi, Shih-Cheng Huang, et al. Ai recognition of patient race in medical imaging: a modelling study. The Lancet Digital Health, 4(6):e406-e414, 2022. 1" + ], + "bbox": [ + 93, + 261, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Yuanbo Guo, Zheyu Yan, Xiaoting Yu, Qingpeng Kong, Joy Xie, Kevin Luo, Dewen Zeng, Yawen Wu, Zhenge Jia, and Yiyu Shi. Hardware design and the fairness of a neural network. Nature Electronics, 7(8):714-723, 2024. 2", + "[12] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9924-9935, 2022. 3, 6, 7, 8", + "[13] Mohamed Huti, Tiarna Lee, Elinor Sawyer, and Andrew P King. An investigation into race bias in random forest models based on breast dce-mri derived radiomics features. In Workshop on Clinical Image-Based Procedures, pages 225-234. Springer, 2023. 1", + "[14] Suyog Dutt Jain and Kristen Grauman. Active image segmentation propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2864-2873, 2016. 3", + "[15] Ruinan Jin, Zikang Xu, Yuan Zhong, Qionsong Yao, Qi Dou, S Kevin Zhou, and Xiaoxiao Li. Fairmedfm: fairness benchmarking for medical imaging foundation models. arXiv preprint arXiv:2407.00983, 2024. 2", + "[16] Thorsten Kalb, Kaisar Kushibar, Celia Cintas, Karim Lekadir, Oliver Diaz, and Richard Osuala. Revisiting skin tone fairness in dermatological lesion classification. In Workshop on Clinical Image-Based Procedures, pages 246-255. Springer, 2023. 1", + "[17] Michael P Kim, Amirata Ghorbani, and James Zou. Multiaccuracy: Black-box post-processing for fairness in classification. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 247-254, 2019. 2", + "[18] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[19] Newton M Kinyanjui, Timothy Odonga, Celia Cintas, Noel CF Codella, Rameswar Panda, Prasanna Sattigeri, and Kush R Varshney. Fairness of classifiers across skin tones in dermatology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 320-329. Springer, 2020. 1", + "[20] Qingpeng Kong, Ching-Hao Chiu, Dewen Zeng, Yu-Jen Chen, Tsung-Yi Ho, Jingtong Hu, and Yiyu Shi. Achieving fairness through channel pruning for dermatological disease diagnosis. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 24-34. Springer, 2024. 2", + "[21] David D Lewis and Jason Catlett. Heterogeneous uncertainty sampling for supervised learning. In Machine learning proceedings 1994, pages 148-156. Elsevier, 1994. 3", + "[22] Pranay K Lohia, Karthikeyan Natesan Ramamurthy, Manish Bhide, Diptikalyan Saha, Kush R Varshney, and Ruchir Puri. Bias mitigation post-processing for individual and group fairness. In Icassp 2019-2019 iee international conference on acoustics, speech and signal processing (icassp), pages 2847-2851. IEEE, 2019. 2", + "[23] Xiangde Luo, Hongqiu Wang, Jinfeng Xu, Lu Li, Yue Zhao, Yuan He, Hui Huang, Jianghong Xiao, Tao Song, Shichuan Zhang, et al. Generalizable magnetic resonance imaging-based nasopharyngeal carcinoma delineation: Bridging gaps" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24110", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "across multiple centers and raters with active learning. International Journal of Radiation Oncology* Biology* Physics, 121(5):1384-1393, 2025. 1", + "[24] Yan Luo, Min Shi, Muhammad Osama Khan, Muhammad Muneeb Afzal, Hao Huang, Shuaihang Yuan, Yu Tian, Luo Song, Ava Kouhana, Tobias Elze, et al. Fairclip: Harnessing fairness in vision-language learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12289-12301, 2024. 1, 2", + "[25] Yan Luo, Yu Tian, Min Shi, Louis R Pasquale, Lucy Q Shen, Nazlee Zebardast, Tobias Elze, and Mengyu Wang. Harvard glaucoma fairness: a retinal nerve disease dataset for fairness learning and fair identity normalization. IEEE Transactions on Medical Imaging, 2024. 2", + "[26] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM computing surveys (CSUR), 54 (6):1-35, 2021. 2", + "[27] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 415-430. Springer, 2020. 3", + "[28] Luke Melas-Kyriazi and Arjun K Manrai. Pixmatch: Unsupervised domain adaptation via pixelwise consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12435-12445, 2021. 6, 7, 8", + "[29] Esther Puyol-Anton, Bram Ruijsink, Jorge Mariscal Harana, Stefan K Piechnik, Stefan Neubauer, Steffen E Petersen, Reza Razavi, Phil Chowienczyk, and Andrew P King. Fairness in cardiac magnetic resonance imaging: assessing sex and racial bias in deep learning-based segmentation. Frontiers in cardiovascular medicine, 9:859310, 2022. 1, 2", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 6", + "[31] Fernanda Ribeiro, Valentina Shumovskaia, Thomas Davies, and Ira Ktena. How fair is your graph? exploring fairness concerns in neuroimaging studies. In Machine Learning for Healthcare Conference, pages 459-478. PMLR, 2022. 1", + "[32] María Agustina Ricci Lara, Rodrigo Echeveste, and Enzo Ferrante. Addressing fairness in artificial intelligence for medical imaging. nature communications, 13(1):4581, 2022. 1", + "[33] Zohaib Salahuddin, Yi Chen, Xian Zhong, Henry C Woodruff, Nastaran Mohammadian Rad, Shruti Atul Mali, and Philippe Lambin. From head and neck tumour and lymph node segmentation to survival prediction on pet/ct: an end-to-end framework featuring uncertainty, fairness, and multi-region multi-modal radiomics. Cancers, 15(7):1932, 2023. 1", + "[34] Candice Schwartz, Ifeanyi Beverly Chukwudozie, Silvia Tejeda, Ganga Vijayasiri, Ivy Abraham, Mylene Remo, Hiral A Shah, Maria Rojas, Alicia Carillo, Loraine Moreno," + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "et al. Association of population screening for breast cancer risk with use of mammography among women in medically underserved racial and ethnic minority groups. JAMA network open, 4(9):e2123751-e2123751, 2021. 1", + "[35] Laleh Seyyed-Kalantari, Guanxiong Liu, Matthew McDermott, Irene Y Chen, and Marzyeh Ghassemi. Chexclusion: Fairness gaps in deep chest x-ray classifiers. In BIOCOMPUTING 2021: proceedings of the Pacific symposium, pages 232-243. World Scientific, 2020. 1", + "[36] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA McDermott, Irene Y Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12):2176-2182, 2021. 1", + "[37] Dinggang Shen, Guorong Wu, and Heung-II Suk. Deep learning in medical image analysis. Annual review of biomedical engineering, 19(1):221-248, 2017. 1", + "[38] Min Shi, Muhammad Muneeb Afzal, Hao Huang, Congcong Wen, Yan Luo, Muhammad Osama Khan, Yu Tian, Leo Kim, Tobias Elze, Yi Fang, et al. Equitable deep learning for diabetic retinopathy detection using multi-dimensional retinal imaging with fair adaptive scaling: a retrospective study. medRxiv, pages 2024-04, 2024. 2", + "[39] Emma AM Stanley, Matthias Wilms, Pauline Mouches, and Nils D Forkert. Fairness-related performance and explainability effects in deep learning models for brain image analysis. Journal of Medical Imaging, 9(6):061102-061102, 2022. 1", + "[40] Yu Tian, Min Shi, Yan Luo, Ava Kouhana, Tobias Elze, and Mengyu Wang. Fairseg: A large-scale medical image segmentation dataset for fairness learning using segment anything model with fair error-bound scaling. In ICLR, 2024. 1, 2, 3, 4, 6", + "[41] Yu Tian, Congcong Wen, Min Shi, Muhammad Muneeb Afzal, Hao Huang, Muhammad Osama Khan, Yan Luo, Yi Fang, and Mengyu Wang. Fairdomain: Achieving fairness in cross-domain medical image segmentation and classification. In European Conference on Computer Vision, pages 251-271. Springer, 2024. 1, 2, 3, 4, 6, 7, 8", + "[42] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7472-7481, 2018. 2", + "[43] Arvind Vepa, Zukang Yang, Andrew Choi, Jungseock Joo, Fabien Scalzo, and Yizhou Sun. Integrating deep metric learning with coreset for active learning in 3d segmentation. Advances in Neural Information Processing Systems, 37:71643-71671, 2024. 7, 8", + "[44] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 2, 7, 8", + "[45] Fan Wang, Zhongyi Han, Zhiyan Zhang, Rundong He, and Yilong Yin. Mhpl: Minimum happy points learning for" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24111", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "active source free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20008-20018, 2023. 2, 3, 6, 7, 8", + "[46] Hongqiu Wang, Yueming Jin, and Lei Zhu. Dynamic interactive relation capturing via scene graph learning for robotic surgical report generation. arXiv preprint arXiv:2306.02651, 2023. 1", + "[47] Hongqiu Wang, Jian Chen, Shichen Zhang, Yuan He, Jinfeng Xu, Mengwan Wu, Jinlan He, Wenjun Liao, and Xiangde Luo. Dual-reference source-free active domain adaptation for nasopharyngeal carcinoma tumor segmentation across multiple hospitals. IEEE Transactions on Medical Imaging, 2024. 3, 6, 7, 8", + "[48] Hongqiu Wang, Xiangde Luo, Wu Chen, Qingqing Tang, Mei Xin, Qiong Wang, and Lei Zhu. Advancing uwf-slo vessel segmentation with source-free active domain adaptation and a novel multi-center dataset. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 75-85. Springer, 2024. 3", + "[49] Hongqiu Wang, Zhaohu Xing, Weitong Wu, Yijun Yang, Qingqing Tang, Meixia Zhang, Yanwu Xu, and Lei Zhu. Non-invasive to invasive: Enhancing ffa synthesis from cfp with a benchmark dataset and a novel network. In Proceedings of the 1st International Workshop on Multimedia Computing for Health and Medicine, pages 7-15, 2024. 1", + "[50] Hongqiu Wang, Guang Yang, Shichen Zhang, Jing Qin, Yike Guo, Bo Xu, Yueming Jin, and Lei Zhu. Video-instrument synergistic network for referring video instrument segmentation in robotic surgery. IEEE Transactions on Medical Imaging, 2024.", + "[51] Hongqiu Wang, Yixian Chen, Wu Chen, Huihui Xu, Haoyu Zhao, Bin Sheng, Huazhu Fu, Guang Yang, and Lei Zhu. Serp-mamba: Advancing high-resolution retinal vessel segmentation with selective state-space model. IEEE Transactions on Medical Imaging, 2025. 1", + "[52] Tianxiang Wang and Qun Dai. A patch distribution-based active learning method for multiple instance alzheimer's disease diagnosis. Pattern Recognition, 150:110341, 2024. 3", + "[53] Zhibo Wang, Xiaowei Dong, Henry Xue, Zhifei Zhang, Weifeng Chiu, Tao Wei, and Kui Ren. Fairness-aware adversarial perturbation towards bias mitigation for deployed deep models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10379-10388, 2022. 2", + "[54] Garrett Wilson and Diane J Cook. A survey of unsupervised deep domain adaptation. ACM Transactions on Intelligent Systems and Technology (TIST), 11(5):1-46, 2020. 1, 2", + "[55] Wenshuai Xu, Zhenghui Hu, Yu Lu, Jinzhou Meng, Qingjie Liu, and Yunhong Wang. Activedc: Distribution calibration for active finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16996-17005, 2024. 3", + "[56] Zhao Xu, Kai Yu, Volker Tresp, Xiaowei Xu, and Jizhi Wang. Representative sampling for text classification using support vector machines. In Advances in Information Retrieval: 25th European Conference on IR Research, ECIR 2003, Pisa, Italy, April 14-16, 2003. Proceedings 25, pages 393-407. Springer, 2003. 3" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Zikang Xu, Jun Li, Qingsong Yao, Han Li, Mingyue Zhao, and S Kevin Zhou. Addressing fairness issues in deep learning-based medical image analysis: a systematic review. npj Digital Medicine, 7(1):286, 2024. 1, 2", + "[58] Jenny Yang, Andrew AS Soltan, David W Eyre, Yang Yang, and David A Clifton. An adversarial training framework for mitigating algorithmic biases in clinical machine learning. NPJ digital medicine, 6(1):55, 2023. 2", + "[59] Ruichen Yao, Ziteng Cui, Xiaoxiao Li, and Lin Gu. Improving fairness in image classification via sketching. arXiv preprint arXiv:2211.00168, 2022. 2", + "[60] Chenxi Yuan, Kristin A Linn, and Rebecca A Hubbard. Algorithmic fairness of machine learning models for alzheimer disease progression. JAMA Network Open, 6(11):e2342203-e2342203, 2023. 1", + "[61] Wenqiao Zhang, Zheqi Lv, Hao Zhou, Jia-Wei Liu, Juncheng Li, Mengze Li, Yunfei Li, Dongping Zhang, Yueting Zhuang, and Siliang Tang. Revisiting the domain shift and sample uncertainty in multi-source active domain transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16751-16761, 2024. 3, 7, 8", + "[62] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. 2", + "[63] Haoyu Zhao, Hao Wang, Xingyue Zhao, Hongqiu Wang, Zhiyu Wu, Chengjiang Long, and Hua Zou. Automated 3d physical simulation of open-world scene with gaussian splatt-ing. arXiv e-prints, pages arXiv-2411, 2024. 4", + "[64] Qingyu Zhao, Ehsan Adeli, and Kilian M Pohl. Training confounder-free deep learning models for medical applications. Nature communications, 11(1):6010, 2020. 2", + "[65] Yongshuo Zong, Yongxin Yang, and Timothy Hospedales. Medfair: Benchmarking fairness for medical imaging. arXiv preprint arXiv:2210.01725, 2022. 2", + "[66] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 3" + ], + "bbox": [ + 516, + 92, + 903, + 655 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24112", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/0cab628c-59eb-4cba-b70b-2c418218511f_model.json b/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/0cab628c-59eb-4cba-b70b-2c418218511f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..af88d3bc93f4a7c146fdd19489fbed374efad6de --- /dev/null +++ b/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/0cab628c-59eb-4cba-b70b-2c418218511f_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "image", + "bbox": [ + 0.112, + 0.127, + 0.159, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.136, + 0.887, + 0.182 + ], + "angle": 0, + "content": "Toward Fair and Accurate Cross-Domain Medical Image Segmentation: A VLM-Driven Active Domain Adaptation Paradigm" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.208, + 0.81, + 0.244 + ], + "angle": 0, + "content": "Hongqiu Wang\\(^{1}\\) Wu Chen\\(^{2}\\) Xiangde Luo\\(^{3}\\) Zhaohu Xing\\(^{1}\\) Lihao Liu\\(^{4}\\) Jing Qin\\(^{5}\\) Shaozhi Wu\\(^{2*}\\) Lei Zhu\\(^{1,6*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.246, + 0.861, + 0.278 + ], + "angle": 0, + "content": "\\(^{1}\\)Hong Kong University of Science and Technology (Guangzhou) \\(^{2}\\)University of Electronic Science and Technology of China \\(^{3}\\)West China Hospital \\(^{4}\\)Amazon \\(^{5}\\)Hong Kong Polytechnic University \\(^{6}\\)Hong Kong University of Science and Technology" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.341, + 0.485, + 0.782 + ], + "angle": 0, + "content": "Fairness in AI-assisted medical image analysis is crucial for equitable healthcare, but is often neglected, especially in prevalent cross-domain scenarios (diverse demographics and imaging protocols). Effective and equitable deployment of AI models in these scenarios is critical, yet traditional Unsupervised Domain Adaptation (UDA) methods exhibit limited improvements. Emerging Active Domain Adaptation (ADA) approaches offer more effective enhancements, but all ignore fairness issues. Therefore, in this work, we propose the first fairness-aware ADA paradigm that simultaneously achieves both enhanced fairness and superior overall performance. Our method leverages the multimodal alignment capability of Vision-Language Models (VLMs): By performing medical images (vision) and sensitive attributes (language) learning, VLM inherently captures semantic correlations between visual features and protected attributes, enabling explicit attributes representation. Building on this foundation, we further devise an attribute-aware strategy (FairAP), which dynamically adapts to diverse patient demographics to promote equitable and high-quality outcomes by considering both Attribute and Polysemy. Extensive experiments on the FairDomain benchmark demonstrate that our method significantly reduces bias and maintains leading performance, outperforming existing UDA and ADA methods. This work pioneers a VLM-driven ADA paradigm for fair cross-domain medical segmentation, offering a blueprint for effective and equitable AI deployment in clinical practice. Code will be on (https://github.com/whq-xxh/Fair-AP)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.823, + 0.222, + 0.839 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.848, + 0.483, + 0.879 + ], + "angle": 0, + "content": "\"Fairness is not an attitude. It's a professional skill that must be developed and exercised.\" — Brit Hume." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.309, + 0.907, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.528 + ], + "angle": 0, + "content": "Figure 1. Comparison of previous cross-domain fairness methods with our VLM-Driven fairness-aware ADA paradigm. Our method combines VLM and ADA to achieve both leading model performance and enhanced fairness." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.538, + 0.907, + 0.689 + ], + "angle": 0, + "content": "The advent of AI technologies has significantly transformed medical image analysis (MedIA), providing sophisticated tools for analyzing complex biological patterns within imaging data [37, 46, 49-51]. These advanced algorithms enhance diagnostic precision, support informed clinical decisions, and improve patient outcomes. However, as these models become increasingly integrated into clinical practice, critical challenges have emerged: inherent biases within these models [24, 32, 57] and domain shifts caused by diverse clinical scenarios [23, 41, 54]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Numerous studies have reported that biases in MedIA tasks [13, 16, 19, 29, 31, 33, 35, 39, 60], stemming from training datasets failing to represent diverse populations adequately, raise profound ethical and fairness concerns [10, 34]. These biases cause severe health inequities, impacting marginalized groups in clinical applications. For example, Black individuals face nearly twice the risk of glaucoma than others, highlighting the need for accurately segmenting the optic disc and cup for early diagnosis, yet AI models often perform worst in this subgroup. This issue is particularly critical in cross-domain scenarios, where models deployed across institutions encounter heterogeneous data distributions that amplify performance disparities [36, 40, 41]. Such domain-specific biases not only undermine the per" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.889, + 0.23, + 0.9 + ], + "angle": 0, + "content": "\\*corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24102" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "formance of AI systems but also violate fairness. Thus, ensuring both efficient and fair adaptation in cross-domain real-world medical AI environments is crucial to translate advanced technology to equitable healthcare for all." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.156, + 0.485, + 0.444 + ], + "angle": 0, + "content": "Recently, significant efforts have been made to tackle fairness across various fields [1, 3, 4, 9, 11, 15, 20, 24, 29, 38-40, 58, 62, 64, 65]. These studies explore various attributes across different modalities or body parts. However, real-world cross-domain challenges remain underexplored. While Tian et al. [41] established the multiattribute benchmark dataset to evaluate cross-domain fairness, the limited performance of its Unsupervised Domain Adaptation (UDA) framework (e.g., IoU values \\(< 40\\%\\) in Tables 1 & 3) severely restricts clinical applicability despite its fairness-aware design. Although emerging Active Domain Adaptation (ADA) has demonstrated significant performance gains over UDA by actively selecting valuable target-domain samples [45], its fairness implications remain largely unexplored. This creates a critical gap in achieving both equitable outcomes and clinically viable performance across demographic subgroups, as current methods force an undesirable trade-off between the fairness guarantees and the model effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.448, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Therefore, in this work, we propose the first fairness-aware ADA paradigm to simultaneously enhance fairness across all subgroups and boost overall performance for cross-domain medical image segmentation scenario applications. Specifically, according to the unique characteristics of medical images and corresponding sensitive attribute information (such as race: Black, White, Asian), we present a novel perspective that integrates the Vision-Language Model (VLM) to align image features with sensitive attributes. This alignment endows our model with the capability to perceive the sensitive attribute representations within images accurately. This VLM-attribute learning is essential as it establishes a solid foundation for subsequent representative attributes and polysemy sample selection. Then, we devise an attribute-aware FairAP sample selection strategy, where we select a quite small (only \\(5\\%\\) here) but critical set of samples and annotate them for target fair model fine-tuning. The FairAP strategy allocates equal quotas to each subgroup to ensure balanced attention and leverages VLM attribute-latent space representations for attribute-representative and polysemic sample selection. Our process and its key distinctions from previous methods are visually demonstrated in Fig. 1. It is worth noting that the FairAP strategy is highly practical as it significantly reduces the workload of doctors, while simultaneously achieving both fairness and high performance (even improving IoU by around \\(30\\%\\) than FairDA method [41] for rim segmentation as shown in Table 1). Comprehensive experiments on the FairDomain benchmark [41] confirm that our approach markedly diminishes bias and retains cutting" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "edge performance in segmentation tasks, surpassing traditional DA and ADA methods." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.123, + 0.895, + 0.137 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.139, + 0.905, + 0.198 + ], + "angle": 0, + "content": "- To the best of our knowledge, we are the first to tackle fairness challenges in cross-domain medical image segmentation with ADA paradigm, dynamically adapting to diverse patient subgroups for equitable outcomes." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.2, + 0.905, + 0.258 + ], + "angle": 0, + "content": "- We propose to leverage the Vision-Language Model to seamlessly integrate demographic attributes like race or gender into the learning process, endowing the model with the ability to perceive these sensitive attributes." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.26, + 0.905, + 0.333 + ], + "angle": 0, + "content": "- We devise the attribute-aware FairAP strategy to ensure both fairness and effectiveness. FairAP optimizes representation by balancing samples from underrepresented demographic groups while maintaining diversity to enhance overall performance comprehensively." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.335, + 0.905, + 0.425 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our method surpasses previous DA and ADA approaches by ensuring superior fairness across multiple protected attributes and maintaining top performance. Our method pioneers a novel solution paradigm for fairness and effectiveness in cross-domain medical AI model deployment." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.139, + 0.905, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.441, + 0.664, + 0.457 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.769, + 0.483 + ], + "angle": 0, + "content": "2.1. Fairness in Medical Imaging" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.908, + 0.745 + ], + "angle": 0, + "content": "In medical image processing, fairness aims to reduce model biases to ensure equitable outcomes across patient subgroups. Existing methods fall into three main categories: pre-processing, in-processing, and post-processing [6, 26, 57]. Pre-processing techniques, such as those proposed by Bissoto et al. [5] and others [24, 29, 59], aim to correct training data imbalances, but may affect computational efficiency. In-processing methods mitigate bias during training by incorporating fairness-aware loss functions [40] or fairness-specific modules [20, 25, 41], which can potentially trade off some accuracy. On the other hand, post-processing approaches, like those introduced by Wang et al. [53], Kim et al. [17], and others [22], adjust the model outputs to improve fairness, which may struggle to generalize well during testing. However, few studies focus on fairness in cross-domain scenarios within real-world deployment environments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.704, + 0.773 + ], + "angle": 0, + "content": "2.2. Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Unsupervised Domain Adaptation (UDA): UDA trains on the labeled source and unlabeled target data to bridge domain gaps, typically using adversarial training and minimizing inter-domain discrepancies [54]. In UDA methods, adversarial learning minimizes distributional differences between source and target domains, enhancing performance in the target domain [42, 44]. Additionally, self-training in UDA uses pseudo-labels from target domain data to en" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24103" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.088, + 0.885, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.453, + 0.908, + 0.527 + ], + "angle": 0, + "content": "Figure 2. Overview of the Fair VLM-Driven ADA method (taking race attribute for illustration). Given the target domain images \\( x^T \\) and their corresponding sensitive attribute information \\( \\text{attrs} \\), we first align them through VLM-attribute learning to obtain the vision-based sensitive attribute features \\( F_{attrR} \\). Then, we devise the attribute-aware FairAP Select strategy to select \\( \\alpha \\% \\) of the samples that are representations of Attribute and Polysemy. These selected samples \\( S \\) are then annotated by experts to obtain the annotations \\( y^{TL} \\). Finally, active domain adaptation fine-tuning is performed to obtain the target fair model \\( h_{\\phi}' \\) for top performance and enhanced fairness." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.54, + 0.486, + 0.664 + ], + "angle": 0, + "content": "hance performance in the target domain [12, 27, 66]. Building on the work by Hoyer et al. [12], Tian et al. [41] further explored fairness in cross-domain settings by integrating the fairness module. This new setting can more accurately reflect real-world clinical scenarios and provide a clear quantification of the performance and fairness of cross-domain methods [41]. However, very limited performance gains brought by UDA significantly diminish its clinical utility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Active Domain Adaptation (ADA): ADA combines active learning and domain adaptation to reduce annotation costs and sustain high performance by focusing on annotating the most valuable samples through strategies like uncertainty-based selection and diversity-based sampling [8, 14, 21, 48, 56]. For example, Wang et al. [45], and others [47, 52, 55, 61] enhanced cross-domain performance by strategically selecting target domain samples for model development. However, these current ADA methods have predominantly concentrated on optimizing overall model performance, neglecting the crucial aspect of fairness. Consequently, we are the pioneers in introducing a novel fairness-aware ADA paradigm, effectively minimizing annotation costs while simultaneously achieving both enhanced fairness and leading performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.539, + 0.651, + 0.558 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.569, + 0.676, + 0.586 + ], + "angle": 0, + "content": "3.1. Problem Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Cross-domain MedIA often exhibits not only conventional performance degradation for domain shifts but also issues of unfairness related to sensitive attributes. Fair ADA is an effective and novel approach to address this challenge. Our work specifically targets fairness challenges in domain adaptation, focusing on fairness issues across multiple sensitive attributes in diverse subgroups. The problem is defined as follows: In medical segmentation tasks, consider the source domain data \\( D_{S} = \\{(x_{i}^{S},y_{i}^{S},\\text{attr}_{i})\\}_{i = 1}^{N_{S}} \\), where \\( x_{i}^{S}\\in \\mathbb{R}^{H\\times W} \\) represents the \\( i_{th} \\) image, the \\( y_{i}^{S}\\in \\mathbb{R}^{C\\times H\\times W} \\) denotes the corresponding label with \\( C \\) classes, and the \\( \\text{attr}_{i}\\in \\mathbb{R}^{n} \\) means the sensitive attributes related to the sample, such as gender, race, or ethnicity. Meanwhile, there is an unlabeled dataset \\( D_{T} = \\{(x_{i}^{T},\\text{attr}_{i})\\}_{i = 1}^{N_{T}} \\) whose domain vary, leading to different image features from \\( D_{S} \\). In the fair segmentation setting, note that the target attributes \\( \\text{attr}_{i} \\) are available and accessible, as outlined in [40, 41]. The aim is to design strategies for selecting a small number of samples \\( S = \\alpha \\% \\cdot D_{T}\\ll D_{T} \\) for labeling and active learning, thereby maximizing the fairness" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24104" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "and performance of model \\( h_{\\phi} \\) as much as possible while maintaining \\( \\alpha \\% \\) low." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.129, + 0.33, + 0.145 + ], + "angle": 0, + "content": "3.2. Overview of the Fair-ADA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.151, + 0.484, + 0.378 + ], + "angle": 0, + "content": "The overall structure of our proposed VLM-driven fairness-aware ADA framework is illustrated in Fig. 2. The entire methodology section uses race attribute \\(attrR\\) as an example to illustrate. In the VLM-attribute Learning stage (a), we utilize a Vision-Language model \\(g_{\\theta}\\) to align image samples with sensitive attribute information within the latent space. The model \\(g_{\\theta}\\) performs vision (unlabeled medical images \\(x^{T}\\)) language (corresponding attribute information \\(attrR\\)) learning to achieve alignment. Then, we can extract vision-based sensitive attribute features \\(F_{attrR}\\). These feature representations consist of multiple subgroups corresponding to various attributes. For example, sensitive racial features \\(F_{attrR}\\) include three subgroup features such as Asian features \\(F_{a}\\), Black features \\(F_{b}\\), and White features \\(F_{w}\\). The entire process can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.385, + 0.483, + 0.403 + ], + "angle": 0, + "content": "\\[\nF _ {a t t r R} = g _ {\\theta} \\left(x ^ {T}, a t t r R\\right) = \\left\\{F _ {a}, F _ {b}, F _ {w} \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.59 + ], + "angle": 0, + "content": "Based on the vision-based sensitive attribute features \\( F_{attrR} \\), the process moves into the FairAP Select stage (b), where our approach is organized into three main steps. In the first step, we ensure and amplify fairness within our active learning strategy by allocating an equal quota of samples from each sensitive attribute subcategory. In the second step, we employ clustering methods for each subgroup based on various sensitive attribute representations, followed by selecting attribute representative samples. In the third step, we further optimize the selected samples by expanding the selection to include polysemy representatives. This overall process can be summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.598, + 0.482, + 0.614 + ], + "angle": 0, + "content": "\\[\nS = \\text {F a i r} A P \\left(F _ {\\text {a t t r} R}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.484, + 0.713 + ], + "angle": 0, + "content": "In the Active Domain Adaptation stage (c), we utilize the model \\( h_{\\phi} \\) that has been previously trained on the source domain data \\( D_{S} \\). We then fine-tune this model using the sample set \\( S \\) selected by our FairAP strategy to produce the optimized target fair model \\( h_{\\phi}^{\\prime} \\). The loss function for this model is defined by the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.72, + 0.483, + 0.76 + ], + "angle": 0, + "content": "\\[\nL _ {A D A} \\left(h _ {\\phi} ^ {\\prime} (x ^ {T L}), y ^ {T L}\\right) = - \\sum_ {c = 1} ^ {C} y _ {c} \\log \\left(h _ {\\phi} ^ {\\prime} (x)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.766, + 0.484, + 0.811 + ], + "angle": 0, + "content": "where \\(C\\) represents the number of classes, and selected list \\(S = \\{x^{TL},y^{TL}\\}\\), whose size corresponds to \\(\\alpha \\%\\) of the total dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.319, + 0.835 + ], + "angle": 0, + "content": "3.3. VLM-attribute Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Traditional fairness-aware segmentation methods [40, 41] using the simplistic categorical encoding (e.g., 0/1 for gender male or female) suffer from semantic sparsity and contextual blindness, which fails to capture attribute-visual" + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.089, + 0.673, + 0.102 + ], + "angle": 0, + "content": "VLM-attribute learning" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.103, + 0.896, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.221, + 0.906, + 0.277 + ], + "angle": 0, + "content": "Figure 3. Visualization of racial feature distribution before and after feature alignment via VLM-attribute learning. Our alignment process creates a denser distribution of features within racial subgroups, improving distinguishability across different races." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.906, + 0.383 + ], + "angle": 0, + "content": "correlations. These disjoint representations limit ADA paradigms' ability to perceive subgroup distributions during cross-domain adaptation. To address this, we propose the VLM-based attribute-visual alignment to enable context-aware perception and provide continuous semantic gradients for precise subgroup-aware adaptation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.907, + 0.672 + ], + "angle": 0, + "content": "Fig. 2 (a) provides a detailed illustration of our vision language learning process. Specifically, we utilize a pretrained vision-language model (CLIP [30, 63] here) denoted as \\( g_{\\theta} \\) to process the samples \\( x^{T} \\) along with their associated sensitive attribute information \\( \\text{attr } R \\). The model's vision encoder, \\( g_{\\theta_v}^v \\), extracts image tokens \\( z_{v} \\) from the input \\( x^{T} \\) with parameters \\( \\theta_{v} \\). Meanwhile, the language encoder, \\( g_{\\theta_l}^l \\), derives attribute tokens \\( z_{l} \\) from the input \\( \\text{attr } R \\) with parameters \\( \\theta_{l} \\). Therefore, we can express the model as \\( g_{\\theta} = \\{g_{\\theta_v}^v,g_\\theta_l^l\\} \\) with the parameters as \\( \\theta = \\{\\theta_v,\\theta_l\\} \\). When training the model \\( g_{\\theta} \\) using a batch of \\( N \\) image-attribute token pairs, we generate a similarity matrix \\( A\\in \\mathbb{R}^{N\\times N} \\). Each entry of the matrix, \\( A_{ij} = z_{vi}^{T}z_{lj} \\), represents the cosine similarity between the \\( i_{th} \\) image token \\( z_{vi} \\) and \\( j_{th} \\) attribute token \\( z_{lj} \\) within the batch. The model \\( g_{\\theta} \\) employs a contrastive learning approach that aims to maximize the cosine similarity for positive image-attribute token pairs while minimizing it for negative token pairs. The model \\( g_{\\theta} \\) loss function can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.683, + 0.906, + 0.723 + ], + "angle": 0, + "content": "\\[\nL _ {V L M} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {1}{2} \\log \\left(\\frac {\\exp (2 A _ {i i})}{S _ {i} T _ {i}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.906, + 0.8 + ], + "angle": 0, + "content": "where \\( S_{i} = \\sum_{j=1}^{N} \\exp(A_{ij}) \\) and \\( T_{i} = \\sum_{j=1}^{N} \\exp(A_{ji}) \\). Since our focus is on enhancing the vision-based attribute perception capabilities, we choose to freeze the language encoder \\( g_{\\theta}^{l} \\) to better train \\( g_{\\theta} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.81, + 0.662, + 0.826 + ], + "angle": 0, + "content": "\\[\nA ^ {\\prime} \\leftarrow A,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.829, + 0.905, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\theta_ {v} ^ {\\prime} \\leftarrow \\theta_ {v} - \\eta \\frac {\\partial L _ {V L M}}{\\partial \\theta_ {v}}, \\quad \\theta_ {l} ^ {\\prime} = \\theta_ {l}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the learning rate. We employ a consistent template for sensitive attribute information as:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24105" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.134, + 0.092, + 0.435, + 0.108 + ], + "angle": 0, + "content": "A photo of a fundus and the race is \\(\\langle \\mathbf{RACE}\\rangle\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.19 + ], + "angle": 0, + "content": "The \\(\\langle \\mathbf{RACE}\\rangle\\) represents specific racial information. After completing the training, we utilize a fixed vision encoder \\(g_{\\theta_v'}^v\\) to process all target domain images \\(x^T\\), allowing extraction of the vision-based sensitive attribute features \\(F_{attrR}\\) as following:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.19, + 0.482, + 0.21 + ], + "angle": 0, + "content": "\\[\nF _ {a t t r R} = g _ {\\theta_ {v} ^ {\\prime}} ^ {v} \\left(x ^ {T}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.484, + 0.245 + ], + "angle": 0, + "content": "The visualization of racial characteristic distribution before and after our VLM-attribute learning is shown in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.251, + 0.254, + 0.268 + ], + "angle": 0, + "content": "3.4. FairAP Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.485, + 0.455 + ], + "angle": 0, + "content": "Building on the VLM-attribute learning described earlier, sensitive vision-based attribute features \\( F_{attrR} \\) are successfully extracted. For clarity, we continue using sensitive racial features \\( F_{attrR} \\) as an example to illustrate the Fair Attribute and Polysemy representation (FairAP) Select strategy. The sensitive racial features \\( F_{attrR} \\) encompass three types of racial characteristics: Asian features \\( F_{a} \\), Black features \\( F_{b} \\), and White features \\( F_{w} \\). Thus, we define \\( F_{attrR} = \\{F_{a}, F_{b}, F_{w}\\} \\). Our FairAP Select strategy encompasses three critical components: Fair Quota Select, Attribute Representation Select, and Polysemy Representation Select. We will introduce each component in sequence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.532 + ], + "angle": 0, + "content": "Fair Quota Selection. We recognize the data-driven nature of AI. Therefore, to achieve our fairness objective ('treat all our subgroups with fairness'), we ensure a consistent sample quota for each racial subgroup. We define the sample quota for each subgroup as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.541, + 0.482, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {a} = \\Delta_ {b} = \\Delta_ {w} = \\Delta_ {R}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.484, + 0.718 + ], + "angle": 0, + "content": "where \\(\\Delta_{a}\\), \\(\\Delta_{b}\\), and \\(\\Delta_{w}\\) represent the quotas given to Asians, Blacks, and Whites respectively. The \\(\\Delta_{R}\\) is calculated by dividing the overall quota \\((\\alpha \\% \\cdot D_T)\\) by the number of total racial subgroups, which is three here. Upon completing the evaluations of the subsequent components (Attribute and Polysemy), which assist in determining the specific samples, we compile the final selection list for each racial subgroup, represented as \\(S_{R} = \\{S_{\\hat{R}A}, S_{\\hat{R}P}\\}\\). Finally, we combine the selection lists \\(S_{R}\\) of all racial subgroups and submitted them to doctors for annotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Attribute Representation Selection. To optimally select representative samples for each racial subgroup, we undertake clustering operation (k-means++ algorithm [2] here) to process the feature set \\( F_{R} \\), resulting in the clustered features \\( F_{R}^{\\prime} \\) and the corresponding cluster centers \\( C_{R}^{N_{R}} \\), where \\( R \\) represents a specific racial group (like Asian, Black, or White). This process is represented as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.833, + 0.484, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\left\\{F _ {R} ^ {\\prime}, C _ {R} ^ {N _ {R}} \\right\\} = \\operatorname {C l u s t e r} \\left(F _ {R}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this equation, \\( C_R^{N_R} \\) denotes the cluster centers, and \\( N_R \\) indicates the number of cluster centers for the racial sub" + }, + { + "type": "code_caption", + "bbox": [ + 0.525, + 0.095, + 0.808, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: POLYSEMY-REP SELECT" + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.113, + 0.905, + 0.525 + ], + "angle": 0, + "content": "Input: The \\(i_{th}\\) cluster center of subgroup \\(R\\): \\(C_R^i\\), distance set: \\(D_R^i\\), selected distance set: \\(D_{\\hat{R}A}^i\\), features: \\(F_{R,m}''\\), number of \\(F_{R,m}': M_i\\), other racial subgroups' cluster centers: \\(C_{\\bar{R}}^k\\), quota: \\(\\Delta_R\\), subgroup cluster number: \\(N_R\\). Output: Polysemy-Rep selective racial list: \\(S_{\\hat{R}P}\\) \n1 for \\(i = 1\\) to \\(N_R\\) do \n2 \\({D_R'} \\gets {D_R} \\backslash {D_{\\hat{R}A}}^i // \\text{Remaining distances}\\) \n3 \\(F_{R,m}'' \\gets \\text{Index} \\left( D_R^{i'} \\right) // \\text{Indexing features}\\) \n4 \\(N_i \\gets M_i - \\frac{\\Delta_R}{2N_R} // \\text{Adjust selection range}\\) \n5 Initialize empty list \\(E_{temp}\\); \n6 for \\(m = 1\\) to \\(N_i\\) do \n7 if \\(F_{R,m}'' \\in C_R^i \\land F_{R,m}'' \\notin C_{\\bar{R}}^k\\) then \n8 \\(\\begin{array}{c} e_i^m \\gets \\min \\left( \\| F_{R,m}' - C_{\\bar{R}}^k \\| \\right) // \\text{Min distance to other racial clusters} \\\\ \\text{Append } e_i^m \\text{ to } E_{temp}; \\end{array}\\) \n9 end \n10 end \n11 Sort \\(E_{temp}\\) in rising order; \n12 \\(E_{\\hat{R}P}^i \\gets \\text{top min distances from } E_{temp}\\); \n14 end \n15 \\(F_{\\hat{R}P}' \\gets \\text{Index} \\left( \\{E_{\\hat{R}P}', ..., E_{\\hat{R}P}', ..., E_{\\hat{R}P}^{N_R}\\} \\right)\\) \n16 \\(S_{\\hat{R}P} \\gets F_{\\hat{R}P}' // \\text{Get the final Polysemy-Rep racial subgroup list}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.616 + ], + "angle": 0, + "content": "group. For each racial subgroup, the number of cluster centers \\( C_R^{N_R} \\) is calculated based on their overall distribution in the dataset to comprehensively represent the target domain. This is illustrated by the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.627, + 0.905, + 0.643 + ], + "angle": 0, + "content": "\\[\nN _ {a}: N _ {b}: N _ {w} \\approx \\operatorname {S u m} _ {a}: \\operatorname {S u m} _ {b}: \\operatorname {S u m} _ {w}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.907, + 0.764 + ], + "angle": 0, + "content": "where \\(\\mathrm{Sum}_a\\), \\(\\mathrm{Sum}_b\\), and \\(\\mathrm{Sum}_w\\) denote the proportions of Asians, Blacks, and Whites, respectively. For the \\(i_{th}\\) cluster center \\(C_R^i\\) of race \\(R\\), we define a set of Euclidean distances \\(D_R^i\\), where each element \\(d_i^m\\) in the set is the distance from sample \\(F_{R,m}^{\\prime}\\) (\\(m_{th}\\) sample belongs to the \\(i_{th}\\) cluster) to the cluster center \\(C_R^i\\). These distances are then sorted in ascending order. This can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.772, + 0.905, + 0.838 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D _ {R} ^ {i} = \\operatorname {s o r t} \\left(\\left\\{d _ {i} ^ {m} = \\| F _ {R, m} ^ {\\prime} - C _ {R} ^ {i} \\| \\mid F _ {R, m} ^ {\\prime} \\in C _ {R} ^ {i} \\right\\}\\right) \\\\ = \\left\\{d _ {i} ^ {1}, d _ {i} ^ {2}, \\dots , d _ {i} ^ {m}, \\dots , d _ {i} ^ {M _ {i}} \\right\\}, \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.838, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(d_i^1 \\leq d_i^2 \\leq \\dots \\leq d_i^m \\leq \\dots \\leq d_i^{M_i}\\), and \\(M_{i}\\) is the number of samples \\(F_{R}^{\\prime}\\) that belong to the cluster center \\(C_R^i\\). To ensure that the selected attribute representatives \\(F_{\\hat{R}A}\\) can comprehensively cover the entire distribution of the dataset," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24106" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.485, + 0.19 + ], + "angle": 0, + "content": "we establish an equal selection quota of \\(\\frac{\\Delta_R}{2N_R}\\) for each cluster. Specifically, we select the top \\(\\frac{\\Delta_R}{2N_R}\\) samples, denoted as the subset \\(D_{\\hat{R}A}^i\\), corresponding to the smallest Euclidean distances from the sorted set \\(D_R^i\\). The set \\(D_{\\hat{R}A}\\) represents the distances selected from the \\(N_R\\) clusters of the current racial subgroup. This can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.199, + 0.483, + 0.226 + ], + "angle": 0, + "content": "\\[\nD _ {R \\hat {A}} ^ {i} = \\left\\{d _ {R} ^ {1}, d _ {R} ^ {2}, \\dots , d _ {R} ^ {\\Delta_ {R} / 2 N _ {R}} \\right\\} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.233, + 0.483, + 0.259 + ], + "angle": 0, + "content": "\\[\nD _ {\\hat {R A}} = \\left\\{D _ {\\hat {R A}} ^ {1}, D _ {\\hat {R A}} ^ {2}, \\dots , D _ {\\hat {R A}} ^ {N _ {R}} \\right\\}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.326 + ], + "angle": 0, + "content": "Using the indices from the set \\(D_{\\hat{R}A}\\) that correspond to the smallest distances, we can derive the feature set \\(F_{\\hat{R}A}^{\\prime}\\) for attribute representative selection and obtain the corresponding sample set \\(S_{\\hat{R}A}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.459 + ], + "angle": 0, + "content": "Polysemy Representation Selection. Given that representative samples for each racial subgroup have been selected in the previous phase, we focus on choosing diverse samples in the Polysemy phase to explore a wider range of expressions within a limited quota, thus enhancing performance. For example, as seen in the right (post-VLM learning) of Fig. 3, some samples within the Black clusters are actually from White individuals, which may embody characteristics of both White and similar attributes of Black." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.484, + 0.691 + ], + "angle": 0, + "content": "In detail, We first filter out the samples selected in the previous stage, as indicated by \\( D_{R}^{i'} = D_{R}^{i} \\setminus D_{\\hat{R}A}^{i} \\). Subsequently, we calculate the nearest distance \\( e_i^m \\) from the remaining samples belonging to the \\( i_{th} \\) cluster center \\( C_R^i \\) of race \\( R \\) to the other races' cluster centers. These distances are then sorted in rising order to obtain a new Euclidean distance set \\( E_R^i \\). After this, we will follow Algorithm 1 to identify the final selected set of polysemy representative features, \\( F_{\\hat{R}P}' \\), and samples list, \\( S_{\\hat{R}P} \\). Through the aforementioned selection strategy, the chosen samples not only exhibit characteristics of the race to which they belong but also possess features similar to other races. This demonstrates the polysemy of the selected samples. As illustrated in Fig. 2 (b), we use Asian samples as an example to demonstrate our selection process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.703, + 0.224, + 0.72 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.282, + 0.745 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We use NVIDIA RTX 3090Ti GPUs with 24GB of memory for all experiments. We utilize the official code repository of the CLIP model [30]. We employ the pre-trained ViT-B/16 model with a batch size of 32 and use the Adam optimizer [18]. The initial learning rate is set to 1e-7, and the total number of iterations is 20,000. For segmentation model in the ADA process, following previous works [40, 41], we use the TransUNet [7] as the backbone and set the batch size to 2. The total number of iterations: 40,000 for training on the source domain and" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.924, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.155, + 0.923, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.218, + 0.922, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.28, + 0.66, + 0.288 + ], + "angle": 0, + "content": "GT Baseline" + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.28, + 0.812, + 0.288 + ], + "angle": 0, + "content": "STDR" + }, + { + "type": "image_caption", + "bbox": [ + 0.869, + 0.28, + 0.893, + 0.288 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.292, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Figure 4. Visual comparisons of our method and representative methods. Baseline: cross-domain results without fine-tuning; FIA [41]: fairness-aware DA approach; STDR [47]: ADA approach. Samples from top to bottom: Asian, Black, White subgroups, with numerical results on each image. Please zoom in for the best view." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.905, + 0.409 + ], + "angle": 0, + "content": "20,000 for fine-tuning on the target domain. The selection percentage of \\(\\alpha \\%\\) is set to \\(5\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.41, + 0.906, + 0.501 + ], + "angle": 0, + "content": "Benchmark Dataset. We employ the cross-domain medical segmentation benchmark of the Harvard-FairDomain dataset [41], comprising 10,000 multi-modal samples with sensitive attributes (eg., gender, race, ethnicity). Its rich demographic attributes and paired modalities make it suitable for assessing and quantifying cross-domain model fairness." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.502, + 0.906, + 0.578 + ], + "angle": 0, + "content": "Evaluation Metrics. We quantify fairness in medical image segmentation using metrics from previous studies [40, 41]: Dice, IoU, and the key composite metrics, ES-Dice and ES-IoU, to assess fairness alongside performance. For more details, please refer to the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.59, + 0.806, + 0.606 + ], + "angle": 0, + "content": "4.2. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.764 + ], + "angle": 0, + "content": "Following common practices in prior ADA works [45, 47], we comprehensively compare our approach with established DA methods and the latest ADA techniques. Our analysis primarily focused on fairness and performance comparisons across three sensitive attributes: Race, Ethnicity, and Gender (presented in Table 1, 3 and supplementary materials). Visualization examples of our method's results on racial attributes, compared to other representative methods (Baseline, FIA [41] and STDR [47]), are shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "With SOTA DA Methods. We perform a comprehensive comparison with previous state-of-the-art DA methods (including PixMatch [28], DAFormer [12], and DAFormerFIA [41]) across three demographic attributes. Tables 1, & 3 prove that our method significantly outperforms these approaches across nearly all metrics. Taking the Race attribute (in Table 1) as an example, FairAP significantly outperforms the DA methods for all metrics. This reflects a stable improvement compared to the previous state-of-the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24107" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.117 + ], + "angle": 0, + "content": "Table 1. Cup and rim segmentation performance on the FairDomain-Segmentation benchmark using different DA and ADA methods with Race as the demographic attribute. The * denotes \\( p \\)-value \\( < 0.05 \\) in all paired \\( t \\)-test, indicating statistically significant differences." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.118, + 0.872, + 0.433 + ], + "angle": 0, + "content": "
MethodVenueOverall ES-Dice↑Overall Dice↑Overall ES-IoU↑Overall IoU↑Asian Dice↑Black Dice↑White Dice↑Asian IoU↑Black IoU↑White IoU↑
DACupBaseline (Source)-0.8680.8880.7840.8080.8810.9010.8860.8020.8300.805
Baseline (Target)-0.6730.7000.5350.5640.7030.7310.6940.5670.6060.555
RimBaseline (Source)-0.8200.8610.7160.7620.8410.8380.8680.7370.7310.771
Baseline (Target)-0.4810.4950.3340.3420.4960.4720.4990.3470.3260.345
DACupPixMatch [28]CVPR'210.7390.7750.6140.6600.7520.7930.7830.6330.693
DAFormer [12]CVPR'220.7850.8040.6690.6910.7940.8160.8020.6820.7120.688
DAFormer-FIA [41]ECCV'240.7960.8100.6820.7000.8040.8210.8090.6940.7170.697
RimPixMatch [28]CVPR'210.6270.6730.4930.5230.6530.6400.6930.5220.4980.559
DAFormer [12]CVPR'220.3260.3320.1990.2020.3340.3450.3290.2040.2120.199
DAFormer-FIA [41]ECCV'240.5210.5310.3630.3690.5220.5390.5300.3640.3800.368
CupRandom-0.8130.8340.7160.7340.8140.8310.8370.7120.7350.737
Entropy [44]CVPR'190.8140.8230.7080.7210.8140.8250.8240.7110.7280.720
MHPL [45]CVPR'230.8090.8340.7080.7330.8100.8300.8380.7020.7330.737
DML-Core [43]NeurIPS'240.8110.8390.7170.7400.8120.8350.8420.7120.7420.743
Detective [61]CVPR'240.8140.8370.7170.7400.8140.8410.8390.7180.7480.741
STDR [47]IEEE TMI'240.8130.8370.7190.7380.8140.8330.8400.7150.7370.741
Our FairAP-0.828*0.843*0.731*0.747*0.827*0.8430.845*0.730*0.7510.748*
RimRandom-0.6870.7800.5670.6540.7330.7120.8000.6000.5770.676
Entropy [44]CVPR'190.6860.7720.5650.6430.7270.7090.7900.5950.5730.662
MHPL [45]CVPR'230.6880.7820.5670.6560.7290.7180.8010.5960.5810.678
DML-Core [43]NeurIPS'240.6970.7830.5780.6580.7380.7220.8010.6090.5890.677
Detective [61]CVPR'240.7000.7840.5820.6590.7400.7250.8010.6120.5930.678
STDR [47]IEEE TMI'240.6880.7810.5700.6560.7330.7130.8000.6040.5790.677
Our FairAP-0.711*0.791*0.592*0.667*0.750*0.735*0.807*0.624*0.602*0.686*
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.447, + 0.482, + 0.627 + ], + "angle": 0, + "content": "art DA models. For the rim segmentation, previous DA methods bring very limited performance improvements, and some (DAFormer [12]) even decline. For instance, as indicated in Table 1, the Baseline ES-IoU and IoU scores are 0.334 and 0.342 respectively, whereas DAFormer-FIA [41] achieved only 0.363 and 0.369. In contrast, our approach significantly outperforms them, achieving 0.592 and 0.667, respectively, which is approximately \\(20\\%\\) and \\(30\\%\\) higher. Overall, the previous fairness-aware DA method achieves fairness but with limited accuracy gains, while our FairAP ensures fairness and substantially boosts performance, underscoring its superiority." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.63, + 0.483, + 0.901 + ], + "angle": 0, + "content": "With SOTA ADA Methods. We further perform extensive comparison of these attributes against the ADA methods: Random (randomly select samples), Entropy [44], MHPL [45], DML-Core [43], Detective [61], and STDR [47], with the same labeling quota for a fair comparison. The experimental results indicate that our approach surpasses existing ADA methods in segmentation performance and demonstrates significant improvements in fairness. For instance, when analyzing the Race attribute (in Table 1), our method's overall Dice score of 0.791 outperforms the STDR method's 0.781 by 1 percentage with statistically significant differences. However, when considering both performance and fairness, our ES-Dice score of 0.711 exceeds STDR's 0.688 by 2.3 percentages, demonstrating FairAP's ability to significantly enhance fairness while maintaining performance. These findings are also reflected in Table 1 and Table 3, which assess Race and Ethnicity attributes. Compared to other ADA methods, our approach not only excels in seg" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.445, + 0.904, + 0.472 + ], + "angle": 0, + "content": "Table 2. Ablation studies for VLM-learning on sensitive attributes. The * denotes \\( p \\)-value \\( < 0.05 \\) in all paired \\( t \\)-test." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.473, + 0.9, + 0.56 + ], + "angle": 0, + "content": "
AttributesMethodCUPRim
Fair-APVLMES-Dice ↑ES-IoU ↑ES-Dice ↑ES-IoU ↑
Race✓0.8110.7110.6940.575
✓✓0.828*0.731*0.711*0.592*
Ethnicity✓0.8210.7190.7810.660
✓✓0.836*0.730*0.784*0.663*
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.905, + 0.604 + ], + "angle": 0, + "content": "mentation performance but also demonstrates effective improvements in fairness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.619, + 0.68, + 0.634 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.643, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Ablation Study on VLM-Learning. We conducted ablation studies to evaluate the impact of VLM-learning on model performance. In Table 2, we compared Fair-AP outcomes with and without VLM-learning under two sensitive attributes. The results demonstrate that VLM-learning significantly improves segmentation performance (\\(p\\)-value \\(< 0.05\\)). This statistically significant enhancement validates the effectiveness of VLM-learning in representing demographic attributes through cross-modal vision-language alignment, thereby enabling equitable performance across diverse subgroups." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Ablation Study on Fair-AP. We conducted ablation experiments to assess the impact of Fair-AP's design, focusing on four specific configurations: 1) Fair-Base, which evenly allocates quotas across subgroups (like White, Black, and Asian) and selects samples randomly. 2) Fair-Attr, which selects samples based on Attribute representativeness. 3)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24108" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 3. Cup and rim segmentation performance on the FairDomain-Segmentation benchmark using different DA and ADA methods with Ethnicity as the demographic attribute. The * denotes \\( p \\) -value \\( < {0.05} \\) in all paired \\( t \\) -test,indicating statistically significant differences." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.12, + 0.864, + 0.426 + ], + "angle": 0, + "content": "
MethodVenueOverall ES-Dice↑Overall Dice↑Overall ES-IoU↑Overall IoU↑Hispanic Dice↑Non-Hispanic Dice↑Hispanic IoU↑Non-Hispanic IoU↑
DACupBaseline (Source)-0.8710.8880.7840.8080.8870.9060.8070.839
Baseline (Target)-0.6840.7000.5510.5550.6990.7220.5630.587
RimBaseline (Source)-0.8450.8610.7430.7620.8600.8790.7610.786
Baseline (Target)-0.4890.4950.3400.3420.4950.5060.3420.348
DAPixMatch [28]CVPR'210.7400.7750.6280.6600.7630.8110.6490.699
DAFormer [12]CVPR'220.7730.7960.6540.6800.7940.8240.6770.717
DAFormer-FIA [41]ECCV'240.7900.8100.6740.7000.8080.8340.6970.734
RimPixMatch [28]CVPR'210.6430.6730.4990.5230.6730.7190.5130.562
DAFormer [12]CVPR'220.4050.4180.2640.2700.4160.4480.2690.295
DAFormer-FIA [41]ECCV'240.5160.5310.3590.3690.5290.5580.3680.395
CupRandom-0.8160.8340.7110.7340.8530.8310.7640.731
Entropy [44]CVPR'190.8010.8230.6930.7210.8480.8210.7580.717
MHPL [45]CVPR'230.8230.8340.7180.7330.8460.8320.7520.731
DML-Core [43]NeurIPS'240.8240.8390.7200.7400.8550.8360.7660.737
Detective [61]CVPR'240.8220.8370.7220.7400.8540.8360.7630.738
STDR [47]IEEE TMI'240.8120.8370.7100.7380.8650.8350.7740.735
ADAOur FairAP-0.836*0.8400.730*0.743*0.8440.8390.7590.741*
RimRandom-0.7750.7800.6510.6540.7860.7790.6570.653
Entropy [44]CVPR'190.7610.7720.6350.6430.7850.7700.6540.641
MHPL [45]CVPR'230.7740.7820.6500.6560.7900.7800.6620.654
DML-Core [43]NeurIPS'240.7750.7830.6530.6580.7910.7810.6620.655
Detective [61]CVPR'240.7770.7840.6560.6590.7920.7820.6620.658
STDR [47]IEEE TMI'240.7750.7810.6520.6560.7880.7790.6600.654
Our FairAP-0.784*0.790*0.663*0.668*0.796*0.788*0.673*0.666*
" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.436, + 0.475, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.682 + ], + "angle": 0, + "content": "Figure 5. Top: bar charts showing Dice-related (left) and IoU-related (right) scores across four ablation models with race attribute. Bottom: radar charts showing Dice (left) and IoU (right) values for Cup and Rim regions, divided by racial subgroups." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Fair-Poly, which chooses samples for Polysemy representativeness. 4) Our Fair-AP, which considers the average division of Fair-Base and prioritizes sample selection with both Attribute and Polysemy. Fig. 5 demonstrates the ablation study results. Standalone Fair-Attr or Fair-Poly strategies may be one-sided and show unstable improvements, while our FairAP—jointly modeling Attribute and Polysemy achieves significant gains in all metrics. The radar chart further confirms this across specific sub-groups, demonstrating our design's effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Exploration on Labeling Quota \\(\\alpha\\%\\). We investigate the impact of varying labeling quotas \\(\\alpha\\%\\) on Fair-AP's performance compared to some representative ADA methods. As" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.436, + 0.709, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.435, + 0.898, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.55, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Figure 6. Ablation studies of different labeling quota \\(\\alpha \\%\\) for Race attribute on the key composite metric ES-Dice." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.907, + 0.693 + ], + "angle": 0, + "content": "shown in Fig. 6, increasing \\(\\alpha\\%\\) improves model accuracy with diminishing returns: significant gains emerge from \\(0\\%, \\rightarrow, 2\\%, \\rightarrow, 5\\%\\), followed by marginal improvements beyond \\(5\\%\\). Crucially, our method consistently achieves superior performance across all quotas, justifying \\(\\alpha = 5\\%\\) as the optimal accuracy-cost balance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.72, + 0.634, + 0.736 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we pioneer a novel fairness-aware ADA paradigm to address fairness in real-world cross-domain medical scenarios. By utilizing the VLM to perceive demographic attributes and devising the attribute-aware FairAP strategy, we significantly enhance both fairness and segmentation performance. Extensive experiments confirm that our approach surpasses existing DA and ADA methods, demonstrating superior performance and enhanced fairness, providing a fresh perspective on addressing fairness challenges toward more equitable AI-assisted healthcare." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24109" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "This work is supported by the Guangdong Science and Technology Department (2024ZDZX2004), a General Research Fund of Hong Kong Research Grants Council (project no. 15218521), and a grant under the scheme of Collaborative Research with World-leading Research Groups in The Hong Kong Polytechnic University (project no. G-SACF)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.237, + 0.188, + 0.254 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.262, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[1] Ehsan Adeli, Qingyu Zhao, Adolf Pfefferbaum, Edith V Sullivan, Li Fei-Fei, Juan Carlos Niebles, and Kilian M Pohl. Representation learning with statistical independence to mitigate bias. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2513–2523, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.347, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[2] David Arthur and Sergei Vassilvitskii. k-means++: The advantages of careful seeding. Technical report, Stanford, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.39, + 0.484, + 0.458 + ], + "angle": 0, + "content": "[3] Jonathan T Barron. A generalization of otsu's method and minimum error thresholding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 455-470. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.461, + 0.484, + 0.531 + ], + "angle": 0, + "content": "[4] Peter J Bevan and Amir Atapour-Abarghoueiet. Detecting melanoma fairly: Skin tone detection and debiasing for skin lesion classification. In MICCAI Workshop on Domain Adaptation and Representation Transfer, pages 1-11. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[5] Alceu Bissoto, Michel Fornaciali, Eduardo Valle, and Sandra Avila. (de) constructing bias on skin lesion datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.484, + 0.629 + ], + "angle": 0, + "content": "[6] Simon Caton and Christian Haas. Fairness in machine learning: A survey. ACM Computing Surveys, 56(7):1-38, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[7] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[8] Jiayi Chen, Benteng Ma, Hengfei Cui, and Yong Xia. Think twice before selection: Federated evidential active learning for medical image analysis with domain shifts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11439-11449, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[9] Julia Dressel and Hany Farid. The accuracy, fairness, and limits of predicting recidivism. Science advances, 4(1): eaao5580, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[10] Judy Wawira Gichoya, Imon Banerjee, Ananth Reddy Bhimireddy, John L Burns, Leo Anthony Celi, Li-Ching Chen, Ramon Correa, Natalie Dullerud, Marzyeh Ghassemi, Shih-Cheng Huang, et al. Ai recognition of patient race in medical imaging: a modelling study. The Lancet Digital Health, 4(6):e406-e414, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.262, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[11] Yuanbo Guo, Zheyu Yan, Xiaoting Yu, Qingpeng Kong, Joy Xie, Kevin Luo, Dewen Zeng, Yawen Wu, Zhenge Jia, and Yiyu Shi. Hardware design and the fairness of a neural network. Nature Electronics, 7(8):714-723, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.908, + 0.218 + ], + "angle": 0, + "content": "[12] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9924-9935, 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.907, + 0.288 + ], + "angle": 0, + "content": "[13] Mohamed Huti, Tiarna Lee, Elinor Sawyer, and Andrew P King. An investigation into race bias in random forest models based on breast dce-mri derived radiomics features. In Workshop on Clinical Image-Based Procedures, pages 225-234. Springer, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.342 + ], + "angle": 0, + "content": "[14] Suyog Dutt Jain and Kristen Grauman. Active image segmentation propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2864-2873, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[15] Ruinan Jin, Zikang Xu, Yuan Zhong, Qionsong Yao, Qi Dou, S Kevin Zhou, and Xiaoxiao Li. Fairmedfm: fairness benchmarking for medical imaging foundation models. arXiv preprint arXiv:2407.00983, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.906, + 0.469 + ], + "angle": 0, + "content": "[16] Thorsten Kalb, Kaisar Kushibar, Celia Cintas, Karim Lekadir, Oliver Diaz, and Richard Osuala. Revisiting skin tone fairness in dermatological lesion classification. In Workshop on Clinical Image-Based Procedures, pages 246-255. Springer, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[17] Michael P Kim, Amirata Ghorbani, and James Zou. Multiaccuracy: Black-box post-processing for fairness in classification. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 247-254, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[18] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.906, + 0.637 + ], + "angle": 0, + "content": "[19] Newton M Kinyanjui, Timothy Odonga, Celia Cintas, Noel CF Codella, Rameswar Panda, Prasanna Sattigeri, and Kush R Varshney. Fairness of classifiers across skin tones in dermatology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 320-329. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.906, + 0.719 + ], + "angle": 0, + "content": "[20] Qingpeng Kong, Ching-Hao Chiu, Dewen Zeng, Yu-Jen Chen, Tsung-Yi Ho, Jingtong Hu, and Yiyu Shi. Achieving fairness through channel pruning for dermatological disease diagnosis. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 24-34. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.906, + 0.762 + ], + "angle": 0, + "content": "[21] David D Lewis and Jason Catlett. Heterogeneous uncertainty sampling for supervised learning. In Machine learning proceedings 1994, pages 148-156. Elsevier, 1994. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[22] Pranay K Lohia, Karthikeyan Natesan Ramamurthy, Manish Bhide, Diptikalyan Saha, Kush R Varshney, and Ruchir Puri. Bias mitigation post-processing for individual and group fairness. In Icassp 2019-2019 iee international conference on acoustics, speech and signal processing (icassp), pages 2847-2851. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[23] Xiangde Luo, Hongqiu Wang, Jinfeng Xu, Lu Li, Yue Zhao, Yuan He, Hui Huang, Jianghong Xiao, Tao Song, Shichuan Zhang, et al. Generalizable magnetic resonance imaging-based nasopharyngeal carcinoma delineation: Bridging gaps" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24110" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "across multiple centers and raters with active learning. International Journal of Radiation Oncology* Biology* Physics, 121(5):1384-1393, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[24] Yan Luo, Min Shi, Muhammad Osama Khan, Muhammad Muneeb Afzal, Hao Huang, Shuaihang Yuan, Yu Tian, Luo Song, Ava Kouhana, Tobias Elze, et al. Fairclip: Harnessing fairness in vision-language learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12289-12301, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[25] Yan Luo, Yu Tian, Min Shi, Louis R Pasquale, Lucy Q Shen, Nazlee Zebardast, Tobias Elze, and Mengyu Wang. Harvard glaucoma fairness: a retinal nerve disease dataset for fairness learning and fair identity normalization. IEEE Transactions on Medical Imaging, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[26] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM computing surveys (CSUR), 54 (6):1-35, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[27] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 415-430. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[28] Luke Melas-Kyriazi and Arjun K Manrai. Pixmatch: Unsupervised domain adaptation via pixelwise consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12435-12445, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.565 + ], + "angle": 0, + "content": "[29] Esther Puyol-Anton, Bram Ruijsink, Jorge Mariscal Harana, Stefan K Piechnik, Stefan Neubauer, Steffen E Petersen, Reza Razavi, Phil Chowienczyk, and Andrew P King. Fairness in cardiac magnetic resonance imaging: assessing sex and racial bias in deep learning-based segmentation. Frontiers in cardiovascular medicine, 9:859310, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[31] Fernanda Ribeiro, Valentina Shumovskaia, Thomas Davies, and Ira Ktena. How fair is your graph? exploring fairness concerns in neuroimaging studies. In Machine Learning for Healthcare Conference, pages 459-478. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[32] María Agustina Ricci Lara, Rodrigo Echeveste, and Enzo Ferrante. Addressing fairness in artificial intelligence for medical imaging. nature communications, 13(1):4581, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[33] Zohaib Salahuddin, Yi Chen, Xian Zhong, Henry C Woodruff, Nastaran Mohammadian Rad, Shruti Atul Mali, and Philippe Lambin. From head and neck tumour and lymph node segmentation to survival prediction on pet/ct: an end-to-end framework featuring uncertainty, fairness, and multi-region multi-modal radiomics. Cancers, 15(7):1932, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[34] Candice Schwartz, Ifeanyi Beverly Chukwudozie, Silvia Tejeda, Ganga Vijayasiri, Ivy Abraham, Mylene Remo, Hiral A Shah, Maria Rojas, Alicia Carillo, Loraine Moreno," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "et al. Association of population screening for breast cancer risk with use of mammography among women in medically underserved racial and ethnic minority groups. JAMA network open, 4(9):e2123751-e2123751, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[35] Laleh Seyyed-Kalantari, Guanxiong Liu, Matthew McDermott, Irene Y Chen, and Marzyeh Ghassemi. Chexclusion: Fairness gaps in deep chest x-ray classifiers. In BIOCOMPUTING 2021: proceedings of the Pacific symposium, pages 232-243. World Scientific, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[36] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA McDermott, Irene Y Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12):2176-2182, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.907, + 0.328 + ], + "angle": 0, + "content": "[37] Dinggang Shen, Guorong Wu, and Heung-II Suk. Deep learning in medical image analysis. Annual review of biomedical engineering, 19(1):221-248, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[38] Min Shi, Muhammad Muneeb Afzal, Hao Huang, Congcong Wen, Yan Luo, Muhammad Osama Khan, Yu Tian, Leo Kim, Tobias Elze, Yi Fang, et al. Equitable deep learning for diabetic retinopathy detection using multi-dimensional retinal imaging with fair adaptive scaling: a retrospective study. medRxiv, pages 2024-04, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[39] Emma AM Stanley, Matthias Wilms, Pauline Mouches, and Nils D Forkert. Fairness-related performance and explainability effects in deep learning models for brain image analysis. Journal of Medical Imaging, 9(6):061102-061102, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[40] Yu Tian, Min Shi, Yan Luo, Ava Kouhana, Tobias Elze, and Mengyu Wang. Fairseg: A large-scale medical image segmentation dataset for fairness learning using segment anything model with fair error-bound scaling. In ICLR, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[41] Yu Tian, Congcong Wen, Min Shi, Muhammad Muneeb Afzal, Hao Huang, Muhammad Osama Khan, Yan Luo, Yi Fang, and Mengyu Wang. Fairdomain: Achieving fairness in cross-domain medical image segmentation and classification. In European Conference on Computer Vision, pages 251-271. Springer, 2024. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[42] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7472-7481, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[43] Arvind Vepa, Zukang Yang, Andrew Choi, Jungseock Joo, Fabien Scalzo, and Yizhou Sun. Integrating deep metric learning with coreset for active learning in 3d segmentation. Advances in Neural Information Processing Systems, 37:71643-71671, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[44] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[45] Fan Wang, Zhongyi Han, Zhiyan Zhang, Rundong He, and Yilong Yin. Mhpl: Minimum happy points learning for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24111" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "active source free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20008-20018, 2023. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[46] Hongqiu Wang, Yueming Jin, and Lei Zhu. Dynamic interactive relation capturing via scene graph learning for robotic surgical report generation. arXiv preprint arXiv:2306.02651, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[47] Hongqiu Wang, Jian Chen, Shichen Zhang, Yuan He, Jinfeng Xu, Mengwan Wu, Jinlan He, Wenjun Liao, and Xiangde Luo. Dual-reference source-free active domain adaptation for nasopharyngeal carcinoma tumor segmentation across multiple hospitals. IEEE Transactions on Medical Imaging, 2024. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[48] Hongqiu Wang, Xiangde Luo, Wu Chen, Qingqing Tang, Mei Xin, Qiong Wang, and Lei Zhu. Advancing uwf-slo vessel segmentation with source-free active domain adaptation and a novel multi-center dataset. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 75-85. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[49] Hongqiu Wang, Zhaohu Xing, Weitong Wu, Yijun Yang, Qingqing Tang, Meixia Zhang, Yanwu Xu, and Lei Zhu. Non-invasive to invasive: Enhancing ffa synthesis from cfp with a benchmark dataset and a novel network. In Proceedings of the 1st International Workshop on Multimedia Computing for Health and Medicine, pages 7-15, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[50] Hongqiu Wang, Guang Yang, Shichen Zhang, Jing Qin, Yike Guo, Bo Xu, Yueming Jin, and Lei Zhu. Video-instrument synergistic network for referring video instrument segmentation in robotic surgery. IEEE Transactions on Medical Imaging, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[51] Hongqiu Wang, Yixian Chen, Wu Chen, Huihui Xu, Haoyu Zhao, Bin Sheng, Huazhu Fu, Guang Yang, and Lei Zhu. Serp-mamba: Advancing high-resolution retinal vessel segmentation with selective state-space model. IEEE Transactions on Medical Imaging, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[52] Tianxiang Wang and Qun Dai. A patch distribution-based active learning method for multiple instance alzheimer's disease diagnosis. Pattern Recognition, 150:110341, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[53] Zhibo Wang, Xiaowei Dong, Henry Xue, Zhifei Zhang, Weifeng Chiu, Tao Wei, and Kui Ren. Fairness-aware adversarial perturbation towards bias mitigation for deployed deep models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10379-10388, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[54] Garrett Wilson and Diane J Cook. A survey of unsupervised deep domain adaptation. ACM Transactions on Intelligent Systems and Technology (TIST), 11(5):1-46, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[55] Wenshuai Xu, Zhenghui Hu, Yu Lu, Jinzhou Meng, Qingjie Liu, and Yunhong Wang. Activedc: Distribution calibration for active finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16996-17005, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[56] Zhao Xu, Kai Yu, Volker Tresp, Xiaowei Xu, and Jizhi Wang. Representative sampling for text classification using support vector machines. In Advances in Information Retrieval: 25th European Conference on IR Research, ECIR 2003, Pisa, Italy, April 14-16, 2003. Proceedings 25, pages 393-407. Springer, 2003. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[57] Zikang Xu, Jun Li, Qingsong Yao, Han Li, Mingyue Zhao, and S Kevin Zhou. Addressing fairness issues in deep learning-based medical image analysis: a systematic review. npj Digital Medicine, 7(1):286, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[58] Jenny Yang, Andrew AS Soltan, David W Eyre, Yang Yang, and David A Clifton. An adversarial training framework for mitigating algorithmic biases in clinical machine learning. NPJ digital medicine, 6(1):55, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[59] Ruichen Yao, Ziteng Cui, Xiaoxiao Li, and Lin Gu. Improving fairness in image classification via sketching. arXiv preprint arXiv:2211.00168, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[60] Chenxi Yuan, Kristin A Linn, and Rebecca A Hubbard. Algorithmic fairness of machine learning models for alzheimer disease progression. JAMA Network Open, 6(11):e2342203-e2342203, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[61] Wenqiao Zhang, Zheqi Lv, Hao Zhou, Jia-Wei Liu, Juncheng Li, Mengze Li, Yunfei Li, Dongping Zhang, Yueting Zhuang, and Siliang Tang. Revisiting the domain shift and sample uncertainty in multi-source active domain transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16751-16761, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[62] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[63] Haoyu Zhao, Hao Wang, Xingyue Zhao, Hongqiu Wang, Zhiyu Wu, Chengjiang Long, and Hua Zou. Automated 3d physical simulation of open-world scene with gaussian splatt-ing. arXiv e-prints, pages arXiv-2411, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[64] Qingyu Zhao, Ehsan Adeli, and Kilian M Pohl. Training confounder-free deep learning models for medical applications. Nature communications, 11(1):6010, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[65] Yongshuo Zong, Yongxin Yang, and Timothy Hospedales. Medfair: Benchmarking fairness for medical imaging. arXiv preprint arXiv:2210.01725, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.656 + ], + "angle": 0, + "content": "[66] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24112" + } + ] +] \ No newline at end of file diff --git a/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/0cab628c-59eb-4cba-b70b-2c418218511f_origin.pdf b/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/0cab628c-59eb-4cba-b70b-2c418218511f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6a8571615b552460bdf2e10b629f119694c425db --- /dev/null +++ b/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/0cab628c-59eb-4cba-b70b-2c418218511f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c3d4149cc32474d82fe6701040ceadadd30d7f9646d8884128d0e83e5f4c331b +size 6517850 diff --git a/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/full.md b/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a101b6a83fd7df579d2e81ac3bce2fa4db9e599e --- /dev/null +++ b/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/full.md @@ -0,0 +1,341 @@ +![](images/166718c4d156e9768c5b6b9751f77486acf30a91676fed38e1ffe982abe83859.jpg) + +# Toward Fair and Accurate Cross-Domain Medical Image Segmentation: A VLM-Driven Active Domain Adaptation Paradigm + +Hongqiu Wang $^{1}$ Wu Chen $^{2}$ Xiangde Luo $^{3}$ Zhaohu Xing $^{1}$ Lihao Liu $^{4}$ Jing Qin $^{5}$ Shaozhi Wu $^{2*}$ Lei Zhu $^{1,6*}$ + +$^{1}$ Hong Kong University of Science and Technology (Guangzhou) $^{2}$ University of Electronic Science and Technology of China $^{3}$ West China Hospital $^{4}$ Amazon $^{5}$ Hong Kong Polytechnic University $^{6}$ Hong Kong University of Science and Technology + +# Abstract + +Fairness in AI-assisted medical image analysis is crucial for equitable healthcare, but is often neglected, especially in prevalent cross-domain scenarios (diverse demographics and imaging protocols). Effective and equitable deployment of AI models in these scenarios is critical, yet traditional Unsupervised Domain Adaptation (UDA) methods exhibit limited improvements. Emerging Active Domain Adaptation (ADA) approaches offer more effective enhancements, but all ignore fairness issues. Therefore, in this work, we propose the first fairness-aware ADA paradigm that simultaneously achieves both enhanced fairness and superior overall performance. Our method leverages the multimodal alignment capability of Vision-Language Models (VLMs): By performing medical images (vision) and sensitive attributes (language) learning, VLM inherently captures semantic correlations between visual features and protected attributes, enabling explicit attributes representation. Building on this foundation, we further devise an attribute-aware strategy (FairAP), which dynamically adapts to diverse patient demographics to promote equitable and high-quality outcomes by considering both Attribute and Polysemy. Extensive experiments on the FairDomain benchmark demonstrate that our method significantly reduces bias and maintains leading performance, outperforming existing UDA and ADA methods. This work pioneers a VLM-driven ADA paradigm for fair cross-domain medical segmentation, offering a blueprint for effective and equitable AI deployment in clinical practice. Code will be on (https://github.com/whq-xxh/Fair-AP). + +# 1. Introduction + +"Fairness is not an attitude. It's a professional skill that must be developed and exercised." — Brit Hume. + +![](images/e3bf325cc1380e29fe850a634f2f1ebad8c401afb6b0a29e792f318a49298ac2.jpg) +Figure 1. Comparison of previous cross-domain fairness methods with our VLM-Driven fairness-aware ADA paradigm. Our method combines VLM and ADA to achieve both leading model performance and enhanced fairness. + +The advent of AI technologies has significantly transformed medical image analysis (MedIA), providing sophisticated tools for analyzing complex biological patterns within imaging data [37, 46, 49-51]. These advanced algorithms enhance diagnostic precision, support informed clinical decisions, and improve patient outcomes. However, as these models become increasingly integrated into clinical practice, critical challenges have emerged: inherent biases within these models [24, 32, 57] and domain shifts caused by diverse clinical scenarios [23, 41, 54]. + +Numerous studies have reported that biases in MedIA tasks [13, 16, 19, 29, 31, 33, 35, 39, 60], stemming from training datasets failing to represent diverse populations adequately, raise profound ethical and fairness concerns [10, 34]. These biases cause severe health inequities, impacting marginalized groups in clinical applications. For example, Black individuals face nearly twice the risk of glaucoma than others, highlighting the need for accurately segmenting the optic disc and cup for early diagnosis, yet AI models often perform worst in this subgroup. This issue is particularly critical in cross-domain scenarios, where models deployed across institutions encounter heterogeneous data distributions that amplify performance disparities [36, 40, 41]. Such domain-specific biases not only undermine the per + +formance of AI systems but also violate fairness. Thus, ensuring both efficient and fair adaptation in cross-domain real-world medical AI environments is crucial to translate advanced technology to equitable healthcare for all. + +Recently, significant efforts have been made to tackle fairness across various fields [1, 3, 4, 9, 11, 15, 20, 24, 29, 38-40, 58, 62, 64, 65]. These studies explore various attributes across different modalities or body parts. However, real-world cross-domain challenges remain underexplored. While Tian et al. [41] established the multiattribute benchmark dataset to evaluate cross-domain fairness, the limited performance of its Unsupervised Domain Adaptation (UDA) framework (e.g., IoU values $< 40\%$ in Tables 1 & 3) severely restricts clinical applicability despite its fairness-aware design. Although emerging Active Domain Adaptation (ADA) has demonstrated significant performance gains over UDA by actively selecting valuable target-domain samples [45], its fairness implications remain largely unexplored. This creates a critical gap in achieving both equitable outcomes and clinically viable performance across demographic subgroups, as current methods force an undesirable trade-off between the fairness guarantees and the model effectiveness. + +Therefore, in this work, we propose the first fairness-aware ADA paradigm to simultaneously enhance fairness across all subgroups and boost overall performance for cross-domain medical image segmentation scenario applications. Specifically, according to the unique characteristics of medical images and corresponding sensitive attribute information (such as race: Black, White, Asian), we present a novel perspective that integrates the Vision-Language Model (VLM) to align image features with sensitive attributes. This alignment endows our model with the capability to perceive the sensitive attribute representations within images accurately. This VLM-attribute learning is essential as it establishes a solid foundation for subsequent representative attributes and polysemy sample selection. Then, we devise an attribute-aware FairAP sample selection strategy, where we select a quite small (only $5\%$ here) but critical set of samples and annotate them for target fair model fine-tuning. The FairAP strategy allocates equal quotas to each subgroup to ensure balanced attention and leverages VLM attribute-latent space representations for attribute-representative and polysemic sample selection. Our process and its key distinctions from previous methods are visually demonstrated in Fig. 1. It is worth noting that the FairAP strategy is highly practical as it significantly reduces the workload of doctors, while simultaneously achieving both fairness and high performance (even improving IoU by around $30\%$ than FairDA method [41] for rim segmentation as shown in Table 1). Comprehensive experiments on the FairDomain benchmark [41] confirm that our approach markedly diminishes bias and retains cutting + +edge performance in segmentation tasks, surpassing traditional DA and ADA methods. + +Our main contributions can be summarized as follows: + +- To the best of our knowledge, we are the first to tackle fairness challenges in cross-domain medical image segmentation with ADA paradigm, dynamically adapting to diverse patient subgroups for equitable outcomes. +- We propose to leverage the Vision-Language Model to seamlessly integrate demographic attributes like race or gender into the learning process, endowing the model with the ability to perceive these sensitive attributes. +- We devise the attribute-aware FairAP strategy to ensure both fairness and effectiveness. FairAP optimizes representation by balancing samples from underrepresented demographic groups while maintaining diversity to enhance overall performance comprehensively. +- Extensive experiments demonstrate that our method surpasses previous DA and ADA approaches by ensuring superior fairness across multiple protected attributes and maintaining top performance. Our method pioneers a novel solution paradigm for fairness and effectiveness in cross-domain medical AI model deployment. + +# 2. Related Works + +# 2.1. Fairness in Medical Imaging + +In medical image processing, fairness aims to reduce model biases to ensure equitable outcomes across patient subgroups. Existing methods fall into three main categories: pre-processing, in-processing, and post-processing [6, 26, 57]. Pre-processing techniques, such as those proposed by Bissoto et al. [5] and others [24, 29, 59], aim to correct training data imbalances, but may affect computational efficiency. In-processing methods mitigate bias during training by incorporating fairness-aware loss functions [40] or fairness-specific modules [20, 25, 41], which can potentially trade off some accuracy. On the other hand, post-processing approaches, like those introduced by Wang et al. [53], Kim et al. [17], and others [22], adjust the model outputs to improve fairness, which may struggle to generalize well during testing. However, few studies focus on fairness in cross-domain scenarios within real-world deployment environments. + +# 2.2. Domain Adaptation + +Unsupervised Domain Adaptation (UDA): UDA trains on the labeled source and unlabeled target data to bridge domain gaps, typically using adversarial training and minimizing inter-domain discrepancies [54]. In UDA methods, adversarial learning minimizes distributional differences between source and target domains, enhancing performance in the target domain [42, 44]. Additionally, self-training in UDA uses pseudo-labels from target domain data to en + +![](images/08f6d3ad504d8ea729a41efa9bb0fef3fff8a084b9552ce8ed6366d6073d7e8e.jpg) +Figure 2. Overview of the Fair VLM-Driven ADA method (taking race attribute for illustration). Given the target domain images $x^T$ and their corresponding sensitive attribute information $\text{attrs}$ , we first align them through VLM-attribute learning to obtain the vision-based sensitive attribute features $F_{attrR}$ . Then, we devise the attribute-aware FairAP Select strategy to select $\alpha \%$ of the samples that are representations of Attribute and Polysemy. These selected samples $S$ are then annotated by experts to obtain the annotations $y^{TL}$ . Finally, active domain adaptation fine-tuning is performed to obtain the target fair model $h_{\phi}'$ for top performance and enhanced fairness. + +hance performance in the target domain [12, 27, 66]. Building on the work by Hoyer et al. [12], Tian et al. [41] further explored fairness in cross-domain settings by integrating the fairness module. This new setting can more accurately reflect real-world clinical scenarios and provide a clear quantification of the performance and fairness of cross-domain methods [41]. However, very limited performance gains brought by UDA significantly diminish its clinical utility. + +Active Domain Adaptation (ADA): ADA combines active learning and domain adaptation to reduce annotation costs and sustain high performance by focusing on annotating the most valuable samples through strategies like uncertainty-based selection and diversity-based sampling [8, 14, 21, 48, 56]. For example, Wang et al. [45], and others [47, 52, 55, 61] enhanced cross-domain performance by strategically selecting target domain samples for model development. However, these current ADA methods have predominantly concentrated on optimizing overall model performance, neglecting the crucial aspect of fairness. Consequently, we are the pioneers in introducing a novel fairness-aware ADA paradigm, effectively minimizing annotation costs while simultaneously achieving both enhanced fairness and leading performance. + +# 3. Methodology + +# 3.1. Problem Setting + +Cross-domain MedIA often exhibits not only conventional performance degradation for domain shifts but also issues of unfairness related to sensitive attributes. Fair ADA is an effective and novel approach to address this challenge. Our work specifically targets fairness challenges in domain adaptation, focusing on fairness issues across multiple sensitive attributes in diverse subgroups. The problem is defined as follows: In medical segmentation tasks, consider the source domain data $D_{S} = \{(x_{i}^{S},y_{i}^{S},\text{attr}_{i})\}_{i = 1}^{N_{S}}$ , where $x_{i}^{S}\in \mathbb{R}^{H\times W}$ represents the $i_{th}$ image, the $y_{i}^{S}\in \mathbb{R}^{C\times H\times W}$ denotes the corresponding label with $C$ classes, and the $\text{attr}_{i}\in \mathbb{R}^{n}$ means the sensitive attributes related to the sample, such as gender, race, or ethnicity. Meanwhile, there is an unlabeled dataset $D_{T} = \{(x_{i}^{T},\text{attr}_{i})\}_{i = 1}^{N_{T}}$ whose domain vary, leading to different image features from $D_{S}$ . In the fair segmentation setting, note that the target attributes $\text{attr}_{i}$ are available and accessible, as outlined in [40, 41]. The aim is to design strategies for selecting a small number of samples $S = \alpha \% \cdot D_{T}\ll D_{T}$ for labeling and active learning, thereby maximizing the fairness + +and performance of model $h_{\phi}$ as much as possible while maintaining $\alpha \%$ low. + +# 3.2. Overview of the Fair-ADA + +The overall structure of our proposed VLM-driven fairness-aware ADA framework is illustrated in Fig. 2. The entire methodology section uses race attribute $attrR$ as an example to illustrate. In the VLM-attribute Learning stage (a), we utilize a Vision-Language model $g_{\theta}$ to align image samples with sensitive attribute information within the latent space. The model $g_{\theta}$ performs vision (unlabeled medical images $x^{T}$ ) language (corresponding attribute information $attrR$ ) learning to achieve alignment. Then, we can extract vision-based sensitive attribute features $F_{attrR}$ . These feature representations consist of multiple subgroups corresponding to various attributes. For example, sensitive racial features $F_{attrR}$ include three subgroup features such as Asian features $F_{a}$ , Black features $F_{b}$ , and White features $F_{w}$ . The entire process can be expressed as follows: + +$$ +F _ {a t t r R} = g _ {\theta} \left(x ^ {T}, a t t r R\right) = \left\{F _ {a}, F _ {b}, F _ {w} \right\}. \tag {1} +$$ + +Based on the vision-based sensitive attribute features $F_{attrR}$ , the process moves into the FairAP Select stage (b), where our approach is organized into three main steps. In the first step, we ensure and amplify fairness within our active learning strategy by allocating an equal quota of samples from each sensitive attribute subcategory. In the second step, we employ clustering methods for each subgroup based on various sensitive attribute representations, followed by selecting attribute representative samples. In the third step, we further optimize the selected samples by expanding the selection to include polysemy representatives. This overall process can be summarized as: + +$$ +S = \text {F a i r} A P \left(F _ {\text {a t t r} R}\right). \tag {2} +$$ + +In the Active Domain Adaptation stage (c), we utilize the model $h_{\phi}$ that has been previously trained on the source domain data $D_{S}$ . We then fine-tune this model using the sample set $S$ selected by our FairAP strategy to produce the optimized target fair model $h_{\phi}^{\prime}$ . The loss function for this model is defined by the following formula: + +$$ +L _ {A D A} \left(h _ {\phi} ^ {\prime} (x ^ {T L}), y ^ {T L}\right) = - \sum_ {c = 1} ^ {C} y _ {c} \log \left(h _ {\phi} ^ {\prime} (x)\right), \tag {3} +$$ + +where $C$ represents the number of classes, and selected list $S = \{x^{TL},y^{TL}\}$ , whose size corresponds to $\alpha \%$ of the total dataset. + +# 3.3. VLM-attribute Learning + +Traditional fairness-aware segmentation methods [40, 41] using the simplistic categorical encoding (e.g., 0/1 for gender male or female) suffer from semantic sparsity and contextual blindness, which fails to capture attribute-visual + +![](images/c646274ca138d0181a38b3873a00f629cd9bb42ab340cce4a970902c44a0c6d2.jpg) +VLM-attribute learning +Figure 3. Visualization of racial feature distribution before and after feature alignment via VLM-attribute learning. Our alignment process creates a denser distribution of features within racial subgroups, improving distinguishability across different races. + +correlations. These disjoint representations limit ADA paradigms' ability to perceive subgroup distributions during cross-domain adaptation. To address this, we propose the VLM-based attribute-visual alignment to enable context-aware perception and provide continuous semantic gradients for precise subgroup-aware adaptation. + +Fig. 2 (a) provides a detailed illustration of our vision language learning process. Specifically, we utilize a pretrained vision-language model (CLIP [30, 63] here) denoted as $g_{\theta}$ to process the samples $x^{T}$ along with their associated sensitive attribute information $\text{attr } R$ . The model's vision encoder, $g_{\theta_v}^v$ , extracts image tokens $z_{v}$ from the input $x^{T}$ with parameters $\theta_{v}$ . Meanwhile, the language encoder, $g_{\theta_l}^l$ , derives attribute tokens $z_{l}$ from the input $\text{attr } R$ with parameters $\theta_{l}$ . Therefore, we can express the model as $g_{\theta} = \{g_{\theta_v}^v,g_\theta_l^l\}$ with the parameters as $\theta = \{\theta_v,\theta_l\}$ . When training the model $g_{\theta}$ using a batch of $N$ image-attribute token pairs, we generate a similarity matrix $A\in \mathbb{R}^{N\times N}$ . Each entry of the matrix, $A_{ij} = z_{vi}^{T}z_{lj}$ , represents the cosine similarity between the $i_{th}$ image token $z_{vi}$ and $j_{th}$ attribute token $z_{lj}$ within the batch. The model $g_{\theta}$ employs a contrastive learning approach that aims to maximize the cosine similarity for positive image-attribute token pairs while minimizing it for negative token pairs. The model $g_{\theta}$ loss function can be expressed as follows: + +$$ +L _ {V L M} = - \frac {1}{N} \sum_ {i = 1} ^ {N} \frac {1}{2} \log \left(\frac {\exp (2 A _ {i i})}{S _ {i} T _ {i}}\right), \tag {4} +$$ + +where $S_{i} = \sum_{j=1}^{N} \exp(A_{ij})$ and $T_{i} = \sum_{j=1}^{N} \exp(A_{ji})$ . Since our focus is on enhancing the vision-based attribute perception capabilities, we choose to freeze the language encoder $g_{\theta}^{l}$ to better train $g_{\theta}$ as follows: + +$$ +A ^ {\prime} \leftarrow A, +$$ + +$$ +\theta_ {v} ^ {\prime} \leftarrow \theta_ {v} - \eta \frac {\partial L _ {V L M}}{\partial \theta_ {v}}, \quad \theta_ {l} ^ {\prime} = \theta_ {l}, \tag {5} +$$ + +where $\eta$ is the learning rate. We employ a consistent template for sensitive attribute information as: + +A photo of a fundus and the race is $\langle \mathbf{RACE}\rangle$ + +The $\langle \mathbf{RACE}\rangle$ represents specific racial information. After completing the training, we utilize a fixed vision encoder $g_{\theta_v'}^v$ to process all target domain images $x^T$ , allowing extraction of the vision-based sensitive attribute features $F_{attrR}$ as following: + +$$ +F _ {a t t r R} = g _ {\theta_ {v} ^ {\prime}} ^ {v} \left(x ^ {T}\right). \tag {6} +$$ + +The visualization of racial characteristic distribution before and after our VLM-attribute learning is shown in Fig. 3. + +# 3.4. FairAP Strategy + +Building on the VLM-attribute learning described earlier, sensitive vision-based attribute features $F_{attrR}$ are successfully extracted. For clarity, we continue using sensitive racial features $F_{attrR}$ as an example to illustrate the Fair Attribute and Polysemy representation (FairAP) Select strategy. The sensitive racial features $F_{attrR}$ encompass three types of racial characteristics: Asian features $F_{a}$ , Black features $F_{b}$ , and White features $F_{w}$ . Thus, we define $F_{attrR} = \{F_{a}, F_{b}, F_{w}\}$ . Our FairAP Select strategy encompasses three critical components: Fair Quota Select, Attribute Representation Select, and Polysemy Representation Select. We will introduce each component in sequence. + +Fair Quota Selection. We recognize the data-driven nature of AI. Therefore, to achieve our fairness objective ('treat all our subgroups with fairness'), we ensure a consistent sample quota for each racial subgroup. We define the sample quota for each subgroup as follows: + +$$ +\Delta_ {a} = \Delta_ {b} = \Delta_ {w} = \Delta_ {R}, \tag {7} +$$ + +where $\Delta_{a}$ , $\Delta_{b}$ , and $\Delta_{w}$ represent the quotas given to Asians, Blacks, and Whites respectively. The $\Delta_{R}$ is calculated by dividing the overall quota $(\alpha \% \cdot D_T)$ by the number of total racial subgroups, which is three here. Upon completing the evaluations of the subsequent components (Attribute and Polysemy), which assist in determining the specific samples, we compile the final selection list for each racial subgroup, represented as $S_{R} = \{S_{\hat{R}A}, S_{\hat{R}P}\}$ . Finally, we combine the selection lists $S_{R}$ of all racial subgroups and submitted them to doctors for annotations. + +Attribute Representation Selection. To optimally select representative samples for each racial subgroup, we undertake clustering operation (k-means++ algorithm [2] here) to process the feature set $F_{R}$ , resulting in the clustered features $F_{R}^{\prime}$ and the corresponding cluster centers $C_{R}^{N_{R}}$ , where $R$ represents a specific racial group (like Asian, Black, or White). This process is represented as follows: + +$$ +\left\{F _ {R} ^ {\prime}, C _ {R} ^ {N _ {R}} \right\} = \operatorname {C l u s t e r} \left(F _ {R}\right). \tag {8} +$$ + +In this equation, $C_R^{N_R}$ denotes the cluster centers, and $N_R$ indicates the number of cluster centers for the racial sub + +Algorithm 1: POLYSEMY-REP SELECT +Input: The $i_{th}$ cluster center of subgroup $R$ : $C_R^i$ , distance set: $D_R^i$ , selected distance set: $D_{\hat{R}A}^i$ , features: $F_{R,m}''$ , number of $F_{R,m}': M_i$ , other racial subgroups' cluster centers: $C_{\bar{R}}^k$ , quota: $\Delta_R$ , subgroup cluster number: $N_R$ . Output: Polysemy-Rep selective racial list: $S_{\hat{R}P}$ +1 for $i = 1$ to $N_R$ do +2 ${D_R'} \gets {D_R} \backslash {D_{\hat{R}A}}^i // \text{Remaining distances}$ +3 $F_{R,m}'' \gets \text{Index} \left( D_R^{i'} \right) // \text{Indexing features}$ +4 $N_i \gets M_i - \frac{\Delta_R}{2N_R} // \text{Adjust selection range}$ +5 Initialize empty list $E_{temp}$ ; +6 for $m = 1$ to $N_i$ do +7 if $F_{R,m}'' \in C_R^i \land F_{R,m}'' \notin C_{\bar{R}}^k$ then +8 $\begin{array}{c} e_i^m \gets \min \left( \| F_{R,m}' - C_{\bar{R}}^k \| \right) // \text{Min distance to other racial clusters} \\ \text{Append } e_i^m \text{ to } E_{temp}; \end{array}$ +9 end +10 end +11 Sort $E_{temp}$ in rising order; +12 $E_{\hat{R}P}^i \gets \text{top min distances from } E_{temp}$ ; +14 end +15 $F_{\hat{R}P}' \gets \text{Index} \left( \{E_{\hat{R}P}', ..., E_{\hat{R}P}', ..., E_{\hat{R}P}^{N_R}\} \right)$ +16 $S_{\hat{R}P} \gets F_{\hat{R}P}' // \text{Get the final Polysemy-Rep racial subgroup list}$ + +group. For each racial subgroup, the number of cluster centers $C_R^{N_R}$ is calculated based on their overall distribution in the dataset to comprehensively represent the target domain. This is illustrated by the following formula: + +$$ +N _ {a}: N _ {b}: N _ {w} \approx \operatorname {S u m} _ {a}: \operatorname {S u m} _ {b}: \operatorname {S u m} _ {w}, \tag {9} +$$ + +where $\mathrm{Sum}_a$ , $\mathrm{Sum}_b$ , and $\mathrm{Sum}_w$ denote the proportions of Asians, Blacks, and Whites, respectively. For the $i_{th}$ cluster center $C_R^i$ of race $R$ , we define a set of Euclidean distances $D_R^i$ , where each element $d_i^m$ in the set is the distance from sample $F_{R,m}^{\prime}$ ( $m_{th}$ sample belongs to the $i_{th}$ cluster) to the cluster center $C_R^i$ . These distances are then sorted in ascending order. This can be represented as: + +$$ +\begin{array}{l} D _ {R} ^ {i} = \operatorname {s o r t} \left(\left\{d _ {i} ^ {m} = \| F _ {R, m} ^ {\prime} - C _ {R} ^ {i} \| \mid F _ {R, m} ^ {\prime} \in C _ {R} ^ {i} \right\}\right) \\ = \left\{d _ {i} ^ {1}, d _ {i} ^ {2}, \dots , d _ {i} ^ {m}, \dots , d _ {i} ^ {M _ {i}} \right\}, \tag {10} \\ \end{array} +$$ + +where $d_i^1 \leq d_i^2 \leq \dots \leq d_i^m \leq \dots \leq d_i^{M_i}$ , and $M_{i}$ is the number of samples $F_{R}^{\prime}$ that belong to the cluster center $C_R^i$ . To ensure that the selected attribute representatives $F_{\hat{R}A}$ can comprehensively cover the entire distribution of the dataset, + +we establish an equal selection quota of $\frac{\Delta_R}{2N_R}$ for each cluster. Specifically, we select the top $\frac{\Delta_R}{2N_R}$ samples, denoted as the subset $D_{\hat{R}A}^i$ , corresponding to the smallest Euclidean distances from the sorted set $D_R^i$ . The set $D_{\hat{R}A}$ represents the distances selected from the $N_R$ clusters of the current racial subgroup. This can be expressed as: + +$$ +D _ {R \hat {A}} ^ {i} = \left\{d _ {R} ^ {1}, d _ {R} ^ {2}, \dots , d _ {R} ^ {\Delta_ {R} / 2 N _ {R}} \right\} \tag {11} +$$ + +$$ +D _ {\hat {R A}} = \left\{D _ {\hat {R A}} ^ {1}, D _ {\hat {R A}} ^ {2}, \dots , D _ {\hat {R A}} ^ {N _ {R}} \right\}. \tag {12} +$$ + +Using the indices from the set $D_{\hat{R}A}$ that correspond to the smallest distances, we can derive the feature set $F_{\hat{R}A}^{\prime}$ for attribute representative selection and obtain the corresponding sample set $S_{\hat{R}A}$ . + +Polysemy Representation Selection. Given that representative samples for each racial subgroup have been selected in the previous phase, we focus on choosing diverse samples in the Polysemy phase to explore a wider range of expressions within a limited quota, thus enhancing performance. For example, as seen in the right (post-VLM learning) of Fig. 3, some samples within the Black clusters are actually from White individuals, which may embody characteristics of both White and similar attributes of Black. + +In detail, We first filter out the samples selected in the previous stage, as indicated by $D_{R}^{i'} = D_{R}^{i} \setminus D_{\hat{R}A}^{i}$ . Subsequently, we calculate the nearest distance $e_i^m$ from the remaining samples belonging to the $i_{th}$ cluster center $C_R^i$ of race $R$ to the other races' cluster centers. These distances are then sorted in rising order to obtain a new Euclidean distance set $E_R^i$ . After this, we will follow Algorithm 1 to identify the final selected set of polysemy representative features, $F_{\hat{R}P}'$ , and samples list, $S_{\hat{R}P}$ . Through the aforementioned selection strategy, the chosen samples not only exhibit characteristics of the race to which they belong but also possess features similar to other races. This demonstrates the polysemy of the selected samples. As illustrated in Fig. 2 (b), we use Asian samples as an example to demonstrate our selection process. + +# 4. Experiments + +# 4.1. Experimental Setup + +Implementation Details. We use NVIDIA RTX 3090Ti GPUs with 24GB of memory for all experiments. We utilize the official code repository of the CLIP model [30]. We employ the pre-trained ViT-B/16 model with a batch size of 32 and use the Adam optimizer [18]. The initial learning rate is set to 1e-7, and the total number of iterations is 20,000. For segmentation model in the ADA process, following previous works [40, 41], we use the TransUNet [7] as the backbone and set the batch size to 2. The total number of iterations: 40,000 for training on the source domain and + +![](images/a6532aa695cdeba009e1aff484cb50abec73ad11a36104a210a20af7243db4d4.jpg) + +![](images/9cc126768d95ab511ea64d42ebca6c2784b6dd866963ccfda36deab6701a9414.jpg) + +![](images/fc0a8e1fd698224b24fb0ece4235844e4de92df43211d5bb8d182cc92538eace.jpg) +GT Baseline +STDR +Ours +Figure 4. Visual comparisons of our method and representative methods. Baseline: cross-domain results without fine-tuning; FIA [41]: fairness-aware DA approach; STDR [47]: ADA approach. Samples from top to bottom: Asian, Black, White subgroups, with numerical results on each image. Please zoom in for the best view. + +20,000 for fine-tuning on the target domain. The selection percentage of $\alpha \%$ is set to $5\%$ . + +Benchmark Dataset. We employ the cross-domain medical segmentation benchmark of the Harvard-FairDomain dataset [41], comprising 10,000 multi-modal samples with sensitive attributes (eg., gender, race, ethnicity). Its rich demographic attributes and paired modalities make it suitable for assessing and quantifying cross-domain model fairness. + +Evaluation Metrics. We quantify fairness in medical image segmentation using metrics from previous studies [40, 41]: Dice, IoU, and the key composite metrics, ES-Dice and ES-IoU, to assess fairness alongside performance. For more details, please refer to the supplementary materials. + +# 4.2. Comparison with SOTA Methods + +Following common practices in prior ADA works [45, 47], we comprehensively compare our approach with established DA methods and the latest ADA techniques. Our analysis primarily focused on fairness and performance comparisons across three sensitive attributes: Race, Ethnicity, and Gender (presented in Table 1, 3 and supplementary materials). Visualization examples of our method's results on racial attributes, compared to other representative methods (Baseline, FIA [41] and STDR [47]), are shown in Fig. 4. + +With SOTA DA Methods. We perform a comprehensive comparison with previous state-of-the-art DA methods (including PixMatch [28], DAFormer [12], and DAFormerFIA [41]) across three demographic attributes. Tables 1, & 3 prove that our method significantly outperforms these approaches across nearly all metrics. Taking the Race attribute (in Table 1) as an example, FairAP significantly outperforms the DA methods for all metrics. This reflects a stable improvement compared to the previous state-of-the + +Table 1. Cup and rim segmentation performance on the FairDomain-Segmentation benchmark using different DA and ADA methods with Race as the demographic attribute. The * denotes $p$ -value $< 0.05$ in all paired $t$ -test, indicating statistically significant differences. + +
MethodVenueOverall ES-Dice↑Overall Dice↑Overall ES-IoU↑Overall IoU↑Asian Dice↑Black Dice↑White Dice↑Asian IoU↑Black IoU↑White IoU↑
DACupBaseline (Source)-0.8680.8880.7840.8080.8810.9010.8860.8020.8300.805
Baseline (Target)-0.6730.7000.5350.5640.7030.7310.6940.5670.6060.555
RimBaseline (Source)-0.8200.8610.7160.7620.8410.8380.8680.7370.7310.771
Baseline (Target)-0.4810.4950.3340.3420.4960.4720.4990.3470.3260.345
DACupPixMatch [28]CVPR'210.7390.7750.6140.6600.7520.7930.7830.6330.693
DAFormer [12]CVPR'220.7850.8040.6690.6910.7940.8160.8020.6820.7120.688
DAFormer-FIA [41]ECCV'240.7960.8100.6820.7000.8040.8210.8090.6940.7170.697
RimPixMatch [28]CVPR'210.6270.6730.4930.5230.6530.6400.6930.5220.4980.559
DAFormer [12]CVPR'220.3260.3320.1990.2020.3340.3450.3290.2040.2120.199
DAFormer-FIA [41]ECCV'240.5210.5310.3630.3690.5220.5390.5300.3640.3800.368
CupRandom-0.8130.8340.7160.7340.8140.8310.8370.7120.7350.737
Entropy [44]CVPR'190.8140.8230.7080.7210.8140.8250.8240.7110.7280.720
MHPL [45]CVPR'230.8090.8340.7080.7330.8100.8300.8380.7020.7330.737
DML-Core [43]NeurIPS'240.8110.8390.7170.7400.8120.8350.8420.7120.7420.743
Detective [61]CVPR'240.8140.8370.7170.7400.8140.8410.8390.7180.7480.741
STDR [47]IEEE TMI'240.8130.8370.7190.7380.8140.8330.8400.7150.7370.741
Our FairAP-0.828*0.843*0.731*0.747*0.827*0.8430.845*0.730*0.7510.748*
RimRandom-0.6870.7800.5670.6540.7330.7120.8000.6000.5770.676
Entropy [44]CVPR'190.6860.7720.5650.6430.7270.7090.7900.5950.5730.662
MHPL [45]CVPR'230.6880.7820.5670.6560.7290.7180.8010.5960.5810.678
DML-Core [43]NeurIPS'240.6970.7830.5780.6580.7380.7220.8010.6090.5890.677
Detective [61]CVPR'240.7000.7840.5820.6590.7400.7250.8010.6120.5930.678
STDR [47]IEEE TMI'240.6880.7810.5700.6560.7330.7130.8000.6040.5790.677
Our FairAP-0.711*0.791*0.592*0.667*0.750*0.735*0.807*0.624*0.602*0.686*
+ +art DA models. For the rim segmentation, previous DA methods bring very limited performance improvements, and some (DAFormer [12]) even decline. For instance, as indicated in Table 1, the Baseline ES-IoU and IoU scores are 0.334 and 0.342 respectively, whereas DAFormer-FIA [41] achieved only 0.363 and 0.369. In contrast, our approach significantly outperforms them, achieving 0.592 and 0.667, respectively, which is approximately $20\%$ and $30\%$ higher. Overall, the previous fairness-aware DA method achieves fairness but with limited accuracy gains, while our FairAP ensures fairness and substantially boosts performance, underscoring its superiority. + +With SOTA ADA Methods. We further perform extensive comparison of these attributes against the ADA methods: Random (randomly select samples), Entropy [44], MHPL [45], DML-Core [43], Detective [61], and STDR [47], with the same labeling quota for a fair comparison. The experimental results indicate that our approach surpasses existing ADA methods in segmentation performance and demonstrates significant improvements in fairness. For instance, when analyzing the Race attribute (in Table 1), our method's overall Dice score of 0.791 outperforms the STDR method's 0.781 by 1 percentage with statistically significant differences. However, when considering both performance and fairness, our ES-Dice score of 0.711 exceeds STDR's 0.688 by 2.3 percentages, demonstrating FairAP's ability to significantly enhance fairness while maintaining performance. These findings are also reflected in Table 1 and Table 3, which assess Race and Ethnicity attributes. Compared to other ADA methods, our approach not only excels in seg + +Table 2. Ablation studies for VLM-learning on sensitive attributes. The * denotes $p$ -value $< 0.05$ in all paired $t$ -test. + +
AttributesMethodCUPRim
Fair-APVLMES-Dice ↑ES-IoU ↑ES-Dice ↑ES-IoU ↑
Race✓0.8110.7110.6940.575
✓✓0.828*0.731*0.711*0.592*
Ethnicity✓0.8210.7190.7810.660
✓✓0.836*0.730*0.784*0.663*
+ +mentation performance but also demonstrates effective improvements in fairness. + +# 4.3. Ablation Studies + +Ablation Study on VLM-Learning. We conducted ablation studies to evaluate the impact of VLM-learning on model performance. In Table 2, we compared Fair-AP outcomes with and without VLM-learning under two sensitive attributes. The results demonstrate that VLM-learning significantly improves segmentation performance ( $p$ -value $< 0.05$ ). This statistically significant enhancement validates the effectiveness of VLM-learning in representing demographic attributes through cross-modal vision-language alignment, thereby enabling equitable performance across diverse subgroups. + +Ablation Study on Fair-AP. We conducted ablation experiments to assess the impact of Fair-AP's design, focusing on four specific configurations: 1) Fair-Base, which evenly allocates quotas across subgroups (like White, Black, and Asian) and selects samples randomly. 2) Fair-Attr, which selects samples based on Attribute representativeness. 3) + +Table 3. Cup and rim segmentation performance on the FairDomain-Segmentation benchmark using different DA and ADA methods with Ethnicity as the demographic attribute. The * denotes $p$ -value $< {0.05}$ in all paired $t$ -test,indicating statistically significant differences. + +
MethodVenueOverall ES-Dice↑Overall Dice↑Overall ES-IoU↑Overall IoU↑Hispanic Dice↑Non-Hispanic Dice↑Hispanic IoU↑Non-Hispanic IoU↑
DACupBaseline (Source)-0.8710.8880.7840.8080.8870.9060.8070.839
Baseline (Target)-0.6840.7000.5510.5550.6990.7220.5630.587
RimBaseline (Source)-0.8450.8610.7430.7620.8600.8790.7610.786
Baseline (Target)-0.4890.4950.3400.3420.4950.5060.3420.348
DAPixMatch [28]CVPR'210.7400.7750.6280.6600.7630.8110.6490.699
DAFormer [12]CVPR'220.7730.7960.6540.6800.7940.8240.6770.717
DAFormer-FIA [41]ECCV'240.7900.8100.6740.7000.8080.8340.6970.734
RimPixMatch [28]CVPR'210.6430.6730.4990.5230.6730.7190.5130.562
DAFormer [12]CVPR'220.4050.4180.2640.2700.4160.4480.2690.295
DAFormer-FIA [41]ECCV'240.5160.5310.3590.3690.5290.5580.3680.395
CupRandom-0.8160.8340.7110.7340.8530.8310.7640.731
Entropy [44]CVPR'190.8010.8230.6930.7210.8480.8210.7580.717
MHPL [45]CVPR'230.8230.8340.7180.7330.8460.8320.7520.731
DML-Core [43]NeurIPS'240.8240.8390.7200.7400.8550.8360.7660.737
Detective [61]CVPR'240.8220.8370.7220.7400.8540.8360.7630.738
STDR [47]IEEE TMI'240.8120.8370.7100.7380.8650.8350.7740.735
ADAOur FairAP-0.836*0.8400.730*0.743*0.8440.8390.7590.741*
RimRandom-0.7750.7800.6510.6540.7860.7790.6570.653
Entropy [44]CVPR'190.7610.7720.6350.6430.7850.7700.6540.641
MHPL [45]CVPR'230.7740.7820.6500.6560.7900.7800.6620.654
DML-Core [43]NeurIPS'240.7750.7830.6530.6580.7910.7810.6620.655
Detective [61]CVPR'240.7770.7840.6560.6590.7920.7820.6620.658
STDR [47]IEEE TMI'240.7750.7810.6520.6560.7880.7790.6600.654
Our FairAP-0.784*0.790*0.663*0.668*0.796*0.788*0.673*0.666*
+ +![](images/4a6e528e7aa9a6b94e0bd9cadec8eb436fbd6d7488ca5899858ecb8d9fe42336.jpg) +Figure 5. Top: bar charts showing Dice-related (left) and IoU-related (right) scores across four ablation models with race attribute. Bottom: radar charts showing Dice (left) and IoU (right) values for Cup and Rim regions, divided by racial subgroups. + +Fair-Poly, which chooses samples for Polysemy representativeness. 4) Our Fair-AP, which considers the average division of Fair-Base and prioritizes sample selection with both Attribute and Polysemy. Fig. 5 demonstrates the ablation study results. Standalone Fair-Attr or Fair-Poly strategies may be one-sided and show unstable improvements, while our FairAP—jointly modeling Attribute and Polysemy achieves significant gains in all metrics. The radar chart further confirms this across specific sub-groups, demonstrating our design's effectiveness. + +Exploration on Labeling Quota $\alpha\%$ . We investigate the impact of varying labeling quotas $\alpha\%$ on Fair-AP's performance compared to some representative ADA methods. As + +![](images/c866706a14acbd9849e5d095d4eeae3e49c18305200092a06c2e727a642e1b8c.jpg) +Figure 6. Ablation studies of different labeling quota $\alpha \%$ for Race attribute on the key composite metric ES-Dice. + +![](images/1c0ec4b432d0d1e94021d5fd0d55a46ddddad311130ef71b2c8c9656e595d294.jpg) + +shown in Fig. 6, increasing $\alpha\%$ improves model accuracy with diminishing returns: significant gains emerge from $0\%, \rightarrow, 2\%, \rightarrow, 5\%$ , followed by marginal improvements beyond $5\%$ . Crucially, our method consistently achieves superior performance across all quotas, justifying $\alpha = 5\%$ as the optimal accuracy-cost balance. + +# 5. Conclusion + +In this work, we pioneer a novel fairness-aware ADA paradigm to address fairness in real-world cross-domain medical scenarios. By utilizing the VLM to perceive demographic attributes and devising the attribute-aware FairAP strategy, we significantly enhance both fairness and segmentation performance. Extensive experiments confirm that our approach surpasses existing DA and ADA methods, demonstrating superior performance and enhanced fairness, providing a fresh perspective on addressing fairness challenges toward more equitable AI-assisted healthcare. + +# Acknowledgments + +This work is supported by the Guangdong Science and Technology Department (2024ZDZX2004), a General Research Fund of Hong Kong Research Grants Council (project no. 15218521), and a grant under the scheme of Collaborative Research with World-leading Research Groups in The Hong Kong Polytechnic University (project no. G-SACF). + +# References + +[1] Ehsan Adeli, Qingyu Zhao, Adolf Pfefferbaum, Edith V Sullivan, Li Fei-Fei, Juan Carlos Niebles, and Kilian M Pohl. Representation learning with statistical independence to mitigate bias. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2513–2523, 2021. 2 +[2] David Arthur and Sergei Vassilvitskii. k-means++: The advantages of careful seeding. Technical report, Stanford, 2006. 5 +[3] Jonathan T Barron. A generalization of otsu's method and minimum error thresholding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 455-470. Springer, 2020. 2 +[4] Peter J Bevan and Amir Atapour-Abarghoueiet. Detecting melanoma fairly: Skin tone detection and debiasing for skin lesion classification. In MICCAI Workshop on Domain Adaptation and Representation Transfer, pages 1-11. Springer, 2022. 2 +[5] Alceu Bissoto, Michel Fornaciali, Eduardo Valle, and Sandra Avila. (de) constructing bias on skin lesion datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 2 +[6] Simon Caton and Christian Haas. Fairness in machine learning: A survey. ACM Computing Surveys, 56(7):1-38, 2024. 2 +[7] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 6 +[8] Jiayi Chen, Benteng Ma, Hengfei Cui, and Yong Xia. Think twice before selection: Federated evidential active learning for medical image analysis with domain shifts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11439-11449, 2024. 3 +[9] Julia Dressel and Hany Farid. The accuracy, fairness, and limits of predicting recidivism. Science advances, 4(1): eaao5580, 2018. 2 +[10] Judy Wawira Gichoya, Imon Banerjee, Ananth Reddy Bhimireddy, John L Burns, Leo Anthony Celi, Li-Ching Chen, Ramon Correa, Natalie Dullerud, Marzyeh Ghassemi, Shih-Cheng Huang, et al. Ai recognition of patient race in medical imaging: a modelling study. The Lancet Digital Health, 4(6):e406-e414, 2022. 1 + +[11] Yuanbo Guo, Zheyu Yan, Xiaoting Yu, Qingpeng Kong, Joy Xie, Kevin Luo, Dewen Zeng, Yawen Wu, Zhenge Jia, and Yiyu Shi. Hardware design and the fairness of a neural network. Nature Electronics, 7(8):714-723, 2024. 2 +[12] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9924-9935, 2022. 3, 6, 7, 8 +[13] Mohamed Huti, Tiarna Lee, Elinor Sawyer, and Andrew P King. An investigation into race bias in random forest models based on breast dce-mri derived radiomics features. In Workshop on Clinical Image-Based Procedures, pages 225-234. Springer, 2023. 1 +[14] Suyog Dutt Jain and Kristen Grauman. Active image segmentation propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2864-2873, 2016. 3 +[15] Ruinan Jin, Zikang Xu, Yuan Zhong, Qionsong Yao, Qi Dou, S Kevin Zhou, and Xiaoxiao Li. Fairmedfm: fairness benchmarking for medical imaging foundation models. arXiv preprint arXiv:2407.00983, 2024. 2 +[16] Thorsten Kalb, Kaisar Kushibar, Celia Cintas, Karim Lekadir, Oliver Diaz, and Richard Osuala. Revisiting skin tone fairness in dermatological lesion classification. In Workshop on Clinical Image-Based Procedures, pages 246-255. Springer, 2023. 1 +[17] Michael P Kim, Amirata Ghorbani, and James Zou. Multiaccuracy: Black-box post-processing for fairness in classification. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 247-254, 2019. 2 +[18] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[19] Newton M Kinyanjui, Timothy Odonga, Celia Cintas, Noel CF Codella, Rameswar Panda, Prasanna Sattigeri, and Kush R Varshney. Fairness of classifiers across skin tones in dermatology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 320-329. Springer, 2020. 1 +[20] Qingpeng Kong, Ching-Hao Chiu, Dewen Zeng, Yu-Jen Chen, Tsung-Yi Ho, Jingtong Hu, and Yiyu Shi. Achieving fairness through channel pruning for dermatological disease diagnosis. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 24-34. Springer, 2024. 2 +[21] David D Lewis and Jason Catlett. Heterogeneous uncertainty sampling for supervised learning. In Machine learning proceedings 1994, pages 148-156. Elsevier, 1994. 3 +[22] Pranay K Lohia, Karthikeyan Natesan Ramamurthy, Manish Bhide, Diptikalyan Saha, Kush R Varshney, and Ruchir Puri. Bias mitigation post-processing for individual and group fairness. In Icassp 2019-2019 iee international conference on acoustics, speech and signal processing (icassp), pages 2847-2851. IEEE, 2019. 2 +[23] Xiangde Luo, Hongqiu Wang, Jinfeng Xu, Lu Li, Yue Zhao, Yuan He, Hui Huang, Jianghong Xiao, Tao Song, Shichuan Zhang, et al. Generalizable magnetic resonance imaging-based nasopharyngeal carcinoma delineation: Bridging gaps + +across multiple centers and raters with active learning. International Journal of Radiation Oncology* Biology* Physics, 121(5):1384-1393, 2025. 1 +[24] Yan Luo, Min Shi, Muhammad Osama Khan, Muhammad Muneeb Afzal, Hao Huang, Shuaihang Yuan, Yu Tian, Luo Song, Ava Kouhana, Tobias Elze, et al. Fairclip: Harnessing fairness in vision-language learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12289-12301, 2024. 1, 2 +[25] Yan Luo, Yu Tian, Min Shi, Louis R Pasquale, Lucy Q Shen, Nazlee Zebardast, Tobias Elze, and Mengyu Wang. Harvard glaucoma fairness: a retinal nerve disease dataset for fairness learning and fair identity normalization. IEEE Transactions on Medical Imaging, 2024. 2 +[26] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM computing surveys (CSUR), 54 (6):1-35, 2021. 2 +[27] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 415-430. Springer, 2020. 3 +[28] Luke Melas-Kyriazi and Arjun K Manrai. Pixmatch: Unsupervised domain adaptation via pixelwise consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12435-12445, 2021. 6, 7, 8 +[29] Esther Puyol-Anton, Bram Ruijsink, Jorge Mariscal Harana, Stefan K Piechnik, Stefan Neubauer, Steffen E Petersen, Reza Razavi, Phil Chowienczyk, and Andrew P King. Fairness in cardiac magnetic resonance imaging: assessing sex and racial bias in deep learning-based segmentation. Frontiers in cardiovascular medicine, 9:859310, 2022. 1, 2 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 6 +[31] Fernanda Ribeiro, Valentina Shumovskaia, Thomas Davies, and Ira Ktena. How fair is your graph? exploring fairness concerns in neuroimaging studies. In Machine Learning for Healthcare Conference, pages 459-478. PMLR, 2022. 1 +[32] María Agustina Ricci Lara, Rodrigo Echeveste, and Enzo Ferrante. Addressing fairness in artificial intelligence for medical imaging. nature communications, 13(1):4581, 2022. 1 +[33] Zohaib Salahuddin, Yi Chen, Xian Zhong, Henry C Woodruff, Nastaran Mohammadian Rad, Shruti Atul Mali, and Philippe Lambin. From head and neck tumour and lymph node segmentation to survival prediction on pet/ct: an end-to-end framework featuring uncertainty, fairness, and multi-region multi-modal radiomics. Cancers, 15(7):1932, 2023. 1 +[34] Candice Schwartz, Ifeanyi Beverly Chukwudozie, Silvia Tejeda, Ganga Vijayasiri, Ivy Abraham, Mylene Remo, Hiral A Shah, Maria Rojas, Alicia Carillo, Loraine Moreno, + +et al. Association of population screening for breast cancer risk with use of mammography among women in medically underserved racial and ethnic minority groups. JAMA network open, 4(9):e2123751-e2123751, 2021. 1 +[35] Laleh Seyyed-Kalantari, Guanxiong Liu, Matthew McDermott, Irene Y Chen, and Marzyeh Ghassemi. Chexclusion: Fairness gaps in deep chest x-ray classifiers. In BIOCOMPUTING 2021: proceedings of the Pacific symposium, pages 232-243. World Scientific, 2020. 1 +[36] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA McDermott, Irene Y Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12):2176-2182, 2021. 1 +[37] Dinggang Shen, Guorong Wu, and Heung-II Suk. Deep learning in medical image analysis. Annual review of biomedical engineering, 19(1):221-248, 2017. 1 +[38] Min Shi, Muhammad Muneeb Afzal, Hao Huang, Congcong Wen, Yan Luo, Muhammad Osama Khan, Yu Tian, Leo Kim, Tobias Elze, Yi Fang, et al. Equitable deep learning for diabetic retinopathy detection using multi-dimensional retinal imaging with fair adaptive scaling: a retrospective study. medRxiv, pages 2024-04, 2024. 2 +[39] Emma AM Stanley, Matthias Wilms, Pauline Mouches, and Nils D Forkert. Fairness-related performance and explainability effects in deep learning models for brain image analysis. Journal of Medical Imaging, 9(6):061102-061102, 2022. 1 +[40] Yu Tian, Min Shi, Yan Luo, Ava Kouhana, Tobias Elze, and Mengyu Wang. Fairseg: A large-scale medical image segmentation dataset for fairness learning using segment anything model with fair error-bound scaling. In ICLR, 2024. 1, 2, 3, 4, 6 +[41] Yu Tian, Congcong Wen, Min Shi, Muhammad Muneeb Afzal, Hao Huang, Muhammad Osama Khan, Yan Luo, Yi Fang, and Mengyu Wang. Fairdomain: Achieving fairness in cross-domain medical image segmentation and classification. In European Conference on Computer Vision, pages 251-271. Springer, 2024. 1, 2, 3, 4, 6, 7, 8 +[42] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7472-7481, 2018. 2 +[43] Arvind Vepa, Zukang Yang, Andrew Choi, Jungseock Joo, Fabien Scalzo, and Yizhou Sun. Integrating deep metric learning with coreset for active learning in 3d segmentation. Advances in Neural Information Processing Systems, 37:71643-71671, 2024. 7, 8 +[44] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 2, 7, 8 +[45] Fan Wang, Zhongyi Han, Zhiyan Zhang, Rundong He, and Yilong Yin. Mhpl: Minimum happy points learning for + +active source free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20008-20018, 2023. 2, 3, 6, 7, 8 +[46] Hongqiu Wang, Yueming Jin, and Lei Zhu. Dynamic interactive relation capturing via scene graph learning for robotic surgical report generation. arXiv preprint arXiv:2306.02651, 2023. 1 +[47] Hongqiu Wang, Jian Chen, Shichen Zhang, Yuan He, Jinfeng Xu, Mengwan Wu, Jinlan He, Wenjun Liao, and Xiangde Luo. Dual-reference source-free active domain adaptation for nasopharyngeal carcinoma tumor segmentation across multiple hospitals. IEEE Transactions on Medical Imaging, 2024. 3, 6, 7, 8 +[48] Hongqiu Wang, Xiangde Luo, Wu Chen, Qingqing Tang, Mei Xin, Qiong Wang, and Lei Zhu. Advancing uwf-slo vessel segmentation with source-free active domain adaptation and a novel multi-center dataset. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 75-85. Springer, 2024. 3 +[49] Hongqiu Wang, Zhaohu Xing, Weitong Wu, Yijun Yang, Qingqing Tang, Meixia Zhang, Yanwu Xu, and Lei Zhu. Non-invasive to invasive: Enhancing ffa synthesis from cfp with a benchmark dataset and a novel network. In Proceedings of the 1st International Workshop on Multimedia Computing for Health and Medicine, pages 7-15, 2024. 1 +[50] Hongqiu Wang, Guang Yang, Shichen Zhang, Jing Qin, Yike Guo, Bo Xu, Yueming Jin, and Lei Zhu. Video-instrument synergistic network for referring video instrument segmentation in robotic surgery. IEEE Transactions on Medical Imaging, 2024. +[51] Hongqiu Wang, Yixian Chen, Wu Chen, Huihui Xu, Haoyu Zhao, Bin Sheng, Huazhu Fu, Guang Yang, and Lei Zhu. Serp-mamba: Advancing high-resolution retinal vessel segmentation with selective state-space model. IEEE Transactions on Medical Imaging, 2025. 1 +[52] Tianxiang Wang and Qun Dai. A patch distribution-based active learning method for multiple instance alzheimer's disease diagnosis. Pattern Recognition, 150:110341, 2024. 3 +[53] Zhibo Wang, Xiaowei Dong, Henry Xue, Zhifei Zhang, Weifeng Chiu, Tao Wei, and Kui Ren. Fairness-aware adversarial perturbation towards bias mitigation for deployed deep models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10379-10388, 2022. 2 +[54] Garrett Wilson and Diane J Cook. A survey of unsupervised deep domain adaptation. ACM Transactions on Intelligent Systems and Technology (TIST), 11(5):1-46, 2020. 1, 2 +[55] Wenshuai Xu, Zhenghui Hu, Yu Lu, Jinzhou Meng, Qingjie Liu, and Yunhong Wang. Activedc: Distribution calibration for active finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16996-17005, 2024. 3 +[56] Zhao Xu, Kai Yu, Volker Tresp, Xiaowei Xu, and Jizhi Wang. Representative sampling for text classification using support vector machines. In Advances in Information Retrieval: 25th European Conference on IR Research, ECIR 2003, Pisa, Italy, April 14-16, 2003. Proceedings 25, pages 393-407. Springer, 2003. 3 + +[57] Zikang Xu, Jun Li, Qingsong Yao, Han Li, Mingyue Zhao, and S Kevin Zhou. Addressing fairness issues in deep learning-based medical image analysis: a systematic review. npj Digital Medicine, 7(1):286, 2024. 1, 2 +[58] Jenny Yang, Andrew AS Soltan, David W Eyre, Yang Yang, and David A Clifton. An adversarial training framework for mitigating algorithmic biases in clinical machine learning. NPJ digital medicine, 6(1):55, 2023. 2 +[59] Ruichen Yao, Ziteng Cui, Xiaoxiao Li, and Lin Gu. Improving fairness in image classification via sketching. arXiv preprint arXiv:2211.00168, 2022. 2 +[60] Chenxi Yuan, Kristin A Linn, and Rebecca A Hubbard. Algorithmic fairness of machine learning models for alzheimer disease progression. JAMA Network Open, 6(11):e2342203-e2342203, 2023. 1 +[61] Wenqiao Zhang, Zheqi Lv, Hao Zhou, Jia-Wei Liu, Juncheng Li, Mengze Li, Yunfei Li, Dongping Zhang, Yueting Zhuang, and Siliang Tang. Revisiting the domain shift and sample uncertainty in multi-source active domain transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16751-16761, 2024. 3, 7, 8 +[62] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. 2 +[63] Haoyu Zhao, Hao Wang, Xingyue Zhao, Hongqiu Wang, Zhiyu Wu, Chengjiang Long, and Hua Zou. Automated 3d physical simulation of open-world scene with gaussian splatt-ing. arXiv e-prints, pages arXiv-2411, 2024. 4 +[64] Qingyu Zhao, Ehsan Adeli, and Kilian M Pohl. Training confounder-free deep learning models for medical applications. Nature communications, 11(1):6010, 2020. 2 +[65] Yongshuo Zong, Yongxin Yang, and Timothy Hospedales. Medfair: Benchmarking fairness for medical imaging. arXiv preprint arXiv:2210.01725, 2022. 2 +[66] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 3 \ No newline at end of file diff --git a/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/images.zip b/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8bcf2ef87ae76d41c79f08d2cb6d91c53b12a109 --- /dev/null +++ b/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:953bd81aebd95dbb967fdd2d7727eee6c793afe8d19cc5af6f94639f876b3f3d +size 782209 diff --git a/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/layout.json b/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9fe98de1ab566c2ff31b510649fdd53830d96973 --- /dev/null +++ b/2025/Toward Fair and Accurate Cross-Domain Medical Image Segmentation_ A VLM-Driven Active Domain Adaptation Paradigm/layout.json @@ -0,0 +1,10849 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 68, + 100, + 97, + 125 + ], + "blocks": [ + { + "bbox": [ + 68, + 100, + 97, + 125 + ], + "lines": [ + { + "bbox": [ + 68, + 100, + 97, + 125 + ], + "spans": [ + { + "bbox": [ + 68, + 100, + 97, + 125 + ], + "type": "image", + "image_path": "166718c4d156e9768c5b6b9751f77486acf30a91676fed38e1ffe982abe83859.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 107, + 542, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 107, + 542, + 144 + ], + "spans": [ + { + "bbox": [ + 99, + 107, + 542, + 144 + ], + "type": "text", + "content": "Toward Fair and Accurate Cross-Domain Medical Image Segmentation: A VLM-Driven Active Domain Adaptation Paradigm" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "spans": [ + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "text", + "content": "Hongqiu Wang" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "text", + "content": " Wu Chen" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "text", + "content": " Xiangde Luo" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "text", + "content": " Zhaohu Xing" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "text", + "content": " Lihao Liu" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "text", + "content": " Jing Qin" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "text", + "content": " Shaozhi Wu" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "text", + "content": " Lei Zhu" + }, + { + "bbox": [ + 83, + 164, + 495, + 193 + ], + "type": "inline_equation", + "content": "^{1,6*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology (Guangzhou) " + }, + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "text", + "content": "University of Electronic Science and Technology of China " + }, + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "text", + "content": "West China Hospital " + }, + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "text", + "content": "Amazon " + }, + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "text", + "content": "Hong Kong Polytechnic University " + }, + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 62, + 194, + 526, + 220 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 270, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 296, + 619 + ], + "type": "text", + "content": "Fairness in AI-assisted medical image analysis is crucial for equitable healthcare, but is often neglected, especially in prevalent cross-domain scenarios (diverse demographics and imaging protocols). Effective and equitable deployment of AI models in these scenarios is critical, yet traditional Unsupervised Domain Adaptation (UDA) methods exhibit limited improvements. Emerging Active Domain Adaptation (ADA) approaches offer more effective enhancements, but all ignore fairness issues. Therefore, in this work, we propose the first fairness-aware ADA paradigm that simultaneously achieves both enhanced fairness and superior overall performance. Our method leverages the multimodal alignment capability of Vision-Language Models (VLMs): By performing medical images (vision) and sensitive attributes (language) learning, VLM inherently captures semantic correlations between visual features and protected attributes, enabling explicit attributes representation. Building on this foundation, we further devise an attribute-aware strategy (FairAP), which dynamically adapts to diverse patient demographics to promote equitable and high-quality outcomes by considering both Attribute and Polysemy. Extensive experiments on the FairDomain benchmark demonstrate that our method significantly reduces bias and maintains leading performance, outperforming existing UDA and ADA methods. This work pioneers a VLM-driven ADA paradigm for fair cross-domain medical segmentation, offering a blueprint for effective and equitable AI deployment in clinical practice. Code will be on (https://github.com/whq-xxh/Fair-AP)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 651, + 135, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 651, + 135, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 651, + 135, + 664 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "content": "\"Fairness is not an attitude. It's a professional skill that must be developed and exercised.\" — Brit Hume." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 244, + 555, + 373 + ], + "blocks": [ + { + "bbox": [ + 317, + 244, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 317, + 244, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 317, + 244, + 555, + 373 + ], + "type": "image", + "image_path": "e3bf325cc1380e29fe850a634f2f1ebad8c401afb6b0a29e792f318a49298ac2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 375, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 418 + ], + "type": "text", + "content": "Figure 1. Comparison of previous cross-domain fairness methods with our VLM-Driven fairness-aware ADA paradigm. Our method combines VLM and ADA to achieve both leading model performance and enhanced fairness." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 426, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 555, + 545 + ], + "type": "text", + "content": "The advent of AI technologies has significantly transformed medical image analysis (MedIA), providing sophisticated tools for analyzing complex biological patterns within imaging data [37, 46, 49-51]. These advanced algorithms enhance diagnostic precision, support informed clinical decisions, and improve patient outcomes. However, as these models become increasingly integrated into clinical practice, critical challenges have emerged: inherent biases within these models [24, 32, 57] and domain shifts caused by diverse clinical scenarios [23, 41, 54]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": "Numerous studies have reported that biases in MedIA tasks [13, 16, 19, 29, 31, 33, 35, 39, 60], stemming from training datasets failing to represent diverse populations adequately, raise profound ethical and fairness concerns [10, 34]. These biases cause severe health inequities, impacting marginalized groups in clinical applications. For example, Black individuals face nearly twice the risk of glaucoma than others, highlighting the need for accurately segmenting the optic disc and cup for early diagnosis, yet AI models often perform worst in this subgroup. This issue is particularly critical in cross-domain scenarios, where models deployed across institutions encounter heterogeneous data distributions that amplify performance disparities [36, 40, 41]. Such domain-specific biases not only undermine the per" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 704, + 140, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 704, + 140, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 704, + 140, + 712 + ], + "type": "text", + "content": "\\*corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24102" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "formance of AI systems but also violate fairness. Thus, ensuring both efficient and fair adaptation in cross-domain real-world medical AI environments is crucial to translate advanced technology to equitable healthcare for all." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 123, + 296, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 123, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 123, + 296, + 351 + ], + "type": "text", + "content": "Recently, significant efforts have been made to tackle fairness across various fields [1, 3, 4, 9, 11, 15, 20, 24, 29, 38-40, 58, 62, 64, 65]. These studies explore various attributes across different modalities or body parts. However, real-world cross-domain challenges remain underexplored. While Tian et al. [41] established the multiattribute benchmark dataset to evaluate cross-domain fairness, the limited performance of its Unsupervised Domain Adaptation (UDA) framework (e.g., IoU values " + }, + { + "bbox": [ + 56, + 123, + 296, + 351 + ], + "type": "inline_equation", + "content": "< 40\\%" + }, + { + "bbox": [ + 56, + 123, + 296, + 351 + ], + "type": "text", + "content": " in Tables 1 & 3) severely restricts clinical applicability despite its fairness-aware design. Although emerging Active Domain Adaptation (ADA) has demonstrated significant performance gains over UDA by actively selecting valuable target-domain samples [45], its fairness implications remain largely unexplored. This creates a critical gap in achieving both equitable outcomes and clinically viable performance across demographic subgroups, as current methods force an undesirable trade-off between the fairness guarantees and the model effectiveness." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 354, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 715 + ], + "type": "text", + "content": "Therefore, in this work, we propose the first fairness-aware ADA paradigm to simultaneously enhance fairness across all subgroups and boost overall performance for cross-domain medical image segmentation scenario applications. Specifically, according to the unique characteristics of medical images and corresponding sensitive attribute information (such as race: Black, White, Asian), we present a novel perspective that integrates the Vision-Language Model (VLM) to align image features with sensitive attributes. This alignment endows our model with the capability to perceive the sensitive attribute representations within images accurately. This VLM-attribute learning is essential as it establishes a solid foundation for subsequent representative attributes and polysemy sample selection. Then, we devise an attribute-aware FairAP sample selection strategy, where we select a quite small (only " + }, + { + "bbox": [ + 56, + 354, + 296, + 715 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 56, + 354, + 296, + 715 + ], + "type": "text", + "content": " here) but critical set of samples and annotate them for target fair model fine-tuning. The FairAP strategy allocates equal quotas to each subgroup to ensure balanced attention and leverages VLM attribute-latent space representations for attribute-representative and polysemic sample selection. Our process and its key distinctions from previous methods are visually demonstrated in Fig. 1. It is worth noting that the FairAP strategy is highly practical as it significantly reduces the workload of doctors, while simultaneously achieving both fairness and high performance (even improving IoU by around " + }, + { + "bbox": [ + 56, + 354, + 296, + 715 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 56, + 354, + 296, + 715 + ], + "type": "text", + "content": " than FairDA method [41] for rim segmentation as shown in Table 1). Comprehensive experiments on the FairDomain benchmark [41] confirm that our approach markedly diminishes bias and retains cutting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "edge performance in segmentation tasks, surpassing traditional DA and ADA methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 326, + 97, + 547, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 97, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 326, + 97, + 547, + 108 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 110, + 553, + 336 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 110, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 110, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 110, + 553, + 156 + ], + "type": "text", + "content": "- To the best of our knowledge, we are the first to tackle fairness challenges in cross-domain medical image segmentation with ADA paradigm, dynamically adapting to diverse patient subgroups for equitable outcomes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 158, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 158, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 314, + 158, + 553, + 204 + ], + "type": "text", + "content": "- We propose to leverage the Vision-Language Model to seamlessly integrate demographic attributes like race or gender into the learning process, endowing the model with the ability to perceive these sensitive attributes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 205, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 205, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 314, + 205, + 553, + 263 + ], + "type": "text", + "content": "- We devise the attribute-aware FairAP strategy to ensure both fairness and effectiveness. FairAP optimizes representation by balancing samples from underrepresented demographic groups while maintaining diversity to enhance overall performance comprehensively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 265, + 553, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 265, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 314, + 265, + 553, + 336 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our method surpasses previous DA and ADA approaches by ensuring superior fairness across multiple protected attributes and maintaining top performance. Our method pioneers a novel solution paradigm for fairness and effectiveness in cross-domain medical AI model deployment." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 349, + 406, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 349, + 406, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 349, + 406, + 361 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 369, + 470, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 470, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 470, + 382 + ], + "type": "text", + "content": "2.1. Fairness in Medical Imaging" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 590 + ], + "type": "text", + "content": "In medical image processing, fairness aims to reduce model biases to ensure equitable outcomes across patient subgroups. Existing methods fall into three main categories: pre-processing, in-processing, and post-processing [6, 26, 57]. Pre-processing techniques, such as those proposed by Bissoto et al. [5] and others [24, 29, 59], aim to correct training data imbalances, but may affect computational efficiency. In-processing methods mitigate bias during training by incorporating fairness-aware loss functions [40] or fairness-specific modules [20, 25, 41], which can potentially trade off some accuracy. On the other hand, post-processing approaches, like those introduced by Wang et al. [53], Kim et al. [17], and others [22], adjust the model outputs to improve fairness, which may struggle to generalize well during testing. However, few studies focus on fairness in cross-domain scenarios within real-world deployment environments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 600, + 430, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 430, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 430, + 612 + ], + "type": "text", + "content": "2.2. Domain Adaptation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Unsupervised Domain Adaptation (UDA): UDA trains on the labeled source and unlabeled target data to bridge domain gaps, typically using adversarial training and minimizing inter-domain discrepancies [54]. In UDA methods, adversarial learning minimizes distributional differences between source and target domains, enhancing performance in the target domain [42, 44]. Additionally, self-training in UDA uses pseudo-labels from target domain data to en" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24103" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 69, + 541, + 358 + ], + "blocks": [ + { + "bbox": [ + 65, + 69, + 541, + 358 + ], + "lines": [ + { + "bbox": [ + 65, + 69, + 541, + 358 + ], + "spans": [ + { + "bbox": [ + 65, + 69, + 541, + 358 + ], + "type": "image", + "image_path": "08f6d3ad504d8ea729a41efa9bb0fef3fff8a084b9552ce8ed6366d6073d7e8e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "text", + "content": "Figure 2. Overview of the Fair VLM-Driven ADA method (taking race attribute for illustration). Given the target domain images " + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "inline_equation", + "content": "x^T" + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "text", + "content": " and their corresponding sensitive attribute information " + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "inline_equation", + "content": "\\text{attrs}" + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "text", + "content": ", we first align them through VLM-attribute learning to obtain the vision-based sensitive attribute features " + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "inline_equation", + "content": "F_{attrR}" + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "text", + "content": ". Then, we devise the attribute-aware FairAP Select strategy to select " + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "inline_equation", + "content": "\\alpha \\%" + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "text", + "content": " of the samples that are representations of Attribute and Polysemy. These selected samples " + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "text", + "content": " are then annotated by experts to obtain the annotations " + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "inline_equation", + "content": "y^{TL}" + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "text", + "content": ". Finally, active domain adaptation fine-tuning is performed to obtain the target fair model " + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "inline_equation", + "content": "h_{\\phi}'" + }, + { + "bbox": [ + 54, + 358, + 555, + 417 + ], + "type": "text", + "content": " for top performance and enhanced fairness." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 427, + 297, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 427, + 297, + 525 + ], + "spans": [ + { + "bbox": [ + 54, + 427, + 297, + 525 + ], + "type": "text", + "content": "hance performance in the target domain [12, 27, 66]. Building on the work by Hoyer et al. [12], Tian et al. [41] further explored fairness in cross-domain settings by integrating the fairness module. This new setting can more accurately reflect real-world clinical scenarios and provide a clear quantification of the performance and fairness of cross-domain methods [41]. However, very limited performance gains brought by UDA significantly diminish its clinical utility." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 297, + 715 + ], + "type": "text", + "content": "Active Domain Adaptation (ADA): ADA combines active learning and domain adaptation to reduce annotation costs and sustain high performance by focusing on annotating the most valuable samples through strategies like uncertainty-based selection and diversity-based sampling [8, 14, 21, 48, 56]. For example, Wang et al. [45], and others [47, 52, 55, 61] enhanced cross-domain performance by strategically selecting target domain samples for model development. However, these current ADA methods have predominantly concentrated on optimizing overall model performance, neglecting the crucial aspect of fairness. Consequently, we are the pioneers in introducing a novel fairness-aware ADA paradigm, effectively minimizing annotation costs while simultaneously achieving both enhanced fairness and leading performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 426, + 398, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 398, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 398, + 441 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 450, + 413, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 413, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 413, + 464 + ], + "type": "text", + "content": "3.1. Problem Setting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": "Cross-domain MedIA often exhibits not only conventional performance degradation for domain shifts but also issues of unfairness related to sensitive attributes. Fair ADA is an effective and novel approach to address this challenge. Our work specifically targets fairness challenges in domain adaptation, focusing on fairness issues across multiple sensitive attributes in diverse subgroups. The problem is defined as follows: In medical segmentation tasks, consider the source domain data " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "D_{S} = \\{(x_{i}^{S},y_{i}^{S},\\text{attr}_{i})\\}_{i = 1}^{N_{S}}" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{i}^{S}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": " image, the " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "y_{i}^{S}\\in \\mathbb{R}^{C\\times H\\times W}" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": " denotes the corresponding label with " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": " classes, and the " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\text{attr}_{i}\\in \\mathbb{R}^{n}" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": " means the sensitive attributes related to the sample, such as gender, race, or ethnicity. Meanwhile, there is an unlabeled dataset " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "D_{T} = \\{(x_{i}^{T},\\text{attr}_{i})\\}_{i = 1}^{N_{T}}" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": " whose domain vary, leading to different image features from " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": ". In the fair segmentation setting, note that the target attributes " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\text{attr}_{i}" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": " are available and accessible, as outlined in [40, 41]. The aim is to design strategies for selecting a small number of samples " + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "inline_equation", + "content": "S = \\alpha \\% \\cdot D_{T}\\ll D_{T}" + }, + { + "bbox": [ + 313, + 472, + 556, + 715 + ], + "type": "text", + "content": " for labeling and active learning, thereby maximizing the fairness" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24104" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "and performance of model " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "h_{\\phi}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " as much as possible while maintaining " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\alpha \\%" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " low." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 201, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 201, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 201, + 114 + ], + "type": "text", + "content": "3.2. Overview of the Fair-ADA" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "text", + "content": "The overall structure of our proposed VLM-driven fairness-aware ADA framework is illustrated in Fig. 2. The entire methodology section uses race attribute " + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "inline_equation", + "content": "attrR" + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "text", + "content": " as an example to illustrate. In the VLM-attribute Learning stage (a), we utilize a Vision-Language model " + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "text", + "content": " to align image samples with sensitive attribute information within the latent space. The model " + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "text", + "content": " performs vision (unlabeled medical images " + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "inline_equation", + "content": "x^{T}" + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "text", + "content": ") language (corresponding attribute information " + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "inline_equation", + "content": "attrR" + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "text", + "content": ") learning to achieve alignment. Then, we can extract vision-based sensitive attribute features " + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "inline_equation", + "content": "F_{attrR}" + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "text", + "content": ". These feature representations consist of multiple subgroups corresponding to various attributes. For example, sensitive racial features " + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "inline_equation", + "content": "F_{attrR}" + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "text", + "content": " include three subgroup features such as Asian features " + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "text", + "content": ", Black features " + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "inline_equation", + "content": "F_{b}" + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "text", + "content": ", and White features " + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "inline_equation", + "content": "F_{w}" + }, + { + "bbox": [ + 55, + 119, + 296, + 299 + ], + "type": "text", + "content": ". The entire process can be expressed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 304, + 295, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 304, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 87, + 304, + 295, + 319 + ], + "type": "interline_equation", + "content": "F _ {a t t r R} = g _ {\\theta} \\left(x ^ {T}, a t t r R\\right) = \\left\\{F _ {a}, F _ {b}, F _ {w} \\right\\}. \\tag {1}", + "image_path": "2d2bb9edd062c19307924045ae29047a9c83efd9ca6d91edf6f7948609bdc267.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 323, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 467 + ], + "type": "text", + "content": "Based on the vision-based sensitive attribute features " + }, + { + "bbox": [ + 55, + 323, + 296, + 467 + ], + "type": "inline_equation", + "content": "F_{attrR}" + }, + { + "bbox": [ + 55, + 323, + 296, + 467 + ], + "type": "text", + "content": ", the process moves into the FairAP Select stage (b), where our approach is organized into three main steps. In the first step, we ensure and amplify fairness within our active learning strategy by allocating an equal quota of samples from each sensitive attribute subcategory. In the second step, we employ clustering methods for each subgroup based on various sensitive attribute representations, followed by selecting attribute representative samples. In the third step, we further optimize the selected samples by expanding the selection to include polysemy representatives. This overall process can be summarized as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 473, + 294, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 473, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 125, + 473, + 294, + 486 + ], + "type": "interline_equation", + "content": "S = \\text {F a i r} A P \\left(F _ {\\text {a t t r} R}\\right). \\tag {2}", + "image_path": "305c0fdd8c62e81318aa7ed39b02324363294ea98039839d91c58670f14e10e4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 491, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 564 + ], + "type": "text", + "content": "In the Active Domain Adaptation stage (c), we utilize the model " + }, + { + "bbox": [ + 55, + 491, + 296, + 564 + ], + "type": "inline_equation", + "content": "h_{\\phi}" + }, + { + "bbox": [ + 55, + 491, + 296, + 564 + ], + "type": "text", + "content": " that has been previously trained on the source domain data " + }, + { + "bbox": [ + 55, + 491, + 296, + 564 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 55, + 491, + 296, + 564 + ], + "type": "text", + "content": ". We then fine-tune this model using the sample set " + }, + { + "bbox": [ + 55, + 491, + 296, + 564 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 491, + 296, + 564 + ], + "type": "text", + "content": " selected by our FairAP strategy to produce the optimized target fair model " + }, + { + "bbox": [ + 55, + 491, + 296, + 564 + ], + "type": "inline_equation", + "content": "h_{\\phi}^{\\prime}" + }, + { + "bbox": [ + 55, + 491, + 296, + 564 + ], + "type": "text", + "content": ". The loss function for this model is defined by the following formula:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 570, + 295, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 570, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 64, + 570, + 295, + 601 + ], + "type": "interline_equation", + "content": "L _ {A D A} \\left(h _ {\\phi} ^ {\\prime} (x ^ {T L}), y ^ {T L}\\right) = - \\sum_ {c = 1} ^ {C} y _ {c} \\log \\left(h _ {\\phi} ^ {\\prime} (x)\\right), \\tag {3}", + "image_path": "fe82b79be8073e74da1422d16677bc4c9e456eb2715624212c25dc5e35b9516c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 606, + 296, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 296, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 296, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 606, + 296, + 642 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 606, + 296, + 642 + ], + "type": "text", + "content": " represents the number of classes, and selected list " + }, + { + "bbox": [ + 55, + 606, + 296, + 642 + ], + "type": "inline_equation", + "content": "S = \\{x^{TL},y^{TL}\\}" + }, + { + "bbox": [ + 55, + 606, + 296, + 642 + ], + "type": "text", + "content": ", whose size corresponds to " + }, + { + "bbox": [ + 55, + 606, + 296, + 642 + ], + "type": "inline_equation", + "content": "\\alpha \\%" + }, + { + "bbox": [ + 55, + 606, + 296, + 642 + ], + "type": "text", + "content": " of the total dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 195, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 195, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 195, + 661 + ], + "type": "text", + "content": "3.3. VLM-attribute Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Traditional fairness-aware segmentation methods [40, 41] using the simplistic categorical encoding (e.g., 0/1 for gender male or female) suffer from semantic sparsity and contextual blindness, which fails to capture attribute-visual" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 322, + 81, + 548, + 170 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 411, + 80 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 411, + 80 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 411, + 80 + ], + "type": "text", + "content": "VLM-attribute learning" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 322, + 81, + 548, + 170 + ], + "lines": [ + { + "bbox": [ + 322, + 81, + 548, + 170 + ], + "spans": [ + { + "bbox": [ + 322, + 81, + 548, + 170 + ], + "type": "image", + "image_path": "c646274ca138d0181a38b3873a00f629cd9bb42ab340cce4a970902c44a0c6d2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 175, + 554, + 219 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 554, + 219 + ], + "type": "text", + "content": "Figure 3. Visualization of racial feature distribution before and after feature alignment via VLM-attribute learning. Our alignment process creates a denser distribution of features within racial subgroups, improving distinguishability across different races." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 231, + 554, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 554, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 554, + 303 + ], + "type": "text", + "content": "correlations. These disjoint representations limit ADA paradigms' ability to perceive subgroup distributions during cross-domain adaptation. To address this, we propose the VLM-based attribute-visual alignment to enable context-aware perception and provide continuous semantic gradients for precise subgroup-aware adaptation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": "Fig. 2 (a) provides a detailed illustration of our vision language learning process. Specifically, we utilize a pretrained vision-language model (CLIP [30, 63] here) denoted as " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " to process the samples " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "x^{T}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " along with their associated sensitive attribute information " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\text{attr } R" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": ". The model's vision encoder, " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "g_{\\theta_v}^v" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": ", extracts image tokens " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "z_{v}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " from the input " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "x^{T}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\theta_{v}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": ". Meanwhile, the language encoder, " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "g_{\\theta_l}^l" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": ", derives attribute tokens " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "z_{l}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " from the input " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\text{attr } R" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\theta_{l}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": ". Therefore, we can express the model as " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "g_{\\theta} = \\{g_{\\theta_v}^v,g_\\theta_l^l\\}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " with the parameters as " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "\\theta = \\{\\theta_v,\\theta_l\\}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": ". When training the model " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " using a batch of " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " image-attribute token pairs, we generate a similarity matrix " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "A\\in \\mathbb{R}^{N\\times N}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": ". Each entry of the matrix, " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "A_{ij} = z_{vi}^{T}z_{lj}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": ", represents the cosine similarity between the " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " image token " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "z_{vi}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "j_{th}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " attribute token " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "z_{lj}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " within the batch. The model " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " employs a contrastive learning approach that aims to maximize the cosine similarity for positive image-attribute token pairs while minimizing it for negative token pairs. The model " + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 313, + 303, + 555, + 532 + ], + "type": "text", + "content": " loss function can be expressed as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 350, + 540, + 554, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 540, + 554, + 572 + ], + "spans": [ + { + "bbox": [ + 350, + 540, + 554, + 572 + ], + "type": "interline_equation", + "content": "L _ {V L M} = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\frac {1}{2} \\log \\left(\\frac {\\exp (2 A _ {i i})}{S _ {i} T _ {i}}\\right), \\tag {4}", + "image_path": "023c08a9d0e163da79203f74f8eac2d079ec17840e60b188a2c317043c0a1fdc.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 554, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 633 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 554, + 633 + ], + "type": "inline_equation", + "content": "S_{i} = \\sum_{j=1}^{N} \\exp(A_{ij})" + }, + { + "bbox": [ + 313, + 582, + 554, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 554, + 633 + ], + "type": "inline_equation", + "content": "T_{i} = \\sum_{j=1}^{N} \\exp(A_{ji})" + }, + { + "bbox": [ + 313, + 582, + 554, + 633 + ], + "type": "text", + "content": ". Since our focus is on enhancing the vision-based attribute perception capabilities, we choose to freeze the language encoder " + }, + { + "bbox": [ + 313, + 582, + 554, + 633 + ], + "type": "inline_equation", + "content": "g_{\\theta}^{l}" + }, + { + "bbox": [ + 313, + 582, + 554, + 633 + ], + "type": "text", + "content": " to better train " + }, + { + "bbox": [ + 313, + 582, + 554, + 633 + ], + "type": "inline_equation", + "content": "g_{\\theta}" + }, + { + "bbox": [ + 313, + 582, + 554, + 633 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 365, + 641, + 405, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 641, + 405, + 654 + ], + "spans": [ + { + "bbox": [ + 365, + 641, + 405, + 654 + ], + "type": "interline_equation", + "content": "A ^ {\\prime} \\leftarrow A,", + "image_path": "9afbc6fde8f7170b096be1c5cd483beab26ce682b4c4e44664b364f7f37a7b1d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 366, + 656, + 553, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 656, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 366, + 656, + 553, + 682 + ], + "type": "interline_equation", + "content": "\\theta_ {v} ^ {\\prime} \\leftarrow \\theta_ {v} - \\eta \\frac {\\partial L _ {V L M}}{\\partial \\theta_ {v}}, \\quad \\theta_ {l} ^ {\\prime} = \\theta_ {l}, \\tag {5}", + "image_path": "2df0c28c5bfc6c2c94d3e42c27a98e8281eee671525d84e85e82e15e74631710.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " is the learning rate. We employ a consistent template for sensitive attribute information as:" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24105" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 82, + 72, + 266, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 72, + 266, + 85 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 266, + 85 + ], + "type": "text", + "content": "A photo of a fundus and the race is " + }, + { + "bbox": [ + 82, + 72, + 266, + 85 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{RACE}\\rangle" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "type": "inline_equation", + "content": "\\langle \\mathbf{RACE}\\rangle" + }, + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "type": "text", + "content": " represents specific racial information. After completing the training, we utilize a fixed vision encoder " + }, + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "type": "inline_equation", + "content": "g_{\\theta_v'}^v" + }, + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "type": "text", + "content": " to process all target domain images " + }, + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "type": "inline_equation", + "content": "x^T" + }, + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "type": "text", + "content": ", allowing extraction of the vision-based sensitive attribute features " + }, + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "type": "inline_equation", + "content": "F_{attrR}" + }, + { + "bbox": [ + 55, + 91, + 296, + 150 + ], + "type": "text", + "content": " as following:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 150, + 294, + 166 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 150, + 294, + 166 + ], + "spans": [ + { + "bbox": [ + 133, + 150, + 294, + 166 + ], + "type": "interline_equation", + "content": "F _ {a t t r R} = g _ {\\theta_ {v} ^ {\\prime}} ^ {v} \\left(x ^ {T}\\right). \\tag {6}", + "image_path": "75aaff7cb94b9bab6153da1f0ca0c20760d05653a3964d9537a55914baf932dd.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 167, + 296, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 194 + ], + "type": "text", + "content": "The visualization of racial characteristic distribution before and after our VLM-attribute learning is shown in Fig. 3." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 198, + 155, + 212 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 155, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 155, + 212 + ], + "type": "text", + "content": "3.4. FairAP Strategy" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "text", + "content": "Building on the VLM-attribute learning described earlier, sensitive vision-based attribute features " + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "inline_equation", + "content": "F_{attrR}" + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "text", + "content": " are successfully extracted. For clarity, we continue using sensitive racial features " + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "inline_equation", + "content": "F_{attrR}" + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "text", + "content": " as an example to illustrate the Fair Attribute and Polysemy representation (FairAP) Select strategy. The sensitive racial features " + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "inline_equation", + "content": "F_{attrR}" + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "text", + "content": " encompass three types of racial characteristics: Asian features " + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "text", + "content": ", Black features " + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "inline_equation", + "content": "F_{b}" + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "text", + "content": ", and White features " + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "inline_equation", + "content": "F_{w}" + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "text", + "content": ". Thus, we define " + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "inline_equation", + "content": "F_{attrR} = \\{F_{a}, F_{b}, F_{w}\\}" + }, + { + "bbox": [ + 55, + 216, + 296, + 360 + ], + "type": "text", + "content": ". Our FairAP Select strategy encompasses three critical components: Fair Quota Select, Attribute Representation Select, and Polysemy Representation Select. We will introduce each component in sequence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 360, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 421 + ], + "type": "text", + "content": "Fair Quota Selection. We recognize the data-driven nature of AI. Therefore, to achieve our fairness objective ('treat all our subgroups with fairness'), we ensure a consistent sample quota for each racial subgroup. We define the sample quota for each subgroup as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 428, + 294, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 428, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 125, + 428, + 294, + 441 + ], + "type": "interline_equation", + "content": "\\Delta_ {a} = \\Delta_ {b} = \\Delta_ {w} = \\Delta_ {R}, \\tag {7}", + "image_path": "259f9a5ed6f41b4b2872dcf6c90078e9c4ecf32da8150dba20f5591b56907410.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\Delta_{a}" + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\Delta_{b}" + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\Delta_{w}" + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "text", + "content": " represent the quotas given to Asians, Blacks, and Whites respectively. The " + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\Delta_{R}" + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "text", + "content": " is calculated by dividing the overall quota " + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "inline_equation", + "content": "(\\alpha \\% \\cdot D_T)" + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "text", + "content": " by the number of total racial subgroups, which is three here. Upon completing the evaluations of the subsequent components (Attribute and Polysemy), which assist in determining the specific samples, we compile the final selection list for each racial subgroup, represented as " + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "inline_equation", + "content": "S_{R} = \\{S_{\\hat{R}A}, S_{\\hat{R}P}\\}" + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "text", + "content": ". Finally, we combine the selection lists " + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "inline_equation", + "content": "S_{R}" + }, + { + "bbox": [ + 55, + 449, + 296, + 568 + ], + "type": "text", + "content": " of all racial subgroups and submitted them to doctors for annotations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "text", + "content": "Attribute Representation Selection. To optimally select representative samples for each racial subgroup, we undertake clustering operation (k-means++ algorithm [2] here) to process the feature set " + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "inline_equation", + "content": "F_{R}" + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "text", + "content": ", resulting in the clustered features " + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "inline_equation", + "content": "F_{R}^{\\prime}" + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "text", + "content": " and the corresponding cluster centers " + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "inline_equation", + "content": "C_{R}^{N_{R}}" + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "text", + "content": " represents a specific racial group (like Asian, Black, or White). This process is represented as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 111, + 659, + 296, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 659, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 111, + 659, + 296, + 681 + ], + "type": "interline_equation", + "content": "\\left\\{F _ {R} ^ {\\prime}, C _ {R} ^ {N _ {R}} \\right\\} = \\operatorname {C l u s t e r} \\left(F _ {R}\\right). \\tag {8}", + "image_path": "3d58f6cb3fbdc45957f21f5ee1215e678370676c7c504c32616966e961897780.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "In this equation, " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "C_R^{N_R}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " denotes the cluster centers, and " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "N_R" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": " indicates the number of cluster centers for the racial sub" + } + ] + } + ], + "index": 11 + }, + { + "type": "code", + "bbox": [ + 316, + 89, + 553, + 415 + ], + "blocks": [ + { + "bbox": [ + 321, + 75, + 494, + 87 + ], + "lines": [ + { + "bbox": [ + 321, + 75, + 494, + 87 + ], + "spans": [ + { + "bbox": [ + 321, + 75, + 494, + 87 + ], + "type": "text", + "content": "Algorithm 1: POLYSEMY-REP SELECT" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "lines": [ + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": "Input: The " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " cluster center of subgroup " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "C_R^i" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": ", distance set: " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "D_R^i" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": ", selected distance set: " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "D_{\\hat{R}A}^i" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": ", features: " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "F_{R,m}''" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": ", number of " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "F_{R,m}': M_i" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": ", other racial subgroups' cluster centers: " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "C_{\\bar{R}}^k" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": ", quota: " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "\\Delta_R" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": ", subgroup cluster number: " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "N_R" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": ". Output: Polysemy-Rep selective racial list: " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "S_{\\hat{R}P}" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " \n1 for " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "i = 1" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "N_R" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " do \n2 " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "{D_R'} \\gets {D_R} \\backslash {D_{\\hat{R}A}}^i // \\text{Remaining distances}" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " \n3 " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "F_{R,m}'' \\gets \\text{Index} \\left( D_R^{i'} \\right) // \\text{Indexing features}" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " \n4 " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "N_i \\gets M_i - \\frac{\\Delta_R}{2N_R} // \\text{Adjust selection range}" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " \n5 Initialize empty list " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "E_{temp}" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": "; \n6 for " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "m = 1" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "N_i" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " do \n7 if " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "F_{R,m}'' \\in C_R^i \\land F_{R,m}'' \\notin C_{\\bar{R}}^k" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " then \n8 " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "\\begin{array}{c} e_i^m \\gets \\min \\left( \\| F_{R,m}' - C_{\\bar{R}}^k \\| \\right) // \\text{Min distance to other racial clusters} \\\\ \\text{Append } e_i^m \\text{ to } E_{temp}; \\end{array}" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " \n9 end \n10 end \n11 Sort " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "E_{temp}" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " in rising order; \n12 " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "E_{\\hat{R}P}^i \\gets \\text{top min distances from } E_{temp}" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": "; \n14 end \n15 " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "F_{\\hat{R}P}' \\gets \\text{Index} \\left( \\{E_{\\hat{R}P}', ..., E_{\\hat{R}P}', ..., E_{\\hat{R}P}^{N_R}\\} \\right)" + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "text", + "content": " \n16 " + }, + { + "bbox": [ + 316, + 89, + 553, + 415 + ], + "type": "inline_equation", + "content": "S_{\\hat{R}P} \\gets F_{\\hat{R}P}' // \\text{Get the final Polysemy-Rep racial subgroup list}" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + } + ], + "index": 13, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 438, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 487 + ], + "type": "text", + "content": "group. For each racial subgroup, the number of cluster centers " + }, + { + "bbox": [ + 313, + 438, + 555, + 487 + ], + "type": "inline_equation", + "content": "C_R^{N_R}" + }, + { + "bbox": [ + 313, + 438, + 555, + 487 + ], + "type": "text", + "content": " is calculated based on their overall distribution in the dataset to comprehensively represent the target domain. This is illustrated by the following formula:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 353, + 496, + 553, + 509 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 496, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 353, + 496, + 553, + 509 + ], + "type": "interline_equation", + "content": "N _ {a}: N _ {b}: N _ {w} \\approx \\operatorname {S u m} _ {a}: \\operatorname {S u m} _ {b}: \\operatorname {S u m} _ {w}, \\tag {9}", + "image_path": "acb8b708cf59d1e9ba5356328acea59ec11ab6086eb7b8bbfd17f3dcb2137363.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{Sum}_a" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{Sum}_b" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{Sum}_w" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": " denote the proportions of Asians, Blacks, and Whites, respectively. For the " + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": " cluster center " + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "inline_equation", + "content": "C_R^i" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": " of race " + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": ", we define a set of Euclidean distances " + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "inline_equation", + "content": "D_R^i" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": ", where each element " + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "inline_equation", + "content": "d_i^m" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": " in the set is the distance from sample " + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "inline_equation", + "content": "F_{R,m}^{\\prime}" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "inline_equation", + "content": "m_{th}" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": " sample belongs to the " + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": " cluster) to the cluster center " + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "inline_equation", + "content": "C_R^i" + }, + { + "bbox": [ + 313, + 517, + 555, + 605 + ], + "type": "text", + "content": ". These distances are then sorted in ascending order. This can be represented as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 611, + 553, + 663 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 611, + 553, + 663 + ], + "spans": [ + { + "bbox": [ + 326, + 611, + 553, + 663 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D _ {R} ^ {i} = \\operatorname {s o r t} \\left(\\left\\{d _ {i} ^ {m} = \\| F _ {R, m} ^ {\\prime} - C _ {R} ^ {i} \\| \\mid F _ {R, m} ^ {\\prime} \\in C _ {R} ^ {i} \\right\\}\\right) \\\\ = \\left\\{d _ {i} ^ {1}, d _ {i} ^ {2}, \\dots , d _ {i} ^ {m}, \\dots , d _ {i} ^ {M _ {i}} \\right\\}, \\tag {10} \\\\ \\end{array}", + "image_path": "a53eb6383abe73912e519f5e33717bac5d9adb90e0bb7d257375a268aa760822.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "type": "inline_equation", + "content": "d_i^1 \\leq d_i^2 \\leq \\dots \\leq d_i^m \\leq \\dots \\leq d_i^{M_i}" + }, + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "type": "inline_equation", + "content": "M_{i}" + }, + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "type": "text", + "content": " is the number of samples " + }, + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "type": "inline_equation", + "content": "F_{R}^{\\prime}" + }, + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "type": "text", + "content": " that belong to the cluster center " + }, + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "type": "inline_equation", + "content": "C_R^i" + }, + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "type": "text", + "content": ". To ensure that the selected attribute representatives " + }, + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "type": "inline_equation", + "content": "F_{\\hat{R}A}" + }, + { + "bbox": [ + 313, + 663, + 555, + 714 + ], + "type": "text", + "content": " can comprehensively cover the entire distribution of the dataset," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24106" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "text", + "content": "we establish an equal selection quota of " + }, + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "inline_equation", + "content": "\\frac{\\Delta_R}{2N_R}" + }, + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "text", + "content": " for each cluster. Specifically, we select the top " + }, + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "inline_equation", + "content": "\\frac{\\Delta_R}{2N_R}" + }, + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "text", + "content": " samples, denoted as the subset " + }, + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "inline_equation", + "content": "D_{\\hat{R}A}^i" + }, + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "text", + "content": ", corresponding to the smallest Euclidean distances from the sorted set " + }, + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "inline_equation", + "content": "D_R^i" + }, + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "text", + "content": ". The set " + }, + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "inline_equation", + "content": "D_{\\hat{R}A}" + }, + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "text", + "content": " represents the distances selected from the " + }, + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "inline_equation", + "content": "N_R" + }, + { + "bbox": [ + 55, + 71, + 296, + 150 + ], + "type": "text", + "content": " clusters of the current racial subgroup. This can be expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 157, + 295, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 157, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 105, + 157, + 295, + 178 + ], + "type": "interline_equation", + "content": "D _ {R \\hat {A}} ^ {i} = \\left\\{d _ {R} ^ {1}, d _ {R} ^ {2}, \\dots , d _ {R} ^ {\\Delta_ {R} / 2 N _ {R}} \\right\\} \\tag {11}", + "image_path": "ab271c82904bfac3cf48516299c64687fff01a607c52dcbd380eddef99844aea.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 102, + 184, + 295, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 184, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 102, + 184, + 295, + 205 + ], + "type": "interline_equation", + "content": "D _ {\\hat {R A}} = \\left\\{D _ {\\hat {R A}} ^ {1}, D _ {\\hat {R A}} ^ {2}, \\dots , D _ {\\hat {R A}} ^ {N _ {R}} \\right\\}. \\tag {12}", + "image_path": "c967aae0d4d15bd5d4da1eff4acbc52496f7af4348dc2714a4d5602fba325d65.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "type": "text", + "content": "Using the indices from the set " + }, + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "type": "inline_equation", + "content": "D_{\\hat{R}A}" + }, + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "type": "text", + "content": " that correspond to the smallest distances, we can derive the feature set " + }, + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "type": "inline_equation", + "content": "F_{\\hat{R}A}^{\\prime}" + }, + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "type": "text", + "content": " for attribute representative selection and obtain the corresponding sample set " + }, + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "type": "inline_equation", + "content": "S_{\\hat{R}A}" + }, + { + "bbox": [ + 55, + 208, + 295, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 257, + 296, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 363 + ], + "type": "text", + "content": "Polysemy Representation Selection. Given that representative samples for each racial subgroup have been selected in the previous phase, we focus on choosing diverse samples in the Polysemy phase to explore a wider range of expressions within a limited quota, thus enhancing performance. For example, as seen in the right (post-VLM learning) of Fig. 3, some samples within the Black clusters are actually from White individuals, which may embody characteristics of both White and similar attributes of Black." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "text", + "content": "In detail, We first filter out the samples selected in the previous stage, as indicated by " + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "inline_equation", + "content": "D_{R}^{i'} = D_{R}^{i} \\setminus D_{\\hat{R}A}^{i}" + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "text", + "content": ". Subsequently, we calculate the nearest distance " + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "inline_equation", + "content": "e_i^m" + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "text", + "content": " from the remaining samples belonging to the " + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "text", + "content": " cluster center " + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "inline_equation", + "content": "C_R^i" + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "text", + "content": " of race " + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "text", + "content": " to the other races' cluster centers. These distances are then sorted in rising order to obtain a new Euclidean distance set " + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "inline_equation", + "content": "E_R^i" + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "text", + "content": ". After this, we will follow Algorithm 1 to identify the final selected set of polysemy representative features, " + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "inline_equation", + "content": "F_{\\hat{R}P}'" + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "text", + "content": ", and samples list, " + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "inline_equation", + "content": "S_{\\hat{R}P}" + }, + { + "bbox": [ + 55, + 365, + 296, + 547 + ], + "type": "text", + "content": ". Through the aforementioned selection strategy, the chosen samples not only exhibit characteristics of the race to which they belong but also possess features similar to other races. This demonstrates the polysemy of the selected samples. As illustrated in Fig. 2 (b), we use Asian samples as an example to demonstrate our selection process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 556, + 137, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 137, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 137, + 570 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 172, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 172, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 172, + 590 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Implementation Details. We use NVIDIA RTX 3090Ti GPUs with 24GB of memory for all experiments. We utilize the official code repository of the CLIP model [30]. We employ the pre-trained ViT-B/16 model with a batch size of 32 and use the Adam optimizer [18]. The initial learning rate is set to 1e-7, and the total number of iterations is 20,000. For segmentation model in the ADA process, following previous works [40, 41], we use the TransUNet [7] as the backbone and set the batch size to 2. The total number of iterations: 40,000 for training on the source domain and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 565, + 121 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 565, + 121 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 565, + 121 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 565, + 121 + ], + "type": "image", + "image_path": "a6532aa695cdeba009e1aff484cb50abec73ad11a36104a210a20af7243db4d4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 122, + 564, + 172 + ], + "blocks": [ + { + "bbox": [ + 317, + 122, + 564, + 172 + ], + "lines": [ + { + "bbox": [ + 317, + 122, + 564, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 122, + 564, + 172 + ], + "type": "image", + "image_path": "9cc126768d95ab511ea64d42ebca6c2784b6dd866963ccfda36deab6701a9414.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 172, + 564, + 221 + ], + "blocks": [ + { + "bbox": [ + 317, + 172, + 564, + 221 + ], + "lines": [ + { + "bbox": [ + 317, + 172, + 564, + 221 + ], + "spans": [ + { + "bbox": [ + 317, + 172, + 564, + 221 + ], + "type": "image", + "image_path": "fc0a8e1fd698224b24fb0ece4235844e4de92df43211d5bb8d182cc92538eace.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 221, + 403, + 228 + ], + "lines": [ + { + "bbox": [ + 334, + 221, + 403, + 228 + ], + "spans": [ + { + "bbox": [ + 334, + 221, + 403, + 228 + ], + "type": "text", + "content": "GT Baseline" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 478, + 221, + 496, + 228 + ], + "lines": [ + { + "bbox": [ + 478, + 221, + 496, + 228 + ], + "spans": [ + { + "bbox": [ + 478, + 221, + 496, + 228 + ], + "type": "text", + "content": "STDR" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 531, + 221, + 546, + 228 + ], + "lines": [ + { + "bbox": [ + 531, + 221, + 546, + 228 + ], + "spans": [ + { + "bbox": [ + 531, + 221, + 546, + 228 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 231, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 285 + ], + "type": "text", + "content": "Figure 4. Visual comparisons of our method and representative methods. Baseline: cross-domain results without fine-tuning; FIA [41]: fairness-aware DA approach; STDR [47]: ADA approach. Samples from top to bottom: Asian, Black, White subgroups, with numerical results on each image. Please zoom in for the best view." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "type": "text", + "content": "20,000 for fine-tuning on the target domain. The selection percentage of " + }, + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "type": "inline_equation", + "content": "\\alpha \\%" + }, + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 300, + 553, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 324, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 554, + 396 + ], + "type": "text", + "content": "Benchmark Dataset. We employ the cross-domain medical segmentation benchmark of the Harvard-FairDomain dataset [41], comprising 10,000 multi-modal samples with sensitive attributes (eg., gender, race, ethnicity). Its rich demographic attributes and paired modalities make it suitable for assessing and quantifying cross-domain model fairness." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 397, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 554, + 457 + ], + "type": "text", + "content": "Evaluation Metrics. We quantify fairness in medical image segmentation using metrics from previous studies [40, 41]: Dice, IoU, and the key composite metrics, ES-Dice and ES-IoU, to assess fairness alongside performance. For more details, please refer to the supplementary materials." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 467, + 493, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 493, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 493, + 479 + ], + "type": "text", + "content": "4.2. Comparison with SOTA Methods" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 605 + ], + "type": "text", + "content": "Following common practices in prior ADA works [45, 47], we comprehensively compare our approach with established DA methods and the latest ADA techniques. Our analysis primarily focused on fairness and performance comparisons across three sensitive attributes: Race, Ethnicity, and Gender (presented in Table 1, 3 and supplementary materials). Visualization examples of our method's results on racial attributes, compared to other representative methods (Baseline, FIA [41] and STDR [47]), are shown in Fig. 4." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "With SOTA DA Methods. We perform a comprehensive comparison with previous state-of-the-art DA methods (including PixMatch [28], DAFormer [12], and DAFormerFIA [41]) across three demographic attributes. Tables 1, & 3 prove that our method significantly outperforms these approaches across nearly all metrics. Taking the Race attribute (in Table 1) as an example, FairAP significantly outperforms the DA methods for all metrics. This reflects a stable improvement compared to the previous state-of-the" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24107" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 93, + 533, + 342 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "type": "text", + "content": "Table 1. Cup and rim segmentation performance on the FairDomain-Segmentation benchmark using different DA and ADA methods with Race as the demographic attribute. The * denotes " + }, + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "type": "text", + "content": "-value " + }, + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "type": "inline_equation", + "content": "< 0.05" + }, + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "type": "text", + "content": " in all paired " + }, + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "type": "text", + "content": "-test, indicating statistically significant differences." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 93, + 533, + 342 + ], + "lines": [ + { + "bbox": [ + 77, + 93, + 533, + 342 + ], + "spans": [ + { + "bbox": [ + 77, + 93, + 533, + 342 + ], + "type": "table", + "html": "
MethodVenueOverall ES-Dice↑Overall Dice↑Overall ES-IoU↑Overall IoU↑Asian Dice↑Black Dice↑White Dice↑Asian IoU↑Black IoU↑White IoU↑
DACupBaseline (Source)-0.8680.8880.7840.8080.8810.9010.8860.8020.8300.805
Baseline (Target)-0.6730.7000.5350.5640.7030.7310.6940.5670.6060.555
RimBaseline (Source)-0.8200.8610.7160.7620.8410.8380.8680.7370.7310.771
Baseline (Target)-0.4810.4950.3340.3420.4960.4720.4990.3470.3260.345
DACupPixMatch [28]CVPR'210.7390.7750.6140.6600.7520.7930.7830.6330.693
DAFormer [12]CVPR'220.7850.8040.6690.6910.7940.8160.8020.6820.7120.688
DAFormer-FIA [41]ECCV'240.7960.8100.6820.7000.8040.8210.8090.6940.7170.697
RimPixMatch [28]CVPR'210.6270.6730.4930.5230.6530.6400.6930.5220.4980.559
DAFormer [12]CVPR'220.3260.3320.1990.2020.3340.3450.3290.2040.2120.199
DAFormer-FIA [41]ECCV'240.5210.5310.3630.3690.5220.5390.5300.3640.3800.368
CupRandom-0.8130.8340.7160.7340.8140.8310.8370.7120.7350.737
Entropy [44]CVPR'190.8140.8230.7080.7210.8140.8250.8240.7110.7280.720
MHPL [45]CVPR'230.8090.8340.7080.7330.8100.8300.8380.7020.7330.737
DML-Core [43]NeurIPS'240.8110.8390.7170.7400.8120.8350.8420.7120.7420.743
Detective [61]CVPR'240.8140.8370.7170.7400.8140.8410.8390.7180.7480.741
STDR [47]IEEE TMI'240.8130.8370.7190.7380.8140.8330.8400.7150.7370.741
Our FairAP-0.828*0.843*0.731*0.747*0.827*0.8430.845*0.730*0.7510.748*
RimRandom-0.6870.7800.5670.6540.7330.7120.8000.6000.5770.676
Entropy [44]CVPR'190.6860.7720.5650.6430.7270.7090.7900.5950.5730.662
MHPL [45]CVPR'230.6880.7820.5670.6560.7290.7180.8010.5960.5810.678
DML-Core [43]NeurIPS'240.6970.7830.5780.6580.7380.7220.8010.6090.5890.677
Detective [61]CVPR'240.7000.7840.5820.6590.7400.7250.8010.6120.5930.678
STDR [47]IEEE TMI'240.6880.7810.5700.6560.7330.7130.8000.6040.5790.677
Our FairAP-0.711*0.791*0.592*0.667*0.750*0.735*0.807*0.624*0.602*0.686*
", + "image_path": "f0eaa3b3a342d688e188a5593b7b34b23d9d07197dab702c96ad8af50780fb95.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 354, + 294, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 354, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 294, + 496 + ], + "type": "text", + "content": "art DA models. For the rim segmentation, previous DA methods bring very limited performance improvements, and some (DAFormer [12]) even decline. For instance, as indicated in Table 1, the Baseline ES-IoU and IoU scores are 0.334 and 0.342 respectively, whereas DAFormer-FIA [41] achieved only 0.363 and 0.369. In contrast, our approach significantly outperforms them, achieving 0.592 and 0.667, respectively, which is approximately " + }, + { + "bbox": [ + 54, + 354, + 294, + 496 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 354, + 294, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 354, + 294, + 496 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 54, + 354, + 294, + 496 + ], + "type": "text", + "content": " higher. Overall, the previous fairness-aware DA method achieves fairness but with limited accuracy gains, while our FairAP ensures fairness and substantially boosts performance, underscoring its superiority." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 295, + 713 + ], + "type": "text", + "content": "With SOTA ADA Methods. We further perform extensive comparison of these attributes against the ADA methods: Random (randomly select samples), Entropy [44], MHPL [45], DML-Core [43], Detective [61], and STDR [47], with the same labeling quota for a fair comparison. The experimental results indicate that our approach surpasses existing ADA methods in segmentation performance and demonstrates significant improvements in fairness. For instance, when analyzing the Race attribute (in Table 1), our method's overall Dice score of 0.791 outperforms the STDR method's 0.781 by 1 percentage with statistically significant differences. However, when considering both performance and fairness, our ES-Dice score of 0.711 exceeds STDR's 0.688 by 2.3 percentages, demonstrating FairAP's ability to significantly enhance fairness while maintaining performance. These findings are also reflected in Table 1 and Table 3, which assess Race and Ethnicity attributes. Compared to other ADA methods, our approach not only excels in seg" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 316, + 374, + 550, + 443 + ], + "blocks": [ + { + "bbox": [ + 313, + 352, + 553, + 373 + ], + "lines": [ + { + "bbox": [ + 313, + 352, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 553, + 373 + ], + "type": "text", + "content": "Table 2. Ablation studies for VLM-learning on sensitive attributes. The * denotes " + }, + { + "bbox": [ + 313, + 352, + 553, + 373 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 352, + 553, + 373 + ], + "type": "text", + "content": "-value " + }, + { + "bbox": [ + 313, + 352, + 553, + 373 + ], + "type": "inline_equation", + "content": "< 0.05" + }, + { + "bbox": [ + 313, + 352, + 553, + 373 + ], + "type": "text", + "content": " in all paired " + }, + { + "bbox": [ + 313, + 352, + 553, + 373 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 352, + 553, + 373 + ], + "type": "text", + "content": "-test." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 374, + 550, + 443 + ], + "lines": [ + { + "bbox": [ + 316, + 374, + 550, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 550, + 443 + ], + "type": "table", + "html": "
AttributesMethodCUPRim
Fair-APVLMES-Dice ↑ES-IoU ↑ES-Dice ↑ES-IoU ↑
Race✓0.8110.7110.6940.575
✓✓0.828*0.731*0.711*0.592*
Ethnicity✓0.8210.7190.7810.660
✓✓0.836*0.730*0.784*0.663*
", + "image_path": "6db440a0f7045d5d83211b4b210f9c769fc2db1f0323a283f9782f0163f1db9c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 455, + 553, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 553, + 478 + ], + "type": "text", + "content": "mentation performance but also demonstrates effective improvements in fairness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 490, + 416, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 416, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 416, + 502 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 509, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 555, + 640 + ], + "type": "text", + "content": "Ablation Study on VLM-Learning. We conducted ablation studies to evaluate the impact of VLM-learning on model performance. In Table 2, we compared Fair-AP outcomes with and without VLM-learning under two sensitive attributes. The results demonstrate that VLM-learning significantly improves segmentation performance (" + }, + { + "bbox": [ + 312, + 509, + 555, + 640 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 312, + 509, + 555, + 640 + ], + "type": "text", + "content": "-value " + }, + { + "bbox": [ + 312, + 509, + 555, + 640 + ], + "type": "inline_equation", + "content": "< 0.05" + }, + { + "bbox": [ + 312, + 509, + 555, + 640 + ], + "type": "text", + "content": "). This statistically significant enhancement validates the effectiveness of VLM-learning in representing demographic attributes through cross-modal vision-language alignment, thereby enabling equitable performance across diverse subgroups." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Ablation Study on Fair-AP. We conducted ablation experiments to assess the impact of Fair-AP's design, focusing on four specific configurations: 1) Fair-Base, which evenly allocates quotas across subgroups (like White, Black, and Asian) and selects samples randomly. 2) Fair-Attr, which selects samples based on Attribute representativeness. 3)" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24108" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 95, + 528, + 337 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 3. Cup and rim segmentation performance on the FairDomain-Segmentation benchmark using different DA and ADA methods with Ethnicity as the demographic attribute. The * denotes " + }, + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "text", + "content": " -value " + }, + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "inline_equation", + "content": "< {0.05}" + }, + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "text", + "content": " in all paired " + }, + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "text", + "content": " -test,indicating statistically significant differences." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 95, + 528, + 337 + ], + "lines": [ + { + "bbox": [ + 82, + 95, + 528, + 337 + ], + "spans": [ + { + "bbox": [ + 82, + 95, + 528, + 337 + ], + "type": "table", + "html": "
MethodVenueOverall ES-Dice↑Overall Dice↑Overall ES-IoU↑Overall IoU↑Hispanic Dice↑Non-Hispanic Dice↑Hispanic IoU↑Non-Hispanic IoU↑
DACupBaseline (Source)-0.8710.8880.7840.8080.8870.9060.8070.839
Baseline (Target)-0.6840.7000.5510.5550.6990.7220.5630.587
RimBaseline (Source)-0.8450.8610.7430.7620.8600.8790.7610.786
Baseline (Target)-0.4890.4950.3400.3420.4950.5060.3420.348
DAPixMatch [28]CVPR'210.7400.7750.6280.6600.7630.8110.6490.699
DAFormer [12]CVPR'220.7730.7960.6540.6800.7940.8240.6770.717
DAFormer-FIA [41]ECCV'240.7900.8100.6740.7000.8080.8340.6970.734
RimPixMatch [28]CVPR'210.6430.6730.4990.5230.6730.7190.5130.562
DAFormer [12]CVPR'220.4050.4180.2640.2700.4160.4480.2690.295
DAFormer-FIA [41]ECCV'240.5160.5310.3590.3690.5290.5580.3680.395
CupRandom-0.8160.8340.7110.7340.8530.8310.7640.731
Entropy [44]CVPR'190.8010.8230.6930.7210.8480.8210.7580.717
MHPL [45]CVPR'230.8230.8340.7180.7330.8460.8320.7520.731
DML-Core [43]NeurIPS'240.8240.8390.7200.7400.8550.8360.7660.737
Detective [61]CVPR'240.8220.8370.7220.7400.8540.8360.7630.738
STDR [47]IEEE TMI'240.8120.8370.7100.7380.8650.8350.7740.735
ADAOur FairAP-0.836*0.8400.730*0.743*0.8440.8390.7590.741*
RimRandom-0.7750.7800.6510.6540.7860.7790.6570.653
Entropy [44]CVPR'190.7610.7720.6350.6430.7850.7700.6540.641
MHPL [45]CVPR'230.7740.7820.6500.6560.7900.7800.6620.654
DML-Core [43]NeurIPS'240.7750.7830.6530.6580.7910.7810.6620.655
Detective [61]CVPR'240.7770.7840.6560.6590.7920.7820.6620.658
STDR [47]IEEE TMI'240.7750.7810.6520.6560.7880.7790.6600.654
Our FairAP-0.784*0.790*0.663*0.668*0.796*0.788*0.673*0.666*
", + "image_path": "812d3b1322ab5f65dd7b8ce84be786a61a451438e1a923a7f0e09b1f23601b71.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 345, + 290, + 491 + ], + "blocks": [ + { + "bbox": [ + 61, + 345, + 290, + 491 + ], + "lines": [ + { + "bbox": [ + 61, + 345, + 290, + 491 + ], + "spans": [ + { + "bbox": [ + 61, + 345, + 290, + 491 + ], + "type": "image", + "image_path": "4a6e528e7aa9a6b94e0bd9cadec8eb436fbd6d7488ca5899858ecb8d9fe42336.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 495, + 295, + 540 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 540 + ], + "type": "text", + "content": "Figure 5. Top: bar charts showing Dice-related (left) and IoU-related (right) scores across four ablation models with race attribute. Bottom: radar charts showing Dice (left) and IoU (right) values for Cup and Rim regions, divided by racial subgroups." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 554, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 674 + ], + "type": "text", + "content": "Fair-Poly, which chooses samples for Polysemy representativeness. 4) Our Fair-AP, which considers the average division of Fair-Base and prioritizes sample selection with both Attribute and Polysemy. Fig. 5 demonstrates the ablation study results. Standalone Fair-Attr or Fair-Poly strategies may be one-sided and show unstable improvements, while our FairAP—jointly modeling Attribute and Polysemy achieves significant gains in all metrics. The radar chart further confirms this across specific sub-groups, demonstrating our design's effectiveness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Exploration on Labeling Quota " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha\\%" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ". We investigate the impact of varying labeling quotas " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha\\%" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " on Fair-AP's performance compared to some representative ADA methods. As" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 320, + 345, + 433, + 434 + ], + "blocks": [ + { + "bbox": [ + 320, + 345, + 433, + 434 + ], + "lines": [ + { + "bbox": [ + 320, + 345, + 433, + 434 + ], + "spans": [ + { + "bbox": [ + 320, + 345, + 433, + 434 + ], + "type": "image", + "image_path": "c866706a14acbd9849e5d095d4eeae3e49c18305200092a06c2e727a642e1b8c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 435, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 456 + ], + "type": "text", + "content": "Figure 6. Ablation studies of different labeling quota " + }, + { + "bbox": [ + 313, + 435, + 555, + 456 + ], + "type": "inline_equation", + "content": "\\alpha \\%" + }, + { + "bbox": [ + 313, + 435, + 555, + 456 + ], + "type": "text", + "content": " for Race attribute on the key composite metric ES-Dice." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 436, + 344, + 549, + 434 + ], + "blocks": [ + { + "bbox": [ + 436, + 344, + 549, + 434 + ], + "lines": [ + { + "bbox": [ + 436, + 344, + 549, + 434 + ], + "spans": [ + { + "bbox": [ + 436, + 344, + 549, + 434 + ], + "type": "image", + "image_path": "1c0ec4b432d0d1e94021d5fd0d55a46ddddad311130ef71b2c8c9656e595d294.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "type": "text", + "content": "shown in Fig. 6, increasing " + }, + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\alpha\\%" + }, + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "type": "text", + "content": " improves model accuracy with diminishing returns: significant gains emerge from " + }, + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "type": "inline_equation", + "content": "0\\%, \\rightarrow, 2\\%, \\rightarrow, 5\\%" + }, + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "type": "text", + "content": ", followed by marginal improvements beyond " + }, + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "type": "text", + "content": ". Crucially, our method consistently achieves superior performance across all quotas, justifying " + }, + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\alpha = 5\\%" + }, + { + "bbox": [ + 313, + 476, + 555, + 548 + ], + "type": "text", + "content": " as the optimal accuracy-cost balance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 388, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 388, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 388, + 582 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "In this work, we pioneer a novel fairness-aware ADA paradigm to address fairness in real-world cross-domain medical scenarios. By utilizing the VLM to perceive demographic attributes and devising the attribute-aware FairAP strategy, we significantly enhance both fairness and segmentation performance. Extensive experiments confirm that our approach surpasses existing DA and ADA methods, demonstrating superior performance and enhanced fairness, providing a fresh perspective on addressing fairness challenges toward more equitable AI-assisted healthcare." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24109" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "This work is supported by the Guangdong Science and Technology Department (2024ZDZX2004), a General Research Fund of Hong Kong Research Grants Council (project no. 15218521), and a grant under the scheme of Collaborative Research with World-leading Research Groups in The Hong Kong Polytechnic University (project no. G-SACF)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "spans": [ + { + "bbox": [ + 56, + 187, + 115, + 201 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 207, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 207, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 207, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 207, + 296, + 273 + ], + "type": "text", + "content": "[1] Ehsan Adeli, Qingyu Zhao, Adolf Pfefferbaum, Edith V Sullivan, Li Fei-Fei, Juan Carlos Niebles, and Kilian M Pohl. Representation learning with statistical independence to mitigate bias. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2513–2523, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 274, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 306 + ], + "type": "text", + "content": "[2] David Arthur and Sergei Vassilvitskii. k-means++: The advantages of careful seeding. Technical report, Stanford, 2006. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 308, + 296, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 308, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 61, + 308, + 296, + 362 + ], + "type": "text", + "content": "[3] Jonathan T Barron. A generalization of otsu's method and minimum error thresholding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part V 16, pages 455-470. Springer, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 365, + 296, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 365, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 61, + 365, + 296, + 420 + ], + "type": "text", + "content": "[4] Peter J Bevan and Amir Atapour-Abarghoueiet. Detecting melanoma fairly: Skin tone detection and debiasing for skin lesion classification. In MICCAI Workshop on Domain Adaptation and Representation Transfer, pages 1-11. Springer, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 421, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 296, + 465 + ], + "type": "text", + "content": "[5] Alceu Bissoto, Michel Fornaciali, Eduardo Valle, and Sandra Avila. (de) constructing bias on skin lesion datasets. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 467, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 296, + 498 + ], + "type": "text", + "content": "[6] Simon Caton and Christian Haas. Fairness in machine learning: A survey. ACM Computing Surveys, 56(7):1-38, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 501, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 296, + 555 + ], + "type": "text", + "content": "[7] Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 557, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 612 + ], + "type": "text", + "content": "[8] Jiayi Chen, Benteng Ma, Hengfei Cui, and Yong Xia. Think twice before selection: Federated evidential active learning for medical image analysis with domain shifts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11439-11449, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 613, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 645 + ], + "type": "text", + "content": "[9] Julia Dressel and Hany Farid. The accuracy, fairness, and limits of predicting recidivism. Science advances, 4(1): eaao5580, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "text", + "content": "[10] Judy Wawira Gichoya, Imon Banerjee, Ananth Reddy Bhimireddy, John L Burns, Leo Anthony Celi, Li-Ching Chen, Ramon Correa, Natalie Dullerud, Marzyeh Ghassemi, Shih-Cheng Huang, et al. Ai recognition of patient race in medical imaging: a modelling study. The Lancet Digital Health, 4(6):e406-e414, 2022. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[11] Yuanbo Guo, Zheyu Yan, Xiaoting Yu, Qingpeng Kong, Joy Xie, Kevin Luo, Dewen Zeng, Yawen Wu, Zhenge Jia, and Yiyu Shi. Hardware design and the fairness of a neural network. Nature Electronics, 7(8):714-723, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 172 + ], + "type": "text", + "content": "[12] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9924-9935, 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 173, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 555, + 228 + ], + "type": "text", + "content": "[13] Mohamed Huti, Tiarna Lee, Elinor Sawyer, and Andrew P King. An investigation into race bias in random forest models based on breast dce-mri derived radiomics features. In Workshop on Clinical Image-Based Procedures, pages 225-234. Springer, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 554, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 270 + ], + "type": "text", + "content": "[14] Suyog Dutt Jain and Kristen Grauman. Active image segmentation propagation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2864-2873, 2016. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "text", + "content": "[15] Ruinan Jin, Zikang Xu, Yuan Zhong, Qionsong Yao, Qi Dou, S Kevin Zhou, and Xiaoxiao Li. Fairmedfm: fairness benchmarking for medical imaging foundation models. arXiv preprint arXiv:2407.00983, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 554, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 554, + 371 + ], + "type": "text", + "content": "[16] Thorsten Kalb, Kaisar Kushibar, Celia Cintas, Karim Lekadir, Oliver Diaz, and Richard Osuala. Revisiting skin tone fairness in dermatological lesion classification. In Workshop on Clinical Image-Based Procedures, pages 246-255. Springer, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 415 + ], + "type": "text", + "content": "[17] Michael P Kim, Amirata Ghorbani, and James Zou. Multiaccuracy: Black-box post-processing for fairness in classification. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, pages 247-254, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 416, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 437 + ], + "type": "text", + "content": "[18] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 438, + 554, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 554, + 504 + ], + "type": "text", + "content": "[19] Newton M Kinyanjui, Timothy Odonga, Celia Cintas, Noel CF Codella, Rameswar Panda, Prasanna Sattigeri, and Kush R Varshney. Fairness of classifiers across skin tones in dermatology. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 320-329. Springer, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 505, + 554, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 554, + 569 + ], + "type": "text", + "content": "[20] Qingpeng Kong, Ching-Hao Chiu, Dewen Zeng, Yu-Jen Chen, Tsung-Yi Ho, Jingtong Hu, and Yiyu Shi. Achieving fairness through channel pruning for dermatological disease diagnosis. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 24-34. Springer, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 571, + 554, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 554, + 603 + ], + "type": "text", + "content": "[21] David D Lewis and Jason Catlett. Heterogeneous uncertainty sampling for supervised learning. In Machine learning proceedings 1994, pages 148-156. Elsevier, 1994. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 604, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 554, + 669 + ], + "type": "text", + "content": "[22] Pranay K Lohia, Karthikeyan Natesan Ramamurthy, Manish Bhide, Diptikalyan Saha, Kush R Varshney, and Ruchir Puri. Bias mitigation post-processing for individual and group fairness. In Icassp 2019-2019 iee international conference on acoustics, speech and signal processing (icassp), pages 2847-2851. IEEE, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 714 + ], + "type": "text", + "content": "[23] Xiangde Luo, Hongqiu Wang, Jinfeng Xu, Lu Li, Yue Zhao, Yuan He, Hui Huang, Jianghong Xiao, Tao Song, Shichuan Zhang, et al. Generalizable magnetic resonance imaging-based nasopharyngeal carcinoma delineation: Bridging gaps" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24110" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 105 + ], + "type": "text", + "content": "across multiple centers and raters with active learning. International Journal of Radiation Oncology* Biology* Physics, 121(5):1384-1393, 2025. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 172 + ], + "type": "text", + "content": "[24] Yan Luo, Min Shi, Muhammad Osama Khan, Muhammad Muneeb Afzal, Hao Huang, Shuaihang Yuan, Yu Tian, Luo Song, Ava Kouhana, Tobias Elze, et al. Fairclip: Harnessing fairness in vision-language learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12289-12301, 2024. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "type": "text", + "content": "[25] Yan Luo, Yu Tian, Min Shi, Louis R Pasquale, Lucy Q Shen, Nazlee Zebardast, Tobias Elze, and Mengyu Wang. Harvard glaucoma fairness: a retinal nerve disease dataset for fairness learning and fair identity normalization. IEEE Transactions on Medical Imaging, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "text", + "content": "[26] Ninareh Mehrabi, Fred Morstatter, Nripsuta Saxena, Kristina Lerman, and Aram Galstyan. A survey on bias and fairness in machine learning. ACM computing surveys (CSUR), 54 (6):1-35, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 327 + ], + "type": "text", + "content": "[27] Ke Mei, Chuang Zhu, Jiaqi Zou, and Shanghang Zhang. Instance adaptive self-training for unsupervised domain adaptation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 415-430. Springer, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 380 + ], + "type": "text", + "content": "[28] Luke Melas-Kyriazi and Arjun K Manrai. Pixmatch: Unsupervised domain adaptation via pixelwise consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12435-12445, 2021. 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 447 + ], + "type": "text", + "content": "[29] Esther Puyol-Anton, Bram Ruijsink, Jorge Mariscal Harana, Stefan K Piechnik, Stefan Neubauer, Steffen E Petersen, Reza Razavi, Phil Chowienczyk, and Andrew P King. Fairness in cardiac magnetic resonance imaging: assessing sex and racial bias in deep learning-based segmentation. Frontiers in cardiovascular medicine, 9:859310, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 513 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 558 + ], + "type": "text", + "content": "[31] Fernanda Ribeiro, Valentina Shumovskaia, Thomas Davies, and Ira Ktena. How fair is your graph? exploring fairness concerns in neuroimaging studies. In Machine Learning for Healthcare Conference, pages 459-478. PMLR, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 601 + ], + "type": "text", + "content": "[32] María Agustina Ricci Lara, Rodrigo Echeveste, and Enzo Ferrante. Addressing fairness in artificial intelligence for medical imaging. nature communications, 13(1):4581, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 678 + ], + "type": "text", + "content": "[33] Zohaib Salahuddin, Yi Chen, Xian Zhong, Henry C Woodruff, Nastaran Mohammadian Rad, Shruti Atul Mali, and Philippe Lambin. From head and neck tumour and lymph node segmentation to survival prediction on pet/ct: an end-to-end framework featuring uncertainty, fairness, and multi-region multi-modal radiomics. Cancers, 15(7):1932, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[34] Candice Schwartz, Ifeanyi Beverly Chukwudozie, Silvia Tejeda, Ganga Vijayasiri, Ivy Abraham, Mylene Remo, Hiral A Shah, Maria Rojas, Alicia Carillo, Loraine Moreno," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "et al. Association of population screening for breast cancer risk with use of mammography among women in medically underserved racial and ethnic minority groups. JAMA network open, 4(9):e2123751-e2123751, 2021. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[35] Laleh Seyyed-Kalantari, Guanxiong Liu, Matthew McDermott, Irene Y Chen, and Marzyeh Ghassemi. Chexclusion: Fairness gaps in deep chest x-ray classifiers. In BIOCOMPUTING 2021: proceedings of the Pacific symposium, pages 232-243. World Scientific, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 173, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 227 + ], + "type": "text", + "content": "[36] Laleh Seyyed-Kalantari, Haoran Zhang, Matthew BA McDermott, Irene Y Chen, and Marzyeh Ghassemi. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature medicine, 27(12):2176-2182, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 228, + 555, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 259 + ], + "type": "text", + "content": "[37] Dinggang Shen, Guorong Wu, and Heung-II Suk. Deep learning in medical image analysis. Annual review of biomedical engineering, 19(1):221-248, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 261, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 327 + ], + "type": "text", + "content": "[38] Min Shi, Muhammad Muneeb Afzal, Hao Huang, Congcong Wen, Yan Luo, Muhammad Osama Khan, Yu Tian, Leo Kim, Tobias Elze, Yi Fang, et al. Equitable deep learning for diabetic retinopathy detection using multi-dimensional retinal imaging with fair adaptive scaling: a retrospective study. medRxiv, pages 2024-04, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "type": "text", + "content": "[39] Emma AM Stanley, Matthias Wilms, Pauline Mouches, and Nils D Forkert. Fairness-related performance and explainability effects in deep learning models for brain image analysis. Journal of Medical Imaging, 9(6):061102-061102, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 383, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 436 + ], + "type": "text", + "content": "[40] Yu Tian, Min Shi, Yan Luo, Ava Kouhana, Tobias Elze, and Mengyu Wang. Fairseg: A large-scale medical image segmentation dataset for fairness learning using segment anything model with fair error-bound scaling. In ICLR, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 437, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 502 + ], + "type": "text", + "content": "[41] Yu Tian, Congcong Wen, Min Shi, Muhammad Muneeb Afzal, Hao Huang, Muhammad Osama Khan, Yan Luo, Yi Fang, and Mengyu Wang. Fairdomain: Achieving fairness in cross-domain medical image segmentation and classification. In European Conference on Computer Vision, pages 251-271. Springer, 2024. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 504, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 568 + ], + "type": "text", + "content": "[42] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7472-7481, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "text", + "content": "[43] Arvind Vepa, Zukang Yang, Andrew Choi, Jungseock Joo, Fabien Scalzo, and Yizhou Sun. Integrating deep metric learning with coreset for active learning in 3d segmentation. Advances in Neural Information Processing Systems, 37:71643-71671, 2024. 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 689 + ], + "type": "text", + "content": "[44] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 2, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[45] Fan Wang, Zhongyi Han, Zhiyan Zhang, Rundong He, and Yilong Yin. Mhpl: Minimum happy points learning for" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24111" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 106 + ], + "type": "text", + "content": "active source free domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20008-20018, 2023. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 148 + ], + "type": "text", + "content": "[46] Hongqiu Wang, Yueming Jin, and Lei Zhu. Dynamic interactive relation capturing via scene graph learning for robotic surgical report generation. arXiv preprint arXiv:2306.02651, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 215 + ], + "type": "text", + "content": "[47] Hongqiu Wang, Jian Chen, Shichen Zhang, Yuan He, Jinfeng Xu, Mengwan Wu, Jinlan He, Wenjun Liao, and Xiangde Luo. Dual-reference source-free active domain adaptation for nasopharyngeal carcinoma tumor segmentation across multiple hospitals. IEEE Transactions on Medical Imaging, 2024. 3, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 281 + ], + "type": "text", + "content": "[48] Hongqiu Wang, Xiangde Luo, Wu Chen, Qingqing Tang, Mei Xin, Qiong Wang, and Lei Zhu. Advancing uwf-slo vessel segmentation with source-free active domain adaptation and a novel multi-center dataset. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 75-85. Springer, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 348 + ], + "type": "text", + "content": "[49] Hongqiu Wang, Zhaohu Xing, Weitong Wu, Yijun Yang, Qingqing Tang, Meixia Zhang, Yanwu Xu, and Lei Zhu. Non-invasive to invasive: Enhancing ffa synthesis from cfp with a benchmark dataset and a novel network. In Proceedings of the 1st International Workshop on Multimedia Computing for Health and Medicine, pages 7-15, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "text", + "content": "[50] Hongqiu Wang, Guang Yang, Shichen Zhang, Jing Qin, Yike Guo, Bo Xu, Yueming Jin, and Lei Zhu. Video-instrument synergistic network for referring video instrument segmentation in robotic surgery. IEEE Transactions on Medical Imaging, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 458 + ], + "type": "text", + "content": "[51] Hongqiu Wang, Yixian Chen, Wu Chen, Huihui Xu, Haoyu Zhao, Bin Sheng, Huazhu Fu, Guang Yang, and Lei Zhu. Serp-mamba: Advancing high-resolution retinal vessel segmentation with selective state-space model. IEEE Transactions on Medical Imaging, 2025. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 491 + ], + "type": "text", + "content": "[52] Tianxiang Wang and Qun Dai. A patch distribution-based active learning method for multiple instance alzheimer's disease diagnosis. Pattern Recognition, 150:110341, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 557 + ], + "type": "text", + "content": "[53] Zhibo Wang, Xiaowei Dong, Henry Xue, Zhifei Zhang, Weifeng Chiu, Tao Wei, and Kui Ren. Fairness-aware adversarial perturbation towards bias mitigation for deployed deep models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10379-10388, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "type": "text", + "content": "[54] Garrett Wilson and Diane J Cook. A survey of unsupervised deep domain adaptation. ACM Transactions on Intelligent Systems and Technology (TIST), 11(5):1-46, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 593, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 645 + ], + "type": "text", + "content": "[55] Wenshuai Xu, Zhenghui Hu, Yu Lu, Jinzhou Meng, Qingjie Liu, and Yunhong Wang. Activedc: Distribution calibration for active finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16996-17005, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[56] Zhao Xu, Kai Yu, Volker Tresp, Xiaowei Xu, and Jizhi Wang. Representative sampling for text classification using support vector machines. In Advances in Information Retrieval: 25th European Conference on IR Research, ECIR 2003, Pisa, Italy, April 14-16, 2003. Proceedings 25, pages 393-407. Springer, 2003. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 519 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[57] Zikang Xu, Jun Li, Qingsong Yao, Han Li, Mingyue Zhao, and S Kevin Zhou. Addressing fairness issues in deep learning-based medical image analysis: a systematic review. npj Digital Medicine, 7(1):286, 2024. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[58] Jenny Yang, Andrew AS Soltan, David W Eyre, Yang Yang, and David A Clifton. An adversarial training framework for mitigating algorithmic biases in clinical machine learning. NPJ digital medicine, 6(1):55, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[59] Ruichen Yao, Ziteng Cui, Xiaoxiao Li, and Lin Gu. Improving fairness in image classification via sketching. arXiv preprint arXiv:2211.00168, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[60] Chenxi Yuan, Kristin A Linn, and Rebecca A Hubbard. Algorithmic fairness of machine learning models for alzheimer disease progression. JAMA Network Open, 6(11):e2342203-e2342203, 2023. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 307 + ], + "type": "text", + "content": "[61] Wenqiao Zhang, Zheqi Lv, Hao Zhou, Jia-Wei Liu, Juncheng Li, Mengze Li, Yunfei Li, Dongping Zhang, Yueting Zhuang, and Siliang Tang. Revisiting the domain shift and sample uncertainty in multi-source active domain transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16751-16761, 2024. 3, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[62] Zhifei Zhang, Yang Song, and Hairong Qi. Age progression/regression by conditional adversarial autoencoder. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5810-5818, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "text", + "content": "[63] Haoyu Zhao, Hao Wang, Xingyue Zhao, Hongqiu Wang, Zhiyu Wu, Chengjiang Long, and Hua Zou. Automated 3d physical simulation of open-world scene with gaussian splatt-ing. arXiv e-prints, pages arXiv-2411, 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 398, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 430 + ], + "type": "text", + "content": "[64] Qingyu Zhao, Ehsan Adeli, and Kilian M Pohl. Training confounder-free deep learning models for medical applications. Nature communications, 11(1):6010, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 432, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 464 + ], + "type": "text", + "content": "[65] Yongshuo Zong, Yongxin Yang, and Timothy Hospedales. Medfair: Benchmarking fairness for medical imaging. arXiv preprint arXiv:2210.01725, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 466, + 553, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 519 + ], + "type": "text", + "content": "[66] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24112" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/331fec76-66a3-4c9b-96fc-eeb1ae13b1b6_content_list.json b/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/331fec76-66a3-4c9b-96fc-eeb1ae13b1b6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7a79d22da1a359439abc3db7847d9339730d3bf2 --- /dev/null +++ b/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/331fec76-66a3-4c9b-96fc-eeb1ae13b1b6_content_list.json @@ -0,0 +1,1868 @@ +[ + { + "type": "text", + "text": "Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts", + "text_level": 1, + "bbox": [ + 259, + 130, + 738, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chiao-An Yang $^{1}$ Kuan-Chuan Peng $^{2}$ Raymond A. Yeh $^{1}$", + "bbox": [ + 256, + 202, + 736, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Computer Science, Purdue University $^{2}$ Mitsubishi Electric Research Laboratories", + "bbox": [ + 104, + 222, + 890, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1{yang2300,rayyeh}@purdue.edu 2kpeng@merl.com", + "bbox": [ + 290, + 241, + 700, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 295, + 326, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anomaly detection (AD) identifies the defect regions of a given image. Recent works have studied AD, focusing on learning AD without abnormal images, with long-tailed distributed training data, and using a unified model for all classes. In addition, online AD learning has also been explored. In this work, we expand in both directions to a realistic setting by considering the novel task of long-tailed online AD (LTOAD). We first identified that the offline state-of-the-art LTAD methods cannot be directly applied to the online setting. Specifically, LTAD is class-aware, requiring class labels that are not available in the online setting. To address this challenge, we propose a class-agnostic framework for LTAD and then adapt it to our online learning setting. Our method outperforms the SOTA baselines in most offline LTAD settings, including both the industrial manufacturing and the medical domain. In particular, we observe $+4.63\\%$ image-AUROC on MVTec even compared to methods that have access to class labels and the number of classes. In the most challenging long-tailed online setting, we achieve $+0.53\\%$ image-AUROC compared to baselines. Our LTOAD benchmark is released here1.", + "bbox": [ + 91, + 327, + 483, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 672, + 220, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Anomaly detection (AD) [8, 65, 96] is the task of identifying defect regions in a given image. This task is important due to its high valued applications in industrial manufacturing [1, 5, 18, 47, 58, 74, 86, 103] and medical [28, 42, 43, 47] settings.", + "bbox": [ + 89, + 696, + 482, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, LTAD [34] explored the long-tailed (LT) [92] setting of unsupervised one-class AD, where abnormal images are not given during training and the training set is unevenly distributed. This long-tailed distribution assumes that images of certain classes significantly outnumber others, which is a realistic assumption for real-world applications. Importantly, LTAD shows that prior work on evenly", + "bbox": [ + 89, + 772, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "distributed data performs suboptimally on LT data. Meanwhile, recent works [16, 22, 55, 61, 79, 84] approach AD with online learning to leverage abnormal images during test time. However, they do not study the case where the classes follow a long-tail distribution, where the performance gap of head and tail classes in the online setting may be significant during offline training. This motivates us to study the long-tailed online AD (LTOAD) task by proposing a benchmark and a novel model.", + "bbox": [ + 511, + 296, + 903, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "At first glance, it may seem that one can take a state-of-the-art (SOTA) long-tailed offline model, e.g., LTAD [34], and combine it with online learning to solve LTOAD. However, several limitations within the LTAD's framework prevent it from being effective in the online setting. First, LTAD is a class-aware method, where they assume that the class information is given. However, class labels are not available in the online setting. Instead, a class-agnostic method is required as illustrated in Fig. 1. Second, LTAD's encoder-decoder architecture does not leverage the latest vector quantization VAE [20, 57, 59], which is more effective, as shown in recent work. Third, the prompt learning designs of LTAD and other AD works [7, 9, 10, 37, 40, 51, 52, 67, 100–102] ignore some aspects of abnormal prompts, e.g., LTAD does not support class-specific abnormal prompts.", + "bbox": [ + 511, + 436, + 905, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address these challenges, we propose a novel model that is class-agnostic, with a vector quantized VAE, and a comprehensive prompt learning framework, for the task of long-tailed online anomaly detection. Experiments show that our method outperforms the SOTA baselines significantly by at least $4\\%$ on several detection benchmarks without access to the class labels and the number of classes.", + "bbox": [ + 511, + 681, + 905, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the following sections, we first discuss how to make existing class-aware methods class-agnostic, specifically on LTAD (Sec. 2). Next, we define the task of long-tailed online AD and present our solution (Sec. 3). Finally, we provide details of each proposed module (Sec. 4), followed by the experimental results (Sec. 5) and related works (Sec. 6).", + "bbox": [ + 511, + 791, + 903, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 511, + 885, + 836, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "MethodCAexp100exp200step100step200Det.Seg.Det.Seg.Det.Seg.Det.Seg.RegAD [36]X82.4395.20N/AN/A81.5495.10N/AN/AAnomalyGPT [25]X87.4489.6885.8090.1585.9589.2882.4789.45LTAD [34]X88.8694.4686.0594.1887.3693.8385.6092.12HVQ [57]X87.4395.2585.3494.7985.3994.1783.0092.61LTOAD*X93.1295.0191.8894.3692.0294.7287.9593.34UniAD [89]✓87.7093.9586.2193.2683.3791.4781.3292.29MoEAD [62]✓84.7394.3483.2393.9684.4093.7683.1392.76LTOAD✓93.4295.2192.0294.9492.3395.1188.6294.00", + "bbox": [ + 93, + 89, + 478, + 218 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/aad2d68b0c8382742b071dcd0d7e31e216eac6033e1694a018425ea562374ccd.jpg", + "table_caption": [ + "Table 2. Comparison (↑) on offline MVTec in image-level AUROC for anomaly detection (Det.) and pixel-level AUROC for anomaly segmentation (Seg.). The column CA (class-agnostic) indicates whether a method requires class names or the number of classes during training or not (require: X; not require: √). RegAD needs at least 2 images per class and thus is not applicable on exp200 and step200. We mark the best and second best performances in bold and underline." + ], + "table_footnote": [], + "table_body": "
MethodCAexp100exp200step100step200
Det.Seg.Det.Seg.Det.Seg.Det.Seg.
RegAD [36]X71.3694.4071.3694.4071.8094.9971.6594.52
AnomalyGPT [25]X70.3480.3269.7879.4871.9882.3069.7881.97
LTAD [34]X80.0095.5680.2195.3684.8096.5784.0396.27
HVQ [57]X78.6396.1878.0395.3276.6895.6375.5295.50
LTOAD*X84.8996.4284.8496.4487.7197.0786.5296.89
UniAD [89]✓77.3195.0376.8794.8078.8396.0477.6495.66
MoEAD [62]✓73.8994.3473.5893.9874.1294.3776.7794.82
LTOAD✓85.2696.9185.2496.9788.3297.5487.0497.23
", + "bbox": [ + 93, + 344, + 480, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines & implementation details. We compare with RegAD [36], UniAD [89], and LTAD [34]. We benchmark the SOTA MoEAD [62] and HVQ [57] to highlight the gain of our concept-specific quantization modules to their class-specific ones. Only UniAD, MoEAD, and LTOAD are class-agnostic and do not need a class label at the test time. We provide a class-aware version of our model, LTOAD*.", + "bbox": [ + 89, + 516, + 483, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative & qualitative comparison. Tab. 2 and 3 show the comparisons of LTOAD with the baselines on MVTec and VisA. We observe that LTOAD outperforms the SOTA on most settings while using no class information. LTOAD achieves a larger gain over Det. On average across all settings, LTOAD achieves $+4.63\\% / +0.58\\%$ gain versus the second best baseline excluding LTOAD* on MVTec Det. / Seg. As for VisA, LTOAD achieves $+4.30\\% / +1.29\\%$ gain.", + "bbox": [ + 89, + 623, + 483, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, comparing LTAD and LTOAD*, we find that being class-agnostic is beneficial. We hypothesize that this is because the classes can now share knowledge through the soft labels. Finally, we show qualitative results in Fig. 3. We observe that LTOAD predictions are closer to ground truth than HVQ and LTAD.", + "bbox": [ + 89, + 760, + 483, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Online long-tailed benchmark", + "text_level": 1, + "bbox": [ + 89, + 862, + 357, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment setup. For the online setting, we report on the", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a6e92571081f1ab204c28fdef9b447b672f5b0d207bcde266f18cf7ecfd61254.jpg", + "image_caption": [ + "truth anomaly masks are circled in pink." + ], + "image_footnote": [], + "bbox": [ + 514, + 85, + 996, + 305 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b66fc2cdead360336d356ec8f90fd310af447ada72f49e9b9c02cc0995c33166.jpg", + "table_caption": [ + "Table 3. Comparison (↑) on offline VisA. Format follows Tab. 2." + ], + "table_footnote": [], + "table_body": "
MethodCA OnlineBB-HFB-TFD2-HFD2-TFD5-HFD5-TFD5-MAvg.
LTAD [34]XX93.4693.4693.4693.4693.4693.4693.4693.4693.46
LTAD [34]XA^N94.1293.8094.0793.7094.0893.4993.8493.6593.84
HVQ [57]XX95.0295.0295.0295.0295.0295.0295.0295.0295.02
HVQ [57]XA^N95.4495.3395.2895.2295.2995.0395.2195.0595.23
LTOAD✓X95.0895.0895.0895.0895.0895.0895.0895.0895.08
LTOAD✓A^N95.4195.2295.3295.2395.0494.9194.7294.7995.08
LTOAD✓A^N95.3695.3095.3295.2395.2895.2695.2595.2195.28
", + "bbox": [ + 514, + 335, + 903, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Comparison (↑) on online MVTec in pixel-level AU-ROC for anomaly segmentation across 8 configurations of $\\mathcal{D}^0$ . All models are pre-trained on MVTec exp100 [34].", + "bbox": [ + 511, + 449, + 906, + 491 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2bb728d42b8f60bad3a2ff0ec501b37237e7cf22da23bc68b962552bdccf7d09.jpg", + "image_caption": [ + "Figure 4. Performance curve of $\\mathcal{A}^{\\mathrm{N}}$ and $\\mathcal{A}^{\\mathrm{AA}}$ in pixel-level AUROC on $\\mathcal{C}^{\\mathrm{head}}$ and $\\mathcal{C}^{\\mathrm{tail}}$ . They are offline-trained on $exp100$ [34] and online-trained on D5-HF or D5-TF." + ], + "image_footnote": [], + "bbox": [ + 517, + 493, + 901, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "same datasets as the offline benchmark [34], i.e., MVtec, VisA, and DAGM. We split the testing set into $\\mathcal{D}^0$ and $\\mathcal{D}^{\\mathrm{E}}$ with a 3:7 ratio. For each dataset, we create the 8 settings defined in Tab. 1.", + "bbox": [ + 511, + 702, + 906, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For baseline comparisons, we extend HVQ and LTAD to have online capability by applying our naive online learning algorithm $\\mathcal{A}^{\\mathsf{N}}$ since they are the best offline models on MVTec. Note that both methods are not class-agnostic and needs class labels as extra information in online learning, which is not available/necessary for our approach.", + "bbox": [ + 511, + 763, + 905, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison. We report the online Seg. performance at final time step $t = T$ on MVtec exp100 in Tab. 4. The results on other settings/datasets are in the", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23424", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1251b746d9d96673fabad7953842162f0d9ce57c87e6a168c35246692c41383c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodOnlineMVTecVisADAGM
→ VisA→ DAGM→ MVTec→ DAGM→ MVTec→ VisA
LTOADX84.3891.2782.7379.1679.5181.13
LTOAD✓92.3796.8789.8496.7988.3792.41
", + "bbox": [ + 91, + 88, + 480, + 151 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ab40cf4e8a3475732ea536bac2ecb0b6f53cdd46304ffe09ee63b37730248803.jpg", + "table_caption": [ + "Table 5. Cross-dataset anomaly segmentation in pixel-level AUROC. All models are pre-trained on their $exp100$ [34]." + ], + "table_footnote": [], + "table_body": "
Set IDDomainSizeElements
CN/A15zipper, pill, capsule, grid, transistor, carpet, metal nut, wood, leather, screw, tile, cable, toothbrush, hazelnut, bottle*
RandomOut10refinery, finger, vault, exam, salad, sailor, dove, dealer, well, humanoid
SmallIn3semiconductor, zipper, beech
LTOAD CIn10semiconductor, zipper, beech, microscopy*, antibiotics, medicines, circuit , mahogany, hardwood, walnut
", + "bbox": [ + 91, + 196, + 478, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Appendix A4. Overall, $\\mathcal{A}^{\\mathrm{N}}$ performs well in the simplest setting $(B)$ , but degrades on more difficult settings, e.g., $D2$ -TF, $D5$ -HF, etc. On average, LTOAD with $\\mathcal{A}^{\\mathrm{AA}}$ performs the best even without class labels as extra information. We plot LTOAD's performance curve of $\\mathsf{F}_{\\theta_t}$ over $t$ with $\\Delta = 16$ under the two most difficult settings ( $D5$ -HF and $D5$ -TF) in Fig. 4, where $\\mathcal{A}^{\\mathrm{N}}$ falls off on $\\mathcal{C}^{\\mathrm{head}} / \\mathcal{C}^{\\mathrm{head}}$ under $D5$ -HF/D5-TF while $\\mathcal{A}^{\\mathrm{AA}}$ improves steadily. Empirically, LTOAD is also more efficient than LTAD because $\\widehat{K} < K$ . Cross-dataset evaluation. Tab. 5 shows that LTOAD can adapt to unseen domains even if it is not designed for cross-dataset evaluation. For each $\\mathcal{D}_1 \\to \\mathcal{D}_2$ experiment, we pre-train LTOAD on $exp100$ of $\\mathcal{D}_1$ then evaluated on $B$ of $\\mathcal{D}_2$ without and with online learning. LTOAD with online learning outperforms the one without it.", + "bbox": [ + 88, + 428, + 483, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Additional analysis", + "text_level": 1, + "bbox": [ + 89, + 665, + 276, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on initialization of $\\widehat{C}$ . We now analyze the effect of our concept learning component. All experiments are done under MVtec exp100 and $B$ . We highlight the importance of our concept learning compared to two baselines: initializing with a random concept set or with a small one.", + "bbox": [ + 89, + 686, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 6 shows the learned initialization of $\\widehat{C}$ versus the ground truth $C$ . We observe that our concept learning can find the $\\widehat{C}$ initialization which represents similar semantics in $C$ , e.g., the bottle* images (see the 1st image in Fig. 3) in MVtec are bottom-only views, which look like an microscopy*.", + "bbox": [ + 89, + 763, + 482, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we experiment with the two baselines to generate initialization for $\\widehat{\\mathcal{C}}$ . In Random of Tab. 6, instead of learning from $\\mathcal{D}^{\\mathrm{T}}$ , i.e., in-domain concepts, we randomly select 10", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/033b6727d68c9985376f027ff31bf91f9b1988de76c851f56cdd3b13eb999560.jpg", + "table_caption": [ + "Table 6. Classes vs. concepts on MVTec. Words with similar semantics have the same color, e.g., wood-related words are colored in brown and electronics-related words are colored in purple." + ], + "table_footnote": [], + "table_body": "
Exp.\\( \\widehat{C} \\)\\( \\widehat{K} \\)RSGDetectionSegmentation
\\( C^{tail} \\)\\( C^{head} \\)\\( C \\)\\( C^{tail} \\)\\( C^{head} \\)\\( C \\)
RandomO-D10✓✓✓84.7399.3191.5494.4695.3994.89
SmallI-D3✓✓✓85.6698.8991.8394.2495.3094.74
NoRI-D10✘✓✘53.8456.2554.9760.7363.4662.00
NoSI-D10✓✘✓85.1699.2191.7294.5795.2494.88
NoGI-D10✓✓✘85.6399.5692.1394.0895.6294.80
LTOADI-D10✓✓✓88.1499.4593.7294.8395.6495.36
", + "bbox": [ + 514, + 89, + 901, + 219 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f548e723d6ec71b069ea1e3032c668d55e3dd98fb3d14ff58ac53098645ac9d3.jpg", + "table_caption": [ + "Table 7. Analysis and ablation on LTOAD using the same metrics as Tab. 4 under MVTec $\\exp {100}$ [34] and $B$ . We report the performances on ${\\mathcal{C}}^{\\text{head }}$ and ${\\mathcal{C}}^{\\text{tail }}$ . I-D/O-D: in/out-domain." + ], + "table_footnote": [], + "table_body": "
Exp.C̅LPaL MCSDetectionSegmentation
CtailCheadCCtailCheadC
FixCX√√√84.6598.1390.9494.57
FixP√X√√85.5999.4592.0694.52
SingleP√√X√85.4199.3791.9294.64
ShareP√√√X87.0399.3592.7895.00
LTOAD√√√√88.1499.4593.7294.83
", + "bbox": [ + 514, + 275, + 903, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Ablations on concept $\\hat{C}$ & abnormal prompt $\\mathcal{P}^{\\mathrm{a}}$ learning under MVtec exp100 [34] on the same metrics as Tab. 7. Acronyms: L: learned; M: multiple; CS: concept-specific.", + "bbox": [ + 511, + 404, + 906, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "nouns in ALIGN's $\\mathcal{V}$ unrelated to the industrial settings of MVTec, i.e., out-domain concepts. In other words, the semantics of $\\mathcal{C}$ and Random do not overlap. In Small of Tab. 6, instead of choosing $\\widehat{K} = 10$ , we select only the top- $\\widehat{K} = 3$ vocabularies as $\\widehat{\\mathcal{C}}$ . The semantics of Small partially covers the semantics of $\\mathcal{C}$ . Tab. 7 shows the quantitative results.", + "bbox": [ + 511, + 460, + 906, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We report the performance of Random and Small in Tab. 7. Even with an out-domain initialization, compared to HVQ in Tab. 2, which uses $\\mathcal{C}$ for their class-specific quantization modules, Random still performs competitively. Specifically, we observe a $+4.11\\%$ gain in Det. versus HVQ. On top of that, with in-domain $\\widehat{\\mathcal{C}}$ as a more aligned representation, LTOAD in Tab. 7 achieves the best image-level AUROC. In particular, we further improve it by $+1.88\\%$ . We have a similar observation initializing with Small, which outperforms HVQ on Det. Having sufficiently large $\\widehat{\\mathcal{C}}$ , LTOAD with $\\widehat{K} = 10$ gives more gain.", + "bbox": [ + 511, + 551, + 905, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, although some $c \\in \\mathcal{C}$ are not presented in our $\\widehat{\\mathcal{C}}$ , LTOAD still adapts and outperforms the baselines in Tab. 7. This suggests that LTOAD can work even with a foundation model whose $\\mathcal{V}$ and $\\mathcal{C}$ are disjoint. However, this is not necessarily true for prior class-aware methods as they implicitly assume that $\\mathcal{C} \\subseteq \\mathcal{V}$ .", + "bbox": [ + 511, + 718, + 903, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation studies on R, S, and G. In Tab. 7, we also ablate LTOAD components (see Sec. 4.2 and 4.3) to justify their design. All experiments are done under the MVTec exp100 and $B$ setting. NoR ablates R. We also remove G, $\\widehat{\\mathbf{Y}}^{\\mathrm{R}}$ from Eq. 4 and related losses (see Appendix A1.1). In contrast, NoS ablates S. We also remove $\\widehat{\\mathbf{Y}}^{\\mathrm{S}}$ and related", + "bbox": [ + 511, + 810, + 905, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23425", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/31b2ab0c9aeef87017dfc43d28cf902c252525c9a870451462b3b7a417755f13.jpg", + "table_caption": [], + "table_footnote": [ + "Table 9. Characteristic comparison between LTOAD and prior unsupervised AD works (e.g., $\\mathcal{G}_0$ : [12, 14, 15, 19, 26, 27, 38, 48, 54, 59, 63, 70, 72, 73, 75, 80, 87, 90, 91, 93, 97], $\\mathcal{G}_1$ : [2, 11, 21, 30, 31, 33, 39, 46, 50, 62, 66, 76, 77, 87-89, 95, 99, 102], $\\mathcal{G}_2$ : [57], $\\mathcal{G}_3$ : [37, 51, 52], $\\mathcal{G}_4$ : [7, 9, 10, 25, 40, 49, 67, 94], $\\mathcal{G}_5$ : [34], $\\mathcal{G}_6$ : [22, 61, 84], $\\mathcal{G}_7$ : [55, 79])." + ], + "table_body": "
Unsupervised AD Method ConditionsUnsupervised AD Categories
G0G1G2G3G4G5G6G7LTOAD
unified modelX✓✓X✓✓X✓✓
online learning capabilityXXXXXX✓✓✓
incorporate VLM/LLM(s)XXX✓✓✓XX✓
can handle long-tailed dataXXXXX✓XX✓
class-agnosticX✓XXXXX✓✓
", + "bbox": [ + 94, + 88, + 478, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "losses. We find from NoR that S alone is not an effective branch while NoS is far more competitive. However, when trained jointly (LTOAD) and compared to NoS, S helps R improve the Det. accuracy on $\\mathcal{C}^{\\text {tail }}$ notably. Specifically, S improves the image-level AUROC on $\\mathcal{C}^{\\text {tail }}$ by $2.98\\%$ .", + "bbox": [ + 89, + 287, + 480, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NoG replaces our prompt-conditioned $\\mathsf{G}$ with unconditional feature jittering [34, 57, 89] for data augmentation. With the guidance of $\\mathcal{P}^{\\mathrm{n}}$ and $\\mathcal{P}^{\\mathrm{a}}$ , $\\mathsf{G}$ creates high-quality pseudo-features especially beneficial for $\\mathcal{C}^{\\mathrm{tail}}$ . We observe a gain of $+2.51\\% / +0.75\\%$ in Det. / Seg. on $\\mathcal{C}^{\\mathrm{tail}}$ .", + "bbox": [ + 89, + 364, + 480, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation studies on learnable concepts/prompts. Tab. 8 analyzes the effect of each attribute detailed in Tab. 10, where we demonstrate that our prompt learning design is comprehensive. All experiments are performed in the setting MVTec exp100 and B. FixC freezes $\\hat{\\mathcal{C}}$ . FixP freezes $\\mathcal{P}^a$ . In SingleP, $\\mathcal{P}^a$ stores only one abnormal prompt for each concept. In ShareP, all concepts share the same abnormal prompts. Each attribute contributes significantly to the Det. performance of $\\mathcal{C}^{\\mathrm{tail}}$ and $\\mathcal{C}$ .", + "bbox": [ + 89, + 439, + 482, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Related work", + "text_level": 1, + "bbox": [ + 89, + 588, + 227, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 9, we group the literature of unsupervised AD based on their assumptions or target tasks. Notably, LTOAD (1) is a unified model, (2) has online learning capability, (3) uses VLM/LLM priors, (4) handles long-tailed training data distributions, and (5) is class-agnostic.", + "bbox": [ + 89, + 613, + 482, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Unsupervised anomaly detection. $\\mathcal{G}_0$ methods train an independent predictor per class. In contrast, UniAD [89] challenges to learn one unified predictor for all classes. More follow-up research falls into this group $(\\mathcal{G}_1)$ . Inspired by prior work [56, 60, 71], MoEAD [62] uses mixture-of-experts. HVQ [57] uses class-specific quantization modules in their VQ-VAE [82] architecture, but HVQ also induces the constraint of requiring class information as inputs, making it not class-agnostic, i.e., $\\mathcal{G}_2$ .", + "bbox": [ + 89, + 688, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Foundation model based AD. With advancement in foundation models, recent works focused on zero-shot [7, 9, 24, 40, 50, 67, 100, 102] or few-shot [2, 25, 51-53, 101] unsupervised AD where the normal training images are limited (or none), they [7, 9, 25, 40, 51, 52, 67, 100] often", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e1a47d6d9eb93515b1561a61644950a0f393086c230064aa810cb72a0b9ab881.jpg", + "table_caption": [], + "table_footnote": [ + "Table 10. Characteristic comparison w.r.t. prompt design between LTOAD and prior unsupervised AD works which use LLM(s) with prompts (e.g., $\\mathcal{G}_8$ : [34, 67], $\\mathcal{G}_9$ : [10], $\\mathcal{G}_{10}$ : [9, 37, 40, 52, 101], $\\mathcal{G}_{11}$ : [7], $\\mathcal{G}_{12}$ : [51, 102]). AP: abnormal prompts." + ], + "table_body": "
Prompt Design Conditions of Unsupervised AD MethodsPrompt Design Categories in AD
G8G9G10G11G12LTOAD
learnable class/concept names✓XXXX✓
support >1 APsXX✓✓✓✓
learnable APsX✓XX✓✓
class/concept specific APsXXX✓✓✓
", + "bbox": [ + 517, + 88, + 903, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "turn themselves to the rich knowledge in foundation models [29, 41, 68, 81, 98] or conversational AI [64]. These methods can be grouped into either $\\mathcal{G}_3$ or $\\mathcal{G}_4$ , depending on whether they are unified model. To query the language models precisely, $\\mathcal{G}_3$ and $\\mathcal{G}_4$ need the class names or labels as extra inputs and are thus not class-agnostic.", + "bbox": [ + 511, + 267, + 903, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although normal images are limited in zero/few-shot AD, the distribution is even since the number of images per class is the same. Consequently, LTAD [34], i.e., $\\mathcal{G}_5$ , explores the long-tailed setting where the training set is highly unevenly distributed. They rely on the pre-trained ALIGN [41] and learn the pseudo-class names. We highlight the differences between LTOAD prompt learning to related works in Tab. 10. In summary, LTOAD supports (a) learnable concept names, (b) learnable abnormal prompts (AP), (c) multiple APs, and (d) concept-specific APs.", + "bbox": [ + 511, + 358, + 905, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Online learning AD. Online (or continual/incremental) learning [23, 44, 69, 83] investigates how the model learns continuously on a data stream. Recent works [22, 55, 61, 79, 84], i.e. $\\mathcal{G}_6$ and $\\mathcal{G}_7$ , explore unsupervised online learning AD. Differently, we propose the task of long-tailed online anomaly detection to study the effect of head/tail classes.", + "bbox": [ + 511, + 508, + 905, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 614, + 633, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose the benchmark of long-tailed online anomaly detection, studying the online performance of head and tail classes. We identify the shortcomings of class-aware methods and propose class-agnostic LTOAD. It learns to approximate the ground truth class set while using foundation models for our novel concept and abnormal prompt learning framework. We also propose $\\mathcal{A}^{\\mathrm{AA}}$ , a learning algorithm to adapt the online stream with potentially abnormal inputs. LTOAD outperforms the SOTA in most offline and online settings without additional class information. Although not specifically designed for domain generalization, LTOAD can adapt to unseen domains. We believe that LTOAD will advance the field toward a more realistic anomaly detection setting that is closer to real-world applications.", + "bbox": [ + 511, + 638, + 905, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. CAY and KCP were supported by Mitsubishi Electric Research Laboratories. CAY and RAY were in part supported by an NSF Award #2420724.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23426", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Akshatha Arodi, Margaux Luck, Jean-Luc Bedwani, Aldo Zaimi, Ge Li, Nicolas Pouliot, Julien Beaudry, and Gaétan Marceau Caron. CableInspect-AD: An expert-annotated anomaly detection dataset. In Proc. NeurIPS, 2024. 1", + "[2] Yuhu Bai, Jiangning Zhang, Zhaofeng Chen, Yuhang Dong, Yunkang Cao, and Guanzhong Tian. Dual-path frequency discriminators for few-shot anomaly detection. Knowledge-Based Systems, 2024. 8", + "[3] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 5", + "[4] Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxiang Zhang, and Xingyu Li. BMAD: Benchmarks for medical anomaly detection. In Proc. CVPR, 2024. 5, 1", + "[5] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. MVTec AD-a comprehensive real-world dataset for unsupervised anomaly detection. In Proc. CVPR, 2019. 1, 5, 3, 4", + "[6] Patrick Bilic, Patrick Christ, Hongwei Bran Li, Eugene Vorontsov, Avi Ben-Cohen, Georgios Kaissis, Adi Szeskin, Colin Jacobs, Gabriel Efrain Humpire Mamani, Gabriel Chartrand, et al. The liver tumor segmentation benchmark (lits). Medical Image Analysis, 2023. 5", + "[7] Yunkang Cao, Xiaohao Xu, Chen Sun, Yuqi Cheng, Zongwei Du, Liang Gao, and Weiming Shen. Segment any anomaly without training via hybrid prompt regularization. arXiv preprint arXiv:2305.10724, 2023. 1, 2, 8", + "[8] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1", + "[9] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. AdaCLIP: Adapting CLIP with hybrid learnable prompts for zero-shot anomaly detection. In Proc. ECCV, 2024. 1, 2, 8", + "[10] Pi-Wei Chen, Jerry Chun-Wei Lin, Jia Ji, Feng-Hao Yeh, Zih-Ching Chen, and Chao-Chun Chen. Human-free automated prompting for vision-language anomaly detection: Prompt optimization with meta-guiding prompt scheme. arXiv preprint arXiv:2406.18197, 2024. 1, 2, 8", + "[11] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. In Proc. ECCV, 2024. 8", + "[12] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 8", + "[13] Microsoft Corporation. Microsoft Copilot, 2023. 4, 2", + "[14] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. PaDiM: a patch distribution modeling framework for anomaly detection and localization. In Proc. ICPR, 2021. 8" + ], + "bbox": [ + 99, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proc. CVPR, 2022. 8", + "[16] Keval Doshi and Yasin Yilmaz. Continual learning for anomaly detection in surveillance videos. In Proc. CVPR Workshop, 2020. 1", + "[17] Hugging Face. Transformers: Align model documentation. https://huggingface.co/docs/ transformers/model_doc align.2", + "[18] Yehonatan Fridman, Matan Rusanovsky, and Gal Oren. ChangeChip: A reference-based unsupervised change detection for PCB defect detection. In Proc. PAINE, 2021. 1", + "[19] Matic Fučka, Vitjan Zavrtanik, and Danijel Skočaj. TransFusion-a transparency-based diffusion model for anomaly detection. In Proc. ECCV, 2024. 8", + "[20] Hugo Gangloff, Minh-Tan Pham, Luc Courtrai, and Sébastien Lefèvre. Leveraging vector-quantized variational autoencoder inner metrics for anomaly detection. In Proc. ICPR, 2022. 1, 5", + "[21] Bin-Bin Gao. Learning to detect multi-class anomalies with just one normal image prompt. In Proc. ECCV, 2024. 5, 8", + "[22] Han Gao, Huiyuan Luo, Fei Shen, and Zhengtao Zhang. Towards total online unsupervised anomaly detection and localization in industrial vision. arXiv preprint arXiv:2305.15652, 2023. 1, 2, 8", + "[23] Yasir Ghunaim, Adel Bibi, Kumail Alhamoud, Motasem Alfarra, Hasan Abed Al Kader Hammoud, Ameya Prabhu, Philip HS Torr, and Bernard Ghanem. Real-time evaluation in online continual learning: A new hope. In Proc. CVPR, 2023. 8", + "[24] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Hao Li, Ming Tang, and Jinqiao Wang. Filo: Zero-shot anomaly detection by fine-grained description and high-quality localization. In Proc. ACM MM, 2024. 8", + "[25] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. AnomalyGPT: Detecting industrial anomalies using large vision-language models. In Proc. AAAI, 2024. 2, 6, 8, 5, 7, 9, 10, 11, 12, 13, 14", + "[26] Guan Gui, Bin-Bin Gao, Jun Liu, Chengjie Wang, and Yunsheng Wu. Few-shot anomaly-driven generation for anomaly classification and segmentation. In Proc. ECCV, 2024. 8", + "[27] Jia Guo, Haonan Han, Shuai Lu, Weihang Zhang, and Huiqi Li. Absolute-unified multi-class anomaly detection via class-agnostic distribution alignment. arXiv preprint arXiv:2404.00724, 2024.8", + "[28] Ahmed Hamada. Br35H: Brain tumor detection 2020, 2020. 1", + "[29] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xi-angyu Yue. OneLLM: One framework to align all modalities with language. In Proc. CVPR, 2024. 8", + "[30] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. MambaAD: Exploring state space models for multi-class unsupervised anomaly detection. In Proc. NeurIPS, 2024. 2, 8" + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23427", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proc. AAAI, 2024. 8", + "[32] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, 2016. 4", + "[33] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. In Proc. ECCV, 2024. 8", + "[34] Chih-Hui Ho, Kuan-Chuan Peng, and Nuno Vasconcelos. Long-tailed anomaly detection with learnable class names. In Proc. CVPR, 2024. 1, 2, 3, 5, 6, 7, 8, 4, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21", + "[35] Junjie Hu, Yuanyuan Chen, and Zhang Yi. Automated segmentation of macular edema in oct using deep neural networks. Medical image analysis, 2019. 5", + "[36] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In Proc. ECCV, 2022. 6, 5, 7, 8, 9, 10, 11, 12, 13, 14", + "[37] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xinchao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proc. CVPR, 2024. 1, 2, 8", + "[38] Jeeho Hyun, Sangyun Kim, Giyoung Jeon, Seung Hwan Kim, Kyunghoon Bae, and Byung Jun Kang. ReConPatch: Contrastive patch representation learning for industrial anomaly detection. In Proc. WACV, 2024. 8", + "[39] Brian KS Isaac-Medina, Yona Falinie A Gaus, Neelanjan Bhowmik, and Toby P Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In Proc. ECCV, 2024. 8", + "[40] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. WinCLIP: Zero/few-shot anomaly classification and segmentation. In Proc. CVPR, 2023. 1, 2, 5, 8", + "[41] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proc. ICML, 2021. 8, 2, 4", + "[42] Pranita Balaji Kanade and PP Gumaste. Brain tumor detection using MRI images. *Brain*, 2015. 1", + "[43] Felipe Campos Kitamura. Head CT - hemorrhage, 2018. 1", + "[44] Hyunseo Koh, Dahyun Kim, Jung-Woo Ha, and Jonghyun Choi. Online continual learning on class incremental blurry task configuration with anytime inference. In Proc. ICLR, 2022. 8", + "[45] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, 2015. 5" + ], + "bbox": [ + 99, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Joo Chan Lee, Taejune Kim, Eunbyung Park, Simon S Woo, and Jong Hwan Ko. Continuous memory representation for anomaly detection. In Proc. ECCV, 2024. 8", + "[47] Jan Lehr, Jan Philipps, Alik Sargsyan, Martin Pape, and Jörg Krüger. AD3: Introducing a score for anomaly detection dataset difficulty assessment using VIADUCT dataset. In Proc. ECCV, 2024. 1", + "[48] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. CutPaste: Self-supervised learning for anomaly detection and localization. In Proc. CVPR, 2021. 8", + "[49] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. ClipSAM: CLIP and SAM collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 2, 8", + "[50] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. MuSc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. In Proc. ICLR, 2024. 8", + "[51] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. PromptAD: Learning prompts with only normal samples for few-shot anomaly detection. In Proc. CVPR, 2024. 1, 2, 5, 8", + "[52] Yuanwei Li, Elizaveta Ivanova, and Martins Bruveris. FADE: Few-shot/zero-shot anomaly detection engine using large vision-language model. In Proc. BMVC, 2024. 1, 2, 8", + "[53] Jingyi Liao, Xun Xu, Manh Cuong Nguyen, Adam Goodge, and Chuan Sheng Foo. COFT-AD: Contrastive fine-tuning for few-shot anomaly detection. IEEE TIP, 2024. 8", + "[54] Chieh Liu, Yu-Min Chu, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning diffusion models for multi-view anomaly detection. In Proc. ECCV, 2024. 8", + "[55] Jiaqi Liu, Kai Wu, Qiang Nie, Ying Chen, Bin-Bin Gao, Yong Liu, Jinbao Wang, Chengjie Wang, and Feng Zheng. Unsupervised continual anomaly detection with contrastively-learned prompt. In Proc. AAAI, 2024. 1, 8", + "[56] Ruiying Lu, Bo Chen, Jianqiao Sun, Wenchao Chen, Penghui Wang, Yuanwei Chen, Hongwei Liu, and Pramod K Varshney. Heterogeneity-aware recurrent neural network for hyperspectral and multispectral image fusion. IEEE Journal of Selected Topics in Signal Processing, 2022. 8", + "[57] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. In Proc. NeurIPS, 2023. 1, 2, 4, 5, 6, 8, 7, 9, 10, 11, 12, 13, 14", + "[58] Jianhong Ma, Tao Zhang, Cong Yang, Yangjie Cao, Lipeng Xie, Hui Tian, and Xuexiang Li. Review of wafer surface defect detection methods. *Electronics*, 2023. 1", + "[59] Sergio Naval Marimont and Giacomo Tarroni. Anomaly detection through latent space restoration using vector quantized variational autoencoders. In Proc. ISBI, 2021. 1, 5, 8", + "[60] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review, 2014. 8" + ], + "bbox": [ + 524, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23428", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Declan GD McIntosh and Alexandra Branzan Albu. Unsupervised, online and on-the-fly anomaly detection for nonstationary image distributions. In Proc. ECCV, 2024. 1, 2, 8", + "[62] Shiyuan Meng, Wenchao Meng, Qihang Zhou, Shizhong Li, Weiye Hou, and Shibo He. MoEAD: A parameter-efficient model for multi-class anomaly detection. In Proc. ECCV, 2024. 2, 6, 8, 5, 7, 9, 10, 11, 12, 13, 14", + "[63] Shancong Mou, Xiaoyi Gu, Meng Cao, Haoping Bai, Ping Huang, Jiulong Shan, and Jianjun Shi. RGI: Robust GAN-inversion for mask-free image inpainting and unsupervised pixel-wise anomaly detection. In Proc. ICLR, 2023. 8", + "[64] OpenAI. ChatGPT, 2023. 4, 8, 2", + "[65] Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM computing surveys (CSUR), 2021. 1", + "[66] Sukanya Patra and Souhaib Ben Taieb. Revisiting deep feature reconstruction for logical and structural industrial anomaly detection. TMLR, 2024. 8", + "[67] Zhen Qu, Xian Tao, Mukesh Prasad, Fei Shen, Zhengtao Zhang, Xinyi Gong, and Guiguang Ding. VCP-CLIP: A visual context prompting model for zero-shot anomaly segmentation. In Proc. ECCV, 2024. 1, 2, 8", + "[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. ICML, 2021. 8", + "[69] Siddeshwar Raghavan, Jiangpeng He, and Fengqing Zhu. DELTA: Decoupling long-tailed online continual learning. In Proc. CVPR, 2024. 8", + "[70] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In Proc. ICML, 2015. 8", + "[71] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Proc. NeurIPS, 2021. 8", + "[72] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proc. CVPR, 2022. 8", + "[73] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proc. WACV, 2021. 8", + "[74] Radhwan AA Saleh, Mehmet Zeki Konyar, Kaplan Kaplan, and H Metin Ertunç. Tire defect detection model using machine learning. In Proc. eSmarTA, 2022. 1", + "[75] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proc. CVPR, 2021. 8, 6, 7", + "[76] Jian Shi, Pengyi Zhang, Ni Zhang, Hakim Ghazzai, and Peter Wonka. Dissolving is amplifying: Towards fine-grained anomaly detection. In Proc. ECCV, 2024. 8", + "[77] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. GeneralAD: Anomaly detection across domains by attending to distorted features. In Proc. ECCV, 2024. 8" + ], + "bbox": [ + 99, + 92, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[78] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, 2019. 2", + "[79] Jiaqi Tang, Hao Lu, Xiaogang Xu, Ruizheng Wu, Sixing Hu, Tong Zhang, Tsz Wa Cheng, Ming Ge, Ying-Cong Chen, and Fugee Tsung. An incremental unified framework for small defect inspection. In Proc. ECCV, 2024. 1, 8", + "[80] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan Duong, Chanh D Tr Nguyen, and Steven QH Truong. Revisiting reverse distillation for anomaly detection. In Proc. CVPR, 2023. 8", + "[81] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 8", + "[82] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In Proc. NeurIPS, 2017. 5, 8", + "[83] Liuyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. IEEE TPAMI, 2024. 8", + "[84] Shenxing Wei, Xing Wei, Zhiheng Ma, Shaochen Zhang Songlin Dong, and Yihong Gong. Few-shot online anomaly detection and segmentation. Knowledge-Based Systems, 2024. 1, 2, 8", + "[85] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, 2007. 5, 3, 4", + "[86] Shuai Yang, Zhifei Chen, Pengguang Chen, Xi Fang, Yixun Liang, Shu Liu, and Yingcong Chen. Defect spectrum: A granular look of large-scale defect datasets with rich semantics. In Proc. ECCV, 2024. 1", + "[87] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. GLAD: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. In Proc. ECCV, 2024. 8", + "[88] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. In Proc. ECCV, 2024.", + "[89] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Proc. NeurIPS, 2022. 5, 6, 8, 7, 9, 10, 11, 12, 13, 14", + "[90] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. FastFlow: Unsupervised anomaly detection and localization via 2D normalizing flows. arXiv preprint arXiv:2111.07677, 2021.8", + "[91] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoçaj. DRAEM-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proc. ICCV, 2021. 8, 6, 7", + "[92] Chongsheng Zhang, George Almpanidis, Gaojuan Fan, Binquan Deng, Yanbo Zhang, Ji Liu, Aouaidjia Kamel, Paolo Soda, and João Gama. A systematic review on long-tailed learning. arXiv preprint arXiv:2408.00483, 2024. 1" + ], + "bbox": [ + 524, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23429", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[93] Gongjie Zhang, Kaiwen Cui, Tzu-Yi Hung, and Shijian Lu. Defect-GAN: High-fidelity defect synthesis for automated defect inspection. In Proc. WACV, 2021. 8", + "[94] Jiangning Zhang, Xuhai Chen, Zhucun Xue, Yabiao Wang, Chengjie Wang, and Yong Liu. Exploring grounding potential of VQA-oriented GPT-4V for zero-shot anomaly detection. arXiv preprint arXiv:2311.02612, 2023. 2, 8", + "[95] Jiangning Zhang, Xuhai Chen, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, Ming-Hsuan Yang, and Dacheng Tao. Exploring plain ViT reconstruction for multiclass unsupervised anomaly detection. In Proc. IJCAI Workshop on Anomaly Detection with Foundation Models, 2024. 5, 8, 1", + "[96] Jiangning Zhang, Haoyang He, Zhenye Gan, Qingdong He, Yuxuan Cai, Zhucun Xue, Yabiao Wang, Chengjie Wang, Lei Xie, and Yong Liu. A Der: A comprehensive benchmark for multi-class visual anomaly detection. arXiv preprint arXiv:2406.03262, 2024. 1, 4", + "[97] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. DeSTSeg: Segmentation guided denoising student-teacher for anomaly detection. In Proc. CVPR, 2023. 8", + "[98] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. MetaTransformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 8", + "[99] Ying Zhao. OmniAL: A unified cnn framework for unsupervised anomaly localization. In Proc. CVPR, 2023. 8", + "[100] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In Proc. ICLR, 2024. 1, 2, 8", + "[101] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proc. CVPR, 2024. 8", + "[102] Jiawen Zhu, Yew-Soon Ong, Chunhua Shen, and Guansong Pang. Fine-grained abnormality prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2410.10289, 2024. 1, 8", + "[103] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. In Proc. ECCV, 2022. 1, 5, 3, 4" + ], + "bbox": [ + 93, + 92, + 482, + 698 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "23430", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/331fec76-66a3-4c9b-96fc-eeb1ae13b1b6_model.json b/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/331fec76-66a3-4c9b-96fc-eeb1ae13b1b6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3aaf191b3e7f71a4c575310e1448497181bc6727 --- /dev/null +++ b/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/331fec76-66a3-4c9b-96fc-eeb1ae13b1b6_model.json @@ -0,0 +1,3018 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.26, + 0.131, + 0.74, + 0.176 + ], + "angle": 0, + "content": "Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.203, + 0.737, + 0.222 + ], + "angle": 0, + "content": "Chiao-An Yang\\(^{1}\\) Kuan-Chuan Peng\\(^{2}\\) Raymond A. Yeh\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.223, + 0.892, + 0.241 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Computer Science, Purdue University \\(^{2}\\)Mitsubishi Electric Research Laboratories" + }, + { + "type": "text", + "bbox": [ + 0.291, + 0.242, + 0.702, + 0.26 + ], + "angle": 0, + "content": "1{yang2300,rayyeh}@purdue.edu 2kpeng@merl.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.296, + 0.327, + 0.312 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.328, + 0.485, + 0.645 + ], + "angle": 0, + "content": "Anomaly detection (AD) identifies the defect regions of a given image. Recent works have studied AD, focusing on learning AD without abnormal images, with long-tailed distributed training data, and using a unified model for all classes. In addition, online AD learning has also been explored. In this work, we expand in both directions to a realistic setting by considering the novel task of long-tailed online AD (LTOAD). We first identified that the offline state-of-the-art LTAD methods cannot be directly applied to the online setting. Specifically, LTAD is class-aware, requiring class labels that are not available in the online setting. To address this challenge, we propose a class-agnostic framework for LTAD and then adapt it to our online learning setting. Our method outperforms the SOTA baselines in most offline LTAD settings, including both the industrial manufacturing and the medical domain. In particular, we observe \\(+4.63\\%\\) image-AUROC on MVTec even compared to methods that have access to class labels and the number of classes. In the most challenging long-tailed online setting, we achieve \\(+0.53\\%\\) image-AUROC compared to baselines. Our LTOAD benchmark is released here1." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.673, + 0.222, + 0.688 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Anomaly detection (AD) [8, 65, 96] is the task of identifying defect regions in a given image. This task is important due to its high valued applications in industrial manufacturing [1, 5, 18, 47, 58, 74, 86, 103] and medical [28, 42, 43, 47] settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.773, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Recently, LTAD [34] explored the long-tailed (LT) [92] setting of unsupervised one-class AD, where abnormal images are not given during training and the training set is unevenly distributed. This long-tailed distribution assumes that images of certain classes significantly outnumber others, which is a realistic assumption for real-world applications. Importantly, LTAD shows that prior work on evenly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.905, + 0.433 + ], + "angle": 0, + "content": "distributed data performs suboptimally on LT data. Meanwhile, recent works [16, 22, 55, 61, 79, 84] approach AD with online learning to leverage abnormal images during test time. However, they do not study the case where the classes follow a long-tail distribution, where the performance gap of head and tail classes in the online setting may be significant during offline training. This motivates us to study the long-tailed online AD (LTOAD) task by proposing a benchmark and a novel model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.906, + 0.679 + ], + "angle": 0, + "content": "At first glance, it may seem that one can take a state-of-the-art (SOTA) long-tailed offline model, e.g., LTAD [34], and combine it with online learning to solve LTOAD. However, several limitations within the LTAD's framework prevent it from being effective in the online setting. First, LTAD is a class-aware method, where they assume that the class information is given. However, class labels are not available in the online setting. Instead, a class-agnostic method is required as illustrated in Fig. 1. Second, LTAD's encoder-decoder architecture does not leverage the latest vector quantization VAE [20, 57, 59], which is more effective, as shown in recent work. Third, the prompt learning designs of LTAD and other AD works [7, 9, 10, 37, 40, 51, 52, 67, 100–102] ignore some aspects of abnormal prompts, e.g., LTAD does not support class-specific abnormal prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.906, + 0.788 + ], + "angle": 0, + "content": "To address these challenges, we propose a novel model that is class-agnostic, with a vector quantized VAE, and a comprehensive prompt learning framework, for the task of long-tailed online anomaly detection. Experiments show that our method outperforms the SOTA baselines significantly by at least \\(4\\%\\) on several detection benchmarks without access to the class labels and the number of classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.792, + 0.905, + 0.882 + ], + "angle": 0, + "content": "In the following sections, we first discuss how to make existing class-aware methods class-agnostic, specifically on LTAD (Sec. 2). Next, we define the task of long-tailed online AD and present our solution (Sec. 3). Finally, we provide details of each proposed module (Sec. 4), followed by the experimental results (Sec. 5) and related works (Sec. 6)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.838, + 0.9 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.424, + 0.9 + ], + "angle": 0, + "content": "MethodCAexp100exp200step100step200Det.Seg.Det.Seg.Det.Seg.Det.Seg.RegAD [36]X82.4395.20N/AN/A81.5495.10N/AN/AAnomalyGPT [25]X87.4489.6885.8090.1585.9589.2882.4789.45LTAD [34]X88.8694.4686.0594.1887.3693.8385.6092.12HVQ [57]X87.4395.2585.3494.7985.3994.1783.0092.61LTOAD*X93.1295.0191.8894.3692.0294.7287.9593.34UniAD [89]✓87.7093.9586.2193.2683.3791.4781.3292.29MoEAD [62]✓84.7394.3483.2393.9684.4093.7683.1392.76LTOAD✓93.4295.2192.0294.9492.3395.1188.6294.00" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.225, + 0.484, + 0.336 + ], + "angle": 0, + "content": "Table 2. Comparison (↑) on offline MVTec in image-level AUROC for anomaly detection (Det.) and pixel-level AUROC for anomaly segmentation (Seg.). The column CA (class-agnostic) indicates whether a method requires class names or the number of classes during training or not (require: X; not require: √). RegAD needs at least 2 images per class and thus is not applicable on exp200 and step200. We mark the best and second best performances in bold and underline." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.345, + 0.481, + 0.475 + ], + "angle": 0, + "content": "
MethodCAexp100exp200step100step200
Det.Seg.Det.Seg.Det.Seg.Det.Seg.
RegAD [36]X71.3694.4071.3694.4071.8094.9971.6594.52
AnomalyGPT [25]X70.3480.3269.7879.4871.9882.3069.7881.97
LTAD [34]X80.0095.5680.2195.3684.8096.5784.0396.27
HVQ [57]X78.6396.1878.0395.3276.6895.6375.5295.50
LTOAD*X84.8996.4284.8496.4487.7197.0786.5296.89
UniAD [89]✓77.3195.0376.8794.8078.8396.0477.6495.66
MoEAD [62]✓73.8994.3473.5893.9874.1294.3776.7794.82
LTOAD✓85.2696.9185.2496.9788.3297.5487.0497.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.481, + 0.483, + 0.495 + ], + "angle": 0, + "content": "Table 3. Comparison (↑) on offline VisA. Format follows Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.622 + ], + "angle": 0, + "content": "Baselines & implementation details. We compare with RegAD [36], UniAD [89], and LTAD [34]. We benchmark the SOTA MoEAD [62] and HVQ [57] to highlight the gain of our concept-specific quantization modules to their class-specific ones. Only UniAD, MoEAD, and LTOAD are class-agnostic and do not need a class label at the test time. We provide a class-aware version of our model, LTOAD*." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.759 + ], + "angle": 0, + "content": "Quantitative & qualitative comparison. Tab. 2 and 3 show the comparisons of LTOAD with the baselines on MVTec and VisA. We observe that LTOAD outperforms the SOTA on most settings while using no class information. LTOAD achieves a larger gain over Det. On average across all settings, LTOAD achieves \\(+4.63\\% / +0.58\\%\\) gain versus the second best baseline excluding LTOAD* on MVTec Det. / Seg. As for VisA, LTOAD achieves \\(+4.30\\% / +1.29\\%\\) gain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.484, + 0.851 + ], + "angle": 0, + "content": "Next, comparing LTAD and LTOAD*, we find that being class-agnostic is beneficial. We hypothesize that this is because the classes can now share knowledge through the soft labels. Finally, we show qualitative results in Fig. 3. We observe that LTOAD predictions are closer to ground truth than HVQ and LTAD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.863, + 0.358, + 0.879 + ], + "angle": 0, + "content": "5.2. Online long-tailed benchmark" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Experiment setup. For the online setting, we report on the" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.087, + 0.998, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.307, + 0.755, + 0.32 + ], + "angle": 0, + "content": "truth anomaly masks are circled in pink." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.337, + 0.905, + 0.439 + ], + "angle": 0, + "content": "
MethodCA OnlineBB-HFB-TFD2-HFD2-TFD5-HFD5-TFD5-MAvg.
LTAD [34]XX93.4693.4693.4693.4693.4693.4693.4693.4693.46
LTAD [34]XA^N94.1293.8094.0793.7094.0893.4993.8493.6593.84
HVQ [57]XX95.0295.0295.0295.0295.0295.0295.0295.0295.02
HVQ [57]XA^N95.4495.3395.2895.2295.2995.0395.2195.0595.23
LTOAD✓X95.0895.0895.0895.0895.0895.0895.0895.0895.08
LTOAD✓A^N95.4195.2295.3295.2395.0494.9194.7294.7995.08
LTOAD✓A^N95.3695.3095.3295.2395.2895.2695.2595.2195.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.45, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Table 4. Comparison (↑) on online MVTec in pixel-level AU-ROC for anomaly segmentation across 8 configurations of \\(\\mathcal{D}^0\\). All models are pre-trained on MVTec exp100 [34]." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.494, + 0.902, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.639, + 0.907, + 0.681 + ], + "angle": 0, + "content": "Figure 4. Performance curve of \\(\\mathcal{A}^{\\mathrm{N}}\\) and \\(\\mathcal{A}^{\\mathrm{AA}}\\) in pixel-level AUROC on \\(\\mathcal{C}^{\\mathrm{head}}\\) and \\(\\mathcal{C}^{\\mathrm{tail}}\\). They are offline-trained on \\(exp100\\) [34] and online-trained on D5-HF or D5-TF." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.763 + ], + "angle": 0, + "content": "same datasets as the offline benchmark [34], i.e., MVtec, VisA, and DAGM. We split the testing set into \\(\\mathcal{D}^0\\) and \\(\\mathcal{D}^{\\mathrm{E}}\\) with a 3:7 ratio. For each dataset, we create the 8 settings defined in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.906, + 0.855 + ], + "angle": 0, + "content": "For baseline comparisons, we extend HVQ and LTAD to have online capability by applying our naive online learning algorithm \\(\\mathcal{A}^{\\mathsf{N}}\\) since they are the best offline models on MVTec. Note that both methods are not class-agnostic and needs class labels as extra information in online learning, which is not available/necessary for our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Quantitative comparison. We report the online Seg. performance at final time step \\( t = T \\) on MVtec exp100 in Tab. 4. The results on other settings/datasets are in the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23424" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.482, + 0.152 + ], + "angle": 0, + "content": "
MethodOnlineMVTecVisADAGM
→ VisA→ DAGM→ MVTec→ DAGM→ MVTec→ VisA
LTOADX84.3891.2782.7379.1679.5181.13
LTOAD✓92.3796.8789.8496.7988.3792.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.157, + 0.483, + 0.187 + ], + "angle": 0, + "content": "Table 5. Cross-dataset anomaly segmentation in pixel-level AUROC. All models are pre-trained on their \\(exp100\\) [34]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.198, + 0.48, + 0.365 + ], + "angle": 0, + "content": "
Set IDDomainSizeElements
CN/A15zipper, pill, capsule, grid, transistor, carpet, metal nut, wood, leather, screw, tile, cable, toothbrush, hazelnut, bottle*
RandomOut10refinery, finger, vault, exam, salad, sailor, dove, dealer, well, humanoid
SmallIn3semiconductor, zipper, beech
LTOAD CIn10semiconductor, zipper, beech, microscopy*, antibiotics, medicines, circuit , mahogany, hardwood, walnut
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.415 + ], + "angle": 0, + "content": "Table 6. Classes vs. concepts on MVTec. Words with similar semantics have the same color, e.g., wood-related words are colored in brown and electronics-related words are colored in purple." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.429, + 0.485, + 0.657 + ], + "angle": 0, + "content": "Appendix A4. Overall, \\(\\mathcal{A}^{\\mathrm{N}}\\) performs well in the simplest setting \\((B)\\), but degrades on more difficult settings, e.g., \\(D2\\)-TF, \\(D5\\)-HF, etc. On average, LTOAD with \\(\\mathcal{A}^{\\mathrm{AA}}\\) performs the best even without class labels as extra information. We plot LTOAD's performance curve of \\(\\mathsf{F}_{\\theta_t}\\) over \\(t\\) with \\(\\Delta = 16\\) under the two most difficult settings (\\(D5\\)-HF and \\(D5\\)-TF) in Fig. 4, where \\(\\mathcal{A}^{\\mathrm{N}}\\) falls off on \\(\\mathcal{C}^{\\mathrm{head}} / \\mathcal{C}^{\\mathrm{head}}\\) under \\(D5\\)-HF/D5-TF while \\(\\mathcal{A}^{\\mathrm{AA}}\\) improves steadily. Empirically, LTOAD is also more efficient than LTAD because \\(\\widehat{K} < K\\). Cross-dataset evaluation. Tab. 5 shows that LTOAD can adapt to unseen domains even if it is not designed for cross-dataset evaluation. For each \\(\\mathcal{D}_1 \\to \\mathcal{D}_2\\) experiment, we pre-train LTOAD on \\(exp100\\) of \\(\\mathcal{D}_1\\) then evaluated on \\(B\\) of \\(\\mathcal{D}_2\\) without and with online learning. LTOAD with online learning outperforms the one without it." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.277, + 0.683 + ], + "angle": 0, + "content": "5.3. Additional analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Analysis on initialization of \\(\\widehat{C}\\). We now analyze the effect of our concept learning component. All experiments are done under MVtec exp100 and \\(B\\). We highlight the importance of our concept learning compared to two baselines: initializing with a random concept set or with a small one." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Tab. 6 shows the learned initialization of \\(\\widehat{C}\\) versus the ground truth \\(C\\). We observe that our concept learning can find the \\(\\widehat{C}\\) initialization which represents similar semantics in \\(C\\), e.g., the bottle* images (see the 1st image in Fig. 3) in MVtec are bottom-only views, which look like an microscopy*." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Next, we experiment with the two baselines to generate initialization for \\(\\widehat{\\mathcal{C}}\\). In Random of Tab. 6, instead of learning from \\(\\mathcal{D}^{\\mathrm{T}}\\), i.e., in-domain concepts, we randomly select 10" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.09, + 0.903, + 0.22 + ], + "angle": 0, + "content": "
Exp.\\( \\widehat{C} \\)\\( \\widehat{K} \\)RSGDetectionSegmentation
\\( C^{tail} \\)\\( C^{head} \\)\\( C \\)\\( C^{tail} \\)\\( C^{head} \\)\\( C \\)
RandomO-D10✓✓✓84.7399.3191.5494.4695.3994.89
SmallI-D3✓✓✓85.6698.8991.8394.2495.3094.74
NoRI-D10✘✓✘53.8456.2554.9760.7363.4662.00
NoSI-D10✓✘✓85.1699.2191.7294.5795.2494.88
NoGI-D10✓✓✘85.6399.5692.1394.0895.6294.80
LTOADI-D10✓✓✓88.1499.4593.7294.8395.6495.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.226, + 0.905, + 0.268 + ], + "angle": 0, + "content": "Table 7. Analysis and ablation on LTOAD using the same metrics as Tab. 4 under MVTec \\( \\exp {100} \\) [34] and \\( B \\) . We report the performances on \\( {\\mathcal{C}}^{\\text{head }} \\) and \\( {\\mathcal{C}}^{\\text{tail }} \\) . I-D/O-D: in/out-domain." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.276, + 0.905, + 0.398 + ], + "angle": 0, + "content": "
Exp.C̅LPaL MCSDetectionSegmentation
CtailCheadCCtailCheadC
FixCX√√√84.6598.1390.9494.57
FixP√X√√85.5999.4592.0694.52
SingleP√√X√85.4199.3791.9294.64
ShareP√√√X87.0399.3592.7895.00
LTOAD√√√√88.1499.4593.7294.83
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.405, + 0.907, + 0.448 + ], + "angle": 0, + "content": "Table 8. Ablations on concept \\(\\hat{C}\\) & abnormal prompt \\(\\mathcal{P}^{\\mathrm{a}}\\) learning under MVtec exp100 [34] on the same metrics as Tab. 7. Acronyms: L: learned; M: multiple; CS: concept-specific." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.553 + ], + "angle": 0, + "content": "nouns in ALIGN's \\(\\mathcal{V}\\) unrelated to the industrial settings of MVTec, i.e., out-domain concepts. In other words, the semantics of \\(\\mathcal{C}\\) and Random do not overlap. In Small of Tab. 6, instead of choosing \\(\\widehat{K} = 10\\), we select only the top-\\(\\widehat{K} = 3\\) vocabularies as \\(\\widehat{\\mathcal{C}}\\). The semantics of Small partially covers the semantics of \\(\\mathcal{C}\\). Tab. 7 shows the quantitative results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.719 + ], + "angle": 0, + "content": "We report the performance of Random and Small in Tab. 7. Even with an out-domain initialization, compared to HVQ in Tab. 2, which uses \\(\\mathcal{C}\\) for their class-specific quantization modules, Random still performs competitively. Specifically, we observe a \\(+4.11\\%\\) gain in Det. versus HVQ. On top of that, with in-domain \\(\\widehat{\\mathcal{C}}\\) as a more aligned representation, LTOAD in Tab. 7 achieves the best image-level AUROC. In particular, we further improve it by \\(+1.88\\%\\). We have a similar observation initializing with Small, which outperforms HVQ on Det. Having sufficiently large \\(\\widehat{\\mathcal{C}}\\), LTOAD with \\(\\widehat{K} = 10\\) gives more gain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Finally, although some \\(c \\in \\mathcal{C}\\) are not presented in our \\(\\widehat{\\mathcal{C}}\\), LTOAD still adapts and outperforms the baselines in Tab. 7. This suggests that LTOAD can work even with a foundation model whose \\(\\mathcal{V}\\) and \\(\\mathcal{C}\\) are disjoint. However, this is not necessarily true for prior class-aware methods as they implicitly assume that \\(\\mathcal{C} \\subseteq \\mathcal{V}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Ablation studies on R, S, and G. In Tab. 7, we also ablate LTOAD components (see Sec. 4.2 and 4.3) to justify their design. All experiments are done under the MVTec exp100 and \\(B\\) setting. NoR ablates R. We also remove G, \\(\\widehat{\\mathbf{Y}}^{\\mathrm{R}}\\) from Eq. 4 and related losses (see Appendix A1.1). In contrast, NoS ablates S. We also remove \\(\\widehat{\\mathbf{Y}}^{\\mathrm{S}}\\) and related" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23425" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.189 + ], + "angle": 0, + "content": "
Unsupervised AD Method ConditionsUnsupervised AD Categories
G0G1G2G3G4G5G6G7LTOAD
unified modelX✓✓X✓✓X✓✓
online learning capabilityXXXXXX✓✓✓
incorporate VLM/LLM(s)XXX✓✓✓XX✓
can handle long-tailed dataXXXXX✓XX✓
class-agnosticX✓XXXXX✓✓
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.193, + 0.483, + 0.277 + ], + "angle": 0, + "content": "Table 9. Characteristic comparison between LTOAD and prior unsupervised AD works (e.g., \\(\\mathcal{G}_0\\): [12, 14, 15, 19, 26, 27, 38, 48, 54, 59, 63, 70, 72, 73, 75, 80, 87, 90, 91, 93, 97], \\(\\mathcal{G}_1\\): [2, 11, 21, 30, 31, 33, 39, 46, 50, 62, 66, 76, 77, 87-89, 95, 99, 102], \\(\\mathcal{G}_2\\): [57], \\(\\mathcal{G}_3\\): [37, 51, 52], \\(\\mathcal{G}_4\\): [7, 9, 10, 25, 40, 49, 67, 94], \\(\\mathcal{G}_5\\): [34], \\(\\mathcal{G}_6\\): [22, 61, 84], \\(\\mathcal{G}_7\\): [55, 79])." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.482, + 0.365 + ], + "angle": 0, + "content": "losses. We find from NoR that S alone is not an effective branch while NoS is far more competitive. However, when trained jointly (LTOAD) and compared to NoS, S helps R improve the Det. accuracy on \\(\\mathcal{C}^{\\text {tail }}\\) notably. Specifically, S improves the image-level AUROC on \\(\\mathcal{C}^{\\text {tail }}\\) by \\(2.98\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.482, + 0.44 + ], + "angle": 0, + "content": "NoG replaces our prompt-conditioned \\(\\mathsf{G}\\) with unconditional feature jittering [34, 57, 89] for data augmentation. With the guidance of \\(\\mathcal{P}^{\\mathrm{n}}\\) and \\(\\mathcal{P}^{\\mathrm{a}}\\), \\(\\mathsf{G}\\) creates high-quality pseudo-features especially beneficial for \\(\\mathcal{C}^{\\mathrm{tail}}\\). We observe a gain of \\(+2.51\\% / +0.75\\%\\) in Det. / Seg. on \\(\\mathcal{C}^{\\mathrm{tail}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.483, + 0.576 + ], + "angle": 0, + "content": "Ablation studies on learnable concepts/prompts. Tab. 8 analyzes the effect of each attribute detailed in Tab. 10, where we demonstrate that our prompt learning design is comprehensive. All experiments are performed in the setting MVTec exp100 and B. FixC freezes \\(\\hat{\\mathcal{C}}\\). FixP freezes \\(\\mathcal{P}^a\\). In SingleP, \\(\\mathcal{P}^a\\) stores only one abnormal prompt for each concept. In ShareP, all concepts share the same abnormal prompts. Each attribute contributes significantly to the Det. performance of \\(\\mathcal{C}^{\\mathrm{tail}}\\) and \\(\\mathcal{C}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.228, + 0.604 + ], + "angle": 0, + "content": "6. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.689 + ], + "angle": 0, + "content": "In Tab. 9, we group the literature of unsupervised AD based on their assumptions or target tasks. Notably, LTOAD (1) is a unified model, (2) has online learning capability, (3) uses VLM/LLM priors, (4) handles long-tailed training data distributions, and (5) is class-agnostic." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Unsupervised anomaly detection. \\(\\mathcal{G}_0\\) methods train an independent predictor per class. In contrast, UniAD [89] challenges to learn one unified predictor for all classes. More follow-up research falls into this group \\((\\mathcal{G}_1)\\). Inspired by prior work [56, 60, 71], MoEAD [62] uses mixture-of-experts. HVQ [57] uses class-specific quantization modules in their VQ-VAE [82] architecture, but HVQ also induces the constraint of requiring class information as inputs, making it not class-agnostic, i.e., \\(\\mathcal{G}_2\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Foundation model based AD. With advancement in foundation models, recent works focused on zero-shot [7, 9, 24, 40, 50, 67, 100, 102] or few-shot [2, 25, 51-53, 101] unsupervised AD where the normal training images are limited (or none), they [7, 9, 25, 40, 51, 52, 67, 100] often" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.188 + ], + "angle": 0, + "content": "
Prompt Design Conditions of Unsupervised AD MethodsPrompt Design Categories in AD
G8G9G10G11G12LTOAD
learnable class/concept names✓XXXX✓
support >1 APsXX✓✓✓✓
learnable APsX✓XX✓✓
class/concept specific APsXXX✓✓✓
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.192, + 0.905, + 0.249 + ], + "angle": 0, + "content": "Table 10. Characteristic comparison w.r.t. prompt design between LTOAD and prior unsupervised AD works which use LLM(s) with prompts (e.g., \\(\\mathcal{G}_8\\): [34, 67], \\(\\mathcal{G}_9\\): [10], \\(\\mathcal{G}_{10}\\): [9, 37, 40, 52, 101], \\(\\mathcal{G}_{11}\\): [7], \\(\\mathcal{G}_{12}\\): [51, 102]). AP: abnormal prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.905, + 0.358 + ], + "angle": 0, + "content": "turn themselves to the rich knowledge in foundation models [29, 41, 68, 81, 98] or conversational AI [64]. These methods can be grouped into either \\(\\mathcal{G}_3\\) or \\(\\mathcal{G}_4\\), depending on whether they are unified model. To query the language models precisely, \\(\\mathcal{G}_3\\) and \\(\\mathcal{G}_4\\) need the class names or labels as extra inputs and are thus not class-agnostic." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.906, + 0.509 + ], + "angle": 0, + "content": "Although normal images are limited in zero/few-shot AD, the distribution is even since the number of images per class is the same. Consequently, LTAD [34], i.e., \\(\\mathcal{G}_5\\), explores the long-tailed setting where the training set is highly unevenly distributed. They rely on the pre-trained ALIGN [41] and learn the pseudo-class names. We highlight the differences between LTOAD prompt learning to related works in Tab. 10. In summary, LTOAD supports (a) learnable concept names, (b) learnable abnormal prompts (AP), (c) multiple APs, and (d) concept-specific APs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.51, + 0.906, + 0.601 + ], + "angle": 0, + "content": "Online learning AD. Online (or continual/incremental) learning [23, 44, 69, 83] investigates how the model learns continuously on a data stream. Recent works [22, 55, 61, 79, 84], i.e. \\(\\mathcal{G}_6\\) and \\(\\mathcal{G}_7\\), explore unsupervised online learning AD. Differently, we propose the task of long-tailed online anomaly detection to study the effect of head/tail classes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.615, + 0.634, + 0.631 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.906, + 0.852 + ], + "angle": 0, + "content": "We propose the benchmark of long-tailed online anomaly detection, studying the online performance of head and tail classes. We identify the shortcomings of class-aware methods and propose class-agnostic LTOAD. It learns to approximate the ground truth class set while using foundation models for our novel concept and abnormal prompt learning framework. We also propose \\(\\mathcal{A}^{\\mathrm{AA}}\\), a learning algorithm to adapt the online stream with potentially abnormal inputs. LTOAD outperforms the SOTA in most offline and online settings without additional class information. Although not specifically designed for domain generalization, LTOAD can adapt to unseen domains. We believe that LTOAD will advance the field toward a more realistic anomaly detection setting that is closer to real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Acknowledgments. CAY and KCP were supported by Mitsubishi Electric Research Laboratories. CAY and RAY were in part supported by an NSF Award #2420724." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23426" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.484, + 0.183 + ], + "angle": 0, + "content": "[1] Akshatha Arodi, Margaux Luck, Jean-Luc Bedwani, Aldo Zaimi, Ge Li, Nicolas Pouliot, Julien Beaudry, and Gaétan Marceau Caron. CableInspect-AD: An expert-annotated anomaly detection dataset. In Proc. NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.185, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Yuhu Bai, Jiangning Zhang, Zhaofeng Chen, Yuhang Dong, Yunkang Cao, and Guanzhong Tian. Dual-path frequency discriminators for few-shot anomaly detection. Knowledge-Based Systems, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.242, + 0.483, + 0.324 + ], + "angle": 0, + "content": "[3] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.326, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[4] Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxiang Zhang, and Xingyu Li. BMAD: Benchmarks for medical anomaly detection. In Proc. CVPR, 2024. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.368, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[5] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. MVTec AD-a comprehensive real-world dataset for unsupervised anomaly detection. In Proc. CVPR, 2019. 1, 5, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.424, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[6] Patrick Bilic, Patrick Christ, Hongwei Bran Li, Eugene Vorontsov, Avi Ben-Cohen, Georgios Kaissis, Adi Szeskin, Colin Jacobs, Gabriel Efrain Humpire Mamani, Gabriel Chartrand, et al. The liver tumor segmentation benchmark (lits). Medical Image Analysis, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.494, + 0.483, + 0.549 + ], + "angle": 0, + "content": "[7] Yunkang Cao, Xiaohao Xu, Chen Sun, Yuqi Cheng, Zongwei Du, Liang Gao, and Weiming Shen. Segment any anomaly without training via hybrid prompt regularization. arXiv preprint arXiv:2305.10724, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.55, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[8] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.107, + 0.607, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[9] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. AdaCLIP: Adapting CLIP with hybrid learnable prompts for zero-shot anomaly detection. In Proc. ECCV, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.662, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[10] Pi-Wei Chen, Jerry Chun-Wei Lin, Jia Ji, Feng-Hao Yeh, Zih-Ching Chen, and Chao-Chun Chen. Human-free automated prompting for vision-language anomaly detection: Prompt optimization with meta-guiding prompt scheme. arXiv preprint arXiv:2406.18197, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.733, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[11] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. In Proc. ECCV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.789, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[12] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.831, + 0.45, + 0.844 + ], + "angle": 0, + "content": "[13] Microsoft Corporation. Microsoft Copilot, 2023. 4, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.845, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[14] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. PaDiM: a patch distribution modeling framework for anomaly detection and localization. In Proc. ICPR, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[15] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proc. CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[16] Keval Doshi and Yasin Yilmaz. Continual learning for anomaly detection in surveillance videos. In Proc. CVPR Workshop, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.177, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[17] Hugging Face. Transformers: Align model documentation. https://huggingface.co/docs/ transformers/model_doc align.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[18] Yehonatan Fridman, Matan Rusanovsky, and Gal Oren. ChangeChip: A reference-based unsupervised change detection for PCB defect detection. In Proc. PAINE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.274, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[19] Matic Fučka, Vitjan Zavrtanik, and Danijel Skočaj. TransFusion-a transparency-based diffusion model for anomaly detection. In Proc. ECCV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.317, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[20] Hugo Gangloff, Minh-Tan Pham, Luc Courtrai, and Sébastien Lefèvre. Leveraging vector-quantized variational autoencoder inner metrics for anomaly detection. In Proc. ICPR, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.372, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[21] Bin-Bin Gao. Learning to detect multi-class anomalies with just one normal image prompt. In Proc. ECCV, 2024. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.401, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[22] Han Gao, Huiyuan Luo, Fei Shen, and Zhengtao Zhang. Towards total online unsupervised anomaly detection and localization in industrial vision. arXiv preprint arXiv:2305.15652, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.456, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[23] Yasir Ghunaim, Adel Bibi, Kumail Alhamoud, Motasem Alfarra, Hasan Abed Al Kader Hammoud, Ameya Prabhu, Philip HS Torr, and Bernard Ghanem. Real-time evaluation in online continual learning: A new hope. In Proc. CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.525, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[24] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Hao Li, Ming Tang, and Jinqiao Wang. Filo: Zero-shot anomaly detection by fine-grained description and high-quality localization. In Proc. ACM MM, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.581, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[25] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. AnomalyGPT: Detecting industrial anomalies using large vision-language models. In Proc. AAAI, 2024. 2, 6, 8, 5, 7, 9, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.637, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[26] Guan Gui, Bin-Bin Gao, Jun Liu, Chengjie Wang, and Yunsheng Wu. Few-shot anomaly-driven generation for anomaly classification and segmentation. In Proc. ECCV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.692, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[27] Jia Guo, Haonan Han, Shuai Lu, Weihang Zhang, and Huiqi Li. Absolute-unified multi-class anomaly detection via class-agnostic distribution alignment. arXiv preprint arXiv:2404.00724, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.748, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[28] Ahmed Hamada. Br35H: Brain tumor detection 2020, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[29] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xi-angyu Yue. OneLLM: One framework to align all modalities with language. In Proc. CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. MambaAD: Exploring state space models for multi-class unsupervised anomaly detection. In Proc. NeurIPS, 2024. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23427" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[31] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proc. AAAI, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.151, + 0.482, + 0.191 + ], + "angle": 0, + "content": "[32] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.194, + 0.482, + 0.263 + ], + "angle": 0, + "content": "[33] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. In Proc. ECCV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.267, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[34] Chih-Hui Ho, Kuan-Chuan Peng, and Nuno Vasconcelos. Long-tailed anomaly detection with learnable class names. In Proc. CVPR, 2024. 1, 2, 3, 5, 6, 7, 8, 4, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.324, + 0.482, + 0.365 + ], + "angle": 0, + "content": "[35] Junjie Hu, Yuanyuan Chen, and Zhang Yi. Automated segmentation of macular edema in oct using deep neural networks. Medical image analysis, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[36] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In Proc. ECCV, 2022. 6, 5, 7, 8, 9, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.482, + 0.48 + ], + "angle": 0, + "content": "[37] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xinchao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proc. CVPR, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.483, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[38] Jeeho Hyun, Sangyun Kim, Giyoung Jeon, Seung Hwan Kim, Kyunghoon Bae, and Byung Jun Kang. ReConPatch: Contrastive patch representation learning for industrial anomaly detection. In Proc. WACV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.482, + 0.595 + ], + "angle": 0, + "content": "[39] Brian KS Isaac-Medina, Yona Falinie A Gaus, Neelanjan Bhowmik, and Toby P Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In Proc. ECCV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.599, + 0.482, + 0.653 + ], + "angle": 0, + "content": "[40] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. WinCLIP: Zero/few-shot anomaly classification and segmentation. In Proc. CVPR, 2023. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.657, + 0.482, + 0.724 + ], + "angle": 0, + "content": "[41] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proc. ICML, 2021. 8, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.728, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[42] Pranita Balaji Kanade and PP Gumaste. Brain tumor detection using MRI images. *Brain*, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.758, + 0.481, + 0.772 + ], + "angle": 0, + "content": "[43] Felipe Campos Kitamura. Head CT - hemorrhage, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[44] Hyunseo Koh, Dahyun Kim, Jung-Woo Ha, and Jonghyun Choi. Online continual learning on class incremental blurry task configuration with anytime inference. In Proc. ICLR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[45] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[46] Joo Chan Lee, Taejune Kim, Eunbyung Park, Simon S Woo, and Jong Hwan Ko. Continuous memory representation for anomaly detection. In Proc. ECCV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.137, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[47] Jan Lehr, Jan Philipps, Alik Sargsyan, Martin Pape, and Jörg Krüger. AD3: Introducing a score for anomaly detection dataset difficulty assessment using VIADUCT dataset. In Proc. ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.195, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[48] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. CutPaste: Self-supervised learning for anomaly detection and localization. In Proc. CVPR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.239, + 0.905, + 0.292 + ], + "angle": 0, + "content": "[49] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. ClipSAM: CLIP and SAM collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.297, + 0.905, + 0.35 + ], + "angle": 0, + "content": "[50] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. MuSc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. In Proc. ICLR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.354, + 0.905, + 0.409 + ], + "angle": 0, + "content": "[51] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. PromptAD: Learning prompts with only normal samples for few-shot anomaly detection. In Proc. CVPR, 2024. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.412, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[52] Yuanwei Li, Elizaveta Ivanova, and Martins Bruveris. FADE: Few-shot/zero-shot anomaly detection engine using large vision-language model. In Proc. BMVC, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.456, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[53] Jingyi Liao, Xun Xu, Manh Cuong Nguyen, Adam Goodge, and Chuan Sheng Foo. COFT-AD: Contrastive fine-tuning for few-shot anomaly detection. IEEE TIP, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.499, + 0.905, + 0.541 + ], + "angle": 0, + "content": "[54] Chieh Liu, Yu-Min Chu, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning diffusion models for multi-view anomaly detection. In Proc. ECCV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.543, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[55] Jiaqi Liu, Kai Wu, Qiang Nie, Ying Chen, Bin-Bin Gao, Yong Liu, Jinbao Wang, Chengjie Wang, and Feng Zheng. Unsupervised continual anomaly detection with contrastively-learned prompt. In Proc. AAAI, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.601, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[56] Ruiying Lu, Bo Chen, Jianqiao Sun, Wenchao Chen, Penghui Wang, Yuanwei Chen, Hongwei Liu, and Pramod K Varshney. Heterogeneity-aware recurrent neural network for hyperspectral and multispectral image fusion. IEEE Journal of Selected Topics in Signal Processing, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.686, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[57] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. In Proc. NeurIPS, 2023. 1, 2, 4, 5, 6, 8, 7, 9, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.758, + 0.905, + 0.798 + ], + "angle": 0, + "content": "[58] Jianhong Ma, Tao Zhang, Cong Yang, Yangjie Cao, Lipeng Xie, Hui Tian, and Xuexiang Li. Review of wafer surface defect detection methods. *Electronics*, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.802, + 0.905, + 0.855 + ], + "angle": 0, + "content": "[59] Sergio Naval Marimont and Giacomo Tarroni. Anomaly detection through latent space restoration using vector quantized variational autoencoders. In Proc. ISBI, 2021. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.859, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[60] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review, 2014. 8" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23428" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.146 + ], + "angle": 0, + "content": "[61] Declan GD McIntosh and Alexandra Branzan Albu. Unsupervised, online and on-the-fly anomaly detection for nonstationary image distributions. In Proc. ECCV, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.148, + 0.483, + 0.203 + ], + "angle": 0, + "content": "[62] Shiyuan Meng, Wenchao Meng, Qihang Zhou, Shizhong Li, Weiye Hou, and Shibo He. MoEAD: A parameter-efficient model for multi-class anomaly detection. In Proc. ECCV, 2024. 2, 6, 8, 5, 7, 9, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.205, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[63] Shancong Mou, Xiaoyi Gu, Meng Cao, Haoping Bai, Ping Huang, Jiulong Shan, and Jianjun Shi. RGI: Robust GAN-inversion for mask-free image inpainting and unsupervised pixel-wise anomaly detection. In Proc. ICLR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.331, + 0.273 + ], + "angle": 0, + "content": "[64] OpenAI. ChatGPT, 2023. 4, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[65] Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM computing surveys (CSUR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[66] Sukanya Patra and Souhaib Ben Taieb. Revisiting deep feature reconstruction for logical and structural industrial anomaly detection. TMLR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[67] Zhen Qu, Xian Tao, Mukesh Prasad, Fei Shen, Zhengtao Zhang, Xinyi Gong, and Guiguang Ding. VCP-CLIP: A visual context prompting model for zero-shot anomaly segmentation. In Proc. ECCV, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. ICML, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[69] Siddeshwar Raghavan, Jiangpeng He, and Fengqing Zhu. DELTA: Decoupling long-tailed online continual learning. In Proc. CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[70] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In Proc. ICML, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.554, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[71] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Proc. NeurIPS, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[72] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proc. CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[73] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proc. WACV, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[74] Radhwan AA Saleh, Mehmet Zeki Konyar, Kaplan Kaplan, and H Metin Ertunç. Tire defect detection model using machine learning. In Proc. eSmarTA, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[75] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proc. CVPR, 2021. 8, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[76] Jian Shi, Pengyi Zhang, Ni Zhang, Hakim Ghazzai, and Peter Wonka. Dissolving is amplifying: Towards fine-grained anomaly detection. In Proc. ECCV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[77] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. GeneralAD: Anomaly detection across domains by attending to distorted features. In Proc. ECCV, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[78] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[79] Jiaqi Tang, Hao Lu, Xiaogang Xu, Ruizheng Wu, Sixing Hu, Tong Zhang, Tsz Wa Cheng, Ming Ge, Ying-Cong Chen, and Fugee Tsung. An incremental unified framework for small defect inspection. In Proc. ECCV, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[80] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan Duong, Chanh D Tr Nguyen, and Steven QH Truong. Revisiting reverse distillation for anomaly detection. In Proc. CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.25, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[81] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.321, + 0.904, + 0.347 + ], + "angle": 0, + "content": "[82] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In Proc. NeurIPS, 2017. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.349, + 0.904, + 0.39 + ], + "angle": 0, + "content": "[83] Liuyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. IEEE TPAMI, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.392, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[84] Shenxing Wei, Xing Wei, Zhiheng Ma, Shaochen Zhang Songlin Dong, and Yihong Gong. Few-shot online anomaly detection and segmentation. Knowledge-Based Systems, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.449, + 0.904, + 0.49 + ], + "angle": 0, + "content": "[85] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, 2007. 5, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.492, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[86] Shuai Yang, Zhifei Chen, Pengguang Chen, Xi Fang, Yixun Liang, Shu Liu, and Yingcong Chen. Defect spectrum: A granular look of large-scale defect datasets with rich semantics. In Proc. ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.548, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[87] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. GLAD: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. In Proc. ECCV, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.619, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[88] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. In Proc. ECCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[89] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Proc. NeurIPS, 2022. 5, 6, 8, 7, 9, 10, 11, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.733, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[90] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. FastFlow: Unsupervised anomaly detection and localization via 2D normalizing flows. arXiv preprint arXiv:2111.07677, 2021.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.789, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[91] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoçaj. DRAEM-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proc. ICCV, 2021. 8, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[92] Chongsheng Zhang, George Almpanidis, Gaojuan Fan, Binquan Deng, Yanbo Zhang, Ji Liu, Aouaidjia Kamel, Paolo Soda, and João Gama. A systematic review on long-tailed learning. arXiv preprint arXiv:2408.00483, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23429" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[93] Gongjie Zhang, Kaiwen Cui, Tzu-Yi Hung, and Shijian Lu. Defect-GAN: High-fidelity defect synthesis for automated defect inspection. In Proc. WACV, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[94] Jiangning Zhang, Xuhai Chen, Zhucun Xue, Yabiao Wang, Chengjie Wang, and Yong Liu. Exploring grounding potential of VQA-oriented GPT-4V for zero-shot anomaly detection. arXiv preprint arXiv:2311.02612, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.192, + 0.483, + 0.274 + ], + "angle": 0, + "content": "[95] Jiangning Zhang, Xuhai Chen, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, Ming-Hsuan Yang, and Dacheng Tao. Exploring plain ViT reconstruction for multiclass unsupervised anomaly detection. In Proc. IJCAI Workshop on Anomaly Detection with Foundation Models, 2024. 5, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.277, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[96] Jiangning Zhang, Haoyang He, Zhenye Gan, Qingdong He, Yuxuan Cai, Zhucun Xue, Yabiao Wang, Chengjie Wang, Lei Xie, and Yong Liu. A Der: A comprehensive benchmark for multi-class visual anomaly detection. arXiv preprint arXiv:2406.03262, 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[97] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. DeSTSeg: Segmentation guided denoising student-teacher for anomaly detection. In Proc. CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[98] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. MetaTransformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[99] Ying Zhao. OmniAL: A unified cnn framework for unsupervised anomaly localization. In Proc. CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.489, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[100] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In Proc. ICLR, 2024. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.546, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[101] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proc. CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.589, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[102] Jiawen Zhu, Yew-Soon Ong, Chunhua Shen, and Guansong Pang. Fine-grained abnormality prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2410.10289, 2024. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[103] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. In Proc. ECCV, 2022. 1, 5, 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23430" + } + ] +] \ No newline at end of file diff --git a/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/331fec76-66a3-4c9b-96fc-eeb1ae13b1b6_origin.pdf b/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/331fec76-66a3-4c9b-96fc-eeb1ae13b1b6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6e396fcaf48fd77a9e5eb8d06da5113fecf05be8 --- /dev/null +++ b/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/331fec76-66a3-4c9b-96fc-eeb1ae13b1b6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c87f2b5fe0b27d75c4d3253274be853430e5d15d64069d97cedf9b3dbfb6742 +size 4778470 diff --git a/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/full.md b/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d827c7c082b25715a387c0518cec29a206a3132b --- /dev/null +++ b/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/full.md @@ -0,0 +1,419 @@ +# Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts + +Chiao-An Yang $^{1}$ Kuan-Chuan Peng $^{2}$ Raymond A. Yeh $^{1}$ + +$^{1}$ Department of Computer Science, Purdue University $^{2}$ Mitsubishi Electric Research Laboratories + +1{yang2300,rayyeh}@purdue.edu 2kpeng@merl.com + +# Abstract + +Anomaly detection (AD) identifies the defect regions of a given image. Recent works have studied AD, focusing on learning AD without abnormal images, with long-tailed distributed training data, and using a unified model for all classes. In addition, online AD learning has also been explored. In this work, we expand in both directions to a realistic setting by considering the novel task of long-tailed online AD (LTOAD). We first identified that the offline state-of-the-art LTAD methods cannot be directly applied to the online setting. Specifically, LTAD is class-aware, requiring class labels that are not available in the online setting. To address this challenge, we propose a class-agnostic framework for LTAD and then adapt it to our online learning setting. Our method outperforms the SOTA baselines in most offline LTAD settings, including both the industrial manufacturing and the medical domain. In particular, we observe $+4.63\%$ image-AUROC on MVTec even compared to methods that have access to class labels and the number of classes. In the most challenging long-tailed online setting, we achieve $+0.53\%$ image-AUROC compared to baselines. Our LTOAD benchmark is released here1. + +# 1. Introduction + +Anomaly detection (AD) [8, 65, 96] is the task of identifying defect regions in a given image. This task is important due to its high valued applications in industrial manufacturing [1, 5, 18, 47, 58, 74, 86, 103] and medical [28, 42, 43, 47] settings. + +Recently, LTAD [34] explored the long-tailed (LT) [92] setting of unsupervised one-class AD, where abnormal images are not given during training and the training set is unevenly distributed. This long-tailed distribution assumes that images of certain classes significantly outnumber others, which is a realistic assumption for real-world applications. Importantly, LTAD shows that prior work on evenly + +distributed data performs suboptimally on LT data. Meanwhile, recent works [16, 22, 55, 61, 79, 84] approach AD with online learning to leverage abnormal images during test time. However, they do not study the case where the classes follow a long-tail distribution, where the performance gap of head and tail classes in the online setting may be significant during offline training. This motivates us to study the long-tailed online AD (LTOAD) task by proposing a benchmark and a novel model. + +At first glance, it may seem that one can take a state-of-the-art (SOTA) long-tailed offline model, e.g., LTAD [34], and combine it with online learning to solve LTOAD. However, several limitations within the LTAD's framework prevent it from being effective in the online setting. First, LTAD is a class-aware method, where they assume that the class information is given. However, class labels are not available in the online setting. Instead, a class-agnostic method is required as illustrated in Fig. 1. Second, LTAD's encoder-decoder architecture does not leverage the latest vector quantization VAE [20, 57, 59], which is more effective, as shown in recent work. Third, the prompt learning designs of LTAD and other AD works [7, 9, 10, 37, 40, 51, 52, 67, 100–102] ignore some aspects of abnormal prompts, e.g., LTAD does not support class-specific abnormal prompts. + +To address these challenges, we propose a novel model that is class-agnostic, with a vector quantized VAE, and a comprehensive prompt learning framework, for the task of long-tailed online anomaly detection. Experiments show that our method outperforms the SOTA baselines significantly by at least $4\%$ on several detection benchmarks without access to the class labels and the number of classes. + +In the following sections, we first discuss how to make existing class-aware methods class-agnostic, specifically on LTAD (Sec. 2). Next, we define the task of long-tailed online AD and present our solution (Sec. 3). Finally, we provide details of each proposed module (Sec. 4), followed by the experimental results (Sec. 5) and related works (Sec. 6). + +Our contributions are summarized as follows: + +![](images/20cb2d241de4f3bc7ba6d5ecd50ea8c58f6a56ae60b1fc51f72d3f30dfea936e.jpg) +(a) class-aware offline learning + +![](images/595e5781067a46ce1900a46aa8faef3fcbf054dc131740b3254b001111813ea5.jpg) +(c)ours offline learning + +![](images/df1a074b2bf483e936c909220bf062352879fdb379fb68aaf547db1b2dd9584d.jpg) +(b) class-aware online learning + +![](images/501cf3b0e2aa87c7e0109537f6810ad71d7e82902b6bdffa4740a054959915bd.jpg) +(d)ours online learning +Figure 1. Comparison of LTOAD to class-aware anomaly detection methods on offline and online learning. (a) Class-aware methods have a class-specific module for each class $c$ in the class set $\mathcal{C}$ . (b) These methods cannot work in online learning when the class labels are unavailable. (c) LTOAD solves this problem by learning a concept set $\widehat{\mathcal{C}}$ to approximate $\mathcal{C}$ . We note that $\widehat{K} = |\widehat{\mathcal{C}}|$ does not need to match $K = |\mathcal{C}|$ . (d) In online learning, LTOAD can weight input images of seen and unseen classes with existing concept-specific modules, i.e., $\{p_c\}_{\widehat{c} \in \widehat{\mathcal{C}}}$ . + +- We propose the task and benchmark for long-tailed online anomaly detection (LTOAD). +- We propose a class-agnostic framework for both offline and online learning that removes the constraint of requiring additional class information in previous work. Along with vector-quantization VAE and a comprehensive prompt learning scheme that benefits both offline and online learning. +- LTOAD outperforms SOTA offline long-tailed methods and online baselines on MVTec, VisA, DAGM, and UniMedical. It also shows promising generalizability to cross-dataset settings. + +# 2. Making LTAD class-agnostic + +# 2.1. Background + +Anomaly Detection (AD). Given an image $X \in \mathbb{R}^{H \times W \times 3}$ of height $H$ and width $W$ , an AD model $\mathrm{F}_{\theta}$ , parametrized by $\theta$ , aims to predict an abnormal map $\widehat{\pmb{Y}} \in \{0,1\}^{H \times W}$ or an abnormal label $y \in \{0,1\}$ indicating whether the image is abnormal or not, where $\widehat{\pmb{Y}_i} = \mathrm{F}_{\theta}(\pmb{X}_i)$ . + +Beyond the standard setting, LTAD [34] proposes a more challenging setup of unsupervised unified AD that focuses on an unevenly distributed training set. Each class $c$ in the class set $\mathcal{C}$ belongs to either the set of "head classes" $\mathcal{C}^{\text{head}}$ or "tail classes" $\mathcal{C}^{\text{tail}}$ depending on the number of times a class appears in the data set. That is, images from the classes in $\mathcal{C}^{\text{head}}$ significantly outnumber those from $\mathcal{C}^{\text{tail}}$ . + +Vision and language models (VLM) in AD. Recent works + +[7, 9, 10, 25, 34, 37, 40, 49, 51, 52, 67, 94, 100] have shown the benefit of foundation VLM for AD. A foundation model consists of a image encoder $\mathbf{E}^{\mathrm{I}}$ and an text encoder $\mathbf{E}^{\mathrm{T}}$ . The encoders encode images and text to a shared space such that cosine similarly between the embeddings captures the relationship between image and text. We use the notation $\langle \cdot \rangle$ for cosine similarity that implicitly supports broadcasting, i.e., when $\langle \cdot \rangle$ is between a matrix with a vector, we broadcast on the spatial dimension. + +In more detail, recent works [30, 34, 57, 62] extract fine-grained image features from intermediate layers or blocks of $\mathbb{E}^{\mathrm{I}}$ . We denote these intermediate visual feature maps as $F^{\mathrm{i}}$ and the final visual feature vectors as $f^{\mathrm{f}}$ , i.e., $(F^{\mathrm{i}}, f^{\mathrm{f}}) = \mathbb{E}^{\mathrm{I}}(X)$ . The precise descriptions of $F^{\mathrm{i}}$ and $f^{\mathrm{f}}$ are provided in Appendix A1.5. + +Class-aware AD. Existing works [7, 9, 10, 22, 25, 34, 37, 40, 49, 51, 52, 57, 61, 67, 84, 94] consist of class-aware modules in their architecture. These works require an additional class label $c \in \mathcal{C}$ for each image as input. Specifically, they assume that the class $c$ of the input image is also provided to the model, i.e., $\widehat{\pmb{Y}}_i = \mathbb{F}_\theta(\pmb{X}_i, c_i)$ . As depicted in Fig. 1, a hard switching mechanism is used in their framework to select the output of a specific class-aware module, i.e., $M_{c_i}$ , to use for each instance. In this work, we do not assume that the class label is provided, as it would typically be unavailable in the online setting. Instead, we develop a class-agnostic approach, which we describe next. + +# 2.2. Concept set for removing class-awareness + +To remove the requirement of having the class label $c$ , we introduce a concept set $\widehat{\mathcal{C}}$ where we assume that the class information $c$ can be represented as a composition of multiple concepts in $\widehat{\mathcal{C}}$ . For example, the class 'transistor' is related to and derived from concepts 'semiconductors' and 'circuits'. In other words, for each image of class $c$ , instead of applying a hard one-hot label, we employ a soft weighting mechanism and assign a soft label $p \in \mathbb{R}^{\widehat{K}}$ where $\widehat{K} = |\widehat{\mathcal{C}}|$ . There are a few advantages. First, we can reduce $\widehat{K}$ for model efficiency, since $\widehat{\mathcal{C}}$ does not need to match $\mathcal{C}$ exactly and can be much smaller. Next, this soft weighting mechanism is applicable when an unseen class $\notin \mathcal{C}$ appears in a realistic testing scenario, while previous class-aware works [34, 57] would fail to function. + +For this approach to be effective, the concept set $\widehat{C}$ should be representative enough to cover the image classes $\mathcal{C}$ of interest. Instead of manually selecting the set $\widehat{C}$ , we leverage the zero-shot capability of foundation models where $\widehat{C}$ is learned with only the visual information of the training set and without seeing any class labels. + +Given the training dataset $\mathcal{D}^{\mathrm{T}} = \{\pmb{X}_i\}_{i=1}^N$ and a vocabulary set $\mathcal{V} = \{v_j\}_{j=1}^{|\mathcal{V}|}$ , we first measure the pairwise similarity score $S_{ij}$ between $\pmb{f}_i^{\mathrm{f}}$ and each vocabulary $v_j$ , i.e., $S_{ij} = \langle \pmb{f}_i^{\mathrm{f}}, \mathbb{E}^{\mathrm{T}}(v_j) \rangle$ . We then conduct a majority vote to + +![](images/8334c713c5c11b349ca2ac79a9020e77c77f433c4349defef23765013161ce0a.jpg) + +![](images/0967cc4716a63926e6fb6ce8e0a6ba99c293ee1a63b44acbdebcddf97d0e55f7.jpg) +Figure 2. Proposed class-agnostic pipeline. We construct $\widehat{C}$ , and the correspondent normal prompts $\mathcal{P}^{\mathrm{n}}$ and abnormal prompts $\mathcal{P}^{\mathrm{a}}$ . The concept score $p$ is assigned to each image $X$ by computing the similarity between $X$ and each $\widehat{c} \in \widehat{C}$ . It then controls the soft switching mechanism in our class-agnostic reconstruction module R and semantics module S. R reconstructs $F^{\mathrm{i}}$ into $F^{\mathrm{r}}$ through Concept VQ and output $\widehat{Y}^{\mathrm{R}}$ by measuring the dissimilarity between $F^{\mathrm{i}}$ and $F^{\mathrm{r}}$ . S compares the similarity map $S^{\mathrm{n}}$ of $F^{\mathrm{i}}$ and normal prompt features and the similarity map $S^{\mathrm{a}}$ of $F^{\mathrm{i}}$ and abnormal prompt features to output $\widehat{Y}^{\mathrm{S}}$ . The final prediction $\widehat{Y}$ is aggregated from $\widehat{Y}^{\mathrm{R}}$ and $\widehat{Y}^{\mathrm{S}}$ . + +pick the top- $\widehat{K}$ vocabularies with the overall highest similarity scores across all images. This voting result is used to initialize $\widehat{\mathcal{C}}$ . Specifically, we create embeddings $t_{\widehat{c}} = \mathbb{E}^{\mathrm{T}}(\widehat{c})$ for all $\widehat{c} \in \widehat{\mathcal{C}}$ and make them learnable so that we can find more representative concept features during training. + +At test time, a soft label, i.e., concept score $\pmb{p} \in [0,1]^{\widehat{K}}$ , is predicted for each image $\pmb{X}$ . We compute $\pmb{p}$ by measuring its similarity to all concepts $\widehat{c} \in \widehat{\mathcal{C}}$ , i.e., + +$$ +\boldsymbol {p} = \operatorname {S o f t M a x} \left(\left\{\left\langle \boldsymbol {f} ^ {\mathrm {f}}, \boldsymbol {t} _ {\widehat {c}} \right\rangle \right\} _ {\widehat {c} \in \widehat {\mathcal {C}}}\right). \tag {1} +$$ + +As shown in Fig. 1, we incorporate a soft switching mechanism and the final output is a composition of each $M_{\widehat{c}}$ weighted by $p$ , i.e., $\sum_{\widehat{c} = 1}^{\widehat{K}} p_{\widehat{c}} M_{\widehat{c}}$ . We will now describe the high-level idea of how to incorporate soft switching into the AD pipeline. + +# 2.3. Proposed class-agnostic AD pipeline + +Our class-agnostic model is based on the general two-branch pipeline of LTAD [34] as it is the SOTA method, however, we note that the idea of soft switching can be applied to other class-aware models as well. Leveraging the previously acquired $\pmb{p}$ , our soft switching mechanism can learn to composite outputs resulting from any multiple concepts or class-specific submodules. + +As shown in Fig. 2, we deploy a similar overall pipeline as LTAD but with several key differences: (a) we introduce $\widehat{C}$ as the cornerstone of our class-agnostic framework; (b) we replace the LTAD's class-aware modules with our class-agnostic LTOAD design; (c) for the model architecture in R, we use our concept-based vector quantization VAE (Concept VQ); (d) we implement a comprehensive prompt learning. We provide the details of each module in Sec. 4. + +# 3. Long-tailed online anomaly detection + +# 3.1. Task formulation + +Given a model $\mathbb{F}_{\theta_0}$ where $\theta_0$ is the parameters trained offline on an LTAD dataset, the goal of LTOAD is to update the parameters $\theta_t$ in an online manner to improve the performance on a data stream $\widetilde{\pmb{X}}_{\leq t} = \left[\widetilde{\pmb{X}}_1,\dots ,\widetilde{\pmb{X}}_t\right]$ where each image $\widetilde{\pmb{X}}$ comes sequentially. Formally, we focus on improving the accuracy of $\mathbb{F}_{\theta_t}(\widetilde{\pmb{X}}_{\leq t}) = \widehat{\pmb{Y}}_t$ , where $\widehat{\pmb{Y}}_t$ is the prediction at $t$ . Note that $\widetilde{\pmb{X}}_{\leq t}$ is an ordered list, i.e., LTOAD is evaluated sequentially. + +Benchmark. We evaluate LTOAD by considering the any- $\Delta$ inference setting where the model is updated on small batches of data samples of size $\Delta$ . We split the data into two sets: (a) an online training set $\mathcal{D}^0 = \widetilde{\boldsymbol{X}}_{\leq N^0}$ where $N^0$ denotes the number of online training samples and (b) an evaluation set $\mathcal{D}^E$ . Given the training set $\mathcal{D}^0$ and the model parameters $\theta_{t+1} \gets \mathcal{A}(\theta_t; \mathcal{D}_t^0)$ is updated based on the $t$ -th batch of samples $\mathcal{D}_t^0 = \widetilde{\boldsymbol{X}}_{(t-1)\Delta:t\Delta}$ using an online learning algorithm $\mathcal{A}$ . The model is then evaluated on $\mathcal{D}^E$ at every step $t$ . + +Following standard metrics [34], we use AUROC to evaluate the performance of detection (Det.) and segmentation (Seg.). We report the average, over $\mathcal{C}^{\mathrm{head}}$ , $\mathcal{C}^{\mathrm{tail}}$ , and $\mathcal{C}$ , at each intermediate steps $t' \in [1, T-1]$ in the data stream. + +Dataset setup. As the data comes in a stream, a model's online learning performance is highly related to the ordering data in $\mathcal{D}^0$ . To study the effect, we sequentially split $\mathcal{D}^0$ into sessions which corresponds to a sublist of the dataset. In each session, the distribution of the data is similar. + +We consider two settings, disjoint and blurry, based on whether the class subset in each session overlaps or not. In a disjoint setting, the class subset in each session does not + +Config. Definition +$B$ a blurry stream where the whole stream is ordered randomly + $B-HF$ a blurry stream where $\mathcal{C}^{\text {head }}$ are more likely to appear first + $B-TF$ a blurry stream where $\mathcal{C}^{\text {tail }}$ are more likely to appear first + $D2-HF$ a disjoint stream with 2 sessions where all $\mathcal{C}^{\text {head }}$ are in the first session + $D2-TF$ a disjoint stream with 2 sessions where all $\mathcal{C}^{\text {tail }}$ are in the first session + $D5-HF$ a disjoint stream with 5 sessions where $\mathcal{C}^{\text {head }}$ concentrate in the first few sessions + $D5-TF$ a disjoint stream with 5 sessions where $\mathcal{C}^{\text {tail }}$ concentrate in the first few sessions + $D5-M$ a disjoint stream with 5 sessions where $\mathcal{C}^{\text {head }}$ and $\mathcal{C}^{\text {tail }}$ are mixed in each session + +Table 1. Configurations for $\mathcal{D}^0$ . We define 8 configurations with combinations of different session type $\in \{$ blurry, disjoint $\}$ and ordering type $\in \{$ head-first, head-first, else $\}$ . + +overlap with that of any other sessions, but in a blurry setting, the class subset of each session can overlap. + +Next, data order is also important in the performance due to overfitting and forgetting, e.g., if $\mathcal{D}^0$ has more $\mathcal{C}^{\mathrm{tail}}$ at the beginning, the performance of $\mathcal{C}^{\mathrm{head}}$ might drop first due to overfitting while the performance of $\mathcal{C}^{\mathrm{tail}}$ plateaus at the end of the stream due to forgetting. Therefore, we also consider whether $\mathcal{C}^{\mathrm{head}}$ or $\mathcal{C}^{\mathrm{tail}}$ are more likely to appear first in $\mathcal{D}^0$ to study the effect of forgetting and overfitting. Overall, we design 8 settings for $\mathcal{D}^0$ with different session and ordering types combinations, as defined in Tab. 1. Please refer to our Appendix A2 for more details. + +# 3.2. Anomaly adaptive online learning + +We propose an Anomalous Adaptive learning algorithm $\mathcal{A}^{\mathrm{AA}}$ to tackle the issues of overfitting and forgetting for LTOAD. As LTOAD is an unsupervised one-class classification task, a naive online algorithm $\mathcal{A}^{\mathrm{N}}$ is to directly use the same loss function as offline learning. Note that $\mathcal{A}^{\mathrm{N}}$ is not ideal as it does not leverage the unlabeled abnormal image in the online learning data. + +To take advantage of this difference, our anomaly adaptive online learning algorithm $\mathcal{A}^{\mathrm{AA}}$ computes a pseudo abnormal map $\widehat{M} = \mathcal{T}(\widehat{Y})$ acquired by thresholding $\widehat{Y}$ with a function $\mathcal{T}$ to adjust the behavior of the original losses. The algorithm is summarized in Alg. 1 where we denote these adjustments as a mask-dependent loss function $\widetilde{\mathcal{L}} (\widehat{M})$ . As abnormal images in the online training set are unavailable in offline training, we value these samples more in the loss calculation and assign a larger weight $\beta$ to images with higher $\widehat{y}$ . Finally, we use the exponential moving average (EMA) for more stable updates. Please see the Appendix A1.5 for details. + +# 4. Model details + +We now provide the description for each of the modules as overviewed in Fig. 2. Additional implementation details are provided in the Appendix A1. + +# 4.1. Prompt learning + +Based on the acquired $\hat{\mathcal{C}}$ , we construct normal and abnormal prompt sets $\mathcal{P}^{\mathrm{n}}$ and $\mathcal{P}^{\mathrm{a}}$ for LTOAD to learn the textual + +Algorithm 1 Anomaly Adaptive $(\mathcal{A}^{\mathrm{AA}})$ Online learning +Input: offline-trained model with its parameters $\mathbb{F}_{\theta_0}$ , online training data stream $\mathcal{D}^0$ , batch size $\Delta$ , threshold function $\mathcal{T}$ , mask-dependent loss function $\widehat{\mathcal{L}}$ , reduction operation $r$ , hyper-parameters $\tau$ , $\beta$ , $\gamma$ +Output: $\Theta = [\theta_{1},\dots ,\theta_{T}]$ , where $T = |\mathcal{D}^0 | / \Delta$ +1: Initialize $\Theta = []$ $\widetilde{\theta}_0 = \theta_0$ +2: for $t = 1$ to $T$ do +3: Acquire batch $\mathcal{D}_t^0 = \left[\widetilde{\pmb{X}}_i\right]_{i = (t - 1)\Delta ,\dots ,t\Delta}$ +4: for $j = 1$ to $\Delta$ do +5: Forward pass $\widehat{\pmb{Y}}_j = \mathbb{F}_{\widetilde{\theta}_{t - 1}}(\mathcal{D}_t^0 [j])$ +6: Compute pseudo abnormal label $\widehat{M}_j = \mathcal{T}(\widehat{Y}_j)$ +7: Compute mask-dependent loss $\tilde{\mathcal{L}}_j = \tilde{\mathcal{L}} (\widehat{\mathcal{M}}_j)$ +8: Compute gradient weight $\lambda_{j}\gets \beta$ if $r(\widehat{Y}_j)\geq \tau$ else 1 +9: end for +10: $\widetilde{\theta}_t\gets \widetilde{\theta}_{t - 1} + \frac{1}{\Delta}\sum_j\lambda_j\nabla \widetilde{\mathcal{L}}_j$ +11: Update with EMA $\theta_t\gets \gamma \theta_{t - 1} + (1 - \gamma)\widetilde{\theta}_t$ +12: $\Theta .\mathrm{append}(\theta_t)$ +13: end for +14: return $\Theta$ + +semantics. $\mathcal{P}^{\mathrm{n}}$ contains $\widehat{K}$ prompts where each $P_{\widehat{c}}^{\mathrm{n}} \in \mathcal{P}^{\mathrm{n}}$ is initialized with the template of "a normal $\widehat{c}$ ". In contrast, $\mathcal{P}^{\mathrm{a}}$ is initialized by asking a conversational AI [13, 64] "For $\widehat{c}$ , what are the 5 anomalies we are likely to observe?" for each $\widehat{c}$ . Each normal prompt feature is defined as $\boldsymbol{n}_{\widehat{c}} = \mathbb{E}^{\mathrm{T}}(P_{\widehat{c}}^{\mathrm{m}})$ while each abnormal prompt feature is defined as $\boldsymbol{a}_{\widehat{c},i} = \mathbb{E}^{\mathrm{T}}(P_{\widehat{c},i}^{\mathrm{a}})$ for $i \in [1,5]$ . In addition to the learnable $\widehat{C}$ , $\boldsymbol{n}_{\widehat{c}}$ and $\boldsymbol{a}_{\widehat{c},i}$ are also learnable so that LTOAD can find more accurate abnormal semantics for each concept. In general, we learn $\boldsymbol{t}_{\widehat{c}}, \boldsymbol{a}_{\widehat{c}}$ , and multiple $\boldsymbol{n}_{\widehat{c}}$ for each concept. + +# 4.2. Reconstruction module + +The goal of the reconstruction module R is to reconstruct the features $\pmb{F}^{\mathrm{i}}$ of normal images for extracting cues about whether an image is anomalous. Since R is trained only on the representation of normal features, it would be unable to accurately reconstruct abnormal features. Therefore, a discrepancy between the input and reconstructed features provides a useful cue that the input is abnormal. In more detail, given an input feature $\pmb{F}^{\mathrm{i}}$ , R reconstructs it to $\pmb{F}^{\mathrm{r}}$ and output an abnormal map prediction: + +$$ +\widehat {\boldsymbol {Y}} ^ {\mathrm {R}} = \frac {1}{2} \left(\mathbf {1} - \left\langle \boldsymbol {F} ^ {\mathrm {i}}, \boldsymbol {F} ^ {\mathrm {r}} \right\rangle\right) \tag {2} +$$ + +by measuring the discrepancy between $F^{\mathrm{i}}$ and $F^{\mathrm{r}}$ via cosine similarity. We now introduce the background of HVQ [57] before we describe the module's architecture and concept codebooks used to create $F^{\mathrm{r}}$ . + +Background. Lu et al. [57] proposes hierarchical vector-quantization (HVQ) to construct multiple expertise modules + +for each class to solve unified AD. A hard switching mechanism is applied to choose the correspondent modules for each class. They also introduce a hierarchical architecture to connect high-level and low-level feature maps. Given multiscale feature maps $\{Z_{l}\}_{l = 1,\dots ,L}$ , where $L$ is the number of layers, they build a hierarchical class-specific module $\mathbb{Q}_{l,c}$ for each $c\in \mathcal{C}$ . During training, each $\mathbb{Q}_{l,c}$ only sees $Z_{l}$ that comes from the images of class $c$ . The output of each $\mathbb{Q}_{l,c}$ is denoted as $Q_{l}$ . Each $Q_{l}$ is conditioned on $Z_{l}$ , and the output of the previous module, i.e. $Q_{l - 1}$ . Specifically, + +$$ +\boldsymbol {Q} _ {l} = \left\{ \begin{array}{l l} \mathbb {Q} _ {l, c} (\boldsymbol {Z} _ {l}), & \text {i f} l = 1 \\ \mathbb {Q} _ {l, c} (\boldsymbol {Z} _ {l} \oplus \boldsymbol {Q} _ {l - 1}), & \text {o t h e r w i s e .} \end{array} \right. \tag {3} +$$ + +Here, $\oplus$ denotes concatenation on the channel dimension. The outputs $\{Q_{l}\}_{l = 1,\dots ,L}$ are then fed into their proposed hierarchical decoder [57] for reconstruction. + +The class-specific hierarchical modules $\mathbb{Q}_{l,c}$ are quantization modules [20, 59, 82] that learn to quantify visual feature representations. For readability, we remove the subscripts of $\mathbb{Q}_{l,c}$ . Each $\mathbb{Q}$ stores a learnable codebook $B$ . The codebook contains $M$ codes of $D$ dimension, i.e. $B \in \mathbb{R}^{M \times D}$ . Given a 2D input feature $Z$ , $\mathbb{Q}$ queries its codebook and finds the nearest prototypes $Q$ . For each pixel location $(h, w)$ , they find $Q[h, w] = b_{m^*}$ , where $m^* \triangleq \arg \min_m \| Z[h, w] - B[m] \|_2^2$ . + +Module architecture and concept codebooks. Given an input $F^{\mathrm{i}}$ , we first project it onto a visual-text shared space through a layer-wise projection module $\mathbb{E}^{\mathrm{p}}$ [34], i.e., $F^{\mathrm{p}} = \mathbb{E}^{\mathrm{p}}(F^{\mathrm{i}})$ . For feature reconstruction, we use a VQ-autoencoder dubbed "Concept-VQ," modified from class-specific HVQ [57] for our concept-specific objective. We also use a prompt-conditioned generator $G$ for data augmentation to generate pseudo-normal features $F^{\mathrm{n}}$ and pseudoabnormal features $F^{\mathrm{a}}$ . These pseudo features provide additional guidance in training $R$ . Finally, we use the soft label $p$ as a weighting mechanism for our concept-specific designs in Concept-VQ and $G$ . + +Our Concept-VQ contains an encoder $\mathbb{E}^{\mathrm{R}}$ , a decoder $\mathsf{D}^{\mathbb{R}}$ , and hierarchical quantization modules between them. For each $l \in [1, \dots, L]$ and each $\widehat{c} \in \widehat{\mathcal{C}}$ , we construct a concept-specific quantization module $\mathbb{Q}_{l,\widehat{c}}$ . To capture the text-visual representation of $\mathcal{D}^{\mathrm{T}}$ , our concept codebook $\pmb{B}_{l,\widehat{c}}$ in each $\mathbb{Q}_{l,\widehat{c}}$ is initialize $\pmb{B}_{l,\widehat{c}}$ by sampling $M$ codes around its correspondent $\pmb{t}_{\widehat{c}}$ . + +Prompt-conditioned data augmentation. During training, prior works [21, 34, 57, 89] create pseudo-abnormal features for data augmentation by adding noise to the input features, i.e., feature jittering. These prior works ignore the semantics of the $F^{\mathrm{p}}$ . Instead, we build a generator $G$ that synthesizes pseudo features conditioned on prompt features $\mathbf{a}$ and $\mathbf{n}$ . We generate pseudo-normal features $F^{\mathrm{n}}$ conditioned on $F^{\mathrm{p}}$ and $\mathbf{n}$ while we generate pseudo-abnormal + +features $\mathbf{F}^{\mathrm{a}}$ conditioned on $\mathbf{F}^{\mathrm{p}}$ and $\mathbf{a}$ . + +# 4.3. Semantics module + +The semantics module S leverages the cross-modality knowledge of abnormality in $\mathbf{E}^{\mathrm{I}}$ and $\mathbf{E}^{\mathrm{T}}$ to extract cues on whether an image is anomalous. It learns the semantic similarities or dissimilarities between visual features and normal texts, e.g., "normal," versus abnormal texts, e.g., "broken" by comparing the aggregated normal similarity scores $S^n$ and abnormal scores $S^a$ . The higher $S^a - S^n$ is, the more likely it is abnormal. Given the projected feature $F^{\mathrm{p}}$ , we denote the similarity difference as a functionality of S, i.e., $\mathrm{S}(F^{\mathrm{p}}) = S^{\mathrm{a}} - S^{\mathrm{n}}$ . Then, the final prediction $\widehat{Y}^{\mathrm{S}} \triangleq (\mathrm{S}(F^{\mathrm{p}}) + 2)/4$ , where we scale to the range of [0, 1]. + +Module architecture. Given $\pmb{F}^{\mathrm{p}}$ , we first compute the similarity map $S_{\hat{c}}^{\mathrm{n}}$ between it and the concept-specific normal prompt features, i.e., $S_{\hat{c}}^{\mathrm{n}} = \langle \pmb{F}^{\mathrm{p}}, \pmb{n}_{\hat{c}} \rangle$ . The overall normal similarity map $S^{\mathrm{n}}$ is then $\sum_{\hat{c}} p_{\hat{c}} S_{\hat{c}}^{\mathrm{n}}$ . Similarly, we also compute the overall abnormal similarity map $S^{\mathrm{a}}$ for $F^{\mathrm{p}}$ . + +For segmentation, the final prediction $\widehat{\pmb{Y}}$ is the elementwise harmonic mean [40, 51] of the two intermediate predictions, i.e., + +$$ +\hat {\boldsymbol {Y}} = \left(\alpha \left(\hat {\boldsymbol {Y}} ^ {\mathrm {R}}\right) ^ {- 1} + (1 - \alpha) \left(\hat {\boldsymbol {Y}} ^ {\mathrm {S}}\right) ^ {- 1}\right) ^ {- 1}, \tag {4} +$$ + +where $\alpha$ controls the weight between them. For detection, we apply an AvgPool2d with a kernel size of 16 on $\widehat{\pmb{Y}}$ before we reduce the map to its maximum value as $\widehat{y}$ . We denote this reduction operation as $\widehat{y} = r(\widehat{\pmb{Y}})$ . + +# 5. Experiments + +We first study the offline LTAD setting as an offline model parameter $\theta_0$ is the starting point for an online method. Importantly, we show that class-agnostic models are effective in offline LTAD without degrading performance. Finally, we conduct experiments on LTOAD with our $\mathcal{A}^{\mathrm{AA}}$ in the online setting and evaluate cross-dataset generalizability. Additional experiment details are in the Appendix A3. + +# 5.1. Offline long-tailed settings + +Experiment setup. Following the LTAD benchmark [34], we report on MVtec [5] and VisA [103]. Each dataset has multiple long-tailed settings dependent on the degree of imbalance between $\mathcal{C}^{\mathrm{head}}$ and $\mathcal{C}^{\mathrm{tail}}$ . The imbalance type $\in \{exp, step\}$ describes whether the numbers of classes decay exponentially or drop like a step function; while the imbalance factor $\in \{100, 200\}$ measures the ratio between the number of the most populated class to the least one. A larger imbalance factor means a more challenging setting. We follow exactly the same LTAD training and testing settings unless otherwise specified. Additional results on DAGM [85] and Uni-Medical [4, 95] are provided in the Appendix A3. + +
MethodCAexp100exp200step100step200
Det.Seg.Det.Seg.Det.Seg.Det.Seg.
RegAD [36]X82.4395.20N/AN/A81.5495.10N/AN/A
AnomalyGPT [25]X87.4489.6885.8090.1585.9589.2882.4789.45
LTAD [34]X88.8694.4686.0594.1887.3693.8385.6092.12
HVQ [57]X87.4395.2585.3494.7985.3994.1783.0092.61
LTOAD*X93.1295.0191.8894.3692.0294.7287.9593.34
UniAD [89]✓87.7093.9586.2193.2683.3791.4781.3292.29
MoEAD [62]✓84.7394.3483.2393.9684.4093.7683.1392.76
LTOAD✓93.4295.2192.0294.9492.3395.1188.6294.00
+ +Table 2. Comparison (↑) on offline MVTec in image-level AUROC for anomaly detection (Det.) and pixel-level AUROC for anomaly segmentation (Seg.). The column CA (class-agnostic) indicates whether a method requires class names or the number of classes during training or not (require: X; not require: √). RegAD needs at least 2 images per class and thus is not applicable on exp200 and step200. We mark the best and second best performances in bold and underline. + +
MethodCAexp100exp200step100step200
Det.Seg.Det.Seg.Det.Seg.Det.Seg.
RegAD [36]X71.3694.4071.3694.4071.8094.9971.6594.52
AnomalyGPT [25]X70.3480.3269.7879.4871.9882.3069.7881.97
LTAD [34]X80.0095.5680.2195.3684.8096.5784.0396.27
HVQ [57]X78.6396.1878.0395.3276.6895.6375.5295.50
LTOAD*X84.8996.4284.8496.4487.7197.0786.5296.89
UniAD [89]✓77.3195.0376.8794.8078.8396.0477.6495.66
MoEAD [62]✓73.8994.3473.5893.9874.1294.3776.7794.82
LTOAD✓85.2696.9185.2496.9788.3297.5487.0497.23
+ +Baselines & implementation details. We compare with RegAD [36], UniAD [89], and LTAD [34]. We benchmark the SOTA MoEAD [62] and HVQ [57] to highlight the gain of our concept-specific quantization modules to their class-specific ones. Only UniAD, MoEAD, and LTOAD are class-agnostic and do not need a class label at the test time. We provide a class-aware version of our model, LTOAD*. + +Quantitative & qualitative comparison. Tab. 2 and 3 show the comparisons of LTOAD with the baselines on MVTec and VisA. We observe that LTOAD outperforms the SOTA on most settings while using no class information. LTOAD achieves a larger gain over Det. On average across all settings, LTOAD achieves $+4.63\% / +0.58\%$ gain versus the second best baseline excluding LTOAD* on MVTec Det. / Seg. As for VisA, LTOAD achieves $+4.30\% / +1.29\%$ gain. + +Next, comparing LTAD and LTOAD*, we find that being class-agnostic is beneficial. We hypothesize that this is because the classes can now share knowledge through the soft labels. Finally, we show qualitative results in Fig. 3. We observe that LTOAD predictions are closer to ground truth than HVQ and LTAD. + +# 5.2. Online long-tailed benchmark + +Experiment setup. For the online setting, we report on the + +![](images/a6e92571081f1ab204c28fdef9b447b672f5b0d207bcde266f18cf7ecfd61254.jpg) +truth anomaly masks are circled in pink. + +Table 3. Comparison (↑) on offline VisA. Format follows Tab. 2. + +
MethodCA OnlineBB-HFB-TFD2-HFD2-TFD5-HFD5-TFD5-MAvg.
LTAD [34]XX93.4693.4693.4693.4693.4693.4693.4693.4693.46
LTAD [34]XA^N94.1293.8094.0793.7094.0893.4993.8493.6593.84
HVQ [57]XX95.0295.0295.0295.0295.0295.0295.0295.0295.02
HVQ [57]XA^N95.4495.3395.2895.2295.2995.0395.2195.0595.23
LTOAD✓X95.0895.0895.0895.0895.0895.0895.0895.0895.08
LTOAD✓A^N95.4195.2295.3295.2395.0494.9194.7294.7995.08
LTOAD✓A^N95.3695.3095.3295.2395.2895.2695.2595.2195.28
+ +Table 4. Comparison (↑) on online MVTec in pixel-level AU-ROC for anomaly segmentation across 8 configurations of $\mathcal{D}^0$ . All models are pre-trained on MVTec exp100 [34]. + +![](images/2bb728d42b8f60bad3a2ff0ec501b37237e7cf22da23bc68b962552bdccf7d09.jpg) +Figure 4. Performance curve of $\mathcal{A}^{\mathrm{N}}$ and $\mathcal{A}^{\mathrm{AA}}$ in pixel-level AUROC on $\mathcal{C}^{\mathrm{head}}$ and $\mathcal{C}^{\mathrm{tail}}$ . They are offline-trained on $exp100$ [34] and online-trained on D5-HF or D5-TF. + +same datasets as the offline benchmark [34], i.e., MVtec, VisA, and DAGM. We split the testing set into $\mathcal{D}^0$ and $\mathcal{D}^{\mathrm{E}}$ with a 3:7 ratio. For each dataset, we create the 8 settings defined in Tab. 1. + +For baseline comparisons, we extend HVQ and LTAD to have online capability by applying our naive online learning algorithm $\mathcal{A}^{\mathsf{N}}$ since they are the best offline models on MVTec. Note that both methods are not class-agnostic and needs class labels as extra information in online learning, which is not available/necessary for our approach. + +Quantitative comparison. We report the online Seg. performance at final time step $t = T$ on MVtec exp100 in Tab. 4. The results on other settings/datasets are in the + +
MethodOnlineMVTecVisADAGM
→ VisA→ DAGM→ MVTec→ DAGM→ MVTec→ VisA
LTOADX84.3891.2782.7379.1679.5181.13
LTOAD✓92.3796.8789.8496.7988.3792.41
+ +Table 5. Cross-dataset anomaly segmentation in pixel-level AUROC. All models are pre-trained on their $exp100$ [34]. + +
Set IDDomainSizeElements
CN/A15zipper, pill, capsule, grid, transistor, carpet, metal nut, wood, leather, screw, tile, cable, toothbrush, hazelnut, bottle*
RandomOut10refinery, finger, vault, exam, salad, sailor, dove, dealer, well, humanoid
SmallIn3semiconductor, zipper, beech
LTOAD CIn10semiconductor, zipper, beech, microscopy*, antibiotics, medicines, circuit , mahogany, hardwood, walnut
+ +Appendix A4. Overall, $\mathcal{A}^{\mathrm{N}}$ performs well in the simplest setting $(B)$ , but degrades on more difficult settings, e.g., $D2$ -TF, $D5$ -HF, etc. On average, LTOAD with $\mathcal{A}^{\mathrm{AA}}$ performs the best even without class labels as extra information. We plot LTOAD's performance curve of $\mathsf{F}_{\theta_t}$ over $t$ with $\Delta = 16$ under the two most difficult settings ( $D5$ -HF and $D5$ -TF) in Fig. 4, where $\mathcal{A}^{\mathrm{N}}$ falls off on $\mathcal{C}^{\mathrm{head}} / \mathcal{C}^{\mathrm{head}}$ under $D5$ -HF/D5-TF while $\mathcal{A}^{\mathrm{AA}}$ improves steadily. Empirically, LTOAD is also more efficient than LTAD because $\widehat{K} < K$ . Cross-dataset evaluation. Tab. 5 shows that LTOAD can adapt to unseen domains even if it is not designed for cross-dataset evaluation. For each $\mathcal{D}_1 \to \mathcal{D}_2$ experiment, we pre-train LTOAD on $exp100$ of $\mathcal{D}_1$ then evaluated on $B$ of $\mathcal{D}_2$ without and with online learning. LTOAD with online learning outperforms the one without it. + +# 5.3. Additional analysis + +Analysis on initialization of $\widehat{C}$ . We now analyze the effect of our concept learning component. All experiments are done under MVtec exp100 and $B$ . We highlight the importance of our concept learning compared to two baselines: initializing with a random concept set or with a small one. + +Tab. 6 shows the learned initialization of $\widehat{C}$ versus the ground truth $C$ . We observe that our concept learning can find the $\widehat{C}$ initialization which represents similar semantics in $C$ , e.g., the bottle* images (see the 1st image in Fig. 3) in MVtec are bottom-only views, which look like an microscopy*. + +Next, we experiment with the two baselines to generate initialization for $\widehat{\mathcal{C}}$ . In Random of Tab. 6, instead of learning from $\mathcal{D}^{\mathrm{T}}$ , i.e., in-domain concepts, we randomly select 10 + +Table 6. Classes vs. concepts on MVTec. Words with similar semantics have the same color, e.g., wood-related words are colored in brown and electronics-related words are colored in purple. + +
Exp.\( \widehat{C} \)\( \widehat{K} \)RSGDetectionSegmentation
\( C^{tail} \)\( C^{head} \)\( C \)\( C^{tail} \)\( C^{head} \)\( C \)
RandomO-D10✓✓✓84.7399.3191.5494.4695.3994.89
SmallI-D3✓✓✓85.6698.8991.8394.2495.3094.74
NoRI-D10✘✓✘53.8456.2554.9760.7363.4662.00
NoSI-D10✓✘✓85.1699.2191.7294.5795.2494.88
NoGI-D10✓✓✘85.6399.5692.1394.0895.6294.80
LTOADI-D10✓✓✓88.1499.4593.7294.8395.6495.36
+ +Table 7. Analysis and ablation on LTOAD using the same metrics as Tab. 4 under MVTec $\exp {100}$ [34] and $B$ . We report the performances on ${\mathcal{C}}^{\text{head }}$ and ${\mathcal{C}}^{\text{tail }}$ . I-D/O-D: in/out-domain. + +
Exp.C̅LPaL MCSDetectionSegmentation
CtailCheadCCtailCheadC
FixCX√√√84.6598.1390.9494.57
FixP√X√√85.5999.4592.0694.52
SingleP√√X√85.4199.3791.9294.64
ShareP√√√X87.0399.3592.7895.00
LTOAD√√√√88.1499.4593.7294.83
+ +Table 8. Ablations on concept $\hat{C}$ & abnormal prompt $\mathcal{P}^{\mathrm{a}}$ learning under MVtec exp100 [34] on the same metrics as Tab. 7. Acronyms: L: learned; M: multiple; CS: concept-specific. + +nouns in ALIGN's $\mathcal{V}$ unrelated to the industrial settings of MVTec, i.e., out-domain concepts. In other words, the semantics of $\mathcal{C}$ and Random do not overlap. In Small of Tab. 6, instead of choosing $\widehat{K} = 10$ , we select only the top- $\widehat{K} = 3$ vocabularies as $\widehat{\mathcal{C}}$ . The semantics of Small partially covers the semantics of $\mathcal{C}$ . Tab. 7 shows the quantitative results. + +We report the performance of Random and Small in Tab. 7. Even with an out-domain initialization, compared to HVQ in Tab. 2, which uses $\mathcal{C}$ for their class-specific quantization modules, Random still performs competitively. Specifically, we observe a $+4.11\%$ gain in Det. versus HVQ. On top of that, with in-domain $\widehat{\mathcal{C}}$ as a more aligned representation, LTOAD in Tab. 7 achieves the best image-level AUROC. In particular, we further improve it by $+1.88\%$ . We have a similar observation initializing with Small, which outperforms HVQ on Det. Having sufficiently large $\widehat{\mathcal{C}}$ , LTOAD with $\widehat{K} = 10$ gives more gain. + +Finally, although some $c \in \mathcal{C}$ are not presented in our $\widehat{\mathcal{C}}$ , LTOAD still adapts and outperforms the baselines in Tab. 7. This suggests that LTOAD can work even with a foundation model whose $\mathcal{V}$ and $\mathcal{C}$ are disjoint. However, this is not necessarily true for prior class-aware methods as they implicitly assume that $\mathcal{C} \subseteq \mathcal{V}$ . + +Ablation studies on R, S, and G. In Tab. 7, we also ablate LTOAD components (see Sec. 4.2 and 4.3) to justify their design. All experiments are done under the MVTec exp100 and $B$ setting. NoR ablates R. We also remove G, $\widehat{\mathbf{Y}}^{\mathrm{R}}$ from Eq. 4 and related losses (see Appendix A1.1). In contrast, NoS ablates S. We also remove $\widehat{\mathbf{Y}}^{\mathrm{S}}$ and related + +
Unsupervised AD Method ConditionsUnsupervised AD Categories
G0G1G2G3G4G5G6G7LTOAD
unified modelX✓✓X✓✓X✓✓
online learning capabilityXXXXXX✓✓✓
incorporate VLM/LLM(s)XXX✓✓✓XX✓
can handle long-tailed dataXXXXX✓XX✓
class-agnosticX✓XXXXX✓✓
+ +Table 9. Characteristic comparison between LTOAD and prior unsupervised AD works (e.g., $\mathcal{G}_0$ : [12, 14, 15, 19, 26, 27, 38, 48, 54, 59, 63, 70, 72, 73, 75, 80, 87, 90, 91, 93, 97], $\mathcal{G}_1$ : [2, 11, 21, 30, 31, 33, 39, 46, 50, 62, 66, 76, 77, 87-89, 95, 99, 102], $\mathcal{G}_2$ : [57], $\mathcal{G}_3$ : [37, 51, 52], $\mathcal{G}_4$ : [7, 9, 10, 25, 40, 49, 67, 94], $\mathcal{G}_5$ : [34], $\mathcal{G}_6$ : [22, 61, 84], $\mathcal{G}_7$ : [55, 79]). + +losses. We find from NoR that S alone is not an effective branch while NoS is far more competitive. However, when trained jointly (LTOAD) and compared to NoS, S helps R improve the Det. accuracy on $\mathcal{C}^{\text {tail }}$ notably. Specifically, S improves the image-level AUROC on $\mathcal{C}^{\text {tail }}$ by $2.98\%$ . + +NoG replaces our prompt-conditioned $\mathsf{G}$ with unconditional feature jittering [34, 57, 89] for data augmentation. With the guidance of $\mathcal{P}^{\mathrm{n}}$ and $\mathcal{P}^{\mathrm{a}}$ , $\mathsf{G}$ creates high-quality pseudo-features especially beneficial for $\mathcal{C}^{\mathrm{tail}}$ . We observe a gain of $+2.51\% / +0.75\%$ in Det. / Seg. on $\mathcal{C}^{\mathrm{tail}}$ . + +Ablation studies on learnable concepts/prompts. Tab. 8 analyzes the effect of each attribute detailed in Tab. 10, where we demonstrate that our prompt learning design is comprehensive. All experiments are performed in the setting MVTec exp100 and B. FixC freezes $\hat{\mathcal{C}}$ . FixP freezes $\mathcal{P}^a$ . In SingleP, $\mathcal{P}^a$ stores only one abnormal prompt for each concept. In ShareP, all concepts share the same abnormal prompts. Each attribute contributes significantly to the Det. performance of $\mathcal{C}^{\mathrm{tail}}$ and $\mathcal{C}$ . + +# 6. Related work + +In Tab. 9, we group the literature of unsupervised AD based on their assumptions or target tasks. Notably, LTOAD (1) is a unified model, (2) has online learning capability, (3) uses VLM/LLM priors, (4) handles long-tailed training data distributions, and (5) is class-agnostic. + +Unsupervised anomaly detection. $\mathcal{G}_0$ methods train an independent predictor per class. In contrast, UniAD [89] challenges to learn one unified predictor for all classes. More follow-up research falls into this group $(\mathcal{G}_1)$ . Inspired by prior work [56, 60, 71], MoEAD [62] uses mixture-of-experts. HVQ [57] uses class-specific quantization modules in their VQ-VAE [82] architecture, but HVQ also induces the constraint of requiring class information as inputs, making it not class-agnostic, i.e., $\mathcal{G}_2$ . + +Foundation model based AD. With advancement in foundation models, recent works focused on zero-shot [7, 9, 24, 40, 50, 67, 100, 102] or few-shot [2, 25, 51-53, 101] unsupervised AD where the normal training images are limited (or none), they [7, 9, 25, 40, 51, 52, 67, 100] often + +
Prompt Design Conditions of Unsupervised AD MethodsPrompt Design Categories in AD
G8G9G10G11G12LTOAD
learnable class/concept names✓XXXX✓
support >1 APsXX✓✓✓✓
learnable APsX✓XX✓✓
class/concept specific APsXXX✓✓✓
+ +Table 10. Characteristic comparison w.r.t. prompt design between LTOAD and prior unsupervised AD works which use LLM(s) with prompts (e.g., $\mathcal{G}_8$ : [34, 67], $\mathcal{G}_9$ : [10], $\mathcal{G}_{10}$ : [9, 37, 40, 52, 101], $\mathcal{G}_{11}$ : [7], $\mathcal{G}_{12}$ : [51, 102]). AP: abnormal prompts. + +turn themselves to the rich knowledge in foundation models [29, 41, 68, 81, 98] or conversational AI [64]. These methods can be grouped into either $\mathcal{G}_3$ or $\mathcal{G}_4$ , depending on whether they are unified model. To query the language models precisely, $\mathcal{G}_3$ and $\mathcal{G}_4$ need the class names or labels as extra inputs and are thus not class-agnostic. + +Although normal images are limited in zero/few-shot AD, the distribution is even since the number of images per class is the same. Consequently, LTAD [34], i.e., $\mathcal{G}_5$ , explores the long-tailed setting where the training set is highly unevenly distributed. They rely on the pre-trained ALIGN [41] and learn the pseudo-class names. We highlight the differences between LTOAD prompt learning to related works in Tab. 10. In summary, LTOAD supports (a) learnable concept names, (b) learnable abnormal prompts (AP), (c) multiple APs, and (d) concept-specific APs. + +Online learning AD. Online (or continual/incremental) learning [23, 44, 69, 83] investigates how the model learns continuously on a data stream. Recent works [22, 55, 61, 79, 84], i.e. $\mathcal{G}_6$ and $\mathcal{G}_7$ , explore unsupervised online learning AD. Differently, we propose the task of long-tailed online anomaly detection to study the effect of head/tail classes. + +# 7. Conclusion + +We propose the benchmark of long-tailed online anomaly detection, studying the online performance of head and tail classes. We identify the shortcomings of class-aware methods and propose class-agnostic LTOAD. It learns to approximate the ground truth class set while using foundation models for our novel concept and abnormal prompt learning framework. We also propose $\mathcal{A}^{\mathrm{AA}}$ , a learning algorithm to adapt the online stream with potentially abnormal inputs. LTOAD outperforms the SOTA in most offline and online settings without additional class information. Although not specifically designed for domain generalization, LTOAD can adapt to unseen domains. We believe that LTOAD will advance the field toward a more realistic anomaly detection setting that is closer to real-world applications. + +Acknowledgments. CAY and KCP were supported by Mitsubishi Electric Research Laboratories. CAY and RAY were in part supported by an NSF Award #2420724. + +# References + +[1] Akshatha Arodi, Margaux Luck, Jean-Luc Bedwani, Aldo Zaimi, Ge Li, Nicolas Pouliot, Julien Beaudry, and Gaétan Marceau Caron. CableInspect-AD: An expert-annotated anomaly detection dataset. In Proc. NeurIPS, 2024. 1 +[2] Yuhu Bai, Jiangning Zhang, Zhaofeng Chen, Yuhang Dong, Yunkang Cao, and Guanzhong Tian. Dual-path frequency discriminators for few-shot anomaly detection. Knowledge-Based Systems, 2024. 8 +[3] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 5 +[4] Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxiang Zhang, and Xingyu Li. BMAD: Benchmarks for medical anomaly detection. In Proc. CVPR, 2024. 5, 1 +[5] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. MVTec AD-a comprehensive real-world dataset for unsupervised anomaly detection. In Proc. CVPR, 2019. 1, 5, 3, 4 +[6] Patrick Bilic, Patrick Christ, Hongwei Bran Li, Eugene Vorontsov, Avi Ben-Cohen, Georgios Kaissis, Adi Szeskin, Colin Jacobs, Gabriel Efrain Humpire Mamani, Gabriel Chartrand, et al. The liver tumor segmentation benchmark (lits). Medical Image Analysis, 2023. 5 +[7] Yunkang Cao, Xiaohao Xu, Chen Sun, Yuqi Cheng, Zongwei Du, Liang Gao, and Weiming Shen. Segment any anomaly without training via hybrid prompt regularization. arXiv preprint arXiv:2305.10724, 2023. 1, 2, 8 +[8] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1 +[9] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. AdaCLIP: Adapting CLIP with hybrid learnable prompts for zero-shot anomaly detection. In Proc. ECCV, 2024. 1, 2, 8 +[10] Pi-Wei Chen, Jerry Chun-Wei Lin, Jia Ji, Feng-Hao Yeh, Zih-Ching Chen, and Chao-Chun Chen. Human-free automated prompting for vision-language anomaly detection: Prompt optimization with meta-guiding prompt scheme. arXiv preprint arXiv:2406.18197, 2024. 1, 2, 8 +[11] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. In Proc. ECCV, 2024. 8 +[12] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 8 +[13] Microsoft Corporation. Microsoft Copilot, 2023. 4, 2 +[14] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. PaDiM: a patch distribution modeling framework for anomaly detection and localization. In Proc. ICPR, 2021. 8 + +[15] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proc. CVPR, 2022. 8 +[16] Keval Doshi and Yasin Yilmaz. Continual learning for anomaly detection in surveillance videos. In Proc. CVPR Workshop, 2020. 1 +[17] Hugging Face. Transformers: Align model documentation. https://huggingface.co/docs/ transformers/model_doc align.2 +[18] Yehonatan Fridman, Matan Rusanovsky, and Gal Oren. ChangeChip: A reference-based unsupervised change detection for PCB defect detection. In Proc. PAINE, 2021. 1 +[19] Matic Fučka, Vitjan Zavrtanik, and Danijel Skočaj. TransFusion-a transparency-based diffusion model for anomaly detection. In Proc. ECCV, 2024. 8 +[20] Hugo Gangloff, Minh-Tan Pham, Luc Courtrai, and Sébastien Lefèvre. Leveraging vector-quantized variational autoencoder inner metrics for anomaly detection. In Proc. ICPR, 2022. 1, 5 +[21] Bin-Bin Gao. Learning to detect multi-class anomalies with just one normal image prompt. In Proc. ECCV, 2024. 5, 8 +[22] Han Gao, Huiyuan Luo, Fei Shen, and Zhengtao Zhang. Towards total online unsupervised anomaly detection and localization in industrial vision. arXiv preprint arXiv:2305.15652, 2023. 1, 2, 8 +[23] Yasir Ghunaim, Adel Bibi, Kumail Alhamoud, Motasem Alfarra, Hasan Abed Al Kader Hammoud, Ameya Prabhu, Philip HS Torr, and Bernard Ghanem. Real-time evaluation in online continual learning: A new hope. In Proc. CVPR, 2023. 8 +[24] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Hao Li, Ming Tang, and Jinqiao Wang. Filo: Zero-shot anomaly detection by fine-grained description and high-quality localization. In Proc. ACM MM, 2024. 8 +[25] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. AnomalyGPT: Detecting industrial anomalies using large vision-language models. In Proc. AAAI, 2024. 2, 6, 8, 5, 7, 9, 10, 11, 12, 13, 14 +[26] Guan Gui, Bin-Bin Gao, Jun Liu, Chengjie Wang, and Yunsheng Wu. Few-shot anomaly-driven generation for anomaly classification and segmentation. In Proc. ECCV, 2024. 8 +[27] Jia Guo, Haonan Han, Shuai Lu, Weihang Zhang, and Huiqi Li. Absolute-unified multi-class anomaly detection via class-agnostic distribution alignment. arXiv preprint arXiv:2404.00724, 2024.8 +[28] Ahmed Hamada. Br35H: Brain tumor detection 2020, 2020. 1 +[29] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xi-angyu Yue. OneLLM: One framework to align all modalities with language. In Proc. CVPR, 2024. 8 +[30] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. MambaAD: Exploring state space models for multi-class unsupervised anomaly detection. In Proc. NeurIPS, 2024. 2, 8 + +[31] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proc. AAAI, 2024. 8 +[32] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, 2016. 4 +[33] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. In Proc. ECCV, 2024. 8 +[34] Chih-Hui Ho, Kuan-Chuan Peng, and Nuno Vasconcelos. Long-tailed anomaly detection with learnable class names. In Proc. CVPR, 2024. 1, 2, 3, 5, 6, 7, 8, 4, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21 +[35] Junjie Hu, Yuanyuan Chen, and Zhang Yi. Automated segmentation of macular edema in oct using deep neural networks. Medical image analysis, 2019. 5 +[36] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In Proc. ECCV, 2022. 6, 5, 7, 8, 9, 10, 11, 12, 13, 14 +[37] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xinchao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proc. CVPR, 2024. 1, 2, 8 +[38] Jeeho Hyun, Sangyun Kim, Giyoung Jeon, Seung Hwan Kim, Kyunghoon Bae, and Byung Jun Kang. ReConPatch: Contrastive patch representation learning for industrial anomaly detection. In Proc. WACV, 2024. 8 +[39] Brian KS Isaac-Medina, Yona Falinie A Gaus, Neelanjan Bhowmik, and Toby P Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In Proc. ECCV, 2024. 8 +[40] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. WinCLIP: Zero/few-shot anomaly classification and segmentation. In Proc. CVPR, 2023. 1, 2, 5, 8 +[41] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proc. ICML, 2021. 8, 2, 4 +[42] Pranita Balaji Kanade and PP Gumaste. Brain tumor detection using MRI images. *Brain*, 2015. 1 +[43] Felipe Campos Kitamura. Head CT - hemorrhage, 2018. 1 +[44] Hyunseo Koh, Dahyun Kim, Jung-Woo Ha, and Jonghyun Choi. Online continual learning on class incremental blurry task configuration with anytime inference. In Proc. ICLR, 2022. 8 +[45] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, 2015. 5 + +[46] Joo Chan Lee, Taejune Kim, Eunbyung Park, Simon S Woo, and Jong Hwan Ko. Continuous memory representation for anomaly detection. In Proc. ECCV, 2024. 8 +[47] Jan Lehr, Jan Philipps, Alik Sargsyan, Martin Pape, and Jörg Krüger. AD3: Introducing a score for anomaly detection dataset difficulty assessment using VIADUCT dataset. In Proc. ECCV, 2024. 1 +[48] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. CutPaste: Self-supervised learning for anomaly detection and localization. In Proc. CVPR, 2021. 8 +[49] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. ClipSAM: CLIP and SAM collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 2, 8 +[50] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. MuSc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. In Proc. ICLR, 2024. 8 +[51] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. PromptAD: Learning prompts with only normal samples for few-shot anomaly detection. In Proc. CVPR, 2024. 1, 2, 5, 8 +[52] Yuanwei Li, Elizaveta Ivanova, and Martins Bruveris. FADE: Few-shot/zero-shot anomaly detection engine using large vision-language model. In Proc. BMVC, 2024. 1, 2, 8 +[53] Jingyi Liao, Xun Xu, Manh Cuong Nguyen, Adam Goodge, and Chuan Sheng Foo. COFT-AD: Contrastive fine-tuning for few-shot anomaly detection. IEEE TIP, 2024. 8 +[54] Chieh Liu, Yu-Min Chu, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning diffusion models for multi-view anomaly detection. In Proc. ECCV, 2024. 8 +[55] Jiaqi Liu, Kai Wu, Qiang Nie, Ying Chen, Bin-Bin Gao, Yong Liu, Jinbao Wang, Chengjie Wang, and Feng Zheng. Unsupervised continual anomaly detection with contrastively-learned prompt. In Proc. AAAI, 2024. 1, 8 +[56] Ruiying Lu, Bo Chen, Jianqiao Sun, Wenchao Chen, Penghui Wang, Yuanwei Chen, Hongwei Liu, and Pramod K Varshney. Heterogeneity-aware recurrent neural network for hyperspectral and multispectral image fusion. IEEE Journal of Selected Topics in Signal Processing, 2022. 8 +[57] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. In Proc. NeurIPS, 2023. 1, 2, 4, 5, 6, 8, 7, 9, 10, 11, 12, 13, 14 +[58] Jianhong Ma, Tao Zhang, Cong Yang, Yangjie Cao, Lipeng Xie, Hui Tian, and Xuexiang Li. Review of wafer surface defect detection methods. *Electronics*, 2023. 1 +[59] Sergio Naval Marimont and Giacomo Tarroni. Anomaly detection through latent space restoration using vector quantized variational autoencoders. In Proc. ISBI, 2021. 1, 5, 8 +[60] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review, 2014. 8 + +[61] Declan GD McIntosh and Alexandra Branzan Albu. Unsupervised, online and on-the-fly anomaly detection for nonstationary image distributions. In Proc. ECCV, 2024. 1, 2, 8 +[62] Shiyuan Meng, Wenchao Meng, Qihang Zhou, Shizhong Li, Weiye Hou, and Shibo He. MoEAD: A parameter-efficient model for multi-class anomaly detection. In Proc. ECCV, 2024. 2, 6, 8, 5, 7, 9, 10, 11, 12, 13, 14 +[63] Shancong Mou, Xiaoyi Gu, Meng Cao, Haoping Bai, Ping Huang, Jiulong Shan, and Jianjun Shi. RGI: Robust GAN-inversion for mask-free image inpainting and unsupervised pixel-wise anomaly detection. In Proc. ICLR, 2023. 8 +[64] OpenAI. ChatGPT, 2023. 4, 8, 2 +[65] Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM computing surveys (CSUR), 2021. 1 +[66] Sukanya Patra and Souhaib Ben Taieb. Revisiting deep feature reconstruction for logical and structural industrial anomaly detection. TMLR, 2024. 8 +[67] Zhen Qu, Xian Tao, Mukesh Prasad, Fei Shen, Zhengtao Zhang, Xinyi Gong, and Guiguang Ding. VCP-CLIP: A visual context prompting model for zero-shot anomaly segmentation. In Proc. ECCV, 2024. 1, 2, 8 +[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. ICML, 2021. 8 +[69] Siddeshwar Raghavan, Jiangpeng He, and Fengqing Zhu. DELTA: Decoupling long-tailed online continual learning. In Proc. CVPR, 2024. 8 +[70] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In Proc. ICML, 2015. 8 +[71] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Proc. NeurIPS, 2021. 8 +[72] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proc. CVPR, 2022. 8 +[73] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proc. WACV, 2021. 8 +[74] Radhwan AA Saleh, Mehmet Zeki Konyar, Kaplan Kaplan, and H Metin Ertunç. Tire defect detection model using machine learning. In Proc. eSmarTA, 2022. 1 +[75] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proc. CVPR, 2021. 8, 6, 7 +[76] Jian Shi, Pengyi Zhang, Ni Zhang, Hakim Ghazzai, and Peter Wonka. Dissolving is amplifying: Towards fine-grained anomaly detection. In Proc. ECCV, 2024. 8 +[77] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. GeneralAD: Anomaly detection across domains by attending to distorted features. In Proc. ECCV, 2024. 8 + +[78] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, 2019. 2 +[79] Jiaqi Tang, Hao Lu, Xiaogang Xu, Ruizheng Wu, Sixing Hu, Tong Zhang, Tsz Wa Cheng, Ming Ge, Ying-Cong Chen, and Fugee Tsung. An incremental unified framework for small defect inspection. In Proc. ECCV, 2024. 1, 8 +[80] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan Duong, Chanh D Tr Nguyen, and Steven QH Truong. Revisiting reverse distillation for anomaly detection. In Proc. CVPR, 2023. 8 +[81] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 8 +[82] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In Proc. NeurIPS, 2017. 5, 8 +[83] Liuyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. IEEE TPAMI, 2024. 8 +[84] Shenxing Wei, Xing Wei, Zhiheng Ma, Shaochen Zhang Songlin Dong, and Yihong Gong. Few-shot online anomaly detection and segmentation. Knowledge-Based Systems, 2024. 1, 2, 8 +[85] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, 2007. 5, 3, 4 +[86] Shuai Yang, Zhifei Chen, Pengguang Chen, Xi Fang, Yixun Liang, Shu Liu, and Yingcong Chen. Defect spectrum: A granular look of large-scale defect datasets with rich semantics. In Proc. ECCV, 2024. 1 +[87] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. GLAD: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. In Proc. ECCV, 2024. 8 +[88] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. In Proc. ECCV, 2024. +[89] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Proc. NeurIPS, 2022. 5, 6, 8, 7, 9, 10, 11, 12, 13, 14 +[90] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. FastFlow: Unsupervised anomaly detection and localization via 2D normalizing flows. arXiv preprint arXiv:2111.07677, 2021.8 +[91] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoçaj. DRAEM-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proc. ICCV, 2021. 8, 6, 7 +[92] Chongsheng Zhang, George Almpanidis, Gaojuan Fan, Binquan Deng, Yanbo Zhang, Ji Liu, Aouaidjia Kamel, Paolo Soda, and João Gama. A systematic review on long-tailed learning. arXiv preprint arXiv:2408.00483, 2024. 1 + +[93] Gongjie Zhang, Kaiwen Cui, Tzu-Yi Hung, and Shijian Lu. Defect-GAN: High-fidelity defect synthesis for automated defect inspection. In Proc. WACV, 2021. 8 +[94] Jiangning Zhang, Xuhai Chen, Zhucun Xue, Yabiao Wang, Chengjie Wang, and Yong Liu. Exploring grounding potential of VQA-oriented GPT-4V for zero-shot anomaly detection. arXiv preprint arXiv:2311.02612, 2023. 2, 8 +[95] Jiangning Zhang, Xuhai Chen, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, Ming-Hsuan Yang, and Dacheng Tao. Exploring plain ViT reconstruction for multiclass unsupervised anomaly detection. In Proc. IJCAI Workshop on Anomaly Detection with Foundation Models, 2024. 5, 8, 1 +[96] Jiangning Zhang, Haoyang He, Zhenye Gan, Qingdong He, Yuxuan Cai, Zhucun Xue, Yabiao Wang, Chengjie Wang, Lei Xie, and Yong Liu. A Der: A comprehensive benchmark for multi-class visual anomaly detection. arXiv preprint arXiv:2406.03262, 2024. 1, 4 +[97] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. DeSTSeg: Segmentation guided denoising student-teacher for anomaly detection. In Proc. CVPR, 2023. 8 +[98] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. MetaTransformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 8 +[99] Ying Zhao. OmniAL: A unified cnn framework for unsupervised anomaly localization. In Proc. CVPR, 2023. 8 +[100] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In Proc. ICLR, 2024. 1, 2, 8 +[101] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proc. CVPR, 2024. 8 +[102] Jiawen Zhu, Yew-Soon Ong, Chunhua Shen, and Guansong Pang. Fine-grained abnormality prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2410.10289, 2024. 1, 8 +[103] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. In Proc. ECCV, 2022. 1, 5, 3, 4 \ No newline at end of file diff --git a/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/images.zip b/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fb7dcebe144ef9aeef4ecd2bc57ccc63e2683493 --- /dev/null +++ b/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:76a0bac49828ad14a9795b5b53b3f3a3b4a669b101fabae56d995d6755628b8d +size 555477 diff --git a/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/layout.json b/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..72db5ae441d55068bb12717875b5ada1dcb7f685 --- /dev/null +++ b/2025/Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts/layout.json @@ -0,0 +1,17235 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 103, + 452, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 103, + 452, + 139 + ], + "spans": [ + { + "bbox": [ + 159, + 103, + 452, + 139 + ], + "type": "text", + "content": "Toward Long-Tailed Online Anomaly Detection through Class-Agnostic Concepts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 160, + 451, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 160, + 451, + 175 + ], + "spans": [ + { + "bbox": [ + 157, + 160, + 451, + 175 + ], + "type": "text", + "content": "Chiao-An Yang" + }, + { + "bbox": [ + 157, + 160, + 451, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 160, + 451, + 175 + ], + "type": "text", + "content": " Kuan-Chuan Peng" + }, + { + "bbox": [ + 157, + 160, + 451, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 160, + 451, + 175 + ], + "type": "text", + "content": " Raymond A. Yeh" + }, + { + "bbox": [ + 157, + 160, + 451, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 176, + 545, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 176, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 64, + 176, + 545, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 176, + 545, + 190 + ], + "type": "text", + "content": "Department of Computer Science, Purdue University " + }, + { + "bbox": [ + 64, + 176, + 545, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 176, + 545, + 190 + ], + "type": "text", + "content": "Mitsubishi Electric Research Laboratories" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 178, + 191, + 429, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 178, + 191, + 429, + 205 + ], + "spans": [ + { + "bbox": [ + 178, + 191, + 429, + 205 + ], + "type": "text", + "content": "1{yang2300,rayyeh}@purdue.edu 2kpeng@merl.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 234, + 200, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 234, + 200, + 247 + ], + "spans": [ + { + "bbox": [ + 151, + 234, + 200, + 247 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 259, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 259, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 259, + 296, + 510 + ], + "type": "text", + "content": "Anomaly detection (AD) identifies the defect regions of a given image. Recent works have studied AD, focusing on learning AD without abnormal images, with long-tailed distributed training data, and using a unified model for all classes. In addition, online AD learning has also been explored. In this work, we expand in both directions to a realistic setting by considering the novel task of long-tailed online AD (LTOAD). We first identified that the offline state-of-the-art LTAD methods cannot be directly applied to the online setting. Specifically, LTAD is class-aware, requiring class labels that are not available in the online setting. To address this challenge, we propose a class-agnostic framework for LTAD and then adapt it to our online learning setting. Our method outperforms the SOTA baselines in most offline LTAD settings, including both the industrial manufacturing and the medical domain. In particular, we observe " + }, + { + "bbox": [ + 56, + 259, + 296, + 510 + ], + "type": "inline_equation", + "content": "+4.63\\%" + }, + { + "bbox": [ + 56, + 259, + 296, + 510 + ], + "type": "text", + "content": " image-AUROC on MVTec even compared to methods that have access to class labels and the number of classes. In the most challenging long-tailed online setting, we achieve " + }, + { + "bbox": [ + 56, + 259, + 296, + 510 + ], + "type": "inline_equation", + "content": "+0.53\\%" + }, + { + "bbox": [ + 56, + 259, + 296, + 510 + ], + "type": "text", + "content": " image-AUROC compared to baselines. Our LTOAD benchmark is released here1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 533, + 135, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 135, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 135, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 552, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 612 + ], + "type": "text", + "content": "Anomaly detection (AD) [8, 65, 96] is the task of identifying defect regions in a given image. This task is important due to its high valued applications in industrial manufacturing [1, 5, 18, 47, 58, 74, 86, 103] and medical [28, 42, 43, 47] settings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 612, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 295, + 696 + ], + "type": "text", + "content": "Recently, LTAD [34] explored the long-tailed (LT) [92] setting of unsupervised one-class AD, where abnormal images are not given during training and the training set is unevenly distributed. This long-tailed distribution assumes that images of certain classes significantly outnumber others, which is a realistic assumption for real-world applications. Importantly, LTAD shows that prior work on evenly" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 235, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 553, + 342 + ], + "type": "text", + "content": "distributed data performs suboptimally on LT data. Meanwhile, recent works [16, 22, 55, 61, 79, 84] approach AD with online learning to leverage abnormal images during test time. However, they do not study the case where the classes follow a long-tail distribution, where the performance gap of head and tail classes in the online setting may be significant during offline training. This motivates us to study the long-tailed online AD (LTOAD) task by proposing a benchmark and a novel model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 346, + 554, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 554, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 554, + 537 + ], + "type": "text", + "content": "At first glance, it may seem that one can take a state-of-the-art (SOTA) long-tailed offline model, e.g., LTAD [34], and combine it with online learning to solve LTOAD. However, several limitations within the LTAD's framework prevent it from being effective in the online setting. First, LTAD is a class-aware method, where they assume that the class information is given. However, class labels are not available in the online setting. Instead, a class-agnostic method is required as illustrated in Fig. 1. Second, LTAD's encoder-decoder architecture does not leverage the latest vector quantization VAE [20, 57, 59], which is more effective, as shown in recent work. Third, the prompt learning designs of LTAD and other AD works [7, 9, 10, 37, 40, 51, 52, 67, 100–102] ignore some aspects of abnormal prompts, e.g., LTAD does not support class-specific abnormal prompts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "type": "text", + "content": "To address these challenges, we propose a novel model that is class-agnostic, with a vector quantized VAE, and a comprehensive prompt learning framework, for the task of long-tailed online anomaly detection. Experiments show that our method outperforms the SOTA baselines significantly by at least " + }, + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 313, + 540, + 554, + 624 + ], + "type": "text", + "content": " on several detection benchmarks without access to the class labels and the number of classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 627, + 553, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 553, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 553, + 698 + ], + "type": "text", + "content": "In the following sections, we first discuss how to make existing class-aware methods class-agnostic, specifically on LTAD (Sec. 2). Next, we define the task of long-tailed online AD and present our solution (Sec. 3). Finally, we provide details of each proposed module (Sec. 4), followed by the experimental results (Sec. 5) and related works (Sec. 6)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 701, + 512, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 512, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 512, + 712 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 259, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 259, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 259, + 712 + ], + "type": "text", + "content": "MethodCAexp100exp200step100step200Det.Seg.Det.Seg.Det.Seg.Det.Seg.RegAD [36]X82.4395.20N/AN/A81.5495.10N/AN/AAnomalyGPT [25]X87.4489.6885.8090.1585.9589.2882.4789.45LTAD [34]X88.8694.4686.0594.1887.3693.8385.6092.12HVQ [57]X87.4395.2585.3494.7985.3994.1783.0092.61LTOAD*X93.1295.0191.8894.3692.0294.7287.9593.34UniAD [89]✓87.7093.9586.2193.2683.3791.4781.3292.29MoEAD [62]✓84.7394.3483.2393.9684.4093.7683.1392.76LTOAD✓93.4295.2192.0294.9492.3395.1188.6294.00", + "image_path": "5c415d9087a51cc53e641fc692629a32ba5a131a001d5c02e8f82df2ca729f01.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 273, + 294, + 376 + ], + "blocks": [ + { + "bbox": [ + 55, + 178, + 296, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 296, + 266 + ], + "type": "text", + "content": "Table 2. Comparison (↑) on offline MVTec in image-level AUROC for anomaly detection (Det.) and pixel-level AUROC for anomaly segmentation (Seg.). The column CA (class-agnostic) indicates whether a method requires class names or the number of classes during training or not (require: X; not require: √). RegAD needs at least 2 images per class and thus is not applicable on exp200 and step200. We mark the best and second best performances in bold and underline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 273, + 294, + 376 + ], + "lines": [ + { + "bbox": [ + 57, + 273, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 57, + 273, + 294, + 376 + ], + "type": "table", + "html": "
MethodCAexp100exp200step100step200
Det.Seg.Det.Seg.Det.Seg.Det.Seg.
RegAD [36]X71.3694.4071.3694.4071.8094.9971.6594.52
AnomalyGPT [25]X70.3480.3269.7879.4871.9882.3069.7881.97
LTAD [34]X80.0095.5680.2195.3684.8096.5784.0396.27
HVQ [57]X78.6396.1878.0395.3276.6895.6375.5295.50
LTOAD*X84.8996.4284.8496.4487.7197.0786.5296.89
UniAD [89]✓77.3195.0376.8794.8078.8396.0477.6495.66
MoEAD [62]✓73.8994.3473.5893.9874.1294.3776.7794.82
LTOAD✓85.2696.9185.2496.9788.3297.5487.0497.23
", + "image_path": "aad2d68b0c8382742b071dcd0d7e31e216eac6033e1694a018425ea562374ccd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 409, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 492 + ], + "type": "text", + "content": "Baselines & implementation details. We compare with RegAD [36], UniAD [89], and LTAD [34]. We benchmark the SOTA MoEAD [62] and HVQ [57] to highlight the gain of our concept-specific quantization modules to their class-specific ones. Only UniAD, MoEAD, and LTOAD are class-agnostic and do not need a class label at the test time. We provide a class-aware version of our model, LTOAD*." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 494, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 601 + ], + "type": "text", + "content": "Quantitative & qualitative comparison. Tab. 2 and 3 show the comparisons of LTOAD with the baselines on MVTec and VisA. We observe that LTOAD outperforms the SOTA on most settings while using no class information. LTOAD achieves a larger gain over Det. On average across all settings, LTOAD achieves " + }, + { + "bbox": [ + 55, + 494, + 296, + 601 + ], + "type": "inline_equation", + "content": "+4.63\\% / +0.58\\%" + }, + { + "bbox": [ + 55, + 494, + 296, + 601 + ], + "type": "text", + "content": " gain versus the second best baseline excluding LTOAD* on MVTec Det. / Seg. As for VisA, LTOAD achieves " + }, + { + "bbox": [ + 55, + 494, + 296, + 601 + ], + "type": "inline_equation", + "content": "+4.30\\% / +1.29\\%" + }, + { + "bbox": [ + 55, + 494, + 296, + 601 + ], + "type": "text", + "content": " gain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 602, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 296, + 673 + ], + "type": "text", + "content": "Next, comparing LTAD and LTOAD*, we find that being class-agnostic is beneficial. We hypothesize that this is because the classes can now share knowledge through the soft labels. Finally, we show qualitative results in Fig. 3. We observe that LTOAD predictions are closer to ground truth than HVQ and LTAD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 683, + 219, + 696 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 219, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 219, + 696 + ], + "type": "text", + "content": "5.2. Online long-tailed benchmark" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Experiment setup. For the online setting, we report on the" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 68, + 610, + 242 + ], + "blocks": [ + { + "bbox": [ + 315, + 68, + 610, + 242 + ], + "lines": [ + { + "bbox": [ + 315, + 68, + 610, + 242 + ], + "spans": [ + { + "bbox": [ + 315, + 68, + 610, + 242 + ], + "type": "image", + "image_path": "a6e92571081f1ab204c28fdef9b447b672f5b0d207bcde266f18cf7ecfd61254.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 243, + 462, + 253 + ], + "lines": [ + { + "bbox": [ + 315, + 243, + 462, + 253 + ], + "spans": [ + { + "bbox": [ + 315, + 243, + 462, + 253 + ], + "type": "text", + "content": "truth anomaly masks are circled in pink." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 266, + 553, + 347 + ], + "blocks": [ + { + "bbox": [ + 56, + 380, + 295, + 392 + ], + "lines": [ + { + "bbox": [ + 56, + 380, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 380, + 295, + 392 + ], + "type": "text", + "content": "Table 3. Comparison (↑) on offline VisA. Format follows Tab. 2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 266, + 553, + 347 + ], + "lines": [ + { + "bbox": [ + 315, + 266, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 315, + 266, + 553, + 347 + ], + "type": "table", + "html": "
MethodCA OnlineBB-HFB-TFD2-HFD2-TFD5-HFD5-TFD5-MAvg.
LTAD [34]XX93.4693.4693.4693.4693.4693.4693.4693.4693.46
LTAD [34]XA^N94.1293.8094.0793.7094.0893.4993.8493.6593.84
HVQ [57]XX95.0295.0295.0295.0295.0295.0295.0295.0295.02
HVQ [57]XA^N95.4495.3395.2895.2295.2995.0395.2195.0595.23
LTOAD✓X95.0895.0895.0895.0895.0895.0895.0895.0895.08
LTOAD✓A^N95.4195.2295.3295.2395.0494.9194.7294.7995.08
LTOAD✓A^N95.3695.3095.3295.2395.2895.2695.2595.2195.28
", + "image_path": "b66fc2cdead360336d356ec8f90fd310af447ada72f49e9b9c02cc0995c33166.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 356, + 555, + 389 + ], + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 389 + ], + "type": "text", + "content": "Table 4. Comparison (↑) on online MVTec in pixel-level AU-ROC for anomaly segmentation across 8 configurations of " + }, + { + "bbox": [ + 313, + 356, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^0" + }, + { + "bbox": [ + 313, + 356, + 555, + 389 + ], + "type": "text", + "content": ". All models are pre-trained on MVTec exp100 [34]." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 391, + 552, + 502 + ], + "blocks": [ + { + "bbox": [ + 317, + 391, + 552, + 502 + ], + "lines": [ + { + "bbox": [ + 317, + 391, + 552, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 391, + 552, + 502 + ], + "type": "image", + "image_path": "2bb728d42b8f60bad3a2ff0ec501b37237e7cf22da23bc68b962552bdccf7d09.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "type": "text", + "content": "Figure 4. Performance curve of " + }, + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{\\mathrm{N}}" + }, + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{\\mathrm{AA}}" + }, + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "type": "text", + "content": " in pixel-level AUROC on " + }, + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{head}}" + }, + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{tail}}" + }, + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "type": "text", + "content": ". They are offline-trained on " + }, + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "type": "inline_equation", + "content": "exp100" + }, + { + "bbox": [ + 313, + 506, + 555, + 539 + ], + "type": "text", + "content": " [34] and online-trained on D5-HF or D5-TF." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 556, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 604 + ], + "type": "text", + "content": "same datasets as the offline benchmark [34], i.e., MVtec, VisA, and DAGM. We split the testing set into " + }, + { + "bbox": [ + 313, + 556, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^0" + }, + { + "bbox": [ + 313, + 556, + 555, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 556, + 555, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{E}}" + }, + { + "bbox": [ + 313, + 556, + 555, + 604 + ], + "type": "text", + "content": " with a 3:7 ratio. For each dataset, we create the 8 settings defined in Tab. 1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": "For baseline comparisons, we extend HVQ and LTAD to have online capability by applying our naive online learning algorithm " + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{\\mathsf{N}}" + }, + { + "bbox": [ + 313, + 605, + 554, + 677 + ], + "type": "text", + "content": " since they are the best offline models on MVTec. Note that both methods are not class-agnostic and needs class labels as extra information in online learning, which is not available/necessary for our approach." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Quantitative comparison. We report the online Seg. performance at final time step " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " on MVtec exp100 in Tab. 4. The results on other settings/datasets are in the" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23424" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 294, + 120 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 120 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 120 + ], + "type": "table", + "html": "
MethodOnlineMVTecVisADAGM
→ VisA→ DAGM→ MVTec→ DAGM→ MVTec→ VisA
LTOADX84.3891.2782.7379.1679.5181.13
LTOAD✓92.3796.8789.8496.7988.3792.41
", + "image_path": "1251b746d9d96673fabad7953842162f0d9ce57c87e6a168c35246692c41383c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 156, + 293, + 289 + ], + "blocks": [ + { + "bbox": [ + 56, + 124, + 295, + 148 + ], + "lines": [ + { + "bbox": [ + 56, + 124, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 124, + 295, + 148 + ], + "type": "text", + "content": "Table 5. Cross-dataset anomaly segmentation in pixel-level AUROC. All models are pre-trained on their " + }, + { + "bbox": [ + 56, + 124, + 295, + 148 + ], + "type": "inline_equation", + "content": "exp100" + }, + { + "bbox": [ + 56, + 124, + 295, + 148 + ], + "type": "text", + "content": " [34]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 156, + 293, + 289 + ], + "lines": [ + { + "bbox": [ + 56, + 156, + 293, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 293, + 289 + ], + "type": "table", + "html": "
Set IDDomainSizeElements
CN/A15zipper, pill, capsule, grid, transistor, carpet, metal nut, wood, leather, screw, tile, cable, toothbrush, hazelnut, bottle*
RandomOut10refinery, finger, vault, exam, salad, sailor, dove, dealer, well, humanoid
SmallIn3semiconductor, zipper, beech
LTOAD CIn10semiconductor, zipper, beech, microscopy*, antibiotics, medicines, circuit , mahogany, hardwood, walnut
", + "image_path": "ab40cf4e8a3475732ea536bac2ecb0b6f53cdd46304ffe09ee63b37730248803.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": "Appendix A4. Overall, " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{\\mathrm{N}}" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " performs well in the simplest setting " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "(B)" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": ", but degrades on more difficult settings, e.g., " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "D2" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": "-TF, " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "D5" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": "-HF, etc. On average, LTOAD with " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{\\mathrm{AA}}" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " performs the best even without class labels as extra information. We plot LTOAD's performance curve of " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathsf{F}_{\\theta_t}" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\Delta = 16" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " under the two most difficult settings (" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "D5" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": "-HF and " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "D5" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": "-TF) in Fig. 4, where " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{\\mathrm{N}}" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " falls off on " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{head}} / \\mathcal{C}^{\\mathrm{head}}" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " under " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "D5" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": "-HF/D5-TF while " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{\\mathrm{AA}}" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " improves steadily. Empirically, LTOAD is also more efficient than LTAD because " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\widehat{K} < K" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": ". Cross-dataset evaluation. Tab. 5 shows that LTOAD can adapt to unseen domains even if it is not designed for cross-dataset evaluation. For each " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1 \\to \\mathcal{D}_2" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " experiment, we pre-train LTOAD on " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "exp100" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_1" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " then evaluated on " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_2" + }, + { + "bbox": [ + 54, + 339, + 296, + 520 + ], + "type": "text", + "content": " without and with online learning. LTOAD with online learning outperforms the one without it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 527, + 169, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 169, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 169, + 540 + ], + "type": "text", + "content": "5.3. Additional analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "content": "Analysis on initialization of " + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\widehat{C}" + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "content": ". We now analyze the effect of our concept learning component. All experiments are done under MVtec exp100 and " + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 544, + 295, + 605 + ], + "type": "text", + "content": ". We highlight the importance of our concept learning compared to two baselines: initializing with a random concept set or with a small one." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": "Tab. 6 shows the learned initialization of " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\widehat{C}" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": " versus the ground truth " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": ". We observe that our concept learning can find the " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\widehat{C}" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": " initialization which represents similar semantics in " + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 605, + 295, + 677 + ], + "type": "text", + "content": ", e.g., the bottle* images (see the 1st image in Fig. 3) in MVtec are bottom-only views, which look like an microscopy*." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Next, we experiment with the two baselines to generate initialization for " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{C}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ". In Random of Tab. 6, instead of learning from " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ", i.e., in-domain concepts, we randomly select 10" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 71, + 552, + 174 + ], + "blocks": [ + { + "bbox": [ + 55, + 294, + 295, + 328 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 328 + ], + "type": "text", + "content": "Table 6. Classes vs. concepts on MVTec. Words with similar semantics have the same color, e.g., wood-related words are colored in brown and electronics-related words are colored in purple." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 71, + 552, + 174 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 552, + 174 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 552, + 174 + ], + "type": "table", + "html": "
Exp.\\( \\widehat{C} \\)\\( \\widehat{K} \\)RSGDetectionSegmentation
\\( C^{tail} \\)\\( C^{head} \\)\\( C \\)\\( C^{tail} \\)\\( C^{head} \\)\\( C \\)
RandomO-D10✓✓✓84.7399.3191.5494.4695.3994.89
SmallI-D3✓✓✓85.6698.8991.8394.2495.3094.74
NoRI-D10✘✓✘53.8456.2554.9760.7363.4662.00
NoSI-D10✓✘✓85.1699.2191.7294.5795.2494.88
NoGI-D10✓✓✘85.6399.5692.1394.0895.6294.80
LTOADI-D10✓✓✓88.1499.4593.7294.8395.6495.36
", + "image_path": "033b6727d68c9985376f027ff31bf91f9b1988de76c851f56cdd3b13eb999560.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 218, + 553, + 315 + ], + "blocks": [ + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "lines": [ + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "type": "text", + "content": "Table 7. Analysis and ablation on LTOAD using the same metrics as Tab. 4 under MVTec " + }, + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "type": "inline_equation", + "content": "\\exp {100}" + }, + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "type": "text", + "content": " [34] and " + }, + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "type": "text", + "content": " . We report the performances on " + }, + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "type": "inline_equation", + "content": "{\\mathcal{C}}^{\\text{head }}" + }, + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "type": "inline_equation", + "content": "{\\mathcal{C}}^{\\text{tail }}" + }, + { + "bbox": [ + 313, + 178, + 553, + 212 + ], + "type": "text", + "content": " . I-D/O-D: in/out-domain." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 218, + 553, + 315 + ], + "lines": [ + { + "bbox": [ + 315, + 218, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 315, + 218, + 553, + 315 + ], + "type": "table", + "html": "
Exp.C̅LPaL MCSDetectionSegmentation
CtailCheadCCtailCheadC
FixCX√√√84.6598.1390.9494.57
FixP√X√√85.5999.4592.0694.52
SingleP√√X√85.4199.3791.9294.64
ShareP√√√X87.0399.3592.7895.00
LTOAD√√√√88.1499.4593.7294.83
", + "image_path": "f548e723d6ec71b069ea1e3032c668d55e3dd98fb3d14ff58ac53098645ac9d3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 320, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 354 + ], + "type": "text", + "content": "Table 8. Ablations on concept " + }, + { + "bbox": [ + 313, + 320, + 555, + 354 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 313, + 320, + 555, + 354 + ], + "type": "text", + "content": " & abnormal prompt " + }, + { + "bbox": [ + 313, + 320, + 555, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{a}}" + }, + { + "bbox": [ + 313, + 320, + 555, + 354 + ], + "type": "text", + "content": " learning under MVtec exp100 [34] on the same metrics as Tab. 7. Acronyms: L: learned; M: multiple; CS: concept-specific." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": "nouns in ALIGN's " + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": " unrelated to the industrial settings of MVTec, i.e., out-domain concepts. In other words, the semantics of " + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": " and Random do not overlap. In Small of Tab. 6, instead of choosing " + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\widehat{K} = 10" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": ", we select only the top-" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\widehat{K} = 3" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": " vocabularies as " + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": ". The semantics of Small partially covers the semantics of " + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 365, + 555, + 437 + ], + "type": "text", + "content": ". Tab. 7 shows the quantitative results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": "We report the performance of Random and Small in Tab. 7. Even with an out-domain initialization, compared to HVQ in Tab. 2, which uses " + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": " for their class-specific quantization modules, Random still performs competitively. Specifically, we observe a " + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "inline_equation", + "content": "+4.11\\%" + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": " gain in Det. versus HVQ. On top of that, with in-domain " + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": " as a more aligned representation, LTOAD in Tab. 7 achieves the best image-level AUROC. In particular, we further improve it by " + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "inline_equation", + "content": "+1.88\\%" + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": ". We have a similar observation initializing with Small, which outperforms HVQ on Det. Having sufficiently large " + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": ", LTOAD with " + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\widehat{K} = 10" + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": " gives more gain." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "text", + "content": "Finally, although some " + }, + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "text", + "content": " are not presented in our " + }, + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{C}}" + }, + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "text", + "content": ", LTOAD still adapts and outperforms the baselines in Tab. 7. This suggests that LTOAD can work even with a foundation model whose " + }, + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "text", + "content": " are disjoint. However, this is not necessarily true for prior class-aware methods as they implicitly assume that " + }, + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\subseteq \\mathcal{V}" + }, + { + "bbox": [ + 313, + 569, + 553, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": "Ablation studies on R, S, and G. In Tab. 7, we also ablate LTOAD components (see Sec. 4.2 and 4.3) to justify their design. All experiments are done under the MVTec exp100 and " + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": " setting. NoR ablates R. We also remove G, " + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{Y}}^{\\mathrm{R}}" + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": " from Eq. 4 and related losses (see Appendix A1.1). In contrast, NoS ablates S. We also remove " + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{Y}}^{\\mathrm{S}}" + }, + { + "bbox": [ + 313, + 642, + 554, + 712 + ], + "type": "text", + "content": " and related" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23425" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 149 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 149 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 149 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 149 + ], + "type": "table", + "html": "
Unsupervised AD Method ConditionsUnsupervised AD Categories
G0G1G2G3G4G5G6G7LTOAD
unified modelX✓✓X✓✓X✓✓
online learning capabilityXXXXXX✓✓✓
incorporate VLM/LLM(s)XXX✓✓✓XX✓
can handle long-tailed dataXXXXX✓XX✓
class-agnosticX✓XXXXX✓✓
", + "image_path": "31b2ab0c9aeef87017dfc43d28cf902c252525c9a870451462b3b7a417755f13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "text", + "content": "Table 9. Characteristic comparison between LTOAD and prior unsupervised AD works (e.g., " + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_0" + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "text", + "content": ": [12, 14, 15, 19, 26, 27, 38, 48, 54, 59, 63, 70, 72, 73, 75, 80, 87, 90, 91, 93, 97], " + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_1" + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "text", + "content": ": [2, 11, 21, 30, 31, 33, 39, 46, 50, 62, 66, 76, 77, 87-89, 95, 99, 102], " + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_2" + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "text", + "content": ": [57], " + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_3" + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "text", + "content": ": [37, 51, 52], " + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_4" + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "text", + "content": ": [7, 9, 10, 25, 40, 49, 67, 94], " + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_5" + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "text", + "content": ": [34], " + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_6" + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "text", + "content": ": [22, 61, 84], " + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_7" + }, + { + "bbox": [ + 55, + 152, + 295, + 219 + ], + "type": "text", + "content": ": [55, 79])." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 294, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 294, + 289 + ], + "type": "text", + "content": "losses. We find from NoR that S alone is not an effective branch while NoS is far more competitive. However, when trained jointly (LTOAD) and compared to NoS, S helps R improve the Det. accuracy on " + }, + { + "bbox": [ + 55, + 228, + 294, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\text {tail }}" + }, + { + "bbox": [ + 55, + 228, + 294, + 289 + ], + "type": "text", + "content": " notably. Specifically, S improves the image-level AUROC on " + }, + { + "bbox": [ + 55, + 228, + 294, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\text {tail }}" + }, + { + "bbox": [ + 55, + 228, + 294, + 289 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 228, + 294, + 289 + ], + "type": "inline_equation", + "content": "2.98\\%" + }, + { + "bbox": [ + 55, + 228, + 294, + 289 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "text", + "content": "NoG replaces our prompt-conditioned " + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "inline_equation", + "content": "\\mathsf{G}" + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "text", + "content": " with unconditional feature jittering [34, 57, 89] for data augmentation. With the guidance of " + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{n}}" + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^{\\mathrm{a}}" + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "inline_equation", + "content": "\\mathsf{G}" + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "text", + "content": " creates high-quality pseudo-features especially beneficial for " + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{tail}}" + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "text", + "content": ". We observe a gain of " + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "inline_equation", + "content": "+2.51\\% / +0.75\\%" + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "text", + "content": " in Det. / Seg. on " + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{tail}}" + }, + { + "bbox": [ + 55, + 289, + 294, + 348 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "text", + "content": "Ablation studies on learnable concepts/prompts. Tab. 8 analyzes the effect of each attribute detailed in Tab. 10, where we demonstrate that our prompt learning design is comprehensive. All experiments are performed in the setting MVTec exp100 and B. FixC freezes " + }, + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{C}}" + }, + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "text", + "content": ". FixP freezes " + }, + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^a" + }, + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "text", + "content": ". In SingleP, " + }, + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^a" + }, + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "text", + "content": " stores only one abnormal prompt for each concept. In ShareP, all concepts share the same abnormal prompts. Each attribute contributes significantly to the Det. performance of " + }, + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{tail}}" + }, + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 348, + 295, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 466, + 139, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 139, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 139, + 478 + ], + "type": "text", + "content": "6. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 545 + ], + "type": "text", + "content": "In Tab. 9, we group the literature of unsupervised AD based on their assumptions or target tasks. Notably, LTOAD (1) is a unified model, (2) has online learning capability, (3) uses VLM/LLM priors, (4) handles long-tailed training data distributions, and (5) is class-agnostic." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": "Unsupervised anomaly detection. " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_0" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": " methods train an independent predictor per class. In contrast, UniAD [89] challenges to learn one unified predictor for all classes. More follow-up research falls into this group " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "(\\mathcal{G}_1)" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": ". Inspired by prior work [56, 60, 71], MoEAD [62] uses mixture-of-experts. HVQ [57] uses class-specific quantization modules in their VQ-VAE [82] architecture, but HVQ also induces the constraint of requiring class information as inputs, making it not class-agnostic, i.e., " + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_2" + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "Foundation model based AD. With advancement in foundation models, recent works focused on zero-shot [7, 9, 24, 40, 50, 67, 100, 102] or few-shot [2, 25, 51-53, 101] unsupervised AD where the normal training images are limited (or none), they [7, 9, 25, 40, 51, 52, 67, 100] often" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 148 + ], + "type": "table", + "html": "
Prompt Design Conditions of Unsupervised AD MethodsPrompt Design Categories in AD
G8G9G10G11G12LTOAD
learnable class/concept names✓XXXX✓
support >1 APsXX✓✓✓✓
learnable APsX✓XX✓✓
class/concept specific APsXXX✓✓✓
", + "image_path": "e1a47d6d9eb93515b1561a61644950a0f393086c230064aa810cb72a0b9ab881.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "type": "text", + "content": "Table 10. Characteristic comparison w.r.t. prompt design between LTOAD and prior unsupervised AD works which use LLM(s) with prompts (e.g., " + }, + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_8" + }, + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "type": "text", + "content": ": [34, 67], " + }, + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_9" + }, + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "type": "text", + "content": ": [10], " + }, + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{10}" + }, + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "type": "text", + "content": ": [9, 37, 40, 52, 101], " + }, + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{11}" + }, + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "type": "text", + "content": ": [7], " + }, + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{12}" + }, + { + "bbox": [ + 313, + 152, + 553, + 197 + ], + "type": "text", + "content": ": [51, 102]). AP: abnormal prompts." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 212, + 553, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 553, + 283 + ], + "type": "text", + "content": "turn themselves to the rich knowledge in foundation models [29, 41, 68, 81, 98] or conversational AI [64]. These methods can be grouped into either " + }, + { + "bbox": [ + 313, + 212, + 553, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_3" + }, + { + "bbox": [ + 313, + 212, + 553, + 283 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 212, + 553, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_4" + }, + { + "bbox": [ + 313, + 212, + 553, + 283 + ], + "type": "text", + "content": ", depending on whether they are unified model. To query the language models precisely, " + }, + { + "bbox": [ + 313, + 212, + 553, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_3" + }, + { + "bbox": [ + 313, + 212, + 553, + 283 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 212, + 553, + 283 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_4" + }, + { + "bbox": [ + 313, + 212, + 553, + 283 + ], + "type": "text", + "content": " need the class names or labels as extra inputs and are thus not class-agnostic." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 284, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 554, + 403 + ], + "type": "text", + "content": "Although normal images are limited in zero/few-shot AD, the distribution is even since the number of images per class is the same. Consequently, LTAD [34], i.e., " + }, + { + "bbox": [ + 313, + 284, + 554, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_5" + }, + { + "bbox": [ + 313, + 284, + 554, + 403 + ], + "type": "text", + "content": ", explores the long-tailed setting where the training set is highly unevenly distributed. They rely on the pre-trained ALIGN [41] and learn the pseudo-class names. We highlight the differences between LTOAD prompt learning to related works in Tab. 10. In summary, LTOAD supports (a) learnable concept names, (b) learnable abnormal prompts (AP), (c) multiple APs, and (d) concept-specific APs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "content": "Online learning AD. Online (or continual/incremental) learning [23, 44, 69, 83] investigates how the model learns continuously on a data stream. Recent works [22, 55, 61, 79, 84], i.e. " + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_6" + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_7" + }, + { + "bbox": [ + 313, + 403, + 554, + 475 + ], + "type": "text", + "content": ", explore unsupervised online learning AD. Differently, we propose the task of long-tailed online anomaly detection to study the effect of head/tail classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 487, + 388, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 388, + 499 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 388, + 499 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 506, + 554, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 674 + ], + "type": "text", + "content": "We propose the benchmark of long-tailed online anomaly detection, studying the online performance of head and tail classes. We identify the shortcomings of class-aware methods and propose class-agnostic LTOAD. It learns to approximate the ground truth class set while using foundation models for our novel concept and abnormal prompt learning framework. We also propose " + }, + { + "bbox": [ + 313, + 506, + 554, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{\\mathrm{AA}}" + }, + { + "bbox": [ + 313, + 506, + 554, + 674 + ], + "type": "text", + "content": ", a learning algorithm to adapt the online stream with potentially abnormal inputs. LTOAD outperforms the SOTA in most offline and online settings without additional class information. Although not specifically designed for domain generalization, LTOAD can adapt to unseen domains. We believe that LTOAD will advance the field toward a more realistic anomaly detection setting that is closer to real-world applications." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Acknowledgments. CAY and KCP were supported by Mitsubishi Electric Research Laboratories. CAY and RAY were in part supported by an NSF Award #2420724." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23426" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 91, + 296, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 296, + 144 + ], + "type": "text", + "content": "[1] Akshatha Arodi, Margaux Luck, Jean-Luc Bedwani, Aldo Zaimi, Ge Li, Nicolas Pouliot, Julien Beaudry, and Gaétan Marceau Caron. CableInspect-AD: An expert-annotated anomaly detection dataset. In Proc. NeurIPS, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 65, + 146, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 146, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 65, + 146, + 296, + 190 + ], + "type": "text", + "content": "[2] Yuhu Bai, Jiangning Zhang, Zhaofeng Chen, Yuhang Dong, Yunkang Cao, and Guanzhong Tian. Dual-path frequency discriminators for few-shot anomaly detection. Knowledge-Based Systems, 2024. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 191, + 295, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 191, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 65, + 191, + 295, + 256 + ], + "type": "text", + "content": "[3] Ujjwal Baid, Satyam Ghodasara, Suyash Mohan, Michel Bilello, Evan Calabrese, Errol Colak, Keyvan Farahani, Jayashree Kalpathy-Cramer, Felipe C Kitamura, Sarthak Pati, et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification. arXiv preprint arXiv:2107.02314, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 258, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 258, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 65, + 258, + 295, + 290 + ], + "type": "text", + "content": "[4] Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxiang Zhang, and Xingyu Li. BMAD: Benchmarks for medical anomaly detection. In Proc. CVPR, 2024. 5, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 291, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 291, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 65, + 291, + 295, + 334 + ], + "type": "text", + "content": "[5] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. MVTec AD-a comprehensive real-world dataset for unsupervised anomaly detection. In Proc. CVPR, 2019. 1, 5, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 335, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 335, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 65, + 335, + 295, + 390 + ], + "type": "text", + "content": "[6] Patrick Bilic, Patrick Christ, Hongwei Bran Li, Eugene Vorontsov, Avi Ben-Cohen, Georgios Kaissis, Adi Szeskin, Colin Jacobs, Gabriel Efrain Humpire Mamani, Gabriel Chartrand, et al. The liver tumor segmentation benchmark (lits). Medical Image Analysis, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 391, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 391, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 65, + 391, + 295, + 434 + ], + "type": "text", + "content": "[7] Yunkang Cao, Xiaohao Xu, Chen Sun, Yuqi Cheng, Zongwei Du, Liang Gao, and Weiming Shen. Segment any anomaly without training via hybrid prompt regularization. arXiv preprint arXiv:2305.10724, 2023. 1, 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 435, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 435, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 65, + 435, + 295, + 479 + ], + "type": "text", + "content": "[8] Yunkang Cao, Xiaohao Xu, Jiangning Zhang, Yuqi Cheng, Xiaonan Huang, Guansong Pang, and Weiming Shen. A survey on visual anomaly detection: Challenge, approach, and prospect. arXiv preprint arXiv:2401.16402, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 480, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 480, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 65, + 480, + 295, + 523 + ], + "type": "text", + "content": "[9] Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi. AdaCLIP: Adapting CLIP with hybrid learnable prompts for zero-shot anomaly detection. In Proc. ECCV, 2024. 1, 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 524, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 524, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 61, + 524, + 295, + 579 + ], + "type": "text", + "content": "[10] Pi-Wei Chen, Jerry Chun-Wei Lin, Jia Ji, Feng-Hao Yeh, Zih-Ching Chen, and Chao-Chun Chen. Human-free automated prompting for vision-language anomaly detection: Prompt optimization with meta-guiding prompt scheme. arXiv preprint arXiv:2406.18197, 2024. 1, 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 580, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 580, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 61, + 580, + 295, + 623 + ], + "type": "text", + "content": "[11] Qiyu Chen, Huiyuan Luo, Chengkan Lv, and Zhengtao Zhang. A unified anomaly synthesis strategy with gradient ascent for industrial anomaly detection and localization. In Proc. ECCV, 2024. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 624, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 61, + 624, + 295, + 656 + ], + "type": "text", + "content": "[12] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. arXiv preprint arXiv:2005.02357, 2020. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 658, + 275, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 658, + 275, + 668 + ], + "spans": [ + { + "bbox": [ + 61, + 658, + 275, + 668 + ], + "type": "text", + "content": "[13] Microsoft Corporation. Microsoft Copilot, 2023. 4, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 669, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 669, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 669, + 295, + 712 + ], + "type": "text", + "content": "[14] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. PaDiM: a patch distribution modeling framework for anomaly detection and localization. In Proc. ICPR, 2021. 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one-class embedding. In Proc. CVPR, 2022. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 553, + 139 + ], + "type": "text", + "content": "[16] Keval Doshi and Yasin Yilmaz. Continual learning for anomaly detection in surveillance videos. In Proc. CVPR Workshop, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 140, + 553, + 172 + ], + "type": "text", + "content": "[17] Hugging Face. Transformers: Align model documentation. https://huggingface.co/docs/ transformers/model_doc align.2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 320, + 173, + 553, + 215 + ], + "type": "text", + "content": "[18] Yehonatan Fridman, Matan Rusanovsky, and Gal Oren. ChangeChip: A reference-based unsupervised change detection for PCB defect detection. In Proc. PAINE, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 217, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 217, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 320, + 217, + 553, + 250 + ], + "type": "text", + "content": "[19] Matic Fučka, Vitjan Zavrtanik, and Danijel Skočaj. TransFusion-a transparency-based diffusion model for anomaly detection. In Proc. ECCV, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 320, + 251, + 553, + 293 + ], + "type": "text", + "content": "[20] Hugo Gangloff, Minh-Tan Pham, Luc Courtrai, and Sébastien Lefèvre. Leveraging vector-quantized variational autoencoder inner metrics for anomaly detection. In Proc. ICPR, 2022. 1, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 294, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 294, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 320, + 294, + 553, + 316 + ], + "type": "text", + "content": "[21] Bin-Bin Gao. Learning to detect multi-class anomalies with just one normal image prompt. In Proc. ECCV, 2024. 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 317, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 317, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 317, + 553, + 360 + ], + "type": "text", + "content": "[22] Han Gao, Huiyuan Luo, Fei Shen, and Zhengtao Zhang. Towards total online unsupervised anomaly detection and localization in industrial vision. arXiv preprint arXiv:2305.15652, 2023. 1, 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 361, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 361, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 320, + 361, + 553, + 414 + ], + "type": "text", + "content": "[23] Yasir Ghunaim, Adel Bibi, Kumail Alhamoud, Motasem Alfarra, Hasan Abed Al Kader Hammoud, Ameya Prabhu, Philip HS Torr, and Bernard Ghanem. Real-time evaluation in online continual learning: A new hope. In Proc. CVPR, 2023. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 415, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 415, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 320, + 415, + 553, + 459 + ], + "type": "text", + "content": "[24] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Hao Li, Ming Tang, and Jinqiao Wang. Filo: Zero-shot anomaly detection by fine-grained description and high-quality localization. In Proc. ACM MM, 2024. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 460, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 460, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 320, + 460, + 553, + 503 + ], + "type": "text", + "content": "[25] Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. AnomalyGPT: Detecting industrial anomalies using large vision-language models. In Proc. AAAI, 2024. 2, 6, 8, 5, 7, 9, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 504, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 504, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 320, + 504, + 553, + 547 + ], + "type": "text", + "content": "[26] Guan Gui, Bin-Bin Gao, Jun Liu, Chengjie Wang, and Yunsheng Wu. Few-shot anomaly-driven generation for anomaly classification and segmentation. In Proc. ECCV, 2024. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 548, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 548, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 320, + 548, + 553, + 591 + ], + "type": "text", + "content": "[27] Jia Guo, Haonan Han, Shuai Lu, Weihang Zhang, and Huiqi Li. Absolute-unified multi-class anomaly detection via class-agnostic distribution alignment. arXiv preprint arXiv:2404.00724, 2024.8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 592, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 592, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 320, + 592, + 553, + 613 + ], + "type": "text", + "content": "[28] Ahmed Hamada. Br35H: Brain tumor detection 2020, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 320, + 614, + 553, + 658 + ], + "type": "text", + "content": "[29] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xi-angyu Yue. OneLLM: One framework to align all modalities with language. In Proc. CVPR, 2024. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 553, + 712 + ], + "type": "text", + "content": "[30] Haoyang He, Yuhu Bai, Jiangning Zhang, Qingdong He, Hongxu Chen, Zhenye Gan, Chengjie Wang, Xiangtai Li, Guanzhong Tian, and Lei Xie. MambaAD: Exploring state space models for multi-class unsupervised anomaly detection. In Proc. NeurIPS, 2024. 2, 8" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23427" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[31] Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi-class anomaly detection. In Proc. AAAI, 2024. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 119, + 294, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 119, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 61, + 119, + 294, + 151 + ], + "type": "text", + "content": "[32] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proc. CVPR, 2016. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 153, + 294, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 153, + 294, + 208 + ], + "spans": [ + { + "bbox": [ + 62, + 153, + 294, + 208 + ], + "type": "text", + "content": "[33] Liren He, Zhengkai Jiang, Jinlong Peng, Liang Liu, Qianggang Du, Xiaobin Hu, Wenbing Zhu, Mingmin Chi, Yabiao Wang, and Chengjie Wang. Learning unified reference representation for unsupervised multi-class anomaly detection. In Proc. ECCV, 2024. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 211, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 211, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 62, + 211, + 294, + 253 + ], + "type": "text", + "content": "[34] Chih-Hui Ho, Kuan-Chuan Peng, and Nuno Vasconcelos. Long-tailed anomaly detection with learnable class names. In Proc. CVPR, 2024. 1, 2, 3, 5, 6, 7, 8, 4, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 256, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 256, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 256, + 294, + 289 + ], + "type": "text", + "content": "[35] Junjie Hu, Yuanyuan Chen, and Zhang Yi. Automated segmentation of macular edema in oct using deep neural networks. Medical image analysis, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 294, + 334 + ], + "type": "text", + "content": "[36] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. In Proc. ECCV, 2022. 6, 5, 7, 8, 9, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 294, + 380 + ], + "type": "text", + "content": "[37] Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xinchao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in medical images. In Proc. CVPR, 2024. 1, 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 382, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 382, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 382, + 294, + 425 + ], + "type": "text", + "content": "[38] Jeeho Hyun, Sangyun Kim, Giyoung Jeon, Seung Hwan Kim, Kyunghoon Bae, and Byung Jun Kang. ReConPatch: Contrastive patch representation learning for industrial anomaly detection. In Proc. WACV, 2024. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "type": "text", + "content": "[39] Brian KS Isaac-Medina, Yona Falinie A Gaus, Neelanjan Bhowmik, and Toby P Breckon. Towards open-world object-based anomaly detection via self-supervised outlier synthesis. In Proc. ECCV, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 474, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 474, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 62, + 474, + 294, + 517 + ], + "type": "text", + "content": "[40] Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer. WinCLIP: Zero/few-shot anomaly classification and segmentation. In Proc. CVPR, 2023. 1, 2, 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 520, + 294, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 520, + 294, + 573 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 294, + 573 + ], + "type": "text", + "content": "[41] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In Proc. ICML, 2021. 8, 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 576, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 576, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 576, + 294, + 597 + ], + "type": "text", + "content": "[42] Pranita Balaji Kanade and PP Gumaste. Brain tumor detection using MRI images. *Brain*, 2015. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 600, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 600, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 62, + 600, + 294, + 611 + ], + "type": "text", + "content": "[43] Felipe Campos Kitamura. Head CT - hemorrhage, 2018. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 613, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 294, + 655 + ], + "type": "text", + "content": "[44] Hyunseo Koh, Dahyun Kim, Jung-Woo Ha, and Jonghyun Choi. Online continual learning on class incremental blurry task configuration with anytime inference. In Proc. ICLR, 2022. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "text", + "content": "[45] Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge. In Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, 2015. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "text", + "content": "[46] Joo Chan Lee, Taejune Kim, Eunbyung Park, Simon S Woo, and Jong Hwan Ko. Continuous memory representation for anomaly detection. In Proc. ECCV, 2024. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 108, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 108, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 321, + 108, + 553, + 151 + ], + "type": "text", + "content": "[47] Jan Lehr, Jan Philipps, Alik Sargsyan, Martin Pape, and Jörg Krüger. AD3: Introducing a score for anomaly detection dataset difficulty assessment using VIADUCT dataset. In Proc. ECCV, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 154, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 154, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 321, + 154, + 553, + 186 + ], + "type": "text", + "content": "[48] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. CutPaste: Self-supervised learning for anomaly detection and localization. In Proc. CVPR, 2021. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 189, + 553, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 189, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 321, + 189, + 553, + 231 + ], + "type": "text", + "content": "[49] Shengze Li, Jianjian Cao, Peng Ye, Yuhan Ding, Chongjun Tu, and Tao Chen. ClipSAM: CLIP and SAM collaboration for zero-shot anomaly segmentation. arXiv preprint arXiv:2401.12665, 2024. 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 235, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 235, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 321, + 235, + 553, + 277 + ], + "type": "text", + "content": "[50] Xurui Li, Ziming Huang, Feng Xue, and Yu Zhou. MuSc: Zero-shot industrial anomaly classification and segmentation with mutual scoring of the unlabeled images. In Proc. ICLR, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 280, + 553, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 280, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 321, + 280, + 553, + 323 + ], + "type": "text", + "content": "[51] Xiaofan Li, Zhizhong Zhang, Xin Tan, Chengwei Chen, Yanyun Qu, Yuan Xie, and Lizhuang Ma. PromptAD: Learning prompts with only normal samples for few-shot anomaly detection. In Proc. CVPR, 2024. 1, 2, 5, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 326, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 326, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 321, + 326, + 553, + 358 + ], + "type": "text", + "content": "[52] Yuanwei Li, Elizaveta Ivanova, and Martins Bruveris. FADE: Few-shot/zero-shot anomaly detection engine using large vision-language model. In Proc. BMVC, 2024. 1, 2, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 361, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 361, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 321, + 361, + 553, + 392 + ], + "type": "text", + "content": "[53] Jingyi Liao, Xun Xu, Manh Cuong Nguyen, Adam Goodge, and Chuan Sheng Foo. COFT-AD: Contrastive fine-tuning for few-shot anomaly detection. IEEE TIP, 2024. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 395, + 553, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 395, + 553, + 428 + ], + "spans": [ + { + "bbox": [ + 321, + 395, + 553, + 428 + ], + "type": "text", + "content": "[54] Chieh Liu, Yu-Min Chu, Ting-I Hsieh, Hwann-Tzong Chen, and Tyng-Luh Liu. Learning diffusion models for multi-view anomaly detection. In Proc. ECCV, 2024. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 430, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 430, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 321, + 430, + 553, + 473 + ], + "type": "text", + "content": "[55] Jiaqi Liu, Kai Wu, Qiang Nie, Ying Chen, Bin-Bin Gao, Yong Liu, Jinbao Wang, Chengjie Wang, and Feng Zheng. Unsupervised continual anomaly detection with contrastively-learned prompt. In Proc. AAAI, 2024. 1, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 475, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 475, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 321, + 475, + 553, + 540 + ], + "type": "text", + "content": "[56] Ruiying Lu, Bo Chen, Jianqiao Sun, Wenchao Chen, Penghui Wang, Yuanwei Chen, Hongwei Liu, and Pramod K Varshney. Heterogeneity-aware recurrent neural network for hyperspectral and multispectral image fusion. IEEE Journal of Selected Topics in Signal Processing, 2022. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 543, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 543, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 321, + 543, + 553, + 597 + ], + "type": "text", + "content": "[57] Ruiying Lu, YuJie Wu, Long Tian, Dongsheng Wang, Bo Chen, Xiyang Liu, and Ruimin Hu. Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection. In Proc. NeurIPS, 2023. 1, 2, 4, 5, 6, 8, 7, 9, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 600, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 600, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 321, + 600, + 553, + 632 + ], + "type": "text", + "content": "[58] Jianhong Ma, Tao Zhang, Cong Yang, Yangjie Cao, Lipeng Xie, Hui Tian, and Xuexiang Li. Review of wafer surface defect detection methods. *Electronics*, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 635, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 635, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 321, + 635, + 553, + 677 + ], + "type": "text", + "content": "[59] Sergio Naval Marimont and Giacomo Tarroni. Anomaly detection through latent space restoration using vector quantized variational autoencoders. In Proc. ISBI, 2021. 1, 5, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 680, + 553, + 712 + ], + "type": "text", + "content": "[60] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review, 2014. 8" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23428" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 296, + 115 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 296, + 115 + ], + "type": "text", + "content": "[61] Declan GD McIntosh and Alexandra Branzan Albu. Unsupervised, online and on-the-fly anomaly detection for nonstationary image distributions. In Proc. ECCV, 2024. 1, 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 117, + 295, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 117, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 61, + 117, + 295, + 160 + ], + "type": "text", + "content": "[62] Shiyuan Meng, Wenchao Meng, Qihang Zhou, Shizhong Li, Weiye Hou, and Shibo He. MoEAD: A parameter-efficient model for multi-class anomaly detection. In Proc. ECCV, 2024. 2, 6, 8, 5, 7, 9, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 162, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 162, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 162, + 295, + 205 + ], + "type": "text", + "content": "[63] Shancong Mou, Xiaoyi Gu, Meng Cao, Haoping Bai, Ping Huang, Jiulong Shan, and Jianjun Shi. RGI: Robust GAN-inversion for mask-free image inpainting and unsupervised pixel-wise anomaly detection. In Proc. ICLR, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 202, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 202, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 202, + 216 + ], + "type": "text", + "content": "[64] OpenAI. ChatGPT, 2023. 4, 8, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 294, + 249 + ], + "type": "text", + "content": "[65] Guansong Pang, Chunhua Shen, Longbing Cao, and Anton Van Den Hengel. Deep learning for anomaly detection: A review. ACM computing surveys (CSUR), 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 294, + 282 + ], + "type": "text", + "content": "[66] Sukanya Patra and Souhaib Ben Taieb. Revisiting deep feature reconstruction for logical and structural industrial anomaly detection. TMLR, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 284, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 326 + ], + "type": "text", + "content": "[67] Zhen Qu, Xian Tao, Mukesh Prasad, Fei Shen, Zhengtao Zhang, Xinyi Gong, and Guiguang Ding. VCP-CLIP: A visual context prompting model for zero-shot anomaly segmentation. In Proc. ECCV, 2024. 1, 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 327, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 380 + ], + "type": "text", + "content": "[68] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In Proc. ICML, 2021. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 383, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 294, + 414 + ], + "type": "text", + "content": "[69] Siddeshwar Raghavan, Jiangpeng He, and Fengqing Zhu. DELTA: Decoupling long-tailed online continual learning. In Proc. CVPR, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 415, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 294, + 437 + ], + "type": "text", + "content": "[70] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In Proc. ICML, 2015. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 438, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 438, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 438, + 294, + 480 + ], + "type": "text", + "content": "[71] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. In Proc. NeurIPS, 2021. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 482, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 294, + 524 + ], + "type": "text", + "content": "[72] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In Proc. CVPR, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 526, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 294, + 558 + ], + "type": "text", + "content": "[73] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In Proc. WACV, 2021. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 559, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 591 + ], + "type": "text", + "content": "[74] Radhwan AA Saleh, Mehmet Zeki Konyar, Kaplan Kaplan, and H Metin Ertunç. Tire defect detection model using machine learning. In Proc. eSmarTA, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 593, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 294, + 635 + ], + "type": "text", + "content": "[75] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H Rohban, and Hamid R Rabiee. Multiresolution knowledge distillation for anomaly detection. In Proc. CVPR, 2021. 8, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 636, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 294, + 668 + ], + "type": "text", + "content": "[76] Jian Shi, Pengyi Zhang, Ni Zhang, Hakim Ghazzai, and Peter Wonka. Dissolving is amplifying: Towards fine-grained anomaly detection. In Proc. ECCV, 2024. 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[77] Luc PJ Sträter, Mohammadreza Salehi, Efstratios Gavves, Cees GM Snoek, and Yuki M Asano. GeneralAD: Anomaly detection across domains by attending to distorted features. In Proc. ECCV, 2024. 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 105 + ], + "type": "text", + "content": "[78] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Proc. ICML, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 107, + 553, + 150 + ], + "type": "text", + "content": "[79] Jiaqi Tang, Hao Lu, Xiaogang Xu, Ruizheng Wu, Sixing Hu, Tong Zhang, Tsz Wa Cheng, Ming Ge, Ying-Cong Chen, and Fugee Tsung. An incremental unified framework for small defect inspection. In Proc. ECCV, 2024. 1, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 321, + 152, + 553, + 195 + ], + "type": "text", + "content": "[80] Tran Dinh Tien, Anh Tuan Nguyen, Nguyen Hoang Tran, Ta Duc Huy, Soan Duong, Chanh D Tr Nguyen, and Steven QH Truong. Revisiting reverse distillation for anomaly detection. In Proc. CVPR, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 321, + 198, + 553, + 251 + ], + "type": "text", + "content": "[81] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, et al. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 254, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 254, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 254, + 553, + 274 + ], + "type": "text", + "content": "[82] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. In Proc. NeurIPS, 2017. 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 321, + 276, + 553, + 308 + ], + "type": "text", + "content": "[83] Liuyuan Wang, Xingxing Zhang, Hang Su, and Jun Zhu. A comprehensive survey of continual learning: theory, method and application. IEEE TPAMI, 2024. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 310, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 310, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 321, + 310, + 553, + 353 + ], + "type": "text", + "content": "[84] Shenxing Wei, Xing Wei, Zhiheng Ma, Shaochen Zhang Songlin Dong, and Yihong Gong. Few-shot online anomaly detection and segmentation. Knowledge-Based Systems, 2024. 1, 2, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 355, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 355, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 321, + 355, + 553, + 388 + ], + "type": "text", + "content": "[85] Matthias Wieler and Tobias Hahn. Weakly supervised learning for industrial optical inspection. In DAGM symposium in, 2007. 5, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 389, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 389, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 321, + 389, + 553, + 431 + ], + "type": "text", + "content": "[86] Shuai Yang, Zhifei Chen, Pengguang Chen, Xi Fang, Yixun Liang, Shu Liu, and Yingcong Chen. Defect spectrum: A granular look of large-scale defect datasets with rich semantics. In Proc. ECCV, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 321, + 434, + 553, + 487 + ], + "type": "text", + "content": "[87] Hang Yao, Ming Liu, Haolin Wang, Zhicun Yin, Zifei Yan, Xiaopeng Hong, and Wangmeng Zuo. GLAD: Towards better reconstruction with global and local adaptive diffusion models for unsupervised anomaly detection. In Proc. ECCV, 2024. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 321, + 490, + 553, + 533 + ], + "type": "text", + "content": "[88] Xincheng Yao, Ruoqi Li, Zefeng Qian, Lu Wang, and Chongyang Zhang. Hierarchical gaussian mixture normalizing flow modeling for unified anomaly detection. In Proc. ECCV, 2024." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 321, + 535, + 553, + 578 + ], + "type": "text", + "content": "[89] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. In Proc. NeurIPS, 2022. 5, 6, 8, 7, 9, 10, 11, 12, 13, 14" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 580, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 580, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 321, + 580, + 553, + 623 + ], + "type": "text", + "content": "[90] Jiawei Yu, Ye Zheng, Xiang Wang, Wei Li, Yushuang Wu, Rui Zhao, and Liwei Wu. FastFlow: Unsupervised anomaly detection and localization via 2D normalizing flows. arXiv preprint arXiv:2111.07677, 2021.8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 321, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 321, + 624, + 553, + 667 + ], + "type": "text", + "content": "[91] Vitjan Zavrtanik, Matej Kristan, and Danijel Skoçaj. DRAEM-a discriminatively trained reconstruction embedding for surface anomaly detection. In Proc. ICCV, 2021. 8, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 321, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 670, + 553, + 713 + ], + "type": "text", + "content": "[92] Chongsheng Zhang, George Almpanidis, Gaojuan Fan, Binquan Deng, Yanbo Zhang, Ji Liu, Aouaidjia Kamel, Paolo Soda, and João Gama. A systematic review on long-tailed learning. arXiv preprint arXiv:2408.00483, 2024. 1" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23429" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 553 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 106 + ], + "type": "text", + "content": "[93] Gongjie Zhang, Kaiwen Cui, Tzu-Yi Hung, and Shijian Lu. Defect-GAN: High-fidelity defect synthesis for automated defect inspection. In Proc. WACV, 2021. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 295, + 150 + ], + "type": "text", + "content": "[94] Jiangning Zhang, Xuhai Chen, Zhucun Xue, Yabiao Wang, Chengjie Wang, and Yong Liu. Exploring grounding potential of VQA-oriented GPT-4V for zero-shot anomaly detection. arXiv preprint arXiv:2311.02612, 2023. 2, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 152, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 152, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 152, + 295, + 217 + ], + "type": "text", + "content": "[95] Jiangning Zhang, Xuhai Chen, Yabiao Wang, Chengjie Wang, Yong Liu, Xiangtai Li, Ming-Hsuan Yang, and Dacheng Tao. Exploring plain ViT reconstruction for multiclass unsupervised anomaly detection. In Proc. IJCAI Workshop on Anomaly Detection with Foundation Models, 2024. 5, 8, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 219, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 219, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 219, + 294, + 272 + ], + "type": "text", + "content": "[96] Jiangning Zhang, Haoyang He, Zhenye Gan, Qingdong He, Yuxuan Cai, Zhucun Xue, Yabiao Wang, Chengjie Wang, Lei Xie, and Yong Liu. A Der: A comprehensive benchmark for multi-class visual anomaly detection. arXiv preprint arXiv:2406.03262, 2024. 1, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 274, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 294, + 317 + ], + "type": "text", + "content": "[97] Xuan Zhang, Shiyu Li, Xi Li, Ping Huang, Jiulong Shan, and Ting Chen. DeSTSeg: Segmentation guided denoising student-teacher for anomaly detection. In Proc. CVPR, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 319, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 363 + ], + "type": "text", + "content": "[98] Yiyuan Zhang, Kaixiong Gong, Kaipeng Zhang, Hongsheng Li, Yu Qiao, Wanli Ouyang, and Xiangyu Yue. MetaTransformer: A unified framework for multimodal learning. arXiv preprint arXiv:2307.10802, 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 365, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 294, + 385 + ], + "type": "text", + "content": "[99] Ying Zhao. OmniAL: A unified cnn framework for unsupervised anomaly localization. In Proc. CVPR, 2023. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 387, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 387, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 57, + 387, + 294, + 430 + ], + "type": "text", + "content": "[100] Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jiming Chen. AnomalyCLIP: Object-agnostic prompt learning for zero-shot anomaly detection. In Proc. ICLR, 2024. 1, 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 432, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 432, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 57, + 432, + 294, + 464 + ], + "type": "text", + "content": "[101] Jiawen Zhu and Guansong Pang. Toward generalist anomaly detection via in-context residual learning with few-shot sample prompts. In Proc. CVPR, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 466, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 466, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 57, + 466, + 294, + 508 + ], + "type": "text", + "content": "[102] Jiawen Zhu, Yew-Soon Ong, Chunhua Shen, and Guansong Pang. Fine-grained abnormality prompt learning for zero-shot anomaly detection. arXiv preprint arXiv:2410.10289, 2024. 1, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 510, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 294, + 553 + ], + "type": "text", + "content": "[103] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pre-training for anomaly detection and segmentation. In Proc. ECCV, 2022. 1, 5, 3, 4" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23430" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Toward Material-Agnostic System Identification from Videos/58fcfc32-a696-4a6d-a6cd-3c218994c3ba_content_list.json b/2025/Toward Material-Agnostic System Identification from Videos/58fcfc32-a696-4a6d-a6cd-3c218994c3ba_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..436f567439ec7f28df8e63d294e0b1a23862db5e --- /dev/null +++ b/2025/Toward Material-Agnostic System Identification from Videos/58fcfc32-a696-4a6d-a6cd-3c218994c3ba_content_list.json @@ -0,0 +1,1658 @@ +[ + { + "type": "text", + "text": "Toward Material-Agnostic System Identification from Videos", + "text_level": 1, + "bbox": [ + 197, + 140, + 799, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yizhou Zhao1 Haoyu Chen1 Chunjiang Liu1 Zhenyang Li2 Charles Herrmann3", + "bbox": [ + 156, + 194, + 834, + 212 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junhwa $\\mathsf{Hur}^3$ Yinxiao Li3 Ming-Hsuan Yang3,4 Bhiksha Raj1 Min $\\mathbf{X}\\mathbf{u}^{1*}$", + "bbox": [ + 192, + 212, + 810, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University $^{2}$ University of Alabama at Birmingham $^{3}$ Google $^{4}$ UC Merced", + "bbox": [ + 124, + 229, + 875, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/Skaldak/MASIV", + "bbox": [ + 313, + 247, + 683, + 263 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54aeab1a885dda3cf57aa77b1e3cc10ce1e8177a2e18b1495742ca14509692bc.jpg", + "image_caption": [ + "Figure 1. System identification from video observations seeks to understand the intrinsic dynamics of a given scene. (a) Existing methods rely on scene-specific material priors. (b) Our MASIV removes this requirement, grounding object motion in a material-agnostic manner." + ], + "image_footnote": [], + "bbox": [ + 140, + 290, + 854, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 254, + 628, + 330, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "System identification from videos aims to recover object geometry and governing physical laws. Existing methods integrate differentiable rendering with simulation but rely on predefined material priors, limiting their ability to handle unknown ones. We introduce MASIV, the first vision-based framework for material-agnostic system identification. Unlike existing approaches that depend on hand-crafted constitutive laws, MASIV employs learnable neural constitutive models, inferring object dynamics without assuming a scene-specific material prior. However, the absence of full particle state information imposes unique challenges, leading to unstable optimization and physically implausible behaviors. To address this, we introduce dense geometric guidance by reconstructing continuum particle trajectory", + "bbox": [ + 99, + 656, + 485, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ries, providing temporally rich motion constraints beyond sparse visual cues. Comprehensive experiments show that MASIV achieves state-of-the-art performance in geometric accuracy, rendering quality, and generalization ability.", + "bbox": [ + 511, + 628, + 895, + 689 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 511, + 703, + 640, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "From the flow of water down a stream to the deformation of a rubber ball upon impact, the physical world operates with a set of underlying principles that govern motion and interaction. Humans intuitively grasp these principles through observation, allowing us to see a scene and instinctively predict its outcome or re-imagine it with different initial dynamics. Material identification forms the basis of this ability, requiring us to determine an object's composition and understand how that influences its response to forces. This capability is fundamental to world modeling, the process of constructing internal representations that help us understand", + "bbox": [ + 509, + 727, + 895, + 890 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 116, + 16, + 189, + 55 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 245, + 13, + 799, + 29 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 328, + 30, + 715, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 297, + 44, + 746, + 56 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 119, + 876, + 240, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5944", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and anticipate the evolution of the environment. However, replicating the ability to generalize physical laws to diverse scenarios remains a challenge for artificial intelligence.", + "bbox": [ + 99, + 102, + 480, + 147 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this gap, researchers have explored vision-based system identification [6, 8, 9, 20, 23, 30, 38], which aims to infer governing physical laws from visual observations. Typically, they optimize a parameterized physical model, by end-to-end integrating differentiable renderers, such as Neural Radiance Fields (NeRF) [60] and 3D Gaussian Splatting (3DGS) [34], with differentiable simulators [26, 27, 31, 78, 91]. To enforce physically meaningful constraints on the simulation, existing approaches [6, 38, 100] often rely on material-specific constitutive laws, such as elasticity, plasticity, or viscosity models. These laws define how an object's material properties influence its response to external forces, allowing system identification methods to estimate parameters such as stiffness, damping, or friction. While effective in controlled settings, this dependence on hand-crafted material models limits adaptability. Moreover, these methods assume knowing the material type for a given scene, making it difficult for them to generalize across in-the-wild scenarios where material properties are unknown.", + "bbox": [ + 99, + 148, + 482, + 443 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Directly targeting these limitations, we introduce Material-Agnostic System Identification from Videos (MASIV), a vision-based approach that infers object dynamics free of predefined material priors. As illustrated in Fig. 1, rather than relying on expert-designed elasticity and plasticity models, MASIV employs fully learnable neural constitutive models inspired by NCLaw [59]. However, directly applying neural constitutive models to vision-based system identification poses significant challenges. Similar to the learning process of neural PDE solvers [5, 22, 44, 50], the original formulation of NCLaw assumes access to complete state observations of simulated particles, including position, velocity, deformation gradient, and affine momentum. Since visual observation fails to provide such complete information, simply relying on frame-wise pixel supervision provides insufficient constraints. This leads to unstable optimization and poor convergence [48].", + "bbox": [ + 99, + 446, + 482, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We combine two main insights to enable this optimization. First, despite the lack of ground truth full particle state, frame-wise observations still allow reasonable estimates for some of them, such as position and velocity. Existing dynamic reconstruction methods [42, 55, 56] factorize moving objects with their canonical representation and corresponding deformations, which enables per-frame position tracking. Further, the commonly adopted motion basis [61, 94, 101] provides a low-rank encoding of object motion, facilitating state interpolation between frames. This formulation allows us to reconstruct particle trajectories over dense time steps. Second, although trajectory estimations are inherently limited to visible object regions,", + "bbox": [ + 99, + 699, + 482, + 890 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "internal deformations can be inferred by propagating surface motions inward. Our intuition is that nearby particles inside an object should be deformed similarly, akin to the As-Rigid-As-Possible (ARAP) prior.", + "bbox": [ + 511, + 102, + 893, + 161 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We use these insights to guide the design of our framework. Concretely, we reconstruct dynamic Gaussians from multi-view video inputs, simultaneously establishing geometry and deformation models of the scene. The geometry model is then transformed into a solid continuum, where the deformation model is fine-tuned to encode both the exterior and the interior motion. To identify intrinsic dynamics, we guide the learning of the neural constitutive model using two complementary cues: appearance cues from video frames and geometric constraints from the fine-tuned deformation model. In this way, MASIV distills dense particle motion from sparse visual observations and guides the learning of neural constitutive models for stable, material-agnostic system identification.", + "bbox": [ + 509, + 162, + 893, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summed up as follows:", + "bbox": [ + 529, + 369, + 843, + 383 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a novel framework, MASIV, that determines the intrinsic dynamics of an object from videos in a material-agnostic manner. In contrast, all other current vision-based system identification methods [6, 8, 38] necessitate predefined material priors.", + "- We introduce temporally dense geometry guidance to constrain object behaviors with neural constitutive models, facilitating smoother convergence and improved performance for material-agnostic system identification.", + "- Through extensive qualitative and quantitative experiments, we demonstrate that MASIV achieves superior performance in both observable state recovery and future state prediction, even without explicit material priors." + ], + "bbox": [ + 511, + 386, + 893, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 595, + 650, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic reconstruction is a fundamental task in computer vision, aimed at reconstructing high-fidelity representations of complex dynamic scenes from diverse inputs such as monocular and multi-view videos [1, 7, 19, 41, 53, 76, 83, 88, 92, 95, 96]. Earlier research primarily focused on augmenting canonical Neural Radiance Fields (NeRF) [60] with a deformation field [64], which were later enhanced by incorporating volume-preserving regularization techniques [62, 63]. While primarily focused on the novel view synthesis task, their neural implicit representations often introduce noisy deformations, limiting the accuracy of recovered geometries required for reliable physics-based property estimation [38]. With the explicit representation of the scene using Gaussian ellipsoids, 3D Gaussian Splatting (3DGS) [34] achieves efficient dynamic scene reconstruction. Building on the principles of 3DGS, several studies have extended its application to 4D dynamic reconstruction by handling each frame separately [56] or mod", + "bbox": [ + 509, + 625, + 893, + 890 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5945", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "eling the entire scene with a canonical 3D Gaussian point cloud coupled with a robust deformation framework that accurately maps it to the target scene [36, 86, 93].", + "bbox": [ + 99, + 102, + 483, + 147 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic simulation integrates physics principles into 3D representations to generate realistic motion and interactions. A typical approach to synthesizing dynamic 3D scenes involves integrating a 3D generation pipeline with a video generation model [3, 49, 69, 75]. Recent research incorporates 3D Gaussian kernels as both visual and physical representations, embedding Newtonian dynamics into the Gaussian framework to enable seamless rendering and simulation while introducing physical constraints to ensure plausible dynamics [4, 16, 38, 48, 51, 67, 89, 100]. Building on this, Gaussian Splashing [14] introduces a position-based dynamics framework within 3D Gaussian Splating, enabling interactions among solids, fluids, and deformable objects. Additionally, motion-conditioned simulation methods [18, 43, 74, 85, 87] use trajectory-based guidance to predict object motion, showcasing how generative models can enhance dynamic scene synthesis. However, these methods primarily rely on pre-trained generative models, which often lack a fundamental understanding of physics. More recent works leverage neural physics modeling, incorporating video priors and generative models to infer object dynamics, allowing physics-driven motion synthesis without explicit solvers [15, 28, 54, 81, 98]. However, these methods are constrained to specific material types, limiting their generalization across diverse objects. To address this, we introduce a material-agnostic approach that unifies system identification across different object types, ensuring broader applicability beyond predefined materials.", + "bbox": [ + 99, + 151, + 483, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "System identification aims to understand the physical laws governing the 3D world and is an essential task for simulation [37, 40, 45, 68, 79, 100] and robotic manipulation [46, 66, 72, 73, 99]. However, accurately recovering both object geometries and physical properties remains a significant challenge. Existing methods often assume known geometries and rely on finite element methods (FEM) or mass-spring systems for dynamic simulation [80, 82], limiting their ability to handle complex, real-world materials. Neural network-based methods [39, 70, 90] offer greater flexibility by leveraging data-driven models for system identification, but they struggle to generalize across diverse material properties and environmental conditions. More recently, differentiable physics-based simulations have emerged as a promising direction [11, 12, 17, 21, 27, 29, 30, 32, 58, 65]. However, these methods often require intricate modeling to bridge the gap between simulated and real-world behavior. To improve adaptability, some approaches focus on neural constitutive modeling, embedding learned material properties within physical simulations to refine expert-designed models [8, 59]. Other methods leverage the Material Point Method (MPM)", + "bbox": [ + 99, + 566, + 483, + 890 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Algorithm 1 MPM state transfer $\\mathcal{M}$", + "text_level": 1, + "bbox": [ + 514, + 102, + 750, + 116 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input: $\\mathbf{s}_n = \\{\\mathbf{x}_n,\\mathbf{v}_n,\\mathbf{F}_n^e\\}$", + "bbox": [ + 514, + 119, + 691, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Output: $\\mathbf{s}_{n + 1} = \\{\\mathbf{x}_{n + 1},\\mathbf{v}_{n + 1},\\mathbf{F}_{n + 1}^{e}\\}$", + "bbox": [ + 514, + 137, + 766, + 151 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for $i = 1 \\dots Q$ , $\\mathbf{P}_n(i) = \\mathcal{E}(\\mathbf{F}_n^e (i))$", + "2: $\\mathbf{x}_{n + 1},\\mathbf{v}_{n + 1},\\mathbf{F}_{n + 1}^{e,\\mathrm{trial}} = \\mathcal{I}(\\mathbf{x}_n,\\mathbf{v}_n,\\mathbf{F}_n^e,\\mathbf{P}_n)$", + "3: for $i = 1 \\dots Q$ , $\\mathbf{F}_{n + 1}^{e}(i) = \\mathcal{P}(\\mathbf{F}_{n + 1}^{e,\\mathrm{trial}}(i))$" + ], + "bbox": [ + 524, + 151, + 807, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and 3D Gaussian-based representations to enhance geometry reconstruction and material property estimation [6, 38, 71, 100]. Building upon GIC [6], our method presents a material-agnostic approach that eliminates reliance on predefined material priors. Rather than assuming material properties, our model infers object dynamics directly from visual observations, making it adapt better to a wide range of unknown or heterogeneous materials.", + "bbox": [ + 509, + 228, + 895, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 358, + 601, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given visual observations $\\mathbf{I} = \\{\\mathbf{I}_0, \\mathbf{I}_1, \\dots, \\mathbf{I}_T\\}$ over $T$ frames capturing a dynamic object, our goal is to recover its explicit geometric representation and the governing evolution rules of the underlying dynamical system. Formally, we represent the object's geometry using a set of Gaussian kernels [34] and integrate neural constitutive laws [59] with the material point method (MPM) [26, 31, 78] to roll out their motion over time. Unlike [6, 8, 38, 100], we do not assume a known material type of the object, while exploring Material-Agnostic System Identification from Videos (MASIV). Fig. 2 shows an overview of our pipeline.", + "bbox": [ + 509, + 381, + 895, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 551, + 650, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Material Point Method (MPM) discretizes the object of interest into $Q$ material points and models their evolution over time via a state transfer equation", + "bbox": [ + 511, + 575, + 893, + 621 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {n + 1} = \\mathcal {M} (\\mathbf {s} _ {n}), \\forall n = 0, 1, \\dots , N. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 630, + 893, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{s}_n$ denotes the particle state at time step $n$ . $N$ is the total number of simulation steps with the simulation interval $\\tau$ and observed $T$ frames, where $N = T / \\tau \\gg T$ .", + "bbox": [ + 509, + 654, + 893, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The transition process $\\mathcal{M}$ follows a time-stepping scheme, as outlined in Algorithm 1. The state of particle $i$ at time step $n$ is described by its position $\\mathbf{x}_n(i)$ , velocity $\\mathbf{v}_n(i)$ , and the elastic component $\\mathbf{F}_n^e (i)$ of its deformation gradient $\\mathbf{F}_n(i)$ . At each time step, the elastic constitutive law $\\mathcal{E}$ computes the first Piola-Kirchhoff stress $\\mathbf{P}_n(i)$ from the elastic deformation gradient $\\mathbf{F}_n^e (i)$ . Subsequently, the time integration scheme $\\mathcal{I}$ updates all particle states. The updated trial elastic deformation gradient, denoted as $\\mathbf{F}_{n + 1}^{e,\\mathrm{trial}}(i)$ , needs to be further corrected by the plastic constitutive law $\\mathcal{P}$ , which applies plasticity constraints and yields the final elastic deformation gradient $\\mathbf{F}_{n + 1}^{e}(i)$ . More details of MPM can be found in Sec. A.2 of Supp. Mat.", + "bbox": [ + 509, + 699, + 893, + 890 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5946", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/01d83b78a20996eb69288d40cd33267e5f5a99259e915331ce3d41a5b5aaba90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 99, + 893, + 223 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0578d2a5e614b23dae067b60d80ca85803c40286b31c8cc2d057d7b89abe5c2f.jpg", + "image_caption": [ + "Figure 2. The overview of MASIV. Our pipeline comprises three phases. Phase I, Geometrical Reconstruction, infers geometrical representations (Sec. 3.2) and dense trajectories (Sec. 3.3) for the captured object. Phase II, Material-Agnostic System Identification (Sec. 3.4), grounds dynamic characteristics by exploiting visual observations with reconstructed motion clues. From this, we obtain a generalizable digital twin in Phase III, which can simulate novel interactions, such as new velocities or forces." + ], + "image_footnote": [], + "bbox": [ + 102, + 224, + 893, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing efforts [6, 8, 38, 100] for video-based system identification commonly assume known constitutive laws $\\mathcal{E}$ and $\\mathcal{P}$ , with a particular backbone form, e.g., neo-Hookean elasticity with identity plasticity return function. This simplifies the problem by estimating a small number of physical properties, such as Young's modulus, fluid viscosity, friction angles, etc., at the cost of sacrificing flexibility. In contrast, we explore the same problem without this material-specific assumption.", + "bbox": [ + 99, + 439, + 483, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dynamic Gaussian Reconstruction", + "text_level": 1, + "bbox": [ + 101, + 580, + 395, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We start by reconstructing dynamic Gaussians over $T$ frames. We consider this reconstruction as temporally sparse because the number of simulation steps is hundreds of times more than the number of observed video frames $T$ , i.e., $N = T / \\tau \\gg T$ . Following [6, 36, 86, 93], we maintain a set of canonical Gaussians $\\mathcal{G}^* = \\{\\pmb{\\mu}^*, \\pmb{\\Sigma}^*, \\mathbf{c}^*, \\sigma^*\\}$ and use a neural network $\\mathcal{D}$ to warp them over time to represent observed frames. Specifically, a basis network maps a given time step $t$ to $B$ bases, producing deformation basis for position $\\mathbf{B}^{\\mu}(t) \\in \\mathbb{R}^{B \\times 3}$ and scale $\\mathbf{B}^s(t) \\in \\mathbb{R}^{B \\times 1}$ . A coefficient network then estimates the weight of each basis, $\\mathbf{w}(\\pmb{\\mu}^*, t) \\in \\mathbb{R}^B$ , from the center coordinates of canonical Gaussian kernels $\\pmb{\\mu}^*$ and the time step $t$ . This results in deformed Gaussian kernels at each time step $t$ :", + "bbox": [ + 99, + 601, + 483, + 806 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {t} = \\boldsymbol {\\mu} ^ {*} + \\mathbf {w} (\\boldsymbol {\\mu} ^ {*}, t) \\mathbf {B} ^ {\\boldsymbol {\\mu}} (t), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 816, + 482, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {t} = s ^ {*} + \\mathbf {w} \\left(\\boldsymbol {\\mu} ^ {*}, t\\right) \\mathbf {B} ^ {s} (t). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 835, + 482, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ the same optimization strategy in [6, 93] to minimize the L1 and Structural Similarity Index Mea", + "bbox": [ + 99, + 859, + 483, + 888 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sure (SSIM) losses between rendered frames $\\hat{\\mathbf{I}}_t$ and corresponding ground-truth observations $\\mathbf{I}_t$ . We also regard all Gaussian kernels as isotropic and apply L1 regularization on scales for simplicity and reconstruction quality [10, 97, 100]. The overall objective function is formulated as", + "bbox": [ + 509, + 439, + 893, + 526 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {G} ^ {*}, \\mathcal {D}} \\left[ \\mathcal {L} _ {1} (\\hat {\\mathbf {I}} _ {t}, \\mathbf {I} _ {t}) + \\lambda_ {\\mathrm {S S I M}} \\mathcal {L} _ {\\mathrm {S S I M}} (\\hat {\\mathbf {I}} _ {t}, \\mathbf {I} _ {t}) + \\lambda_ {s} \\| s _ {t} \\| _ {1} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 532, + 893, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with hyperparameters $\\lambda_{\\mathrm{SSIM}}$ and $\\lambda_{s}$", + "bbox": [ + 511, + 566, + 746, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Continuum Trajectory Estimation", + "text_level": 1, + "bbox": [ + 511, + 588, + 802, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To bridge the gap between unevenly distributed Gaussian particles and the corresponding solid continuum, we follow [6] to fill internal volumes, forming continuum particles $\\{\\mathbf{p}_t\\}$ with uniform scale $s$ and density $\\sigma$ . Existing methods for material-specific system identification often rely on sparse visual [6, 8, 38] or geometrical [6] cues along $T$ frames to optimize a small set of parameters. However, since we do not assume a known constitutive backbone or material types, these sparse signals may lead to unconstrained behaviors at unobserved time steps or even cause training instability. Inspired by system identification [22, 33, 59] using state data per simulation step, we propose leveraging particle trajectories over $N$ simulation steps as dense geometrical supervision to better constrain the simulated behaviors of continuum particles.", + "bbox": [ + 509, + 609, + 893, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given that the deformation network trained in Sec. 3.2 can produce rough trajectories for dynamic Gaussians, we fine-tune it to establish trajectories for continuum particles. Specifically, we initialize the canonical space of particles", + "bbox": [ + 509, + 830, + 893, + 888 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5947", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{x}_0$ with the filled continuum $\\mathbf{p}_1$ at $t = 1$ and apply deformations similar to Eq. (2),", + "bbox": [ + 99, + 102, + 485, + 132 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} = \\mathbf {x} ^ {*} + \\mathbf {w} \\left(\\mathbf {x} ^ {*}, t\\right) \\mathbf {B} ^ {\\mathbf {x}} (t), \\forall t = 0, 1, \\dots , T, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 141, + 482, + 157 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "resulting in per-frame deformed particles $\\{\\mathbf{x}_t\\}$ . Then we optimize this deformation with all filled continuums $\\{\\mathbf{p}_t\\}$ using the Chamfer Distance [13, 57]", + "bbox": [ + 99, + 167, + 482, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {x} ^ {*}, \\mathcal {D} ^ {\\prime}} \\left[ \\mathcal {L} _ {\\mathrm {C D}} \\left(\\left\\{\\mathbf {x} _ {t} \\right\\} _ {t = 0} ^ {T}, \\left\\{\\mathbf {p} _ {t} \\right\\} _ {t = 0} ^ {T}\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 220, + 482, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{D}'$ denotes the fine-tuned deformation network. Unlike the deformation network $\\mathcal{D}$ trained in Sec. 3.2, this fine-tuned version $\\mathcal{D}'$ is aware of the internal particles absence in Gaussians. Further, the learned motion basis with temporal positional embeddings enables it for temporal interpolation, allowing us to probe the motion between observed discrete time steps $t = 0,1,\\dots,T$ . Formally, we have particle positions at each simulation time step", + "bbox": [ + 99, + 253, + 483, + 371 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} = \\mathbf {x} ^ {*} + \\mathbf {w} (\\mathbf {x} ^ {*}, t) \\mathbf {B} ^ {\\mathbf {x}} (t), \\forall t = 0, \\tau , \\dots , N \\tau , \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 381, + 482, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is the time interval between two simulation steps. We then consider these temporally dense particle trajectories as pseudo ground truth for later optimization.", + "bbox": [ + 99, + 407, + 483, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Material-Agnostic System Identification", + "text_level": 1, + "bbox": [ + 101, + 459, + 434, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by NCLaw [59], we parameterize the elastic and plastic constitutive laws using neural networks, denoted as $\\mathcal{E}_{\\theta^e}$ and $\\mathcal{P}_{\\theta^p}$ , where $\\theta^e$ and $\\theta^p$ are their respective network parameters. Both networks share the same multi-layer perceptron (MLP) architecture with a single hidden layer. Further, they incorporate two physical priors, frame indifference and undeformed state equilibrium, enforced by using rotation-invariant input representations and eliminating bias terms, respectively. While we adopt the same parameterization for material-agnostic modeling, we introduce a more challenging setting by relaxing the input requirements from dense, complete state sequences $\\{\\mathbf{s}_t\\}_{n=0}^N$ to sparse visual observations $\\{\\mathbf{I}_t\\}_{t=0}^T$ . To this end, we follow an analysis-by-synthesis scheme. For the initial particle state $\\mathbf{s}_0$ , we use positions $\\mathbf{x}_0$ derived from Eq. (7), an estimizable velocity $\\mathbf{v}_0$ , and identity deformation gradients $\\mathbf{F}_0$ . After we evolve particle positions $\\{\\hat{\\mathbf{x}}_t\\}_{t=\\tau}^{N\\tau}$ through time integration as described in Eq. (1), we optimize $\\theta^e$ and $\\theta^p$ with", + "bbox": [ + 99, + 479, + 483, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta^ {e}, \\theta^ {p}} \\left(\\mathcal {L} _ {\\mathrm {g e o}} + \\mathcal {L} _ {\\mathrm {s i l}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 755, + 482, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{geo}}$ supervises particle positions $\\hat{\\mathbf{x}}_t$ with estimated trajectories $\\mathbf{x}_t$ and $\\mathcal{L}_{\\mathrm{sil}}$ compares rendered masks $\\hat{\\mathbf{M}}_t$ with object silhouettes $\\mathbf{M}_t$ extracted from frame $\\mathbf{I}_t$", + "bbox": [ + 99, + 786, + 483, + 833 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g e o}} = \\mathcal {L} _ {\\mathrm {t r a j}} = \\mathcal {L} _ {1} \\left(\\left\\{\\hat {\\mathbf {x}} _ {t} \\right\\} _ {t = 0} ^ {N \\tau}, \\left\\{\\mathbf {x} _ {t} \\right\\} _ {t = 0} ^ {N \\tau}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 840, + 482, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s i l}} = \\mathcal {L} _ {1} \\left(\\left\\{\\hat {\\mathbf {M}} _ {t} \\right\\} _ {t = 0} ^ {N \\tau}, \\left\\{\\mathbf {M} _ {t} \\right\\} _ {t = 0} ^ {N \\tau}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 863, + 482, + 886 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While not assuming known material types in our problem setting, we note that MASIV still requires a pre-trained constitutive model from [59] as a stable initialization.", + "bbox": [ + 511, + 102, + 893, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 165, + 640, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 191, + 714, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct our experiments on the PAC-NeRF dataset [38] and the Spring-Gaus dataset [100]. The PAC-NeRF dataset consists of synthetic objects simulated using the Material Point Method (MPM) and includes a variety of material types such as elastic bodies, plasticine, sand, and Newtonian and non-Newtonian fluids. This dataset provides 45 multi-view RGB video sequences that allow for the estimation of both object geometry and physical properties and approximately 14 frames per viewpoint. The Spring-Gaus dataset contains two subsets, namely, Synthetic and Real. The Synthetic subset focuses on the reconstruction and simulation of elastic objects, containing 30 frames in each of 10 viewpoints. The Real subset provides both dense-view static and sparse-view dynamic captures of elastic objects. We evaluate the performance of observable state simulation using both datasets, while the Spring-Gaus dataset is additionally utilized for future state prediction. To obtain object masks, we apply off-the-shelf matting with [35, 47].", + "bbox": [ + 511, + 217, + 893, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. To assess the performance of our approach, we compare it against four state-of-the-art methods for system identification. PAC-NeRF [38] combines neural radiance fields with differentiable physics to jointly infer geometry and material properties. Spring-Gaus [100] integrates a spring-mass system into Gaussian representations for efficient elastic object simulation. NeuMA [8] fine-tunes a neural adapter on expert-designed physical models to better align reconstructed dynamics with visual observations. GIC [6] leverages a two-stage pipeline to obtain the geometry representation and physical properties sequentially. These baselines represent the most relevant prior work and provide a strong comparison for evaluating our method. We note that they all necessitate material priors. To be concrete, PAC-NeRF and GIC roughly categorize materials into five groups and adopt a certain set of physical models for each group. Spring-Gaus generally focuses on elastic objects, thereby being material-specific. And NeuMA leverages a scene-specific constitutive model as the material prior.", + "bbox": [ + 511, + 487, + 893, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We evaluate our method using several standard metrics to assess both geometric reconstruction quality and physical property estimation accuracy. Metrics including 1) Chamfer Distance (CD) [13, 57] with units expressed in $10^{3}\\mathrm{mm}^{2}$ ; 2) Peak Signal-to-Noise Ratio (PSNR) [24]; 3) Structural Similarity Index Measure (SSIM) [84]. The CD quantifies the similarity between two point distributions, while PSNR and SSIM assess the visual quality.", + "bbox": [ + 511, + 771, + 893, + 888 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5948", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/02b29894ae1f136d9bac105ae91d7c34e71d8b629004350d1ef42a3d9c400834.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPAC-NeRFSpring-Gaus
NewtonianNon-NewtonianElasticityPlasticineSandMeanTorusCrossCreamApplePasteChessBananaMean
\\(\\rightarrow\\)DPAC-NeRF [38]0.2770.2360.2380.4290.2120.2784.921.100.771.113.140.962.772.11
Spring-Gaus [100]------0.170.480.360.380.191.802.600.85
NeuMA [8]------4.590.060.080.040.710.050.030.79
GIC [6]0.2430.1950.1780.1960.2500.2120.130.130.140.150.170.410.030.17
MASIV (Ours)0.2330.1980.1920.2010.2290.2100.080.100.190.160.130.190.080.13
", + "bbox": [ + 102, + 99, + 893, + 195 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4f3ac5522d5d199ba855db3dc67322054f0bca2332dc7ec93450823ea4bb5862.jpg", + "table_caption": [ + "Table 1. Observable state simulation on PAC-NeRF and Spring-Gaus Synthetic datasets. Despite being the only technique that does not use expert-designed physical models (noted by the gray row), MASIV performs better than or on par with existing baselines." + ], + "table_footnote": [], + "table_body": "
MethodBunBurgerDogPigPotatoMean
PSNR ↑Spring-Gaus [100]30.6934.0132.1034.9732.7232.90
NeuMA [8]31.2723.7825.6125.40-26.51
GIC [6]34.6840.4537.1738.3239.9538.11
Ours40.2642.1238.9740.8443.4041.12
SSIM ↑Spring-Gaus0.9920.9940.9940.9960.9920.994
NeuMA0.9940.9930.9950.995-0.994
GIC0.9950.9970.9970.9970.9960.996
Ours0.9980.9970.9970.9980.9980.997
", + "bbox": [ + 106, + 244, + 482, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. In dynamic reconstruction, we follow [6] to optimize canonical Gaussians $\\mathcal{G}^*$ and the deformation network $\\mathcal{D}$ for 40K iterations. Then the deformation network is fine-tuned to estimate particle trajectories from filled continuum sequences for 10K iterations. In system identification, we set the simulation time interval $\\tau$ to $1/200$ the frame duration. We adopt the RAdam [52] optimizer to optimize the initial velocity $\\mathbf{v}_0$ for 100 steps and constitutive model parameters $\\theta^e, \\theta^p$ for 1000 steps. All experiments are conducted on a single NVIDIA A100 GPU.", + "bbox": [ + 99, + 426, + 482, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison Results", + "text_level": 1, + "bbox": [ + 101, + 582, + 287, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Observable state simulation. In Tab. 1, we compare MASIV with existing vision-based system identification methods for observable state simulation. The results indicate that MASIV outperforms all other methods in terms of CD, highlighting its effectiveness in reconstructing object dynamics across diverse materials. Further, Tab. 2 evaluates on real data. Thanks to the more flexible constitutive representation, MASIV achieves state-of-the-art results. This demonstrates the potential of MASIV as an adaptive and generalizable approach for system identification.", + "bbox": [ + 99, + 606, + 482, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Future state simulation. In Tab. 3, we provide a detailed evaluation of future state simulation performance. The results indicate that MASIV consistently outperforms all other methods while slightly underperforming GIC. Since NeuMA [8] does not originally experiment on the synthetic subset of Spring-Gaus, we report our reproduction results. To be concrete, we use the closest pre-trained model, Jelly, from NCLaw [59] as the material prior and conduct LoRA [25] fine-tuning with hyperparameters from its orig-", + "bbox": [ + 99, + 757, + 482, + 888 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2642955ee551d28255fa6522410aaa9a0e3bbb68e1dd68ade0807f7e6315f173.jpg", + "table_caption": [ + "Table 2. Observable state simulation on Spring-Gaus Real dataset. MASIV achieves state-of-the-art performance." + ], + "table_footnote": [], + "table_body": "
MethodTorusCrossCreamApplePasteChessBananaMean
CD\\(\\rightarrow\\)PAC-NeRF [38]2.473.872.214.6937.708.2066.4317.94
Spring-Gaus [100]2.381.572.221.877.032.5918.485.16
NeuMA [8]1.170.401147.40.0544.260.330.30170.56
GIC [6]0.751.090.940.222.790.770.120.95
MASIV (Ours)0.160.381.512.641.751.700.241.20
PSNR\\(\\dagger\\)PAC-NeRF [38]17.4614.1515.3719.9412.3215.0816.0415.77
Spring-Gaus [100]16.8316.9315.4221.5514.7116.0817.8917.06
NeuMA [8]16.6119.672.3226.2513.4220.8523.1617.47
GIC [6]20.2430.5119.1526.8916.3118.4429.2922.98
MASIV (Ours)22.4233.7316.4521.1915.7317.2826.5621.91
SSIM\\(\\dagger\\)PAC-NeRF [38]0.9190.9060.8580.8780.8190.8480.8860.870
Spring-Gaus [100]0.9190.9400.8620.9020.8720.8810.9040.897
NeuMA [8]0.9420.9480.8890.9640.8890.9330.9640.933
GIC [6]0.9420.9390.9090.9480.8940.9120.9640.930
MASIV (Ours)0.9550.9670.8520.8980.8910.8780.9650.915
", + "bbox": [ + 516, + 244, + 893, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Future state simulation on Spring-Gaus Synthetic dataset. MASIV outperforms all baselines except for GIC.", + "bbox": [ + 511, + 429, + 893, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "inal implementation. However, we observe unstable optimization processes for most categories. This is probably rooted in its sparse pixel supervision, which only operates once per frame. With dense geometrical constraints supervising hundreds of times per frame, MASIV achieves significantly better results on both geometrical accuracy and rendering quality. This suggests that MASIV effectively captures the underlying dynamics governing future states, allowing it to generate predictions that remain structurally and visually consistent over time. Unlike baseline approaches that rely on predefined material priors, MASIV demonstrates the ability to infer relevant properties directly from visual observations, making it more adaptable to a wide range of scenarios. On the other hand, while dense trajectory guidance enables overfitting to observations, the lack of category-level regularization, unlike hand-crafted constitutive models, limits its generalization under scarce data. This leads to inferior performance than GIC, which can benefit from its pre-known constitutive model as an oracle. More analysis on the enhanced generalization ability of MASIV with more data is in Sec. C of Supp. Mat.", + "bbox": [ + 509, + 474, + 893, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 792, + 661, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We investigate the impact of geometrical objectives here, while referring the readers to more ablation studies in Sec. B of Supp. Mat. Fig. 3 validates the effectiveness of dense trajectory supervision. As shown in Fig. 4, we evaluate three geometrical objectives by replacing the $\\mathcal{L}_{\\mathrm{geo}}$ term in opti", + "bbox": [ + 509, + 815, + 893, + 888 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5949", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a8d08b06a43e59ee3dad00c5bdeb89db87af2e923380dcfed7a22b675df5e87a.jpg", + "image_caption": [ + "Figure 3. Ablation study of geometrical objectives $\\mathcal{L}_{\\mathrm{geo}}$ on PAC-NeRF dataset. We compare three types of geometrical objectives against a baseline without geometrical guidance and plot their simulation errors using Chamfer Distance (CD). Overall, dense trajectory guidance achieves the lowest and most stable errors. The figure is best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 125, + 104, + 883, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7567d0477707b026ddf94e294ef0df503a49d1bbaf64417e9ecf864da56cae28.jpg", + "image_caption": [ + "(a) Surfaces", + "(b) Continuums" + ], + "image_footnote": [], + "bbox": [ + 101, + 303, + 483, + 358 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f45a87c5acba9c8fcf011c8e9bebe3196d7968da1d2c1f1bb1aaa165547e13f.jpg", + "image_caption": [ + "(c) Gaussian Splats", + "(d) Trajectories" + ], + "image_footnote": [], + "bbox": [ + 101, + 383, + 485, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mizing the neural constitutive model objective, Eq. (8). We define them formally as follows. (1) Surface: per-frame surface alignment between the simulated continuum and the filled continuum,", + "bbox": [ + 99, + 542, + 483, + 602 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s u r f}} = \\mathcal {L} _ {\\mathrm {C D}} \\left(\\left\\{\\mathcal {S} (\\hat {\\mathbf {x}} _ {t}) \\right\\} _ {t = 0} ^ {T}, \\left\\{\\mathcal {S} (\\mathbf {p} _ {t}) \\right\\} _ {t = 0} ^ {T}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 609, + 482, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\mathcal{S}$ denotes the surface extraction by thresholding Gaussian opacities [6]. (2) Continuum: similar to Surface while between solid continuums instead of surfaces,", + "bbox": [ + 99, + 642, + 483, + 686 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t}} = \\mathcal {L} _ {\\mathrm {C D}} \\left(\\left\\{\\hat {\\mathbf {x}} _ {t} \\right\\} _ {t = 0} ^ {T}, \\left\\{\\mathbf {p} _ {t} \\right\\} _ {t = 0} ^ {T}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 695, + 482, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(3) Trajectory: per-simulation-step trajectory alignment between the simulated continuum $\\{\\hat{\\mathbf{x}}_t\\}_{t=0}^{N\\tau}$ and the continuum predicted by the deformation network $\\{\\mathbf{x}_t\\}_{t=0}^{N\\tau}$ , as defined in Eq. (9). By analyzing per-frame errors, we first confirm the necessity of geometrical guidance, as the baseline (red curve) exhibits significantly higher CD errors across all material types. Surface-based supervision (yellow curve) shows moderate error reduction but remains less stable over time, especially in Plasticine and Sand. Continuum-based supervision (blue curve) further reduces errors but struggles with certain materials, e.g., Sand. These error overshoots", + "bbox": [ + 99, + 727, + 483, + 888 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df93c4895806c327cddbdc010c34deef9c591e70875751216b002522252b80f6.jpg", + "image_caption": [ + "Figure 4. Visualization of geometrical objectives $\\mathcal{L}_{\\mathrm{geo}}$ on PAC-NeRF Bird. The fine-tuned deformation network $\\mathcal{D}'$ captures smooth and accurate motion for continuums, including low-frequency translation of the body and high-frequency swing of the tail. We space adjacent frames in the horizontal axis for clarity.", + "Ground Truth Surface $\\mathcal{L}_{\\mathrm{surf}}$ Continuum $\\mathcal{L}_{\\mathrm{cont}}$ Trajectory $\\mathcal{L}_{\\mathrm{traj}}$", + "Figure 5. Qualitative comparison of geometrical objectives $\\mathcal{L}_{\\mathrm{geo}}$ on PAC-NeRF Cream. Using sparse surfaces or solid continuum as geometrical cues can lead to physically implausible deformations, while trajectory guidance mitigates these issues by introducing temporally denser constraints. Better viewed when zoomed in." + ], + "image_footnote": [], + "bbox": [ + 511, + 321, + 875, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are probably due to supervising relatively complex neural constitutive models with sparse signals, which might lead to overfitting since the simulated behaviors between observed discrete time steps are not well constrained. Finally, by enforcing dense alignment across simulation steps, trajectory-based supervision (green curve) achieves the lowest and most stable errors across all materials.", + "bbox": [ + 509, + 671, + 895, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 511, + 787, + 700, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dynamics reconstruction. We visualize the reconstructed Cream object by using the alternative objectives in Fig. 5. Here, we observe physically implausible deformations in results for surface and continuum supervision. Under uniform force, the object exhibits inconsistent distortion, espe", + "bbox": [ + 509, + 815, + 895, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5950", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42886a4924893fee5eba6c102b9c3f58652849db4c7b2ef5dfe4fda829f36fd2.jpg", + "image_caption": [ + "Figure 6. Intra-class generalization. We first optimize the geometry for PAC-NeRF Bird and then use the constitutive models optimized for reference elastic objects to predict novel dynamic behaviors, thereby softening or hardening the Bird respectively." + ], + "image_footnote": [], + "bbox": [ + 102, + 98, + 486, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cially in the boundary between blue stripes and red stripes. We attribute this to two main reasons. 1) Chamfer Distance based geometrical constraint focuses mainly on the overall structure. While it encourages alignment between simulated and reference shapes, they do not enforce point-wise correspondence. A single particle can thus be matched with different ground truth particles in different time steps. As a result, local mismatches can accumulate, leading to inaccurate deformations. 2) Sparse supervision fails to constrain intermediate motion dynamics. Surface and continuum objectives provide alignment only at discrete time steps, leaving large gaps in supervision. Without explicit constraints on how the object deforms over time, the neural constitutive model may interpolate arbitrarily, leading to unstable and unrealistic motion patterns.", + "bbox": [ + 99, + 414, + 483, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamics generalization. We further study the generalization ability for MASIV. Fig. 6 examines how MASIV can be used to modify the dynamic characteristics of a single object (PAC-NeRF Bird). Specifically, we first reconstruct the geometrical representation of the Bird and optimize neural constitutive models for two elastic cross-shaped objects. These optimized constitutive models are then used as references to drive the simulation of the Bird, which deforms accordingly, exhibiting fluid-like softness on the left and increased rigidity and bounce on the right. Additionally, we see continuous transitions between soft and rigid behaviors when interpolating the neural constitutive model weights. In Fig. 7, we extend this study to inter-class generalization by applying neural constitutive models trained on different materials to a novel object, the PAC-NeRF Trophy. The results show that despite sharing the same initial state, each material model induces distinct deformation patterns. These", + "bbox": [ + 99, + 638, + 483, + 890 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96852bb1d70c3f30255e28d3469a5bfa833545729202d706c56af8b9aca9bce7.jpg", + "image_caption": [ + "Figure 7. Inter-class generalization. We train multiple neural constitutive models with different materials and apply them to the PAC-NeRF Trophy. While sharing the same initial state, the Trophies with different materials exhibit distinct behaviors over time." + ], + "image_footnote": [], + "bbox": [ + 514, + 98, + 890, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "observations confirm that MASIV effectively captures generalizable intrinsic dynamics. More qualitative results can be found in Sec. D of Supp. Mat.", + "bbox": [ + 511, + 411, + 893, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Discussion", + "text_level": 1, + "bbox": [ + 511, + 463, + 627, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While MASIV excels in material-agnostic system identification, several limitations remain. First, it relies on multiview video input, restricting applicability to single-view or sparse observations. Second, our approach assumes that internal geometries and deformations can be inferred from surface motion, following a spatially coherent model. This assumption may break down in materials with complex internal flows, such as granular media or multiphase fluids. Finally, MASIV does not explicitly disentangle material properties from external forces in system identification, assuming a uniform gravity field for all objects. Future works aim to relax these assumptions for broader applicability.", + "bbox": [ + 509, + 484, + 893, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 672, + 627, + 688 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present MASIV, a material-agnostic system identification framework. By integrating neural constitutive models with dense geometry guidance, MASIV effectively captures object dynamics across diverse materials. Our approach reconstructs dynamic Gaussians from multi-view videos and fine-tunes a deformation model within a solid continuum. This enables us to extract dense motion cues, which, combined with visual observations, guide the learning of neural constitutive models. Through this formulation, MASIV generalizes beyond traditional system identification methods that rely on pre-defined material priors. Extensive experiments highlight its ability to reconstruct observable dynamics and to generalize for novel interactions.", + "bbox": [ + 509, + 698, + 893, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5951", + "bbox": [ + 483, + 926, + 513, + 938 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 102, + 101, + 250, + 117 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by U.S. NSF grant DBI-2238093, and in part by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project).", + "bbox": [ + 99, + 125, + 483, + 194 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 102, + 219, + 194, + 234 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 2", + "[2] Genesis Authors. Genesis: A universal and generative physics engine for robotics and beyond, 2024. 3", + "[3] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B Lindell. 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7996-8006, 2024. 3", + "[4] Piotr Borycki, Weronika Smolak, Joanna Waczyńska, Marcin Mazur, Sławomir Tadeja, and Przemysław Spurek. Gasp: Gaussian splatting for physic-based simulations. arXiv preprint arXiv:2409.05819, 2024. 3", + "[5] Johannes Brandstetter, Daniel Worrall, and Max Welling. Message passing neural pde solvers. arXiv preprint arXiv:2202.03376, 2022. 2", + "[6] Junhao Cai, Yuji Yang, Weihao Yuan, Yisheng He, Zilong Dong, Liefeng Bo, Hui Cheng, and Qifeng Chen. Gic: Gaussian-informed continuum for physical property identification and simulation. arXiv preprint arXiv:2406.14927, 2024. 2, 3, 4, 5, 6, 7", + "[7] Yuanhao Cai, Jiahao Wang, Alan Yuille, Zongwei Zhou, and Angtian Wang. Structure-aware sparse-view x-ray 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11174-11183, 2024. 2", + "[8] Junyi Cao, Shanyan Guan, Yanhao Ge, Wei Li, Xiaokang Yang, and Chao Ma. Neuma: Neural material adaptor for visual grounding of intrinsic dynamics. Advances in Neural Information Processing Systems, 37:65643-65669, 2025. 2, 3, 4, 5, 6", + "[9] Pradyumna Chari, Chinmay Talegaonkar, Yunhao Ba, and Achuta Kadambi. Visual physics: Discovering physical laws from videos. arXiv preprint arXiv:1911.11893, 2019. 2", + "[10] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splattering guidance. arXiv preprint arXiv:2312.00846, 2023. 4", + "[11] Hsiao-yu Chen, Edith Tretschk, Tuur Stuyck, Petr Kadlecek, Ladislav Kavan, Etienne Vouga, and Christoph Lassner. Virtual elastic objects. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 109, + 243, + 483, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 15827-15837, 2022. 3", + "[12] Tao Du, Kui Wu, Pingchuan Ma, Sebastien Wah, Andrew Spielberg, Daniela Rus, and Wojciech Matusik. Diffpd: Differentiable projective dynamics. ACM Transactions on Graphics (TOG), 41(2):1-21, 2021. 3", + "[13] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In European Conference on Computer Vision, pages 108-124. Springer, 2020. 5", + "[14] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, et al. Gaussian splashing: Unified particles for versatile motion synthesis and rendering. arXiv preprint arXiv:2401.15318, 3, 2024. 3", + "[15] Yutao Feng, Yintong Shang, Xuan Li, Tianjia Shao, Chenfanfu Jiang, and Yin Yang. Pie-nerf: Physics-based interactive elastodynamics with nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4450–4461, 2024. 3", + "[16] Zhoujie Fu, Jiacheng Wei, Wenhao Shen, Chaoyue Song, Xiaofeng Yang, Fayao Liu, Xulei Yang, and Guosheng Lin. Sync4d: Video guided controllable dynamics for physics-based 4d generation. arXiv preprint arXiv:2405.16849, 2024.3", + "[17] Moritz Geilinger, David Hahn, Jonas Zehnder, Moritz Bächer, Bernhard Thomaszewski, and Stelian Coros. Add: Analytically differentiable dynamics for multi-body systems with frictional contact. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 3", + "[18] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, et al. Motion prompting: Controlling video generation with motion trajectories. arXiv preprint arXiv:2412.02700, 2024. 3", + "[19] Sharath Girish, Tianye Li, Amrita Mazumdar, Abhinav Shrivastava, David Luebke, and Shalini De Mello. Queen: Quantized efficient encoding of dynamic gaussians for streaming free-viewpoint videos. arXiv preprint arXiv:2412.04469, 2024. 2", + "[20] Shanyan Guan, Huayu Deng, Yunbo Wang, and Xiaokang Yang. Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields. In International conference on machine learning, pages 7919-7929. PMLR, 2022. 2", + "[21] Eric Heiden, Miles Macklin, Yashraj Narang, Dieter Fox, Animesh Garg, and Fabio Ramos. Disect: A differentiable simulation engine for autonomous robotic cutting. arXiv preprint arXiv:2105.12244, 2021. 3", + "[22] Maximilian Herde, Bogdan Raonic, Tobias Rohner, Roger Käppeli, Roberto Molinaro, Emmanuel de Bézenac, and Siddhartha Mishra. Poseidon: Efficient foundation models for pdes. Advances in Neural Information Processing Systems, 37:72525-72624, 2025. 2, 4", + "[23] Florian Hofherr, Lukas Koestler, Florian Bernard, and Daniel Cremers. Neural implicit representations for physical parameter inference from a single video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2093-2103, 2023. 2" + ], + "bbox": [ + 521, + 103, + 893, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5952", + "bbox": [ + 483, + 926, + 514, + 938 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 5", + "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 6", + "[26] Yuanming Hu, Yu Fang, Ziheng Ge, Ziyin Qu, Yixin Zhu, Andre Pradhana, and Chenfanfu Jiang. A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2, 3", + "[27] Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Fredo Durand. Difftaichi: Differentiable programming for physical simulation. arXiv preprint arXiv:1910.00935, 2019. 2, 3", + "[28] Tianyu Huang, Haoze Zhang, Yihan Zeng, Zhilu Zhang, Hui Li, Wangmeng Zuo, and Rynson WH Lau. Dreamphysics: Learning physical properties of dynamic 3d gaussians with video diffusion priors. arXiv preprint arXiv:2406.01476, 2024.3", + "[29] Zhiao Huang, Yuanming Hu, Tao Du, Siyuan Zhou, Hao Su, Joshua B Tenenbaum, and Chuang Gan. Plasticinelab: A soft-body manipulation benchmark with differentiable physics. arXiv preprint arXiv:2104.03311, 2021. 3", + "[30] Krishna Murthy Jatavallabhula, Miles Macklin, Florian Golemo, Vikram Voleti, Linda Petrini, Martin Weiss, Breandan Considine, Jérôme Parent-Lévesque, Kevin Xie, Kenny Erleben, et al. gradsim: Differentiable simulation for system identification and visuomotor control. arXiv preprint arXiv:2104.02646, 2021. 2, 3", + "[31] Chenfanfu Jiang, Craig Schroeder, Joseph Teran, Alexey Stomakhin, and Andrew Selle. The material point method for simulating continuum materials. In ACM SIGGRAPH 2016 Courses, pages 1-52, 2016. 2, 3, 1", + "[32] Takuhiro Kaneko. Improving physics-augmented continuum neural radiance field-based geometry-agnostic system identification with lagrangian particle optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5480, 2024. 3", + "[33] Namgyu Kang, Jaemin Oh, Youngjoon Hong, and Eunbyung Park. Pig: Physics-informed gaussians as adaptive parametric mesh representations. arXiv preprint arXiv:2412.05994, 2024. 4", + "[34] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 1", + "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 5", + "[36] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. In European Conference on Computer Vision, pages 252-269. Springer, 2025. 3, 4" + ], + "bbox": [ + 109, + 103, + 483, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Jinxi Li, Ziyang Song, and Bo Yang. Nvfi: neural velocity fields for 3d physics learning from dynamic videos. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[38] Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Krishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang, and Chuang Gan. Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. arXiv preprint arXiv:2303.05512, 2023. 2, 3, 4, 5, 6", + "[39] Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B Tenenbaum, and Antonio Torralba. Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. arXiv preprint arXiv:1810.01566, 2018. 3", + "[40] Yifei Li, Tao Du, Kui Wu, Jie Xu, and Wojciech Matusik. Diffcloth: Differentiable cloth simulation with dry frictional contact. ACM Transactions on Graphics (TOG), 42 (1):1-20, 2022. 3", + "[41] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2", + "[42] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. 2", + "[43] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 3", + "[44] Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Azizzadenesheli, and Anima Anandkumar. Physics-informed neural operator for learning partial differential equations. ACM/JMS Journal of Data Science, 1(3):1-27, 2024. 2", + "[45] Junbang Liang, Ming Lin, and Vladlen Koltun. Differentiable cloth simulation for inverse problems. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[46] Xiao Liang, Fei Liu, Yutong Zhang, Yuelei Li, Shan Lin, and Michael Yip. Real-to-sim deformable object manipulation: Optimizing physics models with residual mappings for robotic surgery. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 15471-15477. IEEE, 2024. 3", + "[47] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 5", + "[48] Yuchen Lin, Chenguo Lin, Jianjin Xu, and Yadong Mu. Omniphysgs: 3d constitutive gaussians for general physics-based dynamics generation. arXiv preprint arXiv:2501.18982, 2025. 2, 3", + "[49] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d" + ], + "bbox": [ + 522, + 103, + 893, + 888 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5953", + "bbox": [ + 483, + 928, + 513, + 938 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576-8588, 2024. 3", + "[50] Phillip Lippe, Bas Veeling, Paris Perdikaris, Richard Turner, and Johannes Brandstetter. Pde-refiner: Achieving accurate long rollouts with neural pde solvers. Advances in Neural Information Processing Systems, 36:67398-67433, 2023. 2", + "[51] Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. Physics3d: Learning physical properties of 3d gaussians via video diffusion. arXiv preprint arXiv:2406.04338, 2024. 3", + "[52] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond. arXiv preprint arXiv:1908.03265, 2019. 6", + "[53] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splatting from causally-captured monocular videos. arXiv preprint arXiv:2406.00434, 2024. 2", + "[54] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imageto-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 3", + "[55] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2", + "[56] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2", + "[57] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. 5", + "[58] Pingchuan Ma, Tao Du, Joshua B Tenenbaum, Wojciech Matusik, and Chuang Gan. Risp: Rendering-invariant state predictor with differentiable simulation and rendering for cross-domain parameter estimation. arXiv preprint arXiv:2205.05678, 2022.3", + "[59] Pingchuan Ma, Peter Yichen Chen, Bolei Deng, Joshua B Tenenbaum, Tao Du, Chuang Gan, and Wojciech Matusik. Learning neural constitutive laws from motion observations for generalizable pde dynamics. In International Conference on Machine Learning, pages 23279-23300. PMLR, 2023. 2, 3, 4, 5, 6", + "[60] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[61] Clinton A Mo, Kun Hu, Chengjiang Long, and Zhiyong Wang. Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In Proceedings of the IEEE/CVF Conference on" + ], + "bbox": [ + 109, + 103, + 483, + 888 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition, pages 13894-13903, 2023. 2", + "[62] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2", + "[63] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2", + "[64] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 2", + "[65] Yiling Qiao, Junbang Liang, Vladlen Koltun, and Ming Lin. Differentiable simulation of soft multi-body systems. Advances in Neural Information Processing Systems, 34: 17123-17135, 2021. 3", + "[66] Yi-Ling Qiao, Alexander Gao, and Ming Lin. Neuphysics: Editable neural geometry and physics from monocular videos. Advances in Neural Information Processing Systems, 35:12841-12854, 2022. 3", + "[67] Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Language-driven physics-based scene synthesis and editing via feature splatting. In European Conference on Computer Vision, pages 368-383. Springer, 2024. 3", + "[68] Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378:686-707, 2019. 3", + "[69] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 3", + "[70] Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In International conference on machine learning, pages 8459-8468. PMLR, 2020. 3", + "[71] Yidi Shao, Mu Huang, Chen Change Loy, and Bo Dai. Gausim: Registering elastic objects into digital world by gaussian simulator. arXiv preprint arXiv:2412.17804, 2024.3", + "[72] Haochen Shi, Huazhe Xu, Samuel Clarke, Yunzhu Li, and Jiajun Wu. Robocook: Long-horizon elasto-plastic object manipulation with diverse tools. arXiv preprint arXiv:2306.14447, 2023. 3", + "[73] Haochen Shi, Huazhe Xu, Zhiao Huang, Yunzhu Li, and Ji-jun Wu. Robocraft: Learning to see, simulate, and shape elasto-plastic objects in 3d with graph networks. The International Journal of Robotics Research, 43(4):533-549, 2024. 3", + "[74] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun" + ], + "bbox": [ + 522, + 103, + 893, + 888 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5954", + "bbox": [ + 483, + 928, + 514, + 938 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3", + "[75] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 3", + "[76] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 2", + "[77] Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 1", + "[78] Deborah Sulsky, Shi-Jian Zhou, and Howard L Schreyer. Application of a particle-in-cell method to solid mechanics. Computer physics communications, 87(1-2):236-252, 1995. 2, 3", + "[79] Priya Sundaresan, Rika Antonova, and Jeannette Bohgl. Diffcloud: Real-to-sim from point clouds with differentiable simulation and rendering of deformable objects. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10828-10835. IEEE, 2022. 3", + "[80] Tetsuya Takahashi and Ming C Lin. Video-guided real-to-virtual parameter transfer for viscous fluids. ACM Transactions on Graphics (TOG), 38(6):1-12, 2019. 3", + "[81] Xiyang Tan, Ying Jiang, Xuan Li, Zeshun Zong, Tianyi Xie, Yin Yang, and Chenfanfu Jiang. Physmotion: Physics-grounded dynamics from a single image. arXiv preprint arXiv:2411.17189, 2024. 3", + "[82] Bin Wang, Longhua Wu, KangKang Yin, Uri M Ascher, Libin Liu, and Hui Huang. Deformation capture and modeling of soft objects. ACM Trans. Graph., 34(4):94-1, 2015. 3", + "[83] Chaoyang Wang, Lachlan Ewen MacDonald, Laszlo A Jeni, and Simon Lucey. Flow supervision for deformable nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21128-21137, 2023. 2", + "[84] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[85] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3", + "[86] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splattering for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 3, 4" + ], + "bbox": [ + 109, + 103, + 483, + 888 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[87] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2024. 3", + "[88] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9421-9431, 2021. 2", + "[89] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 3", + "[90] Zhenjia Xu, Jiajun Wu, Andy Zeng, Joshua B Tenenbaum, and Shuran Song. Densephysnet: Learning dense physical object representations via multi-step dynamic interactions. arXiv preprint arXiv:1906.03853, 2019. 3", + "[91] Tianju Xue, Shuheng Liao, Zhengtao Gan, Chanwook Park, Xiaoyu Xie, Wing Kam Liu, and Jian Cao. Jax-fem: A differentiablegpu-accelerated 3d finite element solver for automatic inverse design and mechanistic data science. Computer Physics Communications, 291:108802, 2023. 2", + "[92] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2", + "[93] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3, 4", + "[94] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13117-13125, 2021. 2", + "[95] Heng Yu, Joel Julin, Zoltan A Milacski, Koichiro Niinuma, and László A Jeni. Dylin: Making light field networks dynamic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12397-12406, 2023. 2", + "[96] Heng Yu, Joel Julin, Zoltán Á Milacski, Koichiro Niinuma, and László A Jeni. Cogs: Controllable gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21624-21633, 2024. 2", + "[97] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 4", + "[98] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2025. 3" + ], + "bbox": [ + 522, + 103, + 893, + 888 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "5955", + "bbox": [ + 483, + 928, + 513, + 938 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[99] Dongzhe Zheng, Siqiong Yao, Wenqiang Xu, and Cewu Lu. Differentiable cloth parameter identification and state estimation in manipulation. IEEE Robotics and Automation Letters, 2024. 3", + "[100] Licheng Zhong, Hong-Xing Yu, Jiajun Wu, and Yunzhu Li. Reconstruction and simulation of elastic objects with spring-mass 3d gaussians. In European Conference on Computer Vision, pages 407-423. Springer, 2025. 2, 3, 4, 5, 6", + "[101] Xiaowei Zhou, Can Yang, and Weichuan Yu. Moving object detection by detecting contiguous outliers in the low-rank representation. IEEE transactions on pattern analysis and machine intelligence, 35(3):597-610, 2012. 2" + ], + "bbox": [ + 104, + 103, + 482, + 280 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "5956", + "bbox": [ + 483, + 928, + 514, + 938 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Toward Material-Agnostic System Identification from Videos/58fcfc32-a696-4a6d-a6cd-3c218994c3ba_model.json b/2025/Toward Material-Agnostic System Identification from Videos/58fcfc32-a696-4a6d-a6cd-3c218994c3ba_model.json new file mode 100644 index 0000000000000000000000000000000000000000..04cf3fbb24a960fbd3193ca5d2c634e91ea5875c --- /dev/null +++ b/2025/Toward Material-Agnostic System Identification from Videos/58fcfc32-a696-4a6d-a6cd-3c218994c3ba_model.json @@ -0,0 +1,2844 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.117, + 0.017, + 0.191, + 0.056 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.246, + 0.015, + 0.8, + 0.03 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.33, + 0.031, + 0.717, + 0.044 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.299, + 0.045, + 0.748, + 0.058 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.199, + 0.141, + 0.8, + 0.163 + ], + "angle": 0, + "content": "Toward Material-Agnostic System Identification from Videos" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.195, + 0.835, + 0.213 + ], + "angle": 0, + "content": "Yizhou Zhao1 Haoyu Chen1 Chunjiang Liu1 Zhenyang Li2 Charles Herrmann3" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.213, + 0.812, + 0.23 + ], + "angle": 0, + "content": "Junhwa \\(\\mathsf{Hur}^3\\) Yinxiao Li3 Ming-Hsuan Yang3,4 Bhiksha Raj1 Min \\(\\mathbf{X}\\mathbf{u}^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.23, + 0.876, + 0.248 + ], + "angle": 0, + "content": "\\(^{1}\\)Carnegie Mellon University \\(^{2}\\)University of Alabama at Birmingham \\(^{3}\\)Google \\(^{4}\\)UC Merced" + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.248, + 0.684, + 0.265 + ], + "angle": 0, + "content": "https://github.com/Skaldak/MASIV" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.291, + 0.855, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.589, + 0.896, + 0.618 + ], + "angle": 0, + "content": "Figure 1. System identification from video observations seeks to understand the intrinsic dynamics of a given scene. (a) Existing methods rely on scene-specific material priors. (b) Our MASIV removes this requirement, grounding object motion in a material-agnostic manner." + }, + { + "type": "title", + "bbox": [ + 0.256, + 0.629, + 0.331, + 0.643 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.657, + 0.486, + 0.864 + ], + "angle": 0, + "content": "System identification from videos aims to recover object geometry and governing physical laws. Existing methods integrate differentiable rendering with simulation but rely on predefined material priors, limiting their ability to handle unknown ones. We introduce MASIV, the first vision-based framework for material-agnostic system identification. Unlike existing approaches that depend on hand-crafted constitutive laws, MASIV employs learnable neural constitutive models, inferring object dynamics without assuming a scene-specific material prior. However, the absence of full particle state information imposes unique challenges, leading to unstable optimization and physically implausible behaviors. To address this, we introduce dense geometric guidance by reconstructing continuum particle trajectory" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.896, + 0.69 + ], + "angle": 0, + "content": "ries, providing temporally rich motion constraints beyond sparse visual cues. Comprehensive experiments show that MASIV achieves state-of-the-art performance in geometric accuracy, rendering quality, and generalization ability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.704, + 0.641, + 0.718 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.728, + 0.896, + 0.891 + ], + "angle": 0, + "content": "From the flow of water down a stream to the deformation of a rubber ball upon impact, the physical world operates with a set of underlying principles that govern motion and interaction. Humans intuitively grasp these principles through observation, allowing us to see a scene and instinctively predict its outcome or re-imagine it with different initial dynamics. Material identification forms the basis of this ability, requiring us to determine an object's composition and understand how that influences its response to forces. This capability is fundamental to world modeling, the process of constructing internal representations that help us understand" + }, + { + "type": "page_footnote", + "bbox": [ + 0.12, + 0.877, + 0.241, + 0.889 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5944" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.482, + 0.148 + ], + "angle": 0, + "content": "and anticipate the evolution of the environment. However, replicating the ability to generalize physical laws to diverse scenarios remains a challenge for artificial intelligence." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.15, + 0.483, + 0.444 + ], + "angle": 0, + "content": "To bridge this gap, researchers have explored vision-based system identification [6, 8, 9, 20, 23, 30, 38], which aims to infer governing physical laws from visual observations. Typically, they optimize a parameterized physical model, by end-to-end integrating differentiable renderers, such as Neural Radiance Fields (NeRF) [60] and 3D Gaussian Splatting (3DGS) [34], with differentiable simulators [26, 27, 31, 78, 91]. To enforce physically meaningful constraints on the simulation, existing approaches [6, 38, 100] often rely on material-specific constitutive laws, such as elasticity, plasticity, or viscosity models. These laws define how an object's material properties influence its response to external forces, allowing system identification methods to estimate parameters such as stiffness, damping, or friction. While effective in controlled settings, this dependence on hand-crafted material models limits adaptability. Moreover, these methods assume knowing the material type for a given scene, making it difficult for them to generalize across in-the-wild scenarios where material properties are unknown." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.447, + 0.483, + 0.697 + ], + "angle": 0, + "content": "Directly targeting these limitations, we introduce Material-Agnostic System Identification from Videos (MASIV), a vision-based approach that infers object dynamics free of predefined material priors. As illustrated in Fig. 1, rather than relying on expert-designed elasticity and plasticity models, MASIV employs fully learnable neural constitutive models inspired by NCLaw [59]. However, directly applying neural constitutive models to vision-based system identification poses significant challenges. Similar to the learning process of neural PDE solvers [5, 22, 44, 50], the original formulation of NCLaw assumes access to complete state observations of simulated particles, including position, velocity, deformation gradient, and affine momentum. Since visual observation fails to provide such complete information, simply relying on frame-wise pixel supervision provides insufficient constraints. This leads to unstable optimization and poor convergence [48]." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.7, + 0.483, + 0.891 + ], + "angle": 0, + "content": "We combine two main insights to enable this optimization. First, despite the lack of ground truth full particle state, frame-wise observations still allow reasonable estimates for some of them, such as position and velocity. Existing dynamic reconstruction methods [42, 55, 56] factorize moving objects with their canonical representation and corresponding deformations, which enables per-frame position tracking. Further, the commonly adopted motion basis [61, 94, 101] provides a low-rank encoding of object motion, facilitating state interpolation between frames. This formulation allows us to reconstruct particle trajectories over dense time steps. Second, although trajectory estimations are inherently limited to visible object regions," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.894, + 0.162 + ], + "angle": 0, + "content": "internal deformations can be inferred by propagating surface motions inward. Our intuition is that nearby particles inside an object should be deformed similarly, akin to the As-Rigid-As-Possible (ARAP) prior." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.164, + 0.895, + 0.368 + ], + "angle": 0, + "content": "We use these insights to guide the design of our framework. Concretely, we reconstruct dynamic Gaussians from multi-view video inputs, simultaneously establishing geometry and deformation models of the scene. The geometry model is then transformed into a solid continuum, where the deformation model is fine-tuned to encode both the exterior and the interior motion. To identify intrinsic dynamics, we guide the learning of the neural constitutive model using two complementary cues: appearance cues from video frames and geometric constraints from the fine-tuned deformation model. In this way, MASIV distills dense particle motion from sparse visual observations and guides the learning of neural constitutive models for stable, material-agnostic system identification." + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.371, + 0.844, + 0.385 + ], + "angle": 0, + "content": "Our contributions can be summed up as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.387, + 0.894, + 0.459 + ], + "angle": 0, + "content": "- We present a novel framework, MASIV, that determines the intrinsic dynamics of an object from videos in a material-agnostic manner. In contrast, all other current vision-based system identification methods [6, 8, 38] necessitate predefined material priors." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.46, + 0.894, + 0.518 + ], + "angle": 0, + "content": "- We introduce temporally dense geometry guidance to constrain object behaviors with neural constitutive models, facilitating smoother convergence and improved performance for material-agnostic system identification." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.52, + 0.894, + 0.577 + ], + "angle": 0, + "content": "- Through extensive qualitative and quantitative experiments, we demonstrate that MASIV achieves superior performance in both observable state recovery and future state prediction, even without explicit material priors." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.387, + 0.894, + 0.577 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.596, + 0.651, + 0.611 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.625, + 0.895, + 0.891 + ], + "angle": 0, + "content": "Dynamic reconstruction is a fundamental task in computer vision, aimed at reconstructing high-fidelity representations of complex dynamic scenes from diverse inputs such as monocular and multi-view videos [1, 7, 19, 41, 53, 76, 83, 88, 92, 95, 96]. Earlier research primarily focused on augmenting canonical Neural Radiance Fields (NeRF) [60] with a deformation field [64], which were later enhanced by incorporating volume-preserving regularization techniques [62, 63]. While primarily focused on the novel view synthesis task, their neural implicit representations often introduce noisy deformations, limiting the accuracy of recovered geometries required for reliable physics-based property estimation [38]. With the explicit representation of the scene using Gaussian ellipsoids, 3D Gaussian Splatting (3DGS) [34] achieves efficient dynamic scene reconstruction. Building on the principles of 3DGS, several studies have extended its application to 4D dynamic reconstruction by handling each frame separately [56] or mod" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5945" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.484, + 0.148 + ], + "angle": 0, + "content": "eling the entire scene with a canonical 3D Gaussian point cloud coupled with a robust deformation framework that accurately maps it to the target scene [36, 86, 93]." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.152, + 0.484, + 0.563 + ], + "angle": 0, + "content": "Dynamic simulation integrates physics principles into 3D representations to generate realistic motion and interactions. A typical approach to synthesizing dynamic 3D scenes involves integrating a 3D generation pipeline with a video generation model [3, 49, 69, 75]. Recent research incorporates 3D Gaussian kernels as both visual and physical representations, embedding Newtonian dynamics into the Gaussian framework to enable seamless rendering and simulation while introducing physical constraints to ensure plausible dynamics [4, 16, 38, 48, 51, 67, 89, 100]. Building on this, Gaussian Splashing [14] introduces a position-based dynamics framework within 3D Gaussian Splating, enabling interactions among solids, fluids, and deformable objects. Additionally, motion-conditioned simulation methods [18, 43, 74, 85, 87] use trajectory-based guidance to predict object motion, showcasing how generative models can enhance dynamic scene synthesis. However, these methods primarily rely on pre-trained generative models, which often lack a fundamental understanding of physics. More recent works leverage neural physics modeling, incorporating video priors and generative models to infer object dynamics, allowing physics-driven motion synthesis without explicit solvers [15, 28, 54, 81, 98]. However, these methods are constrained to specific material types, limiting their generalization across diverse objects. To address this, we introduce a material-agnostic approach that unifies system identification across different object types, ensuring broader applicability beyond predefined materials." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.567, + 0.484, + 0.891 + ], + "angle": 0, + "content": "System identification aims to understand the physical laws governing the 3D world and is an essential task for simulation [37, 40, 45, 68, 79, 100] and robotic manipulation [46, 66, 72, 73, 99]. However, accurately recovering both object geometries and physical properties remains a significant challenge. Existing methods often assume known geometries and rely on finite element methods (FEM) or mass-spring systems for dynamic simulation [80, 82], limiting their ability to handle complex, real-world materials. Neural network-based methods [39, 70, 90] offer greater flexibility by leveraging data-driven models for system identification, but they struggle to generalize across diverse material properties and environmental conditions. More recently, differentiable physics-based simulations have emerged as a promising direction [11, 12, 17, 21, 27, 29, 30, 32, 58, 65]. However, these methods often require intricate modeling to bridge the gap between simulated and real-world behavior. To improve adaptability, some approaches focus on neural constitutive modeling, embedding learned material properties within physical simulations to refine expert-designed models [8, 59]. Other methods leverage the Material Point Method (MPM)" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.103, + 0.751, + 0.117 + ], + "angle": 0, + "content": "Algorithm 1 MPM state transfer \\(\\mathcal{M}\\)" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.121, + 0.692, + 0.138 + ], + "angle": 0, + "content": "Input: \\(\\mathbf{s}_n = \\{\\mathbf{x}_n,\\mathbf{v}_n,\\mathbf{F}_n^e\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.138, + 0.767, + 0.152 + ], + "angle": 0, + "content": "Output: \\(\\mathbf{s}_{n + 1} = \\{\\mathbf{x}_{n + 1},\\mathbf{v}_{n + 1},\\mathbf{F}_{n + 1}^{e}\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.152, + 0.765, + 0.167 + ], + "angle": 0, + "content": "1: for \\( i = 1 \\dots Q \\), \\( \\mathbf{P}_n(i) = \\mathcal{E}(\\mathbf{F}_n^e (i)) \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.167, + 0.808, + 0.183 + ], + "angle": 0, + "content": "2: \\(\\mathbf{x}_{n + 1},\\mathbf{v}_{n + 1},\\mathbf{F}_{n + 1}^{e,\\mathrm{trial}} = \\mathcal{I}(\\mathbf{x}_n,\\mathbf{v}_n,\\mathbf{F}_n^e,\\mathbf{P}_n)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.184, + 0.802, + 0.201 + ], + "angle": 0, + "content": "3: for \\( i = 1 \\dots Q \\), \\( \\mathbf{F}_{n + 1}^{e}(i) = \\mathcal{P}(\\mathbf{F}_{n + 1}^{e,\\mathrm{trial}}(i)) \\)" + }, + { + "type": "list", + "bbox": [ + 0.526, + 0.152, + 0.808, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.229, + 0.896, + 0.348 + ], + "angle": 0, + "content": "and 3D Gaussian-based representations to enhance geometry reconstruction and material property estimation [6, 38, 71, 100]. Building upon GIC [6], our method presents a material-agnostic approach that eliminates reliance on predefined material priors. Rather than assuming material properties, our model infers object dynamics directly from visual observations, making it adapt better to a wide range of unknown or heterogeneous materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.359, + 0.602, + 0.373 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.382, + 0.896, + 0.545 + ], + "angle": 0, + "content": "Given visual observations \\(\\mathbf{I} = \\{\\mathbf{I}_0, \\mathbf{I}_1, \\dots, \\mathbf{I}_T\\}\\) over \\(T\\) frames capturing a dynamic object, our goal is to recover its explicit geometric representation and the governing evolution rules of the underlying dynamical system. Formally, we represent the object's geometry using a set of Gaussian kernels [34] and integrate neural constitutive laws [59] with the material point method (MPM) [26, 31, 78] to roll out their motion over time. Unlike [6, 8, 38, 100], we do not assume a known material type of the object, while exploring Material-Agnostic System Identification from Videos (MASIV). Fig. 2 shows an overview of our pipeline." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.552, + 0.651, + 0.567 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.895, + 0.622 + ], + "angle": 0, + "content": "Material Point Method (MPM) discretizes the object of interest into \\( Q \\) material points and models their evolution over time via a state transfer equation" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.631, + 0.894, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {n + 1} = \\mathcal {M} (\\mathbf {s} _ {n}), \\forall n = 0, 1, \\dots , N. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.655, + 0.895, + 0.699 + ], + "angle": 0, + "content": "\\(\\mathbf{s}_n\\) denotes the particle state at time step \\(n\\). \\(N\\) is the total number of simulation steps with the simulation interval \\(\\tau\\) and observed \\(T\\) frames, where \\(N = T / \\tau \\gg T\\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.7, + 0.895, + 0.891 + ], + "angle": 0, + "content": "The transition process \\(\\mathcal{M}\\) follows a time-stepping scheme, as outlined in Algorithm 1. The state of particle \\(i\\) at time step \\(n\\) is described by its position \\(\\mathbf{x}_n(i)\\), velocity \\(\\mathbf{v}_n(i)\\), and the elastic component \\(\\mathbf{F}_n^e (i)\\) of its deformation gradient \\(\\mathbf{F}_n(i)\\). At each time step, the elastic constitutive law \\(\\mathcal{E}\\) computes the first Piola-Kirchhoff stress \\(\\mathbf{P}_n(i)\\) from the elastic deformation gradient \\(\\mathbf{F}_n^e (i)\\). Subsequently, the time integration scheme \\(\\mathcal{I}\\) updates all particle states. The updated trial elastic deformation gradient, denoted as \\(\\mathbf{F}_{n + 1}^{e,\\mathrm{trial}}(i)\\), needs to be further corrected by the plastic constitutive law \\(\\mathcal{P}\\), which applies plasticity constraints and yields the final elastic deformation gradient \\(\\mathbf{F}_{n + 1}^{e}(i)\\). More details of MPM can be found in Sec. A.2 of Supp. Mat." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5946" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.1, + 0.894, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.226, + 0.894, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.372, + 0.897, + 0.427 + ], + "angle": 0, + "content": "Figure 2. The overview of MASIV. Our pipeline comprises three phases. Phase I, Geometrical Reconstruction, infers geometrical representations (Sec. 3.2) and dense trajectories (Sec. 3.3) for the captured object. Phase II, Material-Agnostic System Identification (Sec. 3.4), grounds dynamic characteristics by exploiting visual observations with reconstructed motion clues. From this, we obtain a generalizable digital twin in Phase III, which can simulate novel interactions, such as new velocities or forces." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.44, + 0.484, + 0.573 + ], + "angle": 0, + "content": "Existing efforts [6, 8, 38, 100] for video-based system identification commonly assume known constitutive laws \\(\\mathcal{E}\\) and \\(\\mathcal{P}\\), with a particular backbone form, e.g., neo-Hookean elasticity with identity plasticity return function. This simplifies the problem by estimating a small number of physical properties, such as Young's modulus, fluid viscosity, friction angles, etc., at the cost of sacrificing flexibility. In contrast, we explore the same problem without this material-specific assumption." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.581, + 0.396, + 0.595 + ], + "angle": 0, + "content": "3.2. Dynamic Gaussian Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.602, + 0.484, + 0.807 + ], + "angle": 0, + "content": "We start by reconstructing dynamic Gaussians over \\(T\\) frames. We consider this reconstruction as temporally sparse because the number of simulation steps is hundreds of times more than the number of observed video frames \\(T\\), i.e., \\(N = T / \\tau \\gg T\\). Following [6, 36, 86, 93], we maintain a set of canonical Gaussians \\(\\mathcal{G}^* = \\{\\pmb{\\mu}^*, \\pmb{\\Sigma}^*, \\mathbf{c}^*, \\sigma^*\\}\\) and use a neural network \\(\\mathcal{D}\\) to warp them over time to represent observed frames. Specifically, a basis network maps a given time step \\(t\\) to \\(B\\) bases, producing deformation basis for position \\(\\mathbf{B}^{\\mu}(t) \\in \\mathbb{R}^{B \\times 3}\\) and scale \\(\\mathbf{B}^s(t) \\in \\mathbb{R}^{B \\times 1}\\). A coefficient network then estimates the weight of each basis, \\(\\mathbf{w}(\\pmb{\\mu}^*, t) \\in \\mathbb{R}^B\\), from the center coordinates of canonical Gaussian kernels \\(\\pmb{\\mu}^*\\) and the time step \\(t\\). This results in deformed Gaussian kernels at each time step \\(t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.817, + 0.483, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {t} = \\boldsymbol {\\mu} ^ {*} + \\mathbf {w} (\\boldsymbol {\\mu} ^ {*}, t) \\mathbf {B} ^ {\\boldsymbol {\\mu}} (t), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.836, + 0.483, + 0.852 + ], + "angle": 0, + "content": "\\[\ns _ {t} = s ^ {*} + \\mathbf {w} \\left(\\boldsymbol {\\mu} ^ {*}, t\\right) \\mathbf {B} ^ {s} (t). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.86, + 0.484, + 0.89 + ], + "angle": 0, + "content": "We employ the same optimization strategy in [6, 93] to minimize the L1 and Structural Similarity Index Mea" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.44, + 0.895, + 0.527 + ], + "angle": 0, + "content": "sure (SSIM) losses between rendered frames \\(\\hat{\\mathbf{I}}_t\\) and corresponding ground-truth observations \\(\\mathbf{I}_t\\). We also regard all Gaussian kernels as isotropic and apply L1 regularization on scales for simplicity and reconstruction quality [10, 97, 100]. The overall objective function is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.533, + 0.894, + 0.56 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {G} ^ {*}, \\mathcal {D}} \\left[ \\mathcal {L} _ {1} (\\hat {\\mathbf {I}} _ {t}, \\mathbf {I} _ {t}) + \\lambda_ {\\mathrm {S S I M}} \\mathcal {L} _ {\\mathrm {S S I M}} (\\hat {\\mathbf {I}} _ {t}, \\mathbf {I} _ {t}) + \\lambda_ {s} \\| s _ {t} \\| _ {1} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.567, + 0.747, + 0.582 + ], + "angle": 0, + "content": "with hyperparameters \\(\\lambda_{\\mathrm{SSIM}}\\) and \\(\\lambda_{s}\\)" + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.589, + 0.803, + 0.605 + ], + "angle": 0, + "content": "3.3. Continuum Trajectory Estimation" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.61, + 0.895, + 0.831 + ], + "angle": 0, + "content": "To bridge the gap between unevenly distributed Gaussian particles and the corresponding solid continuum, we follow [6] to fill internal volumes, forming continuum particles \\(\\{\\mathbf{p}_t\\}\\) with uniform scale \\(s\\) and density \\(\\sigma\\). Existing methods for material-specific system identification often rely on sparse visual [6, 8, 38] or geometrical [6] cues along \\(T\\) frames to optimize a small set of parameters. However, since we do not assume a known constitutive backbone or material types, these sparse signals may lead to unconstrained behaviors at unobserved time steps or even cause training instability. Inspired by system identification [22, 33, 59] using state data per simulation step, we propose leveraging particle trajectories over \\(N\\) simulation steps as dense geometrical supervision to better constrain the simulated behaviors of continuum particles." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.831, + 0.895, + 0.89 + ], + "angle": 0, + "content": "Given that the deformation network trained in Sec. 3.2 can produce rough trajectories for dynamic Gaussians, we fine-tune it to establish trajectories for continuum particles. Specifically, we initialize the canonical space of particles" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5947" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.486, + 0.133 + ], + "angle": 0, + "content": "\\(\\mathbf{x}_0\\) with the filled continuum \\(\\mathbf{p}_1\\) at \\(t = 1\\) and apply deformations similar to Eq. (2)," + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.142, + 0.483, + 0.159 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} = \\mathbf {x} ^ {*} + \\mathbf {w} \\left(\\mathbf {x} ^ {*}, t\\right) \\mathbf {B} ^ {\\mathbf {x}} (t), \\forall t = 0, 1, \\dots , T, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.168, + 0.483, + 0.213 + ], + "angle": 0, + "content": "resulting in per-frame deformed particles \\(\\{\\mathbf{x}_t\\}\\). Then we optimize this deformation with all filled continuums \\(\\{\\mathbf{p}_t\\}\\) using the Chamfer Distance [13, 57]" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.221, + 0.483, + 0.244 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {x} ^ {*}, \\mathcal {D} ^ {\\prime}} \\left[ \\mathcal {L} _ {\\mathrm {C D}} \\left(\\left\\{\\mathbf {x} _ {t} \\right\\} _ {t = 0} ^ {T}, \\left\\{\\mathbf {p} _ {t} \\right\\} _ {t = 0} ^ {T}\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.254, + 0.485, + 0.372 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}'\\) denotes the fine-tuned deformation network. Unlike the deformation network \\(\\mathcal{D}\\) trained in Sec. 3.2, this fine-tuned version \\(\\mathcal{D}'\\) is aware of the internal particles absence in Gaussians. Further, the learned motion basis with temporal positional embeddings enables it for temporal interpolation, allowing us to probe the motion between observed discrete time steps \\(t = 0,1,\\dots,T\\). Formally, we have particle positions at each simulation time step" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.382, + 0.483, + 0.398 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} = \\mathbf {x} ^ {*} + \\mathbf {w} (\\mathbf {x} ^ {*}, t) \\mathbf {B} ^ {\\mathbf {x}} (t), \\forall t = 0, \\tau , \\dots , N \\tau , \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.408, + 0.484, + 0.452 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the time interval between two simulation steps. We then consider these temporally dense particle trajectories as pseudo ground truth for later optimization." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.46, + 0.436, + 0.475 + ], + "angle": 0, + "content": "3.4. Material-Agnostic System Identification" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.481, + 0.484, + 0.746 + ], + "angle": 0, + "content": "Inspired by NCLaw [59], we parameterize the elastic and plastic constitutive laws using neural networks, denoted as \\(\\mathcal{E}_{\\theta^e}\\) and \\(\\mathcal{P}_{\\theta^p}\\), where \\(\\theta^e\\) and \\(\\theta^p\\) are their respective network parameters. Both networks share the same multi-layer perceptron (MLP) architecture with a single hidden layer. Further, they incorporate two physical priors, frame indifference and undeformed state equilibrium, enforced by using rotation-invariant input representations and eliminating bias terms, respectively. While we adopt the same parameterization for material-agnostic modeling, we introduce a more challenging setting by relaxing the input requirements from dense, complete state sequences \\(\\{\\mathbf{s}_t\\}_{n=0}^N\\) to sparse visual observations \\(\\{\\mathbf{I}_t\\}_{t=0}^T\\). To this end, we follow an analysis-by-synthesis scheme. For the initial particle state \\(\\mathbf{s}_0\\), we use positions \\(\\mathbf{x}_0\\) derived from Eq. (7), an estimizable velocity \\(\\mathbf{v}_0\\), and identity deformation gradients \\(\\mathbf{F}_0\\). After we evolve particle positions \\(\\{\\hat{\\mathbf{x}}_t\\}_{t=\\tau}^{N\\tau}\\) through time integration as described in Eq. (1), we optimize \\(\\theta^e\\) and \\(\\theta^p\\) with" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.756, + 0.483, + 0.777 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta^ {e}, \\theta^ {p}} \\left(\\mathcal {L} _ {\\mathrm {g e o}} + \\mathcal {L} _ {\\mathrm {s i l}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.787, + 0.484, + 0.834 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{geo}}\\) supervises particle positions \\(\\hat{\\mathbf{x}}_t\\) with estimated trajectories \\(\\mathbf{x}_t\\) and \\(\\mathcal{L}_{\\mathrm{sil}}\\) compares rendered masks \\(\\hat{\\mathbf{M}}_t\\) with object silhouettes \\(\\mathbf{M}_t\\) extracted from frame \\(\\mathbf{I}_t\\)" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.842, + 0.483, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g e o}} = \\mathcal {L} _ {\\mathrm {t r a j}} = \\mathcal {L} _ {1} \\left(\\left\\{\\hat {\\mathbf {x}} _ {t} \\right\\} _ {t = 0} ^ {N \\tau}, \\left\\{\\mathbf {x} _ {t} \\right\\} _ {t = 0} ^ {N \\tau}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.864, + 0.483, + 0.887 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s i l}} = \\mathcal {L} _ {1} \\left(\\left\\{\\hat {\\mathbf {M}} _ {t} \\right\\} _ {t = 0} ^ {N \\tau}, \\left\\{\\mathbf {M} _ {t} \\right\\} _ {t = 0} ^ {N \\tau}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.895, + 0.147 + ], + "angle": 0, + "content": "While not assuming known material types in our problem setting, we note that MASIV still requires a pre-trained constitutive model from [59] as a stable initialization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.166, + 0.642, + 0.183 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.192, + 0.715, + 0.208 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.218, + 0.895, + 0.483 + ], + "angle": 0, + "content": "Datasets. We conduct our experiments on the PAC-NeRF dataset [38] and the Spring-Gaus dataset [100]. The PAC-NeRF dataset consists of synthetic objects simulated using the Material Point Method (MPM) and includes a variety of material types such as elastic bodies, plasticine, sand, and Newtonian and non-Newtonian fluids. This dataset provides 45 multi-view RGB video sequences that allow for the estimation of both object geometry and physical properties and approximately 14 frames per viewpoint. The Spring-Gaus dataset contains two subsets, namely, Synthetic and Real. The Synthetic subset focuses on the reconstruction and simulation of elastic objects, containing 30 frames in each of 10 viewpoints. The Real subset provides both dense-view static and sparse-view dynamic captures of elastic objects. We evaluate the performance of observable state simulation using both datasets, while the Spring-Gaus dataset is additionally utilized for future state prediction. To obtain object masks, we apply off-the-shelf matting with [35, 47]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.895, + 0.768 + ], + "angle": 0, + "content": "Baselines. To assess the performance of our approach, we compare it against four state-of-the-art methods for system identification. PAC-NeRF [38] combines neural radiance fields with differentiable physics to jointly infer geometry and material properties. Spring-Gaus [100] integrates a spring-mass system into Gaussian representations for efficient elastic object simulation. NeuMA [8] fine-tunes a neural adapter on expert-designed physical models to better align reconstructed dynamics with visual observations. GIC [6] leverages a two-stage pipeline to obtain the geometry representation and physical properties sequentially. These baselines represent the most relevant prior work and provide a strong comparison for evaluating our method. We note that they all necessitate material priors. To be concrete, PAC-NeRF and GIC roughly categorize materials into five groups and adopt a certain set of physical models for each group. Spring-Gaus generally focuses on elastic objects, thereby being material-specific. And NeuMA leverages a scene-specific constitutive model as the material prior." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.895, + 0.89 + ], + "angle": 0, + "content": "Metrics. We evaluate our method using several standard metrics to assess both geometric reconstruction quality and physical property estimation accuracy. Metrics including 1) Chamfer Distance (CD) [13, 57] with units expressed in \\(10^{3}\\mathrm{mm}^{2}\\); 2) Peak Signal-to-Noise Ratio (PSNR) [24]; 3) Structural Similarity Index Measure (SSIM) [84]. The CD quantifies the similarity between two point distributions, while PSNR and SSIM assess the visual quality." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5948" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.101, + 0.894, + 0.196 + ], + "angle": 0, + "content": "
MethodPAC-NeRFSpring-Gaus
NewtonianNon-NewtonianElasticityPlasticineSandMeanTorusCrossCreamApplePasteChessBananaMean
\\(\\rightarrow\\)DPAC-NeRF [38]0.2770.2360.2380.4290.2120.2784.921.100.771.113.140.962.772.11
Spring-Gaus [100]------0.170.480.360.380.191.802.600.85
NeuMA [8]------4.590.060.080.040.710.050.030.79
GIC [6]0.2430.1950.1780.1960.2500.2120.130.130.140.150.170.410.030.17
MASIV (Ours)0.2330.1980.1920.2010.2290.2100.080.100.190.160.130.190.080.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.205, + 0.893, + 0.233 + ], + "angle": 0, + "content": "Table 1. Observable state simulation on PAC-NeRF and Spring-Gaus Synthetic datasets. Despite being the only technique that does not use expert-designed physical models (noted by the gray row), MASIV performs better than or on par with existing baselines." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.246, + 0.483, + 0.375 + ], + "angle": 0, + "content": "
MethodBunBurgerDogPigPotatoMean
PSNR ↑Spring-Gaus [100]30.6934.0132.1034.9732.7232.90
NeuMA [8]31.2723.7825.6125.40-26.51
GIC [6]34.6840.4537.1738.3239.9538.11
Ours40.2642.1238.9740.8443.4041.12
SSIM ↑Spring-Gaus0.9920.9940.9940.9960.9920.994
NeuMA0.9940.9930.9950.995-0.994
GIC0.9950.9970.9970.9970.9960.996
Ours0.9980.9970.9970.9980.9980.997
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.385, + 0.483, + 0.412 + ], + "angle": 0, + "content": "Table 2. Observable state simulation on Spring-Gaus Real dataset. MASIV achieves state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.427, + 0.483, + 0.574 + ], + "angle": 0, + "content": "Implementation details. In dynamic reconstruction, we follow [6] to optimize canonical Gaussians \\(\\mathcal{G}^*\\) and the deformation network \\(\\mathcal{D}\\) for 40K iterations. Then the deformation network is fine-tuned to estimate particle trajectories from filled continuum sequences for 10K iterations. In system identification, we set the simulation time interval \\(\\tau\\) to \\(1/200\\) the frame duration. We adopt the RAdam [52] optimizer to optimize the initial velocity \\(\\mathbf{v}_0\\) for 100 steps and constitutive model parameters \\(\\theta^e, \\theta^p\\) for 1000 steps. All experiments are conducted on a single NVIDIA A100 GPU." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.583, + 0.289, + 0.598 + ], + "angle": 0, + "content": "4.2. Comparison Results" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.607, + 0.483, + 0.754 + ], + "angle": 0, + "content": "Observable state simulation. In Tab. 1, we compare MASIV with existing vision-based system identification methods for observable state simulation. The results indicate that MASIV outperforms all other methods in terms of CD, highlighting its effectiveness in reconstructing object dynamics across diverse materials. Further, Tab. 2 evaluates on real data. Thanks to the more flexible constitutive representation, MASIV achieves state-of-the-art results. This demonstrates the potential of MASIV as an adaptive and generalizable approach for system identification." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.758, + 0.483, + 0.89 + ], + "angle": 0, + "content": "Future state simulation. In Tab. 3, we provide a detailed evaluation of future state simulation performance. The results indicate that MASIV consistently outperforms all other methods while slightly underperforming GIC. Since NeuMA [8] does not originally experiment on the synthetic subset of Spring-Gaus, we report our reproduction results. To be concrete, we use the closest pre-trained model, Jelly, from NCLaw [59] as the material prior and conduct LoRA [25] fine-tuning with hyperparameters from its orig-" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.246, + 0.895, + 0.42 + ], + "angle": 0, + "content": "
MethodTorusCrossCreamApplePasteChessBananaMean
CD\\(\\rightarrow\\)PAC-NeRF [38]2.473.872.214.6937.708.2066.4317.94
Spring-Gaus [100]2.381.572.221.877.032.5918.485.16
NeuMA [8]1.170.401147.40.0544.260.330.30170.56
GIC [6]0.751.090.940.222.790.770.120.95
MASIV (Ours)0.160.381.512.641.751.700.241.20
PSNR\\(\\dagger\\)PAC-NeRF [38]17.4614.1515.3719.9412.3215.0816.0415.77
Spring-Gaus [100]16.8316.9315.4221.5514.7116.0817.8917.06
NeuMA [8]16.6119.672.3226.2513.4220.8523.1617.47
GIC [6]20.2430.5119.1526.8916.3118.4429.2922.98
MASIV (Ours)22.4233.7316.4521.1915.7317.2826.5621.91
SSIM\\(\\dagger\\)PAC-NeRF [38]0.9190.9060.8580.8780.8190.8480.8860.870
Spring-Gaus [100]0.9190.9400.8620.9020.8720.8810.9040.897
NeuMA [8]0.9420.9480.8890.9640.8890.9330.9640.933
GIC [6]0.9420.9390.9090.9480.8940.9120.9640.930
MASIV (Ours)0.9550.9670.8520.8980.8910.8780.9650.915
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.43, + 0.894, + 0.458 + ], + "angle": 0, + "content": "Table 3. Future state simulation on Spring-Gaus Synthetic dataset. MASIV outperforms all baselines except for GIC." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.475, + 0.895, + 0.784 + ], + "angle": 0, + "content": "inal implementation. However, we observe unstable optimization processes for most categories. This is probably rooted in its sparse pixel supervision, which only operates once per frame. With dense geometrical constraints supervising hundreds of times per frame, MASIV achieves significantly better results on both geometrical accuracy and rendering quality. This suggests that MASIV effectively captures the underlying dynamics governing future states, allowing it to generate predictions that remain structurally and visually consistent over time. Unlike baseline approaches that rely on predefined material priors, MASIV demonstrates the ability to infer relevant properties directly from visual observations, making it more adaptable to a wide range of scenarios. On the other hand, while dense trajectory guidance enables overfitting to observations, the lack of category-level regularization, unlike hand-crafted constitutive models, limits its generalization under scarce data. This leads to inferior performance than GIC, which can benefit from its pre-known constitutive model as an oracle. More analysis on the enhanced generalization ability of MASIV with more data is in Sec. C of Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.794, + 0.663, + 0.81 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.816, + 0.894, + 0.89 + ], + "angle": 0, + "content": "We investigate the impact of geometrical objectives here, while referring the readers to more ablation studies in Sec. B of Supp. Mat. Fig. 3 validates the effectiveness of dense trajectory supervision. As shown in Fig. 4, we evaluate three geometrical objectives by replacing the \\(\\mathcal{L}_{\\mathrm{geo}}\\) term in opti" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.106, + 0.885, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.254, + 0.897, + 0.296 + ], + "angle": 0, + "content": "Figure 3. Ablation study of geometrical objectives \\(\\mathcal{L}_{\\mathrm{geo}}\\) on PAC-NeRF dataset. We compare three types of geometrical objectives against a baseline without geometrical guidance and plot their simulation errors using Chamfer Distance (CD). Overall, dense trajectory guidance achieves the lowest and most stable errors. The figure is best viewed in color." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.304, + 0.485, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.363, + 0.234, + 0.376 + ], + "angle": 0, + "content": "(a) Surfaces" + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.363, + 0.44, + 0.375 + ], + "angle": 0, + "content": "(b) Continuums" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.385, + 0.486, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.444, + 0.255, + 0.456 + ], + "angle": 0, + "content": "(c) Gaussian Splats" + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.444, + 0.439, + 0.456 + ], + "angle": 0, + "content": "(d) Trajectories" + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.457, + 0.484, + 0.526 + ], + "angle": 0, + "content": "Figure 4. Visualization of geometrical objectives \\(\\mathcal{L}_{\\mathrm{geo}}\\) on PAC-NeRF Bird. The fine-tuned deformation network \\(\\mathcal{D}'\\) captures smooth and accurate motion for continuums, including low-frequency translation of the body and high-frequency swing of the tail. We space adjacent frames in the horizontal axis for clarity." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.544, + 0.484, + 0.603 + ], + "angle": 0, + "content": "mizing the neural constitutive model objective, Eq. (8). We define them formally as follows. (1) Surface: per-frame surface alignment between the simulated continuum and the filled continuum," + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.61, + 0.483, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s u r f}} = \\mathcal {L} _ {\\mathrm {C D}} \\left(\\left\\{\\mathcal {S} (\\hat {\\mathbf {x}} _ {t}) \\right\\} _ {t = 0} ^ {T}, \\left\\{\\mathcal {S} (\\mathbf {p} _ {t}) \\right\\} _ {t = 0} ^ {T}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.643, + 0.485, + 0.687 + ], + "angle": 0, + "content": "where \\(\\mathcal{S}\\) denotes the surface extraction by thresholding Gaussian opacities [6]. (2) Continuum: similar to Surface while between solid continuums instead of surfaces," + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.696, + 0.483, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n t}} = \\mathcal {L} _ {\\mathrm {C D}} \\left(\\left\\{\\hat {\\mathbf {x}} _ {t} \\right\\} _ {t = 0} ^ {T}, \\left\\{\\mathbf {p} _ {t} \\right\\} _ {t = 0} ^ {T}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.728, + 0.484, + 0.89 + ], + "angle": 0, + "content": "(3) Trajectory: per-simulation-step trajectory alignment between the simulated continuum \\(\\{\\hat{\\mathbf{x}}_t\\}_{t=0}^{N\\tau}\\) and the continuum predicted by the deformation network \\(\\{\\mathbf{x}_t\\}_{t=0}^{N\\tau}\\), as defined in Eq. (9). By analyzing per-frame errors, we first confirm the necessity of geometrical guidance, as the baseline (red curve) exhibits significantly higher CD errors across all material types. Surface-based supervision (yellow curve) shows moderate error reduction but remains less stable over time, especially in Plasticine and Sand. Continuum-based supervision (blue curve) further reduces errors but struggles with certain materials, e.g., Sand. These error overshoots" + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.304, + 0.896, + 0.321 + ], + "angle": 0, + "content": "Ground Truth Surface \\(\\mathcal{L}_{\\mathrm{surf}}\\) Continuum \\(\\mathcal{L}_{\\mathrm{cont}}\\) Trajectory \\(\\mathcal{L}_{\\mathrm{traj}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.322, + 0.877, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.587, + 0.895, + 0.655 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of geometrical objectives \\(\\mathcal{L}_{\\mathrm{geo}}\\) on PAC-NeRF Cream. Using sparse surfaces or solid continuum as geometrical cues can lead to physically implausible deformations, while trajectory guidance mitigates these issues by introducing temporally denser constraints. Better viewed when zoomed in." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.672, + 0.896, + 0.775 + ], + "angle": 0, + "content": "are probably due to supervising relatively complex neural constitutive models with sparse signals, which might lead to overfitting since the simulated behaviors between observed discrete time steps are not well constrained. Finally, by enforcing dense alignment across simulation steps, trajectory-based supervision (green curve) achieves the lowest and most stable errors across all materials." + }, + { + "type": "title", + "bbox": [ + 0.512, + 0.789, + 0.701, + 0.805 + ], + "angle": 0, + "content": "4.4. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.816, + 0.896, + 0.89 + ], + "angle": 0, + "content": "Dynamics reconstruction. We visualize the reconstructed Cream object by using the alternative objectives in Fig. 5. Here, we observe physically implausible deformations in results for surface and continuum supervision. Under uniform force, the object exhibits inconsistent distortion, espe" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5950" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.099, + 0.487, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.334, + 0.484, + 0.389 + ], + "angle": 0, + "content": "Figure 6. Intra-class generalization. We first optimize the geometry for PAC-NeRF Bird and then use the constitutive models optimized for reference elastic objects to predict novel dynamic behaviors, thereby softening or hardening the Bird respectively." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.415, + 0.484, + 0.636 + ], + "angle": 0, + "content": "cially in the boundary between blue stripes and red stripes. We attribute this to two main reasons. 1) Chamfer Distance based geometrical constraint focuses mainly on the overall structure. While it encourages alignment between simulated and reference shapes, they do not enforce point-wise correspondence. A single particle can thus be matched with different ground truth particles in different time steps. As a result, local mismatches can accumulate, leading to inaccurate deformations. 2) Sparse supervision fails to constrain intermediate motion dynamics. Surface and continuum objectives provide alignment only at discrete time steps, leaving large gaps in supervision. Without explicit constraints on how the object deforms over time, the neural constitutive model may interpolate arbitrarily, leading to unstable and unrealistic motion patterns." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.64, + 0.484, + 0.891 + ], + "angle": 0, + "content": "Dynamics generalization. We further study the generalization ability for MASIV. Fig. 6 examines how MASIV can be used to modify the dynamic characteristics of a single object (PAC-NeRF Bird). Specifically, we first reconstruct the geometrical representation of the Bird and optimize neural constitutive models for two elastic cross-shaped objects. These optimized constitutive models are then used as references to drive the simulation of the Bird, which deforms accordingly, exhibiting fluid-like softness on the left and increased rigidity and bounce on the right. Additionally, we see continuous transitions between soft and rigid behaviors when interpolating the neural constitutive model weights. In Fig. 7, we extend this study to inter-class generalization by applying neural constitutive models trained on different materials to a novel object, the PAC-NeRF Trophy. The results show that despite sharing the same initial state, each material model induces distinct deformation patterns. These" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.099, + 0.892, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.334, + 0.895, + 0.388 + ], + "angle": 0, + "content": "Figure 7. Inter-class generalization. We train multiple neural constitutive models with different materials and apply them to the PAC-NeRF Trophy. While sharing the same initial state, the Trophies with different materials exhibit distinct behaviors over time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.894, + 0.456 + ], + "angle": 0, + "content": "observations confirm that MASIV effectively captures generalizable intrinsic dynamics. More qualitative results can be found in Sec. D of Supp. Mat." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.464, + 0.628, + 0.478 + ], + "angle": 0, + "content": "4.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.485, + 0.895, + 0.662 + ], + "angle": 0, + "content": "While MASIV excels in material-agnostic system identification, several limitations remain. First, it relies on multiview video input, restricting applicability to single-view or sparse observations. Second, our approach assumes that internal geometries and deformations can be inferred from surface motion, following a spatially coherent model. This assumption may break down in materials with complex internal flows, such as granular media or multiphase fluids. Finally, MASIV does not explicitly disentangle material properties from external forces in system identification, assuming a uniform gravity field for all objects. Future works aim to relax these assumptions for broader applicability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.674, + 0.629, + 0.689 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.699, + 0.895, + 0.89 + ], + "angle": 0, + "content": "We present MASIV, a material-agnostic system identification framework. By integrating neural constitutive models with dense geometry guidance, MASIV effectively captures object dynamics across diverse materials. Our approach reconstructs dynamic Gaussians from multi-view videos and fine-tunes a deformation model within a solid continuum. This enables us to extract dense motion cues, which, combined with visual observations, guide the learning of neural constitutive models. Through this formulation, MASIV generalizes beyond traditional system identification methods that rely on pre-defined material priors. Extensive experiments highlight its ability to reconstruct observable dynamics and to generalize for novel interactions." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.514, + 0.939 + ], + "angle": 0, + "content": "5951" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.103, + 0.102, + 0.25, + 0.118 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.125, + 0.484, + 0.195 + ], + "angle": 0, + "content": "This work was supported in part by U.S. NSF grant DBI-2238093, and in part by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project)." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.22, + 0.196, + 0.236 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.244, + 0.484, + 0.326 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.327, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[2] Genesis Authors. Genesis: A universal and generative physics engine for robotics and beyond, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.355, + 0.484, + 0.448 + ], + "angle": 0, + "content": "[3] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B Lindell. 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7996-8006, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.45, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[4] Piotr Borycki, Weronika Smolak, Joanna Waczyńska, Marcin Mazur, Sławomir Tadeja, and Przemysław Spurek. Gasp: Gaussian splatting for physic-based simulations. arXiv preprint arXiv:2409.05819, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.505, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[5] Johannes Brandstetter, Daniel Worrall, and Max Welling. Message passing neural pde solvers. arXiv preprint arXiv:2202.03376, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.547, + 0.484, + 0.613 + ], + "angle": 0, + "content": "[6] Junhao Cai, Yuji Yang, Weihao Yuan, Yisheng He, Zilong Dong, Liefeng Bo, Hui Cheng, and Qifeng Chen. Gic: Gaussian-informed continuum for physical property identification and simulation. arXiv preprint arXiv:2406.14927, 2024. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.615, + 0.484, + 0.681 + ], + "angle": 0, + "content": "[7] Yuanhao Cai, Jiahao Wang, Alan Yuille, Zongwei Zhou, and Angtian Wang. Structure-aware sparse-view x-ray 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11174-11183, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.684, + 0.484, + 0.75 + ], + "angle": 0, + "content": "[8] Junyi Cao, Shanyan Guan, Yanhao Ge, Wei Li, Xiaokang Yang, and Chao Ma. Neuma: Neural material adaptor for visual grounding of intrinsic dynamics. Advances in Neural Information Processing Systems, 37:65643-65669, 2025. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.752, + 0.484, + 0.805 + ], + "angle": 0, + "content": "[9] Pradyumna Chari, Chinmay Talegaonkar, Yunhao Ba, and Achuta Kadambi. Visual physics: Discovering physical laws from videos. arXiv preprint arXiv:1911.11893, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.807, + 0.482, + 0.848 + ], + "angle": 0, + "content": "[10] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splattering guidance. arXiv preprint arXiv:2312.00846, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.849, + 0.484, + 0.89 + ], + "angle": 0, + "content": "[11] Hsiao-yu Chen, Edith Tretschk, Tuur Stuyck, Petr Kadlecek, Ladislav Kavan, Etienne Vouga, and Christoph Lassner. Virtual elastic objects. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.111, + 0.244, + 0.484, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.104, + 0.893, + 0.131 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 15827-15837, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.132, + 0.894, + 0.185 + ], + "angle": 0, + "content": "[12] Tao Du, Kui Wu, Pingchuan Ma, Sebastien Wah, Andrew Spielberg, Daniela Rus, and Wojciech Matusik. Diffpd: Differentiable projective dynamics. ACM Transactions on Graphics (TOG), 41(2):1-21, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.186, + 0.895, + 0.24 + ], + "angle": 0, + "content": "[13] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In European Conference on Computer Vision, pages 108-124. Springer, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.24, + 0.895, + 0.307 + ], + "angle": 0, + "content": "[14] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, et al. Gaussian splashing: Unified particles for versatile motion synthesis and rendering. arXiv preprint arXiv:2401.15318, 3, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.308, + 0.895, + 0.375 + ], + "angle": 0, + "content": "[15] Yutao Feng, Yintong Shang, Xuan Li, Tianjia Shao, Chenfanfu Jiang, and Yin Yang. Pie-nerf: Physics-based interactive elastodynamics with nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4450–4461, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.376, + 0.894, + 0.442 + ], + "angle": 0, + "content": "[16] Zhoujie Fu, Jiacheng Wei, Wenhao Shen, Chaoyue Song, Xiaofeng Yang, Fayao Liu, Xulei Yang, and Guosheng Lin. Sync4d: Video guided controllable dynamics for physics-based 4d generation. arXiv preprint arXiv:2405.16849, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.443, + 0.894, + 0.51 + ], + "angle": 0, + "content": "[17] Moritz Geilinger, David Hahn, Jonas Zehnder, Moritz Bächer, Bernhard Thomaszewski, and Stelian Coros. Add: Analytically differentiable dynamics for multi-body systems with frictional contact. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.511, + 0.894, + 0.579 + ], + "angle": 0, + "content": "[18] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, et al. Motion prompting: Controlling video generation with motion trajectories. arXiv preprint arXiv:2412.02700, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.579, + 0.894, + 0.645 + ], + "angle": 0, + "content": "[19] Sharath Girish, Tianye Li, Amrita Mazumdar, Abhinav Shrivastava, David Luebke, and Shalini De Mello. Queen: Quantized efficient encoding of dynamic gaussians for streaming free-viewpoint videos. arXiv preprint arXiv:2412.04469, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.646, + 0.894, + 0.699 + ], + "angle": 0, + "content": "[20] Shanyan Guan, Huayu Deng, Yunbo Wang, and Xiaokang Yang. Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields. In International conference on machine learning, pages 7919-7929. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.7, + 0.894, + 0.755 + ], + "angle": 0, + "content": "[21] Eric Heiden, Miles Macklin, Yashraj Narang, Dieter Fox, Animesh Garg, and Fabio Ramos. Disect: A differentiable simulation engine for autonomous robotic cutting. arXiv preprint arXiv:2105.12244, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.755, + 0.894, + 0.821 + ], + "angle": 0, + "content": "[22] Maximilian Herde, Bogdan Raonic, Tobias Rohner, Roger Käppeli, Roberto Molinaro, Emmanuel de Bézenac, and Siddhartha Mishra. Poseidon: Efficient foundation models for pdes. Advances in Neural Information Processing Systems, 37:72525-72624, 2025. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.522, + 0.822, + 0.894, + 0.89 + ], + "angle": 0, + "content": "[23] Florian Hofherr, Lukas Koestler, Florian Bernard, and Daniel Cremers. Neural implicit representations for physical parameter inference from a single video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2093-2103, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.104, + 0.895, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.928, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5952" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.104, + 0.484, + 0.144 + ], + "angle": 0, + "content": "[24] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.146, + 0.483, + 0.198 + ], + "angle": 0, + "content": "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.2, + 0.483, + 0.267 + ], + "angle": 0, + "content": "[26] Yuanming Hu, Yu Fang, Ziheng Ge, Ziyin Qu, Yixin Zhu, Andre Pradhana, and Chenfanfu Jiang. A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.268, + 0.483, + 0.321 + ], + "angle": 0, + "content": "[27] Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Fredo Durand. Difftaichi: Differentiable programming for physical simulation. arXiv preprint arXiv:1910.00935, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.322, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[28] Tianyu Huang, Haoze Zhang, Yihan Zeng, Zhilu Zhang, Hui Li, Wangmeng Zuo, and Rynson WH Lau. Dreamphysics: Learning physical properties of dynamic 3d gaussians with video diffusion priors. arXiv preprint arXiv:2406.01476, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.389, + 0.483, + 0.443 + ], + "angle": 0, + "content": "[29] Zhiao Huang, Yuanming Hu, Tao Du, Siyuan Zhou, Hao Su, Joshua B Tenenbaum, and Chuang Gan. Plasticinelab: A soft-body manipulation benchmark with differentiable physics. arXiv preprint arXiv:2104.03311, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.444, + 0.483, + 0.524 + ], + "angle": 0, + "content": "[30] Krishna Murthy Jatavallabhula, Miles Macklin, Florian Golemo, Vikram Voleti, Linda Petrini, Martin Weiss, Breandan Considine, Jérôme Parent-Lévesque, Kevin Xie, Kenny Erleben, et al. gradsim: Differentiable simulation for system identification and visuomotor control. arXiv preprint arXiv:2104.02646, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.525, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[31] Chenfanfu Jiang, Craig Schroeder, Joseph Teran, Alexey Stomakhin, and Andrew Selle. The material point method for simulating continuum materials. In ACM SIGGRAPH 2016 Courses, pages 1-52, 2016. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.58, + 0.483, + 0.646 + ], + "angle": 0, + "content": "[32] Takuhiro Kaneko. Improving physics-augmented continuum neural radiance field-based geometry-agnostic system identification with lagrangian particle optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5480, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.647, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[33] Namgyu Kang, Jaemin Oh, Youngjoon Hong, and Eunbyung Park. Pig: Physics-informed gaussians as adaptive parametric mesh representations. arXiv preprint arXiv:2412.05994, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.701, + 0.483, + 0.753 + ], + "angle": 0, + "content": "[34] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.755, + 0.483, + 0.822 + ], + "angle": 0, + "content": "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.823, + 0.483, + 0.889 + ], + "angle": 0, + "content": "[36] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. In European Conference on Computer Vision, pages 252-269. Springer, 2025. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.111, + 0.104, + 0.484, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.104, + 0.894, + 0.157 + ], + "angle": 0, + "content": "[37] Jinxi Li, Ziyang Song, and Bo Yang. Nvfi: neural velocity fields for 3d physics learning from dynamic videos. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.159, + 0.894, + 0.238 + ], + "angle": 0, + "content": "[38] Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Krishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang, and Chuang Gan. Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. arXiv preprint arXiv:2303.05512, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.24, + 0.894, + 0.293 + ], + "angle": 0, + "content": "[39] Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B Tenenbaum, and Antonio Torralba. Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. arXiv preprint arXiv:1810.01566, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.295, + 0.894, + 0.347 + ], + "angle": 0, + "content": "[40] Yifei Li, Tao Du, Kui Wu, Jie Xu, and Wojciech Matusik. Diffcloth: Differentiable cloth simulation with dry frictional contact. ACM Transactions on Graphics (TOG), 42 (1):1-20, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.349, + 0.894, + 0.415 + ], + "angle": 0, + "content": "[41] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.417, + 0.894, + 0.482 + ], + "angle": 0, + "content": "[42] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.484, + 0.894, + 0.537 + ], + "angle": 0, + "content": "[43] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.538, + 0.894, + 0.605 + ], + "angle": 0, + "content": "[44] Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Azizzadenesheli, and Anima Anandkumar. Physics-informed neural operator for learning partial differential equations. ACM/JMS Journal of Data Science, 1(3):1-27, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.606, + 0.894, + 0.646 + ], + "angle": 0, + "content": "[45] Junbang Liang, Ming Lin, and Vladlen Koltun. Differentiable cloth simulation for inverse problems. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.647, + 0.894, + 0.727 + ], + "angle": 0, + "content": "[46] Xiao Liang, Fei Liu, Yutong Zhang, Yuelei Li, Shan Lin, and Michael Yip. Real-to-sim deformable object manipulation: Optimizing physics models with residual mappings for robotic surgery. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 15471-15477. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.728, + 0.894, + 0.807 + ], + "angle": 0, + "content": "[47] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.809, + 0.894, + 0.862 + ], + "angle": 0, + "content": "[48] Yuchen Lin, Chenguo Lin, Jianjin Xu, and Yadong Mu. Omniphysgs: 3d constitutive gaussians for general physics-based dynamics generation. arXiv preprint arXiv:2501.18982, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.863, + 0.894, + 0.89 + ], + "angle": 0, + "content": "[49] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.104, + 0.894, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.929, + 0.514, + 0.939 + ], + "angle": 0, + "content": "5953" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.141, + 0.104, + 0.484, + 0.144 + ], + "angle": 0, + "content": "with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576-8588, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.146, + 0.483, + 0.211 + ], + "angle": 0, + "content": "[50] Phillip Lippe, Bas Veeling, Paris Perdikaris, Richard Turner, and Johannes Brandstetter. Pde-refiner: Achieving accurate long rollouts with neural pde solvers. Advances in Neural Information Processing Systems, 36:67398-67433, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.215, + 0.483, + 0.268 + ], + "angle": 0, + "content": "[51] Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. Physics3d: Learning physical properties of 3d gaussians via video diffusion. arXiv preprint arXiv:2406.04338, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.271, + 0.482, + 0.323 + ], + "angle": 0, + "content": "[52] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond. arXiv preprint arXiv:1908.03265, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.326, + 0.483, + 0.379 + ], + "angle": 0, + "content": "[53] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splatting from causally-captured monocular videos. arXiv preprint arXiv:2406.00434, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.381, + 0.483, + 0.435 + ], + "angle": 0, + "content": "[54] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imageto-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.437, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[55] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.505, + 0.483, + 0.558 + ], + "angle": 0, + "content": "[56] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.561, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[57] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.616, + 0.483, + 0.682 + ], + "angle": 0, + "content": "[58] Pingchuan Ma, Tao Du, Joshua B Tenenbaum, Wojciech Matusik, and Chuang Gan. Risp: Rendering-invariant state predictor with differentiable simulation and rendering for cross-domain parameter estimation. arXiv preprint arXiv:2205.05678, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.685, + 0.483, + 0.764 + ], + "angle": 0, + "content": "[59] Pingchuan Ma, Peter Yichen Chen, Bolei Deng, Joshua B Tenenbaum, Tao Du, Chuang Gan, and Wojciech Matusik. Learning neural constitutive laws from motion observations for generalizable pde dynamics. In International Conference on Machine Learning, pages 23279-23300. PMLR, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.767, + 0.483, + 0.833 + ], + "angle": 0, + "content": "[60] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.112, + 0.836, + 0.483, + 0.889 + ], + "angle": 0, + "content": "[61] Clinton A Mo, Kun Hu, Chengjiang Long, and Zhiyong Wang. Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In Proceedings of the IEEE/CVF Conference on" + }, + { + "type": "list", + "bbox": [ + 0.111, + 0.104, + 0.484, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.104, + 0.894, + 0.129 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition, pages 13894-13903, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.132, + 0.894, + 0.198 + ], + "angle": 0, + "content": "[62] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.2, + 0.894, + 0.266 + ], + "angle": 0, + "content": "[63] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.268, + 0.894, + 0.332 + ], + "angle": 0, + "content": "[64] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.334, + 0.894, + 0.387 + ], + "angle": 0, + "content": "[65] Yiling Qiao, Junbang Liang, Vladlen Koltun, and Ming Lin. Differentiable simulation of soft multi-body systems. Advances in Neural Information Processing Systems, 34: 17123-17135, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.389, + 0.893, + 0.441 + ], + "angle": 0, + "content": "[66] Yi-Ling Qiao, Alexander Gao, and Ming Lin. Neuphysics: Editable neural geometry and physics from monocular videos. Advances in Neural Information Processing Systems, 35:12841-12854, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.444, + 0.894, + 0.496 + ], + "angle": 0, + "content": "[67] Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Language-driven physics-based scene synthesis and editing via feature splatting. In European Conference on Computer Vision, pages 368-383. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.498, + 0.893, + 0.563 + ], + "angle": 0, + "content": "[68] Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378:686-707, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.566, + 0.893, + 0.617 + ], + "angle": 0, + "content": "[69] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.619, + 0.893, + 0.685 + ], + "angle": 0, + "content": "[70] Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In International conference on machine learning, pages 8459-8468. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.687, + 0.893, + 0.739 + ], + "angle": 0, + "content": "[71] Yidi Shao, Mu Huang, Chen Change Loy, and Bo Dai. Gausim: Registering elastic objects into digital world by gaussian simulator. arXiv preprint arXiv:2412.17804, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.741, + 0.894, + 0.793 + ], + "angle": 0, + "content": "[72] Haochen Shi, Huazhe Xu, Samuel Clarke, Yunzhu Li, and Jiajun Wu. Robocook: Long-horizon elasto-plastic object manipulation with diverse tools. arXiv preprint arXiv:2306.14447, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.796, + 0.894, + 0.86 + ], + "angle": 0, + "content": "[73] Haochen Shi, Huazhe Xu, Zhiao Huang, Yunzhu Li, and Ji-jun Wu. Robocraft: Learning to see, simulate, and shape elasto-plastic objects in 3d with graph networks. The International Journal of Robotics Research, 43(4):533-549, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.862, + 0.894, + 0.89 + ], + "angle": 0, + "content": "[74] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.104, + 0.894, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.929, + 0.515, + 0.939 + ], + "angle": 0, + "content": "5954" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.141, + 0.104, + 0.484, + 0.158 + ], + "angle": 0, + "content": "Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.159, + 0.483, + 0.223 + ], + "angle": 0, + "content": "[75] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.226, + 0.483, + 0.293 + ], + "angle": 0, + "content": "[76] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.295, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[77] Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.349, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[78] Deborah Sulsky, Shi-Jian Zhou, and Howard L Schreyer. Application of a particle-in-cell method to solid mechanics. Computer physics communications, 87(1-2):236-252, 1995. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.403, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[79] Priya Sundaresan, Rika Antonova, and Jeannette Bohgl. Diffcloud: Real-to-sim from point clouds with differentiable simulation and rendering of deformable objects. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10828-10835. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.484, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[80] Tetsuya Takahashi and Ming C Lin. Video-guided real-to-virtual parameter transfer for viscous fluids. ACM Transactions on Graphics (TOG), 38(6):1-12, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.525, + 0.482, + 0.577 + ], + "angle": 0, + "content": "[81] Xiyang Tan, Ying Jiang, Xuan Li, Zeshun Zong, Tianyi Xie, Yin Yang, and Chenfanfu Jiang. Physmotion: Physics-grounded dynamics from a single image. arXiv preprint arXiv:2411.17189, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.579, + 0.482, + 0.619 + ], + "angle": 0, + "content": "[82] Bin Wang, Longhua Wu, KangKang Yin, Uri M Ascher, Libin Liu, and Hui Huang. Deformation capture and modeling of soft objects. ACM Trans. Graph., 34(4):94-1, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.621, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[83] Chaoyang Wang, Lachlan Ewen MacDonald, Laszlo A Jeni, and Simon Lucey. Flow supervision for deformable nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21128-21137, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.688, + 0.482, + 0.74 + ], + "angle": 0, + "content": "[84] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.742, + 0.482, + 0.808 + ], + "angle": 0, + "content": "[85] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.81, + 0.482, + 0.889 + ], + "angle": 0, + "content": "[86] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splattering for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.111, + 0.104, + 0.484, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.104, + 0.894, + 0.171 + ], + "angle": 0, + "content": "[87] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.173, + 0.894, + 0.238 + ], + "angle": 0, + "content": "[88] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9421-9431, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.24, + 0.894, + 0.307 + ], + "angle": 0, + "content": "[89] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.308, + 0.893, + 0.361 + ], + "angle": 0, + "content": "[90] Zhenjia Xu, Jiajun Wu, Andy Zeng, Joshua B Tenenbaum, and Shuran Song. Densephysnet: Learning dense physical object representations via multi-step dynamic interactions. arXiv preprint arXiv:1906.03853, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.362, + 0.893, + 0.429 + ], + "angle": 0, + "content": "[91] Tianju Xue, Shuheng Liao, Zhengtao Gan, Chanwook Park, Xiaoyu Xie, Wing Kam Liu, and Jian Cao. Jax-fem: A differentiablegpu-accelerated 3d finite element solver for automatic inverse design and mechanistic data science. Computer Physics Communications, 291:108802, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.43, + 0.894, + 0.482 + ], + "angle": 0, + "content": "[92] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.484, + 0.893, + 0.563 + ], + "angle": 0, + "content": "[93] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.565, + 0.893, + 0.632 + ], + "angle": 0, + "content": "[94] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13117-13125, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.633, + 0.893, + 0.699 + ], + "angle": 0, + "content": "[95] Heng Yu, Joel Julin, Zoltan A Milacski, Koichiro Niinuma, and László A Jeni. Dylin: Making light field networks dynamic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12397-12406, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.7, + 0.893, + 0.766 + ], + "angle": 0, + "content": "[96] Heng Yu, Joel Julin, Zoltán Á Milacski, Koichiro Niinuma, and László A Jeni. Cogs: Controllable gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21624-21633, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.768, + 0.893, + 0.808 + ], + "angle": 0, + "content": "[97] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.81, + 0.893, + 0.889 + ], + "angle": 0, + "content": "[98] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.104, + 0.894, + 0.889 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.929, + 0.514, + 0.939 + ], + "angle": 0, + "content": "5955" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.104, + 0.482, + 0.158 + ], + "angle": 0, + "content": "[99] Dongzhe Zheng, Siqiong Yao, Wenqiang Xu, and Cewu Lu. Differentiable cloth parameter identification and state estimation in manipulation. IEEE Robotics and Automation Letters, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.16, + 0.483, + 0.225 + ], + "angle": 0, + "content": "[100] Licheng Zhong, Hong-Xing Yu, Jiajun Wu, and Yunzhu Li. Reconstruction and simulation of elastic objects with spring-mass 3d gaussians. In European Conference on Computer Vision, pages 407-423. Springer, 2025. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.105, + 0.228, + 0.483, + 0.281 + ], + "angle": 0, + "content": "[101] Xiaowei Zhou, Can Yang, and Weichuan Yu. Moving object detection by detecting contiguous outliers in the low-rank representation. IEEE transactions on pattern analysis and machine intelligence, 35(3):597-610, 2012. 2" + }, + { + "type": "list", + "bbox": [ + 0.105, + 0.104, + 0.483, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.929, + 0.516, + 0.939 + ], + "angle": 0, + "content": "5956" + } + ] +] \ No newline at end of file diff --git a/2025/Toward Material-Agnostic System Identification from Videos/58fcfc32-a696-4a6d-a6cd-3c218994c3ba_origin.pdf b/2025/Toward Material-Agnostic System Identification from Videos/58fcfc32-a696-4a6d-a6cd-3c218994c3ba_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ab5011e9113570e92891caf32d3830951d85bc87 --- /dev/null +++ b/2025/Toward Material-Agnostic System Identification from Videos/58fcfc32-a696-4a6d-a6cd-3c218994c3ba_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:774ac814a632e6bd65898faa17bf98dddf0a8022bb1488aed120b1cf5aeff606 +size 2166305 diff --git a/2025/Toward Material-Agnostic System Identification from Videos/full.md b/2025/Toward Material-Agnostic System Identification from Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4722a293306222fbb2c40d836cbda6878560d74a --- /dev/null +++ b/2025/Toward Material-Agnostic System Identification from Videos/full.md @@ -0,0 +1,365 @@ +# Toward Material-Agnostic System Identification from Videos + +Yizhou Zhao1 Haoyu Chen1 Chunjiang Liu1 Zhenyang Li2 Charles Herrmann3 + +Junhwa $\mathsf{Hur}^3$ Yinxiao Li3 Ming-Hsuan Yang3,4 Bhiksha Raj1 Min $\mathbf{X}\mathbf{u}^{1*}$ + +$^{1}$ Carnegie Mellon University $^{2}$ University of Alabama at Birmingham $^{3}$ Google $^{4}$ UC Merced + +https://github.com/Skaldak/MASIV + +![](images/54aeab1a885dda3cf57aa77b1e3cc10ce1e8177a2e18b1495742ca14509692bc.jpg) +Figure 1. System identification from video observations seeks to understand the intrinsic dynamics of a given scene. (a) Existing methods rely on scene-specific material priors. (b) Our MASIV removes this requirement, grounding object motion in a material-agnostic manner. + +# Abstract + +System identification from videos aims to recover object geometry and governing physical laws. Existing methods integrate differentiable rendering with simulation but rely on predefined material priors, limiting their ability to handle unknown ones. We introduce MASIV, the first vision-based framework for material-agnostic system identification. Unlike existing approaches that depend on hand-crafted constitutive laws, MASIV employs learnable neural constitutive models, inferring object dynamics without assuming a scene-specific material prior. However, the absence of full particle state information imposes unique challenges, leading to unstable optimization and physically implausible behaviors. To address this, we introduce dense geometric guidance by reconstructing continuum particle trajectory + +ries, providing temporally rich motion constraints beyond sparse visual cues. Comprehensive experiments show that MASIV achieves state-of-the-art performance in geometric accuracy, rendering quality, and generalization ability. + +# 1. Introduction + +From the flow of water down a stream to the deformation of a rubber ball upon impact, the physical world operates with a set of underlying principles that govern motion and interaction. Humans intuitively grasp these principles through observation, allowing us to see a scene and instinctively predict its outcome or re-imagine it with different initial dynamics. Material identification forms the basis of this ability, requiring us to determine an object's composition and understand how that influences its response to forces. This capability is fundamental to world modeling, the process of constructing internal representations that help us understand + +and anticipate the evolution of the environment. However, replicating the ability to generalize physical laws to diverse scenarios remains a challenge for artificial intelligence. + +To bridge this gap, researchers have explored vision-based system identification [6, 8, 9, 20, 23, 30, 38], which aims to infer governing physical laws from visual observations. Typically, they optimize a parameterized physical model, by end-to-end integrating differentiable renderers, such as Neural Radiance Fields (NeRF) [60] and 3D Gaussian Splatting (3DGS) [34], with differentiable simulators [26, 27, 31, 78, 91]. To enforce physically meaningful constraints on the simulation, existing approaches [6, 38, 100] often rely on material-specific constitutive laws, such as elasticity, plasticity, or viscosity models. These laws define how an object's material properties influence its response to external forces, allowing system identification methods to estimate parameters such as stiffness, damping, or friction. While effective in controlled settings, this dependence on hand-crafted material models limits adaptability. Moreover, these methods assume knowing the material type for a given scene, making it difficult for them to generalize across in-the-wild scenarios where material properties are unknown. + +Directly targeting these limitations, we introduce Material-Agnostic System Identification from Videos (MASIV), a vision-based approach that infers object dynamics free of predefined material priors. As illustrated in Fig. 1, rather than relying on expert-designed elasticity and plasticity models, MASIV employs fully learnable neural constitutive models inspired by NCLaw [59]. However, directly applying neural constitutive models to vision-based system identification poses significant challenges. Similar to the learning process of neural PDE solvers [5, 22, 44, 50], the original formulation of NCLaw assumes access to complete state observations of simulated particles, including position, velocity, deformation gradient, and affine momentum. Since visual observation fails to provide such complete information, simply relying on frame-wise pixel supervision provides insufficient constraints. This leads to unstable optimization and poor convergence [48]. + +We combine two main insights to enable this optimization. First, despite the lack of ground truth full particle state, frame-wise observations still allow reasonable estimates for some of them, such as position and velocity. Existing dynamic reconstruction methods [42, 55, 56] factorize moving objects with their canonical representation and corresponding deformations, which enables per-frame position tracking. Further, the commonly adopted motion basis [61, 94, 101] provides a low-rank encoding of object motion, facilitating state interpolation between frames. This formulation allows us to reconstruct particle trajectories over dense time steps. Second, although trajectory estimations are inherently limited to visible object regions, + +internal deformations can be inferred by propagating surface motions inward. Our intuition is that nearby particles inside an object should be deformed similarly, akin to the As-Rigid-As-Possible (ARAP) prior. + +We use these insights to guide the design of our framework. Concretely, we reconstruct dynamic Gaussians from multi-view video inputs, simultaneously establishing geometry and deformation models of the scene. The geometry model is then transformed into a solid continuum, where the deformation model is fine-tuned to encode both the exterior and the interior motion. To identify intrinsic dynamics, we guide the learning of the neural constitutive model using two complementary cues: appearance cues from video frames and geometric constraints from the fine-tuned deformation model. In this way, MASIV distills dense particle motion from sparse visual observations and guides the learning of neural constitutive models for stable, material-agnostic system identification. + +Our contributions can be summed up as follows: + +- We present a novel framework, MASIV, that determines the intrinsic dynamics of an object from videos in a material-agnostic manner. In contrast, all other current vision-based system identification methods [6, 8, 38] necessitate predefined material priors. +- We introduce temporally dense geometry guidance to constrain object behaviors with neural constitutive models, facilitating smoother convergence and improved performance for material-agnostic system identification. +- Through extensive qualitative and quantitative experiments, we demonstrate that MASIV achieves superior performance in both observable state recovery and future state prediction, even without explicit material priors. + +# 2. Related Work + +Dynamic reconstruction is a fundamental task in computer vision, aimed at reconstructing high-fidelity representations of complex dynamic scenes from diverse inputs such as monocular and multi-view videos [1, 7, 19, 41, 53, 76, 83, 88, 92, 95, 96]. Earlier research primarily focused on augmenting canonical Neural Radiance Fields (NeRF) [60] with a deformation field [64], which were later enhanced by incorporating volume-preserving regularization techniques [62, 63]. While primarily focused on the novel view synthesis task, their neural implicit representations often introduce noisy deformations, limiting the accuracy of recovered geometries required for reliable physics-based property estimation [38]. With the explicit representation of the scene using Gaussian ellipsoids, 3D Gaussian Splatting (3DGS) [34] achieves efficient dynamic scene reconstruction. Building on the principles of 3DGS, several studies have extended its application to 4D dynamic reconstruction by handling each frame separately [56] or mod + +eling the entire scene with a canonical 3D Gaussian point cloud coupled with a robust deformation framework that accurately maps it to the target scene [36, 86, 93]. + +Dynamic simulation integrates physics principles into 3D representations to generate realistic motion and interactions. A typical approach to synthesizing dynamic 3D scenes involves integrating a 3D generation pipeline with a video generation model [3, 49, 69, 75]. Recent research incorporates 3D Gaussian kernels as both visual and physical representations, embedding Newtonian dynamics into the Gaussian framework to enable seamless rendering and simulation while introducing physical constraints to ensure plausible dynamics [4, 16, 38, 48, 51, 67, 89, 100]. Building on this, Gaussian Splashing [14] introduces a position-based dynamics framework within 3D Gaussian Splating, enabling interactions among solids, fluids, and deformable objects. Additionally, motion-conditioned simulation methods [18, 43, 74, 85, 87] use trajectory-based guidance to predict object motion, showcasing how generative models can enhance dynamic scene synthesis. However, these methods primarily rely on pre-trained generative models, which often lack a fundamental understanding of physics. More recent works leverage neural physics modeling, incorporating video priors and generative models to infer object dynamics, allowing physics-driven motion synthesis without explicit solvers [15, 28, 54, 81, 98]. However, these methods are constrained to specific material types, limiting their generalization across diverse objects. To address this, we introduce a material-agnostic approach that unifies system identification across different object types, ensuring broader applicability beyond predefined materials. + +System identification aims to understand the physical laws governing the 3D world and is an essential task for simulation [37, 40, 45, 68, 79, 100] and robotic manipulation [46, 66, 72, 73, 99]. However, accurately recovering both object geometries and physical properties remains a significant challenge. Existing methods often assume known geometries and rely on finite element methods (FEM) or mass-spring systems for dynamic simulation [80, 82], limiting their ability to handle complex, real-world materials. Neural network-based methods [39, 70, 90] offer greater flexibility by leveraging data-driven models for system identification, but they struggle to generalize across diverse material properties and environmental conditions. More recently, differentiable physics-based simulations have emerged as a promising direction [11, 12, 17, 21, 27, 29, 30, 32, 58, 65]. However, these methods often require intricate modeling to bridge the gap between simulated and real-world behavior. To improve adaptability, some approaches focus on neural constitutive modeling, embedding learned material properties within physical simulations to refine expert-designed models [8, 59]. Other methods leverage the Material Point Method (MPM) + +# Algorithm 1 MPM state transfer $\mathcal{M}$ + +Input: $\mathbf{s}_n = \{\mathbf{x}_n,\mathbf{v}_n,\mathbf{F}_n^e\}$ + +Output: $\mathbf{s}_{n + 1} = \{\mathbf{x}_{n + 1},\mathbf{v}_{n + 1},\mathbf{F}_{n + 1}^{e}\}$ + +1: for $i = 1 \dots Q$ , $\mathbf{P}_n(i) = \mathcal{E}(\mathbf{F}_n^e (i))$ +2: $\mathbf{x}_{n + 1},\mathbf{v}_{n + 1},\mathbf{F}_{n + 1}^{e,\mathrm{trial}} = \mathcal{I}(\mathbf{x}_n,\mathbf{v}_n,\mathbf{F}_n^e,\mathbf{P}_n)$ +3: for $i = 1 \dots Q$ , $\mathbf{F}_{n + 1}^{e}(i) = \mathcal{P}(\mathbf{F}_{n + 1}^{e,\mathrm{trial}}(i))$ + +and 3D Gaussian-based representations to enhance geometry reconstruction and material property estimation [6, 38, 71, 100]. Building upon GIC [6], our method presents a material-agnostic approach that eliminates reliance on predefined material priors. Rather than assuming material properties, our model infers object dynamics directly from visual observations, making it adapt better to a wide range of unknown or heterogeneous materials. + +# 3. Method + +Given visual observations $\mathbf{I} = \{\mathbf{I}_0, \mathbf{I}_1, \dots, \mathbf{I}_T\}$ over $T$ frames capturing a dynamic object, our goal is to recover its explicit geometric representation and the governing evolution rules of the underlying dynamical system. Formally, we represent the object's geometry using a set of Gaussian kernels [34] and integrate neural constitutive laws [59] with the material point method (MPM) [26, 31, 78] to roll out their motion over time. Unlike [6, 8, 38, 100], we do not assume a known material type of the object, while exploring Material-Agnostic System Identification from Videos (MASIV). Fig. 2 shows an overview of our pipeline. + +# 3.1. Preliminaries + +Material Point Method (MPM) discretizes the object of interest into $Q$ material points and models their evolution over time via a state transfer equation + +$$ +\mathbf {s} _ {n + 1} = \mathcal {M} (\mathbf {s} _ {n}), \forall n = 0, 1, \dots , N. \tag {1} +$$ + +$\mathbf{s}_n$ denotes the particle state at time step $n$ . $N$ is the total number of simulation steps with the simulation interval $\tau$ and observed $T$ frames, where $N = T / \tau \gg T$ . + +The transition process $\mathcal{M}$ follows a time-stepping scheme, as outlined in Algorithm 1. The state of particle $i$ at time step $n$ is described by its position $\mathbf{x}_n(i)$ , velocity $\mathbf{v}_n(i)$ , and the elastic component $\mathbf{F}_n^e (i)$ of its deformation gradient $\mathbf{F}_n(i)$ . At each time step, the elastic constitutive law $\mathcal{E}$ computes the first Piola-Kirchhoff stress $\mathbf{P}_n(i)$ from the elastic deformation gradient $\mathbf{F}_n^e (i)$ . Subsequently, the time integration scheme $\mathcal{I}$ updates all particle states. The updated trial elastic deformation gradient, denoted as $\mathbf{F}_{n + 1}^{e,\mathrm{trial}}(i)$ , needs to be further corrected by the plastic constitutive law $\mathcal{P}$ , which applies plasticity constraints and yields the final elastic deformation gradient $\mathbf{F}_{n + 1}^{e}(i)$ . More details of MPM can be found in Sec. A.2 of Supp. Mat. + +![](images/01d83b78a20996eb69288d40cd33267e5f5a99259e915331ce3d41a5b5aaba90.jpg) + +![](images/0578d2a5e614b23dae067b60d80ca85803c40286b31c8cc2d057d7b89abe5c2f.jpg) +Figure 2. The overview of MASIV. Our pipeline comprises three phases. Phase I, Geometrical Reconstruction, infers geometrical representations (Sec. 3.2) and dense trajectories (Sec. 3.3) for the captured object. Phase II, Material-Agnostic System Identification (Sec. 3.4), grounds dynamic characteristics by exploiting visual observations with reconstructed motion clues. From this, we obtain a generalizable digital twin in Phase III, which can simulate novel interactions, such as new velocities or forces. + +Existing efforts [6, 8, 38, 100] for video-based system identification commonly assume known constitutive laws $\mathcal{E}$ and $\mathcal{P}$ , with a particular backbone form, e.g., neo-Hookean elasticity with identity plasticity return function. This simplifies the problem by estimating a small number of physical properties, such as Young's modulus, fluid viscosity, friction angles, etc., at the cost of sacrificing flexibility. In contrast, we explore the same problem without this material-specific assumption. + +# 3.2. Dynamic Gaussian Reconstruction + +We start by reconstructing dynamic Gaussians over $T$ frames. We consider this reconstruction as temporally sparse because the number of simulation steps is hundreds of times more than the number of observed video frames $T$ , i.e., $N = T / \tau \gg T$ . Following [6, 36, 86, 93], we maintain a set of canonical Gaussians $\mathcal{G}^* = \{\pmb{\mu}^*, \pmb{\Sigma}^*, \mathbf{c}^*, \sigma^*\}$ and use a neural network $\mathcal{D}$ to warp them over time to represent observed frames. Specifically, a basis network maps a given time step $t$ to $B$ bases, producing deformation basis for position $\mathbf{B}^{\mu}(t) \in \mathbb{R}^{B \times 3}$ and scale $\mathbf{B}^s(t) \in \mathbb{R}^{B \times 1}$ . A coefficient network then estimates the weight of each basis, $\mathbf{w}(\pmb{\mu}^*, t) \in \mathbb{R}^B$ , from the center coordinates of canonical Gaussian kernels $\pmb{\mu}^*$ and the time step $t$ . This results in deformed Gaussian kernels at each time step $t$ : + +$$ +\boldsymbol {\mu} _ {t} = \boldsymbol {\mu} ^ {*} + \mathbf {w} (\boldsymbol {\mu} ^ {*}, t) \mathbf {B} ^ {\boldsymbol {\mu}} (t), \tag {2} +$$ + +$$ +s _ {t} = s ^ {*} + \mathbf {w} \left(\boldsymbol {\mu} ^ {*}, t\right) \mathbf {B} ^ {s} (t). \tag {3} +$$ + +We employ the same optimization strategy in [6, 93] to minimize the L1 and Structural Similarity Index Mea + +sure (SSIM) losses between rendered frames $\hat{\mathbf{I}}_t$ and corresponding ground-truth observations $\mathbf{I}_t$ . We also regard all Gaussian kernels as isotropic and apply L1 regularization on scales for simplicity and reconstruction quality [10, 97, 100]. The overall objective function is formulated as + +$$ +\min _ {\mathcal {G} ^ {*}, \mathcal {D}} \left[ \mathcal {L} _ {1} (\hat {\mathbf {I}} _ {t}, \mathbf {I} _ {t}) + \lambda_ {\mathrm {S S I M}} \mathcal {L} _ {\mathrm {S S I M}} (\hat {\mathbf {I}} _ {t}, \mathbf {I} _ {t}) + \lambda_ {s} \| s _ {t} \| _ {1} \right], \tag {4} +$$ + +with hyperparameters $\lambda_{\mathrm{SSIM}}$ and $\lambda_{s}$ + +# 3.3. Continuum Trajectory Estimation + +To bridge the gap between unevenly distributed Gaussian particles and the corresponding solid continuum, we follow [6] to fill internal volumes, forming continuum particles $\{\mathbf{p}_t\}$ with uniform scale $s$ and density $\sigma$ . Existing methods for material-specific system identification often rely on sparse visual [6, 8, 38] or geometrical [6] cues along $T$ frames to optimize a small set of parameters. However, since we do not assume a known constitutive backbone or material types, these sparse signals may lead to unconstrained behaviors at unobserved time steps or even cause training instability. Inspired by system identification [22, 33, 59] using state data per simulation step, we propose leveraging particle trajectories over $N$ simulation steps as dense geometrical supervision to better constrain the simulated behaviors of continuum particles. + +Given that the deformation network trained in Sec. 3.2 can produce rough trajectories for dynamic Gaussians, we fine-tune it to establish trajectories for continuum particles. Specifically, we initialize the canonical space of particles + +$\mathbf{x}_0$ with the filled continuum $\mathbf{p}_1$ at $t = 1$ and apply deformations similar to Eq. (2), + +$$ +\mathbf {x} _ {t} = \mathbf {x} ^ {*} + \mathbf {w} \left(\mathbf {x} ^ {*}, t\right) \mathbf {B} ^ {\mathbf {x}} (t), \forall t = 0, 1, \dots , T, \tag {5} +$$ + +resulting in per-frame deformed particles $\{\mathbf{x}_t\}$ . Then we optimize this deformation with all filled continuums $\{\mathbf{p}_t\}$ using the Chamfer Distance [13, 57] + +$$ +\min _ {\mathbf {x} ^ {*}, \mathcal {D} ^ {\prime}} \left[ \mathcal {L} _ {\mathrm {C D}} \left(\left\{\mathbf {x} _ {t} \right\} _ {t = 0} ^ {T}, \left\{\mathbf {p} _ {t} \right\} _ {t = 0} ^ {T}\right) \right], \tag {6} +$$ + +where $\mathcal{D}'$ denotes the fine-tuned deformation network. Unlike the deformation network $\mathcal{D}$ trained in Sec. 3.2, this fine-tuned version $\mathcal{D}'$ is aware of the internal particles absence in Gaussians. Further, the learned motion basis with temporal positional embeddings enables it for temporal interpolation, allowing us to probe the motion between observed discrete time steps $t = 0,1,\dots,T$ . Formally, we have particle positions at each simulation time step + +$$ +\mathbf {x} _ {t} = \mathbf {x} ^ {*} + \mathbf {w} (\mathbf {x} ^ {*}, t) \mathbf {B} ^ {\mathbf {x}} (t), \forall t = 0, \tau , \dots , N \tau , \tag {7} +$$ + +where $\tau$ is the time interval between two simulation steps. We then consider these temporally dense particle trajectories as pseudo ground truth for later optimization. + +# 3.4. Material-Agnostic System Identification + +Inspired by NCLaw [59], we parameterize the elastic and plastic constitutive laws using neural networks, denoted as $\mathcal{E}_{\theta^e}$ and $\mathcal{P}_{\theta^p}$ , where $\theta^e$ and $\theta^p$ are their respective network parameters. Both networks share the same multi-layer perceptron (MLP) architecture with a single hidden layer. Further, they incorporate two physical priors, frame indifference and undeformed state equilibrium, enforced by using rotation-invariant input representations and eliminating bias terms, respectively. While we adopt the same parameterization for material-agnostic modeling, we introduce a more challenging setting by relaxing the input requirements from dense, complete state sequences $\{\mathbf{s}_t\}_{n=0}^N$ to sparse visual observations $\{\mathbf{I}_t\}_{t=0}^T$ . To this end, we follow an analysis-by-synthesis scheme. For the initial particle state $\mathbf{s}_0$ , we use positions $\mathbf{x}_0$ derived from Eq. (7), an estimizable velocity $\mathbf{v}_0$ , and identity deformation gradients $\mathbf{F}_0$ . After we evolve particle positions $\{\hat{\mathbf{x}}_t\}_{t=\tau}^{N\tau}$ through time integration as described in Eq. (1), we optimize $\theta^e$ and $\theta^p$ with + +$$ +\min _ {\theta^ {e}, \theta^ {p}} \left(\mathcal {L} _ {\mathrm {g e o}} + \mathcal {L} _ {\mathrm {s i l}}\right), \tag {8} +$$ + +where $\mathcal{L}_{\mathrm{geo}}$ supervises particle positions $\hat{\mathbf{x}}_t$ with estimated trajectories $\mathbf{x}_t$ and $\mathcal{L}_{\mathrm{sil}}$ compares rendered masks $\hat{\mathbf{M}}_t$ with object silhouettes $\mathbf{M}_t$ extracted from frame $\mathbf{I}_t$ + +$$ +\mathcal {L} _ {\mathrm {g e o}} = \mathcal {L} _ {\mathrm {t r a j}} = \mathcal {L} _ {1} \left(\left\{\hat {\mathbf {x}} _ {t} \right\} _ {t = 0} ^ {N \tau}, \left\{\mathbf {x} _ {t} \right\} _ {t = 0} ^ {N \tau}\right), \tag {9} +$$ + +$$ +\mathcal {L} _ {\mathrm {s i l}} = \mathcal {L} _ {1} \left(\left\{\hat {\mathbf {M}} _ {t} \right\} _ {t = 0} ^ {N \tau}, \left\{\mathbf {M} _ {t} \right\} _ {t = 0} ^ {N \tau}\right). \tag {10} +$$ + +While not assuming known material types in our problem setting, we note that MASIV still requires a pre-trained constitutive model from [59] as a stable initialization. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. We conduct our experiments on the PAC-NeRF dataset [38] and the Spring-Gaus dataset [100]. The PAC-NeRF dataset consists of synthetic objects simulated using the Material Point Method (MPM) and includes a variety of material types such as elastic bodies, plasticine, sand, and Newtonian and non-Newtonian fluids. This dataset provides 45 multi-view RGB video sequences that allow for the estimation of both object geometry and physical properties and approximately 14 frames per viewpoint. The Spring-Gaus dataset contains two subsets, namely, Synthetic and Real. The Synthetic subset focuses on the reconstruction and simulation of elastic objects, containing 30 frames in each of 10 viewpoints. The Real subset provides both dense-view static and sparse-view dynamic captures of elastic objects. We evaluate the performance of observable state simulation using both datasets, while the Spring-Gaus dataset is additionally utilized for future state prediction. To obtain object masks, we apply off-the-shelf matting with [35, 47]. + +Baselines. To assess the performance of our approach, we compare it against four state-of-the-art methods for system identification. PAC-NeRF [38] combines neural radiance fields with differentiable physics to jointly infer geometry and material properties. Spring-Gaus [100] integrates a spring-mass system into Gaussian representations for efficient elastic object simulation. NeuMA [8] fine-tunes a neural adapter on expert-designed physical models to better align reconstructed dynamics with visual observations. GIC [6] leverages a two-stage pipeline to obtain the geometry representation and physical properties sequentially. These baselines represent the most relevant prior work and provide a strong comparison for evaluating our method. We note that they all necessitate material priors. To be concrete, PAC-NeRF and GIC roughly categorize materials into five groups and adopt a certain set of physical models for each group. Spring-Gaus generally focuses on elastic objects, thereby being material-specific. And NeuMA leverages a scene-specific constitutive model as the material prior. + +Metrics. We evaluate our method using several standard metrics to assess both geometric reconstruction quality and physical property estimation accuracy. Metrics including 1) Chamfer Distance (CD) [13, 57] with units expressed in $10^{3}\mathrm{mm}^{2}$ ; 2) Peak Signal-to-Noise Ratio (PSNR) [24]; 3) Structural Similarity Index Measure (SSIM) [84]. The CD quantifies the similarity between two point distributions, while PSNR and SSIM assess the visual quality. + +
MethodPAC-NeRFSpring-Gaus
NewtonianNon-NewtonianElasticityPlasticineSandMeanTorusCrossCreamApplePasteChessBananaMean
\(\rightarrow\)DPAC-NeRF [38]0.2770.2360.2380.4290.2120.2784.921.100.771.113.140.962.772.11
Spring-Gaus [100]------0.170.480.360.380.191.802.600.85
NeuMA [8]------4.590.060.080.040.710.050.030.79
GIC [6]0.2430.1950.1780.1960.2500.2120.130.130.140.150.170.410.030.17
MASIV (Ours)0.2330.1980.1920.2010.2290.2100.080.100.190.160.130.190.080.13
+ +Table 1. Observable state simulation on PAC-NeRF and Spring-Gaus Synthetic datasets. Despite being the only technique that does not use expert-designed physical models (noted by the gray row), MASIV performs better than or on par with existing baselines. + +
MethodBunBurgerDogPigPotatoMean
PSNR ↑Spring-Gaus [100]30.6934.0132.1034.9732.7232.90
NeuMA [8]31.2723.7825.6125.40-26.51
GIC [6]34.6840.4537.1738.3239.9538.11
Ours40.2642.1238.9740.8443.4041.12
SSIM ↑Spring-Gaus0.9920.9940.9940.9960.9920.994
NeuMA0.9940.9930.9950.995-0.994
GIC0.9950.9970.9970.9970.9960.996
Ours0.9980.9970.9970.9980.9980.997
+ +Implementation details. In dynamic reconstruction, we follow [6] to optimize canonical Gaussians $\mathcal{G}^*$ and the deformation network $\mathcal{D}$ for 40K iterations. Then the deformation network is fine-tuned to estimate particle trajectories from filled continuum sequences for 10K iterations. In system identification, we set the simulation time interval $\tau$ to $1/200$ the frame duration. We adopt the RAdam [52] optimizer to optimize the initial velocity $\mathbf{v}_0$ for 100 steps and constitutive model parameters $\theta^e, \theta^p$ for 1000 steps. All experiments are conducted on a single NVIDIA A100 GPU. + +# 4.2. Comparison Results + +Observable state simulation. In Tab. 1, we compare MASIV with existing vision-based system identification methods for observable state simulation. The results indicate that MASIV outperforms all other methods in terms of CD, highlighting its effectiveness in reconstructing object dynamics across diverse materials. Further, Tab. 2 evaluates on real data. Thanks to the more flexible constitutive representation, MASIV achieves state-of-the-art results. This demonstrates the potential of MASIV as an adaptive and generalizable approach for system identification. + +Future state simulation. In Tab. 3, we provide a detailed evaluation of future state simulation performance. The results indicate that MASIV consistently outperforms all other methods while slightly underperforming GIC. Since NeuMA [8] does not originally experiment on the synthetic subset of Spring-Gaus, we report our reproduction results. To be concrete, we use the closest pre-trained model, Jelly, from NCLaw [59] as the material prior and conduct LoRA [25] fine-tuning with hyperparameters from its orig- + +Table 2. Observable state simulation on Spring-Gaus Real dataset. MASIV achieves state-of-the-art performance. + +
MethodTorusCrossCreamApplePasteChessBananaMean
CD\(\rightarrow\)PAC-NeRF [38]2.473.872.214.6937.708.2066.4317.94
Spring-Gaus [100]2.381.572.221.877.032.5918.485.16
NeuMA [8]1.170.401147.40.0544.260.330.30170.56
GIC [6]0.751.090.940.222.790.770.120.95
MASIV (Ours)0.160.381.512.641.751.700.241.20
PSNR\(\dagger\)PAC-NeRF [38]17.4614.1515.3719.9412.3215.0816.0415.77
Spring-Gaus [100]16.8316.9315.4221.5514.7116.0817.8917.06
NeuMA [8]16.6119.672.3226.2513.4220.8523.1617.47
GIC [6]20.2430.5119.1526.8916.3118.4429.2922.98
MASIV (Ours)22.4233.7316.4521.1915.7317.2826.5621.91
SSIM\(\dagger\)PAC-NeRF [38]0.9190.9060.8580.8780.8190.8480.8860.870
Spring-Gaus [100]0.9190.9400.8620.9020.8720.8810.9040.897
NeuMA [8]0.9420.9480.8890.9640.8890.9330.9640.933
GIC [6]0.9420.9390.9090.9480.8940.9120.9640.930
MASIV (Ours)0.9550.9670.8520.8980.8910.8780.9650.915
+ +Table 3. Future state simulation on Spring-Gaus Synthetic dataset. MASIV outperforms all baselines except for GIC. + +inal implementation. However, we observe unstable optimization processes for most categories. This is probably rooted in its sparse pixel supervision, which only operates once per frame. With dense geometrical constraints supervising hundreds of times per frame, MASIV achieves significantly better results on both geometrical accuracy and rendering quality. This suggests that MASIV effectively captures the underlying dynamics governing future states, allowing it to generate predictions that remain structurally and visually consistent over time. Unlike baseline approaches that rely on predefined material priors, MASIV demonstrates the ability to infer relevant properties directly from visual observations, making it more adaptable to a wide range of scenarios. On the other hand, while dense trajectory guidance enables overfitting to observations, the lack of category-level regularization, unlike hand-crafted constitutive models, limits its generalization under scarce data. This leads to inferior performance than GIC, which can benefit from its pre-known constitutive model as an oracle. More analysis on the enhanced generalization ability of MASIV with more data is in Sec. C of Supp. Mat. + +# 4.3. Ablation Study + +We investigate the impact of geometrical objectives here, while referring the readers to more ablation studies in Sec. B of Supp. Mat. Fig. 3 validates the effectiveness of dense trajectory supervision. As shown in Fig. 4, we evaluate three geometrical objectives by replacing the $\mathcal{L}_{\mathrm{geo}}$ term in opti + +![](images/a8d08b06a43e59ee3dad00c5bdeb89db87af2e923380dcfed7a22b675df5e87a.jpg) +Figure 3. Ablation study of geometrical objectives $\mathcal{L}_{\mathrm{geo}}$ on PAC-NeRF dataset. We compare three types of geometrical objectives against a baseline without geometrical guidance and plot their simulation errors using Chamfer Distance (CD). Overall, dense trajectory guidance achieves the lowest and most stable errors. The figure is best viewed in color. + +![](images/7567d0477707b026ddf94e294ef0df503a49d1bbaf64417e9ecf864da56cae28.jpg) +(a) Surfaces +(b) Continuums + +![](images/1f45a87c5acba9c8fcf011c8e9bebe3196d7968da1d2c1f1bb1aaa165547e13f.jpg) +(c) Gaussian Splats +(d) Trajectories + +mizing the neural constitutive model objective, Eq. (8). We define them formally as follows. (1) Surface: per-frame surface alignment between the simulated continuum and the filled continuum, + +$$ +\mathcal {L} _ {\text {s u r f}} = \mathcal {L} _ {\mathrm {C D}} \left(\left\{\mathcal {S} (\hat {\mathbf {x}} _ {t}) \right\} _ {t = 0} ^ {T}, \left\{\mathcal {S} (\mathbf {p} _ {t}) \right\} _ {t = 0} ^ {T}\right), \tag {11} +$$ + +where $\mathcal{S}$ denotes the surface extraction by thresholding Gaussian opacities [6]. (2) Continuum: similar to Surface while between solid continuums instead of surfaces, + +$$ +\mathcal {L} _ {\text {c o n t}} = \mathcal {L} _ {\mathrm {C D}} \left(\left\{\hat {\mathbf {x}} _ {t} \right\} _ {t = 0} ^ {T}, \left\{\mathbf {p} _ {t} \right\} _ {t = 0} ^ {T}\right). \tag {12} +$$ + +(3) Trajectory: per-simulation-step trajectory alignment between the simulated continuum $\{\hat{\mathbf{x}}_t\}_{t=0}^{N\tau}$ and the continuum predicted by the deformation network $\{\mathbf{x}_t\}_{t=0}^{N\tau}$ , as defined in Eq. (9). By analyzing per-frame errors, we first confirm the necessity of geometrical guidance, as the baseline (red curve) exhibits significantly higher CD errors across all material types. Surface-based supervision (yellow curve) shows moderate error reduction but remains less stable over time, especially in Plasticine and Sand. Continuum-based supervision (blue curve) further reduces errors but struggles with certain materials, e.g., Sand. These error overshoots + +![](images/df93c4895806c327cddbdc010c34deef9c591e70875751216b002522252b80f6.jpg) +Figure 4. Visualization of geometrical objectives $\mathcal{L}_{\mathrm{geo}}$ on PAC-NeRF Bird. The fine-tuned deformation network $\mathcal{D}'$ captures smooth and accurate motion for continuums, including low-frequency translation of the body and high-frequency swing of the tail. We space adjacent frames in the horizontal axis for clarity. +Ground Truth Surface $\mathcal{L}_{\mathrm{surf}}$ Continuum $\mathcal{L}_{\mathrm{cont}}$ Trajectory $\mathcal{L}_{\mathrm{traj}}$ +Figure 5. Qualitative comparison of geometrical objectives $\mathcal{L}_{\mathrm{geo}}$ on PAC-NeRF Cream. Using sparse surfaces or solid continuum as geometrical cues can lead to physically implausible deformations, while trajectory guidance mitigates these issues by introducing temporally denser constraints. Better viewed when zoomed in. + +are probably due to supervising relatively complex neural constitutive models with sparse signals, which might lead to overfitting since the simulated behaviors between observed discrete time steps are not well constrained. Finally, by enforcing dense alignment across simulation steps, trajectory-based supervision (green curve) achieves the lowest and most stable errors across all materials. + +# 4.4. Qualitative Analysis + +Dynamics reconstruction. We visualize the reconstructed Cream object by using the alternative objectives in Fig. 5. Here, we observe physically implausible deformations in results for surface and continuum supervision. Under uniform force, the object exhibits inconsistent distortion, espe + +![](images/42886a4924893fee5eba6c102b9c3f58652849db4c7b2ef5dfe4fda829f36fd2.jpg) +Figure 6. Intra-class generalization. We first optimize the geometry for PAC-NeRF Bird and then use the constitutive models optimized for reference elastic objects to predict novel dynamic behaviors, thereby softening or hardening the Bird respectively. + +cially in the boundary between blue stripes and red stripes. We attribute this to two main reasons. 1) Chamfer Distance based geometrical constraint focuses mainly on the overall structure. While it encourages alignment between simulated and reference shapes, they do not enforce point-wise correspondence. A single particle can thus be matched with different ground truth particles in different time steps. As a result, local mismatches can accumulate, leading to inaccurate deformations. 2) Sparse supervision fails to constrain intermediate motion dynamics. Surface and continuum objectives provide alignment only at discrete time steps, leaving large gaps in supervision. Without explicit constraints on how the object deforms over time, the neural constitutive model may interpolate arbitrarily, leading to unstable and unrealistic motion patterns. + +Dynamics generalization. We further study the generalization ability for MASIV. Fig. 6 examines how MASIV can be used to modify the dynamic characteristics of a single object (PAC-NeRF Bird). Specifically, we first reconstruct the geometrical representation of the Bird and optimize neural constitutive models for two elastic cross-shaped objects. These optimized constitutive models are then used as references to drive the simulation of the Bird, which deforms accordingly, exhibiting fluid-like softness on the left and increased rigidity and bounce on the right. Additionally, we see continuous transitions between soft and rigid behaviors when interpolating the neural constitutive model weights. In Fig. 7, we extend this study to inter-class generalization by applying neural constitutive models trained on different materials to a novel object, the PAC-NeRF Trophy. The results show that despite sharing the same initial state, each material model induces distinct deformation patterns. These + +![](images/96852bb1d70c3f30255e28d3469a5bfa833545729202d706c56af8b9aca9bce7.jpg) +Figure 7. Inter-class generalization. We train multiple neural constitutive models with different materials and apply them to the PAC-NeRF Trophy. While sharing the same initial state, the Trophies with different materials exhibit distinct behaviors over time. + +observations confirm that MASIV effectively captures generalizable intrinsic dynamics. More qualitative results can be found in Sec. D of Supp. Mat. + +# 4.5. Discussion + +While MASIV excels in material-agnostic system identification, several limitations remain. First, it relies on multiview video input, restricting applicability to single-view or sparse observations. Second, our approach assumes that internal geometries and deformations can be inferred from surface motion, following a spatially coherent model. This assumption may break down in materials with complex internal flows, such as granular media or multiphase fluids. Finally, MASIV does not explicitly disentangle material properties from external forces in system identification, assuming a uniform gravity field for all objects. Future works aim to relax these assumptions for broader applicability. + +# 5. Conclusion + +We present MASIV, a material-agnostic system identification framework. By integrating neural constitutive models with dense geometry guidance, MASIV effectively captures object dynamics across diverse materials. Our approach reconstructs dynamic Gaussians from multi-view videos and fine-tunes a deformation model within a solid continuum. This enables us to extract dense motion cues, which, combined with visual observations, guide the learning of neural constitutive models. Through this formulation, MASIV generalizes beyond traditional system identification methods that rely on pre-defined material priors. Extensive experiments highlight its ability to reconstruct observable dynamics and to generalize for novel interactions. + +# Acknowledgment + +This work was supported in part by U.S. NSF grant DBI-2238093, and in part by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project). + +# References + +[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 2 +[2] Genesis Authors. Genesis: A universal and generative physics engine for robotics and beyond, 2024. 3 +[3] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B Lindell. 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7996-8006, 2024. 3 +[4] Piotr Borycki, Weronika Smolak, Joanna Waczyńska, Marcin Mazur, Sławomir Tadeja, and Przemysław Spurek. Gasp: Gaussian splatting for physic-based simulations. arXiv preprint arXiv:2409.05819, 2024. 3 +[5] Johannes Brandstetter, Daniel Worrall, and Max Welling. Message passing neural pde solvers. arXiv preprint arXiv:2202.03376, 2022. 2 +[6] Junhao Cai, Yuji Yang, Weihao Yuan, Yisheng He, Zilong Dong, Liefeng Bo, Hui Cheng, and Qifeng Chen. Gic: Gaussian-informed continuum for physical property identification and simulation. arXiv preprint arXiv:2406.14927, 2024. 2, 3, 4, 5, 6, 7 +[7] Yuanhao Cai, Jiahao Wang, Alan Yuille, Zongwei Zhou, and Angtian Wang. Structure-aware sparse-view x-ray 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11174-11183, 2024. 2 +[8] Junyi Cao, Shanyan Guan, Yanhao Ge, Wei Li, Xiaokang Yang, and Chao Ma. Neuma: Neural material adaptor for visual grounding of intrinsic dynamics. Advances in Neural Information Processing Systems, 37:65643-65669, 2025. 2, 3, 4, 5, 6 +[9] Pradyumna Chari, Chinmay Talegaonkar, Yunhao Ba, and Achuta Kadambi. Visual physics: Discovering physical laws from videos. arXiv preprint arXiv:1911.11893, 2019. 2 +[10] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splattering guidance. arXiv preprint arXiv:2312.00846, 2023. 4 +[11] Hsiao-yu Chen, Edith Tretschk, Tuur Stuyck, Petr Kadlecek, Ladislav Kavan, Etienne Vouga, and Christoph Lassner. Virtual elastic objects. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 15827-15837, 2022. 3 +[12] Tao Du, Kui Wu, Pingchuan Ma, Sebastien Wah, Andrew Spielberg, Daniela Rus, and Wojciech Matusik. Diffpd: Differentiable projective dynamics. ACM Transactions on Graphics (TOG), 41(2):1-21, 2021. 3 +[13] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In European Conference on Computer Vision, pages 108-124. Springer, 2020. 5 +[14] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, et al. Gaussian splashing: Unified particles for versatile motion synthesis and rendering. arXiv preprint arXiv:2401.15318, 3, 2024. 3 +[15] Yutao Feng, Yintong Shang, Xuan Li, Tianjia Shao, Chenfanfu Jiang, and Yin Yang. Pie-nerf: Physics-based interactive elastodynamics with nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4450–4461, 2024. 3 +[16] Zhoujie Fu, Jiacheng Wei, Wenhao Shen, Chaoyue Song, Xiaofeng Yang, Fayao Liu, Xulei Yang, and Guosheng Lin. Sync4d: Video guided controllable dynamics for physics-based 4d generation. arXiv preprint arXiv:2405.16849, 2024.3 +[17] Moritz Geilinger, David Hahn, Jonas Zehnder, Moritz Bächer, Bernhard Thomaszewski, and Stelian Coros. Add: Analytically differentiable dynamics for multi-body systems with frictional contact. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 3 +[18] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, et al. Motion prompting: Controlling video generation with motion trajectories. arXiv preprint arXiv:2412.02700, 2024. 3 +[19] Sharath Girish, Tianye Li, Amrita Mazumdar, Abhinav Shrivastava, David Luebke, and Shalini De Mello. Queen: Quantized efficient encoding of dynamic gaussians for streaming free-viewpoint videos. arXiv preprint arXiv:2412.04469, 2024. 2 +[20] Shanyan Guan, Huayu Deng, Yunbo Wang, and Xiaokang Yang. Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields. In International conference on machine learning, pages 7919-7929. PMLR, 2022. 2 +[21] Eric Heiden, Miles Macklin, Yashraj Narang, Dieter Fox, Animesh Garg, and Fabio Ramos. Disect: A differentiable simulation engine for autonomous robotic cutting. arXiv preprint arXiv:2105.12244, 2021. 3 +[22] Maximilian Herde, Bogdan Raonic, Tobias Rohner, Roger Käppeli, Roberto Molinaro, Emmanuel de Bézenac, and Siddhartha Mishra. Poseidon: Efficient foundation models for pdes. Advances in Neural Information Processing Systems, 37:72525-72624, 2025. 2, 4 +[23] Florian Hofherr, Lukas Koestler, Florian Bernard, and Daniel Cremers. Neural implicit representations for physical parameter inference from a single video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2093-2103, 2023. 2 + +[24] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 5 +[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 6 +[26] Yuanming Hu, Yu Fang, Ziheng Ge, Ziyin Qu, Yixin Zhu, Andre Pradhana, and Chenfanfu Jiang. A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2, 3 +[27] Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Fredo Durand. Difftaichi: Differentiable programming for physical simulation. arXiv preprint arXiv:1910.00935, 2019. 2, 3 +[28] Tianyu Huang, Haoze Zhang, Yihan Zeng, Zhilu Zhang, Hui Li, Wangmeng Zuo, and Rynson WH Lau. Dreamphysics: Learning physical properties of dynamic 3d gaussians with video diffusion priors. arXiv preprint arXiv:2406.01476, 2024.3 +[29] Zhiao Huang, Yuanming Hu, Tao Du, Siyuan Zhou, Hao Su, Joshua B Tenenbaum, and Chuang Gan. Plasticinelab: A soft-body manipulation benchmark with differentiable physics. arXiv preprint arXiv:2104.03311, 2021. 3 +[30] Krishna Murthy Jatavallabhula, Miles Macklin, Florian Golemo, Vikram Voleti, Linda Petrini, Martin Weiss, Breandan Considine, Jérôme Parent-Lévesque, Kevin Xie, Kenny Erleben, et al. gradsim: Differentiable simulation for system identification and visuomotor control. arXiv preprint arXiv:2104.02646, 2021. 2, 3 +[31] Chenfanfu Jiang, Craig Schroeder, Joseph Teran, Alexey Stomakhin, and Andrew Selle. The material point method for simulating continuum materials. In ACM SIGGRAPH 2016 Courses, pages 1-52, 2016. 2, 3, 1 +[32] Takuhiro Kaneko. Improving physics-augmented continuum neural radiance field-based geometry-agnostic system identification with lagrangian particle optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5480, 2024. 3 +[33] Namgyu Kang, Jaemin Oh, Youngjoon Hong, and Eunbyung Park. Pig: Physics-informed gaussians as adaptive parametric mesh representations. arXiv preprint arXiv:2412.05994, 2024. 4 +[34] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 1 +[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 5 +[36] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. In European Conference on Computer Vision, pages 252-269. Springer, 2025. 3, 4 + +[37] Jinxi Li, Ziyang Song, and Bo Yang. Nvfi: neural velocity fields for 3d physics learning from dynamic videos. Advances in Neural Information Processing Systems, 36, 2024. 3 +[38] Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Krishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang, and Chuang Gan. Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. arXiv preprint arXiv:2303.05512, 2023. 2, 3, 4, 5, 6 +[39] Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B Tenenbaum, and Antonio Torralba. Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. arXiv preprint arXiv:1810.01566, 2018. 3 +[40] Yifei Li, Tao Du, Kui Wu, Jie Xu, and Wojciech Matusik. Diffcloth: Differentiable cloth simulation with dry frictional contact. ACM Transactions on Graphics (TOG), 42 (1):1-20, 2022. 3 +[41] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2 +[42] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. 2 +[43] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 3 +[44] Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Azizzadenesheli, and Anima Anandkumar. Physics-informed neural operator for learning partial differential equations. ACM/JMS Journal of Data Science, 1(3):1-27, 2024. 2 +[45] Junbang Liang, Ming Lin, and Vladlen Koltun. Differentiable cloth simulation for inverse problems. Advances in Neural Information Processing Systems, 32, 2019. 3 +[46] Xiao Liang, Fei Liu, Yutong Zhang, Yuelei Li, Shan Lin, and Michael Yip. Real-to-sim deformable object manipulation: Optimizing physics models with residual mappings for robotic surgery. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 15471-15477. IEEE, 2024. 3 +[47] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 5 +[48] Yuchen Lin, Chenguo Lin, Jianjin Xu, and Yadong Mu. Omniphysgs: 3d constitutive gaussians for general physics-based dynamics generation. arXiv preprint arXiv:2501.18982, 2025. 2, 3 +[49] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d + +with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576-8588, 2024. 3 +[50] Phillip Lippe, Bas Veeling, Paris Perdikaris, Richard Turner, and Johannes Brandstetter. Pde-refiner: Achieving accurate long rollouts with neural pde solvers. Advances in Neural Information Processing Systems, 36:67398-67433, 2023. 2 +[51] Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. Physics3d: Learning physical properties of 3d gaussians via video diffusion. arXiv preprint arXiv:2406.04338, 2024. 3 +[52] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond. arXiv preprint arXiv:1908.03265, 2019. 6 +[53] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splatting from causally-captured monocular videos. arXiv preprint arXiv:2406.00434, 2024. 2 +[54] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imageto-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 3 +[55] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2 +[56] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2 +[57] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. 5 +[58] Pingchuan Ma, Tao Du, Joshua B Tenenbaum, Wojciech Matusik, and Chuang Gan. Risp: Rendering-invariant state predictor with differentiable simulation and rendering for cross-domain parameter estimation. arXiv preprint arXiv:2205.05678, 2022.3 +[59] Pingchuan Ma, Peter Yichen Chen, Bolei Deng, Joshua B Tenenbaum, Tao Du, Chuang Gan, and Wojciech Matusik. Learning neural constitutive laws from motion observations for generalizable pde dynamics. In International Conference on Machine Learning, pages 23279-23300. PMLR, 2023. 2, 3, 4, 5, 6 +[60] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[61] Clinton A Mo, Kun Hu, Chengjiang Long, and Zhiyong Wang. Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In Proceedings of the IEEE/CVF Conference on + +Computer Vision and Pattern Recognition, pages 13894-13903, 2023. 2 +[62] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2 +[63] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2 +[64] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 2 +[65] Yiling Qiao, Junbang Liang, Vladlen Koltun, and Ming Lin. Differentiable simulation of soft multi-body systems. Advances in Neural Information Processing Systems, 34: 17123-17135, 2021. 3 +[66] Yi-Ling Qiao, Alexander Gao, and Ming Lin. Neuphysics: Editable neural geometry and physics from monocular videos. Advances in Neural Information Processing Systems, 35:12841-12854, 2022. 3 +[67] Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Language-driven physics-based scene synthesis and editing via feature splatting. In European Conference on Computer Vision, pages 368-383. Springer, 2024. 3 +[68] Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378:686-707, 2019. 3 +[69] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 3 +[70] Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In International conference on machine learning, pages 8459-8468. PMLR, 2020. 3 +[71] Yidi Shao, Mu Huang, Chen Change Loy, and Bo Dai. Gausim: Registering elastic objects into digital world by gaussian simulator. arXiv preprint arXiv:2412.17804, 2024.3 +[72] Haochen Shi, Huazhe Xu, Samuel Clarke, Yunzhu Li, and Jiajun Wu. Robocook: Long-horizon elasto-plastic object manipulation with diverse tools. arXiv preprint arXiv:2306.14447, 2023. 3 +[73] Haochen Shi, Huazhe Xu, Zhiao Huang, Yunzhu Li, and Ji-jun Wu. Robocraft: Learning to see, simulate, and shape elasto-plastic objects in 3d with graph networks. The International Journal of Robotics Research, 43(4):533-549, 2024. 3 +[74] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun + +Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[75] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 3 +[76] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 2 +[77] Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 1 +[78] Deborah Sulsky, Shi-Jian Zhou, and Howard L Schreyer. Application of a particle-in-cell method to solid mechanics. Computer physics communications, 87(1-2):236-252, 1995. 2, 3 +[79] Priya Sundaresan, Rika Antonova, and Jeannette Bohgl. Diffcloud: Real-to-sim from point clouds with differentiable simulation and rendering of deformable objects. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10828-10835. IEEE, 2022. 3 +[80] Tetsuya Takahashi and Ming C Lin. Video-guided real-to-virtual parameter transfer for viscous fluids. ACM Transactions on Graphics (TOG), 38(6):1-12, 2019. 3 +[81] Xiyang Tan, Ying Jiang, Xuan Li, Zeshun Zong, Tianyi Xie, Yin Yang, and Chenfanfu Jiang. Physmotion: Physics-grounded dynamics from a single image. arXiv preprint arXiv:2411.17189, 2024. 3 +[82] Bin Wang, Longhua Wu, KangKang Yin, Uri M Ascher, Libin Liu, and Hui Huang. Deformation capture and modeling of soft objects. ACM Trans. Graph., 34(4):94-1, 2015. 3 +[83] Chaoyang Wang, Lachlan Ewen MacDonald, Laszlo A Jeni, and Simon Lucey. Flow supervision for deformable nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21128-21137, 2023. 2 +[84] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[85] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3 +[86] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splattering for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 3, 4 + +[87] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2024. 3 +[88] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9421-9431, 2021. 2 +[89] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 3 +[90] Zhenjia Xu, Jiajun Wu, Andy Zeng, Joshua B Tenenbaum, and Shuran Song. Densephysnet: Learning dense physical object representations via multi-step dynamic interactions. arXiv preprint arXiv:1906.03853, 2019. 3 +[91] Tianju Xue, Shuheng Liao, Zhengtao Gan, Chanwook Park, Xiaoyu Xie, Wing Kam Liu, and Jian Cao. Jax-fem: A differentiablegpu-accelerated 3d finite element solver for automatic inverse design and mechanistic data science. Computer Physics Communications, 291:108802, 2023. 2 +[92] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2 +[93] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3, 4 +[94] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13117-13125, 2021. 2 +[95] Heng Yu, Joel Julin, Zoltan A Milacski, Koichiro Niinuma, and László A Jeni. Dylin: Making light field networks dynamic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12397-12406, 2023. 2 +[96] Heng Yu, Joel Julin, Zoltán Á Milacski, Koichiro Niinuma, and László A Jeni. Cogs: Controllable gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21624-21633, 2024. 2 +[97] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 4 +[98] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2025. 3 + +[99] Dongzhe Zheng, Siqiong Yao, Wenqiang Xu, and Cewu Lu. Differentiable cloth parameter identification and state estimation in manipulation. IEEE Robotics and Automation Letters, 2024. 3 +[100] Licheng Zhong, Hong-Xing Yu, Jiajun Wu, and Yunzhu Li. Reconstruction and simulation of elastic objects with spring-mass 3d gaussians. In European Conference on Computer Vision, pages 407-423. Springer, 2025. 2, 3, 4, 5, 6 +[101] Xiaowei Zhou, Can Yang, and Weichuan Yu. Moving object detection by detecting contiguous outliers in the low-rank representation. IEEE transactions on pattern analysis and machine intelligence, 35(3):597-610, 2012. 2 \ No newline at end of file diff --git a/2025/Toward Material-Agnostic System Identification from Videos/images.zip b/2025/Toward Material-Agnostic System Identification from Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3e55dc8514b02ce48d88187f39f65622a083dbd9 --- /dev/null +++ b/2025/Toward Material-Agnostic System Identification from Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6c3dca8cc923949bfdcf3d277f63204030e62e179a19b7925ab7892774075328 +size 503156 diff --git a/2025/Toward Material-Agnostic System Identification from Videos/layout.json b/2025/Toward Material-Agnostic System Identification from Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..165edebe6ca09ed5fda50d28ea9432c6d380aaf7 --- /dev/null +++ b/2025/Toward Material-Agnostic System Identification from Videos/layout.json @@ -0,0 +1,10597 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 111, + 489, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 111, + 489, + 129 + ], + "spans": [ + { + "bbox": [ + 121, + 111, + 489, + 129 + ], + "type": "text", + "content": "Toward Material-Agnostic System Identification from Videos" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 154, + 511, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 154, + 511, + 168 + ], + "spans": [ + { + "bbox": [ + 96, + 154, + 511, + 168 + ], + "type": "text", + "content": "Yizhou Zhao1 Haoyu Chen1 Chunjiang Liu1 Zhenyang Li2 Charles Herrmann3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 118, + 168, + 496, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 168, + 496, + 182 + ], + "spans": [ + { + "bbox": [ + 118, + 168, + 496, + 182 + ], + "type": "text", + "content": "Junhwa " + }, + { + "bbox": [ + 118, + 168, + 496, + 182 + ], + "type": "inline_equation", + "content": "\\mathsf{Hur}^3" + }, + { + "bbox": [ + 118, + 168, + 496, + 182 + ], + "type": "text", + "content": " Yinxiao Li3 Ming-Hsuan Yang3,4 Bhiksha Raj1 Min " + }, + { + "bbox": [ + 118, + 168, + 496, + 182 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\mathbf{u}^{1*}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 76, + 182, + 536, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 182, + 536, + 196 + ], + "spans": [ + { + "bbox": [ + 76, + 182, + 536, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 76, + 182, + 536, + 196 + ], + "type": "text", + "content": "Carnegie Mellon University " + }, + { + "bbox": [ + 76, + 182, + 536, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 76, + 182, + 536, + 196 + ], + "type": "text", + "content": "University of Alabama at Birmingham " + }, + { + "bbox": [ + 76, + 182, + 536, + 196 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 76, + 182, + 536, + 196 + ], + "type": "text", + "content": "Google " + }, + { + "bbox": [ + 76, + 182, + 536, + 196 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 76, + 182, + 536, + 196 + ], + "type": "text", + "content": "UC Merced" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 192, + 196, + 418, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 196, + 418, + 209 + ], + "spans": [ + { + "bbox": [ + 192, + 196, + 418, + 209 + ], + "type": "text", + "content": "https://github.com/Skaldak/MASIV" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 86, + 230, + 523, + 459 + ], + "blocks": [ + { + "bbox": [ + 86, + 230, + 523, + 459 + ], + "lines": [ + { + "bbox": [ + 86, + 230, + 523, + 459 + ], + "spans": [ + { + "bbox": [ + 86, + 230, + 523, + 459 + ], + "type": "image", + "image_path": "54aeab1a885dda3cf57aa77b1e3cc10ce1e8177a2e18b1495742ca14509692bc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 466, + 548, + 489 + ], + "lines": [ + { + "bbox": [ + 61, + 466, + 548, + 489 + ], + "spans": [ + { + "bbox": [ + 61, + 466, + 548, + 489 + ], + "type": "text", + "content": "Figure 1. System identification from video observations seeks to understand the intrinsic dynamics of a given scene. (a) Existing methods rely on scene-specific material priors. (b) Our MASIV removes this requirement, grounding object motion in a material-agnostic manner." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 156, + 498, + 202, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 498, + 202, + 509 + ], + "spans": [ + { + "bbox": [ + 156, + 498, + 202, + 509 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 520, + 297, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 520, + 297, + 684 + ], + "spans": [ + { + "bbox": [ + 61, + 520, + 297, + 684 + ], + "type": "text", + "content": "System identification from videos aims to recover object geometry and governing physical laws. Existing methods integrate differentiable rendering with simulation but rely on predefined material priors, limiting their ability to handle unknown ones. We introduce MASIV, the first vision-based framework for material-agnostic system identification. Unlike existing approaches that depend on hand-crafted constitutive laws, MASIV employs learnable neural constitutive models, inferring object dynamics without assuming a scene-specific material prior. However, the absence of full particle state information imposes unique challenges, leading to unstable optimization and physically implausible behaviors. To address this, we introduce dense geometric guidance by reconstructing continuum particle trajectory" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 548, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 548, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 548, + 546 + ], + "type": "text", + "content": "ries, providing temporally rich motion constraints beyond sparse visual cues. Comprehensive experiments show that MASIV achieves state-of-the-art performance in geometric accuracy, rendering quality, and generalization ability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 392, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 392, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 392, + 568 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 576, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 576, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 576, + 548, + 705 + ], + "type": "text", + "content": "From the flow of water down a stream to the deformation of a rubber ball upon impact, the physical world operates with a set of underlying principles that govern motion and interaction. Humans intuitively grasp these principles through observation, allowing us to see a scene and instinctively predict its outcome or re-imagine it with different initial dynamics. Material identification forms the basis of this ability, requiring us to determine an object's composition and understand how that influences its response to forces. This capability is fundamental to world modeling, the process of constructing internal representations that help us understand" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "spans": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 11, + 489, + 23 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 11, + 489, + 23 + ], + "spans": [ + { + "bbox": [ + 150, + 11, + 489, + 23 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 201, + 24, + 438, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 24, + 438, + 34 + ], + "spans": [ + { + "bbox": [ + 201, + 24, + 438, + 34 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 182, + 35, + 457, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 35, + 457, + 45 + ], + "spans": [ + { + "bbox": [ + 182, + 35, + 457, + 45 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 73, + 694, + 147, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 694, + 147, + 704 + ], + "spans": [ + { + "bbox": [ + 73, + 694, + 147, + 704 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5944" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 294, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 294, + 117 + ], + "type": "text", + "content": "and anticipate the evolution of the environment. However, replicating the ability to generalize physical laws to diverse scenarios remains a challenge for artificial intelligence." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 295, + 351 + ], + "type": "text", + "content": "To bridge this gap, researchers have explored vision-based system identification [6, 8, 9, 20, 23, 30, 38], which aims to infer governing physical laws from visual observations. Typically, they optimize a parameterized physical model, by end-to-end integrating differentiable renderers, such as Neural Radiance Fields (NeRF) [60] and 3D Gaussian Splatting (3DGS) [34], with differentiable simulators [26, 27, 31, 78, 91]. To enforce physically meaningful constraints on the simulation, existing approaches [6, 38, 100] often rely on material-specific constitutive laws, such as elasticity, plasticity, or viscosity models. These laws define how an object's material properties influence its response to external forces, allowing system identification methods to estimate parameters such as stiffness, damping, or friction. While effective in controlled settings, this dependence on hand-crafted material models limits adaptability. Moreover, these methods assume knowing the material type for a given scene, making it difficult for them to generalize across in-the-wild scenarios where material properties are unknown." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 354, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 354, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 61, + 354, + 295, + 552 + ], + "type": "text", + "content": "Directly targeting these limitations, we introduce Material-Agnostic System Identification from Videos (MASIV), a vision-based approach that infers object dynamics free of predefined material priors. As illustrated in Fig. 1, rather than relying on expert-designed elasticity and plasticity models, MASIV employs fully learnable neural constitutive models inspired by NCLaw [59]. However, directly applying neural constitutive models to vision-based system identification poses significant challenges. Similar to the learning process of neural PDE solvers [5, 22, 44, 50], the original formulation of NCLaw assumes access to complete state observations of simulated particles, including position, velocity, deformation gradient, and affine momentum. Since visual observation fails to provide such complete information, simply relying on frame-wise pixel supervision provides insufficient constraints. This leads to unstable optimization and poor convergence [48]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 554, + 295, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 554, + 295, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 554, + 295, + 705 + ], + "type": "text", + "content": "We combine two main insights to enable this optimization. First, despite the lack of ground truth full particle state, frame-wise observations still allow reasonable estimates for some of them, such as position and velocity. Existing dynamic reconstruction methods [42, 55, 56] factorize moving objects with their canonical representation and corresponding deformations, which enables per-frame position tracking. Further, the commonly adopted motion basis [61, 94, 101] provides a low-rank encoding of object motion, facilitating state interpolation between frames. This formulation allows us to reconstruct particle trajectories over dense time steps. Second, although trajectory estimations are inherently limited to visible object regions," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 81, + 547, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 128 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 128 + ], + "type": "text", + "content": "internal deformations can be inferred by propagating surface motions inward. Our intuition is that nearby particles inside an object should be deformed similarly, akin to the As-Rigid-As-Possible (ARAP) prior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 312, + 129, + 547, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 547, + 291 + ], + "type": "text", + "content": "We use these insights to guide the design of our framework. Concretely, we reconstruct dynamic Gaussians from multi-view video inputs, simultaneously establishing geometry and deformation models of the scene. The geometry model is then transformed into a solid continuum, where the deformation model is fine-tuned to encode both the exterior and the interior motion. To identify intrinsic dynamics, we guide the learning of the neural constitutive model using two complementary cues: appearance cues from video frames and geometric constraints from the fine-tuned deformation model. In this way, MASIV distills dense particle motion from sparse visual observations and guides the learning of neural constitutive models for stable, material-agnostic system identification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 324, + 293, + 516, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 293, + 516, + 304 + ], + "spans": [ + { + "bbox": [ + 324, + 293, + 516, + 304 + ], + "type": "text", + "content": "Our contributions can be summed up as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 306, + 547, + 456 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 306, + 547, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 547, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 547, + 363 + ], + "type": "text", + "content": "- We present a novel framework, MASIV, that determines the intrinsic dynamics of an object from videos in a material-agnostic manner. In contrast, all other current vision-based system identification methods [6, 8, 38] necessitate predefined material priors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 364, + 547, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 547, + 410 + ], + "type": "text", + "content": "- We introduce temporally dense geometry guidance to constrain object behaviors with neural constitutive models, facilitating smoother convergence and improved performance for material-agnostic system identification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 411, + 547, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 547, + 456 + ], + "type": "text", + "content": "- Through extensive qualitative and quantitative experiments, we demonstrate that MASIV achieves superior performance in both observable state recovery and future state prediction, even without explicit material priors." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 472, + 398, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 398, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 398, + 483 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 495, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 495, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 495, + 547, + 705 + ], + "type": "text", + "content": "Dynamic reconstruction is a fundamental task in computer vision, aimed at reconstructing high-fidelity representations of complex dynamic scenes from diverse inputs such as monocular and multi-view videos [1, 7, 19, 41, 53, 76, 83, 88, 92, 95, 96]. Earlier research primarily focused on augmenting canonical Neural Radiance Fields (NeRF) [60] with a deformation field [64], which were later enhanced by incorporating volume-preserving regularization techniques [62, 63]. While primarily focused on the novel view synthesis task, their neural implicit representations often introduce noisy deformations, limiting the accuracy of recovered geometries required for reliable physics-based property estimation [38]. With the explicit representation of the scene using Gaussian ellipsoids, 3D Gaussian Splatting (3DGS) [34] achieves efficient dynamic scene reconstruction. Building on the principles of 3DGS, several studies have extended its application to 4D dynamic reconstruction by handling each frame separately [56] or mod" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5945" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 296, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 296, + 117 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 296, + 117 + ], + "type": "text", + "content": "eling the entire scene with a canonical 3D Gaussian point cloud coupled with a robust deformation framework that accurately maps it to the target scene [36, 86, 93]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 120, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 120, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 61, + 120, + 296, + 445 + ], + "type": "text", + "content": "Dynamic simulation integrates physics principles into 3D representations to generate realistic motion and interactions. A typical approach to synthesizing dynamic 3D scenes involves integrating a 3D generation pipeline with a video generation model [3, 49, 69, 75]. Recent research incorporates 3D Gaussian kernels as both visual and physical representations, embedding Newtonian dynamics into the Gaussian framework to enable seamless rendering and simulation while introducing physical constraints to ensure plausible dynamics [4, 16, 38, 48, 51, 67, 89, 100]. Building on this, Gaussian Splashing [14] introduces a position-based dynamics framework within 3D Gaussian Splating, enabling interactions among solids, fluids, and deformable objects. Additionally, motion-conditioned simulation methods [18, 43, 74, 85, 87] use trajectory-based guidance to predict object motion, showcasing how generative models can enhance dynamic scene synthesis. However, these methods primarily rely on pre-trained generative models, which often lack a fundamental understanding of physics. More recent works leverage neural physics modeling, incorporating video priors and generative models to infer object dynamics, allowing physics-driven motion synthesis without explicit solvers [15, 28, 54, 81, 98]. However, these methods are constrained to specific material types, limiting their generalization across diverse objects. To address this, we introduce a material-agnostic approach that unifies system identification across different object types, ensuring broader applicability beyond predefined materials." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 449, + 296, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 449, + 296, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 449, + 296, + 705 + ], + "type": "text", + "content": "System identification aims to understand the physical laws governing the 3D world and is an essential task for simulation [37, 40, 45, 68, 79, 100] and robotic manipulation [46, 66, 72, 73, 99]. However, accurately recovering both object geometries and physical properties remains a significant challenge. Existing methods often assume known geometries and rely on finite element methods (FEM) or mass-spring systems for dynamic simulation [80, 82], limiting their ability to handle complex, real-world materials. Neural network-based methods [39, 70, 90] offer greater flexibility by leveraging data-driven models for system identification, but they struggle to generalize across diverse material properties and environmental conditions. More recently, differentiable physics-based simulations have emerged as a promising direction [11, 12, 17, 21, 27, 29, 30, 32, 58, 65]. However, these methods often require intricate modeling to bridge the gap between simulated and real-world behavior. To improve adaptability, some approaches focus on neural constitutive modeling, embedding learned material properties within physical simulations to refine expert-designed models [8, 59]. Other methods leverage the Material Point Method (MPM)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 315, + 81, + 459, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 81, + 459, + 92 + ], + "spans": [ + { + "bbox": [ + 315, + 81, + 459, + 92 + ], + "type": "text", + "content": "Algorithm 1 MPM state transfer " + }, + { + "bbox": [ + 315, + 81, + 459, + 92 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 315, + 95, + 423, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 95, + 423, + 109 + ], + "spans": [ + { + "bbox": [ + 315, + 95, + 423, + 109 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 315, + 95, + 423, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_n = \\{\\mathbf{x}_n,\\mathbf{v}_n,\\mathbf{F}_n^e\\}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 109, + 469, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 109, + 469, + 120 + ], + "spans": [ + { + "bbox": [ + 315, + 109, + 469, + 120 + ], + "type": "text", + "content": "Output: " + }, + { + "bbox": [ + 315, + 109, + 469, + 120 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{n + 1} = \\{\\mathbf{x}_{n + 1},\\mathbf{v}_{n + 1},\\mathbf{F}_{n + 1}^{e}\\}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 321, + 120, + 494, + 159 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 321, + 120, + 468, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 120, + 468, + 132 + ], + "spans": [ + { + "bbox": [ + 321, + 120, + 468, + 132 + ], + "type": "text", + "content": "1: for " + }, + { + "bbox": [ + 321, + 120, + 468, + 132 + ], + "type": "inline_equation", + "content": "i = 1 \\dots Q" + }, + { + "bbox": [ + 321, + 120, + 468, + 132 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 321, + 120, + 468, + 132 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_n(i) = \\mathcal{E}(\\mathbf{F}_n^e (i))" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 321, + 132, + 494, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 132, + 494, + 144 + ], + "spans": [ + { + "bbox": [ + 321, + 132, + 494, + 144 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 321, + 132, + 494, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{n + 1},\\mathbf{v}_{n + 1},\\mathbf{F}_{n + 1}^{e,\\mathrm{trial}} = \\mathcal{I}(\\mathbf{x}_n,\\mathbf{v}_n,\\mathbf{F}_n^e,\\mathbf{P}_n)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 145, + 490, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 145, + 490, + 159 + ], + "spans": [ + { + "bbox": [ + 321, + 145, + 490, + 159 + ], + "type": "text", + "content": "3: for " + }, + { + "bbox": [ + 321, + 145, + 490, + 159 + ], + "type": "inline_equation", + "content": "i = 1 \\dots Q" + }, + { + "bbox": [ + 321, + 145, + 490, + 159 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 321, + 145, + 490, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{n + 1}^{e}(i) = \\mathcal{P}(\\mathbf{F}_{n + 1}^{e,\\mathrm{trial}}(i))" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 181, + 548, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 181, + 548, + 275 + ], + "spans": [ + { + "bbox": [ + 312, + 181, + 548, + 275 + ], + "type": "text", + "content": "and 3D Gaussian-based representations to enhance geometry reconstruction and material property estimation [6, 38, 71, 100]. Building upon GIC [6], our method presents a material-agnostic approach that eliminates reliance on predefined material priors. Rather than assuming material properties, our model infers object dynamics directly from visual observations, making it adapt better to a wide range of unknown or heterogeneous materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 284, + 368, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 368, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 368, + 295 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 302, + 548, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 302, + 548, + 431 + ], + "spans": [ + { + "bbox": [ + 312, + 302, + 548, + 431 + ], + "type": "text", + "content": "Given visual observations " + }, + { + "bbox": [ + 312, + 302, + 548, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\{\\mathbf{I}_0, \\mathbf{I}_1, \\dots, \\mathbf{I}_T\\}" + }, + { + "bbox": [ + 312, + 302, + 548, + 431 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 312, + 302, + 548, + 431 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 312, + 302, + 548, + 431 + ], + "type": "text", + "content": " frames capturing a dynamic object, our goal is to recover its explicit geometric representation and the governing evolution rules of the underlying dynamical system. Formally, we represent the object's geometry using a set of Gaussian kernels [34] and integrate neural constitutive laws [59] with the material point method (MPM) [26, 31, 78] to roll out their motion over time. Unlike [6, 8, 38, 100], we do not assume a known material type of the object, while exploring Material-Agnostic System Identification from Videos (MASIV). Fig. 2 shows an overview of our pipeline." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 398, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 398, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 398, + 449 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 456, + 547, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 547, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 547, + 492 + ], + "type": "text", + "content": "Material Point Method (MPM) discretizes the object of interest into " + }, + { + "bbox": [ + 313, + 456, + 547, + 492 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 456, + 547, + 492 + ], + "type": "text", + "content": " material points and models their evolution over time via a state transfer equation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 358, + 499, + 547, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 499, + 547, + 512 + ], + "spans": [ + { + "bbox": [ + 358, + 499, + 547, + 512 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {n + 1} = \\mathcal {M} (\\mathbf {s} _ {n}), \\forall n = 0, 1, \\dots , N. \\tag {1}", + "image_path": "45e5bbe4490551252dac76da6627d7e91123f7c6f694245de3da44d463e8e2c8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_n" + }, + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "text", + "content": " denotes the particle state at time step " + }, + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "text", + "content": " is the total number of simulation steps with the simulation interval " + }, + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "text", + "content": " and observed " + }, + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "text", + "content": " frames, where " + }, + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "inline_equation", + "content": "N = T / \\tau \\gg T" + }, + { + "bbox": [ + 312, + 518, + 547, + 553 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": "The transition process " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": " follows a time-stepping scheme, as outlined in Algorithm 1. The state of particle " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": " at time step " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": " is described by its position " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_n(i)" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": ", velocity " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_n(i)" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": ", and the elastic component " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_n^e (i)" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": " of its deformation gradient " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_n(i)" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": ". At each time step, the elastic constitutive law " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": " computes the first Piola-Kirchhoff stress " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_n(i)" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": " from the elastic deformation gradient " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_n^e (i)" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": ". Subsequently, the time integration scheme " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": " updates all particle states. The updated trial elastic deformation gradient, denoted as " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{n + 1}^{e,\\mathrm{trial}}(i)" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": ", needs to be further corrected by the plastic constitutive law " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": ", which applies plasticity constraints and yields the final elastic deformation gradient " + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{n + 1}^{e}(i)" + }, + { + "bbox": [ + 312, + 554, + 547, + 705 + ], + "type": "text", + "content": ". More details of MPM can be found in Sec. A.2 of Supp. Mat." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5946" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 79, + 547, + 177 + ], + "blocks": [ + { + "bbox": [ + 63, + 79, + 547, + 177 + ], + "lines": [ + { + "bbox": [ + 63, + 79, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 63, + 79, + 547, + 177 + ], + "type": "image", + "image_path": "01d83b78a20996eb69288d40cd33267e5f5a99259e915331ce3d41a5b5aaba90.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 178, + 547, + 288 + ], + "blocks": [ + { + "bbox": [ + 63, + 178, + 547, + 288 + ], + "lines": [ + { + "bbox": [ + 63, + 178, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 63, + 178, + 547, + 288 + ], + "type": "image", + "image_path": "0578d2a5e614b23dae067b60d80ca85803c40286b31c8cc2d057d7b89abe5c2f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 294, + 548, + 338 + ], + "lines": [ + { + "bbox": [ + 61, + 294, + 548, + 338 + ], + "spans": [ + { + "bbox": [ + 61, + 294, + 548, + 338 + ], + "type": "text", + "content": "Figure 2. The overview of MASIV. Our pipeline comprises three phases. Phase I, Geometrical Reconstruction, infers geometrical representations (Sec. 3.2) and dense trajectories (Sec. 3.3) for the captured object. Phase II, Material-Agnostic System Identification (Sec. 3.4), grounds dynamic characteristics by exploiting visual observations with reconstructed motion clues. From this, we obtain a generalizable digital twin in Phase III, which can simulate novel interactions, such as new velocities or forces." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 348, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 348, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 61, + 348, + 296, + 453 + ], + "type": "text", + "content": "Existing efforts [6, 8, 38, 100] for video-based system identification commonly assume known constitutive laws " + }, + { + "bbox": [ + 61, + 348, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 61, + 348, + 296, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 348, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 61, + 348, + 296, + 453 + ], + "type": "text", + "content": ", with a particular backbone form, e.g., neo-Hookean elasticity with identity plasticity return function. This simplifies the problem by estimating a small number of physical properties, such as Young's modulus, fluid viscosity, friction angles, etc., at the cost of sacrificing flexibility. In contrast, we explore the same problem without this material-specific assumption." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 460, + 242, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 242, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 242, + 471 + ], + "type": "text", + "content": "3.2. Dynamic Gaussian Reconstruction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": "We start by reconstructing dynamic Gaussians over " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": " frames. We consider this reconstruction as temporally sparse because the number of simulation steps is hundreds of times more than the number of observed video frames " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "N = T / \\tau \\gg T" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": ". Following [6, 36, 86, 93], we maintain a set of canonical Gaussians " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^* = \\{\\pmb{\\mu}^*, \\pmb{\\Sigma}^*, \\mathbf{c}^*, \\sigma^*\\}" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": " and use a neural network " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": " to warp them over time to represent observed frames. Specifically, a basis network maps a given time step " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": " bases, producing deformation basis for position " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^{\\mu}(t) \\in \\mathbb{R}^{B \\times 3}" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^s(t) \\in \\mathbb{R}^{B \\times 1}" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": ". A coefficient network then estimates the weight of each basis, " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{w}(\\pmb{\\mu}^*, t) \\in \\mathbb{R}^B" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": ", from the center coordinates of canonical Gaussian kernels " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^*" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": " and the time step " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": ". This results in deformed Gaussian kernels at each time step " + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 61, + 476, + 296, + 639 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 122, + 647, + 295, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 647, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 122, + 647, + 295, + 659 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {t} = \\boldsymbol {\\mu} ^ {*} + \\mathbf {w} (\\boldsymbol {\\mu} ^ {*}, t) \\mathbf {B} ^ {\\boldsymbol {\\mu}} (t), \\tag {2}", + "image_path": "b105ad4e585137fd36e58db531f8e9cf95da2b2be9734b6a3801d7e37594b68d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 662, + 295, + 674 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 662, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 125, + 662, + 295, + 674 + ], + "type": "interline_equation", + "content": "s _ {t} = s ^ {*} + \\mathbf {w} \\left(\\boldsymbol {\\mu} ^ {*}, t\\right) \\mathbf {B} ^ {s} (t). \\tag {3}", + "image_path": "d9f4cf7342c8dec3e4d6cf809ba41b2bc8d17e10007f0ad199fac3a3fffba00b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 681, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 681, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 681, + 296, + 704 + ], + "type": "text", + "content": "We employ the same optimization strategy in [6, 93] to minimize the L1 and Structural Similarity Index Mea" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 348, + 547, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 348, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 312, + 348, + 547, + 417 + ], + "type": "text", + "content": "sure (SSIM) losses between rendered frames " + }, + { + "bbox": [ + 312, + 348, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}_t" + }, + { + "bbox": [ + 312, + 348, + 547, + 417 + ], + "type": "text", + "content": " and corresponding ground-truth observations " + }, + { + "bbox": [ + 312, + 348, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + }, + { + "bbox": [ + 312, + 348, + 547, + 417 + ], + "type": "text", + "content": ". We also regard all Gaussian kernels as isotropic and apply L1 regularization on scales for simplicity and reconstruction quality [10, 97, 100]. The overall objective function is formulated as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 322, + 422, + 547, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 422, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 322, + 422, + 547, + 443 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {G} ^ {*}, \\mathcal {D}} \\left[ \\mathcal {L} _ {1} (\\hat {\\mathbf {I}} _ {t}, \\mathbf {I} _ {t}) + \\lambda_ {\\mathrm {S S I M}} \\mathcal {L} _ {\\mathrm {S S I M}} (\\hat {\\mathbf {I}} _ {t}, \\mathbf {I} _ {t}) + \\lambda_ {s} \\| s _ {t} \\| _ {1} \\right], \\tag {4}", + "image_path": "b872065f809bb6ae2c0a90894e2323a198ec0d56c7b2e52bfa02ef7decd95363.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 457, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 457, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 457, + 460 + ], + "type": "text", + "content": "with hyperparameters " + }, + { + "bbox": [ + 313, + 449, + 457, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SSIM}}" + }, + { + "bbox": [ + 313, + 449, + 457, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 449, + 457, + 460 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 466, + 491, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 491, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 491, + 479 + ], + "type": "text", + "content": "3.3. Continuum Trajectory Estimation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "type": "text", + "content": "To bridge the gap between unevenly distributed Gaussian particles and the corresponding solid continuum, we follow [6] to fill internal volumes, forming continuum particles " + }, + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_t\\}" + }, + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "type": "text", + "content": " with uniform scale " + }, + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "type": "text", + "content": ". Existing methods for material-specific system identification often rely on sparse visual [6, 8, 38] or geometrical [6] cues along " + }, + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "type": "text", + "content": " frames to optimize a small set of parameters. However, since we do not assume a known constitutive backbone or material types, these sparse signals may lead to unconstrained behaviors at unobserved time steps or even cause training instability. Inspired by system identification [22, 33, 59] using state data per simulation step, we propose leveraging particle trajectories over " + }, + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 312, + 483, + 547, + 658 + ], + "type": "text", + "content": " simulation steps as dense geometrical supervision to better constrain the simulated behaviors of continuum particles." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 658, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 547, + 704 + ], + "type": "text", + "content": "Given that the deformation network trained in Sec. 3.2 can produce rough trajectories for dynamic Gaussians, we fine-tune it to establish trajectories for continuum particles. Specifically, we initialize the canonical space of particles" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5947" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 297, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 297, + 105 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 61, + 81, + 297, + 105 + ], + "type": "text", + "content": " with the filled continuum " + }, + { + "bbox": [ + 61, + 81, + 297, + 105 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_1" + }, + { + "bbox": [ + 61, + 81, + 297, + 105 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 61, + 81, + 297, + 105 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 61, + 81, + 297, + 105 + ], + "type": "text", + "content": " and apply deformations similar to Eq. (2)," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 88, + 112, + 295, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 112, + 295, + 125 + ], + "spans": [ + { + "bbox": [ + 88, + 112, + 295, + 125 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} = \\mathbf {x} ^ {*} + \\mathbf {w} \\left(\\mathbf {x} ^ {*}, t\\right) \\mathbf {B} ^ {\\mathbf {x}} (t), \\forall t = 0, 1, \\dots , T, \\tag {5}", + "image_path": "1ac30b87fc1216e7ee16bd8c029f4b74523fa125660b2f4d7a6a36687dd715a5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 133, + 295, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 133, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 133, + 295, + 168 + ], + "type": "text", + "content": "resulting in per-frame deformed particles " + }, + { + "bbox": [ + 61, + 133, + 295, + 168 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_t\\}" + }, + { + "bbox": [ + 61, + 133, + 295, + 168 + ], + "type": "text", + "content": ". Then we optimize this deformation with all filled continuums " + }, + { + "bbox": [ + 61, + 133, + 295, + 168 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_t\\}" + }, + { + "bbox": [ + 61, + 133, + 295, + 168 + ], + "type": "text", + "content": " using the Chamfer Distance [13, 57]" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 175, + 295, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 175, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 113, + 175, + 295, + 193 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {x} ^ {*}, \\mathcal {D} ^ {\\prime}} \\left[ \\mathcal {L} _ {\\mathrm {C D}} \\left(\\left\\{\\mathbf {x} _ {t} \\right\\} _ {t = 0} ^ {T}, \\left\\{\\mathbf {p} _ {t} \\right\\} _ {t = 0} ^ {T}\\right) \\right], \\tag {6}", + "image_path": "0731232feccf6e287a6c0cf8265e52999fffd2cda6e4798fe4eb734f84dae369.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 201, + 296, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 201, + 296, + 294 + ], + "spans": [ + { + "bbox": [ + 61, + 201, + 296, + 294 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 201, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{D}'" + }, + { + "bbox": [ + 61, + 201, + 296, + 294 + ], + "type": "text", + "content": " denotes the fine-tuned deformation network. Unlike the deformation network " + }, + { + "bbox": [ + 61, + 201, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 61, + 201, + 296, + 294 + ], + "type": "text", + "content": " trained in Sec. 3.2, this fine-tuned version " + }, + { + "bbox": [ + 61, + 201, + 296, + 294 + ], + "type": "inline_equation", + "content": "\\mathcal{D}'" + }, + { + "bbox": [ + 61, + 201, + 296, + 294 + ], + "type": "text", + "content": " is aware of the internal particles absence in Gaussians. Further, the learned motion basis with temporal positional embeddings enables it for temporal interpolation, allowing us to probe the motion between observed discrete time steps " + }, + { + "bbox": [ + 61, + 201, + 296, + 294 + ], + "type": "inline_equation", + "content": "t = 0,1,\\dots,T" + }, + { + "bbox": [ + 61, + 201, + 296, + 294 + ], + "type": "text", + "content": ". Formally, we have particle positions at each simulation time step" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 302, + 295, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 302, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 85, + 302, + 295, + 315 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} = \\mathbf {x} ^ {*} + \\mathbf {w} (\\mathbf {x} ^ {*}, t) \\mathbf {B} ^ {\\mathbf {x}} (t), \\forall t = 0, \\tau , \\dots , N \\tau , \\tag {7}", + "image_path": "89ead90d8de91561a3e7893fd4df3bd71d28c686d25796d2b5bb56190fe2c8c1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 323, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 323, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 61, + 323, + 296, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 323, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 61, + 323, + 296, + 357 + ], + "type": "text", + "content": " is the time interval between two simulation steps. We then consider these temporally dense particle trajectories as pseudo ground truth for later optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 364, + 266, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 266, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 266, + 376 + ], + "type": "text", + "content": "3.4. Material-Agnostic System Identification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": "Inspired by NCLaw [59], we parameterize the elastic and plastic constitutive laws using neural networks, denoted as " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\theta^e}" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\theta^p}" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\theta^e" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\theta^p" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": " are their respective network parameters. Both networks share the same multi-layer perceptron (MLP) architecture with a single hidden layer. Further, they incorporate two physical priors, frame indifference and undeformed state equilibrium, enforced by using rotation-invariant input representations and eliminating bias terms, respectively. While we adopt the same parameterization for material-agnostic modeling, we introduce a more challenging setting by relaxing the input requirements from dense, complete state sequences " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_t\\}_{n=0}^N" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": " to sparse visual observations " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_t\\}_{t=0}^T" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": ". To this end, we follow an analysis-by-synthesis scheme. For the initial particle state " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_0" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": ", we use positions " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": " derived from Eq. (7), an estimizable velocity " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": ", and identity deformation gradients " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_0" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": ". After we evolve particle positions " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{x}}_t\\}_{t=\\tau}^{N\\tau}" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": " through time integration as described in Eq. (1), we optimize " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\theta^e" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\theta^p" + }, + { + "bbox": [ + 61, + 380, + 296, + 590 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 140, + 598, + 295, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 598, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 140, + 598, + 295, + 615 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta^ {e}, \\theta^ {p}} \\left(\\mathcal {L} _ {\\mathrm {g e o}} + \\mathcal {L} _ {\\mathrm {s i l}}\\right), \\tag {8}", + "image_path": "5eaf5f7ed5eeec94ff5b84c9a684ba7698c865df06e8df4ff1a88389b8b9fa3b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "spans": [ + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "text", + "content": " supervises particle positions " + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_t" + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "text", + "content": " with estimated trajectories " + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sil}}" + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "text", + "content": " compares rendered masks " + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_t" + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "text", + "content": " with object silhouettes " + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_t" + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "text", + "content": " extracted from frame " + }, + { + "bbox": [ + 61, + 623, + 296, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_t" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 100, + 666, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 666, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 100, + 666, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g e o}} = \\mathcal {L} _ {\\mathrm {t r a j}} = \\mathcal {L} _ {1} \\left(\\left\\{\\hat {\\mathbf {x}} _ {t} \\right\\} _ {t = 0} ^ {N \\tau}, \\left\\{\\mathbf {x} _ {t} \\right\\} _ {t = 0} ^ {N \\tau}\\right), \\tag {9}", + "image_path": "fc18e52543aa47a76c1b64b2ff3528bb69dae2e92aadc40f804a6f34154e5b9f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 104, + 684, + 295, + 702 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 684, + 295, + 702 + ], + "spans": [ + { + "bbox": [ + 104, + 684, + 295, + 702 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s i l}} = \\mathcal {L} _ {1} \\left(\\left\\{\\hat {\\mathbf {M}} _ {t} \\right\\} _ {t = 0} ^ {N \\tau}, \\left\\{\\mathbf {M} _ {t} \\right\\} _ {t = 0} ^ {N \\tau}\\right). \\tag {10}", + "image_path": "3b4e3324c1e385b9c7c539f89363756892140c583b406d15e3d17ce3a2e4db8b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 81, + 547, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 116 + ], + "type": "text", + "content": "While not assuming known material types in our problem setting, we note that MASIV still requires a pre-trained constitutive model from [59] as a stable initialization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 131, + 392, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 131, + 392, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 131, + 392, + 144 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 152, + 437, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 437, + 164 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 437, + 164 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 172, + 547, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 172, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 547, + 382 + ], + "type": "text", + "content": "Datasets. We conduct our experiments on the PAC-NeRF dataset [38] and the Spring-Gaus dataset [100]. The PAC-NeRF dataset consists of synthetic objects simulated using the Material Point Method (MPM) and includes a variety of material types such as elastic bodies, plasticine, sand, and Newtonian and non-Newtonian fluids. This dataset provides 45 multi-view RGB video sequences that allow for the estimation of both object geometry and physical properties and approximately 14 frames per viewpoint. The Spring-Gaus dataset contains two subsets, namely, Synthetic and Real. The Synthetic subset focuses on the reconstruction and simulation of elastic objects, containing 30 frames in each of 10 viewpoints. The Real subset provides both dense-view static and sparse-view dynamic captures of elastic objects. We evaluate the performance of observable state simulation using both datasets, while the Spring-Gaus dataset is additionally utilized for future state prediction. To obtain object masks, we apply off-the-shelf matting with [35, 47]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 386, + 547, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 547, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 547, + 608 + ], + "type": "text", + "content": "Baselines. To assess the performance of our approach, we compare it against four state-of-the-art methods for system identification. PAC-NeRF [38] combines neural radiance fields with differentiable physics to jointly infer geometry and material properties. Spring-Gaus [100] integrates a spring-mass system into Gaussian representations for efficient elastic object simulation. NeuMA [8] fine-tunes a neural adapter on expert-designed physical models to better align reconstructed dynamics with visual observations. GIC [6] leverages a two-stage pipeline to obtain the geometry representation and physical properties sequentially. These baselines represent the most relevant prior work and provide a strong comparison for evaluating our method. We note that they all necessitate material priors. To be concrete, PAC-NeRF and GIC roughly categorize materials into five groups and adopt a certain set of physical models for each group. Spring-Gaus generally focuses on elastic objects, thereby being material-specific. And NeuMA leverages a scene-specific constitutive model as the material prior." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 611, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 547, + 704 + ], + "type": "text", + "content": "Metrics. We evaluate our method using several standard metrics to assess both geometric reconstruction quality and physical property estimation accuracy. Metrics including 1) Chamfer Distance (CD) [13, 57] with units expressed in " + }, + { + "bbox": [ + 313, + 611, + 547, + 704 + ], + "type": "inline_equation", + "content": "10^{3}\\mathrm{mm}^{2}" + }, + { + "bbox": [ + 313, + 611, + 547, + 704 + ], + "type": "text", + "content": "; 2) Peak Signal-to-Noise Ratio (PSNR) [24]; 3) Structural Similarity Index Measure (SSIM) [84]. The CD quantifies the similarity between two point distributions, while PSNR and SSIM assess the visual quality." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5948" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 79, + 547, + 155 + ], + "blocks": [ + { + "bbox": [ + 63, + 79, + 547, + 155 + ], + "lines": [ + { + "bbox": [ + 63, + 79, + 547, + 155 + ], + "spans": [ + { + "bbox": [ + 63, + 79, + 547, + 155 + ], + "type": "table", + "html": "
MethodPAC-NeRFSpring-Gaus
NewtonianNon-NewtonianElasticityPlasticineSandMeanTorusCrossCreamApplePasteChessBananaMean
\\(\\rightarrow\\)DPAC-NeRF [38]0.2770.2360.2380.4290.2120.2784.921.100.771.113.140.962.772.11
Spring-Gaus [100]------0.170.480.360.380.191.802.600.85
NeuMA [8]------4.590.060.080.040.710.050.030.79
GIC [6]0.2430.1950.1780.1960.2500.2120.130.130.140.150.170.410.030.17
MASIV (Ours)0.2330.1980.1920.2010.2290.2100.080.100.190.160.130.190.080.13
", + "image_path": "02b29894ae1f136d9bac105ae91d7c34e71d8b629004350d1ef42a3d9c400834.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 194, + 295, + 297 + ], + "blocks": [ + { + "bbox": [ + 61, + 162, + 546, + 184 + ], + "lines": [ + { + "bbox": [ + 61, + 162, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 162, + 546, + 184 + ], + "type": "text", + "content": "Table 1. Observable state simulation on PAC-NeRF and Spring-Gaus Synthetic datasets. Despite being the only technique that does not use expert-designed physical models (noted by the gray row), MASIV performs better than or on par with existing baselines." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 194, + 295, + 297 + ], + "lines": [ + { + "bbox": [ + 65, + 194, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 65, + 194, + 295, + 297 + ], + "type": "table", + "html": "
MethodBunBurgerDogPigPotatoMean
PSNR ↑Spring-Gaus [100]30.6934.0132.1034.9732.7232.90
NeuMA [8]31.2723.7825.6125.40-26.51
GIC [6]34.6840.4537.1738.3239.9538.11
Ours40.2642.1238.9740.8443.4041.12
SSIM ↑Spring-Gaus0.9920.9940.9940.9960.9920.994
NeuMA0.9940.9930.9950.995-0.994
GIC0.9950.9970.9970.9970.9960.996
Ours0.9980.9970.9970.9980.9980.997
", + "image_path": "4f3ac5522d5d199ba855db3dc67322054f0bca2332dc7ec93450823ea4bb5862.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "text", + "content": "Implementation details. In dynamic reconstruction, we follow [6] to optimize canonical Gaussians " + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^*" + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "text", + "content": " and the deformation network " + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "text", + "content": " for 40K iterations. Then the deformation network is fine-tuned to estimate particle trajectories from filled continuum sequences for 10K iterations. In system identification, we set the simulation time interval " + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "inline_equation", + "content": "1/200" + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "text", + "content": " the frame duration. We adopt the RAdam [52] optimizer to optimize the initial velocity " + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_0" + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "text", + "content": " for 100 steps and constitutive model parameters " + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "inline_equation", + "content": "\\theta^e, \\theta^p" + }, + { + "bbox": [ + 61, + 338, + 295, + 454 + ], + "type": "text", + "content": " for 1000 steps. All experiments are conducted on a single NVIDIA A100 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 461, + 176, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 461, + 176, + 473 + ], + "spans": [ + { + "bbox": [ + 62, + 461, + 176, + 473 + ], + "type": "text", + "content": "4.2. Comparison Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 480, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 480, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 61, + 480, + 295, + 597 + ], + "type": "text", + "content": "Observable state simulation. In Tab. 1, we compare MASIV with existing vision-based system identification methods for observable state simulation. The results indicate that MASIV outperforms all other methods in terms of CD, highlighting its effectiveness in reconstructing object dynamics across diverse materials. Further, Tab. 2 evaluates on real data. Thanks to the more flexible constitutive representation, MASIV achieves state-of-the-art results. This demonstrates the potential of MASIV as an adaptive and generalizable approach for system identification." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 600, + 295, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 600, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 600, + 295, + 704 + ], + "type": "text", + "content": "Future state simulation. In Tab. 3, we provide a detailed evaluation of future state simulation performance. The results indicate that MASIV consistently outperforms all other methods while slightly underperforming GIC. Since NeuMA [8] does not originally experiment on the synthetic subset of Spring-Gaus, we report our reproduction results. To be concrete, we use the closest pre-trained model, Jelly, from NCLaw [59] as the material prior and conduct LoRA [25] fine-tuning with hyperparameters from its orig-" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 194, + 547, + 332 + ], + "blocks": [ + { + "bbox": [ + 61, + 304, + 295, + 326 + ], + "lines": [ + { + "bbox": [ + 61, + 304, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 61, + 304, + 295, + 326 + ], + "type": "text", + "content": "Table 2. Observable state simulation on Spring-Gaus Real dataset. MASIV achieves state-of-the-art performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 194, + 547, + 332 + ], + "lines": [ + { + "bbox": [ + 316, + 194, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 547, + 332 + ], + "type": "table", + "html": "
MethodTorusCrossCreamApplePasteChessBananaMean
CD\\(\\rightarrow\\)PAC-NeRF [38]2.473.872.214.6937.708.2066.4317.94
Spring-Gaus [100]2.381.572.221.877.032.5918.485.16
NeuMA [8]1.170.401147.40.0544.260.330.30170.56
GIC [6]0.751.090.940.222.790.770.120.95
MASIV (Ours)0.160.381.512.641.751.700.241.20
PSNR\\(\\dagger\\)PAC-NeRF [38]17.4614.1515.3719.9412.3215.0816.0415.77
Spring-Gaus [100]16.8316.9315.4221.5514.7116.0817.8917.06
NeuMA [8]16.6119.672.3226.2513.4220.8523.1617.47
GIC [6]20.2430.5119.1526.8916.3118.4429.2922.98
MASIV (Ours)22.4233.7316.4521.1915.7317.2826.5621.91
SSIM\\(\\dagger\\)PAC-NeRF [38]0.9190.9060.8580.8780.8190.8480.8860.870
Spring-Gaus [100]0.9190.9400.8620.9020.8720.8810.9040.897
NeuMA [8]0.9420.9480.8890.9640.8890.9330.9640.933
GIC [6]0.9420.9390.9090.9480.8940.9120.9640.930
MASIV (Ours)0.9550.9670.8520.8980.8910.8780.9650.915
", + "image_path": "2642955ee551d28255fa6522410aaa9a0e3bbb68e1dd68ade0807f7e6315f173.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 340, + 547, + 362 + ], + "lines": [ + { + "bbox": [ + 313, + 340, + 547, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 547, + 362 + ], + "type": "text", + "content": "Table 3. Future state simulation on Spring-Gaus Synthetic dataset. MASIV outperforms all baselines except for GIC." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 312, + 376, + 547, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 376, + 547, + 620 + ], + "spans": [ + { + "bbox": [ + 312, + 376, + 547, + 620 + ], + "type": "text", + "content": "inal implementation. However, we observe unstable optimization processes for most categories. This is probably rooted in its sparse pixel supervision, which only operates once per frame. With dense geometrical constraints supervising hundreds of times per frame, MASIV achieves significantly better results on both geometrical accuracy and rendering quality. This suggests that MASIV effectively captures the underlying dynamics governing future states, allowing it to generate predictions that remain structurally and visually consistent over time. Unlike baseline approaches that rely on predefined material priors, MASIV demonstrates the ability to infer relevant properties directly from visual observations, making it more adaptable to a wide range of scenarios. On the other hand, while dense trajectory guidance enables overfitting to observations, the lack of category-level regularization, unlike hand-crafted constitutive models, limits its generalization under scarce data. This leads to inferior performance than GIC, which can benefit from its pre-known constitutive model as an oracle. More analysis on the enhanced generalization ability of MASIV with more data is in Sec. C of Supp. Mat." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 628, + 405, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 405, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 405, + 641 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 646, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 646, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 646, + 547, + 704 + ], + "type": "text", + "content": "We investigate the impact of geometrical objectives here, while referring the readers to more ablation studies in Sec. B of Supp. Mat. Fig. 3 validates the effectiveness of dense trajectory supervision. As shown in Fig. 4, we evaluate three geometrical objectives by replacing the " + }, + { + "bbox": [ + 312, + 646, + 547, + 704 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 312, + 646, + 547, + 704 + ], + "type": "text", + "content": " term in opti" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5949" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 83, + 541, + 190 + ], + "blocks": [ + { + "bbox": [ + 77, + 83, + 541, + 190 + ], + "lines": [ + { + "bbox": [ + 77, + 83, + 541, + 190 + ], + "spans": [ + { + "bbox": [ + 77, + 83, + 541, + 190 + ], + "type": "image", + "image_path": "a8d08b06a43e59ee3dad00c5bdeb89db87af2e923380dcfed7a22b675df5e87a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 201, + 548, + 234 + ], + "lines": [ + { + "bbox": [ + 61, + 201, + 548, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 201, + 548, + 234 + ], + "type": "text", + "content": "Figure 3. Ablation study of geometrical objectives " + }, + { + "bbox": [ + 61, + 201, + 548, + 234 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 61, + 201, + 548, + 234 + ], + "type": "text", + "content": " on PAC-NeRF dataset. We compare three types of geometrical objectives against a baseline without geometrical guidance and plot their simulation errors using Chamfer Distance (CD). Overall, dense trajectory guidance achieves the lowest and most stable errors. The figure is best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 240, + 296, + 284 + ], + "blocks": [ + { + "bbox": [ + 62, + 240, + 296, + 284 + ], + "lines": [ + { + "bbox": [ + 62, + 240, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 296, + 284 + ], + "type": "image", + "image_path": "7567d0477707b026ddf94e294ef0df503a49d1bbaf64417e9ecf864da56cae28.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 287, + 143, + 297 + ], + "lines": [ + { + "bbox": [ + 100, + 287, + 143, + 297 + ], + "spans": [ + { + "bbox": [ + 100, + 287, + 143, + 297 + ], + "type": "text", + "content": "(a) Surfaces" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 213, + 287, + 269, + 297 + ], + "lines": [ + { + "bbox": [ + 213, + 287, + 269, + 297 + ], + "spans": [ + { + "bbox": [ + 213, + 287, + 269, + 297 + ], + "type": "text", + "content": "(b) Continuums" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 62, + 304, + 297, + 347 + ], + "blocks": [ + { + "bbox": [ + 62, + 304, + 297, + 347 + ], + "lines": [ + { + "bbox": [ + 62, + 304, + 297, + 347 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 297, + 347 + ], + "type": "image", + "image_path": "1f45a87c5acba9c8fcf011c8e9bebe3196d7968da1d2c1f1bb1aaa165547e13f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 351, + 156, + 361 + ], + "lines": [ + { + "bbox": [ + 88, + 351, + 156, + 361 + ], + "spans": [ + { + "bbox": [ + 88, + 351, + 156, + 361 + ], + "type": "text", + "content": "(c) Gaussian Splats" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 214, + 351, + 268, + 361 + ], + "lines": [ + { + "bbox": [ + 214, + 351, + 268, + 361 + ], + "spans": [ + { + "bbox": [ + 214, + 351, + 268, + 361 + ], + "type": "text", + "content": "(d) Trajectories" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 430, + 296, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 430, + 296, + 477 + ], + "spans": [ + { + "bbox": [ + 61, + 430, + 296, + 477 + ], + "type": "text", + "content": "mizing the neural constitutive model objective, Eq. (8). We define them formally as follows. (1) Surface: per-frame surface alignment between the simulated continuum and the filled continuum," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 85, + 483, + 295, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 483, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 85, + 483, + 295, + 503 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s u r f}} = \\mathcal {L} _ {\\mathrm {C D}} \\left(\\left\\{\\mathcal {S} (\\hat {\\mathbf {x}} _ {t}) \\right\\} _ {t = 0} ^ {T}, \\left\\{\\mathcal {S} (\\mathbf {p} _ {t}) \\right\\} _ {t = 0} ^ {T}\\right), \\tag {11}", + "image_path": "972c9618f3d4005d0be7335195ad73a17d58a2d2190e89742fc20ca837e55a7a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 509, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 509, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 61, + 509, + 296, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 509, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 61, + 509, + 296, + 544 + ], + "type": "text", + "content": " denotes the surface extraction by thresholding Gaussian opacities [6]. (2) Continuum: similar to Surface while between solid continuums instead of surfaces," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 108, + 551, + 295, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 551, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 108, + 551, + 295, + 571 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n t}} = \\mathcal {L} _ {\\mathrm {C D}} \\left(\\left\\{\\hat {\\mathbf {x}} _ {t} \\right\\} _ {t = 0} ^ {T}, \\left\\{\\mathbf {p} _ {t} \\right\\} _ {t = 0} ^ {T}\\right). \\tag {12}", + "image_path": "ab12e0cfc83d3bb6c9bc02f125f47336a50fe765b9689bc92cb3290d8c23e20d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 576, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 576, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 576, + 296, + 704 + ], + "type": "text", + "content": "(3) Trajectory: per-simulation-step trajectory alignment between the simulated continuum " + }, + { + "bbox": [ + 61, + 576, + 296, + 704 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{x}}_t\\}_{t=0}^{N\\tau}" + }, + { + "bbox": [ + 61, + 576, + 296, + 704 + ], + "type": "text", + "content": " and the continuum predicted by the deformation network " + }, + { + "bbox": [ + 61, + 576, + 296, + 704 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_t\\}_{t=0}^{N\\tau}" + }, + { + "bbox": [ + 61, + 576, + 296, + 704 + ], + "type": "text", + "content": ", as defined in Eq. (9). By analyzing per-frame errors, we first confirm the necessity of geometrical guidance, as the baseline (red curve) exhibits significantly higher CD errors across all material types. Surface-based supervision (yellow curve) shows moderate error reduction but remains less stable over time, especially in Plasticine and Sand. Continuum-based supervision (blue curve) further reduces errors but struggles with certain materials, e.g., Sand. These error overshoots" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 313, + 255, + 536, + 457 + ], + "blocks": [ + { + "bbox": [ + 61, + 361, + 296, + 416 + ], + "lines": [ + { + "bbox": [ + 61, + 361, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 61, + 361, + 296, + 416 + ], + "type": "text", + "content": "Figure 4. Visualization of geometrical objectives " + }, + { + "bbox": [ + 61, + 361, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 61, + 361, + 296, + 416 + ], + "type": "text", + "content": " on PAC-NeRF Bird. The fine-tuned deformation network " + }, + { + "bbox": [ + 61, + 361, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{D}'" + }, + { + "bbox": [ + 61, + 361, + 296, + 416 + ], + "type": "text", + "content": " captures smooth and accurate motion for continuums, including low-frequency translation of the body and high-frequency swing of the tail. We space adjacent frames in the horizontal axis for clarity." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 240, + 548, + 254 + ], + "lines": [ + { + "bbox": [ + 314, + 240, + 548, + 254 + ], + "spans": [ + { + "bbox": [ + 314, + 240, + 548, + 254 + ], + "type": "text", + "content": "Ground Truth Surface " + }, + { + "bbox": [ + 314, + 240, + 548, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{surf}}" + }, + { + "bbox": [ + 314, + 240, + 548, + 254 + ], + "type": "text", + "content": " Continuum " + }, + { + "bbox": [ + 314, + 240, + 548, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cont}}" + }, + { + "bbox": [ + 314, + 240, + 548, + 254 + ], + "type": "text", + "content": " Trajectory " + }, + { + "bbox": [ + 314, + 240, + 548, + 254 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{traj}}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 255, + 536, + 457 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 536, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 536, + 457 + ], + "type": "image", + "image_path": "df93c4895806c327cddbdc010c34deef9c591e70875751216b002522252b80f6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 464, + 547, + 518 + ], + "lines": [ + { + "bbox": [ + 313, + 464, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 547, + 518 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of geometrical objectives " + }, + { + "bbox": [ + 313, + 464, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 313, + 464, + 547, + 518 + ], + "type": "text", + "content": " on PAC-NeRF Cream. Using sparse surfaces or solid continuum as geometrical cues can lead to physically implausible deformations, while trajectory guidance mitigates these issues by introducing temporally denser constraints. Better viewed when zoomed in." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 532, + 548, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 532, + 548, + 613 + ], + "spans": [ + { + "bbox": [ + 312, + 532, + 548, + 613 + ], + "type": "text", + "content": "are probably due to supervising relatively complex neural constitutive models with sparse signals, which might lead to overfitting since the simulated behaviors between observed discrete time steps are not well constrained. Finally, by enforcing dense alignment across simulation steps, trajectory-based supervision (green curve) achieves the lowest and most stable errors across all materials." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 624, + 429, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 429, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 429, + 637 + ], + "type": "text", + "content": "4.4. Qualitative Analysis" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 646, + 548, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 646, + 548, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 646, + 548, + 704 + ], + "type": "text", + "content": "Dynamics reconstruction. We visualize the reconstructed Cream object by using the alternative objectives in Fig. 5. Here, we observe physically implausible deformations in results for surface and continuum supervision. Under uniform force, the object exhibits inconsistent distortion, espe" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5950" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 78, + 298, + 257 + ], + "blocks": [ + { + "bbox": [ + 63, + 78, + 298, + 257 + ], + "lines": [ + { + "bbox": [ + 63, + 78, + 298, + 257 + ], + "spans": [ + { + "bbox": [ + 63, + 78, + 298, + 257 + ], + "type": "image", + "image_path": "42886a4924893fee5eba6c102b9c3f58652849db4c7b2ef5dfe4fda829f36fd2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 264, + 296, + 308 + ], + "lines": [ + { + "bbox": [ + 61, + 264, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 264, + 296, + 308 + ], + "type": "text", + "content": "Figure 6. Intra-class generalization. We first optimize the geometry for PAC-NeRF Bird and then use the constitutive models optimized for reference elastic objects to predict novel dynamic behaviors, thereby softening or hardening the Bird respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 328, + 296, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 328, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 61, + 328, + 296, + 503 + ], + "type": "text", + "content": "cially in the boundary between blue stripes and red stripes. We attribute this to two main reasons. 1) Chamfer Distance based geometrical constraint focuses mainly on the overall structure. While it encourages alignment between simulated and reference shapes, they do not enforce point-wise correspondence. A single particle can thus be matched with different ground truth particles in different time steps. As a result, local mismatches can accumulate, leading to inaccurate deformations. 2) Sparse supervision fails to constrain intermediate motion dynamics. Surface and continuum objectives provide alignment only at discrete time steps, leaving large gaps in supervision. Without explicit constraints on how the object deforms over time, the neural constitutive model may interpolate arbitrarily, leading to unstable and unrealistic motion patterns." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 506, + 296, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 506, + 296, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 506, + 296, + 705 + ], + "type": "text", + "content": "Dynamics generalization. We further study the generalization ability for MASIV. Fig. 6 examines how MASIV can be used to modify the dynamic characteristics of a single object (PAC-NeRF Bird). Specifically, we first reconstruct the geometrical representation of the Bird and optimize neural constitutive models for two elastic cross-shaped objects. These optimized constitutive models are then used as references to drive the simulation of the Bird, which deforms accordingly, exhibiting fluid-like softness on the left and increased rigidity and bounce on the right. Additionally, we see continuous transitions between soft and rigid behaviors when interpolating the neural constitutive model weights. In Fig. 7, we extend this study to inter-class generalization by applying neural constitutive models trained on different materials to a novel object, the PAC-NeRF Trophy. The results show that despite sharing the same initial state, each material model induces distinct deformation patterns. These" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 315, + 78, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 315, + 78, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 315, + 78, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 315, + 78, + 545, + 257 + ], + "type": "image", + "image_path": "96852bb1d70c3f30255e28d3469a5bfa833545729202d706c56af8b9aca9bce7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 264, + 547, + 307 + ], + "lines": [ + { + "bbox": [ + 313, + 264, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 547, + 307 + ], + "type": "text", + "content": "Figure 7. Inter-class generalization. We train multiple neural constitutive models with different materials and apply them to the PAC-NeRF Trophy. While sharing the same initial state, the Trophies with different materials exhibit distinct behaviors over time." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 326, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 547, + 361 + ], + "type": "text", + "content": "observations confirm that MASIV effectively captures generalizable intrinsic dynamics. More qualitative results can be found in Sec. D of Supp. Mat." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 367, + 384, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 384, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 384, + 378 + ], + "type": "text", + "content": "4.5. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 384, + 547, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 384, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 312, + 384, + 547, + 524 + ], + "type": "text", + "content": "While MASIV excels in material-agnostic system identification, several limitations remain. First, it relies on multiview video input, restricting applicability to single-view or sparse observations. Second, our approach assumes that internal geometries and deformations can be inferred from surface motion, following a spatially coherent model. This assumption may break down in materials with complex internal flows, such as granular media or multiphase fluids. Finally, MASIV does not explicitly disentangle material properties from external forces in system identification, assuming a uniform gravity field for all objects. Future works aim to relax these assumptions for broader applicability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 384, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 384, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 384, + 545 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 553, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 553, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 553, + 547, + 704 + ], + "type": "text", + "content": "We present MASIV, a material-agnostic system identification framework. By integrating neural constitutive models with dense geometry guidance, MASIV effectively captures object dynamics across diverse materials. Our approach reconstructs dynamic Gaussians from multi-view videos and fine-tunes a deformation model within a solid continuum. This enables us to extract dense motion cues, which, combined with visual observations, guide the learning of neural constitutive models. Through this formulation, MASIV generalizes beyond traditional system identification methods that rely on pre-defined material priors. Extensive experiments highlight its ability to reconstruct observable dynamics and to generalize for novel interactions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 314, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 314, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 314, + 743 + ], + "type": "text", + "content": "5951" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 80, + 153, + 93 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 80, + 153, + 93 + ], + "spans": [ + { + "bbox": [ + 63, + 80, + 153, + 93 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 99, + 296, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 99, + 296, + 154 + ], + "spans": [ + { + "bbox": [ + 61, + 99, + 296, + 154 + ], + "type": "text", + "content": "This work was supported in part by U.S. NSF grant DBI-2238093, and in part by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 174, + 119, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 174, + 119, + 186 + ], + "spans": [ + { + "bbox": [ + 63, + 174, + 119, + 186 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 193, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 193, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 193, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 72, + 193, + 296, + 258 + ], + "type": "text", + "content": "[1] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16610-16620, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 258, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 258, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 72, + 258, + 296, + 280 + ], + "type": "text", + "content": "[2] Genesis Authors. Genesis: A universal and generative physics engine for robotics and beyond, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 281, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 281, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 72, + 281, + 296, + 354 + ], + "type": "text", + "content": "[3] Sherwin Bahmani, Ivan Skorokhodov, Victor Rong, Gordon Wetzstein, Leonidas Guibas, Peter Wonka, Sergey Tulyakov, Jeong Joon Park, Andrea Tagliasacchi, and David B Lindell. 4d-fy: Text-to-4d generation using hybrid score distillation sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7996-8006, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 72, + 356, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 356, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 72, + 356, + 294, + 399 + ], + "type": "text", + "content": "[4] Piotr Borycki, Weronika Smolak, Joanna Waczyńska, Marcin Mazur, Sławomir Tadeja, and Przemysław Spurek. Gasp: Gaussian splatting for physic-based simulations. arXiv preprint arXiv:2409.05819, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 72, + 399, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 399, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 72, + 399, + 296, + 431 + ], + "type": "text", + "content": "[5] Johannes Brandstetter, Daniel Worrall, and Max Welling. Message passing neural pde solvers. arXiv preprint arXiv:2202.03376, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 72, + 433, + 296, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 433, + 296, + 485 + ], + "spans": [ + { + "bbox": [ + 72, + 433, + 296, + 485 + ], + "type": "text", + "content": "[6] Junhao Cai, Yuji Yang, Weihao Yuan, Yisheng He, Zilong Dong, Liefeng Bo, Hui Cheng, and Qifeng Chen. Gic: Gaussian-informed continuum for physical property identification and simulation. arXiv preprint arXiv:2406.14927, 2024. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 72, + 487, + 296, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 487, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 72, + 487, + 296, + 539 + ], + "type": "text", + "content": "[7] Yuanhao Cai, Jiahao Wang, Alan Yuille, Zongwei Zhou, and Angtian Wang. Structure-aware sparse-view x-ray 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11174-11183, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 72, + 541, + 296, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 541, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 72, + 541, + 296, + 594 + ], + "type": "text", + "content": "[8] Junyi Cao, Shanyan Guan, Yanhao Ge, Wei Li, Xiaokang Yang, and Chao Ma. Neuma: Neural material adaptor for visual grounding of intrinsic dynamics. Advances in Neural Information Processing Systems, 37:65643-65669, 2025. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 72, + 595, + 296, + 637 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 595, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 72, + 595, + 296, + 637 + ], + "type": "text", + "content": "[9] Pradyumna Chari, Chinmay Talegaonkar, Yunhao Ba, and Achuta Kadambi. Visual physics: Discovering physical laws from videos. arXiv preprint arXiv:1911.11893, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 639, + 294, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 639, + 294, + 671 + ], + "spans": [ + { + "bbox": [ + 67, + 639, + 294, + 671 + ], + "type": "text", + "content": "[10] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splattering guidance. arXiv preprint arXiv:2312.00846, 2023. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 672, + 296, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 672, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 67, + 672, + 296, + 704 + ], + "type": "text", + "content": "[11] Hsiao-yu Chen, Edith Tretschk, Tuur Stuyck, Petr Kadlecek, Ladislav Kavan, Etienne Vouga, and Christoph Lassner. Virtual elastic objects. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 319, + 82, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 338, + 82, + 546, + 103 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 82, + 546, + 103 + ], + "spans": [ + { + "bbox": [ + 338, + 82, + 546, + 103 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 15827-15837, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 104, + 547, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 104, + 547, + 146 + ], + "spans": [ + { + "bbox": [ + 319, + 104, + 547, + 146 + ], + "type": "text", + "content": "[12] Tao Du, Kui Wu, Pingchuan Ma, Sebastien Wah, Andrew Spielberg, Daniela Rus, and Wojciech Matusik. Diffpd: Differentiable projective dynamics. ACM Transactions on Graphics (TOG), 41(2):1-21, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 147, + 547, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 147, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 319, + 147, + 547, + 190 + ], + "type": "text", + "content": "[13] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In European Conference on Computer Vision, pages 108-124. Springer, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 190, + 547, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 190, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 319, + 190, + 547, + 243 + ], + "type": "text", + "content": "[14] Yutao Feng, Xiang Feng, Yintong Shang, Ying Jiang, Chang Yu, Zeshun Zong, Tianjia Shao, Hongzhi Wu, Kun Zhou, Chenfanfu Jiang, et al. Gaussian splashing: Unified particles for versatile motion synthesis and rendering. arXiv preprint arXiv:2401.15318, 3, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 319, + 243, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 243, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 319, + 243, + 547, + 297 + ], + "type": "text", + "content": "[15] Yutao Feng, Yintong Shang, Xuan Li, Tianjia Shao, Chenfanfu Jiang, and Yin Yang. Pie-nerf: Physics-based interactive elastodynamics with nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4450–4461, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 319, + 297, + 547, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 297, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 319, + 297, + 547, + 350 + ], + "type": "text", + "content": "[16] Zhoujie Fu, Jiacheng Wei, Wenhao Shen, Chaoyue Song, Xiaofeng Yang, Fayao Liu, Xulei Yang, and Guosheng Lin. Sync4d: Video guided controllable dynamics for physics-based 4d generation. arXiv preprint arXiv:2405.16849, 2024.3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 319, + 350, + 547, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 350, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 319, + 350, + 547, + 403 + ], + "type": "text", + "content": "[17] Moritz Geilinger, David Hahn, Jonas Zehnder, Moritz Bächer, Bernhard Thomaszewski, and Stelian Coros. Add: Analytically differentiable dynamics for multi-body systems with frictional contact. ACM Transactions on Graphics (TOG), 39(6):1-15, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 319, + 404, + 547, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 404, + 547, + 458 + ], + "spans": [ + { + "bbox": [ + 319, + 404, + 547, + 458 + ], + "type": "text", + "content": "[18] Daniel Geng, Charles Herrmann, Junhwa Hur, Forrester Cole, Serena Zhang, Tobias Pfaff, Tatiana Lopez-Guevara, Carl Doersch, Yusuf Aytar, Michael Rubinstein, et al. Motion prompting: Controlling video generation with motion trajectories. arXiv preprint arXiv:2412.02700, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 319, + 458, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 458, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 319, + 458, + 547, + 510 + ], + "type": "text", + "content": "[19] Sharath Girish, Tianye Li, Amrita Mazumdar, Abhinav Shrivastava, David Luebke, and Shalini De Mello. Queen: Quantized efficient encoding of dynamic gaussians for streaming free-viewpoint videos. arXiv preprint arXiv:2412.04469, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 319, + 511, + 547, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 511, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 319, + 511, + 547, + 553 + ], + "type": "text", + "content": "[20] Shanyan Guan, Huayu Deng, Yunbo Wang, and Xiaokang Yang. Neurofluid: Fluid dynamics grounding with particle-driven neural radiance fields. In International conference on machine learning, pages 7919-7929. PMLR, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 319, + 554, + 547, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 554, + 547, + 597 + ], + "spans": [ + { + "bbox": [ + 319, + 554, + 547, + 597 + ], + "type": "text", + "content": "[21] Eric Heiden, Miles Macklin, Yashraj Narang, Dieter Fox, Animesh Garg, and Fabio Ramos. Disect: A differentiable simulation engine for autonomous robotic cutting. arXiv preprint arXiv:2105.12244, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 319, + 597, + 547, + 650 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 597, + 547, + 650 + ], + "spans": [ + { + "bbox": [ + 319, + 597, + 547, + 650 + ], + "type": "text", + "content": "[22] Maximilian Herde, Bogdan Raonic, Tobias Rohner, Roger Käppeli, Roberto Molinaro, Emmanuel de Bézenac, and Siddhartha Mishra. Poseidon: Efficient foundation models for pdes. Advances in Neural Information Processing Systems, 37:72525-72624, 2025. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 319, + 651, + 547, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 651, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 319, + 651, + 547, + 704 + ], + "type": "text", + "content": "[23] Florian Hofherr, Lukas Koestler, Florian Bernard, and Daniel Cremers. Neural implicit representations for physical parameter inference from a single video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2093-2103, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 734, + 315, + 743 + ], + "type": "text", + "content": "5952" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 82, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 82, + 296, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 82, + 296, + 114 + ], + "spans": [ + { + "bbox": [ + 67, + 82, + 296, + 114 + ], + "type": "text", + "content": "[24] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 115, + 295, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 115, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 67, + 115, + 295, + 156 + ], + "type": "text", + "content": "[25] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 158, + 295, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 158, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 67, + 158, + 295, + 211 + ], + "type": "text", + "content": "[26] Yuanming Hu, Yu Fang, Ziheng Ge, Ziyin Qu, Yixin Zhu, Andre Pradhana, and Chenfanfu Jiang. A moving least squares material point method with displacement discontinuity and two-way rigid body coupling. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 212, + 295, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 212, + 295, + 254 + ], + "spans": [ + { + "bbox": [ + 67, + 212, + 295, + 254 + ], + "type": "text", + "content": "[27] Yuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun, Nathan Carr, Jonathan Ragan-Kelley, and Fredo Durand. Difftaichi: Differentiable programming for physical simulation. arXiv preprint arXiv:1910.00935, 2019. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 255, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 255, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 67, + 255, + 295, + 306 + ], + "type": "text", + "content": "[28] Tianyu Huang, Haoze Zhang, Yihan Zeng, Zhilu Zhang, Hui Li, Wangmeng Zuo, and Rynson WH Lau. Dreamphysics: Learning physical properties of dynamic 3d gaussians with video diffusion priors. arXiv preprint arXiv:2406.01476, 2024.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 308, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 308, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 67, + 308, + 295, + 350 + ], + "type": "text", + "content": "[29] Zhiao Huang, Yuanming Hu, Tao Du, Siyuan Zhou, Hao Su, Joshua B Tenenbaum, and Chuang Gan. Plasticinelab: A soft-body manipulation benchmark with differentiable physics. arXiv preprint arXiv:2104.03311, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 351, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 351, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 67, + 351, + 295, + 415 + ], + "type": "text", + "content": "[30] Krishna Murthy Jatavallabhula, Miles Macklin, Florian Golemo, Vikram Voleti, Linda Petrini, Martin Weiss, Breandan Considine, Jérôme Parent-Lévesque, Kevin Xie, Kenny Erleben, et al. gradsim: Differentiable simulation for system identification and visuomotor control. arXiv preprint arXiv:2104.02646, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 415, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 415, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 67, + 415, + 295, + 458 + ], + "type": "text", + "content": "[31] Chenfanfu Jiang, Craig Schroeder, Joseph Teran, Alexey Stomakhin, and Andrew Selle. The material point method for simulating continuum materials. In ACM SIGGRAPH 2016 Courses, pages 1-52, 2016. 2, 3, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 459, + 295, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 459, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 67, + 459, + 295, + 511 + ], + "type": "text", + "content": "[32] Takuhiro Kaneko. Improving physics-augmented continuum neural radiance field-based geometry-agnostic system identification with lagrangian particle optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470-5480, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 512, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 512, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 67, + 512, + 295, + 553 + ], + "type": "text", + "content": "[33] Namgyu Kang, Jaemin Oh, Youngjoon Hong, and Eunbyung Park. Pig: Physics-informed gaussians as adaptive parametric mesh representations. arXiv preprint arXiv:2412.05994, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 555, + 295, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 555, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 67, + 555, + 295, + 596 + ], + "type": "text", + "content": "[34] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 597, + 295, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 295, + 651 + ], + "type": "text", + "content": "[35] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 651, + 295, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 651, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 67, + 651, + 295, + 704 + ], + "type": "text", + "content": "[36] Agelos Kratimenos, Jiahui Lei, and Kostas Daniilidis. Dynmf: Neural motion factorization for real-time dynamic view synthesis with 3d gaussian splattering. In European Conference on Computer Vision, pages 252-269. Springer, 2025. 3, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 82, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 320, + 82, + 547, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 82, + 547, + 124 + ], + "spans": [ + { + "bbox": [ + 320, + 82, + 547, + 124 + ], + "type": "text", + "content": "[37] Jinxi Li, Ziyang Song, and Bo Yang. Nvfi: neural velocity fields for 3d physics learning from dynamic videos. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 125, + 547, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 125, + 547, + 188 + ], + "spans": [ + { + "bbox": [ + 320, + 125, + 547, + 188 + ], + "type": "text", + "content": "[38] Xuan Li, Yi-Ling Qiao, Peter Yichen Chen, Krishna Murthy Jatavallabhula, Ming Lin, Chenfanfu Jiang, and Chuang Gan. Pac-nerf: Physics augmented continuum neural radiance fields for geometry-agnostic system identification. arXiv preprint arXiv:2303.05512, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 190, + 547, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 190, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 320, + 190, + 547, + 232 + ], + "type": "text", + "content": "[39] Yunzhu Li, Jiajun Wu, Russ Tedrake, Joshua B Tenenbaum, and Antonio Torralba. Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids. arXiv preprint arXiv:1810.01566, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 233, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 233, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 320, + 233, + 547, + 274 + ], + "type": "text", + "content": "[40] Yifei Li, Tao Du, Kui Wu, Jie Xu, and Wojciech Matusik. Diffcloth: Differentiable cloth simulation with dry frictional contact. ACM Transactions on Graphics (TOG), 42 (1):1-20, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 276, + 547, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 276, + 547, + 328 + ], + "spans": [ + { + "bbox": [ + 320, + 276, + 547, + 328 + ], + "type": "text", + "content": "[41] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4273-4284, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 330, + 547, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 330, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 320, + 330, + 547, + 381 + ], + "type": "text", + "content": "[42] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8508-8520, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 383, + 547, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 383, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 320, + 383, + 547, + 425 + ], + "type": "text", + "content": "[43] Zhengqi Li, Richard Tucker, Noah Snavely, and Aleksander Holynski. Generative image dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24142-24153, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 426, + 547, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 426, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 320, + 426, + 547, + 479 + ], + "type": "text", + "content": "[44] Zongyi Li, Hongkai Zheng, Nikola Kovachki, David Jin, Haoxuan Chen, Burigede Liu, Kamyar Azizzadenesheli, and Anima Anandkumar. Physics-informed neural operator for learning partial differential equations. ACM/JMS Journal of Data Science, 1(3):1-27, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 479, + 547, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 479, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 320, + 479, + 547, + 511 + ], + "type": "text", + "content": "[45] Junbang Liang, Ming Lin, and Vladlen Koltun. Differentiable cloth simulation for inverse problems. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 512, + 547, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 512, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 320, + 512, + 547, + 575 + ], + "type": "text", + "content": "[46] Xiao Liang, Fei Liu, Yutong Zhang, Yuelei Li, Shan Lin, and Michael Yip. Real-to-sim deformable object manipulation: Optimizing physics models with residual mappings for robotic surgery. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 15471-15477. IEEE, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 576, + 547, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 576, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 320, + 576, + 547, + 639 + ], + "type": "text", + "content": "[47] Shanchuan Lin, Andrey Ryabtsev, Soumyadip Sengupta, Brian L Curless, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Real-time high-resolution background matting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8762-8771, 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 640, + 547, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 640, + 547, + 682 + ], + "spans": [ + { + "bbox": [ + 320, + 640, + 547, + 682 + ], + "type": "text", + "content": "[48] Yuchen Lin, Chenguo Lin, Jianjin Xu, and Yadong Mu. Omniphysgs: 3d constitutive gaussians for general physics-based dynamics generation. arXiv preprint arXiv:2501.18982, 2025. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 683, + 547, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 683, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 320, + 683, + 547, + 704 + ], + "type": "text", + "content": "[49] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 735, + 314, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 735, + 314, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 735, + 314, + 743 + ], + "type": "text", + "content": "5953" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 82, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 86, + 82, + 296, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 82, + 296, + 114 + ], + "spans": [ + { + "bbox": [ + 86, + 82, + 296, + 114 + ], + "type": "text", + "content": "with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576-8588, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 115, + 295, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 115, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 68, + 115, + 295, + 167 + ], + "type": "text", + "content": "[50] Phillip Lippe, Bas Veeling, Paris Perdikaris, Richard Turner, and Johannes Brandstetter. Pde-refiner: Achieving accurate long rollouts with neural pde solvers. Advances in Neural Information Processing Systems, 36:67398-67433, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 170, + 295, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 170, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 67, + 170, + 295, + 212 + ], + "type": "text", + "content": "[51] Fangfu Liu, Hanyang Wang, Shunyu Yao, Shengjun Zhang, Jie Zhou, and Yueqi Duan. Physics3d: Learning physical properties of 3d gaussians via video diffusion. arXiv preprint arXiv:2406.04338, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 214, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 214, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 68, + 214, + 294, + 255 + ], + "type": "text", + "content": "[52] Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond. arXiv preprint arXiv:1908.03265, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 258, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 258, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 68, + 258, + 295, + 300 + ], + "type": "text", + "content": "[53] Qingming Liu, Yuan Liu, Jiepeng Wang, Xianqiang Lv, Peng Wang, Wenping Wang, and Junhui Hou. Modgs: Dynamic gaussian splatting from causally-captured monocular videos. arXiv preprint arXiv:2406.00434, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 301, + 295, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 301, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 68, + 301, + 295, + 344 + ], + "type": "text", + "content": "[54] Shaowei Liu, Zhongzheng Ren, Saurabh Gupta, and Shenlong Wang. Physgen: Rigid-body physics-grounded imageto-video generation. In European Conference on Computer Vision, pages 360-378. Springer, 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 68, + 346, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 346, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 68, + 346, + 295, + 398 + ], + "type": "text", + "content": "[55] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splatting for dynamic view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8900-8910, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 68, + 399, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 399, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 68, + 399, + 295, + 441 + ], + "type": "text", + "content": "[56] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. arXiv preprint arXiv:2308.09713, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 68, + 444, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 444, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 68, + 444, + 295, + 485 + ], + "type": "text", + "content": "[57] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance functions from point clouds by learning to pull space onto surfaces. arXiv preprint arXiv:2011.13495, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 68, + 487, + 295, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 487, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 68, + 487, + 295, + 540 + ], + "type": "text", + "content": "[58] Pingchuan Ma, Tao Du, Joshua B Tenenbaum, Wojciech Matusik, and Chuang Gan. Risp: Rendering-invariant state predictor with differentiable simulation and rendering for cross-domain parameter estimation. arXiv preprint arXiv:2205.05678, 2022.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 68, + 542, + 295, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 542, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 68, + 542, + 295, + 605 + ], + "type": "text", + "content": "[59] Pingchuan Ma, Peter Yichen Chen, Bolei Deng, Joshua B Tenenbaum, Tao Du, Chuang Gan, and Wojciech Matusik. Learning neural constitutive laws from motion observations for generalizable pde dynamics. In International Conference on Machine Learning, pages 23279-23300. PMLR, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 68, + 607, + 295, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 607, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 68, + 607, + 295, + 659 + ], + "type": "text", + "content": "[60] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 68, + 662, + 295, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 662, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 68, + 662, + 295, + 704 + ], + "type": "text", + "content": "[61] Clinton A Mo, Kun Hu, Chengjiang Long, and Zhiyong Wang. Continuous intermediate token learning with implicit motion manifold for keyframe based motion interpolation. In Proceedings of the IEEE/CVF Conference on" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 82, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 339, + 82, + 547, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 82, + 547, + 102 + ], + "spans": [ + { + "bbox": [ + 339, + 82, + 547, + 102 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition, pages 13894-13903, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 104, + 547, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 104, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 320, + 104, + 547, + 156 + ], + "type": "text", + "content": "[62] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 158, + 547, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 158, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 320, + 158, + 547, + 210 + ], + "type": "text", + "content": "[63] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. arXiv preprint arXiv:2106.13228, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 212, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 212, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 212, + 547, + 262 + ], + "type": "text", + "content": "[64] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318-10327, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 264, + 547, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 264, + 547, + 306 + ], + "spans": [ + { + "bbox": [ + 320, + 264, + 547, + 306 + ], + "type": "text", + "content": "[65] Yiling Qiao, Junbang Liang, Vladlen Koltun, and Ming Lin. Differentiable simulation of soft multi-body systems. Advances in Neural Information Processing Systems, 34: 17123-17135, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 308, + 546, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 308, + 546, + 349 + ], + "spans": [ + { + "bbox": [ + 320, + 308, + 546, + 349 + ], + "type": "text", + "content": "[66] Yi-Ling Qiao, Alexander Gao, and Ming Lin. Neuphysics: Editable neural geometry and physics from monocular videos. Advances in Neural Information Processing Systems, 35:12841-12854, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 351, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 351, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 320, + 351, + 547, + 392 + ], + "type": "text", + "content": "[67] Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Language-driven physics-based scene synthesis and editing via feature splatting. In European Conference on Computer Vision, pages 368-383. Springer, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 394, + 546, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 394, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 320, + 394, + 546, + 445 + ], + "type": "text", + "content": "[68] Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational physics, 378:686-707, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 448, + 546, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 448, + 546, + 488 + ], + "spans": [ + { + "bbox": [ + 320, + 448, + 546, + 488 + ], + "type": "text", + "content": "[69] Jiawei Ren, Liang Pan, Jiaxiang Tang, Chi Zhang, Ang Cao, Gang Zeng, and Ziwei Liu. Dreamgaussian4d: Generative 4d gaussian splatting. arXiv preprint arXiv:2312.17142, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 490, + 546, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 490, + 546, + 542 + ], + "spans": [ + { + "bbox": [ + 320, + 490, + 546, + 542 + ], + "type": "text", + "content": "[70] Alvaro Sanchez-Gonzalez, Jonathan Godwin, Tobias Pfaff, Rex Ying, Jure Leskovec, and Peter Battaglia. Learning to simulate complex physics with graph networks. In International conference on machine learning, pages 8459-8468. PMLR, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 544, + 546, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 544, + 546, + 585 + ], + "spans": [ + { + "bbox": [ + 320, + 544, + 546, + 585 + ], + "type": "text", + "content": "[71] Yidi Shao, Mu Huang, Chen Change Loy, and Bo Dai. Gausim: Registering elastic objects into digital world by gaussian simulator. arXiv preprint arXiv:2412.17804, 2024.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 586, + 547, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 586, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 320, + 586, + 547, + 628 + ], + "type": "text", + "content": "[72] Haochen Shi, Huazhe Xu, Samuel Clarke, Yunzhu Li, and Jiajun Wu. Robocook: Long-horizon elasto-plastic object manipulation with diverse tools. arXiv preprint arXiv:2306.14447, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 630, + 547, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 630, + 547, + 681 + ], + "spans": [ + { + "bbox": [ + 320, + 630, + 547, + 681 + ], + "type": "text", + "content": "[73] Haochen Shi, Huazhe Xu, Zhiao Huang, Yunzhu Li, and Ji-jun Wu. Robocraft: Learning to see, simulate, and shape elasto-plastic objects in 3d with graph networks. The International Journal of Robotics Research, 43(4):533-549, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 682, + 547, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 682, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 320, + 682, + 547, + 704 + ], + "type": "text", + "content": "[74] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 735, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 735, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 735, + 315, + 743 + ], + "type": "text", + "content": "5954" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 82, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 86, + 82, + 296, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 82, + 296, + 125 + ], + "spans": [ + { + "bbox": [ + 86, + 82, + 296, + 125 + ], + "type": "text", + "content": "Cheung, Simon See, Hongwei Qin, et al. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 125, + 295, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 125, + 295, + 176 + ], + "spans": [ + { + "bbox": [ + 67, + 125, + 295, + 176 + ], + "type": "text", + "content": "[75] Uriel Singer, Shelly Sheynin, Adam Polyak, Oron Ashual, Iurii Makarov, Filippos Kokkinos, Naman Goyal, Andrea Vedaldi, Devi Parikh, Justin Johnson, et al. Text-to-4d dynamic scene generation. arXiv preprint arXiv:2301.11280, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 178, + 295, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 178, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 178, + 295, + 232 + ], + "type": "text", + "content": "[76] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 233, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 233, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 67, + 233, + 294, + 274 + ], + "type": "text", + "content": "[77] Alexey Stomakhin, Craig Schroeder, Lawrence Chai, Joseph Teran, and Andrew Selle. A material point method for snow simulation. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 276, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 276, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 67, + 276, + 294, + 317 + ], + "type": "text", + "content": "[78] Deborah Sulsky, Shi-Jian Zhou, and Howard L Schreyer. Application of a particle-in-cell method to solid mechanics. Computer physics communications, 87(1-2):236-252, 1995. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 319, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 319, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 67, + 319, + 294, + 380 + ], + "type": "text", + "content": "[79] Priya Sundaresan, Rika Antonova, and Jeannette Bohgl. Diffcloud: Real-to-sim from point clouds with differentiable simulation and rendering of deformable objects. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10828-10835. IEEE, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 383, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 383, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 67, + 383, + 294, + 414 + ], + "type": "text", + "content": "[80] Tetsuya Takahashi and Ming C Lin. Video-guided real-to-virtual parameter transfer for viscous fluids. ACM Transactions on Graphics (TOG), 38(6):1-12, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 415, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 415, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 67, + 415, + 294, + 456 + ], + "type": "text", + "content": "[81] Xiyang Tan, Ying Jiang, Xuan Li, Zeshun Zong, Tianyi Xie, Yin Yang, and Chenfanfu Jiang. Physmotion: Physics-grounded dynamics from a single image. arXiv preprint arXiv:2411.17189, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 458, + 294, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 458, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 67, + 458, + 294, + 490 + ], + "type": "text", + "content": "[82] Bin Wang, Longhua Wu, KangKang Yin, Uri M Ascher, Libin Liu, and Hui Huang. Deformation capture and modeling of soft objects. ACM Trans. Graph., 34(4):94-1, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 491, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 491, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 67, + 491, + 294, + 542 + ], + "type": "text", + "content": "[83] Chaoyang Wang, Lachlan Ewen MacDonald, Laszlo A Jeni, and Simon Lucey. Flow supervision for deformable nerf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21128-21137, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 544, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 544, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 67, + 544, + 294, + 586 + ], + "type": "text", + "content": "[84] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 587, + 294, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 587, + 294, + 639 + ], + "spans": [ + { + "bbox": [ + 67, + 587, + 294, + 639 + ], + "type": "text", + "content": "[85] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 641, + 294, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 641, + 294, + 704 + ], + "spans": [ + { + "bbox": [ + 67, + 641, + 294, + 704 + ], + "type": "text", + "content": "[86] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splattering for real-time dynamic scene rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20310-20320, 2024. 3, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 82, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 320, + 82, + 547, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 82, + 547, + 135 + ], + "spans": [ + { + "bbox": [ + 320, + 82, + 547, + 135 + ], + "type": "text", + "content": "[87] Weijia Wu, Zhuang Li, Yuchao Gu, Rui Zhao, Yefei He, David Junhao Zhang, Mike Zheng Shou, Yan Li, Tingting Gao, and Di Zhang. Draganything: Motion control for anything using entity representation. In European Conference on Computer Vision, pages 331-348. Springer, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 137, + 547, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 137, + 547, + 188 + ], + "spans": [ + { + "bbox": [ + 320, + 137, + 547, + 188 + ], + "type": "text", + "content": "[88] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9421-9431, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 190, + 547, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 190, + 547, + 243 + ], + "spans": [ + { + "bbox": [ + 320, + 190, + 547, + 243 + ], + "type": "text", + "content": "[89] Tianyi Xie, Zeshun Zong, Yuxing Qiu, Xuan Li, Yutao Feng, Yin Yang, and Chenfanfu Jiang. Physgaussian: Physics-integrated 3d gaussians for generative dynamics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4389-4398, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 243, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 243, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 320, + 243, + 546, + 285 + ], + "type": "text", + "content": "[90] Zhenjia Xu, Jiajun Wu, Andy Zeng, Joshua B Tenenbaum, and Shuran Song. Densephysnet: Learning dense physical object representations via multi-step dynamic interactions. arXiv preprint arXiv:1906.03853, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 286, + 546, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 286, + 546, + 339 + ], + "spans": [ + { + "bbox": [ + 320, + 286, + 546, + 339 + ], + "type": "text", + "content": "[91] Tianju Xue, Shuheng Liao, Zhengtao Gan, Chanwook Park, Xiaoyu Xie, Wing Kam Liu, and Jian Cao. Jax-fem: A differentiablegpu-accelerated 3d finite element solver for automatic inverse design and mechanistic data science. Computer Physics Communications, 291:108802, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 340, + 547, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 340, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 320, + 340, + 547, + 381 + ], + "type": "text", + "content": "[92] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. arXiv preprint arXiv:2310.10642, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 383, + 546, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 383, + 546, + 445 + ], + "spans": [ + { + "bbox": [ + 320, + 383, + 546, + 445 + ], + "type": "text", + "content": "[93] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20331-20341, 2024. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 447, + 546, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 447, + 546, + 500 + ], + "spans": [ + { + "bbox": [ + 320, + 447, + 546, + 500 + ], + "type": "text", + "content": "[94] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13117-13125, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 501, + 546, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 501, + 546, + 553 + ], + "spans": [ + { + "bbox": [ + 320, + 501, + 546, + 553 + ], + "type": "text", + "content": "[95] Heng Yu, Joel Julin, Zoltan A Milacski, Koichiro Niinuma, and László A Jeni. Dylin: Making light field networks dynamic. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12397-12406, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 554, + 546, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 554, + 546, + 606 + ], + "spans": [ + { + "bbox": [ + 320, + 554, + 546, + 606 + ], + "type": "text", + "content": "[96] Heng Yu, Joel Julin, Zoltán Á Milacski, Koichiro Niinuma, and László A Jeni. Cogs: Controllable gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21624-21633, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 608, + 546, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 608, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 320, + 608, + 546, + 639 + ], + "type": "text", + "content": "[97] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 641, + 546, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 641, + 546, + 704 + ], + "spans": [ + { + "bbox": [ + 320, + 641, + 546, + 704 + ], + "type": "text", + "content": "[98] Tianyuan Zhang, Hong-Xing Yu, Rundi Wu, Brandon Y Feng, Changxi Zheng, Noah Snavely, Jiajun Wu, and William T Freeman. Physdreamer: Physics-based interaction with 3d objects via video generation. In European Conference on Computer Vision, pages 388-406. Springer, 2025. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 735, + 314, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 735, + 314, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 735, + 314, + 743 + ], + "type": "text", + "content": "5955" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 82, + 295, + 222 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 67, + 82, + 294, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 82, + 294, + 125 + ], + "spans": [ + { + "bbox": [ + 67, + 82, + 294, + 125 + ], + "type": "text", + "content": "[99] Dongzhe Zheng, Siqiong Yao, Wenqiang Xu, and Cewu Lu. Differentiable cloth parameter identification and state estimation in manipulation. IEEE Robotics and Automation Letters, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 126, + 295, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 126, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 64, + 126, + 295, + 178 + ], + "type": "text", + "content": "[100] Licheng Zhong, Hong-Xing Yu, Jiajun Wu, and Yunzhu Li. Reconstruction and simulation of elastic objects with spring-mass 3d gaussians. In European Conference on Computer Vision, pages 407-423. Springer, 2025. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 180, + 295, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 180, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 64, + 180, + 295, + 222 + ], + "type": "text", + "content": "[101] Xiaowei Zhou, Can Yang, and Weichuan Yu. Moving object detection by detecting contiguous outliers in the low-rank representation. IEEE transactions on pattern analysis and machine intelligence, 35(3):597-610, 2012. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 735, + 315, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 735, + 315, + 743 + ], + "spans": [ + { + "bbox": [ + 296, + 735, + 315, + 743 + ], + "type": "text", + "content": "5956" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/8be189e3-e5ec-449b-9eb4-67198582e1c8_content_list.json b/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/8be189e3-e5ec-449b-9eb4-67198582e1c8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..dd6409450f3c6b45e32317175c09d5c49bc4de90 --- /dev/null +++ b/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/8be189e3-e5ec-449b-9eb4-67198582e1c8_content_list.json @@ -0,0 +1,1365 @@ +[ + { + "type": "text", + "text": "Towards Accurate and Efficient 3D Object Detection for Autonomous Driving: A Mixture of Experts Computing System on Edge", + "text_level": 1, + "bbox": [ + 104, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Linshen Liu $^{1,*}$ Boyan Su $^{1,*}$ Junyue Jiang $^{1}$ Guanlin Wu $^{1}$ Cong Guo $^{2}$ Ceyu Xu $^{3}$ Hao Frank Yang $^{1,\\dagger}$ $^{1}$ Johns Hopkins University $^{2}$ Duke University $^{3}$ HKUST {11iu148, bsul1}@jh.edu, haofrankyang@jhu.edu", + "bbox": [ + 264, + 202, + 736, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents Edge-based Mixture of Experts (MoE) Collaborative Computing $(EMC2^{\\ddagger})$ , an optimal computing system designed for autonomous vehicles (AVs) that simultaneously achieves low-latency and high-accuracy 3D object detection. Unlike existing works, the EMC2 introduces a novel scenario-aware MoE architecture optimized for fusing complementary sparse 3D point clouds and dense 2D images to achieve robust multimodal representations for detection. Furthermore, EMC2 integrates an adaptive multimodal data bridge with multi-scale region proposing and scenario-aware routing, dynamically dispatching features to complementary experts based on object visibility and distance. In addition, EMC2 integrates joint hardware-software optimizations, including hardware resource utilization optimization and computational graph simplification, to ensure efficient and real-time inference on resource-constrained edge devices. Experiments on open-source benchmarks clearly show the EMC2 advancements as an end-to-end system. On the KITTI dataset, it achieves an average accuracy improvement of $3.58\\%$ and a $159.06\\%$ inference speedup compared to 15 baseline methods on Jetson platforms, with similar performance gains on the nuScenes dataset, highlighting its capability to advance reliable, real-time 3D object detection tasks for AVs.", + "bbox": [ + 88, + 340, + 485, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 718, + 220, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traffic safety is a fundamental concern for both human drivers and Autonomous Driving Systems (ADS). Within an ADS, the perception module plays a pivotal role by serving as the \"eyes\" of the autonomous vehicle (AV), sensing the surrounding environment and delivering critical information to downstream modules such as motion planning, control, and safety guarantees. Existing conceptual safety of an AV perception system is generally defined by two core", + "bbox": [ + 89, + 743, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7037e73d354f54394db0640ae8f77d2d26d60527926c16fe7628081fb33e055b.jpg", + "image_caption": [ + "Figure 1. Comparison of 3D object detection models in accuracy and inference efficiency on the KITTI and nuScenes dataset. Most existing methods improve accuracy at the expense of inference efficiency. The proposed EMC2 achieves win-win performance by customizing MoE for autonomous driving." + ], + "image_footnote": [], + "bbox": [ + 542, + 292, + 879, + 584 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dimensions: accuracy and efficiency. Accuracy ensures reliable object detection and tracking capabilities, enabling a credible understanding of the environment. Efficiency, characterized by low perception delay, is equally essential, as real-time high-level planning and control rely on fast information processing. Perception delay, defined as the sum of data acquisition, computation, and algorithm processing time, directly affects an ADS's responsiveness to dynamic environments. Research indicates that an ADS must process sensory inputs and issue feedback control commands within 100 milliseconds to guarantee prompt and safe vehicle operations [1]. Detection precision and perception delay together define a critical safety boundary: a precise but sluggish system cannot react in time, while a fast but inaccurate system may lead to unsafe decisions.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution, †Corresponding author. \n‡EMC2 Code: https://github.com/LinshenLiu622/EMC2", + "bbox": [ + 109, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25903", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As illustrated in Fig. 1, achieving both high accuracy and low latency remains a fundamental challenge. Existing approaches often face inherent trade-offs between these objectives due to limited computational resources, insufficient adaptability of perception systems, and the inherent complexity of dynamic driving scenarios. Urban traffic environments require low-latency and accurate sensing capabilities to handle complex, densely interactive surroundings, while highway scenarios demand reliable detection of distant objects and prompt responses to unexpected events. Moreover, even for the same object, recognition priorities may vary dynamically within a scene depending on distance and scene complexity, posing notable challenges for single-modality models in balancing efficiency and accuracy [2]. Multimodal models, such as large-scale CNN and Transformer-based architectures that integrate LiDAR, radar, and camera data [3-5], leverage the complementary advantages of diverse sensors to enhance perception accuracy. Nevertheless, these approaches [6-9] often impose substantial hardware overhead and exhibit limited adaptability, particularly when deployed on resource-constrained edge platforms [10]. Even with the assistance of low-power machine learning compilation frameworks [11], such systems may still execute redundant computations under straightforward scenarios, introducing additional delays that undermine timely safety-critical decision-making.", + "bbox": [ + 93, + 90, + 480, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, multimodal approaches provide a natural foundation for addressing the accuracy challenges faced by ADS. Rather than relying on a single integrated model designed to operate uniformly across diverse driving scenarios, the MoE framework [12] dynamically activates specialized sub-models, or experts, based on input characteristics and prior knowledge. MoE has demonstrated considerable effectiveness in large-scale applications, such as Large Language Models (LLMs) for natural language processing [13, 14], where it routes different tokens to appropriate experts, reducing redundant computations without sacrificing efficiency [15]. As model complexity grows, MoE enables models to scale their capacity while keeping inference hardware costs manageable. Deploying MoE in ADS, however, poses notable challenges [16], including strict real-time constraints, the need for multimodal data processing, and seamless end-to-end autonomous feedback control. Successful deployment requires efficient data preprocessing, adaptive modality selection, and robust expert routing mechanisms.", + "bbox": [ + 93, + 488, + 480, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By leveraging diverse sensor data and a scenario-adaptive MoE dispatcher, we dynamically assign expert sub-models based on scenario data and prior statistical knowledge, providing an efficient, energy-efficient, and safe solution for ADS. To this end, we introduce an Edge-based MoE Collaborative Computing (EMC2) framework - an algorithm-hardware-software co-design solution that not", + "bbox": [ + 93, + 795, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "only improves accuracy through complex multimodal experts but also effectively reduces latency by utilizing an adaptive MoE dispatcher and lightweight single-modal experts. Our main contributions are summarized as follows:", + "bbox": [ + 517, + 90, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1) A Multimodal 3D Object Detection Solution with MoE. The proposed EMC2 system achieves both high detection accuracy and inference efficiency, outperforming 15 existing one-size-fits-all baseline models. Designed for real-world ADS applications, EMC2 is optimized for deployment on resource-constrained platforms, making it more suitable for edge computing environments than generic alternatives.", + "bbox": [ + 517, + 152, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2) Robust and Efficient Multimodal Perception through Expert Collaboration and System-Level Optimization. The proposed EMC2 framework incorporates scenario-specific experts and employs an adaptive expert selection strategy to meet heterogeneous perception demands. To mitigate the training instability commonly observed in MoE-based systems, we design a hierarchical optimization scheme that combines expert-level and global EMC2 backpropagation for effective collaborative learning. Furthermore, a distribution-aware resampling technique enhances robustness under real-world data imbalance. In addition to these designs, a tailored multiscale pooling module improves the efficiency of LiDAR-camera fusion, ensuring both efficient and accurate multimodal perception.", + "bbox": [ + 517, + 276, + 903, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3) System-Level Optimization for Edge Inference. The original PyTorch-based MoE inference framework [17] suffers from inefficient management of memory and computational resources, leading to high latency on edge devices. To overcome these challenges, we propose an algorithm-hardware-software co-optimization strategy within EMC2, which incorporates efficient memory management techniques to improve L1/L2 cache hit rates and applies computational graph fusion to eliminate redundant data loading across layers. These combined optimizations notably improve inference efficiency.", + "bbox": [ + 517, + 488, + 903, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "4) State-of-the-art (SOTA) Accuracy and Efficiency Combined. On the KITTI dataset, EMC2 achieves $3.28\\%$ higher accuracy for pedestrians and $6.03\\%$ for bicycles under hard-level detection challenge compared to 15 baseline models. On the nuScenes dataset, it yields a $3.9\\%$ increase in mAP and a $1.8\\%$ boost in NDS. In addition, EMC2 achieves a $159.06\\%$ reduction in inference latency on the Jetson platform for KITTI, supporting its suitability for real-time edge deployment.", + "bbox": [ + 517, + 659, + 903, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 517, + 801, + 653, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LiDAR-only Method. The active 3D sensing of LiDAR makes it dominant in ADS 3D detection [18-20]. LiDAR-based methods process point clouds via voxelization [21], pillarization [20], or direct raw input [22]. Anchor-based heads, used for region proposal [21], and", + "bbox": [ + 517, + 825, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25904", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "center-based representations[23] are widely adopted. These methods achieve high accuracy on KITTI, as demonstrated by 3DSSD [24], PartA2 [19], and Voxel R-CNN [3]. However, weak LiDAR reflections for distant or occluded objects degrade detection performance [25, 26].", + "bbox": [ + 88, + 90, + 480, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodal Fusion Method. Combining LiDAR and camera RGB input enhances detection performance, especially for sparse LiDAR objects [27]. Fusion approaches include projection-based methods, which align image features with point clouds in 3D space [28], and model-based techniques that incorporate cross-attention [29] or unified feature spaces [30]. Recent multimodal models, such as PPF-Det [31] and RoboFusion [32], achieve top accuracy on KITTI benchmarks. However, these models often trade off accuracy for speed or vice versa.", + "bbox": [ + 89, + 170, + 480, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Edge-based Compilation. ML compilation platforms are essential for improving the inference efficiency of deep learning models across diverse hardware architectures [33, 34]. TensorRT [35] is an NVIDIA inference optimizer that accelerates ML models on GPUs through layer fusion, precision calibration, and kernel auto-tuning. XLA [36] is a domain-specific compiler for TensorFlow enhancing efficiency by fusing operations and optimizing computation graphs. Open Neural Network Exchange (ONNX) Runtime [37] provides an open standard for model interoperability, enabling efficient execution across CPUs, GPUs, and accelerators. TVM [11] is an open-source deep learning compiler that automates model tuning and tensor scheduling for efficient deployment across CPUs and GPUs.", + "bbox": [ + 91, + 325, + 483, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. EMC2 Computing Methodology", + "text_level": 1, + "bbox": [ + 89, + 542, + 383, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce EMC2, a multimodal MoE framework that minimizes latency and ensures reliable accuracy. Fig. 2 shows the EMC2 overview, EMC2 consists of: 1) Adaptive Multimodal Data Bridge (AMDB) to preprocess multimodal input (Sec. 3.1); 2) Scenario-Adaptive Dispatcher to adaptively dispatch a suitable expert for each scenario (Sec. 3.2); and 3) Three Scenario-Optimized Experts to decode AMDB-output features into final results (Sec. 3.3). We also design an effective training strategy to address the long-tail effect in the training process (Sec. 3.4). In addition, we deploy EMC2 on edge devices with self-designed 3D sparse convolution and Multiscale Pooling.", + "bbox": [ + 89, + 568, + 482, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Adaptive Multimodal Data Bridge", + "text_level": 1, + "bbox": [ + 89, + 757, + 388, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the multimodal nature of EMC2's sensor inputs, a module providing suitable preprocessed data for different experts is required. Thus, we propose AMDB, a preprocessing unit to efficiently process the required input data for each expert. The workflow of AMDB begins with a UNet [38] and a sparse CNN to extract LiDAR features, followed by fully connected layers that generate proposal regions along with their confidence scores. These results", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "serve as inputs to Latency-Prioritized Expert and Versatile Efficiency Expert. The extraction of image features depends on the Scenario-Adaptive Dispatcher. If further information is required for reliable detection, AMDB will extract image features by ResNet. Based on depth information, it projects relevant image pixels into 3D space, applying Multiscale Pooling to reduce computational overhead, and fuses them with LiDAR voxel features to produce a multimodal output. These results are inputs for Accuracy-Prioritized Experts. The Multiscale Pooling and multimodal fusion processes are demonstrated in Fig. 5.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Scenario-Adaptive Dispatcher", + "text_level": 1, + "bbox": [ + 511, + 266, + 781, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Scenario-Adaptive Dispatcher (SAD) dynamically dispatches each scenario to the most suitable expert based on traffic conditions and latency-accuracy requirements. Our expert selection algorithm considers real-world traffic safety needs: specifically, how far away objects are and how clearly they can be perceived.", + "bbox": [ + 511, + 289, + 903, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "So, the switch among experts is determined by two key parameters: object distance $\\mathcal{D}$ and clarity, which is inferred from the confidence $\\mathcal{C}$ of proposal regions provided by the AMDB. We classify traffic scenarios into three categories: 1) Close and Distinct Cases. All objects are clearly visible and no distant objects are present. This is characterized by all proposal regions $< \\mathcal{D}$ and having confidence $\\geq \\mathcal{C}$ . In such scenarios, as near-field objects provide abundant Li-DAR information, even when compressed into bird's eye view (BEV), a 2D CNN can effectively capture key patterns such as object surfaces and heights [20], ensuring reliable detection accuracy with minimal latency (Sec. 4.4). These cases will be routed to the Latency-Prioritized Expert. 2) Mixed Visibility Cases. Some objects are either distant but clearly visible or near but unclear. Mathematically, this corresponds to cases where i) some proposal regions $< \\mathcal{D}$ with confidence $< \\mathcal{C}$ , or ii) the distance of some objects $\\geq \\mathcal{D}$ with confidence $\\geq \\mathcal{C}$ . Here, objects have insufficient voxel data, requiring more complex 3D convolution to extract meaningful patterns, and they will be dispatched to the Versatile Efficiency Expert. 3) Distant and Uncertain Cases. Some objects are both distant and unclear. This occurs when some proposal regions $\\geq \\mathcal{D}$ with confidence $< \\mathcal{C}$ . In these challenging scenarios, significant voxel information is missing due to long distances, steep reflection angles, or occlusions, necessitating the integration of image data to compensate for missing LiDAR details. These cases will be processed by the Accuracy-Prioritized Expert.", + "bbox": [ + 511, + 381, + 903, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Scenario-Optimized Experts", + "text_level": 1, + "bbox": [ + 511, + 816, + 767, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latency-Prioritized Expert (LPE). LPE uses 2D CNNs for object detection based on BEV projections. It first projects proposal regions from the AMDB into BEV space, then extracts 2D features for object localization and", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25905", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f13f7eecda8ca25f2db6748d167f0ac62896c17fa918088cec8473745c50189c.jpg", + "image_caption": [ + "Figure 2. System Structure of EMC2. The central row illustrates the architecture of EMC2, while the top and bottom rows provide visual examples of core modules and relevant legends. The system integrates five main components: an Adaptive Multimodal Data Bridge (AMDB) for preprocessing raw multimodal input and generating expert-specific features; a Scenario-Adaptive Dispatcher that dispatches suitable experts based on scenario-specific features (see Fig. 3); the Latency-Prioritized Expert designed for simple scenarios requiring minimal inference delay; the Versatile Efficiency Expert designed for uncertain or distant object detection scenarios; and the Accuracy-Prioritized Expert responsible for handling complex, high-precision detection scenarios. Additionally, an Emergency Expert API is designed to support rapid response to hazardous or unseen situations. The overall training process incorporates loss terms for both final detection results and intermediate outputs from the AMDB, as described in Sec. 3.4." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 425 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8cb87474ddee652c12453ab79550d6e0ea89ea00809e0beb2ac65cf22d4d29ee.jpg", + "image_caption": [ + "Figure 3. Routing Strategy of the Scenario-Adaptive Dispatcher. The Routing criteria is detailed mathematically in Sec. 3.2." + ], + "image_footnote": [], + "bbox": [ + 94, + 553, + 480, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "classification. It finally maps detection results back to 3D space. Compared to sparse 3D convolutions, its use of 2D operations reduces computation in outdoor ADS scenarios [20], allowing LPE to improve inference efficiency with minimal accuracy loss.", + "bbox": [ + 89, + 763, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Versatile Efficiency Expert (VEE). VEE uses a 3D CNN to process proposal regions and confidence scores from the AMDB. After sparse convolution, voxel features are decoded into per-voxel classifications and bounding", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "boxes. Compared to LPE, 3D convolutions better preserve spatial structure, making it more robust for scenarios with insufficient LiDAR signal.", + "bbox": [ + 511, + 554, + 906, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accuracy-Prioritized Expert (APE). APE extends VEE by incorporating the multimodal signal. It employs a 3D CNN and decoder with the same structure but different parameters to process multimodal proposal regions (from AMDB) and fuses image data to recover missing LiDAR information. Image features compressed via Multiscale Pooling (Sec.3.5) are projected into 3D space and fused with corresponding voxels of LiDAR features, ensuring more robust understanding for occluded and distant objects while keeping computation feasible. The Multiscale Pooling and multimodal fusion are detailed in Fig. 5.", + "bbox": [ + 511, + 599, + 908, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Collaborative Training of Multimodal Experts", + "text_level": 1, + "bbox": [ + 511, + 772, + 903, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hierarchical Training Strategy: Multimodal Supervision with Triple Back-Propagation. During training, the performance of experts is highly dependent on the proposal regions produced by the AMDB, yet these proposals tend to be unreliable in the early training phase, leading to an unstable expert learning process and unstable SAD decisions. To address this, we adopt a hierarchical supervision stra", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25906", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bf1cea59c5f10b342486ad5d82aa8d91aa2198df29496c17f3124dc349209ef1.jpg", + "image_caption": [ + "Figure 4. Illustration of Hierarchical Training. The three backpropagation paths originate from: 1) the final detection output, 2) the LiDAR branch output of AMDB, and 3) the image branch output of AMDB. Since expert training relies on initially unreliable AMDB outputs, direct supervision accelerates convergence and enhances representation learning." + ], + "image_footnote": [], + "bbox": [ + 130, + 89, + 441, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "egy with three distinct back-propagation pathways, as illustrated in Fig. 4: 1) supervision of the LiDAR branch outputs from the AMDB; 2) supervision of the image branch outputs from the AMDB; and 3) supervision of expert predictions and associated confidence scores. To further improve training stability, the LiDAR and image branches of the AMDB are pre-trained separately before joint optimization. Each back-propagation route is supervised by the following loss function, which collectively forms Multimodal Loss:", + "bbox": [ + 89, + 470, + 483, + 606 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r o u t e}} = \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {r e g}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 617, + 480, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{cls}}$ uses cross-entropy loss and $\\mathcal{L}_{\\mathrm{reg}}$ uses smooth- $\\ell 1$ loss. The Hierarchical Training not only stabilizes expert learning but also enhances the model's adaptability to various data granularities. In turn, even under compressed data representations, it enables the effective use of our Multiscale Pooling for memory efficiency, as detailed in Sec. 3.5.", + "bbox": [ + 89, + 643, + 482, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Addressing Long-tail Effects in MoE Training. The MoE framework often suffers from long-tail effects, where certain experts receive insufficient training data, leading to imbalances in sub-datasets, and therefore affecting performance [39]. To address this, we propose a threefold training strategy: 1) Data subset division. We partition the dataset into sub-datasets based on target and auxiliary samples. Each expert $\\mathcal{E}_i$ is dispatched a sub-dataset $S_{i}$ $(i = 1,2,3)$ containing target samples $\\mathcal{T}_{\\mathcal{S}_i}$ , which are expected to be routed to the corresponding expert, and auxiliary samples $\\mathcal{A}_{\\mathcal{S}_i}$ , which are not. We ensure that $\\mathcal{T}_{\\mathcal{S}_i} \\cap \\mathcal{T}_{\\mathcal{S}_j} = \\emptyset$ for any", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/370e7b99ef66e8ccacbdbb6143a6bc15182d6a6bf5f37353615c9a6ae204e431.jpg", + "image_caption": [ + "Figure 5. Illustration of Multiscale Pooling and Multimodal Fusion. Proposal regions with image features are first obtained from the 3D proposal space generated by the LiDAR branch of AMDB. After extracting the relevant image features from the corresponding image regions and projecting them into 3D space, pooling and fusion are then applied. Following the pooling step, image features of pixels that share the same 3D coordinate as a voxel are concatenated with the corresponding LiDAR voxel features, while those without matching voxel coordinates are concatenated with zero-filled features." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 905, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$i \\neq j$ . The inclusion of auxiliary samples, which are not intended for a given expert, helps prevent overfitting and enhances generalization. 2) Balanced sampling strategy. We adjust the selection probability $\\mathcal{P}_i$ for the sub-datasets which have fewer samples to ensure balanced training:", + "bbox": [ + 511, + 398, + 905, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {i} \\times \\mathcal {N} _ {\\mathcal {S} _ {i}} = \\mathcal {P} _ {j} \\times \\mathcal {N} _ {\\mathcal {S} _ {j}}, \\sum_ {i = 1} ^ {3} \\mathcal {P} _ {i} = 1 \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 478, + 903, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{N}_{S_i}$ represents the number of samples in sub-datasets $S_{i}$ . This ensures that experts dispatched to less frequent scenarios receive sufficient training iterations. 3) Adaptive optimizer. We introduce an optimizer that adapts the learning rate based on the proportion of target samples in a batch. Denoting $\\alpha_0$ as the base learning rate, the adaptive learning rate applied to an expert with a given batch is computed as:", + "bbox": [ + 511, + 523, + 905, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = \\left(1 + \\frac {1}{\\mathcal {N}} \\sum_ {i = 1} ^ {\\mathcal {N}} \\mathbf {1} _ {i \\in \\mathcal {T}}\\right) \\cdot \\alpha_ {0} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 638, + 903, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{N}$ is the batch size, $p$ is the proportion of target samples within current mini-batch, $\\mathcal{T}$ represents target samples in the batch. By regulating the gradient update rate based on the batch composition, this adaptive optimizer encourages each expert to prioritize its designated target samples while preventing excessive updates from auxiliary samples.", + "bbox": [ + 511, + 688, + 905, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Algorithm Edge-adaptivity Optimization", + "text_level": 1, + "bbox": [ + 511, + 787, + 862, + 804 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Edge devices face challenges such as incomplete function libraries and limited cache. We adopt the open-source ONNX Runtime for compiling EMC2 on edge, which provides modules that help address these challenges and facilitate edge-oriented optimization. As 3D convolutions account for the majority of computation during inference, we", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25907", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "customize 3D sparse convolution and Multiscale Pooling to improve the algorithmic efficiency of EMC2.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D Sparse Convolution. Since the 3D sparse convolution library is incomplete on most edge platforms, we develop a customized 3D sparse convolution library with parallel threads execution under ONNX Runtime. Let $\\mathcal{H}$ denote spatial resolution, $\\mathcal{N}$ the number of non-empty voxels, and $\\mathcal{C}_{\\mathrm{v}}$ the per-voxel computation, including channels and kernel size. By computing nonzero voxels only, 3D sparse convolution reduces complexity from $O(\\mathcal{H}^3\\times \\mathcal{C}_{\\mathrm{v}})$ for dense convolutions to $O(\\mathcal{N}\\times \\mathcal{C}_{\\mathrm{v}})$ , where $\\mathcal{N}\\ll \\mathcal{H}^3$ , cutting redundant operations by $65 - 80\\%$ [40].", + "bbox": [ + 89, + 126, + 483, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multiscale Pooling. To alleviate cache limitations, Multiscale 3D Sparse Pooling adaptively pools image features based on available memory before fusing them with LiDAR features (Fig.5). The pooling size is user-configurable, allowing flexibility for maximizing memory utilization. When combined with Hierarchical Training (Sec.3.4), Multiscale Pooling effectively reduces resource consumption while maintaining model reliability (Sec. 4.4).", + "bbox": [ + 89, + 282, + 483, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. EMC2 Computing System Optimization", + "text_level": 1, + "bbox": [ + 89, + 417, + 434, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building upon the algorithm-level optimizations introduced in Sec. 3.5, we further enhance the ONNX execution efficiency through system-level improvements, specifically targeting memory management and computation graph optimization.", + "bbox": [ + 89, + 440, + 483, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Memory Optimization. To further improve memory efficiency during inference, we implement three system-level techniques. 1) Overlapping communication and computation [41]. Matrix data is partitioned into segments processed and transferred in parallel, overlapping computation and communication. This reduces global memory access, improves cache usage, and mitigates memory bottlenecks. 2) Thread scheduling. We implement a staged thread management mechanism, where active threads are periodically terminated and re-initialized (Fig. 6). This approach limits unnecessary memory retention between stages, reducing system memory footprint. 3) Prefix-sum for sparse convolution. Parallel prefix-sum reduces index search time from $\\mathcal{O}(\\mathcal{N})$ to $\\mathcal{O}(\\log \\mathcal{N})$ using GPU parallelism, where $\\mathcal{N}$ is the number of nonzero elements.", + "bbox": [ + 89, + 521, + 482, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Computational Graph Optimization. ONNX partitions the computation graph into segments and assigns each to the most suitable hardware. Our optimization includes three stages: 1) Model pruning. Redundant parameters and operations are removed to simplify the graph; 2) Model quantization. Low-precision arithmetic replaces floating-point computation to improve hardware efficiency; 3) Computational graph fusion. Consecutive operators are merged to reduce memory access and improve execution, benefiting memory-bound CPUs and reducing kernel", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "launch overhead on GPUs.", + "bbox": [ + 513, + 90, + 692, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment and Result", + "text_level": 1, + "bbox": [ + 511, + 118, + 732, + 135 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiment Settings and Datasets", + "text_level": 1, + "bbox": [ + 511, + 140, + 808, + 156 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiment Platform. Experiments are conducted on the Jetson AGX Orin, an edge platform with limited computational resources, an NVIDIA Ampere GPU, a 4MB shared L2 cache, and hierarchical memory. For comparison, EMC2 is also evaluated on an NVIDIA A4000 GPU.", + "bbox": [ + 511, + 162, + 903, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and Sample Difficulties. We evaluate 3D object detection on two public benchmarks: KITTI [42] and nuScenes [43]. KITTI provides 3,712 training and 3,769 validation samples, with object difficulty levels categorized as $30\\%$ easy, $40\\%$ moderate, and $30\\%$ hard. nuScenes contains 850 training and 150 test scenes, covering 10 object categories with 3D bounding box annotations. Before splitting the data, we analyze the confidence distribution of proposal regions generated by the pre-trained AMDB. A two-sample Kolmogorov-Smirnov test reveals a statistically significant difference between objects within and beyond a distance threshold $D$ , which is empirically set to 23.5 meters for KITTI and 35 meters for nuScenes. The training subsets for each expert on the KITTI and nuScenes datasets are defined as follows: $\\mathcal{T}_{S_1}$ includes scenes where all objects are within $D$ meters and labeled as \"easy\" or \"moderate\"; $\\mathcal{T}_{S_2}$ includes scenes with at least one object beyond $D$ meters labeled as \"easy\" or \"moderate,\" or any object within $D$ meters labeled as \"hard\"; and $\\mathcal{T}_{S_3}$ includes scenes where at least one object is both beyond $D$ meters and labeled as \"hard,\" requiring multimodal processing. For each expert, the auxiliary subset $\\mathcal{A}_{S_i}$ is uniformly sampled from the other experts' target subsets, with its size equal to that of the corresponding target subset.", + "bbox": [ + 511, + 243, + 906, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. KITTI uses R40 3D Average Precision (AP) with IoU thresholds of 0.7 for cars and 0.5 for pedestrians and cyclists. nuScenes adopts the NDS metric, combining mAP—computed via center distance matching $(0.5\\mathrm{m}$ to $4.0\\mathrm{m}$ thresholds)—and five error terms: mean Average Translation (mATE), Scale (mASE), Orientation (mAOE), Velocity (mAVE), and Attribute (mAAE) Errors. Inference latency on Jetson AGX Orin is also reported to assess real-time performance.", + "bbox": [ + 511, + 609, + 905, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hierarchical Training Details. Allowing all experts to propagate gradients to AMDB causes instability due to overlapping training data, leading to repeated updates. To address this, AMDB is first pre-trained for 20 epochs, followed by joint training where gradients are propagated only from APE, which receives full multimodal input.", + "bbox": [ + 511, + 744, + 905, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Existing SOTA Methods", + "text_level": 1, + "bbox": [ + 511, + 847, + 872, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 1 and Tab. 2, EMC2 achieves SOTA accuracy and latency on KITTI and nuScenes. On the Jetson", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25908", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d62ca3cdfe936cfa47d6ab4d81ab4d294870be561966d0dd6faf48048b96e60b.jpg", + "image_caption": [ + "Figure 6. Illustration of Thread Management. Each solid box represents a module occupying threads. During inference, four boundaries are defined at which preceding threads are terminated, and their thread-specific address space is released upon crossing each boundary." + ], + "image_footnote": [], + "bbox": [ + 130, + 90, + 861, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c51465c9b24cc8b4e0c05547de5828eb66ab9ef9357ae08a3e9df1f8c130a1cb.jpg", + "image_caption": [ + "Table 1. We evaluate different 3D object detection methods on the KITTI dataset. Some cells are denoted as $N / A$ when the corresponding open-source implementation lacks support for the specific object class or is not compatible with Jetson devices." + ], + "image_footnote": [], + "bbox": [ + 91, + 284, + 908, + 551 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5a1e6dee6a34a22d1c59ddd3bb508a780f1eb162ba06db3f63d895f056d8b6ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSizemAP↑NDS↑Latency (ms)↓
FocalFormer [55]189M0.66400.7090N/A
BEVFusion [56]156M0.68520.7138N/A
EMC2 (ours)87M0.72410.7316229.3
", + "bbox": [ + 96, + 595, + 475, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison of different methods on the nuScenes dataset.", + "bbox": [ + 91, + 665, + 482, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "AGX Orin platform, it also outperforms existing baseline models across all difficulty levels.", + "bbox": [ + 89, + 707, + 482, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on KITTI. Pedestrians and cyclists face higher traffic risks [54], which remain challenging for most SOTA methods. EMC2 improves hard-set detection by $5\\%$ and $7\\%$ for pedestrians and cyclists, respectively, outperforming TED and LoGoNet (Tab. 1). For cars, it achieves $88.15\\%$ on the hard set, exceeding prior multimodal methods by up to $9\\%$ . It runs at $372.5 \\mathrm{~ms}$ on Jetson, $2.5 \\times -2.5 \\times$ faster than methods requiring $965 - 1$ , $787 \\mathrm{~ms}$ .", + "bbox": [ + 89, + 743, + 482, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on nuScenes. EMC2 improves mAP and NDS by $3.9\\%$ and $1.8\\%$ , respectively, outperforming recent", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods FocalFormer [55] and BEVFusion [56] (Tab. 2). It runs at $229.3\\mathrm{ms}$ on Jetson AGX Orin, enabling real-time use on edge devices.", + "bbox": [ + 511, + 597, + 906, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis and Discussion", + "text_level": 1, + "bbox": [ + 511, + 656, + 733, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benefits of Mixture-of-Experts. The balance between accuracy and efficiency observed in Sec. 4.2 is achieved through dynamic expert selection based on scenario characteristics, combined with algorithmic and system-level optimizations designed for edge devices, such as Jetson.", + "bbox": [ + 511, + 679, + 906, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Empirical Dispatcher. The EMC2 dispatcher uses fixed distance and confidence thresholds to dispatch scenes to experts. The protocol is adapted per dataset format. This strategy is validated via expert activations on the nuScenes validation set (Fig. 7): LPE performs best in close-range, high-confidence scenes; APE in distant, low-confidence cases; and VEE in all others, supporting the effectiveness of empirical routing.", + "bbox": [ + 511, + 758, + 908, + 881 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Optimized Parameter Size. EMC2 has 226M param-", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25909", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0783037485353f1a17162a96132bcc397f417a468d0c2ef34a3fefe2db0f6c3c.jpg", + "image_caption": [ + "Figure 7. Distribution of expert activation. During EMC2 inference on the nuScenes validation set, scenes are grouped into a 2D grid based on two metrics: the y-axis indicates the number of AMDB-generated proposals beyond the distance threshold $D$ (i.e., distant objects), and the x-axis represents the average confidence score of these proposals (higher values indicate clearer scenes). Each cell is labeled with the name of the expert that performs best for scenes falling into that bin." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 475, + 232 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8ad0f87d16bfc084782903972c7c98a5c73af69fe8048b9bcf032eadfe65aa1.jpg", + "image_caption": [ + "Figure 8. Median Detection Confidence vs. Object Distances. Four configurations are compared on the KITTI validation set: LPE (blue), VEE (peach), APE (rose), and EMC2 (green). More complex experts yield higher confidence but slower inference. By switching experts based on distance and clarity, EMC2 balances accuracy and efficiency." + ], + "image_footnote": [], + "bbox": [ + 96, + 361, + 475, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "eters, but its MoE framework activates only a subset during inference, averaging 87M on nuScenes and 146M on KITTI, which is fewer than those used by the full model and other SOTA baselines. System-level optimizations instead prioritize computational throughput and memory efficiency over parameter reduction.", + "bbox": [ + 89, + 622, + 482, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hardware Flexibility. The EMC2 is implemented in the ONNX format, an open format supported by various compilers and enables model compilation across diverse hardware platforms. As a result, EMC2 is not limited to Jetson and can be deployed on other edge platforms.", + "bbox": [ + 89, + 715, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 801, + 243, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on the Roles of Three Experts. We evaluate each expert on its target scenarios using the KITTI validation set. As shown in Tab. 3, LPE, VEE, and APE achieve comparable accuracy in LPE scenes, but LPE has the lowest inference latency on Jetson. In VEE scenarios, incor", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b8dd0d61261168fbbc2a23a2ec6c81ad2d2c6b2ef6e290e6111ed79b98e7e993.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAP↑Latency (ms) ↓
Ped.CarCyclist
Ablation on LPE
LPE66.3289.0488.54219
VEE74.4992.2391.69468
APE74.3493.2192.41526
Ablation on VEE
W/ VEE70.5590.5084.07372.5
W/O VEE68.2686.4079.17512.7
Ablation on APE
VEE67.2487.6981.85468
APE70.5590.5084.18526
Ablation on Hierarchical Training and Multiscale Pooling
HT √, MP √70.5590.5084.07526
HT √, MP ×70.5590.5084.181320
HT ×, MP √63.1188.6875.98529
HT ×, MP ×66.3489.1080.831367
", + "bbox": [ + 514, + 88, + 893, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Results Across Three Ablation Configurations. The main body of this table is divided into four sections corresponding to each part of the ablation study. In the third section, HT and MP are abbreviated for Hierarchical Training and Multiscale Pooling, respectively.", + "bbox": [ + 511, + 338, + 903, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "porating VEE improves both accuracy and latency over the baseline. For APE, which incorporates image features, accuracy improves further with only a minor latency increase, supported by the score distribution in Fig. 8.", + "bbox": [ + 511, + 420, + 903, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Hierarchical Training (HT) and Multiscale Pooling (MP). We evaluate four combinations of HT and MP on the full KITTI validation set: both applied, HT only, MP only, and neither. Tab. 3 shows that combining HT and MP reduces inference time on Jetson from 1320 ms to 526 ms without accuracy loss. For the car class, MP alone reduces accuracy from $80.83\\%$ to $75.98\\%$ , while HT alone improves it from $66.34\\%$ to $70.55\\%$ . HT facilitates multigranular feature learning, while MP enhances memory efficiency via compression.", + "bbox": [ + 511, + 484, + 906, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 648, + 633, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present EMC2, a multimodal 3D object detection solution for ADS based on a customized MoE architecture. This design enables adaptive expert selection across diverse traffic scenarios, balancing detection accuracy and efficiency. Experiments on KITTI and nuScenes show strong performance, particularly under challenging conditions. On the Jetson platform, EMC2 achieves a $159.06\\%$ speedup in inference, supporting real-time edge deployment. Future work will explore fully adaptive expert selection to enhance system responsiveness and resource efficiency.", + "bbox": [ + 511, + 670, + 906, + 821 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25910", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shih-Chieh Lin, Yunqi Zhang, Chang-Hong Hsu, Matt Skach, Md E. Haque, Lingjia Tang, and Jason Mars. The architectural implications of autonomous driving: Constraints and acceleration. In Proceedings of the Twenty-Third International Conference on Architectural Support for Programming Languages and Operating Systems, ASPLOS '18, page 751-766, New York, NY, USA, 2018. Association for Computing Machinery. 1", + "[2] Yi Xiao, Felipe Codevilla, Akhil Gurram, Onay Urfalioglu, and Antonio M López. Multimodal end-to-end autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(1):537-547, 2020. 2", + "[3] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. arXiv:2012.15712, 2020. 2, 3, 7", + "[4] Hao Frank Yang, Yifan Ling, Cole Kopca, Sam Ricord, and Yinhai Wang. Cooperative traffic signal assistance system for non-motorized users and disabilities empowered by computer vision and edge artificial intelligence. Transportation Research Part C: Emerging Technologies, 145:103896, 2022. Elsevier.", + "[5] Hao Yang, Jiarui Cai, Meixin Zhu, Chenxi Liu, and Yinhai Wang. Traffic-informed multi-camera sensing (TIMS) system based on vehicle re-identification. IEEE Transactions on Intelligent Transportation Systems, 23(10):17189-17200, 2022. IEEE. 2", + "[6] Guibiao Liao, Jiankun Li, and Xiaqing Ye. Vlm2scene: Self-supervised image-text-lidar learning with foundation models for autonomous driving scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 3351-3359, 2024. 2", + "[7] Hao Frank Yang, Yang Zhao, Jiarui Cai, Meixin Zhu, JenqNeng Hwang, and Yiran Chen. Mitigating bias of deep neural networks for trustworthy traffic perception in autonomous systems. In Proceedings of the 2024 IEEE Intelligent Vehicles Symposium (IV), pages 633-638, 2024. IEEE.", + "[8] Chenxi Liu, Hao Yang, Ruimin Ke, Wei Sun, Julia Wang, and Yinhai Wang. Cooperative and comprehensive multitask surveillance sensing and interaction system empowered by edge artificial intelligence. Transportation Research Record, 2677(9):652-668, 2023. SAGE Publications.", + "[9] Hao Frank Yang, Jiarui Cai, Chenxi Liu, Ruimin Ke, and Yinhai Wang. Cooperative multi-camera vehicle tracking and traffic surveillance with edge artificial intelligence and representation learning. Transportation Research Part C: Emerging Technologies, 148:103982, 2023. Elsevier. 2", + "[10] Chen Zhang, Peng Li, Guangyu Sun, Yijin Guan, Bingjun Xiao, and Jason Cong. Optimizing fpga-based accelerator design for deep convolutional neural networks. In Proceedings of the 2015 ACM/SIGDA international symposium on field-programmable gate arrays, pages 161-170, 2015. 2", + "[11] Tianqi Chen, Thierry Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Yan, Haichen Shen, Meghan Cowan, Leyuan Wang, Yuwei Hu, Luis Ceze, et al. {TVM}: An automated {End-to-End} optimizing compiler for deep learning. In 13th" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "USENIX Symposium on Operating Systems Design and Implementation (OSDI 18), pages 578-594, 2018. 2, 3", + "[12] Jiarui Cai, Yizhou Wang, and Jenq-Neng Hwang. Ace: Ally complementary experts for solving long-tailed recognition in one-shot. In Proceedings of the IEEE/CVF international conference on computer vision, pages 112-121, 2021. 2", + "[13] Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. Unimoe: Scaling unified multimodal llms with mixture of experts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2", + "[14] Jiachen Li, Xinyao Wang, Sijie Zhu, Chia-Wen Kuo, Lu Xu, Fan Chen, Jitesh Jain, Humphrey Shi, and Longyin Wen. Cumo: Scaling multimodal llm with co-upcycled mixture-of-experts. Advances in Neural Information Processing Systems, 37:131224-131246, 2024. 2", + "[15] Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts. arXiv preprint arXiv:2407.06204, 2024. 2", + "[16] Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. Time-moe: Billion-scale time series foundation models with mixture of experts. arXiv preprint arXiv:2409.16040, 2024. 2", + "[17] Jiaao He, Jiezhong Qiu, Aohan Zeng, Zhilin Yang, Jidong Zhai, and Jie Tang. Fastmoe: A fast mixture-of-expert training system. arXiv preprint arXiv:2103.13262, 2021. 2", + "[18] Charles R. Qi, Xinlei Chen, Or Litany, and Leonidas J. Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4403-4412, 2020. 2", + "[19] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. arXiv preprint arXiv:1907.03670, 2019. 3", + "[20] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689-12697, 2019. 2, 3, 4, 7", + "[21] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2018. 2", + "[22] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10526–10535, 2020. 2", + "[23] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11779-11788, 2021. 3", + "[24] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11037-11045, 2020. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25911", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Andrew M Wallace, Abderrahim Halimi, and Gerald S Buller. Full waveform lidar for adverse weather conditions. IEEE transactions on vehicular technology, 69(7):7064-7077, 2020. 3", + "[26] You Li, Pierre Duthon, Michele Colomb, and Javier Ibanez-Guzman. What happens for a tof lidar in fog? IEEE Transactions on Intelligent Transportation Systems, 22(11):6670-6681, 2020. 3", + "[27] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In 2019 international conference on robotics and automation (ICRA), pages 3288-3295. IEEE, 2019. 3", + "[28] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. E-net: Enhancing point features with image semantics for 3d object detection. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 35–52, Cham, 2020. Springer International Publishing. 3", + "[29] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quaan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V. Le, Alan Yuille, and Mingxing Tan. Deep-fusion: Lidar-camera deep fusion for multi-modal 3d object detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17161-17170, 2022. 3", + "[30] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022. 3", + "[31] Guotao Xie, Zhiyuan Chen, Ming Gao, Manjiang Hu, and Xiaohui Qin. Ppf-det: Point-pixel fusion for multi-modal 3d object detection. IEEE Transactions on Intelligent Transportation Systems, 25(6):5598-5611, 2024. 3", + "[32] Ziying Song, Guoxing Zhang, Lin Liu, Lei Yang, Shaoqing Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Robofusion: Towards robust multi-modal 3d object detection via sam. arXiv preprint arXiv:2401.03907, 2024. 3", + "[33] Zheng Wang and Michael O'Boyle. Machine learning in compiler optimization. Proceedings of the IEEE, 106(11):1879-1901, 2018. 3", + "[34] Hugh Leather, Edwin Bonilla, and Michael O'boyle. Automatic feature generation for machine learning-based optimising compilation. ACM Transactions on Architecture and Code Optimization (TACO), 11(1):1-32, 2014. 3", + "[35] NVIDIA Corporation. NVIDIA TensorRT Developer Guide, 2024. Accessed: 2025-02-13. 3", + "[36] Amit Sabne. XLA: Compiling Machine Learning for Peak Performance, 2020. 3", + "[37] ONNX Community. Onnx: Open neural network exchange. https://github.com/onnx/onnx, 2024. Accessed: YYYYY-MM-DD. 3", + "[38] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. *Neural computation*, 3(1):79–87, 1991. 5", + "[40] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10), 2018. 6, 7", + "[41] Juraj Hromkovič. Communication complexity and parallel computing. Springer Science & Business Media, 2013. 6", + "[42] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 6", + "[43] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6", + "[44] Muhammad Aslam. Introducing kolmogorov-smirnov tests under uncertainty: an application to radioactive data. ACS omega, 5(1):914-917, 2019.", + "[45] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 772-782, 2022. 7", + "[46] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 7", + "[47] Jiajia Liao, Yujun Liu, Yingchao Piao, Jinhe Su, Guorong Cai, and Yundong Wu. Gle-net: A global and local ensemble network for aerial object detection. International Journal of Computational Intelligence Systems, 15(1):2, 2022. 7", + "[48] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence, 43(8):2647-2664, 2020. 7", + "[49] Hiep Anh Hoang and MyungsiK Yoo. 3onet: 3d detector for occluded object under obstructed conditions. IEEE Sensors Journal, 2023. 7", + "[50] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023. 7", + "[51] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7", + "[52] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25912", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision and Pattern Recognition, pages 908-917, 2022. 7", + "[53] Hai Wu, Chenglu Wen, Wei Li, Xin Li, Ruigang Yang, and Cheng Wang. Transformation-equivariant 3d object detection for autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2795-2802, 2023. 7", + "[54] Nils Lubbe, Yi Wu, and Hanna Jeppsson. Safe speeds: fatality and injury risks of pedestrians, cyclists, motorcyclists, and car drivers impacting the front of another passenger car as a function of closing speed and age. Traffic safety research, 2:000006-000006, 2022. 7", + "[55] Yilun Chen, Zhiding Yu, Yukang Chen, Shiyi Lan, Anima Anandkumar, Jiaya Jia, and Jose M Alvarez. Focalformer3d: focusing on hard instance for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8394-8405, 2023. 7", + "[56] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781, 2023. 7" + ], + "bbox": [ + 91, + 90, + 482, + 412 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25913", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/8be189e3-e5ec-449b-9eb4-67198582e1c8_model.json b/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/8be189e3-e5ec-449b-9eb4-67198582e1c8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f8b72df8165c25bf07af65a119716b1f8bacc97d --- /dev/null +++ b/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/8be189e3-e5ec-449b-9eb4-67198582e1c8_model.json @@ -0,0 +1,1993 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.895, + 0.178 + ], + "angle": 0, + "content": "Towards Accurate and Efficient 3D Object Detection for Autonomous Driving: A Mixture of Experts Computing System on Edge" + }, + { + "type": "text", + "bbox": [ + 0.265, + 0.203, + 0.737, + 0.277 + ], + "angle": 0, + "content": "Linshen Liu\\(^{1,*}\\) Boyan Su\\(^{1,*}\\) Junyue Jiang\\(^{1}\\) Guanlin Wu\\(^{1}\\) Cong Guo\\(^{2}\\) Ceyu Xu\\(^{3}\\) Hao Frank Yang\\(^{1,\\dagger}\\) \n\\(^{1}\\)Johns Hopkins University \\(^{2}\\)Duke University \\(^{3}\\)HKUST {11iu148, bsul1}@jh.edu, haofrankyang@jhu.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.309, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.486, + 0.705 + ], + "angle": 0, + "content": "This paper presents Edge-based Mixture of Experts (MoE) Collaborative Computing \\((EMC2^{\\ddagger})\\), an optimal computing system designed for autonomous vehicles (AVs) that simultaneously achieves low-latency and high-accuracy 3D object detection. Unlike existing works, the EMC2 introduces a novel scenario-aware MoE architecture optimized for fusing complementary sparse 3D point clouds and dense 2D images to achieve robust multimodal representations for detection. Furthermore, EMC2 integrates an adaptive multimodal data bridge with multi-scale region proposing and scenario-aware routing, dynamically dispatching features to complementary experts based on object visibility and distance. In addition, EMC2 integrates joint hardware-software optimizations, including hardware resource utilization optimization and computational graph simplification, to ensure efficient and real-time inference on resource-constrained edge devices. Experiments on open-source benchmarks clearly show the EMC2 advancements as an end-to-end system. On the KITTI dataset, it achieves an average accuracy improvement of \\(3.58\\%\\) and a \\(159.06\\%\\) inference speedup compared to 15 baseline methods on Jetson platforms, with similar performance gains on the nuScenes dataset, highlighting its capability to advance reliable, real-time 3D object detection tasks for AVs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.719, + 0.222, + 0.735 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.485, + 0.866 + ], + "angle": 0, + "content": "Traffic safety is a fundamental concern for both human drivers and Autonomous Driving Systems (ADS). Within an ADS, the perception module plays a pivotal role by serving as the \"eyes\" of the autonomous vehicle (AV), sensing the surrounding environment and delivering critical information to downstream modules such as motion planning, control, and safety guarantees. Existing conceptual safety of an AV perception system is generally defined by two core" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.293, + 0.88, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.597, + 0.908, + 0.669 + ], + "angle": 0, + "content": "Figure 1. Comparison of 3D object detection models in accuracy and inference efficiency on the KITTI and nuScenes dataset. Most existing methods improve accuracy at the expense of inference efficiency. The proposed EMC2 achieves win-win performance by customizing MoE for autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "dimensions: accuracy and efficiency. Accuracy ensures reliable object detection and tracking capabilities, enabling a credible understanding of the environment. Efficiency, characterized by low perception delay, is equally essential, as real-time high-level planning and control rely on fast information processing. Perception delay, defined as the sum of data acquisition, computation, and algorithm processing time, directly affects an ADS's responsiveness to dynamic environments. Research indicates that an ADS must process sensory inputs and issue feedback control commands within 100 milliseconds to guarantee prompt and safe vehicle operations [1]. Detection precision and perception delay together define a critical safety boundary: a precise but sluggish system cannot react in time, while a fast but inaccurate system may lead to unsafe decisions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution, †Corresponding author. \n‡EMC2 Code: https://github.com/LinshenLiu622/EMC2" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25903" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.484 + ], + "angle": 0, + "content": "As illustrated in Fig. 1, achieving both high accuracy and low latency remains a fundamental challenge. Existing approaches often face inherent trade-offs between these objectives due to limited computational resources, insufficient adaptability of perception systems, and the inherent complexity of dynamic driving scenarios. Urban traffic environments require low-latency and accurate sensing capabilities to handle complex, densely interactive surroundings, while highway scenarios demand reliable detection of distant objects and prompt responses to unexpected events. Moreover, even for the same object, recognition priorities may vary dynamically within a scene depending on distance and scene complexity, posing notable challenges for single-modality models in balancing efficiency and accuracy [2]. Multimodal models, such as large-scale CNN and Transformer-based architectures that integrate LiDAR, radar, and camera data [3-5], leverage the complementary advantages of diverse sensors to enhance perception accuracy. Nevertheless, these approaches [6-9] often impose substantial hardware overhead and exhibit limited adaptability, particularly when deployed on resource-constrained edge platforms [10]. Even with the assistance of low-power machine learning compilation frameworks [11], such systems may still execute redundant computations under straightforward scenarios, introducing additional delays that undermine timely safety-critical decision-making." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.489, + 0.482, + 0.789 + ], + "angle": 0, + "content": "Nevertheless, multimodal approaches provide a natural foundation for addressing the accuracy challenges faced by ADS. Rather than relying on a single integrated model designed to operate uniformly across diverse driving scenarios, the MoE framework [12] dynamically activates specialized sub-models, or experts, based on input characteristics and prior knowledge. MoE has demonstrated considerable effectiveness in large-scale applications, such as Large Language Models (LLMs) for natural language processing [13, 14], where it routes different tokens to appropriate experts, reducing redundant computations without sacrificing efficiency [15]. As model complexity grows, MoE enables models to scale their capacity while keeping inference hardware costs manageable. Deploying MoE in ADS, however, poses notable challenges [16], including strict real-time constraints, the need for multimodal data processing, and seamless end-to-end autonomous feedback control. Successful deployment requires efficient data preprocessing, adaptive modality selection, and robust expert routing mechanisms." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.796, + 0.482, + 0.9 + ], + "angle": 0, + "content": "By leveraging diverse sensor data and a scenario-adaptive MoE dispatcher, we dynamically assign expert sub-models based on scenario data and prior statistical knowledge, providing an efficient, energy-efficient, and safe solution for ADS. To this end, we introduce an Edge-based MoE Collaborative Computing (EMC2) framework - an algorithm-hardware-software co-design solution that not" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "only improves accuracy through complex multimodal experts but also effectively reduces latency by utilizing an adaptive MoE dispatcher and lightweight single-modal experts. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.153, + 0.905, + 0.273 + ], + "angle": 0, + "content": "1) A Multimodal 3D Object Detection Solution with MoE. The proposed EMC2 system achieves both high detection accuracy and inference efficiency, outperforming 15 existing one-size-fits-all baseline models. Designed for real-world ADS applications, EMC2 is optimized for deployment on resource-constrained platforms, making it more suitable for edge computing environments than generic alternatives." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.277, + 0.905, + 0.488 + ], + "angle": 0, + "content": "2) Robust and Efficient Multimodal Perception through Expert Collaboration and System-Level Optimization. The proposed EMC2 framework incorporates scenario-specific experts and employs an adaptive expert selection strategy to meet heterogeneous perception demands. To mitigate the training instability commonly observed in MoE-based systems, we design a hierarchical optimization scheme that combines expert-level and global EMC2 backpropagation for effective collaborative learning. Furthermore, a distribution-aware resampling technique enhances robustness under real-world data imbalance. In addition to these designs, a tailored multiscale pooling module improves the efficiency of LiDAR-camera fusion, ensuring both efficient and accurate multimodal perception." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.489, + 0.905, + 0.655 + ], + "angle": 0, + "content": "3) System-Level Optimization for Edge Inference. The original PyTorch-based MoE inference framework [17] suffers from inefficient management of memory and computational resources, leading to high latency on edge devices. To overcome these challenges, we propose an algorithm-hardware-software co-optimization strategy within EMC2, which incorporates efficient memory management techniques to improve L1/L2 cache hit rates and applies computational graph fusion to eliminate redundant data loading across layers. These combined optimizations notably improve inference efficiency." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.66, + 0.905, + 0.795 + ], + "angle": 0, + "content": "4) State-of-the-art (SOTA) Accuracy and Efficiency Combined. On the KITTI dataset, EMC2 achieves \\(3.28\\%\\) higher accuracy for pedestrians and \\(6.03\\%\\) for bicycles under hard-level detection challenge compared to 15 baseline models. On the nuScenes dataset, it yields a \\(3.9\\%\\) increase in mAP and a \\(1.8\\%\\) boost in NDS. In addition, EMC2 achieves a \\(159.06\\%\\) reduction in inference latency on the Jetson platform for KITTI, supporting its suitability for real-time edge deployment." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.803, + 0.654, + 0.818 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.826, + 0.905, + 0.9 + ], + "angle": 0, + "content": "LiDAR-only Method. The active 3D sensing of LiDAR makes it dominant in ADS 3D detection [18-20]. LiDAR-based methods process point clouds via voxelization [21], pillarization [20], or direct raw input [22]. Anchor-based heads, used for region proposal [21], and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25904" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "center-based representations[23] are widely adopted. These methods achieve high accuracy on KITTI, as demonstrated by 3DSSD [24], PartA2 [19], and Voxel R-CNN [3]. However, weak LiDAR reflections for distant or occluded objects degrade detection performance [25, 26]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.171, + 0.482, + 0.322 + ], + "angle": 0, + "content": "Multimodal Fusion Method. Combining LiDAR and camera RGB input enhances detection performance, especially for sparse LiDAR objects [27]. Fusion approaches include projection-based methods, which align image features with point clouds in 3D space [28], and model-based techniques that incorporate cross-attention [29] or unified feature spaces [30]. Recent multimodal models, such as PPF-Det [31] and RoboFusion [32], achieve top accuracy on KITTI benchmarks. However, these models often trade off accuracy for speed or vice versa." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.326, + 0.485, + 0.537 + ], + "angle": 0, + "content": "Edge-based Compilation. ML compilation platforms are essential for improving the inference efficiency of deep learning models across diverse hardware architectures [33, 34]. TensorRT [35] is an NVIDIA inference optimizer that accelerates ML models on GPUs through layer fusion, precision calibration, and kernel auto-tuning. XLA [36] is a domain-specific compiler for TensorFlow enhancing efficiency by fusing operations and optimizing computation graphs. Open Neural Network Exchange (ONNX) Runtime [37] provides an open standard for model interoperability, enabling efficient execution across CPUs, GPUs, and accelerators. TVM [11] is an open-source deep learning compiler that automates model tuning and tensor scheduling for efficient deployment across CPUs and GPUs." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.544, + 0.385, + 0.56 + ], + "angle": 0, + "content": "3. EMC2 Computing Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.751 + ], + "angle": 0, + "content": "We introduce EMC2, a multimodal MoE framework that minimizes latency and ensures reliable accuracy. Fig. 2 shows the EMC2 overview, EMC2 consists of: 1) Adaptive Multimodal Data Bridge (AMDB) to preprocess multimodal input (Sec. 3.1); 2) Scenario-Adaptive Dispatcher to adaptively dispatch a suitable expert for each scenario (Sec. 3.2); and 3) Three Scenario-Optimized Experts to decode AMDB-output features into final results (Sec. 3.3). We also design an effective training strategy to address the long-tail effect in the training process (Sec. 3.4). In addition, we deploy EMC2 on edge devices with self-designed 3D sparse convolution and Multiscale Pooling." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.758, + 0.39, + 0.774 + ], + "angle": 0, + "content": "3.1. Adaptive Multimodal Data Bridge" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Given the multimodal nature of EMC2's sensor inputs, a module providing suitable preprocessed data for different experts is required. Thus, we propose AMDB, a preprocessing unit to efficiently process the required input data for each expert. The workflow of AMDB begins with a UNet [38] and a sparse CNN to extract LiDAR features, followed by fully connected layers that generate proposal regions along with their confidence scores. These results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "serve as inputs to Latency-Prioritized Expert and Versatile Efficiency Expert. The extraction of image features depends on the Scenario-Adaptive Dispatcher. If further information is required for reliable detection, AMDB will extract image features by ResNet. Based on depth information, it projects relevant image pixels into 3D space, applying Multiscale Pooling to reduce computational overhead, and fuses them with LiDAR voxel features to produce a multimodal output. These results are inputs for Accuracy-Prioritized Experts. The Multiscale Pooling and multimodal fusion processes are demonstrated in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.267, + 0.782, + 0.284 + ], + "angle": 0, + "content": "3.2. Scenario-Adaptive Dispatcher" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.905, + 0.381 + ], + "angle": 0, + "content": "The Scenario-Adaptive Dispatcher (SAD) dynamically dispatches each scenario to the most suitable expert based on traffic conditions and latency-accuracy requirements. Our expert selection algorithm considers real-world traffic safety needs: specifically, how far away objects are and how clearly they can be perceived." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.905, + 0.806 + ], + "angle": 0, + "content": "So, the switch among experts is determined by two key parameters: object distance \\(\\mathcal{D}\\) and clarity, which is inferred from the confidence \\(\\mathcal{C}\\) of proposal regions provided by the AMDB. We classify traffic scenarios into three categories: 1) Close and Distinct Cases. All objects are clearly visible and no distant objects are present. This is characterized by all proposal regions \\(< \\mathcal{D}\\) and having confidence \\(\\geq \\mathcal{C}\\). In such scenarios, as near-field objects provide abundant Li-DAR information, even when compressed into bird's eye view (BEV), a 2D CNN can effectively capture key patterns such as object surfaces and heights [20], ensuring reliable detection accuracy with minimal latency (Sec. 4.4). These cases will be routed to the Latency-Prioritized Expert. 2) Mixed Visibility Cases. Some objects are either distant but clearly visible or near but unclear. Mathematically, this corresponds to cases where i) some proposal regions \\(< \\mathcal{D}\\) with confidence \\(< \\mathcal{C}\\), or ii) the distance of some objects \\(\\geq \\mathcal{D}\\) with confidence \\(\\geq \\mathcal{C}\\). Here, objects have insufficient voxel data, requiring more complex 3D convolution to extract meaningful patterns, and they will be dispatched to the Versatile Efficiency Expert. 3) Distant and Uncertain Cases. Some objects are both distant and unclear. This occurs when some proposal regions \\(\\geq \\mathcal{D}\\) with confidence \\(< \\mathcal{C}\\). In these challenging scenarios, significant voxel information is missing due to long distances, steep reflection angles, or occlusions, necessitating the integration of image data to compensate for missing LiDAR details. These cases will be processed by the Accuracy-Prioritized Expert." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.768, + 0.834 + ], + "angle": 0, + "content": "3.3. Scenario-Optimized Experts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Latency-Prioritized Expert (LPE). LPE uses 2D CNNs for object detection based on BEV projections. It first projects proposal regions from the AMDB into BEV space, then extracts 2D features for object localization and" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25905" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.431, + 0.907, + 0.542 + ], + "angle": 0, + "content": "Figure 2. System Structure of EMC2. The central row illustrates the architecture of EMC2, while the top and bottom rows provide visual examples of core modules and relevant legends. The system integrates five main components: an Adaptive Multimodal Data Bridge (AMDB) for preprocessing raw multimodal input and generating expert-specific features; a Scenario-Adaptive Dispatcher that dispatches suitable experts based on scenario-specific features (see Fig. 3); the Latency-Prioritized Expert designed for simple scenarios requiring minimal inference delay; the Versatile Efficiency Expert designed for uncertain or distant object detection scenarios; and the Accuracy-Prioritized Expert responsible for handling complex, high-precision detection scenarios. Additionally, an Emergency Expert API is designed to support rapid response to hazardous or unseen situations. The overall training process incorporates loss terms for both final detection results and intermediate outputs from the AMDB, as described in Sec. 3.4." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.554, + 0.482, + 0.705 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.713, + 0.483, + 0.743 + ], + "angle": 0, + "content": "Figure 3. Routing Strategy of the Scenario-Adaptive Dispatcher. The Routing criteria is detailed mathematically in Sec. 3.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.841 + ], + "angle": 0, + "content": "classification. It finally maps detection results back to 3D space. Compared to sparse 3D convolutions, its use of 2D operations reduces computation in outdoor ADS scenarios [20], allowing LPE to improve inference efficiency with minimal accuracy loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Versatile Efficiency Expert (VEE). VEE uses a 3D CNN to process proposal regions and confidence scores from the AMDB. After sparse convolution, voxel features are decoded into per-voxel classifications and bounding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.907, + 0.6 + ], + "angle": 0, + "content": "boxes. Compared to LPE, 3D convolutions better preserve spatial structure, making it more robust for scenarios with insufficient LiDAR signal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.767 + ], + "angle": 0, + "content": "Accuracy-Prioritized Expert (APE). APE extends VEE by incorporating the multimodal signal. It employs a 3D CNN and decoder with the same structure but different parameters to process multimodal proposal regions (from AMDB) and fuses image data to recover missing LiDAR information. Image features compressed via Multiscale Pooling (Sec.3.5) are projected into 3D space and fused with corresponding voxels of LiDAR features, ensuring more robust understanding for occluded and distant objects while keeping computation feasible. The Multiscale Pooling and multimodal fusion are detailed in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.905, + 0.79 + ], + "angle": 0, + "content": "3.4. Collaborative Training of Multimodal Experts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Hierarchical Training Strategy: Multimodal Supervision with Triple Back-Propagation. During training, the performance of experts is highly dependent on the proposal regions produced by the AMDB, yet these proposals tend to be unreliable in the early training phase, leading to an unstable expert learning process and unstable SAD decisions. To address this, we adopt a hierarchical supervision stra" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25906" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.09, + 0.442, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.374, + 0.484, + 0.457 + ], + "angle": 0, + "content": "Figure 4. Illustration of Hierarchical Training. The three backpropagation paths originate from: 1) the final detection output, 2) the LiDAR branch output of AMDB, and 3) the image branch output of AMDB. Since expert training relies on initially unreliable AMDB outputs, direct supervision accelerates convergence and enhances representation learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.484, + 0.607 + ], + "angle": 0, + "content": "egy with three distinct back-propagation pathways, as illustrated in Fig. 4: 1) supervision of the LiDAR branch outputs from the AMDB; 2) supervision of the image branch outputs from the AMDB; and 3) supervision of expert predictions and associated confidence scores. To further improve training stability, the LiDAR and image branches of the AMDB are pre-trained separately before joint optimization. Each back-propagation route is supervised by the following loss function, which collectively forms Multimodal Loss:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.618, + 0.482, + 0.635 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r o u t e}} = \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {r e g}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.734 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{cls}}\\) uses cross-entropy loss and \\(\\mathcal{L}_{\\mathrm{reg}}\\) uses smooth-\\(\\ell 1\\) loss. The Hierarchical Training not only stabilizes expert learning but also enhances the model's adaptability to various data granularities. In turn, even under compressed data representations, it enables the effective use of our Multiscale Pooling for memory efficiency, as detailed in Sec. 3.5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Addressing Long-tail Effects in MoE Training. The MoE framework often suffers from long-tail effects, where certain experts receive insufficient training data, leading to imbalances in sub-datasets, and therefore affecting performance [39]. To address this, we propose a threefold training strategy: 1) Data subset division. We partition the dataset into sub-datasets based on target and auxiliary samples. Each expert \\(\\mathcal{E}_i\\) is dispatched a sub-dataset \\(S_{i}\\) \\((i = 1,2,3)\\) containing target samples \\(\\mathcal{T}_{\\mathcal{S}_i}\\), which are expected to be routed to the corresponding expert, and auxiliary samples \\(\\mathcal{A}_{\\mathcal{S}_i}\\), which are not. We ensure that \\(\\mathcal{T}_{\\mathcal{S}_i} \\cap \\mathcal{T}_{\\mathcal{S}_j} = \\emptyset\\) for any" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.906, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.257, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Figure 5. Illustration of Multiscale Pooling and Multimodal Fusion. Proposal regions with image features are first obtained from the 3D proposal space generated by the LiDAR branch of AMDB. After extracting the relevant image features from the corresponding image regions and projecting them into 3D space, pooling and fusion are then applied. Following the pooling step, image features of pixels that share the same 3D coordinate as a voxel are concatenated with the corresponding LiDAR voxel features, while those without matching voxel coordinates are concatenated with zero-filled features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.4, + 0.906, + 0.476 + ], + "angle": 0, + "content": "\\(i \\neq j\\). The inclusion of auxiliary samples, which are not intended for a given expert, helps prevent overfitting and enhances generalization. 2) Balanced sampling strategy. We adjust the selection probability \\(\\mathcal{P}_i\\) for the sub-datasets which have fewer samples to ensure balanced training:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.479, + 0.905, + 0.52 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {i} \\times \\mathcal {N} _ {\\mathcal {S} _ {i}} = \\mathcal {P} _ {j} \\times \\mathcal {N} _ {\\mathcal {S} _ {j}}, \\sum_ {i = 1} ^ {3} \\mathcal {P} _ {i} = 1 \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.906, + 0.631 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}_{S_i}\\) represents the number of samples in sub-datasets \\(S_{i}\\). This ensures that experts dispatched to less frequent scenarios receive sufficient training iterations. 3) Adaptive optimizer. We introduce an optimizer that adapts the learning rate based on the proportion of target samples in a batch. Denoting \\(\\alpha_0\\) as the base learning rate, the adaptive learning rate applied to an expert with a given batch is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.639, + 0.905, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\alpha = \\left(1 + \\frac {1}{\\mathcal {N}} \\sum_ {i = 1} ^ {\\mathcal {N}} \\mathbf {1} _ {i \\in \\mathcal {T}}\\right) \\cdot \\alpha_ {0} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.78 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}\\) is the batch size, \\(p\\) is the proportion of target samples within current mini-batch, \\(\\mathcal{T}\\) represents target samples in the batch. By regulating the gradient update rate based on the batch composition, this adaptive optimizer encourages each expert to prioritize its designated target samples while preventing excessive updates from auxiliary samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.863, + 0.805 + ], + "angle": 0, + "content": "3.5. Algorithm Edge-adaptivity Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Edge devices face challenges such as incomplete function libraries and limited cache. We adopt the open-source ONNX Runtime for compiling EMC2 on edge, which provides modules that help address these challenges and facilitate edge-oriented optimization. As 3D convolutions account for the majority of computation during inference, we" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25907" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "customize 3D sparse convolution and Multiscale Pooling to improve the algorithmic efficiency of EMC2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.127, + 0.484, + 0.278 + ], + "angle": 0, + "content": "3D Sparse Convolution. Since the 3D sparse convolution library is incomplete on most edge platforms, we develop a customized 3D sparse convolution library with parallel threads execution under ONNX Runtime. Let \\(\\mathcal{H}\\) denote spatial resolution, \\(\\mathcal{N}\\) the number of non-empty voxels, and \\(\\mathcal{C}_{\\mathrm{v}}\\) the per-voxel computation, including channels and kernel size. By computing nonzero voxels only, 3D sparse convolution reduces complexity from \\(O(\\mathcal{H}^3\\times \\mathcal{C}_{\\mathrm{v}})\\) for dense convolutions to \\(O(\\mathcal{N}\\times \\mathcal{C}_{\\mathrm{v}})\\), where \\(\\mathcal{N}\\ll \\mathcal{H}^3\\), cutting redundant operations by \\(65 - 80\\%\\) [40]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.484, + 0.404 + ], + "angle": 0, + "content": "Multiscale Pooling. To alleviate cache limitations, Multiscale 3D Sparse Pooling adaptively pools image features based on available memory before fusing them with LiDAR features (Fig.5). The pooling size is user-configurable, allowing flexibility for maximizing memory utilization. When combined with Hierarchical Training (Sec.3.4), Multiscale Pooling effectively reduces resource consumption while maintaining model reliability (Sec. 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.418, + 0.436, + 0.435 + ], + "angle": 0, + "content": "3.6. EMC2 Computing System Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.517 + ], + "angle": 0, + "content": "Building upon the algorithm-level optimizations introduced in Sec. 3.5, we further enhance the ONNX execution efficiency through system-level improvements, specifically targeting memory management and computation graph optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Memory Optimization. To further improve memory efficiency during inference, we implement three system-level techniques. 1) Overlapping communication and computation [41]. Matrix data is partitioned into segments processed and transferred in parallel, overlapping computation and communication. This reduces global memory access, improves cache usage, and mitigates memory bottlenecks. 2) Thread scheduling. We implement a staged thread management mechanism, where active threads are periodically terminated and re-initialized (Fig. 6). This approach limits unnecessary memory retention between stages, reducing system memory footprint. 3) Prefix-sum for sparse convolution. Parallel prefix-sum reduces index search time from \\(\\mathcal{O}(\\mathcal{N})\\) to \\(\\mathcal{O}(\\log \\mathcal{N})\\) using GPU parallelism, where \\(\\mathcal{N}\\) is the number of nonzero elements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Computational Graph Optimization. ONNX partitions the computation graph into segments and assigns each to the most suitable hardware. Our optimization includes three stages: 1) Model pruning. Redundant parameters and operations are removed to simplify the graph; 2) Model quantization. Low-precision arithmetic replaces floating-point computation to improve hardware efficiency; 3) Computational graph fusion. Consecutive operators are merged to reduce memory access and improve execution, benefiting memory-bound CPUs and reducing kernel" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.694, + 0.106 + ], + "angle": 0, + "content": "launch overhead on GPUs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.119, + 0.733, + 0.136 + ], + "angle": 0, + "content": "4. Experiment and Result" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.141, + 0.809, + 0.157 + ], + "angle": 0, + "content": "4.1. Experiment Settings and Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.163, + 0.905, + 0.239 + ], + "angle": 0, + "content": "Experiment Platform. Experiments are conducted on the Jetson AGX Orin, an edge platform with limited computational resources, an NVIDIA Ampere GPU, a 4MB shared L2 cache, and hierarchical memory. For comparison, EMC2 is also evaluated on an NVIDIA A4000 GPU." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.907, + 0.606 + ], + "angle": 0, + "content": "Dataset and Sample Difficulties. We evaluate 3D object detection on two public benchmarks: KITTI [42] and nuScenes [43]. KITTI provides 3,712 training and 3,769 validation samples, with object difficulty levels categorized as \\(30\\%\\) easy, \\(40\\%\\) moderate, and \\(30\\%\\) hard. nuScenes contains 850 training and 150 test scenes, covering 10 object categories with 3D bounding box annotations. Before splitting the data, we analyze the confidence distribution of proposal regions generated by the pre-trained AMDB. A two-sample Kolmogorov-Smirnov test reveals a statistically significant difference between objects within and beyond a distance threshold \\(D\\), which is empirically set to 23.5 meters for KITTI and 35 meters for nuScenes. The training subsets for each expert on the KITTI and nuScenes datasets are defined as follows: \\(\\mathcal{T}_{S_1}\\) includes scenes where all objects are within \\(D\\) meters and labeled as \"easy\" or \"moderate\"; \\(\\mathcal{T}_{S_2}\\) includes scenes with at least one object beyond \\(D\\) meters labeled as \"easy\" or \"moderate,\" or any object within \\(D\\) meters labeled as \"hard\"; and \\(\\mathcal{T}_{S_3}\\) includes scenes where at least one object is both beyond \\(D\\) meters and labeled as \"hard,\" requiring multimodal processing. For each expert, the auxiliary subset \\(\\mathcal{A}_{S_i}\\) is uniformly sampled from the other experts' target subsets, with its size equal to that of the corresponding target subset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.746 + ], + "angle": 0, + "content": "Evaluation Metrics. KITTI uses R40 3D Average Precision (AP) with IoU thresholds of 0.7 for cars and 0.5 for pedestrians and cyclists. nuScenes adopts the NDS metric, combining mAP—computed via center distance matching \\((0.5\\mathrm{m}\\) to \\(4.0\\mathrm{m}\\) thresholds)—and five error terms: mean Average Translation (mATE), Scale (mASE), Orientation (mAOE), Velocity (mAVE), and Attribute (mAAE) Errors. Inference latency on Jetson AGX Orin is also reported to assess real-time performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.906, + 0.837 + ], + "angle": 0, + "content": "Hierarchical Training Details. Allowing all experts to propagate gradients to AMDB causes instability due to overlapping training data, leading to repeated updates. To address this, AMDB is first pre-trained for 20 epochs, followed by joint training where gradients are propagated only from APE, which receives full multimodal input." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.874, + 0.865 + ], + "angle": 0, + "content": "4.2. Comparison with Existing SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 1 and Tab. 2, EMC2 achieves SOTA accuracy and latency on KITTI and nuScenes. On the Jetson" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25908" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.092, + 0.862, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.252, + 0.907, + 0.281 + ], + "angle": 0, + "content": "Figure 6. Illustration of Thread Management. Each solid box represents a module occupying threads. During inference, four boundaries are defined at which preceding threads are terminated, and their thread-specific address space is released upon crossing each boundary." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.285, + 0.909, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.558, + 0.908, + 0.587 + ], + "angle": 0, + "content": "Table 1. We evaluate different 3D object detection methods on the KITTI dataset. Some cells are denoted as \\( N / A \\) when the corresponding open-source implementation lacks support for the specific object class or is not compatible with Jetson devices." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.596, + 0.477, + 0.66 + ], + "angle": 0, + "content": "
MethodSizemAP↑NDS↑Latency (ms)↓
FocalFormer [55]189M0.66400.7090N/A
BEVFusion [56]156M0.68520.7138N/A
EMC2 (ours)87M0.72410.7316229.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.666, + 0.483, + 0.681 + ], + "angle": 0, + "content": "Table 2. Comparison of different methods on the nuScenes dataset." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.708, + 0.483, + 0.739 + ], + "angle": 0, + "content": "AGX Orin platform, it also outperforms existing baseline models across all difficulty levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Results on KITTI. Pedestrians and cyclists face higher traffic risks [54], which remain challenging for most SOTA methods. EMC2 improves hard-set detection by \\(5\\%\\) and \\(7\\%\\) for pedestrians and cyclists, respectively, outperforming TED and LoGoNet (Tab. 1). For cars, it achieves \\(88.15\\%\\) on the hard set, exceeding prior multimodal methods by up to \\(9\\%\\). It runs at \\(372.5 \\mathrm{~ms}\\) on Jetson, \\(2.5 \\times -2.5 \\times\\) faster than methods requiring \\(965 - 1\\), \\(787 \\mathrm{~ms}\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Results on nuScenes. EMC2 improves mAP and NDS by \\(3.9\\%\\) and \\(1.8\\%\\), respectively, outperforming recent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.645 + ], + "angle": 0, + "content": "methods FocalFormer [55] and BEVFusion [56] (Tab. 2). It runs at \\(229.3\\mathrm{ms}\\) on Jetson AGX Orin, enabling real-time use on edge devices." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.657, + 0.734, + 0.673 + ], + "angle": 0, + "content": "4.3. Analysis and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.907, + 0.756 + ], + "angle": 0, + "content": "Benefits of Mixture-of-Experts. The balance between accuracy and efficiency observed in Sec. 4.2 is achieved through dynamic expert selection based on scenario characteristics, combined with algorithmic and system-level optimizations designed for edge devices, such as Jetson." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.76, + 0.909, + 0.882 + ], + "angle": 0, + "content": "Empirical Dispatcher. The EMC2 dispatcher uses fixed distance and confidence thresholds to dispatch scenes to experts. The protocol is adapted per dataset format. This strategy is validated via expert activations on the nuScenes validation set (Fig. 7): LPE performs best in close-range, high-confidence scenes; APE in distant, low-confidence cases; and VEE in all others, supporting the effectiveness of empirical routing." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Optimized Parameter Size. EMC2 has 226M param-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25909" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.476, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.238, + 0.485, + 0.351 + ], + "angle": 0, + "content": "Figure 7. Distribution of expert activation. During EMC2 inference on the nuScenes validation set, scenes are grouped into a 2D grid based on two metrics: the y-axis indicates the number of AMDB-generated proposals beyond the distance threshold \\( D \\) (i.e., distant objects), and the x-axis represents the average confidence score of these proposals (higher values indicate clearer scenes). Each cell is labeled with the name of the expert that performs best for scenes falling into that bin." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.362, + 0.476, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.519, + 0.483, + 0.603 + ], + "angle": 0, + "content": "Figure 8. Median Detection Confidence vs. Object Distances. Four configurations are compared on the KITTI validation set: LPE (blue), VEE (peach), APE (rose), and EMC2 (green). More complex experts yield higher confidence but slower inference. By switching experts based on distance and clarity, EMC2 balances accuracy and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.483, + 0.713 + ], + "angle": 0, + "content": "eters, but its MoE framework activates only a subset during inference, averaging 87M on nuScenes and 146M on KITTI, which is fewer than those used by the full model and other SOTA baselines. System-level optimizations instead prioritize computational throughput and memory efficiency over parameter reduction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Hardware Flexibility. The EMC2 is implemented in the ONNX format, an open format supported by various compilers and enables model compilation across diverse hardware platforms. As a result, EMC2 is not limited to Jetson and can be deployed on other edge platforms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.245, + 0.819 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablation on the Roles of Three Experts. We evaluate each expert on its target scenarios using the KITTI validation set. As shown in Tab. 3, LPE, VEE, and APE achieve comparable accuracy in LPE scenes, but LPE has the lowest inference latency on Jetson. In VEE scenarios, incor" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.894, + 0.329 + ], + "angle": 0, + "content": "
MethodAP↑Latency (ms) ↓
Ped.CarCyclist
Ablation on LPE
LPE66.3289.0488.54219
VEE74.4992.2391.69468
APE74.3493.2192.41526
Ablation on VEE
W/ VEE70.5590.5084.07372.5
W/O VEE68.2686.4079.17512.7
Ablation on APE
VEE67.2487.6981.85468
APE70.5590.5084.18526
Ablation on Hierarchical Training and Multiscale Pooling
HT √, MP √70.5590.5084.07526
HT √, MP ×70.5590.5084.181320
HT ×, MP √63.1188.6875.98529
HT ×, MP ×66.3489.1080.831367
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.339, + 0.905, + 0.409 + ], + "angle": 0, + "content": "Table 3. Results Across Three Ablation Configurations. The main body of this table is divided into four sections corresponding to each part of the ablation study. In the third section, HT and MP are abbreviated for Hierarchical Training and Multiscale Pooling, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.905, + 0.482 + ], + "angle": 0, + "content": "porating VEE improves both accuracy and latency over the baseline. For APE, which incorporates image features, accuracy improves further with only a minor latency increase, supported by the score distribution in Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.636 + ], + "angle": 0, + "content": "Ablation on Hierarchical Training (HT) and Multiscale Pooling (MP). We evaluate four combinations of HT and MP on the full KITTI validation set: both applied, HT only, MP only, and neither. Tab. 3 shows that combining HT and MP reduces inference time on Jetson from 1320 ms to 526 ms without accuracy loss. For the car class, MP alone reduces accuracy from \\(80.83\\%\\) to \\(75.98\\%\\), while HT alone improves it from \\(66.34\\%\\) to \\(70.55\\%\\). HT facilitates multigranular feature learning, while MP enhances memory efficiency via compression." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.65, + 0.634, + 0.665 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.907, + 0.823 + ], + "angle": 0, + "content": "We present EMC2, a multimodal 3D object detection solution for ADS based on a customized MoE architecture. This design enables adaptive expert selection across diverse traffic scenarios, balancing detection accuracy and efficiency. Experiments on KITTI and nuScenes show strong performance, particularly under challenging conditions. On the Jetson platform, EMC2 achieves a \\(159.06\\%\\) speedup in inference, supporting real-time edge deployment. Future work will explore fully adaptive expert selection to enhance system responsiveness and resource efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25910" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[1] Shih-Chieh Lin, Yunqi Zhang, Chang-Hong Hsu, Matt Skach, Md E. Haque, Lingjia Tang, and Jason Mars. The architectural implications of autonomous driving: Constraints and acceleration. In Proceedings of the Twenty-Third International Conference on Architectural Support for Programming Languages and Operating Systems, ASPLOS '18, page 751-766, New York, NY, USA, 2018. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.227, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[2] Yi Xiao, Felipe Codevilla, Akhil Gurram, Onay Urfalioglu, and Antonio M López. Multimodal end-to-end autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(1):537-547, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.482, + 0.338 + ], + "angle": 0, + "content": "[3] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. arXiv:2012.15712, 2020. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[4] Hao Frank Yang, Yifan Ling, Cole Kopca, Sam Ricord, and Yinhai Wang. Cooperative traffic signal assistance system for non-motorized users and disabilities empowered by computer vision and edge artificial intelligence. Transportation Research Part C: Emerging Technologies, 145:103896, 2022. Elsevier." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.482, + 0.491 + ], + "angle": 0, + "content": "[5] Hao Yang, Jiarui Cai, Meixin Zhu, Chenxi Liu, and Yinhai Wang. Traffic-informed multi-camera sensing (TIMS) system based on vehicle re-identification. IEEE Transactions on Intelligent Transportation Systems, 23(10):17189-17200, 2022. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[6] Guibiao Liao, Jiankun Li, and Xiaqing Ye. Vlm2scene: Self-supervised image-text-lidar learning with foundation models for autonomous driving scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 3351-3359, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[7] Hao Frank Yang, Yang Zhao, Jiarui Cai, Meixin Zhu, JenqNeng Hwang, and Yiran Chen. Mitigating bias of deep neural networks for trustworthy traffic perception in autonomous systems. In Proceedings of the 2024 IEEE Intelligent Vehicles Symposium (IV), pages 633-638, 2024. IEEE." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[8] Chenxi Liu, Hao Yang, Ruimin Ke, Wei Sun, Julia Wang, and Yinhai Wang. Cooperative and comprehensive multitask surveillance sensing and interaction system empowered by edge artificial intelligence. Transportation Research Record, 2677(9):652-668, 2023. SAGE Publications." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[9] Hao Frank Yang, Jiarui Cai, Chenxi Liu, Ruimin Ke, and Yinhai Wang. Cooperative multi-camera vehicle tracking and traffic surveillance with edge artificial intelligence and representation learning. Transportation Research Part C: Emerging Technologies, 148:103982, 2023. Elsevier. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[10] Chen Zhang, Peng Li, Guangyu Sun, Yijin Guan, Bingjun Xiao, and Jason Cong. Optimizing fpga-based accelerator design for deep convolutional neural networks. In Proceedings of the 2015 ACM/SIGDA international symposium on field-programmable gate arrays, pages 161-170, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Tianqi Chen, Thierry Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Yan, Haichen Shen, Meghan Cowan, Leyuan Wang, Yuwei Hu, Luis Ceze, et al. {TVM}: An automated {End-to-End} optimizing compiler for deep learning. In 13th" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "USENIX Symposium on Operating Systems Design and Implementation (OSDI 18), pages 578-594, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[12] Jiarui Cai, Yizhou Wang, and Jenq-Neng Hwang. Ace: Ally complementary experts for solving long-tailed recognition in one-shot. In Proceedings of the IEEE/CVF international conference on computer vision, pages 112-121, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[13] Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. Unimoe: Scaling unified multimodal llms with mixture of experts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[14] Jiachen Li, Xinyao Wang, Sijie Zhu, Chia-Wen Kuo, Lu Xu, Fan Chen, Jitesh Jain, Humphrey Shi, and Longyin Wen. Cumo: Scaling multimodal llm with co-upcycled mixture-of-experts. Advances in Neural Information Processing Systems, 37:131224-131246, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[15] Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts. arXiv preprint arXiv:2407.06204, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[16] Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. Time-moe: Billion-scale time series foundation models with mixture of experts. arXiv preprint arXiv:2409.16040, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[17] Jiaao He, Jiezhong Qiu, Aohan Zeng, Zhilin Yang, Jidong Zhai, and Jie Tang. Fastmoe: A fast mixture-of-expert training system. arXiv preprint arXiv:2103.13262, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[18] Charles R. Qi, Xinlei Chen, Or Litany, and Leonidas J. Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4403-4412, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[19] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. arXiv preprint arXiv:1907.03670, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[20] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689-12697, 2019. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[21] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[22] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10526–10535, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[23] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11779-11788, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11037-11045, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25911" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[25] Andrew M Wallace, Abderrahim Halimi, and Gerald S Buller. Full waveform lidar for adverse weather conditions. IEEE transactions on vehicular technology, 69(7):7064-7077, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.484, + 0.203 + ], + "angle": 0, + "content": "[26] You Li, Pierre Duthon, Michele Colomb, and Javier Ibanez-Guzman. What happens for a tof lidar in fog? IEEE Transactions on Intelligent Transportation Systems, 22(11):6670-6681, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.484, + 0.273 + ], + "angle": 0, + "content": "[27] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In 2019 international conference on robotics and automation (ICRA), pages 3288-3295. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[28] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. E-net: Enhancing point features with image semantics for 3d object detection. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 35–52, Cham, 2020. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.484, + 0.453 + ], + "angle": 0, + "content": "[29] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quaan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V. Le, Alan Yuille, and Mingxing Tan. Deep-fusion: Lidar-camera deep fusion for multi-modal 3d object detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17161-17170, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[30] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[31] Guotao Xie, Zhiyuan Chen, Ming Gao, Manjiang Hu, and Xiaohui Qin. Ppf-det: Point-pixel fusion for multi-modal 3d object detection. IEEE Transactions on Intelligent Transportation Systems, 25(6):5598-5611, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[32] Ziying Song, Guoxing Zhang, Lin Liu, Lei Yang, Shaoqing Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Robofusion: Towards robust multi-modal 3d object detection via sam. arXiv preprint arXiv:2401.03907, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.484, + 0.663 + ], + "angle": 0, + "content": "[33] Zheng Wang and Michael O'Boyle. Machine learning in compiler optimization. Proceedings of the IEEE, 106(11):1879-1901, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.484, + 0.719 + ], + "angle": 0, + "content": "[34] Hugh Leather, Edwin Bonilla, and Michael O'boyle. Automatic feature generation for machine learning-based optimising compilation. ACM Transactions on Architecture and Code Optimization (TACO), 11(1):1-32, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[35] NVIDIA Corporation. NVIDIA TensorRT Developer Guide, 2024. Accessed: 2025-02-13. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.774 + ], + "angle": 0, + "content": "[36] Amit Sabne. XLA: Compiling Machine Learning for Peak Performance, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[37] ONNX Community. Onnx: Open neural network exchange. https://github.com/onnx/onnx, 2024. Accessed: YYYYY-MM-DD. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[38] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[39] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. *Neural computation*, 3(1):79–87, 1991. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[40] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10), 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.207 + ], + "angle": 0, + "content": "[41] Juraj Hromkovič. Communication complexity and parallel computing. Springer Science & Business Media, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.907, + 0.263 + ], + "angle": 0, + "content": "[42] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.265, + 0.907, + 0.348 + ], + "angle": 0, + "content": "[43] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.391 + ], + "angle": 0, + "content": "[44] Muhammad Aslam. Introducing kolmogorov-smirnov tests under uncertainty: an application to radioactive data. ACS omega, 5(1):914-917, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.907, + 0.474 + ], + "angle": 0, + "content": "[45] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 772-782, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.907, + 0.546 + ], + "angle": 0, + "content": "[46] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.907, + 0.603 + ], + "angle": 0, + "content": "[47] Jiajia Liao, Yujun Liu, Yingchao Piao, Jinhe Su, Guorong Cai, and Yundong Wu. Gle-net: A global and local ensemble network for aerial object detection. International Journal of Computational Intelligence Systems, 15(1):2, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.907, + 0.673 + ], + "angle": 0, + "content": "[48] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence, 43(8):2647-2664, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[49] Hiep Anh Hoang and MyungsiK Yoo. 3onet: 3d detector for occluded object under obstructed conditions. IEEE Sensors Journal, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.907, + 0.8 + ], + "angle": 0, + "content": "[50] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.907, + 0.857 + ], + "angle": 0, + "content": "[51] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[52] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25912" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Computer Vision and Pattern Recognition, pages 908-917, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[53] Hai Wu, Chenglu Wen, Wei Li, Xin Li, Ruigang Yang, and Cheng Wang. Transformation-equivariant 3d object detection for autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2795-2802, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[54] Nils Lubbe, Yi Wu, and Hanna Jeppsson. Safe speeds: fatality and injury risks of pedestrians, cyclists, motorcyclists, and car drivers impacting the front of another passenger car as a function of closing speed and age. Traffic safety research, 2:000006-000006, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[55] Yilun Chen, Zhiding Yu, Yukang Chen, Shiyi Lan, Anima Anandkumar, Jiaya Jia, and Jose M Alvarez. Focalformer3d: focusing on hard instance for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8394-8405, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[56] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "25913" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/8be189e3-e5ec-449b-9eb4-67198582e1c8_origin.pdf b/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/8be189e3-e5ec-449b-9eb4-67198582e1c8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a40861179891efa3cb8529782732cd75788219e1 --- /dev/null +++ b/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/8be189e3-e5ec-449b-9eb4-67198582e1c8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eaac85331e3b8227a52447fe1384a1cdcfc00c0cf97f3679066fa7299eb3c695 +size 1002749 diff --git a/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/full.md b/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d90e593d590fafa35d2787048ee0267954a2006d --- /dev/null +++ b/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/full.md @@ -0,0 +1,263 @@ +# Towards Accurate and Efficient 3D Object Detection for Autonomous Driving: A Mixture of Experts Computing System on Edge + +Linshen Liu $^{1,*}$ Boyan Su $^{1,*}$ Junyue Jiang $^{1}$ Guanlin Wu $^{1}$ Cong Guo $^{2}$ Ceyu Xu $^{3}$ Hao Frank Yang $^{1,\dagger}$ $^{1}$ Johns Hopkins University $^{2}$ Duke University $^{3}$ HKUST {11iu148, bsul1}@jh.edu, haofrankyang@jhu.edu + +# Abstract + +This paper presents Edge-based Mixture of Experts (MoE) Collaborative Computing $(EMC2^{\ddagger})$ , an optimal computing system designed for autonomous vehicles (AVs) that simultaneously achieves low-latency and high-accuracy 3D object detection. Unlike existing works, the EMC2 introduces a novel scenario-aware MoE architecture optimized for fusing complementary sparse 3D point clouds and dense 2D images to achieve robust multimodal representations for detection. Furthermore, EMC2 integrates an adaptive multimodal data bridge with multi-scale region proposing and scenario-aware routing, dynamically dispatching features to complementary experts based on object visibility and distance. In addition, EMC2 integrates joint hardware-software optimizations, including hardware resource utilization optimization and computational graph simplification, to ensure efficient and real-time inference on resource-constrained edge devices. Experiments on open-source benchmarks clearly show the EMC2 advancements as an end-to-end system. On the KITTI dataset, it achieves an average accuracy improvement of $3.58\%$ and a $159.06\%$ inference speedup compared to 15 baseline methods on Jetson platforms, with similar performance gains on the nuScenes dataset, highlighting its capability to advance reliable, real-time 3D object detection tasks for AVs. + +# 1. Introduction + +Traffic safety is a fundamental concern for both human drivers and Autonomous Driving Systems (ADS). Within an ADS, the perception module plays a pivotal role by serving as the "eyes" of the autonomous vehicle (AV), sensing the surrounding environment and delivering critical information to downstream modules such as motion planning, control, and safety guarantees. Existing conceptual safety of an AV perception system is generally defined by two core + +![](images/7037e73d354f54394db0640ae8f77d2d26d60527926c16fe7628081fb33e055b.jpg) +Figure 1. Comparison of 3D object detection models in accuracy and inference efficiency on the KITTI and nuScenes dataset. Most existing methods improve accuracy at the expense of inference efficiency. The proposed EMC2 achieves win-win performance by customizing MoE for autonomous driving. + +dimensions: accuracy and efficiency. Accuracy ensures reliable object detection and tracking capabilities, enabling a credible understanding of the environment. Efficiency, characterized by low perception delay, is equally essential, as real-time high-level planning and control rely on fast information processing. Perception delay, defined as the sum of data acquisition, computation, and algorithm processing time, directly affects an ADS's responsiveness to dynamic environments. Research indicates that an ADS must process sensory inputs and issue feedback control commands within 100 milliseconds to guarantee prompt and safe vehicle operations [1]. Detection precision and perception delay together define a critical safety boundary: a precise but sluggish system cannot react in time, while a fast but inaccurate system may lead to unsafe decisions. + +As illustrated in Fig. 1, achieving both high accuracy and low latency remains a fundamental challenge. Existing approaches often face inherent trade-offs between these objectives due to limited computational resources, insufficient adaptability of perception systems, and the inherent complexity of dynamic driving scenarios. Urban traffic environments require low-latency and accurate sensing capabilities to handle complex, densely interactive surroundings, while highway scenarios demand reliable detection of distant objects and prompt responses to unexpected events. Moreover, even for the same object, recognition priorities may vary dynamically within a scene depending on distance and scene complexity, posing notable challenges for single-modality models in balancing efficiency and accuracy [2]. Multimodal models, such as large-scale CNN and Transformer-based architectures that integrate LiDAR, radar, and camera data [3-5], leverage the complementary advantages of diverse sensors to enhance perception accuracy. Nevertheless, these approaches [6-9] often impose substantial hardware overhead and exhibit limited adaptability, particularly when deployed on resource-constrained edge platforms [10]. Even with the assistance of low-power machine learning compilation frameworks [11], such systems may still execute redundant computations under straightforward scenarios, introducing additional delays that undermine timely safety-critical decision-making. + +Nevertheless, multimodal approaches provide a natural foundation for addressing the accuracy challenges faced by ADS. Rather than relying on a single integrated model designed to operate uniformly across diverse driving scenarios, the MoE framework [12] dynamically activates specialized sub-models, or experts, based on input characteristics and prior knowledge. MoE has demonstrated considerable effectiveness in large-scale applications, such as Large Language Models (LLMs) for natural language processing [13, 14], where it routes different tokens to appropriate experts, reducing redundant computations without sacrificing efficiency [15]. As model complexity grows, MoE enables models to scale their capacity while keeping inference hardware costs manageable. Deploying MoE in ADS, however, poses notable challenges [16], including strict real-time constraints, the need for multimodal data processing, and seamless end-to-end autonomous feedback control. Successful deployment requires efficient data preprocessing, adaptive modality selection, and robust expert routing mechanisms. + +By leveraging diverse sensor data and a scenario-adaptive MoE dispatcher, we dynamically assign expert sub-models based on scenario data and prior statistical knowledge, providing an efficient, energy-efficient, and safe solution for ADS. To this end, we introduce an Edge-based MoE Collaborative Computing (EMC2) framework - an algorithm-hardware-software co-design solution that not + +only improves accuracy through complex multimodal experts but also effectively reduces latency by utilizing an adaptive MoE dispatcher and lightweight single-modal experts. Our main contributions are summarized as follows: + +1) A Multimodal 3D Object Detection Solution with MoE. The proposed EMC2 system achieves both high detection accuracy and inference efficiency, outperforming 15 existing one-size-fits-all baseline models. Designed for real-world ADS applications, EMC2 is optimized for deployment on resource-constrained platforms, making it more suitable for edge computing environments than generic alternatives. + +2) Robust and Efficient Multimodal Perception through Expert Collaboration and System-Level Optimization. The proposed EMC2 framework incorporates scenario-specific experts and employs an adaptive expert selection strategy to meet heterogeneous perception demands. To mitigate the training instability commonly observed in MoE-based systems, we design a hierarchical optimization scheme that combines expert-level and global EMC2 backpropagation for effective collaborative learning. Furthermore, a distribution-aware resampling technique enhances robustness under real-world data imbalance. In addition to these designs, a tailored multiscale pooling module improves the efficiency of LiDAR-camera fusion, ensuring both efficient and accurate multimodal perception. + +3) System-Level Optimization for Edge Inference. The original PyTorch-based MoE inference framework [17] suffers from inefficient management of memory and computational resources, leading to high latency on edge devices. To overcome these challenges, we propose an algorithm-hardware-software co-optimization strategy within EMC2, which incorporates efficient memory management techniques to improve L1/L2 cache hit rates and applies computational graph fusion to eliminate redundant data loading across layers. These combined optimizations notably improve inference efficiency. + +4) State-of-the-art (SOTA) Accuracy and Efficiency Combined. On the KITTI dataset, EMC2 achieves $3.28\%$ higher accuracy for pedestrians and $6.03\%$ for bicycles under hard-level detection challenge compared to 15 baseline models. On the nuScenes dataset, it yields a $3.9\%$ increase in mAP and a $1.8\%$ boost in NDS. In addition, EMC2 achieves a $159.06\%$ reduction in inference latency on the Jetson platform for KITTI, supporting its suitability for real-time edge deployment. + +# 2. Related Work + +LiDAR-only Method. The active 3D sensing of LiDAR makes it dominant in ADS 3D detection [18-20]. LiDAR-based methods process point clouds via voxelization [21], pillarization [20], or direct raw input [22]. Anchor-based heads, used for region proposal [21], and + +center-based representations[23] are widely adopted. These methods achieve high accuracy on KITTI, as demonstrated by 3DSSD [24], PartA2 [19], and Voxel R-CNN [3]. However, weak LiDAR reflections for distant or occluded objects degrade detection performance [25, 26]. + +Multimodal Fusion Method. Combining LiDAR and camera RGB input enhances detection performance, especially for sparse LiDAR objects [27]. Fusion approaches include projection-based methods, which align image features with point clouds in 3D space [28], and model-based techniques that incorporate cross-attention [29] or unified feature spaces [30]. Recent multimodal models, such as PPF-Det [31] and RoboFusion [32], achieve top accuracy on KITTI benchmarks. However, these models often trade off accuracy for speed or vice versa. + +Edge-based Compilation. ML compilation platforms are essential for improving the inference efficiency of deep learning models across diverse hardware architectures [33, 34]. TensorRT [35] is an NVIDIA inference optimizer that accelerates ML models on GPUs through layer fusion, precision calibration, and kernel auto-tuning. XLA [36] is a domain-specific compiler for TensorFlow enhancing efficiency by fusing operations and optimizing computation graphs. Open Neural Network Exchange (ONNX) Runtime [37] provides an open standard for model interoperability, enabling efficient execution across CPUs, GPUs, and accelerators. TVM [11] is an open-source deep learning compiler that automates model tuning and tensor scheduling for efficient deployment across CPUs and GPUs. + +# 3. EMC2 Computing Methodology + +We introduce EMC2, a multimodal MoE framework that minimizes latency and ensures reliable accuracy. Fig. 2 shows the EMC2 overview, EMC2 consists of: 1) Adaptive Multimodal Data Bridge (AMDB) to preprocess multimodal input (Sec. 3.1); 2) Scenario-Adaptive Dispatcher to adaptively dispatch a suitable expert for each scenario (Sec. 3.2); and 3) Three Scenario-Optimized Experts to decode AMDB-output features into final results (Sec. 3.3). We also design an effective training strategy to address the long-tail effect in the training process (Sec. 3.4). In addition, we deploy EMC2 on edge devices with self-designed 3D sparse convolution and Multiscale Pooling. + +# 3.1. Adaptive Multimodal Data Bridge + +Given the multimodal nature of EMC2's sensor inputs, a module providing suitable preprocessed data for different experts is required. Thus, we propose AMDB, a preprocessing unit to efficiently process the required input data for each expert. The workflow of AMDB begins with a UNet [38] and a sparse CNN to extract LiDAR features, followed by fully connected layers that generate proposal regions along with their confidence scores. These results + +serve as inputs to Latency-Prioritized Expert and Versatile Efficiency Expert. The extraction of image features depends on the Scenario-Adaptive Dispatcher. If further information is required for reliable detection, AMDB will extract image features by ResNet. Based on depth information, it projects relevant image pixels into 3D space, applying Multiscale Pooling to reduce computational overhead, and fuses them with LiDAR voxel features to produce a multimodal output. These results are inputs for Accuracy-Prioritized Experts. The Multiscale Pooling and multimodal fusion processes are demonstrated in Fig. 5. + +# 3.2. Scenario-Adaptive Dispatcher + +The Scenario-Adaptive Dispatcher (SAD) dynamically dispatches each scenario to the most suitable expert based on traffic conditions and latency-accuracy requirements. Our expert selection algorithm considers real-world traffic safety needs: specifically, how far away objects are and how clearly they can be perceived. + +So, the switch among experts is determined by two key parameters: object distance $\mathcal{D}$ and clarity, which is inferred from the confidence $\mathcal{C}$ of proposal regions provided by the AMDB. We classify traffic scenarios into three categories: 1) Close and Distinct Cases. All objects are clearly visible and no distant objects are present. This is characterized by all proposal regions $< \mathcal{D}$ and having confidence $\geq \mathcal{C}$ . In such scenarios, as near-field objects provide abundant Li-DAR information, even when compressed into bird's eye view (BEV), a 2D CNN can effectively capture key patterns such as object surfaces and heights [20], ensuring reliable detection accuracy with minimal latency (Sec. 4.4). These cases will be routed to the Latency-Prioritized Expert. 2) Mixed Visibility Cases. Some objects are either distant but clearly visible or near but unclear. Mathematically, this corresponds to cases where i) some proposal regions $< \mathcal{D}$ with confidence $< \mathcal{C}$ , or ii) the distance of some objects $\geq \mathcal{D}$ with confidence $\geq \mathcal{C}$ . Here, objects have insufficient voxel data, requiring more complex 3D convolution to extract meaningful patterns, and they will be dispatched to the Versatile Efficiency Expert. 3) Distant and Uncertain Cases. Some objects are both distant and unclear. This occurs when some proposal regions $\geq \mathcal{D}$ with confidence $< \mathcal{C}$ . In these challenging scenarios, significant voxel information is missing due to long distances, steep reflection angles, or occlusions, necessitating the integration of image data to compensate for missing LiDAR details. These cases will be processed by the Accuracy-Prioritized Expert. + +# 3.3. Scenario-Optimized Experts + +Latency-Prioritized Expert (LPE). LPE uses 2D CNNs for object detection based on BEV projections. It first projects proposal regions from the AMDB into BEV space, then extracts 2D features for object localization and + +![](images/f13f7eecda8ca25f2db6748d167f0ac62896c17fa918088cec8473745c50189c.jpg) +Figure 2. System Structure of EMC2. The central row illustrates the architecture of EMC2, while the top and bottom rows provide visual examples of core modules and relevant legends. The system integrates five main components: an Adaptive Multimodal Data Bridge (AMDB) for preprocessing raw multimodal input and generating expert-specific features; a Scenario-Adaptive Dispatcher that dispatches suitable experts based on scenario-specific features (see Fig. 3); the Latency-Prioritized Expert designed for simple scenarios requiring minimal inference delay; the Versatile Efficiency Expert designed for uncertain or distant object detection scenarios; and the Accuracy-Prioritized Expert responsible for handling complex, high-precision detection scenarios. Additionally, an Emergency Expert API is designed to support rapid response to hazardous or unseen situations. The overall training process incorporates loss terms for both final detection results and intermediate outputs from the AMDB, as described in Sec. 3.4. + +![](images/8cb87474ddee652c12453ab79550d6e0ea89ea00809e0beb2ac65cf22d4d29ee.jpg) +Figure 3. Routing Strategy of the Scenario-Adaptive Dispatcher. The Routing criteria is detailed mathematically in Sec. 3.2. + +classification. It finally maps detection results back to 3D space. Compared to sparse 3D convolutions, its use of 2D operations reduces computation in outdoor ADS scenarios [20], allowing LPE to improve inference efficiency with minimal accuracy loss. + +Versatile Efficiency Expert (VEE). VEE uses a 3D CNN to process proposal regions and confidence scores from the AMDB. After sparse convolution, voxel features are decoded into per-voxel classifications and bounding + +boxes. Compared to LPE, 3D convolutions better preserve spatial structure, making it more robust for scenarios with insufficient LiDAR signal. + +Accuracy-Prioritized Expert (APE). APE extends VEE by incorporating the multimodal signal. It employs a 3D CNN and decoder with the same structure but different parameters to process multimodal proposal regions (from AMDB) and fuses image data to recover missing LiDAR information. Image features compressed via Multiscale Pooling (Sec.3.5) are projected into 3D space and fused with corresponding voxels of LiDAR features, ensuring more robust understanding for occluded and distant objects while keeping computation feasible. The Multiscale Pooling and multimodal fusion are detailed in Fig. 5. + +# 3.4. Collaborative Training of Multimodal Experts + +Hierarchical Training Strategy: Multimodal Supervision with Triple Back-Propagation. During training, the performance of experts is highly dependent on the proposal regions produced by the AMDB, yet these proposals tend to be unreliable in the early training phase, leading to an unstable expert learning process and unstable SAD decisions. To address this, we adopt a hierarchical supervision stra + +![](images/bf1cea59c5f10b342486ad5d82aa8d91aa2198df29496c17f3124dc349209ef1.jpg) +Figure 4. Illustration of Hierarchical Training. The three backpropagation paths originate from: 1) the final detection output, 2) the LiDAR branch output of AMDB, and 3) the image branch output of AMDB. Since expert training relies on initially unreliable AMDB outputs, direct supervision accelerates convergence and enhances representation learning. + +egy with three distinct back-propagation pathways, as illustrated in Fig. 4: 1) supervision of the LiDAR branch outputs from the AMDB; 2) supervision of the image branch outputs from the AMDB; and 3) supervision of expert predictions and associated confidence scores. To further improve training stability, the LiDAR and image branches of the AMDB are pre-trained separately before joint optimization. Each back-propagation route is supervised by the following loss function, which collectively forms Multimodal Loss: + +$$ +\mathcal {L} _ {\text {r o u t e}} = \mathcal {L} _ {\text {c l s}} + \mathcal {L} _ {\text {r e g}} \tag {1} +$$ + +where $\mathcal{L}_{\mathrm{cls}}$ uses cross-entropy loss and $\mathcal{L}_{\mathrm{reg}}$ uses smooth- $\ell 1$ loss. The Hierarchical Training not only stabilizes expert learning but also enhances the model's adaptability to various data granularities. In turn, even under compressed data representations, it enables the effective use of our Multiscale Pooling for memory efficiency, as detailed in Sec. 3.5. + +Addressing Long-tail Effects in MoE Training. The MoE framework often suffers from long-tail effects, where certain experts receive insufficient training data, leading to imbalances in sub-datasets, and therefore affecting performance [39]. To address this, we propose a threefold training strategy: 1) Data subset division. We partition the dataset into sub-datasets based on target and auxiliary samples. Each expert $\mathcal{E}_i$ is dispatched a sub-dataset $S_{i}$ $(i = 1,2,3)$ containing target samples $\mathcal{T}_{\mathcal{S}_i}$ , which are expected to be routed to the corresponding expert, and auxiliary samples $\mathcal{A}_{\mathcal{S}_i}$ , which are not. We ensure that $\mathcal{T}_{\mathcal{S}_i} \cap \mathcal{T}_{\mathcal{S}_j} = \emptyset$ for any + +![](images/370e7b99ef66e8ccacbdbb6143a6bc15182d6a6bf5f37353615c9a6ae204e431.jpg) +Figure 5. Illustration of Multiscale Pooling and Multimodal Fusion. Proposal regions with image features are first obtained from the 3D proposal space generated by the LiDAR branch of AMDB. After extracting the relevant image features from the corresponding image regions and projecting them into 3D space, pooling and fusion are then applied. Following the pooling step, image features of pixels that share the same 3D coordinate as a voxel are concatenated with the corresponding LiDAR voxel features, while those without matching voxel coordinates are concatenated with zero-filled features. + +$i \neq j$ . The inclusion of auxiliary samples, which are not intended for a given expert, helps prevent overfitting and enhances generalization. 2) Balanced sampling strategy. We adjust the selection probability $\mathcal{P}_i$ for the sub-datasets which have fewer samples to ensure balanced training: + +$$ +\mathcal {P} _ {i} \times \mathcal {N} _ {\mathcal {S} _ {i}} = \mathcal {P} _ {j} \times \mathcal {N} _ {\mathcal {S} _ {j}}, \sum_ {i = 1} ^ {3} \mathcal {P} _ {i} = 1 \tag {2} +$$ + +where $\mathcal{N}_{S_i}$ represents the number of samples in sub-datasets $S_{i}$ . This ensures that experts dispatched to less frequent scenarios receive sufficient training iterations. 3) Adaptive optimizer. We introduce an optimizer that adapts the learning rate based on the proportion of target samples in a batch. Denoting $\alpha_0$ as the base learning rate, the adaptive learning rate applied to an expert with a given batch is computed as: + +$$ +\alpha = \left(1 + \frac {1}{\mathcal {N}} \sum_ {i = 1} ^ {\mathcal {N}} \mathbf {1} _ {i \in \mathcal {T}}\right) \cdot \alpha_ {0} \tag {3} +$$ + +where $\mathcal{N}$ is the batch size, $p$ is the proportion of target samples within current mini-batch, $\mathcal{T}$ represents target samples in the batch. By regulating the gradient update rate based on the batch composition, this adaptive optimizer encourages each expert to prioritize its designated target samples while preventing excessive updates from auxiliary samples. + +# 3.5. Algorithm Edge-adaptivity Optimization + +Edge devices face challenges such as incomplete function libraries and limited cache. We adopt the open-source ONNX Runtime for compiling EMC2 on edge, which provides modules that help address these challenges and facilitate edge-oriented optimization. As 3D convolutions account for the majority of computation during inference, we + +customize 3D sparse convolution and Multiscale Pooling to improve the algorithmic efficiency of EMC2. + +3D Sparse Convolution. Since the 3D sparse convolution library is incomplete on most edge platforms, we develop a customized 3D sparse convolution library with parallel threads execution under ONNX Runtime. Let $\mathcal{H}$ denote spatial resolution, $\mathcal{N}$ the number of non-empty voxels, and $\mathcal{C}_{\mathrm{v}}$ the per-voxel computation, including channels and kernel size. By computing nonzero voxels only, 3D sparse convolution reduces complexity from $O(\mathcal{H}^3\times \mathcal{C}_{\mathrm{v}})$ for dense convolutions to $O(\mathcal{N}\times \mathcal{C}_{\mathrm{v}})$ , where $\mathcal{N}\ll \mathcal{H}^3$ , cutting redundant operations by $65 - 80\%$ [40]. + +Multiscale Pooling. To alleviate cache limitations, Multiscale 3D Sparse Pooling adaptively pools image features based on available memory before fusing them with LiDAR features (Fig.5). The pooling size is user-configurable, allowing flexibility for maximizing memory utilization. When combined with Hierarchical Training (Sec.3.4), Multiscale Pooling effectively reduces resource consumption while maintaining model reliability (Sec. 4.4). + +# 3.6. EMC2 Computing System Optimization + +Building upon the algorithm-level optimizations introduced in Sec. 3.5, we further enhance the ONNX execution efficiency through system-level improvements, specifically targeting memory management and computation graph optimization. + +Memory Optimization. To further improve memory efficiency during inference, we implement three system-level techniques. 1) Overlapping communication and computation [41]. Matrix data is partitioned into segments processed and transferred in parallel, overlapping computation and communication. This reduces global memory access, improves cache usage, and mitigates memory bottlenecks. 2) Thread scheduling. We implement a staged thread management mechanism, where active threads are periodically terminated and re-initialized (Fig. 6). This approach limits unnecessary memory retention between stages, reducing system memory footprint. 3) Prefix-sum for sparse convolution. Parallel prefix-sum reduces index search time from $\mathcal{O}(\mathcal{N})$ to $\mathcal{O}(\log \mathcal{N})$ using GPU parallelism, where $\mathcal{N}$ is the number of nonzero elements. + +Computational Graph Optimization. ONNX partitions the computation graph into segments and assigns each to the most suitable hardware. Our optimization includes three stages: 1) Model pruning. Redundant parameters and operations are removed to simplify the graph; 2) Model quantization. Low-precision arithmetic replaces floating-point computation to improve hardware efficiency; 3) Computational graph fusion. Consecutive operators are merged to reduce memory access and improve execution, benefiting memory-bound CPUs and reducing kernel + +launch overhead on GPUs. + +# 4. Experiment and Result + +# 4.1. Experiment Settings and Datasets + +Experiment Platform. Experiments are conducted on the Jetson AGX Orin, an edge platform with limited computational resources, an NVIDIA Ampere GPU, a 4MB shared L2 cache, and hierarchical memory. For comparison, EMC2 is also evaluated on an NVIDIA A4000 GPU. + +Dataset and Sample Difficulties. We evaluate 3D object detection on two public benchmarks: KITTI [42] and nuScenes [43]. KITTI provides 3,712 training and 3,769 validation samples, with object difficulty levels categorized as $30\%$ easy, $40\%$ moderate, and $30\%$ hard. nuScenes contains 850 training and 150 test scenes, covering 10 object categories with 3D bounding box annotations. Before splitting the data, we analyze the confidence distribution of proposal regions generated by the pre-trained AMDB. A two-sample Kolmogorov-Smirnov test reveals a statistically significant difference between objects within and beyond a distance threshold $D$ , which is empirically set to 23.5 meters for KITTI and 35 meters for nuScenes. The training subsets for each expert on the KITTI and nuScenes datasets are defined as follows: $\mathcal{T}_{S_1}$ includes scenes where all objects are within $D$ meters and labeled as "easy" or "moderate"; $\mathcal{T}_{S_2}$ includes scenes with at least one object beyond $D$ meters labeled as "easy" or "moderate," or any object within $D$ meters labeled as "hard"; and $\mathcal{T}_{S_3}$ includes scenes where at least one object is both beyond $D$ meters and labeled as "hard," requiring multimodal processing. For each expert, the auxiliary subset $\mathcal{A}_{S_i}$ is uniformly sampled from the other experts' target subsets, with its size equal to that of the corresponding target subset. + +Evaluation Metrics. KITTI uses R40 3D Average Precision (AP) with IoU thresholds of 0.7 for cars and 0.5 for pedestrians and cyclists. nuScenes adopts the NDS metric, combining mAP—computed via center distance matching $(0.5\mathrm{m}$ to $4.0\mathrm{m}$ thresholds)—and five error terms: mean Average Translation (mATE), Scale (mASE), Orientation (mAOE), Velocity (mAVE), and Attribute (mAAE) Errors. Inference latency on Jetson AGX Orin is also reported to assess real-time performance. + +Hierarchical Training Details. Allowing all experts to propagate gradients to AMDB causes instability due to overlapping training data, leading to repeated updates. To address this, AMDB is first pre-trained for 20 epochs, followed by joint training where gradients are propagated only from APE, which receives full multimodal input. + +# 4.2. Comparison with Existing SOTA Methods + +As shown in Tab. 1 and Tab. 2, EMC2 achieves SOTA accuracy and latency on KITTI and nuScenes. On the Jetson + +![](images/d62ca3cdfe936cfa47d6ab4d81ab4d294870be561966d0dd6faf48048b96e60b.jpg) +Figure 6. Illustration of Thread Management. Each solid box represents a module occupying threads. During inference, four boundaries are defined at which preceding threads are terminated, and their thread-specific address space is released upon crossing each boundary. + +![](images/c51465c9b24cc8b4e0c05547de5828eb66ab9ef9357ae08a3e9df1f8c130a1cb.jpg) +Table 1. We evaluate different 3D object detection methods on the KITTI dataset. Some cells are denoted as $N / A$ when the corresponding open-source implementation lacks support for the specific object class or is not compatible with Jetson devices. + +
MethodSizemAP↑NDS↑Latency (ms)↓
FocalFormer [55]189M0.66400.7090N/A
BEVFusion [56]156M0.68520.7138N/A
EMC2 (ours)87M0.72410.7316229.3
+ +Table 2. Comparison of different methods on the nuScenes dataset. + +AGX Orin platform, it also outperforms existing baseline models across all difficulty levels. + +Results on KITTI. Pedestrians and cyclists face higher traffic risks [54], which remain challenging for most SOTA methods. EMC2 improves hard-set detection by $5\%$ and $7\%$ for pedestrians and cyclists, respectively, outperforming TED and LoGoNet (Tab. 1). For cars, it achieves $88.15\%$ on the hard set, exceeding prior multimodal methods by up to $9\%$ . It runs at $372.5 \mathrm{~ms}$ on Jetson, $2.5 \times -2.5 \times$ faster than methods requiring $965 - 1$ , $787 \mathrm{~ms}$ . + +Results on nuScenes. EMC2 improves mAP and NDS by $3.9\%$ and $1.8\%$ , respectively, outperforming recent + +methods FocalFormer [55] and BEVFusion [56] (Tab. 2). It runs at $229.3\mathrm{ms}$ on Jetson AGX Orin, enabling real-time use on edge devices. + +# 4.3. Analysis and Discussion + +Benefits of Mixture-of-Experts. The balance between accuracy and efficiency observed in Sec. 4.2 is achieved through dynamic expert selection based on scenario characteristics, combined with algorithmic and system-level optimizations designed for edge devices, such as Jetson. + +Empirical Dispatcher. The EMC2 dispatcher uses fixed distance and confidence thresholds to dispatch scenes to experts. The protocol is adapted per dataset format. This strategy is validated via expert activations on the nuScenes validation set (Fig. 7): LPE performs best in close-range, high-confidence scenes; APE in distant, low-confidence cases; and VEE in all others, supporting the effectiveness of empirical routing. + +Optimized Parameter Size. EMC2 has 226M param- + +![](images/0783037485353f1a17162a96132bcc397f417a468d0c2ef34a3fefe2db0f6c3c.jpg) +Figure 7. Distribution of expert activation. During EMC2 inference on the nuScenes validation set, scenes are grouped into a 2D grid based on two metrics: the y-axis indicates the number of AMDB-generated proposals beyond the distance threshold $D$ (i.e., distant objects), and the x-axis represents the average confidence score of these proposals (higher values indicate clearer scenes). Each cell is labeled with the name of the expert that performs best for scenes falling into that bin. + +![](images/a8ad0f87d16bfc084782903972c7c98a5c73af69fe8048b9bcf032eadfe65aa1.jpg) +Figure 8. Median Detection Confidence vs. Object Distances. Four configurations are compared on the KITTI validation set: LPE (blue), VEE (peach), APE (rose), and EMC2 (green). More complex experts yield higher confidence but slower inference. By switching experts based on distance and clarity, EMC2 balances accuracy and efficiency. + +eters, but its MoE framework activates only a subset during inference, averaging 87M on nuScenes and 146M on KITTI, which is fewer than those used by the full model and other SOTA baselines. System-level optimizations instead prioritize computational throughput and memory efficiency over parameter reduction. + +Hardware Flexibility. The EMC2 is implemented in the ONNX format, an open format supported by various compilers and enables model compilation across diverse hardware platforms. As a result, EMC2 is not limited to Jetson and can be deployed on other edge platforms. + +# 4.4. Ablation Study + +Ablation on the Roles of Three Experts. We evaluate each expert on its target scenarios using the KITTI validation set. As shown in Tab. 3, LPE, VEE, and APE achieve comparable accuracy in LPE scenes, but LPE has the lowest inference latency on Jetson. In VEE scenarios, incor + +
MethodAP↑Latency (ms) ↓
Ped.CarCyclist
Ablation on LPE
LPE66.3289.0488.54219
VEE74.4992.2391.69468
APE74.3493.2192.41526
Ablation on VEE
W/ VEE70.5590.5084.07372.5
W/O VEE68.2686.4079.17512.7
Ablation on APE
VEE67.2487.6981.85468
APE70.5590.5084.18526
Ablation on Hierarchical Training and Multiscale Pooling
HT √, MP √70.5590.5084.07526
HT √, MP ×70.5590.5084.181320
HT ×, MP √63.1188.6875.98529
HT ×, MP ×66.3489.1080.831367
+ +Table 3. Results Across Three Ablation Configurations. The main body of this table is divided into four sections corresponding to each part of the ablation study. In the third section, HT and MP are abbreviated for Hierarchical Training and Multiscale Pooling, respectively. + +porating VEE improves both accuracy and latency over the baseline. For APE, which incorporates image features, accuracy improves further with only a minor latency increase, supported by the score distribution in Fig. 8. + +Ablation on Hierarchical Training (HT) and Multiscale Pooling (MP). We evaluate four combinations of HT and MP on the full KITTI validation set: both applied, HT only, MP only, and neither. Tab. 3 shows that combining HT and MP reduces inference time on Jetson from 1320 ms to 526 ms without accuracy loss. For the car class, MP alone reduces accuracy from $80.83\%$ to $75.98\%$ , while HT alone improves it from $66.34\%$ to $70.55\%$ . HT facilitates multigranular feature learning, while MP enhances memory efficiency via compression. + +# 5. Conclusion + +We present EMC2, a multimodal 3D object detection solution for ADS based on a customized MoE architecture. This design enables adaptive expert selection across diverse traffic scenarios, balancing detection accuracy and efficiency. Experiments on KITTI and nuScenes show strong performance, particularly under challenging conditions. On the Jetson platform, EMC2 achieves a $159.06\%$ speedup in inference, supporting real-time edge deployment. Future work will explore fully adaptive expert selection to enhance system responsiveness and resource efficiency. + +# References + +[1] Shih-Chieh Lin, Yunqi Zhang, Chang-Hong Hsu, Matt Skach, Md E. Haque, Lingjia Tang, and Jason Mars. The architectural implications of autonomous driving: Constraints and acceleration. In Proceedings of the Twenty-Third International Conference on Architectural Support for Programming Languages and Operating Systems, ASPLOS '18, page 751-766, New York, NY, USA, 2018. Association for Computing Machinery. 1 +[2] Yi Xiao, Felipe Codevilla, Akhil Gurram, Onay Urfalioglu, and Antonio M López. Multimodal end-to-end autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(1):537-547, 2020. 2 +[3] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. arXiv:2012.15712, 2020. 2, 3, 7 +[4] Hao Frank Yang, Yifan Ling, Cole Kopca, Sam Ricord, and Yinhai Wang. Cooperative traffic signal assistance system for non-motorized users and disabilities empowered by computer vision and edge artificial intelligence. Transportation Research Part C: Emerging Technologies, 145:103896, 2022. Elsevier. +[5] Hao Yang, Jiarui Cai, Meixin Zhu, Chenxi Liu, and Yinhai Wang. Traffic-informed multi-camera sensing (TIMS) system based on vehicle re-identification. IEEE Transactions on Intelligent Transportation Systems, 23(10):17189-17200, 2022. IEEE. 2 +[6] Guibiao Liao, Jiankun Li, and Xiaqing Ye. Vlm2scene: Self-supervised image-text-lidar learning with foundation models for autonomous driving scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 3351-3359, 2024. 2 +[7] Hao Frank Yang, Yang Zhao, Jiarui Cai, Meixin Zhu, JenqNeng Hwang, and Yiran Chen. Mitigating bias of deep neural networks for trustworthy traffic perception in autonomous systems. In Proceedings of the 2024 IEEE Intelligent Vehicles Symposium (IV), pages 633-638, 2024. IEEE. +[8] Chenxi Liu, Hao Yang, Ruimin Ke, Wei Sun, Julia Wang, and Yinhai Wang. Cooperative and comprehensive multitask surveillance sensing and interaction system empowered by edge artificial intelligence. Transportation Research Record, 2677(9):652-668, 2023. SAGE Publications. +[9] Hao Frank Yang, Jiarui Cai, Chenxi Liu, Ruimin Ke, and Yinhai Wang. Cooperative multi-camera vehicle tracking and traffic surveillance with edge artificial intelligence and representation learning. Transportation Research Part C: Emerging Technologies, 148:103982, 2023. Elsevier. 2 +[10] Chen Zhang, Peng Li, Guangyu Sun, Yijin Guan, Bingjun Xiao, and Jason Cong. Optimizing fpga-based accelerator design for deep convolutional neural networks. In Proceedings of the 2015 ACM/SIGDA international symposium on field-programmable gate arrays, pages 161-170, 2015. 2 +[11] Tianqi Chen, Thierry Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Yan, Haichen Shen, Meghan Cowan, Leyuan Wang, Yuwei Hu, Luis Ceze, et al. {TVM}: An automated {End-to-End} optimizing compiler for deep learning. In 13th + +USENIX Symposium on Operating Systems Design and Implementation (OSDI 18), pages 578-594, 2018. 2, 3 +[12] Jiarui Cai, Yizhou Wang, and Jenq-Neng Hwang. Ace: Ally complementary experts for solving long-tailed recognition in one-shot. In Proceedings of the IEEE/CVF international conference on computer vision, pages 112-121, 2021. 2 +[13] Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. Unimoe: Scaling unified multimodal llms with mixture of experts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2 +[14] Jiachen Li, Xinyao Wang, Sijie Zhu, Chia-Wen Kuo, Lu Xu, Fan Chen, Jitesh Jain, Humphrey Shi, and Longyin Wen. Cumo: Scaling multimodal llm with co-upcycled mixture-of-experts. Advances in Neural Information Processing Systems, 37:131224-131246, 2024. 2 +[15] Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts. arXiv preprint arXiv:2407.06204, 2024. 2 +[16] Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. Time-moe: Billion-scale time series foundation models with mixture of experts. arXiv preprint arXiv:2409.16040, 2024. 2 +[17] Jiaao He, Jiezhong Qiu, Aohan Zeng, Zhilin Yang, Jidong Zhai, and Jie Tang. Fastmoe: A fast mixture-of-expert training system. arXiv preprint arXiv:2103.13262, 2021. 2 +[18] Charles R. Qi, Xinlei Chen, Or Litany, and Leonidas J. Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4403-4412, 2020. 2 +[19] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. arXiv preprint arXiv:1907.03670, 2019. 3 +[20] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689-12697, 2019. 2, 3, 4, 7 +[21] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2018. 2 +[22] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10526–10535, 2020. 2 +[23] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11779-11788, 2021. 3 +[24] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11037-11045, 2020. 3 + +[25] Andrew M Wallace, Abderrahim Halimi, and Gerald S Buller. Full waveform lidar for adverse weather conditions. IEEE transactions on vehicular technology, 69(7):7064-7077, 2020. 3 +[26] You Li, Pierre Duthon, Michele Colomb, and Javier Ibanez-Guzman. What happens for a tof lidar in fog? IEEE Transactions on Intelligent Transportation Systems, 22(11):6670-6681, 2020. 3 +[27] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In 2019 international conference on robotics and automation (ICRA), pages 3288-3295. IEEE, 2019. 3 +[28] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. E-net: Enhancing point features with image semantics for 3d object detection. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 35–52, Cham, 2020. Springer International Publishing. 3 +[29] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quaan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V. Le, Alan Yuille, and Mingxing Tan. Deep-fusion: Lidar-camera deep fusion for multi-modal 3d object detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17161-17170, 2022. 3 +[30] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022. 3 +[31] Guotao Xie, Zhiyuan Chen, Ming Gao, Manjiang Hu, and Xiaohui Qin. Ppf-det: Point-pixel fusion for multi-modal 3d object detection. IEEE Transactions on Intelligent Transportation Systems, 25(6):5598-5611, 2024. 3 +[32] Ziying Song, Guoxing Zhang, Lin Liu, Lei Yang, Shaoqing Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Robofusion: Towards robust multi-modal 3d object detection via sam. arXiv preprint arXiv:2401.03907, 2024. 3 +[33] Zheng Wang and Michael O'Boyle. Machine learning in compiler optimization. Proceedings of the IEEE, 106(11):1879-1901, 2018. 3 +[34] Hugh Leather, Edwin Bonilla, and Michael O'boyle. Automatic feature generation for machine learning-based optimising compilation. ACM Transactions on Architecture and Code Optimization (TACO), 11(1):1-32, 2014. 3 +[35] NVIDIA Corporation. NVIDIA TensorRT Developer Guide, 2024. Accessed: 2025-02-13. 3 +[36] Amit Sabne. XLA: Compiling Machine Learning for Peak Performance, 2020. 3 +[37] ONNX Community. Onnx: Open neural network exchange. https://github.com/onnx/onnx, 2024. Accessed: YYYYY-MM-DD. 3 +[38] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3 + +[39] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. *Neural computation*, 3(1):79–87, 1991. 5 +[40] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10), 2018. 6, 7 +[41] Juraj Hromkovič. Communication complexity and parallel computing. Springer Science & Business Media, 2013. 6 +[42] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 6 +[43] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6 +[44] Muhammad Aslam. Introducing kolmogorov-smirnov tests under uncertainty: an application to radioactive data. ACS omega, 5(1):914-917, 2019. +[45] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 772-782, 2022. 7 +[46] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 7 +[47] Jiajia Liao, Yujun Liu, Yingchao Piao, Jinhe Su, Guorong Cai, and Yundong Wu. Gle-net: A global and local ensemble network for aerial object detection. International Journal of Computational Intelligence Systems, 15(1):2, 2022. 7 +[48] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence, 43(8):2647-2664, 2020. 7 +[49] Hiep Anh Hoang and MyungsiK Yoo. 3onet: 3d detector for occluded object under obstructed conditions. IEEE Sensors Journal, 2023. 7 +[50] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023. 7 +[51] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7 +[52] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on + +Computer Vision and Pattern Recognition, pages 908-917, 2022. 7 +[53] Hai Wu, Chenglu Wen, Wei Li, Xin Li, Ruigang Yang, and Cheng Wang. Transformation-equivariant 3d object detection for autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2795-2802, 2023. 7 +[54] Nils Lubbe, Yi Wu, and Hanna Jeppsson. Safe speeds: fatality and injury risks of pedestrians, cyclists, motorcyclists, and car drivers impacting the front of another passenger car as a function of closing speed and age. Traffic safety research, 2:000006-000006, 2022. 7 +[55] Yilun Chen, Zhiding Yu, Yukang Chen, Shiyi Lan, Anima Anandkumar, Jiaya Jia, and Jose M Alvarez. Focalformer3d: focusing on hard instance for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8394-8405, 2023. 7 +[56] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781, 2023. 7 \ No newline at end of file diff --git a/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/images.zip b/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..09a8125ff8a91f19813b772afe85b999a7d26c78 --- /dev/null +++ b/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:68cf3e244f97ea90080bb3a7c1eeb8c2165b307e7539146169d97d6d02300e1f +size 572320 diff --git a/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/layout.json b/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2d84b261edb72bd236cb31c7bcd68c0b3cb4bdc4 --- /dev/null +++ b/2025/Towards Accurate and Efficient 3D Object Detection for Autonomous Driving_ A Mixture of Experts Computing System on Edge/layout.json @@ -0,0 +1,7786 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 547, + 140 + ], + "type": "text", + "content": "Towards Accurate and Efficient 3D Object Detection for Autonomous Driving: A Mixture of Experts Computing System on Edge" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "spans": [ + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "text", + "content": "Linshen Liu" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "text", + "content": " Boyan Su" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "text", + "content": " Junyue Jiang" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "text", + "content": " Guanlin Wu" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "text", + "content": " Cong Guo" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "text", + "content": " Ceyu Xu" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "text", + "content": " Hao Frank Yang" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "text", + "content": "Johns Hopkins University " + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "text", + "content": "Duke University " + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 162, + 160, + 451, + 219 + ], + "type": "text", + "content": "HKUST {11iu148, bsul1}@jh.edu, haofrankyang@jhu.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 152, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 297, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 558 + ], + "type": "text", + "content": "This paper presents Edge-based Mixture of Experts (MoE) Collaborative Computing " + }, + { + "bbox": [ + 54, + 270, + 297, + 558 + ], + "type": "inline_equation", + "content": "(EMC2^{\\ddagger})" + }, + { + "bbox": [ + 54, + 270, + 297, + 558 + ], + "type": "text", + "content": ", an optimal computing system designed for autonomous vehicles (AVs) that simultaneously achieves low-latency and high-accuracy 3D object detection. Unlike existing works, the EMC2 introduces a novel scenario-aware MoE architecture optimized for fusing complementary sparse 3D point clouds and dense 2D images to achieve robust multimodal representations for detection. Furthermore, EMC2 integrates an adaptive multimodal data bridge with multi-scale region proposing and scenario-aware routing, dynamically dispatching features to complementary experts based on object visibility and distance. In addition, EMC2 integrates joint hardware-software optimizations, including hardware resource utilization optimization and computational graph simplification, to ensure efficient and real-time inference on resource-constrained edge devices. Experiments on open-source benchmarks clearly show the EMC2 advancements as an end-to-end system. On the KITTI dataset, it achieves an average accuracy improvement of " + }, + { + "bbox": [ + 54, + 270, + 297, + 558 + ], + "type": "inline_equation", + "content": "3.58\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 558 + ], + "type": "text", + "content": " and a " + }, + { + "bbox": [ + 54, + 270, + 297, + 558 + ], + "type": "inline_equation", + "content": "159.06\\%" + }, + { + "bbox": [ + 54, + 270, + 297, + 558 + ], + "type": "text", + "content": " inference speedup compared to 15 baseline methods on Jetson platforms, with similar performance gains on the nuScenes dataset, highlighting its capability to advance reliable, real-time 3D object detection tasks for AVs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 569, + 135, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 135, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 135, + 582 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": "Traffic safety is a fundamental concern for both human drivers and Autonomous Driving Systems (ADS). Within an ADS, the perception module plays a pivotal role by serving as the \"eyes\" of the autonomous vehicle (AV), sensing the surrounding environment and delivering critical information to downstream modules such as motion planning, control, and safety guarantees. Existing conceptual safety of an AV perception system is generally defined by two core" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 332, + 232, + 538, + 463 + ], + "blocks": [ + { + "bbox": [ + 332, + 232, + 538, + 463 + ], + "lines": [ + { + "bbox": [ + 332, + 232, + 538, + 463 + ], + "spans": [ + { + "bbox": [ + 332, + 232, + 538, + 463 + ], + "type": "image", + "image_path": "7037e73d354f54394db0640ae8f77d2d26d60527926c16fe7628081fb33e055b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 472, + 555, + 529 + ], + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 529 + ], + "type": "text", + "content": "Figure 1. Comparison of 3D object detection models in accuracy and inference efficiency on the KITTI and nuScenes dataset. Most existing methods improve accuracy at the expense of inference efficiency. The proposed EMC2 achieves win-win performance by customizing MoE for autonomous driving." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "dimensions: accuracy and efficiency. Accuracy ensures reliable object detection and tracking capabilities, enabling a credible understanding of the environment. Efficiency, characterized by low perception delay, is equally essential, as real-time high-level planning and control rely on fast information processing. Perception delay, defined as the sum of data acquisition, computation, and algorithm processing time, directly affects an ADS's responsiveness to dynamic environments. Research indicates that an ADS must process sensory inputs and issue feedback control commands within 100 milliseconds to guarantee prompt and safe vehicle operations [1]. Detection precision and perception delay together define a critical safety boundary: a precise but sluggish system cannot react in time, while a fast but inaccurate system may lead to unsafe decisions." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 295, + 713 + ], + "type": "text", + "content": "*Equal contribution, †Corresponding author. \n‡EMC2 Code: https://github.com/LinshenLiu622/EMC2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25903" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 383 + ], + "type": "text", + "content": "As illustrated in Fig. 1, achieving both high accuracy and low latency remains a fundamental challenge. Existing approaches often face inherent trade-offs between these objectives due to limited computational resources, insufficient adaptability of perception systems, and the inherent complexity of dynamic driving scenarios. Urban traffic environments require low-latency and accurate sensing capabilities to handle complex, densely interactive surroundings, while highway scenarios demand reliable detection of distant objects and prompt responses to unexpected events. Moreover, even for the same object, recognition priorities may vary dynamically within a scene depending on distance and scene complexity, posing notable challenges for single-modality models in balancing efficiency and accuracy [2]. Multimodal models, such as large-scale CNN and Transformer-based architectures that integrate LiDAR, radar, and camera data [3-5], leverage the complementary advantages of diverse sensors to enhance perception accuracy. Nevertheless, these approaches [6-9] often impose substantial hardware overhead and exhibit limited adaptability, particularly when deployed on resource-constrained edge platforms [10]. Even with the assistance of low-power machine learning compilation frameworks [11], such systems may still execute redundant computations under straightforward scenarios, introducing additional delays that undermine timely safety-critical decision-making." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 387, + 294, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 387, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 387, + 294, + 624 + ], + "type": "text", + "content": "Nevertheless, multimodal approaches provide a natural foundation for addressing the accuracy challenges faced by ADS. Rather than relying on a single integrated model designed to operate uniformly across diverse driving scenarios, the MoE framework [12] dynamically activates specialized sub-models, or experts, based on input characteristics and prior knowledge. MoE has demonstrated considerable effectiveness in large-scale applications, such as Large Language Models (LLMs) for natural language processing [13, 14], where it routes different tokens to appropriate experts, reducing redundant computations without sacrificing efficiency [15]. As model complexity grows, MoE enables models to scale their capacity while keeping inference hardware costs manageable. Deploying MoE in ADS, however, poses notable challenges [16], including strict real-time constraints, the need for multimodal data processing, and seamless end-to-end autonomous feedback control. Successful deployment requires efficient data preprocessing, adaptive modality selection, and robust expert routing mechanisms." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 630, + 294, + 712 + ], + "type": "text", + "content": "By leveraging diverse sensor data and a scenario-adaptive MoE dispatcher, we dynamically assign expert sub-models based on scenario data and prior statistical knowledge, providing an efficient, energy-efficient, and safe solution for ADS. To this end, we introduce an Edge-based MoE Collaborative Computing (EMC2) framework - an algorithm-hardware-software co-design solution that not" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 317, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 119 + ], + "type": "text", + "content": "only improves accuracy through complex multimodal experts but also effectively reduces latency by utilizing an adaptive MoE dispatcher and lightweight single-modal experts. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 121, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 121, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 121, + 553, + 216 + ], + "type": "text", + "content": "1) A Multimodal 3D Object Detection Solution with MoE. The proposed EMC2 system achieves both high detection accuracy and inference efficiency, outperforming 15 existing one-size-fits-all baseline models. Designed for real-world ADS applications, EMC2 is optimized for deployment on resource-constrained platforms, making it more suitable for edge computing environments than generic alternatives." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 219, + 553, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 219, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 219, + 553, + 386 + ], + "type": "text", + "content": "2) Robust and Efficient Multimodal Perception through Expert Collaboration and System-Level Optimization. The proposed EMC2 framework incorporates scenario-specific experts and employs an adaptive expert selection strategy to meet heterogeneous perception demands. To mitigate the training instability commonly observed in MoE-based systems, we design a hierarchical optimization scheme that combines expert-level and global EMC2 backpropagation for effective collaborative learning. Furthermore, a distribution-aware resampling technique enhances robustness under real-world data imbalance. In addition to these designs, a tailored multiscale pooling module improves the efficiency of LiDAR-camera fusion, ensuring both efficient and accurate multimodal perception." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 387, + 553, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 387, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 317, + 387, + 553, + 518 + ], + "type": "text", + "content": "3) System-Level Optimization for Edge Inference. The original PyTorch-based MoE inference framework [17] suffers from inefficient management of memory and computational resources, leading to high latency on edge devices. To overcome these challenges, we propose an algorithm-hardware-software co-optimization strategy within EMC2, which incorporates efficient memory management techniques to improve L1/L2 cache hit rates and applies computational graph fusion to eliminate redundant data loading across layers. These combined optimizations notably improve inference efficiency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "type": "text", + "content": "4) State-of-the-art (SOTA) Accuracy and Efficiency Combined. On the KITTI dataset, EMC2 achieves " + }, + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "type": "inline_equation", + "content": "3.28\\%" + }, + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "type": "text", + "content": " higher accuracy for pedestrians and " + }, + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "type": "inline_equation", + "content": "6.03\\%" + }, + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "type": "text", + "content": " for bicycles under hard-level detection challenge compared to 15 baseline models. On the nuScenes dataset, it yields a " + }, + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "type": "text", + "content": " increase in mAP and a " + }, + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "type": "text", + "content": " boost in NDS. In addition, EMC2 achieves a " + }, + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "type": "inline_equation", + "content": "159.06\\%" + }, + { + "bbox": [ + 317, + 522, + 553, + 629 + ], + "type": "text", + "content": " reduction in inference latency on the Jetson platform for KITTI, supporting its suitability for real-time edge deployment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 635, + 400, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 400, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 400, + 647 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 654, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 654, + 553, + 712 + ], + "type": "text", + "content": "LiDAR-only Method. The active 3D sensing of LiDAR makes it dominant in ADS 3D detection [18-20]. LiDAR-based methods process point clouds via voxelization [21], pillarization [20], or direct raw input [22]. Anchor-based heads, used for region proposal [21], and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25904" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "content": "center-based representations[23] are widely adopted. These methods achieve high accuracy on KITTI, as demonstrated by 3DSSD [24], PartA2 [19], and Voxel R-CNN [3]. However, weak LiDAR reflections for distant or occluded objects degrade detection performance [25, 26]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 135, + 294, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 135, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 135, + 294, + 255 + ], + "type": "text", + "content": "Multimodal Fusion Method. Combining LiDAR and camera RGB input enhances detection performance, especially for sparse LiDAR objects [27]. Fusion approaches include projection-based methods, which align image features with point clouds in 3D space [28], and model-based techniques that incorporate cross-attention [29] or unified feature spaces [30]. Recent multimodal models, such as PPF-Det [31] and RoboFusion [32], achieve top accuracy on KITTI benchmarks. However, these models often trade off accuracy for speed or vice versa." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 258, + 296, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 296, + 425 + ], + "type": "text", + "content": "Edge-based Compilation. ML compilation platforms are essential for improving the inference efficiency of deep learning models across diverse hardware architectures [33, 34]. TensorRT [35] is an NVIDIA inference optimizer that accelerates ML models on GPUs through layer fusion, precision calibration, and kernel auto-tuning. XLA [36] is a domain-specific compiler for TensorFlow enhancing efficiency by fusing operations and optimizing computation graphs. Open Neural Network Exchange (ONNX) Runtime [37] provides an open standard for model interoperability, enabling efficient execution across CPUs, GPUs, and accelerators. TVM [11] is an open-source deep learning compiler that automates model tuning and tensor scheduling for efficient deployment across CPUs and GPUs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 430, + 235, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 235, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 235, + 443 + ], + "type": "text", + "content": "3. EMC2 Computing Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 594 + ], + "type": "text", + "content": "We introduce EMC2, a multimodal MoE framework that minimizes latency and ensures reliable accuracy. Fig. 2 shows the EMC2 overview, EMC2 consists of: 1) Adaptive Multimodal Data Bridge (AMDB) to preprocess multimodal input (Sec. 3.1); 2) Scenario-Adaptive Dispatcher to adaptively dispatch a suitable expert for each scenario (Sec. 3.2); and 3) Three Scenario-Optimized Experts to decode AMDB-output features into final results (Sec. 3.3). We also design an effective training strategy to address the long-tail effect in the training process (Sec. 3.4). In addition, we deploy EMC2 on edge devices with self-designed 3D sparse convolution and Multiscale Pooling." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 238, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 238, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 238, + 613 + ], + "type": "text", + "content": "3.1. Adaptive Multimodal Data Bridge" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Given the multimodal nature of EMC2's sensor inputs, a module providing suitable preprocessed data for different experts is required. Thus, we propose AMDB, a preprocessing unit to efficiently process the required input data for each expert. The workflow of AMDB begins with a UNet [38] and a sparse CNN to extract LiDAR features, followed by fully connected layers that generate proposal regions along with their confidence scores. These results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "serve as inputs to Latency-Prioritized Expert and Versatile Efficiency Expert. The extraction of image features depends on the Scenario-Adaptive Dispatcher. If further information is required for reliable detection, AMDB will extract image features by ResNet. Based on depth information, it projects relevant image pixels into 3D space, applying Multiscale Pooling to reduce computational overhead, and fuses them with LiDAR voxel features to produce a multimodal output. These results are inputs for Accuracy-Prioritized Experts. The Multiscale Pooling and multimodal fusion processes are demonstrated in Fig. 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 211, + 478, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 478, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 478, + 224 + ], + "type": "text", + "content": "3.2. Scenario-Adaptive Dispatcher" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 229, + 553, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 553, + 301 + ], + "type": "text", + "content": "The Scenario-Adaptive Dispatcher (SAD) dynamically dispatches each scenario to the most suitable expert based on traffic conditions and latency-accuracy requirements. Our expert selection algorithm considers real-world traffic safety needs: specifically, how far away objects are and how clearly they can be perceived." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "text", + "content": "So, the switch among experts is determined by two key parameters: object distance " + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "text", + "content": " and clarity, which is inferred from the confidence " + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "text", + "content": " of proposal regions provided by the AMDB. We classify traffic scenarios into three categories: 1) Close and Distinct Cases. All objects are clearly visible and no distant objects are present. This is characterized by all proposal regions " + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "inline_equation", + "content": "< \\mathcal{D}" + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "text", + "content": " and having confidence " + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "inline_equation", + "content": "\\geq \\mathcal{C}" + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "text", + "content": ". In such scenarios, as near-field objects provide abundant Li-DAR information, even when compressed into bird's eye view (BEV), a 2D CNN can effectively capture key patterns such as object surfaces and heights [20], ensuring reliable detection accuracy with minimal latency (Sec. 4.4). These cases will be routed to the Latency-Prioritized Expert. 2) Mixed Visibility Cases. Some objects are either distant but clearly visible or near but unclear. Mathematically, this corresponds to cases where i) some proposal regions " + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "inline_equation", + "content": "< \\mathcal{D}" + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "text", + "content": " with confidence " + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "inline_equation", + "content": "< \\mathcal{C}" + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "text", + "content": ", or ii) the distance of some objects " + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "inline_equation", + "content": "\\geq \\mathcal{D}" + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "text", + "content": " with confidence " + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "inline_equation", + "content": "\\geq \\mathcal{C}" + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "text", + "content": ". Here, objects have insufficient voxel data, requiring more complex 3D convolution to extract meaningful patterns, and they will be dispatched to the Versatile Efficiency Expert. 3) Distant and Uncertain Cases. Some objects are both distant and unclear. This occurs when some proposal regions " + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "inline_equation", + "content": "\\geq \\mathcal{D}" + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "text", + "content": " with confidence " + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "inline_equation", + "content": "< \\mathcal{C}" + }, + { + "bbox": [ + 313, + 302, + 553, + 638 + ], + "type": "text", + "content": ". In these challenging scenarios, significant voxel information is missing due to long distances, steep reflection angles, or occlusions, necessitating the integration of image data to compensate for missing LiDAR details. These cases will be processed by the Accuracy-Prioritized Expert." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 647, + 470, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 470, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 470, + 660 + ], + "type": "text", + "content": "3.3. Scenario-Optimized Experts" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Latency-Prioritized Expert (LPE). LPE uses 2D CNNs for object detection based on BEV projections. It first projects proposal regions from the AMDB into BEV space, then extracts 2D features for object localization and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25905" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 337 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 337 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 337 + ], + "type": "image", + "image_path": "f13f7eecda8ca25f2db6748d167f0ac62896c17fa918088cec8473745c50189c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 341, + 555, + 429 + ], + "lines": [ + { + "bbox": [ + 55, + 341, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 555, + 429 + ], + "type": "text", + "content": "Figure 2. System Structure of EMC2. The central row illustrates the architecture of EMC2, while the top and bottom rows provide visual examples of core modules and relevant legends. The system integrates five main components: an Adaptive Multimodal Data Bridge (AMDB) for preprocessing raw multimodal input and generating expert-specific features; a Scenario-Adaptive Dispatcher that dispatches suitable experts based on scenario-specific features (see Fig. 3); the Latency-Prioritized Expert designed for simple scenarios requiring minimal inference delay; the Versatile Efficiency Expert designed for uncertain or distant object detection scenarios; and the Accuracy-Prioritized Expert responsible for handling complex, high-precision detection scenarios. Additionally, an Emergency Expert API is designed to support rapid response to hazardous or unseen situations. The overall training process incorporates loss terms for both final detection results and intermediate outputs from the AMDB, as described in Sec. 3.4." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 438, + 294, + 558 + ], + "blocks": [ + { + "bbox": [ + 58, + 438, + 294, + 558 + ], + "lines": [ + { + "bbox": [ + 58, + 438, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 58, + 438, + 294, + 558 + ], + "type": "image", + "image_path": "8cb87474ddee652c12453ab79550d6e0ea89ea00809e0beb2ac65cf22d4d29ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 564, + 295, + 588 + ], + "lines": [ + { + "bbox": [ + 55, + 564, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 295, + 588 + ], + "type": "text", + "content": "Figure 3. Routing Strategy of the Scenario-Adaptive Dispatcher. The Routing criteria is detailed mathematically in Sec. 3.2." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": "classification. It finally maps detection results back to 3D space. Compared to sparse 3D convolutions, its use of 2D operations reduces computation in outdoor ADS scenarios [20], allowing LPE to improve inference efficiency with minimal accuracy loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Versatile Efficiency Expert (VEE). VEE uses a 3D CNN to process proposal regions and confidence scores from the AMDB. After sparse convolution, voxel features are decoded into per-voxel classifications and bounding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 439, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 475 + ], + "type": "text", + "content": "boxes. Compared to LPE, 3D convolutions better preserve spatial structure, making it more robust for scenarios with insufficient LiDAR signal." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 475, + 556, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 607 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 607 + ], + "type": "text", + "content": "Accuracy-Prioritized Expert (APE). APE extends VEE by incorporating the multimodal signal. It employs a 3D CNN and decoder with the same structure but different parameters to process multimodal proposal regions (from AMDB) and fuses image data to recover missing LiDAR information. Image features compressed via Multiscale Pooling (Sec.3.5) are projected into 3D space and fused with corresponding voxels of LiDAR features, ensuring more robust understanding for occluded and distant objects while keeping computation feasible. The Multiscale Pooling and multimodal fusion are detailed in Fig. 5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 612, + 553, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 553, + 625 + ], + "type": "text", + "content": "3.4. Collaborative Training of Multimodal Experts" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Hierarchical Training Strategy: Multimodal Supervision with Triple Back-Propagation. During training, the performance of experts is highly dependent on the proposal regions produced by the AMDB, yet these proposals tend to be unreliable in the early training phase, leading to an unstable expert learning process and unstable SAD decisions. To address this, we adopt a hierarchical supervision stra" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25906" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 71, + 270, + 291 + ], + "blocks": [ + { + "bbox": [ + 80, + 71, + 270, + 291 + ], + "lines": [ + { + "bbox": [ + 80, + 71, + 270, + 291 + ], + "spans": [ + { + "bbox": [ + 80, + 71, + 270, + 291 + ], + "type": "image", + "image_path": "bf1cea59c5f10b342486ad5d82aa8d91aa2198df29496c17f3124dc349209ef1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 296, + 296, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 296, + 361 + ], + "type": "text", + "content": "Figure 4. Illustration of Hierarchical Training. The three backpropagation paths originate from: 1) the final detection output, 2) the LiDAR branch output of AMDB, and 3) the image branch output of AMDB. Since expert training relies on initially unreliable AMDB outputs, direct supervision accelerates convergence and enhances representation learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 373, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 480 + ], + "type": "text", + "content": "egy with three distinct back-propagation pathways, as illustrated in Fig. 4: 1) supervision of the LiDAR branch outputs from the AMDB; 2) supervision of the image branch outputs from the AMDB; and 3) supervision of expert predictions and associated confidence scores. To further improve training stability, the LiDAR and image branches of the AMDB are pre-trained separately before joint optimization. Each back-propagation route is supervised by the following loss function, which collectively forms Multimodal Loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 489, + 294, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 489, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 134, + 489, + 294, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r o u t e}} = \\mathcal {L} _ {\\text {c l s}} + \\mathcal {L} _ {\\text {r e g}} \\tag {1}", + "image_path": "8cdbbff99ee367e3de1a7ce07b8c23f371316989f1dc3a4d7283ee238d538ec6.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "content": " uses cross-entropy loss and " + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "content": " uses smooth-" + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\ell 1" + }, + { + "bbox": [ + 55, + 510, + 295, + 581 + ], + "type": "text", + "content": " loss. The Hierarchical Training not only stabilizes expert learning but also enhances the model's adaptability to various data granularities. In turn, even under compressed data representations, it enables the effective use of our Multiscale Pooling for memory efficiency, as detailed in Sec. 3.5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Addressing Long-tail Effects in MoE Training. The MoE framework often suffers from long-tail effects, where certain experts receive insufficient training data, leading to imbalances in sub-datasets, and therefore affecting performance [39]. To address this, we propose a threefold training strategy: 1) Data subset division. We partition the dataset into sub-datasets based on target and auxiliary samples. Each expert " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " is dispatched a sub-dataset " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "(i = 1,2,3)" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " containing target samples " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{S}_i}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ", which are expected to be routed to the corresponding expert, and auxiliary samples " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{\\mathcal{S}_i}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ", which are not. We ensure that " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathcal{S}_i} \\cap \\mathcal{T}_{\\mathcal{S}_j} = \\emptyset" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " for any" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 554, + 198 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 554, + 198 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 554, + 198 + ], + "type": "image", + "image_path": "370e7b99ef66e8ccacbdbb6143a6bc15182d6a6bf5f37353615c9a6ae204e431.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 203, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 312 + ], + "type": "text", + "content": "Figure 5. Illustration of Multiscale Pooling and Multimodal Fusion. Proposal regions with image features are first obtained from the 3D proposal space generated by the LiDAR branch of AMDB. After extracting the relevant image features from the corresponding image regions and projecting them into 3D space, pooling and fusion are then applied. Following the pooling step, image features of pixels that share the same 3D coordinate as a voxel are concatenated with the corresponding LiDAR voxel features, while those without matching voxel coordinates are concatenated with zero-filled features." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 316, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 376 + ], + "type": "inline_equation", + "content": "i \\neq j" + }, + { + "bbox": [ + 313, + 316, + 554, + 376 + ], + "type": "text", + "content": ". The inclusion of auxiliary samples, which are not intended for a given expert, helps prevent overfitting and enhances generalization. 2) Balanced sampling strategy. We adjust the selection probability " + }, + { + "bbox": [ + 313, + 316, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 313, + 316, + 554, + 376 + ], + "type": "text", + "content": " for the sub-datasets which have fewer samples to ensure balanced training:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 379, + 553, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 379, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 359, + 379, + 553, + 411 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {i} \\times \\mathcal {N} _ {\\mathcal {S} _ {i}} = \\mathcal {P} _ {j} \\times \\mathcal {N} _ {\\mathcal {S} _ {j}}, \\sum_ {i = 1} ^ {3} \\mathcal {P} _ {i} = 1 \\tag {2}", + "image_path": "13f160df87c346ff2856cb441797278af1358c86454a7a3eb512ddec114307a1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_{S_i}" + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "text", + "content": " represents the number of samples in sub-datasets " + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "text", + "content": ". This ensures that experts dispatched to less frequent scenarios receive sufficient training iterations. 3) Adaptive optimizer. We introduce an optimizer that adapts the learning rate based on the proportion of target samples in a batch. Denoting " + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\alpha_0" + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "text", + "content": " as the base learning rate, the adaptive learning rate applied to an expert with a given batch is computed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 372, + 506, + 553, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 506, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 372, + 506, + 553, + 539 + ], + "type": "interline_equation", + "content": "\\alpha = \\left(1 + \\frac {1}{\\mathcal {N}} \\sum_ {i = 1} ^ {\\mathcal {N}} \\mathbf {1} _ {i \\in \\mathcal {T}}\\right) \\cdot \\alpha_ {0} \\tag {3}", + "image_path": "8b33283adefa08c907e4b37b7d6c9dceab3482267f37eb69c87892e173bdb367.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "text", + "content": " is the batch size, " + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "text", + "content": " is the proportion of target samples within current mini-batch, " + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 545, + 554, + 617 + ], + "type": "text", + "content": " represents target samples in the batch. By regulating the gradient update rate based on the batch composition, this adaptive optimizer encourages each expert to prioritize its designated target samples while preventing excessive updates from auxiliary samples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 624, + 528, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 528, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 528, + 637 + ], + "type": "text", + "content": "3.5. Algorithm Edge-adaptivity Optimization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Edge devices face challenges such as incomplete function libraries and limited cache. We adopt the open-source ONNX Runtime for compiling EMC2 on edge, which provides modules that help address these challenges and facilitate edge-oriented optimization. As 3D convolutions account for the majority of computation during inference, we" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25907" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "customize 3D sparse convolution and Multiscale Pooling to improve the algorithmic efficiency of EMC2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "text", + "content": "3D Sparse Convolution. Since the 3D sparse convolution library is incomplete on most edge platforms, we develop a customized 3D sparse convolution library with parallel threads execution under ONNX Runtime. Let " + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "text", + "content": " denote spatial resolution, " + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "text", + "content": " the number of non-empty voxels, and " + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_{\\mathrm{v}}" + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "text", + "content": " the per-voxel computation, including channels and kernel size. By computing nonzero voxels only, 3D sparse convolution reduces complexity from " + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "inline_equation", + "content": "O(\\mathcal{H}^3\\times \\mathcal{C}_{\\mathrm{v}})" + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "text", + "content": " for dense convolutions to " + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "inline_equation", + "content": "O(\\mathcal{N}\\times \\mathcal{C}_{\\mathrm{v}})" + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "inline_equation", + "content": "\\mathcal{N}\\ll \\mathcal{H}^3" + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "text", + "content": ", cutting redundant operations by " + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "inline_equation", + "content": "65 - 80\\%" + }, + { + "bbox": [ + 55, + 100, + 296, + 220 + ], + "type": "text", + "content": " [40]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 224, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 319 + ], + "type": "text", + "content": "Multiscale Pooling. To alleviate cache limitations, Multiscale 3D Sparse Pooling adaptively pools image features based on available memory before fusing them with LiDAR features (Fig.5). The pooling size is user-configurable, allowing flexibility for maximizing memory utilization. When combined with Hierarchical Training (Sec.3.4), Multiscale Pooling effectively reduces resource consumption while maintaining model reliability (Sec. 4.4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 331, + 266, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 266, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 266, + 344 + ], + "type": "text", + "content": "3.6. EMC2 Computing System Optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 349, + 296, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 409 + ], + "type": "text", + "content": "Building upon the algorithm-level optimizations introduced in Sec. 3.5, we further enhance the ONNX execution efficiency through system-level improvements, specifically targeting memory management and computation graph optimization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 592 + ], + "type": "text", + "content": "Memory Optimization. To further improve memory efficiency during inference, we implement three system-level techniques. 1) Overlapping communication and computation [41]. Matrix data is partitioned into segments processed and transferred in parallel, overlapping computation and communication. This reduces global memory access, improves cache usage, and mitigates memory bottlenecks. 2) Thread scheduling. We implement a staged thread management mechanism, where active threads are periodically terminated and re-initialized (Fig. 6). This approach limits unnecessary memory retention between stages, reducing system memory footprint. 3) Prefix-sum for sparse convolution. Parallel prefix-sum reduces index search time from " + }, + { + "bbox": [ + 55, + 413, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\mathcal{N})" + }, + { + "bbox": [ + 55, + 413, + 295, + 592 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 413, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\log \\mathcal{N})" + }, + { + "bbox": [ + 55, + 413, + 295, + 592 + ], + "type": "text", + "content": " using GPU parallelism, where " + }, + { + "bbox": [ + 55, + 413, + 295, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 413, + 295, + 592 + ], + "type": "text", + "content": " is the number of nonzero elements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Computational Graph Optimization. ONNX partitions the computation graph into segments and assigns each to the most suitable hardware. Our optimization includes three stages: 1) Model pruning. Redundant parameters and operations are removed to simplify the graph; 2) Model quantization. Low-precision arithmetic replaces floating-point computation to improve hardware efficiency; 3) Computational graph fusion. Consecutive operators are merged to reduce memory access and improve execution, benefiting memory-bound CPUs and reducing kernel" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 424, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 424, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 424, + 83 + ], + "type": "text", + "content": "launch overhead on GPUs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 94, + 448, + 107 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 94, + 448, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 94, + 448, + 107 + ], + "type": "text", + "content": "4. Experiment and Result" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 111, + 495, + 124 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 111, + 495, + 124 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 495, + 124 + ], + "type": "text", + "content": "4.1. Experiment Settings and Datasets" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 129, + 553, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 129, + 553, + 189 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 553, + 189 + ], + "type": "text", + "content": "Experiment Platform. Experiments are conducted on the Jetson AGX Orin, an edge platform with limited computational resources, an NVIDIA Ampere GPU, a 4MB shared L2 cache, and hierarchical memory. For comparison, EMC2 is also evaluated on an NVIDIA A4000 GPU." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": "Dataset and Sample Difficulties. We evaluate 3D object detection on two public benchmarks: KITTI [42] and nuScenes [43]. KITTI provides 3,712 training and 3,769 validation samples, with object difficulty levels categorized as " + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": " easy, " + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": " moderate, and " + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": " hard. nuScenes contains 850 training and 150 test scenes, covering 10 object categories with 3D bounding box annotations. Before splitting the data, we analyze the confidence distribution of proposal regions generated by the pre-trained AMDB. A two-sample Kolmogorov-Smirnov test reveals a statistically significant difference between objects within and beyond a distance threshold " + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": ", which is empirically set to 23.5 meters for KITTI and 35 meters for nuScenes. The training subsets for each expert on the KITTI and nuScenes datasets are defined as follows: " + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{S_1}" + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": " includes scenes where all objects are within " + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": " meters and labeled as \"easy\" or \"moderate\"; " + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{S_2}" + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": " includes scenes with at least one object beyond " + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": " meters labeled as \"easy\" or \"moderate,\" or any object within " + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": " meters labeled as \"hard\"; and " + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{S_3}" + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": " includes scenes where at least one object is both beyond " + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": " meters and labeled as \"hard,\" requiring multimodal processing. For each expert, the auxiliary subset " + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{S_i}" + }, + { + "bbox": [ + 313, + 193, + 555, + 479 + ], + "type": "text", + "content": " is uniformly sampled from the other experts' target subsets, with its size equal to that of the corresponding target subset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 590 + ], + "type": "text", + "content": "Evaluation Metrics. KITTI uses R40 3D Average Precision (AP) with IoU thresholds of 0.7 for cars and 0.5 for pedestrians and cyclists. nuScenes adopts the NDS metric, combining mAP—computed via center distance matching " + }, + { + "bbox": [ + 313, + 483, + 554, + 590 + ], + "type": "inline_equation", + "content": "(0.5\\mathrm{m}" + }, + { + "bbox": [ + 313, + 483, + 554, + 590 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 483, + 554, + 590 + ], + "type": "inline_equation", + "content": "4.0\\mathrm{m}" + }, + { + "bbox": [ + 313, + 483, + 554, + 590 + ], + "type": "text", + "content": " thresholds)—and five error terms: mean Average Translation (mATE), Scale (mASE), Orientation (mAOE), Velocity (mAVE), and Attribute (mAAE) Errors. Inference latency on Jetson AGX Orin is also reported to assess real-time performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "type": "text", + "content": "Hierarchical Training Details. Allowing all experts to propagate gradients to AMDB causes instability due to overlapping training data, leading to repeated updates. To address this, AMDB is first pre-trained for 20 epochs, followed by joint training where gradients are propagated only from APE, which receives full multimodal input." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 671, + 534, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 534, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 534, + 685 + ], + "type": "text", + "content": "4.2. Comparison with Existing SOTA Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "As shown in Tab. 1 and Tab. 2, EMC2 achieves SOTA accuracy and latency on KITTI and nuScenes. On the Jetson" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25908" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 72, + 527, + 194 + ], + "blocks": [ + { + "bbox": [ + 80, + 72, + 527, + 194 + ], + "lines": [ + { + "bbox": [ + 80, + 72, + 527, + 194 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 527, + 194 + ], + "type": "image", + "image_path": "d62ca3cdfe936cfa47d6ab4d81ab4d294870be561966d0dd6faf48048b96e60b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 199, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 199, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 555, + 222 + ], + "type": "text", + "content": "Figure 6. Illustration of Thread Management. Each solid box represents a module occupying threads. During inference, four boundaries are defined at which preceding threads are terminated, and their thread-specific address space is released upon crossing each boundary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 225, + 556, + 437 + ], + "blocks": [ + { + "bbox": [ + 56, + 225, + 556, + 437 + ], + "lines": [ + { + "bbox": [ + 56, + 225, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 225, + 556, + 437 + ], + "type": "image", + "image_path": "c51465c9b24cc8b4e0c05547de5828eb66ab9ef9357ae08a3e9df1f8c130a1cb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 441, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 555, + 464 + ], + "type": "text", + "content": "Table 1. We evaluate different 3D object detection methods on the KITTI dataset. Some cells are denoted as " + }, + { + "bbox": [ + 55, + 441, + 555, + 464 + ], + "type": "inline_equation", + "content": "N / A" + }, + { + "bbox": [ + 55, + 441, + 555, + 464 + ], + "type": "text", + "content": " when the corresponding open-source implementation lacks support for the specific object class or is not compatible with Jetson devices." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 472, + 291, + 522 + ], + "blocks": [ + { + "bbox": [ + 59, + 472, + 291, + 522 + ], + "lines": [ + { + "bbox": [ + 59, + 472, + 291, + 522 + ], + "spans": [ + { + "bbox": [ + 59, + 472, + 291, + 522 + ], + "type": "table", + "html": "
MethodSizemAP↑NDS↑Latency (ms)↓
FocalFormer [55]189M0.66400.7090N/A
BEVFusion [56]156M0.68520.7138N/A
EMC2 (ours)87M0.72410.7316229.3
", + "image_path": "5a1e6dee6a34a22d1c59ddd3bb508a780f1eb162ba06db3f63d895f056d8b6ff.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 527, + 295, + 539 + ], + "lines": [ + { + "bbox": [ + 56, + 527, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 56, + 527, + 295, + 539 + ], + "type": "text", + "content": "Table 2. Comparison of different methods on the nuScenes dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 560, + 295, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 295, + 585 + ], + "type": "text", + "content": "AGX Orin platform, it also outperforms existing baseline models across all difficulty levels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "text", + "content": "Results on KITTI. Pedestrians and cyclists face higher traffic risks [54], which remain challenging for most SOTA methods. EMC2 improves hard-set detection by " + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "text", + "content": " for pedestrians and cyclists, respectively, outperforming TED and LoGoNet (Tab. 1). For cars, it achieves " + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "inline_equation", + "content": "88.15\\%" + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "text", + "content": " on the hard set, exceeding prior multimodal methods by up to " + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "text", + "content": ". It runs at " + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "inline_equation", + "content": "372.5 \\mathrm{~ms}" + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "text", + "content": " on Jetson, " + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "inline_equation", + "content": "2.5 \\times -2.5 \\times" + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "text", + "content": " faster than methods requiring " + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "inline_equation", + "content": "965 - 1" + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "inline_equation", + "content": "787 \\mathrm{~ms}" + }, + { + "bbox": [ + 55, + 589, + 295, + 685 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Results on nuScenes. EMC2 improves mAP and NDS by " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", respectively, outperforming recent" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 473, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 510 + ], + "type": "text", + "content": "methods FocalFormer [55] and BEVFusion [56] (Tab. 2). It runs at " + }, + { + "bbox": [ + 313, + 473, + 555, + 510 + ], + "type": "inline_equation", + "content": "229.3\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 473, + 555, + 510 + ], + "type": "text", + "content": " on Jetson AGX Orin, enabling real-time use on edge devices." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 520, + 449, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 449, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 449, + 533 + ], + "type": "text", + "content": "4.3. Analysis and Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 538, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 555, + 598 + ], + "type": "text", + "content": "Benefits of Mixture-of-Experts. The balance between accuracy and efficiency observed in Sec. 4.2 is achieved through dynamic expert selection based on scenario characteristics, combined with algorithmic and system-level optimizations designed for edge devices, such as Jetson." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 601, + 556, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 556, + 698 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 556, + 698 + ], + "type": "text", + "content": "Empirical Dispatcher. The EMC2 dispatcher uses fixed distance and confidence thresholds to dispatch scenes to experts. The protocol is adapted per dataset format. This strategy is validated via expert activations on the nuScenes validation set (Fig. 7): LPE performs best in close-range, high-confidence scenes; APE in distant, low-confidence cases; and VEE in all others, supporting the effectiveness of empirical routing." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "Optimized Parameter Size. EMC2 has 226M param-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25909" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 291, + 184 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 291, + 184 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 291, + 184 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 291, + 184 + ], + "type": "image", + "image_path": "0783037485353f1a17162a96132bcc397f417a468d0c2ef34a3fefe2db0f6c3c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 188, + 296, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 296, + 277 + ], + "type": "text", + "content": "Figure 7. Distribution of expert activation. During EMC2 inference on the nuScenes validation set, scenes are grouped into a 2D grid based on two metrics: the y-axis indicates the number of AMDB-generated proposals beyond the distance threshold " + }, + { + "bbox": [ + 55, + 188, + 296, + 277 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 188, + 296, + 277 + ], + "type": "text", + "content": " (i.e., distant objects), and the x-axis represents the average confidence score of these proposals (higher values indicate clearer scenes). Each cell is labeled with the name of the expert that performs best for scenes falling into that bin." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 286, + 291, + 403 + ], + "blocks": [ + { + "bbox": [ + 59, + 286, + 291, + 403 + ], + "lines": [ + { + "bbox": [ + 59, + 286, + 291, + 403 + ], + "spans": [ + { + "bbox": [ + 59, + 286, + 291, + 403 + ], + "type": "image", + "image_path": "a8ad0f87d16bfc084782903972c7c98a5c73af69fe8048b9bcf032eadfe65aa1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 411, + 295, + 477 + ], + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 477 + ], + "type": "text", + "content": "Figure 8. Median Detection Confidence vs. Object Distances. Four configurations are compared on the KITTI validation set: LPE (blue), VEE (peach), APE (rose), and EMC2 (green). More complex experts yield higher confidence but slower inference. By switching experts based on distance and clarity, EMC2 balances accuracy and efficiency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 493, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 295, + 564 + ], + "type": "text", + "content": "eters, but its MoE framework activates only a subset during inference, averaging 87M on nuScenes and 146M on KITTI, which is fewer than those used by the full model and other SOTA baselines. System-level optimizations instead prioritize computational throughput and memory efficiency over parameter reduction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 628 + ], + "type": "text", + "content": "Hardware Flexibility. The EMC2 is implemented in the ONNX format, an open format supported by various compilers and enables model compilation across diverse hardware platforms. As a result, EMC2 is not limited to Jetson and can be deployed on other edge platforms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 149, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 149, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 149, + 648 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "Ablation on the Roles of Three Experts. We evaluate each expert on its target scenarios using the KITTI validation set. As shown in Tab. 3, LPE, VEE, and APE achieve comparable accuracy in LPE scenes, but LPE has the lowest inference latency on Jetson. In VEE scenarios, incor" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 547, + 260 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 547, + 260 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 547, + 260 + ], + "type": "table", + "html": "
MethodAP↑Latency (ms) ↓
Ped.CarCyclist
Ablation on LPE
LPE66.3289.0488.54219
VEE74.4992.2391.69468
APE74.3493.2192.41526
Ablation on VEE
W/ VEE70.5590.5084.07372.5
W/O VEE68.2686.4079.17512.7
Ablation on APE
VEE67.2487.6981.85468
APE70.5590.5084.18526
Ablation on Hierarchical Training and Multiscale Pooling
HT √, MP √70.5590.5084.07526
HT √, MP ×70.5590.5084.181320
HT ×, MP √63.1188.6875.98529
HT ×, MP ×66.3489.1080.831367
", + "image_path": "b8dd0d61261168fbbc2a23a2ec6c81ad2d2c6b2ef6e290e6111ed79b98e7e993.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 268, + 553, + 323 + ], + "lines": [ + { + "bbox": [ + 313, + 268, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 553, + 323 + ], + "type": "text", + "content": "Table 3. Results Across Three Ablation Configurations. The main body of this table is divided into four sections corresponding to each part of the ablation study. In the third section, HT and MP are abbreviated for Hierarchical Training and Multiscale Pooling, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 333, + 553, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 553, + 381 + ], + "type": "text", + "content": "porating VEE improves both accuracy and latency over the baseline. For APE, which incorporates image features, accuracy improves further with only a minor latency increase, supported by the score distribution in Fig. 8." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 384, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 503 + ], + "type": "text", + "content": "Ablation on Hierarchical Training (HT) and Multiscale Pooling (MP). We evaluate four combinations of HT and MP on the full KITTI validation set: both applied, HT only, MP only, and neither. Tab. 3 shows that combining HT and MP reduces inference time on Jetson from 1320 ms to 526 ms without accuracy loss. For the car class, MP alone reduces accuracy from " + }, + { + "bbox": [ + 313, + 384, + 555, + 503 + ], + "type": "inline_equation", + "content": "80.83\\%" + }, + { + "bbox": [ + 313, + 384, + 555, + 503 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 384, + 555, + 503 + ], + "type": "inline_equation", + "content": "75.98\\%" + }, + { + "bbox": [ + 313, + 384, + 555, + 503 + ], + "type": "text", + "content": ", while HT alone improves it from " + }, + { + "bbox": [ + 313, + 384, + 555, + 503 + ], + "type": "inline_equation", + "content": "66.34\\%" + }, + { + "bbox": [ + 313, + 384, + 555, + 503 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 384, + 555, + 503 + ], + "type": "inline_equation", + "content": "70.55\\%" + }, + { + "bbox": [ + 313, + 384, + 555, + 503 + ], + "type": "text", + "content": ". HT facilitates multigranular feature learning, while MP enhances memory efficiency via compression." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 531, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 651 + ], + "type": "text", + "content": "We present EMC2, a multimodal 3D object detection solution for ADS based on a customized MoE architecture. This design enables adaptive expert selection across diverse traffic scenarios, balancing detection accuracy and efficiency. Experiments on KITTI and nuScenes show strong performance, particularly under challenging conditions. On the Jetson platform, EMC2 achieves a " + }, + { + "bbox": [ + 313, + 531, + 555, + 651 + ], + "type": "inline_equation", + "content": "159.06\\%" + }, + { + "bbox": [ + 313, + 531, + 555, + 651 + ], + "type": "text", + "content": " speedup in inference, supporting real-time edge deployment. Future work will explore fully adaptive expert selection to enhance system responsiveness and resource efficiency." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25910" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 178 + ], + "type": "text", + "content": "[1] Shih-Chieh Lin, Yunqi Zhang, Chang-Hong Hsu, Matt Skach, Md E. Haque, Lingjia Tang, and Jason Mars. The architectural implications of autonomous driving: Constraints and acceleration. In Proceedings of the Twenty-Third International Conference on Architectural Support for Programming Languages and Operating Systems, ASPLOS '18, page 751-766, New York, NY, USA, 2018. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 179, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 179, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 179, + 296, + 223 + ], + "type": "text", + "content": "[2] Yi Xiao, Felipe Codevilla, Akhil Gurram, Onay Urfalioglu, and Antonio M López. Multimodal end-to-end autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 23(1):537-547, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "type": "text", + "content": "[3] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. arXiv:2012.15712, 2020. 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 270, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 294, + 334 + ], + "type": "text", + "content": "[4] Hao Frank Yang, Yifan Ling, Cole Kopca, Sam Ricord, and Yinhai Wang. Cooperative traffic signal assistance system for non-motorized users and disabilities empowered by computer vision and edge artificial intelligence. Transportation Research Part C: Emerging Technologies, 145:103896, 2022. Elsevier." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 335, + 294, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 294, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 294, + 388 + ], + "type": "text", + "content": "[5] Hao Yang, Jiarui Cai, Meixin Zhu, Chenxi Liu, and Yinhai Wang. Traffic-informed multi-camera sensing (TIMS) system based on vehicle re-identification. IEEE Transactions on Intelligent Transportation Systems, 23(10):17189-17200, 2022. IEEE. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 391, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 294, + 446 + ], + "type": "text", + "content": "[6] Guibiao Liao, Jiankun Li, and Xiaqing Ye. Vlm2scene: Self-supervised image-text-lidar learning with foundation models for autonomous driving scene understanding. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 3351-3359, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 447, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 501 + ], + "type": "text", + "content": "[7] Hao Frank Yang, Yang Zhao, Jiarui Cai, Meixin Zhu, JenqNeng Hwang, and Yiran Chen. Mitigating bias of deep neural networks for trustworthy traffic perception in autonomous systems. In Proceedings of the 2024 IEEE Intelligent Vehicles Symposium (IV), pages 633-638, 2024. IEEE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 502, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 556 + ], + "type": "text", + "content": "[8] Chenxi Liu, Hao Yang, Ruimin Ke, Wei Sun, Julia Wang, and Yinhai Wang. Cooperative and comprehensive multitask surveillance sensing and interaction system empowered by edge artificial intelligence. Transportation Research Record, 2677(9):652-668, 2023. SAGE Publications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 558, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 294, + 612 + ], + "type": "text", + "content": "[9] Hao Frank Yang, Jiarui Cai, Chenxi Liu, Ruimin Ke, and Yinhai Wang. Cooperative multi-camera vehicle tracking and traffic surveillance with edge artificial intelligence and representation learning. Transportation Research Part C: Emerging Technologies, 148:103982, 2023. Elsevier. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 614, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 668 + ], + "type": "text", + "content": "[10] Chen Zhang, Peng Li, Guangyu Sun, Yijin Guan, Bingjun Xiao, and Jason Cong. Optimizing fpga-based accelerator design for deep convolutional neural networks. In Proceedings of the 2015 ACM/SIGDA international symposium on field-programmable gate arrays, pages 161-170, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[11] Tianqi Chen, Thierry Moreau, Ziheng Jiang, Lianmin Zheng, Eddie Yan, Haichen Shen, Meghan Cowan, Leyuan Wang, Yuwei Hu, Luis Ceze, et al. {TVM}: An automated {End-to-End} optimizing compiler for deep learning. In 13th" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "USENIX Symposium on Operating Systems Design and Implementation (OSDI 18), pages 578-594, 2018. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": "[12] Jiarui Cai, Yizhou Wang, and Jenq-Neng Hwang. Ace: Ally complementary experts for solving long-tailed recognition in one-shot. In Proceedings of the IEEE/CVF international conference on computer vision, pages 112-121, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "text", + "content": "[13] Yunxin Li, Shenyuan Jiang, Baotian Hu, Longyue Wang, Wanqi Zhong, Wenhan Luo, Lin Ma, and Min Zhang. Unimoe: Scaling unified multimodal llms with mixture of experts. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[14] Jiachen Li, Xinyao Wang, Sijie Zhu, Chia-Wen Kuo, Lu Xu, Fan Chen, Jitesh Jain, Humphrey Shi, and Longyin Wen. Cumo: Scaling multimodal llm with co-upcycled mixture-of-experts. Advances in Neural Information Processing Systems, 37:131224-131246, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[15] Weilin Cai, Juyong Jiang, Fan Wang, Jing Tang, Sunghun Kim, and Jiayi Huang. A survey on mixture of experts. arXiv preprint arXiv:2407.06204, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 331 + ], + "type": "text", + "content": "[16] Xiaoming Shi, Shiyu Wang, Yuqi Nie, Dianqi Li, Zhou Ye, Qingsong Wen, and Ming Jin. Time-moe: Billion-scale time series foundation models with mixture of experts. arXiv preprint arXiv:2409.16040, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "text", + "content": "[17] Jiaao He, Jiezhong Qiu, Aohan Zeng, Zhilin Yang, Jidong Zhai, and Jie Tang. Fastmoe: A fast mixture-of-expert training system. arXiv preprint arXiv:2103.13262, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 366, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 420 + ], + "type": "text", + "content": "[18] Charles R. Qi, Xinlei Chen, Or Litany, and Leonidas J. Guibas. Imvotenet: Boosting 3d object detection in point clouds with image votes. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4403-4412, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 466 + ], + "type": "text", + "content": "[19] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. arXiv preprint arXiv:1907.03670, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "text", + "content": "[20] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689-12697, 2019. 2, 3, 4, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[21] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 624 + ], + "type": "text", + "content": "[22] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10526–10535, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "text", + "content": "[23] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11779-11788, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[24] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11037-11045, 2020. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25911" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[25] Andrew M Wallace, Abderrahim Halimi, and Gerald S Buller. Full waveform lidar for adverse weather conditions. IEEE transactions on vehicular technology, 69(7):7064-7077, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 296, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 296, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 296, + 160 + ], + "type": "text", + "content": "[26] You Li, Pierre Duthon, Michele Colomb, and Javier Ibanez-Guzman. What happens for a tof lidar in fog? IEEE Transactions on Intelligent Transportation Systems, 22(11):6670-6681, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 296, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 296, + 216 + ], + "type": "text", + "content": "[27] Fangchang Ma, Guilherme Venturelli Cavalheiro, and Sertac Karaman. Self-supervised sparse-to-dense: Self-supervised depth completion from lidar and monocular camera. In 2019 international conference on robotics and automation (ICRA), pages 3288-3295. IEEE, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 281 + ], + "type": "text", + "content": "[28] Tengteng Huang, Zhe Liu, Xiwu Chen, and Xiang Bai. E-net: Enhancing point features with image semantics for 3d object detection. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 35–52, Cham, 2020. Springer International Publishing. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 296, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 296, + 358 + ], + "type": "text", + "content": "[29] Yingwei Li, Adams Wei Yu, Tianjian Meng, Ben Caine, Ji-quaan Ngiam, Daiyi Peng, Junyang Shen, Yifeng Lu, Denny Zhou, Quoc V. Le, Alan Yuille, and Mingxing Tan. Deep-fusion: Lidar-camera deep fusion for multi-modal 3d object detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17161-17170, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 360, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 296, + 403 + ], + "type": "text", + "content": "[30] Yanwei Li, Yilun Chen, Xiaojuan Qi, Zeming Li, Jian Sun, and Jiaya Jia. Unifying voxel-based representation with transformer for 3d object detection. In Advances in Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "type": "text", + "content": "[31] Guotao Xie, Zhiyuan Chen, Ming Gao, Manjiang Hu, and Xiaohui Qin. Ppf-det: Point-pixel fusion for multi-modal 3d object detection. IEEE Transactions on Intelligent Transportation Systems, 25(6):5598-5611, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 491 + ], + "type": "text", + "content": "[32] Ziying Song, Guoxing Zhang, Lin Liu, Lei Yang, Shaoqing Xu, Caiyan Jia, Feiyang Jia, and Li Wang. Robofusion: Towards robust multi-modal 3d object detection via sam. arXiv preprint arXiv:2401.03907, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 525 + ], + "type": "text", + "content": "[33] Zheng Wang and Michael O'Boyle. Machine learning in compiler optimization. Proceedings of the IEEE, 106(11):1879-1901, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 296, + 569 + ], + "type": "text", + "content": "[34] Hugh Leather, Edwin Bonilla, and Michael O'boyle. Automatic feature generation for machine learning-based optimising compilation. ACM Transactions on Architecture and Code Optimization (TACO), 11(1):1-32, 2014. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 296, + 590 + ], + "type": "text", + "content": "[35] NVIDIA Corporation. NVIDIA TensorRT Developer Guide, 2024. Accessed: 2025-02-13. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 613 + ], + "type": "text", + "content": "[36] Amit Sabne. XLA: Compiling Machine Learning for Peak Performance, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 646 + ], + "type": "text", + "content": "[37] ONNX Community. Onnx: Open neural network exchange. https://github.com/onnx/onnx, 2024. Accessed: YYYYY-MM-DD. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 713 + ], + "type": "text", + "content": "[38] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "type": "text", + "content": "[39] Robert A Jacobs, Michael I Jordan, Steven J Nowlan, and Geoffrey E Hinton. Adaptive mixtures of local experts. *Neural computation*, 3(1):79–87, 1991. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 139 + ], + "type": "text", + "content": "[40] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10), 2018. 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 555, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 163 + ], + "type": "text", + "content": "[41] Juraj Hromkovič. Communication complexity and parallel computing. Springer Science & Business Media, 2013. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 164, + 555, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 555, + 208 + ], + "type": "text", + "content": "[42] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 555, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 275 + ], + "type": "text", + "content": "[43] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 276, + 555, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 309 + ], + "type": "text", + "content": "[44] Muhammad Aslam. Introducing kolmogorov-smirnov tests under uncertainty: an application to radioactive data. ACS omega, 5(1):914-917, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 310, + 555, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 555, + 375 + ], + "type": "text", + "content": "[45] Tianrui Guan, Jun Wang, Shiyi Lan, Rohan Chandra, Zuxuan Wu, Larry Davis, and Dinesh Manocha. M3detr: Multi-representation, multi-scale, mutual-relation 3d object detection with transformers. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 772-782, 2022. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 555, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 555, + 432 + ], + "type": "text", + "content": "[46] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Pointvoxel feature set abstraction for 3d object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10529–10538, 2020. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 477 + ], + "type": "text", + "content": "[47] Jiajia Liao, Yujun Liu, Yingchao Piao, Jinhe Su, Guorong Cai, and Yundong Wu. Gle-net: A global and local ensemble network for aerial object detection. International Journal of Computational Intelligence Systems, 15(1):2, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 478, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 555, + 533 + ], + "type": "text", + "content": "[48] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE transactions on pattern analysis and machine intelligence, 43(8):2647-2664, 2020. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 534, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 555, + 567 + ], + "type": "text", + "content": "[49] Hiep Anh Hoang and MyungsiK Yoo. 3onet: 3d detector for occluded object under obstructed conditions. IEEE Sensors Journal, 2023. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 568, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 555, + 633 + ], + "type": "text", + "content": "[50] Xin Li, Tao Ma, Yuenan Hou, Botian Shi, Yuchen Yang, Youquan Liu, Xingjiao Wu, Qin Chen, Yikang Li, Yu Qiao, et al. Logonet: Towards accurate 3d object detection with local-to-global cross-modal fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17524-17534, 2023. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 678 + ], + "type": "text", + "content": "[51] Yanwei Li, Xiaojuan Qi, Yukang Chen, Liwei Wang, Zeming Li, Jian Sun, and Jiaya Jia. Voxel field fusion for 3d object detection. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 555, + 713 + ], + "type": "text", + "content": "[52] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In Proceedings of the IEEE/CVF Conference on" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25912" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 327 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Computer Vision and Pattern Recognition, pages 908-917, 2022. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[53] Hai Wu, Chenglu Wen, Wei Li, Xin Li, Ruigang Yang, and Cheng Wang. Transformation-equivariant 3d object detection for autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 2795-2802, 2023. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[54] Nils Lubbe, Yi Wu, and Hanna Jeppsson. Safe speeds: fatality and injury risks of pedestrians, cyclists, motorcyclists, and car drivers impacting the front of another passenger car as a function of closing speed and age. Traffic safety research, 2:000006-000006, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[55] Yilun Chen, Zhiding Yu, Yukang Chen, Shiyi Lan, Anima Anandkumar, Jiaya Jia, and Jose M Alvarez. Focalformer3d: focusing on hard instance for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8394-8405, 2023. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 327 + ], + "type": "text", + "content": "[56] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L. Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 2774-2781, 2023. 7" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25913" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/be7e63af-2995-4db4-b164-1814191af90b_content_list.json b/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/be7e63af-2995-4db4-b164-1814191af90b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..01124f478c379c3f05e49ed584337d1d6944dcf5 --- /dev/null +++ b/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/be7e63af-2995-4db4-b164-1814191af90b_content_list.json @@ -0,0 +1,1827 @@ +[ + { + "type": "text", + "text": "Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kejia Zhang $^{1}$ , Juanjuan Weng $^{2*}$ , Shaozi Li $^{1}$ , Zhiming Luo $^{1,3}$", + "bbox": [ + 246, + 179, + 748, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Department of Artificial Intelligence, Xiamen University, Xiamen, China", + "bbox": [ + 277, + 200, + 720, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ College of Information Science and Technology, Jinan University, Guangzhou, China", + "bbox": [ + 241, + 217, + 754, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University", + "bbox": [ + 114, + 234, + 880, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kejiaz171@gmail.com, jjweng@jnu.edu.cn, {szlig,zhiming.luo}@xmu.edu.cn", + "bbox": [ + 184, + 253, + 803, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the remarkable progress of deep neural networks (DNNs) in various visual tasks, their vulnerability to adversarial examples raises significant security concerns. Recent adversarial training methods leverage inverse adversarial attacks to generate high-confidence examples, aiming to align adversarial distributions with high-confidence class regions. However, our investigation reveals that under inverse adversarial attacks, high-confidence outputs are influenced by biased feature activations, causing models to rely on background features that lack a causal relationship with the labels. This spurious correlation bias leads to overfitting irrelevant background features during adversarial training, thereby degrading the model's robust performance and generalization capabilities. To address this issue, we propose Debiased High-Confidence Adversarial Training (DHAT), a novel approach that aligns adversarial logits with debiased high-confidence logits and restores proper attention by enhancing foreground logit orthogonality. Extensive experiments demonstrate that DHAT achieves state-of-the-art robustness on both CIFAR and ImageNet1K benchmarks, while significantly improving generalization by mitigating the feature bias inherent in inverse adversarial training approaches. Code is available at https://github.com/KejiaZhang-Robust/DHAT.", + "bbox": [ + 86, + 335, + 485, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 728, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have demonstrated exceptional performance across a range of visual tasks [13, 25, 35, 48]. However, these networks remain susceptible to adversarial attacks, which present significant security challenges [11, 39]. Adversarial attacks introduce subtle perturbations that are imperceptible to humans but can significantly disrupt the inference process of DNNs, leading to incorrect predictions. Adversarial training (AT) is widely rec", + "bbox": [ + 89, + 753, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7a0dce767e07cf34d7d22817ca08fc8e2230eb59c6d073ca86ccfe60d02ddb0a.jpg", + "image_caption": [ + "(a)", + "Figure 1. Inverse adversarial attack exploits spurious correlations via attention misallocation to non-causal background features. The negative $\\epsilon$ denotes inverse adversarial perturbation strength. (a) Grad-CAM [32] visualizations show that adversarial attacks disrupt attention and cause errors, while inverse adversarial attacks amplify activation and improve accuracy but expose a critical flaw: attention shifts from foreground to irrelevant background regions. (b) Quantitative comparison of IoU(\\%) between attention regions and the true foreground and background regions using UIAT [9] and AT. Increasing inverse attack strength $-\\epsilon$ demonstrates UIAT's disproportionate focus on non-causal background features, exposing reliance on spurious correlations [3, 28, 33]." + ], + "image_footnote": [], + "bbox": [ + 516, + 301, + 676, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b1a5bc78731f31768cd4a98b8596db93676016165cb0860f3daa8fca165af52.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 674, + 303, + 903, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ognized as one of the most effective methods for defending against these attacks by incorporating adversarial examples into the training process [14, 34].", + "bbox": [ + 511, + 671, + 906, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, numerous advanced AT-based techniques have been proposed to address the vulnerabilities of DNNs [18, 23, 24, 38]. Prominent methods in this domain focus on aligning the distributions of adversarial examples to their true classes, thereby enhancing classification accuracy under adversarial attacks. For example, MART [41] and TRADES [49] aim to align the logits of adversarial examples with their corresponding natural examples, promoting consistency in predictions. Expanding on this, UIAT [9] and ACR [5] extend this approach by generating inverse adversarial examples [30] that exhibit higher confidence compared to their natural counterparts. This strategy helps al", + "bbox": [ + 509, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2783", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "leviate the negative impact of misclassification of natural examples in alignment and directs adversarial examples toward regions with correct high-confidence classifications.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Upon investigation, we uncover a remarkable yet concerning phenomenon: inverse adversarial examples produce high-confidence outputs by triggering biased feature activation. To illuminate this effect, we visualize attention maps from a conventional adversarial training (AT) model under both adversarial and inverse adversarial attacks (see Figure 1 (a)). Adversarial attacks disrupt normal attention patterns, causing incorrect predictions. In contrast, inverse adversarial attacks amplify the model's attention on features, with stronger attacks even improving prediction accuracy [9, 43]. However, this apparent enhancement masks a critical flaw: the model's attention systematically shifts from discriminative foreground features to irrelevant background regions—akin to identifying a sheep by detecting grass rather than the animal itself. This reliance on spurious correlations undermines the model's causal reasoning capabilities and compromises its robustness [1, 6, 19].", + "bbox": [ + 89, + 138, + 482, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To delve deeper into this issue, we conduct a statistical analysis using the Intersection over Union (IoU) metric to measure the alignment between attention maps and foreground/background regions on a subset of ImageNet [17]. Leveraging the Class Activation Mapping (CAM)-based technique inspired by [10], we define foreground as regions with high-class activation (i.e., areas critical for predictions) and background as regions with low activation. Comparing the UIAT model [9] (trained with inverse adversarial examples) to a standard AT model reveals a concerning trend. As evidenced in Figure 1 (b), while increasing the magnitude $\\epsilon$ of inverse adversarial attacks enhances overall feature attention, the UIAT model fails to improve IoU for foreground features compared to the AT model. Instead, it demonstrates a marked increase in IoU for background features—a clear indicator of attention misallocation.", + "bbox": [ + 89, + 397, + 482, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This leads to a pivotal insight: the inverse adversarial training disproportionately skews attention toward background features, resulting in biased feature activation. This shift engenders a fundamental vulnerability: a spurious correlation bias where the model becomes overly dependent on contextual background cues rather than intrinsic object features. Much like a classifier that associates beaches with birds rather than wings and beaks, this misplaced reliance significantly degrades robustness and generalization performance [15, 50], as conclusively demonstrated in Table 5.", + "bbox": [ + 89, + 642, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building upon the above findings, we introduce a method named Debiased High-Confidence Adversarial Training (DHAT). DHAT implements two key techniques: Debiased High-Confidence Logit Regularization (DHLR) and Foreground Logit Orthogonal Enhancement (FLOE). DHLR quantifies the bias towards background feature activation and debiase the spurious correlations by recalibrating the", + "bbox": [ + 89, + 795, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "biased high-confidence logits derived from inverse adversarial examples. This regularization aligns the logits of adversarial examples with debiased high-confidence logits to improve adversarial robustness and mitigate spurious correlation bias. FLOE further refines this process by reducing the correlation between high-confidence logits and background features in affine space, which helps restore the model's attention to its normal state. Extensive experiments demonstrate that DHAT outperforms state-of-the-art techniques across various datasets, providing superior adversarial robustness and generalization. Additionally, our DHAT can be seamlessly integrated with existing advanced adversarial training methods.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Below, we outline our principal contributions:", + "bbox": [ + 532, + 287, + 836, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify that the adversarial training model shows a biased feature activation under inverse adversarial attacks. Training with inverse adversarial examples causes the model's attention shifting towards background features, thus lead to spurious correlation bias.", + "- We introduce a novel method, Debiased High-Confidence Adversarial Training (DHAT). This method aligns the logits of adversarial examples with debiased high-confidence logits corresponding to their classes, effectively mitigating spurious correlation bias.", + "- Extensive experiments validate the superior robustness and generalization of our method against adversarial attacks. Additionally, our approach integrates smoothly with existing advanced adversarial training strategies." + ], + "bbox": [ + 511, + 304, + 903, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 531, + 653, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Adversarial Attack", + "text_level": 1, + "bbox": [ + 513, + 556, + 696, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial attacks introduce subtle perturbations $\\delta$ into natural examples $x$ , causing DNNs to produce incorrect predictions rather than the true class $y$ . These attacks are typically formulated as an optimization problem:", + "bbox": [ + 511, + 579, + 903, + 640 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} = \\max _ {| | \\delta | | _ {p} \\leq \\epsilon} \\left(\\mathcal {L} _ {\\text {a t t a c k}} (\\theta ; x + \\delta , y)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 652, + 903, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\hat{x}$ represents the adversarial example, $\\mathcal{L}_{attack}$ is the adversarial attack loss function, $\\theta$ denotes the model parameters, and $\\delta$ is constrained to the $p$ -norm ball of radius $\\epsilon$ .", + "bbox": [ + 511, + 690, + 903, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Numerous studies have investigated the vulnerability of DNNs and have proposed various adversarial attack methods. Goodfellow et al. [14] introduced the Fast Gradient Sign Method (FGSM), which compute the adversarial perturbation in the direction of the gradient:", + "bbox": [ + 511, + 736, + 903, + 811 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\delta = \\eta \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} \\mathcal {L} _ {\\text {a t t a c k}} (\\theta ; x, y)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 825, + 903, + 842 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\eta$ denotes the attack step size. Building on this, Madry et al. [27] proposed Projected Gradient Descent (PGD), which iteratively updates the gradient of the input", + "bbox": [ + 511, + 854, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2784", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "data to generate more potent adversarial examples, thereby increasing the success rate of attacks:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} ^ {t + 1} = \\Pi_ {\\mathbb {B} (\\epsilon)} \\{\\hat {x} ^ {t} + \\eta \\cdot \\operatorname {s i g n} (\\nabla_ {\\hat {x} ^ {t}} \\mathcal {L} _ {\\text {a t t a c k}} (\\theta ; \\hat {x} ^ {t}, y)) \\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 127, + 482, + 147 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Pi_{\\mathbb{B}(\\epsilon)}$ projects the perturbation onto the $p$ -norm ball of radius $\\epsilon$ . Furthermore, Carlini and Wagner [4] developed the Carlini-Wagner (C&W) attack, an optimization-based method capable of producing more imperceptible adversarial examples. Croce and Hein [7] introduced AutoAttack, an ensemble attack strategy that is user-independent and parameter-free.", + "bbox": [ + 89, + 152, + 483, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Adversarial Defense", + "text_level": 1, + "bbox": [ + 89, + 266, + 282, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial training (AT) is a widely used strategy for improving model robustness. The core principle of AT is to include adversarial examples in the training process, thereby aligning the model's outputs towards the true label distribution [12, 45]. Mathematically, conventional AT methods are formulated as a min-max optimization problem:", + "bbox": [ + 89, + 287, + 483, + 378 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} \\mathbb {E} _ {(x, y) \\sim \\mathcal {D}} \\max _ {| | \\delta | | _ {p} \\leq \\epsilon} \\mathcal {L} _ {A T} (\\theta ; \\mathcal {G} _ {\\theta} (x, \\delta), y), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 385, + 482, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $(x,y)$ denotes a clean image-label pair sampled from the data distribution $\\mathcal{D}$ , $\\delta$ is constrained to the maximum $p$ -norm magnitude $\\epsilon$ , $\\mathcal{L}_{AT}$ represents the training strategy function, and $\\mathcal{G}_{\\theta}(x,\\delta)$ denotes the attack generation procedure incorporating the perturbation $\\delta$ . The outer minimization optimizes the model parameters $\\theta$ to minimize the expected loss over adversarial examples, while the inner maximization identifies the perturbations that maximize this loss, thereby enhancing the model's robustness.", + "bbox": [ + 89, + 417, + 483, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent advancements in AT have introduced several techniques to further improve adversarial robustness. Wu et al. [44] introduced Adversarial Weight Perturbation (AWP), which enhances robustness by flattening the loss landscape with respect to input variations. Kim et al. [20] proposed Feature Separation and Recalibration (FSR) to capture useful logits from non-robust features. Wei et al. [42] developed Class-wise Calibrated Fair Adversarial Training (CFA), which adaptively configures training for each class to achieve category-level robust fairness. Li et al. [26] proposed SGLR, which uses self-distillation to refine the soft-label distribution logits, thereby calibrating the adversarial training.", + "bbox": [ + 89, + 553, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Methods similar to our approach include MART [41] and UIAT [9]. These methods use Kullback-Leibler (KL) divergence to align the logits of adversarial examples with high-confidence logits. MART encourages sharing a similar logit distribution between adversarial examples and their corresponding natural examples. UIAT generates inverse adversarial examples [30] that exhibit higher confidence compared to natural examples, thereby enhancing the alignment of adversarial examples towards the high-confidence regions of the decision surface.", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/71b56ad4483a89b149d57fd64cd126af45e34fc490f06b0eb7c868b3f9f88469.jpg", + "image_caption": [ + "Figure 2. Overview of Debiased High-Confidence Logits Regularization (DHLR). The method first quantifies the degree of biased activation towards background features and recalibrates biased logits, then aligns the logits of adversarial examples with biased high-confidence logits." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 906, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 351, + 650, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the basic notation and present our approach to mitigate adversarial bias and improve robustness. Our method recalibrates high-confidence logits and reduces reliance on spurious background features. The optimization strategy is outlined, with pseudocode provided in Section 7 of the supplementary material.", + "bbox": [ + 511, + 377, + 906, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Notation", + "text_level": 1, + "bbox": [ + 513, + 478, + 617, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We denote the training dataset as $(X,Y) = \\{(x_{i},y_{i})|i\\in (1,2,\\ldots ,N)\\}$ , where $x$ and $y$ denote the natural input examples and their corresponding labels, respectively. $N$ represents the total number of examples in the dataset. Let $\\theta$ denote the parameters of a convolutional neural network (CNN) for image classification. The network's output logits are denoted as $f_{\\theta}(\\cdot)$ , and the final predicted label is given by $\\arg \\max f_{\\theta}(\\cdot)$ . Inverse adversarial examples $\\tilde{x}$ can be regarded as examples for which the model correctly predicts the label with high-confidence. These examples are generated by applying image perturbations aimed at minimizing the attack loss [30]:", + "bbox": [ + 511, + 500, + 906, + 681 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\check {x} = \\min _ {| | \\delta | | _ {p} \\leq \\epsilon} \\left(\\mathcal {L} _ {\\mathrm {I n v}} (\\theta ; x - \\delta , y)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 694, + 906, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{Inv}}$ denotes the loss function used to generate inverse adversarial attack. The logits of the inverse adversarial example $\\tilde{x}$ and the adversarial example $\\hat{x}$ are computed as $\\tilde{z} = f_{\\theta}(\\tilde{x})$ and $\\hat{z} = f_{\\theta}(\\hat{x})$ , respectively.", + "bbox": [ + 511, + 731, + 906, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Debiased High-Confidence Logit Regularization", + "text_level": 1, + "bbox": [ + 511, + 801, + 905, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To mitigate the spurious correlation bias introduced by adversarial training with inverse adversarial examples, we propose utilizing debiased high-confidence logits for regularization. Specifically, this approach involves separating and", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2785", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "quantifying the spurious bias towards background features under inverse adversarial attacks, and then recalibrating the biased high-confidence logits to align with those of adversarial examples, as illustrated in Figure 2.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we quantify the model's biased activation degree towards background features under inverse adversarial attacks. Following the approach proposed by [10], we use using Grad-CAM [31] or alternative methods such as SAM [21] to generate pixel-wise attention maps $M \\in [0,1]^{n \\times n}$ for the natural example $x$ . We separate the background feature $\\tilde{x}_{(B)}$ from the inverse adversarial examples $\\tilde{x}$ based on the $M$ :", + "bbox": [ + 89, + 152, + 482, + 271 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n[ \\check {x} _ {(B)} ] _ {(i, j)} = \\mathbb {I} _ {\\left(M _ {i, j} < \\omega\\right)} \\cdot \\check {x} _ {(i, j)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 277, + 482, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{I}_{(M_{i,j} < \\omega)}$ is an indicator function and $\\omega$ is a predefined threshold. Next, we compute the logits for the background feature $\\check{x}_{(B)}$ :", + "bbox": [ + 89, + 301, + 480, + 347 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\check {z} _ {(B)} = f _ {\\theta} \\left(\\check {x} _ {(B)}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 227, + 353, + 482, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These logits can quantify the biased degree introduced by background feature activation during network inference.", + "bbox": [ + 89, + 375, + 480, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To recalibrate the biased high-confidence logits, we refine the logits of inverse adversarial examples by subtracting the background feature logits:", + "bbox": [ + 89, + 406, + 480, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\check {z} ^ {*} = \\check {z} - \\check {z} _ {(B)}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 458, + 482, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\check{z}^*$ represents the debiased high-confidence logits.", + "bbox": [ + 89, + 479, + 464, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we introduce a regularizer to align the logits of adversarial examples with these debiased high-confidence logits. This regularization encourages the model to focus on foreground features and mitigates the biased activation of background features. Specifically, our Debiased High-Confidence Logit Regularization (DHLR) is defined as:", + "bbox": [ + 89, + 494, + 482, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D H L R} \\left(\\check {z} ^ {*}, \\hat {z}\\right) = \\mathcal {L} _ {K L} \\left(\\phi \\left(\\check {z} ^ {*}\\right) | | \\phi (\\hat {z})\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 592, + 480, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi$ denotes the softmax function, $\\mathcal{L}_{KL}$ represents the Kullback-Leibler (KL) divergence.", + "bbox": [ + 89, + 614, + 482, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This regularization term $\\mathcal{L}_{DHLR}$ promotes consistency between the logits of adversarial examples $\\hat{z}$ and the debiased high-confidence logits $\\tilde{z}^*$ . This alignment mitigates spurious correlation bias, thereby improving adversarial robustness and generalization.", + "bbox": [ + 89, + 645, + 482, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Foreground Logit Orthogonal Enhancement", + "text_level": 1, + "bbox": [ + 89, + 727, + 467, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While DHLR effectively recalibrates high-confidence logits to mitigate spurious correlations, it fails to fully restore the model's focus on foreground features when handling inverse adversarial examples $\\check{x}$ . This is because DHLR primarily aligns logits with debiased high-confidence targets, without directly addressing the model's persistent bias toward background features. Under inverse adversarial attacks, the model still exhibits strong activation in irrelevant background regions, undermining its ability to prioritize critical foreground cues.", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, we introduce Foreground Logit Orthogonal Enhancement (FLOE), a targeted method designed to minimize the correlation between high-confidence logits $\\tilde{z}$ and background feature logits $\\tilde{z}_{(B)}$ . The intuition behind FLOE is straightforward: by reducing the projection of $\\tilde{z}$ onto $\\tilde{z}_{(B)}$ in the affine space, we make $\\tilde{z}$ less explainable by $\\tilde{z}_{(B)}$ , thereby encouraging the model to rely more on foreground features. This projection-based approach ensures that the model's predictions are driven by meaningful foreground information rather than spurious background correlations. The optimization objective is formulated as:", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {F L O E} \\left(\\check {z}, \\check {z} _ {(B)}\\right) = - \\left| \\check {z} - \\frac {\\check {z} \\cdot \\check {z} _ {(B)}}{\\left| \\check {z} _ {(B)} \\right| ^ {2}} \\cdot \\check {z} _ {(B)} \\right| _ {p}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 267, + 903, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $p$ indicates the norm exponent. The term $\\mathcal{L}_{FLOE}$ is designed to alleviate biased feature activation under inverse adversarial attacks.", + "bbox": [ + 511, + 311, + 903, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Optimization Strategy", + "text_level": 1, + "bbox": [ + 511, + 367, + 720, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we integrate the DHLR and FLOE techniques into a unified training framework. Debiased High-Confidence Adversarial Training (DHAT). The overall optimization objective for DHAT is formulated as:", + "bbox": [ + 511, + 388, + 903, + 448 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {D H A T} = \\mathcal {L} _ {A T} (\\hat {z}, y) + \\lambda_ {1} \\cdot \\mathcal {L} _ {D H L R} (\\hat {z} ^ {*}, \\hat {z}) \\tag {11} \\\\ + \\lambda_ {2} \\cdot \\mathcal {L} _ {F L O E} (\\tilde {z}, \\tilde {z} _ {(B)}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 458, + 903, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{AT}(\\hat{z},y)$ represents the standard adversarial training loss, and $\\lambda_{1}$ and $\\lambda_{2}$ are hyperparameters that control the relative contributions of the DHLR and FLOE terms, respectively. Besides, our DHAT can be seamlessly integrated with various advanced adversarial training strategies represented by $\\mathcal{L}_{AT}$ to improve the robustness.", + "bbox": [ + 511, + 506, + 903, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 611, + 645, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 636, + 720, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1. Datasets and models.", + "text_level": 1, + "bbox": [ + 511, + 657, + 705, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conduct experiments on the CIFAR-10, CIFAR-100 [22], and ImageNet-1K [29] datasets. The architecture for our experiments include the mainstream models WRN28-10 [47], ResNet-18 [16], ResNet-50 [16], Inception-V3 [40], and VGG16 [36]. Detailed information is provided in the supplementary material.", + "bbox": [ + 511, + 676, + 903, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.2. Training setup", + "text_level": 1, + "bbox": [ + 511, + 776, + 661, + 791 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For CIFAR datasets, each model is trained for 100 epochs, with the learning rate reduced by a factor of 0.1 at the 80th and 90th epochs. For adversarial example generation, we use a fixed 10-iteration attack. The maximum $\\ell_{\\infty}$ -norm of the adversarial perturbation is set to $\\epsilon = 8 / 255$ , while the inverse adversarial perturbation is set to $\\epsilon = 4 / 255$ . Both perturbations use a fixed step size of $\\alpha = 2 / 255$ . For the", + "bbox": [ + 511, + 794, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2786", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/bf755635fc4e570063477661b011d38d7e8d76fcbdb2a9860a0e54829edf8915.jpg", + "table_caption": [ + "Table 1. Comparison of robustness (%) and robust generalization gap (%) . The number in bold indicates the best performance." + ], + "table_footnote": [], + "table_body": "
Architecture: WRN28-10\nImageNet-1KAttack (ε = 4/255)Robust Gap↓
Clean↑PGD-10↑PGD-20↑PGD-50↑C&W↑AA↑
MART [41]62.3143.5642.8942.5440.6238.9413.76
AWP [44]64.2545.1344.6744.2141.9640.0212.82
FSR [20]64.7044.3143.8243.4541.2339.3014.17
CFA [42]65.3844.8744.3243.9641.5839.6515.49
UIAT [9]62.6445.2944.8544.5742.1340.1814.68
SGLR [26]63.4644.5244.0543.7841.3639.4719.72
DHAT (Ours)65.9046.8346.4246.1143.2541.709.53
DHAT-AWP (Ours)65.6247.3847.0446.7943.6242.0710.27
DHAT-CFA (Ours)66.2648.2747.8947.5344.1842.4511.64
Architecture: WRN28-10\nCIFAR-10Attack (ε = 8/255)Robust Gap↓
Clean↑PGD-10↑PGD-20↑PGD-50↑C&W↑AA↑
MART [41]82.9956.2555.4855.4552.2650.679.52
AWP [44]82.6757.8057.2157.0754.8251.906.90
FSR [20]82.9256.6955.9455.5153.9351.747.42
CFA [42]84.4357.8756.9056.6454.6051.8510.36
UIAT [9]82.9458.6658.1258.0554.1152.177.92
SGLR [26]85.7657.5356.9156.6654.2852.079.38
DHAT (Ours)83.9560.4959.9559.8755.2753.103.51
DHAT-AWP (Ours)83.2161.6161.3561.2755.7153.694.54
DHAT-CFA (Ours)84.4962.6762.3862.2255.9554.056.33
Architecture: WRN28-10\nCIFAR-100Attack (ε = 8/255)Robust Gap↓
Clean↑PGD-10↑PGD-20↑PGD-50↑C&W↑AA↑
MART [41]54.6932.0631.9031.8828.7727.259.96
AWP [44]57.9434.0133.7533.7230.7428.907.87
FSR [20]57.4832.9332.3032.2629.1627.047.84
CFA [42]60.9233.1032.5632.4130.4928.0410.47
UIAT [9]57.6534.2733.9133.8530.9729.0311.70
SGLR [26]61.0233.4332.9832.8230.7228.5015.67
DHAT (Ours)59.1435.8235.3335.0231.7230.174.24
DHAT-AWP (Ours)59.3036.3936.1535.9832.0530.384.85
DHAT-CFA (Ours)61.5437.6737.1536.9932.4030.935.93
", + "bbox": [ + 151, + 114, + 836, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ImageNet dataset, we adopt the approach in [37] with 2-iteration PGD over 50 epochs, and set $\\epsilon = 4 / 255$ and $\\alpha = 1 / 255$ for adversarial perturbations. The hyperparameters $\\lambda_{1}$ and $\\lambda_{2}$ are set to 1.0 in all data sets. Comprehensive training details are available in the supplementary material.", + "bbox": [ + 89, + 660, + 480, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.3. Evaluation setup.", + "text_level": 1, + "bbox": [ + 89, + 744, + 256, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For robustness performance evaluation, we employ PGD [27], C&W [4], and AutoAttack (AA) [7] attacks under the $\\ell_{\\infty}$ norm. AutoAttack comprises APGD-DLR [7], APGD-CE [7], FAB [8], and Square [2]. \"Clean\" denotes natural examples without adversarial perturbations. Additionally, we introduce the concept of the robust generalization gap, referred to as the \"Robust Gap\", which measures the difference in robustness performance between the training and test sets as a direct indicator of the generalization", + "bbox": [ + 88, + 763, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ability [46]. This metric quantifies the strength of spurious correlations, where larger gaps indicate greater reliance on features that lack causal relationships with target labels.", + "bbox": [ + 511, + 660, + 905, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Experimental Result", + "text_level": 1, + "bbox": [ + 511, + 715, + 707, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this part, we conducted comprehensive evaluations of our proposed method across various visual datasets and compared it with mainstream adversarial training (AT) methods. Additionally, DHAT can be integrated with advanced AT strategies. The results are shown in Table 1.", + "bbox": [ + 511, + 738, + 905, + 813 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Robust performance.", + "text_level": 1, + "bbox": [ + 511, + 821, + 705, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DHAT consistently demonstrates superior robustness across all datasets. For example, on CIFAR-10, DHAT outperforms the second-best method by $1.93\\%$ and $1.16\\%$ under PGD-10 and C&W attacks, respectively. DHAT also", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2787", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/05098f207c7da2f937fa014021b4462ade9ec239115cc946f55af6d99cae8cb3.jpg", + "table_caption": [ + "Table 2. Comparison of robustness (%) and robust generalization gap (%). The number in bold indicates the best performance." + ], + "table_footnote": [], + "table_body": "
Dataset: CIFAR-10\nMethodResNet-50VGG-16Inception-V3
Clean↑AA↑Robust Gap↓Clean↑AA↑Robust Gap↓Clean↑AA↑Robust Gap↓
MART [41]75.0549.054.5766.7244.273.6076.0549.595.91
AWP [44]75.5950.803.9167.1445.973.4776.7450.464.07
FSR [20]78.3850.593.9769.5944.903.3077.2850.024.73
CFA [42]77.8850.644.2770.4545.534.6676.7049.974.46
UIAT [9]78.4251.002.9968.4645.273.5775.1951.233.70
SGLR [26]80.3950.555.2170.9045.043.9480.6250.735.00
DHAT (Ours)79.9451.921.8570.0146.401.3477.8551.811.92
DHAT-AWP (Ours)76.8452.052.3467.4147.212.1877.1252.082.88
DHAT-CFA (Ours)80.9752.382.0970.5247.832.2178.5252.672.44
", + "bbox": [ + 116, + 114, + 883, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shows the best performance on more complex datasets such as CIFAR-100 and high-resolution Imagenette datasets. Specifically, on ImageNet-1K, DHAT achieves improvements of $1.54\\%$ and $1.52\\%$ over the second-best method under PGD-10 and AA attacks, respectively.", + "bbox": [ + 89, + 315, + 483, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Combine with other AT methods.", + "text_level": 1, + "bbox": [ + 89, + 407, + 369, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DHAT can be effectively combined with various AT methods to further enhance adversarial performance. Notably, DHAT can be combined with AWP and CFA, denoted as DHAT-AWP and DHAT-CFA, respectively. However, DHAT is not suitable for combining with methods that rely on spurious correlation to improve performance (e.g., FSR, SLGR). In particular, DHAT-CFA achieves superior performance across all datasets. For instance, on Imagenette, DHAT-CFA provides a $2.84\\%$ performance improvement under the AA attack compared to the second-best method. Moreover, combining DHAT with advanced AT surpasses the performance of the original baseline methods in terms of both robustness and clean accuracy.", + "bbox": [ + 88, + 428, + 483, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3. Robust overfitting.", + "text_level": 1, + "bbox": [ + 89, + 640, + 266, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method significantly reduces the robust generalization gap across all evaluated datasets, effectively mitigating the robust overfitting phenomenon. For example, on CIFAR-10, DHAT reduces the robust generalization gap by $4.41\\%$ and $6.01\\%$ compared to UIAT and MART, respectively. This reduction highlights DHAT's ability to mitigate spurious correlation bias, thereby improving the model's overall generalization performance. Detailed visualizations of the learning curve are provided in the supplementary material.", + "bbox": [ + 89, + 661, + 483, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Generalization Analysis", + "text_level": 1, + "bbox": [ + 89, + 815, + 310, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this part, we evaluate the generalizability of our proposed method by comprehensively evaluating its performance across various network architectures and under different magnitudes of adversarial attacks.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/916c0ac27ef8715d57b26358cafee580c3f236249463b5ac8a16e47dd6b30019.jpg", + "image_caption": [ + "(a) PGD-10 Attack", + "Figure 3. Comparisons of robustness $(\\%)$ under varying $\\epsilon$ adversarial attacks using ResNet-18 on CIFAR-10." + ], + "image_footnote": [], + "bbox": [ + 519, + 314, + 710, + 459 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71e8fa34bfc78b2ee8bcc7f3b6350ccd91cd8b538282b4c34f307a7223ca6975.jpg", + "image_caption": [ + "(b) C&W Attack" + ], + "image_footnote": [], + "bbox": [ + 712, + 315, + 900, + 459 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1. Performance with diverse network architecture.", + "text_level": 1, + "bbox": [ + 511, + 541, + 890, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the adaptability of our method across different network architectures, we applied DHAT on ResNet-50, VGG-16, and Inception-V3. The performance comparisons are reported in Table 2. The results demonstrate the DHAT's consistently superior robustness and generalization across a variety of model architectures. For example, DHAT-CFA outperforms the second-best method by $1.83\\%$ under AA attacks on ResNet-50. Similarly, on VGG-16, DHAT significantly reduces the robustness generalization gap by $1.96\\%$ compared to the second-best method.", + "bbox": [ + 511, + 561, + 906, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2. Different attack magnitude.", + "text_level": 1, + "bbox": [ + 511, + 728, + 751, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 illustrates the performance of DHAT under varying magnitudes of PGD-10 and C&W attacks. Our findings reveal that DHAT maintains a significant performance advantage across attack magnitudes. Compared to other methods, DHAT exhibits improved stability and robustness, with minimal performance degradation even under more severe attacks. This consistent performance across varying attack levels underscores the effectiveness of our approach in mitigating the impact of adversarial perturbations. Further details are provided in the supplementary material (Section 7).", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2788", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6c28bb72f6d280c1f288e8122719debdae8c0444a1b0fac1a67ed631ea78cdda.jpg", + "table_caption": [ + "Table 3. Comparison of computational costs (# A100 GPU hours) using different adversarial training methods on CIFAR-10." + ], + "table_footnote": [], + "table_body": "
MethodArchitecture
ResNet-18WRN28-10Inception-V3
MART [41]0.0500.2850.331
AWP [44]0.0480.2740.318
FSR [20]0.0510.2910.337
CFA [42]0.0450.2620.303
UIAT [9]0.0400.2450.283
SGLR [26]0.0390.2400.278
DHAT (Ours)0.0410.2480.287
", + "bbox": [ + 114, + 127, + 464, + 266 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/77138da2ef72adfa790ac57d5975efc6bfa752a44b1b5d56da1ac0122413332a.jpg", + "table_caption": [ + "Table 4. Comparison of robustness (\\%) and robust generalization gap (\\%). The number in bold indicates the best performance." + ], + "table_footnote": [], + "table_body": "
DHLRFLOEClean↑PGD-10↑AA↑Robust Gap↓
78.0355.0348.353.84
✓78.7956.5649.471.98
✓78.5056.2249.732.06
✓✓79.2857.2150.250.81
", + "bbox": [ + 109, + 337, + 457, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Computational Efficiency", + "text_level": 1, + "bbox": [ + 89, + 441, + 323, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The computational costs of our method and baseline models are provided in terms of per-epoch training time on a single NVIDIA A100 GPU with a batch size of 128 on the CIFAR-10 dataset, as shown in Table 3. Our method demonstrates significantly improved time efficiency compared to the baseline MART, AWP, FSR, and CFA. While it requires slightly more computational resources than the state-of-the-art methods UIAT and SGLR, this minor increase is justifiable given the performance gains achieved.", + "bbox": [ + 88, + 467, + 482, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Analysis", + "text_level": 1, + "bbox": [ + 89, + 623, + 264, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this part, we conduct a comprehensive analysis of the impact of individual components and parameters of our model.", + "bbox": [ + 89, + 648, + 482, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.1. Impact of individual components.", + "text_level": 1, + "bbox": [ + 89, + 696, + 370, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the contributions of the individual components of DHAT, specifically Debiased High-Confidence Logit Regularization (DHLR) and Foreground Logit Orthogonal Enhancement (FLOE), as shown in Table 4. The results indicate that both DHLR and FLOE significantly enhance model performance compared to the UIAT method. They enhance clean accuracy and adversarial robustness, while also reducing the gap in robustness generalization. This highlights the effectiveness of DHLR and FLOE in mitigating spurious correlation biases. Notably, the integration of both DHLR and FLOE results in a more substantial overall improvement in model performance.", + "bbox": [ + 88, + 719, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cac6452ce8fdd4fdfd37b168ff8b527a222f4fb60f3508e21988f18abf63416f.jpg", + "image_caption": [ + "(a) $\\mathcal{L}_{DHLR}$" + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 700, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e71e2437acc5f51a0dc6dbdf3b9e7c7dd3a02eef436a73ea4f70e790fb2ed842.jpg", + "image_caption": [ + "(b) $\\mathcal{L}_{FLOE}$" + ], + "image_footnote": [], + "bbox": [ + 681, + 89, + 898, + 207 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c3a4d2a24a52fb478104250c76b64b0327766b92ebb43a60af45c8eaf64bb7a.jpg", + "image_caption": [ + "Figure 4. Analysis of hyper-parameters $\\lambda_{1}$ and $\\lambda_{2}$ using ResNet-18 on CIFAR-10. The robust accuracy denotes the accuracy under PGD-10 attack.", + "(a) CIFAR-10", + "Figure 5. Analysis of hyper-parameters $\\omega$ using ResNet-18 on CIFAR-10 and Imagenette. The robust accuracy denotes the accuracy under PGD-10 attack." + ], + "image_footnote": [], + "bbox": [ + 517, + 290, + 702, + 407 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85e08e3ef216b6ab5b2d70937c646b3d911ceb622e51576184fd1d65acf419f6.jpg", + "image_caption": [ + "(b) Imagenette" + ], + "image_footnote": [], + "bbox": [ + 705, + 290, + 898, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.2. Analysis of parameter sensitivity.", + "text_level": 1, + "bbox": [ + 511, + 505, + 790, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 illustrates the sensitivity of the optimization parameters $\\lambda_{1}$ and $\\lambda_{2}$ in DHAT. Our analysis reveals that these parameters influence the trade-off between adversarial robustness and clean accuracy. As $\\lambda_{1}$ increases, adversarial robustness improves, while clean accuracy tends to decrease. This occurs because higher values of $\\lambda_{1}$ increase the regularization strength, which enhances the alignment of adversarial example logits with debiased high-confidence logits. However, this stronger regularization effect may also lead to a reduced emphasis on clean accuracy, as the model focuses more on handling adversarial perturbations. Conversely, increasing $\\lambda_{2}$ generally improves clean accuracy but may adversely affect adversarial robustness. Larger values of $\\lambda_{2}$ help the model restore its attention to the normal state, which can result in less effective alignment of adversarial example logits with the debiased high-confidence logits.", + "bbox": [ + 511, + 523, + 906, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.3. Analysis of threshold factor for quantifying spurious bias.", + "text_level": 1, + "bbox": [ + 511, + 790, + 903, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 illustrates the impact of threshold values $w$ on datasets with varying resolutions. Our analysis reveals that as the threshold value increases, both clean accuracy and robustness initially improve before sharply declining. This initial improvement occurs because higher thresholds help", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2789", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e29c250f1f3638d19b9bca044e9c10b2fa46a1ebe54658e4c0bbc6a941a929a.jpg", + "image_caption": [ + "Figure 6. Visualization of feature activation maps using UIAT and DHAT, utilizing Grad-CAM under inverse attack $(\\epsilon = -8)$ and adversarial attack $(\\epsilon = 8)$ . Left-to-right: the original image; normal feature activation maps; feature activation maps under adversarial attacks for UIAT and DHAT; and feature activation maps under inverse adversarial attacks for UIAT and DHAT.", + "Figure 7. IoU of attention regions with normal foreground and background regions. The comparison is conducted among conventional AT, UIAT, and DHAT on the Imagenette Dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in better separating accurate background features. However, further increasing the threshold eventually leads to the extraction of more foreground features, which causes significant performance degradation. For low-resolution datasets, the optimal threshold tends to be higher compared to high-resolution datasets, as high-resolution images more readily separate background features.", + "bbox": [ + 89, + 468, + 482, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Discussion DHAT Effectiveness", + "text_level": 1, + "bbox": [ + 89, + 584, + 369, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this part, we analyze the mechanisms underlying the effectiveness of our proposed method, Debiased High-Confidence Adversarial Training (DHAT), with a focus on its capacity to mitigate spurious correlation bias.", + "bbox": [ + 89, + 606, + 483, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6.1. Feature Activation Map Analysis", + "text_level": 1, + "bbox": [ + 89, + 674, + 366, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 6 visualizes feature activation maps of UIAT and DHAT models under adversarial and inverse adversarial attacks. DHAT successfully maintains normal attention patterns, while UIAT exhibits biased activation with attention shifting significantly to background features during inverse attacks. DHAT achieves this robust performance by recalibrating biased high-confidence logits and systematically reducing correlation between logits and background features.", + "bbox": [ + 89, + 691, + 482, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6.2. Quantitative Evaluation of Bias Mitigation", + "text_level": 1, + "bbox": [ + 89, + 821, + 434, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 7 quantifies DHAT's effectiveness through IoU between attention maps and true foreground/background regions. DHAT achieves higher foreground IoU and lower background IoU than both AT and UIAT, confirming its", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/33666b8d58ca2952974707b1c0b62d3e93e4dbe803fc48724e26b48ae77ba0dc.jpg", + "image_caption": [ + "(a) IoU with Normal Foreground" + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 709, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4472d173a4ea941042d640635544b0b670d3ce8f63732b585fc254e685f001f1.jpg", + "image_caption": [ + "(b) IoU with Normal Background" + ], + "image_footnote": [], + "bbox": [ + 707, + 89, + 897, + 220 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/675e96b1d8c4350db1f0881f643ee88f03fead073c27da82ba69ac2992dec787.jpg", + "table_caption": [ + "Table 5. Comparison of training with different magnitude reverse adversarial samples (UIAT) and our debiased training DHAT using WRN28-10 on the CIFAR-10." + ], + "table_footnote": [], + "table_body": "
Methodε of zTraining SetTest SetRobust Gap↓
PGD-10↑AA↑PGD-10↑AA↑
UIAT-490.5361.4582.9452.079.38
UIAT-892.5762.8782.2651.3311.54
UIAT-1294.9065.2881.7250.8514.43
DHAT (ours)-488.7356.6183.9553.103.51
", + "bbox": [ + 517, + 359, + 901, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "superior feature attention. Table 5 shows stronger inverse attacks in UIAT worsen overfitting, increasing the robust gap from 9.38 to 14.43. In contrast, DHAT achieves better test robustness (83.95% PGD-10, 53.10% AA) with a significantly smaller robust gap (3.51), effectively mitigating overfitting while maintaining strong performance.", + "bbox": [ + 511, + 467, + 905, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 573, + 633, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we identify a critical issue in adversarial training: AT model exhibits biased feature activation under inverse adversarial attacks. Specifically, training with inverse adversarial examples causes the model attention to shift from the foreground to background features, resulting in spurious correlation bias. To tackle this challenge, we propose a novel method called Debiased High-Confidence Adversarial Training (DHAT). DHAT incorporates two pivotal techniques: Debiased High-Confidence Logit Regularization (DHLR) and Foreground Logit Orthogonal Enhancement (FLOE). DHLR quantifies the bias towards background feature activation and recalibrates biased high-confidence logits to align the logits of adversarial examples with these debiased high-confidence logits. FLOE further restores the model's attention to its normal state by reducing the correlation between high-confidence logits and background features in affine space. This work highlights DHAT's effectiveness in enhancing model resilience and presents an adversarial training framework that integrates seamlessly with advanced training strategies.", + "bbox": [ + 511, + 598, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2790", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China under Grant No. 62276221, 62376232; the Fundamental Research Funds for the Central Universities under Grant No. 21625344; Key Laboratory of Equipment Data Security and Guarantee Technology, Ministry of Education under Grant No. 2024020200.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 219, + 186, + 234 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ola Ahmad, Nicolas Béreux, Loïc Baret, Vahid Hashemi, and Freddy Lecue. Causal analysis for robust interpretability of neural networks. In WAVC, pages 4685-4694, 2024. 2", + "[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In ECCV, 2020. 5", + "[3] Saeid Asgari, Aliasghar Khani, Fereshte Khani, Ali Gholami, Linh Tran, Ali Mahdavi Amiri, and Ghassan Hamarneh. Masktune: Mitigating spurious correlations by forcing to explore. In NeurIPS, 2022. 1", + "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. IEEE Symp. Security Privacy, pages 39-57, 2017. 3, 5", + "[5] Hyuna Cho, Yubin Han, and Won Hwa Kim. Antiadversarial consistency regularization for data augmentation: Applications to robust medical image segmentation. In MI-CAI, 2023. 1", + "[6] Jonathan Crabbé and Mihaela van der Schaar. Evaluating the robustness of interpretability methods through explanation invariance and equivariance. In NeurIPS, 2023. 2", + "[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 3, 5", + "[8] Francesco Croce and Matthias Hein. Minimally distorted adversarial examples with a fast adaptive boundary attack. In ICML, 2020. 5", + "[9] Junhao Dong, Seyed-Mohsen Moosavi-Dezfooli, Jianhuang Lai, and Xiaohua Xie. The enemy of my enemy is my friend: Exploring inverse adversaries for improving aversarial training. In CVPR, 2023. 1, 2, 3, 5, 6, 7", + "[10] Xiaoyi Dong, Jiangfan Han, Dongdong Chen, Jiayang Liu, Huanyu Bian, Zehua Ma, Hongsheng Li, Xiaogang Wang, Weiming Zhang, and Nenghai Yu. Robust superpixel-guided attentional adversarial attack. In CVPR, 2020. 2, 4", + "[11] Alhussein Fawzi, Hamza Fawzi, and Omar Fawzi. Adversarial vulnerability for any classifier. In NeurIPS, 2018. 1", + "[12] Liam Fowl, Micah Goldblum, Ping-yeh Chiang, Jonas Geiping, Wojciech Czaja, and Tom Goldstein. Adversarial examples make strong poisons. In NeurIPS, 2021. 3", + "[13] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. Avsegformer: Audio-visual segmentation with transformer. In AAAI, 2024. 1", + "[14] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 2" + ], + "bbox": [ + 93, + 244, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Xinzhe Han, Shuhui Wang, Chi Su, Qingming Huang, and Qi Tian. General greedy de-bias learning. IEEE TPAMI, 45 (8):9789-9805, 2023. 2", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4", + "[17] FastAI Jeremy Howard. Imagenette: A smaller subset of imagenet. In *GibHub*. https://github.com/fastai/imagenette, 2019.2", + "[18] Zhichao Huang, Yanbo Fan, Chen Liu, Weizhong Zhang, Yong Zhang, Mathieu Salzmann, Sabine Susstrunk, and Jue Wang. Fast adversarial training with adaptive step size. IEEE TIP, 32:6102-6114, 2023. 1", + "[19] Pavel Izmailov, Polina Kirichenko, Nate Gruver, and Andrew G Wilson. On feature learning in the presence of spurious correlations. In NeurIPS, 2022. 2", + "[20] Woo Jae Kim, Yoonki Cho, Junsik Jung, and Sung-Eui Yoon. Feature separation and recalibration for adversarial robustness. In CVPR, 2023. 3, 5, 6, 7", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 4", + "[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. In Toronto, ON, Canada, 2009. 4", + "[23] Boqi Li and Weiwei Liu. Wat: improve the worst-class robustness in adversarial training. In AAAI, 2023. 1", + "[24] Bo Li, Haoke Xiao, and Lv Tang. Asam: Boosting segment anything model with adversarial tuning. In CVPR, 2024. 1", + "[25] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Jianwei Yang, Chunyuan Li, et al. Visual in-context prompting. In CVPR, 2024. 1", + "[26] Zhuorong Li, Daiwei Yu, Lina Wei, Canghong Jin, Yun Zhang, and Sixian Chan. Soften to defend: Towards adversarial robustness via self-guided label refinement. In CVPR, 2024. 3, 5, 6, 7", + "[27] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 2, 5", + "[28] Yifei Ming, Hang Yin, and Yixuan Li. On the impact of spurious correlation for out-of-distribution detection. In AAAI, 2022. 1", + "[29] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 4", + "[30] Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, and Ashish Kapoor. Unadversarial examples: Designing objects for robust vision. In NeurIPS, 2021. 1, 3", + "[31] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 4" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2791", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In CVPR, 2017. 1", + "[33] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Information-theoretic bias reduction via causal view of spurious correlation. In AAAI, 2022. 1", + "[34] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! In NeurIPS, 2019. 1", + "[35] Kele Shao, Keda Tao, Can Qin, Haoxuan You, Yang Sui, and Huan Wang. Holitom: Holistic token merging for fast video large language models. arXiv preprint arXiv:2505.21334, 2025. 1", + "[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4", + "[37] Naman Deep Singh, Francesco Croce, and Matthias Hein. Revisiting adversarial training for imagenet: Architectures, training and generalization across threat models. In NeurIPS, 2023. 5", + "[38] Ajay Subramanian, Elena Sizikova, Najib Majaj, and Denis Pelli. Spatial-frequency channels, shape bias, and adversarial robustness. In NeurIPS, 2024. 1", + "[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2014. 1", + "[40] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, 2015. 4", + "[41] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 1, 3, 5, 6, 7", + "[42] Zeming Wei, Yifei Wang, Yiwen Guo, and Yisen Wang. Cfa: Class-wise calibrated fair adversarial training. In CVPR, 2023. 3, 5, 6, 7", + "[43] Jing Wen, Siu-Ming Yiu, and Lucas CK Hui. Defending against model inversion attack by adversarial examples. In IEEE CSR, pages 551-556, 2021. 2", + "[44] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. In NeurIPS, 2020. 3, 5, 6, 7", + "[45] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In CVPR, 2020. 3", + "[46] Shenglin Yin, Kelu Yao, Sheng Shi, Yangzhou Du, and Zhen Xiao. Again: Adversarial training with attribution span enlargement and hybrid feature fusion. In CVPR, 2023. 5", + "[47] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 4", + "[48] Bohan Zeng, Shanglin Li, Xuhui Liu, Sicheng Gao, Xiaolong Jiang, Xu Tang, Yao Hu, Jianzhuang Liu, and Baochang Zhang. Controllable mind visual diffusion model. In AAAI, 2024. 1" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019. 1", + "[50] Tianren Zhang, Chujie Zhao, Guanyu Chen, Yizhou Jiang, and Feng Chen. Feature contamination: Neural networks learn uncorrelated features and fail to generalize. In ICML, 2024. 2" + ], + "bbox": [ + 514, + 92, + 903, + 203 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2792", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/be7e63af-2995-4db4-b164-1814191af90b_model.json b/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/be7e63af-2995-4db4-b164-1814191af90b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9af3726a5b59264f062eb72c01fce948a6619572 --- /dev/null +++ b/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/be7e63af-2995-4db4-b164-1814191af90b_model.json @@ -0,0 +1,2486 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.908, + 0.155 + ], + "angle": 0, + "content": "Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.18, + 0.749, + 0.2 + ], + "angle": 0, + "content": "Kejia Zhang\\(^{1}\\), Juanjuan Weng\\(^{2*}\\), Shaozi Li\\(^{1}\\), Zhiming Luo\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.201, + 0.721, + 0.217 + ], + "angle": 0, + "content": "1Department of Artificial Intelligence, Xiamen University, Xiamen, China" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.218, + 0.756, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)College of Information Science and Technology, Jinan University, Guangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.236, + 0.882, + 0.253 + ], + "angle": 0, + "content": "3Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.255, + 0.804, + 0.27 + ], + "angle": 0, + "content": "kejiaz171@gmail.com, jjweng@jnu.edu.cn, {szlig,zhiming.luo}@xmu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.337, + 0.486, + 0.699 + ], + "angle": 0, + "content": "Despite the remarkable progress of deep neural networks (DNNs) in various visual tasks, their vulnerability to adversarial examples raises significant security concerns. Recent adversarial training methods leverage inverse adversarial attacks to generate high-confidence examples, aiming to align adversarial distributions with high-confidence class regions. However, our investigation reveals that under inverse adversarial attacks, high-confidence outputs are influenced by biased feature activations, causing models to rely on background features that lack a causal relationship with the labels. This spurious correlation bias leads to overfitting irrelevant background features during adversarial training, thereby degrading the model's robust performance and generalization capabilities. To address this issue, we propose Debiased High-Confidence Adversarial Training (DHAT), a novel approach that aligns adversarial logits with debiased high-confidence logits and restores proper attention by enhancing foreground logit orthogonality. Extensive experiments demonstrate that DHAT achieves state-of-the-art robustness on both CIFAR and ImageNet1K benchmarks, while significantly improving generalization by mitigating the feature bias inherent in inverse adversarial training approaches. Code is available at https://github.com/KejiaZhang-Robust/DHAT." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.729, + 0.222, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.485, + 0.877 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have demonstrated exceptional performance across a range of visual tasks [13, 25, 35, 48]. However, these networks remain susceptible to adversarial attacks, which present significant security challenges [11, 39]. Adversarial attacks introduce subtle perturbations that are imperceptible to humans but can significantly disrupt the inference process of DNNs, leading to incorrect predictions. Adversarial training (AT) is widely rec" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.303, + 0.677, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.606, + 0.465, + 0.621, + 0.475 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.304, + 0.905, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.464, + 0.812, + 0.475 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.655 + ], + "angle": 0, + "content": "Figure 1. Inverse adversarial attack exploits spurious correlations via attention misallocation to non-causal background features. The negative \\(\\epsilon\\) denotes inverse adversarial perturbation strength. (a) Grad-CAM [32] visualizations show that adversarial attacks disrupt attention and cause errors, while inverse adversarial attacks amplify activation and improve accuracy but expose a critical flaw: attention shifts from foreground to irrelevant background regions. (b) Quantitative comparison of IoU(\\%) between attention regions and the true foreground and background regions using UIAT [9] and AT. Increasing inverse attack strength \\(-\\epsilon\\) demonstrates UIAT's disproportionate focus on non-causal background features, exposing reliance on spurious correlations [3, 28, 33]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.907, + 0.718 + ], + "angle": 0, + "content": "ognized as one of the most effective methods for defending against these attacks by incorporating adversarial examples into the training process [14, 34]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recently, numerous advanced AT-based techniques have been proposed to address the vulnerabilities of DNNs [18, 23, 24, 38]. Prominent methods in this domain focus on aligning the distributions of adversarial examples to their true classes, thereby enhancing classification accuracy under adversarial attacks. For example, MART [41] and TRADES [49] aim to align the logits of adversarial examples with their corresponding natural examples, promoting consistency in predictions. Expanding on this, UIAT [9] and ACR [5] extend this approach by generating inverse adversarial examples [30] that exhibit higher confidence compared to their natural counterparts. This strategy helps al" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2783" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "leviate the negative impact of misclassification of natural examples in alignment and directs adversarial examples toward regions with correct high-confidence classifications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.139, + 0.483, + 0.397 + ], + "angle": 0, + "content": "Upon investigation, we uncover a remarkable yet concerning phenomenon: inverse adversarial examples produce high-confidence outputs by triggering biased feature activation. To illuminate this effect, we visualize attention maps from a conventional adversarial training (AT) model under both adversarial and inverse adversarial attacks (see Figure 1 (a)). Adversarial attacks disrupt normal attention patterns, causing incorrect predictions. In contrast, inverse adversarial attacks amplify the model's attention on features, with stronger attacks even improving prediction accuracy [9, 43]. However, this apparent enhancement masks a critical flaw: the model's attention systematically shifts from discriminative foreground features to irrelevant background regions—akin to identifying a sheep by detecting grass rather than the animal itself. This reliance on spurious correlations undermines the model's causal reasoning capabilities and compromises its robustness [1, 6, 19]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.64 + ], + "angle": 0, + "content": "To delve deeper into this issue, we conduct a statistical analysis using the Intersection over Union (IoU) metric to measure the alignment between attention maps and foreground/background regions on a subset of ImageNet [17]. Leveraging the Class Activation Mapping (CAM)-based technique inspired by [10], we define foreground as regions with high-class activation (i.e., areas critical for predictions) and background as regions with low activation. Comparing the UIAT model [9] (trained with inverse adversarial examples) to a standard AT model reveals a concerning trend. As evidenced in Figure 1 (b), while increasing the magnitude \\(\\epsilon\\) of inverse adversarial attacks enhances overall feature attention, the UIAT model fails to improve IoU for foreground features compared to the AT model. Instead, it demonstrates a marked increase in IoU for background features—a clear indicator of attention misallocation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.793 + ], + "angle": 0, + "content": "This leads to a pivotal insight: the inverse adversarial training disproportionately skews attention toward background features, resulting in biased feature activation. This shift engenders a fundamental vulnerability: a spurious correlation bias where the model becomes overly dependent on contextual background cues rather than intrinsic object features. Much like a classifier that associates beaches with birds rather than wings and beaks, this misplaced reliance significantly degrades robustness and generalization performance [15, 50], as conclusively demonstrated in Table 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Building upon the above findings, we introduce a method named Debiased High-Confidence Adversarial Training (DHAT). DHAT implements two key techniques: Debiased High-Confidence Logit Regularization (DHLR) and Foreground Logit Orthogonal Enhancement (FLOE). DHLR quantifies the bias towards background feature activation and debiase the spurious correlations by recalibrating the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "biased high-confidence logits derived from inverse adversarial examples. This regularization aligns the logits of adversarial examples with debiased high-confidence logits to improve adversarial robustness and mitigate spurious correlation bias. FLOE further refines this process by reducing the correlation between high-confidence logits and background features in affine space, which helps restore the model's attention to its normal state. Extensive experiments demonstrate that DHAT outperforms state-of-the-art techniques across various datasets, providing superior adversarial robustness and generalization. Additionally, our DHAT can be seamlessly integrated with existing advanced adversarial training methods." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.289, + 0.838, + 0.303 + ], + "angle": 0, + "content": "Below, we outline our principal contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.305, + 0.905, + 0.379 + ], + "angle": 0, + "content": "- We identify that the adversarial training model shows a biased feature activation under inverse adversarial attacks. Training with inverse adversarial examples causes the model's attention shifting towards background features, thus lead to spurious correlation bias." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.38, + 0.905, + 0.456 + ], + "angle": 0, + "content": "- We introduce a novel method, Debiased High-Confidence Adversarial Training (DHAT). This method aligns the logits of adversarial examples with debiased high-confidence logits corresponding to their classes, effectively mitigating spurious correlation bias." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.456, + 0.905, + 0.516 + ], + "angle": 0, + "content": "- Extensive experiments validate the superior robustness and generalization of our method against adversarial attacks. Additionally, our approach integrates smoothly with existing advanced adversarial training strategies." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.305, + 0.905, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.532, + 0.655, + 0.547 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.557, + 0.697, + 0.572 + ], + "angle": 0, + "content": "2.1. Adversarial Attack" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.905, + 0.641 + ], + "angle": 0, + "content": "Adversarial attacks introduce subtle perturbations \\(\\delta\\) into natural examples \\(x\\), causing DNNs to produce incorrect predictions rather than the true class \\(y\\). These attacks are typically formulated as an optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.654, + 0.905, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\hat {x} = \\max _ {| | \\delta | | _ {p} \\leq \\epsilon} \\left(\\mathcal {L} _ {\\text {a t t a c k}} (\\theta ; x + \\delta , y)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.905, + 0.736 + ], + "angle": 0, + "content": "where \\(\\hat{x}\\) represents the adversarial example, \\(\\mathcal{L}_{attack}\\) is the adversarial attack loss function, \\(\\theta\\) denotes the model parameters, and \\(\\delta\\) is constrained to the \\(p\\)-norm ball of radius \\(\\epsilon\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.737, + 0.905, + 0.812 + ], + "angle": 0, + "content": "Numerous studies have investigated the vulnerability of DNNs and have proposed various adversarial attack methods. Goodfellow et al. [14] introduced the Fast Gradient Sign Method (FGSM), which compute the adversarial perturbation in the direction of the gradient:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.826, + 0.905, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\delta = \\eta \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} \\mathcal {L} _ {\\text {a t t a c k}} (\\theta ; x, y)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(\\eta\\) denotes the attack step size. Building on this, Madry et al. [27] proposed Projected Gradient Descent (PGD), which iteratively updates the gradient of the input" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2784" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "data to generate more potent adversarial examples, thereby increasing the success rate of attacks:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.128, + 0.483, + 0.148 + ], + "angle": 0, + "content": "\\[\n\\hat {x} ^ {t + 1} = \\Pi_ {\\mathbb {B} (\\epsilon)} \\{\\hat {x} ^ {t} + \\eta \\cdot \\operatorname {s i g n} (\\nabla_ {\\hat {x} ^ {t}} \\mathcal {L} _ {\\text {a t t a c k}} (\\theta ; \\hat {x} ^ {t}, y)) \\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.484, + 0.259 + ], + "angle": 0, + "content": "where \\(\\Pi_{\\mathbb{B}(\\epsilon)}\\) projects the perturbation onto the \\(p\\)-norm ball of radius \\(\\epsilon\\). Furthermore, Carlini and Wagner [4] developed the Carlini-Wagner (C&W) attack, an optimization-based method capable of producing more imperceptible adversarial examples. Croce and Hein [7] introduced AutoAttack, an ensemble attack strategy that is user-independent and parameter-free." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.267, + 0.284, + 0.282 + ], + "angle": 0, + "content": "2.2. Adversarial Defense" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.484, + 0.38 + ], + "angle": 0, + "content": "Adversarial training (AT) is a widely used strategy for improving model robustness. The core principle of AT is to include adversarial examples in the training process, thereby aligning the model's outputs towards the true label distribution [12, 45]. Mathematically, conventional AT methods are formulated as a min-max optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.386, + 0.483, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} \\mathbb {E} _ {(x, y) \\sim \\mathcal {D}} \\max _ {| | \\delta | | _ {p} \\leq \\epsilon} \\mathcal {L} _ {A T} (\\theta ; \\mathcal {G} _ {\\theta} (x, \\delta), y), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.553 + ], + "angle": 0, + "content": "where \\((x,y)\\) denotes a clean image-label pair sampled from the data distribution \\(\\mathcal{D}\\), \\(\\delta\\) is constrained to the maximum \\(p\\)-norm magnitude \\(\\epsilon\\), \\(\\mathcal{L}_{AT}\\) represents the training strategy function, and \\(\\mathcal{G}_{\\theta}(x,\\delta)\\) denotes the attack generation procedure incorporating the perturbation \\(\\delta\\). The outer minimization optimizes the model parameters \\(\\theta\\) to minimize the expected loss over adversarial examples, while the inner maximization identifies the perturbations that maximize this loss, thereby enhancing the model's robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.75 + ], + "angle": 0, + "content": "Recent advancements in AT have introduced several techniques to further improve adversarial robustness. Wu et al. [44] introduced Adversarial Weight Perturbation (AWP), which enhances robustness by flattening the loss landscape with respect to input variations. Kim et al. [20] proposed Feature Separation and Recalibration (FSR) to capture useful logits from non-robust features. Wei et al. [42] developed Class-wise Calibrated Fair Adversarial Training (CFA), which adaptively configures training for each class to achieve category-level robust fairness. Li et al. [26] proposed SGLR, which uses self-distillation to refine the soft-label distribution logits, thereby calibrating the adversarial training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Methods similar to our approach include MART [41] and UIAT [9]. These methods use Kullback-Leibler (KL) divergence to align the logits of adversarial examples with high-confidence logits. MART encourages sharing a similar logit distribution between adversarial examples and their corresponding natural examples. UIAT generates inverse adversarial examples [30] that exhibit higher confidence compared to natural examples, thereby enhancing the alignment of adversarial examples towards the high-confidence regions of the decision surface." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.907, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.256, + 0.907, + 0.327 + ], + "angle": 0, + "content": "Figure 2. Overview of Debiased High-Confidence Logits Regularization (DHLR). The method first quantifies the degree of biased activation towards background features and recalibrates biased logits, then aligns the logits of adversarial examples with biased high-confidence logits." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.352, + 0.651, + 0.37 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.907, + 0.469 + ], + "angle": 0, + "content": "In this section, we introduce the basic notation and present our approach to mitigate adversarial bias and improve robustness. Our method recalibrates high-confidence logits and reduces reliance on spurious background features. The optimization strategy is outlined, with pseudocode provided in Section 7 of the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.479, + 0.619, + 0.494 + ], + "angle": 0, + "content": "3.1. Notation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.907, + 0.683 + ], + "angle": 0, + "content": "We denote the training dataset as \\((X,Y) = \\{(x_{i},y_{i})|i\\in (1,2,\\ldots ,N)\\}\\), where \\(x\\) and \\(y\\) denote the natural input examples and their corresponding labels, respectively. \\(N\\) represents the total number of examples in the dataset. Let \\(\\theta\\) denote the parameters of a convolutional neural network (CNN) for image classification. The network's output logits are denoted as \\(f_{\\theta}(\\cdot)\\), and the final predicted label is given by \\(\\arg \\max f_{\\theta}(\\cdot)\\). Inverse adversarial examples \\(\\tilde{x}\\) can be regarded as examples for which the model correctly predicts the label with high-confidence. These examples are generated by applying image perturbations aimed at minimizing the attack loss [30]:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.695, + 0.907, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\check {x} = \\min _ {| | \\delta | | _ {p} \\leq \\epsilon} \\left(\\mathcal {L} _ {\\mathrm {I n v}} (\\theta ; x - \\delta , y)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.732, + 0.907, + 0.795 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{Inv}}\\) denotes the loss function used to generate inverse adversarial attack. The logits of the inverse adversarial example \\(\\tilde{x}\\) and the adversarial example \\(\\hat{x}\\) are computed as \\(\\tilde{z} = f_{\\theta}(\\tilde{x})\\) and \\(\\hat{z} = f_{\\theta}(\\hat{x})\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.906, + 0.833 + ], + "angle": 0, + "content": "3.2. Debiased High-Confidence Logit Regularization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To mitigate the spurious correlation bias introduced by adversarial training with inverse adversarial examples, we propose utilizing debiased high-confidence logits for regularization. Specifically, this approach involves separating and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2785" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "quantifying the spurious bias towards background features under inverse adversarial attacks, and then recalibrating the biased high-confidence logits to align with those of adversarial examples, as illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.272 + ], + "angle": 0, + "content": "First, we quantify the model's biased activation degree towards background features under inverse adversarial attacks. Following the approach proposed by [10], we use using Grad-CAM [31] or alternative methods such as SAM [21] to generate pixel-wise attention maps \\( M \\in [0,1]^{n \\times n} \\) for the natural example \\( x \\). We separate the background feature \\( \\tilde{x}_{(B)} \\) from the inverse adversarial examples \\( \\tilde{x} \\) based on the \\( M \\):" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.279, + 0.483, + 0.298 + ], + "angle": 0, + "content": "\\[\n[ \\check {x} _ {(B)} ] _ {(i, j)} = \\mathbb {I} _ {\\left(M _ {i, j} < \\omega\\right)} \\cdot \\check {x} _ {(i, j)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.482, + 0.348 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}_{(M_{i,j} < \\omega)}\\) is an indicator function and \\(\\omega\\) is a predefined threshold. Next, we compute the logits for the background feature \\(\\check{x}_{(B)}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.228, + 0.354, + 0.483, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\check {z} _ {(B)} = f _ {\\theta} \\left(\\check {x} _ {(B)}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.482, + 0.406 + ], + "angle": 0, + "content": "These logits can quantify the biased degree introduced by background feature activation during network inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.482, + 0.452 + ], + "angle": 0, + "content": "To recalibrate the biased high-confidence logits, we refine the logits of inverse adversarial examples by subtracting the background feature logits:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.459, + 0.483, + 0.476 + ], + "angle": 0, + "content": "\\[\n\\check {z} ^ {*} = \\check {z} - \\check {z} _ {(B)}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.481, + 0.465, + 0.495 + ], + "angle": 0, + "content": "where \\(\\check{z}^*\\) represents the debiased high-confidence logits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.483, + 0.587 + ], + "angle": 0, + "content": "Next, we introduce a regularizer to align the logits of adversarial examples with these debiased high-confidence logits. This regularization encourages the model to focus on foreground features and mitigates the biased activation of background features. Specifically, our Debiased High-Confidence Logit Regularization (DHLR) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.593, + 0.482, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D H L R} \\left(\\check {z} ^ {*}, \\hat {z}\\right) = \\mathcal {L} _ {K L} \\left(\\phi \\left(\\check {z} ^ {*}\\right) | | \\phi (\\hat {z})\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.645 + ], + "angle": 0, + "content": "where \\(\\phi\\) denotes the softmax function, \\(\\mathcal{L}_{KL}\\) represents the Kullback-Leibler (KL) divergence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.483, + 0.721 + ], + "angle": 0, + "content": "This regularization term \\(\\mathcal{L}_{DHLR}\\) promotes consistency between the logits of adversarial examples \\(\\hat{z}\\) and the debiased high-confidence logits \\(\\tilde{z}^*\\). This alignment mitigates spurious correlation bias, thereby improving adversarial robustness and generalization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.468, + 0.744 + ], + "angle": 0, + "content": "3.3. Foreground Logit Orthogonal Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "While DHLR effectively recalibrates high-confidence logits to mitigate spurious correlations, it fails to fully restore the model's focus on foreground features when handling inverse adversarial examples \\(\\check{x}\\). This is because DHLR primarily aligns logits with debiased high-confidence targets, without directly addressing the model's persistent bias toward background features. Under inverse adversarial attacks, the model still exhibits strong activation in irrelevant background regions, undermining its ability to prioritize critical foreground cues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "To address this issue, we introduce Foreground Logit Orthogonal Enhancement (FLOE), a targeted method designed to minimize the correlation between high-confidence logits \\(\\tilde{z}\\) and background feature logits \\(\\tilde{z}_{(B)}\\). The intuition behind FLOE is straightforward: by reducing the projection of \\(\\tilde{z}\\) onto \\(\\tilde{z}_{(B)}\\) in the affine space, we make \\(\\tilde{z}\\) less explainable by \\(\\tilde{z}_{(B)}\\), thereby encouraging the model to rely more on foreground features. This projection-based approach ensures that the model's predictions are driven by meaningful foreground information rather than spurious background correlations. The optimization objective is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.268, + 0.905, + 0.303 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {F L O E} \\left(\\check {z}, \\check {z} _ {(B)}\\right) = - \\left| \\check {z} - \\frac {\\check {z} \\cdot \\check {z} _ {(B)}}{\\left| \\check {z} _ {(B)} \\right| ^ {2}} \\cdot \\check {z} _ {(B)} \\right| _ {p}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.905, + 0.357 + ], + "angle": 0, + "content": "where \\(p\\) indicates the norm exponent. The term \\(\\mathcal{L}_{FLOE}\\) is designed to alleviate biased feature activation under inverse adversarial attacks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.368, + 0.722, + 0.384 + ], + "angle": 0, + "content": "3.4. Optimization Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.905, + 0.449 + ], + "angle": 0, + "content": "Finally, we integrate the DHLR and FLOE techniques into a unified training framework. Debiased High-Confidence Adversarial Training (DHAT). The overall optimization objective for DHAT is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.459, + 0.905, + 0.496 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {D H A T} = \\mathcal {L} _ {A T} (\\hat {z}, y) + \\lambda_ {1} \\cdot \\mathcal {L} _ {D H L R} (\\hat {z} ^ {*}, \\hat {z}) \\tag {11} \\\\ + \\lambda_ {2} \\cdot \\mathcal {L} _ {F L O E} (\\tilde {z}, \\tilde {z} _ {(B)}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.905, + 0.598 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{AT}(\\hat{z},y)\\) represents the standard adversarial training loss, and \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyperparameters that control the relative contributions of the DHLR and FLOE terms, respectively. Besides, our DHAT can be seamlessly integrated with various advanced adversarial training strategies represented by \\(\\mathcal{L}_{AT}\\) to improve the robustness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.646, + 0.629 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.721, + 0.653 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.659, + 0.706, + 0.672 + ], + "angle": 0, + "content": "4.1.1. Datasets and models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.905, + 0.769 + ], + "angle": 0, + "content": "We conduct experiments on the CIFAR-10, CIFAR-100 [22], and ImageNet-1K [29] datasets. The architecture for our experiments include the mainstream models WRN28-10 [47], ResNet-18 [16], ResNet-50 [16], Inception-V3 [40], and VGG16 [36]. Detailed information is provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.777, + 0.663, + 0.792 + ], + "angle": 0, + "content": "4.1.2. Training setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.901 + ], + "angle": 0, + "content": "For CIFAR datasets, each model is trained for 100 epochs, with the learning rate reduced by a factor of 0.1 at the 80th and 90th epochs. For adversarial example generation, we use a fixed 10-iteration attack. The maximum \\(\\ell_{\\infty}\\)-norm of the adversarial perturbation is set to \\(\\epsilon = 8 / 255\\), while the inverse adversarial perturbation is set to \\(\\epsilon = 4 / 255\\). Both perturbations use a fixed step size of \\(\\alpha = 2 / 255\\). For the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2786" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.123, + 0.09, + 0.873, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison of robustness (%) and robust generalization gap (%) . The number in bold indicates the best performance." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.116, + 0.837, + 0.637 + ], + "angle": 0, + "content": "
Architecture: WRN28-10\nImageNet-1KAttack (ε = 4/255)Robust Gap↓
Clean↑PGD-10↑PGD-20↑PGD-50↑C&W↑AA↑
MART [41]62.3143.5642.8942.5440.6238.9413.76
AWP [44]64.2545.1344.6744.2141.9640.0212.82
FSR [20]64.7044.3143.8243.4541.2339.3014.17
CFA [42]65.3844.8744.3243.9641.5839.6515.49
UIAT [9]62.6445.2944.8544.5742.1340.1814.68
SGLR [26]63.4644.5244.0543.7841.3639.4719.72
DHAT (Ours)65.9046.8346.4246.1143.2541.709.53
DHAT-AWP (Ours)65.6247.3847.0446.7943.6242.0710.27
DHAT-CFA (Ours)66.2648.2747.8947.5344.1842.4511.64
Architecture: WRN28-10\nCIFAR-10Attack (ε = 8/255)Robust Gap↓
Clean↑PGD-10↑PGD-20↑PGD-50↑C&W↑AA↑
MART [41]82.9956.2555.4855.4552.2650.679.52
AWP [44]82.6757.8057.2157.0754.8251.906.90
FSR [20]82.9256.6955.9455.5153.9351.747.42
CFA [42]84.4357.8756.9056.6454.6051.8510.36
UIAT [9]82.9458.6658.1258.0554.1152.177.92
SGLR [26]85.7657.5356.9156.6654.2852.079.38
DHAT (Ours)83.9560.4959.9559.8755.2753.103.51
DHAT-AWP (Ours)83.2161.6161.3561.2755.7153.694.54
DHAT-CFA (Ours)84.4962.6762.3862.2255.9554.056.33
Architecture: WRN28-10\nCIFAR-100Attack (ε = 8/255)Robust Gap↓
Clean↑PGD-10↑PGD-20↑PGD-50↑C&W↑AA↑
MART [41]54.6932.0631.9031.8828.7727.259.96
AWP [44]57.9434.0133.7533.7230.7428.907.87
FSR [20]57.4832.9332.3032.2629.1627.047.84
CFA [42]60.9233.1032.5632.4130.4928.0410.47
UIAT [9]57.6534.2733.9133.8530.9729.0311.70
SGLR [26]61.0233.4332.9832.8230.7228.5015.67
DHAT (Ours)59.1435.8235.3335.0231.7230.174.24
DHAT-AWP (Ours)59.3036.3936.1535.9832.0530.384.85
DHAT-CFA (Ours)61.5437.6737.1536.9932.4030.935.93
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.482, + 0.737 + ], + "angle": 0, + "content": "ImageNet dataset, we adopt the approach in [37] with 2-iteration PGD over 50 epochs, and set \\(\\epsilon = 4 / 255\\) and \\(\\alpha = 1 / 255\\) for adversarial perturbations. The hyperparameters \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are set to 1.0 in all data sets. Comprehensive training details are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.746, + 0.258, + 0.761 + ], + "angle": 0, + "content": "4.1.3. Evaluation setup." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For robustness performance evaluation, we employ PGD [27], C&W [4], and AutoAttack (AA) [7] attacks under the \\(\\ell_{\\infty}\\) norm. AutoAttack comprises APGD-DLR [7], APGD-CE [7], FAB [8], and Square [2]. \"Clean\" denotes natural examples without adversarial perturbations. Additionally, we introduce the concept of the robust generalization gap, referred to as the \"Robust Gap\", which measures the difference in robustness performance between the training and test sets as a direct indicator of the generalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.906, + 0.708 + ], + "angle": 0, + "content": "ability [46]. This metric quantifies the strength of spurious correlations, where larger gaps indicate greater reliance on features that lack causal relationships with target labels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.716, + 0.709, + 0.731 + ], + "angle": 0, + "content": "4.2. Experimental Result" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.739, + 0.906, + 0.814 + ], + "angle": 0, + "content": "In this part, we conducted comprehensive evaluations of our proposed method across various visual datasets and compared it with mainstream adversarial training (AT) methods. Additionally, DHAT can be integrated with advanced AT strategies. The results are shown in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.822, + 0.707, + 0.837 + ], + "angle": 0, + "content": "4.2.1. Robust performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "DHAT consistently demonstrates superior robustness across all datasets. For example, on CIFAR-10, DHAT outperforms the second-best method by \\(1.93\\%\\) and \\(1.16\\%\\) under PGD-10 and C&W attacks, respectively. DHAT also" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2787" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.122, + 0.089, + 0.877, + 0.105 + ], + "angle": 0, + "content": "Table 2. Comparison of robustness (%) and robust generalization gap (%). The number in bold indicates the best performance." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.115, + 0.885, + 0.293 + ], + "angle": 0, + "content": "
Dataset: CIFAR-10\nMethodResNet-50VGG-16Inception-V3
Clean↑AA↑Robust Gap↓Clean↑AA↑Robust Gap↓Clean↑AA↑Robust Gap↓
MART [41]75.0549.054.5766.7244.273.6076.0549.595.91
AWP [44]75.5950.803.9167.1445.973.4776.7450.464.07
FSR [20]78.3850.593.9769.5944.903.3077.2850.024.73
CFA [42]77.8850.644.2770.4545.534.6676.7049.974.46
UIAT [9]78.4251.002.9968.4645.273.5775.1951.233.70
SGLR [26]80.3950.555.2170.9045.043.9480.6250.735.00
DHAT (Ours)79.9451.921.8570.0146.401.3477.8551.811.92
DHAT-AWP (Ours)76.8452.052.3467.4147.212.1877.1252.082.88
DHAT-CFA (Ours)80.9752.382.0970.5247.832.2178.5252.672.44
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.484, + 0.392 + ], + "angle": 0, + "content": "shows the best performance on more complex datasets such as CIFAR-100 and high-resolution Imagenette datasets. Specifically, on ImageNet-1K, DHAT achieves improvements of \\(1.54\\%\\) and \\(1.52\\%\\) over the second-best method under PGD-10 and AA attacks, respectively." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.408, + 0.37, + 0.421 + ], + "angle": 0, + "content": "4.2.2. Combine with other AT methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.429, + 0.484, + 0.626 + ], + "angle": 0, + "content": "DHAT can be effectively combined with various AT methods to further enhance adversarial performance. Notably, DHAT can be combined with AWP and CFA, denoted as DHAT-AWP and DHAT-CFA, respectively. However, DHAT is not suitable for combining with methods that rely on spurious correlation to improve performance (e.g., FSR, SLGR). In particular, DHAT-CFA achieves superior performance across all datasets. For instance, on Imagenette, DHAT-CFA provides a \\(2.84\\%\\) performance improvement under the AA attack compared to the second-best method. Moreover, combining DHAT with advanced AT surpasses the performance of the original baseline methods in terms of both robustness and clean accuracy." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.641, + 0.267, + 0.657 + ], + "angle": 0, + "content": "4.2.3. Robust overfitting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.662, + 0.484, + 0.8 + ], + "angle": 0, + "content": "Our method significantly reduces the robust generalization gap across all evaluated datasets, effectively mitigating the robust overfitting phenomenon. For example, on CIFAR-10, DHAT reduces the robust generalization gap by \\(4.41\\%\\) and \\(6.01\\%\\) compared to UIAT and MART, respectively. This reduction highlights DHAT's ability to mitigate spurious correlation bias, thereby improving the model's overall generalization performance. Detailed visualizations of the learning curve are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.816, + 0.312, + 0.832 + ], + "angle": 0, + "content": "4.3. Generalization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this part, we evaluate the generalizability of our proposed method by comprehensively evaluating its performance across various network architectures and under different magnitudes of adversarial attacks." + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.315, + 0.712, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.468, + 0.673, + 0.48 + ], + "angle": 0, + "content": "(a) PGD-10 Attack" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.316, + 0.901, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.77, + 0.468, + 0.861, + 0.48 + ], + "angle": 0, + "content": "(b) C&W Attack" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.494, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Figure 3. Comparisons of robustness \\((\\%)\\) under varying \\(\\epsilon\\) adversarial attacks using ResNet-18 on CIFAR-10." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.892, + 0.556 + ], + "angle": 0, + "content": "4.3.1. Performance with diverse network architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.714 + ], + "angle": 0, + "content": "To evaluate the adaptability of our method across different network architectures, we applied DHAT on ResNet-50, VGG-16, and Inception-V3. The performance comparisons are reported in Table 2. The results demonstrate the DHAT's consistently superior robustness and generalization across a variety of model architectures. For example, DHAT-CFA outperforms the second-best method by \\(1.83\\%\\) under AA attacks on ResNet-50. Similarly, on VGG-16, DHAT significantly reduces the robustness generalization gap by \\(1.96\\%\\) compared to the second-best method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.729, + 0.753, + 0.744 + ], + "angle": 0, + "content": "4.3.2. Different attack magnitude." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Figure 3 illustrates the performance of DHAT under varying magnitudes of PGD-10 and C&W attacks. Our findings reveal that DHAT maintains a significant performance advantage across attack magnitudes. Compared to other methods, DHAT exhibits improved stability and robustness, with minimal performance degradation even under more severe attacks. This consistent performance across varying attack levels underscores the effectiveness of our approach in mitigating the impact of adversarial perturbations. Further details are provided in the supplementary material (Section 7)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2788" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.483, + 0.119 + ], + "angle": 0, + "content": "Table 3. Comparison of computational costs (# A100 GPU hours) using different adversarial training methods on CIFAR-10." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.128, + 0.465, + 0.267 + ], + "angle": 0, + "content": "
MethodArchitecture
ResNet-18WRN28-10Inception-V3
MART [41]0.0500.2850.331
AWP [44]0.0480.2740.318
FSR [20]0.0510.2910.337
CFA [42]0.0450.2620.303
UIAT [9]0.0400.2450.283
SGLR [26]0.0390.2400.278
DHAT (Ours)0.0410.2480.287
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.299, + 0.483, + 0.328 + ], + "angle": 0, + "content": "Table 4. Comparison of robustness (\\%) and robust generalization gap (\\%). The number in bold indicates the best performance." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.338, + 0.458, + 0.412 + ], + "angle": 0, + "content": "
DHLRFLOEClean↑PGD-10↑AA↑Robust Gap↓
78.0355.0348.353.84
✓78.7956.5649.471.98
✓78.5056.2249.732.06
✓✓79.2857.2150.250.81
" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.442, + 0.325, + 0.459 + ], + "angle": 0, + "content": "4.4. Computational Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.468, + 0.483, + 0.605 + ], + "angle": 0, + "content": "The computational costs of our method and baseline models are provided in terms of per-epoch training time on a single NVIDIA A100 GPU with a batch size of 128 on the CIFAR-10 dataset, as shown in Table 3. Our method demonstrates significantly improved time efficiency compared to the baseline MART, AWP, FSR, and CFA. While it requires slightly more computational resources than the state-of-the-art methods UIAT and SGLR, this minor increase is justifiable given the performance gains achieved." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.624, + 0.266, + 0.64 + ], + "angle": 0, + "content": "4.5. Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.483, + 0.681 + ], + "angle": 0, + "content": "In this part, we conduct a comprehensive analysis of the impact of individual components and parameters of our model." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.698, + 0.371, + 0.714 + ], + "angle": 0, + "content": "4.5.1. Impact of individual components." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We evaluate the contributions of the individual components of DHAT, specifically Debiased High-Confidence Logit Regularization (DHLR) and Foreground Logit Orthogonal Enhancement (FLOE), as shown in Table 4. The results indicate that both DHLR and FLOE significantly enhance model performance compared to the UIAT method. They enhance clean accuracy and adversarial robustness, while also reducing the gap in robustness generalization. This highlights the effectiveness of DHLR and FLOE in mitigating spurious correlation biases. Notably, the integration of both DHLR and FLOE results in a more substantial overall improvement in model performance." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.701, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.215, + 0.652, + 0.227 + ], + "angle": 0, + "content": "(a) \\(\\mathcal{L}_{DHLR}\\)" + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.09, + 0.899, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.214, + 0.84, + 0.227 + ], + "angle": 0, + "content": "(b) \\(\\mathcal{L}_{FLOE}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.234, + 0.905, + 0.275 + ], + "angle": 0, + "content": "Figure 4. Analysis of hyper-parameters \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) using ResNet-18 on CIFAR-10. The robust accuracy denotes the accuracy under PGD-10 attack." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.291, + 0.703, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.416, + 0.652, + 0.428 + ], + "angle": 0, + "content": "(a) CIFAR-10" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.291, + 0.899, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.415, + 0.843, + 0.428 + ], + "angle": 0, + "content": "(b) Imagenette" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.435, + 0.908, + 0.478 + ], + "angle": 0, + "content": "Figure 5. Analysis of hyper-parameters \\(\\omega\\) using ResNet-18 on CIFAR-10 and Imagenette. The robust accuracy denotes the accuracy under PGD-10 attack." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.506, + 0.791, + 0.521 + ], + "angle": 0, + "content": "4.5.2. Analysis of parameter sensitivity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Figure 4 illustrates the sensitivity of the optimization parameters \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) in DHAT. Our analysis reveals that these parameters influence the trade-off between adversarial robustness and clean accuracy. As \\(\\lambda_{1}\\) increases, adversarial robustness improves, while clean accuracy tends to decrease. This occurs because higher values of \\(\\lambda_{1}\\) increase the regularization strength, which enhances the alignment of adversarial example logits with debiased high-confidence logits. However, this stronger regularization effect may also lead to a reduced emphasis on clean accuracy, as the model focuses more on handling adversarial perturbations. Conversely, increasing \\(\\lambda_{2}\\) generally improves clean accuracy but may adversely affect adversarial robustness. Larger values of \\(\\lambda_{2}\\) help the model restore its attention to the normal state, which can result in less effective alignment of adversarial example logits with the debiased high-confidence logits." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.791, + 0.905, + 0.82 + ], + "angle": 0, + "content": "4.5.3. Analysis of threshold factor for quantifying spurious bias." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Figure 5 illustrates the impact of threshold values \\( w \\) on datasets with varying resolutions. Our analysis reveals that as the threshold value increases, both clean accuracy and robustness initially improve before sharply declining. This initial improvement occurs because higher thresholds help" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2789" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.365, + 0.485, + 0.448 + ], + "angle": 0, + "content": "Figure 6. Visualization of feature activation maps using UIAT and DHAT, utilizing Grad-CAM under inverse attack \\((\\epsilon = -8)\\) and adversarial attack \\((\\epsilon = 8)\\). Left-to-right: the original image; normal feature activation maps; feature activation maps under adversarial attacks for UIAT and DHAT; and feature activation maps under inverse adversarial attacks for UIAT and DHAT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.576 + ], + "angle": 0, + "content": "in better separating accurate background features. However, further increasing the threshold eventually leads to the extraction of more foreground features, which causes significant performance degradation. For low-resolution datasets, the optimal threshold tends to be higher compared to high-resolution datasets, as high-resolution images more readily separate background features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.585, + 0.37, + 0.6 + ], + "angle": 0, + "content": "4.6. Discussion DHAT Effectiveness" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.669 + ], + "angle": 0, + "content": "In this part, we analyze the mechanisms underlying the effectiveness of our proposed method, Debiased High-Confidence Adversarial Training (DHAT), with a focus on its capacity to mitigate spurious correlation bias." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.675, + 0.367, + 0.69 + ], + "angle": 0, + "content": "4.6.1. Feature Activation Map Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.693, + 0.483, + 0.815 + ], + "angle": 0, + "content": "Figure 6 visualizes feature activation maps of UIAT and DHAT models under adversarial and inverse adversarial attacks. DHAT successfully maintains normal attention patterns, while UIAT exhibits biased activation with attention shifting significantly to background features during inverse attacks. DHAT achieves this robust performance by recalibrating biased high-confidence logits and systematically reducing correlation between logits and background features." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.435, + 0.837 + ], + "angle": 0, + "content": "4.6.2. Quantitative Evaluation of Bias Mitigation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Figure 7 quantifies DHAT's effectiveness through IoU between attention maps and true foreground/background regions. DHAT achieves higher foreground IoU and lower background IoU than both AT and UIAT, confirming its" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.71, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.228, + 0.702, + 0.241 + ], + "angle": 0, + "content": "(a) IoU with Normal Foreground" + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.09, + 0.898, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.228, + 0.898, + 0.241 + ], + "angle": 0, + "content": "(b) IoU with Normal Background" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.249, + 0.905, + 0.291 + ], + "angle": 0, + "content": "Figure 7. IoU of attention regions with normal foreground and background regions. The comparison is conducted among conventional AT, UIAT, and DHAT on the Imagenette Dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.306, + 0.906, + 0.348 + ], + "angle": 0, + "content": "Table 5. Comparison of training with different magnitude reverse adversarial samples (UIAT) and our debiased training DHAT using WRN28-10 on the CIFAR-10." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.36, + 0.903, + 0.442 + ], + "angle": 0, + "content": "
Methodε of zTraining SetTest SetRobust Gap↓
PGD-10↑AA↑PGD-10↑AA↑
UIAT-490.5361.4582.9452.079.38
UIAT-892.5762.8782.2651.3311.54
UIAT-1294.9065.2881.7250.8514.43
DHAT (ours)-488.7356.6183.9553.103.51
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.906, + 0.558 + ], + "angle": 0, + "content": "superior feature attention. Table 5 shows stronger inverse attacks in UIAT worsen overfitting, increasing the robust gap from 9.38 to 14.43. In contrast, DHAT achieves better test robustness (83.95% PGD-10, 53.10% AA) with a significantly smaller robust gap (3.51), effectively mitigating overfitting while maintaining strong performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.574, + 0.634, + 0.589 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we identify a critical issue in adversarial training: AT model exhibits biased feature activation under inverse adversarial attacks. Specifically, training with inverse adversarial examples causes the model attention to shift from the foreground to background features, resulting in spurious correlation bias. To tackle this challenge, we propose a novel method called Debiased High-Confidence Adversarial Training (DHAT). DHAT incorporates two pivotal techniques: Debiased High-Confidence Logit Regularization (DHLR) and Foreground Logit Orthogonal Enhancement (FLOE). DHLR quantifies the bias towards background feature activation and recalibrates biased high-confidence logits to align the logits of adversarial examples with these debiased high-confidence logits. FLOE further restores the model's attention to its normal state by reducing the correlation between high-confidence logits and background features in affine space. This work highlights DHAT's effectiveness in enhancing model resilience and presents an adversarial training framework that integrates seamlessly with advanced training strategies." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2790" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.206 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China under Grant No. 62276221, 62376232; the Fundamental Research Funds for the Central Universities under Grant No. 21625344; Key Laboratory of Equipment Data Security and Guarantee Technology, Ministry of Education under Grant No. 2024020200." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.22, + 0.187, + 0.236 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[1] Ola Ahmad, Nicolas Béreux, Loïc Baret, Vahid Hashemi, and Freddy Lecue. Causal analysis for robust interpretability of neural networks. In WAVC, pages 4685-4694, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.289, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In ECCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[3] Saeid Asgari, Aliasghar Khani, Fereshte Khani, Ali Gholami, Linh Tran, Ali Mahdavi Amiri, and Ghassan Hamarneh. Masktune: Mitigating spurious correlations by forcing to explore. In NeurIPS, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.482, + 0.443 + ], + "angle": 0, + "content": "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. IEEE Symp. Security Privacy, pages 39-57, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[5] Hyuna Cho, Yubin Han, and Won Hwa Kim. Antiadversarial consistency regularization for data augmentation: Applications to robust medical image segmentation. In MI-CAI, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[6] Jonathan Crabbé and Mihaela van der Schaar. Evaluating the robustness of interpretability methods through explanation invariance and equivariance. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[8] Francesco Croce and Matthias Hein. Minimally distorted adversarial examples with a fast adaptive boundary attack. In ICML, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[9] Junhao Dong, Seyed-Mohsen Moosavi-Dezfooli, Jianhuang Lai, and Xiaohua Xie. The enemy of my enemy is my friend: Exploring inverse adversaries for improving aversarial training. In CVPR, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[10] Xiaoyi Dong, Jiangfan Han, Dongdong Chen, Jiayang Liu, Huanyu Bian, Zehua Ma, Hongsheng Li, Xiaogang Wang, Weiming Zhang, and Nenghai Yu. Robust superpixel-guided attentional adversarial attack. In CVPR, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[11] Alhussein Fawzi, Hamza Fawzi, and Omar Fawzi. Adversarial vulnerability for any classifier. In NeurIPS, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[12] Liam Fowl, Micah Goldblum, Ping-yeh Chiang, Jonas Geiping, Wojciech Czaja, and Tom Goldstein. Adversarial examples make strong poisons. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[13] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. Avsegformer: Audio-visual segmentation with transformer. In AAAI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.246, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[15] Xinzhe Han, Shuhui Wang, Chi Su, Qingming Huang, and Qi Tian. General greedy de-bias learning. IEEE TPAMI, 45 (8):9789-9805, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[17] FastAI Jeremy Howard. Imagenette: A smaller subset of imagenet. In *GibHub*. https://github.com/fastai/imagenette, 2019.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[18] Zhichao Huang, Yanbo Fan, Chen Liu, Weizhong Zhang, Yong Zhang, Mathieu Salzmann, Sabine Susstrunk, and Jue Wang. Fast adversarial training with adaptive step size. IEEE TIP, 32:6102-6114, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.904, + 0.318 + ], + "angle": 0, + "content": "[19] Pavel Izmailov, Polina Kirichenko, Nate Gruver, and Andrew G Wilson. On feature learning in the presence of spurious correlations. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.904, + 0.36 + ], + "angle": 0, + "content": "[20] Woo Jae Kim, Yoonki Cho, Junsik Jung, and Sung-Eui Yoon. Feature separation and recalibration for adversarial robustness. In CVPR, 2023. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.904, + 0.417 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.904, + 0.459 + ], + "angle": 0, + "content": "[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. In Toronto, ON, Canada, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.904, + 0.49 + ], + "angle": 0, + "content": "[23] Boqi Li and Weiwei Liu. Wat: improve the worst-class robustness in adversarial training. In AAAI, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.904, + 0.518 + ], + "angle": 0, + "content": "[24] Bo Li, Haoke Xiao, and Lv Tang. Asam: Boosting segment anything model with adversarial tuning. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.521, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[25] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Jianwei Yang, Chunyuan Li, et al. Visual in-context prompting. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[26] Zhuorong Li, Daiwei Yu, Lina Wei, Canghong Jin, Yun Zhang, and Sixian Chan. Soften to defend: Towards adversarial robustness via self-guided label refinement. In CVPR, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.634, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[27] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.677, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[28] Yifei Ming, Hang Yin, and Yixuan Li. On the impact of spurious correlation for out-of-distribution detection. In AAAI, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[29] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[30] Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, and Ashish Kapoor. Unadversarial examples: Designing objects for robust vision. In NeurIPS, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2791" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[32] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[33] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Information-theoretic bias reduction via causal view of spurious correlation. In AAAI, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[34] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! In NeurIPS, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[35] Kele Shao, Keda Tao, Can Qin, Haoxuan You, Yang Sui, and Huan Wang. Holitom: Holistic token merging for fast video large language models. arXiv preprint arXiv:2505.21334, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[37] Naman Deep Singh, Francesco Croce, and Matthias Hein. Revisiting adversarial training for imagenet: Architectures, training and generalization across threat models. In NeurIPS, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[38] Ajay Subramanian, Elena Sizikova, Najib Majaj, and Denis Pelli. Spatial-frequency channels, shape bias, and adversarial robustness. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[40] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[41] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[42] Zeming Wei, Yifei Wang, Yiwen Guo, and Yisen Wang. Cfa: Class-wise calibrated fair adversarial training. In CVPR, 2023. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[43] Jing Wen, Siu-Ming Yiu, and Lucas CK Hui. Defending against model inversion attack by adversarial examples. In IEEE CSR, pages 551-556, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[44] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. In NeurIPS, 2020. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[45] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[46] Shenglin Yin, Kelu Yao, Sheng Shi, Yangzhou Du, and Zhen Xiao. Again: Adversarial training with attribution span enlargement and hybrid feature fusion. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[47] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[48] Bohan Zeng, Shanglin Li, Xuhui Liu, Sicheng Gao, Xiaolong Jiang, Xu Tang, Yao Hu, Jianzhuang Liu, and Baochang Zhang. Controllable mind visual diffusion model. In AAAI, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[49] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[50] Tianren Zhang, Chujie Zhao, Guanyu Chen, Yizhou Jiang, and Feng Chen. Feature contamination: Neural networks learn uncorrelated features and fail to generalize. In ICML, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2792" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/be7e63af-2995-4db4-b164-1814191af90b_origin.pdf b/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/be7e63af-2995-4db4-b164-1814191af90b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d0273debf99d5383e0a26b7530cff7faf0c79fe6 --- /dev/null +++ b/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/be7e63af-2995-4db4-b164-1814191af90b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3b0651fd3a9757ed7809cded73cd33426ed9905bfa12af86aef1eac8fe4aee8 +size 3389374 diff --git a/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/full.md b/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a4790a0fc386cb0c8b96c7cd257d9ca1391c7b50 --- /dev/null +++ b/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/full.md @@ -0,0 +1,372 @@ +# Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment + +Kejia Zhang $^{1}$ , Juanjuan Weng $^{2*}$ , Shaozi Li $^{1}$ , Zhiming Luo $^{1,3}$ + +1Department of Artificial Intelligence, Xiamen University, Xiamen, China + +$^{2}$ College of Information Science and Technology, Jinan University, Guangzhou, China + +3Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University + +kejiaz171@gmail.com, jjweng@jnu.edu.cn, {szlig,zhiming.luo}@xmu.edu.cn + +# Abstract + +Despite the remarkable progress of deep neural networks (DNNs) in various visual tasks, their vulnerability to adversarial examples raises significant security concerns. Recent adversarial training methods leverage inverse adversarial attacks to generate high-confidence examples, aiming to align adversarial distributions with high-confidence class regions. However, our investigation reveals that under inverse adversarial attacks, high-confidence outputs are influenced by biased feature activations, causing models to rely on background features that lack a causal relationship with the labels. This spurious correlation bias leads to overfitting irrelevant background features during adversarial training, thereby degrading the model's robust performance and generalization capabilities. To address this issue, we propose Debiased High-Confidence Adversarial Training (DHAT), a novel approach that aligns adversarial logits with debiased high-confidence logits and restores proper attention by enhancing foreground logit orthogonality. Extensive experiments demonstrate that DHAT achieves state-of-the-art robustness on both CIFAR and ImageNet1K benchmarks, while significantly improving generalization by mitigating the feature bias inherent in inverse adversarial training approaches. Code is available at https://github.com/KejiaZhang-Robust/DHAT. + +# 1. Introduction + +Deep neural networks (DNNs) have demonstrated exceptional performance across a range of visual tasks [13, 25, 35, 48]. However, these networks remain susceptible to adversarial attacks, which present significant security challenges [11, 39]. Adversarial attacks introduce subtle perturbations that are imperceptible to humans but can significantly disrupt the inference process of DNNs, leading to incorrect predictions. Adversarial training (AT) is widely rec + +![](images/7a0dce767e07cf34d7d22817ca08fc8e2230eb59c6d073ca86ccfe60d02ddb0a.jpg) +(a) +Figure 1. Inverse adversarial attack exploits spurious correlations via attention misallocation to non-causal background features. The negative $\epsilon$ denotes inverse adversarial perturbation strength. (a) Grad-CAM [32] visualizations show that adversarial attacks disrupt attention and cause errors, while inverse adversarial attacks amplify activation and improve accuracy but expose a critical flaw: attention shifts from foreground to irrelevant background regions. (b) Quantitative comparison of IoU(\%) between attention regions and the true foreground and background regions using UIAT [9] and AT. Increasing inverse attack strength $-\epsilon$ demonstrates UIAT's disproportionate focus on non-causal background features, exposing reliance on spurious correlations [3, 28, 33]. + +![](images/4b1a5bc78731f31768cd4a98b8596db93676016165cb0860f3daa8fca165af52.jpg) +(b) + +ognized as one of the most effective methods for defending against these attacks by incorporating adversarial examples into the training process [14, 34]. + +Recently, numerous advanced AT-based techniques have been proposed to address the vulnerabilities of DNNs [18, 23, 24, 38]. Prominent methods in this domain focus on aligning the distributions of adversarial examples to their true classes, thereby enhancing classification accuracy under adversarial attacks. For example, MART [41] and TRADES [49] aim to align the logits of adversarial examples with their corresponding natural examples, promoting consistency in predictions. Expanding on this, UIAT [9] and ACR [5] extend this approach by generating inverse adversarial examples [30] that exhibit higher confidence compared to their natural counterparts. This strategy helps al + +leviate the negative impact of misclassification of natural examples in alignment and directs adversarial examples toward regions with correct high-confidence classifications. + +Upon investigation, we uncover a remarkable yet concerning phenomenon: inverse adversarial examples produce high-confidence outputs by triggering biased feature activation. To illuminate this effect, we visualize attention maps from a conventional adversarial training (AT) model under both adversarial and inverse adversarial attacks (see Figure 1 (a)). Adversarial attacks disrupt normal attention patterns, causing incorrect predictions. In contrast, inverse adversarial attacks amplify the model's attention on features, with stronger attacks even improving prediction accuracy [9, 43]. However, this apparent enhancement masks a critical flaw: the model's attention systematically shifts from discriminative foreground features to irrelevant background regions—akin to identifying a sheep by detecting grass rather than the animal itself. This reliance on spurious correlations undermines the model's causal reasoning capabilities and compromises its robustness [1, 6, 19]. + +To delve deeper into this issue, we conduct a statistical analysis using the Intersection over Union (IoU) metric to measure the alignment between attention maps and foreground/background regions on a subset of ImageNet [17]. Leveraging the Class Activation Mapping (CAM)-based technique inspired by [10], we define foreground as regions with high-class activation (i.e., areas critical for predictions) and background as regions with low activation. Comparing the UIAT model [9] (trained with inverse adversarial examples) to a standard AT model reveals a concerning trend. As evidenced in Figure 1 (b), while increasing the magnitude $\epsilon$ of inverse adversarial attacks enhances overall feature attention, the UIAT model fails to improve IoU for foreground features compared to the AT model. Instead, it demonstrates a marked increase in IoU for background features—a clear indicator of attention misallocation. + +This leads to a pivotal insight: the inverse adversarial training disproportionately skews attention toward background features, resulting in biased feature activation. This shift engenders a fundamental vulnerability: a spurious correlation bias where the model becomes overly dependent on contextual background cues rather than intrinsic object features. Much like a classifier that associates beaches with birds rather than wings and beaks, this misplaced reliance significantly degrades robustness and generalization performance [15, 50], as conclusively demonstrated in Table 5. + +Building upon the above findings, we introduce a method named Debiased High-Confidence Adversarial Training (DHAT). DHAT implements two key techniques: Debiased High-Confidence Logit Regularization (DHLR) and Foreground Logit Orthogonal Enhancement (FLOE). DHLR quantifies the bias towards background feature activation and debiase the spurious correlations by recalibrating the + +biased high-confidence logits derived from inverse adversarial examples. This regularization aligns the logits of adversarial examples with debiased high-confidence logits to improve adversarial robustness and mitigate spurious correlation bias. FLOE further refines this process by reducing the correlation between high-confidence logits and background features in affine space, which helps restore the model's attention to its normal state. Extensive experiments demonstrate that DHAT outperforms state-of-the-art techniques across various datasets, providing superior adversarial robustness and generalization. Additionally, our DHAT can be seamlessly integrated with existing advanced adversarial training methods. + +Below, we outline our principal contributions: + +- We identify that the adversarial training model shows a biased feature activation under inverse adversarial attacks. Training with inverse adversarial examples causes the model's attention shifting towards background features, thus lead to spurious correlation bias. +- We introduce a novel method, Debiased High-Confidence Adversarial Training (DHAT). This method aligns the logits of adversarial examples with debiased high-confidence logits corresponding to their classes, effectively mitigating spurious correlation bias. +- Extensive experiments validate the superior robustness and generalization of our method against adversarial attacks. Additionally, our approach integrates smoothly with existing advanced adversarial training strategies. + +# 2. Related Work + +# 2.1. Adversarial Attack + +Adversarial attacks introduce subtle perturbations $\delta$ into natural examples $x$ , causing DNNs to produce incorrect predictions rather than the true class $y$ . These attacks are typically formulated as an optimization problem: + +$$ +\hat {x} = \max _ {| | \delta | | _ {p} \leq \epsilon} \left(\mathcal {L} _ {\text {a t t a c k}} (\theta ; x + \delta , y)\right), \tag {1} +$$ + +where $\hat{x}$ represents the adversarial example, $\mathcal{L}_{attack}$ is the adversarial attack loss function, $\theta$ denotes the model parameters, and $\delta$ is constrained to the $p$ -norm ball of radius $\epsilon$ . + +Numerous studies have investigated the vulnerability of DNNs and have proposed various adversarial attack methods. Goodfellow et al. [14] introduced the Fast Gradient Sign Method (FGSM), which compute the adversarial perturbation in the direction of the gradient: + +$$ +\delta = \eta \cdot \operatorname {s i g n} \left(\nabla_ {x} \mathcal {L} _ {\text {a t t a c k}} (\theta ; x, y)\right), \tag {2} +$$ + +where $\eta$ denotes the attack step size. Building on this, Madry et al. [27] proposed Projected Gradient Descent (PGD), which iteratively updates the gradient of the input + +data to generate more potent adversarial examples, thereby increasing the success rate of attacks: + +$$ +\hat {x} ^ {t + 1} = \Pi_ {\mathbb {B} (\epsilon)} \{\hat {x} ^ {t} + \eta \cdot \operatorname {s i g n} (\nabla_ {\hat {x} ^ {t}} \mathcal {L} _ {\text {a t t a c k}} (\theta ; \hat {x} ^ {t}, y)) \}, \tag {3} +$$ + +where $\Pi_{\mathbb{B}(\epsilon)}$ projects the perturbation onto the $p$ -norm ball of radius $\epsilon$ . Furthermore, Carlini and Wagner [4] developed the Carlini-Wagner (C&W) attack, an optimization-based method capable of producing more imperceptible adversarial examples. Croce and Hein [7] introduced AutoAttack, an ensemble attack strategy that is user-independent and parameter-free. + +# 2.2. Adversarial Defense + +Adversarial training (AT) is a widely used strategy for improving model robustness. The core principle of AT is to include adversarial examples in the training process, thereby aligning the model's outputs towards the true label distribution [12, 45]. Mathematically, conventional AT methods are formulated as a min-max optimization problem: + +$$ +\min _ {\theta} \mathbb {E} _ {(x, y) \sim \mathcal {D}} \max _ {| | \delta | | _ {p} \leq \epsilon} \mathcal {L} _ {A T} (\theta ; \mathcal {G} _ {\theta} (x, \delta), y), \tag {4} +$$ + +where $(x,y)$ denotes a clean image-label pair sampled from the data distribution $\mathcal{D}$ , $\delta$ is constrained to the maximum $p$ -norm magnitude $\epsilon$ , $\mathcal{L}_{AT}$ represents the training strategy function, and $\mathcal{G}_{\theta}(x,\delta)$ denotes the attack generation procedure incorporating the perturbation $\delta$ . The outer minimization optimizes the model parameters $\theta$ to minimize the expected loss over adversarial examples, while the inner maximization identifies the perturbations that maximize this loss, thereby enhancing the model's robustness. + +Recent advancements in AT have introduced several techniques to further improve adversarial robustness. Wu et al. [44] introduced Adversarial Weight Perturbation (AWP), which enhances robustness by flattening the loss landscape with respect to input variations. Kim et al. [20] proposed Feature Separation and Recalibration (FSR) to capture useful logits from non-robust features. Wei et al. [42] developed Class-wise Calibrated Fair Adversarial Training (CFA), which adaptively configures training for each class to achieve category-level robust fairness. Li et al. [26] proposed SGLR, which uses self-distillation to refine the soft-label distribution logits, thereby calibrating the adversarial training. + +Methods similar to our approach include MART [41] and UIAT [9]. These methods use Kullback-Leibler (KL) divergence to align the logits of adversarial examples with high-confidence logits. MART encourages sharing a similar logit distribution between adversarial examples and their corresponding natural examples. UIAT generates inverse adversarial examples [30] that exhibit higher confidence compared to natural examples, thereby enhancing the alignment of adversarial examples towards the high-confidence regions of the decision surface. + +![](images/71b56ad4483a89b149d57fd64cd126af45e34fc490f06b0eb7c868b3f9f88469.jpg) +Figure 2. Overview of Debiased High-Confidence Logits Regularization (DHLR). The method first quantifies the degree of biased activation towards background features and recalibrates biased logits, then aligns the logits of adversarial examples with biased high-confidence logits. + +# 3. Methodology + +In this section, we introduce the basic notation and present our approach to mitigate adversarial bias and improve robustness. Our method recalibrates high-confidence logits and reduces reliance on spurious background features. The optimization strategy is outlined, with pseudocode provided in Section 7 of the supplementary material. + +# 3.1. Notation + +We denote the training dataset as $(X,Y) = \{(x_{i},y_{i})|i\in (1,2,\ldots ,N)\}$ , where $x$ and $y$ denote the natural input examples and their corresponding labels, respectively. $N$ represents the total number of examples in the dataset. Let $\theta$ denote the parameters of a convolutional neural network (CNN) for image classification. The network's output logits are denoted as $f_{\theta}(\cdot)$ , and the final predicted label is given by $\arg \max f_{\theta}(\cdot)$ . Inverse adversarial examples $\tilde{x}$ can be regarded as examples for which the model correctly predicts the label with high-confidence. These examples are generated by applying image perturbations aimed at minimizing the attack loss [30]: + +$$ +\check {x} = \min _ {| | \delta | | _ {p} \leq \epsilon} \left(\mathcal {L} _ {\mathrm {I n v}} (\theta ; x - \delta , y)\right), \tag {5} +$$ + +where $\mathcal{L}_{\mathrm{Inv}}$ denotes the loss function used to generate inverse adversarial attack. The logits of the inverse adversarial example $\tilde{x}$ and the adversarial example $\hat{x}$ are computed as $\tilde{z} = f_{\theta}(\tilde{x})$ and $\hat{z} = f_{\theta}(\hat{x})$ , respectively. + +# 3.2. Debiased High-Confidence Logit Regularization + +To mitigate the spurious correlation bias introduced by adversarial training with inverse adversarial examples, we propose utilizing debiased high-confidence logits for regularization. Specifically, this approach involves separating and + +quantifying the spurious bias towards background features under inverse adversarial attacks, and then recalibrating the biased high-confidence logits to align with those of adversarial examples, as illustrated in Figure 2. + +First, we quantify the model's biased activation degree towards background features under inverse adversarial attacks. Following the approach proposed by [10], we use using Grad-CAM [31] or alternative methods such as SAM [21] to generate pixel-wise attention maps $M \in [0,1]^{n \times n}$ for the natural example $x$ . We separate the background feature $\tilde{x}_{(B)}$ from the inverse adversarial examples $\tilde{x}$ based on the $M$ : + +$$ +[ \check {x} _ {(B)} ] _ {(i, j)} = \mathbb {I} _ {\left(M _ {i, j} < \omega\right)} \cdot \check {x} _ {(i, j)}, \tag {6} +$$ + +where $\mathbb{I}_{(M_{i,j} < \omega)}$ is an indicator function and $\omega$ is a predefined threshold. Next, we compute the logits for the background feature $\check{x}_{(B)}$ : + +$$ +\check {z} _ {(B)} = f _ {\theta} \left(\check {x} _ {(B)}\right). \tag {7} +$$ + +These logits can quantify the biased degree introduced by background feature activation during network inference. + +To recalibrate the biased high-confidence logits, we refine the logits of inverse adversarial examples by subtracting the background feature logits: + +$$ +\check {z} ^ {*} = \check {z} - \check {z} _ {(B)}, \tag {8} +$$ + +where $\check{z}^*$ represents the debiased high-confidence logits. + +Next, we introduce a regularizer to align the logits of adversarial examples with these debiased high-confidence logits. This regularization encourages the model to focus on foreground features and mitigates the biased activation of background features. Specifically, our Debiased High-Confidence Logit Regularization (DHLR) is defined as: + +$$ +\mathcal {L} _ {D H L R} \left(\check {z} ^ {*}, \hat {z}\right) = \mathcal {L} _ {K L} \left(\phi \left(\check {z} ^ {*}\right) | | \phi (\hat {z})\right), \tag {9} +$$ + +where $\phi$ denotes the softmax function, $\mathcal{L}_{KL}$ represents the Kullback-Leibler (KL) divergence. + +This regularization term $\mathcal{L}_{DHLR}$ promotes consistency between the logits of adversarial examples $\hat{z}$ and the debiased high-confidence logits $\tilde{z}^*$ . This alignment mitigates spurious correlation bias, thereby improving adversarial robustness and generalization. + +# 3.3. Foreground Logit Orthogonal Enhancement + +While DHLR effectively recalibrates high-confidence logits to mitigate spurious correlations, it fails to fully restore the model's focus on foreground features when handling inverse adversarial examples $\check{x}$ . This is because DHLR primarily aligns logits with debiased high-confidence targets, without directly addressing the model's persistent bias toward background features. Under inverse adversarial attacks, the model still exhibits strong activation in irrelevant background regions, undermining its ability to prioritize critical foreground cues. + +To address this issue, we introduce Foreground Logit Orthogonal Enhancement (FLOE), a targeted method designed to minimize the correlation between high-confidence logits $\tilde{z}$ and background feature logits $\tilde{z}_{(B)}$ . The intuition behind FLOE is straightforward: by reducing the projection of $\tilde{z}$ onto $\tilde{z}_{(B)}$ in the affine space, we make $\tilde{z}$ less explainable by $\tilde{z}_{(B)}$ , thereby encouraging the model to rely more on foreground features. This projection-based approach ensures that the model's predictions are driven by meaningful foreground information rather than spurious background correlations. The optimization objective is formulated as: + +$$ +\mathcal {L} _ {F L O E} \left(\check {z}, \check {z} _ {(B)}\right) = - \left| \check {z} - \frac {\check {z} \cdot \check {z} _ {(B)}}{\left| \check {z} _ {(B)} \right| ^ {2}} \cdot \check {z} _ {(B)} \right| _ {p}, \tag {10} +$$ + +where $p$ indicates the norm exponent. The term $\mathcal{L}_{FLOE}$ is designed to alleviate biased feature activation under inverse adversarial attacks. + +# 3.4. Optimization Strategy + +Finally, we integrate the DHLR and FLOE techniques into a unified training framework. Debiased High-Confidence Adversarial Training (DHAT). The overall optimization objective for DHAT is formulated as: + +$$ +\begin{array}{l} \mathcal {L} _ {D H A T} = \mathcal {L} _ {A T} (\hat {z}, y) + \lambda_ {1} \cdot \mathcal {L} _ {D H L R} (\hat {z} ^ {*}, \hat {z}) \tag {11} \\ + \lambda_ {2} \cdot \mathcal {L} _ {F L O E} (\tilde {z}, \tilde {z} _ {(B)}), \\ \end{array} +$$ + +where $\mathcal{L}_{AT}(\hat{z},y)$ represents the standard adversarial training loss, and $\lambda_{1}$ and $\lambda_{2}$ are hyperparameters that control the relative contributions of the DHLR and FLOE terms, respectively. Besides, our DHAT can be seamlessly integrated with various advanced adversarial training strategies represented by $\mathcal{L}_{AT}$ to improve the robustness. + +# 4. Experiments + +# 4.1. Experimental Settings + +# 4.1.1. Datasets and models. + +We conduct experiments on the CIFAR-10, CIFAR-100 [22], and ImageNet-1K [29] datasets. The architecture for our experiments include the mainstream models WRN28-10 [47], ResNet-18 [16], ResNet-50 [16], Inception-V3 [40], and VGG16 [36]. Detailed information is provided in the supplementary material. + +# 4.1.2. Training setup + +For CIFAR datasets, each model is trained for 100 epochs, with the learning rate reduced by a factor of 0.1 at the 80th and 90th epochs. For adversarial example generation, we use a fixed 10-iteration attack. The maximum $\ell_{\infty}$ -norm of the adversarial perturbation is set to $\epsilon = 8 / 255$ , while the inverse adversarial perturbation is set to $\epsilon = 4 / 255$ . Both perturbations use a fixed step size of $\alpha = 2 / 255$ . For the + +Table 1. Comparison of robustness (%) and robust generalization gap (%) . The number in bold indicates the best performance. + +
Architecture: WRN28-10 +ImageNet-1KAttack (ε = 4/255)Robust Gap↓
Clean↑PGD-10↑PGD-20↑PGD-50↑C&W↑AA↑
MART [41]62.3143.5642.8942.5440.6238.9413.76
AWP [44]64.2545.1344.6744.2141.9640.0212.82
FSR [20]64.7044.3143.8243.4541.2339.3014.17
CFA [42]65.3844.8744.3243.9641.5839.6515.49
UIAT [9]62.6445.2944.8544.5742.1340.1814.68
SGLR [26]63.4644.5244.0543.7841.3639.4719.72
DHAT (Ours)65.9046.8346.4246.1143.2541.709.53
DHAT-AWP (Ours)65.6247.3847.0446.7943.6242.0710.27
DHAT-CFA (Ours)66.2648.2747.8947.5344.1842.4511.64
Architecture: WRN28-10 +CIFAR-10Attack (ε = 8/255)Robust Gap↓
Clean↑PGD-10↑PGD-20↑PGD-50↑C&W↑AA↑
MART [41]82.9956.2555.4855.4552.2650.679.52
AWP [44]82.6757.8057.2157.0754.8251.906.90
FSR [20]82.9256.6955.9455.5153.9351.747.42
CFA [42]84.4357.8756.9056.6454.6051.8510.36
UIAT [9]82.9458.6658.1258.0554.1152.177.92
SGLR [26]85.7657.5356.9156.6654.2852.079.38
DHAT (Ours)83.9560.4959.9559.8755.2753.103.51
DHAT-AWP (Ours)83.2161.6161.3561.2755.7153.694.54
DHAT-CFA (Ours)84.4962.6762.3862.2255.9554.056.33
Architecture: WRN28-10 +CIFAR-100Attack (ε = 8/255)Robust Gap↓
Clean↑PGD-10↑PGD-20↑PGD-50↑C&W↑AA↑
MART [41]54.6932.0631.9031.8828.7727.259.96
AWP [44]57.9434.0133.7533.7230.7428.907.87
FSR [20]57.4832.9332.3032.2629.1627.047.84
CFA [42]60.9233.1032.5632.4130.4928.0410.47
UIAT [9]57.6534.2733.9133.8530.9729.0311.70
SGLR [26]61.0233.4332.9832.8230.7228.5015.67
DHAT (Ours)59.1435.8235.3335.0231.7230.174.24
DHAT-AWP (Ours)59.3036.3936.1535.9832.0530.384.85
DHAT-CFA (Ours)61.5437.6737.1536.9932.4030.935.93
+ +ImageNet dataset, we adopt the approach in [37] with 2-iteration PGD over 50 epochs, and set $\epsilon = 4 / 255$ and $\alpha = 1 / 255$ for adversarial perturbations. The hyperparameters $\lambda_{1}$ and $\lambda_{2}$ are set to 1.0 in all data sets. Comprehensive training details are available in the supplementary material. + +# 4.1.3. Evaluation setup. + +For robustness performance evaluation, we employ PGD [27], C&W [4], and AutoAttack (AA) [7] attacks under the $\ell_{\infty}$ norm. AutoAttack comprises APGD-DLR [7], APGD-CE [7], FAB [8], and Square [2]. "Clean" denotes natural examples without adversarial perturbations. Additionally, we introduce the concept of the robust generalization gap, referred to as the "Robust Gap", which measures the difference in robustness performance between the training and test sets as a direct indicator of the generalization + +ability [46]. This metric quantifies the strength of spurious correlations, where larger gaps indicate greater reliance on features that lack causal relationships with target labels. + +# 4.2. Experimental Result + +In this part, we conducted comprehensive evaluations of our proposed method across various visual datasets and compared it with mainstream adversarial training (AT) methods. Additionally, DHAT can be integrated with advanced AT strategies. The results are shown in Table 1. + +# 4.2.1. Robust performance. + +DHAT consistently demonstrates superior robustness across all datasets. For example, on CIFAR-10, DHAT outperforms the second-best method by $1.93\%$ and $1.16\%$ under PGD-10 and C&W attacks, respectively. DHAT also + +Table 2. Comparison of robustness (%) and robust generalization gap (%). The number in bold indicates the best performance. + +
Dataset: CIFAR-10 +MethodResNet-50VGG-16Inception-V3
Clean↑AA↑Robust Gap↓Clean↑AA↑Robust Gap↓Clean↑AA↑Robust Gap↓
MART [41]75.0549.054.5766.7244.273.6076.0549.595.91
AWP [44]75.5950.803.9167.1445.973.4776.7450.464.07
FSR [20]78.3850.593.9769.5944.903.3077.2850.024.73
CFA [42]77.8850.644.2770.4545.534.6676.7049.974.46
UIAT [9]78.4251.002.9968.4645.273.5775.1951.233.70
SGLR [26]80.3950.555.2170.9045.043.9480.6250.735.00
DHAT (Ours)79.9451.921.8570.0146.401.3477.8551.811.92
DHAT-AWP (Ours)76.8452.052.3467.4147.212.1877.1252.082.88
DHAT-CFA (Ours)80.9752.382.0970.5247.832.2178.5252.672.44
+ +shows the best performance on more complex datasets such as CIFAR-100 and high-resolution Imagenette datasets. Specifically, on ImageNet-1K, DHAT achieves improvements of $1.54\%$ and $1.52\%$ over the second-best method under PGD-10 and AA attacks, respectively. + +# 4.2.2. Combine with other AT methods. + +DHAT can be effectively combined with various AT methods to further enhance adversarial performance. Notably, DHAT can be combined with AWP and CFA, denoted as DHAT-AWP and DHAT-CFA, respectively. However, DHAT is not suitable for combining with methods that rely on spurious correlation to improve performance (e.g., FSR, SLGR). In particular, DHAT-CFA achieves superior performance across all datasets. For instance, on Imagenette, DHAT-CFA provides a $2.84\%$ performance improvement under the AA attack compared to the second-best method. Moreover, combining DHAT with advanced AT surpasses the performance of the original baseline methods in terms of both robustness and clean accuracy. + +# 4.2.3. Robust overfitting. + +Our method significantly reduces the robust generalization gap across all evaluated datasets, effectively mitigating the robust overfitting phenomenon. For example, on CIFAR-10, DHAT reduces the robust generalization gap by $4.41\%$ and $6.01\%$ compared to UIAT and MART, respectively. This reduction highlights DHAT's ability to mitigate spurious correlation bias, thereby improving the model's overall generalization performance. Detailed visualizations of the learning curve are provided in the supplementary material. + +# 4.3. Generalization Analysis + +In this part, we evaluate the generalizability of our proposed method by comprehensively evaluating its performance across various network architectures and under different magnitudes of adversarial attacks. + +![](images/916c0ac27ef8715d57b26358cafee580c3f236249463b5ac8a16e47dd6b30019.jpg) +(a) PGD-10 Attack +Figure 3. Comparisons of robustness $(\%)$ under varying $\epsilon$ adversarial attacks using ResNet-18 on CIFAR-10. + +![](images/71e8fa34bfc78b2ee8bcc7f3b6350ccd91cd8b538282b4c34f307a7223ca6975.jpg) +(b) C&W Attack + +# 4.3.1. Performance with diverse network architecture. + +To evaluate the adaptability of our method across different network architectures, we applied DHAT on ResNet-50, VGG-16, and Inception-V3. The performance comparisons are reported in Table 2. The results demonstrate the DHAT's consistently superior robustness and generalization across a variety of model architectures. For example, DHAT-CFA outperforms the second-best method by $1.83\%$ under AA attacks on ResNet-50. Similarly, on VGG-16, DHAT significantly reduces the robustness generalization gap by $1.96\%$ compared to the second-best method. + +# 4.3.2. Different attack magnitude. + +Figure 3 illustrates the performance of DHAT under varying magnitudes of PGD-10 and C&W attacks. Our findings reveal that DHAT maintains a significant performance advantage across attack magnitudes. Compared to other methods, DHAT exhibits improved stability and robustness, with minimal performance degradation even under more severe attacks. This consistent performance across varying attack levels underscores the effectiveness of our approach in mitigating the impact of adversarial perturbations. Further details are provided in the supplementary material (Section 7). + +Table 3. Comparison of computational costs (# A100 GPU hours) using different adversarial training methods on CIFAR-10. + +
MethodArchitecture
ResNet-18WRN28-10Inception-V3
MART [41]0.0500.2850.331
AWP [44]0.0480.2740.318
FSR [20]0.0510.2910.337
CFA [42]0.0450.2620.303
UIAT [9]0.0400.2450.283
SGLR [26]0.0390.2400.278
DHAT (Ours)0.0410.2480.287
+ +Table 4. Comparison of robustness (\%) and robust generalization gap (\%). The number in bold indicates the best performance. + +
DHLRFLOEClean↑PGD-10↑AA↑Robust Gap↓
78.0355.0348.353.84
✓78.7956.5649.471.98
✓78.5056.2249.732.06
✓✓79.2857.2150.250.81
+ +# 4.4. Computational Efficiency + +The computational costs of our method and baseline models are provided in terms of per-epoch training time on a single NVIDIA A100 GPU with a batch size of 128 on the CIFAR-10 dataset, as shown in Table 3. Our method demonstrates significantly improved time efficiency compared to the baseline MART, AWP, FSR, and CFA. While it requires slightly more computational resources than the state-of-the-art methods UIAT and SGLR, this minor increase is justifiable given the performance gains achieved. + +# 4.5. Ablation Analysis + +In this part, we conduct a comprehensive analysis of the impact of individual components and parameters of our model. + +# 4.5.1. Impact of individual components. + +We evaluate the contributions of the individual components of DHAT, specifically Debiased High-Confidence Logit Regularization (DHLR) and Foreground Logit Orthogonal Enhancement (FLOE), as shown in Table 4. The results indicate that both DHLR and FLOE significantly enhance model performance compared to the UIAT method. They enhance clean accuracy and adversarial robustness, while also reducing the gap in robustness generalization. This highlights the effectiveness of DHLR and FLOE in mitigating spurious correlation biases. Notably, the integration of both DHLR and FLOE results in a more substantial overall improvement in model performance. + +![](images/cac6452ce8fdd4fdfd37b168ff8b527a222f4fb60f3508e21988f18abf63416f.jpg) +(a) $\mathcal{L}_{DHLR}$ + +![](images/e71e2437acc5f51a0dc6dbdf3b9e7c7dd3a02eef436a73ea4f70e790fb2ed842.jpg) +(b) $\mathcal{L}_{FLOE}$ + +![](images/7c3a4d2a24a52fb478104250c76b64b0327766b92ebb43a60af45c8eaf64bb7a.jpg) +Figure 4. Analysis of hyper-parameters $\lambda_{1}$ and $\lambda_{2}$ using ResNet-18 on CIFAR-10. The robust accuracy denotes the accuracy under PGD-10 attack. +(a) CIFAR-10 +Figure 5. Analysis of hyper-parameters $\omega$ using ResNet-18 on CIFAR-10 and Imagenette. The robust accuracy denotes the accuracy under PGD-10 attack. + +![](images/85e08e3ef216b6ab5b2d70937c646b3d911ceb622e51576184fd1d65acf419f6.jpg) +(b) Imagenette + +# 4.5.2. Analysis of parameter sensitivity. + +Figure 4 illustrates the sensitivity of the optimization parameters $\lambda_{1}$ and $\lambda_{2}$ in DHAT. Our analysis reveals that these parameters influence the trade-off between adversarial robustness and clean accuracy. As $\lambda_{1}$ increases, adversarial robustness improves, while clean accuracy tends to decrease. This occurs because higher values of $\lambda_{1}$ increase the regularization strength, which enhances the alignment of adversarial example logits with debiased high-confidence logits. However, this stronger regularization effect may also lead to a reduced emphasis on clean accuracy, as the model focuses more on handling adversarial perturbations. Conversely, increasing $\lambda_{2}$ generally improves clean accuracy but may adversely affect adversarial robustness. Larger values of $\lambda_{2}$ help the model restore its attention to the normal state, which can result in less effective alignment of adversarial example logits with the debiased high-confidence logits. + +# 4.5.3. Analysis of threshold factor for quantifying spurious bias. + +Figure 5 illustrates the impact of threshold values $w$ on datasets with varying resolutions. Our analysis reveals that as the threshold value increases, both clean accuracy and robustness initially improve before sharply declining. This initial improvement occurs because higher thresholds help + +![](images/9e29c250f1f3638d19b9bca044e9c10b2fa46a1ebe54658e4c0bbc6a941a929a.jpg) +Figure 6. Visualization of feature activation maps using UIAT and DHAT, utilizing Grad-CAM under inverse attack $(\epsilon = -8)$ and adversarial attack $(\epsilon = 8)$ . Left-to-right: the original image; normal feature activation maps; feature activation maps under adversarial attacks for UIAT and DHAT; and feature activation maps under inverse adversarial attacks for UIAT and DHAT. +Figure 7. IoU of attention regions with normal foreground and background regions. The comparison is conducted among conventional AT, UIAT, and DHAT on the Imagenette Dataset. + +in better separating accurate background features. However, further increasing the threshold eventually leads to the extraction of more foreground features, which causes significant performance degradation. For low-resolution datasets, the optimal threshold tends to be higher compared to high-resolution datasets, as high-resolution images more readily separate background features. + +# 4.6. Discussion DHAT Effectiveness + +In this part, we analyze the mechanisms underlying the effectiveness of our proposed method, Debiased High-Confidence Adversarial Training (DHAT), with a focus on its capacity to mitigate spurious correlation bias. + +# 4.6.1. Feature Activation Map Analysis + +Figure 6 visualizes feature activation maps of UIAT and DHAT models under adversarial and inverse adversarial attacks. DHAT successfully maintains normal attention patterns, while UIAT exhibits biased activation with attention shifting significantly to background features during inverse attacks. DHAT achieves this robust performance by recalibrating biased high-confidence logits and systematically reducing correlation between logits and background features. + +# 4.6.2. Quantitative Evaluation of Bias Mitigation + +Figure 7 quantifies DHAT's effectiveness through IoU between attention maps and true foreground/background regions. DHAT achieves higher foreground IoU and lower background IoU than both AT and UIAT, confirming its + +![](images/33666b8d58ca2952974707b1c0b62d3e93e4dbe803fc48724e26b48ae77ba0dc.jpg) +(a) IoU with Normal Foreground + +![](images/4472d173a4ea941042d640635544b0b670d3ce8f63732b585fc254e685f001f1.jpg) +(b) IoU with Normal Background + +Table 5. Comparison of training with different magnitude reverse adversarial samples (UIAT) and our debiased training DHAT using WRN28-10 on the CIFAR-10. + +
Methodε of zTraining SetTest SetRobust Gap↓
PGD-10↑AA↑PGD-10↑AA↑
UIAT-490.5361.4582.9452.079.38
UIAT-892.5762.8782.2651.3311.54
UIAT-1294.9065.2881.7250.8514.43
DHAT (ours)-488.7356.6183.9553.103.51
+ +superior feature attention. Table 5 shows stronger inverse attacks in UIAT worsen overfitting, increasing the robust gap from 9.38 to 14.43. In contrast, DHAT achieves better test robustness (83.95% PGD-10, 53.10% AA) with a significantly smaller robust gap (3.51), effectively mitigating overfitting while maintaining strong performance. + +# 5. Conclusion + +In this work, we identify a critical issue in adversarial training: AT model exhibits biased feature activation under inverse adversarial attacks. Specifically, training with inverse adversarial examples causes the model attention to shift from the foreground to background features, resulting in spurious correlation bias. To tackle this challenge, we propose a novel method called Debiased High-Confidence Adversarial Training (DHAT). DHAT incorporates two pivotal techniques: Debiased High-Confidence Logit Regularization (DHLR) and Foreground Logit Orthogonal Enhancement (FLOE). DHLR quantifies the bias towards background feature activation and recalibrates biased high-confidence logits to align the logits of adversarial examples with these debiased high-confidence logits. FLOE further restores the model's attention to its normal state by reducing the correlation between high-confidence logits and background features in affine space. This work highlights DHAT's effectiveness in enhancing model resilience and presents an adversarial training framework that integrates seamlessly with advanced training strategies. + +# Acknowledgment + +This work is supported by the National Natural Science Foundation of China under Grant No. 62276221, 62376232; the Fundamental Research Funds for the Central Universities under Grant No. 21625344; Key Laboratory of Equipment Data Security and Guarantee Technology, Ministry of Education under Grant No. 2024020200. + +# References + +[1] Ola Ahmad, Nicolas Béreux, Loïc Baret, Vahid Hashemi, and Freddy Lecue. Causal analysis for robust interpretability of neural networks. In WAVC, pages 4685-4694, 2024. 2 +[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In ECCV, 2020. 5 +[3] Saeid Asgari, Aliasghar Khani, Fereshte Khani, Ali Gholami, Linh Tran, Ali Mahdavi Amiri, and Ghassan Hamarneh. Masktune: Mitigating spurious correlations by forcing to explore. In NeurIPS, 2022. 1 +[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. IEEE Symp. Security Privacy, pages 39-57, 2017. 3, 5 +[5] Hyuna Cho, Yubin Han, and Won Hwa Kim. Antiadversarial consistency regularization for data augmentation: Applications to robust medical image segmentation. In MI-CAI, 2023. 1 +[6] Jonathan Crabbé and Mihaela van der Schaar. Evaluating the robustness of interpretability methods through explanation invariance and equivariance. In NeurIPS, 2023. 2 +[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 3, 5 +[8] Francesco Croce and Matthias Hein. Minimally distorted adversarial examples with a fast adaptive boundary attack. In ICML, 2020. 5 +[9] Junhao Dong, Seyed-Mohsen Moosavi-Dezfooli, Jianhuang Lai, and Xiaohua Xie. The enemy of my enemy is my friend: Exploring inverse adversaries for improving aversarial training. In CVPR, 2023. 1, 2, 3, 5, 6, 7 +[10] Xiaoyi Dong, Jiangfan Han, Dongdong Chen, Jiayang Liu, Huanyu Bian, Zehua Ma, Hongsheng Li, Xiaogang Wang, Weiming Zhang, and Nenghai Yu. Robust superpixel-guided attentional adversarial attack. In CVPR, 2020. 2, 4 +[11] Alhussein Fawzi, Hamza Fawzi, and Omar Fawzi. Adversarial vulnerability for any classifier. In NeurIPS, 2018. 1 +[12] Liam Fowl, Micah Goldblum, Ping-yeh Chiang, Jonas Geiping, Wojciech Czaja, and Tom Goldstein. Adversarial examples make strong poisons. In NeurIPS, 2021. 3 +[13] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. Avsegformer: Audio-visual segmentation with transformer. In AAAI, 2024. 1 +[14] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 2 + +[15] Xinzhe Han, Shuhui Wang, Chi Su, Qingming Huang, and Qi Tian. General greedy de-bias learning. IEEE TPAMI, 45 (8):9789-9805, 2023. 2 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4 +[17] FastAI Jeremy Howard. Imagenette: A smaller subset of imagenet. In *GibHub*. https://github.com/fastai/imagenette, 2019.2 +[18] Zhichao Huang, Yanbo Fan, Chen Liu, Weizhong Zhang, Yong Zhang, Mathieu Salzmann, Sabine Susstrunk, and Jue Wang. Fast adversarial training with adaptive step size. IEEE TIP, 32:6102-6114, 2023. 1 +[19] Pavel Izmailov, Polina Kirichenko, Nate Gruver, and Andrew G Wilson. On feature learning in the presence of spurious correlations. In NeurIPS, 2022. 2 +[20] Woo Jae Kim, Yoonki Cho, Junsik Jung, and Sung-Eui Yoon. Feature separation and recalibration for adversarial robustness. In CVPR, 2023. 3, 5, 6, 7 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 4 +[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. In Toronto, ON, Canada, 2009. 4 +[23] Boqi Li and Weiwei Liu. Wat: improve the worst-class robustness in adversarial training. In AAAI, 2023. 1 +[24] Bo Li, Haoke Xiao, and Lv Tang. Asam: Boosting segment anything model with adversarial tuning. In CVPR, 2024. 1 +[25] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Jianwei Yang, Chunyuan Li, et al. Visual in-context prompting. In CVPR, 2024. 1 +[26] Zhuorong Li, Daiwei Yu, Lina Wei, Canghong Jin, Yun Zhang, and Sixian Chan. Soften to defend: Towards adversarial robustness via self-guided label refinement. In CVPR, 2024. 3, 5, 6, 7 +[27] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 2, 5 +[28] Yifei Ming, Hang Yin, and Yixuan Li. On the impact of spurious correlation for out-of-distribution detection. In AAAI, 2022. 1 +[29] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 4 +[30] Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, and Ashish Kapoor. Unadversarial examples: Designing objects for robust vision. In NeurIPS, 2021. 1, 3 +[31] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 4 + +[32] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In CVPR, 2017. 1 +[33] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Information-theoretic bias reduction via causal view of spurious correlation. In AAAI, 2022. 1 +[34] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! In NeurIPS, 2019. 1 +[35] Kele Shao, Keda Tao, Can Qin, Haoxuan You, Yang Sui, and Huan Wang. Holitom: Holistic token merging for fast video large language models. arXiv preprint arXiv:2505.21334, 2025. 1 +[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4 +[37] Naman Deep Singh, Francesco Croce, and Matthias Hein. Revisiting adversarial training for imagenet: Architectures, training and generalization across threat models. In NeurIPS, 2023. 5 +[38] Ajay Subramanian, Elena Sizikova, Najib Majaj, and Denis Pelli. Spatial-frequency channels, shape bias, and adversarial robustness. In NeurIPS, 2024. 1 +[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2014. 1 +[40] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, 2015. 4 +[41] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 1, 3, 5, 6, 7 +[42] Zeming Wei, Yifei Wang, Yiwen Guo, and Yisen Wang. Cfa: Class-wise calibrated fair adversarial training. In CVPR, 2023. 3, 5, 6, 7 +[43] Jing Wen, Siu-Ming Yiu, and Lucas CK Hui. Defending against model inversion attack by adversarial examples. In IEEE CSR, pages 551-556, 2021. 2 +[44] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. In NeurIPS, 2020. 3, 5, 6, 7 +[45] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In CVPR, 2020. 3 +[46] Shenglin Yin, Kelu Yao, Sheng Shi, Yangzhou Du, and Zhen Xiao. Again: Adversarial training with attribution span enlargement and hybrid feature fusion. In CVPR, 2023. 5 +[47] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 4 +[48] Bohan Zeng, Shanglin Li, Xuhui Liu, Sicheng Gao, Xiaolong Jiang, Xu Tang, Yao Hu, Jianzhuang Liu, and Baochang Zhang. Controllable mind visual diffusion model. In AAAI, 2024. 1 + +[49] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019. 1 +[50] Tianren Zhang, Chujie Zhao, Guanyu Chen, Yizhou Jiang, and Feng Chen. Feature contamination: Neural networks learn uncorrelated features and fail to generalize. In ICML, 2024. 2 \ No newline at end of file diff --git a/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/images.zip b/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..490677fbe543ca0cccde1dcdb8d20d98066bb9af --- /dev/null +++ b/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:543aa2dbd0f2ad357197090bae21da7219cccb7f69aee7dc615dda2945c92879 +size 658016 diff --git a/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/layout.json b/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dad28e22014cc27ff74edb679c97a455515570aa --- /dev/null +++ b/2025/Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment/layout.json @@ -0,0 +1,9673 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 122 + ], + "type": "text", + "content": "Towards Adversarial Robustness via Debiased High-Confidence Logit Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 142, + 458, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 142, + 458, + 158 + ], + "spans": [ + { + "bbox": [ + 151, + 142, + 458, + 158 + ], + "type": "text", + "content": "Kejia Zhang" + }, + { + "bbox": [ + 151, + 142, + 458, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 142, + 458, + 158 + ], + "type": "text", + "content": ", Juanjuan Weng" + }, + { + "bbox": [ + 151, + 142, + 458, + 158 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 151, + 142, + 458, + 158 + ], + "type": "text", + "content": ", Shaozi Li" + }, + { + "bbox": [ + 151, + 142, + 458, + 158 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 151, + 142, + 458, + 158 + ], + "type": "text", + "content": ", Zhiming Luo" + }, + { + "bbox": [ + 151, + 142, + 458, + 158 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 170, + 159, + 441, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 159, + 441, + 171 + ], + "spans": [ + { + "bbox": [ + 170, + 159, + 441, + 171 + ], + "type": "text", + "content": "1Department of Artificial Intelligence, Xiamen University, Xiamen, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 172, + 462, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 172, + 462, + 185 + ], + "spans": [ + { + "bbox": [ + 148, + 172, + 462, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 148, + 172, + 462, + 185 + ], + "type": "text", + "content": "College of Information Science and Technology, Jinan University, Guangzhou, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 186, + 539, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 186, + 539, + 200 + ], + "spans": [ + { + "bbox": [ + 70, + 186, + 539, + 200 + ], + "type": "text", + "content": "3Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education of China, Xiamen University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 113, + 201, + 492, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 201, + 492, + 213 + ], + "spans": [ + { + "bbox": [ + 113, + 201, + 492, + 213 + ], + "type": "text", + "content": "kejiaz171@gmail.com, jjweng@jnu.edu.cn, {szlig,zhiming.luo}@xmu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 266, + 297, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 266, + 297, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 266, + 297, + 553 + ], + "type": "text", + "content": "Despite the remarkable progress of deep neural networks (DNNs) in various visual tasks, their vulnerability to adversarial examples raises significant security concerns. Recent adversarial training methods leverage inverse adversarial attacks to generate high-confidence examples, aiming to align adversarial distributions with high-confidence class regions. However, our investigation reveals that under inverse adversarial attacks, high-confidence outputs are influenced by biased feature activations, causing models to rely on background features that lack a causal relationship with the labels. This spurious correlation bias leads to overfitting irrelevant background features during adversarial training, thereby degrading the model's robust performance and generalization capabilities. To address this issue, we propose Debiased High-Confidence Adversarial Training (DHAT), a novel approach that aligns adversarial logits with debiased high-confidence logits and restores proper attention by enhancing foreground logit orthogonality. Extensive experiments demonstrate that DHAT achieves state-of-the-art robustness on both CIFAR and ImageNet1K benchmarks, while significantly improving generalization by mitigating the feature bias inherent in inverse adversarial training approaches. Code is available at https://github.com/KejiaZhang-Robust/DHAT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 577, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 597, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 694 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have demonstrated exceptional performance across a range of visual tasks [13, 25, 35, 48]. However, these networks remain susceptible to adversarial attacks, which present significant security challenges [11, 39]. Adversarial attacks introduce subtle perturbations that are imperceptible to humans but can significantly disrupt the inference process of DNNs, leading to incorrect predictions. Adversarial training (AT) is widely rec" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 239, + 414, + 366 + ], + "blocks": [ + { + "bbox": [ + 316, + 239, + 414, + 366 + ], + "lines": [ + { + "bbox": [ + 316, + 239, + 414, + 366 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 414, + 366 + ], + "type": "image", + "image_path": "7a0dce767e07cf34d7d22817ca08fc8e2230eb59c6d073ca86ccfe60d02ddb0a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 368, + 380, + 376 + ], + "lines": [ + { + "bbox": [ + 370, + 368, + 380, + 376 + ], + "spans": [ + { + "bbox": [ + 370, + 368, + 380, + 376 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 385, + 555, + 518 + ], + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 518 + ], + "type": "text", + "content": "Figure 1. Inverse adversarial attack exploits spurious correlations via attention misallocation to non-causal background features. The negative " + }, + { + "bbox": [ + 313, + 385, + 555, + 518 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 385, + 555, + 518 + ], + "type": "text", + "content": " denotes inverse adversarial perturbation strength. (a) Grad-CAM [32] visualizations show that adversarial attacks disrupt attention and cause errors, while inverse adversarial attacks amplify activation and improve accuracy but expose a critical flaw: attention shifts from foreground to irrelevant background regions. (b) Quantitative comparison of IoU(\\%) between attention regions and the true foreground and background regions using UIAT [9] and AT. Increasing inverse attack strength " + }, + { + "bbox": [ + 313, + 385, + 555, + 518 + ], + "type": "inline_equation", + "content": "-\\epsilon" + }, + { + "bbox": [ + 313, + 385, + 555, + 518 + ], + "type": "text", + "content": " demonstrates UIAT's disproportionate focus on non-causal background features, exposing reliance on spurious correlations [3, 28, 33]." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 413, + 240, + 553, + 365 + ], + "blocks": [ + { + "bbox": [ + 413, + 240, + 553, + 365 + ], + "lines": [ + { + "bbox": [ + 413, + 240, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 413, + 240, + 553, + 365 + ], + "type": "image", + "image_path": "4b1a5bc78731f31768cd4a98b8596db93676016165cb0860f3daa8fca165af52.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 367, + 496, + 376 + ], + "lines": [ + { + "bbox": [ + 486, + 367, + 496, + 376 + ], + "spans": [ + { + "bbox": [ + 486, + 367, + 496, + 376 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 532, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 555, + 568 + ], + "type": "text", + "content": "ognized as one of the most effective methods for defending against these attacks by incorporating adversarial examples into the training process [14, 34]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 556, + 715 + ], + "type": "text", + "content": "Recently, numerous advanced AT-based techniques have been proposed to address the vulnerabilities of DNNs [18, 23, 24, 38]. Prominent methods in this domain focus on aligning the distributions of adversarial examples to their true classes, thereby enhancing classification accuracy under adversarial attacks. For example, MART [41] and TRADES [49] aim to align the logits of adversarial examples with their corresponding natural examples, promoting consistency in predictions. Expanding on this, UIAT [9] and ACR [5] extend this approach by generating inverse adversarial examples [30] that exhibit higher confidence compared to their natural counterparts. This strategy helps al" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2783" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "leviate the negative impact of misclassification of natural examples in alignment and directs adversarial examples toward regions with correct high-confidence classifications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 314 + ], + "type": "text", + "content": "Upon investigation, we uncover a remarkable yet concerning phenomenon: inverse adversarial examples produce high-confidence outputs by triggering biased feature activation. To illuminate this effect, we visualize attention maps from a conventional adversarial training (AT) model under both adversarial and inverse adversarial attacks (see Figure 1 (a)). Adversarial attacks disrupt normal attention patterns, causing incorrect predictions. In contrast, inverse adversarial attacks amplify the model's attention on features, with stronger attacks even improving prediction accuracy [9, 43]. However, this apparent enhancement masks a critical flaw: the model's attention systematically shifts from discriminative foreground features to irrelevant background regions—akin to identifying a sheep by detecting grass rather than the animal itself. This reliance on spurious correlations undermines the model's causal reasoning capabilities and compromises its robustness [1, 6, 19]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 315, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 506 + ], + "type": "text", + "content": "To delve deeper into this issue, we conduct a statistical analysis using the Intersection over Union (IoU) metric to measure the alignment between attention maps and foreground/background regions on a subset of ImageNet [17]. Leveraging the Class Activation Mapping (CAM)-based technique inspired by [10], we define foreground as regions with high-class activation (i.e., areas critical for predictions) and background as regions with low activation. Comparing the UIAT model [9] (trained with inverse adversarial examples) to a standard AT model reveals a concerning trend. As evidenced in Figure 1 (b), while increasing the magnitude " + }, + { + "bbox": [ + 55, + 315, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 315, + 295, + 506 + ], + "type": "text", + "content": " of inverse adversarial attacks enhances overall feature attention, the UIAT model fails to improve IoU for foreground features compared to the AT model. Instead, it demonstrates a marked increase in IoU for background features—a clear indicator of attention misallocation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 509, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 628 + ], + "type": "text", + "content": "This leads to a pivotal insight: the inverse adversarial training disproportionately skews attention toward background features, resulting in biased feature activation. This shift engenders a fundamental vulnerability: a spurious correlation bias where the model becomes overly dependent on contextual background cues rather than intrinsic object features. Much like a classifier that associates beaches with birds rather than wings and beaks, this misplaced reliance significantly degrades robustness and generalization performance [15, 50], as conclusively demonstrated in Table 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 714 + ], + "type": "text", + "content": "Building upon the above findings, we introduce a method named Debiased High-Confidence Adversarial Training (DHAT). DHAT implements two key techniques: Debiased High-Confidence Logit Regularization (DHLR) and Foreground Logit Orthogonal Enhancement (FLOE). DHLR quantifies the bias towards background feature activation and debiase the spurious correlations by recalibrating the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "biased high-confidence logits derived from inverse adversarial examples. This regularization aligns the logits of adversarial examples with debiased high-confidence logits to improve adversarial robustness and mitigate spurious correlation bias. FLOE further refines this process by reducing the correlation between high-confidence logits and background features in affine space, which helps restore the model's attention to its normal state. Extensive experiments demonstrate that DHAT outperforms state-of-the-art techniques across various datasets, providing superior adversarial robustness and generalization. Additionally, our DHAT can be seamlessly integrated with existing advanced adversarial training methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 228, + 512, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 228, + 512, + 239 + ], + "spans": [ + { + "bbox": [ + 326, + 228, + 512, + 239 + ], + "type": "text", + "content": "Below, we outline our principal contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 241, + 553, + 408 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 241, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 241, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 241, + 553, + 300 + ], + "type": "text", + "content": "- We identify that the adversarial training model shows a biased feature activation under inverse adversarial attacks. Training with inverse adversarial examples causes the model's attention shifting towards background features, thus lead to spurious correlation bias." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 361 + ], + "type": "text", + "content": "- We introduce a novel method, Debiased High-Confidence Adversarial Training (DHAT). This method aligns the logits of adversarial examples with debiased high-confidence logits corresponding to their classes, effectively mitigating spurious correlation bias." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 361, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 553, + 408 + ], + "type": "text", + "content": "- Extensive experiments validate the superior robustness and generalization of our method against adversarial attacks. Additionally, our approach integrates smoothly with existing advanced adversarial training strategies." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 421, + 400, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 400, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 400, + 433 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 441, + 426, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 426, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 426, + 453 + ], + "type": "text", + "content": "2.1. Adversarial Attack" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 459, + 553, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 553, + 507 + ], + "type": "text", + "content": "Adversarial attacks introduce subtle perturbations " + }, + { + "bbox": [ + 313, + 459, + 553, + 507 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 459, + 553, + 507 + ], + "type": "text", + "content": " into natural examples " + }, + { + "bbox": [ + 313, + 459, + 553, + 507 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 459, + 553, + 507 + ], + "type": "text", + "content": ", causing DNNs to produce incorrect predictions rather than the true class " + }, + { + "bbox": [ + 313, + 459, + 553, + 507 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 459, + 553, + 507 + ], + "type": "text", + "content": ". These attacks are typically formulated as an optimization problem:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 364, + 517, + 553, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 517, + 553, + 537 + ], + "spans": [ + { + "bbox": [ + 364, + 517, + 553, + 537 + ], + "type": "interline_equation", + "content": "\\hat {x} = \\max _ {| | \\delta | | _ {p} \\leq \\epsilon} \\left(\\mathcal {L} _ {\\text {a t t a c k}} (\\theta ; x + \\delta , y)\\right), \\tag {1}", + "image_path": "27887e0dcca338a1adac5d38c8e6b0e9205adf96bef3fda16648ebe102113000.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "text", + "content": " represents the adversarial example, " + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{attack}" + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "text", + "content": " is the adversarial attack loss function, " + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "text", + "content": " denotes the model parameters, and " + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "text", + "content": " is constrained to the " + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "text", + "content": "-norm ball of radius " + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 547, + 553, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 583, + 553, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 583, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 583, + 553, + 643 + ], + "type": "text", + "content": "Numerous studies have investigated the vulnerability of DNNs and have proposed various adversarial attack methods. Goodfellow et al. [14] introduced the Fast Gradient Sign Method (FGSM), which compute the adversarial perturbation in the direction of the gradient:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 365, + 654, + 553, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 654, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 365, + 654, + 553, + 667 + ], + "type": "interline_equation", + "content": "\\delta = \\eta \\cdot \\operatorname {s i g n} \\left(\\nabla_ {x} \\mathcal {L} _ {\\text {a t t a c k}} (\\theta ; x, y)\\right), \\tag {2}", + "image_path": "f82bc219d2452aa5c2a69ef608c3f4636bbd8ef82c1c1d14c94cc9f14ad07778.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " denotes the attack step size. Building on this, Madry et al. [27] proposed Projected Gradient Descent (PGD), which iteratively updates the gradient of the input" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2784" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "data to generate more potent adversarial examples, thereby increasing the success rate of attacks:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 101, + 295, + 117 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 101, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 63, + 101, + 295, + 117 + ], + "type": "interline_equation", + "content": "\\hat {x} ^ {t + 1} = \\Pi_ {\\mathbb {B} (\\epsilon)} \\{\\hat {x} ^ {t} + \\eta \\cdot \\operatorname {s i g n} (\\nabla_ {\\hat {x} ^ {t}} \\mathcal {L} _ {\\text {a t t a c k}} (\\theta ; \\hat {x} ^ {t}, y)) \\}, \\tag {3}", + "image_path": "11b37b193cb9e1ad4bd8a7c4bd0c6e3eaf0a8e243e18408f1d3108bb9f72abe0.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\Pi_{\\mathbb{B}(\\epsilon)}" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": " projects the perturbation onto the " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": "-norm ball of radius " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": ". Furthermore, Carlini and Wagner [4] developed the Carlini-Wagner (C&W) attack, an optimization-based method capable of producing more imperceptible adversarial examples. Croce and Hein [7] introduced AutoAttack, an ensemble attack strategy that is user-independent and parameter-free." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 211, + 173, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 173, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 173, + 223 + ], + "type": "text", + "content": "2.2. Adversarial Defense" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "content": "Adversarial training (AT) is a widely used strategy for improving model robustness. The core principle of AT is to include adversarial examples in the training process, thereby aligning the model's outputs towards the true label distribution [12, 45]. Mathematically, conventional AT methods are formulated as a min-max optimization problem:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 305, + 295, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 305, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 91, + 305, + 295, + 326 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} \\mathbb {E} _ {(x, y) \\sim \\mathcal {D}} \\max _ {| | \\delta | | _ {p} \\leq \\epsilon} \\mathcal {L} _ {A T} (\\theta ; \\mathcal {G} _ {\\theta} (x, \\delta), y), \\tag {4}", + "image_path": "7efb40d9c7e78ea4619ed721f1b922826418c3b8410d0aa172fad0949b8c228a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "text", + "content": " denotes a clean image-label pair sampled from the data distribution " + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "text", + "content": " is constrained to the maximum " + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "text", + "content": "-norm magnitude " + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AT}" + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "text", + "content": " represents the training strategy function, and " + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\theta}(x,\\delta)" + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "text", + "content": " denotes the attack generation procedure incorporating the perturbation " + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "text", + "content": ". The outer minimization optimizes the model parameters " + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 331, + 296, + 437 + ], + "type": "text", + "content": " to minimize the expected loss over adversarial examples, while the inner maximization identifies the perturbations that maximize this loss, thereby enhancing the model's robustness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 438, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 594 + ], + "type": "text", + "content": "Recent advancements in AT have introduced several techniques to further improve adversarial robustness. Wu et al. [44] introduced Adversarial Weight Perturbation (AWP), which enhances robustness by flattening the loss landscape with respect to input variations. Kim et al. [20] proposed Feature Separation and Recalibration (FSR) to capture useful logits from non-robust features. Wei et al. [42] developed Class-wise Calibrated Fair Adversarial Training (CFA), which adaptively configures training for each class to achieve category-level robust fairness. Li et al. [26] proposed SGLR, which uses self-distillation to refine the soft-label distribution logits, thereby calibrating the adversarial training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Methods similar to our approach include MART [41] and UIAT [9]. These methods use Kullback-Leibler (KL) divergence to align the logits of adversarial examples with high-confidence logits. MART encourages sharing a similar logit distribution between adversarial examples and their corresponding natural examples. UIAT generates inverse adversarial examples [30] that exhibit higher confidence compared to natural examples, thereby enhancing the alignment of adversarial examples towards the high-confidence regions of the decision surface." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 555, + 196 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 555, + 196 + ], + "type": "image", + "image_path": "71b56ad4483a89b149d57fd64cd126af45e34fc490f06b0eb7c868b3f9f88469.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 202, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 313, + 202, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 555, + 258 + ], + "type": "text", + "content": "Figure 2. Overview of Debiased High-Confidence Logits Regularization (DHLR). The method first quantifies the degree of biased activation towards background features and recalibrates biased logits, then aligns the logits of adversarial examples with biased high-confidence logits." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 278, + 398, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 278, + 398, + 293 + ], + "spans": [ + { + "bbox": [ + 314, + 278, + 398, + 293 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 299, + 555, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 371 + ], + "type": "text", + "content": "In this section, we introduce the basic notation and present our approach to mitigate adversarial bias and improve robustness. Our method recalibrates high-confidence logits and reduces reliance on spurious background features. The optimization strategy is outlined, with pseudocode provided in Section 7 of the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 379, + 378, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 379, + 378, + 391 + ], + "spans": [ + { + "bbox": [ + 314, + 379, + 378, + 391 + ], + "type": "text", + "content": "3.1. Notation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "text", + "content": "We denote the training dataset as " + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "inline_equation", + "content": "(X,Y) = \\{(x_{i},y_{i})|i\\in (1,2,\\ldots ,N)\\}" + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "text", + "content": " denote the natural input examples and their corresponding labels, respectively. " + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "text", + "content": " represents the total number of examples in the dataset. Let " + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "text", + "content": " denote the parameters of a convolutional neural network (CNN) for image classification. The network's output logits are denoted as " + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "text", + "content": ", and the final predicted label is given by " + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\arg \\max f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "text", + "content": ". Inverse adversarial examples " + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 313, + 396, + 555, + 540 + ], + "type": "text", + "content": " can be regarded as examples for which the model correctly predicts the label with high-confidence. These examples are generated by applying image perturbations aimed at minimizing the attack loss [30]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 370, + 550, + 555, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 550, + 555, + 571 + ], + "spans": [ + { + "bbox": [ + 370, + 550, + 555, + 571 + ], + "type": "interline_equation", + "content": "\\check {x} = \\min _ {| | \\delta | | _ {p} \\leq \\epsilon} \\left(\\mathcal {L} _ {\\mathrm {I n v}} (\\theta ; x - \\delta , y)\\right), \\tag {5}", + "image_path": "a750f29994d2f554815390d033f3e0f57fcd30a49fd8dfaef600d27c0997f620.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Inv}}" + }, + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "type": "text", + "content": " denotes the loss function used to generate inverse adversarial attack. The logits of the inverse adversarial example " + }, + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "type": "text", + "content": " and the adversarial example " + }, + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "type": "text", + "content": " are computed as " + }, + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{z} = f_{\\theta}(\\tilde{x})" + }, + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\hat{z} = f_{\\theta}(\\hat{x})" + }, + { + "bbox": [ + 313, + 579, + 555, + 629 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 635, + 554, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 554, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 554, + 659 + ], + "type": "text", + "content": "3.2. Debiased High-Confidence Logit Regularization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "To mitigate the spurious correlation bias introduced by adversarial training with inverse adversarial examples, we propose utilizing debiased high-confidence logits for regularization. Specifically, this approach involves separating and" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2785" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "quantifying the spurious bias towards background features under inverse adversarial attacks, and then recalibrating the biased high-confidence logits to align with those of adversarial examples, as illustrated in Figure 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "text", + "content": "First, we quantify the model's biased activation degree towards background features under inverse adversarial attacks. Following the approach proposed by [10], we use using Grad-CAM [31] or alternative methods such as SAM [21] to generate pixel-wise attention maps " + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "inline_equation", + "content": "M \\in [0,1]^{n \\times n}" + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "text", + "content": " for the natural example " + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "text", + "content": ". We separate the background feature " + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "inline_equation", + "content": "\\tilde{x}_{(B)}" + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "text", + "content": " from the inverse adversarial examples " + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "inline_equation", + "content": "\\tilde{x}" + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "text", + "content": " based on the " + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 121, + 295, + 215 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 112, + 220, + 295, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 220, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 112, + 220, + 295, + 236 + ], + "type": "interline_equation", + "content": "[ \\check {x} _ {(B)} ] _ {(i, j)} = \\mathbb {I} _ {\\left(M _ {i, j} < \\omega\\right)} \\cdot \\check {x} _ {(i, j)}, \\tag {6}", + "image_path": "f056d2c55f5b86588e5f129dff1223748942cba4e1a602d7b04159f78206bb77.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 239, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 294, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 239, + 294, + 275 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_{(M_{i,j} < \\omega)}" + }, + { + "bbox": [ + 55, + 239, + 294, + 275 + ], + "type": "text", + "content": " is an indicator function and " + }, + { + "bbox": [ + 55, + 239, + 294, + 275 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 239, + 294, + 275 + ], + "type": "text", + "content": " is a predefined threshold. Next, we compute the logits for the background feature " + }, + { + "bbox": [ + 55, + 239, + 294, + 275 + ], + "type": "inline_equation", + "content": "\\check{x}_{(B)}" + }, + { + "bbox": [ + 55, + 239, + 294, + 275 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 139, + 280, + 295, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 280, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 139, + 280, + 295, + 294 + ], + "type": "interline_equation", + "content": "\\check {z} _ {(B)} = f _ {\\theta} \\left(\\check {x} _ {(B)}\\right). \\tag {7}", + "image_path": "f8cc96f40594b62efa4b76e408aaaa39d4c91616e4d2d98adb7bb35a6bba5c73.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 297, + 294, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 294, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 294, + 321 + ], + "type": "text", + "content": "These logits can quantify the biased degree introduced by background feature activation during network inference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 322, + 294, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 294, + 357 + ], + "type": "text", + "content": "To recalibrate the biased high-confidence logits, we refine the logits of inverse adversarial examples by subtracting the background feature logits:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 363, + 295, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 363, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 143, + 363, + 295, + 376 + ], + "type": "interline_equation", + "content": "\\check {z} ^ {*} = \\check {z} - \\check {z} _ {(B)}, \\tag {8}", + "image_path": "32f85bd58cc1b3dee7247fba44d1fb860f1f94b5bfb2de2279d661651bb4621f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 380, + 284, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 284, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 284, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 380, + 284, + 392 + ], + "type": "inline_equation", + "content": "\\check{z}^*" + }, + { + "bbox": [ + 55, + 380, + 284, + 392 + ], + "type": "text", + "content": " represents the debiased high-confidence logits." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 392, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 464 + ], + "type": "text", + "content": "Next, we introduce a regularizer to align the logits of adversarial examples with these debiased high-confidence logits. This regularization encourages the model to focus on foreground features and mitigates the biased activation of background features. Specifically, our Debiased High-Confidence Logit Regularization (DHLR) is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 97, + 469, + 294, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 469, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 97, + 469, + 294, + 483 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D H L R} \\left(\\check {z} ^ {*}, \\hat {z}\\right) = \\mathcal {L} _ {K L} \\left(\\phi \\left(\\check {z} ^ {*}\\right) | | \\phi (\\hat {z})\\right), \\tag {9}", + "image_path": "2af43a56d767c3a88be174f877b680fa32e6bceb1838bc12e5f01ffda5dfeae0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 487, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 487, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 487, + 295, + 510 + ], + "type": "text", + "content": " denotes the softmax function, " + }, + { + "bbox": [ + 55, + 487, + 295, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{KL}" + }, + { + "bbox": [ + 55, + 487, + 295, + 510 + ], + "type": "text", + "content": " represents the Kullback-Leibler (KL) divergence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 511, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 295, + 571 + ], + "type": "text", + "content": "This regularization term " + }, + { + "bbox": [ + 55, + 511, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DHLR}" + }, + { + "bbox": [ + 55, + 511, + 295, + 571 + ], + "type": "text", + "content": " promotes consistency between the logits of adversarial examples " + }, + { + "bbox": [ + 55, + 511, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 55, + 511, + 295, + 571 + ], + "type": "text", + "content": " and the debiased high-confidence logits " + }, + { + "bbox": [ + 55, + 511, + 295, + 571 + ], + "type": "inline_equation", + "content": "\\tilde{z}^*" + }, + { + "bbox": [ + 55, + 511, + 295, + 571 + ], + "type": "text", + "content": ". This alignment mitigates spurious correlation bias, thereby improving adversarial robustness and generalization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 576, + 286, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 286, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 286, + 589 + ], + "type": "text", + "content": "3.3. Foreground Logit Orthogonal Enhancement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "While DHLR effectively recalibrates high-confidence logits to mitigate spurious correlations, it fails to fully restore the model's focus on foreground features when handling inverse adversarial examples " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\check{x}" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": ". This is because DHLR primarily aligns logits with debiased high-confidence targets, without directly addressing the model's persistent bias toward background features. Under inverse adversarial attacks, the model still exhibits strong activation in irrelevant background regions, undermining its ability to prioritize critical foreground cues." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "To address this issue, we introduce Foreground Logit Orthogonal Enhancement (FLOE), a targeted method designed to minimize the correlation between high-confidence logits " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " and background feature logits " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{(B)}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": ". The intuition behind FLOE is straightforward: by reducing the projection of " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " onto " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{(B)}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " in the affine space, we make " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " less explainable by " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\tilde{z}_{(B)}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": ", thereby encouraging the model to rely more on foreground features. This projection-based approach ensures that the model's predictions are driven by meaningful foreground information rather than spurious background correlations. The optimization objective is formulated as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 336, + 212, + 553, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 212, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 336, + 212, + 553, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {F L O E} \\left(\\check {z}, \\check {z} _ {(B)}\\right) = - \\left| \\check {z} - \\frac {\\check {z} \\cdot \\check {z} _ {(B)}}{\\left| \\check {z} _ {(B)} \\right| ^ {2}} \\cdot \\check {z} _ {(B)} \\right| _ {p}, \\tag {10}", + "image_path": "96f06f51144b541596f472959c4b576c185893eef53e0f413d9d719fd2b0f727.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 247, + 553, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 247, + 553, + 282 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 247, + 553, + 282 + ], + "type": "text", + "content": " indicates the norm exponent. The term " + }, + { + "bbox": [ + 313, + 247, + 553, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{FLOE}" + }, + { + "bbox": [ + 313, + 247, + 553, + 282 + ], + "type": "text", + "content": " is designed to alleviate biased feature activation under inverse adversarial attacks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 291, + 441, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 441, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 441, + 304 + ], + "type": "text", + "content": "3.4. Optimization Strategy" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 308, + 553, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 553, + 355 + ], + "type": "text", + "content": "Finally, we integrate the DHLR and FLOE techniques into a unified training framework. Debiased High-Confidence Adversarial Training (DHAT). The overall optimization objective for DHAT is formulated as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 336, + 363, + 553, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 363, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 336, + 363, + 553, + 392 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {D H A T} = \\mathcal {L} _ {A T} (\\hat {z}, y) + \\lambda_ {1} \\cdot \\mathcal {L} _ {D H L R} (\\hat {z} ^ {*}, \\hat {z}) \\tag {11} \\\\ + \\lambda_ {2} \\cdot \\mathcal {L} _ {F L O E} (\\tilde {z}, \\tilde {z} _ {(B)}), \\\\ \\end{array}", + "image_path": "cd355a5dd2c0353c0b60a6182ac3d3529367c3902f8bcfc74851889e9829f86c.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AT}(\\hat{z},y)" + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "text", + "content": " represents the standard adversarial training loss, and " + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "text", + "content": " are hyperparameters that control the relative contributions of the DHLR and FLOE terms, respectively. Besides, our DHAT can be seamlessly integrated with various advanced adversarial training strategies represented by " + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{AT}" + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "text", + "content": " to improve the robustness." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 484, + 395, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 395, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 395, + 498 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 504, + 441, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 441, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 441, + 517 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 521, + 432, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 432, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 432, + 532 + ], + "type": "text", + "content": "4.1.1. Datasets and models." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 536, + 553, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 553, + 609 + ], + "type": "text", + "content": "We conduct experiments on the CIFAR-10, CIFAR-100 [22], and ImageNet-1K [29] datasets. The architecture for our experiments include the mainstream models WRN28-10 [47], ResNet-18 [16], ResNet-50 [16], Inception-V3 [40], and VGG16 [36]. Detailed information is provided in the supplementary material." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 615, + 405, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 405, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 405, + 627 + ], + "type": "text", + "content": "4.1.2. Training setup" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "For CIFAR datasets, each model is trained for 100 epochs, with the learning rate reduced by a factor of 0.1 at the 80th and 90th epochs. For adversarial example generation, we use a fixed 10-iteration attack. The maximum " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": "-norm of the adversarial perturbation is set to " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": ", while the inverse adversarial perturbation is set to " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": ". Both perturbations use a fixed step size of " + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = 2 / 255" + }, + { + "bbox": [ + 313, + 629, + 553, + 713 + ], + "type": "text", + "content": ". For the" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2786" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 91, + 512, + 504 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 534, + 82 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 534, + 82 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 534, + 82 + ], + "type": "text", + "content": "Table 1. Comparison of robustness (%) and robust generalization gap (%) . The number in bold indicates the best performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 91, + 512, + 504 + ], + "lines": [ + { + "bbox": [ + 93, + 91, + 512, + 504 + ], + "spans": [ + { + "bbox": [ + 93, + 91, + 512, + 504 + ], + "type": "table", + "html": "
Architecture: WRN28-10\nImageNet-1KAttack (ε = 4/255)Robust Gap↓
Clean↑PGD-10↑PGD-20↑PGD-50↑C&W↑AA↑
MART [41]62.3143.5642.8942.5440.6238.9413.76
AWP [44]64.2545.1344.6744.2141.9640.0212.82
FSR [20]64.7044.3143.8243.4541.2339.3014.17
CFA [42]65.3844.8744.3243.9641.5839.6515.49
UIAT [9]62.6445.2944.8544.5742.1340.1814.68
SGLR [26]63.4644.5244.0543.7841.3639.4719.72
DHAT (Ours)65.9046.8346.4246.1143.2541.709.53
DHAT-AWP (Ours)65.6247.3847.0446.7943.6242.0710.27
DHAT-CFA (Ours)66.2648.2747.8947.5344.1842.4511.64
Architecture: WRN28-10\nCIFAR-10Attack (ε = 8/255)Robust Gap↓
Clean↑PGD-10↑PGD-20↑PGD-50↑C&W↑AA↑
MART [41]82.9956.2555.4855.4552.2650.679.52
AWP [44]82.6757.8057.2157.0754.8251.906.90
FSR [20]82.9256.6955.9455.5153.9351.747.42
CFA [42]84.4357.8756.9056.6454.6051.8510.36
UIAT [9]82.9458.6658.1258.0554.1152.177.92
SGLR [26]85.7657.5356.9156.6654.2852.079.38
DHAT (Ours)83.9560.4959.9559.8755.2753.103.51
DHAT-AWP (Ours)83.2161.6161.3561.2755.7153.694.54
DHAT-CFA (Ours)84.4962.6762.3862.2255.9554.056.33
Architecture: WRN28-10\nCIFAR-100Attack (ε = 8/255)Robust Gap↓
Clean↑PGD-10↑PGD-20↑PGD-50↑C&W↑AA↑
MART [41]54.6932.0631.9031.8828.7727.259.96
AWP [44]57.9434.0133.7533.7230.7428.907.87
FSR [20]57.4832.9332.3032.2629.1627.047.84
CFA [42]60.9233.1032.5632.4130.4928.0410.47
UIAT [9]57.6534.2733.9133.8530.9729.0311.70
SGLR [26]61.0233.4332.9832.8230.7228.5015.67
DHAT (Ours)59.1435.8235.3335.0231.7230.174.24
DHAT-AWP (Ours)59.3036.3936.1535.9832.0530.384.85
DHAT-CFA (Ours)61.5437.6737.1536.9932.4030.935.93
", + "image_path": "bf755635fc4e570063477661b011d38d7e8d76fcbdb2a9860a0e54829edf8915.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 523, + 294, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 294, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 294, + 583 + ], + "type": "text", + "content": "ImageNet dataset, we adopt the approach in [37] with 2-iteration PGD over 50 epochs, and set " + }, + { + "bbox": [ + 55, + 523, + 294, + 583 + ], + "type": "inline_equation", + "content": "\\epsilon = 4 / 255" + }, + { + "bbox": [ + 55, + 523, + 294, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 523, + 294, + 583 + ], + "type": "inline_equation", + "content": "\\alpha = 1 / 255" + }, + { + "bbox": [ + 55, + 523, + 294, + 583 + ], + "type": "text", + "content": " for adversarial perturbations. The hyperparameters " + }, + { + "bbox": [ + 55, + 523, + 294, + 583 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 523, + 294, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 523, + 294, + 583 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 523, + 294, + 583 + ], + "type": "text", + "content": " are set to 1.0 in all data sets. Comprehensive training details are available in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 590, + 157, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 157, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 157, + 602 + ], + "type": "text", + "content": "4.1.3. Evaluation setup." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": "For robustness performance evaluation, we employ PGD [27], C&W [4], and AutoAttack (AA) [7] attacks under the " + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": " norm. AutoAttack comprises APGD-DLR [7], APGD-CE [7], FAB [8], and Square [2]. \"Clean\" denotes natural examples without adversarial perturbations. Additionally, we introduce the concept of the robust generalization gap, referred to as the \"Robust Gap\", which measures the difference in robustness performance between the training and test sets as a direct indicator of the generalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 523, + 554, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 554, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 554, + 560 + ], + "type": "text", + "content": "ability [46]. This metric quantifies the strength of spurious correlations, where larger gaps indicate greater reliance on features that lack causal relationships with target labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 567, + 433, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 433, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 433, + 578 + ], + "type": "text", + "content": "4.2. Experimental Result" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 554, + 644 + ], + "type": "text", + "content": "In this part, we conducted comprehensive evaluations of our proposed method across various visual datasets and compared it with mainstream adversarial training (AT) methods. Additionally, DHAT can be integrated with advanced AT strategies. The results are shown in Table 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 651, + 432, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 432, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 432, + 662 + ], + "type": "text", + "content": "4.2.1. Robust performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "DHAT consistently demonstrates superior robustness across all datasets. For example, on CIFAR-10, DHAT outperforms the second-best method by " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "1.93\\%" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "1.16\\%" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " under PGD-10 and C&W attacks, respectively. DHAT also" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2787" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 91, + 541, + 232 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 536, + 83 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 536, + 83 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 536, + 83 + ], + "type": "text", + "content": "Table 2. Comparison of robustness (%) and robust generalization gap (%). The number in bold indicates the best performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 91, + 541, + 232 + ], + "lines": [ + { + "bbox": [ + 71, + 91, + 541, + 232 + ], + "spans": [ + { + "bbox": [ + 71, + 91, + 541, + 232 + ], + "type": "table", + "html": "
Dataset: CIFAR-10\nMethodResNet-50VGG-16Inception-V3
Clean↑AA↑Robust Gap↓Clean↑AA↑Robust Gap↓Clean↑AA↑Robust Gap↓
MART [41]75.0549.054.5766.7244.273.6076.0549.595.91
AWP [44]75.5950.803.9167.1445.973.4776.7450.464.07
FSR [20]78.3850.593.9769.5944.903.3077.2850.024.73
CFA [42]77.8850.644.2770.4545.534.6676.7049.974.46
UIAT [9]78.4251.002.9968.4645.273.5775.1951.233.70
SGLR [26]80.3950.555.2170.9045.043.9480.6250.735.00
DHAT (Ours)79.9451.921.8570.0146.401.3477.8551.811.92
DHAT-AWP (Ours)76.8452.052.3467.4147.212.1877.1252.082.88
DHAT-CFA (Ours)80.9752.382.0970.5247.832.2178.5252.672.44
", + "image_path": "05098f207c7da2f937fa014021b4462ade9ec239115cc946f55af6d99cae8cb3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "content": "shows the best performance on more complex datasets such as CIFAR-100 and high-resolution Imagenette datasets. Specifically, on ImageNet-1K, DHAT achieves improvements of " + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "inline_equation", + "content": "1.54\\%" + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "inline_equation", + "content": "1.52\\%" + }, + { + "bbox": [ + 55, + 250, + 296, + 310 + ], + "type": "text", + "content": " over the second-best method under PGD-10 and AA attacks, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 323, + 226, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 226, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 226, + 333 + ], + "type": "text", + "content": "4.2.2. Combine with other AT methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 339, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 339, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 339, + 296, + 495 + ], + "type": "text", + "content": "DHAT can be effectively combined with various AT methods to further enhance adversarial performance. Notably, DHAT can be combined with AWP and CFA, denoted as DHAT-AWP and DHAT-CFA, respectively. However, DHAT is not suitable for combining with methods that rely on spurious correlation to improve performance (e.g., FSR, SLGR). In particular, DHAT-CFA achieves superior performance across all datasets. For instance, on Imagenette, DHAT-CFA provides a " + }, + { + "bbox": [ + 54, + 339, + 296, + 495 + ], + "type": "inline_equation", + "content": "2.84\\%" + }, + { + "bbox": [ + 54, + 339, + 296, + 495 + ], + "type": "text", + "content": " performance improvement under the AA attack compared to the second-best method. Moreover, combining DHAT with advanced AT surpasses the performance of the original baseline methods in terms of both robustness and clean accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 507, + 163, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 163, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 163, + 520 + ], + "type": "text", + "content": "4.2.3. Robust overfitting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 524, + 296, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 296, + 633 + ], + "type": "text", + "content": "Our method significantly reduces the robust generalization gap across all evaluated datasets, effectively mitigating the robust overfitting phenomenon. For example, on CIFAR-10, DHAT reduces the robust generalization gap by " + }, + { + "bbox": [ + 55, + 524, + 296, + 633 + ], + "type": "inline_equation", + "content": "4.41\\%" + }, + { + "bbox": [ + 55, + 524, + 296, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 524, + 296, + 633 + ], + "type": "inline_equation", + "content": "6.01\\%" + }, + { + "bbox": [ + 55, + 524, + 296, + 633 + ], + "type": "text", + "content": " compared to UIAT and MART, respectively. This reduction highlights DHAT's ability to mitigate spurious correlation bias, thereby improving the model's overall generalization performance. Detailed visualizations of the learning curve are provided in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 646, + 190, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 190, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 190, + 658 + ], + "type": "text", + "content": "4.3. Generalization Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "In this part, we evaluate the generalizability of our proposed method by comprehensively evaluating its performance across various network architectures and under different magnitudes of adversarial attacks." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 249, + 435, + 364 + ], + "blocks": [ + { + "bbox": [ + 318, + 249, + 435, + 364 + ], + "lines": [ + { + "bbox": [ + 318, + 249, + 435, + 364 + ], + "spans": [ + { + "bbox": [ + 318, + 249, + 435, + 364 + ], + "type": "image", + "image_path": "916c0ac27ef8715d57b26358cafee580c3f236249463b5ac8a16e47dd6b30019.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 370, + 411, + 380 + ], + "lines": [ + { + "bbox": [ + 349, + 370, + 411, + 380 + ], + "spans": [ + { + "bbox": [ + 349, + 370, + 411, + 380 + ], + "type": "text", + "content": "(a) PGD-10 Attack" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 391, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 413 + ], + "type": "text", + "content": "Figure 3. Comparisons of robustness " + }, + { + "bbox": [ + 313, + 391, + 555, + 413 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 391, + 555, + 413 + ], + "type": "text", + "content": " under varying " + }, + { + "bbox": [ + 313, + 391, + 555, + 413 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 391, + 555, + 413 + ], + "type": "text", + "content": " adversarial attacks using ResNet-18 on CIFAR-10." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 436, + 250, + 551, + 364 + ], + "blocks": [ + { + "bbox": [ + 436, + 250, + 551, + 364 + ], + "lines": [ + { + "bbox": [ + 436, + 250, + 551, + 364 + ], + "spans": [ + { + "bbox": [ + 436, + 250, + 551, + 364 + ], + "type": "image", + "image_path": "71e8fa34bfc78b2ee8bcc7f3b6350ccd91cd8b538282b4c34f307a7223ca6975.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 370, + 526, + 380 + ], + "lines": [ + { + "bbox": [ + 471, + 370, + 526, + 380 + ], + "spans": [ + { + "bbox": [ + 471, + 370, + 526, + 380 + ], + "type": "text", + "content": "(b) C&W Attack" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 429, + 545, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 545, + 440 + ], + "type": "text", + "content": "4.3.1. Performance with diverse network architecture." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 445, + 555, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 565 + ], + "type": "text", + "content": "To evaluate the adaptability of our method across different network architectures, we applied DHAT on ResNet-50, VGG-16, and Inception-V3. The performance comparisons are reported in Table 2. The results demonstrate the DHAT's consistently superior robustness and generalization across a variety of model architectures. For example, DHAT-CFA outperforms the second-best method by " + }, + { + "bbox": [ + 313, + 445, + 555, + 565 + ], + "type": "inline_equation", + "content": "1.83\\%" + }, + { + "bbox": [ + 313, + 445, + 555, + 565 + ], + "type": "text", + "content": " under AA attacks on ResNet-50. Similarly, on VGG-16, DHAT significantly reduces the robustness generalization gap by " + }, + { + "bbox": [ + 313, + 445, + 555, + 565 + ], + "type": "inline_equation", + "content": "1.96\\%" + }, + { + "bbox": [ + 313, + 445, + 555, + 565 + ], + "type": "text", + "content": " compared to the second-best method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 577, + 460, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 460, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 460, + 589 + ], + "type": "text", + "content": "4.3.2. Different attack magnitude." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Figure 3 illustrates the performance of DHAT under varying magnitudes of PGD-10 and C&W attacks. Our findings reveal that DHAT maintains a significant performance advantage across attack magnitudes. Compared to other methods, DHAT exhibits improved stability and robustness, with minimal performance degradation even under more severe attacks. This consistent performance across varying attack levels underscores the effectiveness of our approach in mitigating the impact of adversarial perturbations. Further details are provided in the supplementary material (Section 7)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2788" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 101, + 284, + 211 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 94 + ], + "type": "text", + "content": "Table 3. Comparison of computational costs (# A100 GPU hours) using different adversarial training methods on CIFAR-10." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 101, + 284, + 211 + ], + "lines": [ + { + "bbox": [ + 70, + 101, + 284, + 211 + ], + "spans": [ + { + "bbox": [ + 70, + 101, + 284, + 211 + ], + "type": "table", + "html": "
MethodArchitecture
ResNet-18WRN28-10Inception-V3
MART [41]0.0500.2850.331
AWP [44]0.0480.2740.318
FSR [20]0.0510.2910.337
CFA [42]0.0450.2620.303
UIAT [9]0.0400.2450.283
SGLR [26]0.0390.2400.278
DHAT (Ours)0.0410.2480.287
", + "image_path": "6c28bb72f6d280c1f288e8122719debdae8c0444a1b0fac1a67ed631ea78cdda.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 67, + 267, + 280, + 326 + ], + "blocks": [ + { + "bbox": [ + 55, + 236, + 295, + 259 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 295, + 259 + ], + "type": "text", + "content": "Table 4. Comparison of robustness (\\%) and robust generalization gap (\\%). The number in bold indicates the best performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 267, + 280, + 326 + ], + "lines": [ + { + "bbox": [ + 67, + 267, + 280, + 326 + ], + "spans": [ + { + "bbox": [ + 67, + 267, + 280, + 326 + ], + "type": "table", + "html": "
DHLRFLOEClean↑PGD-10↑AA↑Robust Gap↓
78.0355.0348.353.84
✓78.7956.5649.471.98
✓78.5056.2249.732.06
✓✓79.2857.2150.250.81
", + "image_path": "77138da2ef72adfa790ac57d5975efc6bfa752a44b1b5d56da1ac0122413332a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 350, + 198, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 198, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 198, + 363 + ], + "type": "text", + "content": "4.4. Computational Efficiency" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 370, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 370, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 370, + 295, + 479 + ], + "type": "text", + "content": "The computational costs of our method and baseline models are provided in terms of per-epoch training time on a single NVIDIA A100 GPU with a batch size of 128 on the CIFAR-10 dataset, as shown in Table 3. Our method demonstrates significantly improved time efficiency compared to the baseline MART, AWP, FSR, and CFA. While it requires slightly more computational resources than the state-of-the-art methods UIAT and SGLR, this minor increase is justifiable given the performance gains achieved." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 494, + 162, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 162, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 162, + 506 + ], + "type": "text", + "content": "4.5. Ablation Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 539 + ], + "type": "text", + "content": "In this part, we conduct a comprehensive analysis of the impact of individual components and parameters of our model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 227, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 227, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 227, + 565 + ], + "type": "text", + "content": "4.5.1. Impact of individual components." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": "We evaluate the contributions of the individual components of DHAT, specifically Debiased High-Confidence Logit Regularization (DHLR) and Foreground Logit Orthogonal Enhancement (FLOE), as shown in Table 4. The results indicate that both DHLR and FLOE significantly enhance model performance compared to the UIAT method. They enhance clean accuracy and adversarial robustness, while also reducing the gap in robustness generalization. This highlights the effectiveness of DHLR and FLOE in mitigating spurious correlation biases. Notably, the integration of both DHLR and FLOE results in a more substantial overall improvement in model performance." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 429, + 163 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 429, + 163 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 429, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 429, + 163 + ], + "type": "image", + "image_path": "cac6452ce8fdd4fdfd37b168ff8b527a222f4fb60f3508e21988f18abf63416f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 170, + 399, + 179 + ], + "lines": [ + { + "bbox": [ + 356, + 170, + 399, + 179 + ], + "spans": [ + { + "bbox": [ + 356, + 170, + 399, + 179 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 356, + 170, + 399, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DHLR}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 417, + 71, + 550, + 164 + ], + "blocks": [ + { + "bbox": [ + 417, + 71, + 550, + 164 + ], + "lines": [ + { + "bbox": [ + 417, + 71, + 550, + 164 + ], + "spans": [ + { + "bbox": [ + 417, + 71, + 550, + 164 + ], + "type": "image", + "image_path": "e71e2437acc5f51a0dc6dbdf3b9e7c7dd3a02eef436a73ea4f70e790fb2ed842.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 169, + 514, + 179 + ], + "lines": [ + { + "bbox": [ + 473, + 169, + 514, + 179 + ], + "spans": [ + { + "bbox": [ + 473, + 169, + 514, + 179 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 473, + 169, + 514, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{FLOE}" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 230, + 430, + 323 + ], + "blocks": [ + { + "bbox": [ + 313, + 185, + 553, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 185, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 553, + 217 + ], + "type": "text", + "content": "Figure 4. Analysis of hyper-parameters " + }, + { + "bbox": [ + 313, + 185, + 553, + 217 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 185, + 553, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 185, + 553, + 217 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 185, + 553, + 217 + ], + "type": "text", + "content": " using ResNet-18 on CIFAR-10. The robust accuracy denotes the accuracy under PGD-10 attack." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 230, + 430, + 323 + ], + "lines": [ + { + "bbox": [ + 317, + 230, + 430, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 430, + 323 + ], + "type": "image", + "image_path": "7c3a4d2a24a52fb478104250c76b64b0327766b92ebb43a60af45c8eaf64bb7a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 329, + 399, + 338 + ], + "lines": [ + { + "bbox": [ + 352, + 329, + 399, + 338 + ], + "spans": [ + { + "bbox": [ + 352, + 329, + 399, + 338 + ], + "type": "text", + "content": "(a) CIFAR-10" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 344, + 555, + 378 + ], + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 378 + ], + "type": "text", + "content": "Figure 5. Analysis of hyper-parameters " + }, + { + "bbox": [ + 313, + 344, + 555, + 378 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 344, + 555, + 378 + ], + "type": "text", + "content": " using ResNet-18 on CIFAR-10 and Imagenette. The robust accuracy denotes the accuracy under PGD-10 attack." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 432, + 230, + 550, + 323 + ], + "blocks": [ + { + "bbox": [ + 432, + 230, + 550, + 323 + ], + "lines": [ + { + "bbox": [ + 432, + 230, + 550, + 323 + ], + "spans": [ + { + "bbox": [ + 432, + 230, + 550, + 323 + ], + "type": "image", + "image_path": "85e08e3ef216b6ab5b2d70937c646b3d911ceb622e51576184fd1d65acf419f6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 467, + 328, + 515, + 338 + ], + "lines": [ + { + "bbox": [ + 467, + 328, + 515, + 338 + ], + "spans": [ + { + "bbox": [ + 467, + 328, + 515, + 338 + ], + "type": "text", + "content": "(b) Imagenette" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 400, + 484, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 484, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 484, + 412 + ], + "type": "text", + "content": "4.5.2. Analysis of parameter sensitivity." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "text", + "content": "Figure 4 illustrates the sensitivity of the optimization parameters " + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "text", + "content": " in DHAT. Our analysis reveals that these parameters influence the trade-off between adversarial robustness and clean accuracy. As " + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "text", + "content": " increases, adversarial robustness improves, while clean accuracy tends to decrease. This occurs because higher values of " + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "text", + "content": " increase the regularization strength, which enhances the alignment of adversarial example logits with debiased high-confidence logits. However, this stronger regularization effect may also lead to a reduced emphasis on clean accuracy, as the model focuses more on handling adversarial perturbations. Conversely, increasing " + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "text", + "content": " generally improves clean accuracy but may adversely affect adversarial robustness. Larger values of " + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 415, + 555, + 617 + ], + "type": "text", + "content": " help the model restore its attention to the normal state, which can result in less effective alignment of adversarial example logits with the debiased high-confidence logits." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 626, + 553, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 553, + 649 + ], + "type": "text", + "content": "4.5.3. Analysis of threshold factor for quantifying spurious bias." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Figure 5 illustrates the impact of threshold values " + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": " on datasets with varying resolutions. Our analysis reveals that as the threshold value increases, both clean accuracy and robustness initially improve before sharply declining. This initial improvement occurs because higher thresholds help" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2789" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 281 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 281 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 281 + ], + "type": "image", + "image_path": "9e29c250f1f3638d19b9bca044e9c10b2fa46a1ebe54658e4c0bbc6a941a929a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 289, + 296, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 354 + ], + "type": "text", + "content": "Figure 6. Visualization of feature activation maps using UIAT and DHAT, utilizing Grad-CAM under inverse attack " + }, + { + "bbox": [ + 55, + 289, + 296, + 354 + ], + "type": "inline_equation", + "content": "(\\epsilon = -8)" + }, + { + "bbox": [ + 55, + 289, + 296, + 354 + ], + "type": "text", + "content": " and adversarial attack " + }, + { + "bbox": [ + 55, + 289, + 296, + 354 + ], + "type": "inline_equation", + "content": "(\\epsilon = 8)" + }, + { + "bbox": [ + 55, + 289, + 296, + 354 + ], + "type": "text", + "content": ". Left-to-right: the original image; normal feature activation maps; feature activation maps under adversarial attacks for UIAT and DHAT; and feature activation maps under inverse adversarial attacks for UIAT and DHAT." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 197, + 553, + 230 + ], + "lines": [ + { + "bbox": [ + 313, + 197, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 197, + 553, + 230 + ], + "type": "text", + "content": "Figure 7. IoU of attention regions with normal foreground and background regions. The comparison is conducted among conventional AT, UIAT, and DHAT on the Imagenette Dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 371, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 456 + ], + "type": "text", + "content": "in better separating accurate background features. However, further increasing the threshold eventually leads to the extraction of more foreground features, which causes significant performance degradation. For low-resolution datasets, the optimal threshold tends to be higher compared to high-resolution datasets, as high-resolution images more readily separate background features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 463, + 226, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 226, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 226, + 475 + ], + "type": "text", + "content": "4.6. Discussion DHAT Effectiveness" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 480, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 529 + ], + "type": "text", + "content": "In this part, we analyze the mechanisms underlying the effectiveness of our proposed method, Debiased High-Confidence Adversarial Training (DHAT), with a focus on its capacity to mitigate spurious correlation bias." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 224, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 224, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 224, + 546 + ], + "type": "text", + "content": "4.6.1. Feature Activation Map Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 548, + 295, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 645 + ], + "type": "text", + "content": "Figure 6 visualizes feature activation maps of UIAT and DHAT models under adversarial and inverse adversarial attacks. DHAT successfully maintains normal attention patterns, while UIAT exhibits biased activation with attention shifting significantly to background features during inverse attacks. DHAT achieves this robust performance by recalibrating biased high-confidence logits and systematically reducing correlation between logits and background features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 651, + 266, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 266, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 266, + 662 + ], + "type": "text", + "content": "4.6.2. Quantitative Evaluation of Bias Mitigation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Figure 7 quantifies DHAT's effectiveness through IoU between attention maps and true foreground/background regions. DHAT achieves higher foreground IoU and lower background IoU than both AT and UIAT, confirming its" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 434, + 175 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 434, + 175 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 434, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 434, + 175 + ], + "type": "image", + "image_path": "33666b8d58ca2952974707b1c0b62d3e93e4dbe803fc48724e26b48ae77ba0dc.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 180, + 429, + 190 + ], + "lines": [ + { + "bbox": [ + 322, + 180, + 429, + 190 + ], + "spans": [ + { + "bbox": [ + 322, + 180, + 429, + 190 + ], + "type": "text", + "content": "(a) IoU with Normal Foreground" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 433, + 71, + 549, + 175 + ], + "blocks": [ + { + "bbox": [ + 433, + 71, + 549, + 175 + ], + "lines": [ + { + "bbox": [ + 433, + 71, + 549, + 175 + ], + "spans": [ + { + "bbox": [ + 433, + 71, + 549, + 175 + ], + "type": "image", + "image_path": "4472d173a4ea941042d640635544b0b670d3ce8f63732b585fc254e685f001f1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 180, + 549, + 190 + ], + "lines": [ + { + "bbox": [ + 440, + 180, + 549, + 190 + ], + "spans": [ + { + "bbox": [ + 440, + 180, + 549, + 190 + ], + "type": "text", + "content": "(b) IoU with Normal Background" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 285, + 552, + 350 + ], + "blocks": [ + { + "bbox": [ + 313, + 242, + 554, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 242, + 554, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 554, + 275 + ], + "type": "text", + "content": "Table 5. Comparison of training with different magnitude reverse adversarial samples (UIAT) and our debiased training DHAT using WRN28-10 on the CIFAR-10." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 285, + 552, + 350 + ], + "lines": [ + { + "bbox": [ + 317, + 285, + 552, + 350 + ], + "spans": [ + { + "bbox": [ + 317, + 285, + 552, + 350 + ], + "type": "table", + "html": "
Methodε of zTraining SetTest SetRobust Gap↓
PGD-10↑AA↑PGD-10↑AA↑
UIAT-490.5361.4582.9452.079.38
UIAT-892.5762.8782.2651.3311.54
UIAT-1294.9065.2881.7250.8514.43
DHAT (ours)-488.7356.6183.9553.103.51
", + "image_path": "675e96b1d8c4350db1f0881f643ee88f03fead073c27da82ba69ac2992dec787.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 370, + 554, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 554, + 441 + ], + "type": "text", + "content": "superior feature attention. Table 5 shows stronger inverse attacks in UIAT worsen overfitting, increasing the robust gap from 9.38 to 14.43. In contrast, DHAT achieves better test robustness (83.95% PGD-10, 53.10% AA) with a significantly smaller robust gap (3.51), effectively mitigating overfitting while maintaining strong performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 454, + 388, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 454, + 388, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 454, + 388, + 466 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": "In this work, we identify a critical issue in adversarial training: AT model exhibits biased feature activation under inverse adversarial attacks. Specifically, training with inverse adversarial examples causes the model attention to shift from the foreground to background features, resulting in spurious correlation bias. To tackle this challenge, we propose a novel method called Debiased High-Confidence Adversarial Training (DHAT). DHAT incorporates two pivotal techniques: Debiased High-Confidence Logit Regularization (DHLR) and Foreground Logit Orthogonal Enhancement (FLOE). DHLR quantifies the bias towards background feature activation and recalibrates biased high-confidence logits to align the logits of adversarial examples with these debiased high-confidence logits. FLOE further restores the model's attention to its normal state by reducing the correlation between high-confidence logits and background features in affine space. This work highlights DHAT's effectiveness in enhancing model resilience and presents an adversarial training framework that integrates seamlessly with advanced training strategies." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2790" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China under Grant No. 62276221, 62376232; the Fundamental Research Funds for the Central Universities under Grant No. 21625344; Key Laboratory of Equipment Data Security and Guarantee Technology, Ministry of Education under Grant No. 2024020200." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 114, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 114, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 114, + 186 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 194, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 194, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 295, + 227 + ], + "type": "text", + "content": "[1] Ola Ahmad, Nicolas Béreux, Loïc Baret, Vahid Hashemi, and Freddy Lecue. Causal analysis for robust interpretability of neural networks. In WAVC, pages 4685-4694, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 271 + ], + "type": "text", + "content": "[2] Maksym Andriushchenko, Francesco Croce, Nicolas Flammarion, and Matthias Hein. Square attack: a query-efficient black-box adversarial attack via random search. In ECCV, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "type": "text", + "content": "[3] Saeid Asgari, Aliasghar Khani, Fereshte Khani, Ali Gholami, Linh Tran, Ali Mahdavi Amiri, and Ghassan Hamarneh. Masktune: Mitigating spurious correlations by forcing to explore. In NeurIPS, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 350 + ], + "type": "text", + "content": "[4] Nicholas Carlini and David Wagner. Towards evaluating the robustness of neural networks. IEEE Symp. Security Privacy, pages 39-57, 2017. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 396 + ], + "type": "text", + "content": "[5] Hyuna Cho, Yubin Han, and Won Hwa Kim. Antiadversarial consistency regularization for data augmentation: Applications to robust medical image segmentation. In MI-CAI, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 397, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 294, + 430 + ], + "type": "text", + "content": "[6] Jonathan Crabbé and Mihaela van der Schaar. Evaluating the robustness of interpretability methods through explanation invariance and equivariance. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 432, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 294, + 463 + ], + "type": "text", + "content": "[7] Francesco Croce and Matthias Hein. Reliable evaluation of adversarial robustness with an ensemble of diverse parameter-free attacks. In ICML, 2020. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 465, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 294, + 497 + ], + "type": "text", + "content": "[8] Francesco Croce and Matthias Hein. Minimally distorted adversarial examples with a fast adaptive boundary attack. In ICML, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 499, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 294, + 543 + ], + "type": "text", + "content": "[9] Junhao Dong, Seyed-Mohsen Moosavi-Dezfooli, Jianhuang Lai, and Xiaohua Xie. The enemy of my enemy is my friend: Exploring inverse adversaries for improving aversarial training. In CVPR, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 294, + 588 + ], + "type": "text", + "content": "[10] Xiaoyi Dong, Jiangfan Han, Dongdong Chen, Jiayang Liu, Huanyu Bian, Zehua Ma, Hongsheng Li, Xiaogang Wang, Weiming Zhang, and Nenghai Yu. Robust superpixel-guided attentional adversarial attack. In CVPR, 2020. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 611 + ], + "type": "text", + "content": "[11] Alhussein Fawzi, Hamza Fawzi, and Omar Fawzi. Adversarial vulnerability for any classifier. In NeurIPS, 2018. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 613, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 645 + ], + "type": "text", + "content": "[12] Liam Fowl, Micah Goldblum, Ping-yeh Chiang, Jonas Geiping, Wojciech Czaja, and Tom Goldstein. Adversarial examples make strong poisons. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "text", + "content": "[13] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. Avsegformer: Audio-visual segmentation with transformer. In AAAI, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[14] Ian J Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[15] Xinzhe Han, Shuhui Wang, Chi Su, Qingming Huang, and Qi Tian. General greedy de-bias learning. IEEE TPAMI, 45 (8):9789-9805, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 141, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 172 + ], + "type": "text", + "content": "[17] FastAI Jeremy Howard. Imagenette: A smaller subset of imagenet. In *GibHub*. https://github.com/fastai/imagenette, 2019.2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "text", + "content": "[18] Zhichao Huang, Yanbo Fan, Chen Liu, Weizhong Zhang, Yong Zhang, Mathieu Salzmann, Sabine Susstrunk, and Jue Wang. Fast adversarial training with adaptive step size. IEEE TIP, 32:6102-6114, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 251 + ], + "type": "text", + "content": "[19] Pavel Izmailov, Polina Kirichenko, Nate Gruver, and Andrew G Wilson. On feature learning in the presence of spurious correlations. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 285 + ], + "type": "text", + "content": "[20] Woo Jae Kim, Yoonki Cho, Junsik Jung, and Sung-Eui Yoon. Feature separation and recalibration for adversarial robustness. In CVPR, 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 288, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 330 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 332, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 363 + ], + "type": "text", + "content": "[22] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. In Toronto, ON, Canada, 2009. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 366, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 388 + ], + "type": "text", + "content": "[23] Boqi Li and Weiwei Liu. Wat: improve the worst-class robustness in adversarial training. In AAAI, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 389, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 410 + ], + "type": "text", + "content": "[24] Bo Li, Haoke Xiao, and Lv Tang. Asam: Boosting segment anything model with adversarial tuning. In CVPR, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 412, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 412, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 412, + 553, + 454 + ], + "type": "text", + "content": "[25] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Jianwei Yang, Chunyuan Li, et al. Visual in-context prompting. In CVPR, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 499 + ], + "type": "text", + "content": "[26] Zhuorong Li, Daiwei Yu, Lina Wei, Canghong Jin, Yun Zhang, and Sixian Chan. Soften to defend: Towards adversarial robustness via self-guided label refinement. In CVPR, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 502, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 553, + 533 + ], + "type": "text", + "content": "[27] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. 2, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 536, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 567 + ], + "type": "text", + "content": "[28] Yifei Ming, Hang Yin, and Yixuan Li. On the impact of spurious correlation for out-of-distribution detection. In AAAI, 2022. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 569, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 622 + ], + "type": "text", + "content": "[29] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "text", + "content": "[30] Hadi Salman, Andrew Ilyas, Logan Engstrom, Sai Vemprala, Aleksander Madry, and Ashish Kapoor. Unadversarial examples: Designing objects for robust vision. In NeurIPS, 2021. 1, 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[31] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 4" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2791" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[32] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In CVPR, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[33] Seonguk Seo, Joon-Young Lee, and Bohyung Han. Information-theoretic bias reduction via causal view of spurious correlation. In AAAI, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[34] Ali Shafahi, Mahyar Najibi, Mohammad Amin Ghiasi, Zheng Xu, John Dickerson, Christoph Studer, Larry S Davis, Gavin Taylor, and Tom Goldstein. Adversarial training for free! In NeurIPS, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 239 + ], + "type": "text", + "content": "[35] Kele Shao, Keda Tao, Can Qin, Haoxuan You, Yang Sui, and Huan Wang. Holitom: Holistic token merging for fast video large language models. arXiv preprint arXiv:2505.21334, 2025. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "text", + "content": "[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 318 + ], + "type": "text", + "content": "[37] Naman Deep Singh, Francesco Croce, and Matthias Hein. Revisiting adversarial training for imagenet: Architectures, training and generalization across threat models. In NeurIPS, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 319, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 352 + ], + "type": "text", + "content": "[38] Ajay Subramanian, Elena Sizikova, Najib Majaj, and Denis Pelli. Spatial-frequency channels, shape bias, and adversarial robustness. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "text", + "content": "[39] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2014. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[40] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. In CVPR, 2015. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 433, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 475 + ], + "type": "text", + "content": "[41] Yisen Wang, Difan Zou, Jinfeng Yi, James Bailey, Xingjun Ma, and Quanquan Gu. Improving adversarial robustness requires revisiting misclassified examples. In ICLR, 2020. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 477, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 294, + 510 + ], + "type": "text", + "content": "[42] Zeming Wei, Yifei Wang, Yiwen Guo, and Yisen Wang. Cfa: Class-wise calibrated fair adversarial training. In CVPR, 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "type": "text", + "content": "[43] Jing Wen, Siu-Ming Yiu, and Lucas CK Hui. Defending against model inversion attack by adversarial examples. In IEEE CSR, pages 551-556, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 545, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 577 + ], + "type": "text", + "content": "[44] Dongxian Wu, Shu-Tao Xia, and Yisen Wang. Adversarial weight perturbation helps robust generalization. In NeurIPS, 2020. 3, 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 578, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 612 + ], + "type": "text", + "content": "[45] Cihang Xie, Mingxing Tan, Boqing Gong, Jiang Wang, Alan L Yuille, and Quoc V Le. Adversarial examples improve image recognition. In CVPR, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 613, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 646 + ], + "type": "text", + "content": "[46] Shenglin Yin, Kelu Yao, Sheng Shi, Yangzhou Du, and Zhen Xiao. Again: Adversarial training with attribution span enlargement and hybrid feature fusion. In CVPR, 2023. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 647, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 669 + ], + "type": "text", + "content": "[47] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[48] Bohan Zeng, Shanglin Li, Xuhui Liu, Sicheng Gao, Xiaolong Jiang, Xu Tang, Yao Hu, Jianzhuang Liu, and Baochang Zhang. Controllable mind visual diffusion model. In AAAI, 2024. 1" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 161 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "text", + "content": "[49] Hongyang Zhang, Yaodong Yu, Jiantao Jiao, Eric Xing, Laurent El Ghaoui, and Michael Jordan. Theoretically principled trade-off between robustness and accuracy. In ICML, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[50] Tianren Zhang, Chujie Zhao, Guanyu Chen, Yizhou Jiang, and Feng Chen. Feature contamination: Neural networks learn uncorrelated features and fail to generalize. In ICML, 2024. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2792" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/c6c6dbbf-396b-41fc-b50d-460da67eb4ab_content_list.json b/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/c6c6dbbf-396b-41fc-b50d-460da67eb4ab_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e89531ebfeefe0fcdf3c3d5c1169233e3b71f815 --- /dev/null +++ b/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/c6c6dbbf-396b-41fc-b50d-460da67eb4ab_content_list.json @@ -0,0 +1,1574 @@ +[ + { + "type": "text", + "text": "Towards Annotation-Free Evaluation: KPAScore for Human Keypoint Detection", + "text_level": 1, + "bbox": [ + 91, + 130, + 903, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoxiao Wang $^{1*}$ Chunxiao Li $^{2*}$ Peng Sun $^{3}$ Boming Miao $^{2}$ Yunjian Zhang $^{1\\dagger}$ Yao Zhu $^{4\\dagger}$ $^{1}$ University of Chinese Academy of Sciences, Beijing, China $^{2}$ Beijing Normal University, Beijing, China $^{3}$ Central University of Finance and Economics, Beijing, China $^{4}$ Zhejiang University, Hangzhou, China", + "bbox": [ + 99, + 179, + 944, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wangxiaoxiao@mails.ucas.ac.cn, ee-zhuy@zju.edu.cn", + "bbox": [ + 303, + 236, + 736, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human keypoint detection is fundamental in computer vision, with applications in pose estimation and action recognition. However, existing evaluation metrics (e.g., OKS, PCP, PDJ) rely on human-annotated ground truth, a labor-intensive process that increases costs, limits scalability. To address this, we propose KPAScore (KeyPoint Answering Score), an annotation-free metric independent of ground truth. It evaluates keypoint detection using a two-stage VLM-based question-answering process: first, the VLM identifies the presence of keypoints within the image, and second, visual prompts are introduced to query the likelihood of each keypoint being accurately localized within a predefined boundary. To validate the rationale behind KPAScore, we propose KPUBench (KeyPoint Understanding Benchmark), which comprehensively evaluates the VLM's ability to determine keypoint presence and localization. Extensive experiments demonstrate KPAScore's effectiveness from three perspectives: consistency to keypoint variation, correlation with traditional metrics, alignment with human perception. We hope KPAScore will reduce reliance on manual annotations, facilitating broader adoption of keypoint detection in real-world applications.", + "bbox": [ + 89, + 319, + 485, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 684, + 220, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The primary evaluation metrics for human keypoint detection include Object Keypoint Similarity (OKS) [17], Percentage of Detected Joints (PDJ) [28], and Percentage of Correct Keypoints (PCK) [1], all of which require ground truth for computation. Typically, ground truth is obtained in two ways. The first approach uses optical motion capture during dataset collection to generate precise ground truth, as seen in datasets like Human3.6M [11]; however, the limitations of motion capture equipment often result in datasets with limited scene diversity. The second approach relies on", + "bbox": [ + 89, + 710, + 483, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "manual annotation, which is highly time-consuming, labor-intensive and prone to inaccuracies. For example, MPII [1] and COCO [17] require over 400k and 5.1 million keypoint annotations. As shown in Fig. 1, missing ground truth annotations can lead to the exclusion of some person keypoints from evaluation. Thus, developing a new metric that is annotation-free, automated, and more objective in reflecting detection performance is essential.", + "bbox": [ + 511, + 287, + 906, + 409 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the continuous advancement of vision-language models, VLMs [2, 16, 19] have developed robust semantic understanding and strong generalization capabilities. These VLMs can tackle a variety of computer vision tasks, such as image comprehension [10] and localization [5]. Leveraging vision-language models, Lin et al. [18] propose the VQA score, which evaluates the alignment between image-text prompts using the responses generated by VLMs. This approach outperforms the widely used CLIPScore [9, 21], particularly when handling complex prompts involving combinations of objects, attributes, and relationships. Inspired by VQA score [18], we propose the Keypoint Answering Score (KPAScore), a novel keypoint detection metric based on a two-stage question-answering process and visual prompts. In the first stage, the Visual Language Model (VLM) is asked whether a specific keypoint exists in the image, and based on the response probability of presence, different weights are assigned to each keypoint. In the second stage, we introduce a visual prompt to guide the VLM's focus towards keypoint regions. Specifically, a visual prompt in the form of a blue dot is generated on the estimated keypoint, and the VLM is queried whether the keypoint accurately corresponds to this prompt. The product of the probability of answering \"yes\" in the second stage and the weight determines the KPAscore for the evaluated detection algorithms.", + "bbox": [ + 511, + 412, + 908, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We conducted extensive experiments to validate the rationale, consistency, correlation with ground-truth-based metrics, and alignment with human perception of the proposed KPAScore, thereby demonstrating its effectiveness. Regarding its rationale, we constructed the KeyPoint Understanding Benchmark (KPubBench) for visual language models (VLMs), which includes over 1.8 million questions.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8441", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/252797c1828c2b1d5fa2386e9660d1092a10db5653f07f4dcecb833f12731ad1.jpg", + "image_caption": [ + "Figure 1. Based on the visual understanding capabilities of VLM, the KPAScore is proposed as an annotation-free evaluation metric for human keypoint detection, avoiding the issues of omissions and errors commonly associated with ground truth annotations. This makes it suitable for extensive evaluation in wild settings." + ], + "image_footnote": [], + "bbox": [ + 116, + 95, + 877, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This benchmark evaluates the performance of 18 existing VLM models across multiple dimensions, including keypoint presence detection and precise localization. The evaluation results indicate that existing VLM models perform well in understanding the existence of keypoints but exhibit some deficiencies in the accuracy of keypoint localization. To optimize the performance of KPAScore in downstream tasks, we fine-tuned the model using 10,000 keypoint localization data points. The fine-tuned VLM demonstrated its capability to understand human keypoints, thus validating the rationale of using the VLM model to construct keypoint detection evaluation metrics.", + "bbox": [ + 88, + 349, + 480, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding consistency, we evaluated the overall consistency between KPAScore and OKS across keypoint detection algorithms. Additionally, we assessed KPAScore's responsiveness to keypoint displacement, finding that it exhibits high variation consistency in response to changes in predicted keypoint positions. In terms of correlation with ground-truth-based metrics, we examined the relationship between KPAScore and existing evaluation metrics (OKS) on the COCO [17]. The results show that KPAScore, independent of annotations, exhibits a strong positive correlation with existing evaluation metrics. Then, we explore the broader implications of KPAScore in guiding keypoint detection and integration strategies. Finally, we validated the alignment of KPAScore with human perception on complex human images in real-world scenarios, where ground-truth annotations were unavailable, and found that KPAScore aligns well with human perception.", + "bbox": [ + 89, + 532, + 482, + 790 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 89, + 792, + 415, + 806 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce KPAScore, an annotation-free, groundtruth-independent metric for keypoint detection evaluation, showing a high positive correlation with conventional groundtruth-based metrics.", + "- We propose a two-stage question-answering method and visual prompting approach, which effectively enhance the" + ], + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VLM's keypoint understanding capability, thereby improving the reliability of the KPAScore.", + "bbox": [ + 526, + 349, + 903, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We establish KPU-Bench, using over 1.8 million questions to evaluate 18 VLMs' understanding capability of human keypoints. And through fine-tuning, the performance of VLM on KPAScore is further improved.", + "bbox": [ + 511, + 380, + 906, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 454, + 653, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Keypoint Detection Metrics: In human keypoint detection tasks, commonly used evaluation metrics include PCK [1], PDJ [28], OKS [17], and OKS-based metrics like mAP and mAR [17]. OKS-based metrics are more widely adopted, with the OKS calculation formula as follows:", + "bbox": [ + 511, + 479, + 905, + 555 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {O K S} = \\frac {\\sum_ {i} \\exp \\left(- \\frac {d _ {i} ^ {2}}{2 s ^ {2} \\kappa_ {i} ^ {2}}\\right) \\delta \\left(v _ {i} > 0\\right)}{\\sum_ {i} \\delta \\left(v _ {i} > 0\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 565, + 903, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $d_{i}$ represents the Euclidean distance between the predicted and ground-truth keypoints, $s$ is the object scale, $\\kappa_{i}$ is a constant representing keypoint-specific falloff, and $v_{i}$ is the visibility flag of the ground-truth keypoint $i$ . The use of the negative exponential function $\\exp(-x)$ makes OKS less sensitive to the detection results. Even if there are certain errors in detecting certain key points, OKS still tends to have a higher score as long as the error is not too large.", + "bbox": [ + 511, + 613, + 905, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on different OKS thresholds, predicted keypoints are matched with ground-truth keypoints to calculate the precision and recall for each detection result, forming a Precision-Recall curve. AP represents the area under this curve, while mAP is the average AP across a set of thresholds, providing an overall measure of detection accuracy. mAR, in turn, reflects the model's recall capability for ground-truth keypoints across multiple OKS thresholds.", + "bbox": [ + 511, + 734, + 905, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human keypoint detection algorithm: Human keypoint detection algorithm has evolved rapidly. Early approaches like CPM [26] focused on sequentially refining keypoint", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8442", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8f3613895074f8931179e4717475c5208031619f5c0530415d5877463ff63889.jpg", + "image_caption": [ + "Segmentation from detection method" + ], + "image_footnote": [], + "bbox": [ + 112, + 122, + 370, + 279 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cd38116ec0a17dfeed3b691083a9ac2cb946845ebe0b07c3310b64b6205ef5f5.jpg", + "image_caption": [ + "Phase1: Presence Detection", + "Does the image include the location of the person's {left shoulder}?" + ], + "image_footnote": [], + "bbox": [ + 424, + 135, + 496, + 234 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/739e8fce916d4383c6a2ea6f02c2bc5bdc3d3eac95311fe46491bd1fac81df48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 172, + 573, + 195 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/32fc663f756aec074d72c0a8bdb2d6005bbf138cf159b89aaade7f663a6c1980.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 148, + 684, + 219 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} = \\operatorname {P r o b} \\left(\" Y s e \"\\right) = 0. 6 5 4 3\n$$\n", + "text_format": "latex", + "bbox": [ + 705, + 159, + 864, + 170 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f12db0ef0faa0c78c4788d3ee3625e39dabe9449c78091cdd7f47a93ad2f472b.jpg", + "image_caption": [ + "Finetuning for $VLM_{2}$" + ], + "image_footnote": [], + "bbox": [ + 138, + 323, + 349, + 381 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/018fc78cb72d4dab55ee27a58ef1083eea4295f06f7a7188f2a2abf8db523822.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 328, + 410, + 357 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/142d52ee48fb2d2a33400a033a6da20bb319d3e3278c75c478aff4a4bb6600f3.jpg", + "image_caption": [ + "Phase2: Precision Localization", + "Does the blue dot correspond to the location of the {left shoulder}?", + "Figure 2. We introduce the two-stage KPAScore computation for multiple occluded persons. First, bounding boxes obtained from the detection method are used to segment individuals in the image. In Stage 1, $VLM_{1}$ is queried to determine the presence of keypoints, with the probability of a \"Yes\" response serving as the visibility score. To enhance accuracy, $VLM_{2}$ is fine-tuned using 10k training samples before being deployed in the second stage. In Stage 2, the predicted keypoint coordinates are visualized as prompts, and $VLM_{2}$ is queried again to assess accuracy. The probability of a \"Yes\" response serves as the accuracy score. Finally, KPAScore integrates both visibility and accuracy scores to provide a comprehensive evaluation of keypoint detection performance." + ], + "image_footnote": [], + "bbox": [ + 428, + 280, + 501, + 397 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/39520483cf3d47cb190d1507db6c3e6cd94078f2400a81b314ff04aa8f102ea4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 309, + 573, + 333 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a259cc492e8b656d79ae9f88898232413072224e38f9ca7b1945a1078a5b2730.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 311, + 676, + 334 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {k} _ {i} = \\operatorname {P r o b} \\left(\" Y s e \"\\right) = 0. 8 9 5 5\n$$\n", + "text_format": "latex", + "bbox": [ + 694, + 311, + 852, + 323 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nK P A S c o r e _ {p} = \\frac {\\sum_ {i = 1} ^ {n} \\sigma \\left(v _ {i}\\right) \\sqrt {k _ {i} v _ {i}}}{\\sum_ {i = 1} ^ {n} \\sigma \\left(v _ {i}\\right)} = 0. 9 0 1 2\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 359, + 799, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "predictions through multi-stage processing. HRNet [22] proposed a high-resolution network architecture that maintains spatial detail across the network. LiteHRNet [30] later advanced keypoint detection by maintaining high-resolution features while optimizing for computational efficiency. RSN [3] introduced a method for aggregating intra-level features, preserving rich spatial information. More recent methods, such as RTMPose [12] and AlphaPose [6], adopt a top-down framework, utilizing object detectors to first locate each individual before estimating poses. RTMPose [12] excels in complex environments by isolating instances, while AlphaPose [6] handles occlusions and challenging poses efficiently, making it ideal for real-time scenarios. ViTPose [27], applies vision transformers as non-hierarchical backbones to capture global features. VLM has shown significant progress in keypoint detection, LocLLM [24] achieved comparable performance to ViT-Pose [27] on COCO dataset, highlighting its capability for keypoint understanding, though it has not yet been applied in evaluation tasks. Appendix. A give a supplement to related work.", + "bbox": [ + 91, + 516, + 483, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 829, + 181, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the KPAScore, to the best of our knowledge, the first model-based evaluation metric for keypoint detection. This metric leverages the visual under", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "standing capabilities of large vision-language models, eliminating the reliance on human annotations inherent in traditional groundtruth-based metrics. An overview of the KPA-Score calculation process is illustrated in Fig. 2 and the KPAScore can be formulated as:", + "bbox": [ + 511, + 516, + 906, + 590 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {K P A S c o r e} _ {p} = \\frac {\\sum_ {i = 1} ^ {n} \\sigma \\left(v _ {i}\\right) \\sqrt {k _ {i} \\cdot v _ {i}}}{\\sum_ {i = 1} ^ {n} \\sigma \\left(v _ {i}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 602, + 903, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the indicator function $\\sigma(v_{i})$ is defined as:", + "bbox": [ + 513, + 643, + 831, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma \\left(v _ {i}\\right) = \\left\\{ \\begin{array}{l l} 1, & v _ {i} > 0. 5 \\\\ 0, & v _ {i} \\leq 0. 5 \\end{array} . \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 669, + 903, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $v_{i}$ represents the keypoint visibility score, obtained from the first-stage presence detection. If $v_{i} < 0.5$ , the keypoint is considered not existed. For values between 0.5 and 1, $v_{i}$ reflects the confidence of the VLM in the keypoint's presence: lower values indicate that the keypoint is likely occluded, while higher values suggest it is unobstructed. Similarly, $k_{i}$ represents the keypoint accuracy score, derived from the second-stage precision localization. The denominator ensures the normalization of contributions from all visible keypoints and the square operation avoids that the product of two numbers less than 1 is too small, maintaining a balanced evaluation.", + "bbox": [ + 511, + 719, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8443", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. First-Stage Presence Detection", + "text_level": 1, + "bbox": [ + 89, + 90, + 361, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the first stage, we leverage vision-language models (VLMs) to determine whether specific keypoints are present in the image. Image $I$ is the segmented images of individual persons which obtained from object detection. Given the detection task's keypoint list with $N$ keypoints $P = [p_{1}, p_{2}, \\ldots, p_{N}]$ , for instance:", + "bbox": [ + 89, + 112, + 483, + 204 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP = \\left[ ^ {\\prime \\prime} \\text {N o s e} ^ {\\prime \\prime}, ^ {\\prime \\prime} \\text {L e f t E y} ^ {\\prime \\prime}, \\dots , ^ {\\prime \\prime} \\text {R i g h t W i r s t} ^ {\\prime \\prime} \\right], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 210, + 482, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we query the VLM for each keypoint $p_i$ in the detection task's keypoint list with the question $q(p_i)$ as:", + "bbox": [ + 89, + 233, + 483, + 263 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} = \\text {\" n o s e \"}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 271, + 217, + 287 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {1} \\left(p _ {i}\\right) = \\text {D o e s t h e i m a g e i n c l u d e t h e l o c a t i o n o f t h e} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 289, + 482, + 306 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {p e r s o n ' s \" n o s e \"} \\text {? P l a c e a n s w e r y e s o r n o .}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 309, + 439, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The VLM returns a probability score for answering \"Yes,\" which is taken as the visibility score $v_{i}$ :", + "bbox": [ + 89, + 330, + 483, + 361 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {i} = \\operatorname {P r o b} \\left(\" \\text {Y e s}\" \\mid I, q _ {1} \\left(k _ {i}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 368, + 482, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If $v_{i} \\leq 0.5$ , the keypoint is considered not existential and is excluded from subsequent precision evaluation.", + "bbox": [ + 89, + 391, + 483, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Second-Stage Precision Localization", + "text_level": 1, + "bbox": [ + 89, + 429, + 405, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the second stage, we perform a more refined evaluation of the detection method by assessing the accuracy of the estimated keypoint coordinates.", + "bbox": [ + 89, + 450, + 483, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VLMs are generally not well-suited for directly evaluating keypoint coordinate accuracy. This limitation likely arises from the inherent difficulty VLMs face in establishing a precise mapping between textual descriptions and specific spatial locations within an image. To overcome this challenge, we incorporate visual prompts [13, 15, 33] to enhance the VLMs' ability to perceive and interpret keypoint locations. Specifically, we overlay a blue dot $d_{i}$ at the keypoint coordinates predicted by the detection algorithm, using it as a visual cue to facilitate keypoint accuracy evaluation.", + "bbox": [ + 89, + 496, + 483, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an image embedded with a visual prompt, denoted as $I_{d_i}$ , we query the VLM with the following question:", + "bbox": [ + 89, + 662, + 483, + 693 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} = \\text {\" n o s e \"}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 700, + 210, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {2} \\left(p _ {i}\\right) = \\text {D o e s t h e b l u e d o t c o r r e s p o n d t o t h e l o c a t i o n (7)}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 719, + 482, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\text {o f t h e \" n o s e\" ? P l a c e a n s w e r Y e s o r N o .} \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 738, + 424, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The generative likelihood of the VLM responding \"Yes\" is then computed as:", + "bbox": [ + 89, + 758, + 482, + 790 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nk _ {i} = \\operatorname {P r o b} \\left(\" \\text {Y e s}\" \\mid I _ {d _ {i}}, q _ {2} \\left(p _ {i}\\right)\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 805, + 482, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k_{i}$ serves as the accuracy score for the given keypoint in this second stage. Finally, the model's overall KPAScore is obtained by computing a weighted average of all keypoint scores, integrating the evaluation results from both stages as shown in Eq. (2).", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section conducts extensive experiments to answer the following questions:", + "bbox": [ + 511, + 114, + 903, + 143 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Can VLMs comprehend human keypoints? Additionally, how can we further enhance VLMs' capability to understand keypoints? See Sec. 4.1.", + "- Can KPAScore consistently reflect changes in predicted keypoints relative to the ground truth? See Sec. 4.2.", + "- Is there a correlation between KPAScore and existing annotation-dependent metrics? See Sec. 4.3.", + "- Can KPAScore be considered as a guide for keypoint detection ensemble methods? See Sec. 4.4", + "- Does the KPAScore align with human perception in complex scenarios without ground truth? See Sec. 4.5." + ], + "bbox": [ + 513, + 145, + 903, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Keypoint Understanding Benchmark", + "text_level": 1, + "bbox": [ + 513, + 319, + 833, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the widespread application of VLMs in computer vision tasks, their ability to accurately understand human keypoints remains an open question. We evaluate the keypoint understanding capability of VLMs from two perspectives: presence detection and precision localization. To facilitate this evaluation, we introduce the Keypoint Understanding Benchmark (KPUBench), a large-scale dataset comprising 1.8 million structured questions derived from COCO [17], MPII [1], and Human3.6M [11]. The dataset is systematically split into training and test sets with an approximate ratio of 9.5:0.5, and all evaluations are conducted exclusively on the test set.", + "bbox": [ + 511, + 340, + 906, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The relationship between KPUBench and KPAScore can be summarized as follows: KPUBench serves as a benchmark for assessing the ability of existing VLMs to comprehend human keypoints in an unsupervised or zero-shot setting, measuring their fine-grained semantic and visual perception capabilities. The best-performing VLM from this evaluation is then employed in downstream tasks, including the construction of KPAScore. The architecture of KPUBench is illustrated in Figure 3 and the evaluation results of 18 visual-language models is presented in Tab. 1.", + "bbox": [ + 511, + 523, + 905, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(a) Presence Detection assesses the VLM's ability to identify whether a specific keypoint is present in an image. We query the VLM: \"Does the image include the location of the person's {keypoint name}?\". Ideally, for keypoints annotated as present (COCO visibility scores of 1 or 2), the VLM should return \"Yes\" with a probability above 0.5, whereas for absent keypoints (visibility score of 0), this probability should be below 0.5. The positive-to-negative sample ratio is approximately 5.31:4.69. To evaluate VLM reliability and confidence in keypoint presence detection, we compute accuracy $(ACC)$ and probabilistic expectation $(E(Prob))$ .", + "bbox": [ + 511, + 674, + 906, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in the third and fourth columns of Tab. 1, most VLMs demonstrate a fundamental ability to recognize keypoints, though model performance does not clearly correlate", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8444", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b7208c297a8c898e031abad7c49fbada1727edfcc150df2975c42c5e58acd06f.jpg", + "image_caption": [ + "(c) Precision Localization Detection—Visual Prompt Understanding Capability", + "Figure 3. Overview of the KPUBench evaluation framework. (a): Evaluates whether a VLM can determine the existence of a keypoint in an image. (b): Assesses the VLM's ability to recognize keypoints based on explicit numerical coordinates. (c): Measures whether the VLM can accurately associate a visual cue (a blue dot) with the corresponding keypoint. The expected answer is \"Yes\" for correct localizations and \"No\" for incorrect ones." + ], + "image_footnote": [], + "bbox": [ + 148, + 89, + 852, + 352 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2c3e909f772d67968d3b745f341a77e48610512f0b4d0f87432e81c6bf4c5f64.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSizePresence DetectionPrecision Localization Detection
ACC(%)E(Prob)CoordinateVisual Prompt
ACC(%)E(Prob)ACC(%)E(Prob)
LLaVA-OneVersion [14]0.5B55.5849.7150.0050.2452.3250.30
DeepSeek-Coder-Base [8]1.3B62.0352.7749.9149.9655.5750.86
InternVL2-2B [4]2B56.3047.3350.0950.0655.2551.75
Qwen2-VL-2B [25]2B73.6758.7050.0050.0456.5451.21
MiniCPM-V 2.0 [29]2.8B58.4344.4249.9449.9454.8652.25
LLaVA-V1.5-7B [19]7B61.0057.1849.8649.7053.3250.72
LLaVA-V1.6-7B [19]7B75.1651.2149.7349.9254.8451.21
Qwen2-VL-7B [25]7B66.9857.6150.2250.1160.5552.71
InternVL2-8B [4]8B56.3047.3350.9550.0655.2551.76
MiniCPM-V-2.6 [29]8B54.8451.2150.8550.1153.0450.97
GLM-4V-9B [7]9B76.9949.8749.9849.7662.5758.17
LLama3.2-11B-coco [23]11B61.2055.0848.7549.0156.0254.53
LLaVA-V1.5-13B [19]13B69.8758.8250.0049.9153.5050.61
LLaVA-V1.6-13B [19]13B71.5255.9349.9949.0458.9651.32
Cogglm2-LLama3-chat-19B [10]19B70.3354.0649.9849.9854.7550.86
InternVL2-26B [4]26B61.8553.1750.0350.1957.2353.89
LLaVA-V1.6-34B [19]34B84.6765.1249.9549.9050.0050.35
Qwen2-VL-72B [25]72B79.5463.6050.8450.1556.0752.96
GLM-4V-9B-finetuig9B----80.1574.32
LLaVA + GLM-finetuig43B84.6765.12--84.5378.10
", + "bbox": [ + 114, + 464, + 888, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Performance of 18 VLM models on KPUBench. KPUBench evaluates VLMs' understanding of keypoint detection from two perspectives: presence detection and precision localization detection, providing a comprehensive measurement using accuracy and probabilistic expectation.", + "bbox": [ + 89, + 830, + 906, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8445", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with model size. Both the LLaVA-V1.6 and Qwen families perform well, exceeding $66.98\\%$ accuracy, likely due to task-related training data or effective architectures. Among them, LLaVA-V1.6-34B [19] achieves the highest accuracy $(84.67\\%)$ and probabilistic expectation $(65.12\\%)$ , indicating strong keypoint perception. In contrast, GLM-4V [7] attains $76.99\\%$ accuracy but a lower probabilistic expectation $(49.87\\%)$ , suggesting lower confidence despite correct predictions, which may affect reliability in practical use.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Precision Localization Detection consists of two components: coordinate understanding and visual prompt understanding. We use human-annotated keypoint coordinates as ground truth, treating them as positive samples. Negative samples are generated by shifting each keypoint randomly by $10\\%$ of the bounding box size, maintaining a 1:1 positive-to-negative ratio.", + "bbox": [ + 89, + 227, + 480, + 333 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Coordinate Understanding We ask the VLM: \"Does the coordinate $(x, y)$ correspond to the location of the {keypoint name}?\" Accuracy and probabilistic expectation are computed, where higher accuracy indicates better alignment with human judgment, and higher probabilistic expectation reflects the model's confidence.", + "- Visual Prompt Understanding Using the same positive and negative samples, we mark a blue dot at the keypoint location and ask: \"Does the blue dot correspond to the location of the {keypoint name}?\" Accuracy and probabilistic expectation are measured similarly. The effect of the color of the visual prompt is illustrated in Appendix. B." + ], + "bbox": [ + 89, + 337, + 483, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on the fifth and sixth columns of Tab. 1, VLMs show limited ability in interpreting numerical keypoint coordinates, with accuracy around $50\\%$ , near random guessing. This suggests difficulty in mapping numerical coordinates to spatial locations in an image. Introducing visual prompts significantly improves keypoint comprehension, increasing accuracy by an average of $5.42\\%$ . Notably, GLM-4V's accuracy rose from $49.98\\%$ to $62.57\\%$ , with its probabilistic expectation improving from $49.76\\%$ to $58.17\\%$ , confirming that visual prompts help direct model attention to spatial locations. However, despite these gains, performance remains insufficient for practical applications.", + "bbox": [ + 89, + 537, + 482, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further enhance the model's precision, we fine-tuned GLM-4V-9B using the LoRA method with 10k Precision Localization Detection training samples containing visual prompts (detailed hyperparameter settings are provided in Appendix. C). The fine-tuned model, GLM-4V-9B-finetuned, achieved a remarkable accuracy of $80.15\\%$ and a probabilistic expectation of $78.10\\%$ , demonstrating that even a simple fine-tuning process can significantly enhance VLM performance on specialized tasks.", + "bbox": [ + 89, + 718, + 482, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, we implemented a two-stage approach by employing LLaVA-V1.6-34B as the Presence Detection module in the first stage. We retained only the keypoints identi", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fied as present by LLaVA-V1.6-34B and fed them into the Precision Localization Detection stage. This LLaVA-V1.6-34B + GLM-4V-9B-finetuned two-stage framework further enhanced the model's keypoint localization understanding, achieving an accuracy of $84.53\\%$ and a probabilistic expectation of $78.10\\%$ . Therefore, we propose a two-stage evaluation framework that integrates presence detection and precision localization to systematically assess keypoint detection models using VLMs.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a2d8729ede95ec94215b447e6f726fafc9685393bd43082665c9325c66d8223d.jpg", + "table_caption": [ + "4.2. The Consistency of KPAscore" + ], + "table_footnote": [ + "Table 2. The overall scores of seven keypoint detection algorithms on the COCO dataset." + ], + "table_body": "
ModelKPAScoreOKSmAP
ViTPose-H* [27]60.9872.9579.20
RTMPose-1 [12]59.0572.4777.30
RSN [3]59.0472.1075.00
HRNet [22]59.2772.1474.91
Alphapose [6]58.1371.4670.61
CPM [26]57.2470.1465.23
LiteHRNet [30]57.0870.3064.24
", + "bbox": [ + 542, + 273, + 877, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After establishing the LLaVA-V1.6-34B + GLM-4V-9B-finetuned two-stage KPAScore framework, we conducted a comprehensive evaluation of seven mainstream keypoint detection methods on the COCO dataset. The selected models include ViTPose [27], RTMPose [12], Alphapose [6], HRNet [22], RSN [3], LiteHRNet [30], and CPM [26].", + "bbox": [ + 511, + 445, + 903, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consistency in response to changes is a critical factor in evaluating the usability of a metric. Here, we demonstrate that KPAScore possesses both overall consistency and change consistency from two perspectives.", + "bbox": [ + 511, + 536, + 903, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we present the average KPAScore, OKS, and mAP of the seven methods on COCO (see Table 2). As shown, KPAScore exhibits strong alignment with traditional keypoint metrics (OKS and $mAP$ ). Models with higher OKS and mAP scores, such as ViTPose-H and RTMPose-1, also achieve the highest KPAScore, while lower-performing models, like CPM and LiteHRNet, similarly rank lower under KPAScore. This consistency validates KPAScore as an effective metric that aligns with established evaluation criteria while leveraging VLM-based assessment to provide a more holistic keypoint detection evaluation.", + "bbox": [ + 511, + 597, + 905, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, we apply varying intensities of offset to the keypoint detection results for each detection algorithm, then recalculated the KPAScore. We analyzed the KPAScore across different offset intensities, where the offset is defined as the square root of the bounding box area (a standard approach on the COCO dataset) multiplied by a factor ranging from $0\\%$ to $50\\%$ , with random offset directions which shown in Fig. 4a", + "bbox": [ + 511, + 763, + 903, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 4b, the KPAScore for all 7 keypoint", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8446", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f4818e4fa8cdc639b531a6d7a7b2d8e706fe4e1b5fc9dc6b0f7c93bda8878e6.jpg", + "image_caption": [ + "(a) Offset Example." + ], + "image_footnote": [ + "- 3% offset\n- 5% offset\n- 10% offset\n- 15% offset\n- 20% offset\n- 25% offset\n- 30% offset\n- 50% offset" + ], + "bbox": [ + 254, + 95, + 472, + 250 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a39f267bc48dc7382fd9edc78c097dd580ff37cb70032b308f0a25639a92e17.jpg", + "image_caption": [ + "(b) Variation of KPAScore and OKS.", + "Figure 4. (a): Circles of different colors represent the corresponding offset keypoint positions at various offset intensities. (b): A line chart illustrating the variation in values of KPAScore (red line) and OKS (blue line) at different offset intensities." + ], + "image_footnote": [], + "bbox": [ + 478, + 90, + 821, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bcfbedcc9a88cb04043fd9b6ad43947fcc05b398b3067545339a1acbdf37cd35.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodViTPose-H* [27]RSN [3]HRNet [22]RTMPose-1 [12]Alphapose [6]CPM [26]LiteHRNet [30]Average
Corr0.720.710.720.710.700.690.730.71
", + "bbox": [ + 93, + 321, + 906, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. The Pearson correlation coefficients between the KPAScore and OKS.", + "bbox": [ + 263, + 378, + 732, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "detection methods steadily decreases with increasing offset intensity, indicating that the KPAScore effectively responds to variations in keypoint positions and can serve as a reliable new evaluation metric. Specifically, offset intensities of $3\\%$ and $5\\%$ result in a decrease in the KPAScore, but the decline is not significant. As illustrated in Fig. 4a, offsets within $5\\%$ are not easily perceptible to the human eye, and therefore the minor drop in score is consistent with human perception. When the offset intensity ranges from $10\\%$ to $30\\%$ , the KPAScore decreases more markedly, and at a $50\\%$ offset intensity, the scores for all algorithms converge, KPAScore is almost 0. OKS also decreases steadily with the offset, but when the offset strength is $50\\%$ , OKS is still above 50, which shows that OKS is not very sensitive to the feedback of the offset.", + "bbox": [ + 88, + 406, + 485, + 632 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. The Correlation Between KPAScore and OKS", + "text_level": 1, + "bbox": [ + 89, + 648, + 480, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Beyond evaluating the consistency of KPAScore, we examine its relationship with OKS by computing their Pearson correlation across various keypoint detection methods, as shown in Tab. 3. The results show a strong positive correlation, ranging from 0.69 to 0.73 (average 0.71), indicating that KPAScore aligns well with OKS in assessing keypoint detection quality, further validating its reliability as an evaluation metric.", + "bbox": [ + 89, + 672, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "However, it is worth noting that the correlation is not perfect $(\\mathrm{Corr} < 1.0)$ , indicating that KPAScore captures aspects of keypoint detection performance that may not be fully reflected by OKS. Unlike OKS, which primarily relies on Euclidean distance-based similarity, KPAScore incorporates a human-centric evaluation approach using visual-language models (VLMs). This difference allows KPAS-", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "core to assess keypoint detection from a perceptual perspective, potentially capturing nuanced failure cases that traditional distance-based metrics overlook.", + "bbox": [ + 511, + 407, + 906, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further investigate these differences, we analyze cases where OKS and KPAScore exhibit discrepancies (see Appendix. D). We find that OKS tends to penalize predictions that deviate slightly from ground truth, even if they are perceptually acceptable, while KPAScore—by leveraging VLM-based reasoning—can tolerate minor deviations if the predicted keypoints remain semantically correct. Conversely, KPAScore may assign lower scores to cases where OKS is relatively high, particularly when the predicted keypoint location appears anatomically implausible despite being close to the ground truth. These findings highlight KPAScore's potential to complement OKS by incorporating a perceptual validation mechanism, making it a promising metric for real-world pose estimation tasks where human perception plays a crucial role.", + "bbox": [ + 511, + 453, + 908, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. KPAScore as a Guide for Keypoint Detection Ensembles", + "text_level": 1, + "bbox": [ + 511, + 694, + 906, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we explore whether KPAScore can serve as a guiding metric for keypoint detection ensemble methods. Since KPAScore is annotation-free and evaluates keypoints in a perceptually meaningful manner, we propose an ensemble strategy that selects keypoints with the highest KPAScore from each algorithm, forming a new method we refer to as KPAPose.", + "bbox": [ + 511, + 733, + 906, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further assess the effectiveness of KPAScore in guiding keypoint detection improvements, we incorporate it as an additional supervisory signal for ViTPose- $\\mathbf{H}^*$ and RTMPose-1 on the COCO dataset, selecting keypoints that", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8447", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieve the highest estimated KPAScore across different locations.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 4, KPAPose achieves the highest scores in both KPAScore (62.65) and OKS (73.63), surpassing ViTPose-H* [27] and RTMPose-1 [12]. This demonstrates that KPAScore can serve as an effective fusion strategy for integrating multiple detection methods, producing results that not only align closely with traditional evaluation metrics (e.g., OKS) but also better capture human perceptual judgments—an advantage that annotation-dependent metrics fail to achieve.", + "bbox": [ + 89, + 123, + 483, + 257 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e1006a1ab6127dc01fdde5cb18d34aa16ba8901d1212e484c3fc40549c10341b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelKPAScoreOKS
ViTPose-H* [27]60.9872.95
RTMPose-1 [12]59.0572.47
KPAPose62.6573.63
", + "bbox": [ + 161, + 272, + 411, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. KPAPose performance that integrates both methods", + "bbox": [ + 106, + 366, + 465, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Alignment with Human Perception", + "text_level": 1, + "bbox": [ + 89, + 405, + 395, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further verify the alignment between KPAScore's perception verification mechanism and human judgment, we constructed a test set featuring complex scenes and challenging actions in both indoor and outdoor environments. These include activities such as yoga, boxing, modern dance, and gymnastics, with participants wearing diverse clothing, making manual annotation of ground truth extremely difficult.", + "bbox": [ + 89, + 429, + 483, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the performance of various keypoint detection algorithms in these challenging settings, we employed KPAScore alongside human ratings. A group of 20 researchers independently rated the inference results from different algorithms on a scale of 0 to 10, averaging the scores for each algorithm per image. The details of the human rating strategy are provided in Appendix E.", + "bbox": [ + 89, + 551, + 483, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5 presents the evaluation results of three algorithms on three actions, along with KPAPose, an ensemble of these models, arranged in decreasing difficulty. The results show that in the absence of ground-truth annotations, KPAScore closely aligns with human evaluations. RTM [12] performed the worst in both KPAScore and human ratings, while KPAPose consistently achieved the highest scores, demonstrating KPAScore's robustness in challenging scenarios. For harder-to-estimate actions, KPAPose significantly improved KPAScore from 63.09 to 89.03 and human ratings from 6.55 to 7.65. In contrast, simpler actions saw marginal improvements, with KPAScore increasing by only 0.36, making the difference barely noticeable to humans. Moving forward, we aim to further explore KPAScore's role in evaluating complex scenarios and its potential to enhance keypoint detection algorithms.", + "bbox": [ + 89, + 659, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/212337c1d4e4b5a9c9ec8d189741cdfb66d8a8487fb1922095e3653efa3d5bce.jpg", + "image_caption": [ + "Figure 5. Alignment of KPA-Score and Human Rating in Complex Scenes with Complex Actions. Here, we select ViTPose, CPM, RTM, and the KPA-Pose generated based on these three algorithms, with additional visualizations provided in Appendix. F" + ], + "image_footnote": [], + "bbox": [ + 522, + 90, + 895, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 539, + 633, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce the KPAScore, a method that explores the keypoint understanding capabilities of vision-language models (VLMs) and proposes, to the best of our knowledge, the first human keypoint evaluation metric that eliminates the reliance on ground truth through a two-stage visual prompt question-answering approach. To validate the rationale behind this method, we introduce the Keypoint Understanding Benchmark (KPubBench), which contains over 1.8 million questions. The results demonstrate that our two-stage questioning and visual prompt approach significantly enhance the keypoint comprehension capabilities of VLMs. Furthermore, our analysis demonstrates that KPAScore exhibits variation consistency, maintains a strong correlation with traditional evaluation metrics such as OKS, and aligns closely with human perception—all while being entirely annotation-free. By leveraging VLMs for keypoint assessment, KPAScore provides a scalable, automated, and human-aligned alternative to existing evaluation methods. We hope this work inspires further research on utilizing VLMs for structured vision tasks and fosters the development of more robust, perception-driven keypoint evaluation metrics in computer vision applications.", + "bbox": [ + 511, + 580, + 906, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8448", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 1, 2, 4", + "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 1", + "[3] Yuanhao Cai, Zhicheng Wang, Zhengxiong Luo, Binyi Yin, Angang Du, Haoqian Wang, Xiangyu Zhang, Xinyu Zhou, Erjin Zhou, and Jian Sun. Learning delicate local representations for multi-person pose estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 455-472. Springer, 2020. 3, 6, 7", + "[4] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023.5", + "[5] Michael Dorkenwald, Nimrod Barazani, Cees GM Snoek, and Yuki M Asano. Pin: Positional insert unlocks object localisation abilities in vlms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13548-13558, 2024. 1, 11", + "[6] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3, 6, 7", + "[7] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. Chatglm: A family of large language models from glm-130b to glm-4 all tools, 2024. 5, 6, 11", + "[8] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y. Wu, Y.K. Li, Fuli Luo, Yingfei Xiong, and Wenfeng Liang. Deepseek-coder: When the large language model meets programming – the rise of code intelligence, 2024. 5", + "[9] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 1" + ], + "bbox": [ + 99, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 1, 5, 11", + "[11] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 1, 4", + "[12] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Real-time multi-person pose estimation based on mmpose, 2023. 3, 6, 7, 8", + "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4", + "[14] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 5", + "[15] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 4", + "[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1", + "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1, 2, 4", + "[18] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2025. 1, 11", + "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 5, 6, 11", + "[20] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 11", + "[21] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1", + "[22] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE conference on computer" + ], + "bbox": [ + 516, + 92, + 905, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8449", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vision and pattern recognition, pages 5693-5703, 2019. 3, 6, 7", + "[23] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 11", + "[24] Dongkai Wang, Shiyu Xuan, and Shiliang Zhang. Locllm: Exploiting generalizable human keypoint localization via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[25] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 11", + "[26] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7", + "[27] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In Advances in Neural Information Processing Systems, 2022. 3, 6, 7, 8", + "[28] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. 1, 2", + "[29] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 5", + "[30] Changqian Yu, Bin Xiao, Changxin Gao, Lu Yuan, Lei Zhang, Nong Sang, and Jingdong Wang. Lite-hrnet: A lightweight high-resolution network. In CVPR, 2021. 3, 6, 7", + "[31] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 11", + "[32] Hongyi Zhu, Jia-Hong Huang, Stevan Rudinac, and Evangelos Kanoulas. Enhancing interactive image retrieval with query rewriting using large language models and vision language models. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 978-987, 2024. 11", + "[33] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 4" + ], + "bbox": [ + 91, + 90, + 482, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8450", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/c6c6dbbf-396b-41fc-b50d-460da67eb4ab_model.json b/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/c6c6dbbf-396b-41fc-b50d-460da67eb4ab_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c27c5cecfacfa65edbcc0fcbe479a967b5685227 --- /dev/null +++ b/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/c6c6dbbf-396b-41fc-b50d-460da67eb4ab_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.905, + 0.153 + ], + "angle": 0, + "content": "Towards Annotation-Free Evaluation: KPAScore for Human Keypoint Detection" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.18, + 0.946, + 0.235 + ], + "angle": 0, + "content": "Xiaoxiao Wang\\(^{1*}\\) Chunxiao Li\\(^{2*}\\) Peng Sun\\(^{3}\\) Boming Miao\\(^{2}\\) Yunjian Zhang\\(^{1\\dagger}\\) Yao Zhu\\(^{4\\dagger}\\) \n\\(^{1}\\)University of Chinese Academy of Sciences, Beijing, China \\(^{2}\\)Beijing Normal University, Beijing, China \\(^{3}\\)Central University of Finance and Economics, Beijing, China \\(^{4}\\)Zhejiang University, Hangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.237, + 0.737, + 0.251 + ], + "angle": 0, + "content": "wangxiaoxiao@mails.ucas.ac.cn, ee-zhuy@zju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.486, + 0.653 + ], + "angle": 0, + "content": "Human keypoint detection is fundamental in computer vision, with applications in pose estimation and action recognition. However, existing evaluation metrics (e.g., OKS, PCP, PDJ) rely on human-annotated ground truth, a labor-intensive process that increases costs, limits scalability. To address this, we propose KPAScore (KeyPoint Answering Score), an annotation-free metric independent of ground truth. It evaluates keypoint detection using a two-stage VLM-based question-answering process: first, the VLM identifies the presence of keypoints within the image, and second, visual prompts are introduced to query the likelihood of each keypoint being accurately localized within a predefined boundary. To validate the rationale behind KPAScore, we propose KPUBench (KeyPoint Understanding Benchmark), which comprehensively evaluates the VLM's ability to determine keypoint presence and localization. Extensive experiments demonstrate KPAScore's effectiveness from three perspectives: consistency to keypoint variation, correlation with traditional metrics, alignment with human perception. We hope KPAScore will reduce reliance on manual annotations, facilitating broader adoption of keypoint detection in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.685, + 0.222, + 0.7 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.484, + 0.862 + ], + "angle": 0, + "content": "The primary evaluation metrics for human keypoint detection include Object Keypoint Similarity (OKS) [17], Percentage of Detected Joints (PDJ) [28], and Percentage of Correct Keypoints (PCK) [1], all of which require ground truth for computation. Typically, ground truth is obtained in two ways. The first approach uses optical motion capture during dataset collection to generate precise ground truth, as seen in datasets like Human3.6M [11]; however, the limitations of motion capture equipment often result in datasets with limited scene diversity. The second approach relies on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.907, + 0.41 + ], + "angle": 0, + "content": "manual annotation, which is highly time-consuming, labor-intensive and prone to inaccuracies. For example, MPII [1] and COCO [17] require over 400k and 5.1 million keypoint annotations. As shown in Fig. 1, missing ground truth annotations can lead to the exclusion of some person keypoints from evaluation. Thus, developing a new metric that is annotation-free, automated, and more objective in reflecting detection performance is essential." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.909, + 0.791 + ], + "angle": 0, + "content": "With the continuous advancement of vision-language models, VLMs [2, 16, 19] have developed robust semantic understanding and strong generalization capabilities. These VLMs can tackle a variety of computer vision tasks, such as image comprehension [10] and localization [5]. Leveraging vision-language models, Lin et al. [18] propose the VQA score, which evaluates the alignment between image-text prompts using the responses generated by VLMs. This approach outperforms the widely used CLIPScore [9, 21], particularly when handling complex prompts involving combinations of objects, attributes, and relationships. Inspired by VQA score [18], we propose the Keypoint Answering Score (KPAScore), a novel keypoint detection metric based on a two-stage question-answering process and visual prompts. In the first stage, the Visual Language Model (VLM) is asked whether a specific keypoint exists in the image, and based on the response probability of presence, different weights are assigned to each keypoint. In the second stage, we introduce a visual prompt to guide the VLM's focus towards keypoint regions. Specifically, a visual prompt in the form of a blue dot is generated on the estimated keypoint, and the VLM is queried whether the keypoint accurately corresponds to this prompt. The product of the probability of answering \"yes\" in the second stage and the weight determines the KPAscore for the evaluated detection algorithms." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We conducted extensive experiments to validate the rationale, consistency, correlation with ground-truth-based metrics, and alignment with human perception of the proposed KPAScore, thereby demonstrating its effectiveness. Regarding its rationale, we constructed the KeyPoint Understanding Benchmark (KPubBench) for visual language models (VLMs), which includes over 1.8 million questions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.236, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8441" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.117, + 0.096, + 0.878, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.294, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 1. Based on the visual understanding capabilities of VLM, the KPAScore is proposed as an annotation-free evaluation metric for human keypoint detection, avoiding the issues of omissions and errors commonly associated with ground truth annotations. This makes it suitable for extensive evaluation in wild settings." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.351, + 0.482, + 0.531 + ], + "angle": 0, + "content": "This benchmark evaluates the performance of 18 existing VLM models across multiple dimensions, including keypoint presence detection and precise localization. The evaluation results indicate that existing VLM models perform well in understanding the existence of keypoints but exhibit some deficiencies in the accuracy of keypoint localization. To optimize the performance of KPAScore in downstream tasks, we fine-tuned the model using 10,000 keypoint localization data points. The fine-tuned VLM demonstrated its capability to understand human keypoints, thus validating the rationale of using the VLM model to construct keypoint detection evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.483, + 0.791 + ], + "angle": 0, + "content": "Regarding consistency, we evaluated the overall consistency between KPAScore and OKS across keypoint detection algorithms. Additionally, we assessed KPAScore's responsiveness to keypoint displacement, finding that it exhibits high variation consistency in response to changes in predicted keypoint positions. In terms of correlation with ground-truth-based metrics, we examined the relationship between KPAScore and existing evaluation metrics (OKS) on the COCO [17]. The results show that KPAScore, independent of annotations, exhibits a strong positive correlation with existing evaluation metrics. Then, we explore the broader implications of KPAScore in guiding keypoint detection and integration strategies. Finally, we validated the alignment of KPAScore with human perception on complex human images in real-world scenarios, where ground-truth annotations were unavailable, and found that KPAScore aligns well with human perception." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.793, + 0.416, + 0.807 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.483, + 0.87 + ], + "angle": 0, + "content": "- We introduce KPAScore, an annotation-free, groundtruth-independent metric for keypoint detection evaluation, showing a high positive correlation with conventional groundtruth-based metrics." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- We propose a two-stage question-answering method and visual prompting approach, which effectively enhance the" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.351, + 0.905, + 0.38 + ], + "angle": 0, + "content": "VLM's keypoint understanding capability, thereby improving the reliability of the KPAScore." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.381, + 0.907, + 0.442 + ], + "angle": 0, + "content": "- We establish KPU-Bench, using over 1.8 million questions to evaluate 18 VLMs' understanding capability of human keypoints. And through fine-tuning, the performance of VLM on KPAScore is further improved." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.455, + 0.655, + 0.47 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.906, + 0.556 + ], + "angle": 0, + "content": "Human Keypoint Detection Metrics: In human keypoint detection tasks, commonly used evaluation metrics include PCK [1], PDJ [28], OKS [17], and OKS-based metrics like mAP and mAR [17]. OKS-based metrics are more widely adopted, with the OKS calculation formula as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.566, + 0.905, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\mathrm {O K S} = \\frac {\\sum_ {i} \\exp \\left(- \\frac {d _ {i} ^ {2}}{2 s ^ {2} \\kappa_ {i} ^ {2}}\\right) \\delta \\left(v _ {i} > 0\\right)}{\\sum_ {i} \\delta \\left(v _ {i} > 0\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.734 + ], + "angle": 0, + "content": "where \\( d_{i} \\) represents the Euclidean distance between the predicted and ground-truth keypoints, \\( s \\) is the object scale, \\( \\kappa_{i} \\) is a constant representing keypoint-specific falloff, and \\( v_{i} \\) is the visibility flag of the ground-truth keypoint \\( i \\). The use of the negative exponential function \\( \\exp(-x) \\) makes OKS less sensitive to the detection results. Even if there are certain errors in detecting certain key points, OKS still tends to have a higher score as long as the error is not too large." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Based on different OKS thresholds, predicted keypoints are matched with ground-truth keypoints to calculate the precision and recall for each detection result, forming a Precision-Recall curve. AP represents the area under this curve, while mAP is the average AP across a set of thresholds, providing an overall measure of detection accuracy. mAR, in turn, reflects the model's recall capability for ground-truth keypoints across multiple OKS thresholds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Human keypoint detection algorithm: Human keypoint detection algorithm has evolved rapidly. Early approaches like CPM [26] focused on sequentially refining keypoint" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8442" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.104, + 0.367, + 0.117 + ], + "angle": 0, + "content": "Segmentation from detection method" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.123, + 0.371, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.103, + 0.614, + 0.116 + ], + "angle": 0, + "content": "Phase1: Presence Detection" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.122, + 0.879, + 0.135 + ], + "angle": 0, + "content": "Does the image include the location of the person's {left shoulder}?" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.136, + 0.498, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.173, + 0.574, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.15, + 0.686, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.706, + 0.16, + 0.865, + 0.171 + ], + "angle": 0, + "content": "\\[\nv _ {i} = \\operatorname {P r o b} \\left(\" Y s e \"\\right) = 0. 6 5 4 3\n\\]" + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.257, + 0.637, + 0.27 + ], + "angle": 0, + "content": "Phase2: Precision Localization" + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.275, + 0.882, + 0.288 + ], + "angle": 0, + "content": "Does the blue dot correspond to the location of the {left shoulder}?" + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.303, + 0.321, + 0.318 + ], + "angle": 0, + "content": "Finetuning for \\(VLM_{2}\\)" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.324, + 0.35, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.329, + 0.411, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.281, + 0.502, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.31, + 0.574, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.312, + 0.678, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.695, + 0.312, + 0.854, + 0.324 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {k} _ {i} = \\operatorname {P r o b} \\left(\" Y s e \"\\right) = 0. 8 9 5 5\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.361, + 0.8, + 0.389 + ], + "angle": 0, + "content": "\\[\nK P A S c o r e _ {p} = \\frac {\\sum_ {i = 1} ^ {n} \\sigma \\left(v _ {i}\\right) \\sqrt {k _ {i} v _ {i}}}{\\sum_ {i = 1} ^ {n} \\sigma \\left(v _ {i}\\right)} = 0. 9 0 1 2\n\\]" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.419, + 0.907, + 0.503 + ], + "angle": 0, + "content": "Figure 2. We introduce the two-stage KPAScore computation for multiple occluded persons. First, bounding boxes obtained from the detection method are used to segment individuals in the image. In Stage 1, \\( VLM_{1} \\) is queried to determine the presence of keypoints, with the probability of a \"Yes\" response serving as the visibility score. To enhance accuracy, \\( VLM_{2} \\) is fine-tuned using 10k training samples before being deployed in the second stage. In Stage 2, the predicted keypoint coordinates are visualized as prompts, and \\( VLM_{2} \\) is queried again to assess accuracy. The probability of a \"Yes\" response serves as the accuracy score. Finally, KPAScore integrates both visibility and accuracy scores to provide a comprehensive evaluation of keypoint detection performance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.517, + 0.485, + 0.82 + ], + "angle": 0, + "content": "predictions through multi-stage processing. HRNet [22] proposed a high-resolution network architecture that maintains spatial detail across the network. LiteHRNet [30] later advanced keypoint detection by maintaining high-resolution features while optimizing for computational efficiency. RSN [3] introduced a method for aggregating intra-level features, preserving rich spatial information. More recent methods, such as RTMPose [12] and AlphaPose [6], adopt a top-down framework, utilizing object detectors to first locate each individual before estimating poses. RTMPose [12] excels in complex environments by isolating instances, while AlphaPose [6] handles occlusions and challenging poses efficiently, making it ideal for real-time scenarios. ViTPose [27], applies vision transformers as non-hierarchical backbones to capture global features. VLM has shown significant progress in keypoint detection, LocLLM [24] achieved comparable performance to ViT-Pose [27] on COCO dataset, highlighting its capability for keypoint understanding, though it has not yet been applied in evaluation tasks. Appendix. A give a supplement to related work." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.182, + 0.845 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we introduce the KPAScore, to the best of our knowledge, the first model-based evaluation metric for keypoint detection. This metric leverages the visual under" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.907, + 0.592 + ], + "angle": 0, + "content": "standing capabilities of large vision-language models, eliminating the reliance on human annotations inherent in traditional groundtruth-based metrics. An overview of the KPA-Score calculation process is illustrated in Fig. 2 and the KPAScore can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.603, + 0.905, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathrm {K P A S c o r e} _ {p} = \\frac {\\sum_ {i = 1} ^ {n} \\sigma \\left(v _ {i}\\right) \\sqrt {k _ {i} \\cdot v _ {i}}}{\\sum_ {i = 1} ^ {n} \\sigma \\left(v _ {i}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.644, + 0.833, + 0.659 + ], + "angle": 0, + "content": "where the indicator function \\(\\sigma(v_{i})\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.67, + 0.905, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\sigma \\left(v _ {i}\\right) = \\left\\{ \\begin{array}{l l} 1, & v _ {i} > 0. 5 \\\\ 0, & v _ {i} \\leq 0. 5 \\end{array} . \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Here, \\( v_{i} \\) represents the keypoint visibility score, obtained from the first-stage presence detection. If \\( v_{i} < 0.5 \\), the keypoint is considered not existed. For values between 0.5 and 1, \\( v_{i} \\) reflects the confidence of the VLM in the keypoint's presence: lower values indicate that the keypoint is likely occluded, while higher values suggest it is unobstructed. Similarly, \\( k_{i} \\) represents the keypoint accuracy score, derived from the second-stage precision localization. The denominator ensures the normalization of contributions from all visible keypoints and the square operation avoids that the product of two numbers less than 1 is too small, maintaining a balanced evaluation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8443" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.362, + 0.108 + ], + "angle": 0, + "content": "3.1. First-Stage Presence Detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.205 + ], + "angle": 0, + "content": "In the first stage, we leverage vision-language models (VLMs) to determine whether specific keypoints are present in the image. Image \\(I\\) is the segmented images of individual persons which obtained from object detection. Given the detection task's keypoint list with \\(N\\) keypoints \\(P = [p_{1}, p_{2}, \\ldots, p_{N}]\\), for instance:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.211, + 0.483, + 0.228 + ], + "angle": 0, + "content": "\\[\nP = \\left[ ^ {\\prime \\prime} \\text {N o s e} ^ {\\prime \\prime}, ^ {\\prime \\prime} \\text {L e f t E y} ^ {\\prime \\prime}, \\dots , ^ {\\prime \\prime} \\text {R i g h t W i r s t} ^ {\\prime \\prime} \\right], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.234, + 0.484, + 0.265 + ], + "angle": 0, + "content": "we query the VLM for each keypoint \\( p_i \\) in the detection task's keypoint list with the question \\( q(p_i) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.272, + 0.218, + 0.288 + ], + "angle": 0, + "content": "\\[\np _ {i} = \\text {\" n o s e \"}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.29, + 0.483, + 0.307 + ], + "angle": 0, + "content": "\\[\nq _ {1} \\left(p _ {i}\\right) = \\text {D o e s t h e i m a g e i n c l u d e t h e l o c a t i o n o f t h e} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.31, + 0.441, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\text {p e r s o n ' s \" n o s e \"} \\text {? P l a c e a n s w e r y e s o r n o .}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.331, + 0.484, + 0.362 + ], + "angle": 0, + "content": "The VLM returns a probability score for answering \"Yes,\" which is taken as the visibility score \\( v_{i} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.369, + 0.483, + 0.386 + ], + "angle": 0, + "content": "\\[\nv _ {i} = \\operatorname {P r o b} \\left(\" \\text {Y e s}\" \\mid I, q _ {1} \\left(k _ {i}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.423 + ], + "angle": 0, + "content": "If \\( v_{i} \\leq 0.5 \\), the keypoint is considered not existential and is excluded from subsequent precision evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.43, + 0.406, + 0.446 + ], + "angle": 0, + "content": "3.2. Second-Stage Precision Localization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.484, + 0.496 + ], + "angle": 0, + "content": "In the second stage, we perform a more refined evaluation of the detection method by assessing the accuracy of the estimated keypoint coordinates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.484, + 0.661 + ], + "angle": 0, + "content": "VLMs are generally not well-suited for directly evaluating keypoint coordinate accuracy. This limitation likely arises from the inherent difficulty VLMs face in establishing a precise mapping between textual descriptions and specific spatial locations within an image. To overcome this challenge, we incorporate visual prompts [13, 15, 33] to enhance the VLMs' ability to perceive and interpret keypoint locations. Specifically, we overlay a blue dot \\( d_{i} \\) at the keypoint coordinates predicted by the detection algorithm, using it as a visual cue to facilitate keypoint accuracy evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.484, + 0.694 + ], + "angle": 0, + "content": "Given an image embedded with a visual prompt, denoted as \\( I_{d_i} \\), we query the VLM with the following question:" + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.701, + 0.212, + 0.716 + ], + "angle": 0, + "content": "\\[\np _ {i} = \\text {\" n o s e \"}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.72, + 0.483, + 0.735 + ], + "angle": 0, + "content": "\\[\nq _ {2} \\left(p _ {i}\\right) = \\text {D o e s t h e b l u e d o t c o r r e s p o n d t o t h e l o c a t i o n (7)}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.739, + 0.425, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\text {o f t h e \" n o s e\" ? P l a c e a n s w e r Y e s o r N o .} \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.483, + 0.791 + ], + "angle": 0, + "content": "The generative likelihood of the VLM responding \"Yes\" is then computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.806, + 0.483, + 0.822 + ], + "angle": 0, + "content": "\\[\nk _ {i} = \\operatorname {P r o b} \\left(\" \\text {Y e s}\" \\mid I _ {d _ {i}}, q _ {2} \\left(p _ {i}\\right)\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( k_{i} \\) serves as the accuracy score for the given keypoint in this second stage. Finally, the model's overall KPAScore is obtained by computing a weighted average of all keypoint scores, integrating the evaluation results from both stages as shown in Eq. (2)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.145 + ], + "angle": 0, + "content": "This section conducts extensive experiments to answer the following questions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.146, + 0.905, + 0.19 + ], + "angle": 0, + "content": "- Can VLMs comprehend human keypoints? Additionally, how can we further enhance VLMs' capability to understand keypoints? See Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.191, + 0.905, + 0.221 + ], + "angle": 0, + "content": "- Can KPAScore consistently reflect changes in predicted keypoints relative to the ground truth? See Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.222, + 0.905, + 0.25 + ], + "angle": 0, + "content": "- Is there a correlation between KPAScore and existing annotation-dependent metrics? See Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.252, + 0.904, + 0.28 + ], + "angle": 0, + "content": "- Can KPAScore be considered as a guide for keypoint detection ensemble methods? See Sec. 4.4" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.282, + 0.905, + 0.312 + ], + "angle": 0, + "content": "- Does the KPAScore align with human perception in complex scenarios without ground truth? See Sec. 4.5." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.146, + 0.905, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.32, + 0.834, + 0.337 + ], + "angle": 0, + "content": "4.1. Keypoint Understanding Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.522 + ], + "angle": 0, + "content": "With the widespread application of VLMs in computer vision tasks, their ability to accurately understand human keypoints remains an open question. We evaluate the keypoint understanding capability of VLMs from two perspectives: presence detection and precision localization. To facilitate this evaluation, we introduce the Keypoint Understanding Benchmark (KPUBench), a large-scale dataset comprising 1.8 million structured questions derived from COCO [17], MPII [1], and Human3.6M [11]. The dataset is systematically split into training and test sets with an approximate ratio of 9.5:0.5, and all evaluations are conducted exclusively on the test set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.674 + ], + "angle": 0, + "content": "The relationship between KPUBench and KPAScore can be summarized as follows: KPUBench serves as a benchmark for assessing the ability of existing VLMs to comprehend human keypoints in an unsupervised or zero-shot setting, measuring their fine-grained semantic and visual perception capabilities. The best-performing VLM from this evaluation is then employed in downstream tasks, including the construction of KPAScore. The architecture of KPUBench is illustrated in Figure 3 and the evaluation results of 18 visual-language models is presented in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.907, + 0.855 + ], + "angle": 0, + "content": "(a) Presence Detection assesses the VLM's ability to identify whether a specific keypoint is present in an image. We query the VLM: \"Does the image include the location of the person's {keypoint name}?\". Ideally, for keypoints annotated as present (COCO visibility scores of 1 or 2), the VLM should return \"Yes\" with a probability above 0.5, whereas for absent keypoints (visibility score of 0), this probability should be below 0.5. The positive-to-negative sample ratio is approximately 5.31:4.69. To evaluate VLM reliability and confidence in keypoint presence detection, we compute accuracy \\((ACC)\\) and probabilistic expectation \\((E(Prob))\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "As shown in the third and fourth columns of Tab. 1, most VLMs demonstrate a fundamental ability to recognize keypoints, though model performance does not clearly correlate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8444" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.09, + 0.854, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.362, + 0.766, + 0.377 + ], + "angle": 0, + "content": "(c) Precision Localization Detection—Visual Prompt Understanding Capability" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.396, + 0.908, + 0.455 + ], + "angle": 0, + "content": "Figure 3. Overview of the KPUBench evaluation framework. (a): Evaluates whether a VLM can determine the existence of a keypoint in an image. (b): Assesses the VLM's ability to recognize keypoints based on explicit numerical coordinates. (c): Measures whether the VLM can accurately associate a visual cue (a blue dot) with the corresponding keypoint. The expected answer is \"Yes\" for correct localizations and \"No\" for incorrect ones." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.465, + 0.889, + 0.822 + ], + "angle": 0, + "content": "
ModelSizePresence DetectionPrecision Localization Detection
ACC(%)E(Prob)CoordinateVisual Prompt
ACC(%)E(Prob)ACC(%)E(Prob)
LLaVA-OneVersion [14]0.5B55.5849.7150.0050.2452.3250.30
DeepSeek-Coder-Base [8]1.3B62.0352.7749.9149.9655.5750.86
InternVL2-2B [4]2B56.3047.3350.0950.0655.2551.75
Qwen2-VL-2B [25]2B73.6758.7050.0050.0456.5451.21
MiniCPM-V 2.0 [29]2.8B58.4344.4249.9449.9454.8652.25
LLaVA-V1.5-7B [19]7B61.0057.1849.8649.7053.3250.72
LLaVA-V1.6-7B [19]7B75.1651.2149.7349.9254.8451.21
Qwen2-VL-7B [25]7B66.9857.6150.2250.1160.5552.71
InternVL2-8B [4]8B56.3047.3350.9550.0655.2551.76
MiniCPM-V-2.6 [29]8B54.8451.2150.8550.1153.0450.97
GLM-4V-9B [7]9B76.9949.8749.9849.7662.5758.17
LLama3.2-11B-coco [23]11B61.2055.0848.7549.0156.0254.53
LLaVA-V1.5-13B [19]13B69.8758.8250.0049.9153.5050.61
LLaVA-V1.6-13B [19]13B71.5255.9349.9949.0458.9651.32
Cogglm2-LLama3-chat-19B [10]19B70.3354.0649.9849.9854.7550.86
InternVL2-26B [4]26B61.8553.1750.0350.1957.2353.89
LLaVA-V1.6-34B [19]34B84.6765.1249.9549.9050.0050.35
Qwen2-VL-72B [25]72B79.5463.6050.8450.1556.0752.96
GLM-4V-9B-finetuig9B----80.1574.32
LLaVA + GLM-finetuig43B84.6765.12--84.5378.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.832, + 0.908, + 0.875 + ], + "angle": 0, + "content": "Table 1. Performance of 18 VLM models on KPUBench. KPUBench evaluates VLMs' understanding of keypoint detection from two perspectives: presence detection and precision localization detection, providing a comprehensive measurement using accuracy and probabilistic expectation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8445" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "with model size. Both the LLaVA-V1.6 and Qwen families perform well, exceeding \\(66.98\\%\\) accuracy, likely due to task-related training data or effective architectures. Among them, LLaVA-V1.6-34B [19] achieves the highest accuracy \\((84.67\\%)\\) and probabilistic expectation \\((65.12\\%)\\), indicating strong keypoint perception. In contrast, GLM-4V [7] attains \\(76.99\\%\\) accuracy but a lower probabilistic expectation \\((49.87\\%)\\), suggesting lower confidence despite correct predictions, which may affect reliability in practical use." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.482, + 0.334 + ], + "angle": 0, + "content": "Precision Localization Detection consists of two components: coordinate understanding and visual prompt understanding. We use human-annotated keypoint coordinates as ground truth, treating them as positive samples. Negative samples are generated by shifting each keypoint randomly by \\(10\\%\\) of the bounding box size, maintaining a 1:1 positive-to-negative ratio." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.338, + 0.483, + 0.429 + ], + "angle": 0, + "content": "- Coordinate Understanding We ask the VLM: \"Does the coordinate \\((x, y)\\) correspond to the location of the {keypoint name}?\" Accuracy and probabilistic expectation are computed, where higher accuracy indicates better alignment with human judgment, and higher probabilistic expectation reflects the model's confidence." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.429, + 0.484, + 0.533 + ], + "angle": 0, + "content": "- Visual Prompt Understanding Using the same positive and negative samples, we mark a blue dot at the keypoint location and ask: \"Does the blue dot correspond to the location of the {keypoint name}?\" Accuracy and probabilistic expectation are measured similarly. The effect of the color of the visual prompt is illustrated in Appendix. B." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.338, + 0.484, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Based on the fifth and sixth columns of Tab. 1, VLMs show limited ability in interpreting numerical keypoint coordinates, with accuracy around \\(50\\%\\), near random guessing. This suggests difficulty in mapping numerical coordinates to spatial locations in an image. Introducing visual prompts significantly improves keypoint comprehension, increasing accuracy by an average of \\(5.42\\%\\). Notably, GLM-4V's accuracy rose from \\(49.98\\%\\) to \\(62.57\\%\\), with its probabilistic expectation improving from \\(49.76\\%\\) to \\(58.17\\%\\), confirming that visual prompts help direct model attention to spatial locations. However, despite these gains, performance remains insufficient for practical applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.856 + ], + "angle": 0, + "content": "To further enhance the model's precision, we fine-tuned GLM-4V-9B using the LoRA method with 10k Precision Localization Detection training samples containing visual prompts (detailed hyperparameter settings are provided in Appendix. C). The fine-tuned model, GLM-4V-9B-finetuned, achieved a remarkable accuracy of \\(80.15\\%\\) and a probabilistic expectation of \\(78.10\\%\\), demonstrating that even a simple fine-tuning process can significantly enhance VLM performance on specialized tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Finally, we implemented a two-stage approach by employing LLaVA-V1.6-34B as the Presence Detection module in the first stage. We retained only the keypoints identi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "fied as present by LLaVA-V1.6-34B and fed them into the Precision Localization Detection stage. This LLaVA-V1.6-34B + GLM-4V-9B-finetuned two-stage framework further enhanced the model's keypoint localization understanding, achieving an accuracy of \\(84.53\\%\\) and a probabilistic expectation of \\(78.10\\%\\). Therefore, we propose a two-stage evaluation framework that integrates presence detection and precision localization to systematically assess keypoint detection models using VLMs." + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.239, + 0.776, + 0.255 + ], + "angle": 0, + "content": "4.2. The Consistency of KPAscore" + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.274, + 0.879, + 0.401 + ], + "angle": 0, + "content": "
ModelKPAScoreOKSmAP
ViTPose-H* [27]60.9872.9579.20
RTMPose-1 [12]59.0572.4777.30
RSN [3]59.0472.1075.00
HRNet [22]59.2772.1474.91
Alphapose [6]58.1371.4670.61
CPM [26]57.2470.1465.23
LiteHRNet [30]57.0870.3064.24
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.41, + 0.905, + 0.438 + ], + "angle": 0, + "content": "Table 2. The overall scores of seven keypoint detection algorithms on the COCO dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.905, + 0.537 + ], + "angle": 0, + "content": "After establishing the LLaVA-V1.6-34B + GLM-4V-9B-finetuned two-stage KPAScore framework, we conducted a comprehensive evaluation of seven mainstream keypoint detection methods on the COCO dataset. The selected models include ViTPose [27], RTMPose [12], Alphapose [6], HRNet [22], RSN [3], LiteHRNet [30], and CPM [26]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.905, + 0.597 + ], + "angle": 0, + "content": "Consistency in response to changes is a critical factor in evaluating the usability of a metric. Here, we demonstrate that KPAScore possesses both overall consistency and change consistency from two perspectives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.764 + ], + "angle": 0, + "content": "First, we present the average KPAScore, OKS, and mAP of the seven methods on COCO (see Table 2). As shown, KPAScore exhibits strong alignment with traditional keypoint metrics (OKS and \\(mAP\\)). Models with higher OKS and mAP scores, such as ViTPose-H and RTMPose-1, also achieve the highest KPAScore, while lower-performing models, like CPM and LiteHRNet, similarly rank lower under KPAScore. This consistency validates KPAScore as an effective metric that aligns with established evaluation criteria while leveraging VLM-based assessment to provide a more holistic keypoint detection evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.884 + ], + "angle": 0, + "content": "Second, we apply varying intensities of offset to the keypoint detection results for each detection algorithm, then recalculated the KPAScore. We analyzed the KPAScore across different offset intensities, where the offset is defined as the square root of the bounding box area (a standard approach on the COCO dataset) multiplied by a factor ranging from \\(0\\%\\) to \\(50\\%\\), with random offset directions which shown in Fig. 4a" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 4b, the KPAScore for all 7 keypoint" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8446" + } + ], + [ + { + "type": "image_footnote", + "bbox": [ + 0.185, + 0.119, + 0.241, + 0.224 + ], + "angle": 0, + "content": "- 3% offset\n- 5% offset\n- 10% offset\n- 15% offset\n- 20% offset\n- 25% offset\n- 30% offset\n- 50% offset" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.096, + 0.473, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.255, + 0.378, + 0.267 + ], + "angle": 0, + "content": "(a) Offset Example." + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.091, + 0.822, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.255, + 0.747, + 0.267 + ], + "angle": 0, + "content": "(b) Variation of KPAScore and OKS." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.279, + 0.908, + 0.308 + ], + "angle": 0, + "content": "Figure 4. (a): Circles of different colors represent the corresponding offset keypoint positions at various offset intensities. (b): A line chart illustrating the variation in values of KPAScore (red line) and OKS (blue line) at different offset intensities." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.322, + 0.907, + 0.369 + ], + "angle": 0, + "content": "
MethodViTPose-H* [27]RSN [3]HRNet [22]RTMPose-1 [12]Alphapose [6]CPM [26]LiteHRNet [30]Average
Corr0.720.710.720.710.700.690.730.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.264, + 0.379, + 0.733, + 0.392 + ], + "angle": 0, + "content": "Table 3. The Pearson correlation coefficients between the KPAScore and OKS." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.407, + 0.486, + 0.633 + ], + "angle": 0, + "content": "detection methods steadily decreases with increasing offset intensity, indicating that the KPAScore effectively responds to variations in keypoint positions and can serve as a reliable new evaluation metric. Specifically, offset intensities of \\(3\\%\\) and \\(5\\%\\) result in a decrease in the KPAScore, but the decline is not significant. As illustrated in Fig. 4a, offsets within \\(5\\%\\) are not easily perceptible to the human eye, and therefore the minor drop in score is consistent with human perception. When the offset intensity ranges from \\(10\\%\\) to \\(30\\%\\), the KPAScore decreases more markedly, and at a \\(50\\%\\) offset intensity, the scores for all algorithms converge, KPAScore is almost 0. OKS also decreases steadily with the offset, but when the offset strength is \\(50\\%\\), OKS is still above 50, which shows that OKS is not very sensitive to the feedback of the offset." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.649, + 0.481, + 0.665 + ], + "angle": 0, + "content": "4.3. The Correlation Between KPAScore and OKS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Beyond evaluating the consistency of KPAScore, we examine its relationship with OKS by computing their Pearson correlation across various keypoint detection methods, as shown in Tab. 3. The results show a strong positive correlation, ranging from 0.69 to 0.73 (average 0.71), indicating that KPAScore aligns well with OKS in assessing keypoint detection quality, further validating its reliability as an evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "However, it is worth noting that the correlation is not perfect \\((\\mathrm{Corr} < 1.0)\\), indicating that KPAScore captures aspects of keypoint detection performance that may not be fully reflected by OKS. Unlike OKS, which primarily relies on Euclidean distance-based similarity, KPAScore incorporates a human-centric evaluation approach using visual-language models (VLMs). This difference allows KPAS-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.452 + ], + "angle": 0, + "content": "core to assess keypoint detection from a perceptual perspective, potentially capturing nuanced failure cases that traditional distance-based metrics overlook." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.909, + 0.681 + ], + "angle": 0, + "content": "To further investigate these differences, we analyze cases where OKS and KPAScore exhibit discrepancies (see Appendix. D). We find that OKS tends to penalize predictions that deviate slightly from ground truth, even if they are perceptually acceptable, while KPAScore—by leveraging VLM-based reasoning—can tolerate minor deviations if the predicted keypoints remain semantically correct. Conversely, KPAScore may assign lower scores to cases where OKS is relatively high, particularly when the predicted keypoint location appears anatomically implausible despite being close to the ground truth. These findings highlight KPAScore's potential to complement OKS by incorporating a perceptual validation mechanism, making it a promising metric for real-world pose estimation tasks where human perception plays a crucial role." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.907, + 0.725 + ], + "angle": 0, + "content": "4.4. KPAScore as a Guide for Keypoint Detection Ensembles" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.838 + ], + "angle": 0, + "content": "In this section, we explore whether KPAScore can serve as a guiding metric for keypoint detection ensemble methods. Since KPAScore is annotation-free and evaluates keypoints in a perceptually meaningful manner, we propose an ensemble strategy that selects keypoints with the highest KPAScore from each algorithm, forming a new method we refer to as KPAPose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To further assess the effectiveness of KPAScore in guiding keypoint detection improvements, we incorporate it as an additional supervisory signal for ViTPose-\\(\\mathbf{H}^*\\) and RTMPose-1 on the COCO dataset, selecting keypoints that" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8447" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "achieve the highest estimated KPAScore across different locations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.124, + 0.484, + 0.258 + ], + "angle": 0, + "content": "As shown in Table 4, KPAPose achieves the highest scores in both KPAScore (62.65) and OKS (73.63), surpassing ViTPose-H* [27] and RTMPose-1 [12]. This demonstrates that KPAScore can serve as an effective fusion strategy for integrating multiple detection methods, producing results that not only align closely with traditional evaluation metrics (e.g., OKS) but also better capture human perceptual judgments—an advantage that annotation-dependent metrics fail to achieve." + }, + { + "type": "table", + "bbox": [ + 0.163, + 0.273, + 0.413, + 0.356 + ], + "angle": 0, + "content": "
ModelKPAScoreOKS
ViTPose-H* [27]60.9872.95
RTMPose-1 [12]59.0572.47
KPAPose62.6573.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.367, + 0.466, + 0.382 + ], + "angle": 0, + "content": "Table 4. KPAPose performance that integrates both methods" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.406, + 0.396, + 0.423 + ], + "angle": 0, + "content": "4.5. Alignment with Human Perception" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.55 + ], + "angle": 0, + "content": "To further verify the alignment between KPAScore's perception verification mechanism and human judgment, we constructed a test set featuring complex scenes and challenging actions in both indoor and outdoor environments. These include activities such as yoga, boxing, modern dance, and gymnastics, with participants wearing diverse clothing, making manual annotation of ground truth extremely difficult." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.658 + ], + "angle": 0, + "content": "To evaluate the performance of various keypoint detection algorithms in these challenging settings, we employed KPAScore alongside human ratings. A group of 20 researchers independently rated the inference results from different algorithms on a scale of 0 to 10, averaging the scores for each algorithm per image. The details of the human rating strategy are provided in Appendix E." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Figure 5 presents the evaluation results of three algorithms on three actions, along with KPAPose, an ensemble of these models, arranged in decreasing difficulty. The results show that in the absence of ground-truth annotations, KPAScore closely aligns with human evaluations. RTM [12] performed the worst in both KPAScore and human ratings, while KPAPose consistently achieved the highest scores, demonstrating KPAScore's robustness in challenging scenarios. For harder-to-estimate actions, KPAPose significantly improved KPAScore from 63.09 to 89.03 and human ratings from 6.55 to 7.65. In contrast, simpler actions saw marginal improvements, with KPAScore increasing by only 0.36, making the difference barely noticeable to humans. Moving forward, we aim to further explore KPAScore's role in evaluating complex scenarios and its potential to enhance keypoint detection algorithms." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.091, + 0.896, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.436, + 0.906, + 0.493 + ], + "angle": 0, + "content": "Figure 5. Alignment of KPA-Score and Human Rating in Complex Scenes with Complex Actions. Here, we select ViTPose, CPM, RTM, and the KPA-Pose generated based on these three algorithms, with additional visualizations provided in Appendix. F" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.54, + 0.634, + 0.556 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.889 + ], + "angle": 0, + "content": "In this paper, we introduce the KPAScore, a method that explores the keypoint understanding capabilities of vision-language models (VLMs) and proposes, to the best of our knowledge, the first human keypoint evaluation metric that eliminates the reliance on ground truth through a two-stage visual prompt question-answering approach. To validate the rationale behind this method, we introduce the Keypoint Understanding Benchmark (KPubBench), which contains over 1.8 million questions. The results demonstrate that our two-stage questioning and visual prompt approach significantly enhance the keypoint comprehension capabilities of VLMs. Furthermore, our analysis demonstrates that KPAScore exhibits variation consistency, maintains a strong correlation with traditional evaluation metrics such as OKS, and aligns closely with human perception—all while being entirely annotation-free. By leveraging VLMs for keypoint assessment, KPAScore provides a scalable, automated, and human-aligned alternative to existing evaluation methods. We hope this work inspires further research on utilizing VLMs for structured vision tasks and fosters the development of more robust, perception-driven keypoint evaluation metrics in computer vision applications." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8448" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[3] Yuanhao Cai, Zhicheng Wang, Zhengxiong Luo, Binyi Yin, Angang Du, Haoqian Wang, Xiangyu Zhang, Xinyu Zhou, Erjin Zhou, and Jian Sun. Learning delicate local representations for multi-person pose estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 455-472. Springer, 2020. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.355, + 0.484, + 0.437 + ], + "angle": 0, + "content": "[4] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.484, + 0.509 + ], + "angle": 0, + "content": "[5] Michael Dorkenwald, Nimrod Barazani, Cees GM Snoek, and Yuki M Asano. Pin: Positional insert unlocks object localisation abilities in vlms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13548-13558, 2024. 1, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.484, + 0.579 + ], + "angle": 0, + "content": "[6] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[7] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. Chatglm: A family of large language models from glm-130b to glm-4 all tools, 2024. 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.775, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[8] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y. Wu, Y.K. Li, Fuli Luo, Yingfei Xiong, and Wenfeng Liang. Deepseek-coder: When the large language model meets programming – the rise of code intelligence, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[9] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[10] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 1, 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[11] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.906, + 0.289 + ], + "angle": 0, + "content": "[12] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Real-time multi-person pose estimation based on mmpose, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.906, + 0.361 + ], + "angle": 0, + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[14] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.906, + 0.475 + ], + "angle": 0, + "content": "[15] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.547, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.631, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[18] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2025. 1, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 5, 6, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[20] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[21] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[22] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE conference on computer" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8449" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "vision and pattern recognition, pages 5693-5703, 2019. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[23] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[24] Dongkai Wang, Shiyu Xuan, and Shiliang Zhang. Locllm: Exploiting generalizable human keypoint localization via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[25] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.416 + ], + "angle": 0, + "content": "[26] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.472 + ], + "angle": 0, + "content": "[27] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In Advances in Neural Information Processing Systems, 2022. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[28] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[29] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[30] Changqian Yu, Bin Xiao, Changxin Gao, Lu Yuan, Lei Zhang, Nong Sang, and Jingdong Wang. Lite-hrnet: A lightweight high-resolution network. In CVPR, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[31] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.768 + ], + "angle": 0, + "content": "[32] Hongyi Zhu, Jia-Hong Huang, Stevan Rudinac, and Evangelos Kanoulas. Enhancing interactive image retrieval with query rewriting using large language models and vision language models. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 978-987, 2024. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.482, + 0.826 + ], + "angle": 0, + "content": "[33] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "8450" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/c6c6dbbf-396b-41fc-b50d-460da67eb4ab_origin.pdf b/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/c6c6dbbf-396b-41fc-b50d-460da67eb4ab_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f54e6ad9d48777797e729bfa6331e4e1fff01e15 --- /dev/null +++ b/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/c6c6dbbf-396b-41fc-b50d-460da67eb4ab_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0e5d2a3ee20289b256eb3f79894d63c22e7c4838f38dc3a44f0d1d07c31daa8 +size 4856469 diff --git a/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/full.md b/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5437f69870ea8f8b22c5cf8511adc2264f508f82 --- /dev/null +++ b/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/full.md @@ -0,0 +1,322 @@ +# Towards Annotation-Free Evaluation: KPAScore for Human Keypoint Detection + +Xiaoxiao Wang $^{1*}$ Chunxiao Li $^{2*}$ Peng Sun $^{3}$ Boming Miao $^{2}$ Yunjian Zhang $^{1\dagger}$ Yao Zhu $^{4\dagger}$ $^{1}$ University of Chinese Academy of Sciences, Beijing, China $^{2}$ Beijing Normal University, Beijing, China $^{3}$ Central University of Finance and Economics, Beijing, China $^{4}$ Zhejiang University, Hangzhou, China + +wangxiaoxiao@mails.ucas.ac.cn, ee-zhuy@zju.edu.cn + +# Abstract + +Human keypoint detection is fundamental in computer vision, with applications in pose estimation and action recognition. However, existing evaluation metrics (e.g., OKS, PCP, PDJ) rely on human-annotated ground truth, a labor-intensive process that increases costs, limits scalability. To address this, we propose KPAScore (KeyPoint Answering Score), an annotation-free metric independent of ground truth. It evaluates keypoint detection using a two-stage VLM-based question-answering process: first, the VLM identifies the presence of keypoints within the image, and second, visual prompts are introduced to query the likelihood of each keypoint being accurately localized within a predefined boundary. To validate the rationale behind KPAScore, we propose KPUBench (KeyPoint Understanding Benchmark), which comprehensively evaluates the VLM's ability to determine keypoint presence and localization. Extensive experiments demonstrate KPAScore's effectiveness from three perspectives: consistency to keypoint variation, correlation with traditional metrics, alignment with human perception. We hope KPAScore will reduce reliance on manual annotations, facilitating broader adoption of keypoint detection in real-world applications. + +# 1. Introduction + +The primary evaluation metrics for human keypoint detection include Object Keypoint Similarity (OKS) [17], Percentage of Detected Joints (PDJ) [28], and Percentage of Correct Keypoints (PCK) [1], all of which require ground truth for computation. Typically, ground truth is obtained in two ways. The first approach uses optical motion capture during dataset collection to generate precise ground truth, as seen in datasets like Human3.6M [11]; however, the limitations of motion capture equipment often result in datasets with limited scene diversity. The second approach relies on + +manual annotation, which is highly time-consuming, labor-intensive and prone to inaccuracies. For example, MPII [1] and COCO [17] require over 400k and 5.1 million keypoint annotations. As shown in Fig. 1, missing ground truth annotations can lead to the exclusion of some person keypoints from evaluation. Thus, developing a new metric that is annotation-free, automated, and more objective in reflecting detection performance is essential. + +With the continuous advancement of vision-language models, VLMs [2, 16, 19] have developed robust semantic understanding and strong generalization capabilities. These VLMs can tackle a variety of computer vision tasks, such as image comprehension [10] and localization [5]. Leveraging vision-language models, Lin et al. [18] propose the VQA score, which evaluates the alignment between image-text prompts using the responses generated by VLMs. This approach outperforms the widely used CLIPScore [9, 21], particularly when handling complex prompts involving combinations of objects, attributes, and relationships. Inspired by VQA score [18], we propose the Keypoint Answering Score (KPAScore), a novel keypoint detection metric based on a two-stage question-answering process and visual prompts. In the first stage, the Visual Language Model (VLM) is asked whether a specific keypoint exists in the image, and based on the response probability of presence, different weights are assigned to each keypoint. In the second stage, we introduce a visual prompt to guide the VLM's focus towards keypoint regions. Specifically, a visual prompt in the form of a blue dot is generated on the estimated keypoint, and the VLM is queried whether the keypoint accurately corresponds to this prompt. The product of the probability of answering "yes" in the second stage and the weight determines the KPAscore for the evaluated detection algorithms. + +We conducted extensive experiments to validate the rationale, consistency, correlation with ground-truth-based metrics, and alignment with human perception of the proposed KPAScore, thereby demonstrating its effectiveness. Regarding its rationale, we constructed the KeyPoint Understanding Benchmark (KPubBench) for visual language models (VLMs), which includes over 1.8 million questions. + +![](images/252797c1828c2b1d5fa2386e9660d1092a10db5653f07f4dcecb833f12731ad1.jpg) +Figure 1. Based on the visual understanding capabilities of VLM, the KPAScore is proposed as an annotation-free evaluation metric for human keypoint detection, avoiding the issues of omissions and errors commonly associated with ground truth annotations. This makes it suitable for extensive evaluation in wild settings. + +This benchmark evaluates the performance of 18 existing VLM models across multiple dimensions, including keypoint presence detection and precise localization. The evaluation results indicate that existing VLM models perform well in understanding the existence of keypoints but exhibit some deficiencies in the accuracy of keypoint localization. To optimize the performance of KPAScore in downstream tasks, we fine-tuned the model using 10,000 keypoint localization data points. The fine-tuned VLM demonstrated its capability to understand human keypoints, thus validating the rationale of using the VLM model to construct keypoint detection evaluation metrics. + +Regarding consistency, we evaluated the overall consistency between KPAScore and OKS across keypoint detection algorithms. Additionally, we assessed KPAScore's responsiveness to keypoint displacement, finding that it exhibits high variation consistency in response to changes in predicted keypoint positions. In terms of correlation with ground-truth-based metrics, we examined the relationship between KPAScore and existing evaluation metrics (OKS) on the COCO [17]. The results show that KPAScore, independent of annotations, exhibits a strong positive correlation with existing evaluation metrics. Then, we explore the broader implications of KPAScore in guiding keypoint detection and integration strategies. Finally, we validated the alignment of KPAScore with human perception on complex human images in real-world scenarios, where ground-truth annotations were unavailable, and found that KPAScore aligns well with human perception. + +Our contributions can be summarized as follows: + +- We introduce KPAScore, an annotation-free, groundtruth-independent metric for keypoint detection evaluation, showing a high positive correlation with conventional groundtruth-based metrics. +- We propose a two-stage question-answering method and visual prompting approach, which effectively enhance the + +VLM's keypoint understanding capability, thereby improving the reliability of the KPAScore. + +- We establish KPU-Bench, using over 1.8 million questions to evaluate 18 VLMs' understanding capability of human keypoints. And through fine-tuning, the performance of VLM on KPAScore is further improved. + +# 2. Related Work + +Human Keypoint Detection Metrics: In human keypoint detection tasks, commonly used evaluation metrics include PCK [1], PDJ [28], OKS [17], and OKS-based metrics like mAP and mAR [17]. OKS-based metrics are more widely adopted, with the OKS calculation formula as follows: + +$$ +\mathrm {O K S} = \frac {\sum_ {i} \exp \left(- \frac {d _ {i} ^ {2}}{2 s ^ {2} \kappa_ {i} ^ {2}}\right) \delta \left(v _ {i} > 0\right)}{\sum_ {i} \delta \left(v _ {i} > 0\right)}, \tag {1} +$$ + +where $d_{i}$ represents the Euclidean distance between the predicted and ground-truth keypoints, $s$ is the object scale, $\kappa_{i}$ is a constant representing keypoint-specific falloff, and $v_{i}$ is the visibility flag of the ground-truth keypoint $i$ . The use of the negative exponential function $\exp(-x)$ makes OKS less sensitive to the detection results. Even if there are certain errors in detecting certain key points, OKS still tends to have a higher score as long as the error is not too large. + +Based on different OKS thresholds, predicted keypoints are matched with ground-truth keypoints to calculate the precision and recall for each detection result, forming a Precision-Recall curve. AP represents the area under this curve, while mAP is the average AP across a set of thresholds, providing an overall measure of detection accuracy. mAR, in turn, reflects the model's recall capability for ground-truth keypoints across multiple OKS thresholds. + +Human keypoint detection algorithm: Human keypoint detection algorithm has evolved rapidly. Early approaches like CPM [26] focused on sequentially refining keypoint + +![](images/8f3613895074f8931179e4717475c5208031619f5c0530415d5877463ff63889.jpg) +Segmentation from detection method + +![](images/cd38116ec0a17dfeed3b691083a9ac2cb946845ebe0b07c3310b64b6205ef5f5.jpg) +Phase1: Presence Detection +Does the image include the location of the person's {left shoulder}? + +![](images/739e8fce916d4383c6a2ea6f02c2bc5bdc3d3eac95311fe46491bd1fac81df48.jpg) + +![](images/32fc663f756aec074d72c0a8bdb2d6005bbf138cf159b89aaade7f663a6c1980.jpg) + +$$ +v _ {i} = \operatorname {P r o b} \left(" Y s e "\right) = 0. 6 5 4 3 +$$ + +![](images/f12db0ef0faa0c78c4788d3ee3625e39dabe9449c78091cdd7f47a93ad2f472b.jpg) +Finetuning for $VLM_{2}$ + +![](images/018fc78cb72d4dab55ee27a58ef1083eea4295f06f7a7188f2a2abf8db523822.jpg) + +![](images/142d52ee48fb2d2a33400a033a6da20bb319d3e3278c75c478aff4a4bb6600f3.jpg) +Phase2: Precision Localization +Does the blue dot correspond to the location of the {left shoulder}? +Figure 2. We introduce the two-stage KPAScore computation for multiple occluded persons. First, bounding boxes obtained from the detection method are used to segment individuals in the image. In Stage 1, $VLM_{1}$ is queried to determine the presence of keypoints, with the probability of a "Yes" response serving as the visibility score. To enhance accuracy, $VLM_{2}$ is fine-tuned using 10k training samples before being deployed in the second stage. In Stage 2, the predicted keypoint coordinates are visualized as prompts, and $VLM_{2}$ is queried again to assess accuracy. The probability of a "Yes" response serves as the accuracy score. Finally, KPAScore integrates both visibility and accuracy scores to provide a comprehensive evaluation of keypoint detection performance. + +![](images/39520483cf3d47cb190d1507db6c3e6cd94078f2400a81b314ff04aa8f102ea4.jpg) + +![](images/a259cc492e8b656d79ae9f88898232413072224e38f9ca7b1945a1078a5b2730.jpg) + +$$ +\boldsymbol {k} _ {i} = \operatorname {P r o b} \left(" Y s e "\right) = 0. 8 9 5 5 +$$ + +$$ +K P A S c o r e _ {p} = \frac {\sum_ {i = 1} ^ {n} \sigma \left(v _ {i}\right) \sqrt {k _ {i} v _ {i}}}{\sum_ {i = 1} ^ {n} \sigma \left(v _ {i}\right)} = 0. 9 0 1 2 +$$ + +predictions through multi-stage processing. HRNet [22] proposed a high-resolution network architecture that maintains spatial detail across the network. LiteHRNet [30] later advanced keypoint detection by maintaining high-resolution features while optimizing for computational efficiency. RSN [3] introduced a method for aggregating intra-level features, preserving rich spatial information. More recent methods, such as RTMPose [12] and AlphaPose [6], adopt a top-down framework, utilizing object detectors to first locate each individual before estimating poses. RTMPose [12] excels in complex environments by isolating instances, while AlphaPose [6] handles occlusions and challenging poses efficiently, making it ideal for real-time scenarios. ViTPose [27], applies vision transformers as non-hierarchical backbones to capture global features. VLM has shown significant progress in keypoint detection, LocLLM [24] achieved comparable performance to ViT-Pose [27] on COCO dataset, highlighting its capability for keypoint understanding, though it has not yet been applied in evaluation tasks. Appendix. A give a supplement to related work. + +# 3. Method + +In this section, we introduce the KPAScore, to the best of our knowledge, the first model-based evaluation metric for keypoint detection. This metric leverages the visual under + +standing capabilities of large vision-language models, eliminating the reliance on human annotations inherent in traditional groundtruth-based metrics. An overview of the KPA-Score calculation process is illustrated in Fig. 2 and the KPAScore can be formulated as: + +$$ +\mathrm {K P A S c o r e} _ {p} = \frac {\sum_ {i = 1} ^ {n} \sigma \left(v _ {i}\right) \sqrt {k _ {i} \cdot v _ {i}}}{\sum_ {i = 1} ^ {n} \sigma \left(v _ {i}\right)}, \tag {2} +$$ + +where the indicator function $\sigma(v_{i})$ is defined as: + +$$ +\sigma \left(v _ {i}\right) = \left\{ \begin{array}{l l} 1, & v _ {i} > 0. 5 \\ 0, & v _ {i} \leq 0. 5 \end{array} . \right. \tag {3} +$$ + +Here, $v_{i}$ represents the keypoint visibility score, obtained from the first-stage presence detection. If $v_{i} < 0.5$ , the keypoint is considered not existed. For values between 0.5 and 1, $v_{i}$ reflects the confidence of the VLM in the keypoint's presence: lower values indicate that the keypoint is likely occluded, while higher values suggest it is unobstructed. Similarly, $k_{i}$ represents the keypoint accuracy score, derived from the second-stage precision localization. The denominator ensures the normalization of contributions from all visible keypoints and the square operation avoids that the product of two numbers less than 1 is too small, maintaining a balanced evaluation. + +# 3.1. First-Stage Presence Detection + +In the first stage, we leverage vision-language models (VLMs) to determine whether specific keypoints are present in the image. Image $I$ is the segmented images of individual persons which obtained from object detection. Given the detection task's keypoint list with $N$ keypoints $P = [p_{1}, p_{2}, \ldots, p_{N}]$ , for instance: + +$$ +P = \left[ ^ {\prime \prime} \text {N o s e} ^ {\prime \prime}, ^ {\prime \prime} \text {L e f t E y} ^ {\prime \prime}, \dots , ^ {\prime \prime} \text {R i g h t W i r s t} ^ {\prime \prime} \right], \tag {4} +$$ + +we query the VLM for each keypoint $p_i$ in the detection task's keypoint list with the question $q(p_i)$ as: + +$$ +p _ {i} = \text {" n o s e "} +$$ + +$$ +q _ {1} \left(p _ {i}\right) = \text {D o e s t h e i m a g e i n c l u d e t h e l o c a t i o n o f t h e} \tag {5} +$$ + +$$ +\text {p e r s o n ' s " n o s e "} \text {? P l a c e a n s w e r y e s o r n o .} +$$ + +The VLM returns a probability score for answering "Yes," which is taken as the visibility score $v_{i}$ : + +$$ +v _ {i} = \operatorname {P r o b} \left(" \text {Y e s}" \mid I, q _ {1} \left(k _ {i}\right)\right), \tag {6} +$$ + +If $v_{i} \leq 0.5$ , the keypoint is considered not existential and is excluded from subsequent precision evaluation. + +# 3.2. Second-Stage Precision Localization + +In the second stage, we perform a more refined evaluation of the detection method by assessing the accuracy of the estimated keypoint coordinates. + +VLMs are generally not well-suited for directly evaluating keypoint coordinate accuracy. This limitation likely arises from the inherent difficulty VLMs face in establishing a precise mapping between textual descriptions and specific spatial locations within an image. To overcome this challenge, we incorporate visual prompts [13, 15, 33] to enhance the VLMs' ability to perceive and interpret keypoint locations. Specifically, we overlay a blue dot $d_{i}$ at the keypoint coordinates predicted by the detection algorithm, using it as a visual cue to facilitate keypoint accuracy evaluation. + +Given an image embedded with a visual prompt, denoted as $I_{d_i}$ , we query the VLM with the following question: + +$$ +p _ {i} = \text {" n o s e "} +$$ + +$$ +q _ {2} \left(p _ {i}\right) = \text {D o e s t h e b l u e d o t c o r r e s p o n d t o t h e l o c a t i o n (7)} +$$ + +$$ +\begin{array}{l} \text {o f t h e " n o s e" ? P l a c e a n s w e r Y e s o r N o .} \end{array} +$$ + +The generative likelihood of the VLM responding "Yes" is then computed as: + +$$ +k _ {i} = \operatorname {P r o b} \left(" \text {Y e s}" \mid I _ {d _ {i}}, q _ {2} \left(p _ {i}\right)\right), \tag {8} +$$ + +where $k_{i}$ serves as the accuracy score for the given keypoint in this second stage. Finally, the model's overall KPAScore is obtained by computing a weighted average of all keypoint scores, integrating the evaluation results from both stages as shown in Eq. (2). + +# 4. Experiments + +This section conducts extensive experiments to answer the following questions: + +- Can VLMs comprehend human keypoints? Additionally, how can we further enhance VLMs' capability to understand keypoints? See Sec. 4.1. +- Can KPAScore consistently reflect changes in predicted keypoints relative to the ground truth? See Sec. 4.2. +- Is there a correlation between KPAScore and existing annotation-dependent metrics? See Sec. 4.3. +- Can KPAScore be considered as a guide for keypoint detection ensemble methods? See Sec. 4.4 +- Does the KPAScore align with human perception in complex scenarios without ground truth? See Sec. 4.5. + +# 4.1. Keypoint Understanding Benchmark + +With the widespread application of VLMs in computer vision tasks, their ability to accurately understand human keypoints remains an open question. We evaluate the keypoint understanding capability of VLMs from two perspectives: presence detection and precision localization. To facilitate this evaluation, we introduce the Keypoint Understanding Benchmark (KPUBench), a large-scale dataset comprising 1.8 million structured questions derived from COCO [17], MPII [1], and Human3.6M [11]. The dataset is systematically split into training and test sets with an approximate ratio of 9.5:0.5, and all evaluations are conducted exclusively on the test set. + +The relationship between KPUBench and KPAScore can be summarized as follows: KPUBench serves as a benchmark for assessing the ability of existing VLMs to comprehend human keypoints in an unsupervised or zero-shot setting, measuring their fine-grained semantic and visual perception capabilities. The best-performing VLM from this evaluation is then employed in downstream tasks, including the construction of KPAScore. The architecture of KPUBench is illustrated in Figure 3 and the evaluation results of 18 visual-language models is presented in Tab. 1. + +(a) Presence Detection assesses the VLM's ability to identify whether a specific keypoint is present in an image. We query the VLM: "Does the image include the location of the person's {keypoint name}?". Ideally, for keypoints annotated as present (COCO visibility scores of 1 or 2), the VLM should return "Yes" with a probability above 0.5, whereas for absent keypoints (visibility score of 0), this probability should be below 0.5. The positive-to-negative sample ratio is approximately 5.31:4.69. To evaluate VLM reliability and confidence in keypoint presence detection, we compute accuracy $(ACC)$ and probabilistic expectation $(E(Prob))$ . + +As shown in the third and fourth columns of Tab. 1, most VLMs demonstrate a fundamental ability to recognize keypoints, though model performance does not clearly correlate + +![](images/b7208c297a8c898e031abad7c49fbada1727edfcc150df2975c42c5e58acd06f.jpg) +(c) Precision Localization Detection—Visual Prompt Understanding Capability +Figure 3. Overview of the KPUBench evaluation framework. (a): Evaluates whether a VLM can determine the existence of a keypoint in an image. (b): Assesses the VLM's ability to recognize keypoints based on explicit numerical coordinates. (c): Measures whether the VLM can accurately associate a visual cue (a blue dot) with the corresponding keypoint. The expected answer is "Yes" for correct localizations and "No" for incorrect ones. + +
ModelSizePresence DetectionPrecision Localization Detection
ACC(%)E(Prob)CoordinateVisual Prompt
ACC(%)E(Prob)ACC(%)E(Prob)
LLaVA-OneVersion [14]0.5B55.5849.7150.0050.2452.3250.30
DeepSeek-Coder-Base [8]1.3B62.0352.7749.9149.9655.5750.86
InternVL2-2B [4]2B56.3047.3350.0950.0655.2551.75
Qwen2-VL-2B [25]2B73.6758.7050.0050.0456.5451.21
MiniCPM-V 2.0 [29]2.8B58.4344.4249.9449.9454.8652.25
LLaVA-V1.5-7B [19]7B61.0057.1849.8649.7053.3250.72
LLaVA-V1.6-7B [19]7B75.1651.2149.7349.9254.8451.21
Qwen2-VL-7B [25]7B66.9857.6150.2250.1160.5552.71
InternVL2-8B [4]8B56.3047.3350.9550.0655.2551.76
MiniCPM-V-2.6 [29]8B54.8451.2150.8550.1153.0450.97
GLM-4V-9B [7]9B76.9949.8749.9849.7662.5758.17
LLama3.2-11B-coco [23]11B61.2055.0848.7549.0156.0254.53
LLaVA-V1.5-13B [19]13B69.8758.8250.0049.9153.5050.61
LLaVA-V1.6-13B [19]13B71.5255.9349.9949.0458.9651.32
Cogglm2-LLama3-chat-19B [10]19B70.3354.0649.9849.9854.7550.86
InternVL2-26B [4]26B61.8553.1750.0350.1957.2353.89
LLaVA-V1.6-34B [19]34B84.6765.1249.9549.9050.0050.35
Qwen2-VL-72B [25]72B79.5463.6050.8450.1556.0752.96
GLM-4V-9B-finetuig9B----80.1574.32
LLaVA + GLM-finetuig43B84.6765.12--84.5378.10
+ +Table 1. Performance of 18 VLM models on KPUBench. KPUBench evaluates VLMs' understanding of keypoint detection from two perspectives: presence detection and precision localization detection, providing a comprehensive measurement using accuracy and probabilistic expectation. + +with model size. Both the LLaVA-V1.6 and Qwen families perform well, exceeding $66.98\%$ accuracy, likely due to task-related training data or effective architectures. Among them, LLaVA-V1.6-34B [19] achieves the highest accuracy $(84.67\%)$ and probabilistic expectation $(65.12\%)$ , indicating strong keypoint perception. In contrast, GLM-4V [7] attains $76.99\%$ accuracy but a lower probabilistic expectation $(49.87\%)$ , suggesting lower confidence despite correct predictions, which may affect reliability in practical use. + +Precision Localization Detection consists of two components: coordinate understanding and visual prompt understanding. We use human-annotated keypoint coordinates as ground truth, treating them as positive samples. Negative samples are generated by shifting each keypoint randomly by $10\%$ of the bounding box size, maintaining a 1:1 positive-to-negative ratio. + +- Coordinate Understanding We ask the VLM: "Does the coordinate $(x, y)$ correspond to the location of the {keypoint name}?" Accuracy and probabilistic expectation are computed, where higher accuracy indicates better alignment with human judgment, and higher probabilistic expectation reflects the model's confidence. +- Visual Prompt Understanding Using the same positive and negative samples, we mark a blue dot at the keypoint location and ask: "Does the blue dot correspond to the location of the {keypoint name}?" Accuracy and probabilistic expectation are measured similarly. The effect of the color of the visual prompt is illustrated in Appendix. B. + +Based on the fifth and sixth columns of Tab. 1, VLMs show limited ability in interpreting numerical keypoint coordinates, with accuracy around $50\%$ , near random guessing. This suggests difficulty in mapping numerical coordinates to spatial locations in an image. Introducing visual prompts significantly improves keypoint comprehension, increasing accuracy by an average of $5.42\%$ . Notably, GLM-4V's accuracy rose from $49.98\%$ to $62.57\%$ , with its probabilistic expectation improving from $49.76\%$ to $58.17\%$ , confirming that visual prompts help direct model attention to spatial locations. However, despite these gains, performance remains insufficient for practical applications. + +To further enhance the model's precision, we fine-tuned GLM-4V-9B using the LoRA method with 10k Precision Localization Detection training samples containing visual prompts (detailed hyperparameter settings are provided in Appendix. C). The fine-tuned model, GLM-4V-9B-finetuned, achieved a remarkable accuracy of $80.15\%$ and a probabilistic expectation of $78.10\%$ , demonstrating that even a simple fine-tuning process can significantly enhance VLM performance on specialized tasks. + +Finally, we implemented a two-stage approach by employing LLaVA-V1.6-34B as the Presence Detection module in the first stage. We retained only the keypoints identi + +fied as present by LLaVA-V1.6-34B and fed them into the Precision Localization Detection stage. This LLaVA-V1.6-34B + GLM-4V-9B-finetuned two-stage framework further enhanced the model's keypoint localization understanding, achieving an accuracy of $84.53\%$ and a probabilistic expectation of $78.10\%$ . Therefore, we propose a two-stage evaluation framework that integrates presence detection and precision localization to systematically assess keypoint detection models using VLMs. + +4.2. The Consistency of KPAscore + +
ModelKPAScoreOKSmAP
ViTPose-H* [27]60.9872.9579.20
RTMPose-1 [12]59.0572.4777.30
RSN [3]59.0472.1075.00
HRNet [22]59.2772.1474.91
Alphapose [6]58.1371.4670.61
CPM [26]57.2470.1465.23
LiteHRNet [30]57.0870.3064.24
+ +Table 2. The overall scores of seven keypoint detection algorithms on the COCO dataset. + +After establishing the LLaVA-V1.6-34B + GLM-4V-9B-finetuned two-stage KPAScore framework, we conducted a comprehensive evaluation of seven mainstream keypoint detection methods on the COCO dataset. The selected models include ViTPose [27], RTMPose [12], Alphapose [6], HRNet [22], RSN [3], LiteHRNet [30], and CPM [26]. + +Consistency in response to changes is a critical factor in evaluating the usability of a metric. Here, we demonstrate that KPAScore possesses both overall consistency and change consistency from two perspectives. + +First, we present the average KPAScore, OKS, and mAP of the seven methods on COCO (see Table 2). As shown, KPAScore exhibits strong alignment with traditional keypoint metrics (OKS and $mAP$ ). Models with higher OKS and mAP scores, such as ViTPose-H and RTMPose-1, also achieve the highest KPAScore, while lower-performing models, like CPM and LiteHRNet, similarly rank lower under KPAScore. This consistency validates KPAScore as an effective metric that aligns with established evaluation criteria while leveraging VLM-based assessment to provide a more holistic keypoint detection evaluation. + +Second, we apply varying intensities of offset to the keypoint detection results for each detection algorithm, then recalculated the KPAScore. We analyzed the KPAScore across different offset intensities, where the offset is defined as the square root of the bounding box area (a standard approach on the COCO dataset) multiplied by a factor ranging from $0\%$ to $50\%$ , with random offset directions which shown in Fig. 4a + +As shown in Fig. 4b, the KPAScore for all 7 keypoint + +(a) Offset Example. +![](images/5f4818e4fa8cdc639b531a6d7a7b2d8e706fe4e1b5fc9dc6b0f7c93bda8878e6.jpg) +- 3% offset +- 5% offset +- 10% offset +- 15% offset +- 20% offset +- 25% offset +- 30% offset +- 50% offset + +![](images/1a39f267bc48dc7382fd9edc78c097dd580ff37cb70032b308f0a25639a92e17.jpg) +(b) Variation of KPAScore and OKS. +Figure 4. (a): Circles of different colors represent the corresponding offset keypoint positions at various offset intensities. (b): A line chart illustrating the variation in values of KPAScore (red line) and OKS (blue line) at different offset intensities. + +
MethodViTPose-H* [27]RSN [3]HRNet [22]RTMPose-1 [12]Alphapose [6]CPM [26]LiteHRNet [30]Average
Corr0.720.710.720.710.700.690.730.71
+ +Table 3. The Pearson correlation coefficients between the KPAScore and OKS. + +detection methods steadily decreases with increasing offset intensity, indicating that the KPAScore effectively responds to variations in keypoint positions and can serve as a reliable new evaluation metric. Specifically, offset intensities of $3\%$ and $5\%$ result in a decrease in the KPAScore, but the decline is not significant. As illustrated in Fig. 4a, offsets within $5\%$ are not easily perceptible to the human eye, and therefore the minor drop in score is consistent with human perception. When the offset intensity ranges from $10\%$ to $30\%$ , the KPAScore decreases more markedly, and at a $50\%$ offset intensity, the scores for all algorithms converge, KPAScore is almost 0. OKS also decreases steadily with the offset, but when the offset strength is $50\%$ , OKS is still above 50, which shows that OKS is not very sensitive to the feedback of the offset. + +# 4.3. The Correlation Between KPAScore and OKS + +Beyond evaluating the consistency of KPAScore, we examine its relationship with OKS by computing their Pearson correlation across various keypoint detection methods, as shown in Tab. 3. The results show a strong positive correlation, ranging from 0.69 to 0.73 (average 0.71), indicating that KPAScore aligns well with OKS in assessing keypoint detection quality, further validating its reliability as an evaluation metric. + +However, it is worth noting that the correlation is not perfect $(\mathrm{Corr} < 1.0)$ , indicating that KPAScore captures aspects of keypoint detection performance that may not be fully reflected by OKS. Unlike OKS, which primarily relies on Euclidean distance-based similarity, KPAScore incorporates a human-centric evaluation approach using visual-language models (VLMs). This difference allows KPAS- + +core to assess keypoint detection from a perceptual perspective, potentially capturing nuanced failure cases that traditional distance-based metrics overlook. + +To further investigate these differences, we analyze cases where OKS and KPAScore exhibit discrepancies (see Appendix. D). We find that OKS tends to penalize predictions that deviate slightly from ground truth, even if they are perceptually acceptable, while KPAScore—by leveraging VLM-based reasoning—can tolerate minor deviations if the predicted keypoints remain semantically correct. Conversely, KPAScore may assign lower scores to cases where OKS is relatively high, particularly when the predicted keypoint location appears anatomically implausible despite being close to the ground truth. These findings highlight KPAScore's potential to complement OKS by incorporating a perceptual validation mechanism, making it a promising metric for real-world pose estimation tasks where human perception plays a crucial role. + +# 4.4. KPAScore as a Guide for Keypoint Detection Ensembles + +In this section, we explore whether KPAScore can serve as a guiding metric for keypoint detection ensemble methods. Since KPAScore is annotation-free and evaluates keypoints in a perceptually meaningful manner, we propose an ensemble strategy that selects keypoints with the highest KPAScore from each algorithm, forming a new method we refer to as KPAPose. + +To further assess the effectiveness of KPAScore in guiding keypoint detection improvements, we incorporate it as an additional supervisory signal for ViTPose- $\mathbf{H}^*$ and RTMPose-1 on the COCO dataset, selecting keypoints that + +achieve the highest estimated KPAScore across different locations. + +As shown in Table 4, KPAPose achieves the highest scores in both KPAScore (62.65) and OKS (73.63), surpassing ViTPose-H* [27] and RTMPose-1 [12]. This demonstrates that KPAScore can serve as an effective fusion strategy for integrating multiple detection methods, producing results that not only align closely with traditional evaluation metrics (e.g., OKS) but also better capture human perceptual judgments—an advantage that annotation-dependent metrics fail to achieve. + +
ModelKPAScoreOKS
ViTPose-H* [27]60.9872.95
RTMPose-1 [12]59.0572.47
KPAPose62.6573.63
+ +Table 4. KPAPose performance that integrates both methods + +# 4.5. Alignment with Human Perception + +To further verify the alignment between KPAScore's perception verification mechanism and human judgment, we constructed a test set featuring complex scenes and challenging actions in both indoor and outdoor environments. These include activities such as yoga, boxing, modern dance, and gymnastics, with participants wearing diverse clothing, making manual annotation of ground truth extremely difficult. + +To evaluate the performance of various keypoint detection algorithms in these challenging settings, we employed KPAScore alongside human ratings. A group of 20 researchers independently rated the inference results from different algorithms on a scale of 0 to 10, averaging the scores for each algorithm per image. The details of the human rating strategy are provided in Appendix E. + +Figure 5 presents the evaluation results of three algorithms on three actions, along with KPAPose, an ensemble of these models, arranged in decreasing difficulty. The results show that in the absence of ground-truth annotations, KPAScore closely aligns with human evaluations. RTM [12] performed the worst in both KPAScore and human ratings, while KPAPose consistently achieved the highest scores, demonstrating KPAScore's robustness in challenging scenarios. For harder-to-estimate actions, KPAPose significantly improved KPAScore from 63.09 to 89.03 and human ratings from 6.55 to 7.65. In contrast, simpler actions saw marginal improvements, with KPAScore increasing by only 0.36, making the difference barely noticeable to humans. Moving forward, we aim to further explore KPAScore's role in evaluating complex scenarios and its potential to enhance keypoint detection algorithms. + +![](images/212337c1d4e4b5a9c9ec8d189741cdfb66d8a8487fb1922095e3653efa3d5bce.jpg) +Figure 5. Alignment of KPA-Score and Human Rating in Complex Scenes with Complex Actions. Here, we select ViTPose, CPM, RTM, and the KPA-Pose generated based on these three algorithms, with additional visualizations provided in Appendix. F + +# 5. Conclusion + +In this paper, we introduce the KPAScore, a method that explores the keypoint understanding capabilities of vision-language models (VLMs) and proposes, to the best of our knowledge, the first human keypoint evaluation metric that eliminates the reliance on ground truth through a two-stage visual prompt question-answering approach. To validate the rationale behind this method, we introduce the Keypoint Understanding Benchmark (KPubBench), which contains over 1.8 million questions. The results demonstrate that our two-stage questioning and visual prompt approach significantly enhance the keypoint comprehension capabilities of VLMs. Furthermore, our analysis demonstrates that KPAScore exhibits variation consistency, maintains a strong correlation with traditional evaluation metrics such as OKS, and aligns closely with human perception—all while being entirely annotation-free. By leveraging VLMs for keypoint assessment, KPAScore provides a scalable, automated, and human-aligned alternative to existing evaluation methods. We hope this work inspires further research on utilizing VLMs for structured vision tasks and fosters the development of more robust, perception-driven keypoint evaluation metrics in computer vision applications. + +# References + +[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 1, 2, 4 +[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 1 +[3] Yuanhao Cai, Zhicheng Wang, Zhengxiong Luo, Binyi Yin, Angang Du, Haoqian Wang, Xiangyu Zhang, Xinyu Zhou, Erjin Zhou, and Jian Sun. Learning delicate local representations for multi-person pose estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 455-472. Springer, 2020. 3, 6, 7 +[4] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023.5 +[5] Michael Dorkenwald, Nimrod Barazani, Cees GM Snoek, and Yuki M Asano. Pin: Positional insert unlocks object localisation abilities in vlms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13548-13558, 2024. 1, 11 +[6] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3, 6, 7 +[7] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. Chatglm: A family of large language models from glm-130b to glm-4 all tools, 2024. 5, 6, 11 +[8] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y. Wu, Y.K. Li, Fuli Luo, Yingfei Xiong, and Wenfeng Liang. Deepseek-coder: When the large language model meets programming – the rise of code intelligence, 2024. 5 +[9] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 1 + +[10] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 1, 5, 11 +[11] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 1, 4 +[12] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Real-time multi-person pose estimation based on mmpose, 2023. 3, 6, 7, 8 +[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4 +[14] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 5 +[15] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 4 +[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1 +[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1, 2, 4 +[18] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2025. 1, 11 +[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 5, 6, 11 +[20] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 11 +[21] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1 +[22] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE conference on computer + +vision and pattern recognition, pages 5693-5703, 2019. 3, 6, 7 +[23] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 11 +[24] Dongkai Wang, Shiyu Xuan, and Shiliang Zhang. Locllm: Exploiting generalizable human keypoint localization via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[25] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 11 +[26] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7 +[27] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In Advances in Neural Information Processing Systems, 2022. 3, 6, 7, 8 +[28] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. 1, 2 +[29] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 5 +[30] Changqian Yu, Bin Xiao, Changxin Gao, Lu Yuan, Lei Zhang, Nong Sang, and Jingdong Wang. Lite-hrnet: A lightweight high-resolution network. In CVPR, 2021. 3, 6, 7 +[31] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 11 +[32] Hongyi Zhu, Jia-Hong Huang, Stevan Rudinac, and Evangelos Kanoulas. Enhancing interactive image retrieval with query rewriting using large language models and vision language models. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 978-987, 2024. 11 +[33] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 4 \ No newline at end of file diff --git a/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/images.zip b/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..44f40e119d56d9df65e603a05a7406efd452b7b7 --- /dev/null +++ b/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3188afb77578302c06d087b7f6c0655b7a14fdf0e7750c5f95ee772f194cbac8 +size 655458 diff --git a/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/layout.json b/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..738201ca525ba4829942c608ff206112bffaad82 --- /dev/null +++ b/2025/Towards Annotation-Free Evaluation_ KPAScore for Human Keypoint Detection/layout.json @@ -0,0 +1,7774 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 553, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 553, + 121 + ], + "type": "text", + "content": "Towards Annotation-Free Evaluation: KPAScore for Human Keypoint Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "text", + "content": "Xiaoxiao Wang" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "text", + "content": " Chunxiao Li" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "text", + "content": " Peng Sun" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "text", + "content": " Boming Miao" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "text", + "content": " Yunjian Zhang" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "text", + "content": " Yao Zhu" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences, Beijing, China " + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "text", + "content": "Beijing Normal University, Beijing, China " + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "text", + "content": "Central University of Finance and Economics, Beijing, China " + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 578, + 186 + ], + "type": "text", + "content": "Zhejiang University, Hangzhou, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 187, + 451, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 187, + 451, + 198 + ], + "spans": [ + { + "bbox": [ + 186, + 187, + 451, + 198 + ], + "type": "text", + "content": "wangxiaoxiao@mails.ucas.ac.cn, ee-zhuy@zju.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 253, + 297, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 297, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 297, + 517 + ], + "type": "text", + "content": "Human keypoint detection is fundamental in computer vision, with applications in pose estimation and action recognition. However, existing evaluation metrics (e.g., OKS, PCP, PDJ) rely on human-annotated ground truth, a labor-intensive process that increases costs, limits scalability. To address this, we propose KPAScore (KeyPoint Answering Score), an annotation-free metric independent of ground truth. It evaluates keypoint detection using a two-stage VLM-based question-answering process: first, the VLM identifies the presence of keypoints within the image, and second, visual prompts are introduced to query the likelihood of each keypoint being accurately localized within a predefined boundary. To validate the rationale behind KPAScore, we propose KPUBench (KeyPoint Understanding Benchmark), which comprehensively evaluates the VLM's ability to determine keypoint presence and localization. Extensive experiments demonstrate KPAScore's effectiveness from three perspectives: consistency to keypoint variation, correlation with traditional metrics, alignment with human perception. We hope KPAScore will reduce reliance on manual annotations, facilitating broader adoption of keypoint detection in real-world applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 542, + 135, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 542, + 135, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 542, + 135, + 554 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 563, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 682 + ], + "type": "text", + "content": "The primary evaluation metrics for human keypoint detection include Object Keypoint Similarity (OKS) [17], Percentage of Detected Joints (PDJ) [28], and Percentage of Correct Keypoints (PCK) [1], all of which require ground truth for computation. Typically, ground truth is obtained in two ways. The first approach uses optical motion capture during dataset collection to generate precise ground truth, as seen in datasets like Human3.6M [11]; however, the limitations of motion capture equipment often result in datasets with limited scene diversity. The second approach relies on" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 324 + ], + "type": "text", + "content": "manual annotation, which is highly time-consuming, labor-intensive and prone to inaccuracies. For example, MPII [1] and COCO [17] require over 400k and 5.1 million keypoint annotations. As shown in Fig. 1, missing ground truth annotations can lead to the exclusion of some person keypoints from evaluation. Thus, developing a new metric that is annotation-free, automated, and more objective in reflecting detection performance is essential." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 327, + 556, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 556, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 556, + 626 + ], + "type": "text", + "content": "With the continuous advancement of vision-language models, VLMs [2, 16, 19] have developed robust semantic understanding and strong generalization capabilities. These VLMs can tackle a variety of computer vision tasks, such as image comprehension [10] and localization [5]. Leveraging vision-language models, Lin et al. [18] propose the VQA score, which evaluates the alignment between image-text prompts using the responses generated by VLMs. This approach outperforms the widely used CLIPScore [9, 21], particularly when handling complex prompts involving combinations of objects, attributes, and relationships. Inspired by VQA score [18], we propose the Keypoint Answering Score (KPAScore), a novel keypoint detection metric based on a two-stage question-answering process and visual prompts. In the first stage, the Visual Language Model (VLM) is asked whether a specific keypoint exists in the image, and based on the response probability of presence, different weights are assigned to each keypoint. In the second stage, we introduce a visual prompt to guide the VLM's focus towards keypoint regions. Specifically, a visual prompt in the form of a blue dot is generated on the estimated keypoint, and the VLM is queried whether the keypoint accurately corresponds to this prompt. The product of the probability of answering \"yes\" in the second stage and the weight determines the KPAscore for the evaluated detection algorithms." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "We conducted extensive experiments to validate the rationale, consistency, correlation with ground-truth-based metrics, and alignment with human perception of the proposed KPAScore, thereby demonstrating its effectiveness. Regarding its rationale, we constructed the KeyPoint Understanding Benchmark (KPubBench) for visual language models (VLMs), which includes over 1.8 million questions." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8441" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 76, + 537, + 216 + ], + "blocks": [ + { + "bbox": [ + 71, + 76, + 537, + 216 + ], + "lines": [ + { + "bbox": [ + 71, + 76, + 537, + 216 + ], + "spans": [ + { + "bbox": [ + 71, + 76, + 537, + 216 + ], + "type": "image", + "image_path": "252797c1828c2b1d5fa2386e9660d1092a10db5653f07f4dcecb833f12731ad1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 232, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 54, + 232, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 54, + 232, + 555, + 266 + ], + "type": "text", + "content": "Figure 1. Based on the visual understanding capabilities of VLM, the KPAScore is proposed as an annotation-free evaluation metric for human keypoint detection, avoiding the issues of omissions and errors commonly associated with ground truth annotations. This makes it suitable for extensive evaluation in wild settings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 277, + 294, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 277, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 277, + 294, + 420 + ], + "type": "text", + "content": "This benchmark evaluates the performance of 18 existing VLM models across multiple dimensions, including keypoint presence detection and precise localization. The evaluation results indicate that existing VLM models perform well in understanding the existence of keypoints but exhibit some deficiencies in the accuracy of keypoint localization. To optimize the performance of KPAScore in downstream tasks, we fine-tuned the model using 10,000 keypoint localization data points. The fine-tuned VLM demonstrated its capability to understand human keypoints, thus validating the rationale of using the VLM model to construct keypoint detection evaluation metrics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 422, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 626 + ], + "type": "text", + "content": "Regarding consistency, we evaluated the overall consistency between KPAScore and OKS across keypoint detection algorithms. Additionally, we assessed KPAScore's responsiveness to keypoint displacement, finding that it exhibits high variation consistency in response to changes in predicted keypoint positions. In terms of correlation with ground-truth-based metrics, we examined the relationship between KPAScore and existing evaluation metrics (OKS) on the COCO [17]. The results show that KPAScore, independent of annotations, exhibits a strong positive correlation with existing evaluation metrics. Then, we explore the broader implications of KPAScore in guiding keypoint detection and integration strategies. Finally, we validated the alignment of KPAScore with human perception on complex human images in real-world scenarios, where ground-truth annotations were unavailable, and found that KPAScore aligns well with human perception." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 628, + 254, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 254, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 254, + 639 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "type": "text", + "content": "- We introduce KPAScore, an annotation-free, groundtruth-independent metric for keypoint detection evaluation, showing a high positive correlation with conventional groundtruth-based metrics." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "- We propose a two-stage question-answering method and visual prompting approach, which effectively enhance the" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 277, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 277, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 322, + 277, + 553, + 300 + ], + "type": "text", + "content": "VLM's keypoint understanding capability, thereby improving the reliability of the KPAScore." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 301, + 555, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 350 + ], + "type": "text", + "content": "- We establish KPU-Bench, using over 1.8 million questions to evaluate 18 VLMs' understanding capability of human keypoints. And through fine-tuning, the performance of VLM on KPAScore is further improved." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 360, + 400, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 400, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 400, + 372 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 380, + 554, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 440 + ], + "type": "text", + "content": "Human Keypoint Detection Metrics: In human keypoint detection tasks, commonly used evaluation metrics include PCK [1], PDJ [28], OKS [17], and OKS-based metrics like mAP and mAR [17]. OKS-based metrics are more widely adopted, with the OKS calculation formula as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 448, + 553, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 448, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 354, + 448, + 553, + 483 + ], + "type": "interline_equation", + "content": "\\mathrm {O K S} = \\frac {\\sum_ {i} \\exp \\left(- \\frac {d _ {i} ^ {2}}{2 s ^ {2} \\kappa_ {i} ^ {2}}\\right) \\delta \\left(v _ {i} > 0\\right)}{\\sum_ {i} \\delta \\left(v _ {i} > 0\\right)}, \\tag {1}", + "image_path": "bf8d143093d316738f37b78ded557e58242ec8ee9a2d87b4adc4018d8213bff2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": " represents the Euclidean distance between the predicted and ground-truth keypoints, " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": " is the object scale, " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\kappa_{i}" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": " is a constant representing keypoint-specific falloff, and " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": " is the visibility flag of the ground-truth keypoint " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": ". The use of the negative exponential function " + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "inline_equation", + "content": "\\exp(-x)" + }, + { + "bbox": [ + 313, + 486, + 554, + 581 + ], + "type": "text", + "content": " makes OKS less sensitive to the detection results. Even if there are certain errors in detecting certain key points, OKS still tends to have a higher score as long as the error is not too large." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": "Based on different OKS thresholds, predicted keypoints are matched with ground-truth keypoints to calculate the precision and recall for each detection result, forming a Precision-Recall curve. AP represents the area under this curve, while mAP is the average AP across a set of thresholds, providing an overall measure of detection accuracy. mAR, in turn, reflects the model's recall capability for ground-truth keypoints across multiple OKS thresholds." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Human keypoint detection algorithm: Human keypoint detection algorithm has evolved rapidly. Early approaches like CPM [26] focused on sequentially refining keypoint" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8442" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 97, + 227, + 221 + ], + "blocks": [ + { + "bbox": [ + 72, + 82, + 224, + 92 + ], + "lines": [ + { + "bbox": [ + 72, + 82, + 224, + 92 + ], + "spans": [ + { + "bbox": [ + 72, + 82, + 224, + 92 + ], + "type": "text", + "content": "Segmentation from detection method" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 97, + 227, + 221 + ], + "lines": [ + { + "bbox": [ + 69, + 97, + 227, + 221 + ], + "spans": [ + { + "bbox": [ + 69, + 97, + 227, + 221 + ], + "type": "image", + "image_path": "8f3613895074f8931179e4717475c5208031619f5c0530415d5877463ff63889.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 260, + 107, + 304, + 186 + ], + "blocks": [ + { + "bbox": [ + 261, + 81, + 375, + 91 + ], + "lines": [ + { + "bbox": [ + 261, + 81, + 375, + 91 + ], + "spans": [ + { + "bbox": [ + 261, + 81, + 375, + 91 + ], + "type": "text", + "content": "Phase1: Presence Detection" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 96, + 537, + 106 + ], + "lines": [ + { + "bbox": [ + 306, + 96, + 537, + 106 + ], + "spans": [ + { + "bbox": [ + 306, + 96, + 537, + 106 + ], + "type": "text", + "content": "Does the image include the location of the person's {left shoulder}?" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 260, + 107, + 304, + 186 + ], + "lines": [ + { + "bbox": [ + 260, + 107, + 304, + 186 + ], + "spans": [ + { + "bbox": [ + 260, + 107, + 304, + 186 + ], + "type": "image", + "image_path": "cd38116ec0a17dfeed3b691083a9ac2cb946845ebe0b07c3310b64b6205ef5f5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 137, + 351, + 155 + ], + "blocks": [ + { + "bbox": [ + 315, + 137, + 351, + 155 + ], + "lines": [ + { + "bbox": [ + 315, + 137, + 351, + 155 + ], + "spans": [ + { + "bbox": [ + 315, + 137, + 351, + 155 + ], + "type": "image", + "image_path": "739e8fce916d4383c6a2ea6f02c2bc5bdc3d3eac95311fe46491bd1fac81df48.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 358, + 118, + 419, + 174 + ], + "blocks": [ + { + "bbox": [ + 358, + 118, + 419, + 174 + ], + "lines": [ + { + "bbox": [ + 358, + 118, + 419, + 174 + ], + "spans": [ + { + "bbox": [ + 358, + 118, + 419, + 174 + ], + "type": "image", + "image_path": "32fc663f756aec074d72c0a8bdb2d6005bbf138cf159b89aaade7f663a6c1980.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 432, + 126, + 529, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 432, + 126, + 529, + 135 + ], + "spans": [ + { + "bbox": [ + 432, + 126, + 529, + 135 + ], + "type": "interline_equation", + "content": "v _ {i} = \\operatorname {P r o b} \\left(\" Y s e \"\\right) = 0. 6 5 4 3", + "image_path": "4a1bc991a8b53801b6acbfbc9113905572071d2d592af99dac44403499e1f550.jpg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 85, + 256, + 214, + 302 + ], + "blocks": [ + { + "bbox": [ + 106, + 239, + 196, + 251 + ], + "lines": [ + { + "bbox": [ + 106, + 239, + 196, + 251 + ], + "spans": [ + { + "bbox": [ + 106, + 239, + 196, + 251 + ], + "type": "text", + "content": "Finetuning for " + }, + { + "bbox": [ + 106, + 239, + 196, + 251 + ], + "type": "inline_equation", + "content": "VLM_{2}" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 256, + 214, + 302 + ], + "lines": [ + { + "bbox": [ + 85, + 256, + 214, + 302 + ], + "spans": [ + { + "bbox": [ + 85, + 256, + 214, + 302 + ], + "type": "image", + "image_path": "f12db0ef0faa0c78c4788d3ee3625e39dabe9449c78091cdd7f47a93ad2f472b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 230, + 260, + 251, + 283 + ], + "blocks": [ + { + "bbox": [ + 230, + 260, + 251, + 283 + ], + "lines": [ + { + "bbox": [ + 230, + 260, + 251, + 283 + ], + "spans": [ + { + "bbox": [ + 230, + 260, + 251, + 283 + ], + "type": "image", + "image_path": "018fc78cb72d4dab55ee27a58ef1083eea4295f06f7a7188f2a2abf8db523822.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 262, + 222, + 307, + 315 + ], + "blocks": [ + { + "bbox": [ + 262, + 203, + 389, + 213 + ], + "lines": [ + { + "bbox": [ + 262, + 203, + 389, + 213 + ], + "spans": [ + { + "bbox": [ + 262, + 203, + 389, + 213 + ], + "type": "text", + "content": "Phase2: Precision Localization" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 217, + 539, + 228 + ], + "lines": [ + { + "bbox": [ + 309, + 217, + 539, + 228 + ], + "spans": [ + { + "bbox": [ + 309, + 217, + 539, + 228 + ], + "type": "text", + "content": "Does the blue dot correspond to the location of the {left shoulder}?" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 262, + 222, + 307, + 315 + ], + "lines": [ + { + "bbox": [ + 262, + 222, + 307, + 315 + ], + "spans": [ + { + "bbox": [ + 262, + 222, + 307, + 315 + ], + "type": "image", + "image_path": "142d52ee48fb2d2a33400a033a6da20bb319d3e3278c75c478aff4a4bb6600f3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 555, + 398 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 555, + 398 + ], + "type": "text", + "content": "Figure 2. We introduce the two-stage KPAScore computation for multiple occluded persons. First, bounding boxes obtained from the detection method are used to segment individuals in the image. In Stage 1, " + }, + { + "bbox": [ + 55, + 331, + 555, + 398 + ], + "type": "inline_equation", + "content": "VLM_{1}" + }, + { + "bbox": [ + 55, + 331, + 555, + 398 + ], + "type": "text", + "content": " is queried to determine the presence of keypoints, with the probability of a \"Yes\" response serving as the visibility score. To enhance accuracy, " + }, + { + "bbox": [ + 55, + 331, + 555, + 398 + ], + "type": "inline_equation", + "content": "VLM_{2}" + }, + { + "bbox": [ + 55, + 331, + 555, + 398 + ], + "type": "text", + "content": " is fine-tuned using 10k training samples before being deployed in the second stage. In Stage 2, the predicted keypoint coordinates are visualized as prompts, and " + }, + { + "bbox": [ + 55, + 331, + 555, + 398 + ], + "type": "inline_equation", + "content": "VLM_{2}" + }, + { + "bbox": [ + 55, + 331, + 555, + 398 + ], + "type": "text", + "content": " is queried again to assess accuracy. The probability of a \"Yes\" response serves as the accuracy score. Finally, KPAScore integrates both visibility and accuracy scores to provide a comprehensive evaluation of keypoint detection performance." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 245, + 351, + 264 + ], + "blocks": [ + { + "bbox": [ + 317, + 245, + 351, + 264 + ], + "lines": [ + { + "bbox": [ + 317, + 245, + 351, + 264 + ], + "spans": [ + { + "bbox": [ + 317, + 245, + 351, + 264 + ], + "type": "image", + "image_path": "39520483cf3d47cb190d1507db6c3e6cd94078f2400a81b314ff04aa8f102ea4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 354, + 247, + 414, + 265 + ], + "blocks": [ + { + "bbox": [ + 354, + 247, + 414, + 265 + ], + "lines": [ + { + "bbox": [ + 354, + 247, + 414, + 265 + ], + "spans": [ + { + "bbox": [ + 354, + 247, + 414, + 265 + ], + "type": "image", + "image_path": "a259cc492e8b656d79ae9f88898232413072224e38f9ca7b1945a1078a5b2730.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 425, + 247, + 522, + 256 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 425, + 247, + 522, + 256 + ], + "spans": [ + { + "bbox": [ + 425, + 247, + 522, + 256 + ], + "type": "interline_equation", + "content": "\\boldsymbol {k} _ {i} = \\operatorname {P r o b} \\left(\" Y s e \"\\right) = 0. 8 9 5 5", + "image_path": "6bd84918dafc314f0d97f5295a704c25fe568a89a1d1f39d786682a3d1163326.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 349, + 285, + 489, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 285, + 489, + 308 + ], + "spans": [ + { + "bbox": [ + 349, + 285, + 489, + 308 + ], + "type": "interline_equation", + "content": "K P A S c o r e _ {p} = \\frac {\\sum_ {i = 1} ^ {n} \\sigma \\left(v _ {i}\\right) \\sqrt {k _ {i} v _ {i}}}{\\sum_ {i = 1} ^ {n} \\sigma \\left(v _ {i}\\right)} = 0. 9 0 1 2", + "image_path": "700f49eb14cdc2c113ef4f8330854394d5ffb8cf68c7e9b15358440ac6872578.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 56, + 409, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 296, + 649 + ], + "type": "text", + "content": "predictions through multi-stage processing. HRNet [22] proposed a high-resolution network architecture that maintains spatial detail across the network. LiteHRNet [30] later advanced keypoint detection by maintaining high-resolution features while optimizing for computational efficiency. RSN [3] introduced a method for aggregating intra-level features, preserving rich spatial information. More recent methods, such as RTMPose [12] and AlphaPose [6], adopt a top-down framework, utilizing object detectors to first locate each individual before estimating poses. RTMPose [12] excels in complex environments by isolating instances, while AlphaPose [6] handles occlusions and challenging poses efficiently, making it ideal for real-time scenarios. ViTPose [27], applies vision transformers as non-hierarchical backbones to capture global features. VLM has shown significant progress in keypoint detection, LocLLM [24] achieved comparable performance to ViT-Pose [27] on COCO dataset, highlighting its capability for keypoint understanding, though it has not yet been applied in evaluation tasks. Appendix. A give a supplement to related work." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 657, + 111, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 111, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 111, + 669 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "In this section, we introduce the KPAScore, to the best of our knowledge, the first model-based evaluation metric for keypoint detection. This metric leverages the visual under" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 409, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 468 + ], + "type": "text", + "content": "standing capabilities of large vision-language models, eliminating the reliance on human annotations inherent in traditional groundtruth-based metrics. An overview of the KPA-Score calculation process is illustrated in Fig. 2 and the KPAScore can be formulated as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 359, + 477, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 477, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 359, + 477, + 553, + 506 + ], + "type": "interline_equation", + "content": "\\mathrm {K P A S c o r e} _ {p} = \\frac {\\sum_ {i = 1} ^ {n} \\sigma \\left(v _ {i}\\right) \\sqrt {k _ {i} \\cdot v _ {i}}}{\\sum_ {i = 1} ^ {n} \\sigma \\left(v _ {i}\\right)}, \\tag {2}", + "image_path": "eece74a07b37b0fe80fa4a3fbea4c044e2995e9056f63101dfebf2aca3e4a596.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 510, + 509, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 510, + 509, + 521 + ], + "spans": [ + { + "bbox": [ + 314, + 510, + 509, + 521 + ], + "type": "text", + "content": "where the indicator function " + }, + { + "bbox": [ + 314, + 510, + 509, + 521 + ], + "type": "inline_equation", + "content": "\\sigma(v_{i})" + }, + { + "bbox": [ + 314, + 510, + 509, + 521 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 381, + 530, + 553, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 530, + 553, + 562 + ], + "spans": [ + { + "bbox": [ + 381, + 530, + 553, + 562 + ], + "type": "interline_equation", + "content": "\\sigma \\left(v _ {i}\\right) = \\left\\{ \\begin{array}{l l} 1, & v _ {i} > 0. 5 \\\\ 0, & v _ {i} \\leq 0. 5 \\end{array} . \\right. \\tag {3}", + "image_path": "596c228d79d8bd0e19dc4b529536e6c804003ea5a98297630fffee60e3655396.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "text", + "content": " represents the keypoint visibility score, obtained from the first-stage presence detection. If " + }, + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "inline_equation", + "content": "v_{i} < 0.5" + }, + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "text", + "content": ", the keypoint is considered not existed. For values between 0.5 and 1, " + }, + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "text", + "content": " reflects the confidence of the VLM in the keypoint's presence: lower values indicate that the keypoint is likely occluded, while higher values suggest it is unobstructed. Similarly, " + }, + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 313, + 570, + 555, + 712 + ], + "type": "text", + "content": " represents the keypoint accuracy score, derived from the second-stage precision localization. The denominator ensures the normalization of contributions from all visible keypoints and the square operation avoids that the product of two numbers less than 1 is too small, maintaining a balanced evaluation." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8443" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 221, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 221, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 221, + 85 + ], + "type": "text", + "content": "3.1. First-Stage Presence Detection" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 162 + ], + "type": "text", + "content": "In the first stage, we leverage vision-language models (VLMs) to determine whether specific keypoints are present in the image. Image " + }, + { + "bbox": [ + 55, + 89, + 296, + 162 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 89, + 296, + 162 + ], + "type": "text", + "content": " is the segmented images of individual persons which obtained from object detection. Given the detection task's keypoint list with " + }, + { + "bbox": [ + 55, + 89, + 296, + 162 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 89, + 296, + 162 + ], + "type": "text", + "content": " keypoints " + }, + { + "bbox": [ + 55, + 89, + 296, + 162 + ], + "type": "inline_equation", + "content": "P = [p_{1}, p_{2}, \\ldots, p_{N}]" + }, + { + "bbox": [ + 55, + 89, + 296, + 162 + ], + "type": "text", + "content": ", for instance:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 167, + 295, + 180 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 167, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 81, + 167, + 295, + 180 + ], + "type": "interline_equation", + "content": "P = \\left[ ^ {\\prime \\prime} \\text {N o s e} ^ {\\prime \\prime}, ^ {\\prime \\prime} \\text {L e f t E y} ^ {\\prime \\prime}, \\dots , ^ {\\prime \\prime} \\text {R i g h t W i r s t} ^ {\\prime \\prime} \\right], \\tag {4}", + "image_path": "34b3e8a0a97532462e574529ff61c48213ddcd92b36fdda3a0c7634be8636388.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 185, + 296, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 185, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 296, + 209 + ], + "type": "text", + "content": "we query the VLM for each keypoint " + }, + { + "bbox": [ + 55, + 185, + 296, + 209 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 185, + 296, + 209 + ], + "type": "text", + "content": " in the detection task's keypoint list with the question " + }, + { + "bbox": [ + 55, + 185, + 296, + 209 + ], + "type": "inline_equation", + "content": "q(p_i)" + }, + { + "bbox": [ + 55, + 185, + 296, + 209 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 215, + 133, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 215, + 133, + 228 + ], + "spans": [ + { + "bbox": [ + 80, + 215, + 133, + 228 + ], + "type": "interline_equation", + "content": "p _ {i} = \\text {\" n o s e \"}", + "image_path": "e258a4e0fa972afd4038b02a2ba65492a6ff3f2a94b6a3e712ea41347154b6b4.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 229, + 295, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 229, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 63, + 229, + 295, + 243 + ], + "type": "interline_equation", + "content": "q _ {1} \\left(p _ {i}\\right) = \\text {D o e s t h e i m a g e i n c l u d e t h e l o c a t i o n o f t h e} \\tag {5}", + "image_path": "c66e3f488cbaf51ae6d50aee27eb4a3146d234ac3b6d6aa094adfe4e1e601397.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 245, + 269, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 245, + 269, + 257 + ], + "spans": [ + { + "bbox": [ + 99, + 245, + 269, + 257 + ], + "type": "interline_equation", + "content": "\\text {p e r s o n ' s \" n o s e \"} \\text {? P l a c e a n s w e r y e s o r n o .}", + "image_path": "2c1b91774e528d0998ba60e681d97c4253582da0e6c36a17406a9e0570c07da1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 262, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 286 + ], + "type": "text", + "content": "The VLM returns a probability score for answering \"Yes,\" which is taken as the visibility score " + }, + { + "bbox": [ + 55, + 262, + 296, + 286 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 262, + 296, + 286 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 292, + 295, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 292, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 115, + 292, + 295, + 305 + ], + "type": "interline_equation", + "content": "v _ {i} = \\operatorname {P r o b} \\left(\" \\text {Y e s}\" \\mid I, q _ {1} \\left(k _ {i}\\right)\\right), \\tag {6}", + "image_path": "327d706067ff115d4b4793d1109e1ba3acd65d47bc8de443ff7a4021fb4a5c7c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 310, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 335 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 55, + 310, + 296, + 335 + ], + "type": "inline_equation", + "content": "v_{i} \\leq 0.5" + }, + { + "bbox": [ + 55, + 310, + 296, + 335 + ], + "type": "text", + "content": ", the keypoint is considered not existential and is excluded from subsequent precision evaluation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 340, + 248, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 248, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 248, + 353 + ], + "type": "text", + "content": "3.2. Second-Stage Precision Localization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 357, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 392 + ], + "type": "text", + "content": "In the second stage, we perform a more refined evaluation of the detection method by assessing the accuracy of the estimated keypoint coordinates." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 393, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 296, + 523 + ], + "type": "text", + "content": "VLMs are generally not well-suited for directly evaluating keypoint coordinate accuracy. This limitation likely arises from the inherent difficulty VLMs face in establishing a precise mapping between textual descriptions and specific spatial locations within an image. To overcome this challenge, we incorporate visual prompts [13, 15, 33] to enhance the VLMs' ability to perceive and interpret keypoint locations. Specifically, we overlay a blue dot " + }, + { + "bbox": [ + 55, + 393, + 296, + 523 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 55, + 393, + 296, + 523 + ], + "type": "text", + "content": " at the keypoint coordinates predicted by the detection algorithm, using it as a visual cue to facilitate keypoint accuracy evaluation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 525, + 296, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 549 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 549 + ], + "type": "text", + "content": "Given an image embedded with a visual prompt, denoted as " + }, + { + "bbox": [ + 55, + 525, + 296, + 549 + ], + "type": "inline_equation", + "content": "I_{d_i}" + }, + { + "bbox": [ + 55, + 525, + 296, + 549 + ], + "type": "text", + "content": ", we query the VLM with the following question:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 77, + 555, + 129, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 555, + 129, + 567 + ], + "spans": [ + { + "bbox": [ + 77, + 555, + 129, + 567 + ], + "type": "interline_equation", + "content": "p _ {i} = \\text {\" n o s e \"}", + "image_path": "ccdb0da7bc3e8e773982f383146a26663f7c2e02543237e8182f189abd21041c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 60, + 570, + 295, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 570, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 60, + 570, + 295, + 582 + ], + "type": "interline_equation", + "content": "q _ {2} \\left(p _ {i}\\right) = \\text {D o e s t h e b l u e d o t c o r r e s p o n d t o t h e l o c a t i o n (7)}", + "image_path": "a918bc0b422d958c6c4f9cf86c9c7cb0575656e2848ddc263aeec6dd7f872ecd.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 95, + 585, + 260, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 585, + 260, + 596 + ], + "spans": [ + { + "bbox": [ + 95, + 585, + 260, + 596 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\text {o f t h e \" n o s e\" ? P l a c e a n s w e r Y e s o r N o .} \\end{array}", + "image_path": "ad50a101205629a622ea2b9001fa28217a73c4d367d3fbb57ef568d43a325e11.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 601, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 626 + ], + "type": "text", + "content": "The generative likelihood of the VLM responding \"Yes\" is then computed as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 111, + 638, + 295, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 638, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 111, + 638, + 295, + 651 + ], + "type": "interline_equation", + "content": "k _ {i} = \\operatorname {P r o b} \\left(\" \\text {Y e s}\" \\mid I _ {d _ {i}}, q _ {2} \\left(p _ {i}\\right)\\right), \\tag {8}", + "image_path": "0ce7c8fb77c82da5ed8160e6a13fd4ab8817965ca791640711f4b3e19d956566.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "k_{i}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " serves as the accuracy score for the given keypoint in this second stage. Finally, the model's overall KPAScore is obtained by computing a weighted average of all keypoint scores, integrating the evaluation results from both stages as shown in Eq. (2)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 91, + 553, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 114 + ], + "type": "text", + "content": "This section conducts extensive experiments to answer the following questions:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 115, + 553, + 247 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 314, + 115, + 553, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 115, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 314, + 115, + 553, + 150 + ], + "type": "text", + "content": "- Can VLMs comprehend human keypoints? Additionally, how can we further enhance VLMs' capability to understand keypoints? See Sec. 4.1." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 151, + 553, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 151, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 314, + 151, + 553, + 175 + ], + "type": "text", + "content": "- Can KPAScore consistently reflect changes in predicted keypoints relative to the ground truth? See Sec. 4.2." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 175, + 553, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 175, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 314, + 175, + 553, + 198 + ], + "type": "text", + "content": "- Is there a correlation between KPAScore and existing annotation-dependent metrics? See Sec. 4.3." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 199, + 553, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 199, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 314, + 199, + 553, + 221 + ], + "type": "text", + "content": "- Can KPAScore be considered as a guide for keypoint detection ensemble methods? See Sec. 4.4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 223, + 553, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 223, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 314, + 223, + 553, + 247 + ], + "type": "text", + "content": "- Does the KPAScore align with human perception in complex scenarios without ground truth? See Sec. 4.5." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 253, + 510, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 510, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 510, + 266 + ], + "type": "text", + "content": "4.1. Keypoint Understanding Benchmark" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 270, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 413 + ], + "type": "text", + "content": "With the widespread application of VLMs in computer vision tasks, their ability to accurately understand human keypoints remains an open question. We evaluate the keypoint understanding capability of VLMs from two perspectives: presence detection and precision localization. To facilitate this evaluation, we introduce the Keypoint Understanding Benchmark (KPUBench), a large-scale dataset comprising 1.8 million structured questions derived from COCO [17], MPII [1], and Human3.6M [11]. The dataset is systematically split into training and test sets with an approximate ratio of 9.5:0.5, and all evaluations are conducted exclusively on the test set." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 415, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 533 + ], + "type": "text", + "content": "The relationship between KPUBench and KPAScore can be summarized as follows: KPUBench serves as a benchmark for assessing the ability of existing VLMs to comprehend human keypoints in an unsupervised or zero-shot setting, measuring their fine-grained semantic and visual perception capabilities. The best-performing VLM from this evaluation is then employed in downstream tasks, including the construction of KPAScore. The architecture of KPUBench is illustrated in Figure 3 and the evaluation results of 18 visual-language models is presented in Tab. 1." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "type": "text", + "content": "(a) Presence Detection assesses the VLM's ability to identify whether a specific keypoint is present in an image. We query the VLM: \"Does the image include the location of the person's {keypoint name}?\". Ideally, for keypoints annotated as present (COCO visibility scores of 1 or 2), the VLM should return \"Yes\" with a probability above 0.5, whereas for absent keypoints (visibility score of 0), this probability should be below 0.5. The positive-to-negative sample ratio is approximately 5.31:4.69. To evaluate VLM reliability and confidence in keypoint presence detection, we compute accuracy " + }, + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "type": "inline_equation", + "content": "(ACC)" + }, + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "type": "text", + "content": " and probabilistic expectation " + }, + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "type": "inline_equation", + "content": "(E(Prob))" + }, + { + "bbox": [ + 313, + 534, + 555, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "As shown in the third and fourth columns of Tab. 1, most VLMs demonstrate a fundamental ability to recognize keypoints, though model performance does not clearly correlate" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8444" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 71, + 522, + 279 + ], + "blocks": [ + { + "bbox": [ + 91, + 71, + 522, + 279 + ], + "lines": [ + { + "bbox": [ + 91, + 71, + 522, + 279 + ], + "spans": [ + { + "bbox": [ + 91, + 71, + 522, + 279 + ], + "type": "image", + "image_path": "b7208c297a8c898e031abad7c49fbada1727edfcc150df2975c42c5e58acd06f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 286, + 468, + 298 + ], + "lines": [ + { + "bbox": [ + 187, + 286, + 468, + 298 + ], + "spans": [ + { + "bbox": [ + 187, + 286, + 468, + 298 + ], + "type": "text", + "content": "(c) Precision Localization Detection—Visual Prompt Understanding Capability" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 313, + 555, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 313, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 555, + 360 + ], + "type": "text", + "content": "Figure 3. Overview of the KPUBench evaluation framework. (a): Evaluates whether a VLM can determine the existence of a keypoint in an image. (b): Assesses the VLM's ability to recognize keypoints based on explicit numerical coordinates. (c): Measures whether the VLM can accurately associate a visual cue (a blue dot) with the corresponding keypoint. The expected answer is \"Yes\" for correct localizations and \"No\" for incorrect ones." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 368, + 544, + 651 + ], + "blocks": [ + { + "bbox": [ + 70, + 368, + 544, + 651 + ], + "lines": [ + { + "bbox": [ + 70, + 368, + 544, + 651 + ], + "spans": [ + { + "bbox": [ + 70, + 368, + 544, + 651 + ], + "type": "table", + "html": "
ModelSizePresence DetectionPrecision Localization Detection
ACC(%)E(Prob)CoordinateVisual Prompt
ACC(%)E(Prob)ACC(%)E(Prob)
LLaVA-OneVersion [14]0.5B55.5849.7150.0050.2452.3250.30
DeepSeek-Coder-Base [8]1.3B62.0352.7749.9149.9655.5750.86
InternVL2-2B [4]2B56.3047.3350.0950.0655.2551.75
Qwen2-VL-2B [25]2B73.6758.7050.0050.0456.5451.21
MiniCPM-V 2.0 [29]2.8B58.4344.4249.9449.9454.8652.25
LLaVA-V1.5-7B [19]7B61.0057.1849.8649.7053.3250.72
LLaVA-V1.6-7B [19]7B75.1651.2149.7349.9254.8451.21
Qwen2-VL-7B [25]7B66.9857.6150.2250.1160.5552.71
InternVL2-8B [4]8B56.3047.3350.9550.0655.2551.76
MiniCPM-V-2.6 [29]8B54.8451.2150.8550.1153.0450.97
GLM-4V-9B [7]9B76.9949.8749.9849.7662.5758.17
LLama3.2-11B-coco [23]11B61.2055.0848.7549.0156.0254.53
LLaVA-V1.5-13B [19]13B69.8758.8250.0049.9153.5050.61
LLaVA-V1.6-13B [19]13B71.5255.9349.9949.0458.9651.32
Cogglm2-LLama3-chat-19B [10]19B70.3354.0649.9849.9854.7550.86
InternVL2-26B [4]26B61.8553.1750.0350.1957.2353.89
LLaVA-V1.6-34B [19]34B84.6765.1249.9549.9050.0050.35
Qwen2-VL-72B [25]72B79.5463.6050.8450.1556.0752.96
GLM-4V-9B-finetuig9B----80.1574.32
LLaVA + GLM-finetuig43B84.6765.12--84.5378.10
", + "image_path": "2c3e909f772d67968d3b745f341a77e48610512f0b4d0f87432e81c6bf4c5f64.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 658, + 555, + 693 + ], + "lines": [ + { + "bbox": [ + 55, + 658, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 555, + 693 + ], + "type": "text", + "content": "Table 1. Performance of 18 VLM models on KPUBench. KPUBench evaluates VLMs' understanding of keypoint detection from two perspectives: presence detection and precision localization detection, providing a comprehensive measurement using accuracy and probabilistic expectation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8445" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "with model size. Both the LLaVA-V1.6 and Qwen families perform well, exceeding " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "66.98\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " accuracy, likely due to task-related training data or effective architectures. Among them, LLaVA-V1.6-34B [19] achieves the highest accuracy " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "(84.67\\%)" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " and probabilistic expectation " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "(65.12\\%)" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": ", indicating strong keypoint perception. In contrast, GLM-4V [7] attains " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "76.99\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " accuracy but a lower probabilistic expectation " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "(49.87\\%)" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": ", suggesting lower confidence despite correct predictions, which may affect reliability in practical use." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 294, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 294, + 264 + ], + "type": "text", + "content": "Precision Localization Detection consists of two components: coordinate understanding and visual prompt understanding. We use human-annotated keypoint coordinates as ground truth, treating them as positive samples. Negative samples are generated by shifting each keypoint randomly by " + }, + { + "bbox": [ + 55, + 180, + 294, + 264 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 180, + 294, + 264 + ], + "type": "text", + "content": " of the bounding box size, maintaining a 1:1 positive-to-negative ratio." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 267, + 296, + 422 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 267, + 295, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 295, + 339 + ], + "type": "text", + "content": "- Coordinate Understanding We ask the VLM: \"Does the coordinate " + }, + { + "bbox": [ + 55, + 267, + 295, + 339 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 55, + 267, + 295, + 339 + ], + "type": "text", + "content": " correspond to the location of the {keypoint name}?\" Accuracy and probabilistic expectation are computed, where higher accuracy indicates better alignment with human judgment, and higher probabilistic expectation reflects the model's confidence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 339, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 422 + ], + "type": "text", + "content": "- Visual Prompt Understanding Using the same positive and negative samples, we mark a blue dot at the keypoint location and ask: \"Does the blue dot correspond to the location of the {keypoint name}?\" Accuracy and probabilistic expectation are measured similarly. The effect of the color of the visual prompt is illustrated in Appendix. B." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "text", + "content": "Based on the fifth and sixth columns of Tab. 1, VLMs show limited ability in interpreting numerical keypoint coordinates, with accuracy around " + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "text", + "content": ", near random guessing. This suggests difficulty in mapping numerical coordinates to spatial locations in an image. Introducing visual prompts significantly improves keypoint comprehension, increasing accuracy by an average of " + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "inline_equation", + "content": "5.42\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "text", + "content": ". Notably, GLM-4V's accuracy rose from " + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "inline_equation", + "content": "49.98\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "inline_equation", + "content": "62.57\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "text", + "content": ", with its probabilistic expectation improving from " + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "inline_equation", + "content": "49.76\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "inline_equation", + "content": "58.17\\%" + }, + { + "bbox": [ + 55, + 426, + 295, + 568 + ], + "type": "text", + "content": ", confirming that visual prompts help direct model attention to spatial locations. However, despite these gains, performance remains insufficient for practical applications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 569, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 677 + ], + "type": "text", + "content": "To further enhance the model's precision, we fine-tuned GLM-4V-9B using the LoRA method with 10k Precision Localization Detection training samples containing visual prompts (detailed hyperparameter settings are provided in Appendix. C). The fine-tuned model, GLM-4V-9B-finetuned, achieved a remarkable accuracy of " + }, + { + "bbox": [ + 55, + 569, + 295, + 677 + ], + "type": "inline_equation", + "content": "80.15\\%" + }, + { + "bbox": [ + 55, + 569, + 295, + 677 + ], + "type": "text", + "content": " and a probabilistic expectation of " + }, + { + "bbox": [ + 55, + 569, + 295, + 677 + ], + "type": "inline_equation", + "content": "78.10\\%" + }, + { + "bbox": [ + 55, + 569, + 295, + 677 + ], + "type": "text", + "content": ", demonstrating that even a simple fine-tuning process can significantly enhance VLM performance on specialized tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Finally, we implemented a two-stage approach by employing LLaVA-V1.6-34B as the Presence Detection module in the first stage. We retained only the keypoints identi" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "fied as present by LLaVA-V1.6-34B and fed them into the Precision Localization Detection stage. This LLaVA-V1.6-34B + GLM-4V-9B-finetuned two-stage framework further enhanced the model's keypoint localization understanding, achieving an accuracy of " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "84.53\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " and a probabilistic expectation of " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "78.10\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ". Therefore, we propose a two-stage evaluation framework that integrates presence detection and precision localization to systematically assess keypoint detection models using VLMs." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 332, + 217, + 537, + 317 + ], + "blocks": [ + { + "bbox": [ + 314, + 189, + 474, + 201 + ], + "lines": [ + { + "bbox": [ + 314, + 189, + 474, + 201 + ], + "spans": [ + { + "bbox": [ + 314, + 189, + 474, + 201 + ], + "type": "text", + "content": "4.2. The Consistency of KPAscore" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 217, + 537, + 317 + ], + "lines": [ + { + "bbox": [ + 332, + 217, + 537, + 317 + ], + "spans": [ + { + "bbox": [ + 332, + 217, + 537, + 317 + ], + "type": "table", + "html": "
ModelKPAScoreOKSmAP
ViTPose-H* [27]60.9872.9579.20
RTMPose-1 [12]59.0572.4777.30
RSN [3]59.0472.1075.00
HRNet [22]59.2772.1474.91
Alphapose [6]58.1371.4670.61
CPM [26]57.2470.1465.23
LiteHRNet [30]57.0870.3064.24
", + "image_path": "a2d8729ede95ec94215b447e6f726fafc9685393bd43082665c9325c66d8223d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 324, + 553, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 324, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 553, + 346 + ], + "type": "text", + "content": "Table 2. The overall scores of seven keypoint detection algorithms on the COCO dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 353, + 553, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 553, + 425 + ], + "type": "text", + "content": "After establishing the LLaVA-V1.6-34B + GLM-4V-9B-finetuned two-stage KPAScore framework, we conducted a comprehensive evaluation of seven mainstream keypoint detection methods on the COCO dataset. The selected models include ViTPose [27], RTMPose [12], Alphapose [6], HRNet [22], RSN [3], LiteHRNet [30], and CPM [26]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 425, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 553, + 472 + ], + "type": "text", + "content": "Consistency in response to changes is a critical factor in evaluating the usability of a metric. Here, we demonstrate that KPAScore possesses both overall consistency and change consistency from two perspectives." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 473, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 605 + ], + "type": "text", + "content": "First, we present the average KPAScore, OKS, and mAP of the seven methods on COCO (see Table 2). As shown, KPAScore exhibits strong alignment with traditional keypoint metrics (OKS and " + }, + { + "bbox": [ + 313, + 473, + 554, + 605 + ], + "type": "inline_equation", + "content": "mAP" + }, + { + "bbox": [ + 313, + 473, + 554, + 605 + ], + "type": "text", + "content": "). Models with higher OKS and mAP scores, such as ViTPose-H and RTMPose-1, also achieve the highest KPAScore, while lower-performing models, like CPM and LiteHRNet, similarly rank lower under KPAScore. This consistency validates KPAScore as an effective metric that aligns with established evaluation criteria while leveraging VLM-based assessment to provide a more holistic keypoint detection evaluation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 700 + ], + "type": "text", + "content": "Second, we apply varying intensities of offset to the keypoint detection results for each detection algorithm, then recalculated the KPAScore. We analyzed the KPAScore across different offset intensities, where the offset is defined as the square root of the bounding box area (a standard approach on the COCO dataset) multiplied by a factor ranging from " + }, + { + "bbox": [ + 313, + 605, + 553, + 700 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 313, + 605, + 553, + 700 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 605, + 553, + 700 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 605, + 553, + 700 + ], + "type": "text", + "content": ", with random offset directions which shown in Fig. 4a" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "As shown in Fig. 4b, the KPAScore for all 7 keypoint" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8446" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 156, + 76, + 289, + 198 + ], + "blocks": [ + { + "bbox": [ + 113, + 94, + 147, + 177 + ], + "lines": [ + { + "bbox": [ + 113, + 94, + 147, + 177 + ], + "spans": [ + { + "bbox": [ + 113, + 94, + 147, + 177 + ], + "type": "text", + "content": "- 3% offset\n- 5% offset\n- 10% offset\n- 15% offset\n- 20% offset\n- 25% offset\n- 30% offset\n- 50% offset" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 156, + 76, + 289, + 198 + ], + "lines": [ + { + "bbox": [ + 156, + 76, + 289, + 198 + ], + "spans": [ + { + "bbox": [ + 156, + 76, + 289, + 198 + ], + "type": "image", + "image_path": "5f4818e4fa8cdc639b531a6d7a7b2d8e706fe4e1b5fc9dc6b0f7c93bda8878e6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 201, + 231, + 211 + ], + "lines": [ + { + "bbox": [ + 166, + 201, + 231, + 211 + ], + "spans": [ + { + "bbox": [ + 166, + 201, + 231, + 211 + ], + "type": "text", + "content": "(a) Offset Example." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 293, + 72, + 503, + 198 + ], + "blocks": [ + { + "bbox": [ + 293, + 72, + 503, + 198 + ], + "lines": [ + { + "bbox": [ + 293, + 72, + 503, + 198 + ], + "spans": [ + { + "bbox": [ + 293, + 72, + 503, + 198 + ], + "type": "image", + "image_path": "1a39f267bc48dc7382fd9edc78c097dd580ff37cb70032b308f0a25639a92e17.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 201, + 457, + 211 + ], + "lines": [ + { + "bbox": [ + 337, + 201, + 457, + 211 + ], + "spans": [ + { + "bbox": [ + 337, + 201, + 457, + 211 + ], + "type": "text", + "content": "(b) Variation of KPAScore and OKS." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 243 + ], + "type": "text", + "content": "Figure 4. (a): Circles of different colors represent the corresponding offset keypoint positions at various offset intensities. (b): A line chart illustrating the variation in values of KPAScore (red line) and OKS (blue line) at different offset intensities." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 57, + 255, + 555, + 292 + ], + "blocks": [ + { + "bbox": [ + 57, + 255, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 57, + 255, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 57, + 255, + 555, + 292 + ], + "type": "table", + "html": "
MethodViTPose-H* [27]RSN [3]HRNet [22]RTMPose-1 [12]Alphapose [6]CPM [26]LiteHRNet [30]Average
Corr0.720.710.720.710.700.690.730.71
", + "image_path": "bcfbedcc9a88cb04043fd9b6ad43947fcc05b398b3067545339a1acbdf37cd35.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 161, + 300, + 448, + 310 + ], + "lines": [ + { + "bbox": [ + 161, + 300, + 448, + 310 + ], + "spans": [ + { + "bbox": [ + 161, + 300, + 448, + 310 + ], + "type": "text", + "content": "Table 3. The Pearson correlation coefficients between the KPAScore and OKS." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "spans": [ + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "text", + "content": "detection methods steadily decreases with increasing offset intensity, indicating that the KPAScore effectively responds to variations in keypoint positions and can serve as a reliable new evaluation metric. Specifically, offset intensities of " + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "text", + "content": " result in a decrease in the KPAScore, but the decline is not significant. As illustrated in Fig. 4a, offsets within " + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "text", + "content": " are not easily perceptible to the human eye, and therefore the minor drop in score is consistent with human perception. When the offset intensity ranges from " + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "text", + "content": ", the KPAScore decreases more markedly, and at a " + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "text", + "content": " offset intensity, the scores for all algorithms converge, KPAScore is almost 0. OKS also decreases steadily with the offset, but when the offset strength is " + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 322, + 297, + 501 + ], + "type": "text", + "content": ", OKS is still above 50, which shows that OKS is not very sensitive to the feedback of the offset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 514, + 294, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 294, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 294, + 526 + ], + "type": "text", + "content": "4.3. The Correlation Between KPAScore and OKS" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 628 + ], + "type": "text", + "content": "Beyond evaluating the consistency of KPAScore, we examine its relationship with OKS by computing their Pearson correlation across various keypoint detection methods, as shown in Tab. 3. The results show a strong positive correlation, ranging from 0.69 to 0.73 (average 0.71), indicating that KPAScore aligns well with OKS in assessing keypoint detection quality, further validating its reliability as an evaluation metric." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "However, it is worth noting that the correlation is not perfect " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "(\\mathrm{Corr} < 1.0)" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", indicating that KPAScore captures aspects of keypoint detection performance that may not be fully reflected by OKS. Unlike OKS, which primarily relies on Euclidean distance-based similarity, KPAScore incorporates a human-centric evaluation approach using visual-language models (VLMs). This difference allows KPAS-" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 323, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 357 + ], + "type": "text", + "content": "core to assess keypoint detection from a perceptual perspective, potentially capturing nuanced failure cases that traditional distance-based metrics overlook." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 359, + 556, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 556, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 556, + 539 + ], + "type": "text", + "content": "To further investigate these differences, we analyze cases where OKS and KPAScore exhibit discrepancies (see Appendix. D). We find that OKS tends to penalize predictions that deviate slightly from ground truth, even if they are perceptually acceptable, while KPAScore—by leveraging VLM-based reasoning—can tolerate minor deviations if the predicted keypoints remain semantically correct. Conversely, KPAScore may assign lower scores to cases where OKS is relatively high, particularly when the predicted keypoint location appears anatomically implausible despite being close to the ground truth. These findings highlight KPAScore's potential to complement OKS by incorporating a perceptual validation mechanism, making it a promising metric for real-world pose estimation tasks where human perception plays a crucial role." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 550, + 555, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 555, + 574 + ], + "type": "text", + "content": "4.4. KPAScore as a Guide for Keypoint Detection Ensembles" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 581, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 663 + ], + "type": "text", + "content": "In this section, we explore whether KPAScore can serve as a guiding metric for keypoint detection ensemble methods. Since KPAScore is annotation-free and evaluates keypoints in a perceptually meaningful manner, we propose an ensemble strategy that selects keypoints with the highest KPAScore from each algorithm, forming a new method we refer to as KPAPose." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "To further assess the effectiveness of KPAScore in guiding keypoint detection improvements, we incorporate it as an additional supervisory signal for ViTPose-" + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^*" + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": " and RTMPose-1 on the COCO dataset, selecting keypoints that" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8447" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "achieve the highest estimated KPAScore across different locations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 98, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 98, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 98, + 296, + 204 + ], + "type": "text", + "content": "As shown in Table 4, KPAPose achieves the highest scores in both KPAScore (62.65) and OKS (73.63), surpassing ViTPose-H* [27] and RTMPose-1 [12]. This demonstrates that KPAScore can serve as an effective fusion strategy for integrating multiple detection methods, producing results that not only align closely with traditional evaluation metrics (e.g., OKS) but also better capture human perceptual judgments—an advantage that annotation-dependent metrics fail to achieve." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 99, + 216, + 252, + 281 + ], + "blocks": [ + { + "bbox": [ + 99, + 216, + 252, + 281 + ], + "lines": [ + { + "bbox": [ + 99, + 216, + 252, + 281 + ], + "spans": [ + { + "bbox": [ + 99, + 216, + 252, + 281 + ], + "type": "table", + "html": "
ModelKPAScoreOKS
ViTPose-H* [27]60.9872.95
RTMPose-1 [12]59.0572.47
KPAPose62.6573.63
", + "image_path": "e1006a1ab6127dc01fdde5cb18d34aa16ba8901d1212e484c3fc40549c10341b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 290, + 285, + 302 + ], + "lines": [ + { + "bbox": [ + 65, + 290, + 285, + 302 + ], + "spans": [ + { + "bbox": [ + 65, + 290, + 285, + 302 + ], + "type": "text", + "content": "Table 4. KPAPose performance that integrates both methods" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 321, + 242, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 242, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 242, + 335 + ], + "type": "text", + "content": "4.5. Alignment with Human Perception" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 340, + 296, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 435 + ], + "type": "text", + "content": "To further verify the alignment between KPAScore's perception verification mechanism and human judgment, we constructed a test set featuring complex scenes and challenging actions in both indoor and outdoor environments. These include activities such as yoga, boxing, modern dance, and gymnastics, with participants wearing diverse clothing, making manual annotation of ground truth extremely difficult." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 521 + ], + "type": "text", + "content": "To evaluate the performance of various keypoint detection algorithms in these challenging settings, we employed KPAScore alongside human ratings. A group of 20 researchers independently rated the inference results from different algorithms on a scale of 0 to 10, averaging the scores for each algorithm per image. The details of the human rating strategy are provided in Appendix E." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": "Figure 5 presents the evaluation results of three algorithms on three actions, along with KPAPose, an ensemble of these models, arranged in decreasing difficulty. The results show that in the absence of ground-truth annotations, KPAScore closely aligns with human evaluations. RTM [12] performed the worst in both KPAScore and human ratings, while KPAPose consistently achieved the highest scores, demonstrating KPAScore's robustness in challenging scenarios. For harder-to-estimate actions, KPAPose significantly improved KPAScore from 63.09 to 89.03 and human ratings from 6.55 to 7.65. In contrast, simpler actions saw marginal improvements, with KPAScore increasing by only 0.36, making the difference barely noticeable to humans. Moving forward, we aim to further explore KPAScore's role in evaluating complex scenarios and its potential to enhance keypoint detection algorithms." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 548, + 335 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 548, + 335 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 548, + 335 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 548, + 335 + ], + "type": "image", + "image_path": "212337c1d4e4b5a9c9ec8d189741cdfb66d8a8487fb1922095e3653efa3d5bce.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 345, + 554, + 390 + ], + "lines": [ + { + "bbox": [ + 313, + 345, + 554, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 554, + 390 + ], + "type": "text", + "content": "Figure 5. Alignment of KPA-Score and Human Rating in Complex Scenes with Complex Actions. Here, we select ViTPose, CPM, RTM, and the KPA-Pose generated based on these three algorithms, with additional visualizations provided in Appendix. F" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 427, + 388, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 427, + 388, + 440 + ], + "spans": [ + { + "bbox": [ + 314, + 427, + 388, + 440 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 704 + ], + "type": "text", + "content": "In this paper, we introduce the KPAScore, a method that explores the keypoint understanding capabilities of vision-language models (VLMs) and proposes, to the best of our knowledge, the first human keypoint evaluation metric that eliminates the reliance on ground truth through a two-stage visual prompt question-answering approach. To validate the rationale behind this method, we introduce the Keypoint Understanding Benchmark (KPubBench), which contains over 1.8 million questions. The results demonstrate that our two-stage questioning and visual prompt approach significantly enhance the keypoint comprehension capabilities of VLMs. Furthermore, our analysis demonstrates that KPAScore exhibits variation consistency, maintains a strong correlation with traditional evaluation metrics such as OKS, and aligns closely with human perception—all while being entirely annotation-free. By leveraging VLMs for keypoint assessment, KPAScore provides a scalable, automated, and human-aligned alternative to existing evaluation methods. We hope this work inspires further research on utilizing VLMs for structured vision tasks and fosters the development of more robust, perception-driven keypoint evaluation metrics in computer vision applications." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8448" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. In Proceedings of the IEEE Conference on computer Vision and Pattern Recognition, pages 3686-3693, 2014. 1, 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 201 + ], + "type": "text", + "content": "[2] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 203, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 296, + 280 + ], + "type": "text", + "content": "[3] Yuanhao Cai, Zhicheng Wang, Zhengxiong Luo, Binyi Yin, Angang Du, Haoqian Wang, Xiangyu Zhang, Xinyu Zhou, Erjin Zhou, and Jian Sun. Learning delicate local representations for multi-person pose estimation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 455-472. Springer, 2020. 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 281, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 281, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 296, + 346 + ], + "type": "text", + "content": "[4] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, Bin Li, Ping Luo, Tong Lu, Yu Qiao, and Jifeng Dai. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. arXiv preprint arXiv:2312.14238, 2023.5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 347, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 296, + 403 + ], + "type": "text", + "content": "[5] Michael Dorkenwald, Nimrod Barazani, Cees GM Snoek, and Yuki M Asano. Pin: Positional insert unlocks object localisation abilities in vlms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13548-13558, 2024. 1, 11" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 403, + 296, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 458 + ], + "type": "text", + "content": "[6] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 459, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 296, + 612 + ], + "type": "text", + "content": "[7] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. Chatglm: A family of large language models from glm-130b to glm-4 all tools, 2024. 5, 6, 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 613, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 613, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 613, + 296, + 668 + ], + "type": "text", + "content": "[8] Daya Guo, Qihao Zhu, Dejian Yang, Zhenda Xie, Kai Dong, Wentao Zhang, Guanting Chen, Xiao Bi, Y. Wu, Y.K. Li, Fuli Luo, Yingfei Xiong, and Wenfeng Liang. Deepseek-coder: When the large language model meets programming – the rise of code intelligence, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "text", + "content": "[9] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 1" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[10] Wenyi Hong, Weihan Wang, Ming Ding, Wenmeng Yu, Qingsong Lv, Yan Wang, Yean Cheng, Shiyu Huang, Junhui Ji, Zhao Xue, et al. Cogvlm2: Visual language models for image and video understanding. arXiv preprint arXiv:2408.16500, 2024. 1, 5, 11" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 129, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 184 + ], + "type": "text", + "content": "[11] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, 2014. 1, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 185, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 554, + 228 + ], + "type": "text", + "content": "[12] Tao Jiang, Peng Lu, Li Zhang, Ningsheng Ma, Rui Han, Chengqi Lyu, Yining Li, and Kai Chen. Rtmpose: Real-time multi-person pose estimation based on mmpose, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 230, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 554, + 285 + ], + "type": "text", + "content": "[13] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "type": "text", + "content": "[14] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer, 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 331, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 376 + ], + "type": "text", + "content": "[15] Feng Li, Hao Zhang, Peize Sun, Xueyan Zou, Shilong Liu, Jianwei Yang, Chunyuan Li, Lei Zhang, and Jianfeng Gao. Semantic-sam: Segment and recognize anything at any granularity. arXiv preprint arXiv:2307.04767, 2023. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 376, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 431 + ], + "type": "text", + "content": "[16] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 433, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 433, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 433, + 554, + 498 + ], + "type": "text", + "content": "[17] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 1, 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 499, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 499, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 499, + 554, + 555 + ], + "type": "text", + "content": "[18] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2025. 1, 11" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 556, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 554, + 578 + ], + "type": "text", + "content": "[19] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1, 5, 6, 11" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 579, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 554, + 612 + ], + "type": "text", + "content": "[20] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 11" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 613, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 554, + 679 + ], + "type": "text", + "content": "[21] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 680, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 714 + ], + "type": "text", + "content": "[22] Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE conference on computer" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8449" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 654 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "vision and pattern recognition, pages 5693-5703, 2019. 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[23] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 5, 11" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 206 + ], + "type": "text", + "content": "[24] Dongkai Wang, Shiyu Xuan, and Shiliang Zhang. Locllm: Exploiting generalizable human keypoint localization via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 284 + ], + "type": "text", + "content": "[25] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 5, 11" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 329 + ], + "type": "text", + "content": "[26] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 373 + ], + "type": "text", + "content": "[27] Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In Advances in Neural Information Processing Systems, 2022. 3, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 418 + ], + "type": "text", + "content": "[28] Yi Yang and Deva Ramanan. Articulated human detection with flexible mixtures of parts. IEEE transactions on pattern analysis and machine intelligence, 35(12):2878-2890, 2012. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 464 + ], + "type": "text", + "content": "[29] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 465, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 294, + 497 + ], + "type": "text", + "content": "[30] Changqian Yu, Bin Xiao, Changxin Gao, Lu Yuan, Lei Zhang, Nong Sang, and Jingdong Wang. Lite-hrnet: A lightweight high-resolution network. In CVPR, 2021. 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "type": "text", + "content": "[31] Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 11" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 294, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 608 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 608 + ], + "type": "text", + "content": "[32] Hongyi Zhu, Jia-Hong Huang, Stevan Rudinac, and Evangelos Kanoulas. Enhancing interactive image retrieval with query rewriting using large language models and vision language models. In Proceedings of the 2024 International Conference on Multimedia Retrieval, pages 978-987, 2024. 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 610, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 294, + 654 + ], + "type": "text", + "content": "[33] Xueyan Zou, Jianwei Yang, Hao Zhang, Feng Li, Linjie Li, Jianfeng Wang, Lijuan Wang, Jianfeng Gao, and Yong Jae Lee. Segment everything everywhere all at once. Advances in Neural Information Processing Systems, 36, 2024. 4" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8450" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/58e52446-3fb4-48af-8a99-0aa5fc34cde3_content_list.json b/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/58e52446-3fb4-48af-8a99-0aa5fc34cde3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4c787974212e9a87c25b2320650e90cc9a312e79 --- /dev/null +++ b/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/58e52446-3fb4-48af-8a99-0aa5fc34cde3_content_list.json @@ -0,0 +1,1774 @@ +[ + { + "type": "text", + "text": "Towards Comprehensive Lecture Slides Understanding: Large-scale Dataset and Effective Method", + "text_level": 1, + "bbox": [ + 94, + 128, + 903, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Enming Zhang $^{1}$ , Yuzhe Li $^{1}$ , Yuliang Liu $^{1}$ , Yingying Zhu $^{1*}$ , Xiang Bai $^{1}$ $^{1}$ Huazhong University of Science and Technology \n{emzhang,yzli,ylliu,yyzhu,xbai} $@$ hust.edu.cn", + "bbox": [ + 218, + 202, + 776, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Online education has been widespread in worldwide universities and educational institutions. Lecture slides, a fundamental component of online education, contain a wealth of information, playing a crucial role in learning. However, previous works have not yet paid sufficient attention to understanding lecture slides, including the absence of the large-scale dataset and comprehensive understanding tasks. To facilitate the research about lecture slides understanding, we establish the LecSlides-370K, which consists of 25,542 lectures with 370,078 slides across 15 areas. We also introduce two comprehensive tasks, Lecture Summary and Lecture Question Answering (QA), for providing different perspectives of slides understanding. Furthermore, complex and flexible text relations can hinder the understanding of the internal logic of slides. To address this challenge, we propose a novel method, named SlideParser, which includes an auxiliary branch to predict text relations within slides and enhance attention between related texts, thereby improving slides understanding. With extensive experiments, we show the superiority of our proposed method on both LecSlides-370k and SlideVQA. Dataset and Codes are available at https://github.com/zamling/LecSlides_370K", + "bbox": [ + 89, + 321, + 485, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 699, + 220, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid progress in online education, an increasing number of universities and educational institutions have adopted online learning in many platforms [20, 21, 25]. Among diverse online education formats, such as lecture videos and live streams, slides often serve as a key medium for conveying significant knowledge. Consequently, more researchers have turned their attention to this field. However, there are two obstacles to understand lecture slides: 1) Texts in slides are often conclusive and headline-style statements, which complicate the task of providing detailed", + "bbox": [ + 89, + 724, + 483, + 877 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "content. 2) The complex text relations pose additional challenges for models attempting to grasp the internal logic and generate reasonable responses.", + "bbox": [ + 511, + 292, + 906, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, several works [4, 5, 7, 14, 17, 27] have noticed the importance of lecture slides and introduced corresponding datasets, such as text detection in slides [5], slide segmentation [7] and image-text retrieval for slides and corresponding speech [14]. However, none of them has yet paid attention on comprehensive understanding slides. SlideVQA [27], a recent slides dataset, provides a Question-Answering (QA) benchmark for slides, yet it still overlooks the global understanding task, such as generating a detailed summary for a complete set of slides. Additionally, the restricted scales and limited number of lecture areas also constrain the development of lecture slides understanding.", + "bbox": [ + 511, + 339, + 908, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aiming to address the limitations in the lecture slides dataset, we establish a large-scale and multi-task dataset for comprehensive slides understanding, named LecSlides-370K. Specifically, we collect 25,542 lectures composed of 370,078 slides. These lectures span 15 lecture areas, including business, natural sciences, computer sciences, and so on, significantly surpassing previous datasets in both scale and diversity, as shown in Fig. 1 (a). Furthermore, we introduce two tasks based on our dataset, termed Lecture Summary and Lecture Question Answering (QA). These two tasks represent the most frequently utilized functions for students, offering varied perspectives for slides understanding. Specifically, Lecture Summary focuses on extracting key information and fostering a global understanding, while Lecture QA focuses on locating problems and understanding fine-grained content. Finally, our dataset includes 114,472 summary annotations and 259,461 QA pairs.", + "bbox": [ + 511, + 521, + 908, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Based on LecSlides-370K dataset, we evaluate the existing NLP methods [15, 22], document understanding methods [10, 28, 32, 33] and vision-language models [16, 34]. The results reveal the inferior performance of these methods. We argue that this inferiority can be attributed to the fact that these methods have not yet adequately considered the unique characteristic of lecture slides, that text relations in slides are complex and contain a wealth of logic and rel", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 109, + 886, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4455", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dce8870acaf00311263f9d70cd056d989c5a01c24e1226f0bc702777ffd06985.jpg", + "image_caption": [ + "(a) Data scales of LecSlides-370K and related datasets" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 377, + 276 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2578ba197cded1e26310be408419f364f44f6ba4867fb178fdbd4b55ed45c5b8.jpg", + "image_caption": [ + "(b) An example of LecSlides-370K with various tasks", + "Figure 1. LecSlides-370K is a large-scale and multi-task lecture slides dataset. It contains 25,542 lectures and 370,078 slides across 15 lecture areas, which has around 9.7 times lecture numbers and 7.1 times slide numbers compared with SlideVQA [27]. (a) A comparison between LecSlides-370K and other slides dataset in terms of the number of lectures, slides and lecture areas. The detailed information is shown in Table 1. (b) An example of LecSlides-370K, including Lecture Summary and Lecture QA tasks." + ], + "image_footnote": [], + "bbox": [ + 379, + 88, + 903, + 276 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "evance. For instance, as shown in Fig. 1 (b), texts displayed in the upper left corner indicate detailed requirements for forecasting, while texts on the right side of each block explain each stage in the flow chart. Therefore, understanding complex text relations in slides can grasp the internal logic, and is the key part of understanding slides.", + "bbox": [ + 88, + 383, + 480, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To extract the complex text relations and enhance slides understanding, we propose a novel method called SlideParser. The key of SlideParser is how to effectively predict and utilize complex text relations in slides. Specifically, we propose an auxiliary branch, termed Text Relation Prediction Module, comprising the Relation Prediction Head (RPH) to predict relations among texts and the Relation Attention Layer (RAL) to leverage predicted text relations, enhancing attention between related texts while reducing it between unrelated texts. Additionally, to enhance the quality of predictions, we build a weak supervision framework for RPH with a pseudo relation label generation pipeline, named DL-Cluster, which considers both text spatial and semantic relations, yielding superior text relation labels compared to traditional clustering methods [2, 6, 9].", + "bbox": [ + 89, + 476, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our experiments on LecSlides-370K, related baselines struggle to extract the main point of lecture slides and pinpoint the correct content corresponding to the question, revealing the challenges of our dataset. By effectively predicting and utilizing text relations, SlideParser outperforms all the related methods in the LecSlides-370K Lecture Summary and Lecture QA benchmarks. Additionally, we evaluate SlideParser on the SlideVQA [27] dataset, further validating the effectiveness of our text relation design.", + "bbox": [ + 89, + 703, + 482, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In conclusion, our contributions are:", + "bbox": [ + 109, + 839, + 352, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We establish a large-scale and multi-task slides understanding dataset, termed LecSlides-370K, comprising 25,542 lectures across 15 lecture areas and 370,078", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "slides. We also introduce two comprehensive tasks, Lecture Summary and Lecture QA.", + "bbox": [ + 526, + 383, + 903, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a novel method, SlideParser, including an auxiliary brach to effectively predict and utilize complex text relations in slides, enhancing slides understanding.", + "bbox": [ + 513, + 415, + 903, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 494, + 653, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Lecture Slides Datasets", + "text_level": 1, + "bbox": [ + 513, + 525, + 728, + 539 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, several datasets based on lecture slides have been proposed, which are most relevant to our current work, including the LectureBank Dataset [17], ALV [7], LectureVideoDB [5], LPM Dataset [14], $\\mathbf{M}^3\\mathrm{AV}$ [4] and SlideVQA [27], as shown in Table 1.", + "bbox": [ + 511, + 554, + 905, + 630 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LectureBank [17] is annotated for prerequisite relationships among topics, providing a structure of connected knowledge. ALV dataset [7] primarily focuses on segmenting the entire video lecture. LectureVideoDB is formulated to investigate text detection and recognition in video lectures. Furthermore, LPM [14] attempts to explore the retrieval of spoken words with corresponding visual cues. SlideVQA, a recent work for slides, proposes a Visual Question Answering (VQA) task on it.", + "bbox": [ + 511, + 637, + 905, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, previous datasets contain two primary limitations: 1) Most of datasets do not pay sufficient attention to slides understanding, and the tasks they define also do not provide multiple perspectives for understanding. 2) The limited data scale hinders the advancement of understanding lecture slides. Therefore, a substantial requirement in the research community emerges for a large-scale slides dataset with comprehensive understanding tasks.", + "bbox": [ + 511, + 780, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4456", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/e508a39255bc9c2167651a9255b6121a8dd2a5e84eda82551c0fca5a51b3f6a7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSizeAnnotationTask
LecturesSlidesOCRQASummary
LectureVideoDB [5]245,000✓--Text Detection
LectureBank [17]1,35251,939✓--Prerequisite Chain
ALV Dataset [7]-1,498✗--Slide Segmentation
LPM [14]3349,031✓--Image Retrieval
M3AV [4]111324,956✓--ASR & TTS
SlideVQA [27]2,61952,380✓14,500-QA
LecSlides-370K (Ours)25,542370,078✓259,461114,472QA & Summary
", + "bbox": [ + 181, + 88, + 816, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Data scales, annotations and tasks for different lecture slides datasets. LecSlides-370K contain the largest data scale, the most comprehensive annotations and two understanding tasks to evaluate slides understanding from multiple perspectives.", + "bbox": [ + 89, + 260, + 906, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Slides Understanding Methods", + "text_level": 1, + "bbox": [ + 89, + 314, + 362, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although there is a lack of specific methods for slide understanding, other area methods can be used in slide understanding, mainly including three categories: vision-language methods [16, 31, 34, 36], NLP methods [15, 22], and document understanding methods [10, 28].", + "bbox": [ + 89, + 337, + 482, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-language methods typically begin with large-scale pre-training tasks and then fine-tune on downstream multi-modal understanding tasks. These methods often integrate a vision-language fused transformer encoder [16, 31, 36]. NLP methods provide a plain text solution, such as T5 [22] and BART [15], following an encoder-decoder architecture to process textual contents and generate textual outputs. Furthermore, inspired by NLP approaches, document understanding methods replace traditional transformer encoder to multi-modal encoder, processing and fusing textual and visual content. For example, LayoutLM series [10, 32, 33] integrate text, image features, and 2D positional embeddings in the multi-modal encoder.", + "bbox": [ + 88, + 414, + 482, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, these related methods do not consider the complex text relations in slides, a crucial part of slides understanding, thereby yielding inferior performance.", + "bbox": [ + 89, + 611, + 483, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. LecSlides-370K Dataset", + "text_level": 1, + "bbox": [ + 89, + 672, + 313, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection", + "text_level": 1, + "bbox": [ + 89, + 699, + 246, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To build a large-scale and multi-modal lecture slides dataset, we collect slides from Slideshare*. Our raw data consists of 720,673 slides from 39,784 lectures at the beginning. However, some of these slides are irrelevant to the lecture content, such as advertisements, or contain extraneous information, which is not essential for comprehension. Thus, we filter these slides by their contents. Finally, LecSlides-370K contains 370,078 slides from 25,542 lectures across 15 areas, such as computer science, marketing, medicine, and so on. As shown in Table 1, compared with", + "bbox": [ + 89, + 722, + 482, + 875 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the previous largest datasets, SlideVQA [27], LecSlides-370K has around 9.7 times lecture numbers and 7.1 times slide numbers. The superiority of data scale and diversity renders our dataset more comprehensive and challenging.", + "bbox": [ + 511, + 314, + 903, + 376 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, since the length of these raw lecture slides varies widely, from a few to hundreds, which makes it difficult for researchers to use them, we segment these raw slides into 114,472 lecture chunks, following the principle that each chunk belongs to the same sub-topic. For instance, if one slide initially displays the first point of a list, and the subsequent slide reveals the following points, these slides should belong to the same chunk. The distribution of slides number in each lecture chunk is shown in Fig. 2 (b). Additionally, we annotate the positions of texts, illustrations, and tables in slides to facilitate the study and encourage researchers to focus more on developing new methods rather than information extraction and data processing. Furthermore, we also annotate the textual contents in slides, which can be utilized by the models to understand slides better.", + "bbox": [ + 511, + 377, + 906, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Task Definition", + "text_level": 1, + "bbox": [ + 511, + 619, + 666, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define two tasks: Lecture Summary and Lecture QA. These two tasks not only involve most application scenarios, but also take into account the global and detailed understanding of slides, fully reflecting the understanding ability of models. In the context of Lecture Summary, given a lecture chunk $\\mathbf{I} = \\{I_1, I_2, \\dots, I_j\\}^M$ , comprising $M$ slide images, the model's objective is to analyze the content of each slide, integrate contextual information, retain key details, and eliminate redundancies. Finally, the model outputs a summary of all slides. Contrary to the global scope of Lecture Summary, Lecture QA primarily concentrates on locating and comprehending specific content within a lecture chunk. Given a lecture chunk and a query $q$ , the model need to identify the slide corresponding to the question and then provide an answer based on the content of that slide. QA serves to assess the model's fine-grained understanding of slides. In total, We annotate 114,472 summary labels and", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "\\*https://www.slideshare.net", + "bbox": [ + 109, + 886, + 256, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4457", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c08e6c9169570e6bf5c56f30f5c5dade21740a5eb4f6d5d9528b09bb1990e4e1.jpg", + "image_caption": [ + "Figure 2. Examples and statistics of LecSlides-370K. (a) Lecture numbers in each discipline for LecSlides-370K, SlideVQA [27] and LPM [14]. (b) The number of slides for each chunk. (c) The distribution of OCRs in slides. (d) The distribution of summary length. (e) The percentage of each QA type." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 464, + 321 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fd1f0a697e404d64196040d4d3fdefd28389d04d2e5bdf52a8c1af4528406c87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 464, + 88, + 671, + 202 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c8a26ff8849eee5283b2dda52594a9195039edf05fb45c86cd5a43a438a64a6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 678, + 89, + 893, + 203 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "259,461 QA pairs.", + "bbox": [ + 89, + 402, + 217, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dataset Statistics and Analysis", + "text_level": 1, + "bbox": [ + 89, + 425, + 362, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we aim to provide a comprehensive analysis and understanding of the LecSlides-370K dataset.", + "bbox": [ + 89, + 446, + 482, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Lecture Diversity", + "text_level": 1, + "bbox": [ + 89, + 484, + 256, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We collect slides from various disciplines, as shown in Fig. 2 (a). There are 15 disciplines, and the highest proportion of data pertains to business, while law and marketing lectures constitute a relatively small percentage. Compared with previous slides datasets, such as SlideVQA [27] and LPM [14], LecSlides-370K contains more disciplines, such as marketing, agriculture, and so on. The diversity of disciplines contains rich teaching content, bringing challenges that models need to handle more diverse scenes.", + "bbox": [ + 89, + 502, + 483, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Professionalism", + "text_level": 1, + "bbox": [ + 89, + 645, + 246, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lecture slides usually contain specialized knowledge and terminologies, bringing more challenges. Therefore, we use the Flesch Reading Ease (FRES) score [13] to evaluate the difficulty of contents and get an average score of 34.3 under this metric. This indicates that the content of our dataset is on 'Hard to Read' level. Moreover, we also use the Flesch-Kincaid Grade Level [13] to reflect which education level our dataset is appropriate for and obtain a 14.1 grade on average. These show there is a professional difficulty in our dataset.", + "bbox": [ + 89, + 662, + 482, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.3. Slide Image", + "text_level": 1, + "bbox": [ + 89, + 821, + 218, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For LecSlides-370K, there are 114,472 lecture chunks corresponding to 370,078 slide images. Each chunk contains a varying number of slide images, with an average of 3.2 images per chunk. The distribution of chunks containing", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "different numbers of slide images is shown in Fig. 2 (b), and most chunks contain two slide images. Moreover, slide images are rich in visual information. By our count, there are 94,906 diagrams, 34,562 illustrations, and 18,938 tables in LecSlides-370K dataset. We further show the statistics for textual content in slides. We totally annotate 28,220,274 words, with an average of 76.2 words. The size of the vocabulary for our dataset is 84,630. In Fig. 2 (c), we show the distribution of texts across slides, revealing that the majority of slides contain fewer than 100 words.", + "bbox": [ + 511, + 402, + 906, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.4. Characteristics of Annotation", + "text_level": 1, + "bbox": [ + 511, + 561, + 766, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For summary annotations, Fig. 2 (d) illustrates the length distribution of the annotations. The average length of the summary is 124.6 words. We further comprehensively analyze QA annotations, including four distinct question types: binary choice, visual analysis, data interpretation and textual reasoning. As shown in Fig. 2 (e), the highest proportion of question types is textual reasoning, while binary choice constitutes the smallest percentage. This diverse set of question types covers the whole slide content, enabling a comprehensive evaluation of the model's QA capabilities. In addition, the average length of answer is 4.1 words.", + "bbox": [ + 511, + 580, + 906, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 511, + 757, + 702, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "LecSlides-370K dataset contains 109,472 slide chunks for the training set, 2,000 chunks for the validation set and 3,000 chunks for the test set. We adapt the ROUGE-1, ROUGE-L [18] and METEOR [3] as evaluation metrics for Lecture Summary task. These metrics are widely used in image caption and text generation, as they consider the quality of generated texts in different aspects. Following SlideVQA [27], we choose EM (Exact Matching) and F1", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4458", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/77a82f29de098fe3925f47adbf0e1703f41e4c8cb5372a4d86334600716c29b8.jpg", + "image_caption": [ + "(a) Overview of SlideParser Architecture" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 480, + 246 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6174a25993604cd65c0b87875edfefe565c05f5177a7a3b0c0137c1e8dfe9e76.jpg", + "image_caption": [ + "(b) Architecture of Text Relation Prediction Module", + "Figure 3. (a) The architecture of our proposed method, SlideParser. It consists of a Transformer Encoder, a Text Relation Prediction Module, and a Text Decoder. (b) The details of Text Relation Prediction Module, which can predict and utilize text relations effectively, including the Relation Prediction Head (RPH) with weak supervision and the Relation Attention Layer (RAL)." + ], + "image_footnote": [], + "bbox": [ + 491, + 85, + 903, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "score as evaluation metrics for the Lecture QA task.", + "bbox": [ + 89, + 342, + 434, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 371, + 181, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3 (a), SlideParser consists of a Transformer Encoder, a Text Relation Prediction Module, and a Text Decoder. Texts and image patches from slides are fed into the Transformer Encoder, and the output text embeddings are forwarded to the Text Relation Prediction Module to predict text relations and utilize them to update text embeddings. Finally, updated text embeddings combine with output image embeddings and are fed into the Text Decoder to generate the summary or answer.", + "bbox": [ + 88, + 397, + 483, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Input Schema", + "text_level": 1, + "bbox": [ + 89, + 542, + 235, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SlideParser is built on a unified text-image multi-modal Transformer Encoder to learn cross-modal representations. Similar to previous document understanding works [10, 28], the Transformer Encoder adopts a multi-layer structure, with each layer primarily comprising multi-head self-attention and feed-forward networks. The input of Transformer Encoder consists of text embeddings $\\mathbf{X} = x_{1:M}$ and image embedding $\\mathbf{V} = v_{1:M}$ , where $L$ and $M$ are sequence length of textual and visual embeddings respectively. Specifically, for text embedding $\\mathbf{X}$ , $L$ text tokens $\\{s_i\\}_{i=1}^L$ are encoded by word embedding with 1D position and 2D layout position embedding [28]. Meanwhile, the slide image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}$ is divided into $M = \\frac{H}{P} \\times \\frac{H}{P}$ image patches. These patches are then processed through Patch Embedding to generate the image embeddings, $\\mathbf{V}$ .", + "bbox": [ + 89, + 565, + 483, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Text Relation Prediction Module", + "text_level": 1, + "bbox": [ + 89, + 801, + 377, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Transformer Encoder, 2D position embedding, which can be used to comprehend layout and text relations, has been demonstrated in previous document understanding works [10, 28]. However, in the context of slides, text relations tend to be more complex compared to traditional doc", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "uments, such as examples in Fig. 2. Merely incorporating 2D position embeddings into the transformer encoder may not suffice for comprehensive relation comprehension and could potentially introduce inaccuracies.", + "bbox": [ + 511, + 342, + 906, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this issue, we introduce the Text Relation Prediction Module, including the Relation Prediction Head (RPH) with weak supervision and the Relation Attention Layer (RAL), as shown in Fig. 3 (b). This module predicts text relations by RPH and utilizes relations to enhance attention between related texts while reducing attention between unrelated texts by RAL. Additionally, to improve the quality of predicted relations, we establish a weak supervision pipeline. We will introduce each block in detail.", + "bbox": [ + 511, + 404, + 908, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Relation Prediction Head", + "text_level": 1, + "bbox": [ + 511, + 554, + 736, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To predict the accurate text relations, inspired by [30], RPH leverages text position and semantic information to predict text relations. Given text bounding boxes $\\{b_i\\}_{i=1}^L$ , we define the box delta as:", + "bbox": [ + 511, + 575, + 906, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Delta \\left(b _ {i}, b _ {j}\\right) = \\left\\{\\left(x _ {i} ^ {c t r} - x _ {j} ^ {c t r}\\right) / w _ {i}, \\left(y _ {i} ^ {c t r} - y _ {j} ^ {c t r}\\right) / h _ {i}, \\right. \\tag {1} \\\\ \\left. \\log \\left(w _ {i} / w _ {j}\\right), \\log \\left(w _ {i} / w _ {j}\\right) \\right\\} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 647, + 903, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{i}^{ctr}$ and $y_{i}^{ctr}$ represent the position of the center point for bounding box, and $w$ and $h$ are the width and height of box respectively. Moreover, we present the spatial relation of two boxes as: $\\mathbf{r}_{\\mathbf{bi},\\mathbf{bj}} = \\{\\Delta (b_i,b_j),\\Delta (b_i,b_{ij}),\\Delta (b_{ij},b_j)\\}$ , which is a 12-D vector. Specifically, $b_{ij}$ is a union bounding box for $b_{i}$ and $b_{j}$ . Then, we employ an MLP layer and integrate all box relations to get the spatial relation matrix $\\mathbf{R}_{\\mathbf{spa}} \\in \\mathbb{R}^{L\\times L}$ . In addition, RPH also considers semantic relations. Given output text embeddings from Transformer Encoder, $\\mathbf{E_T} \\in \\mathbb{R}^{L\\times D}$ , we compute the semantic relation between two texts as", + "bbox": [ + 511, + 702, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R _ {s e m}} = \\left(\\mathbf {W _ {q} E _ {T}} \\times \\mathbf {W _ {k} E _ {T} ^ {T}}\\right) / \\sqrt {D} \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 883, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4459", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/005dc09c7e3e684565abd43a5b65b6a987075f72cf5ee7b14d246c37f9095602.jpg", + "image_caption": [ + "Figure 4. The Pipeline of DL-Cluster to generate pseudo relation label. We combine spatial and semantic text relations to get pseudo labels." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 888, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The predicted relation matrix $\\hat{\\mathbf{M}}_{\\mathbf{R}} \\in \\mathbb{R}^{L \\times L}$ is computed as", + "bbox": [ + 89, + 276, + 482, + 292 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {\\hat {M} _ {R}} = \\operatorname {S i g m o i d} \\left(\\mathbf {R} _ {\\mathbf {s p a}} + \\mathbf {R} _ {\\mathbf {s e m}}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 308, + 482, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Weak Supervised Label Generation", + "text_level": 1, + "bbox": [ + 89, + 335, + 383, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our experiment, we find that the RPH struggles to generate reasonable text relations without the presence of relation supervision. As a solution, we introduce a weak supervision pipeline to enhance the quality of predictions. Previous works [1, 8] rely on traditional cluster methods [6, 9] to predict text relations. However, these traditional methods can not handle diverse layout format in slides effectively. Moreover, our observations suggest that both spatial and semantic relations play crucial roles in slides.", + "bbox": [ + 89, + 354, + 483, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Thus, we introduce a dual-stream pseudo relation label generation pipeline, termed DL-Cluster (Dynamic Layout Cluster), as shown in Fig. 4. To obtain spatial relations, we first sample $N$ reference points as cluster points for each text box. Because of the diverse nature of text relations in slides, a predefined number of clusters may not be ideal for accurate relation predictions. Hence, we opt for a dynamic approach to determine the cluster number for each slide.", + "bbox": [ + 89, + 491, + 483, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, based on Kmeans++ [2], we iterate over the number of cluster $k$ from 2 to $\\mathcal{K}$ and determine the optimal cluster number by Silhouette method [24]. Silhouette analyzes the distances of each cluster point to its own cluster and its closest neighboring cluster. The optimal $k$ is obtained when the average distance of a cluster point to all the other points in its own cluster is minimum, while the average distance of a point to the nearest neighboring cluster's points is maximum. In our experiment, $\\mathcal{K} = 8$ .", + "bbox": [ + 89, + 613, + 483, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subsequently, we assign a cluster ID to each box through a voting mechanism among reference points. Then, the spatial relation score $s_{spa}$ is computed as", + "bbox": [ + 89, + 750, + 482, + 795 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {\\mathbf {s p a}} = \\left\\{ \\begin{array}{l} 1 - \\frac {1}{N} \\sum_ {\\varepsilon = 1} ^ {N} \\left\\| x _ {i, \\varepsilon} - \\mu_ {j} \\right\\|, \\text {w h e n} C _ {i} = C _ {j} \\\\ 0, \\text {w h e n} C _ {i} \\neq C _ {j} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 806, + 482, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $C_i$ and $C_j$ are clusters and $\\mu_j$ is cluster centroid of $C_j$ . Specifically, scores are assigned based on the average distance between reference points and the cluster centroid", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "when these texts are deemed related and are placed within the same cluster. Conversely, when texts are considered unrelated, the score is set to zero.", + "bbox": [ + 511, + 277, + 903, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Simultaneously, to obtain semantic relation, we extract textual contents in each text box, and compute the semantic similarity by the cosine distance of Sentence Transformer [23] and get the semantic relation score $s_{sem}$ . These two relation scores are combined as", + "bbox": [ + 511, + 323, + 905, + 398 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} = \\beta_ {1} \\operatorname {N o r m} \\left(\\mathbf {s} _ {\\mathbf {s p a}}\\right) + \\beta_ {2} \\operatorname {N o r m} \\left(\\mathbf {s} _ {\\mathbf {s e m}}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 410, + 903, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we set the default values of $\\beta_{1} = 0.7$ , $\\beta_{2} = 0.3$ . Finally, the relation matrix $M_{R} = \\{s_{ij}\\}^{L\\times L}$ is binarized with 0.5 threshold. Then, the text relation loss can be calculated as", + "bbox": [ + 511, + 436, + 905, + 482 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e l} = B C E L o s s \\left(\\hat {\\mathbf {M}} _ {\\mathbf {R}}, \\mathbf {M} _ {\\mathbf {R}}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 493, + 903, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3. Relation Attention Layer", + "text_level": 1, + "bbox": [ + 511, + 522, + 735, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the predicted text relations, we utilize relations by Self Attention [29], as shown in Fig. 3 (b). Specifically, for predicted relation matrix $\\hat{\\mathbf{M}}_{\\mathbf{R}}$ , we generate attention mask by assigning $-\\infty$ when $m_{ij} < 0.5$ , $m_{ij} \\in \\hat{\\mathbf{M}}_{\\mathbf{R}}$ . Subsequently, the output text embedding $\\mathbf{E}_{\\mathbf{T}}$ are fed into N-layer self-attention blocks with this attention mask, resulting in the updated embeddings, $\\mathbf{E}_{\\mathbf{T}}'$ . It enhances attention between related texts and isolates attention between unrelated texts. Finally, we combine $\\mathbf{E}_{\\mathbf{T}}$ and $\\mathbf{E}_{\\mathbf{T}}'$ by residual connection.", + "bbox": [ + 511, + 540, + 906, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Text Decoder", + "text_level": 1, + "bbox": [ + 511, + 686, + 653, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following previous work [28], we adapt a T5 decoder as our text decoder, a uni-directional Transformer decoder that performs the next prediction task to generate texts, such as summary or answer. We choose Cross Entropy loss to supervise text generation. The total loss of SlideParser is", + "bbox": [ + 511, + 708, + 905, + 784 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c e} + \\mathcal {L} _ {r e l} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 795, + 903, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 823, + 645, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implement details", + "text_level": 1, + "bbox": [ + 511, + 848, + 689, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For SlideParser, both the Transformer Encoder and the Text Decoder are initialized from the pre-trained UDOP [28]", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4460", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/68fa788b9579795cfee2e43b99108133c9624d88745da03a0018ea7d45f6bf88.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method#Paramsvaltest
R-1R-LMETEORR-1R-LMETEOR
BART-base [15]217M33.121.121.932.820.121.5
BART-large [15]509M35.921.924.235.021.022.9
T5-base [22]272M34.621.622.834.320.722.4
T5-large [22]803M36.222.124.935.121.323.2
Vid2Seq [34]314M16.011.812.917.813.114.1
mPLUG-base [16]350M27.119.021.627.619.321.7
mPLUG-large [16]600M30.020.524.029.820.323.6
Donut [12]543M27.418.621.327.618.521.1
LayoutLMv3-base [10]533M33.621.023.831.719.822.6
LayoutLMv3-large [10]768M34.621.624.833.921.023.2
UDOP [28]794M36.122.126.435.321.525.6
SlideParser (Ours)810M42.225.331.341.724.730.4
", + "bbox": [ + 94, + 88, + 500, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/87e7b3a0f61d2f94be6f7a91bcb56af5f7b1ea216f3ba7bfa91b6ac6648edfa0.jpg", + "table_caption": [ + "Table 2. Performance on Lecture Summary in LecSlides-370K" + ], + "table_footnote": [], + "table_body": "
Method#Paramsvaltest
EMF1EMF1
BART-base [15]217M18.925.118.524.7
BART-large [15]509M20.627.119.826.3
T5-base [22]272M19.126.018.725.3
T5-large [22]803M20.927.420.027.1
Vid2Seq [34]314M7.014.36.713.6
mPLUG-base [16]350M8.410.98.510.1
mPLUG-large [16]600M9.514.39.514.2
LayoutLMv3-base [10]533M12.515.812.415.1
LayoutLMv3-large [10]768M13.619.313.519.1
UDOP [28]794M19.226.018.725.6
SlideParser (Ours)810M23.733.023.632.1
", + "bbox": [ + 98, + 286, + 475, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "model weights. We adapt AdamW [19] optimizer with $\\beta_{1}$ of 0.9, the $\\beta_{2}$ of 0.999, and the weight decay of 0.02. We use the cosine learning rate scheduler and the initial learning rate is $4\\times 10^{-5}$ . The total number of iterations is 41K. Moreover, we choose a simple concat operator for encoder outputs to handle multiple slide situations for all models.", + "bbox": [ + 89, + 525, + 482, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To provide a full view of LecSlides-370K, we evaluate other related methods from various domains: NLP models such as T5 [22] and BART [15], vision-language models such as mPLUG [16] and Vid2Seq [34], and document understanding models such as LayoutLMv3 [10], Donut [12] and UDOP [28]. Specifically, LayoutLMv3 lacks a text decoder, which is unable to handle text generation tasks. To address this limitation, we incorporate the same text decoder as our proposed method.", + "bbox": [ + 89, + 617, + 482, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 89, + 766, + 184, + 780 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.1. Performance on LecSlides-370K", + "text_level": 1, + "bbox": [ + 89, + 789, + 362, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "At the beginning, we analyze Lecture Summary task in LecSlides-370K. As shown in Table 2, SlideParser achieves notable scores of 42.2 in ROUGE-1, 25.3 in ROUGE-L, and 31.3 in METEOR. This performance surpasses other baseline methods, attributed to its effective understanding text relations. Specifically, the methods without processing tex", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ad8512deb9a800685f74358fa29cbab858aea441505f28261510129b699c69f6.jpg", + "table_caption": [ + "Table 3. Performance on Lecture QA in LecSlides-370K" + ], + "table_footnote": [], + "table_body": "
Model Namedevtest
EMF1EMF1
VinVL [36]9.411.410.713.5
PreasM [35]36.341.930.738.2
T5 [22]35.241.331.039.7
LayoutT5 [26]38.944.831.739.9
LayoutLMv2 [33]26.533.421.429.3
FiD [11]37.642.930.438.9
M3D [27]41.347.133.541.7
SlideParser (Ours)44.149.837.543.9
SlideParser (Ours)†49.853.942.447.7
", + "bbox": [ + 563, + 88, + 857, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Performance on SlideVQA benchmark. † represents pretraining on LecSlides-370K and fine-tuning on SlideVQA.", + "bbox": [ + 511, + 255, + 903, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tual modality, including Vid2Seq [34], mPLUG series [16] and Donut [12], exhibit the worst results. We argue that only processing visual modality may lead to a lack of detailed content extraction in slides. In addition, NLP methods, such as T5 [22] and BART [15], perform relatively well but still lag behind our proposed method, containing excellent multi-modal fusion and text relation utilization.", + "bbox": [ + 511, + 311, + 903, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In contrast to Lecture Summary, which primarily evaluates global understanding of slides, Lecture QA task measures methods for in-depth understanding of specific content. As shown in Table 3, SlideParser achieves superior performance compared to other baselines. Furthermore, visual-language models such as mPLUG and Vid2Seq still perform the worst among all baselines. It shows that detailed text information is still crucial for QA task.", + "bbox": [ + 511, + 417, + 905, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.2. Performance on SlideVQA", + "text_level": 1, + "bbox": [ + 511, + 547, + 745, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further evaluate our proposed method in QA task, we also fine-tune and evaluate SlideParser on SlideVQA, a VQA benchmark in multi-image slides. As shown in Table 4, SlideParser achieves the best performance compared with other methods, which further demonstrates the effective designs of SlideParser. In an extensive experiment, we pre-train SlidePraser in our dataset and fine-tune on SlideVQA, denoted as SlideParser†. Results show that largescale and comprehensive LecSlides-370K yield a great performance enhancement, demonstrating our dataset can also serve as a valuable large-scale pre-training dataset.", + "bbox": [ + 511, + 566, + 905, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 743, + 666, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To gain a deeper analysis of SlideParser, we conduct ablation experiments in LecSlides-370K Lecture Summary task.", + "bbox": [ + 511, + 766, + 903, + 796 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.1. Effect of Text Relation Prediction Module.", + "text_level": 1, + "bbox": [ + 511, + 805, + 852, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first evaluate the impact of incorporating the Text Relation Prediction Module. As shown in Table 5, results demonstrate the removal of the entire Text Relation Prediction Module (in r1) yields the poorest performance. Furthermore, as we add RPH and RAL to predict and utilize rela", + "bbox": [ + 511, + 824, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4461", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c02792924cdeda7fe787033e8e6145ecd1348554a23cbc4a5a70b8ca5793d571.jpg", + "image_caption": [ + "Figure 5. Visualization of predicted text relations in RPH" + ], + "image_footnote": [], + "bbox": [ + 124, + 87, + 450, + 203 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/71d5dd6a62a40fc3c44fe3be0cc67fdb7e81cdde4f2de762072afd4408139ff0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RPHRALPseudo LabelR-1R-LMETEOR
r1XXX36.322.626.9
r2✓✓X37.523.227.7
r3X✓✓41.825.030.8
r4✓✓✓42.225.331.3
", + "bbox": [ + 112, + 243, + 460, + 323 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/90a67d1390ddb96ca4a0dc3eed782a81e18852686728d04e6f4f81ac52857c07.jpg", + "table_caption": [ + "Table 5. Ablation on Text Relation Prediction Module" + ], + "table_footnote": [], + "table_body": "
Cluster TypeR-1R-LMETEOR
Kmeans (k=3)40.924.829.8
Kmeans (k=4)40.724.729.6
HAC(n=3)41.024.830.1
DBSCAN41.324.830.6
DL-Cluster42.225.331.3
", + "bbox": [ + 166, + 363, + 406, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tions without any supervision (in r2), the performance only has a slight improvement. It demonstrates the RPH struggles to converge effectively in generating reasonable text relations. In addition, we also explore the approach of directly feeding pseudo relation labels generated by DL-Cluster into the subsequent RAL, instead of utilizing learnable relations (in r3). Although it exhibits substantial improvement, yet it continues to fall short in comparison to leveraging learnable text predictions and establishing weak supervision (r3 vs r4). It shows that learnable relations could be refined during the training process.", + "bbox": [ + 89, + 508, + 482, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, we provide visualization results in Fig. 5. We assign the same color for two texts if their relation score is larger than 0.5. The results illustrate that the RPH can predict reasonable text relations under weak supervision.", + "bbox": [ + 89, + 675, + 482, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.2. Different Methods for Pseudo Label Generation.", + "text_level": 1, + "bbox": [ + 89, + 744, + 473, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Traditional cluster methods are widely used in previous works for relation prediction [1, 8]. Therefore, we compare our proposed DL-Cluster with other cluster methods. As shown in Table 6, DL-Cluster surpasses clustering techniques with predefined cluster numbers like K-means and Hierarchical Agglomerating Clustering (HAC). Additionally, DL-Cluster outperforms methods without specified cluster numbers, such as DBSCAN [6], due to its capacity to consider spatial and semantic aspects concurrently.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7f67174f07536019fe83af6782212bd13b307d919acadf911d3055cf3985747a.jpg", + "image_caption": [ + "Figure 6. Qualitative results of Lecture QA in LecSlides-370K." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 908, + 198 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a5f64f3b6b417d21dbaf4fc9020aa24a1f3827b4a9ea73c6a22773010c5354e2.jpg", + "table_caption": [ + "Table 6. Ablation on cluster types in pseudo label generation" + ], + "table_footnote": [], + "table_body": "
SpatialSemanticR-1R-LMETEOR
×✓38.323.728.5
✓×41.424.830.7
✓✓42.225.331.3
", + "bbox": [ + 576, + 243, + 843, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation on designs of DL-Cluster", + "bbox": [ + 578, + 320, + 839, + 334 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.3. Design of DL-Cluster.", + "text_level": 1, + "bbox": [ + 511, + 369, + 709, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further investigate the impact of text relations represented in spatial and semantic dimensions. As shown in Tab. 7, removing either of these dimensions leads to a decline in performance, and spatial relations have a larger impact. The best results are achieved when both aspects were used simultaneously, indicating that both dimensions complement each other to enhance overall performance.", + "bbox": [ + 511, + 392, + 906, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Visualization", + "text_level": 1, + "bbox": [ + 511, + 518, + 650, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In a side-by-side qualitative analysis for Lecture QA in LecSlides-370K, we compare our proposed method with other baselines in related fields. As shown in Fig. 6, results show that other methods face two challenges: 1) They can not grasp correct internal logic due to the complex text relations. 2) As the number of diagrams on slides increases, other methods struggle to pinpoint the correct diagram corresponding to the question. In contrast, SlideParser excels in accurate responses by comprehending text relations effectively and easily determining the answer's location within the slides.", + "bbox": [ + 511, + 544, + 906, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 734, + 633, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce a large-scale and multi-task lecture slides dataset, LecSlides-370K, which has a larger data scale and more diverse disciplines. Furthermore, considering the pivotal role of text relation prediction and utilization for slides understanding, we propose SlideParser. The results demonstrate the effect of enhanced text relation design, and our method outperforms other models. We believe LecSlides-370K and SlideParser can prompt lecture slides research.", + "bbox": [ + 511, + 763, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4462", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the NSFC (62206103 and 62225603)", + "bbox": [ + 89, + 114, + 483, + 147 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 162, + 187, + 178 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rhys Agombar, Max Luebbering, and Rafet Sifa. A clustering backed deep learning approach for document layout analysis. In Machine Learning and Knowledge Extraction: 4th IFIP TC 5, TC 12, WG 8.4, WG 8.9, WG 12.9 International Cross-Domain Conference, CD-MAKE 2020, Dublin, Ireland, August 25–28, 2020, Proceedings 4, pages 423–430. Springer, 2020. 6, 8", + "[2] David Arthur and Sergei Vassilvitskii. k-means++: the advantages of careful seeding. In Society for Industrial and Applied Mathematics, page 1027-1035, 2007. 2, 6", + "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 4", + "[4] Zhe Chen, Heyang Liu, Wenyi Yu, Guangzhi Sun, Hongcheng Liu, Ji Wu, Chao Zhang, Yu Wang, and Yanfeng Wang. M3av: A multimodal, multigenre, and multipurpose audio-visual academic lecture dataset. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 9041–9060, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1, 2, 3", + "[5] Kartik Dutta, Minesh Mathew, Praveen Krishnan, and CV Jawahar. Localizing and recognizing text in lecture videos. In 2018 16th international conference on frontiers in handwriting recognition (ICFHR), pages 235-240. IEEE, 2018. 1, 2, 3", + "[6] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, pages 226-231, 1996. 2, 6, 8", + "[7] Damianos Galanopoulos and Vasileios Mezaris. Temporal lecture video fragmentation using word embeddings. In MultiMedia Modeling: 25th International Conference, MMM 2019, Thessaloniki, Greece, January 8-11, 2019, Proceedings, Part II 25, pages 254-265. Springer, 2019. 1, 2, 3", + "[8] Zhangxuan Gu, Changhua Meng, Ke Wang, Jun Lan, Weiqiang Wang, Ming Gu, and Liqing Zhang. Xylayoutlm: Towards layout-aware multimodal networks for visually-rich document understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4583-4592, 2022. 6, 8", + "[9] Jaekyu Ha, R.M. Haralick, and I.T. Phillips. Recursive x-y cut using bounding boxes of connected components. In Proceedings of 3rd International Conference on Document Analysis and Recognition, pages 952-955, 1995. 2, 6", + "[10] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM" + ], + "bbox": [ + 93, + 186, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Multimedia, pages 4083-4091, 2022. 1, 3, 5, 7", + "[11] Gautier Izacard and Edouard Grave. Leveraging passage retrieval with generative models for open domain question answering. arXiv preprint arXiv:2007.01282, pages 1-6, 2020. 7", + "[12] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 7", + "[13] JP Kincaid. Derivation of new readability formulas (automated readability index, fog count and flesch reading ease formula) for navy enlisted personnel. Chief of Naval Technical Training, pages 1-45, 1975. 4", + "[14] Dong Won Lee, Chaitanya Ahuja, Paul Pu Liang, Sanika Natu, and Louis-Philippe Morency. Lecture presentations multimodal dataset: Towards understanding multimodality in educational videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20087-20098, 2023. 1, 2, 3, 4", + "[15] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online, 2020. Association for Computational Linguistics. 1, 3, 7", + "[16] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, He Chen, Guohai Xu, Zheng Cao, Ji Zhang, Songfang Huang, Fei Huang, Jingren Zhou, and Luo Si. mPLUG: Effective and efficient vision-language learning by cross-modal skip-connections. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7241–7259, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 1, 3, 7", + "[17] Irene Li, Alexander R Fabbri, Robert R Tung, and Dragomir R Radev. What should i learn first: Introducing lecturebank for nlp education and prerequisite chain learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6674-6681, 2019. 1, 2, 3", + "[18] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 4", + "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, pages 1-19, 2017. 7", + "[20] T. Muthuprasad, S. Aiswarya, K.S. Aditya, and Girish K. Jha. Students' perception and preference for online education in india during covid -19 pandemic. Social Sciences Humanities Open, 3(1):100101, 2021. 1", + "[21] Pitambar Paudel. Online education: Benefits, challenges and strategies during and after Covid-19 in higher education. International Journal on Studies in Education (IJonSE), 3:70–85, 2021. 1" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4463", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1, 3, 7", + "[23] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3982-3992. Association for Computational Linguistics, 2019. 6", + "[24] Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53-65, 1987. 6", + "[25] Anna Sun and Xiufang Chen. Online education and its effective practice: A research review. Journal of information technology education: Research, 15:1-34, 2016. 1", + "[26] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 7", + "[27] Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. Slidevqa: A dataset for document visual question answering on multiple images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13636-13645, 2023. 1, 2, 3, 4, 7", + "[28] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19254-19264, 2023. 1, 3, 5, 6, 7", + "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 6000-6010, 2017. 6", + "[30] Jiawei Wang, Kai Hu, Zhuoyao Zhong, Lei Sun, and Qiang Huo. Detect-order-construct: A tree construction based approach for hierarchical document structure analysis. arXiv preprint arXiv:2401.11874, pages 1-35, 2024. 5", + "[31] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 3", + "[32] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192-1200, 2020. 1, 3", + "[33] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. LayoutLMv2: Multi-modal pre-training for visually-rich document understanding. In Proceedings of the 59th Annual Meeting of the" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2579-2591, Online, 2021. Association for Computational Linguistics. 1, 3, 7", + "[34] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 1, 3, 7", + "[35] Ori Yoran, Alon Talmor, and Jonathan Berant. Turning tables: Generating examples from semi-structured tables for endowing language models with reasoning skills. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pages 6016–6031. Association for Computational Linguistics, 2022. 7", + "[36] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5579-5588, 2021. 3, 7" + ], + "bbox": [ + 516, + 92, + 903, + 398 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4464", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/58e52446-3fb4-48af-8a99-0aa5fc34cde3_model.json b/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/58e52446-3fb4-48af-8a99-0aa5fc34cde3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0d13794912919135c97f4b4affa7d4a03d1f68c7 --- /dev/null +++ b/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/58e52446-3fb4-48af-8a99-0aa5fc34cde3_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.13, + 0.905, + 0.175 + ], + "angle": 0, + "content": "Towards Comprehensive Lecture Slides Understanding: Large-scale Dataset and Effective Method" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.203, + 0.777, + 0.258 + ], + "angle": 0, + "content": "Enming Zhang\\(^{1}\\), Yuzhe Li\\(^{1}\\), Yuliang Liu\\(^{1}\\), Yingying Zhu\\(^{1*}\\), Xiang Bai\\(^{1}\\) \n\\(^{1}\\)Huazhong University of Science and Technology \n{emzhang,yzli,ylliu,yyzhu,xbai} \\(@\\) hust.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.486, + 0.671 + ], + "angle": 0, + "content": "Online education has been widespread in worldwide universities and educational institutions. Lecture slides, a fundamental component of online education, contain a wealth of information, playing a crucial role in learning. However, previous works have not yet paid sufficient attention to understanding lecture slides, including the absence of the large-scale dataset and comprehensive understanding tasks. To facilitate the research about lecture slides understanding, we establish the LecSlides-370K, which consists of 25,542 lectures with 370,078 slides across 15 areas. We also introduce two comprehensive tasks, Lecture Summary and Lecture Question Answering (QA), for providing different perspectives of slides understanding. Furthermore, complex and flexible text relations can hinder the understanding of the internal logic of slides. To address this challenge, we propose a novel method, named SlideParser, which includes an auxiliary branch to predict text relations within slides and enhance attention between related texts, thereby improving slides understanding. With extensive experiments, we show the superiority of our proposed method on both LecSlides-370k and SlideVQA. Dataset and Codes are available at https://github.com/zamling/LecSlides_370K" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.7, + 0.222, + 0.716 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.485, + 0.878 + ], + "angle": 0, + "content": "With the rapid progress in online education, an increasing number of universities and educational institutions have adopted online learning in many platforms [20, 21, 25]. Among diverse online education formats, such as lecture videos and live streams, slides often serve as a key medium for conveying significant knowledge. Consequently, more researchers have turned their attention to this field. However, there are two obstacles to understand lecture slides: 1) Texts in slides are often conclusive and headline-style statements, which complicate the task of providing detailed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.339 + ], + "angle": 0, + "content": "content. 2) The complex text relations pose additional challenges for models attempting to grasp the internal logic and generate reasonable responses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.909, + 0.521 + ], + "angle": 0, + "content": "Recently, several works [4, 5, 7, 14, 17, 27] have noticed the importance of lecture slides and introduced corresponding datasets, such as text detection in slides [5], slide segmentation [7] and image-text retrieval for slides and corresponding speech [14]. However, none of them has yet paid attention on comprehensive understanding slides. SlideVQA [27], a recent slides dataset, provides a Question-Answering (QA) benchmark for slides, yet it still overlooks the global understanding task, such as generating a detailed summary for a complete set of slides. Additionally, the restricted scales and limited number of lecture areas also constrain the development of lecture slides understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.909, + 0.779 + ], + "angle": 0, + "content": "Aiming to address the limitations in the lecture slides dataset, we establish a large-scale and multi-task dataset for comprehensive slides understanding, named LecSlides-370K. Specifically, we collect 25,542 lectures composed of 370,078 slides. These lectures span 15 lecture areas, including business, natural sciences, computer sciences, and so on, significantly surpassing previous datasets in both scale and diversity, as shown in Fig. 1 (a). Furthermore, we introduce two tasks based on our dataset, termed Lecture Summary and Lecture Question Answering (QA). These two tasks represent the most frequently utilized functions for students, offering varied perspectives for slides understanding. Specifically, Lecture Summary focuses on extracting key information and fostering a global understanding, while Lecture QA focuses on locating problems and understanding fine-grained content. Finally, our dataset includes 114,472 summary annotations and 259,461 QA pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Based on LecSlides-370K dataset, we evaluate the existing NLP methods [15, 22], document understanding methods [10, 28, 32, 33] and vision-language models [16, 34]. The results reveal the inferior performance of these methods. We argue that this inferiority can be attributed to the fact that these methods have not yet adequately considered the unique characteristic of lecture slides, that text relations in slides are complex and contain a wealth of logic and rel" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.887, + 0.238, + 0.901 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4455" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.379, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.28, + 0.368, + 0.291 + ], + "angle": 0, + "content": "(a) Data scales of LecSlides-370K and related datasets" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.089, + 0.905, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.28, + 0.783, + 0.291 + ], + "angle": 0, + "content": "(b) An example of LecSlides-370K with various tasks" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.302, + 0.907, + 0.358 + ], + "angle": 0, + "content": "Figure 1. LecSlides-370K is a large-scale and multi-task lecture slides dataset. It contains 25,542 lectures and 370,078 slides across 15 lecture areas, which has around 9.7 times lecture numbers and 7.1 times slide numbers compared with SlideVQA [27]. (a) A comparison between LecSlides-370K and other slides dataset in terms of the number of lectures, slides and lecture areas. The detailed information is shown in Table 1. (b) An example of LecSlides-370K, including Lecture Summary and Lecture QA tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.385, + 0.482, + 0.476 + ], + "angle": 0, + "content": "evance. For instance, as shown in Fig. 1 (b), texts displayed in the upper left corner indicate detailed requirements for forecasting, while texts on the right side of each block explain each stage in the flow chart. Therefore, understanding complex text relations in slides can grasp the internal logic, and is the key part of understanding slides." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.702 + ], + "angle": 0, + "content": "To extract the complex text relations and enhance slides understanding, we propose a novel method called SlideParser. The key of SlideParser is how to effectively predict and utilize complex text relations in slides. Specifically, we propose an auxiliary branch, termed Text Relation Prediction Module, comprising the Relation Prediction Head (RPH) to predict relations among texts and the Relation Attention Layer (RAL) to leverage predicted text relations, enhancing attention between related texts while reducing it between unrelated texts. Additionally, to enhance the quality of predictions, we build a weak supervision framework for RPH with a pseudo relation label generation pipeline, named DL-Cluster, which considers both text spatial and semantic relations, yielding superior text relation labels compared to traditional clustering methods [2, 6, 9]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.839 + ], + "angle": 0, + "content": "In our experiments on LecSlides-370K, related baselines struggle to extract the main point of lecture slides and pinpoint the correct content corresponding to the question, revealing the challenges of our dataset. By effectively predicting and utilizing text relations, SlideParser outperforms all the related methods in the LecSlides-370K Lecture Summary and Lecture QA benchmarks. Additionally, we evaluate SlideParser on the SlideVQA [27] dataset, further validating the effectiveness of our text relation design." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.84, + 0.353, + 0.854 + ], + "angle": 0, + "content": "In conclusion, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We establish a large-scale and multi-task slides understanding dataset, termed LecSlides-370K, comprising 25,542 lectures across 15 lecture areas and 370,078" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.385, + 0.905, + 0.414 + ], + "angle": 0, + "content": "slides. We also introduce two comprehensive tasks, Lecture Summary and Lecture QA." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.416, + 0.905, + 0.461 + ], + "angle": 0, + "content": "- We propose a novel method, SlideParser, including an auxiliary brach to effectively predict and utilize complex text relations in slides, enhancing slides understanding." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.495, + 0.655, + 0.51 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.526, + 0.729, + 0.54 + ], + "angle": 0, + "content": "2.1. Lecture Slides Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.906, + 0.631 + ], + "angle": 0, + "content": "Recently, several datasets based on lecture slides have been proposed, which are most relevant to our current work, including the LectureBank Dataset [17], ALV [7], LectureVideoDB [5], LPM Dataset [14], \\(\\mathbf{M}^3\\mathrm{AV}\\) [4] and SlideVQA [27], as shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.906, + 0.774 + ], + "angle": 0, + "content": "LectureBank [17] is annotated for prerequisite relationships among topics, providing a structure of connected knowledge. ALV dataset [7] primarily focuses on segmenting the entire video lecture. LectureVideoDB is formulated to investigate text detection and recognition in video lectures. Furthermore, LPM [14] attempts to explore the retrieval of spoken words with corresponding visual cues. SlideVQA, a recent work for slides, proposes a Visual Question Answering (VQA) task on it." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In summary, previous datasets contain two primary limitations: 1) Most of datasets do not pay sufficient attention to slides understanding, and the tasks they define also do not provide multiple perspectives for understanding. 2) The limited data scale hinders the advancement of understanding lecture slides. Therefore, a substantial requirement in the research community emerges for a large-scale slides dataset with comprehensive understanding tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4456" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.182, + 0.089, + 0.817, + 0.251 + ], + "angle": 0, + "content": "
DatasetSizeAnnotationTask
LecturesSlidesOCRQASummary
LectureVideoDB [5]245,000✓--Text Detection
LectureBank [17]1,35251,939✓--Prerequisite Chain
ALV Dataset [7]-1,498✗--Slide Segmentation
LPM [14]3349,031✓--Image Retrieval
M3AV [4]111324,956✓--ASR & TTS
SlideVQA [27]2,61952,380✓14,500-QA
LecSlides-370K (Ours)25,542370,078✓259,461114,472QA & Summary
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.261, + 0.908, + 0.292 + ], + "angle": 0, + "content": "Table 1. Data scales, annotations and tasks for different lecture slides datasets. LecSlides-370K contain the largest data scale, the most comprehensive annotations and two understanding tasks to evaluate slides understanding from multiple perspectives." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.315, + 0.363, + 0.332 + ], + "angle": 0, + "content": "2.2. Slides Understanding Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.483, + 0.414 + ], + "angle": 0, + "content": "Although there is a lack of specific methods for slide understanding, other area methods can be used in slide understanding, mainly including three categories: vision-language methods [16, 31, 34, 36], NLP methods [15, 22], and document understanding methods [10, 28]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.415, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Vision-language methods typically begin with large-scale pre-training tasks and then fine-tune on downstream multi-modal understanding tasks. These methods often integrate a vision-language fused transformer encoder [16, 31, 36]. NLP methods provide a plain text solution, such as T5 [22] and BART [15], following an encoder-decoder architecture to process textual contents and generate textual outputs. Furthermore, inspired by NLP approaches, document understanding methods replace traditional transformer encoder to multi-modal encoder, processing and fusing textual and visual content. For example, LayoutLM series [10, 32, 33] integrate text, image features, and 2D positional embeddings in the multi-modal encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.484, + 0.658 + ], + "angle": 0, + "content": "However, these related methods do not consider the complex text relations in slides, a crucial part of slides understanding, thereby yielding inferior performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.674, + 0.315, + 0.69 + ], + "angle": 0, + "content": "3. LecSlides-370K Dataset" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.7, + 0.248, + 0.715 + ], + "angle": 0, + "content": "3.1. Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.483, + 0.875 + ], + "angle": 0, + "content": "To build a large-scale and multi-modal lecture slides dataset, we collect slides from Slideshare*. Our raw data consists of 720,673 slides from 39,784 lectures at the beginning. However, some of these slides are irrelevant to the lecture content, such as advertisements, or contain extraneous information, which is not essential for comprehension. Thus, we filter these slides by their contents. Finally, LecSlides-370K contains 370,078 slides from 25,542 lectures across 15 areas, such as computer science, marketing, medicine, and so on. As shown in Table 1, compared with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.905, + 0.377 + ], + "angle": 0, + "content": "the previous largest datasets, SlideVQA [27], LecSlides-370K has around 9.7 times lecture numbers and 7.1 times slide numbers. The superiority of data scale and diversity renders our dataset more comprehensive and challenging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.907, + 0.605 + ], + "angle": 0, + "content": "Moreover, since the length of these raw lecture slides varies widely, from a few to hundreds, which makes it difficult for researchers to use them, we segment these raw slides into 114,472 lecture chunks, following the principle that each chunk belongs to the same sub-topic. For instance, if one slide initially displays the first point of a list, and the subsequent slide reveals the following points, these slides should belong to the same chunk. The distribution of slides number in each lecture chunk is shown in Fig. 2 (b). Additionally, we annotate the positions of texts, illustrations, and tables in slides to facilitate the study and encourage researchers to focus more on developing new methods rather than information extraction and data processing. Furthermore, we also annotate the textual contents in slides, which can be utilized by the models to understand slides better." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.62, + 0.667, + 0.635 + ], + "angle": 0, + "content": "3.2. Task Definition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We define two tasks: Lecture Summary and Lecture QA. These two tasks not only involve most application scenarios, but also take into account the global and detailed understanding of slides, fully reflecting the understanding ability of models. In the context of Lecture Summary, given a lecture chunk \\(\\mathbf{I} = \\{I_1, I_2, \\dots, I_j\\}^M\\), comprising \\(M\\) slide images, the model's objective is to analyze the content of each slide, integrate contextual information, retain key details, and eliminate redundancies. Finally, the model outputs a summary of all slides. Contrary to the global scope of Lecture Summary, Lecture QA primarily concentrates on locating and comprehending specific content within a lecture chunk. Given a lecture chunk and a query \\(q\\), the model need to identify the slide corresponding to the question and then provide an answer based on the content of that slide. QA serves to assess the model's fine-grained understanding of slides. In total, We annotate 114,472 summary labels and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.258, + 0.901 + ], + "angle": 0, + "content": "\\*https://www.slideshare.net" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4457" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.465, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.089, + 0.672, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.09, + 0.895, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.334, + 0.908, + 0.378 + ], + "angle": 0, + "content": "Figure 2. Examples and statistics of LecSlides-370K. (a) Lecture numbers in each discipline for LecSlides-370K, SlideVQA [27] and LPM [14]. (b) The number of slides for each chunk. (c) The distribution of OCRs in slides. (d) The distribution of summary length. (e) The percentage of each QA type." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.403, + 0.218, + 0.418 + ], + "angle": 0, + "content": "259,461 QA pairs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.426, + 0.363, + 0.442 + ], + "angle": 0, + "content": "3.3. Dataset Statistics and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.483, + 0.479 + ], + "angle": 0, + "content": "In this section, we aim to provide a comprehensive analysis and understanding of the LecSlides-370K dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.485, + 0.258, + 0.5 + ], + "angle": 0, + "content": "3.3.1. Lecture Diversity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.639 + ], + "angle": 0, + "content": "We collect slides from various disciplines, as shown in Fig. 2 (a). There are 15 disciplines, and the highest proportion of data pertains to business, while law and marketing lectures constitute a relatively small percentage. Compared with previous slides datasets, such as SlideVQA [27] and LPM [14], LecSlides-370K contains more disciplines, such as marketing, agriculture, and so on. The diversity of disciplines contains rich teaching content, bringing challenges that models need to handle more diverse scenes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.646, + 0.247, + 0.659 + ], + "angle": 0, + "content": "3.3.2. Professionalism" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.483, + 0.815 + ], + "angle": 0, + "content": "Lecture slides usually contain specialized knowledge and terminologies, bringing more challenges. Therefore, we use the Flesch Reading Ease (FRES) score [13] to evaluate the difficulty of contents and get an average score of 34.3 under this metric. This indicates that the content of our dataset is on 'Hard to Read' level. Moreover, we also use the Flesch-Kincaid Grade Level [13] to reflect which education level our dataset is appropriate for and obtain a 14.1 grade on average. These show there is a professional difficulty in our dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.22, + 0.837 + ], + "angle": 0, + "content": "3.3.3. Slide Image" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For LecSlides-370K, there are 114,472 lecture chunks corresponding to 370,078 slide images. Each chunk contains a varying number of slide images, with an average of 3.2 images per chunk. The distribution of chunks containing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.907, + 0.553 + ], + "angle": 0, + "content": "different numbers of slide images is shown in Fig. 2 (b), and most chunks contain two slide images. Moreover, slide images are rich in visual information. By our count, there are 94,906 diagrams, 34,562 illustrations, and 18,938 tables in LecSlides-370K dataset. We further show the statistics for textual content in slides. We totally annotate 28,220,274 words, with an average of 76.2 words. The size of the vocabulary for our dataset is 84,630. In Fig. 2 (c), we show the distribution of texts across slides, revealing that the majority of slides contain fewer than 100 words." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.563, + 0.767, + 0.575 + ], + "angle": 0, + "content": "3.3.4. Characteristics of Annotation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.747 + ], + "angle": 0, + "content": "For summary annotations, Fig. 2 (d) illustrates the length distribution of the annotations. The average length of the summary is 124.6 words. We further comprehensively analyze QA annotations, including four distinct question types: binary choice, visual analysis, data interpretation and textual reasoning. As shown in Fig. 2 (e), the highest proportion of question types is textual reasoning, while binary choice constitutes the smallest percentage. This diverse set of question types covers the whole slide content, enabling a comprehensive evaluation of the model's QA capabilities. In addition, the average length of answer is 4.1 words." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.704, + 0.772 + ], + "angle": 0, + "content": "3.4. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "LecSlides-370K dataset contains 109,472 slide chunks for the training set, 2,000 chunks for the validation set and 3,000 chunks for the test set. We adapt the ROUGE-1, ROUGE-L [18] and METEOR [3] as evaluation metrics for Lecture Summary task. These metrics are widely used in image caption and text generation, as they consider the quality of generated texts in different aspects. Following SlideVQA [27], we choose EM (Exact Matching) and F1" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4458" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.482, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.25, + 0.378, + 0.262 + ], + "angle": 0, + "content": "(a) Overview of SlideParser Architecture" + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.087, + 0.905, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.249, + 0.828, + 0.261 + ], + "angle": 0, + "content": "(b) Architecture of Text Relation Prediction Module" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.274, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 3. (a) The architecture of our proposed method, SlideParser. It consists of a Transformer Encoder, a Text Relation Prediction Module, and a Text Decoder. (b) The details of Text Relation Prediction Module, which can predict and utilize text relations effectively, including the Relation Prediction Head (RPH) with weak supervision and the Relation Attention Layer (RAL)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.435, + 0.358 + ], + "angle": 0, + "content": "score as evaluation metrics for the Lecture QA task." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.372, + 0.182, + 0.387 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.398, + 0.484, + 0.534 + ], + "angle": 0, + "content": "As shown in Fig. 3 (a), SlideParser consists of a Transformer Encoder, a Text Relation Prediction Module, and a Text Decoder. Texts and image patches from slides are fed into the Transformer Encoder, and the output text embeddings are forwarded to the Text Relation Prediction Module to predict text relations and utilize them to update text embeddings. Finally, updated text embeddings combine with output image embeddings and are fed into the Text Decoder to generate the summary or answer." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.544, + 0.236, + 0.56 + ], + "angle": 0, + "content": "4.1. Input Schema" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.793 + ], + "angle": 0, + "content": "SlideParser is built on a unified text-image multi-modal Transformer Encoder to learn cross-modal representations. Similar to previous document understanding works [10, 28], the Transformer Encoder adopts a multi-layer structure, with each layer primarily comprising multi-head self-attention and feed-forward networks. The input of Transformer Encoder consists of text embeddings \\(\\mathbf{X} = x_{1:M}\\) and image embedding \\(\\mathbf{V} = v_{1:M}\\), where \\(L\\) and \\(M\\) are sequence length of textual and visual embeddings respectively. Specifically, for text embedding \\(\\mathbf{X}\\), \\(L\\) text tokens \\(\\{s_i\\}_{i=1}^L\\) are encoded by word embedding with 1D position and 2D layout position embedding [28]. Meanwhile, the slide image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}\\) is divided into \\(M = \\frac{H}{P} \\times \\frac{H}{P}\\) image patches. These patches are then processed through Patch Embedding to generate the image embeddings, \\(\\mathbf{V}\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.803, + 0.379, + 0.818 + ], + "angle": 0, + "content": "4.2. Text Relation Prediction Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In Transformer Encoder, 2D position embedding, which can be used to comprehend layout and text relations, has been demonstrated in previous document understanding works [10, 28]. However, in the context of slides, text relations tend to be more complex compared to traditional doc" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.907, + 0.403 + ], + "angle": 0, + "content": "uments, such as examples in Fig. 2. Merely incorporating 2D position embeddings into the transformer encoder may not suffice for comprehensive relation comprehension and could potentially introduce inaccuracies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.909, + 0.542 + ], + "angle": 0, + "content": "To address this issue, we introduce the Text Relation Prediction Module, including the Relation Prediction Head (RPH) with weak supervision and the Relation Attention Layer (RAL), as shown in Fig. 3 (b). This module predicts text relations by RPH and utilizes relations to enhance attention between related texts while reducing attention between unrelated texts by RAL. Additionally, to improve the quality of predicted relations, we establish a weak supervision pipeline. We will introduce each block in detail." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.555, + 0.738, + 0.569 + ], + "angle": 0, + "content": "4.2.1. Relation Prediction Head" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.907, + 0.636 + ], + "angle": 0, + "content": "To predict the accurate text relations, inspired by [30], RPH leverages text position and semantic information to predict text relations. Given text bounding boxes \\(\\{b_i\\}_{i=1}^L\\), we define the box delta as:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.648, + 0.905, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Delta \\left(b _ {i}, b _ {j}\\right) = \\left\\{\\left(x _ {i} ^ {c t r} - x _ {j} ^ {c t r}\\right) / w _ {i}, \\left(y _ {i} ^ {c t r} - y _ {j} ^ {c t r}\\right) / h _ {i}, \\right. \\tag {1} \\\\ \\left. \\log \\left(w _ {i} / w _ {j}\\right), \\log \\left(w _ {i} / w _ {j}\\right) \\right\\} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.87 + ], + "angle": 0, + "content": "where \\( x_{i}^{ctr} \\) and \\( y_{i}^{ctr} \\) represent the position of the center point for bounding box, and \\( w \\) and \\( h \\) are the width and height of box respectively. Moreover, we present the spatial relation of two boxes as: \\( \\mathbf{r}_{\\mathbf{bi},\\mathbf{bj}} = \\{\\Delta (b_i,b_j),\\Delta (b_i,b_{ij}),\\Delta (b_{ij},b_j)\\} \\), which is a 12-D vector. Specifically, \\( b_{ij} \\) is a union bounding box for \\( b_{i} \\) and \\( b_{j} \\). Then, we employ an MLP layer and integrate all box relations to get the spatial relation matrix \\( \\mathbf{R}_{\\mathbf{spa}} \\in \\mathbb{R}^{L\\times L} \\). In addition, RPH also considers semantic relations. Given output text embeddings from Transformer Encoder, \\( \\mathbf{E_T} \\in \\mathbb{R}^{L\\times D} \\), we compute the semantic relation between two texts as" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.884, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R _ {s e m}} = \\left(\\mathbf {W _ {q} E _ {T}} \\times \\mathbf {W _ {k} E _ {T} ^ {T}}\\right) / \\sqrt {D} \\qquad (2)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4459" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.889, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.238, + 0.905, + 0.253 + ], + "angle": 0, + "content": "Figure 4. The Pipeline of DL-Cluster to generate pseudo relation label. We combine spatial and semantic text relations to get pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.277, + 0.483, + 0.294 + ], + "angle": 0, + "content": "The predicted relation matrix \\(\\hat{\\mathbf{M}}_{\\mathbf{R}} \\in \\mathbb{R}^{L \\times L}\\) is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.309, + 0.483, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\mathbf {\\hat {M} _ {R}} = \\operatorname {S i g m o i d} \\left(\\mathbf {R} _ {\\mathbf {s p a}} + \\mathbf {R} _ {\\mathbf {s e m}}\\right) \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.336, + 0.385, + 0.351 + ], + "angle": 0, + "content": "4.2.2. Weak Supervised Label Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.484, + 0.491 + ], + "angle": 0, + "content": "In our experiment, we find that the RPH struggles to generate reasonable text relations without the presence of relation supervision. As a solution, we introduce a weak supervision pipeline to enhance the quality of predictions. Previous works [1, 8] rely on traditional cluster methods [6, 9] to predict text relations. However, these traditional methods can not handle diverse layout format in slides effectively. Moreover, our observations suggest that both spatial and semantic relations play crucial roles in slides." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.612 + ], + "angle": 0, + "content": "Thus, we introduce a dual-stream pseudo relation label generation pipeline, termed DL-Cluster (Dynamic Layout Cluster), as shown in Fig. 4. To obtain spatial relations, we first sample \\(N\\) reference points as cluster points for each text box. Because of the diverse nature of text relations in slides, a predefined number of clusters may not be ideal for accurate relation predictions. Hence, we opt for a dynamic approach to determine the cluster number for each slide." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.749 + ], + "angle": 0, + "content": "Specifically, based on Kmeans++ [2], we iterate over the number of cluster \\( k \\) from 2 to \\( \\mathcal{K} \\) and determine the optimal cluster number by Silhouette method [24]. Silhouette analyzes the distances of each cluster point to its own cluster and its closest neighboring cluster. The optimal \\( k \\) is obtained when the average distance of a cluster point to all the other points in its own cluster is minimum, while the average distance of a point to the nearest neighboring cluster's points is maximum. In our experiment, \\( \\mathcal{K} = 8 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.796 + ], + "angle": 0, + "content": "Subsequently, we assign a cluster ID to each box through a voting mechanism among reference points. Then, the spatial relation score \\( s_{spa} \\) is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.808, + 0.483, + 0.854 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {\\mathbf {s p a}} = \\left\\{ \\begin{array}{l} 1 - \\frac {1}{N} \\sum_ {\\varepsilon = 1} ^ {N} \\left\\| x _ {i, \\varepsilon} - \\mu_ {j} \\right\\|, \\text {w h e n} C _ {i} = C _ {j} \\\\ 0, \\text {w h e n} C _ {i} \\neq C _ {j} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\( C_i \\) and \\( C_j \\) are clusters and \\( \\mu_j \\) is cluster centroid of \\( C_j \\). Specifically, scores are assigned based on the average distance between reference points and the cluster centroid" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.905, + 0.322 + ], + "angle": 0, + "content": "when these texts are deemed related and are placed within the same cluster. Conversely, when texts are considered unrelated, the score is set to zero." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.906, + 0.399 + ], + "angle": 0, + "content": "Simultaneously, to obtain semantic relation, we extract textual contents in each text box, and compute the semantic similarity by the cosine distance of Sentence Transformer [23] and get the semantic relation score \\( s_{sem} \\). These two relation scores are combined as" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.411, + 0.905, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} = \\beta_ {1} \\operatorname {N o r m} \\left(\\mathbf {s} _ {\\mathbf {s p a}}\\right) + \\beta_ {2} \\operatorname {N o r m} \\left(\\mathbf {s} _ {\\mathbf {s e m}}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.438, + 0.906, + 0.483 + ], + "angle": 0, + "content": "we set the default values of \\(\\beta_{1} = 0.7\\), \\(\\beta_{2} = 0.3\\). Finally, the relation matrix \\(M_{R} = \\{s_{ij}\\}^{L\\times L}\\) is binarized with 0.5 threshold. Then, the text relation loss can be calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.494, + 0.905, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e l} = B C E L o s s \\left(\\hat {\\mathbf {M}} _ {\\mathbf {R}}, \\mathbf {M} _ {\\mathbf {R}}\\right) \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.523, + 0.736, + 0.538 + ], + "angle": 0, + "content": "4.2.3. Relation Attention Layer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.907, + 0.678 + ], + "angle": 0, + "content": "With the predicted text relations, we utilize relations by Self Attention [29], as shown in Fig. 3 (b). Specifically, for predicted relation matrix \\(\\hat{\\mathbf{M}}_{\\mathbf{R}}\\), we generate attention mask by assigning \\(-\\infty\\) when \\(m_{ij} < 0.5\\), \\(m_{ij} \\in \\hat{\\mathbf{M}}_{\\mathbf{R}}\\). Subsequently, the output text embedding \\(\\mathbf{E}_{\\mathbf{T}}\\) are fed into N-layer self-attention blocks with this attention mask, resulting in the updated embeddings, \\(\\mathbf{E}_{\\mathbf{T}}'\\). It enhances attention between related texts and isolates attention between unrelated texts. Finally, we combine \\(\\mathbf{E}_{\\mathbf{T}}\\) and \\(\\mathbf{E}_{\\mathbf{T}}'\\) by residual connection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.687, + 0.654, + 0.701 + ], + "angle": 0, + "content": "4.3. Text Decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.709, + 0.906, + 0.785 + ], + "angle": 0, + "content": "Following previous work [28], we adapt a T5 decoder as our text decoder, a uni-directional Transformer decoder that performs the next prediction task to generate texts, such as summary or answer. We choose Cross Entropy loss to supervise text generation. The total loss of SlideParser is" + }, + { + "type": "equation", + "bbox": [ + 0.654, + 0.796, + 0.905, + 0.812 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c e} + \\mathcal {L} _ {r e l} \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.646, + 0.841 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.691, + 0.865 + ], + "angle": 0, + "content": "5.1. Implement details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For SlideParser, both the Transformer Encoder and the Text Decoder are initialized from the pre-trained UDOP [28]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4460" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.5, + 0.245 + ], + "angle": 0, + "content": "
Method#Paramsvaltest
R-1R-LMETEORR-1R-LMETEOR
BART-base [15]217M33.121.121.932.820.121.5
BART-large [15]509M35.921.924.235.021.022.9
T5-base [22]272M34.621.622.834.320.722.4
T5-large [22]803M36.222.124.935.121.323.2
Vid2Seq [34]314M16.011.812.917.813.114.1
mPLUG-base [16]350M27.119.021.627.619.321.7
mPLUG-large [16]600M30.020.524.029.820.323.6
Donut [12]543M27.418.621.327.618.521.1
LayoutLMv3-base [10]533M33.621.023.831.719.822.6
LayoutLMv3-large [10]768M34.621.624.833.921.023.2
UDOP [28]794M36.122.126.435.321.525.6
SlideParser (Ours)810M42.225.331.341.724.730.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.256, + 0.473, + 0.27 + ], + "angle": 0, + "content": "Table 2. Performance on Lecture Summary in LecSlides-370K" + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.287, + 0.476, + 0.472 + ], + "angle": 0, + "content": "
Method#Paramsvaltest
EMF1EMF1
BART-base [15]217M18.925.118.524.7
BART-large [15]509M20.627.119.826.3
T5-base [22]272M19.126.018.725.3
T5-large [22]803M20.927.420.027.1
Vid2Seq [34]314M7.014.36.713.6
mPLUG-base [16]350M8.410.98.510.1
mPLUG-large [16]600M9.514.39.514.2
LayoutLMv3-base [10]533M12.515.812.415.1
LayoutLMv3-large [10]768M13.619.313.519.1
UDOP [28]794M19.226.018.725.6
SlideParser (Ours)810M23.733.023.632.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.119, + 0.483, + 0.455, + 0.497 + ], + "angle": 0, + "content": "Table 3. Performance on Lecture QA in LecSlides-370K" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.483, + 0.617 + ], + "angle": 0, + "content": "model weights. We adapt AdamW [19] optimizer with \\(\\beta_{1}\\) of 0.9, the \\(\\beta_{2}\\) of 0.999, and the weight decay of 0.02. We use the cosine learning rate scheduler and the initial learning rate is \\(4\\times 10^{-5}\\). The total number of iterations is 41K. Moreover, we choose a simple concat operator for encoder outputs to handle multiple slide situations for all models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.618, + 0.483, + 0.755 + ], + "angle": 0, + "content": "To provide a full view of LecSlides-370K, we evaluate other related methods from various domains: NLP models such as T5 [22] and BART [15], vision-language models such as mPLUG [16] and Vid2Seq [34], and document understanding models such as LayoutLMv3 [10], Donut [12] and UDOP [28]. Specifically, LayoutLMv3 lacks a text decoder, which is unable to handle text generation tasks. To address this limitation, we incorporate the same text decoder as our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.767, + 0.185, + 0.781 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.79, + 0.363, + 0.805 + ], + "angle": 0, + "content": "5.2.1. Performance on LecSlides-370K" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "At the beginning, we analyze Lecture Summary task in LecSlides-370K. As shown in Table 2, SlideParser achieves notable scores of 42.2 in ROUGE-1, 25.3 in ROUGE-L, and 31.3 in METEOR. This performance surpasses other baseline methods, attributed to its effective understanding text relations. Specifically, the methods without processing tex" + }, + { + "type": "table", + "bbox": [ + 0.564, + 0.089, + 0.858, + 0.245 + ], + "angle": 0, + "content": "
Model Namedevtest
EMF1EMF1
VinVL [36]9.411.410.713.5
PreasM [35]36.341.930.738.2
T5 [22]35.241.331.039.7
LayoutT5 [26]38.944.831.739.9
LayoutLMv2 [33]26.533.421.429.3
FiD [11]37.642.930.438.9
M3D [27]41.347.133.541.7
SlideParser (Ours)44.149.837.543.9
SlideParser (Ours)†49.853.942.447.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.256, + 0.905, + 0.284 + ], + "angle": 0, + "content": "Table 4. Performance on SlideVQA benchmark. † represents pretraining on LecSlides-370K and fine-tuning on SlideVQA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.905, + 0.417 + ], + "angle": 0, + "content": "tual modality, including Vid2Seq [34], mPLUG series [16] and Donut [12], exhibit the worst results. We argue that only processing visual modality may lead to a lack of detailed content extraction in slides. In addition, NLP methods, such as T5 [22] and BART [15], perform relatively well but still lag behind our proposed method, containing excellent multi-modal fusion and text relation utilization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.906, + 0.539 + ], + "angle": 0, + "content": "In contrast to Lecture Summary, which primarily evaluates global understanding of slides, Lecture QA task measures methods for in-depth understanding of specific content. As shown in Table 3, SlideParser achieves superior performance compared to other baselines. Furthermore, visual-language models such as mPLUG and Vid2Seq still perform the worst among all baselines. It shows that detailed text information is still crucial for QA task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.746, + 0.563 + ], + "angle": 0, + "content": "5.2.2. Performance on SlideVQA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.906, + 0.734 + ], + "angle": 0, + "content": "To further evaluate our proposed method in QA task, we also fine-tune and evaluate SlideParser on SlideVQA, a VQA benchmark in multi-image slides. As shown in Table 4, SlideParser achieves the best performance compared with other methods, which further demonstrates the effective designs of SlideParser. In an extensive experiment, we pre-train SlidePraser in our dataset and fine-tune on SlideVQA, denoted as SlideParser†. Results show that largescale and comprehensive LecSlides-370K yield a great performance enhancement, demonstrating our dataset can also serve as a valuable large-scale pre-training dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.744, + 0.667, + 0.76 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.905, + 0.797 + ], + "angle": 0, + "content": "To gain a deeper analysis of SlideParser, we conduct ablation experiments in LecSlides-370K Lecture Summary task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.806, + 0.854, + 0.82 + ], + "angle": 0, + "content": "5.3.1. Effect of Text Relation Prediction Module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We first evaluate the impact of incorporating the Text Relation Prediction Module. As shown in Table 5, results demonstrate the removal of the entire Text Relation Prediction Module (in r1) yields the poorest performance. Furthermore, as we add RPH and RAL to predict and utilize rela" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4461" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.088, + 0.451, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.214, + 0.457, + 0.228 + ], + "angle": 0, + "content": "Figure 5. Visualization of predicted text relations in RPH" + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.244, + 0.462, + 0.324 + ], + "angle": 0, + "content": "
RPHRALPseudo LabelR-1R-LMETEOR
r1XXX36.322.626.9
r2✓✓X37.523.227.7
r3X✓✓41.825.030.8
r4✓✓✓42.225.331.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.335, + 0.448, + 0.348 + ], + "angle": 0, + "content": "Table 5. Ablation on Text Relation Prediction Module" + }, + { + "type": "table", + "bbox": [ + 0.168, + 0.364, + 0.408, + 0.456 + ], + "angle": 0, + "content": "
Cluster TypeR-1R-LMETEOR
Kmeans (k=3)40.924.829.8
Kmeans (k=4)40.724.729.6
HAC(n=3)41.024.830.1
DBSCAN41.324.830.6
DL-Cluster42.225.331.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.467, + 0.468, + 0.481 + ], + "angle": 0, + "content": "Table 6. Ablation on cluster types in pseudo label generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.675 + ], + "angle": 0, + "content": "tions without any supervision (in r2), the performance only has a slight improvement. It demonstrates the RPH struggles to converge effectively in generating reasonable text relations. In addition, we also explore the approach of directly feeding pseudo relation labels generated by DL-Cluster into the subsequent RAL, instead of utilizing learnable relations (in r3). Although it exhibits substantial improvement, yet it continues to fall short in comparison to leveraging learnable text predictions and establishing weak supervision (r3 vs r4). It shows that learnable relations could be refined during the training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.676, + 0.483, + 0.737 + ], + "angle": 0, + "content": "Moreover, we provide visualization results in Fig. 5. We assign the same color for two texts if their relation score is larger than 0.5. The results illustrate that the RPH can predict reasonable text relations under weak supervision." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.746, + 0.474, + 0.76 + ], + "angle": 0, + "content": "5.3.2. Different Methods for Pseudo Label Generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Traditional cluster methods are widely used in previous works for relation prediction [1, 8]. Therefore, we compare our proposed DL-Cluster with other cluster methods. As shown in Table 6, DL-Cluster surpasses clustering techniques with predefined cluster numbers like K-means and Hierarchical Agglomerating Clustering (HAC). Additionally, DL-Cluster outperforms methods without specified cluster numbers, such as DBSCAN [6], due to its capacity to consider spatial and semantic aspects concurrently." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.909, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.208, + 0.898, + 0.222 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of Lecture QA in LecSlides-370K." + }, + { + "type": "table", + "bbox": [ + 0.578, + 0.244, + 0.844, + 0.31 + ], + "angle": 0, + "content": "
SpatialSemanticR-1R-LMETEOR
×✓38.323.728.5
✓×41.424.830.7
✓✓42.225.331.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.58, + 0.321, + 0.84, + 0.335 + ], + "angle": 0, + "content": "Table 7. Ablation on designs of DL-Cluster" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.37, + 0.71, + 0.385 + ], + "angle": 0, + "content": "5.3.3. Design of DL-Cluster." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.499 + ], + "angle": 0, + "content": "We further investigate the impact of text relations represented in spatial and semantic dimensions. As shown in Tab. 7, removing either of these dimensions leads to a decline in performance, and spatial relations have a larger impact. The best results are achieved when both aspects were used simultaneously, indicating that both dimensions complement each other to enhance overall performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.519, + 0.651, + 0.533 + ], + "angle": 0, + "content": "5.4. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.71 + ], + "angle": 0, + "content": "In a side-by-side qualitative analysis for Lecture QA in LecSlides-370K, we compare our proposed method with other baselines in related fields. As shown in Fig. 6, results show that other methods face two challenges: 1) They can not grasp correct internal logic due to the complex text relations. 2) As the number of diagrams on slides increases, other methods struggle to pinpoint the correct diagram corresponding to the question. In contrast, SlideParser excels in accurate responses by comprehending text relations effectively and easily determining the answer's location within the slides." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.736, + 0.634, + 0.751 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.9 + ], + "angle": 0, + "content": "In this work, we introduce a large-scale and multi-task lecture slides dataset, LecSlides-370K, which has a larger data scale and more diverse disciplines. Furthermore, considering the pivotal role of text relation prediction and utilization for slides understanding, we propose SlideParser. The results demonstrate the effect of enhanced text relation design, and our method outperforms other models. We believe LecSlides-370K and SlideParser can prompt lecture slides research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4462" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.148 + ], + "angle": 0, + "content": "This work was supported by the NSFC (62206103 and 62225603)" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.163, + 0.188, + 0.179 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.188, + 0.484, + 0.285 + ], + "angle": 0, + "content": "[1] Rhys Agombar, Max Luebbering, and Rafet Sifa. A clustering backed deep learning approach for document layout analysis. In Machine Learning and Knowledge Extraction: 4th IFIP TC 5, TC 12, WG 8.4, WG 8.9, WG 12.9 International Cross-Domain Conference, CD-MAKE 2020, Dublin, Ireland, August 25–28, 2020, Proceedings 4, pages 423–430. Springer, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.288, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[2] David Arthur and Sergei Vassilvitskii. k-means++: the advantages of careful seeding. In Society for Industrial and Applied Mathematics, page 1027-1035, 2007. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.332, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[4] Zhe Chen, Heyang Liu, Wenyi Yu, Guangzhi Sun, Hongcheng Liu, Ji Wu, Chao Zhang, Yu Wang, and Yanfeng Wang. M3av: A multimodal, multigenre, and multipurpose audio-visual academic lecture dataset. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 9041–9060, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.484, + 0.571 + ], + "angle": 0, + "content": "[5] Kartik Dutta, Minesh Mathew, Praveen Krishnan, and CV Jawahar. Localizing and recognizing text in lecture videos. In 2018 16th international conference on frontiers in handwriting recognition (ICFHR), pages 235-240. IEEE, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[6] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, pages 226-231, 1996. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[7] Damianos Galanopoulos and Vasileios Mezaris. Temporal lecture video fragmentation using word embeddings. In MultiMedia Modeling: 25th International Conference, MMM 2019, Thessaloniki, Greece, January 8-11, 2019, Proceedings, Part II 25, pages 254-265. Springer, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[8] Zhangxuan Gu, Changhua Meng, Ke Wang, Jun Lan, Weiqiang Wang, Ming Gu, and Liqing Zhang. Xylayoutlm: Towards layout-aware multimodal networks for visually-rich document understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4583-4592, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[9] Jaekyu Ha, R.M. Haralick, and I.T. Phillips. Recursive x-y cut using bounding boxes of connected components. In Proceedings of 3rd International Conference on Document Analysis and Recognition, pages 952-955, 1995. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[10] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.188, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "International Conference on Multimedia, pages 4083-4091, 2022. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[11] Gautier Izacard and Edouard Grave. Leveraging passage retrieval with generative models for open domain question answering. arXiv preprint arXiv:2007.01282, pages 1-6, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[12] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[13] JP Kincaid. Derivation of new readability formulas (automated readability index, fog count and flesch reading ease formula) for navy enlisted personnel. Chief of Naval Technical Training, pages 1-45, 1975. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[14] Dong Won Lee, Chaitanya Ahuja, Paul Pu Liang, Sanika Natu, and Louis-Philippe Morency. Lecture presentations multimodal dataset: Towards understanding multimodality in educational videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20087-20098, 2023. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[15] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online, 2020. Association for Computational Linguistics. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.634 + ], + "angle": 0, + "content": "[16] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, He Chen, Guohai Xu, Zheng Cao, Ji Zhang, Songfang Huang, Fei Huang, Jingren Zhou, and Luo Si. mPLUG: Effective and efficient vision-language learning by cross-modal skip-connections. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7241–7259, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.636, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[17] Irene Li, Alexander R Fabbri, Robert R Tung, and Dragomir R Radev. What should i learn first: Introducing lecturebank for nlp education and prerequisite chain learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6674-6681, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.706, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[18] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, pages 1-19, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[20] T. Muthuprasad, S. Aiswarya, K.S. Aditya, and Girish K. Jha. Students' perception and preference for online education in india during covid -19 pandemic. Social Sciences Humanities Open, 3(1):100101, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[21] Pitambar Paudel. Online education: Benefits, challenges and strategies during and after Covid-19 in higher education. International Journal on Studies in Education (IJonSE), 3:70–85, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4463" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[22] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[23] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3982-3992. Association for Computational Linguistics, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[24] Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53-65, 1987. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[25] Anna Sun and Xiufang Chen. Online education and its effective practice: A research review. Journal of information technology education: Research, 15:1-34, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[26] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[27] Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. Slidevqa: A dataset for document visual question answering on multiple images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13636-13645, 2023. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[28] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19254-19264, 2023. 1, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 6000-6010, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[30] Jiawei Wang, Kai Hu, Zhuoyao Zhong, Lei Sun, and Qiang Huo. Detect-order-construct: A tree construction based approach for hierarchical document structure analysis. arXiv preprint arXiv:2401.11874, pages 1-35, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[31] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[32] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192-1200, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[33] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. LayoutLMv2: Multi-modal pre-training for visually-rich document understanding. In Proceedings of the 59th Annual Meeting of the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2579-2591, Online, 2021. Association for Computational Linguistics. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[34] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[35] Ori Yoran, Alon Talmor, and Jonathan Berant. Turning tables: Generating examples from semi-structured tables for endowing language models with reasoning skills. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pages 6016–6031. Association for Computational Linguistics, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.318, + 0.905, + 0.4 + ], + "angle": 0, + "content": "[36] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5579-5588, 2021. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4464" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/58e52446-3fb4-48af-8a99-0aa5fc34cde3_origin.pdf b/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/58e52446-3fb4-48af-8a99-0aa5fc34cde3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4c69ef502f5d5b25ce1865ad0f9b6b66dd01946d --- /dev/null +++ b/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/58e52446-3fb4-48af-8a99-0aa5fc34cde3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7bb52168156a77f954ee40638652432dc23eace10333b6fb23095bb9e9528bf3 +size 3219563 diff --git a/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/full.md b/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0dbf0ed2e366a6415d1af1a626173fec42c3c049 --- /dev/null +++ b/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/full.md @@ -0,0 +1,332 @@ +# Towards Comprehensive Lecture Slides Understanding: Large-scale Dataset and Effective Method + +Enming Zhang $^{1}$ , Yuzhe Li $^{1}$ , Yuliang Liu $^{1}$ , Yingying Zhu $^{1*}$ , Xiang Bai $^{1}$ $^{1}$ Huazhong University of Science and Technology +{emzhang,yzli,ylliu,yyzhu,xbai} $@$ hust.edu.cn + +# Abstract + +Online education has been widespread in worldwide universities and educational institutions. Lecture slides, a fundamental component of online education, contain a wealth of information, playing a crucial role in learning. However, previous works have not yet paid sufficient attention to understanding lecture slides, including the absence of the large-scale dataset and comprehensive understanding tasks. To facilitate the research about lecture slides understanding, we establish the LecSlides-370K, which consists of 25,542 lectures with 370,078 slides across 15 areas. We also introduce two comprehensive tasks, Lecture Summary and Lecture Question Answering (QA), for providing different perspectives of slides understanding. Furthermore, complex and flexible text relations can hinder the understanding of the internal logic of slides. To address this challenge, we propose a novel method, named SlideParser, which includes an auxiliary branch to predict text relations within slides and enhance attention between related texts, thereby improving slides understanding. With extensive experiments, we show the superiority of our proposed method on both LecSlides-370k and SlideVQA. Dataset and Codes are available at https://github.com/zamling/LecSlides_370K + +# 1. Introduction + +With the rapid progress in online education, an increasing number of universities and educational institutions have adopted online learning in many platforms [20, 21, 25]. Among diverse online education formats, such as lecture videos and live streams, slides often serve as a key medium for conveying significant knowledge. Consequently, more researchers have turned their attention to this field. However, there are two obstacles to understand lecture slides: 1) Texts in slides are often conclusive and headline-style statements, which complicate the task of providing detailed + +content. 2) The complex text relations pose additional challenges for models attempting to grasp the internal logic and generate reasonable responses. + +Recently, several works [4, 5, 7, 14, 17, 27] have noticed the importance of lecture slides and introduced corresponding datasets, such as text detection in slides [5], slide segmentation [7] and image-text retrieval for slides and corresponding speech [14]. However, none of them has yet paid attention on comprehensive understanding slides. SlideVQA [27], a recent slides dataset, provides a Question-Answering (QA) benchmark for slides, yet it still overlooks the global understanding task, such as generating a detailed summary for a complete set of slides. Additionally, the restricted scales and limited number of lecture areas also constrain the development of lecture slides understanding. + +Aiming to address the limitations in the lecture slides dataset, we establish a large-scale and multi-task dataset for comprehensive slides understanding, named LecSlides-370K. Specifically, we collect 25,542 lectures composed of 370,078 slides. These lectures span 15 lecture areas, including business, natural sciences, computer sciences, and so on, significantly surpassing previous datasets in both scale and diversity, as shown in Fig. 1 (a). Furthermore, we introduce two tasks based on our dataset, termed Lecture Summary and Lecture Question Answering (QA). These two tasks represent the most frequently utilized functions for students, offering varied perspectives for slides understanding. Specifically, Lecture Summary focuses on extracting key information and fostering a global understanding, while Lecture QA focuses on locating problems and understanding fine-grained content. Finally, our dataset includes 114,472 summary annotations and 259,461 QA pairs. + +Based on LecSlides-370K dataset, we evaluate the existing NLP methods [15, 22], document understanding methods [10, 28, 32, 33] and vision-language models [16, 34]. The results reveal the inferior performance of these methods. We argue that this inferiority can be attributed to the fact that these methods have not yet adequately considered the unique characteristic of lecture slides, that text relations in slides are complex and contain a wealth of logic and rel + +![](images/dce8870acaf00311263f9d70cd056d989c5a01c24e1226f0bc702777ffd06985.jpg) +(a) Data scales of LecSlides-370K and related datasets + +![](images/2578ba197cded1e26310be408419f364f44f6ba4867fb178fdbd4b55ed45c5b8.jpg) +(b) An example of LecSlides-370K with various tasks +Figure 1. LecSlides-370K is a large-scale and multi-task lecture slides dataset. It contains 25,542 lectures and 370,078 slides across 15 lecture areas, which has around 9.7 times lecture numbers and 7.1 times slide numbers compared with SlideVQA [27]. (a) A comparison between LecSlides-370K and other slides dataset in terms of the number of lectures, slides and lecture areas. The detailed information is shown in Table 1. (b) An example of LecSlides-370K, including Lecture Summary and Lecture QA tasks. + +evance. For instance, as shown in Fig. 1 (b), texts displayed in the upper left corner indicate detailed requirements for forecasting, while texts on the right side of each block explain each stage in the flow chart. Therefore, understanding complex text relations in slides can grasp the internal logic, and is the key part of understanding slides. + +To extract the complex text relations and enhance slides understanding, we propose a novel method called SlideParser. The key of SlideParser is how to effectively predict and utilize complex text relations in slides. Specifically, we propose an auxiliary branch, termed Text Relation Prediction Module, comprising the Relation Prediction Head (RPH) to predict relations among texts and the Relation Attention Layer (RAL) to leverage predicted text relations, enhancing attention between related texts while reducing it between unrelated texts. Additionally, to enhance the quality of predictions, we build a weak supervision framework for RPH with a pseudo relation label generation pipeline, named DL-Cluster, which considers both text spatial and semantic relations, yielding superior text relation labels compared to traditional clustering methods [2, 6, 9]. + +In our experiments on LecSlides-370K, related baselines struggle to extract the main point of lecture slides and pinpoint the correct content corresponding to the question, revealing the challenges of our dataset. By effectively predicting and utilizing text relations, SlideParser outperforms all the related methods in the LecSlides-370K Lecture Summary and Lecture QA benchmarks. Additionally, we evaluate SlideParser on the SlideVQA [27] dataset, further validating the effectiveness of our text relation design. + +In conclusion, our contributions are: + +- We establish a large-scale and multi-task slides understanding dataset, termed LecSlides-370K, comprising 25,542 lectures across 15 lecture areas and 370,078 + +slides. We also introduce two comprehensive tasks, Lecture Summary and Lecture QA. + +- We propose a novel method, SlideParser, including an auxiliary brach to effectively predict and utilize complex text relations in slides, enhancing slides understanding. + +# 2. Related Work + +# 2.1. Lecture Slides Datasets + +Recently, several datasets based on lecture slides have been proposed, which are most relevant to our current work, including the LectureBank Dataset [17], ALV [7], LectureVideoDB [5], LPM Dataset [14], $\mathbf{M}^3\mathrm{AV}$ [4] and SlideVQA [27], as shown in Table 1. + +LectureBank [17] is annotated for prerequisite relationships among topics, providing a structure of connected knowledge. ALV dataset [7] primarily focuses on segmenting the entire video lecture. LectureVideoDB is formulated to investigate text detection and recognition in video lectures. Furthermore, LPM [14] attempts to explore the retrieval of spoken words with corresponding visual cues. SlideVQA, a recent work for slides, proposes a Visual Question Answering (VQA) task on it. + +In summary, previous datasets contain two primary limitations: 1) Most of datasets do not pay sufficient attention to slides understanding, and the tasks they define also do not provide multiple perspectives for understanding. 2) The limited data scale hinders the advancement of understanding lecture slides. Therefore, a substantial requirement in the research community emerges for a large-scale slides dataset with comprehensive understanding tasks. + +
DatasetSizeAnnotationTask
LecturesSlidesOCRQASummary
LectureVideoDB [5]245,000✓--Text Detection
LectureBank [17]1,35251,939✓--Prerequisite Chain
ALV Dataset [7]-1,498✗--Slide Segmentation
LPM [14]3349,031✓--Image Retrieval
M3AV [4]111324,956✓--ASR & TTS
SlideVQA [27]2,61952,380✓14,500-QA
LecSlides-370K (Ours)25,542370,078✓259,461114,472QA & Summary
+ +Table 1. Data scales, annotations and tasks for different lecture slides datasets. LecSlides-370K contain the largest data scale, the most comprehensive annotations and two understanding tasks to evaluate slides understanding from multiple perspectives. + +# 2.2. Slides Understanding Methods + +Although there is a lack of specific methods for slide understanding, other area methods can be used in slide understanding, mainly including three categories: vision-language methods [16, 31, 34, 36], NLP methods [15, 22], and document understanding methods [10, 28]. + +Vision-language methods typically begin with large-scale pre-training tasks and then fine-tune on downstream multi-modal understanding tasks. These methods often integrate a vision-language fused transformer encoder [16, 31, 36]. NLP methods provide a plain text solution, such as T5 [22] and BART [15], following an encoder-decoder architecture to process textual contents and generate textual outputs. Furthermore, inspired by NLP approaches, document understanding methods replace traditional transformer encoder to multi-modal encoder, processing and fusing textual and visual content. For example, LayoutLM series [10, 32, 33] integrate text, image features, and 2D positional embeddings in the multi-modal encoder. + +However, these related methods do not consider the complex text relations in slides, a crucial part of slides understanding, thereby yielding inferior performance. + +# 3. LecSlides-370K Dataset + +# 3.1. Data Collection + +To build a large-scale and multi-modal lecture slides dataset, we collect slides from Slideshare*. Our raw data consists of 720,673 slides from 39,784 lectures at the beginning. However, some of these slides are irrelevant to the lecture content, such as advertisements, or contain extraneous information, which is not essential for comprehension. Thus, we filter these slides by their contents. Finally, LecSlides-370K contains 370,078 slides from 25,542 lectures across 15 areas, such as computer science, marketing, medicine, and so on. As shown in Table 1, compared with + +the previous largest datasets, SlideVQA [27], LecSlides-370K has around 9.7 times lecture numbers and 7.1 times slide numbers. The superiority of data scale and diversity renders our dataset more comprehensive and challenging. + +Moreover, since the length of these raw lecture slides varies widely, from a few to hundreds, which makes it difficult for researchers to use them, we segment these raw slides into 114,472 lecture chunks, following the principle that each chunk belongs to the same sub-topic. For instance, if one slide initially displays the first point of a list, and the subsequent slide reveals the following points, these slides should belong to the same chunk. The distribution of slides number in each lecture chunk is shown in Fig. 2 (b). Additionally, we annotate the positions of texts, illustrations, and tables in slides to facilitate the study and encourage researchers to focus more on developing new methods rather than information extraction and data processing. Furthermore, we also annotate the textual contents in slides, which can be utilized by the models to understand slides better. + +# 3.2. Task Definition + +We define two tasks: Lecture Summary and Lecture QA. These two tasks not only involve most application scenarios, but also take into account the global and detailed understanding of slides, fully reflecting the understanding ability of models. In the context of Lecture Summary, given a lecture chunk $\mathbf{I} = \{I_1, I_2, \dots, I_j\}^M$ , comprising $M$ slide images, the model's objective is to analyze the content of each slide, integrate contextual information, retain key details, and eliminate redundancies. Finally, the model outputs a summary of all slides. Contrary to the global scope of Lecture Summary, Lecture QA primarily concentrates on locating and comprehending specific content within a lecture chunk. Given a lecture chunk and a query $q$ , the model need to identify the slide corresponding to the question and then provide an answer based on the content of that slide. QA serves to assess the model's fine-grained understanding of slides. In total, We annotate 114,472 summary labels and + +![](images/c08e6c9169570e6bf5c56f30f5c5dade21740a5eb4f6d5d9528b09bb1990e4e1.jpg) +Figure 2. Examples and statistics of LecSlides-370K. (a) Lecture numbers in each discipline for LecSlides-370K, SlideVQA [27] and LPM [14]. (b) The number of slides for each chunk. (c) The distribution of OCRs in slides. (d) The distribution of summary length. (e) The percentage of each QA type. + +![](images/fd1f0a697e404d64196040d4d3fdefd28389d04d2e5bdf52a8c1af4528406c87.jpg) + +![](images/c8a26ff8849eee5283b2dda52594a9195039edf05fb45c86cd5a43a438a64a6b.jpg) + +259,461 QA pairs. + +# 3.3. Dataset Statistics and Analysis + +In this section, we aim to provide a comprehensive analysis and understanding of the LecSlides-370K dataset. + +# 3.3.1. Lecture Diversity + +We collect slides from various disciplines, as shown in Fig. 2 (a). There are 15 disciplines, and the highest proportion of data pertains to business, while law and marketing lectures constitute a relatively small percentage. Compared with previous slides datasets, such as SlideVQA [27] and LPM [14], LecSlides-370K contains more disciplines, such as marketing, agriculture, and so on. The diversity of disciplines contains rich teaching content, bringing challenges that models need to handle more diverse scenes. + +# 3.3.2. Professionalism + +Lecture slides usually contain specialized knowledge and terminologies, bringing more challenges. Therefore, we use the Flesch Reading Ease (FRES) score [13] to evaluate the difficulty of contents and get an average score of 34.3 under this metric. This indicates that the content of our dataset is on 'Hard to Read' level. Moreover, we also use the Flesch-Kincaid Grade Level [13] to reflect which education level our dataset is appropriate for and obtain a 14.1 grade on average. These show there is a professional difficulty in our dataset. + +# 3.3.3. Slide Image + +For LecSlides-370K, there are 114,472 lecture chunks corresponding to 370,078 slide images. Each chunk contains a varying number of slide images, with an average of 3.2 images per chunk. The distribution of chunks containing + +different numbers of slide images is shown in Fig. 2 (b), and most chunks contain two slide images. Moreover, slide images are rich in visual information. By our count, there are 94,906 diagrams, 34,562 illustrations, and 18,938 tables in LecSlides-370K dataset. We further show the statistics for textual content in slides. We totally annotate 28,220,274 words, with an average of 76.2 words. The size of the vocabulary for our dataset is 84,630. In Fig. 2 (c), we show the distribution of texts across slides, revealing that the majority of slides contain fewer than 100 words. + +# 3.3.4. Characteristics of Annotation + +For summary annotations, Fig. 2 (d) illustrates the length distribution of the annotations. The average length of the summary is 124.6 words. We further comprehensively analyze QA annotations, including four distinct question types: binary choice, visual analysis, data interpretation and textual reasoning. As shown in Fig. 2 (e), the highest proportion of question types is textual reasoning, while binary choice constitutes the smallest percentage. This diverse set of question types covers the whole slide content, enabling a comprehensive evaluation of the model's QA capabilities. In addition, the average length of answer is 4.1 words. + +# 3.4. Evaluation Protocol + +LecSlides-370K dataset contains 109,472 slide chunks for the training set, 2,000 chunks for the validation set and 3,000 chunks for the test set. We adapt the ROUGE-1, ROUGE-L [18] and METEOR [3] as evaluation metrics for Lecture Summary task. These metrics are widely used in image caption and text generation, as they consider the quality of generated texts in different aspects. Following SlideVQA [27], we choose EM (Exact Matching) and F1 + +![](images/77a82f29de098fe3925f47adbf0e1703f41e4c8cb5372a4d86334600716c29b8.jpg) +(a) Overview of SlideParser Architecture + +![](images/6174a25993604cd65c0b87875edfefe565c05f5177a7a3b0c0137c1e8dfe9e76.jpg) +(b) Architecture of Text Relation Prediction Module +Figure 3. (a) The architecture of our proposed method, SlideParser. It consists of a Transformer Encoder, a Text Relation Prediction Module, and a Text Decoder. (b) The details of Text Relation Prediction Module, which can predict and utilize text relations effectively, including the Relation Prediction Head (RPH) with weak supervision and the Relation Attention Layer (RAL). + +score as evaluation metrics for the Lecture QA task. + +# 4. Method + +As shown in Fig. 3 (a), SlideParser consists of a Transformer Encoder, a Text Relation Prediction Module, and a Text Decoder. Texts and image patches from slides are fed into the Transformer Encoder, and the output text embeddings are forwarded to the Text Relation Prediction Module to predict text relations and utilize them to update text embeddings. Finally, updated text embeddings combine with output image embeddings and are fed into the Text Decoder to generate the summary or answer. + +# 4.1. Input Schema + +SlideParser is built on a unified text-image multi-modal Transformer Encoder to learn cross-modal representations. Similar to previous document understanding works [10, 28], the Transformer Encoder adopts a multi-layer structure, with each layer primarily comprising multi-head self-attention and feed-forward networks. The input of Transformer Encoder consists of text embeddings $\mathbf{X} = x_{1:M}$ and image embedding $\mathbf{V} = v_{1:M}$ , where $L$ and $M$ are sequence length of textual and visual embeddings respectively. Specifically, for text embedding $\mathbf{X}$ , $L$ text tokens $\{s_i\}_{i=1}^L$ are encoded by word embedding with 1D position and 2D layout position embedding [28]. Meanwhile, the slide image $\mathbf{I} \in \mathbb{R}^{H \times W \times C}$ is divided into $M = \frac{H}{P} \times \frac{H}{P}$ image patches. These patches are then processed through Patch Embedding to generate the image embeddings, $\mathbf{V}$ . + +# 4.2. Text Relation Prediction Module + +In Transformer Encoder, 2D position embedding, which can be used to comprehend layout and text relations, has been demonstrated in previous document understanding works [10, 28]. However, in the context of slides, text relations tend to be more complex compared to traditional doc + +uments, such as examples in Fig. 2. Merely incorporating 2D position embeddings into the transformer encoder may not suffice for comprehensive relation comprehension and could potentially introduce inaccuracies. + +To address this issue, we introduce the Text Relation Prediction Module, including the Relation Prediction Head (RPH) with weak supervision and the Relation Attention Layer (RAL), as shown in Fig. 3 (b). This module predicts text relations by RPH and utilizes relations to enhance attention between related texts while reducing attention between unrelated texts by RAL. Additionally, to improve the quality of predicted relations, we establish a weak supervision pipeline. We will introduce each block in detail. + +# 4.2.1. Relation Prediction Head + +To predict the accurate text relations, inspired by [30], RPH leverages text position and semantic information to predict text relations. Given text bounding boxes $\{b_i\}_{i=1}^L$ , we define the box delta as: + +$$ +\begin{array}{l} \Delta \left(b _ {i}, b _ {j}\right) = \left\{\left(x _ {i} ^ {c t r} - x _ {j} ^ {c t r}\right) / w _ {i}, \left(y _ {i} ^ {c t r} - y _ {j} ^ {c t r}\right) / h _ {i}, \right. \tag {1} \\ \left. \log \left(w _ {i} / w _ {j}\right), \log \left(w _ {i} / w _ {j}\right) \right\} \\ \end{array} +$$ + +where $x_{i}^{ctr}$ and $y_{i}^{ctr}$ represent the position of the center point for bounding box, and $w$ and $h$ are the width and height of box respectively. Moreover, we present the spatial relation of two boxes as: $\mathbf{r}_{\mathbf{bi},\mathbf{bj}} = \{\Delta (b_i,b_j),\Delta (b_i,b_{ij}),\Delta (b_{ij},b_j)\}$ , which is a 12-D vector. Specifically, $b_{ij}$ is a union bounding box for $b_{i}$ and $b_{j}$ . Then, we employ an MLP layer and integrate all box relations to get the spatial relation matrix $\mathbf{R}_{\mathbf{spa}} \in \mathbb{R}^{L\times L}$ . In addition, RPH also considers semantic relations. Given output text embeddings from Transformer Encoder, $\mathbf{E_T} \in \mathbb{R}^{L\times D}$ , we compute the semantic relation between two texts as + +$$ +\mathbf {R _ {s e m}} = \left(\mathbf {W _ {q} E _ {T}} \times \mathbf {W _ {k} E _ {T} ^ {T}}\right) / \sqrt {D} \qquad (2) +$$ + +![](images/005dc09c7e3e684565abd43a5b65b6a987075f72cf5ee7b14d246c37f9095602.jpg) +Figure 4. The Pipeline of DL-Cluster to generate pseudo relation label. We combine spatial and semantic text relations to get pseudo labels. + +The predicted relation matrix $\hat{\mathbf{M}}_{\mathbf{R}} \in \mathbb{R}^{L \times L}$ is computed as + +$$ +\mathbf {\hat {M} _ {R}} = \operatorname {S i g m o i d} \left(\mathbf {R} _ {\mathbf {s p a}} + \mathbf {R} _ {\mathbf {s e m}}\right) \tag {3} +$$ + +# 4.2.2. Weak Supervised Label Generation + +In our experiment, we find that the RPH struggles to generate reasonable text relations without the presence of relation supervision. As a solution, we introduce a weak supervision pipeline to enhance the quality of predictions. Previous works [1, 8] rely on traditional cluster methods [6, 9] to predict text relations. However, these traditional methods can not handle diverse layout format in slides effectively. Moreover, our observations suggest that both spatial and semantic relations play crucial roles in slides. + +Thus, we introduce a dual-stream pseudo relation label generation pipeline, termed DL-Cluster (Dynamic Layout Cluster), as shown in Fig. 4. To obtain spatial relations, we first sample $N$ reference points as cluster points for each text box. Because of the diverse nature of text relations in slides, a predefined number of clusters may not be ideal for accurate relation predictions. Hence, we opt for a dynamic approach to determine the cluster number for each slide. + +Specifically, based on Kmeans++ [2], we iterate over the number of cluster $k$ from 2 to $\mathcal{K}$ and determine the optimal cluster number by Silhouette method [24]. Silhouette analyzes the distances of each cluster point to its own cluster and its closest neighboring cluster. The optimal $k$ is obtained when the average distance of a cluster point to all the other points in its own cluster is minimum, while the average distance of a point to the nearest neighboring cluster's points is maximum. In our experiment, $\mathcal{K} = 8$ . + +Subsequently, we assign a cluster ID to each box through a voting mechanism among reference points. Then, the spatial relation score $s_{spa}$ is computed as + +$$ +\mathbf {s} _ {\mathbf {s p a}} = \left\{ \begin{array}{l} 1 - \frac {1}{N} \sum_ {\varepsilon = 1} ^ {N} \left\| x _ {i, \varepsilon} - \mu_ {j} \right\|, \text {w h e n} C _ {i} = C _ {j} \\ 0, \text {w h e n} C _ {i} \neq C _ {j} \end{array} \right. \tag {4} +$$ + +where $C_i$ and $C_j$ are clusters and $\mu_j$ is cluster centroid of $C_j$ . Specifically, scores are assigned based on the average distance between reference points and the cluster centroid + +when these texts are deemed related and are placed within the same cluster. Conversely, when texts are considered unrelated, the score is set to zero. + +Simultaneously, to obtain semantic relation, we extract textual contents in each text box, and compute the semantic similarity by the cosine distance of Sentence Transformer [23] and get the semantic relation score $s_{sem}$ . These two relation scores are combined as + +$$ +\mathbf {s} = \beta_ {1} \operatorname {N o r m} \left(\mathbf {s} _ {\mathbf {s p a}}\right) + \beta_ {2} \operatorname {N o r m} \left(\mathbf {s} _ {\mathbf {s e m}}\right) \tag {5} +$$ + +we set the default values of $\beta_{1} = 0.7$ , $\beta_{2} = 0.3$ . Finally, the relation matrix $M_{R} = \{s_{ij}\}^{L\times L}$ is binarized with 0.5 threshold. Then, the text relation loss can be calculated as + +$$ +\mathcal {L} _ {r e l} = B C E L o s s \left(\hat {\mathbf {M}} _ {\mathbf {R}}, \mathbf {M} _ {\mathbf {R}}\right) \tag {6} +$$ + +# 4.2.3. Relation Attention Layer + +With the predicted text relations, we utilize relations by Self Attention [29], as shown in Fig. 3 (b). Specifically, for predicted relation matrix $\hat{\mathbf{M}}_{\mathbf{R}}$ , we generate attention mask by assigning $-\infty$ when $m_{ij} < 0.5$ , $m_{ij} \in \hat{\mathbf{M}}_{\mathbf{R}}$ . Subsequently, the output text embedding $\mathbf{E}_{\mathbf{T}}$ are fed into N-layer self-attention blocks with this attention mask, resulting in the updated embeddings, $\mathbf{E}_{\mathbf{T}}'$ . It enhances attention between related texts and isolates attention between unrelated texts. Finally, we combine $\mathbf{E}_{\mathbf{T}}$ and $\mathbf{E}_{\mathbf{T}}'$ by residual connection. + +# 4.3. Text Decoder + +Following previous work [28], we adapt a T5 decoder as our text decoder, a uni-directional Transformer decoder that performs the next prediction task to generate texts, such as summary or answer. We choose Cross Entropy loss to supervise text generation. The total loss of SlideParser is + +$$ +\mathcal {L} = \mathcal {L} _ {c e} + \mathcal {L} _ {r e l} \tag {7} +$$ + +# 5. Experiments + +# 5.1. Implement details + +For SlideParser, both the Transformer Encoder and the Text Decoder are initialized from the pre-trained UDOP [28] + +
Method#Paramsvaltest
R-1R-LMETEORR-1R-LMETEOR
BART-base [15]217M33.121.121.932.820.121.5
BART-large [15]509M35.921.924.235.021.022.9
T5-base [22]272M34.621.622.834.320.722.4
T5-large [22]803M36.222.124.935.121.323.2
Vid2Seq [34]314M16.011.812.917.813.114.1
mPLUG-base [16]350M27.119.021.627.619.321.7
mPLUG-large [16]600M30.020.524.029.820.323.6
Donut [12]543M27.418.621.327.618.521.1
LayoutLMv3-base [10]533M33.621.023.831.719.822.6
LayoutLMv3-large [10]768M34.621.624.833.921.023.2
UDOP [28]794M36.122.126.435.321.525.6
SlideParser (Ours)810M42.225.331.341.724.730.4
+ +Table 2. Performance on Lecture Summary in LecSlides-370K + +
Method#Paramsvaltest
EMF1EMF1
BART-base [15]217M18.925.118.524.7
BART-large [15]509M20.627.119.826.3
T5-base [22]272M19.126.018.725.3
T5-large [22]803M20.927.420.027.1
Vid2Seq [34]314M7.014.36.713.6
mPLUG-base [16]350M8.410.98.510.1
mPLUG-large [16]600M9.514.39.514.2
LayoutLMv3-base [10]533M12.515.812.415.1
LayoutLMv3-large [10]768M13.619.313.519.1
UDOP [28]794M19.226.018.725.6
SlideParser (Ours)810M23.733.023.632.1
+ +model weights. We adapt AdamW [19] optimizer with $\beta_{1}$ of 0.9, the $\beta_{2}$ of 0.999, and the weight decay of 0.02. We use the cosine learning rate scheduler and the initial learning rate is $4\times 10^{-5}$ . The total number of iterations is 41K. Moreover, we choose a simple concat operator for encoder outputs to handle multiple slide situations for all models. + +To provide a full view of LecSlides-370K, we evaluate other related methods from various domains: NLP models such as T5 [22] and BART [15], vision-language models such as mPLUG [16] and Vid2Seq [34], and document understanding models such as LayoutLMv3 [10], Donut [12] and UDOP [28]. Specifically, LayoutLMv3 lacks a text decoder, which is unable to handle text generation tasks. To address this limitation, we incorporate the same text decoder as our proposed method. + +# 5.2. Results + +# 5.2.1. Performance on LecSlides-370K + +At the beginning, we analyze Lecture Summary task in LecSlides-370K. As shown in Table 2, SlideParser achieves notable scores of 42.2 in ROUGE-1, 25.3 in ROUGE-L, and 31.3 in METEOR. This performance surpasses other baseline methods, attributed to its effective understanding text relations. Specifically, the methods without processing tex + +Table 3. Performance on Lecture QA in LecSlides-370K + +
Model Namedevtest
EMF1EMF1
VinVL [36]9.411.410.713.5
PreasM [35]36.341.930.738.2
T5 [22]35.241.331.039.7
LayoutT5 [26]38.944.831.739.9
LayoutLMv2 [33]26.533.421.429.3
FiD [11]37.642.930.438.9
M3D [27]41.347.133.541.7
SlideParser (Ours)44.149.837.543.9
SlideParser (Ours)†49.853.942.447.7
+ +Table 4. Performance on SlideVQA benchmark. † represents pretraining on LecSlides-370K and fine-tuning on SlideVQA. + +tual modality, including Vid2Seq [34], mPLUG series [16] and Donut [12], exhibit the worst results. We argue that only processing visual modality may lead to a lack of detailed content extraction in slides. In addition, NLP methods, such as T5 [22] and BART [15], perform relatively well but still lag behind our proposed method, containing excellent multi-modal fusion and text relation utilization. + +In contrast to Lecture Summary, which primarily evaluates global understanding of slides, Lecture QA task measures methods for in-depth understanding of specific content. As shown in Table 3, SlideParser achieves superior performance compared to other baselines. Furthermore, visual-language models such as mPLUG and Vid2Seq still perform the worst among all baselines. It shows that detailed text information is still crucial for QA task. + +# 5.2.2. Performance on SlideVQA + +To further evaluate our proposed method in QA task, we also fine-tune and evaluate SlideParser on SlideVQA, a VQA benchmark in multi-image slides. As shown in Table 4, SlideParser achieves the best performance compared with other methods, which further demonstrates the effective designs of SlideParser. In an extensive experiment, we pre-train SlidePraser in our dataset and fine-tune on SlideVQA, denoted as SlideParser†. Results show that largescale and comprehensive LecSlides-370K yield a great performance enhancement, demonstrating our dataset can also serve as a valuable large-scale pre-training dataset. + +# 5.3. Ablation Study + +To gain a deeper analysis of SlideParser, we conduct ablation experiments in LecSlides-370K Lecture Summary task. + +# 5.3.1. Effect of Text Relation Prediction Module. + +We first evaluate the impact of incorporating the Text Relation Prediction Module. As shown in Table 5, results demonstrate the removal of the entire Text Relation Prediction Module (in r1) yields the poorest performance. Furthermore, as we add RPH and RAL to predict and utilize rela + +![](images/c02792924cdeda7fe787033e8e6145ecd1348554a23cbc4a5a70b8ca5793d571.jpg) +Figure 5. Visualization of predicted text relations in RPH + +
RPHRALPseudo LabelR-1R-LMETEOR
r1XXX36.322.626.9
r2✓✓X37.523.227.7
r3X✓✓41.825.030.8
r4✓✓✓42.225.331.3
+ +Table 5. Ablation on Text Relation Prediction Module + +
Cluster TypeR-1R-LMETEOR
Kmeans (k=3)40.924.829.8
Kmeans (k=4)40.724.729.6
HAC(n=3)41.024.830.1
DBSCAN41.324.830.6
DL-Cluster42.225.331.3
+ +tions without any supervision (in r2), the performance only has a slight improvement. It demonstrates the RPH struggles to converge effectively in generating reasonable text relations. In addition, we also explore the approach of directly feeding pseudo relation labels generated by DL-Cluster into the subsequent RAL, instead of utilizing learnable relations (in r3). Although it exhibits substantial improvement, yet it continues to fall short in comparison to leveraging learnable text predictions and establishing weak supervision (r3 vs r4). It shows that learnable relations could be refined during the training process. + +Moreover, we provide visualization results in Fig. 5. We assign the same color for two texts if their relation score is larger than 0.5. The results illustrate that the RPH can predict reasonable text relations under weak supervision. + +# 5.3.2. Different Methods for Pseudo Label Generation. + +Traditional cluster methods are widely used in previous works for relation prediction [1, 8]. Therefore, we compare our proposed DL-Cluster with other cluster methods. As shown in Table 6, DL-Cluster surpasses clustering techniques with predefined cluster numbers like K-means and Hierarchical Agglomerating Clustering (HAC). Additionally, DL-Cluster outperforms methods without specified cluster numbers, such as DBSCAN [6], due to its capacity to consider spatial and semantic aspects concurrently. + +![](images/7f67174f07536019fe83af6782212bd13b307d919acadf911d3055cf3985747a.jpg) +Figure 6. Qualitative results of Lecture QA in LecSlides-370K. + +Table 6. Ablation on cluster types in pseudo label generation + +
SpatialSemanticR-1R-LMETEOR
×✓38.323.728.5
✓×41.424.830.7
✓✓42.225.331.3
+ +Table 7. Ablation on designs of DL-Cluster + +# 5.3.3. Design of DL-Cluster. + +We further investigate the impact of text relations represented in spatial and semantic dimensions. As shown in Tab. 7, removing either of these dimensions leads to a decline in performance, and spatial relations have a larger impact. The best results are achieved when both aspects were used simultaneously, indicating that both dimensions complement each other to enhance overall performance. + +# 5.4. Visualization + +In a side-by-side qualitative analysis for Lecture QA in LecSlides-370K, we compare our proposed method with other baselines in related fields. As shown in Fig. 6, results show that other methods face two challenges: 1) They can not grasp correct internal logic due to the complex text relations. 2) As the number of diagrams on slides increases, other methods struggle to pinpoint the correct diagram corresponding to the question. In contrast, SlideParser excels in accurate responses by comprehending text relations effectively and easily determining the answer's location within the slides. + +# 6. Conclusion + +In this work, we introduce a large-scale and multi-task lecture slides dataset, LecSlides-370K, which has a larger data scale and more diverse disciplines. Furthermore, considering the pivotal role of text relation prediction and utilization for slides understanding, we propose SlideParser. The results demonstrate the effect of enhanced text relation design, and our method outperforms other models. We believe LecSlides-370K and SlideParser can prompt lecture slides research. + +# Acknowledgments + +This work was supported by the NSFC (62206103 and 62225603) + +# References + +[1] Rhys Agombar, Max Luebbering, and Rafet Sifa. A clustering backed deep learning approach for document layout analysis. In Machine Learning and Knowledge Extraction: 4th IFIP TC 5, TC 12, WG 8.4, WG 8.9, WG 12.9 International Cross-Domain Conference, CD-MAKE 2020, Dublin, Ireland, August 25–28, 2020, Proceedings 4, pages 423–430. Springer, 2020. 6, 8 +[2] David Arthur and Sergei Vassilvitskii. k-means++: the advantages of careful seeding. In Society for Industrial and Applied Mathematics, page 1027-1035, 2007. 2, 6 +[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 4 +[4] Zhe Chen, Heyang Liu, Wenyi Yu, Guangzhi Sun, Hongcheng Liu, Ji Wu, Chao Zhang, Yu Wang, and Yanfeng Wang. M3av: A multimodal, multigenre, and multipurpose audio-visual academic lecture dataset. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 9041–9060, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1, 2, 3 +[5] Kartik Dutta, Minesh Mathew, Praveen Krishnan, and CV Jawahar. Localizing and recognizing text in lecture videos. In 2018 16th international conference on frontiers in handwriting recognition (ICFHR), pages 235-240. IEEE, 2018. 1, 2, 3 +[6] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, pages 226-231, 1996. 2, 6, 8 +[7] Damianos Galanopoulos and Vasileios Mezaris. Temporal lecture video fragmentation using word embeddings. In MultiMedia Modeling: 25th International Conference, MMM 2019, Thessaloniki, Greece, January 8-11, 2019, Proceedings, Part II 25, pages 254-265. Springer, 2019. 1, 2, 3 +[8] Zhangxuan Gu, Changhua Meng, Ke Wang, Jun Lan, Weiqiang Wang, Ming Gu, and Liqing Zhang. Xylayoutlm: Towards layout-aware multimodal networks for visually-rich document understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4583-4592, 2022. 6, 8 +[9] Jaekyu Ha, R.M. Haralick, and I.T. Phillips. Recursive x-y cut using bounding boxes of connected components. In Proceedings of 3rd International Conference on Document Analysis and Recognition, pages 952-955, 1995. 2, 6 +[10] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM + +International Conference on Multimedia, pages 4083-4091, 2022. 1, 3, 5, 7 +[11] Gautier Izacard and Edouard Grave. Leveraging passage retrieval with generative models for open domain question answering. arXiv preprint arXiv:2007.01282, pages 1-6, 2020. 7 +[12] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 7 +[13] JP Kincaid. Derivation of new readability formulas (automated readability index, fog count and flesch reading ease formula) for navy enlisted personnel. Chief of Naval Technical Training, pages 1-45, 1975. 4 +[14] Dong Won Lee, Chaitanya Ahuja, Paul Pu Liang, Sanika Natu, and Louis-Philippe Morency. Lecture presentations multimodal dataset: Towards understanding multimodality in educational videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20087-20098, 2023. 1, 2, 3, 4 +[15] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online, 2020. Association for Computational Linguistics. 1, 3, 7 +[16] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, He Chen, Guohai Xu, Zheng Cao, Ji Zhang, Songfang Huang, Fei Huang, Jingren Zhou, and Luo Si. mPLUG: Effective and efficient vision-language learning by cross-modal skip-connections. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7241–7259, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 1, 3, 7 +[17] Irene Li, Alexander R Fabbri, Robert R Tung, and Dragomir R Radev. What should i learn first: Introducing lecturebank for nlp education and prerequisite chain learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6674-6681, 2019. 1, 2, 3 +[18] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 4 +[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, pages 1-19, 2017. 7 +[20] T. Muthuprasad, S. Aiswarya, K.S. Aditya, and Girish K. Jha. Students' perception and preference for online education in india during covid -19 pandemic. Social Sciences Humanities Open, 3(1):100101, 2021. 1 +[21] Pitambar Paudel. Online education: Benefits, challenges and strategies during and after Covid-19 in higher education. International Journal on Studies in Education (IJonSE), 3:70–85, 2021. 1 + +[22] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1, 3, 7 +[23] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3982-3992. Association for Computational Linguistics, 2019. 6 +[24] Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53-65, 1987. 6 +[25] Anna Sun and Xiufang Chen. Online education and its effective practice: A research review. Journal of information technology education: Research, 15:1-34, 2016. 1 +[26] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 7 +[27] Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. Slidevqa: A dataset for document visual question answering on multiple images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13636-13645, 2023. 1, 2, 3, 4, 7 +[28] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19254-19264, 2023. 1, 3, 5, 6, 7 +[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 6000-6010, 2017. 6 +[30] Jiawei Wang, Kai Hu, Zhuoyao Zhong, Lei Sun, and Qiang Huo. Detect-order-construct: A tree construction based approach for hierarchical document structure analysis. arXiv preprint arXiv:2401.11874, pages 1-35, 2024. 5 +[31] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 3 +[32] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192-1200, 2020. 1, 3 +[33] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. LayoutLMv2: Multi-modal pre-training for visually-rich document understanding. In Proceedings of the 59th Annual Meeting of the + +Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2579-2591, Online, 2021. Association for Computational Linguistics. 1, 3, 7 +[34] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 1, 3, 7 +[35] Ori Yoran, Alon Talmor, and Jonathan Berant. Turning tables: Generating examples from semi-structured tables for endowing language models with reasoning skills. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pages 6016–6031. Association for Computational Linguistics, 2022. 7 +[36] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5579-5588, 2021. 3, 7 \ No newline at end of file diff --git a/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/images.zip b/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b5a057b9be1b5baf1489a6c94fde12163b5bea2a --- /dev/null +++ b/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:741b4aca1acedf85b62a200196932a766562ec62893f09de7f8fb4ba6e69439a +size 605712 diff --git a/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/layout.json b/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..39a5b22bbd08c4bf11b1d466970507771537d0e7 --- /dev/null +++ b/2025/Towards Comprehensive Lecture Slides Understanding_ Large-scale Dataset and Effective Method/layout.json @@ -0,0 +1,7975 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 553, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 553, + 138 + ], + "type": "text", + "content": "Towards Comprehensive Lecture Slides Understanding: Large-scale Dataset and Effective Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "spans": [ + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "text", + "content": "Enming Zhang" + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "text", + "content": ", Yuzhe Li" + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "text", + "content": ", Yuliang Liu" + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "text", + "content": ", Yingying Zhu" + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "text", + "content": ", Xiang Bai" + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "text", + "content": "Huazhong University of Science and Technology \n{emzhang,yzli,ylliu,yyzhu,xbai} " + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 134, + 160, + 475, + 204 + ], + "type": "text", + "content": " hust.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 255, + 297, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 297, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 297, + 531 + ], + "type": "text", + "content": "Online education has been widespread in worldwide universities and educational institutions. Lecture slides, a fundamental component of online education, contain a wealth of information, playing a crucial role in learning. However, previous works have not yet paid sufficient attention to understanding lecture slides, including the absence of the large-scale dataset and comprehensive understanding tasks. To facilitate the research about lecture slides understanding, we establish the LecSlides-370K, which consists of 25,542 lectures with 370,078 slides across 15 areas. We also introduce two comprehensive tasks, Lecture Summary and Lecture Question Answering (QA), for providing different perspectives of slides understanding. Furthermore, complex and flexible text relations can hinder the understanding of the internal logic of slides. To address this challenge, we propose a novel method, named SlideParser, which includes an auxiliary branch to predict text relations within slides and enhance attention between related texts, thereby improving slides understanding. With extensive experiments, we show the superiority of our proposed method on both LecSlides-370k and SlideVQA. Dataset and Codes are available at https://github.com/zamling/LecSlides_370K" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 554, + 135, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 554, + 135, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 554, + 135, + 567 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 574, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 695 + ], + "type": "text", + "content": "With the rapid progress in online education, an increasing number of universities and educational institutions have adopted online learning in many platforms [20, 21, 25]. Among diverse online education formats, such as lecture videos and live streams, slides often serve as a key medium for conveying significant knowledge. Consequently, more researchers have turned their attention to this field. However, there are two obstacles to understand lecture slides: 1) Texts in slides are often conclusive and headline-style statements, which complicate the task of providing detailed" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 232, + 555, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 268 + ], + "type": "text", + "content": "content. 2) The complex text relations pose additional challenges for models attempting to grasp the internal logic and generate reasonable responses." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 269, + 556, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 556, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 556, + 412 + ], + "type": "text", + "content": "Recently, several works [4, 5, 7, 14, 17, 27] have noticed the importance of lecture slides and introduced corresponding datasets, such as text detection in slides [5], slide segmentation [7] and image-text retrieval for slides and corresponding speech [14]. However, none of them has yet paid attention on comprehensive understanding slides. SlideVQA [27], a recent slides dataset, provides a Question-Answering (QA) benchmark for slides, yet it still overlooks the global understanding task, such as generating a detailed summary for a complete set of slides. Additionally, the restricted scales and limited number of lecture areas also constrain the development of lecture slides understanding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 413, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 556, + 616 + ], + "type": "text", + "content": "Aiming to address the limitations in the lecture slides dataset, we establish a large-scale and multi-task dataset for comprehensive slides understanding, named LecSlides-370K. Specifically, we collect 25,542 lectures composed of 370,078 slides. These lectures span 15 lecture areas, including business, natural sciences, computer sciences, and so on, significantly surpassing previous datasets in both scale and diversity, as shown in Fig. 1 (a). Furthermore, we introduce two tasks based on our dataset, termed Lecture Summary and Lecture Question Answering (QA). These two tasks represent the most frequently utilized functions for students, offering varied perspectives for slides understanding. Specifically, Lecture Summary focuses on extracting key information and fostering a global understanding, while Lecture QA focuses on locating problems and understanding fine-grained content. Finally, our dataset includes 114,472 summary annotations and 259,461 QA pairs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "Based on LecSlides-370K dataset, we evaluate the existing NLP methods [15, 22], document understanding methods [10, 28, 32, 33] and vision-language models [16, 34]. The results reveal the inferior performance of these methods. We argue that this inferiority can be attributed to the fact that these methods have not yet adequately considered the unique characteristic of lecture slides, that text relations in slides are complex and contain a wealth of logic and rel" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4455" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 231, + 219 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 231, + 219 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 231, + 219 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 231, + 219 + ], + "type": "image", + "image_path": "dce8870acaf00311263f9d70cd056d989c5a01c24e1226f0bc702777ffd06985.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 221, + 225, + 230 + ], + "lines": [ + { + "bbox": [ + 63, + 221, + 225, + 230 + ], + "spans": [ + { + "bbox": [ + 63, + 221, + 225, + 230 + ], + "type": "text", + "content": "(a) Data scales of LecSlides-370K and related datasets" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 232, + 70, + 553, + 219 + ], + "blocks": [ + { + "bbox": [ + 232, + 70, + 553, + 219 + ], + "lines": [ + { + "bbox": [ + 232, + 70, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 232, + 70, + 553, + 219 + ], + "type": "image", + "image_path": "2578ba197cded1e26310be408419f364f44f6ba4867fb178fdbd4b55ed45c5b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 221, + 479, + 230 + ], + "lines": [ + { + "bbox": [ + 318, + 221, + 479, + 230 + ], + "spans": [ + { + "bbox": [ + 318, + 221, + 479, + 230 + ], + "type": "text", + "content": "(b) An example of LecSlides-370K with various tasks" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 239, + 555, + 283 + ], + "lines": [ + { + "bbox": [ + 54, + 239, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 555, + 283 + ], + "type": "text", + "content": "Figure 1. LecSlides-370K is a large-scale and multi-task lecture slides dataset. It contains 25,542 lectures and 370,078 slides across 15 lecture areas, which has around 9.7 times lecture numbers and 7.1 times slide numbers compared with SlideVQA [27]. (a) A comparison between LecSlides-370K and other slides dataset in terms of the number of lectures, slides and lecture areas. The detailed information is shown in Table 1. (b) An example of LecSlides-370K, including Lecture Summary and Lecture QA tasks." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 304, + 294, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 294, + 376 + ], + "type": "text", + "content": "evance. For instance, as shown in Fig. 1 (b), texts displayed in the upper left corner indicate detailed requirements for forecasting, while texts on the right side of each block explain each stage in the flow chart. Therefore, understanding complex text relations in slides can grasp the internal logic, and is the key part of understanding slides." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 377, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 555 + ], + "type": "text", + "content": "To extract the complex text relations and enhance slides understanding, we propose a novel method called SlideParser. The key of SlideParser is how to effectively predict and utilize complex text relations in slides. Specifically, we propose an auxiliary branch, termed Text Relation Prediction Module, comprising the Relation Prediction Head (RPH) to predict relations among texts and the Relation Attention Layer (RAL) to leverage predicted text relations, enhancing attention between related texts while reducing it between unrelated texts. Additionally, to enhance the quality of predictions, we build a weak supervision framework for RPH with a pseudo relation label generation pipeline, named DL-Cluster, which considers both text spatial and semantic relations, yielding superior text relation labels compared to traditional clustering methods [2, 6, 9]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 557, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 664 + ], + "type": "text", + "content": "In our experiments on LecSlides-370K, related baselines struggle to extract the main point of lecture slides and pinpoint the correct content corresponding to the question, revealing the challenges of our dataset. By effectively predicting and utilizing text relations, SlideParser outperforms all the related methods in the LecSlides-370K Lecture Summary and Lecture QA benchmarks. Additionally, we evaluate SlideParser on the SlideVQA [27] dataset, further validating the effectiveness of our text relation design." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 665, + 216, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 665, + 216, + 676 + ], + "spans": [ + { + "bbox": [ + 67, + 665, + 216, + 676 + ], + "type": "text", + "content": "In conclusion, our contributions are:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "- We establish a large-scale and multi-task slides understanding dataset, termed LecSlides-370K, comprising 25,542 lectures across 15 lecture areas and 370,078" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 322, + 304, + 553, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 304, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 322, + 304, + 553, + 327 + ], + "type": "text", + "content": "slides. We also introduce two comprehensive tasks, Lecture Summary and Lecture QA." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 329, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 329, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 314, + 329, + 553, + 365 + ], + "type": "text", + "content": "- We propose a novel method, SlideParser, including an auxiliary brach to effectively predict and utilize complex text relations in slides, enhancing slides understanding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 392, + 400, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 392, + 400, + 403 + ], + "spans": [ + { + "bbox": [ + 314, + 392, + 400, + 403 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 416, + 446, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 416, + 446, + 427 + ], + "spans": [ + { + "bbox": [ + 314, + 416, + 446, + 427 + ], + "type": "text", + "content": "2.1. Lecture Slides Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 439, + 554, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 554, + 499 + ], + "type": "text", + "content": "Recently, several datasets based on lecture slides have been proposed, which are most relevant to our current work, including the LectureBank Dataset [17], ALV [7], LectureVideoDB [5], LPM Dataset [14], " + }, + { + "bbox": [ + 313, + 439, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^3\\mathrm{AV}" + }, + { + "bbox": [ + 313, + 439, + 554, + 499 + ], + "type": "text", + "content": " [4] and SlideVQA [27], as shown in Table 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 505, + 554, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 554, + 613 + ], + "type": "text", + "content": "LectureBank [17] is annotated for prerequisite relationships among topics, providing a structure of connected knowledge. ALV dataset [7] primarily focuses on segmenting the entire video lecture. LectureVideoDB is formulated to investigate text detection and recognition in video lectures. Furthermore, LPM [14] attempts to explore the retrieval of spoken words with corresponding visual cues. SlideVQA, a recent work for slides, proposes a Visual Question Answering (VQA) task on it." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": "In summary, previous datasets contain two primary limitations: 1) Most of datasets do not pay sufficient attention to slides understanding, and the tasks they define also do not provide multiple perspectives for understanding. 2) The limited data scale hinders the advancement of understanding lecture slides. Therefore, a substantial requirement in the research community emerges for a large-scale slides dataset with comprehensive understanding tasks." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4456" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 70, + 500, + 198 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 500, + 198 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 500, + 198 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 500, + 198 + ], + "type": "table", + "html": "
DatasetSizeAnnotationTask
LecturesSlidesOCRQASummary
LectureVideoDB [5]245,000✓--Text Detection
LectureBank [17]1,35251,939✓--Prerequisite Chain
ALV Dataset [7]-1,498✗--Slide Segmentation
LPM [14]3349,031✓--Image Retrieval
M3AV [4]111324,956✓--ASR & TTS
SlideVQA [27]2,61952,380✓14,500-QA
LecSlides-370K (Ours)25,542370,078✓259,461114,472QA & Summary
", + "image_path": "e508a39255bc9c2167651a9255b6121a8dd2a5e84eda82551c0fca5a51b3f6a7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "type": "text", + "content": "Table 1. Data scales, annotations and tasks for different lecture slides datasets. LecSlides-370K contain the largest data scale, the most comprehensive annotations and two understanding tasks to evaluate slides understanding from multiple perspectives." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 249, + 222, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 222, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 222, + 262 + ], + "type": "text", + "content": "2.2. Slides Understanding Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 267, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 295, + 327 + ], + "type": "text", + "content": "Although there is a lack of specific methods for slide understanding, other area methods can be used in slide understanding, mainly including three categories: vision-language methods [16, 31, 34, 36], NLP methods [15, 22], and document understanding methods [10, 28]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 328, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 328, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 328, + 295, + 483 + ], + "type": "text", + "content": "Vision-language methods typically begin with large-scale pre-training tasks and then fine-tune on downstream multi-modal understanding tasks. These methods often integrate a vision-language fused transformer encoder [16, 31, 36]. NLP methods provide a plain text solution, such as T5 [22] and BART [15], following an encoder-decoder architecture to process textual contents and generate textual outputs. Furthermore, inspired by NLP approaches, document understanding methods replace traditional transformer encoder to multi-modal encoder, processing and fusing textual and visual content. For example, LayoutLM series [10, 32, 33] integrate text, image features, and 2D positional embeddings in the multi-modal encoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 484, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 296, + 521 + ], + "type": "text", + "content": "However, these related methods do not consider the complex text relations in slides, a crucial part of slides understanding, thereby yielding inferior performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 533, + 192, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 192, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 192, + 546 + ], + "type": "text", + "content": "3. LecSlides-370K Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 554, + 151, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 151, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 151, + 566 + ], + "type": "text", + "content": "3.1. Data Collection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 572, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 295, + 693 + ], + "type": "text", + "content": "To build a large-scale and multi-modal lecture slides dataset, we collect slides from Slideshare*. Our raw data consists of 720,673 slides from 39,784 lectures at the beginning. However, some of these slides are irrelevant to the lecture content, such as advertisements, or contain extraneous information, which is not essential for comprehension. Thus, we filter these slides by their contents. Finally, LecSlides-370K contains 370,078 slides from 25,542 lectures across 15 areas, such as computer science, marketing, medicine, and so on. As shown in Table 1, compared with" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 249, + 553, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 553, + 298 + ], + "type": "text", + "content": "the previous largest datasets, SlideVQA [27], LecSlides-370K has around 9.7 times lecture numbers and 7.1 times slide numbers. The superiority of data scale and diversity renders our dataset more comprehensive and challenging." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 299, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 479 + ], + "type": "text", + "content": "Moreover, since the length of these raw lecture slides varies widely, from a few to hundreds, which makes it difficult for researchers to use them, we segment these raw slides into 114,472 lecture chunks, following the principle that each chunk belongs to the same sub-topic. For instance, if one slide initially displays the first point of a list, and the subsequent slide reveals the following points, these slides should belong to the same chunk. The distribution of slides number in each lecture chunk is shown in Fig. 2 (b). Additionally, we annotate the positions of texts, illustrations, and tables in slides to facilitate the study and encourage researchers to focus more on developing new methods rather than information extraction and data processing. Furthermore, we also annotate the textual contents in slides, which can be utilized by the models to understand slides better." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 491, + 408, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 408, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 408, + 502 + ], + "type": "text", + "content": "3.2. Task Definition" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "We define two tasks: Lecture Summary and Lecture QA. These two tasks not only involve most application scenarios, but also take into account the global and detailed understanding of slides, fully reflecting the understanding ability of models. In the context of Lecture Summary, given a lecture chunk " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I} = \\{I_1, I_2, \\dots, I_j\\}^M" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": ", comprising " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": " slide images, the model's objective is to analyze the content of each slide, integrate contextual information, retain key details, and eliminate redundancies. Finally, the model outputs a summary of all slides. Contrary to the global scope of Lecture Summary, Lecture QA primarily concentrates on locating and comprehending specific content within a lecture chunk. Given a lecture chunk and a query " + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": ", the model need to identify the slide corresponding to the question and then provide an answer based on the content of that slide. QA serves to assess the model's fine-grained understanding of slides. In total, We annotate 114,472 summary labels and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 67, + 702, + 157, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 157, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 157, + 713 + ], + "type": "text", + "content": "\\*https://www.slideshare.net" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4457" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 284, + 255 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 284, + 255 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 284, + 255 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 284, + 255 + ], + "type": "image", + "image_path": "c08e6c9169570e6bf5c56f30f5c5dade21740a5eb4f6d5d9528b09bb1990e4e1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 264, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 555, + 299 + ], + "type": "text", + "content": "Figure 2. Examples and statistics of LecSlides-370K. (a) Lecture numbers in each discipline for LecSlides-370K, SlideVQA [27] and LPM [14]. (b) The number of slides for each chunk. (c) The distribution of OCRs in slides. (d) The distribution of summary length. (e) The percentage of each QA type." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 284, + 70, + 411, + 160 + ], + "blocks": [ + { + "bbox": [ + 284, + 70, + 411, + 160 + ], + "lines": [ + { + "bbox": [ + 284, + 70, + 411, + 160 + ], + "spans": [ + { + "bbox": [ + 284, + 70, + 411, + 160 + ], + "type": "image", + "image_path": "fd1f0a697e404d64196040d4d3fdefd28389d04d2e5bdf52a8c1af4528406c87.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 415, + 71, + 547, + 161 + ], + "blocks": [ + { + "bbox": [ + 415, + 71, + 547, + 161 + ], + "lines": [ + { + "bbox": [ + 415, + 71, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 415, + 71, + 547, + 161 + ], + "type": "image", + "image_path": "c8a26ff8849eee5283b2dda52594a9195039edf05fb45c86cd5a43a438a64a6b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 133, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 133, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 133, + 331 + ], + "type": "text", + "content": "259,461 QA pairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 337, + 222, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 222, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 222, + 350 + ], + "type": "text", + "content": "3.3. Dataset Statistics and Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 354, + 295, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 295, + 379 + ], + "type": "text", + "content": "In this section, we aim to provide a comprehensive analysis and understanding of the LecSlides-370K dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 384, + 157, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 157, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 157, + 396 + ], + "type": "text", + "content": "3.3.1. Lecture Diversity" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 398, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 506 + ], + "type": "text", + "content": "We collect slides from various disciplines, as shown in Fig. 2 (a). There are 15 disciplines, and the highest proportion of data pertains to business, while law and marketing lectures constitute a relatively small percentage. Compared with previous slides datasets, such as SlideVQA [27] and LPM [14], LecSlides-370K contains more disciplines, such as marketing, agriculture, and so on. The diversity of disciplines contains rich teaching content, bringing challenges that models need to handle more diverse scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 511, + 151, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 151, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 151, + 521 + ], + "type": "text", + "content": "3.3.2. Professionalism" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 525, + 295, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 645 + ], + "type": "text", + "content": "Lecture slides usually contain specialized knowledge and terminologies, bringing more challenges. Therefore, we use the Flesch Reading Ease (FRES) score [13] to evaluate the difficulty of contents and get an average score of 34.3 under this metric. This indicates that the content of our dataset is on 'Hard to Read' level. Moreover, we also use the Flesch-Kincaid Grade Level [13] to reflect which education level our dataset is appropriate for and obtain a 14.1 grade on average. These show there is a professional difficulty in our dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 651, + 134, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 134, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 134, + 662 + ], + "type": "text", + "content": "3.3.3. Slide Image" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "For LecSlides-370K, there are 114,472 lecture chunks corresponding to 370,078 slide images. Each chunk contains a varying number of slide images, with an average of 3.2 images per chunk. The distribution of chunks containing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 437 + ], + "type": "text", + "content": "different numbers of slide images is shown in Fig. 2 (b), and most chunks contain two slide images. Moreover, slide images are rich in visual information. By our count, there are 94,906 diagrams, 34,562 illustrations, and 18,938 tables in LecSlides-370K dataset. We further show the statistics for textual content in slides. We totally annotate 28,220,274 words, with an average of 76.2 words. The size of the vocabulary for our dataset is 84,630. In Fig. 2 (c), we show the distribution of texts across slides, revealing that the majority of slides contain fewer than 100 words." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 445, + 469, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 469, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 469, + 455 + ], + "type": "text", + "content": "3.3.4. Characteristics of Annotation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 460, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 591 + ], + "type": "text", + "content": "For summary annotations, Fig. 2 (d) illustrates the length distribution of the annotations. The average length of the summary is 124.6 words. We further comprehensively analyze QA annotations, including four distinct question types: binary choice, visual analysis, data interpretation and textual reasoning. As shown in Fig. 2 (e), the highest proportion of question types is textual reasoning, while binary choice constitutes the smallest percentage. This diverse set of question types covers the whole slide content, enabling a comprehensive evaluation of the model's QA capabilities. In addition, the average length of answer is 4.1 words." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 600, + 430, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 430, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 430, + 611 + ], + "type": "text", + "content": "3.4. Evaluation Protocol" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "LecSlides-370K dataset contains 109,472 slide chunks for the training set, 2,000 chunks for the validation set and 3,000 chunks for the test set. We adapt the ROUGE-1, ROUGE-L [18] and METEOR [3] as evaluation metrics for Lecture Summary task. These metrics are widely used in image caption and text generation, as they consider the quality of generated texts in different aspects. Following SlideVQA [27], we choose EM (Exact Matching) and F1" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4458" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 294, + 195 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 294, + 195 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 294, + 195 + ], + "type": "image", + "image_path": "77a82f29de098fe3925f47adbf0e1703f41e4c8cb5372a4d86334600716c29b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 198, + 231, + 207 + ], + "lines": [ + { + "bbox": [ + 100, + 198, + 231, + 207 + ], + "spans": [ + { + "bbox": [ + 100, + 198, + 231, + 207 + ], + "type": "text", + "content": "(a) Overview of SlideParser Architecture" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 68, + 553, + 192 + ], + "blocks": [ + { + "bbox": [ + 301, + 68, + 553, + 192 + ], + "lines": [ + { + "bbox": [ + 301, + 68, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 301, + 68, + 553, + 192 + ], + "type": "image", + "image_path": "6174a25993604cd65c0b87875edfefe565c05f5177a7a3b0c0137c1e8dfe9e76.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 197, + 506, + 206 + ], + "lines": [ + { + "bbox": [ + 340, + 197, + 506, + 206 + ], + "spans": [ + { + "bbox": [ + 340, + 197, + 506, + 206 + ], + "type": "text", + "content": "(b) Architecture of Text Relation Prediction Module" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 555, + 251 + ], + "type": "text", + "content": "Figure 3. (a) The architecture of our proposed method, SlideParser. It consists of a Transformer Encoder, a Text Relation Prediction Module, and a Text Decoder. (b) The details of Text Relation Prediction Module, which can predict and utilize text relations effectively, including the Relation Prediction Head (RPH) with weak supervision and the Relation Attention Layer (RAL)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 271, + 266, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 266, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 266, + 283 + ], + "type": "text", + "content": "score as evaluation metrics for the Lecture QA task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 294, + 111, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 111, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 111, + 306 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 296, + 422 + ], + "type": "text", + "content": "As shown in Fig. 3 (a), SlideParser consists of a Transformer Encoder, a Text Relation Prediction Module, and a Text Decoder. Texts and image patches from slides are fed into the Transformer Encoder, and the output text embeddings are forwarded to the Text Relation Prediction Module to predict text relations and utilize them to update text embeddings. Finally, updated text embeddings combine with output image embeddings and are fed into the Text Decoder to generate the summary or answer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 430, + 144, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 144, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 144, + 443 + ], + "type": "text", + "content": "4.1. Input Schema" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "content": "SlideParser is built on a unified text-image multi-modal Transformer Encoder to learn cross-modal representations. Similar to previous document understanding works [10, 28], the Transformer Encoder adopts a multi-layer structure, with each layer primarily comprising multi-head self-attention and feed-forward networks. The input of Transformer Encoder consists of text embeddings " + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = x_{1:M}" + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "content": " and image embedding " + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = v_{1:M}" + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "content": " are sequence length of textual and visual embeddings respectively. Specifically, for text embedding " + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "content": " text tokens " + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^L" + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "content": " are encoded by word embedding with 1D position and 2D layout position embedding [28]. Meanwhile, the slide image " + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "content": " is divided into " + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "inline_equation", + "content": "M = \\frac{H}{P} \\times \\frac{H}{P}" + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "content": " image patches. These patches are then processed through Patch Embedding to generate the image embeddings, " + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 448, + 296, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 635, + 231, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 231, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 231, + 647 + ], + "type": "text", + "content": "4.2. Text Relation Prediction Module" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "In Transformer Encoder, 2D position embedding, which can be used to comprehend layout and text relations, has been demonstrated in previous document understanding works [10, 28]. However, in the context of slides, text relations tend to be more complex compared to traditional doc" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 271, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 319 + ], + "type": "text", + "content": "uments, such as examples in Fig. 2. Merely incorporating 2D position embeddings into the transformer encoder may not suffice for comprehensive relation comprehension and could potentially introduce inaccuracies." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 320, + 556, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 556, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 556, + 429 + ], + "type": "text", + "content": "To address this issue, we introduce the Text Relation Prediction Module, including the Relation Prediction Head (RPH) with weak supervision and the Relation Attention Layer (RAL), as shown in Fig. 3 (b). This module predicts text relations by RPH and utilizes relations to enhance attention between related texts while reducing attention between unrelated texts by RAL. Additionally, to improve the quality of predicted relations, we establish a weak supervision pipeline. We will introduce each block in detail." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 439, + 451, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 451, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 451, + 450 + ], + "type": "text", + "content": "4.2.1. Relation Prediction Head" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 456, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 503 + ], + "type": "text", + "content": "To predict the accurate text relations, inspired by [30], RPH leverages text position and semantic information to predict text relations. Given text bounding boxes " + }, + { + "bbox": [ + 313, + 456, + 555, + 503 + ], + "type": "inline_equation", + "content": "\\{b_i\\}_{i=1}^L" + }, + { + "bbox": [ + 313, + 456, + 555, + 503 + ], + "type": "text", + "content": ", we define the box delta as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 328, + 513, + 553, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 513, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 328, + 513, + 553, + 544 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Delta \\left(b _ {i}, b _ {j}\\right) = \\left\\{\\left(x _ {i} ^ {c t r} - x _ {j} ^ {c t r}\\right) / w _ {i}, \\left(y _ {i} ^ {c t r} - y _ {j} ^ {c t r}\\right) / h _ {i}, \\right. \\tag {1} \\\\ \\left. \\log \\left(w _ {i} / w _ {j}\\right), \\log \\left(w _ {i} / w _ {j}\\right) \\right\\} \\\\ \\end{array}", + "image_path": "f024720dc625e884da2acf79d1591149789f67920d2694e1cfbd9c97d7c11496.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "x_{i}^{ctr}" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "y_{i}^{ctr}" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": " represent the position of the center point for bounding box, and " + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": " are the width and height of box respectively. Moreover, we present the spatial relation of two boxes as: " + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathbf{bi},\\mathbf{bj}} = \\{\\Delta (b_i,b_j),\\Delta (b_i,b_{ij}),\\Delta (b_{ij},b_j)\\}" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": ", which is a 12-D vector. Specifically, " + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "b_{ij}" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": " is a union bounding box for " + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "b_{j}" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": ". Then, we employ an MLP layer and integrate all box relations to get the spatial relation matrix " + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathbf{spa}} \\in \\mathbb{R}^{L\\times L}" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": ". In addition, RPH also considers semantic relations. Given output text embeddings from Transformer Encoder, " + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{E_T} \\in \\mathbb{R}^{L\\times D}" + }, + { + "bbox": [ + 313, + 556, + 555, + 689 + ], + "type": "text", + "content": ", we compute the semantic relation between two texts as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 361, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 361, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {R _ {s e m}} = \\left(\\mathbf {W _ {q} E _ {T}} \\times \\mathbf {W _ {k} E _ {T} ^ {T}}\\right) / \\sqrt {D} \\qquad (2)", + "image_path": "b3f087176cc903ab39f9f58d614ddccc5e506470d4e19d0363542b378a8afb0c.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4459" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 544, + 178 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 544, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 544, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 544, + 178 + ], + "type": "image", + "image_path": "005dc09c7e3e684565abd43a5b65b6a987075f72cf5ee7b14d246c37f9095602.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 188, + 553, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 553, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 553, + 200 + ], + "type": "text", + "content": "Figure 4. The Pipeline of DL-Cluster to generate pseudo relation label. We combine spatial and semantic text relations to get pseudo labels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 219, + 295, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 295, + 232 + ], + "type": "text", + "content": "The predicted relation matrix " + }, + { + "bbox": [ + 55, + 219, + 295, + 232 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_{\\mathbf{R}} \\in \\mathbb{R}^{L \\times L}" + }, + { + "bbox": [ + 55, + 219, + 295, + 232 + ], + "type": "text", + "content": " is computed as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 244, + 295, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 244, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 106, + 244, + 295, + 259 + ], + "type": "interline_equation", + "content": "\\mathbf {\\hat {M} _ {R}} = \\operatorname {S i g m o i d} \\left(\\mathbf {R} _ {\\mathbf {s p a}} + \\mathbf {R} _ {\\mathbf {s e m}}\\right) \\tag {3}", + "image_path": "3ca01b479be553905b7f688a6028a93cba57bb73077cf7e86e8ac5ae2ca287d4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 266, + 235, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 235, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 235, + 277 + ], + "type": "text", + "content": "4.2.2. Weak Supervised Label Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 388 + ], + "type": "text", + "content": "In our experiment, we find that the RPH struggles to generate reasonable text relations without the presence of relation supervision. As a solution, we introduce a weak supervision pipeline to enhance the quality of predictions. Previous works [1, 8] rely on traditional cluster methods [6, 9] to predict text relations. However, these traditional methods can not handle diverse layout format in slides effectively. Moreover, our observations suggest that both spatial and semantic relations play crucial roles in slides." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 484 + ], + "type": "text", + "content": "Thus, we introduce a dual-stream pseudo relation label generation pipeline, termed DL-Cluster (Dynamic Layout Cluster), as shown in Fig. 4. To obtain spatial relations, we first sample " + }, + { + "bbox": [ + 55, + 389, + 296, + 484 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 389, + 296, + 484 + ], + "type": "text", + "content": " reference points as cluster points for each text box. Because of the diverse nature of text relations in slides, a predefined number of clusters may not be ideal for accurate relation predictions. Hence, we opt for a dynamic approach to determine the cluster number for each slide." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "content": "Specifically, based on Kmeans++ [2], we iterate over the number of cluster " + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "content": " from 2 to " + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "content": " and determine the optimal cluster number by Silhouette method [24]. Silhouette analyzes the distances of each cluster point to its own cluster and its closest neighboring cluster. The optimal " + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "content": " is obtained when the average distance of a cluster point to all the other points in its own cluster is minimum, while the average distance of a point to the nearest neighboring cluster's points is maximum. In our experiment, " + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{K} = 8" + }, + { + "bbox": [ + 55, + 486, + 296, + 593 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "content": "Subsequently, we assign a cluster ID to each box through a voting mechanism among reference points. Then, the spatial relation score " + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "inline_equation", + "content": "s_{spa}" + }, + { + "bbox": [ + 55, + 594, + 295, + 630 + ], + "type": "text", + "content": " is computed as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 639, + 295, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 639, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 64, + 639, + 295, + 676 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {\\mathbf {s p a}} = \\left\\{ \\begin{array}{l} 1 - \\frac {1}{N} \\sum_ {\\varepsilon = 1} ^ {N} \\left\\| x _ {i, \\varepsilon} - \\mu_ {j} \\right\\|, \\text {w h e n} C _ {i} = C _ {j} \\\\ 0, \\text {w h e n} C _ {i} \\neq C _ {j} \\end{array} \\right. \\tag {4}", + "image_path": "de49b6eec9743c5f2548043aec3979cd1a7f4bafcd004b6e5ac960094b00a9fb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " are clusters and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mu_j" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " is cluster centroid of " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ". Specifically, scores are assigned based on the average distance between reference points and the cluster centroid" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 220, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 255 + ], + "type": "text", + "content": "when these texts are deemed related and are placed within the same cluster. Conversely, when texts are considered unrelated, the score is set to zero." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 256, + 554, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 554, + 316 + ], + "type": "text", + "content": "Simultaneously, to obtain semantic relation, we extract textual contents in each text box, and compute the semantic similarity by the cosine distance of Sentence Transformer [23] and get the semantic relation score " + }, + { + "bbox": [ + 313, + 256, + 554, + 316 + ], + "type": "inline_equation", + "content": "s_{sem}" + }, + { + "bbox": [ + 313, + 256, + 554, + 316 + ], + "type": "text", + "content": ". These two relation scores are combined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 325, + 553, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 325, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 358, + 325, + 553, + 338 + ], + "type": "interline_equation", + "content": "\\mathbf {s} = \\beta_ {1} \\operatorname {N o r m} \\left(\\mathbf {s} _ {\\mathbf {s p a}}\\right) + \\beta_ {2} \\operatorname {N o r m} \\left(\\mathbf {s} _ {\\mathbf {s e m}}\\right) \\tag {5}", + "image_path": "2c42ff34504699b07590272030e8f7bbdc7e4a0fafdebdc632a76021d7a145b6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 346, + 554, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 554, + 382 + ], + "type": "text", + "content": "we set the default values of " + }, + { + "bbox": [ + 313, + 346, + 554, + 382 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.7" + }, + { + "bbox": [ + 313, + 346, + 554, + 382 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 346, + 554, + 382 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.3" + }, + { + "bbox": [ + 313, + 346, + 554, + 382 + ], + "type": "text", + "content": ". Finally, the relation matrix " + }, + { + "bbox": [ + 313, + 346, + 554, + 382 + ], + "type": "inline_equation", + "content": "M_{R} = \\{s_{ij}\\}^{L\\times L}" + }, + { + "bbox": [ + 313, + 346, + 554, + 382 + ], + "type": "text", + "content": " is binarized with 0.5 threshold. Then, the text relation loss can be calculated as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 370, + 391, + 553, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 391, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 370, + 391, + 553, + 405 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e l} = B C E L o s s \\left(\\hat {\\mathbf {M}} _ {\\mathbf {R}}, \\mathbf {M} _ {\\mathbf {R}}\\right) \\tag {6}", + "image_path": "57c1d7f25a8b811be788abe1468661b786f9d777b4098d83b3cc0dcbe9b9a672.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 414, + 450, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 450, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 450, + 426 + ], + "type": "text", + "content": "4.2.3. Relation Attention Layer" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "text", + "content": "With the predicted text relations, we utilize relations by Self Attention [29], as shown in Fig. 3 (b). Specifically, for predicted relation matrix " + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}_{\\mathbf{R}}" + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "text", + "content": ", we generate attention mask by assigning " + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "inline_equation", + "content": "-\\infty" + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "inline_equation", + "content": "m_{ij} < 0.5" + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "inline_equation", + "content": "m_{ij} \\in \\hat{\\mathbf{M}}_{\\mathbf{R}}" + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "text", + "content": ". Subsequently, the output text embedding " + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{T}}" + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "text", + "content": " are fed into N-layer self-attention blocks with this attention mask, resulting in the updated embeddings, " + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{T}}'" + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "text", + "content": ". It enhances attention between related texts and isolates attention between unrelated texts. Finally, we combine " + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{T}}" + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathbf{T}}'" + }, + { + "bbox": [ + 313, + 428, + 555, + 536 + ], + "type": "text", + "content": " by residual connection." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 544, + 400, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 400, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 400, + 555 + ], + "type": "text", + "content": "4.3. Text Decoder" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 561, + 554, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 554, + 621 + ], + "type": "text", + "content": "Following previous work [28], we adapt a T5 decoder as our text decoder, a uni-directional Transformer decoder that performs the next prediction task to generate texts, such as summary or answer. We choose Cross Entropy loss to supervise text generation. The total loss of SlideParser is" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 400, + 630, + 553, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 630, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 400, + 630, + 553, + 643 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c e} + \\mathcal {L} _ {r e l} \\tag {7}", + "image_path": "241bd42ca23508e7b0fb9782587819d18bd61e6a47dd3528f671325adc04ad6b.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 672, + 422, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 422, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 422, + 685 + ], + "type": "text", + "content": "5.1. Implement details" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "For SlideParser, both the Transformer Encoder and the Text Decoder are initialized from the pre-trained UDOP [28]" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4460" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 306, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 306, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 306, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 306, + 194 + ], + "type": "table", + "html": "
Method#Paramsvaltest
R-1R-LMETEORR-1R-LMETEOR
BART-base [15]217M33.121.121.932.820.121.5
BART-large [15]509M35.921.924.235.021.022.9
T5-base [22]272M34.621.622.834.320.722.4
T5-large [22]803M36.222.124.935.121.323.2
Vid2Seq [34]314M16.011.812.917.813.114.1
mPLUG-base [16]350M27.119.021.627.619.321.7
mPLUG-large [16]600M30.020.524.029.820.323.6
Donut [12]543M27.418.621.327.618.521.1
LayoutLMv3-base [10]533M33.621.023.831.719.822.6
LayoutLMv3-large [10]768M34.621.624.833.921.023.2
UDOP [28]794M36.122.126.435.321.525.6
SlideParser (Ours)810M42.225.331.341.724.730.4
", + "image_path": "68fa788b9579795cfee2e43b99108133c9624d88745da03a0018ea7d45f6bf88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 227, + 291, + 373 + ], + "blocks": [ + { + "bbox": [ + 61, + 202, + 289, + 213 + ], + "lines": [ + { + "bbox": [ + 61, + 202, + 289, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 202, + 289, + 213 + ], + "type": "text", + "content": "Table 2. Performance on Lecture Summary in LecSlides-370K" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 227, + 291, + 373 + ], + "lines": [ + { + "bbox": [ + 60, + 227, + 291, + 373 + ], + "spans": [ + { + "bbox": [ + 60, + 227, + 291, + 373 + ], + "type": "table", + "html": "
Method#Paramsvaltest
EMF1EMF1
BART-base [15]217M18.925.118.524.7
BART-large [15]509M20.627.119.826.3
T5-base [22]272M19.126.018.725.3
T5-large [22]803M20.927.420.027.1
Vid2Seq [34]314M7.014.36.713.6
mPLUG-base [16]350M8.410.98.510.1
mPLUG-large [16]600M9.514.39.514.2
LayoutLMv3-base [10]533M12.515.812.415.1
LayoutLMv3-large [10]768M13.619.313.519.1
UDOP [28]794M19.226.018.725.6
SlideParser (Ours)810M23.733.023.632.1
", + "image_path": "87e7b3a0f61d2f94be6f7a91bcb56af5f7b1ea216f3ba7bfa91b6ac6648edfa0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 416, + 295, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 295, + 488 + ], + "type": "text", + "content": "model weights. We adapt AdamW [19] optimizer with " + }, + { + "bbox": [ + 55, + 416, + 295, + 488 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 55, + 416, + 295, + 488 + ], + "type": "text", + "content": " of 0.9, the " + }, + { + "bbox": [ + 55, + 416, + 295, + 488 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 55, + 416, + 295, + 488 + ], + "type": "text", + "content": " of 0.999, and the weight decay of 0.02. We use the cosine learning rate scheduler and the initial learning rate is " + }, + { + "bbox": [ + 55, + 416, + 295, + 488 + ], + "type": "inline_equation", + "content": "4\\times 10^{-5}" + }, + { + "bbox": [ + 55, + 416, + 295, + 488 + ], + "type": "text", + "content": ". The total number of iterations is 41K. Moreover, we choose a simple concat operator for encoder outputs to handle multiple slide situations for all models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 489, + 295, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 295, + 597 + ], + "type": "text", + "content": "To provide a full view of LecSlides-370K, we evaluate other related methods from various domains: NLP models such as T5 [22] and BART [15], vision-language models such as mPLUG [16] and Vid2Seq [34], and document understanding models such as LayoutLMv3 [10], Donut [12] and UDOP [28]. Specifically, LayoutLMv3 lacks a text decoder, which is unable to handle text generation tasks. To address this limitation, we incorporate the same text decoder as our proposed method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 607, + 113, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 607, + 113, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 607, + 113, + 618 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 625, + 222, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 222, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 222, + 637 + ], + "type": "text", + "content": "5.2.1. Performance on LecSlides-370K" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "At the beginning, we analyze Lecture Summary task in LecSlides-370K. As shown in Table 2, SlideParser achieves notable scores of 42.2 in ROUGE-1, 25.3 in ROUGE-L, and 31.3 in METEOR. This performance surpasses other baseline methods, attributed to its effective understanding text relations. Specifically, the methods without processing tex" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 345, + 70, + 525, + 194 + ], + "blocks": [ + { + "bbox": [ + 72, + 382, + 278, + 393 + ], + "lines": [ + { + "bbox": [ + 72, + 382, + 278, + 393 + ], + "spans": [ + { + "bbox": [ + 72, + 382, + 278, + 393 + ], + "type": "text", + "content": "Table 3. Performance on Lecture QA in LecSlides-370K" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 70, + 525, + 194 + ], + "lines": [ + { + "bbox": [ + 345, + 70, + 525, + 194 + ], + "spans": [ + { + "bbox": [ + 345, + 70, + 525, + 194 + ], + "type": "table", + "html": "
Model Namedevtest
EMF1EMF1
VinVL [36]9.411.410.713.5
PreasM [35]36.341.930.738.2
T5 [22]35.241.331.039.7
LayoutT5 [26]38.944.831.739.9
LayoutLMv2 [33]26.533.421.429.3
FiD [11]37.642.930.438.9
M3D [27]41.347.133.541.7
SlideParser (Ours)44.149.837.543.9
SlideParser (Ours)†49.853.942.447.7
", + "image_path": "ad8512deb9a800685f74358fa29cbab858aea441505f28261510129b699c69f6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 202, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 313, + 202, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 553, + 224 + ], + "type": "text", + "content": "Table 4. Performance on SlideVQA benchmark. † represents pretraining on LecSlides-370K and fine-tuning on SlideVQA." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 247, + 553, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 330 + ], + "type": "text", + "content": "tual modality, including Vid2Seq [34], mPLUG series [16] and Donut [12], exhibit the worst results. We argue that only processing visual modality may lead to a lack of detailed content extraction in slides. In addition, NLP methods, such as T5 [22] and BART [15], perform relatively well but still lag behind our proposed method, containing excellent multi-modal fusion and text relation utilization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": "In contrast to Lecture Summary, which primarily evaluates global understanding of slides, Lecture QA task measures methods for in-depth understanding of specific content. As shown in Table 3, SlideParser achieves superior performance compared to other baselines. Furthermore, visual-language models such as mPLUG and Vid2Seq still perform the worst among all baselines. It shows that detailed text information is still crucial for QA task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 434, + 456, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 456, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 456, + 445 + ], + "type": "text", + "content": "5.2.2. Performance on SlideVQA" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 449, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 581 + ], + "type": "text", + "content": "To further evaluate our proposed method in QA task, we also fine-tune and evaluate SlideParser on SlideVQA, a VQA benchmark in multi-image slides. As shown in Table 4, SlideParser achieves the best performance compared with other methods, which further demonstrates the effective designs of SlideParser. In an extensive experiment, we pre-train SlidePraser in our dataset and fine-tune on SlideVQA, denoted as SlideParser†. Results show that largescale and comprehensive LecSlides-370K yield a great performance enhancement, demonstrating our dataset can also serve as a valuable large-scale pre-training dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 589, + 408, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 408, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 408, + 601 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 607, + 553, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 553, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 553, + 631 + ], + "type": "text", + "content": "To gain a deeper analysis of SlideParser, we conduct ablation experiments in LecSlides-370K Lecture Summary task." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 638, + 522, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 522, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 522, + 649 + ], + "type": "text", + "content": "5.3.1. Effect of Text Relation Prediction Module." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 713 + ], + "type": "text", + "content": "We first evaluate the impact of incorporating the Text Relation Prediction Module. As shown in Table 5, results demonstrate the removal of the entire Text Relation Prediction Module (in r1) yields the poorest performance. Furthermore, as we add RPH and RAL to predict and utilize rela" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4461" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 69, + 276, + 161 + ], + "blocks": [ + { + "bbox": [ + 76, + 69, + 276, + 161 + ], + "lines": [ + { + "bbox": [ + 76, + 69, + 276, + 161 + ], + "spans": [ + { + "bbox": [ + 76, + 69, + 276, + 161 + ], + "type": "image", + "image_path": "c02792924cdeda7fe787033e8e6145ecd1348554a23cbc4a5a70b8ca5793d571.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 169, + 279, + 180 + ], + "lines": [ + { + "bbox": [ + 71, + 169, + 279, + 180 + ], + "spans": [ + { + "bbox": [ + 71, + 169, + 279, + 180 + ], + "type": "text", + "content": "Figure 5. Visualization of predicted text relations in RPH" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 69, + 193, + 282, + 256 + ], + "blocks": [ + { + "bbox": [ + 69, + 193, + 282, + 256 + ], + "lines": [ + { + "bbox": [ + 69, + 193, + 282, + 256 + ], + "spans": [ + { + "bbox": [ + 69, + 193, + 282, + 256 + ], + "type": "table", + "html": "
RPHRALPseudo LabelR-1R-LMETEOR
r1XXX36.322.626.9
r2✓✓X37.523.227.7
r3X✓✓41.825.030.8
r4✓✓✓42.225.331.3
", + "image_path": "71d5dd6a62a40fc3c44fe3be0cc67fdb7e81cdde4f2de762072afd4408139ff0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 102, + 288, + 249, + 361 + ], + "blocks": [ + { + "bbox": [ + 77, + 265, + 274, + 275 + ], + "lines": [ + { + "bbox": [ + 77, + 265, + 274, + 275 + ], + "spans": [ + { + "bbox": [ + 77, + 265, + 274, + 275 + ], + "type": "text", + "content": "Table 5. Ablation on Text Relation Prediction Module" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 102, + 288, + 249, + 361 + ], + "lines": [ + { + "bbox": [ + 102, + 288, + 249, + 361 + ], + "spans": [ + { + "bbox": [ + 102, + 288, + 249, + 361 + ], + "type": "table", + "html": "
Cluster TypeR-1R-LMETEOR
Kmeans (k=3)40.924.829.8
Kmeans (k=4)40.724.729.6
HAC(n=3)41.024.830.1
DBSCAN41.324.830.6
DL-Cluster42.225.331.3
", + "image_path": "90a67d1390ddb96ca4a0dc3eed782a81e18852686728d04e6f4f81ac52857c07.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 534 + ], + "type": "text", + "content": "tions without any supervision (in r2), the performance only has a slight improvement. It demonstrates the RPH struggles to converge effectively in generating reasonable text relations. In addition, we also explore the approach of directly feeding pseudo relation labels generated by DL-Cluster into the subsequent RAL, instead of utilizing learnable relations (in r3). Although it exhibits substantial improvement, yet it continues to fall short in comparison to leveraging learnable text predictions and establishing weak supervision (r3 vs r4). It shows that learnable relations could be refined during the training process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 535, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 295, + 583 + ], + "type": "text", + "content": "Moreover, we provide visualization results in Fig. 5. We assign the same color for two texts if their relation score is larger than 0.5. The results illustrate that the RPH can predict reasonable text relations under weak supervision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 590, + 290, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 290, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 290, + 601 + ], + "type": "text", + "content": "5.3.2. Different Methods for Pseudo Label Generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Traditional cluster methods are widely used in previous works for relation prediction [1, 8]. Therefore, we compare our proposed DL-Cluster with other cluster methods. As shown in Table 6, DL-Cluster surpasses clustering techniques with predefined cluster numbers like K-means and Hierarchical Agglomerating Clustering (HAC). Additionally, DL-Cluster outperforms methods without specified cluster numbers, such as DBSCAN [6], due to its capacity to consider spatial and semantic aspects concurrently." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 556, + 157 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 556, + 157 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 556, + 157 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 556, + 157 + ], + "type": "image", + "image_path": "7f67174f07536019fe83af6782212bd13b307d919acadf911d3055cf3985747a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 164, + 549, + 175 + ], + "lines": [ + { + "bbox": [ + 318, + 164, + 549, + 175 + ], + "spans": [ + { + "bbox": [ + 318, + 164, + 549, + 175 + ], + "type": "text", + "content": "Figure 6. Qualitative results of Lecture QA in LecSlides-370K." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 353, + 193, + 516, + 245 + ], + "blocks": [ + { + "bbox": [ + 64, + 369, + 286, + 380 + ], + "lines": [ + { + "bbox": [ + 64, + 369, + 286, + 380 + ], + "spans": [ + { + "bbox": [ + 64, + 369, + 286, + 380 + ], + "type": "text", + "content": "Table 6. Ablation on cluster types in pseudo label generation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 193, + 516, + 245 + ], + "lines": [ + { + "bbox": [ + 353, + 193, + 516, + 245 + ], + "spans": [ + { + "bbox": [ + 353, + 193, + 516, + 245 + ], + "type": "table", + "html": "
SpatialSemanticR-1R-LMETEOR
×✓38.323.728.5
✓×41.424.830.7
✓✓42.225.331.3
", + "image_path": "a5f64f3b6b417d21dbaf4fc9020aa24a1f3827b4a9ea73c6a22773010c5354e2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 354, + 254, + 514, + 265 + ], + "lines": [ + { + "bbox": [ + 354, + 254, + 514, + 265 + ], + "spans": [ + { + "bbox": [ + 354, + 254, + 514, + 265 + ], + "type": "text", + "content": "Table 7. Ablation on designs of DL-Cluster" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 293, + 434, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 434, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 434, + 304 + ], + "type": "text", + "content": "5.3.3. Design of DL-Cluster." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 311, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 395 + ], + "type": "text", + "content": "We further investigate the impact of text relations represented in spatial and semantic dimensions. As shown in Tab. 7, removing either of these dimensions leads to a decline in performance, and spatial relations have a larger impact. The best results are achieved when both aspects were used simultaneously, indicating that both dimensions complement each other to enhance overall performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 411, + 398, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 398, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 398, + 422 + ], + "type": "text", + "content": "5.4. Visualization" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 431, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 562 + ], + "type": "text", + "content": "In a side-by-side qualitative analysis for Lecture QA in LecSlides-370K, we compare our proposed method with other baselines in related fields. As shown in Fig. 6, results show that other methods face two challenges: 1) They can not grasp correct internal logic due to the complex text relations. 2) As the number of diagrams on slides increases, other methods struggle to pinpoint the correct diagram corresponding to the question. In contrast, SlideParser excels in accurate responses by comprehending text relations effectively and easily determining the answer's location within the slides." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 388, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 388, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 388, + 594 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "text", + "content": "In this work, we introduce a large-scale and multi-task lecture slides dataset, LecSlides-370K, which has a larger data scale and more diverse disciplines. Furthermore, considering the pivotal role of text relation prediction and utilization for slides understanding, we propose SlideParser. The results demonstrate the effect of enhanced text relation design, and our method outperforms other models. We believe LecSlides-370K and SlideParser can prompt lecture slides research." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4462" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 117 + ], + "type": "text", + "content": "This work was supported by the NSFC (62206103 and 62225603)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 115, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 115, + 141 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 115, + 141 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 148, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 148, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 225 + ], + "type": "text", + "content": "[1] Rhys Agombar, Max Luebbering, and Rafet Sifa. A clustering backed deep learning approach for document layout analysis. In Machine Learning and Knowledge Extraction: 4th IFIP TC 5, TC 12, WG 8.4, WG 8.9, WG 12.9 International Cross-Domain Conference, CD-MAKE 2020, Dublin, Ireland, August 25–28, 2020, Proceedings 4, pages 423–430. Springer, 2020. 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 261 + ], + "type": "text", + "content": "[2] David Arthur and Sergei Vassilvitskii. k-means++: the advantages of careful seeding. In Society for Industrial and Applied Mathematics, page 1027-1035, 2007. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 317 + ], + "type": "text", + "content": "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 396 + ], + "type": "text", + "content": "[4] Zhe Chen, Heyang Liu, Wenyi Yu, Guangzhi Sun, Hongcheng Liu, Ji Wu, Chao Zhang, Yu Wang, and Yanfeng Wang. M3av: A multimodal, multigenre, and multipurpose audio-visual academic lecture dataset. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 9041–9060, Bangkok, Thailand, 2024. Association for Computational Linguistics. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 397, + 296, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 296, + 452 + ], + "type": "text", + "content": "[5] Kartik Dutta, Minesh Mathew, Praveen Krishnan, and CV Jawahar. Localizing and recognizing text in lecture videos. In 2018 16th international conference on frontiers in handwriting recognition (ICFHR), pages 235-240. IEEE, 2018. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 454, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 296, + 509 + ], + "type": "text", + "content": "[6] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, pages 226-231, 1996. 2, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 511, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 296, + 565 + ], + "type": "text", + "content": "[7] Damianos Galanopoulos and Vasileios Mezaris. Temporal lecture video fragmentation using word embeddings. In MultiMedia Modeling: 25th International Conference, MMM 2019, Thessaloniki, Greece, January 8-11, 2019, Proceedings, Part II 25, pages 254-265. Springer, 2019. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 567, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 633 + ], + "type": "text", + "content": "[8] Zhangxuan Gu, Changhua Meng, Ke Wang, Jun Lan, Weiqiang Wang, Ming Gu, and Liqing Zhang. Xylayoutlm: Towards layout-aware multimodal networks for visually-rich document understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4583-4592, 2022. 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 678 + ], + "type": "text", + "content": "[9] Jaekyu Ha, R.M. Haralick, and I.T. Phillips. Recursive x-y cut using bounding boxes of connected components. In Proceedings of 3rd International Conference on Document Analysis and Recognition, pages 952-955, 1995. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "text", + "content": "[10] Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, and Furu Wei. Layoutlmv3: Pre-training for document ai with unified text and image masking. In Proceedings of the 30th ACM" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "International Conference on Multimedia, pages 4083-4091, 2022. 1, 3, 5, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[11] Gautier Izacard and Edouard Grave. Leveraging passage retrieval with generative models for open domain question answering. arXiv preprint arXiv:2007.01282, pages 1-6, 2020. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 205 + ], + "type": "text", + "content": "[12] Geewook Kim, Teakgyu Hong, Moonbin Yim, JeongYeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 250 + ], + "type": "text", + "content": "[13] JP Kincaid. Derivation of new readability formulas (automated readability index, fog count and flesch reading ease formula) for navy enlisted personnel. Chief of Naval Technical Training, pages 1-45, 1975. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 251, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 315 + ], + "type": "text", + "content": "[14] Dong Won Lee, Chaitanya Ahuja, Paul Pu Liang, Sanika Natu, and Louis-Philippe Morency. Lecture presentations multimodal dataset: Towards understanding multimodality in educational videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 20087-20098, 2023. 1, 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 316, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 403 + ], + "type": "text", + "content": "[15] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7871-7880, Online, 2020. Association for Computational Linguistics. 1, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 404, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 502 + ], + "type": "text", + "content": "[16] Chenliang Li, Haiyang Xu, Junfeng Tian, Wei Wang, Ming Yan, Bin Bi, Jiabo Ye, He Chen, Guohai Xu, Zheng Cao, Ji Zhang, Songfang Huang, Fei Huang, Jingren Zhou, and Luo Si. mPLUG: Effective and efficient vision-language learning by cross-modal skip-connections. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7241–7259, Abu Dhabi, United Arab Emirates, 2022. Association for Computational Linguistics. 1, 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 503, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 503, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 503, + 553, + 557 + ], + "type": "text", + "content": "[17] Irene Li, Alexander R Fabbri, Robert R Tung, and Dragomir R Radev. What should i learn first: Introducing lecturebank for nlp education and prerequisite chain learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6674-6681, 2019. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 559, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 591 + ], + "type": "text", + "content": "[18] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 592, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 553, + 624 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, pages 1-19, 2017. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 669 + ], + "type": "text", + "content": "[20] T. Muthuprasad, S. Aiswarya, K.S. Aditya, and Girish K. Jha. Students' perception and preference for online education in india during covid -19 pandemic. Social Sciences Humanities Open, 3(1):100101, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[21] Pitambar Paudel. Online education: Benefits, challenges and strategies during and after Covid-19 in higher education. International Journal on Studies in Education (IJonSE), 3:70–85, 2021. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4463" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[22] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1, 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 195 + ], + "type": "text", + "content": "[23] Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3982-3992. Association for Computational Linguistics, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "text", + "content": "[24] Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53-65, 1987. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 263 + ], + "type": "text", + "content": "[25] Anna Sun and Xiufang Chen. Online education and its effective practice: A research review. Journal of information technology education: Research, 15:1-34, 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "type": "text", + "content": "[26] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13878-13888, 2021. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 365 + ], + "type": "text", + "content": "[27] Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa, Itsumi Saito, and Kuniko Saito. Slidevqa: A dataset for document visual question answering on multiple images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13636-13645, 2023. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 366, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 431 + ], + "type": "text", + "content": "[28] Zineng Tang, Ziyi Yang, Guoxin Wang, Yuwei Fang, Yang Liu, Chenguang Zhu, Michael Zeng, Cha Zhang, and Mohit Bansal. Unifying vision, text, and layout for universal document processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19254-19264, 2023. 1, 3, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "type": "text", + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, page 6000-6010, 2017. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[30] Jiawei Wang, Kai Hu, Zhuoyao Zhong, Lei Sun, and Qiang Huo. Detect-order-construct: A tree construction based approach for hierarchical document structure analysis. arXiv preprint arXiv:2401.11874, pages 1-35, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 535, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 601 + ], + "type": "text", + "content": "[31] Peng Wang, An Yang, Rui Men, Junyang Lin, Shuai Bai, Zhikang Li, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. Ofa: Unifying architectures, tasks, and modalities through a simple sequence-to-sequence learning framework. In International Conference on Machine Learning, pages 23318-23340. PMLR, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "text", + "content": "[32] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1192-1200, 2020. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[33] Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, and Lidong Zhou. LayoutLMv2: Multi-modal pre-training for visually-rich document understanding. In Proceedings of the 59th Annual Meeting of the" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 316 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 117 + ], + "type": "text", + "content": "Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2579-2591, Online, 2021. Association for Computational Linguistics. 1, 3, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 184 + ], + "type": "text", + "content": "[34] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10714-10726, 2023. 1, 3, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 185, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 250 + ], + "type": "text", + "content": "[35] Ori Yoran, Alon Talmor, and Jonathan Berant. Turning tables: Generating examples from semi-structured tables for endowing language models with reasoning skills. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics, pages 6016–6031. Association for Computational Linguistics, 2022. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 251, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 316 + ], + "type": "text", + "content": "[36] Pengchuan Zhang, Xiujun Li, Xiaowei Hu, Jianwei Yang, Lei Zhang, Lijuan Wang, Yejin Choi, and Jianfeng Gao. Vinvl: Revisiting visual representations in vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5579-5588, 2021. 3, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4464" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/5c932632-b5a7-42ba-9baf-b3ca38b774d6_content_list.json b/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/5c932632-b5a7-42ba-9baf-b3ca38b774d6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..af3082f51021bb36003f62a744fb8a24a2514a24 --- /dev/null +++ b/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/5c932632-b5a7-42ba-9baf-b3ca38b774d6_content_list.json @@ -0,0 +1,1567 @@ +[ + { + "type": "text", + "text": "Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models", + "text_level": 1, + "bbox": [ + 194, + 130, + 803, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Young Kyun Jang1", + "bbox": [ + 267, + 203, + 426, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google DeepMind", + "bbox": [ + 267, + 220, + 426, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kyun0914@gmail.com", + "bbox": [ + 267, + 242, + 429, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ser-nam $\\mathrm{Lim}^2$", + "bbox": [ + 553, + 203, + 669, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ University of Central Florida", + "bbox": [ + 491, + 220, + 730, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sernam@gmail.com", + "bbox": [ + 537, + 242, + 681, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern retrieval systems often struggle with upgrading to new and more powerful models due to the incompatibility of embeddings between the old and new models. This necessitates a costly process known as backfilling, which involves re-computing the embeddings for a large number of data samples. In vision, Backward-compatible Training (BT) has been proposed to ensure that the new model aligns with the old model's embeddings. This paper extends the concept of vision-only BT to the field of cross-modal retrieval, marking the first attempt to address Cross-modal BT (XBT). Our goal is to achieve backward-compatibility between Vision-Language Pretraining (VLP) models, such as CLIP, for the cross-modal retrieval task. To address XBT challenges, we propose an efficient solution: a projection module that maps the new model's embeddings to those of the old model. This module, pretrained solely with text data, significantly reduces the number of image-text pairs required for XBT learning, and, once it is pretrained, it avoids using the old model during training. Furthermore, we utilize parameter-efficient training strategies that improve efficiency and preserve the off-the-shelf new model's knowledge by avoiding any modifications. Experimental results on cross-modal retrieval datasets demonstrate the effectiveness of XBT and its potential to enable backfill-free upgrades when a new VLP model emerges.", + "bbox": [ + 89, + 325, + 485, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 737, + 220, + 752 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As the volume and variety of data grow exponentially in our multimedia-rich era, developing and maintaining efficient multi-modal retrieval systems becomes increasingly challenging. These systems, which provide data samples relevant to a user's query, must handle diverse data types, from text and images to audio and video. This growth puts a premium on the scalability and performance of these systems, necessitating continuous advancements in algorithms and technology. Embedding-based deep models for", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e0e2ad6479550047c00477deca97bf74e0efc4a26999cc0a8d59d5b9f88c00db.jpg", + "image_caption": [ + "Backward-compatible Training (BT)" + ], + "image_footnote": [], + "bbox": [ + 519, + 311, + 901, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/014ff1aa669b621ec3ffc021639fc59f131d860af096d99852d58d8164515017.jpg", + "image_caption": [ + "Cross-modal Backward-compatible Training (XBT)", + "Figure 1. A conceptual visualization of Backward-compatible Training (BT, left) and its extension, the Cross(X)-modal version (XBT, right). Circles and squares denote data samples of images and text, respectively. XBT uses Vision-Language Pretraining (VLP) models as baselines to achieve cross-modal, backward-compatible representation learning, allowing the new, improved model to be compatible with the fixed old model." + ], + "image_footnote": [], + "bbox": [ + 516, + 482, + 903, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "retrieval have emerged as a key solution, transforming high-dimensional data into a lower-dimensional dense embedding space [11, 15, 16, 30, 37, 46]. These models capture the semantic meanings of data samples, enabling the quantification of similarities for retrieval.", + "bbox": [ + 511, + 776, + 906, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, it is important to note that the embedding spaces generated by different deep models are not inherently compatible with each other. This incompatibility restricts the", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1783", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "reuse of an existing gallery that has been embedded with an older model when a new, better-performing model is introduced. Consequently, this necessitates \"backfilling,\" a process where the entire gallery must be rebuilt using the embeddings from the new model. Such a requirement is resource-intensive and time-consuming, posing a significant challenge when building retrieval systems.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backward-compatible Training (BT) [13, 34, 38, 43, 44] has been developed to tackle this issue, specifically focusing on image retrieval systems. The main objective of BT is to train a new model from scratch in a manner that ensures its compatibility with an old model that was used to create the existing gallery. A successful BT model must consequently demonstrate that retrieving from the old gallery using a query embedded with the new model outperforms that using an embedding from the old model. This enhancement is crucial for justifying the application of BT.", + "bbox": [ + 89, + 198, + 480, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Taking the BT problem a step further, we propose a new challenging task called Cross(X)-modal Backward-compatible Training (XBT) as shown in Figure 1. Our focus here is on applying BT principles within the realm of cross-modal retrieval, specifically exploring the interaction and compatibility between image and text embeddings of different Vision-Language Pretraining (VLP) models [17, 21, 22, 28, 35].", + "bbox": [ + 89, + 351, + 482, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve XBT, just like the BT problem, we need to resolve the incompatibility between the embeddings of the old, inferior model and the new, improved model. If we follow most prior practices of training from scratch with an additional compatibility loss term, a substantial number of supervised image-text pairs would be needed. The ideal quantity would be in the hundreds of millions, approximating the amount utilized in the establishment of VLP models. However, accessing the original training samples is often impossible, and training on that scale is prohibitively expensive and impractical. We propose an efficient solution instead: a text-only pretrained projection module, $\\phi$ , to align a given pretrained new model's embeddings with those of the old model.", + "bbox": [ + 89, + 474, + 482, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our focus is on using only text data to estimate the entire distribution of embeddings in the VLP space. Specifically, we train $\\phi$ to learn the similarity between old and new text embeddings in a contrastive manner. By increasing the number of text samples, which is simpler than preparing image-text pairs, $\\phi$ can approximate the oracle projection between the intra-modal distributions of both old and new embeddings for texts. Assuming that the intra-modal distribution of texts in the VLP embedding space can mirror that of images, we can simply apply the same $\\phi$ to the new image embeddings to synthesize the corresponding old image embeddings.", + "bbox": [ + 89, + 686, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With these generated synthetic old image and text embeddings, which can be considered as aligned with the new", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "embeddings, we aim to fine-tune the new model. In this process, far fewer supervised image-text pairs are required than in the original training dataset of the VLP. This also allows us to avoid using old model parameters during training, thereby enhancing training efficiency. In addition, we incorporate parameter-efficient fine-tuning schemes [12, 19, 20] into the new model, which add small trainable parameters and do not harm any of the new model's original parameters during training. This not only accelerates the new model's fine-tuning process but also allows for an easy reversion to the original new-to-new retrieval by simply removing the additional parameters.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate our approach on various cross-modal benchmarks, highlighting the effectiveness of XBT in cross-modal retrieval protocols. In response to the rapid advancement of VLP models, XBT offers an environmentally friendly solution by enabling backfill-free systems.", + "bbox": [ + 511, + 275, + 903, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as:", + "bbox": [ + 532, + 352, + 803, + 364 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The Cross-modal Backward-compatible Training (XBT) concept is introduced for the first time to solve the backfilling problem that stems from real-world cross-modal retrieval systems.", + "- A novel XBT solution is proposed, which uses a text-only pretrained projection module, $\\phi$ , to efficiently align the embeddings of new and old models using only text samples.", + "- The proposal is demonstrated on various datasets and protocols, showcasing XBT's effectiveness in building backfill-free cross-modal retrieval systems." + ], + "bbox": [ + 513, + 369, + 903, + 534 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 556, + 661, + 570 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backward-compatible Training. The concept of Backward-compatible Training (BT) was first introduced in the study ([34]). This approach influences a new model with the learned classifier of the old model, thereby achieving backward compatibility between the old and new models. However, BT can degrade the original representational performance of the new model. To overcome this, [26] proposed aligning class-wise centers presented by the old and new models. Another approach to achieve backward compatibility is to use an additional projection to map the old embedding into the new embedding space by adding a lightweight transformation, as suggested in [13, 38]. The work of [29] further adds an auxiliary feature in preparation for future updates, while [47] uses additional dimensionality in the embedding to obtain compatibility. An online strategy that backfills the gallery on the fly is explored in [43], and [42] addresses the model regression problem. Despite this progress, the scenario considering cross-modal retrieval between image and text, which has many real-world applications, remains unexplored. Our study on XBT in this paper aims to fill this gap.", + "bbox": [ + 511, + 583, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1784", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Continual and Transfer Learning The fields of continual learning [2, 5, 7] and transfer learning [25, 48] bear similarities to backward-compatible representation learning, as all aim to update an existing model to boost performance. In the realm of multi-modal representation learning for cross-modal retrieval, continual learning approaches like [40] propose methods to prevent catastrophic forgetting across different modalities. Transfer learning approaches like [45] suggest strategies for transferring knowledge from previously labeled categories (source modality) to new, unlabeled categories (target modality). However, our proposed XBT stands apart from these methods as it specifically tackles the challenge of maintaining backward compatibility between old and new models. This unique attribute makes XBT ideally suited for backfill-free retrieval scenarios, where the embeddings of the old model remain unchanged, yet we can still leverage the enhanced performance of the new model.", + "bbox": [ + 89, + 90, + 486, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 376, + 181, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to construct a backfill-free, embedding-based, cross-modal retrieval system using VLP models, which are configured with an image encoder $E^{I}$ and a text encoder $E^{T}$ . When a new better performing VLP model $\\{E_{new}^{I}, E_{new}^{T}\\}$ emerges, we aim to ensure its compatibility with the old model $\\{E_{old}^{I}, E_{old}^{T}\\}$ that was used to construct the gallery. To achieve this, we introduce Cross(X)-modal Backward-compatible Training (XBT).", + "bbox": [ + 89, + 401, + 483, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an image $x$ and text caption $t$ , XBT enables retrieval between a new image embedding $v_{new} = E_{new}^{I}(x)$ and the text embeddings $(w_{old})$ , as well as between a new text embedding $w_{new} = E_{new}^{T}(t)$ and the image embeddings $(v_{old})$ . We denote backward compatible embeddings as $\\bar{v}$ and $\\bar{w}$ . All embeddings we utilize in this work are l2-normalized.", + "bbox": [ + 89, + 522, + 483, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Criterion for Cross-modal Backward Compatibility", + "text_level": 1, + "bbox": [ + 89, + 637, + 482, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the definition of backward compatibility in BT work [34], we can construct strict constraints that ensure cross-modal backward compatibility as:", + "bbox": [ + 89, + 674, + 483, + 720 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d \\left(w _ {n e w _ {i}}, v _ {o l d _ {j}}\\right) \\leq d \\left(w _ {o l d _ {i}}, v _ {o l d _ {j}}\\right), \\forall y _ {i} = y _ {j}, \\\\ d \\left(w _ {\\text {n e w} _ {i}}, v _ {\\text {o l d} _ {j}}\\right) \\geq d \\left(w _ {\\text {o l d} _ {i}}, v _ {\\text {o l d} _ {j}}\\right), \\forall y _ {i} \\neq y _ {j}, \\tag {1} \\\\ d (v _ {n e w _ {i}}, w _ {o l d _ {j}}) \\leq d (v _ {o l d _ {i}}, w _ {o l d _ {j}}), \\forall y _ {i} = y _ {j}, \\\\ d (v _ {n e w _ {i}}, w _ {o l d _ {j}}) \\geq d (v _ {o l d _ {i}}, w _ {o l d _ {j}}), \\forall y _ {i} \\neq y _ {j}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 742, + 482, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y_{i}$ and $y_{j}$ represent whether an image and text are paired ( $y_{i} = y_{j}$ ) or not ( $y_{i} \\neq y_{j}$ ). The term $d(\\cdot, \\cdot)$ represents a distance metric in the embedding space, and we choose cosine distance as the baseline. The constraints in Eqn. 1 formally express the requirement that the new embedding", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "must perform at least as well as the old embedding in terms of correctly matching image-text pairs.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, exhaustively satisfying these constraints is intractable due to the potential for the old embeddings to outperform the new embeddings in certain retrieval cases. As a result, we relax the criteria by defining an alternative evaluation metric as:", + "bbox": [ + 511, + 122, + 908, + 196 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\begin{array}{l} \\mathcal {M} (w _ {\\text {n e w}}, v _ {\\text {o l d}}; \\mathrm {Q}, \\mathrm {G}) > \\mathcal {M} (w _ {\\text {o l d}}, v _ {\\text {o l d}}; \\mathrm {Q}, \\mathrm {G}), \\\\ \\mathcal {A} ^ {\\prime} (\\mathrm {Q}, \\mathrm {G}) = \\mathcal {A} ^ {\\prime} (\\mathrm {Q}, \\mathrm {G}) \\end{array} \\tag {2} \\\\ \\mathcal {M} (v _ {n e w}, w _ {o l d}; \\mathrm {Q}, \\mathrm {G}) > \\mathcal {M} (v _ {o l d}, w _ {o l d}; \\mathrm {Q}, \\mathrm {G}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 224, + 906, + 260 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where a metric, such as recall, $\\mathcal{M}(q,g;\\mathrm{Q},\\mathrm{G})$ takes query embedding $q$ and gallery embedding $g$ over query set Q and gallery set G. This criterion suggests that the general performance is enhanced when the query embedding from the XBT-trained new model is used to perform retrieval with the gallery of the old model, compared to the performance of the old model alone. In essence, fulfilling Eqn. 2 indicates that the new model has achieved backward compatibility and can feasibly update without backfilling gallery.", + "bbox": [ + 511, + 273, + 906, + 410 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, it is important to highlight that VLP models function as zero-shot learners. This leads us to define the XBT problem differently from classic BT, which trains separate models for specific image domain retrieval tasks such as ImageNet [31] or VGGFace2 [4]. In contrast, XBT tackles a more challenging task, aiming to bridge a new VLP model with a frozen old VLP model while retaining the new model's original zero shot capability. In this paper, we therefore assess the performance of XBT using retrieval and classification benchmarks in a zero-shot manner.", + "bbox": [ + 511, + 411, + 908, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Text-only Pretraining", + "text_level": 1, + "bbox": [ + 511, + 574, + 718, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The vast and diverse image-text pairs used to build VLP models significantly enhance their ability to connect semantically similar visual content and natural language [6]. However, this creates a complex embedding distribution that is difficult to predict and understand, thereby complicating the XBT process. A straightforward solution is to use a large number of supervised image-text pairs, similar to the approach used when building VLP models from scratch, to estimate the entire distribution of image and text embeddings during XBT.", + "bbox": [ + 511, + 597, + 908, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, acquiring a sufficient number of accurate supervised pairs is extremely costly. To be far more efficient, we employ a small sized projection module, $\\phi$ , and train it exclusively with text samples, as shown in Figure 2. We hypothesize that the distribution of text embeddings in VLP models, which is determined by their semantic similarity, is similarly mirrored in the distribution of their corresponding matched images. With this in mind, we aim to train $\\phi$ to cover the broad spectrum of the text embedding distribution between the new and old VLP models, making embeddings", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1785", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/39336d708d8dd5f82225376033f56341510ecb6f23cff56980107b9b142ccf28.jpg", + "image_caption": [ + "Text-only Pretraining" + ], + "image_footnote": [], + "bbox": [ + 93, + 104, + 483, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5540558dc3f4240ed4f4e8c844614f06cb25c9e88fc1897a1a3ed78cf7071704.jpg", + "image_caption": [ + "Figure 2. An illustration of the text-only pretraining of $\\phi$ (above), and XBT with $\\phi$ (below). Using only text samples, $\\phi$ is trained to approximate distribution of old text embeddings from that of new ones. After pretraining, the same $\\phi$ is used to generate both of synthetic old image and text embeddings from the new VLP embeddings and train to learn cross-modal backward-compatible representation." + ], + "image_footnote": [], + "bbox": [ + 93, + 227, + 483, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from the same text sample similar and others dissimilar in a contrastive way:", + "bbox": [ + 89, + 540, + 482, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p r e} = \\mathbb {E} _ {t \\sim D _ {T}} \\left[ \\mathcal {L} _ {c} \\left(\\phi \\left(w _ {\\text {n e w}} + \\epsilon\\right), w _ {\\text {o l d}}; \\tau_ {p r e}\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 603, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ denotes text data sampled from text corpus $D_{T} = \\{t_{i}\\}_{i=1}^{N_{D_{T}}} \\text{ of } N_{D_{T}} \\text{ text samples, and } \\mathcal{L}_{c}(\\cdot, \\cdot) \\text{ is a standard batch-wise contrastive loss as defined in CLIP [28] of temperature } \\tau_{pre}.$", + "bbox": [ + 89, + 637, + 483, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To prevent overfitting towards the text domain, which could widen the discrepancy between the image and text embedding representations, we inject noise $\\epsilon$ sampled from a Gaussian distribution $(\\epsilon \\sim N(0,\\sigma^2))$ , inspired by the approaches [8, 27]. Notably, $\\phi$ is designed to take the new VLP model's embeddings, which means that $w_{new} + \\epsilon$ is also $l2$ -normalized.", + "bbox": [ + 89, + 702, + 483, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During this stage, only $\\phi$ is updated, while all encoders remain fixed. By expanding the scale of $D_{T}$ , we can effectively replicate the complete text embedding spaces of both new and old VLP models. Additionally, $\\phi$ can be applied to new image embeddings to generate synthetic old image embeddings that closely approximate the actual old ones.", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b9379e2bd6f070cf88c957d8a78f1f13e57f99c4976019a8179f7966262acf8b.jpg", + "image_caption": [ + "XBT Learning Process", + "Figure 3. Our proposed learning process to achieve XBT. Notably, the old VLP model's encoders are not required in this stage, enhancing efficiency in training." + ], + "image_footnote": [], + "bbox": [ + 516, + 99, + 903, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Cross-modal Backward-compatible Training", + "text_level": 1, + "bbox": [ + 511, + 229, + 892, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Loss. The objective of XBT is to ensure cross-modal compatibility, specifically between $w_{new}$ and $v_{old}$ , as well as between $v_{new}$ and $w_{old}$ . For a given dataset $D = \\{x_i, t_i\\}_{i=1}^{N_D}$ of $N_D$ supervised image-text pairs, we aim to train the new VLP model encoders, $E_{new}^I$ and $E_{new}^T$ . Note that in our base setting, the text corpus is significantly larger than the supervised dataset, i.e., $N_{D_T} \\gg N_D$ .", + "bbox": [ + 511, + 251, + 906, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the dimensionality of new and old VLP embeddings may differ, and even if they are the same, they are not be directly compatible. To address this, we apply the pretrained $\\phi$ to ensure that the new embeddings match the dimension of the old ones and project into the compatible embedding space, as shown in Figure 2. This can be represented as:", + "bbox": [ + 511, + 357, + 908, + 462 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (v _ {\\text {n e w}}) = \\bar {v} _ {\\text {n e w}}, \\phi (w _ {\\text {n e w}}) = \\bar {w} _ {\\text {n e w}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 474, + 906, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\bar{v}_{new}$ and $\\bar{w}_{new}$ denote synthetic old embeddings and $\\{\\bar{v}_{new},\\bar{w}_{new},v_{old},w_{old}\\} \\subset \\mathbb{R}^K$ , with $K$ denoting the dimensionality of the old VLP model's embedding.", + "bbox": [ + 511, + 501, + 906, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then apply XBT loss $\\mathcal{L}_X$ as follows:", + "bbox": [ + 532, + 546, + 800, + 560 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {X} = \\mathbb {E} _ {x, t \\sim D} \\left[ \\mathcal {L} _ {c} \\left(\\bar {v} _ {\\text {n e w}}, \\bar {w} _ {\\text {n e w}}; \\tau_ {X}\\right) \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 571, + 906, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathcal{L}_c(\\cdot ,\\cdot)$ is the same contrastive loss used in Eqn. 3, with temperature $\\tau_{X}$ . As shown in Figure 3, the new VLP encoders, $E_{new}^{I}$ and $E_{new}^{T}$ , and $\\phi$ are trained through $\\mathcal{L}_X$ in an end-to-end manner. Ultimately, $\\bar{v}_{new}$ and $\\bar{w}_{new}$ become cross-modal backward-compatible with existing old gallery embeddings, $v_{old}$ and $w_{old}$ , as all embeddings are distributed in the compatible space through $\\phi$ .", + "bbox": [ + 511, + 598, + 906, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Efficient Training. The efficiency of XBT can be attributed to two design choices that avoid: (1) any dependence on the old model when conducting XBT, and (2) updating the new VLP model off the shelf. We have already seen how $\\phi$ can effectively achieve (1). (2) is simply facilitated by applying small-sized additional parameters, namely Soft prompt [20] and LoRA[12].", + "bbox": [ + 511, + 704, + 905, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To be more specific, we incorporate the concept of soft prompt tuning, as outlined in [18, 20]. Soft prompting is applied to the vision encoder $E_{new}^{I}$ by adding trainable prompts as input, along with image patch tokens. We exclude soft prompt tuning on the text side, as we use the text-only pretrained module $\\phi$ , which requires maintaining the original", + "bbox": [ + 511, + 810, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1786", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "distribution of text embeddings. We adopt the LoRA strategy [12] for the new VLP model's encoders $E_{new}^{I}$ and $E_{new}^{T}$ to avoid modifying original parameters. These two factors, in conjunction with $\\phi$ , offer an on-off solution: we can retrieve old samples using additional parameters, and we can easily revert to the original model by removing these parameters for new-to-new retrieval.", + "bbox": [ + 89, + 90, + 480, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further, regarding $\\phi$ during XBT (Sec. 3.3), we only finetune the parameters of layer normalization. This is done with the aim of preserving the knowledge learned during the text-only pretraining stage (Sec. 3.2) and facilitating easy adaptation towards the image-text joint representation space", + "bbox": [ + 89, + 196, + 480, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the end, our XBT framework offers a parameter-efficient solution that enables rapid convergence with fewer training iterations (a single epoch is sufficient); maintains the new model's power; and requires far fewer supervised training samples than the scale of VLP model pretraining (approximately $1\\%$ of the level required to train CLIP [28] from scratch, which uses a 400M dataset, and $0.2\\%$ of the level required for LAION-2B based CLIP models [32], which utilize a 2B dataset). All these help to avoid any full-scale retraining of the new VLP models, paving the way for XBT to simply use new VLP models off-the-shelf as mentioned.", + "bbox": [ + 89, + 272, + 480, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 452, + 220, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The evaluation of the proposed cross-modal backward compatibility includes image-text zero-shot retrieval. Pretrained VLP models are used as our baseline, with the goal to assess performance in a zero-shot environment by tuning a given pretrained new VLP model that is stronger than the old model. The results highlight the potential of XBT, showing improved performance when old model outputs are matched with the XBT-learned new model.", + "bbox": [ + 89, + 476, + 480, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setup", + "text_level": 1, + "bbox": [ + 89, + 606, + 171, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model Training. Our XBT process involves two distinct training stages: text-only pretraining (See Sec. 3.2) and image-text supervised training (See Sec. 3.3). For the text-only pretraining stage, we utilize the text samples from a subset of the 115M filtered web-collected image-text paired dataset from BLIP [21], comprising around 67M available pairs (58.2% of the total). We construct a text corpus $D_{T}$ in Eqn. 3 using these text samples. For the subsequent image-text supervised training stage, we use a smaller subset of 4M image-text pairs from the same dataset to construct a supervised dataset $D$ in Eqn. 5. It is important to note that these subsets are significantly smaller than the dataset scale (400M, 2B or more) used to build CLIP models [28, 32].", + "bbox": [ + 89, + 628, + 482, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize 8-A100-80GB GPUs to train and evaluate the models. For the text-only pretraining stage in Sec. 3.2, the batch size is set to 8,192 (1,024 batch per GPU) and during this stage, the entire set of trainable parameters of $\\phi$ are trained while both VLP text encoders are fixed.", + "bbox": [ + 89, + 824, + 482, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moving on to the image-text supervised training in Sec. 3.3, the batch size is reduced to 1,024 (128 batch per GPU), and in this stage, layer normalization is set to be the only trainable component for all VLP image and text encoders with $\\phi$ . Temperature hyper-parameters $\\tau_{pre}$ , $\\tau_{X}$ , and $\\tau_{N}$ are fixed at 2.6592. We employ the AdamW optimizer [24] with a fixed learning rate of 1e-4 and a weight decay of 0.01. For soft prompts, we use 10 prompts and apply 100 times larger learning rate, 1e-2. The training iteration is determined by the dataset size, and the entire pipeline is trained for a single epoch. For image augmentation, we begin with a random resized crop, adjusting the image scale between 0.5 and 1.0. Additionally, we apply a random horizontal flip and make random adjustments to the image's contrast, brightness, and sharpness. To incorporate different perspectives and angles, we modify the image's translation and rotation. The pretrained weights provided by HuggingFace [41] are applied to baseline VLP models. We employ as: openai/clip-vit-base-patch32, openai/clip-vit-target-patch14, laion/CLIP-ViT-H-14-laion2B-s32B-b79K.", + "bbox": [ + 511, + 90, + 906, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model Evaluation. We validate the effectiveness of XBT with a benchmark comprising three popular image-text paired datasets for cross-modal retrieval evaluation. The first is the nocaps dataset [1], which offers a diverse distribution of object names and facilitates a more detailed analysis. We employ the validation split of this dataset, which consists of images each paired with 10 relevant textual captions, totaling 4,500 images and 45,000 captions.", + "bbox": [ + 511, + 429, + 906, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate XBT at a larger scale, we also include the Flickr [14] and COCO [23] datasets, which consist of images that are each paired with 5 relevant textual captions. For the Flickr dataset, we use the entire dataset, encompassing 31,783 images and 158,915 captions. For the COCO dataset, we use the validation split, which includes 35,136 images and 175,680 captions. We note the two datasets as Large-Flickr and Large-COCO respectively for the remainder of the paper.", + "bbox": [ + 511, + 551, + 906, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For retrieval, we utilize all the samples in each dataset for both query and gallery, following the cross-modal benchmark used in [21, 28]. For evaluation purpose, we adopt recall scores at top K retrieval results $(\\mathbb{R}@\\mathbb{K},\\%)$ to estimate the cross-modal backward compatibility (See Eqn. 2).", + "bbox": [ + 511, + 686, + 906, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. In our work, we primarily use the popular CLIP models [28], based on a Transformer [36] backbone, as our baseline VLP models. The models, listed in ascending order of scale and performance, are: CLIP-ViT-B32, CLIP-ViT-L14, and CLIP-ViT-H14. For simplicity, we will refer to them as B32, L14, and H14, respec", + "bbox": [ + 511, + 785, + 908, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://huggingface.co/models", + "bbox": [ + 531, + 886, + 699, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1787", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a46a815a5fbb8e08a6bc535f9d97094a66d948a48eca668339e185e4c67b03e9.jpg", + "table_caption": [ + "Table 1. Cross-modal retrieval results on nocaps. Note that for comparison, the B32/L14 case should be compared with B32, and the L14/H14 case should be compared with L14, respectively." + ], + "table_footnote": [], + "table_body": "
Old Model / New ModelMethodText Query(w, w)/Image Gallery(v)Image Query(v, v)/Text Gallery(w)
CaseR@1R@5R@10R@50CaseR@1R@5R@10R@50
Original
CLIP-ViT-B32-w/v45.1474.9884.5196.00v/w71.3892.0296.3399.67
CLIP-ViT-L14-w/v47.7776.5085.1395.95v/w73.2993.2497.4099.78
Cross-modal Backward Compatible Training
CLIP-ViT-B32 / CLIP-ViT-L14Full-tune\\( \\bar{w}_{new}/v_{old} \\)41.6073.5484.2696.45\\( \\bar{v}_{new}/w_{old} \\)63.7687.0093.5199.09
LoRA-only43.4074.6684.9496.7368.8290.8796.0099.76
Base43.4874.8785.0496.7469.7891.3196.0299.78
XBT48.0279.0088.2197.6675.0293.2797.3199.91
Full-tune\\( \\bar{w}_{new}/\\bar{v}_{new} \\)48.9279.3188.2697.64\\( \\bar{v}_{new}/\\bar{w}_{new} \\)61.2285.8992.6299.04
LoRA-only55.4483.8291.2498.3470.1191.1396.4999.73
Base56.1884.3191.4498.3771.0791.4496.2299.76
XBT63.4888.6193.9898.8677.2294.5697.9199.89
CLIP-ViT-L14 / CLIP-ViT-H14Full-tune\\( \\bar{w}_{new}/v_{old} \\)45.1877.8285.3497.65\\( \\bar{v}_{new}/w_{old} \\)66.3888.5593.5998.90
LoRA-only47.0076.9486.0297.6372.0492.4296.5899.67
Base47.0677.1586.1297.6473.0092.8696.6099.69
XBT51.1480.8288.9398.7276.6294.1197.5899.84
Full-tune\\( \\bar{w}_{new}/\\bar{v}_{new} \\)54.3982.2689.0698.95\\( \\bar{v}_{new}/\\bar{w}_{new} \\)65.5787.5293.5199.13
LoRA-only60.9186.7792.0498.6574.4692.7397.3899.82
Base61.6587.2692.2498.6875.4293.0497.1199.85
XBT66.4790.2194.9299.0580.0296.1198.87100.0
", + "bbox": [ + 150, + 128, + 844, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tively. Notably, while B32 and L14 are trained with the same dataset at 400M scale, H14 is trained with a larger dataset at 2B scale. We apply additional LoRA [12] parameters for each new model encoder configured as follows: $\\mathrm{LoRA}_{\\alpha} = 16$ , $\\mathrm{rank} = 16$ , and dropout $= 0.1$ . The proposed projection module $\\phi$ consists of three Linear layers with layer normalization (LN) [3] and the GELU nonlinearity function [9]. Detailed architecture is as follows: Linear-LN-GELU-Linear-LN-GELU-Linear. Dropout is not applied as it was found to degrade performance empirically. The intermediate hidden dimension of the Linear layer is set to be four times the dimensionality of the output embedding.", + "bbox": [ + 88, + 462, + 486, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 667, + 230, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Comparisons. To simplify notation, we use $w / v$ to denote the retrieval results obtained using $w$ as the query embeddings and $v$ as the gallery embeddings. We establish two protocols: assessing cross-modal backward compatibility with new/old ( $\\bar{w}_{\\text{new}} / v_{\\text{old}}$ , $\\bar{v}_{\\text{new}} / w_{\\text{old}}$ ), and maintaining the new VLP's original performance with new/new ( $\\bar{w}_{\\text{new}} / \\bar{v}_{\\text{new}}$ , $\\bar{v}_{\\text{new}} / \\bar{w}_{\\text{new}}$ ). All of the above are similarly applied to $v / w$ .", + "bbox": [ + 89, + 688, + 482, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To the best of our knowledge, since this paper presents the first method that aims to solve the XBT problem, there are no direct previous works to compare to. Nevertheless, we compare our XBT against three baselines: Full-tune, LoRA-only, and Base, which we elaborate on below.", + "bbox": [ + 89, + 810, + 483, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Naive solution - Direct backward contrastive learning:", + "bbox": [ + 109, + 886, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The pretrained $\\phi$ in Eqn. 3, which maps the new embeddings to the old via text based contrastive learning, constitutes a major contribution of XBT. We therefore would like to compare with a solution that does not utilize such a pretrained $\\phi$ . A straightforward solution for this is to fine-tune the new VLP encoders $(E_{new}^{I}$ and $E_{new}^{T})$ to be compatible with the old ones by minimizing the following loss:", + "bbox": [ + 511, + 462, + 908, + 570 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {D i r e c t} = \\mathbb {E} _ {x, t \\sim D} [ \\mathcal {L} _ {c} (\\bar {v} _ {\\text {n e w}}, w _ {\\text {o l d}}; \\tau_ {N}) \\\\ \\left. + \\mathcal {L} _ {c} \\left(\\bar {w} _ {\\text {n e w}}, v _ {\\text {o l d}}; \\tau_ {N}\\right) \\right] \\tag {0} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 583, + 905, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_c(\\cdot, \\cdot)$ is the same contrastive loss used in Eqn. 5, and $\\tau_N$ is a temperature. Here, we apply a randomly initialized $\\phi$ with the same configuration as XBT for projection, as per Eqn. 4, to generate cross-modal backward compatible embeddings, $\\bar{w}$ and $\\bar{v}$ .", + "bbox": [ + 511, + 627, + 906, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different training options: Based on Eqn. 6, we consider multiple training setups: (1) full fine-tuning of all trainable components (Full-tune), (2) tuning LoRA parameters-only (LoRA-only), and (3) starting from LoRA-only, adding an extra learnable prompt (Base). Models produced by Base would therefore be akin to XBT without the pretrained $\\phi$ . For these, we train the $E_{new}^{I}$ , $E_{new}^{T}$ and $\\phi$ using $\\mathcal{L}_{Direct}$ with 4M image-text pairs. Note that, XBT only trains layer-normalization layers including pretrained $\\phi$ , however, Full-tune, LoRA-only and Base setups are training entire parameters of $\\phi$ since it is not trained before. The results of these baselines are presented in Table 1, tested with nocaps dataset. The performance of the original VLP models is also reported", + "bbox": [ + 511, + 704, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1788", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8481aad00a18e3d66c1700cea5c38c6e32126392e3e6dd04b332633546f5a4bd.jpg", + "table_caption": [ + "Table 2. Cross-modal retrieval results on Large-Flickr and Large-COCO." + ], + "table_footnote": [], + "table_body": "
DatasetOld Model / New ModelMethodText Query(w, w)/Image Gallery(v)Image Query(v, v)/Text Gallery(w)
CaseR@1R@5R@10R@50CaseR@1R@5R@10R@50
Original
CLIP-ViT-B32-w/v21.5441.2450.7272.69v/w40.3564.4273.3989.91
Cross-modal Backward Compatible Training
CLIP-ViT-B32 / CLIP-ViT-L14Base\\( \\bar{w}_{\\text{new}}/v_{\\text{old}} \\)19.0237.2546.6369.32\\( \\bar{v}_{\\text{new}}/w_{\\text{old}} \\)36.7460.6170.2288.34
XBT22.3842.7152.4174.8042.4766.0474.8790.79
Base\\( \\bar{w}_{\\text{new}}/\\bar{v}_{\\text{new}} \\)32.2955.1864.6883.60\\( \\bar{v}_{\\text{new}}/\\bar{w}_{\\text{new}} \\)36.1059.8069.9589.01
XBT39.5962.8971.7588.1343.5068.5077.9993.21
CLIP-ViT-B32 / CLIP-ViT-H14Base\\( \\bar{w}_{\\text{new}}/v_{\\text{old}} \\)20.2739.1748.5970.67\\( \\bar{v}_{\\text{new}}/w_{\\text{old}} \\)35.5359.4469.0687.67
XBT23.7044.4754.1976.0341.0965.1874.3290.16
Base\\( \\bar{w}_{\\text{new}}/\\bar{v}_{\\text{new}} \\)32.9455.9465.2683.77\\( \\bar{v}_{\\text{new}}/\\bar{w}_{\\text{new}} \\)34.8159.5569.9688.81
XBT40.4263.9072.7488.4746.3571.1880.0194.08
Original
CLIP-ViT-B32-w/v14.4430.2038.9662.30v/w28.6250.1759.6780.22
Cross-modal Backward Compatible Training
CLIP-ViT-B32 / CLIP-ViT-L14Base\\( \\bar{w}_{\\text{new}}/v_{\\text{old}} \\)12.9728.1436.4859.78\\( \\bar{v}_{\\text{new}}/w_{\\text{old}} \\)26.3246.9156.5477.86
XBT15.5532.2741.3465.1130.7352.5862.3081.83
Base\\( \\bar{w}_{\\text{new}}/\\bar{v}_{\\text{new}} \\)22.1641.6550.9873.06\\( \\bar{v}_{\\text{new}}/\\bar{w}_{\\text{new}} \\)27.2648.1657.8079.57
XBT27.6148.6958.0878.8332.9255.3764.8984.67
CLIP-ViT-B32 / CLIP-ViT-H14Base\\( \\bar{w}_{\\text{new}}/v_{\\text{old}} \\)13.8829.5538.2361.57\\( \\bar{v}_{\\text{new}}/w_{\\text{old}} \\)26.1946.7556.3077.96
XBT15.8232.7941.8465.4530.3351.2859.9781.66
Base\\( \\bar{w}_{\\text{new}}/\\bar{v}_{\\text{new}} \\)23.2043.2252.7074.68\\( \\bar{v}_{\\text{new}}/\\bar{w}_{\\text{new}} \\)27.3148.5658.6080.09
XBT28.1449.4858.8879.3234.5457.2966.7985.92
", + "bbox": [ + 148, + 114, + 841, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "for a clear comparison. We highlight our XBT method.", + "bbox": [ + 89, + 472, + 464, + 487 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the context of cross-modal backward compatibility, where recall scores of $\\bar{w}_{\\text{new}} / v_{\\text{old}}$ and $\\bar{v}_{\\text{new}} / w_{\\text{old}}$ should be higher than those of $w / v$ and $v / w$ of old model, respectively, Full-tune significantly underperforms compared to the old VLP's. This indicates that full fine-tuning is not an appropriate solution. While both LoRA-only and Base improve performance over Full-tune, they still fall short of the old VLP's performance and fail to meet the criterion. In contrast, XBT significantly outperforms these baselines and even surpasses the old VLP's performance, satisfying the cross-modal backward compatibility for all recall metrics (see Eqn. 2).", + "bbox": [ + 88, + 489, + 482, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In terms of maintaining new VLP's performance, where recall scores of $\\bar{w}_{\\text{new}} / v_{\\text{new}}$ and $\\bar{v}_{\\text{new}} / w_{\\text{new}}$ could be similar to those of $w / v$ and $v / w$ of new model, respectively, all baselines show decent performance in $\\bar{w}_{\\text{new}} / v_{\\text{new}}$ case but only XBT achieves better performance than new VLP model in $\\bar{v}_{\\text{new}} / w_{\\text{new}}$ case. These results support that XBT not only maintains the performance of the new VLP model but also enhances it in certain cases. The superior performance of XBT in both cases underscores its promise as a robust solution for cross-modal retrieval tasks.", + "bbox": [ + 89, + 672, + 482, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Large Scale Retrieval. In this experiment, we assess the performance of cross-modal backward compatibility in more practical use cases. We use larger datasets, namely LargeFlickr and Large-COCO, as shown in Table 2. These datasets provide more than six times the number of samples typically", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "used in image-text retrieval literature [6, 37], which make the retrieval process more challenging. We use the second-best performing Base method from Table 1 as a basis for comparison. To validate the applicability of XBT across various VLP models, we utilize three CLIP baselines: B32, L14, and H14. Additionally, we explore different combinations, such as B32/L14 and B32/H14, to further analyze the effectiveness of XBT, beyond the comparisons made in Table 1 (B32/L14 and L14/H32).", + "bbox": [ + 511, + 472, + 906, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When comparing B32/L14 with B32/H14, it is observed that every retrieval cases of improved models (L14 and H14) successfully maintain cross-modal backward compatibility, regardless of their scale. Especially, retrieval performance of B32/H14 is better than that of B32/L14, which demonstrates XBT's ability to preserve and align original model's power with old gallery.", + "bbox": [ + 511, + 608, + 908, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Upon examining the results, we consistently observe that XBT outperforms the Base model across all instances and both retrieval scenarios, often by a significant margin. Furthermore, XBT maintains the performance of new VLPs better than Base, achieving even larger margins for $\\bar{w}_{new} / \\bar{v}_{new}$ and $\\bar{v}_{new} / w_{old}$ . In summary, XBT consistently demonstrates robust and superior retrieval performance, even in large-scale scenarios.", + "bbox": [ + 511, + 715, + 908, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 847, + 681, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Continual learning. In line with the literature on BT works [29, 34], we set up a continual learning scenario for a se", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1789", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9fe6b79b35257d7ed04243d1d2c9b6d39c3baba7c8e192101b91f0d1c115e6ce.jpg", + "table_caption": [ + "Table 3. Continual learning scenario experimental results with XBT on nocaps. Note that for H14 training, we use L14, which has been previously adapted with B32, and H14 never encounters B14." + ], + "table_footnote": [], + "table_body": "
Old Model / New ModelText Query(w, w)/Image Gallery(v)Image Query(v, v)/Text Gallery(w)
CaseR@1R@5R@10R@50CaseR@1R@5R@10R@50
CLIP-ViT-B32 / CLIP-ViT-L14wnew/vold47.6878.0088.0097.50vnew/wold73.8992.0097.2299.86
wnew/vnew62.2286.1893.0098.12vnew/wnew76.2792.9096.7699.11
CLIP-ViT-L14 / CLIP-ViT-H14wnew/vold50.9780.6888.9897.72vnew/wold77.3894.8297.8099.84
wnew/vnew66.0989.7894.7799.01vnew/wnew80.6096.3698.7199.96
CLIP-ViT-B32 / CLIP-ViT-H14wnew/vold49.1479.8888.5997.67vnew/wold74.8493.7397.2299.80
wnew/vnew65.9589.6894.6898.99vnew/wnew80.9696.0098.7199.96
", + "bbox": [ + 163, + 128, + 834, + 261 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86a116e654f354f096d1759859cedcff74f79fd43916b63a565fc65f1286ac48.jpg", + "image_caption": [ + "Figure 4. New query vs. Old gallery retrieval results on nocaps. B32 as old, and L14 as new model." + ], + "image_footnote": [], + "bbox": [ + 98, + 272, + 183, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Retrieved from Old Text Gallery", + "text_level": 1, + "bbox": [ + 263, + 271, + 397, + 281 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "R#1 - A dog with colorful sunglasses on and a bandanna.", + "R#2 - A happy dog is wearing a blue bandanna and sunglasses.", + "R#3 - A dog wearing sunglasses and a blue and white bandanna.", + "R#4 - An old brown dog is wearing mirrored sunglasses and a blue bandana.", + "R#5 - A dog has its tongue out as it wears sunglasses and bandana." + ], + "bbox": [ + 196, + 282, + 485, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "New Query Text", + "text_level": 1, + "bbox": [ + 504, + 279, + 576, + 287 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "The cute brown", + "tabby and white", + "kitten, sits on the", + "hard wood floor." + ], + "bbox": [ + 504, + 279, + 576, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Retrieved from Old Image Gallery", + "text_level": 1, + "bbox": [ + 663, + 271, + 805, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/776d7b9f9867f9f5e67996b4f82968b0904338f89af2208207e70a89a081c620.jpg", + "image_caption": [ + "R#1" + ], + "image_footnote": [], + "bbox": [ + 594, + 281, + 900, + 327 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2d99f85d593a7494acf5348176395bf299f88b7a2b1f39ee7a29717f560dd19a.jpg", + "table_caption": [ + "Table 4. Ablation study tested on nocaps, with B32 as old and L14 as new model." + ], + "table_footnote": [], + "table_body": "
Method\\(\\bar{w}(\\text{text})/v(\\text{image})\\)\\(\\bar{v}(\\text{image})/w(\\text{text})\\)
R@1R@10R@1R@10
(a) Baseline48.0279.0075.0293.27
(b) w.o noise47.2478.0573.3192.49
(c) \\(0.25 \\times D_T\\)46.8977.9173.2092.18
(d) \\(0.5 \\times D_T\\)47.6878.0073.8992.00
(e) \\(2 \\times D\\)47.6378.6974.9893.98
(f) \\(0.5 \\times D\\)46.8077.8971.6992.00
(g) \\(D=\\text{CC3M}\\)46.4477.7671.3391.84
(h) Image-only45.3276.9274.0292.35
", + "bbox": [ + 125, + 412, + 444, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "quence of model updates (old model B32, new model L14, and better new model H14) and present the retrieval results in Table 3. Initially, we apply XBT to L14 to ensure compatibility with B32. Subsequently, we apply XBT to H14 to ensure compatibility with the previously learned L32, which is already compatible with B32. For this process, we divide both the image-text pairs and the text-only pretraining train sets into two halves, using each separately for each case. A comparison with the original results reported in Table 3 confirms that our XBT performs well in the continual learning scenario either. It achieves cross-modal backward compatibility while leveraging the power of the improved new model.", + "bbox": [ + 88, + 625, + 482, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study. To validate XBT, we conduct further analysis as shown in Table 4. By comparing (a) and (b), we observe that introducing noise during $\\phi$ training aids in generalization. The comparison between (c) and (d) demonstrates that the scale of $D_{T}$ is important, supporting our", + "bbox": [ + 88, + 824, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "assumption that a sufficient number of text samples can help build a robust $\\phi$ . (e) outperforms (a), confirming that utilizing more image-text pairs can enhance XBT. The lower performance of (f) also aligns with the notion that the number of image-text pairs is crucial. In (g), when we replace $D$ with CC3M [33], from which we can obtain around 2.4M, the performance is similar to (f), suggesting that XBT can be effectively applied with other datasets. In (h), we replace text-only pretraining with image-only pretraining of the same scale and observe that text-only pretraining performs better in both retrieval scenarios, demonstrating its efficiency and superiority.", + "bbox": [ + 511, + 376, + 906, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. We facilitate retrieval by utilizing new query embeddings and old gallery embeddings. The results in Figure 4 demonstrate accurate cross-modal backward-compatible retrieval.", + "bbox": [ + 511, + 560, + 906, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 645, + 633, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced Cross-modal Backward-compatible Training (XBT), a novel task for cross-modal retrieval that focuses on the compatibility between image and text embeddings of different Vision-Language Pretraining (VLP) models. We proposed an efficient solution using a text-only pretrained projection module, $\\phi$ , to align the new model's embeddings with those of the old model, thereby enhancing training efficiency. By integrating parameter-efficient training schemes into the XBT framework, we were able to accelerate the model's training while preserving the original VLP's zero-shot capabilities. Our approach, demonstrated on various cross-modal benchmarks, effectively builds cross-modal retrieval systems without backfilling, offering an efficient and environmentally friendly solution in response to the VLP improvements.", + "bbox": [ + 509, + 674, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1790", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The first author performed the initial research and development of this work in collaboration with Ser-nam Lim. The finalization and refinement of this work were completed while the first author was affiliated with Google DeepMind.", + "bbox": [ + 89, + 113, + 485, + 172 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 181, + 187, + 196 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In ICCV, 2019.", + "[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In CVPR, 2017.", + "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016.", + "[4] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In IEEE international conference on automatic face & gesture recognition. IEEE, 2018.", + "[5] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. In ICLR, 2019.", + "[6] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023.", + "[7] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 2021.", + "[8] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, Yoohoon Kang, and Sangdoo Yun. Language-only efficient training of zero-shot composed image retrieval. arXiv preprint arXiv:2312.01998, 2023.", + "[9] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016.", + "[10] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, 2021.", + "[11] Jiun Tian Hoe, Kam Woh Ng, Tianyu Zhang, Chee Seng Chan, Yi-Zhe Song, and Tao Xiang. One loss for all: Deep hashing with a single cosine similarity based learning objective. NeurIPS, 2021.", + "[12] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021.", + "[13] Weihua Hu, Rajas Bansal, Kaidi Cao, Nikhil Rao, Karthik Subbian, and Jure Leskovec. Learning backward compatible embeddings. In ACM SIGKDD, 2022.", + "[14] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In ACM ICMR, pages 39-43, 2008.", + "[15] Young Kyun Jang and Nam Ik Cho. Generalized product quantization network for semi-supervised image retrieval. In CVPR, 2020." + ], + "bbox": [ + 93, + 205, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Young Kyun Jang, Geonmo Gu, Byungsoo Ko, Isaac Kang, and Nam Ik Cho. Deep hash distillation for image retrieval. In ECCV, 2022.", + "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021.", + "[18] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV. Springer, 2022.", + "[19] Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. How to adapt your large-scale vision-and-language model. 2021.", + "[20] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021.", + "[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022.", + "[22] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022.", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014.", + "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2017.", + "[25] Jie Lu, Vahid Behbood, Peng Hao, Hua Zuo, Shan Xue, and Guangquan Zhang. Transfer learning using computational intelligence: A survey. Knowledge-Based Systems, 2015.", + "[26] Qiang Meng, Chixiang Zhang, Xiaogiang Xu, and Feng Zhou. Learning compatible embeddings. In ICCV, 2021.", + "[27] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. In EMNLP, 2022.", + "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021.", + "[29] Vivek Ramanujan, Pavan Kumar Anasosalu Vasu, Ali Farhadi, Oncel Tuzel, and Hadi Pouransari. Forward compatible training for large-scale embedding retrieval systems. In CVPR, 2022.", + "[30] Mehwish Rehman, Muhammad Iqbal, Muhammad Sharif, and Mudassar Raza. Content based image retrieval: survey. World Applied Sciences Journal, 2012.", + "[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015.", + "[32] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022." + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1791", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, pages 2556-2565, 2018.", + "[34] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In CVPR, 2020.", + "[35] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022.", + "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017.", + "[37] Ji Wan, Dayong Wang, Steven Chu Hong Hoi, Pengcheng Wu, Jianke Zhu, Yongdong Zhang, and Jintao Li. Deep learning for content-based image retrieval: A comprehensive study. In ACMMM, 2014.", + "[38] Chien-Yi Wang, Ya-Liang Chang, Shang-Ta Yang, Dong Chen, and Shang-Hong Lai. Unified representation learning for cross model compatibility. In BMVC, 2020.", + "[39] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. NeurIPS, 32, 2019.", + "[40] Kai Wang, Luis Herranz, and Joost van de Weijer. Continual learning in cross-modal retrieval. In CVPR, 2021.", + "[41] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, 2020.", + "[42] Sijie Yan, Yuanjun Xiong, Kaustav Kundu, Shuo Yang, Siqi Deng, Meng Wang, Wei Xia, and Stefano Soatto. Positive-congruent training: Towards regression-free model updates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14299-14308, 2021.", + "[43] Binjie Zhang, Yixiao Ge, Yantao Shen, Yu Li, Chun Yuan, Xuyuan Xu, Yexin Wang, and Ying Shan. Hot-refresh model upgrades with regression-alleviating compatible training in image retrieval. In ICLR, 2022.", + "[44] Binjie Zhang, Yixiao Ge, Yantao Shen, Shupeng Su, Chun Yuan, Xuyuan Xu, Yexin Wang, and Ying Shan. Towards universal backward-compatible representation learning. In IJCAI, 2022.", + "[45] Liangli Zhen, Peng Hu, Xi Peng, Rick Siow Mong Goh, and Joey Tianyi Zhou. Deep multimodal transfer learning for cross-modal retrieval. TNNLS, 2020.", + "[46] Wengang Zhou, Houqiang Li, and Qi Tian. Recent advance in content-based image retrieval: A literature survey. arXiv preprint arXiv:1706.06064, 2017.", + "[47] Yifei Zhou, Zilu Li, Abhinav Shrivastava, Hengshuang Zhao, Antonio Torralba, Taipeng Tian, and Ser-Nam Lim. bt2: Backward-compatible training with basis transformation. arXiv preprint arXiv:2211.03989, 2022.", + "[48] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "comprehensive survey on transfer learning. Proceedings of the IEEE, 2020.", + "bbox": [ + 545, + 92, + 906, + 119 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1792", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/5c932632-b5a7-42ba-9baf-b3ca38b774d6_model.json b/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/5c932632-b5a7-42ba-9baf-b3ca38b774d6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ac89fb623214f7b794cd1bcf6928e9665a76088f --- /dev/null +++ b/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/5c932632-b5a7-42ba-9baf-b3ca38b774d6_model.json @@ -0,0 +1,2244 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.131, + 0.804, + 0.177 + ], + "angle": 0, + "content": "Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.204, + 0.428, + 0.222 + ], + "angle": 0, + "content": "Young Kyun Jang1" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.222, + 0.427, + 0.24 + ], + "angle": 0, + "content": "Google DeepMind" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.243, + 0.43, + 0.256 + ], + "angle": 0, + "content": "kyun0914@gmail.com" + }, + { + "type": "text", + "bbox": [ + 0.554, + 0.204, + 0.671, + 0.221 + ], + "angle": 0, + "content": "Ser-nam \\(\\mathrm{Lim}^2\\)" + }, + { + "type": "text", + "bbox": [ + 0.493, + 0.222, + 0.731, + 0.24 + ], + "angle": 0, + "content": "\\(^{2}\\)University of Central Florida" + }, + { + "type": "text", + "bbox": [ + 0.538, + 0.243, + 0.682, + 0.256 + ], + "angle": 0, + "content": "sernam@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.486, + 0.705 + ], + "angle": 0, + "content": "Modern retrieval systems often struggle with upgrading to new and more powerful models due to the incompatibility of embeddings between the old and new models. This necessitates a costly process known as backfilling, which involves re-computing the embeddings for a large number of data samples. In vision, Backward-compatible Training (BT) has been proposed to ensure that the new model aligns with the old model's embeddings. This paper extends the concept of vision-only BT to the field of cross-modal retrieval, marking the first attempt to address Cross-modal BT (XBT). Our goal is to achieve backward-compatibility between Vision-Language Pretraining (VLP) models, such as CLIP, for the cross-modal retrieval task. To address XBT challenges, we propose an efficient solution: a projection module that maps the new model's embeddings to those of the old model. This module, pretrained solely with text data, significantly reduces the number of image-text pairs required for XBT learning, and, once it is pretrained, it avoids using the old model during training. Furthermore, we utilize parameter-efficient training strategies that improve efficiency and preserve the off-the-shelf new model's knowledge by avoiding any modifications. Experimental results on cross-modal retrieval datasets demonstrate the effectiveness of XBT and its potential to enable backfill-free upgrades when a new VLP model emerges." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.738, + 0.222, + 0.753 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.485, + 0.902 + ], + "angle": 0, + "content": "As the volume and variety of data grow exponentially in our multimedia-rich era, developing and maintaining efficient multi-modal retrieval systems becomes increasingly challenging. These systems, which provide data samples relevant to a user's query, must handle diverse data types, from text and images to audio and video. This growth puts a premium on the scalability and performance of these systems, necessitating continuous advancements in algorithms and technology. Embedding-based deep models for" + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.29, + 0.824, + 0.304 + ], + "angle": 0, + "content": "Backward-compatible Training (BT)" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.312, + 0.902, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.461, + 0.87, + 0.475 + ], + "angle": 0, + "content": "Cross-modal Backward-compatible Training (XBT)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.646, + 0.908, + 0.744 + ], + "angle": 0, + "content": "Figure 1. A conceptual visualization of Backward-compatible Training (BT, left) and its extension, the Cross(X)-modal version (XBT, right). Circles and squares denote data samples of images and text, respectively. XBT uses Vision-Language Pretraining (VLP) models as baselines to achieve cross-modal, backward-compatible representation learning, allowing the new, improved model to be compatible with the fixed old model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.777, + 0.908, + 0.852 + ], + "angle": 0, + "content": "retrieval have emerged as a key solution, transforming high-dimensional data into a lower-dimensional dense embedding space [11, 15, 16, 30, 37, 46]. These models capture the semantic meanings of data samples, enabling the quantification of similarities for retrieval." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.902 + ], + "angle": 0, + "content": "However, it is important to note that the embedding spaces generated by different deep models are not inherently compatible with each other. This incompatibility restricts the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1783" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "reuse of an existing gallery that has been embedded with an older model when a new, better-performing model is introduced. Consequently, this necessitates \"backfilling,\" a process where the entire gallery must be rebuilt using the embeddings from the new model. Such a requirement is resource-intensive and time-consuming, posing a significant challenge when building retrieval systems." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.482, + 0.35 + ], + "angle": 0, + "content": "Backward-compatible Training (BT) [13, 34, 38, 43, 44] has been developed to tackle this issue, specifically focusing on image retrieval systems. The main objective of BT is to train a new model from scratch in a manner that ensures its compatibility with an old model that was used to create the existing gallery. A successful BT model must consequently demonstrate that retrieving from the old gallery using a query embedded with the new model outperforms that using an embedding from the old model. This enhancement is crucial for justifying the application of BT." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.483, + 0.472 + ], + "angle": 0, + "content": "Taking the BT problem a step further, we propose a new challenging task called Cross(X)-modal Backward-compatible Training (XBT) as shown in Figure 1. Our focus here is on applying BT principles within the realm of cross-modal retrieval, specifically exploring the interaction and compatibility between image and text embeddings of different Vision-Language Pretraining (VLP) models [17, 21, 22, 28, 35]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.483, + 0.685 + ], + "angle": 0, + "content": "To achieve XBT, just like the BT problem, we need to resolve the incompatibility between the embeddings of the old, inferior model and the new, improved model. If we follow most prior practices of training from scratch with an additional compatibility loss term, a substantial number of supervised image-text pairs would be needed. The ideal quantity would be in the hundreds of millions, approximating the amount utilized in the establishment of VLP models. However, accessing the original training samples is often impossible, and training on that scale is prohibitively expensive and impractical. We propose an efficient solution instead: a text-only pretrained projection module, \\(\\phi\\), to align a given pretrained new model's embeddings with those of the old model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Our focus is on using only text data to estimate the entire distribution of embeddings in the VLP space. Specifically, we train \\(\\phi\\) to learn the similarity between old and new text embeddings in a contrastive manner. By increasing the number of text samples, which is simpler than preparing image-text pairs, \\(\\phi\\) can approximate the oracle projection between the intra-modal distributions of both old and new embeddings for texts. Assuming that the intra-modal distribution of texts in the VLP embedding space can mirror that of images, we can simply apply the same \\(\\phi\\) to the new image embeddings to synthesize the corresponding old image embeddings." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "With these generated synthetic old image and text embeddings, which can be considered as aligned with the new" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "embeddings, we aim to fine-tune the new model. In this process, far fewer supervised image-text pairs are required than in the original training dataset of the VLP. This also allows us to avoid using old model parameters during training, thereby enhancing training efficiency. In addition, we incorporate parameter-efficient fine-tuning schemes [12, 19, 20] into the new model, which add small trainable parameters and do not harm any of the new model's original parameters during training. This not only accelerates the new model's fine-tuning process but also allows for an easy reversion to the original new-to-new retrieval by simply removing the additional parameters." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.276, + 0.905, + 0.351 + ], + "angle": 0, + "content": "We demonstrate our approach on various cross-modal benchmarks, highlighting the effectiveness of XBT in cross-modal retrieval protocols. In response to the rapid advancement of VLP models, XBT offers an environmentally friendly solution by enabling backfill-free systems." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.353, + 0.805, + 0.366 + ], + "angle": 0, + "content": "Our contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.37, + 0.905, + 0.429 + ], + "angle": 0, + "content": "- The Cross-modal Backward-compatible Training (XBT) concept is introduced for the first time to solve the backfilling problem that stems from real-world cross-modal retrieval systems." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.431, + 0.905, + 0.49 + ], + "angle": 0, + "content": "- A novel XBT solution is proposed, which uses a text-only pretrained projection module, \\(\\phi\\), to efficiently align the embeddings of new and old models using only text samples." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.491, + 0.904, + 0.535 + ], + "angle": 0, + "content": "- The proposal is demonstrated on various datasets and protocols, showcasing XBT's effectiveness in building backfill-free cross-modal retrieval systems." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.37, + 0.905, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.557, + 0.662, + 0.571 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Backward-compatible Training. The concept of Backward-compatible Training (BT) was first introduced in the study ([34]). This approach influences a new model with the learned classifier of the old model, thereby achieving backward compatibility between the old and new models. However, BT can degrade the original representational performance of the new model. To overcome this, [26] proposed aligning class-wise centers presented by the old and new models. Another approach to achieve backward compatibility is to use an additional projection to map the old embedding into the new embedding space by adding a lightweight transformation, as suggested in [13, 38]. The work of [29] further adds an auxiliary feature in preparation for future updates, while [47] uses additional dimensionality in the embedding to obtain compatibility. An online strategy that backfills the gallery on the fly is explored in [43], and [42] addresses the model regression problem. Despite this progress, the scenario considering cross-modal retrieval between image and text, which has many real-world applications, remains unexplored. Our study on XBT in this paper aims to fill this gap." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1784" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.487, + 0.364 + ], + "angle": 0, + "content": "Vision-Language Continual and Transfer Learning The fields of continual learning [2, 5, 7] and transfer learning [25, 48] bear similarities to backward-compatible representation learning, as all aim to update an existing model to boost performance. In the realm of multi-modal representation learning for cross-modal retrieval, continual learning approaches like [40] propose methods to prevent catastrophic forgetting across different modalities. Transfer learning approaches like [45] suggest strategies for transferring knowledge from previously labeled categories (source modality) to new, unlabeled categories (target modality). However, our proposed XBT stands apart from these methods as it specifically tackles the challenge of maintaining backward compatibility between old and new models. This unique attribute makes XBT ideally suited for backfill-free retrieval scenarios, where the embeddings of the old model remain unchanged, yet we can still leverage the enhanced performance of the new model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.377, + 0.182, + 0.392 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.485, + 0.522 + ], + "angle": 0, + "content": "Our goal is to construct a backfill-free, embedding-based, cross-modal retrieval system using VLP models, which are configured with an image encoder \\( E^{I} \\) and a text encoder \\( E^{T} \\). When a new better performing VLP model \\( \\{E_{new}^{I}, E_{new}^{T}\\} \\) emerges, we aim to ensure its compatibility with the old model \\( \\{E_{old}^{I}, E_{old}^{T}\\} \\) that was used to construct the gallery. To achieve this, we introduce Cross(X)-modal Backward-compatible Training (XBT)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.485, + 0.629 + ], + "angle": 0, + "content": "Given an image \\( x \\) and text caption \\( t \\), XBT enables retrieval between a new image embedding \\( v_{new} = E_{new}^{I}(x) \\) and the text embeddings \\( (w_{old}) \\), as well as between a new text embedding \\( w_{new} = E_{new}^{T}(t) \\) and the image embeddings \\( (v_{old}) \\). We denote backward compatible embeddings as \\( \\bar{v} \\) and \\( \\bar{w} \\). All embeddings we utilize in this work are l2-normalized." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.483, + 0.669 + ], + "angle": 0, + "content": "3.1. Criterion for Cross-modal Backward Compatibility" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.721 + ], + "angle": 0, + "content": "Following the definition of backward compatibility in BT work [34], we can construct strict constraints that ensure cross-modal backward compatibility as:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.743, + 0.483, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d \\left(w _ {n e w _ {i}}, v _ {o l d _ {j}}\\right) \\leq d \\left(w _ {o l d _ {i}}, v _ {o l d _ {j}}\\right), \\forall y _ {i} = y _ {j}, \\\\ d \\left(w _ {\\text {n e w} _ {i}}, v _ {\\text {o l d} _ {j}}\\right) \\geq d \\left(w _ {\\text {o l d} _ {i}}, v _ {\\text {o l d} _ {j}}\\right), \\forall y _ {i} \\neq y _ {j}, \\tag {1} \\\\ d (v _ {n e w _ {i}}, w _ {o l d _ {j}}) \\leq d (v _ {o l d _ {i}}, w _ {o l d _ {j}}), \\forall y _ {i} = y _ {j}, \\\\ d (v _ {n e w _ {i}}, w _ {o l d _ {j}}) \\geq d (v _ {o l d _ {i}}, w _ {o l d _ {j}}), \\forall y _ {i} \\neq y _ {j}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( y_{i} \\) and \\( y_{j} \\) represent whether an image and text are paired (\\( y_{i} = y_{j} \\)) or not (\\( y_{i} \\neq y_{j} \\)). The term \\( d(\\cdot, \\cdot) \\) represents a distance metric in the embedding space, and we choose cosine distance as the baseline. The constraints in Eqn. 1 formally express the requirement that the new embedding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "must perform at least as well as the old embedding in terms of correctly matching image-text pairs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.909, + 0.198 + ], + "angle": 0, + "content": "However, exhaustively satisfying these constraints is intractable due to the potential for the old embeddings to outperform the new embeddings in certain retrieval cases. As a result, we relax the criteria by defining an alternative evaluation metric as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.225, + 0.907, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\begin{array}{l} \\mathcal {M} (w _ {\\text {n e w}}, v _ {\\text {o l d}}; \\mathrm {Q}, \\mathrm {G}) > \\mathcal {M} (w _ {\\text {o l d}}, v _ {\\text {o l d}}; \\mathrm {Q}, \\mathrm {G}), \\\\ \\mathcal {A} ^ {\\prime} (\\mathrm {Q}, \\mathrm {G}) = \\mathcal {A} ^ {\\prime} (\\mathrm {Q}, \\mathrm {G}) \\end{array} \\tag {2} \\\\ \\mathcal {M} (v _ {n e w}, w _ {o l d}; \\mathrm {Q}, \\mathrm {G}) > \\mathcal {M} (v _ {o l d}, w _ {o l d}; \\mathrm {Q}, \\mathrm {G}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.411 + ], + "angle": 0, + "content": "where a metric, such as recall, \\(\\mathcal{M}(q,g;\\mathrm{Q},\\mathrm{G})\\) takes query embedding \\(q\\) and gallery embedding \\(g\\) over query set Q and gallery set G. This criterion suggests that the general performance is enhanced when the query embedding from the XBT-trained new model is used to perform retrieval with the gallery of the old model, compared to the performance of the old model alone. In essence, fulfilling Eqn. 2 indicates that the new model has achieved backward compatibility and can feasibly update without backfilling gallery." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.909, + 0.563 + ], + "angle": 0, + "content": "Additionally, it is important to highlight that VLP models function as zero-shot learners. This leads us to define the XBT problem differently from classic BT, which trains separate models for specific image domain retrieval tasks such as ImageNet [31] or VGGFace2 [4]. In contrast, XBT tackles a more challenging task, aiming to bridge a new VLP model with a frozen old VLP model while retaining the new model's original zero shot capability. In this paper, we therefore assess the performance of XBT using retrieval and classification benchmarks in a zero-shot manner." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.719, + 0.592 + ], + "angle": 0, + "content": "3.2. Text-only Pretraining" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.909, + 0.748 + ], + "angle": 0, + "content": "The vast and diverse image-text pairs used to build VLP models significantly enhance their ability to connect semantically similar visual content and natural language [6]. However, this creates a complex embedding distribution that is difficult to predict and understand, thereby complicating the XBT process. A straightforward solution is to use a large number of supervised image-text pairs, similar to the approach used when building VLP models from scratch, to estimate the entire distribution of image and text embeddings during XBT." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "However, acquiring a sufficient number of accurate supervised pairs is extremely costly. To be far more efficient, we employ a small sized projection module, \\(\\phi\\), and train it exclusively with text samples, as shown in Figure 2. We hypothesize that the distribution of text embeddings in VLP models, which is determined by their semantic similarity, is similarly mirrored in the distribution of their corresponding matched images. With this in mind, we aim to train \\(\\phi\\) to cover the broad spectrum of the text embedding distribution between the new and old VLP models, making embeddings" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1785" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.088, + 0.354, + 0.102 + ], + "angle": 0, + "content": "Text-only Pretraining" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.106, + 0.484, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.228, + 0.484, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.511 + ], + "angle": 0, + "content": "Figure 2. An illustration of the text-only pretraining of \\(\\phi\\) (above), and XBT with \\(\\phi\\) (below). Using only text samples, \\(\\phi\\) is trained to approximate distribution of old text embeddings from that of new ones. After pretraining, the same \\(\\phi\\) is used to generate both of synthetic old image and text embeddings from the new VLP embeddings and train to learn cross-modal backward-compatible representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.541, + 0.483, + 0.572 + ], + "angle": 0, + "content": "from the same text sample similar and others dissimilar in a contrastive way:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.604, + 0.484, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p r e} = \\mathbb {E} _ {t \\sim D _ {T}} \\left[ \\mathcal {L} _ {c} \\left(\\phi \\left(w _ {\\text {n e w}} + \\epsilon\\right), w _ {\\text {o l d}}; \\tau_ {p r e}\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.485, + 0.701 + ], + "angle": 0, + "content": "where \\(t\\) denotes text data sampled from text corpus \\(D_{T} = \\{t_{i}\\}_{i=1}^{N_{D_{T}}} \\text{ of } N_{D_{T}} \\text{ text samples, and } \\mathcal{L}_{c}(\\cdot, \\cdot) \\text{ is a standard batch-wise contrastive loss as defined in CLIP [28] of temperature } \\tau_{pre}.\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.809 + ], + "angle": 0, + "content": "To prevent overfitting towards the text domain, which could widen the discrepancy between the image and text embedding representations, we inject noise \\(\\epsilon\\) sampled from a Gaussian distribution \\((\\epsilon \\sim N(0,\\sigma^2))\\), inspired by the approaches [8, 27]. Notably, \\(\\phi\\) is designed to take the new VLP model's embeddings, which means that \\(w_{new} + \\epsilon\\) is also \\(l2\\)-normalized." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "During this stage, only \\(\\phi\\) is updated, while all encoders remain fixed. By expanding the scale of \\(D_{T}\\), we can effectively replicate the complete text embedding spaces of both new and old VLP models. Additionally, \\(\\phi\\) can be applied to new image embeddings to generate synthetic old image embeddings that closely approximate the actual old ones." + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.088, + 0.767, + 0.099 + ], + "angle": 0, + "content": "XBT Learning Process" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.1, + 0.905, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.164, + 0.908, + 0.206 + ], + "angle": 0, + "content": "Figure 3. Our proposed learning process to achieve XBT. Notably, the old VLP model's encoders are not required in this stage, enhancing efficiency in training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.23, + 0.893, + 0.246 + ], + "angle": 0, + "content": "3.3. Cross-modal Backward-compatible Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.907, + 0.358 + ], + "angle": 0, + "content": "Training Loss. The objective of XBT is to ensure cross-modal compatibility, specifically between \\( w_{new} \\) and \\( v_{old} \\), as well as between \\( v_{new} \\) and \\( w_{old} \\). For a given dataset \\( D = \\{x_i, t_i\\}_{i=1}^{N_D} \\) of \\( N_D \\) supervised image-text pairs, we aim to train the new VLP model encoders, \\( E_{new}^I \\) and \\( E_{new}^T \\). Note that in our base setting, the text corpus is significantly larger than the supervised dataset, i.e., \\( N_{D_T} \\gg N_D \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.909, + 0.463 + ], + "angle": 0, + "content": "However, the dimensionality of new and old VLP embeddings may differ, and even if they are the same, they are not be directly compatible. To address this, we apply the pretrained \\(\\phi\\) to ensure that the new embeddings match the dimension of the old ones and project into the compatible embedding space, as shown in Figure 2. This can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.475, + 0.907, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\phi (v _ {\\text {n e w}}) = \\bar {v} _ {\\text {n e w}}, \\phi (w _ {\\text {n e w}}) = \\bar {w} _ {\\text {n e w}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.502, + 0.908, + 0.546 + ], + "angle": 0, + "content": "where \\(\\bar{v}_{new}\\) and \\(\\bar{w}_{new}\\) denote synthetic old embeddings and \\(\\{\\bar{v}_{new},\\bar{w}_{new},v_{old},w_{old}\\} \\subset \\mathbb{R}^K\\) , with \\(K\\) denoting the dimensionality of the old VLP model's embedding." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.547, + 0.801, + 0.561 + ], + "angle": 0, + "content": "We then apply XBT loss \\(\\mathcal{L}_X\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.573, + 0.907, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {X} = \\mathbb {E} _ {x, t \\sim D} \\left[ \\mathcal {L} _ {c} \\left(\\bar {v} _ {\\text {n e w}}, \\bar {w} _ {\\text {n e w}}; \\tau_ {X}\\right) \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.704 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{L}_c(\\cdot ,\\cdot)\\) is the same contrastive loss used in Eqn. 3, with temperature \\(\\tau_{X}\\). As shown in Figure 3, the new VLP encoders, \\(E_{new}^{I}\\) and \\(E_{new}^{T}\\), and \\(\\phi\\) are trained through \\(\\mathcal{L}_X\\) in an end-to-end manner. Ultimately, \\(\\bar{v}_{new}\\) and \\(\\bar{w}_{new}\\) become cross-modal backward-compatible with existing old gallery embeddings, \\(v_{old}\\) and \\(w_{old}\\), as all embeddings are distributed in the compatible space through \\(\\phi\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.81 + ], + "angle": 0, + "content": "Efficient Training. The efficiency of XBT can be attributed to two design choices that avoid: (1) any dependence on the old model when conducting XBT, and (2) updating the new VLP model off the shelf. We have already seen how \\(\\phi\\) can effectively achieve (1). (2) is simply facilitated by applying small-sized additional parameters, namely Soft prompt [20] and LoRA[12]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To be more specific, we incorporate the concept of soft prompt tuning, as outlined in [18, 20]. Soft prompting is applied to the vision encoder \\( E_{new}^{I} \\) by adding trainable prompts as input, along with image patch tokens. We exclude soft prompt tuning on the text side, as we use the text-only pretrained module \\( \\phi \\), which requires maintaining the original" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1786" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.196 + ], + "angle": 0, + "content": "distribution of text embeddings. We adopt the LoRA strategy [12] for the new VLP model's encoders \\( E_{new}^{I} \\) and \\( E_{new}^{T} \\) to avoid modifying original parameters. These two factors, in conjunction with \\( \\phi \\), offer an on-off solution: we can retrieve old samples using additional parameters, and we can easily revert to the original model by removing these parameters for new-to-new retrieval." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.197, + 0.482, + 0.273 + ], + "angle": 0, + "content": "Further, regarding \\(\\phi\\) during XBT (Sec. 3.3), we only finetune the parameters of layer normalization. This is done with the aim of preserving the knowledge learned during the text-only pretraining stage (Sec. 3.2) and facilitating easy adaptation towards the image-text joint representation space" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.273, + 0.482, + 0.438 + ], + "angle": 0, + "content": "In the end, our XBT framework offers a parameter-efficient solution that enables rapid convergence with fewer training iterations (a single epoch is sufficient); maintains the new model's power; and requires far fewer supervised training samples than the scale of VLP model pretraining (approximately \\(1\\%\\) of the level required to train CLIP [28] from scratch, which uses a 400M dataset, and \\(0.2\\%\\) of the level required for LAION-2B based CLIP models [32], which utilize a 2B dataset). All these help to avoid any full-scale retraining of the new VLP models, paving the way for XBT to simply use new VLP models off-the-shelf as mentioned." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.222, + 0.469 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.482, + 0.597 + ], + "angle": 0, + "content": "The evaluation of the proposed cross-modal backward compatibility includes image-text zero-shot retrieval. Pretrained VLP models are used as our baseline, with the goal to assess performance in a zero-shot environment by tuning a given pretrained new VLP model that is stronger than the old model. The results highlight the potential of XBT, showing improved performance when old model outputs are matched with the XBT-learned new model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.172, + 0.623 + ], + "angle": 0, + "content": "4.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.824 + ], + "angle": 0, + "content": "Model Training. Our XBT process involves two distinct training stages: text-only pretraining (See Sec. 3.2) and image-text supervised training (See Sec. 3.3). For the text-only pretraining stage, we utilize the text samples from a subset of the 115M filtered web-collected image-text paired dataset from BLIP [21], comprising around 67M available pairs (58.2% of the total). We construct a text corpus \\(D_{T}\\) in Eqn. 3 using these text samples. For the subsequent image-text supervised training stage, we use a smaller subset of 4M image-text pairs from the same dataset to construct a supervised dataset \\(D\\) in Eqn. 5. It is important to note that these subsets are significantly smaller than the dataset scale (400M, 2B or more) used to build CLIP models [28, 32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.899 + ], + "angle": 0, + "content": "We utilize 8-A100-80GB GPUs to train and evaluate the models. For the text-only pretraining stage in Sec. 3.2, the batch size is set to 8,192 (1,024 batch per GPU) and during this stage, the entire set of trainable parameters of \\(\\phi\\) are trained while both VLP text encoders are fixed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.409 + ], + "angle": 0, + "content": "Moving on to the image-text supervised training in Sec. 3.3, the batch size is reduced to 1,024 (128 batch per GPU), and in this stage, layer normalization is set to be the only trainable component for all VLP image and text encoders with \\(\\phi\\). Temperature hyper-parameters \\(\\tau_{pre}\\), \\(\\tau_{X}\\), and \\(\\tau_{N}\\) are fixed at 2.6592. We employ the AdamW optimizer [24] with a fixed learning rate of 1e-4 and a weight decay of 0.01. For soft prompts, we use 10 prompts and apply 100 times larger learning rate, 1e-2. The training iteration is determined by the dataset size, and the entire pipeline is trained for a single epoch. For image augmentation, we begin with a random resized crop, adjusting the image scale between 0.5 and 1.0. Additionally, we apply a random horizontal flip and make random adjustments to the image's contrast, brightness, and sharpness. To incorporate different perspectives and angles, we modify the image's translation and rotation. The pretrained weights provided by HuggingFace [41] are applied to baseline VLP models. We employ as: openai/clip-vit-base-patch32, openai/clip-vit-target-patch14, laion/CLIP-ViT-H-14-laion2B-s32B-b79K." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.55 + ], + "angle": 0, + "content": "Model Evaluation. We validate the effectiveness of XBT with a benchmark comprising three popular image-text paired datasets for cross-modal retrieval evaluation. The first is the nocaps dataset [1], which offers a diverse distribution of object names and facilitates a more detailed analysis. We employ the validation split of this dataset, which consists of images each paired with 10 relevant textual captions, totaling 4,500 images and 45,000 captions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.687 + ], + "angle": 0, + "content": "To evaluate XBT at a larger scale, we also include the Flickr [14] and COCO [23] datasets, which consist of images that are each paired with 5 relevant textual captions. For the Flickr dataset, we use the entire dataset, encompassing 31,783 images and 158,915 captions. For the COCO dataset, we use the validation split, which includes 35,136 images and 175,680 captions. We note the two datasets as Large-Flickr and Large-COCO respectively for the remainder of the paper." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.907, + 0.763 + ], + "angle": 0, + "content": "For retrieval, we utilize all the samples in each dataset for both query and gallery, following the cross-modal benchmark used in [21, 28]. For evaluation purpose, we adopt recall scores at top K retrieval results \\((\\mathbb{R}@\\mathbb{K},\\%)\\) to estimate the cross-modal backward compatibility (See Eqn. 2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.786, + 0.909, + 0.876 + ], + "angle": 0, + "content": "Implementation Details. In our work, we primarily use the popular CLIP models [28], based on a Transformer [36] backbone, as our baseline VLP models. The models, listed in ascending order of scale and performance, are: CLIP-ViT-B32, CLIP-ViT-L14, and CLIP-ViT-H14. For simplicity, we will refer to them as B32, L14, and H14, respec" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.7, + 0.9 + ], + "angle": 0, + "content": "1https://huggingface.co/models" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1787" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 1. Cross-modal retrieval results on nocaps. Note that for comparison, the B32/L14 case should be compared with B32, and the L14/H14 case should be compared with L14, respectively." + }, + { + "type": "table", + "bbox": [ + 0.151, + 0.129, + 0.846, + 0.44 + ], + "angle": 0, + "content": "
Old Model / New ModelMethodText Query(w, w)/Image Gallery(v)Image Query(v, v)/Text Gallery(w)
CaseR@1R@5R@10R@50CaseR@1R@5R@10R@50
Original
CLIP-ViT-B32-w/v45.1474.9884.5196.00v/w71.3892.0296.3399.67
CLIP-ViT-L14-w/v47.7776.5085.1395.95v/w73.2993.2497.4099.78
Cross-modal Backward Compatible Training
CLIP-ViT-B32 / CLIP-ViT-L14Full-tune\\( \\bar{w}_{new}/v_{old} \\)41.6073.5484.2696.45\\( \\bar{v}_{new}/w_{old} \\)63.7687.0093.5199.09
LoRA-only43.4074.6684.9496.7368.8290.8796.0099.76
Base43.4874.8785.0496.7469.7891.3196.0299.78
XBT48.0279.0088.2197.6675.0293.2797.3199.91
Full-tune\\( \\bar{w}_{new}/\\bar{v}_{new} \\)48.9279.3188.2697.64\\( \\bar{v}_{new}/\\bar{w}_{new} \\)61.2285.8992.6299.04
LoRA-only55.4483.8291.2498.3470.1191.1396.4999.73
Base56.1884.3191.4498.3771.0791.4496.2299.76
XBT63.4888.6193.9898.8677.2294.5697.9199.89
CLIP-ViT-L14 / CLIP-ViT-H14Full-tune\\( \\bar{w}_{new}/v_{old} \\)45.1877.8285.3497.65\\( \\bar{v}_{new}/w_{old} \\)66.3888.5593.5998.90
LoRA-only47.0076.9486.0297.6372.0492.4296.5899.67
Base47.0677.1586.1297.6473.0092.8696.6099.69
XBT51.1480.8288.9398.7276.6294.1197.5899.84
Full-tune\\( \\bar{w}_{new}/\\bar{v}_{new} \\)54.3982.2689.0698.95\\( \\bar{v}_{new}/\\bar{w}_{new} \\)65.5787.5293.5199.13
LoRA-only60.9186.7792.0498.6574.4692.7397.3899.82
Base61.6587.2692.2498.6875.4293.0497.1199.85
XBT66.4790.2194.9299.0580.0296.1198.87100.0
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.463, + 0.487, + 0.661 + ], + "angle": 0, + "content": "tively. Notably, while B32 and L14 are trained with the same dataset at 400M scale, H14 is trained with a larger dataset at 2B scale. We apply additional LoRA [12] parameters for each new model encoder configured as follows: \\(\\mathrm{LoRA}_{\\alpha} = 16\\), \\(\\mathrm{rank} = 16\\), and dropout \\(= 0.1\\). The proposed projection module \\(\\phi\\) consists of three Linear layers with layer normalization (LN) [3] and the GELU nonlinearity function [9]. Detailed architecture is as follows: Linear-LN-GELU-Linear-LN-GELU-Linear. Dropout is not applied as it was found to degrade performance empirically. The intermediate hidden dimension of the Linear layer is set to be four times the dimensionality of the output embedding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.231, + 0.683 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Baseline Comparisons. To simplify notation, we use \\( w / v \\) to denote the retrieval results obtained using \\( w \\) as the query embeddings and \\( v \\) as the gallery embeddings. We establish two protocols: assessing cross-modal backward compatibility with new/old (\\( \\bar{w}_{\\text{new}} / v_{\\text{old}} \\), \\( \\bar{v}_{\\text{new}} / w_{\\text{old}} \\)), and maintaining the new VLP's original performance with new/new (\\( \\bar{w}_{\\text{new}} / \\bar{v}_{\\text{new}} \\), \\( \\bar{v}_{\\text{new}} / \\bar{w}_{\\text{new}} \\)). All of the above are similarly applied to \\( v / w \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.885 + ], + "angle": 0, + "content": "To the best of our knowledge, since this paper presents the first method that aims to solve the XBT problem, there are no direct previous works to compare to. Nevertheless, we compare our XBT against three baselines: Full-tune, LoRA-only, and Base, which we elaborate on below." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.887, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Naive solution - Direct backward contrastive learning:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.909, + 0.571 + ], + "angle": 0, + "content": "The pretrained \\(\\phi\\) in Eqn. 3, which maps the new embeddings to the old via text based contrastive learning, constitutes a major contribution of XBT. We therefore would like to compare with a solution that does not utilize such a pretrained \\(\\phi\\). A straightforward solution for this is to fine-tune the new VLP encoders \\((E_{new}^{I}\\) and \\(E_{new}^{T})\\) to be compatible with the old ones by minimizing the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.584, + 0.906, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {D i r e c t} = \\mathbb {E} _ {x, t \\sim D} [ \\mathcal {L} _ {c} (\\bar {v} _ {\\text {n e w}}, w _ {\\text {o l d}}; \\tau_ {N}) \\\\ \\left. + \\mathcal {L} _ {c} \\left(\\bar {w} _ {\\text {n e w}}, v _ {\\text {o l d}}; \\tau_ {N}\\right) \\right] \\tag {0} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.703 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_c(\\cdot, \\cdot)\\) is the same contrastive loss used in Eqn. 5, and \\(\\tau_N\\) is a temperature. Here, we apply a randomly initialized \\(\\phi\\) with the same configuration as XBT for projection, as per Eqn. 4, to generate cross-modal backward compatible embeddings, \\(\\bar{w}\\) and \\(\\bar{v}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Different training options: Based on Eqn. 6, we consider multiple training setups: (1) full fine-tuning of all trainable components (Full-tune), (2) tuning LoRA parameters-only (LoRA-only), and (3) starting from LoRA-only, adding an extra learnable prompt (Base). Models produced by Base would therefore be akin to XBT without the pretrained \\(\\phi\\). For these, we train the \\(E_{new}^{I}\\), \\(E_{new}^{T}\\) and \\(\\phi\\) using \\(\\mathcal{L}_{Direct}\\) with 4M image-text pairs. Note that, XBT only trains layer-normalization layers including pretrained \\(\\phi\\), however, Full-tune, LoRA-only and Base setups are training entire parameters of \\(\\phi\\) since it is not trained before. The results of these baselines are presented in Table 1, tested with nocaps dataset. The performance of the original VLP models is also reported" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1788" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.281, + 0.09, + 0.717, + 0.104 + ], + "angle": 0, + "content": "Table 2. Cross-modal retrieval results on Large-Flickr and Large-COCO." + }, + { + "type": "table", + "bbox": [ + 0.149, + 0.115, + 0.843, + 0.461 + ], + "angle": 0, + "content": "
DatasetOld Model / New ModelMethodText Query(w, w)/Image Gallery(v)Image Query(v, v)/Text Gallery(w)
CaseR@1R@5R@10R@50CaseR@1R@5R@10R@50
Original
CLIP-ViT-B32-w/v21.5441.2450.7272.69v/w40.3564.4273.3989.91
Cross-modal Backward Compatible Training
CLIP-ViT-B32 / CLIP-ViT-L14Base\\( \\bar{w}_{\\text{new}}/v_{\\text{old}} \\)19.0237.2546.6369.32\\( \\bar{v}_{\\text{new}}/w_{\\text{old}} \\)36.7460.6170.2288.34
XBT22.3842.7152.4174.8042.4766.0474.8790.79
Base\\( \\bar{w}_{\\text{new}}/\\bar{v}_{\\text{new}} \\)32.2955.1864.6883.60\\( \\bar{v}_{\\text{new}}/\\bar{w}_{\\text{new}} \\)36.1059.8069.9589.01
XBT39.5962.8971.7588.1343.5068.5077.9993.21
CLIP-ViT-B32 / CLIP-ViT-H14Base\\( \\bar{w}_{\\text{new}}/v_{\\text{old}} \\)20.2739.1748.5970.67\\( \\bar{v}_{\\text{new}}/w_{\\text{old}} \\)35.5359.4469.0687.67
XBT23.7044.4754.1976.0341.0965.1874.3290.16
Base\\( \\bar{w}_{\\text{new}}/\\bar{v}_{\\text{new}} \\)32.9455.9465.2683.77\\( \\bar{v}_{\\text{new}}/\\bar{w}_{\\text{new}} \\)34.8159.5569.9688.81
XBT40.4263.9072.7488.4746.3571.1880.0194.08
Original
CLIP-ViT-B32-w/v14.4430.2038.9662.30v/w28.6250.1759.6780.22
Cross-modal Backward Compatible Training
CLIP-ViT-B32 / CLIP-ViT-L14Base\\( \\bar{w}_{\\text{new}}/v_{\\text{old}} \\)12.9728.1436.4859.78\\( \\bar{v}_{\\text{new}}/w_{\\text{old}} \\)26.3246.9156.5477.86
XBT15.5532.2741.3465.1130.7352.5862.3081.83
Base\\( \\bar{w}_{\\text{new}}/\\bar{v}_{\\text{new}} \\)22.1641.6550.9873.06\\( \\bar{v}_{\\text{new}}/\\bar{w}_{\\text{new}} \\)27.2648.1657.8079.57
XBT27.6148.6958.0878.8332.9255.3764.8984.67
CLIP-ViT-B32 / CLIP-ViT-H14Base\\( \\bar{w}_{\\text{new}}/v_{\\text{old}} \\)13.8829.5538.2361.57\\( \\bar{v}_{\\text{new}}/w_{\\text{old}} \\)26.1946.7556.3077.96
XBT15.8232.7941.8465.4530.3351.2859.9781.66
Base\\( \\bar{w}_{\\text{new}}/\\bar{v}_{\\text{new}} \\)23.2043.2252.7074.68\\( \\bar{v}_{\\text{new}}/\\bar{w}_{\\text{new}} \\)27.3148.5658.6080.09
XBT28.1449.4858.8879.3234.5457.2966.7985.92
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.465, + 0.488 + ], + "angle": 0, + "content": "for a clear comparison. We highlight our XBT method." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.49, + 0.483, + 0.67 + ], + "angle": 0, + "content": "In the context of cross-modal backward compatibility, where recall scores of \\( \\bar{w}_{\\text{new}} / v_{\\text{old}} \\) and \\( \\bar{v}_{\\text{new}} / w_{\\text{old}} \\) should be higher than those of \\( w / v \\) and \\( v / w \\) of old model, respectively, Full-tune significantly underperforms compared to the old VLP's. This indicates that full fine-tuning is not an appropriate solution. While both LoRA-only and Base improve performance over Full-tune, they still fall short of the old VLP's performance and fail to meet the criterion. In contrast, XBT significantly outperforms these baselines and even surpasses the old VLP's performance, satisfying the cross-modal backward compatibility for all recall metrics (see Eqn. 2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.822 + ], + "angle": 0, + "content": "In terms of maintaining new VLP's performance, where recall scores of \\( \\bar{w}_{\\text{new}} / v_{\\text{new}} \\) and \\( \\bar{v}_{\\text{new}} / w_{\\text{new}} \\) could be similar to those of \\( w / v \\) and \\( v / w \\) of new model, respectively, all baselines show decent performance in \\( \\bar{w}_{\\text{new}} / v_{\\text{new}} \\) case but only XBT achieves better performance than new VLP model in \\( \\bar{v}_{\\text{new}} / w_{\\text{new}} \\) case. These results support that XBT not only maintains the performance of the new VLP model but also enhances it in certain cases. The superior performance of XBT in both cases underscores its promise as a robust solution for cross-modal retrieval tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Large Scale Retrieval. In this experiment, we assess the performance of cross-modal backward compatibility in more practical use cases. We use larger datasets, namely LargeFlickr and Large-COCO, as shown in Table 2. These datasets provide more than six times the number of samples typically" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.908, + 0.608 + ], + "angle": 0, + "content": "used in image-text retrieval literature [6, 37], which make the retrieval process more challenging. We use the second-best performing Base method from Table 1 as a basis for comparison. To validate the applicability of XBT across various VLP models, we utilize three CLIP baselines: B32, L14, and H14. Additionally, we explore different combinations, such as B32/L14 and B32/H14, to further analyze the effectiveness of XBT, beyond the comparisons made in Table 1 (B32/L14 and L14/H32)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.909, + 0.715 + ], + "angle": 0, + "content": "When comparing B32/L14 with B32/H14, it is observed that every retrieval cases of improved models (L14 and H14) successfully maintain cross-modal backward compatibility, regardless of their scale. Especially, retrieval performance of B32/H14 is better than that of B32/L14, which demonstrates XBT's ability to preserve and align original model's power with old gallery." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.909, + 0.837 + ], + "angle": 0, + "content": "Upon examining the results, we consistently observe that XBT outperforms the Base model across all instances and both retrieval scenarios, often by a significant margin. Furthermore, XBT maintains the performance of new VLPs better than Base, achieving even larger margins for \\(\\bar{w}_{new} / \\bar{v}_{new}\\) and \\(\\bar{v}_{new} / w_{old}\\). In summary, XBT consistently demonstrates robust and superior retrieval performance, even in large-scale scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.682, + 0.864 + ], + "angle": 0, + "content": "4.3. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Continual learning. In line with the literature on BT works [29, 34], we set up a continual learning scenario for a se" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1789" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 3. Continual learning scenario experimental results with XBT on nocaps. Note that for H14 training, we use L14, which has been previously adapted with B32, and H14 never encounters B14." + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.129, + 0.835, + 0.262 + ], + "angle": 0, + "content": "
Old Model / New ModelText Query(w, w)/Image Gallery(v)Image Query(v, v)/Text Gallery(w)
CaseR@1R@5R@10R@50CaseR@1R@5R@10R@50
CLIP-ViT-B32 / CLIP-ViT-L14wnew/vold47.6878.0088.0097.50vnew/wold73.8992.0097.2299.86
wnew/vnew62.2286.1893.0098.12vnew/wnew76.2792.9096.7699.11
CLIP-ViT-L14 / CLIP-ViT-H14wnew/vold50.9780.6888.9897.72vnew/wold77.3894.8297.8099.84
wnew/vnew66.0989.7894.7799.01vnew/wnew80.6096.3698.7199.96
CLIP-ViT-B32 / CLIP-ViT-H14wnew/vold49.1479.8888.5997.67vnew/wold74.8493.7397.2299.80
wnew/vnew65.9589.6894.6898.99vnew/wnew80.9696.0098.7199.96
" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.273, + 0.184, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.264, + 0.272, + 0.398, + 0.282 + ], + "angle": 0, + "content": "Retrieved from Old Text Gallery" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.284, + 0.412, + 0.293 + ], + "angle": 0, + "content": "R#1 - A dog with colorful sunglasses on and a bandanna." + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.293, + 0.437, + 0.302 + ], + "angle": 0, + "content": "R#2 - A happy dog is wearing a blue bandanna and sunglasses." + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.302, + 0.442, + 0.31 + ], + "angle": 0, + "content": "R#3 - A dog wearing sunglasses and a blue and white bandanna." + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.31, + 0.486, + 0.319 + ], + "angle": 0, + "content": "R#4 - An old brown dog is wearing mirrored sunglasses and a blue bandana." + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.319, + 0.452, + 0.327 + ], + "angle": 0, + "content": "R#5 - A dog has its tongue out as it wears sunglasses and bandana." + }, + { + "type": "list", + "bbox": [ + 0.197, + 0.284, + 0.486, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.506, + 0.28, + 0.577, + 0.289 + ], + "angle": 0, + "content": "New Query Text" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.289, + 0.576, + 0.3 + ], + "angle": 0, + "content": "The cute brown" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.3, + 0.575, + 0.309 + ], + "angle": 0, + "content": "tabby and white" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.309, + 0.575, + 0.318 + ], + "angle": 0, + "content": "kitten, sits on the" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.318, + 0.575, + 0.327 + ], + "angle": 0, + "content": "hard wood floor." + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.28, + 0.577, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.665, + 0.272, + 0.807, + 0.282 + ], + "angle": 0, + "content": "Retrieved from Old Image Gallery" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.282, + 0.901, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.622, + 0.328, + 0.637, + 0.337 + ], + "angle": 0, + "content": "R#1" + }, + { + "type": "image_caption", + "bbox": [ + 0.201, + 0.348, + 0.794, + 0.363 + ], + "angle": 0, + "content": "Figure 4. New query vs. Old gallery retrieval results on nocaps. B32 as old, and L14 as new model." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.375, + 0.483, + 0.401 + ], + "angle": 0, + "content": "Table 4. Ablation study tested on nocaps, with B32 as old and L14 as new model." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.414, + 0.446, + 0.593 + ], + "angle": 0, + "content": "
Method\\(\\bar{w}(\\text{text})/v(\\text{image})\\)\\(\\bar{v}(\\text{image})/w(\\text{text})\\)
R@1R@10R@1R@10
(a) Baseline48.0279.0075.0293.27
(b) w.o noise47.2478.0573.3192.49
(c) \\(0.25 \\times D_T\\)46.8977.9173.2092.18
(d) \\(0.5 \\times D_T\\)47.6878.0073.8992.00
(e) \\(2 \\times D\\)47.6378.6974.9893.98
(f) \\(0.5 \\times D\\)46.8077.8971.6992.00
(g) \\(D=\\text{CC3M}\\)46.4477.7671.3391.84
(h) Image-only45.3276.9274.0292.35
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.625, + 0.483, + 0.82 + ], + "angle": 0, + "content": "quence of model updates (old model B32, new model L14, and better new model H14) and present the retrieval results in Table 3. Initially, we apply XBT to L14 to ensure compatibility with B32. Subsequently, we apply XBT to H14 to ensure compatibility with the previously learned L32, which is already compatible with B32. For this process, we divide both the image-text pairs and the text-only pretraining train sets into two halves, using each separately for each case. A comparison with the original results reported in Table 3 confirms that our XBT performs well in the continual learning scenario either. It achieves cross-modal backward compatibility while leveraging the power of the improved new model." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.825, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Ablation Study. To validate XBT, we conduct further analysis as shown in Table 4. By comparing (a) and (b), we observe that introducing noise during \\(\\phi\\) training aids in generalization. The comparison between (c) and (d) demonstrates that the scale of \\(D_{T}\\) is important, supporting our" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.377, + 0.908, + 0.558 + ], + "angle": 0, + "content": "assumption that a sufficient number of text samples can help build a robust \\(\\phi\\). (e) outperforms (a), confirming that utilizing more image-text pairs can enhance XBT. The lower performance of (f) also aligns with the notion that the number of image-text pairs is crucial. In (g), when we replace \\(D\\) with CC3M [33], from which we can obtain around 2.4M, the performance is similar to (f), suggesting that XBT can be effectively applied with other datasets. In (h), we replace text-only pretraining with image-only pretraining of the same scale and observe that text-only pretraining performs better in both retrieval scenarios, demonstrating its efficiency and superiority." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.908, + 0.622 + ], + "angle": 0, + "content": "Qualitative Results. We facilitate retrieval by utilizing new query embeddings and old gallery embeddings. The results in Figure 4 demonstrate accurate cross-modal backward-compatible retrieval." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.646, + 0.634, + 0.661 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced Cross-modal Backward-compatible Training (XBT), a novel task for cross-modal retrieval that focuses on the compatibility between image and text embeddings of different Vision-Language Pretraining (VLP) models. We proposed an efficient solution using a text-only pretrained projection module, \\(\\phi\\), to align the new model's embeddings with those of the old model, thereby enhancing training efficiency. By integrating parameter-efficient training schemes into the XBT framework, we were able to accelerate the model's training while preserving the original VLP's zero-shot capabilities. Our approach, demonstrated on various cross-modal benchmarks, effectively builds cross-modal retrieval systems without backfilling, offering an efficient and environmentally friendly solution in response to the VLP improvements." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1790" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.173 + ], + "angle": 0, + "content": "The first author performed the initial research and development of this work in collaboration with Ser-nam Lim. The finalization and refinement of this work were completed while the first author was affiliated with Google DeepMind." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.182, + 0.188, + 0.198 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.486, + 0.262 + ], + "angle": 0, + "content": "[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.264, + 0.484, + 0.303 + ], + "angle": 0, + "content": "[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.305, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.334, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[4] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In IEEE international conference on automatic face & gesture recognition. IEEE, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[5] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. In ICLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.484, + 0.484 + ], + "angle": 0, + "content": "[6] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.484, + 0.539 + ], + "angle": 0, + "content": "[7] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.484, + 0.595 + ], + "angle": 0, + "content": "[8] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, Yoohoon Kang, and Sangdoo Yun. Language-only efficient training of zero-shot composed image retrieval. arXiv preprint arXiv:2312.01998, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.484, + 0.623 + ], + "angle": 0, + "content": "[9] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.624, + 0.486, + 0.691 + ], + "angle": 0, + "content": "[10] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.693, + 0.486, + 0.747 + ], + "angle": 0, + "content": "[11] Jiun Tian Hoe, Kam Woh Ng, Tianyu Zhang, Chee Seng Chan, Yi-Zhe Song, and Tao Xiang. One loss for all: Deep hashing with a single cosine similarity based learning objective. NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.749, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[12] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[13] Weihua Hu, Rajas Bansal, Kaidi Cao, Nikhil Rao, Karthik Subbian, and Jure Leskovec. Learning backward compatible embeddings. In ACM SIGKDD, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.859 + ], + "angle": 0, + "content": "[14] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In ACM ICMR, pages 39-43, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[15] Young Kyun Jang and Nam Ik Cho. Generalized product quantization network for semi-supervised image retrieval. In CVPR, 2020." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.207, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.133 + ], + "angle": 0, + "content": "[16] Young Kyun Jang, Geonmo Gu, Byungsoo Ko, Isaac Kang, and Nam Ik Cho. Deep hash distillation for image retrieval. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.907, + 0.189 + ], + "angle": 0, + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.907, + 0.232 + ], + "angle": 0, + "content": "[18] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.906, + 0.272 + ], + "angle": 0, + "content": "[19] Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. How to adapt your large-scale vision-and-language model. 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.907, + 0.314 + ], + "angle": 0, + "content": "[20] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.908, + 0.357 + ], + "angle": 0, + "content": "[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.907, + 0.413 + ], + "angle": 0, + "content": "[22] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.906, + 0.467 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.906, + 0.496 + ], + "angle": 0, + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[25] Jie Lu, Vahid Behbood, Peng Hao, Hua Zuo, Shan Xue, and Guangquan Zhang. Transfer learning using computational intelligence: A survey. Knowledge-Based Systems, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.907, + 0.566 + ], + "angle": 0, + "content": "[26] Qiang Meng, Chixiang Zhang, Xiaogiang Xu, and Feng Zhou. Learning compatible embeddings. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[27] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. In EMNLP, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.908, + 0.678 + ], + "angle": 0, + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.908, + 0.732 + ], + "angle": 0, + "content": "[29] Vivek Ramanujan, Pavan Kumar Anasosalu Vasu, Ali Farhadi, Oncel Tuzel, and Hadi Pouransari. Forward compatible training for large-scale embedding retrieval systems. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.908, + 0.775 + ], + "angle": 0, + "content": "[30] Mehwish Rehman, Muhammad Iqbal, Muhammad Sharif, and Mudassar Raza. Content based image retrieval: survey. World Applied Sciences Journal, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.908, + 0.831 + ], + "angle": 0, + "content": "[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[32] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1791" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[33] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, pages 2556-2565, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[34] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[35] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[37] Ji Wan, Dayong Wang, Steven Chu Hong Hoi, Pengcheng Wu, Jianke Zhu, Yongdong Zhang, and Jintao Li. Deep learning for content-based image retrieval: A comprehensive study. In ACMMM, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[38] Chien-Yi Wang, Ya-Liang Chang, Shang-Ta Yang, Dong Chen, and Shang-Hong Lai. Unified representation learning for cross model compatibility. In BMVC, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[39] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. NeurIPS, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[40] Kai Wang, Luis Herranz, and Joost van de Weijer. Continual learning in cross-modal retrieval. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[41] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[42] Sijie Yan, Yuanjun Xiong, Kaustav Kundu, Shuo Yang, Siqi Deng, Meng Wang, Wei Xia, and Stefano Soatto. Positive-congruent training: Towards regression-free model updates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14299-14308, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[43] Binjie Zhang, Yixiao Ge, Yantao Shen, Yu Li, Chun Yuan, Xuyuan Xu, Yexin Wang, and Ying Shan. Hot-refresh model upgrades with regression-alleviating compatible training in image retrieval. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[44] Binjie Zhang, Yixiao Ge, Yantao Shen, Shupeng Su, Chun Yuan, Xuyuan Xu, Yexin Wang, and Ying Shan. Towards universal backward-compatible representation learning. In IJCAI, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.772 + ], + "angle": 0, + "content": "[45] Liangli Zhen, Peng Hu, Xi Peng, Rick Siow Mong Goh, and Joey Tianyi Zhou. Deep multimodal transfer learning for cross-modal retrieval. TNNLS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[46] Wengang Zhou, Houqiang Li, and Qi Tian. Recent advance in content-based image retrieval: A literature survey. arXiv preprint arXiv:1706.06064, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[47] Yifei Zhou, Zilu Li, Abhinav Shrivastava, Hengshuang Zhao, Antonio Torralba, Taipeng Tian, and Ser-Nam Lim. bt2: Backward-compatible training with basis transformation. arXiv preprint arXiv:2211.03989, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[48] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "comprehensive survey on transfer learning. Proceedings of the IEEE, 2020." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1792" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/5c932632-b5a7-42ba-9baf-b3ca38b774d6_origin.pdf b/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/5c932632-b5a7-42ba-9baf-b3ca38b774d6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8601266c1400f1308d624a10f982b6a34d5fddf2 --- /dev/null +++ b/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/5c932632-b5a7-42ba-9baf-b3ca38b774d6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8954c980556105e756fa5c90c1fae9e6568bdfc9444901d8debb38570f9090a2 +size 2038347 diff --git a/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/full.md b/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0b7d3f7d8fc1768d2602f4b2a82eda556c262d83 --- /dev/null +++ b/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/full.md @@ -0,0 +1,314 @@ +# Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models + +Young Kyun Jang1 + +Google DeepMind + +kyun0914@gmail.com + +Ser-nam $\mathrm{Lim}^2$ + +$^{2}$ University of Central Florida + +sernam@gmail.com + +# Abstract + +Modern retrieval systems often struggle with upgrading to new and more powerful models due to the incompatibility of embeddings between the old and new models. This necessitates a costly process known as backfilling, which involves re-computing the embeddings for a large number of data samples. In vision, Backward-compatible Training (BT) has been proposed to ensure that the new model aligns with the old model's embeddings. This paper extends the concept of vision-only BT to the field of cross-modal retrieval, marking the first attempt to address Cross-modal BT (XBT). Our goal is to achieve backward-compatibility between Vision-Language Pretraining (VLP) models, such as CLIP, for the cross-modal retrieval task. To address XBT challenges, we propose an efficient solution: a projection module that maps the new model's embeddings to those of the old model. This module, pretrained solely with text data, significantly reduces the number of image-text pairs required for XBT learning, and, once it is pretrained, it avoids using the old model during training. Furthermore, we utilize parameter-efficient training strategies that improve efficiency and preserve the off-the-shelf new model's knowledge by avoiding any modifications. Experimental results on cross-modal retrieval datasets demonstrate the effectiveness of XBT and its potential to enable backfill-free upgrades when a new VLP model emerges. + +# 1. Introduction + +As the volume and variety of data grow exponentially in our multimedia-rich era, developing and maintaining efficient multi-modal retrieval systems becomes increasingly challenging. These systems, which provide data samples relevant to a user's query, must handle diverse data types, from text and images to audio and video. This growth puts a premium on the scalability and performance of these systems, necessitating continuous advancements in algorithms and technology. Embedding-based deep models for + +![](images/e0e2ad6479550047c00477deca97bf74e0efc4a26999cc0a8d59d5b9f88c00db.jpg) +Backward-compatible Training (BT) + +![](images/014ff1aa669b621ec3ffc021639fc59f131d860af096d99852d58d8164515017.jpg) +Cross-modal Backward-compatible Training (XBT) +Figure 1. A conceptual visualization of Backward-compatible Training (BT, left) and its extension, the Cross(X)-modal version (XBT, right). Circles and squares denote data samples of images and text, respectively. XBT uses Vision-Language Pretraining (VLP) models as baselines to achieve cross-modal, backward-compatible representation learning, allowing the new, improved model to be compatible with the fixed old model. + +retrieval have emerged as a key solution, transforming high-dimensional data into a lower-dimensional dense embedding space [11, 15, 16, 30, 37, 46]. These models capture the semantic meanings of data samples, enabling the quantification of similarities for retrieval. + +However, it is important to note that the embedding spaces generated by different deep models are not inherently compatible with each other. This incompatibility restricts the + +reuse of an existing gallery that has been embedded with an older model when a new, better-performing model is introduced. Consequently, this necessitates "backfilling," a process where the entire gallery must be rebuilt using the embeddings from the new model. Such a requirement is resource-intensive and time-consuming, posing a significant challenge when building retrieval systems. + +Backward-compatible Training (BT) [13, 34, 38, 43, 44] has been developed to tackle this issue, specifically focusing on image retrieval systems. The main objective of BT is to train a new model from scratch in a manner that ensures its compatibility with an old model that was used to create the existing gallery. A successful BT model must consequently demonstrate that retrieving from the old gallery using a query embedded with the new model outperforms that using an embedding from the old model. This enhancement is crucial for justifying the application of BT. + +Taking the BT problem a step further, we propose a new challenging task called Cross(X)-modal Backward-compatible Training (XBT) as shown in Figure 1. Our focus here is on applying BT principles within the realm of cross-modal retrieval, specifically exploring the interaction and compatibility between image and text embeddings of different Vision-Language Pretraining (VLP) models [17, 21, 22, 28, 35]. + +To achieve XBT, just like the BT problem, we need to resolve the incompatibility between the embeddings of the old, inferior model and the new, improved model. If we follow most prior practices of training from scratch with an additional compatibility loss term, a substantial number of supervised image-text pairs would be needed. The ideal quantity would be in the hundreds of millions, approximating the amount utilized in the establishment of VLP models. However, accessing the original training samples is often impossible, and training on that scale is prohibitively expensive and impractical. We propose an efficient solution instead: a text-only pretrained projection module, $\phi$ , to align a given pretrained new model's embeddings with those of the old model. + +Our focus is on using only text data to estimate the entire distribution of embeddings in the VLP space. Specifically, we train $\phi$ to learn the similarity between old and new text embeddings in a contrastive manner. By increasing the number of text samples, which is simpler than preparing image-text pairs, $\phi$ can approximate the oracle projection between the intra-modal distributions of both old and new embeddings for texts. Assuming that the intra-modal distribution of texts in the VLP embedding space can mirror that of images, we can simply apply the same $\phi$ to the new image embeddings to synthesize the corresponding old image embeddings. + +With these generated synthetic old image and text embeddings, which can be considered as aligned with the new + +embeddings, we aim to fine-tune the new model. In this process, far fewer supervised image-text pairs are required than in the original training dataset of the VLP. This also allows us to avoid using old model parameters during training, thereby enhancing training efficiency. In addition, we incorporate parameter-efficient fine-tuning schemes [12, 19, 20] into the new model, which add small trainable parameters and do not harm any of the new model's original parameters during training. This not only accelerates the new model's fine-tuning process but also allows for an easy reversion to the original new-to-new retrieval by simply removing the additional parameters. + +We demonstrate our approach on various cross-modal benchmarks, highlighting the effectiveness of XBT in cross-modal retrieval protocols. In response to the rapid advancement of VLP models, XBT offers an environmentally friendly solution by enabling backfill-free systems. + +Our contributions can be summarized as: + +- The Cross-modal Backward-compatible Training (XBT) concept is introduced for the first time to solve the backfilling problem that stems from real-world cross-modal retrieval systems. +- A novel XBT solution is proposed, which uses a text-only pretrained projection module, $\phi$ , to efficiently align the embeddings of new and old models using only text samples. +- The proposal is demonstrated on various datasets and protocols, showcasing XBT's effectiveness in building backfill-free cross-modal retrieval systems. + +# 2. Related Works + +Backward-compatible Training. The concept of Backward-compatible Training (BT) was first introduced in the study ([34]). This approach influences a new model with the learned classifier of the old model, thereby achieving backward compatibility between the old and new models. However, BT can degrade the original representational performance of the new model. To overcome this, [26] proposed aligning class-wise centers presented by the old and new models. Another approach to achieve backward compatibility is to use an additional projection to map the old embedding into the new embedding space by adding a lightweight transformation, as suggested in [13, 38]. The work of [29] further adds an auxiliary feature in preparation for future updates, while [47] uses additional dimensionality in the embedding to obtain compatibility. An online strategy that backfills the gallery on the fly is explored in [43], and [42] addresses the model regression problem. Despite this progress, the scenario considering cross-modal retrieval between image and text, which has many real-world applications, remains unexplored. Our study on XBT in this paper aims to fill this gap. + +Vision-Language Continual and Transfer Learning The fields of continual learning [2, 5, 7] and transfer learning [25, 48] bear similarities to backward-compatible representation learning, as all aim to update an existing model to boost performance. In the realm of multi-modal representation learning for cross-modal retrieval, continual learning approaches like [40] propose methods to prevent catastrophic forgetting across different modalities. Transfer learning approaches like [45] suggest strategies for transferring knowledge from previously labeled categories (source modality) to new, unlabeled categories (target modality). However, our proposed XBT stands apart from these methods as it specifically tackles the challenge of maintaining backward compatibility between old and new models. This unique attribute makes XBT ideally suited for backfill-free retrieval scenarios, where the embeddings of the old model remain unchanged, yet we can still leverage the enhanced performance of the new model. + +# 3. Method + +Our goal is to construct a backfill-free, embedding-based, cross-modal retrieval system using VLP models, which are configured with an image encoder $E^{I}$ and a text encoder $E^{T}$ . When a new better performing VLP model $\{E_{new}^{I}, E_{new}^{T}\}$ emerges, we aim to ensure its compatibility with the old model $\{E_{old}^{I}, E_{old}^{T}\}$ that was used to construct the gallery. To achieve this, we introduce Cross(X)-modal Backward-compatible Training (XBT). + +Given an image $x$ and text caption $t$ , XBT enables retrieval between a new image embedding $v_{new} = E_{new}^{I}(x)$ and the text embeddings $(w_{old})$ , as well as between a new text embedding $w_{new} = E_{new}^{T}(t)$ and the image embeddings $(v_{old})$ . We denote backward compatible embeddings as $\bar{v}$ and $\bar{w}$ . All embeddings we utilize in this work are l2-normalized. + +# 3.1. Criterion for Cross-modal Backward Compatibility + +Following the definition of backward compatibility in BT work [34], we can construct strict constraints that ensure cross-modal backward compatibility as: + +$$ +\begin{array}{l} d \left(w _ {n e w _ {i}}, v _ {o l d _ {j}}\right) \leq d \left(w _ {o l d _ {i}}, v _ {o l d _ {j}}\right), \forall y _ {i} = y _ {j}, \\ d \left(w _ {\text {n e w} _ {i}}, v _ {\text {o l d} _ {j}}\right) \geq d \left(w _ {\text {o l d} _ {i}}, v _ {\text {o l d} _ {j}}\right), \forall y _ {i} \neq y _ {j}, \tag {1} \\ d (v _ {n e w _ {i}}, w _ {o l d _ {j}}) \leq d (v _ {o l d _ {i}}, w _ {o l d _ {j}}), \forall y _ {i} = y _ {j}, \\ d (v _ {n e w _ {i}}, w _ {o l d _ {j}}) \geq d (v _ {o l d _ {i}}, w _ {o l d _ {j}}), \forall y _ {i} \neq y _ {j}, \\ \end{array} +$$ + +where $y_{i}$ and $y_{j}$ represent whether an image and text are paired ( $y_{i} = y_{j}$ ) or not ( $y_{i} \neq y_{j}$ ). The term $d(\cdot, \cdot)$ represents a distance metric in the embedding space, and we choose cosine distance as the baseline. The constraints in Eqn. 1 formally express the requirement that the new embedding + +must perform at least as well as the old embedding in terms of correctly matching image-text pairs. + +However, exhaustively satisfying these constraints is intractable due to the potential for the old embeddings to outperform the new embeddings in certain retrieval cases. As a result, we relax the criteria by defining an alternative evaluation metric as: + +$$ +\begin{array}{l} \begin{array}{l} \mathcal {M} (w _ {\text {n e w}}, v _ {\text {o l d}}; \mathrm {Q}, \mathrm {G}) > \mathcal {M} (w _ {\text {o l d}}, v _ {\text {o l d}}; \mathrm {Q}, \mathrm {G}), \\ \mathcal {A} ^ {\prime} (\mathrm {Q}, \mathrm {G}) = \mathcal {A} ^ {\prime} (\mathrm {Q}, \mathrm {G}) \end{array} \tag {2} \\ \mathcal {M} (v _ {n e w}, w _ {o l d}; \mathrm {Q}, \mathrm {G}) > \mathcal {M} (v _ {o l d}, w _ {o l d}; \mathrm {Q}, \mathrm {G}), \\ \end{array} +$$ + +where a metric, such as recall, $\mathcal{M}(q,g;\mathrm{Q},\mathrm{G})$ takes query embedding $q$ and gallery embedding $g$ over query set Q and gallery set G. This criterion suggests that the general performance is enhanced when the query embedding from the XBT-trained new model is used to perform retrieval with the gallery of the old model, compared to the performance of the old model alone. In essence, fulfilling Eqn. 2 indicates that the new model has achieved backward compatibility and can feasibly update without backfilling gallery. + +Additionally, it is important to highlight that VLP models function as zero-shot learners. This leads us to define the XBT problem differently from classic BT, which trains separate models for specific image domain retrieval tasks such as ImageNet [31] or VGGFace2 [4]. In contrast, XBT tackles a more challenging task, aiming to bridge a new VLP model with a frozen old VLP model while retaining the new model's original zero shot capability. In this paper, we therefore assess the performance of XBT using retrieval and classification benchmarks in a zero-shot manner. + +# 3.2. Text-only Pretraining + +The vast and diverse image-text pairs used to build VLP models significantly enhance their ability to connect semantically similar visual content and natural language [6]. However, this creates a complex embedding distribution that is difficult to predict and understand, thereby complicating the XBT process. A straightforward solution is to use a large number of supervised image-text pairs, similar to the approach used when building VLP models from scratch, to estimate the entire distribution of image and text embeddings during XBT. + +However, acquiring a sufficient number of accurate supervised pairs is extremely costly. To be far more efficient, we employ a small sized projection module, $\phi$ , and train it exclusively with text samples, as shown in Figure 2. We hypothesize that the distribution of text embeddings in VLP models, which is determined by their semantic similarity, is similarly mirrored in the distribution of their corresponding matched images. With this in mind, we aim to train $\phi$ to cover the broad spectrum of the text embedding distribution between the new and old VLP models, making embeddings + +![](images/39336d708d8dd5f82225376033f56341510ecb6f23cff56980107b9b142ccf28.jpg) +Text-only Pretraining + +![](images/5540558dc3f4240ed4f4e8c844614f06cb25c9e88fc1897a1a3ed78cf7071704.jpg) +Figure 2. An illustration of the text-only pretraining of $\phi$ (above), and XBT with $\phi$ (below). Using only text samples, $\phi$ is trained to approximate distribution of old text embeddings from that of new ones. After pretraining, the same $\phi$ is used to generate both of synthetic old image and text embeddings from the new VLP embeddings and train to learn cross-modal backward-compatible representation. + +from the same text sample similar and others dissimilar in a contrastive way: + +$$ +\mathcal {L} _ {p r e} = \mathbb {E} _ {t \sim D _ {T}} \left[ \mathcal {L} _ {c} \left(\phi \left(w _ {\text {n e w}} + \epsilon\right), w _ {\text {o l d}}; \tau_ {p r e}\right) \right], \tag {3} +$$ + +where $t$ denotes text data sampled from text corpus $D_{T} = \{t_{i}\}_{i=1}^{N_{D_{T}}} \text{ of } N_{D_{T}} \text{ text samples, and } \mathcal{L}_{c}(\cdot, \cdot) \text{ is a standard batch-wise contrastive loss as defined in CLIP [28] of temperature } \tau_{pre}.$ + +To prevent overfitting towards the text domain, which could widen the discrepancy between the image and text embedding representations, we inject noise $\epsilon$ sampled from a Gaussian distribution $(\epsilon \sim N(0,\sigma^2))$ , inspired by the approaches [8, 27]. Notably, $\phi$ is designed to take the new VLP model's embeddings, which means that $w_{new} + \epsilon$ is also $l2$ -normalized. + +During this stage, only $\phi$ is updated, while all encoders remain fixed. By expanding the scale of $D_{T}$ , we can effectively replicate the complete text embedding spaces of both new and old VLP models. Additionally, $\phi$ can be applied to new image embeddings to generate synthetic old image embeddings that closely approximate the actual old ones. + +![](images/b9379e2bd6f070cf88c957d8a78f1f13e57f99c4976019a8179f7966262acf8b.jpg) +XBT Learning Process +Figure 3. Our proposed learning process to achieve XBT. Notably, the old VLP model's encoders are not required in this stage, enhancing efficiency in training. + +# 3.3. Cross-modal Backward-compatible Training + +Training Loss. The objective of XBT is to ensure cross-modal compatibility, specifically between $w_{new}$ and $v_{old}$ , as well as between $v_{new}$ and $w_{old}$ . For a given dataset $D = \{x_i, t_i\}_{i=1}^{N_D}$ of $N_D$ supervised image-text pairs, we aim to train the new VLP model encoders, $E_{new}^I$ and $E_{new}^T$ . Note that in our base setting, the text corpus is significantly larger than the supervised dataset, i.e., $N_{D_T} \gg N_D$ . + +However, the dimensionality of new and old VLP embeddings may differ, and even if they are the same, they are not be directly compatible. To address this, we apply the pretrained $\phi$ to ensure that the new embeddings match the dimension of the old ones and project into the compatible embedding space, as shown in Figure 2. This can be represented as: + +$$ +\phi (v _ {\text {n e w}}) = \bar {v} _ {\text {n e w}}, \phi (w _ {\text {n e w}}) = \bar {w} _ {\text {n e w}}, \tag {4} +$$ + +where $\bar{v}_{new}$ and $\bar{w}_{new}$ denote synthetic old embeddings and $\{\bar{v}_{new},\bar{w}_{new},v_{old},w_{old}\} \subset \mathbb{R}^K$ , with $K$ denoting the dimensionality of the old VLP model's embedding. + +We then apply XBT loss $\mathcal{L}_X$ as follows: + +$$ +\mathcal {L} _ {X} = \mathbb {E} _ {x, t \sim D} \left[ \mathcal {L} _ {c} \left(\bar {v} _ {\text {n e w}}, \bar {w} _ {\text {n e w}}; \tau_ {X}\right) \right]. \tag {5} +$$ + +Here, $\mathcal{L}_c(\cdot ,\cdot)$ is the same contrastive loss used in Eqn. 3, with temperature $\tau_{X}$ . As shown in Figure 3, the new VLP encoders, $E_{new}^{I}$ and $E_{new}^{T}$ , and $\phi$ are trained through $\mathcal{L}_X$ in an end-to-end manner. Ultimately, $\bar{v}_{new}$ and $\bar{w}_{new}$ become cross-modal backward-compatible with existing old gallery embeddings, $v_{old}$ and $w_{old}$ , as all embeddings are distributed in the compatible space through $\phi$ . + +Efficient Training. The efficiency of XBT can be attributed to two design choices that avoid: (1) any dependence on the old model when conducting XBT, and (2) updating the new VLP model off the shelf. We have already seen how $\phi$ can effectively achieve (1). (2) is simply facilitated by applying small-sized additional parameters, namely Soft prompt [20] and LoRA[12]. + +To be more specific, we incorporate the concept of soft prompt tuning, as outlined in [18, 20]. Soft prompting is applied to the vision encoder $E_{new}^{I}$ by adding trainable prompts as input, along with image patch tokens. We exclude soft prompt tuning on the text side, as we use the text-only pretrained module $\phi$ , which requires maintaining the original + +distribution of text embeddings. We adopt the LoRA strategy [12] for the new VLP model's encoders $E_{new}^{I}$ and $E_{new}^{T}$ to avoid modifying original parameters. These two factors, in conjunction with $\phi$ , offer an on-off solution: we can retrieve old samples using additional parameters, and we can easily revert to the original model by removing these parameters for new-to-new retrieval. + +Further, regarding $\phi$ during XBT (Sec. 3.3), we only finetune the parameters of layer normalization. This is done with the aim of preserving the knowledge learned during the text-only pretraining stage (Sec. 3.2) and facilitating easy adaptation towards the image-text joint representation space + +In the end, our XBT framework offers a parameter-efficient solution that enables rapid convergence with fewer training iterations (a single epoch is sufficient); maintains the new model's power; and requires far fewer supervised training samples than the scale of VLP model pretraining (approximately $1\%$ of the level required to train CLIP [28] from scratch, which uses a 400M dataset, and $0.2\%$ of the level required for LAION-2B based CLIP models [32], which utilize a 2B dataset). All these help to avoid any full-scale retraining of the new VLP models, paving the way for XBT to simply use new VLP models off-the-shelf as mentioned. + +# 4. Experiments + +The evaluation of the proposed cross-modal backward compatibility includes image-text zero-shot retrieval. Pretrained VLP models are used as our baseline, with the goal to assess performance in a zero-shot environment by tuning a given pretrained new VLP model that is stronger than the old model. The results highlight the potential of XBT, showing improved performance when old model outputs are matched with the XBT-learned new model. + +# 4.1. Setup + +Model Training. Our XBT process involves two distinct training stages: text-only pretraining (See Sec. 3.2) and image-text supervised training (See Sec. 3.3). For the text-only pretraining stage, we utilize the text samples from a subset of the 115M filtered web-collected image-text paired dataset from BLIP [21], comprising around 67M available pairs (58.2% of the total). We construct a text corpus $D_{T}$ in Eqn. 3 using these text samples. For the subsequent image-text supervised training stage, we use a smaller subset of 4M image-text pairs from the same dataset to construct a supervised dataset $D$ in Eqn. 5. It is important to note that these subsets are significantly smaller than the dataset scale (400M, 2B or more) used to build CLIP models [28, 32]. + +We utilize 8-A100-80GB GPUs to train and evaluate the models. For the text-only pretraining stage in Sec. 3.2, the batch size is set to 8,192 (1,024 batch per GPU) and during this stage, the entire set of trainable parameters of $\phi$ are trained while both VLP text encoders are fixed. + +Moving on to the image-text supervised training in Sec. 3.3, the batch size is reduced to 1,024 (128 batch per GPU), and in this stage, layer normalization is set to be the only trainable component for all VLP image and text encoders with $\phi$ . Temperature hyper-parameters $\tau_{pre}$ , $\tau_{X}$ , and $\tau_{N}$ are fixed at 2.6592. We employ the AdamW optimizer [24] with a fixed learning rate of 1e-4 and a weight decay of 0.01. For soft prompts, we use 10 prompts and apply 100 times larger learning rate, 1e-2. The training iteration is determined by the dataset size, and the entire pipeline is trained for a single epoch. For image augmentation, we begin with a random resized crop, adjusting the image scale between 0.5 and 1.0. Additionally, we apply a random horizontal flip and make random adjustments to the image's contrast, brightness, and sharpness. To incorporate different perspectives and angles, we modify the image's translation and rotation. The pretrained weights provided by HuggingFace [41] are applied to baseline VLP models. We employ as: openai/clip-vit-base-patch32, openai/clip-vit-target-patch14, laion/CLIP-ViT-H-14-laion2B-s32B-b79K. + +Model Evaluation. We validate the effectiveness of XBT with a benchmark comprising three popular image-text paired datasets for cross-modal retrieval evaluation. The first is the nocaps dataset [1], which offers a diverse distribution of object names and facilitates a more detailed analysis. We employ the validation split of this dataset, which consists of images each paired with 10 relevant textual captions, totaling 4,500 images and 45,000 captions. + +To evaluate XBT at a larger scale, we also include the Flickr [14] and COCO [23] datasets, which consist of images that are each paired with 5 relevant textual captions. For the Flickr dataset, we use the entire dataset, encompassing 31,783 images and 158,915 captions. For the COCO dataset, we use the validation split, which includes 35,136 images and 175,680 captions. We note the two datasets as Large-Flickr and Large-COCO respectively for the remainder of the paper. + +For retrieval, we utilize all the samples in each dataset for both query and gallery, following the cross-modal benchmark used in [21, 28]. For evaluation purpose, we adopt recall scores at top K retrieval results $(\mathbb{R}@\mathbb{K},\%)$ to estimate the cross-modal backward compatibility (See Eqn. 2). + +Implementation Details. In our work, we primarily use the popular CLIP models [28], based on a Transformer [36] backbone, as our baseline VLP models. The models, listed in ascending order of scale and performance, are: CLIP-ViT-B32, CLIP-ViT-L14, and CLIP-ViT-H14. For simplicity, we will refer to them as B32, L14, and H14, respec + +Table 1. Cross-modal retrieval results on nocaps. Note that for comparison, the B32/L14 case should be compared with B32, and the L14/H14 case should be compared with L14, respectively. + +
Old Model / New ModelMethodText Query(w, w)/Image Gallery(v)Image Query(v, v)/Text Gallery(w)
CaseR@1R@5R@10R@50CaseR@1R@5R@10R@50
Original
CLIP-ViT-B32-w/v45.1474.9884.5196.00v/w71.3892.0296.3399.67
CLIP-ViT-L14-w/v47.7776.5085.1395.95v/w73.2993.2497.4099.78
Cross-modal Backward Compatible Training
CLIP-ViT-B32 / CLIP-ViT-L14Full-tune\( \bar{w}_{new}/v_{old} \)41.6073.5484.2696.45\( \bar{v}_{new}/w_{old} \)63.7687.0093.5199.09
LoRA-only43.4074.6684.9496.7368.8290.8796.0099.76
Base43.4874.8785.0496.7469.7891.3196.0299.78
XBT48.0279.0088.2197.6675.0293.2797.3199.91
Full-tune\( \bar{w}_{new}/\bar{v}_{new} \)48.9279.3188.2697.64\( \bar{v}_{new}/\bar{w}_{new} \)61.2285.8992.6299.04
LoRA-only55.4483.8291.2498.3470.1191.1396.4999.73
Base56.1884.3191.4498.3771.0791.4496.2299.76
XBT63.4888.6193.9898.8677.2294.5697.9199.89
CLIP-ViT-L14 / CLIP-ViT-H14Full-tune\( \bar{w}_{new}/v_{old} \)45.1877.8285.3497.65\( \bar{v}_{new}/w_{old} \)66.3888.5593.5998.90
LoRA-only47.0076.9486.0297.6372.0492.4296.5899.67
Base47.0677.1586.1297.6473.0092.8696.6099.69
XBT51.1480.8288.9398.7276.6294.1197.5899.84
Full-tune\( \bar{w}_{new}/\bar{v}_{new} \)54.3982.2689.0698.95\( \bar{v}_{new}/\bar{w}_{new} \)65.5787.5293.5199.13
LoRA-only60.9186.7792.0498.6574.4692.7397.3899.82
Base61.6587.2692.2498.6875.4293.0497.1199.85
XBT66.4790.2194.9299.0580.0296.1198.87100.0
+ +tively. Notably, while B32 and L14 are trained with the same dataset at 400M scale, H14 is trained with a larger dataset at 2B scale. We apply additional LoRA [12] parameters for each new model encoder configured as follows: $\mathrm{LoRA}_{\alpha} = 16$ , $\mathrm{rank} = 16$ , and dropout $= 0.1$ . The proposed projection module $\phi$ consists of three Linear layers with layer normalization (LN) [3] and the GELU nonlinearity function [9]. Detailed architecture is as follows: Linear-LN-GELU-Linear-LN-GELU-Linear. Dropout is not applied as it was found to degrade performance empirically. The intermediate hidden dimension of the Linear layer is set to be four times the dimensionality of the output embedding. + +# 4.2. Main Results + +Baseline Comparisons. To simplify notation, we use $w / v$ to denote the retrieval results obtained using $w$ as the query embeddings and $v$ as the gallery embeddings. We establish two protocols: assessing cross-modal backward compatibility with new/old ( $\bar{w}_{\text{new}} / v_{\text{old}}$ , $\bar{v}_{\text{new}} / w_{\text{old}}$ ), and maintaining the new VLP's original performance with new/new ( $\bar{w}_{\text{new}} / \bar{v}_{\text{new}}$ , $\bar{v}_{\text{new}} / \bar{w}_{\text{new}}$ ). All of the above are similarly applied to $v / w$ . + +To the best of our knowledge, since this paper presents the first method that aims to solve the XBT problem, there are no direct previous works to compare to. Nevertheless, we compare our XBT against three baselines: Full-tune, LoRA-only, and Base, which we elaborate on below. + +Naive solution - Direct backward contrastive learning: + +The pretrained $\phi$ in Eqn. 3, which maps the new embeddings to the old via text based contrastive learning, constitutes a major contribution of XBT. We therefore would like to compare with a solution that does not utilize such a pretrained $\phi$ . A straightforward solution for this is to fine-tune the new VLP encoders $(E_{new}^{I}$ and $E_{new}^{T})$ to be compatible with the old ones by minimizing the following loss: + +$$ +\begin{array}{l} \mathcal {L} _ {D i r e c t} = \mathbb {E} _ {x, t \sim D} [ \mathcal {L} _ {c} (\bar {v} _ {\text {n e w}}, w _ {\text {o l d}}; \tau_ {N}) \\ \left. + \mathcal {L} _ {c} \left(\bar {w} _ {\text {n e w}}, v _ {\text {o l d}}; \tau_ {N}\right) \right] \tag {0} \\ \end{array} +$$ + +where $\mathcal{L}_c(\cdot, \cdot)$ is the same contrastive loss used in Eqn. 5, and $\tau_N$ is a temperature. Here, we apply a randomly initialized $\phi$ with the same configuration as XBT for projection, as per Eqn. 4, to generate cross-modal backward compatible embeddings, $\bar{w}$ and $\bar{v}$ . + +Different training options: Based on Eqn. 6, we consider multiple training setups: (1) full fine-tuning of all trainable components (Full-tune), (2) tuning LoRA parameters-only (LoRA-only), and (3) starting from LoRA-only, adding an extra learnable prompt (Base). Models produced by Base would therefore be akin to XBT without the pretrained $\phi$ . For these, we train the $E_{new}^{I}$ , $E_{new}^{T}$ and $\phi$ using $\mathcal{L}_{Direct}$ with 4M image-text pairs. Note that, XBT only trains layer-normalization layers including pretrained $\phi$ , however, Full-tune, LoRA-only and Base setups are training entire parameters of $\phi$ since it is not trained before. The results of these baselines are presented in Table 1, tested with nocaps dataset. The performance of the original VLP models is also reported + +Table 2. Cross-modal retrieval results on Large-Flickr and Large-COCO. + +
DatasetOld Model / New ModelMethodText Query(w, w)/Image Gallery(v)Image Query(v, v)/Text Gallery(w)
CaseR@1R@5R@10R@50CaseR@1R@5R@10R@50
Original
CLIP-ViT-B32-w/v21.5441.2450.7272.69v/w40.3564.4273.3989.91
Cross-modal Backward Compatible Training
CLIP-ViT-B32 / CLIP-ViT-L14Base\( \bar{w}_{\text{new}}/v_{\text{old}} \)19.0237.2546.6369.32\( \bar{v}_{\text{new}}/w_{\text{old}} \)36.7460.6170.2288.34
XBT22.3842.7152.4174.8042.4766.0474.8790.79
Base\( \bar{w}_{\text{new}}/\bar{v}_{\text{new}} \)32.2955.1864.6883.60\( \bar{v}_{\text{new}}/\bar{w}_{\text{new}} \)36.1059.8069.9589.01
XBT39.5962.8971.7588.1343.5068.5077.9993.21
CLIP-ViT-B32 / CLIP-ViT-H14Base\( \bar{w}_{\text{new}}/v_{\text{old}} \)20.2739.1748.5970.67\( \bar{v}_{\text{new}}/w_{\text{old}} \)35.5359.4469.0687.67
XBT23.7044.4754.1976.0341.0965.1874.3290.16
Base\( \bar{w}_{\text{new}}/\bar{v}_{\text{new}} \)32.9455.9465.2683.77\( \bar{v}_{\text{new}}/\bar{w}_{\text{new}} \)34.8159.5569.9688.81
XBT40.4263.9072.7488.4746.3571.1880.0194.08
Original
CLIP-ViT-B32-w/v14.4430.2038.9662.30v/w28.6250.1759.6780.22
Cross-modal Backward Compatible Training
CLIP-ViT-B32 / CLIP-ViT-L14Base\( \bar{w}_{\text{new}}/v_{\text{old}} \)12.9728.1436.4859.78\( \bar{v}_{\text{new}}/w_{\text{old}} \)26.3246.9156.5477.86
XBT15.5532.2741.3465.1130.7352.5862.3081.83
Base\( \bar{w}_{\text{new}}/\bar{v}_{\text{new}} \)22.1641.6550.9873.06\( \bar{v}_{\text{new}}/\bar{w}_{\text{new}} \)27.2648.1657.8079.57
XBT27.6148.6958.0878.8332.9255.3764.8984.67
CLIP-ViT-B32 / CLIP-ViT-H14Base\( \bar{w}_{\text{new}}/v_{\text{old}} \)13.8829.5538.2361.57\( \bar{v}_{\text{new}}/w_{\text{old}} \)26.1946.7556.3077.96
XBT15.8232.7941.8465.4530.3351.2859.9781.66
Base\( \bar{w}_{\text{new}}/\bar{v}_{\text{new}} \)23.2043.2252.7074.68\( \bar{v}_{\text{new}}/\bar{w}_{\text{new}} \)27.3148.5658.6080.09
XBT28.1449.4858.8879.3234.5457.2966.7985.92
+ +for a clear comparison. We highlight our XBT method. + +In the context of cross-modal backward compatibility, where recall scores of $\bar{w}_{\text{new}} / v_{\text{old}}$ and $\bar{v}_{\text{new}} / w_{\text{old}}$ should be higher than those of $w / v$ and $v / w$ of old model, respectively, Full-tune significantly underperforms compared to the old VLP's. This indicates that full fine-tuning is not an appropriate solution. While both LoRA-only and Base improve performance over Full-tune, they still fall short of the old VLP's performance and fail to meet the criterion. In contrast, XBT significantly outperforms these baselines and even surpasses the old VLP's performance, satisfying the cross-modal backward compatibility for all recall metrics (see Eqn. 2). + +In terms of maintaining new VLP's performance, where recall scores of $\bar{w}_{\text{new}} / v_{\text{new}}$ and $\bar{v}_{\text{new}} / w_{\text{new}}$ could be similar to those of $w / v$ and $v / w$ of new model, respectively, all baselines show decent performance in $\bar{w}_{\text{new}} / v_{\text{new}}$ case but only XBT achieves better performance than new VLP model in $\bar{v}_{\text{new}} / w_{\text{new}}$ case. These results support that XBT not only maintains the performance of the new VLP model but also enhances it in certain cases. The superior performance of XBT in both cases underscores its promise as a robust solution for cross-modal retrieval tasks. + +Large Scale Retrieval. In this experiment, we assess the performance of cross-modal backward compatibility in more practical use cases. We use larger datasets, namely LargeFlickr and Large-COCO, as shown in Table 2. These datasets provide more than six times the number of samples typically + +used in image-text retrieval literature [6, 37], which make the retrieval process more challenging. We use the second-best performing Base method from Table 1 as a basis for comparison. To validate the applicability of XBT across various VLP models, we utilize three CLIP baselines: B32, L14, and H14. Additionally, we explore different combinations, such as B32/L14 and B32/H14, to further analyze the effectiveness of XBT, beyond the comparisons made in Table 1 (B32/L14 and L14/H32). + +When comparing B32/L14 with B32/H14, it is observed that every retrieval cases of improved models (L14 and H14) successfully maintain cross-modal backward compatibility, regardless of their scale. Especially, retrieval performance of B32/H14 is better than that of B32/L14, which demonstrates XBT's ability to preserve and align original model's power with old gallery. + +Upon examining the results, we consistently observe that XBT outperforms the Base model across all instances and both retrieval scenarios, often by a significant margin. Furthermore, XBT maintains the performance of new VLPs better than Base, achieving even larger margins for $\bar{w}_{new} / \bar{v}_{new}$ and $\bar{v}_{new} / w_{old}$ . In summary, XBT consistently demonstrates robust and superior retrieval performance, even in large-scale scenarios. + +# 4.3. Further Analysis + +Continual learning. In line with the literature on BT works [29, 34], we set up a continual learning scenario for a se + +Table 3. Continual learning scenario experimental results with XBT on nocaps. Note that for H14 training, we use L14, which has been previously adapted with B32, and H14 never encounters B14. + +
Old Model / New ModelText Query(w, w)/Image Gallery(v)Image Query(v, v)/Text Gallery(w)
CaseR@1R@5R@10R@50CaseR@1R@5R@10R@50
CLIP-ViT-B32 / CLIP-ViT-L14wnew/vold47.6878.0088.0097.50vnew/wold73.8992.0097.2299.86
wnew/vnew62.2286.1893.0098.12vnew/wnew76.2792.9096.7699.11
CLIP-ViT-L14 / CLIP-ViT-H14wnew/vold50.9780.6888.9897.72vnew/wold77.3894.8297.8099.84
wnew/vnew66.0989.7894.7799.01vnew/wnew80.6096.3698.7199.96
CLIP-ViT-B32 / CLIP-ViT-H14wnew/vold49.1479.8888.5997.67vnew/wold74.8493.7397.2299.80
wnew/vnew65.9589.6894.6898.99vnew/wnew80.9696.0098.7199.96
+ +![](images/86a116e654f354f096d1759859cedcff74f79fd43916b63a565fc65f1286ac48.jpg) +Figure 4. New query vs. Old gallery retrieval results on nocaps. B32 as old, and L14 as new model. + +# Retrieved from Old Text Gallery + +R#1 - A dog with colorful sunglasses on and a bandanna. +R#2 - A happy dog is wearing a blue bandanna and sunglasses. +R#3 - A dog wearing sunglasses and a blue and white bandanna. +R#4 - An old brown dog is wearing mirrored sunglasses and a blue bandana. +R#5 - A dog has its tongue out as it wears sunglasses and bandana. + +# New Query Text + +The cute brown +tabby and white +kitten, sits on the +hard wood floor. + +# Retrieved from Old Image Gallery + +![](images/776d7b9f9867f9f5e67996b4f82968b0904338f89af2208207e70a89a081c620.jpg) +R#1 + +Table 4. Ablation study tested on nocaps, with B32 as old and L14 as new model. + +
Method\(\bar{w}(\text{text})/v(\text{image})\)\(\bar{v}(\text{image})/w(\text{text})\)
R@1R@10R@1R@10
(a) Baseline48.0279.0075.0293.27
(b) w.o noise47.2478.0573.3192.49
(c) \(0.25 \times D_T\)46.8977.9173.2092.18
(d) \(0.5 \times D_T\)47.6878.0073.8992.00
(e) \(2 \times D\)47.6378.6974.9893.98
(f) \(0.5 \times D\)46.8077.8971.6992.00
(g) \(D=\text{CC3M}\)46.4477.7671.3391.84
(h) Image-only45.3276.9274.0292.35
+ +quence of model updates (old model B32, new model L14, and better new model H14) and present the retrieval results in Table 3. Initially, we apply XBT to L14 to ensure compatibility with B32. Subsequently, we apply XBT to H14 to ensure compatibility with the previously learned L32, which is already compatible with B32. For this process, we divide both the image-text pairs and the text-only pretraining train sets into two halves, using each separately for each case. A comparison with the original results reported in Table 3 confirms that our XBT performs well in the continual learning scenario either. It achieves cross-modal backward compatibility while leveraging the power of the improved new model. + +Ablation Study. To validate XBT, we conduct further analysis as shown in Table 4. By comparing (a) and (b), we observe that introducing noise during $\phi$ training aids in generalization. The comparison between (c) and (d) demonstrates that the scale of $D_{T}$ is important, supporting our + +assumption that a sufficient number of text samples can help build a robust $\phi$ . (e) outperforms (a), confirming that utilizing more image-text pairs can enhance XBT. The lower performance of (f) also aligns with the notion that the number of image-text pairs is crucial. In (g), when we replace $D$ with CC3M [33], from which we can obtain around 2.4M, the performance is similar to (f), suggesting that XBT can be effectively applied with other datasets. In (h), we replace text-only pretraining with image-only pretraining of the same scale and observe that text-only pretraining performs better in both retrieval scenarios, demonstrating its efficiency and superiority. + +Qualitative Results. We facilitate retrieval by utilizing new query embeddings and old gallery embeddings. The results in Figure 4 demonstrate accurate cross-modal backward-compatible retrieval. + +# 5. Conclusion + +In this paper, we introduced Cross-modal Backward-compatible Training (XBT), a novel task for cross-modal retrieval that focuses on the compatibility between image and text embeddings of different Vision-Language Pretraining (VLP) models. We proposed an efficient solution using a text-only pretrained projection module, $\phi$ , to align the new model's embeddings with those of the old model, thereby enhancing training efficiency. By integrating parameter-efficient training schemes into the XBT framework, we were able to accelerate the model's training while preserving the original VLP's zero-shot capabilities. Our approach, demonstrated on various cross-modal benchmarks, effectively builds cross-modal retrieval systems without backfilling, offering an efficient and environmentally friendly solution in response to the VLP improvements. + +# Acknowledgments + +The first author performed the initial research and development of this work in collaboration with Ser-nam Lim. The finalization and refinement of this work were completed while the first author was affiliated with Google DeepMind. + +# References + +[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In ICCV, 2019. +[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In CVPR, 2017. +[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016. +[4] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In IEEE international conference on automatic face & gesture recognition. IEEE, 2018. +[5] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. In ICLR, 2019. +[6] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023. +[7] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 2021. +[8] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, Yoohoon Kang, and Sangdoo Yun. Language-only efficient training of zero-shot composed image retrieval. arXiv preprint arXiv:2312.01998, 2023. +[9] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. +[10] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, 2021. +[11] Jiun Tian Hoe, Kam Woh Ng, Tianyu Zhang, Chee Seng Chan, Yi-Zhe Song, and Tao Xiang. One loss for all: Deep hashing with a single cosine similarity based learning objective. NeurIPS, 2021. +[12] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021. +[13] Weihua Hu, Rajas Bansal, Kaidi Cao, Nikhil Rao, Karthik Subbian, and Jure Leskovec. Learning backward compatible embeddings. In ACM SIGKDD, 2022. +[14] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In ACM ICMR, pages 39-43, 2008. +[15] Young Kyun Jang and Nam Ik Cho. Generalized product quantization network for semi-supervised image retrieval. In CVPR, 2020. + +[16] Young Kyun Jang, Geonmo Gu, Byungsoo Ko, Isaac Kang, and Nam Ik Cho. Deep hash distillation for image retrieval. In ECCV, 2022. +[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. +[18] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV. Springer, 2022. +[19] Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. How to adapt your large-scale vision-and-language model. 2021. +[20] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021. +[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. +[22] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. +[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2017. +[25] Jie Lu, Vahid Behbood, Peng Hao, Hua Zuo, Shan Xue, and Guangquan Zhang. Transfer learning using computational intelligence: A survey. Knowledge-Based Systems, 2015. +[26] Qiang Meng, Chixiang Zhang, Xiaogiang Xu, and Feng Zhou. Learning compatible embeddings. In ICCV, 2021. +[27] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. In EMNLP, 2022. +[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. +[29] Vivek Ramanujan, Pavan Kumar Anasosalu Vasu, Ali Farhadi, Oncel Tuzel, and Hadi Pouransari. Forward compatible training for large-scale embedding retrieval systems. In CVPR, 2022. +[30] Mehwish Rehman, Muhammad Iqbal, Muhammad Sharif, and Mudassar Raza. Content based image retrieval: survey. World Applied Sciences Journal, 2012. +[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. +[32] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022. + +[33] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, pages 2556-2565, 2018. +[34] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In CVPR, 2020. +[35] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022. +[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. +[37] Ji Wan, Dayong Wang, Steven Chu Hong Hoi, Pengcheng Wu, Jianke Zhu, Yongdong Zhang, and Jintao Li. Deep learning for content-based image retrieval: A comprehensive study. In ACMMM, 2014. +[38] Chien-Yi Wang, Ya-Liang Chang, Shang-Ta Yang, Dong Chen, and Shang-Hong Lai. Unified representation learning for cross model compatibility. In BMVC, 2020. +[39] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. NeurIPS, 32, 2019. +[40] Kai Wang, Luis Herranz, and Joost van de Weijer. Continual learning in cross-modal retrieval. In CVPR, 2021. +[41] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, 2020. +[42] Sijie Yan, Yuanjun Xiong, Kaustav Kundu, Shuo Yang, Siqi Deng, Meng Wang, Wei Xia, and Stefano Soatto. Positive-congruent training: Towards regression-free model updates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14299-14308, 2021. +[43] Binjie Zhang, Yixiao Ge, Yantao Shen, Yu Li, Chun Yuan, Xuyuan Xu, Yexin Wang, and Ying Shan. Hot-refresh model upgrades with regression-alleviating compatible training in image retrieval. In ICLR, 2022. +[44] Binjie Zhang, Yixiao Ge, Yantao Shen, Shupeng Su, Chun Yuan, Xuyuan Xu, Yexin Wang, and Ying Shan. Towards universal backward-compatible representation learning. In IJCAI, 2022. +[45] Liangli Zhen, Peng Hu, Xi Peng, Rick Siow Mong Goh, and Joey Tianyi Zhou. Deep multimodal transfer learning for cross-modal retrieval. TNNLS, 2020. +[46] Wengang Zhou, Houqiang Li, and Qi Tian. Recent advance in content-based image retrieval: A literature survey. arXiv preprint arXiv:1706.06064, 2017. +[47] Yifei Zhou, Zilu Li, Abhinav Shrivastava, Hengshuang Zhao, Antonio Torralba, Taipeng Tian, and Ser-Nam Lim. bt2: Backward-compatible training with basis transformation. arXiv preprint arXiv:2211.03989, 2022. +[48] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A + +comprehensive survey on transfer learning. Proceedings of the IEEE, 2020. \ No newline at end of file diff --git a/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/images.zip b/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6c08583f2a87dd7edd1ffa0600a3cae8bd658916 --- /dev/null +++ b/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f34fbe2ce51a4d44ddbfa802d2e10b10f5681a97c0a5f3bbb234bab071db318 +size 581214 diff --git a/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/layout.json b/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3a01589f8084616c63f6f633b1dab1843e312c14 --- /dev/null +++ b/2025/Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models/layout.json @@ -0,0 +1,9400 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 492, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 492, + 140 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 492, + 140 + ], + "type": "text", + "content": "Towards Cross-modal Backward-compatible Representation Learning for Vision-Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 164, + 161, + 261, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 161, + 261, + 175 + ], + "spans": [ + { + "bbox": [ + 164, + 161, + 261, + 175 + ], + "type": "text", + "content": "Young Kyun Jang1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 175, + 261, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 175, + 261, + 190 + ], + "spans": [ + { + "bbox": [ + 164, + 175, + 261, + 190 + ], + "type": "text", + "content": "Google DeepMind" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 192, + 263, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 192, + 263, + 202 + ], + "spans": [ + { + "bbox": [ + 164, + 192, + 263, + 202 + ], + "type": "text", + "content": "kyun0914@gmail.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 339, + 161, + 410, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 161, + 410, + 175 + ], + "spans": [ + { + "bbox": [ + 339, + 161, + 410, + 175 + ], + "type": "text", + "content": "Ser-nam " + }, + { + "bbox": [ + 339, + 161, + 410, + 175 + ], + "type": "inline_equation", + "content": "\\mathrm{Lim}^2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 301, + 175, + 447, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 301, + 175, + 447, + 190 + ], + "spans": [ + { + "bbox": [ + 301, + 175, + 447, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 301, + 175, + 447, + 190 + ], + "type": "text", + "content": "University of Central Florida" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 329, + 192, + 417, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 192, + 417, + 202 + ], + "spans": [ + { + "bbox": [ + 329, + 192, + 417, + 202 + ], + "type": "text", + "content": "sernam@gmail.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 152, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 258, + 297, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 297, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 297, + 558 + ], + "type": "text", + "content": "Modern retrieval systems often struggle with upgrading to new and more powerful models due to the incompatibility of embeddings between the old and new models. This necessitates a costly process known as backfilling, which involves re-computing the embeddings for a large number of data samples. In vision, Backward-compatible Training (BT) has been proposed to ensure that the new model aligns with the old model's embeddings. This paper extends the concept of vision-only BT to the field of cross-modal retrieval, marking the first attempt to address Cross-modal BT (XBT). Our goal is to achieve backward-compatibility between Vision-Language Pretraining (VLP) models, such as CLIP, for the cross-modal retrieval task. To address XBT challenges, we propose an efficient solution: a projection module that maps the new model's embeddings to those of the old model. This module, pretrained solely with text data, significantly reduces the number of image-text pairs required for XBT learning, and, once it is pretrained, it avoids using the old model during training. Furthermore, we utilize parameter-efficient training strategies that improve efficiency and preserve the off-the-shelf new model's knowledge by avoiding any modifications. Experimental results on cross-modal retrieval datasets demonstrate the effectiveness of XBT and its potential to enable backfill-free upgrades when a new VLP model emerges." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "spans": [ + { + "bbox": [ + 56, + 584, + 135, + 596 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "As the volume and variety of data grow exponentially in our multimedia-rich era, developing and maintaining efficient multi-modal retrieval systems becomes increasingly challenging. These systems, which provide data samples relevant to a user's query, must handle diverse data types, from text and images to audio and video. This growth puts a premium on the scalability and performance of these systems, necessitating continuous advancements in algorithms and technology. Embedding-based deep models for" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 247, + 552, + 358 + ], + "blocks": [ + { + "bbox": [ + 363, + 229, + 504, + 240 + ], + "lines": [ + { + "bbox": [ + 363, + 229, + 504, + 240 + ], + "spans": [ + { + "bbox": [ + 363, + 229, + 504, + 240 + ], + "type": "text", + "content": "Backward-compatible Training (BT)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 247, + 552, + 358 + ], + "lines": [ + { + "bbox": [ + 318, + 247, + 552, + 358 + ], + "spans": [ + { + "bbox": [ + 318, + 247, + 552, + 358 + ], + "type": "image", + "image_path": "e0e2ad6479550047c00477deca97bf74e0efc4a26999cc0a8d59d5b9f88c00db.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 316, + 382, + 553, + 504 + ], + "blocks": [ + { + "bbox": [ + 334, + 365, + 532, + 376 + ], + "lines": [ + { + "bbox": [ + 334, + 365, + 532, + 376 + ], + "spans": [ + { + "bbox": [ + 334, + 365, + 532, + 376 + ], + "type": "text", + "content": "Cross-modal Backward-compatible Training (XBT)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 382, + 553, + 504 + ], + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 504 + ], + "type": "image", + "image_path": "014ff1aa669b621ec3ffc021639fc59f131d860af096d99852d58d8164515017.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 511, + 555, + 589 + ], + "lines": [ + { + "bbox": [ + 313, + 511, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 555, + 589 + ], + "type": "text", + "content": "Figure 1. A conceptual visualization of Backward-compatible Training (BT, left) and its extension, the Cross(X)-modal version (XBT, right). Circles and squares denote data samples of images and text, respectively. XBT uses Vision-Language Pretraining (VLP) models as baselines to achieve cross-modal, backward-compatible representation learning, allowing the new, improved model to be compatible with the fixed old model." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 615, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 555, + 674 + ], + "type": "text", + "content": "retrieval have emerged as a key solution, transforming high-dimensional data into a lower-dimensional dense embedding space [11, 15, 16, 30, 37, 46]. These models capture the semantic meanings of data samples, enabling the quantification of similarities for retrieval." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "However, it is important to note that the embedding spaces generated by different deep models are not inherently compatible with each other. This incompatibility restricts the" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1783" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "reuse of an existing gallery that has been embedded with an older model when a new, better-performing model is introduced. Consequently, this necessitates \"backfilling,\" a process where the entire gallery must be rebuilt using the embeddings from the new model. Such a requirement is resource-intensive and time-consuming, posing a significant challenge when building retrieval systems." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 294, + 277 + ], + "type": "text", + "content": "Backward-compatible Training (BT) [13, 34, 38, 43, 44] has been developed to tackle this issue, specifically focusing on image retrieval systems. The main objective of BT is to train a new model from scratch in a manner that ensures its compatibility with an old model that was used to create the existing gallery. A successful BT model must consequently demonstrate that retrieving from the old gallery using a query embedded with the new model outperforms that using an embedding from the old model. This enhancement is crucial for justifying the application of BT." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 278, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 295, + 373 + ], + "type": "text", + "content": "Taking the BT problem a step further, we propose a new challenging task called Cross(X)-modal Backward-compatible Training (XBT) as shown in Figure 1. Our focus here is on applying BT principles within the realm of cross-modal retrieval, specifically exploring the interaction and compatibility between image and text embeddings of different Vision-Language Pretraining (VLP) models [17, 21, 22, 28, 35]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 542 + ], + "type": "text", + "content": "To achieve XBT, just like the BT problem, we need to resolve the incompatibility between the embeddings of the old, inferior model and the new, improved model. If we follow most prior practices of training from scratch with an additional compatibility loss term, a substantial number of supervised image-text pairs would be needed. The ideal quantity would be in the hundreds of millions, approximating the amount utilized in the establishment of VLP models. However, accessing the original training samples is often impossible, and training on that scale is prohibitively expensive and impractical. We propose an efficient solution instead: a text-only pretrained projection module, " + }, + { + "bbox": [ + 55, + 376, + 295, + 542 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 376, + 295, + 542 + ], + "type": "text", + "content": ", to align a given pretrained new model's embeddings with those of the old model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "content": "Our focus is on using only text data to estimate the entire distribution of embeddings in the VLP space. Specifically, we train " + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "content": " to learn the similarity between old and new text embeddings in a contrastive manner. By increasing the number of text samples, which is simpler than preparing image-text pairs, " + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "content": " can approximate the oracle projection between the intra-modal distributions of both old and new embeddings for texts. Assuming that the intra-modal distribution of texts in the VLP embedding space can mirror that of images, we can simply apply the same " + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 544, + 295, + 687 + ], + "type": "text", + "content": " to the new image embeddings to synthesize the corresponding old image embeddings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "With these generated synthetic old image and text embeddings, which can be considered as aligned with the new" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "embeddings, we aim to fine-tune the new model. In this process, far fewer supervised image-text pairs are required than in the original training dataset of the VLP. This also allows us to avoid using old model parameters during training, thereby enhancing training efficiency. In addition, we incorporate parameter-efficient fine-tuning schemes [12, 19, 20] into the new model, which add small trainable parameters and do not harm any of the new model's original parameters during training. This not only accelerates the new model's fine-tuning process but also allows for an easy reversion to the original new-to-new retrieval by simply removing the additional parameters." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 218, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 553, + 277 + ], + "type": "text", + "content": "We demonstrate our approach on various cross-modal benchmarks, highlighting the effectiveness of XBT in cross-modal retrieval protocols. In response to the rapid advancement of VLP models, XBT offers an environmentally friendly solution by enabling backfill-free systems." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 279, + 492, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 279, + 492, + 289 + ], + "spans": [ + { + "bbox": [ + 326, + 279, + 492, + 289 + ], + "type": "text", + "content": "Our contributions can be summarized as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 293, + 553, + 423 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 293, + 553, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 293, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 314, + 293, + 553, + 339 + ], + "type": "text", + "content": "- The Cross-modal Backward-compatible Training (XBT) concept is introduced for the first time to solve the backfilling problem that stems from real-world cross-modal retrieval systems." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 341, + 553, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 341, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 314, + 341, + 553, + 388 + ], + "type": "text", + "content": "- A novel XBT solution is proposed, which uses a text-only pretrained projection module, " + }, + { + "bbox": [ + 314, + 341, + 553, + 388 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 314, + 341, + 553, + 388 + ], + "type": "text", + "content": ", to efficiently align the embeddings of new and old models using only text samples." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 388, + 553, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 388, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 314, + 388, + 553, + 423 + ], + "type": "text", + "content": "- The proposal is demonstrated on various datasets and protocols, showcasing XBT's effectiveness in building backfill-free cross-modal retrieval systems." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 441, + 405, + 452 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 405, + 452 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 405, + 452 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 462, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 553, + 713 + ], + "type": "text", + "content": "Backward-compatible Training. The concept of Backward-compatible Training (BT) was first introduced in the study ([34]). This approach influences a new model with the learned classifier of the old model, thereby achieving backward compatibility between the old and new models. However, BT can degrade the original representational performance of the new model. To overcome this, [26] proposed aligning class-wise centers presented by the old and new models. Another approach to achieve backward compatibility is to use an additional projection to map the old embedding into the new embedding space by adding a lightweight transformation, as suggested in [13, 38]. The work of [29] further adds an auxiliary feature in preparation for future updates, while [47] uses additional dimensionality in the embedding to obtain compatibility. An online strategy that backfills the gallery on the fly is explored in [43], and [42] addresses the model regression problem. Despite this progress, the scenario considering cross-modal retrieval between image and text, which has many real-world applications, remains unexplored. Our study on XBT in this paper aims to fill this gap." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1784" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 298, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 298, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 298, + 288 + ], + "type": "text", + "content": "Vision-Language Continual and Transfer Learning The fields of continual learning [2, 5, 7] and transfer learning [25, 48] bear similarities to backward-compatible representation learning, as all aim to update an existing model to boost performance. In the realm of multi-modal representation learning for cross-modal retrieval, continual learning approaches like [40] propose methods to prevent catastrophic forgetting across different modalities. Transfer learning approaches like [45] suggest strategies for transferring knowledge from previously labeled categories (source modality) to new, unlabeled categories (target modality). However, our proposed XBT stands apart from these methods as it specifically tackles the challenge of maintaining backward compatibility between old and new models. This unique attribute makes XBT ideally suited for backfill-free retrieval scenarios, where the embeddings of the old model remain unchanged, yet we can still leverage the enhanced performance of the new model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 298, + 111, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 111, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 111, + 310 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "text", + "content": "Our goal is to construct a backfill-free, embedding-based, cross-modal retrieval system using VLP models, which are configured with an image encoder " + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "inline_equation", + "content": "E^{I}" + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "inline_equation", + "content": "E^{T}" + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "text", + "content": ". When a new better performing VLP model " + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\{E_{new}^{I}, E_{new}^{T}\\}" + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "text", + "content": " emerges, we aim to ensure its compatibility with the old model " + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\{E_{old}^{I}, E_{old}^{T}\\}" + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "text", + "content": " that was used to construct the gallery. To achieve this, we introduce Cross(X)-modal Backward-compatible Training (XBT)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "text", + "content": "Given an image " + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "text", + "content": " and text caption " + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "text", + "content": ", XBT enables retrieval between a new image embedding " + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "inline_equation", + "content": "v_{new} = E_{new}^{I}(x)" + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "text", + "content": " and the text embeddings " + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "inline_equation", + "content": "(w_{old})" + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "text", + "content": ", as well as between a new text embedding " + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "inline_equation", + "content": "w_{new} = E_{new}^{T}(t)" + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "text", + "content": " and the image embeddings " + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "inline_equation", + "content": "(v_{old})" + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "text", + "content": ". We denote backward compatible embeddings as " + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\bar{v}" + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 55, + 414, + 296, + 498 + ], + "type": "text", + "content": ". All embeddings we utilize in this work are l2-normalized." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 295, + 529 + ], + "type": "text", + "content": "3.1. Criterion for Cross-modal Backward Compatibility" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 571 + ], + "type": "text", + "content": "Following the definition of backward compatibility in BT work [34], we can construct strict constraints that ensure cross-modal backward compatibility as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 84, + 588, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 588, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 84, + 588, + 295, + 647 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d \\left(w _ {n e w _ {i}}, v _ {o l d _ {j}}\\right) \\leq d \\left(w _ {o l d _ {i}}, v _ {o l d _ {j}}\\right), \\forall y _ {i} = y _ {j}, \\\\ d \\left(w _ {\\text {n e w} _ {i}}, v _ {\\text {o l d} _ {j}}\\right) \\geq d \\left(w _ {\\text {o l d} _ {i}}, v _ {\\text {o l d} _ {j}}\\right), \\forall y _ {i} \\neq y _ {j}, \\tag {1} \\\\ d (v _ {n e w _ {i}}, w _ {o l d _ {j}}) \\leq d (v _ {o l d _ {i}}, w _ {o l d _ {j}}), \\forall y _ {i} = y _ {j}, \\\\ d (v _ {n e w _ {i}}, w _ {o l d _ {j}}) \\geq d (v _ {o l d _ {i}}, w _ {o l d _ {j}}), \\forall y _ {i} \\neq y _ {j}, \\\\ \\end{array}", + "image_path": "fb409886b34f57a2b2286c6c2326b9d0d733ec776b74265bcf947ac1757ed06b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "y_{j}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " represent whether an image and text are paired (" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "y_{i} = y_{j}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ") or not (" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "y_{i} \\neq y_{j}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "). The term " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " represents a distance metric in the embedding space, and we choose cosine distance as the baseline. The constraints in Eqn. 1 formally express the requirement that the new embedding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "must perform at least as well as the old embedding in terms of correctly matching image-text pairs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 97, + 556, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 556, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 556, + 156 + ], + "type": "text", + "content": "However, exhaustively satisfying these constraints is intractable due to the potential for the old embeddings to outperform the new embeddings in certain retrieval cases. As a result, we relax the criteria by defining an alternative evaluation metric as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 178, + 555, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 178, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 339, + 178, + 555, + 206 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\begin{array}{l} \\mathcal {M} (w _ {\\text {n e w}}, v _ {\\text {o l d}}; \\mathrm {Q}, \\mathrm {G}) > \\mathcal {M} (w _ {\\text {o l d}}, v _ {\\text {o l d}}; \\mathrm {Q}, \\mathrm {G}), \\\\ \\mathcal {A} ^ {\\prime} (\\mathrm {Q}, \\mathrm {G}) = \\mathcal {A} ^ {\\prime} (\\mathrm {Q}, \\mathrm {G}) \\end{array} \\tag {2} \\\\ \\mathcal {M} (v _ {n e w}, w _ {o l d}; \\mathrm {Q}, \\mathrm {G}) > \\mathcal {M} (v _ {o l d}, w _ {o l d}; \\mathrm {Q}, \\mathrm {G}), \\\\ \\end{array}", + "image_path": "df29a2b17b5ba8a1f5d641c507ffb2d6b5a82a69d691c703c290a6f87f2eb9de.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "text", + "content": "where a metric, such as recall, " + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(q,g;\\mathrm{Q},\\mathrm{G})" + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "text", + "content": " takes query embedding " + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "text", + "content": " and gallery embedding " + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "text", + "content": " over query set Q and gallery set G. This criterion suggests that the general performance is enhanced when the query embedding from the XBT-trained new model is used to perform retrieval with the gallery of the old model, compared to the performance of the old model alone. In essence, fulfilling Eqn. 2 indicates that the new model has achieved backward compatibility and can feasibly update without backfilling gallery." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 326, + 556, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 556, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 556, + 445 + ], + "type": "text", + "content": "Additionally, it is important to highlight that VLP models function as zero-shot learners. This leads us to define the XBT problem differently from classic BT, which trains separate models for specific image domain retrieval tasks such as ImageNet [31] or VGGFace2 [4]. In contrast, XBT tackles a more challenging task, aiming to bridge a new VLP model with a frozen old VLP model while retaining the new model's original zero shot capability. In this paper, we therefore assess the performance of XBT using retrieval and classification benchmarks in a zero-shot manner." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 455, + 440, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 440, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 440, + 468 + ], + "type": "text", + "content": "3.2. Text-only Pretraining" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 473, + 556, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 556, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 556, + 592 + ], + "type": "text", + "content": "The vast and diverse image-text pairs used to build VLP models significantly enhance their ability to connect semantically similar visual content and natural language [6]. However, this creates a complex embedding distribution that is difficult to predict and understand, thereby complicating the XBT process. A straightforward solution is to use a large number of supervised image-text pairs, similar to the approach used when building VLP models from scratch, to estimate the entire distribution of image and text embeddings during XBT." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "However, acquiring a sufficient number of accurate supervised pairs is extremely costly. To be far more efficient, we employ a small sized projection module, " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": ", and train it exclusively with text samples, as shown in Figure 2. We hypothesize that the distribution of text embeddings in VLP models, which is determined by their semantic similarity, is similarly mirrored in the distribution of their corresponding matched images. With this in mind, we aim to train " + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": " to cover the broad spectrum of the text embedding distribution between the new and old VLP models, making embeddings" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1785" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 83, + 296, + 178 + ], + "blocks": [ + { + "bbox": [ + 134, + 69, + 216, + 80 + ], + "lines": [ + { + "bbox": [ + 134, + 69, + 216, + 80 + ], + "spans": [ + { + "bbox": [ + 134, + 69, + 216, + 80 + ], + "type": "text", + "content": "Text-only Pretraining" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 83, + 296, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 83, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 83, + 296, + 178 + ], + "type": "image", + "image_path": "39336d708d8dd5f82225376033f56341510ecb6f23cff56980107b9b142ccf28.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 180, + 296, + 319 + ], + "blocks": [ + { + "bbox": [ + 57, + 180, + 296, + 319 + ], + "lines": [ + { + "bbox": [ + 57, + 180, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 57, + 180, + 296, + 319 + ], + "type": "image", + "image_path": "5540558dc3f4240ed4f4e8c844614f06cb25c9e88fc1897a1a3ed78cf7071704.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 327, + 296, + 404 + ], + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 404 + ], + "type": "text", + "content": "Figure 2. An illustration of the text-only pretraining of " + }, + { + "bbox": [ + 55, + 327, + 296, + 404 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 327, + 296, + 404 + ], + "type": "text", + "content": " (above), and XBT with " + }, + { + "bbox": [ + 55, + 327, + 296, + 404 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 327, + 296, + 404 + ], + "type": "text", + "content": " (below). Using only text samples, " + }, + { + "bbox": [ + 55, + 327, + 296, + 404 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 327, + 296, + 404 + ], + "type": "text", + "content": " is trained to approximate distribution of old text embeddings from that of new ones. After pretraining, the same " + }, + { + "bbox": [ + 55, + 327, + 296, + 404 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 327, + 296, + 404 + ], + "type": "text", + "content": " is used to generate both of synthetic old image and text embeddings from the new VLP embeddings and train to learn cross-modal backward-compatible representation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 428, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 295, + 453 + ], + "type": "text", + "content": "from the same text sample similar and others dissimilar in a contrastive way:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 478, + 296, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 478, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 82, + 478, + 296, + 491 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p r e} = \\mathbb {E} _ {t \\sim D _ {T}} \\left[ \\mathcal {L} _ {c} \\left(\\phi \\left(w _ {\\text {n e w}} + \\epsilon\\right), w _ {\\text {o l d}}; \\tau_ {p r e}\\right) \\right], \\tag {3}", + "image_path": "27017db36e3ca7bcbb0337fec7def9a534affcf81f8e545a781758260e5397c6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 505, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 296, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 505, + 296, + 555 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 505, + 296, + 555 + ], + "type": "text", + "content": " denotes text data sampled from text corpus " + }, + { + "bbox": [ + 55, + 505, + 296, + 555 + ], + "type": "inline_equation", + "content": "D_{T} = \\{t_{i}\\}_{i=1}^{N_{D_{T}}} \\text{ of } N_{D_{T}} \\text{ text samples, and } \\mathcal{L}_{c}(\\cdot, \\cdot) \\text{ is a standard batch-wise contrastive loss as defined in CLIP [28] of temperature } \\tau_{pre}." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "text", + "content": "To prevent overfitting towards the text domain, which could widen the discrepancy between the image and text embedding representations, we inject noise " + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "text", + "content": " sampled from a Gaussian distribution " + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "inline_equation", + "content": "(\\epsilon \\sim N(0,\\sigma^2))" + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "text", + "content": ", inspired by the approaches [8, 27]. Notably, " + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "text", + "content": " is designed to take the new VLP model's embeddings, which means that " + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "inline_equation", + "content": "w_{new} + \\epsilon" + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "text", + "content": " is also " + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "inline_equation", + "content": "l2" + }, + { + "bbox": [ + 55, + 556, + 296, + 640 + ], + "type": "text", + "content": "-normalized." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "During this stage, only " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " is updated, while all encoders remain fixed. By expanding the scale of " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": ", we can effectively replicate the complete text embedding spaces of both new and old VLP models. Additionally, " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " can be applied to new image embeddings to generate synthetic old image embeddings that closely approximate the actual old ones." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 79, + 553, + 119 + ], + "blocks": [ + { + "bbox": [ + 396, + 69, + 469, + 78 + ], + "lines": [ + { + "bbox": [ + 396, + 69, + 469, + 78 + ], + "spans": [ + { + "bbox": [ + 396, + 69, + 469, + 78 + ], + "type": "text", + "content": "XBT Learning Process" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 79, + 553, + 119 + ], + "lines": [ + { + "bbox": [ + 316, + 79, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 316, + 79, + 553, + 119 + ], + "type": "image", + "image_path": "b9379e2bd6f070cf88c957d8a78f1f13e57f99c4976019a8179f7966262acf8b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 129, + 555, + 163 + ], + "lines": [ + { + "bbox": [ + 313, + 129, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 555, + 163 + ], + "type": "text", + "content": "Figure 3. Our proposed learning process to achieve XBT. Notably, the old VLP model's encoders are not required in this stage, enhancing efficiency in training." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 182, + 546, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 546, + 194 + ], + "type": "text", + "content": "3.3. Cross-modal Backward-compatible Training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "text", + "content": "Training Loss. The objective of XBT is to ensure cross-modal compatibility, specifically between " + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "inline_equation", + "content": "w_{new}" + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "inline_equation", + "content": "v_{old}" + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "text", + "content": ", as well as between " + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "inline_equation", + "content": "v_{new}" + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "inline_equation", + "content": "w_{old}" + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "text", + "content": ". For a given dataset " + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "inline_equation", + "content": "D = \\{x_i, t_i\\}_{i=1}^{N_D}" + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "inline_equation", + "content": "N_D" + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "text", + "content": " supervised image-text pairs, we aim to train the new VLP model encoders, " + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "inline_equation", + "content": "E_{new}^I" + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "inline_equation", + "content": "E_{new}^T" + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "text", + "content": ". Note that in our base setting, the text corpus is significantly larger than the supervised dataset, i.e., " + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "inline_equation", + "content": "N_{D_T} \\gg N_D" + }, + { + "bbox": [ + 313, + 199, + 555, + 283 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 283, + 556, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 556, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 556, + 366 + ], + "type": "text", + "content": "However, the dimensionality of new and old VLP embeddings may differ, and even if they are the same, they are not be directly compatible. To address this, we apply the pretrained " + }, + { + "bbox": [ + 313, + 283, + 556, + 366 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 283, + 556, + 366 + ], + "type": "text", + "content": " to ensure that the new embeddings match the dimension of the old ones and project into the compatible embedding space, as shown in Figure 2. This can be represented as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 376, + 555, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 376, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 359, + 376, + 555, + 389 + ], + "type": "interline_equation", + "content": "\\phi (v _ {\\text {n e w}}) = \\bar {v} _ {\\text {n e w}}, \\phi (w _ {\\text {n e w}}) = \\bar {w} _ {\\text {n e w}}, \\tag {4}", + "image_path": "c9b9f3d5c0ff01928bca3562d042e836090ff9da1b213a95225e86adbf071470.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 397, + 555, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 397, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\bar{v}_{new}" + }, + { + "bbox": [ + 313, + 397, + 555, + 432 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 397, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\bar{w}_{new}" + }, + { + "bbox": [ + 313, + 397, + 555, + 432 + ], + "type": "text", + "content": " denote synthetic old embeddings and " + }, + { + "bbox": [ + 313, + 397, + 555, + 432 + ], + "type": "inline_equation", + "content": "\\{\\bar{v}_{new},\\bar{w}_{new},v_{old},w_{old}\\} \\subset \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 397, + 555, + 432 + ], + "type": "text", + "content": " , with " + }, + { + "bbox": [ + 313, + 397, + 555, + 432 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 397, + 555, + 432 + ], + "type": "text", + "content": " denoting the dimensionality of the old VLP model's embedding." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 433, + 490, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 433, + 490, + 444 + ], + "spans": [ + { + "bbox": [ + 326, + 433, + 490, + 444 + ], + "type": "text", + "content": "We then apply XBT loss " + }, + { + "bbox": [ + 326, + 433, + 490, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_X" + }, + { + "bbox": [ + 326, + 433, + 490, + 444 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 358, + 453, + 555, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 453, + 555, + 467 + ], + "spans": [ + { + "bbox": [ + 358, + 453, + 555, + 467 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {X} = \\mathbb {E} _ {x, t \\sim D} \\left[ \\mathcal {L} _ {c} \\left(\\bar {v} _ {\\text {n e w}}, \\bar {w} _ {\\text {n e w}}; \\tau_ {X}\\right) \\right]. \\tag {5}", + "image_path": "92e4cf79723d0145f59ec76d868011bd1d54318093e98f69efbb71fa5299cf43.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": " is the same contrastive loss used in Eqn. 3, with temperature " + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\tau_{X}" + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": ". As shown in Figure 3, the new VLP encoders, " + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "inline_equation", + "content": "E_{new}^{I}" + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "inline_equation", + "content": "E_{new}^{T}" + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": " are trained through " + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_X" + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": " in an end-to-end manner. Ultimately, " + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\bar{v}_{new}" + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\bar{w}_{new}" + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": " become cross-modal backward-compatible with existing old gallery embeddings, " + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "inline_equation", + "content": "v_{old}" + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "inline_equation", + "content": "w_{old}" + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": ", as all embeddings are distributed in the compatible space through " + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 474, + 555, + 557 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "type": "text", + "content": "Efficient Training. The efficiency of XBT can be attributed to two design choices that avoid: (1) any dependence on the old model when conducting XBT, and (2) updating the new VLP model off the shelf. We have already seen how " + }, + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "type": "text", + "content": " can effectively achieve (1). (2) is simply facilitated by applying small-sized additional parameters, namely Soft prompt [20] and LoRA[12]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": "To be more specific, we incorporate the concept of soft prompt tuning, as outlined in [18, 20]. Soft prompting is applied to the vision encoder " + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "inline_equation", + "content": "E_{new}^{I}" + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": " by adding trainable prompts as input, along with image patch tokens. We exclude soft prompt tuning on the text side, as we use the text-only pretrained module " + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 642, + 556, + 713 + ], + "type": "text", + "content": ", which requires maintaining the original" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1786" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "text", + "content": "distribution of text embeddings. We adopt the LoRA strategy [12] for the new VLP model's encoders " + }, + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "inline_equation", + "content": "E_{new}^{I}" + }, + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "inline_equation", + "content": "E_{new}^{T}" + }, + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "text", + "content": " to avoid modifying original parameters. These two factors, in conjunction with " + }, + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "text", + "content": ", offer an on-off solution: we can retrieve old samples using additional parameters, and we can easily revert to the original model by removing these parameters for new-to-new retrieval." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 294, + 216 + ], + "type": "text", + "content": "Further, regarding " + }, + { + "bbox": [ + 55, + 156, + 294, + 216 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 156, + 294, + 216 + ], + "type": "text", + "content": " during XBT (Sec. 3.3), we only finetune the parameters of layer normalization. This is done with the aim of preserving the knowledge learned during the text-only pretraining stage (Sec. 3.2) and facilitating easy adaptation towards the image-text joint representation space" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 216, + 294, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 216, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 294, + 346 + ], + "type": "text", + "content": "In the end, our XBT framework offers a parameter-efficient solution that enables rapid convergence with fewer training iterations (a single epoch is sufficient); maintains the new model's power; and requires far fewer supervised training samples than the scale of VLP model pretraining (approximately " + }, + { + "bbox": [ + 55, + 216, + 294, + 346 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 216, + 294, + 346 + ], + "type": "text", + "content": " of the level required to train CLIP [28] from scratch, which uses a 400M dataset, and " + }, + { + "bbox": [ + 55, + 216, + 294, + 346 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 55, + 216, + 294, + 346 + ], + "type": "text", + "content": " of the level required for LAION-2B based CLIP models [32], which utilize a 2B dataset). All these help to avoid any full-scale retraining of the new VLP models, paving the way for XBT to simply use new VLP models off-the-shelf as mentioned." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 135, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 135, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 135, + 371 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 377, + 294, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 294, + 472 + ], + "type": "text", + "content": "The evaluation of the proposed cross-modal backward compatibility includes image-text zero-shot retrieval. Pretrained VLP models are used as our baseline, with the goal to assess performance in a zero-shot environment by tuning a given pretrained new VLP model that is stronger than the old model. The results highlight the potential of XBT, showing improved performance when old model outputs are matched with the XBT-learned new model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 105, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 105, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 105, + 493 + ], + "type": "text", + "content": "4.1. Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "type": "text", + "content": "Model Training. Our XBT process involves two distinct training stages: text-only pretraining (See Sec. 3.2) and image-text supervised training (See Sec. 3.3). For the text-only pretraining stage, we utilize the text samples from a subset of the 115M filtered web-collected image-text paired dataset from BLIP [21], comprising around 67M available pairs (58.2% of the total). We construct a text corpus " + }, + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "type": "text", + "content": " in Eqn. 3 using these text samples. For the subsequent image-text supervised training stage, we use a smaller subset of 4M image-text pairs from the same dataset to construct a supervised dataset " + }, + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 498, + 295, + 652 + ], + "type": "text", + "content": " in Eqn. 5. It is important to note that these subsets are significantly smaller than the dataset scale (400M, 2B or more) used to build CLIP models [28, 32]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 712 + ], + "type": "text", + "content": "We utilize 8-A100-80GB GPUs to train and evaluate the models. For the text-only pretraining stage in Sec. 3.2, the batch size is set to 8,192 (1,024 batch per GPU) and during this stage, the entire set of trainable parameters of " + }, + { + "bbox": [ + 55, + 653, + 295, + 712 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 653, + 295, + 712 + ], + "type": "text", + "content": " are trained while both VLP text encoders are fixed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "text", + "content": "Moving on to the image-text supervised training in Sec. 3.3, the batch size is reduced to 1,024 (128 batch per GPU), and in this stage, layer normalization is set to be the only trainable component for all VLP image and text encoders with " + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "text", + "content": ". Temperature hyper-parameters " + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\tau_{pre}" + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\tau_{X}" + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\tau_{N}" + }, + { + "bbox": [ + 313, + 72, + 555, + 323 + ], + "type": "text", + "content": " are fixed at 2.6592. We employ the AdamW optimizer [24] with a fixed learning rate of 1e-4 and a weight decay of 0.01. For soft prompts, we use 10 prompts and apply 100 times larger learning rate, 1e-2. The training iteration is determined by the dataset size, and the entire pipeline is trained for a single epoch. For image augmentation, we begin with a random resized crop, adjusting the image scale between 0.5 and 1.0. Additionally, we apply a random horizontal flip and make random adjustments to the image's contrast, brightness, and sharpness. To incorporate different perspectives and angles, we modify the image's translation and rotation. The pretrained weights provided by HuggingFace [41] are applied to baseline VLP models. We employ as: openai/clip-vit-base-patch32, openai/clip-vit-target-patch14, laion/CLIP-ViT-H-14-laion2B-s32B-b79K." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 435 + ], + "type": "text", + "content": "Model Evaluation. We validate the effectiveness of XBT with a benchmark comprising three popular image-text paired datasets for cross-modal retrieval evaluation. The first is the nocaps dataset [1], which offers a diverse distribution of object names and facilitates a more detailed analysis. We employ the validation split of this dataset, which consists of images each paired with 10 relevant textual captions, totaling 4,500 images and 45,000 captions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 544 + ], + "type": "text", + "content": "To evaluate XBT at a larger scale, we also include the Flickr [14] and COCO [23] datasets, which consist of images that are each paired with 5 relevant textual captions. For the Flickr dataset, we use the entire dataset, encompassing 31,783 images and 158,915 captions. For the COCO dataset, we use the validation split, which includes 35,136 images and 175,680 captions. We note the two datasets as Large-Flickr and Large-COCO respectively for the remainder of the paper." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "type": "text", + "content": "For retrieval, we utilize all the samples in each dataset for both query and gallery, following the cross-modal benchmark used in [21, 28]. For evaluation purpose, we adopt recall scores at top K retrieval results " + }, + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}@\\mathbb{K},\\%)" + }, + { + "bbox": [ + 313, + 544, + 555, + 604 + ], + "type": "text", + "content": " to estimate the cross-modal backward compatibility (See Eqn. 2)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 622, + 556, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 556, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 556, + 693 + ], + "type": "text", + "content": "Implementation Details. In our work, we primarily use the popular CLIP models [28], based on a Transformer [36] backbone, as our baseline VLP models. The models, listed in ascending order of scale and performance, are: CLIP-ViT-B32, CLIP-ViT-L14, and CLIP-ViT-H14. For simplicity, we will refer to them as B32, L14, and H14, respec" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 428, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 428, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 428, + 712 + ], + "type": "text", + "content": "1https://huggingface.co/models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1787" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 92, + 102, + 517, + 348 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 1. Cross-modal retrieval results on nocaps. Note that for comparison, the B32/L14 case should be compared with B32, and the L14/H14 case should be compared with L14, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 92, + 102, + 517, + 348 + ], + "lines": [ + { + "bbox": [ + 92, + 102, + 517, + 348 + ], + "spans": [ + { + "bbox": [ + 92, + 102, + 517, + 348 + ], + "type": "table", + "html": "
Old Model / New ModelMethodText Query(w, w)/Image Gallery(v)Image Query(v, v)/Text Gallery(w)
CaseR@1R@5R@10R@50CaseR@1R@5R@10R@50
Original
CLIP-ViT-B32-w/v45.1474.9884.5196.00v/w71.3892.0296.3399.67
CLIP-ViT-L14-w/v47.7776.5085.1395.95v/w73.2993.2497.4099.78
Cross-modal Backward Compatible Training
CLIP-ViT-B32 / CLIP-ViT-L14Full-tune\\( \\bar{w}_{new}/v_{old} \\)41.6073.5484.2696.45\\( \\bar{v}_{new}/w_{old} \\)63.7687.0093.5199.09
LoRA-only43.4074.6684.9496.7368.8290.8796.0099.76
Base43.4874.8785.0496.7469.7891.3196.0299.78
XBT48.0279.0088.2197.6675.0293.2797.3199.91
Full-tune\\( \\bar{w}_{new}/\\bar{v}_{new} \\)48.9279.3188.2697.64\\( \\bar{v}_{new}/\\bar{w}_{new} \\)61.2285.8992.6299.04
LoRA-only55.4483.8291.2498.3470.1191.1396.4999.73
Base56.1884.3191.4498.3771.0791.4496.2299.76
XBT63.4888.6193.9898.8677.2294.5697.9199.89
CLIP-ViT-L14 / CLIP-ViT-H14Full-tune\\( \\bar{w}_{new}/v_{old} \\)45.1877.8285.3497.65\\( \\bar{v}_{new}/w_{old} \\)66.3888.5593.5998.90
LoRA-only47.0076.9486.0297.6372.0492.4296.5899.67
Base47.0677.1586.1297.6473.0092.8696.6099.69
XBT51.1480.8288.9398.7276.6294.1197.5899.84
Full-tune\\( \\bar{w}_{new}/\\bar{v}_{new} \\)54.3982.2689.0698.95\\( \\bar{v}_{new}/\\bar{w}_{new} \\)65.5787.5293.5199.13
LoRA-only60.9186.7792.0498.6574.4692.7397.3899.82
Base61.6587.2692.2498.6875.4293.0497.1199.85
XBT66.4790.2194.9299.0580.0296.1198.87100.0
", + "image_path": "a46a815a5fbb8e08a6bc535f9d97094a66d948a48eca668339e185e4c67b03e9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 366, + 298, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 366, + 298, + 523 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 298, + 523 + ], + "type": "text", + "content": "tively. Notably, while B32 and L14 are trained with the same dataset at 400M scale, H14 is trained with a larger dataset at 2B scale. We apply additional LoRA [12] parameters for each new model encoder configured as follows: " + }, + { + "bbox": [ + 54, + 366, + 298, + 523 + ], + "type": "inline_equation", + "content": "\\mathrm{LoRA}_{\\alpha} = 16" + }, + { + "bbox": [ + 54, + 366, + 298, + 523 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 366, + 298, + 523 + ], + "type": "inline_equation", + "content": "\\mathrm{rank} = 16" + }, + { + "bbox": [ + 54, + 366, + 298, + 523 + ], + "type": "text", + "content": ", and dropout " + }, + { + "bbox": [ + 54, + 366, + 298, + 523 + ], + "type": "inline_equation", + "content": "= 0.1" + }, + { + "bbox": [ + 54, + 366, + 298, + 523 + ], + "type": "text", + "content": ". The proposed projection module " + }, + { + "bbox": [ + 54, + 366, + 298, + 523 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 54, + 366, + 298, + 523 + ], + "type": "text", + "content": " consists of three Linear layers with layer normalization (LN) [3] and the GELU nonlinearity function [9]. Detailed architecture is as follows: Linear-LN-GELU-Linear-LN-GELU-Linear. Dropout is not applied as it was found to degrade performance empirically. The intermediate hidden dimension of the Linear layer is set to be four times the dimensionality of the output embedding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 529, + 141, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 141, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 141, + 540 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": "Baseline Comparisons. To simplify notation, we use " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "w / v" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": " to denote the retrieval results obtained using " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": " as the query embeddings and " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": " as the gallery embeddings. We establish two protocols: assessing cross-modal backward compatibility with new/old (" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\bar{w}_{\\text{new}} / v_{\\text{old}}" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\bar{v}_{\\text{new}} / w_{\\text{old}}" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": "), and maintaining the new VLP's original performance with new/new (" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\bar{w}_{\\text{new}} / \\bar{v}_{\\text{new}}" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\bar{v}_{\\text{new}} / \\bar{w}_{\\text{new}}" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": "). All of the above are similarly applied to " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "v / w" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 700 + ], + "type": "text", + "content": "To the best of our knowledge, since this paper presents the first method that aims to solve the XBT problem, there are no direct previous works to compare to. Nevertheless, we compare our XBT against three baselines: Full-tune, LoRA-only, and Base, which we elaborate on below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 702, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 296, + 714 + ], + "type": "text", + "content": "Naive solution - Direct backward contrastive learning:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 366, + 556, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 556, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 556, + 452 + ], + "type": "text", + "content": "The pretrained " + }, + { + "bbox": [ + 313, + 366, + 556, + 452 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 366, + 556, + 452 + ], + "type": "text", + "content": " in Eqn. 3, which maps the new embeddings to the old via text based contrastive learning, constitutes a major contribution of XBT. We therefore would like to compare with a solution that does not utilize such a pretrained " + }, + { + "bbox": [ + 313, + 366, + 556, + 452 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 366, + 556, + 452 + ], + "type": "text", + "content": ". A straightforward solution for this is to fine-tune the new VLP encoders " + }, + { + "bbox": [ + 313, + 366, + 556, + 452 + ], + "type": "inline_equation", + "content": "(E_{new}^{I}" + }, + { + "bbox": [ + 313, + 366, + 556, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 366, + 556, + 452 + ], + "type": "inline_equation", + "content": "E_{new}^{T})" + }, + { + "bbox": [ + 313, + 366, + 556, + 452 + ], + "type": "text", + "content": " to be compatible with the old ones by minimizing the following loss:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 354, + 462, + 554, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 462, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 354, + 462, + 554, + 491 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {D i r e c t} = \\mathbb {E} _ {x, t \\sim D} [ \\mathcal {L} _ {c} (\\bar {v} _ {\\text {n e w}}, w _ {\\text {o l d}}; \\tau_ {N}) \\\\ \\left. + \\mathcal {L} _ {c} \\left(\\bar {w} _ {\\text {n e w}}, v _ {\\text {o l d}}; \\tau_ {N}\\right) \\right] \\tag {0} \\\\ \\end{array}", + "image_path": "59dbb2bff070d1426a64d54039d7d202eb25e381c813624295a50ac7c73d27ae.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "text", + "content": " is the same contrastive loss used in Eqn. 5, and " + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\tau_N" + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "text", + "content": " is a temperature. Here, we apply a randomly initialized " + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "text", + "content": " with the same configuration as XBT for projection, as per Eqn. 4, to generate cross-modal backward compatible embeddings, " + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\bar{w}" + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "inline_equation", + "content": "\\bar{v}" + }, + { + "bbox": [ + 313, + 497, + 555, + 556 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": "Different training options: Based on Eqn. 6, we consider multiple training setups: (1) full fine-tuning of all trainable components (Full-tune), (2) tuning LoRA parameters-only (LoRA-only), and (3) starting from LoRA-only, adding an extra learnable prompt (Base). Models produced by Base would therefore be akin to XBT without the pretrained " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": ". For these, we train the " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "E_{new}^{I}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "E_{new}^{T}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{Direct}" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": " with 4M image-text pairs. Note that, XBT only trains layer-normalization layers including pretrained " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": ", however, Full-tune, LoRA-only and Base setups are training entire parameters of " + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 558, + 556, + 714 + ], + "type": "text", + "content": " since it is not trained before. The results of these baselines are presented in Table 1, tested with nocaps dataset. The performance of the original VLP models is also reported" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1788" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 91, + 515, + 365 + ], + "blocks": [ + { + "bbox": [ + 171, + 71, + 438, + 82 + ], + "lines": [ + { + "bbox": [ + 171, + 71, + 438, + 82 + ], + "spans": [ + { + "bbox": [ + 171, + 71, + 438, + 82 + ], + "type": "text", + "content": "Table 2. Cross-modal retrieval results on Large-Flickr and Large-COCO." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 91, + 91, + 515, + 365 + ], + "lines": [ + { + "bbox": [ + 91, + 91, + 515, + 365 + ], + "spans": [ + { + "bbox": [ + 91, + 91, + 515, + 365 + ], + "type": "table", + "html": "
DatasetOld Model / New ModelMethodText Query(w, w)/Image Gallery(v)Image Query(v, v)/Text Gallery(w)
CaseR@1R@5R@10R@50CaseR@1R@5R@10R@50
Original
CLIP-ViT-B32-w/v21.5441.2450.7272.69v/w40.3564.4273.3989.91
Cross-modal Backward Compatible Training
CLIP-ViT-B32 / CLIP-ViT-L14Base\\( \\bar{w}_{\\text{new}}/v_{\\text{old}} \\)19.0237.2546.6369.32\\( \\bar{v}_{\\text{new}}/w_{\\text{old}} \\)36.7460.6170.2288.34
XBT22.3842.7152.4174.8042.4766.0474.8790.79
Base\\( \\bar{w}_{\\text{new}}/\\bar{v}_{\\text{new}} \\)32.2955.1864.6883.60\\( \\bar{v}_{\\text{new}}/\\bar{w}_{\\text{new}} \\)36.1059.8069.9589.01
XBT39.5962.8971.7588.1343.5068.5077.9993.21
CLIP-ViT-B32 / CLIP-ViT-H14Base\\( \\bar{w}_{\\text{new}}/v_{\\text{old}} \\)20.2739.1748.5970.67\\( \\bar{v}_{\\text{new}}/w_{\\text{old}} \\)35.5359.4469.0687.67
XBT23.7044.4754.1976.0341.0965.1874.3290.16
Base\\( \\bar{w}_{\\text{new}}/\\bar{v}_{\\text{new}} \\)32.9455.9465.2683.77\\( \\bar{v}_{\\text{new}}/\\bar{w}_{\\text{new}} \\)34.8159.5569.9688.81
XBT40.4263.9072.7488.4746.3571.1880.0194.08
Original
CLIP-ViT-B32-w/v14.4430.2038.9662.30v/w28.6250.1759.6780.22
Cross-modal Backward Compatible Training
CLIP-ViT-B32 / CLIP-ViT-L14Base\\( \\bar{w}_{\\text{new}}/v_{\\text{old}} \\)12.9728.1436.4859.78\\( \\bar{v}_{\\text{new}}/w_{\\text{old}} \\)26.3246.9156.5477.86
XBT15.5532.2741.3465.1130.7352.5862.3081.83
Base\\( \\bar{w}_{\\text{new}}/\\bar{v}_{\\text{new}} \\)22.1641.6550.9873.06\\( \\bar{v}_{\\text{new}}/\\bar{w}_{\\text{new}} \\)27.2648.1657.8079.57
XBT27.6148.6958.0878.8332.9255.3764.8984.67
CLIP-ViT-B32 / CLIP-ViT-H14Base\\( \\bar{w}_{\\text{new}}/v_{\\text{old}} \\)13.8829.5538.2361.57\\( \\bar{v}_{\\text{new}}/w_{\\text{old}} \\)26.1946.7556.3077.96
XBT15.8232.7941.8465.4530.3351.2859.9781.66
Base\\( \\bar{w}_{\\text{new}}/\\bar{v}_{\\text{new}} \\)23.2043.2252.7074.68\\( \\bar{v}_{\\text{new}}/\\bar{w}_{\\text{new}} \\)27.3148.5658.6080.09
XBT28.1449.4858.8879.3234.5457.2966.7985.92
", + "image_path": "8481aad00a18e3d66c1700cea5c38c6e32126392e3e6dd04b332633546f5a4bd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 374, + 284, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 284, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 284, + 386 + ], + "type": "text", + "content": "for a clear comparison. We highlight our XBT method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "type": "text", + "content": "In the context of cross-modal backward compatibility, where recall scores of " + }, + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "type": "inline_equation", + "content": "\\bar{w}_{\\text{new}} / v_{\\text{old}}" + }, + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "type": "inline_equation", + "content": "\\bar{v}_{\\text{new}} / w_{\\text{old}}" + }, + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "type": "text", + "content": " should be higher than those of " + }, + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "type": "inline_equation", + "content": "w / v" + }, + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "type": "inline_equation", + "content": "v / w" + }, + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "type": "text", + "content": " of old model, respectively, Full-tune significantly underperforms compared to the old VLP's. This indicates that full fine-tuning is not an appropriate solution. While both LoRA-only and Base improve performance over Full-tune, they still fall short of the old VLP's performance and fail to meet the criterion. In contrast, XBT significantly outperforms these baselines and even surpasses the old VLP's performance, satisfying the cross-modal backward compatibility for all recall metrics (see Eqn. 2)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "text", + "content": "In terms of maintaining new VLP's performance, where recall scores of " + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\bar{w}_{\\text{new}} / v_{\\text{new}}" + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\bar{v}_{\\text{new}} / w_{\\text{new}}" + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "text", + "content": " could be similar to those of " + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "inline_equation", + "content": "w / v" + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "inline_equation", + "content": "v / w" + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "text", + "content": " of new model, respectively, all baselines show decent performance in " + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\bar{w}_{\\text{new}} / v_{\\text{new}}" + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "text", + "content": " case but only XBT achieves better performance than new VLP model in " + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\bar{v}_{\\text{new}} / w_{\\text{new}}" + }, + { + "bbox": [ + 55, + 533, + 295, + 651 + ], + "type": "text", + "content": " case. These results support that XBT not only maintains the performance of the new VLP model but also enhances it in certain cases. The superior performance of XBT in both cases underscores its promise as a robust solution for cross-modal retrieval tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "Large Scale Retrieval. In this experiment, we assess the performance of cross-modal backward compatibility in more practical use cases. We use larger datasets, namely LargeFlickr and Large-COCO, as shown in Table 2. These datasets provide more than six times the number of samples typically" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 374, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 481 + ], + "type": "text", + "content": "used in image-text retrieval literature [6, 37], which make the retrieval process more challenging. We use the second-best performing Base method from Table 1 as a basis for comparison. To validate the applicability of XBT across various VLP models, we utilize three CLIP baselines: B32, L14, and H14. Additionally, we explore different combinations, such as B32/L14 and B32/H14, to further analyze the effectiveness of XBT, beyond the comparisons made in Table 1 (B32/L14 and L14/H32)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 482, + 556, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 556, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 556, + 566 + ], + "type": "text", + "content": "When comparing B32/L14 with B32/H14, it is observed that every retrieval cases of improved models (L14 and H14) successfully maintain cross-modal backward compatibility, regardless of their scale. Especially, retrieval performance of B32/H14 is better than that of B32/L14, which demonstrates XBT's ability to preserve and align original model's power with old gallery." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 567, + 556, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 556, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 556, + 662 + ], + "type": "text", + "content": "Upon examining the results, we consistently observe that XBT outperforms the Base model across all instances and both retrieval scenarios, often by a significant margin. Furthermore, XBT maintains the performance of new VLPs better than Base, achieving even larger margins for " + }, + { + "bbox": [ + 313, + 567, + 556, + 662 + ], + "type": "inline_equation", + "content": "\\bar{w}_{new} / \\bar{v}_{new}" + }, + { + "bbox": [ + 313, + 567, + 556, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 567, + 556, + 662 + ], + "type": "inline_equation", + "content": "\\bar{v}_{new} / w_{old}" + }, + { + "bbox": [ + 313, + 567, + 556, + 662 + ], + "type": "text", + "content": ". In summary, XBT consistently demonstrates robust and superior retrieval performance, even in large-scale scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 671, + 417, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 417, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 417, + 684 + ], + "type": "text", + "content": "4.3. Further Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Continual learning. In line with the literature on BT works [29, 34], we set up a continual learning scenario for a se" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1789" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 102, + 511, + 207 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 3. Continual learning scenario experimental results with XBT on nocaps. Note that for H14 training, we use L14, which has been previously adapted with B32, and H14 never encounters B14." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 102, + 511, + 207 + ], + "lines": [ + { + "bbox": [ + 100, + 102, + 511, + 207 + ], + "spans": [ + { + "bbox": [ + 100, + 102, + 511, + 207 + ], + "type": "table", + "html": "
Old Model / New ModelText Query(w, w)/Image Gallery(v)Image Query(v, v)/Text Gallery(w)
CaseR@1R@5R@10R@50CaseR@1R@5R@10R@50
CLIP-ViT-B32 / CLIP-ViT-L14wnew/vold47.6878.0088.0097.50vnew/wold73.8992.0097.2299.86
wnew/vnew62.2286.1893.0098.12vnew/wnew76.2792.9096.7699.11
CLIP-ViT-L14 / CLIP-ViT-H14wnew/vold50.9780.6888.9897.72vnew/wold77.3894.8297.8099.84
wnew/vnew66.0989.7894.7799.01vnew/wnew80.6096.3698.7199.96
CLIP-ViT-B32 / CLIP-ViT-H14wnew/vold49.1479.8888.5997.67vnew/wold74.8493.7397.2299.80
wnew/vnew65.9589.6894.6898.99vnew/wnew80.9696.0098.7199.96
", + "image_path": "9fe6b79b35257d7ed04243d1d2c9b6d39c3baba7c8e192101b91f0d1c115e6ce.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 60, + 216, + 112, + 266 + ], + "blocks": [ + { + "bbox": [ + 60, + 216, + 112, + 266 + ], + "lines": [ + { + "bbox": [ + 60, + 216, + 112, + 266 + ], + "spans": [ + { + "bbox": [ + 60, + 216, + 112, + 266 + ], + "type": "image", + "image_path": "86a116e654f354f096d1759859cedcff74f79fd43916b63a565fc65f1286ac48.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 275, + 485, + 287 + ], + "lines": [ + { + "bbox": [ + 123, + 275, + 485, + 287 + ], + "spans": [ + { + "bbox": [ + 123, + 275, + 485, + 287 + ], + "type": "text", + "content": "Figure 4. New query vs. Old gallery retrieval results on nocaps. B32 as old, and L14 as new model." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 215, + 243, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 215, + 243, + 223 + ], + "spans": [ + { + "bbox": [ + 161, + 215, + 243, + 223 + ], + "type": "text", + "content": "Retrieved from Old Text Gallery" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 224, + 297, + 258 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 120, + 224, + 252, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 224, + 252, + 232 + ], + "spans": [ + { + "bbox": [ + 120, + 224, + 252, + 232 + ], + "type": "text", + "content": "R#1 - A dog with colorful sunglasses on and a bandanna." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 120, + 232, + 267, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 232, + 267, + 239 + ], + "spans": [ + { + "bbox": [ + 120, + 232, + 267, + 239 + ], + "type": "text", + "content": "R#2 - A happy dog is wearing a blue bandanna and sunglasses." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 120, + 239, + 270, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 239, + 270, + 245 + ], + "spans": [ + { + "bbox": [ + 120, + 239, + 270, + 245 + ], + "type": "text", + "content": "R#3 - A dog wearing sunglasses and a blue and white bandanna." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 120, + 245, + 297, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 245, + 297, + 252 + ], + "spans": [ + { + "bbox": [ + 120, + 245, + 297, + 252 + ], + "type": "text", + "content": "R#4 - An old brown dog is wearing mirrored sunglasses and a blue bandana." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 120, + 252, + 276, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 252, + 276, + 258 + ], + "spans": [ + { + "bbox": [ + 120, + 252, + 276, + 258 + ], + "type": "text", + "content": "R#5 - A dog has its tongue out as it wears sunglasses and bandana." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 309, + 221, + 353, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 221, + 353, + 228 + ], + "spans": [ + { + "bbox": [ + 309, + 221, + 353, + 228 + ], + "type": "text", + "content": "New Query Text" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 309, + 221, + 353, + 258 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 309, + 228, + 352, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 228, + 352, + 237 + ], + "spans": [ + { + "bbox": [ + 309, + 228, + 352, + 237 + ], + "type": "text", + "content": "The cute brown" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 309, + 237, + 351, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 237, + 351, + 244 + ], + "spans": [ + { + "bbox": [ + 309, + 237, + 351, + 244 + ], + "type": "text", + "content": "tabby and white" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 309, + 244, + 351, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 244, + 351, + 251 + ], + "spans": [ + { + "bbox": [ + 309, + 244, + 351, + 251 + ], + "type": "text", + "content": "kitten, sits on the" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 309, + 251, + 351, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 251, + 351, + 258 + ], + "spans": [ + { + "bbox": [ + 309, + 251, + 351, + 258 + ], + "type": "text", + "content": "hard wood floor." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 406, + 215, + 493, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 215, + 493, + 223 + ], + "spans": [ + { + "bbox": [ + 406, + 215, + 493, + 223 + ], + "type": "text", + "content": "Retrieved from Old Image Gallery" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 364, + 223, + 551, + 259 + ], + "blocks": [ + { + "bbox": [ + 364, + 223, + 551, + 259 + ], + "lines": [ + { + "bbox": [ + 364, + 223, + 551, + 259 + ], + "spans": [ + { + "bbox": [ + 364, + 223, + 551, + 259 + ], + "type": "image", + "image_path": "776d7b9f9867f9f5e67996b4f82968b0904338f89af2208207e70a89a081c620.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 380, + 259, + 389, + 266 + ], + "lines": [ + { + "bbox": [ + 380, + 259, + 389, + 266 + ], + "spans": [ + { + "bbox": [ + 380, + 259, + 389, + 266 + ], + "type": "text", + "content": "R#1" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 77, + 327, + 272, + 469 + ], + "blocks": [ + { + "bbox": [ + 55, + 297, + 295, + 317 + ], + "lines": [ + { + "bbox": [ + 55, + 297, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 295, + 317 + ], + "type": "text", + "content": "Table 4. Ablation study tested on nocaps, with B32 as old and L14 as new model." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 327, + 272, + 469 + ], + "lines": [ + { + "bbox": [ + 77, + 327, + 272, + 469 + ], + "spans": [ + { + "bbox": [ + 77, + 327, + 272, + 469 + ], + "type": "table", + "html": "
Method\\(\\bar{w}(\\text{text})/v(\\text{image})\\)\\(\\bar{v}(\\text{image})/w(\\text{text})\\)
R@1R@10R@1R@10
(a) Baseline48.0279.0075.0293.27
(b) w.o noise47.2478.0573.3192.49
(c) \\(0.25 \\times D_T\\)46.8977.9173.2092.18
(d) \\(0.5 \\times D_T\\)47.6878.0073.8992.00
(e) \\(2 \\times D\\)47.6378.6974.9893.98
(f) \\(0.5 \\times D\\)46.8077.8971.6992.00
(g) \\(D=\\text{CC3M}\\)46.4477.7671.3391.84
(h) Image-only45.3276.9274.0292.35
", + "image_path": "2d99f85d593a7494acf5348176395bf299f88b7a2b1f39ee7a29717f560dd19a.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 54, + 495, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 495, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 54, + 495, + 295, + 649 + ], + "type": "text", + "content": "quence of model updates (old model B32, new model L14, and better new model H14) and present the retrieval results in Table 3. Initially, we apply XBT to L14 to ensure compatibility with B32. Subsequently, we apply XBT to H14 to ensure compatibility with the previously learned L32, which is already compatible with B32. For this process, we divide both the image-text pairs and the text-only pretraining train sets into two halves, using each separately for each case. A comparison with the original results reported in Table 3 confirms that our XBT performs well in the continual learning scenario either. It achieves cross-modal backward compatibility while leveraging the power of the improved new model." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 54, + 653, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 653, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 653, + 297, + 715 + ], + "type": "text", + "content": "Ablation Study. To validate XBT, we conduct further analysis as shown in Table 4. By comparing (a) and (b), we observe that introducing noise during " + }, + { + "bbox": [ + 54, + 653, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 54, + 653, + 297, + 715 + ], + "type": "text", + "content": " training aids in generalization. The comparison between (c) and (d) demonstrates that the scale of " + }, + { + "bbox": [ + 54, + 653, + 297, + 715 + ], + "type": "inline_equation", + "content": "D_{T}" + }, + { + "bbox": [ + 54, + 653, + 297, + 715 + ], + "type": "text", + "content": " is important, supporting our" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "type": "text", + "content": "assumption that a sufficient number of text samples can help build a robust " + }, + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "type": "text", + "content": ". (e) outperforms (a), confirming that utilizing more image-text pairs can enhance XBT. The lower performance of (f) also aligns with the notion that the number of image-text pairs is crucial. In (g), when we replace " + }, + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 298, + 555, + 441 + ], + "type": "text", + "content": " with CC3M [33], from which we can obtain around 2.4M, the performance is similar to (f), suggesting that XBT can be effectively applied with other datasets. In (h), we replace text-only pretraining with image-only pretraining of the same scale and observe that text-only pretraining performs better in both retrieval scenarios, demonstrating its efficiency and superiority." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 444, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 492 + ], + "type": "text", + "content": "Qualitative Results. We facilitate retrieval by utilizing new query embeddings and old gallery embeddings. The results in Figure 4 demonstrate accurate cross-modal backward-compatible retrieval." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 511, + 388, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 388, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 388, + 523 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we introduced Cross-modal Backward-compatible Training (XBT), a novel task for cross-modal retrieval that focuses on the compatibility between image and text embeddings of different Vision-Language Pretraining (VLP) models. We proposed an efficient solution using a text-only pretrained projection module, " + }, + { + "bbox": [ + 312, + 534, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 312, + 534, + 556, + 713 + ], + "type": "text", + "content": ", to align the new model's embeddings with those of the old model, thereby enhancing training efficiency. By integrating parameter-efficient training schemes into the XBT framework, we were able to accelerate the model's training while preserving the original VLP's zero-shot capabilities. Our approach, demonstrated on various cross-modal benchmarks, effectively builds cross-modal retrieval systems without backfilling, offering an efficient and environmentally friendly solution in response to the VLP improvements." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1790" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 137 + ], + "type": "text", + "content": "The first author performed the initial research and development of this work in collaboration with Ser-nam Lim. The finalization and refinement of this work were completed while the first author was affiliated with Google DeepMind." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 163, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 163, + 297, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 297, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 297, + 207 + ], + "type": "text", + "content": "[1] Harsh Agrawal, Karan Desai, Yufei Wang, Xinlei Chen, Rishabh Jain, Mark Johnson, Dhruv Batra, Devi Parikh, Stefan Lee, and Peter Anderson. Nocaps: Novel object captioning at scale. In ICCV, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 209, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 209, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 209, + 296, + 239 + ], + "type": "text", + "content": "[2] Rahaf Aljundi, Punarjay Chakravarty, and Tinne Tuytelaars. Expert gate: Lifelong learning with a network of experts. In CVPR, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 241, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 241, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 296, + 262 + ], + "type": "text", + "content": "[3] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 264, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 264, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 264, + 296, + 306 + ], + "type": "text", + "content": "[4] Qiong Cao, Li Shen, Weidi Xie, Omkar M Parkhi, and Andrew Zisserman. Vggface2: A dataset for recognising faces across pose and age. In IEEE international conference on automatic face & gesture recognition. IEEE, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 308, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 339 + ], + "type": "text", + "content": "[5] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a-gem. In ICLR, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 341, + 296, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 296, + 383 + ], + "type": "text", + "content": "[6] Fei-Long Chen, Du-Zhen Zhang, Ming-Lun Han, Xiu-Yi Chen, Jing Shi, Shuang Xu, and Bo Xu. Vlp: A survey on vision-language pre-training. Machine Intelligence Research, 20(1):38-56, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 384, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 296, + 426 + ], + "type": "text", + "content": "[7] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. TPAMI, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 428, + 296, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 296, + 471 + ], + "type": "text", + "content": "[8] Geonmo Gu, Sanghyuk Chun, Wonjae Kim, Yoohoon Kang, and Sangdoo Yun. Language-only efficient training of zero-shot composed image retrieval. arXiv preprint arXiv:2312.01998, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 472, + 296, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 296, + 493 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 296, + 493 + ], + "type": "text", + "content": "[9] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 494, + 297, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 494, + 297, + 547 + ], + "spans": [ + { + "bbox": [ + 57, + 494, + 297, + 547 + ], + "type": "text", + "content": "[10] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 548, + 297, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 297, + 591 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 297, + 591 + ], + "type": "text", + "content": "[11] Jiun Tian Hoe, Kam Woh Ng, Tianyu Zhang, Chee Seng Chan, Yi-Zhe Song, and Tao Xiang. One loss for all: Deep hashing with a single cosine similarity based learning objective. NeurIPS, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 593, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 593, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 593, + 296, + 624 + ], + "type": "text", + "content": "[12] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In ICLR, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 625, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 296, + 658 + ], + "type": "text", + "content": "[13] Weihua Hu, Rajas Bansal, Kaidi Cao, Nikhil Rao, Karthik Subbian, and Jure Leskovec. Learning backward compatible embeddings. In ACM SIGKDD, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 658, + 296, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 680 + ], + "type": "text", + "content": "[14] Mark J Huiskes and Michael S Lew. The mir flickr retrieval evaluation. In ACM ICMR, pages 39-43, 2008." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "type": "text", + "content": "[15] Young Kyun Jang and Nam Ik Cho. Generalized product quantization network for semi-supervised image retrieval. In CVPR, 2020." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "type": "text", + "content": "[16] Young Kyun Jang, Geonmo Gu, Byungsoo Ko, Isaac Kang, and Nam Ik Cho. Deep hash distillation for image retrieval. In ECCV, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 107, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 149 + ], + "type": "text", + "content": "[17] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 151, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 555, + 183 + ], + "type": "text", + "content": "[18] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV. Springer, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 184, + 554, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 554, + 215 + ], + "type": "text", + "content": "[19] Konwoo Kim, Michael Laskin, Igor Mordatch, and Deepak Pathak. How to adapt your large-scale vision-and-language model. 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 217, + 555, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 248 + ], + "type": "text", + "content": "[20] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. arXiv preprint arXiv:2104.08691, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 251, + 555, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 282 + ], + "type": "text", + "content": "[21] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 284, + 555, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 555, + 327 + ], + "type": "text", + "content": "[22] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 328, + 554, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 554, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 554, + 369 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 372, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 554, + 392 + ], + "type": "text", + "content": "[24] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2017." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 394, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 554, + 426 + ], + "type": "text", + "content": "[25] Jie Lu, Vahid Behbood, Peng Hao, Hua Zuo, Shan Xue, and Guangquan Zhang. Transfer learning using computational intelligence: A survey. Knowledge-Based Systems, 2015." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 427, + 555, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 555, + 448 + ], + "type": "text", + "content": "[26] Qiang Meng, Chixiang Zhang, Xiaogiang Xu, and Feng Zhou. Learning compatible embeddings. In ICCV, 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 449, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 554, + 480 + ], + "type": "text", + "content": "[27] David Nukrai, Ron Mokady, and Amir Globerson. Text-only training for image captioning using noise-injected clip. In EMNLP, 2022." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 482, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 555, + 536 + ], + "type": "text", + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 537, + 555, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 555, + 579 + ], + "type": "text", + "content": "[29] Vivek Ramanujan, Pavan Kumar Anasosalu Vasu, Ali Farhadi, Oncel Tuzel, and Hadi Pouransari. Forward compatible training for large-scale embedding retrieval systems. In CVPR, 2022." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 581, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 555, + 613 + ], + "type": "text", + "content": "[30] Mehwish Rehman, Muhammad Iqbal, Muhammad Sharif, and Mudassar Raza. Content based image retrieval: survey. World Applied Sciences Journal, 2012." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 615, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 555, + 658 + ], + "type": "text", + "content": "[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[32] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 2022." + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1791" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[33] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, pages 2556-2565, 2018." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 149 + ], + "type": "text", + "content": "[34] Yantao Shen, Yuanjun Xiong, Wei Xia, and Stefano Soatto. Towards backward-compatible representation learning. In CVPR, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 195 + ], + "type": "text", + "content": "[35] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. Flava: A foundational language and vision alignment model. In CVPR, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 229 + ], + "type": "text", + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 295, + 274 + ], + "type": "text", + "content": "[37] Ji Wan, Dayong Wang, Steven Chu Hong Hoi, Pengcheng Wu, Jianke Zhu, Yongdong Zhang, and Jintao Li. Deep learning for content-based image retrieval: A comprehensive study. In ACMMM, 2014." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 308 + ], + "type": "text", + "content": "[38] Chien-Yi Wang, Ya-Liang Chang, Shang-Ta Yang, Dong Chen, and Shang-Hong Lai. Unified representation learning for cross model compatibility. In BMVC, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 310, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 295, + 342 + ], + "type": "text", + "content": "[39] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. NeurIPS, 32, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 295, + 365 + ], + "type": "text", + "content": "[40] Kai Wang, Luis Herranz, and Joost van de Weijer. Continual learning in cross-modal retrieval. In CVPR, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 366, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 295, + 432 + ], + "type": "text", + "content": "[41] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumont, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 487 + ], + "type": "text", + "content": "[42] Sijie Yan, Yuanjun Xiong, Kaustav Kundu, Shuo Yang, Siqi Deng, Meng Wang, Wei Xia, and Stefano Soatto. Positive-congruent training: Towards regression-free model updates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14299-14308, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 489, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 532 + ], + "type": "text", + "content": "[43] Binjie Zhang, Yixiao Ge, Yantao Shen, Yu Li, Chun Yuan, Xuyuan Xu, Yexin Wang, and Ying Shan. Hot-refresh model upgrades with regression-alleviating compatible training in image retrieval. In ICLR, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 534, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 577 + ], + "type": "text", + "content": "[44] Binjie Zhang, Yixiao Ge, Yantao Shen, Shupeng Su, Chun Yuan, Xuyuan Xu, Yexin Wang, and Ying Shan. Towards universal backward-compatible representation learning. In IJCAI, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 579, + 295, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 611 + ], + "type": "text", + "content": "[45] Liangli Zhen, Peng Hu, Xi Peng, Rick Siow Mong Goh, and Joey Tianyi Zhou. Deep multimodal transfer learning for cross-modal retrieval. TNNLS, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 613, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 645 + ], + "type": "text", + "content": "[46] Wengang Zhou, Houqiang Li, and Qi Tian. Recent advance in content-based image retrieval: A literature survey. arXiv preprint arXiv:1706.06064, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "text", + "content": "[47] Yifei Zhou, Zilu Li, Abhinav Shrivastava, Hengshuang Zhao, Antonio Torralba, Taipeng Tian, and Ser-Nam Lim. bt2: Backward-compatible training with basis transformation. arXiv preprint arXiv:2211.03989, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[48] Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 73, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 334, + 73, + 555, + 95 + ], + "type": "text", + "content": "comprehensive survey on transfer learning. Proceedings of the IEEE, 2020." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1792" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/2b5d04f4-f92c-4faf-b657-95623ea87551_content_list.json b/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/2b5d04f4-f92c-4faf-b657-95623ea87551_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cee69c75b193c50d895e6cc316f129d9ce140872 --- /dev/null +++ b/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/2b5d04f4-f92c-4faf-b657-95623ea87551_content_list.json @@ -0,0 +1,1540 @@ +[ + { + "type": "text", + "text": "Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning", + "text_level": 1, + "bbox": [ + 112, + 128, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fei Zhou $^{1}$ Peng Wang $^{2}$ Lei Zhang $^{1*}$ Wei Wei $^{1}$ Chen Ding $^{3}$ Guosheng Lin $^{4}$ Yanning Zhang $^{1}$ $^{1}$ School of Computer Science, Northwestern Polytechnical University", + "bbox": [ + 104, + 202, + 934, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Computer Science and Engineering, University of Electronic Science and Technology of China \n $^{3}$ School of Computer Science, Xi'an University of Posts & Telecommunications \n $^{4}$ College of Computing and Data Science, Nanyang Technological University", + "bbox": [ + 99, + 239, + 941, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zhoufei@mail.nwpu.edu.cn wangpeng8619@gmail.com dingchen@xupt.edu.cn gslin@ntu.edu.sg {nwpuzhanglei,weiweinwpu,ynzhang}@nwpu.edu.cn", + "bbox": [ + 204, + 295, + 839, + 327 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 361, + 326, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale pre-trained foundation models have demonstrated remarkable generalization capabilities across diverse computer vision tasks through fine-tuning. However, existing fine-tuning approaches often encounter challenges in extreme cross-domain few-shot learning scenarios, primarily due to the significant domain shift between pre-training data and target tasks, as well as the scarcity of annotated target samples. To mitigate this issue, we propose a novel absorption adaptation learning framework which meticulously regularizes the fine-tuning procedure of foundation model using an expert model with the same architecture but trained from scratch on the targeted data in two aspects. On one hand, we first design a masked cross-model unidirectional reconstruction scheme, which forces the foundation model to recover the intermediate feature of the expert model in a randomly masked manner. On the other hand, a decision graph association loss is developed to encourage the consistency of token similarity matrix between these two models. By doing these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels are appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation. Sufficient experiments with further observations and analyses underpin our observation and argument. The code is available at https://github.com/NWPUZhoufei/FMA.", + "bbox": [ + 88, + 395, + 485, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 811, + 223, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent breakthroughs in foundation models [1-4] have exhibited compelling generalization performance across diverse computer vision tasks. Their efficacy is fundamentally", + "bbox": [ + 89, + 829, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b94f1880682c8b5e68b1d3e7b6a3b064d82a67a16c6759cc760de808fc868d3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 358, + 890, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bbd94ef97a28081471bf0d34f57b3301712df4c64faf41a2465c571ce95a72ac.jpg", + "image_caption": [ + "(a) Comparing the performance of foundation models on both in-domain (e.g., mini-ImageNet, CIFAR-FS) and extreme cross-domain (e.g., ISIC, Chest) FSL.", + "(b) Comparing the performance of ours with existing adaptation approaches on extreme cross-domain (e.g., ISIC) FSL." + ], + "image_footnote": [], + "bbox": [ + 531, + 527, + 890, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Performance of foundation models on extreme CD-FSL. rooted in pre-training massive-scale learnable parameters on extensive datasets, endowing these models with rich semantic knowledge adaptable to multiple domains—especially those aligned with the pre-training distribution. Even straightforward fine-tuning of these models has demonstrated superior performance in downstream tasks. Consequently, there has been a burgeoning interest in developing advanced fine-tuning methodologies [5-9], aiming to enhance the adaptability and efficiency of foundation models in specialized scenarios.", + "bbox": [ + 511, + 686, + 906, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Insight and motivation. While effective in many computer vision tasks, existing fine-tuning strategies encounter significant limitations in extreme cross-domain few-shot", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 116, + 16, + 189, + 54 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 245, + 13, + 799, + 56 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4582", + "bbox": [ + 482, + 931, + 514, + 944 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "learning (CD-FSL) scenarios [10-18]. These tasks are characterized by substantial domain discrepancy from the pretraining data of foundation models, compounded by the scarcity of annotated samples [7, 19-22]. As visualized in Fig. 1(a), foundation models (e.g., DINO[2]) achieve near-saturated few-shot performance on in-domain benchmarks like mini-[ImageNet] [23] and CIFAR-FS [24], which share distributional similarity with pre-training data. However, when applied to extreme cross-domain FSL tasks such as ISIC [25] (medical imaging) and Chest [26] (radiology), their accuracy degrades sharply. We further examine several existing fine-tuning-based adaptation methods for foundation models and compare their performance experimentally, as shown in Fig.1(b). Among these methods, prototype inference (PI\") [27], which directly utilizes the features extracted by the foundation model for classification, exhibits the worst performance. Linear probing (LP\"), which fine-tunes a linear mapping for the features produced by the foundation model in the target domain, shows a slight improvement over prototype inference. Notably, fine-tuning foundation models using the LoRA scheme[5] performs even worse than linear probing. This result suggests that fine-tuning only a small subset of parameters is ineffective, as it fails to mitigate the model's excessive memorization of the source domain. In contrast, standard fine-tuning (\"FT\"), which updates all parameters of the foundation model, proves to be the most effective baseline. However, its performance remains suboptimal compared to the results achieved by foundation models in in-domain scenarios.", + "bbox": [ + 89, + 90, + 485, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "According to the empirical observation above, we find that most existing fine-tuning strategies struggle to effectively address extreme CD-FSL tasks. This limitation stems from the substantial domain gap between the target tasks and the pre-training data of foundation models, coupled with the scarcity of annotated examples in the few-shot setting. Specifically, bridging the domain gap may necessitate updating all parameters of the foundation model, yet the limited supervision provided by the few annotated examples is insufficient to achieve effective adaptation.", + "bbox": [ + 89, + 534, + 483, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, a potential solution is to incorporate pure task-specific knowledge (e.g., a model trained from scratch using the limited annotated samples from the target task) to regularize the fine-tuning process of the foundation model. To validate this perspective, we present visualization results in Fig. 2. Specifically, \"F-ViT\" represents the pre-trained foundation model, while \"T-ViT\" denotes a task-specific expert model with the same architecture as \"F-ViT\" but trained from scratch on the target task's training examples. The following key observations can be made: 1) Although \"F-ViT\" provides high-quality initial features without severe noise corruption, it fails to focus on the primary semantic regions. 2) \"T-ViT\" captures patterns closely related to the target task but introduces substantial noise,", + "bbox": [ + 89, + 689, + 485, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0d76aa9e566790347db0322f0dcba948a0861b26419a9b75d9e7cd7eab2f9847.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 89, + 612, + 157 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8ac476ca11df4d52bed52b92e5002aebf2da2d880cdf181feacc966d9f0d8e7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 616, + 89, + 707, + 157 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9dd8aefff41de6e467daa435a82782f5c0275a3807b4f564856d2ddaf6ba577a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 88, + 803, + 157 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/da1bd50c248a0f473357e5d36ea0b676dd6daf7614400cfd911012b670f4f858.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 807, + 88, + 900, + 157 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4406a74a209e41d8a3894c2a2c150eab75fde1b018b4905ffdba30731a115c32.jpg", + "image_caption": [ + "(a) Image", + "(e) Image", + "Figure 2. Visualization of the class activation mapping (CAM) computed by different models. Results of \"F-ViT\" and \"T-ViT\" in (b) and (c) show complementarity, while \"Ours\" integrates the knowledge of both \"F-ViT\" and \"T-ViT\", resulting in a better response. In terms of average accuracy, \"T-ViT\" scores $27.71\\%$ , \"F-ViT\" scores $39.70\\%$ , and \"Ours\" scores $43.35\\%$ ." + ], + "image_footnote": [], + "bbox": [ + 519, + 172, + 612, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d6c57677ea8f7874e17d5218eede1f0ea55a3448466900363c683a02e23ffecf.jpg", + "image_caption": [ + "(b) F-ViT", + "(f) F-ViT" + ], + "image_footnote": [], + "bbox": [ + 616, + 172, + 709, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/214d3528eb697b28eaf4f24877a5b24476dc4c529f57dd70817b6c127ed89107.jpg", + "image_caption": [ + "(c) T-ViT", + "(g) T-ViT" + ], + "image_footnote": [], + "bbox": [ + 710, + 172, + 803, + 244 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/32ced301ecfa761038f5251e808b612c0331fbc65c149ad61c8e4fbfefcca360.jpg", + "image_caption": [ + "(d) Ours", + "(h) Ours" + ], + "image_footnote": [], + "bbox": [ + 807, + 172, + 900, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "leading to poor generalization. 3) More importantly, the highlighted regions in \"F-ViT\" and \"T-ViT\" exhibit clear complementarity. These findings suggest that adapting the foundation model to extreme CD-FSL tasks can be effectively achieved by leveraging task-relevant knowledge from the expert model during fine-tuning.", + "bbox": [ + 511, + 353, + 906, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Framework. Following this idea, we present a novel absorption adaptation learning framework, which attempt to enhance the extreme cross-domain generalization capacity of a pre-trained vision foundation model (e.g., \"F-ViT\") through absorbing some task-relevant knowledge from an expert model (e.g., \"T-ViT\") during fine-tuning. However, it is difficult to collaborate the learning procedure of these two models, due to their huge difference in feature representation capacity. In a specific, the adequate training of a massive amount of parameters on large-scale pre-training data empowers the foundation model to exhibit powerful and overconfidence semantic representational capabilities, which will inevitably overshadow the task-relevant knowledge exploited by the expert model. Besides, the extensive feature noise corruption produced by the expert model may mislead the fine-tuning of the foundation model.", + "bbox": [ + 511, + 446, + 908, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we propose two innovative strategies: masked cross-model unidirectional reconstruction and decision graph association. The former learns to reconstruct the token features of the expert model using the masked token features of the foundation model with a light-weight decoder. The advantage of this strategy is twofold. On one hand, the reconstruction explicitly injects task-relevant knowledge from the expert model to the foundation model, thus alleviating its overconfidence. On the other hand, the masked input scheme encourages the semantic relationship between local patterns. The latter first defines the decision graphs using the similarity matrix computed by the class token and the patch tokens for each model, and then introduce loss to constrain their consistency. By doing", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4583", + "bbox": [ + 482, + 931, + 514, + 944 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels can be appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation. Extensive results verify our approach and claims.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contributions. The contributions of this work as follows. 1) We propose a novel absorption adaptation learning framework that meticulously regularizes the fine-tuning process of the foundation model using an expert model. 2) We introduce two innovative strategies, including masked cross-model unidirectional reconstruction and decision graph association, to empower the foundation model effectively absorb the task-relevant semantic knowledge from the exert model. 3) We achieve new state-of-the-art results on multiple benchmarks.", + "bbox": [ + 89, + 184, + 486, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 335, + 232, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cross-domain Few-shot learning. Recent works [10-15, 28-31] have developed few-shot learners capable of cross-domain transfer, employing techniques like global-local semantic alignment [10] to enhance representation learning and adversarial training [11, 12, 14] to mitigate the interference of image styles. Some approaches [31, 32] focus on knowledge distillation for effective feature extraction, while others tackle sharp minima in loss landscapes [33] to facilitate knowledge transfer. Additionally, innovative methods like dual adaptive representation alignment [34] and meta-memory frameworks [35] have emerged to bridge domain gaps. Despite some progress, these methods are generally constrained by the use of shallow backbone networks (e.g., ResNet-10) to learn feature initialization models from small-scale source domains like mini-[ImageNet]. In the age of large models, this labor-intensive source domain training process becomes redundant. With the availability of open-source foundation models that offer robust feature initialization, the emphasis of CD-FSL should now shift towards the effective adaptation of foundation models to target domains. Consequently, this study concentrates on addressing the challenges of adapting foundation models in extreme cross-domain scenarios.", + "bbox": [ + 89, + 354, + 486, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Foundation models and adaptation strategies. The emergence of scalable foundation models, facilitated by Transformers [1] and self-supervised learning techniques, marks a significant advancement in FSL. Models like DINO [2], Masked AutoEncoders [3] (MAE), and CLIP[4] learn transferable feature representations from vast unlabeled datasets. Recent research [7, 8, 8, 9, 9, 36-41] focuses on effectively leveraging these foundation models in downstream few-shot tasks while minimizing data requirements for fine-tuning. Techniques such as LoRA [5], Visual Prompt Tuning [6], and automated CLIP prompt engineering [36] aim to enhance efficiency. Additionally, some approaches utilize semantic-guided vision-language alignment to overcome modality", + "bbox": [ + 89, + 705, + 486, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "discrepancies [42]. Others focus on improving ViT-based models by mitigating domain sensitivity in CLS tokens [43] and applying adaptive attention temperature mechanisms to enhance feature stability [44]. Meanwhile, integrating contrastive learning with entropy minimization has been shown to boost feature discrimination in source-free scenarios [45], and frequency-aware prompting strengthens robustness by manipulating high-frequency components [46]. Inspired by in-context learning, context-aware meta-learning [47] facilitates rapid, gradient-free adaptation through modeling relationships between support and query samples. Orthogonally, this study proposes a novel adaptation framework for extreme cross-domain scenarios, enhancing the generalization ability of pre-trained vision foundation models by incorporating task-relevant knowledge during fine-tuning.", + "bbox": [ + 511, + 90, + 908, + 319 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 340, + 648, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. As illustrated in Fig. 3, our framework consists of two structurally identical Transformers, namely Foundation ViT (F-ViT) and Task-specific expert ViT (T-ViT), initialized respectively by the pre-trained vision foundation model and random parameters. Among them, the F-ViT separately takes the original image and its masked image as inputs and extracts their features, respectively. Similarly, the T-ViT receives the original image and then generates its features. More importantly, we meticulously designed a Masked Cross-model unidirectional Reconstruction (MCR) module and a Decision Graph Association (DGA) module to adapt the F-ViT using the task-relevant knowledge exploited by the T-ViT. Specifically, in the MCR module, the masked features from F-ViT are fed into a lightweight decoder to reconstruct the semantic features produced by T-ViT. Then, a reconstruction loss is introduced to encourage the reconstruction accuracy. In the DGA module, we first define the decision graphs using the similarity matrix computed by the class token and the feature tokens (i.e., defined by the ViT architecture) for each model, and then introduce an association loss to constrain their consistency. In the end, the cross-entropy loss, along with reconstruction loss and association loss, are employed for end-to-end training of the entire framework. Once training is completed, we only preserve the F-ViT and its corresponding linear classifier for testing.", + "bbox": [ + 511, + 362, + 908, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Main framework. Given an image $X$ from the training set $\\mathcal{T}$ , we first transform it into patches and then randomly mask out some of the patches. We denote the masked parts as $\\tilde{X} = \\{\\tilde{x}_i | i = 1, \\dots, I\\}$ and the retained parts as $\\hat{X} = \\{\\hat{x}_j | j = 1, \\dots, J\\}$ . The full set of patches corresponding to the original image is denoted as $X = \\hat{X} \\cup \\tilde{X}$ . The masking ratio $\\sigma = \\frac{I}{J + I}$ . Subsequently, we utilize F-ViT to extract features from both $X$ and $\\hat{X}$ separately. Similarly, we also utilize T-ViT to extract the features corresponding to $X$ . The", + "bbox": [ + 511, + 758, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4584", + "bbox": [ + 482, + 931, + 514, + 944 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e6cdb3c68037ac6e68a574a026462ef4a06aaad56a2f0b811fbd5e03a0058a50.jpg", + "image_caption": [ + "Figure 3. We present a novel absorption adaptation learning framework, which attempt to enhance the extreme cross-domain generalization capacity of a pre-trained vision foundation model (e.g., \"F-ViT\"). This framework achieves this by through absorbing some task-relevant knowledge from an expert model (e.g., \"T-ViT\") during fine-tuning. To this end, we introduce two innovative strategies, including masked cross-model unidirectional reconstruction and decision graph association. By doing these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels can be appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation." + ], + "image_footnote": [], + "bbox": [ + 176, + 87, + 433, + 344 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/316178569c307b6e5376c2e08f4fcc64c3d1e341c56252e9e494b06ec2cf4092.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 88, + 823, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "aforementioned process can be formalized as follows:", + "bbox": [ + 89, + 436, + 447, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\{f _ {c l s} ^ {\\mathrm {F}} \\right\\} \\cup \\left\\{\\bar {f} _ {1} ^ {F}, \\dots , \\bar {f} _ {J} ^ {F} \\right\\} \\cup \\left\\{\\tilde {f} _ {1} ^ {F}, \\dots , \\tilde {f} _ {I} ^ {F} \\right\\} = \\mathrm {F} - \\mathrm {V i T} (X), \\\\ \\left\\{\\hat {f} _ {1} ^ {F}, \\dots , \\hat {f} _ {J} ^ {F} \\right\\} = F - V i T (\\hat {X}), \\\\ \\left\\{f _ {c l s} ^ {T} \\right\\} \\cup \\left\\{\\hat {f} _ {1} ^ {T}, \\dots , \\hat {f} _ {J} ^ {T} \\right\\} \\cup \\left\\{\\tilde {f} _ {1} ^ {T}, \\dots , \\tilde {f} _ {I} ^ {T} \\right\\} = \\mathrm {T} - \\mathrm {V i T} (X), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 465, + 483, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{cls}^{\\mathrm{F}}$ and $f_{cls}^{\\mathrm{T}}$ respectively represent the class token features from F-ViT and T-ViT, $\\{\\tilde{f}_1^F,\\dots ,\\tilde{f}_J^F\\} \\cup \\left\\{\\tilde{f}_1^F,\\dots ,\\hat{f}_I^F\\right\\}$ and $\\{\\hat{f}_1^T,\\dots ,\\hat{f}_J^T\\} \\cup \\{\\tilde{f}_1^T,\\dots ,\\tilde{f}_I^T\\}$ respectively represent full patch features of $X$ from F-ViT and T-ViT, $\\{\\hat{f}_1^F,\\dots ,\\hat{f}_J^F\\}$ represents the features w.r.t retained patches after randomly masking.", + "bbox": [ + 89, + 574, + 485, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masked cross-model unidirectional reconstruction. We propose the masked cross-model unidirectional reconstruction (MCR) approach, which aims to reconstruct the masked information in F-ViT under the guidance of T-ViT to enhance the understanding of task-relevant semantic knowledge. In terms of technical details, the MCR module takes the masked features from F-ViT as input and employs a lightweight decoder (e.g., a self-attention layer) to reconstruct the local semantic features of T-ViT. Specifically, we first utilize a set of random vectors $\\{\\tilde{\\varnothing}_1,\\dots ,\\tilde{\\varnothing}_I\\}$ to pad the masked patches. After merging the retained patch features, we can obtain $V = \\{\\hat{f}_1^F,\\dots ,\\hat{f}_J^F\\} \\cup \\{\\tilde{\\varnothing}_1,\\dots ,\\tilde{\\varnothing}_I\\}$ . Afterward, to preserve the spatial structure of the original image patches, we", + "bbox": [ + 89, + 691, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "add positional encoding to $V$ . Subsequently, we utilize the decoder to reconstruct the features of the masked patches as follows:", + "bbox": [ + 511, + 436, + 906, + 482 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\hat {f} _ {1} ^ {D}, \\dots , \\hat {f} _ {J} ^ {D} \\right\\} \\cup \\left\\{\\tilde {f} _ {1} ^ {D}, \\dots , \\tilde {f} _ {I} ^ {D} \\right\\} = \\operatorname {D e c o d e r} (\\mathrm {P E} (V)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 491, + 905, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\left\\{\\tilde{f}_1^D,\\dots ,\\tilde{f}_I^D\\right\\}$ represents the reconstructed features corresponding to the masked patches, PE $(\\cdot)$ represents the positional encoding function. We use the patch features from the corresponding positions in T-ViT $\\left\\{\\tilde{f}_1^T,\\dots ,\\tilde{f}_I^T\\right\\}$ as the reconstruction targets, and the loss $\\mathcal{L}_{rec}$ can be computed as:", + "bbox": [ + 511, + 521, + 905, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\frac {1}{I} \\sum_ {i = 1} ^ {I} \\left\\| \\tilde {f} _ {i} ^ {T} - \\tilde {f} _ {i} ^ {D} \\right\\|. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 625, + 906, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decision graph association. Beyond the MCR module, we expect to introduce a decision-level constraint to further guide the foundation model to absorb task-relevant knowledge from the expert model. While a straightforward approach might involve imposing a logits-level consistency between the foundation model and the expert model, this method is sub-optimal. The primary reason is that these two models have different perceptual decision-making capabilities. Directly enforcing alignment of their logits could cause the entire framework to collapse into a compromise between the two decision states, which might severely misguide the adaptation process of the foundation model. To overcome this issue, we propose a more sophisticated strategy, namely decision graph association (DGA). This method promotes consistency in the token similarity matrix between the two", + "bbox": [ + 511, + 672, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4585", + "bbox": [ + 482, + 931, + 514, + 944 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "models, focusing on aligning how these models relate different features to each other internally. The decision graph is constructed from the token similarity matrix, which captures the relationships between different tokens within the models. By ensuring consistency in these token relationships (the \"graph\") between the foundation model and the expert model, the framework effectively transfers the high-order semantic structures learned by the expert model to the foundation model. In a specific, we first define the decision graphs using the similarity between the class token and the feature tokens for each model, and then learns to association them using a consistency loss. Formally, for the F-ViT branch, we compute the cosine similarity between each patch features $f_{l}^{\\mathrm{F}} \\in \\{\\bar{f}_{1}^{F},\\dots ,\\bar{f}_{J}^{F}\\} \\cup \\{\\tilde{f}_{1}^{F},\\dots ,\\tilde{f}_{I}^{F}\\}$ and the class semantic token features $f_{cls}^{\\mathrm{F}}$ :", + "bbox": [ + 89, + 90, + 483, + 329 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {l} ^ {F} = \\frac {f _ {c l s} ^ {\\mathrm {F}} \\cdot f _ {l} ^ {\\mathrm {F}}}{\\left\\| f _ {c l s} ^ {\\mathrm {F}} \\right\\| \\times \\left\\| f _ {l} ^ {\\mathrm {F}} \\right\\|}, l \\in [ 1, \\dots , J + I ]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 338, + 483, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Afterwards, we process the T-ViT branch in the same manner:", + "bbox": [ + 89, + 376, + 483, + 404 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns _ {l} ^ {T} = \\frac {f _ {c l s} ^ {T} \\cdot f _ {l} ^ {T}}{\\left\\| f _ {c l s} ^ {T} \\right\\| \\times \\left\\| f _ {l} ^ {T} \\right\\|}, l \\in [ 1, \\dots , J + I ], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 414, + 483, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\forall f_{l}^{T}\\in \\left\\{\\hat{f}_{1}^{T},\\dots ,\\hat{f}_{J}^{T}\\right\\} \\cup \\left\\{\\tilde{f}_{1}^{T},\\dots ,\\tilde{f}_{I}^{T}\\right\\}$ . Finally, we calculate the association loss $\\mathcal{L}_{aso}$ as:", + "bbox": [ + 89, + 452, + 483, + 489 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {a s o} = \\frac {1}{J + I} \\sqrt {\\sum_ {l = 1} ^ {J + I} \\left(s _ {l} ^ {F} - s _ {l} ^ {T}\\right) ^ {2}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 502, + 483, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training and testing. Besides MCR and DGA, we employ linear classifiers to predict class affiliation based on the class token features from F-ViT and T-ViT, respectively:", + "bbox": [ + 89, + 560, + 483, + 606 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np ^ {F} = F - \\text {C l a s s i f i e r} \\left(f _ {c l s} ^ {F}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 613, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np ^ {T} = \\mathrm {T - C l a s s i f i e r} \\left(f _ {c l s} ^ {T}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 633, + 383, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p^F$ and $p^T$ respectively represent the predictions. We assign the class corresponding to the maximum value of $p^F$ as the predicted label $y^{F}$ by the F-ViT for input image $X$ . Similarly, we denote the predicted label from T-ViT as $y^{T}$ . Then, we calculate the cross-entropy loss $\\mathcal{L}_{ce}$ based on the predicted labels and the ground truth as:", + "bbox": [ + 89, + 664, + 485, + 756 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} = H \\left(y ^ {F}, y\\right) + H \\left(y ^ {T}, y\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 767, + 483, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y$ represents the ground truth for input image $X$ , and $H(\\cdot)$ denotes the cross-entropy calculation function. In the end, we obtain the total loss as:", + "bbox": [ + 89, + 787, + 483, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {c e} + \\lambda_ {1} \\cdot \\mathcal {L} _ {r e c} + \\lambda_ {2} \\cdot \\mathcal {L} _ {a s o}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 845, + 483, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are hyper-parameters. We optimize the entire framework end-to-end based on the loss $\\mathcal{L}$ . After training, we retain only F-ViT and its classifier for testing.", + "bbox": [ + 89, + 863, + 483, + 907 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/624acfbe92b368f1e9d1fc1ef9282d1c2ad9063cc963f47e3ea644e16473356c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 934, + 165 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f8c18c4e392f169ba34897291d1c0daa3271003515836b17fb9644297e2a7ac2.jpg", + "image_caption": [ + "Figure 4. The differences between ours and MAE." + ], + "image_footnote": [], + "bbox": [ + 517, + 172, + 934, + 236 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a914acf03bcf25e9b66ff053fcdce058a991494cc0fa3ee445628d335ac50bcd.jpg", + "image_caption": [ + "Figure 5. The differences between ours and distillation." + ], + "image_footnote": [], + "bbox": [ + 517, + 253, + 906, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further analysis. Our framework provides deep insights into how to adapt foundation models to extreme CD-FSL tasks. It goes beyond original Masked Auto-Encoder (MAE) [3] or knowledge distillation [10, 48]. We compare our method with MAE and distillation in Fig. 4 and Fig. 5, respectively. Firstly, the original MAE tends to overfit in such few-shot tasks, leading to poor generalization. In contrast, the proposed masked cross-model unidirectional reconstruction scheme offers two advantages. For one thing, our approach explicitly injects task-relevant expert knowledge into the adaptation process, and forces the foundation model to focus on task-relevant semantic patterns. For another, our method can utilize T-ViT to reduce F-ViT's overconfidence. Secondly, simple knowledge distillation is insufficient to assist in adapting foundation models to extreme cross-domain few-shot tasks. The reason is that F-ViT and T-ViT have different perception capabilities for few-shot tasks, and forcibly aligning their features or logits may lead to the model collapsing into a trivial state. In contrast, the proposed decision graph association is a higher-order semantic structural cue with robust properties, leading to better generalization. Thirdly, we jointly train T-ViT and F-ViT, employing DGA for mutual regularization to mitigate overfitting in both models.", + "bbox": [ + 511, + 422, + 908, + 785 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 511, + 796, + 748, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 511, + 814, + 714, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the effectiveness of the proposed method in two distinct settings: extreme CD-FSL and traditional CD-FSL.", + "bbox": [ + 511, + 837, + 906, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extreme setting. In extreme CD-FSL setting, we utilize benchmarks that include ISIC [25], ChestX [26], Eu", + "bbox": [ + 511, + 869, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4586", + "bbox": [ + 482, + 931, + 514, + 944 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a9d59b220cfd56f81b7d0450d7c48c8bb3aaa842f0d86ac69ef747adfa10c149.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-art methods. \"B\" denotes the backbone. \"F\" denotes the initialization with foundation model." + ], + "table_footnote": [], + "table_body": "
ISICChestEuroSATCropDiseaseAve.
MethodBF1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
ProtoNet [27]ViT-SDINO24.1434.0616.4319.1159.0578.7039.3973.4034.7551.31
Linear probing [49]ViT-SDINO24.9451.8818.2925.2960.3883.0251.0184.0738.6561.06
LoRA [5]ViT-SDINO24.4450.2918.2925.3560.0883.0952.8184.0738.9060.70
VPT [6]ViT-SDINO27.7953.6817.7624.8560.3481.9251.1883.5839.2661.00
RotNet [50]ViT-SDINO34.3960.4220.8428.0559.0581.9650.1684.5141.1163.73
SimCLR [51]ViT-SDINO28.7258.5618.8726.0760.5883.6752.8383.3040.2562.90
BYOL [52]ViT-SDINO28.0059.5319.9028.0958.0382.9852.0385.3639.4963.99
SetFeat [53]ViT-SDINO24.1433.5716.4319.2459.0572.5339.3966.6234.7547.99
MaskCLIP [48]ViT-SDINO28.1260.4919.0826.0858.8283.2353.9485.8139.9963.90
OursViT-SDINO37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
", + "bbox": [ + 191, + 80, + 803, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "roSAT [54], and CropDisease [55]. These datasets demonstrate substantial distributional differences compared to the natural image domain [56]. Unlike the previous practice, we re-organize these datasets into a more challenging All-way $K$ -shot configuration, where all categories are considered in the few-shot task, making it a more rigorous test of cross-domain generalization. This setup allows us to more effectively evaluate the ability of our method to handle diverse, real-world domain shifts. Specifically, we randomly sample $K$ samples from all categories in the target domain for model adaptation, and then perform model evaluation on all remaining samples. To mitigate the effects of randomness, we conduct five runs for each target domain and utilize the average result as the metric.", + "bbox": [ + 89, + 202, + 485, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Traditional setting. In the traditional CD-FSL setting, we follow the benchmarks established by Guo et al.[56] and Tseng et al.[57] for experimentation. Specifically, this involves the $N$ -way $K$ -shot $M$ -query setup, where $N \\times K$ support samples are used to adapt the model, and $N \\times M$ query samples are used for evaluation. In this context, $N$ represents a subset randomly sampled from the target domain categories, typically set to 5, while $K$ denotes the number of labeled samples per category, and $M$ represents the number of test samples per category, usually set to 15. For evaluation, we adopt the same protocol as prior studies [10-12, 14, 58], sampling 600 random $N$ -way $K$ -shot tasks from each target domain and using the average accuracy across these tasks as the performance metric. Additionally, previous methods [10-12, 14, 58] typically use two types of inference: inductive and transductive. Inductive inference makes predictions solely based on the support set of few-shot tasks, while transductive inference leverages query data in an unsupervised manner to assist in inference. To ensure a comprehensive comparison, we follow [10, 11, 58] and implement both inductive and transductive inference.", + "bbox": [ + 89, + 412, + 483, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to space constraints, more implementation details can be found in the supplementary document.", + "bbox": [ + 89, + 729, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 89, + 765, + 459, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on extreme CD-FSL. We compare the proposed method with state-of-the-art works to clarify our advantages. Among these comparison methods, ProtoNet [27] and SetFeat [53] represent the classical baseline in FSL. VPT [6] and LoRA [5] are representative methods that achieves knowledge transfer from the foundation model to downstream task. SimCLR [51] and BYOL [52] are two exemplary transfer learning methods. MaskCLIP [48] is a robust representation", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "learning method characterized by its effective generalization potential. We re-implemented these methods under the same settings for comparison in extreme CD-FSL tasks. The experimental results are as shown in Table 1. Overall, the proposed method achieves the best performance compared to these state-of-the-art approaches. Specifically, our method consistently outperforms methods based on transfer learning (e.g., VPT, LoRA, SimCLR, MaskCLIP, etc.). This indicates the strong generalization capability of the proposed method in handling extreme cross-domain scenarios. In addition, compared to ProtoNet [27] and SetFeat [53], our method demonstrates significant performance advantages. This suggests that the robustness of advanced few-shot learners is limited without the introduction of explicit cross-domain learning mechanisms. The above results demonstrate the advantages of our method.", + "bbox": [ + 511, + 203, + 906, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on traditional CD-FSL. Table 2 presents the results for both the 5-way 1-shot and 5-way 5-shot traditional CD-FSL tasks. Overall, our method outperforms advanced comparison methods, further validating its effectiveness.", + "bbox": [ + 511, + 445, + 908, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 508, + 666, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparing our method with baselines. We compare our method with F-ViT and T-ViT baselines to highlight our advantages. Among them, F-ViT refers to the standard fine-tuning of the foundation model. We experimentally found that standard fine-tuning is relatively the best approach. Therefore, we consider F-ViT as our strong baseline. T-ViT refers to a model that learns from scratch using few-shot data from the target domain. We conduct experiments under 1-shot and 5-shot settings across four datasets, initializing F-ViT with DINO unless otherwise specified. The experimental results are shown in Table 3. The proposed method exhibits significant performance advantages in average accuracy across four datasets compared to the T-ViT and F-ViT baselines. Furthermore, this experiment shows that relying solely on T-ViT is insufficient for effectively addressing such extreme cross-domain few-shot tasks. Nevertheless, due to the task-specific intrinsic properties of T-ViT, it still holds some decision-making instructions. Therefore, our framework, which leverages T-ViT to enhance the generalization of F-ViT, possesses practical application value.", + "bbox": [ + 511, + 523, + 908, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluating the effectiveness of MCR and DGA. Our framework consists of two core components: Masked Cross-model unidirectional Reconstruction (MCR) and Decision Graph Association (DGA). We conduct four experimental cases to evaluate the contributions of them. Case 1: The pro", + "bbox": [ + 511, + 825, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4587", + "bbox": [ + 482, + 931, + 514, + 944 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c55c2834a770a83950f7b21d6c2dd5e2c6537cd9dde7e2ad9c8e7d4d4ce08535.jpg", + "table_caption": [ + "Table 2. Comparison with state-of-the-art methods on traditional 5-way CD-FSL. † stands for using query data in an unsupervised manner to assist in inference. “B” denotes the backbone. “F” denotes the initialization with foundation model. “R10” denotes ResNet-10." + ], + "table_footnote": [], + "table_body": "
MethodBFShotCUBCarsPlacesPlantaeChestISICEuroSATCropDiseaseAve.
Dara [34]R10-1-shot52.7035.2551.2542.0822.9236.4267.4280.7448.60
GNN+FAP [46]R10-1-shot46.9134.6954.4136.7322.3635.6362.9669.9745.46
IM-DCL†[46]ViT-B-1-shot----22.6838.2074.6885.98-
PMF [59]ViT-SDINO1-shot78.1337.2471.1153.6021.7330.3670.7480.7955.46
MEM [60]ViT-SDINO1-shot----22.7632.9768.1181.11-
StyleAdv†[11]ViT-SDINO1-shot84.0140.4872.6455.5222.9233.0572.1581.2257.75
StyleAdv-FT†[11]ViT-SDINO1-shot84.0140.4872.6455.5222.9233.9974.9384.1158.57
FLoR†[33]ViT-SDINO1-shot85.4043.4274.6952.2923.2635.4973.0983.5558.90
AttnTemp [44]ViT-SDINO1-shot----23.6338.0575.0984.78-
AttnTemp†[44]ViT-SDINO1-shot----23.9640.1377.4087.58-
CD-CLS [43]ViT-SDINO1-shot----23.3935.5674.9784.54-
CD-CLS†[43]ViT-SDINO1-shot----23.8837.2078.4187.39-
OursViT-SDINO1-shot85.2644.1374.3057.5423.1334.4872.0882.1359.13
Ours†ViT-SDINO1-shot94.1946.4582.4163.8423.1235.0282.3891.3664.84
Dara [34]R10-5-shot77.5158.4472.1565.4027.5456.2885.8495.3267.31
GNN+FAP [46]R10-5-shot67.6650.2074.9854.5425.3147.6082.5291.7961.83
IM-DCL†[46]ViT-B-5-shot----27.9853.8987.7695.93-
PMF [59]ViT-SDINO5-shot----27.2750.1285.9892.96-
MEM [60]ViT-SDINO5-shot----26.6747.3886.4993.74-
StyleAdv†[11]ViT-SDINO5-shot95.8261.7388.3375.5526.9747.7388.5794.8572.44
StyleAdv-FT†[11]ViT-SDINO5-shot95.8266.0288.3378.0126.9751.2390.1295.9974.06
FLoR†[33]ViT-SDINO5-shot96.5368.4489.4876.2227.0253.0690.7596.4774.75
AttnTemp [44]ViT-SDINO5-shot----28.0354.9190.8296.66-
AttnTemp†[44]ViT-SDINO5-shot----28.4155.2291.3496.74-
CD-CLS [43]ViT-SDINO5-shot----27.6654.6991.5396.27-
CD-CLS†[43]ViT-SDINO5-shot----28.2555.6691.6896.62-
OursViT-SDINO5-shot97.5574.7689.1581.7326.2550.5589.9995.8175.72
Ours†ViT-SDINO5-shot98.3978.0091.4684.7426.1151.1992.7997.2677.49
", + "bbox": [ + 176, + 116, + 823, + 385 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c44ff43aa740e5bc6ef71cc24c4144bfe9865fb9e46b9cbf203ad89b9e6d5b36.jpg", + "table_caption": [ + "Table 3. Ablation study. $\\checkmark$ indicates that the component is used." + ], + "table_footnote": [], + "table_body": "
ISICChestEuroSATCropDiseaseAve.
Method1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
T-ViT baseline26.7153.6821.1526.1638.3449.3324.6655.4927.7146.16
F-ViT baseline26.3560.9520.0526.4160.2482.9852.1886.4839.7064.20
Ours37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
MCRDGA1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
XX26.3560.9520.0526.4160.2482.9852.1886.4839.7064.20
✓X32.2762.4921.2827.7260.2583.5354.0586.5241.9665.06
X✓33.1162.2821.1627.3260.2983.9653.5886.4242.0364.99
✓✓37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
", + "bbox": [ + 202, + 401, + 797, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "posed method without the MCR and DGA modules, equivalent to a standard fine-tuning baseline. Case 2: The proposed method using only the MCR module on top of the baseline. Case 3: The proposed method using only the DGA module on top of the baseline. Case 4: The proposed method using both MCR and DGA modules simultaneously. As can be seen in Table 3, compared to the stand fine-tuning baseline (\"F-ViT baseline\"), utilizing either the MCR or the DGA alone can achieve certain performance gains. For example, across the four datasets, the MCR module achieves improvements of $2.26\\%$ and $0.86\\%$ under the 1-shot and 5-shot settings respectively. Similarly, the DGA module achieves improvements of $2.33\\%$ and $0.79\\%$ under the 1-shot and 5-shot settings respectively. Notably, on the ISIC dataset under 1-shot setting, MCR and DGA achieve performance gains of $5.92\\%$ and $6.76\\%$ respectively. Moreover, integrating these two components into a unified framework can unleash greater potential. For instance, compared to using a single component, simultaneously using them further increases the average results by $1.32\\%$ in the 1-shot setting. These results confirm the benefits of the proposed approaches.", + "bbox": [ + 88, + 523, + 486, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 89, + 853, + 259, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing our method with MAE variants. We compare the proposed method with the standard MAE, which", + "bbox": [ + 89, + 869, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "involves adding a masked reconstruction loss on top of the fine-tuning F-ViT baseline. Additionally, our method reconstructs the masked features in F-ViT using T-ViT as the target, aiming to enhance F-ViT's focus on task-specific semantic patterns. Our approach is essentially a unidirectional cross-model masked reconstruction. To demonstrate the benefits of this design, we also compare it with a bidirectional MAE variant, where the masked parts of F-ViT and T-ViT are reconstructed using T-ViT and F-ViT as targets, respectively. As show in Table 4, our method shows clear performance advantages over both of these MAE variants. Firstly, the original MAE tends to over-fit in such FSL tasks, leading to poor generalization. In contrast, our masked cross-model unidirectional reconstruction offers two advantages. For one thing, our approach explicitly injects task-specific model knowledge into the foundation model's adaptation process. For another, our method utilizes T-ViT to relax F-ViT's overconfidence, mitigating over-fitting to some extent. Secondly, the optimization objective of bidirectional MAE is ambiguous. In contrast, our unidirectional cross-modality masked reconstruction approach is goal-oriented, which has resulted in improved performance.", + "bbox": [ + 511, + 523, + 908, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing our method with distillation or fusion. We compare the proposed method with knowledge distillation and model fusion approaches to clarify our previous claims.", + "bbox": [ + 511, + 854, + 911, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4588", + "bbox": [ + 482, + 931, + 514, + 944 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c487f5163aa979278d867b6daaaab1d440bfd9f8f39e6d960ebcf9bbc9f14c16.jpg", + "table_caption": [ + "Table 4. Comparing ours with MAE and distillation based methods." + ], + "table_footnote": [], + "table_body": "
ISICChestEuroSATCropDiseaseAve.
Method1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
MAE35.6462.3921.1927.3356.7877.5152.3885.7341.4963.24
Bidirectional MAE30.3862.0919.3928.6759.8983.0652.5985.0240.5664.71
Feature-level Distillation27.7662.3122.0127.2656.2879.8153.5886.1239.9063.87
Logits-level Distillation27.5361.5420.6725.9960.8883.1854.8786.5140.9864.30
Cross-attention Fusion29.3462.1920.7826.7358.7482.4053.7185.6440.6464.24
Decision Fusion25.4456.9619.8525.5058.9482.7636.2961.9835.1356.80
Ours37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
", + "bbox": [ + 205, + 95, + 792, + 191 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/80376bc072b08dbedb051df53e9f99cb40238c8ba96435ca7ef1995ba35c5e4f.jpg", + "table_caption": [ + "Table 5. Comparison with SoTA methods on mini-ImageNet." + ], + "table_footnote": [], + "table_body": "
MethodBackbone5-way 1-shot5-way 5-shot
CAML [47]ViT-B96.2098.60
SgVA-CLIP [42]ViT-B97.9598.72
PMF [59]ViT-S93.1098.00
OursViT-S94.6898.86
", + "bbox": [ + 140, + 209, + 434, + 267 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/602967c363fd1d8c1f5997803f21eb746e153476f4b1f3b04dbbcd44974897ae.jpg", + "table_caption": [ + "Table 6. Comparison with SoTA methods on tiered-ImageNet." + ], + "table_footnote": [], + "table_body": "
MethodBackbone5-way 1-shot5-way 5-shot
FewTURE [61]ViT-S72.9686.43
HCT [62]ViT-S79.6791.72
OursViT-S82.7893.78
", + "bbox": [ + 143, + 282, + 431, + 330 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/491776c1eb722760b1b8840d11c77072a62fa730a104420fb436e4b7809ada54.jpg", + "table_caption": [ + "Table 7. Comparison with SoTA methods on CIFAR-FS." + ], + "table_footnote": [], + "table_body": "
MethodBackbone5-way 1-shot5-way 5-shot
HCT [62]ViT-S78.8990.50
PMF [59]ViT-S81.1092.50
CAML [47]ViT-B70.8085.50
OursViT-S87.6496.03
", + "bbox": [ + 151, + 349, + 424, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, we conduct variant experiments atop the baseline. Variant 1: Feature-level Distillation (FD), aligns the features of the foundation model and the task-specific model directly. Variant 2: Logits-level Distillation (LD), aligns the class prediction scores of the two models directly. Variant 3: Cross-attention Fusion (CF), which involves integrating the features of the foundation model and task-specific model through cross-attention mechanisms. Variant 4: Decision Fusion (DF), directly integrating the baseline model and task-specific model for decision-making. The results are presented in Table 4. It can be observed that our method consistently outperforms these variants. For instance, across four datasets, our method exhibits improvements ranging from $2.37\\%$ to $8.22\\%$ in 1-shot setting and from $1.22\\%$ to $8.77\\%$ in 5-shot setting, on average. The insights it brings encompass two aspects. Firstly, our framework embodies sophisticated absorption adaptation mechanisms, which are neither simple knowledge distillation nor plain model fusion. Secondly, conventional technical approaches such as knowledge distillation, model fusion, are insufficient to tackle complex extreme CD-FSL. In contrast, our meticulously crafted framework can provide valuable technical contributions.", + "bbox": [ + 89, + 409, + 485, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extending the proposed method to in-domain FSL. We further extend our method to in-domain FSL. We follow common experimental setups from previous works [59, 61, 62] to conduct in-domain FSL experiments on three representative benchmarks: mini-[ImageNet [63], tiered-[ImageNet [64] and CIFAR-FS [24]. To ensure fairness, we maintain the same data settings as previous works [59, 61, 62]. As shown in Table 5, Table 6, and Table 7, the proposed method excels in standard in-domain FSL as well. This reinforces the robustness and versatility of our approach.", + "bbox": [ + 89, + 750, + 486, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/288fffd904cc8d6392a8dfaa3948fc90d5b996d960f14104c4006d11d3773bf6.jpg", + "table_caption": [ + "Table 8. Evaluating the efficiency of our method." + ], + "table_footnote": [], + "table_body": "
TrainingInference
MethodTime (S)Memory (M)Time (S)Memory (M)Accuracy (%)
F-ViT837.40347514.10251639.70
Ours939.50642614.10251643.35
", + "bbox": [ + 517, + 210, + 905, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Efficiency. The efficiency of our method is evaluated in Table 8, with all experiments conducted in the All-way 1-shot setting on a 3090 GPU. The accuracy reported in Table 8 represents the average results across four datasets: ISIC, Chest, EuroSAT, and CropDisease. The training time refers to the duration of each training epoch, while the testing time indicates the total time required to process all testing samples. As shown, our method introduces minimal overhead during training, matches the F-ViT baseline in terms of inference efficiency, and significantly outperforms it in accuracy.", + "bbox": [ + 511, + 273, + 906, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 438, + 638, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we propose a novel absorption adaptation learning framework, where a foundation model is guided by an expert model trained from scratch on targeted data. Central to our method are two innovative approaches: masked cross-model unidirectional reconstruction and decision graph association. These strategies enable the foundation model to absorb task-relevant knowledge from the expert model, both at the intermediate feature and final decision levels, thereby mitigating performance drops caused by domain gaps and limited annotations. Both qualitative and quantitative analyses substantiate the effectiveness of our approach.", + "bbox": [ + 511, + 463, + 908, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. First, our framework employs two parallel ViTs for target domain adaptation, leading to increased computation and storage during training. Second, while it shows promising results, there is still room for improvement, particularly in extreme scenarios. Third, being tailored for ViT models, our approach inherits their limitations.", + "bbox": [ + 511, + 638, + 908, + 729 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This work was supported in part by the National Natural Science Foundation of China under Grand 62372379, Grant 62472359, and Grant 62472350; in part by the Xi'an's Key Industrial Chain Core Technology Breakthrough Project: AI Core Technology Breakthrough under Grand 23ZDCYJSGG0003-2023; in part by Innovation Foundation for Doctor Dissertation of Northwestern Polytechnical University under Grant CX2024017; in part by National Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing fundation under Grant TJ-04-23-04.", + "bbox": [ + 511, + 734, + 910, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4589", + "bbox": [ + 482, + 931, + 514, + 944 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 3", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 3", + "[3] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 5", + "[4] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3", + "[5] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 1, 2, 3, 6", + "[6] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3, 6", + "[7] Samyadeep Basu, Shell Hu, Daniela Massiceti, and Soheil Feizi. Strong baselines for parameter-efficient few-shot fine-tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 11024-11031, 2024. 2, 3", + "[8] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23681-23690, 2024. 3", + "[9] Shuai Shao, Yu Bai, Yan Wang, Baodi Liu, and Bin Liu. Collaborative consortium of foundation models for open-world few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4740-4747, 2024. 1, 3", + "[10] Fei Zhou, Peng Wang, Lei Zhang, Wei Wei, and Yanning Zhang. Revisiting prototypical network for cross" + ], + "bbox": [ + 93, + 114, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20061-20070, 2023. 2, 3, 5, 6", + "[11] Yuqian Fu, Yu Xie, Yanwei Fu, and Yu-Gang Jiang. Styleadv: Meta style adversarial training for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24575-24584, 2023. 3, 6, 7", + "[12] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1075–1081. International Joint Conferences on Artificial Intelligence Organization, 8 2021. Main Track. 3, 6", + "[13] Yi Wu, Ziqiang Li, Chaoyue Wang, Heliang Zheng, Shanshan Zhao, Bin Li, and Dacheng Tao. Domain remodulation for few-shot generative domain adaptation. volume 36, 2024.", + "[14] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 3, 6", + "[15] Pin-Yu Chen. Model reprogramming: Resource-efficient cross-domain machine learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 22584-22591, 2024. 3", + "[16] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple domains for few-shot classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9526-9535, 2021.", + "[17] Yuqian Fu, Yanwei Fu, and Yu-Gang Jiang. Meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. In Proceedings of the 29th ACM international conference on multimedia, pages 5326-5334, 2021.", + "[18] Jiamei Sun, Sebastian Lapuschkin, Wojciech Samek, Yunqing Zhao, Ngai-Man Cheung, and Alexander Binder. Explanation-guided training for cross-domain few-shot classification. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7609-7616. IEEE, 2021. 2", + "[19] Baoquan Zhang, Chuyao Luo, Demin Yu, Xutao Li, Huiwei Lin, Yunming Ye, and Bowen Zhang. Metadata: Meta-learning with conditional diffusion for few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 16687-16695, 2024. 2", + "[20] Ilia Sucholutsky and Thomas L Griffiths. Alignment with human representations supports robust few-shot learning. In Proceedings of the 37th International" + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4590", + "bbox": [ + 482, + 931, + 514, + 944 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Neural Information Processing Systems, pages 73464-73479, 2023.", + "[21] Meiqi Sun, Zhonghan Zhao, Wenhao Chai, Hanjun Luo, Shidong Cao, Yanting Zhang, Jenq-Neng Hwang, and Gaoang Wang. Uniap: Towards universal animal perception in vision via few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5008-5016, 2024.", + "[22] Xingyu Zhu, Shuo Wang, Jinda Lu, Yanbin Hao, Haifeng Liu, and Xiangnan He. Boosting few-shot learning via attentive feature regularization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 7793-7801, 2024. 2", + "[23] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International conference on learning representations, 2016. 2", + "[24] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In International Conference on Learning Representations, 2019. 2, 8", + "[25] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multisource dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. 2, 5", + "[26] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 2, 5", + "[27] Jake Snell, Kevin Swersky, and Richard Zemel. Phototypical networks for few-shot learning. In Advances in Neural Information Processing Systems, pages 4077-4087, 2017. 2, 6", + "[28] Yurong Guo, Ruoyi Du, Yuan Dong, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Task-aware adaptive learning for cross-domain few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1590-1599, 2023. 3", + "[29] Rashindrie Perera and Saman Halgamuge. Discriminative sample-guided and parameter-efficient feature space adaptation for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23794-23804, 2024.", + "[30] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Cross-domain few-shot learning with task-specific adapters. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7161-7170, 2022." + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Hao ZHENG, Runqi Wang, Jianzhuang Liu, and Asako Kanezaki. Cross-level distillation and feature denoising for cross-domain few-shot classification. In The Eleventh International Conference on Learning Representations. 3", + "[32] Yongjin Yang, Taehyeon Kim, and Se-Young Yun. Leveraging normalization layer in adapters with progressive learning and adaptive distillation for cross-domain few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 16370–16378, 2024. 3", + "[33] Yixiong Zou, Yicong Liu, Yiman Hu, Yuhua Li, and Ruixuan Li. Flatten long-range loss landscapes for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23575-23584, 2024. 3, 7", + "[34] Yifan Zhao, Tong Zhang, Jia Li, and Yonghong Tian. Dual adaptive representation alignment for cross-domain few-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11720-11732, 2023. 3, 7", + "[35] Wenjian Wang, Lijuan Duan, Yuxi Wang, Junsong Fan, and Zhaoxiang Zhang. Mmt: cross domain few-shot learning via meta-memory transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3", + "[36] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3", + "[37] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19325-19337, 2023.", + "[38] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15211-15222, 2023.", + "[39] Lin Song, Ruoyi Xue, Hang Wang, Hongbin Sun, Yixiao Ge, Ying Shan, et al. Meta-adapter: An online few-shot learner for vision-language model. Advances in Neural Information Processing Systems, 36:55361-55374, 2023.", + "[40] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024.", + "[41] Jiamin Wu, Xin Liu, Xiaotian Yin, Tianzhu Zhang, and Yongdong Zhang. Task-adaptive prompted transformer" + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4591", + "bbox": [ + 482, + 931, + 513, + 944 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for cross-domain few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 6012-6020, 2024. 3", + "[42] Fang Peng, Xiaoshan Yang, Linhui Xiao, Yaowei Wang, and Changsheng Xu. Sgva-clip: Semantic-guided visual adapting of vision-language models for few-shot image classification. IEEE Transactions on Multimedia, 26:3469-3480, 2023. 3, 8", + "[43] Yixiong Zou, Shuai Yi, Yuhua Li, and Ruixuan Li. A closer look at thecls token for cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:85523-85545, 2024. 3, 7", + "[44] Yixiong Zou, Ran Ma, Yuhua Li, and Ruixuan Li. Attention temperature matters in vit-based cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:116332-116354, 2024. 3, 7", + "[45] Huali Xu, Li Liu, Shuaifeng Zhi, Shaojing Fu, Zhuo Su, Ming-Ming Cheng, and Yongxiang Liu. Enhancing information maximization with distance-aware contrastive learning for source-free cross-domain few-shot learning. IEEE Transactions on Image Processing, 33:2058-2073, 2024. 3", + "[46] Tiange Zhang, Qing Cai, Feng Gao, Lin Qi, and Junyu Dong. Exploring cross-domain few-shot classification via frequency-aware prompting. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 5490–5498, 2024. 3, 7", + "[47] Christopher Fifty, Dennis Duan, Ronald Guenther Junkins, Ehsan Amid, Jure Leskovec, Christopher Re, and Sebastian Thrun. Context-aware meta-learning. In The Twelfth International Conference on Learning Representations. 3, 8", + "[48] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995-11005, 2023. 5, 6", + "[49] Yonatan Belinkov. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207-219, 2022. 6", + "[50] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR 2018, 2018. 6", + "[51] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6", + "[52] Jean-Bastien Grill, Florian Strub, Florent Altché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo," + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6", + "[53] Arman Afrasiyabi, Hugo Larochelle, Jean-François Lalonde, and Christian Gagné. Matching feature sets for few-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9014-9024, 2022. 6", + "[54] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6", + "[55] Punam Bedi, Pushkar Gole, and Sumit Kumar Agarwal. 18 using deep learning for image-based plant disease detection. _Internet Of things and machine learning in agriculture_, pages 369–402, 2021. 6", + "[56] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In European conference on computer vision, pages 124-141. Springer, 2020. 6", + "[57] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In International Conference on Learning Representations, 2019. 6", + "[58] Pan Li, Shaogang Gong, Chengjie Wang, and Yanwei Fu. Ranking distance calibration for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9099-9108, 2022. 6", + "[59] Nenad Tomasev, Ioana Bica, Brian McWilliams, Lars Holger Buesing, Razvan Pascanu, Charles Blundell, and Jovana Mitrovic. Pushing the limits of self-supervised resnets: Can we outperform supervised learning without labels onImagenet? In First Workshop on Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022. 7, 8", + "[60] Reece Walsh, Islam Osman, and Mohamed S Shehata. Masked embedding modeling with rapid domain adjustment for few-shot image classification. IEEE Transactions on Image Processing, 32:4907-4920, 2023. 7", + "[61] Markus Hiller, Rongkai Ma, Mehrtash Harandi, and Tom Drummond. Rethinking generalization in few-shot classification. Advances in Neural Information Processing Systems, 35:3582-3595, 2022. 8", + "[62] Yangji He, Weihan Liang, Dongyang Zhao, Hong-Yu Zhou, Weifeng Ge, Yizhou Yu, and Wenqiang Zhang. Attribute surrogates learning and spectral tokens pooling in transformers for few-shot learning. In Proceed-" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4592", + "bbox": [ + 482, + 931, + 514, + 944 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9119-9129, 2022. 8", + "bbox": [ + 127, + 90, + 480, + 119 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in neural information processing systems, pages 3630-3638, 2016. 8", + "[64] Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. In International Conference on Learning Representations, 2019. 8" + ], + "bbox": [ + 91, + 123, + 483, + 258 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "4593", + "bbox": [ + 482, + 931, + 514, + 944 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/2b5d04f4-f92c-4faf-b657-95623ea87551_model.json b/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/2b5d04f4-f92c-4faf-b657-95623ea87551_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3531a8d93533d750a0f911d9be43956290b49cd2 --- /dev/null +++ b/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/2b5d04f4-f92c-4faf-b657-95623ea87551_model.json @@ -0,0 +1,2358 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.117, + 0.017, + 0.191, + 0.055 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.246, + 0.015, + 0.8, + 0.058 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.13, + 0.885, + 0.177 + ], + "angle": 0, + "content": "Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.203, + 0.936, + 0.24 + ], + "angle": 0, + "content": "Fei Zhou\\(^{1}\\) Peng Wang\\(^{2}\\) Lei Zhang\\(^{1*}\\) Wei Wei\\(^{1}\\) Chen Ding\\(^{3}\\) Guosheng Lin\\(^{4}\\) Yanning Zhang\\(^{1}\\) \n\\(^{1}\\) School of Computer Science, Northwestern Polytechnical University" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.24, + 0.942, + 0.294 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Computer Science and Engineering, University of Electronic Science and Technology of China \n\\(^{3}\\) School of Computer Science, Xi'an University of Posts & Telecommunications \n\\(^{4}\\) College of Computing and Data Science, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.205, + 0.296, + 0.841, + 0.328 + ], + "angle": 0, + "content": "zhoufei@mail.nwpu.edu.cn wangpeng8619@gmail.com dingchen@xupt.edu.cn gslin@ntu.edu.sg {nwpuzhanglei,weiweinwpu,ynzhang}@nwpu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.362, + 0.327, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.396, + 0.486, + 0.789 + ], + "angle": 0, + "content": "Large-scale pre-trained foundation models have demonstrated remarkable generalization capabilities across diverse computer vision tasks through fine-tuning. However, existing fine-tuning approaches often encounter challenges in extreme cross-domain few-shot learning scenarios, primarily due to the significant domain shift between pre-training data and target tasks, as well as the scarcity of annotated target samples. To mitigate this issue, we propose a novel absorption adaptation learning framework which meticulously regularizes the fine-tuning procedure of foundation model using an expert model with the same architecture but trained from scratch on the targeted data in two aspects. On one hand, we first design a masked cross-model unidirectional reconstruction scheme, which forces the foundation model to recover the intermediate feature of the expert model in a randomly masked manner. On the other hand, a decision graph association loss is developed to encourage the consistency of token similarity matrix between these two models. By doing these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels are appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation. Sufficient experiments with further observations and analyses underpin our observation and argument. The code is available at https://github.com/NWPUZhoufei/FMA." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.812, + 0.224, + 0.827 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.83, + 0.485, + 0.876 + ], + "angle": 0, + "content": "Recent breakthroughs in foundation models [1-4] have exhibited compelling generalization performance across diverse computer vision tasks. Their efficacy is fundamentally" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.359, + 0.891, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.492, + 0.89, + 0.529 + ], + "angle": 0, + "content": "(a) Comparing the performance of foundation models on both in-domain (e.g., mini-ImageNet, CIFAR-FS) and extreme cross-domain (e.g., ISIC, Chest) FSL." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.529, + 0.891, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.661, + 0.888, + 0.685 + ], + "angle": 0, + "content": "(b) Comparing the performance of ours with existing adaptation approaches on extreme cross-domain (e.g., ISIC) FSL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.908, + 0.853 + ], + "angle": 0, + "content": "Figure 1. Performance of foundation models on extreme CD-FSL. rooted in pre-training massive-scale learnable parameters on extensive datasets, endowing these models with rich semantic knowledge adaptable to multiple domains—especially those aligned with the pre-training distribution. Even straightforward fine-tuning of these models has demonstrated superior performance in downstream tasks. Consequently, there has been a burgeoning interest in developing advanced fine-tuning methodologies [5-9], aiming to enhance the adaptability and efficiency of foundation models in specialized scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Insight and motivation. While effective in many computer vision tasks, existing fine-tuning strategies encounter significant limitations in extreme cross-domain few-shot" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.933, + 0.516, + 0.945 + ], + "angle": 0, + "content": "4582" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.53 + ], + "angle": 0, + "content": "learning (CD-FSL) scenarios [10-18]. These tasks are characterized by substantial domain discrepancy from the pretraining data of foundation models, compounded by the scarcity of annotated samples [7, 19-22]. As visualized in Fig. 1(a), foundation models (e.g., DINO[2]) achieve near-saturated few-shot performance on in-domain benchmarks like mini-[ImageNet] [23] and CIFAR-FS [24], which share distributional similarity with pre-training data. However, when applied to extreme cross-domain FSL tasks such as ISIC [25] (medical imaging) and Chest [26] (radiology), their accuracy degrades sharply. We further examine several existing fine-tuning-based adaptation methods for foundation models and compare their performance experimentally, as shown in Fig.1(b). Among these methods, prototype inference (PI\") [27], which directly utilizes the features extracted by the foundation model for classification, exhibits the worst performance. Linear probing (LP\"), which fine-tunes a linear mapping for the features produced by the foundation model in the target domain, shows a slight improvement over prototype inference. Notably, fine-tuning foundation models using the LoRA scheme[5] performs even worse than linear probing. This result suggests that fine-tuning only a small subset of parameters is ineffective, as it fails to mitigate the model's excessive memorization of the source domain. In contrast, standard fine-tuning (\"FT\"), which updates all parameters of the foundation model, proves to be the most effective baseline. However, its performance remains suboptimal compared to the results achieved by foundation models in in-domain scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.485, + 0.686 + ], + "angle": 0, + "content": "According to the empirical observation above, we find that most existing fine-tuning strategies struggle to effectively address extreme CD-FSL tasks. This limitation stems from the substantial domain gap between the target tasks and the pre-training data of foundation models, coupled with the scarcity of annotated examples in the few-shot setting. Specifically, bridging the domain gap may necessitate updating all parameters of the foundation model, yet the limited supervision provided by the few annotated examples is insufficient to achieve effective adaptation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.486, + 0.901 + ], + "angle": 0, + "content": "To address this challenge, a potential solution is to incorporate pure task-specific knowledge (e.g., a model trained from scratch using the limited annotated samples from the target task) to regularize the fine-tuning process of the foundation model. To validate this perspective, we present visualization results in Fig. 2. Specifically, \"F-ViT\" represents the pre-trained foundation model, while \"T-ViT\" denotes a task-specific expert model with the same architecture as \"F-ViT\" but trained from scratch on the target task's training examples. The following key observations can be made: 1) Although \"F-ViT\" provides high-quality initial features without severe noise corruption, it fails to focus on the primary semantic regions. 2) \"T-ViT\" captures patterns closely related to the target task but introduces substantial noise," + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.09, + 0.614, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.162, + 0.593, + 0.174 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.09, + 0.709, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.162, + 0.689, + 0.173 + ], + "angle": 0, + "content": "(b) F-ViT" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.089, + 0.805, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.162, + 0.783, + 0.173 + ], + "angle": 0, + "content": "(c) T-ViT" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.089, + 0.901, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.162, + 0.876, + 0.173 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.174, + 0.614, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.248, + 0.593, + 0.259 + ], + "angle": 0, + "content": "(e) Image" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.174, + 0.71, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.247, + 0.688, + 0.258 + ], + "angle": 0, + "content": "(f) F-ViT" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.174, + 0.805, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.733, + 0.247, + 0.783, + 0.259 + ], + "angle": 0, + "content": "(g) T-ViT" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.174, + 0.901, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.832, + 0.247, + 0.876, + 0.258 + ], + "angle": 0, + "content": "(h) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.264, + 0.908, + 0.348 + ], + "angle": 0, + "content": "Figure 2. Visualization of the class activation mapping (CAM) computed by different models. Results of \"F-ViT\" and \"T-ViT\" in (b) and (c) show complementarity, while \"Ours\" integrates the knowledge of both \"F-ViT\" and \"T-ViT\", resulting in a better response. In terms of average accuracy, \"T-ViT\" scores \\(27.71\\%\\), \"F-ViT\" scores \\(39.70\\%\\), and \"Ours\" scores \\(43.35\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.445 + ], + "angle": 0, + "content": "leading to poor generalization. 3) More importantly, the highlighted regions in \"F-ViT\" and \"T-ViT\" exhibit clear complementarity. These findings suggest that adapting the foundation model to extreme CD-FSL tasks can be effectively achieved by leveraging task-relevant knowledge from the expert model during fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.909, + 0.689 + ], + "angle": 0, + "content": "Framework. Following this idea, we present a novel absorption adaptation learning framework, which attempt to enhance the extreme cross-domain generalization capacity of a pre-trained vision foundation model (e.g., \"F-ViT\") through absorbing some task-relevant knowledge from an expert model (e.g., \"T-ViT\") during fine-tuning. However, it is difficult to collaborate the learning procedure of these two models, due to their huge difference in feature representation capacity. In a specific, the adequate training of a massive amount of parameters on large-scale pre-training data empowers the foundation model to exhibit powerful and overconfidence semantic representational capabilities, which will inevitably overshadow the task-relevant knowledge exploited by the expert model. Besides, the extensive feature noise corruption produced by the expert model may mislead the fine-tuning of the foundation model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To address this challenge, we propose two innovative strategies: masked cross-model unidirectional reconstruction and decision graph association. The former learns to reconstruct the token features of the expert model using the masked token features of the foundation model with a light-weight decoder. The advantage of this strategy is twofold. On one hand, the reconstruction explicitly injects task-relevant knowledge from the expert model to the foundation model, thus alleviating its overconfidence. On the other hand, the masked input scheme encourages the semantic relationship between local patterns. The latter first defines the decision graphs using the similarity matrix computed by the class token and the patch tokens for each model, and then introduce loss to constrain their consistency. By doing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.933, + 0.515, + 0.945 + ], + "angle": 0, + "content": "4583" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels can be appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation. Extensive results verify our approach and claims." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.487, + 0.322 + ], + "angle": 0, + "content": "Contributions. The contributions of this work as follows. 1) We propose a novel absorption adaptation learning framework that meticulously regularizes the fine-tuning process of the foundation model using an expert model. 2) We introduce two innovative strategies, including masked cross-model unidirectional reconstruction and decision graph association, to empower the foundation model effectively absorb the task-relevant semantic knowledge from the exert model. 3) We achieve new state-of-the-art results on multiple benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.336, + 0.233, + 0.351 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.487, + 0.701 + ], + "angle": 0, + "content": "Cross-domain Few-shot learning. Recent works [10-15, 28-31] have developed few-shot learners capable of cross-domain transfer, employing techniques like global-local semantic alignment [10] to enhance representation learning and adversarial training [11, 12, 14] to mitigate the interference of image styles. Some approaches [31, 32] focus on knowledge distillation for effective feature extraction, while others tackle sharp minima in loss landscapes [33] to facilitate knowledge transfer. Additionally, innovative methods like dual adaptive representation alignment [34] and meta-memory frameworks [35] have emerged to bridge domain gaps. Despite some progress, these methods are generally constrained by the use of shallow backbone networks (e.g., ResNet-10) to learn feature initialization models from small-scale source domains like mini-[ImageNet]. In the age of large models, this labor-intensive source domain training process becomes redundant. With the availability of open-source foundation models that offer robust feature initialization, the emphasis of CD-FSL should now shift towards the effective adaptation of foundation models to target domains. Consequently, this study concentrates on addressing the challenges of adapting foundation models in extreme cross-domain scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Foundation models and adaptation strategies. The emergence of scalable foundation models, facilitated by Transformers [1] and self-supervised learning techniques, marks a significant advancement in FSL. Models like DINO [2], Masked AutoEncoders [3] (MAE), and CLIP[4] learn transferable feature representations from vast unlabeled datasets. Recent research [7, 8, 8, 9, 9, 36-41] focuses on effectively leveraging these foundation models in downstream few-shot tasks while minimizing data requirements for fine-tuning. Techniques such as LoRA [5], Visual Prompt Tuning [6], and automated CLIP prompt engineering [36] aim to enhance efficiency. Additionally, some approaches utilize semantic-guided vision-language alignment to overcome modality" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.32 + ], + "angle": 0, + "content": "discrepancies [42]. Others focus on improving ViT-based models by mitigating domain sensitivity in CLS tokens [43] and applying adaptive attention temperature mechanisms to enhance feature stability [44]. Meanwhile, integrating contrastive learning with entropy minimization has been shown to boost feature discrimination in source-free scenarios [45], and frequency-aware prompting strengthens robustness by manipulating high-frequency components [46]. Inspired by in-context learning, context-aware meta-learning [47] facilitates rapid, gradient-free adaptation through modeling relationships between support and query samples. Orthogonally, this study proposes a novel adaptation framework for extreme cross-domain scenarios, enhancing the generalization ability of pre-trained vision foundation models by incorporating task-relevant knowledge during fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.342, + 0.649, + 0.359 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.909, + 0.742 + ], + "angle": 0, + "content": "Overview. As illustrated in Fig. 3, our framework consists of two structurally identical Transformers, namely Foundation ViT (F-ViT) and Task-specific expert ViT (T-ViT), initialized respectively by the pre-trained vision foundation model and random parameters. Among them, the F-ViT separately takes the original image and its masked image as inputs and extracts their features, respectively. Similarly, the T-ViT receives the original image and then generates its features. More importantly, we meticulously designed a Masked Cross-model unidirectional Reconstruction (MCR) module and a Decision Graph Association (DGA) module to adapt the F-ViT using the task-relevant knowledge exploited by the T-ViT. Specifically, in the MCR module, the masked features from F-ViT are fed into a lightweight decoder to reconstruct the semantic features produced by T-ViT. Then, a reconstruction loss is introduced to encourage the reconstruction accuracy. In the DGA module, we first define the decision graphs using the similarity matrix computed by the class token and the feature tokens (i.e., defined by the ViT architecture) for each model, and then introduce an association loss to constrain their consistency. In the end, the cross-entropy loss, along with reconstruction loss and association loss, are employed for end-to-end training of the entire framework. Once training is completed, we only preserve the F-ViT and its corresponding linear classifier for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.759, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Main framework. Given an image \\(X\\) from the training set \\(\\mathcal{T}\\), we first transform it into patches and then randomly mask out some of the patches. We denote the masked parts as \\(\\tilde{X} = \\{\\tilde{x}_i | i = 1, \\dots, I\\}\\) and the retained parts as \\(\\hat{X} = \\{\\hat{x}_j | j = 1, \\dots, J\\}\\). The full set of patches corresponding to the original image is denoted as \\(X = \\hat{X} \\cup \\tilde{X}\\). The masking ratio \\(\\sigma = \\frac{I}{J + I}\\). Subsequently, we utilize F-ViT to extract features from both \\(X\\) and \\(\\hat{X}\\) separately. Similarly, we also utilize T-ViT to extract the features corresponding to \\(X\\). The" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.933, + 0.516, + 0.945 + ], + "angle": 0, + "content": "4584" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.177, + 0.088, + 0.434, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.089, + 0.825, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.349, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 3. We present a novel absorption adaptation learning framework, which attempt to enhance the extreme cross-domain generalization capacity of a pre-trained vision foundation model (e.g., \"F-ViT\"). This framework achieves this by through absorbing some task-relevant knowledge from an expert model (e.g., \"T-ViT\") during fine-tuning. To this end, we introduce two innovative strategies, including masked cross-model unidirectional reconstruction and decision graph association. By doing these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels can be appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.438, + 0.449, + 0.455 + ], + "angle": 0, + "content": "aforementioned process can be formalized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.466, + 0.484, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\{f _ {c l s} ^ {\\mathrm {F}} \\right\\} \\cup \\left\\{\\bar {f} _ {1} ^ {F}, \\dots , \\bar {f} _ {J} ^ {F} \\right\\} \\cup \\left\\{\\tilde {f} _ {1} ^ {F}, \\dots , \\tilde {f} _ {I} ^ {F} \\right\\} = \\mathrm {F} - \\mathrm {V i T} (X), \\\\ \\left\\{\\hat {f} _ {1} ^ {F}, \\dots , \\hat {f} _ {J} ^ {F} \\right\\} = F - V i T (\\hat {X}), \\\\ \\left\\{f _ {c l s} ^ {T} \\right\\} \\cup \\left\\{\\hat {f} _ {1} ^ {T}, \\dots , \\hat {f} _ {J} ^ {T} \\right\\} \\cup \\left\\{\\tilde {f} _ {1} ^ {T}, \\dots , \\tilde {f} _ {I} ^ {T} \\right\\} = \\mathrm {T} - \\mathrm {V i T} (X), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.486, + 0.684 + ], + "angle": 0, + "content": "where \\( f_{cls}^{\\mathrm{F}} \\) and \\( f_{cls}^{\\mathrm{T}} \\) respectively represent the class token features from F-ViT and T-ViT, \\( \\{\\tilde{f}_1^F,\\dots ,\\tilde{f}_J^F\\} \\cup \\left\\{\\tilde{f}_1^F,\\dots ,\\hat{f}_I^F\\right\\} \\) and \\( \\{\\hat{f}_1^T,\\dots ,\\hat{f}_J^T\\} \\cup \\{\\tilde{f}_1^T,\\dots ,\\tilde{f}_I^T\\} \\) respectively represent full patch features of \\( X \\) from F-ViT and T-ViT, \\( \\{\\hat{f}_1^F,\\dots ,\\hat{f}_J^F\\} \\) represents the features w.r.t retained patches after randomly masking." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Masked cross-model unidirectional reconstruction. We propose the masked cross-model unidirectional reconstruction (MCR) approach, which aims to reconstruct the masked information in F-ViT under the guidance of T-ViT to enhance the understanding of task-relevant semantic knowledge. In terms of technical details, the MCR module takes the masked features from F-ViT as input and employs a lightweight decoder (e.g., a self-attention layer) to reconstruct the local semantic features of T-ViT. Specifically, we first utilize a set of random vectors \\(\\{\\tilde{\\varnothing}_1,\\dots ,\\tilde{\\varnothing}_I\\}\\) to pad the masked patches. After merging the retained patch features, we can obtain \\(V = \\{\\hat{f}_1^F,\\dots ,\\hat{f}_J^F\\} \\cup \\{\\tilde{\\varnothing}_1,\\dots ,\\tilde{\\varnothing}_I\\}\\). Afterward, to preserve the spatial structure of the original image patches, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.907, + 0.483 + ], + "angle": 0, + "content": "add positional encoding to \\( V \\). Subsequently, we utilize the decoder to reconstruct the features of the masked patches as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.492, + 0.906, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\hat {f} _ {1} ^ {D}, \\dots , \\hat {f} _ {J} ^ {D} \\right\\} \\cup \\left\\{\\tilde {f} _ {1} ^ {D}, \\dots , \\tilde {f} _ {I} ^ {D} \\right\\} = \\operatorname {D e c o d e r} (\\mathrm {P E} (V)), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.625 + ], + "angle": 0, + "content": "where \\(\\left\\{\\tilde{f}_1^D,\\dots ,\\tilde{f}_I^D\\right\\}\\) represents the reconstructed features corresponding to the masked patches, PE \\((\\cdot)\\) represents the positional encoding function. We use the patch features from the corresponding positions in T-ViT \\(\\left\\{\\tilde{f}_1^T,\\dots ,\\tilde{f}_I^T\\right\\}\\) as the reconstruction targets, and the loss \\(\\mathcal{L}_{rec}\\) can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.625, + 0.907, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\frac {1}{I} \\sum_ {i = 1} ^ {I} \\left\\| \\tilde {f} _ {i} ^ {T} - \\tilde {f} _ {i} ^ {D} \\right\\|. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Decision graph association. Beyond the MCR module, we expect to introduce a decision-level constraint to further guide the foundation model to absorb task-relevant knowledge from the expert model. While a straightforward approach might involve imposing a logits-level consistency between the foundation model and the expert model, this method is sub-optimal. The primary reason is that these two models have different perceptual decision-making capabilities. Directly enforcing alignment of their logits could cause the entire framework to collapse into a compromise between the two decision states, which might severely misguide the adaptation process of the foundation model. To overcome this issue, we propose a more sophisticated strategy, namely decision graph association (DGA). This method promotes consistency in the token similarity matrix between the two" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.933, + 0.515, + 0.945 + ], + "angle": 0, + "content": "4585" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.33 + ], + "angle": 0, + "content": "models, focusing on aligning how these models relate different features to each other internally. The decision graph is constructed from the token similarity matrix, which captures the relationships between different tokens within the models. By ensuring consistency in these token relationships (the \"graph\") between the foundation model and the expert model, the framework effectively transfers the high-order semantic structures learned by the expert model to the foundation model. In a specific, we first define the decision graphs using the similarity between the class token and the feature tokens for each model, and then learns to association them using a consistency loss. Formally, for the F-ViT branch, we compute the cosine similarity between each patch features \\( f_{l}^{\\mathrm{F}} \\in \\{\\bar{f}_{1}^{F},\\dots ,\\bar{f}_{J}^{F}\\} \\cup \\{\\tilde{f}_{1}^{F},\\dots ,\\tilde{f}_{I}^{F}\\} \\) and the class semantic token features \\( f_{cls}^{\\mathrm{F}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.339, + 0.484, + 0.376 + ], + "angle": 0, + "content": "\\[\ns _ {l} ^ {F} = \\frac {f _ {c l s} ^ {\\mathrm {F}} \\cdot f _ {l} ^ {\\mathrm {F}}}{\\left\\| f _ {c l s} ^ {\\mathrm {F}} \\right\\| \\times \\left\\| f _ {l} ^ {\\mathrm {F}} \\right\\|}, l \\in [ 1, \\dots , J + I ]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.405 + ], + "angle": 0, + "content": "Afterwards, we process the T-ViT branch in the same manner:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.415, + 0.484, + 0.452 + ], + "angle": 0, + "content": "\\[\ns _ {l} ^ {T} = \\frac {f _ {c l s} ^ {T} \\cdot f _ {l} ^ {T}}{\\left\\| f _ {c l s} ^ {T} \\right\\| \\times \\left\\| f _ {l} ^ {T} \\right\\|}, l \\in [ 1, \\dots , J + I ], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.491 + ], + "angle": 0, + "content": "where \\(\\forall f_{l}^{T}\\in \\left\\{\\hat{f}_{1}^{T},\\dots ,\\hat{f}_{J}^{T}\\right\\} \\cup \\left\\{\\tilde{f}_{1}^{T},\\dots ,\\tilde{f}_{I}^{T}\\right\\}\\). Finally, we calculate the association loss \\(\\mathcal{L}_{aso}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.503, + 0.484, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {a s o} = \\frac {1}{J + I} \\sqrt {\\sum_ {l = 1} ^ {J + I} \\left(s _ {l} ^ {F} - s _ {l} ^ {T}\\right) ^ {2}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.484, + 0.607 + ], + "angle": 0, + "content": "Training and testing. Besides MCR and DGA, we employ linear classifiers to predict class affiliation based on the class token features from F-ViT and T-ViT, respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.614, + 0.484, + 0.64 + ], + "angle": 0, + "content": "\\[\np ^ {F} = F - \\text {C l a s s i f i e r} \\left(f _ {c l s} ^ {F}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.635, + 0.384, + 0.655 + ], + "angle": 0, + "content": "\\[\np ^ {T} = \\mathrm {T - C l a s s i f i e r} \\left(f _ {c l s} ^ {T}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.486, + 0.757 + ], + "angle": 0, + "content": "where \\(p^F\\) and \\(p^T\\) respectively represent the predictions. We assign the class corresponding to the maximum value of \\(p^F\\) as the predicted label \\(y^{F}\\) by the F-ViT for input image \\(X\\). Similarly, we denote the predicted label from T-ViT as \\(y^{T}\\). Then, we calculate the cross-entropy loss \\(\\mathcal{L}_{ce}\\) based on the predicted labels and the ground truth as:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.768, + 0.484, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} = H \\left(y ^ {F}, y\\right) + H \\left(y ^ {T}, y\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.484, + 0.832 + ], + "angle": 0, + "content": "where \\(y\\) represents the ground truth for input image \\(X\\), and \\(H(\\cdot)\\) denotes the cross-entropy calculation function. In the end, we obtain the total loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.846, + 0.484, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {c e} + \\lambda_ {1} \\cdot \\mathcal {L} _ {r e c} + \\lambda_ {2} \\cdot \\mathcal {L} _ {a s o}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.864, + 0.484, + 0.909 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyper-parameters. We optimize the entire framework end-to-end based on the loss \\(\\mathcal{L}\\). After training, we retain only F-ViT and its classifier for testing." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.936, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.174, + 0.936, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.238, + 0.858, + 0.251 + ], + "angle": 0, + "content": "Figure 4. The differences between ours and MAE." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.255, + 0.907, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.403, + 0.874, + 0.416 + ], + "angle": 0, + "content": "Figure 5. The differences between ours and distillation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.909, + 0.786 + ], + "angle": 0, + "content": "Further analysis. Our framework provides deep insights into how to adapt foundation models to extreme CD-FSL tasks. It goes beyond original Masked Auto-Encoder (MAE) [3] or knowledge distillation [10, 48]. We compare our method with MAE and distillation in Fig. 4 and Fig. 5, respectively. Firstly, the original MAE tends to overfit in such few-shot tasks, leading to poor generalization. In contrast, the proposed masked cross-model unidirectional reconstruction scheme offers two advantages. For one thing, our approach explicitly injects task-relevant expert knowledge into the adaptation process, and forces the foundation model to focus on task-relevant semantic patterns. For another, our method can utilize T-ViT to reduce F-ViT's overconfidence. Secondly, simple knowledge distillation is insufficient to assist in adapting foundation models to extreme cross-domain few-shot tasks. The reason is that F-ViT and T-ViT have different perception capabilities for few-shot tasks, and forcibly aligning their features or logits may lead to the model collapsing into a trivial state. In contrast, the proposed decision graph association is a higher-order semantic structural cue with robust properties, leading to better generalization. Thirdly, we jointly train T-ViT and F-ViT, employing DGA for mutual regularization to mitigate overfitting in both models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.797, + 0.749, + 0.813 + ], + "angle": 0, + "content": "4. Experimental Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.715, + 0.831 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.838, + 0.907, + 0.867 + ], + "angle": 0, + "content": "We evaluate the effectiveness of the proposed method in two distinct settings: extreme CD-FSL and traditional CD-FSL." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Extreme setting. In extreme CD-FSL setting, we utilize benchmarks that include ISIC [25], ChestX [26], Eu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.933, + 0.516, + 0.945 + ], + "angle": 0, + "content": "4586" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.068, + 0.888, + 0.081 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art methods. \"B\" denotes the backbone. \"F\" denotes the initialization with foundation model." + }, + { + "type": "table", + "bbox": [ + 0.192, + 0.082, + 0.805, + 0.197 + ], + "angle": 0, + "content": "
ISICChestEuroSATCropDiseaseAve.
MethodBF1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
ProtoNet [27]ViT-SDINO24.1434.0616.4319.1159.0578.7039.3973.4034.7551.31
Linear probing [49]ViT-SDINO24.9451.8818.2925.2960.3883.0251.0184.0738.6561.06
LoRA [5]ViT-SDINO24.4450.2918.2925.3560.0883.0952.8184.0738.9060.70
VPT [6]ViT-SDINO27.7953.6817.7624.8560.3481.9251.1883.5839.2661.00
RotNet [50]ViT-SDINO34.3960.4220.8428.0559.0581.9650.1684.5141.1163.73
SimCLR [51]ViT-SDINO28.7258.5618.8726.0760.5883.6752.8383.3040.2562.90
BYOL [52]ViT-SDINO28.0059.5319.9028.0958.0382.9852.0385.3639.4963.99
SetFeat [53]ViT-SDINO24.1433.5716.4319.2459.0572.5339.3966.6234.7547.99
MaskCLIP [48]ViT-SDINO28.1260.4919.0826.0858.8283.2353.9485.8139.9963.90
OursViT-SDINO37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.203, + 0.486, + 0.413 + ], + "angle": 0, + "content": "roSAT [54], and CropDisease [55]. These datasets demonstrate substantial distributional differences compared to the natural image domain [56]. Unlike the previous practice, we re-organize these datasets into a more challenging All-way \\(K\\)-shot configuration, where all categories are considered in the few-shot task, making it a more rigorous test of cross-domain generalization. This setup allows us to more effectively evaluate the ability of our method to handle diverse, real-world domain shifts. Specifically, we randomly sample \\(K\\) samples from all categories in the target domain for model adaptation, and then perform model evaluation on all remaining samples. To mitigate the effects of randomness, we conduct five runs for each target domain and utilize the average result as the metric." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.485, + 0.73 + ], + "angle": 0, + "content": "Traditional setting. In the traditional CD-FSL setting, we follow the benchmarks established by Guo et al.[56] and Tseng et al.[57] for experimentation. Specifically, this involves the \\(N\\)-way \\(K\\)-shot \\(M\\)-query setup, where \\(N \\times K\\) support samples are used to adapt the model, and \\(N \\times M\\) query samples are used for evaluation. In this context, \\(N\\) represents a subset randomly sampled from the target domain categories, typically set to 5, while \\(K\\) denotes the number of labeled samples per category, and \\(M\\) represents the number of test samples per category, usually set to 15. For evaluation, we adopt the same protocol as prior studies [10-12, 14, 58], sampling 600 random \\(N\\)-way \\(K\\)-shot tasks from each target domain and using the average accuracy across these tasks as the performance metric. Additionally, previous methods [10-12, 14, 58] typically use two types of inference: inductive and transductive. Inductive inference makes predictions solely based on the support set of few-shot tasks, while transductive inference leverages query data in an unsupervised manner to assist in inference. To ensure a comprehensive comparison, we follow [10, 11, 58] and implement both inductive and transductive inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Due to space constraints, more implementation details can be found in the supplementary document." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.766, + 0.46, + 0.78 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results on extreme CD-FSL. We compare the proposed method with state-of-the-art works to clarify our advantages. Among these comparison methods, ProtoNet [27] and SetFeat [53] represent the classical baseline in FSL. VPT [6] and LoRA [5] are representative methods that achieves knowledge transfer from the foundation model to downstream task. SimCLR [51] and BYOL [52] are two exemplary transfer learning methods. MaskCLIP [48] is a robust representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.204, + 0.908, + 0.444 + ], + "angle": 0, + "content": "learning method characterized by its effective generalization potential. We re-implemented these methods under the same settings for comparison in extreme CD-FSL tasks. The experimental results are as shown in Table 1. Overall, the proposed method achieves the best performance compared to these state-of-the-art approaches. Specifically, our method consistently outperforms methods based on transfer learning (e.g., VPT, LoRA, SimCLR, MaskCLIP, etc.). This indicates the strong generalization capability of the proposed method in handling extreme cross-domain scenarios. In addition, compared to ProtoNet [27] and SetFeat [53], our method demonstrates significant performance advantages. This suggests that the robustness of advanced few-shot learners is limited without the introduction of explicit cross-domain learning mechanisms. The above results demonstrate the advantages of our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.909, + 0.506 + ], + "angle": 0, + "content": "Results on traditional CD-FSL. Table 2 presents the results for both the 5-way 1-shot and 5-way 5-shot traditional CD-FSL tasks. Overall, our method outperforms advanced comparison methods, further validating its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.51, + 0.667, + 0.525 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.909, + 0.827 + ], + "angle": 0, + "content": "Comparing our method with baselines. We compare our method with F-ViT and T-ViT baselines to highlight our advantages. Among them, F-ViT refers to the standard fine-tuning of the foundation model. We experimentally found that standard fine-tuning is relatively the best approach. Therefore, we consider F-ViT as our strong baseline. T-ViT refers to a model that learns from scratch using few-shot data from the target domain. We conduct experiments under 1-shot and 5-shot settings across four datasets, initializing F-ViT with DINO unless otherwise specified. The experimental results are shown in Table 3. The proposed method exhibits significant performance advantages in average accuracy across four datasets compared to the T-ViT and F-ViT baselines. Furthermore, this experiment shows that relying solely on T-ViT is insufficient for effectively addressing such extreme cross-domain few-shot tasks. Nevertheless, due to the task-specific intrinsic properties of T-ViT, it still holds some decision-making instructions. Therefore, our framework, which leverages T-ViT to enhance the generalization of F-ViT, possesses practical application value." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Evaluating the effectiveness of MCR and DGA. Our framework consists of two core components: Masked Cross-model unidirectional Reconstruction (MCR) and Decision Graph Association (DGA). We conduct four experimental cases to evaluate the contributions of them. Case 1: The pro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.933, + 0.515, + 0.945 + ], + "angle": 0, + "content": "4587" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.117 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art methods on traditional 5-way CD-FSL. † stands for using query data in an unsupervised manner to assist in inference. “B” denotes the backbone. “F” denotes the initialization with foundation model. “R10” denotes ResNet-10." + }, + { + "type": "table", + "bbox": [ + 0.178, + 0.117, + 0.825, + 0.386 + ], + "angle": 0, + "content": "
MethodBFShotCUBCarsPlacesPlantaeChestISICEuroSATCropDiseaseAve.
Dara [34]R10-1-shot52.7035.2551.2542.0822.9236.4267.4280.7448.60
GNN+FAP [46]R10-1-shot46.9134.6954.4136.7322.3635.6362.9669.9745.46
IM-DCL†[46]ViT-B-1-shot----22.6838.2074.6885.98-
PMF [59]ViT-SDINO1-shot78.1337.2471.1153.6021.7330.3670.7480.7955.46
MEM [60]ViT-SDINO1-shot----22.7632.9768.1181.11-
StyleAdv†[11]ViT-SDINO1-shot84.0140.4872.6455.5222.9233.0572.1581.2257.75
StyleAdv-FT†[11]ViT-SDINO1-shot84.0140.4872.6455.5222.9233.9974.9384.1158.57
FLoR†[33]ViT-SDINO1-shot85.4043.4274.6952.2923.2635.4973.0983.5558.90
AttnTemp [44]ViT-SDINO1-shot----23.6338.0575.0984.78-
AttnTemp†[44]ViT-SDINO1-shot----23.9640.1377.4087.58-
CD-CLS [43]ViT-SDINO1-shot----23.3935.5674.9784.54-
CD-CLS†[43]ViT-SDINO1-shot----23.8837.2078.4187.39-
OursViT-SDINO1-shot85.2644.1374.3057.5423.1334.4872.0882.1359.13
Ours†ViT-SDINO1-shot94.1946.4582.4163.8423.1235.0282.3891.3664.84
Dara [34]R10-5-shot77.5158.4472.1565.4027.5456.2885.8495.3267.31
GNN+FAP [46]R10-5-shot67.6650.2074.9854.5425.3147.6082.5291.7961.83
IM-DCL†[46]ViT-B-5-shot----27.9853.8987.7695.93-
PMF [59]ViT-SDINO5-shot----27.2750.1285.9892.96-
MEM [60]ViT-SDINO5-shot----26.6747.3886.4993.74-
StyleAdv†[11]ViT-SDINO5-shot95.8261.7388.3375.5526.9747.7388.5794.8572.44
StyleAdv-FT†[11]ViT-SDINO5-shot95.8266.0288.3378.0126.9751.2390.1295.9974.06
FLoR†[33]ViT-SDINO5-shot96.5368.4489.4876.2227.0253.0690.7596.4774.75
AttnTemp [44]ViT-SDINO5-shot----28.0354.9190.8296.66-
AttnTemp†[44]ViT-SDINO5-shot----28.4155.2291.3496.74-
CD-CLS [43]ViT-SDINO5-shot----27.6654.6991.5396.27-
CD-CLS†[43]ViT-SDINO5-shot----28.2555.6691.6896.62-
OursViT-SDINO5-shot97.5574.7689.1581.7326.2550.5589.9995.8175.72
Ours†ViT-SDINO5-shot98.3978.0091.4684.7426.1151.1992.7997.2677.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.308, + 0.388, + 0.688, + 0.402 + ], + "angle": 0, + "content": "Table 3. Ablation study. \\(\\checkmark\\) indicates that the component is used." + }, + { + "type": "table", + "bbox": [ + 0.203, + 0.402, + 0.798, + 0.521 + ], + "angle": 0, + "content": "
ISICChestEuroSATCropDiseaseAve.
Method1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
T-ViT baseline26.7153.6821.1526.1638.3449.3324.6655.4927.7146.16
F-ViT baseline26.3560.9520.0526.4160.2482.9852.1886.4839.7064.20
Ours37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
MCRDGA1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
XX26.3560.9520.0526.4160.2482.9852.1886.4839.7064.20
✓X32.2762.4921.2827.7260.2583.5354.0586.5241.9665.06
X✓33.1162.2821.1627.3260.2983.9653.5886.4242.0364.99
✓✓37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.487, + 0.844 + ], + "angle": 0, + "content": "posed method without the MCR and DGA modules, equivalent to a standard fine-tuning baseline. Case 2: The proposed method using only the MCR module on top of the baseline. Case 3: The proposed method using only the DGA module on top of the baseline. Case 4: The proposed method using both MCR and DGA modules simultaneously. As can be seen in Table 3, compared to the stand fine-tuning baseline (\"F-ViT baseline\"), utilizing either the MCR or the DGA alone can achieve certain performance gains. For example, across the four datasets, the MCR module achieves improvements of \\(2.26\\%\\) and \\(0.86\\%\\) under the 1-shot and 5-shot settings respectively. Similarly, the DGA module achieves improvements of \\(2.33\\%\\) and \\(0.79\\%\\) under the 1-shot and 5-shot settings respectively. Notably, on the ISIC dataset under 1-shot setting, MCR and DGA achieve performance gains of \\(5.92\\%\\) and \\(6.76\\%\\) respectively. Moreover, integrating these two components into a unified framework can unleash greater potential. For instance, compared to using a single component, simultaneously using them further increases the average results by \\(1.32\\%\\) in the 1-shot setting. These results confirm the benefits of the proposed approaches." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.854, + 0.26, + 0.869 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Comparing our method with MAE variants. We compare the proposed method with the standard MAE, which" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.909, + 0.854 + ], + "angle": 0, + "content": "involves adding a masked reconstruction loss on top of the fine-tuning F-ViT baseline. Additionally, our method reconstructs the masked features in F-ViT using T-ViT as the target, aiming to enhance F-ViT's focus on task-specific semantic patterns. Our approach is essentially a unidirectional cross-model masked reconstruction. To demonstrate the benefits of this design, we also compare it with a bidirectional MAE variant, where the masked parts of F-ViT and T-ViT are reconstructed using T-ViT and F-ViT as targets, respectively. As show in Table 4, our method shows clear performance advantages over both of these MAE variants. Firstly, the original MAE tends to over-fit in such FSL tasks, leading to poor generalization. In contrast, our masked cross-model unidirectional reconstruction offers two advantages. For one thing, our approach explicitly injects task-specific model knowledge into the foundation model's adaptation process. For another, our method utilizes T-ViT to relax F-ViT's overconfidence, mitigating over-fitting to some extent. Secondly, the optimization objective of bidirectional MAE is ambiguous. In contrast, our unidirectional cross-modality masked reconstruction approach is goal-oriented, which has resulted in improved performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.912, + 0.903 + ], + "angle": 0, + "content": "Comparing our method with distillation or fusion. We compare the proposed method with knowledge distillation and model fusion approaches to clarify our previous claims." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.933, + 0.516, + 0.945 + ], + "angle": 0, + "content": "4588" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.297, + 0.084, + 0.699, + 0.096 + ], + "angle": 0, + "content": "Table 4. Comparing ours with MAE and distillation based methods." + }, + { + "type": "table", + "bbox": [ + 0.207, + 0.096, + 0.794, + 0.193 + ], + "angle": 0, + "content": "
ISICChestEuroSATCropDiseaseAve.
Method1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
MAE35.6462.3921.1927.3356.7877.5152.3885.7341.4963.24
Bidirectional MAE30.3862.0919.3928.6759.8983.0652.5985.0240.5664.71
Feature-level Distillation27.7662.3122.0127.2656.2879.8153.5886.1239.9063.87
Logits-level Distillation27.5361.5420.6725.9960.8883.1854.8786.5140.9864.30
Cross-attention Fusion29.3462.1920.7826.7358.7482.4053.7185.6440.6464.24
Decision Fusion25.4456.9619.8525.5058.9482.7636.2961.9835.1356.80
Ours37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.197, + 0.468, + 0.21 + ], + "angle": 0, + "content": "Table 5. Comparison with SoTA methods on mini-ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.21, + 0.436, + 0.268 + ], + "angle": 0, + "content": "
MethodBackbone5-way 1-shot5-way 5-shot
CAML [47]ViT-B96.2098.60
SgVA-CLIP [42]ViT-B97.9598.72
PMF [59]ViT-S93.1098.00
OursViT-S94.6898.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.102, + 0.271, + 0.471, + 0.284 + ], + "angle": 0, + "content": "Table 6. Comparison with SoTA methods on tiered-ImageNet." + }, + { + "type": "table", + "bbox": [ + 0.144, + 0.284, + 0.432, + 0.332 + ], + "angle": 0, + "content": "
MethodBackbone5-way 1-shot5-way 5-shot
FewTURE [61]ViT-S72.9686.43
HCT [62]ViT-S79.6791.72
OursViT-S82.7893.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.119, + 0.337, + 0.455, + 0.35 + ], + "angle": 0, + "content": "Table 7. Comparison with SoTA methods on CIFAR-FS." + }, + { + "type": "table", + "bbox": [ + 0.152, + 0.351, + 0.425, + 0.406 + ], + "angle": 0, + "content": "
MethodBackbone5-way 1-shot5-way 5-shot
HCT [62]ViT-S78.8990.50
PMF [59]ViT-S81.1092.50
CAML [47]ViT-B70.8085.50
OursViT-S87.6496.03
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.486, + 0.741 + ], + "angle": 0, + "content": "Specifically, we conduct variant experiments atop the baseline. Variant 1: Feature-level Distillation (FD), aligns the features of the foundation model and the task-specific model directly. Variant 2: Logits-level Distillation (LD), aligns the class prediction scores of the two models directly. Variant 3: Cross-attention Fusion (CF), which involves integrating the features of the foundation model and task-specific model through cross-attention mechanisms. Variant 4: Decision Fusion (DF), directly integrating the baseline model and task-specific model for decision-making. The results are presented in Table 4. It can be observed that our method consistently outperforms these variants. For instance, across four datasets, our method exhibits improvements ranging from \\(2.37\\%\\) to \\(8.22\\%\\) in 1-shot setting and from \\(1.22\\%\\) to \\(8.77\\%\\) in 5-shot setting, on average. The insights it brings encompass two aspects. Firstly, our framework embodies sophisticated absorption adaptation mechanisms, which are neither simple knowledge distillation nor plain model fusion. Secondly, conventional technical approaches such as knowledge distillation, model fusion, are insufficient to tackle complex extreme CD-FSL. In contrast, our meticulously crafted framework can provide valuable technical contributions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Extending the proposed method to in-domain FSL. We further extend our method to in-domain FSL. We follow common experimental setups from previous works [59, 61, 62] to conduct in-domain FSL experiments on three representative benchmarks: mini-[ImageNet [63], tiered-[ImageNet [64] and CIFAR-FS [24]. To ensure fairness, we maintain the same data settings as previous works [59, 61, 62]. As shown in Table 5, Table 6, and Table 7, the proposed method excels in standard in-domain FSL as well. This reinforces the robustness and versatility of our approach." + }, + { + "type": "table_caption", + "bbox": [ + 0.563, + 0.197, + 0.856, + 0.21 + ], + "angle": 0, + "content": "Table 8. Evaluating the efficiency of our method." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.211, + 0.906, + 0.264 + ], + "angle": 0, + "content": "
TrainingInference
MethodTime (S)Memory (M)Time (S)Memory (M)Accuracy (%)
F-ViT837.40347514.10251639.70
Ours939.50642614.10251643.35
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.908, + 0.427 + ], + "angle": 0, + "content": "Efficiency. The efficiency of our method is evaluated in Table 8, with all experiments conducted in the All-way 1-shot setting on a 3090 GPU. The accuracy reported in Table 8 represents the average results across four datasets: ISIC, Chest, EuroSAT, and CropDisease. The training time refers to the duration of each training epoch, while the testing time indicates the total time required to process all testing samples. As shown, our method introduces minimal overhead during training, matches the F-ViT baseline in terms of inference efficiency, and significantly outperforms it in accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.439, + 0.64, + 0.454 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.909, + 0.63 + ], + "angle": 0, + "content": "In this study, we propose a novel absorption adaptation learning framework, where a foundation model is guided by an expert model trained from scratch on targeted data. Central to our method are two innovative approaches: masked cross-model unidirectional reconstruction and decision graph association. These strategies enable the foundation model to absorb task-relevant knowledge from the expert model, both at the intermediate feature and final decision levels, thereby mitigating performance drops caused by domain gaps and limited annotations. Both qualitative and quantitative analyses substantiate the effectiveness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.909, + 0.73 + ], + "angle": 0, + "content": "Limitations. First, our framework employs two parallel ViTs for target domain adaptation, leading to increased computation and storage during training. Second, while it shows promising results, there is still room for improvement, particularly in extreme scenarios. Third, being tailored for ViT models, our approach inherits their limitations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Acknowledgment. This work was supported in part by the National Natural Science Foundation of China under Grand 62372379, Grant 62472359, and Grant 62472350; in part by the Xi'an's Key Industrial Chain Core Technology Breakthrough Project: AI Core Technology Breakthrough under Grand 23ZDCYJSGG0003-2023; in part by Innovation Foundation for Doctor Dissertation of Northwestern Polytechnical University under Grant CX2024017; in part by National Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing fundation under Grant TJ-04-23-04." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.933, + 0.516, + 0.945 + ], + "angle": 0, + "content": "4589" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.486, + 0.221 + ], + "angle": 0, + "content": "[1] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.222, + 0.486, + 0.311 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.485, + 0.387 + ], + "angle": 0, + "content": "[3] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.485, + 0.491 + ], + "angle": 0, + "content": "[4] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.493, + 0.485, + 0.568 + ], + "angle": 0, + "content": "[5] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.569, + 0.485, + 0.643 + ], + "angle": 0, + "content": "[6] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.485, + 0.718 + ], + "angle": 0, + "content": "[7] Samyadeep Basu, Shell Hu, Daniela Massiceti, and Soheil Feizi. Strong baselines for parameter-efficient few-shot fine-tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 11024-11031, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.72, + 0.485, + 0.795 + ], + "angle": 0, + "content": "[8] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23681-23690, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.796, + 0.485, + 0.87 + ], + "angle": 0, + "content": "[9] Shuai Shao, Yu Bai, Yan Wang, Baodi Liu, and Bin Liu. Collaborative consortium of foundation models for open-world few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4740-4747, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.871, + 0.485, + 0.902 + ], + "angle": 0, + "content": "[10] Fei Zhou, Peng Wang, Lei Zhang, Wei Wei, and Yanning Zhang. Revisiting prototypical network for cross" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.551, + 0.092, + 0.908, + 0.151 + ], + "angle": 0, + "content": "domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20061-20070, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.154, + 0.908, + 0.228 + ], + "angle": 0, + "content": "[11] Yuqian Fu, Yu Xie, Yanwei Fu, and Yu-Gang Jiang. Styleadv: Meta style adversarial training for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24575-24584, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.23, + 0.908, + 0.334 + ], + "angle": 0, + "content": "[12] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1075–1081. International Joint Conferences on Artificial Intelligence Organization, 8 2021. Main Track. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.336, + 0.908, + 0.395 + ], + "angle": 0, + "content": "[13] Yi Wu, Ziqiang Li, Chaoyue Wang, Heliang Zheng, Shanshan Zhao, Bin Li, and Dacheng Tao. Domain remodulation for few-shot generative domain adaptation. volume 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.398, + 0.908, + 0.457 + ], + "angle": 0, + "content": "[14] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.908, + 0.519 + ], + "angle": 0, + "content": "[15] Pin-Yu Chen. Model reprogramming: Resource-efficient cross-domain machine learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 22584-22591, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.908, + 0.594 + ], + "angle": 0, + "content": "[16] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple domains for few-shot classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9526-9535, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.597, + 0.908, + 0.671 + ], + "angle": 0, + "content": "[17] Yuqian Fu, Yanwei Fu, and Yu-Gang Jiang. Meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. In Proceedings of the 29th ACM international conference on multimedia, pages 5326-5334, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.908, + 0.762 + ], + "angle": 0, + "content": "[18] Jiamei Sun, Sebastian Lapuschkin, Wojciech Samek, Yunqing Zhao, Ngai-Man Cheung, and Alexander Binder. Explanation-guided training for cross-domain few-shot classification. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7609-7616. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.765, + 0.908, + 0.853 + ], + "angle": 0, + "content": "[19] Baoquan Zhang, Chuyao Luo, Demin Yu, Xutao Li, Huiwei Lin, Yunming Ye, and Bowen Zhang. Metadata: Meta-learning with conditional diffusion for few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 16687-16695, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.856, + 0.908, + 0.902 + ], + "angle": 0, + "content": "[20] Ilia Sucholutsky and Thomas L Griffiths. Alignment with human representations supports robust few-shot learning. In Proceedings of the 37th International" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.933, + 0.516, + 0.945 + ], + "angle": 0, + "content": "4590" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.128, + 0.092, + 0.484, + 0.122 + ], + "angle": 0, + "content": "Conference on Neural Information Processing Systems, pages 73464-73479, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.124, + 0.484, + 0.215 + ], + "angle": 0, + "content": "[21] Meiqi Sun, Zhonghan Zhao, Wenhao Chai, Hanjun Luo, Shidong Cao, Yanting Zhang, Jenq-Neng Hwang, and Gaoang Wang. Uniap: Towards universal animal perception in vision via few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5008-5016, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.484, + 0.293 + ], + "angle": 0, + "content": "[22] Xingyu Zhu, Shuo Wang, Jinda Lu, Yanbin Hao, Haifeng Liu, and Xiangnan He. Boosting few-shot learning via attentive feature regularization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 7793-7801, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.295, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[23] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International conference on learning representations, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[24] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In International Conference on Learning Representations, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[25] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multisource dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.484, + 0.575 + ], + "angle": 0, + "content": "[26] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.484, + 0.636 + ], + "angle": 0, + "content": "[27] Jake Snell, Kevin Swersky, and Richard Zemel. Phototypical networks for few-shot learning. In Advances in Neural Information Processing Systems, pages 4077-4087, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.64, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[28] Yurong Guo, Ruoyi Du, Yuan Dong, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Task-aware adaptive learning for cross-domain few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1590-1599, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.484, + 0.822 + ], + "angle": 0, + "content": "[29] Rashindrie Perera and Saman Halgamuge. Discriminative sample-guided and parameter-efficient feature space adaptation for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23794-23804, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.825, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[30] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Cross-domain few-shot learning with task-specific adapters. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7161-7170, 2022." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.166 + ], + "angle": 0, + "content": "[31] Hao ZHENG, Runqi Wang, Jianzhuang Liu, and Asako Kanezaki. Cross-level distillation and feature denoising for cross-domain few-shot classification. In The Eleventh International Conference on Learning Representations. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.169, + 0.908, + 0.258 + ], + "angle": 0, + "content": "[32] Yongjin Yang, Taehyeon Kim, and Se-Young Yun. Leveraging normalization layer in adapters with progressive learning and adaptive distillation for cross-domain few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 16370–16378, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.908, + 0.335 + ], + "angle": 0, + "content": "[33] Yixiong Zou, Yicong Liu, Yiman Hu, Yuhua Li, and Ruixuan Li. Flatten long-range loss landscapes for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23575-23584, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.337, + 0.908, + 0.411 + ], + "angle": 0, + "content": "[34] Yifan Zhao, Tong Zhang, Jia Li, and Yonghong Tian. Dual adaptive representation alignment for cross-domain few-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11720-11732, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.908, + 0.472 + ], + "angle": 0, + "content": "[35] Wenjian Wang, Lijuan Duan, Yuxi Wang, Junsong Fan, and Zhaoxiang Zhang. Mmt: cross domain few-shot learning via meta-memory transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.908, + 0.533 + ], + "angle": 0, + "content": "[36] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.536, + 0.908, + 0.624 + ], + "angle": 0, + "content": "[37] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19325-19337, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.627, + 0.908, + 0.716 + ], + "angle": 0, + "content": "[38] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15211-15222, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.908, + 0.792 + ], + "angle": 0, + "content": "[39] Lin Song, Ruoyi Xue, Hang Wang, Hongbin Sun, Yixiao Ge, Ying Shan, et al. Meta-adapter: An online few-shot learner for vision-language model. Advances in Neural Information Processing Systems, 36:55361-55374, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.795, + 0.908, + 0.869 + ], + "angle": 0, + "content": "[40] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.908, + 0.902 + ], + "angle": 0, + "content": "[41] Jiamin Wu, Xin Liu, Xiaotian Yin, Tianzhu Zhang, and Yongdong Zhang. Task-adaptive prompted transformer" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.933, + 0.514, + 0.945 + ], + "angle": 0, + "content": "4591" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.128, + 0.092, + 0.484, + 0.137 + ], + "angle": 0, + "content": "for cross-domain few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 6012-6020, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.138, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[42] Fang Peng, Xiaoshan Yang, Linhui Xiao, Yaowei Wang, and Changsheng Xu. Sgva-clip: Semantic-guided visual adapting of vision-language models for few-shot image classification. IEEE Transactions on Multimedia, 26:3469-3480, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.215, + 0.483, + 0.274 + ], + "angle": 0, + "content": "[43] Yixiong Zou, Shuai Yi, Yuhua Li, and Ruixuan Li. A closer look at thecls token for cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:85523-85545, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.276, + 0.483, + 0.335 + ], + "angle": 0, + "content": "[44] Yixiong Zou, Ran Ma, Yuhua Li, and Ruixuan Li. Attention temperature matters in vit-based cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:116332-116354, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.337, + 0.483, + 0.426 + ], + "angle": 0, + "content": "[45] Huali Xu, Li Liu, Shuaifeng Zhi, Shaojing Fu, Zhuo Su, Ming-Ming Cheng, and Yongxiang Liu. Enhancing information maximization with distance-aware contrastive learning for source-free cross-domain few-shot learning. IEEE Transactions on Image Processing, 33:2058-2073, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.429, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[46] Tiange Zhang, Qing Cai, Feng Gao, Lin Qi, and Junyu Dong. Exploring cross-domain few-shot classification via frequency-aware prompting. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 5490–5498, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[47] Christopher Fifty, Dennis Duan, Ronald Guenther Junkins, Ehsan Amid, Jure Leskovec, Christopher Re, and Sebastian Thrun. Context-aware meta-learning. In The Twelfth International Conference on Learning Representations. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[48] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995-11005, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[49] Yonatan Belinkov. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207-219, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.483, + 0.778 + ], + "angle": 0, + "content": "[50] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR 2018, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.78, + 0.483, + 0.853 + ], + "angle": 0, + "content": "[51] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.856, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[52] Jean-Bastien Grill, Florian Strub, Florent Altché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.55, + 0.092, + 0.907, + 0.15 + ], + "angle": 0, + "content": "Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.154, + 0.907, + 0.228 + ], + "angle": 0, + "content": "[53] Arman Afrasiyabi, Hugo Larochelle, Jean-François Lalonde, and Christian Gagné. Matching feature sets for few-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9014-9024, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.23, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[54] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.907, + 0.382 + ], + "angle": 0, + "content": "[55] Punam Bedi, Pushkar Gole, and Sumit Kumar Agarwal. 18 using deep learning for image-based plant disease detection. _Internet Of things and machine learning in agriculture_, pages 369–402, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.383, + 0.907, + 0.457 + ], + "angle": 0, + "content": "[56] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In European conference on computer vision, pages 124-141. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.907, + 0.533 + ], + "angle": 0, + "content": "[57] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In International Conference on Learning Representations, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.535, + 0.907, + 0.61 + ], + "angle": 0, + "content": "[58] Pan Li, Shaogang Gong, Chengjie Wang, and Yanwei Fu. Ranking distance calibration for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9099-9108, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.612, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[59] Nenad Tomasev, Ioana Bica, Brian McWilliams, Lars Holger Buesing, Razvan Pascanu, Charles Blundell, and Jovana Mitrovic. Pushing the limits of self-supervised resnets: Can we outperform supervised learning without labels onImagenet? In First Workshop on Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.907, + 0.778 + ], + "angle": 0, + "content": "[60] Reece Walsh, Islam Osman, and Mohamed S Shehata. Masked embedding modeling with rapid domain adjustment for few-shot image classification. IEEE Transactions on Image Processing, 32:4907-4920, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.78, + 0.907, + 0.839 + ], + "angle": 0, + "content": "[61] Markus Hiller, Rongkai Ma, Mehrtash Harandi, and Tom Drummond. Rethinking generalization in few-shot classification. Advances in Neural Information Processing Systems, 35:3582-3595, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.841, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[62] Yangji He, Weihan Liang, Dongyang Zhao, Hong-Yu Zhou, Weifeng Ge, Yizhou Yu, and Wenqiang Zhang. Attribute surrogates learning and spectral tokens pooling in transformers for few-shot learning. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.933, + 0.515, + 0.945 + ], + "angle": 0, + "content": "4592" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.128, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9119-9129, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.124, + 0.484, + 0.183 + ], + "angle": 0, + "content": "[63] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in neural information processing systems, pages 3630-3638, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.185, + 0.485, + 0.26 + ], + "angle": 0, + "content": "[64] Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. In International Conference on Learning Representations, 2019. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.124, + 0.485, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.933, + 0.515, + 0.945 + ], + "angle": 0, + "content": "4593" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/2b5d04f4-f92c-4faf-b657-95623ea87551_origin.pdf b/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/2b5d04f4-f92c-4faf-b657-95623ea87551_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5c97891dcd5216cb5a154feeb56280c16d62c9f2 --- /dev/null +++ b/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/2b5d04f4-f92c-4faf-b657-95623ea87551_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea4736912cc51bc8765014c84dd2b9c5ddf802d8141047984ce4ab912c35258d +size 790518 diff --git a/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/full.md b/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0f9061135413ae73bfa3d4163f5bd7696880c8e7 --- /dev/null +++ b/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/full.md @@ -0,0 +1,326 @@ +# Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning + +Fei Zhou $^{1}$ Peng Wang $^{2}$ Lei Zhang $^{1*}$ Wei Wei $^{1}$ Chen Ding $^{3}$ Guosheng Lin $^{4}$ Yanning Zhang $^{1}$ $^{1}$ School of Computer Science, Northwestern Polytechnical University + +$^{2}$ School of Computer Science and Engineering, University of Electronic Science and Technology of China + $^{3}$ School of Computer Science, Xi'an University of Posts & Telecommunications + $^{4}$ College of Computing and Data Science, Nanyang Technological University + +zhoufei@mail.nwpu.edu.cn wangpeng8619@gmail.com dingchen@xupt.edu.cn gslin@ntu.edu.sg {nwpuzhanglei,weiweinwpu,ynzhang}@nwpu.edu.cn + +# Abstract + +Large-scale pre-trained foundation models have demonstrated remarkable generalization capabilities across diverse computer vision tasks through fine-tuning. However, existing fine-tuning approaches often encounter challenges in extreme cross-domain few-shot learning scenarios, primarily due to the significant domain shift between pre-training data and target tasks, as well as the scarcity of annotated target samples. To mitigate this issue, we propose a novel absorption adaptation learning framework which meticulously regularizes the fine-tuning procedure of foundation model using an expert model with the same architecture but trained from scratch on the targeted data in two aspects. On one hand, we first design a masked cross-model unidirectional reconstruction scheme, which forces the foundation model to recover the intermediate feature of the expert model in a randomly masked manner. On the other hand, a decision graph association loss is developed to encourage the consistency of token similarity matrix between these two models. By doing these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels are appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation. Sufficient experiments with further observations and analyses underpin our observation and argument. The code is available at https://github.com/NWPUZhoufei/FMA. + +# 1. Introduction + +Recent breakthroughs in foundation models [1-4] have exhibited compelling generalization performance across diverse computer vision tasks. Their efficacy is fundamentally + +![](images/b94f1880682c8b5e68b1d3e7b6a3b064d82a67a16c6759cc760de808fc868d3b.jpg) + +![](images/bbd94ef97a28081471bf0d34f57b3301712df4c64faf41a2465c571ce95a72ac.jpg) +(a) Comparing the performance of foundation models on both in-domain (e.g., mini-ImageNet, CIFAR-FS) and extreme cross-domain (e.g., ISIC, Chest) FSL. +(b) Comparing the performance of ours with existing adaptation approaches on extreme cross-domain (e.g., ISIC) FSL. + +Figure 1. Performance of foundation models on extreme CD-FSL. rooted in pre-training massive-scale learnable parameters on extensive datasets, endowing these models with rich semantic knowledge adaptable to multiple domains—especially those aligned with the pre-training distribution. Even straightforward fine-tuning of these models has demonstrated superior performance in downstream tasks. Consequently, there has been a burgeoning interest in developing advanced fine-tuning methodologies [5-9], aiming to enhance the adaptability and efficiency of foundation models in specialized scenarios. + +Insight and motivation. While effective in many computer vision tasks, existing fine-tuning strategies encounter significant limitations in extreme cross-domain few-shot + +learning (CD-FSL) scenarios [10-18]. These tasks are characterized by substantial domain discrepancy from the pretraining data of foundation models, compounded by the scarcity of annotated samples [7, 19-22]. As visualized in Fig. 1(a), foundation models (e.g., DINO[2]) achieve near-saturated few-shot performance on in-domain benchmarks like mini-[ImageNet] [23] and CIFAR-FS [24], which share distributional similarity with pre-training data. However, when applied to extreme cross-domain FSL tasks such as ISIC [25] (medical imaging) and Chest [26] (radiology), their accuracy degrades sharply. We further examine several existing fine-tuning-based adaptation methods for foundation models and compare their performance experimentally, as shown in Fig.1(b). Among these methods, prototype inference (PI") [27], which directly utilizes the features extracted by the foundation model for classification, exhibits the worst performance. Linear probing (LP"), which fine-tunes a linear mapping for the features produced by the foundation model in the target domain, shows a slight improvement over prototype inference. Notably, fine-tuning foundation models using the LoRA scheme[5] performs even worse than linear probing. This result suggests that fine-tuning only a small subset of parameters is ineffective, as it fails to mitigate the model's excessive memorization of the source domain. In contrast, standard fine-tuning ("FT"), which updates all parameters of the foundation model, proves to be the most effective baseline. However, its performance remains suboptimal compared to the results achieved by foundation models in in-domain scenarios. + +According to the empirical observation above, we find that most existing fine-tuning strategies struggle to effectively address extreme CD-FSL tasks. This limitation stems from the substantial domain gap between the target tasks and the pre-training data of foundation models, coupled with the scarcity of annotated examples in the few-shot setting. Specifically, bridging the domain gap may necessitate updating all parameters of the foundation model, yet the limited supervision provided by the few annotated examples is insufficient to achieve effective adaptation. + +To address this challenge, a potential solution is to incorporate pure task-specific knowledge (e.g., a model trained from scratch using the limited annotated samples from the target task) to regularize the fine-tuning process of the foundation model. To validate this perspective, we present visualization results in Fig. 2. Specifically, "F-ViT" represents the pre-trained foundation model, while "T-ViT" denotes a task-specific expert model with the same architecture as "F-ViT" but trained from scratch on the target task's training examples. The following key observations can be made: 1) Although "F-ViT" provides high-quality initial features without severe noise corruption, it fails to focus on the primary semantic regions. 2) "T-ViT" captures patterns closely related to the target task but introduces substantial noise, + +![](images/0d76aa9e566790347db0322f0dcba948a0861b26419a9b75d9e7cd7eab2f9847.jpg) + +![](images/8ac476ca11df4d52bed52b92e5002aebf2da2d880cdf181feacc966d9f0d8e7a.jpg) + +![](images/9dd8aefff41de6e467daa435a82782f5c0275a3807b4f564856d2ddaf6ba577a.jpg) + +![](images/da1bd50c248a0f473357e5d36ea0b676dd6daf7614400cfd911012b670f4f858.jpg) + +![](images/4406a74a209e41d8a3894c2a2c150eab75fde1b018b4905ffdba30731a115c32.jpg) +(a) Image +(e) Image +Figure 2. Visualization of the class activation mapping (CAM) computed by different models. Results of "F-ViT" and "T-ViT" in (b) and (c) show complementarity, while "Ours" integrates the knowledge of both "F-ViT" and "T-ViT", resulting in a better response. In terms of average accuracy, "T-ViT" scores $27.71\%$ , "F-ViT" scores $39.70\%$ , and "Ours" scores $43.35\%$ . + +![](images/d6c57677ea8f7874e17d5218eede1f0ea55a3448466900363c683a02e23ffecf.jpg) +(b) F-ViT +(f) F-ViT + +![](images/214d3528eb697b28eaf4f24877a5b24476dc4c529f57dd70817b6c127ed89107.jpg) +(c) T-ViT +(g) T-ViT + +![](images/32ced301ecfa761038f5251e808b612c0331fbc65c149ad61c8e4fbfefcca360.jpg) +(d) Ours +(h) Ours + +leading to poor generalization. 3) More importantly, the highlighted regions in "F-ViT" and "T-ViT" exhibit clear complementarity. These findings suggest that adapting the foundation model to extreme CD-FSL tasks can be effectively achieved by leveraging task-relevant knowledge from the expert model during fine-tuning. + +Framework. Following this idea, we present a novel absorption adaptation learning framework, which attempt to enhance the extreme cross-domain generalization capacity of a pre-trained vision foundation model (e.g., "F-ViT") through absorbing some task-relevant knowledge from an expert model (e.g., "T-ViT") during fine-tuning. However, it is difficult to collaborate the learning procedure of these two models, due to their huge difference in feature representation capacity. In a specific, the adequate training of a massive amount of parameters on large-scale pre-training data empowers the foundation model to exhibit powerful and overconfidence semantic representational capabilities, which will inevitably overshadow the task-relevant knowledge exploited by the expert model. Besides, the extensive feature noise corruption produced by the expert model may mislead the fine-tuning of the foundation model. + +To address this challenge, we propose two innovative strategies: masked cross-model unidirectional reconstruction and decision graph association. The former learns to reconstruct the token features of the expert model using the masked token features of the foundation model with a light-weight decoder. The advantage of this strategy is twofold. On one hand, the reconstruction explicitly injects task-relevant knowledge from the expert model to the foundation model, thus alleviating its overconfidence. On the other hand, the masked input scheme encourages the semantic relationship between local patterns. The latter first defines the decision graphs using the similarity matrix computed by the class token and the patch tokens for each model, and then introduce loss to constrain their consistency. By doing + +these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels can be appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation. Extensive results verify our approach and claims. + +Contributions. The contributions of this work as follows. 1) We propose a novel absorption adaptation learning framework that meticulously regularizes the fine-tuning process of the foundation model using an expert model. 2) We introduce two innovative strategies, including masked cross-model unidirectional reconstruction and decision graph association, to empower the foundation model effectively absorb the task-relevant semantic knowledge from the exert model. 3) We achieve new state-of-the-art results on multiple benchmarks. + +# 2. Related Work + +Cross-domain Few-shot learning. Recent works [10-15, 28-31] have developed few-shot learners capable of cross-domain transfer, employing techniques like global-local semantic alignment [10] to enhance representation learning and adversarial training [11, 12, 14] to mitigate the interference of image styles. Some approaches [31, 32] focus on knowledge distillation for effective feature extraction, while others tackle sharp minima in loss landscapes [33] to facilitate knowledge transfer. Additionally, innovative methods like dual adaptive representation alignment [34] and meta-memory frameworks [35] have emerged to bridge domain gaps. Despite some progress, these methods are generally constrained by the use of shallow backbone networks (e.g., ResNet-10) to learn feature initialization models from small-scale source domains like mini-[ImageNet]. In the age of large models, this labor-intensive source domain training process becomes redundant. With the availability of open-source foundation models that offer robust feature initialization, the emphasis of CD-FSL should now shift towards the effective adaptation of foundation models to target domains. Consequently, this study concentrates on addressing the challenges of adapting foundation models in extreme cross-domain scenarios. + +Foundation models and adaptation strategies. The emergence of scalable foundation models, facilitated by Transformers [1] and self-supervised learning techniques, marks a significant advancement in FSL. Models like DINO [2], Masked AutoEncoders [3] (MAE), and CLIP[4] learn transferable feature representations from vast unlabeled datasets. Recent research [7, 8, 8, 9, 9, 36-41] focuses on effectively leveraging these foundation models in downstream few-shot tasks while minimizing data requirements for fine-tuning. Techniques such as LoRA [5], Visual Prompt Tuning [6], and automated CLIP prompt engineering [36] aim to enhance efficiency. Additionally, some approaches utilize semantic-guided vision-language alignment to overcome modality + +discrepancies [42]. Others focus on improving ViT-based models by mitigating domain sensitivity in CLS tokens [43] and applying adaptive attention temperature mechanisms to enhance feature stability [44]. Meanwhile, integrating contrastive learning with entropy minimization has been shown to boost feature discrimination in source-free scenarios [45], and frequency-aware prompting strengthens robustness by manipulating high-frequency components [46]. Inspired by in-context learning, context-aware meta-learning [47] facilitates rapid, gradient-free adaptation through modeling relationships between support and query samples. Orthogonally, this study proposes a novel adaptation framework for extreme cross-domain scenarios, enhancing the generalization ability of pre-trained vision foundation models by incorporating task-relevant knowledge during fine-tuning. + +# 3. Methodology + +Overview. As illustrated in Fig. 3, our framework consists of two structurally identical Transformers, namely Foundation ViT (F-ViT) and Task-specific expert ViT (T-ViT), initialized respectively by the pre-trained vision foundation model and random parameters. Among them, the F-ViT separately takes the original image and its masked image as inputs and extracts their features, respectively. Similarly, the T-ViT receives the original image and then generates its features. More importantly, we meticulously designed a Masked Cross-model unidirectional Reconstruction (MCR) module and a Decision Graph Association (DGA) module to adapt the F-ViT using the task-relevant knowledge exploited by the T-ViT. Specifically, in the MCR module, the masked features from F-ViT are fed into a lightweight decoder to reconstruct the semantic features produced by T-ViT. Then, a reconstruction loss is introduced to encourage the reconstruction accuracy. In the DGA module, we first define the decision graphs using the similarity matrix computed by the class token and the feature tokens (i.e., defined by the ViT architecture) for each model, and then introduce an association loss to constrain their consistency. In the end, the cross-entropy loss, along with reconstruction loss and association loss, are employed for end-to-end training of the entire framework. Once training is completed, we only preserve the F-ViT and its corresponding linear classifier for testing. + +Main framework. Given an image $X$ from the training set $\mathcal{T}$ , we first transform it into patches and then randomly mask out some of the patches. We denote the masked parts as $\tilde{X} = \{\tilde{x}_i | i = 1, \dots, I\}$ and the retained parts as $\hat{X} = \{\hat{x}_j | j = 1, \dots, J\}$ . The full set of patches corresponding to the original image is denoted as $X = \hat{X} \cup \tilde{X}$ . The masking ratio $\sigma = \frac{I}{J + I}$ . Subsequently, we utilize F-ViT to extract features from both $X$ and $\hat{X}$ separately. Similarly, we also utilize T-ViT to extract the features corresponding to $X$ . The + +![](images/e6cdb3c68037ac6e68a574a026462ef4a06aaad56a2f0b811fbd5e03a0058a50.jpg) +Figure 3. We present a novel absorption adaptation learning framework, which attempt to enhance the extreme cross-domain generalization capacity of a pre-trained vision foundation model (e.g., "F-ViT"). This framework achieves this by through absorbing some task-relevant knowledge from an expert model (e.g., "T-ViT") during fine-tuning. To this end, we introduce two innovative strategies, including masked cross-model unidirectional reconstruction and decision graph association. By doing these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels can be appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation. + +![](images/316178569c307b6e5376c2e08f4fcc64c3d1e341c56252e9e494b06ec2cf4092.jpg) + +aforementioned process can be formalized as follows: + +$$ +\begin{array}{l} \left\{f _ {c l s} ^ {\mathrm {F}} \right\} \cup \left\{\bar {f} _ {1} ^ {F}, \dots , \bar {f} _ {J} ^ {F} \right\} \cup \left\{\tilde {f} _ {1} ^ {F}, \dots , \tilde {f} _ {I} ^ {F} \right\} = \mathrm {F} - \mathrm {V i T} (X), \\ \left\{\hat {f} _ {1} ^ {F}, \dots , \hat {f} _ {J} ^ {F} \right\} = F - V i T (\hat {X}), \\ \left\{f _ {c l s} ^ {T} \right\} \cup \left\{\hat {f} _ {1} ^ {T}, \dots , \hat {f} _ {J} ^ {T} \right\} \cup \left\{\tilde {f} _ {1} ^ {T}, \dots , \tilde {f} _ {I} ^ {T} \right\} = \mathrm {T} - \mathrm {V i T} (X), \tag {1} \\ \end{array} +$$ + +where $f_{cls}^{\mathrm{F}}$ and $f_{cls}^{\mathrm{T}}$ respectively represent the class token features from F-ViT and T-ViT, $\{\tilde{f}_1^F,\dots ,\tilde{f}_J^F\} \cup \left\{\tilde{f}_1^F,\dots ,\hat{f}_I^F\right\}$ and $\{\hat{f}_1^T,\dots ,\hat{f}_J^T\} \cup \{\tilde{f}_1^T,\dots ,\tilde{f}_I^T\}$ respectively represent full patch features of $X$ from F-ViT and T-ViT, $\{\hat{f}_1^F,\dots ,\hat{f}_J^F\}$ represents the features w.r.t retained patches after randomly masking. + +Masked cross-model unidirectional reconstruction. We propose the masked cross-model unidirectional reconstruction (MCR) approach, which aims to reconstruct the masked information in F-ViT under the guidance of T-ViT to enhance the understanding of task-relevant semantic knowledge. In terms of technical details, the MCR module takes the masked features from F-ViT as input and employs a lightweight decoder (e.g., a self-attention layer) to reconstruct the local semantic features of T-ViT. Specifically, we first utilize a set of random vectors $\{\tilde{\varnothing}_1,\dots ,\tilde{\varnothing}_I\}$ to pad the masked patches. After merging the retained patch features, we can obtain $V = \{\hat{f}_1^F,\dots ,\hat{f}_J^F\} \cup \{\tilde{\varnothing}_1,\dots ,\tilde{\varnothing}_I\}$ . Afterward, to preserve the spatial structure of the original image patches, we + +add positional encoding to $V$ . Subsequently, we utilize the decoder to reconstruct the features of the masked patches as follows: + +$$ +\left\{\hat {f} _ {1} ^ {D}, \dots , \hat {f} _ {J} ^ {D} \right\} \cup \left\{\tilde {f} _ {1} ^ {D}, \dots , \tilde {f} _ {I} ^ {D} \right\} = \operatorname {D e c o d e r} (\mathrm {P E} (V)), \tag {2} +$$ + +where $\left\{\tilde{f}_1^D,\dots ,\tilde{f}_I^D\right\}$ represents the reconstructed features corresponding to the masked patches, PE $(\cdot)$ represents the positional encoding function. We use the patch features from the corresponding positions in T-ViT $\left\{\tilde{f}_1^T,\dots ,\tilde{f}_I^T\right\}$ as the reconstruction targets, and the loss $\mathcal{L}_{rec}$ can be computed as: + +$$ +\mathcal {L} _ {r e c} = \frac {1}{I} \sum_ {i = 1} ^ {I} \left\| \tilde {f} _ {i} ^ {T} - \tilde {f} _ {i} ^ {D} \right\|. \tag {3} +$$ + +Decision graph association. Beyond the MCR module, we expect to introduce a decision-level constraint to further guide the foundation model to absorb task-relevant knowledge from the expert model. While a straightforward approach might involve imposing a logits-level consistency between the foundation model and the expert model, this method is sub-optimal. The primary reason is that these two models have different perceptual decision-making capabilities. Directly enforcing alignment of their logits could cause the entire framework to collapse into a compromise between the two decision states, which might severely misguide the adaptation process of the foundation model. To overcome this issue, we propose a more sophisticated strategy, namely decision graph association (DGA). This method promotes consistency in the token similarity matrix between the two + +models, focusing on aligning how these models relate different features to each other internally. The decision graph is constructed from the token similarity matrix, which captures the relationships between different tokens within the models. By ensuring consistency in these token relationships (the "graph") between the foundation model and the expert model, the framework effectively transfers the high-order semantic structures learned by the expert model to the foundation model. In a specific, we first define the decision graphs using the similarity between the class token and the feature tokens for each model, and then learns to association them using a consistency loss. Formally, for the F-ViT branch, we compute the cosine similarity between each patch features $f_{l}^{\mathrm{F}} \in \{\bar{f}_{1}^{F},\dots ,\bar{f}_{J}^{F}\} \cup \{\tilde{f}_{1}^{F},\dots ,\tilde{f}_{I}^{F}\}$ and the class semantic token features $f_{cls}^{\mathrm{F}}$ : + +$$ +s _ {l} ^ {F} = \frac {f _ {c l s} ^ {\mathrm {F}} \cdot f _ {l} ^ {\mathrm {F}}}{\left\| f _ {c l s} ^ {\mathrm {F}} \right\| \times \left\| f _ {l} ^ {\mathrm {F}} \right\|}, l \in [ 1, \dots , J + I ]. \tag {4} +$$ + +Afterwards, we process the T-ViT branch in the same manner: + +$$ +s _ {l} ^ {T} = \frac {f _ {c l s} ^ {T} \cdot f _ {l} ^ {T}}{\left\| f _ {c l s} ^ {T} \right\| \times \left\| f _ {l} ^ {T} \right\|}, l \in [ 1, \dots , J + I ], \tag {5} +$$ + +where $\forall f_{l}^{T}\in \left\{\hat{f}_{1}^{T},\dots ,\hat{f}_{J}^{T}\right\} \cup \left\{\tilde{f}_{1}^{T},\dots ,\tilde{f}_{I}^{T}\right\}$ . Finally, we calculate the association loss $\mathcal{L}_{aso}$ as: + +$$ +\mathcal {L} _ {a s o} = \frac {1}{J + I} \sqrt {\sum_ {l = 1} ^ {J + I} \left(s _ {l} ^ {F} - s _ {l} ^ {T}\right) ^ {2}}. \tag {6} +$$ + +Training and testing. Besides MCR and DGA, we employ linear classifiers to predict class affiliation based on the class token features from F-ViT and T-ViT, respectively: + +$$ +p ^ {F} = F - \text {C l a s s i f i e r} \left(f _ {c l s} ^ {F}\right), \tag {7} +$$ + +$$ +p ^ {T} = \mathrm {T - C l a s s i f i e r} \left(f _ {c l s} ^ {T}\right), +$$ + +where $p^F$ and $p^T$ respectively represent the predictions. We assign the class corresponding to the maximum value of $p^F$ as the predicted label $y^{F}$ by the F-ViT for input image $X$ . Similarly, we denote the predicted label from T-ViT as $y^{T}$ . Then, we calculate the cross-entropy loss $\mathcal{L}_{ce}$ based on the predicted labels and the ground truth as: + +$$ +\mathcal {L} _ {c e} = H \left(y ^ {F}, y\right) + H \left(y ^ {T}, y\right), \tag {8} +$$ + +where $y$ represents the ground truth for input image $X$ , and $H(\cdot)$ denotes the cross-entropy calculation function. In the end, we obtain the total loss as: + +$$ +\mathcal {L} = \mathcal {L} _ {c e} + \lambda_ {1} \cdot \mathcal {L} _ {r e c} + \lambda_ {2} \cdot \mathcal {L} _ {a s o}, \tag {9} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are hyper-parameters. We optimize the entire framework end-to-end based on the loss $\mathcal{L}$ . After training, we retain only F-ViT and its classifier for testing. + +![](images/624acfbe92b368f1e9d1fc1ef9282d1c2ad9063cc963f47e3ea644e16473356c.jpg) + +![](images/f8c18c4e392f169ba34897291d1c0daa3271003515836b17fb9644297e2a7ac2.jpg) +Figure 4. The differences between ours and MAE. + +![](images/a914acf03bcf25e9b66ff053fcdce058a991494cc0fa3ee445628d335ac50bcd.jpg) +Figure 5. The differences between ours and distillation. + +Further analysis. Our framework provides deep insights into how to adapt foundation models to extreme CD-FSL tasks. It goes beyond original Masked Auto-Encoder (MAE) [3] or knowledge distillation [10, 48]. We compare our method with MAE and distillation in Fig. 4 and Fig. 5, respectively. Firstly, the original MAE tends to overfit in such few-shot tasks, leading to poor generalization. In contrast, the proposed masked cross-model unidirectional reconstruction scheme offers two advantages. For one thing, our approach explicitly injects task-relevant expert knowledge into the adaptation process, and forces the foundation model to focus on task-relevant semantic patterns. For another, our method can utilize T-ViT to reduce F-ViT's overconfidence. Secondly, simple knowledge distillation is insufficient to assist in adapting foundation models to extreme cross-domain few-shot tasks. The reason is that F-ViT and T-ViT have different perception capabilities for few-shot tasks, and forcibly aligning their features or logits may lead to the model collapsing into a trivial state. In contrast, the proposed decision graph association is a higher-order semantic structural cue with robust properties, leading to better generalization. Thirdly, we jointly train T-ViT and F-ViT, employing DGA for mutual regularization to mitigate overfitting in both models. + +# 4. Experimental Evaluation + +# 4.1. Experimental Setting + +We evaluate the effectiveness of the proposed method in two distinct settings: extreme CD-FSL and traditional CD-FSL. + +Extreme setting. In extreme CD-FSL setting, we utilize benchmarks that include ISIC [25], ChestX [26], Eu + +Table 1. Comparison with state-of-the-art methods. "B" denotes the backbone. "F" denotes the initialization with foundation model. + +
ISICChestEuroSATCropDiseaseAve.
MethodBF1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
ProtoNet [27]ViT-SDINO24.1434.0616.4319.1159.0578.7039.3973.4034.7551.31
Linear probing [49]ViT-SDINO24.9451.8818.2925.2960.3883.0251.0184.0738.6561.06
LoRA [5]ViT-SDINO24.4450.2918.2925.3560.0883.0952.8184.0738.9060.70
VPT [6]ViT-SDINO27.7953.6817.7624.8560.3481.9251.1883.5839.2661.00
RotNet [50]ViT-SDINO34.3960.4220.8428.0559.0581.9650.1684.5141.1163.73
SimCLR [51]ViT-SDINO28.7258.5618.8726.0760.5883.6752.8383.3040.2562.90
BYOL [52]ViT-SDINO28.0059.5319.9028.0958.0382.9852.0385.3639.4963.99
SetFeat [53]ViT-SDINO24.1433.5716.4319.2459.0572.5339.3966.6234.7547.99
MaskCLIP [48]ViT-SDINO28.1260.4919.0826.0858.8283.2353.9485.8139.9963.90
OursViT-SDINO37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
+ +roSAT [54], and CropDisease [55]. These datasets demonstrate substantial distributional differences compared to the natural image domain [56]. Unlike the previous practice, we re-organize these datasets into a more challenging All-way $K$ -shot configuration, where all categories are considered in the few-shot task, making it a more rigorous test of cross-domain generalization. This setup allows us to more effectively evaluate the ability of our method to handle diverse, real-world domain shifts. Specifically, we randomly sample $K$ samples from all categories in the target domain for model adaptation, and then perform model evaluation on all remaining samples. To mitigate the effects of randomness, we conduct five runs for each target domain and utilize the average result as the metric. + +Traditional setting. In the traditional CD-FSL setting, we follow the benchmarks established by Guo et al.[56] and Tseng et al.[57] for experimentation. Specifically, this involves the $N$ -way $K$ -shot $M$ -query setup, where $N \times K$ support samples are used to adapt the model, and $N \times M$ query samples are used for evaluation. In this context, $N$ represents a subset randomly sampled from the target domain categories, typically set to 5, while $K$ denotes the number of labeled samples per category, and $M$ represents the number of test samples per category, usually set to 15. For evaluation, we adopt the same protocol as prior studies [10-12, 14, 58], sampling 600 random $N$ -way $K$ -shot tasks from each target domain and using the average accuracy across these tasks as the performance metric. Additionally, previous methods [10-12, 14, 58] typically use two types of inference: inductive and transductive. Inductive inference makes predictions solely based on the support set of few-shot tasks, while transductive inference leverages query data in an unsupervised manner to assist in inference. To ensure a comprehensive comparison, we follow [10, 11, 58] and implement both inductive and transductive inference. + +Due to space constraints, more implementation details can be found in the supplementary document. + +# 4.2. Comparison with State-of-the-Art Methods + +Results on extreme CD-FSL. We compare the proposed method with state-of-the-art works to clarify our advantages. Among these comparison methods, ProtoNet [27] and SetFeat [53] represent the classical baseline in FSL. VPT [6] and LoRA [5] are representative methods that achieves knowledge transfer from the foundation model to downstream task. SimCLR [51] and BYOL [52] are two exemplary transfer learning methods. MaskCLIP [48] is a robust representation + +learning method characterized by its effective generalization potential. We re-implemented these methods under the same settings for comparison in extreme CD-FSL tasks. The experimental results are as shown in Table 1. Overall, the proposed method achieves the best performance compared to these state-of-the-art approaches. Specifically, our method consistently outperforms methods based on transfer learning (e.g., VPT, LoRA, SimCLR, MaskCLIP, etc.). This indicates the strong generalization capability of the proposed method in handling extreme cross-domain scenarios. In addition, compared to ProtoNet [27] and SetFeat [53], our method demonstrates significant performance advantages. This suggests that the robustness of advanced few-shot learners is limited without the introduction of explicit cross-domain learning mechanisms. The above results demonstrate the advantages of our method. + +Results on traditional CD-FSL. Table 2 presents the results for both the 5-way 1-shot and 5-way 5-shot traditional CD-FSL tasks. Overall, our method outperforms advanced comparison methods, further validating its effectiveness. + +# 4.3. Ablation Study + +Comparing our method with baselines. We compare our method with F-ViT and T-ViT baselines to highlight our advantages. Among them, F-ViT refers to the standard fine-tuning of the foundation model. We experimentally found that standard fine-tuning is relatively the best approach. Therefore, we consider F-ViT as our strong baseline. T-ViT refers to a model that learns from scratch using few-shot data from the target domain. We conduct experiments under 1-shot and 5-shot settings across four datasets, initializing F-ViT with DINO unless otherwise specified. The experimental results are shown in Table 3. The proposed method exhibits significant performance advantages in average accuracy across four datasets compared to the T-ViT and F-ViT baselines. Furthermore, this experiment shows that relying solely on T-ViT is insufficient for effectively addressing such extreme cross-domain few-shot tasks. Nevertheless, due to the task-specific intrinsic properties of T-ViT, it still holds some decision-making instructions. Therefore, our framework, which leverages T-ViT to enhance the generalization of F-ViT, possesses practical application value. + +Evaluating the effectiveness of MCR and DGA. Our framework consists of two core components: Masked Cross-model unidirectional Reconstruction (MCR) and Decision Graph Association (DGA). We conduct four experimental cases to evaluate the contributions of them. Case 1: The pro + +Table 2. Comparison with state-of-the-art methods on traditional 5-way CD-FSL. † stands for using query data in an unsupervised manner to assist in inference. “B” denotes the backbone. “F” denotes the initialization with foundation model. “R10” denotes ResNet-10. + +
MethodBFShotCUBCarsPlacesPlantaeChestISICEuroSATCropDiseaseAve.
Dara [34]R10-1-shot52.7035.2551.2542.0822.9236.4267.4280.7448.60
GNN+FAP [46]R10-1-shot46.9134.6954.4136.7322.3635.6362.9669.9745.46
IM-DCL†[46]ViT-B-1-shot----22.6838.2074.6885.98-
PMF [59]ViT-SDINO1-shot78.1337.2471.1153.6021.7330.3670.7480.7955.46
MEM [60]ViT-SDINO1-shot----22.7632.9768.1181.11-
StyleAdv†[11]ViT-SDINO1-shot84.0140.4872.6455.5222.9233.0572.1581.2257.75
StyleAdv-FT†[11]ViT-SDINO1-shot84.0140.4872.6455.5222.9233.9974.9384.1158.57
FLoR†[33]ViT-SDINO1-shot85.4043.4274.6952.2923.2635.4973.0983.5558.90
AttnTemp [44]ViT-SDINO1-shot----23.6338.0575.0984.78-
AttnTemp†[44]ViT-SDINO1-shot----23.9640.1377.4087.58-
CD-CLS [43]ViT-SDINO1-shot----23.3935.5674.9784.54-
CD-CLS†[43]ViT-SDINO1-shot----23.8837.2078.4187.39-
OursViT-SDINO1-shot85.2644.1374.3057.5423.1334.4872.0882.1359.13
Ours†ViT-SDINO1-shot94.1946.4582.4163.8423.1235.0282.3891.3664.84
Dara [34]R10-5-shot77.5158.4472.1565.4027.5456.2885.8495.3267.31
GNN+FAP [46]R10-5-shot67.6650.2074.9854.5425.3147.6082.5291.7961.83
IM-DCL†[46]ViT-B-5-shot----27.9853.8987.7695.93-
PMF [59]ViT-SDINO5-shot----27.2750.1285.9892.96-
MEM [60]ViT-SDINO5-shot----26.6747.3886.4993.74-
StyleAdv†[11]ViT-SDINO5-shot95.8261.7388.3375.5526.9747.7388.5794.8572.44
StyleAdv-FT†[11]ViT-SDINO5-shot95.8266.0288.3378.0126.9751.2390.1295.9974.06
FLoR†[33]ViT-SDINO5-shot96.5368.4489.4876.2227.0253.0690.7596.4774.75
AttnTemp [44]ViT-SDINO5-shot----28.0354.9190.8296.66-
AttnTemp†[44]ViT-SDINO5-shot----28.4155.2291.3496.74-
CD-CLS [43]ViT-SDINO5-shot----27.6654.6991.5396.27-
CD-CLS†[43]ViT-SDINO5-shot----28.2555.6691.6896.62-
OursViT-SDINO5-shot97.5574.7689.1581.7326.2550.5589.9995.8175.72
Ours†ViT-SDINO5-shot98.3978.0091.4684.7426.1151.1992.7997.2677.49
+ +Table 3. Ablation study. $\checkmark$ indicates that the component is used. + +
ISICChestEuroSATCropDiseaseAve.
Method1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
T-ViT baseline26.7153.6821.1526.1638.3449.3324.6655.4927.7146.16
F-ViT baseline26.3560.9520.0526.4160.2482.9852.1886.4839.7064.20
Ours37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
MCRDGA1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
XX26.3560.9520.0526.4160.2482.9852.1886.4839.7064.20
✓X32.2762.4921.2827.7260.2583.5354.0586.5241.9665.06
X✓33.1162.2821.1627.3260.2983.9653.5886.4242.0364.99
✓✓37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
+ +posed method without the MCR and DGA modules, equivalent to a standard fine-tuning baseline. Case 2: The proposed method using only the MCR module on top of the baseline. Case 3: The proposed method using only the DGA module on top of the baseline. Case 4: The proposed method using both MCR and DGA modules simultaneously. As can be seen in Table 3, compared to the stand fine-tuning baseline ("F-ViT baseline"), utilizing either the MCR or the DGA alone can achieve certain performance gains. For example, across the four datasets, the MCR module achieves improvements of $2.26\%$ and $0.86\%$ under the 1-shot and 5-shot settings respectively. Similarly, the DGA module achieves improvements of $2.33\%$ and $0.79\%$ under the 1-shot and 5-shot settings respectively. Notably, on the ISIC dataset under 1-shot setting, MCR and DGA achieve performance gains of $5.92\%$ and $6.76\%$ respectively. Moreover, integrating these two components into a unified framework can unleash greater potential. For instance, compared to using a single component, simultaneously using them further increases the average results by $1.32\%$ in the 1-shot setting. These results confirm the benefits of the proposed approaches. + +# 4.4. Further Analysis + +Comparing our method with MAE variants. We compare the proposed method with the standard MAE, which + +involves adding a masked reconstruction loss on top of the fine-tuning F-ViT baseline. Additionally, our method reconstructs the masked features in F-ViT using T-ViT as the target, aiming to enhance F-ViT's focus on task-specific semantic patterns. Our approach is essentially a unidirectional cross-model masked reconstruction. To demonstrate the benefits of this design, we also compare it with a bidirectional MAE variant, where the masked parts of F-ViT and T-ViT are reconstructed using T-ViT and F-ViT as targets, respectively. As show in Table 4, our method shows clear performance advantages over both of these MAE variants. Firstly, the original MAE tends to over-fit in such FSL tasks, leading to poor generalization. In contrast, our masked cross-model unidirectional reconstruction offers two advantages. For one thing, our approach explicitly injects task-specific model knowledge into the foundation model's adaptation process. For another, our method utilizes T-ViT to relax F-ViT's overconfidence, mitigating over-fitting to some extent. Secondly, the optimization objective of bidirectional MAE is ambiguous. In contrast, our unidirectional cross-modality masked reconstruction approach is goal-oriented, which has resulted in improved performance. + +Comparing our method with distillation or fusion. We compare the proposed method with knowledge distillation and model fusion approaches to clarify our previous claims. + +Table 4. Comparing ours with MAE and distillation based methods. + +
ISICChestEuroSATCropDiseaseAve.
Method1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
MAE35.6462.3921.1927.3356.7877.5152.3885.7341.4963.24
Bidirectional MAE30.3862.0919.3928.6759.8983.0652.5985.0240.5664.71
Feature-level Distillation27.7662.3122.0127.2656.2879.8153.5886.1239.9063.87
Logits-level Distillation27.5361.5420.6725.9960.8883.1854.8786.5140.9864.30
Cross-attention Fusion29.3462.1920.7826.7358.7482.4053.7185.6440.6464.24
Decision Fusion25.4456.9619.8525.5058.9482.7636.2961.9835.1356.80
Ours37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
+ +Table 5. Comparison with SoTA methods on mini-ImageNet. + +
MethodBackbone5-way 1-shot5-way 5-shot
CAML [47]ViT-B96.2098.60
SgVA-CLIP [42]ViT-B97.9598.72
PMF [59]ViT-S93.1098.00
OursViT-S94.6898.86
+ +Table 6. Comparison with SoTA methods on tiered-ImageNet. + +
MethodBackbone5-way 1-shot5-way 5-shot
FewTURE [61]ViT-S72.9686.43
HCT [62]ViT-S79.6791.72
OursViT-S82.7893.78
+ +Table 7. Comparison with SoTA methods on CIFAR-FS. + +
MethodBackbone5-way 1-shot5-way 5-shot
HCT [62]ViT-S78.8990.50
PMF [59]ViT-S81.1092.50
CAML [47]ViT-B70.8085.50
OursViT-S87.6496.03
+ +Specifically, we conduct variant experiments atop the baseline. Variant 1: Feature-level Distillation (FD), aligns the features of the foundation model and the task-specific model directly. Variant 2: Logits-level Distillation (LD), aligns the class prediction scores of the two models directly. Variant 3: Cross-attention Fusion (CF), which involves integrating the features of the foundation model and task-specific model through cross-attention mechanisms. Variant 4: Decision Fusion (DF), directly integrating the baseline model and task-specific model for decision-making. The results are presented in Table 4. It can be observed that our method consistently outperforms these variants. For instance, across four datasets, our method exhibits improvements ranging from $2.37\%$ to $8.22\%$ in 1-shot setting and from $1.22\%$ to $8.77\%$ in 5-shot setting, on average. The insights it brings encompass two aspects. Firstly, our framework embodies sophisticated absorption adaptation mechanisms, which are neither simple knowledge distillation nor plain model fusion. Secondly, conventional technical approaches such as knowledge distillation, model fusion, are insufficient to tackle complex extreme CD-FSL. In contrast, our meticulously crafted framework can provide valuable technical contributions. + +Extending the proposed method to in-domain FSL. We further extend our method to in-domain FSL. We follow common experimental setups from previous works [59, 61, 62] to conduct in-domain FSL experiments on three representative benchmarks: mini-[ImageNet [63], tiered-[ImageNet [64] and CIFAR-FS [24]. To ensure fairness, we maintain the same data settings as previous works [59, 61, 62]. As shown in Table 5, Table 6, and Table 7, the proposed method excels in standard in-domain FSL as well. This reinforces the robustness and versatility of our approach. + +Table 8. Evaluating the efficiency of our method. + +
TrainingInference
MethodTime (S)Memory (M)Time (S)Memory (M)Accuracy (%)
F-ViT837.40347514.10251639.70
Ours939.50642614.10251643.35
+ +Efficiency. The efficiency of our method is evaluated in Table 8, with all experiments conducted in the All-way 1-shot setting on a 3090 GPU. The accuracy reported in Table 8 represents the average results across four datasets: ISIC, Chest, EuroSAT, and CropDisease. The training time refers to the duration of each training epoch, while the testing time indicates the total time required to process all testing samples. As shown, our method introduces minimal overhead during training, matches the F-ViT baseline in terms of inference efficiency, and significantly outperforms it in accuracy. + +# 5. Conclusions + +In this study, we propose a novel absorption adaptation learning framework, where a foundation model is guided by an expert model trained from scratch on targeted data. Central to our method are two innovative approaches: masked cross-model unidirectional reconstruction and decision graph association. These strategies enable the foundation model to absorb task-relevant knowledge from the expert model, both at the intermediate feature and final decision levels, thereby mitigating performance drops caused by domain gaps and limited annotations. Both qualitative and quantitative analyses substantiate the effectiveness of our approach. + +Limitations. First, our framework employs two parallel ViTs for target domain adaptation, leading to increased computation and storage during training. Second, while it shows promising results, there is still room for improvement, particularly in extreme scenarios. Third, being tailored for ViT models, our approach inherits their limitations. + +Acknowledgment. This work was supported in part by the National Natural Science Foundation of China under Grand 62372379, Grant 62472359, and Grant 62472350; in part by the Xi'an's Key Industrial Chain Core Technology Breakthrough Project: AI Core Technology Breakthrough under Grand 23ZDCYJSGG0003-2023; in part by Innovation Foundation for Doctor Dissertation of Northwestern Polytechnical University under Grant CX2024017; in part by National Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing fundation under Grant TJ-04-23-04. + +# References + +[1] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 3 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 3 +[3] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 5 +[4] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3 +[5] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 1, 2, 3, 6 +[6] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3, 6 +[7] Samyadeep Basu, Shell Hu, Daniela Massiceti, and Soheil Feizi. Strong baselines for parameter-efficient few-shot fine-tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 11024-11031, 2024. 2, 3 +[8] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23681-23690, 2024. 3 +[9] Shuai Shao, Yu Bai, Yan Wang, Baodi Liu, and Bin Liu. Collaborative consortium of foundation models for open-world few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4740-4747, 2024. 1, 3 +[10] Fei Zhou, Peng Wang, Lei Zhang, Wei Wei, and Yanning Zhang. Revisiting prototypical network for cross + +domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20061-20070, 2023. 2, 3, 5, 6 +[11] Yuqian Fu, Yu Xie, Yanwei Fu, and Yu-Gang Jiang. Styleadv: Meta style adversarial training for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24575-24584, 2023. 3, 6, 7 +[12] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1075–1081. International Joint Conferences on Artificial Intelligence Organization, 8 2021. Main Track. 3, 6 +[13] Yi Wu, Ziqiang Li, Chaoyue Wang, Heliang Zheng, Shanshan Zhao, Bin Li, and Dacheng Tao. Domain remodulation for few-shot generative domain adaptation. volume 36, 2024. +[14] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 3, 6 +[15] Pin-Yu Chen. Model reprogramming: Resource-efficient cross-domain machine learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 22584-22591, 2024. 3 +[16] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple domains for few-shot classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9526-9535, 2021. +[17] Yuqian Fu, Yanwei Fu, and Yu-Gang Jiang. Meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. In Proceedings of the 29th ACM international conference on multimedia, pages 5326-5334, 2021. +[18] Jiamei Sun, Sebastian Lapuschkin, Wojciech Samek, Yunqing Zhao, Ngai-Man Cheung, and Alexander Binder. Explanation-guided training for cross-domain few-shot classification. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7609-7616. IEEE, 2021. 2 +[19] Baoquan Zhang, Chuyao Luo, Demin Yu, Xutao Li, Huiwei Lin, Yunming Ye, and Bowen Zhang. Metadata: Meta-learning with conditional diffusion for few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 16687-16695, 2024. 2 +[20] Ilia Sucholutsky and Thomas L Griffiths. Alignment with human representations supports robust few-shot learning. In Proceedings of the 37th International + +Conference on Neural Information Processing Systems, pages 73464-73479, 2023. +[21] Meiqi Sun, Zhonghan Zhao, Wenhao Chai, Hanjun Luo, Shidong Cao, Yanting Zhang, Jenq-Neng Hwang, and Gaoang Wang. Uniap: Towards universal animal perception in vision via few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5008-5016, 2024. +[22] Xingyu Zhu, Shuo Wang, Jinda Lu, Yanbin Hao, Haifeng Liu, and Xiangnan He. Boosting few-shot learning via attentive feature regularization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 7793-7801, 2024. 2 +[23] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International conference on learning representations, 2016. 2 +[24] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In International Conference on Learning Representations, 2019. 2, 8 +[25] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multisource dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. 2, 5 +[26] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 2, 5 +[27] Jake Snell, Kevin Swersky, and Richard Zemel. Phototypical networks for few-shot learning. In Advances in Neural Information Processing Systems, pages 4077-4087, 2017. 2, 6 +[28] Yurong Guo, Ruoyi Du, Yuan Dong, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Task-aware adaptive learning for cross-domain few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1590-1599, 2023. 3 +[29] Rashindrie Perera and Saman Halgamuge. Discriminative sample-guided and parameter-efficient feature space adaptation for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23794-23804, 2024. +[30] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Cross-domain few-shot learning with task-specific adapters. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7161-7170, 2022. + +[31] Hao ZHENG, Runqi Wang, Jianzhuang Liu, and Asako Kanezaki. Cross-level distillation and feature denoising for cross-domain few-shot classification. In The Eleventh International Conference on Learning Representations. 3 +[32] Yongjin Yang, Taehyeon Kim, and Se-Young Yun. Leveraging normalization layer in adapters with progressive learning and adaptive distillation for cross-domain few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 16370–16378, 2024. 3 +[33] Yixiong Zou, Yicong Liu, Yiman Hu, Yuhua Li, and Ruixuan Li. Flatten long-range loss landscapes for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23575-23584, 2024. 3, 7 +[34] Yifan Zhao, Tong Zhang, Jia Li, and Yonghong Tian. Dual adaptive representation alignment for cross-domain few-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11720-11732, 2023. 3, 7 +[35] Wenjian Wang, Lijuan Duan, Yuxi Wang, Junsong Fan, and Zhaoxiang Zhang. Mmt: cross domain few-shot learning via meta-memory transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3 +[36] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3 +[37] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19325-19337, 2023. +[38] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15211-15222, 2023. +[39] Lin Song, Ruoyi Xue, Hang Wang, Hongbin Sun, Yixiao Ge, Ying Shan, et al. Meta-adapter: An online few-shot learner for vision-language model. Advances in Neural Information Processing Systems, 36:55361-55374, 2023. +[40] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024. +[41] Jiamin Wu, Xin Liu, Xiaotian Yin, Tianzhu Zhang, and Yongdong Zhang. Task-adaptive prompted transformer + +for cross-domain few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 6012-6020, 2024. 3 +[42] Fang Peng, Xiaoshan Yang, Linhui Xiao, Yaowei Wang, and Changsheng Xu. Sgva-clip: Semantic-guided visual adapting of vision-language models for few-shot image classification. IEEE Transactions on Multimedia, 26:3469-3480, 2023. 3, 8 +[43] Yixiong Zou, Shuai Yi, Yuhua Li, and Ruixuan Li. A closer look at thecls token for cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:85523-85545, 2024. 3, 7 +[44] Yixiong Zou, Ran Ma, Yuhua Li, and Ruixuan Li. Attention temperature matters in vit-based cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:116332-116354, 2024. 3, 7 +[45] Huali Xu, Li Liu, Shuaifeng Zhi, Shaojing Fu, Zhuo Su, Ming-Ming Cheng, and Yongxiang Liu. Enhancing information maximization with distance-aware contrastive learning for source-free cross-domain few-shot learning. IEEE Transactions on Image Processing, 33:2058-2073, 2024. 3 +[46] Tiange Zhang, Qing Cai, Feng Gao, Lin Qi, and Junyu Dong. Exploring cross-domain few-shot classification via frequency-aware prompting. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 5490–5498, 2024. 3, 7 +[47] Christopher Fifty, Dennis Duan, Ronald Guenther Junkins, Ehsan Amid, Jure Leskovec, Christopher Re, and Sebastian Thrun. Context-aware meta-learning. In The Twelfth International Conference on Learning Representations. 3, 8 +[48] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995-11005, 2023. 5, 6 +[49] Yonatan Belinkov. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207-219, 2022. 6 +[50] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR 2018, 2018. 6 +[51] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6 +[52] Jean-Bastien Grill, Florian Strub, Florent Altché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, + +Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6 +[53] Arman Afrasiyabi, Hugo Larochelle, Jean-François Lalonde, and Christian Gagné. Matching feature sets for few-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9014-9024, 2022. 6 +[54] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6 +[55] Punam Bedi, Pushkar Gole, and Sumit Kumar Agarwal. 18 using deep learning for image-based plant disease detection. _Internet Of things and machine learning in agriculture_, pages 369–402, 2021. 6 +[56] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In European conference on computer vision, pages 124-141. Springer, 2020. 6 +[57] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In International Conference on Learning Representations, 2019. 6 +[58] Pan Li, Shaogang Gong, Chengjie Wang, and Yanwei Fu. Ranking distance calibration for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9099-9108, 2022. 6 +[59] Nenad Tomasev, Ioana Bica, Brian McWilliams, Lars Holger Buesing, Razvan Pascanu, Charles Blundell, and Jovana Mitrovic. Pushing the limits of self-supervised resnets: Can we outperform supervised learning without labels onImagenet? In First Workshop on Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022. 7, 8 +[60] Reece Walsh, Islam Osman, and Mohamed S Shehata. Masked embedding modeling with rapid domain adjustment for few-shot image classification. IEEE Transactions on Image Processing, 32:4907-4920, 2023. 7 +[61] Markus Hiller, Rongkai Ma, Mehrtash Harandi, and Tom Drummond. Rethinking generalization in few-shot classification. Advances in Neural Information Processing Systems, 35:3582-3595, 2022. 8 +[62] Yangji He, Weihan Liang, Dongyang Zhao, Hong-Yu Zhou, Weifeng Ge, Yizhou Yu, and Wenqiang Zhang. Attribute surrogates learning and spectral tokens pooling in transformers for few-shot learning. In Proceed- + +ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9119-9129, 2022. 8 + +[63] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in neural information processing systems, pages 3630-3638, 2016. 8 +[64] Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. In International Conference on Learning Representations, 2019. 8 \ No newline at end of file diff --git a/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/images.zip b/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e8508feef91089b85999d3f15bba9f33ef293c1 --- /dev/null +++ b/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd1b47b6a9091402798cf9f34bb51ff0872a06b25ecda0931caf67c4c3065c7c +size 648279 diff --git a/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/layout.json b/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4552183755968699466da7bd2160f8321ceff531 --- /dev/null +++ b/2025/Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning/layout.json @@ -0,0 +1,8815 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 102, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 102, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 541, + 140 + ], + "type": "text", + "content": "Towards Effective Foundation Model Adaptation for Extreme Cross-Domain Few-Shot Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "spans": [ + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "text", + "content": "Fei Zhou" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "text", + "content": " Peng Wang" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "text", + "content": " Lei Zhang" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "text", + "content": " Wei Wei" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "text", + "content": " Chen Ding" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "text", + "content": " Guosheng Lin" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "text", + "content": " Yanning Zhang" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 160, + 572, + 190 + ], + "type": "text", + "content": " School of Computer Science, Northwestern Polytechnical University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 190, + 576, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 190, + 576, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 190, + 576, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 190, + 576, + 232 + ], + "type": "text", + "content": " School of Computer Science and Engineering, University of Electronic Science and Technology of China \n" + }, + { + "bbox": [ + 61, + 190, + 576, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 190, + 576, + 232 + ], + "type": "text", + "content": " School of Computer Science, Xi'an University of Posts & Telecommunications \n" + }, + { + "bbox": [ + 61, + 190, + 576, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 190, + 576, + 232 + ], + "type": "text", + "content": " College of Computing and Data Science, Nanyang Technological University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 125, + 234, + 514, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 234, + 514, + 259 + ], + "spans": [ + { + "bbox": [ + 125, + 234, + 514, + 259 + ], + "type": "text", + "content": "zhoufei@mail.nwpu.edu.cn wangpeng8619@gmail.com dingchen@xupt.edu.cn gslin@ntu.edu.sg {nwpuzhanglei,weiweinwpu,ynzhang}@nwpu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 313, + 297, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 313, + 297, + 624 + ], + "spans": [ + { + "bbox": [ + 54, + 313, + 297, + 624 + ], + "type": "text", + "content": "Large-scale pre-trained foundation models have demonstrated remarkable generalization capabilities across diverse computer vision tasks through fine-tuning. However, existing fine-tuning approaches often encounter challenges in extreme cross-domain few-shot learning scenarios, primarily due to the significant domain shift between pre-training data and target tasks, as well as the scarcity of annotated target samples. To mitigate this issue, we propose a novel absorption adaptation learning framework which meticulously regularizes the fine-tuning procedure of foundation model using an expert model with the same architecture but trained from scratch on the targeted data in two aspects. On one hand, we first design a masked cross-model unidirectional reconstruction scheme, which forces the foundation model to recover the intermediate feature of the expert model in a randomly masked manner. On the other hand, a decision graph association loss is developed to encourage the consistency of token similarity matrix between these two models. By doing these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels are appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation. Sufficient experiments with further observations and analyses underpin our observation and argument. The code is available at https://github.com/NWPUZhoufei/FMA." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 643, + 137, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 137, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 137, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "content": "Recent breakthroughs in foundation models [1-4] have exhibited compelling generalization performance across diverse computer vision tasks. Their efficacy is fundamentally" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 325, + 284, + 545, + 388 + ], + "blocks": [ + { + "bbox": [ + 325, + 284, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 325, + 284, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 325, + 284, + 545, + 388 + ], + "type": "image", + "image_path": "b94f1880682c8b5e68b1d3e7b6a3b064d82a67a16c6759cc760de808fc868d3b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 325, + 418, + 545, + 521 + ], + "blocks": [ + { + "bbox": [ + 324, + 389, + 544, + 418 + ], + "lines": [ + { + "bbox": [ + 324, + 389, + 544, + 418 + ], + "spans": [ + { + "bbox": [ + 324, + 389, + 544, + 418 + ], + "type": "text", + "content": "(a) Comparing the performance of foundation models on both in-domain (e.g., mini-ImageNet, CIFAR-FS) and extreme cross-domain (e.g., ISIC, Chest) FSL." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 418, + 545, + 521 + ], + "lines": [ + { + "bbox": [ + 325, + 418, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 325, + 418, + 545, + 521 + ], + "type": "image", + "image_path": "bbd94ef97a28081471bf0d34f57b3301712df4c64faf41a2465c571ce95a72ac.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 523, + 543, + 542 + ], + "lines": [ + { + "bbox": [ + 324, + 523, + 543, + 542 + ], + "spans": [ + { + "bbox": [ + 324, + 523, + 543, + 542 + ], + "type": "text", + "content": "(b) Comparing the performance of ours with existing adaptation approaches on extreme cross-domain (e.g., ISIC) FSL." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 544, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 675 + ], + "type": "text", + "content": "Figure 1. Performance of foundation models on extreme CD-FSL. rooted in pre-training massive-scale learnable parameters on extensive datasets, endowing these models with rich semantic knowledge adaptable to multiple domains—especially those aligned with the pre-training distribution. Even straightforward fine-tuning of these models has demonstrated superior performance in downstream tasks. Consequently, there has been a burgeoning interest in developing advanced fine-tuning methodologies [5-9], aiming to enhance the adaptability and efficiency of foundation models in specialized scenarios." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Insight and motivation. While effective in many computer vision tasks, existing fine-tuning strategies encounter significant limitations in extreme cross-domain few-shot" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 13, + 116, + 43 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 13, + 116, + 43 + ], + "spans": [ + { + "bbox": [ + 71, + 13, + 116, + 43 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "spans": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "spans": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "text", + "content": "4582" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 419 + ], + "type": "text", + "content": "learning (CD-FSL) scenarios [10-18]. These tasks are characterized by substantial domain discrepancy from the pretraining data of foundation models, compounded by the scarcity of annotated samples [7, 19-22]. As visualized in Fig. 1(a), foundation models (e.g., DINO[2]) achieve near-saturated few-shot performance on in-domain benchmarks like mini-[ImageNet] [23] and CIFAR-FS [24], which share distributional similarity with pre-training data. However, when applied to extreme cross-domain FSL tasks such as ISIC [25] (medical imaging) and Chest [26] (radiology), their accuracy degrades sharply. We further examine several existing fine-tuning-based adaptation methods for foundation models and compare their performance experimentally, as shown in Fig.1(b). Among these methods, prototype inference (PI\") [27], which directly utilizes the features extracted by the foundation model for classification, exhibits the worst performance. Linear probing (LP\"), which fine-tunes a linear mapping for the features produced by the foundation model in the target domain, shows a slight improvement over prototype inference. Notably, fine-tuning foundation models using the LoRA scheme[5] performs even worse than linear probing. This result suggests that fine-tuning only a small subset of parameters is ineffective, as it fails to mitigate the model's excessive memorization of the source domain. In contrast, standard fine-tuning (\"FT\"), which updates all parameters of the foundation model, proves to be the most effective baseline. However, its performance remains suboptimal compared to the results achieved by foundation models in in-domain scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 423, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 543 + ], + "type": "text", + "content": "According to the empirical observation above, we find that most existing fine-tuning strategies struggle to effectively address extreme CD-FSL tasks. This limitation stems from the substantial domain gap between the target tasks and the pre-training data of foundation models, coupled with the scarcity of annotated examples in the few-shot setting. Specifically, bridging the domain gap may necessitate updating all parameters of the foundation model, yet the limited supervision provided by the few annotated examples is insufficient to achieve effective adaptation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 297, + 713 + ], + "type": "text", + "content": "To address this challenge, a potential solution is to incorporate pure task-specific knowledge (e.g., a model trained from scratch using the limited annotated samples from the target task) to regularize the fine-tuning process of the foundation model. To validate this perspective, we present visualization results in Fig. 2. Specifically, \"F-ViT\" represents the pre-trained foundation model, while \"T-ViT\" denotes a task-specific expert model with the same architecture as \"F-ViT\" but trained from scratch on the target task's training examples. The following key observations can be made: 1) Although \"F-ViT\" provides high-quality initial features without severe noise corruption, it fails to focus on the primary semantic regions. 2) \"T-ViT\" captures patterns closely related to the target task but introduces substantial noise," + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 319, + 71, + 375, + 125 + ], + "blocks": [ + { + "bbox": [ + 319, + 71, + 375, + 125 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 375, + 125 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 375, + 125 + ], + "type": "image", + "image_path": "0d76aa9e566790347db0322f0dcba948a0861b26419a9b75d9e7cd7eab2f9847.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 377, + 71, + 433, + 125 + ], + "blocks": [ + { + "bbox": [ + 377, + 71, + 433, + 125 + ], + "lines": [ + { + "bbox": [ + 377, + 71, + 433, + 125 + ], + "spans": [ + { + "bbox": [ + 377, + 71, + 433, + 125 + ], + "type": "image", + "image_path": "8ac476ca11df4d52bed52b92e5002aebf2da2d880cdf181feacc966d9f0d8e7a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 436, + 70, + 492, + 125 + ], + "blocks": [ + { + "bbox": [ + 436, + 70, + 492, + 125 + ], + "lines": [ + { + "bbox": [ + 436, + 70, + 492, + 125 + ], + "spans": [ + { + "bbox": [ + 436, + 70, + 492, + 125 + ], + "type": "image", + "image_path": "9dd8aefff41de6e467daa435a82782f5c0275a3807b4f564856d2ddaf6ba577a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 494, + 70, + 551, + 125 + ], + "blocks": [ + { + "bbox": [ + 494, + 70, + 551, + 125 + ], + "lines": [ + { + "bbox": [ + 494, + 70, + 551, + 125 + ], + "spans": [ + { + "bbox": [ + 494, + 70, + 551, + 125 + ], + "type": "image", + "image_path": "da1bd50c248a0f473357e5d36ea0b676dd6daf7614400cfd911012b670f4f858.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 137, + 375, + 194 + ], + "blocks": [ + { + "bbox": [ + 331, + 128, + 362, + 137 + ], + "lines": [ + { + "bbox": [ + 331, + 128, + 362, + 137 + ], + "spans": [ + { + "bbox": [ + 331, + 128, + 362, + 137 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 137, + 375, + 194 + ], + "lines": [ + { + "bbox": [ + 318, + 137, + 375, + 194 + ], + "spans": [ + { + "bbox": [ + 318, + 137, + 375, + 194 + ], + "type": "image", + "image_path": "4406a74a209e41d8a3894c2a2c150eab75fde1b018b4905ffdba30731a115c32.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 196, + 362, + 205 + ], + "lines": [ + { + "bbox": [ + 331, + 196, + 362, + 205 + ], + "spans": [ + { + "bbox": [ + 331, + 196, + 362, + 205 + ], + "type": "text", + "content": "(e) Image" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 209, + 555, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 275 + ], + "type": "text", + "content": "Figure 2. Visualization of the class activation mapping (CAM) computed by different models. Results of \"F-ViT\" and \"T-ViT\" in (b) and (c) show complementarity, while \"Ours\" integrates the knowledge of both \"F-ViT\" and \"T-ViT\", resulting in a better response. In terms of average accuracy, \"T-ViT\" scores " + }, + { + "bbox": [ + 313, + 209, + 555, + 275 + ], + "type": "inline_equation", + "content": "27.71\\%" + }, + { + "bbox": [ + 313, + 209, + 555, + 275 + ], + "type": "text", + "content": ", \"F-ViT\" scores " + }, + { + "bbox": [ + 313, + 209, + 555, + 275 + ], + "type": "inline_equation", + "content": "39.70\\%" + }, + { + "bbox": [ + 313, + 209, + 555, + 275 + ], + "type": "text", + "content": ", and \"Ours\" scores " + }, + { + "bbox": [ + 313, + 209, + 555, + 275 + ], + "type": "inline_equation", + "content": "43.35\\%" + }, + { + "bbox": [ + 313, + 209, + 555, + 275 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 377, + 137, + 434, + 194 + ], + "blocks": [ + { + "bbox": [ + 390, + 128, + 421, + 137 + ], + "lines": [ + { + "bbox": [ + 390, + 128, + 421, + 137 + ], + "spans": [ + { + "bbox": [ + 390, + 128, + 421, + 137 + ], + "type": "text", + "content": "(b) F-ViT" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 377, + 137, + 434, + 194 + ], + "lines": [ + { + "bbox": [ + 377, + 137, + 434, + 194 + ], + "spans": [ + { + "bbox": [ + 377, + 137, + 434, + 194 + ], + "type": "image", + "image_path": "d6c57677ea8f7874e17d5218eede1f0ea55a3448466900363c683a02e23ffecf.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 195, + 421, + 204 + ], + "lines": [ + { + "bbox": [ + 390, + 195, + 421, + 204 + ], + "spans": [ + { + "bbox": [ + 390, + 195, + 421, + 204 + ], + "type": "text", + "content": "(f) F-ViT" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 435, + 137, + 492, + 194 + ], + "blocks": [ + { + "bbox": [ + 449, + 128, + 479, + 137 + ], + "lines": [ + { + "bbox": [ + 449, + 128, + 479, + 137 + ], + "spans": [ + { + "bbox": [ + 449, + 128, + 479, + 137 + ], + "type": "text", + "content": "(c) T-ViT" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 435, + 137, + 492, + 194 + ], + "lines": [ + { + "bbox": [ + 435, + 137, + 492, + 194 + ], + "spans": [ + { + "bbox": [ + 435, + 137, + 492, + 194 + ], + "type": "image", + "image_path": "214d3528eb697b28eaf4f24877a5b24476dc4c529f57dd70817b6c127ed89107.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 448, + 195, + 479, + 205 + ], + "lines": [ + { + "bbox": [ + 448, + 195, + 479, + 205 + ], + "spans": [ + { + "bbox": [ + 448, + 195, + 479, + 205 + ], + "type": "text", + "content": "(g) T-ViT" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 494, + 137, + 551, + 194 + ], + "blocks": [ + { + "bbox": [ + 509, + 128, + 536, + 137 + ], + "lines": [ + { + "bbox": [ + 509, + 128, + 536, + 137 + ], + "spans": [ + { + "bbox": [ + 509, + 128, + 536, + 137 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 494, + 137, + 551, + 194 + ], + "lines": [ + { + "bbox": [ + 494, + 137, + 551, + 194 + ], + "spans": [ + { + "bbox": [ + 494, + 137, + 551, + 194 + ], + "type": "image", + "image_path": "32ced301ecfa761038f5251e808b612c0331fbc65c149ad61c8e4fbfefcca360.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 195, + 536, + 204 + ], + "lines": [ + { + "bbox": [ + 509, + 195, + 536, + 204 + ], + "spans": [ + { + "bbox": [ + 509, + 195, + 536, + 204 + ], + "type": "text", + "content": "(h) Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 280, + 555, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 352 + ], + "type": "text", + "content": "leading to poor generalization. 3) More importantly, the highlighted regions in \"F-ViT\" and \"T-ViT\" exhibit clear complementarity. These findings suggest that adapting the foundation model to extreme CD-FSL tasks can be effectively achieved by leveraging task-relevant knowledge from the expert model during fine-tuning." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 354, + 556, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 545 + ], + "type": "text", + "content": "Framework. Following this idea, we present a novel absorption adaptation learning framework, which attempt to enhance the extreme cross-domain generalization capacity of a pre-trained vision foundation model (e.g., \"F-ViT\") through absorbing some task-relevant knowledge from an expert model (e.g., \"T-ViT\") during fine-tuning. However, it is difficult to collaborate the learning procedure of these two models, due to their huge difference in feature representation capacity. In a specific, the adequate training of a massive amount of parameters on large-scale pre-training data empowers the foundation model to exhibit powerful and overconfidence semantic representational capabilities, which will inevitably overshadow the task-relevant knowledge exploited by the expert model. Besides, the extensive feature noise corruption produced by the expert model may mislead the fine-tuning of the foundation model." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "To address this challenge, we propose two innovative strategies: masked cross-model unidirectional reconstruction and decision graph association. The former learns to reconstruct the token features of the expert model using the masked token features of the foundation model with a light-weight decoder. The advantage of this strategy is twofold. On one hand, the reconstruction explicitly injects task-relevant knowledge from the expert model to the foundation model, thus alleviating its overconfidence. On the other hand, the masked input scheme encourages the semantic relationship between local patterns. The latter first defines the decision graphs using the similarity matrix computed by the class token and the patch tokens for each model, and then introduce loss to constrain their consistency. By doing" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "spans": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "text", + "content": "4583" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels can be appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation. Extensive results verify our approach and claims." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 298, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 298, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 298, + 255 + ], + "type": "text", + "content": "Contributions. The contributions of this work as follows. 1) We propose a novel absorption adaptation learning framework that meticulously regularizes the fine-tuning process of the foundation model using an expert model. 2) We introduce two innovative strategies, including masked cross-model unidirectional reconstruction and decision graph association, to empower the foundation model effectively absorb the task-relevant semantic knowledge from the exert model. 3) We achieve new state-of-the-art results on multiple benchmarks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 266, + 142, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 142, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 142, + 277 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 281, + 298, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 298, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 298, + 555 + ], + "type": "text", + "content": "Cross-domain Few-shot learning. Recent works [10-15, 28-31] have developed few-shot learners capable of cross-domain transfer, employing techniques like global-local semantic alignment [10] to enhance representation learning and adversarial training [11, 12, 14] to mitigate the interference of image styles. Some approaches [31, 32] focus on knowledge distillation for effective feature extraction, while others tackle sharp minima in loss landscapes [33] to facilitate knowledge transfer. Additionally, innovative methods like dual adaptive representation alignment [34] and meta-memory frameworks [35] have emerged to bridge domain gaps. Despite some progress, these methods are generally constrained by the use of shallow backbone networks (e.g., ResNet-10) to learn feature initialization models from small-scale source domains like mini-[ImageNet]. In the age of large models, this labor-intensive source domain training process becomes redundant. With the availability of open-source foundation models that offer robust feature initialization, the emphasis of CD-FSL should now shift towards the effective adaptation of foundation models to target domains. Consequently, this study concentrates on addressing the challenges of adapting foundation models in extreme cross-domain scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 559, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 298, + 714 + ], + "type": "text", + "content": "Foundation models and adaptation strategies. The emergence of scalable foundation models, facilitated by Transformers [1] and self-supervised learning techniques, marks a significant advancement in FSL. Models like DINO [2], Masked AutoEncoders [3] (MAE), and CLIP[4] learn transferable feature representations from vast unlabeled datasets. Recent research [7, 8, 8, 9, 9, 36-41] focuses on effectively leveraging these foundation models in downstream few-shot tasks while minimizing data requirements for fine-tuning. Techniques such as LoRA [5], Visual Prompt Tuning [6], and automated CLIP prompt engineering [36] aim to enhance efficiency. Additionally, some approaches utilize semantic-guided vision-language alignment to overcome modality" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 556, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 253 + ], + "type": "text", + "content": "discrepancies [42]. Others focus on improving ViT-based models by mitigating domain sensitivity in CLS tokens [43] and applying adaptive attention temperature mechanisms to enhance feature stability [44]. Meanwhile, integrating contrastive learning with entropy minimization has been shown to boost feature discrimination in source-free scenarios [45], and frequency-aware prompting strengthens robustness by manipulating high-frequency components [46]. Inspired by in-context learning, context-aware meta-learning [47] facilitates rapid, gradient-free adaptation through modeling relationships between support and query samples. Orthogonally, this study proposes a novel adaptation framework for extreme cross-domain scenarios, enhancing the generalization ability of pre-trained vision foundation models by incorporating task-relevant knowledge during fine-tuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 270, + 397, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 397, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 397, + 284 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 287, + 556, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 556, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 556, + 587 + ], + "type": "text", + "content": "Overview. As illustrated in Fig. 3, our framework consists of two structurally identical Transformers, namely Foundation ViT (F-ViT) and Task-specific expert ViT (T-ViT), initialized respectively by the pre-trained vision foundation model and random parameters. Among them, the F-ViT separately takes the original image and its masked image as inputs and extracts their features, respectively. Similarly, the T-ViT receives the original image and then generates its features. More importantly, we meticulously designed a Masked Cross-model unidirectional Reconstruction (MCR) module and a Decision Graph Association (DGA) module to adapt the F-ViT using the task-relevant knowledge exploited by the T-ViT. Specifically, in the MCR module, the masked features from F-ViT are fed into a lightweight decoder to reconstruct the semantic features produced by T-ViT. Then, a reconstruction loss is introduced to encourage the reconstruction accuracy. In the DGA module, we first define the decision graphs using the similarity matrix computed by the class token and the feature tokens (i.e., defined by the ViT architecture) for each model, and then introduce an association loss to constrain their consistency. In the end, the cross-entropy loss, along with reconstruction loss and association loss, are employed for end-to-end training of the entire framework. Once training is completed, we only preserve the F-ViT and its corresponding linear classifier for testing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "text", + "content": "Main framework. Given an image " + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "text", + "content": " from the training set " + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "text", + "content": ", we first transform it into patches and then randomly mask out some of the patches. We denote the masked parts as " + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{X} = \\{\\tilde{x}_i | i = 1, \\dots, I\\}" + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "text", + "content": " and the retained parts as " + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{X} = \\{\\hat{x}_j | j = 1, \\dots, J\\}" + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "text", + "content": ". The full set of patches corresponding to the original image is denoted as " + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "inline_equation", + "content": "X = \\hat{X} \\cup \\tilde{X}" + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "text", + "content": ". The masking ratio " + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma = \\frac{I}{J + I}" + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "text", + "content": ". Subsequently, we utilize F-ViT to extract features from both " + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{X}" + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "text", + "content": " separately. Similarly, we also utilize T-ViT to extract the features corresponding to " + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 601, + 556, + 715 + ], + "type": "text", + "content": ". The" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "spans": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "text", + "content": "4584" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 108, + 69, + 265, + 273 + ], + "blocks": [ + { + "bbox": [ + 108, + 69, + 265, + 273 + ], + "lines": [ + { + "bbox": [ + 108, + 69, + 265, + 273 + ], + "spans": [ + { + "bbox": [ + 108, + 69, + 265, + 273 + ], + "type": "image", + "image_path": "e6cdb3c68037ac6e68a574a026462ef4a06aaad56a2f0b811fbd5e03a0058a50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 276, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 54, + 276, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 276, + 555, + 344 + ], + "type": "text", + "content": "Figure 3. We present a novel absorption adaptation learning framework, which attempt to enhance the extreme cross-domain generalization capacity of a pre-trained vision foundation model (e.g., \"F-ViT\"). This framework achieves this by through absorbing some task-relevant knowledge from an expert model (e.g., \"T-ViT\") during fine-tuning. To this end, we introduce two innovative strategies, including masked cross-model unidirectional reconstruction and decision graph association. By doing these, the task-relevant semantic knowledge in the expert model from both intermediate feature and the final decision levels can be appropriately extracted and absorbed by the foundation model during its fine-tuning, thus mitigating the performance drop caused by domain gap and limited annotation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 293, + 70, + 504, + 273 + ], + "blocks": [ + { + "bbox": [ + 293, + 70, + 504, + 273 + ], + "lines": [ + { + "bbox": [ + 293, + 70, + 504, + 273 + ], + "spans": [ + { + "bbox": [ + 293, + 70, + 504, + 273 + ], + "type": "image", + "image_path": "316178569c307b6e5376c2e08f4fcc64c3d1e341c56252e9e494b06ec2cf4092.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 346, + 274, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 274, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 274, + 360 + ], + "type": "text", + "content": "aforementioned process can be formalized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 369, + 296, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 369, + 296, + 444 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 296, + 444 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\{f _ {c l s} ^ {\\mathrm {F}} \\right\\} \\cup \\left\\{\\bar {f} _ {1} ^ {F}, \\dots , \\bar {f} _ {J} ^ {F} \\right\\} \\cup \\left\\{\\tilde {f} _ {1} ^ {F}, \\dots , \\tilde {f} _ {I} ^ {F} \\right\\} = \\mathrm {F} - \\mathrm {V i T} (X), \\\\ \\left\\{\\hat {f} _ {1} ^ {F}, \\dots , \\hat {f} _ {J} ^ {F} \\right\\} = F - V i T (\\hat {X}), \\\\ \\left\\{f _ {c l s} ^ {T} \\right\\} \\cup \\left\\{\\hat {f} _ {1} ^ {T}, \\dots , \\hat {f} _ {J} ^ {T} \\right\\} \\cup \\left\\{\\tilde {f} _ {1} ^ {T}, \\dots , \\tilde {f} _ {I} ^ {T} \\right\\} = \\mathrm {T} - \\mathrm {V i T} (X), \\tag {1} \\\\ \\end{array}", + "image_path": "d84390f6040b094462e00fddc7441da36077836b119532163716977e71d8964f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "inline_equation", + "content": "f_{cls}^{\\mathrm{F}}" + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "inline_equation", + "content": "f_{cls}^{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "text", + "content": " respectively represent the class token features from F-ViT and T-ViT, " + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "inline_equation", + "content": "\\{\\tilde{f}_1^F,\\dots ,\\tilde{f}_J^F\\} \\cup \\left\\{\\tilde{f}_1^F,\\dots ,\\hat{f}_I^F\\right\\}" + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "inline_equation", + "content": "\\{\\hat{f}_1^T,\\dots ,\\hat{f}_J^T\\} \\cup \\{\\tilde{f}_1^T,\\dots ,\\tilde{f}_I^T\\}" + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "text", + "content": " respectively represent full patch features of " + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "text", + "content": " from F-ViT and T-ViT, " + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "inline_equation", + "content": "\\{\\hat{f}_1^F,\\dots ,\\hat{f}_J^F\\}" + }, + { + "bbox": [ + 55, + 455, + 297, + 541 + ], + "type": "text", + "content": " represents the features w.r.t retained patches after randomly masking." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 548, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 297, + 714 + ], + "type": "text", + "content": "Masked cross-model unidirectional reconstruction. We propose the masked cross-model unidirectional reconstruction (MCR) approach, which aims to reconstruct the masked information in F-ViT under the guidance of T-ViT to enhance the understanding of task-relevant semantic knowledge. In terms of technical details, the MCR module takes the masked features from F-ViT as input and employs a lightweight decoder (e.g., a self-attention layer) to reconstruct the local semantic features of T-ViT. Specifically, we first utilize a set of random vectors " + }, + { + "bbox": [ + 55, + 548, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\{\\tilde{\\varnothing}_1,\\dots ,\\tilde{\\varnothing}_I\\}" + }, + { + "bbox": [ + 55, + 548, + 297, + 714 + ], + "type": "text", + "content": " to pad the masked patches. After merging the retained patch features, we can obtain " + }, + { + "bbox": [ + 55, + 548, + 297, + 714 + ], + "type": "inline_equation", + "content": "V = \\{\\hat{f}_1^F,\\dots ,\\hat{f}_J^F\\} \\cup \\{\\tilde{\\varnothing}_1,\\dots ,\\tilde{\\varnothing}_I\\}" + }, + { + "bbox": [ + 55, + 548, + 297, + 714 + ], + "type": "text", + "content": ". Afterward, to preserve the spatial structure of the original image patches, we" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 346, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 382 + ], + "type": "text", + "content": "add positional encoding to " + }, + { + "bbox": [ + 313, + 346, + 555, + 382 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 346, + 555, + 382 + ], + "type": "text", + "content": ". Subsequently, we utilize the decoder to reconstruct the features of the masked patches as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 389, + 554, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 389, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 318, + 389, + 554, + 418 + ], + "type": "interline_equation", + "content": "\\left\\{\\hat {f} _ {1} ^ {D}, \\dots , \\hat {f} _ {J} ^ {D} \\right\\} \\cup \\left\\{\\tilde {f} _ {1} ^ {D}, \\dots , \\tilde {f} _ {I} ^ {D} \\right\\} = \\operatorname {D e c o d e r} (\\mathrm {P E} (V)), \\tag {2}", + "image_path": "831679d743fbc739d9eb255960670092a592b45cf146e5952cbdc553e6f9315a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 413, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 413, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\left\\{\\tilde{f}_1^D,\\dots ,\\tilde{f}_I^D\\right\\}" + }, + { + "bbox": [ + 313, + 413, + 554, + 495 + ], + "type": "text", + "content": " represents the reconstructed features corresponding to the masked patches, PE " + }, + { + "bbox": [ + 313, + 413, + 554, + 495 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 313, + 413, + 554, + 495 + ], + "type": "text", + "content": " represents the positional encoding function. We use the patch features from the corresponding positions in T-ViT " + }, + { + "bbox": [ + 313, + 413, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\left\\{\\tilde{f}_1^T,\\dots ,\\tilde{f}_I^T\\right\\}" + }, + { + "bbox": [ + 313, + 413, + 554, + 495 + ], + "type": "text", + "content": " as the reconstruction targets, and the loss " + }, + { + "bbox": [ + 313, + 413, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 313, + 413, + 554, + 495 + ], + "type": "text", + "content": " can be computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 378, + 495, + 555, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 495, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 378, + 495, + 555, + 528 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\frac {1}{I} \\sum_ {i = 1} ^ {I} \\left\\| \\tilde {f} _ {i} ^ {T} - \\tilde {f} _ {i} ^ {D} \\right\\|. \\tag {3}", + "image_path": "85e7381d8b6a02486a9517c4f23eb5daa922c83f56d40f9913059350412df595.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 714 + ], + "type": "text", + "content": "Decision graph association. Beyond the MCR module, we expect to introduce a decision-level constraint to further guide the foundation model to absorb task-relevant knowledge from the expert model. While a straightforward approach might involve imposing a logits-level consistency between the foundation model and the expert model, this method is sub-optimal. The primary reason is that these two models have different perceptual decision-making capabilities. Directly enforcing alignment of their logits could cause the entire framework to collapse into a compromise between the two decision states, which might severely misguide the adaptation process of the foundation model. To overcome this issue, we propose a more sophisticated strategy, namely decision graph association (DGA). This method promotes consistency in the token similarity matrix between the two" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "spans": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "text", + "content": "4585" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 261 + ], + "type": "text", + "content": "models, focusing on aligning how these models relate different features to each other internally. The decision graph is constructed from the token similarity matrix, which captures the relationships between different tokens within the models. By ensuring consistency in these token relationships (the \"graph\") between the foundation model and the expert model, the framework effectively transfers the high-order semantic structures learned by the expert model to the foundation model. In a specific, we first define the decision graphs using the similarity between the class token and the feature tokens for each model, and then learns to association them using a consistency loss. Formally, for the F-ViT branch, we compute the cosine similarity between each patch features " + }, + { + "bbox": [ + 55, + 72, + 296, + 261 + ], + "type": "inline_equation", + "content": "f_{l}^{\\mathrm{F}} \\in \\{\\bar{f}_{1}^{F},\\dots ,\\bar{f}_{J}^{F}\\} \\cup \\{\\tilde{f}_{1}^{F},\\dots ,\\tilde{f}_{I}^{F}\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 261 + ], + "type": "text", + "content": " and the class semantic token features " + }, + { + "bbox": [ + 55, + 72, + 296, + 261 + ], + "type": "inline_equation", + "content": "f_{cls}^{\\mathrm{F}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 261 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 88, + 268, + 296, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 268, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 88, + 268, + 296, + 297 + ], + "type": "interline_equation", + "content": "s _ {l} ^ {F} = \\frac {f _ {c l s} ^ {\\mathrm {F}} \\cdot f _ {l} ^ {\\mathrm {F}}}{\\left\\| f _ {c l s} ^ {\\mathrm {F}} \\right\\| \\times \\left\\| f _ {l} ^ {\\mathrm {F}} \\right\\|}, l \\in [ 1, \\dots , J + I ]. \\tag {4}", + "image_path": "1a08d0c38ffc467398188b949cd0416fd4819ea0f8fef82611265695d1ce374b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 298, + 296, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 320 + ], + "type": "text", + "content": "Afterwards, we process the T-ViT branch in the same manner:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 328, + 296, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 328, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 88, + 328, + 296, + 357 + ], + "type": "interline_equation", + "content": "s _ {l} ^ {T} = \\frac {f _ {c l s} ^ {T} \\cdot f _ {l} ^ {T}}{\\left\\| f _ {c l s} ^ {T} \\right\\| \\times \\left\\| f _ {l} ^ {T} \\right\\|}, l \\in [ 1, \\dots , J + I ], \\tag {5}", + "image_path": "e58ad6acb3d3a4be41962da346fc88b03c1d9f321c2c4f40b26cd2f33d9ece03.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 358, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 358, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\forall f_{l}^{T}\\in \\left\\{\\hat{f}_{1}^{T},\\dots ,\\hat{f}_{J}^{T}\\right\\} \\cup \\left\\{\\tilde{f}_{1}^{T},\\dots ,\\tilde{f}_{I}^{T}\\right\\}" + }, + { + "bbox": [ + 55, + 358, + 296, + 388 + ], + "type": "text", + "content": ". Finally, we calculate the association loss " + }, + { + "bbox": [ + 55, + 358, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{aso}" + }, + { + "bbox": [ + 55, + 358, + 296, + 388 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 398, + 296, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 398, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 105, + 398, + 296, + 436 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {a s o} = \\frac {1}{J + I} \\sqrt {\\sum_ {l = 1} ^ {J + I} \\left(s _ {l} ^ {F} - s _ {l} ^ {T}\\right) ^ {2}}. \\tag {6}", + "image_path": "c0f8c3af358bb7c463192f81de8513c63f3b8054966ca1a6b33018d70c96ea82.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 444, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 480 + ], + "type": "text", + "content": "Training and testing. Besides MCR and DGA, we employ linear classifiers to predict class affiliation based on the class token features from F-ViT and T-ViT, respectively:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 112, + 486, + 296, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 486, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 112, + 486, + 296, + 506 + ], + "type": "interline_equation", + "content": "p ^ {F} = F - \\text {C l a s s i f i e r} \\left(f _ {c l s} ^ {F}\\right), \\tag {7}", + "image_path": "0bfd3f82bc4c45ddd0cdc1e417ac8d5b0d87396a720e5adfc817e328ba393489.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 502, + 235, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 502, + 235, + 518 + ], + "spans": [ + { + "bbox": [ + 115, + 502, + 235, + 518 + ], + "type": "interline_equation", + "content": "p ^ {T} = \\mathrm {T - C l a s s i f i e r} \\left(f _ {c l s} ^ {T}\\right),", + "image_path": "f8aa83c6ca8cad0438d5df2d76cc66b5a7fdc27c05d1fa15664fe30e9601a06a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "inline_equation", + "content": "p^F" + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "inline_equation", + "content": "p^T" + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "text", + "content": " respectively represent the predictions. We assign the class corresponding to the maximum value of " + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "inline_equation", + "content": "p^F" + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "text", + "content": " as the predicted label " + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "inline_equation", + "content": "y^{F}" + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "text", + "content": " by the F-ViT for input image " + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "text", + "content": ". Similarly, we denote the predicted label from T-ViT as " + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "inline_equation", + "content": "y^{T}" + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "text", + "content": ". Then, we calculate the cross-entropy loss " + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 55, + 526, + 297, + 599 + ], + "type": "text", + "content": " based on the predicted labels and the ground truth as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 114, + 608, + 296, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 608, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 114, + 608, + 296, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} = H \\left(y ^ {F}, y\\right) + H \\left(y ^ {T}, y\\right), \\tag {8}", + "image_path": "52a49f308824fcb9cd2c9b6453ebab87025a2490855c9da04559caf0f46782c1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 624, + 296, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 296, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 624, + 296, + 658 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 624, + 296, + 658 + ], + "type": "text", + "content": " represents the ground truth for input image " + }, + { + "bbox": [ + 55, + 624, + 296, + 658 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 624, + 296, + 658 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 624, + 296, + 658 + ], + "type": "inline_equation", + "content": "H(\\cdot)" + }, + { + "bbox": [ + 55, + 624, + 296, + 658 + ], + "type": "text", + "content": " denotes the cross-entropy calculation function. In the end, we obtain the total loss as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 106, + 670, + 296, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 670, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 106, + 670, + 296, + 682 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {c e} + \\lambda_ {1} \\cdot \\mathcal {L} _ {r e c} + \\lambda_ {2} \\cdot \\mathcal {L} _ {a s o}, \\tag {9}", + "image_path": "2bac79001d18a64d678b3b2860224959c1aba3b00dc3181ba4d836262c4c172d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 684, + 296, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 684, + 296, + 719 + ], + "spans": [ + { + "bbox": [ + 55, + 684, + 296, + 719 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 684, + 296, + 719 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 684, + 296, + 719 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 684, + 296, + 719 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 684, + 296, + 719 + ], + "type": "text", + "content": " are hyper-parameters. We optimize the entire framework end-to-end based on the loss " + }, + { + "bbox": [ + 55, + 684, + 296, + 719 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 684, + 296, + 719 + ], + "type": "text", + "content": ". After training, we retain only F-ViT and its classifier for testing." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 572, + 131 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 572, + 131 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 572, + 131 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 572, + 131 + ], + "type": "image", + "image_path": "624acfbe92b368f1e9d1fc1ef9282d1c2ad9063cc963f47e3ea644e16473356c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 137, + 572, + 187 + ], + "blocks": [ + { + "bbox": [ + 317, + 137, + 572, + 187 + ], + "lines": [ + { + "bbox": [ + 317, + 137, + 572, + 187 + ], + "spans": [ + { + "bbox": [ + 317, + 137, + 572, + 187 + ], + "type": "image", + "image_path": "f8c18c4e392f169ba34897291d1c0daa3271003515836b17fb9644297e2a7ac2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 188, + 525, + 198 + ], + "lines": [ + { + "bbox": [ + 342, + 188, + 525, + 198 + ], + "spans": [ + { + "bbox": [ + 342, + 188, + 525, + 198 + ], + "type": "text", + "content": "Figure 4. The differences between ours and MAE." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 201, + 555, + 318 + ], + "blocks": [ + { + "bbox": [ + 317, + 201, + 555, + 318 + ], + "lines": [ + { + "bbox": [ + 317, + 201, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 317, + 201, + 555, + 318 + ], + "type": "image", + "image_path": "a914acf03bcf25e9b66ff053fcdce058a991494cc0fa3ee445628d335ac50bcd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 319, + 534, + 329 + ], + "lines": [ + { + "bbox": [ + 332, + 319, + 534, + 329 + ], + "spans": [ + { + "bbox": [ + 332, + 319, + 534, + 329 + ], + "type": "text", + "content": "Figure 5. The differences between ours and distillation." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 335, + 556, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 556, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 556, + 622 + ], + "type": "text", + "content": "Further analysis. Our framework provides deep insights into how to adapt foundation models to extreme CD-FSL tasks. It goes beyond original Masked Auto-Encoder (MAE) [3] or knowledge distillation [10, 48]. We compare our method with MAE and distillation in Fig. 4 and Fig. 5, respectively. Firstly, the original MAE tends to overfit in such few-shot tasks, leading to poor generalization. In contrast, the proposed masked cross-model unidirectional reconstruction scheme offers two advantages. For one thing, our approach explicitly injects task-relevant expert knowledge into the adaptation process, and forces the foundation model to focus on task-relevant semantic patterns. For another, our method can utilize T-ViT to reduce F-ViT's overconfidence. Secondly, simple knowledge distillation is insufficient to assist in adapting foundation models to extreme cross-domain few-shot tasks. The reason is that F-ViT and T-ViT have different perception capabilities for few-shot tasks, and forcibly aligning their features or logits may lead to the model collapsing into a trivial state. In contrast, the proposed decision graph association is a higher-order semantic structural cue with robust properties, leading to better generalization. Thirdly, we jointly train T-ViT and F-ViT, employing DGA for mutual regularization to mitigate overfitting in both models." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 631, + 458, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 458, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 458, + 643 + ], + "type": "text", + "content": "4. Experimental Evaluation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 645, + 437, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 437, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 437, + 658 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 663, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 555, + 686 + ], + "type": "text", + "content": "We evaluate the effectiveness of the proposed method in two distinct settings: extreme CD-FSL and traditional CD-FSL." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "content": "Extreme setting. In extreme CD-FSL setting, we utilize benchmarks that include ISIC [25], ChestX [26], Eu" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "spans": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "text", + "content": "4586" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 117, + 64, + 492, + 156 + ], + "blocks": [ + { + "bbox": [ + 66, + 53, + 543, + 64 + ], + "lines": [ + { + "bbox": [ + 66, + 53, + 543, + 64 + ], + "spans": [ + { + "bbox": [ + 66, + 53, + 543, + 64 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art methods. \"B\" denotes the backbone. \"F\" denotes the initialization with foundation model." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 117, + 64, + 492, + 156 + ], + "lines": [ + { + "bbox": [ + 117, + 64, + 492, + 156 + ], + "spans": [ + { + "bbox": [ + 117, + 64, + 492, + 156 + ], + "type": "table", + "html": "
ISICChestEuroSATCropDiseaseAve.
MethodBF1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
ProtoNet [27]ViT-SDINO24.1434.0616.4319.1159.0578.7039.3973.4034.7551.31
Linear probing [49]ViT-SDINO24.9451.8818.2925.2960.3883.0251.0184.0738.6561.06
LoRA [5]ViT-SDINO24.4450.2918.2925.3560.0883.0952.8184.0738.9060.70
VPT [6]ViT-SDINO27.7953.6817.7624.8560.3481.9251.1883.5839.2661.00
RotNet [50]ViT-SDINO34.3960.4220.8428.0559.0581.9650.1684.5141.1163.73
SimCLR [51]ViT-SDINO28.7258.5618.8726.0760.5883.6752.8383.3040.2562.90
BYOL [52]ViT-SDINO28.0059.5319.9028.0958.0382.9852.0385.3639.4963.99
SetFeat [53]ViT-SDINO24.1433.5716.4319.2459.0572.5339.3966.6234.7547.99
MaskCLIP [48]ViT-SDINO28.1260.4919.0826.0858.8283.2353.9485.8139.9963.90
OursViT-SDINO37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
", + "image_path": "a9d59b220cfd56f81b7d0450d7c48c8bb3aaa842f0d86ac69ef747adfa10c149.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 160, + 297, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 160, + 297, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 297, + 327 + ], + "type": "text", + "content": "roSAT [54], and CropDisease [55]. These datasets demonstrate substantial distributional differences compared to the natural image domain [56]. Unlike the previous practice, we re-organize these datasets into a more challenging All-way " + }, + { + "bbox": [ + 55, + 160, + 297, + 327 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 160, + 297, + 327 + ], + "type": "text", + "content": "-shot configuration, where all categories are considered in the few-shot task, making it a more rigorous test of cross-domain generalization. This setup allows us to more effectively evaluate the ability of our method to handle diverse, real-world domain shifts. Specifically, we randomly sample " + }, + { + "bbox": [ + 55, + 160, + 297, + 327 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 160, + 297, + 327 + ], + "type": "text", + "content": " samples from all categories in the target domain for model adaptation, and then perform model evaluation on all remaining samples. To mitigate the effects of randomness, we conduct five runs for each target domain and utilize the average result as the metric." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "text", + "content": "Traditional setting. In the traditional CD-FSL setting, we follow the benchmarks established by Guo et al.[56] and Tseng et al.[57] for experimentation. Specifically, this involves the " + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "text", + "content": "-shot " + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "text", + "content": "-query setup, where " + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "inline_equation", + "content": "N \\times K" + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "text", + "content": " support samples are used to adapt the model, and " + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "inline_equation", + "content": "N \\times M" + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "text", + "content": " query samples are used for evaluation. In this context, " + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "text", + "content": " represents a subset randomly sampled from the target domain categories, typically set to 5, while " + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "text", + "content": " denotes the number of labeled samples per category, and " + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "text", + "content": " represents the number of test samples per category, usually set to 15. For evaluation, we adopt the same protocol as prior studies [10-12, 14, 58], sampling 600 random " + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "text", + "content": "-way " + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 327, + 296, + 578 + ], + "type": "text", + "content": "-shot tasks from each target domain and using the average accuracy across these tasks as the performance metric. Additionally, previous methods [10-12, 14, 58] typically use two types of inference: inductive and transductive. Inductive inference makes predictions solely based on the support set of few-shot tasks, while transductive inference leverages query data in an unsupervised manner to assist in inference. To ensure a comprehensive comparison, we follow [10, 11, 58] and implement both inductive and transductive inference." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 578, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 602 + ], + "type": "text", + "content": "Due to space constraints, more implementation details can be found in the supplementary document." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 606, + 281, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 281, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 281, + 617 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Results on extreme CD-FSL. We compare the proposed method with state-of-the-art works to clarify our advantages. Among these comparison methods, ProtoNet [27] and SetFeat [53] represent the classical baseline in FSL. VPT [6] and LoRA [5] are representative methods that achieves knowledge transfer from the foundation model to downstream task. SimCLR [51] and BYOL [52] are two exemplary transfer learning methods. MaskCLIP [48] is a robust representation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 161, + 555, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 555, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 555, + 351 + ], + "type": "text", + "content": "learning method characterized by its effective generalization potential. We re-implemented these methods under the same settings for comparison in extreme CD-FSL tasks. The experimental results are as shown in Table 1. Overall, the proposed method achieves the best performance compared to these state-of-the-art approaches. Specifically, our method consistently outperforms methods based on transfer learning (e.g., VPT, LoRA, SimCLR, MaskCLIP, etc.). This indicates the strong generalization capability of the proposed method in handling extreme cross-domain scenarios. In addition, compared to ProtoNet [27] and SetFeat [53], our method demonstrates significant performance advantages. This suggests that the robustness of advanced few-shot learners is limited without the introduction of explicit cross-domain learning mechanisms. The above results demonstrate the advantages of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 353, + 556, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 556, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 556, + 400 + ], + "type": "text", + "content": "Results on traditional CD-FSL. Table 2 presents the results for both the 5-way 1-shot and 5-way 5-shot traditional CD-FSL tasks. Overall, our method outperforms advanced comparison methods, further validating its effectiveness." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 403, + 408, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 408, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 408, + 415 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 415, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 654 + ], + "type": "text", + "content": "Comparing our method with baselines. We compare our method with F-ViT and T-ViT baselines to highlight our advantages. Among them, F-ViT refers to the standard fine-tuning of the foundation model. We experimentally found that standard fine-tuning is relatively the best approach. Therefore, we consider F-ViT as our strong baseline. T-ViT refers to a model that learns from scratch using few-shot data from the target domain. We conduct experiments under 1-shot and 5-shot settings across four datasets, initializing F-ViT with DINO unless otherwise specified. The experimental results are shown in Table 3. The proposed method exhibits significant performance advantages in average accuracy across four datasets compared to the T-ViT and F-ViT baselines. Furthermore, this experiment shows that relying solely on T-ViT is insufficient for effectively addressing such extreme cross-domain few-shot tasks. Nevertheless, due to the task-specific intrinsic properties of T-ViT, it still holds some decision-making instructions. Therefore, our framework, which leverages T-ViT to enhance the generalization of F-ViT, possesses practical application value." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 713 + ], + "type": "text", + "content": "Evaluating the effectiveness of MCR and DGA. Our framework consists of two core components: Masked Cross-model unidirectional Reconstruction (MCR) and Decision Graph Association (DGA). We conduct four experimental cases to evaluate the contributions of them. Case 1: The pro" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "spans": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "text", + "content": "4587" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 92, + 504, + 305 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 92 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art methods on traditional 5-way CD-FSL. † stands for using query data in an unsupervised manner to assist in inference. “B” denotes the backbone. “F” denotes the initialization with foundation model. “R10” denotes ResNet-10." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 108, + 92, + 504, + 305 + ], + "lines": [ + { + "bbox": [ + 108, + 92, + 504, + 305 + ], + "spans": [ + { + "bbox": [ + 108, + 92, + 504, + 305 + ], + "type": "table", + "html": "
MethodBFShotCUBCarsPlacesPlantaeChestISICEuroSATCropDiseaseAve.
Dara [34]R10-1-shot52.7035.2551.2542.0822.9236.4267.4280.7448.60
GNN+FAP [46]R10-1-shot46.9134.6954.4136.7322.3635.6362.9669.9745.46
IM-DCL†[46]ViT-B-1-shot----22.6838.2074.6885.98-
PMF [59]ViT-SDINO1-shot78.1337.2471.1153.6021.7330.3670.7480.7955.46
MEM [60]ViT-SDINO1-shot----22.7632.9768.1181.11-
StyleAdv†[11]ViT-SDINO1-shot84.0140.4872.6455.5222.9233.0572.1581.2257.75
StyleAdv-FT†[11]ViT-SDINO1-shot84.0140.4872.6455.5222.9233.9974.9384.1158.57
FLoR†[33]ViT-SDINO1-shot85.4043.4274.6952.2923.2635.4973.0983.5558.90
AttnTemp [44]ViT-SDINO1-shot----23.6338.0575.0984.78-
AttnTemp†[44]ViT-SDINO1-shot----23.9640.1377.4087.58-
CD-CLS [43]ViT-SDINO1-shot----23.3935.5674.9784.54-
CD-CLS†[43]ViT-SDINO1-shot----23.8837.2078.4187.39-
OursViT-SDINO1-shot85.2644.1374.3057.5423.1334.4872.0882.1359.13
Ours†ViT-SDINO1-shot94.1946.4582.4163.8423.1235.0282.3891.3664.84
Dara [34]R10-5-shot77.5158.4472.1565.4027.5456.2885.8495.3267.31
GNN+FAP [46]R10-5-shot67.6650.2074.9854.5425.3147.6082.5291.7961.83
IM-DCL†[46]ViT-B-5-shot----27.9853.8987.7695.93-
PMF [59]ViT-SDINO5-shot----27.2750.1285.9892.96-
MEM [60]ViT-SDINO5-shot----26.6747.3886.4993.74-
StyleAdv†[11]ViT-SDINO5-shot95.8261.7388.3375.5526.9747.7388.5794.8572.44
StyleAdv-FT†[11]ViT-SDINO5-shot95.8266.0288.3378.0126.9751.2390.1295.9974.06
FLoR†[33]ViT-SDINO5-shot96.5368.4489.4876.2227.0253.0690.7596.4774.75
AttnTemp [44]ViT-SDINO5-shot----28.0354.9190.8296.66-
AttnTemp†[44]ViT-SDINO5-shot----28.4155.2291.3496.74-
CD-CLS [43]ViT-SDINO5-shot----27.6654.6991.5396.27-
CD-CLS†[43]ViT-SDINO5-shot----28.2555.6691.6896.62-
OursViT-SDINO5-shot97.5574.7689.1581.7326.2550.5589.9995.8175.72
Ours†ViT-SDINO5-shot98.3978.0091.4684.7426.1151.1992.7997.2677.49
", + "image_path": "c55c2834a770a83950f7b21d6c2dd5e2c6537cd9dde7e2ad9c8e7d4d4ce08535.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 124, + 318, + 488, + 412 + ], + "blocks": [ + { + "bbox": [ + 188, + 307, + 421, + 318 + ], + "lines": [ + { + "bbox": [ + 188, + 307, + 421, + 318 + ], + "spans": [ + { + "bbox": [ + 188, + 307, + 421, + 318 + ], + "type": "text", + "content": "Table 3. Ablation study. " + }, + { + "bbox": [ + 188, + 307, + 421, + 318 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 188, + 307, + 421, + 318 + ], + "type": "text", + "content": " indicates that the component is used." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 124, + 318, + 488, + 412 + ], + "lines": [ + { + "bbox": [ + 124, + 318, + 488, + 412 + ], + "spans": [ + { + "bbox": [ + 124, + 318, + 488, + 412 + ], + "type": "table", + "html": "
ISICChestEuroSATCropDiseaseAve.
Method1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
T-ViT baseline26.7153.6821.1526.1638.3449.3324.6655.4927.7146.16
F-ViT baseline26.3560.9520.0526.4160.2482.9852.1886.4839.7064.20
Ours37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
MCRDGA1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
XX26.3560.9520.0526.4160.2482.9852.1886.4839.7064.20
✓X32.2762.4921.2827.7260.2583.5354.0586.5241.9665.06
X✓33.1162.2821.1627.3260.2983.9653.5886.4242.0364.99
✓✓37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
", + "image_path": "c44ff43aa740e5bc6ef71cc24c4144bfe9865fb9e46b9cbf203ad89b9e6d5b36.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "text", + "content": "posed method without the MCR and DGA modules, equivalent to a standard fine-tuning baseline. Case 2: The proposed method using only the MCR module on top of the baseline. Case 3: The proposed method using only the DGA module on top of the baseline. Case 4: The proposed method using both MCR and DGA modules simultaneously. As can be seen in Table 3, compared to the stand fine-tuning baseline (\"F-ViT baseline\"), utilizing either the MCR or the DGA alone can achieve certain performance gains. For example, across the four datasets, the MCR module achieves improvements of " + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "inline_equation", + "content": "2.26\\%" + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "inline_equation", + "content": "0.86\\%" + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "text", + "content": " under the 1-shot and 5-shot settings respectively. Similarly, the DGA module achieves improvements of " + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "inline_equation", + "content": "2.33\\%" + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "inline_equation", + "content": "0.79\\%" + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "text", + "content": " under the 1-shot and 5-shot settings respectively. Notably, on the ISIC dataset under 1-shot setting, MCR and DGA achieve performance gains of " + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "inline_equation", + "content": "5.92\\%" + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "inline_equation", + "content": "6.76\\%" + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "text", + "content": " respectively. Moreover, integrating these two components into a unified framework can unleash greater potential. For instance, compared to using a single component, simultaneously using them further increases the average results by " + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "inline_equation", + "content": "1.32\\%" + }, + { + "bbox": [ + 54, + 415, + 298, + 668 + ], + "type": "text", + "content": " in the 1-shot setting. These results confirm the benefits of the proposed approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 676, + 159, + 688 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 676, + 159, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 676, + 159, + 688 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 297, + 715 + ], + "type": "text", + "content": "Comparing our method with MAE variants. We compare the proposed method with the standard MAE, which" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 415, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 676 + ], + "type": "text", + "content": "involves adding a masked reconstruction loss on top of the fine-tuning F-ViT baseline. Additionally, our method reconstructs the masked features in F-ViT using T-ViT as the target, aiming to enhance F-ViT's focus on task-specific semantic patterns. Our approach is essentially a unidirectional cross-model masked reconstruction. To demonstrate the benefits of this design, we also compare it with a bidirectional MAE variant, where the masked parts of F-ViT and T-ViT are reconstructed using T-ViT and F-ViT as targets, respectively. As show in Table 4, our method shows clear performance advantages over both of these MAE variants. Firstly, the original MAE tends to over-fit in such FSL tasks, leading to poor generalization. In contrast, our masked cross-model unidirectional reconstruction offers two advantages. For one thing, our approach explicitly injects task-specific model knowledge into the foundation model's adaptation process. For another, our method utilizes T-ViT to relax F-ViT's overconfidence, mitigating over-fitting to some extent. Secondly, the optimization objective of bidirectional MAE is ambiguous. In contrast, our unidirectional cross-modality masked reconstruction approach is goal-oriented, which has resulted in improved performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 677, + 558, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 558, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 558, + 715 + ], + "type": "text", + "content": "Comparing our method with distillation or fusion. We compare the proposed method with knowledge distillation and model fusion approaches to clarify our previous claims." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "spans": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "text", + "content": "4588" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 126, + 76, + 485, + 152 + ], + "blocks": [ + { + "bbox": [ + 181, + 66, + 427, + 76 + ], + "lines": [ + { + "bbox": [ + 181, + 66, + 427, + 76 + ], + "spans": [ + { + "bbox": [ + 181, + 66, + 427, + 76 + ], + "type": "text", + "content": "Table 4. Comparing ours with MAE and distillation based methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 126, + 76, + 485, + 152 + ], + "lines": [ + { + "bbox": [ + 126, + 76, + 485, + 152 + ], + "spans": [ + { + "bbox": [ + 126, + 76, + 485, + 152 + ], + "type": "table", + "html": "
ISICChestEuroSATCropDiseaseAve.
Method1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot1-shot5-shot
MAE35.6462.3921.1927.3356.7877.5152.3885.7341.4963.24
Bidirectional MAE30.3862.0919.3928.6759.8983.0652.5985.0240.5664.71
Feature-level Distillation27.7662.3122.0127.2656.2879.8153.5886.1239.9063.87
Logits-level Distillation27.5361.5420.6725.9960.8883.1854.8786.5140.9864.30
Cross-attention Fusion29.3462.1920.7826.7358.7482.4053.7185.6440.6464.24
Decision Fusion25.4456.9619.8525.5058.9482.7636.2961.9835.1356.80
Ours37.4962.9021.3628.7460.2883.9454.2786.7043.3565.57
", + "image_path": "c487f5163aa979278d867b6daaaab1d440bfd9f8f39e6d960ebcf9bbc9f14c16.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 86, + 166, + 266, + 212 + ], + "blocks": [ + { + "bbox": [ + 64, + 156, + 286, + 166 + ], + "lines": [ + { + "bbox": [ + 64, + 156, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 64, + 156, + 286, + 166 + ], + "type": "text", + "content": "Table 5. Comparison with SoTA methods on mini-ImageNet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 166, + 266, + 212 + ], + "lines": [ + { + "bbox": [ + 86, + 166, + 266, + 212 + ], + "spans": [ + { + "bbox": [ + 86, + 166, + 266, + 212 + ], + "type": "table", + "html": "
MethodBackbone5-way 1-shot5-way 5-shot
CAML [47]ViT-B96.2098.60
SgVA-CLIP [42]ViT-B97.9598.72
PMF [59]ViT-S93.1098.00
OursViT-S94.6898.86
", + "image_path": "80376bc072b08dbedb051df53e9f99cb40238c8ba96435ca7ef1995ba35c5e4f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 88, + 224, + 264, + 262 + ], + "blocks": [ + { + "bbox": [ + 62, + 214, + 288, + 224 + ], + "lines": [ + { + "bbox": [ + 62, + 214, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 288, + 224 + ], + "type": "text", + "content": "Table 6. Comparison with SoTA methods on tiered-ImageNet." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 224, + 264, + 262 + ], + "lines": [ + { + "bbox": [ + 88, + 224, + 264, + 262 + ], + "spans": [ + { + "bbox": [ + 88, + 224, + 264, + 262 + ], + "type": "table", + "html": "
MethodBackbone5-way 1-shot5-way 5-shot
FewTURE [61]ViT-S72.9686.43
HCT [62]ViT-S79.6791.72
OursViT-S82.7893.78
", + "image_path": "602967c363fd1d8c1f5997803f21eb746e153476f4b1f3b04dbbcd44974897ae.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 93, + 277, + 260, + 321 + ], + "blocks": [ + { + "bbox": [ + 72, + 266, + 278, + 277 + ], + "lines": [ + { + "bbox": [ + 72, + 266, + 278, + 277 + ], + "spans": [ + { + "bbox": [ + 72, + 266, + 278, + 277 + ], + "type": "text", + "content": "Table 7. Comparison with SoTA methods on CIFAR-FS." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 277, + 260, + 321 + ], + "lines": [ + { + "bbox": [ + 93, + 277, + 260, + 321 + ], + "spans": [ + { + "bbox": [ + 93, + 277, + 260, + 321 + ], + "type": "table", + "html": "
MethodBackbone5-way 1-shot5-way 5-shot
HCT [62]ViT-S78.8990.50
PMF [59]ViT-S81.1092.50
CAML [47]ViT-B70.8085.50
OursViT-S87.6496.03
", + "image_path": "491776c1eb722760b1b8840d11c77072a62fa730a104420fb436e4b7809ada54.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 324, + 297, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 297, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 297, + 586 + ], + "type": "text", + "content": "Specifically, we conduct variant experiments atop the baseline. Variant 1: Feature-level Distillation (FD), aligns the features of the foundation model and the task-specific model directly. Variant 2: Logits-level Distillation (LD), aligns the class prediction scores of the two models directly. Variant 3: Cross-attention Fusion (CF), which involves integrating the features of the foundation model and task-specific model through cross-attention mechanisms. Variant 4: Decision Fusion (DF), directly integrating the baseline model and task-specific model for decision-making. The results are presented in Table 4. It can be observed that our method consistently outperforms these variants. For instance, across four datasets, our method exhibits improvements ranging from " + }, + { + "bbox": [ + 55, + 324, + 297, + 586 + ], + "type": "inline_equation", + "content": "2.37\\%" + }, + { + "bbox": [ + 55, + 324, + 297, + 586 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 324, + 297, + 586 + ], + "type": "inline_equation", + "content": "8.22\\%" + }, + { + "bbox": [ + 55, + 324, + 297, + 586 + ], + "type": "text", + "content": " in 1-shot setting and from " + }, + { + "bbox": [ + 55, + 324, + 297, + 586 + ], + "type": "inline_equation", + "content": "1.22\\%" + }, + { + "bbox": [ + 55, + 324, + 297, + 586 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 324, + 297, + 586 + ], + "type": "inline_equation", + "content": "8.77\\%" + }, + { + "bbox": [ + 55, + 324, + 297, + 586 + ], + "type": "text", + "content": " in 5-shot setting, on average. The insights it brings encompass two aspects. Firstly, our framework embodies sophisticated absorption adaptation mechanisms, which are neither simple knowledge distillation nor plain model fusion. Secondly, conventional technical approaches such as knowledge distillation, model fusion, are insufficient to tackle complex extreme CD-FSL. In contrast, our meticulously crafted framework can provide valuable technical contributions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 298, + 714 + ], + "type": "text", + "content": "Extending the proposed method to in-domain FSL. We further extend our method to in-domain FSL. We follow common experimental setups from previous works [59, 61, 62] to conduct in-domain FSL experiments on three representative benchmarks: mini-[ImageNet [63], tiered-[ImageNet [64] and CIFAR-FS [24]. To ensure fairness, we maintain the same data settings as previous works [59, 61, 62]. As shown in Table 5, Table 6, and Table 7, the proposed method excels in standard in-domain FSL as well. This reinforces the robustness and versatility of our approach." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 167, + 554, + 209 + ], + "blocks": [ + { + "bbox": [ + 344, + 156, + 523, + 166 + ], + "lines": [ + { + "bbox": [ + 344, + 156, + 523, + 166 + ], + "spans": [ + { + "bbox": [ + 344, + 156, + 523, + 166 + ], + "type": "text", + "content": "Table 8. Evaluating the efficiency of our method." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 167, + 554, + 209 + ], + "lines": [ + { + "bbox": [ + 317, + 167, + 554, + 209 + ], + "spans": [ + { + "bbox": [ + 317, + 167, + 554, + 209 + ], + "type": "table", + "html": "
TrainingInference
MethodTime (S)Memory (M)Time (S)Memory (M)Accuracy (%)
F-ViT837.40347514.10251639.70
Ours939.50642614.10251643.35
", + "image_path": "288fffd904cc8d6392a8dfaa3948fc90d5b996d960f14104c4006d11d3773bf6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 338 + ], + "type": "text", + "content": "Efficiency. The efficiency of our method is evaluated in Table 8, with all experiments conducted in the All-way 1-shot setting on a 3090 GPU. The accuracy reported in Table 8 represents the average results across four datasets: ISIC, Chest, EuroSAT, and CropDisease. The training time refers to the duration of each training epoch, while the testing time indicates the total time required to process all testing samples. As shown, our method introduces minimal overhead during training, matches the F-ViT baseline in terms of inference efficiency, and significantly outperforms it in accuracy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 347, + 391, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 391, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 391, + 359 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": "In this study, we propose a novel absorption adaptation learning framework, where a foundation model is guided by an expert model trained from scratch on targeted data. Central to our method are two innovative approaches: masked cross-model unidirectional reconstruction and decision graph association. These strategies enable the foundation model to absorb task-relevant knowledge from the expert model, both at the intermediate feature and final decision levels, thereby mitigating performance drops caused by domain gaps and limited annotations. Both qualitative and quantitative analyses substantiate the effectiveness of our approach." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 506, + 556, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 578 + ], + "type": "text", + "content": "Limitations. First, our framework employs two parallel ViTs for target domain adaptation, leading to increased computation and storage during training. Second, while it shows promising results, there is still room for improvement, particularly in extreme scenarios. Third, being tailored for ViT models, our approach inherits their limitations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 713 + ], + "type": "text", + "content": "Acknowledgment. This work was supported in part by the National Natural Science Foundation of China under Grand 62372379, Grant 62472359, and Grant 62472350; in part by the Xi'an's Key Industrial Chain Core Technology Breakthrough Project: AI Core Technology Breakthrough under Grand 23ZDCYJSGG0003-2023; in part by Innovation Foundation for Doctor Dissertation of Northwestern Polytechnical University under Grant CX2024017; in part by National Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing fundation under Grant TJ-04-23-04." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "spans": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "text", + "content": "4589" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 175 + ], + "type": "text", + "content": "[1] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 175, + 297, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 297, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 297, + 246 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 247, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 296, + 306 + ], + "type": "text", + "content": "[3] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 307, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 296, + 388 + ], + "type": "text", + "content": "[4] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 390, + 296, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 390, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 390, + 296, + 449 + ], + "type": "text", + "content": "[5] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 450, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 450, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 450, + 296, + 509 + ], + "type": "text", + "content": "[6] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 510, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 296, + 568 + ], + "type": "text", + "content": "[7] Samyadeep Basu, Shell Hu, Daniela Massiceti, and Soheil Feizi. Strong baselines for parameter-efficient few-shot fine-tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 11024-11031, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 570, + 296, + 629 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 570, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 62, + 570, + 296, + 629 + ], + "type": "text", + "content": "[8] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23681-23690, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 630, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 630, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 630, + 296, + 689 + ], + "type": "text", + "content": "[9] Shuai Shao, Yu Bai, Yan Wang, Baodi Liu, and Bin Liu. Collaborative consortium of foundation models for open-world few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4740-4747, 2024. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 689, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 689, + 296, + 714 + ], + "type": "text", + "content": "[10] Fei Zhou, Peng Wang, Lei Zhang, Wei Wei, and Yanning Zhang. Revisiting prototypical network for cross" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 337, + 72, + 555, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 72, + 555, + 119 + ], + "spans": [ + { + "bbox": [ + 337, + 72, + 555, + 119 + ], + "type": "text", + "content": "domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20061-20070, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 121, + 555, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 121, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 316, + 121, + 555, + 180 + ], + "type": "text", + "content": "[11] Yuqian Fu, Yu Xie, Yanwei Fu, and Yu-Gang Jiang. Styleadv: Meta style adversarial training for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24575-24584, 2023. 3, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 182, + 555, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 182, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 182, + 555, + 264 + ], + "type": "text", + "content": "[12] Haoqing Wang and Zhi-Hong Deng. Cross-domain few-shot classification via adversarial task augmentation. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 1075–1081. International Joint Conferences on Artificial Intelligence Organization, 8 2021. Main Track. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 266, + 555, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 555, + 312 + ], + "type": "text", + "content": "[13] Yi Wu, Ziqiang Li, Chaoyue Wang, Heliang Zheng, Shanshan Zhao, Bin Li, and Dacheng Tao. Domain remodulation for few-shot generative domain adaptation. volume 36, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 315, + 555, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 315, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 555, + 361 + ], + "type": "text", + "content": "[14] Yanxu Hu and Andy J Ma. Adversarial feature augmentation for cross-domain few-shot classification. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 363, + 555, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 555, + 411 + ], + "type": "text", + "content": "[15] Pin-Yu Chen. Model reprogramming: Resource-efficient cross-domain machine learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 22584-22591, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 412, + 555, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 555, + 470 + ], + "type": "text", + "content": "[16] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Universal representation learning from multiple domains for few-shot classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9526-9535, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 472, + 555, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 555, + 531 + ], + "type": "text", + "content": "[17] Yuqian Fu, Yanwei Fu, and Yu-Gang Jiang. Meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. In Proceedings of the 29th ACM international conference on multimedia, pages 5326-5334, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 533, + 555, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 555, + 603 + ], + "type": "text", + "content": "[18] Jiamei Sun, Sebastian Lapuschkin, Wojciech Samek, Yunqing Zhao, Ngai-Man Cheung, and Alexander Binder. Explanation-guided training for cross-domain few-shot classification. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 7609-7616. IEEE, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 605, + 555, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 605, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 316, + 605, + 555, + 675 + ], + "type": "text", + "content": "[19] Baoquan Zhang, Chuyao Luo, Demin Yu, Xutao Li, Huiwei Lin, Yunming Ye, and Bowen Zhang. Metadata: Meta-learning with conditional diffusion for few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 16687-16695, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 677, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 555, + 714 + ], + "type": "text", + "content": "[20] Ilia Sucholutsky and Thomas L Griffiths. Alignment with human representations supports robust few-shot learning. In Proceedings of the 37th International" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "spans": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "text", + "content": "4590" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 78, + 72, + 296, + 96 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 72, + 296, + 96 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 296, + 96 + ], + "type": "text", + "content": "Conference on Neural Information Processing Systems, pages 73464-73479, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 98, + 296, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 98, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 56, + 98, + 296, + 170 + ], + "type": "text", + "content": "[21] Meiqi Sun, Zhonghan Zhao, Wenhao Chai, Hanjun Luo, Shidong Cao, Yanting Zhang, Jenq-Neng Hwang, and Gaoang Wang. Uniap: Towards universal animal perception in vision via few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5008-5016, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 232 + ], + "type": "text", + "content": "[22] Xingyu Zhu, Shuo Wang, Jinda Lu, Yanbin Hao, Haifeng Liu, and Xiangnan He. Boosting few-shot learning via attentive feature regularization. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 7793-7801, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 233, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 233, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 56, + 233, + 296, + 269 + ], + "type": "text", + "content": "[23] Sachin Ravi and Hugo Larochelle. Optimization as a model for few-shot learning. In International conference on learning representations, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 271, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 296, + 319 + ], + "type": "text", + "content": "[24] Luca Bertinetto, Joao F Henriques, Philip HS Torr, and Andrea Vedaldi. Meta-learning with differentiable closed-form solvers. In International Conference on Learning Representations, 2019. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 296, + 369 + ], + "type": "text", + "content": "[25] Philipp Tschandl, Cliff Rosendahl, and Harald Kittler. The ham10000 dataset, a large collection of multisource dermatoscopic images of common pigmented skin lesions. Scientific data, 5(1):1-9, 2018. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 296, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 296, + 455 + ], + "type": "text", + "content": "[26] Xiaosong Wang, Yifan Peng, Le Lu, Zhiyong Lu, Mohammadhadi Bagheri, and Ronald M Summers. Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2097-2106, 2017. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 456, + 296, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 503 + ], + "type": "text", + "content": "[27] Jake Snell, Kevin Swersky, and Richard Zemel. Phototypical networks for few-shot learning. In Advances in Neural Information Processing Systems, pages 4077-4087, 2017. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 506, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 506, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 506, + 296, + 577 + ], + "type": "text", + "content": "[28] Yurong Guo, Ruoyi Du, Yuan Dong, Timothy Hospedales, Yi-Zhe Song, and Zhanyu Ma. Task-aware adaptive learning for cross-domain few-shot learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1590-1599, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 580, + 296, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 296, + 651 + ], + "type": "text", + "content": "[29] Rashindrie Perera and Saman Halgamuge. Discriminative sample-guided and parameter-efficient feature space adaptation for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23794-23804, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 653, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 653, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 653, + 296, + 712 + ], + "type": "text", + "content": "[30] Wei-Hong Li, Xialei Liu, and Hakan Bilen. Cross-domain few-shot learning with task-specific adapters. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7161-7170, 2022." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 131 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 131 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 131 + ], + "type": "text", + "content": "[31] Hao ZHENG, Runqi Wang, Jianzhuang Liu, and Asako Kanezaki. Cross-level distillation and feature denoising for cross-domain few-shot classification. In The Eleventh International Conference on Learning Representations. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 133, + 555, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 133, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 316, + 133, + 555, + 204 + ], + "type": "text", + "content": "[32] Yongjin Yang, Taehyeon Kim, and Se-Young Yun. Leveraging normalization layer in adapters with progressive learning and adaptive distillation for cross-domain few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 16370–16378, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 206, + 555, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 555, + 265 + ], + "type": "text", + "content": "[33] Yixiong Zou, Yicong Liu, Yiman Hu, Yuhua Li, and Ruixuan Li. Flatten long-range loss landscapes for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23575-23584, 2024. 3, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 266, + 555, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 555, + 325 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 555, + 325 + ], + "type": "text", + "content": "[34] Yifan Zhao, Tong Zhang, Jia Li, and Yonghong Tian. Dual adaptive representation alignment for cross-domain few-shot learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11720-11732, 2023. 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 327, + 555, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 373 + ], + "type": "text", + "content": "[35] Wenjian Wang, Lijuan Duan, Yuxi Wang, Junsong Fan, and Zhaoxiang Zhang. Mmt: cross domain few-shot learning via meta-memory transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 376, + 555, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 555, + 422 + ], + "type": "text", + "content": "[36] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 424, + 555, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 424, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 316, + 424, + 555, + 494 + ], + "type": "text", + "content": "[37] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19325-19337, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 496, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 496, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 496, + 555, + 567 + ], + "type": "text", + "content": "[38] Renrui Zhang, Xiangfei Hu, Bohao Li, Siyuan Huang, Hanqiu Deng, Yu Qiao, Peng Gao, and Hongsheng Li. Prompt, generate, then cache: Cascade of foundation models makes strong few-shot learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15211-15222, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 569, + 555, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 555, + 627 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 555, + 627 + ], + "type": "text", + "content": "[39] Lin Song, Ruoyi Xue, Hang Wang, Hongbin Sun, Yixiao Ge, Ying Shan, et al. Meta-adapter: An online few-shot learner for vision-language model. Advances in Neural Information Processing Systems, 36:55361-55374, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 629, + 555, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 629, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 316, + 629, + 555, + 688 + ], + "type": "text", + "content": "[40] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2):581-595, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 690, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 555, + 714 + ], + "type": "text", + "content": "[41] Jiamin Wu, Xin Liu, Xiaotian Yin, Tianzhu Zhang, and Yongdong Zhang. Task-adaptive prompted transformer" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 738, + 314, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 738, + 314, + 748 + ], + "spans": [ + { + "bbox": [ + 295, + 738, + 314, + 748 + ], + "type": "text", + "content": "4591" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 78, + 72, + 296, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 72, + 296, + 108 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 296, + 108 + ], + "type": "text", + "content": "for cross-domain few-shot learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 6012-6020, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 109, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 109, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 109, + 296, + 167 + ], + "type": "text", + "content": "[42] Fang Peng, Xiaoshan Yang, Linhui Xiao, Yaowei Wang, and Changsheng Xu. Sgva-clip: Semantic-guided visual adapting of vision-language models for few-shot image classification. IEEE Transactions on Multimedia, 26:3469-3480, 2023. 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 170, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 295, + 217 + ], + "type": "text", + "content": "[43] Yixiong Zou, Shuai Yi, Yuhua Li, and Ruixuan Li. A closer look at thecls token for cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:85523-85545, 2024. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 218, + 295, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 295, + 265 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 295, + 265 + ], + "type": "text", + "content": "[44] Yixiong Zou, Ran Ma, Yuhua Li, and Ruixuan Li. Attention temperature matters in vit-based cross-domain few-shot learning. Advances in Neural Information Processing Systems, 37:116332-116354, 2024. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 266, + 295, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 295, + 337 + ], + "type": "text", + "content": "[45] Huali Xu, Li Liu, Shuaifeng Zhi, Shaojing Fu, Zhuo Su, Ming-Ming Cheng, and Yongxiang Liu. Enhancing information maximization with distance-aware contrastive learning for source-free cross-domain few-shot learning. IEEE Transactions on Image Processing, 33:2058-2073, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 339, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 339, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 339, + 295, + 398 + ], + "type": "text", + "content": "[46] Tiange Zhang, Qing Cai, Feng Gao, Lin Qi, and Junyu Dong. Exploring cross-domain few-shot classification via frequency-aware prompting. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 5490–5498, 2024. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 458 + ], + "type": "text", + "content": "[47] Christopher Fifty, Dennis Duan, Ronald Guenther Junkins, Ehsan Amid, Jure Leskovec, Christopher Re, and Sebastian Thrun. Context-aware meta-learning. In The Twelfth International Conference on Learning Representations. 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 543 + ], + "type": "text", + "content": "[48] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995-11005, 2023. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 544, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 578 + ], + "type": "text", + "content": "[49] Yonatan Belinkov. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1):207-219, 2022. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 581, + 295, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 295, + 616 + ], + "type": "text", + "content": "[50] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR 2018, 2018. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 617, + 295, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 617, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 56, + 617, + 295, + 675 + ], + "type": "text", + "content": "[51] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 677, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 295, + 712 + ], + "type": "text", + "content": "[52] Jean-Bastien Grill, Florian Strub, Florent Altché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 336, + 72, + 555, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 72, + 555, + 118 + ], + "spans": [ + { + "bbox": [ + 336, + 72, + 555, + 118 + ], + "type": "text", + "content": "Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 121, + 555, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 121, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 316, + 121, + 555, + 180 + ], + "type": "text", + "content": "[53] Arman Afrasiyabi, Hugo Larochelle, Jean-François Lalonde, and Christian Gagné. Matching feature sets for few-shot image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9014-9024, 2022. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 182, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 182, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 182, + 555, + 252 + ], + "type": "text", + "content": "[54] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 254, + 555, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 555, + 302 + ], + "type": "text", + "content": "[55] Punam Bedi, Pushkar Gole, and Sumit Kumar Agarwal. 18 using deep learning for image-based plant disease detection. _Internet Of things and machine learning in agriculture_, pages 369–402, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 303, + 555, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 303, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 555, + 361 + ], + "type": "text", + "content": "[56] Yunhui Guo, Noel C Codella, Leonid Karlinsky, James V Codella, John R Smith, Kate Saenko, Tajana Rosing, and Rogerio Feris. A broader study of cross-domain few-shot learning. In European conference on computer vision, pages 124-141. Springer, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 363, + 555, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 555, + 422 + ], + "type": "text", + "content": "[57] Hung-Yu Tseng, Hsin-Ying Lee, Jia-Bin Huang, and Ming-Hsuan Yang. Cross-domain few-shot classification via learned feature-wise transformation. In International Conference on Learning Representations, 2019. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 423, + 555, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 423, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 423, + 555, + 483 + ], + "type": "text", + "content": "[58] Pan Li, Shaogang Gong, Chengjie Wang, and Yanwei Fu. Ranking distance calibration for cross-domain few-shot learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9099-9108, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 484, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 484, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 484, + 555, + 567 + ], + "type": "text", + "content": "[59] Nenad Tomasev, Ioana Bica, Brian McWilliams, Lars Holger Buesing, Razvan Pascanu, Charles Blundell, and Jovana Mitrovic. Pushing the limits of self-supervised resnets: Can we outperform supervised learning without labels onImagenet? In First Workshop on Pre-training: Perspectives, Pitfalls, and Paths Forward at ICML 2022. 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 568, + 555, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 555, + 616 + ], + "type": "text", + "content": "[60] Reece Walsh, Islam Osman, and Mohamed S Shehata. Masked embedding modeling with rapid domain adjustment for few-shot image classification. IEEE Transactions on Image Processing, 32:4907-4920, 2023. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 617, + 555, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 617, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 316, + 617, + 555, + 664 + ], + "type": "text", + "content": "[61] Markus Hiller, Rongkai Ma, Mehrtash Harandi, and Tom Drummond. Rethinking generalization in few-shot classification. Advances in Neural Information Processing Systems, 35:3582-3595, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 666, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 666, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 666, + 555, + 712 + ], + "type": "text", + "content": "[62] Yangji He, Weihan Liang, Dongyang Zhao, Hong-Yu Zhou, Weifeng Ge, Yizhou Yu, and Wenqiang Zhang. Attribute surrogates learning and spectral tokens pooling in transformers for few-shot learning. In Proceed-" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "spans": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "text", + "content": "4592" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 78, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 294, + 95 + ], + "type": "text", + "content": "ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9119-9129, 2022. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 98, + 296, + 205 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 98, + 296, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 98, + 296, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 98, + 296, + 144 + ], + "type": "text", + "content": "[63] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. In Advances in neural information processing systems, pages 3630-3638, 2016. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 146, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 146, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 146, + 296, + 205 + ], + "type": "text", + "content": "[64] Andrei A Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. Meta-learning with latent embedding optimization. In International Conference on Learning Representations, 2019. 8" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "spans": [ + { + "bbox": [ + 295, + 738, + 315, + 748 + ], + "type": "text", + "content": "4593" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/23b6091d-1a61-4a72-bde7-0fb203b07465_content_list.json b/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/23b6091d-1a61-4a72-bde7-0fb203b07465_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..373e8ff1724d8e05e05da92431400bb2ea7e69f8 --- /dev/null +++ b/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/23b6091d-1a61-4a72-bde7-0fb203b07465_content_list.json @@ -0,0 +1,1458 @@ +[ + { + "type": "text", + "text": "Towards Efficient General Feature Prediction in Masked Skeleton Modeling", + "text_level": 1, + "bbox": [ + 116, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengkai Sun $^{1}$ , Zefan Zhang $^{2}$ , Jianfeng Dong $^{3}$ , Zhiyong Cheng $^{1*}$ , Xiaojun Chang $^{4}$ , Meng Wang $^{1}$ $^{1}$ Hefei University of Technology $^{2}$ Jilin University", + "bbox": [ + 116, + 179, + 875, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Zhejiang Gongshang University $^{4}$ University of Science and Technology of China", + "bbox": [ + 169, + 215, + 826, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in the masked autoencoder (MAE) paradigm have significantly propelled self-supervised skeleton-based action recognition. However, most existing approaches limit reconstruction targets to raw joint coordinates or their simple variants, resulting in computational redundancy and limited semantic representation. To address this, we propose a novel General Feature Prediction framework (GFP) for efficient mask skeleton modeling. Our key innovation is replacing conventional low-level reconstruction with high-level feature prediction that spans from local motion patterns to global semantic representations. Specifically, we introduce a collaborative learning framework where a lightweight target generation network dynamically produces diversified supervision signals across spatial-temporal hierarchies, avoiding reliance on pre-computed offline features. The framework incorporates constrained optimization to ensure feature diversity while preventing model collapse. Experiments on NTU RGB+D 60, NTU RGB+D 120 and PKU-MMD demonstrate the benefits of our approach: Computational efficiency (with $6.2\\times$ faster training than standard masked skeleton modeling methods) and superior representation quality, achieving state-of-the-art performance in various downstream tasks.", + "bbox": [ + 86, + 301, + 485, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 680, + 220, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding human action [8, 11, 43] has long been a fundamental research task in computer vision, demonstrating significant application value in human-computer interaction, intelligent surveillance, and autonomous driving, etc. In contrast to conventional RGB video representations [13, 28] that densely encode visual information, skeletal representations [17, 24] capture body postures as topological joint sets, employing sparse coordinate sequences to characterize human actions instead of raw pixel values. Owing to the compact data structure and explicit motion representation, skeleton-based action understanding", + "bbox": [ + 89, + 705, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8a708eb943763f3e8f5b1c42f46bc667609a4bafe9a240a77c705891145d6953.jpg", + "image_caption": [ + "Figure 1. Frameworks for general feature prediction (GFP). Instead of low-level targets, GFP predicts the hierarchical abstract representations of input skeleton sequence, embedded by a target generation network, effectively learning more informative high-level features for the downstream tasks. This paradigm achieves an optimal balance between model performance and computational efficiency." + ], + "image_footnote": [], + "bbox": [ + 557, + 271, + 864, + 483 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[4, 25, 26, 36] methods not only significantly reduce computational costs, but also effectively mitigate noise from dynamic backgrounds and lighting changes. However, despite these advantages, early approaches based on fully supervised learning faced notable limitations due to their heavy reliance on manually labeled skeleton data, which requires labor-intensive and costly annotation processes. To overcome this, the research community has increasingly focused on self-supervised learning [6, 30, 44], designing pretext tasks that can automatically learn spatiotemporal patterns in skeleton data without manual labels. The resulting representations can then be efficiently adapted to downstream tasks with a few labeled examples.", + "bbox": [ + 511, + 626, + 906, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised skeleton-based action learning evolved from early handcrafted methods (e.g., colorization, temporal segment rearrangement) to instance discrimination. Contrastive learning [35], a prominent instance discrimination approach, gained dominance due to its simplicity and", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: jason.zy.cheng@gmail.com", + "bbox": [ + 107, + 886, + 385, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12212", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "efficiency. However, current research reveals some limitations in contrastive learning: excessive dependence on data augmentation strategies and persistent false negative sample issues. Recently, transformers have achieved remarkable success in visual tasks [7, 32, 42], and transformer-based masked image modeling has emerged as a novel generative pretraining paradigm, offering fresh insights for self-supervised skeleton action learning.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The masked skeleton modeling methods follow the classic MAE framework by randomly masking some joints in the input skeleton sequence. Its encoder processes only the visible joints, while the decoder reconstructs missing joint coordinates from latent representations and learnable mask tokens. However, due to the intrinsic differences between skeleton and image data, skeleton MAE-like methods significantly differ in implementation. These methods employ a $90\\%$ masking ratio to address temporal redundancy and adopt deeper decoders (instead of MAE's lightweight design) to handle complex spatiotemporal dependencies. Despite their impressive performance, the current design and learning objectives have clear shortcomings: (1) Heavy decoder computation: Due to the reconstruction target involving many joints and the transformer being used as the decoder, excessively long decoding sequences significantly slow down training; (2) Lacking semantic guidance: The low-level target reconstruction objective lacks explicit supervision from higher-level spatiotemporal features like motion semantics, which creates a semantic gap between pretraining and downstream tasks. These limitations motivate our exploration of new masked skeleton modeling paradigms that achieve a balance between representation quality and computational efficiency.", + "bbox": [ + 91, + 217, + 486, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel General Feature Prediction framework (GFP) for efficient mask skeleton modeling. As shown in Figure 1, unlike methods targeting numerous redundant low-level features, GFP focuses on a smaller set of hierarchical high-level features, learning through progressive semantic prediction. Specifically, during decoding, it inputs elements of varying spatiotemporal granularity to predict from local motion patterns (e.g., a joint's movement over 10 frames) to global action semantics. This approach offers two advantages: First, the significantly reduced number of targets cuts computational costs in decoding and optimization. Second, it balances local details and global context, capturing both micromotion changes and overall action structures, thereby substantially improving feature quality. To obtain these high-level semantic features as learning targets, we designed a lightweight target generation network. This network processes elements with varying spatiotemporal granularities, using a multiple feature extraction module to capture both local details and global context, providing hierarchical supervision for the decoder. This supervisory mechanism", + "bbox": [ + 89, + 584, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "operates bidirectionally: The decoder infers high-level semantic features from unmasked joint features for learning, while the target generation network uses the decoder's output as an optimization target, guiding feature representation through a consistency learning manner. However, this design alone is insufficient, as joint optimization can lead to trivial solutions where both networks output constant values regardless of the input. Inspired by the recent information maximization representation learning method [2], we introduced a constraint mechanism to ensure meaningful feature outputs, enabling iterative optimization that converges to a semantically consistent stable state. Ultimately, we developed a flexible mask skeleton modeling framework that significantly accelerates training, enabling the model to learn more discriminative and robust feature representations, yielding superior performance across various downstream tasks.", + "bbox": [ + 511, + 90, + 903, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 531, + 348, + 834, + 362 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We extend conventional mask skeleton modeling from low-level reconstruction to general feature prediction, improving decoding efficiency while introducing local and global semantic guidance.", + "- We design a lightweight target generation network that provides online supervision for the decoder. By introducing an information maximization constraint, we enable collaborative learning between networks, achieving a non-trivial self-supervised skeleton action representation learning scheme.", + "- Extensive experiments across three datasets demonstrate the effectiveness of our approach. The results show that our framework is highly efficient, learns richer semantic information, and outperforms previous methods on multiple downstream tasks." + ], + "bbox": [ + 513, + 364, + 903, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 608, + 653, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Self-supervised Skeleton-based Action Representation Learning", + "text_level": 1, + "bbox": [ + 511, + 633, + 903, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised skeleton-based action representation learning aims to learn discriminative features without manual annotation. The core challenge lies in designing meaningful supervisory signals to guide the model in capturing action semantics.", + "bbox": [ + 511, + 672, + 903, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early approaches to self-supervised skeleton action representation learning predominantly relied on handcrafted pretext tasks within encoder-decoder frameworks. Reconstruction-based methods formed the cornerstone: [44] reconstructed corrupted skeletons, while [14] and [29] employed autoencoders for full-sequence reconstruction. [21] pioneered feature disentanglement into pose/view components before reconstruction. Subsequent innovations leveraged spatio-temporal relationships—[38] predicted joint coloring which is pre-defined by spatial-", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12213", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "temporal order, [5] forecasted inter-frame motions, and [12] estimated multi-scale pose displacements. [37] advanced view invariance through motion retargeting-driven reconstruction.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The paradigm shifted with the advent of contrastive representation learning. [22] established the baseline by contrasting augmented skeleton pairs via dual encoders. [31] enhanced this with skeleton-specific augmentations and cross-architectural contrast. Later works diversified positive pair generation: [15, 40] mined contextual positives, [23] synthesized latent-space positives. Multi-modal fusion emerged through [19] and [30]. [10, 16] introduced guidance mechanisms focusing on action-critical regions to enhance semantic feature learning. [41] pioneered dual-prompted pretraining combining masked reconstruction and feature contrast.", + "bbox": [ + 89, + 151, + 482, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on the success of masked autoencoders, recent works [1, 20, 34] have adapted this paradigm to skeleton-based action representation learning, demonstrating impressive results in capturing action semantics through partial observation and reconstruction.", + "bbox": [ + 89, + 332, + 482, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Masked Modeling for Skeleton-based Action", + "text_level": 1, + "bbox": [ + 89, + 446, + 465, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[34] pioneers the adaptation of Masked Autoencoder to transformer-based skeleton-based action representation learning, employing direct reconstruction of masked joint coordinates as its pretext task. [20] extended SkeletonMAE to motion feature prediction, proposing a motion-aware masking strategy that prioritizes high movement intensity regions during input masking. This input masking forces the model to focus on action-critical patches, achieving substantial performance improvements over conventional random masking approaches. Despite these advancements, the reliance on low-level reconstruction targets inherently limits high-level semantic learning, creating a performance gap in tasks requiring holistic action representation. [1] attempted to address this by using model-generated features as learning targets. However, their patch-level feature prediction necessitates a bulky EMA-updated encoder of massive computations for target generation. While achieving accuracy gains, this approach incurs higher computational costs and much slower convergence than conventional methods.", + "bbox": [ + 89, + 468, + 482, + 770 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By contrast, our method employs spatiotemporally high-level semantic prediction paired with a lightweight target generation network. This design achieves much faster training versus [1] and higher downstream accuracy.", + "bbox": [ + 89, + 771, + 482, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 844, + 181, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first reformulate masked skeleton modeling by replacing its learning objective from conventional", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "patch-level reconstruction to high-level semantic feature prediction. Next, we develop a hierarchical feature prediction framework through multi-granularity representation learning, jointly predicting local details and global semantics. Finally, we design a dedicated target generation network to provide high-quality supervision for prediction tasks, coupled with an information constraint mechanism ensuring stable model optimization.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 220, + 624, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pipeline of skeleton masked autoencoders operates as follows: A predefined masking strategy randomly masks joints, training the network to reconstruct masked joints for representation learning. Specifically, given an input skeleton sequence $X \\in \\mathbb{R}^{T \\times V \\times C}$ , where $T$ is the number of frames, $V$ is the number of joints, and $C$ is the number of channels, temporal patchify first transforms it into $X_{e} \\in \\mathbb{R}^{(T_{e} \\times V) \\times (l \\times C)}$ . This operation segments the temporal streams of every joint into non-overlapping $l$ -frame patches based on consideration of computational efficiency, where $T_{e} = T / l$ . The patches are embedded into feature vectors $E \\in \\mathbb{R}^{(T_{e} \\times V) \\times C_{e}}$ of $C_{e}$ dimension. Most patches are randomly masked, leaving only a small subset visible as input to the transformer encoder for feature extraction. The decoder subsequently predicts masked regions using sequence $E_{N} \\in \\mathbb{R}^{(T_{e} \\times V) \\times C_{e}}$ which combines extracted visible features with learnable mask tokens:", + "bbox": [ + 511, + 241, + 906, + 498 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\frac {1}{N} \\| f (E _ {N}) - X _ {e} \\| _ {F} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 503, + 903, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N = T_{e} \\times V$ , $f(E_{N})$ is the prediction of decoder $f(\\cdot)$ . Prediction targets in the current works extend beyond raw coordinates to include motion trajectories (i.e., frame differences [20]) or patch features from other networks [1].", + "bbox": [ + 511, + 539, + 903, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, such low-level targets share critical limitations: high computational costs and inadequate high-level semantic capture, creating significant gaps for downstream tasks. This drives our extension of prediction targets to different high-level features through:", + "bbox": [ + 511, + 601, + 903, + 676 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\frac {1}{M} \\| f (E _ {N}) - g (X) \\| _ {F} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 683, + 903, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M$ is the number of target features, $g$ is a network to extract various features from the input skeleton action as targets of the decoder, such as action-aware representative features. This high-level feature prediction enhances feature transferability for downstream tasks, particularly those demanding global semantic understanding like recognition and retrieval.", + "bbox": [ + 511, + 719, + 903, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. High-level Feature Prediction", + "text_level": 1, + "bbox": [ + 511, + 833, + 774, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve the above learning objectives, we develop a hierarchical architectural instantiation that strategically balances local motion details with global contextual awareness.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12214", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/daf966f60e946001f06a95ab48c4c027582a00d811479ab56bbb3dc06b40c72c.jpg", + "image_caption": [ + "Figure 2. The framework of our proposed GFP. The GFP framework establishes a bidirectional learning paradigm between two core networks: an encoder-decoder architecture and a Target Generation Network (TGN). The encoder extracts features from partially visible skeleton joints, while the decoder progressively predicts hierarchical high-level features — spanning short-term motion patterns to global action semantics. Simultaneously, the TGN generates semantic targets by encoding transformed inputs and performs consistency learning. This mutual guidance is stabilized through variance-covariance regularization, an information-maximization constraint that ensures feature diversity." + ], + "image_footnote": [], + "bbox": [ + 117, + 97, + 872, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach focuses on selecting learning objectives with richer semantic information across larger spatial-temporal scales, such as the motion patterns of specific nodes over 10-frame sequences. To handle numerous candidate features, we designed a progressive modeling strategy. It starts with short-term local joint patterns, gradually expands to full cycles, and finally integrates local features for holistic action prediction.", + "bbox": [ + 88, + 507, + 480, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, the decoder takes $E_N \\in \\mathbb{R}^{(T_e \\times V) \\times C_e}$ that combines unmasked joint features from the encoder with mask tokens as input. The pipeline first employs temporal average pooling with kernel size $t_1$ , then processes the sequence through a transformer decoder to build the local motion prediction model, producing feature prediction $Z_{p_1} \\in \\mathbb{R}^{((T_e / t_1) \\times V) \\times C_p}$ . Following the initial stage, our hierarchical architecture iteratively scales temporal receptive fields through cascaded processing blocks. For the second stage, temporal average pooling with another kernel size is applied to $Z_{p_1}$ , followed by another transformer decoder that models mid-term dynamics, yielding $Z_{p_2} \\in \\mathbb{R}^{((T_e / t_2) \\times V) \\times C_p}$ . This pyramidal reduction continues until capturing complete action cycles, where the temporal dimension collapses to a singleton, gaining different time granularity feature predictions $\\{Z_{p_1}, Z_{p_2}, \\dots, Z_{p_k}\\}$ . The compressed representation $Z_{p_k} \\in \\mathbb{R}^{V \\times C_p}$ then passes through feature aggregation, followed by an MLP for action-level semantic pre", + "bbox": [ + 93, + 628, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "diction $Z_{p_g}$ . This multi-granularity fusion enables simultaneous preservation of fine-grained joint kinematics and high-level semantic context. Upon obtaining the hierarchical features, our encoder-decoder network learns by minimizing discrepancies between predicted features and target features across hierarchy levels. Given the target set $\\{Z_{t_1}, Z_{t_2}, \\dots, Z_{t_k}, Z_{t_g}\\}$ and $\\mathcal{J} = 1, \\dots, k, g$ , the loss function is formulated as:", + "bbox": [ + 511, + 507, + 903, + 627 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p} = \\frac {1}{M} \\sum_ {j \\in \\mathcal {J}} \\| Z _ {p _ {j}} - Z _ {t _ {j}} \\| _ {F} ^ {2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 637, + 906, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $M$ is the number of all targets. All predicted features are dimensionally aligned to target features through an MLP projector, retaining the original notation for brevity.", + "bbox": [ + 511, + 685, + 906, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is worth noting that previous work predicting patch features from other networks can be viewed as a special case of our method when $t_1 = 1$ with only a single hierarchy level. However, such exclusive reliance on the finest-grained (yet most redundant) local information inevitably leads to significant computational overhead while struggling to capture global contextual patterns.", + "bbox": [ + 511, + 732, + 903, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Target Generation", + "text_level": 1, + "bbox": [ + 511, + 848, + 691, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Semantic Extraction: To obtain the aforementioned high-level semantic targets, we introduce a jointly optimized", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12215", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "target generation network (TGN) composed of multiple lightweight MLPs, where each MLP specializes in producing hierarchical semantic features at specific granularity. Specifically, for each hierarchy level, we adopt simple spatiotemporal aggregation operations that process input signals within corresponding receptive fields: local-level semantic extractors operate on joints that merge across consecutive frames, while the global-level semantic extractor performs holistic feature extraction by integrating all spatiotemporal context. As an example, for the target feature $Z_{t_1}$ , the corresponding input is represented as $X_{t_1} \\in \\mathbb{R}^{((T_e / t_1)\\times V)\\times (t_1\\times l\\times C)}$ .", + "bbox": [ + 89, + 90, + 483, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As for the input of TGN, feeding identical encoder input $\\mathbf{X}_e$ to both networks amplifies bias towards frequently unmasked joints. We devise a simple transformation strategy to get $X_{e} = X_{e}[1:] - X_{e}[: - 1]$ . This geometrically grounded transformation preserves high-level motion semantics while avoiding bias. ur experiments demonstrate the remarkable effectiveness of this parameter-efficient design.", + "bbox": [ + 89, + 275, + 483, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Information Maximization: Under the current framework where the decoder learns from TGN-generated targets while the TGN simultaneously adapts to the decoder's evolving feature representations, this co-learning scheme enables consistent high-level semantic alignment between both networks. However, such bidirectional dependency inevitably risks collapsing into trivial solutions where both networks output constant feature vectors independent of input signals, thereby losing all discriminative power. To prevent this degenerative collapse, we introduce a self-constrained feature space regularization mechanism inspired by VICReg [2], a joint embedding architecture that enforces information-maximization constraints.", + "bbox": [ + 89, + 398, + 483, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Variance Regularization. Taking the global target feature $Z_{t_g} \\in \\mathbb{R}^{B \\times C_t}$ as an example, variance regularization enforces each feature dimension to maintain batch-wise variance above a predefined threshold $\\gamma$ , ensuring persistent diversity in the learned representations and preventing pathological collapse where disparate inputs are mapped to identical feature vectors. Formally, this is implemented through a variance regularization term:", + "bbox": [ + 89, + 597, + 483, + 717 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v a r} = \\frac {1}{C _ {t}} \\sum_ {i = 1} ^ {C _ {t}} \\max \\left(0, \\gamma - \\sqrt {\\operatorname {V a r} \\left(Z _ {t _ {g}} [ : , i ]\\right)}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 734, + 483, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname{Var}(\\cdot)$ computes empirical variance over the batch dimension $B$ . Following [2], $\\gamma$ is set 1 in our experiments.", + "bbox": [ + 89, + 792, + 482, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Covariance Regularization. The covariance constraint drives the batch-wise covariance between every feature dimension pair toward zero, eliminating inter-dimensional correlations to prevent redundant information encoding across channels. This is formalized through a covariance", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "minimization term:", + "bbox": [ + 513, + 92, + 643, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c o v} = \\frac {1}{C _ {t}} \\sum_ {i \\neq j} \\left[ \\operatorname {C o v} \\left(Z _ {t _ {g}}\\right) \\right] _ {i, j} ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 113, + 906, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\text{Cov}(Z_{t_g})$ denotes covariance matrix of $Z_{t_g}$ .", + "bbox": [ + 511, + 160, + 852, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimization Objectives: The final information-maximization constraints are applied to all target features across hierarchy levels through unified regularization:", + "bbox": [ + 511, + 176, + 906, + 222 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} = \\sum_ {j \\in \\mathcal {J}} \\left(\\alpha \\mathcal {L} _ {c o v} \\left(Z _ {t _ {j}}\\right) + \\beta \\mathcal {L} _ {v a r} \\left(Z _ {t _ {j}}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 231, + 906, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are trade-off coefficients. The collaborative optimization objective for both networks combines the prediction loss with information-maximization constraints through weighted aggregation:", + "bbox": [ + 511, + 273, + 905, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda \\mathcal {L} _ {\\text {p r e d}} + \\mathcal {L} _ {\\text {r e g}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 345, + 906, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ denotes trade-off parameters, and $L_{pred}$ denotes the prediction loss computed over the entire batch, i.e., $L_{pred} = \\frac{1}{B}\\sum \\mathcal{L}_p$ . This unified objective drives the coevolution of the decoder and TGN towards both precise hierarchical alignment and collapse-resistant feature extraction.", + "bbox": [ + 511, + 372, + 905, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 476, + 645, + 493 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 501, + 616, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This paper conducts experiments on three widely-used skeleton action benchmarks: NTU RGB+D 60 (NTU-60), NTU RGB+D 120 (NTU-120), and PKU-MMD II (PKU-II). Technical specifications and evaluation protocols for these datasets are summarized as follows.", + "bbox": [ + 511, + 523, + 905, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NTU RGB+D 60 [24] features 60 action classes with 56,880 sequences captured from 40 participants using multi-view cameras. Two standard evaluation settings are adopted: 1) Cross-Subject (x-sub) divides data by participant IDs, using 20 subjects for training and 20 for testing; 2) Cross-View (x-view) employs camera views 2-3 for training and view 1 for testing.", + "bbox": [ + 511, + 598, + 905, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NTU RGB+D 120 [17] extends NTU-60 to 120 action classes with 114,480 sequences from 106 subjects across 32 capture setups. Its evaluation protocols include: 1) Cross-Subject (x-sub) split with 53 subjects each for training/testing; 2) Cross-Setup (x-setup) partitioning using even-numbered setups for training and odd-numbered for testing.", + "bbox": [ + 511, + 704, + 905, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PKU-MMD II [18] focuses on viewpoint-robust action analysis with 51 categories containing 6,952 sequences. The dataset's complexity stems from significant viewpoint diversity. Following established convention, we report results under the standard cross-subject split (5,339 training/1,613 test samples).", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12216", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/87d33a644813c9705529ea0b48c0079f47a22cb02a6d4bb73f9a7852c9e1dd5c.jpg", + "table_caption": [ + "Table 1. Skeleton-based action recognition. Comparisons to the state-of-the-art MAE-like methods for skeleton-based action recognition downstream task on NTU-60. All pre-training runs were completed under identical hardware conditions using a single RTX 4090 GPU. Enc, Dec, and TGN denote the encoder, decoder, and target generation network, respectively." + ], + "table_footnote": [], + "table_body": "
MethodTargetEncDecTGNHoursSpeedupNTU-60
x-subx-view
SkeletonMAE [34]Joint1.97G17.70G-20h27m1×74.877.7
MAMP [20]Motion1.97G17.70G-20h27m1×84.989.1
S-JEPA [1]Patch-level1.97G17.70G28.32G90h57m0.2×85.389.8
GFP (Ours)Hierarchical1.97G1.57G0.64G3h14m6.2×85.992.0
", + "bbox": [ + 122, + 142, + 874, + 256 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2837f32c17e69eb40bfa6a38e62f43dd0776e38eb58532b48c91868e8f02fc4e.jpg", + "table_caption": [ + "Table 2. Skeleton-based action recognition. Comparisons to the state-of-the-art methods for skeleton-based action recognition downstream task on NTU-120 and PKU-MMD II. Bold and underline formatting indicate the best and second-best methods in each group, respectively." + ], + "table_footnote": [], + "table_body": "
MethodNTU-120PKU-II
x-subx-setupx-sub
other pretext task:
P&C [29]42.741.725.5
ISC [31]67.167.936.0
HaLP[23]71.172.243.5
HiCo[6]72.874.149.4
CMD [19]70.371.543.0
UmURL[30]73.574.352.1
USDRL [33]76.678.154.4
masked modeling:
SkeletonMAE[39]72.573.536.1
MAMP [20]78.679.153.8
S-JEPA [1]79.679.953.5
GFP (Ours)79.180.356.2
", + "bbox": [ + 112, + 345, + 464, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consistent with recent literature [19, 31], all experiments on three datasets employ top-1 accuracy as the evaluation metric.", + "bbox": [ + 89, + 612, + 482, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 666, + 308, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Architecture. Follow [1, 20], our model utilizes a standard vision transformer backbone with learnable spatiotemporal position encoding. The encoder stacks $L_{e} = 8$ identical layers, each containing multi-head attention (8 heads, 256-dim) and a feed-forward network (1024 hidden units). For pre-training, the decoder retains the same dimensions but simplifies the architecture: each local feature decoder uses one layer, while the global feature decoder adopts a 2-layer MLP with 2048-dimensional hidden units. The target generation network employs a 3-layer MLP (512 hidden units) for local semantic extraction and a 3-layer MLP (2048 hidden units) for global semantic extraction. All projectors in the decoder are 3-layer MLPs that project features to match the dimensions of the target features.", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ec396af795345b540664ede44c90fbb110b46be40961e58f25910815e753ac4f.jpg", + "table_caption": [ + "Table 3. Comparisons to the state-of-the-art methods for skeleton-based action retrieval. **Bold** indicates the best methods in each group." + ], + "table_footnote": [], + "table_body": "
MethodNTU-60
x-subx-view
other pretext task:
LongT GAN[44]39.148.1
P&C[29]50.776.3
ISC[31]62.582.6
HaLP[23]65.883.6
HiCo[6]68.384.8
SkeAttnCLR[10]69.476.8
UmURL[30]71.388.3
masked modeling:
MAMP [20]62.070.0
GFP (Ours)70.987.1
", + "bbox": [ + 576, + 316, + 839, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Augmentation. For skeleton sequence processing, as in previous works [1, 20], we adopt temporal cropping with dynamic scaling by randomly selecting a continuous clip spanning $50\\% - 100\\%$ of the original sequence length. All clips are resampled to frames via bilinear interpolation to ensure temporal consistency.", + "bbox": [ + 511, + 580, + 906, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pretraining Setup. We employ the motion-aware masking strategy proposed in [20] and segment length $l$ to 4 in the pertaining. Optimization uses AdamW $(\\beta_{1} = 0.9, \\beta_{2} = 0.95$ , weight decay 0.05) under a 400-epoch schedule. Learning rates ramp up linearly from 0 to 1e-3 during 20 warmup epochs, then decay to 5e-4 via cosine annealing. Hyperparameters $\\lambda, \\alpha$ and $\\beta$ are set to 5, 5, 1, respectively. For the target features, we employ a four-level hierarchical feature corresponding to local features at $t_{1} = 5$ , $t_{2} = 10$ , $t_{3} = 30$ and global feature.", + "bbox": [ + 511, + 672, + 908, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison to the State-of-the-art", + "text_level": 1, + "bbox": [ + 511, + 832, + 816, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our method with state-of-the-art approaches on two skeleton-based tasks: action recognition and action retrieval.", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12217", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e59af5232d4892bf3a8f85cf8ec8c98dcc05b8ba18f64ba72296b151d06d289.jpg", + "image_caption": [ + "Figure 3. Impact of Different Targets. The $5,10,30,\\mathrm{g}$ configurations represent local features from 5-, 10-, 30-frame temporal spans and global representations respectively, all derived from high-level representations generated by the TGN. All results are obtained under the NTU-60 x-view protocol." + ], + "image_footnote": [], + "bbox": [ + 106, + 97, + 464, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Skeleton-based Action Recognition. Following the standard practice from previous works [1, 20, 31], we train a linear classifier on top of the frozen encoder pre-trained with our proposed method. First, we compare our method against prior MAE-like approaches [1, 20, 34], on the NTU-60 dataset. Beyond recognition performance, we quantify computational efficiency by measuring computational FLOPs per action sequence for encoder, decoder, and target generation network, full pre-training time under the x-sub protocol (identical hardware), and training speedup relative to SkeletonMAE [34]. As shown in Table 1, our method achieves significant performance gains while substantially accelerating training. This improvement stems from reducing the optimization targets from 750 low-level targets $(\\mathbf{N} = 750)$ in prior works to 251 hierarchical high-level semantic targets $(\\mathbf{M} = 251)$ . Notably, while S-JPEA also adopts learned features as objectives, its patch-level implementation necessitates redundant feature encoding. This design choice, coupled with extremely slow convergence (requiring 1200 epochs versus 400 epochs for other methods), severely degrades training speed compared to SkeletonMAE. These results demonstrate our method's balanced efficacy in feature quality and computational efficiency.", + "bbox": [ + 91, + 416, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further validate our approach on larger-scale NTU-120 and more challenging PKU-MMD II datasets, comparing against methods using alternative pretext tasks. As shown in Table 2, our method outperforms most existing approaches, confirming its broad effectiveness.", + "bbox": [ + 89, + 763, + 482, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Skeleton-based Action Retrieval. In this experiment, the encoder operates in a direct feature extraction mode without requiring any training. Similar to video-to-video retrieval, we use cosine similarity in the feature space to", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f81d9450731cd77fbb267283a53b34246a8794d95d80832db96bfa36c45072ac.jpg", + "image_caption": [ + "Figure 4. Performance of various methods with different targets. All results are obtained under the NTU-60 x-sub protocol." + ], + "image_footnote": [], + "bbox": [ + 524, + 92, + 893, + 267 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "retrieve the nearest action samples for a given query. As detailed in Table 3, conventional MAE-like methods exhibit poor performance on this task due to their focus on local reconstruction objectives without global semantic learning. In contrast, our method achieves substantial improvements (8.9% on x-sub and 17.1% on x-view protocols) over prior approaches like MAMP by incorporating high-level semantic objectives. Notably, our results remain competitive with contrastive learning methods [6, 30] that explicitly optimize global representations, demonstrating the superiority of high-level semantic targets over low-level reconstruction objectives.", + "bbox": [ + 511, + 324, + 906, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 515, + 666, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All ablation studies were conducted on the NTU-60 dataset for skeleton-based action recognition.", + "bbox": [ + 511, + 537, + 906, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Prediction Targets: Our method employs local joint features (5-, 10-, and 30-frame spans) and global semantic features as learning targets. To investigate their individual impacts, we progressively ablated these components. As shown in Figure 3, sequential removal of global and local semantic learning objectives caused performance degradation, demonstrating the complementary nature of hierarchical feature learning. Furthermore, we explored using low-level targets instead of learned representations by removing the TGN from our framework, which forces direct prediction vectors of all spatiotemporal information within targets (e.g., flattening input into a single vector for global target). The corresponding results (labeled $\\{5,10,30,g\\}^*$ in Figure 3) reveal significant performance drops, proving that high-level semantics cannot be effectively captured through direct prediction of low-level signals.", + "bbox": [ + 511, + 568, + 908, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Input of Target Generation Network: For the target network inputs, we evaluated three variants: (1) raw joint coordinates, (2) masked joint data, and (3) motion features derived via linear transformation of joints. As shown in Table 4, the motion features achieved optimal performance, validating our design choice. Notably, masked", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12218", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4e4b49aa793f299068f8a46f161e187ee2a22f3b53f2de786073e0b8b195a499.jpg", + "table_caption": [ + "Table 4. Ablation study on the input of target generation network." + ], + "table_footnote": [], + "table_body": "
InputNTU 60
x-subx-view
joint85.090.9
masked joint84.290.3
motion85.992.0
", + "bbox": [ + 99, + 128, + 331, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/908205a8c1879a65bcf7a66f17bb758047b48fc67e7f10a222ab529f55b6994a.jpg", + "table_caption": [ + "Table 5. Ablation study on the architecture of target generation network." + ], + "table_footnote": [], + "table_body": "
Arch.NTU 60
x-subx-view
Base85.992.0
Shadow85.891.8
Wider85.891.3
", + "bbox": [ + 395, + 128, + 596, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/01dd9592bbb0a05d1ddf1c415936f1df6c32c2105fd2ac67199d729c7fbf3cc5.jpg", + "table_caption": [ + "Table 6. Ablation study on different number of layers used in projector." + ], + "table_footnote": [], + "table_body": "
LayersNTU 60
x-subx-view
285.591.3
385.992.0
485.691.6
", + "bbox": [ + 679, + 128, + 874, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/31ff49715f46f248315fcfaa65deb27ee353d0a6267e3544094ad41a2e60b525.jpg", + "table_caption": [ + "Table 7. Comparisons to the state-of-the-art methods with semi-supervised learning on NTU-60 dataset." + ], + "table_footnote": [], + "table_body": "
Methodx-subx-view
1%10%1%10%
ASSL [27]-64.3-69.8
ISC [31]35.765.938.172.5
Colorization [38]48.371.752.578.9
Hi-TRS [3]-70.7-74.8
GL-Transformer [12])-68.6-74.9
HaLP [23]46.672.648.777.1
CMD [19]50.675.453.080.2
CPM [40]56.773.057.577.1
HYSP [9]-76.2-80.4
\\( PCM^3 \\) [41])53.877.753.182.8
HiCo [6]54.473.054.878.3
UmURL [30]58.1-58.3-
USDRL [33]57.380.260.784.0
SkeletonMAE [34]54.480.654.683.5
MAMP [20]66.088.068.791.5
S-JEPA [1]67.588.469.191.4
GFP (Ours)71.888.772.992.1
", + "bbox": [ + 101, + 273, + 464, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "inputs hinder semantic feature extraction by TGN, aligning with observations in prior work [1]. We evaluate the epoch-wise performance of different methods with varying learning objectives: SkeletonMAE learns to reconstruct raw joints, MAMP targets motion features (i.e., temporal difference of joint coordinates), and our method predicts high-level semantic representations. As shown in Figure 4, our approach consistently outperforms prior works across all training epochs, demonstrating the superiority of hierarchical semantic learning over low-level reconstruction objectives.", + "bbox": [ + 88, + 657, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Architecture Design: We conducted architectural experiments on the target generation network and projector. For TGN: (1)Shallow variant: 2-layer MLP; (2)Wider variant: 4-layer MLP with doubled hidden dimensions. (3)Base: Standard 3-layer MLP. As shown in Table 5,", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the baseline 3-layer configuration achieves optimal performance, though all variants show marginal differences, demonstrating TGN's robustness to architectural variations. For the projector, we evaluated MLPs with 2-4 layers. The 3-layer configuration yields superior performance (Table 6).", + "bbox": [ + 511, + 237, + 903, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Extended Performance Validation", + "text_level": 1, + "bbox": [ + 511, + 323, + 808, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To assess representation transferability, we conduct limited-label experiments on NTU-60 under two data settings: $1\\%$ and $5\\%$ labeled training data. Following semi-supervised learning benchmarks [31], we initialize the model with weights pre-trained through our model, then attach and fine-tune a classifier using available labels.", + "bbox": [ + 511, + 345, + 905, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Table 7, our approach establishes new state-of-the-art performance across both data scales. This significant performance gap confirms the effectiveness of our learned representations in label-efficient scenarios.", + "bbox": [ + 511, + 436, + 905, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 511, + 633, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an efficient general feature prediction framework for masked skeleton modeling. Unlike conventional approaches that reconstruct low-level targets, our method employs hierarchical high-level semantic features as learning targets, significantly reducing decoder complexity. The framework integrates a target generation network that provides online learning objectives. We further introduce the information maximization constraints, enabling non-trivial co-optimization between networks. This mechanism simultaneously preserves local motion details and global semantics. Extensive experiments across three datasets demonstrate the effectiveness of our method, achieving state-of-the-art performance with significant training speedup over prior masked skeleton modeling approaches.", + "bbox": [ + 511, + 537, + 906, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 777, + 671, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the Pioneer and Leading Goose R&D Program of Zhejiang (No. 2024C01110), National Natural Science Foundation of China (No. 72188101, 62020106007, and 62472385). The computation is completed on the HPC Platform of Hefei University of Technology.", + "bbox": [ + 511, + 801, + 905, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12219", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mohamed Abdelfattah and Alexandre Alahi. S-jepa: A joint embedding predictive architecture for skeletal action recognition. In European Conference on Computer Vision, pages 367-384. Springer, 2024. 3, 6, 7, 8", + "[2] Adrien Bardes, Jean Ponce, and Yann Lecun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR 2022-International Conference on Learning Representations, 2022. 2, 5", + "[3] Yuxiao Chen, Long Zhao, Jianbo Yuan, Yu Tian, Zhaoyang Xia, Shijie Geng, Ligong Han, and Dimitris N Metaxas. Hierarchically self-supervised transformer for human skeleton representation learning. In European Conference on Computer Vision, pages 185-202, 2022. 8", + "[4] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183-192, 2020. 1", + "[5] Yi-Bin Cheng, Xipeng Chen, Junhong Chen, Pengxu Wei, Dongyu Zhang, and Liang Lin. Hierarchical transformer: Unsupervised representation learning for skeleton-based human action recognition. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2021. 3", + "[6] Jianfeng Dong, Shengkai Sun, Zhonglin Liu, Shujie Chen, Baolong Liu, and Xun Wang. Hierarchical contrast for unsupervised skeleton-based action representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 525-533, 2023. 1, 6, 7, 8", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1", + "[9] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. arXiv preprint arXiv:2303.06242, 2023.8", + "[10] Yilei Hua, Wenhan Wu, Ce Zheng, Aidong Lu, Mengyuan Liu, Chen Chen, and Shiqian Wu. Part aware contrastive learning for self-supervised action recognition. In International Joint Conference on Artificial Intelligence, 2023. 3, 6", + "[11] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence, 35(1):221-231, 2012. 1", + "[12] Boeun Kim, Hyung Jin Chang, Jungho Kim, and Jin Young Choi. Global-local motion transformer for unsupervised skeleton-based action learning. In European Conference on Computer Vision, pages 209-225, 2022. 3, 8" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. 1", + "[14] Jogendra Nath Kundu, Maharshi Gor, Phani Krishna Uppala, and Venkatesh Babu Radhakrishnan. Unsupervised feature learning of human actions as trajectories in pose embedding manifold. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1459-1467, 2019. 2", + "[15] Linguuo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4741-4750, 2021. 3", + "[16] Lilang Lin, Jiahang Zhang, and Jiaying Liu. Actionlet-dependent contrastive learning for unsupervised skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2363-2372, 2023. 3", + "[17] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684-2701, 2019. 1, 5", + "[18] Jiaying Liu, Sijie Song, Chunhui Liu, Yanghao Li, and Yueyu Hu. A benchmark dataset and comparison study for multi-modal human action analytics. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 16(2):1-24, 2020. 5", + "[19] Yunyao Mao, Wengang Zhou, Zhenbo Lu, Jiajun Deng, and Houqiang Li. Cmd: Self-supervised 3d action representation learning with cross-modal mutual distillation. In European Conference on Computer Vision, pages 734-752. Springer, 2022. 3, 6, 8", + "[20] Yunyao Mao, Jiajun Deng, Wengang Zhou, Yao Fang, Wanli Ouyang, and Houqiang Li. Masked motion predictors are strong 3d action representation learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10191, 2023. 3, 6, 7, 8", + "[21] Qiang Nie, Ziwei Liu, and Yunhui Liu. Unsupervised 3d human pose representation with viewpoint and pose disentanglement. In European Conference on Computer Vision, pages 102-118, 2020. 2", + "[22] Haocong Rao, Shihao Xu, Xiping Hu, Jun Cheng, and Bin Hu. Augmented skeleton based contrastive action learning with momentum lstm for unsupervised action recognition. Information Sciences, 569:90-109, 2021. 3", + "[23] Anshul Shah, Aniket Roy, Ketul Shah, Shlok Mishra, David Jacobs, Anoop Cherian, and Rama Chellappa. Halp: Hallucinating latent positives for skeleton-based self-supervised learning of actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18846-18856, 2023. 3, 6, 8", + "[24] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE Conference on Computer" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12220", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition, pages 1010-1019, 2016. 1, 5", + "[25] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7912-7921, 2019. 1", + "[26] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12026-12035, 2019. 1", + "[27] Chenyang Si, Xuecheng Nie, Wei Wang, Liang Wang, Ti-niu Tan, and Jiashi Feng. Adversarial self-supervised learning for semi-supervised 3d action recognition. In European Conference on Computer Vision, pages 35-51, 2020. 8", + "[28] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1", + "[29] Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9631-9640, 2020. 2, 6", + "[30] Shengkai Sun, Daizong Liu, Jianfeng Dong, Xiaoye Qu, Junyu Gao, Xun Yang, Xun Wang, and Meng Wang. Unified multi-modal unsupervised representation learning for skeleton-based action understanding. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2973-2984, 2023. 1, 3, 6, 7, 8", + "[31] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1655–1663, 2021. 3, 6, 7, 8", + "[32] Fei Wang, Dan Guo, Kun Li, and Meng Wang. Eulermormer: Robust eulerian motion magnification via dynamic filtering within transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5345-5353, 2024. 2", + "[33] Wanjiang Weng, Hongsong Wang, Junbo He, Lei He, and Guosen Xie. Usdrl: Unified skeleton-based dense representation learning with multi-grained feature decorrelation. arXiv preprint arXiv:2412.09220, 2024. 6, 8", + "[34] Wenhan Wu, Yilei Hua, Ce Zheng, Shiqian Wu, Chen Chen, and Aidong Lu. Skeletonmae: Spatial-temporal masked autoencoders for self-supervised skeleton action recognition. In 2023 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 224-229. IEEE, 2023. 3, 6, 7, 8", + "[35] Yuankun Xu, Dong Huang, Chang-Dong Wang, and Jian-Huang Lai. Deep image clustering with contrastive learning and multi-scale graph convolutional networks. Pattern Recognition, 2024. 1", + "[36] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI Conference on Artificial Intelligence, 2018. 1", + "[37] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and François Brémond. View-invariant skeleton action representation learning via motion" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "retargeting. International Journal of Computer Vision, pages 1-16, 2024. 3", + "[38] Siyuan Yang, Jun Liu, Shijian Lu, Meng Hwa Er, and Alex C Kot. Skeleton cloud colorization for unsupervised 3d action representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13423-13433, 2021. 2, 8", + "[39] Siyuan Yang, Jun Liu, Shijian Lu, Er Meng Hwa, Yongjian Hu, and Alex C Kot. Self-supervised 3d action representation learning with skeleton cloud colorization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 6", + "[40] Haoyuan Zhang, Yonghong Hou, Wenjing Zhang, and Wanqing Li. Contrastive positive mining for unsupervised 3d action representation learning. In European Conference on Computer Vision, pages 36-51. Springer, 2022. 3, 8", + "[41] Jiahang Zhang, Lilang Lin, and Jiaying Liu. Prompted contrast with masked motion modeling: Towards versatile 3d action representation learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7175-7183, 2023. 3, 8", + "[42] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber: Tubelet transformer for video action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13598-13607, 2022. 2", + "[43] Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In Proceedings of the IEEE international conference on computer vision, pages 2914-2923, 2017. 1", + "[44] Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2644-2651, 2018. 1, 2, 6" + ], + "bbox": [ + 516, + 92, + 903, + 613 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12221", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/23b6091d-1a61-4a72-bde7-0fb203b07465_model.json b/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/23b6091d-1a61-4a72-bde7-0fb203b07465_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1c1bf2537b56001d38e80280c67b23db8c706567 --- /dev/null +++ b/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/23b6091d-1a61-4a72-bde7-0fb203b07465_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.117, + 0.131, + 0.882, + 0.154 + ], + "angle": 0, + "content": "Towards Efficient General Feature Prediction in Masked Skeleton Modeling" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.18, + 0.877, + 0.217 + ], + "angle": 0, + "content": "Shengkai Sun\\(^{1}\\), Zefan Zhang\\(^{2}\\), Jianfeng Dong\\(^{3}\\), Zhiyong Cheng\\(^{1*}\\), Xiaojun Chang\\(^{4}\\), Meng Wang\\(^{1}\\) \n\\(^{1}\\)Hefei University of Technology \\(^{2}\\)Jilin University" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.217, + 0.828, + 0.236 + ], + "angle": 0, + "content": "\\(^{3}\\)Zhejiang Gongshang University \\(^{4}\\)University of Science and Technology of China" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.302, + 0.486, + 0.651 + ], + "angle": 0, + "content": "Recent advances in the masked autoencoder (MAE) paradigm have significantly propelled self-supervised skeleton-based action recognition. However, most existing approaches limit reconstruction targets to raw joint coordinates or their simple variants, resulting in computational redundancy and limited semantic representation. To address this, we propose a novel General Feature Prediction framework (GFP) for efficient mask skeleton modeling. Our key innovation is replacing conventional low-level reconstruction with high-level feature prediction that spans from local motion patterns to global semantic representations. Specifically, we introduce a collaborative learning framework where a lightweight target generation network dynamically produces diversified supervision signals across spatial-temporal hierarchies, avoiding reliance on pre-computed offline features. The framework incorporates constrained optimization to ensure feature diversity while preventing model collapse. Experiments on NTU RGB+D 60, NTU RGB+D 120 and PKU-MMD demonstrate the benefits of our approach: Computational efficiency (with \\(6.2\\times\\) faster training than standard masked skeleton modeling methods) and superior representation quality, achieving state-of-the-art performance in various downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.222, + 0.697 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Understanding human action [8, 11, 43] has long been a fundamental research task in computer vision, demonstrating significant application value in human-computer interaction, intelligent surveillance, and autonomous driving, etc. In contrast to conventional RGB video representations [13, 28] that densely encode visual information, skeletal representations [17, 24] capture body postures as topological joint sets, employing sparse coordinate sequences to characterize human actions instead of raw pixel values. Owing to the compact data structure and explicit motion representation, skeleton-based action understanding" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.272, + 0.865, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.498, + 0.908, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Frameworks for general feature prediction (GFP). Instead of low-level targets, GFP predicts the hierarchical abstract representations of input skeleton sequence, embedded by a target generation network, effectively learning more informative high-level features for the downstream tasks. This paradigm achieves an optimal balance between model performance and computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.824 + ], + "angle": 0, + "content": "[4, 25, 26, 36] methods not only significantly reduce computational costs, but also effectively mitigate noise from dynamic backgrounds and lighting changes. However, despite these advantages, early approaches based on fully supervised learning faced notable limitations due to their heavy reliance on manually labeled skeleton data, which requires labor-intensive and costly annotation processes. To overcome this, the research community has increasingly focused on self-supervised learning [6, 30, 44], designing pretext tasks that can automatically learn spatiotemporal patterns in skeleton data without manual labels. The resulting representations can then be efficiently adapted to downstream tasks with a few labeled examples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Self-supervised skeleton-based action learning evolved from early handcrafted methods (e.g., colorization, temporal segment rearrangement) to instance discrimination. Contrastive learning [35], a prominent instance discrimination approach, gained dominance due to its simplicity and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.386, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: jason.zy.cheng@gmail.com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12212" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "efficiency. However, current research reveals some limitations in contrastive learning: excessive dependence on data augmentation strategies and persistent false negative sample issues. Recently, transformers have achieved remarkable success in visual tasks [7, 32, 42], and transformer-based masked image modeling has emerged as a novel generative pretraining paradigm, offering fresh insights for self-supervised skeleton action learning." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.218, + 0.487, + 0.581 + ], + "angle": 0, + "content": "The masked skeleton modeling methods follow the classic MAE framework by randomly masking some joints in the input skeleton sequence. Its encoder processes only the visible joints, while the decoder reconstructs missing joint coordinates from latent representations and learnable mask tokens. However, due to the intrinsic differences between skeleton and image data, skeleton MAE-like methods significantly differ in implementation. These methods employ a \\(90\\%\\) masking ratio to address temporal redundancy and adopt deeper decoders (instead of MAE's lightweight design) to handle complex spatiotemporal dependencies. Despite their impressive performance, the current design and learning objectives have clear shortcomings: (1) Heavy decoder computation: Due to the reconstruction target involving many joints and the transformer being used as the decoder, excessively long decoding sequences significantly slow down training; (2) Lacking semantic guidance: The low-level target reconstruction objective lacks explicit supervision from higher-level spatiotemporal features like motion semantics, which creates a semantic gap between pretraining and downstream tasks. These limitations motivate our exploration of new masked skeleton modeling paradigms that achieve a balance between representation quality and computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose a novel General Feature Prediction framework (GFP) for efficient mask skeleton modeling. As shown in Figure 1, unlike methods targeting numerous redundant low-level features, GFP focuses on a smaller set of hierarchical high-level features, learning through progressive semantic prediction. Specifically, during decoding, it inputs elements of varying spatiotemporal granularity to predict from local motion patterns (e.g., a joint's movement over 10 frames) to global action semantics. This approach offers two advantages: First, the significantly reduced number of targets cuts computational costs in decoding and optimization. Second, it balances local details and global context, capturing both micromotion changes and overall action structures, thereby substantially improving feature quality. To obtain these high-level semantic features as learning targets, we designed a lightweight target generation network. This network processes elements with varying spatiotemporal granularities, using a multiple feature extraction module to capture both local details and global context, providing hierarchical supervision for the decoder. This supervisory mechanism" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.346 + ], + "angle": 0, + "content": "operates bidirectionally: The decoder infers high-level semantic features from unmasked joint features for learning, while the target generation network uses the decoder's output as an optimization target, guiding feature representation through a consistency learning manner. However, this design alone is insufficient, as joint optimization can lead to trivial solutions where both networks output constant values regardless of the input. Inspired by the recent information maximization representation learning method [2], we introduced a constraint mechanism to ensure meaningful feature outputs, enabling iterative optimization that converges to a semantically consistent stable state. Ultimately, we developed a flexible mask skeleton modeling framework that significantly accelerates training, enabling the model to learn more discriminative and robust feature representations, yielding superior performance across various downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.349, + 0.836, + 0.363 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.366, + 0.905, + 0.425 + ], + "angle": 0, + "content": "- We extend conventional mask skeleton modeling from low-level reconstruction to general feature prediction, improving decoding efficiency while introducing local and global semantic guidance." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.426, + 0.905, + 0.516 + ], + "angle": 0, + "content": "- We design a lightweight target generation network that provides online supervision for the decoder. By introducing an information maximization constraint, we enable collaborative learning between networks, achieving a non-trivial self-supervised skeleton action representation learning scheme." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.517, + 0.905, + 0.591 + ], + "angle": 0, + "content": "- Extensive experiments across three datasets demonstrate the effectiveness of our approach. The results show that our framework is highly efficient, learns richer semantic information, and outperforms previous methods on multiple downstream tasks." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.366, + 0.905, + 0.591 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.609, + 0.655, + 0.625 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.905, + 0.667 + ], + "angle": 0, + "content": "2.1. Self-supervised Skeleton-based Action Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.905, + 0.748 + ], + "angle": 0, + "content": "Self-supervised skeleton-based action representation learning aims to learn discriminative features without manual annotation. The core challenge lies in designing meaningful supervisory signals to guide the model in capturing action semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Early approaches to self-supervised skeleton action representation learning predominantly relied on handcrafted pretext tasks within encoder-decoder frameworks. Reconstruction-based methods formed the cornerstone: [44] reconstructed corrupted skeletons, while [14] and [29] employed autoencoders for full-sequence reconstruction. [21] pioneered feature disentanglement into pose/view components before reconstruction. Subsequent innovations leveraged spatio-temporal relationships—[38] predicted joint coloring which is pre-defined by spatial-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12213" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "temporal order, [5] forecasted inter-frame motions, and [12] estimated multi-scale pose displacements. [37] advanced view invariance through motion retargeting-driven reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.152, + 0.483, + 0.332 + ], + "angle": 0, + "content": "The paradigm shifted with the advent of contrastive representation learning. [22] established the baseline by contrasting augmented skeleton pairs via dual encoders. [31] enhanced this with skeleton-specific augmentations and cross-architectural contrast. Later works diversified positive pair generation: [15, 40] mined contextual positives, [23] synthesized latent-space positives. Multi-modal fusion emerged through [19] and [30]. [10, 16] introduced guidance mechanisms focusing on action-critical regions to enhance semantic feature learning. [41] pioneered dual-prompted pretraining combining masked reconstruction and feature contrast." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.483, + 0.409 + ], + "angle": 0, + "content": "Building on the success of masked autoencoders, recent works [1, 20, 34] have adapted this paradigm to skeleton-based action representation learning, demonstrating impressive results in capturing action semantics through partial observation and reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.448, + 0.467, + 0.464 + ], + "angle": 0, + "content": "2.2. Masked Modeling for Skeleton-based Action" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[34] pioneers the adaptation of Masked Autoencoder to transformer-based skeleton-based action representation learning, employing direct reconstruction of masked joint coordinates as its pretext task. [20] extended SkeletonMAE to motion feature prediction, proposing a motion-aware masking strategy that prioritizes high movement intensity regions during input masking. This input masking forces the model to focus on action-critical patches, achieving substantial performance improvements over conventional random masking approaches. Despite these advancements, the reliance on low-level reconstruction targets inherently limits high-level semantic learning, creating a performance gap in tasks requiring holistic action representation. [1] attempted to address this by using model-generated features as learning targets. However, their patch-level feature prediction necessitates a bulky EMA-updated encoder of massive computations for target generation. While achieving accuracy gains, this approach incurs higher computational costs and much slower convergence than conventional methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.833 + ], + "angle": 0, + "content": "By contrast, our method employs spatiotemporally high-level semantic prediction paired with a lightweight target generation network. This design achieves much faster training versus [1] and higher downstream accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.182, + 0.86 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we first reformulate masked skeleton modeling by replacing its learning objective from conventional" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "patch-level reconstruction to high-level semantic feature prediction. Next, we develop a hierarchical feature prediction framework through multi-granularity representation learning, jointly predicting local details and global semantics. Finally, we design a dedicated target generation network to provide high-quality supervision for prediction tasks, coupled with an information constraint mechanism ensuring stable model optimization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.221, + 0.625, + 0.235 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.242, + 0.907, + 0.499 + ], + "angle": 0, + "content": "The pipeline of skeleton masked autoencoders operates as follows: A predefined masking strategy randomly masks joints, training the network to reconstruct masked joints for representation learning. Specifically, given an input skeleton sequence \\( X \\in \\mathbb{R}^{T \\times V \\times C} \\), where \\( T \\) is the number of frames, \\( V \\) is the number of joints, and \\( C \\) is the number of channels, temporal patchify first transforms it into \\( X_{e} \\in \\mathbb{R}^{(T_{e} \\times V) \\times (l \\times C)} \\). This operation segments the temporal streams of every joint into non-overlapping \\( l \\)-frame patches based on consideration of computational efficiency, where \\( T_{e} = T / l \\). The patches are embedded into feature vectors \\( E \\in \\mathbb{R}^{(T_{e} \\times V) \\times C_{e}} \\) of \\( C_{e} \\) dimension. Most patches are randomly masked, leaving only a small subset visible as input to the transformer encoder for feature extraction. The decoder subsequently predicts masked regions using sequence \\( E_{N} \\in \\mathbb{R}^{(T_{e} \\times V) \\times C_{e}} \\) which combines extracted visible features with learnable mask tokens:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.504, + 0.905, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\frac {1}{N} \\| f (E _ {N}) - X _ {e} \\| _ {F} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.54, + 0.905, + 0.601 + ], + "angle": 0, + "content": "where \\(N = T_{e} \\times V\\), \\(f(E_{N})\\) is the prediction of decoder \\(f(\\cdot)\\). Prediction targets in the current works extend beyond raw coordinates to include motion trajectories (i.e., frame differences [20]) or patch features from other networks [1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.602, + 0.905, + 0.677 + ], + "angle": 0, + "content": "However, such low-level targets share critical limitations: high computational costs and inadequate high-level semantic capture, creating significant gaps for downstream tasks. This drives our extension of prediction targets to different high-level features through:" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.684, + 0.905, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\frac {1}{M} \\| f (E _ {N}) - g (X) \\| _ {F} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.825 + ], + "angle": 0, + "content": "where \\( M \\) is the number of target features, \\( g \\) is a network to extract various features from the input skeleton action as targets of the decoder, such as action-aware representative features. This high-level feature prediction enhances feature transferability for downstream tasks, particularly those demanding global semantic understanding like recognition and retrieval." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.776, + 0.849 + ], + "angle": 0, + "content": "3.2. High-level Feature Prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To achieve the above learning objectives, we develop a hierarchical architectural instantiation that strategically balances local motion details with global contextual awareness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12214" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.098, + 0.873, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.398, + 0.908, + 0.482 + ], + "angle": 0, + "content": "Figure 2. The framework of our proposed GFP. The GFP framework establishes a bidirectional learning paradigm between two core networks: an encoder-decoder architecture and a Target Generation Network (TGN). The encoder extracts features from partially visible skeleton joints, while the decoder progressively predicts hierarchical high-level features — spanning short-term motion patterns to global action semantics. Simultaneously, the TGN generates semantic targets by encoding transformed inputs and performs consistency learning. This mutual guidance is stabilized through variance-covariance regularization, an information-maximization constraint that ensures feature diversity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.482, + 0.629 + ], + "angle": 0, + "content": "Our approach focuses on selecting learning objectives with richer semantic information across larger spatial-temporal scales, such as the motion patterns of specific nodes over 10-frame sequences. To handle numerous candidate features, we designed a progressive modeling strategy. It starts with short-term local joint patterns, gradually expands to full cycles, and finally integrates local features for holistic action prediction." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.63, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Formally, the decoder takes \\( E_N \\in \\mathbb{R}^{(T_e \\times V) \\times C_e} \\) that combines unmasked joint features from the encoder with mask tokens as input. The pipeline first employs temporal average pooling with kernel size \\( t_1 \\), then processes the sequence through a transformer decoder to build the local motion prediction model, producing feature prediction \\( Z_{p_1} \\in \\mathbb{R}^{((T_e / t_1) \\times V) \\times C_p} \\). Following the initial stage, our hierarchical architecture iteratively scales temporal receptive fields through cascaded processing blocks. For the second stage, temporal average pooling with another kernel size is applied to \\( Z_{p_1} \\), followed by another transformer decoder that models mid-term dynamics, yielding \\( Z_{p_2} \\in \\mathbb{R}^{((T_e / t_2) \\times V) \\times C_p} \\). This pyramidal reduction continues until capturing complete action cycles, where the temporal dimension collapses to a singleton, gaining different time granularity feature predictions \\( \\{Z_{p_1}, Z_{p_2}, \\dots, Z_{p_k}\\} \\). The compressed representation \\( Z_{p_k} \\in \\mathbb{R}^{V \\times C_p} \\) then passes through feature aggregation, followed by an MLP for action-level semantic pre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.905, + 0.628 + ], + "angle": 0, + "content": "diction \\( Z_{p_g} \\). This multi-granularity fusion enables simultaneous preservation of fine-grained joint kinematics and high-level semantic context. Upon obtaining the hierarchical features, our encoder-decoder network learns by minimizing discrepancies between predicted features and target features across hierarchy levels. Given the target set \\( \\{Z_{t_1}, Z_{t_2}, \\dots, Z_{t_k}, Z_{t_g}\\} \\) and \\( \\mathcal{J} = 1, \\dots, k, g \\), the loss function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.638, + 0.907, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p} = \\frac {1}{M} \\sum_ {j \\in \\mathcal {J}} \\| Z _ {p _ {j}} - Z _ {t _ {j}} \\| _ {F} ^ {2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.907, + 0.732 + ], + "angle": 0, + "content": "where \\( M \\) is the number of all targets. All predicted features are dimensionally aligned to target features through an MLP projector, retaining the original notation for brevity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.905, + 0.839 + ], + "angle": 0, + "content": "It is worth noting that previous work predicting patch features from other networks can be viewed as a special case of our method when \\( t_1 = 1 \\) with only a single hierarchy level. However, such exclusive reliance on the finest-grained (yet most redundant) local information inevitably leads to significant computational overhead while struggling to capture global contextual patterns." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.692, + 0.864 + ], + "angle": 0, + "content": "3.3. Target Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Semantic Extraction: To obtain the aforementioned high-level semantic targets, we introduce a jointly optimized" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12215" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.271 + ], + "angle": 0, + "content": "target generation network (TGN) composed of multiple lightweight MLPs, where each MLP specializes in producing hierarchical semantic features at specific granularity. Specifically, for each hierarchy level, we adopt simple spatiotemporal aggregation operations that process input signals within corresponding receptive fields: local-level semantic extractors operate on joints that merge across consecutive frames, while the global-level semantic extractor performs holistic feature extraction by integrating all spatiotemporal context. As an example, for the target feature \\( Z_{t_1} \\), the corresponding input is represented as \\( X_{t_1} \\in \\mathbb{R}^{((T_e / t_1)\\times V)\\times (t_1\\times l\\times C)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.484, + 0.396 + ], + "angle": 0, + "content": "As for the input of TGN, feeding identical encoder input \\(\\mathbf{X}_e\\) to both networks amplifies bias towards frequently unmasked joints. We devise a simple transformation strategy to get \\(X_{e} = X_{e}[1:] - X_{e}[: - 1]\\). This geometrically grounded transformation preserves high-level motion semantics while avoiding bias. ur experiments demonstrate the remarkable effectiveness of this parameter-efficient design." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Information Maximization: Under the current framework where the decoder learns from TGN-generated targets while the TGN simultaneously adapts to the decoder's evolving feature representations, this co-learning scheme enables consistent high-level semantic alignment between both networks. However, such bidirectional dependency inevitably risks collapsing into trivial solutions where both networks output constant feature vectors independent of input signals, thereby losing all discriminative power. To prevent this degenerative collapse, we introduce a self-constrained feature space regularization mechanism inspired by VICReg [2], a joint embedding architecture that enforces information-maximization constraints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Variance Regularization. Taking the global target feature \\( Z_{t_g} \\in \\mathbb{R}^{B \\times C_t} \\) as an example, variance regularization enforces each feature dimension to maintain batch-wise variance above a predefined threshold \\( \\gamma \\), ensuring persistent diversity in the learned representations and preventing pathological collapse where disparate inputs are mapped to identical feature vectors. Formally, this is implemented through a variance regularization term:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.735, + 0.484, + 0.777 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v a r} = \\frac {1}{C _ {t}} \\sum_ {i = 1} ^ {C _ {t}} \\max \\left(0, \\gamma - \\sqrt {\\operatorname {V a r} \\left(Z _ {t _ {g}} [ : , i ]\\right)}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.483, + 0.824 + ], + "angle": 0, + "content": "where \\(\\operatorname{Var}(\\cdot)\\) computes empirical variance over the batch dimension \\(B\\). Following [2], \\(\\gamma\\) is set 1 in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Covariance Regularization. The covariance constraint drives the batch-wise covariance between every feature dimension pair toward zero, eliminating inter-dimensional correlations to prevent redundant information encoding across channels. This is formalized through a covariance" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.645, + 0.106 + ], + "angle": 0, + "content": "minimization term:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.114, + 0.907, + 0.151 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c o v} = \\frac {1}{C _ {t}} \\sum_ {i \\neq j} \\left[ \\operatorname {C o v} \\left(Z _ {t _ {g}}\\right) \\right] _ {i, j} ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.161, + 0.854, + 0.178 + ], + "angle": 0, + "content": "where \\( \\text{Cov}(Z_{t_g}) \\) denotes covariance matrix of \\( Z_{t_g} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.178, + 0.907, + 0.223 + ], + "angle": 0, + "content": "Optimization Objectives: The final information-maximization constraints are applied to all target features across hierarchy levels through unified regularization:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.232, + 0.907, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} = \\sum_ {j \\in \\mathcal {J}} \\left(\\alpha \\mathcal {L} _ {c o v} \\left(Z _ {t _ {j}}\\right) + \\beta \\mathcal {L} _ {v a r} \\left(Z _ {t _ {j}}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.275, + 0.906, + 0.336 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are trade-off coefficients. The collaborative optimization objective for both networks combines the prediction loss with information-maximization constraints through weighted aggregation:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.347, + 0.907, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda \\mathcal {L} _ {\\text {p r e d}} + \\mathcal {L} _ {\\text {r e g}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.373, + 0.906, + 0.463 + ], + "angle": 0, + "content": "where \\(\\lambda\\) denotes trade-off parameters, and \\(L_{pred}\\) denotes the prediction loss computed over the entire batch, i.e., \\(L_{pred} = \\frac{1}{B}\\sum \\mathcal{L}_p\\). This unified objective drives the coevolution of the decoder and TGN towards both precise hierarchical alignment and collapse-resistant feature extraction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.477, + 0.647, + 0.494 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.502, + 0.617, + 0.516 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.598 + ], + "angle": 0, + "content": "This paper conducts experiments on three widely-used skeleton action benchmarks: NTU RGB+D 60 (NTU-60), NTU RGB+D 120 (NTU-120), and PKU-MMD II (PKU-II). Technical specifications and evaluation protocols for these datasets are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.704 + ], + "angle": 0, + "content": "NTU RGB+D 60 [24] features 60 action classes with 56,880 sequences captured from 40 participants using multi-view cameras. Two standard evaluation settings are adopted: 1) Cross-Subject (x-sub) divides data by participant IDs, using 20 subjects for training and 20 for testing; 2) Cross-View (x-view) employs camera views 2-3 for training and view 1 for testing." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.705, + 0.906, + 0.81 + ], + "angle": 0, + "content": "NTU RGB+D 120 [17] extends NTU-60 to 120 action classes with 114,480 sequences from 106 subjects across 32 capture setups. Its evaluation protocols include: 1) Cross-Subject (x-sub) split with 53 subjects each for training/testing; 2) Cross-Setup (x-setup) partitioning using even-numbered setups for training and odd-numbered for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "PKU-MMD II [18] focuses on viewpoint-robust action analysis with 51 categories containing 6,952 sequences. The dataset's complexity stems from significant viewpoint diversity. Following established convention, we report results under the standard cross-subject split (5,339 training/1,613 test samples)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12216" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.133 + ], + "angle": 0, + "content": "Table 1. Skeleton-based action recognition. Comparisons to the state-of-the-art MAE-like methods for skeleton-based action recognition downstream task on NTU-60. All pre-training runs were completed under identical hardware conditions using a single RTX 4090 GPU. Enc, Dec, and TGN denote the encoder, decoder, and target generation network, respectively." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.143, + 0.875, + 0.257 + ], + "angle": 0, + "content": "
MethodTargetEncDecTGNHoursSpeedupNTU-60
x-subx-view
SkeletonMAE [34]Joint1.97G17.70G-20h27m1×74.877.7
MAMP [20]Motion1.97G17.70G-20h27m1×84.989.1
S-JEPA [1]Patch-level1.97G17.70G28.32G90h57m0.2×85.389.8
GFP (Ours)Hierarchical1.97G1.57G0.64G3h14m6.2×85.992.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.265, + 0.486, + 0.336 + ], + "angle": 0, + "content": "Table 2. Skeleton-based action recognition. Comparisons to the state-of-the-art methods for skeleton-based action recognition downstream task on NTU-120 and PKU-MMD II. Bold and underline formatting indicate the best and second-best methods in each group, respectively." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.346, + 0.465, + 0.602 + ], + "angle": 0, + "content": "
MethodNTU-120PKU-II
x-subx-setupx-sub
other pretext task:
P&C [29]42.741.725.5
ISC [31]67.167.936.0
HaLP[23]71.172.243.5
HiCo[6]72.874.149.4
CMD [19]70.371.543.0
UmURL[30]73.574.352.1
USDRL [33]76.678.154.4
masked modeling:
SkeletonMAE[39]72.573.536.1
MAMP [20]78.679.153.8
S-JEPA [1]79.679.953.5
GFP (Ours)79.180.356.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.658 + ], + "angle": 0, + "content": "Consistent with recent literature [19, 31], all experiments on three datasets employ top-1 accuracy as the evaluation metric." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.31, + 0.684 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Model Architecture. Follow [1, 20], our model utilizes a standard vision transformer backbone with learnable spatiotemporal position encoding. The encoder stacks \\( L_{e} = 8 \\) identical layers, each containing multi-head attention (8 heads, 256-dim) and a feed-forward network (1024 hidden units). For pre-training, the decoder retains the same dimensions but simplifies the architecture: each local feature decoder uses one layer, while the global feature decoder adopts a 2-layer MLP with 2048-dimensional hidden units. The target generation network employs a 3-layer MLP (512 hidden units) for local semantic extraction and a 3-layer MLP (2048 hidden units) for global semantic extraction. All projectors in the decoder are 3-layer MLPs that project features to match the dimensions of the target features." + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.265, + 0.908, + 0.308 + ], + "angle": 0, + "content": "Table 3. Comparisons to the state-of-the-art methods for skeleton-based action retrieval. **Bold** indicates the best methods in each group." + }, + { + "type": "table", + "bbox": [ + 0.578, + 0.318, + 0.84, + 0.557 + ], + "angle": 0, + "content": "
MethodNTU-60
x-subx-view
other pretext task:
LongT GAN[44]39.148.1
P&C[29]50.776.3
ISC[31]62.582.6
HaLP[23]65.883.6
HiCo[6]68.384.8
SkeAttnCLR[10]69.476.8
UmURL[30]71.388.3
masked modeling:
MAMP [20]62.070.0
GFP (Ours)70.987.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.673 + ], + "angle": 0, + "content": "Data Augmentation. For skeleton sequence processing, as in previous works [1, 20], we adopt temporal cropping with dynamic scaling by randomly selecting a continuous clip spanning \\(50\\% - 100\\%\\) of the original sequence length. All clips are resampled to frames via bilinear interpolation to ensure temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Pretraining Setup. We employ the motion-aware masking strategy proposed in [20] and segment length \\(l\\) to 4 in the pertaining. Optimization uses AdamW \\((\\beta_{1} = 0.9, \\beta_{2} = 0.95\\), weight decay 0.05) under a 400-epoch schedule. Learning rates ramp up linearly from 0 to 1e-3 during 20 warmup epochs, then decay to 5e-4 via cosine annealing. Hyperparameters \\(\\lambda, \\alpha\\) and \\(\\beta\\) are set to 5, 5, 1, respectively. For the target features, we employ a four-level hierarchical feature corresponding to local features at \\(t_{1} = 5\\), \\(t_{2} = 10\\), \\(t_{3} = 30\\) and global feature." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.818, + 0.85 + ], + "angle": 0, + "content": "4.3. Comparison to the State-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "In this section, we compare our method with state-of-the-art approaches on two skeleton-based tasks: action recognition and action retrieval." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12217" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.098, + 0.465, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.32, + 0.487, + 0.392 + ], + "angle": 0, + "content": "Figure 3. Impact of Different Targets. The \\(5,10,30,\\mathrm{g}\\) configurations represent local features from 5-, 10-, 30-frame temporal spans and global representations respectively, all derived from high-level representations generated by the TGN. All results are obtained under the NTU-60 x-view protocol." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.417, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Skeleton-based Action Recognition. Following the standard practice from previous works [1, 20, 31], we train a linear classifier on top of the frozen encoder pre-trained with our proposed method. First, we compare our method against prior MAE-like approaches [1, 20, 34], on the NTU-60 dataset. Beyond recognition performance, we quantify computational efficiency by measuring computational FLOPs per action sequence for encoder, decoder, and target generation network, full pre-training time under the x-sub protocol (identical hardware), and training speedup relative to SkeletonMAE [34]. As shown in Table 1, our method achieves significant performance gains while substantially accelerating training. This improvement stems from reducing the optimization targets from 750 low-level targets \\((\\mathbf{N} = 750)\\) in prior works to 251 hierarchical high-level semantic targets \\((\\mathbf{M} = 251)\\). Notably, while S-JPEA also adopts learned features as objectives, its patch-level implementation necessitates redundant feature encoding. This design choice, coupled with extremely slow convergence (requiring 1200 epochs versus 400 epochs for other methods), severely degrades training speed compared to SkeletonMAE. These results demonstrate our method's balanced efficacy in feature quality and computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.841 + ], + "angle": 0, + "content": "We further validate our approach on larger-scale NTU-120 and more challenging PKU-MMD II datasets, comparing against methods using alternative pretext tasks. As shown in Table 2, our method outperforms most existing approaches, confirming its broad effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Skeleton-based Action Retrieval. In this experiment, the encoder operates in a direct feature extraction mode without requiring any training. Similar to video-to-video retrieval, we use cosine similarity in the feature space to" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.093, + 0.895, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.27, + 0.905, + 0.298 + ], + "angle": 0, + "content": "Figure 4. Performance of various methods with different targets. All results are obtained under the NTU-60 x-sub protocol." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.908, + 0.506 + ], + "angle": 0, + "content": "retrieve the nearest action samples for a given query. As detailed in Table 3, conventional MAE-like methods exhibit poor performance on this task due to their focus on local reconstruction objectives without global semantic learning. In contrast, our method achieves substantial improvements (8.9% on x-sub and 17.1% on x-view protocols) over prior approaches like MAMP by incorporating high-level semantic objectives. Notably, our results remain competitive with contrastive learning methods [6, 30] that explicitly optimize global representations, demonstrating the superiority of high-level semantic targets over low-level reconstruction objectives." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.516, + 0.667, + 0.532 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.538, + 0.907, + 0.568 + ], + "angle": 0, + "content": "All ablation studies were conducted on the NTU-60 dataset for skeleton-based action recognition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.81 + ], + "angle": 0, + "content": "Prediction Targets: Our method employs local joint features (5-, 10-, and 30-frame spans) and global semantic features as learning targets. To investigate their individual impacts, we progressively ablated these components. As shown in Figure 3, sequential removal of global and local semantic learning objectives caused performance degradation, demonstrating the complementary nature of hierarchical feature learning. Furthermore, we explored using low-level targets instead of learned representations by removing the TGN from our framework, which forces direct prediction vectors of all spatiotemporal information within targets (e.g., flattening input into a single vector for global target). The corresponding results (labeled \\(\\{5,10,30,g\\}^*\\) in Figure 3) reveal significant performance drops, proving that high-level semantics cannot be effectively captured through direct prediction of low-level signals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Input of Target Generation Network: For the target network inputs, we evaluated three variants: (1) raw joint coordinates, (2) masked joint data, and (3) motion features derived via linear transformation of joints. As shown in Table 4, the motion features achieved optimal performance, validating our design choice. Notably, masked" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12218" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.342, + 0.118 + ], + "angle": 0, + "content": "Table 4. Ablation study on the input of target generation network." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.129, + 0.332, + 0.228 + ], + "angle": 0, + "content": "
InputNTU 60
x-subx-view
joint85.090.9
masked joint84.290.3
motion85.992.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.372, + 0.089, + 0.621, + 0.118 + ], + "angle": 0, + "content": "Table 5. Ablation study on the architecture of target generation network." + }, + { + "type": "table", + "bbox": [ + 0.397, + 0.129, + 0.597, + 0.228 + ], + "angle": 0, + "content": "
Arch.NTU 60
x-subx-view
Base85.992.0
Shadow85.891.8
Wider85.891.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.653, + 0.089, + 0.901, + 0.118 + ], + "angle": 0, + "content": "Table 6. Ablation study on different number of layers used in projector." + }, + { + "type": "table", + "bbox": [ + 0.68, + 0.129, + 0.875, + 0.228 + ], + "angle": 0, + "content": "
LayersNTU 60
x-subx-view
285.591.3
385.992.0
485.691.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.235, + 0.483, + 0.265 + ], + "angle": 0, + "content": "Table 7. Comparisons to the state-of-the-art methods with semi-supervised learning on NTU-60 dataset." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.275, + 0.465, + 0.649 + ], + "angle": 0, + "content": "
Methodx-subx-view
1%10%1%10%
ASSL [27]-64.3-69.8
ISC [31]35.765.938.172.5
Colorization [38]48.371.752.578.9
Hi-TRS [3]-70.7-74.8
GL-Transformer [12])-68.6-74.9
HaLP [23]46.672.648.777.1
CMD [19]50.675.453.080.2
CPM [40]56.773.057.577.1
HYSP [9]-76.2-80.4
\\( PCM^3 \\) [41])53.877.753.182.8
HiCo [6]54.473.054.878.3
UmURL [30]58.1-58.3-
USDRL [33]57.380.260.784.0
SkeletonMAE [34]54.480.654.683.5
MAMP [20]66.088.068.791.5
S-JEPA [1]67.588.469.191.4
GFP (Ours)71.888.772.992.1
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.825 + ], + "angle": 0, + "content": "inputs hinder semantic feature extraction by TGN, aligning with observations in prior work [1]. We evaluate the epoch-wise performance of different methods with varying learning objectives: SkeletonMAE learns to reconstruct raw joints, MAMP targets motion features (i.e., temporal difference of joint coordinates), and our method predicts high-level semantic representations. As shown in Figure 4, our approach consistently outperforms prior works across all training epochs, demonstrating the superiority of hierarchical semantic learning over low-level reconstruction objectives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Architecture Design: We conducted architectural experiments on the target generation network and projector. For TGN: (1)Shallow variant: 2-layer MLP; (2)Wider variant: 4-layer MLP with doubled hidden dimensions. (3)Base: Standard 3-layer MLP. As shown in Table 5," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.238, + 0.905, + 0.315 + ], + "angle": 0, + "content": "the baseline 3-layer configuration achieves optimal performance, though all variants show marginal differences, demonstrating TGN's robustness to architectural variations. For the projector, we evaluated MLPs with 2-4 layers. The 3-layer configuration yields superior performance (Table 6)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.324, + 0.81, + 0.339 + ], + "angle": 0, + "content": "4.5. Extended Performance Validation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.906, + 0.436 + ], + "angle": 0, + "content": "To assess representation transferability, we conduct limited-label experiments on NTU-60 under two data settings: \\(1\\%\\) and \\(5\\%\\) labeled training data. Following semi-supervised learning benchmarks [31], we initialize the model with weights pre-trained through our model, then attach and fine-tune a classifier using available labels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.906, + 0.498 + ], + "angle": 0, + "content": "As shown in Table 7, our approach establishes new state-of-the-art performance across both data scales. This significant performance gap confirms the effectiveness of our learned representations in label-efficient scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.512, + 0.634, + 0.527 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.765 + ], + "angle": 0, + "content": "In this paper, we propose an efficient general feature prediction framework for masked skeleton modeling. Unlike conventional approaches that reconstruct low-level targets, our method employs hierarchical high-level semantic features as learning targets, significantly reducing decoder complexity. The framework integrates a target generation network that provides online learning objectives. We further introduce the information maximization constraints, enabling non-trivial co-optimization between networks. This mechanism simultaneously preserves local motion details and global semantics. Extensive experiments across three datasets demonstrate the effectiveness of our method, achieving state-of-the-art performance with significant training speedup over prior masked skeleton modeling approaches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.779, + 0.673, + 0.796 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.906, + 0.889 + ], + "angle": 0, + "content": "This work was supported by the Pioneer and Leading Goose R&D Program of Zhejiang (No. 2024C01110), National Natural Science Foundation of China (No. 72188101, 62020106007, and 62472385). The computation is completed on the HPC Platform of Hefei University of Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12219" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Mohamed Abdelfattah and Alexandre Alahi. S-jepa: A joint embedding predictive architecture for skeletal action recognition. In European Conference on Computer Vision, pages 367-384. Springer, 2024. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Adrien Bardes, Jean Ponce, and Yann Lecun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR 2022-International Conference on Learning Representations, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.297 + ], + "angle": 0, + "content": "[3] Yuxiao Chen, Long Zhao, Jianbo Yuan, Yu Tian, Zhaoyang Xia, Shijie Geng, Ligong Han, and Dimitris N Metaxas. Hierarchically self-supervised transformer for human skeleton representation learning. In European Conference on Computer Vision, pages 185-202, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.484, + 0.367 + ], + "angle": 0, + "content": "[4] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183-192, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.449 + ], + "angle": 0, + "content": "[5] Yi-Bin Cheng, Xipeng Chen, Junhong Chen, Pengxu Wei, Dongyu Zhang, and Liang Lin. Hierarchical transformer: Unsupervised representation learning for skeleton-based human action recognition. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.483, + 0.521 + ], + "angle": 0, + "content": "[6] Jianfeng Dong, Shengkai Sun, Zhonglin Liu, Shujie Chen, Baolong Liu, and Xun Wang. Hierarchical contrast for unsupervised skeleton-based action representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 525-533, 2023. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.661 + ], + "angle": 0, + "content": "[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[9] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. arXiv preprint arXiv:2303.06242, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[10] Yilei Hua, Wenhan Wu, Ce Zheng, Aidong Lu, Mengyuan Liu, Chen Chen, and Shiqian Wu. Part aware contrastive learning for self-supervised action recognition. In International Joint Conference on Artificial Intelligence, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[11] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence, 35(1):221-231, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Boeun Kim, Hyung Jin Chang, Jungho Kim, and Jin Young Choi. Global-local motion transformer for unsupervised skeleton-based action learning. In European Conference on Computer Vision, pages 209-225, 2022. 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[13] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[14] Jogendra Nath Kundu, Maharshi Gor, Phani Krishna Uppala, and Venkatesh Babu Radhakrishnan. Unsupervised feature learning of human actions as trajectories in pose embedding manifold. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1459-1467, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[15] Linguuo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4741-4750, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.305, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[16] Lilang Lin, Jiahang Zhang, and Jiaying Liu. Actionlet-dependent contrastive learning for unsupervised skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2363-2372, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[17] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684-2701, 2019. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[18] Jiaying Liu, Sijie Song, Chunhui Liu, Yanghao Li, and Yueyu Hu. A benchmark dataset and comparison study for multi-modal human action analytics. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 16(2):1-24, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.518, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[19] Yunyao Mao, Wengang Zhou, Zhenbo Lu, Jiajun Deng, and Houqiang Li. Cmd: Self-supervised 3d action representation learning with cross-modal mutual distillation. In European Conference on Computer Vision, pages 734-752. Springer, 2022. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[20] Yunyao Mao, Jiajun Deng, Wengang Zhou, Yao Fang, Wanli Ouyang, and Houqiang Li. Masked motion predictors are strong 3d action representation learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10191, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.66, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[21] Qiang Nie, Ziwei Liu, and Yunhui Liu. Unsupervised 3d human pose representation with viewpoint and pose disentanglement. In European Conference on Computer Vision, pages 102-118, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[22] Haocong Rao, Shihao Xu, Xiping Hu, Jun Cheng, and Bin Hu. Augmented skeleton based contrastive action learning with momentum lstm for unsupervised action recognition. Information Sciences, 569:90-109, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[23] Anshul Shah, Aniket Roy, Ketul Shah, Shlok Mishra, David Jacobs, Anoop Cherian, and Rama Chellappa. Halp: Hallucinating latent positives for skeleton-based self-supervised learning of actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18846-18856, 2023. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12220" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Vision and Pattern Recognition, pages 1010-1019, 2016. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[25] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7912-7921, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[26] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12026-12035, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[27] Chenyang Si, Xuecheng Nie, Wei Wang, Liang Wang, Ti-niu Tan, and Jiashi Feng. Adversarial self-supervised learning for semi-supervised 3d action recognition. In European Conference on Computer Vision, pages 35-51, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[28] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[29] Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9631-9640, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[30] Shengkai Sun, Daizong Liu, Jianfeng Dong, Xiaoye Qu, Junyu Gao, Xun Yang, Xun Wang, and Meng Wang. Unified multi-modal unsupervised representation learning for skeleton-based action understanding. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2973-2984, 2023. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[31] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1655–1663, 2021. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[32] Fei Wang, Dan Guo, Kun Li, and Meng Wang. Eulermormer: Robust eulerian motion magnification via dynamic filtering within transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5345-5353, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[33] Wanjiang Weng, Hongsong Wang, Junbo He, Lei He, and Guosen Xie. Usdrl: Unified skeleton-based dense representation learning with multi-grained feature decorrelation. arXiv preprint arXiv:2412.09220, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[34] Wenhan Wu, Yilei Hua, Ce Zheng, Shiqian Wu, Chen Chen, and Aidong Lu. Skeletonmae: Spatial-temporal masked autoencoders for self-supervised skeleton action recognition. In 2023 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 224-229. IEEE, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[35] Yuankun Xu, Dong Huang, Chang-Dong Wang, and Jian-Huang Lai. Deep image clustering with contrastive learning and multi-scale graph convolutional networks. Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[36] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI Conference on Artificial Intelligence, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and François Brémond. View-invariant skeleton action representation learning via motion" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "retargeting. International Journal of Computer Vision, pages 1-16, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[38] Siyuan Yang, Jun Liu, Shijian Lu, Meng Hwa Er, and Alex C Kot. Skeleton cloud colorization for unsupervised 3d action representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13423-13433, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[39] Siyuan Yang, Jun Liu, Shijian Lu, Er Meng Hwa, Yongjian Hu, and Alex C Kot. Self-supervised 3d action representation learning with skeleton cloud colorization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[40] Haoyuan Zhang, Yonghong Hou, Wenjing Zhang, and Wanqing Li. Contrastive positive mining for unsupervised 3d action representation learning. In European Conference on Computer Vision, pages 36-51. Springer, 2022. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[41] Jiahang Zhang, Lilang Lin, and Jiaying Liu. Prompted contrast with masked motion modeling: Towards versatile 3d action representation learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7175-7183, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.471 + ], + "angle": 0, + "content": "[42] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber: Tubelet transformer for video action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13598-13607, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.542 + ], + "angle": 0, + "content": "[43] Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In Proceedings of the IEEE international conference on computer vision, pages 2914-2923, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.544, + 0.905, + 0.614 + ], + "angle": 0, + "content": "[44] Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2644-2651, 2018. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12221" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/23b6091d-1a61-4a72-bde7-0fb203b07465_origin.pdf b/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/23b6091d-1a61-4a72-bde7-0fb203b07465_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..be8826a542f09615bbc38d9f6918b348ddf79aca --- /dev/null +++ b/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/23b6091d-1a61-4a72-bde7-0fb203b07465_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d38fda31aed5797073ac8cb4badbc917fdf63623b520724caad16bd78dcae86 +size 416558 diff --git a/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/full.md b/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..09d056086c2d9f78f3a8ec3b9c32d1e33f27b12e --- /dev/null +++ b/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/full.md @@ -0,0 +1,289 @@ +# Towards Efficient General Feature Prediction in Masked Skeleton Modeling + +Shengkai Sun $^{1}$ , Zefan Zhang $^{2}$ , Jianfeng Dong $^{3}$ , Zhiyong Cheng $^{1*}$ , Xiaojun Chang $^{4}$ , Meng Wang $^{1}$ $^{1}$ Hefei University of Technology $^{2}$ Jilin University + +$^{3}$ Zhejiang Gongshang University $^{4}$ University of Science and Technology of China + +# Abstract + +Recent advances in the masked autoencoder (MAE) paradigm have significantly propelled self-supervised skeleton-based action recognition. However, most existing approaches limit reconstruction targets to raw joint coordinates or their simple variants, resulting in computational redundancy and limited semantic representation. To address this, we propose a novel General Feature Prediction framework (GFP) for efficient mask skeleton modeling. Our key innovation is replacing conventional low-level reconstruction with high-level feature prediction that spans from local motion patterns to global semantic representations. Specifically, we introduce a collaborative learning framework where a lightweight target generation network dynamically produces diversified supervision signals across spatial-temporal hierarchies, avoiding reliance on pre-computed offline features. The framework incorporates constrained optimization to ensure feature diversity while preventing model collapse. Experiments on NTU RGB+D 60, NTU RGB+D 120 and PKU-MMD demonstrate the benefits of our approach: Computational efficiency (with $6.2\times$ faster training than standard masked skeleton modeling methods) and superior representation quality, achieving state-of-the-art performance in various downstream tasks. + +# 1. Introduction + +Understanding human action [8, 11, 43] has long been a fundamental research task in computer vision, demonstrating significant application value in human-computer interaction, intelligent surveillance, and autonomous driving, etc. In contrast to conventional RGB video representations [13, 28] that densely encode visual information, skeletal representations [17, 24] capture body postures as topological joint sets, employing sparse coordinate sequences to characterize human actions instead of raw pixel values. Owing to the compact data structure and explicit motion representation, skeleton-based action understanding + +![](images/8a708eb943763f3e8f5b1c42f46bc667609a4bafe9a240a77c705891145d6953.jpg) +Figure 1. Frameworks for general feature prediction (GFP). Instead of low-level targets, GFP predicts the hierarchical abstract representations of input skeleton sequence, embedded by a target generation network, effectively learning more informative high-level features for the downstream tasks. This paradigm achieves an optimal balance between model performance and computational efficiency. + +[4, 25, 26, 36] methods not only significantly reduce computational costs, but also effectively mitigate noise from dynamic backgrounds and lighting changes. However, despite these advantages, early approaches based on fully supervised learning faced notable limitations due to their heavy reliance on manually labeled skeleton data, which requires labor-intensive and costly annotation processes. To overcome this, the research community has increasingly focused on self-supervised learning [6, 30, 44], designing pretext tasks that can automatically learn spatiotemporal patterns in skeleton data without manual labels. The resulting representations can then be efficiently adapted to downstream tasks with a few labeled examples. + +Self-supervised skeleton-based action learning evolved from early handcrafted methods (e.g., colorization, temporal segment rearrangement) to instance discrimination. Contrastive learning [35], a prominent instance discrimination approach, gained dominance due to its simplicity and + +efficiency. However, current research reveals some limitations in contrastive learning: excessive dependence on data augmentation strategies and persistent false negative sample issues. Recently, transformers have achieved remarkable success in visual tasks [7, 32, 42], and transformer-based masked image modeling has emerged as a novel generative pretraining paradigm, offering fresh insights for self-supervised skeleton action learning. + +The masked skeleton modeling methods follow the classic MAE framework by randomly masking some joints in the input skeleton sequence. Its encoder processes only the visible joints, while the decoder reconstructs missing joint coordinates from latent representations and learnable mask tokens. However, due to the intrinsic differences between skeleton and image data, skeleton MAE-like methods significantly differ in implementation. These methods employ a $90\%$ masking ratio to address temporal redundancy and adopt deeper decoders (instead of MAE's lightweight design) to handle complex spatiotemporal dependencies. Despite their impressive performance, the current design and learning objectives have clear shortcomings: (1) Heavy decoder computation: Due to the reconstruction target involving many joints and the transformer being used as the decoder, excessively long decoding sequences significantly slow down training; (2) Lacking semantic guidance: The low-level target reconstruction objective lacks explicit supervision from higher-level spatiotemporal features like motion semantics, which creates a semantic gap between pretraining and downstream tasks. These limitations motivate our exploration of new masked skeleton modeling paradigms that achieve a balance between representation quality and computational efficiency. + +In this paper, we propose a novel General Feature Prediction framework (GFP) for efficient mask skeleton modeling. As shown in Figure 1, unlike methods targeting numerous redundant low-level features, GFP focuses on a smaller set of hierarchical high-level features, learning through progressive semantic prediction. Specifically, during decoding, it inputs elements of varying spatiotemporal granularity to predict from local motion patterns (e.g., a joint's movement over 10 frames) to global action semantics. This approach offers two advantages: First, the significantly reduced number of targets cuts computational costs in decoding and optimization. Second, it balances local details and global context, capturing both micromotion changes and overall action structures, thereby substantially improving feature quality. To obtain these high-level semantic features as learning targets, we designed a lightweight target generation network. This network processes elements with varying spatiotemporal granularities, using a multiple feature extraction module to capture both local details and global context, providing hierarchical supervision for the decoder. This supervisory mechanism + +operates bidirectionally: The decoder infers high-level semantic features from unmasked joint features for learning, while the target generation network uses the decoder's output as an optimization target, guiding feature representation through a consistency learning manner. However, this design alone is insufficient, as joint optimization can lead to trivial solutions where both networks output constant values regardless of the input. Inspired by the recent information maximization representation learning method [2], we introduced a constraint mechanism to ensure meaningful feature outputs, enabling iterative optimization that converges to a semantically consistent stable state. Ultimately, we developed a flexible mask skeleton modeling framework that significantly accelerates training, enabling the model to learn more discriminative and robust feature representations, yielding superior performance across various downstream tasks. + +Our contributions are summarized as follows: + +- We extend conventional mask skeleton modeling from low-level reconstruction to general feature prediction, improving decoding efficiency while introducing local and global semantic guidance. +- We design a lightweight target generation network that provides online supervision for the decoder. By introducing an information maximization constraint, we enable collaborative learning between networks, achieving a non-trivial self-supervised skeleton action representation learning scheme. +- Extensive experiments across three datasets demonstrate the effectiveness of our approach. The results show that our framework is highly efficient, learns richer semantic information, and outperforms previous methods on multiple downstream tasks. + +# 2. Related Work + +# 2.1. Self-supervised Skeleton-based Action Representation Learning + +Self-supervised skeleton-based action representation learning aims to learn discriminative features without manual annotation. The core challenge lies in designing meaningful supervisory signals to guide the model in capturing action semantics. + +Early approaches to self-supervised skeleton action representation learning predominantly relied on handcrafted pretext tasks within encoder-decoder frameworks. Reconstruction-based methods formed the cornerstone: [44] reconstructed corrupted skeletons, while [14] and [29] employed autoencoders for full-sequence reconstruction. [21] pioneered feature disentanglement into pose/view components before reconstruction. Subsequent innovations leveraged spatio-temporal relationships—[38] predicted joint coloring which is pre-defined by spatial- + +temporal order, [5] forecasted inter-frame motions, and [12] estimated multi-scale pose displacements. [37] advanced view invariance through motion retargeting-driven reconstruction. + +The paradigm shifted with the advent of contrastive representation learning. [22] established the baseline by contrasting augmented skeleton pairs via dual encoders. [31] enhanced this with skeleton-specific augmentations and cross-architectural contrast. Later works diversified positive pair generation: [15, 40] mined contextual positives, [23] synthesized latent-space positives. Multi-modal fusion emerged through [19] and [30]. [10, 16] introduced guidance mechanisms focusing on action-critical regions to enhance semantic feature learning. [41] pioneered dual-prompted pretraining combining masked reconstruction and feature contrast. + +Building on the success of masked autoencoders, recent works [1, 20, 34] have adapted this paradigm to skeleton-based action representation learning, demonstrating impressive results in capturing action semantics through partial observation and reconstruction. + +# 2.2. Masked Modeling for Skeleton-based Action + +[34] pioneers the adaptation of Masked Autoencoder to transformer-based skeleton-based action representation learning, employing direct reconstruction of masked joint coordinates as its pretext task. [20] extended SkeletonMAE to motion feature prediction, proposing a motion-aware masking strategy that prioritizes high movement intensity regions during input masking. This input masking forces the model to focus on action-critical patches, achieving substantial performance improvements over conventional random masking approaches. Despite these advancements, the reliance on low-level reconstruction targets inherently limits high-level semantic learning, creating a performance gap in tasks requiring holistic action representation. [1] attempted to address this by using model-generated features as learning targets. However, their patch-level feature prediction necessitates a bulky EMA-updated encoder of massive computations for target generation. While achieving accuracy gains, this approach incurs higher computational costs and much slower convergence than conventional methods. + +By contrast, our method employs spatiotemporally high-level semantic prediction paired with a lightweight target generation network. This design achieves much faster training versus [1] and higher downstream accuracy. + +# 3. Method + +In this section, we first reformulate masked skeleton modeling by replacing its learning objective from conventional + +patch-level reconstruction to high-level semantic feature prediction. Next, we develop a hierarchical feature prediction framework through multi-granularity representation learning, jointly predicting local details and global semantics. Finally, we design a dedicated target generation network to provide high-quality supervision for prediction tasks, coupled with an information constraint mechanism ensuring stable model optimization. + +# 3.1. Overview + +The pipeline of skeleton masked autoencoders operates as follows: A predefined masking strategy randomly masks joints, training the network to reconstruct masked joints for representation learning. Specifically, given an input skeleton sequence $X \in \mathbb{R}^{T \times V \times C}$ , where $T$ is the number of frames, $V$ is the number of joints, and $C$ is the number of channels, temporal patchify first transforms it into $X_{e} \in \mathbb{R}^{(T_{e} \times V) \times (l \times C)}$ . This operation segments the temporal streams of every joint into non-overlapping $l$ -frame patches based on consideration of computational efficiency, where $T_{e} = T / l$ . The patches are embedded into feature vectors $E \in \mathbb{R}^{(T_{e} \times V) \times C_{e}}$ of $C_{e}$ dimension. Most patches are randomly masked, leaving only a small subset visible as input to the transformer encoder for feature extraction. The decoder subsequently predicts masked regions using sequence $E_{N} \in \mathbb{R}^{(T_{e} \times V) \times C_{e}}$ which combines extracted visible features with learnable mask tokens: + +$$ +\mathcal {L} _ {p} = \frac {1}{N} \| f (E _ {N}) - X _ {e} \| _ {F} ^ {2}, \tag {1} +$$ + +where $N = T_{e} \times V$ , $f(E_{N})$ is the prediction of decoder $f(\cdot)$ . Prediction targets in the current works extend beyond raw coordinates to include motion trajectories (i.e., frame differences [20]) or patch features from other networks [1]. + +However, such low-level targets share critical limitations: high computational costs and inadequate high-level semantic capture, creating significant gaps for downstream tasks. This drives our extension of prediction targets to different high-level features through: + +$$ +\mathcal {L} _ {p} = \frac {1}{M} \| f (E _ {N}) - g (X) \| _ {F} ^ {2}, \tag {2} +$$ + +where $M$ is the number of target features, $g$ is a network to extract various features from the input skeleton action as targets of the decoder, such as action-aware representative features. This high-level feature prediction enhances feature transferability for downstream tasks, particularly those demanding global semantic understanding like recognition and retrieval. + +# 3.2. High-level Feature Prediction + +To achieve the above learning objectives, we develop a hierarchical architectural instantiation that strategically balances local motion details with global contextual awareness. + +![](images/daf966f60e946001f06a95ab48c4c027582a00d811479ab56bbb3dc06b40c72c.jpg) +Figure 2. The framework of our proposed GFP. The GFP framework establishes a bidirectional learning paradigm between two core networks: an encoder-decoder architecture and a Target Generation Network (TGN). The encoder extracts features from partially visible skeleton joints, while the decoder progressively predicts hierarchical high-level features — spanning short-term motion patterns to global action semantics. Simultaneously, the TGN generates semantic targets by encoding transformed inputs and performs consistency learning. This mutual guidance is stabilized through variance-covariance regularization, an information-maximization constraint that ensures feature diversity. + +Our approach focuses on selecting learning objectives with richer semantic information across larger spatial-temporal scales, such as the motion patterns of specific nodes over 10-frame sequences. To handle numerous candidate features, we designed a progressive modeling strategy. It starts with short-term local joint patterns, gradually expands to full cycles, and finally integrates local features for holistic action prediction. + +Formally, the decoder takes $E_N \in \mathbb{R}^{(T_e \times V) \times C_e}$ that combines unmasked joint features from the encoder with mask tokens as input. The pipeline first employs temporal average pooling with kernel size $t_1$ , then processes the sequence through a transformer decoder to build the local motion prediction model, producing feature prediction $Z_{p_1} \in \mathbb{R}^{((T_e / t_1) \times V) \times C_p}$ . Following the initial stage, our hierarchical architecture iteratively scales temporal receptive fields through cascaded processing blocks. For the second stage, temporal average pooling with another kernel size is applied to $Z_{p_1}$ , followed by another transformer decoder that models mid-term dynamics, yielding $Z_{p_2} \in \mathbb{R}^{((T_e / t_2) \times V) \times C_p}$ . This pyramidal reduction continues until capturing complete action cycles, where the temporal dimension collapses to a singleton, gaining different time granularity feature predictions $\{Z_{p_1}, Z_{p_2}, \dots, Z_{p_k}\}$ . The compressed representation $Z_{p_k} \in \mathbb{R}^{V \times C_p}$ then passes through feature aggregation, followed by an MLP for action-level semantic pre + +diction $Z_{p_g}$ . This multi-granularity fusion enables simultaneous preservation of fine-grained joint kinematics and high-level semantic context. Upon obtaining the hierarchical features, our encoder-decoder network learns by minimizing discrepancies between predicted features and target features across hierarchy levels. Given the target set $\{Z_{t_1}, Z_{t_2}, \dots, Z_{t_k}, Z_{t_g}\}$ and $\mathcal{J} = 1, \dots, k, g$ , the loss function is formulated as: + +$$ +\mathcal {L} _ {p} = \frac {1}{M} \sum_ {j \in \mathcal {J}} \| Z _ {p _ {j}} - Z _ {t _ {j}} \| _ {F} ^ {2}. \tag {3} +$$ + +where $M$ is the number of all targets. All predicted features are dimensionally aligned to target features through an MLP projector, retaining the original notation for brevity. + +It is worth noting that previous work predicting patch features from other networks can be viewed as a special case of our method when $t_1 = 1$ with only a single hierarchy level. However, such exclusive reliance on the finest-grained (yet most redundant) local information inevitably leads to significant computational overhead while struggling to capture global contextual patterns. + +# 3.3. Target Generation + +Semantic Extraction: To obtain the aforementioned high-level semantic targets, we introduce a jointly optimized + +target generation network (TGN) composed of multiple lightweight MLPs, where each MLP specializes in producing hierarchical semantic features at specific granularity. Specifically, for each hierarchy level, we adopt simple spatiotemporal aggregation operations that process input signals within corresponding receptive fields: local-level semantic extractors operate on joints that merge across consecutive frames, while the global-level semantic extractor performs holistic feature extraction by integrating all spatiotemporal context. As an example, for the target feature $Z_{t_1}$ , the corresponding input is represented as $X_{t_1} \in \mathbb{R}^{((T_e / t_1)\times V)\times (t_1\times l\times C)}$ . + +As for the input of TGN, feeding identical encoder input $\mathbf{X}_e$ to both networks amplifies bias towards frequently unmasked joints. We devise a simple transformation strategy to get $X_{e} = X_{e}[1:] - X_{e}[: - 1]$ . This geometrically grounded transformation preserves high-level motion semantics while avoiding bias. ur experiments demonstrate the remarkable effectiveness of this parameter-efficient design. + +Information Maximization: Under the current framework where the decoder learns from TGN-generated targets while the TGN simultaneously adapts to the decoder's evolving feature representations, this co-learning scheme enables consistent high-level semantic alignment between both networks. However, such bidirectional dependency inevitably risks collapsing into trivial solutions where both networks output constant feature vectors independent of input signals, thereby losing all discriminative power. To prevent this degenerative collapse, we introduce a self-constrained feature space regularization mechanism inspired by VICReg [2], a joint embedding architecture that enforces information-maximization constraints. + +Variance Regularization. Taking the global target feature $Z_{t_g} \in \mathbb{R}^{B \times C_t}$ as an example, variance regularization enforces each feature dimension to maintain batch-wise variance above a predefined threshold $\gamma$ , ensuring persistent diversity in the learned representations and preventing pathological collapse where disparate inputs are mapped to identical feature vectors. Formally, this is implemented through a variance regularization term: + +$$ +\mathcal {L} _ {v a r} = \frac {1}{C _ {t}} \sum_ {i = 1} ^ {C _ {t}} \max \left(0, \gamma - \sqrt {\operatorname {V a r} \left(Z _ {t _ {g}} [ : , i ]\right)}\right) \tag {4} +$$ + +where $\operatorname{Var}(\cdot)$ computes empirical variance over the batch dimension $B$ . Following [2], $\gamma$ is set 1 in our experiments. + +Covariance Regularization. The covariance constraint drives the batch-wise covariance between every feature dimension pair toward zero, eliminating inter-dimensional correlations to prevent redundant information encoding across channels. This is formalized through a covariance + +minimization term: + +$$ +\mathcal {L} _ {c o v} = \frac {1}{C _ {t}} \sum_ {i \neq j} \left[ \operatorname {C o v} \left(Z _ {t _ {g}}\right) \right] _ {i, j} ^ {2} \tag {5} +$$ + +where $\text{Cov}(Z_{t_g})$ denotes covariance matrix of $Z_{t_g}$ . + +Optimization Objectives: The final information-maximization constraints are applied to all target features across hierarchy levels through unified regularization: + +$$ +\mathcal {L} _ {r e g} = \sum_ {j \in \mathcal {J}} \left(\alpha \mathcal {L} _ {c o v} \left(Z _ {t _ {j}}\right) + \beta \mathcal {L} _ {v a r} \left(Z _ {t _ {j}}\right)\right), \tag {6} +$$ + +where $\alpha$ and $\beta$ are trade-off coefficients. The collaborative optimization objective for both networks combines the prediction loss with information-maximization constraints through weighted aggregation: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda \mathcal {L} _ {\text {p r e d}} + \mathcal {L} _ {\text {r e g}} \tag {7} +$$ + +where $\lambda$ denotes trade-off parameters, and $L_{pred}$ denotes the prediction loss computed over the entire batch, i.e., $L_{pred} = \frac{1}{B}\sum \mathcal{L}_p$ . This unified objective drives the coevolution of the decoder and TGN towards both precise hierarchical alignment and collapse-resistant feature extraction. + +# 4. Experiments + +# 4.1. Datasets + +This paper conducts experiments on three widely-used skeleton action benchmarks: NTU RGB+D 60 (NTU-60), NTU RGB+D 120 (NTU-120), and PKU-MMD II (PKU-II). Technical specifications and evaluation protocols for these datasets are summarized as follows. + +NTU RGB+D 60 [24] features 60 action classes with 56,880 sequences captured from 40 participants using multi-view cameras. Two standard evaluation settings are adopted: 1) Cross-Subject (x-sub) divides data by participant IDs, using 20 subjects for training and 20 for testing; 2) Cross-View (x-view) employs camera views 2-3 for training and view 1 for testing. + +NTU RGB+D 120 [17] extends NTU-60 to 120 action classes with 114,480 sequences from 106 subjects across 32 capture setups. Its evaluation protocols include: 1) Cross-Subject (x-sub) split with 53 subjects each for training/testing; 2) Cross-Setup (x-setup) partitioning using even-numbered setups for training and odd-numbered for testing. + +PKU-MMD II [18] focuses on viewpoint-robust action analysis with 51 categories containing 6,952 sequences. The dataset's complexity stems from significant viewpoint diversity. Following established convention, we report results under the standard cross-subject split (5,339 training/1,613 test samples). + +Table 1. Skeleton-based action recognition. Comparisons to the state-of-the-art MAE-like methods for skeleton-based action recognition downstream task on NTU-60. All pre-training runs were completed under identical hardware conditions using a single RTX 4090 GPU. Enc, Dec, and TGN denote the encoder, decoder, and target generation network, respectively. + +
MethodTargetEncDecTGNHoursSpeedupNTU-60
x-subx-view
SkeletonMAE [34]Joint1.97G17.70G-20h27m1×74.877.7
MAMP [20]Motion1.97G17.70G-20h27m1×84.989.1
S-JEPA [1]Patch-level1.97G17.70G28.32G90h57m0.2×85.389.8
GFP (Ours)Hierarchical1.97G1.57G0.64G3h14m6.2×85.992.0
+ +Table 2. Skeleton-based action recognition. Comparisons to the state-of-the-art methods for skeleton-based action recognition downstream task on NTU-120 and PKU-MMD II. Bold and underline formatting indicate the best and second-best methods in each group, respectively. + +
MethodNTU-120PKU-II
x-subx-setupx-sub
other pretext task:
P&C [29]42.741.725.5
ISC [31]67.167.936.0
HaLP[23]71.172.243.5
HiCo[6]72.874.149.4
CMD [19]70.371.543.0
UmURL[30]73.574.352.1
USDRL [33]76.678.154.4
masked modeling:
SkeletonMAE[39]72.573.536.1
MAMP [20]78.679.153.8
S-JEPA [1]79.679.953.5
GFP (Ours)79.180.356.2
+ +Consistent with recent literature [19, 31], all experiments on three datasets employ top-1 accuracy as the evaluation metric. + +# 4.2. Implementation Details + +Model Architecture. Follow [1, 20], our model utilizes a standard vision transformer backbone with learnable spatiotemporal position encoding. The encoder stacks $L_{e} = 8$ identical layers, each containing multi-head attention (8 heads, 256-dim) and a feed-forward network (1024 hidden units). For pre-training, the decoder retains the same dimensions but simplifies the architecture: each local feature decoder uses one layer, while the global feature decoder adopts a 2-layer MLP with 2048-dimensional hidden units. The target generation network employs a 3-layer MLP (512 hidden units) for local semantic extraction and a 3-layer MLP (2048 hidden units) for global semantic extraction. All projectors in the decoder are 3-layer MLPs that project features to match the dimensions of the target features. + +Table 3. Comparisons to the state-of-the-art methods for skeleton-based action retrieval. **Bold** indicates the best methods in each group. + +
MethodNTU-60
x-subx-view
other pretext task:
LongT GAN[44]39.148.1
P&C[29]50.776.3
ISC[31]62.582.6
HaLP[23]65.883.6
HiCo[6]68.384.8
SkeAttnCLR[10]69.476.8
UmURL[30]71.388.3
masked modeling:
MAMP [20]62.070.0
GFP (Ours)70.987.1
+ +Data Augmentation. For skeleton sequence processing, as in previous works [1, 20], we adopt temporal cropping with dynamic scaling by randomly selecting a continuous clip spanning $50\% - 100\%$ of the original sequence length. All clips are resampled to frames via bilinear interpolation to ensure temporal consistency. + +Pretraining Setup. We employ the motion-aware masking strategy proposed in [20] and segment length $l$ to 4 in the pertaining. Optimization uses AdamW $(\beta_{1} = 0.9, \beta_{2} = 0.95$ , weight decay 0.05) under a 400-epoch schedule. Learning rates ramp up linearly from 0 to 1e-3 during 20 warmup epochs, then decay to 5e-4 via cosine annealing. Hyperparameters $\lambda, \alpha$ and $\beta$ are set to 5, 5, 1, respectively. For the target features, we employ a four-level hierarchical feature corresponding to local features at $t_{1} = 5$ , $t_{2} = 10$ , $t_{3} = 30$ and global feature. + +# 4.3. Comparison to the State-of-the-art + +In this section, we compare our method with state-of-the-art approaches on two skeleton-based tasks: action recognition and action retrieval. + +![](images/3e59af5232d4892bf3a8f85cf8ec8c98dcc05b8ba18f64ba72296b151d06d289.jpg) +Figure 3. Impact of Different Targets. The $5,10,30,\mathrm{g}$ configurations represent local features from 5-, 10-, 30-frame temporal spans and global representations respectively, all derived from high-level representations generated by the TGN. All results are obtained under the NTU-60 x-view protocol. + +Skeleton-based Action Recognition. Following the standard practice from previous works [1, 20, 31], we train a linear classifier on top of the frozen encoder pre-trained with our proposed method. First, we compare our method against prior MAE-like approaches [1, 20, 34], on the NTU-60 dataset. Beyond recognition performance, we quantify computational efficiency by measuring computational FLOPs per action sequence for encoder, decoder, and target generation network, full pre-training time under the x-sub protocol (identical hardware), and training speedup relative to SkeletonMAE [34]. As shown in Table 1, our method achieves significant performance gains while substantially accelerating training. This improvement stems from reducing the optimization targets from 750 low-level targets $(\mathbf{N} = 750)$ in prior works to 251 hierarchical high-level semantic targets $(\mathbf{M} = 251)$ . Notably, while S-JPEA also adopts learned features as objectives, its patch-level implementation necessitates redundant feature encoding. This design choice, coupled with extremely slow convergence (requiring 1200 epochs versus 400 epochs for other methods), severely degrades training speed compared to SkeletonMAE. These results demonstrate our method's balanced efficacy in feature quality and computational efficiency. + +We further validate our approach on larger-scale NTU-120 and more challenging PKU-MMD II datasets, comparing against methods using alternative pretext tasks. As shown in Table 2, our method outperforms most existing approaches, confirming its broad effectiveness. + +Skeleton-based Action Retrieval. In this experiment, the encoder operates in a direct feature extraction mode without requiring any training. Similar to video-to-video retrieval, we use cosine similarity in the feature space to + +![](images/f81d9450731cd77fbb267283a53b34246a8794d95d80832db96bfa36c45072ac.jpg) +Figure 4. Performance of various methods with different targets. All results are obtained under the NTU-60 x-sub protocol. + +retrieve the nearest action samples for a given query. As detailed in Table 3, conventional MAE-like methods exhibit poor performance on this task due to their focus on local reconstruction objectives without global semantic learning. In contrast, our method achieves substantial improvements (8.9% on x-sub and 17.1% on x-view protocols) over prior approaches like MAMP by incorporating high-level semantic objectives. Notably, our results remain competitive with contrastive learning methods [6, 30] that explicitly optimize global representations, demonstrating the superiority of high-level semantic targets over low-level reconstruction objectives. + +# 4.4. Ablation Study + +All ablation studies were conducted on the NTU-60 dataset for skeleton-based action recognition. + +Prediction Targets: Our method employs local joint features (5-, 10-, and 30-frame spans) and global semantic features as learning targets. To investigate their individual impacts, we progressively ablated these components. As shown in Figure 3, sequential removal of global and local semantic learning objectives caused performance degradation, demonstrating the complementary nature of hierarchical feature learning. Furthermore, we explored using low-level targets instead of learned representations by removing the TGN from our framework, which forces direct prediction vectors of all spatiotemporal information within targets (e.g., flattening input into a single vector for global target). The corresponding results (labeled $\{5,10,30,g\}^*$ in Figure 3) reveal significant performance drops, proving that high-level semantics cannot be effectively captured through direct prediction of low-level signals. + +Input of Target Generation Network: For the target network inputs, we evaluated three variants: (1) raw joint coordinates, (2) masked joint data, and (3) motion features derived via linear transformation of joints. As shown in Table 4, the motion features achieved optimal performance, validating our design choice. Notably, masked + +Table 4. Ablation study on the input of target generation network. + +
InputNTU 60
x-subx-view
joint85.090.9
masked joint84.290.3
motion85.992.0
+ +Table 5. Ablation study on the architecture of target generation network. + +
Arch.NTU 60
x-subx-view
Base85.992.0
Shadow85.891.8
Wider85.891.3
+ +Table 6. Ablation study on different number of layers used in projector. + +
LayersNTU 60
x-subx-view
285.591.3
385.992.0
485.691.6
+ +Table 7. Comparisons to the state-of-the-art methods with semi-supervised learning on NTU-60 dataset. + +
Methodx-subx-view
1%10%1%10%
ASSL [27]-64.3-69.8
ISC [31]35.765.938.172.5
Colorization [38]48.371.752.578.9
Hi-TRS [3]-70.7-74.8
GL-Transformer [12])-68.6-74.9
HaLP [23]46.672.648.777.1
CMD [19]50.675.453.080.2
CPM [40]56.773.057.577.1
HYSP [9]-76.2-80.4
\( PCM^3 \) [41])53.877.753.182.8
HiCo [6]54.473.054.878.3
UmURL [30]58.1-58.3-
USDRL [33]57.380.260.784.0
SkeletonMAE [34]54.480.654.683.5
MAMP [20]66.088.068.791.5
S-JEPA [1]67.588.469.191.4
GFP (Ours)71.888.772.992.1
+ +inputs hinder semantic feature extraction by TGN, aligning with observations in prior work [1]. We evaluate the epoch-wise performance of different methods with varying learning objectives: SkeletonMAE learns to reconstruct raw joints, MAMP targets motion features (i.e., temporal difference of joint coordinates), and our method predicts high-level semantic representations. As shown in Figure 4, our approach consistently outperforms prior works across all training epochs, demonstrating the superiority of hierarchical semantic learning over low-level reconstruction objectives. + +Architecture Design: We conducted architectural experiments on the target generation network and projector. For TGN: (1)Shallow variant: 2-layer MLP; (2)Wider variant: 4-layer MLP with doubled hidden dimensions. (3)Base: Standard 3-layer MLP. As shown in Table 5, + +the baseline 3-layer configuration achieves optimal performance, though all variants show marginal differences, demonstrating TGN's robustness to architectural variations. For the projector, we evaluated MLPs with 2-4 layers. The 3-layer configuration yields superior performance (Table 6). + +# 4.5. Extended Performance Validation + +To assess representation transferability, we conduct limited-label experiments on NTU-60 under two data settings: $1\%$ and $5\%$ labeled training data. Following semi-supervised learning benchmarks [31], we initialize the model with weights pre-trained through our model, then attach and fine-tune a classifier using available labels. + +As shown in Table 7, our approach establishes new state-of-the-art performance across both data scales. This significant performance gap confirms the effectiveness of our learned representations in label-efficient scenarios. + +# 5. Conclusion + +In this paper, we propose an efficient general feature prediction framework for masked skeleton modeling. Unlike conventional approaches that reconstruct low-level targets, our method employs hierarchical high-level semantic features as learning targets, significantly reducing decoder complexity. The framework integrates a target generation network that provides online learning objectives. We further introduce the information maximization constraints, enabling non-trivial co-optimization between networks. This mechanism simultaneously preserves local motion details and global semantics. Extensive experiments across three datasets demonstrate the effectiveness of our method, achieving state-of-the-art performance with significant training speedup over prior masked skeleton modeling approaches. + +# Acknowledgments + +This work was supported by the Pioneer and Leading Goose R&D Program of Zhejiang (No. 2024C01110), National Natural Science Foundation of China (No. 72188101, 62020106007, and 62472385). The computation is completed on the HPC Platform of Hefei University of Technology. + +# References + +[1] Mohamed Abdelfattah and Alexandre Alahi. S-jepa: A joint embedding predictive architecture for skeletal action recognition. In European Conference on Computer Vision, pages 367-384. Springer, 2024. 3, 6, 7, 8 +[2] Adrien Bardes, Jean Ponce, and Yann Lecun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR 2022-International Conference on Learning Representations, 2022. 2, 5 +[3] Yuxiao Chen, Long Zhao, Jianbo Yuan, Yu Tian, Zhaoyang Xia, Shijie Geng, Ligong Han, and Dimitris N Metaxas. Hierarchically self-supervised transformer for human skeleton representation learning. In European Conference on Computer Vision, pages 185-202, 2022. 8 +[4] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183-192, 2020. 1 +[5] Yi-Bin Cheng, Xipeng Chen, Junhong Chen, Pengxu Wei, Dongyu Zhang, and Liang Lin. Hierarchical transformer: Unsupervised representation learning for skeleton-based human action recognition. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2021. 3 +[6] Jianfeng Dong, Shengkai Sun, Zhonglin Liu, Shujie Chen, Baolong Liu, and Xun Wang. Hierarchical contrast for unsupervised skeleton-based action representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 525-533, 2023. 1, 6, 7, 8 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1 +[9] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. arXiv preprint arXiv:2303.06242, 2023.8 +[10] Yilei Hua, Wenhan Wu, Ce Zheng, Aidong Lu, Mengyuan Liu, Chen Chen, and Shiqian Wu. Part aware contrastive learning for self-supervised action recognition. In International Joint Conference on Artificial Intelligence, 2023. 3, 6 +[11] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence, 35(1):221-231, 2012. 1 +[12] Boeun Kim, Hyung Jin Chang, Jungho Kim, and Jin Young Choi. Global-local motion transformer for unsupervised skeleton-based action learning. In European Conference on Computer Vision, pages 209-225, 2022. 3, 8 + +[13] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. 1 +[14] Jogendra Nath Kundu, Maharshi Gor, Phani Krishna Uppala, and Venkatesh Babu Radhakrishnan. Unsupervised feature learning of human actions as trajectories in pose embedding manifold. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1459-1467, 2019. 2 +[15] Linguuo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4741-4750, 2021. 3 +[16] Lilang Lin, Jiahang Zhang, and Jiaying Liu. Actionlet-dependent contrastive learning for unsupervised skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2363-2372, 2023. 3 +[17] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684-2701, 2019. 1, 5 +[18] Jiaying Liu, Sijie Song, Chunhui Liu, Yanghao Li, and Yueyu Hu. A benchmark dataset and comparison study for multi-modal human action analytics. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 16(2):1-24, 2020. 5 +[19] Yunyao Mao, Wengang Zhou, Zhenbo Lu, Jiajun Deng, and Houqiang Li. Cmd: Self-supervised 3d action representation learning with cross-modal mutual distillation. In European Conference on Computer Vision, pages 734-752. Springer, 2022. 3, 6, 8 +[20] Yunyao Mao, Jiajun Deng, Wengang Zhou, Yao Fang, Wanli Ouyang, and Houqiang Li. Masked motion predictors are strong 3d action representation learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10191, 2023. 3, 6, 7, 8 +[21] Qiang Nie, Ziwei Liu, and Yunhui Liu. Unsupervised 3d human pose representation with viewpoint and pose disentanglement. In European Conference on Computer Vision, pages 102-118, 2020. 2 +[22] Haocong Rao, Shihao Xu, Xiping Hu, Jun Cheng, and Bin Hu. Augmented skeleton based contrastive action learning with momentum lstm for unsupervised action recognition. Information Sciences, 569:90-109, 2021. 3 +[23] Anshul Shah, Aniket Roy, Ketul Shah, Shlok Mishra, David Jacobs, Anoop Cherian, and Rama Chellappa. Halp: Hallucinating latent positives for skeleton-based self-supervised learning of actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18846-18856, 2023. 3, 6, 8 +[24] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE Conference on Computer + +Vision and Pattern Recognition, pages 1010-1019, 2016. 1, 5 +[25] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7912-7921, 2019. 1 +[26] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12026-12035, 2019. 1 +[27] Chenyang Si, Xuecheng Nie, Wei Wang, Liang Wang, Ti-niu Tan, and Jiashi Feng. Adversarial self-supervised learning for semi-supervised 3d action recognition. In European Conference on Computer Vision, pages 35-51, 2020. 8 +[28] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1 +[29] Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9631-9640, 2020. 2, 6 +[30] Shengkai Sun, Daizong Liu, Jianfeng Dong, Xiaoye Qu, Junyu Gao, Xun Yang, Xun Wang, and Meng Wang. Unified multi-modal unsupervised representation learning for skeleton-based action understanding. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2973-2984, 2023. 1, 3, 6, 7, 8 +[31] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1655–1663, 2021. 3, 6, 7, 8 +[32] Fei Wang, Dan Guo, Kun Li, and Meng Wang. Eulermormer: Robust eulerian motion magnification via dynamic filtering within transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5345-5353, 2024. 2 +[33] Wanjiang Weng, Hongsong Wang, Junbo He, Lei He, and Guosen Xie. Usdrl: Unified skeleton-based dense representation learning with multi-grained feature decorrelation. arXiv preprint arXiv:2412.09220, 2024. 6, 8 +[34] Wenhan Wu, Yilei Hua, Ce Zheng, Shiqian Wu, Chen Chen, and Aidong Lu. Skeletonmae: Spatial-temporal masked autoencoders for self-supervised skeleton action recognition. In 2023 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 224-229. IEEE, 2023. 3, 6, 7, 8 +[35] Yuankun Xu, Dong Huang, Chang-Dong Wang, and Jian-Huang Lai. Deep image clustering with contrastive learning and multi-scale graph convolutional networks. Pattern Recognition, 2024. 1 +[36] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI Conference on Artificial Intelligence, 2018. 1 +[37] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and François Brémond. View-invariant skeleton action representation learning via motion + +retargeting. International Journal of Computer Vision, pages 1-16, 2024. 3 +[38] Siyuan Yang, Jun Liu, Shijian Lu, Meng Hwa Er, and Alex C Kot. Skeleton cloud colorization for unsupervised 3d action representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13423-13433, 2021. 2, 8 +[39] Siyuan Yang, Jun Liu, Shijian Lu, Er Meng Hwa, Yongjian Hu, and Alex C Kot. Self-supervised 3d action representation learning with skeleton cloud colorization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 6 +[40] Haoyuan Zhang, Yonghong Hou, Wenjing Zhang, and Wanqing Li. Contrastive positive mining for unsupervised 3d action representation learning. In European Conference on Computer Vision, pages 36-51. Springer, 2022. 3, 8 +[41] Jiahang Zhang, Lilang Lin, and Jiaying Liu. Prompted contrast with masked motion modeling: Towards versatile 3d action representation learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7175-7183, 2023. 3, 8 +[42] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber: Tubelet transformer for video action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13598-13607, 2022. 2 +[43] Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In Proceedings of the IEEE international conference on computer vision, pages 2914-2923, 2017. 1 +[44] Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2644-2651, 2018. 1, 2, 6 \ No newline at end of file diff --git a/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/images.zip b/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..904c2ff4621bd9eb1a3db61f45cf1ed989e78158 --- /dev/null +++ b/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c57179049d7585eaf5be9e09a3bd1fc0ed3604b69371eb495a3ccfc747983856 +size 369776 diff --git a/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/layout.json b/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e618975d867f66507ced8ed651e76b108fe360a9 --- /dev/null +++ b/2025/Towards Efficient General Feature Prediction in Masked Skeleton Modeling/layout.json @@ -0,0 +1,7487 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 71, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 71, + 103, + 539, + 121 + ], + "type": "text", + "content": "Towards Efficient General Feature Prediction in Masked Skeleton Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "spans": [ + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "text", + "content": "Shengkai Sun" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "text", + "content": ", Zefan Zhang" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "text", + "content": ", Jianfeng Dong" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "text", + "content": ", Zhiyong Cheng" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "text", + "content": ", Xiaojun Chang" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "text", + "content": ", Meng Wang" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "text", + "content": "Hefei University of Technology " + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 71, + 142, + 536, + 171 + ], + "type": "text", + "content": "Jilin University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 171, + 506, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 171, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 104, + 171, + 506, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 104, + 171, + 506, + 186 + ], + "type": "text", + "content": "Zhejiang Gongshang University " + }, + { + "bbox": [ + 104, + 171, + 506, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 104, + 171, + 506, + 186 + ], + "type": "text", + "content": "University of Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 239, + 297, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 297, + 515 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 297, + 515 + ], + "type": "text", + "content": "Recent advances in the masked autoencoder (MAE) paradigm have significantly propelled self-supervised skeleton-based action recognition. However, most existing approaches limit reconstruction targets to raw joint coordinates or their simple variants, resulting in computational redundancy and limited semantic representation. To address this, we propose a novel General Feature Prediction framework (GFP) for efficient mask skeleton modeling. Our key innovation is replacing conventional low-level reconstruction with high-level feature prediction that spans from local motion patterns to global semantic representations. Specifically, we introduce a collaborative learning framework where a lightweight target generation network dynamically produces diversified supervision signals across spatial-temporal hierarchies, avoiding reliance on pre-computed offline features. The framework incorporates constrained optimization to ensure feature diversity while preventing model collapse. Experiments on NTU RGB+D 60, NTU RGB+D 120 and PKU-MMD demonstrate the benefits of our approach: Computational efficiency (with " + }, + { + "bbox": [ + 53, + 239, + 297, + 515 + ], + "type": "inline_equation", + "content": "6.2\\times" + }, + { + "bbox": [ + 53, + 239, + 297, + 515 + ], + "type": "text", + "content": " faster training than standard masked skeleton modeling methods) and superior representation quality, achieving state-of-the-art performance in various downstream tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 539, + 135, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 135, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 135, + 552 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 559, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 693 + ], + "type": "text", + "content": "Understanding human action [8, 11, 43] has long been a fundamental research task in computer vision, demonstrating significant application value in human-computer interaction, intelligent surveillance, and autonomous driving, etc. In contrast to conventional RGB video representations [13, 28] that densely encode visual information, skeletal representations [17, 24] capture body postures as topological joint sets, employing sparse coordinate sequences to characterize human actions instead of raw pixel values. Owing to the compact data structure and explicit motion representation, skeleton-based action understanding" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 341, + 215, + 529, + 383 + ], + "blocks": [ + { + "bbox": [ + 341, + 215, + 529, + 383 + ], + "lines": [ + { + "bbox": [ + 341, + 215, + 529, + 383 + ], + "spans": [ + { + "bbox": [ + 341, + 215, + 529, + 383 + ], + "type": "image", + "image_path": "8a708eb943763f3e8f5b1c42f46bc667609a4bafe9a240a77c705891145d6953.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 394, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 472 + ], + "type": "text", + "content": "Figure 1. Frameworks for general feature prediction (GFP). Instead of low-level targets, GFP predicts the hierarchical abstract representations of input skeleton sequence, embedded by a target generation network, effectively learning more informative high-level features for the downstream tasks. This paradigm achieves an optimal balance between model performance and computational efficiency." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 496, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 652 + ], + "type": "text", + "content": "[4, 25, 26, 36] methods not only significantly reduce computational costs, but also effectively mitigate noise from dynamic backgrounds and lighting changes. However, despite these advantages, early approaches based on fully supervised learning faced notable limitations due to their heavy reliance on manually labeled skeleton data, which requires labor-intensive and costly annotation processes. To overcome this, the research community has increasingly focused on self-supervised learning [6, 30, 44], designing pretext tasks that can automatically learn spatiotemporal patterns in skeleton data without manual labels. The resulting representations can then be efficiently adapted to downstream tasks with a few labeled examples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Self-supervised skeleton-based action learning evolved from early handcrafted methods (e.g., colorization, temporal segment rearrangement) to instance discrimination. Contrastive learning [35], a prominent instance discrimination approach, gained dominance due to its simplicity and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 236, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 236, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 236, + 713 + ], + "type": "text", + "content": "*Corresponding author: jason.zy.cheng@gmail.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12212" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "efficiency. However, current research reveals some limitations in contrastive learning: excessive dependence on data augmentation strategies and persistent false negative sample issues. Recently, transformers have achieved remarkable success in visual tasks [7, 32, 42], and transformer-based masked image modeling has emerged as a novel generative pretraining paradigm, offering fresh insights for self-supervised skeleton action learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 172, + 298, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 298, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 298, + 460 + ], + "type": "text", + "content": "The masked skeleton modeling methods follow the classic MAE framework by randomly masking some joints in the input skeleton sequence. Its encoder processes only the visible joints, while the decoder reconstructs missing joint coordinates from latent representations and learnable mask tokens. However, due to the intrinsic differences between skeleton and image data, skeleton MAE-like methods significantly differ in implementation. These methods employ a " + }, + { + "bbox": [ + 56, + 172, + 298, + 460 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 56, + 172, + 298, + 460 + ], + "type": "text", + "content": " masking ratio to address temporal redundancy and adopt deeper decoders (instead of MAE's lightweight design) to handle complex spatiotemporal dependencies. Despite their impressive performance, the current design and learning objectives have clear shortcomings: (1) Heavy decoder computation: Due to the reconstruction target involving many joints and the transformer being used as the decoder, excessively long decoding sequences significantly slow down training; (2) Lacking semantic guidance: The low-level target reconstruction objective lacks explicit supervision from higher-level spatiotemporal features like motion semantics, which creates a semantic gap between pretraining and downstream tasks. These limitations motivate our exploration of new masked skeleton modeling paradigms that achieve a balance between representation quality and computational efficiency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 463, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we propose a novel General Feature Prediction framework (GFP) for efficient mask skeleton modeling. As shown in Figure 1, unlike methods targeting numerous redundant low-level features, GFP focuses on a smaller set of hierarchical high-level features, learning through progressive semantic prediction. Specifically, during decoding, it inputs elements of varying spatiotemporal granularity to predict from local motion patterns (e.g., a joint's movement over 10 frames) to global action semantics. This approach offers two advantages: First, the significantly reduced number of targets cuts computational costs in decoding and optimization. Second, it balances local details and global context, capturing both micromotion changes and overall action structures, thereby substantially improving feature quality. To obtain these high-level semantic features as learning targets, we designed a lightweight target generation network. This network processes elements with varying spatiotemporal granularities, using a multiple feature extraction module to capture both local details and global context, providing hierarchical supervision for the decoder. This supervisory mechanism" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 274 + ], + "type": "text", + "content": "operates bidirectionally: The decoder infers high-level semantic features from unmasked joint features for learning, while the target generation network uses the decoder's output as an optimization target, guiding feature representation through a consistency learning manner. However, this design alone is insufficient, as joint optimization can lead to trivial solutions where both networks output constant values regardless of the input. Inspired by the recent information maximization representation learning method [2], we introduced a constraint mechanism to ensure meaningful feature outputs, enabling iterative optimization that converges to a semantically consistent stable state. Ultimately, we developed a flexible mask skeleton modeling framework that significantly accelerates training, enabling the model to learn more discriminative and robust feature representations, yielding superior performance across various downstream tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 325, + 276, + 511, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 276, + 511, + 287 + ], + "spans": [ + { + "bbox": [ + 325, + 276, + 511, + 287 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 289, + 553, + 468 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 289, + 553, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 289, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 314, + 289, + 553, + 336 + ], + "type": "text", + "content": "- We extend conventional mask skeleton modeling from low-level reconstruction to general feature prediction, improving decoding efficiency while introducing local and global semantic guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 337, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 337, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 315, + 337, + 553, + 408 + ], + "type": "text", + "content": "- We design a lightweight target generation network that provides online supervision for the decoder. By introducing an information maximization constraint, we enable collaborative learning between networks, achieving a non-trivial self-supervised skeleton action representation learning scheme." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 409, + 553, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 409, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 315, + 409, + 553, + 468 + ], + "type": "text", + "content": "- Extensive experiments across three datasets demonstrate the effectiveness of our approach. The results show that our framework is highly efficient, learns richer semantic information, and outperforms previous methods on multiple downstream tasks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 482, + 400, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 482, + 400, + 495 + ], + "spans": [ + { + "bbox": [ + 314, + 482, + 400, + 495 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 502, + 553, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 553, + 528 + ], + "type": "text", + "content": "2.1. Self-supervised Skeleton-based Action Representation Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 553, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 553, + 592 + ], + "type": "text", + "content": "Self-supervised skeleton-based action representation learning aims to learn discriminative features without manual annotation. The core challenge lies in designing meaningful supervisory signals to guide the model in capturing action semantics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Early approaches to self-supervised skeleton action representation learning predominantly relied on handcrafted pretext tasks within encoder-decoder frameworks. Reconstruction-based methods formed the cornerstone: [44] reconstructed corrupted skeletons, while [14] and [29] employed autoencoders for full-sequence reconstruction. [21] pioneered feature disentanglement into pose/view components before reconstruction. Subsequent innovations leveraged spatio-temporal relationships—[38] predicted joint coloring which is pre-defined by spatial-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12213" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "temporal order, [5] forecasted inter-frame motions, and [12] estimated multi-scale pose displacements. [37] advanced view invariance through motion retargeting-driven reconstruction." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 120, + 295, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 295, + 262 + ], + "type": "text", + "content": "The paradigm shifted with the advent of contrastive representation learning. [22] established the baseline by contrasting augmented skeleton pairs via dual encoders. [31] enhanced this with skeleton-specific augmentations and cross-architectural contrast. Later works diversified positive pair generation: [15, 40] mined contextual positives, [23] synthesized latent-space positives. Multi-modal fusion emerged through [19] and [30]. [10, 16] introduced guidance mechanisms focusing on action-critical regions to enhance semantic feature learning. [41] pioneered dual-prompted pretraining combining masked reconstruction and feature contrast." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 263, + 295, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 295, + 323 + ], + "type": "text", + "content": "Building on the success of masked autoencoders, recent works [1, 20, 34] have adapted this paradigm to skeleton-based action representation learning, demonstrating impressive results in capturing action semantics through partial observation and reconstruction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 354, + 285, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 285, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 285, + 367 + ], + "type": "text", + "content": "2.2. Masked Modeling for Skeleton-based Action" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 371, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 610 + ], + "type": "text", + "content": "[34] pioneers the adaptation of Masked Autoencoder to transformer-based skeleton-based action representation learning, employing direct reconstruction of masked joint coordinates as its pretext task. [20] extended SkeletonMAE to motion feature prediction, proposing a motion-aware masking strategy that prioritizes high movement intensity regions during input masking. This input masking forces the model to focus on action-critical patches, achieving substantial performance improvements over conventional random masking approaches. Despite these advancements, the reliance on low-level reconstruction targets inherently limits high-level semantic learning, creating a performance gap in tasks requiring holistic action representation. [1] attempted to address this by using model-generated features as learning targets. However, their patch-level feature prediction necessitates a bulky EMA-updated encoder of massive computations for target generation. While achieving accuracy gains, this approach incurs higher computational costs and much slower convergence than conventional methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 659 + ], + "type": "text", + "content": "By contrast, our method employs spatiotemporally high-level semantic prediction paired with a lightweight target generation network. This design achieves much faster training versus [1] and higher downstream accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 669, + 111, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 111, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 111, + 681 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "In this section, we first reformulate masked skeleton modeling by replacing its learning objective from conventional" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "patch-level reconstruction to high-level semantic feature prediction. Next, we develop a hierarchical feature prediction framework through multi-granularity representation learning, jointly predicting local details and global semantics. Finally, we design a dedicated target generation network to provide high-quality supervision for prediction tasks, coupled with an information constraint mechanism ensuring stable model optimization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 175, + 382, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 382, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 382, + 186 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "content": "The pipeline of skeleton masked autoencoders operates as follows: A predefined masking strategy randomly masks joints, training the network to reconstruct masked joints for representation learning. Specifically, given an input skeleton sequence " + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{T \\times V \\times C}" + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "content": " is the number of frames, " + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "content": " is the number of joints, and " + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "content": " is the number of channels, temporal patchify first transforms it into " + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "inline_equation", + "content": "X_{e} \\in \\mathbb{R}^{(T_{e} \\times V) \\times (l \\times C)}" + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "content": ". This operation segments the temporal streams of every joint into non-overlapping " + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "content": "-frame patches based on consideration of computational efficiency, where " + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "inline_equation", + "content": "T_{e} = T / l" + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "content": ". The patches are embedded into feature vectors " + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "inline_equation", + "content": "E \\in \\mathbb{R}^{(T_{e} \\times V) \\times C_{e}}" + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "inline_equation", + "content": "C_{e}" + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "content": " dimension. Most patches are randomly masked, leaving only a small subset visible as input to the transformer encoder for feature extraction. The decoder subsequently predicts masked regions using sequence " + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "inline_equation", + "content": "E_{N} \\in \\mathbb{R}^{(T_{e} \\times V) \\times C_{e}}" + }, + { + "bbox": [ + 313, + 191, + 555, + 395 + ], + "type": "text", + "content": " which combines extracted visible features with learnable mask tokens:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 377, + 399, + 553, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 399, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 377, + 399, + 553, + 422 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\frac {1}{N} \\| f (E _ {N}) - X _ {e} \\| _ {F} ^ {2}, \\tag {1}", + "image_path": "0ae791b26e1d5f2b7ce85c885639a148cd0af9e7174bc1681a9495b7989b22e6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 427, + 553, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 553, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 427, + 553, + 475 + ], + "type": "inline_equation", + "content": "N = T_{e} \\times V" + }, + { + "bbox": [ + 313, + 427, + 553, + 475 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 427, + 553, + 475 + ], + "type": "inline_equation", + "content": "f(E_{N})" + }, + { + "bbox": [ + 313, + 427, + 553, + 475 + ], + "type": "text", + "content": " is the prediction of decoder " + }, + { + "bbox": [ + 313, + 427, + 553, + 475 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 427, + 553, + 475 + ], + "type": "text", + "content": ". Prediction targets in the current works extend beyond raw coordinates to include motion trajectories (i.e., frame differences [20]) or patch features from other networks [1]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 476, + 553, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 553, + 536 + ], + "type": "text", + "content": "However, such low-level targets share critical limitations: high computational costs and inadequate high-level semantic capture, creating significant gaps for downstream tasks. This drives our extension of prediction targets to different high-level features through:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 372, + 541, + 553, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 541, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 372, + 541, + 553, + 564 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\frac {1}{M} \\| f (E _ {N}) - g (X) \\| _ {F} ^ {2}, \\tag {2}", + "image_path": "525643e605e5a9af5530b75d0075229575795504ae348fa18aefbf1e658b247c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 553, + 653 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 570, + 553, + 653 + ], + "type": "text", + "content": " is the number of target features, " + }, + { + "bbox": [ + 313, + 570, + 553, + 653 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 570, + 553, + 653 + ], + "type": "text", + "content": " is a network to extract various features from the input skeleton action as targets of the decoder, such as action-aware representative features. This high-level feature prediction enhances feature transferability for downstream tasks, particularly those demanding global semantic understanding like recognition and retrieval." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 660, + 474, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 474, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 474, + 672 + ], + "type": "text", + "content": "3.2. High-level Feature Prediction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "To achieve the above learning objectives, we develop a hierarchical architectural instantiation that strategically balances local motion details with global contextual awareness." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12214" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 77, + 534, + 305 + ], + "blocks": [ + { + "bbox": [ + 72, + 77, + 534, + 305 + ], + "lines": [ + { + "bbox": [ + 72, + 77, + 534, + 305 + ], + "spans": [ + { + "bbox": [ + 72, + 77, + 534, + 305 + ], + "type": "image", + "image_path": "daf966f60e946001f06a95ab48c4c027582a00d811479ab56bbb3dc06b40c72c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 315, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 54, + 315, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 54, + 315, + 555, + 381 + ], + "type": "text", + "content": "Figure 2. The framework of our proposed GFP. The GFP framework establishes a bidirectional learning paradigm between two core networks: an encoder-decoder architecture and a Target Generation Network (TGN). The encoder extracts features from partially visible skeleton joints, while the decoder progressively predicts hierarchical high-level features — spanning short-term motion patterns to global action semantics. Simultaneously, the TGN generates semantic targets by encoding transformed inputs and performs consistency learning. This mutual guidance is stabilized through variance-covariance regularization, an information-maximization constraint that ensures feature diversity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 402, + 294, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 294, + 498 + ], + "type": "text", + "content": "Our approach focuses on selecting learning objectives with richer semantic information across larger spatial-temporal scales, such as the motion patterns of specific nodes over 10-frame sequences. To handle numerous candidate features, we designed a progressive modeling strategy. It starts with short-term local joint patterns, gradually expands to full cycles, and finally integrates local features for holistic action prediction." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "text", + "content": "Formally, the decoder takes " + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "E_N \\in \\mathbb{R}^{(T_e \\times V) \\times C_e}" + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "text", + "content": " that combines unmasked joint features from the encoder with mask tokens as input. The pipeline first employs temporal average pooling with kernel size " + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "text", + "content": ", then processes the sequence through a transformer decoder to build the local motion prediction model, producing feature prediction " + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "Z_{p_1} \\in \\mathbb{R}^{((T_e / t_1) \\times V) \\times C_p}" + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "text", + "content": ". Following the initial stage, our hierarchical architecture iteratively scales temporal receptive fields through cascaded processing blocks. For the second stage, temporal average pooling with another kernel size is applied to " + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "Z_{p_1}" + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "text", + "content": ", followed by another transformer decoder that models mid-term dynamics, yielding " + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "Z_{p_2} \\in \\mathbb{R}^{((T_e / t_2) \\times V) \\times C_p}" + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "text", + "content": ". This pyramidal reduction continues until capturing complete action cycles, where the temporal dimension collapses to a singleton, gaining different time granularity feature predictions " + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\{Z_{p_1}, Z_{p_2}, \\dots, Z_{p_k}\\}" + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "text", + "content": ". The compressed representation " + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "inline_equation", + "content": "Z_{p_k} \\in \\mathbb{R}^{V \\times C_p}" + }, + { + "bbox": [ + 57, + 498, + 295, + 713 + ], + "type": "text", + "content": " then passes through feature aggregation, followed by an MLP for action-level semantic pre" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 402, + 553, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 553, + 497 + ], + "type": "text", + "content": "diction " + }, + { + "bbox": [ + 313, + 402, + 553, + 497 + ], + "type": "inline_equation", + "content": "Z_{p_g}" + }, + { + "bbox": [ + 313, + 402, + 553, + 497 + ], + "type": "text", + "content": ". This multi-granularity fusion enables simultaneous preservation of fine-grained joint kinematics and high-level semantic context. Upon obtaining the hierarchical features, our encoder-decoder network learns by minimizing discrepancies between predicted features and target features across hierarchy levels. Given the target set " + }, + { + "bbox": [ + 313, + 402, + 553, + 497 + ], + "type": "inline_equation", + "content": "\\{Z_{t_1}, Z_{t_2}, \\dots, Z_{t_k}, Z_{t_g}\\}" + }, + { + "bbox": [ + 313, + 402, + 553, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 402, + 553, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{J} = 1, \\dots, k, g" + }, + { + "bbox": [ + 313, + 402, + 553, + 497 + ], + "type": "text", + "content": ", the loss function is formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 373, + 505, + 555, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 505, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 373, + 505, + 555, + 534 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p} = \\frac {1}{M} \\sum_ {j \\in \\mathcal {J}} \\| Z _ {p _ {j}} - Z _ {t _ {j}} \\| _ {F} ^ {2}. \\tag {3}", + "image_path": "8992f8668fbdbd46bdd7f6931f33efe0a5188d9c1e9c4277982262b5c21776a3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 543, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 555, + 579 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 543, + 555, + 579 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 543, + 555, + 579 + ], + "type": "text", + "content": " is the number of all targets. All predicted features are dimensionally aligned to target features through an MLP projector, retaining the original notation for brevity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 580, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 553, + 664 + ], + "type": "text", + "content": "It is worth noting that previous work predicting patch features from other networks can be viewed as a special case of our method when " + }, + { + "bbox": [ + 313, + 580, + 553, + 664 + ], + "type": "inline_equation", + "content": "t_1 = 1" + }, + { + "bbox": [ + 313, + 580, + 553, + 664 + ], + "type": "text", + "content": " with only a single hierarchy level. However, such exclusive reliance on the finest-grained (yet most redundant) local information inevitably leads to significant computational overhead while struggling to capture global contextual patterns." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 672, + 423, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 423, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 423, + 684 + ], + "type": "text", + "content": "3.3. Target Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Semantic Extraction: To obtain the aforementioned high-level semantic targets, we introduce a jointly optimized" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12215" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": "target generation network (TGN) composed of multiple lightweight MLPs, where each MLP specializes in producing hierarchical semantic features at specific granularity. Specifically, for each hierarchy level, we adopt simple spatiotemporal aggregation operations that process input signals within corresponding receptive fields: local-level semantic extractors operate on joints that merge across consecutive frames, while the global-level semantic extractor performs holistic feature extraction by integrating all spatiotemporal context. As an example, for the target feature " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "Z_{t_1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": ", the corresponding input is represented as " + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "inline_equation", + "content": "X_{t_1} \\in \\mathbb{R}^{((T_e / t_1)\\times V)\\times (t_1\\times l\\times C)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 214 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 218, + 296, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 296, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 296, + 313 + ], + "type": "text", + "content": "As for the input of TGN, feeding identical encoder input " + }, + { + "bbox": [ + 55, + 218, + 296, + 313 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_e" + }, + { + "bbox": [ + 55, + 218, + 296, + 313 + ], + "type": "text", + "content": " to both networks amplifies bias towards frequently unmasked joints. We devise a simple transformation strategy to get " + }, + { + "bbox": [ + 55, + 218, + 296, + 313 + ], + "type": "inline_equation", + "content": "X_{e} = X_{e}[1:] - X_{e}[: - 1]" + }, + { + "bbox": [ + 55, + 218, + 296, + 313 + ], + "type": "text", + "content": ". This geometrically grounded transformation preserves high-level motion semantics while avoiding bias. ur experiments demonstrate the remarkable effectiveness of this parameter-efficient design." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 471 + ], + "type": "text", + "content": "Information Maximization: Under the current framework where the decoder learns from TGN-generated targets while the TGN simultaneously adapts to the decoder's evolving feature representations, this co-learning scheme enables consistent high-level semantic alignment between both networks. However, such bidirectional dependency inevitably risks collapsing into trivial solutions where both networks output constant feature vectors independent of input signals, thereby losing all discriminative power. To prevent this degenerative collapse, we introduce a self-constrained feature space regularization mechanism inspired by VICReg [2], a joint embedding architecture that enforces information-maximization constraints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 473, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 568 + ], + "type": "text", + "content": "Variance Regularization. Taking the global target feature " + }, + { + "bbox": [ + 55, + 473, + 296, + 568 + ], + "type": "inline_equation", + "content": "Z_{t_g} \\in \\mathbb{R}^{B \\times C_t}" + }, + { + "bbox": [ + 55, + 473, + 296, + 568 + ], + "type": "text", + "content": " as an example, variance regularization enforces each feature dimension to maintain batch-wise variance above a predefined threshold " + }, + { + "bbox": [ + 55, + 473, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 473, + 296, + 568 + ], + "type": "text", + "content": ", ensuring persistent diversity in the learned representations and preventing pathological collapse where disparate inputs are mapped to identical feature vectors. Formally, this is implemented through a variance regularization term:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 582, + 296, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 582, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 80, + 582, + 296, + 615 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v a r} = \\frac {1}{C _ {t}} \\sum_ {i = 1} ^ {C _ {t}} \\max \\left(0, \\gamma - \\sqrt {\\operatorname {V a r} \\left(Z _ {t _ {g}} [ : , i ]\\right)}\\right) \\tag {4}", + "image_path": "d5c11be8380a18546f7cd1373f9ad0bcccc3c1e24abd70a763a96bdf6b7467de.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\operatorname{Var}(\\cdot)" + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "content": " computes empirical variance over the batch dimension " + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "content": ". Following [2], " + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 628, + 295, + 652 + ], + "type": "text", + "content": " is set 1 in our experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Covariance Regularization. The covariance constraint drives the batch-wise covariance between every feature dimension pair toward zero, eliminating inter-dimensional correlations to prevent redundant information encoding across channels. This is formalized through a covariance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 73, + 394, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 394, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 394, + 83 + ], + "type": "text", + "content": "minimization term:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 372, + 90, + 555, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 90, + 555, + 119 + ], + "spans": [ + { + "bbox": [ + 372, + 90, + 555, + 119 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c o v} = \\frac {1}{C _ {t}} \\sum_ {i \\neq j} \\left[ \\operatorname {C o v} \\left(Z _ {t _ {g}}\\right) \\right] _ {i, j} ^ {2} \\tag {5}", + "image_path": "32a974eafaea29fa3b0c2e604694488b286f78268ce64f337e3ce6ba3f37cde9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 127, + 522, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 522, + 140 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 522, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 127, + 522, + 140 + ], + "type": "inline_equation", + "content": "\\text{Cov}(Z_{t_g})" + }, + { + "bbox": [ + 313, + 127, + 522, + 140 + ], + "type": "text", + "content": " denotes covariance matrix of " + }, + { + "bbox": [ + 313, + 127, + 522, + 140 + ], + "type": "inline_equation", + "content": "Z_{t_g}" + }, + { + "bbox": [ + 313, + 127, + 522, + 140 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 140, + 555, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 555, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 555, + 176 + ], + "type": "text", + "content": "Optimization Objectives: The final information-maximization constraints are applied to all target features across hierarchy levels through unified regularization:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 347, + 183, + 555, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 183, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 347, + 183, + 555, + 209 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} = \\sum_ {j \\in \\mathcal {J}} \\left(\\alpha \\mathcal {L} _ {c o v} \\left(Z _ {t _ {j}}\\right) + \\beta \\mathcal {L} _ {v a r} \\left(Z _ {t _ {j}}\\right)\\right), \\tag {6}", + "image_path": "5315cc538a254d7664d3e101d5e84ad0fb4c904397d5e1af3190ba6f03235587.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 217, + 554, + 266 + ], + "type": "text", + "content": " are trade-off coefficients. The collaborative optimization objective for both networks combines the prediction loss with information-maximization constraints through weighted aggregation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 383, + 274, + 555, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 274, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 383, + 274, + 555, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda \\mathcal {L} _ {\\text {p r e d}} + \\mathcal {L} _ {\\text {r e g}} \\tag {7}", + "image_path": "f4c938216fffaa4bda72a0903b17501b395b7d49a23d83cdae661f33539bb710.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 295, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 554, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 295, + 554, + 366 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 295, + 554, + 366 + ], + "type": "text", + "content": " denotes trade-off parameters, and " + }, + { + "bbox": [ + 313, + 295, + 554, + 366 + ], + "type": "inline_equation", + "content": "L_{pred}" + }, + { + "bbox": [ + 313, + 295, + 554, + 366 + ], + "type": "text", + "content": " denotes the prediction loss computed over the entire batch, i.e., " + }, + { + "bbox": [ + 313, + 295, + 554, + 366 + ], + "type": "inline_equation", + "content": "L_{pred} = \\frac{1}{B}\\sum \\mathcal{L}_p" + }, + { + "bbox": [ + 313, + 295, + 554, + 366 + ], + "type": "text", + "content": ". This unified objective drives the coevolution of the decoder and TGN towards both precise hierarchical alignment and collapse-resistant feature extraction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 377, + 395, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 395, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 395, + 391 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 397, + 377, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 377, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 377, + 408 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 415, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 473 + ], + "type": "text", + "content": "This paper conducts experiments on three widely-used skeleton action benchmarks: NTU RGB+D 60 (NTU-60), NTU RGB+D 120 (NTU-120), and PKU-MMD II (PKU-II). Technical specifications and evaluation protocols for these datasets are summarized as follows." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 557 + ], + "type": "text", + "content": "NTU RGB+D 60 [24] features 60 action classes with 56,880 sequences captured from 40 participants using multi-view cameras. Two standard evaluation settings are adopted: 1) Cross-Subject (x-sub) divides data by participant IDs, using 20 subjects for training and 20 for testing; 2) Cross-View (x-view) employs camera views 2-3 for training and view 1 for testing." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 641 + ], + "type": "text", + "content": "NTU RGB+D 120 [17] extends NTU-60 to 120 action classes with 114,480 sequences from 106 subjects across 32 capture setups. Its evaluation protocols include: 1) Cross-Subject (x-sub) split with 53 subjects each for training/testing; 2) Cross-Setup (x-setup) partitioning using even-numbered setups for training and odd-numbered for testing." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "PKU-MMD II [18] focuses on viewpoint-robust action analysis with 51 categories containing 6,952 sequences. The dataset's complexity stems from significant viewpoint diversity. Following established convention, we report results under the standard cross-subject split (5,339 training/1,613 test samples)." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12216" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 113, + 535, + 203 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Skeleton-based action recognition. Comparisons to the state-of-the-art MAE-like methods for skeleton-based action recognition downstream task on NTU-60. All pre-training runs were completed under identical hardware conditions using a single RTX 4090 GPU. Enc, Dec, and TGN denote the encoder, decoder, and target generation network, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 113, + 535, + 203 + ], + "lines": [ + { + "bbox": [ + 75, + 113, + 535, + 203 + ], + "spans": [ + { + "bbox": [ + 75, + 113, + 535, + 203 + ], + "type": "table", + "html": "
MethodTargetEncDecTGNHoursSpeedupNTU-60
x-subx-view
SkeletonMAE [34]Joint1.97G17.70G-20h27m1×74.877.7
MAMP [20]Motion1.97G17.70G-20h27m1×84.989.1
S-JEPA [1]Patch-level1.97G17.70G28.32G90h57m0.2×85.389.8
GFP (Ours)Hierarchical1.97G1.57G0.64G3h14m6.2×85.992.0
", + "image_path": "87d33a644813c9705529ea0b48c0079f47a22cb02a6d4bb73f9a7852c9e1dd5c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 274, + 284, + 476 + ], + "blocks": [ + { + "bbox": [ + 55, + 209, + 297, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 297, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 297, + 266 + ], + "type": "text", + "content": "Table 2. Skeleton-based action recognition. Comparisons to the state-of-the-art methods for skeleton-based action recognition downstream task on NTU-120 and PKU-MMD II. Bold and underline formatting indicate the best and second-best methods in each group, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 274, + 284, + 476 + ], + "lines": [ + { + "bbox": [ + 69, + 274, + 284, + 476 + ], + "spans": [ + { + "bbox": [ + 69, + 274, + 284, + 476 + ], + "type": "table", + "html": "
MethodNTU-120PKU-II
x-subx-setupx-sub
other pretext task:
P&C [29]42.741.725.5
ISC [31]67.167.936.0
HaLP[23]71.172.243.5
HiCo[6]72.874.149.4
CMD [19]70.371.543.0
UmURL[30]73.574.352.1
USDRL [33]76.678.154.4
masked modeling:
SkeletonMAE[39]72.573.536.1
MAMP [20]78.679.153.8
S-JEPA [1]79.679.953.5
GFP (Ours)79.180.356.2
", + "image_path": "2837f32c17e69eb40bfa6a38e62f43dd0776e38eb58532b48c91868e8f02fc4e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 521 + ], + "type": "text", + "content": "Consistent with recent literature [19, 31], all experiments on three datasets employ top-1 accuracy as the evaluation metric." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 528, + 189, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 189, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 189, + 541 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Model Architecture. Follow [1, 20], our model utilizes a standard vision transformer backbone with learnable spatiotemporal position encoding. The encoder stacks " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "L_{e} = 8" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " identical layers, each containing multi-head attention (8 heads, 256-dim) and a feed-forward network (1024 hidden units). For pre-training, the decoder retains the same dimensions but simplifies the architecture: each local feature decoder uses one layer, while the global feature decoder adopts a 2-layer MLP with 2048-dimensional hidden units. The target generation network employs a 3-layer MLP (512 hidden units) for local semantic extraction and a 3-layer MLP (2048 hidden units) for global semantic extraction. All projectors in the decoder are 3-layer MLPs that project features to match the dimensions of the target features." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 353, + 251, + 514, + 441 + ], + "blocks": [ + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "type": "text", + "content": "Table 3. Comparisons to the state-of-the-art methods for skeleton-based action retrieval. **Bold** indicates the best methods in each group." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 353, + 251, + 514, + 441 + ], + "lines": [ + { + "bbox": [ + 353, + 251, + 514, + 441 + ], + "spans": [ + { + "bbox": [ + 353, + 251, + 514, + 441 + ], + "type": "table", + "html": "
MethodNTU-60
x-subx-view
other pretext task:
LongT GAN[44]39.148.1
P&C[29]50.776.3
ISC[31]62.582.6
HaLP[23]65.883.6
HiCo[6]68.384.8
SkeAttnCLR[10]69.476.8
UmURL[30]71.388.3
masked modeling:
MAMP [20]62.070.0
GFP (Ours)70.987.1
", + "image_path": "ec396af795345b540664ede44c90fbb110b46be40961e58f25910815e753ac4f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "text", + "content": "Data Augmentation. For skeleton sequence processing, as in previous works [1, 20], we adopt temporal cropping with dynamic scaling by randomly selecting a continuous clip spanning " + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "inline_equation", + "content": "50\\% - 100\\%" + }, + { + "bbox": [ + 313, + 460, + 555, + 533 + ], + "type": "text", + "content": " of the original sequence length. All clips are resampled to frames via bilinear interpolation to ensure temporal consistency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": "Pretraining Setup. We employ the motion-aware masking strategy proposed in [20] and segment length " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": " to 4 in the pertaining. Optimization uses AdamW " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "(\\beta_{1} = 0.9, \\beta_{2} = 0.95" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": ", weight decay 0.05) under a 400-epoch schedule. Learning rates ramp up linearly from 0 to 1e-3 during 20 warmup epochs, then decay to 5e-4 via cosine annealing. Hyperparameters " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\lambda, \\alpha" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": " are set to 5, 5, 1, respectively. For the target features, we employ a four-level hierarchical feature corresponding to local features at " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "t_{1} = 5" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "t_{2} = 10" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "inline_equation", + "content": "t_{3} = 30" + }, + { + "bbox": [ + 313, + 533, + 556, + 653 + ], + "type": "text", + "content": " and global feature." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 659, + 500, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 500, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 500, + 673 + ], + "type": "text", + "content": "4.3. Comparison to the State-of-the-art" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "In this section, we compare our method with state-of-the-art approaches on two skeleton-based tasks: action recognition and action retrieval." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12217" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 77, + 284, + 251 + ], + "blocks": [ + { + "bbox": [ + 65, + 77, + 284, + 251 + ], + "lines": [ + { + "bbox": [ + 65, + 77, + 284, + 251 + ], + "spans": [ + { + "bbox": [ + 65, + 77, + 284, + 251 + ], + "type": "image", + "image_path": "3e59af5232d4892bf3a8f85cf8ec8c98dcc05b8ba18f64ba72296b151d06d289.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 253, + 298, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 298, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 298, + 310 + ], + "type": "text", + "content": "Figure 3. Impact of Different Targets. The " + }, + { + "bbox": [ + 55, + 253, + 298, + 310 + ], + "type": "inline_equation", + "content": "5,10,30,\\mathrm{g}" + }, + { + "bbox": [ + 55, + 253, + 298, + 310 + ], + "type": "text", + "content": " configurations represent local features from 5-, 10-, 30-frame temporal spans and global representations respectively, all derived from high-level representations generated by the TGN. All results are obtained under the NTU-60 x-view protocol." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 330, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 330, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 56, + 330, + 295, + 605 + ], + "type": "text", + "content": "Skeleton-based Action Recognition. Following the standard practice from previous works [1, 20, 31], we train a linear classifier on top of the frozen encoder pre-trained with our proposed method. First, we compare our method against prior MAE-like approaches [1, 20, 34], on the NTU-60 dataset. Beyond recognition performance, we quantify computational efficiency by measuring computational FLOPs per action sequence for encoder, decoder, and target generation network, full pre-training time under the x-sub protocol (identical hardware), and training speedup relative to SkeletonMAE [34]. As shown in Table 1, our method achieves significant performance gains while substantially accelerating training. This improvement stems from reducing the optimization targets from 750 low-level targets " + }, + { + "bbox": [ + 56, + 330, + 295, + 605 + ], + "type": "inline_equation", + "content": "(\\mathbf{N} = 750)" + }, + { + "bbox": [ + 56, + 330, + 295, + 605 + ], + "type": "text", + "content": " in prior works to 251 hierarchical high-level semantic targets " + }, + { + "bbox": [ + 56, + 330, + 295, + 605 + ], + "type": "inline_equation", + "content": "(\\mathbf{M} = 251)" + }, + { + "bbox": [ + 56, + 330, + 295, + 605 + ], + "type": "text", + "content": ". Notably, while S-JPEA also adopts learned features as objectives, its patch-level implementation necessitates redundant feature encoding. This design choice, coupled with extremely slow convergence (requiring 1200 epochs versus 400 epochs for other methods), severely degrades training speed compared to SkeletonMAE. These results demonstrate our method's balanced efficacy in feature quality and computational efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": "We further validate our approach on larger-scale NTU-120 and more challenging PKU-MMD II datasets, comparing against methods using alternative pretext tasks. As shown in Table 2, our method outperforms most existing approaches, confirming its broad effectiveness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Skeleton-based Action Retrieval. In this experiment, the encoder operates in a direct feature extraction mode without requiring any training. Similar to video-to-video retrieval, we use cosine similarity in the feature space to" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 321, + 73, + 547, + 212 + ], + "blocks": [ + { + "bbox": [ + 321, + 73, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 321, + 73, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 547, + 212 + ], + "type": "image", + "image_path": "f81d9450731cd77fbb267283a53b34246a8794d95d80832db96bfa36c45072ac.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 213, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 314, + 213, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 314, + 213, + 553, + 236 + ], + "type": "text", + "content": "Figure 4. Performance of various methods with different targets. All results are obtained under the NTU-60 x-sub protocol." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 257, + 555, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 555, + 400 + ], + "type": "text", + "content": "retrieve the nearest action samples for a given query. As detailed in Table 3, conventional MAE-like methods exhibit poor performance on this task due to their focus on local reconstruction objectives without global semantic learning. In contrast, our method achieves substantial improvements (8.9% on x-sub and 17.1% on x-view protocols) over prior approaches like MAMP by incorporating high-level semantic objectives. Notably, our results remain competitive with contrastive learning methods [6, 30] that explicitly optimize global representations, demonstrating the superiority of high-level semantic targets over low-level reconstruction objectives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 408, + 408, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 408, + 408, + 421 + ], + "spans": [ + { + "bbox": [ + 314, + 408, + 408, + 421 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 449 + ], + "type": "text", + "content": "All ablation studies were conducted on the NTU-60 dataset for skeleton-based action recognition." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 450, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 641 + ], + "type": "text", + "content": "Prediction Targets: Our method employs local joint features (5-, 10-, and 30-frame spans) and global semantic features as learning targets. To investigate their individual impacts, we progressively ablated these components. As shown in Figure 3, sequential removal of global and local semantic learning objectives caused performance degradation, demonstrating the complementary nature of hierarchical feature learning. Furthermore, we explored using low-level targets instead of learned representations by removing the TGN from our framework, which forces direct prediction vectors of all spatiotemporal information within targets (e.g., flattening input into a single vector for global target). The corresponding results (labeled " + }, + { + "bbox": [ + 313, + 450, + 556, + 641 + ], + "type": "inline_equation", + "content": "\\{5,10,30,g\\}^*" + }, + { + "bbox": [ + 313, + 450, + 556, + 641 + ], + "type": "text", + "content": " in Figure 3) reveal significant performance drops, proving that high-level semantics cannot be effectively captured through direct prediction of low-level signals." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Input of Target Generation Network: For the target network inputs, we evaluated three variants: (1) raw joint coordinates, (2) masked joint data, and (3) motion features derived via linear transformation of joints. As shown in Table 4, the motion features achieved optimal performance, validating our design choice. Notably, masked" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12218" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 102, + 203, + 180 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 209, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 209, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 209, + 93 + ], + "type": "text", + "content": "Table 4. Ablation study on the input of target generation network." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 102, + 203, + 180 + ], + "lines": [ + { + "bbox": [ + 61, + 102, + 203, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 203, + 180 + ], + "type": "table", + "html": "
InputNTU 60
x-subx-view
joint85.090.9
masked joint84.290.3
motion85.992.0
", + "image_path": "4e4b49aa793f299068f8a46f161e187ee2a22f3b53f2de786073e0b8b195a499.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 242, + 102, + 365, + 180 + ], + "blocks": [ + { + "bbox": [ + 227, + 70, + 380, + 93 + ], + "lines": [ + { + "bbox": [ + 227, + 70, + 380, + 93 + ], + "spans": [ + { + "bbox": [ + 227, + 70, + 380, + 93 + ], + "type": "text", + "content": "Table 5. Ablation study on the architecture of target generation network." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 242, + 102, + 365, + 180 + ], + "lines": [ + { + "bbox": [ + 242, + 102, + 365, + 180 + ], + "spans": [ + { + "bbox": [ + 242, + 102, + 365, + 180 + ], + "type": "table", + "html": "
Arch.NTU 60
x-subx-view
Base85.992.0
Shadow85.891.8
Wider85.891.3
", + "image_path": "908205a8c1879a65bcf7a66f17bb758047b48fc67e7f10a222ab529f55b6994a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 416, + 102, + 535, + 180 + ], + "blocks": [ + { + "bbox": [ + 399, + 70, + 551, + 93 + ], + "lines": [ + { + "bbox": [ + 399, + 70, + 551, + 93 + ], + "spans": [ + { + "bbox": [ + 399, + 70, + 551, + 93 + ], + "type": "text", + "content": "Table 6. Ablation study on different number of layers used in projector." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 416, + 102, + 535, + 180 + ], + "lines": [ + { + "bbox": [ + 416, + 102, + 535, + 180 + ], + "spans": [ + { + "bbox": [ + 416, + 102, + 535, + 180 + ], + "type": "table", + "html": "
LayersNTU 60
x-subx-view
285.591.3
385.992.0
485.691.6
", + "image_path": "01dd9592bbb0a05d1ddf1c415936f1df6c32c2105fd2ac67199d729c7fbf3cc5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 62, + 217, + 284, + 514 + ], + "blocks": [ + { + "bbox": [ + 55, + 186, + 295, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 295, + 209 + ], + "type": "text", + "content": "Table 7. Comparisons to the state-of-the-art methods with semi-supervised learning on NTU-60 dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 217, + 284, + 514 + ], + "lines": [ + { + "bbox": [ + 62, + 217, + 284, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 284, + 514 + ], + "type": "table", + "html": "
Methodx-subx-view
1%10%1%10%
ASSL [27]-64.3-69.8
ISC [31]35.765.938.172.5
Colorization [38]48.371.752.578.9
Hi-TRS [3]-70.7-74.8
GL-Transformer [12])-68.6-74.9
HaLP [23]46.672.648.777.1
CMD [19]50.675.453.080.2
CPM [40]56.773.057.577.1
HYSP [9]-76.2-80.4
\\( PCM^3 \\) [41])53.877.753.182.8
HiCo [6]54.473.054.878.3
UmURL [30]58.1-58.3-
USDRL [33]57.380.260.784.0
SkeletonMAE [34]54.480.654.683.5
MAMP [20]66.088.068.791.5
S-JEPA [1]67.588.469.191.4
GFP (Ours)71.888.772.992.1
", + "image_path": "31ff49715f46f248315fcfaa65deb27ee353d0a6267e3544094ad41a2e60b525.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 521, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 653 + ], + "type": "text", + "content": "inputs hinder semantic feature extraction by TGN, aligning with observations in prior work [1]. We evaluate the epoch-wise performance of different methods with varying learning objectives: SkeletonMAE learns to reconstruct raw joints, MAMP targets motion features (i.e., temporal difference of joint coordinates), and our method predicts high-level semantic representations. As shown in Figure 4, our approach consistently outperforms prior works across all training epochs, demonstrating the superiority of hierarchical semantic learning over low-level reconstruction objectives." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Architecture Design: We conducted architectural experiments on the target generation network and projector. For TGN: (1)Shallow variant: 2-layer MLP; (2)Wider variant: 4-layer MLP with doubled hidden dimensions. (3)Base: Standard 3-layer MLP. As shown in Table 5," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 188, + 553, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 188, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 553, + 249 + ], + "type": "text", + "content": "the baseline 3-layer configuration achieves optimal performance, though all variants show marginal differences, demonstrating TGN's robustness to architectural variations. For the projector, we evaluated MLPs with 2-4 layers. The 3-layer configuration yields superior performance (Table 6)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 256, + 495, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 495, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 495, + 268 + ], + "type": "text", + "content": "4.5. Extended Performance Validation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 274, + 554, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 554, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 554, + 345 + ], + "type": "text", + "content": "To assess representation transferability, we conduct limited-label experiments on NTU-60 under two data settings: " + }, + { + "bbox": [ + 313, + 274, + 554, + 345 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 274, + 554, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 274, + 554, + 345 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 274, + 554, + 345 + ], + "type": "text", + "content": " labeled training data. Following semi-supervised learning benchmarks [31], we initialize the model with weights pre-trained through our model, then attach and fine-tune a classifier using available labels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 346, + 554, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 554, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 554, + 394 + ], + "type": "text", + "content": "As shown in Table 7, our approach establishes new state-of-the-art performance across both data scales. This significant performance gap confirms the effectiveness of our learned representations in label-efficient scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 405, + 388, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 388, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 388, + 417 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "type": "text", + "content": "In this paper, we propose an efficient general feature prediction framework for masked skeleton modeling. Unlike conventional approaches that reconstruct low-level targets, our method employs hierarchical high-level semantic features as learning targets, significantly reducing decoder complexity. The framework integrates a target generation network that provides online learning objectives. We further introduce the information maximization constraints, enabling non-trivial co-optimization between networks. This mechanism simultaneously preserves local motion details and global semantics. Extensive experiments across three datasets demonstrate the effectiveness of our method, achieving state-of-the-art performance with significant training speedup over prior masked skeleton modeling approaches." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 616, + 411, + 630 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 411, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 411, + 630 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 635, + 554, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 554, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 554, + 704 + ], + "type": "text", + "content": "This work was supported by the Pioneer and Leading Goose R&D Program of Zhejiang (No. 2024C01110), National Natural Science Foundation of China (No. 72188101, 62020106007, and 62472385). The computation is completed on the HPC Platform of Hefei University of Technology." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12219" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Mohamed Abdelfattah and Alexandre Alahi. S-jepa: A joint embedding predictive architecture for skeletal action recognition. In European Conference on Computer Vision, pages 367-384. Springer, 2024. 3, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 179 + ], + "type": "text", + "content": "[2] Adrien Bardes, Jean Ponce, and Yann Lecun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR 2022-International Conference on Learning Representations, 2022. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 235 + ], + "type": "text", + "content": "[3] Yuxiao Chen, Long Zhao, Jianbo Yuan, Yu Tian, Zhaoyang Xia, Shijie Geng, Ligong Han, and Dimitris N Metaxas. Hierarchically self-supervised transformer for human skeleton representation learning. In European Conference on Computer Vision, pages 185-202, 2022. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 290 + ], + "type": "text", + "content": "[4] Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183-192, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 355 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 355 + ], + "type": "text", + "content": "[5] Yi-Bin Cheng, Xipeng Chen, Junhong Chen, Pengxu Wei, Dongyu Zhang, and Liang Lin. Hierarchical transformer: Unsupervised representation learning for skeleton-based human action recognition. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "type": "text", + "content": "[6] Jianfeng Dong, Shengkai Sun, Zhonglin Liu, Shujie Chen, Baolong Liu, and Xun Wang. Hierarchical contrast for unsupervised skeleton-based action representation learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 525-533, 2023. 1, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 413, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 295, + 479 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 523 + ], + "type": "text", + "content": "[8] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 295, + 567 + ], + "type": "text", + "content": "[9] Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. arXiv preprint arXiv:2303.06242, 2023.8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 569, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 295, + 623 + ], + "type": "text", + "content": "[10] Yilei Hua, Wenhan Wu, Ce Zheng, Aidong Lu, Mengyuan Liu, Chen Chen, and Shiqian Wu. Part aware contrastive learning for self-supervised action recognition. In International Joint Conference on Artificial Intelligence, 2023. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "text", + "content": "[11] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3d convolutional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence, 35(1):221-231, 2012. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 295, + 713 + ], + "type": "text", + "content": "[12] Boeun Kim, Hyung Jin Chang, Jungho Kim, and Jin Young Choi. Global-local motion transformer for unsupervised skeleton-based action learning. In European Conference on Computer Vision, pages 209-225, 2022. 3, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[13] Hildegard Kuehne, Hueihan Jhuang, Estíbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 International conference on computer vision, pages 2556-2563. IEEE, 2011. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "text", + "content": "[14] Jogendra Nath Kundu, Maharshi Gor, Phani Krishna Uppala, and Venkatesh Babu Radhakrishnan. Unsupervised feature learning of human actions as trajectories in pose embedding manifold. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1459-1467, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 186, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 240 + ], + "type": "text", + "content": "[15] Linguuo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4741-4750, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 241, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 241, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 241, + 553, + 297 + ], + "type": "text", + "content": "[16] Lilang Lin, Jiahang Zhang, and Jiaying Liu. Actionlet-dependent contrastive learning for unsupervised skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2363-2372, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 298, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 352 + ], + "type": "text", + "content": "[17] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+d 120: A large-scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684-2701, 2019. 1, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 354, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 408 + ], + "type": "text", + "content": "[18] Jiaying Liu, Sijie Song, Chunhui Liu, Yanghao Li, and Yueyu Hu. A benchmark dataset and comparison study for multi-modal human action analytics. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 16(2):1-24, 2020. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 410, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 410, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 410, + 553, + 464 + ], + "type": "text", + "content": "[19] Yunyao Mao, Wengang Zhou, Zhenbo Lu, Jiajun Deng, and Houqiang Li. Cmd: Self-supervised 3d action representation learning with cross-modal mutual distillation. In European Conference on Computer Vision, pages 734-752. Springer, 2022. 3, 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 466, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 553, + 521 + ], + "type": "text", + "content": "[20] Yunyao Mao, Jiajun Deng, Wengang Zhou, Yao Fang, Wanli Ouyang, and Houqiang Li. Masked motion predictors are strong 3d action representation learners. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10181-10191, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 522, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 553, + 567 + ], + "type": "text", + "content": "[21] Qiang Nie, Ziwei Liu, and Yunhui Liu. Unsupervised 3d human pose representation with viewpoint and pose disentanglement. In European Conference on Computer Vision, pages 102-118, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 568, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 611 + ], + "type": "text", + "content": "[22] Haocong Rao, Shihao Xu, Xiping Hu, Jun Cheng, and Bin Hu. Augmented skeleton based contrastive action learning with momentum lstm for unsupervised action recognition. Information Sciences, 569:90-109, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 613, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 678 + ], + "type": "text", + "content": "[23] Anshul Shah, Aniket Roy, Ketul Shah, Shlok Mishra, David Jacobs, Anoop Cherian, and Rama Chellappa. Halp: Hallucinating latent positives for skeleton-based self-supervised learning of actions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18846-18856, 2023. 3, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "text", + "content": "[24] Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity analysis. In Proceedings of the IEEE Conference on Computer" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12220" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Vision and Pattern Recognition, pages 1010-1019, 2016. 1, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[25] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7912-7921, 2019. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "text", + "content": "[26] Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12026-12035, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "text", + "content": "[27] Chenyang Si, Xuecheng Nie, Wei Wang, Liang Wang, Ti-niu Tan, and Jiashi Feng. Adversarial self-supervised learning for semi-supervised 3d action recognition. In European Conference on Computer Vision, pages 35-51, 2020. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 282 + ], + "type": "text", + "content": "[28] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "type": "text", + "content": "[29] Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9631-9640, 2020. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "text", + "content": "[30] Shengkai Sun, Daizong Liu, Jianfeng Dong, Xiaoye Qu, Junyu Gao, Xun Yang, Xun Wang, and Meng Wang. Unified multi-modal unsupervised representation learning for skeleton-based action understanding. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2973-2984, 2023. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 437 + ], + "type": "text", + "content": "[31] Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1655–1663, 2021. 3, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 438, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 480 + ], + "type": "text", + "content": "[32] Fei Wang, Dan Guo, Kun Li, and Meng Wang. Eulermormer: Robust eulerian motion magnification via dynamic filtering within transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5345-5353, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 525 + ], + "type": "text", + "content": "[33] Wanjiang Weng, Hongsong Wang, Junbo He, Lei He, and Guosen Xie. Usdrl: Unified skeleton-based dense representation learning with multi-grained feature decorrelation. arXiv preprint arXiv:2412.09220, 2024. 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 590 + ], + "type": "text", + "content": "[34] Wenhan Wu, Yilei Hua, Ce Zheng, Shiqian Wu, Chen Chen, and Aidong Lu. Skeletonmae: Spatial-temporal masked autoencoders for self-supervised skeleton action recognition. In 2023 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 224-229. IEEE, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[35] Yuankun Xu, Dong Huang, Chang-Dong Wang, and Jian-Huang Lai. Deep image clustering with contrastive learning and multi-scale graph convolutional networks. Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 679 + ], + "type": "text", + "content": "[36] Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI Conference on Artificial Intelligence, 2018. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[37] Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garattoni, Gianpiero Francesca, and François Brémond. View-invariant skeleton action representation learning via motion" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 486 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "retargeting. International Journal of Computer Vision, pages 1-16, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[38] Siyuan Yang, Jun Liu, Shijian Lu, Meng Hwa Er, and Alex C Kot. Skeleton cloud colorization for unsupervised 3d action representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13423-13433, 2021. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[39] Siyuan Yang, Jun Liu, Shijian Lu, Er Meng Hwa, Yongjian Hu, and Alex C Kot. Self-supervised 3d action representation learning with skeleton cloud colorization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "text", + "content": "[40] Haoyuan Zhang, Yonghong Hou, Wenjing Zhang, and Wanqing Li. Contrastive positive mining for unsupervised 3d action representation learning. In European Conference on Computer Vision, pages 36-51. Springer, 2022. 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "text", + "content": "[41] Jiahang Zhang, Lilang Lin, and Jiaying Liu. Prompted contrast with masked motion modeling: Towards versatile 3d action representation learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7175-7183, 2023. 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "type": "text", + "content": "[42] Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber: Tubelet transformer for video action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13598-13607, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 375, + 553, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 429 + ], + "type": "text", + "content": "[43] Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xiaou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In Proceedings of the IEEE international conference on computer vision, pages 2914-2923, 2017. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 430, + 553, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 486 + ], + "type": "text", + "content": "[44] Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2644-2651, 2018. 1, 2, 6" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12221" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/83fba3ab-8c7e-4045-994b-8f70e4fc8d14_content_list.json b/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/83fba3ab-8c7e-4045-994b-8f70e4fc8d14_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..433a72e5957947679b39681e931f19337ca43c38 --- /dev/null +++ b/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/83fba3ab-8c7e-4045-994b-8f70e4fc8d14_content_list.json @@ -0,0 +1,1606 @@ +[ + { + "type": "text", + "text": "Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars", + "text_level": 1, + "bbox": [ + 94, + 128, + 903, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yifan Zhan $^{1,2*}$ Qingtian Zhu $^{2}$ Muyao Niu $^{2}$ Mingze Ma $^{2}$ Jiancheng Zhao $^{2}$ Zhihang Zhong $^{1\\dagger}$ Xiao Sun $^{1\\dagger}$ Yu Qiao $^{1}$ Yinqiang Zheng $^{2}$", + "bbox": [ + 173, + 200, + 823, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Artificial Intelligence Laboratory $^{2}$ The University of Tokyo", + "bbox": [ + 199, + 251, + 792, + 272 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7fcea9dc1d75506f622186a76e8e80f51a270bf5b1ddda0b24eb1872536cdb06.jpg", + "image_caption": [ + "Figure 1. We show two complex scenarios in 3D human modeling: with hand-held objects and loose-fitting clothing, which cannot be accurately represented by the standard SMPL model. Our ToMiE can realize adaptive growth to enhance the representation capability of SMPL without needing for time-consuming case-specific customization, achieving state-of-the-art results in both rendering and human (including complex scenarios) animation." + ], + "image_footnote": [], + "bbox": [ + 96, + 286, + 640, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fae2fc5ceec61ad8c8edb91c423520bcda27587f4d765ad551779d288cf1b36f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 287, + 901, + 517 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 602, + 326, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we highlight a critical yet often overlooked factor in most 3D human tasks, namely modeling complicated 3D human with with hand-held objects or loose-fitting clothing. It is known that the parameterized formulation of SMPL is able to fit human skin; while hand-held objects and loose-fitting clothing, are difficult to get modeled within the unified framework, since their movements are usually decoupled with the human body. To enhance the capability of SMPL skeleton in response to this situation, we propose a growth strategy that enables the joint tree of the skeleton to expand adaptively. Specifically, our method, called ToMiE, consists of parent joints localization and external joints optimization. For parent joints localization, we employ a gradient-based approach guided by both LBS blending weights and motion kernels. Once", + "bbox": [ + 88, + 635, + 485, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the external joints are obtained, we proceed to optimize their transformations in $SE(3)$ across different frames, enabling rendering and explicit animation. ToMiE manages to outperform other methods across various cases with hand-held objects and loose-fitting clothing, not only in rendering quality but also by offering free animation of grown joints, thereby enhancing the expressive ability of SMPL skeleton for a broader range of applications. The code is available at https://github.com/Yifever20002/ToMiE.", + "bbox": [ + 511, + 604, + 906, + 741 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 768, + 645, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D human reconstruction endeavors to model high-fidelity digital avatars based on real-world characters for virtual rendering and animating, which has been of long-term research value in areas such as gaming, virtual reality (VR), and beyond. Traditional methods, such as SMPL [36], achieve human body parameterization by performing principal component analysis (PCA) on large sets", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done during the author's internship at the Shanghai Artificial Intelligence Laboratory.† denotes co-corresponding authors.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14259", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of 3D scanned meshes, allowing for the fitting of a specified identity. Recent neural rendering techniques have enabled implicit digital human modeling guided by Linear Blend Skinning (LBS) and SMPL skeleton, realizing lifelike rendering and animating from video inputs.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The neural-based 3D human rendering has been empowered by the cutting-edge technique, 3D Gaussian Splitting (3DGS) [25], for its real-time and high-quality novel view synthesis performance. By representing 3D gaussians under the canonical T-pose and utilizing the pre-extracted SMPL skeleton in the observation space, 3D human rendering results can be obtained from novel views in any frame. This stream of approaches proves high quality in rendering 3D humans that conform to the SMPL paradigm (e.g., avatars in tight-fitting clothing). However, we raise concerns regarding its ability to handle complicated human modeling involving hand-held objects or loose-fitting clothing.", + "bbox": [ + 89, + 169, + 480, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Fig. 1, we show two cases to illustrate the limitations of current 3D human modeling. On the one hand, the movements of hand-held objects, e.g., feather duster, are highly decoupled from the human body and thus cannot be represented by the SMPL. On the other hand, characters shot in-the-wild are dressed in clothing with high complexity in dynamics, rather than the tight-fitting clothing configured under strict experimental conditions. These complicated scenarios break the existing methods' paradigm by assuming that surface of avatars should be bounded to the motion warping of SMPL skeleton in the same way as the human skin is. Rooted in the aforementioned scenarios, SMPL fails to accurately fit such 3D human models.", + "bbox": [ + 89, + 352, + 482, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we break through the limitations of modeling complicated 3D human gaussians by maintaining an extended joint tree from SMPL skeleton. Although existing SMPL model has the potential capability to manually customize additional skeleton information, time-consuming and case-by-case adjustments are necessary in this case. To overcome this issue, we extend the SMPL skeleton with additional joints for each individual case adaptively. The growth is performed in an explicit and adaptive manner, and enables the fitting of complicated 3D human avatars, offering high-quality rendering with hand-held objects and loose-fitting clothing (e.g., novel view synthesis results in Fig. 1).", + "bbox": [ + 89, + 550, + 482, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main challenge of extending the SMPL skeleton is to determine where and how to grow additional joints. To avoid the unnecessary memory consumption and potential overfitting caused by an arbitrary growth, we first determine which of the joints are supposed to serve as parent joints by a localization strategy. We have empirically observed that parent joints requiring growth witness larger backpropagation gradients in their associated gaussians due to underfitting. However, determining the association of gaussians with different joints is non-trivial, as the SMPL's LBS", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "blending weights may not work with the human with handheld objects or clothing. To more precisely define such association, we introduce the concept of Motion Kernels based on rigid body priors and combine them with LBS weights, resulting in more accurate gradient-based localization. After growth, we adaptively maintain an extended joint tree and update the extra joints by optimizing two MLP decoders for joints positions and rotations. The proposed method, termed as ToMiE, allows for explicit rendering and animation of complicated human avatars represented by extended joints.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By experiments on complicated cases of the DNA-Rendering dataset [7], ToMiE exhibits state-of-the-art rendering quality while maintaining the animatability that is significant for downstream productions. To summarize, our contributions are three-fold as follows:", + "bbox": [ + 511, + 265, + 903, + 339 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) ToMiE, a method for creating an enhanced SMPL joint tree via an adaptive growth strategy, which is able to decouple complicated parts from the human body, thereby achieving state-of-the-art results in both rendering and explicit animation on target cases;", + "2) a hybrid assignment strategy for gaussians utilizing LBS weights and Motion Kernels, combined with gradient-driven parent joints localization, to guide the growth of external joints;", + "3) a joints optimization approach fitting local rotations across different frames while sharing joints positions." + ], + "bbox": [ + 506, + 347, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 550, + 653, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. SMPL-based Human Mesh Avatars", + "text_level": 1, + "bbox": [ + 513, + 583, + 821, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most of the recent success in digital human modeling can be attributed to the contributions of the SMPL [36] series, which parameterizes the human body as individual shape components and motion-related human poses through 3D mesh scanning and PCA. The pose blend shapes in SMPL describe human body deformations as a linear weighted blending of different joint poses, significantly improving the efficiency of editing and animating digital humans. Furthermore, it has been widely adopted for human body animation, thanks to the methods [9, 49] of estimating SMPL parameters from 2D inputs. Despite their wide range of applications, SMPL and its family still suffer from inherent limitations. Since the originally scanned 3D meshes are skin-tight, according to which the pose blend shapes are learned, the model is unable to handle significantly outlying meshes, such as human with hand-held objects and clothing like skirts. Although this could potentially be solved by auto-rigging [1, 20, 28, 43, 54, 55], it requires 2D/3D shapes as inputs, which is not accessible in human videos.", + "bbox": [ + 511, + 613, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14260", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Neural Representation for 3D Human", + "text_level": 1, + "bbox": [ + 89, + 90, + 416, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Methods based on neural representations, such as NeRF [38] and 3DGS [25], have also been playing an important part in digital human reconstruction for their high-quality rendering capabilities. Early NeRF-based methods [3-6, 10-13, 27, 46, 52] aim to reconstruct human avatars by inputting monocular or multi-view synchronized videos. [50] enforce smooth priors based on neural Signed Distance Function (SDF) to obtain more accurate human geometry. Recent breakthroughs [16-18, 21, 24, 26, 29-32, 34, 35, 39-41, 47, 56, 58, 59, 61] rely on 3DGS, enabling faster and more accurate rendering. All these methods register the T-pose in a canonical space and use LBS weights to guide the rigid transformations.", + "bbox": [ + 89, + 112, + 483, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Rendering & Editing of Intricate 3D Human", + "text_level": 1, + "bbox": [ + 89, + 318, + 470, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We revisit the methods attempting to implicitly improve modeling of complicated 3D human. Animatable NeRF [45] defines a per-frame latent code to capture appearance variations across each frame. Simulation methods [2, 48] physically construct simple clothing but struggle with complex one and objects. [14, 29] additionally register global latent bones to compensate for the limitations in clothing rendering. They fail to explicitly decouple the clothing from human body, making precise control infeasible. Another stream [6, 18] leverages a human poses sequence as contexts to resolve appearance ambiguities. But correspondingly, they require a sequence of human poses for animation, adding to the challenges of editing. Moreover, these methods struggle to fit the object-level pose independent of the human poses sequence, such as hand-held items. We also note that some works [15, 37] introduce diffusion-based generative methods to enhance the realism of garment rendering, but these methods are restricted by the traditional SMPL and overlook the editing of complex clothing. Although SMPLlicit [8] enables implicit interpolation of clothing types, it remains dependent on SMPL's LBS process to generate the observed mesh. This limitation prevents localized explicit animation of loose-fitting clothing and restricts its application to external objects. Rendering-based human reconstruction method [19, 22, 23, 33, 53, 58, 60] can not achieve animation. Our approach enables explicit decoupling of hand-held objects and clothing with the human body by extending the SMPL joint tree, allowing for high-quality rendering and explicit animation in complicated scenarios.", + "bbox": [ + 89, + 340, + 483, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 808, + 230, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. SMPL(-X) Revisited", + "text_level": 1, + "bbox": [ + 89, + 832, + 284, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-trained on scanned meshes, the SMPL(-X) family [36, 44] employs a parameterized model to fit human bodies of different shapes and under different poses. The", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "human mesh in each frame evolves from a canonical human mesh and is controlled by shape and pose parameters. Specifically, a 3D point $\\pmb{x}_c$ on the canonical mesh will be warped to obtain point in the observation space as", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {o} = \\sum_ {k = 1} ^ {K} \\omega_ {k} \\left(\\boldsymbol {x} _ {c}\\right) \\left(R _ {k} \\left(\\boldsymbol {r} ^ {0}\\right) \\boldsymbol {x} _ {c} + t _ {k} \\left(\\boldsymbol {j} ^ {0}, \\beta\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 157, + 906, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $K$ is the total number of joints, $R_{k}$ is per-joint global rotation controlled by local joint rotations $\\pmb{r}^{0}$ , and $t_k$ is per-joint translation controlled by joint positions $\\pmb{j}^0$ and human shape $\\beta$ . Notice that linear blending weight $\\omega_{k}$ is a function of $\\pmb{x}_c$ , which is regressed from large human assets.", + "bbox": [ + 511, + 205, + 905, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main issue with this LBS-based prior model is that it can only model tight-fitting avatars conforming to the SMPL(-X) paradigm, making it unsuitable for modeling complex human clothing in more generic cases. What even more challenging is its inability to handle hand-held objects that are fully decoupled from the human pose.", + "bbox": [ + 511, + 282, + 905, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Human Gaussians Revisited", + "text_level": 1, + "bbox": [ + 511, + 380, + 766, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Human gaussians achieve high-quality real-time human rendering by combining the SMPL prior with 3DGS as the representation. The SMPL model naturally obtains the T-posed (i.e., all human poses are identity transformations) mesh in the canonical space and then mesh vertices are used to initialize the canonical gaussian units. Each gaussian is defined as", + "bbox": [ + 511, + 402, + 905, + 506 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (\\boldsymbol {x}) = \\frac {1}{(2 \\pi) ^ {\\frac {3}{2}} | \\boldsymbol {\\Sigma} | ^ {\\frac {1}{2}}} e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 508, + 903, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu$ is the 3D gaussian center, and $\\Sigma$ is the 3D covariance matrix, which will be further decomposed into learnable rotation $\\pmb{R}$ and scale $\\pmb{S}$ . Now we have $\\boldsymbol{\\Sigma} = \\boldsymbol{RSS}^{\\top} \\boldsymbol{R}^{\\top}$ , which is performed by optimizing a quaternion $\\boldsymbol{r}_g$ for rotation and a 3D vector $\\boldsymbol{s}_g$ for scaling. Each gaussian is further assigned with color $c$ and opacity $\\alpha$ .", + "bbox": [ + 511, + 546, + 905, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once the 3D gaussians in the canonical space are obtained, each position $\\mu$ will be warped to the observation space according to Eq. (1). Next, the 3D gaussians of each frame are projected into 2D gaussians, followed by tile-based rasterization. Color of each pixel can be calculated by blending $N$ ordered gaussians following", + "bbox": [ + 511, + 637, + 905, + 728 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 733, + 903, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The 3D gaussians can be optimized and updated by adaptive density control, which primarily includes cloning, splitting, and pruning. Cloning and splitting are guided by gradients to control the number of gaussians, while pruning removes empty gaussians based on current opacity $\\alpha$ . The supervision upon human gaussians is derived from multiview or monocular videos, enabling high-quality rendering and avatar animation.", + "bbox": [ + 511, + 779, + 906, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14261", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 89, + 89, + 189, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 illustrates ToMiE's adaptive joint growth and gaussian training strategy. Our goal is to extend the SMPL skeleton to handle complicated human scenarios. However, the abuse of growth can lead to unnecessary computational and memory overheads. For an efficient adaptive growth, we first propose a localization strategy of parent joints to ensure that only necessary joints are grown. Furthermore, we explicitly define the hand-held joints in the $SE(3)$ space and optimize them end-to-end through an MLP, ensuring alignment with the original skeleton of SMPL. The extended skeleton can thus support rendering and explicit animation. To address the limitation of LBS in guiding gaussian attributes apart from the positions, we further fine-tune the rotation and scale during training with a deformation field to achieve better non-rigid warping. Next, we will elaborate on these modules and the training strategy.", + "bbox": [ + 89, + 114, + 483, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Motion Kernels-guided Joint Gradient Accumulation", + "text_level": 1, + "bbox": [ + 89, + 368, + 483, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A quantitative metric needs to be identified to determine whether a joint requires growth. We notice that, due to the poor fitting ability of existing SMPL-based human gaussian, it will leave larger gradients in complex human regions (e.g., the hand-held object in Fig. 2 ③). In other words, identifying joints with larger accumulated gradients can help to indicate which parent joints are more likely to require growth. Bounded with each gaussian, the accumulated gradients will be first assigned to their corresponding joints. Let $g = \\|(g_x, g_y, g_z)\\|_2$ represent the L2 norm of the gradient at each gaussian position. The gradients accumulation $g_J$ for the $k$ -th joint can be computed following", + "bbox": [ + 89, + 406, + 483, + 588 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {J _ {k}} = \\frac {\\sum_ {i \\in N} \\omega_ {k} \\left(\\boldsymbol {x} _ {c}\\right) g}{\\sum_ {i \\in N} \\omega_ {k} \\left(\\boldsymbol {x} _ {c}\\right)}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 590, + 482, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use $\\pmb{x}_c$ to stand for gaussian position in canonical space and $\\pmb{x}_o$ in the observation space.", + "bbox": [ + 89, + 631, + 482, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is worth emphasizing that $\\omega_{k}$ here is a weight term determining the assignment of a gaussian to the $k$ -th joint. Under the paradigm of SMPL representation, this weight corresponds to the LBS weight $\\omega_{\\mathrm{lbs0}}$ , and guides the rigid transformation of vertices on the SMPL mesh according to the human pose. To account for the differences between the human mesh with clothing and the vanilla SMPL mesh, [17] calculate $\\omega_{\\mathrm{lbs}}$ by keeping using the LBS weight prior $\\omega_{\\mathrm{lbs0}}$ and adding an extra learnable network $\\Phi_{\\mathrm{lbs}}$ for fine-tuning. This formulation (with $k$ omitted) can be summarized as", + "bbox": [ + 89, + 661, + 483, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {\\mathrm {l b s}} \\left(\\boldsymbol {x} _ {c}\\right) = \\omega_ {\\mathrm {l b s 0}} \\left(\\mathrm {N N} _ {1} \\left(\\boldsymbol {x} _ {c}, \\boldsymbol {V}\\right)\\right) + \\Phi_ {\\mathrm {l b s}} \\left(\\boldsymbol {x} _ {c}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 818, + 482, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{NN}_1$ stands for top-1 nearest-neighbor search algorithm and $V$ represents the canonical standard SMPL vertices. It will be, however, clarified by us, that this nearest neighbor-based assigning method will no longer be feasible", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in cases with handheld objects and complex clothing.", + "bbox": [ + 511, + 90, + 866, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Fig. 3, we present an example of misclassification by the nearest-neighbor search algorithm in Eq. (5). Due to the lack of human topology constraints in the canonical space, incorrect classification of hand-held objects can occur, as shown in Fig. 3 (a). This part, following the naive nearest-neighbor search, would be assigned to the leg by mistake. Fig. 3 (b) shows the points belonging to the hand joint (the correct parent joint where the handheld object should grow). Yet being solely guided by LBS weights results in the voids by misclassification.", + "bbox": [ + 511, + 106, + 903, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To mitigate this issue, we propose a more robust assignment method based on motion priors, which we call Motion Kernels. Specifically, in the observation space, the motion kernel of each point $\\pmb{x}_o$ with respect to the $k$ -th joint position $j_k$ is defined based on the changes in their pairwise Euclidean distances through all input $N$ frames, following", + "bbox": [ + 511, + 257, + 903, + 348 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M K} \\left(\\boldsymbol {x} _ {c}, \\boldsymbol {j} _ {k}\\right) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\left\\| \\boldsymbol {x} _ {o i} - \\boldsymbol {j} _ {k i} \\right\\| _ {2} - \\mu\\right) ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 353, + 903, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| \\boldsymbol {x} _ {o i} - \\boldsymbol {j} _ {k i} \\| _ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 398, + 903, + 440 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our motion kernel (MK) reflects the relative motion between each gaussian and joint. A smaller MK indicates that the pair of gaussian and joint is relatively stationary to each other, signifying a stronger association, while a larger MK suggests a higher degree of relative motion, indicating a weaker association. We further represent the assignment weight reflected by the MK as $\\omega_{\\mathrm{MK}}(\\boldsymbol{x}_c) = \\mathrm{Normalize}_k(\\mathrm{MK}^{-1}(\\boldsymbol{x}_c,\\boldsymbol{j}_k))$ , and the final assignment weight in Eq. (4) (with index $k$ omitted) becomes", + "bbox": [ + 511, + 443, + 905, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\omega \\left(\\boldsymbol {x} _ {c}\\right) = \\lambda \\omega_ {\\mathrm {M K}} \\left(\\boldsymbol {x} _ {c}\\right) + (1 - \\lambda) \\omega_ {\\mathrm {l b s}} \\left(\\boldsymbol {x} _ {c}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 584, + 903, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter to balance the MK weight and LBS weight. Note that we do not completely abandon the LBS weight, because the MK cannot differentiate the association of points on either side of a joint (\"misdirect\" in Fig. 3), requiring the LBS weight to compensate.", + "bbox": [ + 511, + 606, + 905, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Gradient-based Parent Joint Localization", + "text_level": 1, + "bbox": [ + 511, + 689, + 867, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By combining Eq. (4) and Eq. (8), the parent joints $J_{s} \\subseteq J$ that require growth can be located. Basically, we have $g_{J} = (g_{J_{1}}, g_{J_{2}}, \\ldots, g_{J_{K}})$ to represent the gradient accumulation of total $K$ human joints. As mentioned in Sec. 4.1, the joints with larger gradients accumulated are more likely to require the growth of child joints. This is achieved by sorting $g_{J}$ in descending order $\\pi$ to get $g_{J}^{\\mathrm{sorted}} = (g_{J_{\\pi(1)}}, g_{J_{\\pi(2)}}, \\ldots, g_{J_{\\pi(K)}})$ .", + "bbox": [ + 511, + 710, + 905, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We set a gradient threshold $\\epsilon_{J}$ to identify the $J\\in J_{s}$ that requires growth, following", + "bbox": [ + 511, + 832, + 903, + 862 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {J} _ {s} = \\left(J _ {\\pi (1)}, J _ {\\pi (2)}, \\dots , J _ {\\pi (N)}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 867, + 903, + 891 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\quad g _ {J _ {\\pi (N)}} \\geq \\epsilon_ {J} \\text {a n d} g _ {J _ {\\pi (N + 1)}} < \\epsilon_ {J},\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 888, + 834, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14262", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5e78f3ac3a3766ff8f0dce7ee65b712815d4950a870f54df892dd7313de2956c.jpg", + "image_caption": [ + "Figure 2. The pipeline of ToMiE. $①$ We initialize the gaussians in the canonical space with a standard SMPL vertices. $②$ (Sec. 4.4) We apply Linear Blend Skinning (LBS) to the gaussian position and utilize a network for rotation and scale correction. During the warmup phase, Adaptive LBS only utilizes the original SMPL skeleton. After adaptive growth, it further includes the newly grown external skeleton. $③$ Gaussian rasterization and gradients backpropagation. $④$ (Sec. 4.1, Sec. 4.2) We employ a gradient-based parent joints localization method, and a motion kernel to optimize the gradient assignment process. $⑤$ (Sec. 4.3) We maintain an extra joint book with MLPs, which generates explicit human pose, enabling the decoupling and explicit animating of hand-held objects and loose-fitting clothing." + ], + "image_footnote": [], + "bbox": [ + 117, + 89, + 883, + 444 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eca421488364b38a297a7f32e72bb5af3c3725a9d0e1c3f56b5fc7b3d06d8f5c.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 96, + 553, + 267, + 650 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/53c973ec0b4b56750e6faf1ae362145bddc3502ac77ef6866a7a1970a903ace8.jpg", + "image_caption": [ + "(b)", + "Figure 3. Principle of the Motion Kernel. Relying solely on LBS weights will bring the misclassification in the canonical space shown in (a) to the observation space in (b), resulting in voids. Our proposed motion kernel focuses on motion-dependent priors in the observation space, offering better robustness and being less sensitive to misclassifications in the canonical space. This aids for point assignment process in parent joints localization." + ], + "image_footnote": [], + "bbox": [ + 269, + 553, + 478, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we can safely assume $N < K$ .", + "bbox": [ + 89, + 792, + 339, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each joint in $J_{s}$ , we designate it as a parent joint requiring growth and proceed with the initialization of its child joint. In order to explicitly model each child joint and ensure the consistency with the SMPL paradigm for ease of animation, we maintain an extra joint book $B^{e} = (\\text{parent}, j^{e}, r^{e})$ , that includes the indices of parent joints,", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the extra joint positions $j^{e}$ in canonical space and the extra rotations $r^{e}$ in its parent joint coordinate. We initialize the joint position to its parent joint's position and set the rotation to the identity rotation. The entire parent joint localization and child joint initialization process is guided by the gradients, effectively preventing unnecessary overgrowth and ensuring a dense distribution of the extra joints.", + "bbox": [ + 511, + 555, + 906, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Extra Joint Optimization", + "text_level": 1, + "bbox": [ + 511, + 676, + 743, + 693 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The joint positions and rotations in the extra joint book are set as estimizable, which will be stored and later decoded by two shallow MLPs. According to the SMPL paradigm, the canonical joint position is a time-invariant quantity, therefore the joint position optimization network $\\Phi_p$ is defined as", + "bbox": [ + 511, + 700, + 906, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd \\boldsymbol {j} ^ {e} (i) = \\Phi_ {p} (P. E. (i)), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 796, + 903, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P.E.$ is a positional encoding function as is in [38] and $i$ is the joint index in $B^{e}$ . Rotations are also dependent on the timestamp $t$ of each frame, thus the rotation optimization network $\\Phi_{r}$ is defined as", + "bbox": [ + 511, + 818, + 906, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {r} ^ {e} (i, t) = \\Phi_ {r} (P. E. (i), P. E. (t)). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 883, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14263", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Now we have the positions of extra joints $j^{e}$ (optimized by the offset $d\\boldsymbol{j}^{e}$ ) and the rotations $\\boldsymbol{r}^{e}$ . Each extra joint position is defined in the canonical space, representing the intrinsic properties of the extended skeleton, while the extra joint rotation in parent joint coordinate varies in each frame.", + "bbox": [ + 88, + 90, + 480, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although both the extra joint positions and rotations are stored in the MLPs, their inputs and outputs are explicit features with real physical meaning, which allows for both implicit and explicit editing. For example, we can interpolate over timestamp $t$ based on $\\Phi_r$ or directly use explicit inputs to take the place of $\\Phi_r$ during animation. The MLPs here function as decoders, deriving joint-related values from indices and timestamps, thus effectively circumventing the need for explicit storage of joint values employed in SMPL.", + "bbox": [ + 88, + 167, + 480, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Inference Process and Training Strategy", + "text_level": 1, + "bbox": [ + 89, + 316, + 434, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we explain how our adaptive growth is integrated with the training process of human Gaussians.", + "bbox": [ + 89, + 339, + 482, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we initialize the canonical gaussians with standard SMPL vertices. At the beginning of training, we set up a number of warm-up iterations during which no joint growth occurs, and the gaussian fitting is performed following the traditional human gaussian methods. This prevents underfitting due to insufficient training, which could further affect the joint localization in Sec. 4.2.", + "bbox": [ + 89, + 371, + 482, + 476 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "During the warm-up iterations, canonical human gaussians will be first warped to the observation space according to the LBS weight in Eq. (5). To compensate for the inability of the LBS model to represent gaussians rotation and scale, a deformable network $\\Phi_d$ is employed to correct the rotation and scale during the LBS process. To distinguish it from the joint rotation $\\pmb{r}$ of the SMPL human pose, we denote the rotation of the gaussian with subscript $g$ , and", + "bbox": [ + 89, + 478, + 482, + 599 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd \\boldsymbol {r} _ {g}, d \\boldsymbol {s} _ {g} = \\Phi_ {d} \\left(\\boldsymbol {x} _ {g}, \\boldsymbol {r} ^ {0}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 603, + 480, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The gaussians rotation $\\boldsymbol{r}_g$ and scale $s_g$ are modified with offsets $d\\boldsymbol{r}_g$ and $ds_g$ to get the final gaussians in the observation space. In the observation space, we obtain the rendered images through the rasterisation of gaussians and compute the image loss to supervise canonical gaussians.", + "bbox": [ + 89, + 626, + 482, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Once the warm-up iterations completes, we begin the adaptive joint growth. With the MK calculated during the warm-up phase, we can locate the parent joints $J_{s}$ that require growth. Then we add grown joints to the extra joint book $B^{e}$ , optimizing its positions $j^{e}$ and rotations $r^{e}$ during the subsequent training. Notably, $\\Phi_{\\mathrm{lbs}}$ needs to extend its output dimensions to include the blending weights for extra joints, following $K = K^0 + K^e$ . Since the extra joints do not have prior LBS weights, their blending weights are entirely learned through $\\Phi_{\\mathrm{lbs}}$ .", + "bbox": [ + 89, + 703, + 482, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Both the warm-up stage and the post-growth learning stage adopt adaptive density control to manage the updates of the gaussians. We also dynamically adjust the threshold", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of gradient for densification in [25] based on the number of gaussians, in order to balance memory consumption. Please check our supplemental material for details of this design.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 160, + 645, + 176 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Dataset", + "text_level": 1, + "bbox": [ + 513, + 188, + 609, + 204 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method focuses on hand-held objects and loose-fitting clothing, so we select datasets for experiments accordingly. We notice that the DNA-Rendering dataset [7], by capturing complex scenes of the human body, meets our requirements. Specifically, we selected 8 cases that align with our hypothesis, namely 0041_10, 0090_06, 0176_07, 0800_07 (hand-held objects) and 0007_04, 0014_06, 0051_09, 0811_06 (loose-fitting clothing). For each case, we use 24 surrounding views for training and 6 novel surrounding views for testing. All views are synchronized and contained 100 frames each.", + "bbox": [ + 511, + 214, + 903, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to tackle complicated scenarios, it is essential to ensure the model's performance in typical scenarios involving tight-fitting clothes. Therefore, we additionally test our method on the ZJU-MoCap [46] dataset. Although the tight-clothing cases are too simple to require joint extension, our overall framework still achieves optimal results. Since this part of the experiment is not directly related to adaptive growth, we refer the readers to the supplemental material for further visualizations.", + "bbox": [ + 511, + 383, + 905, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Baselines and Metrics", + "text_level": 1, + "bbox": [ + 513, + 539, + 718, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We select the most cutting-edge and representative works from each focus area for a fair comparison. Since 3DGS is currently the leading representation for novel view synthesis, we compare 3DGS-based methods, including 3DGS-Avatar [47], GART [29], and GauHuman [17]. Among them, GART is expected to offer extra animatability for its modeling of implicit global auxiliary bones. Additionally, there is another category of human modeling without incorporating SMPL-like pose priors. Although these methods don't guarantee an animatable human avatar, they can achieve high-quality rendering, among which, we compare the rendering quality of Im4D [33] with our method.", + "bbox": [ + 511, + 565, + 905, + 746 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct a comprehensive comparison of our ToMiE against these methods. We report three key metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [51], and learned perceptual image patch similarity (LPIPS) [57]. Per-scene results can be found in the supplemental material. In addition to comparing the rendering results, we also demonstrate the animatability of the complicated human parts enabled by our method. We strongly recommend readers to watch the supplemental video for a more intuitive understanding of the animating.", + "bbox": [ + 511, + 750, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14264", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5fe11b50405fec08130ae01efdccb76a4b4c12d64eebc1aea431479086230f59.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on the DNA-Rendering dataset [7] with animatable baselines. We show two cases of hand-held objects (0800_07) and loose-fitting clothing (0811_06) (from top to bottom). Im4d* [33] achieves high-quality rendering but cannot be animated, and we compare it in the supplemental materials. Please check the supplemental video for better visualization." + ], + "image_footnote": [], + "bbox": [ + 109, + 90, + 883, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Novel View Synthesis Results", + "text_level": 1, + "bbox": [ + 89, + 635, + 352, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4, Tab. 1 and Tab. 2 present the results of our method compared to other baselines. In the tables, we showcase two evaluation protocols. The first evaluates the entire image, reflecting the overall rendering quality. The second uses a binary mask to specifically compare the complicated regions, demonstrating how our method outperforms others in these challenging cases. The mask is shown in Fig. 4, whose details can be found in the supplemental material.", + "bbox": [ + 89, + 657, + 482, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 787, + 253, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A. Adaptive Growth Ablation. We remove the adaptive growth to ablate its impact on the rendering results. Tab. 3 \"w/o Adaptive Growth\" shows a decline in rendering quality, while the hand-held objects and loose-fitting clothing also become not animatable.", + "bbox": [ + 89, + 810, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "B. Non-rigid Design Ablation. In Sec. 4.4, we apply a", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/73eb51f03109980f4d59769194ab313c398e4cad1acfcf5cb1381a94d3a3b42e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset Method | MetricAnimatablePSNR(full) ↑DNA-RenderingPSNR(masked) ↑
BGSSIM ↑LPIPS ↓
3DGS-Avatar✓✘28.930.9530.045917.69
GART✓✓29.250.9580.048017.12
Im4D✘✘26.280.9650.030814.90
GauHuman✓✘30.220.9620.040518.26
ToMiE(ours)✓✓31.280.9660.037419.75
", + "bbox": [ + 517, + 635, + 905, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison between our method and other methods on complicated DNA-Rendering dataset [7]. $\\mathbb{B}$ and $\\mathbb{G}$ stand for human body and complex clothing (including handheld objects). We color each result as best, second best and third best. ToMiE achieves the best performance in PSNR and SSIM, while ranking second only to rendering-based Im4D [33] in LPIPS, mainly because rendering-based methods do not consider human structural constraints, resulting in higher visual fidelity.", + "bbox": [ + 511, + 715, + 906, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "non-rigid deformation network $\\Phi_d$ to correct gaussian rotation and scale. As shown in Tab. 3 \"w/o Non-rigid Design\", this significantly improves the final rendering quality.", + "bbox": [ + 511, + 854, + 903, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14265", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e6efcaac15714d75a22180e6a5dafeaa08eac5482fb2eb864066e717b5c5f9bb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset Method | MetricAnimatableZJU-Mocap
BGPSNR(full) ↑SSIM ↑LPIPS ↓
3DGS-Avatar✓×30.610.9700.0296
GART✓✓30.910.9620.0318
Im4D××28.990.9730.0620
GauHuman✓×30.820.9620.0326
ToMiE(ours)✓✓31.100.9630.0312
", + "bbox": [ + 114, + 88, + 464, + 176 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2d5b83be56019a44ab0da57c3a3fe494448321d4fda6c2ca1e984398d7a25d92.jpg", + "table_caption": [ + "Table 2. Quantitative comparison between our method and other methods on ZJU-Mocap dataset [46] with tight-fitting clothing. $\\mathbb{B}$ and $\\mathbb{G}$ stand for human body and clothing (including hand-held objects). We color each result as best, second best and third best. This comparison demonstrates that ToMiE can achieve comparable (SSIM and LPIPS) or even superior (PSNR) performance to other approaches in tight-fitting scenarios where joint growth is not required." + ], + "table_footnote": [], + "table_body": "
PSNR(full) ↑SSIM ↑LPIPS ↓PSNR(masked) ↑
w/o Adaptive Growth30.990.9640.039019.21
w/o Non-rigid Design30.780.9640.036319.38
Full31.280.9660.037419.75
", + "bbox": [ + 93, + 311, + 480, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on DNA-Rendering dataset [7]. We independently ablate the adaptive growth strategy and non-rigid design to validate their impact on the overall performance.", + "bbox": [ + 89, + 371, + 482, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Animating Hand-held Objects and Loose-fitting clothing", + "text_level": 1, + "bbox": [ + 89, + 429, + 482, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We demonstrate the uniqueness of our method, specifically its ability to explicitly animate hand-held objects and loose-fitting clothing. Our animating approach can be implemented in two ways. On the one hand, we can utilize the transformation already recorded in the extra joint book to replay clothing motions. The visualization of this part is shown in Fig. 5. On the other hand, ToMiE also supports bypassing the decoding process of the extra joint book by directly inputting the transformation explicitly. This allows us to customize the motion trajectory of external joints. Due to the limitations of the image's expressive capabilities, we strongly recommend the readers to watch the supplemental video to check the animated results.", + "bbox": [ + 88, + 465, + 482, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 5, we edit the extra joints while keeping the body poses under the SMPL paradigm stationary. Since the implicit auxiliary bones of GART [29] is controlled by the traditional SMPL poses, only the identical appearance can be output when SMPL poses are stationary. In contrast, our method explicitly models hand-held objects and loose-fitting clothing, fully decoupling them from the traditional SMPL poses, enabling free animating.", + "bbox": [ + 89, + 662, + 482, + 785 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Conclusion", + "text_level": 1, + "bbox": [ + 89, + 799, + 346, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Although our method enhances the modeling for rigid and non-rigid clothing, it cannot address scenarios involving drastic changes in the topology (e.g., taking off clothes or opening a book). This is because topological changes disrupt the one-to-one correspondence be", + "bbox": [ + 89, + 824, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c9bb521bca23a855504ace1df68a04ec2f22d9f4e028c7a110c35bfefab242a.jpg", + "image_caption": [ + "Figure 5. Animating Results of ToMiE. Our explicit modeling fully decouples hand-held objects and loose-fitting clothing from the human body, enabling part-specific animating." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tween frames, making the human modeling centered on the canonical space become downgraded. We notice that [42] addresses topology issues by introducing high-dimensional mappings, which could be adapted to build our non-rigid deformation. However, this is not the main scope of this paper and can be explored as future work.", + "bbox": [ + 511, + 522, + 906, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this paper, we introduce ToMiE, an adaptive growth method designed to extend traditional SMPL skeleton for better modeling of hand-held objects and loose-fitting clothing. In the first stage, we assign the gradient of gaussian points to different joints by combining LBS weights with the motion kernel based on motion priors. This allows us to accurately locate the parent joints that need to grow, avoiding redundant growth. In the second stage, we design an extra joint book to achieve explicit joint modeling and optimize the transformation of the newly grown joints in an end-to-end manner. With the improved designs mentioned above, our ToMiE stands out among numerous state-of-the-art methods, achieving the best rendering quality and animatability of hand-held objects and loose-fitting clothing. We hope the adaptive growing method will spark a renewed discussion on the current capabilities of digital human modeling. What's more, it is expected to offer some insights for subsequent works related to topology and skeleton generation.", + "bbox": [ + 511, + 613, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14266", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported in part by the Tateisi Science and Technology Foundation, JSPS KAKENHI Grant Numbers 24KK0209 and 22H00529. This research was also supported by Shanghai Artificial Intelligence Laboratory.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 205, + 187, + 220 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] I. Baran and J. Popović. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG), 26(3):72-es, 2007. 2", + "[2] H. Bertiche, M. Madadi, and S. Escalera. Neural cloth simulation. ACM Transactions on Graphics (TOG), 41(6):1-14, 2022. 3", + "[3] J. Chen, Y. Zhang, D. Kang, X. Zhe, L. Bao, X. Jia, and H. Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 3", + "[4] X. Chen, T. Jiang, J. Song, M. Rietmann, A. Geiger, M. J. Black, and O. Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11796-11809, 2023.", + "[5] X. Chen, Y. Zheng, M. J. Black, O. Hilliges, and A. Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11594-11604, 2021.", + "[6] Y. Chen, Y. Zhan, Z. Zhong, W. Wang, X. Sun, Y. Qiao, and Y. Zheng. Within the Dynamic Context: Inertia-aware 3D Human Modeling with Pose Sequence. arXiv preprint arXiv:2403.19160, 2024. 3", + "[7] W. Cheng, R. Chen, S. Fan, W. Yin, K. Chen, Z. Cai, J. Wang, Y. Gao, Z. Yu, Z. Lin, et al. DNA-Rendering: A Diverse Neural Actor Repository for High-fidelity Human-centric Rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 2, 6, 7, 8", + "[8] E. Corona, A. Pumarola, G. Alenya, G. Pons-Moll, and F. Moreno-Noguer. SMPlicit: Topology-aware generative model for clothed people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11875-11885, 2021. 3", + "[9] J. Dong, Q. Fang, W. Jiang, Y. Yang, H. Bao, and X. Zhou. Fast and robust multi-person 3d pose estimation and tracking from multiple views. In T-PAMI, 2021. 2", + "[10] G. Gafni, J. Thies, M. Zollhofer, and M. Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3", + "[11] Q. Gao, Y. Wang, L. Liu, L. Liu, C. Theobalt, and B. Chen. Neural novel actor: Learning a generalized animatable neural representation for human actors. IEEE Transactions on Visualization and Computer Graphics, 2023.", + "[12] C. Geng, S. Peng, Z. Xu, H. Bao, and X. Zhou. Learning neural volumetric representations of dynamic humans" + ], + "bbox": [ + 93, + 231, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023.", + "[13] S. Goel, G. Pavlakos, J. Rajasegaran, A. Kanazawa, and J. Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 3", + "[14] C. Guo, T. Jiang, M. Kaufmann, C. Zheng, J. Valentin, J. Song, and O. Hilliges. ReLoo: Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild. arXiv preprint arXiv:2409.15269, 2024. 3", + "[15] L. Hu. animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3", + "[16] L. Hu, H. Zhang, Y. Zhang, B. Zhou, B. Liu, S. Zhang, and L. Nie. Gaussianavatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 3", + "[17] S. Hu, T. Hu, and Z. Liu. Gauhuman: Articulated Gaussian Splatting from Monocular Human Videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 4, 6", + "[18] T. Hu, F. Hong, and Z. Liu. Surmo: Surface-based 4d motion modeling for dynamic human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6550-6560, 2024. 3", + "[19] M. Isik, M. Rünz, M. Georgopoulos, T. Khakhulin, J. Starck, L. Agapito, and M. Nießner. Humanrf: High-fidelity Neural Radiance Fields for Humans in Motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 3", + "[20] A. Jacobson, I. Baran, L. Kavan, J. Popovic, and O. Sorkine. Fast automatic skinning transformations. ACM Transactions on Graphics (ToG), 31(4):1-10, 2012. 2", + "[21] R. Jena, G. S. Iyer, S. Choudhary, B. Smith, P. Chaudhari, and J. Gee. Splatarmor: Articulated Gaussian Splatting for Animatable Humans from Monocular RGB Videos. arXiv preprint arXiv:2311.10812, 2023. 3", + "[22] Y. Jiang, Z. Shen, Y. Hong, C. Guo, Y. Wu, Y. Zhang, J. Yu, and L. Xu. Robust Dual Gaussian Splatting for Immersive Human-centric Volumetric Videos. ACM Transactions on Graphics (TOG), 43(6):1-15, 2024. 3", + "[23] Y. Jiang, Z. Shen, P. Wang, Z. Su, Y. Hong, Y. Zhang, J. Yu, and L. Xu. Hifi4g: High-fidelity Human Performance Rendering via Compact Gaussian Splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19734-19745, 2024. 3", + "[24] H. Jung, N. Brasch, J. Song, E. Perez-Pellitero, Y. Zhou, Z. Li, N. Navab, and B. Busam. Deformable 3D Gaussian Splitting for Animatable Human Avatars. arXiv preprint arXiv:2312.15059, 2023. 3", + "[25] B. Kerbl, G. Kopanas, T. Leimkuhler, and G. Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 3, 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14267", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] M. Kocabas, J.-H. R. Chang, J. Gabriel, O. Tuzel, and A. Ranjan. Hugs: Human Gaussian Splats. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 505-515, 2024. 3", + "[27] Y. Kwon, D. Kim, D. Ceylan, and H. Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 3", + "[28] B. H. Le and Z. Deng. Two-layer sparse compression of dense-weight blend skinning. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 2", + "[29] J. Lei, Y. Wang, G. Pavlakos, L. Liu, and K. Daniilidis. Gart: Gaussian Articulated Template Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19876-19887, 2024. 3, 6, 8", + "[30] M. Li, J. Tao, Z. Yang, and Y. Yang. Human101: Training 100+ fps Human Gaussians in 100s from 1 View. arXiv preprint arXiv:2312.15258, 2023.", + "[31] M. Li, S. Yao, Z. Xie, K. Chen, and Y.-G. Jiang. Gaussian-body: Clothed human reconstruction via 3d gaussian splatt-ing. arXiv preprint arXiv:2401.09720, 2024.", + "[32] Z. Li, Z. Zheng, L. Wang, and Y. Liu. Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 3", + "[33] H. Lin, S. Peng, Z. Xu, T. Xie, X. He, H. Bao, and X. Zhou. High-Fidelity and Real-Time Novel View Synthesis for Dynamic Scenes. In SIGGRAPH Asia Conference Proceedings, 2023. 3, 6, 7", + "[34] X. Liu, C. Wu, X. Liu, J. Liu, J. Wu, C. Zhao, H. Feng, E. Ding, and J. Wang. Gea: Reconstructing expressive 3d gaussian avatar from monocular video. arXiv preprint arXiv:2402.16607, 2024. 3", + "[35] Y. Liu, X. Huang, M. Qin, Q. Lin, and H. Wang. Animatable 3d gaussian: Fast and high-quality reconstruction of multiple human avatars. arXiv preprint arXiv:2311.16482, 2023. 3", + "[36] M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black. SMPL: A Skinned Multi-Person Linear Model. ACM Transactions on Graphics, 34(6), 2015. 1, 2, 3", + "[37] Y. Men, Y. Yao, M. Cui, and L. Bo. Mimo: Controllable character video synthesis with spatial decomposed modeling. arXiv preprint arXiv:2409.16160, 2024. 3", + "[38] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 3, 5", + "[39] A. Moreau, J. Song, H. Dhamo, R. Shaw, Y. Zhou, and E. Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 3", + "[40] M. Niu, M. Cao, Y. Zhan, Q. Zhu, M. Ma, J. Zhao, Y. Zeng, Z. Zhong, X. Sun, and Y. Zheng. Anicrafter: Customizing realistic human-centric animation via avatar-background conditioning in video diffusion models. arXiv preprint arXiv:2505.20255, 2025." + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] H. Pang, H. Zhu, A. Kortylewski, C. Theobalt, and M. Habermann. Ash: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1165-1175, 2024. 3", + "[42] K. Park, U. Sinha, P. Hedman, J. T. Barron, S. Bouaziz, D. B. Goldman, R. Martin-Brualla, and S. M. Seitz. Hypernerf: A Higher-dimensional Representation for Topologically Varying Neural Radiance Fields. arXiv preprint arXiv:2106.13228, 2021. 8", + "[43] D. Paschalidou, A. Katharopoulos, A. Geiger, and S. Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3204-3215, 2021. 2", + "[44] G. Pavlakos, V. Choutas, N. Ghorbani, T. Bolkart, A. A. A. Osman, D. Tzionas, and M. J. Black. Expressive Body Capture: 3D Hands, Face, and Body from a Single Image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 3", + "[45] S. Peng, J. Dong, Q. Wang, S. Zhang, Q. Shuai, X. Zhou, and H. Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 3", + "[46] S. Peng, Y. Zhang, Y. Xu, Q. Wang, Q. Shuai, H. Bao, and X. Zhou. Neural Body: Implicit Neural Representations with Structured Latent Codes for Novel View Synthesis of Dynamic Humans. In CVPR, 2021. 3, 6, 8", + "[47] Z. Qian, S. Wang, M. Mihajlovic, A. Geiger, and S. Tang. 3dgs-avatar: Animatable Avatars via Deformable 3d Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020–5030, 2024. 3, 6", + "[48] I. Santesteban, M. A. Otaduy, and D. Casas. Snug: self-supervised neural dynamic garments. IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), volume 2, page 9, 2022. 3", + "[49] Q. Shuai, C. Geng, Q. Fang, S. Peng, W. Shen, X. Zhou, and H. Bao. Novel view synthesis of human interactions from sparse multi-view videos. In SIGGRAPH Conference Proceedings, 2022. 2", + "[50] S. Wang, K. Schwarz, A. Geiger, and S. Tang. A随机变量的统计学，1987年，第2版，M. H. Lai, The asymptotic analysis of a random variable with a finite distribution, 1987. Springer, 2003.", + "[51] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[52] C.-Y. Weng, B. Curless, P. P. Srinivasan, J. T. Barron, and I. Kemelmacher-Shlizerman. Humannerf: Free-viewpoint Rendering of Moving People from Monocular Video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 3", + "[53] Z. Xu, S. Peng, H. Lin, G. He, J. Sun, Y. Shen, H. Bao, and X. Zhou. 4K4D: Real-time 4D View Synthesis at 4K Resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20029-20040, 2024. 3" + ], + "bbox": [ + 514, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14268", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Z. Xu, Y. Zhou, E. Kalogerakis, C. Landreth, and K. Singh. Rignet: Neural rigging for articulated characters. arXiv preprint arXiv:2005.00559, 2020. 2", + "[55] C.-H. Yao, W.-C. Hung, Y. Li, M. Rubinstein, M.-H. Yang, and V. Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4853-4862, 2023. 2", + "[56] Y. Zhan, W. Xu, Q. Zhu, M. Niu, M. Ma, Y. Liu, Z. Zhong, X. Sun, and Y. Zheng. R3-avatar: Record and retrieve temporal codebook for reconstructing photorealistic human avatars. arXiv preprint arXiv:2503.12751, 2025. 3", + "[57] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[58] S. Zheng, B. Zhou, R. Shao, B. Liu, S. Zhang, L. Nie, and Y. Liu. Gps-gaussian: Generalizable Pixel-wise 3D Gaussian Splatting for Real-time Human Novel View Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 3", + "[59] Y. Zheng, Q. Zhao, G. Yang, W. Yifan, D. Xiang, F. Dubost, D. Lagun, T. Beeler, F. Tombari, L. Guibas, et al. Physavatar: Learning the physics of dressed 3d avatars from visual observations. arXiv preprint arXiv:2404.04421, 2024. 3", + "[60] B. Zhou, S. Zheng, H. Tu, R. Shao, B. Liu, S. Zhang, L. Nie, and Y. Liu. GPS-Gaussian+: Generalizable Pixel-wise 3D Gaussian Splatting for Real-Time Human-Scene Rendering from Sparse Views. arXiv preprint arXiv:2411.11363, 2024. 3", + "[61] W. Zielonka, T. Bagautdinov, S. Saito, M. Zollhöfer, J. Thies, and J. Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + ], + "bbox": [ + 91, + 90, + 480, + 584 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14269", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/83fba3ab-8c7e-4045-994b-8f70e4fc8d14_model.json b/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/83fba3ab-8c7e-4045-994b-8f70e4fc8d14_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1c50303a28a250ca7e1b8e50fb7378b4ac2a9469 --- /dev/null +++ b/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/83fba3ab-8c7e-4045-994b-8f70e4fc8d14_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.096, + 0.13, + 0.904, + 0.175 + ], + "angle": 0, + "content": "Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.202, + 0.825, + 0.245 + ], + "angle": 0, + "content": "Yifan Zhan\\(^{1,2*}\\) Qingtian Zhu\\(^{2}\\) Muyao Niu\\(^{2}\\) Mingze Ma\\(^{2}\\) Jiancheng Zhao\\(^{2}\\) Zhihang Zhong\\(^{1\\dagger}\\) Xiao Sun\\(^{1\\dagger}\\) Yu Qiao\\(^{1}\\) Yinqiang Zheng\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.252, + 0.794, + 0.273 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Artificial Intelligence Laboratory \\(^{2}\\)The University of Tokyo" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.287, + 0.641, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.289, + 0.902, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.53, + 0.908, + 0.586 + ], + "angle": 0, + "content": "Figure 1. We show two complex scenarios in 3D human modeling: with hand-held objects and loose-fitting clothing, which cannot be accurately represented by the standard SMPL model. Our ToMiE can realize adaptive growth to enhance the representation capability of SMPL without needing for time-consuming case-specific customization, achieving state-of-the-art results in both rendering and human (including complex scenarios) animation." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.603, + 0.327, + 0.619 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.636, + 0.486, + 0.864 + ], + "angle": 0, + "content": "In this paper, we highlight a critical yet often overlooked factor in most 3D human tasks, namely modeling complicated 3D human with with hand-held objects or loose-fitting clothing. It is known that the parameterized formulation of SMPL is able to fit human skin; while hand-held objects and loose-fitting clothing, are difficult to get modeled within the unified framework, since their movements are usually decoupled with the human body. To enhance the capability of SMPL skeleton in response to this situation, we propose a growth strategy that enables the joint tree of the skeleton to expand adaptively. Specifically, our method, called ToMiE, consists of parent joints localization and external joints optimization. For parent joints localization, we employ a gradient-based approach guided by both LBS blending weights and motion kernels. Once" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.907, + 0.742 + ], + "angle": 0, + "content": "the external joints are obtained, we proceed to optimize their transformations in \\( SE(3) \\) across different frames, enabling rendering and explicit animation. ToMiE manages to outperform other methods across various cases with hand-held objects and loose-fitting clothing, not only in rendering quality but also by offering free animation of grown joints, thereby enhancing the expressive ability of SMPL skeleton for a broader range of applications. The code is available at https://github.com/Yifever20002/ToMiE." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.77, + 0.646, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "3D human reconstruction endeavors to model high-fidelity digital avatars based on real-world characters for virtual rendering and animating, which has been of long-term research value in areas such as gaming, virtual reality (VR), and beyond. Traditional methods, such as SMPL [36], achieve human body parameterization by performing principal component analysis (PCA) on large sets" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*This work was done during the author's internship at the Shanghai Artificial Intelligence Laboratory.† denotes co-corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "14259" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "of 3D scanned meshes, allowing for the fitting of a specified identity. Recent neural rendering techniques have enabled implicit digital human modeling guided by Linear Blend Skinning (LBS) and SMPL skeleton, realizing lifelike rendering and animating from video inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.482, + 0.351 + ], + "angle": 0, + "content": "The neural-based 3D human rendering has been empowered by the cutting-edge technique, 3D Gaussian Splitting (3DGS) [25], for its real-time and high-quality novel view synthesis performance. By representing 3D gaussians under the canonical T-pose and utilizing the pre-extracted SMPL skeleton in the observation space, 3D human rendering results can be obtained from novel views in any frame. This stream of approaches proves high quality in rendering 3D humans that conform to the SMPL paradigm (e.g., avatars in tight-fitting clothing). However, we raise concerns regarding its ability to handle complicated human modeling involving hand-held objects or loose-fitting clothing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.549 + ], + "angle": 0, + "content": "In Fig. 1, we show two cases to illustrate the limitations of current 3D human modeling. On the one hand, the movements of hand-held objects, e.g., feather duster, are highly decoupled from the human body and thus cannot be represented by the SMPL. On the other hand, characters shot in-the-wild are dressed in clothing with high complexity in dynamics, rather than the tight-fitting clothing configured under strict experimental conditions. These complicated scenarios break the existing methods' paradigm by assuming that surface of avatars should be bounded to the motion warping of SMPL skeleton in the same way as the human skin is. Rooted in the aforementioned scenarios, SMPL fails to accurately fit such 3D human models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.748 + ], + "angle": 0, + "content": "To this end, we break through the limitations of modeling complicated 3D human gaussians by maintaining an extended joint tree from SMPL skeleton. Although existing SMPL model has the potential capability to manually customize additional skeleton information, time-consuming and case-by-case adjustments are necessary in this case. To overcome this issue, we extend the SMPL skeleton with additional joints for each individual case adaptively. The growth is performed in an explicit and adaptive manner, and enables the fitting of complicated 3D human avatars, offering high-quality rendering with hand-held objects and loose-fitting clothing (e.g., novel view synthesis results in Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The main challenge of extending the SMPL skeleton is to determine where and how to grow additional joints. To avoid the unnecessary memory consumption and potential overfitting caused by an arbitrary growth, we first determine which of the joints are supposed to serve as parent joints by a localization strategy. We have empirically observed that parent joints requiring growth witness larger backpropagation gradients in their associated gaussians due to underfitting. However, determining the association of gaussians with different joints is non-trivial, as the SMPL's LBS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "blending weights may not work with the human with handheld objects or clothing. To more precisely define such association, we introduce the concept of Motion Kernels based on rigid body priors and combine them with LBS weights, resulting in more accurate gradient-based localization. After growth, we adaptively maintain an extended joint tree and update the extra joints by optimizing two MLP decoders for joints positions and rotations. The proposed method, termed as ToMiE, allows for explicit rendering and animation of complicated human avatars represented by extended joints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.905, + 0.34 + ], + "angle": 0, + "content": "By experiments on complicated cases of the DNA-Rendering dataset [7], ToMiE exhibits state-of-the-art rendering quality while maintaining the animatability that is significant for downstream productions. To summarize, our contributions are three-fold as follows:" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.348, + 0.905, + 0.424 + ], + "angle": 0, + "content": "1) ToMiE, a method for creating an enhanced SMPL joint tree via an adaptive growth strategy, which is able to decouple complicated parts from the human body, thereby achieving state-of-the-art results in both rendering and explicit animation on target cases;" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.424, + 0.905, + 0.484 + ], + "angle": 0, + "content": "2) a hybrid assignment strategy for gaussians utilizing LBS weights and Motion Kernels, combined with gradient-driven parent joints localization, to guide the growth of external joints;" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.485, + 0.905, + 0.514 + ], + "angle": 0, + "content": "3) a joints optimization approach fitting local rotations across different frames while sharing joints positions." + }, + { + "type": "list", + "bbox": [ + 0.507, + 0.348, + 0.905, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.551, + 0.655, + 0.567 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.584, + 0.822, + 0.6 + ], + "angle": 0, + "content": "2.1. SMPL-based Human Mesh Avatars" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Most of the recent success in digital human modeling can be attributed to the contributions of the SMPL [36] series, which parameterizes the human body as individual shape components and motion-related human poses through 3D mesh scanning and PCA. The pose blend shapes in SMPL describe human body deformations as a linear weighted blending of different joint poses, significantly improving the efficiency of editing and animating digital humans. Furthermore, it has been widely adopted for human body animation, thanks to the methods [9, 49] of estimating SMPL parameters from 2D inputs. Despite their wide range of applications, SMPL and its family still suffer from inherent limitations. Since the originally scanned 3D meshes are skin-tight, according to which the pose blend shapes are learned, the model is unable to handle significantly outlying meshes, such as human with hand-held objects and clothing like skirts. Although this could potentially be solved by auto-rigging [1, 20, 28, 43, 54, 55], it requires 2D/3D shapes as inputs, which is not accessible in human videos." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14260" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.417, + 0.108 + ], + "angle": 0, + "content": "2.2. Neural Representation for 3D Human" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.31 + ], + "angle": 0, + "content": "Methods based on neural representations, such as NeRF [38] and 3DGS [25], have also been playing an important part in digital human reconstruction for their high-quality rendering capabilities. Early NeRF-based methods [3-6, 10-13, 27, 46, 52] aim to reconstruct human avatars by inputting monocular or multi-view synchronized videos. [50] enforce smooth priors based on neural Signed Distance Function (SDF) to obtain more accurate human geometry. Recent breakthroughs [16-18, 21, 24, 26, 29-32, 34, 35, 39-41, 47, 56, 58, 59, 61] rely on 3DGS, enabling faster and more accurate rendering. All these methods register the T-pose in a canonical space and use LBS weights to guide the rigid transformations." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.319, + 0.471, + 0.336 + ], + "angle": 0, + "content": "2.3. Rendering & Editing of Intricate 3D Human" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.485, + 0.795 + ], + "angle": 0, + "content": "We revisit the methods attempting to implicitly improve modeling of complicated 3D human. Animatable NeRF [45] defines a per-frame latent code to capture appearance variations across each frame. Simulation methods [2, 48] physically construct simple clothing but struggle with complex one and objects. [14, 29] additionally register global latent bones to compensate for the limitations in clothing rendering. They fail to explicitly decouple the clothing from human body, making precise control infeasible. Another stream [6, 18] leverages a human poses sequence as contexts to resolve appearance ambiguities. But correspondingly, they require a sequence of human poses for animation, adding to the challenges of editing. Moreover, these methods struggle to fit the object-level pose independent of the human poses sequence, such as hand-held items. We also note that some works [15, 37] introduce diffusion-based generative methods to enhance the realism of garment rendering, but these methods are restricted by the traditional SMPL and overlook the editing of complex clothing. Although SMPLlicit [8] enables implicit interpolation of clothing types, it remains dependent on SMPL's LBS process to generate the observed mesh. This limitation prevents localized explicit animation of loose-fitting clothing and restricts its application to external objects. Rendering-based human reconstruction method [19, 22, 23, 33, 53, 58, 60] can not achieve animation. Our approach enables explicit decoupling of hand-held objects and clothing with the human body by extending the SMPL joint tree, allowing for high-quality rendering and explicit animation in complicated scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.809, + 0.231, + 0.825 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.285, + 0.849 + ], + "angle": 0, + "content": "3.1. SMPL(-X) Revisited" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Pre-trained on scanned meshes, the SMPL(-X) family [36, 44] employs a parameterized model to fit human bodies of different shapes and under different poses. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "human mesh in each frame evolves from a canonical human mesh and is controlled by shape and pose parameters. Specifically, a 3D point \\( \\pmb{x}_c \\) on the canonical mesh will be warped to obtain point in the observation space as" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.159, + 0.907, + 0.201 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {o} = \\sum_ {k = 1} ^ {K} \\omega_ {k} \\left(\\boldsymbol {x} _ {c}\\right) \\left(R _ {k} \\left(\\boldsymbol {r} ^ {0}\\right) \\boldsymbol {x} _ {c} + t _ {k} \\left(\\boldsymbol {j} ^ {0}, \\beta\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.906, + 0.282 + ], + "angle": 0, + "content": "where \\(K\\) is the total number of joints, \\(R_{k}\\) is per-joint global rotation controlled by local joint rotations \\(\\pmb{r}^{0}\\), and \\(t_k\\) is per-joint translation controlled by joint positions \\(\\pmb{j}^0\\) and human shape \\(\\beta\\). Notice that linear blending weight \\(\\omega_{k}\\) is a function of \\(\\pmb{x}_c\\), which is regressed from large human assets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.283, + 0.906, + 0.373 + ], + "angle": 0, + "content": "The main issue with this LBS-based prior model is that it can only model tight-fitting avatars conforming to the SMPL(-X) paradigm, making it unsuitable for modeling complex human clothing in more generic cases. What even more challenging is its inability to handle hand-held objects that are fully decoupled from the human pose." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.767, + 0.395 + ], + "angle": 0, + "content": "3.2. Human Gaussians Revisited" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.507 + ], + "angle": 0, + "content": "Human gaussians achieve high-quality real-time human rendering by combining the SMPL prior with 3DGS as the representation. The SMPL model naturally obtains the T-posed (i.e., all human poses are identity transformations) mesh in the canonical space and then mesh vertices are used to initialize the canonical gaussian units. Each gaussian is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.51, + 0.905, + 0.543 + ], + "angle": 0, + "content": "\\[\nG (\\boldsymbol {x}) = \\frac {1}{(2 \\pi) ^ {\\frac {3}{2}} | \\boldsymbol {\\Sigma} | ^ {\\frac {1}{2}}} e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.906, + 0.638 + ], + "angle": 0, + "content": "where \\(\\mu\\) is the 3D gaussian center, and \\(\\Sigma\\) is the 3D covariance matrix, which will be further decomposed into learnable rotation \\(\\pmb{R}\\) and scale \\(\\pmb{S}\\). Now we have \\(\\boldsymbol{\\Sigma} = \\boldsymbol{RSS}^{\\top} \\boldsymbol{R}^{\\top}\\), which is performed by optimizing a quaternion \\(\\boldsymbol{r}_g\\) for rotation and a 3D vector \\(\\boldsymbol{s}_g\\) for scaling. Each gaussian is further assigned with color \\(c\\) and opacity \\(\\alpha\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.906, + 0.729 + ], + "angle": 0, + "content": "Once the 3D gaussians in the canonical space are obtained, each position \\(\\mu\\) will be warped to the observation space according to Eq. (1). Next, the 3D gaussians of each frame are projected into 2D gaussians, followed by tile-based rasterization. Color of each pixel can be calculated by blending \\(N\\) ordered gaussians following" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.734, + 0.905, + 0.775 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.9 + ], + "angle": 0, + "content": "The 3D gaussians can be optimized and updated by adaptive density control, which primarily includes cloning, splitting, and pruning. Cloning and splitting are guided by gradients to control the number of gaussians, while pruning removes empty gaussians based on current opacity \\(\\alpha\\). The supervision upon human gaussians is derived from multiview or monocular videos, enabling high-quality rendering and avatar animation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14261" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.191, + 0.106 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.485, + 0.358 + ], + "angle": 0, + "content": "Fig. 2 illustrates ToMiE's adaptive joint growth and gaussian training strategy. Our goal is to extend the SMPL skeleton to handle complicated human scenarios. However, the abuse of growth can lead to unnecessary computational and memory overheads. For an efficient adaptive growth, we first propose a localization strategy of parent joints to ensure that only necessary joints are grown. Furthermore, we explicitly define the hand-held joints in the \\(SE(3)\\) space and optimize them end-to-end through an MLP, ensuring alignment with the original skeleton of SMPL. The extended skeleton can thus support rendering and explicit animation. To address the limitation of LBS in guiding gaussian attributes apart from the positions, we further fine-tune the rotation and scale during training with a deformation field to achieve better non-rigid warping. Next, we will elaborate on these modules and the training strategy." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.369, + 0.484, + 0.399 + ], + "angle": 0, + "content": "4.1. Motion Kernels-guided Joint Gradient Accumulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.589 + ], + "angle": 0, + "content": "A quantitative metric needs to be identified to determine whether a joint requires growth. We notice that, due to the poor fitting ability of existing SMPL-based human gaussian, it will leave larger gradients in complex human regions (e.g., the hand-held object in Fig. 2 ③). In other words, identifying joints with larger accumulated gradients can help to indicate which parent joints are more likely to require growth. Bounded with each gaussian, the accumulated gradients will be first assigned to their corresponding joints. Let \\( g = \\|(g_x, g_y, g_z)\\|_2 \\) represent the L2 norm of the gradient at each gaussian position. The gradients accumulation \\( g_J \\) for the \\( k \\)-th joint can be computed following" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.592, + 0.483, + 0.629 + ], + "angle": 0, + "content": "\\[\ng _ {J _ {k}} = \\frac {\\sum_ {i \\in N} \\omega_ {k} \\left(\\boldsymbol {x} _ {c}\\right) g}{\\sum_ {i \\in N} \\omega_ {k} \\left(\\boldsymbol {x} _ {c}\\right)}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.483, + 0.661 + ], + "angle": 0, + "content": "We use \\( \\pmb{x}_c \\) to stand for gaussian position in canonical space and \\( \\pmb{x}_o \\) in the observation space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.662, + 0.484, + 0.812 + ], + "angle": 0, + "content": "It is worth emphasizing that \\(\\omega_{k}\\) here is a weight term determining the assignment of a gaussian to the \\(k\\)-th joint. Under the paradigm of SMPL representation, this weight corresponds to the LBS weight \\(\\omega_{\\mathrm{lbs0}}\\), and guides the rigid transformation of vertices on the SMPL mesh according to the human pose. To account for the differences between the human mesh with clothing and the vanilla SMPL mesh, [17] calculate \\(\\omega_{\\mathrm{lbs}}\\) by keeping using the LBS weight prior \\(\\omega_{\\mathrm{lbs0}}\\) and adding an extra learnable network \\(\\Phi_{\\mathrm{lbs}}\\) for fine-tuning. This formulation (with \\(k\\) omitted) can be summarized as" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.819, + 0.483, + 0.836 + ], + "angle": 0, + "content": "\\[\n\\omega_ {\\mathrm {l b s}} \\left(\\boldsymbol {x} _ {c}\\right) = \\omega_ {\\mathrm {l b s 0}} \\left(\\mathrm {N N} _ {1} \\left(\\boldsymbol {x} _ {c}, \\boldsymbol {V}\\right)\\right) + \\Phi_ {\\mathrm {l b s}} \\left(\\boldsymbol {x} _ {c}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathrm{NN}_1\\) stands for top-1 nearest-neighbor search algorithm and \\(V\\) represents the canonical standard SMPL vertices. It will be, however, clarified by us, that this nearest neighbor-based assigning method will no longer be feasible" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.867, + 0.107 + ], + "angle": 0, + "content": "in cases with handheld objects and complex clothing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.107, + 0.905, + 0.257 + ], + "angle": 0, + "content": "In Fig. 3, we present an example of misclassification by the nearest-neighbor search algorithm in Eq. (5). Due to the lack of human topology constraints in the canonical space, incorrect classification of hand-held objects can occur, as shown in Fig. 3 (a). This part, following the naive nearest-neighbor search, would be assigned to the leg by mistake. Fig. 3 (b) shows the points belonging to the hand joint (the correct parent joint where the handheld object should grow). Yet being solely guided by LBS weights results in the voids by misclassification." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.349 + ], + "angle": 0, + "content": "To mitigate this issue, we propose a more robust assignment method based on motion priors, which we call Motion Kernels. Specifically, in the observation space, the motion kernel of each point \\( \\pmb{x}_o \\) with respect to the \\( k \\)-th joint position \\( j_k \\) is defined based on the changes in their pairwise Euclidean distances through all input \\( N \\) frames, following" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.354, + 0.905, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M K} \\left(\\boldsymbol {x} _ {c}, \\boldsymbol {j} _ {k}\\right) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\left\\| \\boldsymbol {x} _ {o i} - \\boldsymbol {j} _ {k i} \\right\\| _ {2} - \\mu\\right) ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.4, + 0.905, + 0.441 + ], + "angle": 0, + "content": "\\[\n\\mu = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| \\boldsymbol {x} _ {o i} - \\boldsymbol {j} _ {k i} \\| _ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.906, + 0.58 + ], + "angle": 0, + "content": "Our motion kernel (MK) reflects the relative motion between each gaussian and joint. A smaller MK indicates that the pair of gaussian and joint is relatively stationary to each other, signifying a stronger association, while a larger MK suggests a higher degree of relative motion, indicating a weaker association. We further represent the assignment weight reflected by the MK as \\(\\omega_{\\mathrm{MK}}(\\boldsymbol{x}_c) = \\mathrm{Normalize}_k(\\mathrm{MK}^{-1}(\\boldsymbol{x}_c,\\boldsymbol{j}_k))\\), and the final assignment weight in Eq. (4) (with index \\(k\\) omitted) becomes" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.585, + 0.905, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\omega \\left(\\boldsymbol {x} _ {c}\\right) = \\lambda \\omega_ {\\mathrm {M K}} \\left(\\boldsymbol {x} _ {c}\\right) + (1 - \\lambda) \\omega_ {\\mathrm {l b s}} \\left(\\boldsymbol {x} _ {c}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.683 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter to balance the MK weight and LBS weight. Note that we do not completely abandon the LBS weight, because the MK cannot differentiate the association of points on either side of a joint (\"misdirect\" in Fig. 3), requiring the LBS weight to compensate." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.69, + 0.868, + 0.705 + ], + "angle": 0, + "content": "4.2. Gradient-based Parent Joint Localization" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.712, + 0.906, + 0.834 + ], + "angle": 0, + "content": "By combining Eq. (4) and Eq. (8), the parent joints \\( J_{s} \\subseteq J \\) that require growth can be located. Basically, we have \\( g_{J} = (g_{J_{1}}, g_{J_{2}}, \\ldots, g_{J_{K}}) \\) to represent the gradient accumulation of total \\( K \\) human joints. As mentioned in Sec. 4.1, the joints with larger gradients accumulated are more likely to require the growth of child joints. This is achieved by sorting \\( g_{J} \\) in descending order \\( \\pi \\) to get \\( g_{J}^{\\mathrm{sorted}} = (g_{J_{\\pi(1)}}, g_{J_{\\pi(2)}}, \\ldots, g_{J_{\\pi(K)}}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.833, + 0.905, + 0.863 + ], + "angle": 0, + "content": "We set a gradient threshold \\(\\epsilon_{J}\\) to identify the \\(J\\in J_{s}\\) that requires growth, following" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.868, + 0.905, + 0.892 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {J} _ {s} = \\left(J _ {\\pi (1)}, J _ {\\pi (2)}, \\dots , J _ {\\pi (N)}\\right) \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.89, + 0.835, + 0.905 + ], + "angle": 0, + "content": "\\[\ns. t. \\quad g _ {J _ {\\pi (N)}} \\geq \\epsilon_ {J} \\text {a n d} g _ {J _ {\\pi (N + 1)}} < \\epsilon_ {J},\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14262" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.09, + 0.885, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.456, + 0.908, + 0.542 + ], + "angle": 0, + "content": "Figure 2. The pipeline of ToMiE. \\(①\\) We initialize the gaussians in the canonical space with a standard SMPL vertices. \\(②\\) (Sec. 4.4) We apply Linear Blend Skinning (LBS) to the gaussian position and utilize a network for rotation and scale correction. During the warmup phase, Adaptive LBS only utilizes the original SMPL skeleton. After adaptive growth, it further includes the newly grown external skeleton. \\(③\\) Gaussian rasterization and gradients backpropagation. \\(④\\) (Sec. 4.1, Sec. 4.2) We employ a gradient-based parent joints localization method, and a motion kernel to optimize the gradient assignment process. \\(⑤\\) (Sec. 4.3) We maintain an extra joint book with MLPs, which generates explicit human pose, enabling the decoupling and explicit animating of hand-held objects and loose-fitting clothing." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.554, + 0.268, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.654, + 0.185, + 0.664 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.554, + 0.48, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.654, + 0.384, + 0.665 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.773 + ], + "angle": 0, + "content": "Figure 3. Principle of the Motion Kernel. Relying solely on LBS weights will bring the misclassification in the canonical space shown in (a) to the observation space in (b), resulting in voids. Our proposed motion kernel focuses on motion-dependent priors in the observation space, offering better robustness and being less sensitive to misclassifications in the canonical space. This aids for point assignment process in parent joints localization." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.793, + 0.341, + 0.808 + ], + "angle": 0, + "content": "where we can safely assume \\(N < K\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.903 + ], + "angle": 0, + "content": "For each joint in \\( J_{s} \\), we designate it as a parent joint requiring growth and proceed with the initialization of its child joint. In order to explicitly model each child joint and ensure the consistency with the SMPL paradigm for ease of animation, we maintain an extra joint book \\( B^{e} = (\\text{parent}, j^{e}, r^{e}) \\), that includes the indices of parent joints," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.556, + 0.907, + 0.663 + ], + "angle": 0, + "content": "the extra joint positions \\( j^{e} \\) in canonical space and the extra rotations \\( r^{e} \\) in its parent joint coordinate. We initialize the joint position to its parent joint's position and set the rotation to the identity rotation. The entire parent joint localization and child joint initialization process is guided by the gradients, effectively preventing unnecessary overgrowth and ensuring a dense distribution of the extra joints." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.677, + 0.744, + 0.694 + ], + "angle": 0, + "content": "4.3. Extra Joint Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.907, + 0.793 + ], + "angle": 0, + "content": "The joint positions and rotations in the extra joint book are set as estimizable, which will be stored and later decoded by two shallow MLPs. According to the SMPL paradigm, the canonical joint position is a time-invariant quantity, therefore the joint position optimization network \\(\\Phi_p\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.797, + 0.905, + 0.815 + ], + "angle": 0, + "content": "\\[\nd \\boldsymbol {j} ^ {e} (i) = \\Phi_ {p} (P. E. (i)), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.819, + 0.907, + 0.879 + ], + "angle": 0, + "content": "where \\( P.E. \\) is a positional encoding function as is in [38] and \\( i \\) is the joint index in \\( B^{e} \\). Rotations are also dependent on the timestamp \\( t \\) of each frame, thus the rotation optimization network \\( \\Phi_{r} \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {r} ^ {e} (i, t) = \\Phi_ {r} (P. E. (i), P. E. (t)). \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14263" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "Now we have the positions of extra joints \\( j^{e} \\) (optimized by the offset \\( d\\boldsymbol{j}^{e} \\)) and the rotations \\( \\boldsymbol{r}^{e} \\). Each extra joint position is defined in the canonical space, representing the intrinsic properties of the extended skeleton, while the extra joint rotation in parent joint coordinate varies in each frame." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.169, + 0.482, + 0.306 + ], + "angle": 0, + "content": "Although both the extra joint positions and rotations are stored in the MLPs, their inputs and outputs are explicit features with real physical meaning, which allows for both implicit and explicit editing. For example, we can interpolate over timestamp \\( t \\) based on \\( \\Phi_r \\) or directly use explicit inputs to take the place of \\( \\Phi_r \\) during animation. The MLPs here function as decoders, deriving joint-related values from indices and timestamps, thus effectively circumventing the need for explicit storage of joint values employed in SMPL." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.317, + 0.436, + 0.334 + ], + "angle": 0, + "content": "4.4. Inference Process and Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.483, + 0.371 + ], + "angle": 0, + "content": "In this subsection, we explain how our adaptive growth is integrated with the training process of human Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.483, + 0.477 + ], + "angle": 0, + "content": "First, we initialize the canonical gaussians with standard SMPL vertices. At the beginning of training, we set up a number of warm-up iterations during which no joint growth occurs, and the gaussian fitting is performed following the traditional human gaussian methods. This prevents underfitting due to insufficient training, which could further affect the joint localization in Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.601 + ], + "angle": 0, + "content": "During the warm-up iterations, canonical human gaussians will be first warped to the observation space according to the LBS weight in Eq. (5). To compensate for the inability of the LBS model to represent gaussians rotation and scale, a deformable network \\(\\Phi_d\\) is employed to correct the rotation and scale during the LBS process. To distinguish it from the joint rotation \\(\\pmb{r}\\) of the SMPL human pose, we denote the rotation of the gaussian with subscript \\(g\\), and" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.604, + 0.482, + 0.623 + ], + "angle": 0, + "content": "\\[\nd \\boldsymbol {r} _ {g}, d \\boldsymbol {s} _ {g} = \\Phi_ {d} \\left(\\boldsymbol {x} _ {g}, \\boldsymbol {r} ^ {0}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.703 + ], + "angle": 0, + "content": "The gaussians rotation \\( \\boldsymbol{r}_g \\) and scale \\( s_g \\) are modified with offsets \\( d\\boldsymbol{r}_g \\) and \\( ds_g \\) to get the final gaussians in the observation space. In the observation space, we obtain the rendered images through the rasterisation of gaussians and compute the image loss to supervise canonical gaussians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.855 + ], + "angle": 0, + "content": "Once the warm-up iterations completes, we begin the adaptive joint growth. With the MK calculated during the warm-up phase, we can locate the parent joints \\( J_{s} \\) that require growth. Then we add grown joints to the extra joint book \\( B^{e} \\), optimizing its positions \\( j^{e} \\) and rotations \\( r^{e} \\) during the subsequent training. Notably, \\( \\Phi_{\\mathrm{lbs}} \\) needs to extend its output dimensions to include the blending weights for extra joints, following \\( K = K^0 + K^e \\). Since the extra joints do not have prior LBS weights, their blending weights are entirely learned through \\( \\Phi_{\\mathrm{lbs}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Both the warm-up stage and the post-growth learning stage adopt adaptive density control to manage the updates of the gaussians. We also dynamically adjust the threshold" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "of gradient for densification in [25] based on the number of gaussians, in order to balance memory consumption. Please check our supplemental material for details of this design." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.161, + 0.646, + 0.178 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.189, + 0.61, + 0.205 + ], + "angle": 0, + "content": "5.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.215, + 0.905, + 0.38 + ], + "angle": 0, + "content": "Our method focuses on hand-held objects and loose-fitting clothing, so we select datasets for experiments accordingly. We notice that the DNA-Rendering dataset [7], by capturing complex scenes of the human body, meets our requirements. Specifically, we selected 8 cases that align with our hypothesis, namely 0041_10, 0090_06, 0176_07, 0800_07 (hand-held objects) and 0007_04, 0014_06, 0051_09, 0811_06 (loose-fitting clothing). For each case, we use 24 surrounding views for training and 6 novel surrounding views for testing. All views are synchronized and contained 100 frames each." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.384, + 0.906, + 0.52 + ], + "angle": 0, + "content": "In addition to tackle complicated scenarios, it is essential to ensure the model's performance in typical scenarios involving tight-fitting clothes. Therefore, we additionally test our method on the ZJU-MoCap [46] dataset. Although the tight-clothing cases are too simple to require joint extension, our overall framework still achieves optimal results. Since this part of the experiment is not directly related to adaptive growth, we refer the readers to the supplemental material for further visualizations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.54, + 0.719, + 0.555 + ], + "angle": 0, + "content": "5.2. Baselines and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.566, + 0.906, + 0.747 + ], + "angle": 0, + "content": "We select the most cutting-edge and representative works from each focus area for a fair comparison. Since 3DGS is currently the leading representation for novel view synthesis, we compare 3DGS-based methods, including 3DGS-Avatar [47], GART [29], and GauHuman [17]. Among them, GART is expected to offer extra animatability for its modeling of implicit global auxiliary bones. Additionally, there is another category of human modeling without incorporating SMPL-like pose priors. Although these methods don't guarantee an animatable human avatar, they can achieve high-quality rendering, among which, we compare the rendering quality of Im4D [33] with our method." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.75, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We conduct a comprehensive comparison of our ToMiE against these methods. We report three key metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [51], and learned perceptual image patch similarity (LPIPS) [57]. Per-scene results can be found in the supplemental material. In addition to comparing the rendering results, we also demonstrate the animatability of the complicated human parts enabled by our method. We strongly recommend readers to watch the supplemental video for a more intuitive understanding of the animating." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14264" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.092, + 0.885, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.58, + 0.908, + 0.622 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on the DNA-Rendering dataset [7] with animatable baselines. We show two cases of hand-held objects (0800_07) and loose-fitting clothing (0811_06) (from top to bottom). Im4d* [33] achieves high-quality rendering but cannot be animated, and we compare it in the supplemental materials. Please check the supplemental video for better visualization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.353, + 0.652 + ], + "angle": 0, + "content": "5.3. Novel View Synthesis Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Fig. 4, Tab. 1 and Tab. 2 present the results of our method compared to other baselines. In the tables, we showcase two evaluation protocols. The first evaluates the entire image, reflecting the overall rendering quality. The second uses a binary mask to specifically compare the complicated regions, demonstrating how our method outperforms others in these challenging cases. The mask is shown in Fig. 4, whose details can be found in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.254, + 0.803 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.884 + ], + "angle": 0, + "content": "A. Adaptive Growth Ablation. We remove the adaptive growth to ablate its impact on the rendering results. Tab. 3 \"w/o Adaptive Growth\" shows a decline in rendering quality, while the hand-held objects and loose-fitting clothing also become not animatable." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "B. Non-rigid Design Ablation. In Sec. 4.4, we apply a" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.636, + 0.906, + 0.713 + ], + "angle": 0, + "content": "
Dataset Method | MetricAnimatablePSNR(full) ↑DNA-RenderingPSNR(masked) ↑
BGSSIM ↑LPIPS ↓
3DGS-Avatar✓✘28.930.9530.045917.69
GART✓✓29.250.9580.048017.12
Im4D✘✘26.280.9650.030814.90
GauHuman✓✘30.220.9620.040518.26
ToMiE(ours)✓✓31.280.9660.037419.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.717, + 0.907, + 0.835 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison between our method and other methods on complicated DNA-Rendering dataset [7]. \\(\\mathbb{B}\\) and \\(\\mathbb{G}\\) stand for human body and complex clothing (including handheld objects). We color each result as best, second best and third best. ToMiE achieves the best performance in PSNR and SSIM, while ranking second only to rendering-based Im4D [33] in LPIPS, mainly because rendering-based methods do not consider human structural constraints, resulting in higher visual fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.903 + ], + "angle": 0, + "content": "non-rigid deformation network \\(\\Phi_d\\) to correct gaussian rotation and scale. As shown in Tab. 3 \"w/o Non-rigid Design\", this significantly improves the final rendering quality." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14265" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.465, + 0.178 + ], + "angle": 0, + "content": "
Dataset Method | MetricAnimatableZJU-Mocap
BGPSNR(full) ↑SSIM ↑LPIPS ↓
3DGS-Avatar✓×30.610.9700.0296
GART✓✓30.910.9620.0318
Im4D××28.990.9730.0620
GauHuman✓×30.820.9620.0326
ToMiE(ours)✓✓31.100.9630.0312
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.182, + 0.483, + 0.298 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison between our method and other methods on ZJU-Mocap dataset [46] with tight-fitting clothing. \\(\\mathbb{B}\\) and \\(\\mathbb{G}\\) stand for human body and clothing (including hand-held objects). We color each result as best, second best and third best. This comparison demonstrates that ToMiE can achieve comparable (SSIM and LPIPS) or even superior (PSNR) performance to other approaches in tight-fitting scenarios where joint growth is not required." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.313, + 0.482, + 0.369 + ], + "angle": 0, + "content": "
PSNR(full) ↑SSIM ↑LPIPS ↓PSNR(masked) ↑
w/o Adaptive Growth30.990.9640.039019.21
w/o Non-rigid Design30.780.9640.036319.38
Full31.280.9660.037419.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.372, + 0.483, + 0.414 + ], + "angle": 0, + "content": "Table 3. Ablation studies on DNA-Rendering dataset [7]. We independently ablate the adaptive growth strategy and non-rigid design to validate their impact on the overall performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.43, + 0.483, + 0.461 + ], + "angle": 0, + "content": "5.5. Animating Hand-held Objects and Loose-fitting clothing" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.467, + 0.483, + 0.663 + ], + "angle": 0, + "content": "We demonstrate the uniqueness of our method, specifically its ability to explicitly animate hand-held objects and loose-fitting clothing. Our animating approach can be implemented in two ways. On the one hand, we can utilize the transformation already recorded in the extra joint book to replay clothing motions. The visualization of this part is shown in Fig. 5. On the other hand, ToMiE also supports bypassing the decoding process of the extra joint book by directly inputting the transformation explicitly. This allows us to customize the motion trajectory of external joints. Due to the limitations of the image's expressive capabilities, we strongly recommend the readers to watch the supplemental video to check the animated results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.483, + 0.786 + ], + "angle": 0, + "content": "In Fig. 5, we edit the extra joints while keeping the body poses under the SMPL paradigm stationary. Since the implicit auxiliary bones of GART [29] is controlled by the traditional SMPL poses, only the identical appearance can be output when SMPL poses are stationary. In contrast, our method explicitly models hand-held objects and loose-fitting clothing, fully decoupling them from the traditional SMPL poses, enabling free animating." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.347, + 0.815 + ], + "angle": 0, + "content": "6. Limitations and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Limitations. Although our method enhances the modeling for rigid and non-rigid clothing, it cannot address scenarios involving drastic changes in the topology (e.g., taking off clothes or opening a book). This is because topological changes disrupt the one-to-one correspondence be" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.465, + 0.907, + 0.508 + ], + "angle": 0, + "content": "Figure 5. Animating Results of ToMiE. Our explicit modeling fully decouples hand-held objects and loose-fitting clothing from the human body, enabling part-specific animating." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.613 + ], + "angle": 0, + "content": "tween frames, making the human modeling centered on the canonical space become downgraded. We notice that [42] addresses topology issues by introducing high-dimensional mappings, which could be adapted to build our non-rigid deformation. However, this is not the main scope of this paper and can be explored as future work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Conclusion. In this paper, we introduce ToMiE, an adaptive growth method designed to extend traditional SMPL skeleton for better modeling of hand-held objects and loose-fitting clothing. In the first stage, we assign the gradient of gaussian points to different joints by combining LBS weights with the motion kernel based on motion priors. This allows us to accurately locate the parent joints that need to grow, avoiding redundant growth. In the second stage, we design an extra joint book to achieve explicit joint modeling and optimize the transformation of the newly grown joints in an end-to-end manner. With the improved designs mentioned above, our ToMiE stands out among numerous state-of-the-art methods, achieving the best rendering quality and animatability of hand-held objects and loose-fitting clothing. We hope the adaptive growing method will spark a renewed discussion on the current capabilities of digital human modeling. What's more, it is expected to offer some insights for subsequent works related to topology and skeleton generation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14266" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.193 + ], + "angle": 0, + "content": "This research was supported in part by the Tateisi Science and Technology Foundation, JSPS KAKENHI Grant Numbers 24KK0209 and 22H00529. This research was also supported by Shanghai Artificial Intelligence Laboratory." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.206, + 0.188, + 0.222 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.232, + 0.484, + 0.273 + ], + "angle": 0, + "content": "[1] I. Baran and J. Popović. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG), 26(3):72-es, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.484, + 0.316 + ], + "angle": 0, + "content": "[2] H. Bertiche, M. Madadi, and S. Escalera. Neural cloth simulation. ACM Transactions on Graphics (TOG), 41(6):1-14, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[3] J. Chen, Y. Zhang, D. Kang, X. Zhe, L. Bao, X. Jia, and H. Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[4] X. Chen, T. Jiang, J. Song, M. Rietmann, A. Geiger, M. J. Black, and O. Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11796-11809, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[5] X. Chen, Y. Zheng, M. J. Black, O. Hilliges, and A. Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11594-11604, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.484, + 0.544 + ], + "angle": 0, + "content": "[6] Y. Chen, Y. Zhan, Z. Zhong, W. Wang, X. Sun, Y. Qiao, and Y. Zheng. Within the Dynamic Context: Inertia-aware 3D Human Modeling with Pose Sequence. arXiv preprint arXiv:2403.19160, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.484, + 0.629 + ], + "angle": 0, + "content": "[7] W. Cheng, R. Chen, S. Fan, W. Yin, K. Chen, Z. Cai, J. Wang, Y. Gao, Z. Yu, Z. Lin, et al. DNA-Rendering: A Diverse Neural Actor Repository for High-fidelity Human-centric Rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[8] E. Corona, A. Pumarola, G. Alenya, G. Pons-Moll, and F. Moreno-Noguer. SMPlicit: Topology-aware generative model for clothed people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11875-11885, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.702, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[9] J. Dong, Q. Fang, W. Jiang, Y. Yang, H. Bao, and X. Zhou. Fast and robust multi-person 3d pose estimation and tracking from multiple views. In T-PAMI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[10] G. Gafni, J. Thies, M. Zollhofer, and M. Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[11] Q. Gao, Y. Wang, L. Liu, L. Liu, C. Theobalt, and B. Chen. Neural novel actor: Learning a generalized animatable neural representation for human actors. IEEE Transactions on Visualization and Computer Graphics, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] C. Geng, S. Peng, Z. Xu, H. Bao, and X. Zhou. Learning neural volumetric representations of dynamic humans" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.232, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "in minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[13] S. Goel, G. Pavlakos, J. Rajasegaran, A. Kanazawa, and J. Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[14] C. Guo, T. Jiang, M. Kaufmann, C. Zheng, J. Valentin, J. Song, and O. Hilliges. ReLoo: Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild. arXiv preprint arXiv:2409.15269, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.908, + 0.319 + ], + "angle": 0, + "content": "[15] L. Hu. animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.907, + 0.401 + ], + "angle": 0, + "content": "[16] L. Hu, H. Zhang, Y. Zhang, B. Zhou, B. Liu, S. Zhang, and L. Nie. Gaussianavatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.404, + 0.908, + 0.461 + ], + "angle": 0, + "content": "[17] S. Hu, T. Hu, and Z. Liu. Gauhuman: Articulated Gaussian Splatting from Monocular Human Videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.908, + 0.517 + ], + "angle": 0, + "content": "[18] T. Hu, F. Hong, and Z. Liu. Surmo: Surface-based 4d motion modeling for dynamic human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6550-6560, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[19] M. Isik, M. Rünz, M. Georgopoulos, T. Khakhulin, J. Starck, L. Agapito, and M. Nießner. Humanrf: High-fidelity Neural Radiance Fields for Humans in Motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[20] A. Jacobson, I. Baran, L. Kavan, J. Popovic, and O. Sorkine. Fast automatic skinning transformations. ACM Transactions on Graphics (ToG), 31(4):1-10, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.906, + 0.674 + ], + "angle": 0, + "content": "[21] R. Jena, G. S. Iyer, S. Choudhary, B. Smith, P. Chaudhari, and J. Gee. Splatarmor: Articulated Gaussian Splatting for Animatable Humans from Monocular RGB Videos. arXiv preprint arXiv:2311.10812, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[22] Y. Jiang, Z. Shen, Y. Hong, C. Guo, Y. Wu, Y. Zhang, J. Yu, and L. Xu. Robust Dual Gaussian Splatting for Immersive Human-centric Volumetric Videos. ACM Transactions on Graphics (TOG), 43(6):1-15, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.908, + 0.801 + ], + "angle": 0, + "content": "[23] Y. Jiang, Z. Shen, P. Wang, Z. Su, Y. Hong, Y. Zhang, J. Yu, and L. Xu. Hifi4g: High-fidelity Human Performance Rendering via Compact Gaussian Splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19734-19745, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[24] H. Jung, N. Brasch, J. Song, E. Perez-Pellitero, Y. Zhou, Z. Li, N. Navab, and B. Busam. Deformable 3D Gaussian Splitting for Animatable Human Avatars. arXiv preprint arXiv:2312.15059, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] B. Kerbl, G. Kopanas, T. Leimkuhler, and G. Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14267" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] M. Kocabas, J.-H. R. Chang, J. Gabriel, O. Tuzel, and A. Ranjan. Hugs: Human Gaussian Splats. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 505-515, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[27] Y. Kwon, D. Kim, D. Ceylan, and H. Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[28] B. H. Le and Z. Deng. Two-layer sparse compression of dense-weight blend skinning. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[29] J. Lei, Y. Wang, G. Pavlakos, L. Liu, and K. Daniilidis. Gart: Gaussian Articulated Template Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19876-19887, 2024. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.306, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[30] M. Li, J. Tao, Z. Yang, and Y. Yang. Human101: Training 100+ fps Human Gaussians in 100s from 1 View. arXiv preprint arXiv:2312.15258, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[31] M. Li, S. Yao, Z. Xie, K. Chen, and Y.-G. Jiang. Gaussian-body: Clothed human reconstruction via 3d gaussian splatt-ing. arXiv preprint arXiv:2401.09720, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[32] Z. Li, Z. Zheng, L. Wang, and Y. Liu. Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[33] H. Lin, S. Peng, Z. Xu, T. Xie, X. He, H. Bao, and X. Zhou. High-Fidelity and Real-Time Novel View Synthesis for Dynamic Scenes. In SIGGRAPH Asia Conference Proceedings, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[34] X. Liu, C. Wu, X. Liu, J. Liu, J. Wu, C. Zhao, H. Feng, E. Ding, and J. Wang. Gea: Reconstructing expressive 3d gaussian avatar from monocular video. arXiv preprint arXiv:2402.16607, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[35] Y. Liu, X. Huang, M. Qin, Q. Lin, and H. Wang. Animatable 3d gaussian: Fast and high-quality reconstruction of multiple human avatars. arXiv preprint arXiv:2311.16482, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[36] M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black. SMPL: A Skinned Multi-Person Linear Model. ACM Transactions on Graphics, 34(6), 2015. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.703 + ], + "angle": 0, + "content": "[37] Y. Men, Y. Yao, M. Cui, and L. Bo. Mimo: Controllable character video synthesis with spatial decomposed modeling. arXiv preprint arXiv:2409.16160, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.482, + 0.76 + ], + "angle": 0, + "content": "[38] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[39] A. Moreau, J. Song, H. Dhamo, R. Shaw, Y. Zhou, and E. Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[40] M. Niu, M. Cao, Y. Zhan, Q. Zhu, M. Ma, J. Zhao, Y. Zeng, Z. Zhong, X. Sun, and Y. Zheng. Anicrafter: Customizing realistic human-centric animation via avatar-background conditioning in video diffusion models. arXiv preprint arXiv:2505.20255, 2025." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.907, + 0.161 + ], + "angle": 0, + "content": "[41] H. Pang, H. Zhu, A. Kortylewski, C. Theobalt, and M. Habermann. Ash: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1165-1175, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.163, + 0.906, + 0.231 + ], + "angle": 0, + "content": "[42] K. Park, U. Sinha, P. Hedman, J. T. Barron, S. Bouaziz, D. B. Goldman, R. Martin-Brualla, and S. M. Seitz. Hypernerf: A Higher-dimensional Representation for Topologically Varying Neural Radiance Fields. arXiv preprint arXiv:2106.13228, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.233, + 0.906, + 0.3 + ], + "angle": 0, + "content": "[43] D. Paschalidou, A. Katharopoulos, A. Geiger, and S. Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3204-3215, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.302, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[44] G. Pavlakos, V. Choutas, N. Ghorbani, T. Bolkart, A. A. A. Osman, D. Tzionas, and M. J. Black. Expressive Body Capture: 3D Hands, Face, and Body from a Single Image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.372, + 0.906, + 0.412 + ], + "angle": 0, + "content": "[45] S. Peng, J. Dong, Q. Wang, S. Zhang, Q. Shuai, X. Zhou, and H. Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.414, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[46] S. Peng, Y. Zhang, Y. Xu, Q. Wang, Q. Shuai, H. Bao, and X. Zhou. Neural Body: Implicit Neural Representations with Structured Latent Codes for Novel View Synthesis of Dynamic Humans. In CVPR, 2021. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.47, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[47] Z. Qian, S. Wang, M. Mihajlovic, A. Geiger, and S. Tang. 3dgs-avatar: Animatable Avatars via Deformable 3d Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020–5030, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.539, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[48] I. Santesteban, M. A. Otaduy, and D. Casas. Snug: self-supervised neural dynamic garments. IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), volume 2, page 9, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.596, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[49] Q. Shuai, C. Geng, Q. Fang, S. Peng, W. Shen, X. Zhou, and H. Bao. Novel view synthesis of human interactions from sparse multi-view videos. In SIGGRAPH Conference Proceedings, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.651, + 0.906, + 0.704 + ], + "angle": 0, + "content": "[50] S. Wang, K. Schwarz, A. Geiger, and S. Tang. A随机变量的统计学，1987年，第2版，M. H. Lai, The asymptotic analysis of a random variable with a finite distribution, 1987. Springer, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.706, + 0.906, + 0.761 + ], + "angle": 0, + "content": "[51] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.763, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[52] C.-Y. Weng, B. Curless, P. P. Srinivasan, J. T. Barron, and I. Kemelmacher-Shlizerman. Humannerf: Free-viewpoint Rendering of Moving People from Monocular Video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.832, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[53] Z. Xu, S. Peng, H. Lin, G. He, J. Sun, Y. Shen, H. Bao, and X. Zhou. 4K4D: Real-time 4D View Synthesis at 4K Resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20029-20040, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14268" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[54] Z. Xu, Y. Zhou, E. Kalogerakis, C. Landreth, and K. Singh. Rignet: Neural rigging for articulated characters. arXiv preprint arXiv:2005.00559, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[55] C.-H. Yao, W.-C. Hung, Y. Li, M. Rubinstein, M.-H. Yang, and V. Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4853-4862, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[56] Y. Zhan, W. Xu, Q. Zhu, M. Niu, M. Ma, Y. Liu, Z. Zhong, X. Sun, and Y. Zheng. R3-avatar: Record and retrieve temporal codebook for reconstructing photorealistic human avatars. arXiv preprint arXiv:2503.12751, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[57] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.414 + ], + "angle": 0, + "content": "[58] S. Zheng, B. Zhou, R. Shao, B. Liu, S. Zhang, L. Nie, and Y. Liu. Gps-gaussian: Generalizable Pixel-wise 3D Gaussian Splatting for Real-time Human Novel View Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.417, + 0.482, + 0.472 + ], + "angle": 0, + "content": "[59] Y. Zheng, Q. Zhao, G. Yang, W. Yifan, D. Xiang, F. Dubost, D. Lagun, T. Beeler, F. Tombari, L. Guibas, et al. Physavatar: Learning the physics of dressed 3d avatars from visual observations. arXiv preprint arXiv:2404.04421, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.474, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[60] B. Zhou, S. Zheng, H. Tu, R. Shao, B. Liu, S. Zhang, L. Nie, and Y. Liu. GPS-Gaussian+: Generalizable Pixel-wise 3D Gaussian Splatting for Real-Time Human-Scene Rendering from Sparse Views. arXiv preprint arXiv:2411.11363, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[61] W. Zielonka, T. Bagautdinov, S. Saito, M. Zollhöfer, J. Thies, and J. Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14269" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/83fba3ab-8c7e-4045-994b-8f70e4fc8d14_origin.pdf b/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/83fba3ab-8c7e-4045-994b-8f70e4fc8d14_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e1d6951a1f22336de23d08d0bcd10eb76647168 --- /dev/null +++ b/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/83fba3ab-8c7e-4045-994b-8f70e4fc8d14_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f68f09072fc0f6f5371c67cc943adb3728cfbbf8b380d97e276e3eb25799fce +size 1944811 diff --git a/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/full.md b/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e6e30c82fa307fa2aa6c599f23a4478037cd5025 --- /dev/null +++ b/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/full.md @@ -0,0 +1,329 @@ +# Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars + +Yifan Zhan $^{1,2*}$ Qingtian Zhu $^{2}$ Muyao Niu $^{2}$ Mingze Ma $^{2}$ Jiancheng Zhao $^{2}$ Zhihang Zhong $^{1\dagger}$ Xiao Sun $^{1\dagger}$ Yu Qiao $^{1}$ Yinqiang Zheng $^{2}$ + +$^{1}$ Shanghai Artificial Intelligence Laboratory $^{2}$ The University of Tokyo + +![](images/7fcea9dc1d75506f622186a76e8e80f51a270bf5b1ddda0b24eb1872536cdb06.jpg) +Figure 1. We show two complex scenarios in 3D human modeling: with hand-held objects and loose-fitting clothing, which cannot be accurately represented by the standard SMPL model. Our ToMiE can realize adaptive growth to enhance the representation capability of SMPL without needing for time-consuming case-specific customization, achieving state-of-the-art results in both rendering and human (including complex scenarios) animation. + +![](images/fae2fc5ceec61ad8c8edb91c423520bcda27587f4d765ad551779d288cf1b36f.jpg) + +# Abstract + +In this paper, we highlight a critical yet often overlooked factor in most 3D human tasks, namely modeling complicated 3D human with with hand-held objects or loose-fitting clothing. It is known that the parameterized formulation of SMPL is able to fit human skin; while hand-held objects and loose-fitting clothing, are difficult to get modeled within the unified framework, since their movements are usually decoupled with the human body. To enhance the capability of SMPL skeleton in response to this situation, we propose a growth strategy that enables the joint tree of the skeleton to expand adaptively. Specifically, our method, called ToMiE, consists of parent joints localization and external joints optimization. For parent joints localization, we employ a gradient-based approach guided by both LBS blending weights and motion kernels. Once + +the external joints are obtained, we proceed to optimize their transformations in $SE(3)$ across different frames, enabling rendering and explicit animation. ToMiE manages to outperform other methods across various cases with hand-held objects and loose-fitting clothing, not only in rendering quality but also by offering free animation of grown joints, thereby enhancing the expressive ability of SMPL skeleton for a broader range of applications. The code is available at https://github.com/Yifever20002/ToMiE. + +# 1. Introduction + +3D human reconstruction endeavors to model high-fidelity digital avatars based on real-world characters for virtual rendering and animating, which has been of long-term research value in areas such as gaming, virtual reality (VR), and beyond. Traditional methods, such as SMPL [36], achieve human body parameterization by performing principal component analysis (PCA) on large sets + +of 3D scanned meshes, allowing for the fitting of a specified identity. Recent neural rendering techniques have enabled implicit digital human modeling guided by Linear Blend Skinning (LBS) and SMPL skeleton, realizing lifelike rendering and animating from video inputs. + +The neural-based 3D human rendering has been empowered by the cutting-edge technique, 3D Gaussian Splitting (3DGS) [25], for its real-time and high-quality novel view synthesis performance. By representing 3D gaussians under the canonical T-pose and utilizing the pre-extracted SMPL skeleton in the observation space, 3D human rendering results can be obtained from novel views in any frame. This stream of approaches proves high quality in rendering 3D humans that conform to the SMPL paradigm (e.g., avatars in tight-fitting clothing). However, we raise concerns regarding its ability to handle complicated human modeling involving hand-held objects or loose-fitting clothing. + +In Fig. 1, we show two cases to illustrate the limitations of current 3D human modeling. On the one hand, the movements of hand-held objects, e.g., feather duster, are highly decoupled from the human body and thus cannot be represented by the SMPL. On the other hand, characters shot in-the-wild are dressed in clothing with high complexity in dynamics, rather than the tight-fitting clothing configured under strict experimental conditions. These complicated scenarios break the existing methods' paradigm by assuming that surface of avatars should be bounded to the motion warping of SMPL skeleton in the same way as the human skin is. Rooted in the aforementioned scenarios, SMPL fails to accurately fit such 3D human models. + +To this end, we break through the limitations of modeling complicated 3D human gaussians by maintaining an extended joint tree from SMPL skeleton. Although existing SMPL model has the potential capability to manually customize additional skeleton information, time-consuming and case-by-case adjustments are necessary in this case. To overcome this issue, we extend the SMPL skeleton with additional joints for each individual case adaptively. The growth is performed in an explicit and adaptive manner, and enables the fitting of complicated 3D human avatars, offering high-quality rendering with hand-held objects and loose-fitting clothing (e.g., novel view synthesis results in Fig. 1). + +The main challenge of extending the SMPL skeleton is to determine where and how to grow additional joints. To avoid the unnecessary memory consumption and potential overfitting caused by an arbitrary growth, we first determine which of the joints are supposed to serve as parent joints by a localization strategy. We have empirically observed that parent joints requiring growth witness larger backpropagation gradients in their associated gaussians due to underfitting. However, determining the association of gaussians with different joints is non-trivial, as the SMPL's LBS + +blending weights may not work with the human with handheld objects or clothing. To more precisely define such association, we introduce the concept of Motion Kernels based on rigid body priors and combine them with LBS weights, resulting in more accurate gradient-based localization. After growth, we adaptively maintain an extended joint tree and update the extra joints by optimizing two MLP decoders for joints positions and rotations. The proposed method, termed as ToMiE, allows for explicit rendering and animation of complicated human avatars represented by extended joints. + +By experiments on complicated cases of the DNA-Rendering dataset [7], ToMiE exhibits state-of-the-art rendering quality while maintaining the animatability that is significant for downstream productions. To summarize, our contributions are three-fold as follows: + +1) ToMiE, a method for creating an enhanced SMPL joint tree via an adaptive growth strategy, which is able to decouple complicated parts from the human body, thereby achieving state-of-the-art results in both rendering and explicit animation on target cases; +2) a hybrid assignment strategy for gaussians utilizing LBS weights and Motion Kernels, combined with gradient-driven parent joints localization, to guide the growth of external joints; +3) a joints optimization approach fitting local rotations across different frames while sharing joints positions. + +# 2. Related Work + +# 2.1. SMPL-based Human Mesh Avatars + +Most of the recent success in digital human modeling can be attributed to the contributions of the SMPL [36] series, which parameterizes the human body as individual shape components and motion-related human poses through 3D mesh scanning and PCA. The pose blend shapes in SMPL describe human body deformations as a linear weighted blending of different joint poses, significantly improving the efficiency of editing and animating digital humans. Furthermore, it has been widely adopted for human body animation, thanks to the methods [9, 49] of estimating SMPL parameters from 2D inputs. Despite their wide range of applications, SMPL and its family still suffer from inherent limitations. Since the originally scanned 3D meshes are skin-tight, according to which the pose blend shapes are learned, the model is unable to handle significantly outlying meshes, such as human with hand-held objects and clothing like skirts. Although this could potentially be solved by auto-rigging [1, 20, 28, 43, 54, 55], it requires 2D/3D shapes as inputs, which is not accessible in human videos. + +# 2.2. Neural Representation for 3D Human + +Methods based on neural representations, such as NeRF [38] and 3DGS [25], have also been playing an important part in digital human reconstruction for their high-quality rendering capabilities. Early NeRF-based methods [3-6, 10-13, 27, 46, 52] aim to reconstruct human avatars by inputting monocular or multi-view synchronized videos. [50] enforce smooth priors based on neural Signed Distance Function (SDF) to obtain more accurate human geometry. Recent breakthroughs [16-18, 21, 24, 26, 29-32, 34, 35, 39-41, 47, 56, 58, 59, 61] rely on 3DGS, enabling faster and more accurate rendering. All these methods register the T-pose in a canonical space and use LBS weights to guide the rigid transformations. + +# 2.3. Rendering & Editing of Intricate 3D Human + +We revisit the methods attempting to implicitly improve modeling of complicated 3D human. Animatable NeRF [45] defines a per-frame latent code to capture appearance variations across each frame. Simulation methods [2, 48] physically construct simple clothing but struggle with complex one and objects. [14, 29] additionally register global latent bones to compensate for the limitations in clothing rendering. They fail to explicitly decouple the clothing from human body, making precise control infeasible. Another stream [6, 18] leverages a human poses sequence as contexts to resolve appearance ambiguities. But correspondingly, they require a sequence of human poses for animation, adding to the challenges of editing. Moreover, these methods struggle to fit the object-level pose independent of the human poses sequence, such as hand-held items. We also note that some works [15, 37] introduce diffusion-based generative methods to enhance the realism of garment rendering, but these methods are restricted by the traditional SMPL and overlook the editing of complex clothing. Although SMPLlicit [8] enables implicit interpolation of clothing types, it remains dependent on SMPL's LBS process to generate the observed mesh. This limitation prevents localized explicit animation of loose-fitting clothing and restricts its application to external objects. Rendering-based human reconstruction method [19, 22, 23, 33, 53, 58, 60] can not achieve animation. Our approach enables explicit decoupling of hand-held objects and clothing with the human body by extending the SMPL joint tree, allowing for high-quality rendering and explicit animation in complicated scenarios. + +# 3. Preliminaries + +# 3.1. SMPL(-X) Revisited + +Pre-trained on scanned meshes, the SMPL(-X) family [36, 44] employs a parameterized model to fit human bodies of different shapes and under different poses. The + +human mesh in each frame evolves from a canonical human mesh and is controlled by shape and pose parameters. Specifically, a 3D point $\pmb{x}_c$ on the canonical mesh will be warped to obtain point in the observation space as + +$$ +\boldsymbol {x} _ {o} = \sum_ {k = 1} ^ {K} \omega_ {k} \left(\boldsymbol {x} _ {c}\right) \left(R _ {k} \left(\boldsymbol {r} ^ {0}\right) \boldsymbol {x} _ {c} + t _ {k} \left(\boldsymbol {j} ^ {0}, \beta\right)\right), \tag {1} +$$ + +where $K$ is the total number of joints, $R_{k}$ is per-joint global rotation controlled by local joint rotations $\pmb{r}^{0}$ , and $t_k$ is per-joint translation controlled by joint positions $\pmb{j}^0$ and human shape $\beta$ . Notice that linear blending weight $\omega_{k}$ is a function of $\pmb{x}_c$ , which is regressed from large human assets. + +The main issue with this LBS-based prior model is that it can only model tight-fitting avatars conforming to the SMPL(-X) paradigm, making it unsuitable for modeling complex human clothing in more generic cases. What even more challenging is its inability to handle hand-held objects that are fully decoupled from the human pose. + +# 3.2. Human Gaussians Revisited + +Human gaussians achieve high-quality real-time human rendering by combining the SMPL prior with 3DGS as the representation. The SMPL model naturally obtains the T-posed (i.e., all human poses are identity transformations) mesh in the canonical space and then mesh vertices are used to initialize the canonical gaussian units. Each gaussian is defined as + +$$ +G (\boldsymbol {x}) = \frac {1}{(2 \pi) ^ {\frac {3}{2}} | \boldsymbol {\Sigma} | ^ {\frac {1}{2}}} e ^ {- \frac {1}{2} (\boldsymbol {x} - \boldsymbol {\mu}) ^ {T} \boldsymbol {\Sigma} ^ {- 1} (\boldsymbol {x} - \boldsymbol {\mu})}, \tag {2} +$$ + +where $\mu$ is the 3D gaussian center, and $\Sigma$ is the 3D covariance matrix, which will be further decomposed into learnable rotation $\pmb{R}$ and scale $\pmb{S}$ . Now we have $\boldsymbol{\Sigma} = \boldsymbol{RSS}^{\top} \boldsymbol{R}^{\top}$ , which is performed by optimizing a quaternion $\boldsymbol{r}_g$ for rotation and a 3D vector $\boldsymbol{s}_g$ for scaling. Each gaussian is further assigned with color $c$ and opacity $\alpha$ . + +Once the 3D gaussians in the canonical space are obtained, each position $\mu$ will be warped to the observation space according to Eq. (1). Next, the 3D gaussians of each frame are projected into 2D gaussians, followed by tile-based rasterization. Color of each pixel can be calculated by blending $N$ ordered gaussians following + +$$ +C = \sum_ {i \in N} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}). \tag {3} +$$ + +The 3D gaussians can be optimized and updated by adaptive density control, which primarily includes cloning, splitting, and pruning. Cloning and splitting are guided by gradients to control the number of gaussians, while pruning removes empty gaussians based on current opacity $\alpha$ . The supervision upon human gaussians is derived from multiview or monocular videos, enabling high-quality rendering and avatar animation. + +# 4. Methods + +Fig. 2 illustrates ToMiE's adaptive joint growth and gaussian training strategy. Our goal is to extend the SMPL skeleton to handle complicated human scenarios. However, the abuse of growth can lead to unnecessary computational and memory overheads. For an efficient adaptive growth, we first propose a localization strategy of parent joints to ensure that only necessary joints are grown. Furthermore, we explicitly define the hand-held joints in the $SE(3)$ space and optimize them end-to-end through an MLP, ensuring alignment with the original skeleton of SMPL. The extended skeleton can thus support rendering and explicit animation. To address the limitation of LBS in guiding gaussian attributes apart from the positions, we further fine-tune the rotation and scale during training with a deformation field to achieve better non-rigid warping. Next, we will elaborate on these modules and the training strategy. + +# 4.1. Motion Kernels-guided Joint Gradient Accumulation + +A quantitative metric needs to be identified to determine whether a joint requires growth. We notice that, due to the poor fitting ability of existing SMPL-based human gaussian, it will leave larger gradients in complex human regions (e.g., the hand-held object in Fig. 2 ③). In other words, identifying joints with larger accumulated gradients can help to indicate which parent joints are more likely to require growth. Bounded with each gaussian, the accumulated gradients will be first assigned to their corresponding joints. Let $g = \|(g_x, g_y, g_z)\|_2$ represent the L2 norm of the gradient at each gaussian position. The gradients accumulation $g_J$ for the $k$ -th joint can be computed following + +$$ +g _ {J _ {k}} = \frac {\sum_ {i \in N} \omega_ {k} \left(\boldsymbol {x} _ {c}\right) g}{\sum_ {i \in N} \omega_ {k} \left(\boldsymbol {x} _ {c}\right)}. \tag {4} +$$ + +We use $\pmb{x}_c$ to stand for gaussian position in canonical space and $\pmb{x}_o$ in the observation space. + +It is worth emphasizing that $\omega_{k}$ here is a weight term determining the assignment of a gaussian to the $k$ -th joint. Under the paradigm of SMPL representation, this weight corresponds to the LBS weight $\omega_{\mathrm{lbs0}}$ , and guides the rigid transformation of vertices on the SMPL mesh according to the human pose. To account for the differences between the human mesh with clothing and the vanilla SMPL mesh, [17] calculate $\omega_{\mathrm{lbs}}$ by keeping using the LBS weight prior $\omega_{\mathrm{lbs0}}$ and adding an extra learnable network $\Phi_{\mathrm{lbs}}$ for fine-tuning. This formulation (with $k$ omitted) can be summarized as + +$$ +\omega_ {\mathrm {l b s}} \left(\boldsymbol {x} _ {c}\right) = \omega_ {\mathrm {l b s 0}} \left(\mathrm {N N} _ {1} \left(\boldsymbol {x} _ {c}, \boldsymbol {V}\right)\right) + \Phi_ {\mathrm {l b s}} \left(\boldsymbol {x} _ {c}\right), \tag {5} +$$ + +where $\mathrm{NN}_1$ stands for top-1 nearest-neighbor search algorithm and $V$ represents the canonical standard SMPL vertices. It will be, however, clarified by us, that this nearest neighbor-based assigning method will no longer be feasible + +in cases with handheld objects and complex clothing. + +In Fig. 3, we present an example of misclassification by the nearest-neighbor search algorithm in Eq. (5). Due to the lack of human topology constraints in the canonical space, incorrect classification of hand-held objects can occur, as shown in Fig. 3 (a). This part, following the naive nearest-neighbor search, would be assigned to the leg by mistake. Fig. 3 (b) shows the points belonging to the hand joint (the correct parent joint where the handheld object should grow). Yet being solely guided by LBS weights results in the voids by misclassification. + +To mitigate this issue, we propose a more robust assignment method based on motion priors, which we call Motion Kernels. Specifically, in the observation space, the motion kernel of each point $\pmb{x}_o$ with respect to the $k$ -th joint position $j_k$ is defined based on the changes in their pairwise Euclidean distances through all input $N$ frames, following + +$$ +\operatorname {M K} \left(\boldsymbol {x} _ {c}, \boldsymbol {j} _ {k}\right) = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\left\| \boldsymbol {x} _ {o i} - \boldsymbol {j} _ {k i} \right\| _ {2} - \mu\right) ^ {2}, \tag {6} +$$ + +$$ +\mu = \frac {1}{N} \sum_ {i = 1} ^ {N} \| \boldsymbol {x} _ {o i} - \boldsymbol {j} _ {k i} \| _ {2}. \tag {7} +$$ + +Our motion kernel (MK) reflects the relative motion between each gaussian and joint. A smaller MK indicates that the pair of gaussian and joint is relatively stationary to each other, signifying a stronger association, while a larger MK suggests a higher degree of relative motion, indicating a weaker association. We further represent the assignment weight reflected by the MK as $\omega_{\mathrm{MK}}(\boldsymbol{x}_c) = \mathrm{Normalize}_k(\mathrm{MK}^{-1}(\boldsymbol{x}_c,\boldsymbol{j}_k))$ , and the final assignment weight in Eq. (4) (with index $k$ omitted) becomes + +$$ +\omega \left(\boldsymbol {x} _ {c}\right) = \lambda \omega_ {\mathrm {M K}} \left(\boldsymbol {x} _ {c}\right) + (1 - \lambda) \omega_ {\mathrm {l b s}} \left(\boldsymbol {x} _ {c}\right), \tag {8} +$$ + +where $\lambda$ is a hyperparameter to balance the MK weight and LBS weight. Note that we do not completely abandon the LBS weight, because the MK cannot differentiate the association of points on either side of a joint ("misdirect" in Fig. 3), requiring the LBS weight to compensate. + +# 4.2. Gradient-based Parent Joint Localization + +By combining Eq. (4) and Eq. (8), the parent joints $J_{s} \subseteq J$ that require growth can be located. Basically, we have $g_{J} = (g_{J_{1}}, g_{J_{2}}, \ldots, g_{J_{K}})$ to represent the gradient accumulation of total $K$ human joints. As mentioned in Sec. 4.1, the joints with larger gradients accumulated are more likely to require the growth of child joints. This is achieved by sorting $g_{J}$ in descending order $\pi$ to get $g_{J}^{\mathrm{sorted}} = (g_{J_{\pi(1)}}, g_{J_{\pi(2)}}, \ldots, g_{J_{\pi(K)}})$ . + +We set a gradient threshold $\epsilon_{J}$ to identify the $J\in J_{s}$ that requires growth, following + +$$ +\boldsymbol {J} _ {s} = \left(J _ {\pi (1)}, J _ {\pi (2)}, \dots , J _ {\pi (N)}\right) \tag {9} +$$ + +$$ +s. t. \quad g _ {J _ {\pi (N)}} \geq \epsilon_ {J} \text {a n d} g _ {J _ {\pi (N + 1)}} < \epsilon_ {J}, +$$ + +![](images/5e78f3ac3a3766ff8f0dce7ee65b712815d4950a870f54df892dd7313de2956c.jpg) +Figure 2. The pipeline of ToMiE. $①$ We initialize the gaussians in the canonical space with a standard SMPL vertices. $②$ (Sec. 4.4) We apply Linear Blend Skinning (LBS) to the gaussian position and utilize a network for rotation and scale correction. During the warmup phase, Adaptive LBS only utilizes the original SMPL skeleton. After adaptive growth, it further includes the newly grown external skeleton. $③$ Gaussian rasterization and gradients backpropagation. $④$ (Sec. 4.1, Sec. 4.2) We employ a gradient-based parent joints localization method, and a motion kernel to optimize the gradient assignment process. $⑤$ (Sec. 4.3) We maintain an extra joint book with MLPs, which generates explicit human pose, enabling the decoupling and explicit animating of hand-held objects and loose-fitting clothing. + +![](images/eca421488364b38a297a7f32e72bb5af3c3725a9d0e1c3f56b5fc7b3d06d8f5c.jpg) +(a) + +![](images/53c973ec0b4b56750e6faf1ae362145bddc3502ac77ef6866a7a1970a903ace8.jpg) +(b) +Figure 3. Principle of the Motion Kernel. Relying solely on LBS weights will bring the misclassification in the canonical space shown in (a) to the observation space in (b), resulting in voids. Our proposed motion kernel focuses on motion-dependent priors in the observation space, offering better robustness and being less sensitive to misclassifications in the canonical space. This aids for point assignment process in parent joints localization. + +where we can safely assume $N < K$ . + +For each joint in $J_{s}$ , we designate it as a parent joint requiring growth and proceed with the initialization of its child joint. In order to explicitly model each child joint and ensure the consistency with the SMPL paradigm for ease of animation, we maintain an extra joint book $B^{e} = (\text{parent}, j^{e}, r^{e})$ , that includes the indices of parent joints, + +the extra joint positions $j^{e}$ in canonical space and the extra rotations $r^{e}$ in its parent joint coordinate. We initialize the joint position to its parent joint's position and set the rotation to the identity rotation. The entire parent joint localization and child joint initialization process is guided by the gradients, effectively preventing unnecessary overgrowth and ensuring a dense distribution of the extra joints. + +# 4.3. Extra Joint Optimization + +The joint positions and rotations in the extra joint book are set as estimizable, which will be stored and later decoded by two shallow MLPs. According to the SMPL paradigm, the canonical joint position is a time-invariant quantity, therefore the joint position optimization network $\Phi_p$ is defined as + +$$ +d \boldsymbol {j} ^ {e} (i) = \Phi_ {p} (P. E. (i)), \tag {10} +$$ + +where $P.E.$ is a positional encoding function as is in [38] and $i$ is the joint index in $B^{e}$ . Rotations are also dependent on the timestamp $t$ of each frame, thus the rotation optimization network $\Phi_{r}$ is defined as + +$$ +\boldsymbol {r} ^ {e} (i, t) = \Phi_ {r} (P. E. (i), P. E. (t)). \tag {11} +$$ + +Now we have the positions of extra joints $j^{e}$ (optimized by the offset $d\boldsymbol{j}^{e}$ ) and the rotations $\boldsymbol{r}^{e}$ . Each extra joint position is defined in the canonical space, representing the intrinsic properties of the extended skeleton, while the extra joint rotation in parent joint coordinate varies in each frame. + +Although both the extra joint positions and rotations are stored in the MLPs, their inputs and outputs are explicit features with real physical meaning, which allows for both implicit and explicit editing. For example, we can interpolate over timestamp $t$ based on $\Phi_r$ or directly use explicit inputs to take the place of $\Phi_r$ during animation. The MLPs here function as decoders, deriving joint-related values from indices and timestamps, thus effectively circumventing the need for explicit storage of joint values employed in SMPL. + +# 4.4. Inference Process and Training Strategy + +In this subsection, we explain how our adaptive growth is integrated with the training process of human Gaussians. + +First, we initialize the canonical gaussians with standard SMPL vertices. At the beginning of training, we set up a number of warm-up iterations during which no joint growth occurs, and the gaussian fitting is performed following the traditional human gaussian methods. This prevents underfitting due to insufficient training, which could further affect the joint localization in Sec. 4.2. + +During the warm-up iterations, canonical human gaussians will be first warped to the observation space according to the LBS weight in Eq. (5). To compensate for the inability of the LBS model to represent gaussians rotation and scale, a deformable network $\Phi_d$ is employed to correct the rotation and scale during the LBS process. To distinguish it from the joint rotation $\pmb{r}$ of the SMPL human pose, we denote the rotation of the gaussian with subscript $g$ , and + +$$ +d \boldsymbol {r} _ {g}, d \boldsymbol {s} _ {g} = \Phi_ {d} \left(\boldsymbol {x} _ {g}, \boldsymbol {r} ^ {0}\right). \tag {12} +$$ + +The gaussians rotation $\boldsymbol{r}_g$ and scale $s_g$ are modified with offsets $d\boldsymbol{r}_g$ and $ds_g$ to get the final gaussians in the observation space. In the observation space, we obtain the rendered images through the rasterisation of gaussians and compute the image loss to supervise canonical gaussians. + +Once the warm-up iterations completes, we begin the adaptive joint growth. With the MK calculated during the warm-up phase, we can locate the parent joints $J_{s}$ that require growth. Then we add grown joints to the extra joint book $B^{e}$ , optimizing its positions $j^{e}$ and rotations $r^{e}$ during the subsequent training. Notably, $\Phi_{\mathrm{lbs}}$ needs to extend its output dimensions to include the blending weights for extra joints, following $K = K^0 + K^e$ . Since the extra joints do not have prior LBS weights, their blending weights are entirely learned through $\Phi_{\mathrm{lbs}}$ . + +Both the warm-up stage and the post-growth learning stage adopt adaptive density control to manage the updates of the gaussians. We also dynamically adjust the threshold + +of gradient for densification in [25] based on the number of gaussians, in order to balance memory consumption. Please check our supplemental material for details of this design. + +# 5. Experiments + +# 5.1. Dataset + +Our method focuses on hand-held objects and loose-fitting clothing, so we select datasets for experiments accordingly. We notice that the DNA-Rendering dataset [7], by capturing complex scenes of the human body, meets our requirements. Specifically, we selected 8 cases that align with our hypothesis, namely 0041_10, 0090_06, 0176_07, 0800_07 (hand-held objects) and 0007_04, 0014_06, 0051_09, 0811_06 (loose-fitting clothing). For each case, we use 24 surrounding views for training and 6 novel surrounding views for testing. All views are synchronized and contained 100 frames each. + +In addition to tackle complicated scenarios, it is essential to ensure the model's performance in typical scenarios involving tight-fitting clothes. Therefore, we additionally test our method on the ZJU-MoCap [46] dataset. Although the tight-clothing cases are too simple to require joint extension, our overall framework still achieves optimal results. Since this part of the experiment is not directly related to adaptive growth, we refer the readers to the supplemental material for further visualizations. + +# 5.2. Baselines and Metrics + +We select the most cutting-edge and representative works from each focus area for a fair comparison. Since 3DGS is currently the leading representation for novel view synthesis, we compare 3DGS-based methods, including 3DGS-Avatar [47], GART [29], and GauHuman [17]. Among them, GART is expected to offer extra animatability for its modeling of implicit global auxiliary bones. Additionally, there is another category of human modeling without incorporating SMPL-like pose priors. Although these methods don't guarantee an animatable human avatar, they can achieve high-quality rendering, among which, we compare the rendering quality of Im4D [33] with our method. + +We conduct a comprehensive comparison of our ToMiE against these methods. We report three key metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [51], and learned perceptual image patch similarity (LPIPS) [57]. Per-scene results can be found in the supplemental material. In addition to comparing the rendering results, we also demonstrate the animatability of the complicated human parts enabled by our method. We strongly recommend readers to watch the supplemental video for a more intuitive understanding of the animating. + +![](images/5fe11b50405fec08130ae01efdccb76a4b4c12d64eebc1aea431479086230f59.jpg) +Figure 4. Qualitative comparison on the DNA-Rendering dataset [7] with animatable baselines. We show two cases of hand-held objects (0800_07) and loose-fitting clothing (0811_06) (from top to bottom). Im4d* [33] achieves high-quality rendering but cannot be animated, and we compare it in the supplemental materials. Please check the supplemental video for better visualization. + +# 5.3. Novel View Synthesis Results + +Fig. 4, Tab. 1 and Tab. 2 present the results of our method compared to other baselines. In the tables, we showcase two evaluation protocols. The first evaluates the entire image, reflecting the overall rendering quality. The second uses a binary mask to specifically compare the complicated regions, demonstrating how our method outperforms others in these challenging cases. The mask is shown in Fig. 4, whose details can be found in the supplemental material. + +# 5.4. Ablation Studies + +A. Adaptive Growth Ablation. We remove the adaptive growth to ablate its impact on the rendering results. Tab. 3 "w/o Adaptive Growth" shows a decline in rendering quality, while the hand-held objects and loose-fitting clothing also become not animatable. + +B. Non-rigid Design Ablation. In Sec. 4.4, we apply a + +
Dataset Method | MetricAnimatablePSNR(full) ↑DNA-RenderingPSNR(masked) ↑
BGSSIM ↑LPIPS ↓
3DGS-Avatar✓✘28.930.9530.045917.69
GART✓✓29.250.9580.048017.12
Im4D✘✘26.280.9650.030814.90
GauHuman✓✘30.220.9620.040518.26
ToMiE(ours)✓✓31.280.9660.037419.75
+ +Table 1. Quantitative comparison between our method and other methods on complicated DNA-Rendering dataset [7]. $\mathbb{B}$ and $\mathbb{G}$ stand for human body and complex clothing (including handheld objects). We color each result as best, second best and third best. ToMiE achieves the best performance in PSNR and SSIM, while ranking second only to rendering-based Im4D [33] in LPIPS, mainly because rendering-based methods do not consider human structural constraints, resulting in higher visual fidelity. + +non-rigid deformation network $\Phi_d$ to correct gaussian rotation and scale. As shown in Tab. 3 "w/o Non-rigid Design", this significantly improves the final rendering quality. + +
Dataset Method | MetricAnimatableZJU-Mocap
BGPSNR(full) ↑SSIM ↑LPIPS ↓
3DGS-Avatar✓×30.610.9700.0296
GART✓✓30.910.9620.0318
Im4D××28.990.9730.0620
GauHuman✓×30.820.9620.0326
ToMiE(ours)✓✓31.100.9630.0312
+ +Table 2. Quantitative comparison between our method and other methods on ZJU-Mocap dataset [46] with tight-fitting clothing. $\mathbb{B}$ and $\mathbb{G}$ stand for human body and clothing (including hand-held objects). We color each result as best, second best and third best. This comparison demonstrates that ToMiE can achieve comparable (SSIM and LPIPS) or even superior (PSNR) performance to other approaches in tight-fitting scenarios where joint growth is not required. + +
PSNR(full) ↑SSIM ↑LPIPS ↓PSNR(masked) ↑
w/o Adaptive Growth30.990.9640.039019.21
w/o Non-rigid Design30.780.9640.036319.38
Full31.280.9660.037419.75
+ +Table 3. Ablation studies on DNA-Rendering dataset [7]. We independently ablate the adaptive growth strategy and non-rigid design to validate their impact on the overall performance. + +# 5.5. Animating Hand-held Objects and Loose-fitting clothing + +We demonstrate the uniqueness of our method, specifically its ability to explicitly animate hand-held objects and loose-fitting clothing. Our animating approach can be implemented in two ways. On the one hand, we can utilize the transformation already recorded in the extra joint book to replay clothing motions. The visualization of this part is shown in Fig. 5. On the other hand, ToMiE also supports bypassing the decoding process of the extra joint book by directly inputting the transformation explicitly. This allows us to customize the motion trajectory of external joints. Due to the limitations of the image's expressive capabilities, we strongly recommend the readers to watch the supplemental video to check the animated results. + +In Fig. 5, we edit the extra joints while keeping the body poses under the SMPL paradigm stationary. Since the implicit auxiliary bones of GART [29] is controlled by the traditional SMPL poses, only the identical appearance can be output when SMPL poses are stationary. In contrast, our method explicitly models hand-held objects and loose-fitting clothing, fully decoupling them from the traditional SMPL poses, enabling free animating. + +# 6. Limitations and Conclusion + +Limitations. Although our method enhances the modeling for rigid and non-rigid clothing, it cannot address scenarios involving drastic changes in the topology (e.g., taking off clothes or opening a book). This is because topological changes disrupt the one-to-one correspondence be + +![](images/2c9bb521bca23a855504ace1df68a04ec2f22d9f4e028c7a110c35bfefab242a.jpg) +Figure 5. Animating Results of ToMiE. Our explicit modeling fully decouples hand-held objects and loose-fitting clothing from the human body, enabling part-specific animating. + +tween frames, making the human modeling centered on the canonical space become downgraded. We notice that [42] addresses topology issues by introducing high-dimensional mappings, which could be adapted to build our non-rigid deformation. However, this is not the main scope of this paper and can be explored as future work. + +Conclusion. In this paper, we introduce ToMiE, an adaptive growth method designed to extend traditional SMPL skeleton for better modeling of hand-held objects and loose-fitting clothing. In the first stage, we assign the gradient of gaussian points to different joints by combining LBS weights with the motion kernel based on motion priors. This allows us to accurately locate the parent joints that need to grow, avoiding redundant growth. In the second stage, we design an extra joint book to achieve explicit joint modeling and optimize the transformation of the newly grown joints in an end-to-end manner. With the improved designs mentioned above, our ToMiE stands out among numerous state-of-the-art methods, achieving the best rendering quality and animatability of hand-held objects and loose-fitting clothing. We hope the adaptive growing method will spark a renewed discussion on the current capabilities of digital human modeling. What's more, it is expected to offer some insights for subsequent works related to topology and skeleton generation. + +# Acknowledgement + +This research was supported in part by the Tateisi Science and Technology Foundation, JSPS KAKENHI Grant Numbers 24KK0209 and 22H00529. This research was also supported by Shanghai Artificial Intelligence Laboratory. + +# References + +[1] I. Baran and J. Popović. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG), 26(3):72-es, 2007. 2 +[2] H. Bertiche, M. Madadi, and S. Escalera. Neural cloth simulation. ACM Transactions on Graphics (TOG), 41(6):1-14, 2022. 3 +[3] J. Chen, Y. Zhang, D. Kang, X. Zhe, L. Bao, X. Jia, and H. Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 3 +[4] X. Chen, T. Jiang, J. Song, M. Rietmann, A. Geiger, M. J. Black, and O. Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11796-11809, 2023. +[5] X. Chen, Y. Zheng, M. J. Black, O. Hilliges, and A. Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11594-11604, 2021. +[6] Y. Chen, Y. Zhan, Z. Zhong, W. Wang, X. Sun, Y. Qiao, and Y. Zheng. Within the Dynamic Context: Inertia-aware 3D Human Modeling with Pose Sequence. arXiv preprint arXiv:2403.19160, 2024. 3 +[7] W. Cheng, R. Chen, S. Fan, W. Yin, K. Chen, Z. Cai, J. Wang, Y. Gao, Z. Yu, Z. Lin, et al. DNA-Rendering: A Diverse Neural Actor Repository for High-fidelity Human-centric Rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 2, 6, 7, 8 +[8] E. Corona, A. Pumarola, G. Alenya, G. Pons-Moll, and F. Moreno-Noguer. SMPlicit: Topology-aware generative model for clothed people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11875-11885, 2021. 3 +[9] J. Dong, Q. Fang, W. Jiang, Y. Yang, H. Bao, and X. Zhou. Fast and robust multi-person 3d pose estimation and tracking from multiple views. In T-PAMI, 2021. 2 +[10] G. Gafni, J. Thies, M. Zollhofer, and M. Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3 +[11] Q. Gao, Y. Wang, L. Liu, L. Liu, C. Theobalt, and B. Chen. Neural novel actor: Learning a generalized animatable neural representation for human actors. IEEE Transactions on Visualization and Computer Graphics, 2023. +[12] C. Geng, S. Peng, Z. Xu, H. Bao, and X. Zhou. Learning neural volumetric representations of dynamic humans + +in minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023. +[13] S. Goel, G. Pavlakos, J. Rajasegaran, A. Kanazawa, and J. Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 3 +[14] C. Guo, T. Jiang, M. Kaufmann, C. Zheng, J. Valentin, J. Song, and O. Hilliges. ReLoo: Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild. arXiv preprint arXiv:2409.15269, 2024. 3 +[15] L. Hu. animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3 +[16] L. Hu, H. Zhang, Y. Zhang, B. Zhou, B. Liu, S. Zhang, and L. Nie. Gaussianavatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 3 +[17] S. Hu, T. Hu, and Z. Liu. Gauhuman: Articulated Gaussian Splatting from Monocular Human Videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 4, 6 +[18] T. Hu, F. Hong, and Z. Liu. Surmo: Surface-based 4d motion modeling for dynamic human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6550-6560, 2024. 3 +[19] M. Isik, M. Rünz, M. Georgopoulos, T. Khakhulin, J. Starck, L. Agapito, and M. Nießner. Humanrf: High-fidelity Neural Radiance Fields for Humans in Motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 3 +[20] A. Jacobson, I. Baran, L. Kavan, J. Popovic, and O. Sorkine. Fast automatic skinning transformations. ACM Transactions on Graphics (ToG), 31(4):1-10, 2012. 2 +[21] R. Jena, G. S. Iyer, S. Choudhary, B. Smith, P. Chaudhari, and J. Gee. Splatarmor: Articulated Gaussian Splatting for Animatable Humans from Monocular RGB Videos. arXiv preprint arXiv:2311.10812, 2023. 3 +[22] Y. Jiang, Z. Shen, Y. Hong, C. Guo, Y. Wu, Y. Zhang, J. Yu, and L. Xu. Robust Dual Gaussian Splatting for Immersive Human-centric Volumetric Videos. ACM Transactions on Graphics (TOG), 43(6):1-15, 2024. 3 +[23] Y. Jiang, Z. Shen, P. Wang, Z. Su, Y. Hong, Y. Zhang, J. Yu, and L. Xu. Hifi4g: High-fidelity Human Performance Rendering via Compact Gaussian Splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19734-19745, 2024. 3 +[24] H. Jung, N. Brasch, J. Song, E. Perez-Pellitero, Y. Zhou, Z. Li, N. Navab, and B. Busam. Deformable 3D Gaussian Splitting for Animatable Human Avatars. arXiv preprint arXiv:2312.15059, 2023. 3 +[25] B. Kerbl, G. Kopanas, T. Leimkuhler, and G. Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 3, 6 + +[26] M. Kocabas, J.-H. R. Chang, J. Gabriel, O. Tuzel, and A. Ranjan. Hugs: Human Gaussian Splats. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 505-515, 2024. 3 +[27] Y. Kwon, D. Kim, D. Ceylan, and H. Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 3 +[28] B. H. Le and Z. Deng. Two-layer sparse compression of dense-weight blend skinning. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 2 +[29] J. Lei, Y. Wang, G. Pavlakos, L. Liu, and K. Daniilidis. Gart: Gaussian Articulated Template Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19876-19887, 2024. 3, 6, 8 +[30] M. Li, J. Tao, Z. Yang, and Y. Yang. Human101: Training 100+ fps Human Gaussians in 100s from 1 View. arXiv preprint arXiv:2312.15258, 2023. +[31] M. Li, S. Yao, Z. Xie, K. Chen, and Y.-G. Jiang. Gaussian-body: Clothed human reconstruction via 3d gaussian splatt-ing. arXiv preprint arXiv:2401.09720, 2024. +[32] Z. Li, Z. Zheng, L. Wang, and Y. Liu. Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 3 +[33] H. Lin, S. Peng, Z. Xu, T. Xie, X. He, H. Bao, and X. Zhou. High-Fidelity and Real-Time Novel View Synthesis for Dynamic Scenes. In SIGGRAPH Asia Conference Proceedings, 2023. 3, 6, 7 +[34] X. Liu, C. Wu, X. Liu, J. Liu, J. Wu, C. Zhao, H. Feng, E. Ding, and J. Wang. Gea: Reconstructing expressive 3d gaussian avatar from monocular video. arXiv preprint arXiv:2402.16607, 2024. 3 +[35] Y. Liu, X. Huang, M. Qin, Q. Lin, and H. Wang. Animatable 3d gaussian: Fast and high-quality reconstruction of multiple human avatars. arXiv preprint arXiv:2311.16482, 2023. 3 +[36] M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black. SMPL: A Skinned Multi-Person Linear Model. ACM Transactions on Graphics, 34(6), 2015. 1, 2, 3 +[37] Y. Men, Y. Yao, M. Cui, and L. Bo. Mimo: Controllable character video synthesis with spatial decomposed modeling. arXiv preprint arXiv:2409.16160, 2024. 3 +[38] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 3, 5 +[39] A. Moreau, J. Song, H. Dhamo, R. Shaw, Y. Zhou, and E. Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 3 +[40] M. Niu, M. Cao, Y. Zhan, Q. Zhu, M. Ma, J. Zhao, Y. Zeng, Z. Zhong, X. Sun, and Y. Zheng. Anicrafter: Customizing realistic human-centric animation via avatar-background conditioning in video diffusion models. arXiv preprint arXiv:2505.20255, 2025. + +[41] H. Pang, H. Zhu, A. Kortylewski, C. Theobalt, and M. Habermann. Ash: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1165-1175, 2024. 3 +[42] K. Park, U. Sinha, P. Hedman, J. T. Barron, S. Bouaziz, D. B. Goldman, R. Martin-Brualla, and S. M. Seitz. Hypernerf: A Higher-dimensional Representation for Topologically Varying Neural Radiance Fields. arXiv preprint arXiv:2106.13228, 2021. 8 +[43] D. Paschalidou, A. Katharopoulos, A. Geiger, and S. Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3204-3215, 2021. 2 +[44] G. Pavlakos, V. Choutas, N. Ghorbani, T. Bolkart, A. A. A. Osman, D. Tzionas, and M. J. Black. Expressive Body Capture: 3D Hands, Face, and Body from a Single Image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 3 +[45] S. Peng, J. Dong, Q. Wang, S. Zhang, Q. Shuai, X. Zhou, and H. Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 3 +[46] S. Peng, Y. Zhang, Y. Xu, Q. Wang, Q. Shuai, H. Bao, and X. Zhou. Neural Body: Implicit Neural Representations with Structured Latent Codes for Novel View Synthesis of Dynamic Humans. In CVPR, 2021. 3, 6, 8 +[47] Z. Qian, S. Wang, M. Mihajlovic, A. Geiger, and S. Tang. 3dgs-avatar: Animatable Avatars via Deformable 3d Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020–5030, 2024. 3, 6 +[48] I. Santesteban, M. A. Otaduy, and D. Casas. Snug: self-supervised neural dynamic garments. IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), volume 2, page 9, 2022. 3 +[49] Q. Shuai, C. Geng, Q. Fang, S. Peng, W. Shen, X. Zhou, and H. Bao. Novel view synthesis of human interactions from sparse multi-view videos. In SIGGRAPH Conference Proceedings, 2022. 2 +[50] S. Wang, K. Schwarz, A. Geiger, and S. Tang. A随机变量的统计学，1987年，第2版，M. H. Lai, The asymptotic analysis of a random variable with a finite distribution, 1987. Springer, 2003. +[51] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[52] C.-Y. Weng, B. Curless, P. P. Srinivasan, J. T. Barron, and I. Kemelmacher-Shlizerman. Humannerf: Free-viewpoint Rendering of Moving People from Monocular Video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 3 +[53] Z. Xu, S. Peng, H. Lin, G. He, J. Sun, Y. Shen, H. Bao, and X. Zhou. 4K4D: Real-time 4D View Synthesis at 4K Resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20029-20040, 2024. 3 + +[54] Z. Xu, Y. Zhou, E. Kalogerakis, C. Landreth, and K. Singh. Rignet: Neural rigging for articulated characters. arXiv preprint arXiv:2005.00559, 2020. 2 +[55] C.-H. Yao, W.-C. Hung, Y. Li, M. Rubinstein, M.-H. Yang, and V. Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4853-4862, 2023. 2 +[56] Y. Zhan, W. Xu, Q. Zhu, M. Niu, M. Ma, Y. Liu, Z. Zhong, X. Sun, and Y. Zheng. R3-avatar: Record and retrieve temporal codebook for reconstructing photorealistic human avatars. arXiv preprint arXiv:2503.12751, 2025. 3 +[57] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[58] S. Zheng, B. Zhou, R. Shao, B. Liu, S. Zhang, L. Nie, and Y. Liu. Gps-gaussian: Generalizable Pixel-wise 3D Gaussian Splatting for Real-time Human Novel View Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 3 +[59] Y. Zheng, Q. Zhao, G. Yang, W. Yifan, D. Xiang, F. Dubost, D. Lagun, T. Beeler, F. Tombari, L. Guibas, et al. Physavatar: Learning the physics of dressed 3d avatars from visual observations. arXiv preprint arXiv:2404.04421, 2024. 3 +[60] B. Zhou, S. Zheng, H. Tu, R. Shao, B. Liu, S. Zhang, L. Nie, and Y. Liu. GPS-Gaussian+: Generalizable Pixel-wise 3D Gaussian Splatting for Real-Time Human-Scene Rendering from Sparse Views. arXiv preprint arXiv:2411.11363, 2024. 3 +[61] W. Zielonka, T. Bagautdinov, S. Saito, M. Zollhöfer, J. Thies, and J. Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3 \ No newline at end of file diff --git a/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/images.zip b/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f443c8f4b533a2e8936fd6d806a368f21768a6de --- /dev/null +++ b/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:852369f37bbef3f202ce65f3b64bb4a7372fad392e29a706303a7f0fcf4654e5 +size 479829 diff --git a/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/layout.json b/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8f74b311a17127aad848153aa9843bce1d4005f8 --- /dev/null +++ b/2025/Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars/layout.json @@ -0,0 +1,8974 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 553, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 553, + 138 + ], + "type": "text", + "content": "Towards Explicit Exoskeleton for the Reconstruction of Complicated 3D Human Avatars" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "spans": [ + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "text", + "content": "Yifan Zhan" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "text", + "content": " Qingtian Zhu" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "text", + "content": " Muyao Niu" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "text", + "content": " Mingze Ma" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "text", + "content": " Jiancheng Zhao" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "text", + "content": " Zhihang Zhong" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "text", + "content": " Xiao Sun" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "text", + "content": " Yu Qiao" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "text", + "content": " Yinqiang Zheng" + }, + { + "bbox": [ + 106, + 159, + 504, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 122, + 199, + 485, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 199, + 485, + 216 + ], + "spans": [ + { + "bbox": [ + 122, + 199, + 485, + 216 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 122, + 199, + 485, + 216 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory " + }, + { + "bbox": [ + 122, + 199, + 485, + 216 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 122, + 199, + 485, + 216 + ], + "type": "text", + "content": "The University of Tokyo" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 227, + 392, + 411 + ], + "blocks": [ + { + "bbox": [ + 59, + 227, + 392, + 411 + ], + "lines": [ + { + "bbox": [ + 59, + 227, + 392, + 411 + ], + "spans": [ + { + "bbox": [ + 59, + 227, + 392, + 411 + ], + "type": "image", + "image_path": "7fcea9dc1d75506f622186a76e8e80f51a270bf5b1ddda0b24eb1872536cdb06.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 419, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 54, + 419, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 54, + 419, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. We show two complex scenarios in 3D human modeling: with hand-held objects and loose-fitting clothing, which cannot be accurately represented by the standard SMPL model. Our ToMiE can realize adaptive growth to enhance the representation capability of SMPL without needing for time-consuming case-specific customization, achieving state-of-the-art results in both rendering and human (including complex scenarios) animation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 393, + 228, + 552, + 410 + ], + "blocks": [ + { + "bbox": [ + 393, + 228, + 552, + 410 + ], + "lines": [ + { + "bbox": [ + 393, + 228, + 552, + 410 + ], + "spans": [ + { + "bbox": [ + 393, + 228, + 552, + 410 + ], + "type": "image", + "image_path": "fae2fc5ceec61ad8c8edb91c423520bcda27587f4d765ad551779d288cf1b36f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 477, + 200, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 477, + 200, + 490 + ], + "spans": [ + { + "bbox": [ + 151, + 477, + 200, + 490 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 503, + 297, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 503, + 297, + 684 + ], + "spans": [ + { + "bbox": [ + 54, + 503, + 297, + 684 + ], + "type": "text", + "content": "In this paper, we highlight a critical yet often overlooked factor in most 3D human tasks, namely modeling complicated 3D human with with hand-held objects or loose-fitting clothing. It is known that the parameterized formulation of SMPL is able to fit human skin; while hand-held objects and loose-fitting clothing, are difficult to get modeled within the unified framework, since their movements are usually decoupled with the human body. To enhance the capability of SMPL skeleton in response to this situation, we propose a growth strategy that enables the joint tree of the skeleton to expand adaptively. Specifically, our method, called ToMiE, consists of parent joints localization and external joints optimization. For parent joints localization, we employ a gradient-based approach guided by both LBS blending weights and motion kernels. Once" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 479, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 587 + ], + "type": "text", + "content": "the external joints are obtained, we proceed to optimize their transformations in " + }, + { + "bbox": [ + 313, + 479, + 555, + 587 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 313, + 479, + 555, + 587 + ], + "type": "text", + "content": " across different frames, enabling rendering and explicit animation. ToMiE manages to outperform other methods across various cases with hand-held objects and loose-fitting clothing, not only in rendering quality but also by offering free animation of grown joints, thereby enhancing the expressive ability of SMPL skeleton for a broader range of applications. The code is available at https://github.com/Yifever20002/ToMiE." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 609, + 395, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "3D human reconstruction endeavors to model high-fidelity digital avatars based on real-world characters for virtual rendering and animating, which has been of long-term research value in areas such as gaming, virtual reality (VR), and beyond. Traditional methods, such as SMPL [36], achieve human body parameterization by performing principal component analysis (PCA) on large sets" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "*This work was done during the author's internship at the Shanghai Artificial Intelligence Laboratory.† denotes co-corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "14259" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "of 3D scanned meshes, allowing for the fitting of a specified identity. Recent neural rendering techniques have enabled implicit digital human modeling guided by Linear Blend Skinning (LBS) and SMPL skeleton, realizing lifelike rendering and animating from video inputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 294, + 277 + ], + "type": "text", + "content": "The neural-based 3D human rendering has been empowered by the cutting-edge technique, 3D Gaussian Splitting (3DGS) [25], for its real-time and high-quality novel view synthesis performance. By representing 3D gaussians under the canonical T-pose and utilizing the pre-extracted SMPL skeleton in the observation space, 3D human rendering results can be obtained from novel views in any frame. This stream of approaches proves high quality in rendering 3D humans that conform to the SMPL paradigm (e.g., avatars in tight-fitting clothing). However, we raise concerns regarding its ability to handle complicated human modeling involving hand-held objects or loose-fitting clothing." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 434 + ], + "type": "text", + "content": "In Fig. 1, we show two cases to illustrate the limitations of current 3D human modeling. On the one hand, the movements of hand-held objects, e.g., feather duster, are highly decoupled from the human body and thus cannot be represented by the SMPL. On the other hand, characters shot in-the-wild are dressed in clothing with high complexity in dynamics, rather than the tight-fitting clothing configured under strict experimental conditions. These complicated scenarios break the existing methods' paradigm by assuming that surface of avatars should be bounded to the motion warping of SMPL skeleton in the same way as the human skin is. Rooted in the aforementioned scenarios, SMPL fails to accurately fit such 3D human models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 436, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 592 + ], + "type": "text", + "content": "To this end, we break through the limitations of modeling complicated 3D human gaussians by maintaining an extended joint tree from SMPL skeleton. Although existing SMPL model has the potential capability to manually customize additional skeleton information, time-consuming and case-by-case adjustments are necessary in this case. To overcome this issue, we extend the SMPL skeleton with additional joints for each individual case adaptively. The growth is performed in an explicit and adaptive manner, and enables the fitting of complicated 3D human avatars, offering high-quality rendering with hand-held objects and loose-fitting clothing (e.g., novel view synthesis results in Fig. 1)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "The main challenge of extending the SMPL skeleton is to determine where and how to grow additional joints. To avoid the unnecessary memory consumption and potential overfitting caused by an arbitrary growth, we first determine which of the joints are supposed to serve as parent joints by a localization strategy. We have empirically observed that parent joints requiring growth witness larger backpropagation gradients in their associated gaussians due to underfitting. However, determining the association of gaussians with different joints is non-trivial, as the SMPL's LBS" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "blending weights may not work with the human with handheld objects or clothing. To more precisely define such association, we introduce the concept of Motion Kernels based on rigid body priors and combine them with LBS weights, resulting in more accurate gradient-based localization. After growth, we adaptively maintain an extended joint tree and update the extra joints by optimizing two MLP decoders for joints positions and rotations. The proposed method, termed as ToMiE, allows for explicit rendering and animation of complicated human avatars represented by extended joints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 210, + 553, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 553, + 269 + ], + "type": "text", + "content": "By experiments on complicated cases of the DNA-Rendering dataset [7], ToMiE exhibits state-of-the-art rendering quality while maintaining the animatability that is significant for downstream productions. To summarize, our contributions are three-fold as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 310, + 275, + 553, + 407 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 310, + 275, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 275, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 310, + 275, + 553, + 335 + ], + "type": "text", + "content": "1) ToMiE, a method for creating an enhanced SMPL joint tree via an adaptive growth strategy, which is able to decouple complicated parts from the human body, thereby achieving state-of-the-art results in both rendering and explicit animation on target cases;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 310, + 335, + 553, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 335, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 310, + 335, + 553, + 383 + ], + "type": "text", + "content": "2) a hybrid assignment strategy for gaussians utilizing LBS weights and Motion Kernels, combined with gradient-driven parent joints localization, to guide the growth of external joints;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 310, + 384, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 384, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 310, + 384, + 553, + 407 + ], + "type": "text", + "content": "3) a joints optimization approach fitting local rotations across different frames while sharing joints positions." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 436, + 400, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 436, + 400, + 449 + ], + "spans": [ + { + "bbox": [ + 314, + 436, + 400, + 449 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 462, + 503, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 462, + 503, + 475 + ], + "spans": [ + { + "bbox": [ + 314, + 462, + 503, + 475 + ], + "type": "text", + "content": "2.1. SMPL-based Human Mesh Avatars" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "text", + "content": "Most of the recent success in digital human modeling can be attributed to the contributions of the SMPL [36] series, which parameterizes the human body as individual shape components and motion-related human poses through 3D mesh scanning and PCA. The pose blend shapes in SMPL describe human body deformations as a linear weighted blending of different joint poses, significantly improving the efficiency of editing and animating digital humans. Furthermore, it has been widely adopted for human body animation, thanks to the methods [9, 49] of estimating SMPL parameters from 2D inputs. Despite their wide range of applications, SMPL and its family still suffer from inherent limitations. Since the originally scanned 3D meshes are skin-tight, according to which the pose blend shapes are learned, the model is unable to handle significantly outlying meshes, such as human with hand-held objects and clothing like skirts. Although this could potentially be solved by auto-rigging [1, 20, 28, 43, 54, 55], it requires 2D/3D shapes as inputs, which is not accessible in human videos." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14260" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 255, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 255, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 255, + 85 + ], + "type": "text", + "content": "2.2. Neural Representation for 3D Human" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 245 + ], + "type": "text", + "content": "Methods based on neural representations, such as NeRF [38] and 3DGS [25], have also been playing an important part in digital human reconstruction for their high-quality rendering capabilities. Early NeRF-based methods [3-6, 10-13, 27, 46, 52] aim to reconstruct human avatars by inputting monocular or multi-view synchronized videos. [50] enforce smooth priors based on neural Signed Distance Function (SDF) to obtain more accurate human geometry. Recent breakthroughs [16-18, 21, 24, 26, 29-32, 34, 35, 39-41, 47, 56, 58, 59, 61] rely on 3DGS, enabling faster and more accurate rendering. All these methods register the T-pose in a canonical space and use LBS weights to guide the rigid transformations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 252, + 288, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 288, + 266 + ], + "type": "text", + "content": "2.3. Rendering & Editing of Intricate 3D Human" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 629 + ], + "type": "text", + "content": "We revisit the methods attempting to implicitly improve modeling of complicated 3D human. Animatable NeRF [45] defines a per-frame latent code to capture appearance variations across each frame. Simulation methods [2, 48] physically construct simple clothing but struggle with complex one and objects. [14, 29] additionally register global latent bones to compensate for the limitations in clothing rendering. They fail to explicitly decouple the clothing from human body, making precise control infeasible. Another stream [6, 18] leverages a human poses sequence as contexts to resolve appearance ambiguities. But correspondingly, they require a sequence of human poses for animation, adding to the challenges of editing. Moreover, these methods struggle to fit the object-level pose independent of the human poses sequence, such as hand-held items. We also note that some works [15, 37] introduce diffusion-based generative methods to enhance the realism of garment rendering, but these methods are restricted by the traditional SMPL and overlook the editing of complex clothing. Although SMPLlicit [8] enables implicit interpolation of clothing types, it remains dependent on SMPL's LBS process to generate the observed mesh. This limitation prevents localized explicit animation of loose-fitting clothing and restricts its application to external objects. Rendering-based human reconstruction method [19, 22, 23, 33, 53, 58, 60] can not achieve animation. Our approach enables explicit decoupling of hand-held objects and clothing with the human body by extending the SMPL joint tree, allowing for high-quality rendering and explicit animation in complicated scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 640, + 141, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 141, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 141, + 653 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 659, + 174, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 174, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 174, + 672 + ], + "type": "text", + "content": "3.1. SMPL(-X) Revisited" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Pre-trained on scanned meshes, the SMPL(-X) family [36, 44] employs a parameterized model to fit human bodies of different shapes and under different poses. The" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "human mesh in each frame evolves from a canonical human mesh and is controlled by shape and pose parameters. Specifically, a 3D point " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\pmb{x}_c" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " on the canonical mesh will be warped to obtain point in the observation space as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 347, + 125, + 555, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 125, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 347, + 125, + 555, + 159 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {o} = \\sum_ {k = 1} ^ {K} \\omega_ {k} \\left(\\boldsymbol {x} _ {c}\\right) \\left(R _ {k} \\left(\\boldsymbol {r} ^ {0}\\right) \\boldsymbol {x} _ {c} + t _ {k} \\left(\\boldsymbol {j} ^ {0}, \\beta\\right)\\right), \\tag {1}", + "image_path": "5e882e6ff35f190b255d1388f452de556414ad3a6344b60dc3f2aa92f2d1a02e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "text", + "content": " is the total number of joints, " + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "inline_equation", + "content": "R_{k}" + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "text", + "content": " is per-joint global rotation controlled by local joint rotations " + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "inline_equation", + "content": "\\pmb{r}^{0}" + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "text", + "content": " is per-joint translation controlled by joint positions " + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "inline_equation", + "content": "\\pmb{j}^0" + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "text", + "content": " and human shape " + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "text", + "content": ". Notice that linear blending weight " + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "inline_equation", + "content": "\\omega_{k}" + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "text", + "content": " is a function of " + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "inline_equation", + "content": "\\pmb{x}_c" + }, + { + "bbox": [ + 313, + 163, + 554, + 223 + ], + "type": "text", + "content": ", which is regressed from large human assets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 224, + 554, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 554, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 554, + 295 + ], + "type": "text", + "content": "The main issue with this LBS-based prior model is that it can only model tight-fitting avatars conforming to the SMPL(-X) paradigm, making it unsuitable for modeling complex human clothing in more generic cases. What even more challenging is its inability to handle hand-held objects that are fully decoupled from the human pose." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 301, + 469, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 469, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 469, + 312 + ], + "type": "text", + "content": "3.2. Human Gaussians Revisited" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 319, + 554, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 401 + ], + "type": "text", + "content": "Human gaussians achieve high-quality real-time human rendering by combining the SMPL prior with 3DGS as the representation. The SMPL model naturally obtains the T-posed (i.e., all human poses are identity transformations) mesh in the canonical space and then mesh vertices are used to initialize the canonical gaussian units. Each gaussian is defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 403, + 553, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 403, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 347, + 403, + 553, + 430 + ], + "type": "interline_equation", + "content": "G (\\boldsymbol {x}) = \\frac {1}{(2 \\pi) ^ {\\frac {3}{2}} | \\boldsymbol {\\Sigma} | ^ {\\frac {1}{2}}} e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {T} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})}, \\tag {2}", + "image_path": "cae589d7e83fe9b9b840cf56ade1f6f7ebe16d3267883195a2703ae11e11148c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "text", + "content": " is the 3D gaussian center, and " + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "text", + "content": " is the 3D covariance matrix, which will be further decomposed into learnable rotation " + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "inline_equation", + "content": "\\pmb{S}" + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "text", + "content": ". Now we have " + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma} = \\boldsymbol{RSS}^{\\top} \\boldsymbol{R}^{\\top}" + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "text", + "content": ", which is performed by optimizing a quaternion " + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_g" + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "text", + "content": " for rotation and a 3D vector " + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "inline_equation", + "content": "\\boldsymbol{s}_g" + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "text", + "content": " for scaling. Each gaussian is further assigned with color " + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 433, + 554, + 505 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "type": "text", + "content": "Once the 3D gaussians in the canonical space are obtained, each position " + }, + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "type": "text", + "content": " will be warped to the observation space according to Eq. (1). Next, the 3D gaussians of each frame are projected into 2D gaussians, followed by tile-based rasterization. Color of each pixel can be calculated by blending " + }, + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 505, + 554, + 577 + ], + "type": "text", + "content": " ordered gaussians following" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 377, + 581, + 553, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 581, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 377, + 581, + 553, + 613 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {3}", + "image_path": "24df7e404a13874c9c9c2dc3a2dfa1c8b3e7bf11e5899f972819faddcd933ce2.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "type": "text", + "content": "The 3D gaussians can be optimized and updated by adaptive density control, which primarily includes cloning, splitting, and pruning. Cloning and splitting are guided by gradients to control the number of gaussians, while pruning removes empty gaussians based on current opacity " + }, + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "type": "text", + "content": ". The supervision upon human gaussians is derived from multiview or monocular videos, enabling high-quality rendering and avatar animation." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14261" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 116, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 116, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 116, + 83 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 283 + ], + "type": "text", + "content": "Fig. 2 illustrates ToMiE's adaptive joint growth and gaussian training strategy. Our goal is to extend the SMPL skeleton to handle complicated human scenarios. However, the abuse of growth can lead to unnecessary computational and memory overheads. For an efficient adaptive growth, we first propose a localization strategy of parent joints to ensure that only necessary joints are grown. Furthermore, we explicitly define the hand-held joints in the " + }, + { + "bbox": [ + 55, + 91, + 296, + 283 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 55, + 91, + 296, + 283 + ], + "type": "text", + "content": " space and optimize them end-to-end through an MLP, ensuring alignment with the original skeleton of SMPL. The extended skeleton can thus support rendering and explicit animation. To address the limitation of LBS in guiding gaussian attributes apart from the positions, we further fine-tune the rotation and scale during training with a deformation field to achieve better non-rigid warping. Next, we will elaborate on these modules and the training strategy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 292, + 296, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 316 + ], + "type": "text", + "content": "4.1. Motion Kernels-guided Joint Gradient Accumulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "text", + "content": "A quantitative metric needs to be identified to determine whether a joint requires growth. We notice that, due to the poor fitting ability of existing SMPL-based human gaussian, it will leave larger gradients in complex human regions (e.g., the hand-held object in Fig. 2 ③). In other words, identifying joints with larger accumulated gradients can help to indicate which parent joints are more likely to require growth. Bounded with each gaussian, the accumulated gradients will be first assigned to their corresponding joints. Let " + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "inline_equation", + "content": "g = \\|(g_x, g_y, g_z)\\|_2" + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "text", + "content": " represent the L2 norm of the gradient at each gaussian position. The gradients accumulation " + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "inline_equation", + "content": "g_J" + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "text", + "content": "-th joint can be computed following" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 126, + 468, + 295, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 468, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 126, + 468, + 295, + 498 + ], + "type": "interline_equation", + "content": "g _ {J _ {k}} = \\frac {\\sum_ {i \\in N} \\omega_ {k} \\left(\\boldsymbol {x} _ {c}\\right) g}{\\sum_ {i \\in N} \\omega_ {k} \\left(\\boldsymbol {x} _ {c}\\right)}. \\tag {4}", + "image_path": "3a1c7507d1495137bed3469c81b7f722c42814ffffe1cbee612827d9a60a5546.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 500, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 295, + 523 + ], + "type": "text", + "content": "We use " + }, + { + "bbox": [ + 55, + 500, + 295, + 523 + ], + "type": "inline_equation", + "content": "\\pmb{x}_c" + }, + { + "bbox": [ + 55, + 500, + 295, + 523 + ], + "type": "text", + "content": " to stand for gaussian position in canonical space and " + }, + { + "bbox": [ + 55, + 500, + 295, + 523 + ], + "type": "inline_equation", + "content": "\\pmb{x}_o" + }, + { + "bbox": [ + 55, + 500, + 295, + 523 + ], + "type": "text", + "content": " in the observation space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "text", + "content": "It is worth emphasizing that " + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\omega_{k}" + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "text", + "content": " here is a weight term determining the assignment of a gaussian to the " + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "text", + "content": "-th joint. Under the paradigm of SMPL representation, this weight corresponds to the LBS weight " + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{lbs0}}" + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "text", + "content": ", and guides the rigid transformation of vertices on the SMPL mesh according to the human pose. To account for the differences between the human mesh with clothing and the vanilla SMPL mesh, [17] calculate " + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{lbs}}" + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "text", + "content": " by keeping using the LBS weight prior " + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{lbs0}}" + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "text", + "content": " and adding an extra learnable network " + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{lbs}}" + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "text", + "content": " for fine-tuning. This formulation (with " + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 524, + 296, + 643 + ], + "type": "text", + "content": " omitted) can be summarized as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 648, + 295, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 648, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 86, + 648, + 295, + 662 + ], + "type": "interline_equation", + "content": "\\omega_ {\\mathrm {l b s}} \\left(\\boldsymbol {x} _ {c}\\right) = \\omega_ {\\mathrm {l b s 0}} \\left(\\mathrm {N N} _ {1} \\left(\\boldsymbol {x} _ {c}, \\boldsymbol {V}\\right)\\right) + \\Phi_ {\\mathrm {l b s}} \\left(\\boldsymbol {x} _ {c}\\right), \\tag {5}", + "image_path": "cd55fc6af3a2e421d79bf20dae5c03d9e5af0128a8eb470e81b157fb3ba20008.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{NN}_1" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " stands for top-1 nearest-neighbor search algorithm and " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " represents the canonical standard SMPL vertices. It will be, however, clarified by us, that this nearest neighbor-based assigning method will no longer be feasible" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 530, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 530, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 530, + 84 + ], + "type": "text", + "content": "in cases with handheld objects and complex clothing." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 84, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 203 + ], + "type": "text", + "content": "In Fig. 3, we present an example of misclassification by the nearest-neighbor search algorithm in Eq. (5). Due to the lack of human topology constraints in the canonical space, incorrect classification of hand-held objects can occur, as shown in Fig. 3 (a). This part, following the naive nearest-neighbor search, would be assigned to the leg by mistake. Fig. 3 (b) shows the points belonging to the hand joint (the correct parent joint where the handheld object should grow). Yet being solely guided by LBS weights results in the voids by misclassification." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "text", + "content": "To mitigate this issue, we propose a more robust assignment method based on motion priors, which we call Motion Kernels. Specifically, in the observation space, the motion kernel of each point " + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "inline_equation", + "content": "\\pmb{x}_o" + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "text", + "content": " with respect to the " + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "text", + "content": "-th joint position " + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "inline_equation", + "content": "j_k" + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "text", + "content": " is defined based on the changes in their pairwise Euclidean distances through all input " + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 204, + 553, + 276 + ], + "type": "text", + "content": " frames, following" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 342, + 280, + 553, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 280, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 342, + 280, + 553, + 312 + ], + "type": "interline_equation", + "content": "\\operatorname {M K} \\left(\\boldsymbol {x} _ {c}, \\boldsymbol {j} _ {k}\\right) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\left\\| \\boldsymbol {x} _ {o i} - \\boldsymbol {j} _ {k i} \\right\\| _ {2} - \\mu\\right) ^ {2}, \\tag {6}", + "image_path": "be3c0ede179a807aa058043e8b4a7f5ac06e20ec7a42ffa988c6ab82554f0ef2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 378, + 316, + 553, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 316, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 378, + 316, + 553, + 349 + ], + "type": "interline_equation", + "content": "\\mu = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| \\boldsymbol {x} _ {o i} - \\boldsymbol {j} _ {k i} \\| _ {2}. \\tag {7}", + "image_path": "9faaab17a96731d043a60914f626d9e1b2326ab5b6bf7a96a79fbb5021c18e6c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 351, + 554, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 554, + 459 + ], + "type": "text", + "content": "Our motion kernel (MK) reflects the relative motion between each gaussian and joint. A smaller MK indicates that the pair of gaussian and joint is relatively stationary to each other, signifying a stronger association, while a larger MK suggests a higher degree of relative motion, indicating a weaker association. We further represent the assignment weight reflected by the MK as " + }, + { + "bbox": [ + 313, + 351, + 554, + 459 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{MK}}(\\boldsymbol{x}_c) = \\mathrm{Normalize}_k(\\mathrm{MK}^{-1}(\\boldsymbol{x}_c,\\boldsymbol{j}_k))" + }, + { + "bbox": [ + 313, + 351, + 554, + 459 + ], + "type": "text", + "content": ", and the final assignment weight in Eq. (4) (with index " + }, + { + "bbox": [ + 313, + 351, + 554, + 459 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 351, + 554, + 459 + ], + "type": "text", + "content": " omitted) becomes" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 352, + 463, + 553, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 463, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 352, + 463, + 553, + 476 + ], + "type": "interline_equation", + "content": "\\omega \\left(\\boldsymbol {x} _ {c}\\right) = \\lambda \\omega_ {\\mathrm {M K}} \\left(\\boldsymbol {x} _ {c}\\right) + (1 - \\lambda) \\omega_ {\\mathrm {l b s}} \\left(\\boldsymbol {x} _ {c}\\right), \\tag {8}", + "image_path": "984eaf49f27b986fe461b1502dadd870abd85d054a7b2710478437cd2c38e741.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 480, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 480, + 554, + 540 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 480, + 554, + 540 + ], + "type": "text", + "content": " is a hyperparameter to balance the MK weight and LBS weight. Note that we do not completely abandon the LBS weight, because the MK cannot differentiate the association of points on either side of a joint (\"misdirect\" in Fig. 3), requiring the LBS weight to compensate." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 546, + 531, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 531, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 531, + 558 + ], + "type": "text", + "content": "4.2. Gradient-based Parent Joint Localization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "text", + "content": "By combining Eq. (4) and Eq. (8), the parent joints " + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "inline_equation", + "content": "J_{s} \\subseteq J" + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "text", + "content": " that require growth can be located. Basically, we have " + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "inline_equation", + "content": "g_{J} = (g_{J_{1}}, g_{J_{2}}, \\ldots, g_{J_{K}})" + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "text", + "content": " to represent the gradient accumulation of total " + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "text", + "content": " human joints. As mentioned in Sec. 4.1, the joints with larger gradients accumulated are more likely to require the growth of child joints. This is achieved by sorting " + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "inline_equation", + "content": "g_{J}" + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "text", + "content": " in descending order " + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "inline_equation", + "content": "g_{J}^{\\mathrm{sorted}} = (g_{J_{\\pi(1)}}, g_{J_{\\pi(2)}}, \\ldots, g_{J_{\\pi(K)}})" + }, + { + "bbox": [ + 313, + 563, + 554, + 660 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "type": "text", + "content": "We set a gradient threshold " + }, + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "type": "inline_equation", + "content": "\\epsilon_{J}" + }, + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "type": "text", + "content": " to identify the " + }, + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "type": "inline_equation", + "content": "J\\in J_{s}" + }, + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "type": "text", + "content": " that requires growth, following" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 372, + 687, + 553, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 687, + 553, + 706 + ], + "spans": [ + { + "bbox": [ + 372, + 687, + 553, + 706 + ], + "type": "interline_equation", + "content": "\\boldsymbol {J} _ {s} = \\left(J _ {\\pi (1)}, J _ {\\pi (2)}, \\dots , J _ {\\pi (N)}\\right) \\tag {9}", + "image_path": "b7ba36a6da44cd9f9afb0c45d61141286b7812b199abc1c7b56a9377837f4dae.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 358, + 704, + 511, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 704, + 511, + 716 + ], + "spans": [ + { + "bbox": [ + 358, + 704, + 511, + 716 + ], + "type": "interline_equation", + "content": "s. t. \\quad g _ {J _ {\\pi (N)}} \\geq \\epsilon_ {J} \\text {a n d} g _ {J _ {\\pi (N + 1)}} < \\epsilon_ {J},", + "image_path": "60b1d6d65bf4d3184b26c0c1680f26b3b987e5eb8db4fcf094225e913537d33b.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14262" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 71, + 541, + 352 + ], + "blocks": [ + { + "bbox": [ + 72, + 71, + 541, + 352 + ], + "lines": [ + { + "bbox": [ + 72, + 71, + 541, + 352 + ], + "spans": [ + { + "bbox": [ + 72, + 71, + 541, + 352 + ], + "type": "image", + "image_path": "5e78f3ac3a3766ff8f0dce7ee65b712815d4950a870f54df892dd7313de2956c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "lines": [ + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "type": "text", + "content": "Figure 2. The pipeline of ToMiE. " + }, + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "type": "inline_equation", + "content": "①" + }, + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "type": "text", + "content": " We initialize the gaussians in the canonical space with a standard SMPL vertices. " + }, + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "type": "inline_equation", + "content": "②" + }, + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "type": "text", + "content": " (Sec. 4.4) We apply Linear Blend Skinning (LBS) to the gaussian position and utilize a network for rotation and scale correction. During the warmup phase, Adaptive LBS only utilizes the original SMPL skeleton. After adaptive growth, it further includes the newly grown external skeleton. " + }, + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "type": "inline_equation", + "content": "③" + }, + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "type": "text", + "content": " Gaussian rasterization and gradients backpropagation. " + }, + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "type": "inline_equation", + "content": "④" + }, + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "type": "text", + "content": " (Sec. 4.1, Sec. 4.2) We employ a gradient-based parent joints localization method, and a motion kernel to optimize the gradient assignment process. " + }, + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "type": "inline_equation", + "content": "⑤" + }, + { + "bbox": [ + 54, + 361, + 555, + 429 + ], + "type": "text", + "content": " (Sec. 4.3) We maintain an extra joint book with MLPs, which generates explicit human pose, enabling the decoupling and explicit animating of hand-held objects and loose-fitting clothing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 438, + 164, + 515 + ], + "blocks": [ + { + "bbox": [ + 59, + 438, + 164, + 515 + ], + "lines": [ + { + "bbox": [ + 59, + 438, + 164, + 515 + ], + "spans": [ + { + "bbox": [ + 59, + 438, + 164, + 515 + ], + "type": "image", + "image_path": "eca421488364b38a297a7f32e72bb5af3c3725a9d0e1c3f56b5fc7b3d06d8f5c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 517, + 113, + 525 + ], + "lines": [ + { + "bbox": [ + 102, + 517, + 113, + 525 + ], + "spans": [ + { + "bbox": [ + 102, + 517, + 113, + 525 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 165, + 438, + 293, + 515 + ], + "blocks": [ + { + "bbox": [ + 165, + 438, + 293, + 515 + ], + "lines": [ + { + "bbox": [ + 165, + 438, + 293, + 515 + ], + "spans": [ + { + "bbox": [ + 165, + 438, + 293, + 515 + ], + "type": "image", + "image_path": "53c973ec0b4b56750e6faf1ae362145bddc3502ac77ef6866a7a1970a903ace8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 517, + 235, + 526 + ], + "lines": [ + { + "bbox": [ + 223, + 517, + 235, + 526 + ], + "spans": [ + { + "bbox": [ + 223, + 517, + 235, + 526 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 534, + 295, + 612 + ], + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 612 + ], + "type": "text", + "content": "Figure 3. Principle of the Motion Kernel. Relying solely on LBS weights will bring the misclassification in the canonical space shown in (a) to the observation space in (b), resulting in voids. Our proposed motion kernel focuses on motion-dependent priors in the observation space, offering better robustness and being less sensitive to misclassifications in the canonical space. This aids for point assignment process in parent joints localization." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 628, + 208, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 208, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 208, + 639 + ], + "type": "text", + "content": "where we can safely assume " + }, + { + "bbox": [ + 55, + 628, + 208, + 639 + ], + "type": "inline_equation", + "content": "N < K" + }, + { + "bbox": [ + 55, + 628, + 208, + 639 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "For each joint in " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "J_{s}" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": ", we designate it as a parent joint requiring growth and proceed with the initialization of its child joint. In order to explicitly model each child joint and ensure the consistency with the SMPL paradigm for ease of animation, we maintain an extra joint book " + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "inline_equation", + "content": "B^{e} = (\\text{parent}, j^{e}, r^{e})" + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": ", that includes the indices of parent joints," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 440, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 555, + 525 + ], + "type": "text", + "content": "the extra joint positions " + }, + { + "bbox": [ + 313, + 440, + 555, + 525 + ], + "type": "inline_equation", + "content": "j^{e}" + }, + { + "bbox": [ + 313, + 440, + 555, + 525 + ], + "type": "text", + "content": " in canonical space and the extra rotations " + }, + { + "bbox": [ + 313, + 440, + 555, + 525 + ], + "type": "inline_equation", + "content": "r^{e}" + }, + { + "bbox": [ + 313, + 440, + 555, + 525 + ], + "type": "text", + "content": " in its parent joint coordinate. We initialize the joint position to its parent joint's position and set the rotation to the identity rotation. The entire parent joint localization and child joint initialization process is guided by the gradients, effectively preventing unnecessary overgrowth and ensuring a dense distribution of the extra joints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 536, + 455, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 455, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 455, + 549 + ], + "type": "text", + "content": "4.3. Extra Joint Optimization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 555, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 628 + ], + "type": "text", + "content": "The joint positions and rotations in the extra joint book are set as estimizable, which will be stored and later decoded by two shallow MLPs. According to the SMPL paradigm, the canonical joint position is a time-invariant quantity, therefore the joint position optimization network " + }, + { + "bbox": [ + 313, + 555, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\Phi_p" + }, + { + "bbox": [ + 313, + 555, + 555, + 628 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 388, + 631, + 553, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 631, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 388, + 631, + 553, + 645 + ], + "type": "interline_equation", + "content": "d \\boldsymbol {j} ^ {e} (i) = \\Phi_ {p} (P. E. (i)), \\tag {10}", + "image_path": "64f1fcaa40618161ac0d1643fd731ef4229214afd12c77af584602b138239f16.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "type": "inline_equation", + "content": "P.E." + }, + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "type": "text", + "content": " is a positional encoding function as is in [38] and " + }, + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "type": "text", + "content": " is the joint index in " + }, + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "type": "inline_equation", + "content": "B^{e}" + }, + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "type": "text", + "content": ". Rotations are also dependent on the timestamp " + }, + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "type": "text", + "content": " of each frame, thus the rotation optimization network " + }, + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\Phi_{r}" + }, + { + "bbox": [ + 313, + 648, + 555, + 696 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 370, + 700, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 370, + 700, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\boldsymbol {r} ^ {e} (i, t) = \\Phi_ {r} (P. E. (i), P. E. (t)). \\tag {11}", + "image_path": "4ec777b34847c3dda0bf9d17b76d01f5a3674ad20b0f66b6555a009bda476ac2.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14263" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "content": "Now we have the positions of extra joints " + }, + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "j^{e}" + }, + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "content": " (optimized by the offset " + }, + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "d\\boldsymbol{j}^{e}" + }, + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "content": ") and the rotations " + }, + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}^{e}" + }, + { + "bbox": [ + 54, + 72, + 294, + 133 + ], + "type": "text", + "content": ". Each extra joint position is defined in the canonical space, representing the intrinsic properties of the extended skeleton, while the extra joint rotation in parent joint coordinate varies in each frame." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 133, + 294, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 133, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 54, + 133, + 294, + 242 + ], + "type": "text", + "content": "Although both the extra joint positions and rotations are stored in the MLPs, their inputs and outputs are explicit features with real physical meaning, which allows for both implicit and explicit editing. For example, we can interpolate over timestamp " + }, + { + "bbox": [ + 54, + 133, + 294, + 242 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 133, + 294, + 242 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 54, + 133, + 294, + 242 + ], + "type": "inline_equation", + "content": "\\Phi_r" + }, + { + "bbox": [ + 54, + 133, + 294, + 242 + ], + "type": "text", + "content": " or directly use explicit inputs to take the place of " + }, + { + "bbox": [ + 54, + 133, + 294, + 242 + ], + "type": "inline_equation", + "content": "\\Phi_r" + }, + { + "bbox": [ + 54, + 133, + 294, + 242 + ], + "type": "text", + "content": " during animation. The MLPs here function as decoders, deriving joint-related values from indices and timestamps, thus effectively circumventing the need for explicit storage of joint values employed in SMPL." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 266, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 266, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 266, + 264 + ], + "type": "text", + "content": "4.4. Inference Process and Training Strategy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 269, + 295, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 295, + 293 + ], + "type": "text", + "content": "In this subsection, we explain how our adaptive growth is integrated with the training process of human Gaussians." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 294, + 295, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 377 + ], + "type": "text", + "content": "First, we initialize the canonical gaussians with standard SMPL vertices. At the beginning of training, we set up a number of warm-up iterations during which no joint growth occurs, and the gaussian fitting is performed following the traditional human gaussian methods. This prevents underfitting due to insufficient training, which could further affect the joint localization in Sec. 4.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 379, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 475 + ], + "type": "text", + "content": "During the warm-up iterations, canonical human gaussians will be first warped to the observation space according to the LBS weight in Eq. (5). To compensate for the inability of the LBS model to represent gaussians rotation and scale, a deformable network " + }, + { + "bbox": [ + 55, + 379, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\Phi_d" + }, + { + "bbox": [ + 55, + 379, + 295, + 475 + ], + "type": "text", + "content": " is employed to correct the rotation and scale during the LBS process. To distinguish it from the joint rotation " + }, + { + "bbox": [ + 55, + 379, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\pmb{r}" + }, + { + "bbox": [ + 55, + 379, + 295, + 475 + ], + "type": "text", + "content": " of the SMPL human pose, we denote the rotation of the gaussian with subscript " + }, + { + "bbox": [ + 55, + 379, + 295, + 475 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 55, + 379, + 295, + 475 + ], + "type": "text", + "content": ", and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 478, + 294, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 478, + 294, + 493 + ], + "spans": [ + { + "bbox": [ + 125, + 478, + 294, + 493 + ], + "type": "interline_equation", + "content": "d \\boldsymbol {r} _ {g}, d \\boldsymbol {s} _ {g} = \\Phi_ {d} \\left(\\boldsymbol {x} _ {g}, \\boldsymbol {r} ^ {0}\\right). \\tag {12}", + "image_path": "5129e4fd43cf14393c9ccea7bf39ff3b48aebcb71644fa782e5a133f9a2af8fc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 496, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 556 + ], + "type": "text", + "content": "The gaussians rotation " + }, + { + "bbox": [ + 55, + 496, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_g" + }, + { + "bbox": [ + 55, + 496, + 295, + 556 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 55, + 496, + 295, + 556 + ], + "type": "inline_equation", + "content": "s_g" + }, + { + "bbox": [ + 55, + 496, + 295, + 556 + ], + "type": "text", + "content": " are modified with offsets " + }, + { + "bbox": [ + 55, + 496, + 295, + 556 + ], + "type": "inline_equation", + "content": "d\\boldsymbol{r}_g" + }, + { + "bbox": [ + 55, + 496, + 295, + 556 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 496, + 295, + 556 + ], + "type": "inline_equation", + "content": "ds_g" + }, + { + "bbox": [ + 55, + 496, + 295, + 556 + ], + "type": "text", + "content": " to get the final gaussians in the observation space. In the observation space, we obtain the rendered images through the rasterisation of gaussians and compute the image loss to supervise canonical gaussians." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "content": "Once the warm-up iterations completes, we begin the adaptive joint growth. With the MK calculated during the warm-up phase, we can locate the parent joints " + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "inline_equation", + "content": "J_{s}" + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "content": " that require growth. Then we add grown joints to the extra joint book " + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "inline_equation", + "content": "B^{e}" + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "content": ", optimizing its positions " + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "inline_equation", + "content": "j^{e}" + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "content": " and rotations " + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "inline_equation", + "content": "r^{e}" + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "content": " during the subsequent training. Notably, " + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{lbs}}" + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "content": " needs to extend its output dimensions to include the blending weights for extra joints, following " + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "inline_equation", + "content": "K = K^0 + K^e" + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "content": ". Since the extra joints do not have prior LBS weights, their blending weights are entirely learned through " + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{lbs}}" + }, + { + "bbox": [ + 55, + 557, + 295, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Both the warm-up stage and the post-growth learning stage adopt adaptive density control to manage the updates of the gaussians. We also dynamically adjust the threshold" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "of gradient for densification in [25] based on the number of gaussians, in order to balance memory consumption. Please check our supplemental material for details of this design." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 127, + 395, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 127, + 395, + 140 + ], + "spans": [ + { + "bbox": [ + 314, + 127, + 395, + 140 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 149, + 373, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 149, + 373, + 162 + ], + "spans": [ + { + "bbox": [ + 314, + 149, + 373, + 162 + ], + "type": "text", + "content": "5.1. Dataset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 170, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 300 + ], + "type": "text", + "content": "Our method focuses on hand-held objects and loose-fitting clothing, so we select datasets for experiments accordingly. We notice that the DNA-Rendering dataset [7], by capturing complex scenes of the human body, meets our requirements. Specifically, we selected 8 cases that align with our hypothesis, namely 0041_10, 0090_06, 0176_07, 0800_07 (hand-held objects) and 0007_04, 0014_06, 0051_09, 0811_06 (loose-fitting clothing). For each case, we use 24 surrounding views for training and 6 novel surrounding views for testing. All views are synchronized and contained 100 frames each." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 411 + ], + "type": "text", + "content": "In addition to tackle complicated scenarios, it is essential to ensure the model's performance in typical scenarios involving tight-fitting clothes. Therefore, we additionally test our method on the ZJU-MoCap [46] dataset. Although the tight-clothing cases are too simple to require joint extension, our overall framework still achieves optimal results. Since this part of the experiment is not directly related to adaptive growth, we refer the readers to the supplemental material for further visualizations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 427, + 440, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 427, + 440, + 439 + ], + "spans": [ + { + "bbox": [ + 314, + 427, + 440, + 439 + ], + "type": "text", + "content": "5.2. Baselines and Metrics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 448, + 554, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 554, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 554, + 591 + ], + "type": "text", + "content": "We select the most cutting-edge and representative works from each focus area for a fair comparison. Since 3DGS is currently the leading representation for novel view synthesis, we compare 3DGS-based methods, including 3DGS-Avatar [47], GART [29], and GauHuman [17]. Among them, GART is expected to offer extra animatability for its modeling of implicit global auxiliary bones. Additionally, there is another category of human modeling without incorporating SMPL-like pose priors. Although these methods don't guarantee an animatable human avatar, they can achieve high-quality rendering, among which, we compare the rendering quality of Im4D [33] with our method." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": "We conduct a comprehensive comparison of our ToMiE against these methods. We report three key metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [51], and learned perceptual image patch similarity (LPIPS) [57]. Per-scene results can be found in the supplemental material. In addition to comparing the rendering results, we also demonstrate the animatability of the complicated human parts enabled by our method. We strongly recommend readers to watch the supplemental video for a more intuitive understanding of the animating." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14264" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 72, + 541, + 456 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 541, + 456 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 541, + 456 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 541, + 456 + ], + "type": "image", + "image_path": "5fe11b50405fec08130ae01efdccb76a4b4c12d64eebc1aea431479086230f59.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 459, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 55, + 459, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 555, + 492 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on the DNA-Rendering dataset [7] with animatable baselines. We show two cases of hand-held objects (0800_07) and loose-fitting clothing (0811_06) (from top to bottom). Im4d* [33] achieves high-quality rendering but cannot be animated, and we compare it in the supplemental materials. Please check the supplemental video for better visualization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 503, + 216, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 216, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 216, + 516 + ], + "type": "text", + "content": "5.3. Novel View Synthesis Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "content": "Fig. 4, Tab. 1 and Tab. 2 present the results of our method compared to other baselines. In the tables, we showcase two evaluation protocols. The first evaluates the entire image, reflecting the overall rendering quality. The second uses a binary mask to specifically compare the complicated regions, demonstrating how our method outperforms others in these challenging cases. The mask is shown in Fig. 4, whose details can be found in the supplemental material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 624, + 155, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 155, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 155, + 635 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "content": "A. Adaptive Growth Ablation. We remove the adaptive growth to ablate its impact on the rendering results. Tab. 3 \"w/o Adaptive Growth\" shows a decline in rendering quality, while the hand-held objects and loose-fitting clothing also become not animatable." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "B. Non-rigid Design Ablation. In Sec. 4.4, we apply a" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 503, + 554, + 564 + ], + "blocks": [ + { + "bbox": [ + 317, + 503, + 554, + 564 + ], + "lines": [ + { + "bbox": [ + 317, + 503, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 317, + 503, + 554, + 564 + ], + "type": "table", + "html": "
Dataset Method | MetricAnimatablePSNR(full) ↑DNA-RenderingPSNR(masked) ↑
BGSSIM ↑LPIPS ↓
3DGS-Avatar✓✘28.930.9530.045917.69
GART✓✓29.250.9580.048017.12
Im4D✘✘26.280.9650.030814.90
GauHuman✓✘30.220.9620.040518.26
ToMiE(ours)✓✓31.280.9660.037419.75
", + "image_path": "73eb51f03109980f4d59769194ab313c398e4cad1acfcf5cb1381a94d3a3b42e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 567, + 555, + 661 + ], + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 661 + ], + "type": "text", + "content": "Table 1. Quantitative comparison between our method and other methods on complicated DNA-Rendering dataset [7]. " + }, + { + "bbox": [ + 313, + 567, + 555, + 661 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 313, + 567, + 555, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 567, + 555, + 661 + ], + "type": "inline_equation", + "content": "\\mathbb{G}" + }, + { + "bbox": [ + 313, + 567, + 555, + 661 + ], + "type": "text", + "content": " stand for human body and complex clothing (including handheld objects). We color each result as best, second best and third best. ToMiE achieves the best performance in PSNR and SSIM, while ranking second only to rendering-based Im4D [33] in LPIPS, mainly because rendering-based methods do not consider human structural constraints, resulting in higher visual fidelity." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "content": "non-rigid deformation network " + }, + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "inline_equation", + "content": "\\Phi_d" + }, + { + "bbox": [ + 313, + 677, + 553, + 715 + ], + "type": "text", + "content": " to correct gaussian rotation and scale. As shown in Tab. 3 \"w/o Non-rigid Design\", this significantly improves the final rendering quality." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14265" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 284, + 140 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 284, + 140 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 284, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 284, + 140 + ], + "type": "table", + "html": "
Dataset Method | MetricAnimatableZJU-Mocap
BGPSNR(full) ↑SSIM ↑LPIPS ↓
3DGS-Avatar✓×30.610.9700.0296
GART✓✓30.910.9620.0318
Im4D××28.990.9730.0620
GauHuman✓×30.820.9620.0326
ToMiE(ours)✓✓31.100.9630.0312
", + "image_path": "e6efcaac15714d75a22180e6a5dafeaa08eac5482fb2eb864066e717b5c5f9bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 247, + 294, + 292 + ], + "blocks": [ + { + "bbox": [ + 55, + 144, + 295, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 236 + ], + "type": "text", + "content": "Table 2. Quantitative comparison between our method and other methods on ZJU-Mocap dataset [46] with tight-fitting clothing. " + }, + { + "bbox": [ + 55, + 144, + 295, + 236 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 55, + 144, + 295, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 144, + 295, + 236 + ], + "type": "inline_equation", + "content": "\\mathbb{G}" + }, + { + "bbox": [ + 55, + 144, + 295, + 236 + ], + "type": "text", + "content": " stand for human body and clothing (including hand-held objects). We color each result as best, second best and third best. This comparison demonstrates that ToMiE can achieve comparable (SSIM and LPIPS) or even superior (PSNR) performance to other approaches in tight-fitting scenarios where joint growth is not required." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 247, + 294, + 292 + ], + "lines": [ + { + "bbox": [ + 57, + 247, + 294, + 292 + ], + "spans": [ + { + "bbox": [ + 57, + 247, + 294, + 292 + ], + "type": "table", + "html": "
PSNR(full) ↑SSIM ↑LPIPS ↓PSNR(masked) ↑
w/o Adaptive Growth30.990.9640.039019.21
w/o Non-rigid Design30.780.9640.036319.38
Full31.280.9660.037419.75
", + "image_path": "2d5b83be56019a44ab0da57c3a3fe494448321d4fda6c2ca1e984398d7a25d92.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 294, + 295, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 295, + 327 + ], + "type": "text", + "content": "Table 3. Ablation studies on DNA-Rendering dataset [7]. We independently ablate the adaptive growth strategy and non-rigid design to validate their impact on the overall performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 340, + 295, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 365 + ], + "type": "text", + "content": "5.5. Animating Hand-held Objects and Loose-fitting clothing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 369, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 369, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 295, + 525 + ], + "type": "text", + "content": "We demonstrate the uniqueness of our method, specifically its ability to explicitly animate hand-held objects and loose-fitting clothing. Our animating approach can be implemented in two ways. On the one hand, we can utilize the transformation already recorded in the extra joint book to replay clothing motions. The visualization of this part is shown in Fig. 5. On the other hand, ToMiE also supports bypassing the decoding process of the extra joint book by directly inputting the transformation explicitly. This allows us to customize the motion trajectory of external joints. Due to the limitations of the image's expressive capabilities, we strongly recommend the readers to watch the supplemental video to check the animated results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 525, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 622 + ], + "type": "text", + "content": "In Fig. 5, we edit the extra joints while keeping the body poses under the SMPL paradigm stationary. Since the implicit auxiliary bones of GART [29] is controlled by the traditional SMPL poses, only the identical appearance can be output when SMPL poses are stationary. In contrast, our method explicitly models hand-held objects and loose-fitting clothing, fully decoupling them from the traditional SMPL poses, enabling free animating." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 633, + 212, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 212, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 212, + 645 + ], + "type": "text", + "content": "6. Limitations and Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 714 + ], + "type": "text", + "content": "Limitations. Although our method enhances the modeling for rigid and non-rigid clothing, it cannot address scenarios involving drastic changes in the topology (e.g., taking off clothes or opening a book). This is because topological changes disrupt the one-to-one correspondence be" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 365 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 365 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 365 + ], + "type": "image", + "image_path": "2c9bb521bca23a855504ace1df68a04ec2f22d9f4e028c7a110c35bfefab242a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 368, + 555, + 402 + ], + "lines": [ + { + "bbox": [ + 313, + 368, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 555, + 402 + ], + "type": "text", + "content": "Figure 5. Animating Results of ToMiE. Our explicit modeling fully decouples hand-held objects and loose-fitting clothing from the human body, enabling part-specific animating." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 414, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 485 + ], + "type": "text", + "content": "tween frames, making the human modeling centered on the canonical space become downgraded. We notice that [42] addresses topology issues by introducing high-dimensional mappings, which could be adapted to build our non-rigid deformation. However, this is not the main scope of this paper and can be explored as future work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "content": "Conclusion. In this paper, we introduce ToMiE, an adaptive growth method designed to extend traditional SMPL skeleton for better modeling of hand-held objects and loose-fitting clothing. In the first stage, we assign the gradient of gaussian points to different joints by combining LBS weights with the motion kernel based on motion priors. This allows us to accurately locate the parent joints that need to grow, avoiding redundant growth. In the second stage, we design an extra joint book to achieve explicit joint modeling and optimize the transformation of the newly grown joints in an end-to-end manner. With the improved designs mentioned above, our ToMiE stands out among numerous state-of-the-art methods, achieving the best rendering quality and animatability of hand-held objects and loose-fitting clothing. We hope the adaptive growing method will spark a renewed discussion on the current capabilities of digital human modeling. What's more, it is expected to offer some insights for subsequent works related to topology and skeleton generation." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14266" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This research was supported in part by the Tateisi Science and Technology Foundation, JSPS KAKENHI Grant Numbers 24KK0209 and 22H00529. This research was also supported by Shanghai Artificial Intelligence Laboratory." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 183, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 183, + 296, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 183, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 296, + 216 + ], + "type": "text", + "content": "[1] I. Baran and J. Popović. Automatic rigging and animation of 3d characters. ACM Transactions on graphics (TOG), 26(3):72-es, 2007. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 217, + 296, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 250 + ], + "type": "text", + "content": "[2] H. Bertiche, M. Madadi, and S. Escalera. Neural cloth simulation. ACM Transactions on Graphics (TOG), 41(6):1-14, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 251, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 295, + 285 + ], + "type": "text", + "content": "[3] J. Chen, Y. Zhang, D. Kang, X. Zhe, L. Bao, X. Jia, and H. Lu. Animatable neural radiance fields from monocular rgb videos. arXiv preprint arXiv:2106.13629, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 330 + ], + "type": "text", + "content": "[4] X. Chen, T. Jiang, J. Song, M. Rietmann, A. Geiger, M. J. Black, and O. Hilliges. Fast-snarf: A fast deformer for articulated neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):11796-11809, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 331, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 296, + 386 + ], + "type": "text", + "content": "[5] X. Chen, Y. Zheng, M. J. Black, O. Hilliges, and A. Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11594-11604, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "text", + "content": "[6] Y. Chen, Y. Zhan, Z. Zhong, W. Wang, X. Sun, Y. Qiao, and Y. Zheng. Within the Dynamic Context: Inertia-aware 3D Human Modeling with Pose Sequence. arXiv preprint arXiv:2403.19160, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 432, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 296, + 498 + ], + "type": "text", + "content": "[7] W. Cheng, R. Chen, S. Fan, W. Yin, K. Chen, Z. Cai, J. Wang, Y. Gao, Z. Yu, Z. Lin, et al. DNA-Rendering: A Diverse Neural Actor Repository for High-fidelity Human-centric Rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19982-19993, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 499, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 296, + 555 + ], + "type": "text", + "content": "[8] E. Corona, A. Pumarola, G. Alenya, G. Pons-Moll, and F. Moreno-Noguer. SMPlicit: Topology-aware generative model for clothed people. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11875-11885, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 555, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 296, + 588 + ], + "type": "text", + "content": "[9] J. Dong, Q. Fang, W. Jiang, Y. Yang, H. Bao, and X. Zhou. Fast and robust multi-person 3d pose estimation and tracking from multiple views. In T-PAMI, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 296, + 644 + ], + "type": "text", + "content": "[10] G. Gafni, J. Thies, M. Zollhofer, and M. Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 646, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 296, + 689 + ], + "type": "text", + "content": "[11] Q. Gao, Y. Wang, L. Liu, L. Liu, C. Theobalt, and B. Chen. Neural novel actor: Learning a generalized animatable neural representation for human actors. IEEE Transactions on Visualization and Computer Graphics, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[12] C. Geng, S. Peng, Z. Xu, H. Bao, and X. Zhou. Learning neural volumetric representations of dynamic humans" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "in minutes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8759-8770, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "text", + "content": "[13] S. Goel, G. Pavlakos, J. Rajasegaran, A. Kanazawa, and J. Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783-14794, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 207 + ], + "type": "text", + "content": "[14] C. Guo, T. Jiang, M. Kaufmann, C. Zheng, J. Valentin, J. Song, and O. Hilliges. ReLoo: Reconstructing Humans Dressed in Loose Garments from Monocular Video in the Wild. arXiv preprint arXiv:2409.15269, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 555, + 252 + ], + "type": "text", + "content": "[15] L. Hu. animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 254, + 555, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 555, + 317 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 555, + 317 + ], + "type": "text", + "content": "[16] L. Hu, H. Zhang, Y. Zhang, B. Zhou, B. Liu, S. Zhang, and L. Nie. Gaussianavatar: Towards Realistic Human Avatar Modeling from a Single Video via Animatable 3D Gaussians. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 634-644, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 319, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 319, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 317, + 319, + 555, + 365 + ], + "type": "text", + "content": "[17] S. Hu, T. Hu, and Z. Liu. Gauhuman: Articulated Gaussian Splatting from Monocular Human Videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 365, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 409 + ], + "type": "text", + "content": "[18] T. Hu, F. Hong, and Z. Liu. Surmo: Surface-based 4d motion modeling for dynamic human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6550-6560, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 554, + 453 + ], + "type": "text", + "content": "[19] M. Isik, M. Rünz, M. Georgopoulos, T. Khakhulin, J. Starck, L. Agapito, and M. Nießner. Humanrf: High-fidelity Neural Radiance Fields for Humans in Motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 455, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 488 + ], + "type": "text", + "content": "[20] A. Jacobson, I. Baran, L. Kavan, J. Popovic, and O. Sorkine. Fast automatic skinning transformations. ACM Transactions on Graphics (ToG), 31(4):1-10, 2012. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 489, + 554, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 554, + 533 + ], + "type": "text", + "content": "[21] R. Jena, G. S. Iyer, S. Choudhary, B. Smith, P. Chaudhari, and J. Gee. Splatarmor: Articulated Gaussian Splatting for Animatable Humans from Monocular RGB Videos. arXiv preprint arXiv:2311.10812, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 534, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 554, + 578 + ], + "type": "text", + "content": "[22] Y. Jiang, Z. Shen, Y. Hong, C. Guo, Y. Wu, Y. Zhang, J. Yu, and L. Xu. Robust Dual Gaussian Splatting for Immersive Human-centric Volumetric Videos. ACM Transactions on Graphics (TOG), 43(6):1-15, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 579, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 555, + 634 + ], + "type": "text", + "content": "[23] Y. Jiang, Z. Shen, P. Wang, Z. Su, Y. Hong, Y. Zhang, J. Yu, and L. Xu. Hifi4g: High-fidelity Human Performance Rendering via Compact Gaussian Splatting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19734-19745, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 678 + ], + "type": "text", + "content": "[24] H. Jung, N. Brasch, J. Song, E. Perez-Pellitero, Y. Zhou, Z. Li, N. Navab, and B. Busam. Deformable 3D Gaussian Splitting for Animatable Human Avatars. arXiv preprint arXiv:2312.15059, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[25] B. Kerbl, G. Kopanas, T. Leimkuhler, and G. Drettakis. 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4), 2023. 2, 3, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14267" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[26] M. Kocabas, J.-H. R. Chang, J. Gabriel, O. Tuzel, and A. Ranjan. Hugs: Human Gaussian Splats. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 505-515, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[27] Y. Kwon, D. Kim, D. Ceylan, and H. Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "text", + "content": "[28] B. H. Le and Z. Deng. Two-layer sparse compression of dense-weight blend skinning. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 241 + ], + "type": "text", + "content": "[29] J. Lei, Y. Wang, G. Pavlakos, L. Liu, and K. Daniilidis. Gart: Gaussian Articulated Template Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19876-19887, 2024. 3, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 242, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 294, + 274 + ], + "type": "text", + "content": "[30] M. Li, J. Tao, Z. Yang, and Y. Yang. Human101: Training 100+ fps Human Gaussians in 100s from 1 View. arXiv preprint arXiv:2312.15258, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "text", + "content": "[31] M. Li, S. Yao, Z. Xie, K. Chen, and Y.-G. Jiang. Gaussian-body: Clothed human reconstruction via 3d gaussian splatt-ing. arXiv preprint arXiv:2401.09720, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 310, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 365 + ], + "type": "text", + "content": "[32] Z. Li, Z. Zheng, L. Wang, and Y. Liu. Animatable Gaussians: Learning Pose-dependent Gaussian Maps for High-fidelity Human Avatar Modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19711-19722, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 409 + ], + "type": "text", + "content": "[33] H. Lin, S. Peng, Z. Xu, T. Xie, X. He, H. Bao, and X. Zhou. High-Fidelity and Real-Time Novel View Synthesis for Dynamic Scenes. In SIGGRAPH Asia Conference Proceedings, 2023. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 453 + ], + "type": "text", + "content": "[34] X. Liu, C. Wu, X. Liu, J. Liu, J. Wu, C. Zhao, H. Feng, E. Ding, and J. Wang. Gea: Reconstructing expressive 3d gaussian avatar from monocular video. arXiv preprint arXiv:2402.16607, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "text", + "content": "[35] Y. Liu, X. Huang, M. Qin, Q. Lin, and H. Wang. Animatable 3d gaussian: Fast and high-quality reconstruction of multiple human avatars. arXiv preprint arXiv:2311.16482, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 521 + ], + "type": "text", + "content": "[36] M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black. SMPL: A Skinned Multi-Person Linear Model. ACM Transactions on Graphics, 34(6), 2015. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 294, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 556 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 556 + ], + "type": "text", + "content": "[37] Y. Men, Y. Yao, M. Cui, and L. Bo. Mimo: Controllable character video synthesis with spatial decomposed modeling. arXiv preprint arXiv:2409.16160, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 558, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 294, + 601 + ], + "type": "text", + "content": "[38] B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In European Conference on Computer Vision, pages 405-421, 2020. 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 657 + ], + "type": "text", + "content": "[39] A. Moreau, J. Song, H. Dhamo, R. Shaw, Y. Zhou, and E. Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[40] M. Niu, M. Cao, Y. Zhan, Q. Zhu, M. Ma, J. Zhao, Y. Zeng, Z. Zhong, X. Sun, and Y. Zheng. Anicrafter: Customizing realistic human-centric animation via avatar-background conditioning in video diffusion models. arXiv preprint arXiv:2505.20255, 2025." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 127 + ], + "type": "text", + "content": "[41] H. Pang, H. Zhu, A. Kortylewski, C. Theobalt, and M. Habermann. Ash: Animatable Gaussian Splats for Efficient and Photoreal Human Rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1165-1175, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 129, + 554, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 554, + 182 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 554, + 182 + ], + "type": "text", + "content": "[42] K. Park, U. Sinha, P. Hedman, J. T. Barron, S. Bouaziz, D. B. Goldman, R. Martin-Brualla, and S. M. Seitz. Hypernerf: A Higher-dimensional Representation for Topologically Varying Neural Radiance Fields. arXiv preprint arXiv:2106.13228, 2021. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 184, + 554, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 184, + 554, + 237 + ], + "spans": [ + { + "bbox": [ + 315, + 184, + 554, + 237 + ], + "type": "text", + "content": "[43] D. Paschalidou, A. Katharopoulos, A. Geiger, and S. Fidler. Neural parts: Learning expressive 3d shape abstractions with invertible neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3204-3215, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 239, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 239, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 315, + 239, + 554, + 293 + ], + "type": "text", + "content": "[44] G. Pavlakos, V. Choutas, N. Ghorbani, T. Bolkart, A. A. A. Osman, D. Tzionas, and M. J. Black. Expressive Body Capture: 3D Hands, Face, and Body from a Single Image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 294, + 554, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 294, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 315, + 294, + 554, + 326 + ], + "type": "text", + "content": "[45] S. Peng, J. Dong, Q. Wang, S. Zhang, Q. Shuai, X. Zhou, and H. Bao. Animatable neural radiance fields for modeling dynamic human bodies. In ICCV, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 327, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 327, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 315, + 327, + 554, + 370 + ], + "type": "text", + "content": "[46] S. Peng, Y. Zhang, Y. Xu, Q. Wang, Q. Shuai, H. Bao, and X. Zhou. Neural Body: Implicit Neural Representations with Structured Latent Codes for Novel View Synthesis of Dynamic Humans. In CVPR, 2021. 3, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 372, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 372, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 315, + 372, + 554, + 425 + ], + "type": "text", + "content": "[47] Z. Qian, S. Wang, M. Mihajlovic, A. Geiger, and S. Tang. 3dgs-avatar: Animatable Avatars via Deformable 3d Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5020–5030, 2024. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 426, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 426, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 315, + 426, + 554, + 470 + ], + "type": "text", + "content": "[48] I. Santesteban, M. A. Otaduy, and D. Casas. Snug: self-supervised neural dynamic garments. IEEE. In CVF Conference on Computer Vision and Pattern Recognition (CVPR), volume 2, page 9, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 472, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 472, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 315, + 472, + 554, + 514 + ], + "type": "text", + "content": "[49] Q. Shuai, C. Geng, Q. Fang, S. Peng, W. Shen, X. Zhou, and H. Bao. Novel view synthesis of human interactions from sparse multi-view videos. In SIGGRAPH Conference Proceedings, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 515, + 554, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 515, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 315, + 515, + 554, + 557 + ], + "type": "text", + "content": "[50] S. Wang, K. Schwarz, A. Geiger, and S. Tang. A随机变量的统计学，1987年，第2版，M. H. Lai, The asymptotic analysis of a random variable with a finite distribution, 1987. Springer, 2003." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 559, + 554, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 559, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 315, + 559, + 554, + 602 + ], + "type": "text", + "content": "[51] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image Quality Assessment: from Error Visibility to Structural Similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 604, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 604, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 315, + 604, + 554, + 658 + ], + "type": "text", + "content": "[52] C.-Y. Weng, B. Curless, P. P. Srinivasan, J. T. Barron, and I. Kemelmacher-Shlizerman. Humannerf: Free-viewpoint Rendering of Moving People from Monocular Video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 658, + 554, + 712 + ], + "type": "text", + "content": "[53] Z. Xu, S. Peng, H. Lin, G. He, J. Sun, Y. Shen, H. Bao, and X. Zhou. 4K4D: Real-time 4D View Synthesis at 4K Resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20029-20040, 2024. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14268" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 463 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[54] Z. Xu, Y. Zhou, E. Kalogerakis, C. Landreth, and K. Singh. Rignet: Neural rigging for articulated characters. arXiv preprint arXiv:2005.00559, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "text", + "content": "[55] C.-H. Yao, W.-C. Hung, Y. Li, M. Rubinstein, M.-H. Yang, and V. Jampani. Hi-lassie: High-fidelity articulated shape and skeleton discovery from sparse image ensemble. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4853-4862, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[56] Y. Zhan, W. Xu, Q. Zhu, M. Niu, M. Ma, Y. Liu, Z. Zhong, X. Sun, and Y. Zheng. R3-avatar: Record and retrieve temporal codebook for reconstructing photorealistic human avatars. arXiv preprint arXiv:2503.12751, 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 261 + ], + "type": "text", + "content": "[57] R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 327 + ], + "type": "text", + "content": "[58] S. Zheng, B. Zhou, R. Shao, B. Liu, S. Zhang, L. Nie, and Y. Liu. Gps-gaussian: Generalizable Pixel-wise 3D Gaussian Splatting for Real-time Human Novel View Synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 330, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 330, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 330, + 294, + 373 + ], + "type": "text", + "content": "[59] Y. Zheng, Q. Zhao, G. Yang, W. Yifan, D. Xiang, F. Dubost, D. Lagun, T. Beeler, F. Tombari, L. Guibas, et al. Physavatar: Learning the physics of dressed 3d avatars from visual observations. arXiv preprint arXiv:2404.04421, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 375, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 375, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 375, + 294, + 430 + ], + "type": "text", + "content": "[60] B. Zhou, S. Zheng, H. Tu, R. Shao, B. Liu, S. Zhang, L. Nie, and Y. Liu. GPS-Gaussian+: Generalizable Pixel-wise 3D Gaussian Splatting for Real-Time Human-Scene Rendering from Sparse Views. arXiv preprint arXiv:2411.11363, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 463 + ], + "type": "text", + "content": "[61] W. Zielonka, T. Bagautdinov, S. Saito, M. Zollhöfer, J. Thies, and J. Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14269" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/b24a62fe-1619-470e-918c-9d3158f1e3f4_content_list.json b/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/b24a62fe-1619-470e-918c-9d3158f1e3f4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c948ae862f8f8ef10027b6e627f8d48a70784bab --- /dev/null +++ b/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/b24a62fe-1619-470e-918c-9d3158f1e3f4_content_list.json @@ -0,0 +1,1275 @@ +[ + { + "type": "text", + "text": "Towards Fine-grained Interactive Segmentation in Images and Videos", + "text_level": 1, + "bbox": [ + 148, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuan Yao Qiushi Yang Miaomiao Cui Liefeng Bo", + "bbox": [ + 271, + 181, + 725, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tongyi Lab, Alibaba Group", + "bbox": [ + 387, + 205, + 609, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ryan.yy, yangqiushi.yqs, miaomiao.cmm, liefeng.bo}@alibaba-inc.com", + "bbox": [ + 199, + 224, + 792, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb0427cde5fe37f9e5e300d8ec466acfa54465046eb51e840600f6413185aa1c.jpg", + "image_caption": [ + "Figure 1. Segmentation results of the proposed framework SAM2Refiner and previous methods. The proposed SAM2Refiner achieves the best performance on both image and video segmentation, especially in fine-grained details." + ], + "image_footnote": [], + "bbox": [ + 94, + 250, + 906, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 573, + 326, + 589 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The recent Segment Anything Models (SAMs) have emerged as foundational visual models for general interactive segmentation. Despite demonstrating robust generalization abilities, they still suffer from performance degradations in scenarios that demand accurate masks. Existing methods for high-precision interactive segmentation face a trade-off between perceiving intricate local details and maintaining stable prompting capability, which hinders the applicability and effectiveness of foundational segmentation models. In this paper, we present a SAM2Refiner framework built upon the SAM2 backbone. This architecture allows SAM2 to generate fine-grained segmentation masks for both images and videos while preserving its inherent strengths. Specifically, we design a localization augment module, which incorporates local contextual cues to enhance global features via a cross-attention mechanism, thereby exploiting potential detailed patterns while maintaining semantic information. Moreover, to strengthen the prompting ability toward the enhanced object embeddings, we introduce a prompt", + "bbox": [ + 88, + 613, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "retargeting module that renews the embedding with spatially aligned prompt features. In addition, to obtain accurate high resolution segmentation masks, a mask refinement module is devised by employing a multi-scale cascaded structure to fuse mask features with hierarchical representations from the encoder. Extensive experiments demonstrate the effectiveness of our approach, revealing that the proposed method can produce highly precise masks for both images and videos, surpassing state-of-the-art methods.", + "bbox": [ + 511, + 575, + 906, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 739, + 645, + 756 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Interactive segmentation is a long-standing computer vision task that allows users to segment objects from the guidance of user interaction, typically in the form of clicks, strokes, or bounding boxes. In recent years, the Segment Anything Model (SAM) [16] has emerged as a foundational visual model for general interactive segmentation. Trained on their proposed SA-1B dataset, which consists of more than 1 billion masks and 11 million images, SAM can generate masks in various scenes and has been widely applied", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22509", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to a wide range of downstream vision tasks, including image editing [37, 44], 3D vision [31, 39] and medical image analysis [23, 36]. The following work Segment Anything Model 2 (SAM2) [29] further introduces the streaming memory mechanism to extend the ability of SAM to videos, providing a potential boost to video applications such as AR / VR, robotics and autonomous vehicles.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite demonstrating robust generalization abilities, SAMs encounter the problems in scenario demanding fine-grained segmentation. In particular, as shown in Figure 1, they struggle to predict a complete mask for objects with fine structures or sharp boundaries, leading to overly smooth results or coarse boundaries. Thus, these types of failure may hinder the applicability of SAMs. Valuable efforts [14, 20, 23, 29] have been devoted to solve these challenges by introducing extra input tokens and explicitly fusing the multi-level features, however, they are still limited by SAMs' ability, thus failing to achieve satisfactory fine-grained segmentation.", + "bbox": [ + 89, + 199, + 480, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we aim to start with a holistic analysis on the three key points of fine-grained segmentation toward mining detailed cues and global semantic information guided by the user prompts. Firstly, considering that the input image contains global information of semantics while it lacks the detailed representations on details, we are expected to dynamically harvest the complementary local and global features within the input image. Secondly, SAMs directly combine the prompt with the object features via the decoder, which fail to build strong correlation among prompts with fine-grained cues and are hard to activate the response on object details, producing biased segmentation. Hence, strengthening the correspondences between prompt features and object details is crucial to yield a high response intensity of prompts on local details, which is beneficial to fine-grained segmentation. Thirdly, the object features with low resolution projected by the SAMs' encoder may lose much detailed cues, to recover the missed fine-grained local patterns for yielding high-resolution segmentation maps, the multi-scale features within SAMs' encoder are important to be re-used. A cocurrent research Pi-SAM [20] is the most similar work to ours, which proposes an additional mask decoder to expand the capability of SAM in high-resolution images. Nevertheless, it compromises with diminished responsiveness to visual prompt, necessitating a complementary module for error correction, and exhibit biased preferences for background clicks. Differently, our goal is to facilitate prompt response and consider both local and global presentations for efficient fine-grained segmentation.", + "bbox": [ + 91, + 383, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve the aforementioned goals, we propose the SAM2Refiner architecture, which includes three components for fine-grained interactive segmentation. Specifically, we introduce Localization Augment (LA) to refine the details of global features with local contextual feature", + "bbox": [ + 89, + 825, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cues. We split the input image into four sub-images, and send to SAM2's image encoder together with raw image. A global-local cross-attention mechanism is designed to highlight the local information of the object within the global representation among deep features. Moreover, to improve the response intensity of prompting for segmenting objects with intricate structures, we propose a Prompt Retargeting (PR) strategy by providing precise alignment between the prompts and enhanced object features. PR performs spatial alignment between prompt features and object features, and then incorporates them to yield an enhanced prompt response. In addition, to generate higher resolution masks with finer details, the Mask Refinement (MR) module is devised by incorporating the object features and multi-scale features from the original image. In MR, the intermediate features within the SAM2 encoder is re-used to fully preserve the original zero-shot capabilities. Although designed for image segmentation, the SAM2-based architecture enables native interoperability with video streaming workflows, allowing fine-grained segmentation for both images and videos (as in Figure 1). Extensive experiments on both image and video segmentation benchmarks demonstrate the effectiveness of our approach over state-of-the-art methods.", + "bbox": [ + 511, + 90, + 903, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows.", + "bbox": [ + 529, + 439, + 893, + 454 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SAM2Refiner, a new architecture capable of generating fine-grained segmentation masks with strong perception for both global semantic and local details, and simultaneously maintaining stable prompting ability and powerful zero-shot capability.", + "- In SAM2Refiner, the localization augment module is specifically designed to achieve a holistic representation of input image. By incorporating sub-images as additional inputs, it effectively enhances the global features through improved object localization.", + "- To boost the responsiveness of prompting in intricate regions, we propose a prompt retargeting module that establishes precise alignment between spatially-aware dense prompt features and the enhanced object features.", + "- To produce high-resolution segmentation maps with fine details, the mask refinement module is crafted to integrate multi-scale representations with object features.", + "- Building upon the SAM2 backbone, our architecture exhibits strong compatibility to the video streaming pipeline, exhibiting state-of-the-art fine-grained segmentation on both images and videos benchmarks." + ], + "bbox": [ + 513, + 455, + 903, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 790, + 653, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Interactive Segmentation", + "text_level": 1, + "bbox": [ + 513, + 816, + 741, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interactive segmentation enables users to provide cues for target regions and guide the segmentation process. User interactions can take various forms, including scribbles [14, 16, 28, 35], bounding boxes [14, 16, 20], clicks [16,", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22510", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "45], or language prompts [35, 40]. Traditional methods utilize pixel-level energy minimization methods, capturing low-level appearance features through unary potentials and ensuring consistent segmentation results with pairwise ones [33]. In recent years, many studies leverage prompts as input features and inject them into the models to produce segmentation results [10, 16, 28, 40]. The Segment Anything Model (SAM) [16] pre-trained on a large-scale datasets has emerged as a benchmark for interactive segmentation. Considering that SAM cannot perform semantic prediction, many works [10, 40] utilize category labels to further fine-tune SAM and enable it to achieve the semantic segmentation task. For example, SEEM [40] further trains SAM using labeled segmentation data through a bipartite matching constraint, thus providing the model with the ability to predict semantics. Semantic-SAM [10] proposes a multi-choice learning scheme via multi-task training across diverse datasets, enabling the model to segment at various granularities while predicting semantic labels. Other works [23, 24, 39] apply SAM to diverse specific domains to solve application tasks. MedSAM [23] finetunes SAM by collecting a large-scale dataset of medical image samples in multiple imaging modalities and types of cancer, achieving superior results in disease segmentation. Gaussian-Grouping [24] leverages the SAM as supervision model to jointly train a reconstruction and segmentation of objects in open-world 3D scenes for high-quality 3D edit.", + "bbox": [ + 93, + 90, + 480, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. High-Quality Segmentation", + "text_level": 1, + "bbox": [ + 93, + 513, + 334, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "High-quality segmentation as a fundamental computer vision task focuses on accurately segmenting various complex and detail-rich objects, which ranges from many subtasks including semantic segmentation [22, 47], instance segmentation [9, 12] and panoptic segmentation [15]. Traditional methods [5, 7, 18, 27, 34] on high-quality segmentation dedicate to designing elaborate structures upon on CNN-based networks in order to capture both low-level local patterns and semantic global features through various reception fields towards holistic representation. Deplabv3 [18] designs a set of multi-level convolution kernels to capture features with diverse reception fields and HRNet [34] employs cascaded structures to fuse features from different layers to learn both low-level and high-level information. More recently, many approaches [2, 3, 6, 35, 38] built on the transformer backbone focus on introducing local patterns into the model by narrow the reception field of self-attention module. For example, Mask2Former [6] presents an efficient versatile architecture by using masked attention to extract local features for high-quality image segmentation tasks. Swin-Unet [3] adopts the SwinTransformer network to the UNet like structures to adaptively capture object features with various scales via hierarchical attention. Latest, the Pi-SAM [20] designs a mask", + "bbox": [ + 93, + 539, + 480, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "decoder to expands SAMs ability at high resolution images, however, it may diminish the response of the input prompt and exhibit biased preferences for background clicks.", + "bbox": [ + 516, + 92, + 903, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 516, + 150, + 601, + 164 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose SAM2Refiner, which effectively generates fine-grained segmentation masks with strong perception for both global semantic and local details, and simultaneously maintaining stable prompting ability. In Sec 3.1, we first briefly review the architecture of SAM and SAM2. Then, in Sec 3.2, we introduce the three modules proposed upon SAM2's backbone to effectively improve the precision for segmentation. Finally, in Sec 3.3, we describe the training and inference process of the method, we also depict how our method can be seamlessly inserted into SAM2's video streaming pipeline for zero-shot inference. The core framework of our SAM2Refiner is shown in Figure 2.", + "bbox": [ + 516, + 176, + 903, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries: SAMs", + "text_level": 1, + "bbox": [ + 516, + 364, + 710, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SAM consists of an image encoder, a prompt encoder, and a mask decoder. The image encoder works for image feature extraction with a ViT [1] backbone. The prompt encoder is employed to encode the points, boxes, or masks into prompt tokens. The mask decoder utilizes a two-layer transformer-based decoder that takes both the extracted image embedding with the concatenated output and prompt tokens for final mask prediction.", + "bbox": [ + 516, + 387, + 903, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SAM2 extends SAMs ability to support both images and videos. For image setting, the three modules are largely follow SAM except for two differences, firstly, SAM2 adopt an MAE pretrained Hiera [30] image encoder, which is hierarchical compared with the plane structure in SAM's encoder, and secondly, to solve ambiguity in images without valid object, SAM2 predicts an additional token in mask decoder to indicate whether the object of interest is present. SAM2 introduces additional memory modules to support video streaming, including memory attention, memory encoder, and memory bank. The role of memory attention is to condition the current frame features on the past frame features and predictions as well as on any new prompts. The memory encoder is employed to obtain a memory embedding from masks generated from mask decoder, these memory embeddings are appended to a memory bank in a first-in-first-out (FIFO) way.", + "bbox": [ + 516, + 508, + 903, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Ours: SAM2Refiner", + "text_level": 1, + "bbox": [ + 516, + 773, + 705, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the architecture of SAM2Refiner, which is built upon SAM2's backbone. This design offers two key advantages, 1) the hierarchical image encoder facilitates multi-scale feature capturing, which aids in feature refinement. 2) it allows for flexible integration into the video streaming pipeline. The proposed architecture comprises three modules, Localization Augment (LA)", + "bbox": [ + 516, + 796, + 903, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22511", + "bbox": [ + 478, + 945, + 516, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e42ad3091e8e7a20164e5e90f639d4f01846ef01642c3bfaf3ac1275c3fc2d01.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed SAM2Refiner framework. It contains a localization augment module to balance the localization details and global semantic representations, a prompt retargeting module to enhance responsiveness towards input prompts, and a mask refinement module to boost the quality of mask outcomes. The blue line denotes the SAM2[28] pipeline, and the black line denotes the proposed SAM2Refiner pipeline." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 893, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is inserted into the middle layer of image encoder and mask decoder to refine the details of global features with local contextual feature cues. Prompt Retargeting (PR) aligns the prompt feature with enhanced object embedding to provide flexible prompting. Mask Refinement (MR) further incorporates object features and multi-scale features to generate higher-resolution masks with finer details.", + "bbox": [ + 88, + 530, + 483, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Localization Augment", + "text_level": 1, + "bbox": [ + 89, + 650, + 290, + 665 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We evenly crop the input image $\\mathbf{I} \\in \\mathbb{R}^{B \\times 3 \\times H \\times W}$ into four non-overlapping sub-images $\\{\\mathbf{s}_m\\}_{m=1}^4 \\in \\mathbb{R}^{B \\times 3 \\times \\frac{H}{2} \\times \\frac{W}{2}}$ . To align with input resolution of SAM2's image encoder, we upsample the sub-images to the same resolution with $\\mathbf{I}$ and obtain $\\{\\mathbf{S}_m\\}_{m=1}^4 \\in \\mathbb{R}^{B \\times 3 \\times H \\times W}$ . $\\mathbf{I}$ and $\\{\\mathbf{S}_m\\}_{m=1}^M$ are then fed in batches to the image encoder and generate the highest-level feature maps $\\{F_i\\}_{i=1}^5 \\in \\mathbb{R}^{B \\times C \\times \\frac{H}{16} \\times \\frac{W}{16}}$ . We then split the feature maps into global feature $G = F_1$ and local feature $\\{L_m\\}_{m=1}^4 = \\{F_i\\}_{i=2}^5$ . The local features contain fine-grained details of input images and are used to enhance global feature with detailed clues. The local features are first assembled into a unified concatenated feature $L \\in \\mathbb{R}^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}}$ following their respective positions in the original image. We then apply a multi-granularity average pooling operation as [43] to cap", + "bbox": [ + 89, + 669, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ture important contextual feature cues with different scales. The pooling functions are denote as:", + "bbox": [ + 511, + 530, + 903, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {i} = \\operatorname {A v g P o o l} _ {i} (L), i \\in \\{1, \\dots , n \\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 571, + 903, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $\\mathrm{AvgPool}_i$ is the $i$ -th pooling function, $P_i$ is the $i$ -th feature map. We set $n = 3$ in our experiments, and the receptive fields for each pooling function are set 2, 4, 8 respectively. The feature maps $P_i$ are then tokenized and concatenated into a multi-granularity local features $L_t$ . For global feature $G$ , we tokenized it as $G_t$ . We conduct global-local cross attention between $G_t$ and $L_t$ to obtain the refined global feature $G_{re}$ :", + "bbox": [ + 511, + 597, + 906, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {r e} = \\operatorname {C r o s s A t t n} \\left(G _ {t}, L _ {t}, L _ {t}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 728, + 903, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before sending into attention layer, the positional encoding is added to the features to access critical geometric information. $G_{re}$ is able to highlight the local information about the object within the global representation.", + "bbox": [ + 511, + 753, + 906, + 814 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Prompt Retargeting", + "text_level": 1, + "bbox": [ + 511, + 821, + 700, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We replace the global feature $G$ with $G_{re}$ as the input to the mask decoder. The output mask $M \\in \\mathbb{R}^{B \\times 3 \\times h \\times w}$ and object embedding $E \\in \\mathbb{R}^{B \\times C \\times h \\times w}$ of the mask decoder therefore contain richer representations of the local details.", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22512", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3c87b1bb6e0b674d522c8cb37e40b122c139998cd9023e3651ff593ae555401f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 88, + 890, + 183 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/03f1738bc194a4e26eeabd64570e1cf9b1984e745bbb6d1545d1169b479166ba.jpg", + "image_caption": [ + "Figure 3. A Comparison between SAM2 and SAM2Refiner. The proposed three modules in SAM2Refiner can be seamlessly embedded into SAM2's video streaming pipeline, allowing to support fine-grained interactive segmentation for both images and videos." + ], + "image_footnote": [], + "bbox": [ + 99, + 186, + 890, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance the discriminability of features in small patches, we introduce a modified version of multi-branch convolutional block RFB [21], which applies different kernels to achieve multi-size receptive fields. To be specific, we first employ the bottleneck structure in each branch, consisting of a $1 \\times 1$ convolutional layer to decrease the channel size in object embedding. Secondly, we apply three parallelized blocks with $1 \\times n$ plus a $n \\times 1$ convolutional layer to each branch, where we set $n = 3,5,7$ , respectively. Thirdly, three branches are concatenated and encoded via a $1 \\times 1$ convolutional layer. Subsequently, we obtain the augmented embedding $E_{a}$ by adding the concatenated embedding and a shortcut embedding encoded from another $1 \\times 1$ convolutional layer. Although the ability to perceive local details is significantly enhanced, $E_{a}$ exhibits reduced sensitivity to the input visual prompts. To maintain a stable prompting ability towards the enhanced embedding, we employ a three-channel prompt map to align the prompt feature with the augmented object embedding $E_{a}$ . The map resolution is set as $256 \\times 256$ to reduce the computational cost and only two visual prompts ( clicks and masks) are encoded. We represent clicks as a fixed size radius, where positive clicks are encoded in the first channel, and negative clicks are encoded in the second channel. The mask $M$ is then concatenated into the third channel to form a three-channel dense prompt map $D_{m}$ . We propose a dense prompt encoder to transform the $D_{p}$ to a dense prompt embedding $E_{p}$ . The dense prompt encoder consists of two $2 \\times 2$ convolution layers followed by GELU activations and layer normalization, and one $1 \\times 1$ convolution layer to map the channel dimension to 256. The dense prompt embedding $E_{p}$ is added to $E_{a}$ and obtain $E_{ap}$ in an element-wise manner, followed by a retargeting modules with several attention operations as in SAM2's mask decoder. We re-use the sparse prompt", + "bbox": [ + 91, + 376, + 483, + 891 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "embedding and output token from SAM2's mask decoder to provide positional and class information for the renewing of the object embedding. The retargeting module contains four steps, (1) self-attention on the tokens concatenated from the the sparse prompt embedding and output token; (2) cross-attention from token to $E_{ap}$ ; (3) MLP updated on token, (4) cross-attention from $E_{ap}$ to token. The renewed object embedding allows to possess a higher sensitivity towards the input visual prompts.", + "bbox": [ + 511, + 377, + 906, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Mask Refinement", + "text_level": 1, + "bbox": [ + 511, + 531, + 683, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final mask of SAM2 is directly interpolated from the output prediction of its mask decoder, with a resolution of $256 \\times 256$ . This interpolation process may inevitably damage the details of the output prediction, particularly in high-resolution images. Our mask refinement module is designed to generate higher-resolution masks while enhancing the result quality. We take the object embedding from the prompt retargeting module and the multi-scale features from SAM2's hierarchical features to predict the final mask. The encoder blocks in Figure 3 represent the sub-modules within the image encoder, which generate multi-scale features with resolutions $256 \\times 256$ , $128 \\times 128$ and $64 \\times 64$ , respectively. We incorporate the UNet-like structure to fuse the object features and multi-scale image features. Each of decoder blocks contains an upsampling layer and a convolutional layer, and the upsampling feature along with image feature are concatenated via the channel dimension before sending into the convolutional layer. After the final decoder block, we conduct two transposed convolution layers to obtain mask feature with $1024 \\times 1024$ resolution. The low-level feature generated from input image $\\mathbf{I}$ is added further enhance image quality, and a convolutional layer is then used to generate the final segmentation map.", + "bbox": [ + 511, + 553, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22513", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ab48040f9725d92bd7628ee07ced4047d786b81e359459b9339368a314790ca8.jpg", + "table_caption": [ + "Table 1. Results comparison with state-of-the-art methods on four fine-grained image segmentation tasks." + ], + "table_footnote": [], + "table_body": "
ModelDISCOIFTHRSODThinObjectAverage
mIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoU
SAM (vit_l) [16]55.547.988.483.485.278.466.155.673.866.3
SAM (vit_h) [16]50.643.986.581.780.474.257.649.468.862.3
SAM2 (hiera_b+) [28]57.249.993.288.785.276.366.365.175.570.0
SAM2 (hiera_l) [28]57.650.792.488.086.379.577.867.778.571.5
HQ-SAM (vit_l) [14]77.469.294.489.791.984.687.076.687.780.0
HQ-SAM (vit_h) [14]77.969.794.489.591.983.588.578.688.280.3
SAM2Refiner (hiera_b+)83.478.596.593.394.189.495.190.192.387.8
SAM2Refiner (hiera_l)85.280.696.693.694.790.295.991.093.188.9
", + "bbox": [ + 94, + 119, + 906, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Training and Inference", + "text_level": 1, + "bbox": [ + 89, + 314, + 303, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. We train our SAM2Refiner on image datasets only. We utilize the HQSeg-44K [14] as the training database, which includes six image datasets with extremely fine-grained mask labeling, including DIS [26](train set), ThinObject-5K [19] (train set), FSS-1000 [17], ECSSD [32], MSRA-10K [8], DUT-OMRON [41]. During training, we fix the model parameters of the pre-trained SAM2 model. We train by sampling mixed types of prompt including bounding boxes, randomly sampled points (either positive or negative), and coarse mask input. We use a learning rate of 0.001 and train for 20 epochs, with a learning rate decay schedule. We train on 8 NVIDIA V100 GPUs with a total batch size of 32. If not otherwise specified, we use the hierarchical large model as the default backbone.", + "bbox": [ + 88, + 335, + 480, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss Function. Following HQ-SAM [14], we use a combination of binary cross-entropy (BCE) loss and dice loss for supervision. In addition to calculating the loss at the final prediction, we introduce intermediate supervision at each layer of the three decoder blocks. Before the loss calculation, each of the output features is sent to a separate convolutional layer to align the channel size with ground-truth mask. Our final loss function is:", + "bbox": [ + 88, + 549, + 480, + 670 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {f} \\left(\\mathcal {L} _ {f} ^ {b c e} + \\mathcal {L} _ {f} ^ {d i c e}\\right) + \\lambda_ {i} \\sum_ {i = 1} ^ {3} \\left(\\mathcal {L} _ {i} ^ {b c e} + \\mathcal {L} _ {i} ^ {d i c e}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 680, + 482, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{f}$ and $\\lambda_{i}$ are set to 1.0 and 0.3 in our experiment.", + "bbox": [ + 89, + 733, + 467, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. For image setting, we follow the same inference pipeline as SAM. In the SAM2 temporal streaming pipeline, we extend our SAM2Refiner to the video segmentation by adding the cross-frame interaction via the memory bank. As suggested in Figure 3, compared to the original SAM2 pipeline, SAM2Refiner, including three proposed components, leverages the mask feature memory of previous frames to perform segmentation for the current frame, achieving accurate video segmentation. Note that for each frame except the first frame, the model uses the mem", + "bbox": [ + 88, + 750, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ory feature obtained by the first frame and the past nearest frames to perform frame-by-frame segmentation.", + "bbox": [ + 511, + 314, + 906, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 361, + 648, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 387, + 705, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Following the prior work [14], we evaluate the model ability in both image and video segmentation tasks respectively. In image segmentation, all models are assessed in four challenging fine-grained segmentation benchmarks, including DIS-validation [26], ThinObject5K-test [19], COIFT [19] and HR-SOD [46]. For video object segmentation, we leverage two long-term video segmentation datasets, LVOS-v1 and LVOS-v2 [11], to evaluate our framework.", + "bbox": [ + 511, + 409, + 906, + 545 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate the results of the image segmentation task, we echo previous works [14, 16] and adopt mIoU and mBIoU scores as metrics. For video object segmentation, we utilize $J$ (region similarity), $F$ (contour accuracy), and the combined $J \\& F$ scores as metrics, where evaluations are performed in a semi-supervised setting with the first frame prompt given.", + "bbox": [ + 511, + 547, + 905, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons and Analysis", + "text_level": 1, + "bbox": [ + 511, + 665, + 756, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform extensive comparisons between our SAM2Refiner with SAM [16], SAM2 [28], HQ-SAM [14] and its video variant HQSAM2 on 6 benchmarks across image and video segmentation tasks. We do not compare Pi-SAM as they have not open-sourced their code.", + "bbox": [ + 511, + 688, + 905, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Open-world Image Segmentation. To evaluate the segmentation results in the open-world environment, we compare our SAM2Refiner with previous methods on four image segmentation benchmarks. In Table 1, our SAM2Refiner with large backbone achieves state-of-the-art performance of $85.2\\%$ , $96.6\\%$ , $94.7\\%$ , $95.9\\%$ mIoU scores and $80.6\\%$ , $93.6\\%$ , $90.2\\%$ , $91.0\\%$ mBiOUscores on DIS, COIFT, HRSOD and ThinObject datasets, respectively. In average, SAM2Refiner with a smaller backbone", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22514", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77a26a816ed0cd89a468f98d396fade48a98acd20995cd424c7d424cb0c85c25.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison with previous methods. Given the blue box as visual prompt, our proposed SAM2Refiner produces more accurate results with correct structure and clear boundaries. Zoom in for better visualization." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 261, + 532 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d993c40f9776eaa24539639daac917a607656cd46ce50ac4a2f92dba284cfcd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 88, + 580, + 532 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/58474c52edd750e1c50ea3469ab5d9c3b0944bddf99e27890ee12b060fc48f7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 580, + 88, + 897, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(hiera_1) performs $93.1\\%$ and $88.9\\%$ average mIoU and mBiOU scores, significantly outperforming the second-best one HQ-SAM (vit_h) with $4.9\\%$ and $8.6\\%$ scores. To clear represent the segmentation results, we show the comparison of the visualization results of segmentation maps. From Figure 4, our SAM2Refiner can produce much more accurate segmentation maps than other methods, specifically on detailed objects such as boundary and small regions. These results demonstrate that the SAM2Refiner benefitting from the proposed modules achieves superior fine-grained image segmentation and strong generalization ability.", + "bbox": [ + 89, + 618, + 483, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Zero-Shot Video Object Segmentation. To evaluate the ability of the proposed method on video segmentation, we extend our model on semi-supervised video object segmentation task upon the pipeline of SAM2. Towards the zero-shot segmentation in the open-world scenario, we compare our model with previous video segmentation works in two benchmarks, including LVOS v1 and LVOS v2. As sug-", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6f731a00f27e4cd68b8e500bdc828a3c1b11df40da415b1e18279dd85773ed86.jpg", + "table_caption": [ + "Table 2. Comparison of the proposed method with prior works on video object segmentation task." + ], + "table_footnote": [], + "table_body": "
MethodsLVOS v1LVOS v2
J&FJFJ&FJsFsJuFu
LWL [13]56.451.860.960.658.064.357.262.9
CFBI [13]51.546.256.755.052.959.251.756.2
STCN [13]48.943.954.060.657.264.057.563.8
RDE [25]53.748.359.262.256.764.160.867.2
SwinB-DeAOT-L [42]---63.961.569.058.466.6
XMem [4]52.948.157.764.562.669.160.665.6
SAM2 [28]77.272.282.278.375.287.068.579.4
HQ-SAM2 [14]79.674.984.379.176.588.369.482.2
SAM2Refiner80.975.186.879.978.787.971.681.5
", + "bbox": [ + 514, + 656, + 903, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gested in Table 2, our model yields $80.9\\%$ , $75.1\\%$ , $86.8\\%$ scores of $\\mathcal{J} \\& \\mathcal{F}$ , $\\mathcal{J}$ , $\\mathcal{F}$ and $79.9\\%$ , $78.7\\%$ , $87.9\\%$ , $71.6\\%$ , $81.5\\%$ of $\\mathcal{J} \\& \\mathcal{F}$ , $\\mathcal{J}_s$ , $\\mathcal{F}_s$ , $\\mathcal{J}_u$ , $\\mathcal{F}_u$ in two datasets respectively, significantly outperforming over previous compar", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22515", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1437c48a0ebd0cf3a9c4accc5638be7eeca039f70507c617fcd55088cd6bfd68.jpg", + "table_caption": [ + "Table 3. Analysis of the point prompt response in interactive image segmentation." + ], + "table_footnote": [], + "table_body": "
Point Num.MetricsSAMSAM2HQ-SAMSAM2Refiner
1mIoU8.711.417.725.2
mB IoU6.39.914.320.6
2mIoU30.835.644.652.5
mB IoU26.231.740.247.4
5mIoU70.173.984.290.0
mB IoU62.366.475.584.9
10mIoU71.776.786.191.6
mB IoU64.870.478.588.6
", + "bbox": [ + 94, + 125, + 483, + 258 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8f07ed50e28be5eb1f101a1fcea00eef1749bb457032cdf54e35dd738a01fbda.jpg", + "table_caption": [ + "Table 4. Ablation study on the three modules of SAM2Refiner, including Localization Augment (LA), Prompt Retargeting (PR) and Mask Refinement (MR). The results are averaged over the four evaluated datasets." + ], + "table_footnote": [], + "table_body": "
LAPRMRmIoUmBIoU
SAM2 [29]78.571.5
SAM2Refiner✓85.067.4
✓✓91.186.0
✓✓92.788.3
✓✓91.486.2
✓✓✓93.188.9
", + "bbox": [ + 112, + 335, + 460, + 457 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "isons [4, 13, 14, 25, 28, 42]. These results demonstrate that the proposed framework can not only enhance fine-grained image segmentation, but also boost the video segmentation with refined object details.", + "bbox": [ + 89, + 474, + 482, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on prompt number. To evaluate the response intensity to input prompts in SAM2Refiner compared to other models, we input varying numbers of point prompts into the model in image segmentation task. In Table 3, we observe that the average segmentation results decline as the number of prompt points decreases. However, our SAM2Refiner indicates consistent advantages in segmentation accuracy across varying numbers of point prompts, while SAM, SAM2 and HQ-SAM reflect relatively low results. This clearly suggests that the proposed method with the capability of prompt re-targeting can effectively achieve a strong response for input prompts.", + "bbox": [ + 88, + 535, + 482, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 726, + 243, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct ablation experiments to evaluate the effectiveness of the proposed modules within our SAM2Refiner, i.e., Localization Augment (LA), Prompt Retargeting (PR) and Mask Refinement (MR), and we study the affect of the number of input prompts.", + "bbox": [ + 89, + 750, + 482, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the LA module, the original global features attend to local features from sub-images, thereby enriching the representation with more detailed local information. As shown in Table 4, compared to SAM2, the LA module markedly enhances the mIoU from $78.5\\%$ to $85.0\\%$ , despite a slight", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/457a3a98be44b41efd4869141b4c86174f222a3b37f6c3ce459e6c2702888403.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 645, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a041b8ec8f7ba3dc3e628221dcaad5bacb7920b9e1c0684f3205a7994d3ebf38.jpg", + "image_caption": [ + "w/o LA" + ], + "image_footnote": [], + "bbox": [ + 645, + 89, + 771, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/72f001d2ceb68d11a35995e42c5618413045b845107cfac188be7149c9aa88a4.jpg", + "image_caption": [ + "w/LA" + ], + "image_footnote": [], + "bbox": [ + 771, + 89, + 898, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1174d49dee81c4aff286576d051bc763d1ae850a2d66c6895b4c7954379251d6.jpg", + "image_caption": [ + "Input", + "Figure 6. Effectiveness of PR module. PR allows precise response for point prompts in ambiguous regions. Green point and red point denote positive and negative prompt, respectively." + ], + "image_footnote": [], + "bbox": [ + 517, + 252, + 645, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8328686e622b037acc63c011fc90bdf7a80a7cbf9f419e79491725c86cf64a51.jpg", + "image_caption": [ + "w/o PR" + ], + "image_footnote": [], + "bbox": [ + 645, + 252, + 771, + 318 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/431ac431aa31c1ddcab1f5ac8e2eb2788c314d67ee81d050d1cd974b07d8d4d4.jpg", + "image_caption": [ + "Figure 5. Effectiveness of LA module. LA suggest remarkable improvement on local details although it displays some flickering boundaries.", + "W/PR" + ], + "image_footnote": [], + "bbox": [ + 771, + 252, + 898, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "decrease in boundary accuracy. This is because the LA module strengthens the perception of fine structures, but this can come at the cost of boundary precision. Figure 5 verifies this phenomenon, showing that results with LA exhibit more detailed local features but also show flickering boundaries. The second to fourth rows of the SAM2Refiner in Table 4 highlight the effectiveness of each individual module for fine-grained segmentation. The PR module allows the model to maintain strong sensibility towards point prompting in intricate regions. As shown in Figure 6, result with PR module is able to correctly respond to the input visual prompt. Compared to PR, both the LA and MR modules deliver greater performance improvements in mIoU and mBIoU, with about 2 points gain in mIoU and 3 points gain in mBIoU, underscoring their crucial roles in enhancing segmentation precision.", + "bbox": [ + 511, + 414, + 906, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 669, + 633, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose SAM2Refiner to enhance interactive segmentation for images and videos. SAM2Refiner balances local detail perception and prompting stability, achieving fine-grained visual representation. By integrating local contextual cues through a cross-attention mechanism and employing a UNet-like structure for feature fusion, SAM2Refiner generates high-quality segmentation masks. In addition, a prompt retargeting module is designed to improve the alignment of object embeddings with spatial features. Extensive experiments demonstrate that our SAM2Refiner achieves superior segmentation performance on both image and video segmentation tasks, significantly outperforming other state-of-the-art methods.", + "bbox": [ + 511, + 695, + 906, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22516", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020.", + "[2] Alexander Kirillov Bowen Cheng, Alex Schwing. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021.", + "[3] Wang Y. Chen J. Jiang D. Zhang X. Tian Q. Cao, H. and Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In ECCV, 2022.", + "[4] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022.", + "[5] Ho Kei Cheng, Jihoon Chung, Yu-Wing Tai, and Chi-Keung Tang. CascadePSP: Toward class-agnostic and very high-resolution segmentation via global and local refinement. In CVPR, 2020.", + "[6] Ishan Misra Alexander G. Schwing Alexander Kirillov Cheng, Bowen and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022.", + "[7] Maxwell D. Collins Yukun Zhu Ting Liu Thomas S. Huang Hartwig Adam Cheng, Bowen and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020.", + "[8] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. TPAMI, 37(3):569-582, 2014.", + "[9] Jifeng Dai, Kaiming He, and Jian Sun. Instance-aware semantic segmentation via multi-task network cascades. In CVPR, 2016.", + "[10] Peize Sun Xueyan Zou Shilong Liu Chunyuan Li Jianwei Yang Lei Zhang Jianfeng Gao Feng Li, Hao Zhang. Segment and recognize anything at any granularity. In ECCV, 2024.", + "[11] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019.", + "[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017.", + "[13] Yu-Wing Tai Ho Kei Cheng and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021.", + "[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. In NeurIPS, 2023.", + "[15] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019.", + "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dolkar, and Ross Girshick. Segment anything. In ICCV, 2023.", + "[17] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In CVPR, 2020.", + "[18] Florian Schroff Hartwig Adam Liang-Chieh Chen, George Papandreou. Rethinking atrous convolution" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for semantic image segmentation. In arXiv preprint arXiv:1706.05587, 2017.", + "[19] Jun Hao Liew, Scott Cohen, Brian Price, Long Mai, and Jiashi Feng. Deep interactive thin object selection. In WACV, 2021.", + "[20] Mengzhen Liu, Mengyu Wang, Henghui Ding, Yilong Xu, Yao Zhao, and Yunchao Wei. Segment anything with precise interaction. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3790-3799, 2024.", + "[21] Songtao Liu, Di Huang, et al. Receptive field block net for accurate and fast object detection. In Proceedings of the European conference on computer vision (ECCV), pages 385-400, 2018.", + "[22] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015.", + "[23] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024.", + "[24] Fisher Yu Mingqiao Ye, Martin Danelljan and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In ECCV, 2024.", + "[25] Zhiwei Xiong Bang Zhang Pan Pan Mingxing Li, Liucheng Hu and Dong Liu. Recurrent dynamic embedding for video object segmentation. In CVPR, 2022.", + "[26] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In ECCV, 2022.", + "[27] Yixuan Yuan Qiushi Yang. Learning dynamic convolutions for multi-modal 3d mri brain tumor segmentation. In MIC-CAI Workshop, 2020.", + "[28] Gabeur V. Hu Y.T. Hu R. Ryali C. Ma T. Khedr H. Radle R. Rolland C. Gustafson L. Ravi, N. and E. Mintun. Sam2: Segment anything in images and videos. In arXiv preprint arXiv:2408.00714, 2024.", + "[29] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024.", + "[30] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023.", + "[31] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023.", + "[32] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended.cssd. TPAMI, 38(4): 717-729, 2015.", + "[33] Mengyu Wang, Henghui Ding, Jun Hao Liew, Jiajun Liu, Yao Zhao, and Yunchao Wei. Segrefiner: Towards model-agnostic segmentation refinement with discrete diffusion process. arXiv preprint arXiv:2312.12425, 2023." + ], + "bbox": [ + 516, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22517", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Sun K. Cheng T. Jiang B. Deng C. Zhao Y. Liu D. Mu Y. Tan M. Wang X. Wang, J. and W. Liu. Deep high-resolution representation learning for visual recognition. In TPAMI, 2020.", + "[35] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023.", + "[36] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023.", + "[37] Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023.", + "[38] Wenhai Wang Zhiding Yu Anima Anandkumar Jose M. Alvarez Xie, Enze and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021.", + "[39] Xiuwei Xu, Huangxing Chen, Linqing Zhao, Ziwei Wang, Jie Zhou, and Jiwen Lu. Embodiedsam: Online segment any 3d thing in real time. arXiv preprint arXiv:2408.11811, 2024.", + "[40] Hao Zhang Feng Li Linjie Li Jianfeng Wang Lijuan Wang Jianfeng Gao Yong Jae Lee Xueyan Zou, Jianwei Yang. Segment everything everywhere all at once. In NeurlPS, 2023.", + "[41] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013.", + "[42] Zongxin Yang and Yi Yang. Decoupling features in hierarchical propagation for video object segmentation. In NeurIPS, 2022.", + "[43] Qian Yu, Xiaqi Zhao, Youwei Pang, Lihe Zhang, and Huchuan Lu. Multi-view aggregation network for dichotomous image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3921-3930, 2024.", + "[44] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023.", + "[45] Yuhui Yuan, Jingyi Xie, Xilin Chen, and Jingdong Wang. Segfix: Model-agnostic boundary refinement for segmentation. In ECCV, 2020.", + "[46] Yi Zeng, Pingping Zhang, Jianming Zhang, Zhe Lin, and Huchuan Lu. Towards high-resolution salient object detection. In ICCV, 2019.", + "[47] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017." + ], + "bbox": [ + 91, + 90, + 480, + 785 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22518", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/b24a62fe-1619-470e-918c-9d3158f1e3f4_model.json b/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/b24a62fe-1619-470e-918c-9d3158f1e3f4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0c96e2461d63db983ed4786710bb987762b96986 --- /dev/null +++ b/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/b24a62fe-1619-470e-918c-9d3158f1e3f4_model.json @@ -0,0 +1,1903 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.131, + 0.851, + 0.154 + ], + "angle": 0, + "content": "Towards Fine-grained Interactive Segmentation in Images and Videos" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.182, + 0.726, + 0.201 + ], + "angle": 0, + "content": "Yuan Yao Qiushi Yang Miaomiao Cui Liefeng Bo" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.206, + 0.61, + 0.224 + ], + "angle": 0, + "content": "Tongyi Lab, Alibaba Group" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.226, + 0.794, + 0.24 + ], + "angle": 0, + "content": "{ryan.yy, yangqiushi.yqs, miaomiao.cmm, liefeng.bo}@alibaba-inc.com" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.251, + 0.907, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.521, + 0.908, + 0.549 + ], + "angle": 0, + "content": "Figure 1. Segmentation results of the proposed framework SAM2Refiner and previous methods. The proposed SAM2Refiner achieves the best performance on both image and video segmentation, especially in fine-grained details." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.574, + 0.327, + 0.59 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The recent Segment Anything Models (SAMs) have emerged as foundational visual models for general interactive segmentation. Despite demonstrating robust generalization abilities, they still suffer from performance degradations in scenarios that demand accurate masks. Existing methods for high-precision interactive segmentation face a trade-off between perceiving intricate local details and maintaining stable prompting capability, which hinders the applicability and effectiveness of foundational segmentation models. In this paper, we present a SAM2Refiner framework built upon the SAM2 backbone. This architecture allows SAM2 to generate fine-grained segmentation masks for both images and videos while preserving its inherent strengths. Specifically, we design a localization augment module, which incorporates local contextual cues to enhance global features via a cross-attention mechanism, thereby exploiting potential detailed patterns while maintaining semantic information. Moreover, to strengthen the prompting ability toward the enhanced object embeddings, we introduce a prompt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.907, + 0.712 + ], + "angle": 0, + "content": "retargeting module that renews the embedding with spatially aligned prompt features. In addition, to obtain accurate high resolution segmentation masks, a mask refinement module is devised by employing a multi-scale cascaded structure to fuse mask features with hierarchical representations from the encoder. Extensive experiments demonstrate the effectiveness of our approach, revealing that the proposed method can produce highly precise masks for both images and videos, surpassing state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.741, + 0.646, + 0.757 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Interactive segmentation is a long-standing computer vision task that allows users to segment objects from the guidance of user interaction, typically in the form of clicks, strokes, or bounding boxes. In recent years, the Segment Anything Model (SAM) [16] has emerged as a foundational visual model for general interactive segmentation. Trained on their proposed SA-1B dataset, which consists of more than 1 billion masks and 11 million images, SAM can generate masks in various scenes and has been widely applied" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "22509" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "to a wide range of downstream vision tasks, including image editing [37, 44], 3D vision [31, 39] and medical image analysis [23, 36]. The following work Segment Anything Model 2 (SAM2) [29] further introduces the streaming memory mechanism to extend the ability of SAM to videos, providing a potential boost to video applications such as AR / VR, robotics and autonomous vehicles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.482, + 0.382 + ], + "angle": 0, + "content": "Despite demonstrating robust generalization abilities, SAMs encounter the problems in scenario demanding fine-grained segmentation. In particular, as shown in Figure 1, they struggle to predict a complete mask for objects with fine structures or sharp boundaries, leading to overly smooth results or coarse boundaries. Thus, these types of failure may hinder the applicability of SAMs. Valuable efforts [14, 20, 23, 29] have been devoted to solve these challenges by introducing extra input tokens and explicitly fusing the multi-level features, however, they are still limited by SAMs' ability, thus failing to achieve satisfactory fine-grained segmentation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.385, + 0.483, + 0.822 + ], + "angle": 0, + "content": "To this end, we aim to start with a holistic analysis on the three key points of fine-grained segmentation toward mining detailed cues and global semantic information guided by the user prompts. Firstly, considering that the input image contains global information of semantics while it lacks the detailed representations on details, we are expected to dynamically harvest the complementary local and global features within the input image. Secondly, SAMs directly combine the prompt with the object features via the decoder, which fail to build strong correlation among prompts with fine-grained cues and are hard to activate the response on object details, producing biased segmentation. Hence, strengthening the correspondences between prompt features and object details is crucial to yield a high response intensity of prompts on local details, which is beneficial to fine-grained segmentation. Thirdly, the object features with low resolution projected by the SAMs' encoder may lose much detailed cues, to recover the missed fine-grained local patterns for yielding high-resolution segmentation maps, the multi-scale features within SAMs' encoder are important to be re-used. A cocurrent research Pi-SAM [20] is the most similar work to ours, which proposes an additional mask decoder to expand the capability of SAM in high-resolution images. Nevertheless, it compromises with diminished responsiveness to visual prompt, necessitating a complementary module for error correction, and exhibit biased preferences for background clicks. Differently, our goal is to facilitate prompt response and consider both local and global presentations for efficient fine-grained segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To achieve the aforementioned goals, we propose the SAM2Refiner architecture, which includes three components for fine-grained interactive segmentation. Specifically, we introduce Localization Augment (LA) to refine the details of global features with local contextual feature" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.439 + ], + "angle": 0, + "content": "cues. We split the input image into four sub-images, and send to SAM2's image encoder together with raw image. A global-local cross-attention mechanism is designed to highlight the local information of the object within the global representation among deep features. Moreover, to improve the response intensity of prompting for segmenting objects with intricate structures, we propose a Prompt Retargeting (PR) strategy by providing precise alignment between the prompts and enhanced object features. PR performs spatial alignment between prompt features and object features, and then incorporates them to yield an enhanced prompt response. In addition, to generate higher resolution masks with finer details, the Mask Refinement (MR) module is devised by incorporating the object features and multi-scale features from the original image. In MR, the intermediate features within the SAM2 encoder is re-used to fully preserve the original zero-shot capabilities. Although designed for image segmentation, the SAM2-based architecture enables native interoperability with video streaming workflows, allowing fine-grained segmentation for both images and videos (as in Figure 1). Extensive experiments on both image and video segmentation benchmarks demonstrate the effectiveness of our approach over state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.44, + 0.895, + 0.455 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.457, + 0.905, + 0.532 + ], + "angle": 0, + "content": "- We propose SAM2Refiner, a new architecture capable of generating fine-grained segmentation masks with strong perception for both global semantic and local details, and simultaneously maintaining stable prompting ability and powerful zero-shot capability." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.533, + 0.905, + 0.607 + ], + "angle": 0, + "content": "- In SAM2Refiner, the localization augment module is specifically designed to achieve a holistic representation of input image. By incorporating sub-images as additional inputs, it effectively enhances the global features through improved object localization." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.608, + 0.905, + 0.668 + ], + "angle": 0, + "content": "- To boost the responsiveness of prompting in intricate regions, we propose a prompt retargeting module that establishes precise alignment between spatially-aware dense prompt features and the enhanced object features." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.668, + 0.905, + 0.713 + ], + "angle": 0, + "content": "- To produce high-resolution segmentation maps with fine details, the mask refinement module is crafted to integrate multi-scale representations with object features." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.714, + 0.905, + 0.773 + ], + "angle": 0, + "content": "- Building upon the SAM2 backbone, our architecture exhibits strong compatibility to the video streaming pipeline, exhibiting state-of-the-art fine-grained segmentation on both images and videos benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.457, + 0.905, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.791, + 0.655, + 0.807 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.817, + 0.743, + 0.833 + ], + "angle": 0, + "content": "2.1. Interactive Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Interactive segmentation enables users to provide cues for target regions and guide the segmentation process. User interactions can take various forms, including scribbles [14, 16, 28, 35], bounding boxes [14, 16, 20], clicks [16," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22510" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.5 + ], + "angle": 0, + "content": "45], or language prompts [35, 40]. Traditional methods utilize pixel-level energy minimization methods, capturing low-level appearance features through unary potentials and ensuring consistent segmentation results with pairwise ones [33]. In recent years, many studies leverage prompts as input features and inject them into the models to produce segmentation results [10, 16, 28, 40]. The Segment Anything Model (SAM) [16] pre-trained on a large-scale datasets has emerged as a benchmark for interactive segmentation. Considering that SAM cannot perform semantic prediction, many works [10, 40] utilize category labels to further fine-tune SAM and enable it to achieve the semantic segmentation task. For example, SEEM [40] further trains SAM using labeled segmentation data through a bipartite matching constraint, thus providing the model with the ability to predict semantics. Semantic-SAM [10] proposes a multi-choice learning scheme via multi-task training across diverse datasets, enabling the model to segment at various granularities while predicting semantic labels. Other works [23, 24, 39] apply SAM to diverse specific domains to solve application tasks. MedSAM [23] finetunes SAM by collecting a large-scale dataset of medical image samples in multiple imaging modalities and types of cancer, achieving superior results in disease segmentation. Gaussian-Grouping [24] leverages the SAM as supervision model to jointly train a reconstruction and segmentation of objects in open-world 3D scenes for high-quality 3D edit." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.515, + 0.336, + 0.531 + ], + "angle": 0, + "content": "2.2. High-Quality Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.54, + 0.482, + 0.9 + ], + "angle": 0, + "content": "High-quality segmentation as a fundamental computer vision task focuses on accurately segmenting various complex and detail-rich objects, which ranges from many subtasks including semantic segmentation [22, 47], instance segmentation [9, 12] and panoptic segmentation [15]. Traditional methods [5, 7, 18, 27, 34] on high-quality segmentation dedicate to designing elaborate structures upon on CNN-based networks in order to capture both low-level local patterns and semantic global features through various reception fields towards holistic representation. Deplabv3 [18] designs a set of multi-level convolution kernels to capture features with diverse reception fields and HRNet [34] employs cascaded structures to fuse features from different layers to learn both low-level and high-level information. More recently, many approaches [2, 3, 6, 35, 38] built on the transformer backbone focus on introducing local patterns into the model by narrow the reception field of self-attention module. For example, Mask2Former [6] presents an efficient versatile architecture by using masked attention to extract local features for high-quality image segmentation tasks. Swin-Unet [3] adopts the SwinTransformer network to the UNet like structures to adaptively capture object features with various scales via hierarchical attention. Latest, the Pi-SAM [20] designs a mask" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.136 + ], + "angle": 0, + "content": "decoder to expands SAMs ability at high resolution images, however, it may diminish the response of the input prompt and exhibit biased preferences for background clicks." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.151, + 0.602, + 0.165 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.356 + ], + "angle": 0, + "content": "We propose SAM2Refiner, which effectively generates fine-grained segmentation masks with strong perception for both global semantic and local details, and simultaneously maintaining stable prompting ability. In Sec 3.1, we first briefly review the architecture of SAM and SAM2. Then, in Sec 3.2, we introduce the three modules proposed upon SAM2's backbone to effectively improve the precision for segmentation. Finally, in Sec 3.3, we describe the training and inference process of the method, we also depict how our method can be seamlessly inserted into SAM2's video streaming pipeline for zero-shot inference. The core framework of our SAM2Refiner is shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.366, + 0.712, + 0.38 + ], + "angle": 0, + "content": "3.1. Preliminaries: SAMs" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.388, + 0.905, + 0.508 + ], + "angle": 0, + "content": "SAM consists of an image encoder, a prompt encoder, and a mask decoder. The image encoder works for image feature extraction with a ViT [1] backbone. The prompt encoder is employed to encode the points, boxes, or masks into prompt tokens. The mask decoder utilizes a two-layer transformer-based decoder that takes both the extracted image embedding with the concatenated output and prompt tokens for final mask prediction." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.509, + 0.905, + 0.765 + ], + "angle": 0, + "content": "SAM2 extends SAMs ability to support both images and videos. For image setting, the three modules are largely follow SAM except for two differences, firstly, SAM2 adopt an MAE pretrained Hiera [30] image encoder, which is hierarchical compared with the plane structure in SAM's encoder, and secondly, to solve ambiguity in images without valid object, SAM2 predicts an additional token in mask decoder to indicate whether the object of interest is present. SAM2 introduces additional memory modules to support video streaming, including memory attention, memory encoder, and memory bank. The role of memory attention is to condition the current frame features on the past frame features and predictions as well as on any new prompts. The memory encoder is employed to obtain a memory embedding from masks generated from mask decoder, these memory embeddings are appended to a memory bank in a first-in-first-out (FIFO) way." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.774, + 0.706, + 0.788 + ], + "angle": 0, + "content": "3.2. Ours: SAM2Refiner" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.797, + 0.905, + 0.9 + ], + "angle": 0, + "content": "In this section, we describe the architecture of SAM2Refiner, which is built upon SAM2's backbone. This design offers two key advantages, 1) the hierarchical image encoder facilitates multi-scale feature capturing, which aids in feature refinement. 2) it allows for flexible integration into the video streaming pipeline. The proposed architecture comprises three modules, Localization Augment (LA)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "22511" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.895, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.446, + 0.908, + 0.504 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed SAM2Refiner framework. It contains a localization augment module to balance the localization details and global semantic representations, a prompt retargeting module to enhance responsiveness towards input prompts, and a mask refinement module to boost the quality of mask outcomes. The blue line denotes the SAM2[28] pipeline, and the black line denotes the proposed SAM2Refiner pipeline." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.531, + 0.484, + 0.639 + ], + "angle": 0, + "content": "is inserted into the middle layer of image encoder and mask decoder to refine the details of global features with local contextual feature cues. Prompt Retargeting (PR) aligns the prompt feature with enhanced object embedding to provide flexible prompting. Mask Refinement (MR) further incorporates object features and multi-scale features to generate higher-resolution masks with finer details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.651, + 0.292, + 0.666 + ], + "angle": 0, + "content": "3.2.1. Localization Augment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We evenly crop the input image \\(\\mathbf{I} \\in \\mathbb{R}^{B \\times 3 \\times H \\times W}\\) into four non-overlapping sub-images \\(\\{\\mathbf{s}_m\\}_{m=1}^4 \\in \\mathbb{R}^{B \\times 3 \\times \\frac{H}{2} \\times \\frac{W}{2}}\\). To align with input resolution of SAM2's image encoder, we upsample the sub-images to the same resolution with \\(\\mathbf{I}\\) and obtain \\(\\{\\mathbf{S}_m\\}_{m=1}^4 \\in \\mathbb{R}^{B \\times 3 \\times H \\times W}\\). \\(\\mathbf{I}\\) and \\(\\{\\mathbf{S}_m\\}_{m=1}^M\\) are then fed in batches to the image encoder and generate the highest-level feature maps \\(\\{F_i\\}_{i=1}^5 \\in \\mathbb{R}^{B \\times C \\times \\frac{H}{16} \\times \\frac{W}{16}}\\). We then split the feature maps into global feature \\(G = F_1\\) and local feature \\(\\{L_m\\}_{m=1}^4 = \\{F_i\\}_{i=2}^5\\). The local features contain fine-grained details of input images and are used to enhance global feature with detailed clues. The local features are first assembled into a unified concatenated feature \\(L \\in \\mathbb{R}^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}}\\) following their respective positions in the original image. We then apply a multi-granularity average pooling operation as [43] to cap" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.531, + 0.905, + 0.562 + ], + "angle": 0, + "content": "ture important contextual feature cues with different scales. The pooling functions are denote as:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.572, + 0.905, + 0.589 + ], + "angle": 0, + "content": "\\[\nP _ {i} = \\operatorname {A v g P o o l} _ {i} (L), i \\in \\{1, \\dots , n \\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Where \\(\\mathrm{AvgPool}_i\\) is the \\(i\\)-th pooling function, \\(P_i\\) is the \\(i\\)-th feature map. We set \\(n = 3\\) in our experiments, and the receptive fields for each pooling function are set 2, 4, 8 respectively. The feature maps \\(P_i\\) are then tokenized and concatenated into a multi-granularity local features \\(L_t\\). For global feature \\(G\\), we tokenized it as \\(G_t\\). We conduct global-local cross attention between \\(G_t\\) and \\(L_t\\) to obtain the refined global feature \\(G_{re}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.729, + 0.905, + 0.745 + ], + "angle": 0, + "content": "\\[\nG _ {r e} = \\operatorname {C r o s s A t t n} \\left(G _ {t}, L _ {t}, L _ {t}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.908, + 0.815 + ], + "angle": 0, + "content": "Before sending into attention layer, the positional encoding is added to the features to access critical geometric information. \\( G_{re} \\) is able to highlight the local information about the object within the global representation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.822, + 0.702, + 0.837 + ], + "angle": 0, + "content": "3.2.2. Prompt Retargeting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.901 + ], + "angle": 0, + "content": "We replace the global feature \\(G\\) with \\(G_{re}\\) as the input to the mask decoder. The output mask \\(M \\in \\mathbb{R}^{B \\times 3 \\times h \\times w}\\) and object embedding \\(E \\in \\mathbb{R}^{B \\times C \\times h \\times w}\\) of the mask decoder therefore contain richer representations of the local details." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22512" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.891, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.187, + 0.891, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.31, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 3. A Comparison between SAM2 and SAM2Refiner. The proposed three modules in SAM2Refiner can be seamlessly embedded into SAM2's video streaming pipeline, allowing to support fine-grained interactive segmentation for both images and videos." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.377, + 0.485, + 0.892 + ], + "angle": 0, + "content": "To enhance the discriminability of features in small patches, we introduce a modified version of multi-branch convolutional block RFB [21], which applies different kernels to achieve multi-size receptive fields. To be specific, we first employ the bottleneck structure in each branch, consisting of a \\( 1 \\times 1 \\) convolutional layer to decrease the channel size in object embedding. Secondly, we apply three parallelized blocks with \\( 1 \\times n \\) plus a \\( n \\times 1 \\) convolutional layer to each branch, where we set \\( n = 3,5,7 \\), respectively. Thirdly, three branches are concatenated and encoded via a \\( 1 \\times 1 \\) convolutional layer. Subsequently, we obtain the augmented embedding \\( E_{a} \\) by adding the concatenated embedding and a shortcut embedding encoded from another \\( 1 \\times 1 \\) convolutional layer. Although the ability to perceive local details is significantly enhanced, \\( E_{a} \\) exhibits reduced sensitivity to the input visual prompts. To maintain a stable prompting ability towards the enhanced embedding, we employ a three-channel prompt map to align the prompt feature with the augmented object embedding \\( E_{a} \\). The map resolution is set as \\( 256 \\times 256 \\) to reduce the computational cost and only two visual prompts ( clicks and masks) are encoded. We represent clicks as a fixed size radius, where positive clicks are encoded in the first channel, and negative clicks are encoded in the second channel. The mask \\( M \\) is then concatenated into the third channel to form a three-channel dense prompt map \\( D_{m} \\). We propose a dense prompt encoder to transform the \\( D_{p} \\) to a dense prompt embedding \\( E_{p} \\). The dense prompt encoder consists of two \\( 2 \\times 2 \\) convolution layers followed by GELU activations and layer normalization, and one \\( 1 \\times 1 \\) convolution layer to map the channel dimension to 256. The dense prompt embedding \\( E_{p} \\) is added to \\( E_{a} \\) and obtain \\( E_{ap} \\) in an element-wise manner, followed by a retargeting modules with several attention operations as in SAM2's mask decoder. We re-use the sparse prompt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.908, + 0.514 + ], + "angle": 0, + "content": "embedding and output token from SAM2's mask decoder to provide positional and class information for the renewing of the object embedding. The retargeting module contains four steps, (1) self-attention on the tokens concatenated from the the sparse prompt embedding and output token; (2) cross-attention from token to \\( E_{ap} \\); (3) MLP updated on token, (4) cross-attention from \\( E_{ap} \\) to token. The renewed object embedding allows to possess a higher sensitivity towards the input visual prompts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.684, + 0.545 + ], + "angle": 0, + "content": "3.2.3. Mask Refinement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The final mask of SAM2 is directly interpolated from the output prediction of its mask decoder, with a resolution of \\(256 \\times 256\\). This interpolation process may inevitably damage the details of the output prediction, particularly in high-resolution images. Our mask refinement module is designed to generate higher-resolution masks while enhancing the result quality. We take the object embedding from the prompt retargeting module and the multi-scale features from SAM2's hierarchical features to predict the final mask. The encoder blocks in Figure 3 represent the sub-modules within the image encoder, which generate multi-scale features with resolutions \\(256 \\times 256\\), \\(128 \\times 128\\) and \\(64 \\times 64\\), respectively. We incorporate the UNet-like structure to fuse the object features and multi-scale image features. Each of decoder blocks contains an upsampling layer and a convolutional layer, and the upsampling feature along with image feature are concatenated via the channel dimension before sending into the convolutional layer. After the final decoder block, we conduct two transposed convolution layers to obtain mask feature with \\(1024 \\times 1024\\) resolution. The low-level feature generated from input image \\(\\mathbf{I}\\) is added further enhance image quality, and a convolutional layer is then used to generate the final segmentation map." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22513" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.186, + 0.09, + 0.812, + 0.104 + ], + "angle": 0, + "content": "Table 1. Results comparison with state-of-the-art methods on four fine-grained image segmentation tasks." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.12, + 0.907, + 0.299 + ], + "angle": 0, + "content": "
ModelDISCOIFTHRSODThinObjectAverage
mIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoU
SAM (vit_l) [16]55.547.988.483.485.278.466.155.673.866.3
SAM (vit_h) [16]50.643.986.581.780.474.257.649.468.862.3
SAM2 (hiera_b+) [28]57.249.993.288.785.276.366.365.175.570.0
SAM2 (hiera_l) [28]57.650.792.488.086.379.577.867.778.571.5
HQ-SAM (vit_l) [14]77.469.294.489.791.984.687.076.687.780.0
HQ-SAM (vit_h) [14]77.969.794.489.591.983.588.578.688.280.3
SAM2Refiner (hiera_b+)83.478.596.593.394.189.495.190.192.387.8
SAM2Refiner (hiera_l)85.280.696.693.694.790.295.991.093.188.9
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.315, + 0.305, + 0.331 + ], + "angle": 0, + "content": "3.3. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.482, + 0.548 + ], + "angle": 0, + "content": "Training. We train our SAM2Refiner on image datasets only. We utilize the HQSeg-44K [14] as the training database, which includes six image datasets with extremely fine-grained mask labeling, including DIS [26](train set), ThinObject-5K [19] (train set), FSS-1000 [17], ECSSD [32], MSRA-10K [8], DUT-OMRON [41]. During training, we fix the model parameters of the pre-trained SAM2 model. We train by sampling mixed types of prompt including bounding boxes, randomly sampled points (either positive or negative), and coarse mask input. We use a learning rate of 0.001 and train for 20 epochs, with a learning rate decay schedule. We train on 8 NVIDIA V100 GPUs with a total batch size of 32. If not otherwise specified, we use the hierarchical large model as the default backbone." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.55, + 0.482, + 0.671 + ], + "angle": 0, + "content": "Loss Function. Following HQ-SAM [14], we use a combination of binary cross-entropy (BCE) loss and dice loss for supervision. In addition to calculating the loss at the final prediction, we introduce intermediate supervision at each layer of the three decoder blocks. Before the loss calculation, each of the output features is sent to a separate convolutional layer to align the channel size with ground-truth mask. Our final loss function is:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.681, + 0.483, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {f} \\left(\\mathcal {L} _ {f} ^ {b c e} + \\mathcal {L} _ {f} ^ {d i c e}\\right) + \\lambda_ {i} \\sum_ {i = 1} ^ {3} \\left(\\mathcal {L} _ {i} ^ {b c e} + \\mathcal {L} _ {i} ^ {d i c e}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.468, + 0.749 + ], + "angle": 0, + "content": "where \\(\\lambda_{f}\\) and \\(\\lambda_{i}\\) are set to 1.0 and 0.3 in our experiment." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Inference. For image setting, we follow the same inference pipeline as SAM. In the SAM2 temporal streaming pipeline, we extend our SAM2Refiner to the video segmentation by adding the cross-frame interaction via the memory bank. As suggested in Figure 3, compared to the original SAM2 pipeline, SAM2Refiner, including three proposed components, leverages the mask feature memory of previous frames to perform segmentation for the current frame, achieving accurate video segmentation. Note that for each frame except the first frame, the model uses the mem" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.315, + 0.907, + 0.347 + ], + "angle": 0, + "content": "ory feature obtained by the first frame and the past nearest frames to perform frame-by-frame segmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.362, + 0.649, + 0.379 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.388, + 0.707, + 0.404 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.546 + ], + "angle": 0, + "content": "Datasets. Following the prior work [14], we evaluate the model ability in both image and video segmentation tasks respectively. In image segmentation, all models are assessed in four challenging fine-grained segmentation benchmarks, including DIS-validation [26], ThinObject5K-test [19], COIFT [19] and HR-SOD [46]. For video object segmentation, we leverage two long-term video segmentation datasets, LVOS-v1 and LVOS-v2 [11], to evaluate our framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.655 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate the results of the image segmentation task, we echo previous works [14, 16] and adopt mIoU and mBIoU scores as metrics. For video object segmentation, we utilize \\( J \\) (region similarity), \\( F \\) (contour accuracy), and the combined \\( J \\& F \\) scores as metrics, where evaluations are performed in a semi-supervised setting with the first frame prompt given." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.757, + 0.682 + ], + "angle": 0, + "content": "4.2. Comparisons and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.764 + ], + "angle": 0, + "content": "We perform extensive comparisons between our SAM2Refiner with SAM [16], SAM2 [28], HQ-SAM [14] and its video variant HQSAM2 on 6 benchmarks across image and video segmentation tasks. We do not compare Pi-SAM as they have not open-sourced their code." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Open-world Image Segmentation. To evaluate the segmentation results in the open-world environment, we compare our SAM2Refiner with previous methods on four image segmentation benchmarks. In Table 1, our SAM2Refiner with large backbone achieves state-of-the-art performance of \\(85.2\\%\\), \\(96.6\\%\\), \\(94.7\\%\\), \\(95.9\\%\\) mIoU scores and \\(80.6\\%\\), \\(93.6\\%\\), \\(90.2\\%\\), \\(91.0\\%\\) mBiOUscores on DIS, COIFT, HRSOD and ThinObject datasets, respectively. In average, SAM2Refiner with a smaller backbone" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22514" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.262, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.089, + 0.581, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.089, + 0.898, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.552, + 0.908, + 0.581 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with previous methods. Given the blue box as visual prompt, our proposed SAM2Refiner produces more accurate results with correct structure and clear boundaries. Zoom in for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.786 + ], + "angle": 0, + "content": "(hiera_1) performs \\(93.1\\%\\) and \\(88.9\\%\\) average mIoU and mBiOU scores, significantly outperforming the second-best one HQ-SAM (vit_h) with \\(4.9\\%\\) and \\(8.6\\%\\) scores. To clear represent the segmentation results, we show the comparison of the visualization results of segmentation maps. From Figure 4, our SAM2Refiner can produce much more accurate segmentation maps than other methods, specifically on detailed objects such as boundary and small regions. These results demonstrate that the SAM2Refiner benefitting from the proposed modules achieves superior fine-grained image segmentation and strong generalization ability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Zero-Shot Video Object Segmentation. To evaluate the ability of the proposed method on video segmentation, we extend our model on semi-supervised video object segmentation task upon the pipeline of SAM2. Towards the zero-shot segmentation in the open-world scenario, we compare our model with previous video segmentation works in two benchmarks, including LVOS v1 and LVOS v2. As sug-" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.618, + 0.907, + 0.646 + ], + "angle": 0, + "content": "Table 2. Comparison of the proposed method with prior works on video object segmentation task." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.657, + 0.905, + 0.813 + ], + "angle": 0, + "content": "
MethodsLVOS v1LVOS v2
J&FJFJ&FJsFsJuFu
LWL [13]56.451.860.960.658.064.357.262.9
CFBI [13]51.546.256.755.052.959.251.756.2
STCN [13]48.943.954.060.657.264.057.563.8
RDE [25]53.748.359.262.256.764.160.867.2
SwinB-DeAOT-L [42]---63.961.569.058.466.6
XMem [4]52.948.157.764.562.669.160.665.6
SAM2 [28]77.272.282.278.375.287.068.579.4
HQ-SAM2 [14]79.674.984.379.176.588.369.482.2
SAM2Refiner80.975.186.879.978.787.971.681.5
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "gested in Table 2, our model yields \\(80.9\\%\\), \\(75.1\\%\\), \\(86.8\\%\\) scores of \\(\\mathcal{J} \\& \\mathcal{F}\\), \\(\\mathcal{J}\\), \\(\\mathcal{F}\\) and \\(79.9\\%\\), \\(78.7\\%\\), \\(87.9\\%\\), \\(71.6\\%\\), \\(81.5\\%\\) of \\(\\mathcal{J} \\& \\mathcal{F}\\), \\(\\mathcal{J}_s\\), \\(\\mathcal{F}_s\\), \\(\\mathcal{J}_u\\), \\(\\mathcal{F}_u\\) in two datasets respectively, significantly outperforming over previous compar" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22515" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.119 + ], + "angle": 0, + "content": "Table 3. Analysis of the point prompt response in interactive image segmentation." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.125, + 0.485, + 0.259 + ], + "angle": 0, + "content": "
Point Num.MetricsSAMSAM2HQ-SAMSAM2Refiner
1mIoU8.711.417.725.2
mB IoU6.39.914.320.6
2mIoU30.835.644.652.5
mB IoU26.231.740.247.4
5mIoU70.173.984.290.0
mB IoU62.366.475.584.9
10mIoU71.776.786.191.6
mB IoU64.870.478.588.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.273, + 0.483, + 0.33 + ], + "angle": 0, + "content": "Table 4. Ablation study on the three modules of SAM2Refiner, including Localization Augment (LA), Prompt Retargeting (PR) and Mask Refinement (MR). The results are averaged over the four evaluated datasets." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.336, + 0.462, + 0.458 + ], + "angle": 0, + "content": "
LAPRMRmIoUmBIoU
SAM2 [29]78.571.5
SAM2Refiner✓85.067.4
✓✓91.186.0
✓✓92.788.3
✓✓91.486.2
✓✓✓93.188.9
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.483, + 0.536 + ], + "angle": 0, + "content": "isons [4, 13, 14, 25, 28, 42]. These results demonstrate that the proposed framework can not only enhance fine-grained image segmentation, but also boost the video segmentation with refined object details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.536, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Analysis on prompt number. To evaluate the response intensity to input prompts in SAM2Refiner compared to other models, we input varying numbers of point prompts into the model in image segmentation task. In Table 3, we observe that the average segmentation results decline as the number of prompt points decreases. However, our SAM2Refiner indicates consistent advantages in segmentation accuracy across varying numbers of point prompts, while SAM, SAM2 and HQ-SAM reflect relatively low results. This clearly suggests that the proposed method with the capability of prompt re-targeting can effectively achieve a strong response for input prompts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.245, + 0.743 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.826 + ], + "angle": 0, + "content": "We conduct ablation experiments to evaluate the effectiveness of the proposed modules within our SAM2Refiner, i.e., Localization Augment (LA), Prompt Retargeting (PR) and Mask Refinement (MR), and we study the affect of the number of input prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In the LA module, the original global features attend to local features from sub-images, thereby enriching the representation with more detailed local information. As shown in Table 4, compared to SAM2, the LA module markedly enhances the mIoU from \\(78.5\\%\\) to \\(85.0\\%\\), despite a slight" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.646, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.188, + 0.597, + 0.199 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.09, + 0.772, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.188, + 0.726, + 0.198 + ], + "angle": 0, + "content": "w/o LA" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.09, + 0.899, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.188, + 0.85, + 0.198 + ], + "angle": 0, + "content": "w/LA" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.203, + 0.907, + 0.245 + ], + "angle": 0, + "content": "Figure 5. Effectiveness of LA module. LA suggest remarkable improvement on local details although it displays some flickering boundaries." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.253, + 0.646, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.32, + 0.597, + 0.33 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.253, + 0.772, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.32, + 0.734, + 0.329 + ], + "angle": 0, + "content": "w/o PR" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.253, + 0.9, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.32, + 0.851, + 0.329 + ], + "angle": 0, + "content": "W/PR" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.334, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Figure 6. Effectiveness of PR module. PR allows precise response for point prompts in ambiguous regions. Green point and red point denote positive and negative prompt, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.658 + ], + "angle": 0, + "content": "decrease in boundary accuracy. This is because the LA module strengthens the perception of fine structures, but this can come at the cost of boundary precision. Figure 5 verifies this phenomenon, showing that results with LA exhibit more detailed local features but also show flickering boundaries. The second to fourth rows of the SAM2Refiner in Table 4 highlight the effectiveness of each individual module for fine-grained segmentation. The PR module allows the model to maintain strong sensibility towards point prompting in intricate regions. As shown in Figure 6, result with PR module is able to correctly respond to the input visual prompt. Compared to PR, both the LA and MR modules deliver greater performance improvements in mIoU and mBIoU, with about 2 points gain in mIoU and 3 points gain in mBIoU, underscoring their crucial roles in enhancing segmentation precision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.67, + 0.634, + 0.687 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.696, + 0.907, + 0.893 + ], + "angle": 0, + "content": "In this work, we propose SAM2Refiner to enhance interactive segmentation for images and videos. SAM2Refiner balances local detail perception and prompting stability, achieving fine-grained visual representation. By integrating local contextual cues through a cross-attention mechanism and employing a UNet-like structure for feature fusion, SAM2Refiner generates high-quality segmentation masks. In addition, a prompt retargeting module is designed to improve the alignment of object embeddings with spatial features. Extensive experiments demonstrate that our SAM2Refiner achieves superior segmentation performance on both image and video segmentation tasks, significantly outperforming other state-of-the-art methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22516" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.155 + ], + "angle": 0, + "content": "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.483, + 0.198 + ], + "angle": 0, + "content": "[2] Alexander Kirillov Bowen Cheng, Alex Schwing. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.2, + 0.482, + 0.241 + ], + "angle": 0, + "content": "[3] Wang Y. Chen J. Jiang D. Zhang X. Tian Q. Cao, H. and Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.482, + 0.282 + ], + "angle": 0, + "content": "[4] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.482, + 0.338 + ], + "angle": 0, + "content": "[5] Ho Kei Cheng, Jihoon Chung, Yu-Wing Tai, and Chi-Keung Tang. CascadePSP: Toward class-agnostic and very high-resolution segmentation via global and local refinement. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.482, + 0.394 + ], + "angle": 0, + "content": "[6] Ishan Misra Alexander G. Schwing Alexander Kirillov Cheng, Bowen and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.482, + 0.451 + ], + "angle": 0, + "content": "[7] Maxwell D. Collins Yukun Zhu Ting Liu Thomas S. Huang Hartwig Adam Cheng, Bowen and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.453, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[8] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. TPAMI, 37(3):569-582, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.482, + 0.534 + ], + "angle": 0, + "content": "[9] Jifeng Dai, Kaiming He, and Jian Sun. Instance-aware semantic segmentation via multi-task network cascades. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.537, + 0.483, + 0.578 + ], + "angle": 0, + "content": "[10] Peize Sun Xueyan Zou Shilong Liu Chunyuan Li Jianwei Yang Lei Zhang Jianfeng Gao Feng Li, Hao Zhang. Segment and recognize anything at any granularity. In ECCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.579, + 0.482, + 0.619 + ], + "angle": 0, + "content": "[11] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.482, + 0.647 + ], + "angle": 0, + "content": "[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[13] Yu-Wing Tai Ho Kei Cheng and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. In NeurIPS, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[15] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dolkar, and Ross Girshick. Segment anything. In ICCV, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[17] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[18] Florian Schroff Hartwig Adam Liang-Chieh Chen, George Papandreou. Rethinking atrous convolution" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "for semantic image segmentation. In arXiv preprint arXiv:1706.05587, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[19] Jun Hao Liew, Scott Cohen, Brian Price, Long Mai, and Jiashi Feng. Deep interactive thin object selection. In WACV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.167, + 0.905, + 0.222 + ], + "angle": 0, + "content": "[20] Mengzhen Liu, Mengyu Wang, Henghui Ding, Yilong Xu, Yao Zhao, and Yunchao Wei. Segment anything with precise interaction. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3790-3799, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.224, + 0.905, + 0.278 + ], + "angle": 0, + "content": "[21] Songtao Liu, Di Huang, et al. Receptive field block net for accurate and fast object detection. In Proceedings of the European conference on computer vision (ECCV), pages 385-400, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.904, + 0.322 + ], + "angle": 0, + "content": "[22] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.326, + 0.905, + 0.366 + ], + "angle": 0, + "content": "[23] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.369, + 0.904, + 0.41 + ], + "angle": 0, + "content": "[24] Fisher Yu Mingqiao Ye, Martin Danelljan and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In ECCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.904, + 0.453 + ], + "angle": 0, + "content": "[25] Zhiwei Xiong Bang Zhang Pan Pan Mingxing Li, Liucheng Hu and Dong Liu. Recurrent dynamic embedding for video object segmentation. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.904, + 0.498 + ], + "angle": 0, + "content": "[26] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.501, + 0.904, + 0.541 + ], + "angle": 0, + "content": "[27] Yixuan Yuan Qiushi Yang. Learning dynamic convolutions for multi-modal 3d mri brain tumor segmentation. In MIC-CAI Workshop, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.544, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[28] Gabeur V. Hu Y.T. Hu R. Ryali C. Ma T. Khedr H. Radle R. Rolland C. Gustafson L. Ravi, N. and E. Mintun. Sam2: Segment anything in images and videos. In arXiv preprint arXiv:2408.00714, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.602, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[29] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.673, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[30] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.904, + 0.799 + ], + "angle": 0, + "content": "[31] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.904, + 0.842 + ], + "angle": 0, + "content": "[32] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended.cssd. TPAMI, 38(4): 717-729, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[33] Mengyu Wang, Henghui Ding, Jun Hao Liew, Jiajun Liu, Yao Zhao, and Yunchao Wei. Segrefiner: Towards model-agnostic segmentation refinement with discrete diffusion process. arXiv preprint arXiv:2312.12425, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22517" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[34] Sun K. Cheng T. Jiang B. Deng C. Zhao Y. Liu D. Mu Y. Tan M. Wang X. Wang, J. and W. Liu. Deep high-resolution representation learning for visual recognition. In TPAMI, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[35] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[36] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[37] Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[38] Wenhai Wang Zhiding Yu Anima Anandkumar Jose M. Alvarez Xie, Enze and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[39] Xiuwei Xu, Huangxing Chen, Linqing Zhao, Ziwei Wang, Jie Zhou, and Jiwen Lu. Embodiedsam: Online segment any 3d thing in real time. arXiv preprint arXiv:2408.11811, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[40] Hao Zhang Feng Li Linjie Li Jianfeng Wang Lijuan Wang Jianfeng Gao Yong Jae Lee Xueyan Zou, Jianwei Yang. Segment everything everywhere all at once. In NeurlPS, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[41] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[42] Zongxin Yang and Yi Yang. Decoupling features in hierarchical propagation for video object segmentation. In NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[43] Qian Yu, Xiaqi Zhao, Youwei Pang, Lihe Zhang, and Huchuan Lu. Multi-view aggregation network for dichotomous image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3921-3930, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[44] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[45] Yuhui Yuan, Jingyi Xie, Xilin Chen, and Jingdong Wang. Segfix: Model-agnostic boundary refinement for segmentation. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[46] Yi Zeng, Pingping Zhang, Jianming Zhang, Zhe Lin, and Huchuan Lu. Towards high-resolution salient object detection. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[47] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22518" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/b24a62fe-1619-470e-918c-9d3158f1e3f4_origin.pdf b/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/b24a62fe-1619-470e-918c-9d3158f1e3f4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fc9921d6558cbb03565011488f92dec4db77f427 --- /dev/null +++ b/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/b24a62fe-1619-470e-918c-9d3158f1e3f4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3b59f9b21735a82ab3cc2e0776735f283eca99e11f9995de890b0407e2af8f00 +size 4092295 diff --git a/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/full.md b/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5dc1610a163b723c24a0c88c2588380425abbcbd --- /dev/null +++ b/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/full.md @@ -0,0 +1,254 @@ +# Towards Fine-grained Interactive Segmentation in Images and Videos + +Yuan Yao Qiushi Yang Miaomiao Cui Liefeng Bo + +Tongyi Lab, Alibaba Group + +{ryan.yy, yangqiushi.yqs, miaomiao.cmm, liefeng.bo}@alibaba-inc.com + +![](images/bb0427cde5fe37f9e5e300d8ec466acfa54465046eb51e840600f6413185aa1c.jpg) +Figure 1. Segmentation results of the proposed framework SAM2Refiner and previous methods. The proposed SAM2Refiner achieves the best performance on both image and video segmentation, especially in fine-grained details. + +# Abstract + +The recent Segment Anything Models (SAMs) have emerged as foundational visual models for general interactive segmentation. Despite demonstrating robust generalization abilities, they still suffer from performance degradations in scenarios that demand accurate masks. Existing methods for high-precision interactive segmentation face a trade-off between perceiving intricate local details and maintaining stable prompting capability, which hinders the applicability and effectiveness of foundational segmentation models. In this paper, we present a SAM2Refiner framework built upon the SAM2 backbone. This architecture allows SAM2 to generate fine-grained segmentation masks for both images and videos while preserving its inherent strengths. Specifically, we design a localization augment module, which incorporates local contextual cues to enhance global features via a cross-attention mechanism, thereby exploiting potential detailed patterns while maintaining semantic information. Moreover, to strengthen the prompting ability toward the enhanced object embeddings, we introduce a prompt + +retargeting module that renews the embedding with spatially aligned prompt features. In addition, to obtain accurate high resolution segmentation masks, a mask refinement module is devised by employing a multi-scale cascaded structure to fuse mask features with hierarchical representations from the encoder. Extensive experiments demonstrate the effectiveness of our approach, revealing that the proposed method can produce highly precise masks for both images and videos, surpassing state-of-the-art methods. + +# 1. Introduction + +Interactive segmentation is a long-standing computer vision task that allows users to segment objects from the guidance of user interaction, typically in the form of clicks, strokes, or bounding boxes. In recent years, the Segment Anything Model (SAM) [16] has emerged as a foundational visual model for general interactive segmentation. Trained on their proposed SA-1B dataset, which consists of more than 1 billion masks and 11 million images, SAM can generate masks in various scenes and has been widely applied + +to a wide range of downstream vision tasks, including image editing [37, 44], 3D vision [31, 39] and medical image analysis [23, 36]. The following work Segment Anything Model 2 (SAM2) [29] further introduces the streaming memory mechanism to extend the ability of SAM to videos, providing a potential boost to video applications such as AR / VR, robotics and autonomous vehicles. + +Despite demonstrating robust generalization abilities, SAMs encounter the problems in scenario demanding fine-grained segmentation. In particular, as shown in Figure 1, they struggle to predict a complete mask for objects with fine structures or sharp boundaries, leading to overly smooth results or coarse boundaries. Thus, these types of failure may hinder the applicability of SAMs. Valuable efforts [14, 20, 23, 29] have been devoted to solve these challenges by introducing extra input tokens and explicitly fusing the multi-level features, however, they are still limited by SAMs' ability, thus failing to achieve satisfactory fine-grained segmentation. + +To this end, we aim to start with a holistic analysis on the three key points of fine-grained segmentation toward mining detailed cues and global semantic information guided by the user prompts. Firstly, considering that the input image contains global information of semantics while it lacks the detailed representations on details, we are expected to dynamically harvest the complementary local and global features within the input image. Secondly, SAMs directly combine the prompt with the object features via the decoder, which fail to build strong correlation among prompts with fine-grained cues and are hard to activate the response on object details, producing biased segmentation. Hence, strengthening the correspondences between prompt features and object details is crucial to yield a high response intensity of prompts on local details, which is beneficial to fine-grained segmentation. Thirdly, the object features with low resolution projected by the SAMs' encoder may lose much detailed cues, to recover the missed fine-grained local patterns for yielding high-resolution segmentation maps, the multi-scale features within SAMs' encoder are important to be re-used. A cocurrent research Pi-SAM [20] is the most similar work to ours, which proposes an additional mask decoder to expand the capability of SAM in high-resolution images. Nevertheless, it compromises with diminished responsiveness to visual prompt, necessitating a complementary module for error correction, and exhibit biased preferences for background clicks. Differently, our goal is to facilitate prompt response and consider both local and global presentations for efficient fine-grained segmentation. + +To achieve the aforementioned goals, we propose the SAM2Refiner architecture, which includes three components for fine-grained interactive segmentation. Specifically, we introduce Localization Augment (LA) to refine the details of global features with local contextual feature + +cues. We split the input image into four sub-images, and send to SAM2's image encoder together with raw image. A global-local cross-attention mechanism is designed to highlight the local information of the object within the global representation among deep features. Moreover, to improve the response intensity of prompting for segmenting objects with intricate structures, we propose a Prompt Retargeting (PR) strategy by providing precise alignment between the prompts and enhanced object features. PR performs spatial alignment between prompt features and object features, and then incorporates them to yield an enhanced prompt response. In addition, to generate higher resolution masks with finer details, the Mask Refinement (MR) module is devised by incorporating the object features and multi-scale features from the original image. In MR, the intermediate features within the SAM2 encoder is re-used to fully preserve the original zero-shot capabilities. Although designed for image segmentation, the SAM2-based architecture enables native interoperability with video streaming workflows, allowing fine-grained segmentation for both images and videos (as in Figure 1). Extensive experiments on both image and video segmentation benchmarks demonstrate the effectiveness of our approach over state-of-the-art methods. + +Our main contributions can be summarized as follows. + +- We propose SAM2Refiner, a new architecture capable of generating fine-grained segmentation masks with strong perception for both global semantic and local details, and simultaneously maintaining stable prompting ability and powerful zero-shot capability. +- In SAM2Refiner, the localization augment module is specifically designed to achieve a holistic representation of input image. By incorporating sub-images as additional inputs, it effectively enhances the global features through improved object localization. +- To boost the responsiveness of prompting in intricate regions, we propose a prompt retargeting module that establishes precise alignment between spatially-aware dense prompt features and the enhanced object features. +- To produce high-resolution segmentation maps with fine details, the mask refinement module is crafted to integrate multi-scale representations with object features. +- Building upon the SAM2 backbone, our architecture exhibits strong compatibility to the video streaming pipeline, exhibiting state-of-the-art fine-grained segmentation on both images and videos benchmarks. + +# 2. Related Work + +# 2.1. Interactive Segmentation + +Interactive segmentation enables users to provide cues for target regions and guide the segmentation process. User interactions can take various forms, including scribbles [14, 16, 28, 35], bounding boxes [14, 16, 20], clicks [16, + +45], or language prompts [35, 40]. Traditional methods utilize pixel-level energy minimization methods, capturing low-level appearance features through unary potentials and ensuring consistent segmentation results with pairwise ones [33]. In recent years, many studies leverage prompts as input features and inject them into the models to produce segmentation results [10, 16, 28, 40]. The Segment Anything Model (SAM) [16] pre-trained on a large-scale datasets has emerged as a benchmark for interactive segmentation. Considering that SAM cannot perform semantic prediction, many works [10, 40] utilize category labels to further fine-tune SAM and enable it to achieve the semantic segmentation task. For example, SEEM [40] further trains SAM using labeled segmentation data through a bipartite matching constraint, thus providing the model with the ability to predict semantics. Semantic-SAM [10] proposes a multi-choice learning scheme via multi-task training across diverse datasets, enabling the model to segment at various granularities while predicting semantic labels. Other works [23, 24, 39] apply SAM to diverse specific domains to solve application tasks. MedSAM [23] finetunes SAM by collecting a large-scale dataset of medical image samples in multiple imaging modalities and types of cancer, achieving superior results in disease segmentation. Gaussian-Grouping [24] leverages the SAM as supervision model to jointly train a reconstruction and segmentation of objects in open-world 3D scenes for high-quality 3D edit. + +# 2.2. High-Quality Segmentation + +High-quality segmentation as a fundamental computer vision task focuses on accurately segmenting various complex and detail-rich objects, which ranges from many subtasks including semantic segmentation [22, 47], instance segmentation [9, 12] and panoptic segmentation [15]. Traditional methods [5, 7, 18, 27, 34] on high-quality segmentation dedicate to designing elaborate structures upon on CNN-based networks in order to capture both low-level local patterns and semantic global features through various reception fields towards holistic representation. Deplabv3 [18] designs a set of multi-level convolution kernels to capture features with diverse reception fields and HRNet [34] employs cascaded structures to fuse features from different layers to learn both low-level and high-level information. More recently, many approaches [2, 3, 6, 35, 38] built on the transformer backbone focus on introducing local patterns into the model by narrow the reception field of self-attention module. For example, Mask2Former [6] presents an efficient versatile architecture by using masked attention to extract local features for high-quality image segmentation tasks. Swin-Unet [3] adopts the SwinTransformer network to the UNet like structures to adaptively capture object features with various scales via hierarchical attention. Latest, the Pi-SAM [20] designs a mask + +decoder to expands SAMs ability at high resolution images, however, it may diminish the response of the input prompt and exhibit biased preferences for background clicks. + +# 3. Method + +We propose SAM2Refiner, which effectively generates fine-grained segmentation masks with strong perception for both global semantic and local details, and simultaneously maintaining stable prompting ability. In Sec 3.1, we first briefly review the architecture of SAM and SAM2. Then, in Sec 3.2, we introduce the three modules proposed upon SAM2's backbone to effectively improve the precision for segmentation. Finally, in Sec 3.3, we describe the training and inference process of the method, we also depict how our method can be seamlessly inserted into SAM2's video streaming pipeline for zero-shot inference. The core framework of our SAM2Refiner is shown in Figure 2. + +# 3.1. Preliminaries: SAMs + +SAM consists of an image encoder, a prompt encoder, and a mask decoder. The image encoder works for image feature extraction with a ViT [1] backbone. The prompt encoder is employed to encode the points, boxes, or masks into prompt tokens. The mask decoder utilizes a two-layer transformer-based decoder that takes both the extracted image embedding with the concatenated output and prompt tokens for final mask prediction. + +SAM2 extends SAMs ability to support both images and videos. For image setting, the three modules are largely follow SAM except for two differences, firstly, SAM2 adopt an MAE pretrained Hiera [30] image encoder, which is hierarchical compared with the plane structure in SAM's encoder, and secondly, to solve ambiguity in images without valid object, SAM2 predicts an additional token in mask decoder to indicate whether the object of interest is present. SAM2 introduces additional memory modules to support video streaming, including memory attention, memory encoder, and memory bank. The role of memory attention is to condition the current frame features on the past frame features and predictions as well as on any new prompts. The memory encoder is employed to obtain a memory embedding from masks generated from mask decoder, these memory embeddings are appended to a memory bank in a first-in-first-out (FIFO) way. + +# 3.2. Ours: SAM2Refiner + +In this section, we describe the architecture of SAM2Refiner, which is built upon SAM2's backbone. This design offers two key advantages, 1) the hierarchical image encoder facilitates multi-scale feature capturing, which aids in feature refinement. 2) it allows for flexible integration into the video streaming pipeline. The proposed architecture comprises three modules, Localization Augment (LA) + +![](images/e42ad3091e8e7a20164e5e90f639d4f01846ef01642c3bfaf3ac1275c3fc2d01.jpg) +Figure 2. An overview of the proposed SAM2Refiner framework. It contains a localization augment module to balance the localization details and global semantic representations, a prompt retargeting module to enhance responsiveness towards input prompts, and a mask refinement module to boost the quality of mask outcomes. The blue line denotes the SAM2[28] pipeline, and the black line denotes the proposed SAM2Refiner pipeline. + +is inserted into the middle layer of image encoder and mask decoder to refine the details of global features with local contextual feature cues. Prompt Retargeting (PR) aligns the prompt feature with enhanced object embedding to provide flexible prompting. Mask Refinement (MR) further incorporates object features and multi-scale features to generate higher-resolution masks with finer details. + +# 3.2.1. Localization Augment + +We evenly crop the input image $\mathbf{I} \in \mathbb{R}^{B \times 3 \times H \times W}$ into four non-overlapping sub-images $\{\mathbf{s}_m\}_{m=1}^4 \in \mathbb{R}^{B \times 3 \times \frac{H}{2} \times \frac{W}{2}}$ . To align with input resolution of SAM2's image encoder, we upsample the sub-images to the same resolution with $\mathbf{I}$ and obtain $\{\mathbf{S}_m\}_{m=1}^4 \in \mathbb{R}^{B \times 3 \times H \times W}$ . $\mathbf{I}$ and $\{\mathbf{S}_m\}_{m=1}^M$ are then fed in batches to the image encoder and generate the highest-level feature maps $\{F_i\}_{i=1}^5 \in \mathbb{R}^{B \times C \times \frac{H}{16} \times \frac{W}{16}}$ . We then split the feature maps into global feature $G = F_1$ and local feature $\{L_m\}_{m=1}^4 = \{F_i\}_{i=2}^5$ . The local features contain fine-grained details of input images and are used to enhance global feature with detailed clues. The local features are first assembled into a unified concatenated feature $L \in \mathbb{R}^{B \times C \times \frac{H}{8} \times \frac{W}{8}}$ following their respective positions in the original image. We then apply a multi-granularity average pooling operation as [43] to cap + +ture important contextual feature cues with different scales. The pooling functions are denote as: + +$$ +P _ {i} = \operatorname {A v g P o o l} _ {i} (L), i \in \{1, \dots , n \}. \tag {1} +$$ + +Where $\mathrm{AvgPool}_i$ is the $i$ -th pooling function, $P_i$ is the $i$ -th feature map. We set $n = 3$ in our experiments, and the receptive fields for each pooling function are set 2, 4, 8 respectively. The feature maps $P_i$ are then tokenized and concatenated into a multi-granularity local features $L_t$ . For global feature $G$ , we tokenized it as $G_t$ . We conduct global-local cross attention between $G_t$ and $L_t$ to obtain the refined global feature $G_{re}$ : + +$$ +G _ {r e} = \operatorname {C r o s s A t t n} \left(G _ {t}, L _ {t}, L _ {t}\right) \tag {2} +$$ + +Before sending into attention layer, the positional encoding is added to the features to access critical geometric information. $G_{re}$ is able to highlight the local information about the object within the global representation. + +# 3.2.2. Prompt Retargeting + +We replace the global feature $G$ with $G_{re}$ as the input to the mask decoder. The output mask $M \in \mathbb{R}^{B \times 3 \times h \times w}$ and object embedding $E \in \mathbb{R}^{B \times C \times h \times w}$ of the mask decoder therefore contain richer representations of the local details. + +![](images/3c87b1bb6e0b674d522c8cb37e40b122c139998cd9023e3651ff593ae555401f.jpg) + +![](images/03f1738bc194a4e26eeabd64570e1cf9b1984e745bbb6d1545d1169b479166ba.jpg) +Figure 3. A Comparison between SAM2 and SAM2Refiner. The proposed three modules in SAM2Refiner can be seamlessly embedded into SAM2's video streaming pipeline, allowing to support fine-grained interactive segmentation for both images and videos. + +To enhance the discriminability of features in small patches, we introduce a modified version of multi-branch convolutional block RFB [21], which applies different kernels to achieve multi-size receptive fields. To be specific, we first employ the bottleneck structure in each branch, consisting of a $1 \times 1$ convolutional layer to decrease the channel size in object embedding. Secondly, we apply three parallelized blocks with $1 \times n$ plus a $n \times 1$ convolutional layer to each branch, where we set $n = 3,5,7$ , respectively. Thirdly, three branches are concatenated and encoded via a $1 \times 1$ convolutional layer. Subsequently, we obtain the augmented embedding $E_{a}$ by adding the concatenated embedding and a shortcut embedding encoded from another $1 \times 1$ convolutional layer. Although the ability to perceive local details is significantly enhanced, $E_{a}$ exhibits reduced sensitivity to the input visual prompts. To maintain a stable prompting ability towards the enhanced embedding, we employ a three-channel prompt map to align the prompt feature with the augmented object embedding $E_{a}$ . The map resolution is set as $256 \times 256$ to reduce the computational cost and only two visual prompts ( clicks and masks) are encoded. We represent clicks as a fixed size radius, where positive clicks are encoded in the first channel, and negative clicks are encoded in the second channel. The mask $M$ is then concatenated into the third channel to form a three-channel dense prompt map $D_{m}$ . We propose a dense prompt encoder to transform the $D_{p}$ to a dense prompt embedding $E_{p}$ . The dense prompt encoder consists of two $2 \times 2$ convolution layers followed by GELU activations and layer normalization, and one $1 \times 1$ convolution layer to map the channel dimension to 256. The dense prompt embedding $E_{p}$ is added to $E_{a}$ and obtain $E_{ap}$ in an element-wise manner, followed by a retargeting modules with several attention operations as in SAM2's mask decoder. We re-use the sparse prompt + +embedding and output token from SAM2's mask decoder to provide positional and class information for the renewing of the object embedding. The retargeting module contains four steps, (1) self-attention on the tokens concatenated from the the sparse prompt embedding and output token; (2) cross-attention from token to $E_{ap}$ ; (3) MLP updated on token, (4) cross-attention from $E_{ap}$ to token. The renewed object embedding allows to possess a higher sensitivity towards the input visual prompts. + +# 3.2.3. Mask Refinement + +The final mask of SAM2 is directly interpolated from the output prediction of its mask decoder, with a resolution of $256 \times 256$ . This interpolation process may inevitably damage the details of the output prediction, particularly in high-resolution images. Our mask refinement module is designed to generate higher-resolution masks while enhancing the result quality. We take the object embedding from the prompt retargeting module and the multi-scale features from SAM2's hierarchical features to predict the final mask. The encoder blocks in Figure 3 represent the sub-modules within the image encoder, which generate multi-scale features with resolutions $256 \times 256$ , $128 \times 128$ and $64 \times 64$ , respectively. We incorporate the UNet-like structure to fuse the object features and multi-scale image features. Each of decoder blocks contains an upsampling layer and a convolutional layer, and the upsampling feature along with image feature are concatenated via the channel dimension before sending into the convolutional layer. After the final decoder block, we conduct two transposed convolution layers to obtain mask feature with $1024 \times 1024$ resolution. The low-level feature generated from input image $\mathbf{I}$ is added further enhance image quality, and a convolutional layer is then used to generate the final segmentation map. + +Table 1. Results comparison with state-of-the-art methods on four fine-grained image segmentation tasks. + +
ModelDISCOIFTHRSODThinObjectAverage
mIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoU
SAM (vit_l) [16]55.547.988.483.485.278.466.155.673.866.3
SAM (vit_h) [16]50.643.986.581.780.474.257.649.468.862.3
SAM2 (hiera_b+) [28]57.249.993.288.785.276.366.365.175.570.0
SAM2 (hiera_l) [28]57.650.792.488.086.379.577.867.778.571.5
HQ-SAM (vit_l) [14]77.469.294.489.791.984.687.076.687.780.0
HQ-SAM (vit_h) [14]77.969.794.489.591.983.588.578.688.280.3
SAM2Refiner (hiera_b+)83.478.596.593.394.189.495.190.192.387.8
SAM2Refiner (hiera_l)85.280.696.693.694.790.295.991.093.188.9
+ +# 3.3. Training and Inference + +Training. We train our SAM2Refiner on image datasets only. We utilize the HQSeg-44K [14] as the training database, which includes six image datasets with extremely fine-grained mask labeling, including DIS [26](train set), ThinObject-5K [19] (train set), FSS-1000 [17], ECSSD [32], MSRA-10K [8], DUT-OMRON [41]. During training, we fix the model parameters of the pre-trained SAM2 model. We train by sampling mixed types of prompt including bounding boxes, randomly sampled points (either positive or negative), and coarse mask input. We use a learning rate of 0.001 and train for 20 epochs, with a learning rate decay schedule. We train on 8 NVIDIA V100 GPUs with a total batch size of 32. If not otherwise specified, we use the hierarchical large model as the default backbone. + +Loss Function. Following HQ-SAM [14], we use a combination of binary cross-entropy (BCE) loss and dice loss for supervision. In addition to calculating the loss at the final prediction, we introduce intermediate supervision at each layer of the three decoder blocks. Before the loss calculation, each of the output features is sent to a separate convolutional layer to align the channel size with ground-truth mask. Our final loss function is: + +$$ +\mathcal {L} = \lambda_ {f} \left(\mathcal {L} _ {f} ^ {b c e} + \mathcal {L} _ {f} ^ {d i c e}\right) + \lambda_ {i} \sum_ {i = 1} ^ {3} \left(\mathcal {L} _ {i} ^ {b c e} + \mathcal {L} _ {i} ^ {d i c e}\right) \tag {3} +$$ + +where $\lambda_{f}$ and $\lambda_{i}$ are set to 1.0 and 0.3 in our experiment. + +Inference. For image setting, we follow the same inference pipeline as SAM. In the SAM2 temporal streaming pipeline, we extend our SAM2Refiner to the video segmentation by adding the cross-frame interaction via the memory bank. As suggested in Figure 3, compared to the original SAM2 pipeline, SAM2Refiner, including three proposed components, leverages the mask feature memory of previous frames to perform segmentation for the current frame, achieving accurate video segmentation. Note that for each frame except the first frame, the model uses the mem + +ory feature obtained by the first frame and the past nearest frames to perform frame-by-frame segmentation. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. Following the prior work [14], we evaluate the model ability in both image and video segmentation tasks respectively. In image segmentation, all models are assessed in four challenging fine-grained segmentation benchmarks, including DIS-validation [26], ThinObject5K-test [19], COIFT [19] and HR-SOD [46]. For video object segmentation, we leverage two long-term video segmentation datasets, LVOS-v1 and LVOS-v2 [11], to evaluate our framework. + +Evaluation Metrics. To evaluate the results of the image segmentation task, we echo previous works [14, 16] and adopt mIoU and mBIoU scores as metrics. For video object segmentation, we utilize $J$ (region similarity), $F$ (contour accuracy), and the combined $J \& F$ scores as metrics, where evaluations are performed in a semi-supervised setting with the first frame prompt given. + +# 4.2. Comparisons and Analysis + +We perform extensive comparisons between our SAM2Refiner with SAM [16], SAM2 [28], HQ-SAM [14] and its video variant HQSAM2 on 6 benchmarks across image and video segmentation tasks. We do not compare Pi-SAM as they have not open-sourced their code. + +Open-world Image Segmentation. To evaluate the segmentation results in the open-world environment, we compare our SAM2Refiner with previous methods on four image segmentation benchmarks. In Table 1, our SAM2Refiner with large backbone achieves state-of-the-art performance of $85.2\%$ , $96.6\%$ , $94.7\%$ , $95.9\%$ mIoU scores and $80.6\%$ , $93.6\%$ , $90.2\%$ , $91.0\%$ mBiOUscores on DIS, COIFT, HRSOD and ThinObject datasets, respectively. In average, SAM2Refiner with a smaller backbone + +![](images/77a26a816ed0cd89a468f98d396fade48a98acd20995cd424c7d424cb0c85c25.jpg) +Figure 4. Qualitative comparison with previous methods. Given the blue box as visual prompt, our proposed SAM2Refiner produces more accurate results with correct structure and clear boundaries. Zoom in for better visualization. + +![](images/8d993c40f9776eaa24539639daac917a607656cd46ce50ac4a2f92dba284cfcd.jpg) + +![](images/58474c52edd750e1c50ea3469ab5d9c3b0944bddf99e27890ee12b060fc48f7b.jpg) + +(hiera_1) performs $93.1\%$ and $88.9\%$ average mIoU and mBiOU scores, significantly outperforming the second-best one HQ-SAM (vit_h) with $4.9\%$ and $8.6\%$ scores. To clear represent the segmentation results, we show the comparison of the visualization results of segmentation maps. From Figure 4, our SAM2Refiner can produce much more accurate segmentation maps than other methods, specifically on detailed objects such as boundary and small regions. These results demonstrate that the SAM2Refiner benefitting from the proposed modules achieves superior fine-grained image segmentation and strong generalization ability. + +Zero-Shot Video Object Segmentation. To evaluate the ability of the proposed method on video segmentation, we extend our model on semi-supervised video object segmentation task upon the pipeline of SAM2. Towards the zero-shot segmentation in the open-world scenario, we compare our model with previous video segmentation works in two benchmarks, including LVOS v1 and LVOS v2. As sug- + +Table 2. Comparison of the proposed method with prior works on video object segmentation task. + +
MethodsLVOS v1LVOS v2
J&FJFJ&FJsFsJuFu
LWL [13]56.451.860.960.658.064.357.262.9
CFBI [13]51.546.256.755.052.959.251.756.2
STCN [13]48.943.954.060.657.264.057.563.8
RDE [25]53.748.359.262.256.764.160.867.2
SwinB-DeAOT-L [42]---63.961.569.058.466.6
XMem [4]52.948.157.764.562.669.160.665.6
SAM2 [28]77.272.282.278.375.287.068.579.4
HQ-SAM2 [14]79.674.984.379.176.588.369.482.2
SAM2Refiner80.975.186.879.978.787.971.681.5
+ +gested in Table 2, our model yields $80.9\%$ , $75.1\%$ , $86.8\%$ scores of $\mathcal{J} \& \mathcal{F}$ , $\mathcal{J}$ , $\mathcal{F}$ and $79.9\%$ , $78.7\%$ , $87.9\%$ , $71.6\%$ , $81.5\%$ of $\mathcal{J} \& \mathcal{F}$ , $\mathcal{J}_s$ , $\mathcal{F}_s$ , $\mathcal{J}_u$ , $\mathcal{F}_u$ in two datasets respectively, significantly outperforming over previous compar + +Table 3. Analysis of the point prompt response in interactive image segmentation. + +
Point Num.MetricsSAMSAM2HQ-SAMSAM2Refiner
1mIoU8.711.417.725.2
mB IoU6.39.914.320.6
2mIoU30.835.644.652.5
mB IoU26.231.740.247.4
5mIoU70.173.984.290.0
mB IoU62.366.475.584.9
10mIoU71.776.786.191.6
mB IoU64.870.478.588.6
+ +Table 4. Ablation study on the three modules of SAM2Refiner, including Localization Augment (LA), Prompt Retargeting (PR) and Mask Refinement (MR). The results are averaged over the four evaluated datasets. + +
LAPRMRmIoUmBIoU
SAM2 [29]78.571.5
SAM2Refiner✓85.067.4
✓✓91.186.0
✓✓92.788.3
✓✓91.486.2
✓✓✓93.188.9
+ +isons [4, 13, 14, 25, 28, 42]. These results demonstrate that the proposed framework can not only enhance fine-grained image segmentation, but also boost the video segmentation with refined object details. + +Analysis on prompt number. To evaluate the response intensity to input prompts in SAM2Refiner compared to other models, we input varying numbers of point prompts into the model in image segmentation task. In Table 3, we observe that the average segmentation results decline as the number of prompt points decreases. However, our SAM2Refiner indicates consistent advantages in segmentation accuracy across varying numbers of point prompts, while SAM, SAM2 and HQ-SAM reflect relatively low results. This clearly suggests that the proposed method with the capability of prompt re-targeting can effectively achieve a strong response for input prompts. + +# 4.3. Ablation Study + +We conduct ablation experiments to evaluate the effectiveness of the proposed modules within our SAM2Refiner, i.e., Localization Augment (LA), Prompt Retargeting (PR) and Mask Refinement (MR), and we study the affect of the number of input prompts. + +In the LA module, the original global features attend to local features from sub-images, thereby enriching the representation with more detailed local information. As shown in Table 4, compared to SAM2, the LA module markedly enhances the mIoU from $78.5\%$ to $85.0\%$ , despite a slight + +![](images/457a3a98be44b41efd4869141b4c86174f222a3b37f6c3ce459e6c2702888403.jpg) +Input + +![](images/a041b8ec8f7ba3dc3e628221dcaad5bacb7920b9e1c0684f3205a7994d3ebf38.jpg) +w/o LA + +![](images/72f001d2ceb68d11a35995e42c5618413045b845107cfac188be7149c9aa88a4.jpg) +w/LA + +![](images/1174d49dee81c4aff286576d051bc763d1ae850a2d66c6895b4c7954379251d6.jpg) +Input +Figure 6. Effectiveness of PR module. PR allows precise response for point prompts in ambiguous regions. Green point and red point denote positive and negative prompt, respectively. + +![](images/8328686e622b037acc63c011fc90bdf7a80a7cbf9f419e79491725c86cf64a51.jpg) +w/o PR + +![](images/431ac431aa31c1ddcab1f5ac8e2eb2788c314d67ee81d050d1cd974b07d8d4d4.jpg) +Figure 5. Effectiveness of LA module. LA suggest remarkable improvement on local details although it displays some flickering boundaries. +W/PR + +decrease in boundary accuracy. This is because the LA module strengthens the perception of fine structures, but this can come at the cost of boundary precision. Figure 5 verifies this phenomenon, showing that results with LA exhibit more detailed local features but also show flickering boundaries. The second to fourth rows of the SAM2Refiner in Table 4 highlight the effectiveness of each individual module for fine-grained segmentation. The PR module allows the model to maintain strong sensibility towards point prompting in intricate regions. As shown in Figure 6, result with PR module is able to correctly respond to the input visual prompt. Compared to PR, both the LA and MR modules deliver greater performance improvements in mIoU and mBIoU, with about 2 points gain in mIoU and 3 points gain in mBIoU, underscoring their crucial roles in enhancing segmentation precision. + +# 5. Conclusion + +In this work, we propose SAM2Refiner to enhance interactive segmentation for images and videos. SAM2Refiner balances local detail perception and prompting stability, achieving fine-grained visual representation. By integrating local contextual cues through a cross-attention mechanism and employing a UNet-like structure for feature fusion, SAM2Refiner generates high-quality segmentation masks. In addition, a prompt retargeting module is designed to improve the alignment of object embeddings with spatial features. Extensive experiments demonstrate that our SAM2Refiner achieves superior segmentation performance on both image and video segmentation tasks, significantly outperforming other state-of-the-art methods. + +# References + +[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020. +[2] Alexander Kirillov Bowen Cheng, Alex Schwing. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021. +[3] Wang Y. Chen J. Jiang D. Zhang X. Tian Q. Cao, H. and Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In ECCV, 2022. +[4] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022. +[5] Ho Kei Cheng, Jihoon Chung, Yu-Wing Tai, and Chi-Keung Tang. CascadePSP: Toward class-agnostic and very high-resolution segmentation via global and local refinement. In CVPR, 2020. +[6] Ishan Misra Alexander G. Schwing Alexander Kirillov Cheng, Bowen and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. +[7] Maxwell D. Collins Yukun Zhu Ting Liu Thomas S. Huang Hartwig Adam Cheng, Bowen and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020. +[8] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. TPAMI, 37(3):569-582, 2014. +[9] Jifeng Dai, Kaiming He, and Jian Sun. Instance-aware semantic segmentation via multi-task network cascades. In CVPR, 2016. +[10] Peize Sun Xueyan Zou Shilong Liu Chunyuan Li Jianwei Yang Lei Zhang Jianfeng Gao Feng Li, Hao Zhang. Segment and recognize anything at any granularity. In ECCV, 2024. +[11] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. +[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. +[13] Yu-Wing Tai Ho Kei Cheng and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021. +[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. In NeurIPS, 2023. +[15] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019. +[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dolkar, and Ross Girshick. Segment anything. In ICCV, 2023. +[17] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In CVPR, 2020. +[18] Florian Schroff Hartwig Adam Liang-Chieh Chen, George Papandreou. Rethinking atrous convolution + +for semantic image segmentation. In arXiv preprint arXiv:1706.05587, 2017. +[19] Jun Hao Liew, Scott Cohen, Brian Price, Long Mai, and Jiashi Feng. Deep interactive thin object selection. In WACV, 2021. +[20] Mengzhen Liu, Mengyu Wang, Henghui Ding, Yilong Xu, Yao Zhao, and Yunchao Wei. Segment anything with precise interaction. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3790-3799, 2024. +[21] Songtao Liu, Di Huang, et al. Receptive field block net for accurate and fast object detection. In Proceedings of the European conference on computer vision (ECCV), pages 385-400, 2018. +[22] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015. +[23] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024. +[24] Fisher Yu Mingqiao Ye, Martin Danelljan and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In ECCV, 2024. +[25] Zhiwei Xiong Bang Zhang Pan Pan Mingxing Li, Liucheng Hu and Dong Liu. Recurrent dynamic embedding for video object segmentation. In CVPR, 2022. +[26] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In ECCV, 2022. +[27] Yixuan Yuan Qiushi Yang. Learning dynamic convolutions for multi-modal 3d mri brain tumor segmentation. In MIC-CAI Workshop, 2020. +[28] Gabeur V. Hu Y.T. Hu R. Ryali C. Ma T. Khedr H. Radle R. Rolland C. Gustafson L. Ravi, N. and E. Mintun. Sam2: Segment anything in images and videos. In arXiv preprint arXiv:2408.00714, 2024. +[29] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024. +[30] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023. +[31] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023. +[32] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended.cssd. TPAMI, 38(4): 717-729, 2015. +[33] Mengyu Wang, Henghui Ding, Jun Hao Liew, Jiajun Liu, Yao Zhao, and Yunchao Wei. Segrefiner: Towards model-agnostic segmentation refinement with discrete diffusion process. arXiv preprint arXiv:2312.12425, 2023. + +[34] Sun K. Cheng T. Jiang B. Deng C. Zhao Y. Liu D. Mu Y. Tan M. Wang X. Wang, J. and W. Liu. Deep high-resolution representation learning for visual recognition. In TPAMI, 2020. +[35] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. +[36] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023. +[37] Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023. +[38] Wenhai Wang Zhiding Yu Anima Anandkumar Jose M. Alvarez Xie, Enze and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. +[39] Xiuwei Xu, Huangxing Chen, Linqing Zhao, Ziwei Wang, Jie Zhou, and Jiwen Lu. Embodiedsam: Online segment any 3d thing in real time. arXiv preprint arXiv:2408.11811, 2024. +[40] Hao Zhang Feng Li Linjie Li Jianfeng Wang Lijuan Wang Jianfeng Gao Yong Jae Lee Xueyan Zou, Jianwei Yang. Segment everything everywhere all at once. In NeurlPS, 2023. +[41] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013. +[42] Zongxin Yang and Yi Yang. Decoupling features in hierarchical propagation for video object segmentation. In NeurIPS, 2022. +[43] Qian Yu, Xiaqi Zhao, Youwei Pang, Lihe Zhang, and Huchuan Lu. Multi-view aggregation network for dichotomous image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3921-3930, 2024. +[44] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023. +[45] Yuhui Yuan, Jingyi Xie, Xilin Chen, and Jingdong Wang. Segfix: Model-agnostic boundary refinement for segmentation. In ECCV, 2020. +[46] Yi Zeng, Pingping Zhang, Jianming Zhang, Zhe Lin, and Huchuan Lu. Towards high-resolution salient object detection. In ICCV, 2019. +[47] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017. \ No newline at end of file diff --git a/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/images.zip b/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2e418cc1b492d744c07798e54d32a1b6a355618a --- /dev/null +++ b/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:255d28b73352b38fe171006f316231851d7429fb6084bb747a1bd6a8730d3234 +size 719110 diff --git a/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/layout.json b/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8d05db343256522eb41de79a2e4bd09404ce48bc --- /dev/null +++ b/2025/Towards Fine-grained Interactive Segmentation in Images and Videos/layout.json @@ -0,0 +1,7674 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 520, + 121 + ], + "type": "text", + "content": "Towards Fine-grained Interactive Segmentation in Images and Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 166, + 144, + 444, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 144, + 444, + 159 + ], + "spans": [ + { + "bbox": [ + 166, + 144, + 444, + 159 + ], + "type": "text", + "content": "Yuan Yao Qiushi Yang Miaomiao Cui Liefeng Bo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 237, + 163, + 373, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 163, + 373, + 177 + ], + "spans": [ + { + "bbox": [ + 237, + 163, + 373, + 177 + ], + "type": "text", + "content": "Tongyi Lab, Alibaba Group" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 122, + 178, + 485, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 178, + 485, + 190 + ], + "spans": [ + { + "bbox": [ + 122, + 178, + 485, + 190 + ], + "type": "text", + "content": "{ryan.yy, yangqiushi.yqs, miaomiao.cmm, liefeng.bo}@alibaba-inc.com" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 198, + 555, + 408 + ], + "blocks": [ + { + "bbox": [ + 58, + 198, + 555, + 408 + ], + "lines": [ + { + "bbox": [ + 58, + 198, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 58, + 198, + 555, + 408 + ], + "type": "image", + "image_path": "bb0427cde5fe37f9e5e300d8ec466acfa54465046eb51e840600f6413185aa1c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 412, + 555, + 434 + ], + "lines": [ + { + "bbox": [ + 55, + 412, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 555, + 434 + ], + "type": "text", + "content": "Figure 1. Segmentation results of the proposed framework SAM2Refiner and previous methods. The proposed SAM2Refiner achieves the best performance on both image and video segmentation, especially in fine-grained details." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 454, + 200, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 454, + 200, + 467 + ], + "spans": [ + { + "bbox": [ + 152, + 454, + 200, + 467 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 486, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 297, + 714 + ], + "type": "text", + "content": "The recent Segment Anything Models (SAMs) have emerged as foundational visual models for general interactive segmentation. Despite demonstrating robust generalization abilities, they still suffer from performance degradations in scenarios that demand accurate masks. Existing methods for high-precision interactive segmentation face a trade-off between perceiving intricate local details and maintaining stable prompting capability, which hinders the applicability and effectiveness of foundational segmentation models. In this paper, we present a SAM2Refiner framework built upon the SAM2 backbone. This architecture allows SAM2 to generate fine-grained segmentation masks for both images and videos while preserving its inherent strengths. Specifically, we design a localization augment module, which incorporates local contextual cues to enhance global features via a cross-attention mechanism, thereby exploiting potential detailed patterns while maintaining semantic information. Moreover, to strengthen the prompting ability toward the enhanced object embeddings, we introduce a prompt" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 456, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 563 + ], + "type": "text", + "content": "retargeting module that renews the embedding with spatially aligned prompt features. In addition, to obtain accurate high resolution segmentation masks, a mask refinement module is devised by employing a multi-scale cascaded structure to fuse mask features with hierarchical representations from the encoder. Extensive experiments demonstrate the effectiveness of our approach, revealing that the proposed method can produce highly precise masks for both images and videos, surpassing state-of-the-art methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 586, + 395, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 395, + 599 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 395, + 599 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Interactive segmentation is a long-standing computer vision task that allows users to segment objects from the guidance of user interaction, typically in the form of clicks, strokes, or bounding boxes. In recent years, the Segment Anything Model (SAM) [16] has emerged as a foundational visual model for general interactive segmentation. Trained on their proposed SA-1B dataset, which consists of more than 1 billion masks and 11 million images, SAM can generate masks in various scenes and has been widely applied" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22509" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "to a wide range of downstream vision tasks, including image editing [37, 44], 3D vision [31, 39] and medical image analysis [23, 36]. The following work Segment Anything Model 2 (SAM2) [29] further introduces the streaming memory mechanism to extend the ability of SAM to videos, providing a potential boost to video applications such as AR / VR, robotics and autonomous vehicles." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 294, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 294, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 294, + 302 + ], + "type": "text", + "content": "Despite demonstrating robust generalization abilities, SAMs encounter the problems in scenario demanding fine-grained segmentation. In particular, as shown in Figure 1, they struggle to predict a complete mask for objects with fine structures or sharp boundaries, leading to overly smooth results or coarse boundaries. Thus, these types of failure may hinder the applicability of SAMs. Valuable efforts [14, 20, 23, 29] have been devoted to solve these challenges by introducing extra input tokens and explicitly fusing the multi-level features, however, they are still limited by SAMs' ability, thus failing to achieve satisfactory fine-grained segmentation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 304, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 295, + 651 + ], + "type": "text", + "content": "To this end, we aim to start with a holistic analysis on the three key points of fine-grained segmentation toward mining detailed cues and global semantic information guided by the user prompts. Firstly, considering that the input image contains global information of semantics while it lacks the detailed representations on details, we are expected to dynamically harvest the complementary local and global features within the input image. Secondly, SAMs directly combine the prompt with the object features via the decoder, which fail to build strong correlation among prompts with fine-grained cues and are hard to activate the response on object details, producing biased segmentation. Hence, strengthening the correspondences between prompt features and object details is crucial to yield a high response intensity of prompts on local details, which is beneficial to fine-grained segmentation. Thirdly, the object features with low resolution projected by the SAMs' encoder may lose much detailed cues, to recover the missed fine-grained local patterns for yielding high-resolution segmentation maps, the multi-scale features within SAMs' encoder are important to be re-used. A cocurrent research Pi-SAM [20] is the most similar work to ours, which proposes an additional mask decoder to expand the capability of SAM in high-resolution images. Nevertheless, it compromises with diminished responsiveness to visual prompt, necessitating a complementary module for error correction, and exhibit biased preferences for background clicks. Differently, our goal is to facilitate prompt response and consider both local and global presentations for efficient fine-grained segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 294, + 713 + ], + "type": "text", + "content": "To achieve the aforementioned goals, we propose the SAM2Refiner architecture, which includes three components for fine-grained interactive segmentation. Specifically, we introduce Localization Augment (LA) to refine the details of global features with local contextual feature" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 347 + ], + "type": "text", + "content": "cues. We split the input image into four sub-images, and send to SAM2's image encoder together with raw image. A global-local cross-attention mechanism is designed to highlight the local information of the object within the global representation among deep features. Moreover, to improve the response intensity of prompting for segmenting objects with intricate structures, we propose a Prompt Retargeting (PR) strategy by providing precise alignment between the prompts and enhanced object features. PR performs spatial alignment between prompt features and object features, and then incorporates them to yield an enhanced prompt response. In addition, to generate higher resolution masks with finer details, the Mask Refinement (MR) module is devised by incorporating the object features and multi-scale features from the original image. In MR, the intermediate features within the SAM2 encoder is re-used to fully preserve the original zero-shot capabilities. Although designed for image segmentation, the SAM2-based architecture enables native interoperability with video streaming workflows, allowing fine-grained segmentation for both images and videos (as in Figure 1). Extensive experiments on both image and video segmentation benchmarks demonstrate the effectiveness of our approach over state-of-the-art methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 324, + 348, + 547, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 348, + 547, + 360 + ], + "spans": [ + { + "bbox": [ + 324, + 348, + 547, + 360 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 361, + 553, + 612 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 361, + 553, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 553, + 421 + ], + "type": "text", + "content": "- We propose SAM2Refiner, a new architecture capable of generating fine-grained segmentation masks with strong perception for both global semantic and local details, and simultaneously maintaining stable prompting ability and powerful zero-shot capability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 422, + 553, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 553, + 480 + ], + "type": "text", + "content": "- In SAM2Refiner, the localization augment module is specifically designed to achieve a holistic representation of input image. By incorporating sub-images as additional inputs, it effectively enhances the global features through improved object localization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 481, + 553, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 481, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 314, + 481, + 553, + 529 + ], + "type": "text", + "content": "- To boost the responsiveness of prompting in intricate regions, we propose a prompt retargeting module that establishes precise alignment between spatially-aware dense prompt features and the enhanced object features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 529, + 553, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 529, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 314, + 529, + 553, + 564 + ], + "type": "text", + "content": "- To produce high-resolution segmentation maps with fine details, the mask refinement module is crafted to integrate multi-scale representations with object features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 565, + 553, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 565, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 314, + 565, + 553, + 612 + ], + "type": "text", + "content": "- Building upon the SAM2 backbone, our architecture exhibits strong compatibility to the video streaming pipeline, exhibiting state-of-the-art fine-grained segmentation on both images and videos benchmarks." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 626, + 400, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 626, + 400, + 639 + ], + "spans": [ + { + "bbox": [ + 314, + 626, + 400, + 639 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 647, + 454, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 647, + 454, + 659 + ], + "spans": [ + { + "bbox": [ + 314, + 647, + 454, + 659 + ], + "type": "text", + "content": "2.1. Interactive Segmentation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "Interactive segmentation enables users to provide cues for target regions and guide the segmentation process. User interactions can take various forms, including scribbles [14, 16, 28, 35], bounding boxes [14, 16, 20], clicks [16," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22510" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 396 + ], + "type": "text", + "content": "45], or language prompts [35, 40]. Traditional methods utilize pixel-level energy minimization methods, capturing low-level appearance features through unary potentials and ensuring consistent segmentation results with pairwise ones [33]. In recent years, many studies leverage prompts as input features and inject them into the models to produce segmentation results [10, 16, 28, 40]. The Segment Anything Model (SAM) [16] pre-trained on a large-scale datasets has emerged as a benchmark for interactive segmentation. Considering that SAM cannot perform semantic prediction, many works [10, 40] utilize category labels to further fine-tune SAM and enable it to achieve the semantic segmentation task. For example, SEEM [40] further trains SAM using labeled segmentation data through a bipartite matching constraint, thus providing the model with the ability to predict semantics. Semantic-SAM [10] proposes a multi-choice learning scheme via multi-task training across diverse datasets, enabling the model to segment at various granularities while predicting semantic labels. Other works [23, 24, 39] apply SAM to diverse specific domains to solve application tasks. MedSAM [23] finetunes SAM by collecting a large-scale dataset of medical image samples in multiple imaging modalities and types of cancer, achieving superior results in disease segmentation. Gaussian-Grouping [24] leverages the SAM as supervision model to jointly train a reconstruction and segmentation of objects in open-world 3D scenes for high-quality 3D edit." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 407, + 205, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 407, + 205, + 420 + ], + "spans": [ + { + "bbox": [ + 57, + 407, + 205, + 420 + ], + "type": "text", + "content": "2.2. High-Quality Segmentation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 427, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 427, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 427, + 294, + 712 + ], + "type": "text", + "content": "High-quality segmentation as a fundamental computer vision task focuses on accurately segmenting various complex and detail-rich objects, which ranges from many subtasks including semantic segmentation [22, 47], instance segmentation [9, 12] and panoptic segmentation [15]. Traditional methods [5, 7, 18, 27, 34] on high-quality segmentation dedicate to designing elaborate structures upon on CNN-based networks in order to capture both low-level local patterns and semantic global features through various reception fields towards holistic representation. Deplabv3 [18] designs a set of multi-level convolution kernels to capture features with diverse reception fields and HRNet [34] employs cascaded structures to fuse features from different layers to learn both low-level and high-level information. More recently, many approaches [2, 3, 6, 35, 38] built on the transformer backbone focus on introducing local patterns into the model by narrow the reception field of self-attention module. For example, Mask2Former [6] presents an efficient versatile architecture by using masked attention to extract local features for high-quality image segmentation tasks. Swin-Unet [3] adopts the SwinTransformer network to the UNet like structures to adaptively capture object features with various scales via hierarchical attention. Latest, the Pi-SAM [20] designs a mask" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 316, + 73, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 107 + ], + "type": "text", + "content": "decoder to expands SAMs ability at high resolution images, however, it may diminish the response of the input prompt and exhibit biased preferences for background clicks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 119, + 368, + 130 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 368, + 130 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 368, + 130 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 140, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 281 + ], + "type": "text", + "content": "We propose SAM2Refiner, which effectively generates fine-grained segmentation masks with strong perception for both global semantic and local details, and simultaneously maintaining stable prompting ability. In Sec 3.1, we first briefly review the architecture of SAM and SAM2. Then, in Sec 3.2, we introduce the three modules proposed upon SAM2's backbone to effectively improve the precision for segmentation. Finally, in Sec 3.3, we describe the training and inference process of the method, we also depict how our method can be seamlessly inserted into SAM2's video streaming pipeline for zero-shot inference. The core framework of our SAM2Refiner is shown in Figure 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 289, + 435, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 289, + 435, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 289, + 435, + 300 + ], + "type": "text", + "content": "3.1. Preliminaries: SAMs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 307, + 553, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 307, + 553, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 307, + 553, + 402 + ], + "type": "text", + "content": "SAM consists of an image encoder, a prompt encoder, and a mask decoder. The image encoder works for image feature extraction with a ViT [1] backbone. The prompt encoder is employed to encode the points, boxes, or masks into prompt tokens. The mask decoder utilizes a two-layer transformer-based decoder that takes both the extracted image embedding with the concatenated output and prompt tokens for final mask prediction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 403, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 403, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 316, + 403, + 553, + 605 + ], + "type": "text", + "content": "SAM2 extends SAMs ability to support both images and videos. For image setting, the three modules are largely follow SAM except for two differences, firstly, SAM2 adopt an MAE pretrained Hiera [30] image encoder, which is hierarchical compared with the plane structure in SAM's encoder, and secondly, to solve ambiguity in images without valid object, SAM2 predicts an additional token in mask decoder to indicate whether the object of interest is present. SAM2 introduces additional memory modules to support video streaming, including memory attention, memory encoder, and memory bank. The role of memory attention is to condition the current frame features on the past frame features and predictions as well as on any new prompts. The memory encoder is employed to obtain a memory embedding from masks generated from mask decoder, these memory embeddings are appended to a memory bank in a first-in-first-out (FIFO) way." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 613, + 432, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 432, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 432, + 624 + ], + "type": "text", + "content": "3.2. Ours: SAM2Refiner" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 631, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 631, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 631, + 553, + 712 + ], + "type": "text", + "content": "In this section, we describe the architecture of SAM2Refiner, which is built upon SAM2's backbone. This design offers two key advantages, 1) the hierarchical image encoder facilitates multi-scale feature capturing, which aids in feature refinement. 2) it allows for flexible integration into the video streaming pipeline. The proposed architecture comprises three modules, Localization Augment (LA)" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 316, + 757 + ], + "type": "text", + "content": "22511" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 547, + 342 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 547, + 342 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 547, + 342 + ], + "type": "image", + "image_path": "e42ad3091e8e7a20164e5e90f639d4f01846ef01642c3bfaf3ac1275c3fc2d01.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 353, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 54, + 353, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 555, + 399 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed SAM2Refiner framework. It contains a localization augment module to balance the localization details and global semantic representations, a prompt retargeting module to enhance responsiveness towards input prompts, and a mask refinement module to boost the quality of mask outcomes. The blue line denotes the SAM2[28] pipeline, and the black line denotes the proposed SAM2Refiner pipeline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 420, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 420, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 420, + 296, + 506 + ], + "type": "text", + "content": "is inserted into the middle layer of image encoder and mask decoder to refine the details of global features with local contextual feature cues. Prompt Retargeting (PR) aligns the prompt feature with enhanced object embedding to provide flexible prompting. Mask Refinement (MR) further incorporates object features and multi-scale features to generate higher-resolution masks with finer details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 515, + 178, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 178, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 178, + 527 + ], + "type": "text", + "content": "3.2.1. Localization Augment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "text", + "content": "We evenly crop the input image " + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{B \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "text", + "content": " into four non-overlapping sub-images " + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s}_m\\}_{m=1}^4 \\in \\mathbb{R}^{B \\times 3 \\times \\frac{H}{2} \\times \\frac{W}{2}}" + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "text", + "content": ". To align with input resolution of SAM2's image encoder, we upsample the sub-images to the same resolution with " + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_m\\}_{m=1}^4 \\in \\mathbb{R}^{B \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{S}_m\\}_{m=1}^M" + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "text", + "content": " are then fed in batches to the image encoder and generate the highest-level feature maps " + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{F_i\\}_{i=1}^5 \\in \\mathbb{R}^{B \\times C \\times \\frac{H}{16} \\times \\frac{W}{16}}" + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "text", + "content": ". We then split the feature maps into global feature " + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "inline_equation", + "content": "G = F_1" + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "text", + "content": " and local feature " + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{L_m\\}_{m=1}^4 = \\{F_i\\}_{i=2}^5" + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "text", + "content": ". The local features contain fine-grained details of input images and are used to enhance global feature with detailed clues. The local features are first assembled into a unified concatenated feature " + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "inline_equation", + "content": "L \\in \\mathbb{R}^{B \\times C \\times \\frac{H}{8} \\times \\frac{W}{8}}" + }, + { + "bbox": [ + 55, + 530, + 296, + 714 + ], + "type": "text", + "content": " following their respective positions in the original image. We then apply a multi-granularity average pooling operation as [43] to cap" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 420, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 553, + 445 + ], + "type": "text", + "content": "ture important contextual feature cues with different scales. The pooling functions are denote as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 363, + 453, + 553, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 453, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 363, + 453, + 553, + 466 + ], + "type": "interline_equation", + "content": "P _ {i} = \\operatorname {A v g P o o l} _ {i} (L), i \\in \\{1, \\dots , n \\}. \\tag {1}", + "image_path": "bebdc750202b39b398af20a0e722ff795d0b469f32d141b1be1b8766e617254a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "\\mathrm{AvgPool}_i" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": "-th pooling function, " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": "-th feature map. We set " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": " in our experiments, and the receptive fields for each pooling function are set 2, 4, 8 respectively. The feature maps " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": " are then tokenized and concatenated into a multi-granularity local features " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "L_t" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": ". For global feature " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": ", we tokenized it as " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "G_t" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": ". We conduct global-local cross attention between " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "G_t" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "L_t" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": " to obtain the refined global feature " + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "inline_equation", + "content": "G_{re}" + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 369, + 577, + 553, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 577, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 369, + 577, + 553, + 590 + ], + "type": "interline_equation", + "content": "G _ {r e} = \\operatorname {C r o s s A t t n} \\left(G _ {t}, L _ {t}, L _ {t}\\right) \\tag {2}", + "image_path": "aa36ef9be8ab10540f31cbcfc514e503988df5ec5ad3a92550d03bcba6746199.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 597, + 555, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 555, + 645 + ], + "type": "text", + "content": "Before sending into attention layer, the positional encoding is added to the features to access critical geometric information. " + }, + { + "bbox": [ + 313, + 597, + 555, + 645 + ], + "type": "inline_equation", + "content": "G_{re}" + }, + { + "bbox": [ + 313, + 597, + 555, + 645 + ], + "type": "text", + "content": " is able to highlight the local information about the object within the global representation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 651, + 429, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 429, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 429, + 662 + ], + "type": "text", + "content": "3.2.2. Prompt Retargeting" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "We replace the global feature " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "G_{re}" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " as the input to the mask decoder. The output mask " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{B \\times 3 \\times h \\times w}" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " and object embedding " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "E \\in \\mathbb{R}^{B \\times C \\times h \\times w}" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " of the mask decoder therefore contain richer representations of the local details." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22512" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 545, + 145 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 545, + 145 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 545, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 545, + 145 + ], + "type": "image", + "image_path": "3c87b1bb6e0b674d522c8cb37e40b122c139998cd9023e3651ff593ae555401f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 148, + 545, + 232 + ], + "blocks": [ + { + "bbox": [ + 61, + 148, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 61, + 148, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 545, + 232 + ], + "type": "image", + "image_path": "03f1738bc194a4e26eeabd64570e1cf9b1984e745bbb6d1545d1169b479166ba.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 245, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 55, + 245, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 555, + 269 + ], + "type": "text", + "content": "Figure 3. A Comparison between SAM2 and SAM2Refiner. The proposed three modules in SAM2Refiner can be seamlessly embedded into SAM2's video streaming pipeline, allowing to support fine-grained interactive segmentation for both images and videos." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": "To enhance the discriminability of features in small patches, we introduce a modified version of multi-branch convolutional block RFB [21], which applies different kernels to achieve multi-size receptive fields. To be specific, we first employ the bottleneck structure in each branch, consisting of a " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " convolutional layer to decrease the channel size in object embedding. Secondly, we apply three parallelized blocks with " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "1 \\times n" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " plus a " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "n \\times 1" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " convolutional layer to each branch, where we set " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "n = 3,5,7" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": ", respectively. Thirdly, three branches are concatenated and encoded via a " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " convolutional layer. Subsequently, we obtain the augmented embedding " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "E_{a}" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " by adding the concatenated embedding and a shortcut embedding encoded from another " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " convolutional layer. Although the ability to perceive local details is significantly enhanced, " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "E_{a}" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " exhibits reduced sensitivity to the input visual prompts. To maintain a stable prompting ability towards the enhanced embedding, we employ a three-channel prompt map to align the prompt feature with the augmented object embedding " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "E_{a}" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": ". The map resolution is set as " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " to reduce the computational cost and only two visual prompts ( clicks and masks) are encoded. We represent clicks as a fixed size radius, where positive clicks are encoded in the first channel, and negative clicks are encoded in the second channel. The mask " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " is then concatenated into the third channel to form a three-channel dense prompt map " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "D_{m}" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": ". We propose a dense prompt encoder to transform the " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "D_{p}" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " to a dense prompt embedding " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "E_{p}" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": ". The dense prompt encoder consists of two " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " convolution layers followed by GELU activations and layer normalization, and one " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " convolution layer to map the channel dimension to 256. The dense prompt embedding " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "E_{p}" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " is added to " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "E_{a}" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "inline_equation", + "content": "E_{ap}" + }, + { + "bbox": [ + 56, + 298, + 296, + 706 + ], + "type": "text", + "content": " in an element-wise manner, followed by a retargeting modules with several attention operations as in SAM2's mask decoder. We re-use the sparse prompt" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 299, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 407 + ], + "type": "text", + "content": "embedding and output token from SAM2's mask decoder to provide positional and class information for the renewing of the object embedding. The retargeting module contains four steps, (1) self-attention on the tokens concatenated from the the sparse prompt embedding and output token; (2) cross-attention from token to " + }, + { + "bbox": [ + 313, + 299, + 555, + 407 + ], + "type": "inline_equation", + "content": "E_{ap}" + }, + { + "bbox": [ + 313, + 299, + 555, + 407 + ], + "type": "text", + "content": "; (3) MLP updated on token, (4) cross-attention from " + }, + { + "bbox": [ + 313, + 299, + 555, + 407 + ], + "type": "inline_equation", + "content": "E_{ap}" + }, + { + "bbox": [ + 313, + 299, + 555, + 407 + ], + "type": "text", + "content": " to token. The renewed object embedding allows to possess a higher sensitivity towards the input visual prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 421, + 418, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 418, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 418, + 431 + ], + "type": "text", + "content": "3.2.3. Mask Refinement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": "The final mask of SAM2 is directly interpolated from the output prediction of its mask decoder, with a resolution of " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": ". This interpolation process may inevitably damage the details of the output prediction, particularly in high-resolution images. Our mask refinement module is designed to generate higher-resolution masks while enhancing the result quality. We take the object embedding from the prompt retargeting module and the multi-scale features from SAM2's hierarchical features to predict the final mask. The encoder blocks in Figure 3 represent the sub-modules within the image encoder, which generate multi-scale features with resolutions " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": ", respectively. We incorporate the UNet-like structure to fuse the object features and multi-scale image features. Each of decoder blocks contains an upsampling layer and a convolutional layer, and the upsampling feature along with image feature are concatenated via the channel dimension before sending into the convolutional layer. After the final decoder block, we conduct two transposed convolution layers to obtain mask feature with " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": " resolution. The low-level feature generated from input image " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": " is added further enhance image quality, and a convolutional layer is then used to generate the final segmentation map." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22513" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 95, + 555, + 236 + ], + "blocks": [ + { + "bbox": [ + 113, + 71, + 496, + 82 + ], + "lines": [ + { + "bbox": [ + 113, + 71, + 496, + 82 + ], + "spans": [ + { + "bbox": [ + 113, + 71, + 496, + 82 + ], + "type": "text", + "content": "Table 1. Results comparison with state-of-the-art methods on four fine-grained image segmentation tasks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 95, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 58, + 95, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 58, + 95, + 555, + 236 + ], + "type": "table", + "html": "
ModelDISCOIFTHRSODThinObjectAverage
mIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoUmIoUmBIoU
SAM (vit_l) [16]55.547.988.483.485.278.466.155.673.866.3
SAM (vit_h) [16]50.643.986.581.780.474.257.649.468.862.3
SAM2 (hiera_b+) [28]57.249.993.288.785.276.366.365.175.570.0
SAM2 (hiera_l) [28]57.650.792.488.086.379.577.867.778.571.5
HQ-SAM (vit_l) [14]77.469.294.489.791.984.687.076.687.780.0
HQ-SAM (vit_h) [14]77.969.794.489.591.983.588.578.688.280.3
SAM2Refiner (hiera_b+)83.478.596.593.394.189.495.190.192.387.8
SAM2Refiner (hiera_l)85.280.696.693.694.790.295.991.093.188.9
", + "image_path": "ab48040f9725d92bd7628ee07ced4047d786b81e359459b9339368a314790ca8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 249, + 186, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 249, + 186, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 186, + 262 + ], + "type": "text", + "content": "3.3. Training and Inference" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 266, + 294, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 294, + 434 + ], + "type": "text", + "content": "Training. We train our SAM2Refiner on image datasets only. We utilize the HQSeg-44K [14] as the training database, which includes six image datasets with extremely fine-grained mask labeling, including DIS [26](train set), ThinObject-5K [19] (train set), FSS-1000 [17], ECSSD [32], MSRA-10K [8], DUT-OMRON [41]. During training, we fix the model parameters of the pre-trained SAM2 model. We train by sampling mixed types of prompt including bounding boxes, randomly sampled points (either positive or negative), and coarse mask input. We use a learning rate of 0.001 and train for 20 epochs, with a learning rate decay schedule. We train on 8 NVIDIA V100 GPUs with a total batch size of 32. If not otherwise specified, we use the hierarchical large model as the default backbone." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 435, + 294, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 435, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 54, + 435, + 294, + 531 + ], + "type": "text", + "content": "Loss Function. Following HQ-SAM [14], we use a combination of binary cross-entropy (BCE) loss and dice loss for supervision. In addition to calculating the loss at the final prediction, we introduce intermediate supervision at each layer of the three decoder blocks. Before the loss calculation, each of the output features is sent to a separate convolutional layer to align the channel size with ground-truth mask. Our final loss function is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 539, + 295, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 539, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 79, + 539, + 295, + 572 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {f} \\left(\\mathcal {L} _ {f} ^ {b c e} + \\mathcal {L} _ {f} ^ {d i c e}\\right) + \\lambda_ {i} \\sum_ {i = 1} ^ {3} \\left(\\mathcal {L} _ {i} ^ {b c e} + \\mathcal {L} _ {i} ^ {d i c e}\\right) \\tag {3}", + "image_path": "b25b9656e2e9b462b49ae01eeba7cb40165cb14982ab81dbfe0c1e02d177278d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 581, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 286, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 581, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\lambda_{f}" + }, + { + "bbox": [ + 55, + 581, + 286, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 581, + 286, + 593 + ], + "type": "inline_equation", + "content": "\\lambda_{i}" + }, + { + "bbox": [ + 55, + 581, + 286, + 593 + ], + "type": "text", + "content": " are set to 1.0 and 0.3 in our experiment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": "Inference. For image setting, we follow the same inference pipeline as SAM. In the SAM2 temporal streaming pipeline, we extend our SAM2Refiner to the video segmentation by adding the cross-frame interaction via the memory bank. As suggested in Figure 3, compared to the original SAM2 pipeline, SAM2Refiner, including three proposed components, leverages the mask feature memory of previous frames to perform segmentation for the current frame, achieving accurate video segmentation. Note that for each frame except the first frame, the model uses the mem" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 249, + 555, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 555, + 274 + ], + "type": "text", + "content": "ory feature obtained by the first frame and the past nearest frames to perform frame-by-frame segmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 286, + 397, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 397, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 397, + 300 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 307, + 432, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 432, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 432, + 319 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 324, + 555, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 432 + ], + "type": "text", + "content": "Datasets. Following the prior work [14], we evaluate the model ability in both image and video segmentation tasks respectively. In image segmentation, all models are assessed in four challenging fine-grained segmentation benchmarks, including DIS-validation [26], ThinObject5K-test [19], COIFT [19] and HR-SOD [46]. For video object segmentation, we leverage two long-term video segmentation datasets, LVOS-v1 and LVOS-v2 [11], to evaluate our framework." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate the results of the image segmentation task, we echo previous works [14, 16] and adopt mIoU and mBIoU scores as metrics. For video object segmentation, we utilize " + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": " (region similarity), " + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": " (contour accuracy), and the combined " + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "inline_equation", + "content": "J \\& F" + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": " scores as metrics, where evaluations are performed in a semi-supervised setting with the first frame prompt given." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 527, + 463, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 463, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 463, + 540 + ], + "type": "text", + "content": "4.2. Comparisons and Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 605 + ], + "type": "text", + "content": "We perform extensive comparisons between our SAM2Refiner with SAM [16], SAM2 [28], HQ-SAM [14] and its video variant HQSAM2 on 6 benchmarks across image and video segmentation tasks. We do not compare Pi-SAM as they have not open-sourced their code." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Open-world Image Segmentation. To evaluate the segmentation results in the open-world environment, we compare our SAM2Refiner with previous methods on four image segmentation benchmarks. In Table 1, our SAM2Refiner with large backbone achieves state-of-the-art performance of " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "85.2\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "96.6\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "94.7\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "95.9\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " mIoU scores and " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "80.6\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "93.6\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "90.2\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "91.0\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " mBiOUscores on DIS, COIFT, HRSOD and ThinObject datasets, respectively. In average, SAM2Refiner with a smaller backbone" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22514" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 160, + 422 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 160, + 422 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 160, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 160, + 422 + ], + "type": "image", + "image_path": "77a26a816ed0cd89a468f98d396fade48a98acd20995cd424c7d424cb0c85c25.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 437, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 437, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 555, + 460 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with previous methods. Given the blue box as visual prompt, our proposed SAM2Refiner produces more accurate results with correct structure and clear boundaries. Zoom in for better visualization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 160, + 70, + 355, + 422 + ], + "blocks": [ + { + "bbox": [ + 160, + 70, + 355, + 422 + ], + "lines": [ + { + "bbox": [ + 160, + 70, + 355, + 422 + ], + "spans": [ + { + "bbox": [ + 160, + 70, + 355, + 422 + ], + "type": "image", + "image_path": "8d993c40f9776eaa24539639daac917a607656cd46ce50ac4a2f92dba284cfcd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 355, + 70, + 549, + 422 + ], + "blocks": [ + { + "bbox": [ + 355, + 70, + 549, + 422 + ], + "lines": [ + { + "bbox": [ + 355, + 70, + 549, + 422 + ], + "spans": [ + { + "bbox": [ + 355, + 70, + 549, + 422 + ], + "type": "image", + "image_path": "58474c52edd750e1c50ea3469ab5d9c3b0944bddf99e27890ee12b060fc48f7b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 490, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 622 + ], + "type": "text", + "content": "(hiera_1) performs " + }, + { + "bbox": [ + 55, + 490, + 296, + 622 + ], + "type": "inline_equation", + "content": "93.1\\%" + }, + { + "bbox": [ + 55, + 490, + 296, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 490, + 296, + 622 + ], + "type": "inline_equation", + "content": "88.9\\%" + }, + { + "bbox": [ + 55, + 490, + 296, + 622 + ], + "type": "text", + "content": " average mIoU and mBiOU scores, significantly outperforming the second-best one HQ-SAM (vit_h) with " + }, + { + "bbox": [ + 55, + 490, + 296, + 622 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 55, + 490, + 296, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 490, + 296, + 622 + ], + "type": "inline_equation", + "content": "8.6\\%" + }, + { + "bbox": [ + 55, + 490, + 296, + 622 + ], + "type": "text", + "content": " scores. To clear represent the segmentation results, we show the comparison of the visualization results of segmentation maps. From Figure 4, our SAM2Refiner can produce much more accurate segmentation maps than other methods, specifically on detailed objects such as boundary and small regions. These results demonstrate that the SAM2Refiner benefitting from the proposed modules achieves superior fine-grained image segmentation and strong generalization ability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Zero-Shot Video Object Segmentation. To evaluate the ability of the proposed method on video segmentation, we extend our model on semi-supervised video object segmentation task upon the pipeline of SAM2. Towards the zero-shot segmentation in the open-world scenario, we compare our model with previous video segmentation works in two benchmarks, including LVOS v1 and LVOS v2. As sug-" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 520, + 553, + 643 + ], + "blocks": [ + { + "bbox": [ + 313, + 489, + 555, + 511 + ], + "lines": [ + { + "bbox": [ + 313, + 489, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 555, + 511 + ], + "type": "text", + "content": "Table 2. Comparison of the proposed method with prior works on video object segmentation task." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 520, + 553, + 643 + ], + "lines": [ + { + "bbox": [ + 315, + 520, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 315, + 520, + 553, + 643 + ], + "type": "table", + "html": "
MethodsLVOS v1LVOS v2
J&FJFJ&FJsFsJuFu
LWL [13]56.451.860.960.658.064.357.262.9
CFBI [13]51.546.256.755.052.959.251.756.2
STCN [13]48.943.954.060.657.264.057.563.8
RDE [25]53.748.359.262.256.764.160.867.2
SwinB-DeAOT-L [42]---63.961.569.058.466.6
XMem [4]52.948.157.764.562.669.160.665.6
SAM2 [28]77.272.282.278.375.287.068.579.4
HQ-SAM2 [14]79.674.984.379.176.588.369.482.2
SAM2Refiner80.975.186.879.978.787.971.681.5
", + "image_path": "6f731a00f27e4cd68b8e500bdc828a3c1b11df40da415b1e18279dd85773ed86.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "gested in Table 2, our model yields " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "80.9\\%" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "75.1\\%" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "86.8\\%" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": " scores of " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "79.9\\%" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "78.7\\%" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "87.9\\%" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "71.6\\%" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "81.5\\%" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{J}_s" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_s" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{J}_u" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_u" + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": " in two datasets respectively, significantly outperforming over previous compar" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22515" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 99, + 296, + 205 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 94 + ], + "type": "text", + "content": "Table 3. Analysis of the point prompt response in interactive image segmentation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 99, + 296, + 205 + ], + "lines": [ + { + "bbox": [ + 58, + 99, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 99, + 296, + 205 + ], + "type": "table", + "html": "
Point Num.MetricsSAMSAM2HQ-SAMSAM2Refiner
1mIoU8.711.417.725.2
mB IoU6.39.914.320.6
2mIoU30.835.644.652.5
mB IoU26.231.740.247.4
5mIoU70.173.984.290.0
mB IoU62.366.475.584.9
10mIoU71.776.786.191.6
mB IoU64.870.478.588.6
", + "image_path": "1437c48a0ebd0cf3a9c4accc5638be7eeca039f70507c617fcd55088cd6bfd68.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 69, + 266, + 282, + 362 + ], + "blocks": [ + { + "bbox": [ + 55, + 216, + 295, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 295, + 261 + ], + "type": "text", + "content": "Table 4. Ablation study on the three modules of SAM2Refiner, including Localization Augment (LA), Prompt Retargeting (PR) and Mask Refinement (MR). The results are averaged over the four evaluated datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 266, + 282, + 362 + ], + "lines": [ + { + "bbox": [ + 69, + 266, + 282, + 362 + ], + "spans": [ + { + "bbox": [ + 69, + 266, + 282, + 362 + ], + "type": "table", + "html": "
LAPRMRmIoUmBIoU
SAM2 [29]78.571.5
SAM2Refiner✓85.067.4
✓✓91.186.0
✓✓92.788.3
✓✓91.486.2
✓✓✓93.188.9
", + "image_path": "8f07ed50e28be5eb1f101a1fcea00eef1749bb457032cdf54e35dd738a01fbda.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 376, + 295, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 424 + ], + "type": "text", + "content": "isons [4, 13, 14, 25, 28, 42]. These results demonstrate that the proposed framework can not only enhance fine-grained image segmentation, but also boost the video segmentation with refined object details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 424, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 424, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 424, + 295, + 568 + ], + "type": "text", + "content": "Analysis on prompt number. To evaluate the response intensity to input prompts in SAM2Refiner compared to other models, we input varying numbers of point prompts into the model in image segmentation task. In Table 3, we observe that the average segmentation results decline as the number of prompt points decreases. However, our SAM2Refiner indicates consistent advantages in segmentation accuracy across varying numbers of point prompts, while SAM, SAM2 and HQ-SAM reflect relatively low results. This clearly suggests that the proposed method with the capability of prompt re-targeting can effectively achieve a strong response for input prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 575, + 149, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 149, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 149, + 588 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 654 + ], + "type": "text", + "content": "We conduct ablation experiments to evaluate the effectiveness of the proposed modules within our SAM2Refiner, i.e., Localization Augment (LA), Prompt Retargeting (PR) and Mask Refinement (MR), and we study the affect of the number of input prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "In the LA module, the original global features attend to local features from sub-images, thereby enriching the representation with more detailed local information. As shown in Table 4, compared to SAM2, the LA module markedly enhances the mIoU from " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "78.5\\%" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "85.0\\%" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ", despite a slight" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 395, + 148 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 395, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 395, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 395, + 148 + ], + "type": "image", + "image_path": "457a3a98be44b41efd4869141b4c86174f222a3b37f6c3ce459e6c2702888403.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 148, + 365, + 157 + ], + "lines": [ + { + "bbox": [ + 347, + 148, + 365, + 157 + ], + "spans": [ + { + "bbox": [ + 347, + 148, + 365, + 157 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 395, + 71, + 472, + 148 + ], + "blocks": [ + { + "bbox": [ + 395, + 71, + 472, + 148 + ], + "lines": [ + { + "bbox": [ + 395, + 71, + 472, + 148 + ], + "spans": [ + { + "bbox": [ + 395, + 71, + 472, + 148 + ], + "type": "image", + "image_path": "a041b8ec8f7ba3dc3e628221dcaad5bacb7920b9e1c0684f3205a7994d3ebf38.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 148, + 444, + 156 + ], + "lines": [ + { + "bbox": [ + 421, + 148, + 444, + 156 + ], + "spans": [ + { + "bbox": [ + 421, + 148, + 444, + 156 + ], + "type": "text", + "content": "w/o LA" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 472, + 71, + 550, + 148 + ], + "blocks": [ + { + "bbox": [ + 472, + 71, + 550, + 148 + ], + "lines": [ + { + "bbox": [ + 472, + 71, + 550, + 148 + ], + "spans": [ + { + "bbox": [ + 472, + 71, + 550, + 148 + ], + "type": "image", + "image_path": "72f001d2ceb68d11a35995e42c5618413045b845107cfac188be7149c9aa88a4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 148, + 520, + 156 + ], + "lines": [ + { + "bbox": [ + 501, + 148, + 520, + 156 + ], + "spans": [ + { + "bbox": [ + 501, + 148, + 520, + 156 + ], + "type": "text", + "content": "w/LA" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 200, + 395, + 252 + ], + "blocks": [ + { + "bbox": [ + 317, + 200, + 395, + 252 + ], + "lines": [ + { + "bbox": [ + 317, + 200, + 395, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 200, + 395, + 252 + ], + "type": "image", + "image_path": "1174d49dee81c4aff286576d051bc763d1ae850a2d66c6895b4c7954379251d6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 253, + 365, + 261 + ], + "lines": [ + { + "bbox": [ + 347, + 253, + 365, + 261 + ], + "spans": [ + { + "bbox": [ + 347, + 253, + 365, + 261 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 264, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 299 + ], + "type": "text", + "content": "Figure 6. Effectiveness of PR module. PR allows precise response for point prompts in ambiguous regions. Green point and red point denote positive and negative prompt, respectively." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 395, + 200, + 472, + 252 + ], + "blocks": [ + { + "bbox": [ + 395, + 200, + 472, + 252 + ], + "lines": [ + { + "bbox": [ + 395, + 200, + 472, + 252 + ], + "spans": [ + { + "bbox": [ + 395, + 200, + 472, + 252 + ], + "type": "image", + "image_path": "8328686e622b037acc63c011fc90bdf7a80a7cbf9f419e79491725c86cf64a51.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 253, + 449, + 260 + ], + "lines": [ + { + "bbox": [ + 424, + 253, + 449, + 260 + ], + "spans": [ + { + "bbox": [ + 424, + 253, + 449, + 260 + ], + "type": "text", + "content": "w/o PR" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 472, + 200, + 550, + 252 + ], + "blocks": [ + { + "bbox": [ + 313, + 160, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 160, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 555, + 194 + ], + "type": "text", + "content": "Figure 5. Effectiveness of LA module. LA suggest remarkable improvement on local details although it displays some flickering boundaries." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 472, + 200, + 550, + 252 + ], + "lines": [ + { + "bbox": [ + 472, + 200, + 550, + 252 + ], + "spans": [ + { + "bbox": [ + 472, + 200, + 550, + 252 + ], + "type": "image", + "image_path": "431ac431aa31c1ddcab1f5ac8e2eb2788c314d67ee81d050d1cd974b07d8d4d4.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 253, + 520, + 260 + ], + "lines": [ + { + "bbox": [ + 500, + 253, + 520, + 260 + ], + "spans": [ + { + "bbox": [ + 500, + 253, + 520, + 260 + ], + "type": "text", + "content": "W/PR" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 328, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 521 + ], + "type": "text", + "content": "decrease in boundary accuracy. This is because the LA module strengthens the perception of fine structures, but this can come at the cost of boundary precision. Figure 5 verifies this phenomenon, showing that results with LA exhibit more detailed local features but also show flickering boundaries. The second to fourth rows of the SAM2Refiner in Table 4 highlight the effectiveness of each individual module for fine-grained segmentation. The PR module allows the model to maintain strong sensibility towards point prompting in intricate regions. As shown in Figure 6, result with PR module is able to correctly respond to the input visual prompt. Compared to PR, both the LA and MR modules deliver greater performance improvements in mIoU and mBIoU, with about 2 points gain in mIoU and 3 points gain in mBIoU, underscoring their crucial roles in enhancing segmentation precision." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 530, + 388, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 388, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 388, + 544 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 551, + 555, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 555, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 555, + 707 + ], + "type": "text", + "content": "In this work, we propose SAM2Refiner to enhance interactive segmentation for images and videos. SAM2Refiner balances local detail perception and prompting stability, achieving fine-grained visual representation. By integrating local contextual cues through a cross-attention mechanism and employing a UNet-like structure for feature fusion, SAM2Refiner generates high-quality segmentation masks. In addition, a prompt retargeting module is designed to improve the alignment of object embeddings with spatial features. Extensive experiments demonstrate that our SAM2Refiner achieves superior segmentation performance on both image and video segmentation tasks, significantly outperforming other state-of-the-art methods." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22516" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "type": "text", + "content": "[1] Dosovitskiy Alexey. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv: 2010.11929, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 295, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 295, + 156 + ], + "type": "text", + "content": "[2] Alexander Kirillov Bowen Cheng, Alex Schwing. Per-pixel classification is not all you need for semantic segmentation. In NeurIPS, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 158, + 294, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 158, + 294, + 190 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 294, + 190 + ], + "type": "text", + "content": "[3] Wang Y. Chen J. Jiang D. Zhang X. Tian Q. Cao, H. and Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. In ECCV, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 192, + 294, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 294, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 294, + 223 + ], + "type": "text", + "content": "[4] Ho Kei Cheng and Alexander G Schwing. Xmem: Long-term video object segmentation with an atkinson-shiffrin memory model. In ECCV, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 294, + 267 + ], + "type": "text", + "content": "[5] Ho Kei Cheng, Jihoon Chung, Yu-Wing Tai, and Chi-Keung Tang. CascadePSP: Toward class-agnostic and very high-resolution segmentation via global and local refinement. In CVPR, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 270, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 294, + 312 + ], + "type": "text", + "content": "[6] Ishan Misra Alexander G. Schwing Alexander Kirillov Cheng, Bowen and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 314, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 294, + 357 + ], + "type": "text", + "content": "[7] Maxwell D. Collins Yukun Zhu Ting Liu Thomas S. Huang Hartwig Adam Cheng, Bowen and Liang-Chieh Chen. Panoptic-deeplab: A simple, strong, and fast baseline for bottom-up panoptic segmentation. In CVPR, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 358, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 358, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 358, + 294, + 390 + ], + "type": "text", + "content": "[8] Ming-Ming Cheng, Niloy J Mitra, Xiaolei Huang, Philip HS Torr, and Shi-Min Hu. Global contrast based salient region detection. TPAMI, 37(3):569-582, 2014." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 392, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 422 + ], + "type": "text", + "content": "[9] Jifeng Dai, Kaiming He, and Jian Sun. Instance-aware semantic segmentation via multi-task network cascades. In CVPR, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 425, + 295, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 425, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 57, + 425, + 295, + 457 + ], + "type": "text", + "content": "[10] Peize Sun Xueyan Zou Shilong Liu Chunyuan Li Jianwei Yang Lei Zhang Jianfeng Gao Feng Li, Hao Zhang. Segment and recognize anything at any granularity. In ECCV, 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 458, + 294, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 458, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 57, + 458, + 294, + 490 + ], + "type": "text", + "content": "[11] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 491, + 294, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 294, + 512 + ], + "type": "text", + "content": "[12] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 546 + ], + "type": "text", + "content": "[13] Yu-Wing Tai Ho Kei Cheng and Chi-Keung Tang. Rethinking space-time networks with improved memory coverage for efficient video object segmentation. In NeurIPS, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 547, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 294, + 579 + ], + "type": "text", + "content": "[14] Lei Ke, Mingqiao Ye, Martin Danelljan, Yu-Wing Tai, Chi-Keung Tang, Fisher Yu, et al. Segment anything in high quality. In NeurIPS, 2023." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 581, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 294, + 612 + ], + "type": "text", + "content": "[15] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólár. Panoptic segmentation. In CVPR, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 614, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 657 + ], + "type": "text", + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dolkar, and Ross Girshick. Segment anything. In ICCV, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "type": "text", + "content": "[17] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. Fss-1000: A 1000-class dataset for few-shot segmentation. In CVPR, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[18] Florian Schroff Hartwig Adam Liang-Chieh Chen, George Papandreou. Rethinking atrous convolution" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "for semantic image segmentation. In arXiv preprint arXiv:1706.05587, 2017." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "text", + "content": "[19] Jun Hao Liew, Scott Cohen, Brian Price, Long Mai, and Jiashi Feng. Deep interactive thin object selection. In WACV, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 132, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 132, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 132, + 553, + 175 + ], + "type": "text", + "content": "[20] Mengzhen Liu, Mengyu Wang, Henghui Ding, Yilong Xu, Yao Zhao, and Yunchao Wei. Segment anything with precise interaction. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3790-3799, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 177, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 177, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 177, + 553, + 220 + ], + "type": "text", + "content": "[21] Songtao Liu, Di Huang, et al. Receptive field block net for accurate and fast object detection. In Proceedings of the European conference on computer vision (ECCV), pages 385-400, 2018." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "type": "text", + "content": "[22] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 258, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 553, + 289 + ], + "type": "text", + "content": "[23] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images. Nature Communications, 15(1):654, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 292, + 553, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 292, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 292, + 553, + 324 + ], + "type": "text", + "content": "[24] Fisher Yu Mingqiao Ye, Martin Danelljan and Lei Ke. Gaussian grouping: Segment and edit anything in 3d scenes. In ECCV, 2024." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 327, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 358 + ], + "type": "text", + "content": "[25] Zhiwei Xiong Bang Zhang Pan Pan Mingxing Li, Liucheng Hu and Dong Liu. Recurrent dynamic embedding for video object segmentation. In CVPR, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 361, + 553, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 394 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 394 + ], + "type": "text", + "content": "[26] Xuebin Qin, Hang Dai, Xiaobin Hu, Deng-Ping Fan, Ling Shao, and Luc Van Gool. Highly accurate dichotomous image segmentation. In ECCV, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 396, + 553, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 396, + 553, + 428 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 553, + 428 + ], + "type": "text", + "content": "[27] Yixuan Yuan Qiushi Yang. Learning dynamic convolutions for multi-modal 3d mri brain tumor segmentation. In MIC-CAI Workshop, 2020." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 430, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 473 + ], + "type": "text", + "content": "[28] Gabeur V. Hu Y.T. Hu R. Ryali C. Ma T. Khedr H. Radle R. Rolland C. Gustafson L. Ravi, N. and E. Mintun. Sam2: Segment anything in images and videos. In arXiv preprint arXiv:2408.00714, 2024." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 476, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 476, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 476, + 553, + 529 + ], + "type": "text", + "content": "[29] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Radle, Chloe Rolland, Laura Gustafson, et al. Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714, 2024." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 533, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 553, + 598 + ], + "type": "text", + "content": "[30] Chaitanya Ryali, Yuan-Ting Hu, Daniel Bolya, Chen Wei, Haoqi Fan, Po-Yao Huang, Vaibhav Aggarwal, Arkabandhu Chowdhury, Omid Poursaeed, Judy Hoffman, et al. Hiera: A hierarchical vision transformer without the bells-and-whistles. In International Conference on Machine Learning, pages 29441-29454. PMLR, 2023." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 601, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 632 + ], + "type": "text", + "content": "[31] Qiuhong Shen, Xingyi Yang, and Xinchao Wang. Anything-3d: Towards single-view anything reconstruction in the wild. arXiv preprint arXiv:2304.10261, 2023." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 635, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 666 + ], + "type": "text", + "content": "[32] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended.cssd. TPAMI, 38(4): 717-729, 2015." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[33] Mengyu Wang, Henghui Ding, Jun Hao Liew, Jiajun Liu, Yao Zhao, and Yunchao Wei. Segrefiner: Towards model-agnostic segmentation refinement with discrete diffusion process. arXiv preprint arXiv:2312.12425, 2023." + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22517" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 622 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[34] Sun K. Cheng T. Jiang B. Deng C. Zhao Y. Liu D. Mu Y. Tan M. Wang X. Wang, J. and W. Liu. Deep high-resolution representation learning for visual recognition. In TPAMI, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[35] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[36] Junde Wu, Wei Ji, Yuanpei Liu, Huazhu Fu, Min Xu, Yanwu Xu, and Yueming Jin. Medical sam adapter: Adapting segment anything model for medical image segmentation. arXiv preprint arXiv:2304.12620, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "text", + "content": "[37] Defeng Xie, Ruichen Wang, Jian Ma, Chen Chen, Haonan Lu, Dong Yang, Fobo Shi, and Xiaodong Lin. Edit everything: A text-guided generative system for images editing. arXiv preprint arXiv:2304.14006, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[38] Wenhai Wang Zhiding Yu Anima Anandkumar Jose M. Alvarez Xie, Enze and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 275, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 318 + ], + "type": "text", + "content": "[39] Xiuwei Xu, Huangxing Chen, Linqing Zhao, Ziwei Wang, Jie Zhou, and Jiwen Lu. Embodiedsam: Online segment any 3d thing in real time. arXiv preprint arXiv:2408.11811, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 319, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 353 + ], + "type": "text", + "content": "[40] Hao Zhang Feng Li Linjie Li Jianfeng Wang Lijuan Wang Jianfeng Gao Yong Jae Lee Xueyan Zou, Jianwei Yang. Segment everything everywhere all at once. In NeurlPS, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "text", + "content": "[41] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In CVPR, 2013." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "text", + "content": "[42] Zongxin Yang and Yi Yang. Decoupling features in hierarchical propagation for video object segmentation. In NeurIPS, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 476 + ], + "type": "text", + "content": "[43] Qian Yu, Xiaqi Zhao, Youwei Pang, Lihe Zhang, and Huchuan Lu. Multi-view aggregation network for dichotomous image segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3921-3930, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[44] Tao Yu, Runseng Feng, Ruoyu Feng, Jinming Liu, Xin Jin, Wenjun Zeng, and Zhibo Chen. Inpaint anything: Segment anything meets image inpainting. arXiv preprint arXiv:2304.06790, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "type": "text", + "content": "[45] Yuhui Yuan, Jingyi Xie, Xilin Chen, and Jingdong Wang. Segfix: Model-agnostic boundary refinement for segmentation. In ECCV, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 556, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 588 + ], + "type": "text", + "content": "[46] Yi Zeng, Pingping Zhang, Jianming Zhang, Zhe Lin, and Huchuan Lu. Towards high-resolution salient object detection. In ICCV, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "text", + "content": "[47] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In CVPR, 2017." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22518" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Foundational Models for Single-Chip Radar/6536d138-fd2a-4e4b-a53a-8d73d592cae1_content_list.json b/2025/Towards Foundational Models for Single-Chip Radar/6536d138-fd2a-4e4b-a53a-8d73d592cae1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..437c8ea9630cf9e163d12b921a0f6da42d624eee --- /dev/null +++ b/2025/Towards Foundational Models for Single-Chip Radar/6536d138-fd2a-4e4b-a53a-8d73d592cae1_content_list.json @@ -0,0 +1,1659 @@ +[ + { + "type": "text", + "text": "Towards Foundational Models for Single-Chip Radar", + "text_level": 1, + "bbox": [ + 228, + 130, + 769, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tianshu Huang $^{12}$ Akarsh Prabhakara $^{3}$ Chuhan Chen $^{1}$ Jay Karhade $^{1}$", + "bbox": [ + 181, + 178, + 813, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deva Ramanan1 Matthew O'Toole1 Anthony Rowe12", + "bbox": [ + 272, + 199, + 759, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University $^{2}$ Bosch Research $^{3}$ University of Wisconsin-Madison", + "bbox": [ + 147, + 215, + 849, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tianshu2, chuhanc, jkarhade, deva, motoole2, agr}@andrew.cmu.edu, akarsh@cs.wisc.edu", + "bbox": [ + 120, + 236, + 870, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/137c334653dcbcb903d389abe4bdb14e787b23ab3e5147ea3310e919bb2f800e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 258, + 784, + 366 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4fa09d54a8d0a93b1fef2488c4830a391eda11fca1a3efa1cc131284d937c450.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 368, + 370, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ac22028b3e3a658dd2784993604a0382e85469ce80fac6957332e76c96f105d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 369, + 637, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba52efe88dde9657d76fc8d095960a31ea2c4476563ec5a53da98e49f96916a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 369, + 903, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92eab0b4d91da4461c7e13b0f527f5c004dd85a8d2da5b54af5f6789ad73c027.jpg", + "image_caption": [ + "Figure 1. Using only 4D range-Doppler-azimuth-elevation data from a radar with a 3x4 antenna array - equivalent to a 0.26-megapixel camera, we trained (separate) radar transformers on 24 hours (1M radar-lidar-camera samples) of data to predict Bird's Eye View 2D occupancy (left), 3D occupancy (center), semantic segmentation (right), and Ego-Motion." + ], + "image_footnote": [], + "bbox": [ + 91, + 470, + 370, + 571 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0f467d7ce8d354ce9c9465777138412162cd3fd6a4886be26af56110494add1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 470, + 637, + 570 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a7eb955b38d8294b3ed2e64dd8a2da904f74849b8e939066d3535549593dcf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 470, + 903, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 638, + 326, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mmWave radars are compact, inexpensive, and durable sensors that are robust to occlusions and work regardless of environmental conditions, such as weather and darkness. However, this comes at the cost of poor angular resolution, especially for inexpensive single-chip radars, which are typically used in automotive and indoor sensing applications. Although many have proposed learning-based methods to mitigate this weakness, no standardized foundational models or large datasets for the mmWave radar have emerged, and practitioners have largely trained task-specific models from scratch using relatively small datasets.", + "bbox": [ + 88, + 672, + 482, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we collect (to our knowledge) the largest available raw radar dataset with 1M samples (29 hours) and train a foundational model for 4D single-chip radar, which can predict 3D occupancy and semantic segmenta", + "bbox": [ + 88, + 839, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion with quality that is typically only possible with much higher resolution sensors. We demonstrate that our Generalizable Radar Transformer (GRT) generalizes across diverse settings, can be fine-tuned for different tasks, and shows logarithmic data scaling of $20\\%$ per $10\\times$ data. We also run extensive ablations on common design decisions, and find that using raw radar data significantly outperforms widely-used lossy representations, equivalent to a $10\\times$ increase in training data. Finally, we roughly estimate that $\\approx 100M$ samples (3000 hours) of data are required to fully exploit the potential of GRT.", + "bbox": [ + 511, + 641, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 829, + 645, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As a compact, inexpensive [27], and robust solid-state sensor, mmWave radars are ideal for sensing applications ranging from simple automatic door openers [58] to autonomous", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24655", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "drones [11] or vehicles [53, 62]. mmWave radars are rich sensors which can directly measure range and velocity while capturing a unique range of material properties [18]; however, this comes at the cost of poor angular resolution typically on the order of $15^{\\circ}$ - orders of magnitude worse than cameras or lidars [59].", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Radar data are typically processed into radar point clouds (Fig. 2) derived using Constant False Alarm Rate (CFAR) peak detectors [36, 52] combined with Angle-of-Arrival estimation techniques [60]. However, this is a substantially lossy process: while raw radar data suffers from unique noise patterns such as \"bleed\" and side lobes [25], weak reflectors and other signals can be hidden in this noise, which would ordinarily be filtered out.", + "bbox": [ + 89, + 183, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, raw spectrum (4D range-Doppler-azimuth-elevation data cubes [50]) can be unintuitive and difficult to interpret compared to lidar point clouds or camera images, and include properties such as specularity and Doppler which lack straight-forward Cartesian interpretations [18]. As such, many machine learning methods have been proposed [24, 45, 69] to exploit 4D radar data from single-chip radars, achieving remarkable performance on 2D scene understanding tasks. However, due to the dominance of CFAR point clouds in radar processing, as well as the high data rate of raw mmWave radar data, most radar toolchains only process point-cloud data. Tooling for raw I/Q (in-phase/quadrature) data is often brittle, poorly documented, and largely unsupported by radar vendors, severely limiting the availability and scale of both raw mmWave datasets and the models which operate on raw data.", + "bbox": [ + 89, + 305, + 482, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To rectify this limitation, we develop an open-source toolchain and associated large dataset specifically for 4D mmWave radar data. Training a radar-to-lidar model and fine-tuning for a range of other tasks, we demonstrate the surprising effectiveness of mmWave radar models trained at scale (Fig. 1). Going further, just as large foundational models [4] have greatly accelerated the pace of innovation in computer vision and natural language processing, we believe that a foundational model for raw mmWave radar trained at even larger scale could similarly supercharge the advancement of radar sensing techniques.", + "bbox": [ + 89, + 547, + 482, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions In this paper, we develop a full stack1 for collecting data, training, and evaluating a transformer for 4D single-chip radar to quantify both the potential costs and benefits of training a foundational model at scale. To summarize our contributions:", + "bbox": [ + 89, + 720, + 482, + 796 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1) We develop a compact, lightweight multimodal data collection system (Sec. 3.1) capable of collecting synchronized raw radar, Lidar, and camera data which can be operated as a handheld device. Our system can be", + "bbox": [ + 89, + 799, + 482, + 859 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/156b3552d8b134b3ac1dce53b706acf6043173047e707acceb81d3aac7feb320.jpg", + "image_caption": [ + "Figure 2. While a transformer can generate Lidar-like depth and bird's eye view images, traditional CFAR point clouds are noisy, and have poor angular resolution - especially in the elevation axis." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "easily replicated using off-the-shelf components, 3D printed parts, and our open-source software.", + "bbox": [ + 539, + 330, + 903, + 361 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(2) Using this data collection system, we collect a dataset, I/Q-1M (One Million IQ Frames), consisting of 29 hours of data $-8 \\times$ longer than the next largest publicly available raw radar dataset - split between indoor, outdoor-handheld, and bike-mounted settings, each with different radar configurations (Sec. 3.2).", + "(3) Finally, using our dataset, we train a Generalizable Radar Transformer (GRT), which can output depth maps and segmentation images with quality which is typically only possible with much higher resolution radars. Using GRT, we then run ablations on common design choices (Sec. 5.1), quantify the scalability of GRTs with increasing dataset and model size (Sec. 5.2, 5.4), and demonstrate that our GRT can be readily finetuned for other tasks and settings (Sec. 5.3), including obtaining state-of-the-art performance on the Coloradar [23] dataset with 30-minutes of fine-tuning." + ], + "bbox": [ + 514, + 362, + 905, + 618 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Key Findings We summarize our key findings as follows:", + "bbox": [ + 513, + 641, + 903, + 656 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Radar models can generalize to different settings and radar configurations (Sec. 5.2), as well as across objectives (with some fine-tuning). This suggests great potential for a cross-domain foundational model to improve and accelerate the development of new radar models.", + "- Using raw data yields outsized performance gains, equivalent to more than a $10 \\times$ increase in training data (Sec. 5.1). While existing datasets largely focus on CFAR point clouds or other processed representations, we believe that more emphasis should be placed on making raw data available for research.", + "- Existing mmWave radar datasets are vastly undersized. 24 hours of training data is not enough to saturate even a 4M parameter model! Our analysis suggests that at least $100 \\times$ more data is required to exploit the full potential of radar transformer models (Sec. 5.4)." + ], + "bbox": [ + 514, + 659, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1Our data collection system, dataset, code, and model can be found via our project site: https://wiselabcmu.github.io/grt/.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24656", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e21097ab446e6549c043536819c3baccf2cede017d87b679cf6a344f43d96f05.jpg", + "image_caption": [ + "Figure 3. Our data collection rig in its handheld (right) and bike-mounted (left) configurations; see App. A for additional images." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 482, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Limitations Despite its size compared to previous datasets, I/Q-1M is quite small compared to the datasets used to train modern vision transformers, which can exceed billions of samples [68]. I/Q-1M also only includes daylight conditions and fair weather, and lacks the scale to capture \"edge cases\" that would be represented in a larger dataset [22]. Finally, since I/Q-1M uses a single type of radar, we cannot evaluate generalization across different antenna configurations — only radar configurations.", + "bbox": [ + 88, + 342, + 482, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 492, + 232, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4D Solid State Radar Excluding mechanical radars, which perceive the world as lidar-like 2D heatmaps [3, 5, 54] via a rotating antenna, solid-state mmWave radars operate by transmitting and receiving a sequence of frequency-modulated \"chirps\" from an array of transmit (TX) and receive (RX) antenna [19]; this data is typically (losslessly) transformed to a 4D range-Doppler-azimuth-elevation data cube using a 4D FFT [50], whose resolutions are constrained by bandwidth, form factor, and the integration window. We focus on single-chip radars which have compact form factors – and thus poor angular resolution.", + "bbox": [ + 88, + 516, + 482, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning and Datasets for 4D Radar Most radar processing methods use point clouds extracted from the 4D cubes [36, 52] as inputs [1, 7, 32, 44, 55, 56], allowing them to re-use popular Lidar architectures or even pre-trained models such as PointNet [46]. However, since this discards much of the information contained in a 4D radar cube, many competing approaches propose to directly interpret the 4D radar cube using methods and architectures such as feedforward convolutional architectures [9, 38, 43, 49, 69], multiview convolutional networks across different tensor axes [13, 34, 40], U-Nets [26, 39, 45], diffusion models [70], and transformers [2, 15, 21, 21].", + "bbox": [ + 88, + 688, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These prior machine learning-based approaches rely on publicly available datasets with 4D radar data, including", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/2cd822d5d879b3feb7312e967cb6a2f7f70fbb58174ce4944f138655e3bd4121.jpg", + "table_caption": [ + "Table 1. Comparison with other single-chip mmWave radar datasets; a frame refers to the number of unique radar-sensor samples. For comparisons with datasets using other types of radar, see App. A.3. Our dataset is significantly larger than previous single-chip radar datasets, enabling us to explore scaling up models." + ], + "table_footnote": [], + "table_body": "
Dataset4D Data CubeDataset Size
I/Q-1M (Ours)Yes29 hours (1M frames)
MilliPoint [7]No (3D Points)6.3 Hours (545k frames)
RaDICal [29]Yes3.6 Hours (394k frames)
CRUW [63]No (2D Map)3 hours (400k frames)
Coloradar [23]Yes2.4 hours (82k frames)
RadarHD [45]Yes200k frames
CARRADA [41]No (3D Cube)21 Minutes (13k frames)
RADDet [69]Yes10k frames
", + "bbox": [ + 527, + 161, + 888, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from both cascaded [42, 49] and single-chip [23, 29, 69] radars; however, existing datasets are relatively small, with 3.6-hour RaDICal [29] and 2.4-hour Coloradar [23] as the largest (Table 1). Due to the success of powerful but data-hungry [68] transformer models [61] in computer vision [12], we believe that limited data availability imposes a substantial bottleneck on learning for 4D radar.", + "bbox": [ + 511, + 309, + 903, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "High-resolution Imaging from Low-Resolution Radar Due to the low angular resolution of single-chip radars, extracting high-resolution angular information can be challenging; as a result, prior work focuses on recovering 2D spatial information [15, 45, 70]. Thus, while prior methods can extract 3D information using high-resolution cascaded radars [10], 3D imaging from single-chip radars generally requires additional information such as structured motion or multiple views, for example segmentation and maps using a rotating single-chip radar [24], a 3D occupancy map using multiple views [18], or high resolution images from fixed trajectories using Synthetic Aperture Radar [14, 35, 66, 67]. Instead, we show that by leveraging a sufficiently large dataset, even single frames are sufficient to recover dense angular resolution in both azimuth and elevation.", + "bbox": [ + 511, + 422, + 906, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Data Collection System and Dataset", + "text_level": 1, + "bbox": [ + 511, + 662, + 836, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset scale is key to training and evaluating potential foundational models. As such, we developed a scalable data collection system (Fig. 3), which we used to collect a large raw mmWave radar dataset, consisting of 1M radar-lidar-camera samples over 29 hours (Table 1). For additional details on our dataset and data collection rig, see App. A.", + "bbox": [ + 511, + 686, + 903, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection System", + "text_level": 1, + "bbox": [ + 511, + 787, + 728, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our data collection system, red-rover, was built around a TI AWR1843 Radar, Lidar, Camera, and IMU which can be easily operated via a simple web app on a mobile phone. Our system records all data to a single hot-swappable external drive via a single linux computer which handles time synchronization, minimizing turnaround time.", + "bbox": [ + 511, + 810, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24657", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/adc65dd3ccb2ab4bdc50bd1b3f1d524059ca8548c806ffe3d4a4b7a0b911f936.jpg", + "table_caption": [ + "Table 2. Key specifications for each setting. Settings have varying max Doppler $D_{\\mathrm{max}}$ and range $R_{\\mathrm{max}}$ ; all traces used a fixed resolution of 64 Doppler and 256 range bins." + ], + "table_footnote": [], + "table_body": "
SettingSizeLengthAverage SpeedDmaxRmax
indoor310k8.9h1.0m/s1.2m/s11.2m
outdoor372k10.7h1.4m/s1.8m/s22.4m
bike333k9.3h5.4m/s8.0m/s22.4m
", + "bbox": [ + 91, + 133, + 480, + 195 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/fa8a223c402f10115e5931c95778be6a309470b9197511a0eae1551ceb8e517e.jpg", + "table_caption": [ + "Table 3. Transformer sizes. Layers indicates the number of encoder + decoder layers; Speed indicates the (batched) inference throughput of each model on a single RTX 4090." + ], + "table_footnote": [], + "table_body": "
SizeLayersDimensionParamsSpeed
pico2 + 2256 (4 heads)3.9M750 fps
tiny3 + 3384 (6 heads)12.7M320 fps
small4 + 4512 (8 heads)28.9M170 fps
medium6 + 6640 (10 heads)69.4M84 fps
large9 + 9768 (12 heads)149M44 fps
", + "bbox": [ + 116, + 244, + 455, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also designed our data collection system to have a compact, battery-operated form factor to allow for a variety of collection modalities, including handheld and bicycle mounted. This also allows us to collect data relevant for tasks such as indoor sensing and localization, which are underrepresented in existing datasets, while still collecting automotive-like data by mounting our system to an E-bike.", + "bbox": [ + 89, + 359, + 483, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Collected Data", + "text_level": 1, + "bbox": [ + 89, + 474, + 241, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We collected three roughly equally sized splits (Table 2) from indoor handheld, outdoor handheld, and bike-mounted settings on the CMU campus and Pittsburgh area:", + "bbox": [ + 89, + 496, + 482, + 541 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- indoor: inside buildings at a slow to moderate walking pace, visiting multiple floors and areas within each.", + "- outdoor: neighborhoods ranging from single family detached to high density commercial zoning at a moderate to fast walking pace.", + "- bike: bike rides in different directions from a set starting point with a moderate biking pace." + ], + "bbox": [ + 89, + 542, + 482, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each setting features a mobile observer, with radar modulation parameters tuned for typical speeds. For sample data from each setting, see App. A.4.", + "bbox": [ + 89, + 648, + 482, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 707, + 225, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using a transformer architecture (Sec. 4.1), we train our Generalizable Radar Transformer (GRT) for a range of different tasks (Sec. 4.2-4.3), and evaluate it on our dataset using a rigorous statistical methodology (Sec. 4.4).", + "bbox": [ + 89, + 732, + 483, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Model Architecture", + "text_level": 1, + "bbox": [ + 89, + 803, + 277, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While many architectural refinements exist for vision transformers [30, 48], as well as for radar specifically [15, 26], we use a direct adaptation (Fig. 4) of the original Transformer [61] and Vision Transformer [12] to focus on measuring the fundamental properties of Radar transformers.", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Radar Processing From the (slow time, TX, RX, fast time) I/Q stream, we perform a 4-Dimensional FFT to obtain (range, Doppler, azimuth, elevation) dense 4D radar data cubes of size (256, 64, 8, 2), which we provide to the model as two channels consisting of the amplitude and phase angle. This data cube is patched along the range and Doppler dimensions into patches of size $4 \\times 2(\\times 8 \\times 2)$ , yielding an initial set of $64 \\times 32 = 2048$ patch tokens.", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transformer Architecture Crucially, unlike a vision transformer [12], radar models take inputs that have different spatial axes than their outputs, with vastly different relative resolutions where they overlap. As such, we use a standard transformer with a decoder [61], with varying layers and widths (Table 3); for a full specification of our transformer architecture and training procedure, see App. B.1.", + "bbox": [ + 511, + 220, + 906, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoder Query To handle the \"change of basis\" between the input and output space, we use an architecture based on Perceiver I/O [20]. We start by concatenating a (learned) output token to the encoder (similar to standard vision transformer [8]). The encoder output corresponding to the output token is then tiled into the desired decoder shape with a 3D sinusoidal positional encoding applied and is used as the input to the decoder, which attends to the encoder outputs.", + "bbox": [ + 511, + 335, + 906, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Base Task: 3D Occupancy Classification", + "text_level": 1, + "bbox": [ + 511, + 467, + 859, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An ideal foundational model training task should be easy to gather data for (e.g., using self-supervised learning) and closely aligned with a wide range of potential downstream tasks. As such, since we are primarily concerned with understanding the spatial relationship between 4D radar data and 3D space, we use 3D (polar) occupancy classification – predicting the occupancy of $64 \\times 128 \\times 64$ range-azimuth elevation cells – as a base task, with Lidar as a ground truth. Our task uses a binary cross entropy objective, with some weighting to correct for cell sizes; see App. B.3 for details.", + "bbox": [ + 511, + 489, + 905, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, in addition to being fully self-supervised, this task covers all three possible output dimensions (range, azimuth, and elevation), meaning that downstream tasks such as range-azimuth classification or azimuth-elevation segmentation can be cast as 2D slices of this 3D output. This enables us to fine-tune for tasks, even if they have different spatial dimensions, simply by replacing the output head and modifying the output positional encoding queries.", + "bbox": [ + 511, + 641, + 906, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Other Tasks", + "text_level": 1, + "bbox": [ + 511, + 772, + 643, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to evaluate GRT's suitability as a foundational model for downstream fine-tuning, we use three additional tasks, each representing different output dimensions:", + "bbox": [ + 511, + 794, + 905, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Bird's Eye View (BEV) Occupancy: Similarly to [15, 45], we classify the $256 \\times 1024$ range-azimuth polar occupancy using Lidar as a ground truth, with the range normalized to the radar's range resolution.", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24658", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a916548feedfeb7a432a77b7806970f7adf97ff3da10cb152669c32ab994ec0a.jpg", + "image_caption": [ + "Figure 4. The GRT architecture. 4D radar cubes are patched with a linear projection, and a sinusoidal positional encoding is added. A transformer architecture is then used, with a transformer decoder for dense outputs and a MLP decoder for Ego-Motion estimation; different output encodings are used depending on the output axes and resolution." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 220 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Semantic Segmentation: Similar to [24], we train our GRT to output $640 \\times 640$ azimuth-elevation class labels. Since radars cannot feasibly identify many classes (e.g., poster vs. sign vs. wall) which a camera could, we use eight coarse categories: person, sky, vehicle, flat, nature, structure, ceiling, and object.", + "- Ego-Motion Estimation: Since radars can \"natively\" measure velocity², we predict the velocity of the radar relative to its current orientation. Since ego-motion estimation does not require a dense output, we replace the transformer decoder with a multi-layer-perceptron decoder head with 3 layers of 512 units." + ], + "bbox": [ + 89, + 279, + 482, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For more detail about each task, see App. B.3.", + "bbox": [ + 89, + 462, + 395, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 494, + 210, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the cost of scaling foundational models, false positives can result in significant wasted resources, especially if associated with a costly methodological change. As such, since our dataset cannot be treated as having an \"infinite sample size\", we calculate upper-bounded uncertainty estimates wherever possible. In order to ensure these results are statistically accurate, we take the following steps:", + "bbox": [ + 89, + 518, + 483, + 625 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Geo-Split: Within each setting, $\\approx 1.5$ hours of data was reserved as a test set, which we ensured to be geographically disjoint from the training set to prevent leakage [28].", + "- Sample Size correction: Time series signals – e.g. radar-lidar-video tuples – cannot be viewed as independent samples; as such, the effective sample size, which we obtain from an autocorrelation-based estimate [51], must be used when calculating the standard error.", + "- Paired z-Test: Using the fact that models are evaluated on the same test traces, we use a paired z-test on the relative performance of each model with respect to a baseline." + ], + "bbox": [ + 89, + 627, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We report each metric relative to its specified baseline by default, along with error bars for a two-sided $95\\%$ confidence interval; using our procedure and dataset, we can measure differences of $1 - 2\\%$ (App. B.4).", + "bbox": [ + 89, + 795, + 482, + 856 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Validation Split and Data Size Sampling. We used the last $10\\%$ of each training trace for validation (separate from the test set), with the first $90\\%$ used for training. When training on reduced dataset sizes, we use the first $9\\%$ , $18\\%$ , and $45\\%$ of each trace for training for $10\\%$ , $20\\%$ , $50\\%$ dataset sizes respectively; to reduce the variance of our experiments, we always $10\\%$ of each trace for validation.", + "bbox": [ + 511, + 279, + 906, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 513, + 397, + 599, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using our dataset, we first ran extensive ablation (Sec. 5.1), scaling (Sec. 5.2), and fine-tuning (Sec. 5.3) experiments which show the efficacy, scalability, and generalizability of GRT. Our experiments took 874 RTX 4090-hours $^3$ of training time, with the GRT-small model taking 22 RTX 4090-hours to train.", + "bbox": [ + 511, + 422, + 905, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model Performance Despite the low resolution (only 3 $\\mathrm{TX} \\times 4$ TX antenna) of our radar, GRT is able to predict a range of outputs with remarkable quality (Fig. 5); we show additional examples in App. C.1. We also evaluated common metrics for each objective as an absolute reference (App. C.2); GRT achieves a 3D chamfer distance of 4.9 range bins, corresponding to $0.66\\mathrm{m}$ indoors, $1.6\\mathrm{m}$ outdoors, and $1.5\\mathrm{m}$ on bike.", + "bbox": [ + 511, + 518, + 905, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Ablation Studies", + "text_level": 1, + "bbox": [ + 513, + 647, + 678, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using our dataset and GRT, we performed ablation studies on several parameters that are independent of the underlying architecture and the task (Table 4). In particular, we find that several common practices – omitting Doppler information, using Angle-of-Arrival Estimates, and applying CFAR thresholding – result in degraded performance equivalent to more than a $10 \\times$ reduction in training data.", + "bbox": [ + 511, + 667, + 905, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input Representation Our models use complex 4D radar cubes that losslessly capture all information measured by a radar; however, the common practice in radar models and data sets is to use processed higher-level representations. We benchmark three common approaches:", + "bbox": [ + 511, + 779, + 906, + 854 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3Our experiments were run on a range of different machines with varying compute capacity, which we normalize with respect to a single RTX 4090. We only tracked training and validation time, with testing excluded.", + "bbox": [ + 511, + 862, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{2}$ Sensors which provide \"absolute\" pose, e.g. Lidars and Cameras, must differentiate, while IMUs must integrate.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24659", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/34204fc446f05cb39fec29fd2b2438e1bfd0d5fc9ab884034ec0c67b22bae0a0.jpg", + "image_caption": [ + "Figure 5. Training with more data and fine tuning instead of starting from scratch lead to much higher performance. 3D occupancy maps are less noisy as seen in rendered depth images (left), semantic segmentation is cleaner (center left), bird's eye view occupancy is sharper (center right), and velocity estimation is more accurate (right). However, scaling model size (left) does not have a large impact." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 295, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/65bba12e933c69d364c63d49ce00aef092522ed72932741a9b5432da7e9275c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 90, + 496, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3152cc6949a027a0bcfc98f66254f9a13e5bfde6306fa38ae0e1fd0f5c1602ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 90, + 700, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dbf7f6ead04dc07b3af3c0890e115c36473dd1a40511afb4cf6b2a9f355fbc92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 90, + 903, + 393 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Real (amplitude-only) 4D data cubes are not significantly different from complex data, indicating that the \"leftover\" phase from a Doppler FFT carries little additional information. Since using complex data has negligible compute overhead, we default to complex inputs.", + "- Angle of Arrival Estimates can be used to replace dense antenna measurements. Since our radar has only 2 elevation bins, we reduce the azimuth axis (8 bins) into an AoA estimate. This discards a substantial amount of information, leading to a $28.0\\%$ loss increase, equivalent to more than a $10\\times$ decrease in the size of the dataset.", + "- Constant False Alarm Rate (CFAR) processing removes weak or noisy reflectors based on local estimates of background noise [36]; this ablation uses a p-value threshold of 0.05, and zeros out rejected points. This leads to an even higher $31.5\\%$ loss increase." + ], + "bbox": [ + 89, + 459, + 482, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Impact of Doppler With limited angular resolution, GRT is highly dependent on Doppler information, which can capture higher resolution geometry [18]. We find that removing the Doppler FFT from our processing pipeline (i.e. treating each 4D radar cube as a time series of 64 3D frames [45]) leads to a $22.5\\%$ loss increase. As an additional ablation, we also shuffle the slow-time axis to fully destroy any Doppler information; this does not lead to a further significant loss increase, suggesting that off-the-shelf transformers cannot easily learn FFTs. Finally, we observe worse performance at low speeds since less Doppler information is available at slow speeds (App. C.3).", + "bbox": [ + 88, + 719, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c55036d9e68b567c0bd17d3b0cacef008b6efc980b03d97f1df4d5530a84463f.jpg", + "table_caption": [ + "Table 4. Test loss for each ablation (smaller is better) relative to GRT-small trained on our full dataset, along with $95\\%$ confidence intervals for the relative differences." + ], + "table_footnote": [], + "table_body": "
AblationRelative Test Loss
InputsAmplitude Only+0.04 ± 0.85%
Angle of Arrival+28.0 ± 2.12%
CFAR Thresholding+31.5 ± 2.38%
DopplerWithout Doppler FFT+22.5 ± 1.76%
Slow Time Shuffled+23.10 ± 1.94%
Post-Patch AxesDoppler-Az-EI+6.22 ± 1.11%
Range-Az-EI+6.27 ± 1.10%
Range-Doppler-Az-EI+4.18 ± 1.09%
AugmentationsNone+5.87 ± 1.32%
Scale, Phase, and Flip Only+3.89 ± 1.19%
Separate ModelsIndoor Data Only+5.76 ± 1.85%
Outdoor Data Only+5.58 ± 1.26%
Bike Data Only+2.77 ± 1.41%
", + "bbox": [ + 514, + 503, + 903, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Patch Axes Since 4D range-Doppler-azimuth-elevation radar data cubes have four axes with different properties, they do not have an obvious counterpart to the square patches used in Vision Transformers. Benchmarking four alternatives (App. B.1), with each resulting in 2048 total patches, we find that Range-Doppler patching where the azimuth and elevation axes are \"patched out\" (similar to [15, 49]) is the most effective, performing $\\approx 5\\%$ better.", + "bbox": [ + 511, + 719, + 906, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Augmentations We develop a range of data aug- mentations that together provide a modest but significant performance improvement (5.87±1.32%; Table 4); we pro", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24660", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1424cedc69f52fdfe5cc11b94620d3f35952def931b5567028af627c37ea51e5.jpg", + "image_caption": [ + "Figure 6. Scaling laws for mmWave radar transformers across indoor, outdoor, and bike test splits. Models and confidence intervals are measured relative to the small transformer trained on $10\\%$ of the dataset. While our models show weak scaling over model size when trained on our dataset, we see strong log-linear scaling across dataset size of $\\approx 20\\%$ loss decrease per $10\\times$ increase in data size. For variants of this graph with respect to absolute metrics, see App. C.2." + ], + "image_footnote": [], + "bbox": [ + 102, + 95, + 890, + 251 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f3a8a275912cf64f0adaae5fa3fdf887528a12d5ef894aaf39583bd3e4fb7f2.jpg", + "image_caption": [ + "Figure 7. GRT compared to Baselines, measured with respect to Chamfer distance (in range bins) and test loss across different training set sizes, averaged across our dataset." + ], + "image_footnote": [], + "bbox": [ + 101, + 327, + 473, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vide details about each augmentation in App. B.2.", + "bbox": [ + 89, + 530, + 423, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Towards a Radar Foundational Model", + "text_level": 1, + "bbox": [ + 89, + 554, + 419, + 569 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scalability vs. Baselines Since prior work on learning for single-chip radar focuses on 2D outputs, we benchmark our transformer-based approach against two prior architectures for 2D BEV Occupancy prediction: a U-Net-based model (RadarHD [45]), and a Swin Transformer-based model (T-FFTRadNet [15]), with minor architecture modifications to conform to our data dimensions (App. B.5). We find that GRT-small outperforms both baselines at all training splits (Fig. 7), demonstrating the suitability of transformer architectures for scalability.", + "bbox": [ + 89, + 577, + 483, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling Laws Training 20 different models for 5 different sizes (Table 3) and dataset sizes ranging from $10 - 100\\%$ of our data (Fig. 6), we observe a logarithmic improvement with data size of approximately $20\\%$ improvement per $10\\times$ increase in data, similar to early observations in computer vision [57]. This can also be seen qualitatively (Fig. 5), where models trained on more data produce much higher quality predictions. Similarly to vision transformers [57, 68], we also observe that larger models are more data efficient, although the magnitude of difference that we observe is much smaller due to our limited dataset size.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/afad5eb06eea25731d9df42a1071b40049b575ee816fe9757ad91d941855c1da.jpg", + "table_caption": [ + "Table 5. Chamfer Distance (in meters) for 2D BEV occupancy prediction on the Coloradar dataset [23] by location; the geometric mean is listed to account for the varying difficulty of each location. A fine-tuned GRT model outperforms baselines trained only on Coloradar, including a state of the art diffusion-based model [70] and a U-Net based model [45]." + ], + "table_footnote": [], + "table_body": "
TraceGRT (Ours)Diffusion [70]RadarHD [45]
Geometric Mean0.981.191.73
ARPG Lab0.780.961.73
EC Hallways1.041.041.69
Aspen0.610.510.91
Longboard2.635.475.40
Outdoors1.842.373.10
Edgar0.360.440.60
", + "bbox": [ + 516, + 421, + 901, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalizability across different settings We evaluate the ability of GRT to generalize across different settings by comparing a baseline model trained on combined indoor, outdoor, and bike data with models trained on each setting separately (Table 4). Despite the differences in these settings, the jointly trained model performs significantly better than models trained on each setting separately, confirming that data from different settings can be combined to train a single, stronger model.", + "bbox": [ + 511, + 561, + 906, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. As a Base Model for Downstream Tasks", + "text_level": 1, + "bbox": [ + 511, + 710, + 854, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset Fine-tuning We fine-tuned a small GRT model on the Coloradoar [23] dataset using a BEV Occupancy objective, and benchmarked the resulting model against two prior approaches trained only on ColoRadar, including a state-of-the-art diffusion model, using the same data splits and evaluation procedure [70]. Notably, despite using a modulation and resolution (128 range $\\times$ 128 Doppler) which is not present in our dataset, GRT can be run without any architectural modifications, such as modifying the number of upsampling stages, as would be required for a convolutional architecture (App. B.5). After fine-tuning until vali", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24661", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2080bec3029056116a0cdb45bc4aee236ef9ae367a01697420184a20dc477371.jpg", + "image_caption": [ + "Figure 8. Fine tuning mmWave radar transformers to different downstream tasks. Models and confidence intervals are measured relative to the small transformer trained (from scratch) on $10\\%$ of the dataset. Pre-training and fine tuning strongly impacts data efficiency, equivalent to up to a $5\\times$ increase in dataset size (observed in the Segmentation task)." + ], + "image_footnote": [], + "bbox": [ + 127, + 95, + 380, + 237 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dda890857ac31b4718f1091d4f907fa3ddf26dc18e9ec681a9f1b882ab7b8041.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 95, + 620, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8515653f915d2417d8f4577eb33934689676576b11ab10f0a4d57d0be252bbc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 95, + 864, + 239 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/60514d9ebc7b350aa40b2d511aa9c15631759c8c67de3aa4a45d762a78888f01.jpg", + "image_caption": [ + "Figure 9. A linear projection of the observed logarithmic scaling to a test loss lower bound suggests that logarithmic scaling cannot continue beyond $100 \\times$ our current dataset" + ], + "image_footnote": [], + "bbox": [ + 114, + 311, + 455, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dation loss convergence ( $\\approx 30$ minutes of training using a single RTX 4090), GRT achieves substantially lower Chamfer distance than baselines trained only on ColoRadar, showing the value of easily tunable foundational models (Table 5).", + "bbox": [ + 89, + 534, + 483, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Task Fine-tuning We also fine-tuned GRT-small for each of our secondary tasks using $10 - 100\\%$ of our dataset, and compared the results with models trained from scratch on the same proportions of the dataset. Following this procedure, we find substantial performance gains equivalent to up to a $5\\times$ increase in dataset size compared to training from scratch (Fig. 8). This effect is especially pronounced when less data is available, with the performance benefits of fine tuning disappearing as the dataset is scaled for the BEV Occupancy objective but staying more or less constant for the Semantic Segmentation objective. We also observe this effect as a clear qualitative difference: fine-tuned models produce sharper and more accurate predictions than their counterparts trained from scratch (Fig. 5).", + "bbox": [ + 89, + 609, + 483, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. How Much More Data is Needed?", + "text_level": 1, + "bbox": [ + 89, + 832, + 387, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although we cannot directly observe performance saturation, we project how much data would be required to saturate a Radar Transformer using two different methods to ar", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rive at a best guess of approximately 100M samples - $100\\times$ our current dataset.", + "bbox": [ + 511, + 311, + 903, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Linear Projection of Scaling Laws To lower-bound the possible test loss in our dataset, we trained a small model on the test set to approximate convergence. Assuming that the rate of improvement in model performance with increased training data cannot decrease, we extend our observed (Fig. 6) logarithmic scaling law to this lower bound to, in turn, estimate an upper bound for when the logarithmic trend will no longer hold (Fig. 9). This yields an estimate of $100 \\times$ our current dataset size. For additional details justifying our estimation of this bound, see App. C.4.", + "bbox": [ + 511, + 357, + 906, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Validation Curve Trends We observe that GRTs tend to stop improving (with respect to validation loss) after $\\approx 10$ training epochs, regardless of model or dataset size (App. C.4); this is similar to trends observed in the training of data-constrained LLMs, which are also observed to saturate around 10 epochs [37]. Using vision transformers, whose scaling laws are well studied [68] due to the availability of internet data, as a reference, we expect training saturation to occur around $10^{2} - 10^{4}\\mathrm{M}$ samples seen. Since each epoch in our dataset corresponds to $\\approx 1\\mathrm{M}$ samples seen, this implies that $10\\times$ to $1000\\times$ our current dataset size is required to delay overfitting beyond this point.", + "bbox": [ + 511, + 525, + 908, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 723, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we train a Generalizable Radar Transformer using a large, 29 hour (1M sample) dataset collected using our open-source data collection system and demonstrate that our Radar Transformer can generalize across datasets and settings, can be readily fine-tuned, and exhibits logarithmic scaling. While we believe that substantial gains are still possible through further data scaling, we hope that our dataset and baseline models will enable the community to revisit previous methods in new context and explore new capabilities made possible by a much larger dataset.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24662", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Karim Armanious, Maurice Quach, Michael Ulrich, Timo Winterling, Johannes Friesen, Sascha Braun, Daniel Jenet, Yuri Feldman, Eitan Kosman, Philipp Rapp, et al. Bosch street dataset: A multi-modal dataset with imaging radar for automated driving. arXiv preprint arXiv:2407.12803, 2024. 3", + "[2] Jie Bai, Lianqing Zheng, Sen Li, Bin Tan, Sihan Chen, and Libo Huang. Radar transformer: An object classification network based on 4d mmw imaging radar. Sensors, 21(11): 3854, 2021. 3", + "[3] Dan Barnes, Matthew Gadd, Paul Murcutt, Paul Newman, and Ingmar Posner. The oxford radar robotcar dataset: A radar extension to the oxford robotcar dataset. In 2020 IEEE international conference on robotics and automation (ICRA), pages 6433-6438. IEEE, 2020. 3", + "[4] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2", + "[5] Keenan Burnett, David J Yoon, Yuchen Wu, Andrew Z Li, Haowei Zhang, Shichen Lu, Jingxing Qian, Wei-Kang Tseng, Andrew Lambert, Keith YK Leung, et al. Boreas: A multi-season autonomous driving dataset. The International Journal of Robotics Research, 42(1-2):33-42, 2023. 3", + "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3", + "[7] Han Cui, Shu Zhong, Jiacheng Wu, Zichao Shen, Naim Dahnoun, and Yiren Zhao. Milipoint: A point cloud dataset for mmwave radar. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 4", + "[9] Colin Decourt, Rufin VanRullen, Didier Salle, and Thomas Oberlin. Darod: A deep automotive radar object detector on range-doppler maps. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 112-118. IEEE, 2022. 3", + "[10] Fangqiang Ding, Xiangyu Wen, Yunzhou Zhu, Yiming Li, and Chris Xiaoxuan Lu. Radar: Robust 3d occupancy prediction with 4d imaging radar. Advances in Neural Information Processing Systems, 37:101589-101617, 2025. 3", + "[11] Christopher Doer and Gert F Trommer. Yaw aided radar inertial odometry using manhattan world assumptions. In 2021 28th Saint Petersburg International Conference on Integrated Navigation Systems (ICINS), pages 1-9. IEEE, 2021. 2", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "scale. In International Conference on Learning Representations, 2021. 3, 4", + "[13] Xiangyu Gao, Guanbin Xing, Sumit Roy, and Hui Liu. Ramp-cnn: A novel neural network for enhanced automotive radar object recognition. IEEE Sensors Journal, 21(4): 5119-5132, 2020. 3", + "[14] Xiangyu Gao, Sumit Roy, and Guanbin Xing. Mimo-sar: A hierarchical high-resolution imaging algorithm for mmwave fmew radar in autonomous driving. IEEE Transactions on Vehicular Technology, 70(8):7322-7334, 2021. 3", + "[15] James Giroux, Martin Bouchard, and Robert Laganiere. Tfftradnet: Object detection with swin vision transformers from raw adc radar signals. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4030-4039, 2023. 3, 4, 6, 7, 8", + "[16] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 6", + "[17] Wolfgang Hess, Damon Kohler, Holger Rapp, and Daniel Andor. Real-time loop closure in 2d lidar slam. In 2016 IEEE international conference on robotics and automation (ICRA), pages 1271-1278. IEEE, 2016. 7", + "[18] Tianshu Huang, John Miller, Akarsh Prabhakara, Tao Jin, Tarana Laroia, Zico Kolter, and Anthony Rowe. Dart: Implicit doppler tomography for radar novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24118-24129, 2024. 2, 3, 6, 1, 8", + "[19] Cesar Iovescu and Sandeep Rao. The fundamentals of millimeter wave sensors. *Texas Instruments*, pages 1–8, 2017. 3", + "[20] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 4", + "[21] Yi Jin, Marcel Hoffmann, Anastasios Deligiannis, JuanCarlos Fuentes-Michel, and Martin Vossiek. Semantic segmentation-based occupancy grid map learning with automotive radar raw data. IEEE Transactions on Intelligent Vehicles, 9(1):216-230, 2023. 3", + "[22] Nidhi Kalra and Susan M Paddock. Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability? Transportation Research Part A: Policy and Practice, 94:182-193, 2016. 3", + "[23] Andrew Kramer, Kyle Harlow, Christopher Williams, and Christoffer Heckman. Coloradar: The direct 3d millimeter wave radar dataset. The International Journal of Robotics Research, 41(4):351-360, 2022. 2, 3, 7", + "[24] Haowen Lai, Gaoxiang Luo, Yifei Liu, and Mingmin Zhao. Enabling visual recognition at radio frequency. In Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, pages 388-403, 2024. 2, 3, 5, 9", + "[25] Xinrong Li, Xiaodong Wang, Qing Yang, and Song Fu. Signal processing for tdm mimo fmcw millimeter-wave radar sensors. IEEE Access, 9:167959-167971, 2021. 2", + "[26] Yu-Jhe Li, Shawn Hunt, Jinhyung Park, Matthew O'Toole, and Kris Kitani. Azimuth super-resolution for fmcw radar in" + ], + "bbox": [ + 516, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24663", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17504-17513, 2023. 3, 4", + "[27] Jaime Lien, Nicholas Gillian, M Emre Karagozler, Patrick Amihood, Carsten Schwesig, Erik Olson, Hakim Raja, and Ivan Poupyrev. Soli: Ubiquitous gesture sensing with millimeter wave radar. ACM Transactions on Graphics (TOG), 35(4):1-19, 2016. 1", + "[28] Adam Lilja, Junsheng Fu, Erik Stenberg, and Lars Hammarstrand. Localization is all you evaluate: Data leakage in online mapping datasets and how to fix it. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22150-22159, 2024. 5", + "[29] Teck-Yian Lim, Spencer A Markowitz, and Minh N Do. Radical: A synchronized fmcw radar, depth, imu and rgb camera data dataset with low-level fmcw radar signals. IEEE Journal of Selected Topics in Signal Processing, 15(4):941-953, 2021. 3, 1", + "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4", + "[31] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017.6", + "[32] Chris Xiaoxuan Lu, Muhamad Risqi U Saputra, Peijun Zhao, Yasin Almalioglu, Pedro PB De Gusmao, Changhao Chen, Ke Sun, Niki Trigoni, and Andrew Markham. milliego: single-chip mmwave radar aided egomotion estimation via deep sensor fusion. In Proceedings of the 18th Conference on Embedded Networked Sensor Systems, pages 109-122, 2020. 3", + "[33] Sohrab Madani, Jayden Guan, Waleed Ahmed, Saurabh Gupta, and Haitham Hassanieh. Radatron: Accurate detection using multi-resolution cascaded mimo radar. In European Conference on Computer Vision, pages 160-178. Springer, 2022. 3", + "[34] Bence Major, Daniel Fontijne, Amin Ansari, Ravi Teja Sukhavasi, Radhika Gowaikar, Michael Hamilton, Sean Lee, Slawomir Grzechnik, and Sundar Subramanian. Vehicle detection with automotive radar using deep learning on range-azimuth-doppler tensors. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 3", + "[35] Babak Mamandipoor, Greg Malysa, Amin Arababian, Upamanyu Madhow, and Karam Noujeim. 60 ghz synthetic aperture radar for short-range imaging: Theory and experiments. In 2014 48th Asilomar Conference on Signals, Systems and Computers, pages 553-558. IEEE, 2014. 3", + "[36] Gary Minkler and Jing Minkler. Cfar: the principles of automatic radar detection in clutter. *Nasa sti/recon technical report a*, 90:23371, 1990. 2, 3, 6", + "[37] Niklas Muennighoff, Alexander Rush, Boaz Barak, Teven Le Scao, Nouamane Tazi, Aleksandra Piktus, Sampo Pyysalo, Thomas Wolf, and Colin A Raffel. Scaling data-constrained language models. Advances in Neural Information Processing Systems, 36:50358-50376, 2023. 8" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Farzan Erlik Nowruzi, Dhanvin Kolhatkar, Prince Kapoor, Fahed Al Hassanat, Elnaz Jahani Heravi, Robert Laganiere, Julien Rebut, and Waqas Malik. Deep open space segmentation using automotive radar. In 2020 IEEE MTT-S International Conference on Microwaves for Intelligent Mobility (ICMIM), pages 1-4. IEEE, 2020. 3", + "[39] Itai Orr, Moshik Cohen, and Zeev Zalevsky. High-resolution radar road segmentation using weakly supervised learning. Nature Machine Intelligence, 3(3):239-246, 2021. 3", + "[40] Arthur Ouaknine, Alasdair Newson, Patrick Pérez, Florence Tupin, and Julien Rebut. Multi-view radar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15671-15680, 2021. 3", + "[41] Arthur Ouaknine, Alasdair Newson, Julien Rebut, Florence Tupin, and Patrick Pérez. Carrada dataset: Camera and automotive radar with range-angle-doppler annotations. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 5068-5075. IEEE, 2021. 3", + "[42] Dong-Hee Paek, Seung-Hyun Kong, and Kevin Tirta Wijaya. K-radar: 4d radar object detection for autonomous driving in various weather conditions. Advances in Neural Information Processing Systems, 35:3819-3829, 2022. 3", + "[43] Andras Palffy, Jiaao Dong, Julian FP Kooij, and Dariu M Gavrila. Cnn based road user detection using the 3d radar cube. IEEE Robotics and Automation Letters, 5(2):1263-1270, 2020. 3", + "[44] Akarsh Prabhakara, Diana Zhang, Chao Li, Sirajum Munir, Aswin C Sankaranarayanan, Anthony Rowe, and Swarun Kumar. Exploring mmwave radar and camera fusion for high-resolution and long-range depth imaging. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3995-4002. IEEE, 2022. 3", + "[45] Akarsh Prabhakara, Tao Jin, Arnav Das, Gantavya Bhatt, Lilly Kumari, Elahe Soltanaghai, Jeff Bilmes, Swarun Kumar, and Anthony Rowe. High resolution point clouds from mmwave radar. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4135-4142. IEEE, 2023. 2, 3, 4, 6, 7, 1, 8", + "[46] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 13", + "[48] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 4", + "[49] Julien Rebut, Arthur Ouaknine, Waqas Malik, and Patrick Pérez. Raw high-definition radar for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17021-17030, 2022. 3, 6" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24664", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Mark A Richards et al. Fundamentals of radar signal processing. McGraw-hill New York, 2005. 2, 3", + "[51] Christian P Robert, George Casella, and George Casella. Monte Carlo statistical methods. Springer, 1999. 5, 8", + "[52] Hermann Rohling. Radar cfar thresholding in clutter and multiple target situations. IEEE Transactions on Aerospace and Electronic Systems, AES-19(4):608-621, 1983. 2, 3", + "[53] Mark E Russell, Arthur Crain, Anthony Curran, Richard A Campbell, Clifford A Drubin, and Willian F Miccioli. Millimeter-wave radar sensor for automotive intelligent cruise control (icc). IEEE Transactions on microwave theory and techniques, 45(12):2444-2453, 1997. 2", + "[54] Marcel Sheeny, Emanuele De Pellegrin, Saptarshi Mukherjee, Alireza Ahrabian, Sen Wang, and Andrew Wallace. Radiate: A radar dataset for automotive perception in bad weather. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 1-7. IEEE, 2021. 3", + "[55] Xian Shuai, Yulin Shen, Yi Tang, Shuyao Shi, Luping Ji, and Guoliang Xing. millieye: A lightweight mmwave radar and camera fusion system for robust object detection. In Proceedings of the International Conference on Internet-of-Things Design and Implementation, pages 145–157, 2021. 3", + "[56] Akash Deep Singh, Sandeep Singh Sandha, Luis Garcia, and Mani Srivastava. Radhar: Human activity recognition from point clouds generated through a millimeter-wave radar. In Proceedings of the 3rd ACM Workshop on Millimeter-wave Networks and Sensing Systems, pages 51-56, 2019. 3", + "[57] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 7", + "[58] Texas Instruments. Automated doors reference design using ti mmwave sensors. Design Guide TIDEP-01018, Texas Instruments, Dallas, TX, 2018. 1", + "[59] AWR1843AOP Single-chip 77- and 79-GHz FMCW mmWave Sensor Antennas-OnPackage (AOP). Texas Instruments, 2021. 2, 3", + "[60] Claudia Vasanelli, Fabian Roos, Andre Durr, Johannes Schlichenmaier, Philipp Hugler, Benedikt Meinecke, Maximilian Steiner, and Christian Waldschmidt. Calibration and direction-of-arrival estimation of millimeter-wave radars: A practical introduction. IEEE Antennas and Propagation Magazine, 62(6):34-45, 2020. 2", + "[61] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3, 4, 6", + "[62] Christian Waldschmidt, Juergen Hasch, and Wolfgang Menzel. Automotive radar—from first efforts to future systems. IEEE Journal of Microwaves, 1(1):135-148, 2021. 2", + "[63] Yizhou Wang, Zhongyu Jiang, Xiangyu Gao, Jenq-Neng Hwang, Guanbin Xing, and Hui Liu. Rodnet: Radar object detection using cross-modal supervision. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 504-513, 2021. 3", + "[64] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 7", + "[65] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. On layer normalization in the transformer architecture. In International Conference on Machine Learning, pages 10524-10533. PMLR, 2020. 6", + "[66] Hiroyoshi Yamada, Takumi Kobayashi, Yoshio Yamaguchi, and Yuuichi Sugiyama. High-resolution 2d sar imaging by the millimeter-wave automobile radar. In 2017 IEEE Conference on Antenna Measurements & Applications (CAMA), pages 149-150. IEEE, 2017. 3", + "[67] Muhammet Emin Yanik and Murat Torlak. Near-field mimosar millimeter-wave imaging with sparsely sampled aperture data. IEEE Access, 7:31801-31819, 2019. 3", + "[68] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104-12113, 2022. 3, 7, 8, 13", + "[69] Ao Zhang, Farzan Erlik Nowruzi, and Robert Laganiere. Raddet: Range-azimuth-doppler based radar object detection for dynamic road users. In 2021 18th Conference on Robots and Vision (CRV), pages 95–102. IEEE, 2021. 2, 3", + "[70] Ruibin Zhang, Donglai Xue, Yuhan Wang, Ruixu Geng, and Fei Gao. Towards dense and accurate radar perception via efficient cross-modal diffusion model. IEEE Robotics and Automation Letters, 2024. 3, 7", + "[71] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7" + ], + "bbox": [ + 516, + 92, + 906, + 555 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24665", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards Foundational Models for Single-Chip Radar/6536d138-fd2a-4e4b-a53a-8d73d592cae1_model.json b/2025/Towards Foundational Models for Single-Chip Radar/6536d138-fd2a-4e4b-a53a-8d73d592cae1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..68365f545e0f2cf15d5504054083c5b420b6b405 --- /dev/null +++ b/2025/Towards Foundational Models for Single-Chip Radar/6536d138-fd2a-4e4b-a53a-8d73d592cae1_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.23, + 0.131, + 0.77, + 0.154 + ], + "angle": 0, + "content": "Towards Foundational Models for Single-Chip Radar" + }, + { + "type": "text", + "bbox": [ + 0.183, + 0.179, + 0.814, + 0.199 + ], + "angle": 0, + "content": "Tianshu Huang\\(^{12}\\) Akarsh Prabhakara\\(^{3}\\) Chuhan Chen\\(^{1}\\) Jay Karhade\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.2, + 0.76, + 0.216 + ], + "angle": 0, + "content": "Deva Ramanan1 Matthew O'Toole1 Anthony Rowe12" + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.216, + 0.851, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Carnegie Mellon University \\(^{2}\\)Bosch Research \\(^{3}\\)University of Wisconsin-Madison" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.237, + 0.871, + 0.252 + ], + "angle": 0, + "content": "{tianshu2, chuhanc, jkarhade, deva, motoole2, agr}@andrew.cmu.edu, akarsh@cs.wisc.edu" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.259, + 0.785, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.369, + 0.372, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.37, + 0.638, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.37, + 0.905, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.472, + 0.372, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.472, + 0.638, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.472, + 0.904, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.576, + 0.908, + 0.618 + ], + "angle": 0, + "content": "Figure 1. Using only 4D range-Doppler-azimuth-elevation data from a radar with a 3x4 antenna array - equivalent to a 0.26-megapixel camera, we trained (separate) radar transformers on 24 hours (1M radar-lidar-camera samples) of data to predict Bird's Eye View 2D occupancy (left), 3D occupancy (center), semantic segmentation (right), and Ego-Motion." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.64, + 0.327, + 0.656 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.673, + 0.483, + 0.839 + ], + "angle": 0, + "content": "mmWave radars are compact, inexpensive, and durable sensors that are robust to occlusions and work regardless of environmental conditions, such as weather and darkness. However, this comes at the cost of poor angular resolution, especially for inexpensive single-chip radars, which are typically used in automotive and indoor sensing applications. Although many have proposed learning-based methods to mitigate this weakness, no standardized foundational models or large datasets for the mmWave radar have emerged, and practitioners have largely trained task-specific models from scratch using relatively small datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we collect (to our knowledge) the largest available raw radar dataset with 1M samples (29 hours) and train a foundational model for 4D single-chip radar, which can predict 3D occupancy and semantic segmenta" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.809 + ], + "angle": 0, + "content": "tion with quality that is typically only possible with much higher resolution sensors. We demonstrate that our Generalizable Radar Transformer (GRT) generalizes across diverse settings, can be fine-tuned for different tasks, and shows logarithmic data scaling of \\(20\\%\\) per \\(10\\times\\) data. We also run extensive ablations on common design decisions, and find that using raw radar data significantly outperforms widely-used lossy representations, equivalent to a \\(10\\times\\) increase in training data. Finally, we roughly estimate that \\(\\approx 100M\\) samples (3000 hours) of data are required to fully exploit the potential of GRT." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.83, + 0.646, + 0.846 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "As a compact, inexpensive [27], and robust solid-state sensor, mmWave radars are ideal for sensing applications ranging from simple automatic door openers [58] to autonomous" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24655" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "drones [11] or vehicles [53, 62]. mmWave radars are rich sensors which can directly measure range and velocity while capturing a unique range of material properties [18]; however, this comes at the cost of poor angular resolution typically on the order of \\(15^{\\circ}\\) - orders of magnitude worse than cameras or lidars [59]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.482, + 0.304 + ], + "angle": 0, + "content": "Radar data are typically processed into radar point clouds (Fig. 2) derived using Constant False Alarm Rate (CFAR) peak detectors [36, 52] combined with Angle-of-Arrival estimation techniques [60]. However, this is a substantially lossy process: while raw radar data suffers from unique noise patterns such as \"bleed\" and side lobes [25], weak reflectors and other signals can be hidden in this noise, which would ordinarily be filtered out." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.483, + 0.547 + ], + "angle": 0, + "content": "On the other hand, raw spectrum (4D range-Doppler-azimuth-elevation data cubes [50]) can be unintuitive and difficult to interpret compared to lidar point clouds or camera images, and include properties such as specularity and Doppler which lack straight-forward Cartesian interpretations [18]. As such, many machine learning methods have been proposed [24, 45, 69] to exploit 4D radar data from single-chip radars, achieving remarkable performance on 2D scene understanding tasks. However, due to the dominance of CFAR point clouds in radar processing, as well as the high data rate of raw mmWave radar data, most radar toolchains only process point-cloud data. Tooling for raw I/Q (in-phase/quadrature) data is often brittle, poorly documented, and largely unsupported by radar vendors, severely limiting the availability and scale of both raw mmWave datasets and the models which operate on raw data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.549, + 0.483, + 0.716 + ], + "angle": 0, + "content": "To rectify this limitation, we develop an open-source toolchain and associated large dataset specifically for 4D mmWave radar data. Training a radar-to-lidar model and fine-tuning for a range of other tasks, we demonstrate the surprising effectiveness of mmWave radar models trained at scale (Fig. 1). Going further, just as large foundational models [4] have greatly accelerated the pace of innovation in computer vision and natural language processing, we believe that a foundational model for raw mmWave radar trained at even larger scale could similarly supercharge the advancement of radar sensing techniques." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.722, + 0.483, + 0.797 + ], + "angle": 0, + "content": "Contributions In this paper, we develop a full stack1 for collecting data, training, and evaluating a transformer for 4D single-chip radar to quantify both the potential costs and benefits of training a foundational model at scale. To summarize our contributions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.8, + 0.483, + 0.86 + ], + "angle": 0, + "content": "(1) We develop a compact, lightweight multimodal data collection system (Sec. 3.1) capable of collecting synchronized raw radar, Lidar, and camera data which can be operated as a handheld device. Our system can be" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.258, + 0.905, + 0.3 + ], + "angle": 0, + "content": "Figure 2. While a transformer can generate Lidar-like depth and bird's eye view images, traditional CFAR point clouds are noisy, and have poor angular resolution - especially in the elevation axis." + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.332, + 0.905, + 0.362 + ], + "angle": 0, + "content": "easily replicated using off-the-shelf components, 3D printed parts, and our open-source software." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.363, + 0.905, + 0.452 + ], + "angle": 0, + "content": "(2) Using this data collection system, we collect a dataset, I/Q-1M (One Million IQ Frames), consisting of 29 hours of data \\( -8 \\times \\) longer than the next largest publicly available raw radar dataset - split between indoor, outdoor-handheld, and bike-mounted settings, each with different radar configurations (Sec. 3.2)." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.453, + 0.906, + 0.619 + ], + "angle": 0, + "content": "(3) Finally, using our dataset, we train a Generalizable Radar Transformer (GRT), which can output depth maps and segmentation images with quality which is typically only possible with much higher resolution radars. Using GRT, we then run ablations on common design choices (Sec. 5.1), quantify the scalability of GRTs with increasing dataset and model size (Sec. 5.2, 5.4), and demonstrate that our GRT can be readily finetuned for other tasks and settings (Sec. 5.3), including obtaining state-of-the-art performance on the Coloradar [23] dataset with 30-minutes of fine-tuning." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.363, + 0.906, + 0.619 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.642, + 0.905, + 0.657 + ], + "angle": 0, + "content": "Key Findings We summarize our key findings as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.66, + 0.905, + 0.735 + ], + "angle": 0, + "content": "- Radar models can generalize to different settings and radar configurations (Sec. 5.2), as well as across objectives (with some fine-tuning). This suggests great potential for a cross-domain foundational model to improve and accelerate the development of new radar models." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.735, + 0.905, + 0.824 + ], + "angle": 0, + "content": "- Using raw data yields outsized performance gains, equivalent to more than a \\(10 \\times\\) increase in training data (Sec. 5.1). While existing datasets largely focus on CFAR point clouds or other processed representations, we believe that more emphasis should be placed on making raw data available for research." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Existing mmWave radar datasets are vastly undersized. 24 hours of training data is not enough to saturate even a 4M parameter model! Our analysis suggests that at least \\(100 \\times\\) more data is required to exploit the full potential of radar transformer models (Sec. 5.4)." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.66, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1Our data collection system, dataset, code, and model can be found via our project site: https://wiselabcmu.github.io/grt/." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24656" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.483, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.295, + 0.483, + 0.325 + ], + "angle": 0, + "content": "Figure 3. Our data collection rig in its handheld (right) and bike-mounted (left) configurations; see App. A for additional images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.483, + 0.48 + ], + "angle": 0, + "content": "Limitations Despite its size compared to previous datasets, I/Q-1M is quite small compared to the datasets used to train modern vision transformers, which can exceed billions of samples [68]. I/Q-1M also only includes daylight conditions and fair weather, and lacks the scale to capture \"edge cases\" that would be represented in a larger dataset [22]. Finally, since I/Q-1M uses a single type of radar, we cannot evaluate generalization across different antenna configurations — only radar configurations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.493, + 0.233, + 0.508 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.517, + 0.483, + 0.684 + ], + "angle": 0, + "content": "4D Solid State Radar Excluding mechanical radars, which perceive the world as lidar-like 2D heatmaps [3, 5, 54] via a rotating antenna, solid-state mmWave radars operate by transmitting and receiving a sequence of frequency-modulated \"chirps\" from an array of transmit (TX) and receive (RX) antenna [19]; this data is typically (losslessly) transformed to a 4D range-Doppler-azimuth-elevation data cube using a 4D FFT [50], whose resolutions are constrained by bandwidth, form factor, and the integration window. We focus on single-chip radars which have compact form factors – and thus poor angular resolution." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Learning and Datasets for 4D Radar Most radar processing methods use point clouds extracted from the 4D cubes [36, 52] as inputs [1, 7, 32, 44, 55, 56], allowing them to re-use popular Lidar architectures or even pre-trained models such as PointNet [46]. However, since this discards much of the information contained in a 4D radar cube, many competing approaches propose to directly interpret the 4D radar cube using methods and architectures such as feedforward convolutional architectures [9, 38, 43, 49, 69], multiview convolutional networks across different tensor axes [13, 34, 40], U-Nets [26, 39, 45], diffusion models [70], and transformers [2, 15, 21, 21]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "These prior machine learning-based approaches rely on publicly available datasets with 4D radar data, including" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.16 + ], + "angle": 0, + "content": "Table 1. Comparison with other single-chip mmWave radar datasets; a frame refers to the number of unique radar-sensor samples. For comparisons with datasets using other types of radar, see App. A.3. Our dataset is significantly larger than previous single-chip radar datasets, enabling us to explore scaling up models." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.162, + 0.89, + 0.283 + ], + "angle": 0, + "content": "
Dataset4D Data CubeDataset Size
I/Q-1M (Ours)Yes29 hours (1M frames)
MilliPoint [7]No (3D Points)6.3 Hours (545k frames)
RaDICal [29]Yes3.6 Hours (394k frames)
CRUW [63]No (2D Map)3 hours (400k frames)
Coloradar [23]Yes2.4 hours (82k frames)
RadarHD [45]Yes200k frames
CARRADA [41]No (3D Cube)21 Minutes (13k frames)
RADDet [69]Yes10k frames
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.905, + 0.417 + ], + "angle": 0, + "content": "from both cascaded [42, 49] and single-chip [23, 29, 69] radars; however, existing datasets are relatively small, with 3.6-hour RaDICal [29] and 2.4-hour Coloradar [23] as the largest (Table 1). Due to the success of powerful but data-hungry [68] transformer models [61] in computer vision [12], we believe that limited data availability imposes a substantial bottleneck on learning for 4D radar." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.907, + 0.65 + ], + "angle": 0, + "content": "High-resolution Imaging from Low-Resolution Radar Due to the low angular resolution of single-chip radars, extracting high-resolution angular information can be challenging; as a result, prior work focuses on recovering 2D spatial information [15, 45, 70]. Thus, while prior methods can extract 3D information using high-resolution cascaded radars [10], 3D imaging from single-chip radars generally requires additional information such as structured motion or multiple views, for example segmentation and maps using a rotating single-chip radar [24], a 3D occupancy map using multiple views [18], or high resolution images from fixed trajectories using Synthetic Aperture Radar [14, 35, 66, 67]. Instead, we show that by leveraging a sufficiently large dataset, even single frames are sufficient to recover dense angular resolution in both azimuth and elevation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.663, + 0.838, + 0.68 + ], + "angle": 0, + "content": "3. Data Collection System and Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.905, + 0.78 + ], + "angle": 0, + "content": "Dataset scale is key to training and evaluating potential foundational models. As such, we developed a scalable data collection system (Fig. 3), which we used to collect a large raw mmWave radar dataset, consisting of 1M radar-lidar-camera samples over 29 hours (Table 1). For additional details on our dataset and data collection rig, see App. A." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.729, + 0.804 + ], + "angle": 0, + "content": "3.1. Data Collection System" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Our data collection system, red-rover, was built around a TI AWR1843 Radar, Lidar, Camera, and IMU which can be easily operated via a simple web app on a mobile phone. Our system records all data to a single hot-swappable external drive via a single linux computer which handles time synchronization, minimizing turnaround time." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24657" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.131 + ], + "angle": 0, + "content": "Table 2. Key specifications for each setting. Settings have varying max Doppler \\( D_{\\mathrm{max}} \\) and range \\( R_{\\mathrm{max}} \\); all traces used a fixed resolution of 64 Doppler and 256 range bins." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.134, + 0.481, + 0.196 + ], + "angle": 0, + "content": "
SettingSizeLengthAverage SpeedDmaxRmax
indoor310k8.9h1.0m/s1.2m/s11.2m
outdoor372k10.7h1.4m/s1.8m/s22.4m
bike333k9.3h5.4m/s8.0m/s22.4m
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.202, + 0.483, + 0.243 + ], + "angle": 0, + "content": "Table 3. Transformer sizes. Layers indicates the number of encoder + decoder layers; Speed indicates the (batched) inference throughput of each model on a single RTX 4090." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.246, + 0.457, + 0.332 + ], + "angle": 0, + "content": "
SizeLayersDimensionParamsSpeed
pico2 + 2256 (4 heads)3.9M750 fps
tiny3 + 3384 (6 heads)12.7M320 fps
small4 + 4512 (8 heads)28.9M170 fps
medium6 + 6640 (10 heads)69.4M84 fps
large9 + 9768 (12 heads)149M44 fps
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.484, + 0.466 + ], + "angle": 0, + "content": "We also designed our data collection system to have a compact, battery-operated form factor to allow for a variety of collection modalities, including handheld and bicycle mounted. This also allows us to collect data relevant for tasks such as indoor sensing and localization, which are underrepresented in existing datasets, while still collecting automotive-like data by mounting our system to an E-bike." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.475, + 0.242, + 0.49 + ], + "angle": 0, + "content": "3.2. Collected Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.483, + 0.542 + ], + "angle": 0, + "content": "We collected three roughly equally sized splits (Table 2) from indoor handheld, outdoor handheld, and bike-mounted settings on the CMU campus and Pittsburgh area:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.543, + 0.483, + 0.572 + ], + "angle": 0, + "content": "- indoor: inside buildings at a slow to moderate walking pace, visiting multiple floors and areas within each." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.573, + 0.483, + 0.619 + ], + "angle": 0, + "content": "- outdoor: neighborhoods ranging from single family detached to high density commercial zoning at a moderate to fast walking pace." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.619, + 0.483, + 0.648 + ], + "angle": 0, + "content": "- bike: bike rides in different directions from a set starting point with a moderate biking pace." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.543, + 0.483, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.483, + 0.695 + ], + "angle": 0, + "content": "Each setting features a mobile observer, with radar modulation parameters tuned for typical speeds. For sample data from each setting, see App. A.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.226, + 0.725 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Using a transformer architecture (Sec. 4.1), we train our Generalizable Radar Transformer (GRT) for a range of different tasks (Sec. 4.2-4.3), and evaluate it on our dataset using a rigorous statistical methodology (Sec. 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.279, + 0.818 + ], + "angle": 0, + "content": "4.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "While many architectural refinements exist for vision transformers [30, 48], as well as for radar specifically [15, 26], we use a direct adaptation (Fig. 4) of the original Transformer [61] and Vision Transformer [12] to focus on measuring the fundamental properties of Radar transformers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "Radar Processing From the (slow time, TX, RX, fast time) I/Q stream, we perform a 4-Dimensional FFT to obtain (range, Doppler, azimuth, elevation) dense 4D radar data cubes of size (256, 64, 8, 2), which we provide to the model as two channels consisting of the amplitude and phase angle. This data cube is patched along the range and Doppler dimensions into patches of size \\(4 \\times 2(\\times 8 \\times 2)\\), yielding an initial set of \\(64 \\times 32 = 2048\\) patch tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.221, + 0.907, + 0.328 + ], + "angle": 0, + "content": "Transformer Architecture Crucially, unlike a vision transformer [12], radar models take inputs that have different spatial axes than their outputs, with vastly different relative resolutions where they overlap. As such, we use a standard transformer with a decoder [61], with varying layers and widths (Table 3); for a full specification of our transformer architecture and training procedure, see App. B.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.907, + 0.458 + ], + "angle": 0, + "content": "Decoder Query To handle the \"change of basis\" between the input and output space, we use an architecture based on Perceiver I/O [20]. We start by concatenating a (learned) output token to the encoder (similar to standard vision transformer [8]). The encoder output corresponding to the output token is then tiled into the desired decoder shape with a 3D sinusoidal positional encoding applied and is used as the input to the decoder, which attends to the encoder outputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.86, + 0.484 + ], + "angle": 0, + "content": "4.2. Base Task: 3D Occupancy Classification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.906, + 0.641 + ], + "angle": 0, + "content": "An ideal foundational model training task should be easy to gather data for (e.g., using self-supervised learning) and closely aligned with a wide range of potential downstream tasks. As such, since we are primarily concerned with understanding the spatial relationship between 4D radar data and 3D space, we use 3D (polar) occupancy classification – predicting the occupancy of \\(64 \\times 128 \\times 64\\) range-azimuth elevation cells – as a base task, with Lidar as a ground truth. Our task uses a binary cross entropy objective, with some weighting to correct for cell sizes; see App. B.3 for details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Notably, in addition to being fully self-supervised, this task covers all three possible output dimensions (range, azimuth, and elevation), meaning that downstream tasks such as range-azimuth classification or azimuth-elevation segmentation can be cast as 2D slices of this 3D output. This enables us to fine-tune for tasks, even if they have different spatial dimensions, simply by replacing the output head and modifying the output positional encoding queries." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.644, + 0.787 + ], + "angle": 0, + "content": "4.3. Other Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.84 + ], + "angle": 0, + "content": "In order to evaluate GRT's suitability as a foundational model for downstream fine-tuning, we use three additional tasks, each representing different output dimensions:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Bird's Eye View (BEV) Occupancy: Similarly to [15, 45], we classify the \\(256 \\times 1024\\) range-azimuth polar occupancy using Lidar as a ground truth, with the range normalized to the radar's range resolution." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24658" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.907, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.224, + 0.907, + 0.266 + ], + "angle": 0, + "content": "Figure 4. The GRT architecture. 4D radar cubes are patched with a linear projection, and a sinusoidal positional encoding is added. A transformer architecture is then used, with a transformer decoder for dense outputs and a MLP decoder for Ego-Motion estimation; different output encodings are used depending on the output axes and resolution." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.28, + 0.482, + 0.37 + ], + "angle": 0, + "content": "- Semantic Segmentation: Similar to [24], we train our GRT to output \\(640 \\times 640\\) azimuth-elevation class labels. Since radars cannot feasibly identify many classes (e.g., poster vs. sign vs. wall) which a camera could, we use eight coarse categories: person, sky, vehicle, flat, nature, structure, ceiling, and object." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.37, + 0.483, + 0.461 + ], + "angle": 0, + "content": "- Ego-Motion Estimation: Since radars can \"natively\" measure velocity², we predict the velocity of the radar relative to its current orientation. Since ego-motion estimation does not require a dense output, we replace the transformer decoder with a multi-layer-perceptron decoder head with 3 layers of 512 units." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.28, + 0.483, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.463, + 0.396, + 0.478 + ], + "angle": 0, + "content": "For more detail about each task, see App. B.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.495, + 0.212, + 0.51 + ], + "angle": 0, + "content": "4.4. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.625 + ], + "angle": 0, + "content": "Due to the cost of scaling foundational models, false positives can result in significant wasted resources, especially if associated with a costly methodological change. As such, since our dataset cannot be treated as having an \"infinite sample size\", we calculate upper-bounded uncertainty estimates wherever possible. In order to ensure these results are statistically accurate, we take the following steps:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.628, + 0.483, + 0.673 + ], + "angle": 0, + "content": "- Geo-Split: Within each setting, \\(\\approx 1.5\\) hours of data was reserved as a test set, which we ensured to be geographically disjoint from the training set to prevent leakage [28]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.673, + 0.483, + 0.748 + ], + "angle": 0, + "content": "- Sample Size correction: Time series signals – e.g. radar-lidar-video tuples – cannot be viewed as independent samples; as such, the effective sample size, which we obtain from an autocorrelation-based estimate [51], must be used when calculating the standard error." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.749, + 0.483, + 0.794 + ], + "angle": 0, + "content": "- Paired z-Test: Using the fact that models are evaluated on the same test traces, we use a paired z-test on the relative performance of each model with respect to a baseline." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.628, + 0.483, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.857 + ], + "angle": 0, + "content": "We report each metric relative to its specified baseline by default, along with error bars for a two-sided \\(95\\%\\) confidence interval; using our procedure and dataset, we can measure differences of \\(1 - 2\\%\\) (App. B.4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.907, + 0.386 + ], + "angle": 0, + "content": "Validation Split and Data Size Sampling. We used the last \\(10\\%\\) of each training trace for validation (separate from the test set), with the first \\(90\\%\\) used for training. When training on reduced dataset sizes, we use the first \\(9\\%\\), \\(18\\%\\), and \\(45\\%\\) of each trace for training for \\(10\\%\\), \\(20\\%\\), \\(50\\%\\) dataset sizes respectively; to reduce the variance of our experiments, we always \\(10\\%\\) of each trace for validation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.398, + 0.6, + 0.413 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.906, + 0.513 + ], + "angle": 0, + "content": "Using our dataset, we first ran extensive ablation (Sec. 5.1), scaling (Sec. 5.2), and fine-tuning (Sec. 5.3) experiments which show the efficacy, scalability, and generalizability of GRT. Our experiments took 874 RTX 4090-hours\\(^3\\) of training time, with the GRT-small model taking 22 RTX 4090-hours to train." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.906, + 0.64 + ], + "angle": 0, + "content": "Model Performance Despite the low resolution (only 3 \\(\\mathrm{TX} \\times 4\\) TX antenna) of our radar, GRT is able to predict a range of outputs with remarkable quality (Fig. 5); we show additional examples in App. C.1. We also evaluated common metrics for each objective as an absolute reference (App. C.2); GRT achieves a 3D chamfer distance of 4.9 range bins, corresponding to \\(0.66\\mathrm{m}\\) indoors, \\(1.6\\mathrm{m}\\) outdoors, and \\(1.5\\mathrm{m}\\) on bike." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.648, + 0.679, + 0.662 + ], + "angle": 0, + "content": "5.1. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.906, + 0.776 + ], + "angle": 0, + "content": "Using our dataset and GRT, we performed ablation studies on several parameters that are independent of the underlying architecture and the task (Table 4). In particular, we find that several common practices – omitting Doppler information, using Angle-of-Arrival Estimates, and applying CFAR thresholding – result in degraded performance equivalent to more than a \\(10 \\times\\) reduction in training data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.856 + ], + "angle": 0, + "content": "Input Representation Our models use complex 4D radar cubes that losslessly capture all information measured by a radar; however, the common practice in radar models and data sets is to use processed higher-level representations. We benchmark three common approaches:" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.906, + 0.901 + ], + "angle": 0, + "content": "3Our experiments were run on a range of different machines with varying compute capacity, which we normalize with respect to a single RTX 4090. We only tracked training and validation time, with testing excluded." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)Sensors which provide \"absolute\" pose, e.g. Lidars and Cameras, must differentiate, while IMUs must integrate." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24659" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.297, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.091, + 0.498, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.091, + 0.701, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.091, + 0.904, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.398, + 0.908, + 0.441 + ], + "angle": 0, + "content": "Figure 5. Training with more data and fine tuning instead of starting from scratch lead to much higher performance. 3D occupancy maps are less noisy as seen in rendered depth images (left), semantic segmentation is cleaner (center left), bird's eye view occupancy is sharper (center right), and velocity estimation is more accurate (right). However, scaling model size (left) does not have a large impact." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.46, + 0.483, + 0.536 + ], + "angle": 0, + "content": "- Real (amplitude-only) 4D data cubes are not significantly different from complex data, indicating that the \"leftover\" phase from a Doppler FFT carries little additional information. Since using complex data has negligible compute overhead, we default to complex inputs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.536, + 0.483, + 0.625 + ], + "angle": 0, + "content": "- Angle of Arrival Estimates can be used to replace dense antenna measurements. Since our radar has only 2 elevation bins, we reduce the azimuth axis (8 bins) into an AoA estimate. This discards a substantial amount of information, leading to a \\(28.0\\%\\) loss increase, equivalent to more than a \\(10\\times\\) decrease in the size of the dataset." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.627, + 0.483, + 0.702 + ], + "angle": 0, + "content": "- Constant False Alarm Rate (CFAR) processing removes weak or noisy reflectors based on local estimates of background noise [36]; this ablation uses a p-value threshold of 0.05, and zeros out rejected points. This leads to an even higher \\(31.5\\%\\) loss increase." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.46, + 0.483, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Impact of Doppler With limited angular resolution, GRT is highly dependent on Doppler information, which can capture higher resolution geometry [18]. We find that removing the Doppler FFT from our processing pipeline (i.e. treating each 4D radar cube as a time series of 64 3D frames [45]) leads to a \\(22.5\\%\\) loss increase. As an additional ablation, we also shuffle the slow-time axis to fully destroy any Doppler information; this does not lead to a further significant loss increase, suggesting that off-the-shelf transformers cannot easily learn FFTs. Finally, we observe worse performance at low speeds since less Doppler information is available at slow speeds (App. C.3)." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.458, + 0.907, + 0.499 + ], + "angle": 0, + "content": "Table 4. Test loss for each ablation (smaller is better) relative to GRT-small trained on our full dataset, along with \\(95\\%\\) confidence intervals for the relative differences." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.504, + 0.905, + 0.687 + ], + "angle": 0, + "content": "
AblationRelative Test Loss
InputsAmplitude Only+0.04 ± 0.85%
Angle of Arrival+28.0 ± 2.12%
CFAR Thresholding+31.5 ± 2.38%
DopplerWithout Doppler FFT+22.5 ± 1.76%
Slow Time Shuffled+23.10 ± 1.94%
Post-Patch AxesDoppler-Az-EI+6.22 ± 1.11%
Range-Az-EI+6.27 ± 1.10%
Range-Doppler-Az-EI+4.18 ± 1.09%
AugmentationsNone+5.87 ± 1.32%
Scale, Phase, and Flip Only+3.89 ± 1.19%
Separate ModelsIndoor Data Only+5.76 ± 1.85%
Outdoor Data Only+5.58 ± 1.26%
Bike Data Only+2.77 ± 1.41%
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.842 + ], + "angle": 0, + "content": "Patch Axes Since 4D range-Doppler-azimuth-elevation radar data cubes have four axes with different properties, they do not have an obvious counterpart to the square patches used in Vision Transformers. Benchmarking four alternatives (App. B.1), with each resulting in 2048 total patches, we find that Range-Doppler patching where the azimuth and elevation axes are \"patched out\" (similar to [15, 49]) is the most effective, performing \\(\\approx 5\\%\\) better." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Data Augmentations We develop a range of data aug- mentations that together provide a modest but significant performance improvement (5.87±1.32%; Table 4); we pro" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24660" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.097, + 0.892, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.255, + 0.908, + 0.312 + ], + "angle": 0, + "content": "Figure 6. Scaling laws for mmWave radar transformers across indoor, outdoor, and bike test splits. Models and confidence intervals are measured relative to the small transformer trained on \\(10\\%\\) of the dataset. While our models show weak scaling over model size when trained on our dataset, we see strong log-linear scaling across dataset size of \\(\\approx 20\\%\\) loss decrease per \\(10\\times\\) increase in data size. For variants of this graph with respect to absolute metrics, see App. C.2." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.328, + 0.475, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.468, + 0.486, + 0.511 + ], + "angle": 0, + "content": "Figure 7. GRT compared to Baselines, measured with respect to Chamfer distance (in range bins) and test loss across different training set sizes, averaged across our dataset." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.531, + 0.424, + 0.547 + ], + "angle": 0, + "content": "vide details about each augmentation in App. B.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.555, + 0.421, + 0.57 + ], + "angle": 0, + "content": "5.2. Towards a Radar Foundational Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.578, + 0.484, + 0.73 + ], + "angle": 0, + "content": "Scalability vs. Baselines Since prior work on learning for single-chip radar focuses on 2D outputs, we benchmark our transformer-based approach against two prior architectures for 2D BEV Occupancy prediction: a U-Net-based model (RadarHD [45]), and a Swin Transformer-based model (T-FFTRadNet [15]), with minor architecture modifications to conform to our data dimensions (App. B.5). We find that GRT-small outperforms both baselines at all training splits (Fig. 7), demonstrating the suitability of transformer architectures for scalability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Scaling Laws Training 20 different models for 5 different sizes (Table 3) and dataset sizes ranging from \\(10 - 100\\%\\) of our data (Fig. 6), we observe a logarithmic improvement with data size of approximately \\(20\\%\\) improvement per \\(10\\times\\) increase in data, similar to early observations in computer vision [57]. This can also be seen qualitatively (Fig. 5), where models trained on more data produce much higher quality predictions. Similarly to vision transformers [57, 68], we also observe that larger models are more data efficient, although the magnitude of difference that we observe is much smaller due to our limited dataset size." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.335, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Table 5. Chamfer Distance (in meters) for 2D BEV occupancy prediction on the Coloradar dataset [23] by location; the geometric mean is listed to account for the varying difficulty of each location. A fine-tuned GRT model outperforms baselines trained only on Coloradar, including a state of the art diffusion-based model [70] and a U-Net based model [45]." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.422, + 0.902, + 0.53 + ], + "angle": 0, + "content": "
TraceGRT (Ours)Diffusion [70]RadarHD [45]
Geometric Mean0.981.191.73
ARPG Lab0.780.961.73
EC Hallways1.041.041.69
Aspen0.610.510.91
Longboard2.635.475.40
Outdoors1.842.373.10
Edgar0.360.440.60
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.562, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Generalizability across different settings We evaluate the ability of GRT to generalize across different settings by comparing a baseline model trained on combined indoor, outdoor, and bike data with models trained on each setting separately (Table 4). Despite the differences in these settings, the jointly trained model performs significantly better than models trained on each setting separately, confirming that data from different settings can be combined to train a single, stronger model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.855, + 0.727 + ], + "angle": 0, + "content": "5.3. As a Base Model for Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Dataset Fine-tuning We fine-tuned a small GRT model on the Coloradoar [23] dataset using a BEV Occupancy objective, and benchmarked the resulting model against two prior approaches trained only on ColoRadar, including a state-of-the-art diffusion model, using the same data splits and evaluation procedure [70]. Notably, despite using a modulation and resolution (128 range \\(\\times\\) 128 Doppler) which is not present in our dataset, GRT can be run without any architectural modifications, such as modifying the number of upsampling stages, as would be required for a convolutional architecture (App. B.5). After fine-tuning until vali" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24661" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.096, + 0.382, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.096, + 0.621, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.097, + 0.865, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.243, + 0.908, + 0.287 + ], + "angle": 0, + "content": "Figure 8. Fine tuning mmWave radar transformers to different downstream tasks. Models and confidence intervals are measured relative to the small transformer trained (from scratch) on \\(10\\%\\) of the dataset. Pre-training and fine tuning strongly impacts data efficiency, equivalent to up to a \\(5\\times\\) increase in dataset size (observed in the Segmentation task)." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.313, + 0.456, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.465, + 0.486, + 0.507 + ], + "angle": 0, + "content": "Figure 9. A linear projection of the observed logarithmic scaling to a test loss lower bound suggests that logarithmic scaling cannot continue beyond \\(100 \\times\\) our current dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.484, + 0.596 + ], + "angle": 0, + "content": "dation loss convergence (\\(\\approx 30\\) minutes of training using a single RTX 4090), GRT achieves substantially lower Chamfer distance than baselines trained only on ColoRadar, showing the value of easily tunable foundational models (Table 5)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.822 + ], + "angle": 0, + "content": "Task Fine-tuning We also fine-tuned GRT-small for each of our secondary tasks using \\(10 - 100\\%\\) of our dataset, and compared the results with models trained from scratch on the same proportions of the dataset. Following this procedure, we find substantial performance gains equivalent to up to a \\(5\\times\\) increase in dataset size compared to training from scratch (Fig. 8). This effect is especially pronounced when less data is available, with the performance benefits of fine tuning disappearing as the dataset is scaled for the BEV Occupancy objective but staying more or less constant for the Semantic Segmentation objective. We also observe this effect as a clear qualitative difference: fine-tuned models produce sharper and more accurate predictions than their counterparts trained from scratch (Fig. 5)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.388, + 0.849 + ], + "angle": 0, + "content": "5.4. How Much More Data is Needed?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Although we cannot directly observe performance saturation, we project how much data would be required to saturate a Radar Transformer using two different methods to ar" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.905, + 0.342 + ], + "angle": 0, + "content": "rive at a best guess of approximately 100M samples - \\(100\\times\\) our current dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.907, + 0.511 + ], + "angle": 0, + "content": "Linear Projection of Scaling Laws To lower-bound the possible test loss in our dataset, we trained a small model on the test set to approximate convergence. Assuming that the rate of improvement in model performance with increased training data cannot decrease, we extend our observed (Fig. 6) logarithmic scaling law to this lower bound to, in turn, estimate an upper bound for when the logarithmic trend will no longer hold (Fig. 9). This yields an estimate of \\(100 \\times\\) our current dataset size. For additional details justifying our estimation of this bound, see App. C.4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.909, + 0.709 + ], + "angle": 0, + "content": "Validation Curve Trends We observe that GRTs tend to stop improving (with respect to validation loss) after \\(\\approx 10\\) training epochs, regardless of model or dataset size (App. C.4); this is similar to trends observed in the training of data-constrained LLMs, which are also observed to saturate around 10 epochs [37]. Using vision transformers, whose scaling laws are well studied [68] due to the availability of internet data, as a reference, we expect training saturation to occur around \\(10^{2} - 10^{4}\\mathrm{M}\\) samples seen. Since each epoch in our dataset corresponds to \\(\\approx 1\\mathrm{M}\\) samples seen, this implies that \\(10\\times\\) to \\(1000\\times\\) our current dataset size is required to delay overfitting beyond this point." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.634, + 0.74 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we train a Generalizable Radar Transformer using a large, 29 hour (1M sample) dataset collected using our open-source data collection system and demonstrate that our Radar Transformer can generalize across datasets and settings, can be readily fine-tuned, and exhibits logarithmic scaling. While we believe that substantial gains are still possible through further data scaling, we hope that our dataset and baseline models will enable the community to revisit previous methods in new context and explore new capabilities made possible by a much larger dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24662" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.197 + ], + "angle": 0, + "content": "[1] Karim Armanious, Maurice Quach, Michael Ulrich, Timo Winterling, Johannes Friesen, Sascha Braun, Daniel Jenet, Yuri Feldman, Eitan Kosman, Philipp Rapp, et al. Bosch street dataset: A multi-modal dataset with imaging radar for automated driving. arXiv preprint arXiv:2407.12803, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Jie Bai, Lianqing Zheng, Sen Li, Bin Tan, Sihan Chen, and Libo Huang. Radar transformer: An object classification network based on 4d mmw imaging radar. Sensors, 21(11): 3854, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Dan Barnes, Matthew Gadd, Paul Murcutt, Paul Newman, and Ingmar Posner. The oxford radar robotcar dataset: A radar extension to the oxford robotcar dataset. In 2020 IEEE international conference on robotics and automation (ICRA), pages 6433-6438. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.484, + 0.465 + ], + "angle": 0, + "content": "[5] Keenan Burnett, David J Yoon, Yuchen Wu, Andrew Z Li, Haowei Zhang, Shichen Lu, Jingxing Qian, Wei-Kang Tseng, Andrew Lambert, Keith YK Leung, et al. Boreas: A multi-season autonomous driving dataset. The International Journal of Robotics Research, 42(1-2):33-42, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.485, + 0.549 + ], + "angle": 0, + "content": "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.606 + ], + "angle": 0, + "content": "[7] Han Cui, Shu Zhong, Jiacheng Wu, Zichao Shen, Naim Dahnoun, and Yiren Zhao. Milipoint: A point cloud dataset for mmwave radar. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.648 + ], + "angle": 0, + "content": "[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[9] Colin Decourt, Rufin VanRullen, Didier Salle, and Thomas Oberlin. Darod: A deep automotive radar object detector on range-doppler maps. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 112-118. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[10] Fangqiang Ding, Xiangyu Wen, Yunzhou Zhu, Yiming Li, and Chris Xiaoxuan Lu. Radar: Robust 3d occupancy prediction with 4d imaging radar. Advances in Neural Information Processing Systems, 37:101589-101617, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[11] Christopher Doer and Gert F Trommer. Yaw aided radar inertial odometry using manhattan world assumptions. In 2021 28th Saint Petersburg International Conference on Integrated Navigation Systems (ICINS), pages 1-9. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "scale. In International Conference on Learning Representations, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[13] Xiangyu Gao, Guanbin Xing, Sumit Roy, and Hui Liu. Ramp-cnn: A novel neural network for enhanced automotive radar object recognition. IEEE Sensors Journal, 21(4): 5119-5132, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[14] Xiangyu Gao, Sumit Roy, and Guanbin Xing. Mimo-sar: A hierarchical high-resolution imaging algorithm for mmwave fmew radar in autonomous driving. IEEE Transactions on Vehicular Technology, 70(8):7322-7334, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[15] James Giroux, Martin Bouchard, and Robert Laganiere. Tfftradnet: Object detection with swin vision transformers from raw adc radar signals. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4030-4039, 2023. 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.904, + 0.334 + ], + "angle": 0, + "content": "[16] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[17] Wolfgang Hess, Damon Kohler, Holger Rapp, and Daniel Andor. Real-time loop closure in 2d lidar slam. In 2016 IEEE international conference on robotics and automation (ICRA), pages 1271-1278. IEEE, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[18] Tianshu Huang, John Miller, Akarsh Prabhakara, Tao Jin, Tarana Laroia, Zico Kolter, and Anthony Rowe. Dart: Implicit doppler tomography for radar novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24118-24129, 2024. 2, 3, 6, 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.904, + 0.516 + ], + "angle": 0, + "content": "[19] Cesar Iovescu and Sandeep Rao. The fundamentals of millimeter wave sensors. *Texas Instruments*, pages 1–8, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[20] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[21] Yi Jin, Marcel Hoffmann, Anastasios Deligiannis, JuanCarlos Fuentes-Michel, and Martin Vossiek. Semantic segmentation-based occupancy grid map learning with automotive radar raw data. IEEE Transactions on Intelligent Vehicles, 9(1):216-230, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[22] Nidhi Kalra and Susan M Paddock. Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability? Transportation Research Part A: Policy and Practice, 94:182-193, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.717, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[23] Andrew Kramer, Kyle Harlow, Christopher Williams, and Christoffer Heckman. Coloradar: The direct 3d millimeter wave radar dataset. The International Journal of Robotics Research, 41(4):351-360, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.773, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[24] Haowen Lai, Gaoxiang Luo, Yifei Liu, and Mingmin Zhao. Enabling visual recognition at radio frequency. In Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, pages 388-403, 2024. 2, 3, 5, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[25] Xinrong Li, Xiaodong Wang, Qing Yang, and Song Fu. Signal processing for tdm mimo fmcw millimeter-wave radar sensors. IEEE Access, 9:167959-167971, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Yu-Jhe Li, Shawn Hunt, Jinhyung Park, Matthew O'Toole, and Kris Kitani. Azimuth super-resolution for fmcw radar in" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24663" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17504-17513, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[27] Jaime Lien, Nicholas Gillian, M Emre Karagozler, Patrick Amihood, Carsten Schwesig, Erik Olson, Hakim Raja, and Ivan Poupyrev. Soli: Ubiquitous gesture sensing with millimeter wave radar. ACM Transactions on Graphics (TOG), 35(4):1-19, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[28] Adam Lilja, Junsheng Fu, Erik Stenberg, and Lars Hammarstrand. Localization is all you evaluate: Data leakage in online mapping datasets and how to fix it. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22150-22159, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[29] Teck-Yian Lim, Spencer A Markowitz, and Minh N Do. Radical: A synchronized fmcw radar, depth, imu and rgb camera data dataset with low-level fmcw radar signals. IEEE Journal of Selected Topics in Signal Processing, 15(4):941-953, 2021. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.484, + 0.453 + ], + "angle": 0, + "content": "[31] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[32] Chris Xiaoxuan Lu, Muhamad Risqi U Saputra, Peijun Zhao, Yasin Almalioglu, Pedro PB De Gusmao, Changhao Chen, Ke Sun, Niki Trigoni, and Andrew Markham. milliego: single-chip mmwave radar aided egomotion estimation via deep sensor fusion. In Proceedings of the 18th Conference on Embedded Networked Sensor Systems, pages 109-122, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.552, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[33] Sohrab Madani, Jayden Guan, Waleed Ahmed, Saurabh Gupta, and Haitham Hassanieh. Radatron: Accurate detection using multi-resolution cascaded mimo radar. In European Conference on Computer Vision, pages 160-178. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.484, + 0.719 + ], + "angle": 0, + "content": "[34] Bence Major, Daniel Fontijne, Amin Ansari, Ravi Teja Sukhavasi, Radhika Gowaikar, Michael Hamilton, Sean Lee, Slawomir Grzechnik, and Sundar Subramanian. Vehicle detection with automotive radar using deep learning on range-azimuth-doppler tensors. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[35] Babak Mamandipoor, Greg Malysa, Amin Arababian, Upamanyu Madhow, and Karam Noujeim. 60 ghz synthetic aperture radar for short-range imaging: Theory and experiments. In 2014 48th Asilomar Conference on Signals, Systems and Computers, pages 553-558. IEEE, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[36] Gary Minkler and Jing Minkler. Cfar: the principles of automatic radar detection in clutter. *Nasa sti/recon technical report a*, 90:23371, 1990. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[37] Niklas Muennighoff, Alexander Rush, Boaz Barak, Teven Le Scao, Nouamane Tazi, Aleksandra Piktus, Sampo Pyysalo, Thomas Wolf, and Colin A Raffel. Scaling data-constrained language models. Advances in Neural Information Processing Systems, 36:50358-50376, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[38] Farzan Erlik Nowruzi, Dhanvin Kolhatkar, Prince Kapoor, Fahed Al Hassanat, Elnaz Jahani Heravi, Robert Laganiere, Julien Rebut, and Waqas Malik. Deep open space segmentation using automotive radar. In 2020 IEEE MTT-S International Conference on Microwaves for Intelligent Mobility (ICMIM), pages 1-4. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[39] Itai Orr, Moshik Cohen, and Zeev Zalevsky. High-resolution radar road segmentation using weakly supervised learning. Nature Machine Intelligence, 3(3):239-246, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[40] Arthur Ouaknine, Alasdair Newson, Patrick Pérez, Florence Tupin, and Julien Rebut. Multi-view radar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15671-15680, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.343 + ], + "angle": 0, + "content": "[41] Arthur Ouaknine, Alasdair Newson, Julien Rebut, Florence Tupin, and Patrick Pérez. Carrada dataset: Camera and automotive radar with range-angle-doppler annotations. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 5068-5075. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.906, + 0.399 + ], + "angle": 0, + "content": "[42] Dong-Hee Paek, Seung-Hyun Kong, and Kevin Tirta Wijaya. K-radar: 4d radar object detection for autonomous driving in various weather conditions. Advances in Neural Information Processing Systems, 35:3819-3829, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[43] Andras Palffy, Jiaao Dong, Julian FP Kooij, and Dariu M Gavrila. Cnn based road user detection using the 3d radar cube. IEEE Robotics and Automation Letters, 5(2):1263-1270, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[44] Akarsh Prabhakara, Diana Zhang, Chao Li, Sirajum Munir, Aswin C Sankaranarayanan, Anthony Rowe, and Swarun Kumar. Exploring mmwave radar and camera fusion for high-resolution and long-range depth imaging. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3995-4002. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.906, + 0.621 + ], + "angle": 0, + "content": "[45] Akarsh Prabhakara, Tao Jin, Arnav Das, Gantavya Bhatt, Lilly Kumari, Elahe Soltanaghai, Jeff Bilmes, Swarun Kumar, and Anthony Rowe. High resolution point clouds from mmwave radar. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4135-4142. IEEE, 2023. 2, 3, 4, 6, 7, 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.906, + 0.691 + ], + "angle": 0, + "content": "[46] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.906, + 0.774 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[48] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.906, + 0.899 + ], + "angle": 0, + "content": "[49] Julien Rebut, Arthur Ouaknine, Waqas Malik, and Patrick Pérez. Raw high-definition radar for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17021-17030, 2022. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24664" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "[50] Mark A Richards et al. Fundamentals of radar signal processing. McGraw-hill New York, 2005. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.482, + 0.15 + ], + "angle": 0, + "content": "[51] Christian P Robert, George Casella, and George Casella. Monte Carlo statistical methods. Springer, 1999. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.153, + 0.482, + 0.193 + ], + "angle": 0, + "content": "[52] Hermann Rohling. Radar cfar thresholding in clutter and multiple target situations. IEEE Transactions on Aerospace and Electronic Systems, AES-19(4):608-621, 1983. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.197, + 0.482, + 0.265 + ], + "angle": 0, + "content": "[53] Mark E Russell, Arthur Crain, Anthony Curran, Richard A Campbell, Clifford A Drubin, and Willian F Miccioli. Millimeter-wave radar sensor for automotive intelligent cruise control (icc). IEEE Transactions on microwave theory and techniques, 45(12):2444-2453, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.268, + 0.482, + 0.337 + ], + "angle": 0, + "content": "[54] Marcel Sheeny, Emanuele De Pellegrin, Saptarshi Mukherjee, Alireza Ahrabian, Sen Wang, and Andrew Wallace. Radiate: A radar dataset for automotive perception in bad weather. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 1-7. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.34, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[55] Xian Shuai, Yulin Shen, Yi Tang, Shuyao Shi, Luping Ji, and Guoliang Xing. millieye: A lightweight mmwave radar and camera fusion system for robust object detection. In Proceedings of the International Conference on Internet-of-Things Design and Implementation, pages 145–157, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.425, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[56] Akash Deep Singh, Sandeep Singh Sandha, Luis Garcia, and Mani Srivastava. Radhar: Human activity recognition from point clouds generated through a millimeter-wave radar. In Proceedings of the 3rd ACM Workshop on Millimeter-wave Networks and Sensing Systems, pages 51-56, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[57] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.555, + 0.482, + 0.595 + ], + "angle": 0, + "content": "[58] Texas Instruments. Automated doors reference design using ti mmwave sensors. Design Guide TIDEP-01018, Texas Instruments, Dallas, TX, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.482, + 0.638 + ], + "angle": 0, + "content": "[59] AWR1843AOP Single-chip 77- and 79-GHz FMCW mmWave Sensor Antennas-OnPackage (AOP). Texas Instruments, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.642, + 0.482, + 0.724 + ], + "angle": 0, + "content": "[60] Claudia Vasanelli, Fabian Roos, Andre Durr, Johannes Schlichenmaier, Philipp Hugler, Benedikt Meinecke, Maximilian Steiner, and Christian Waldschmidt. Calibration and direction-of-arrival estimation of millimeter-wave radars: A practical introduction. IEEE Antennas and Propagation Magazine, 62(6):34-45, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[61] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[62] Christian Waldschmidt, Juergen Hasch, and Wolfgang Menzel. Automotive radar—from first efforts to future systems. IEEE Journal of Microwaves, 1(1):135-148, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[63] Yizhou Wang, Zhongyu Jiang, Xiangyu Gao, Jenq-Neng Hwang, Guanbin Xing, and Hui Liu. Rodnet: Radar object detection using cross-modal supervision. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 504-513, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[64] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[65] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. On layer normalization in the transformer architecture. In International Conference on Machine Learning, pages 10524-10533. PMLR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[66] Hiroyoshi Yamada, Takumi Kobayashi, Yoshio Yamaguchi, and Yuuichi Sugiyama. High-resolution 2d sar imaging by the millimeter-wave automobile radar. In 2017 IEEE Conference on Antenna Measurements & Applications (CAMA), pages 149-150. IEEE, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[67] Muhammet Emin Yanik and Murat Torlak. Near-field mimosar millimeter-wave imaging with sparsely sampled aperture data. IEEE Access, 7:31801-31819, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.908, + 0.374 + ], + "angle": 0, + "content": "[68] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104-12113, 2022. 3, 7, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[69] Ao Zhang, Farzan Erlik Nowruzi, and Robert Laganiere. Raddet: Range-azimuth-doppler based radar object detection for dynamic road users. In 2021 18th Conference on Robots and Vision (CRV), pages 95–102. IEEE, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[70] Ruibin Zhang, Donglai Xue, Yuhan Wang, Ruixu Geng, and Fei Gao. Towards dense and accurate radar perception via efficient cross-modal diffusion model. IEEE Robotics and Automation Letters, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.556 + ], + "angle": 0, + "content": "[71] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24665" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Foundational Models for Single-Chip Radar/6536d138-fd2a-4e4b-a53a-8d73d592cae1_origin.pdf b/2025/Towards Foundational Models for Single-Chip Radar/6536d138-fd2a-4e4b-a53a-8d73d592cae1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0c71ece1e282ee13e9a5b448604407b8c367831c --- /dev/null +++ b/2025/Towards Foundational Models for Single-Chip Radar/6536d138-fd2a-4e4b-a53a-8d73d592cae1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:217de1a4132e0b70c125f36e1fe115321410179f5884e89c8a3d04b8be2f859d +size 1550439 diff --git a/2025/Towards Foundational Models for Single-Chip Radar/full.md b/2025/Towards Foundational Models for Single-Chip Radar/full.md new file mode 100644 index 0000000000000000000000000000000000000000..95efdae618f60b78682ec75407cdfdc3b126cfc5 --- /dev/null +++ b/2025/Towards Foundational Models for Single-Chip Radar/full.md @@ -0,0 +1,326 @@ +# Towards Foundational Models for Single-Chip Radar + +Tianshu Huang $^{12}$ Akarsh Prabhakara $^{3}$ Chuhan Chen $^{1}$ Jay Karhade $^{1}$ + +Deva Ramanan1 Matthew O'Toole1 Anthony Rowe12 + +$^{1}$ Carnegie Mellon University $^{2}$ Bosch Research $^{3}$ University of Wisconsin-Madison + +{tianshu2, chuhanc, jkarhade, deva, motoole2, agr}@andrew.cmu.edu, akarsh@cs.wisc.edu + +![](images/137c334653dcbcb903d389abe4bdb14e787b23ab3e5147ea3310e919bb2f800e.jpg) + +![](images/4fa09d54a8d0a93b1fef2488c4830a391eda11fca1a3efa1cc131284d937c450.jpg) + +![](images/2ac22028b3e3a658dd2784993604a0382e85469ce80fac6957332e76c96f105d.jpg) + +![](images/ba52efe88dde9657d76fc8d095960a31ea2c4476563ec5a53da98e49f96916a5.jpg) + +![](images/92eab0b4d91da4461c7e13b0f527f5c004dd85a8d2da5b54af5f6789ad73c027.jpg) +Figure 1. Using only 4D range-Doppler-azimuth-elevation data from a radar with a 3x4 antenna array - equivalent to a 0.26-megapixel camera, we trained (separate) radar transformers on 24 hours (1M radar-lidar-camera samples) of data to predict Bird's Eye View 2D occupancy (left), 3D occupancy (center), semantic segmentation (right), and Ego-Motion. + +![](images/0f467d7ce8d354ce9c9465777138412162cd3fd6a4886be26af56110494add1e.jpg) + +![](images/1a7eb955b38d8294b3ed2e64dd8a2da904f74849b8e939066d3535549593dcf4.jpg) + +# Abstract + +mmWave radars are compact, inexpensive, and durable sensors that are robust to occlusions and work regardless of environmental conditions, such as weather and darkness. However, this comes at the cost of poor angular resolution, especially for inexpensive single-chip radars, which are typically used in automotive and indoor sensing applications. Although many have proposed learning-based methods to mitigate this weakness, no standardized foundational models or large datasets for the mmWave radar have emerged, and practitioners have largely trained task-specific models from scratch using relatively small datasets. + +In this paper, we collect (to our knowledge) the largest available raw radar dataset with 1M samples (29 hours) and train a foundational model for 4D single-chip radar, which can predict 3D occupancy and semantic segmenta + +tion with quality that is typically only possible with much higher resolution sensors. We demonstrate that our Generalizable Radar Transformer (GRT) generalizes across diverse settings, can be fine-tuned for different tasks, and shows logarithmic data scaling of $20\%$ per $10\times$ data. We also run extensive ablations on common design decisions, and find that using raw radar data significantly outperforms widely-used lossy representations, equivalent to a $10\times$ increase in training data. Finally, we roughly estimate that $\approx 100M$ samples (3000 hours) of data are required to fully exploit the potential of GRT. + +# 1. Introduction + +As a compact, inexpensive [27], and robust solid-state sensor, mmWave radars are ideal for sensing applications ranging from simple automatic door openers [58] to autonomous + +drones [11] or vehicles [53, 62]. mmWave radars are rich sensors which can directly measure range and velocity while capturing a unique range of material properties [18]; however, this comes at the cost of poor angular resolution typically on the order of $15^{\circ}$ - orders of magnitude worse than cameras or lidars [59]. + +Radar data are typically processed into radar point clouds (Fig. 2) derived using Constant False Alarm Rate (CFAR) peak detectors [36, 52] combined with Angle-of-Arrival estimation techniques [60]. However, this is a substantially lossy process: while raw radar data suffers from unique noise patterns such as "bleed" and side lobes [25], weak reflectors and other signals can be hidden in this noise, which would ordinarily be filtered out. + +On the other hand, raw spectrum (4D range-Doppler-azimuth-elevation data cubes [50]) can be unintuitive and difficult to interpret compared to lidar point clouds or camera images, and include properties such as specularity and Doppler which lack straight-forward Cartesian interpretations [18]. As such, many machine learning methods have been proposed [24, 45, 69] to exploit 4D radar data from single-chip radars, achieving remarkable performance on 2D scene understanding tasks. However, due to the dominance of CFAR point clouds in radar processing, as well as the high data rate of raw mmWave radar data, most radar toolchains only process point-cloud data. Tooling for raw I/Q (in-phase/quadrature) data is often brittle, poorly documented, and largely unsupported by radar vendors, severely limiting the availability and scale of both raw mmWave datasets and the models which operate on raw data. + +To rectify this limitation, we develop an open-source toolchain and associated large dataset specifically for 4D mmWave radar data. Training a radar-to-lidar model and fine-tuning for a range of other tasks, we demonstrate the surprising effectiveness of mmWave radar models trained at scale (Fig. 1). Going further, just as large foundational models [4] have greatly accelerated the pace of innovation in computer vision and natural language processing, we believe that a foundational model for raw mmWave radar trained at even larger scale could similarly supercharge the advancement of radar sensing techniques. + +Contributions In this paper, we develop a full stack1 for collecting data, training, and evaluating a transformer for 4D single-chip radar to quantify both the potential costs and benefits of training a foundational model at scale. To summarize our contributions: + +(1) We develop a compact, lightweight multimodal data collection system (Sec. 3.1) capable of collecting synchronized raw radar, Lidar, and camera data which can be operated as a handheld device. Our system can be + +![](images/156b3552d8b134b3ac1dce53b706acf6043173047e707acceb81d3aac7feb320.jpg) +Figure 2. While a transformer can generate Lidar-like depth and bird's eye view images, traditional CFAR point clouds are noisy, and have poor angular resolution - especially in the elevation axis. + +easily replicated using off-the-shelf components, 3D printed parts, and our open-source software. + +(2) Using this data collection system, we collect a dataset, I/Q-1M (One Million IQ Frames), consisting of 29 hours of data $-8 \times$ longer than the next largest publicly available raw radar dataset - split between indoor, outdoor-handheld, and bike-mounted settings, each with different radar configurations (Sec. 3.2). +(3) Finally, using our dataset, we train a Generalizable Radar Transformer (GRT), which can output depth maps and segmentation images with quality which is typically only possible with much higher resolution radars. Using GRT, we then run ablations on common design choices (Sec. 5.1), quantify the scalability of GRTs with increasing dataset and model size (Sec. 5.2, 5.4), and demonstrate that our GRT can be readily finetuned for other tasks and settings (Sec. 5.3), including obtaining state-of-the-art performance on the Coloradar [23] dataset with 30-minutes of fine-tuning. + +Key Findings We summarize our key findings as follows: + +- Radar models can generalize to different settings and radar configurations (Sec. 5.2), as well as across objectives (with some fine-tuning). This suggests great potential for a cross-domain foundational model to improve and accelerate the development of new radar models. +- Using raw data yields outsized performance gains, equivalent to more than a $10 \times$ increase in training data (Sec. 5.1). While existing datasets largely focus on CFAR point clouds or other processed representations, we believe that more emphasis should be placed on making raw data available for research. +- Existing mmWave radar datasets are vastly undersized. 24 hours of training data is not enough to saturate even a 4M parameter model! Our analysis suggests that at least $100 \times$ more data is required to exploit the full potential of radar transformer models (Sec. 5.4). + +![](images/e21097ab446e6549c043536819c3baccf2cede017d87b679cf6a344f43d96f05.jpg) +Figure 3. Our data collection rig in its handheld (right) and bike-mounted (left) configurations; see App. A for additional images. + +Limitations Despite its size compared to previous datasets, I/Q-1M is quite small compared to the datasets used to train modern vision transformers, which can exceed billions of samples [68]. I/Q-1M also only includes daylight conditions and fair weather, and lacks the scale to capture "edge cases" that would be represented in a larger dataset [22]. Finally, since I/Q-1M uses a single type of radar, we cannot evaluate generalization across different antenna configurations — only radar configurations. + +# 2. Related Work + +4D Solid State Radar Excluding mechanical radars, which perceive the world as lidar-like 2D heatmaps [3, 5, 54] via a rotating antenna, solid-state mmWave radars operate by transmitting and receiving a sequence of frequency-modulated "chirps" from an array of transmit (TX) and receive (RX) antenna [19]; this data is typically (losslessly) transformed to a 4D range-Doppler-azimuth-elevation data cube using a 4D FFT [50], whose resolutions are constrained by bandwidth, form factor, and the integration window. We focus on single-chip radars which have compact form factors – and thus poor angular resolution. + +Learning and Datasets for 4D Radar Most radar processing methods use point clouds extracted from the 4D cubes [36, 52] as inputs [1, 7, 32, 44, 55, 56], allowing them to re-use popular Lidar architectures or even pre-trained models such as PointNet [46]. However, since this discards much of the information contained in a 4D radar cube, many competing approaches propose to directly interpret the 4D radar cube using methods and architectures such as feedforward convolutional architectures [9, 38, 43, 49, 69], multiview convolutional networks across different tensor axes [13, 34, 40], U-Nets [26, 39, 45], diffusion models [70], and transformers [2, 15, 21, 21]. + +These prior machine learning-based approaches rely on publicly available datasets with 4D radar data, including + +Table 1. Comparison with other single-chip mmWave radar datasets; a frame refers to the number of unique radar-sensor samples. For comparisons with datasets using other types of radar, see App. A.3. Our dataset is significantly larger than previous single-chip radar datasets, enabling us to explore scaling up models. + +
Dataset4D Data CubeDataset Size
I/Q-1M (Ours)Yes29 hours (1M frames)
MilliPoint [7]No (3D Points)6.3 Hours (545k frames)
RaDICal [29]Yes3.6 Hours (394k frames)
CRUW [63]No (2D Map)3 hours (400k frames)
Coloradar [23]Yes2.4 hours (82k frames)
RadarHD [45]Yes200k frames
CARRADA [41]No (3D Cube)21 Minutes (13k frames)
RADDet [69]Yes10k frames
+ +from both cascaded [42, 49] and single-chip [23, 29, 69] radars; however, existing datasets are relatively small, with 3.6-hour RaDICal [29] and 2.4-hour Coloradar [23] as the largest (Table 1). Due to the success of powerful but data-hungry [68] transformer models [61] in computer vision [12], we believe that limited data availability imposes a substantial bottleneck on learning for 4D radar. + +High-resolution Imaging from Low-Resolution Radar Due to the low angular resolution of single-chip radars, extracting high-resolution angular information can be challenging; as a result, prior work focuses on recovering 2D spatial information [15, 45, 70]. Thus, while prior methods can extract 3D information using high-resolution cascaded radars [10], 3D imaging from single-chip radars generally requires additional information such as structured motion or multiple views, for example segmentation and maps using a rotating single-chip radar [24], a 3D occupancy map using multiple views [18], or high resolution images from fixed trajectories using Synthetic Aperture Radar [14, 35, 66, 67]. Instead, we show that by leveraging a sufficiently large dataset, even single frames are sufficient to recover dense angular resolution in both azimuth and elevation. + +# 3. Data Collection System and Dataset + +Dataset scale is key to training and evaluating potential foundational models. As such, we developed a scalable data collection system (Fig. 3), which we used to collect a large raw mmWave radar dataset, consisting of 1M radar-lidar-camera samples over 29 hours (Table 1). For additional details on our dataset and data collection rig, see App. A. + +# 3.1. Data Collection System + +Our data collection system, red-rover, was built around a TI AWR1843 Radar, Lidar, Camera, and IMU which can be easily operated via a simple web app on a mobile phone. Our system records all data to a single hot-swappable external drive via a single linux computer which handles time synchronization, minimizing turnaround time. + +Table 2. Key specifications for each setting. Settings have varying max Doppler $D_{\mathrm{max}}$ and range $R_{\mathrm{max}}$ ; all traces used a fixed resolution of 64 Doppler and 256 range bins. + +
SettingSizeLengthAverage SpeedDmaxRmax
indoor310k8.9h1.0m/s1.2m/s11.2m
outdoor372k10.7h1.4m/s1.8m/s22.4m
bike333k9.3h5.4m/s8.0m/s22.4m
+ +Table 3. Transformer sizes. Layers indicates the number of encoder + decoder layers; Speed indicates the (batched) inference throughput of each model on a single RTX 4090. + +
SizeLayersDimensionParamsSpeed
pico2 + 2256 (4 heads)3.9M750 fps
tiny3 + 3384 (6 heads)12.7M320 fps
small4 + 4512 (8 heads)28.9M170 fps
medium6 + 6640 (10 heads)69.4M84 fps
large9 + 9768 (12 heads)149M44 fps
+ +We also designed our data collection system to have a compact, battery-operated form factor to allow for a variety of collection modalities, including handheld and bicycle mounted. This also allows us to collect data relevant for tasks such as indoor sensing and localization, which are underrepresented in existing datasets, while still collecting automotive-like data by mounting our system to an E-bike. + +# 3.2. Collected Data + +We collected three roughly equally sized splits (Table 2) from indoor handheld, outdoor handheld, and bike-mounted settings on the CMU campus and Pittsburgh area: + +- indoor: inside buildings at a slow to moderate walking pace, visiting multiple floors and areas within each. +- outdoor: neighborhoods ranging from single family detached to high density commercial zoning at a moderate to fast walking pace. +- bike: bike rides in different directions from a set starting point with a moderate biking pace. + +Each setting features a mobile observer, with radar modulation parameters tuned for typical speeds. For sample data from each setting, see App. A.4. + +# 4. Methodology + +Using a transformer architecture (Sec. 4.1), we train our Generalizable Radar Transformer (GRT) for a range of different tasks (Sec. 4.2-4.3), and evaluate it on our dataset using a rigorous statistical methodology (Sec. 4.4). + +# 4.1. Model Architecture + +While many architectural refinements exist for vision transformers [30, 48], as well as for radar specifically [15, 26], we use a direct adaptation (Fig. 4) of the original Transformer [61] and Vision Transformer [12] to focus on measuring the fundamental properties of Radar transformers. + +Radar Processing From the (slow time, TX, RX, fast time) I/Q stream, we perform a 4-Dimensional FFT to obtain (range, Doppler, azimuth, elevation) dense 4D radar data cubes of size (256, 64, 8, 2), which we provide to the model as two channels consisting of the amplitude and phase angle. This data cube is patched along the range and Doppler dimensions into patches of size $4 \times 2(\times 8 \times 2)$ , yielding an initial set of $64 \times 32 = 2048$ patch tokens. + +Transformer Architecture Crucially, unlike a vision transformer [12], radar models take inputs that have different spatial axes than their outputs, with vastly different relative resolutions where they overlap. As such, we use a standard transformer with a decoder [61], with varying layers and widths (Table 3); for a full specification of our transformer architecture and training procedure, see App. B.1. + +Decoder Query To handle the "change of basis" between the input and output space, we use an architecture based on Perceiver I/O [20]. We start by concatenating a (learned) output token to the encoder (similar to standard vision transformer [8]). The encoder output corresponding to the output token is then tiled into the desired decoder shape with a 3D sinusoidal positional encoding applied and is used as the input to the decoder, which attends to the encoder outputs. + +# 4.2. Base Task: 3D Occupancy Classification + +An ideal foundational model training task should be easy to gather data for (e.g., using self-supervised learning) and closely aligned with a wide range of potential downstream tasks. As such, since we are primarily concerned with understanding the spatial relationship between 4D radar data and 3D space, we use 3D (polar) occupancy classification – predicting the occupancy of $64 \times 128 \times 64$ range-azimuth elevation cells – as a base task, with Lidar as a ground truth. Our task uses a binary cross entropy objective, with some weighting to correct for cell sizes; see App. B.3 for details. + +Notably, in addition to being fully self-supervised, this task covers all three possible output dimensions (range, azimuth, and elevation), meaning that downstream tasks such as range-azimuth classification or azimuth-elevation segmentation can be cast as 2D slices of this 3D output. This enables us to fine-tune for tasks, even if they have different spatial dimensions, simply by replacing the output head and modifying the output positional encoding queries. + +# 4.3. Other Tasks + +In order to evaluate GRT's suitability as a foundational model for downstream fine-tuning, we use three additional tasks, each representing different output dimensions: + +- Bird's Eye View (BEV) Occupancy: Similarly to [15, 45], we classify the $256 \times 1024$ range-azimuth polar occupancy using Lidar as a ground truth, with the range normalized to the radar's range resolution. + +![](images/a916548feedfeb7a432a77b7806970f7adf97ff3da10cb152669c32ab994ec0a.jpg) +Figure 4. The GRT architecture. 4D radar cubes are patched with a linear projection, and a sinusoidal positional encoding is added. A transformer architecture is then used, with a transformer decoder for dense outputs and a MLP decoder for Ego-Motion estimation; different output encodings are used depending on the output axes and resolution. + +- Semantic Segmentation: Similar to [24], we train our GRT to output $640 \times 640$ azimuth-elevation class labels. Since radars cannot feasibly identify many classes (e.g., poster vs. sign vs. wall) which a camera could, we use eight coarse categories: person, sky, vehicle, flat, nature, structure, ceiling, and object. +- Ego-Motion Estimation: Since radars can "natively" measure velocity², we predict the velocity of the radar relative to its current orientation. Since ego-motion estimation does not require a dense output, we replace the transformer decoder with a multi-layer-perceptron decoder head with 3 layers of 512 units. + +For more detail about each task, see App. B.3. + +# 4.4. Evaluation + +Due to the cost of scaling foundational models, false positives can result in significant wasted resources, especially if associated with a costly methodological change. As such, since our dataset cannot be treated as having an "infinite sample size", we calculate upper-bounded uncertainty estimates wherever possible. In order to ensure these results are statistically accurate, we take the following steps: + +- Geo-Split: Within each setting, $\approx 1.5$ hours of data was reserved as a test set, which we ensured to be geographically disjoint from the training set to prevent leakage [28]. +- Sample Size correction: Time series signals – e.g. radar-lidar-video tuples – cannot be viewed as independent samples; as such, the effective sample size, which we obtain from an autocorrelation-based estimate [51], must be used when calculating the standard error. +- Paired z-Test: Using the fact that models are evaluated on the same test traces, we use a paired z-test on the relative performance of each model with respect to a baseline. + +We report each metric relative to its specified baseline by default, along with error bars for a two-sided $95\%$ confidence interval; using our procedure and dataset, we can measure differences of $1 - 2\%$ (App. B.4). + +Validation Split and Data Size Sampling. We used the last $10\%$ of each training trace for validation (separate from the test set), with the first $90\%$ used for training. When training on reduced dataset sizes, we use the first $9\%$ , $18\%$ , and $45\%$ of each trace for training for $10\%$ , $20\%$ , $50\%$ dataset sizes respectively; to reduce the variance of our experiments, we always $10\%$ of each trace for validation. + +# 5. Results + +Using our dataset, we first ran extensive ablation (Sec. 5.1), scaling (Sec. 5.2), and fine-tuning (Sec. 5.3) experiments which show the efficacy, scalability, and generalizability of GRT. Our experiments took 874 RTX 4090-hours $^3$ of training time, with the GRT-small model taking 22 RTX 4090-hours to train. + +Model Performance Despite the low resolution (only 3 $\mathrm{TX} \times 4$ TX antenna) of our radar, GRT is able to predict a range of outputs with remarkable quality (Fig. 5); we show additional examples in App. C.1. We also evaluated common metrics for each objective as an absolute reference (App. C.2); GRT achieves a 3D chamfer distance of 4.9 range bins, corresponding to $0.66\mathrm{m}$ indoors, $1.6\mathrm{m}$ outdoors, and $1.5\mathrm{m}$ on bike. + +# 5.1. Ablation Studies + +Using our dataset and GRT, we performed ablation studies on several parameters that are independent of the underlying architecture and the task (Table 4). In particular, we find that several common practices – omitting Doppler information, using Angle-of-Arrival Estimates, and applying CFAR thresholding – result in degraded performance equivalent to more than a $10 \times$ reduction in training data. + +Input Representation Our models use complex 4D radar cubes that losslessly capture all information measured by a radar; however, the common practice in radar models and data sets is to use processed higher-level representations. We benchmark three common approaches: + +![](images/34204fc446f05cb39fec29fd2b2438e1bfd0d5fc9ab884034ec0c67b22bae0a0.jpg) +Figure 5. Training with more data and fine tuning instead of starting from scratch lead to much higher performance. 3D occupancy maps are less noisy as seen in rendered depth images (left), semantic segmentation is cleaner (center left), bird's eye view occupancy is sharper (center right), and velocity estimation is more accurate (right). However, scaling model size (left) does not have a large impact. + +![](images/65bba12e933c69d364c63d49ce00aef092522ed72932741a9b5432da7e9275c9.jpg) + +![](images/3152cc6949a027a0bcfc98f66254f9a13e5bfde6306fa38ae0e1fd0f5c1602ed.jpg) + +![](images/dbf7f6ead04dc07b3af3c0890e115c36473dd1a40511afb4cf6b2a9f355fbc92.jpg) + +- Real (amplitude-only) 4D data cubes are not significantly different from complex data, indicating that the "leftover" phase from a Doppler FFT carries little additional information. Since using complex data has negligible compute overhead, we default to complex inputs. +- Angle of Arrival Estimates can be used to replace dense antenna measurements. Since our radar has only 2 elevation bins, we reduce the azimuth axis (8 bins) into an AoA estimate. This discards a substantial amount of information, leading to a $28.0\%$ loss increase, equivalent to more than a $10\times$ decrease in the size of the dataset. +- Constant False Alarm Rate (CFAR) processing removes weak or noisy reflectors based on local estimates of background noise [36]; this ablation uses a p-value threshold of 0.05, and zeros out rejected points. This leads to an even higher $31.5\%$ loss increase. + +Impact of Doppler With limited angular resolution, GRT is highly dependent on Doppler information, which can capture higher resolution geometry [18]. We find that removing the Doppler FFT from our processing pipeline (i.e. treating each 4D radar cube as a time series of 64 3D frames [45]) leads to a $22.5\%$ loss increase. As an additional ablation, we also shuffle the slow-time axis to fully destroy any Doppler information; this does not lead to a further significant loss increase, suggesting that off-the-shelf transformers cannot easily learn FFTs. Finally, we observe worse performance at low speeds since less Doppler information is available at slow speeds (App. C.3). + +Table 4. Test loss for each ablation (smaller is better) relative to GRT-small trained on our full dataset, along with $95\%$ confidence intervals for the relative differences. + +
AblationRelative Test Loss
InputsAmplitude Only+0.04 ± 0.85%
Angle of Arrival+28.0 ± 2.12%
CFAR Thresholding+31.5 ± 2.38%
DopplerWithout Doppler FFT+22.5 ± 1.76%
Slow Time Shuffled+23.10 ± 1.94%
Post-Patch AxesDoppler-Az-EI+6.22 ± 1.11%
Range-Az-EI+6.27 ± 1.10%
Range-Doppler-Az-EI+4.18 ± 1.09%
AugmentationsNone+5.87 ± 1.32%
Scale, Phase, and Flip Only+3.89 ± 1.19%
Separate ModelsIndoor Data Only+5.76 ± 1.85%
Outdoor Data Only+5.58 ± 1.26%
Bike Data Only+2.77 ± 1.41%
+ +Patch Axes Since 4D range-Doppler-azimuth-elevation radar data cubes have four axes with different properties, they do not have an obvious counterpart to the square patches used in Vision Transformers. Benchmarking four alternatives (App. B.1), with each resulting in 2048 total patches, we find that Range-Doppler patching where the azimuth and elevation axes are "patched out" (similar to [15, 49]) is the most effective, performing $\approx 5\%$ better. + +Data Augmentations We develop a range of data aug- mentations that together provide a modest but significant performance improvement (5.87±1.32%; Table 4); we pro + +![](images/1424cedc69f52fdfe5cc11b94620d3f35952def931b5567028af627c37ea51e5.jpg) +Figure 6. Scaling laws for mmWave radar transformers across indoor, outdoor, and bike test splits. Models and confidence intervals are measured relative to the small transformer trained on $10\%$ of the dataset. While our models show weak scaling over model size when trained on our dataset, we see strong log-linear scaling across dataset size of $\approx 20\%$ loss decrease per $10\times$ increase in data size. For variants of this graph with respect to absolute metrics, see App. C.2. + +![](images/9f3a8a275912cf64f0adaae5fa3fdf887528a12d5ef894aaf39583bd3e4fb7f2.jpg) +Figure 7. GRT compared to Baselines, measured with respect to Chamfer distance (in range bins) and test loss across different training set sizes, averaged across our dataset. + +vide details about each augmentation in App. B.2. + +# 5.2. Towards a Radar Foundational Model + +Scalability vs. Baselines Since prior work on learning for single-chip radar focuses on 2D outputs, we benchmark our transformer-based approach against two prior architectures for 2D BEV Occupancy prediction: a U-Net-based model (RadarHD [45]), and a Swin Transformer-based model (T-FFTRadNet [15]), with minor architecture modifications to conform to our data dimensions (App. B.5). We find that GRT-small outperforms both baselines at all training splits (Fig. 7), demonstrating the suitability of transformer architectures for scalability. + +Scaling Laws Training 20 different models for 5 different sizes (Table 3) and dataset sizes ranging from $10 - 100\%$ of our data (Fig. 6), we observe a logarithmic improvement with data size of approximately $20\%$ improvement per $10\times$ increase in data, similar to early observations in computer vision [57]. This can also be seen qualitatively (Fig. 5), where models trained on more data produce much higher quality predictions. Similarly to vision transformers [57, 68], we also observe that larger models are more data efficient, although the magnitude of difference that we observe is much smaller due to our limited dataset size. + +Table 5. Chamfer Distance (in meters) for 2D BEV occupancy prediction on the Coloradar dataset [23] by location; the geometric mean is listed to account for the varying difficulty of each location. A fine-tuned GRT model outperforms baselines trained only on Coloradar, including a state of the art diffusion-based model [70] and a U-Net based model [45]. + +
TraceGRT (Ours)Diffusion [70]RadarHD [45]
Geometric Mean0.981.191.73
ARPG Lab0.780.961.73
EC Hallways1.041.041.69
Aspen0.610.510.91
Longboard2.635.475.40
Outdoors1.842.373.10
Edgar0.360.440.60
+ +Generalizability across different settings We evaluate the ability of GRT to generalize across different settings by comparing a baseline model trained on combined indoor, outdoor, and bike data with models trained on each setting separately (Table 4). Despite the differences in these settings, the jointly trained model performs significantly better than models trained on each setting separately, confirming that data from different settings can be combined to train a single, stronger model. + +# 5.3. As a Base Model for Downstream Tasks + +Dataset Fine-tuning We fine-tuned a small GRT model on the Coloradoar [23] dataset using a BEV Occupancy objective, and benchmarked the resulting model against two prior approaches trained only on ColoRadar, including a state-of-the-art diffusion model, using the same data splits and evaluation procedure [70]. Notably, despite using a modulation and resolution (128 range $\times$ 128 Doppler) which is not present in our dataset, GRT can be run without any architectural modifications, such as modifying the number of upsampling stages, as would be required for a convolutional architecture (App. B.5). After fine-tuning until vali + +![](images/2080bec3029056116a0cdb45bc4aee236ef9ae367a01697420184a20dc477371.jpg) +Figure 8. Fine tuning mmWave radar transformers to different downstream tasks. Models and confidence intervals are measured relative to the small transformer trained (from scratch) on $10\%$ of the dataset. Pre-training and fine tuning strongly impacts data efficiency, equivalent to up to a $5\times$ increase in dataset size (observed in the Segmentation task). + +![](images/dda890857ac31b4718f1091d4f907fa3ddf26dc18e9ec681a9f1b882ab7b8041.jpg) + +![](images/8515653f915d2417d8f4577eb33934689676576b11ab10f0a4d57d0be252bbc1.jpg) + +![](images/60514d9ebc7b350aa40b2d511aa9c15631759c8c67de3aa4a45d762a78888f01.jpg) +Figure 9. A linear projection of the observed logarithmic scaling to a test loss lower bound suggests that logarithmic scaling cannot continue beyond $100 \times$ our current dataset + +dation loss convergence ( $\approx 30$ minutes of training using a single RTX 4090), GRT achieves substantially lower Chamfer distance than baselines trained only on ColoRadar, showing the value of easily tunable foundational models (Table 5). + +Task Fine-tuning We also fine-tuned GRT-small for each of our secondary tasks using $10 - 100\%$ of our dataset, and compared the results with models trained from scratch on the same proportions of the dataset. Following this procedure, we find substantial performance gains equivalent to up to a $5\times$ increase in dataset size compared to training from scratch (Fig. 8). This effect is especially pronounced when less data is available, with the performance benefits of fine tuning disappearing as the dataset is scaled for the BEV Occupancy objective but staying more or less constant for the Semantic Segmentation objective. We also observe this effect as a clear qualitative difference: fine-tuned models produce sharper and more accurate predictions than their counterparts trained from scratch (Fig. 5). + +# 5.4. How Much More Data is Needed? + +Although we cannot directly observe performance saturation, we project how much data would be required to saturate a Radar Transformer using two different methods to ar + +rive at a best guess of approximately 100M samples - $100\times$ our current dataset. + +Linear Projection of Scaling Laws To lower-bound the possible test loss in our dataset, we trained a small model on the test set to approximate convergence. Assuming that the rate of improvement in model performance with increased training data cannot decrease, we extend our observed (Fig. 6) logarithmic scaling law to this lower bound to, in turn, estimate an upper bound for when the logarithmic trend will no longer hold (Fig. 9). This yields an estimate of $100 \times$ our current dataset size. For additional details justifying our estimation of this bound, see App. C.4. + +Validation Curve Trends We observe that GRTs tend to stop improving (with respect to validation loss) after $\approx 10$ training epochs, regardless of model or dataset size (App. C.4); this is similar to trends observed in the training of data-constrained LLMs, which are also observed to saturate around 10 epochs [37]. Using vision transformers, whose scaling laws are well studied [68] due to the availability of internet data, as a reference, we expect training saturation to occur around $10^{2} - 10^{4}\mathrm{M}$ samples seen. Since each epoch in our dataset corresponds to $\approx 1\mathrm{M}$ samples seen, this implies that $10\times$ to $1000\times$ our current dataset size is required to delay overfitting beyond this point. + +# 6. Conclusion + +In this paper, we train a Generalizable Radar Transformer using a large, 29 hour (1M sample) dataset collected using our open-source data collection system and demonstrate that our Radar Transformer can generalize across datasets and settings, can be readily fine-tuned, and exhibits logarithmic scaling. While we believe that substantial gains are still possible through further data scaling, we hope that our dataset and baseline models will enable the community to revisit previous methods in new context and explore new capabilities made possible by a much larger dataset. + +# References + +[1] Karim Armanious, Maurice Quach, Michael Ulrich, Timo Winterling, Johannes Friesen, Sascha Braun, Daniel Jenet, Yuri Feldman, Eitan Kosman, Philipp Rapp, et al. Bosch street dataset: A multi-modal dataset with imaging radar for automated driving. arXiv preprint arXiv:2407.12803, 2024. 3 +[2] Jie Bai, Lianqing Zheng, Sen Li, Bin Tan, Sihan Chen, and Libo Huang. Radar transformer: An object classification network based on 4d mmw imaging radar. Sensors, 21(11): 3854, 2021. 3 +[3] Dan Barnes, Matthew Gadd, Paul Murcutt, Paul Newman, and Ingmar Posner. The oxford radar robotcar dataset: A radar extension to the oxford robotcar dataset. In 2020 IEEE international conference on robotics and automation (ICRA), pages 6433-6438. IEEE, 2020. 3 +[4] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2 +[5] Keenan Burnett, David J Yoon, Yuchen Wu, Andrew Z Li, Haowei Zhang, Shichen Lu, Jingxing Qian, Wei-Kang Tseng, Andrew Lambert, Keith YK Leung, et al. Boreas: A multi-season autonomous driving dataset. The International Journal of Robotics Research, 42(1-2):33-42, 2023. 3 +[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3 +[7] Han Cui, Shu Zhong, Jiacheng Wu, Zichao Shen, Naim Dahnoun, and Yiren Zhao. Milipoint: A point cloud dataset for mmwave radar. Advances in Neural Information Processing Systems, 36, 2024. 3 +[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 4 +[9] Colin Decourt, Rufin VanRullen, Didier Salle, and Thomas Oberlin. Darod: A deep automotive radar object detector on range-doppler maps. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 112-118. IEEE, 2022. 3 +[10] Fangqiang Ding, Xiangyu Wen, Yunzhou Zhu, Yiming Li, and Chris Xiaoxuan Lu. Radar: Robust 3d occupancy prediction with 4d imaging radar. Advances in Neural Information Processing Systems, 37:101589-101617, 2025. 3 +[11] Christopher Doer and Gert F Trommer. Yaw aided radar inertial odometry using manhattan world assumptions. In 2021 28th Saint Petersburg International Conference on Integrated Navigation Systems (ICINS), pages 1-9. IEEE, 2021. 2 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at + +scale. In International Conference on Learning Representations, 2021. 3, 4 +[13] Xiangyu Gao, Guanbin Xing, Sumit Roy, and Hui Liu. Ramp-cnn: A novel neural network for enhanced automotive radar object recognition. IEEE Sensors Journal, 21(4): 5119-5132, 2020. 3 +[14] Xiangyu Gao, Sumit Roy, and Guanbin Xing. Mimo-sar: A hierarchical high-resolution imaging algorithm for mmwave fmew radar in autonomous driving. IEEE Transactions on Vehicular Technology, 70(8):7322-7334, 2021. 3 +[15] James Giroux, Martin Bouchard, and Robert Laganiere. Tfftradnet: Object detection with swin vision transformers from raw adc radar signals. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4030-4039, 2023. 3, 4, 6, 7, 8 +[16] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 6 +[17] Wolfgang Hess, Damon Kohler, Holger Rapp, and Daniel Andor. Real-time loop closure in 2d lidar slam. In 2016 IEEE international conference on robotics and automation (ICRA), pages 1271-1278. IEEE, 2016. 7 +[18] Tianshu Huang, John Miller, Akarsh Prabhakara, Tao Jin, Tarana Laroia, Zico Kolter, and Anthony Rowe. Dart: Implicit doppler tomography for radar novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24118-24129, 2024. 2, 3, 6, 1, 8 +[19] Cesar Iovescu and Sandeep Rao. The fundamentals of millimeter wave sensors. *Texas Instruments*, pages 1–8, 2017. 3 +[20] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 4 +[21] Yi Jin, Marcel Hoffmann, Anastasios Deligiannis, JuanCarlos Fuentes-Michel, and Martin Vossiek. Semantic segmentation-based occupancy grid map learning with automotive radar raw data. IEEE Transactions on Intelligent Vehicles, 9(1):216-230, 2023. 3 +[22] Nidhi Kalra and Susan M Paddock. Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability? Transportation Research Part A: Policy and Practice, 94:182-193, 2016. 3 +[23] Andrew Kramer, Kyle Harlow, Christopher Williams, and Christoffer Heckman. Coloradar: The direct 3d millimeter wave radar dataset. The International Journal of Robotics Research, 41(4):351-360, 2022. 2, 3, 7 +[24] Haowen Lai, Gaoxiang Luo, Yifei Liu, and Mingmin Zhao. Enabling visual recognition at radio frequency. In Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, pages 388-403, 2024. 2, 3, 5, 9 +[25] Xinrong Li, Xiaodong Wang, Qing Yang, and Song Fu. Signal processing for tdm mimo fmcw millimeter-wave radar sensors. IEEE Access, 9:167959-167971, 2021. 2 +[26] Yu-Jhe Li, Shawn Hunt, Jinhyung Park, Matthew O'Toole, and Kris Kitani. Azimuth super-resolution for fmcw radar in + +autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17504-17513, 2023. 3, 4 +[27] Jaime Lien, Nicholas Gillian, M Emre Karagozler, Patrick Amihood, Carsten Schwesig, Erik Olson, Hakim Raja, and Ivan Poupyrev. Soli: Ubiquitous gesture sensing with millimeter wave radar. ACM Transactions on Graphics (TOG), 35(4):1-19, 2016. 1 +[28] Adam Lilja, Junsheng Fu, Erik Stenberg, and Lars Hammarstrand. Localization is all you evaluate: Data leakage in online mapping datasets and how to fix it. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22150-22159, 2024. 5 +[29] Teck-Yian Lim, Spencer A Markowitz, and Minh N Do. Radical: A synchronized fmcw radar, depth, imu and rgb camera data dataset with low-level fmcw radar signals. IEEE Journal of Selected Topics in Signal Processing, 15(4):941-953, 2021. 3, 1 +[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4 +[31] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017.6 +[32] Chris Xiaoxuan Lu, Muhamad Risqi U Saputra, Peijun Zhao, Yasin Almalioglu, Pedro PB De Gusmao, Changhao Chen, Ke Sun, Niki Trigoni, and Andrew Markham. milliego: single-chip mmwave radar aided egomotion estimation via deep sensor fusion. In Proceedings of the 18th Conference on Embedded Networked Sensor Systems, pages 109-122, 2020. 3 +[33] Sohrab Madani, Jayden Guan, Waleed Ahmed, Saurabh Gupta, and Haitham Hassanieh. Radatron: Accurate detection using multi-resolution cascaded mimo radar. In European Conference on Computer Vision, pages 160-178. Springer, 2022. 3 +[34] Bence Major, Daniel Fontijne, Amin Ansari, Ravi Teja Sukhavasi, Radhika Gowaikar, Michael Hamilton, Sean Lee, Slawomir Grzechnik, and Sundar Subramanian. Vehicle detection with automotive radar using deep learning on range-azimuth-doppler tensors. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 3 +[35] Babak Mamandipoor, Greg Malysa, Amin Arababian, Upamanyu Madhow, and Karam Noujeim. 60 ghz synthetic aperture radar for short-range imaging: Theory and experiments. In 2014 48th Asilomar Conference on Signals, Systems and Computers, pages 553-558. IEEE, 2014. 3 +[36] Gary Minkler and Jing Minkler. Cfar: the principles of automatic radar detection in clutter. *Nasa sti/recon technical report a*, 90:23371, 1990. 2, 3, 6 +[37] Niklas Muennighoff, Alexander Rush, Boaz Barak, Teven Le Scao, Nouamane Tazi, Aleksandra Piktus, Sampo Pyysalo, Thomas Wolf, and Colin A Raffel. Scaling data-constrained language models. Advances in Neural Information Processing Systems, 36:50358-50376, 2023. 8 + +[38] Farzan Erlik Nowruzi, Dhanvin Kolhatkar, Prince Kapoor, Fahed Al Hassanat, Elnaz Jahani Heravi, Robert Laganiere, Julien Rebut, and Waqas Malik. Deep open space segmentation using automotive radar. In 2020 IEEE MTT-S International Conference on Microwaves for Intelligent Mobility (ICMIM), pages 1-4. IEEE, 2020. 3 +[39] Itai Orr, Moshik Cohen, and Zeev Zalevsky. High-resolution radar road segmentation using weakly supervised learning. Nature Machine Intelligence, 3(3):239-246, 2021. 3 +[40] Arthur Ouaknine, Alasdair Newson, Patrick Pérez, Florence Tupin, and Julien Rebut. Multi-view radar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15671-15680, 2021. 3 +[41] Arthur Ouaknine, Alasdair Newson, Julien Rebut, Florence Tupin, and Patrick Pérez. Carrada dataset: Camera and automotive radar with range-angle-doppler annotations. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 5068-5075. IEEE, 2021. 3 +[42] Dong-Hee Paek, Seung-Hyun Kong, and Kevin Tirta Wijaya. K-radar: 4d radar object detection for autonomous driving in various weather conditions. Advances in Neural Information Processing Systems, 35:3819-3829, 2022. 3 +[43] Andras Palffy, Jiaao Dong, Julian FP Kooij, and Dariu M Gavrila. Cnn based road user detection using the 3d radar cube. IEEE Robotics and Automation Letters, 5(2):1263-1270, 2020. 3 +[44] Akarsh Prabhakara, Diana Zhang, Chao Li, Sirajum Munir, Aswin C Sankaranarayanan, Anthony Rowe, and Swarun Kumar. Exploring mmwave radar and camera fusion for high-resolution and long-range depth imaging. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3995-4002. IEEE, 2022. 3 +[45] Akarsh Prabhakara, Tao Jin, Arnav Das, Gantavya Bhatt, Lilly Kumari, Elahe Soltanaghai, Jeff Bilmes, Swarun Kumar, and Anthony Rowe. High resolution point clouds from mmwave radar. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4135-4142. IEEE, 2023. 2, 3, 4, 6, 7, 1, 8 +[46] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 13 +[48] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 4 +[49] Julien Rebut, Arthur Ouaknine, Waqas Malik, and Patrick Pérez. Raw high-definition radar for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17021-17030, 2022. 3, 6 + +[50] Mark A Richards et al. Fundamentals of radar signal processing. McGraw-hill New York, 2005. 2, 3 +[51] Christian P Robert, George Casella, and George Casella. Monte Carlo statistical methods. Springer, 1999. 5, 8 +[52] Hermann Rohling. Radar cfar thresholding in clutter and multiple target situations. IEEE Transactions on Aerospace and Electronic Systems, AES-19(4):608-621, 1983. 2, 3 +[53] Mark E Russell, Arthur Crain, Anthony Curran, Richard A Campbell, Clifford A Drubin, and Willian F Miccioli. Millimeter-wave radar sensor for automotive intelligent cruise control (icc). IEEE Transactions on microwave theory and techniques, 45(12):2444-2453, 1997. 2 +[54] Marcel Sheeny, Emanuele De Pellegrin, Saptarshi Mukherjee, Alireza Ahrabian, Sen Wang, and Andrew Wallace. Radiate: A radar dataset for automotive perception in bad weather. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 1-7. IEEE, 2021. 3 +[55] Xian Shuai, Yulin Shen, Yi Tang, Shuyao Shi, Luping Ji, and Guoliang Xing. millieye: A lightweight mmwave radar and camera fusion system for robust object detection. In Proceedings of the International Conference on Internet-of-Things Design and Implementation, pages 145–157, 2021. 3 +[56] Akash Deep Singh, Sandeep Singh Sandha, Luis Garcia, and Mani Srivastava. Radhar: Human activity recognition from point clouds generated through a millimeter-wave radar. In Proceedings of the 3rd ACM Workshop on Millimeter-wave Networks and Sensing Systems, pages 51-56, 2019. 3 +[57] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 7 +[58] Texas Instruments. Automated doors reference design using ti mmwave sensors. Design Guide TIDEP-01018, Texas Instruments, Dallas, TX, 2018. 1 +[59] AWR1843AOP Single-chip 77- and 79-GHz FMCW mmWave Sensor Antennas-OnPackage (AOP). Texas Instruments, 2021. 2, 3 +[60] Claudia Vasanelli, Fabian Roos, Andre Durr, Johannes Schlichenmaier, Philipp Hugler, Benedikt Meinecke, Maximilian Steiner, and Christian Waldschmidt. Calibration and direction-of-arrival estimation of millimeter-wave radars: A practical introduction. IEEE Antennas and Propagation Magazine, 62(6):34-45, 2020. 2 +[61] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3, 4, 6 +[62] Christian Waldschmidt, Juergen Hasch, and Wolfgang Menzel. Automotive radar—from first efforts to future systems. IEEE Journal of Microwaves, 1(1):135-148, 2021. 2 +[63] Yizhou Wang, Zhongyu Jiang, Xiangyu Gao, Jenq-Neng Hwang, Guanbin Xing, and Hui Liu. Rodnet: Radar object detection using cross-modal supervision. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 504-513, 2021. 3 +[64] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and + +efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 7 +[65] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. On layer normalization in the transformer architecture. In International Conference on Machine Learning, pages 10524-10533. PMLR, 2020. 6 +[66] Hiroyoshi Yamada, Takumi Kobayashi, Yoshio Yamaguchi, and Yuuichi Sugiyama. High-resolution 2d sar imaging by the millimeter-wave automobile radar. In 2017 IEEE Conference on Antenna Measurements & Applications (CAMA), pages 149-150. IEEE, 2017. 3 +[67] Muhammet Emin Yanik and Murat Torlak. Near-field mimosar millimeter-wave imaging with sparsely sampled aperture data. IEEE Access, 7:31801-31819, 2019. 3 +[68] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104-12113, 2022. 3, 7, 8, 13 +[69] Ao Zhang, Farzan Erlik Nowruzi, and Robert Laganiere. Raddet: Range-azimuth-doppler based radar object detection for dynamic road users. In 2021 18th Conference on Robots and Vision (CRV), pages 95–102. IEEE, 2021. 2, 3 +[70] Ruibin Zhang, Donglai Xue, Yuhan Wang, Ruixu Geng, and Fei Gao. Towards dense and accurate radar perception via efficient cross-modal diffusion model. IEEE Robotics and Automation Letters, 2024. 3, 7 +[71] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7 \ No newline at end of file diff --git a/2025/Towards Foundational Models for Single-Chip Radar/images.zip b/2025/Towards Foundational Models for Single-Chip Radar/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aab9f4dea44251ea04a5a779de79f936f24bb7a8 --- /dev/null +++ b/2025/Towards Foundational Models for Single-Chip Radar/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5922560061da65a80364f7daf399a2aac6901b499d6ff35b763b6d343bc9269e +size 616210 diff --git a/2025/Towards Foundational Models for Single-Chip Radar/layout.json b/2025/Towards Foundational Models for Single-Chip Radar/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3b9b1271eb53a1ff977899f42dfa81eea7282340 --- /dev/null +++ b/2025/Towards Foundational Models for Single-Chip Radar/layout.json @@ -0,0 +1,9269 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 103, + 471, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 103, + 471, + 121 + ], + "spans": [ + { + "bbox": [ + 140, + 103, + 471, + 121 + ], + "type": "text", + "content": "Towards Foundational Models for Single-Chip Radar" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 141, + 498, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 141, + 498, + 157 + ], + "spans": [ + { + "bbox": [ + 111, + 141, + 498, + 157 + ], + "type": "text", + "content": "Tianshu Huang" + }, + { + "bbox": [ + 111, + 141, + 498, + 157 + ], + "type": "inline_equation", + "content": "^{12}" + }, + { + "bbox": [ + 111, + 141, + 498, + 157 + ], + "type": "text", + "content": " Akarsh Prabhakara" + }, + { + "bbox": [ + 111, + 141, + 498, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 141, + 498, + 157 + ], + "type": "text", + "content": " Chuhan Chen" + }, + { + "bbox": [ + 111, + 141, + 498, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 111, + 141, + 498, + 157 + ], + "type": "text", + "content": " Jay Karhade" + }, + { + "bbox": [ + 111, + 141, + 498, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 167, + 158, + 465, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 158, + 465, + 171 + ], + "spans": [ + { + "bbox": [ + 167, + 158, + 465, + 171 + ], + "type": "text", + "content": "Deva Ramanan1 Matthew O'Toole1 Anthony Rowe12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 90, + 171, + 520, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 171, + 520, + 186 + ], + "spans": [ + { + "bbox": [ + 90, + 171, + 520, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 90, + 171, + 520, + 186 + ], + "type": "text", + "content": "Carnegie Mellon University " + }, + { + "bbox": [ + 90, + 171, + 520, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 90, + 171, + 520, + 186 + ], + "type": "text", + "content": "Bosch Research " + }, + { + "bbox": [ + 90, + 171, + 520, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 90, + 171, + 520, + 186 + ], + "type": "text", + "content": "University of Wisconsin-Madison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 74, + 187, + 533, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 187, + 533, + 199 + ], + "spans": [ + { + "bbox": [ + 74, + 187, + 533, + 199 + ], + "type": "text", + "content": "{tianshu2, chuhanc, jkarhade, deva, motoole2, agr}@andrew.cmu.edu, akarsh@cs.wisc.edu" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 130, + 205, + 480, + 290 + ], + "blocks": [ + { + "bbox": [ + 130, + 205, + 480, + 290 + ], + "lines": [ + { + "bbox": [ + 130, + 205, + 480, + 290 + ], + "spans": [ + { + "bbox": [ + 130, + 205, + 480, + 290 + ], + "type": "image", + "image_path": "137c334653dcbcb903d389abe4bdb14e787b23ab3e5147ea3310e919bb2f800e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 56, + 292, + 227, + 372 + ], + "blocks": [ + { + "bbox": [ + 56, + 292, + 227, + 372 + ], + "lines": [ + { + "bbox": [ + 56, + 292, + 227, + 372 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 227, + 372 + ], + "type": "image", + "image_path": "4fa09d54a8d0a93b1fef2488c4830a391eda11fca1a3efa1cc131284d937c450.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 228, + 293, + 390, + 372 + ], + "blocks": [ + { + "bbox": [ + 228, + 293, + 390, + 372 + ], + "lines": [ + { + "bbox": [ + 228, + 293, + 390, + 372 + ], + "spans": [ + { + "bbox": [ + 228, + 293, + 390, + 372 + ], + "type": "image", + "image_path": "2ac22028b3e3a658dd2784993604a0382e85469ce80fac6957332e76c96f105d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 391, + 293, + 553, + 372 + ], + "blocks": [ + { + "bbox": [ + 391, + 293, + 553, + 372 + ], + "lines": [ + { + "bbox": [ + 391, + 293, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 391, + 293, + 553, + 372 + ], + "type": "image", + "image_path": "ba52efe88dde9657d76fc8d095960a31ea2c4476563ec5a53da98e49f96916a5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 56, + 373, + 227, + 453 + ], + "blocks": [ + { + "bbox": [ + 56, + 373, + 227, + 453 + ], + "lines": [ + { + "bbox": [ + 56, + 373, + 227, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 227, + 453 + ], + "type": "image", + "image_path": "92eab0b4d91da4461c7e13b0f527f5c004dd85a8d2da5b54af5f6789ad73c027.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 456, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 55, + 456, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 555, + 489 + ], + "type": "text", + "content": "Figure 1. Using only 4D range-Doppler-azimuth-elevation data from a radar with a 3x4 antenna array - equivalent to a 0.26-megapixel camera, we trained (separate) radar transformers on 24 hours (1M radar-lidar-camera samples) of data to predict Bird's Eye View 2D occupancy (left), 3D occupancy (center), semantic segmentation (right), and Ego-Motion." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 228, + 373, + 390, + 452 + ], + "blocks": [ + { + "bbox": [ + 228, + 373, + 390, + 452 + ], + "lines": [ + { + "bbox": [ + 228, + 373, + 390, + 452 + ], + "spans": [ + { + "bbox": [ + 228, + 373, + 390, + 452 + ], + "type": "image", + "image_path": "0f467d7ce8d354ce9c9465777138412162cd3fd6a4886be26af56110494add1e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 391, + 373, + 553, + 452 + ], + "blocks": [ + { + "bbox": [ + 391, + 373, + 553, + 452 + ], + "lines": [ + { + "bbox": [ + 391, + 373, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 391, + 373, + 553, + 452 + ], + "type": "image", + "image_path": "1a7eb955b38d8294b3ed2e64dd8a2da904f74849b8e939066d3535549593dcf4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 151, + 506, + 200, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 506, + 200, + 519 + ], + "spans": [ + { + "bbox": [ + 151, + 506, + 200, + 519 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 533, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 664 + ], + "type": "text", + "content": "mmWave radars are compact, inexpensive, and durable sensors that are robust to occlusions and work regardless of environmental conditions, such as weather and darkness. However, this comes at the cost of poor angular resolution, especially for inexpensive single-chip radars, which are typically used in automotive and indoor sensing applications. Although many have proposed learning-based methods to mitigate this weakness, no standardized foundational models or large datasets for the mmWave radar have emerged, and practitioners have largely trained task-specific models from scratch using relatively small datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 665, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we collect (to our knowledge) the largest available raw radar dataset with 1M samples (29 hours) and train a foundational model for 4D single-chip radar, which can predict 3D occupancy and semantic segmenta" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "type": "text", + "content": "tion with quality that is typically only possible with much higher resolution sensors. We demonstrate that our Generalizable Radar Transformer (GRT) generalizes across diverse settings, can be fine-tuned for different tasks, and shows logarithmic data scaling of " + }, + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "type": "text", + "content": " per " + }, + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "type": "text", + "content": " data. We also run extensive ablations on common design decisions, and find that using raw radar data significantly outperforms widely-used lossy representations, equivalent to a " + }, + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "type": "text", + "content": " increase in training data. Finally, we roughly estimate that " + }, + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\approx 100M" + }, + { + "bbox": [ + 313, + 508, + 555, + 640 + ], + "type": "text", + "content": " samples (3000 hours) of data are required to fully exploit the potential of GRT." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "spans": [ + { + "bbox": [ + 314, + 657, + 395, + 670 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "As a compact, inexpensive [27], and robust solid-state sensor, mmWave radars are ideal for sensing applications ranging from simple automatic door openers [58] to autonomous" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24655" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "drones [11] or vehicles [53, 62]. mmWave radars are rich sensors which can directly measure range and velocity while capturing a unique range of material properties [18]; however, this comes at the cost of poor angular resolution typically on the order of " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " - orders of magnitude worse than cameras or lidars [59]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 145, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 294, + 240 + ], + "type": "text", + "content": "Radar data are typically processed into radar point clouds (Fig. 2) derived using Constant False Alarm Rate (CFAR) peak detectors [36, 52] combined with Angle-of-Arrival estimation techniques [60]. However, this is a substantially lossy process: while raw radar data suffers from unique noise patterns such as \"bleed\" and side lobes [25], weak reflectors and other signals can be hidden in this noise, which would ordinarily be filtered out." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 242, + 295, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 295, + 433 + ], + "type": "text", + "content": "On the other hand, raw spectrum (4D range-Doppler-azimuth-elevation data cubes [50]) can be unintuitive and difficult to interpret compared to lidar point clouds or camera images, and include properties such as specularity and Doppler which lack straight-forward Cartesian interpretations [18]. As such, many machine learning methods have been proposed [24, 45, 69] to exploit 4D radar data from single-chip radars, achieving remarkable performance on 2D scene understanding tasks. However, due to the dominance of CFAR point clouds in radar processing, as well as the high data rate of raw mmWave radar data, most radar toolchains only process point-cloud data. Tooling for raw I/Q (in-phase/quadrature) data is often brittle, poorly documented, and largely unsupported by radar vendors, severely limiting the availability and scale of both raw mmWave datasets and the models which operate on raw data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 567 + ], + "type": "text", + "content": "To rectify this limitation, we develop an open-source toolchain and associated large dataset specifically for 4D mmWave radar data. Training a radar-to-lidar model and fine-tuning for a range of other tasks, we demonstrate the surprising effectiveness of mmWave radar models trained at scale (Fig. 1). Going further, just as large foundational models [4] have greatly accelerated the pace of innovation in computer vision and natural language processing, we believe that a foundational model for raw mmWave radar trained at even larger scale could similarly supercharge the advancement of radar sensing techniques." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 571, + 295, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 631 + ], + "type": "text", + "content": "Contributions In this paper, we develop a full stack1 for collecting data, training, and evaluating a transformer for 4D single-chip radar to quantify both the potential costs and benefits of training a foundational model at scale. To summarize our contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 633, + 295, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 295, + 681 + ], + "type": "text", + "content": "(1) We develop a compact, lightweight multimodal data collection system (Sec. 3.1) capable of collecting synchronized raw radar, Lidar, and camera data which can be operated as a handheld device. Our system can be" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 202 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 202 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 202 + ], + "type": "image", + "image_path": "156b3552d8b134b3ac1dce53b706acf6043173047e707acceb81d3aac7feb320.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 204, + 553, + 237 + ], + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 237 + ], + "type": "text", + "content": "Figure 2. While a transformer can generate Lidar-like depth and bird's eye view images, traditional CFAR point clouds are noisy, and have poor angular resolution - especially in the elevation axis." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 330, + 262, + 553, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 262, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 330, + 262, + 553, + 286 + ], + "type": "text", + "content": "easily replicated using off-the-shelf components, 3D printed parts, and our open-source software." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 287, + 554, + 490 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 315, + 287, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 287, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 315, + 287, + 553, + 357 + ], + "type": "text", + "content": "(2) Using this data collection system, we collect a dataset, I/Q-1M (One Million IQ Frames), consisting of 29 hours of data " + }, + { + "bbox": [ + 315, + 287, + 553, + 357 + ], + "type": "inline_equation", + "content": "-8 \\times" + }, + { + "bbox": [ + 315, + 287, + 553, + 357 + ], + "type": "text", + "content": " longer than the next largest publicly available raw radar dataset - split between indoor, outdoor-handheld, and bike-mounted settings, each with different radar configurations (Sec. 3.2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 358, + 554, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 358, + 554, + 490 + ], + "spans": [ + { + "bbox": [ + 315, + 358, + 554, + 490 + ], + "type": "text", + "content": "(3) Finally, using our dataset, we train a Generalizable Radar Transformer (GRT), which can output depth maps and segmentation images with quality which is typically only possible with much higher resolution radars. Using GRT, we then run ablations on common design choices (Sec. 5.1), quantify the scalability of GRTs with increasing dataset and model size (Sec. 5.2, 5.4), and demonstrate that our GRT can be readily finetuned for other tasks and settings (Sec. 5.3), including obtaining state-of-the-art performance on the Coloradar [23] dataset with 30-minutes of fine-tuning." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 508, + 553, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 508, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 314, + 508, + 553, + 520 + ], + "type": "text", + "content": "Key Findings We summarize our key findings as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 522, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 315, + 522, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 522, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 315, + 522, + 553, + 582 + ], + "type": "text", + "content": "- Radar models can generalize to different settings and radar configurations (Sec. 5.2), as well as across objectives (with some fine-tuning). This suggests great potential for a cross-domain foundational model to improve and accelerate the development of new radar models." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 582, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 582, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 315, + 582, + 553, + 652 + ], + "type": "text", + "content": "- Using raw data yields outsized performance gains, equivalent to more than a " + }, + { + "bbox": [ + 315, + 582, + 553, + 652 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 315, + 582, + 553, + 652 + ], + "type": "text", + "content": " increase in training data (Sec. 5.1). While existing datasets largely focus on CFAR point clouds or other processed representations, we believe that more emphasis should be placed on making raw data available for research." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 654, + 553, + 713 + ], + "type": "text", + "content": "- Existing mmWave radar datasets are vastly undersized. 24 hours of training data is not enough to saturate even a 4M parameter model! Our analysis suggests that at least " + }, + { + "bbox": [ + 315, + 654, + 553, + 713 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 315, + 654, + 553, + 713 + ], + "type": "text", + "content": " more data is required to exploit the full potential of radar transformer models (Sec. 5.4)." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1Our data collection system, dataset, code, and model can be found via our project site: https://wiselabcmu.github.io/grt/." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24656" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 295, + 229 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 295, + 229 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 295, + 229 + ], + "type": "image", + "image_path": "e21097ab446e6549c043536819c3baccf2cede017d87b679cf6a344f43d96f05.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 233, + 295, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 295, + 257 + ], + "type": "text", + "content": "Figure 3. Our data collection rig in its handheld (right) and bike-mounted (left) configurations; see App. A for additional images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 271, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 295, + 380 + ], + "type": "text", + "content": "Limitations Despite its size compared to previous datasets, I/Q-1M is quite small compared to the datasets used to train modern vision transformers, which can exceed billions of samples [68]. I/Q-1M also only includes daylight conditions and fair weather, and lacks the scale to capture \"edge cases\" that would be represented in a larger dataset [22]. Finally, since I/Q-1M uses a single type of radar, we cannot evaluate generalization across different antenna configurations — only radar configurations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 390, + 142, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 142, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 142, + 402 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 409, + 295, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 409, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 54, + 409, + 295, + 541 + ], + "type": "text", + "content": "4D Solid State Radar Excluding mechanical radars, which perceive the world as lidar-like 2D heatmaps [3, 5, 54] via a rotating antenna, solid-state mmWave radars operate by transmitting and receiving a sequence of frequency-modulated \"chirps\" from an array of transmit (TX) and receive (RX) antenna [19]; this data is typically (losslessly) transformed to a 4D range-Doppler-azimuth-elevation data cube using a 4D FFT [50], whose resolutions are constrained by bandwidth, form factor, and the integration window. We focus on single-chip radars which have compact form factors – and thus poor angular resolution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 545, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 295, + 689 + ], + "type": "text", + "content": "Learning and Datasets for 4D Radar Most radar processing methods use point clouds extracted from the 4D cubes [36, 52] as inputs [1, 7, 32, 44, 55, 56], allowing them to re-use popular Lidar architectures or even pre-trained models such as PointNet [46]. However, since this discards much of the information contained in a 4D radar cube, many competing approaches propose to directly interpret the 4D radar cube using methods and architectures such as feedforward convolutional architectures [9, 38, 43, 49, 69], multiview convolutional networks across different tensor axes [13, 34, 40], U-Nets [26, 39, 45], diffusion models [70], and transformers [2, 15, 21, 21]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "These prior machine learning-based approaches rely on publicly available datasets with 4D radar data, including" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 128, + 544, + 224 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 126 + ], + "type": "text", + "content": "Table 1. Comparison with other single-chip mmWave radar datasets; a frame refers to the number of unique radar-sensor samples. For comparisons with datasets using other types of radar, see App. A.3. Our dataset is significantly larger than previous single-chip radar datasets, enabling us to explore scaling up models." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 128, + 544, + 224 + ], + "lines": [ + { + "bbox": [ + 323, + 128, + 544, + 224 + ], + "spans": [ + { + "bbox": [ + 323, + 128, + 544, + 224 + ], + "type": "table", + "html": "
Dataset4D Data CubeDataset Size
I/Q-1M (Ours)Yes29 hours (1M frames)
MilliPoint [7]No (3D Points)6.3 Hours (545k frames)
RaDICal [29]Yes3.6 Hours (394k frames)
CRUW [63]No (2D Map)3 hours (400k frames)
Coloradar [23]Yes2.4 hours (82k frames)
RadarHD [45]Yes200k frames
CARRADA [41]No (3D Cube)21 Minutes (13k frames)
RADDet [69]Yes10k frames
", + "image_path": "2cd822d5d879b3feb7312e967cb6a2f7f70fbb58174ce4944f138655e3bd4121.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 245, + 553, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 553, + 330 + ], + "type": "text", + "content": "from both cascaded [42, 49] and single-chip [23, 29, 69] radars; however, existing datasets are relatively small, with 3.6-hour RaDICal [29] and 2.4-hour Coloradar [23] as the largest (Table 1). Due to the success of powerful but data-hungry [68] transformer models [61] in computer vision [12], we believe that limited data availability imposes a substantial bottleneck on learning for 4D radar." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 335, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 514 + ], + "type": "text", + "content": "High-resolution Imaging from Low-Resolution Radar Due to the low angular resolution of single-chip radars, extracting high-resolution angular information can be challenging; as a result, prior work focuses on recovering 2D spatial information [15, 45, 70]. Thus, while prior methods can extract 3D information using high-resolution cascaded radars [10], 3D imaging from single-chip radars generally requires additional information such as structured motion or multiple views, for example segmentation and maps using a rotating single-chip radar [24], a 3D occupancy map using multiple views [18], or high resolution images from fixed trajectories using Synthetic Aperture Radar [14, 35, 66, 67]. Instead, we show that by leveraging a sufficiently large dataset, even single frames are sufficient to recover dense angular resolution in both azimuth and elevation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 525, + 512, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 512, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 512, + 538 + ], + "type": "text", + "content": "3. Data Collection System and Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 544, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 617 + ], + "type": "text", + "content": "Dataset scale is key to training and evaluating potential foundational models. As such, we developed a scalable data collection system (Fig. 3), which we used to collect a large raw mmWave radar dataset, consisting of 1M radar-lidar-camera samples over 29 hours (Table 1). For additional details on our dataset and data collection rig, see App. A." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 624, + 446, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 446, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 446, + 636 + ], + "type": "text", + "content": "3.1. Data Collection System" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 714 + ], + "type": "text", + "content": "Our data collection system, red-rover, was built around a TI AWR1843 Radar, Lidar, Camera, and IMU which can be easily operated via a simple web app on a mobile phone. Our system records all data to a single hot-swappable external drive via a single linux computer which handles time synchronization, minimizing turnaround time." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24657" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 106, + 294, + 155 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 103 + ], + "type": "text", + "content": "Table 2. Key specifications for each setting. Settings have varying max Doppler " + }, + { + "bbox": [ + 55, + 71, + 295, + 103 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 71, + 295, + 103 + ], + "type": "text", + "content": " and range " + }, + { + "bbox": [ + 55, + 71, + 295, + 103 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 71, + 295, + 103 + ], + "type": "text", + "content": "; all traces used a fixed resolution of 64 Doppler and 256 range bins." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 106, + 294, + 155 + ], + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 155 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 155 + ], + "type": "table", + "html": "
SettingSizeLengthAverage SpeedDmaxRmax
indoor310k8.9h1.0m/s1.2m/s11.2m
outdoor372k10.7h1.4m/s1.8m/s22.4m
bike333k9.3h5.4m/s8.0m/s22.4m
", + "image_path": "adc65dd3ccb2ab4bdc50bd1b3f1d524059ca8548c806ffe3d4a4b7a0b911f936.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 71, + 194, + 279, + 262 + ], + "blocks": [ + { + "bbox": [ + 55, + 159, + 295, + 192 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 295, + 192 + ], + "type": "text", + "content": "Table 3. Transformer sizes. Layers indicates the number of encoder + decoder layers; Speed indicates the (batched) inference throughput of each model on a single RTX 4090." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 194, + 279, + 262 + ], + "lines": [ + { + "bbox": [ + 71, + 194, + 279, + 262 + ], + "spans": [ + { + "bbox": [ + 71, + 194, + 279, + 262 + ], + "type": "table", + "html": "
SizeLayersDimensionParamsSpeed
pico2 + 2256 (4 heads)3.9M750 fps
tiny3 + 3384 (6 heads)12.7M320 fps
small4 + 4512 (8 heads)28.9M170 fps
medium6 + 6640 (10 heads)69.4M84 fps
large9 + 9768 (12 heads)149M44 fps
", + "image_path": "fa8a223c402f10115e5931c95778be6a309470b9197511a0eae1551ceb8e517e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 285, + 296, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 369 + ], + "type": "text", + "content": "We also designed our data collection system to have a compact, battery-operated form factor to allow for a variety of collection modalities, including handheld and bicycle mounted. This also allows us to collect data relevant for tasks such as indoor sensing and localization, which are underrepresented in existing datasets, while still collecting automotive-like data by mounting our system to an E-bike." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 376, + 148, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 148, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 148, + 388 + ], + "type": "text", + "content": "3.2. Collected Data" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 393, + 295, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 295, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 295, + 429 + ], + "type": "text", + "content": "We collected three roughly equally sized splits (Table 2) from indoor handheld, outdoor handheld, and bike-mounted settings on the CMU campus and Pittsburgh area:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 430, + 295, + 513 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "text", + "content": "- indoor: inside buildings at a slow to moderate walking pace, visiting multiple floors and areas within each." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 453, + 295, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 295, + 490 + ], + "type": "text", + "content": "- outdoor: neighborhoods ranging from single family detached to high density commercial zoning at a moderate to fast walking pace." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 295, + 513 + ], + "type": "text", + "content": "- bike: bike rides in different directions from a set starting point with a moderate biking pace." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 514, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 550 + ], + "type": "text", + "content": "Each setting features a mobile observer, with radar modulation parameters tuned for typical speeds. For sample data from each setting, see App. A.4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 560, + 138, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 138, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 138, + 574 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 580, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 296, + 628 + ], + "type": "text", + "content": "Using a transformer architecture (Sec. 4.1), we train our Generalizable Radar Transformer (GRT) for a range of different tasks (Sec. 4.2-4.3), and evaluate it on our dataset using a rigorous statistical methodology (Sec. 4.4)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 636, + 170, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 170, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 170, + 647 + ], + "type": "text", + "content": "4.1. Model Architecture" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "While many architectural refinements exist for vision transformers [30, 48], as well as for radar specifically [15, 26], we use a direct adaptation (Fig. 4) of the original Transformer [61] and Vision Transformer [12] to focus on measuring the fundamental properties of Radar transformers." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "Radar Processing From the (slow time, TX, RX, fast time) I/Q stream, we perform a 4-Dimensional FFT to obtain (range, Doppler, azimuth, elevation) dense 4D radar data cubes of size (256, 64, 8, 2), which we provide to the model as two channels consisting of the amplitude and phase angle. This data cube is patched along the range and Doppler dimensions into patches of size " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "4 \\times 2(\\times 8 \\times 2)" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ", yielding an initial set of " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "64 \\times 32 = 2048" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " patch tokens." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 175, + 555, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 259 + ], + "type": "text", + "content": "Transformer Architecture Crucially, unlike a vision transformer [12], radar models take inputs that have different spatial axes than their outputs, with vastly different relative resolutions where they overlap. As such, we use a standard transformer with a decoder [61], with varying layers and widths (Table 3); for a full specification of our transformer architecture and training procedure, see App. B.1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 266, + 555, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 362 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 362 + ], + "type": "text", + "content": "Decoder Query To handle the \"change of basis\" between the input and output space, we use an architecture based on Perceiver I/O [20]. We start by concatenating a (learned) output token to the encoder (similar to standard vision transformer [8]). The encoder output corresponding to the output token is then tiled into the desired decoder shape with a 3D sinusoidal positional encoding applied and is used as the input to the decoder, which attends to the encoder outputs." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 370, + 526, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 526, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 526, + 383 + ], + "type": "text", + "content": "4.2. Base Task: 3D Occupancy Classification" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 388, + 554, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 507 + ], + "type": "text", + "content": "An ideal foundational model training task should be easy to gather data for (e.g., using self-supervised learning) and closely aligned with a wide range of potential downstream tasks. As such, since we are primarily concerned with understanding the spatial relationship between 4D radar data and 3D space, we use 3D (polar) occupancy classification – predicting the occupancy of " + }, + { + "bbox": [ + 313, + 388, + 554, + 507 + ], + "type": "inline_equation", + "content": "64 \\times 128 \\times 64" + }, + { + "bbox": [ + 313, + 388, + 554, + 507 + ], + "type": "text", + "content": " range-azimuth elevation cells – as a base task, with Lidar as a ground truth. Our task uses a binary cross entropy objective, with some weighting to correct for cell sizes; see App. B.3 for details." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 508, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 604 + ], + "type": "text", + "content": "Notably, in addition to being fully self-supervised, this task covers all three possible output dimensions (range, azimuth, and elevation), meaning that downstream tasks such as range-azimuth classification or azimuth-elevation segmentation can be cast as 2D slices of this 3D output. This enables us to fine-tune for tasks, even if they have different spatial dimensions, simply by replacing the output head and modifying the output positional encoding queries." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 612, + 394, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 394, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 394, + 623 + ], + "type": "text", + "content": "4.3. Other Tasks" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 629, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 665 + ], + "type": "text", + "content": "In order to evaluate GRT's suitability as a foundational model for downstream fine-tuning, we use three additional tasks, each representing different output dimensions:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "- Bird's Eye View (BEV) Occupancy: Similarly to [15, 45], we classify the " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 1024" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " range-azimuth polar occupancy using Lidar as a ground truth, with the range normalized to the radar's range resolution." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24658" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 175 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 175 + ], + "type": "image", + "image_path": "a916548feedfeb7a432a77b7806970f7adf97ff3da10cb152669c32ab994ec0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 177, + 555, + 210 + ], + "lines": [ + { + "bbox": [ + 55, + 177, + 555, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 555, + 210 + ], + "type": "text", + "content": "Figure 4. The GRT architecture. 4D radar cubes are patched with a linear projection, and a sinusoidal positional encoding is added. A transformer architecture is then used, with a transformer decoder for dense outputs and a MLP decoder for Ego-Motion estimation; different output encodings are used depending on the output axes and resolution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 221, + 295, + 365 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 221, + 294, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 294, + 293 + ], + "type": "text", + "content": "- Semantic Segmentation: Similar to [24], we train our GRT to output " + }, + { + "bbox": [ + 55, + 221, + 294, + 293 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 55, + 221, + 294, + 293 + ], + "type": "text", + "content": " azimuth-elevation class labels. Since radars cannot feasibly identify many classes (e.g., poster vs. sign vs. wall) which a camera could, we use eight coarse categories: person, sky, vehicle, flat, nature, structure, ceiling, and object." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 293, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 365 + ], + "type": "text", + "content": "- Ego-Motion Estimation: Since radars can \"natively\" measure velocity², we predict the velocity of the radar relative to its current orientation. Since ego-motion estimation does not require a dense output, we replace the transformer decoder with a multi-layer-perceptron decoder head with 3 layers of 512 units." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 366, + 242, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 242, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 242, + 378 + ], + "type": "text", + "content": "For more detail about each task, see App. B.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 392, + 129, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 129, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 129, + 403 + ], + "type": "text", + "content": "4.4. Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 411, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 495 + ], + "type": "text", + "content": "Due to the cost of scaling foundational models, false positives can result in significant wasted resources, especially if associated with a costly methodological change. As such, since our dataset cannot be treated as having an \"infinite sample size\", we calculate upper-bounded uncertainty estimates wherever possible. In order to ensure these results are statistically accurate, we take the following steps:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 497, + 295, + 628 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "content": "- Geo-Split: Within each setting, " + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\approx 1.5" + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "content": " hours of data was reserved as a test set, which we ensured to be geographically disjoint from the training set to prevent leakage [28]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 592 + ], + "type": "text", + "content": "- Sample Size correction: Time series signals – e.g. radar-lidar-video tuples – cannot be viewed as independent samples; as such, the effective sample size, which we obtain from an autocorrelation-based estimate [51], must be used when calculating the standard error." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 593, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 628 + ], + "type": "text", + "content": "- Paired z-Test: Using the fact that models are evaluated on the same test traces, we use a paired z-test on the relative performance of each model with respect to a baseline." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 630, + 295, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 678 + ], + "type": "text", + "content": "We report each metric relative to its specified baseline by default, along with error bars for a two-sided " + }, + { + "bbox": [ + 55, + 630, + 295, + 678 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 630, + 295, + 678 + ], + "type": "text", + "content": " confidence interval; using our procedure and dataset, we can measure differences of " + }, + { + "bbox": [ + 55, + 630, + 295, + 678 + ], + "type": "inline_equation", + "content": "1 - 2\\%" + }, + { + "bbox": [ + 55, + 630, + 295, + 678 + ], + "type": "text", + "content": " (App. B.4)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "content": "Validation Split and Data Size Sampling. We used the last " + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "content": " of each training trace for validation (separate from the test set), with the first " + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "content": " used for training. When training on reduced dataset sizes, we use the first " + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "inline_equation", + "content": "18\\%" + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "content": " of each trace for training for " + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "content": " dataset sizes respectively; to reduce the variance of our experiments, we always " + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 221, + 555, + 305 + ], + "type": "text", + "content": " of each trace for validation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 315, + 367, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 315, + 367, + 327 + ], + "spans": [ + { + "bbox": [ + 314, + 315, + 367, + 327 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 335, + 554, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 406 + ], + "type": "text", + "content": "Using our dataset, we first ran extensive ablation (Sec. 5.1), scaling (Sec. 5.2), and fine-tuning (Sec. 5.3) experiments which show the efficacy, scalability, and generalizability of GRT. Our experiments took 874 RTX 4090-hours" + }, + { + "bbox": [ + 313, + 335, + 554, + 406 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 313, + 335, + 554, + 406 + ], + "type": "text", + "content": " of training time, with the GRT-small model taking 22 RTX 4090-hours to train." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "content": "Model Performance Despite the low resolution (only 3 " + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{TX} \\times 4" + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "content": " TX antenna) of our radar, GRT is able to predict a range of outputs with remarkable quality (Fig. 5); we show additional examples in App. C.1. We also evaluated common metrics for each objective as an absolute reference (App. C.2); GRT achieves a 3D chamfer distance of 4.9 range bins, corresponding to " + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "inline_equation", + "content": "0.66\\mathrm{m}" + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "content": " indoors, " + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "inline_equation", + "content": "1.6\\mathrm{m}" + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "content": " outdoors, and " + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 313, + 411, + 554, + 506 + ], + "type": "text", + "content": " on bike." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 513, + 415, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 513, + 415, + 524 + ], + "spans": [ + { + "bbox": [ + 314, + 513, + 415, + 524 + ], + "type": "text", + "content": "5.1. Ablation Studies" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 529, + 554, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 614 + ], + "type": "text", + "content": "Using our dataset and GRT, we performed ablation studies on several parameters that are independent of the underlying architecture and the task (Table 4). In particular, we find that several common practices – omitting Doppler information, using Angle-of-Arrival Estimates, and applying CFAR thresholding – result in degraded performance equivalent to more than a " + }, + { + "bbox": [ + 313, + 529, + 554, + 614 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 313, + 529, + 554, + 614 + ], + "type": "text", + "content": " reduction in training data." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 677 + ], + "type": "text", + "content": "Input Representation Our models use complex 4D radar cubes that losslessly capture all information measured by a radar; however, the common practice in radar models and data sets is to use processed higher-level representations. We benchmark three common approaches:" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 554, + 713 + ], + "type": "text", + "content": "3Our experiments were run on a range of different machines with varying compute capacity, which we normalize with respect to a single RTX 4090. We only tracked training and validation time, with testing excluded." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "Sensors which provide \"absolute\" pose, e.g. Lidars and Cameras, must differentiate, while IMUs must integrate." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24659" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 181, + 312 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 181, + 312 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 181, + 312 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 181, + 312 + ], + "type": "image", + "image_path": "34204fc446f05cb39fec29fd2b2438e1bfd0d5fc9ab884034ec0c67b22bae0a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 315, + 555, + 349 + ], + "lines": [ + { + "bbox": [ + 55, + 315, + 555, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 555, + 349 + ], + "type": "text", + "content": "Figure 5. Training with more data and fine tuning instead of starting from scratch lead to much higher performance. 3D occupancy maps are less noisy as seen in rendered depth images (left), semantic segmentation is cleaner (center left), bird's eye view occupancy is sharper (center right), and velocity estimation is more accurate (right). However, scaling model size (left) does not have a large impact." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 72, + 304, + 312 + ], + "blocks": [ + { + "bbox": [ + 181, + 72, + 304, + 312 + ], + "lines": [ + { + "bbox": [ + 181, + 72, + 304, + 312 + ], + "spans": [ + { + "bbox": [ + 181, + 72, + 304, + 312 + ], + "type": "image", + "image_path": "65bba12e933c69d364c63d49ce00aef092522ed72932741a9b5432da7e9275c9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 429, + 312 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 429, + 312 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 429, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 429, + 312 + ], + "type": "image", + "image_path": "3152cc6949a027a0bcfc98f66254f9a13e5bfde6306fa38ae0e1fd0f5c1602ed.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 429, + 72, + 553, + 312 + ], + "blocks": [ + { + "bbox": [ + 429, + 72, + 553, + 312 + ], + "lines": [ + { + "bbox": [ + 429, + 72, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 429, + 72, + 553, + 312 + ], + "type": "image", + "image_path": "dbf7f6ead04dc07b3af3c0890e115c36473dd1a40511afb4cf6b2a9f355fbc92.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 364, + 295, + 555 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 364, + 295, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 424 + ], + "type": "text", + "content": "- Real (amplitude-only) 4D data cubes are not significantly different from complex data, indicating that the \"leftover\" phase from a Doppler FFT carries little additional information. Since using complex data has negligible compute overhead, we default to complex inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 424, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 295, + 495 + ], + "type": "text", + "content": "- Angle of Arrival Estimates can be used to replace dense antenna measurements. Since our radar has only 2 elevation bins, we reduce the azimuth axis (8 bins) into an AoA estimate. This discards a substantial amount of information, leading to a " + }, + { + "bbox": [ + 55, + 424, + 295, + 495 + ], + "type": "inline_equation", + "content": "28.0\\%" + }, + { + "bbox": [ + 55, + 424, + 295, + 495 + ], + "type": "text", + "content": " loss increase, equivalent to more than a " + }, + { + "bbox": [ + 55, + 424, + 295, + 495 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 55, + 424, + 295, + 495 + ], + "type": "text", + "content": " decrease in the size of the dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 496, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 555 + ], + "type": "text", + "content": "- Constant False Alarm Rate (CFAR) processing removes weak or noisy reflectors based on local estimates of background noise [36]; this ablation uses a p-value threshold of 0.05, and zeros out rejected points. This leads to an even higher " + }, + { + "bbox": [ + 55, + 496, + 295, + 555 + ], + "type": "inline_equation", + "content": "31.5\\%" + }, + { + "bbox": [ + 55, + 496, + 295, + 555 + ], + "type": "text", + "content": " loss increase." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": "Impact of Doppler With limited angular resolution, GRT is highly dependent on Doppler information, which can capture higher resolution geometry [18]. We find that removing the Doppler FFT from our processing pipeline (i.e. treating each 4D radar cube as a time series of 64 3D frames [45]) leads to a " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "22.5\\%" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " loss increase. As an additional ablation, we also shuffle the slow-time axis to fully destroy any Doppler information; this does not lead to a further significant loss increase, suggesting that off-the-shelf transformers cannot easily learn FFTs. Finally, we observe worse performance at low speeds since less Doppler information is available at slow speeds (App. C.3)." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 315, + 399, + 553, + 544 + ], + "blocks": [ + { + "bbox": [ + 313, + 362, + 555, + 395 + ], + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 395 + ], + "type": "text", + "content": "Table 4. Test loss for each ablation (smaller is better) relative to GRT-small trained on our full dataset, along with " + }, + { + "bbox": [ + 313, + 362, + 555, + 395 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 362, + 555, + 395 + ], + "type": "text", + "content": " confidence intervals for the relative differences." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 399, + 553, + 544 + ], + "lines": [ + { + "bbox": [ + 315, + 399, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 315, + 399, + 553, + 544 + ], + "type": "table", + "html": "
AblationRelative Test Loss
InputsAmplitude Only+0.04 ± 0.85%
Angle of Arrival+28.0 ± 2.12%
CFAR Thresholding+31.5 ± 2.38%
DopplerWithout Doppler FFT+22.5 ± 1.76%
Slow Time Shuffled+23.10 ± 1.94%
Post-Patch AxesDoppler-Az-EI+6.22 ± 1.11%
Range-Az-EI+6.27 ± 1.10%
Range-Doppler-Az-EI+4.18 ± 1.09%
AugmentationsNone+5.87 ± 1.32%
Scale, Phase, and Flip Only+3.89 ± 1.19%
Separate ModelsIndoor Data Only+5.76 ± 1.85%
Outdoor Data Only+5.58 ± 1.26%
Bike Data Only+2.77 ± 1.41%
", + "image_path": "c55036d9e68b567c0bd17d3b0cacef008b6efc980b03d97f1df4d5530a84463f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 666 + ], + "type": "text", + "content": "Patch Axes Since 4D range-Doppler-azimuth-elevation radar data cubes have four axes with different properties, they do not have an obvious counterpart to the square patches used in Vision Transformers. Benchmarking four alternatives (App. B.1), with each resulting in 2048 total patches, we find that Range-Doppler patching where the azimuth and elevation axes are \"patched out\" (similar to [15, 49]) is the most effective, performing " + }, + { + "bbox": [ + 313, + 570, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\approx 5\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 666 + ], + "type": "text", + "content": " better." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Data Augmentations We develop a range of data aug- mentations that together provide a modest but significant performance improvement (5.87±1.32%; Table 4); we pro" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24660" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 76, + 545, + 199 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 545, + 199 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 545, + 199 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 545, + 199 + ], + "type": "image", + "image_path": "1424cedc69f52fdfe5cc11b94620d3f35952def931b5567028af627c37ea51e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 555, + 247 + ], + "type": "text", + "content": "Figure 6. Scaling laws for mmWave radar transformers across indoor, outdoor, and bike test splits. Models and confidence intervals are measured relative to the small transformer trained on " + }, + { + "bbox": [ + 55, + 201, + 555, + 247 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 201, + 555, + 247 + ], + "type": "text", + "content": " of the dataset. While our models show weak scaling over model size when trained on our dataset, we see strong log-linear scaling across dataset size of " + }, + { + "bbox": [ + 55, + 201, + 555, + 247 + ], + "type": "inline_equation", + "content": "\\approx 20\\%" + }, + { + "bbox": [ + 55, + 201, + 555, + 247 + ], + "type": "text", + "content": " loss decrease per " + }, + { + "bbox": [ + 55, + 201, + 555, + 247 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 55, + 201, + 555, + 247 + ], + "type": "text", + "content": " increase in data size. For variants of this graph with respect to absolute metrics, see App. C.2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 259, + 290, + 367 + ], + "blocks": [ + { + "bbox": [ + 62, + 259, + 290, + 367 + ], + "lines": [ + { + "bbox": [ + 62, + 259, + 290, + 367 + ], + "spans": [ + { + "bbox": [ + 62, + 259, + 290, + 367 + ], + "type": "image", + "image_path": "9f3a8a275912cf64f0adaae5fa3fdf887528a12d5ef894aaf39583bd3e4fb7f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 370, + 297, + 404 + ], + "lines": [ + { + "bbox": [ + 55, + 370, + 297, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 297, + 404 + ], + "type": "text", + "content": "Figure 7. GRT compared to Baselines, measured with respect to Chamfer distance (in range bins) and test loss across different training set sizes, averaged across our dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 420, + 259, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 259, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 259, + 433 + ], + "type": "text", + "content": "vide details about each augmentation in App. B.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 439, + 257, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 257, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 257, + 451 + ], + "type": "text", + "content": "5.2. Towards a Radar Foundational Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 457, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 296, + 578 + ], + "type": "text", + "content": "Scalability vs. Baselines Since prior work on learning for single-chip radar focuses on 2D outputs, we benchmark our transformer-based approach against two prior architectures for 2D BEV Occupancy prediction: a U-Net-based model (RadarHD [45]), and a Swin Transformer-based model (T-FFTRadNet [15]), with minor architecture modifications to conform to our data dimensions (App. B.5). We find that GRT-small outperforms both baselines at all training splits (Fig. 7), demonstrating the suitability of transformer architectures for scalability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Scaling Laws Training 20 different models for 5 different sizes (Table 3) and dataset sizes ranging from " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "10 - 100\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " of our data (Fig. 6), we observe a logarithmic improvement with data size of approximately " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " improvement per " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " increase in data, similar to early observations in computer vision [57]. This can also be seen qualitatively (Fig. 5), where models trained on more data produce much higher quality predictions. Similarly to vision transformers [57, 68], we also observe that larger models are more data efficient, although the magnitude of difference that we observe is much smaller due to our limited dataset size." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 334, + 552, + 419 + ], + "blocks": [ + { + "bbox": [ + 313, + 265, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 332 + ], + "type": "text", + "content": "Table 5. Chamfer Distance (in meters) for 2D BEV occupancy prediction on the Coloradar dataset [23] by location; the geometric mean is listed to account for the varying difficulty of each location. A fine-tuned GRT model outperforms baselines trained only on Coloradar, including a state of the art diffusion-based model [70] and a U-Net based model [45]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 334, + 552, + 419 + ], + "lines": [ + { + "bbox": [ + 316, + 334, + 552, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 334, + 552, + 419 + ], + "type": "table", + "html": "
TraceGRT (Ours)Diffusion [70]RadarHD [45]
Geometric Mean0.981.191.73
ARPG Lab0.780.961.73
EC Hallways1.041.041.69
Aspen0.610.510.91
Longboard2.635.475.40
Outdoors1.842.373.10
Edgar0.360.440.60
", + "image_path": "afad5eb06eea25731d9df42a1071b40049b575ee816fe9757ad91d941855c1da.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 445, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 555 + ], + "type": "text", + "content": "Generalizability across different settings We evaluate the ability of GRT to generalize across different settings by comparing a baseline model trained on combined indoor, outdoor, and bike data with models trained on each setting separately (Table 4). Despite the differences in these settings, the jointly trained model performs significantly better than models trained on each setting separately, confirming that data from different settings can be combined to train a single, stronger model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 563, + 523, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 523, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 523, + 575 + ], + "type": "text", + "content": "5.3. As a Base Model for Downstream Tasks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Dataset Fine-tuning We fine-tuned a small GRT model on the Coloradoar [23] dataset using a BEV Occupancy objective, and benchmarked the resulting model against two prior approaches trained only on ColoRadar, including a state-of-the-art diffusion model, using the same data splits and evaluation procedure [70]. Notably, despite using a modulation and resolution (128 range " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " 128 Doppler) which is not present in our dataset, GRT can be run without any architectural modifications, such as modifying the number of upsampling stages, as would be required for a convolutional architecture (App. B.5). After fine-tuning until vali" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24661" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 76, + 233, + 188 + ], + "blocks": [ + { + "bbox": [ + 78, + 76, + 233, + 188 + ], + "lines": [ + { + "bbox": [ + 78, + 76, + 233, + 188 + ], + "spans": [ + { + "bbox": [ + 78, + 76, + 233, + 188 + ], + "type": "image", + "image_path": "2080bec3029056116a0cdb45bc4aee236ef9ae367a01697420184a20dc477371.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 192, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 555, + 227 + ], + "type": "text", + "content": "Figure 8. Fine tuning mmWave radar transformers to different downstream tasks. Models and confidence intervals are measured relative to the small transformer trained (from scratch) on " + }, + { + "bbox": [ + 55, + 192, + 555, + 227 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 192, + 555, + 227 + ], + "type": "text", + "content": " of the dataset. Pre-training and fine tuning strongly impacts data efficiency, equivalent to up to a " + }, + { + "bbox": [ + 55, + 192, + 555, + 227 + ], + "type": "inline_equation", + "content": "5\\times" + }, + { + "bbox": [ + 55, + 192, + 555, + 227 + ], + "type": "text", + "content": " increase in dataset size (observed in the Segmentation task)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 236, + 76, + 380, + 190 + ], + "blocks": [ + { + "bbox": [ + 236, + 76, + 380, + 190 + ], + "lines": [ + { + "bbox": [ + 236, + 76, + 380, + 190 + ], + "spans": [ + { + "bbox": [ + 236, + 76, + 380, + 190 + ], + "type": "image", + "image_path": "dda890857ac31b4718f1091d4f907fa3ddf26dc18e9ec681a9f1b882ab7b8041.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 384, + 76, + 529, + 190 + ], + "blocks": [ + { + "bbox": [ + 384, + 76, + 529, + 190 + ], + "lines": [ + { + "bbox": [ + 384, + 76, + 529, + 190 + ], + "spans": [ + { + "bbox": [ + 384, + 76, + 529, + 190 + ], + "type": "image", + "image_path": "8515653f915d2417d8f4577eb33934689676576b11ab10f0a4d57d0be252bbc1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 247, + 279, + 366 + ], + "blocks": [ + { + "bbox": [ + 70, + 247, + 279, + 366 + ], + "lines": [ + { + "bbox": [ + 70, + 247, + 279, + 366 + ], + "spans": [ + { + "bbox": [ + 70, + 247, + 279, + 366 + ], + "type": "image", + "image_path": "60514d9ebc7b350aa40b2d511aa9c15631759c8c67de3aa4a45d762a78888f01.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 368, + 297, + 401 + ], + "lines": [ + { + "bbox": [ + 55, + 368, + 297, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 297, + 401 + ], + "type": "text", + "content": "Figure 9. A linear projection of the observed logarithmic scaling to a test loss lower bound suggests that logarithmic scaling cannot continue beyond " + }, + { + "bbox": [ + 55, + 368, + 297, + 401 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 55, + 368, + 297, + 401 + ], + "type": "text", + "content": " our current dataset" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 296, + 472 + ], + "type": "text", + "content": "dation loss convergence (" + }, + { + "bbox": [ + 55, + 423, + 296, + 472 + ], + "type": "inline_equation", + "content": "\\approx 30" + }, + { + "bbox": [ + 55, + 423, + 296, + 472 + ], + "type": "text", + "content": " minutes of training using a single RTX 4090), GRT achieves substantially lower Chamfer distance than baselines trained only on ColoRadar, showing the value of easily tunable foundational models (Table 5)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 483, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 651 + ], + "type": "text", + "content": "Task Fine-tuning We also fine-tuned GRT-small for each of our secondary tasks using " + }, + { + "bbox": [ + 55, + 483, + 296, + 651 + ], + "type": "inline_equation", + "content": "10 - 100\\%" + }, + { + "bbox": [ + 55, + 483, + 296, + 651 + ], + "type": "text", + "content": " of our dataset, and compared the results with models trained from scratch on the same proportions of the dataset. Following this procedure, we find substantial performance gains equivalent to up to a " + }, + { + "bbox": [ + 55, + 483, + 296, + 651 + ], + "type": "inline_equation", + "content": "5\\times" + }, + { + "bbox": [ + 55, + 483, + 296, + 651 + ], + "type": "text", + "content": " increase in dataset size compared to training from scratch (Fig. 8). This effect is especially pronounced when less data is available, with the performance benefits of fine tuning disappearing as the dataset is scaled for the BEV Occupancy objective but staying more or less constant for the Semantic Segmentation objective. We also observe this effect as a clear qualitative difference: fine-tuned models produce sharper and more accurate predictions than their counterparts trained from scratch (Fig. 5)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 237, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 237, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 237, + 672 + ], + "type": "text", + "content": "5.4. How Much More Data is Needed?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Although we cannot directly observe performance saturation, we project how much data would be required to saturate a Radar Transformer using two different methods to ar" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "text", + "content": "rive at a best guess of approximately 100M samples - " + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "inline_equation", + "content": "100\\times" + }, + { + "bbox": [ + 313, + 247, + 553, + 270 + ], + "type": "text", + "content": " our current dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 283, + 555, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 404 + ], + "type": "text", + "content": "Linear Projection of Scaling Laws To lower-bound the possible test loss in our dataset, we trained a small model on the test set to approximate convergence. Assuming that the rate of improvement in model performance with increased training data cannot decrease, we extend our observed (Fig. 6) logarithmic scaling law to this lower bound to, in turn, estimate an upper bound for when the logarithmic trend will no longer hold (Fig. 9). This yields an estimate of " + }, + { + "bbox": [ + 313, + 283, + 555, + 404 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 313, + 283, + 555, + 404 + ], + "type": "text", + "content": " our current dataset size. For additional details justifying our estimation of this bound, see App. C.4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "type": "text", + "content": "Validation Curve Trends We observe that GRTs tend to stop improving (with respect to validation loss) after " + }, + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\approx 10" + }, + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "type": "text", + "content": " training epochs, regardless of model or dataset size (App. C.4); this is similar to trends observed in the training of data-constrained LLMs, which are also observed to saturate around 10 epochs [37]. Using vision transformers, whose scaling laws are well studied [68] due to the availability of internet data, as a reference, we expect training saturation to occur around " + }, + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "type": "inline_equation", + "content": "10^{2} - 10^{4}\\mathrm{M}" + }, + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "type": "text", + "content": " samples seen. Since each epoch in our dataset corresponds to " + }, + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "type": "inline_equation", + "content": "\\approx 1\\mathrm{M}" + }, + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "type": "text", + "content": " samples seen, this implies that " + }, + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "type": "inline_equation", + "content": "1000\\times" + }, + { + "bbox": [ + 313, + 416, + 556, + 561 + ], + "type": "text", + "content": " our current dataset size is required to delay overfitting beyond this point." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 388, + 586 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we train a Generalizable Radar Transformer using a large, 29 hour (1M sample) dataset collected using our open-source data collection system and demonstrate that our Radar Transformer can generalize across datasets and settings, can be readily fine-tuned, and exhibits logarithmic scaling. While we believe that substantial gains are still possible through further data scaling, we hope that our dataset and baseline models will enable the community to revisit previous methods in new context and explore new capabilities made possible by a much larger dataset." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24662" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Karim Armanious, Maurice Quach, Michael Ulrich, Timo Winterling, Johannes Friesen, Sascha Braun, Daniel Jenet, Yuri Feldman, Eitan Kosman, Philipp Rapp, et al. Bosch street dataset: A multi-modal dataset with imaging radar for automated driving. arXiv preprint arXiv:2407.12803, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "text", + "content": "[2] Jie Bai, Lianqing Zheng, Sen Li, Bin Tan, Sihan Chen, and Libo Huang. Radar transformer: An object classification network based on 4d mmw imaging radar. Sensors, 21(11): 3854, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 257 + ], + "type": "text", + "content": "[3] Dan Barnes, Matthew Gadd, Paul Murcutt, Paul Newman, and Ingmar Posner. The oxford radar robotcar dataset: A radar extension to the oxford robotcar dataset. In 2020 IEEE international conference on robotics and automation (ICRA), pages 6433-6438. IEEE, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "text", + "content": "[4] Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 314, + 296, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 296, + 368 + ], + "type": "text", + "content": "[5] Keenan Burnett, David J Yoon, Yuchen Wu, Andrew Z Li, Haowei Zhang, Shichen Lu, Jingxing Qian, Wei-Kang Tseng, Andrew Lambert, Keith YK Leung, et al. Boreas: A multi-season autonomous driving dataset. The International Journal of Robotics Research, 42(1-2):33-42, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 369, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 434 + ], + "type": "text", + "content": "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 479 + ], + "type": "text", + "content": "[7] Han Cui, Shu Zhong, Jiacheng Wu, Zichao Shen, Naim Dahnoun, and Yiren Zhao. Milipoint: A point cloud dataset for mmwave radar. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 480, + 295, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 513 + ], + "type": "text", + "content": "[8] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "text", + "content": "[9] Colin Decourt, Rufin VanRullen, Didier Salle, and Thomas Oberlin. Darod: A deep automotive radar object detector on range-doppler maps. In 2022 IEEE Intelligent Vehicles Symposium (IV), pages 112-118. IEEE, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 602 + ], + "type": "text", + "content": "[10] Fangqiang Ding, Xiangyu Wen, Yunzhou Zhu, Yiming Li, and Chris Xiaoxuan Lu. Radar: Robust 3d occupancy prediction with 4d imaging radar. Advances in Neural Information Processing Systems, 37:101589-101617, 2025. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 603, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 295, + 657 + ], + "type": "text", + "content": "[11] Christopher Doer and Gert F Trommer. Yaw aided radar inertial odometry using manhattan world assumptions. In 2021 28th Saint Petersburg International Conference on Integrated Navigation Systems (ICINS), pages 1-9. IEEE, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "scale. In International Conference on Learning Representations, 2021. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[13] Xiangyu Gao, Guanbin Xing, Sumit Roy, and Hui Liu. Ramp-cnn: A novel neural network for enhanced automotive radar object recognition. IEEE Sensors Journal, 21(4): 5119-5132, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 185 + ], + "type": "text", + "content": "[14] Xiangyu Gao, Sumit Roy, and Guanbin Xing. Mimo-sar: A hierarchical high-resolution imaging algorithm for mmwave fmew radar in autonomous driving. IEEE Transactions on Vehicular Technology, 70(8):7322-7334, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "text", + "content": "[15] James Giroux, Martin Bouchard, and Robert Laganiere. Tfftradnet: Object detection with swin vision transformers from raw adc radar signals. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4030-4039, 2023. 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 553, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 264 + ], + "type": "text", + "content": "[16] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 308 + ], + "type": "text", + "content": "[17] Wolfgang Hess, Damon Kohler, Holger Rapp, and Daniel Andor. Real-time loop closure in 2d lidar slam. In 2016 IEEE international conference on robotics and automation (ICRA), pages 1271-1278. IEEE, 2016. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 310, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 374 + ], + "type": "text", + "content": "[18] Tianshu Huang, John Miller, Akarsh Prabhakara, Tao Jin, Tarana Laroia, Zico Kolter, and Anthony Rowe. Dart: Implicit doppler tomography for radar novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24118-24129, 2024. 2, 3, 6, 1, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 408 + ], + "type": "text", + "content": "[19] Cesar Iovescu and Sandeep Rao. The fundamentals of millimeter wave sensors. *Texas Instruments*, pages 1–8, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 465 + ], + "type": "text", + "content": "[20] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. arXiv preprint arXiv:2107.14795, 2021. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 467, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 521 + ], + "type": "text", + "content": "[21] Yi Jin, Marcel Hoffmann, Anastasios Deligiannis, JuanCarlos Fuentes-Michel, and Martin Vossiek. Semantic segmentation-based occupancy grid map learning with automotive radar raw data. IEEE Transactions on Intelligent Vehicles, 9(1):216-230, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "text", + "content": "[22] Nidhi Kalra and Susan M Paddock. Driving to safety: How many miles of driving would it take to demonstrate autonomous vehicle reliability? Transportation Research Part A: Policy and Practice, 94:182-193, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "type": "text", + "content": "[23] Andrew Kramer, Kyle Harlow, Christopher Williams, and Christoffer Heckman. Coloradar: The direct 3d millimeter wave radar dataset. The International Journal of Robotics Research, 41(4):351-360, 2022. 2, 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 612, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 612, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 612, + 553, + 656 + ], + "type": "text", + "content": "[24] Haowen Lai, Gaoxiang Luo, Yifei Liu, and Mingmin Zhao. Enabling visual recognition at radio frequency. In Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, pages 388-403, 2024. 2, 3, 5, 9" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[25] Xinrong Li, Xiaodong Wang, Qing Yang, and Song Fu. Signal processing for tdm mimo fmcw millimeter-wave radar sensors. IEEE Access, 9:167959-167971, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[26] Yu-Jhe Li, Shawn Hunt, Jinhyung Park, Matthew O'Toole, and Kris Kitani. Azimuth super-resolution for fmcw radar in" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24663" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 106 + ], + "type": "text", + "content": "autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17504-17513, 2023. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 162 + ], + "type": "text", + "content": "[27] Jaime Lien, Nicholas Gillian, M Emre Karagozler, Patrick Amihood, Carsten Schwesig, Erik Olson, Hakim Raja, and Ivan Poupyrev. Soli: Ubiquitous gesture sensing with millimeter wave radar. ACM Transactions on Graphics (TOG), 35(4):1-19, 2016. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 217 + ], + "type": "text", + "content": "[28] Adam Lilja, Junsheng Fu, Erik Stenberg, and Lars Hammarstrand. Localization is all you evaluate: Data leakage in online mapping datasets and how to fix it. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22150-22159, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 271 + ], + "type": "text", + "content": "[29] Teck-Yian Lim, Spencer A Markowitz, and Minh N Do. Radical: A synchronized fmcw radar, depth, imu and rgb camera data dataset with low-level fmcw radar signals. IEEE Journal of Selected Topics in Signal Processing, 15(4):941-953, 2021. 3, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 296, + 327 + ], + "type": "text", + "content": "[30] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 296, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 358 + ], + "type": "text", + "content": "[31] Ilya Loshchilov, Frank Hutter, et al. Fixing weight decay regularization in adam. arXiv preprint arXiv:1711.05101, 5, 2017.6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 296, + 435 + ], + "type": "text", + "content": "[32] Chris Xiaoxuan Lu, Muhamad Risqi U Saputra, Peijun Zhao, Yasin Almalioglu, Pedro PB De Gusmao, Changhao Chen, Ke Sun, Niki Trigoni, and Andrew Markham. milliego: single-chip mmwave radar aided egomotion estimation via deep sensor fusion. In Proceedings of the 18th Conference on Embedded Networked Sensor Systems, pages 109-122, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 491 + ], + "type": "text", + "content": "[33] Sohrab Madani, Jayden Guan, Waleed Ahmed, Saurabh Gupta, and Haitham Hassanieh. Radatron: Accurate detection using multi-resolution cascaded mimo radar. In European Conference on Computer Vision, pages 160-178. Springer, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 569 + ], + "type": "text", + "content": "[34] Bence Major, Daniel Fontijne, Amin Ansari, Ravi Teja Sukhavasi, Radhika Gowaikar, Michael Hamilton, Sean Lee, Slawomir Grzechnik, and Sundar Subramanian. Vehicle detection with automotive radar using deep learning on range-azimuth-doppler tensors. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "type": "text", + "content": "[35] Babak Mamandipoor, Greg Malysa, Amin Arababian, Upamanyu Madhow, and Karam Noujeim. 60 ghz synthetic aperture radar for short-range imaging: Theory and experiments. In 2014 48th Asilomar Conference on Signals, Systems and Computers, pages 553-558. IEEE, 2014. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 625, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 658 + ], + "type": "text", + "content": "[36] Gary Minkler and Jing Minkler. Cfar: the principles of automatic radar detection in clutter. *Nasa sti/recon technical report a*, 90:23371, 1990. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 712 + ], + "type": "text", + "content": "[37] Niklas Muennighoff, Alexander Rush, Boaz Barak, Teven Le Scao, Nouamane Tazi, Aleksandra Piktus, Sampo Pyysalo, Thomas Wolf, and Colin A Raffel. Scaling data-constrained language models. Advances in Neural Information Processing Systems, 36:50358-50376, 2023. 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 139 + ], + "type": "text", + "content": "[38] Farzan Erlik Nowruzi, Dhanvin Kolhatkar, Prince Kapoor, Fahed Al Hassanat, Elnaz Jahani Heravi, Robert Laganiere, Julien Rebut, and Waqas Malik. Deep open space segmentation using automotive radar. In 2020 IEEE MTT-S International Conference on Microwaves for Intelligent Mobility (ICMIM), pages 1-4. IEEE, 2020. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 140, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 171 + ], + "type": "text", + "content": "[39] Itai Orr, Moshik Cohen, and Zeev Zalevsky. High-resolution radar road segmentation using weakly supervised learning. Nature Machine Intelligence, 3(3):239-246, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 172, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 554, + 217 + ], + "type": "text", + "content": "[40] Arthur Ouaknine, Alasdair Newson, Patrick Pérez, Florence Tupin, and Julien Rebut. Multi-view radar semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15671-15680, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 217, + 554, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 271 + ], + "type": "text", + "content": "[41] Arthur Ouaknine, Alasdair Newson, Julien Rebut, Florence Tupin, and Patrick Pérez. Carrada dataset: Camera and automotive radar with range-angle-doppler annotations. In 2020 25th International Conference on Pattern Recognition (ICPR), pages 5068-5075. IEEE, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 554, + 316 + ], + "type": "text", + "content": "[42] Dong-Hee Paek, Seung-Hyun Kong, and Kevin Tirta Wijaya. K-radar: 4d radar object detection for autonomous driving in various weather conditions. Advances in Neural Information Processing Systems, 35:3819-3829, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 316, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 554, + 360 + ], + "type": "text", + "content": "[43] Andras Palffy, Jiaao Dong, Julian FP Kooij, and Dariu M Gavrila. Cnn based road user detection using the 3d radar cube. IEEE Robotics and Automation Letters, 5(2):1263-1270, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 361, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 554, + 426 + ], + "type": "text", + "content": "[44] Akarsh Prabhakara, Diana Zhang, Chao Li, Sirajum Munir, Aswin C Sankaranarayanan, Anthony Rowe, and Swarun Kumar. Exploring mmwave radar and camera fusion for high-resolution and long-range depth imaging. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3995-4002. IEEE, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 426, + 554, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 554, + 491 + ], + "type": "text", + "content": "[45] Akarsh Prabhakara, Tao Jin, Arnav Das, Gantavya Bhatt, Lilly Kumari, Elahe Soltanaghai, Jeff Bilmes, Swarun Kumar, and Anthony Rowe. High resolution point clouds from mmwave radar. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4135-4142. IEEE, 2023. 2, 3, 4, 6, 7, 1, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 492, + 554, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 554, + 547 + ], + "type": "text", + "content": "[46] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 548, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 554, + 613 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 13" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 613, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 658 + ], + "type": "text", + "content": "[48] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 712 + ], + "type": "text", + "content": "[49] Julien Rebut, Arthur Ouaknine, Waqas Malik, and Patrick Pérez. Raw high-definition radar for multi-task learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17021-17030, 2022. 3, 6" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24664" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 95 + ], + "type": "text", + "content": "[50] Mark A Richards et al. Fundamentals of radar signal processing. McGraw-hill New York, 2005. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 294, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 294, + 118 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 294, + 118 + ], + "type": "text", + "content": "[51] Christian P Robert, George Casella, and George Casella. Monte Carlo statistical methods. Springer, 1999. 5, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 121, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 294, + 152 + ], + "type": "text", + "content": "[52] Hermann Rohling. Radar cfar thresholding in clutter and multiple target situations. IEEE Transactions on Aerospace and Electronic Systems, AES-19(4):608-621, 1983. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 156, + 294, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 294, + 209 + ], + "type": "text", + "content": "[53] Mark E Russell, Arthur Crain, Anthony Curran, Richard A Campbell, Clifford A Drubin, and Willian F Miccioli. Millimeter-wave radar sensor for automotive intelligent cruise control (icc). IEEE Transactions on microwave theory and techniques, 45(12):2444-2453, 1997. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 212, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 294, + 266 + ], + "type": "text", + "content": "[54] Marcel Sheeny, Emanuele De Pellegrin, Saptarshi Mukherjee, Alireza Ahrabian, Sen Wang, and Andrew Wallace. Radiate: A radar dataset for automotive perception in bad weather. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 1-7. IEEE, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 269, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 269, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 269, + 294, + 334 + ], + "type": "text", + "content": "[55] Xian Shuai, Yulin Shen, Yi Tang, Shuyao Shi, Luping Ji, and Guoliang Xing. millieye: A lightweight mmwave radar and camera fusion system for robust object detection. In Proceedings of the International Conference on Internet-of-Things Design and Implementation, pages 145–157, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 336, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 336, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 336, + 294, + 392 + ], + "type": "text", + "content": "[56] Akash Deep Singh, Sandeep Singh Sandha, Luis Garcia, and Mani Srivastava. Radhar: Human activity recognition from point clouds generated through a millimeter-wave radar. In Proceedings of the 3rd ACM Workshop on Millimeter-wave Networks and Sensing Systems, pages 51-56, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[57] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 439, + 294, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 439, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 439, + 294, + 471 + ], + "type": "text", + "content": "[58] Texas Instruments. Automated doors reference design using ti mmwave sensors. Design Guide TIDEP-01018, Texas Instruments, Dallas, TX, 2018. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 474, + 294, + 505 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 294, + 505 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 294, + 505 + ], + "type": "text", + "content": "[59] AWR1843AOP Single-chip 77- and 79-GHz FMCW mmWave Sensor Antennas-OnPackage (AOP). Texas Instruments, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 508, + 294, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 508, + 294, + 573 + ], + "spans": [ + { + "bbox": [ + 56, + 508, + 294, + 573 + ], + "type": "text", + "content": "[60] Claudia Vasanelli, Fabian Roos, Andre Durr, Johannes Schlichenmaier, Philipp Hugler, Benedikt Meinecke, Maximilian Steiner, and Christian Waldschmidt. Calibration and direction-of-arrival estimation of millimeter-wave radars: A practical introduction. IEEE Antennas and Propagation Magazine, 62(6):34-45, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 576, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 294, + 597 + ], + "type": "text", + "content": "[61] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 600, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 632 + ], + "type": "text", + "content": "[62] Christian Waldschmidt, Juergen Hasch, and Wolfgang Menzel. Automotive radar—from first efforts to future systems. IEEE Journal of Microwaves, 1(1):135-148, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[63] Yizhou Wang, Zhongyu Jiang, Xiangyu Gao, Jenq-Neng Hwang, Guanbin Xing, and Hui Liu. Rodnet: Radar object detection using cross-modal supervision. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 504-513, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[64] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 440 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[65] Ruibin Xiong, Yunchang Yang, Di He, Kai Zheng, Shuxin Zheng, Chen Xing, Huishuai Zhang, Yanyan Lan, Liwei Wang, and Tieyan Liu. On layer normalization in the transformer architecture. In International Conference on Machine Learning, pages 10524-10533. PMLR, 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[66] Hiroyoshi Yamada, Takumi Kobayashi, Yoshio Yamaguchi, and Yuuichi Sugiyama. High-resolution 2d sar imaging by the millimeter-wave automobile radar. In 2017 IEEE Conference on Antenna Measurements & Applications (CAMA), pages 149-150. IEEE, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "text", + "content": "[67] Muhammet Emin Yanik and Murat Torlak. Near-field mimosar millimeter-wave imaging with sparsely sampled aperture data. IEEE Access, 7:31801-31819, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 253, + 555, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 296 + ], + "type": "text", + "content": "[68] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12104-12113, 2022. 3, 7, 8, 13" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "text", + "content": "[69] Ao Zhang, Farzan Erlik Nowruzi, and Robert Laganiere. Raddet: Range-azimuth-doppler based radar object detection for dynamic road users. In 2021 18th Conference on Robots and Vision (CRV), pages 95–102. IEEE, 2021. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "text", + "content": "[70] Ruibin Zhang, Donglai Xue, Yuhan Wang, Ruixu Geng, and Fei Gao. Towards dense and accurate radar perception via efficient cross-modal diffusion model. IEEE Robotics and Automation Letters, 2024. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 387, + 553, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 440 + ], + "type": "text", + "content": "[71] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24665" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/5863579c-c0a2-49e2-85ef-c6b4d99e52c6_content_list.json b/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/5863579c-c0a2-49e2-85ef-c6b4d99e52c6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cb04ede9368eb5983be771fa80269195e6cd2a43 --- /dev/null +++ b/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/5863579c-c0a2-49e2-85ef-c6b4d99e52c6_content_list.json @@ -0,0 +1,1460 @@ +[ + { + "type": "text", + "text": "Towards Higher Effective Rank in Parameter-efficient Fine-tuning using Khatri-Rao Product", + "text_level": 1, + "bbox": [ + 163, + 128, + 833, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Paul Albert $^{1*}$ Frederic Z. Zhang $^{1*}$ Hemanth Saratchandran $^{2}$ Anton van den Hengel $^{2}$ Ehsan Abbasnejad $^{3}$", + "bbox": [ + 250, + 202, + 746, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Amazon Machine Learning, Australia $^{2}$ Australian Institute for Machine Learning $^{3}$ Monash University {albrtpa, fredzz}@amazon.com https://github.com/PaulAlbert31/KRAdapter", + "bbox": [ + 112, + 239, + 883, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Parameter-efficient fine-tuning (PEFT) has become a standard approach for adapting large pre-trained models. Amongst PEFT methods, low-rank adaptation (LoRA) has achieved notable success. However, recent studies have highlighted its limitations compared against full-rank alternatives, particularly when applied to multimodal and large language models. In this work, we present a quantitative comparison amongst full-rank and low-rank PEFT methods using a synthetic matrix approximation benchmark with controlled spectral properties. Our results confirm that LoRA struggles to approximate matrices with relatively flat spectrums or high frequency components—signs of high effective ranks. To this end, we introduce KRAdapter, a novel PEFT algorithm that leverages the Khatri-Rao product to produce weight updates, which, by construction, tends to produce matrix product with a high effective rank. We demonstrate performance gains with KRAdapter on vision-language models up to 1B parameters and on large language models up to 8B parameters, particularly on unseen common-sense reasoning tasks. In addition, KRAdapter maintains the memory and compute efficiency of LoRA, making it a practical and robust alternative to fine-tune billion-scale parameter models.", + "bbox": [ + 89, + 347, + 485, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 733, + 220, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large pre-trained models have lead to great success in both computer vision and natural language processing [1, 13, 14, 37, 44]. However, fine-tuning these models for specific downstream tasks often demands substantial computational resources due to the sheer number of parameters. Recently, parameter-efficient fine-tuning (PEFT) has emerged as an effective strategy to address this challenge, enabling effi", + "bbox": [ + 89, + 765, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cient adaptation using a fraction of the trainable parameters [18]. In particular, low-rank adaptation (LoRA) [24], a prominent PEFT technique, formulates the learnable weight updates as the product of two low-rank matrices, achieving remarkable memory efficiency with little performance drop. Despite its success, recent studies [2, 26] suggest that the low-rank formulation can limit its ability to effectively fine-tune models for complex tasks. To shed light on this limitation, we compare LoRA to several full-rank PEFT methods by approximating weight matrices with controlled spectral properties. As shown in Figure 1, full-rank methods can outperform LoRA in approximating both synthetic and real weight matrices. Not all full-rank PEFT methods are equally effective—some produce approximations with rapidly decaying singular values, leading to a low effective rank [47]. Building on prior work [7, 50], which links substantial tail singular values to improved generalization and robustness, we hypothesize that a high effective rank is crucial to the performance of full-rank PEFT methods.", + "bbox": [ + 511, + 310, + 908, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we introduce KRAdapter, a novel PEFT approach that leverages the Khatri-Rao matrix product. By design, KRAdapter constructs provably full-rank weight matrices and empirically maintains large tail singular values. The increased effective rank indicates that the weight updates span the parameter space more uniformly, allowing the model to capture complex patterns and generalize better to distribution shifts. Across diverse architectures, including vision-language models and large language models, KRAdapter achieves performance gains while maintaining parameter efficiency. Furthermore, KRAdapter exhibits superior out-of-distribution (OOD) performance compared to LoRA and other full-rank methods. Our key contributions are: (1) a controlled matrix approximation benchmark comparing the representational advantages of full- and rank-constrained PEFT algorithms; (2) the identification of effective rank as a critical factor differentiating full-rank PEFT techniques; (3) KRAdapter, a novel Khatri-Rao product-based PEFT method that consistently achieves high effective", + "bbox": [ + 511, + 603, + 910, + 893 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done at the Australian Institute for Machine Learning.", + "bbox": [ + 107, + 886, + 433, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1292", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9f5d53b3f50b7938274151752709074ce2d67af77e39f571cf30f1a0a4834227.jpg", + "image_caption": [ + "Figure 1. A visualization of the capacity of various PEFT methods to approximate the spectrum of different types of weight matrices. The first row shows the structure of the matrices. In the second row we plot the squared nuclear error as a percentage of that of the LoRA solution, and add the absolute value above each bar. Lower is better. The third row shows the singular value distribution of the solutions compared to the target (square root scale). A flatter distribution indicates a higher effective rank. We compare LoRA [24], SinLoRA [26], Krona [16], RandLoRA [2] and our proposed algorithm KRAdapter. KRAdapter achieves a better approximation to the true matrix in every case except when we force it to have low rank (and thus to match the LoRA model)." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rank; (4) empirical validation of its superior performance and out-of-distribution robustness compared to LoRA and other PEFT methods; and (5) evidence of KRAAdapter's improved parameter scaling efficiency among full-rank approaches.", + "bbox": [ + 88, + 482, + 486, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 564, + 232, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The development of large pre-trained models has revolutionized various fields, yet their deployment and fine-tuning pose significant computational challenges [1, 13, 14, 37, 44]. Parameter-Efficient Fine-Tuning (PEFT) methods have emerged as a crucial paradigm to address these challenges, allowing adaptation of these massive models to downstream tasks while training only a small fraction of the parameters [18].", + "bbox": [ + 88, + 597, + 486, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Low-Rank Adaptation (LoRA)", + "text_level": 1, + "bbox": [ + 89, + 734, + 364, + 752 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LoRA [24] has become a cornerstone of PEFT due to its simplicity and effectiveness. LoRA introduces low-rank matrices to approximate the weight updates, significantly reducing the number of trainable parameters. Its success has led to numerous extensions and applications across various modalities [55, 60, 63, 65]. However, recent studies have begun to explore the limitations of the low-rank constraint, suggesting that it might hinder the model's ability to capture complex relationships, especially for demanding tasks", + "bbox": [ + 88, + 763, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[2, 16, 26]. Furthermore, the spectral properties of weight matrices, which influence a network's capacity for learning and generalization [41, 45, 59], may be constrained by the low-rank nature of LoRA updates. Our work builds upon these observations, confirming the need for methods that can capture richer feature interactions while maintaining parameter efficiency, potentially through weight updates with more desirable spectral characteristics.", + "bbox": [ + 511, + 482, + 908, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Enhancing the training efficiency", + "text_level": 1, + "bbox": [ + 511, + 625, + 803, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To improve convergence speed and overall performance of LoRA, several improvements have been proposed. LoRA+ [19] enhances the training process by applying different learning rates to the low-rank matrices, allowing for more nuanced optimization. DoRA [36] decomposes the low-rank updates into magnitude and direction components, providing finer-grained control over the adaptation process. HydraLoRA [52] employs multiple upscaling matrices, acting as specialized \"experts\" focusing on distinct aspects of the input, potentially capturing more diverse feature interactions.", + "bbox": [ + 511, + 652, + 910, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Further reducing parameters", + "text_level": 1, + "bbox": [ + 511, + 840, + 774, + 858 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While LoRA is inherently parameter-efficient, subsequent research has focused on pushing these limits further. Strategies", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1293", + "bbox": [ + 483, + 944, + 514, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to minimize the trainable parameter count beyond the rank-one case include methods leveraging linear combinations of fixed random matrices, such as VeRA [29] and NoLA [28], where learned one-dimensional vectors combine pre-defined non-trainable random matrices. Another line of work explores initializing or fixing LoRA's projection matrices using insights from the pre-trained weights. Methods like SVFT [35] and Pissa [39] fix or respectively initialize the projection matrices to the first eigenvectors corresponding to the largest singular values of the original weight matrix. Conversely, MiLoRA [57] utilizes the last eigenvectors.", + "bbox": [ + 89, + 90, + 486, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Full-rank updates", + "text_level": 1, + "bbox": [ + 89, + 273, + 266, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the representational bottlenecks and potential spectral limitations inherent in low-rank approximations, several methods explore parameter-efficient ways to perform full-rank updates. The goal is to capture more complex relationships present in the data without incurring the full cost of fine-tuning all parameters. Kronecker Adapters (KronA) [16] utilize the Kronecker product to construct full-rank update matrices from smaller trainable factors. SinLoRA [26] applies a sine function parameterized by a frequency parameter on top of a low-rank update, effectively producing a full-rank update. RandLoRA [2] employs parameter-efficient random matrix combinations to generate full-rank matrices. While these methods theoretically produce full rank updates, we observe their effective rank [47] is usually low (i.e. the tail singular values are very small), suggesting they might not fully capture the desired spectral properties.", + "bbox": [ + 89, + 303, + 486, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.5. Motivations", + "text_level": 1, + "bbox": [ + 89, + 561, + 218, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our analysis of existing work indicates that the low-rank update in LoRA can limit performance on challenging tasks and with complex architectures. While recent advancements aim to overcome this by generating theoretically full-rank updates via random basis combinations, Kronecker products, or sine activations, we question whether this theoretical rank consistently translates to high effective rank. We hypothesize that the Khatri-Rao product offers an alternative method for constructing full-rank update matrices that demonstrably achieve higher effective rank. This increased effective rank would enable to learn more complex feature representations, leading to improved performance across diverse tasks and architectures.", + "bbox": [ + 89, + 590, + 483, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Khatri-Rao Adapters (KRAdapter)", + "text_level": 1, + "bbox": [ + 89, + 806, + 411, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section details the formulation of the proposed Khatri-Rao Adapters (KRAdapter) for parameter-efficient finetuning. We begin by briefly discussing Low-Rank Adaptation (LoRA) before introducing the KRAdapter mechanism", + "bbox": [ + 89, + 839, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and analyzing its parameter efficiency.", + "bbox": [ + 513, + 90, + 767, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 123, + 653, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a pre-trained linear layer in a deep neural network, characterized by a weight matrix $\\mathbf{W}_0\\in \\mathbb{R}^{d_{out}\\times d_{in}}$ , where $d_{in}$ is the input dimension and $d_{out}$ is the output dimension. During standard full fine-tuning, the weight matrix $\\mathbf{W}_0$ is updated directly by gradient descent. Given an input $\\mathbf{x}\\in$ $\\mathbb{R}^{d_{in}}$ , the output $\\mathbf{h}$ is computed as:", + "bbox": [ + 511, + 152, + 906, + 243 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} = \\mathbf {W} _ {0} \\mathbf {x} + \\mathbf {b}, \\tag {3.1}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 266, + 906, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{b} \\in \\mathbb{R}^{d_{out}}$ is the bias term (for simplicity, we omit the bias term in subsequent derivations but it can be easily incorporated).", + "bbox": [ + 511, + 295, + 906, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Low-Rank Adaptation (LoRA) [24] addresses the inefficiency of full fine-tuning by freezing the pre-trained weights $\\mathbf{W}_0$ and introducing a low-rank update. Specifically, a low-rank matrix $\\Delta \\mathbf{W} = \\mathbf{BA}$ is added to the original weights, where $\\mathbf{A} \\in \\mathbb{R}^{r \\times d_{in}}$ and $\\mathbf{B} \\in \\mathbb{R}^{d_{out} \\times r}$ , with $r \\ll \\min(d_{in}, d_{out})$ being the rank. During fine-tuning, only the parameters of $\\mathbf{A}$ and $\\mathbf{B}$ are updated. The output of the LoRA-adapted layer is:", + "bbox": [ + 511, + 348, + 908, + 469 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} = \\left(\\mathbf {W} _ {0} + \\alpha \\mathbf {B A}\\right) \\mathbf {x}, \\tag {3.2}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 492, + 906, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ is a scaling factor that helps in stabilizing training and controlling the magnitude of the adapter.", + "bbox": [ + 511, + 515, + 905, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Method formulation", + "text_level": 1, + "bbox": [ + 513, + 561, + 705, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to LoRAs, our proposed method, Khatri-Rao Adapters (KRAdapter), achieves parameter efficiency without imposing the low-rank constraint. This is achieved by constructing $\\Delta \\mathbf{W}$ using the Khatri-Rao product, otherwise known as a column-wise Kronecker product.", + "bbox": [ + 511, + 590, + 906, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 3.1 (Khatri-Rao Product): Given two matrices $\\mathbf{U} \\in \\mathbb{R}^{a \\times c}$ and $\\mathbf{V} \\in \\mathbb{R}^{b \\times c}$ , their Khatri-Rao product, denoted by $\\mathbf{U} \\odot \\mathbf{V}$ , is a matrix of defined as:", + "bbox": [ + 511, + 674, + 908, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {U} \\odot \\mathbf {V} = \\left[ \\begin{array}{c} u _ {1 1} \\mathbf {v} _ {1}, u _ {1 2} \\mathbf {v} _ {2}, \\dots , u _ {1 c} \\mathbf {v} _ {c} \\\\ \\dots \\\\ u _ {a 1} \\mathbf {v} _ {1}, u _ {a 2} \\mathbf {v} _ {2}, \\dots , u _ {a c} \\mathbf {v} _ {c} \\end{array} \\right] \\in \\mathbb {R} ^ {a b \\times c}, \\tag {3.3}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 729, + 906, + 780 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $u_{ij}$ is the entry in the $i$ -th row and $j$ -th column of $\\mathbf{U}$ and $\\mathbf{v}_j \\in \\mathbb{R}^b$ is the $j$ -th column of $\\mathbf{V}$ .", + "bbox": [ + 511, + 789, + 905, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For a weight update $\\Delta \\mathbf{W}$ of size $d_{out} \\times d_{in}$ , define $\\mathbf{U} \\in \\mathbb{R}^{k_1 \\times d_{in}}$ and $\\mathbf{V} \\in \\mathbb{R}^{k_2 \\times d_{in}}$ , where $k_1 k_2 = d_{out}$ . The forward pass of a linear layer with KRAdapter is as follows", + "bbox": [ + 511, + 827, + 906, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} = \\left(\\mathbf {W} _ {0} + \\alpha \\mathbf {U} \\odot \\mathbf {V}\\right) \\mathbf {x}, \\tag {3.4}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 885, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1294", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ is a scaling factor similar to LoRA that we typically set to 0.1. Prior to training, $\\mathbf{U}$ is initialized as zeros and $\\mathbf{V}$ using a Kaiming uniform initialization [20] with a negative slope coefficient of $\\sqrt{1 / k_1}$ which we empirically find to be the optimal for convergence.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Parameter efficiency", + "text_level": 1, + "bbox": [ + 89, + 181, + 284, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The number of trainable parameters introduced by KRAdapter is determined by the dimensions of the matrices $\\mathbf{U}$ and $\\mathbf{V}$ , and the choice of $k_{1}$ (or $k_{2}$ resp.). The total number of trainable parameters for KRAdapter is:", + "bbox": [ + 89, + 210, + 483, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nN = d _ {i n} \\left(k _ {1} + k _ {2}\\right). \\tag {3.5}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 281, + 483, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$N$ is minimized when $k_{1} = k_{2} = \\sqrt{d_{out}}$ , which we adopt as the default configuration. Details are deferred to Appendix A.1. When $d_{out}$ is not a square number, we set $k_{1} = \\lfloor \\sqrt{d_{out}} \\rfloor$ and $k_{2} = \\lceil \\frac{d_{out}}{k_{1}} \\rceil$ such that $k_{1}k_{2} \\geq d_{out}$ . The resultant product $\\mathbf{U} \\odot \\mathbf{V}$ is then truncated to size $d_{out} \\times d_{in}$ .", + "bbox": [ + 89, + 305, + 483, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "KRAdapter achieves the same parameter reduction as LoRA of rank between 16 and 32 depending on matrix shapes, a common setting [24, 36, 64] without incurring any significant increase in computational cost.", + "bbox": [ + 89, + 388, + 483, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Full-rank guarantee", + "text_level": 1, + "bbox": [ + 89, + 464, + 282, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section establishes a full rank guarantee for the Khatri-Rao product when its constituent matrices are randomly initialized. This full rank property, as we will show, is crucial to the improvement of robustness against distribution shifts.", + "bbox": [ + 89, + 493, + 483, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For brevity of exposition, let us assume $k_{1} = k_{2} = k$ , i.e., $\\mathbf{U} \\in \\mathbb{R}^{k \\times d_{in}}$ and $\\mathbf{V} \\in \\mathbb{R}^{k \\times d_{in}}$ . We assume that each matrix has rank $k$ . Specifically, we consider the case where the entries of $\\mathbf{U}$ and $\\mathbf{V}$ are independently and identically distributed (i.i.d.) random variables drawn from either a standard Gaussian distribution $\\mathcal{N}(0, 1)$ or a uniform distribution on $[-\\delta, \\delta]$ for some $\\delta > 0$ .", + "bbox": [ + 89, + 561, + 483, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 3.1. Let $\\mathbf{U} \\in \\mathbb{R}^{k \\times d_{in}}$ and $\\mathbf{V} \\in \\mathbb{R}^{k \\times d_{in}}$ where $k \\leq d_{in} \\leq k^2$ be matrices whose entries are chosen i.i.d. from a standard Gaussian or uniform distribution. Then, the Khatri-Rao product $\\mathbf{U} \\odot \\mathbf{V}$ has full column rank almost surely, i.e.,", + "bbox": [ + 88, + 672, + 483, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr a n k (\\mathbf {U} \\odot \\mathbf {V}) = d _ {i n}.\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 750, + 357, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The detailed proof of Theorem 3.1 is provided in Appendix A.2. Crucially, this theorem highlights a significant distinction from low-rank adaptation (LoRA) techniques. When employing LoRA updates to matrices $\\mathbf{U}$ and $\\mathbf{V}$ , the resulting matrix, even after combination, can have a maximum rank of at most $k \\leq d_{in}$ , thus inherently producing rank-deficient matrices. In contrast, Theorem 3.1 demonstrates that the Khatri-Rao product of the same randomly", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "initialized matrices $\\mathbf{U}$ and $\\mathbf{V}$ almost surely achieves full column rank, thereby providing a mechanism to construct high-rank matrices in a parameter-efficient manner.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Differences with Kronecker Adapters", + "text_level": 1, + "bbox": [ + 511, + 154, + 836, + 171 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the Khatri-Rao product can be seen as column-wise Kronecker product, our proposed KRAdapter and Kronecker Adapters (KronA) [16] bear some similarity. However, due to the different constructions, the resulting properties differ significantly. Specifically, we find that the Khatri-Rao product leads to different spectral properties in the resultant matrix, which displays significantly higher tail singular values. We find this to be a desirable property for PEFT methods as it 056 increases the representation power when modelling various types of weight matrices (Figure 1) and also improves OOD generalization and robustness in language and vision tasks (Table 2 and 3 3) Further discussion supported by empirical differences is available in Appendix B.", + "bbox": [ + 511, + 184, + 906, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 402, + 645, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Matrix approximation", + "text_level": 1, + "bbox": [ + 511, + 434, + 720, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To isolate and understand the inherent strengths and weaknesses of KRAAdapter compared to LoRA and other full-rank PEFT methods, we design a series of controlled experiments using synthetic weight matrices with distinct structural properties.", + "bbox": [ + 511, + 463, + 906, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We design six matrix patterns to investigate parameter-efficient fine-tuning algorithms. The normally distributed random matrix acts as a high-rank baseline, evaluating general approximation performance. The random $90\\%$ sparse matrix simulates scenarios where critical pre-trained parameters should remain unchanged. The PCA-whitened random matrix assesses algorithms' ability to handle highly decorrelated representations. The low-rank random matrix, constructed by keeping only the top quartile of singular values, tests full-rank algorithms' capacity to model low-rank structures. The CLIP ImageNet fine-tuned weight delta (Vision or Language), obtained by the element-wise difference between the pre-trained CLIP-ViT-L/14 weights and the weights obtained after standard fine-tuning on ImageNet (also known as task vector [63]), represents a realistic target weight for LoRA in transformer in real-world fine-tuning. The High/low frequency features, generated using superposed sinusoidal functions with varying frequencies ([1000, 10000] Hz and [1, 100] Hz respectively), assess algorithms' bias towards feature frequencies.", + "bbox": [ + 511, + 546, + 908, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Methodology For each target matrix pattern, we train the PEFT algorithms to minimize the mean squared error (MSE) between the estimated matrix and the target matrix. We train", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1295", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "on $1024 \\times 768$ matrices which is the dimension of the query part of the attention matrix in ViT-L/14. All algorithms are trained with the same or higher number of trainable parameters than KRAdapter to ensure a fair comparison. We report the absolute nuclear reconstruction error (average absolute element-wise singular value difference) to evaluate each algorithm's capacity to capture the spectrum of the target matrix. Our investigation focuses on comparing with LoRA [24] and its full-rank alternatives, specifically: Sin-LoRA [26], RandLoRA [2], and Krona [16]. Further details are available in appendix C.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results We present results in Figure 1 where our analysis yields several critical insights. First, full-rank PEFT algorithms consistently achieve lower approximation errors compared to LoRA across a range of target matrices (low rank matrices excepted). The advantage of high-rank PEFT methods is particularly apparent when approximating matrices with highly de-correlated features such as random, whitened or sparse noise. Results on approximating traditionally fine-tuned weights for the CLIP vision and language backbones are more nuanced as RandLoRA and Krona struggle to emulate the fine-tuned delta. This difference does not translate to poor performance in practice though which indicates these algorithms probably learn to solve tasks in a different manner than standard fine-tuning does. KRAdapter is especially performant at approximating the CLIP fine-tuned deltas as well as the high frequency matrix, demonstrating a high adaptability to learn highly tailored features. We evidence a key limitation of full-rank methods when approximating explicitly low-rank target matrices, as the algorithms did not demonstrate a substantial advantage over LoRA. Further analysis suggests a potential affinity of LoRA towards low-frequency feature representations as full-rank algorithms struggle to improve over LoRA's solution in this case. The low frequency matrix is however also low rank by nature which would favour LoRA's solution. Overall, KRAdapter provides a strong solution to all cases, low rank matrix excluded, which conforms us in its suitability as a parameter-efficient algorithm. Visualization of the approximated matrices is available in Figure 5 in the appendix.", + "bbox": [ + 91, + 265, + 485, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Vision-language tasks", + "text_level": 1, + "bbox": [ + 89, + 724, + 294, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Common datasets", + "text_level": 1, + "bbox": [ + 89, + 753, + 261, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now move on to real data and compare the performance of several full-rank PEFT algorithms when applied to finetuning CLIP models [44] (both vision and language encoders) across a diverse suite of eleven image classification datasets (detailed in Appendix F). We evaluate Vision Transformer-based (ViT) architectures, initialized with publicly available openCLIP weights [9]. The PEFT algorithms are only trained on the attention heads. As an indicative, we", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "also report the performance of full fine-tuning (FT), where all parameters of the network are updated. We tune the hyperparameter of each method to achieve optimal performance.", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiments are conducted in few-shot settings (1, 4, and 16 shots) and with varying dataset sizes (50% and 100% of the available training data). The training data is exactly the same for each algorithm to ensure fair comparisons. Results are visualized in Figure 2 with detailed results in Table 7 in the appendix.", + "bbox": [ + 511, + 143, + 906, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our results confirm the observations of [2] about the importance of full-rank when fine-tuning vision-language models as these algorithms largely outperform LoRA with equal trainable parameters. We also confirm that full-rank PEFT algorithms can improve performance over standard fine-tuning by limiting over-fitting for larger models (ViT-H/14 especially). KRAdapter's enhanced representational capacity translates to a further performance improvements compared to other full-rank PEFT algorithms. This performance advantage is observed across most settings, although we observe a performance saturation for $\\geq 50\\%$ data settings for the larger ViT-H/14 architecture. KRAdapter's row-wise Kronecker formulation is additionally systematically superior to Krona's direct Kronecker approach.", + "bbox": [ + 511, + 241, + 908, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2. Specialized datasets", + "text_level": 1, + "bbox": [ + 513, + 497, + 697, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further investigate performance on tasks necessitating strong model adaptation, we conduct experiments using the Visual Task Adaptation Benchmark (VTAB1k) [62] (detailed in Appendix F.3). VTAB1k contains 19 datasets grouped into 3 categories. The structured and specialized subsets as especially interesting to us as they contain tasks less likely to be encountered during CLIP's pre-training. The specialized subset focuses on predicting specialist medical or satellite imagery while the structured subset aims to predict object state attributes such as distance, location, count, and rotation. Prior work has shown that CLIP exhibits limitations on these types of tasks [51], making them an ideal benchmark for assessing the enhanced representation learning capacity of full-rank adaptation algorithms. The results are presented in Table 1, where we report average accuracy over the 3 subsets (detailed results are available in appendix F.3.3). We observe that KRAdapter performs especially well on the natural and structured subsets but can struggle to generalize as well as RandLoRA to the structured subset for the larger models. Standard fine-tuning also struggles to generalize to the structured subset for ViT-H/14 which may indicate an implicit regularization in RandLoRA preventing overfitting, thus allowing to perform better in this specific setting. When averaged over the whole benchmark however, KRAdapter performs better across all architectures.", + "bbox": [ + 511, + 523, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1296", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cf3170ca8aae358d994f32c504854310e09d8828468a05e6d50dee4983d82a93.jpg", + "image_caption": [ + "(a) ViT-B/32" + ], + "image_footnote": [], + "bbox": [ + 155, + 90, + 380, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dcc96f8f8caee1d64dfaa0b70ae8db9e412d83663cdd7d4cd7ef01300caa9a5f.jpg", + "image_caption": [ + "(b) ViT-L/14", + "Figure 2. Tuning CLIP's vision and language backbone for image classification. Accuracy $(\\%)$ averaged over 11 datasets." + ], + "image_footnote": [], + "bbox": [ + 385, + 90, + 612, + 239 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1360c4de6681ab747024bafb69f71f282ace55d4053ab1c065579d836728acf8.jpg", + "image_caption": [ + "(c) ViT-H/14" + ], + "image_footnote": [], + "bbox": [ + 617, + 90, + 846, + 239 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bf95ed2fdd2335b395da07e0c394b1968843e53d805dd35681dc37f105b9f6a6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodViT-B/32ViT-L/14ViT-H/14
Nat.Spe.Struc.Avg.Nat.Spe.Struc.Avg.Nat.Spe.Struc.Avg.
LoRA70.682.645.266.182.387.054.774.684.186.756.875.9
SinLoRA72.083.153.169.482.586.856.875.484.286.958.276.4
RandLoRA74.282.453.169.982.887.058.376.083.987.359.576.9
Krona73.783.453.370.183.487.457.576.184.986.758.076.6
KRAdapter76.084.053.371.184.886.957.276.385.887.058.177.0
FT75.783.553.170.883.287.458.376.385.484.843.771.3
", + "bbox": [ + 250, + 297, + 746, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Adaptation performance on VTAB1k's structured datasets", + "bbox": [ + 299, + 448, + 694, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3. Out-of-Distribution Generalization", + "text_level": 1, + "bbox": [ + 89, + 489, + 380, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We expect that the balanced singular value distribution of KRAdapter will allow for better out-of-distribution (OOD) robustness at test time. We thus propose to investigate whether parameter-efficient adaptations, optimized on in-distribution (ID) data, exhibit robustness and transferability to out-of-distribution (OOD) scenarios. We utilize a suite of established OOD benchmark datasets: ImageNet-A [23], ImageNet-Sketch [56], ImageNet-R [22], ImageNet-V2 [46], and CIFAR-100 [30]. These datasets encompass diverse distribution shifts: adversarial, stylized, renditions and domain shift, details in appendix F.4. To quantify generalization, we introduce the ratio $r_{gen} = \\frac{\\Delta_{odd}}{\\Delta_{id}}$ , measuring the relative OOD accuracy gain ( $\\Delta_{odd}$ , averaged over OOD datasets) to ID gain ( $\\Delta_{id}$ , ImageNet validation) over the zero-shot model. Table 2 reports the results where we fine-tune the attentions layers of the vision backbone of CLIP only, maintaining frozen language embeddings to classify. KRAdapter consistently achieves a higher $r_{gen}$ than other PEFT methods. We hypothesize KRAdapter's spectral properties enable it to efficiently work in the parameter space without overfitting to a subset of directions, unlike rank-constrained methods prone to overfitting dominant ID features. To validate this hypothesis and quantify how much the algorithms stray from the pre-trained weights, we evaluate the average Nuclear and Frobenius norm of the learned updates over the attention lay", + "bbox": [ + 89, + 515, + 485, + 892 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ers. We find that KRAdapter produces updates with smaller norms than other PEFT algorithms which is in line with the very small norm of the traditionally fine-tuned model and has been previously reported to be beneficial for OOD generalization [33, 59]. Table 4 in the appendix further reports the effective rank obtained when fine-tuning the vision-language architectures where we observe that KRAdapter systematically leads to higher effective ranks.", + "bbox": [ + 511, + 489, + 908, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Commonsense reasoning", + "text_level": 1, + "bbox": [ + 511, + 630, + 740, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further investigate resource-efficient fine-tuning of Large Language Models (LLMs) for commonsense reasoning through 4-bit quantization-aware training and evaluation. We evaluate this approach on Llama3.1-8B and Qwen2.5-7B, which are well suited to consumer grade GPUs with 24GB VRAM. We fine-tune the key and value projection matrices within the attention layers, optimizing hyper-parameters for each algorithm to maximize performance. To specifically evaluate out-of-distribution (OOD) generalization, we slightly modify the commonsense reasoning benchmark established in previous research [25, 36]. While the standard commonsense benchmark trains and tests on the same tasks with identical question and answer formats, we train on the multi-choice questions with \"answer{1...5}\" format (Science-QA, ARC, and OpenBookQA) and subsequently evaluate on: (1) in-distribution test sets from these training", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1297", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1d5d3d9e66d02f8d57c9f292e23f2e85723aae8c53792d6864c6416d1d8a97f7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ViT-B/32ViT-L/14ViT-H/14
IDOODrgen↑Nuc↓Fro↓IDOODrgen↑Nuc↓Fro↓IDOODrgen↑Nuc↓Fro↓
Zero-shot62.655.5n/an/an/a75.474.1n/an/an/a77.975.7n/an/an/a
LoRA72.958.20.2719.64.183.377.20.3946.26.683.775.90.0470.07.4
SinLoRA72.858.60.31223.54.782.876.30.3061.67.183.677.10.2590.68.1
RandLoRA73.058.40.2846.25.782.876.20.29159.911.882.975.4-0.05508.321.2
Krona72.458.60.3244.14.984.177.80.4242.75.085.078.20.36123.39.2
KRAdapter72.559.90.457.32.383.678.20.499.82.884.778.80.4832.85.5
FT75.558.90.260.70.885.176.70.271.11.085.578.60.394.31.9
", + "bbox": [ + 174, + 88, + 823, + 247 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/29c9ff8cfe7f80e95421a80b91ad63031f1927bc2456d06c391677bf16700928.jpg", + "table_caption": [ + "Table 2. Generalization to distribution shift. We report ID accuracy on ImageNet, OOD accuracy averaged over 7 OOD datasets, the $r_{gen}$ ratio and the nuclear and frobenius norm of the updates. The $\\uparrow$ and $\\downarrow$ indicates if higher or lower scores are better." + ], + "table_footnote": [], + "table_body": "
MethodSIQAArcEArcCOBQAHellaSBoolQPIQAWinoGIDNIDOODAvg.
LLama3.1-8B
Zero-shot20.7326.1822.3516.0025.6943.4645.8717.6021.3225.6935.6427.23
LoRA79.0787.1274.6683.4053.7042.3576.6149.8081.0653.7055.6268.34
DoRA79.7988.2277.5684.4052.1641.7779.8757.3882.4952.1657.8070.14
SinLoRA76.4693.1486.8688.4051.5743.9484.7746.5786.2151.5758.4371.46
RandLoRA79.7987.6776.4582.6057.2252.5179.2254.3881.6357.2260.8371.23
Krona79.1787.9675.4382.4050.0759.1479.6056.6781.2450.0761.3771.31
KRAdapter79.2790.3278.1682.4055.3458.2380.9054.8582.5455.3464.6672.43
Qwen2.5-7B
Zero-shot5.025.184.104.801.7931.444.7327.704.771.7921.2910.59
LoRA51.1392.1385.8488.4071.0143.5583.7940.8179.3771.0159.7969.58
DoRA77.9487.9279.5287.8077.9557.2285.6461.6483.3077.9570.6176.95
SinLoRA77.3388.9374.8385.4058.6135.2681.3454.7881.6258.6157.1269.56
RandLoRA79.7994.8787.6389.0078.2047.4085.6960.8587.8278.2068.0477.93
Krona78.3093.8188.1491.6080.9349.0585.4257.5487.9680.9368.2378.10
KRAdapter79.7994.6186.9589.8075.3162.1484.6661.0187.7975.3170.7879.28
", + "bbox": [ + 183, + 300, + 815, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. 4bit quantized LLama3.1-8B and Qwen2.5-7B on commonsense reasoning tasks. We finetune the Key and Value matrices of the attention layers. We evaluate on in-distribution (ID), near in-distribution (NID) and out-of-distribution (OOD) test sets and bold the best results.", + "bbox": [ + 89, + 582, + 906, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "datasets; (2) near in-distribution HellaSwag test set (multichoice, \"ending{1...4}\" format); and (3) out-of-distribution BoolQ, PiQA, and WinoGrande test sets (binary answers). This evaluation framework is designed to comprehensively assess the generalization capabilities of each PEFT algorithm, a crucial attribute for deploying robust LLMs across diverse real-world scenarios. Further details about our commonsense benchmark are available in appendix G.", + "bbox": [ + 89, + 650, + 485, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 presents a comparative analysis of the PEFT methods. The results demonstrate a clear distinction: rank-restricted algorithms, such as LoRA, exhibit limited generalization to out-of-distribution (OOD) datasets. In contrast, full-rank algorithms significantly enhance the zero-shot model to answer these OOD tasks despite not specifically training for them. Notably, KRAdapter outperforms state-of-the-art methods across architectures. These findings support the", + "bbox": [ + 89, + 779, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "conclusions of Section 4.2.3, reinforcing evidence of the efficacy of KRAdapter in learning generalizable representations for language models.", + "bbox": [ + 511, + 650, + 908, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 715, + 663, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We are interested here in challenging our design choices for KRAdapter and evaluating the impact of scaling the number of trainable parameters further. We study 3 scenario named low, medium and high where we iteratively double the amount of trainable parameters from the base (low) configuration used in other experiments. In KRAdapter, this is achieved by setting $k_{1} = \\{d_{out}^{\\frac{1}{2}}, d_{out}^{\\frac{1}{3}}, d_{out}^{\\frac{1}{4}}\\}$ . Ideally, performance should increase with every parameter increase to suggest favorable scaling laws. Results are reported in Figure 3 for the vision-language task. We find that for smaller", + "bbox": [ + 511, + 744, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1298", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ffcdaa1fbc5e7936a6d4d7fba5b3dce229bfc929f36ca3b853f3e2333c5c0dc6.jpg", + "image_caption": [ + "(a) ViT-B/32" + ], + "image_footnote": [], + "bbox": [ + 178, + 119, + 359, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b15d37b5d3e44bd7f27aa64a105632e4177aa3606dca052bbd5a96d94cc4d8eb.jpg", + "image_caption": [ + "(b) ViT-L/14", + "Figure 3. Increasing the number of trainable parameters in PEFT algorithms. Accuracy $(\\%)$ averaged over 11 datasets." + ], + "image_footnote": [], + "bbox": [ + 364, + 122, + 545, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1dce14862ad3f29d2bd0c6b2d066286bd0e440bdd1c8960e2880e32ac5d1b50c.jpg", + "image_caption": [ + "(c) ViT-H/14" + ], + "image_footnote": [], + "bbox": [ + 550, + 123, + 816, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models, full-rank methods scale better with increases in parameter budgets whereas LoRA saturates faster. For larger models however, performance does not increase beyond the low configuration which ties in with the observed saturation in larger models of Figure 2. This aligns with findings of Albert et al. [2]. SinLoRA is a notable exception as more parameters leads to a decrease in accuracy. For ViT-H/14, KRAAdapter is the only full rank PEFT algorithm to not drop in accuracy with increasing budget. This highlight the robustness of KRAAdapter to over-fitting which helps it in achieving good OOD generalization.", + "bbox": [ + 89, + 306, + 485, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations", + "text_level": 1, + "bbox": [ + 89, + 505, + 212, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although we have shown KRAdapter is a good alternative to LoRA when high effective rank or out-of-distribution generalization is important, we discuss here some limitations of our approach. First KRAdapter in its most parameter-efficient configuration trains more parameters than the most efficient LoRA configurations do. In practice we find that KRAdatper is equivalent to a rank 16 to 32 LoRA, a commonly used configuration. Exploring low-rank approximations for the matrices $U$ and $V$ prior to the Khatri-Rao product could potentially mitigate this, though the theoretical implications for the derived full-rank properties in Section 3.4 require rigorous investigation. Second, empirical evidence from controlled matrix estimation experiments indicates that KRAdapter's performance may be suboptimal when estimating matrices of extremely low rank. In scenarios demanding strict subspace constraints on weight updates, such as continual learning paradigms, LoRA might exhibit superior suitability. Finally, while KRAdapter consistently demonstrates robust out-of-distribution generalization and faster training convergence, we observe instances with larger models where RandLoRA achieves marginally superior performance on purely in-distribution test sets. It is important to note that KRAdapter's in-distribution performance remains highly competitive in these cases, and its advantages in gen", + "bbox": [ + 89, + 537, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "eralization and training efficiency compared to RandLoRA generally outweigh this minor performance delta.", + "bbox": [ + 511, + 306, + 906, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 383, + 633, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces Khatri-Rao Adapters (KRAdapter), a novel parameter-efficient fine-tuning technique designed to overcome the representation limitations of low-rank methods. KRAdapter generate full-rank update matrices with demonstrably higher effective rank, enabling to capture more complex feature interactions and improving OOD generalization by staying close to the pre-trained weights. KRAdapter achieves all this without sacrificing parameter efficiency and training time. Our comprehensive experimental evaluation across vision-language, language understanding, and commonsense reasoning tasks consistently demonstrates KRAdapter produces improved results compared to LoRA and other full-rank PEFT alternatives, especially for specialized VTAB1k tasks that demand nuanced feature adaptation and for larger models that are prone to over-fitting. We specifically highlight that KRAdapter's improved representation space allows to produce representations closer to the zero-shot model that lead to better generalization under distribution shifts compared to other PEFT algorithms. Crucially, KRAdapter maintains the computational advantages of LoRA in terms of training speed and memory footprint, offering a practical and effective solution for fine-tuning billion-scale models as demonstrated in the commonsense experiments. A notable limitation of KRAdapter, however, is that the minimum number of trainable parameter is larger than the most efficient LoRA configurations which currently renders it unadapted to extreme scarcity scenarios. Future work could investigate replacing $U$ and $V$ with low-rank formulations to achieve better efficiency. We however anticipate this formulation would change the efficient rank guarantees of KRAdapter. Our findings underscore the importance of high effective rank weight updates in parameter-efficient adaptation of pre-trained models across diverse downstream applications.", + "bbox": [ + 511, + 414, + 908, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1299", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv:2303.08774, 2023. 1, 2", + "[2] Paul Albert, Frederic Z. Zhang, Hemanth Saratchandran, Cristian Rodriguez-Opazo, Anton van den Hengel, and Ehsan Abbasnejad. RandLoRA: Full rank parameter-efficient finetuning of large models. In International Conference on Learning Representations (ICLR), 2025. 1, 2, 3, 5, 8, 13, 16, 21", + "[3] Charles Beattie, Joel Z Leibo, Denis Teptyashin, Tom Ward, Marcus Wainwright, Heinrich Kuttler, Andrew Lefrancq, Simon Green, Víctor Valdés, Amir Sadik, et al. Deepmind lab. arXiv: 1612.03801, 2016. 18", + "[4] Daniel J. Beutel, Taner Topal, Akhil Mathur, Xinchi Qiu, Titouan Parcollet, and Nicholas D. Lane. Flower: A Friendly Federated Learning Research Framework. arxiv :2007.14390, 2020. 18", + "[5] Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. Piaq: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on Artificial Intelligence (AAAI), 2020. 20", + "[6] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International Conference on Machine Learning (ICML), 2020. 18", + "[7] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International Conference on Machine Learning (ICML), 2019. 1", + "[8] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE, 2017. 18", + "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5", + "[10] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing Textures in the Wild. In Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 17", + "[11] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. BoolQ: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. 20", + "[12] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. 21", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 1, 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 2", + "[15] Emma Dugas, Jorge Jared, and Will Cukierski. Diabetic retinopathy detection (2015). URL https://kaggle.com/competitions/diabetic-retinopathy-detection, 2015. 18", + "[16] Ali Edalati, Marzieh Tahaei, Ivan Kobyzev, Vahid Partovi Nia, James J Clark, and Mehdi Rezagholizadeh. Krona: Parameter efficient tuning with kronecker adapter. In Advances in Neural Information Processing Systems Workshops (NeurIPSW), 2023. 2, 3, 4, 5", + "[17] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. Vision meets Robotics: The KITTI Dataset. International Journal of Robotics Research (IJRR), 2013. 18", + "[18] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 1, 2", + "[19] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. In International Conference on Machine Learning (ICML), 2024. 2", + "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, 2015. 4", + "[21] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Introducing EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. In International Geoscience and Remote Sensing Symposium (IGARSS), 2018. 18", + "[22] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, 2021. 6, 18", + "[23] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15262-15271, 2021. 6, 18", + "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations (ICLR), 2022. 1, 2, 3, 4, 5", + "[25] Zhiqiang Hu, Lei Wang, Yihuai Lan, Wanyu Xu, Ee-Peng Lim, Lidong Bing, Xing Xu, Soujanya Poria, and Roy Ka-Wei Lee. Llm-adapters: An adapter family for parameter-efficient fine-tuning of large language models. arXiv preprint arXiv:2304.01933, 2023. 6, 21", + "[26] Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Sine Activated Low-Rank Matrices for Parameter Efficient Learning. arXiv:2403.19243, 2024. 1, 2, 3, 5", + "[27] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1300", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "diagnostic dataset for compositional language and elementary visual reasoning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 18", + "[28] Soroush Abbasi Koohpayegani, KL Navaneet, Parsa Nooralinejad, Soheil Kolouri, and Hamed Pirsiavash. NOLA: Compressing LoRA using Linear Combination of Random Basis. In International Conference on Learning Representations (ICLR), 2024. 3", + "[29] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki Markus Asano. Vera: Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. 3", + "[30] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 6, 17", + "[31] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (NeurIPS), 2012. 18", + "[32] Yann LeCun, Fu Jie Huang, and Leon Bottou. Learning methods for generic object recognition with invariance to pose and lighting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2004. 18", + "[33] Alex Lewandowski, Michal Bortkiewicz, Saurabh Kumar, András György, Dale Schuurmans, Mateusz Ostaszewski, and Marlos C Machado. Learning continually by spectral regularization. In International Conference on Learning Representations (ICLR), 2025. 6", + "[34] Fei-Fei Li, Marco Andreeto, Marc'Aurelio Ranzato, and Pietro Perona. Caltech 101, 2022. 17", + "[35] Vijay Lingam, Atula Tejaswi, Aditya Vavre, Aneesh Shetty, Gautham Krishna Gudur, Joydeep Ghosh, Alex Dimakis, Eun-sol Choi, Aleksandar Bojchevski, and Sujay Sanghavi. SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors. In International Conference on Machine Learning Workshops (ICMLW), 2024. 3", + "[36] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. In International Conference on Machine Learning (ICML), 2024. 2, 4, 6", + "[37] Y Liu, M Ott, N Goyal, J Du, M Joshi, D Chen, O Levy, M Lewis, L Zettlemoyer, and V Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach. arXiv:1907.11692, 2019. 1, 2, 22", + "[38] Loic Matthew, Irina Higgins, Demis Hassabis, and Alexander Lerchner. dSprites: Disentanglement testingSprites dataset. https://github.com/deepmind/dSprites-dataset/, 2017. 18", + "[39] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. Pissa: Principal singular values and singular vectors adaptation of large language models. Advances in Neural Information Processing Systems (NeurIPS), 2024. 3", + "[40] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. 21" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 2", + "[42] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning (NeurIPSW), 2011. 18", + "[43] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 18", + "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2, 5", + "[45] Maithra Raghu, Ben Poole, Jon Kleinberg, Surya Ganguli, and Jascha Sohl-Dickstein. On the expressive power of deep neural networks. In International Conference on Machine Learning (ICML), 2017. 2", + "[46] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning (ICML), pages 5389-5400. PMLR, 2019. 6, 18", + "[47] Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In European Signal Processing Conference, 2007. 1, 3, 16", + "[48] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 2021. 20", + "[49] Maarten Sap, Hannah Rashkin, Derek Chen, Ronan LeBras, and Yejin Choi. Socialiaq: Commonsense reasoning about social interactions. arXiv preprint arXiv:1904.09728, 2019. 20", + "[50] Houcheng Su, Daixian Liu, Mengzhu Wang, and Wei Wang. Singular Value Penalization and Semantic Data Augmentation for Fully Test-Time Adaptation. In Association for the Advancement of Artificial Intelligence (AAAI), 2024. 1", + "[51] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. In Association for Computational Linguistics (ACL), 2022. 5", + "[52] Chunlin Tian, Zhan Shi, Zhijiang Guo, Li Li, and Chengzhong Xu. HydraLoRA: An Asymmetric LoRA Architecture for Efficient Fine-Tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2", + "[53] Bastiaan S Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling. Rotation Equivariant CNNs for Digital Pathology. arxiv: 1806.03962, 2018. 18", + "[54] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: multi-task benchmark and analysis platform for natural language understanding. In International Conference on Learning Representations (ICLR), 2019. 22" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1301", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Chenglong Wang, Jiangyan Yi, Xiaohui Zhang, Jianhua Tao, Le Xu, and Ruibo Fu. Low-rank adaptation method for wav2vec2-based fake audio detection. In International Joint Conference on Artificial Intelligence Workshops (IJCAIW), 2023. 2", + "[56] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems (NeurIPS), pages 10506-10518, 2019. 6, 18", + "[57] Hanqing Wang, Yixia Li, Shuo Wang, Guanhua Chen, and Yun Chen. Milora: Harnessing minor singular components for parameter-efficient llm finetuning. arXiv:2406.09044, 2024.3", + "[58] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 18", + "[59] Yuichi Yoshida and Takeru Miyato. Spectral norm regularization for improving the generalizability of deep learning. arXiv:1705.10941, 2017. 2, 6", + "[60] Maxime Zanella and Ismail Ben Ayed. Low-Rank Few-Shot Adaptation of Vision-Language Models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[61] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. 20", + "[62] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv: 1910.04867, 2019. 5, 17", + "[63] Frederic Z Zhang, Paul Albert, Cristian Rodriguez-Opazo, Anton van den Hengel, and Ehsan Abbasnejad. Knowledge Composition using Task Vectors with Learned Anisotropic Scaling. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 4, 15, 16", + "[64] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. AdaLoRA: Adaptive budget allocation for parameter-efficient fine-tuning. In International Conference on Learning Representations (ICLR), 2023. 4", + "[65] Ming Zhong, Yelong Shen, Shuohang Wang, Yadong Lu, Yizhu Jiao, Siru Ouyang, Donghan Yu, Jiawei Han, and Weizhu Chen. Multi-lora composition for image generation. Transactions on Machine Learning Research (TMLR), 2024. 2" + ], + "bbox": [ + 91, + 92, + 485, + 781 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1302", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/5863579c-c0a2-49e2-85ef-c6b4d99e52c6_model.json b/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/5863579c-c0a2-49e2-85ef-c6b4d99e52c6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7926eefbd4de492875606a2ec99b5d6dd3e5b689 --- /dev/null +++ b/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/5863579c-c0a2-49e2-85ef-c6b4d99e52c6_model.json @@ -0,0 +1,2169 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.13, + 0.834, + 0.175 + ], + "angle": 0, + "content": "Towards Higher Effective Rank in Parameter-efficient Fine-tuning using Khatri-Rao Product" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.203, + 0.747, + 0.24 + ], + "angle": 0, + "content": "Paul Albert\\(^{1*}\\) Frederic Z. Zhang\\(^{1*}\\) Hemanth Saratchandran\\(^{2}\\) Anton van den Hengel\\(^{2}\\) Ehsan Abbasnejad\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.24, + 0.884, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\)Amazon Machine Learning, Australia \\(^{2}\\)Australian Institute for Machine Learning \\(^{3}\\)Monash University {albrtpa, fredzz}@amazon.com https://github.com/PaulAlbert31/KRAdapter" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.486, + 0.695 + ], + "angle": 0, + "content": "Parameter-efficient fine-tuning (PEFT) has become a standard approach for adapting large pre-trained models. Amongst PEFT methods, low-rank adaptation (LoRA) has achieved notable success. However, recent studies have highlighted its limitations compared against full-rank alternatives, particularly when applied to multimodal and large language models. In this work, we present a quantitative comparison amongst full-rank and low-rank PEFT methods using a synthetic matrix approximation benchmark with controlled spectral properties. Our results confirm that LoRA struggles to approximate matrices with relatively flat spectrums or high frequency components—signs of high effective ranks. To this end, we introduce KRAdapter, a novel PEFT algorithm that leverages the Khatri-Rao product to produce weight updates, which, by construction, tends to produce matrix product with a high effective rank. We demonstrate performance gains with KRAdapter on vision-language models up to 1B parameters and on large language models up to 8B parameters, particularly on unseen common-sense reasoning tasks. In addition, KRAdapter maintains the memory and compute efficiency of LoRA, making it a practical and robust alternative to fine-tune billion-scale parameter models." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.734, + 0.222, + 0.751 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Large pre-trained models have lead to great success in both computer vision and natural language processing [1, 13, 14, 37, 44]. However, fine-tuning these models for specific downstream tasks often demands substantial computational resources due to the sheer number of parameters. Recently, parameter-efficient fine-tuning (PEFT) has emerged as an effective strategy to address this challenge, enabling effi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.909, + 0.599 + ], + "angle": 0, + "content": "cient adaptation using a fraction of the trainable parameters [18]. In particular, low-rank adaptation (LoRA) [24], a prominent PEFT technique, formulates the learnable weight updates as the product of two low-rank matrices, achieving remarkable memory efficiency with little performance drop. Despite its success, recent studies [2, 26] suggest that the low-rank formulation can limit its ability to effectively fine-tune models for complex tasks. To shed light on this limitation, we compare LoRA to several full-rank PEFT methods by approximating weight matrices with controlled spectral properties. As shown in Figure 1, full-rank methods can outperform LoRA in approximating both synthetic and real weight matrices. Not all full-rank PEFT methods are equally effective—some produce approximations with rapidly decaying singular values, leading to a low effective rank [47]. Building on prior work [7, 50], which links substantial tail singular values to improved generalization and robustness, we hypothesize that a high effective rank is crucial to the performance of full-rank PEFT methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.911, + 0.894 + ], + "angle": 0, + "content": "To this end, we introduce KRAdapter, a novel PEFT approach that leverages the Khatri-Rao matrix product. By design, KRAdapter constructs provably full-rank weight matrices and empirically maintains large tail singular values. The increased effective rank indicates that the weight updates span the parameter space more uniformly, allowing the model to capture complex patterns and generalize better to distribution shifts. Across diverse architectures, including vision-language models and large language models, KRAdapter achieves performance gains while maintaining parameter efficiency. Furthermore, KRAdapter exhibits superior out-of-distribution (OOD) performance compared to LoRA and other full-rank methods. Our key contributions are: (1) a controlled matrix approximation benchmark comparing the representational advantages of full- and rank-constrained PEFT algorithms; (2) the identification of effective rank as a critical factor differentiating full-rank PEFT techniques; (3) KRAdapter, a novel Khatri-Rao product-based PEFT method that consistently achieves high effective" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.434, + 0.902 + ], + "angle": 0, + "content": "*Work done at the Australian Institute for Machine Learning." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1292" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.905, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.372, + 0.908, + 0.457 + ], + "angle": 0, + "content": "Figure 1. A visualization of the capacity of various PEFT methods to approximate the spectrum of different types of weight matrices. The first row shows the structure of the matrices. In the second row we plot the squared nuclear error as a percentage of that of the LoRA solution, and add the absolute value above each bar. Lower is better. The third row shows the singular value distribution of the solutions compared to the target (square root scale). A flatter distribution indicates a higher effective rank. We compare LoRA [24], SinLoRA [26], Krona [16], RandLoRA [2] and our proposed algorithm KRAdapter. KRAdapter achieves a better approximation to the true matrix in every case except when we force it to have low rank (and thus to match the LoRA model)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.483, + 0.487, + 0.545 + ], + "angle": 0, + "content": "rank; (4) empirical validation of its superior performance and out-of-distribution robustness compared to LoRA and other PEFT methods; and (5) evidence of KRAAdapter's improved parameter scaling efficiency among full-rank approaches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.565, + 0.233, + 0.58 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.598, + 0.487, + 0.718 + ], + "angle": 0, + "content": "The development of large pre-trained models has revolutionized various fields, yet their deployment and fine-tuning pose significant computational challenges [1, 13, 14, 37, 44]. Parameter-Efficient Fine-Tuning (PEFT) methods have emerged as a crucial paradigm to address these challenges, allowing adaptation of these massive models to downstream tasks while training only a small fraction of the parameters [18]." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.736, + 0.365, + 0.753 + ], + "angle": 0, + "content": "2.1. Low-Rank Adaptation (LoRA)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.486, + 0.903 + ], + "angle": 0, + "content": "LoRA [24] has become a cornerstone of PEFT due to its simplicity and effectiveness. LoRA introduces low-rank matrices to approximate the weight updates, significantly reducing the number of trainable parameters. Its success has led to numerous extensions and applications across various modalities [55, 60, 63, 65]. However, recent studies have begun to explore the limitations of the low-rank constraint, suggesting that it might hinder the model's ability to capture complex relationships, especially for demanding tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.909, + 0.604 + ], + "angle": 0, + "content": "[2, 16, 26]. Furthermore, the spectral properties of weight matrices, which influence a network's capacity for learning and generalization [41, 45, 59], may be constrained by the low-rank nature of LoRA updates. Our work builds upon these observations, confirming the need for methods that can capture richer feature interactions while maintaining parameter efficiency, potentially through weight updates with more desirable spectral characteristics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.625, + 0.804, + 0.642 + ], + "angle": 0, + "content": "2.2. Enhancing the training efficiency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.911, + 0.82 + ], + "angle": 0, + "content": "To improve convergence speed and overall performance of LoRA, several improvements have been proposed. LoRA+ [19] enhances the training process by applying different learning rates to the low-rank matrices, allowing for more nuanced optimization. DoRA [36] decomposes the low-rank updates into magnitude and direction components, providing finer-grained control over the adaptation process. HydraLoRA [52] employs multiple upscaling matrices, acting as specialized \"experts\" focusing on distinct aspects of the input, potentially capturing more diverse feature interactions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.842, + 0.776, + 0.859 + ], + "angle": 0, + "content": "2.3. Further reducing parameters" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "While LoRA is inherently parameter-efficient, subsequent research has focused on pushing these limits further. Strategies" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "1293" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.487, + 0.258 + ], + "angle": 0, + "content": "to minimize the trainable parameter count beyond the rank-one case include methods leveraging linear combinations of fixed random matrices, such as VeRA [29] and NoLA [28], where learned one-dimensional vectors combine pre-defined non-trainable random matrices. Another line of work explores initializing or fixing LoRA's projection matrices using insights from the pre-trained weights. Methods like SVFT [35] and Pissa [39] fix or respectively initialize the projection matrices to the first eigenvectors corresponding to the largest singular values of the original weight matrix. Conversely, MiLoRA [57] utilizes the last eigenvectors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.274, + 0.267, + 0.29 + ], + "angle": 0, + "content": "2.4. Full-rank updates" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.487, + 0.546 + ], + "angle": 0, + "content": "To address the representational bottlenecks and potential spectral limitations inherent in low-rank approximations, several methods explore parameter-efficient ways to perform full-rank updates. The goal is to capture more complex relationships present in the data without incurring the full cost of fine-tuning all parameters. Kronecker Adapters (KronA) [16] utilize the Kronecker product to construct full-rank update matrices from smaller trainable factors. SinLoRA [26] applies a sine function parameterized by a frequency parameter on top of a low-rank update, effectively producing a full-rank update. RandLoRA [2] employs parameter-efficient random matrix combinations to generate full-rank matrices. While these methods theoretically produce full rank updates, we observe their effective rank [47] is usually low (i.e. the tail singular values are very small), suggesting they might not fully capture the desired spectral properties." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.22, + 0.576 + ], + "angle": 0, + "content": "2.5. Motivations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.485, + 0.786 + ], + "angle": 0, + "content": "Our analysis of existing work indicates that the low-rank update in LoRA can limit performance on challenging tasks and with complex architectures. While recent advancements aim to overcome this by generating theoretically full-rank updates via random basis combinations, Kronecker products, or sine activations, we question whether this theoretical rank consistently translates to high effective rank. We hypothesize that the Khatri-Rao product offers an alternative method for constructing full-rank update matrices that demonstrably achieve higher effective rank. This increased effective rank would enable to learn more complex feature representations, leading to improved performance across diverse tasks and architectures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.808, + 0.412, + 0.825 + ], + "angle": 0, + "content": "3. Khatri-Rao Adapters (KRAdapter)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.902 + ], + "angle": 0, + "content": "This section details the formulation of the proposed Khatri-Rao Adapters (KRAdapter) for parameter-efficient finetuning. We begin by briefly discussing Low-Rank Adaptation (LoRA) before introducing the KRAdapter mechanism" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.768, + 0.107 + ], + "angle": 0, + "content": "and analyzing its parameter efficiency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.124, + 0.655, + 0.138 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.908, + 0.244 + ], + "angle": 0, + "content": "Consider a pre-trained linear layer in a deep neural network, characterized by a weight matrix \\(\\mathbf{W}_0\\in \\mathbb{R}^{d_{out}\\times d_{in}}\\) , where \\(d_{in}\\) is the input dimension and \\(d_{out}\\) is the output dimension. During standard full fine-tuning, the weight matrix \\(\\mathbf{W}_0\\) is updated directly by gradient descent. Given an input \\(\\mathbf{x}\\in\\) \\(\\mathbb{R}^{d_{in}}\\) , the output \\(\\mathbf{h}\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.267, + 0.907, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} = \\mathbf {W} _ {0} \\mathbf {x} + \\mathbf {b}, \\tag {3.1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.296, + 0.907, + 0.342 + ], + "angle": 0, + "content": "where \\(\\mathbf{b} \\in \\mathbb{R}^{d_{out}}\\) is the bias term (for simplicity, we omit the bias term in subsequent derivations but it can be easily incorporated)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.909, + 0.47 + ], + "angle": 0, + "content": "Low-Rank Adaptation (LoRA) [24] addresses the inefficiency of full fine-tuning by freezing the pre-trained weights \\(\\mathbf{W}_0\\) and introducing a low-rank update. Specifically, a low-rank matrix \\(\\Delta \\mathbf{W} = \\mathbf{BA}\\) is added to the original weights, where \\(\\mathbf{A} \\in \\mathbb{R}^{r \\times d_{in}}\\) and \\(\\mathbf{B} \\in \\mathbb{R}^{d_{out} \\times r}\\), with \\(r \\ll \\min(d_{in}, d_{out})\\) being the rank. During fine-tuning, only the parameters of \\(\\mathbf{A}\\) and \\(\\mathbf{B}\\) are updated. The output of the LoRA-adapted layer is:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.493, + 0.907, + 0.508 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} = \\left(\\mathbf {W} _ {0} + \\alpha \\mathbf {B A}\\right) \\mathbf {x}, \\tag {3.2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.516, + 0.906, + 0.547 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a scaling factor that helps in stabilizing training and controlling the magnitude of the adapter." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.563, + 0.706, + 0.577 + ], + "angle": 0, + "content": "3.2. Method formulation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.592, + 0.908, + 0.669 + ], + "angle": 0, + "content": "In contrast to LoRAs, our proposed method, Khatri-Rao Adapters (KRAdapter), achieves parameter efficiency without imposing the low-rank constraint. This is achieved by constructing \\(\\Delta \\mathbf{W}\\) using the Khatri-Rao product, otherwise known as a column-wise Kronecker product." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.909, + 0.72 + ], + "angle": 0, + "content": "Definition 3.1 (Khatri-Rao Product): Given two matrices \\(\\mathbf{U} \\in \\mathbb{R}^{a \\times c}\\) and \\(\\mathbf{V} \\in \\mathbb{R}^{b \\times c}\\), their Khatri-Rao product, denoted by \\(\\mathbf{U} \\odot \\mathbf{V}\\), is a matrix of defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.731, + 0.907, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\mathbf {U} \\odot \\mathbf {V} = \\left[ \\begin{array}{c} u _ {1 1} \\mathbf {v} _ {1}, u _ {1 2} \\mathbf {v} _ {2}, \\dots , u _ {1 c} \\mathbf {v} _ {c} \\\\ \\dots \\\\ u _ {a 1} \\mathbf {v} _ {1}, u _ {a 2} \\mathbf {v} _ {2}, \\dots , u _ {a c} \\mathbf {v} _ {c} \\end{array} \\right] \\in \\mathbb {R} ^ {a b \\times c}, \\tag {3.3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.79, + 0.906, + 0.821 + ], + "angle": 0, + "content": "where \\(u_{ij}\\) is the entry in the \\(i\\)-th row and \\(j\\)-th column of \\(\\mathbf{U}\\) and \\(\\mathbf{v}_j \\in \\mathbb{R}^b\\) is the \\(j\\)-th column of \\(\\mathbf{V}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.828, + 0.908, + 0.873 + ], + "angle": 0, + "content": "For a weight update \\(\\Delta \\mathbf{W}\\) of size \\(d_{out} \\times d_{in}\\), define \\(\\mathbf{U} \\in \\mathbb{R}^{k_1 \\times d_{in}}\\) and \\(\\mathbf{V} \\in \\mathbb{R}^{k_2 \\times d_{in}}\\), where \\(k_1 k_2 = d_{out}\\). The forward pass of a linear layer with KRAdapter is as follows" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} = \\left(\\mathbf {W} _ {0} + \\alpha \\mathbf {U} \\odot \\mathbf {V}\\right) \\mathbf {x}, \\tag {3.4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1294" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a scaling factor similar to LoRA that we typically set to 0.1. Prior to training, \\(\\mathbf{U}\\) is initialized as zeros and \\(\\mathbf{V}\\) using a Kaiming uniform initialization [20] with a negative slope coefficient of \\(\\sqrt{1 / k_1}\\) which we empirically find to be the optimal for convergence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.182, + 0.285, + 0.199 + ], + "angle": 0, + "content": "3.3. Parameter efficiency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.485, + 0.273 + ], + "angle": 0, + "content": "The number of trainable parameters introduced by KRAdapter is determined by the dimensions of the matrices \\(\\mathbf{U}\\) and \\(\\mathbf{V}\\), and the choice of \\(k_{1}\\) (or \\(k_{2}\\) resp.). The total number of trainable parameters for KRAdapter is:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.282, + 0.484, + 0.298 + ], + "angle": 0, + "content": "\\[\nN = d _ {i n} \\left(k _ {1} + k _ {2}\\right). \\tag {3.5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.485, + 0.382 + ], + "angle": 0, + "content": "\\(N\\) is minimized when \\(k_{1} = k_{2} = \\sqrt{d_{out}}\\), which we adopt as the default configuration. Details are deferred to Appendix A.1. When \\(d_{out}\\) is not a square number, we set \\(k_{1} = \\lfloor \\sqrt{d_{out}} \\rfloor\\) and \\(k_{2} = \\lceil \\frac{d_{out}}{k_{1}} \\rceil\\) such that \\(k_{1}k_{2} \\geq d_{out}\\). The resultant product \\(\\mathbf{U} \\odot \\mathbf{V}\\) is then truncated to size \\(d_{out} \\times d_{in}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.45 + ], + "angle": 0, + "content": "KRAdapter achieves the same parameter reduction as LoRA of rank between 16 and 32 depending on matrix shapes, a common setting [24, 36, 64] without incurring any significant increase in computational cost." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.465, + 0.283, + 0.481 + ], + "angle": 0, + "content": "3.4. Full-rank guarantee" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.485, + 0.555 + ], + "angle": 0, + "content": "This section establishes a full rank guarantee for the Khatri-Rao product when its constituent matrices are randomly initialized. This full rank property, as we will show, is crucial to the improvement of robustness against distribution shifts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.484, + 0.668 + ], + "angle": 0, + "content": "For brevity of exposition, let us assume \\( k_{1} = k_{2} = k \\), i.e., \\( \\mathbf{U} \\in \\mathbb{R}^{k \\times d_{in}} \\) and \\( \\mathbf{V} \\in \\mathbb{R}^{k \\times d_{in}} \\). We assume that each matrix has rank \\( k \\). Specifically, we consider the case where the entries of \\( \\mathbf{U} \\) and \\( \\mathbf{V} \\) are independently and identically distributed (i.i.d.) random variables drawn from either a standard Gaussian distribution \\( \\mathcal{N}(0, 1) \\) or a uniform distribution on \\( [-\\delta, \\delta] \\) for some \\( \\delta > 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.674, + 0.484, + 0.75 + ], + "angle": 0, + "content": "Theorem 3.1. Let \\(\\mathbf{U} \\in \\mathbb{R}^{k \\times d_{in}}\\) and \\(\\mathbf{V} \\in \\mathbb{R}^{k \\times d_{in}}\\) where \\(k \\leq d_{in} \\leq k^2\\) be matrices whose entries are chosen i.i.d. from a standard Gaussian or uniform distribution. Then, the Khatri-Rao product \\(\\mathbf{U} \\odot \\mathbf{V}\\) has full column rank almost surely, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.751, + 0.359, + 0.767 + ], + "angle": 0, + "content": "\\[\nr a n k (\\mathbf {U} \\odot \\mathbf {V}) = d _ {i n}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The detailed proof of Theorem 3.1 is provided in Appendix A.2. Crucially, this theorem highlights a significant distinction from low-rank adaptation (LoRA) techniques. When employing LoRA updates to matrices \\(\\mathbf{U}\\) and \\(\\mathbf{V}\\), the resulting matrix, even after combination, can have a maximum rank of at most \\(k \\leq d_{in}\\), thus inherently producing rank-deficient matrices. In contrast, Theorem 3.1 demonstrates that the Khatri-Rao product of the same randomly" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "initialized matrices \\(\\mathbf{U}\\) and \\(\\mathbf{V}\\) almost surely achieves full column rank, thereby providing a mechanism to construct high-rank matrices in a parameter-efficient manner." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.155, + 0.837, + 0.172 + ], + "angle": 0, + "content": "3.5. Differences with Kronecker Adapters" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.185, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Since the Khatri-Rao product can be seen as column-wise Kronecker product, our proposed KRAdapter and Kronecker Adapters (KronA) [16] bear some similarity. However, due to the different constructions, the resulting properties differ significantly. Specifically, we find that the Khatri-Rao product leads to different spectral properties in the resultant matrix, which displays significantly higher tail singular values. We find this to be a desirable property for PEFT methods as it 056 increases the representation power when modelling various types of weight matrices (Figure 1) and also improves OOD generalization and robustness in language and vision tasks (Table 2 and 3 3) Further discussion supported by empirical differences is available in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.403, + 0.646, + 0.42 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.435, + 0.722, + 0.451 + ], + "angle": 0, + "content": "4.1. Matrix approximation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.539 + ], + "angle": 0, + "content": "To isolate and understand the inherent strengths and weaknesses of KRAAdapter compared to LoRA and other full-rank PEFT methods, we design a series of controlled experiments using synthetic weight matrices with distinct structural properties." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.909, + 0.849 + ], + "angle": 0, + "content": "We design six matrix patterns to investigate parameter-efficient fine-tuning algorithms. The normally distributed random matrix acts as a high-rank baseline, evaluating general approximation performance. The random \\(90\\%\\) sparse matrix simulates scenarios where critical pre-trained parameters should remain unchanged. The PCA-whitened random matrix assesses algorithms' ability to handle highly decorrelated representations. The low-rank random matrix, constructed by keeping only the top quartile of singular values, tests full-rank algorithms' capacity to model low-rank structures. The CLIP ImageNet fine-tuned weight delta (Vision or Language), obtained by the element-wise difference between the pre-trained CLIP-ViT-L/14 weights and the weights obtained after standard fine-tuning on ImageNet (also known as task vector [63]), represents a realistic target weight for LoRA in transformer in real-world fine-tuning. The High/low frequency features, generated using superposed sinusoidal functions with varying frequencies ([1000, 10000] Hz and [1, 100] Hz respectively), assess algorithms' bias towards feature frequencies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Methodology For each target matrix pattern, we train the PEFT algorithms to minimize the mean squared error (MSE) between the estimated matrix and the target matrix. We train" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1295" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "on \\(1024 \\times 768\\) matrices which is the dimension of the query part of the attention matrix in ViT-L/14. All algorithms are trained with the same or higher number of trainable parameters than KRAdapter to ensure a fair comparison. We report the absolute nuclear reconstruction error (average absolute element-wise singular value difference) to evaluate each algorithm's capacity to capture the spectrum of the target matrix. Our investigation focuses on comparing with LoRA [24] and its full-rank alternatives, specifically: Sin-LoRA [26], RandLoRA [2], and Krona [16]. Further details are available in appendix C." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.266, + 0.486, + 0.705 + ], + "angle": 0, + "content": "Results We present results in Figure 1 where our analysis yields several critical insights. First, full-rank PEFT algorithms consistently achieve lower approximation errors compared to LoRA across a range of target matrices (low rank matrices excepted). The advantage of high-rank PEFT methods is particularly apparent when approximating matrices with highly de-correlated features such as random, whitened or sparse noise. Results on approximating traditionally fine-tuned weights for the CLIP vision and language backbones are more nuanced as RandLoRA and Krona struggle to emulate the fine-tuned delta. This difference does not translate to poor performance in practice though which indicates these algorithms probably learn to solve tasks in a different manner than standard fine-tuning does. KRAdapter is especially performant at approximating the CLIP fine-tuned deltas as well as the high frequency matrix, demonstrating a high adaptability to learn highly tailored features. We evidence a key limitation of full-rank methods when approximating explicitly low-rank target matrices, as the algorithms did not demonstrate a substantial advantage over LoRA. Further analysis suggests a potential affinity of LoRA towards low-frequency feature representations as full-rank algorithms struggle to improve over LoRA's solution in this case. The low frequency matrix is however also low rank by nature which would favour LoRA's solution. Overall, KRAdapter provides a strong solution to all cases, low rank matrix excluded, which conforms us in its suitability as a parameter-efficient algorithm. Visualization of the approximated matrices is available in Figure 5 in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.295, + 0.74 + ], + "angle": 0, + "content": "4.2. Vision-language tasks" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.262, + 0.768 + ], + "angle": 0, + "content": "4.2.1. Common datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We now move on to real data and compare the performance of several full-rank PEFT algorithms when applied to finetuning CLIP models [44] (both vision and language encoders) across a diverse suite of eleven image classification datasets (detailed in Appendix F). We evaluate Vision Transformer-based (ViT) architectures, initialized with publicly available openCLIP weights [9]. The PEFT algorithms are only trained on the attention heads. As an indicative, we" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "also report the performance of full fine-tuning (FT), where all parameters of the network are updated. We tune the hyperparameter of each method to achieve optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.144, + 0.907, + 0.235 + ], + "angle": 0, + "content": "Experiments are conducted in few-shot settings (1, 4, and 16 shots) and with varying dataset sizes (50% and 100% of the available training data). The training data is exactly the same for each algorithm to ensure fair comparisons. Results are visualized in Figure 2 with detailed results in Table 7 in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.242, + 0.909, + 0.455 + ], + "angle": 0, + "content": "Our results confirm the observations of [2] about the importance of full-rank when fine-tuning vision-language models as these algorithms largely outperform LoRA with equal trainable parameters. We also confirm that full-rank PEFT algorithms can improve performance over standard fine-tuning by limiting over-fitting for larger models (ViT-H/14 especially). KRAdapter's enhanced representational capacity translates to a further performance improvements compared to other full-rank PEFT algorithms. This performance advantage is observed across most settings, although we observe a performance saturation for \\(\\geq 50\\%\\) data settings for the larger ViT-H/14 architecture. KRAdapter's row-wise Kronecker formulation is additionally systematically superior to Krona's direct Kronecker approach." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.498, + 0.698, + 0.513 + ], + "angle": 0, + "content": "4.2.2. Specialized datasets" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To further investigate performance on tasks necessitating strong model adaptation, we conduct experiments using the Visual Task Adaptation Benchmark (VTAB1k) [62] (detailed in Appendix F.3). VTAB1k contains 19 datasets grouped into 3 categories. The structured and specialized subsets as especially interesting to us as they contain tasks less likely to be encountered during CLIP's pre-training. The specialized subset focuses on predicting specialist medical or satellite imagery while the structured subset aims to predict object state attributes such as distance, location, count, and rotation. Prior work has shown that CLIP exhibits limitations on these types of tasks [51], making them an ideal benchmark for assessing the enhanced representation learning capacity of full-rank adaptation algorithms. The results are presented in Table 1, where we report average accuracy over the 3 subsets (detailed results are available in appendix F.3.3). We observe that KRAdapter performs especially well on the natural and structured subsets but can struggle to generalize as well as RandLoRA to the structured subset for the larger models. Standard fine-tuning also struggles to generalize to the structured subset for ViT-H/14 which may indicate an implicit regularization in RandLoRA preventing overfitting, thus allowing to perform better in this specific setting. When averaged over the whole benchmark however, KRAdapter performs better across all architectures." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1296" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.092, + 0.382, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.246, + 0.303, + 0.258 + ], + "angle": 0, + "content": "(a) ViT-B/32" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.092, + 0.614, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.246, + 0.534, + 0.258 + ], + "angle": 0, + "content": "(b) ViT-L/14" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.092, + 0.847, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.246, + 0.766, + 0.258 + ], + "angle": 0, + "content": "(c) ViT-H/14" + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.27, + 0.856, + 0.285 + ], + "angle": 0, + "content": "Figure 2. Tuning CLIP's vision and language backbone for image classification. Accuracy \\((\\%)\\) averaged over 11 datasets." + }, + { + "type": "table", + "bbox": [ + 0.25, + 0.298, + 0.747, + 0.438 + ], + "angle": 0, + "content": "
MethodViT-B/32ViT-L/14ViT-H/14
Nat.Spe.Struc.Avg.Nat.Spe.Struc.Avg.Nat.Spe.Struc.Avg.
LoRA70.682.645.266.182.387.054.774.684.186.756.875.9
SinLoRA72.083.153.169.482.586.856.875.484.286.958.276.4
RandLoRA74.282.453.169.982.887.058.376.083.987.359.576.9
Krona73.783.453.370.183.487.457.576.184.986.758.076.6
KRAdapter76.084.053.371.184.886.957.276.385.887.058.177.0
FT75.783.553.170.883.287.458.376.385.484.843.771.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.3, + 0.449, + 0.696, + 0.463 + ], + "angle": 0, + "content": "Table 1. Adaptation performance on VTAB1k's structured datasets" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.49, + 0.382, + 0.504 + ], + "angle": 0, + "content": "4.2.3. Out-of-Distribution Generalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.486, + 0.893 + ], + "angle": 0, + "content": "We expect that the balanced singular value distribution of KRAdapter will allow for better out-of-distribution (OOD) robustness at test time. We thus propose to investigate whether parameter-efficient adaptations, optimized on in-distribution (ID) data, exhibit robustness and transferability to out-of-distribution (OOD) scenarios. We utilize a suite of established OOD benchmark datasets: ImageNet-A [23], ImageNet-Sketch [56], ImageNet-R [22], ImageNet-V2 [46], and CIFAR-100 [30]. These datasets encompass diverse distribution shifts: adversarial, stylized, renditions and domain shift, details in appendix F.4. To quantify generalization, we introduce the ratio \\( r_{gen} = \\frac{\\Delta_{odd}}{\\Delta_{id}} \\), measuring the relative OOD accuracy gain (\\( \\Delta_{odd} \\), averaged over OOD datasets) to ID gain (\\( \\Delta_{id} \\), ImageNet validation) over the zero-shot model. Table 2 reports the results where we fine-tune the attentions layers of the vision backbone of CLIP only, maintaining frozen language embeddings to classify. KRAdapter consistently achieves a higher \\( r_{gen} \\) than other PEFT methods. We hypothesize KRAdapter's spectral properties enable it to efficiently work in the parameter space without overfitting to a subset of directions, unlike rank-constrained methods prone to overfitting dominant ID features. To validate this hypothesis and quantify how much the algorithms stray from the pre-trained weights, we evaluate the average Nuclear and Frobenius norm of the learned updates over the attention lay" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.909, + 0.61 + ], + "angle": 0, + "content": "ers. We find that KRAdapter produces updates with smaller norms than other PEFT algorithms which is in line with the very small norm of the traditionally fine-tuned model and has been previously reported to be beneficial for OOD generalization [33, 59]. Table 4 in the appendix further reports the effective rank obtained when fine-tuning the vision-language architectures where we observe that KRAdapter systematically leads to higher effective ranks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.631, + 0.741, + 0.647 + ], + "angle": 0, + "content": "4.3. Commonsense reasoning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We further investigate resource-efficient fine-tuning of Large Language Models (LLMs) for commonsense reasoning through 4-bit quantization-aware training and evaluation. We evaluate this approach on Llama3.1-8B and Qwen2.5-7B, which are well suited to consumer grade GPUs with 24GB VRAM. We fine-tune the key and value projection matrices within the attention layers, optimizing hyper-parameters for each algorithm to maximize performance. To specifically evaluate out-of-distribution (OOD) generalization, we slightly modify the commonsense reasoning benchmark established in previous research [25, 36]. While the standard commonsense benchmark trains and tests on the same tasks with identical question and answer formats, we train on the multi-choice questions with \"answer{1...5}\" format (Science-QA, ARC, and OpenBookQA) and subsequently evaluate on: (1) in-distribution test sets from these training" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1297" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.175, + 0.089, + 0.825, + 0.248 + ], + "angle": 0, + "content": "
ViT-B/32ViT-L/14ViT-H/14
IDOODrgen↑Nuc↓Fro↓IDOODrgen↑Nuc↓Fro↓IDOODrgen↑Nuc↓Fro↓
Zero-shot62.655.5n/an/an/a75.474.1n/an/an/a77.975.7n/an/an/a
LoRA72.958.20.2719.64.183.377.20.3946.26.683.775.90.0470.07.4
SinLoRA72.858.60.31223.54.782.876.30.3061.67.183.677.10.2590.68.1
RandLoRA73.058.40.2846.25.782.876.20.29159.911.882.975.4-0.05508.321.2
Krona72.458.60.3244.14.984.177.80.4242.75.085.078.20.36123.39.2
KRAdapter72.559.90.457.32.383.678.20.499.82.884.778.80.4832.85.5
FT75.558.90.260.70.885.176.70.271.11.085.578.60.394.31.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.259, + 0.905, + 0.288 + ], + "angle": 0, + "content": "Table 2. Generalization to distribution shift. We report ID accuracy on ImageNet, OOD accuracy averaged over 7 OOD datasets, the \\( r_{gen} \\) ratio and the nuclear and frobenius norm of the updates. The \\( \\uparrow \\) and \\( \\downarrow \\) indicates if higher or lower scores are better." + }, + { + "type": "table", + "bbox": [ + 0.184, + 0.301, + 0.816, + 0.572 + ], + "angle": 0, + "content": "
MethodSIQAArcEArcCOBQAHellaSBoolQPIQAWinoGIDNIDOODAvg.
LLama3.1-8B
Zero-shot20.7326.1822.3516.0025.6943.4645.8717.6021.3225.6935.6427.23
LoRA79.0787.1274.6683.4053.7042.3576.6149.8081.0653.7055.6268.34
DoRA79.7988.2277.5684.4052.1641.7779.8757.3882.4952.1657.8070.14
SinLoRA76.4693.1486.8688.4051.5743.9484.7746.5786.2151.5758.4371.46
RandLoRA79.7987.6776.4582.6057.2252.5179.2254.3881.6357.2260.8371.23
Krona79.1787.9675.4382.4050.0759.1479.6056.6781.2450.0761.3771.31
KRAdapter79.2790.3278.1682.4055.3458.2380.9054.8582.5455.3464.6672.43
Qwen2.5-7B
Zero-shot5.025.184.104.801.7931.444.7327.704.771.7921.2910.59
LoRA51.1392.1385.8488.4071.0143.5583.7940.8179.3771.0159.7969.58
DoRA77.9487.9279.5287.8077.9557.2285.6461.6483.3077.9570.6176.95
SinLoRA77.3388.9374.8385.4058.6135.2681.3454.7881.6258.6157.1269.56
RandLoRA79.7994.8787.6389.0078.2047.4085.6960.8587.8278.2068.0477.93
Krona78.3093.8188.1491.6080.9349.0585.4257.5487.9680.9368.2378.10
KRAdapter79.7994.6186.9589.8075.3162.1484.6661.0187.7975.3170.7879.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.583, + 0.908, + 0.625 + ], + "angle": 0, + "content": "Table 3. 4bit quantized LLama3.1-8B and Qwen2.5-7B on commonsense reasoning tasks. We finetune the Key and Value matrices of the attention layers. We evaluate on in-distribution (ID), near in-distribution (NID) and out-of-distribution (OOD) test sets and bold the best results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.486, + 0.773 + ], + "angle": 0, + "content": "datasets; (2) near in-distribution HellaSwag test set (multichoice, \"ending{1...4}\" format); and (3) out-of-distribution BoolQ, PiQA, and WinoGrande test sets (binary answers). This evaluation framework is designed to comprehensively assess the generalization capabilities of each PEFT algorithm, a crucial attribute for deploying robust LLMs across diverse real-world scenarios. Further details about our commonsense benchmark are available in appendix G." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Table 3 presents a comparative analysis of the PEFT methods. The results demonstrate a clear distinction: rank-restricted algorithms, such as LoRA, exhibit limited generalization to out-of-distribution (OOD) datasets. In contrast, full-rank algorithms significantly enhance the zero-shot model to answer these OOD tasks despite not specifically training for them. Notably, KRAdapter outperforms state-of-the-art methods across architectures. These findings support the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.909, + 0.698 + ], + "angle": 0, + "content": "conclusions of Section 4.2.3, reinforcing evidence of the efficacy of KRAdapter in learning generalizable representations for language models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.717, + 0.665, + 0.734 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We are interested here in challenging our design choices for KRAdapter and evaluating the impact of scaling the number of trainable parameters further. We study 3 scenario named low, medium and high where we iteratively double the amount of trainable parameters from the base (low) configuration used in other experiments. In KRAdapter, this is achieved by setting \\( k_{1} = \\{d_{out}^{\\frac{1}{2}}, d_{out}^{\\frac{1}{3}}, d_{out}^{\\frac{1}{4}}\\} \\). Ideally, performance should increase with every parameter increase to suggest favorable scaling laws. Results are reported in Figure 3 for the vision-language task. We find that for smaller" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1298" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.121, + 0.36, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.241, + 0.303, + 0.253 + ], + "angle": 0, + "content": "(a) ViT-B/32" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.123, + 0.547, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.241, + 0.488, + 0.253 + ], + "angle": 0, + "content": "(b) ViT-L/14" + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.124, + 0.818, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.241, + 0.721, + 0.254 + ], + "angle": 0, + "content": "(c) ViT-H/14" + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.266, + 0.847, + 0.28 + ], + "angle": 0, + "content": "Figure 3. Increasing the number of trainable parameters in PEFT algorithms. Accuracy \\((\\%)\\) averaged over 11 datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.486, + 0.474 + ], + "angle": 0, + "content": "models, full-rank methods scale better with increases in parameter budgets whereas LoRA saturates faster. For larger models however, performance does not increase beyond the low configuration which ties in with the observed saturation in larger models of Figure 2. This aligns with findings of Albert et al. [2]. SinLoRA is a notable exception as more parameters leads to a decrease in accuracy. For ViT-H/14, KRAAdapter is the only full rank PEFT algorithm to not drop in accuracy with increasing budget. This highlight the robustness of KRAAdapter to over-fitting which helps it in achieving good OOD generalization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.506, + 0.214, + 0.521 + ], + "angle": 0, + "content": "5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Although we have shown KRAdapter is a good alternative to LoRA when high effective rank or out-of-distribution generalization is important, we discuss here some limitations of our approach. First KRAdapter in its most parameter-efficient configuration trains more parameters than the most efficient LoRA configurations do. In practice we find that KRAdatper is equivalent to a rank 16 to 32 LoRA, a commonly used configuration. Exploring low-rank approximations for the matrices \\( U \\) and \\( V \\) prior to the Khatri-Rao product could potentially mitigate this, though the theoretical implications for the derived full-rank properties in Section 3.4 require rigorous investigation. Second, empirical evidence from controlled matrix estimation experiments indicates that KRAdapter's performance may be suboptimal when estimating matrices of extremely low rank. In scenarios demanding strict subspace constraints on weight updates, such as continual learning paradigms, LoRA might exhibit superior suitability. Finally, while KRAdapter consistently demonstrates robust out-of-distribution generalization and faster training convergence, we observe instances with larger models where RandLoRA achieves marginally superior performance on purely in-distribution test sets. It is important to note that KRAdapter's in-distribution performance remains highly competitive in these cases, and its advantages in gen" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.307, + 0.907, + 0.339 + ], + "angle": 0, + "content": "eralization and training efficiency compared to RandLoRA generally outweigh this minor performance delta." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.384, + 0.634, + 0.4 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.909, + 0.89 + ], + "angle": 0, + "content": "This paper introduces Khatri-Rao Adapters (KRAdapter), a novel parameter-efficient fine-tuning technique designed to overcome the representation limitations of low-rank methods. KRAdapter generate full-rank update matrices with demonstrably higher effective rank, enabling to capture more complex feature interactions and improving OOD generalization by staying close to the pre-trained weights. KRAdapter achieves all this without sacrificing parameter efficiency and training time. Our comprehensive experimental evaluation across vision-language, language understanding, and commonsense reasoning tasks consistently demonstrates KRAdapter produces improved results compared to LoRA and other full-rank PEFT alternatives, especially for specialized VTAB1k tasks that demand nuanced feature adaptation and for larger models that are prone to over-fitting. We specifically highlight that KRAdapter's improved representation space allows to produce representations closer to the zero-shot model that lead to better generalization under distribution shifts compared to other PEFT algorithms. Crucially, KRAdapter maintains the computational advantages of LoRA in terms of training speed and memory footprint, offering a practical and effective solution for fine-tuning billion-scale models as demonstrated in the commonsense experiments. A notable limitation of KRAdapter, however, is that the minimum number of trainable parameter is larger than the most efficient LoRA configurations which currently renders it unadapted to extreme scarcity scenarios. Future work could investigate replacing \\( U \\) and \\( V \\) with low-rank formulations to achieve better efficiency. We however anticipate this formulation would change the efficient rank guarantees of KRAdapter. Our findings underscore the importance of high effective rank weight updates in parameter-efficient adaptation of pre-trained models across diverse downstream applications." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1299" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv:2303.08774, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Paul Albert, Frederic Z. Zhang, Hemanth Saratchandran, Cristian Rodriguez-Opazo, Anton van den Hengel, and Ehsan Abbasnejad. RandLoRA: Full rank parameter-efficient finetuning of large models. In International Conference on Learning Representations (ICLR), 2025. 1, 2, 3, 5, 8, 13, 16, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.484, + 0.296 + ], + "angle": 0, + "content": "[3] Charles Beattie, Joel Z Leibo, Denis Teptyashin, Tom Ward, Marcus Wainwright, Heinrich Kuttler, Andrew Lefrancq, Simon Green, Víctor Valdés, Amir Sadik, et al. Deepmind lab. arXiv: 1612.03801, 2016. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[4] Daniel J. Beutel, Taner Topal, Akhil Mathur, Xinchi Qiu, Titouan Parcollet, and Nicholas D. Lane. Flower: A Friendly Federated Learning Research Framework. arxiv :2007.14390, 2020. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.484, + 0.409 + ], + "angle": 0, + "content": "[5] Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. Piaq: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on Artificial Intelligence (AAAI), 2020. 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.484, + 0.465 + ], + "angle": 0, + "content": "[6] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International Conference on Machine Learning (ICML), 2020. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.484, + 0.522 + ], + "angle": 0, + "content": "[7] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International Conference on Machine Learning (ICML), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.524, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[8] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE, 2017. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.648 + ], + "angle": 0, + "content": "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.484, + 0.691 + ], + "angle": 0, + "content": "[10] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing Textures in the Wild. In Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[11] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. BoolQ: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[12] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.16 + ], + "angle": 0, + "content": "[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.908, + 0.204 + ], + "angle": 0, + "content": "[15] Emma Dugas, Jorge Jared, and Will Cukierski. Diabetic retinopathy detection (2015). URL https://kaggle.com/competitions/diabetic-retinopathy-detection, 2015. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.908, + 0.272 + ], + "angle": 0, + "content": "[16] Ali Edalati, Marzieh Tahaei, Ivan Kobyzev, Vahid Partovi Nia, James J Clark, and Mehdi Rezagholizadeh. Krona: Parameter efficient tuning with kronecker adapter. In Advances in Neural Information Processing Systems Workshops (NeurIPSW), 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.908, + 0.315 + ], + "angle": 0, + "content": "[17] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. Vision meets Robotics: The KITTI Dataset. International Journal of Robotics Research (IJRR), 2013. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.316, + 0.907, + 0.37 + ], + "angle": 0, + "content": "[18] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.907, + 0.413 + ], + "angle": 0, + "content": "[19] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. In International Conference on Machine Learning (ICML), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.907, + 0.469 + ], + "angle": 0, + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.907, + 0.538 + ], + "angle": 0, + "content": "[21] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Introducing EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. In International Geoscience and Remote Sensing Symposium (IGARSS), 2018. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.908, + 0.621 + ], + "angle": 0, + "content": "[22] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, 2021. 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.908, + 0.677 + ], + "angle": 0, + "content": "[23] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15262-15271, 2021. 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.908, + 0.747 + ], + "angle": 0, + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations (ICLR), 2022. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.908, + 0.816 + ], + "angle": 0, + "content": "[25] Zhiqiang Hu, Lei Wang, Yihuai Lan, Wanyu Xu, Ee-Peng Lim, Lidong Bing, Xing Xu, Soujanya Poria, and Roy Ka-Wei Lee. Llm-adapters: An adapter family for parameter-efficient fine-tuning of large language models. arXiv preprint arXiv:2304.01933, 2023. 6, 21" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[26] Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Sine Activated Low-Rank Matrices for Parameter Efficient Learning. arXiv:2403.19243, 2024. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[27] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1300" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "diagnostic dataset for compositional language and elementary visual reasoning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[28] Soroush Abbasi Koohpayegani, KL Navaneet, Parsa Nooralinejad, Soheil Kolouri, and Hamed Pirsiavash. NOLA: Compressing LoRA using Linear Combination of Random Basis. In International Conference on Learning Representations (ICLR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.209, + 0.484, + 0.263 + ], + "angle": 0, + "content": "[29] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki Markus Asano. Vera: Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.266, + 0.483, + 0.307 + ], + "angle": 0, + "content": "[30] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 6, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.31, + 0.484, + 0.364 + ], + "angle": 0, + "content": "[31] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (NeurIPS), 2012. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.368, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[32] Yann LeCun, Fu Jie Huang, and Leon Bottou. Learning methods for generic object recognition with invariance to pose and lighting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2004. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.484, + 0.495 + ], + "angle": 0, + "content": "[33] Alex Lewandowski, Michal Bortkiewicz, Saurabh Kumar, András György, Dale Schuurmans, Mateusz Ostaszewski, and Marlos C Machado. Learning continually by spectral regularization. In International Conference on Learning Representations (ICLR), 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[34] Fei-Fei Li, Marco Andreeto, Marc'Aurelio Ranzato, and Pietro Perona. Caltech 101, 2022. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.528, + 0.484, + 0.611 + ], + "angle": 0, + "content": "[35] Vijay Lingam, Atula Tejaswi, Aditya Vavre, Aneesh Shetty, Gautham Krishna Gudur, Joydeep Ghosh, Alex Dimakis, Eun-sol Choi, Aleksandar Bojchevski, and Sujay Sanghavi. SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors. In International Conference on Machine Learning Workshops (ICMLW), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.614, + 0.484, + 0.683 + ], + "angle": 0, + "content": "[36] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. In International Conference on Machine Learning (ICML), 2024. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.484, + 0.74 + ], + "angle": 0, + "content": "[37] Y Liu, M Ott, N Goyal, J Du, M Joshi, D Chen, O Levy, M Lewis, L Zettlemoyer, and V Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach. arXiv:1907.11692, 2019. 1, 2, 22" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.743, + 0.484, + 0.785 + ], + "angle": 0, + "content": "[38] Loic Matthew, Irina Higgins, Demis Hassabis, and Alexander Lerchner. dSprites: Disentanglement testingSprites dataset. https://github.com/deepmind/dSprites-dataset/, 2017. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[39] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. Pissa: Principal singular values and singular vectors adaptation of large language models. Advances in Neural Information Processing Systems (NeurIPS), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[40] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. 21" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[41] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.908, + 0.217 + ], + "angle": 0, + "content": "[42] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning (NeurIPSW), 2011. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.907, + 0.259 + ], + "angle": 0, + "content": "[43] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.908, + 0.342 + ], + "angle": 0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.907, + 0.399 + ], + "angle": 0, + "content": "[45] Maithra Raghu, Ben Poole, Jon Kleinberg, Surya Ganguli, and Jascha Sohl-Dickstein. On the expressive power of deep neural networks. In International Conference on Machine Learning (ICML), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.907, + 0.455 + ], + "angle": 0, + "content": "[46] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning (ICML), pages 5389-5400. PMLR, 2019. 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.907, + 0.496 + ], + "angle": 0, + "content": "[47] Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In European Signal Processing Conference, 2007. 1, 3, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.907, + 0.551 + ], + "angle": 0, + "content": "[48] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 2021. 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.907, + 0.607 + ], + "angle": 0, + "content": "[49] Maarten Sap, Hannah Rashkin, Derek Chen, Ronan LeBras, and Yejin Choi. Socialiaq: Commonsense reasoning about social interactions. arXiv preprint arXiv:1904.09728, 2019. 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.907, + 0.663 + ], + "angle": 0, + "content": "[50] Houcheng Su, Daixian Liu, Mengzhu Wang, and Wei Wang. Singular Value Penalization and Semantic Data Augmentation for Fully Test-Time Adaptation. In Association for the Advancement of Artificial Intelligence (AAAI), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.907, + 0.733 + ], + "angle": 0, + "content": "[51] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. In Association for Computational Linguistics (ACL), 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.907, + 0.789 + ], + "angle": 0, + "content": "[52] Chunlin Tian, Zhan Shi, Zhijiang Guo, Li Li, and Chengzhong Xu. HydraLoRA: An Asymmetric LoRA Architecture for Efficient Fine-Tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[53] Bastiaan S Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling. Rotation Equivariant CNNs for Digital Pathology. arxiv: 1806.03962, 2018. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[54] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: multi-task benchmark and analysis platform for natural language understanding. In International Conference on Learning Representations (ICLR), 2019. 22" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1301" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.16 + ], + "angle": 0, + "content": "[55] Chenglong Wang, Jiangyan Yi, Xiaohui Zhang, Jianhua Tao, Le Xu, and Ruibo Fu. Low-rank adaptation method for wav2vec2-based fake audio detection. In International Joint Conference on Artificial Intelligence Workshops (IJCAIW), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[56] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems (NeurIPS), pages 10506-10518, 2019. 6, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[57] Hanqing Wang, Yixia Li, Shuo Wang, Guanhua Chen, and Yun Chen. Milora: Harnessing minor singular components for parameter-efficient llm finetuning. arXiv:2406.09044, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[58] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[59] Yuichi Yoshida and Takeru Miyato. Spectral norm regularization for improving the generalizability of deep learning. arXiv:1705.10941, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.484, + 0.443 + ], + "angle": 0, + "content": "[60] Maxime Zanella and Ismail Ben Ayed. Low-Rank Few-Shot Adaptation of Vision-Language Models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[61] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.571 + ], + "angle": 0, + "content": "[62] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv: 1910.04867, 2019. 5, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.484, + 0.642 + ], + "angle": 0, + "content": "[63] Frederic Z Zhang, Paul Albert, Cristian Rodriguez-Opazo, Anton van den Hengel, and Ehsan Abbasnejad. Knowledge Composition using Task Vectors with Learned Anisotropic Scaling. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 4, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.484, + 0.713 + ], + "angle": 0, + "content": "[64] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. AdaLoRA: Adaptive budget allocation for parameter-efficient fine-tuning. In International Conference on Learning Representations (ICLR), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.484, + 0.782 + ], + "angle": 0, + "content": "[65] Ming Zhong, Yelong Shen, Shuohang Wang, Yadong Lu, Yizhu Jiao, Siru Ouyang, Donghan Yu, Jiawei Han, and Weizhu Chen. Multi-lora composition for image generation. Transactions on Machine Learning Research (TMLR), 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1302" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/5863579c-c0a2-49e2-85ef-c6b4d99e52c6_origin.pdf b/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/5863579c-c0a2-49e2-85ef-c6b4d99e52c6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a46b86eaecb6364bd4ceb117195b45389a5a815c --- /dev/null +++ b/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/5863579c-c0a2-49e2-85ef-c6b4d99e52c6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f1a1dd9e43fbd4ca9d947ba10a4c33f18637b6e7fd0db624913d2bb7b336837 +size 563215 diff --git a/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/full.md b/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1654c91260475e15d6ef85e52b9de1dcda0bacb6 --- /dev/null +++ b/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/full.md @@ -0,0 +1,293 @@ +# Towards Higher Effective Rank in Parameter-efficient Fine-tuning using Khatri-Rao Product + +Paul Albert $^{1*}$ Frederic Z. Zhang $^{1*}$ Hemanth Saratchandran $^{2}$ Anton van den Hengel $^{2}$ Ehsan Abbasnejad $^{3}$ + +$^{1}$ Amazon Machine Learning, Australia $^{2}$ Australian Institute for Machine Learning $^{3}$ Monash University {albrtpa, fredzz}@amazon.com https://github.com/PaulAlbert31/KRAdapter + +# Abstract + +Parameter-efficient fine-tuning (PEFT) has become a standard approach for adapting large pre-trained models. Amongst PEFT methods, low-rank adaptation (LoRA) has achieved notable success. However, recent studies have highlighted its limitations compared against full-rank alternatives, particularly when applied to multimodal and large language models. In this work, we present a quantitative comparison amongst full-rank and low-rank PEFT methods using a synthetic matrix approximation benchmark with controlled spectral properties. Our results confirm that LoRA struggles to approximate matrices with relatively flat spectrums or high frequency components—signs of high effective ranks. To this end, we introduce KRAdapter, a novel PEFT algorithm that leverages the Khatri-Rao product to produce weight updates, which, by construction, tends to produce matrix product with a high effective rank. We demonstrate performance gains with KRAdapter on vision-language models up to 1B parameters and on large language models up to 8B parameters, particularly on unseen common-sense reasoning tasks. In addition, KRAdapter maintains the memory and compute efficiency of LoRA, making it a practical and robust alternative to fine-tune billion-scale parameter models. + +# 1. Introduction + +Large pre-trained models have lead to great success in both computer vision and natural language processing [1, 13, 14, 37, 44]. However, fine-tuning these models for specific downstream tasks often demands substantial computational resources due to the sheer number of parameters. Recently, parameter-efficient fine-tuning (PEFT) has emerged as an effective strategy to address this challenge, enabling effi + +cient adaptation using a fraction of the trainable parameters [18]. In particular, low-rank adaptation (LoRA) [24], a prominent PEFT technique, formulates the learnable weight updates as the product of two low-rank matrices, achieving remarkable memory efficiency with little performance drop. Despite its success, recent studies [2, 26] suggest that the low-rank formulation can limit its ability to effectively fine-tune models for complex tasks. To shed light on this limitation, we compare LoRA to several full-rank PEFT methods by approximating weight matrices with controlled spectral properties. As shown in Figure 1, full-rank methods can outperform LoRA in approximating both synthetic and real weight matrices. Not all full-rank PEFT methods are equally effective—some produce approximations with rapidly decaying singular values, leading to a low effective rank [47]. Building on prior work [7, 50], which links substantial tail singular values to improved generalization and robustness, we hypothesize that a high effective rank is crucial to the performance of full-rank PEFT methods. + +To this end, we introduce KRAdapter, a novel PEFT approach that leverages the Khatri-Rao matrix product. By design, KRAdapter constructs provably full-rank weight matrices and empirically maintains large tail singular values. The increased effective rank indicates that the weight updates span the parameter space more uniformly, allowing the model to capture complex patterns and generalize better to distribution shifts. Across diverse architectures, including vision-language models and large language models, KRAdapter achieves performance gains while maintaining parameter efficiency. Furthermore, KRAdapter exhibits superior out-of-distribution (OOD) performance compared to LoRA and other full-rank methods. Our key contributions are: (1) a controlled matrix approximation benchmark comparing the representational advantages of full- and rank-constrained PEFT algorithms; (2) the identification of effective rank as a critical factor differentiating full-rank PEFT techniques; (3) KRAdapter, a novel Khatri-Rao product-based PEFT method that consistently achieves high effective + +![](images/9f5d53b3f50b7938274151752709074ce2d67af77e39f571cf30f1a0a4834227.jpg) +Figure 1. A visualization of the capacity of various PEFT methods to approximate the spectrum of different types of weight matrices. The first row shows the structure of the matrices. In the second row we plot the squared nuclear error as a percentage of that of the LoRA solution, and add the absolute value above each bar. Lower is better. The third row shows the singular value distribution of the solutions compared to the target (square root scale). A flatter distribution indicates a higher effective rank. We compare LoRA [24], SinLoRA [26], Krona [16], RandLoRA [2] and our proposed algorithm KRAdapter. KRAdapter achieves a better approximation to the true matrix in every case except when we force it to have low rank (and thus to match the LoRA model). + +rank; (4) empirical validation of its superior performance and out-of-distribution robustness compared to LoRA and other PEFT methods; and (5) evidence of KRAAdapter's improved parameter scaling efficiency among full-rank approaches. + +# 2. Related Work + +The development of large pre-trained models has revolutionized various fields, yet their deployment and fine-tuning pose significant computational challenges [1, 13, 14, 37, 44]. Parameter-Efficient Fine-Tuning (PEFT) methods have emerged as a crucial paradigm to address these challenges, allowing adaptation of these massive models to downstream tasks while training only a small fraction of the parameters [18]. + +# 2.1. Low-Rank Adaptation (LoRA) + +LoRA [24] has become a cornerstone of PEFT due to its simplicity and effectiveness. LoRA introduces low-rank matrices to approximate the weight updates, significantly reducing the number of trainable parameters. Its success has led to numerous extensions and applications across various modalities [55, 60, 63, 65]. However, recent studies have begun to explore the limitations of the low-rank constraint, suggesting that it might hinder the model's ability to capture complex relationships, especially for demanding tasks + +[2, 16, 26]. Furthermore, the spectral properties of weight matrices, which influence a network's capacity for learning and generalization [41, 45, 59], may be constrained by the low-rank nature of LoRA updates. Our work builds upon these observations, confirming the need for methods that can capture richer feature interactions while maintaining parameter efficiency, potentially through weight updates with more desirable spectral characteristics. + +# 2.2. Enhancing the training efficiency + +To improve convergence speed and overall performance of LoRA, several improvements have been proposed. LoRA+ [19] enhances the training process by applying different learning rates to the low-rank matrices, allowing for more nuanced optimization. DoRA [36] decomposes the low-rank updates into magnitude and direction components, providing finer-grained control over the adaptation process. HydraLoRA [52] employs multiple upscaling matrices, acting as specialized "experts" focusing on distinct aspects of the input, potentially capturing more diverse feature interactions. + +# 2.3. Further reducing parameters + +While LoRA is inherently parameter-efficient, subsequent research has focused on pushing these limits further. Strategies + +to minimize the trainable parameter count beyond the rank-one case include methods leveraging linear combinations of fixed random matrices, such as VeRA [29] and NoLA [28], where learned one-dimensional vectors combine pre-defined non-trainable random matrices. Another line of work explores initializing or fixing LoRA's projection matrices using insights from the pre-trained weights. Methods like SVFT [35] and Pissa [39] fix or respectively initialize the projection matrices to the first eigenvectors corresponding to the largest singular values of the original weight matrix. Conversely, MiLoRA [57] utilizes the last eigenvectors. + +# 2.4. Full-rank updates + +To address the representational bottlenecks and potential spectral limitations inherent in low-rank approximations, several methods explore parameter-efficient ways to perform full-rank updates. The goal is to capture more complex relationships present in the data without incurring the full cost of fine-tuning all parameters. Kronecker Adapters (KronA) [16] utilize the Kronecker product to construct full-rank update matrices from smaller trainable factors. SinLoRA [26] applies a sine function parameterized by a frequency parameter on top of a low-rank update, effectively producing a full-rank update. RandLoRA [2] employs parameter-efficient random matrix combinations to generate full-rank matrices. While these methods theoretically produce full rank updates, we observe their effective rank [47] is usually low (i.e. the tail singular values are very small), suggesting they might not fully capture the desired spectral properties. + +# 2.5. Motivations + +Our analysis of existing work indicates that the low-rank update in LoRA can limit performance on challenging tasks and with complex architectures. While recent advancements aim to overcome this by generating theoretically full-rank updates via random basis combinations, Kronecker products, or sine activations, we question whether this theoretical rank consistently translates to high effective rank. We hypothesize that the Khatri-Rao product offers an alternative method for constructing full-rank update matrices that demonstrably achieve higher effective rank. This increased effective rank would enable to learn more complex feature representations, leading to improved performance across diverse tasks and architectures. + +# 3. Khatri-Rao Adapters (KRAdapter) + +This section details the formulation of the proposed Khatri-Rao Adapters (KRAdapter) for parameter-efficient finetuning. We begin by briefly discussing Low-Rank Adaptation (LoRA) before introducing the KRAdapter mechanism + +and analyzing its parameter efficiency. + +# 3.1. Preliminaries + +Consider a pre-trained linear layer in a deep neural network, characterized by a weight matrix $\mathbf{W}_0\in \mathbb{R}^{d_{out}\times d_{in}}$ , where $d_{in}$ is the input dimension and $d_{out}$ is the output dimension. During standard full fine-tuning, the weight matrix $\mathbf{W}_0$ is updated directly by gradient descent. Given an input $\mathbf{x}\in$ $\mathbb{R}^{d_{in}}$ , the output $\mathbf{h}$ is computed as: + +$$ +\mathbf {h} = \mathbf {W} _ {0} \mathbf {x} + \mathbf {b}, \tag {3.1} +$$ + +where $\mathbf{b} \in \mathbb{R}^{d_{out}}$ is the bias term (for simplicity, we omit the bias term in subsequent derivations but it can be easily incorporated). + +Low-Rank Adaptation (LoRA) [24] addresses the inefficiency of full fine-tuning by freezing the pre-trained weights $\mathbf{W}_0$ and introducing a low-rank update. Specifically, a low-rank matrix $\Delta \mathbf{W} = \mathbf{BA}$ is added to the original weights, where $\mathbf{A} \in \mathbb{R}^{r \times d_{in}}$ and $\mathbf{B} \in \mathbb{R}^{d_{out} \times r}$ , with $r \ll \min(d_{in}, d_{out})$ being the rank. During fine-tuning, only the parameters of $\mathbf{A}$ and $\mathbf{B}$ are updated. The output of the LoRA-adapted layer is: + +$$ +\mathbf {h} = \left(\mathbf {W} _ {0} + \alpha \mathbf {B A}\right) \mathbf {x}, \tag {3.2} +$$ + +where $\alpha$ is a scaling factor that helps in stabilizing training and controlling the magnitude of the adapter. + +# 3.2. Method formulation + +In contrast to LoRAs, our proposed method, Khatri-Rao Adapters (KRAdapter), achieves parameter efficiency without imposing the low-rank constraint. This is achieved by constructing $\Delta \mathbf{W}$ using the Khatri-Rao product, otherwise known as a column-wise Kronecker product. + +Definition 3.1 (Khatri-Rao Product): Given two matrices $\mathbf{U} \in \mathbb{R}^{a \times c}$ and $\mathbf{V} \in \mathbb{R}^{b \times c}$ , their Khatri-Rao product, denoted by $\mathbf{U} \odot \mathbf{V}$ , is a matrix of defined as: + +$$ +\mathbf {U} \odot \mathbf {V} = \left[ \begin{array}{c} u _ {1 1} \mathbf {v} _ {1}, u _ {1 2} \mathbf {v} _ {2}, \dots , u _ {1 c} \mathbf {v} _ {c} \\ \dots \\ u _ {a 1} \mathbf {v} _ {1}, u _ {a 2} \mathbf {v} _ {2}, \dots , u _ {a c} \mathbf {v} _ {c} \end{array} \right] \in \mathbb {R} ^ {a b \times c}, \tag {3.3} +$$ + +where $u_{ij}$ is the entry in the $i$ -th row and $j$ -th column of $\mathbf{U}$ and $\mathbf{v}_j \in \mathbb{R}^b$ is the $j$ -th column of $\mathbf{V}$ . + +For a weight update $\Delta \mathbf{W}$ of size $d_{out} \times d_{in}$ , define $\mathbf{U} \in \mathbb{R}^{k_1 \times d_{in}}$ and $\mathbf{V} \in \mathbb{R}^{k_2 \times d_{in}}$ , where $k_1 k_2 = d_{out}$ . The forward pass of a linear layer with KRAdapter is as follows + +$$ +\mathbf {h} = \left(\mathbf {W} _ {0} + \alpha \mathbf {U} \odot \mathbf {V}\right) \mathbf {x}, \tag {3.4} +$$ + +where $\alpha$ is a scaling factor similar to LoRA that we typically set to 0.1. Prior to training, $\mathbf{U}$ is initialized as zeros and $\mathbf{V}$ using a Kaiming uniform initialization [20] with a negative slope coefficient of $\sqrt{1 / k_1}$ which we empirically find to be the optimal for convergence. + +# 3.3. Parameter efficiency + +The number of trainable parameters introduced by KRAdapter is determined by the dimensions of the matrices $\mathbf{U}$ and $\mathbf{V}$ , and the choice of $k_{1}$ (or $k_{2}$ resp.). The total number of trainable parameters for KRAdapter is: + +$$ +N = d _ {i n} \left(k _ {1} + k _ {2}\right). \tag {3.5} +$$ + +$N$ is minimized when $k_{1} = k_{2} = \sqrt{d_{out}}$ , which we adopt as the default configuration. Details are deferred to Appendix A.1. When $d_{out}$ is not a square number, we set $k_{1} = \lfloor \sqrt{d_{out}} \rfloor$ and $k_{2} = \lceil \frac{d_{out}}{k_{1}} \rceil$ such that $k_{1}k_{2} \geq d_{out}$ . The resultant product $\mathbf{U} \odot \mathbf{V}$ is then truncated to size $d_{out} \times d_{in}$ . + +KRAdapter achieves the same parameter reduction as LoRA of rank between 16 and 32 depending on matrix shapes, a common setting [24, 36, 64] without incurring any significant increase in computational cost. + +# 3.4. Full-rank guarantee + +This section establishes a full rank guarantee for the Khatri-Rao product when its constituent matrices are randomly initialized. This full rank property, as we will show, is crucial to the improvement of robustness against distribution shifts. + +For brevity of exposition, let us assume $k_{1} = k_{2} = k$ , i.e., $\mathbf{U} \in \mathbb{R}^{k \times d_{in}}$ and $\mathbf{V} \in \mathbb{R}^{k \times d_{in}}$ . We assume that each matrix has rank $k$ . Specifically, we consider the case where the entries of $\mathbf{U}$ and $\mathbf{V}$ are independently and identically distributed (i.i.d.) random variables drawn from either a standard Gaussian distribution $\mathcal{N}(0, 1)$ or a uniform distribution on $[-\delta, \delta]$ for some $\delta > 0$ . + +Theorem 3.1. Let $\mathbf{U} \in \mathbb{R}^{k \times d_{in}}$ and $\mathbf{V} \in \mathbb{R}^{k \times d_{in}}$ where $k \leq d_{in} \leq k^2$ be matrices whose entries are chosen i.i.d. from a standard Gaussian or uniform distribution. Then, the Khatri-Rao product $\mathbf{U} \odot \mathbf{V}$ has full column rank almost surely, i.e., + +$$ +r a n k (\mathbf {U} \odot \mathbf {V}) = d _ {i n}. +$$ + +The detailed proof of Theorem 3.1 is provided in Appendix A.2. Crucially, this theorem highlights a significant distinction from low-rank adaptation (LoRA) techniques. When employing LoRA updates to matrices $\mathbf{U}$ and $\mathbf{V}$ , the resulting matrix, even after combination, can have a maximum rank of at most $k \leq d_{in}$ , thus inherently producing rank-deficient matrices. In contrast, Theorem 3.1 demonstrates that the Khatri-Rao product of the same randomly + +initialized matrices $\mathbf{U}$ and $\mathbf{V}$ almost surely achieves full column rank, thereby providing a mechanism to construct high-rank matrices in a parameter-efficient manner. + +# 3.5. Differences with Kronecker Adapters + +Since the Khatri-Rao product can be seen as column-wise Kronecker product, our proposed KRAdapter and Kronecker Adapters (KronA) [16] bear some similarity. However, due to the different constructions, the resulting properties differ significantly. Specifically, we find that the Khatri-Rao product leads to different spectral properties in the resultant matrix, which displays significantly higher tail singular values. We find this to be a desirable property for PEFT methods as it 056 increases the representation power when modelling various types of weight matrices (Figure 1) and also improves OOD generalization and robustness in language and vision tasks (Table 2 and 3 3) Further discussion supported by empirical differences is available in Appendix B. + +# 4. Experiments + +# 4.1. Matrix approximation + +To isolate and understand the inherent strengths and weaknesses of KRAAdapter compared to LoRA and other full-rank PEFT methods, we design a series of controlled experiments using synthetic weight matrices with distinct structural properties. + +We design six matrix patterns to investigate parameter-efficient fine-tuning algorithms. The normally distributed random matrix acts as a high-rank baseline, evaluating general approximation performance. The random $90\%$ sparse matrix simulates scenarios where critical pre-trained parameters should remain unchanged. The PCA-whitened random matrix assesses algorithms' ability to handle highly decorrelated representations. The low-rank random matrix, constructed by keeping only the top quartile of singular values, tests full-rank algorithms' capacity to model low-rank structures. The CLIP ImageNet fine-tuned weight delta (Vision or Language), obtained by the element-wise difference between the pre-trained CLIP-ViT-L/14 weights and the weights obtained after standard fine-tuning on ImageNet (also known as task vector [63]), represents a realistic target weight for LoRA in transformer in real-world fine-tuning. The High/low frequency features, generated using superposed sinusoidal functions with varying frequencies ([1000, 10000] Hz and [1, 100] Hz respectively), assess algorithms' bias towards feature frequencies. + +Methodology For each target matrix pattern, we train the PEFT algorithms to minimize the mean squared error (MSE) between the estimated matrix and the target matrix. We train + +on $1024 \times 768$ matrices which is the dimension of the query part of the attention matrix in ViT-L/14. All algorithms are trained with the same or higher number of trainable parameters than KRAdapter to ensure a fair comparison. We report the absolute nuclear reconstruction error (average absolute element-wise singular value difference) to evaluate each algorithm's capacity to capture the spectrum of the target matrix. Our investigation focuses on comparing with LoRA [24] and its full-rank alternatives, specifically: Sin-LoRA [26], RandLoRA [2], and Krona [16]. Further details are available in appendix C. + +Results We present results in Figure 1 where our analysis yields several critical insights. First, full-rank PEFT algorithms consistently achieve lower approximation errors compared to LoRA across a range of target matrices (low rank matrices excepted). The advantage of high-rank PEFT methods is particularly apparent when approximating matrices with highly de-correlated features such as random, whitened or sparse noise. Results on approximating traditionally fine-tuned weights for the CLIP vision and language backbones are more nuanced as RandLoRA and Krona struggle to emulate the fine-tuned delta. This difference does not translate to poor performance in practice though which indicates these algorithms probably learn to solve tasks in a different manner than standard fine-tuning does. KRAdapter is especially performant at approximating the CLIP fine-tuned deltas as well as the high frequency matrix, demonstrating a high adaptability to learn highly tailored features. We evidence a key limitation of full-rank methods when approximating explicitly low-rank target matrices, as the algorithms did not demonstrate a substantial advantage over LoRA. Further analysis suggests a potential affinity of LoRA towards low-frequency feature representations as full-rank algorithms struggle to improve over LoRA's solution in this case. The low frequency matrix is however also low rank by nature which would favour LoRA's solution. Overall, KRAdapter provides a strong solution to all cases, low rank matrix excluded, which conforms us in its suitability as a parameter-efficient algorithm. Visualization of the approximated matrices is available in Figure 5 in the appendix. + +# 4.2. Vision-language tasks + +# 4.2.1. Common datasets + +We now move on to real data and compare the performance of several full-rank PEFT algorithms when applied to finetuning CLIP models [44] (both vision and language encoders) across a diverse suite of eleven image classification datasets (detailed in Appendix F). We evaluate Vision Transformer-based (ViT) architectures, initialized with publicly available openCLIP weights [9]. The PEFT algorithms are only trained on the attention heads. As an indicative, we + +also report the performance of full fine-tuning (FT), where all parameters of the network are updated. We tune the hyperparameter of each method to achieve optimal performance. + +Experiments are conducted in few-shot settings (1, 4, and 16 shots) and with varying dataset sizes (50% and 100% of the available training data). The training data is exactly the same for each algorithm to ensure fair comparisons. Results are visualized in Figure 2 with detailed results in Table 7 in the appendix. + +Our results confirm the observations of [2] about the importance of full-rank when fine-tuning vision-language models as these algorithms largely outperform LoRA with equal trainable parameters. We also confirm that full-rank PEFT algorithms can improve performance over standard fine-tuning by limiting over-fitting for larger models (ViT-H/14 especially). KRAdapter's enhanced representational capacity translates to a further performance improvements compared to other full-rank PEFT algorithms. This performance advantage is observed across most settings, although we observe a performance saturation for $\geq 50\%$ data settings for the larger ViT-H/14 architecture. KRAdapter's row-wise Kronecker formulation is additionally systematically superior to Krona's direct Kronecker approach. + +# 4.2.2. Specialized datasets + +To further investigate performance on tasks necessitating strong model adaptation, we conduct experiments using the Visual Task Adaptation Benchmark (VTAB1k) [62] (detailed in Appendix F.3). VTAB1k contains 19 datasets grouped into 3 categories. The structured and specialized subsets as especially interesting to us as they contain tasks less likely to be encountered during CLIP's pre-training. The specialized subset focuses on predicting specialist medical or satellite imagery while the structured subset aims to predict object state attributes such as distance, location, count, and rotation. Prior work has shown that CLIP exhibits limitations on these types of tasks [51], making them an ideal benchmark for assessing the enhanced representation learning capacity of full-rank adaptation algorithms. The results are presented in Table 1, where we report average accuracy over the 3 subsets (detailed results are available in appendix F.3.3). We observe that KRAdapter performs especially well on the natural and structured subsets but can struggle to generalize as well as RandLoRA to the structured subset for the larger models. Standard fine-tuning also struggles to generalize to the structured subset for ViT-H/14 which may indicate an implicit regularization in RandLoRA preventing overfitting, thus allowing to perform better in this specific setting. When averaged over the whole benchmark however, KRAdapter performs better across all architectures. + +![](images/cf3170ca8aae358d994f32c504854310e09d8828468a05e6d50dee4983d82a93.jpg) +(a) ViT-B/32 + +![](images/dcc96f8f8caee1d64dfaa0b70ae8db9e412d83663cdd7d4cd7ef01300caa9a5f.jpg) +(b) ViT-L/14 +Figure 2. Tuning CLIP's vision and language backbone for image classification. Accuracy $(\%)$ averaged over 11 datasets. + +![](images/1360c4de6681ab747024bafb69f71f282ace55d4053ab1c065579d836728acf8.jpg) +(c) ViT-H/14 + +
MethodViT-B/32ViT-L/14ViT-H/14
Nat.Spe.Struc.Avg.Nat.Spe.Struc.Avg.Nat.Spe.Struc.Avg.
LoRA70.682.645.266.182.387.054.774.684.186.756.875.9
SinLoRA72.083.153.169.482.586.856.875.484.286.958.276.4
RandLoRA74.282.453.169.982.887.058.376.083.987.359.576.9
Krona73.783.453.370.183.487.457.576.184.986.758.076.6
KRAdapter76.084.053.371.184.886.957.276.385.887.058.177.0
FT75.783.553.170.883.287.458.376.385.484.843.771.3
+ +Table 1. Adaptation performance on VTAB1k's structured datasets + +# 4.2.3. Out-of-Distribution Generalization + +We expect that the balanced singular value distribution of KRAdapter will allow for better out-of-distribution (OOD) robustness at test time. We thus propose to investigate whether parameter-efficient adaptations, optimized on in-distribution (ID) data, exhibit robustness and transferability to out-of-distribution (OOD) scenarios. We utilize a suite of established OOD benchmark datasets: ImageNet-A [23], ImageNet-Sketch [56], ImageNet-R [22], ImageNet-V2 [46], and CIFAR-100 [30]. These datasets encompass diverse distribution shifts: adversarial, stylized, renditions and domain shift, details in appendix F.4. To quantify generalization, we introduce the ratio $r_{gen} = \frac{\Delta_{odd}}{\Delta_{id}}$ , measuring the relative OOD accuracy gain ( $\Delta_{odd}$ , averaged over OOD datasets) to ID gain ( $\Delta_{id}$ , ImageNet validation) over the zero-shot model. Table 2 reports the results where we fine-tune the attentions layers of the vision backbone of CLIP only, maintaining frozen language embeddings to classify. KRAdapter consistently achieves a higher $r_{gen}$ than other PEFT methods. We hypothesize KRAdapter's spectral properties enable it to efficiently work in the parameter space without overfitting to a subset of directions, unlike rank-constrained methods prone to overfitting dominant ID features. To validate this hypothesis and quantify how much the algorithms stray from the pre-trained weights, we evaluate the average Nuclear and Frobenius norm of the learned updates over the attention lay + +ers. We find that KRAdapter produces updates with smaller norms than other PEFT algorithms which is in line with the very small norm of the traditionally fine-tuned model and has been previously reported to be beneficial for OOD generalization [33, 59]. Table 4 in the appendix further reports the effective rank obtained when fine-tuning the vision-language architectures where we observe that KRAdapter systematically leads to higher effective ranks. + +# 4.3. Commonsense reasoning + +We further investigate resource-efficient fine-tuning of Large Language Models (LLMs) for commonsense reasoning through 4-bit quantization-aware training and evaluation. We evaluate this approach on Llama3.1-8B and Qwen2.5-7B, which are well suited to consumer grade GPUs with 24GB VRAM. We fine-tune the key and value projection matrices within the attention layers, optimizing hyper-parameters for each algorithm to maximize performance. To specifically evaluate out-of-distribution (OOD) generalization, we slightly modify the commonsense reasoning benchmark established in previous research [25, 36]. While the standard commonsense benchmark trains and tests on the same tasks with identical question and answer formats, we train on the multi-choice questions with "answer{1...5}" format (Science-QA, ARC, and OpenBookQA) and subsequently evaluate on: (1) in-distribution test sets from these training + +
ViT-B/32ViT-L/14ViT-H/14
IDOODrgen↑Nuc↓Fro↓IDOODrgen↑Nuc↓Fro↓IDOODrgen↑Nuc↓Fro↓
Zero-shot62.655.5n/an/an/a75.474.1n/an/an/a77.975.7n/an/an/a
LoRA72.958.20.2719.64.183.377.20.3946.26.683.775.90.0470.07.4
SinLoRA72.858.60.31223.54.782.876.30.3061.67.183.677.10.2590.68.1
RandLoRA73.058.40.2846.25.782.876.20.29159.911.882.975.4-0.05508.321.2
Krona72.458.60.3244.14.984.177.80.4242.75.085.078.20.36123.39.2
KRAdapter72.559.90.457.32.383.678.20.499.82.884.778.80.4832.85.5
FT75.558.90.260.70.885.176.70.271.11.085.578.60.394.31.9
+ +Table 2. Generalization to distribution shift. We report ID accuracy on ImageNet, OOD accuracy averaged over 7 OOD datasets, the $r_{gen}$ ratio and the nuclear and frobenius norm of the updates. The $\uparrow$ and $\downarrow$ indicates if higher or lower scores are better. + +
MethodSIQAArcEArcCOBQAHellaSBoolQPIQAWinoGIDNIDOODAvg.
LLama3.1-8B
Zero-shot20.7326.1822.3516.0025.6943.4645.8717.6021.3225.6935.6427.23
LoRA79.0787.1274.6683.4053.7042.3576.6149.8081.0653.7055.6268.34
DoRA79.7988.2277.5684.4052.1641.7779.8757.3882.4952.1657.8070.14
SinLoRA76.4693.1486.8688.4051.5743.9484.7746.5786.2151.5758.4371.46
RandLoRA79.7987.6776.4582.6057.2252.5179.2254.3881.6357.2260.8371.23
Krona79.1787.9675.4382.4050.0759.1479.6056.6781.2450.0761.3771.31
KRAdapter79.2790.3278.1682.4055.3458.2380.9054.8582.5455.3464.6672.43
Qwen2.5-7B
Zero-shot5.025.184.104.801.7931.444.7327.704.771.7921.2910.59
LoRA51.1392.1385.8488.4071.0143.5583.7940.8179.3771.0159.7969.58
DoRA77.9487.9279.5287.8077.9557.2285.6461.6483.3077.9570.6176.95
SinLoRA77.3388.9374.8385.4058.6135.2681.3454.7881.6258.6157.1269.56
RandLoRA79.7994.8787.6389.0078.2047.4085.6960.8587.8278.2068.0477.93
Krona78.3093.8188.1491.6080.9349.0585.4257.5487.9680.9368.2378.10
KRAdapter79.7994.6186.9589.8075.3162.1484.6661.0187.7975.3170.7879.28
+ +Table 3. 4bit quantized LLama3.1-8B and Qwen2.5-7B on commonsense reasoning tasks. We finetune the Key and Value matrices of the attention layers. We evaluate on in-distribution (ID), near in-distribution (NID) and out-of-distribution (OOD) test sets and bold the best results. + +datasets; (2) near in-distribution HellaSwag test set (multichoice, "ending{1...4}" format); and (3) out-of-distribution BoolQ, PiQA, and WinoGrande test sets (binary answers). This evaluation framework is designed to comprehensively assess the generalization capabilities of each PEFT algorithm, a crucial attribute for deploying robust LLMs across diverse real-world scenarios. Further details about our commonsense benchmark are available in appendix G. + +Table 3 presents a comparative analysis of the PEFT methods. The results demonstrate a clear distinction: rank-restricted algorithms, such as LoRA, exhibit limited generalization to out-of-distribution (OOD) datasets. In contrast, full-rank algorithms significantly enhance the zero-shot model to answer these OOD tasks despite not specifically training for them. Notably, KRAdapter outperforms state-of-the-art methods across architectures. These findings support the + +conclusions of Section 4.2.3, reinforcing evidence of the efficacy of KRAdapter in learning generalizable representations for language models. + +# 4.4. Ablation study + +We are interested here in challenging our design choices for KRAdapter and evaluating the impact of scaling the number of trainable parameters further. We study 3 scenario named low, medium and high where we iteratively double the amount of trainable parameters from the base (low) configuration used in other experiments. In KRAdapter, this is achieved by setting $k_{1} = \{d_{out}^{\frac{1}{2}}, d_{out}^{\frac{1}{3}}, d_{out}^{\frac{1}{4}}\}$ . Ideally, performance should increase with every parameter increase to suggest favorable scaling laws. Results are reported in Figure 3 for the vision-language task. We find that for smaller + +![](images/ffcdaa1fbc5e7936a6d4d7fba5b3dce229bfc929f36ca3b853f3e2333c5c0dc6.jpg) +(a) ViT-B/32 + +![](images/b15d37b5d3e44bd7f27aa64a105632e4177aa3606dca052bbd5a96d94cc4d8eb.jpg) +(b) ViT-L/14 +Figure 3. Increasing the number of trainable parameters in PEFT algorithms. Accuracy $(\%)$ averaged over 11 datasets. + +![](images/1dce14862ad3f29d2bd0c6b2d066286bd0e440bdd1c8960e2880e32ac5d1b50c.jpg) +(c) ViT-H/14 + +models, full-rank methods scale better with increases in parameter budgets whereas LoRA saturates faster. For larger models however, performance does not increase beyond the low configuration which ties in with the observed saturation in larger models of Figure 2. This aligns with findings of Albert et al. [2]. SinLoRA is a notable exception as more parameters leads to a decrease in accuracy. For ViT-H/14, KRAAdapter is the only full rank PEFT algorithm to not drop in accuracy with increasing budget. This highlight the robustness of KRAAdapter to over-fitting which helps it in achieving good OOD generalization. + +# 5. Limitations + +Although we have shown KRAdapter is a good alternative to LoRA when high effective rank or out-of-distribution generalization is important, we discuss here some limitations of our approach. First KRAdapter in its most parameter-efficient configuration trains more parameters than the most efficient LoRA configurations do. In practice we find that KRAdatper is equivalent to a rank 16 to 32 LoRA, a commonly used configuration. Exploring low-rank approximations for the matrices $U$ and $V$ prior to the Khatri-Rao product could potentially mitigate this, though the theoretical implications for the derived full-rank properties in Section 3.4 require rigorous investigation. Second, empirical evidence from controlled matrix estimation experiments indicates that KRAdapter's performance may be suboptimal when estimating matrices of extremely low rank. In scenarios demanding strict subspace constraints on weight updates, such as continual learning paradigms, LoRA might exhibit superior suitability. Finally, while KRAdapter consistently demonstrates robust out-of-distribution generalization and faster training convergence, we observe instances with larger models where RandLoRA achieves marginally superior performance on purely in-distribution test sets. It is important to note that KRAdapter's in-distribution performance remains highly competitive in these cases, and its advantages in gen + +eralization and training efficiency compared to RandLoRA generally outweigh this minor performance delta. + +# 6. Conclusion + +This paper introduces Khatri-Rao Adapters (KRAdapter), a novel parameter-efficient fine-tuning technique designed to overcome the representation limitations of low-rank methods. KRAdapter generate full-rank update matrices with demonstrably higher effective rank, enabling to capture more complex feature interactions and improving OOD generalization by staying close to the pre-trained weights. KRAdapter achieves all this without sacrificing parameter efficiency and training time. Our comprehensive experimental evaluation across vision-language, language understanding, and commonsense reasoning tasks consistently demonstrates KRAdapter produces improved results compared to LoRA and other full-rank PEFT alternatives, especially for specialized VTAB1k tasks that demand nuanced feature adaptation and for larger models that are prone to over-fitting. We specifically highlight that KRAdapter's improved representation space allows to produce representations closer to the zero-shot model that lead to better generalization under distribution shifts compared to other PEFT algorithms. Crucially, KRAdapter maintains the computational advantages of LoRA in terms of training speed and memory footprint, offering a practical and effective solution for fine-tuning billion-scale models as demonstrated in the commonsense experiments. A notable limitation of KRAdapter, however, is that the minimum number of trainable parameter is larger than the most efficient LoRA configurations which currently renders it unadapted to extreme scarcity scenarios. Future work could investigate replacing $U$ and $V$ with low-rank formulations to achieve better efficiency. We however anticipate this formulation would change the efficient rank guarantees of KRAdapter. Our findings underscore the importance of high effective rank weight updates in parameter-efficient adaptation of pre-trained models across diverse downstream applications. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv:2303.08774, 2023. 1, 2 +[2] Paul Albert, Frederic Z. Zhang, Hemanth Saratchandran, Cristian Rodriguez-Opazo, Anton van den Hengel, and Ehsan Abbasnejad. RandLoRA: Full rank parameter-efficient finetuning of large models. In International Conference on Learning Representations (ICLR), 2025. 1, 2, 3, 5, 8, 13, 16, 21 +[3] Charles Beattie, Joel Z Leibo, Denis Teptyashin, Tom Ward, Marcus Wainwright, Heinrich Kuttler, Andrew Lefrancq, Simon Green, Víctor Valdés, Amir Sadik, et al. Deepmind lab. arXiv: 1612.03801, 2016. 18 +[4] Daniel J. Beutel, Taner Topal, Akhil Mathur, Xinchi Qiu, Titouan Parcollet, and Nicholas D. Lane. Flower: A Friendly Federated Learning Research Framework. arxiv :2007.14390, 2020. 18 +[5] Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. Piaq: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on Artificial Intelligence (AAAI), 2020. 20 +[6] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International Conference on Machine Learning (ICML), 2020. 18 +[7] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International Conference on Machine Learning (ICML), 2019. 1 +[8] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE, 2017. 18 +[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5 +[10] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing Textures in the Wild. In Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 17 +[11] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. BoolQ: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. 20 +[12] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. 21 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 1, 2 + +[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 2 +[15] Emma Dugas, Jorge Jared, and Will Cukierski. Diabetic retinopathy detection (2015). URL https://kaggle.com/competitions/diabetic-retinopathy-detection, 2015. 18 +[16] Ali Edalati, Marzieh Tahaei, Ivan Kobyzev, Vahid Partovi Nia, James J Clark, and Mehdi Rezagholizadeh. Krona: Parameter efficient tuning with kronecker adapter. In Advances in Neural Information Processing Systems Workshops (NeurIPSW), 2023. 2, 3, 4, 5 +[17] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. Vision meets Robotics: The KITTI Dataset. International Journal of Robotics Research (IJRR), 2013. 18 +[18] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 1, 2 +[19] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. In International Conference on Machine Learning (ICML), 2024. 2 +[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, 2015. 4 +[21] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Introducing EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. In International Geoscience and Remote Sensing Symposium (IGARSS), 2018. 18 +[22] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, 2021. 6, 18 +[23] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15262-15271, 2021. 6, 18 +[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations (ICLR), 2022. 1, 2, 3, 4, 5 +[25] Zhiqiang Hu, Lei Wang, Yihuai Lan, Wanyu Xu, Ee-Peng Lim, Lidong Bing, Xing Xu, Soujanya Poria, and Roy Ka-Wei Lee. Llm-adapters: An adapter family for parameter-efficient fine-tuning of large language models. arXiv preprint arXiv:2304.01933, 2023. 6, 21 +[26] Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Sine Activated Low-Rank Matrices for Parameter Efficient Learning. arXiv:2403.19243, 2024. 1, 2, 3, 5 +[27] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A + +diagnostic dataset for compositional language and elementary visual reasoning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 18 +[28] Soroush Abbasi Koohpayegani, KL Navaneet, Parsa Nooralinejad, Soheil Kolouri, and Hamed Pirsiavash. NOLA: Compressing LoRA using Linear Combination of Random Basis. In International Conference on Learning Representations (ICLR), 2024. 3 +[29] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki Markus Asano. Vera: Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. 3 +[30] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 6, 17 +[31] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (NeurIPS), 2012. 18 +[32] Yann LeCun, Fu Jie Huang, and Leon Bottou. Learning methods for generic object recognition with invariance to pose and lighting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2004. 18 +[33] Alex Lewandowski, Michal Bortkiewicz, Saurabh Kumar, András György, Dale Schuurmans, Mateusz Ostaszewski, and Marlos C Machado. Learning continually by spectral regularization. In International Conference on Learning Representations (ICLR), 2025. 6 +[34] Fei-Fei Li, Marco Andreeto, Marc'Aurelio Ranzato, and Pietro Perona. Caltech 101, 2022. 17 +[35] Vijay Lingam, Atula Tejaswi, Aditya Vavre, Aneesh Shetty, Gautham Krishna Gudur, Joydeep Ghosh, Alex Dimakis, Eun-sol Choi, Aleksandar Bojchevski, and Sujay Sanghavi. SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors. In International Conference on Machine Learning Workshops (ICMLW), 2024. 3 +[36] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. In International Conference on Machine Learning (ICML), 2024. 2, 4, 6 +[37] Y Liu, M Ott, N Goyal, J Du, M Joshi, D Chen, O Levy, M Lewis, L Zettlemoyer, and V Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach. arXiv:1907.11692, 2019. 1, 2, 22 +[38] Loic Matthew, Irina Higgins, Demis Hassabis, and Alexander Lerchner. dSprites: Disentanglement testingSprites dataset. https://github.com/deepmind/dSprites-dataset/, 2017. 18 +[39] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. Pissa: Principal singular values and singular vectors adaptation of large language models. Advances in Neural Information Processing Systems (NeurIPS), 2024. 3 +[40] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. 21 + +[41] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 2 +[42] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning (NeurIPSW), 2011. 18 +[43] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 18 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2, 5 +[45] Maithra Raghu, Ben Poole, Jon Kleinberg, Surya Ganguli, and Jascha Sohl-Dickstein. On the expressive power of deep neural networks. In International Conference on Machine Learning (ICML), 2017. 2 +[46] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning (ICML), pages 5389-5400. PMLR, 2019. 6, 18 +[47] Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In European Signal Processing Conference, 2007. 1, 3, 16 +[48] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 2021. 20 +[49] Maarten Sap, Hannah Rashkin, Derek Chen, Ronan LeBras, and Yejin Choi. Socialiaq: Commonsense reasoning about social interactions. arXiv preprint arXiv:1904.09728, 2019. 20 +[50] Houcheng Su, Daixian Liu, Mengzhu Wang, and Wei Wang. Singular Value Penalization and Semantic Data Augmentation for Fully Test-Time Adaptation. In Association for the Advancement of Artificial Intelligence (AAAI), 2024. 1 +[51] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. In Association for Computational Linguistics (ACL), 2022. 5 +[52] Chunlin Tian, Zhan Shi, Zhijiang Guo, Li Li, and Chengzhong Xu. HydraLoRA: An Asymmetric LoRA Architecture for Efficient Fine-Tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2 +[53] Bastiaan S Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling. Rotation Equivariant CNNs for Digital Pathology. arxiv: 1806.03962, 2018. 18 +[54] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: multi-task benchmark and analysis platform for natural language understanding. In International Conference on Learning Representations (ICLR), 2019. 22 + +[55] Chenglong Wang, Jiangyan Yi, Xiaohui Zhang, Jianhua Tao, Le Xu, and Ruibo Fu. Low-rank adaptation method for wav2vec2-based fake audio detection. In International Joint Conference on Artificial Intelligence Workshops (IJCAIW), 2023. 2 +[56] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems (NeurIPS), pages 10506-10518, 2019. 6, 18 +[57] Hanqing Wang, Yixia Li, Shuo Wang, Guanhua Chen, and Yun Chen. Milora: Harnessing minor singular components for parameter-efficient llm finetuning. arXiv:2406.09044, 2024.3 +[58] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 18 +[59] Yuichi Yoshida and Takeru Miyato. Spectral norm regularization for improving the generalizability of deep learning. arXiv:1705.10941, 2017. 2, 6 +[60] Maxime Zanella and Ismail Ben Ayed. Low-Rank Few-Shot Adaptation of Vision-Language Models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[61] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. 20 +[62] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv: 1910.04867, 2019. 5, 17 +[63] Frederic Z Zhang, Paul Albert, Cristian Rodriguez-Opazo, Anton van den Hengel, and Ehsan Abbasnejad. Knowledge Composition using Task Vectors with Learned Anisotropic Scaling. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 4, 15, 16 +[64] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. AdaLoRA: Adaptive budget allocation for parameter-efficient fine-tuning. In International Conference on Learning Representations (ICLR), 2023. 4 +[65] Ming Zhong, Yelong Shen, Shuohang Wang, Yadong Lu, Yizhu Jiao, Siru Ouyang, Donghan Yu, Jiawei Han, and Weizhu Chen. Multi-lora composition for image generation. Transactions on Machine Learning Research (TMLR), 2024. 2 \ No newline at end of file diff --git a/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/images.zip b/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e20e43a767ad790d74b353b8e85157514183f997 --- /dev/null +++ b/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:077c77c4958fdabc1ad5e0ce0f0bd2130c1c2dbae7f59e91ee9ecbd5b6cba509 +size 483032 diff --git a/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/layout.json b/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3842dd7ff6e2ed36f02804e1f1f32ece7601af9e --- /dev/null +++ b/2025/Towards Higher Effective Rank in Parameter-Efficient Fine-tuning using Khatri-Rao Product/layout.json @@ -0,0 +1,8376 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 102, + 510, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 102, + 510, + 138 + ], + "spans": [ + { + "bbox": [ + 100, + 102, + 510, + 138 + ], + "type": "text", + "content": "Towards Higher Effective Rank in Parameter-efficient Fine-tuning using Khatri-Rao Product" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "spans": [ + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "content": "Paul Albert" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "content": " Frederic Z. Zhang" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "content": " Hemanth Saratchandran" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "content": " Anton van den Hengel" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "text", + "content": " Ehsan Abbasnejad" + }, + { + "bbox": [ + 153, + 160, + 457, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 190, + 541, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 190, + 541, + 217 + ], + "spans": [ + { + "bbox": [ + 69, + 190, + 541, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 190, + 541, + 217 + ], + "type": "text", + "content": "Amazon Machine Learning, Australia " + }, + { + "bbox": [ + 69, + 190, + 541, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 190, + 541, + 217 + ], + "type": "text", + "content": "Australian Institute for Machine Learning " + }, + { + "bbox": [ + 69, + 190, + 541, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 190, + 541, + 217 + ], + "type": "text", + "content": "Monash University {albrtpa, fredzz}@amazon.com https://github.com/PaulAlbert31/KRAdapter" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 275, + 297, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 297, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 297, + 550 + ], + "type": "text", + "content": "Parameter-efficient fine-tuning (PEFT) has become a standard approach for adapting large pre-trained models. Amongst PEFT methods, low-rank adaptation (LoRA) has achieved notable success. However, recent studies have highlighted its limitations compared against full-rank alternatives, particularly when applied to multimodal and large language models. In this work, we present a quantitative comparison amongst full-rank and low-rank PEFT methods using a synthetic matrix approximation benchmark with controlled spectral properties. Our results confirm that LoRA struggles to approximate matrices with relatively flat spectrums or high frequency components—signs of high effective ranks. To this end, we introduce KRAdapter, a novel PEFT algorithm that leverages the Khatri-Rao product to produce weight updates, which, by construction, tends to produce matrix product with a high effective rank. We demonstrate performance gains with KRAdapter on vision-language models up to 1B parameters and on large language models up to 8B parameters, particularly on unseen common-sense reasoning tasks. In addition, KRAdapter maintains the memory and compute efficiency of LoRA, making it a practical and robust alternative to fine-tune billion-scale parameter models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 581, + 135, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 135, + 594 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 135, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 606, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 296, + 693 + ], + "type": "text", + "content": "Large pre-trained models have lead to great success in both computer vision and natural language processing [1, 13, 14, 37, 44]. However, fine-tuning these models for specific downstream tasks often demands substantial computational resources due to the sheer number of parameters. Recently, parameter-efficient fine-tuning (PEFT) has emerged as an effective strategy to address this challenge, enabling effi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 246, + 556, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 556, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 556, + 474 + ], + "type": "text", + "content": "cient adaptation using a fraction of the trainable parameters [18]. In particular, low-rank adaptation (LoRA) [24], a prominent PEFT technique, formulates the learnable weight updates as the product of two low-rank matrices, achieving remarkable memory efficiency with little performance drop. Despite its success, recent studies [2, 26] suggest that the low-rank formulation can limit its ability to effectively fine-tune models for complex tasks. To shed light on this limitation, we compare LoRA to several full-rank PEFT methods by approximating weight matrices with controlled spectral properties. As shown in Figure 1, full-rank methods can outperform LoRA in approximating both synthetic and real weight matrices. Not all full-rank PEFT methods are equally effective—some produce approximations with rapidly decaying singular values, leading to a low effective rank [47]. Building on prior work [7, 50], which links substantial tail singular values to improved generalization and robustness, we hypothesize that a high effective rank is crucial to the performance of full-rank PEFT methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 478, + 557, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 557, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 557, + 708 + ], + "type": "text", + "content": "To this end, we introduce KRAdapter, a novel PEFT approach that leverages the Khatri-Rao matrix product. By design, KRAdapter constructs provably full-rank weight matrices and empirically maintains large tail singular values. The increased effective rank indicates that the weight updates span the parameter space more uniformly, allowing the model to capture complex patterns and generalize better to distribution shifts. Across diverse architectures, including vision-language models and large language models, KRAdapter achieves performance gains while maintaining parameter efficiency. Furthermore, KRAdapter exhibits superior out-of-distribution (OOD) performance compared to LoRA and other full-rank methods. Our key contributions are: (1) a controlled matrix approximation benchmark comparing the representational advantages of full- and rank-constrained PEFT algorithms; (2) the identification of effective rank as a critical factor differentiating full-rank PEFT techniques; (3) KRAdapter, a novel Khatri-Rao product-based PEFT method that consistently achieves high effective" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 265, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 265, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 265, + 714 + ], + "type": "text", + "content": "*Work done at the Australian Institute for Machine Learning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1292" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 289 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 289 + ], + "type": "image", + "image_path": "9f5d53b3f50b7938274151752709074ce2d67af77e39f571cf30f1a0a4834227.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 294, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 54, + 294, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 555, + 361 + ], + "type": "text", + "content": "Figure 1. A visualization of the capacity of various PEFT methods to approximate the spectrum of different types of weight matrices. The first row shows the structure of the matrices. In the second row we plot the squared nuclear error as a percentage of that of the LoRA solution, and add the absolute value above each bar. Lower is better. The third row shows the singular value distribution of the solutions compared to the target (square root scale). A flatter distribution indicates a higher effective rank. We compare LoRA [24], SinLoRA [26], Krona [16], RandLoRA [2] and our proposed algorithm KRAdapter. KRAdapter achieves a better approximation to the true matrix in every case except when we force it to have low rank (and thus to match the LoRA model)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 382, + 298, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 382, + 298, + 431 + ], + "spans": [ + { + "bbox": [ + 54, + 382, + 298, + 431 + ], + "type": "text", + "content": "rank; (4) empirical validation of its superior performance and out-of-distribution robustness compared to LoRA and other PEFT methods; and (5) evidence of KRAAdapter's improved parameter scaling efficiency among full-rank approaches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 447, + 142, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 142, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 142, + 459 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 473, + 298, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 473, + 298, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 473, + 298, + 568 + ], + "type": "text", + "content": "The development of large pre-trained models has revolutionized various fields, yet their deployment and fine-tuning pose significant computational challenges [1, 13, 14, 37, 44]. Parameter-Efficient Fine-Tuning (PEFT) methods have emerged as a crucial paradigm to address these challenges, allowing adaptation of these massive models to downstream tasks while training only a small fraction of the parameters [18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 223, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 223, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 223, + 596 + ], + "type": "text", + "content": "2.1. Low-Rank Adaptation (LoRA)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 297, + 715 + ], + "type": "text", + "content": "LoRA [24] has become a cornerstone of PEFT due to its simplicity and effectiveness. LoRA introduces low-rank matrices to approximate the weight updates, significantly reducing the number of trainable parameters. Its success has led to numerous extensions and applications across various modalities [55, 60, 63, 65]. However, recent studies have begun to explore the limitations of the low-rank constraint, suggesting that it might hinder the model's ability to capture complex relationships, especially for demanding tasks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 382, + 556, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 556, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 556, + 478 + ], + "type": "text", + "content": "[2, 16, 26]. Furthermore, the spectral properties of weight matrices, which influence a network's capacity for learning and generalization [41, 45, 59], may be constrained by the low-rank nature of LoRA updates. Our work builds upon these observations, confirming the need for methods that can capture richer feature interactions while maintaining parameter efficiency, potentially through weight updates with more desirable spectral characteristics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 495, + 492, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 492, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 492, + 508 + ], + "type": "text", + "content": "2.2. Enhancing the training efficiency" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 517, + 557, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 557, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 557, + 649 + ], + "type": "text", + "content": "To improve convergence speed and overall performance of LoRA, several improvements have been proposed. LoRA+ [19] enhances the training process by applying different learning rates to the low-rank matrices, allowing for more nuanced optimization. DoRA [36] decomposes the low-rank updates into magnitude and direction components, providing finer-grained control over the adaptation process. HydraLoRA [52] employs multiple upscaling matrices, acting as specialized \"experts\" focusing on distinct aspects of the input, potentially capturing more diverse feature interactions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 666, + 474, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 474, + 680 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 474, + 680 + ], + "type": "text", + "content": "2.3. Further reducing parameters" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "While LoRA is inherently parameter-efficient, subsequent research has focused on pushing these limits further. Strategies" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 758 + ], + "type": "text", + "content": "1293" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 298, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 298, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 298, + 204 + ], + "type": "text", + "content": "to minimize the trainable parameter count beyond the rank-one case include methods leveraging linear combinations of fixed random matrices, such as VeRA [29] and NoLA [28], where learned one-dimensional vectors combine pre-defined non-trainable random matrices. Another line of work explores initializing or fixing LoRA's projection matrices using insights from the pre-trained weights. Methods like SVFT [35] and Pissa [39] fix or respectively initialize the projection matrices to the first eigenvectors corresponding to the largest singular values of the original weight matrix. Conversely, MiLoRA [57] utilizes the last eigenvectors." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 163, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 163, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 163, + 229 + ], + "type": "text", + "content": "2.4. Full-rank updates" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 240, + 298, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 298, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 298, + 432 + ], + "type": "text", + "content": "To address the representational bottlenecks and potential spectral limitations inherent in low-rank approximations, several methods explore parameter-efficient ways to perform full-rank updates. The goal is to capture more complex relationships present in the data without incurring the full cost of fine-tuning all parameters. Kronecker Adapters (KronA) [16] utilize the Kronecker product to construct full-rank update matrices from smaller trainable factors. SinLoRA [26] applies a sine function parameterized by a frequency parameter on top of a low-rank update, effectively producing a full-rank update. RandLoRA [2] employs parameter-efficient random matrix combinations to generate full-rank matrices. While these methods theoretically produce full rank updates, we observe their effective rank [47] is usually low (i.e. the tail singular values are very small), suggesting they might not fully capture the desired spectral properties." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 134, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 134, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 134, + 456 + ], + "type": "text", + "content": "2.5. Motivations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 622 + ], + "type": "text", + "content": "Our analysis of existing work indicates that the low-rank update in LoRA can limit performance on challenging tasks and with complex architectures. While recent advancements aim to overcome this by generating theoretically full-rank updates via random basis combinations, Kronecker products, or sine activations, we question whether this theoretical rank consistently translates to high effective rank. We hypothesize that the Khatri-Rao product offers an alternative method for constructing full-rank update matrices that demonstrably achieve higher effective rank. This increased effective rank would enable to learn more complex feature representations, leading to improved performance across diverse tasks and architectures." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 639, + 252, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 252, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 252, + 653 + ], + "type": "text", + "content": "3. Khatri-Rao Adapters (KRAdapter)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": "This section details the formulation of the proposed Khatri-Rao Adapters (KRAdapter) for parameter-efficient finetuning. We begin by briefly discussing Low-Rank Adaptation (LoRA) before introducing the KRAdapter mechanism" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 470, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 470, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 470, + 84 + ], + "type": "text", + "content": "and analyzing its parameter efficiency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 98, + 400, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 98, + 400, + 109 + ], + "spans": [ + { + "bbox": [ + 314, + 98, + 400, + 109 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "text", + "content": "Consider a pre-trained linear layer in a deep neural network, characterized by a weight matrix " + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_0\\in \\mathbb{R}^{d_{out}\\times d_{in}}" + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "inline_equation", + "content": "d_{in}" + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "text", + "content": " is the input dimension and " + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "inline_equation", + "content": "d_{out}" + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "text", + "content": " is the output dimension. During standard full fine-tuning, the weight matrix " + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_0" + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "text", + "content": " is updated directly by gradient descent. Given an input " + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in" + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d_{in}}" + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "text", + "content": " , the output " + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{h}" + }, + { + "bbox": [ + 313, + 121, + 555, + 193 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 400, + 211, + 555, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 211, + 555, + 223 + ], + "spans": [ + { + "bbox": [ + 400, + 211, + 555, + 223 + ], + "type": "interline_equation", + "content": "\\mathbf {h} = \\mathbf {W} _ {0} \\mathbf {x} + \\mathbf {b}, \\tag {3.1}", + "image_path": "f757e57c77aef5d0ead96eddbb74e1636f28a3572d797c22514b59d80f33a6d0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 234, + 555, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 234, + 555, + 270 + ], + "type": "inline_equation", + "content": "\\mathbf{b} \\in \\mathbb{R}^{d_{out}}" + }, + { + "bbox": [ + 313, + 234, + 555, + 270 + ], + "type": "text", + "content": " is the bias term (for simplicity, we omit the bias term in subsequent derivations but it can be easily incorporated)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "text", + "content": "Low-Rank Adaptation (LoRA) [24] addresses the inefficiency of full fine-tuning by freezing the pre-trained weights " + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_0" + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "text", + "content": " and introducing a low-rank update. Specifically, a low-rank matrix " + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W} = \\mathbf{BA}" + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "text", + "content": " is added to the original weights, where " + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{r \\times d_{in}}" + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{d_{out} \\times r}" + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "inline_equation", + "content": "r \\ll \\min(d_{in}, d_{out})" + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "text", + "content": " being the rank. During fine-tuning, only the parameters of " + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 276, + 556, + 372 + ], + "type": "text", + "content": " are updated. The output of the LoRA-adapted layer is:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 388, + 390, + 555, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 390, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 388, + 390, + 555, + 402 + ], + "type": "interline_equation", + "content": "\\mathbf {h} = \\left(\\mathbf {W} _ {0} + \\alpha \\mathbf {B A}\\right) \\mathbf {x}, \\tag {3.2}", + "image_path": "043449920b1c92c1d151f7ff4b90566034f2dc806cf1ed7ec12368d325280ff7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 408, + 554, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 433 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 408, + 554, + 433 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 408, + 554, + 433 + ], + "type": "text", + "content": " is a scaling factor that helps in stabilizing training and controlling the magnitude of the adapter." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 445, + 432, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 445, + 432, + 456 + ], + "spans": [ + { + "bbox": [ + 314, + 445, + 432, + 456 + ], + "type": "text", + "content": "3.2. Method formulation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 468, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 529 + ], + "type": "text", + "content": "In contrast to LoRAs, our proposed method, Khatri-Rao Adapters (KRAdapter), achieves parameter efficiency without imposing the low-rank constraint. This is achieved by constructing " + }, + { + "bbox": [ + 313, + 468, + 555, + 529 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}" + }, + { + "bbox": [ + 313, + 468, + 555, + 529 + ], + "type": "text", + "content": " using the Khatri-Rao product, otherwise known as a column-wise Kronecker product." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 556, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 570 + ], + "type": "text", + "content": "Definition 3.1 (Khatri-Rao Product): Given two matrices " + }, + { + "bbox": [ + 313, + 534, + 556, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\in \\mathbb{R}^{a \\times c}" + }, + { + "bbox": [ + 313, + 534, + 556, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 534, + 556, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{b \\times c}" + }, + { + "bbox": [ + 313, + 534, + 556, + 570 + ], + "type": "text", + "content": ", their Khatri-Rao product, denoted by " + }, + { + "bbox": [ + 313, + 534, + 556, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\odot \\mathbf{V}" + }, + { + "bbox": [ + 313, + 534, + 556, + 570 + ], + "type": "text", + "content": ", is a matrix of defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 326, + 578, + 555, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 578, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 326, + 578, + 555, + 618 + ], + "type": "interline_equation", + "content": "\\mathbf {U} \\odot \\mathbf {V} = \\left[ \\begin{array}{c} u _ {1 1} \\mathbf {v} _ {1}, u _ {1 2} \\mathbf {v} _ {2}, \\dots , u _ {1 c} \\mathbf {v} _ {c} \\\\ \\dots \\\\ u _ {a 1} \\mathbf {v} _ {1}, u _ {a 2} \\mathbf {v} _ {2}, \\dots , u _ {a c} \\mathbf {v} _ {c} \\end{array} \\right] \\in \\mathbb {R} ^ {a b \\times c}, \\tag {3.3}", + "image_path": "fab42d98a26706fecd1f549e615f009a0631755d261f49015293eba686f56cab.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "inline_equation", + "content": "u_{ij}" + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "text", + "content": " is the entry in the " + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "text", + "content": "-th row and " + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "text", + "content": "-th column of " + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_j \\in \\mathbb{R}^b" + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "text", + "content": "-th column of " + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 625, + 554, + 650 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "type": "text", + "content": "For a weight update " + }, + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{W}" + }, + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "type": "inline_equation", + "content": "d_{out} \\times d_{in}" + }, + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "type": "text", + "content": ", define " + }, + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\in \\mathbb{R}^{k_1 \\times d_{in}}" + }, + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{k_2 \\times d_{in}}" + }, + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "type": "inline_equation", + "content": "k_1 k_2 = d_{out}" + }, + { + "bbox": [ + 313, + 655, + 555, + 691 + ], + "type": "text", + "content": ". The forward pass of a linear layer with KRAdapter is as follows" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 380, + 701, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 380, + 701, + 555, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {h} = \\left(\\mathbf {W} _ {0} + \\alpha \\mathbf {U} \\odot \\mathbf {V}\\right) \\mathbf {x}, \\tag {3.4}", + "image_path": "02dbc336ddabeb327d5a070db219682369ea5eb9270f130a17e9e251e2c5a870.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1294" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " is a scaling factor similar to LoRA that we typically set to 0.1. Prior to training, " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " is initialized as zeros and " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " using a Kaiming uniform initialization [20] with a negative slope coefficient of " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\sqrt{1 / k_1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " which we empirically find to be the optimal for convergence." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 174, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 174, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 174, + 157 + ], + "type": "text", + "content": "3.3. Parameter efficiency" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 167, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 296, + 216 + ], + "type": "text", + "content": "The number of trainable parameters introduced by KRAdapter is determined by the dimensions of the matrices " + }, + { + "bbox": [ + 55, + 167, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 55, + 167, + 296, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 167, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 167, + 296, + 216 + ], + "type": "text", + "content": ", and the choice of " + }, + { + "bbox": [ + 55, + 167, + 296, + 216 + ], + "type": "inline_equation", + "content": "k_{1}" + }, + { + "bbox": [ + 55, + 167, + 296, + 216 + ], + "type": "text", + "content": " (or " + }, + { + "bbox": [ + 55, + 167, + 296, + 216 + ], + "type": "inline_equation", + "content": "k_{2}" + }, + { + "bbox": [ + 55, + 167, + 296, + 216 + ], + "type": "text", + "content": " resp.). The total number of trainable parameters for KRAdapter is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 223, + 296, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 223, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 134, + 223, + 296, + 236 + ], + "type": "interline_equation", + "content": "N = d _ {i n} \\left(k _ {1} + k _ {2}\\right). \\tag {3.5}", + "image_path": "1652b24302164f662997fba7f6ca8ba9487f9a5e5d0db87f3f9375401efe8297.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "text", + "content": " is minimized when " + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "inline_equation", + "content": "k_{1} = k_{2} = \\sqrt{d_{out}}" + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "text", + "content": ", which we adopt as the default configuration. Details are deferred to Appendix A.1. When " + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "inline_equation", + "content": "d_{out}" + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "text", + "content": " is not a square number, we set " + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "inline_equation", + "content": "k_{1} = \\lfloor \\sqrt{d_{out}} \\rfloor" + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "inline_equation", + "content": "k_{2} = \\lceil \\frac{d_{out}}{k_{1}} \\rceil" + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "inline_equation", + "content": "k_{1}k_{2} \\geq d_{out}" + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "text", + "content": ". The resultant product " + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\odot \\mathbf{V}" + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "text", + "content": " is then truncated to size " + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "inline_equation", + "content": "d_{out} \\times d_{in}" + }, + { + "bbox": [ + 55, + 242, + 296, + 302 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 308, + 296, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 356 + ], + "type": "text", + "content": "KRAdapter achieves the same parameter reduction as LoRA of rank between 16 and 32 depending on matrix shapes, a common setting [24, 36, 64] without incurring any significant increase in computational cost." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 368, + 173, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 368, + 173, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 173, + 380 + ], + "type": "text", + "content": "3.4. Full-rank guarantee" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 391, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 439 + ], + "type": "text", + "content": "This section establishes a full rank guarantee for the Khatri-Rao product when its constituent matrices are randomly initialized. This full rank property, as we will show, is crucial to the improvement of robustness against distribution shifts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": "For brevity of exposition, let us assume " + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "inline_equation", + "content": "k_{1} = k_{2} = k" + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\in \\mathbb{R}^{k \\times d_{in}}" + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{k \\times d_{in}}" + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": ". We assume that each matrix has rank " + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": ". Specifically, we consider the case where the entries of " + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": " are independently and identically distributed (i.i.d.) random variables drawn from either a standard Gaussian distribution " + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, 1)" + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": " or a uniform distribution on " + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "inline_equation", + "content": "[-\\delta, \\delta]" + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": " for some " + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\delta > 0" + }, + { + "bbox": [ + 55, + 445, + 296, + 529 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 533, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 296, + 594 + ], + "type": "text", + "content": "Theorem 3.1. Let " + }, + { + "bbox": [ + 54, + 533, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\in \\mathbb{R}^{k \\times d_{in}}" + }, + { + "bbox": [ + 54, + 533, + 296, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 533, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{V} \\in \\mathbb{R}^{k \\times d_{in}}" + }, + { + "bbox": [ + 54, + 533, + 296, + 594 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 54, + 533, + 296, + 594 + ], + "type": "inline_equation", + "content": "k \\leq d_{in} \\leq k^2" + }, + { + "bbox": [ + 54, + 533, + 296, + 594 + ], + "type": "text", + "content": " be matrices whose entries are chosen i.i.d. from a standard Gaussian or uniform distribution. Then, the Khatri-Rao product " + }, + { + "bbox": [ + 54, + 533, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{U} \\odot \\mathbf{V}" + }, + { + "bbox": [ + 54, + 533, + 296, + 594 + ], + "type": "text", + "content": " has full column rank almost surely, i.e.," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 130, + 594, + 219, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 594, + 219, + 607 + ], + "spans": [ + { + "bbox": [ + 130, + 594, + 219, + 607 + ], + "type": "interline_equation", + "content": "r a n k (\\mathbf {U} \\odot \\mathbf {V}) = d _ {i n}.", + "image_path": "da26cf1039b3d367b5d0d195e2d1ce439e50cb5d0dab70510af5fc7e0c07b746.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "The detailed proof of Theorem 3.1 is provided in Appendix A.2. Crucially, this theorem highlights a significant distinction from low-rank adaptation (LoRA) techniques. When employing LoRA updates to matrices " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": ", the resulting matrix, even after combination, can have a maximum rank of at most " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "k \\leq d_{in}" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": ", thus inherently producing rank-deficient matrices. In contrast, Theorem 3.1 demonstrates that the Khatri-Rao product of the same randomly" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "initialized matrices " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{U}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " almost surely achieves full column rank, thereby providing a mechanism to construct high-rank matrices in a parameter-efficient manner." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 122, + 512, + 136 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 512, + 136 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 512, + 136 + ], + "type": "text", + "content": "3.5. Differences with Kronecker Adapters" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 146, + 555, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 146, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 146, + 555, + 301 + ], + "type": "text", + "content": "Since the Khatri-Rao product can be seen as column-wise Kronecker product, our proposed KRAdapter and Kronecker Adapters (KronA) [16] bear some similarity. However, due to the different constructions, the resulting properties differ significantly. Specifically, we find that the Khatri-Rao product leads to different spectral properties in the resultant matrix, which displays significantly higher tail singular values. We find this to be a desirable property for PEFT methods as it 056 increases the representation power when modelling various types of weight matrices (Figure 1) and also improves OOD generalization and robustness in language and vision tasks (Table 2 and 3 3) Further discussion supported by empirical differences is available in Appendix B." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 319, + 395, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 395, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 395, + 332 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 344, + 441, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 441, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 441, + 357 + ], + "type": "text", + "content": "4.1. Matrix approximation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 367, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 426 + ], + "type": "text", + "content": "To isolate and understand the inherent strengths and weaknesses of KRAAdapter compared to LoRA and other full-rank PEFT methods, we design a series of controlled experiments using synthetic weight matrices with distinct structural properties." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 433, + 556, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 556, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 556, + 672 + ], + "type": "text", + "content": "We design six matrix patterns to investigate parameter-efficient fine-tuning algorithms. The normally distributed random matrix acts as a high-rank baseline, evaluating general approximation performance. The random " + }, + { + "bbox": [ + 313, + 433, + 556, + 672 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 433, + 556, + 672 + ], + "type": "text", + "content": " sparse matrix simulates scenarios where critical pre-trained parameters should remain unchanged. The PCA-whitened random matrix assesses algorithms' ability to handle highly decorrelated representations. The low-rank random matrix, constructed by keeping only the top quartile of singular values, tests full-rank algorithms' capacity to model low-rank structures. The CLIP ImageNet fine-tuned weight delta (Vision or Language), obtained by the element-wise difference between the pre-trained CLIP-ViT-L/14 weights and the weights obtained after standard fine-tuning on ImageNet (also known as task vector [63]), represents a realistic target weight for LoRA in transformer in real-world fine-tuning. The High/low frequency features, generated using superposed sinusoidal functions with varying frequencies ([1000, 10000] Hz and [1, 100] Hz respectively), assess algorithms' bias towards feature frequencies." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Methodology For each target matrix pattern, we train the PEFT algorithms to minimize the mean squared error (MSE) between the estimated matrix and the target matrix. We train" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1295" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "on " + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "inline_equation", + "content": "1024 \\times 768" + }, + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": " matrices which is the dimension of the query part of the attention matrix in ViT-L/14. All algorithms are trained with the same or higher number of trainable parameters than KRAdapter to ensure a fair comparison. We report the absolute nuclear reconstruction error (average absolute element-wise singular value difference) to evaluate each algorithm's capacity to capture the spectrum of the target matrix. Our investigation focuses on comparing with LoRA [24] and its full-rank alternatives, specifically: Sin-LoRA [26], RandLoRA [2], and Krona [16]. Further details are available in appendix C." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 210, + 297, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 210, + 297, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 297, + 558 + ], + "type": "text", + "content": "Results We present results in Figure 1 where our analysis yields several critical insights. First, full-rank PEFT algorithms consistently achieve lower approximation errors compared to LoRA across a range of target matrices (low rank matrices excepted). The advantage of high-rank PEFT methods is particularly apparent when approximating matrices with highly de-correlated features such as random, whitened or sparse noise. Results on approximating traditionally fine-tuned weights for the CLIP vision and language backbones are more nuanced as RandLoRA and Krona struggle to emulate the fine-tuned delta. This difference does not translate to poor performance in practice though which indicates these algorithms probably learn to solve tasks in a different manner than standard fine-tuning does. KRAdapter is especially performant at approximating the CLIP fine-tuned deltas as well as the high frequency matrix, demonstrating a high adaptability to learn highly tailored features. We evidence a key limitation of full-rank methods when approximating explicitly low-rank target matrices, as the algorithms did not demonstrate a substantial advantage over LoRA. Further analysis suggests a potential affinity of LoRA towards low-frequency feature representations as full-rank algorithms struggle to improve over LoRA's solution in this case. The low frequency matrix is however also low rank by nature which would favour LoRA's solution. Overall, KRAdapter provides a strong solution to all cases, low rank matrix excluded, which conforms us in its suitability as a parameter-efficient algorithm. Visualization of the approximated matrices is available in Figure 5 in the appendix." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 574, + 180, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 180, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 180, + 586 + ], + "type": "text", + "content": "4.2. Vision-language tasks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 597, + 160, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 160, + 608 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 160, + 608 + ], + "type": "text", + "content": "4.2.1. Common datasets" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "We now move on to real data and compare the performance of several full-rank PEFT algorithms when applied to finetuning CLIP models [44] (both vision and language encoders) across a diverse suite of eleven image classification datasets (detailed in Appendix F). We evaluate Vision Transformer-based (ViT) architectures, initialized with publicly available openCLIP weights [9]. The PEFT algorithms are only trained on the attention heads. As an indicative, we" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "also report the performance of full fine-tuning (FT), where all parameters of the network are updated. We tune the hyperparameter of each method to achieve optimal performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 114, + 555, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 555, + 186 + ], + "type": "text", + "content": "Experiments are conducted in few-shot settings (1, 4, and 16 shots) and with varying dataset sizes (50% and 100% of the available training data). The training data is exactly the same for each algorithm to ensure fair comparisons. Results are visualized in Figure 2 with detailed results in Table 7 in the appendix." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 191, + 556, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 556, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 556, + 360 + ], + "type": "text", + "content": "Our results confirm the observations of [2] about the importance of full-rank when fine-tuning vision-language models as these algorithms largely outperform LoRA with equal trainable parameters. We also confirm that full-rank PEFT algorithms can improve performance over standard fine-tuning by limiting over-fitting for larger models (ViT-H/14 especially). KRAdapter's enhanced representational capacity translates to a further performance improvements compared to other full-rank PEFT algorithms. This performance advantage is observed across most settings, although we observe a performance saturation for " + }, + { + "bbox": [ + 313, + 191, + 556, + 360 + ], + "type": "inline_equation", + "content": "\\geq 50\\%" + }, + { + "bbox": [ + 313, + 191, + 556, + 360 + ], + "type": "text", + "content": " data settings for the larger ViT-H/14 architecture. KRAdapter's row-wise Kronecker formulation is additionally systematically superior to Krona's direct Kronecker approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 394, + 427, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 394, + 427, + 406 + ], + "spans": [ + { + "bbox": [ + 314, + 394, + 427, + 406 + ], + "type": "text", + "content": "4.2.2. Specialized datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 415, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 715 + ], + "type": "text", + "content": "To further investigate performance on tasks necessitating strong model adaptation, we conduct experiments using the Visual Task Adaptation Benchmark (VTAB1k) [62] (detailed in Appendix F.3). VTAB1k contains 19 datasets grouped into 3 categories. The structured and specialized subsets as especially interesting to us as they contain tasks less likely to be encountered during CLIP's pre-training. The specialized subset focuses on predicting specialist medical or satellite imagery while the structured subset aims to predict object state attributes such as distance, location, count, and rotation. Prior work has shown that CLIP exhibits limitations on these types of tasks [51], making them an ideal benchmark for assessing the enhanced representation learning capacity of full-rank adaptation algorithms. The results are presented in Table 1, where we report average accuracy over the 3 subsets (detailed results are available in appendix F.3.3). We observe that KRAdapter performs especially well on the natural and structured subsets but can struggle to generalize as well as RandLoRA to the structured subset for the larger models. Standard fine-tuning also struggles to generalize to the structured subset for ViT-H/14 which may indicate an implicit regularization in RandLoRA preventing overfitting, thus allowing to perform better in this specific setting. When averaged over the whole benchmark however, KRAdapter performs better across all architectures." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1296" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 72, + 233, + 190 + ], + "blocks": [ + { + "bbox": [ + 95, + 72, + 233, + 190 + ], + "lines": [ + { + "bbox": [ + 95, + 72, + 233, + 190 + ], + "spans": [ + { + "bbox": [ + 95, + 72, + 233, + 190 + ], + "type": "image", + "image_path": "cf3170ca8aae358d994f32c504854310e09d8828468a05e6d50dee4983d82a93.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 194, + 185, + 204 + ], + "lines": [ + { + "bbox": [ + 143, + 194, + 185, + 204 + ], + "spans": [ + { + "bbox": [ + 143, + 194, + 185, + 204 + ], + "type": "text", + "content": "(a) ViT-B/32" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 236, + 72, + 375, + 190 + ], + "blocks": [ + { + "bbox": [ + 236, + 72, + 375, + 190 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 375, + 190 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 375, + 190 + ], + "type": "image", + "image_path": "dcc96f8f8caee1d64dfaa0b70ae8db9e412d83663cdd7d4cd7ef01300caa9a5f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 194, + 326, + 204 + ], + "lines": [ + { + "bbox": [ + 284, + 194, + 326, + 204 + ], + "spans": [ + { + "bbox": [ + 284, + 194, + 326, + 204 + ], + "type": "text", + "content": "(b) ViT-L/14" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 213, + 523, + 225 + ], + "lines": [ + { + "bbox": [ + 85, + 213, + 523, + 225 + ], + "spans": [ + { + "bbox": [ + 85, + 213, + 523, + 225 + ], + "type": "text", + "content": "Figure 2. Tuning CLIP's vision and language backbone for image classification. Accuracy " + }, + { + "bbox": [ + 85, + 213, + 523, + 225 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 85, + 213, + 523, + 225 + ], + "type": "text", + "content": " averaged over 11 datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 378, + 72, + 518, + 190 + ], + "blocks": [ + { + "bbox": [ + 378, + 72, + 518, + 190 + ], + "lines": [ + { + "bbox": [ + 378, + 72, + 518, + 190 + ], + "spans": [ + { + "bbox": [ + 378, + 72, + 518, + 190 + ], + "type": "image", + "image_path": "1360c4de6681ab747024bafb69f71f282ace55d4053ab1c065579d836728acf8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 194, + 468, + 204 + ], + "lines": [ + { + "bbox": [ + 425, + 194, + 468, + 204 + ], + "spans": [ + { + "bbox": [ + 425, + 194, + 468, + 204 + ], + "type": "text", + "content": "(c) ViT-H/14" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 153, + 236, + 457, + 346 + ], + "blocks": [ + { + "bbox": [ + 153, + 236, + 457, + 346 + ], + "lines": [ + { + "bbox": [ + 153, + 236, + 457, + 346 + ], + "spans": [ + { + "bbox": [ + 153, + 236, + 457, + 346 + ], + "type": "table", + "html": "
MethodViT-B/32ViT-L/14ViT-H/14
Nat.Spe.Struc.Avg.Nat.Spe.Struc.Avg.Nat.Spe.Struc.Avg.
LoRA70.682.645.266.182.387.054.774.684.186.756.875.9
SinLoRA72.083.153.169.482.586.856.875.484.286.958.276.4
RandLoRA74.282.453.169.982.887.058.376.083.987.359.576.9
Krona73.783.453.370.183.487.457.576.184.986.758.076.6
KRAdapter76.084.053.371.184.886.957.276.385.887.058.177.0
FT75.783.553.170.883.287.458.376.385.484.843.771.3
", + "image_path": "bf95ed2fdd2335b395da07e0c394b1968843e53d805dd35681dc37f105b9f6a6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 183, + 355, + 425, + 366 + ], + "lines": [ + { + "bbox": [ + 183, + 355, + 425, + 366 + ], + "spans": [ + { + "bbox": [ + 183, + 355, + 425, + 366 + ], + "type": "text", + "content": "Table 1. Adaptation performance on VTAB1k's structured datasets" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 388, + 233, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 233, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 233, + 399 + ], + "type": "text", + "content": "4.2.3. Out-of-Distribution Generalization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 408, + 297, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 297, + 707 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 297, + 707 + ], + "type": "text", + "content": "We expect that the balanced singular value distribution of KRAdapter will allow for better out-of-distribution (OOD) robustness at test time. We thus propose to investigate whether parameter-efficient adaptations, optimized on in-distribution (ID) data, exhibit robustness and transferability to out-of-distribution (OOD) scenarios. We utilize a suite of established OOD benchmark datasets: ImageNet-A [23], ImageNet-Sketch [56], ImageNet-R [22], ImageNet-V2 [46], and CIFAR-100 [30]. These datasets encompass diverse distribution shifts: adversarial, stylized, renditions and domain shift, details in appendix F.4. To quantify generalization, we introduce the ratio " + }, + { + "bbox": [ + 55, + 408, + 297, + 707 + ], + "type": "inline_equation", + "content": "r_{gen} = \\frac{\\Delta_{odd}}{\\Delta_{id}}" + }, + { + "bbox": [ + 55, + 408, + 297, + 707 + ], + "type": "text", + "content": ", measuring the relative OOD accuracy gain (" + }, + { + "bbox": [ + 55, + 408, + 297, + 707 + ], + "type": "inline_equation", + "content": "\\Delta_{odd}" + }, + { + "bbox": [ + 55, + 408, + 297, + 707 + ], + "type": "text", + "content": ", averaged over OOD datasets) to ID gain (" + }, + { + "bbox": [ + 55, + 408, + 297, + 707 + ], + "type": "inline_equation", + "content": "\\Delta_{id}" + }, + { + "bbox": [ + 55, + 408, + 297, + 707 + ], + "type": "text", + "content": ", ImageNet validation) over the zero-shot model. Table 2 reports the results where we fine-tune the attentions layers of the vision backbone of CLIP only, maintaining frozen language embeddings to classify. KRAdapter consistently achieves a higher " + }, + { + "bbox": [ + 55, + 408, + 297, + 707 + ], + "type": "inline_equation", + "content": "r_{gen}" + }, + { + "bbox": [ + 55, + 408, + 297, + 707 + ], + "type": "text", + "content": " than other PEFT methods. We hypothesize KRAdapter's spectral properties enable it to efficiently work in the parameter space without overfitting to a subset of directions, unlike rank-constrained methods prone to overfitting dominant ID features. To validate this hypothesis and quantify how much the algorithms stray from the pre-trained weights, we evaluate the average Nuclear and Frobenius norm of the learned updates over the attention lay" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 388, + 556, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 483 + ], + "type": "text", + "content": "ers. We find that KRAdapter produces updates with smaller norms than other PEFT algorithms which is in line with the very small norm of the traditionally fine-tuned model and has been previously reported to be beneficial for OOD generalization [33, 59]. Table 4 in the appendix further reports the effective rank obtained when fine-tuning the vision-language architectures where we observe that KRAdapter systematically leads to higher effective ranks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 499, + 453, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 453, + 512 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 453, + 512 + ], + "type": "text", + "content": "4.3. Commonsense reasoning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "We further investigate resource-efficient fine-tuning of Large Language Models (LLMs) for commonsense reasoning through 4-bit quantization-aware training and evaluation. We evaluate this approach on Llama3.1-8B and Qwen2.5-7B, which are well suited to consumer grade GPUs with 24GB VRAM. We fine-tune the key and value projection matrices within the attention layers, optimizing hyper-parameters for each algorithm to maximize performance. To specifically evaluate out-of-distribution (OOD) generalization, we slightly modify the commonsense reasoning benchmark established in previous research [25, 36]. While the standard commonsense benchmark trains and tests on the same tasks with identical question and answer formats, we train on the multi-choice questions with \"answer{1...5}\" format (Science-QA, ARC, and OpenBookQA) and subsequently evaluate on: (1) in-distribution test sets from these training" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1297" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 70, + 504, + 196 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 504, + 196 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 504, + 196 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 504, + 196 + ], + "type": "table", + "html": "
ViT-B/32ViT-L/14ViT-H/14
IDOODrgen↑Nuc↓Fro↓IDOODrgen↑Nuc↓Fro↓IDOODrgen↑Nuc↓Fro↓
Zero-shot62.655.5n/an/an/a75.474.1n/an/an/a77.975.7n/an/an/a
LoRA72.958.20.2719.64.183.377.20.3946.26.683.775.90.0470.07.4
SinLoRA72.858.60.31223.54.782.876.30.3061.67.183.677.10.2590.68.1
RandLoRA73.058.40.2846.25.782.876.20.29159.911.882.975.4-0.05508.321.2
Krona72.458.60.3244.14.984.177.80.4242.75.085.078.20.36123.39.2
KRAdapter72.559.90.457.32.383.678.20.499.82.884.778.80.4832.85.5
FT75.558.90.260.70.885.176.70.271.11.085.578.60.394.31.9
", + "image_path": "1d5d3d9e66d02f8d57c9f292e23f2e85723aae8c53792d6864c6416d1d8a97f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 112, + 238, + 499, + 453 + ], + "blocks": [ + { + "bbox": [ + 55, + 205, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 553, + 228 + ], + "type": "text", + "content": "Table 2. Generalization to distribution shift. We report ID accuracy on ImageNet, OOD accuracy averaged over 7 OOD datasets, the " + }, + { + "bbox": [ + 55, + 205, + 553, + 228 + ], + "type": "inline_equation", + "content": "r_{gen}" + }, + { + "bbox": [ + 55, + 205, + 553, + 228 + ], + "type": "text", + "content": " ratio and the nuclear and frobenius norm of the updates. The " + }, + { + "bbox": [ + 55, + 205, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 205, + 553, + 228 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 205, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 55, + 205, + 553, + 228 + ], + "type": "text", + "content": " indicates if higher or lower scores are better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 112, + 238, + 499, + 453 + ], + "lines": [ + { + "bbox": [ + 112, + 238, + 499, + 453 + ], + "spans": [ + { + "bbox": [ + 112, + 238, + 499, + 453 + ], + "type": "table", + "html": "
MethodSIQAArcEArcCOBQAHellaSBoolQPIQAWinoGIDNIDOODAvg.
LLama3.1-8B
Zero-shot20.7326.1822.3516.0025.6943.4645.8717.6021.3225.6935.6427.23
LoRA79.0787.1274.6683.4053.7042.3576.6149.8081.0653.7055.6268.34
DoRA79.7988.2277.5684.4052.1641.7779.8757.3882.4952.1657.8070.14
SinLoRA76.4693.1486.8688.4051.5743.9484.7746.5786.2151.5758.4371.46
RandLoRA79.7987.6776.4582.6057.2252.5179.2254.3881.6357.2260.8371.23
Krona79.1787.9675.4382.4050.0759.1479.6056.6781.2450.0761.3771.31
KRAdapter79.2790.3278.1682.4055.3458.2380.9054.8582.5455.3464.6672.43
Qwen2.5-7B
Zero-shot5.025.184.104.801.7931.444.7327.704.771.7921.2910.59
LoRA51.1392.1385.8488.4071.0143.5583.7940.8179.3771.0159.7969.58
DoRA77.9487.9279.5287.8077.9557.2285.6461.6483.3077.9570.6176.95
SinLoRA77.3388.9374.8385.4058.6135.2681.3454.7881.6258.6157.1269.56
RandLoRA79.7994.8787.6389.0078.2047.4085.6960.8587.8278.2068.0477.93
Krona78.3093.8188.1491.6080.9349.0585.4257.5487.9680.9368.2378.10
KRAdapter79.7994.6186.9589.8075.3162.1484.6661.0187.7975.3170.7879.28
", + "image_path": "29c9ff8cfe7f80e95421a80b91ad63031f1927bc2456d06c391677bf16700928.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 461, + 555, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 461, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 555, + 495 + ], + "type": "text", + "content": "Table 3. 4bit quantized LLama3.1-8B and Qwen2.5-7B on commonsense reasoning tasks. We finetune the Key and Value matrices of the attention layers. We evaluate on in-distribution (ID), near in-distribution (NID) and out-of-distribution (OOD) test sets and bold the best results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 515, + 297, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 297, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 297, + 612 + ], + "type": "text", + "content": "datasets; (2) near in-distribution HellaSwag test set (multichoice, \"ending{1...4}\" format); and (3) out-of-distribution BoolQ, PiQA, and WinoGrande test sets (binary answers). This evaluation framework is designed to comprehensively assess the generalization capabilities of each PEFT algorithm, a crucial attribute for deploying robust LLMs across diverse real-world scenarios. Further details about our commonsense benchmark are available in appendix G." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "content": "Table 3 presents a comparative analysis of the PEFT methods. The results demonstrate a clear distinction: rank-restricted algorithms, such as LoRA, exhibit limited generalization to out-of-distribution (OOD) datasets. In contrast, full-rank algorithms significantly enhance the zero-shot model to answer these OOD tasks despite not specifically training for them. Notably, KRAdapter outperforms state-of-the-art methods across architectures. These findings support the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "type": "text", + "content": "conclusions of Section 4.2.3, reinforcing evidence of the efficacy of KRAdapter in learning generalizable representations for language models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 567, + 406, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 406, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 406, + 581 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "content": "We are interested here in challenging our design choices for KRAdapter and evaluating the impact of scaling the number of trainable parameters further. We study 3 scenario named low, medium and high where we iteratively double the amount of trainable parameters from the base (low) configuration used in other experiments. In KRAdapter, this is achieved by setting " + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "inline_equation", + "content": "k_{1} = \\{d_{out}^{\\frac{1}{2}}, d_{out}^{\\frac{1}{3}}, d_{out}^{\\frac{1}{4}}\\}" + }, + { + "bbox": [ + 313, + 590, + 556, + 715 + ], + "type": "text", + "content": ". Ideally, performance should increase with every parameter increase to suggest favorable scaling laws. Results are reported in Figure 3 for the vision-language task. We find that for smaller" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1298" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 95, + 220, + 186 + ], + "blocks": [ + { + "bbox": [ + 109, + 95, + 220, + 186 + ], + "lines": [ + { + "bbox": [ + 109, + 95, + 220, + 186 + ], + "spans": [ + { + "bbox": [ + 109, + 95, + 220, + 186 + ], + "type": "image", + "image_path": "ffcdaa1fbc5e7936a6d4d7fba5b3dce229bfc929f36ca3b853f3e2333c5c0dc6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 190, + 185, + 200 + ], + "lines": [ + { + "bbox": [ + 141, + 190, + 185, + 200 + ], + "spans": [ + { + "bbox": [ + 141, + 190, + 185, + 200 + ], + "type": "text", + "content": "(a) ViT-B/32" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 223, + 97, + 334, + 186 + ], + "blocks": [ + { + "bbox": [ + 223, + 97, + 334, + 186 + ], + "lines": [ + { + "bbox": [ + 223, + 97, + 334, + 186 + ], + "spans": [ + { + "bbox": [ + 223, + 97, + 334, + 186 + ], + "type": "image", + "image_path": "b15d37b5d3e44bd7f27aa64a105632e4177aa3606dca052bbd5a96d94cc4d8eb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 190, + 298, + 200 + ], + "lines": [ + { + "bbox": [ + 256, + 190, + 298, + 200 + ], + "spans": [ + { + "bbox": [ + 256, + 190, + 298, + 200 + ], + "type": "text", + "content": "(b) ViT-L/14" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 90, + 210, + 518, + 221 + ], + "lines": [ + { + "bbox": [ + 90, + 210, + 518, + 221 + ], + "spans": [ + { + "bbox": [ + 90, + 210, + 518, + 221 + ], + "type": "text", + "content": "Figure 3. Increasing the number of trainable parameters in PEFT algorithms. Accuracy " + }, + { + "bbox": [ + 90, + 210, + 518, + 221 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 90, + 210, + 518, + 221 + ], + "type": "text", + "content": " averaged over 11 datasets." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 337, + 98, + 500, + 186 + ], + "blocks": [ + { + "bbox": [ + 337, + 98, + 500, + 186 + ], + "lines": [ + { + "bbox": [ + 337, + 98, + 500, + 186 + ], + "spans": [ + { + "bbox": [ + 337, + 98, + 500, + 186 + ], + "type": "image", + "image_path": "1dce14862ad3f29d2bd0c6b2d066286bd0e440bdd1c8960e2880e32ac5d1b50c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 398, + 190, + 441, + 201 + ], + "lines": [ + { + "bbox": [ + 398, + 190, + 441, + 201 + ], + "spans": [ + { + "bbox": [ + 398, + 190, + 441, + 201 + ], + "type": "text", + "content": "(c) ViT-H/14" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 243, + 297, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 297, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 297, + 375 + ], + "type": "text", + "content": "models, full-rank methods scale better with increases in parameter budgets whereas LoRA saturates faster. For larger models however, performance does not increase beyond the low configuration which ties in with the observed saturation in larger models of Figure 2. This aligns with findings of Albert et al. [2]. SinLoRA is a notable exception as more parameters leads to a decrease in accuracy. For ViT-H/14, KRAAdapter is the only full rank PEFT algorithm to not drop in accuracy with increasing budget. This highlight the robustness of KRAAdapter to over-fitting which helps it in achieving good OOD generalization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 400, + 130, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 130, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 130, + 412 + ], + "type": "text", + "content": "5. Limitations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": "Although we have shown KRAdapter is a good alternative to LoRA when high effective rank or out-of-distribution generalization is important, we discuss here some limitations of our approach. First KRAdapter in its most parameter-efficient configuration trains more parameters than the most efficient LoRA configurations do. In practice we find that KRAdatper is equivalent to a rank 16 to 32 LoRA, a commonly used configuration. Exploring low-rank approximations for the matrices " + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 426, + 297, + 715 + ], + "type": "text", + "content": " prior to the Khatri-Rao product could potentially mitigate this, though the theoretical implications for the derived full-rank properties in Section 3.4 require rigorous investigation. Second, empirical evidence from controlled matrix estimation experiments indicates that KRAdapter's performance may be suboptimal when estimating matrices of extremely low rank. In scenarios demanding strict subspace constraints on weight updates, such as continual learning paradigms, LoRA might exhibit superior suitability. Finally, while KRAdapter consistently demonstrates robust out-of-distribution generalization and faster training convergence, we observe instances with larger models where RandLoRA achieves marginally superior performance on purely in-distribution test sets. It is important to note that KRAdapter's in-distribution performance remains highly competitive in these cases, and its advantages in gen" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 243, + 555, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 268 + ], + "type": "text", + "content": "eralization and training efficiency compared to RandLoRA generally outweigh this minor performance delta." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 304, + 388, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 388, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 388, + 316 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 328, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 556, + 704 + ], + "type": "text", + "content": "This paper introduces Khatri-Rao Adapters (KRAdapter), a novel parameter-efficient fine-tuning technique designed to overcome the representation limitations of low-rank methods. KRAdapter generate full-rank update matrices with demonstrably higher effective rank, enabling to capture more complex feature interactions and improving OOD generalization by staying close to the pre-trained weights. KRAdapter achieves all this without sacrificing parameter efficiency and training time. Our comprehensive experimental evaluation across vision-language, language understanding, and commonsense reasoning tasks consistently demonstrates KRAdapter produces improved results compared to LoRA and other full-rank PEFT alternatives, especially for specialized VTAB1k tasks that demand nuanced feature adaptation and for larger models that are prone to over-fitting. We specifically highlight that KRAdapter's improved representation space allows to produce representations closer to the zero-shot model that lead to better generalization under distribution shifts compared to other PEFT algorithms. Crucially, KRAdapter maintains the computational advantages of LoRA in terms of training speed and memory footprint, offering a practical and effective solution for fine-tuning billion-scale models as demonstrated in the commonsense experiments. A notable limitation of KRAdapter, however, is that the minimum number of trainable parameter is larger than the most efficient LoRA configurations which currently renders it unadapted to extreme scarcity scenarios. Future work could investigate replacing " + }, + { + "bbox": [ + 313, + 328, + 556, + 704 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 328, + 556, + 704 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 328, + 556, + 704 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 328, + 556, + 704 + ], + "type": "text", + "content": " with low-rank formulations to achieve better efficiency. We however anticipate this formulation would change the efficient rank guarantees of KRAdapter. Our findings underscore the importance of high effective rank weight updates in parameter-efficient adaptation of pre-trained models across diverse downstream applications." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1299" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv:2303.08774, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 190 + ], + "type": "text", + "content": "[2] Paul Albert, Frederic Z. Zhang, Hemanth Saratchandran, Cristian Rodriguez-Opazo, Anton van den Hengel, and Ehsan Abbasnejad. RandLoRA: Full rank parameter-efficient finetuning of large models. In International Conference on Learning Representations (ICLR), 2025. 1, 2, 3, 5, 8, 13, 16, 21" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 296, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 296, + 234 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 296, + 234 + ], + "type": "text", + "content": "[3] Charles Beattie, Joel Z Leibo, Denis Teptyashin, Tom Ward, Marcus Wainwright, Heinrich Kuttler, Andrew Lefrancq, Simon Green, Víctor Valdés, Amir Sadik, et al. Deepmind lab. arXiv: 1612.03801, 2016. 18" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 296, + 278 + ], + "type": "text", + "content": "[4] Daniel J. Beutel, Taner Topal, Akhil Mathur, Xinchi Qiu, Titouan Parcollet, and Nicholas D. Lane. Flower: A Friendly Federated Learning Research Framework. arxiv :2007.14390, 2020. 18" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 296, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 296, + 323 + ], + "type": "text", + "content": "[5] Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. Piaq: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on Artificial Intelligence (AAAI), 2020. 20" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 325, + 296, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 296, + 368 + ], + "type": "text", + "content": "[6] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International Conference on Machine Learning (ICML), 2020. 18" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 296, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 413 + ], + "type": "text", + "content": "[7] Xinyang Chen, Sinan Wang, Mingsheng Long, and Jianmin Wang. Transferability vs. discriminability: Batch spectral penalization for adversarial domain adaptation. In International Conference on Machine Learning (ICML), 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 446 + ], + "type": "text", + "content": "[8] Gong Cheng, Junwei Han, and Xiaoqiang Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proceedings of the IEEE, 2017. 18" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 296, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 513 + ], + "type": "text", + "content": "[9] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 514, + 296, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 296, + 547 + ], + "type": "text", + "content": "[10] M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing Textures in the Wild. In Conference on Computer Vision and Pattern Recognition (CVPR), 2014. 17" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 548, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 296, + 591 + ], + "type": "text", + "content": "[11] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. BoolQ: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. 20" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 592, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 296, + 645 + ], + "type": "text", + "content": "[12] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. 21" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 713 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 126 + ], + "type": "text", + "content": "[14] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 161 + ], + "type": "text", + "content": "[15] Emma Dugas, Jorge Jared, and Will Cukierski. Diabetic retinopathy detection (2015). URL https://kaggle.com/competitions/diabetic-retinopathy-detection, 2015. 18" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 162, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 555, + 215 + ], + "type": "text", + "content": "[16] Ali Edalati, Marzieh Tahaei, Ivan Kobyzev, Vahid Partovi Nia, James J Clark, and Mehdi Rezagholizadeh. Krona: Parameter efficient tuning with kronecker adapter. In Advances in Neural Information Processing Systems Workshops (NeurIPSW), 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 555, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 555, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 555, + 249 + ], + "type": "text", + "content": "[17] A. Geiger, P. Lenz, C. Stiller, and R. Urtasun. Vision meets Robotics: The KITTI Dataset. International Journal of Robotics Research (IJRR), 2013. 18" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 250, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 250, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 250, + 555, + 293 + ], + "type": "text", + "content": "[18] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608, 2024. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 294, + 555, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 555, + 327 + ], + "type": "text", + "content": "[19] Soufiane Hayou, Nikhil Ghosh, and Bin Yu. Lora+: Efficient low rank adaptation of large models. In International Conference on Machine Learning (ICML), 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 327, + 555, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 371 + ], + "type": "text", + "content": "[20] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In Proceedings of the IEEE international conference on computer vision, 2015. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 555, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 555, + 426 + ], + "type": "text", + "content": "[21] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Introducing EuroSAT: A Novel Dataset and Deep Learning Benchmark for Land Use and Land Cover Classification. In International Geoscience and Remote Sensing Symposium (IGARSS), 2018. 18" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 426, + 555, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 555, + 491 + ], + "type": "text", + "content": "[22] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 8340-8349, 2021. 6, 18" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 493, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 555, + 536 + ], + "type": "text", + "content": "[23] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15262-15271, 2021. 6, 18" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 537, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 555, + 591 + ], + "type": "text", + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations (ICLR), 2022. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 592, + 555, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 555, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 555, + 646 + ], + "type": "text", + "content": "[25] Zhiqiang Hu, Lei Wang, Yihuai Lan, Wanyu Xu, Ee-Peng Lim, Lidong Bing, Xing Xu, Soujanya Poria, and Roy Ka-Wei Lee. Llm-adapters: An adapter family for parameter-efficient fine-tuning of large language models. arXiv preprint arXiv:2304.01933, 2023. 6, 21" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "text", + "content": "[26] Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Sine Activated Low-Rank Matrices for Parameter Efficient Learning. arXiv:2403.19243, 2024. 1, 2, 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 712 + ], + "type": "text", + "content": "[27] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1300" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "diagnostic dataset for compositional language and elementary visual reasoning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2017. 18" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 296, + 163 + ], + "type": "text", + "content": "[28] Soroush Abbasi Koohpayegani, KL Navaneet, Parsa Nooralinejad, Soheil Kolouri, and Hamed Pirsiavash. NOLA: Compressing LoRA using Linear Combination of Random Basis. In International Conference on Learning Representations (ICLR), 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 165, + 296, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 165, + 296, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 165, + 296, + 208 + ], + "type": "text", + "content": "[29] Dawid Jan Kopiczko, Tijmen Blankevoort, and Yuki Markus Asano. Vera: Vector-based random matrix adaptation. In International Conference on Learning Representations (ICLR), 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 210, + 295, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 210, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 295, + 243 + ], + "type": "text", + "content": "[30] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 6, 17" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 245, + 296, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 245, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 296, + 288 + ], + "type": "text", + "content": "[31] A. Krizhevsky, I. Sutskever, and G. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems (NeurIPS), 2012. 18" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 291, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 291, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 291, + 295, + 335 + ], + "type": "text", + "content": "[32] Yann LeCun, Fu Jie Huang, and Leon Bottou. Learning methods for generic object recognition with invariance to pose and lighting. In Conference on Computer Vision and Pattern Recognition (CVPR), 2004. 18" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 337, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 296, + 392 + ], + "type": "text", + "content": "[33] Alex Lewandowski, Michal Bortkiewicz, Saurabh Kumar, András György, Dale Schuurmans, Mateusz Ostaszewski, and Marlos C Machado. Learning continually by spectral regularization. In International Conference on Learning Representations (ICLR), 2025. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 295, + 415 + ], + "type": "text", + "content": "[34] Fei-Fei Li, Marco Andreeto, Marc'Aurelio Ranzato, and Pietro Perona. Caltech 101, 2022. 17" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 418, + 296, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 296, + 483 + ], + "type": "text", + "content": "[35] Vijay Lingam, Atula Tejaswi, Aditya Vavre, Aneesh Shetty, Gautham Krishna Gudur, Joydeep Ghosh, Alex Dimakis, Eun-sol Choi, Aleksandar Bojchevski, and Sujay Sanghavi. SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors. In International Conference on Machine Learning Workshops (ICMLW), 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 486, + 296, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 296, + 540 + ], + "type": "text", + "content": "[36] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. Dora: Weight-decomposed low-rank adaptation. In International Conference on Machine Learning (ICML), 2024. 2, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 543, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 296, + 586 + ], + "type": "text", + "content": "[37] Y Liu, M Ott, N Goyal, J Du, M Joshi, D Chen, O Levy, M Lewis, L Zettlemoyer, and V Stoyanov. RoBERTa: A robustly optimized BERT pretraining approach. arXiv:1907.11692, 2019. 1, 2, 22" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 588, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 296, + 621 + ], + "type": "text", + "content": "[38] Loic Matthew, Irina Higgins, Demis Hassabis, and Alexander Lerchner. dSprites: Disentanglement testingSprites dataset. https://github.com/deepmind/dSprites-dataset/, 2017. 18" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 667 + ], + "type": "text", + "content": "[39] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. Pissa: Principal singular values and singular vectors adaptation of large language models. Advances in Neural Information Processing Systems (NeurIPS), 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[40] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. 21" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "type": "text", + "content": "[41] Takeru Miyato, Toshiki Kataoka, Masanori Koyama, and Yuichi Yoshida. Spectral normalization for generative adversarial networks. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 171 + ], + "type": "text", + "content": "[42] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Baolin Wu, Andrew Y Ng, et al. Reading digits in natural images with unsupervised feature learning. In NIPS workshop on deep learning and unsupervised feature learning (NeurIPSW), 2011. 18" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 555, + 205 + ], + "type": "text", + "content": "[43] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In Conference on Computer Vision and Pattern Recognition (CVPR), 2012. 18" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 555, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 555, + 270 + ], + "type": "text", + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 555, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 555, + 316 + ], + "type": "text", + "content": "[45] Maithra Raghu, Ben Poole, Jon Kleinberg, Surya Ganguli, and Jascha Sohl-Dickstein. On the expressive power of deep neural networks. In International Conference on Machine Learning (ICML), 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 555, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 555, + 360 + ], + "type": "text", + "content": "[46] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In International Conference on Machine Learning (ICML), pages 5389-5400. PMLR, 2019. 6, 18" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 555, + 392 + ], + "type": "text", + "content": "[47] Olivier Roy and Martin Vetterli. The effective rank: A measure of effective dimensionality. In European Signal Processing Conference, 2007. 1, 3, 16" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 394, + 555, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 555, + 436 + ], + "type": "text", + "content": "[48] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 2021. 20" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 437, + 555, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 555, + 480 + ], + "type": "text", + "content": "[49] Maarten Sap, Hannah Rashkin, Derek Chen, Ronan LeBras, and Yejin Choi. Socialiaq: Commonsense reasoning about social interactions. arXiv preprint arXiv:1904.09728, 2019. 20" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 482, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 555, + 525 + ], + "type": "text", + "content": "[50] Houcheng Su, Daixian Liu, Mengzhu Wang, and Wei Wang. Singular Value Penalization and Semantic Data Augmentation for Fully Test-Time Adaptation. In Association for the Advancement of Artificial Intelligence (AAAI), 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 526, + 555, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 555, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 555, + 580 + ], + "type": "text", + "content": "[51] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. In Association for Computational Linguistics (ACL), 2022. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 555, + 624 + ], + "type": "text", + "content": "[52] Chunlin Tian, Zhan Shi, Zhijiang Guo, Li Li, and Chengzhong Xu. HydraLoRA: An Asymmetric LoRA Architecture for Efficient Fine-Tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 555, + 658 + ], + "type": "text", + "content": "[53] Bastiaan S Veeling, Jasper Linmans, Jim Winkens, Taco Cohen, and Max Welling. Rotation Equivariant CNNs for Digital Pathology. arxiv: 1806.03962, 2018. 18" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[54] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R. Bowman. GLUE: multi-task benchmark and analysis platform for natural language understanding. In International Conference on Learning Representations (ICLR), 2019. 22" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1301" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 619 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 126 + ], + "type": "text", + "content": "[55] Chenglong Wang, Jiangyan Yi, Xiaohui Zhang, Jianhua Tao, Le Xu, and Ruibo Fu. Low-rank adaptation method for wav2vec2-based fake audio detection. In International Joint Conference on Artificial Intelligence Workshops (IJCAIW), 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "type": "text", + "content": "[56] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems (NeurIPS), pages 10506-10518, 2019. 6, 18" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 227 + ], + "type": "text", + "content": "[57] Hanqing Wang, Yixia Li, Shuo Wang, Guanhua Chen, and Yun Chen. Milora: Harnessing minor singular components for parameter-efficient llm finetuning. arXiv:2406.09044, 2024.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 273 + ], + "type": "text", + "content": "[58] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In Conference on Computer Vision and Pattern Recognition (CVPR), 2010. 18" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 296, + 306 + ], + "type": "text", + "content": "[59] Yuichi Yoshida and Takeru Miyato. Spectral norm regularization for improving the generalizability of deep learning. arXiv:1705.10941, 2017. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 296, + 350 + ], + "type": "text", + "content": "[60] Maxime Zanella and Ismail Ben Ayed. Low-Rank Few-Shot Adaptation of Vision-Language Models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 386 + ], + "type": "text", + "content": "[61] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. 20" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 296, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 452 + ], + "type": "text", + "content": "[62] Xiaohua Zhai, Joan Puigcerver, Alexander Kolesnikov, Pierre Ruyssen, Carlos Riquelme, Mario Lucic, Josip Djolonga, Andre Susano Pinto, Maxim Neumann, Alexey Dosovitskiy, et al. A large-scale study of representation learning with the visual task adaptation benchmark. arXiv: 1910.04867, 2019. 5, 17" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 296, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 296, + 508 + ], + "type": "text", + "content": "[63] Frederic Z Zhang, Paul Albert, Cristian Rodriguez-Opazo, Anton van den Hengel, and Ehsan Abbasnejad. Knowledge Composition using Task Vectors with Learned Anisotropic Scaling. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2, 4, 15, 16" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 296, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 296, + 564 + ], + "type": "text", + "content": "[64] Qingru Zhang, Minshuo Chen, Alexander Bukharin, Nikos Karampatziakis, Pengcheng He, Yu Cheng, Weizhu Chen, and Tuo Zhao. AdaLoRA: Adaptive budget allocation for parameter-efficient fine-tuning. In International Conference on Learning Representations (ICLR), 2023. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 296, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 296, + 619 + ], + "type": "text", + "content": "[65] Ming Zhong, Yelong Shen, Shuohang Wang, Yadong Lu, Yizhu Jiao, Siru Ouyang, Donghan Yu, Jiawei Han, and Weizhu Chen. Multi-lora composition for image generation. Transactions on Machine Learning Research (TMLR), 2024. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1302" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/099d59e1-3cf8-46ca-a7c2-5f8c85602d06_content_list.json b/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/099d59e1-3cf8-46ca-a7c2-5f8c85602d06_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..536898138ee83c7a285b2e2baacfb2d572767163 --- /dev/null +++ b/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/099d59e1-3cf8-46ca-a7c2-5f8c85602d06_content_list.json @@ -0,0 +1,1770 @@ +[ + { + "type": "text", + "text": "Towards Human-like Virtual Beings: Simulating Human Behavior in 3D Scenes", + "text_level": 1, + "bbox": [ + 98, + 130, + 898, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Liang, Wenguan Wang*, Yi Yang", + "bbox": [ + 323, + 180, + 666, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State Key Lab of Brain-Machine Intelligence, Zhejiang University", + "bbox": [ + 299, + 208, + 692, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/leonnnop/BehaviorHub", + "bbox": [ + 321, + 232, + 671, + 247 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/07bd40985b831dd2c2e9e9a990d588f6834ad77a0657139b68a6c44edaf4f3bb.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 94, + 282, + 282, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36b94a56ebda57f74e33e1e20c8365cb7b1e24cf5bf070eec62586faa2b9b3dd.jpg", + "image_caption": [ + "Figure 1. Previous research has primarily focused on: (a) linguistic-based short-horizon task planning, and (b) low-level human-scene interaction. (c) This study investigates the simulation of high-level, long-horizon, abstract goal-driven human behaviors in 3D scenes." + ], + "image_footnote": [], + "bbox": [ + 94, + 393, + 274, + 503 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd3241e930f71c9da4e9b4e61e4111d9a775817de5fb493da588fc2527c0519c.jpg", + "image_caption": [ + "Human Behavior Simulation in 3D Scenes (Ours)" + ], + "image_footnote": [], + "bbox": [ + 300, + 330, + 545, + 486 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b32ffb288e004c4551a45a545ccbb478b03c517913c899d4c5d592053290cef4.jpg", + "image_caption": [ + "High-level, Long-horizon, Abstract Goal: Preparing for work from home" + ], + "image_footnote": [], + "bbox": [ + 553, + 320, + 903, + 486 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 556, + 326, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building autonomous agents that can replicate human behavior in the realistic 3D world is a key step toward artificial general intelligence. This requires agents to be holistic goal achievers and to naturally adapt to environmental dynamics. In this work, we introduce ACTOR, an agent capable of performing high-level, long-horizon, abstract goals in 3D households, guided by its internal value similar to those of humans. ACTOR operates in a perceive-plan-act cycle, extending the ungrounded, scene-agnostic LLM controller with deliberate goal decomposition and decision-making through actively searching the behavior space, generating activity choices based on a hierarchical prior, and evaluating these choices using customizable value functions to determine the subsequent steps. Furthermore, we introduce BEHAVIORHUB, a large-scale human behavior simulation dataset in scene-aware, complicated tasks. Considering the unaffordable acquisition of human-authored 3D human behavior data, we construct BEHAVIORHUB by exploring the commonsense knowledge of LLMs learned from large", + "bbox": [ + 88, + 588, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "corpora, and automatically aligning motion resources with 3D scene for knowledgeable generation. Extensive experiments on our established benchmark demonstrate that the proposed architecture leads to effective behavior planning and simulation. BEHAVIORHUB also proves beneficial for downstream task development.", + "bbox": [ + 511, + 558, + 908, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 679, + 643, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building autonomous agents (e.g., virtual beings, or humanoid robots) that can replicate human behavior in the realistic 3D world, has been a long-standing pursuit since the inception of AI [1]. The study could empower nonplayer game character [2], underpin human-robot interaction [3, 4] and cooperation [5], populate virtual reality communities [6], and accelerate Embodied AI [7-10]. Much progress has been made in vision and language models that imitate human motions (cf. Fig. 1 (b)) or propose linguistic plans (cf. Fig. 1 (a)). However, an effective 3D humanoid agent must go beyond by conquering three major barriers: (i) Holistic goal achievement from perception to action. As depicted in Fig. 1 (c), to accomplish a high-level goal (e.g.,", + "bbox": [ + 509, + 705, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Wenguan Wang.", + "bbox": [ + 107, + 888, + 326, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10753", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "'prepare for work'), the agent must process the perceived information (e.g., 'lying on bed'), decompose the goal into a series of activities (e.g., 'get up', 'use toilet', 'eat breakfast', etc.), and devise appropriate action plans for each; (ii) Environmental dynamics. Agents should be able to actively adjust their plans based on the environment, e.g., determining if the bathroom is occupied when planning to use toilet; (iii) Vast and multifaceted human behavior space, where numerous viable paths exist to achieve even a single goal. For example, agents preparing for work can choose to eat breakfast before using toilet or vice versa. Also, when the bathroom is occupied, an intelligent agent can decide to whether eat breakfast first, or continue waiting, depending on its state and beliefs, e.g., the desire to 'complete goal as soon as possible' or 'save energy costs'. A competent agent must process such value priorities to guide its selection and evaluation of actions and policies. Besides, the absence of a comprehensive testbed further poses great challenge for agent development and evaluation, constituting barrier (iv).", + "bbox": [ + 91, + 90, + 480, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present ACTOR - a large language model (LLM) powered agent towards diligent simulation of human behavior in 3D realistic scenes (§4). ACTOR follows a perceive-plan-act cycle, addressing challenge $(i)$ as envisioned. Using LLM as a central controller, it strategizes plans by searching the human behavior space. It actively maintains a tree of behaviors, where each node represents an intermediate step toward holistic goal achievement. The construction of this tree is guided by a hierarchical prior, i.e., executable low-level actions are grouped into high-level semantic units, called activities, iteratively forming a hierarchical structure (cf. Fig. 1 (c)). The search progress is assessed using a set of customizable value functions that determine the likelihood of different intermediate candidates. In addition to being rational in common sense, ACTOR couples real-valued evaluations (e.g., best efficiency) with personalized priors expressed through language commands (e.g., description of a neat person). These outputs are converted into unified probabilities, allowing for the incorporation of the agent's characteristics and beliefs as value functions, thus addressing challenge (iii). The planning process is dynamic and grounded in specific environmental values, empowering ACTOR to readjust its plans when faced with environmental changes or new language commands, effectively tackling challenge (ii). This formulation facilitates the use of powerful search algorithms, e.g., greedy search, beam search, and Monte Carlo tree search (MCTS), with MCTS proving to be the most effective in our experiments.", + "bbox": [ + 93, + 383, + 480, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we establish a comprehensive environment for development and evaluation of agents like ACTOR, based on our newly proposed large-scale, scene-aware, behavior-rich dataset, dubbed BEHAVIORHUB (§5). One critical issue in constructing the human-authored benchmark is the high cost associated with acquiring and scaling", + "bbox": [ + 93, + 810, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "high-quality, human-generated daily behavior data. Moreover, annotating large-scale behavior data further requires creativity in designing novel tasks and expertise in creating complete plans from scratch, which is also a challenging task for humans [11]. Given this context, we propose to automatically synthesize 3D human behavior data by enhancing existing resources. Initially, we distill the tree-structured linguistic plans of human daily behaviors from LLM using in-context learning [12]. Contactable objects and plausible interactions from the scanned 3D scenes (e.g., ScanNet [13], etc.) and captured motion sequences (e.g., AMASS [14], etc.) are attributed into the in-context prompt to ensure viable plans grounding in certain environment. Then, we align the task-motion-scene data triplets by applying collision and contact constraints [15] for valid translation and rotation parameters. To promote diversity, we sample multiple plausible motion sequences for each high-level goal. In total, BEHAVIORHUB contains more than 1k daily goals over 10k high-quality behavior samples of 15.7 steps on average covering 1.5k scenes, that establishes a comprehensive testbed, that addresses the barrier $(iv)$ .", + "bbox": [ + 516, + 90, + 903, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive evaluations of ACTOR in §6. First, we find that ACTOR produces admissible plans and generates plausible motion sequences to simulate abstract, temporally-extended human behaviors. It outperforms the strong baselines by nearly doubling the overall success rate. This result is further confirmed by human evaluation. Then, we conduct several ablative experiments that limit ACTOR's access to each core design for thorough assessment. Finally, experiments on scene-aware and language-conditioned human motion generation demonstrate how BEHAVIORHUB can benefit the development of downstream task models.", + "bbox": [ + 516, + 409, + 903, + 573 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 516, + 588, + 653, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Behavior Simulation. Simulating human behavior in realistic, open world environment like the one we inhabit is a long-standing topic in artificial intelligence [16]. Historically, the topic has primarily been studied in the game worlds, focusing on enhancing player experiences through intelligent non-player game characters (NPCs) [17-19]. Early approaches rely on rule-based approaches like finite-state machines [20] and behavior trees [21]. They provide a brute force way of manually crafting agent's behaviors, but cannot perform new procedures that were not hard-coded in their script [22], limiting the generalizability. Another strand [23, 24] involves using reinforcement learning, where agents learn its own policy through optimizing the learning algorithm on readily definable rewards over fixed task space, which is often limited to non-open, adversarial games, or blocks worlds only [25-27]. The third line of research, represented by some pioneer works [28-30] in computational cognition, aims to build machines that operate directly in perceive-plan-act cycles, encompassing", + "bbox": [ + 516, + 613, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10754", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the nature of autonomous agents as originally envisioned. This formulation holds potential generalizability to most, if not all, open-world contexts. However, these studies are typically restricted to simplified environments, such as first-person shooter games [29], or 2D gridworlds (e.g., Generative Agents [31]), and focus on a reduced range of behaviors. Our work falls in vein of third category, while pushing frontier towards simulation in realistic 3D environments.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, beyond the game world, new trending on social embodied intelligence, such as assistive robots, draws attention to simulation of human collaborators' behavior. One notable effort is VirtualHome [11], which studies high-level human activities as plain sequences of atomic actions. However, its formulation is inherently environment-agnostic, which has been treated as a purely linguistic procedural planning problem in subsequent studies [32, 33]. Furthermore, the discrete action space built upon manually crafted procedural knowledge also makes it an insufficient testbed. In contrast, with a special focus on environment-aware simulation of contiguous daily behaviors, our work paves one solid step accelerating the development of social embodied intelligence.", + "bbox": [ + 89, + 215, + 482, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLM as Planner. Remarkable progress in LLMs has been observed, demonstrating their emerging capacity to break down complex tasks into more manageable sub-tasks and devise appropriate plans for each [34-36]. LLMs have been successfully applied to solving mathematical problems [37, 38], reasoning on commonsense [39], planning robotics tasks [40, 41], and very recently, manipulating external APIs on a web scale, expanding their capabilities beyond text generation. Based on the observation, we posit that LLMs can serve as a crucial component in extending the perception-decision-action space to construct human-like agents in realistic 3D environments. However, a crucial challenge remains: LLMs lack experience and interaction with their environment [12, 42], preventing them from ordering actionable and rational plans. This paper addresses this issue by incorporating customizable value functions into LLMs to evaluate and prioritize plans. The idea bears some resemblance to recent robotics research [43, 44]. However, those methods necessitate retraining for every new set of primitive robotic skills, making them impractical for the complex and undefined human action space, which is difficult to define in advance. In contrast, we support plan evaluation using real-numbered functions and language-based rules without resource-intensive retraining.", + "bbox": [ + 89, + 429, + 482, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "LLM as Data Generator. Being trained on the large corpus of human-produced language, LLMs are believed to contain a wealth of information about the world [45-47]. Given a handful of task-specific prompts, LLMs can generalize and generate more linguistic data in the same format, with the application of generating tabular data [48], relation triplets [49], sentence pairs [50], instruction data [51],", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "etc. The idea seems naturally to be borrowed for human behavior data acquisition, where the requested human motion and daily activity procedure were commonly crowdsourced with high expense and complexity [11], limiting the scale and coverage of related datasets [52]. However, the process is non-trivial. The generated data is often blamed for low quality and diversity issues [53, 54]. The 3D environment-aware nature of the task also poses unique challenges. To respond, we explore attributed prompts specifically conditioned on the environment that not only mitigates the problems of low informativeness and redundancy, but also offers an effective workflow that can further empower other related domains, such as human-scene interaction [55].", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Environment", + "text_level": 1, + "bbox": [ + 513, + 303, + 648, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Simulating open-ended goals that resemble naturalistic human behaviors necessitates an environment capable of facilitating diverse agent affordances and interactions. Before delving into the detailed agent architecture, we first describe such environment we tailored for agents to instantiate in.", + "bbox": [ + 511, + 328, + 903, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Environmental Setup. The environment features common affordances in a household, including:", + "bbox": [ + 511, + 405, + 903, + 434 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Scene of 3D textured meshes, that spans a house of multiple functional areas (e.g., kitchen, etc.).", + "- Objects embedded in the scene (e.g., stove in kitchen, bed in bedroom), with each constructed of 3D textured geometry and corresponding object state (e.g., fridge: opened).", + "- Humanoid agent(s) defined by SMPL-X [56], an expressive 3D human model of shape and pose of both body and hand. Agents reside in and interact with the scene and can influence the state of objects by their actions (e.g., close the fridge | fridge: opened $\\rightarrow$ closed)." + ], + "bbox": [ + 511, + 435, + 903, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a start point, we demonstrate the environment of indoor household, keep the general formulation open for other environments, such as outdoor streets [57].", + "bbox": [ + 511, + 587, + 903, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Environmental Simulator. Interactions within the environment are driven by the simulator of two components: (i) an engine that manages and evolves the states of objects in the environment; and (ii) a renderer that supports generation of multiple perceptual observations (e.g., RGB, depth, 3D surfaces) for agents. At each time step, the simulator dynamically updates environment and collects egocentric, surround, or third-person-view information based on needs. We build our simulator upon Habitat-Sim [58, 59] to ensure efficient and parallelizable simulation.", + "bbox": [ + 511, + 633, + 905, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. ACTOR Agent", + "text_level": 1, + "bbox": [ + 511, + 799, + 658, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to build humanoid agents that naturally simulate 3D human behaviors to complete daily goals, which can be lengthy, abstract, or ambiguous. Agents receive high-level goal described in language and are tasked with generating plausible motion sequences that align with the given scene.", + "bbox": [ + 511, + 825, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10755", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/383f20063d400ecf9bd55b6c27206e86d702832445b0a037574d7e16aac007c7.jpg", + "image_caption": [ + "Figure 2. Left: Overall agent architecture (§4.1); Right: One-step in perceive-plan-act loop with value-driven behavior planning (§4.2)." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 903, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Agent Architecture", + "text_level": 1, + "bbox": [ + 89, + 273, + 274, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2 shows the workflow of ACTOR. The agent operates in a perceive-plan-act loop, using language as the generic interface connecting the three phases. At each step, it selectively perceives the world based on the target, transforming information into an environment description using heuristics. The observed information assists in decision-making through the LLM core, ultimately resulting in actions in the form of 3D motion sequences based on linguistic plans within the environment. We now delve into the details.", + "bbox": [ + 89, + 296, + 483, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Perception. Our definition of perception extends beyond gathering information, such as 2D images or 3D point clouds, through the simulator (§3). It encompasses attaining a deep understanding of the environment, including object properties, spatial relationships, and scene layouts, etc. In our preliminary implementation, information is perceived using readily available models and converted into linguistic descriptions using heuristic functions (supp. §A.2).", + "bbox": [ + 89, + 431, + 483, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Plan. At the core of ACTOR, a planning module receives the current environment and past behavior trial as input, generating action description as output. For the executable actions, it provides descriptions of motion sequences. In the case of high-level activity, it bypasses the action stage and proceeds to break down the target further. We will provide a detailed explanation of this process in §4.2.", + "bbox": [ + 89, + 554, + 483, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Action. For action, we specifically consider whole-body human actions in 3D scenes to closely resemble human behavior using the off-the-shelf models. It generates whole-body motion based on text and trajectory [60]. For object-interactive actions, we further refine hand grasping using an isolated grasp estimation model [61]. Additionally, for moving actions like walking, trajectory paths are pre-estimated [62]. Detailed trajectory estimation process is provided in supp. §A.3.", + "bbox": [ + 89, + 659, + 483, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Value-driven Behavior Planning", + "text_level": 1, + "bbox": [ + 89, + 803, + 375, + 819 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An LLM, denoted as $p_{\\theta}$ with parameters $\\theta$ , functions as a controller to iteratively decompose the long-term goal $l$ described in language texts into shorter steps of behaviors based on the partial environmental dynamics. To prevent confusion with the global environment $E$ , we denote the", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "partial environmental observation at step $i$ as $E_{i}^{\\prime}$ and its corresponding textual description as $e_{i}$ . For brevity, we omit step indicator $i$ and define $E^{\\prime}$ to include both the environment state and agent's state. We denote textual descriptions of behaviors as $z$ . The problem is formulated as a search over a tree, where each node is a state $s = \\{z_{1\\dots i}\\}$ , representing a partial trial with the input and the sequence of behaviors taken thus far. The instantiation consists of three components: (i) Node expansion, which involves generating $k$ candidate behaviors for the next step search; (ii) Value functions to evaluate each node; (iii) Search algorithm that accounts for branch selection.", + "bbox": [ + 511, + 275, + 906, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Node Expansion. The node is expanded through sampling from LLM with a window size $w$ : $z^{(j)} \\sim p_{\\theta}(z_{i+1}|l,e,s,h)$ , where $j \\in \\{1,\\dots,w\\}$ . We omit the basis prompting parts for brevity, which is also the case in the subsequent context. Here, we introduce a hierarchical heuristic, denoted as $h$ , which is implemented through prompt instructions and demonstrations of specific actions and activity cases. This heuristic ensures that during each expansion, all candidates at the same level are restricted to executable actions or high-level semantic units of activities, that provides a more nuanced representation of behaviors for effectively modeling interchangeable activities.", + "bbox": [ + 511, + 458, + 908, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Value Function. The value function assesses the state by considering the degree to which a specific behavior contributes to the achievement of the target, conditioned on the agents' beliefs reflecting their value. The search algorithm uses the output to determine which nodes to explore further and in what order. The likelihood is given by $p(z_{i+1} | l, E', s) \\propto p_{\\theta}(z_{i+1} | l, e, s) \\cdot p_v(z_{i+1} | l, E', s)$ . Here, the LLM provides us with $p_{\\theta}(z_{i+1} | l, e, s)$ , which represents the likelihood, based on commonsense, that a textual behavior is a valid next step. However, the LLM struggles to generalize or make inferences in the real environment since it is not grounded. On the other hand, $p_v$ provides the likelihood of the behavior being plausible in the current state of both the environment and agent, according to the defined values. We consider a set of values $\\{p_{v_n}\\}_n$ categorized into two types, and $p_v(z_{i+1} | l, E', s) = \\prod_n p_{v_n}(z_{i+1} | l, E', s)$ :", + "bbox": [ + 511, + 641, + 908, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Real-valued function assigns real values as outputs, such", + "bbox": [ + 511, + 885, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10756", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/76b5d3c4930ff393da2ced98b0fc7f3815d0101d5c31190c561e7a652961624d.jpg", + "image_caption": [ + "Figure 3. (a) An illustrative example of our BEHAVIORHUB dataset; (b) Semi-automatic 3D human behavior data generation pipeline ( $\\S 5$ )." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 359, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8ee00300b55710f3d54b0a1c8135779ab602fd33bd84da53d3f0c46e0a94e0da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 90, + 903, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "as the shortest path value, for which the function estimates the distance for each candidate action. The outputs of real-valued functions are normalized into probabilities.", + "bbox": [ + 102, + 301, + 480, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Language-based command is implemented by prompting the LM with a value prompt that conveys the meaning of 'How likely is it for someone who is a neat person to take the following action, considering ...'. The model then reasons about each action to generate the corresponding probability reflecting the value.", + "bbox": [ + 89, + 345, + 482, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Values are also conditioned on the state of the environment and agent. This approach allows for reacting to environmental dynamics in addition to the active planning process. Search Algorithms. Owing to the active planning process we have formulated, it is feasible to use different search algorithms based on the tree structure. We explore greedy search [63], beam search [64] and MCTS [65], and evaluate their performance in experiments (cf. §6.2) where we find MCTS performs best. We therefore use MCTS by default.", + "bbox": [ + 89, + 436, + 483, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Implementation Detail", + "text_level": 1, + "bbox": [ + 89, + 582, + 302, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ GPT-4 and GPT-3.5-turbo API provided by OpenAI as base LLMs. We prompt them with four in-context examples as demonstrations, with decoding temperature as 0 for more deterministic generation. We use official releases of conditional motion generation models and fine-tune them on BEHAVIORHUB using the default parameters for each to promote finer generation quality. By default, we set the window size to $w = 5$ and normalize the shortest path value based on the maximum rollout depth of the search algorithm, which is set to 3. In practice, we find one-time sampling is sufficient for effective tree search, eliminating the need for multiple samplings for branch aggregation. Our algorithm is implemented in PyTorch using Tesla A40 GPUs.", + "bbox": [ + 89, + 604, + 483, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. BEHAVIORHUB Benchmark", + "text_level": 1, + "bbox": [ + 89, + 815, + 352, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To lay a solid foundation for future research, we build a large repository of common behaviors performed in daily household scenarios. Each sample contains three components: $(i)$ a high-level goal (i.e., root node); $(ii)$ a tree", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "structured linguistic plan covering necessary intermediate-level steps (i.e., intermediate nodes) and low-level steps (i.e., leaf nodes) required to accomplish the goal; and (iii) scene-conditioned human motions corresponding to each executable step at either intermediate or low level. We illustrate one example in Fig. 3 (a). Intermediate step set can be empty. Both goal and steps in plans are human activities described with either concrete or abstract language (e.g., 'go to sleep on bed' or 'feel tired'). To generate diverse and high-quality data, we employ a two-step pipeline that utilizes an LLM pretrained on extensive corpora: First, automatic generation of linguistic daily plans ( $\\S 5.1$ ); Second, alignment of the plans with 3D motions and scenes ( $\\S 5.2$ ). The entire data generation process is shown in Fig. 3 (b).", + "bbox": [ + 511, + 301, + 906, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Linguistic Goal-Plan Generation", + "text_level": 1, + "bbox": [ + 511, + 526, + 802, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pipeline consists of three steps: $(i)$ generate potentially incomplete goal-plan trees; $(ii)$ complete and refine each tree; and $(iii)$ filter out low-quality data. We provide detailed prompts in supp. §A.1.", + "bbox": [ + 511, + 550, + 905, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attributed Goal-Plan Tree Generation. In the initial step, we generate new goal-plan trees using a bootstrapping approach based on a small set of seed human-written samples. To ensure broader coverage and facilitate later alignment with specific scenes, we attribute the starting room, candidate objects and actions that interact with the activity in the plans into prompt demonstrations. We start the sample pool with 292 activities from ActivityPrograms [11], and consider 21 room types (e.g., bathroom, gym), $\\sim 10^3$ object types (e.g., TV, table), and $\\sim 10^2$ action types (e.g., open, eat). Refer to the supp. §A.5 for the full list. For each step, we sample eight goal-plans as in-context examples from the pool while restricting the outputs to one room and ten types of objects. Out of the eight plans, six are from the human-written plans, and two are generated by the model in previous steps to enhance diversity.", + "bbox": [ + 511, + 612, + 905, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generated trees are then labeled with sequential order for intermediate nodes in the trees, which encompass multiple sub-step leaf nodes of individual actions. This sequential or", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10757", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "der is represented as interchangeable groups, where nodes within the same set are interchangeable with each other. We prompt the LM in a few-shot way to determine this.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Goal-Plan Tree Refinement. We further improve the constructed trees by addressing two key aspects: (i) we complete any missing internal plan steps, which can often be revised based on commonsense, e.g., opened the fridge without closing it. (ii) we enhance the root node descriptions to be more abstract, e.g., transforming 'use toilet' to 'feel the call of nature'. We accomplish both aspects by querying the LLM for suggestions.", + "bbox": [ + 89, + 137, + 480, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Filtering. To promote diversity, a new tree is added to the pool only if its BERTScore similarity [66] with any existing goal-plan tree is below 0.5. Additionally, we utilize LLM to assess the generated trees by asking the question, 'Is this a valid plan?' Any plans flagged as 'invalid' are filtered out to ensure high-quality data.", + "bbox": [ + 89, + 258, + 480, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Goal-Motion-Scene Alignment", + "text_level": 1, + "bbox": [ + 89, + 359, + 361, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Once the goal-plan trees are formed, we proceed to ground them in the 3D environment. We propose to synthesize 3D behaviors by leveraging existing resources of human motions (i.e., AMASS [14], BABEL [67], GRAB [61]), and indoor scenes (i.e., ScanNet [13], HM3D [68], Replica [69]). For each tree, we first sample actions and objects from resources using corresponding labels to fulfill the executable activities in forms of combinations of actions on objects. Then, given sampled motion, we aim to generate plausible and contiguous interactions to sampled object in scene.", + "bbox": [ + 89, + 382, + 482, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Motion Alignment. We first put motion clips into the scene anchored by the contactable objects. We optimize the translation and rotation parameters matrices by minimizing the collision and contact losses from MOVER [15]. A unified SDF volume is calculated, and all contact vertices for all frames are accumulated in 3D space. The motion is aligned through a transformation of the two matrices. This joint optimization improves human-object contact and resolves 3D interpenetrations between humans and the scene.", + "bbox": [ + 89, + 534, + 482, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sequence Blending. The aligned motion sequence, which may be sparse and not spatially connected, is blended using a Transformer-based motion completion method [70]. This results in a contiguous motion sequence that aligns with both the scene and goal plan. In practice, we find a single network is capable of delivering satisfactory results that align with prior research [71].", + "bbox": [ + 89, + 670, + 482, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Verification. Finally, to ensure the data quality, each example is examined by three verifiers who vote on whether the plan (motion) is complete (valid). If an example receives majority approval, it is accepted; otherwise, it is dropped.", + "bbox": [ + 89, + 776, + 482, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Dataset Statistics", + "text_level": 1, + "bbox": [ + 89, + 848, + 259, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our dataset consists of 10k human behavior samples in 1.5k 3D scenes, covers 2k unique activity over 0.1k actions and", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1k objects, which is an order of magnitude larger than the human-authoring ActivityPrograms [13]. On average, each high-level goal has 15.7 steps, resulting in a total of 10.1 corresponding motion sequences of 83.3 frames, that span over 8.6M motion frames. Each activity corresponds to 4.1 different motion sequences in 1.7 rooms. Owing to the automatic data generation pipeline, our dataset achieves even higher diversity and coverage with low human cost.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Evaluation Metric", + "text_level": 1, + "bbox": [ + 511, + 223, + 689, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our evaluation of the plausible behavior simulation encompasses two key aspects: $(i)$ the simulation should generate linguistically reasonable behavior plans (i.e., behavior planning); and $(ii)$ align them with natural motion sequences within the 3D environment (i.e., behavior simulation).", + "bbox": [ + 511, + 244, + 903, + 321 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For behavior planning, we use Sentence-BLEU [72] and BERTScore [66] to measure the semantic similarity between the ground-truth plans and predictions. We report maximum scores attained across ground-truth plan variants to ensure the evaluation be order-invariant with respect to interchangeable sub-steps.", + "- For behavior simulation, we consider three sets of metrics: (i) Success Rate: The step success rate (SSR) records the percentage of steps where the agent successfully completes the step objective, defined by a contact distance threshold. For example, we consider lying down to be successful if both the hip and head of the humanoid are within $30\\mathrm{cm}$ of the target location [73]. The goal success rate (GSR) is measured to determine whether all steps in the entire plan are successfully executed; (ii) Goal Success Rate Weighted by Path Length (GSRPL): It judges how efficient was the agent at finishing the goal, defined as $\\mathrm{GSR} \\cdot \\frac{g}{\\max(g,l)}$ . Here $g$ is ground-truth path length and $l$ is the agent's path length; (iii) Motion Quality: We evaluate overall quality of generated motions using Frechet Inception Distance (FID) and recognition accuracy measured with the final layer of a pretrained standard RNN action recognition classifier as motion feature extractor." + ], + "bbox": [ + 513, + 321, + 905, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiment", + "text_level": 1, + "bbox": [ + 511, + 686, + 637, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Performance on BEHAVIORHUB Benchmark", + "text_level": 1, + "bbox": [ + 511, + 710, + 877, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset Split and Dynamic Subset. We randomly sample 200 held-out goals as demonstration set from which we select example(s) for prompting language models, and also as training data for fine-tuning conditional generation models. The remaining ones are used for evaluation.", + "bbox": [ + 511, + 733, + 903, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To thoroughly examine the capacity of benchmarked models in managing environmental dynamics, we manually create a subset of 300 samples from the original BEHAVIORHUB, called Dynamic subset. In each sample, we carefully configure the agent's environment state-aware triggers to guarantee a distinct goal-plan solution for a specific goal,", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10758", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/52e701d8fc71dea626c0feba350c53b68da8a2a4e67793841fd1bece7c0660d6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBehavior PlanningBehavior Simulation
S-BLEUBERT-SSSRGSRGSRPLFID ↓Accuracy ↑
Main SetLLMaP [33] [ICML2022]0.0890.825-----
HuggingGPT [34] [NeurIPS23]0.1320.8560.5330.3170.1615.3860.620
ACTOR (Ours)0.1700.8790.6010.4720.3512.0870.773
Dynamic SubsetHuman0.2030.959-----
LLMaP [33] [ICML2022]0.0690.821-----
HuggingGPT [34] [NeurIPS23]0.0990.8300.4070.1640.0739.1160.505
ACTOR (Ours)0.1350.8620.5150.3060.2123.1410.697
", + "bbox": [ + 102, + 88, + 903, + 227 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/105c00aef1fc19ca71363c6bdb91069601bf2a46e8b2b6e2355da2fb4e4cf83f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 253, + 236, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c95c90ce58bc48753305ca24b79a7ad14e1b96d17237c7e7f24a68ef96c34a8b.jpg", + "image_caption": [ + "Original plan: weekend cleaning $\\rightarrow$ clean bedroom $\\rightarrow$ Strategized plan:", + "(occupied signal removed)" + ], + "image_footnote": [], + "bbox": [ + 258, + 276, + 558, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b64adb1e94242751db9dbea8501b12f0db81e7a48fff2aca043a39124028aafd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 301, + 589, + 311 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a97d39aab47f096159cb2557502d9223afb04e2d113cadf24165a3ef007138b3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 276, + 697, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1444e61b2e50bc1aca4e84d1dd643d0d62a48755108d682b1ba98c4c523ca83d.jpg", + "image_caption": [ + "Clean bedroom" + ], + "image_footnote": [], + "bbox": [ + 699, + 286, + 799, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c8ef65b30b3f52532f40ac9912db37642aff8c3b76758d8d83e2252600198e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 799, + 286, + 825, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d22317b8a865bf29be4dbbca2b37b4b8a173016a874a30a5ea680052aef4e630.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 825, + 286, + 898, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d0261e1606befb36ed42293b996b46615ec2713130c7b513b6025371a6ec93e.jpg", + "image_caption": [ + "Goal: prepare for th", + "Figure 4. Qualitative results of ACTOR on the challenging BEHAVIORHUB Dynamic subset. Please refer to §6.1 for more details." + ], + "image_footnote": [], + "bbox": [ + 96, + 359, + 196, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5e1088c145076ebe587cfff6df9130bc3167e7583f1123b3bc43d50068f001c.jpg", + "image_caption": [ + "Language command: he/she is a neat person" + ], + "image_footnote": [], + "bbox": [ + 199, + 359, + 297, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b120b4b17ddf01cdc28ec6a2c4810c165864a21c621c3c92c6773d53ed1d619f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 359, + 398, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2092a6ccacd391c10a7b0f0876f4518fc5c932143cffd6b38b8ea1ae932f32f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 359, + 500, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/651f81261a41e6b2a7676d43e5cd15204931fe2d298fd71c05a97589aa3126c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 359, + 599, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16e8a8fedeee2b016faa9777a90249094806def0e52f525502bde8d90ba102e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 359, + 699, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad8b8ffeab641dedd45303717db750bc5ca68526f4fb725ee7afad5fe60320b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 359, + 799, + 428 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ede2439405ca5e53c44063a00858e96b2f467ba145f1938900533c9e1a49caed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 359, + 898, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "e.g., we designate the bathroom as occupied only after certain pre-request steps have been fulfilled; or incorporate language commands that specify agent characteristics, further contributing to a unique goal-plan preference.", + "bbox": [ + 89, + 468, + 482, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Competitors. We benchmark two top-leading LLM-based models: LLMaP [33], a procedural planning model, and HuggingGPT [34], a general tool agent, to probe the human behavior simulation ability in existing techniques. LLMaP is designed to operate solely on textual inputs, and lacks the capability to perform behavior simulation. We assess its behavior planning ability only.", + "bbox": [ + 89, + 532, + 483, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Evaluation. Table 1 summarizes the automatic evaluation results, from which we take three major observations: (i) ACTOR achieves the best results on both behavior planning and simulation and outperforms the strong baseline by nearly doubling the GSR, indicating the effectiveness of our value-driven behavior planning design. (ii) On the challenging Dynamic subset, both LLMaP and HuggingGPT experience a noticeable decline in terms of BERT-S, whereas our ACTOR performs much better. To illustrate, in the case shown in Fig. 4, both LLMaP and HuggingGPT directly ignore the occupied signal and give an irrational plan of heading directly to the bedroom, causing a failure. This observation confirms our claim that the ungrounded LLM falls short in terms of environment-aware behavior planning. (iii) Although our ACTOR shows more promising results, there still remains a significant gap compared to human performance. This highlights the need for", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c03f11badfbf91cf5e25421db772756952a35be8bf953b34c2af3e6bf8f330f1.jpg", + "table_caption": [ + "Table 1. Quantitative results on Main and Dynamic set of BEHAVIORHUB. $\\downarrow$ ’ indicates smaller values are better. See §6.1 for details." + ], + "table_footnote": [], + "table_body": "
MethodComplete.Rational.Quality
Human4.024.85-
LLMaP [ICML2022]2.232.48-
HuggingGPT [NeuIPS23]2.712.893.18
ACTOR (Ours)3.053.473.75
", + "bbox": [ + 522, + 465, + 901, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Human-evaluated results (§6.1) on Dynamic subset.", + "bbox": [ + 519, + 551, + 898, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "developing more sophisticated behavior simulation models.", + "bbox": [ + 511, + 579, + 903, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human Evaluation. For a comprehensive evaluation, we engage five participants as human evaluators. They rate the model performance based on three aspects: (i) Completeness: assess whether the motion steps can successfully complete the target goal, capturing semantic completeness; (ii) Rationality: evaluate whether the sequence includes necessary steps in the correct sequential order to accomplish the target goal, capturing sequential order correctness; and (iii) Quality: reflect the naturalness and smoothness of the motion sequence, capturing motion quality. The results of the human ratings, based on a 5-point Likert scale, are reported in Table 2. The human subjective judgments generally align with the trends reflected by Table 2, confirming the reliability of our constructed automatic evaluation framework. Also, the results reaffirm that ACTOR yields plausible behavior simulation. However, it is worth noting that human-written plans are consistently preferred over our results, underscoring the challenging nature of our newly proposed BEHAVIORHUB benchmark.", + "bbox": [ + 511, + 595, + 905, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Analysis. Examples from ACTOR on BEHAV", + "bbox": [ + 511, + 885, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10759", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ad52602b83013861bcbebb2b2ecf8fb6115a7efc6f3ed18ef565f8f9b19bba98.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBERT-SGSRGSRPL
Baseline0.8110.1400.062
+ Active Search0.8370.2350.132
+ Hierarchical Prior0.8490.2610.154
+ Value Function0.8620.3060.212
", + "bbox": [ + 101, + 88, + 400, + 169 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/65f5c8b3311ef846ca4651f51b004d659c63e524a4273c4dc3031ddf0c701e26.jpg", + "table_caption": [ + "(a) Key Component Analysis" + ], + "table_footnote": [], + "table_body": "
AlgorithmBERT-SGSRGSRPL
Greedy0.8400.2440.151
Beam0.8530.2870.186
MCTS0.8620.3060.212
", + "bbox": [ + 408, + 103, + 647, + 169 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/02de33cd530696cb809b6b31be13d4edd016b3b078dae59967121a3871e376d9.jpg", + "table_caption": [ + "(b) Search Algorithm" + ], + "table_footnote": [], + "table_body": "
LLMBERT-SGSRGSRPL
Vicuna-7b0.8080.0920.063
GPT-3.50.8330.1760.116
GPT-40.8620.3060.212
", + "bbox": [ + 658, + 103, + 898, + 169 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "IORHUB Dynamic are visualized in Fig. 4. In addition to simulating behavior to strategize and successfully achieve the desired goal, ACTOR is able to respond to environmental changes and adapt to language commands. For instance, to accomplish the goal of 'Sunday cleaning', while someone is still in bed, i.e., the bedroom is occupied, the agent prioritizes scheduling the kitchen first and then the bedroom, waiting for the person to wake up. In the second example, considering the truth that someone is a neat person, he/she engages in home cleaning before eating breakfast.", + "bbox": [ + 89, + 214, + 483, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Diagnostic Experiments", + "text_level": 1, + "bbox": [ + 89, + 377, + 310, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablative studies are conducted on BEHAVIORHUB Dynamic for indepth analyzing each component in ACTOR, using BERT-S, GSR, GSRPL (cf. §5.4) as evaluation metrics", + "bbox": [ + 89, + 400, + 482, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Key Component Analysis. We first validate the importance of our proposed components by attaching them one at a time in Table 4b. The $1^{st}$ row reports the result of a bare baseline model, which produces a global plan based on the given linguistic goal. Next, in the $2^{nd}$ row, we transition from one-pass planning to active tree search, resulting in improved performance and supporting our claim that active planning can mitigate the impact of environmental changes. Moreover, the $3^{rd}$ row gives the score when the hierarchical behavior structure prior is employed when spanning searching branches. As seen, this leads to moderate improvement by constraining the search space with interchangeable semantic units, highlighting its necessity in handling vast and complex human behavior space. Finally, as shown in the $4^{th}$ row, incorporating value function significantly enhances the overall success rate and success rate over path length, aligning with the rational preference for the shortest path.", + "bbox": [ + 89, + 446, + 483, + 703 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Search Algorithm. Table 3b reveals the impact of search algorithms (§4.2), i.e., greedy search, beam search, and MCTS, with window sizes of 5 for the later two algorithms. The default strategy, MCTS, shows optimal results, which aligns with the widely accepted understanding that MCTS is more effective when dealing with a large solution space.", + "bbox": [ + 89, + 704, + 482, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Modular Scalability. In Table 3c, we investigate the modular scalability by employing various LLM cores, including GPT-3.5, GPT-4 and open-source LLM Vicuna-7b [74]. As the LLMs' capabilities improve, ACTOR exhibits a continuous enhancement in its behavior simulation performance. This observation substantiates the notion that our ACTOR system possesses great potential for accommodating the de", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2222aaf3221b6b377edd5f36a62b49c3e0f0328db7655c3233add87a165422dc.jpg", + "table_caption": [ + "(c) Modular Scalability", + "Table 3. Ablative experiments on the Dynamic subset of our proposed BEHAVIORHUB. Please refer to §6.2 for more details." + ], + "table_footnote": [], + "table_body": "
MethodMPJPE↓MPVPE↓
[75] [CVPR2021]242.50222.13
†[75] [CVPR2021]201.56189.21
", + "bbox": [ + 524, + 210, + 906, + 263 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5f0a29e28cbaf1b6e66ad255aca3f4b570585685ff4856a5b7e9dbfc5b347d54.jpg", + "table_caption": [ + "(a) Scene-aware Motion Generation" + ], + "table_footnote": [], + "table_body": "
MethodFID↓R Precision ↑
MDM [76] [ICLR2023]0.5440.611
†MDM [76] [ICLR2023]0.4710.705
", + "bbox": [ + 524, + 279, + 905, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(b) Language-conditioned Motion Generation", + "bbox": [ + 589, + 333, + 831, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Quantitative results (§6.3) on two downstream tasks. † indicates using BEHAVIORHUB for additional training.", + "bbox": [ + 511, + 351, + 903, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "velopment of more powerful LLM cores. Moreover, this shows that BEHAVIORHUB, in conjunction with the human behavior planning task, serves as a robust testbed for evaluating the planning capability of LLMs.", + "bbox": [ + 511, + 396, + 905, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.3. BEHAVIORHUB for Downsteam Task", + "text_level": 1, + "bbox": [ + 511, + 473, + 818, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further probe the effectiveness of BEHAVIORHUB by incorporating it as additional training data for two downstream tasks: scene-aware [75] and language-conditioned motion generation [76], which focus on generation without planning. We follow the official implementation of both methods and begin by pre-training the two models on BEHAVIORHUB. Subsequently, we fine-tune then evaluate these models on PROX [52] and HumanML3D [77], respectively. The evaluation results are reported in Table 4. As observed, BEHAVIORHUB significantly enhances model performance across all evaluation metrics for both tasks, highlighting the potential of our created dataset in facilitating broader applications within the realm of motion generation.", + "bbox": [ + 511, + 498, + 906, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 720, + 633, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present ACTOR, an LLM-powered agent towards realistic simulation of human behavior in 3D scenes. ACTOR integrates an LLM controller to perform complex behavior through planning on goal decomposition guided by hierarchical activity prior. The value-driven mechanism further deepens its understanding of environment. We demonstrate its potential in the simulated 3D indoor environment constructed using our newly created large-scale, scene-aware, behavior-rich dataset, BEHAVIORHUB. Evaluations suggest effective behavior planning and simulation of ACTOR.", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10760", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the National Science and Technology Major Project (No. 2023ZD0121300), National Natural Science Foundation of China (No. 62372405), Fundamental Research Funds for the Central Universities (226-2025-00057), National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University (No. HMHAI-202403), China Postdoctoral Science Foundation (No. 2025T180421), and the Postdoctoral Fellowship Program of CPSF (No. GZC20251066).", + "bbox": [ + 89, + 90, + 485, + 243 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 255, + 187, + 271 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Denis Diderot. Diderot's early philosophical works. Number 4. Open Court, 1911. 1", + "[2] Mark O. Riedl. Interactive narrative: A novel application of artificial intelligence for computer games. In AAAI, 2012. 1", + "[3] Mark O Riedl. Human-centered artificial intelligence and machine learning. Human behavior and emerging technologies, 1(1), 2019. 1", + "[4] Chen Liang, Zhicheng Shi, Wenguan Wang, and Yi Yang. Dual reciprocal learning of language-based human motion understanding and generation. In ICCV, 2025. 1", + "[5] Elias Matsas and George-Christopher Vosniakos. Design of a virtual reality training system for human-robot collaboration in manufacturing tasks. IJIDEM, 11, 2017. 1", + "[6] Joon Sung Park, Lindsay Popowski, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Social simulacra: Creating populated prototypes for social computing systems. In UIST, 2022. 1", + "[7] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724, 2023. 1", + "[8] Xiaohan Wang, Wenguan Wang, Jiayi Shao, and Yi Yang. Lana: A language-capable navigator for instruction following and generation. In CVPR, 2023.", + "[9] Hanqing Wang, Wei Liang, Luc V Gool, and Wenguan Wang. Towards versatile embodied navigation. NeurIPS, 2022.", + "[10] Zhichao Zhai, Guikun Chen, Wenguan Wang, Dong Zheng, and Jun Xiao. Taga: Self-supervised learning for template-free animatable gaussian articulated model. In CVPR, 2025. 1", + "[11] Xavier Puig, Kevin Kyunghwan Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. Virtualhome: Simulating household activities via programs. In CVPR, 2018. 2, 3, 5", + "[12] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 2, 3", + "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 6" + ], + "bbox": [ + 93, + 280, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 2, 6", + "[15] Hongwei Yi, Chun-Hao P Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J Black. Human-aware object placement for visual environment reconstruction. In CVPR, 2022. 2, 6", + "[16] Joseph Bates. The role of emotion in believable agents. Communications of the ACM, 37(7), 1994. 2", + "[17] Robert Zubek. Towards implementation of social interaction. In AAAI, 2002. 2", + "[18] Ruth Aylett. Narrative in virtual environments—towards emergent narrative. In AAAI, 1999.", + "[19] Michael Brenner. Creating dynamic story plots with continual multiagent planning. In AAAI, 2010. 2", + "[20] Ho Chit Siu, Jaime Pena, Edenna Chen, Yutai Zhou, Victor Lopez, Kyle Palko, Kimberlee Chang, and Ross Allen. Evaluation of human-ai teams for learned and rule-based agents in hanabi. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, NeurIPS, 2021. 2", + "[21] Michele Colledanchise and Petter Ögren. Behavior trees in robotics and AI: An introduction. CRC Press, 2018. 2", + "[22] Ilshat Umarov, Mikhail Mozgovoy, and Patrick C. Rogers. Believable and effective ai agents in virtual worlds: Current state and future perspectives. International Journal of Gaming and Computer-Mediated Simulations, 4(2), 2012. 2", + "[23] Christopher Berner, Greg Brockman, Brooke Chan, Vicki Cheung, Przemyslaw Debiak, Christy Dennison, David Farhi, Quirin Fischer, Shariq Hashme, Chris Hesse, Rafal Jozefowicz, Scott Gray, Catherine Olsson, Jakub Pachocki, Michael Petrov, Henrique P. d.O. Pinto, Jonathan Raiman, Tim Salimans, Jeremy Schlatter, Jonas Schneider, Szymon Sidor, Ilya Sutskever, Jie Tang, Filip Wolski, and Susan Zhang. Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680, 2019. 2", + "[24] Oriol Vinyals, Igor Babuschkin, Wojciech M. Czarnecki, and et al. Grandmaster level in starcraft ii using multi-agent reinforcement learning. Nature, 575, 2019. 2", + "[25] Matthew Hausknecht, Prithviraj Ammanabrolu, Marc-Alexandre Cote, and Xinyu Yuan. Interactive fiction games: A colossal adventure. In AAAI, 2020. 2", + "[26] Shohei Miyashita, Xinyu Lian, Xiao Zeng, Takashi Matsubara, and Kuniaki Uehara. Developing game ai agent behaving like human by mixing reinforcement learning and supervised learning. In SNPD, 2017.", + "[27] Chen Tessler, Shahar Givony, Tom Zahavy, Daniel Mankowitz, and Shie Mannor. A deep hierarchical approach to lifelong learning in apache. In AAAI, 2017. 2", + "[28] John Laird. It knows what you're going to do: Adding anticipation to a quakebot. In Proceedings of the 2001 Workshop on Intelligent Cinematography and Editing, 2001. 2", + "[29] Dong kyu Choi, Tolga Konik, Negin Nejati, Chunki Park, and Pat Langley. A believable agent for first-person shooter games. In AAAI, 2021. 3", + "[30] Pat Langley, Dongkyu Choi, and Seth Rogers. Interleaving learning, problem solving, and execution in the icarus architecture. Technical report, Stanford University, Center for the Study of Language and Information, 2005. 2" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10761", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In UIST, 2023. 3", + "[32] Yujie Lu, Weixi Feng, Wanrong Zhu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Neurosymbolic procedural planning with commonsense prompting. In ICLR, 2023. 3", + "[33] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In ICML, 2022. 3, 7", + "[34] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. In NeurIPS, 2023. 3, 7", + "[35] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In NeurIPS, 2023.", + "[36] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In ICML, 2024. 3", + "[37] Shima Imani, Liang Du, and Harsh Shrivastava. Math-prompter: Mathematical reasoning using large language models. In ICLR Workshops, 2023. 3", + "[38] Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen McAleer, Albert Q. Jiang, Jia Deng, Stella Biderman, and Sean Welleck. Llemma: An open language model for mathematics. arXiv preprint arXiv:2310.06786, 2023. 3", + "[39] Xiang Lorraine Li, Adhiguna Kuncoro, Jordan Hoffmann, Cyprien de Masson d'Autume, Phil Blunsom, and Aida Nematzadeh. A systematic investigation of commonsense knowledge in large language models. In EMNLP, 2022. 3", + "[40] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In ICRA, 2023. 3", + "[41] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3", + "[42] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 3", + "[43] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In CoRL, 2023. 3", + "[44] Wenlong Huang, Fei Xia, Dhruv Shah, Danny Driess, Andy Zeng, Yao Lu, Pete Florence, Igor Mordatch, Sergey Levine," + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Karol Hausman, et al. Grounded decoding: Guiding text generation with grounded models for robot control. arXiv preprint arXiv:2303.00855, 2023. 3", + "[45] Belinda Z Li, Maxwell Nye, and Jacob Andreas. Implicit representations of meaning in neural language models. In ACL, 2021. 3", + "[46] Adam Roberts, Colin Raffel, and Noam Shazeer. How much knowledge can you pack into the parameters of a language model? In EMNLP, 2020.", + "[47] Yunqiu Xu, Linchao Zhu, and Yi Yang. Mc-bench: A benchmark for multi-context visual grounding in the era of mllms. In ICCV, 2025. 3", + "[48] Vadim Borisov, Kathrin Sessler, Tobias Leemann, Martin Pawelczyk, and Gjergji Kasneci. Language models are realistic tabular data generators. In ICLR, 2023. 3", + "[49] Yew Ken Chia, Lidong Bing, Soujanya Poria, and Luo Si. Relationprompt: Leveraging prompts to generate synthetic data for zero-shot relation triplet extraction. In ACL, 2022. 3", + "[50] Timo Schick and Hinrich Schütze. Generating datasets with pretrained language models. In EMNLP, 2021. 3", + "[51] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. In ACL, 2023. 3", + "[52] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019. 3, 8", + "[53] Hongming Zhang, Xin Liu, Haojie Pan, Yangqiu Song, and Cane Wing-Ki Leung. Aser: A large-scale eventuality knowledge graph. In WWW, 2020. 3", + "[54] Peter West, Chandra Bhagavatula, Jack Hessel, Jena Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. In NAACL, 2022. 3", + "[55] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. In NeurIPS, 2022. 3", + "[56] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 3", + "[57] Yudi Dai, Yitai Lin, Chenglu Wen, Siqi Shen, Lan Xu, Jingyi Yu, Yuexin Ma, and Cheng Wang. Hsc4d: Human-centered 4d scene capture in large-scale indoor-outdoor space using wearable imus and lidar. In CVPR, 2022. 3", + "[58] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. 3", + "[59] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Xuan Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In NeurIPS, 2021. 3" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10762", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, 2023. 4", + "[61] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In ECCV, 2020. 4, 6", + "[62] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, 2022. 4", + "[63] Thomas A Feo and Mauricio GC Resende. Greedy randomized adaptive search procedures. Journal of global optimization, 6, 1995. 5", + "[64] Markus Freitag and Yaser Al-Onaizan. Beam search strategies for neural machine translation. In ACL, 2017. 5", + "[65] Rémi Coulom. Efficient selectivity and backup operators in monte-carlo tree search. In CG, 2006. 5", + "[66] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In ICLR, 2019. 6", + "[67] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In CVPR, 2021. 6", + "[68] Santhosh K. Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westry, Angel Xuan Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. In NeurIPS Datasets and Benchmarks Track, 2021. 6", + "[69] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6", + "[70] Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan. Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021. 6", + "[71] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 6", + "[72] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6", + "[73] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 6", + "[74] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An opensource chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. https://vicuna.lmsys.org, 2023. 8", + "[75] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In CVPR, 2021. 8" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 8", + "[77] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 8" + ], + "bbox": [ + 516, + 92, + 903, + 175 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10763", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/099d59e1-3cf8-46ca-a7c2-5f8c85602d06_model.json b/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/099d59e1-3cf8-46ca-a7c2-5f8c85602d06_model.json new file mode 100644 index 0000000000000000000000000000000000000000..075cc2fa6af4886e2d751c0e740fd8f6e6f61b40 --- /dev/null +++ b/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/099d59e1-3cf8-46ca-a7c2-5f8c85602d06_model.json @@ -0,0 +1,2686 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.9, + 0.152 + ], + "angle": 0, + "content": "Towards Human-like Virtual Beings: Simulating Human Behavior in 3D Scenes" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.181, + 0.668, + 0.201 + ], + "angle": 0, + "content": "Chen Liang, Wenguan Wang*, Yi Yang" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.209, + 0.694, + 0.224 + ], + "angle": 0, + "content": "State Key Lab of Brain-Machine Intelligence, Zhejiang University" + }, + { + "type": "text", + "bbox": [ + 0.322, + 0.233, + 0.672, + 0.248 + ], + "angle": 0, + "content": "https://github.com/leonnnop/BehaviorHub" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.283, + 0.284, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.394, + 0.276, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.297, + 0.321, + 0.312 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.332, + 0.547, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.49, + 0.753, + 0.505 + ], + "angle": 0, + "content": "Human Behavior Simulation in 3D Scenes (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.285, + 0.747, + 0.315 + ], + "angle": 0, + "content": "High-level, Long-horizon, Abstract Goal: Preparing for work from home" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.321, + 0.904, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.509, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Figure 1. Previous research has primarily focused on: (a) linguistic-based short-horizon task planning, and (b) low-level human-scene interaction. (c) This study investigates the simulation of high-level, long-horizon, abstract goal-driven human behaviors in 3D scenes." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.557, + 0.327, + 0.573 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.589, + 0.486, + 0.877 + ], + "angle": 0, + "content": "Building autonomous agents that can replicate human behavior in the realistic 3D world is a key step toward artificial general intelligence. This requires agents to be holistic goal achievers and to naturally adapt to environmental dynamics. In this work, we introduce ACTOR, an agent capable of performing high-level, long-horizon, abstract goals in 3D households, guided by its internal value similar to those of humans. ACTOR operates in a perceive-plan-act cycle, extending the ungrounded, scene-agnostic LLM controller with deliberate goal decomposition and decision-making through actively searching the behavior space, generating activity choices based on a hierarchical prior, and evaluating these choices using customizable value functions to determine the subsequent steps. Furthermore, we introduce BEHAVIORHUB, a large-scale human behavior simulation dataset in scene-aware, complicated tasks. Considering the unaffordable acquisition of human-authored 3D human behavior data, we construct BEHAVIORHUB by exploring the commonsense knowledge of LLMs learned from large" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.909, + 0.651 + ], + "angle": 0, + "content": "corpora, and automatically aligning motion resources with 3D scene for knowledgeable generation. Extensive experiments on our established benchmark demonstrate that the proposed architecture leads to effective behavior planning and simulation. BEHAVIORHUB also proves beneficial for downstream task development." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.68, + 0.645, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.706, + 0.909, + 0.904 + ], + "angle": 0, + "content": "Building autonomous agents (e.g., virtual beings, or humanoid robots) that can replicate human behavior in the realistic 3D world, has been a long-standing pursuit since the inception of AI [1]. The study could empower nonplayer game character [2], underpin human-robot interaction [3, 4] and cooperation [5], populate virtual reality communities [6], and accelerate Embodied AI [7-10]. Much progress has been made in vision and language models that imitate human motions (cf. Fig. 1 (b)) or propose linguistic plans (cf. Fig. 1 (a)). However, an effective 3D humanoid agent must go beyond by conquering three major barriers: (i) Holistic goal achievement from perception to action. As depicted in Fig. 1 (c), to accomplish a high-level goal (e.g.," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.889, + 0.327, + 0.902 + ], + "angle": 0, + "content": "*Corresponding author: Wenguan Wang." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10753" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.379 + ], + "angle": 0, + "content": "'prepare for work'), the agent must process the perceived information (e.g., 'lying on bed'), decompose the goal into a series of activities (e.g., 'get up', 'use toilet', 'eat breakfast', etc.), and devise appropriate action plans for each; (ii) Environmental dynamics. Agents should be able to actively adjust their plans based on the environment, e.g., determining if the bathroom is occupied when planning to use toilet; (iii) Vast and multifaceted human behavior space, where numerous viable paths exist to achieve even a single goal. For example, agents preparing for work can choose to eat breakfast before using toilet or vice versa. Also, when the bathroom is occupied, an intelligent agent can decide to whether eat breakfast first, or continue waiting, depending on its state and beliefs, e.g., the desire to 'complete goal as soon as possible' or 'save energy costs'. A competent agent must process such value priorities to guide its selection and evaluation of actions and policies. Besides, the absence of a comprehensive testbed further poses great challenge for agent development and evaluation, constituting barrier (iv)." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.385, + 0.482, + 0.806 + ], + "angle": 0, + "content": "In this work, we present ACTOR - a large language model (LLM) powered agent towards diligent simulation of human behavior in 3D realistic scenes (§4). ACTOR follows a perceive-plan-act cycle, addressing challenge \\((i)\\) as envisioned. Using LLM as a central controller, it strategizes plans by searching the human behavior space. It actively maintains a tree of behaviors, where each node represents an intermediate step toward holistic goal achievement. The construction of this tree is guided by a hierarchical prior, i.e., executable low-level actions are grouped into high-level semantic units, called activities, iteratively forming a hierarchical structure (cf. Fig. 1 (c)). The search progress is assessed using a set of customizable value functions that determine the likelihood of different intermediate candidates. In addition to being rational in common sense, ACTOR couples real-valued evaluations (e.g., best efficiency) with personalized priors expressed through language commands (e.g., description of a neat person). These outputs are converted into unified probabilities, allowing for the incorporation of the agent's characteristics and beliefs as value functions, thus addressing challenge (iii). The planning process is dynamic and grounded in specific environmental values, empowering ACTOR to readjust its plans when faced with environmental changes or new language commands, effectively tackling challenge (ii). This formulation facilitates the use of powerful search algorithms, e.g., greedy search, beam search, and Monte Carlo tree search (MCTS), with MCTS proving to be the most effective in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.811, + 0.482, + 0.9 + ], + "angle": 0, + "content": "Furthermore, we establish a comprehensive environment for development and evaluation of agents like ACTOR, based on our newly proposed large-scale, scene-aware, behavior-rich dataset, dubbed BEHAVIORHUB (§5). One critical issue in constructing the human-authored benchmark is the high cost associated with acquiring and scaling" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.408 + ], + "angle": 0, + "content": "high-quality, human-generated daily behavior data. Moreover, annotating large-scale behavior data further requires creativity in designing novel tasks and expertise in creating complete plans from scratch, which is also a challenging task for humans [11]. Given this context, we propose to automatically synthesize 3D human behavior data by enhancing existing resources. Initially, we distill the tree-structured linguistic plans of human daily behaviors from LLM using in-context learning [12]. Contactable objects and plausible interactions from the scanned 3D scenes (e.g., ScanNet [13], etc.) and captured motion sequences (e.g., AMASS [14], etc.) are attributed into the in-context prompt to ensure viable plans grounding in certain environment. Then, we align the task-motion-scene data triplets by applying collision and contact constraints [15] for valid translation and rotation parameters. To promote diversity, we sample multiple plausible motion sequences for each high-level goal. In total, BEHAVIORHUB contains more than 1k daily goals over 10k high-quality behavior samples of 15.7 steps on average covering 1.5k scenes, that establishes a comprehensive testbed, that addresses the barrier \\((iv)\\)." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.41, + 0.905, + 0.574 + ], + "angle": 0, + "content": "We conduct extensive evaluations of ACTOR in §6. First, we find that ACTOR produces admissible plans and generates plausible motion sequences to simulate abstract, temporally-extended human behaviors. It outperforms the strong baselines by nearly doubling the overall success rate. This result is further confirmed by human evaluation. Then, we conduct several ablative experiments that limit ACTOR's access to each core design for thorough assessment. Finally, experiments on scene-aware and language-conditioned human motion generation demonstrate how BEHAVIORHUB can benefit the development of downstream task models." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.589, + 0.654, + 0.604 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.614, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Human Behavior Simulation. Simulating human behavior in realistic, open world environment like the one we inhabit is a long-standing topic in artificial intelligence [16]. Historically, the topic has primarily been studied in the game worlds, focusing on enhancing player experiences through intelligent non-player game characters (NPCs) [17-19]. Early approaches rely on rule-based approaches like finite-state machines [20] and behavior trees [21]. They provide a brute force way of manually crafting agent's behaviors, but cannot perform new procedures that were not hard-coded in their script [22], limiting the generalizability. Another strand [23, 24] involves using reinforcement learning, where agents learn its own policy through optimizing the learning algorithm on readily definable rewards over fixed task space, which is often limited to non-open, adversarial games, or blocks worlds only [25-27]. The third line of research, represented by some pioneer works [28-30] in computational cognition, aims to build machines that operate directly in perceive-plan-act cycles, encompassing" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10754" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "the nature of autonomous agents as originally envisioned. This formulation holds potential generalizability to most, if not all, open-world contexts. However, these studies are typically restricted to simplified environments, such as first-person shooter games [29], or 2D gridworlds (e.g., Generative Agents [31]), and focus on a reduced range of behaviors. Our work falls in vein of third category, while pushing frontier towards simulation in realistic 3D environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.483, + 0.426 + ], + "angle": 0, + "content": "Recently, beyond the game world, new trending on social embodied intelligence, such as assistive robots, draws attention to simulation of human collaborators' behavior. One notable effort is VirtualHome [11], which studies high-level human activities as plain sequences of atomic actions. However, its formulation is inherently environment-agnostic, which has been treated as a purely linguistic procedural planning problem in subsequent studies [32, 33]. Furthermore, the discrete action space built upon manually crafted procedural knowledge also makes it an insufficient testbed. In contrast, with a special focus on environment-aware simulation of contiguous daily behaviors, our work paves one solid step accelerating the development of social embodied intelligence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.793 + ], + "angle": 0, + "content": "LLM as Planner. Remarkable progress in LLMs has been observed, demonstrating their emerging capacity to break down complex tasks into more manageable sub-tasks and devise appropriate plans for each [34-36]. LLMs have been successfully applied to solving mathematical problems [37, 38], reasoning on commonsense [39], planning robotics tasks [40, 41], and very recently, manipulating external APIs on a web scale, expanding their capabilities beyond text generation. Based on the observation, we posit that LLMs can serve as a crucial component in extending the perception-decision-action space to construct human-like agents in realistic 3D environments. However, a crucial challenge remains: LLMs lack experience and interaction with their environment [12, 42], preventing them from ordering actionable and rational plans. This paper addresses this issue by incorporating customizable value functions into LLMs to evaluate and prioritize plans. The idea bears some resemblance to recent robotics research [43, 44]. However, those methods necessitate retraining for every new set of primitive robotic skills, making them impractical for the complex and undefined human action space, which is difficult to define in advance. In contrast, we support plan evaluation using real-numbered functions and language-based rules without resource-intensive retraining." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.904 + ], + "angle": 0, + "content": "LLM as Data Generator. Being trained on the large corpus of human-produced language, LLMs are believed to contain a wealth of information about the world [45-47]. Given a handful of task-specific prompts, LLMs can generalize and generate more linguistic data in the same format, with the application of generating tabular data [48], relation triplets [49], sentence pairs [50], instruction data [51]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.289 + ], + "angle": 0, + "content": "etc. The idea seems naturally to be borrowed for human behavior data acquisition, where the requested human motion and daily activity procedure were commonly crowdsourced with high expense and complexity [11], limiting the scale and coverage of related datasets [52]. However, the process is non-trivial. The generated data is often blamed for low quality and diversity issues [53, 54]. The 3D environment-aware nature of the task also poses unique challenges. To respond, we explore attributed prompts specifically conditioned on the environment that not only mitigates the problems of low informativeness and redundancy, but also offers an effective workflow that can further empower other related domains, such as human-scene interaction [55]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.304, + 0.65, + 0.318 + ], + "angle": 0, + "content": "3. Environment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.905, + 0.405 + ], + "angle": 0, + "content": "Simulating open-ended goals that resemble naturalistic human behaviors necessitates an environment capable of facilitating diverse agent affordances and interactions. Before delving into the detailed agent architecture, we first describe such environment we tailored for agents to instantiate in." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.406, + 0.905, + 0.435 + ], + "angle": 0, + "content": "Environmental Setup. The environment features common affordances in a household, including:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.436, + 0.905, + 0.465 + ], + "angle": 0, + "content": "- Scene of 3D textured meshes, that spans a house of multiple functional areas (e.g., kitchen, etc.)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.467, + 0.905, + 0.512 + ], + "angle": 0, + "content": "- Objects embedded in the scene (e.g., stove in kitchen, bed in bedroom), with each constructed of 3D textured geometry and corresponding object state (e.g., fridge: opened)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.512, + 0.905, + 0.587 + ], + "angle": 0, + "content": "- Humanoid agent(s) defined by SMPL-X [56], an expressive 3D human model of shape and pose of both body and hand. Agents reside in and interact with the scene and can influence the state of objects by their actions (e.g., close the fridge | fridge: opened \\(\\rightarrow\\) closed)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.436, + 0.905, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.905, + 0.633 + ], + "angle": 0, + "content": "As a start point, we demonstrate the environment of indoor household, keep the general formulation open for other environments, such as outdoor streets [57]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.906, + 0.785 + ], + "angle": 0, + "content": "Environmental Simulator. Interactions within the environment are driven by the simulator of two components: (i) an engine that manages and evolves the states of objects in the environment; and (ii) a renderer that supports generation of multiple perceptual observations (e.g., RGB, depth, 3D surfaces) for agents. At each time step, the simulator dynamically updates environment and collects egocentric, surround, or third-person-view information based on needs. We build our simulator upon Habitat-Sim [58, 59] to ensure efficient and parallelizable simulation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.659, + 0.817 + ], + "angle": 0, + "content": "4. ACTOR Agent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.903 + ], + "angle": 0, + "content": "We aim to build humanoid agents that naturally simulate 3D human behaviors to complete daily goals, which can be lengthy, abstract, or ambiguous. Agents receive high-level goal described in language and are tasked with generating plausible motion sequences that align with the given scene." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10755" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.09, + 0.905, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.245, + 0.896, + 0.26 + ], + "angle": 0, + "content": "Figure 2. Left: Overall agent architecture (§4.1); Right: One-step in perceive-plan-act loop with value-driven behavior planning (§4.2)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.275, + 0.276, + 0.291 + ], + "angle": 0, + "content": "4.1. Agent Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.297, + 0.484, + 0.432 + ], + "angle": 0, + "content": "Fig. 2 shows the workflow of ACTOR. The agent operates in a perceive-plan-act loop, using language as the generic interface connecting the three phases. At each step, it selectively perceives the world based on the target, transforming information into an environment description using heuristics. The observed information assists in decision-making through the LLM core, ultimately resulting in actions in the form of 3D motion sequences based on linguistic plans within the environment. We now delve into the details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.554 + ], + "angle": 0, + "content": "Perception. Our definition of perception extends beyond gathering information, such as 2D images or 3D point clouds, through the simulator (§3). It encompasses attaining a deep understanding of the environment, including object properties, spatial relationships, and scene layouts, etc. In our preliminary implementation, information is perceived using readily available models and converted into linguistic descriptions using heuristic functions (supp. §A.2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Plan. At the core of ACTOR, a planning module receives the current environment and past behavior trial as input, generating action description as output. For the executable actions, it provides descriptions of motion sequences. In the case of high-level activity, it bypasses the action stage and proceeds to break down the target further. We will provide a detailed explanation of this process in §4.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.796 + ], + "angle": 0, + "content": "Action. For action, we specifically consider whole-body human actions in 3D scenes to closely resemble human behavior using the off-the-shelf models. It generates whole-body motion based on text and trajectory [60]. For object-interactive actions, we further refine hand grasping using an isolated grasp estimation model [61]. Additionally, for moving actions like walking, trajectory paths are pre-estimated [62]. Detailed trajectory estimation process is provided in supp. §A.3." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.804, + 0.376, + 0.82 + ], + "angle": 0, + "content": "4.2. Value-driven Behavior Planning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "An LLM, denoted as \\( p_{\\theta} \\) with parameters \\( \\theta \\), functions as a controller to iteratively decompose the long-term goal \\( l \\) described in language texts into shorter steps of behaviors based on the partial environmental dynamics. To prevent confusion with the global environment \\( E \\), we denote the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.907, + 0.456 + ], + "angle": 0, + "content": "partial environmental observation at step \\( i \\) as \\( E_{i}^{\\prime} \\) and its corresponding textual description as \\( e_{i} \\). For brevity, we omit step indicator \\( i \\) and define \\( E^{\\prime} \\) to include both the environment state and agent's state. We denote textual descriptions of behaviors as \\( z \\). The problem is formulated as a search over a tree, where each node is a state \\( s = \\{z_{1\\dots i}\\} \\), representing a partial trial with the input and the sequence of behaviors taken thus far. The instantiation consists of three components: (i) Node expansion, which involves generating \\( k \\) candidate behaviors for the next step search; (ii) Value functions to evaluate each node; (iii) Search algorithm that accounts for branch selection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.459, + 0.909, + 0.641 + ], + "angle": 0, + "content": "Node Expansion. The node is expanded through sampling from LLM with a window size \\( w \\): \\( z^{(j)} \\sim p_{\\theta}(z_{i+1}|l,e,s,h) \\), where \\( j \\in \\{1,\\dots,w\\} \\). We omit the basis prompting parts for brevity, which is also the case in the subsequent context. Here, we introduce a hierarchical heuristic, denoted as \\( h \\), which is implemented through prompt instructions and demonstrations of specific actions and activity cases. This heuristic ensures that during each expansion, all candidates at the same level are restricted to executable actions or high-level semantic units of activities, that provides a more nuanced representation of behaviors for effectively modeling interchangeable activities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.885 + ], + "angle": 0, + "content": "Value Function. The value function assesses the state by considering the degree to which a specific behavior contributes to the achievement of the target, conditioned on the agents' beliefs reflecting their value. The search algorithm uses the output to determine which nodes to explore further and in what order. The likelihood is given by \\( p(z_{i+1} | l, E', s) \\propto p_{\\theta}(z_{i+1} | l, e, s) \\cdot p_v(z_{i+1} | l, E', s) \\). Here, the LLM provides us with \\( p_{\\theta}(z_{i+1} | l, e, s) \\), which represents the likelihood, based on commonsense, that a textual behavior is a valid next step. However, the LLM struggles to generalize or make inferences in the real environment since it is not grounded. On the other hand, \\( p_v \\) provides the likelihood of the behavior being plausible in the current state of both the environment and agent, according to the defined values. We consider a set of values \\( \\{p_{v_n}\\}_n \\) categorized into two types, and \\( p_v(z_{i+1} | l, E', s) = \\prod_n p_{v_n}(z_{i+1} | l, E', s) \\):" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Real-valued function assigns real values as outputs, such" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10756" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.361, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.091, + 0.905, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.27, + 0.905, + 0.285 + ], + "angle": 0, + "content": "Figure 3. (a) An illustrative example of our BEHAVIORHUB dataset; (b) Semi-automatic 3D human behavior data generation pipeline (\\(\\S 5\\))." + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.302, + 0.482, + 0.346 + ], + "angle": 0, + "content": "as the shortest path value, for which the function estimates the distance for each candidate action. The outputs of real-valued functions are normalized into probabilities." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.347, + 0.483, + 0.437 + ], + "angle": 0, + "content": "- Language-based command is implemented by prompting the LM with a value prompt that conveys the meaning of 'How likely is it for someone who is a neat person to take the following action, considering ...'. The model then reasons about each action to generate the corresponding probability reflecting the value." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.574 + ], + "angle": 0, + "content": "Values are also conditioned on the state of the environment and agent. This approach allows for reacting to environmental dynamics in addition to the active planning process. Search Algorithms. Owing to the active planning process we have formulated, it is feasible to use different search algorithms based on the tree structure. We explore greedy search [63], beam search [64] and MCTS [65], and evaluate their performance in experiments (cf. §6.2) where we find MCTS performs best. We therefore use MCTS by default." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.583, + 0.303, + 0.599 + ], + "angle": 0, + "content": "4.3. Implementation Detail" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.803 + ], + "angle": 0, + "content": "We employ GPT-4 and GPT-3.5-turbo API provided by OpenAI as base LLMs. We prompt them with four in-context examples as demonstrations, with decoding temperature as 0 for more deterministic generation. We use official releases of conditional motion generation models and fine-tune them on BEHAVIORHUB using the default parameters for each to promote finer generation quality. By default, we set the window size to \\( w = 5 \\) and normalize the shortest path value based on the maximum rollout depth of the search algorithm, which is set to 3. In practice, we find one-time sampling is sufficient for effective tree search, eliminating the need for multiple samplings for branch aggregation. Our algorithm is implemented in PyTorch using Tesla A40 GPUs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.816, + 0.353, + 0.831 + ], + "angle": 0, + "content": "5. BEHAVIORHUB Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To lay a solid foundation for future research, we build a large repository of common behaviors performed in daily household scenarios. Each sample contains three components: \\((i)\\) a high-level goal (i.e., root node); \\((ii)\\) a tree" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.302, + 0.907, + 0.513 + ], + "angle": 0, + "content": "structured linguistic plan covering necessary intermediate-level steps (i.e., intermediate nodes) and low-level steps (i.e., leaf nodes) required to accomplish the goal; and (iii) scene-conditioned human motions corresponding to each executable step at either intermediate or low level. We illustrate one example in Fig. 3 (a). Intermediate step set can be empty. Both goal and steps in plans are human activities described with either concrete or abstract language (e.g., 'go to sleep on bed' or 'feel tired'). To generate diverse and high-quality data, we employ a two-step pipeline that utilizes an LLM pretrained on extensive corpora: First, automatic generation of linguistic daily plans (\\(\\S 5.1\\)); Second, alignment of the plans with 3D motions and scenes (\\(\\S 5.2\\)). The entire data generation process is shown in Fig. 3 (b)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.527, + 0.803, + 0.543 + ], + "angle": 0, + "content": "5.1. Linguistic Goal-Plan Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.906, + 0.612 + ], + "angle": 0, + "content": "The pipeline consists of three steps: \\((i)\\) generate potentially incomplete goal-plan trees; \\((ii)\\) complete and refine each tree; and \\((iii)\\) filter out low-quality data. We provide detailed prompts in supp. §A.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.853 + ], + "angle": 0, + "content": "Attributed Goal-Plan Tree Generation. In the initial step, we generate new goal-plan trees using a bootstrapping approach based on a small set of seed human-written samples. To ensure broader coverage and facilitate later alignment with specific scenes, we attribute the starting room, candidate objects and actions that interact with the activity in the plans into prompt demonstrations. We start the sample pool with 292 activities from ActivityPrograms [11], and consider 21 room types (e.g., bathroom, gym), \\(\\sim 10^3\\) object types (e.g., TV, table), and \\(\\sim 10^2\\) action types (e.g., open, eat). Refer to the supp. §A.5 for the full list. For each step, we sample eight goal-plans as in-context examples from the pool while restricting the outputs to one room and ten types of objects. Out of the eight plans, six are from the human-written plans, and two are generated by the model in previous steps to enhance diversity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Generated trees are then labeled with sequential order for intermediate nodes in the trees, which encompass multiple sub-step leaf nodes of individual actions. This sequential or" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10757" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "der is represented as interchangeable groups, where nodes within the same set are interchangeable with each other. We prompt the LM in a few-shot way to determine this." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.482, + 0.258 + ], + "angle": 0, + "content": "Goal-Plan Tree Refinement. We further improve the constructed trees by addressing two key aspects: (i) we complete any missing internal plan steps, which can often be revised based on commonsense, e.g., opened the fridge without closing it. (ii) we enhance the root node descriptions to be more abstract, e.g., transforming 'use toilet' to 'feel the call of nature'. We accomplish both aspects by querying the LLM for suggestions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.482, + 0.349 + ], + "angle": 0, + "content": "Filtering. To promote diversity, a new tree is added to the pool only if its BERTScore similarity [66] with any existing goal-plan tree is below 0.5. Additionally, we utilize LLM to assess the generated trees by asking the question, 'Is this a valid plan?' Any plans flagged as 'invalid' are filtered out to ensure high-quality data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.36, + 0.362, + 0.377 + ], + "angle": 0, + "content": "5.2. Goal-Motion-Scene Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Once the goal-plan trees are formed, we proceed to ground them in the 3D environment. We propose to synthesize 3D behaviors by leveraging existing resources of human motions (i.e., AMASS [14], BABEL [67], GRAB [61]), and indoor scenes (i.e., ScanNet [13], HM3D [68], Replica [69]). For each tree, we first sample actions and objects from resources using corresponding labels to fulfill the executable activities in forms of combinations of actions on objects. Then, given sampled motion, we aim to generate plausible and contiguous interactions to sampled object in scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.67 + ], + "angle": 0, + "content": "Motion Alignment. We first put motion clips into the scene anchored by the contactable objects. We optimize the translation and rotation parameters matrices by minimizing the collision and contact losses from MOVER [15]. A unified SDF volume is calculated, and all contact vertices for all frames are accumulated in 3D space. The motion is aligned through a transformation of the two matrices. This joint optimization improves human-object contact and resolves 3D interpenetrations between humans and the scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.671, + 0.483, + 0.776 + ], + "angle": 0, + "content": "Sequence Blending. The aligned motion sequence, which may be sparse and not spatially connected, is blended using a Transformer-based motion completion method [70]. This results in a contiguous motion sequence that aligns with both the scene and goal plan. In practice, we find a single network is capable of delivering satisfactory results that align with prior research [71]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.777, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Verification. Finally, to ensure the data quality, each example is examined by three verifiers who vote on whether the plan (motion) is complete (valid). If an example receives majority approval, it is accepted; otherwise, it is dropped." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.26, + 0.864 + ], + "angle": 0, + "content": "5.3. Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our dataset consists of 10k human behavior samples in 1.5k 3D scenes, covers 2k unique activity over 0.1k actions and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "1k objects, which is an order of magnitude larger than the human-authoring ActivityPrograms [13]. On average, each high-level goal has 15.7 steps, resulting in a total of 10.1 corresponding motion sequences of 83.3 frames, that span over 8.6M motion frames. Each activity corresponds to 4.1 different motion sequences in 1.7 rooms. Owing to the automatic data generation pipeline, our dataset achieves even higher diversity and coverage with low human cost." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.224, + 0.691, + 0.239 + ], + "angle": 0, + "content": "5.4. Evaluation Metric" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.905, + 0.322 + ], + "angle": 0, + "content": "Our evaluation of the plausible behavior simulation encompasses two key aspects: \\((i)\\) the simulation should generate linguistically reasonable behavior plans (i.e., behavior planning); and \\((ii)\\) align them with natural motion sequences within the 3D environment (i.e., behavior simulation)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.323, + 0.905, + 0.413 + ], + "angle": 0, + "content": "- For behavior planning, we use Sentence-BLEU [72] and BERTScore [66] to measure the semantic similarity between the ground-truth plans and predictions. We report maximum scores attained across ground-truth plan variants to ensure the evaluation be order-invariant with respect to interchangeable sub-steps." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.413, + 0.906, + 0.672 + ], + "angle": 0, + "content": "- For behavior simulation, we consider three sets of metrics: (i) Success Rate: The step success rate (SSR) records the percentage of steps where the agent successfully completes the step objective, defined by a contact distance threshold. For example, we consider lying down to be successful if both the hip and head of the humanoid are within \\(30\\mathrm{cm}\\) of the target location [73]. The goal success rate (GSR) is measured to determine whether all steps in the entire plan are successfully executed; (ii) Goal Success Rate Weighted by Path Length (GSRPL): It judges how efficient was the agent at finishing the goal, defined as \\(\\mathrm{GSR} \\cdot \\frac{g}{\\max(g,l)}\\). Here \\(g\\) is ground-truth path length and \\(l\\) is the agent's path length; (iii) Motion Quality: We evaluate overall quality of generated motions using Frechet Inception Distance (FID) and recognition accuracy measured with the final layer of a pretrained standard RNN action recognition classifier as motion feature extractor." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.323, + 0.906, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.687, + 0.638, + 0.703 + ], + "angle": 0, + "content": "6. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.878, + 0.727 + ], + "angle": 0, + "content": "6.1. Performance on BEHAVIORHUB Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Dataset Split and Dynamic Subset. We randomly sample 200 held-out goals as demonstration set from which we select example(s) for prompting language models, and also as training data for fine-tuning conditional generation models. The remaining ones are used for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To thoroughly examine the capacity of benchmarked models in managing environmental dynamics, we manually create a subset of 300 samples from the original BEHAVIORHUB, called Dynamic subset. In each sample, we carefully configure the agent's environment state-aware triggers to guarantee a distinct goal-plan solution for a specific goal," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10758" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.089, + 0.905, + 0.228 + ], + "angle": 0, + "content": "
MethodBehavior PlanningBehavior Simulation
S-BLEUBERT-SSSRGSRGSRPLFID ↓Accuracy ↑
Main SetLLMaP [33] [ICML2022]0.0890.825-----
HuggingGPT [34] [NeurIPS23]0.1320.8560.5330.3170.1615.3860.620
ACTOR (Ours)0.1700.8790.6010.4720.3512.0870.773
Dynamic SubsetHuman0.2030.959-----
LLMaP [33] [ICML2022]0.0690.821-----
HuggingGPT [34] [NeurIPS23]0.0990.8300.4070.1640.0739.1160.505
ACTOR (Ours)0.1350.8620.5150.3060.2123.1410.697
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.232, + 0.905, + 0.246 + ], + "angle": 0, + "content": "Table 1. Quantitative results on Main and Dynamic set of BEHAVIORHUB. \\( \\downarrow \\) ’ indicates smaller values are better. See §6.1 for details." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.254, + 0.238, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.254, + 0.542, + 0.277 + ], + "angle": 0, + "content": "Original plan: weekend cleaning \\(\\rightarrow\\) clean bedroom \\(\\rightarrow\\) Strategized plan:" + }, + { + "type": "image", + "bbox": [ + 0.259, + 0.277, + 0.559, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.302, + 0.59, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.313, + 0.624, + 0.333 + ], + "angle": 0, + "content": "(occupied signal removed)" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.277, + 0.698, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.276, + 0.778, + 0.285 + ], + "angle": 0, + "content": "Clean bedroom" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.287, + 0.8, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.287, + 0.826, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.826, + 0.287, + 0.899, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.349, + 0.2, + 0.361 + ], + "angle": 0, + "content": "Goal: prepare for th" + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.349, + 0.49, + 0.361 + ], + "angle": 0, + "content": "Language command: he/she is a neat person" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.361, + 0.197, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.361, + 0.299, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.361, + 0.399, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.361, + 0.5, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.361, + 0.601, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.361, + 0.7, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.361, + 0.8, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.361, + 0.899, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.435, + 0.888, + 0.449 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of ACTOR on the challenging BEHAVIORHUB Dynamic subset. Please refer to §6.1 for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.529 + ], + "angle": 0, + "content": "e.g., we designate the bathroom as occupied only after certain pre-request steps have been fulfilled; or incorporate language commands that specify agent characteristics, further contributing to a unique goal-plan preference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.64 + ], + "angle": 0, + "content": "Competitors. We benchmark two top-leading LLM-based models: LLMaP [33], a procedural planning model, and HuggingGPT [34], a general tool agent, to probe the human behavior simulation ability in existing techniques. LLMaP is designed to operate solely on textual inputs, and lacks the capability to perform behavior simulation. We assess its behavior planning ability only." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Quantitative Evaluation. Table 1 summarizes the automatic evaluation results, from which we take three major observations: (i) ACTOR achieves the best results on both behavior planning and simulation and outperforms the strong baseline by nearly doubling the GSR, indicating the effectiveness of our value-driven behavior planning design. (ii) On the challenging Dynamic subset, both LLMaP and HuggingGPT experience a noticeable decline in terms of BERT-S, whereas our ACTOR performs much better. To illustrate, in the case shown in Fig. 4, both LLMaP and HuggingGPT directly ignore the occupied signal and give an irrational plan of heading directly to the bedroom, causing a failure. This observation confirms our claim that the ungrounded LLM falls short in terms of environment-aware behavior planning. (iii) Although our ACTOR shows more promising results, there still remains a significant gap compared to human performance. This highlights the need for" + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.466, + 0.902, + 0.548 + ], + "angle": 0, + "content": "
MethodComplete.Rational.Quality
Human4.024.85-
LLMaP [ICML2022]2.232.48-
HuggingGPT [NeuIPS23]2.712.893.18
ACTOR (Ours)3.053.473.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.552, + 0.899, + 0.566 + ], + "angle": 0, + "content": "Table 2. Human-evaluated results (§6.1) on Dynamic subset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.905, + 0.594 + ], + "angle": 0, + "content": "developing more sophisticated behavior simulation models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.906, + 0.883 + ], + "angle": 0, + "content": "Human Evaluation. For a comprehensive evaluation, we engage five participants as human evaluators. They rate the model performance based on three aspects: (i) Completeness: assess whether the motion steps can successfully complete the target goal, capturing semantic completeness; (ii) Rationality: evaluate whether the sequence includes necessary steps in the correct sequential order to accomplish the target goal, capturing sequential order correctness; and (iii) Quality: reflect the naturalness and smoothness of the motion sequence, capturing motion quality. The results of the human ratings, based on a 5-point Likert scale, are reported in Table 2. The human subjective judgments generally align with the trends reflected by Table 2, confirming the reliability of our constructed automatic evaluation framework. Also, the results reaffirm that ACTOR yields plausible behavior simulation. However, it is worth noting that human-written plans are consistently preferred over our results, underscoring the challenging nature of our newly proposed BEHAVIORHUB benchmark." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Qualitative Analysis. Examples from ACTOR on BEHAV" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10759" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.401, + 0.17 + ], + "angle": 0, + "content": "
MethodBERT-SGSRGSRPL
Baseline0.8110.1400.062
+ Active Search0.8370.2350.132
+ Hierarchical Prior0.8490.2610.154
+ Value Function0.8620.3060.212
" + }, + { + "type": "table_caption", + "bbox": [ + 0.168, + 0.172, + 0.325, + 0.184 + ], + "angle": 0, + "content": "(a) Key Component Analysis" + }, + { + "type": "table", + "bbox": [ + 0.409, + 0.104, + 0.648, + 0.17 + ], + "angle": 0, + "content": "
AlgorithmBERT-SGSRGSRPL
Greedy0.8400.2440.151
Beam0.8530.2870.186
MCTS0.8620.3060.212
" + }, + { + "type": "table_caption", + "bbox": [ + 0.468, + 0.172, + 0.583, + 0.184 + ], + "angle": 0, + "content": "(b) Search Algorithm" + }, + { + "type": "table", + "bbox": [ + 0.659, + 0.104, + 0.899, + 0.17 + ], + "angle": 0, + "content": "
LLMBERT-SGSRGSRPL
Vicuna-7b0.8080.0920.063
GPT-3.50.8330.1760.116
GPT-40.8620.3060.212
" + }, + { + "type": "table_caption", + "bbox": [ + 0.712, + 0.172, + 0.837, + 0.184 + ], + "angle": 0, + "content": "(c) Modular Scalability" + }, + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.189, + 0.874, + 0.203 + ], + "angle": 0, + "content": "Table 3. Ablative experiments on the Dynamic subset of our proposed BEHAVIORHUB. Please refer to §6.2 for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.215, + 0.484, + 0.366 + ], + "angle": 0, + "content": "IORHUB Dynamic are visualized in Fig. 4. In addition to simulating behavior to strategize and successfully achieve the desired goal, ACTOR is able to respond to environmental changes and adapt to language commands. For instance, to accomplish the goal of 'Sunday cleaning', while someone is still in bed, i.e., the bedroom is occupied, the agent prioritizes scheduling the kitchen first and then the bedroom, waiting for the person to wake up. In the second example, considering the truth that someone is a neat person, he/she engages in home cleaning before eating breakfast." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.378, + 0.312, + 0.394 + ], + "angle": 0, + "content": "6.2. Diagnostic Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.483, + 0.446 + ], + "angle": 0, + "content": "Ablative studies are conducted on BEHAVIORHUB Dynamic for indepth analyzing each component in ACTOR, using BERT-S, GSR, GSRPL (cf. §5.4) as evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.484, + 0.704 + ], + "angle": 0, + "content": "Key Component Analysis. We first validate the importance of our proposed components by attaching them one at a time in Table 4b. The \\(1^{st}\\) row reports the result of a bare baseline model, which produces a global plan based on the given linguistic goal. Next, in the \\(2^{nd}\\) row, we transition from one-pass planning to active tree search, resulting in improved performance and supporting our claim that active planning can mitigate the impact of environmental changes. Moreover, the \\(3^{rd}\\) row gives the score when the hierarchical behavior structure prior is employed when spanning searching branches. As seen, this leads to moderate improvement by constraining the search space with interchangeable semantic units, highlighting its necessity in handling vast and complex human behavior space. Finally, as shown in the \\(4^{th}\\) row, incorporating value function significantly enhances the overall success rate and success rate over path length, aligning with the rational preference for the shortest path." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Search Algorithm. Table 3b reveals the impact of search algorithms (§4.2), i.e., greedy search, beam search, and MCTS, with window sizes of 5 for the later two algorithms. The default strategy, MCTS, shows optimal results, which aligns with the widely accepted understanding that MCTS is more effective when dealing with a large solution space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Modular Scalability. In Table 3c, we investigate the modular scalability by employing various LLM cores, including GPT-3.5, GPT-4 and open-source LLM Vicuna-7b [74]. As the LLMs' capabilities improve, ACTOR exhibits a continuous enhancement in its behavior simulation performance. This observation substantiates the notion that our ACTOR system possesses great potential for accommodating the de" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.212, + 0.907, + 0.264 + ], + "angle": 0, + "content": "
MethodMPJPE↓MPVPE↓
[75] [CVPR2021]242.50222.13
†[75] [CVPR2021]201.56189.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.616, + 0.266, + 0.807, + 0.277 + ], + "angle": 0, + "content": "(a) Scene-aware Motion Generation" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.28, + 0.906, + 0.332 + ], + "angle": 0, + "content": "
MethodFID↓R Precision ↑
MDM [76] [ICLR2023]0.5440.611
†MDM [76] [ICLR2023]0.4710.705
" + }, + { + "type": "table_caption", + "bbox": [ + 0.59, + 0.334, + 0.833, + 0.346 + ], + "angle": 0, + "content": "(b) Language-conditioned Motion Generation" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.352, + 0.905, + 0.38 + ], + "angle": 0, + "content": "Table 4. Quantitative results (§6.3) on two downstream tasks. † indicates using BEHAVIORHUB for additional training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.906, + 0.457 + ], + "angle": 0, + "content": "velopment of more powerful LLM cores. Moreover, this shows that BEHAVIORHUB, in conjunction with the human behavior planning task, serves as a robust testbed for evaluating the planning capability of LLMs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.474, + 0.819, + 0.489 + ], + "angle": 0, + "content": "6.3. BEHAVIORHUB for Downsteam Task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.696 + ], + "angle": 0, + "content": "We further probe the effectiveness of BEHAVIORHUB by incorporating it as additional training data for two downstream tasks: scene-aware [75] and language-conditioned motion generation [76], which focus on generation without planning. We follow the official implementation of both methods and begin by pre-training the two models on BEHAVIORHUB. Subsequently, we fine-tune then evaluate these models on PROX [52] and HumanML3D [77], respectively. The evaluation results are reported in Table 4. As observed, BEHAVIORHUB significantly enhances model performance across all evaluation metrics for both tasks, highlighting the potential of our created dataset in facilitating broader applications within the realm of motion generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.722, + 0.634, + 0.737 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We present ACTOR, an LLM-powered agent towards realistic simulation of human behavior in 3D scenes. ACTOR integrates an LLM controller to perform complex behavior through planning on goal decomposition guided by hierarchical activity prior. The value-driven mechanism further deepens its understanding of environment. We demonstrate its potential in the simulated 3D indoor environment constructed using our newly created large-scale, scene-aware, behavior-rich dataset, BEHAVIORHUB. Evaluations suggest effective behavior planning and simulation of ACTOR." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10760" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.244 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Science and Technology Major Project (No. 2023ZD0121300), National Natural Science Foundation of China (No. 62372405), Fundamental Research Funds for the Central Universities (226-2025-00057), National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University (No. HMHAI-202403), China Postdoctoral Science Foundation (No. 2025T180421), and the Postdoctoral Fellowship Program of CPSF (No. GZC20251066)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.256, + 0.188, + 0.272 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.281, + 0.482, + 0.308 + ], + "angle": 0, + "content": "[1] Denis Diderot. Diderot's early philosophical works. Number 4. Open Court, 1911. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.31, + 0.483, + 0.338 + ], + "angle": 0, + "content": "[2] Mark O. Riedl. Interactive narrative: A novel application of artificial intelligence for computer games. In AAAI, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.339, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[3] Mark O Riedl. Human-centered artificial intelligence and machine learning. Human behavior and emerging technologies, 1(1), 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.381, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[4] Chen Liang, Zhicheng Shi, Wenguan Wang, and Yi Yang. Dual reciprocal learning of language-based human motion understanding and generation. In ICCV, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.423, + 0.482, + 0.465 + ], + "angle": 0, + "content": "[5] Elias Matsas and George-Christopher Vosniakos. Design of a virtual reality training system for human-robot collaboration in manufacturing tasks. IJIDEM, 11, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.482, + 0.522 + ], + "angle": 0, + "content": "[6] Joon Sung Park, Lindsay Popowski, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Social simulacra: Creating populated prototypes for social computing systems. In UIST, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.482, + 0.59 + ], + "angle": 0, + "content": "[7] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[8] Xiaohan Wang, Wenguan Wang, Jiayi Shao, and Yi Yang. Lana: A language-capable navigator for instruction following and generation. In CVPR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.634, + 0.482, + 0.675 + ], + "angle": 0, + "content": "[9] Hanqing Wang, Wei Liang, Luc V Gool, and Wenguan Wang. Towards versatile embodied navigation. NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[10] Zhichao Zhai, Guikun Chen, Wenguan Wang, Dong Zheng, and Jun Xiao. Taga: Self-supervised learning for template-free animatable gaussian articulated model. In CVPR, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[11] Xavier Puig, Kevin Kyunghwan Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. Virtualhome: Simulating household activities via programs. In CVPR, 2018. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[12] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.281, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[14] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[15] Hongwei Yi, Chun-Hao P Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J Black. Human-aware object placement for visual environment reconstruction. In CVPR, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[16] Joseph Bates. The role of emotion in believable agents. Communications of the ACM, 37(7), 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.904, + 0.245 + ], + "angle": 0, + "content": "[17] Robert Zubek. Towards implementation of social interaction. In AAAI, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[18] Ruth Aylett. Narrative in virtual environments—towards emergent narrative. In AAAI, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[19] Michael Brenner. Creating dynamic story plots with continual multiagent planning. In AAAI, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[20] Ho Chit Siu, Jaime Pena, Edenna Chen, Yutai Zhou, Victor Lopez, Kyle Palko, Kimberlee Chang, and Ross Allen. Evaluation of human-ai teams for learned and rule-based agents in hanabi. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[21] Michele Colledanchise and Petter Ögren. Behavior trees in robotics and AI: An introduction. CRC Press, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.401, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[22] Ilshat Umarov, Mikhail Mozgovoy, and Patrick C. Rogers. Believable and effective ai agents in virtual worlds: Current state and future perspectives. International Journal of Gaming and Computer-Mediated Simulations, 4(2), 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[23] Christopher Berner, Greg Brockman, Brooke Chan, Vicki Cheung, Przemyslaw Debiak, Christy Dennison, David Farhi, Quirin Fischer, Shariq Hashme, Chris Hesse, Rafal Jozefowicz, Scott Gray, Catherine Olsson, Jakub Pachocki, Michael Petrov, Henrique P. d.O. Pinto, Jonathan Raiman, Tim Salimans, Jeremy Schlatter, Jonas Schneider, Szymon Sidor, Ilya Sutskever, Jie Tang, Filip Wolski, and Susan Zhang. Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[24] Oriol Vinyals, Igor Babuschkin, Wojciech M. Czarnecki, and et al. Grandmaster level in starcraft ii using multi-agent reinforcement learning. Nature, 575, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[25] Matthew Hausknecht, Prithviraj Ammanabrolu, Marc-Alexandre Cote, and Xinyu Yuan. Interactive fiction games: A colossal adventure. In AAAI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[26] Shohei Miyashita, Xinyu Lian, Xiao Zeng, Takashi Matsubara, and Kuniaki Uehara. Developing game ai agent behaving like human by mixing reinforcement learning and supervised learning. In SNPD, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.72, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[27] Chen Tessler, Shahar Givony, Tom Zahavy, Daniel Mankowitz, and Shie Mannor. A deep hierarchical approach to lifelong learning in apache. In AAAI, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[28] John Laird. It knows what you're going to do: Adding anticipation to a quakebot. In Proceedings of the 2001 Workshop on Intelligent Cinematography and Editing, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[29] Dong kyu Choi, Tolga Konik, Negin Nejati, Chunki Park, and Pat Langley. A believable agent for first-person shooter games. In AAAI, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Pat Langley, Dongkyu Choi, and Seth Rogers. Interleaving learning, problem solving, and execution in the icarus architecture. Technical report, Stanford University, Center for the Study of Language and Information, 2005. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10761" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[31] Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In UIST, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[32] Yujie Lu, Weixi Feng, Wanrong Zhu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Neurosymbolic procedural planning with commonsense prompting. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[33] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In ICML, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[34] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. In NeurIPS, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[35] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In NeurIPS, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[36] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In ICML, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.476 + ], + "angle": 0, + "content": "[37] Shima Imani, Liang Du, and Harsh Shrivastava. Math-prompter: Mathematical reasoning using large language models. In ICLR Workshops, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[38] Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen McAleer, Albert Q. Jiang, Jia Deng, Stella Biderman, and Sean Welleck. Llemma: An open language model for mathematics. arXiv preprint arXiv:2310.06786, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[39] Xiang Lorraine Li, Adhiguna Kuncoro, Jordan Hoffmann, Cyprien de Masson d'Autume, Phil Blunsom, and Aida Nematzadeh. A systematic investigation of commonsense knowledge in large language models. In EMNLP, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[40] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In ICRA, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[41] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[42] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[43] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In CoRL, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[44] Wenlong Huang, Fei Xia, Dhruv Shah, Danny Driess, Andy Zeng, Yao Lu, Pete Florence, Igor Mordatch, Sergey Levine," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Karol Hausman, et al. Grounded decoding: Guiding text generation with grounded models for robot control. arXiv preprint arXiv:2303.00855, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[45] Belinda Z Li, Maxwell Nye, and Jacob Andreas. Implicit representations of meaning in neural language models. In ACL, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[46] Adam Roberts, Colin Raffel, and Noam Shazeer. How much knowledge can you pack into the parameters of a language model? In EMNLP, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[47] Yunqiu Xu, Linchao Zhu, and Yi Yang. Mc-bench: A benchmark for multi-context visual grounding in the era of mllms. In ICCV, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[48] Vadim Borisov, Kathrin Sessler, Tobias Leemann, Martin Pawelczyk, and Gjergji Kasneci. Language models are realistic tabular data generators. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[49] Yew Ken Chia, Lidong Bing, Soujanya Poria, and Luo Si. Relationprompt: Leveraging prompts to generate synthetic data for zero-shot relation triplet extraction. In ACL, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[50] Timo Schick and Hinrich Schütze. Generating datasets with pretrained language models. In EMNLP, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[51] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. In ACL, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[52] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[53] Hongming Zhang, Xin Liu, Haojie Pan, Yangqiu Song, and Cane Wing-Ki Leung. Aser: A large-scale eventuality knowledge graph. In WWW, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[54] Peter West, Chandra Bhagavatula, Jack Hessel, Jena Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. In NAACL, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.905, + 0.623 + ], + "angle": 0, + "content": "[55] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.905, + 0.679 + ], + "angle": 0, + "content": "[56] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.68, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[57] Yudi Dai, Yitai Lin, Chenglu Wen, Siqi Shen, Lan Xu, Jingyi Yu, Yuexin Ma, and Cheng Wang. Hsc4d: Human-centered 4d scene capture in large-scale indoor-outdoor space using wearable imus and lidar. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[58] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.791, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[59] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Xuan Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In NeurIPS, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10762" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[60] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[61] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In ECCV, 2020. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.482, + 0.221 + ], + "angle": 0, + "content": "[62] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.224, + 0.482, + 0.264 + ], + "angle": 0, + "content": "[63] Thomas A Feo and Mauricio GC Resende. Greedy randomized adaptive search procedures. Journal of global optimization, 6, 1995. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.268, + 0.482, + 0.295 + ], + "angle": 0, + "content": "[64] Markus Freitag and Yaser Al-Onaizan. Beam search strategies for neural machine translation. In ACL, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.298, + 0.482, + 0.325 + ], + "angle": 0, + "content": "[65] Rémi Coulom. Efficient selectivity and backup operators in monte-carlo tree search. In CG, 2006. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.328, + 0.482, + 0.368 + ], + "angle": 0, + "content": "[66] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[67] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In CVPR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.429, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[68] Santhosh K. Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westry, Angel Xuan Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. In NeurIPS Datasets and Benchmarks Track, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.528, + 0.482, + 0.596 + ], + "angle": 0, + "content": "[69] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.6, + 0.482, + 0.653 + ], + "angle": 0, + "content": "[70] Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan. Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.697 + ], + "angle": 0, + "content": "[71] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.701, + 0.482, + 0.741 + ], + "angle": 0, + "content": "[72] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[73] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[74] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An opensource chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. https://vicuna.lmsys.org, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[75] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In CVPR, 2021. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[76] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[77] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10763" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/099d59e1-3cf8-46ca-a7c2-5f8c85602d06_origin.pdf b/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/099d59e1-3cf8-46ca-a7c2-5f8c85602d06_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3399f54a46712a57cf26dcfa2899d1fb9e1b680c --- /dev/null +++ b/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/099d59e1-3cf8-46ca-a7c2-5f8c85602d06_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f1ee3d68bc6ff70d8b50058b97e3028d8936cf928bd332f8b4dfbd51841b381c +size 4665810 diff --git a/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/full.md b/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..773a4af6dc0604403f43cd9b7198ca7e62c815f5 --- /dev/null +++ b/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/full.md @@ -0,0 +1,351 @@ +# Towards Human-like Virtual Beings: Simulating Human Behavior in 3D Scenes + +Chen Liang, Wenguan Wang*, Yi Yang + +State Key Lab of Brain-Machine Intelligence, Zhejiang University + +https://github.com/leonnnop/BehaviorHub + +![](images/07bd40985b831dd2c2e9e9a990d588f6834ad77a0657139b68a6c44edaf4f3bb.jpg) +(c) + +![](images/36b94a56ebda57f74e33e1e20c8365cb7b1e24cf5bf070eec62586faa2b9b3dd.jpg) +Figure 1. Previous research has primarily focused on: (a) linguistic-based short-horizon task planning, and (b) low-level human-scene interaction. (c) This study investigates the simulation of high-level, long-horizon, abstract goal-driven human behaviors in 3D scenes. + +![](images/fd3241e930f71c9da4e9b4e61e4111d9a775817de5fb493da588fc2527c0519c.jpg) +Human Behavior Simulation in 3D Scenes (Ours) + +![](images/b32ffb288e004c4551a45a545ccbb478b03c517913c899d4c5d592053290cef4.jpg) +High-level, Long-horizon, Abstract Goal: Preparing for work from home + +# Abstract + +Building autonomous agents that can replicate human behavior in the realistic 3D world is a key step toward artificial general intelligence. This requires agents to be holistic goal achievers and to naturally adapt to environmental dynamics. In this work, we introduce ACTOR, an agent capable of performing high-level, long-horizon, abstract goals in 3D households, guided by its internal value similar to those of humans. ACTOR operates in a perceive-plan-act cycle, extending the ungrounded, scene-agnostic LLM controller with deliberate goal decomposition and decision-making through actively searching the behavior space, generating activity choices based on a hierarchical prior, and evaluating these choices using customizable value functions to determine the subsequent steps. Furthermore, we introduce BEHAVIORHUB, a large-scale human behavior simulation dataset in scene-aware, complicated tasks. Considering the unaffordable acquisition of human-authored 3D human behavior data, we construct BEHAVIORHUB by exploring the commonsense knowledge of LLMs learned from large + +corpora, and automatically aligning motion resources with 3D scene for knowledgeable generation. Extensive experiments on our established benchmark demonstrate that the proposed architecture leads to effective behavior planning and simulation. BEHAVIORHUB also proves beneficial for downstream task development. + +# 1. Introduction + +Building autonomous agents (e.g., virtual beings, or humanoid robots) that can replicate human behavior in the realistic 3D world, has been a long-standing pursuit since the inception of AI [1]. The study could empower nonplayer game character [2], underpin human-robot interaction [3, 4] and cooperation [5], populate virtual reality communities [6], and accelerate Embodied AI [7-10]. Much progress has been made in vision and language models that imitate human motions (cf. Fig. 1 (b)) or propose linguistic plans (cf. Fig. 1 (a)). However, an effective 3D humanoid agent must go beyond by conquering three major barriers: (i) Holistic goal achievement from perception to action. As depicted in Fig. 1 (c), to accomplish a high-level goal (e.g., + +'prepare for work'), the agent must process the perceived information (e.g., 'lying on bed'), decompose the goal into a series of activities (e.g., 'get up', 'use toilet', 'eat breakfast', etc.), and devise appropriate action plans for each; (ii) Environmental dynamics. Agents should be able to actively adjust their plans based on the environment, e.g., determining if the bathroom is occupied when planning to use toilet; (iii) Vast and multifaceted human behavior space, where numerous viable paths exist to achieve even a single goal. For example, agents preparing for work can choose to eat breakfast before using toilet or vice versa. Also, when the bathroom is occupied, an intelligent agent can decide to whether eat breakfast first, or continue waiting, depending on its state and beliefs, e.g., the desire to 'complete goal as soon as possible' or 'save energy costs'. A competent agent must process such value priorities to guide its selection and evaluation of actions and policies. Besides, the absence of a comprehensive testbed further poses great challenge for agent development and evaluation, constituting barrier (iv). + +In this work, we present ACTOR - a large language model (LLM) powered agent towards diligent simulation of human behavior in 3D realistic scenes (§4). ACTOR follows a perceive-plan-act cycle, addressing challenge $(i)$ as envisioned. Using LLM as a central controller, it strategizes plans by searching the human behavior space. It actively maintains a tree of behaviors, where each node represents an intermediate step toward holistic goal achievement. The construction of this tree is guided by a hierarchical prior, i.e., executable low-level actions are grouped into high-level semantic units, called activities, iteratively forming a hierarchical structure (cf. Fig. 1 (c)). The search progress is assessed using a set of customizable value functions that determine the likelihood of different intermediate candidates. In addition to being rational in common sense, ACTOR couples real-valued evaluations (e.g., best efficiency) with personalized priors expressed through language commands (e.g., description of a neat person). These outputs are converted into unified probabilities, allowing for the incorporation of the agent's characteristics and beliefs as value functions, thus addressing challenge (iii). The planning process is dynamic and grounded in specific environmental values, empowering ACTOR to readjust its plans when faced with environmental changes or new language commands, effectively tackling challenge (ii). This formulation facilitates the use of powerful search algorithms, e.g., greedy search, beam search, and Monte Carlo tree search (MCTS), with MCTS proving to be the most effective in our experiments. + +Furthermore, we establish a comprehensive environment for development and evaluation of agents like ACTOR, based on our newly proposed large-scale, scene-aware, behavior-rich dataset, dubbed BEHAVIORHUB (§5). One critical issue in constructing the human-authored benchmark is the high cost associated with acquiring and scaling + +high-quality, human-generated daily behavior data. Moreover, annotating large-scale behavior data further requires creativity in designing novel tasks and expertise in creating complete plans from scratch, which is also a challenging task for humans [11]. Given this context, we propose to automatically synthesize 3D human behavior data by enhancing existing resources. Initially, we distill the tree-structured linguistic plans of human daily behaviors from LLM using in-context learning [12]. Contactable objects and plausible interactions from the scanned 3D scenes (e.g., ScanNet [13], etc.) and captured motion sequences (e.g., AMASS [14], etc.) are attributed into the in-context prompt to ensure viable plans grounding in certain environment. Then, we align the task-motion-scene data triplets by applying collision and contact constraints [15] for valid translation and rotation parameters. To promote diversity, we sample multiple plausible motion sequences for each high-level goal. In total, BEHAVIORHUB contains more than 1k daily goals over 10k high-quality behavior samples of 15.7 steps on average covering 1.5k scenes, that establishes a comprehensive testbed, that addresses the barrier $(iv)$ . + +We conduct extensive evaluations of ACTOR in §6. First, we find that ACTOR produces admissible plans and generates plausible motion sequences to simulate abstract, temporally-extended human behaviors. It outperforms the strong baselines by nearly doubling the overall success rate. This result is further confirmed by human evaluation. Then, we conduct several ablative experiments that limit ACTOR's access to each core design for thorough assessment. Finally, experiments on scene-aware and language-conditioned human motion generation demonstrate how BEHAVIORHUB can benefit the development of downstream task models. + +# 2. Related Work + +Human Behavior Simulation. Simulating human behavior in realistic, open world environment like the one we inhabit is a long-standing topic in artificial intelligence [16]. Historically, the topic has primarily been studied in the game worlds, focusing on enhancing player experiences through intelligent non-player game characters (NPCs) [17-19]. Early approaches rely on rule-based approaches like finite-state machines [20] and behavior trees [21]. They provide a brute force way of manually crafting agent's behaviors, but cannot perform new procedures that were not hard-coded in their script [22], limiting the generalizability. Another strand [23, 24] involves using reinforcement learning, where agents learn its own policy through optimizing the learning algorithm on readily definable rewards over fixed task space, which is often limited to non-open, adversarial games, or blocks worlds only [25-27]. The third line of research, represented by some pioneer works [28-30] in computational cognition, aims to build machines that operate directly in perceive-plan-act cycles, encompassing + +the nature of autonomous agents as originally envisioned. This formulation holds potential generalizability to most, if not all, open-world contexts. However, these studies are typically restricted to simplified environments, such as first-person shooter games [29], or 2D gridworlds (e.g., Generative Agents [31]), and focus on a reduced range of behaviors. Our work falls in vein of third category, while pushing frontier towards simulation in realistic 3D environments. + +Recently, beyond the game world, new trending on social embodied intelligence, such as assistive robots, draws attention to simulation of human collaborators' behavior. One notable effort is VirtualHome [11], which studies high-level human activities as plain sequences of atomic actions. However, its formulation is inherently environment-agnostic, which has been treated as a purely linguistic procedural planning problem in subsequent studies [32, 33]. Furthermore, the discrete action space built upon manually crafted procedural knowledge also makes it an insufficient testbed. In contrast, with a special focus on environment-aware simulation of contiguous daily behaviors, our work paves one solid step accelerating the development of social embodied intelligence. + +LLM as Planner. Remarkable progress in LLMs has been observed, demonstrating their emerging capacity to break down complex tasks into more manageable sub-tasks and devise appropriate plans for each [34-36]. LLMs have been successfully applied to solving mathematical problems [37, 38], reasoning on commonsense [39], planning robotics tasks [40, 41], and very recently, manipulating external APIs on a web scale, expanding their capabilities beyond text generation. Based on the observation, we posit that LLMs can serve as a crucial component in extending the perception-decision-action space to construct human-like agents in realistic 3D environments. However, a crucial challenge remains: LLMs lack experience and interaction with their environment [12, 42], preventing them from ordering actionable and rational plans. This paper addresses this issue by incorporating customizable value functions into LLMs to evaluate and prioritize plans. The idea bears some resemblance to recent robotics research [43, 44]. However, those methods necessitate retraining for every new set of primitive robotic skills, making them impractical for the complex and undefined human action space, which is difficult to define in advance. In contrast, we support plan evaluation using real-numbered functions and language-based rules without resource-intensive retraining. + +LLM as Data Generator. Being trained on the large corpus of human-produced language, LLMs are believed to contain a wealth of information about the world [45-47]. Given a handful of task-specific prompts, LLMs can generalize and generate more linguistic data in the same format, with the application of generating tabular data [48], relation triplets [49], sentence pairs [50], instruction data [51], + +etc. The idea seems naturally to be borrowed for human behavior data acquisition, where the requested human motion and daily activity procedure were commonly crowdsourced with high expense and complexity [11], limiting the scale and coverage of related datasets [52]. However, the process is non-trivial. The generated data is often blamed for low quality and diversity issues [53, 54]. The 3D environment-aware nature of the task also poses unique challenges. To respond, we explore attributed prompts specifically conditioned on the environment that not only mitigates the problems of low informativeness and redundancy, but also offers an effective workflow that can further empower other related domains, such as human-scene interaction [55]. + +# 3. Environment + +Simulating open-ended goals that resemble naturalistic human behaviors necessitates an environment capable of facilitating diverse agent affordances and interactions. Before delving into the detailed agent architecture, we first describe such environment we tailored for agents to instantiate in. + +Environmental Setup. The environment features common affordances in a household, including: + +- Scene of 3D textured meshes, that spans a house of multiple functional areas (e.g., kitchen, etc.). +- Objects embedded in the scene (e.g., stove in kitchen, bed in bedroom), with each constructed of 3D textured geometry and corresponding object state (e.g., fridge: opened). +- Humanoid agent(s) defined by SMPL-X [56], an expressive 3D human model of shape and pose of both body and hand. Agents reside in and interact with the scene and can influence the state of objects by their actions (e.g., close the fridge | fridge: opened $\rightarrow$ closed). + +As a start point, we demonstrate the environment of indoor household, keep the general formulation open for other environments, such as outdoor streets [57]. + +Environmental Simulator. Interactions within the environment are driven by the simulator of two components: (i) an engine that manages and evolves the states of objects in the environment; and (ii) a renderer that supports generation of multiple perceptual observations (e.g., RGB, depth, 3D surfaces) for agents. At each time step, the simulator dynamically updates environment and collects egocentric, surround, or third-person-view information based on needs. We build our simulator upon Habitat-Sim [58, 59] to ensure efficient and parallelizable simulation. + +# 4. ACTOR Agent + +We aim to build humanoid agents that naturally simulate 3D human behaviors to complete daily goals, which can be lengthy, abstract, or ambiguous. Agents receive high-level goal described in language and are tasked with generating plausible motion sequences that align with the given scene. + +![](images/383f20063d400ecf9bd55b6c27206e86d702832445b0a037574d7e16aac007c7.jpg) +Figure 2. Left: Overall agent architecture (§4.1); Right: One-step in perceive-plan-act loop with value-driven behavior planning (§4.2). + +# 4.1. Agent Architecture + +Fig. 2 shows the workflow of ACTOR. The agent operates in a perceive-plan-act loop, using language as the generic interface connecting the three phases. At each step, it selectively perceives the world based on the target, transforming information into an environment description using heuristics. The observed information assists in decision-making through the LLM core, ultimately resulting in actions in the form of 3D motion sequences based on linguistic plans within the environment. We now delve into the details. + +Perception. Our definition of perception extends beyond gathering information, such as 2D images or 3D point clouds, through the simulator (§3). It encompasses attaining a deep understanding of the environment, including object properties, spatial relationships, and scene layouts, etc. In our preliminary implementation, information is perceived using readily available models and converted into linguistic descriptions using heuristic functions (supp. §A.2). + +Plan. At the core of ACTOR, a planning module receives the current environment and past behavior trial as input, generating action description as output. For the executable actions, it provides descriptions of motion sequences. In the case of high-level activity, it bypasses the action stage and proceeds to break down the target further. We will provide a detailed explanation of this process in §4.2. + +Action. For action, we specifically consider whole-body human actions in 3D scenes to closely resemble human behavior using the off-the-shelf models. It generates whole-body motion based on text and trajectory [60]. For object-interactive actions, we further refine hand grasping using an isolated grasp estimation model [61]. Additionally, for moving actions like walking, trajectory paths are pre-estimated [62]. Detailed trajectory estimation process is provided in supp. §A.3. + +# 4.2. Value-driven Behavior Planning + +An LLM, denoted as $p_{\theta}$ with parameters $\theta$ , functions as a controller to iteratively decompose the long-term goal $l$ described in language texts into shorter steps of behaviors based on the partial environmental dynamics. To prevent confusion with the global environment $E$ , we denote the + +partial environmental observation at step $i$ as $E_{i}^{\prime}$ and its corresponding textual description as $e_{i}$ . For brevity, we omit step indicator $i$ and define $E^{\prime}$ to include both the environment state and agent's state. We denote textual descriptions of behaviors as $z$ . The problem is formulated as a search over a tree, where each node is a state $s = \{z_{1\dots i}\}$ , representing a partial trial with the input and the sequence of behaviors taken thus far. The instantiation consists of three components: (i) Node expansion, which involves generating $k$ candidate behaviors for the next step search; (ii) Value functions to evaluate each node; (iii) Search algorithm that accounts for branch selection. + +Node Expansion. The node is expanded through sampling from LLM with a window size $w$ : $z^{(j)} \sim p_{\theta}(z_{i+1}|l,e,s,h)$ , where $j \in \{1,\dots,w\}$ . We omit the basis prompting parts for brevity, which is also the case in the subsequent context. Here, we introduce a hierarchical heuristic, denoted as $h$ , which is implemented through prompt instructions and demonstrations of specific actions and activity cases. This heuristic ensures that during each expansion, all candidates at the same level are restricted to executable actions or high-level semantic units of activities, that provides a more nuanced representation of behaviors for effectively modeling interchangeable activities. + +Value Function. The value function assesses the state by considering the degree to which a specific behavior contributes to the achievement of the target, conditioned on the agents' beliefs reflecting their value. The search algorithm uses the output to determine which nodes to explore further and in what order. The likelihood is given by $p(z_{i+1} | l, E', s) \propto p_{\theta}(z_{i+1} | l, e, s) \cdot p_v(z_{i+1} | l, E', s)$ . Here, the LLM provides us with $p_{\theta}(z_{i+1} | l, e, s)$ , which represents the likelihood, based on commonsense, that a textual behavior is a valid next step. However, the LLM struggles to generalize or make inferences in the real environment since it is not grounded. On the other hand, $p_v$ provides the likelihood of the behavior being plausible in the current state of both the environment and agent, according to the defined values. We consider a set of values $\{p_{v_n}\}_n$ categorized into two types, and $p_v(z_{i+1} | l, E', s) = \prod_n p_{v_n}(z_{i+1} | l, E', s)$ : + +Real-valued function assigns real values as outputs, such + +![](images/76b5d3c4930ff393da2ced98b0fc7f3815d0101d5c31190c561e7a652961624d.jpg) +Figure 3. (a) An illustrative example of our BEHAVIORHUB dataset; (b) Semi-automatic 3D human behavior data generation pipeline ( $\S 5$ ). + +![](images/8ee00300b55710f3d54b0a1c8135779ab602fd33bd84da53d3f0c46e0a94e0da.jpg) + +as the shortest path value, for which the function estimates the distance for each candidate action. The outputs of real-valued functions are normalized into probabilities. + +- Language-based command is implemented by prompting the LM with a value prompt that conveys the meaning of 'How likely is it for someone who is a neat person to take the following action, considering ...'. The model then reasons about each action to generate the corresponding probability reflecting the value. + +Values are also conditioned on the state of the environment and agent. This approach allows for reacting to environmental dynamics in addition to the active planning process. Search Algorithms. Owing to the active planning process we have formulated, it is feasible to use different search algorithms based on the tree structure. We explore greedy search [63], beam search [64] and MCTS [65], and evaluate their performance in experiments (cf. §6.2) where we find MCTS performs best. We therefore use MCTS by default. + +# 4.3. Implementation Detail + +We employ GPT-4 and GPT-3.5-turbo API provided by OpenAI as base LLMs. We prompt them with four in-context examples as demonstrations, with decoding temperature as 0 for more deterministic generation. We use official releases of conditional motion generation models and fine-tune them on BEHAVIORHUB using the default parameters for each to promote finer generation quality. By default, we set the window size to $w = 5$ and normalize the shortest path value based on the maximum rollout depth of the search algorithm, which is set to 3. In practice, we find one-time sampling is sufficient for effective tree search, eliminating the need for multiple samplings for branch aggregation. Our algorithm is implemented in PyTorch using Tesla A40 GPUs. + +# 5. BEHAVIORHUB Benchmark + +To lay a solid foundation for future research, we build a large repository of common behaviors performed in daily household scenarios. Each sample contains three components: $(i)$ a high-level goal (i.e., root node); $(ii)$ a tree + +structured linguistic plan covering necessary intermediate-level steps (i.e., intermediate nodes) and low-level steps (i.e., leaf nodes) required to accomplish the goal; and (iii) scene-conditioned human motions corresponding to each executable step at either intermediate or low level. We illustrate one example in Fig. 3 (a). Intermediate step set can be empty. Both goal and steps in plans are human activities described with either concrete or abstract language (e.g., 'go to sleep on bed' or 'feel tired'). To generate diverse and high-quality data, we employ a two-step pipeline that utilizes an LLM pretrained on extensive corpora: First, automatic generation of linguistic daily plans ( $\S 5.1$ ); Second, alignment of the plans with 3D motions and scenes ( $\S 5.2$ ). The entire data generation process is shown in Fig. 3 (b). + +# 5.1. Linguistic Goal-Plan Generation + +The pipeline consists of three steps: $(i)$ generate potentially incomplete goal-plan trees; $(ii)$ complete and refine each tree; and $(iii)$ filter out low-quality data. We provide detailed prompts in supp. §A.1. + +Attributed Goal-Plan Tree Generation. In the initial step, we generate new goal-plan trees using a bootstrapping approach based on a small set of seed human-written samples. To ensure broader coverage and facilitate later alignment with specific scenes, we attribute the starting room, candidate objects and actions that interact with the activity in the plans into prompt demonstrations. We start the sample pool with 292 activities from ActivityPrograms [11], and consider 21 room types (e.g., bathroom, gym), $\sim 10^3$ object types (e.g., TV, table), and $\sim 10^2$ action types (e.g., open, eat). Refer to the supp. §A.5 for the full list. For each step, we sample eight goal-plans as in-context examples from the pool while restricting the outputs to one room and ten types of objects. Out of the eight plans, six are from the human-written plans, and two are generated by the model in previous steps to enhance diversity. + +Generated trees are then labeled with sequential order for intermediate nodes in the trees, which encompass multiple sub-step leaf nodes of individual actions. This sequential or + +der is represented as interchangeable groups, where nodes within the same set are interchangeable with each other. We prompt the LM in a few-shot way to determine this. + +Goal-Plan Tree Refinement. We further improve the constructed trees by addressing two key aspects: (i) we complete any missing internal plan steps, which can often be revised based on commonsense, e.g., opened the fridge without closing it. (ii) we enhance the root node descriptions to be more abstract, e.g., transforming 'use toilet' to 'feel the call of nature'. We accomplish both aspects by querying the LLM for suggestions. + +Filtering. To promote diversity, a new tree is added to the pool only if its BERTScore similarity [66] with any existing goal-plan tree is below 0.5. Additionally, we utilize LLM to assess the generated trees by asking the question, 'Is this a valid plan?' Any plans flagged as 'invalid' are filtered out to ensure high-quality data. + +# 5.2. Goal-Motion-Scene Alignment + +Once the goal-plan trees are formed, we proceed to ground them in the 3D environment. We propose to synthesize 3D behaviors by leveraging existing resources of human motions (i.e., AMASS [14], BABEL [67], GRAB [61]), and indoor scenes (i.e., ScanNet [13], HM3D [68], Replica [69]). For each tree, we first sample actions and objects from resources using corresponding labels to fulfill the executable activities in forms of combinations of actions on objects. Then, given sampled motion, we aim to generate plausible and contiguous interactions to sampled object in scene. + +Motion Alignment. We first put motion clips into the scene anchored by the contactable objects. We optimize the translation and rotation parameters matrices by minimizing the collision and contact losses from MOVER [15]. A unified SDF volume is calculated, and all contact vertices for all frames are accumulated in 3D space. The motion is aligned through a transformation of the two matrices. This joint optimization improves human-object contact and resolves 3D interpenetrations between humans and the scene. + +Sequence Blending. The aligned motion sequence, which may be sparse and not spatially connected, is blended using a Transformer-based motion completion method [70]. This results in a contiguous motion sequence that aligns with both the scene and goal plan. In practice, we find a single network is capable of delivering satisfactory results that align with prior research [71]. + +Verification. Finally, to ensure the data quality, each example is examined by three verifiers who vote on whether the plan (motion) is complete (valid). If an example receives majority approval, it is accepted; otherwise, it is dropped. + +# 5.3. Dataset Statistics + +Our dataset consists of 10k human behavior samples in 1.5k 3D scenes, covers 2k unique activity over 0.1k actions and + +1k objects, which is an order of magnitude larger than the human-authoring ActivityPrograms [13]. On average, each high-level goal has 15.7 steps, resulting in a total of 10.1 corresponding motion sequences of 83.3 frames, that span over 8.6M motion frames. Each activity corresponds to 4.1 different motion sequences in 1.7 rooms. Owing to the automatic data generation pipeline, our dataset achieves even higher diversity and coverage with low human cost. + +# 5.4. Evaluation Metric + +Our evaluation of the plausible behavior simulation encompasses two key aspects: $(i)$ the simulation should generate linguistically reasonable behavior plans (i.e., behavior planning); and $(ii)$ align them with natural motion sequences within the 3D environment (i.e., behavior simulation). + +- For behavior planning, we use Sentence-BLEU [72] and BERTScore [66] to measure the semantic similarity between the ground-truth plans and predictions. We report maximum scores attained across ground-truth plan variants to ensure the evaluation be order-invariant with respect to interchangeable sub-steps. +- For behavior simulation, we consider three sets of metrics: (i) Success Rate: The step success rate (SSR) records the percentage of steps where the agent successfully completes the step objective, defined by a contact distance threshold. For example, we consider lying down to be successful if both the hip and head of the humanoid are within $30\mathrm{cm}$ of the target location [73]. The goal success rate (GSR) is measured to determine whether all steps in the entire plan are successfully executed; (ii) Goal Success Rate Weighted by Path Length (GSRPL): It judges how efficient was the agent at finishing the goal, defined as $\mathrm{GSR} \cdot \frac{g}{\max(g,l)}$ . Here $g$ is ground-truth path length and $l$ is the agent's path length; (iii) Motion Quality: We evaluate overall quality of generated motions using Frechet Inception Distance (FID) and recognition accuracy measured with the final layer of a pretrained standard RNN action recognition classifier as motion feature extractor. + +# 6. Experiment + +# 6.1. Performance on BEHAVIORHUB Benchmark + +Dataset Split and Dynamic Subset. We randomly sample 200 held-out goals as demonstration set from which we select example(s) for prompting language models, and also as training data for fine-tuning conditional generation models. The remaining ones are used for evaluation. + +To thoroughly examine the capacity of benchmarked models in managing environmental dynamics, we manually create a subset of 300 samples from the original BEHAVIORHUB, called Dynamic subset. In each sample, we carefully configure the agent's environment state-aware triggers to guarantee a distinct goal-plan solution for a specific goal, + +
MethodBehavior PlanningBehavior Simulation
S-BLEUBERT-SSSRGSRGSRPLFID ↓Accuracy ↑
Main SetLLMaP [33] [ICML2022]0.0890.825-----
HuggingGPT [34] [NeurIPS23]0.1320.8560.5330.3170.1615.3860.620
ACTOR (Ours)0.1700.8790.6010.4720.3512.0870.773
Dynamic SubsetHuman0.2030.959-----
LLMaP [33] [ICML2022]0.0690.821-----
HuggingGPT [34] [NeurIPS23]0.0990.8300.4070.1640.0739.1160.505
ACTOR (Ours)0.1350.8620.5150.3060.2123.1410.697
+ +![](images/105c00aef1fc19ca71363c6bdb91069601bf2a46e8b2b6e2355da2fb4e4cf83f.jpg) + +![](images/c95c90ce58bc48753305ca24b79a7ad14e1b96d17237c7e7f24a68ef96c34a8b.jpg) +Original plan: weekend cleaning $\rightarrow$ clean bedroom $\rightarrow$ Strategized plan: +(occupied signal removed) + +![](images/b64adb1e94242751db9dbea8501b12f0db81e7a48fff2aca043a39124028aafd.jpg) + +![](images/a97d39aab47f096159cb2557502d9223afb04e2d113cadf24165a3ef007138b3.jpg) + +![](images/1444e61b2e50bc1aca4e84d1dd643d0d62a48755108d682b1ba98c4c523ca83d.jpg) +Clean bedroom + +![](images/4c8ef65b30b3f52532f40ac9912db37642aff8c3b76758d8d83e2252600198e9.jpg) + +![](images/d22317b8a865bf29be4dbbca2b37b4b8a173016a874a30a5ea680052aef4e630.jpg) + +![](images/0d0261e1606befb36ed42293b996b46615ec2713130c7b513b6025371a6ec93e.jpg) +Goal: prepare for th +Figure 4. Qualitative results of ACTOR on the challenging BEHAVIORHUB Dynamic subset. Please refer to §6.1 for more details. + +![](images/b5e1088c145076ebe587cfff6df9130bc3167e7583f1123b3bc43d50068f001c.jpg) +Language command: he/she is a neat person + +![](images/b120b4b17ddf01cdc28ec6a2c4810c165864a21c621c3c92c6773d53ed1d619f.jpg) + +![](images/2092a6ccacd391c10a7b0f0876f4518fc5c932143cffd6b38b8ea1ae932f32f9.jpg) + +![](images/651f81261a41e6b2a7676d43e5cd15204931fe2d298fd71c05a97589aa3126c4.jpg) + +![](images/16e8a8fedeee2b016faa9777a90249094806def0e52f525502bde8d90ba102e4.jpg) + +![](images/ad8b8ffeab641dedd45303717db750bc5ca68526f4fb725ee7afad5fe60320b0.jpg) + +![](images/ede2439405ca5e53c44063a00858e96b2f467ba145f1938900533c9e1a49caed.jpg) + +e.g., we designate the bathroom as occupied only after certain pre-request steps have been fulfilled; or incorporate language commands that specify agent characteristics, further contributing to a unique goal-plan preference. + +Competitors. We benchmark two top-leading LLM-based models: LLMaP [33], a procedural planning model, and HuggingGPT [34], a general tool agent, to probe the human behavior simulation ability in existing techniques. LLMaP is designed to operate solely on textual inputs, and lacks the capability to perform behavior simulation. We assess its behavior planning ability only. + +Quantitative Evaluation. Table 1 summarizes the automatic evaluation results, from which we take three major observations: (i) ACTOR achieves the best results on both behavior planning and simulation and outperforms the strong baseline by nearly doubling the GSR, indicating the effectiveness of our value-driven behavior planning design. (ii) On the challenging Dynamic subset, both LLMaP and HuggingGPT experience a noticeable decline in terms of BERT-S, whereas our ACTOR performs much better. To illustrate, in the case shown in Fig. 4, both LLMaP and HuggingGPT directly ignore the occupied signal and give an irrational plan of heading directly to the bedroom, causing a failure. This observation confirms our claim that the ungrounded LLM falls short in terms of environment-aware behavior planning. (iii) Although our ACTOR shows more promising results, there still remains a significant gap compared to human performance. This highlights the need for + +Table 1. Quantitative results on Main and Dynamic set of BEHAVIORHUB. $\downarrow$ ’ indicates smaller values are better. See §6.1 for details. + +
MethodComplete.Rational.Quality
Human4.024.85-
LLMaP [ICML2022]2.232.48-
HuggingGPT [NeuIPS23]2.712.893.18
ACTOR (Ours)3.053.473.75
+ +Table 2. Human-evaluated results (§6.1) on Dynamic subset. + +developing more sophisticated behavior simulation models. + +Human Evaluation. For a comprehensive evaluation, we engage five participants as human evaluators. They rate the model performance based on three aspects: (i) Completeness: assess whether the motion steps can successfully complete the target goal, capturing semantic completeness; (ii) Rationality: evaluate whether the sequence includes necessary steps in the correct sequential order to accomplish the target goal, capturing sequential order correctness; and (iii) Quality: reflect the naturalness and smoothness of the motion sequence, capturing motion quality. The results of the human ratings, based on a 5-point Likert scale, are reported in Table 2. The human subjective judgments generally align with the trends reflected by Table 2, confirming the reliability of our constructed automatic evaluation framework. Also, the results reaffirm that ACTOR yields plausible behavior simulation. However, it is worth noting that human-written plans are consistently preferred over our results, underscoring the challenging nature of our newly proposed BEHAVIORHUB benchmark. + +Qualitative Analysis. Examples from ACTOR on BEHAV + +
MethodBERT-SGSRGSRPL
Baseline0.8110.1400.062
+ Active Search0.8370.2350.132
+ Hierarchical Prior0.8490.2610.154
+ Value Function0.8620.3060.212
+ +(a) Key Component Analysis + +
AlgorithmBERT-SGSRGSRPL
Greedy0.8400.2440.151
Beam0.8530.2870.186
MCTS0.8620.3060.212
+ +(b) Search Algorithm + +
LLMBERT-SGSRGSRPL
Vicuna-7b0.8080.0920.063
GPT-3.50.8330.1760.116
GPT-40.8620.3060.212
+ +IORHUB Dynamic are visualized in Fig. 4. In addition to simulating behavior to strategize and successfully achieve the desired goal, ACTOR is able to respond to environmental changes and adapt to language commands. For instance, to accomplish the goal of 'Sunday cleaning', while someone is still in bed, i.e., the bedroom is occupied, the agent prioritizes scheduling the kitchen first and then the bedroom, waiting for the person to wake up. In the second example, considering the truth that someone is a neat person, he/she engages in home cleaning before eating breakfast. + +# 6.2. Diagnostic Experiments + +Ablative studies are conducted on BEHAVIORHUB Dynamic for indepth analyzing each component in ACTOR, using BERT-S, GSR, GSRPL (cf. §5.4) as evaluation metrics + +Key Component Analysis. We first validate the importance of our proposed components by attaching them one at a time in Table 4b. The $1^{st}$ row reports the result of a bare baseline model, which produces a global plan based on the given linguistic goal. Next, in the $2^{nd}$ row, we transition from one-pass planning to active tree search, resulting in improved performance and supporting our claim that active planning can mitigate the impact of environmental changes. Moreover, the $3^{rd}$ row gives the score when the hierarchical behavior structure prior is employed when spanning searching branches. As seen, this leads to moderate improvement by constraining the search space with interchangeable semantic units, highlighting its necessity in handling vast and complex human behavior space. Finally, as shown in the $4^{th}$ row, incorporating value function significantly enhances the overall success rate and success rate over path length, aligning with the rational preference for the shortest path. + +Search Algorithm. Table 3b reveals the impact of search algorithms (§4.2), i.e., greedy search, beam search, and MCTS, with window sizes of 5 for the later two algorithms. The default strategy, MCTS, shows optimal results, which aligns with the widely accepted understanding that MCTS is more effective when dealing with a large solution space. + +Modular Scalability. In Table 3c, we investigate the modular scalability by employing various LLM cores, including GPT-3.5, GPT-4 and open-source LLM Vicuna-7b [74]. As the LLMs' capabilities improve, ACTOR exhibits a continuous enhancement in its behavior simulation performance. This observation substantiates the notion that our ACTOR system possesses great potential for accommodating the de + +(c) Modular Scalability +Table 3. Ablative experiments on the Dynamic subset of our proposed BEHAVIORHUB. Please refer to §6.2 for more details. + +
MethodMPJPE↓MPVPE↓
[75] [CVPR2021]242.50222.13
†[75] [CVPR2021]201.56189.21
+ +(a) Scene-aware Motion Generation + +
MethodFID↓R Precision ↑
MDM [76] [ICLR2023]0.5440.611
†MDM [76] [ICLR2023]0.4710.705
+ +(b) Language-conditioned Motion Generation + +Table 4. Quantitative results (§6.3) on two downstream tasks. † indicates using BEHAVIORHUB for additional training. + +velopment of more powerful LLM cores. Moreover, this shows that BEHAVIORHUB, in conjunction with the human behavior planning task, serves as a robust testbed for evaluating the planning capability of LLMs. + +# 6.3. BEHAVIORHUB for Downsteam Task + +We further probe the effectiveness of BEHAVIORHUB by incorporating it as additional training data for two downstream tasks: scene-aware [75] and language-conditioned motion generation [76], which focus on generation without planning. We follow the official implementation of both methods and begin by pre-training the two models on BEHAVIORHUB. Subsequently, we fine-tune then evaluate these models on PROX [52] and HumanML3D [77], respectively. The evaluation results are reported in Table 4. As observed, BEHAVIORHUB significantly enhances model performance across all evaluation metrics for both tasks, highlighting the potential of our created dataset in facilitating broader applications within the realm of motion generation. + +# 7. Conclusion + +We present ACTOR, an LLM-powered agent towards realistic simulation of human behavior in 3D scenes. ACTOR integrates an LLM controller to perform complex behavior through planning on goal decomposition guided by hierarchical activity prior. The value-driven mechanism further deepens its understanding of environment. We demonstrate its potential in the simulated 3D indoor environment constructed using our newly created large-scale, scene-aware, behavior-rich dataset, BEHAVIORHUB. Evaluations suggest effective behavior planning and simulation of ACTOR. + +Acknowledgement. This work was supported by the National Science and Technology Major Project (No. 2023ZD0121300), National Natural Science Foundation of China (No. 62372405), Fundamental Research Funds for the Central Universities (226-2025-00057), National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University (No. HMHAI-202403), China Postdoctoral Science Foundation (No. 2025T180421), and the Postdoctoral Fellowship Program of CPSF (No. GZC20251066). + +# References + +[1] Denis Diderot. Diderot's early philosophical works. Number 4. Open Court, 1911. 1 +[2] Mark O. Riedl. Interactive narrative: A novel application of artificial intelligence for computer games. In AAAI, 2012. 1 +[3] Mark O Riedl. Human-centered artificial intelligence and machine learning. Human behavior and emerging technologies, 1(1), 2019. 1 +[4] Chen Liang, Zhicheng Shi, Wenguan Wang, and Yi Yang. Dual reciprocal learning of language-based human motion understanding and generation. In ICCV, 2025. 1 +[5] Elias Matsas and George-Christopher Vosniakos. Design of a virtual reality training system for human-robot collaboration in manufacturing tasks. IJIDEM, 11, 2017. 1 +[6] Joon Sung Park, Lindsay Popowski, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Social simulacra: Creating populated prototypes for social computing systems. In UIST, 2022. 1 +[7] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724, 2023. 1 +[8] Xiaohan Wang, Wenguan Wang, Jiayi Shao, and Yi Yang. Lana: A language-capable navigator for instruction following and generation. In CVPR, 2023. +[9] Hanqing Wang, Wei Liang, Luc V Gool, and Wenguan Wang. Towards versatile embodied navigation. NeurIPS, 2022. +[10] Zhichao Zhai, Guikun Chen, Wenguan Wang, Dong Zheng, and Jun Xiao. Taga: Self-supervised learning for template-free animatable gaussian articulated model. In CVPR, 2025. 1 +[11] Xavier Puig, Kevin Kyunghwan Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. Virtualhome: Simulating household activities via programs. In CVPR, 2018. 2, 3, 5 +[12] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 2, 3 +[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 6 + +[14] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 2, 6 +[15] Hongwei Yi, Chun-Hao P Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J Black. Human-aware object placement for visual environment reconstruction. In CVPR, 2022. 2, 6 +[16] Joseph Bates. The role of emotion in believable agents. Communications of the ACM, 37(7), 1994. 2 +[17] Robert Zubek. Towards implementation of social interaction. In AAAI, 2002. 2 +[18] Ruth Aylett. Narrative in virtual environments—towards emergent narrative. In AAAI, 1999. +[19] Michael Brenner. Creating dynamic story plots with continual multiagent planning. In AAAI, 2010. 2 +[20] Ho Chit Siu, Jaime Pena, Edenna Chen, Yutai Zhou, Victor Lopez, Kyle Palko, Kimberlee Chang, and Ross Allen. Evaluation of human-ai teams for learned and rule-based agents in hanabi. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, NeurIPS, 2021. 2 +[21] Michele Colledanchise and Petter Ögren. Behavior trees in robotics and AI: An introduction. CRC Press, 2018. 2 +[22] Ilshat Umarov, Mikhail Mozgovoy, and Patrick C. Rogers. Believable and effective ai agents in virtual worlds: Current state and future perspectives. International Journal of Gaming and Computer-Mediated Simulations, 4(2), 2012. 2 +[23] Christopher Berner, Greg Brockman, Brooke Chan, Vicki Cheung, Przemyslaw Debiak, Christy Dennison, David Farhi, Quirin Fischer, Shariq Hashme, Chris Hesse, Rafal Jozefowicz, Scott Gray, Catherine Olsson, Jakub Pachocki, Michael Petrov, Henrique P. d.O. Pinto, Jonathan Raiman, Tim Salimans, Jeremy Schlatter, Jonas Schneider, Szymon Sidor, Ilya Sutskever, Jie Tang, Filip Wolski, and Susan Zhang. Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680, 2019. 2 +[24] Oriol Vinyals, Igor Babuschkin, Wojciech M. Czarnecki, and et al. Grandmaster level in starcraft ii using multi-agent reinforcement learning. Nature, 575, 2019. 2 +[25] Matthew Hausknecht, Prithviraj Ammanabrolu, Marc-Alexandre Cote, and Xinyu Yuan. Interactive fiction games: A colossal adventure. In AAAI, 2020. 2 +[26] Shohei Miyashita, Xinyu Lian, Xiao Zeng, Takashi Matsubara, and Kuniaki Uehara. Developing game ai agent behaving like human by mixing reinforcement learning and supervised learning. In SNPD, 2017. +[27] Chen Tessler, Shahar Givony, Tom Zahavy, Daniel Mankowitz, and Shie Mannor. A deep hierarchical approach to lifelong learning in apache. In AAAI, 2017. 2 +[28] John Laird. It knows what you're going to do: Adding anticipation to a quakebot. In Proceedings of the 2001 Workshop on Intelligent Cinematography and Editing, 2001. 2 +[29] Dong kyu Choi, Tolga Konik, Negin Nejati, Chunki Park, and Pat Langley. A believable agent for first-person shooter games. In AAAI, 2021. 3 +[30] Pat Langley, Dongkyu Choi, and Seth Rogers. Interleaving learning, problem solving, and execution in the icarus architecture. Technical report, Stanford University, Center for the Study of Language and Information, 2005. 2 + +[31] Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In UIST, 2023. 3 +[32] Yujie Lu, Weixi Feng, Wanrong Zhu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Neurosymbolic procedural planning with commonsense prompting. In ICLR, 2023. 3 +[33] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In ICML, 2022. 3, 7 +[34] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. In NeurIPS, 2023. 3, 7 +[35] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In NeurIPS, 2023. +[36] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In ICML, 2024. 3 +[37] Shima Imani, Liang Du, and Harsh Shrivastava. Math-prompter: Mathematical reasoning using large language models. In ICLR Workshops, 2023. 3 +[38] Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen McAleer, Albert Q. Jiang, Jia Deng, Stella Biderman, and Sean Welleck. Llemma: An open language model for mathematics. arXiv preprint arXiv:2310.06786, 2023. 3 +[39] Xiang Lorraine Li, Adhiguna Kuncoro, Jordan Hoffmann, Cyprien de Masson d'Autume, Phil Blunsom, and Aida Nematzadeh. A systematic investigation of commonsense knowledge in large language models. In EMNLP, 2022. 3 +[40] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In ICRA, 2023. 3 +[41] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3 +[42] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 3 +[43] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In CoRL, 2023. 3 +[44] Wenlong Huang, Fei Xia, Dhruv Shah, Danny Driess, Andy Zeng, Yao Lu, Pete Florence, Igor Mordatch, Sergey Levine, + +Karol Hausman, et al. Grounded decoding: Guiding text generation with grounded models for robot control. arXiv preprint arXiv:2303.00855, 2023. 3 +[45] Belinda Z Li, Maxwell Nye, and Jacob Andreas. Implicit representations of meaning in neural language models. In ACL, 2021. 3 +[46] Adam Roberts, Colin Raffel, and Noam Shazeer. How much knowledge can you pack into the parameters of a language model? In EMNLP, 2020. +[47] Yunqiu Xu, Linchao Zhu, and Yi Yang. Mc-bench: A benchmark for multi-context visual grounding in the era of mllms. In ICCV, 2025. 3 +[48] Vadim Borisov, Kathrin Sessler, Tobias Leemann, Martin Pawelczyk, and Gjergji Kasneci. Language models are realistic tabular data generators. In ICLR, 2023. 3 +[49] Yew Ken Chia, Lidong Bing, Soujanya Poria, and Luo Si. Relationprompt: Leveraging prompts to generate synthetic data for zero-shot relation triplet extraction. In ACL, 2022. 3 +[50] Timo Schick and Hinrich Schütze. Generating datasets with pretrained language models. In EMNLP, 2021. 3 +[51] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. In ACL, 2023. 3 +[52] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019. 3, 8 +[53] Hongming Zhang, Xin Liu, Haojie Pan, Yangqiu Song, and Cane Wing-Ki Leung. Aser: A large-scale eventuality knowledge graph. In WWW, 2020. 3 +[54] Peter West, Chandra Bhagavatula, Jack Hessel, Jena Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. In NAACL, 2022. 3 +[55] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. In NeurIPS, 2022. 3 +[56] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 3 +[57] Yudi Dai, Yitai Lin, Chenglu Wen, Siqi Shen, Lan Xu, Jingyi Yu, Yuexin Ma, and Cheng Wang. Hsc4d: Human-centered 4d scene capture in large-scale indoor-outdoor space using wearable imus and lidar. In CVPR, 2022. 3 +[58] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. 3 +[59] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Xuan Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In NeurIPS, 2021. 3 + +[60] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, 2023. 4 +[61] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In ECCV, 2020. 4, 6 +[62] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, 2022. 4 +[63] Thomas A Feo and Mauricio GC Resende. Greedy randomized adaptive search procedures. Journal of global optimization, 6, 1995. 5 +[64] Markus Freitag and Yaser Al-Onaizan. Beam search strategies for neural machine translation. In ACL, 2017. 5 +[65] Rémi Coulom. Efficient selectivity and backup operators in monte-carlo tree search. In CG, 2006. 5 +[66] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In ICLR, 2019. 6 +[67] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In CVPR, 2021. 6 +[68] Santhosh K. Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westry, Angel Xuan Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. In NeurIPS Datasets and Benchmarks Track, 2021. 6 +[69] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6 +[70] Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan. Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021. 6 +[71] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 6 +[72] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6 +[73] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 6 +[74] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An opensource chatbot impressing gpt-4 with $90\%$ * chatgpt quality. https://vicuna.lmsys.org, 2023. 8 +[75] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In CVPR, 2021. 8 + +[76] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 8 +[77] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 8 \ No newline at end of file diff --git a/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/images.zip b/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cf441d830f7387a9aabf602be850d0fa23dc9574 --- /dev/null +++ b/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:551b5c511b7d3abfb62352826d3032f34f0e3adcd79525f2aa36a92d3e78f363 +size 470515 diff --git a/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/layout.json b/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..af0bf902f1bb015e8ab5368c464a443bcdf12585 --- /dev/null +++ b/2025/Towards Human-like Virtual Beings_ Simulating Human Behavior in 3D Scenes/layout.json @@ -0,0 +1,9296 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 550, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 550, + 120 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 550, + 120 + ], + "type": "text", + "content": "Towards Human-like Virtual Beings: Simulating Human Behavior in 3D Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 143, + 408, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 143, + 408, + 159 + ], + "spans": [ + { + "bbox": [ + 198, + 143, + 408, + 159 + ], + "type": "text", + "content": "Chen Liang, Wenguan Wang*, Yi Yang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 183, + 165, + 424, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 165, + 424, + 177 + ], + "spans": [ + { + "bbox": [ + 183, + 165, + 424, + 177 + ], + "type": "text", + "content": "State Key Lab of Brain-Machine Intelligence, Zhejiang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 197, + 184, + 411, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 197, + 184, + 411, + 196 + ], + "spans": [ + { + "bbox": [ + 197, + 184, + 411, + 196 + ], + "type": "text", + "content": "https://github.com/leonnnop/BehaviorHub" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 224, + 173, + 305 + ], + "blocks": [ + { + "bbox": [ + 58, + 224, + 173, + 305 + ], + "lines": [ + { + "bbox": [ + 58, + 224, + 173, + 305 + ], + "spans": [ + { + "bbox": [ + 58, + 224, + 173, + 305 + ], + "type": "image", + "image_path": "07bd40985b831dd2c2e9e9a990d588f6834ad77a0657139b68a6c44edaf4f3bb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 235, + 196, + 247 + ], + "lines": [ + { + "bbox": [ + 183, + 235, + 196, + 247 + ], + "spans": [ + { + "bbox": [ + 183, + 235, + 196, + 247 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 312, + 168, + 399 + ], + "blocks": [ + { + "bbox": [ + 58, + 312, + 168, + 399 + ], + "lines": [ + { + "bbox": [ + 58, + 312, + 168, + 399 + ], + "spans": [ + { + "bbox": [ + 58, + 312, + 168, + 399 + ], + "type": "image", + "image_path": "36b94a56ebda57f74e33e1e20c8365cb7b1e24cf5bf070eec62586faa2b9b3dd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 403, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 555, + 425 + ], + "type": "text", + "content": "Figure 1. Previous research has primarily focused on: (a) linguistic-based short-horizon task planning, and (b) low-level human-scene interaction. (c) This study investigates the simulation of high-level, long-horizon, abstract goal-driven human behaviors in 3D scenes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 184, + 262, + 334, + 385 + ], + "blocks": [ + { + "bbox": [ + 184, + 262, + 334, + 385 + ], + "lines": [ + { + "bbox": [ + 184, + 262, + 334, + 385 + ], + "spans": [ + { + "bbox": [ + 184, + 262, + 334, + 385 + ], + "type": "image", + "image_path": "fd3241e930f71c9da4e9b4e61e4111d9a775817de5fb493da588fc2527c0519c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 388, + 460, + 399 + ], + "lines": [ + { + "bbox": [ + 268, + 388, + 460, + 399 + ], + "spans": [ + { + "bbox": [ + 268, + 388, + 460, + 399 + ], + "type": "text", + "content": "Human Behavior Simulation in 3D Scenes (Ours)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 339, + 254, + 553, + 385 + ], + "blocks": [ + { + "bbox": [ + 282, + 225, + 457, + 249 + ], + "lines": [ + { + "bbox": [ + 282, + 225, + 457, + 249 + ], + "spans": [ + { + "bbox": [ + 282, + 225, + 457, + 249 + ], + "type": "text", + "content": "High-level, Long-horizon, Abstract Goal: Preparing for work from home" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 339, + 254, + 553, + 385 + ], + "lines": [ + { + "bbox": [ + 339, + 254, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 339, + 254, + 553, + 385 + ], + "type": "image", + "image_path": "b32ffb288e004c4551a45a545ccbb478b03c517913c899d4c5d592053290cef4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 152, + 441, + 200, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 441, + 200, + 453 + ], + "spans": [ + { + "bbox": [ + 152, + 441, + 200, + 453 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 466, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 466, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 54, + 466, + 297, + 694 + ], + "type": "text", + "content": "Building autonomous agents that can replicate human behavior in the realistic 3D world is a key step toward artificial general intelligence. This requires agents to be holistic goal achievers and to naturally adapt to environmental dynamics. In this work, we introduce ACTOR, an agent capable of performing high-level, long-horizon, abstract goals in 3D households, guided by its internal value similar to those of humans. ACTOR operates in a perceive-plan-act cycle, extending the ungrounded, scene-agnostic LLM controller with deliberate goal decomposition and decision-making through actively searching the behavior space, generating activity choices based on a hierarchical prior, and evaluating these choices using customizable value functions to determine the subsequent steps. Furthermore, we introduce BEHAVIORHUB, a large-scale human behavior simulation dataset in scene-aware, complicated tasks. Considering the unaffordable acquisition of human-authored 3D human behavior data, we construct BEHAVIORHUB by exploring the commonsense knowledge of LLMs learned from large" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 442, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 556, + 515 + ], + "type": "text", + "content": "corpora, and automatically aligning motion resources with 3D scene for knowledgeable generation. Extensive experiments on our established benchmark demonstrate that the proposed architecture leads to effective behavior planning and simulation. BEHAVIORHUB also proves beneficial for downstream task development." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 538, + 394, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 538, + 394, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 538, + 394, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 559, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 559, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 559, + 556, + 715 + ], + "type": "text", + "content": "Building autonomous agents (e.g., virtual beings, or humanoid robots) that can replicate human behavior in the realistic 3D world, has been a long-standing pursuit since the inception of AI [1]. The study could empower nonplayer game character [2], underpin human-robot interaction [3, 4] and cooperation [5], populate virtual reality communities [6], and accelerate Embodied AI [7-10]. Much progress has been made in vision and language models that imitate human motions (cf. Fig. 1 (b)) or propose linguistic plans (cf. Fig. 1 (a)). However, an effective 3D humanoid agent must go beyond by conquering three major barriers: (i) Holistic goal achievement from perception to action. As depicted in Fig. 1 (c), to accomplish a high-level goal (e.g.," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 704, + 200, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 704, + 200, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 704, + 200, + 714 + ], + "type": "text", + "content": "*Corresponding author: Wenguan Wang." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10753" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 300 + ], + "type": "text", + "content": "'prepare for work'), the agent must process the perceived information (e.g., 'lying on bed'), decompose the goal into a series of activities (e.g., 'get up', 'use toilet', 'eat breakfast', etc.), and devise appropriate action plans for each; (ii) Environmental dynamics. Agents should be able to actively adjust their plans based on the environment, e.g., determining if the bathroom is occupied when planning to use toilet; (iii) Vast and multifaceted human behavior space, where numerous viable paths exist to achieve even a single goal. For example, agents preparing for work can choose to eat breakfast before using toilet or vice versa. Also, when the bathroom is occupied, an intelligent agent can decide to whether eat breakfast first, or continue waiting, depending on its state and beliefs, e.g., the desire to 'complete goal as soon as possible' or 'save energy costs'. A competent agent must process such value priorities to guide its selection and evaluation of actions and policies. Besides, the absence of a comprehensive testbed further poses great challenge for agent development and evaluation, constituting barrier (iv)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 304, + 294, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 304, + 294, + 638 + ], + "spans": [ + { + "bbox": [ + 57, + 304, + 294, + 638 + ], + "type": "text", + "content": "In this work, we present ACTOR - a large language model (LLM) powered agent towards diligent simulation of human behavior in 3D realistic scenes (§4). ACTOR follows a perceive-plan-act cycle, addressing challenge " + }, + { + "bbox": [ + 57, + 304, + 294, + 638 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 57, + 304, + 294, + 638 + ], + "type": "text", + "content": " as envisioned. Using LLM as a central controller, it strategizes plans by searching the human behavior space. It actively maintains a tree of behaviors, where each node represents an intermediate step toward holistic goal achievement. The construction of this tree is guided by a hierarchical prior, i.e., executable low-level actions are grouped into high-level semantic units, called activities, iteratively forming a hierarchical structure (cf. Fig. 1 (c)). The search progress is assessed using a set of customizable value functions that determine the likelihood of different intermediate candidates. In addition to being rational in common sense, ACTOR couples real-valued evaluations (e.g., best efficiency) with personalized priors expressed through language commands (e.g., description of a neat person). These outputs are converted into unified probabilities, allowing for the incorporation of the agent's characteristics and beliefs as value functions, thus addressing challenge (iii). The planning process is dynamic and grounded in specific environmental values, empowering ACTOR to readjust its plans when faced with environmental changes or new language commands, effectively tackling challenge (ii). This formulation facilitates the use of powerful search algorithms, e.g., greedy search, beam search, and Monte Carlo tree search (MCTS), with MCTS proving to be the most effective in our experiments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 642, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 642, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 642, + 294, + 712 + ], + "type": "text", + "content": "Furthermore, we establish a comprehensive environment for development and evaluation of agents like ACTOR, based on our newly proposed large-scale, scene-aware, behavior-rich dataset, dubbed BEHAVIORHUB (§5). One critical issue in constructing the human-authored benchmark is the high cost associated with acquiring and scaling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 316, + 72, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 323 + ], + "type": "text", + "content": "high-quality, human-generated daily behavior data. Moreover, annotating large-scale behavior data further requires creativity in designing novel tasks and expertise in creating complete plans from scratch, which is also a challenging task for humans [11]. Given this context, we propose to automatically synthesize 3D human behavior data by enhancing existing resources. Initially, we distill the tree-structured linguistic plans of human daily behaviors from LLM using in-context learning [12]. Contactable objects and plausible interactions from the scanned 3D scenes (e.g., ScanNet [13], etc.) and captured motion sequences (e.g., AMASS [14], etc.) are attributed into the in-context prompt to ensure viable plans grounding in certain environment. Then, we align the task-motion-scene data triplets by applying collision and contact constraints [15] for valid translation and rotation parameters. To promote diversity, we sample multiple plausible motion sequences for each high-level goal. In total, BEHAVIORHUB contains more than 1k daily goals over 10k high-quality behavior samples of 15.7 steps on average covering 1.5k scenes, that establishes a comprehensive testbed, that addresses the barrier " + }, + { + "bbox": [ + 316, + 72, + 553, + 323 + ], + "type": "inline_equation", + "content": "(iv)" + }, + { + "bbox": [ + 316, + 72, + 553, + 323 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 324, + 553, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 324, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 324, + 553, + 454 + ], + "type": "text", + "content": "We conduct extensive evaluations of ACTOR in §6. First, we find that ACTOR produces admissible plans and generates plausible motion sequences to simulate abstract, temporally-extended human behaviors. It outperforms the strong baselines by nearly doubling the overall success rate. This result is further confirmed by human evaluation. Then, we conduct several ablative experiments that limit ACTOR's access to each core design for thorough assessment. Finally, experiments on scene-aware and language-conditioned human motion generation demonstrate how BEHAVIORHUB can benefit the development of downstream task models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 466, + 400, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 400, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 400, + 478 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 486, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 486, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 486, + 553, + 712 + ], + "type": "text", + "content": "Human Behavior Simulation. Simulating human behavior in realistic, open world environment like the one we inhabit is a long-standing topic in artificial intelligence [16]. Historically, the topic has primarily been studied in the game worlds, focusing on enhancing player experiences through intelligent non-player game characters (NPCs) [17-19]. Early approaches rely on rule-based approaches like finite-state machines [20] and behavior trees [21]. They provide a brute force way of manually crafting agent's behaviors, but cannot perform new procedures that were not hard-coded in their script [22], limiting the generalizability. Another strand [23, 24] involves using reinforcement learning, where agents learn its own policy through optimizing the learning algorithm on readily definable rewards over fixed task space, which is often limited to non-open, adversarial games, or blocks worlds only [25-27]. The third line of research, represented by some pioneer works [28-30] in computational cognition, aims to build machines that operate directly in perceive-plan-act cycles, encompassing" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10754" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "the nature of autonomous agents as originally envisioned. This formulation holds potential generalizability to most, if not all, open-world contexts. However, these studies are typically restricted to simplified environments, such as first-person shooter games [29], or 2D gridworlds (e.g., Generative Agents [31]), and focus on a reduced range of behaviors. Our work falls in vein of third category, while pushing frontier towards simulation in realistic 3D environments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 295, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 337 + ], + "type": "text", + "content": "Recently, beyond the game world, new trending on social embodied intelligence, such as assistive robots, draws attention to simulation of human collaborators' behavior. One notable effort is VirtualHome [11], which studies high-level human activities as plain sequences of atomic actions. However, its formulation is inherently environment-agnostic, which has been treated as a purely linguistic procedural planning problem in subsequent studies [32, 33]. Furthermore, the discrete action space built upon manually crafted procedural knowledge also makes it an insufficient testbed. In contrast, with a special focus on environment-aware simulation of contiguous daily behaviors, our work paves one solid step accelerating the development of social embodied intelligence." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 340, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 628 + ], + "type": "text", + "content": "LLM as Planner. Remarkable progress in LLMs has been observed, demonstrating their emerging capacity to break down complex tasks into more manageable sub-tasks and devise appropriate plans for each [34-36]. LLMs have been successfully applied to solving mathematical problems [37, 38], reasoning on commonsense [39], planning robotics tasks [40, 41], and very recently, manipulating external APIs on a web scale, expanding their capabilities beyond text generation. Based on the observation, we posit that LLMs can serve as a crucial component in extending the perception-decision-action space to construct human-like agents in realistic 3D environments. However, a crucial challenge remains: LLMs lack experience and interaction with their environment [12, 42], preventing them from ordering actionable and rational plans. This paper addresses this issue by incorporating customizable value functions into LLMs to evaluate and prioritize plans. The idea bears some resemblance to recent robotics research [43, 44]. However, those methods necessitate retraining for every new set of primitive robotic skills, making them impractical for the complex and undefined human action space, which is difficult to define in advance. In contrast, we support plan evaluation using real-numbered functions and language-based rules without resource-intensive retraining." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "LLM as Data Generator. Being trained on the large corpus of human-produced language, LLMs are believed to contain a wealth of information about the world [45-47]. Given a handful of task-specific prompts, LLMs can generalize and generate more linguistic data in the same format, with the application of generating tabular data [48], relation triplets [49], sentence pairs [50], instruction data [51]," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "etc. The idea seems naturally to be borrowed for human behavior data acquisition, where the requested human motion and daily activity procedure were commonly crowdsourced with high expense and complexity [11], limiting the scale and coverage of related datasets [52]. However, the process is non-trivial. The generated data is often blamed for low quality and diversity issues [53, 54]. The 3D environment-aware nature of the task also poses unique challenges. To respond, we explore attributed prompts specifically conditioned on the environment that not only mitigates the problems of low informativeness and redundancy, but also offers an effective workflow that can further empower other related domains, such as human-scene interaction [55]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 240, + 397, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 240, + 397, + 251 + ], + "spans": [ + { + "bbox": [ + 314, + 240, + 397, + 251 + ], + "type": "text", + "content": "3. Environment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 260, + 553, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 553, + 320 + ], + "type": "text", + "content": "Simulating open-ended goals that resemble naturalistic human behaviors necessitates an environment capable of facilitating diverse agent affordances and interactions. Before delving into the detailed agent architecture, we first describe such environment we tailored for agents to instantiate in." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 321, + 553, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 553, + 344 + ], + "type": "text", + "content": "Environmental Setup. The environment features common affordances in a household, including:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 345, + 553, + 464 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 345, + 553, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 345, + 553, + 368 + ], + "spans": [ + { + "bbox": [ + 314, + 345, + 553, + 368 + ], + "type": "text", + "content": "- Scene of 3D textured meshes, that spans a house of multiple functional areas (e.g., kitchen, etc.)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 369, + 553, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 369, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 314, + 369, + 553, + 405 + ], + "type": "text", + "content": "- Objects embedded in the scene (e.g., stove in kitchen, bed in bedroom), with each constructed of 3D textured geometry and corresponding object state (e.g., fridge: opened)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "text", + "content": "- Humanoid agent(s) defined by SMPL-X [56], an expressive 3D human model of shape and pose of both body and hand. Agents reside in and interact with the scene and can influence the state of objects by their actions (e.g., close the fridge | fridge: opened " + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 405, + 553, + 464 + ], + "type": "text", + "content": " closed)." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 465, + 553, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 553, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 553, + 501 + ], + "type": "text", + "content": "As a start point, we demonstrate the environment of indoor household, keep the general formulation open for other environments, such as outdoor streets [57]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 502, + 554, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 554, + 621 + ], + "type": "text", + "content": "Environmental Simulator. Interactions within the environment are driven by the simulator of two components: (i) an engine that manages and evolves the states of objects in the environment; and (ii) a renderer that supports generation of multiple perceptual observations (e.g., RGB, depth, 3D surfaces) for agents. At each time step, the simulator dynamically updates environment and collects egocentric, surround, or third-person-view information based on needs. We build our simulator upon Habitat-Sim [58, 59] to ensure efficient and parallelizable simulation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 633, + 403, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 403, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 403, + 647 + ], + "type": "text", + "content": "4. ACTOR Agent" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 715 + ], + "type": "text", + "content": "We aim to build humanoid agents that naturally simulate 3D human behaviors to complete daily goals, which can be lengthy, abstract, or ambiguous. Agents receive high-level goal described in language and are tasked with generating plausible motion sequences that align with the given scene." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10755" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 553, + 184 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 553, + 184 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 553, + 184 + ], + "type": "image", + "image_path": "383f20063d400ecf9bd55b6c27206e86d702832445b0a037574d7e16aac007c7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 194, + 548, + 205 + ], + "lines": [ + { + "bbox": [ + 60, + 194, + 548, + 205 + ], + "spans": [ + { + "bbox": [ + 60, + 194, + 548, + 205 + ], + "type": "text", + "content": "Figure 2. Left: Overall agent architecture (§4.1); Right: One-step in perceive-plan-act loop with value-driven behavior planning (§4.2)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 168, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 168, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 168, + 230 + ], + "type": "text", + "content": "4.1. Agent Architecture" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 235, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 296, + 342 + ], + "type": "text", + "content": "Fig. 2 shows the workflow of ACTOR. The agent operates in a perceive-plan-act loop, using language as the generic interface connecting the three phases. At each step, it selectively perceives the world based on the target, transforming information into an environment description using heuristics. The observed information assists in decision-making through the LLM core, ultimately resulting in actions in the form of 3D motion sequences based on linguistic plans within the environment. We now delve into the details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 438 + ], + "type": "text", + "content": "Perception. Our definition of perception extends beyond gathering information, such as 2D images or 3D point clouds, through the simulator (§3). It encompasses attaining a deep understanding of the environment, including object properties, spatial relationships, and scene layouts, etc. In our preliminary implementation, information is perceived using readily available models and converted into linguistic descriptions using heuristic functions (supp. §A.2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 439, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 521 + ], + "type": "text", + "content": "Plan. At the core of ACTOR, a planning module receives the current environment and past behavior trial as input, generating action description as output. For the executable actions, it provides descriptions of motion sequences. In the case of high-level activity, it bypasses the action stage and proceeds to break down the target further. We will provide a detailed explanation of this process in §4.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 296, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 630 + ], + "type": "text", + "content": "Action. For action, we specifically consider whole-body human actions in 3D scenes to closely resemble human behavior using the off-the-shelf models. It generates whole-body motion based on text and trajectory [60]. For object-interactive actions, we further refine hand grasping using an isolated grasp estimation model [61]. Additionally, for moving actions like walking, trajectory paths are pre-estimated [62]. Detailed trajectory estimation process is provided in supp. §A.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 636, + 230, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 230, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 230, + 649 + ], + "type": "text", + "content": "4.2. Value-driven Behavior Planning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "An LLM, denoted as " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", functions as a controller to iteratively decompose the long-term goal " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " described in language texts into shorter steps of behaviors based on the partial environmental dynamics. To prevent confusion with the global environment " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": ", we denote the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "text", + "content": "partial environmental observation at step " + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "inline_equation", + "content": "E_{i}^{\\prime}" + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "text", + "content": " and its corresponding textual description as " + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "inline_equation", + "content": "e_{i}" + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "text", + "content": ". For brevity, we omit step indicator " + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "text", + "content": " and define " + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "inline_equation", + "content": "E^{\\prime}" + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "text", + "content": " to include both the environment state and agent's state. We denote textual descriptions of behaviors as " + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "text", + "content": ". The problem is formulated as a search over a tree, where each node is a state " + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "inline_equation", + "content": "s = \\{z_{1\\dots i}\\}" + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "text", + "content": ", representing a partial trial with the input and the sequence of behaviors taken thus far. The instantiation consists of three components: (i) Node expansion, which involves generating " + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 218, + 555, + 361 + ], + "type": "text", + "content": " candidate behaviors for the next step search; (ii) Value functions to evaluate each node; (iii) Search algorithm that accounts for branch selection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 363, + 556, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 556, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 556, + 507 + ], + "type": "text", + "content": "Node Expansion. The node is expanded through sampling from LLM with a window size " + }, + { + "bbox": [ + 313, + 363, + 556, + 507 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 363, + 556, + 507 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 363, + 556, + 507 + ], + "type": "inline_equation", + "content": "z^{(j)} \\sim p_{\\theta}(z_{i+1}|l,e,s,h)" + }, + { + "bbox": [ + 313, + 363, + 556, + 507 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 363, + 556, + 507 + ], + "type": "inline_equation", + "content": "j \\in \\{1,\\dots,w\\}" + }, + { + "bbox": [ + 313, + 363, + 556, + 507 + ], + "type": "text", + "content": ". We omit the basis prompting parts for brevity, which is also the case in the subsequent context. Here, we introduce a hierarchical heuristic, denoted as " + }, + { + "bbox": [ + 313, + 363, + 556, + 507 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 363, + 556, + 507 + ], + "type": "text", + "content": ", which is implemented through prompt instructions and demonstrations of specific actions and activity cases. This heuristic ensures that during each expansion, all candidates at the same level are restricted to executable actions or high-level semantic units of activities, that provides a more nuanced representation of behaviors for effectively modeling interchangeable activities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "text", + "content": "Value Function. The value function assesses the state by considering the degree to which a specific behavior contributes to the achievement of the target, conditioned on the agents' beliefs reflecting their value. The search algorithm uses the output to determine which nodes to explore further and in what order. The likelihood is given by " + }, + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "inline_equation", + "content": "p(z_{i+1} | l, E', s) \\propto p_{\\theta}(z_{i+1} | l, e, s) \\cdot p_v(z_{i+1} | l, E', s)" + }, + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "text", + "content": ". Here, the LLM provides us with " + }, + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "inline_equation", + "content": "p_{\\theta}(z_{i+1} | l, e, s)" + }, + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "text", + "content": ", which represents the likelihood, based on commonsense, that a textual behavior is a valid next step. However, the LLM struggles to generalize or make inferences in the real environment since it is not grounded. On the other hand, " + }, + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "inline_equation", + "content": "p_v" + }, + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "text", + "content": " provides the likelihood of the behavior being plausible in the current state of both the environment and agent, according to the defined values. We consider a set of values " + }, + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "inline_equation", + "content": "\\{p_{v_n}\\}_n" + }, + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "text", + "content": " categorized into two types, and " + }, + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "inline_equation", + "content": "p_v(z_{i+1} | l, E', s) = \\prod_n p_{v_n}(z_{i+1} | l, E', s)" + }, + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 555, + 714 + ], + "type": "text", + "content": "Real-valued function assigns real values as outputs, such" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10756" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 220, + 205 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 220, + 205 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 220, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 220, + 205 + ], + "type": "image", + "image_path": "76b5d3c4930ff393da2ced98b0fc7f3815d0101d5c31190c561e7a652961624d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 213, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 553, + 225 + ], + "type": "text", + "content": "Figure 3. (a) An illustrative example of our BEHAVIORHUB dataset; (b) Semi-automatic 3D human behavior data generation pipeline (" + }, + { + "bbox": [ + 55, + 213, + 553, + 225 + ], + "type": "inline_equation", + "content": "\\S 5" + }, + { + "bbox": [ + 55, + 213, + 553, + 225 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 72, + 553, + 205 + ], + "blocks": [ + { + "bbox": [ + 224, + 72, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 224, + 72, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 224, + 72, + 553, + 205 + ], + "type": "image", + "image_path": "8ee00300b55710f3d54b0a1c8135779ab602fd33bd84da53d3f0c46e0a94e0da.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 239, + 294, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 239, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 63, + 239, + 294, + 274 + ], + "type": "text", + "content": "as the shortest path value, for which the function estimates the distance for each candidate action. The outputs of real-valued functions are normalized into probabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 274, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 346 + ], + "type": "text", + "content": "- Language-based command is implemented by prompting the LM with a value prompt that conveys the meaning of 'How likely is it for someone who is a neat person to take the following action, considering ...'. The model then reasons about each action to generate the corresponding probability reflecting the value." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 346, + 296, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 454 + ], + "type": "text", + "content": "Values are also conditioned on the state of the environment and agent. This approach allows for reacting to environmental dynamics in addition to the active planning process. Search Algorithms. Owing to the active planning process we have formulated, it is feasible to use different search algorithms based on the tree structure. We explore greedy search [63], beam search [64] and MCTS [65], and evaluate their performance in experiments (cf. §6.2) where we find MCTS performs best. We therefore use MCTS by default." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 461, + 185, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 185, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 185, + 474 + ], + "type": "text", + "content": "4.3. Implementation Detail" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": "We employ GPT-4 and GPT-3.5-turbo API provided by OpenAI as base LLMs. We prompt them with four in-context examples as demonstrations, with decoding temperature as 0 for more deterministic generation. We use official releases of conditional motion generation models and fine-tune them on BEHAVIORHUB using the default parameters for each to promote finer generation quality. By default, we set the window size to " + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "inline_equation", + "content": "w = 5" + }, + { + "bbox": [ + 55, + 479, + 296, + 635 + ], + "type": "text", + "content": " and normalize the shortest path value based on the maximum rollout depth of the search algorithm, which is set to 3. In practice, we find one-time sampling is sufficient for effective tree search, eliminating the need for multiple samplings for branch aggregation. Our algorithm is implemented in PyTorch using Tesla A40 GPUs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 646, + 216, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 216, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 216, + 658 + ], + "type": "text", + "content": "5. BEHAVIORHUB Benchmark" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "To lay a solid foundation for future research, we build a large repository of common behaviors performed in daily household scenarios. Each sample contains three components: " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " a high-level goal (i.e., root node); " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " a tree" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 239, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 555, + 406 + ], + "type": "text", + "content": "structured linguistic plan covering necessary intermediate-level steps (i.e., intermediate nodes) and low-level steps (i.e., leaf nodes) required to accomplish the goal; and (iii) scene-conditioned human motions corresponding to each executable step at either intermediate or low level. We illustrate one example in Fig. 3 (a). Intermediate step set can be empty. Both goal and steps in plans are human activities described with either concrete or abstract language (e.g., 'go to sleep on bed' or 'feel tired'). To generate diverse and high-quality data, we employ a two-step pipeline that utilizes an LLM pretrained on extensive corpora: First, automatic generation of linguistic daily plans (" + }, + { + "bbox": [ + 313, + 239, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\S 5.1" + }, + { + "bbox": [ + 313, + 239, + 555, + 406 + ], + "type": "text", + "content": "); Second, alignment of the plans with 3D motions and scenes (" + }, + { + "bbox": [ + 313, + 239, + 555, + 406 + ], + "type": "inline_equation", + "content": "\\S 5.2" + }, + { + "bbox": [ + 313, + 239, + 555, + 406 + ], + "type": "text", + "content": "). The entire data generation process is shown in Fig. 3 (b)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 417, + 491, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 491, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 491, + 430 + ], + "type": "text", + "content": "5.1. Linguistic Goal-Plan Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 436, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 554, + 484 + ], + "type": "text", + "content": "The pipeline consists of three steps: " + }, + { + "bbox": [ + 313, + 436, + 554, + 484 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 313, + 436, + 554, + 484 + ], + "type": "text", + "content": " generate potentially incomplete goal-plan trees; " + }, + { + "bbox": [ + 313, + 436, + 554, + 484 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 313, + 436, + 554, + 484 + ], + "type": "text", + "content": " complete and refine each tree; and " + }, + { + "bbox": [ + 313, + 436, + 554, + 484 + ], + "type": "inline_equation", + "content": "(iii)" + }, + { + "bbox": [ + 313, + 436, + 554, + 484 + ], + "type": "text", + "content": " filter out low-quality data. We provide detailed prompts in supp. §A.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 485, + 554, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 675 + ], + "type": "text", + "content": "Attributed Goal-Plan Tree Generation. In the initial step, we generate new goal-plan trees using a bootstrapping approach based on a small set of seed human-written samples. To ensure broader coverage and facilitate later alignment with specific scenes, we attribute the starting room, candidate objects and actions that interact with the activity in the plans into prompt demonstrations. We start the sample pool with 292 activities from ActivityPrograms [11], and consider 21 room types (e.g., bathroom, gym), " + }, + { + "bbox": [ + 313, + 485, + 554, + 675 + ], + "type": "inline_equation", + "content": "\\sim 10^3" + }, + { + "bbox": [ + 313, + 485, + 554, + 675 + ], + "type": "text", + "content": " object types (e.g., TV, table), and " + }, + { + "bbox": [ + 313, + 485, + 554, + 675 + ], + "type": "inline_equation", + "content": "\\sim 10^2" + }, + { + "bbox": [ + 313, + 485, + 554, + 675 + ], + "type": "text", + "content": " action types (e.g., open, eat). Refer to the supp. §A.5 for the full list. For each step, we sample eight goal-plans as in-context examples from the pool while restricting the outputs to one room and ten types of objects. Out of the eight plans, six are from the human-written plans, and two are generated by the model in previous steps to enhance diversity." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Generated trees are then labeled with sequential order for intermediate nodes in the trees, which encompass multiple sub-step leaf nodes of individual actions. This sequential or" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10757" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "der is represented as interchangeable groups, where nodes within the same set are interchangeable with each other. We prompt the LM in a few-shot way to determine this." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 294, + 204 + ], + "type": "text", + "content": "Goal-Plan Tree Refinement. We further improve the constructed trees by addressing two key aspects: (i) we complete any missing internal plan steps, which can often be revised based on commonsense, e.g., opened the fridge without closing it. (ii) we enhance the root node descriptions to be more abstract, e.g., transforming 'use toilet' to 'feel the call of nature'. We accomplish both aspects by querying the LLM for suggestions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 205, + 294, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 294, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 294, + 276 + ], + "type": "text", + "content": "Filtering. To promote diversity, a new tree is added to the pool only if its BERTScore similarity [66] with any existing goal-plan tree is below 0.5. Additionally, we utilize LLM to assess the generated trees by asking the question, 'Is this a valid plan?' Any plans flagged as 'invalid' are filtered out to ensure high-quality data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 221, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 221, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 221, + 298 + ], + "type": "text", + "content": "5.2. Goal-Motion-Scene Alignment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 303, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 295, + 422 + ], + "type": "text", + "content": "Once the goal-plan trees are formed, we proceed to ground them in the 3D environment. We propose to synthesize 3D behaviors by leveraging existing resources of human motions (i.e., AMASS [14], BABEL [67], GRAB [61]), and indoor scenes (i.e., ScanNet [13], HM3D [68], Replica [69]). For each tree, we first sample actions and objects from resources using corresponding labels to fulfill the executable activities in forms of combinations of actions on objects. Then, given sampled motion, we aim to generate plausible and contiguous interactions to sampled object in scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 530 + ], + "type": "text", + "content": "Motion Alignment. We first put motion clips into the scene anchored by the contactable objects. We optimize the translation and rotation parameters matrices by minimizing the collision and contact losses from MOVER [15]. A unified SDF volume is calculated, and all contact vertices for all frames are accumulated in 3D space. The motion is aligned through a transformation of the two matrices. This joint optimization improves human-object contact and resolves 3D interpenetrations between humans and the scene." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 295, + 614 + ], + "type": "text", + "content": "Sequence Blending. The aligned motion sequence, which may be sparse and not spatially connected, is blended using a Transformer-based motion completion method [70]. This results in a contiguous motion sequence that aligns with both the scene and goal plan. In practice, we find a single network is capable of delivering satisfactory results that align with prior research [71]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 615, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 295, + 664 + ], + "type": "text", + "content": "Verification. Finally, to ensure the data quality, each example is examined by three verifiers who vote on whether the plan (motion) is complete (valid). If an example receives majority approval, it is accepted; otherwise, it is dropped." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 159, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 159, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 159, + 684 + ], + "type": "text", + "content": "5.3. Dataset Statistics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Our dataset consists of 10k human behavior samples in 1.5k 3D scenes, covers 2k unique activity over 0.1k actions and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "1k objects, which is an order of magnitude larger than the human-authoring ActivityPrograms [13]. On average, each high-level goal has 15.7 steps, resulting in a total of 10.1 corresponding motion sequences of 83.3 frames, that span over 8.6M motion frames. Each activity corresponds to 4.1 different motion sequences in 1.7 rooms. Owing to the automatic data generation pipeline, our dataset achieves even higher diversity and coverage with low human cost." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 177, + 422, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 422, + 189 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 422, + 189 + ], + "type": "text", + "content": "5.4. Evaluation Metric" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 194, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 255 + ], + "type": "text", + "content": "Our evaluation of the plausible behavior simulation encompasses two key aspects: " + }, + { + "bbox": [ + 313, + 194, + 553, + 255 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 313, + 194, + 553, + 255 + ], + "type": "text", + "content": " the simulation should generate linguistically reasonable behavior plans (i.e., behavior planning); and " + }, + { + "bbox": [ + 313, + 194, + 553, + 255 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 313, + 194, + 553, + 255 + ], + "type": "text", + "content": " align them with natural motion sequences within the 3D environment (i.e., behavior simulation)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 255, + 554, + 532 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 255, + 553, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 255, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 314, + 255, + 553, + 327 + ], + "type": "text", + "content": "- For behavior planning, we use Sentence-BLEU [72] and BERTScore [66] to measure the semantic similarity between the ground-truth plans and predictions. We report maximum scores attained across ground-truth plan variants to ensure the evaluation be order-invariant with respect to interchangeable sub-steps." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 327, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 327, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 314, + 327, + 554, + 532 + ], + "type": "text", + "content": "- For behavior simulation, we consider three sets of metrics: (i) Success Rate: The step success rate (SSR) records the percentage of steps where the agent successfully completes the step objective, defined by a contact distance threshold. For example, we consider lying down to be successful if both the hip and head of the humanoid are within " + }, + { + "bbox": [ + 314, + 327, + 554, + 532 + ], + "type": "inline_equation", + "content": "30\\mathrm{cm}" + }, + { + "bbox": [ + 314, + 327, + 554, + 532 + ], + "type": "text", + "content": " of the target location [73]. The goal success rate (GSR) is measured to determine whether all steps in the entire plan are successfully executed; (ii) Goal Success Rate Weighted by Path Length (GSRPL): It judges how efficient was the agent at finishing the goal, defined as " + }, + { + "bbox": [ + 314, + 327, + 554, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{GSR} \\cdot \\frac{g}{\\max(g,l)}" + }, + { + "bbox": [ + 314, + 327, + 554, + 532 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 314, + 327, + 554, + 532 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 314, + 327, + 554, + 532 + ], + "type": "text", + "content": " is ground-truth path length and " + }, + { + "bbox": [ + 314, + 327, + 554, + 532 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 314, + 327, + 554, + 532 + ], + "type": "text", + "content": " is the agent's path length; (iii) Motion Quality: We evaluate overall quality of generated motions using Frechet Inception Distance (FID) and recognition accuracy measured with the final layer of a pretrained standard RNN action recognition classifier as motion feature extractor." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 544, + 390, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 390, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 390, + 556 + ], + "type": "text", + "content": "6. Experiment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 563, + 537, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 537, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 537, + 575 + ], + "type": "text", + "content": "6.1. Performance on BEHAVIORHUB Benchmark" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 581, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 553, + 640 + ], + "type": "text", + "content": "Dataset Split and Dynamic Subset. We randomly sample 200 held-out goals as demonstration set from which we select example(s) for prompting language models, and also as training data for fine-tuning conditional generation models. The remaining ones are used for evaluation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "To thoroughly examine the capacity of benchmarked models in managing environmental dynamics, we manually create a subset of 300 samples from the original BEHAVIORHUB, called Dynamic subset. In each sample, we carefully configure the agent's environment state-aware triggers to guarantee a distinct goal-plan solution for a specific goal," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10758" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 70, + 553, + 180 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 553, + 180 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 553, + 180 + ], + "type": "table", + "html": "
MethodBehavior PlanningBehavior Simulation
S-BLEUBERT-SSSRGSRGSRPLFID ↓Accuracy ↑
Main SetLLMaP [33] [ICML2022]0.0890.825-----
HuggingGPT [34] [NeurIPS23]0.1320.8560.5330.3170.1615.3860.620
ACTOR (Ours)0.1700.8790.6010.4720.3512.0870.773
Dynamic SubsetHuman0.2030.959-----
LLMaP [33] [ICML2022]0.0690.821-----
HuggingGPT [34] [NeurIPS23]0.0990.8300.4070.1640.0739.1160.505
ACTOR (Ours)0.1350.8620.5150.3060.2123.1410.697
", + "image_path": "52e701d8fc71dea626c0feba350c53b68da8a2a4e67793841fd1bece7c0660d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 201, + 145, + 274 + ], + "blocks": [ + { + "bbox": [ + 59, + 201, + 145, + 274 + ], + "lines": [ + { + "bbox": [ + 59, + 201, + 145, + 274 + ], + "spans": [ + { + "bbox": [ + 59, + 201, + 145, + 274 + ], + "type": "image", + "image_path": "105c00aef1fc19ca71363c6bdb91069601bf2a46e8b2b6e2355da2fb4e4cf83f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 158, + 219, + 342, + 274 + ], + "blocks": [ + { + "bbox": [ + 158, + 201, + 331, + 219 + ], + "lines": [ + { + "bbox": [ + 158, + 201, + 331, + 219 + ], + "spans": [ + { + "bbox": [ + 158, + 201, + 331, + 219 + ], + "type": "text", + "content": "Original plan: weekend cleaning " + }, + { + "bbox": [ + 158, + 201, + 331, + 219 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 158, + 201, + 331, + 219 + ], + "type": "text", + "content": " clean bedroom " + }, + { + "bbox": [ + 158, + 201, + 331, + 219 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 158, + 201, + 331, + 219 + ], + "type": "text", + "content": " Strategized plan:" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 158, + 219, + 342, + 274 + ], + "lines": [ + { + "bbox": [ + 158, + 219, + 342, + 274 + ], + "spans": [ + { + "bbox": [ + 158, + 219, + 342, + 274 + ], + "type": "image", + "image_path": "c95c90ce58bc48753305ca24b79a7ad14e1b96d17237c7e7f24a68ef96c34a8b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 247, + 381, + 263 + ], + "lines": [ + { + "bbox": [ + 328, + 247, + 381, + 263 + ], + "spans": [ + { + "bbox": [ + 328, + 247, + 381, + 263 + ], + "type": "text", + "content": "(occupied signal removed)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 342, + 239, + 361, + 247 + ], + "blocks": [ + { + "bbox": [ + 342, + 239, + 361, + 247 + ], + "lines": [ + { + "bbox": [ + 342, + 239, + 361, + 247 + ], + "spans": [ + { + "bbox": [ + 342, + 239, + 361, + 247 + ], + "type": "image", + "image_path": "b64adb1e94242751db9dbea8501b12f0db81e7a48fff2aca043a39124028aafd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 365, + 219, + 427, + 274 + ], + "blocks": [ + { + "bbox": [ + 365, + 219, + 427, + 274 + ], + "lines": [ + { + "bbox": [ + 365, + 219, + 427, + 274 + ], + "spans": [ + { + "bbox": [ + 365, + 219, + 427, + 274 + ], + "type": "image", + "image_path": "a97d39aab47f096159cb2557502d9223afb04e2d113cadf24165a3ef007138b3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 428, + 227, + 489, + 274 + ], + "blocks": [ + { + "bbox": [ + 439, + 218, + 476, + 225 + ], + "lines": [ + { + "bbox": [ + 439, + 218, + 476, + 225 + ], + "spans": [ + { + "bbox": [ + 439, + 218, + 476, + 225 + ], + "type": "text", + "content": "Clean bedroom" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 428, + 227, + 489, + 274 + ], + "lines": [ + { + "bbox": [ + 428, + 227, + 489, + 274 + ], + "spans": [ + { + "bbox": [ + 428, + 227, + 489, + 274 + ], + "type": "image", + "image_path": "1444e61b2e50bc1aca4e84d1dd643d0d62a48755108d682b1ba98c4c523ca83d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 489, + 227, + 505, + 274 + ], + "blocks": [ + { + "bbox": [ + 489, + 227, + 505, + 274 + ], + "lines": [ + { + "bbox": [ + 489, + 227, + 505, + 274 + ], + "spans": [ + { + "bbox": [ + 489, + 227, + 505, + 274 + ], + "type": "image", + "image_path": "4c8ef65b30b3f52532f40ac9912db37642aff8c3b76758d8d83e2252600198e9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 505, + 227, + 550, + 274 + ], + "blocks": [ + { + "bbox": [ + 505, + 227, + 550, + 274 + ], + "lines": [ + { + "bbox": [ + 505, + 227, + 550, + 274 + ], + "spans": [ + { + "bbox": [ + 505, + 227, + 550, + 274 + ], + "type": "image", + "image_path": "d22317b8a865bf29be4dbbca2b37b4b8a173016a874a30a5ea680052aef4e630.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 59, + 285, + 120, + 339 + ], + "blocks": [ + { + "bbox": [ + 59, + 276, + 122, + 285 + ], + "lines": [ + { + "bbox": [ + 59, + 276, + 122, + 285 + ], + "spans": [ + { + "bbox": [ + 59, + 276, + 122, + 285 + ], + "type": "text", + "content": "Goal: prepare for th" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 285, + 120, + 339 + ], + "lines": [ + { + "bbox": [ + 59, + 285, + 120, + 339 + ], + "spans": [ + { + "bbox": [ + 59, + 285, + 120, + 339 + ], + "type": "image", + "image_path": "0d0261e1606befb36ed42293b996b46615ec2713130c7b513b6025371a6ec93e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 344, + 543, + 355 + ], + "lines": [ + { + "bbox": [ + 65, + 344, + 543, + 355 + ], + "spans": [ + { + "bbox": [ + 65, + 344, + 543, + 355 + ], + "type": "text", + "content": "Figure 4. Qualitative results of ACTOR on the challenging BEHAVIORHUB Dynamic subset. Please refer to §6.1 for more details." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 122, + 285, + 182, + 339 + ], + "blocks": [ + { + "bbox": [ + 157, + 276, + 299, + 285 + ], + "lines": [ + { + "bbox": [ + 157, + 276, + 299, + 285 + ], + "spans": [ + { + "bbox": [ + 157, + 276, + 299, + 285 + ], + "type": "text", + "content": "Language command: he/she is a neat person" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 122, + 285, + 182, + 339 + ], + "lines": [ + { + "bbox": [ + 122, + 285, + 182, + 339 + ], + "spans": [ + { + "bbox": [ + 122, + 285, + 182, + 339 + ], + "type": "image", + "image_path": "b5e1088c145076ebe587cfff6df9130bc3167e7583f1123b3bc43d50068f001c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 184, + 285, + 244, + 339 + ], + "blocks": [ + { + "bbox": [ + 184, + 285, + 244, + 339 + ], + "lines": [ + { + "bbox": [ + 184, + 285, + 244, + 339 + ], + "spans": [ + { + "bbox": [ + 184, + 285, + 244, + 339 + ], + "type": "image", + "image_path": "b120b4b17ddf01cdc28ec6a2c4810c165864a21c621c3c92c6773d53ed1d619f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 246, + 285, + 306, + 339 + ], + "blocks": [ + { + "bbox": [ + 246, + 285, + 306, + 339 + ], + "lines": [ + { + "bbox": [ + 246, + 285, + 306, + 339 + ], + "spans": [ + { + "bbox": [ + 246, + 285, + 306, + 339 + ], + "type": "image", + "image_path": "2092a6ccacd391c10a7b0f0876f4518fc5c932143cffd6b38b8ea1ae932f32f9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 285, + 367, + 339 + ], + "blocks": [ + { + "bbox": [ + 307, + 285, + 367, + 339 + ], + "lines": [ + { + "bbox": [ + 307, + 285, + 367, + 339 + ], + "spans": [ + { + "bbox": [ + 307, + 285, + 367, + 339 + ], + "type": "image", + "image_path": "651f81261a41e6b2a7676d43e5cd15204931fe2d298fd71c05a97589aa3126c4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 368, + 285, + 428, + 339 + ], + "blocks": [ + { + "bbox": [ + 368, + 285, + 428, + 339 + ], + "lines": [ + { + "bbox": [ + 368, + 285, + 428, + 339 + ], + "spans": [ + { + "bbox": [ + 368, + 285, + 428, + 339 + ], + "type": "image", + "image_path": "16e8a8fedeee2b016faa9777a90249094806def0e52f525502bde8d90ba102e4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 429, + 285, + 489, + 339 + ], + "blocks": [ + { + "bbox": [ + 429, + 285, + 489, + 339 + ], + "lines": [ + { + "bbox": [ + 429, + 285, + 489, + 339 + ], + "spans": [ + { + "bbox": [ + 429, + 285, + 489, + 339 + ], + "type": "image", + "image_path": "ad8b8ffeab641dedd45303717db750bc5ca68526f4fb725ee7afad5fe60320b0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 490, + 285, + 550, + 339 + ], + "blocks": [ + { + "bbox": [ + 490, + 285, + 550, + 339 + ], + "lines": [ + { + "bbox": [ + 490, + 285, + 550, + 339 + ], + "spans": [ + { + "bbox": [ + 490, + 285, + 550, + 339 + ], + "type": "image", + "image_path": "ede2439405ca5e53c44063a00858e96b2f467ba145f1938900533c9e1a49caed.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 371, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 418 + ], + "type": "text", + "content": "e.g., we designate the bathroom as occupied only after certain pre-request steps have been fulfilled; or incorporate language commands that specify agent characteristics, further contributing to a unique goal-plan preference." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 506 + ], + "type": "text", + "content": "Competitors. We benchmark two top-leading LLM-based models: LLMaP [33], a procedural planning model, and HuggingGPT [34], a general tool agent, to probe the human behavior simulation ability in existing techniques. LLMaP is designed to operate solely on textual inputs, and lacks the capability to perform behavior simulation. We assess its behavior planning ability only." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "Quantitative Evaluation. Table 1 summarizes the automatic evaluation results, from which we take three major observations: (i) ACTOR achieves the best results on both behavior planning and simulation and outperforms the strong baseline by nearly doubling the GSR, indicating the effectiveness of our value-driven behavior planning design. (ii) On the challenging Dynamic subset, both LLMaP and HuggingGPT experience a noticeable decline in terms of BERT-S, whereas our ACTOR performs much better. To illustrate, in the case shown in Fig. 4, both LLMaP and HuggingGPT directly ignore the occupied signal and give an irrational plan of heading directly to the bedroom, causing a failure. This observation confirms our claim that the ungrounded LLM falls short in terms of environment-aware behavior planning. (iii) Although our ACTOR shows more promising results, there still remains a significant gap compared to human performance. This highlights the need for" + } + ] + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 320, + 369, + 552, + 434 + ], + "blocks": [ + { + "bbox": [ + 56, + 183, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 183, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 553, + 194 + ], + "type": "text", + "content": "Table 1. Quantitative results on Main and Dynamic set of BEHAVIORHUB. " + }, + { + "bbox": [ + 56, + 183, + 553, + 194 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 56, + 183, + 553, + 194 + ], + "type": "text", + "content": " ’ indicates smaller values are better. See §6.1 for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 369, + 552, + 434 + ], + "lines": [ + { + "bbox": [ + 320, + 369, + 552, + 434 + ], + "spans": [ + { + "bbox": [ + 320, + 369, + 552, + 434 + ], + "type": "table", + "html": "
MethodComplete.Rational.Quality
Human4.024.85-
LLMaP [ICML2022]2.232.48-
HuggingGPT [NeuIPS23]2.712.893.18
ACTOR (Ours)3.053.473.75
", + "image_path": "c03f11badfbf91cf5e25421db772756952a35be8bf953b34c2af3e6bf8f330f1.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 437, + 550, + 448 + ], + "lines": [ + { + "bbox": [ + 318, + 437, + 550, + 448 + ], + "spans": [ + { + "bbox": [ + 318, + 437, + 550, + 448 + ], + "type": "text", + "content": "Table 2. Human-evaluated results (§6.1) on Dynamic subset." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 459, + 553, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 553, + 470 + ], + "type": "text", + "content": "developing more sophisticated behavior simulation models." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 472, + 554, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 699 + ], + "type": "text", + "content": "Human Evaluation. For a comprehensive evaluation, we engage five participants as human evaluators. They rate the model performance based on three aspects: (i) Completeness: assess whether the motion steps can successfully complete the target goal, capturing semantic completeness; (ii) Rationality: evaluate whether the sequence includes necessary steps in the correct sequential order to accomplish the target goal, capturing sequential order correctness; and (iii) Quality: reflect the naturalness and smoothness of the motion sequence, capturing motion quality. The results of the human ratings, based on a 5-point Likert scale, are reported in Table 2. The human subjective judgments generally align with the trends reflected by Table 2, confirming the reliability of our constructed automatic evaluation framework. Also, the results reaffirm that ACTOR yields plausible behavior simulation. However, it is worth noting that human-written plans are consistently preferred over our results, underscoring the challenging nature of our newly proposed BEHAVIORHUB benchmark." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 713 + ], + "type": "text", + "content": "Qualitative Analysis. Examples from ACTOR on BEHAV" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10759" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 245, + 134 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 245, + 134 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 245, + 134 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 245, + 134 + ], + "type": "table", + "html": "
MethodBERT-SGSRGSRPL
Baseline0.8110.1400.062
+ Active Search0.8370.2350.132
+ Hierarchical Prior0.8490.2610.154
+ Value Function0.8620.3060.212
", + "image_path": "ad52602b83013861bcbebb2b2ecf8fb6115a7efc6f3ed18ef565f8f9b19bba98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 250, + 82, + 396, + 134 + ], + "blocks": [ + { + "bbox": [ + 102, + 136, + 198, + 145 + ], + "lines": [ + { + "bbox": [ + 102, + 136, + 198, + 145 + ], + "spans": [ + { + "bbox": [ + 102, + 136, + 198, + 145 + ], + "type": "text", + "content": "(a) Key Component Analysis" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 250, + 82, + 396, + 134 + ], + "lines": [ + { + "bbox": [ + 250, + 82, + 396, + 134 + ], + "spans": [ + { + "bbox": [ + 250, + 82, + 396, + 134 + ], + "type": "table", + "html": "
AlgorithmBERT-SGSRGSRPL
Greedy0.8400.2440.151
Beam0.8530.2870.186
MCTS0.8620.3060.212
", + "image_path": "65f5c8b3311ef846ca4651f51b004d659c63e524a4273c4dc3031ddf0c701e26.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 403, + 82, + 550, + 134 + ], + "blocks": [ + { + "bbox": [ + 286, + 136, + 356, + 145 + ], + "lines": [ + { + "bbox": [ + 286, + 136, + 356, + 145 + ], + "spans": [ + { + "bbox": [ + 286, + 136, + 356, + 145 + ], + "type": "text", + "content": "(b) Search Algorithm" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 403, + 82, + 550, + 134 + ], + "lines": [ + { + "bbox": [ + 403, + 82, + 550, + 134 + ], + "spans": [ + { + "bbox": [ + 403, + 82, + 550, + 134 + ], + "type": "table", + "html": "
LLMBERT-SGSRGSRPL
Vicuna-7b0.8080.0920.063
GPT-3.50.8330.1760.116
GPT-40.8620.3060.212
", + "image_path": "02de33cd530696cb809b6b31be13d4edd016b3b078dae59967121a3871e376d9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 170, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 170, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 170, + 296, + 289 + ], + "type": "text", + "content": "IORHUB Dynamic are visualized in Fig. 4. In addition to simulating behavior to strategize and successfully achieve the desired goal, ACTOR is able to respond to environmental changes and adapt to language commands. For instance, to accomplish the goal of 'Sunday cleaning', while someone is still in bed, i.e., the bedroom is occupied, the agent prioritizes scheduling the kitchen first and then the bedroom, waiting for the person to wake up. In the second example, considering the truth that someone is a neat person, he/she engages in home cleaning before eating breakfast." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 299, + 190, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 190, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 190, + 312 + ], + "type": "text", + "content": "6.2. Diagnostic Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 317, + 295, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 295, + 353 + ], + "type": "text", + "content": "Ablative studies are conducted on BEHAVIORHUB Dynamic for indepth analyzing each component in ACTOR, using BERT-S, GSR, GSRPL (cf. §5.4) as evaluation metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 354, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 557 + ], + "type": "text", + "content": "Key Component Analysis. We first validate the importance of our proposed components by attaching them one at a time in Table 4b. The " + }, + { + "bbox": [ + 55, + 354, + 296, + 557 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 55, + 354, + 296, + 557 + ], + "type": "text", + "content": " row reports the result of a bare baseline model, which produces a global plan based on the given linguistic goal. Next, in the " + }, + { + "bbox": [ + 55, + 354, + 296, + 557 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 55, + 354, + 296, + 557 + ], + "type": "text", + "content": " row, we transition from one-pass planning to active tree search, resulting in improved performance and supporting our claim that active planning can mitigate the impact of environmental changes. Moreover, the " + }, + { + "bbox": [ + 55, + 354, + 296, + 557 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 55, + 354, + 296, + 557 + ], + "type": "text", + "content": " row gives the score when the hierarchical behavior structure prior is employed when spanning searching branches. As seen, this leads to moderate improvement by constraining the search space with interchangeable semantic units, highlighting its necessity in handling vast and complex human behavior space. Finally, as shown in the " + }, + { + "bbox": [ + 55, + 354, + 296, + 557 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 55, + 354, + 296, + 557 + ], + "type": "text", + "content": " row, incorporating value function significantly enhances the overall success rate and success rate over path length, aligning with the rational preference for the shortest path." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "content": "Search Algorithm. Table 3b reveals the impact of search algorithms (§4.2), i.e., greedy search, beam search, and MCTS, with window sizes of 5 for the later two algorithms. The default strategy, MCTS, shows optimal results, which aligns with the widely accepted understanding that MCTS is more effective when dealing with a large solution space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Modular Scalability. In Table 3c, we investigate the modular scalability by employing various LLM cores, including GPT-3.5, GPT-4 and open-source LLM Vicuna-7b [74]. As the LLMs' capabilities improve, ACTOR exhibits a continuous enhancement in its behavior simulation performance. This observation substantiates the notion that our ACTOR system possesses great potential for accommodating the de" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 321, + 167, + 555, + 209 + ], + "blocks": [ + { + "bbox": [ + 435, + 136, + 512, + 145 + ], + "lines": [ + { + "bbox": [ + 435, + 136, + 512, + 145 + ], + "spans": [ + { + "bbox": [ + 435, + 136, + 512, + 145 + ], + "type": "text", + "content": "(c) Modular Scalability" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 149, + 534, + 160 + ], + "lines": [ + { + "bbox": [ + 74, + 149, + 534, + 160 + ], + "spans": [ + { + "bbox": [ + 74, + 149, + 534, + 160 + ], + "type": "text", + "content": "Table 3. Ablative experiments on the Dynamic subset of our proposed BEHAVIORHUB. Please refer to §6.2 for more details." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 167, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 321, + 167, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 321, + 167, + 555, + 209 + ], + "type": "table", + "html": "
MethodMPJPE↓MPVPE↓
[75] [CVPR2021]242.50222.13
†[75] [CVPR2021]201.56189.21
", + "image_path": "2222aaf3221b6b377edd5f36a62b49c3e0f0328db7655c3233add87a165422dc.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 321, + 221, + 554, + 262 + ], + "blocks": [ + { + "bbox": [ + 376, + 210, + 493, + 219 + ], + "lines": [ + { + "bbox": [ + 376, + 210, + 493, + 219 + ], + "spans": [ + { + "bbox": [ + 376, + 210, + 493, + 219 + ], + "type": "text", + "content": "(a) Scene-aware Motion Generation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 221, + 554, + 262 + ], + "lines": [ + { + "bbox": [ + 321, + 221, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 321, + 221, + 554, + 262 + ], + "type": "table", + "html": "
MethodFID↓R Precision ↑
MDM [76] [ICLR2023]0.5440.611
†MDM [76] [ICLR2023]0.4710.705
", + "image_path": "5f0a29e28cbaf1b6e66ad255aca3f4b570585685ff4856a5b7e9dbfc5b347d54.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 361, + 264, + 509, + 274 + ], + "lines": [ + { + "bbox": [ + 361, + 264, + 509, + 274 + ], + "spans": [ + { + "bbox": [ + 361, + 264, + 509, + 274 + ], + "type": "text", + "content": "(b) Language-conditioned Motion Generation" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 278, + 553, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 278, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 553, + 300 + ], + "type": "text", + "content": "Table 4. Quantitative results (§6.3) on two downstream tasks. † indicates using BEHAVIORHUB for additional training." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 314, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 361 + ], + "type": "text", + "content": "velopment of more powerful LLM cores. Moreover, this shows that BEHAVIORHUB, in conjunction with the human behavior planning task, serves as a robust testbed for evaluating the planning capability of LLMs." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 375, + 501, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 501, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 501, + 387 + ], + "type": "text", + "content": "6.3. BEHAVIORHUB for Downsteam Task" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 395, + 555, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 551 + ], + "type": "text", + "content": "We further probe the effectiveness of BEHAVIORHUB by incorporating it as additional training data for two downstream tasks: scene-aware [75] and language-conditioned motion generation [76], which focus on generation without planning. We follow the official implementation of both methods and begin by pre-training the two models on BEHAVIORHUB. Subsequently, we fine-tune then evaluate these models on PROX [52] and HumanML3D [77], respectively. The evaluation results are reported in Table 4. As observed, BEHAVIORHUB significantly enhances model performance across all evaluation metrics for both tasks, highlighting the potential of our created dataset in facilitating broader applications within the realm of motion generation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 571, + 388, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 388, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 388, + 583 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "We present ACTOR, an LLM-powered agent towards realistic simulation of human behavior in 3D scenes. ACTOR integrates an LLM controller to perform complex behavior through planning on goal decomposition guided by hierarchical activity prior. The value-driven mechanism further deepens its understanding of environment. We demonstrate its potential in the simulated 3D indoor environment constructed using our newly created large-scale, scene-aware, behavior-rich dataset, BEHAVIORHUB. Evaluations suggest effective behavior planning and simulation of ACTOR." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10760" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 193 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Science and Technology Major Project (No. 2023ZD0121300), National Natural Science Foundation of China (No. 62372405), Fundamental Research Funds for the Central Universities (226-2025-00057), National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University (No. HMHAI-202403), China Postdoctoral Science Foundation (No. 2025T180421), and the Postdoctoral Fellowship Program of CPSF (No. GZC20251066)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 202, + 115, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 202, + 115, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 202, + 115, + 215 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 222, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 222, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 222, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 222, + 294, + 243 + ], + "type": "text", + "content": "[1] Denis Diderot. Diderot's early philosophical works. Number 4. Open Court, 1911. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 245, + 295, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 245, + 295, + 267 + ], + "spans": [ + { + "bbox": [ + 62, + 245, + 295, + 267 + ], + "type": "text", + "content": "[2] Mark O. Riedl. Interactive narrative: A novel application of artificial intelligence for computer games. In AAAI, 2012. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 268, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 268, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 268, + 295, + 300 + ], + "type": "text", + "content": "[3] Mark O Riedl. Human-centered artificial intelligence and machine learning. Human behavior and emerging technologies, 1(1), 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 301, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 301, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 301, + 294, + 334 + ], + "type": "text", + "content": "[4] Chen Liang, Zhicheng Shi, Wenguan Wang, and Yi Yang. Dual reciprocal learning of language-based human motion understanding and generation. In ICCV, 2025. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 335, + 294, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 294, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 294, + 368 + ], + "type": "text", + "content": "[5] Elias Matsas and George-Christopher Vosniakos. Design of a virtual reality training system for human-robot collaboration in manufacturing tasks. IJIDEM, 11, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 294, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 413 + ], + "type": "text", + "content": "[6] Joon Sung Park, Lindsay Popowski, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Social simulacra: Creating populated prototypes for social computing systems. In UIST, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 413, + 294, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 294, + 467 + ], + "type": "text", + "content": "[7] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 468, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 294, + 501 + ], + "type": "text", + "content": "[8] Xiaohan Wang, Wenguan Wang, Jiayi Shao, and Yi Yang. Lana: A language-capable navigator for instruction following and generation. In CVPR, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 502, + 294, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 534 + ], + "type": "text", + "content": "[9] Hanqing Wang, Wei Liang, Luc V Gool, and Wenguan Wang. Towards versatile embodied navigation. NeurIPS, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 536, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 578 + ], + "type": "text", + "content": "[10] Zhichao Zhai, Guikun Chen, Wenguan Wang, Dong Zheng, and Jun Xiao. Taga: Self-supervised learning for template-free animatable gaussian articulated model. In CVPR, 2025. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 580, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 294, + 624 + ], + "type": "text", + "content": "[11] Xavier Puig, Kevin Kyunghwan Ra, Marko Boben, Jiaman Li, Tingwu Wang, Sanja Fidler, and Antonio Torralba. Virtualhome: Simulating household activities via programs. In CVPR, 2018. 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 669 + ], + "type": "text", + "content": "[12] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[13] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[14] Naureen Mahmood, Nima Ghorbani, Nikolaus F. Troje, Gerard Pons-Moll, and Michael J. Black. Amass: Archive of motion capture as surface shapes. In ICCV, 2019. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 150 + ], + "type": "text", + "content": "[15] Hongwei Yi, Chun-Hao P Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J Black. Human-aware object placement for visual environment reconstruction. In CVPR, 2022. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 151, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 172 + ], + "type": "text", + "content": "[16] Joseph Bates. The role of emotion in believable agents. Communications of the ACM, 37(7), 1994. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 173, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 194 + ], + "type": "text", + "content": "[17] Robert Zubek. Towards implementation of social interaction. In AAAI, 2002. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 195, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 217 + ], + "type": "text", + "content": "[18] Ruth Aylett. Narrative in virtual environments—towards emergent narrative. In AAAI, 1999." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "type": "text", + "content": "[19] Michael Brenner. Creating dynamic story plots with continual multiagent planning. In AAAI, 2010. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 239, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 294 + ], + "type": "text", + "content": "[20] Ho Chit Siu, Jaime Pena, Edenna Chen, Yutai Zhou, Victor Lopez, Kyle Palko, Kimberlee Chang, and Ross Allen. Evaluation of human-ai teams for learned and rule-based agents in hanabi. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors, NeurIPS, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 295, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 553, + 316 + ], + "type": "text", + "content": "[21] Michele Colledanchise and Petter Ögren. Behavior trees in robotics and AI: An introduction. CRC Press, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 317, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 317, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 553, + 360 + ], + "type": "text", + "content": "[22] Ilshat Umarov, Mikhail Mozgovoy, and Patrick C. Rogers. Believable and effective ai agents in virtual worlds: Current state and future perspectives. International Journal of Gaming and Computer-Mediated Simulations, 4(2), 2012. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 361, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 554, + 459 + ], + "type": "text", + "content": "[23] Christopher Berner, Greg Brockman, Brooke Chan, Vicki Cheung, Przemyslaw Debiak, Christy Dennison, David Farhi, Quirin Fischer, Shariq Hashme, Chris Hesse, Rafal Jozefowicz, Scott Gray, Catherine Olsson, Jakub Pachocki, Michael Petrov, Henrique P. d.O. Pinto, Jonathan Raiman, Tim Salimans, Jeremy Schlatter, Jonas Schneider, Szymon Sidor, Ilya Sutskever, Jie Tang, Filip Wolski, and Susan Zhang. Dota 2 with large scale deep reinforcement learning. arXiv preprint arXiv:1912.06680, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "text", + "content": "[24] Oriol Vinyals, Igor Babuschkin, Wojciech M. Czarnecki, and et al. Grandmaster level in starcraft ii using multi-agent reinforcement learning. Nature, 575, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 525 + ], + "type": "text", + "content": "[25] Matthew Hausknecht, Prithviraj Ammanabrolu, Marc-Alexandre Cote, and Xinyu Yuan. Interactive fiction games: A colossal adventure. In AAAI, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 569 + ], + "type": "text", + "content": "[26] Shohei Miyashita, Xinyu Lian, Xiao Zeng, Takashi Matsubara, and Kuniaki Uehara. Developing game ai agent behaving like human by mixing reinforcement learning and supervised learning. In SNPD, 2017." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 570, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 570, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 570, + 553, + 603 + ], + "type": "text", + "content": "[27] Chen Tessler, Shahar Givony, Tom Zahavy, Daniel Mankowitz, and Shie Mannor. A deep hierarchical approach to lifelong learning in apache. In AAAI, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "type": "text", + "content": "[28] John Laird. It knows what you're going to do: Adding anticipation to a quakebot. In Proceedings of the 2001 Workshop on Intelligent Cinematography and Editing, 2001. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 669 + ], + "type": "text", + "content": "[29] Dong kyu Choi, Tolga Konik, Negin Nejati, Chunki Park, and Pat Langley. A believable agent for first-person shooter games. In AAAI, 2021. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[30] Pat Langley, Dongkyu Choi, and Seth Rogers. Interleaving learning, problem solving, and execution in the icarus architecture. Technical report, Stanford University, Center for the Study of Language and Information, 2005. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10761" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[31] Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In UIST, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[32] Yujie Lu, Weixi Feng, Wanrong Zhu, Wenda Xu, Xin Eric Wang, Miguel Eckstein, and William Yang Wang. Neurosymbolic procedural planning with commonsense prompting. In ICLR, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[33] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In ICML, 2022. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "type": "text", + "content": "[34] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang. Huggingppt: Solving ai tasks with chatgpt and its friends in huggingface. In NeurIPS, 2023. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "text", + "content": "[35] Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models. In NeurIPS, 2023." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[36] Zongxin Yang, Guikun Chen, Xiaodi Li, Wenguan Wang, and Yi Yang. Doraemongpt: Toward understanding dynamic scenes with large language models (exemplified as a video agent). In ICML, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 376 + ], + "type": "text", + "content": "[37] Shima Imani, Liang Du, and Harsh Shrivastava. Math-prompter: Mathematical reasoning using large language models. In ICLR Workshops, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 431 + ], + "type": "text", + "content": "[38] Zhangir Azerbayev, Hailey Schoelkopf, Keiran Paster, Marco Dos Santos, Stephen McAleer, Albert Q. Jiang, Jia Deng, Stella Biderman, and Sean Welleck. Llemma: An open language model for mathematics. arXiv preprint arXiv:2310.06786, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "text", + "content": "[39] Xiang Lorraine Li, Adhiguna Kuncoro, Jordan Hoffmann, Cyprien de Masson d'Autume, Phil Blunsom, and Aida Nematzadeh. A systematic investigation of commonsense knowledge in large language models. In EMNLP, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 521 + ], + "type": "text", + "content": "[40] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In ICRA, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[41] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 634 + ], + "type": "text", + "content": "[42] Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, et al. Palm: Scaling language modeling with pathways. arXiv preprint arXiv:2204.02311, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[43] Anthony Brohan, Yevgen Chebotar, Chelsea Finn, Karol Hausman, Alexander Herzog, Daniel Ho, Julian Ibarz, Alex Irpan, Eric Jang, Ryan Julian, et al. Do as i can, not as i say: Grounding language in robotic affordances. In CoRL, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[44] Wenlong Huang, Fei Xia, Dhruv Shah, Danny Driess, Andy Zeng, Yao Lu, Pete Florence, Igor Mordatch, Sergey Levine," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 106 + ], + "type": "text", + "content": "Karol Hausman, et al. Grounded decoding: Guiding text generation with grounded models for robot control. arXiv preprint arXiv:2303.00855, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 138 + ], + "type": "text", + "content": "[45] Belinda Z Li, Maxwell Nye, and Jacob Andreas. Implicit representations of meaning in neural language models. In ACL, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 171 + ], + "type": "text", + "content": "[46] Adam Roberts, Colin Raffel, and Noam Shazeer. How much knowledge can you pack into the parameters of a language model? In EMNLP, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 205 + ], + "type": "text", + "content": "[47] Yunqiu Xu, Linchao Zhu, and Yi Yang. Mc-bench: A benchmark for multi-context visual grounding in the era of mllms. In ICCV, 2025. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 239 + ], + "type": "text", + "content": "[48] Vadim Borisov, Kathrin Sessler, Tobias Leemann, Martin Pawelczyk, and Gjergji Kasneci. Language models are realistic tabular data generators. In ICLR, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 240, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 271 + ], + "type": "text", + "content": "[49] Yew Ken Chia, Lidong Bing, Soujanya Poria, and Luo Si. Relationprompt: Leveraging prompts to generate synthetic data for zero-shot relation triplet extraction. In ACL, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 273, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 294 + ], + "type": "text", + "content": "[50] Timo Schick and Hinrich Schütze. Generating datasets with pretrained language models. In EMNLP, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 295, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 553, + 338 + ], + "type": "text", + "content": "[51] Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language model with self generated instructions. In ACL, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 339, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 553, + 371 + ], + "type": "text", + "content": "[52] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In ICCV, 2019. 3, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 373, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 404 + ], + "type": "text", + "content": "[53] Hongming Zhang, Xin Liu, Haojie Pan, Yangqiu Song, and Cane Wing-Ki Leung. Aser: A large-scale eventuality knowledge graph. In WWW, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 406, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 553, + 459 + ], + "type": "text", + "content": "[54] Peter West, Chandra Bhagavatula, Jack Hessel, Jena Hwang, Liwei Jiang, Ronan Le Bras, Ximing Lu, Sean Welleck, and Yejin Choi. Symbolic knowledge distillation: from general language models to commonsense models. In NAACL, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 460, + 553, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 493 + ], + "type": "text", + "content": "[55] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 494, + 553, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 553, + 537 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 553, + 537 + ], + "type": "text", + "content": "[56] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In CVPR, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 538, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 538, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 538, + 553, + 581 + ], + "type": "text", + "content": "[57] Yudi Dai, Yitai Lin, Chenglu Wen, Siqi Shen, Lan Xu, Jingyi Yu, Yuexin Ma, and Cheng Wang. Hsc4d: Human-centered 4d scene capture in large-scale indoor-outdoor space using wearable imus and lidar. In CVPR, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[58] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, et al. Habitat: A platform for embodied ai research. In ICCV, 2019. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 626, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 626, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 626, + 553, + 713 + ], + "type": "text", + "content": "[59] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Xuan Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10762" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[60] Korrawe Karunratanakul, Konpat Preechakul, Supasorn Suwajanakorn, and Siyu Tang. Guided motion diffusion for controllable human motion synthesis. In ICCV, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[61] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In ECCV, 2020. 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 294, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 175 + ], + "type": "text", + "content": "[62] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In CVPR, 2022. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 177, + 294, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 294, + 209 + ], + "type": "text", + "content": "[63] Thomas A Feo and Mauricio GC Resende. Greedy randomized adaptive search procedures. Journal of global optimization, 6, 1995. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 212, + 294, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 212, + 294, + 233 + ], + "spans": [ + { + "bbox": [ + 56, + 212, + 294, + 233 + ], + "type": "text", + "content": "[64] Markus Freitag and Yaser Al-Onaizan. Beam search strategies for neural machine translation. In ACL, 2017. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 236, + 294, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 236, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 294, + 257 + ], + "type": "text", + "content": "[65] Rémi Coulom. Efficient selectivity and backup operators in monte-carlo tree search. In CG, 2006. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 259, + 294, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 259, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 56, + 259, + 294, + 291 + ], + "type": "text", + "content": "[66] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. In ICLR, 2019. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 293, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 336 + ], + "type": "text", + "content": "[67] Abhinanda R Punnakkal, Arjun Chandrasekaran, Nikos Athanasiou, Alejandra Quiros-Ramirez, and Michael J Black. Babel: Bodies, action and behavior with english labels. In CVPR, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 339, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 339, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 339, + 294, + 415 + ], + "type": "text", + "content": "[68] Santhosh K. Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westry, Angel Xuan Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. In NeurIPS Datasets and Benchmarks Track, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 418, + 294, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 294, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 294, + 472 + ], + "type": "text", + "content": "[69] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 475, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 294, + 517 + ], + "type": "text", + "content": "[70] Yinglin Duan, Tianyang Shi, Zhengxia Zou, Yenan Lin, Zhehui Qian, Bohan Zhang, and Yi Yuan. Single-shot motion completion with transformer. arXiv preprint arXiv:2103.00776, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 520, + 294, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 552 + ], + "type": "text", + "content": "[71] Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. Human motion diffusion as a generative prior. In ICLR, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 555, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 586 + ], + "type": "text", + "content": "[72] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, 2002. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 589, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 294, + 621 + ], + "type": "text", + "content": "[73] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In SIGGRAPH, 2023. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[74] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhang-hao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yong-hao Zhuang, Joseph E Gonzalez, et al. Vicuna: An opensource chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": " * chatgpt quality. https://vicuna.lmsys.org, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 712 + ], + "type": "text", + "content": "[75] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In CVPR, 2021. 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[76] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In ICLR, 2023. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[77] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In CVPR, 2022. 8" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10763" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/4564e033-7cc9-41ba-b940-20704a40b38a_content_list.json b/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/4564e033-7cc9-41ba-b940-20704a40b38a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3023a7cace7f20b45ef83fcdf7660ab4162e8654 --- /dev/null +++ b/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/4564e033-7cc9-41ba-b940-20704a40b38a_content_list.json @@ -0,0 +1,1750 @@ +[ + { + "type": "text", + "text": "Towards Immersive Human-X Interaction: A Real-Time Framework for Physically Plausible Motion Synthesis", + "text_level": 1, + "bbox": [ + 133, + 128, + 864, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kaiyang Ji $^{1}$ Ye Shi $^{1,2,*}$ Zichen Jin $^{1}$ Kangyi Chen $^{1}$ Lan Xu $^{1,2}$ Yuexin Ma $^{1,2}$ Jingyi Yu $^{1,2}$ Jingya Wang $^{1,2,*}$", + "bbox": [ + 184, + 200, + 805, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ ShanghaiTech University", + "bbox": [ + 395, + 239, + 602, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Shanghai Engineering Research Center of Intelligent Vision and Imaging", + "bbox": [ + 205, + 256, + 790, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jiky2024,shiye,jinzch12023,v-chenky,xulan1,mayuexin,yujingyi,wangjingya}@shanghaiitech.edu.cn", + "bbox": [ + 130, + 277, + 861, + 292 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f77c7108da04d1983bd50f0fbc1b3c5ee4539371662cb0c5a1ef5c6c0123245a.jpg", + "image_caption": [ + "Figure 1. We propose Human-X, the first framework designed to enable latency-free interaction between humans and diverse entities, including human-athlete, human-humanoid, and human-robot interaction." + ], + "image_footnote": [], + "bbox": [ + 99, + 309, + 893, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 511, + 326, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Real-time synthesis of physically plausible human interactions remains a critical challenge for immersive VR/AR systems and humanoid robotics. While existing methods demonstrate progress in kinematic motion generation, they often fail to address the fundamental tension between real-time responsiveness, physical feasibility, and safety requirements in dynamic human-machine interactions. We introduce Human-X, a novel framework designed to enable immersive and physically plausible human interactions across diverse entities, including human-ayar, human-humanoid, and human-robot systems. Unlike existing approaches that focus on post-hoc alignment or simplified physics, our method jointly predicts actions and reactions in real-time using an auto-regressive reaction diffusion planner, ensuring seamless synchronization and context-aware responses. To enhance physical realism and safety, we integrate an actor-aware motion tracking policy trained with reinforcement learning, which dynamically adapts to interaction partners' movements while avoiding artifacts like foot sliding and penetration. Extensive experiments on the Inter-X and InterHuman datasets demonstrate significant improvements in motion quality, interaction continuity, and physical", + "bbox": [ + 88, + 542, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plausibility over state-of-the-art methods. Our framework is validated in real-world applications, including virtual reality interface for human-robot interaction, showcasing its potential for advancing human-robot collaboration. Project page: https://humanx-interaction.github.io/", + "bbox": [ + 511, + 512, + 908, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 632, + 645, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human-human interaction synthesis has long been a focal point in the field of motion generation, with widespread applications in social dynamics, gaming, and animation [26, 54]. With the rapid development of virtual reality (VR) and augmented reality (AR), these technologies now empower the creation of deeply immersive and contextually rich interactions between humans and digital avatars. Particularly, the emergence of humanoid robots further bridges the gap between virtual and physical realms, enabling robots to engage with humans in daily life through adaptive and socially aware behaviors. However, achieving real-time prediction of diverse and safety-guaranteed humanoid reactions, an essential capability for applications ranging from daily companionship to assistive care, remains a critical, unsolved challenge.", + "bbox": [ + 509, + 657, + 906, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current approaches to human interaction modeling rely", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 109, + 886, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10173", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/d90d1e1a3d1a5862280c3c87ee3d95b1b3ff3632e0d5f547d86c7b432b7a9f49.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methodsphysics-basedimmersive inferencelatency-freereal-time modelgeneration abilitystrong contacttext-control
InterGen [32]✓✓✓
ReGenNet [66]✓✓✓✓
SoLaMi [26]✓✓✓
PhysReaction [36]✓✓✓✓
Think-Then-React [55]✓✓✓✓
Human-X✓✓✓✓✓✓✓
", + "bbox": [ + 96, + 88, + 906, + 186 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparative analysis of Human-X versus existing human reaction synthesis approaches.", + "bbox": [ + 207, + 198, + 785, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "heavily on offline training data or simplified physics assumptions, leading to limited adaptability in dynamic scenarios. While data-driven methods such as [23, 32, 65] have shown promise in synthesizing kinematic motions interaction, they often fail to account for the intricate physical constraints and safety requirements inherent in human-humanoid cohabitation. For instance, abrupt or unstable motions generated through these techniques pose risks in shared environments. This gap underscores the need for a unified interaction framework that integrates real-time reactivity, physical realism, and safety assurance.", + "bbox": [ + 88, + 239, + 480, + 405 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we explore Human-X Interaction (Fig.1), a systematic study of immersive interactions across three domains: 1) Human-Avatar Interaction using kinematic-driven motion in VR/AR; 2) Human-Humanoid Interaction with physics-based motion synthesis; 3) Human-Robot Interaction ensuring safe physical contact. Unlike existing physics-aware methods limited to replicating training data, our framework balances natural movement with physical feasibility while maintaining real-time performance. This integration enables safer and more natural human-machine collaboration across applications from entertainment to assistive robotics. Achieving real-time synchronization between interacting partners faces major technical challenges. Conventional methods rely exclusively on past actor motions to generate reactions, leading to temporal mismatches. Furthermore, ensuring physical realism is more challenging, as existing solutions focus mainly on basic collision avoidance and neglect dynamic interaction forces.", + "bbox": [ + 91, + 409, + 482, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these limitations, we present Human-X, an Auto-Regressive Action-Reaction Diffusion Planner for real-time interaction. Our framework jointly models both partners' motion histories to predict subsequent actions and reactions. We introduce multiple contact-aware losses to train realistic interactions, and combine the diffusion model with an actor-aware tracking policy to ensure physical feasibility in humanoid/robot scenarios. In summary, our contributions can be summarized as follows:", + "bbox": [ + 89, + 685, + 482, + 819 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The first auto-regressive diffusion model that jointly predicts actions and reactions using both partners' motion history with comprehensive contract losses to ensure real-time, immersive and realistic motion synthesis.", + "- An actor-aware reaction policy trained with reinforce" + ], + "bbox": [ + 89, + 825, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ment learning that dynamically adapts to interaction partners' motions, preventing common artifacts like foot sliding while maintaining physical feasibility.", + "bbox": [ + 524, + 239, + 903, + 284 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- A complete real-time VR system demonstrating practical effectiveness. Experiments on Inter-X and InterHuman datasets show significant improvements in motion quality, interaction continuity, and physical accuracy compared to existing methods.", + "bbox": [ + 511, + 285, + 903, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 371, + 663, + 386 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Human Motion Generation", + "text_level": 1, + "bbox": [ + 511, + 396, + 759, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human-centric motion generation has seen remarkable advancements in recent years, driven by breakthroughs in large language models [1, 25] and generative frameworks [13, 20, 28, 67, 75]. Notably, the introduction of motion diffusion models [58, 71] has significantly elevated the performance of single-person motion generation. Building on these developments, researchers have increasingly focused on enriching interactions with the environment by incorporating diverse modalities, such as text [25, 62, 63, 70], audio [2, 61], and scene information [12, 22, 29, 57]. For instance, recent works [9, 60, 62, 63, 72] have leveraged text-based guidance to facilitate Human-Object Interactions (HOI), while MotionGPT [25] treats motion as a form of language, enabling seamless bidirectional translation between text and motion. Despite these advancements, existing approaches face a critical limitation: their model architectures are inherently ill-suited for real-time reaction synthesis.", + "bbox": [ + 511, + 417, + 906, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Human-Human Interaction", + "text_level": 1, + "bbox": [ + 511, + 696, + 761, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Research on human-human interaction synthesis has evolved around three primary directions: (1) Interactive Behavior Modeling leverages actor-reactor role decomposition to capture interaction dynamics. For example, [56] identifies role-specific patterns, while [32, 64] use transformer architectures to model multi-agent dependencies. (2) Multimodal Interaction integrates text, audio, or scene context to enhance realism [31, 52], but often overlook dynamic physical constraints critical for real-world deployment. (3) Reactive Motion Generation employs sequence models [35] and diffusion models [27, 66] to synthesize temporally coherent reactions. These methods, however, struggle with physical", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10174", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9d3fce72a20fd83d348ef6c4fd7a584078f361a8fc1f08ae40900361a5995903.jpg", + "image_caption": [ + "Figure 2. Overview of our immersive real-time interaction synthesis pipeline: (a) Actor Motion Capture: A human actor's movements are recorded at 30 fps by an RGB-D camera and translated into 3D poses, which are then retargeted to a humanoid character. (b) Realistic Reactor Motion Generation: An auto-regressive diffusion model, guided by optional text prompts (e.g., \"Dancing is what to do\"), generates plausible reaction motions. These motions are tracked by an actor-aware controller, which uses proprioception signals to ensure realistic, synchronized interactions. (c) Real-time VR Interface: The generated and tracked motions are rendered in simulator, providing both a third-person view and a binocular VR view." + ], + "image_footnote": [], + "bbox": [ + 101, + 93, + 908, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "plausibility and temporal consistency in long-horizon interactions. Existing Reactive Generation methods face a fundamental tension: Human interaction models achieve reactivity but lack physical grounding, while physics-based animations enforce plausibility but are agnostic to interaction dynamics. While PhysReaction [36] introduced the first physics-based approach for reaction synthesis using imitation learning, it was limited to mimicking known distributions, lacking generalization ability, diversity, controllability, and the capacity to handle unseen tasks. Our work bridges this gap by unifying interaction-aware reactivity with physically constrained control.", + "bbox": [ + 88, + 476, + 485, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Physics-based Motion Synthesis", + "text_level": 1, + "bbox": [ + 89, + 667, + 372, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to kinematics-based methods, physics-based motion synthesis has emerged as a widely adopted approach for generating realistic and physically plausible character movements by integrating physical principles into the animation process. Early work, such as DeepMimic [45], leveraged deep reinforcement learning to enable characters to imitate diverse motion clips within a physics-based framework. Subsequent advancements [18, 46] utilized generative adversarial networks to enhance stylized motion realism. Progress continued with PHC [38] and PULSE [39], focusing on improving the diversity of generated motions. A significant milestone was achieved with PhysDiff [69], which introduced physics-guided diffusion models to mitigate artifacts such as floating and foot sliding. Despite", + "bbox": [ + 89, + 688, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "these advancements, the physical plausibility of human-human interactions remains largely unexplored.", + "bbox": [ + 511, + 476, + 906, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 520, + 604, + 535 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. System Overview", + "text_level": 1, + "bbox": [ + 511, + 544, + 683, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present a detailed overview of our immersive interaction synthesis system. To facilitate realistic and responsive interactions between two agents in immersive environments, our system takes online actor motion capture data as input, processes it through an advanced motion generation pipeline, and outputs continuous, reactive motion in real-time, as shown in Fig.2. The first key component of our system, detailed in Sec.3.2, is the real-time capture of the actor's pose and the reconstruction of its state in physics simulation. Subsequently, the Auto-regressive Interaction Diffusion Model, as described in Sec.3.3, is employed to generate the reaction of one agent in response to the actor context. Next, we introduce the Reaction Motion Tracker in Sec.3.4, which ensures that the motion of the reactor adheres to physical constraints for realism and natural behavior. Finally, we explain how VR Interface works in Sec.4.1, which enables immersive, real-time interaction with the generated motions.", + "bbox": [ + 511, + 566, + 908, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Human Pose Estimation and Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 848, + 890, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We firstly utilize a real-time RGB-D camera to capture images. Then, HybrIK [30] is employed for latency-free re", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10175", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "construction of the human pose. However, due to the inaccuracies in the depth information provided by HybrIK, the depth map is directly obtained from the RGB-D camera instead. Following [38, 40], we bridge SMPL's kinematic skeleton to physics simulation via applying a simple retargeting method. We utilize Isaac Gym [41] as our physics simulation platform to create a third-person perspective for the interaction between the actor and reactor. Besides, the image reading process, motion generation of the reactor, and the physical simulation within Isaac Gym are handled by separate processes. ROS2 is used as the communication framework to facilitate real-time data transmission.", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Online Human Action-Reaction Synthesis", + "text_level": 1, + "bbox": [ + 89, + 282, + 447, + 299 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of online action-reaction synthesis is to generate realistic and continuous future $k$ frames of reactor motions $\\mathbf{x}^i \\coloneqq \\{x^{n:n + k - 1}\\}$ given $h$ history frames of interaction contexts $\\{\\mathbf{x}^{i - 1},\\mathbf{y}^{i - 1}\\} \\coloneqq \\{x^{n - h:n - 1},y^{n - h:n - 1}\\}$ at current frame number $n$ and current window index $i$ . During each window prediction, an optional text prompt $c$ can be added to guide the reaction generation process.", + "bbox": [ + 89, + 306, + 483, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Reactor-centric Interaction Modeling", + "text_level": 1, + "bbox": [ + 89, + 421, + 398, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reactor-centric Canonicalization. For individual motion synthesis, the initial frame of each sequence is typically aligned at the root to improve the generalizability of the model [14, 59]. However, using the same normalization for human-human interaction data can eliminate key details about the mutual spatial relationship. To retain relative positioning without losing generalization ability, we leverage a reactor-centric canonicalization that the pose of reactor is realigned at the root and oriented toward the forward direction, while the pose of actor is adjusted by the same transformation. This method maintains the essential spatial context between the two, enabling the model to accurately learn their interdependent movement dynamics.", + "bbox": [ + 89, + 440, + 483, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reactor-centric Interaction Representation. We model each frame of interaction as a hybrid representation of both actor and reactor motion. For reaction representation, we follow previous works [14, 58] and use the HumanML3D representation. The reactor motion $x$ in frame $n$ is defined by $x^{n} = (r^{x},\\dot{r}^{x},p^{x},\\dot{p}^{x},\\theta^{x},f^{x})$ , where $r^x\\in \\mathbb{R}$ denotes the height of the roots, $\\dot{r}^x\\in \\mathbb{R}^3$ denotes the angular velocity of the roots along the Y axis and the linear velocity of the roots in the XZ plane, $p^x\\in \\mathbb{R}^{(J - 1)\\times 3}$ are the local joint positions, $\\dot{p}^x\\in \\mathbb{R}^{J\\times 3}$ denotes the temporal difference of the local joint positions, $\\theta^x\\in \\mathbb{R}^{(J - 1)\\times 6}$ is the 6D representation [74] of the joint rotations and $f^x\\in \\mathbb{R}^4$ are binary labels for the detection of contact. For actor motion captured from pose estimation, we define the actor $y$ in frame $n$ by $y^n = (r^{x\\to y},\\dot{r}^{x\\to y},p^{x\\to y},\\dot{p}^{x\\to y},f^y)$ where all the features are in the reactor's reference frame, making it a pure", + "bbox": [ + 89, + 659, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "relative representation consisting of relative root transition, relative root rotation, relative root linear velocity, relative joint positions, relative joint velocity and feet contact labels. To emphasize actor-reactor contacts, we introduce the binary interaction field $\\mathcal{I}^n\\in \\mathbb{R}^{\\bar{J}\\times \\bar{J}}$ obtained by thresholding the relative distance between the selected $\\bar{J}$ joints of each agent in a contact map. For body model selection, we follow the skeleton of SMPL [37] with $J = 22$ joints. The total reactor-centric interaction $\\mathbf{z}$ at frame $n$ representation can be formed as,", + "bbox": [ + 511, + 90, + 906, + 242 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz ^ {n} = \\left(x ^ {n}, y ^ {n}, \\mathcal {I} ^ {n}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 252, + 906, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $x^n$ denotes the reactor motion, $y^n$ denotes the relative actor motion and $\\mathcal{I}^n$ is the joint contact features. Thus, we can model the interaction sequence as $\\mathbf{z}^i = \\{z^{n:n + k - 1}\\}$ given $k$ frames and window index $i$ .", + "bbox": [ + 511, + 279, + 906, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Auto-regressive Interaction Diffusion Planner", + "text_level": 1, + "bbox": [ + 511, + 347, + 877, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the interaction motion representation provided in Sec.3.3.1, we transfer raw motion data into a canonical space suitable for generative learning with denoising networks. For probabilistic generator, we follow the DDPM [20] framework by incrementally adding Gaussian noise to the clean data sample $z_0$ over $T$ steps:", + "bbox": [ + 511, + 366, + 905, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {\\alpha_ {t}} \\mathbf {z} _ {t - 1}, (1 - \\alpha_ {t}) \\mathbf {I}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 467, + 906, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{t} \\in (0,1)$ are constant parameters for variance scheduling. Given the noised data $\\mathbf{z}_t^i$ at the $i$ -th window clip, the diffusion step $t$ , the prefix interaction window $\\mathbf{z}^{i-1} = z^{n-h:n-1}$ , and the optional text prompt $c$ , we train a transformer-based denoiser $\\mathcal{G}$ to predict the cleaned interaction primitive itself instead of predicting the noise [58]:", + "bbox": [ + 511, + 494, + 905, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {z}} _ {0} ^ {i} = \\mathcal {G} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, c\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 598, + 903, + 616 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the text prompt $c$ is randomly masked by a rate of 0.15 during training to enable unconditioned generation and classifier-free guidance [19]. The predicted motion $\\hat{\\mathbf{z}}_0^i$ is then inserted to the interaction history for next interaction window extraction after the diffusion sampling loop [20, 53].", + "bbox": [ + 511, + 621, + 905, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Losses The denoised motion $\\hat{\\mathbf{z}}_0^i$ is supervised by a simple objective in DDPM [20]:", + "bbox": [ + 511, + 729, + 905, + 761 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s i m p l e}} = \\mathbb {E} _ {\\mathbf {z} _ {0} ^ {i} \\sim q \\left(\\mathbf {z} _ {0} ^ {i}\\right), t \\sim [ 1, T ]} \\left[ \\| \\mathbf {z} _ {0} ^ {i} - \\hat {\\mathbf {z}} _ {0} ^ {i} \\| _ {2} ^ {2} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 770, + 903, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We extend the training objective with several auxiliary losses to enhance interaction quality and realism. To optimize foot contact quality, we add a foot contact loss on both agents:", + "bbox": [ + 511, + 797, + 905, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f o o t}} = \\sum_ {j \\in \\{x ^ {\\text {f e e t}}, y ^ {\\text {f e e t}} \\}} \\| \\dot {p} ^ {j} \\left(\\hat {\\mathbf {z}} _ {0} ^ {i}\\right) \\odot f ^ {j} \\left(\\hat {\\mathbf {z}} _ {0} ^ {i}\\right) \\| _ {2} ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 869, + 903, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10176", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\dot{p}^j (\\hat{\\mathbf{z}}_0^i)$ is the corresponding foot joints velocity, $f^{j}(\\hat{\\mathbf{z}}_{0}^{i})$ are foot contact label described in Sec 3.3.1 and $\\odot$ denotes for Hadamard product. Note that $\\dot{p}^j (\\hat{\\mathbf{z}}_0^i)$ is defined by the relative local velocity in the reactor frame, and we transform it to the global velocity in loss calculation. For interaction refinement, we design an interaction loss on contact joints distance:", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i n t e r} = \\sum_ {x \\in \\bar {J} ^ {x}, y \\in \\bar {J} ^ {y}} \\| (p ^ {x} (\\hat {\\mathbf {z}} _ {0} ^ {i}) - p ^ {y} (\\hat {\\mathbf {z}} _ {0} ^ {i})) \\odot \\mathcal {I} ^ {x, y} (\\hat {\\mathbf {z}} _ {0} ^ {i}) \\| _ {2} ^ {2}, (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 208, + 483, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{I}^{x,y}(\\hat{\\mathbf{z}}_0^i)$ are binary body interaction fields and $p^x (\\hat{\\mathbf{z}}_0^i), p^y (\\hat{\\mathbf{z}}_0^i)$ are contact joints positions in the same root frame to enhance strong interaction. To improve the continuity of auto-regressive generation, we finally add a prefix loss $\\mathcal{L}_{prefix}$ to evaluate the difference between the end of prefix motion and the beginning of the generated motion. The full loss applied to the model is then:", + "bbox": [ + 89, + 253, + 483, + 361 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\text {s i m p l e}} + \\lambda_ {\\text {f o o t}} \\mathcal {L} _ {\\text {f o o t}} + \\lambda_ {\\text {i n t e r}} \\mathcal {L} _ {\\text {i n t e r}} \\tag {7} \\\\ + \\lambda_ {p r e f i x} \\mathcal {L} _ {p r e f i x}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 372, + 483, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{foot}$ , $\\lambda_{inter}$ and $\\lambda_{prefix}$ are hyper-parameters to balance the loss terms from different scaling aspects.", + "bbox": [ + 89, + 417, + 483, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training During training, we utilize the scheduled training strategy [3, 48, 68, 73] to enhance auto-regressive generation stability and text prompt controllability. The motion is cropped to sequences of $N$ consecutive interaction windows with a random offset from the dataset. During training, $\\mathbf{z}_0^i$ is noised to $\\mathbf{z}_t^i$ at diffusion timestep $t$ , and interaction history $\\mathbf{z}^{i-1}$ is divide into 3 stages with the sample probability $p$ : ground truth for supervised learning $(p = 1)$ , random sampling for scheduled training (decaying $p$ ), and auto-regressive prediction $(p = 0)$ . Details in Appendix B.1.", + "bbox": [ + 89, + 465, + 483, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference After training a denoise network $\\mathcal{G}$ , we can auto-regressively generate reaction sequences given interaction history $\\mathbf{z}^{i-1}$ , a diffusion sampler $\\mathcal{S}$ like [20, 53] and online text guidance $c$ by sampling with classifier-free guidance [19]:", + "bbox": [ + 89, + 651, + 483, + 727 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {G} _ {w} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, c\\right) = \\mathcal {G} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, \\varnothing\\right) \\\\ + \\quad w \\cdot \\left(\\mathcal {G} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, c\\right) - \\mathcal {G} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, \\varnothing\\right)\\right), \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 736, + 482, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w$ is the guidance hyper-parameter, as shown in Alg.1.", + "bbox": [ + 89, + 792, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Reaction Policy", + "text_level": 1, + "bbox": [ + 89, + 833, + 246, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our actor-motion data captured from virtual reality or existing datasets roughly approximates the dynamic plausibility. In contrast, the generated reactor motions is not always", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Online reaction sampling using auto-regressive motion diffusion model", + "bbox": [ + 513, + 90, + 906, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: denoiser $\\mathcal{G}$ , online actor loader $\\mathbf{Y}$ , online text loader $\\mathbf{C}$ , total diffusion steps $T$ , classifier-free guidance scale $w$ , auto-regressive reaction sampler $S$ .", + "bbox": [ + 527, + 125, + 906, + 170 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: reactor motion sequence $\\mathbf{X}$", + "bbox": [ + 529, + 170, + 774, + 185 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {z} _ {T} ^ {0} \\leftarrow \\mathcal {N} (0, 1) \\\\ \\mathbf {X} \\leftarrow \\mathcal {S} (\\mathcal {G}, \\mathbf {z} _ {T} ^ {0}, T, \\varnothing , \\varnothing , w) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 186, + 702, + 217 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\triangleright$ init with unconditioned sampling", + "bbox": [ + 671, + 217, + 905, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for $i\\gets 1$ to N do", + "bbox": [ + 531, + 231, + 653, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\triangleright$ num of actor windows", + "bbox": [ + 743, + 232, + 905, + 244 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} ^ {i - 1} \\leftarrow \\operatorname {C o l l e c t} (\\mathbf {Y}), \\mathbf {x} ^ {i - 1} \\leftarrow \\operatorname {C o l l e c t} (\\mathbf {X})\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 244, + 867, + 261 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} ^ {i - 1} \\leftarrow \\operatorname {C A N O N I C A L I Z E} \\left(\\mathbf {x} ^ {i - 1}, \\mathbf {y} ^ {i - 1}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 261, + 807, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "> get interaction history", + "bbox": [ + 745, + 277, + 903, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {T} ^ {i} \\leftarrow \\mathcal {N} (0, 1)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 291, + 653, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc ^ {i} \\leftarrow \\operatorname {C O L L E C T} (\\mathbf {C})\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 306, + 689, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {z}} _ {0} ^ {i} \\leftarrow S (\\mathcal {G}, \\mathbf {z} _ {T} ^ {i}, T, \\mathbf {z} ^ {i - 1}, c ^ {i}, w) \\triangleright C F G \\text {s a m p l i n g l o o p}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 321, + 903, + 337 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} ^ {i} \\leftarrow \\operatorname {R e c o v e r} \\left(\\hat {\\mathbf {z}} _ {0} ^ {i}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 337, + 694, + 352 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} \\leftarrow \\mathbf {X} \\cup \\hat {\\mathbf {x}} ^ {i}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 352, + 645, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\triangleright$ insert generated reaction to reactor sequences", + "bbox": [ + 591, + 368, + 903, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "end for", + "bbox": [ + 531, + 382, + 584, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "return X", + "bbox": [ + 531, + 397, + 591, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lying in a dynamically consistent manifold. Under the purpose of building the action-reaction controller in a physical world, we leverage the perpetual humanoid controller (PHC) [38], a universal physics tracker to track the generated reaction.", + "bbox": [ + 511, + 440, + 906, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PHC follows the conventional pipeline of generative adversarial and goal-conditioned reinforcement learning [18, 34], it receives target joint positions or rigid body's keypoints as the policy goals and outputs physically retargeted motions. The policy consists of multiple actor-critic model in a multiplicative manner. Given arbitrary action-reaction generation, we formulate the auto-regressive interaction as a standard Markov Decision Process (MDP), denoted as tuple $\\langle S, A, \\mathcal{P}, r, \\gamma, \\rho_0 \\rangle$ , representing state space, action space, transition distribution, reward, discount factor and initial states' distribution. In our settings, the MDP is also conditioned on action-reaction tuple $\\hat{z}$ , the standard bellman equation of state-action value can be reformulated as:", + "bbox": [ + 511, + 516, + 906, + 714 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Q ^ {\\pi} (s, a, \\hat {z}) = \\mathbb {E} _ {s ^ {\\prime} \\sim \\mathcal {P} (\\cdot | s, a, \\hat {z})} [ r (s, a, s ^ {\\prime}, \\hat {z}) + \\tag {9} \\\\ \\left. \\gamma \\mathbb {E} _ {a ^ {\\prime} \\sim \\pi (\\cdot | s ^ {\\prime}, \\hat {z})} [ Q ^ {\\pi} (s ^ {\\prime}, a ^ {\\prime}, \\hat {z} ^ {\\prime} ] \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 724, + 905, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{z} = \\langle \\hat{x},\\hat{y},y_{real}\\rangle$ is the augmented interaction goal, $\\hat{x}$ means a single predicted reactor frame regarded as the target goal, $\\hat{y}$ and $y_{real}$ are the actor motions coming from the diffusion model generation and real world inputs.", + "bbox": [ + 511, + 773, + 906, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Actor-aware Reaction Policy To conserve computational and communication resources while prioritizing interaction safety, we integrate a low-frequency (0.5 Hz)", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10177", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "diffusion planner with a high-frequency (30 Hz) physics tracker and real-world actor capture interface. This frequency mismatch can cause the reaction policy to pursue stale targets, elevating collision risk when the actor's motion deviates unexpectedly. To safeguard against interpenetration, we incorporate both generated and captured actor motion histories into the policy observation and heavily penalize the trajectory inconsistency in the reward function, thereby enforcing collision-free, safety-critical responses. More details of implementations in Appendix B.2.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Applications", + "text_level": 1, + "bbox": [ + 89, + 253, + 220, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Immersive VR Interface", + "text_level": 1, + "bbox": [ + 89, + 277, + 313, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our system is built upon Unity, enabling real-time interaction and visualization in a VR environment. Leveraging VisionPro [42], we develop a VR platform that allows the user to wear a headset and engage in immersive, egocentric interactions with the character. The camera in Unity is positioned on the actor's head, ensuring that it moves and rotates with the actor's motions. It continuously streams real-time simulations from Unity to the VR headset, providing an immersive experience. The user can perform any actions to guide the character's movement, such as completing a duet dance with the character.", + "bbox": [ + 89, + 300, + 483, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Human-Robot Interaction Interface", + "text_level": 1, + "bbox": [ + 89, + 473, + 401, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benefiting from our robust Reaction Motion Tracker, the generated motions adhere to physical constraints and exhibit high adaptability. As a result, we can seamlessly transfer physically plausible humanoid interaction to human-robot interaction once we retarget the imitation goal from SMPL[37] joints to specific robot joints. An optimization-based method is leveraged following [15, 16, 24]. First we fit the SMPL shape parameters $\\beta$ and use $\\beta$ to solve for optimal robot joint angles $\\theta$ . Then the corresponding joints' goal positions are calculated by forward kinematics. In the VR environment, the character can be replaced with the Unitree H1[49], enabling an immersive and realistic human-robot interaction experience. Furthermore, our approach holds potential for future deployment on physical robots.", + "bbox": [ + 89, + 494, + 483, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 733, + 220, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 89, + 757, + 375, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our model on two datasets: InterX [65], which contains 11,388 interaction SMPL-X [43] motion sequences on 40 interaction categories; InterHuman [32], which consists of 6022 interaction SMPL [37] motion sequences with around 1.7M frames. For data processing, we use the actor-reactor annotations provided by [66]. We canonicalize all the motions into the representation stated in Sec.3.3.1 and set frame rate to 30 FPS.", + "bbox": [ + 89, + 780, + 480, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We assess the generated reactor's sequence using three key aspects: 1) the quality of the reactor's motion independent of the actor, following kinematic motion generation works [14, 58]; 2) the physical plausibility of the reactor's motion, based on physics-based methods [59, 69]; and 3) the interaction quality between the actor and reactor, as discussed in [36, 52].", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For 1), we evaluate using Frechet Inception Distance (FID) [17] to measure the similarity between the generated and ground truth data distributions, Diversity to assess motion variability, and Multi-modal Distance (MMDist) to measure the distance between the generated reaction and the text prompt. For 2), we assess Penetration, the average distance between the ground and the lowest body joint position below the ground, Floating, the average distance between the ground and the lowest body joint above the ground, and Skating, the average horizontal displacement during frames exhibiting skating behavior. For 3), we measure maximum Interpenetration Volume (IV) [36] of reactor vertices intersecting the actor mesh, and cross-distance metrics $FID_{cd}$ and $Div_{cd}$ in a 100-dimensional cross-distance space [52] to evaluate interaction realism. Details of metrics calculation in Appendix C.2.", + "bbox": [ + 511, + 198, + 906, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 453, + 730, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our auto-regressive denoiser, we adopt the DIT architecture [44] consisting of 8 Transformer layers with a 512-dimensional hidden size. We employ a frozen DistilBERT encoder [10, 51] as the text encoder, which is widely used in motion generation tasks [14, 59]. For training and inference, we set the binary interaction field $\\mathcal{I} \\in \\mathbb{R}^{6 \\times 6}$ , the diffusion timestep $T = 8$ , the classifier-free guidance weight $w = 5$ , and the foot-contact, interaction, and prefix loss weights to $\\lambda_{\\mathrm{foot}} = 0.2$ , $\\lambda_{\\mathrm{inter}} = 0.5$ , and $\\lambda_{\\mathrm{prefix}} = 0.1$ , respectively. In the auto-regressive setting, we use an interaction history length of $h = 20$ frames and a prediction window length of $k = 40$ frames. We train for 300K denoising steps on both the InterHuman and Inter-X datasets, and observe optimal checkpoints at 250K steps for InterHuman and 220K steps for Inter-X. All models are trained on a single NVIDIA GeForce RTX 4090d GPU for approximately 30 hours.", + "bbox": [ + 511, + 476, + 905, + 732 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For physics policy training, we use the multiplicative version of PHC network with three primitives, and construct task goal with only target body keypoints. The policy is also trained on the same device and datasets as the former stage for 20 hours. The tracking policy is executed at $30\\mathrm{Hz}$ while physics simulation is processed in NVIDIA's Isaac Gym[41] at $60\\mathrm{Hz}$ . In the starting phase of inference, we capture the first 1 second motion from the actor, and store it in the buffer along with masked reactor motion prefix for the next planning horizon's generation, the downstream reactor tracker receives default joint positions as tracking goals.", + "bbox": [ + 511, + 734, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10178", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1d714e11b8df8ecffa32406a56a9227abf0c06487dc8b8bf4741f2bfc9914ae5.jpg", + "image_caption": [ + "Figure 3. Compared to CAMDM (top row), Human-X (bottom row) achieves more complete hand contact in tasks such as face-hitting and handshaking. Additionally, its foot movement appears more natural, as highlighted in the red and green circles." + ], + "image_footnote": [], + "bbox": [ + 107, + 89, + 486, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/411b3599f11e91dabc4ffa4f42c86d50eda391247161d130da4f0a9f47982e93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 95, + 895, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/33d889fc7db82c7f5e2a32e2ebb7fbe0b27642134625d49d4992158ba51299c0.jpg", + "image_caption": [ + "Figure 4. Visualization results on Human-Robot Interaction. The robot (black skeleton) and human (orange mesh) perform a handshake on a flat plane, from arm extension and palm contact through to shake completion, illustrating our method's spatial coordination and motion coherence." + ], + "image_footnote": [], + "bbox": [ + 93, + 290, + 890, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "After warming-up, the diffusion planner generates reasonable action-reaction motion sequences, the reactor is able to continuously follow the planning and receives both real and generated actor sequences for policy correction.", + "bbox": [ + 89, + 476, + 483, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Actor-To-Reactor Evaluation", + "text_level": 1, + "bbox": [ + 89, + 550, + 352, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. To evaluate the performance of our method, we adopt and reproduce several human reaction generation models as baselines: 1) InterFormer [8]: a transformer-based model with temporal and spatial attention for human reaction generation, 2) InterGen [32]: a diffusion-based model for generating human interaction motions conditioned on text prompts., 3) ReGenNet [66]: a diffusion-based model for generating human reactions conditioned on given human actions. and 4) CAMDM [7]: a transformer-based Conditional Auto-regressive Motion Diffusion Model. However, since it is designed for single-person motion generation, we utilize the action as a condition input to generate the corresponding reaction.", + "bbox": [ + 89, + 573, + 483, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons. As demonstrated in Tab. 2, in the online unconstrained reaction setting, our method exhibits significant improvements over baselines in Reaction Quality, Physical Plausibility, and Interaction Quality on both the Inter-X datasets. Unlike InterFormer's transformer architecture, our diffusion-based approach effectively captures the nuanced structure of data distributions, resulting in re", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "actor motions that closely align with real-world motion patterns and thereby enhancing reaction quality. In comparison to InterGen and ReGenNet, our implementation of Auto-regressive Motion Diffusion Models, augmented with a Physics Tracker, enables real-time interaction with the reactor by predicting future frames. This ensures that the generated motions adhere to fundamental physical laws, producing more natural and believable animations. Furthermore, relative to CAMDM, we introduce an Interaction Loss, leading to substantial improvements in IV and $\\mathrm{FID}_{cd}$ , thereby facilitating more natural interactions. The experiments on the InterHuman dataset can be found in C.1.", + "bbox": [ + 511, + 476, + 906, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 675, + 666, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We primarily conduct ablation studies on three aspects: Losses, Timesteps, and Horizon. For more experiments, please refer to Section C.3.", + "bbox": [ + 511, + 700, + 905, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Losses. To verify the effectiveness of our designed loss function, we ablate $\\mathcal{L}_{inter}$ , $\\mathcal{L}_{foot}$ , and $\\mathcal{L}_{prefix}$ separately. As shown in Tab. 3, removing any of them degrades performance. Notably, excluding $\\mathcal{L}_{inter}$ significantly increases $FID$ and $FID_{cd}$ , underscoring its importance for generating realistic interactions. Removing $\\mathcal{L}_{foot}$ raises Skating and Floating scores, showing its role in reducing foot sliding and enhancing physical plausibility.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10179", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4efe20ad26c42fcf971737ce37e4f66e8bdf246125673cc0e5ce03df1d3184a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsReactionPhysicsInteraction
FID↓Div.→MMDist. ↓Pene. ↓Skate. ↓Float. ↓IV ↓FIDcd ↓Divcd →
Ground Truth0.0026.0283.5240.0000.0237.9560.0240.23511.471
InterFormer [8]6.1504.9228.6480.5240.17524.1090.3326.8258.214
InterGen [32]5.4854.4176.3890.2700.14311.9150.2693.4378.730
ReGenNet [66]2.2125.0475.2350.1410.1189.8100.2181.9509.248
CAMDM [7]1.4305.6575.3180.1530.1279.8640.1452.0869.017
Human-X0.9756.0634.1150.1180.0928.6500.0761.6929.735
Human-X*1.2205.1405.0760.0260.0074.3420.0541.8338.565
", + "bbox": [ + 96, + 85, + 901, + 268 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/082391c46ba6e6105c7e7d3851d8332b6f2394aaa80003e0f34fabd882c69a9b.jpg", + "table_caption": [ + "Table 2. Action-to-Reaction with online unconstrained reaction setting on Inter-X [65] dataset. A higher or lower value is better for ↑ or ↓, and → means the value closer to ground truth is better. * denotes the method with reaction policy designed for humanoid reaction." + ], + "table_footnote": [], + "table_body": "
ClassSettingsReactionPhysicsInteractionLatency (ms)
FID↓Div.→MMDist.→Penetration ↓Skating ↓Floating ↓IV ↓FIDcd ↓Divcd→
Ground Truth0.0026.0283.5240.0000.0237.9560.0240.23511.471-
Lossesw.o. Linter1.4576.1544.2640.1340.1078.7150.1192.2089.91713.6
w.o. Lfoot1.0096.1144.2050.1270.2399.1200.0941.6879.84913.6
w.o. Lprefix1.0236.1284.2250.1330.1138.8570.1001.6979.86513.6
Timesteps21.8095.5175.0070.1440.1139.0080.1532.0087.5133.8
51.1076.2634.5140.1280.1088.9080.1151.8078.0038.2
100.9886.0084.2080.1180.0968.7090.0971.6849.85316.9
1000.9085.9073.9630.1130.0938.6080.0871.6109.708170
Horizonh=10,k=301.0556.1064.1070.1230.1038.8050.1031.6879.8079.6
h=15,k=300.9876.0054.0060.1180.0978.7070.0981.6679.78710.1
h=20,k=200.9575.9563.9580.1130.0948.6580.0931.6269.7569.5
h=15,k=451.0066.0544.1050.1230.1048.8060.1021.6889.80913.6
Human-Xtext0.9265.9513.9090.1120.0878.2180.0921.6079.82213.6
unconstrained0.9756.0634.1150.1180.0928.6500.0761.6929.73513.6
", + "bbox": [ + 94, + 318, + 903, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation studies of online reaction setting on the Inter-X [65] dataset.", + "bbox": [ + 259, + 561, + 733, + 577 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Timesteps. During the denoising process, we experiment with different sampling timesteps: 2, 5, 10, and 100. As shown in the Tab. 3, increasing the number of timesteps improves the model's performance; however, it also leads to higher latency. When timesteps = 10, the model has already achieved satisfactory results. Considering both generation quality and latency, we ultimately select timesteps = 8 as a trade-off, ensuring high-quality generation while meeting real-time requirements.", + "bbox": [ + 89, + 604, + 482, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Horizon. During inference process, the model takes a continuous sequence of $h$ history frames as input to predict $k$ future frames. We conduct experiments with different values of $h$ and $k$ , and the results in Tab. 3 show that the best performance is achieved when $h = 20$ and $k = 40$ .", + "bbox": [ + 89, + 757, + 482, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 845, + 209, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present Human-X Interaction, a novel framework for real-time human interaction synthesis that seamlessly in-", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tegrates autoregressive diffusion modeling with an actor-aware reaction policy. Our approach addresses the critical challenges of real-time synchronization, physical realism, and safety in human-ayar, human-humanoid and humanrobot interactions. Specifically, Human-X jointly predicts actions and reactions using both partner's motion histories, ensuring immersive and realistic motion synthesis. By incorporating comprehensive interaction-aware rewards and training the tracking policy with reinforcement learning, we achieve physically feasible and artifact-free interactions even in dynamic and safety-critical scenarios. Extensive experiments on the Inter-X and InterHuman datasets demonstrate significant improvements in motion quality, interaction continuity, and physical accuracy over existing methods. The current formulation focuses on synthesizing reactive motions and does not account for collaborative task completion between humans and machines. Extending the framework to support joint task execution with shared goals remains an open challenge.", + "bbox": [ + 511, + 604, + 906, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10180", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Shanghai Local College Capacity Building Program(23010503100), NSFC (No.62406195), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (Shanghai), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University), HPC Platform and Core Facility Platform of Computer Science and Communication of ShanghaiTech University and Shanghai Engineering Research Center of Intelligent Vision and Imaging.", + "bbox": [ + 89, + 116, + 483, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 285, + 187, + 301 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2", + "[2] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2", + "[3] Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems, 28, 2015. 5, 1", + "[4] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, et al. Digital life project: Autonomous 3d characters with social intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 582-592, 2024. 6", + "[5] Haoran Chen, Yiteng Xu, Yiming Ren, Yaoqin Ye, Xinran Li, Ning Ding, Peishan Cong, Ziyi Wang, Bushi Liu, Yuhan Chen, et al. Symbiosim: Human-in-the-loop simulation platform for bidirectional continuing learning in human-robot interaction. arXiv preprint arXiv:2502.07358, 2025. 6", + "[6] Jingxiao Chen, Xinyao Li, Jiahang Cao, Zhengbang Zhu, Wentao Dong, Minghuan Liu, Ying Wen, Yong Yu, Liqing Zhang, and Weinan Zhang. Rhino: Learning real-time humanoid-human-object interaction from human demonstrations. arXiv preprint arXiv:2502.13134, 2025. 6", + "[7] Rui Chen, Mingyi Shi, Shaoli Huang, Ping Tan, Taku Komura, and Xuelin Chen. Taming diffusion probabilistic models for character control. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 7, 8, 3, 4, 5", + "[8] Baptiste Chopin, Hao Tang, Naima Otberdout, Mohamed Daoudi, and Nicu Sebe. Interaction transformer for human reaction generation. IEEE Transactions on Multimedia, 25: 8842-8854, 2023. 7, 8, 3", + "[9] Zekai Deng, Ye Shi, Kaiyang Ji, Lan Xu, Shaoli Huang, and Jingya Wang. Human-object interaction via automatically designed vlm-guided motion policy. arXiv preprint arXiv:2503.18349, 2025. 2" + ], + "bbox": [ + 99, + 311, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 6", + "[11] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm- e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning, pages 8469-8488, 2023. 6", + "[12] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. 2", + "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[14] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5152-5161, 2022. 4, 6, 2", + "[15] Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. arXiv preprint arXiv:2406.08858, 2024. 6", + "[16] Tairan He, Zhengyi Luo, Wenli Xiao, Chong Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Learning human-to-humanoid real-time whole-body teleoperation, 2024. 6", + "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[18] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 3, 5", + "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4, 5", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4, 5", + "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 6", + "[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10181", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition, pages 16750-16761, 2023. 2", + "[23] Muhammad Gohar Javed, Chuan Guo, Li Cheng, and Xingyu Li. Intermask: 3d human interaction generation via collaborative masked modelling. arXiv preprint arXiv:2410.10010, 2024. 2", + "[24] Mazeyu Ji, Xuanbin Peng, Fangchen Liu, Jialong Li, Ge Yang, Xuxin Cheng, and Xiaolong Wang. Exbody2: Advanced expressive humanoid whole-body control. arXiv preprint arXiv:2412.13196, 2024. 6", + "[25] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2", + "[26] Jianping Jiang, Weiye Xiao, Zhengyu Lin, Huaizhong Zhang, Tianxiang Ren, Yang Gao, Zhiqian Lin, Zhongang Cai, Lei Yang, and Ziwei Liu. Solami: Social vision-language-action modeling for immersive interaction with 3d autonomous characters. arXiv preprint arXiv:2412.00174, 2024. 1, 2, 6", + "[27] Wentao Jiang, Jingya Wang, Kaiyang Ji, Baoxiong Jia, Siyuan Huang, and Ye Shi. Arflow: Human action-reaction flow matching with physical guidance. arXiv preprint arXiv:2503.16973, 2025. 2", + "[28] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 2", + "[29] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 2", + "[30] Jiefeng Li, Siyuan Bian, Chao Xu, Zhicun Chen, Lixin Yang, and Cewu Lu. Hybrid analytical-neural inverse kinematics for whole-body mesh recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3", + "[31] Ronghui Li, Youliang Zhang, Yachao Zhang, Yuxiang Zhang, Mingyang Su, Jie Guo, Ziwei Liu, Yebin Liu, and Xiu Li. Interdance: Reactive 3d dance generation with realistic duet interactions. arXiv preprint arXiv:2412.16982, 2024. 2", + "[32] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 2, 6, 7, 8, 3, 4, 5", + "[33] Jing Lin, Yao Feng, Weiyang Liu, and Michael J Black. Chathuman: Language-driven 3d human understanding with retrieval-augmented tool reasoning. arXiv preprint arXiv:2405.04533, 2024. 6", + "[34] Minghuan Liu, Menghui Zhu, and Weinan Zhang. Goal-conditioned reinforcement learning: Problems and solutions, 2022. 5", + "[35] Yunze Liu, Changxi Chen, and Li Yi. Interactive humanoid: Online full-body motion reaction synthesis with social affordance canonicalization and forecasting. arXiv preprint arXiv:2312.08983, 2023. 2", + "[36] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-reaction: Physically plausible real-time humanoid reaction" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthesis via forward dynamics guided 4d imitation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3771-3780, 2024. 2, 3, 6", + "[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 4, 6", + "[38] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10895-10904, 2023. 3, 4, 5", + "[39] Zhengyi Luo, Jinkun Cao, Josh Merel, Alexander Winkler, Jing Huang, Kris Kitani, and Weipeng Xu. Universal humanoid motion representations for physics-based control. arXiv preprint arXiv:2310.04582, 2023. 3", + "[40] Zhengyi Luo, Jinkun Cao, Rawal Khirodkar, Alexander Winkler, Kris Kitani, and Weipeng Xu. Real-time simulated avatar from head-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 571-581, 2024. 4", + "[41] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 4, 6", + "[42] Younghyo Park and Pulkit Agrawal. Using apple vision pro to train and control robots, 2024. 6", + "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 6", + "[44] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 6", + "[45] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4):1-14, 2018. 3", + "[46] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4, 5", + "[48] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 5, 1", + "[49] Unitree Robotics. Unitree h1 robot. https://www.unitree.com, 2021. Accessed: 2023-10-10. 6" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10182", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 6", + "[51] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 6, 4", + "[52] Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. arXiv preprint arXiv:2403.18811, 2024. 2, 6", + "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4, 5", + "[54] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39 (4):54-1, 2020. 1", + "[55] Wenhui Tan, Boyuan Li, Chuhao Jin, Wenbing Huang, Xiting Wang, and Ruihua Song. Think then react: Towards unconstrained action-to-reaction motion generation. In The Thirteenth International Conference on Learning Representations, 2025. 2", + "[56] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15999-16009, 2023. 2", + "[57] Jiangnan Tang, Jingya Wang, Kaiyang Ji, Lan Xu, Jingyi Yu, and Ye Shi. A unified diffusion framework for scene-aware human motion estimation from sparse signals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21251-21262, 2024. 2", + "[58] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 6", + "[59] Guy Tevet, Sigal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. Closd: Closing the loop between simulation and diffusion for multi-task character control. arXiv preprint arXiv:2410.03441, 2024. 4, 6, 2", + "[60] Jie Tian, Ran Ji, Lingxiao Yang, Suting Ni, Yuexin Ma, Lan Xu, Jingyi Yu, Ye Shi, and Jingya Wang. Gaze-guided hand-object interaction synthesis: Dataset and method. arXiv preprint arXiv:2403.16169, 2024. 2", + "[61] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 2, 6", + "[62] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 2", + "[63] Zhen Wu, Jiaman Li, and C Karen Liu. Human-object interaction from human-level instructions. CoRR, 2024. 2, 6" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Liang Xu, Ziyang Song, Dongliang Wang, Jing Su, Zhicheng Fang, Chenjing Ding, Weihao Gan, Yichao Yan, Xin Jin, Xiaokang Yang, et al. Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2023. 2", + "[65] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter-x: Towards versatile human-human interaction analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22260-22271, 2024. 2, 6, 8, 4, 5", + "[66] Liang Xu, Yizhou Zhou, Yichao Yan, Xin Jin, Wenhan Zhu, Fengyun Rao, Xiaokang Yang, and Wenjun Zeng. Regennet: Towards human action-reaction synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1759-1769, 2024. 2, 6, 7, 8, 3, 4, 5", + "[67] Lingxiao Yang, Shutong Ding, Yifan Cai, Jingyi Yu, Jingya Wang, and Ye Shi. Guidance with spherical gaussian constraint for conditional diffusion. In Proceedings of the 41st International Conference on Machine Learning, pages 56071-56095, 2024. 2", + "[68] Ye Yuan, Viktor Makoviychuk, Y Guo, S Fidler, XB Peng, and K Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Trans. Graph, 42(4), 2023. 5, 1", + "[69] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16010-16021, 2023. 3, 6, 2", + "[70] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2", + "[71] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE transactions on pattern analysis and machine intelligence, 46(6):4115-4128, 2024. 2", + "[72] Zhenhao Zhang, Ye Shi, Lingxiao Yang, Suting Ni, Qi Ye, and Jingya Wang. Openhoi: Open-world hand-object interaction synthesis with multimodal large language model. arXiv preprint arXiv:2505.18947, 2025. 2", + "[73] Kaifeng Zhao, Gen Li, and Siyu Tang. Dart: A diffusion-based autoregressive motion model for real-time text-driven motion control. arXiv preprint arXiv:2410.05260, 2024. 5, 1", + "[74] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. 4", + "[75] Kaizhen Zhu, Mokai Pan, Yuexin Ma, Yanwei Fu, Jingyi Yu, Jingya Wang, and Ye Shi. Unidb: A unified diffusion bridge framework via stochastic optimal control. arXiv preprint arXiv:2502.05749, 2025. 2" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10183", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/4564e033-7cc9-41ba-b940-20704a40b38a_model.json b/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/4564e033-7cc9-41ba-b940-20704a40b38a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..74305ae835cd240e564cfd1ed7c7f74155803727 --- /dev/null +++ b/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/4564e033-7cc9-41ba-b940-20704a40b38a_model.json @@ -0,0 +1,2532 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.135, + 0.13, + 0.865, + 0.177 + ], + "angle": 0, + "content": "Towards Immersive Human-X Interaction: A Real-Time Framework for Physically Plausible Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.202, + 0.807, + 0.239 + ], + "angle": 0, + "content": "Kaiyang Ji\\(^{1}\\) Ye Shi\\(^{1,2,*}\\) Zichen Jin\\(^{1}\\) Kangyi Chen\\(^{1}\\) Lan Xu\\(^{1,2}\\) Yuexin Ma\\(^{1,2}\\) Jingyi Yu\\(^{1,2}\\) Jingya Wang\\(^{1,2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.397, + 0.24, + 0.603, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)ShanghaiTech University" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.257, + 0.792, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)Shanghai Engineering Research Center of Intelligent Vision and Imaging" + }, + { + "type": "text", + "bbox": [ + 0.131, + 0.278, + 0.862, + 0.293 + ], + "angle": 0, + "content": "{jiky2024,shiye,jinzch12023,v-chenky,xulan1,mayuexin,yujingyi,wangjingya}@shanghaiitech.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.31, + 0.895, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.471, + 0.908, + 0.499 + ], + "angle": 0, + "content": "Figure 1. We propose Human-X, the first framework designed to enable latency-free interaction between humans and diverse entities, including human-athlete, human-humanoid, and human-robot interaction." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.512, + 0.327, + 0.528 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.544, + 0.486, + 0.877 + ], + "angle": 0, + "content": "Real-time synthesis of physically plausible human interactions remains a critical challenge for immersive VR/AR systems and humanoid robotics. While existing methods demonstrate progress in kinematic motion generation, they often fail to address the fundamental tension between real-time responsiveness, physical feasibility, and safety requirements in dynamic human-machine interactions. We introduce Human-X, a novel framework designed to enable immersive and physically plausible human interactions across diverse entities, including human-ayar, human-humanoid, and human-robot systems. Unlike existing approaches that focus on post-hoc alignment or simplified physics, our method jointly predicts actions and reactions in real-time using an auto-regressive reaction diffusion planner, ensuring seamless synchronization and context-aware responses. To enhance physical realism and safety, we integrate an actor-aware motion tracking policy trained with reinforcement learning, which dynamically adapts to interaction partners' movements while avoiding artifacts like foot sliding and penetration. Extensive experiments on the Inter-X and InterHuman datasets demonstrate significant improvements in motion quality, interaction continuity, and physical" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.909, + 0.603 + ], + "angle": 0, + "content": "plausibility over state-of-the-art methods. Our framework is validated in real-world applications, including virtual reality interface for human-robot interaction, showcasing its potential for advancing human-robot collaboration. Project page: https://humanx-interaction.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.633, + 0.646, + 0.649 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Human-human interaction synthesis has long been a focal point in the field of motion generation, with widespread applications in social dynamics, gaming, and animation [26, 54]. With the rapid development of virtual reality (VR) and augmented reality (AR), these technologies now empower the creation of deeply immersive and contextually rich interactions between humans and digital avatars. Particularly, the emergence of humanoid robots further bridges the gap between virtual and physical realms, enabling robots to engage with humans in daily life through adaptive and socially aware behaviors. However, achieving real-time prediction of diverse and safety-guaranteed humanoid reactions, an essential capability for applications ranging from daily companionship to assistive care, remains a critical, unsolved challenge." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Current approaches to human interaction modeling rely" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.236, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10173" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.907, + 0.188 + ], + "angle": 0, + "content": "
Methodsphysics-basedimmersive inferencelatency-freereal-time modelgeneration abilitystrong contacttext-control
InterGen [32]✓✓✓
ReGenNet [66]✓✓✓✓
SoLaMi [26]✓✓✓
PhysReaction [36]✓✓✓✓
Think-Then-React [55]✓✓✓✓
Human-X✓✓✓✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.209, + 0.199, + 0.786, + 0.213 + ], + "angle": 0, + "content": "Table 1. Comparative analysis of Human-X versus existing human reaction synthesis approaches." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.24, + 0.482, + 0.406 + ], + "angle": 0, + "content": "heavily on offline training data or simplified physics assumptions, leading to limited adaptability in dynamic scenarios. While data-driven methods such as [23, 32, 65] have shown promise in synthesizing kinematic motions interaction, they often fail to account for the intricate physical constraints and safety requirements inherent in human-humanoid cohabitation. For instance, abrupt or unstable motions generated through these techniques pose risks in shared environments. This gap underscores the need for a unified interaction framework that integrates real-time reactivity, physical realism, and safety assurance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.41, + 0.483, + 0.681 + ], + "angle": 0, + "content": "In this work, we explore Human-X Interaction (Fig.1), a systematic study of immersive interactions across three domains: 1) Human-Avatar Interaction using kinematic-driven motion in VR/AR; 2) Human-Humanoid Interaction with physics-based motion synthesis; 3) Human-Robot Interaction ensuring safe physical contact. Unlike existing physics-aware methods limited to replicating training data, our framework balances natural movement with physical feasibility while maintaining real-time performance. This integration enables safer and more natural human-machine collaboration across applications from entertainment to assistive robotics. Achieving real-time synchronization between interacting partners faces major technical challenges. Conventional methods rely exclusively on past actor motions to generate reactions, leading to temporal mismatches. Furthermore, ensuring physical realism is more challenging, as existing solutions focus mainly on basic collision avoidance and neglect dynamic interaction forces." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.82 + ], + "angle": 0, + "content": "To overcome these limitations, we present Human-X, an Auto-Regressive Action-Reaction Diffusion Planner for real-time interaction. Our framework jointly models both partners' motion histories to predict subsequent actions and reactions. We introduce multiple contact-aware losses to train realistic interactions, and combine the diffusion model with an actor-aware tracking policy to ensure physical feasibility in humanoid/robot scenarios. In summary, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.884 + ], + "angle": 0, + "content": "- The first auto-regressive diffusion model that jointly predicts actions and reactions using both partners' motion history with comprehensive contract losses to ensure real-time, immersive and realistic motion synthesis." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.887, + 0.483, + 0.9 + ], + "angle": 0, + "content": "- An actor-aware reaction policy trained with reinforce" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.24, + 0.905, + 0.285 + ], + "angle": 0, + "content": "ment learning that dynamically adapts to interaction partners' motions, preventing common artifacts like foot sliding while maintaining physical feasibility." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.286, + 0.905, + 0.361 + ], + "angle": 0, + "content": "- A complete real-time VR system demonstrating practical effectiveness. Experiments on Inter-X and InterHuman datasets show significant improvements in motion quality, interaction continuity, and physical accuracy compared to existing methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.372, + 0.664, + 0.387 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.397, + 0.761, + 0.412 + ], + "angle": 0, + "content": "2.1. Human Motion Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.689 + ], + "angle": 0, + "content": "Human-centric motion generation has seen remarkable advancements in recent years, driven by breakthroughs in large language models [1, 25] and generative frameworks [13, 20, 28, 67, 75]. Notably, the introduction of motion diffusion models [58, 71] has significantly elevated the performance of single-person motion generation. Building on these developments, researchers have increasingly focused on enriching interactions with the environment by incorporating diverse modalities, such as text [25, 62, 63, 70], audio [2, 61], and scene information [12, 22, 29, 57]. For instance, recent works [9, 60, 62, 63, 72] have leveraged text-based guidance to facilitate Human-Object Interactions (HOI), while MotionGPT [25] treats motion as a form of language, enabling seamless bidirectional translation between text and motion. Despite these advancements, existing approaches face a critical limitation: their model architectures are inherently ill-suited for real-time reaction synthesis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.763, + 0.713 + ], + "angle": 0, + "content": "2.2. Human-Human Interaction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Research on human-human interaction synthesis has evolved around three primary directions: (1) Interactive Behavior Modeling leverages actor-reactor role decomposition to capture interaction dynamics. For example, [56] identifies role-specific patterns, while [32, 64] use transformer architectures to model multi-agent dependencies. (2) Multimodal Interaction integrates text, audio, or scene context to enhance realism [31, 52], but often overlook dynamic physical constraints critical for real-world deployment. (3) Reactive Motion Generation employs sequence models [35] and diffusion models [27, 66] to synthesize temporally coherent reactions. These methods, however, struggle with physical" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10174" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.094, + 0.91, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.367, + 0.908, + 0.451 + ], + "angle": 0, + "content": "Figure 2. Overview of our immersive real-time interaction synthesis pipeline: (a) Actor Motion Capture: A human actor's movements are recorded at 30 fps by an RGB-D camera and translated into 3D poses, which are then retargeted to a humanoid character. (b) Realistic Reactor Motion Generation: An auto-regressive diffusion model, guided by optional text prompts (e.g., \"Dancing is what to do\"), generates plausible reaction motions. These motions are tracked by an actor-aware controller, which uses proprioception signals to ensure realistic, synchronized interactions. (c) Real-time VR Interface: The generated and tracked motions are rendered in simulator, providing both a third-person view and a binocular VR view." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.486, + 0.659 + ], + "angle": 0, + "content": "plausibility and temporal consistency in long-horizon interactions. Existing Reactive Generation methods face a fundamental tension: Human interaction models achieve reactivity but lack physical grounding, while physics-based animations enforce plausibility but are agnostic to interaction dynamics. While PhysReaction [36] introduced the first physics-based approach for reaction synthesis using imitation learning, it was limited to mimicking known distributions, lacking generalization ability, diversity, controllability, and the capacity to handle unseen tasks. Our work bridges this gap by unifying interaction-aware reactivity with physically constrained control." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.373, + 0.684 + ], + "angle": 0, + "content": "2.3. Physics-based Motion Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.902 + ], + "angle": 0, + "content": "In contrast to kinematics-based methods, physics-based motion synthesis has emerged as a widely adopted approach for generating realistic and physically plausible character movements by integrating physical principles into the animation process. Early work, such as DeepMimic [45], leveraged deep reinforcement learning to enable characters to imitate diverse motion clips within a physics-based framework. Subsequent advancements [18, 46] utilized generative adversarial networks to enhance stylized motion realism. Progress continued with PHC [38] and PULSE [39], focusing on improving the diversity of generated motions. A significant milestone was achieved with PhysDiff [69], which introduced physics-guided diffusion models to mitigate artifacts such as floating and foot sliding. Despite" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.508 + ], + "angle": 0, + "content": "these advancements, the physical plausibility of human-human interactions remains largely unexplored." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.521, + 0.606, + 0.536 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.684, + 0.561 + ], + "angle": 0, + "content": "3.1. System Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.84 + ], + "angle": 0, + "content": "In this section, we present a detailed overview of our immersive interaction synthesis system. To facilitate realistic and responsive interactions between two agents in immersive environments, our system takes online actor motion capture data as input, processes it through an advanced motion generation pipeline, and outputs continuous, reactive motion in real-time, as shown in Fig.2. The first key component of our system, detailed in Sec.3.2, is the real-time capture of the actor's pose and the reconstruction of its state in physics simulation. Subsequently, the Auto-regressive Interaction Diffusion Model, as described in Sec.3.3, is employed to generate the reaction of one agent in response to the actor context. Next, we introduce the Reaction Motion Tracker in Sec.3.4, which ensures that the motion of the reactor adheres to physical constraints for realism and natural behavior. Finally, we explain how VR Interface works in Sec.4.1, which enables immersive, real-time interaction with the generated motions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.892, + 0.865 + ], + "angle": 0, + "content": "3.2. Human Pose Estimation and Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We firstly utilize a real-time RGB-D camera to capture images. Then, HybrIK [30] is employed for latency-free re" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10175" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "construction of the human pose. However, due to the inaccuracies in the depth information provided by HybrIK, the depth map is directly obtained from the RGB-D camera instead. Following [38, 40], we bridge SMPL's kinematic skeleton to physics simulation via applying a simple retargeting method. We utilize Isaac Gym [41] as our physics simulation platform to create a third-person perspective for the interaction between the actor and reactor. Besides, the image reading process, motion generation of the reactor, and the physical simulation within Isaac Gym are handled by separate processes. ROS2 is used as the communication framework to facilitate real-time data transmission." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.284, + 0.449, + 0.3 + ], + "angle": 0, + "content": "3.3. Online Human Action-Reaction Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.484, + 0.413 + ], + "angle": 0, + "content": "The goal of online action-reaction synthesis is to generate realistic and continuous future \\( k \\) frames of reactor motions \\( \\mathbf{x}^i \\coloneqq \\{x^{n:n + k - 1}\\} \\) given \\( h \\) history frames of interaction contexts \\( \\{\\mathbf{x}^{i - 1},\\mathbf{y}^{i - 1}\\} \\coloneqq \\{x^{n - h:n - 1},y^{n - h:n - 1}\\} \\) at current frame number \\( n \\) and current window index \\( i \\). During each window prediction, an optional text prompt \\( c \\) can be added to guide the reaction generation process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.422, + 0.4, + 0.437 + ], + "angle": 0, + "content": "3.3.1. Reactor-centric Interaction Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.639 + ], + "angle": 0, + "content": "Reactor-centric Canonicalization. For individual motion synthesis, the initial frame of each sequence is typically aligned at the root to improve the generalizability of the model [14, 59]. However, using the same normalization for human-human interaction data can eliminate key details about the mutual spatial relationship. To retain relative positioning without losing generalization ability, we leverage a reactor-centric canonicalization that the pose of reactor is realigned at the root and oriented toward the forward direction, while the pose of actor is adjusted by the same transformation. This method maintains the essential spatial context between the two, enabling the model to accurately learn their interdependent movement dynamics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Reactor-centric Interaction Representation. We model each frame of interaction as a hybrid representation of both actor and reactor motion. For reaction representation, we follow previous works [14, 58] and use the HumanML3D representation. The reactor motion \\( x \\) in frame \\( n \\) is defined by \\( x^{n} = (r^{x},\\dot{r}^{x},p^{x},\\dot{p}^{x},\\theta^{x},f^{x}) \\), where \\( r^x\\in \\mathbb{R} \\) denotes the height of the roots, \\( \\dot{r}^x\\in \\mathbb{R}^3 \\) denotes the angular velocity of the roots along the Y axis and the linear velocity of the roots in the XZ plane, \\( p^x\\in \\mathbb{R}^{(J - 1)\\times 3} \\) are the local joint positions, \\( \\dot{p}^x\\in \\mathbb{R}^{J\\times 3} \\) denotes the temporal difference of the local joint positions, \\( \\theta^x\\in \\mathbb{R}^{(J - 1)\\times 6} \\) is the 6D representation [74] of the joint rotations and \\( f^x\\in \\mathbb{R}^4 \\) are binary labels for the detection of contact. For actor motion captured from pose estimation, we define the actor \\( y \\) in frame \\( n \\) by \\( y^n = (r^{x\\to y},\\dot{r}^{x\\to y},p^{x\\to y},\\dot{p}^{x\\to y},f^y) \\) where all the features are in the reactor's reference frame, making it a pure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.243 + ], + "angle": 0, + "content": "relative representation consisting of relative root transition, relative root rotation, relative root linear velocity, relative joint positions, relative joint velocity and feet contact labels. To emphasize actor-reactor contacts, we introduce the binary interaction field \\(\\mathcal{I}^n\\in \\mathbb{R}^{\\bar{J}\\times \\bar{J}}\\) obtained by thresholding the relative distance between the selected \\(\\bar{J}\\) joints of each agent in a contact map. For body model selection, we follow the skeleton of SMPL [37] with \\(J = 22\\) joints. The total reactor-centric interaction \\(\\mathbf{z}\\) at frame \\(n\\) representation can be formed as," + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.253, + 0.907, + 0.271 + ], + "angle": 0, + "content": "\\[\nz ^ {n} = \\left(x ^ {n}, y ^ {n}, \\mathcal {I} ^ {n}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.28, + 0.907, + 0.34 + ], + "angle": 0, + "content": "with \\(x^n\\) denotes the reactor motion, \\(y^n\\) denotes the relative actor motion and \\(\\mathcal{I}^n\\) is the joint contact features. Thus, we can model the interaction sequence as \\(\\mathbf{z}^i = \\{z^{n:n + k - 1}\\}\\) given \\(k\\) frames and window index \\(i\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.348, + 0.878, + 0.362 + ], + "angle": 0, + "content": "3.3.2. Auto-regressive Interaction Diffusion Planner" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.367, + 0.906, + 0.457 + ], + "angle": 0, + "content": "Based on the interaction motion representation provided in Sec.3.3.1, we transfer raw motion data into a canonical space suitable for generative learning with denoising networks. For probabilistic generator, we follow the DDPM [20] framework by incrementally adding Gaussian noise to the clean data sample \\(z_0\\) over \\(T\\) steps:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.468, + 0.907, + 0.485 + ], + "angle": 0, + "content": "\\[\nq \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {\\alpha_ {t}} \\mathbf {z} _ {t - 1}, (1 - \\alpha_ {t}) \\mathbf {I}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.906, + 0.587 + ], + "angle": 0, + "content": "where \\(\\alpha_{t} \\in (0,1)\\) are constant parameters for variance scheduling. Given the noised data \\(\\mathbf{z}_t^i\\) at the \\(i\\)-th window clip, the diffusion step \\(t\\), the prefix interaction window \\(\\mathbf{z}^{i-1} = z^{n-h:n-1}\\), and the optional text prompt \\(c\\), we train a transformer-based denoiser \\(\\mathcal{G}\\) to predict the cleaned interaction primitive itself instead of predicting the noise [58]:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.599, + 0.905, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {z}} _ {0} ^ {i} = \\mathcal {G} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, c\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.906, + 0.713 + ], + "angle": 0, + "content": "where the text prompt \\( c \\) is randomly masked by a rate of 0.15 during training to enable unconditioned generation and classifier-free guidance [19]. The predicted motion \\( \\hat{\\mathbf{z}}_0^i \\) is then inserted to the interaction history for next interaction window extraction after the diffusion sampling loop [20, 53]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.906, + 0.762 + ], + "angle": 0, + "content": "Losses The denoised motion \\(\\hat{\\mathbf{z}}_0^i\\) is supervised by a simple objective in DDPM [20]:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.771, + 0.905, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s i m p l e}} = \\mathbb {E} _ {\\mathbf {z} _ {0} ^ {i} \\sim q \\left(\\mathbf {z} _ {0} ^ {i}\\right), t \\sim [ 1, T ]} \\left[ \\| \\mathbf {z} _ {0} ^ {i} - \\hat {\\mathbf {z}} _ {0} ^ {i} \\| _ {2} ^ {2} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.799, + 0.906, + 0.86 + ], + "angle": 0, + "content": "We extend the training objective with several auxiliary losses to enhance interaction quality and realism. To optimize foot contact quality, we add a foot contact loss on both agents:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.87, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f o o t}} = \\sum_ {j \\in \\{x ^ {\\text {f e e t}}, y ^ {\\text {f e e t}} \\}} \\| \\dot {p} ^ {j} \\left(\\hat {\\mathbf {z}} _ {0} ^ {i}\\right) \\odot f ^ {j} \\left(\\hat {\\mathbf {z}} _ {0} ^ {i}\\right) \\| _ {2} ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10176" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.198 + ], + "angle": 0, + "content": "where \\(\\dot{p}^j (\\hat{\\mathbf{z}}_0^i)\\) is the corresponding foot joints velocity, \\(f^{j}(\\hat{\\mathbf{z}}_{0}^{i})\\) are foot contact label described in Sec 3.3.1 and \\(\\odot\\) denotes for Hadamard product. Note that \\(\\dot{p}^j (\\hat{\\mathbf{z}}_0^i)\\) is defined by the relative local velocity in the reactor frame, and we transform it to the global velocity in loss calculation. For interaction refinement, we design an interaction loss on contact joints distance:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.209, + 0.484, + 0.245 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i n t e r} = \\sum_ {x \\in \\bar {J} ^ {x}, y \\in \\bar {J} ^ {y}} \\| (p ^ {x} (\\hat {\\mathbf {z}} _ {0} ^ {i}) - p ^ {y} (\\hat {\\mathbf {z}} _ {0} ^ {i})) \\odot \\mathcal {I} ^ {x, y} (\\hat {\\mathbf {z}} _ {0} ^ {i}) \\| _ {2} ^ {2}, (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.255, + 0.485, + 0.362 + ], + "angle": 0, + "content": "where \\(\\mathcal{I}^{x,y}(\\hat{\\mathbf{z}}_0^i)\\) are binary body interaction fields and \\(p^x (\\hat{\\mathbf{z}}_0^i), p^y (\\hat{\\mathbf{z}}_0^i)\\) are contact joints positions in the same root frame to enhance strong interaction. To improve the continuity of auto-regressive generation, we finally add a prefix loss \\(\\mathcal{L}_{prefix}\\) to evaluate the difference between the end of prefix motion and the beginning of the generated motion. The full loss applied to the model is then:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.373, + 0.484, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\text {s i m p l e}} + \\lambda_ {\\text {f o o t}} \\mathcal {L} _ {\\text {f o o t}} + \\lambda_ {\\text {i n t e r}} \\mathcal {L} _ {\\text {i n t e r}} \\tag {7} \\\\ + \\lambda_ {p r e f i x} \\mathcal {L} _ {p r e f i x}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.419, + 0.484, + 0.451 + ], + "angle": 0, + "content": "where \\(\\lambda_{foot}\\), \\(\\lambda_{inter}\\) and \\(\\lambda_{prefix}\\) are hyper-parameters to balance the loss terms from different scaling aspects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.633 + ], + "angle": 0, + "content": "Training During training, we utilize the scheduled training strategy [3, 48, 68, 73] to enhance auto-regressive generation stability and text prompt controllability. The motion is cropped to sequences of \\( N \\) consecutive interaction windows with a random offset from the dataset. During training, \\( \\mathbf{z}_0^i \\) is noised to \\( \\mathbf{z}_t^i \\) at diffusion timestep \\( t \\), and interaction history \\( \\mathbf{z}^{i-1} \\) is divide into 3 stages with the sample probability \\( p \\): ground truth for supervised learning \\( (p = 1) \\), random sampling for scheduled training (decaying \\( p \\)), and auto-regressive prediction \\( (p = 0) \\). Details in Appendix B.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.484, + 0.728 + ], + "angle": 0, + "content": "Inference After training a denoise network \\(\\mathcal{G}\\), we can auto-regressively generate reaction sequences given interaction history \\(\\mathbf{z}^{i-1}\\), a diffusion sampler \\(\\mathcal{S}\\) like [20, 53] and online text guidance \\(c\\) by sampling with classifier-free guidance [19]:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.737, + 0.483, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {G} _ {w} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, c\\right) = \\mathcal {G} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, \\varnothing\\right) \\\\ + \\quad w \\cdot \\left(\\mathcal {G} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, c\\right) - \\mathcal {G} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, \\varnothing\\right)\\right), \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.484, + 0.825 + ], + "angle": 0, + "content": "where \\(w\\) is the guidance hyper-parameter, as shown in Alg.1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.247, + 0.85 + ], + "angle": 0, + "content": "3.4. Reaction Policy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our actor-motion data captured from virtual reality or existing datasets roughly approximates the dynamic plausibility. In contrast, the generated reactor motions is not always" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.091, + 0.907, + 0.122 + ], + "angle": 0, + "content": "Algorithm 1 Online reaction sampling using auto-regressive motion diffusion model" + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.126, + 0.907, + 0.171 + ], + "angle": 0, + "content": "Input: denoiser \\(\\mathcal{G}\\), online actor loader \\(\\mathbf{Y}\\), online text loader \\(\\mathbf{C}\\), total diffusion steps \\(T\\), classifier-free guidance scale \\(w\\), auto-regressive reaction sampler \\(S\\)." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.171, + 0.775, + 0.186 + ], + "angle": 0, + "content": "Output: reactor motion sequence \\(\\mathbf{X}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.187, + 0.703, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {z} _ {T} ^ {0} \\leftarrow \\mathcal {N} (0, 1) \\\\ \\mathbf {X} \\leftarrow \\mathcal {S} (\\mathcal {G}, \\mathbf {z} _ {T} ^ {0}, T, \\varnothing , \\varnothing , w) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.673, + 0.218, + 0.906, + 0.232 + ], + "angle": 0, + "content": "\\(\\triangleright\\) init with unconditioned sampling" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.232, + 0.655, + 0.245 + ], + "angle": 0, + "content": "for \\(i\\gets 1\\) to N do" + }, + { + "type": "text", + "bbox": [ + 0.744, + 0.233, + 0.906, + 0.245 + ], + "angle": 0, + "content": "\\(\\triangleright\\) num of actor windows" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.246, + 0.868, + 0.262 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} ^ {i - 1} \\leftarrow \\operatorname {C o l l e c t} (\\mathbf {Y}), \\mathbf {x} ^ {i - 1} \\leftarrow \\operatorname {C o l l e c t} (\\mathbf {X})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.262, + 0.808, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} ^ {i - 1} \\leftarrow \\operatorname {C A N O N I C A L I Z E} \\left(\\mathbf {x} ^ {i - 1}, \\mathbf {y} ^ {i - 1}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.746, + 0.279, + 0.905, + 0.292 + ], + "angle": 0, + "content": "> get interaction history" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.292, + 0.655, + 0.308 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {T} ^ {i} \\leftarrow \\mathcal {N} (0, 1)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.308, + 0.69, + 0.322 + ], + "angle": 0, + "content": "\\[\nc ^ {i} \\leftarrow \\operatorname {C O L L E C T} (\\mathbf {C})\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.322, + 0.905, + 0.338 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {z}} _ {0} ^ {i} \\leftarrow S (\\mathcal {G}, \\mathbf {z} _ {T} ^ {i}, T, \\mathbf {z} ^ {i - 1}, c ^ {i}, w) \\triangleright C F G \\text {s a m p l i n g l o o p}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.338, + 0.696, + 0.353 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} ^ {i} \\leftarrow \\operatorname {R e c o v e r} \\left(\\hat {\\mathbf {z}} _ {0} ^ {i}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.353, + 0.646, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} \\leftarrow \\mathbf {X} \\cup \\hat {\\mathbf {x}} ^ {i}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.593, + 0.369, + 0.905, + 0.382 + ], + "angle": 0, + "content": "\\(\\triangleright\\) insert generated reaction to reactor sequences" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.383, + 0.586, + 0.396 + ], + "angle": 0, + "content": "end for" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.398, + 0.593, + 0.411 + ], + "angle": 0, + "content": "return X" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.516 + ], + "angle": 0, + "content": "lying in a dynamically consistent manifold. Under the purpose of building the action-reaction controller in a physical world, we leverage the perpetual humanoid controller (PHC) [38], a universal physics tracker to track the generated reaction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.907, + 0.715 + ], + "angle": 0, + "content": "PHC follows the conventional pipeline of generative adversarial and goal-conditioned reinforcement learning [18, 34], it receives target joint positions or rigid body's keypoints as the policy goals and outputs physically retargeted motions. The policy consists of multiple actor-critic model in a multiplicative manner. Given arbitrary action-reaction generation, we formulate the auto-regressive interaction as a standard Markov Decision Process (MDP), denoted as tuple \\(\\langle S, A, \\mathcal{P}, r, \\gamma, \\rho_0 \\rangle\\), representing state space, action space, transition distribution, reward, discount factor and initial states' distribution. In our settings, the MDP is also conditioned on action-reaction tuple \\(\\hat{z}\\), the standard bellman equation of state-action value can be reformulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.725, + 0.906, + 0.764 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Q ^ {\\pi} (s, a, \\hat {z}) = \\mathbb {E} _ {s ^ {\\prime} \\sim \\mathcal {P} (\\cdot | s, a, \\hat {z})} [ r (s, a, s ^ {\\prime}, \\hat {z}) + \\tag {9} \\\\ \\left. \\gamma \\mathbb {E} _ {a ^ {\\prime} \\sim \\pi (\\cdot | s ^ {\\prime}, \\hat {z})} [ Q ^ {\\pi} (s ^ {\\prime}, a ^ {\\prime}, \\hat {z} ^ {\\prime} ] \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.775, + 0.907, + 0.837 + ], + "angle": 0, + "content": "where \\(\\hat{z} = \\langle \\hat{x},\\hat{y},y_{real}\\rangle\\) is the augmented interaction goal, \\(\\hat{x}\\) means a single predicted reactor frame regarded as the target goal, \\(\\hat{y}\\) and \\(y_{real}\\) are the actor motions coming from the diffusion model generation and real world inputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Actor-aware Reaction Policy To conserve computational and communication resources while prioritizing interaction safety, we integrate a low-frequency (0.5 Hz)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10177" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "diffusion planner with a high-frequency (30 Hz) physics tracker and real-world actor capture interface. This frequency mismatch can cause the reaction policy to pursue stale targets, elevating collision risk when the actor's motion deviates unexpectedly. To safeguard against interpenetration, we incorporate both generated and captured actor motion histories into the policy observation and heavily penalize the trajectory inconsistency in the reward function, thereby enforcing collision-free, safety-critical responses. More details of implementations in Appendix B.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.254, + 0.222, + 0.272 + ], + "angle": 0, + "content": "4. Applications" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.279, + 0.314, + 0.294 + ], + "angle": 0, + "content": "4.1. Immersive VR Interface" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.484, + 0.466 + ], + "angle": 0, + "content": "Our system is built upon Unity, enabling real-time interaction and visualization in a VR environment. Leveraging VisionPro [42], we develop a VR platform that allows the user to wear a headset and engage in immersive, egocentric interactions with the character. The camera in Unity is positioned on the actor's head, ensuring that it moves and rotates with the actor's motions. It continuously streams real-time simulations from Unity to the VR headset, providing an immersive experience. The user can perform any actions to guide the character's movement, such as completing a duet dance with the character." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.474, + 0.403, + 0.489 + ], + "angle": 0, + "content": "4.2. Human-Robot Interaction Interface" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.484, + 0.722 + ], + "angle": 0, + "content": "Benefiting from our robust Reaction Motion Tracker, the generated motions adhere to physical constraints and exhibit high adaptability. As a result, we can seamlessly transfer physically plausible humanoid interaction to human-robot interaction once we retarget the imitation goal from SMPL[37] joints to specific robot joints. An optimization-based method is leveraged following [15, 16, 24]. First we fit the SMPL shape parameters \\(\\beta\\) and use \\(\\beta\\) to solve for optimal robot joint angles \\(\\theta\\). Then the corresponding joints' goal positions are calculated by forward kinematics. In the VR environment, the character can be replaced with the Unitree H1[49], enabling an immersive and realistic human-robot interaction experience. Furthermore, our approach holds potential for future deployment on physical robots." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.734, + 0.222, + 0.75 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.376, + 0.773 + ], + "angle": 0, + "content": "5.1. Datasets and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.482, + 0.9 + ], + "angle": 0, + "content": "Datasets. We evaluate our model on two datasets: InterX [65], which contains 11,388 interaction SMPL-X [43] motion sequences on 40 interaction categories; InterHuman [32], which consists of 6022 interaction SMPL [37] motion sequences with around 1.7M frames. For data processing, we use the actor-reactor annotations provided by [66]. We canonicalize all the motions into the representation stated in Sec.3.3.1 and set frame rate to 30 FPS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Evaluation Metrics. We assess the generated reactor's sequence using three key aspects: 1) the quality of the reactor's motion independent of the actor, following kinematic motion generation works [14, 58]; 2) the physical plausibility of the reactor's motion, based on physics-based methods [59, 69]; and 3) the interaction quality between the actor and reactor, as discussed in [36, 52]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.907, + 0.44 + ], + "angle": 0, + "content": "For 1), we evaluate using Frechet Inception Distance (FID) [17] to measure the similarity between the generated and ground truth data distributions, Diversity to assess motion variability, and Multi-modal Distance (MMDist) to measure the distance between the generated reaction and the text prompt. For 2), we assess Penetration, the average distance between the ground and the lowest body joint position below the ground, Floating, the average distance between the ground and the lowest body joint above the ground, and Skating, the average horizontal displacement during frames exhibiting skating behavior. For 3), we measure maximum Interpenetration Volume (IV) [36] of reactor vertices intersecting the actor mesh, and cross-distance metrics \\( FID_{cd} \\) and \\( Div_{cd} \\) in a 100-dimensional cross-distance space [52] to evaluate interaction realism. Details of metrics calculation in Appendix C.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.454, + 0.731, + 0.469 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.733 + ], + "angle": 0, + "content": "In our auto-regressive denoiser, we adopt the DIT architecture [44] consisting of 8 Transformer layers with a 512-dimensional hidden size. We employ a frozen DistilBERT encoder [10, 51] as the text encoder, which is widely used in motion generation tasks [14, 59]. For training and inference, we set the binary interaction field \\(\\mathcal{I} \\in \\mathbb{R}^{6 \\times 6}\\), the diffusion timestep \\(T = 8\\), the classifier-free guidance weight \\(w = 5\\), and the foot-contact, interaction, and prefix loss weights to \\(\\lambda_{\\mathrm{foot}} = 0.2\\), \\(\\lambda_{\\mathrm{inter}} = 0.5\\), and \\(\\lambda_{\\mathrm{prefix}} = 0.1\\), respectively. In the auto-regressive setting, we use an interaction history length of \\(h = 20\\) frames and a prediction window length of \\(k = 40\\) frames. We train for 300K denoising steps on both the InterHuman and Inter-X datasets, and observe optimal checkpoints at 250K steps for InterHuman and 220K steps for Inter-X. All models are trained on a single NVIDIA GeForce RTX 4090d GPU for approximately 30 hours." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.901 + ], + "angle": 0, + "content": "For physics policy training, we use the multiplicative version of PHC network with three primitives, and construct task goal with only target body keypoints. The policy is also trained on the same device and datasets as the former stage for 20 hours. The tracking policy is executed at \\(30\\mathrm{Hz}\\) while physics simulation is processed in NVIDIA's Isaac Gym[41] at \\(60\\mathrm{Hz}\\). In the starting phase of inference, we capture the first 1 second motion from the actor, and store it in the buffer along with masked reactor motion prefix for the next planning horizon's generation, the downstream reactor tracker receives default joint positions as tracking goals." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10178" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.09, + 0.488, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.096, + 0.897, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.25, + 0.908, + 0.281 + ], + "angle": 0, + "content": "Figure 3. Compared to CAMDM (top row), Human-X (bottom row) achieves more complete hand contact in tasks such as face-hitting and handshaking. Additionally, its foot movement appears more natural, as highlighted in the red and green circles." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.291, + 0.891, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.408, + 0.908, + 0.451 + ], + "angle": 0, + "content": "Figure 4. Visualization results on Human-Robot Interaction. The robot (black skeleton) and human (orange mesh) perform a handshake on a flat plane, from arm extension and palm contact through to shake completion, illustrating our method's spatial coordination and motion coherence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.484, + 0.539 + ], + "angle": 0, + "content": "After warming-up, the diffusion planner generates reasonable action-reaction motion sequences, the reactor is able to continuously follow the planning and receives both real and generated actor sequences for policy correction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.551, + 0.353, + 0.567 + ], + "angle": 0, + "content": "5.3. Actor-To-Reactor Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.771 + ], + "angle": 0, + "content": "Baselines. To evaluate the performance of our method, we adopt and reproduce several human reaction generation models as baselines: 1) InterFormer [8]: a transformer-based model with temporal and spatial attention for human reaction generation, 2) InterGen [32]: a diffusion-based model for generating human interaction motions conditioned on text prompts., 3) ReGenNet [66]: a diffusion-based model for generating human reactions conditioned on given human actions. and 4) CAMDM [7]: a transformer-based Conditional Auto-regressive Motion Diffusion Model. However, since it is designed for single-person motion generation, we utilize the action as a condition input to generate the corresponding reaction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Comparisons. As demonstrated in Tab. 2, in the online unconstrained reaction setting, our method exhibits significant improvements over baselines in Reaction Quality, Physical Plausibility, and Interaction Quality on both the Inter-X datasets. Unlike InterFormer's transformer architecture, our diffusion-based approach effectively captures the nuanced structure of data distributions, resulting in re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.658 + ], + "angle": 0, + "content": "actor motions that closely align with real-world motion patterns and thereby enhancing reaction quality. In comparison to InterGen and ReGenNet, our implementation of Auto-regressive Motion Diffusion Models, augmented with a Physics Tracker, enables real-time interaction with the reactor by predicting future frames. This ensures that the generated motions adhere to fundamental physical laws, producing more natural and believable animations. Furthermore, relative to CAMDM, we introduce an Interaction Loss, leading to substantial improvements in IV and \\(\\mathrm{FID}_{cd}\\), thereby facilitating more natural interactions. The experiments on the InterHuman dataset can be found in C.1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.676, + 0.667, + 0.693 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.906, + 0.748 + ], + "angle": 0, + "content": "We primarily conduct ablation studies on three aspects: Losses, Timesteps, and Horizon. For more experiments, please refer to Section C.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Losses. To verify the effectiveness of our designed loss function, we ablate \\(\\mathcal{L}_{inter}\\), \\(\\mathcal{L}_{foot}\\), and \\(\\mathcal{L}_{prefix}\\) separately. As shown in Tab. 3, removing any of them degrades performance. Notably, excluding \\(\\mathcal{L}_{inter}\\) significantly increases \\(FID\\) and \\(FID_{cd}\\), underscoring its importance for generating realistic interactions. Removing \\(\\mathcal{L}_{foot}\\) raises Skating and Floating scores, showing its role in reducing foot sliding and enhancing physical plausibility." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10179" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.087, + 0.903, + 0.27 + ], + "angle": 0, + "content": "
MethodsReactionPhysicsInteraction
FID↓Div.→MMDist. ↓Pene. ↓Skate. ↓Float. ↓IV ↓FIDcd ↓Divcd →
Ground Truth0.0026.0283.5240.0000.0237.9560.0240.23511.471
InterFormer [8]6.1504.9228.6480.5240.17524.1090.3326.8258.214
InterGen [32]5.4854.4176.3890.2700.14311.9150.2693.4378.730
ReGenNet [66]2.2125.0475.2350.1410.1189.8100.2181.9509.248
CAMDM [7]1.4305.6575.3180.1530.1279.8640.1452.0869.017
Human-X0.9756.0634.1150.1180.0928.6500.0761.6929.735
Human-X*1.2205.1405.0760.0260.0074.3420.0541.8338.565
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.279, + 0.907, + 0.31 + ], + "angle": 0, + "content": "Table 2. Action-to-Reaction with online unconstrained reaction setting on Inter-X [65] dataset. A higher or lower value is better for ↑ or ↓, and → means the value closer to ground truth is better. * denotes the method with reaction policy designed for humanoid reaction." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.319, + 0.905, + 0.555 + ], + "angle": 0, + "content": "
ClassSettingsReactionPhysicsInteractionLatency (ms)
FID↓Div.→MMDist.→Penetration ↓Skating ↓Floating ↓IV ↓FIDcd ↓Divcd→
Ground Truth0.0026.0283.5240.0000.0237.9560.0240.23511.471-
Lossesw.o. Linter1.4576.1544.2640.1340.1078.7150.1192.2089.91713.6
w.o. Lfoot1.0096.1144.2050.1270.2399.1200.0941.6879.84913.6
w.o. Lprefix1.0236.1284.2250.1330.1138.8570.1001.6979.86513.6
Timesteps21.8095.5175.0070.1440.1139.0080.1532.0087.5133.8
51.1076.2634.5140.1280.1088.9080.1151.8078.0038.2
100.9886.0084.2080.1180.0968.7090.0971.6849.85316.9
1000.9085.9073.9630.1130.0938.6080.0871.6109.708170
Horizonh=10,k=301.0556.1064.1070.1230.1038.8050.1031.6879.8079.6
h=15,k=300.9876.0054.0060.1180.0978.7070.0981.6679.78710.1
h=20,k=200.9575.9563.9580.1130.0948.6580.0931.6269.7569.5
h=15,k=451.0066.0544.1050.1230.1048.8060.1021.6889.80913.6
Human-Xtext0.9265.9513.9090.1120.0878.2180.0921.6079.82213.6
unconstrained0.9756.0634.1150.1180.0928.6500.0761.6929.73513.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.261, + 0.563, + 0.735, + 0.578 + ], + "angle": 0, + "content": "Table 3. Ablation studies of online reaction setting on the Inter-X [65] dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.483, + 0.741 + ], + "angle": 0, + "content": "Timesteps. During the denoising process, we experiment with different sampling timesteps: 2, 5, 10, and 100. As shown in the Tab. 3, increasing the number of timesteps improves the model's performance; however, it also leads to higher latency. When timesteps = 10, the model has already achieved satisfactory results. Considering both generation quality and latency, we ultimately select timesteps = 8 as a trade-off, ensuring high-quality generation while meeting real-time requirements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.483, + 0.834 + ], + "angle": 0, + "content": "Horizon. During inference process, the model takes a continuous sequence of \\( h \\) history frames as input to predict \\( k \\) future frames. We conduct experiments with different values of \\( h \\) and \\( k \\), and the results in Tab. 3 show that the best performance is achieved when \\( h = 20 \\) and \\( k = 40 \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.846, + 0.21, + 0.861 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We present Human-X Interaction, a novel framework for real-time human interaction synthesis that seamlessly in-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.907, + 0.892 + ], + "angle": 0, + "content": "tegrates autoregressive diffusion modeling with an actor-aware reaction policy. Our approach addresses the critical challenges of real-time synchronization, physical realism, and safety in human-ayar, human-humanoid and humanrobot interactions. Specifically, Human-X jointly predicts actions and reactions using both partner's motion histories, ensuring immersive and realistic motion synthesis. By incorporating comprehensive interaction-aware rewards and training the tracking policy with reinforcement learning, we achieve physically feasible and artifact-free interactions even in dynamic and safety-critical scenarios. Extensive experiments on the Inter-X and InterHuman datasets demonstrate significant improvements in motion quality, interaction continuity, and physical accuracy over existing methods. The current formulation focuses on synthesizing reactive motions and does not account for collaborative task completion between humans and machines. Extending the framework to support joint task execution with shared goals remains an open challenge." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10180" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.271, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.484, + 0.269 + ], + "angle": 0, + "content": "This work was supported by Shanghai Local College Capacity Building Program(23010503100), NSFC (No.62406195), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (Shanghai), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University), HPC Platform and Core Facility Platform of Computer Science and Communication of ShanghaiTech University and Shanghai Engineering Research Center of Intelligent Vision and Imaging." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.286, + 0.188, + 0.303 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.312, + 0.484, + 0.379 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.383, + 0.484, + 0.438 + ], + "angle": 0, + "content": "[2] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.442, + 0.483, + 0.497 + ], + "angle": 0, + "content": "[3] Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems, 28, 2015. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.5, + 0.484, + 0.582 + ], + "angle": 0, + "content": "[4] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, et al. Digital life project: Autonomous 3d characters with social intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 582-592, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.585, + 0.484, + 0.655 + ], + "angle": 0, + "content": "[5] Haoran Chen, Yiteng Xu, Yiming Ren, Yaoqin Ye, Xinran Li, Ning Ding, Peishan Cong, Ziyi Wang, Bushi Liu, Yuhan Chen, et al. Symbiosim: Human-in-the-loop simulation platform for bidirectional continuing learning in human-robot interaction. arXiv preprint arXiv:2502.07358, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.657, + 0.484, + 0.726 + ], + "angle": 0, + "content": "[6] Jingxiao Chen, Xinyao Li, Jiahang Cao, Zhengbang Zhu, Wentao Dong, Minghuan Liu, Ying Wen, Yong Yu, Liqing Zhang, and Weinan Zhang. Rhino: Learning real-time humanoid-human-object interaction from human demonstrations. arXiv preprint arXiv:2502.13134, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.729, + 0.484, + 0.785 + ], + "angle": 0, + "content": "[7] Rui Chen, Mingyi Shi, Shaoli Huang, Ping Tan, Taku Komura, and Xuelin Chen. Taming diffusion probabilistic models for character control. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 7, 8, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.484, + 0.842 + ], + "angle": 0, + "content": "[8] Baptiste Chopin, Hao Tang, Naima Otberdout, Mohamed Daoudi, and Nicu Sebe. Interaction transformer for human reaction generation. IEEE Transactions on Multimedia, 25: 8842-8854, 2023. 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[9] Zekai Deng, Ye Shi, Kaiyang Ji, Lan Xu, Shaoli Huang, and Jingya Wang. Human-object interaction via automatically designed vlm-guided motion policy. arXiv preprint arXiv:2503.18349, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.312, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.188 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.906, + 0.274 + ], + "angle": 0, + "content": "[11] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm- e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning, pages 8469-8488, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.359 + ], + "angle": 0, + "content": "[12] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.906, + 0.416 + ], + "angle": 0, + "content": "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.906, + 0.487 + ], + "angle": 0, + "content": "[14] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5152-5161, 2022. 4, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.906, + 0.558 + ], + "angle": 0, + "content": "[15] Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. arXiv preprint arXiv:2406.08858, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.906, + 0.601 + ], + "angle": 0, + "content": "[16] Tairan He, Zhengyi Luo, Wenli Xiao, Chong Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Learning human-to-humanoid real-time whole-body teleoperation, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.906, + 0.671 + ], + "angle": 0, + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.906, + 0.714 + ], + "angle": 0, + "content": "[18] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10181" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Vision and Pattern Recognition, pages 16750-16761, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[23] Muhammad Gohar Javed, Chuan Guo, Li Cheng, and Xingyu Li. Intermask: 3d human interaction generation via collaborative masked modelling. arXiv preprint arXiv:2410.10010, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.235 + ], + "angle": 0, + "content": "[24] Mazeyu Ji, Xuanbin Peng, Fangchen Liu, Jialong Li, Ge Yang, Xuxin Cheng, and Xiaolong Wang. Exbody2: Advanced expressive humanoid whole-body control. arXiv preprint arXiv:2412.13196, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.238, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[25] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.378 + ], + "angle": 0, + "content": "[26] Jianping Jiang, Weiye Xiao, Zhengyu Lin, Huaizhong Zhang, Tianxiang Ren, Yang Gao, Zhiqian Lin, Zhongang Cai, Lei Yang, and Ziwei Liu. Solami: Social vision-language-action modeling for immersive interaction with 3d autonomous characters. arXiv preprint arXiv:2412.00174, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.381, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[27] Wentao Jiang, Jingya Wang, Kaiyang Ji, Baoxiong Jia, Siyuan Huang, and Ye Shi. Arflow: Human action-reaction flow matching with physical guidance. arXiv preprint arXiv:2503.16973, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.439, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[28] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[29] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[30] Jiefeng Li, Siyuan Bian, Chao Xu, Zhicun Chen, Lixin Yang, and Cewu Lu. Hybrid analytical-neural inverse kinematics for whole-body mesh recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.584, + 0.482, + 0.653 + ], + "angle": 0, + "content": "[31] Ronghui Li, Youliang Zhang, Yachao Zhang, Yuxiang Zhang, Mingyang Su, Jie Guo, Ziwei Liu, Yebin Liu, and Xiu Li. Interdance: Reactive 3d dance generation with realistic duet interactions. arXiv preprint arXiv:2412.16982, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.656, + 0.483, + 0.712 + ], + "angle": 0, + "content": "[32] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 2, 6, 7, 8, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.714, + 0.483, + 0.769 + ], + "angle": 0, + "content": "[33] Jing Lin, Yao Feng, Weiyang Liu, and Michael J Black. Chathuman: Language-driven 3d human understanding with retrieval-augmented tool reasoning. arXiv preprint arXiv:2405.04533, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.483, + 0.812 + ], + "angle": 0, + "content": "[34] Minghuan Liu, Menghui Zhu, and Weinan Zhang. Goal-conditioned reinforcement learning: Problems and solutions, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.814, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[35] Yunze Liu, Changxi Chen, and Li Yi. Interactive humanoid: Online full-body motion reaction synthesis with social affordance canonicalization and forecasting. arXiv preprint arXiv:2312.08983, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-reaction: Physically plausible real-time humanoid reaction" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "synthesis via forward dynamics guided 4d imitation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3771-3780, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[38] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10895-10904, 2023. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[39] Zhengyi Luo, Jinkun Cao, Josh Merel, Alexander Winkler, Jing Huang, Kris Kitani, and Weipeng Xu. Universal humanoid motion representations for physics-based control. arXiv preprint arXiv:2310.04582, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.907, + 0.371 + ], + "angle": 0, + "content": "[40] Zhengyi Luo, Jinkun Cao, Rawal Khirodkar, Alexander Winkler, Kris Kitani, and Weipeng Xu. Real-time simulated avatar from head-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 571-581, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[41] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[42] Younghyo Park and Pulkit Agrawal. Using apple vision pro to train and control robots, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[44] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[45] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4):1-14, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[46] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[48] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.901, + 0.901 + ], + "angle": 0, + "content": "[49] Unitree Robotics. Unitree h1 robot. https://www.unitree.com, 2021. Accessed: 2023-10-10. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10182" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[50] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[51] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[52] Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. arXiv preprint arXiv:2403.18811, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[54] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39 (4):54-1, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[55] Wenhui Tan, Boyuan Li, Chuhao Jin, Wenbing Huang, Xiting Wang, and Ruihua Song. Think then react: Towards unconstrained action-to-reaction motion generation. In The Thirteenth International Conference on Learning Representations, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[56] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15999-16009, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[57] Jiangnan Tang, Jingya Wang, Kaiyang Ji, Lan Xu, Jingyi Yu, and Ye Shi. A unified diffusion framework for scene-aware human motion estimation from sparse signals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21251-21262, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[58] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[59] Guy Tevet, Sigal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. Closd: Closing the loop between simulation and diffusion for multi-task character control. arXiv preprint arXiv:2410.03441, 2024. 4, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[60] Jie Tian, Ran Ji, Lingxiao Yang, Suting Ni, Yuexin Ma, Lan Xu, Jingyi Yu, Ye Shi, and Jingya Wang. Gaze-guided hand-object interaction synthesis: Dataset and method. arXiv preprint arXiv:2403.16169, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[61] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[62] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[63] Zhen Wu, Jiaman Li, and C Karen Liu. Human-object interaction from human-level instructions. CoRR, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[64] Liang Xu, Ziyang Song, Dongliang Wang, Jing Su, Zhicheng Fang, Chenjing Ding, Weihao Gan, Yichao Yan, Xin Jin, Xiaokang Yang, et al. Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[65] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter-x: Towards versatile human-human interaction analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22260-22271, 2024. 2, 6, 8, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.907, + 0.328 + ], + "angle": 0, + "content": "[66] Liang Xu, Yizhou Zhou, Yichao Yan, Xin Jin, Wenhan Zhu, Fengyun Rao, Xiaokang Yang, and Wenjun Zeng. Regennet: Towards human action-reaction synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1759-1769, 2024. 2, 6, 7, 8, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[67] Lingxiao Yang, Shutong Ding, Yifan Cai, Jingyi Yu, Jingya Wang, and Ye Shi. Guidance with spherical gaussian constraint for conditional diffusion. In Proceedings of the 41st International Conference on Machine Learning, pages 56071-56095, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[68] Ye Yuan, Viktor Makoviychuk, Y Guo, S Fidler, XB Peng, and K Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Trans. Graph, 42(4), 2023. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.522 + ], + "angle": 0, + "content": "[69] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16010-16021, 2023. 3, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[70] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[71] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE transactions on pattern analysis and machine intelligence, 46(6):4115-4128, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[72] Zhenhao Zhang, Ye Shi, Lingxiao Yang, Suting Ni, Qi Ye, and Jingya Wang. Openhoi: Open-world hand-object interaction synthesis with multimodal large language model. arXiv preprint arXiv:2505.18947, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[73] Kaifeng Zhao, Gen Li, and Siyu Tang. Dart: A diffusion-based autoregressive motion model for real-time text-driven motion control. arXiv preprint arXiv:2410.05260, 2024. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[74] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[75] Kaizhen Zhu, Mokai Pan, Yuexin Ma, Yanwei Fu, Jingyi Yu, Jingya Wang, and Ye Shi. Unidb: A unified diffusion bridge framework via stochastic optimal control. arXiv preprint arXiv:2502.05749, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10183" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/4564e033-7cc9-41ba-b940-20704a40b38a_origin.pdf b/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/4564e033-7cc9-41ba-b940-20704a40b38a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e378c32fd23bc3dade5fef1379d0e57672bb7e48 --- /dev/null +++ b/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/4564e033-7cc9-41ba-b940-20704a40b38a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:beb578609f8c2900ec1d5505b55143d8b32c475fa130067e73626b964285c989 +size 12440083 diff --git a/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/full.md b/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..63e99ca62e2f708475a6f55bcd9a9c6cf0a6e8c9 --- /dev/null +++ b/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/full.md @@ -0,0 +1,371 @@ +# Towards Immersive Human-X Interaction: A Real-Time Framework for Physically Plausible Motion Synthesis + +Kaiyang Ji $^{1}$ Ye Shi $^{1,2,*}$ Zichen Jin $^{1}$ Kangyi Chen $^{1}$ Lan Xu $^{1,2}$ Yuexin Ma $^{1,2}$ Jingyi Yu $^{1,2}$ Jingya Wang $^{1,2,*}$ + +$^{1}$ ShanghaiTech University + +$^{2}$ Shanghai Engineering Research Center of Intelligent Vision and Imaging + +{jiky2024,shiye,jinzch12023,v-chenky,xulan1,mayuexin,yujingyi,wangjingya}@shanghaiitech.edu.cn + +![](images/f77c7108da04d1983bd50f0fbc1b3c5ee4539371662cb0c5a1ef5c6c0123245a.jpg) +Figure 1. We propose Human-X, the first framework designed to enable latency-free interaction between humans and diverse entities, including human-athlete, human-humanoid, and human-robot interaction. + +# Abstract + +Real-time synthesis of physically plausible human interactions remains a critical challenge for immersive VR/AR systems and humanoid robotics. While existing methods demonstrate progress in kinematic motion generation, they often fail to address the fundamental tension between real-time responsiveness, physical feasibility, and safety requirements in dynamic human-machine interactions. We introduce Human-X, a novel framework designed to enable immersive and physically plausible human interactions across diverse entities, including human-ayar, human-humanoid, and human-robot systems. Unlike existing approaches that focus on post-hoc alignment or simplified physics, our method jointly predicts actions and reactions in real-time using an auto-regressive reaction diffusion planner, ensuring seamless synchronization and context-aware responses. To enhance physical realism and safety, we integrate an actor-aware motion tracking policy trained with reinforcement learning, which dynamically adapts to interaction partners' movements while avoiding artifacts like foot sliding and penetration. Extensive experiments on the Inter-X and InterHuman datasets demonstrate significant improvements in motion quality, interaction continuity, and physical + +plausibility over state-of-the-art methods. Our framework is validated in real-world applications, including virtual reality interface for human-robot interaction, showcasing its potential for advancing human-robot collaboration. Project page: https://humanx-interaction.github.io/ + +# 1. Introduction + +Human-human interaction synthesis has long been a focal point in the field of motion generation, with widespread applications in social dynamics, gaming, and animation [26, 54]. With the rapid development of virtual reality (VR) and augmented reality (AR), these technologies now empower the creation of deeply immersive and contextually rich interactions between humans and digital avatars. Particularly, the emergence of humanoid robots further bridges the gap between virtual and physical realms, enabling robots to engage with humans in daily life through adaptive and socially aware behaviors. However, achieving real-time prediction of diverse and safety-guaranteed humanoid reactions, an essential capability for applications ranging from daily companionship to assistive care, remains a critical, unsolved challenge. + +Current approaches to human interaction modeling rely + +
Methodsphysics-basedimmersive inferencelatency-freereal-time modelgeneration abilitystrong contacttext-control
InterGen [32]✓✓✓
ReGenNet [66]✓✓✓✓
SoLaMi [26]✓✓✓
PhysReaction [36]✓✓✓✓
Think-Then-React [55]✓✓✓✓
Human-X✓✓✓✓✓✓✓
+ +Table 1. Comparative analysis of Human-X versus existing human reaction synthesis approaches. + +heavily on offline training data or simplified physics assumptions, leading to limited adaptability in dynamic scenarios. While data-driven methods such as [23, 32, 65] have shown promise in synthesizing kinematic motions interaction, they often fail to account for the intricate physical constraints and safety requirements inherent in human-humanoid cohabitation. For instance, abrupt or unstable motions generated through these techniques pose risks in shared environments. This gap underscores the need for a unified interaction framework that integrates real-time reactivity, physical realism, and safety assurance. + +In this work, we explore Human-X Interaction (Fig.1), a systematic study of immersive interactions across three domains: 1) Human-Avatar Interaction using kinematic-driven motion in VR/AR; 2) Human-Humanoid Interaction with physics-based motion synthesis; 3) Human-Robot Interaction ensuring safe physical contact. Unlike existing physics-aware methods limited to replicating training data, our framework balances natural movement with physical feasibility while maintaining real-time performance. This integration enables safer and more natural human-machine collaboration across applications from entertainment to assistive robotics. Achieving real-time synchronization between interacting partners faces major technical challenges. Conventional methods rely exclusively on past actor motions to generate reactions, leading to temporal mismatches. Furthermore, ensuring physical realism is more challenging, as existing solutions focus mainly on basic collision avoidance and neglect dynamic interaction forces. + +To overcome these limitations, we present Human-X, an Auto-Regressive Action-Reaction Diffusion Planner for real-time interaction. Our framework jointly models both partners' motion histories to predict subsequent actions and reactions. We introduce multiple contact-aware losses to train realistic interactions, and combine the diffusion model with an actor-aware tracking policy to ensure physical feasibility in humanoid/robot scenarios. In summary, our contributions can be summarized as follows: + +- The first auto-regressive diffusion model that jointly predicts actions and reactions using both partners' motion history with comprehensive contract losses to ensure real-time, immersive and realistic motion synthesis. +- An actor-aware reaction policy trained with reinforce + +ment learning that dynamically adapts to interaction partners' motions, preventing common artifacts like foot sliding while maintaining physical feasibility. + +- A complete real-time VR system demonstrating practical effectiveness. Experiments on Inter-X and InterHuman datasets show significant improvements in motion quality, interaction continuity, and physical accuracy compared to existing methods. + +# 2. Related Works + +# 2.1. Human Motion Generation + +Human-centric motion generation has seen remarkable advancements in recent years, driven by breakthroughs in large language models [1, 25] and generative frameworks [13, 20, 28, 67, 75]. Notably, the introduction of motion diffusion models [58, 71] has significantly elevated the performance of single-person motion generation. Building on these developments, researchers have increasingly focused on enriching interactions with the environment by incorporating diverse modalities, such as text [25, 62, 63, 70], audio [2, 61], and scene information [12, 22, 29, 57]. For instance, recent works [9, 60, 62, 63, 72] have leveraged text-based guidance to facilitate Human-Object Interactions (HOI), while MotionGPT [25] treats motion as a form of language, enabling seamless bidirectional translation between text and motion. Despite these advancements, existing approaches face a critical limitation: their model architectures are inherently ill-suited for real-time reaction synthesis. + +# 2.2. Human-Human Interaction + +Research on human-human interaction synthesis has evolved around three primary directions: (1) Interactive Behavior Modeling leverages actor-reactor role decomposition to capture interaction dynamics. For example, [56] identifies role-specific patterns, while [32, 64] use transformer architectures to model multi-agent dependencies. (2) Multimodal Interaction integrates text, audio, or scene context to enhance realism [31, 52], but often overlook dynamic physical constraints critical for real-world deployment. (3) Reactive Motion Generation employs sequence models [35] and diffusion models [27, 66] to synthesize temporally coherent reactions. These methods, however, struggle with physical + +![](images/9d3fce72a20fd83d348ef6c4fd7a584078f361a8fc1f08ae40900361a5995903.jpg) +Figure 2. Overview of our immersive real-time interaction synthesis pipeline: (a) Actor Motion Capture: A human actor's movements are recorded at 30 fps by an RGB-D camera and translated into 3D poses, which are then retargeted to a humanoid character. (b) Realistic Reactor Motion Generation: An auto-regressive diffusion model, guided by optional text prompts (e.g., "Dancing is what to do"), generates plausible reaction motions. These motions are tracked by an actor-aware controller, which uses proprioception signals to ensure realistic, synchronized interactions. (c) Real-time VR Interface: The generated and tracked motions are rendered in simulator, providing both a third-person view and a binocular VR view. + +plausibility and temporal consistency in long-horizon interactions. Existing Reactive Generation methods face a fundamental tension: Human interaction models achieve reactivity but lack physical grounding, while physics-based animations enforce plausibility but are agnostic to interaction dynamics. While PhysReaction [36] introduced the first physics-based approach for reaction synthesis using imitation learning, it was limited to mimicking known distributions, lacking generalization ability, diversity, controllability, and the capacity to handle unseen tasks. Our work bridges this gap by unifying interaction-aware reactivity with physically constrained control. + +# 2.3. Physics-based Motion Synthesis + +In contrast to kinematics-based methods, physics-based motion synthesis has emerged as a widely adopted approach for generating realistic and physically plausible character movements by integrating physical principles into the animation process. Early work, such as DeepMimic [45], leveraged deep reinforcement learning to enable characters to imitate diverse motion clips within a physics-based framework. Subsequent advancements [18, 46] utilized generative adversarial networks to enhance stylized motion realism. Progress continued with PHC [38] and PULSE [39], focusing on improving the diversity of generated motions. A significant milestone was achieved with PhysDiff [69], which introduced physics-guided diffusion models to mitigate artifacts such as floating and foot sliding. Despite + +these advancements, the physical plausibility of human-human interactions remains largely unexplored. + +# 3. Method + +# 3.1. System Overview + +In this section, we present a detailed overview of our immersive interaction synthesis system. To facilitate realistic and responsive interactions between two agents in immersive environments, our system takes online actor motion capture data as input, processes it through an advanced motion generation pipeline, and outputs continuous, reactive motion in real-time, as shown in Fig.2. The first key component of our system, detailed in Sec.3.2, is the real-time capture of the actor's pose and the reconstruction of its state in physics simulation. Subsequently, the Auto-regressive Interaction Diffusion Model, as described in Sec.3.3, is employed to generate the reaction of one agent in response to the actor context. Next, we introduce the Reaction Motion Tracker in Sec.3.4, which ensures that the motion of the reactor adheres to physical constraints for realism and natural behavior. Finally, we explain how VR Interface works in Sec.4.1, which enables immersive, real-time interaction with the generated motions. + +# 3.2. Human Pose Estimation and Reconstruction + +We firstly utilize a real-time RGB-D camera to capture images. Then, HybrIK [30] is employed for latency-free re + +construction of the human pose. However, due to the inaccuracies in the depth information provided by HybrIK, the depth map is directly obtained from the RGB-D camera instead. Following [38, 40], we bridge SMPL's kinematic skeleton to physics simulation via applying a simple retargeting method. We utilize Isaac Gym [41] as our physics simulation platform to create a third-person perspective for the interaction between the actor and reactor. Besides, the image reading process, motion generation of the reactor, and the physical simulation within Isaac Gym are handled by separate processes. ROS2 is used as the communication framework to facilitate real-time data transmission. + +# 3.3. Online Human Action-Reaction Synthesis + +The goal of online action-reaction synthesis is to generate realistic and continuous future $k$ frames of reactor motions $\mathbf{x}^i \coloneqq \{x^{n:n + k - 1}\}$ given $h$ history frames of interaction contexts $\{\mathbf{x}^{i - 1},\mathbf{y}^{i - 1}\} \coloneqq \{x^{n - h:n - 1},y^{n - h:n - 1}\}$ at current frame number $n$ and current window index $i$ . During each window prediction, an optional text prompt $c$ can be added to guide the reaction generation process. + +# 3.3.1. Reactor-centric Interaction Modeling + +Reactor-centric Canonicalization. For individual motion synthesis, the initial frame of each sequence is typically aligned at the root to improve the generalizability of the model [14, 59]. However, using the same normalization for human-human interaction data can eliminate key details about the mutual spatial relationship. To retain relative positioning without losing generalization ability, we leverage a reactor-centric canonicalization that the pose of reactor is realigned at the root and oriented toward the forward direction, while the pose of actor is adjusted by the same transformation. This method maintains the essential spatial context between the two, enabling the model to accurately learn their interdependent movement dynamics. + +Reactor-centric Interaction Representation. We model each frame of interaction as a hybrid representation of both actor and reactor motion. For reaction representation, we follow previous works [14, 58] and use the HumanML3D representation. The reactor motion $x$ in frame $n$ is defined by $x^{n} = (r^{x},\dot{r}^{x},p^{x},\dot{p}^{x},\theta^{x},f^{x})$ , where $r^x\in \mathbb{R}$ denotes the height of the roots, $\dot{r}^x\in \mathbb{R}^3$ denotes the angular velocity of the roots along the Y axis and the linear velocity of the roots in the XZ plane, $p^x\in \mathbb{R}^{(J - 1)\times 3}$ are the local joint positions, $\dot{p}^x\in \mathbb{R}^{J\times 3}$ denotes the temporal difference of the local joint positions, $\theta^x\in \mathbb{R}^{(J - 1)\times 6}$ is the 6D representation [74] of the joint rotations and $f^x\in \mathbb{R}^4$ are binary labels for the detection of contact. For actor motion captured from pose estimation, we define the actor $y$ in frame $n$ by $y^n = (r^{x\to y},\dot{r}^{x\to y},p^{x\to y},\dot{p}^{x\to y},f^y)$ where all the features are in the reactor's reference frame, making it a pure + +relative representation consisting of relative root transition, relative root rotation, relative root linear velocity, relative joint positions, relative joint velocity and feet contact labels. To emphasize actor-reactor contacts, we introduce the binary interaction field $\mathcal{I}^n\in \mathbb{R}^{\bar{J}\times \bar{J}}$ obtained by thresholding the relative distance between the selected $\bar{J}$ joints of each agent in a contact map. For body model selection, we follow the skeleton of SMPL [37] with $J = 22$ joints. The total reactor-centric interaction $\mathbf{z}$ at frame $n$ representation can be formed as, + +$$ +z ^ {n} = \left(x ^ {n}, y ^ {n}, \mathcal {I} ^ {n}\right), \tag {1} +$$ + +with $x^n$ denotes the reactor motion, $y^n$ denotes the relative actor motion and $\mathcal{I}^n$ is the joint contact features. Thus, we can model the interaction sequence as $\mathbf{z}^i = \{z^{n:n + k - 1}\}$ given $k$ frames and window index $i$ . + +# 3.3.2. Auto-regressive Interaction Diffusion Planner + +Based on the interaction motion representation provided in Sec.3.3.1, we transfer raw motion data into a canonical space suitable for generative learning with denoising networks. For probabilistic generator, we follow the DDPM [20] framework by incrementally adding Gaussian noise to the clean data sample $z_0$ over $T$ steps: + +$$ +q \left(\mathbf {z} _ {t} \mid \mathbf {z} _ {t - 1}\right) = \mathcal {N} \left(\sqrt {\alpha_ {t}} \mathbf {z} _ {t - 1}, (1 - \alpha_ {t}) \mathbf {I}\right), \tag {2} +$$ + +where $\alpha_{t} \in (0,1)$ are constant parameters for variance scheduling. Given the noised data $\mathbf{z}_t^i$ at the $i$ -th window clip, the diffusion step $t$ , the prefix interaction window $\mathbf{z}^{i-1} = z^{n-h:n-1}$ , and the optional text prompt $c$ , we train a transformer-based denoiser $\mathcal{G}$ to predict the cleaned interaction primitive itself instead of predicting the noise [58]: + +$$ +\tilde {\mathbf {z}} _ {0} ^ {i} = \mathcal {G} \left(\mathbf {z} _ {t} ^ {i}, \mathbf {z} ^ {i - 1}, t, c\right), \tag {3} +$$ + +where the text prompt $c$ is randomly masked by a rate of 0.15 during training to enable unconditioned generation and classifier-free guidance [19]. The predicted motion $\hat{\mathbf{z}}_0^i$ is then inserted to the interaction history for next interaction window extraction after the diffusion sampling loop [20, 53]. + +Losses The denoised motion $\hat{\mathbf{z}}_0^i$ is supervised by a simple objective in DDPM [20]: + +$$ +\mathcal {L} _ {\text {s i m p l e}} = \mathbb {E} _ {\mathbf {z} _ {0} ^ {i} \sim q \left(\mathbf {z} _ {0} ^ {i}\right), t \sim [ 1, T ]} \left[ \| \mathbf {z} _ {0} ^ {i} - \hat {\mathbf {z}} _ {0} ^ {i} \| _ {2} ^ {2} \right]. \tag {4} +$$ + +We extend the training objective with several auxiliary losses to enhance interaction quality and realism. To optimize foot contact quality, we add a foot contact loss on both agents: + +$$ +\mathcal {L} _ {\text {f o o t}} = \sum_ {j \in \{x ^ {\text {f e e t}}, y ^ {\text {f e e t}} \}} \| \dot {p} ^ {j} \left(\hat {\mathbf {z}} _ {0} ^ {i}\right) \odot f ^ {j} \left(\hat {\mathbf {z}} _ {0} ^ {i}\right) \| _ {2} ^ {2}, \tag {5} +$$ + +where $\dot{p}^j (\hat{\mathbf{z}}_0^i)$ is the corresponding foot joints velocity, $f^{j}(\hat{\mathbf{z}}_{0}^{i})$ are foot contact label described in Sec 3.3.1 and $\odot$ denotes for Hadamard product. Note that $\dot{p}^j (\hat{\mathbf{z}}_0^i)$ is defined by the relative local velocity in the reactor frame, and we transform it to the global velocity in loss calculation. For interaction refinement, we design an interaction loss on contact joints distance: + +$$ +\mathcal {L} _ {i n t e r} = \sum_ {x \in \bar {J} ^ {x}, y \in \bar {J} ^ {y}} \| (p ^ {x} (\hat {\mathbf {z}} _ {0} ^ {i}) - p ^ {y} (\hat {\mathbf {z}} _ {0} ^ {i})) \odot \mathcal {I} ^ {x, y} (\hat {\mathbf {z}} _ {0} ^ {i}) \| _ {2} ^ {2}, (6) +$$ + +where $\mathcal{I}^{x,y}(\hat{\mathbf{z}}_0^i)$ are binary body interaction fields and $p^x (\hat{\mathbf{z}}_0^i), p^y (\hat{\mathbf{z}}_0^i)$ are contact joints positions in the same root frame to enhance strong interaction. To improve the continuity of auto-regressive generation, we finally add a prefix loss $\mathcal{L}_{prefix}$ to evaluate the difference between the end of prefix motion and the beginning of the generated motion. The full loss applied to the model is then: + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} _ {\text {s i m p l e}} + \lambda_ {\text {f o o t}} \mathcal {L} _ {\text {f o o t}} + \lambda_ {\text {i n t e r}} \mathcal {L} _ {\text {i n t e r}} \tag {7} \\ + \lambda_ {p r e f i x} \mathcal {L} _ {p r e f i x}, \\ \end{array} +$$ + +where $\lambda_{foot}$ , $\lambda_{inter}$ and $\lambda_{prefix}$ are hyper-parameters to balance the loss terms from different scaling aspects. + +Training During training, we utilize the scheduled training strategy [3, 48, 68, 73] to enhance auto-regressive generation stability and text prompt controllability. The motion is cropped to sequences of $N$ consecutive interaction windows with a random offset from the dataset. During training, $\mathbf{z}_0^i$ is noised to $\mathbf{z}_t^i$ at diffusion timestep $t$ , and interaction history $\mathbf{z}^{i-1}$ is divide into 3 stages with the sample probability $p$ : ground truth for supervised learning $(p = 1)$ , random sampling for scheduled training (decaying $p$ ), and auto-regressive prediction $(p = 0)$ . Details in Appendix B.1. + +Inference After training a denoise network $\mathcal{G}$ , we can auto-regressively generate reaction sequences given interaction history $\mathbf{z}^{i-1}$ , a diffusion sampler $\mathcal{S}$ like [20, 53] and online text guidance $c$ by sampling with classifier-free guidance [19]: + +$$ +\begin{array}{l} \mathcal {G} _ {w} \left(\mathbf {z} _ {t} ^ {i}, \mathbf {z} ^ {i - 1}, t, c\right) = \mathcal {G} \left(\mathbf {z} _ {t} ^ {i}, \mathbf {z} ^ {i - 1}, t, \varnothing\right) \\ + \quad w \cdot \left(\mathcal {G} \left(\mathbf {z} _ {t} ^ {i}, \mathbf {z} ^ {i - 1}, t, c\right) - \mathcal {G} \left(\mathbf {z} _ {t} ^ {i}, \mathbf {z} ^ {i - 1}, t, \varnothing\right)\right), \tag {8} \\ \end{array} +$$ + +where $w$ is the guidance hyper-parameter, as shown in Alg.1. + +# 3.4. Reaction Policy + +Our actor-motion data captured from virtual reality or existing datasets roughly approximates the dynamic plausibility. In contrast, the generated reactor motions is not always + +Algorithm 1 Online reaction sampling using auto-regressive motion diffusion model + +Input: denoiser $\mathcal{G}$ , online actor loader $\mathbf{Y}$ , online text loader $\mathbf{C}$ , total diffusion steps $T$ , classifier-free guidance scale $w$ , auto-regressive reaction sampler $S$ . + +Output: reactor motion sequence $\mathbf{X}$ + +$$ +\begin{array}{l} \mathbf {z} _ {T} ^ {0} \leftarrow \mathcal {N} (0, 1) \\ \mathbf {X} \leftarrow \mathcal {S} (\mathcal {G}, \mathbf {z} _ {T} ^ {0}, T, \varnothing , \varnothing , w) \\ \end{array} +$$ + +$\triangleright$ init with unconditioned sampling + +for $i\gets 1$ to N do + +$\triangleright$ num of actor windows + +$$ +\mathbf {y} ^ {i - 1} \leftarrow \operatorname {C o l l e c t} (\mathbf {Y}), \mathbf {x} ^ {i - 1} \leftarrow \operatorname {C o l l e c t} (\mathbf {X}) +$$ + +$$ +\mathbf {z} ^ {i - 1} \leftarrow \operatorname {C A N O N I C A L I Z E} \left(\mathbf {x} ^ {i - 1}, \mathbf {y} ^ {i - 1}\right) +$$ + +> get interaction history + +$$ +\mathbf {z} _ {T} ^ {i} \leftarrow \mathcal {N} (0, 1) +$$ + +$$ +c ^ {i} \leftarrow \operatorname {C O L L E C T} (\mathbf {C}) +$$ + +$$ +\hat {\mathbf {z}} _ {0} ^ {i} \leftarrow S (\mathcal {G}, \mathbf {z} _ {T} ^ {i}, T, \mathbf {z} ^ {i - 1}, c ^ {i}, w) \triangleright C F G \text {s a m p l i n g l o o p} +$$ + +$$ +\hat {\mathbf {x}} ^ {i} \leftarrow \operatorname {R e c o v e r} \left(\hat {\mathbf {z}} _ {0} ^ {i}\right) +$$ + +$$ +\mathbf {X} \leftarrow \mathbf {X} \cup \hat {\mathbf {x}} ^ {i} +$$ + +$\triangleright$ insert generated reaction to reactor sequences + +end for + +return X + +lying in a dynamically consistent manifold. Under the purpose of building the action-reaction controller in a physical world, we leverage the perpetual humanoid controller (PHC) [38], a universal physics tracker to track the generated reaction. + +PHC follows the conventional pipeline of generative adversarial and goal-conditioned reinforcement learning [18, 34], it receives target joint positions or rigid body's keypoints as the policy goals and outputs physically retargeted motions. The policy consists of multiple actor-critic model in a multiplicative manner. Given arbitrary action-reaction generation, we formulate the auto-regressive interaction as a standard Markov Decision Process (MDP), denoted as tuple $\langle S, A, \mathcal{P}, r, \gamma, \rho_0 \rangle$ , representing state space, action space, transition distribution, reward, discount factor and initial states' distribution. In our settings, the MDP is also conditioned on action-reaction tuple $\hat{z}$ , the standard bellman equation of state-action value can be reformulated as: + +$$ +\begin{array}{l} Q ^ {\pi} (s, a, \hat {z}) = \mathbb {E} _ {s ^ {\prime} \sim \mathcal {P} (\cdot | s, a, \hat {z})} [ r (s, a, s ^ {\prime}, \hat {z}) + \tag {9} \\ \left. \gamma \mathbb {E} _ {a ^ {\prime} \sim \pi (\cdot | s ^ {\prime}, \hat {z})} [ Q ^ {\pi} (s ^ {\prime}, a ^ {\prime}, \hat {z} ^ {\prime} ] \right], \\ \end{array} +$$ + +where $\hat{z} = \langle \hat{x},\hat{y},y_{real}\rangle$ is the augmented interaction goal, $\hat{x}$ means a single predicted reactor frame regarded as the target goal, $\hat{y}$ and $y_{real}$ are the actor motions coming from the diffusion model generation and real world inputs. + +Actor-aware Reaction Policy To conserve computational and communication resources while prioritizing interaction safety, we integrate a low-frequency (0.5 Hz) + +diffusion planner with a high-frequency (30 Hz) physics tracker and real-world actor capture interface. This frequency mismatch can cause the reaction policy to pursue stale targets, elevating collision risk when the actor's motion deviates unexpectedly. To safeguard against interpenetration, we incorporate both generated and captured actor motion histories into the policy observation and heavily penalize the trajectory inconsistency in the reward function, thereby enforcing collision-free, safety-critical responses. More details of implementations in Appendix B.2. + +# 4. Applications + +# 4.1. Immersive VR Interface + +Our system is built upon Unity, enabling real-time interaction and visualization in a VR environment. Leveraging VisionPro [42], we develop a VR platform that allows the user to wear a headset and engage in immersive, egocentric interactions with the character. The camera in Unity is positioned on the actor's head, ensuring that it moves and rotates with the actor's motions. It continuously streams real-time simulations from Unity to the VR headset, providing an immersive experience. The user can perform any actions to guide the character's movement, such as completing a duet dance with the character. + +# 4.2. Human-Robot Interaction Interface + +Benefiting from our robust Reaction Motion Tracker, the generated motions adhere to physical constraints and exhibit high adaptability. As a result, we can seamlessly transfer physically plausible humanoid interaction to human-robot interaction once we retarget the imitation goal from SMPL[37] joints to specific robot joints. An optimization-based method is leveraged following [15, 16, 24]. First we fit the SMPL shape parameters $\beta$ and use $\beta$ to solve for optimal robot joint angles $\theta$ . Then the corresponding joints' goal positions are calculated by forward kinematics. In the VR environment, the character can be replaced with the Unitree H1[49], enabling an immersive and realistic human-robot interaction experience. Furthermore, our approach holds potential for future deployment on physical robots. + +# 5. Experiments + +# 5.1. Datasets and Evaluation Metrics + +Datasets. We evaluate our model on two datasets: InterX [65], which contains 11,388 interaction SMPL-X [43] motion sequences on 40 interaction categories; InterHuman [32], which consists of 6022 interaction SMPL [37] motion sequences with around 1.7M frames. For data processing, we use the actor-reactor annotations provided by [66]. We canonicalize all the motions into the representation stated in Sec.3.3.1 and set frame rate to 30 FPS. + +Evaluation Metrics. We assess the generated reactor's sequence using three key aspects: 1) the quality of the reactor's motion independent of the actor, following kinematic motion generation works [14, 58]; 2) the physical plausibility of the reactor's motion, based on physics-based methods [59, 69]; and 3) the interaction quality between the actor and reactor, as discussed in [36, 52]. + +For 1), we evaluate using Frechet Inception Distance (FID) [17] to measure the similarity between the generated and ground truth data distributions, Diversity to assess motion variability, and Multi-modal Distance (MMDist) to measure the distance between the generated reaction and the text prompt. For 2), we assess Penetration, the average distance between the ground and the lowest body joint position below the ground, Floating, the average distance between the ground and the lowest body joint above the ground, and Skating, the average horizontal displacement during frames exhibiting skating behavior. For 3), we measure maximum Interpenetration Volume (IV) [36] of reactor vertices intersecting the actor mesh, and cross-distance metrics $FID_{cd}$ and $Div_{cd}$ in a 100-dimensional cross-distance space [52] to evaluate interaction realism. Details of metrics calculation in Appendix C.2. + +# 5.2. Implementation Details + +In our auto-regressive denoiser, we adopt the DIT architecture [44] consisting of 8 Transformer layers with a 512-dimensional hidden size. We employ a frozen DistilBERT encoder [10, 51] as the text encoder, which is widely used in motion generation tasks [14, 59]. For training and inference, we set the binary interaction field $\mathcal{I} \in \mathbb{R}^{6 \times 6}$ , the diffusion timestep $T = 8$ , the classifier-free guidance weight $w = 5$ , and the foot-contact, interaction, and prefix loss weights to $\lambda_{\mathrm{foot}} = 0.2$ , $\lambda_{\mathrm{inter}} = 0.5$ , and $\lambda_{\mathrm{prefix}} = 0.1$ , respectively. In the auto-regressive setting, we use an interaction history length of $h = 20$ frames and a prediction window length of $k = 40$ frames. We train for 300K denoising steps on both the InterHuman and Inter-X datasets, and observe optimal checkpoints at 250K steps for InterHuman and 220K steps for Inter-X. All models are trained on a single NVIDIA GeForce RTX 4090d GPU for approximately 30 hours. + +For physics policy training, we use the multiplicative version of PHC network with three primitives, and construct task goal with only target body keypoints. The policy is also trained on the same device and datasets as the former stage for 20 hours. The tracking policy is executed at $30\mathrm{Hz}$ while physics simulation is processed in NVIDIA's Isaac Gym[41] at $60\mathrm{Hz}$ . In the starting phase of inference, we capture the first 1 second motion from the actor, and store it in the buffer along with masked reactor motion prefix for the next planning horizon's generation, the downstream reactor tracker receives default joint positions as tracking goals. + +![](images/1d714e11b8df8ecffa32406a56a9227abf0c06487dc8b8bf4741f2bfc9914ae5.jpg) +Figure 3. Compared to CAMDM (top row), Human-X (bottom row) achieves more complete hand contact in tasks such as face-hitting and handshaking. Additionally, its foot movement appears more natural, as highlighted in the red and green circles. + +![](images/411b3599f11e91dabc4ffa4f42c86d50eda391247161d130da4f0a9f47982e93.jpg) + +![](images/33d889fc7db82c7f5e2a32e2ebb7fbe0b27642134625d49d4992158ba51299c0.jpg) +Figure 4. Visualization results on Human-Robot Interaction. The robot (black skeleton) and human (orange mesh) perform a handshake on a flat plane, from arm extension and palm contact through to shake completion, illustrating our method's spatial coordination and motion coherence. + +After warming-up, the diffusion planner generates reasonable action-reaction motion sequences, the reactor is able to continuously follow the planning and receives both real and generated actor sequences for policy correction. + +# 5.3. Actor-To-Reactor Evaluation + +Baselines. To evaluate the performance of our method, we adopt and reproduce several human reaction generation models as baselines: 1) InterFormer [8]: a transformer-based model with temporal and spatial attention for human reaction generation, 2) InterGen [32]: a diffusion-based model for generating human interaction motions conditioned on text prompts., 3) ReGenNet [66]: a diffusion-based model for generating human reactions conditioned on given human actions. and 4) CAMDM [7]: a transformer-based Conditional Auto-regressive Motion Diffusion Model. However, since it is designed for single-person motion generation, we utilize the action as a condition input to generate the corresponding reaction. + +Comparisons. As demonstrated in Tab. 2, in the online unconstrained reaction setting, our method exhibits significant improvements over baselines in Reaction Quality, Physical Plausibility, and Interaction Quality on both the Inter-X datasets. Unlike InterFormer's transformer architecture, our diffusion-based approach effectively captures the nuanced structure of data distributions, resulting in re + +actor motions that closely align with real-world motion patterns and thereby enhancing reaction quality. In comparison to InterGen and ReGenNet, our implementation of Auto-regressive Motion Diffusion Models, augmented with a Physics Tracker, enables real-time interaction with the reactor by predicting future frames. This ensures that the generated motions adhere to fundamental physical laws, producing more natural and believable animations. Furthermore, relative to CAMDM, we introduce an Interaction Loss, leading to substantial improvements in IV and $\mathrm{FID}_{cd}$ , thereby facilitating more natural interactions. The experiments on the InterHuman dataset can be found in C.1. + +# 5.4. Ablation Study + +We primarily conduct ablation studies on three aspects: Losses, Timesteps, and Horizon. For more experiments, please refer to Section C.3. + +Losses. To verify the effectiveness of our designed loss function, we ablate $\mathcal{L}_{inter}$ , $\mathcal{L}_{foot}$ , and $\mathcal{L}_{prefix}$ separately. As shown in Tab. 3, removing any of them degrades performance. Notably, excluding $\mathcal{L}_{inter}$ significantly increases $FID$ and $FID_{cd}$ , underscoring its importance for generating realistic interactions. Removing $\mathcal{L}_{foot}$ raises Skating and Floating scores, showing its role in reducing foot sliding and enhancing physical plausibility. + +
MethodsReactionPhysicsInteraction
FID↓Div.→MMDist. ↓Pene. ↓Skate. ↓Float. ↓IV ↓FIDcd ↓Divcd →
Ground Truth0.0026.0283.5240.0000.0237.9560.0240.23511.471
InterFormer [8]6.1504.9228.6480.5240.17524.1090.3326.8258.214
InterGen [32]5.4854.4176.3890.2700.14311.9150.2693.4378.730
ReGenNet [66]2.2125.0475.2350.1410.1189.8100.2181.9509.248
CAMDM [7]1.4305.6575.3180.1530.1279.8640.1452.0869.017
Human-X0.9756.0634.1150.1180.0928.6500.0761.6929.735
Human-X*1.2205.1405.0760.0260.0074.3420.0541.8338.565
+ +Table 2. Action-to-Reaction with online unconstrained reaction setting on Inter-X [65] dataset. A higher or lower value is better for ↑ or ↓, and → means the value closer to ground truth is better. * denotes the method with reaction policy designed for humanoid reaction. + +
ClassSettingsReactionPhysicsInteractionLatency (ms)
FID↓Div.→MMDist.→Penetration ↓Skating ↓Floating ↓IV ↓FIDcd ↓Divcd→
Ground Truth0.0026.0283.5240.0000.0237.9560.0240.23511.471-
Lossesw.o. Linter1.4576.1544.2640.1340.1078.7150.1192.2089.91713.6
w.o. Lfoot1.0096.1144.2050.1270.2399.1200.0941.6879.84913.6
w.o. Lprefix1.0236.1284.2250.1330.1138.8570.1001.6979.86513.6
Timesteps21.8095.5175.0070.1440.1139.0080.1532.0087.5133.8
51.1076.2634.5140.1280.1088.9080.1151.8078.0038.2
100.9886.0084.2080.1180.0968.7090.0971.6849.85316.9
1000.9085.9073.9630.1130.0938.6080.0871.6109.708170
Horizonh=10,k=301.0556.1064.1070.1230.1038.8050.1031.6879.8079.6
h=15,k=300.9876.0054.0060.1180.0978.7070.0981.6679.78710.1
h=20,k=200.9575.9563.9580.1130.0948.6580.0931.6269.7569.5
h=15,k=451.0066.0544.1050.1230.1048.8060.1021.6889.80913.6
Human-Xtext0.9265.9513.9090.1120.0878.2180.0921.6079.82213.6
unconstrained0.9756.0634.1150.1180.0928.6500.0761.6929.73513.6
+ +Table 3. Ablation studies of online reaction setting on the Inter-X [65] dataset. + +Timesteps. During the denoising process, we experiment with different sampling timesteps: 2, 5, 10, and 100. As shown in the Tab. 3, increasing the number of timesteps improves the model's performance; however, it also leads to higher latency. When timesteps = 10, the model has already achieved satisfactory results. Considering both generation quality and latency, we ultimately select timesteps = 8 as a trade-off, ensuring high-quality generation while meeting real-time requirements. + +Horizon. During inference process, the model takes a continuous sequence of $h$ history frames as input to predict $k$ future frames. We conduct experiments with different values of $h$ and $k$ , and the results in Tab. 3 show that the best performance is achieved when $h = 20$ and $k = 40$ . + +# 6. Conclusion + +We present Human-X Interaction, a novel framework for real-time human interaction synthesis that seamlessly in- + +tegrates autoregressive diffusion modeling with an actor-aware reaction policy. Our approach addresses the critical challenges of real-time synchronization, physical realism, and safety in human-ayar, human-humanoid and humanrobot interactions. Specifically, Human-X jointly predicts actions and reactions using both partner's motion histories, ensuring immersive and realistic motion synthesis. By incorporating comprehensive interaction-aware rewards and training the tracking policy with reinforcement learning, we achieve physically feasible and artifact-free interactions even in dynamic and safety-critical scenarios. Extensive experiments on the Inter-X and InterHuman datasets demonstrate significant improvements in motion quality, interaction continuity, and physical accuracy over existing methods. The current formulation focuses on synthesizing reactive motions and does not account for collaborative task completion between humans and machines. Extending the framework to support joint task execution with shared goals remains an open challenge. + +# 7. Acknowledgement + +This work was supported by Shanghai Local College Capacity Building Program(23010503100), NSFC (No.62406195), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (Shanghai), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University), HPC Platform and Core Facility Platform of Computer Science and Communication of ShanghaiTech University and Shanghai Engineering Research Center of Intelligent Vision and Imaging. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2 +[2] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2 +[3] Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems, 28, 2015. 5, 1 +[4] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, et al. Digital life project: Autonomous 3d characters with social intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 582-592, 2024. 6 +[5] Haoran Chen, Yiteng Xu, Yiming Ren, Yaoqin Ye, Xinran Li, Ning Ding, Peishan Cong, Ziyi Wang, Bushi Liu, Yuhan Chen, et al. Symbiosim: Human-in-the-loop simulation platform for bidirectional continuing learning in human-robot interaction. arXiv preprint arXiv:2502.07358, 2025. 6 +[6] Jingxiao Chen, Xinyao Li, Jiahang Cao, Zhengbang Zhu, Wentao Dong, Minghuan Liu, Ying Wen, Yong Yu, Liqing Zhang, and Weinan Zhang. Rhino: Learning real-time humanoid-human-object interaction from human demonstrations. arXiv preprint arXiv:2502.13134, 2025. 6 +[7] Rui Chen, Mingyi Shi, Shaoli Huang, Ping Tan, Taku Komura, and Xuelin Chen. Taming diffusion probabilistic models for character control. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 7, 8, 3, 4, 5 +[8] Baptiste Chopin, Hao Tang, Naima Otberdout, Mohamed Daoudi, and Nicu Sebe. Interaction transformer for human reaction generation. IEEE Transactions on Multimedia, 25: 8842-8854, 2023. 7, 8, 3 +[9] Zekai Deng, Ye Shi, Kaiyang Ji, Lan Xu, Shaoli Huang, and Jingya Wang. Human-object interaction via automatically designed vlm-guided motion policy. arXiv preprint arXiv:2503.18349, 2025. 2 + +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 6 +[11] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm- e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning, pages 8469-8488, 2023. 6 +[12] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. 2 +[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[14] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5152-5161, 2022. 4, 6, 2 +[15] Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. arXiv preprint arXiv:2406.08858, 2024. 6 +[16] Tairan He, Zhengyi Luo, Wenli Xiao, Chong Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Learning human-to-humanoid real-time whole-body teleoperation, 2024. 6 +[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[18] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 3, 5 +[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4, 5 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4, 5 +[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 6 +[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer + +Vision and Pattern Recognition, pages 16750-16761, 2023. 2 +[23] Muhammad Gohar Javed, Chuan Guo, Li Cheng, and Xingyu Li. Intermask: 3d human interaction generation via collaborative masked modelling. arXiv preprint arXiv:2410.10010, 2024. 2 +[24] Mazeyu Ji, Xuanbin Peng, Fangchen Liu, Jialong Li, Ge Yang, Xuxin Cheng, and Xiaolong Wang. Exbody2: Advanced expressive humanoid whole-body control. arXiv preprint arXiv:2412.13196, 2024. 6 +[25] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2 +[26] Jianping Jiang, Weiye Xiao, Zhengyu Lin, Huaizhong Zhang, Tianxiang Ren, Yang Gao, Zhiqian Lin, Zhongang Cai, Lei Yang, and Ziwei Liu. Solami: Social vision-language-action modeling for immersive interaction with 3d autonomous characters. arXiv preprint arXiv:2412.00174, 2024. 1, 2, 6 +[27] Wentao Jiang, Jingya Wang, Kaiyang Ji, Baoxiong Jia, Siyuan Huang, and Ye Shi. Arflow: Human action-reaction flow matching with physical guidance. arXiv preprint arXiv:2503.16973, 2025. 2 +[28] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 2 +[29] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 2 +[30] Jiefeng Li, Siyuan Bian, Chao Xu, Zhicun Chen, Lixin Yang, and Cewu Lu. Hybrid analytical-neural inverse kinematics for whole-body mesh recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3 +[31] Ronghui Li, Youliang Zhang, Yachao Zhang, Yuxiang Zhang, Mingyang Su, Jie Guo, Ziwei Liu, Yebin Liu, and Xiu Li. Interdance: Reactive 3d dance generation with realistic duet interactions. arXiv preprint arXiv:2412.16982, 2024. 2 +[32] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 2, 6, 7, 8, 3, 4, 5 +[33] Jing Lin, Yao Feng, Weiyang Liu, and Michael J Black. Chathuman: Language-driven 3d human understanding with retrieval-augmented tool reasoning. arXiv preprint arXiv:2405.04533, 2024. 6 +[34] Minghuan Liu, Menghui Zhu, and Weinan Zhang. Goal-conditioned reinforcement learning: Problems and solutions, 2022. 5 +[35] Yunze Liu, Changxi Chen, and Li Yi. Interactive humanoid: Online full-body motion reaction synthesis with social affordance canonicalization and forecasting. arXiv preprint arXiv:2312.08983, 2023. 2 +[36] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-reaction: Physically plausible real-time humanoid reaction + +synthesis via forward dynamics guided 4d imitation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3771-3780, 2024. 2, 3, 6 +[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 4, 6 +[38] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10895-10904, 2023. 3, 4, 5 +[39] Zhengyi Luo, Jinkun Cao, Josh Merel, Alexander Winkler, Jing Huang, Kris Kitani, and Weipeng Xu. Universal humanoid motion representations for physics-based control. arXiv preprint arXiv:2310.04582, 2023. 3 +[40] Zhengyi Luo, Jinkun Cao, Rawal Khirodkar, Alexander Winkler, Kris Kitani, and Weipeng Xu. Real-time simulated avatar from head-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 571-581, 2024. 4 +[41] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 4, 6 +[42] Younghyo Park and Pulkit Agrawal. Using apple vision pro to train and control robots, 2024. 6 +[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 6 +[44] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 6 +[45] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4):1-14, 2018. 3 +[46] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4, 5 +[48] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 5, 1 +[49] Unitree Robotics. Unitree h1 robot. https://www.unitree.com, 2021. Accessed: 2023-10-10. 6 + +[50] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 6 +[51] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 6, 4 +[52] Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. arXiv preprint arXiv:2403.18811, 2024. 2, 6 +[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4, 5 +[54] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39 (4):54-1, 2020. 1 +[55] Wenhui Tan, Boyuan Li, Chuhao Jin, Wenbing Huang, Xiting Wang, and Ruihua Song. Think then react: Towards unconstrained action-to-reaction motion generation. In The Thirteenth International Conference on Learning Representations, 2025. 2 +[56] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15999-16009, 2023. 2 +[57] Jiangnan Tang, Jingya Wang, Kaiyang Ji, Lan Xu, Jingyi Yu, and Ye Shi. A unified diffusion framework for scene-aware human motion estimation from sparse signals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21251-21262, 2024. 2 +[58] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 6 +[59] Guy Tevet, Sigal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. Closd: Closing the loop between simulation and diffusion for multi-task character control. arXiv preprint arXiv:2410.03441, 2024. 4, 6, 2 +[60] Jie Tian, Ran Ji, Lingxiao Yang, Suting Ni, Yuexin Ma, Lan Xu, Jingyi Yu, Ye Shi, and Jingya Wang. Gaze-guided hand-object interaction synthesis: Dataset and method. arXiv preprint arXiv:2403.16169, 2024. 2 +[61] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 2, 6 +[62] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 2 +[63] Zhen Wu, Jiaman Li, and C Karen Liu. Human-object interaction from human-level instructions. CoRR, 2024. 2, 6 + +[64] Liang Xu, Ziyang Song, Dongliang Wang, Jing Su, Zhicheng Fang, Chenjing Ding, Weihao Gan, Yichao Yan, Xin Jin, Xiaokang Yang, et al. Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2023. 2 +[65] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter-x: Towards versatile human-human interaction analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22260-22271, 2024. 2, 6, 8, 4, 5 +[66] Liang Xu, Yizhou Zhou, Yichao Yan, Xin Jin, Wenhan Zhu, Fengyun Rao, Xiaokang Yang, and Wenjun Zeng. Regennet: Towards human action-reaction synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1759-1769, 2024. 2, 6, 7, 8, 3, 4, 5 +[67] Lingxiao Yang, Shutong Ding, Yifan Cai, Jingyi Yu, Jingya Wang, and Ye Shi. Guidance with spherical gaussian constraint for conditional diffusion. In Proceedings of the 41st International Conference on Machine Learning, pages 56071-56095, 2024. 2 +[68] Ye Yuan, Viktor Makoviychuk, Y Guo, S Fidler, XB Peng, and K Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Trans. Graph, 42(4), 2023. 5, 1 +[69] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16010-16021, 2023. 3, 6, 2 +[70] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2 +[71] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE transactions on pattern analysis and machine intelligence, 46(6):4115-4128, 2024. 2 +[72] Zhenhao Zhang, Ye Shi, Lingxiao Yang, Suting Ni, Qi Ye, and Jingya Wang. Openhoi: Open-world hand-object interaction synthesis with multimodal large language model. arXiv preprint arXiv:2505.18947, 2025. 2 +[73] Kaifeng Zhao, Gen Li, and Siyu Tang. Dart: A diffusion-based autoregressive motion model for real-time text-driven motion control. arXiv preprint arXiv:2410.05260, 2024. 5, 1 +[74] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. 4 +[75] Kaizhen Zhu, Mokai Pan, Yuexin Ma, Yanwei Fu, Jingyi Yu, Jingya Wang, and Ye Shi. Unidb: A unified diffusion bridge framework via stochastic optimal control. arXiv preprint arXiv:2502.05749, 2025. 2 \ No newline at end of file diff --git a/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/images.zip b/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d970a54408a51480eb4e0ed5618b128012442663 --- /dev/null +++ b/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07ce8caad165c3208eb4e631d1877361227f15cb9478ec254136a2a11fc2f5a9 +size 543630 diff --git a/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/layout.json b/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a0b8ccef0771c3be2a572c235ece77f584fab3e5 --- /dev/null +++ b/2025/Towards Immersive Human-X Interaction_ A Real-Time Framework for Physically Plausible Motion Synthesis/layout.json @@ -0,0 +1,10004 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 102, + 529, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 102, + 529, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 529, + 140 + ], + "type": "text", + "content": "Towards Immersive Human-X Interaction: A Real-Time Framework for Physically Plausible Motion Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "spans": [ + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "text", + "content": "Kaiyang Ji" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "text", + "content": " Ye Shi" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "text", + "content": " Zichen Jin" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "text", + "content": " Kangyi Chen" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "text", + "content": " Lan Xu" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "text", + "content": " Yuexin Ma" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "text", + "content": " Jingyi Yu" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "text", + "content": " Jingya Wang" + }, + { + "bbox": [ + 113, + 159, + 493, + 189 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 242, + 190, + 369, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 242, + 190, + 369, + 203 + ], + "spans": [ + { + "bbox": [ + 242, + 190, + 369, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 242, + 190, + 369, + 203 + ], + "type": "text", + "content": "ShanghaiTech University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 203, + 484, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 203, + 484, + 217 + ], + "spans": [ + { + "bbox": [ + 126, + 203, + 484, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 126, + 203, + 484, + 217 + ], + "type": "text", + "content": "Shanghai Engineering Research Center of Intelligent Vision and Imaging" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 220, + 527, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 220, + 527, + 232 + ], + "spans": [ + { + "bbox": [ + 80, + 220, + 527, + 232 + ], + "type": "text", + "content": "{jiky2024,shiye,jinzch12023,v-chenky,xulan1,mayuexin,yujingyi,wangjingya}@shanghaiitech.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 61, + 245, + 547, + 365 + ], + "blocks": [ + { + "bbox": [ + 61, + 245, + 547, + 365 + ], + "lines": [ + { + "bbox": [ + 61, + 245, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 61, + 245, + 547, + 365 + ], + "type": "image", + "image_path": "f77c7108da04d1983bd50f0fbc1b3c5ee4539371662cb0c5a1ef5c6c0123245a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 373, + 555, + 395 + ], + "lines": [ + { + "bbox": [ + 54, + 373, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 555, + 395 + ], + "type": "text", + "content": "Figure 1. We propose Human-X, the first framework designed to enable latency-free interaction between humans and diverse entities, including human-athlete, human-humanoid, and human-robot interaction." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 405, + 200, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 405, + 200, + 418 + ], + "spans": [ + { + "bbox": [ + 151, + 405, + 200, + 418 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 430, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 430, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 54, + 430, + 297, + 694 + ], + "type": "text", + "content": "Real-time synthesis of physically plausible human interactions remains a critical challenge for immersive VR/AR systems and humanoid robotics. While existing methods demonstrate progress in kinematic motion generation, they often fail to address the fundamental tension between real-time responsiveness, physical feasibility, and safety requirements in dynamic human-machine interactions. We introduce Human-X, a novel framework designed to enable immersive and physically plausible human interactions across diverse entities, including human-ayar, human-humanoid, and human-robot systems. Unlike existing approaches that focus on post-hoc alignment or simplified physics, our method jointly predicts actions and reactions in real-time using an auto-regressive reaction diffusion planner, ensuring seamless synchronization and context-aware responses. To enhance physical realism and safety, we integrate an actor-aware motion tracking policy trained with reinforcement learning, which dynamically adapts to interaction partners' movements while avoiding artifacts like foot sliding and penetration. Extensive experiments on the Inter-X and InterHuman datasets demonstrate significant improvements in motion quality, interaction continuity, and physical" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 406, + 556, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 556, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 556, + 477 + ], + "type": "text", + "content": "plausibility over state-of-the-art methods. Our framework is validated in real-world applications, including virtual reality interface for human-robot interaction, showcasing its potential for advancing human-robot collaboration. Project page: https://humanx-interaction.github.io/" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 501, + 395, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 501, + 395, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 501, + 395, + 514 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 521, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 700 + ], + "type": "text", + "content": "Human-human interaction synthesis has long been a focal point in the field of motion generation, with widespread applications in social dynamics, gaming, and animation [26, 54]. With the rapid development of virtual reality (VR) and augmented reality (AR), these technologies now empower the creation of deeply immersive and contextually rich interactions between humans and digital avatars. Particularly, the emergence of humanoid robots further bridges the gap between virtual and physical realms, enabling robots to engage with humans in daily life through adaptive and socially aware behaviors. However, achieving real-time prediction of diverse and safety-guaranteed humanoid reactions, an essential capability for applications ranging from daily companionship to assistive care, remains a critical, unsolved challenge." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "Current approaches to human interaction modeling rely" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10173" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 555, + 148 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 555, + 148 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 555, + 148 + ], + "type": "table", + "html": "
Methodsphysics-basedimmersive inferencelatency-freereal-time modelgeneration abilitystrong contacttext-control
InterGen [32]✓✓✓
ReGenNet [66]✓✓✓✓
SoLaMi [26]✓✓✓
PhysReaction [36]✓✓✓✓
Think-Then-React [55]✓✓✓✓
Human-X✓✓✓✓✓✓✓
", + "image_path": "d90d1e1a3d1a5862280c3c87ee3d95b1b3ff3632e0d5f547d86c7b432b7a9f49.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 127, + 157, + 481, + 168 + ], + "lines": [ + { + "bbox": [ + 127, + 157, + 481, + 168 + ], + "spans": [ + { + "bbox": [ + 127, + 157, + 481, + 168 + ], + "type": "text", + "content": "Table 1. Comparative analysis of Human-X versus existing human reaction synthesis approaches." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 190, + 294, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 190, + 294, + 321 + ], + "spans": [ + { + "bbox": [ + 54, + 190, + 294, + 321 + ], + "type": "text", + "content": "heavily on offline training data or simplified physics assumptions, leading to limited adaptability in dynamic scenarios. While data-driven methods such as [23, 32, 65] have shown promise in synthesizing kinematic motions interaction, they often fail to account for the intricate physical constraints and safety requirements inherent in human-humanoid cohabitation. For instance, abrupt or unstable motions generated through these techniques pose risks in shared environments. This gap underscores the need for a unified interaction framework that integrates real-time reactivity, physical realism, and safety assurance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 324, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 324, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 56, + 324, + 295, + 539 + ], + "type": "text", + "content": "In this work, we explore Human-X Interaction (Fig.1), a systematic study of immersive interactions across three domains: 1) Human-Avatar Interaction using kinematic-driven motion in VR/AR; 2) Human-Humanoid Interaction with physics-based motion synthesis; 3) Human-Robot Interaction ensuring safe physical contact. Unlike existing physics-aware methods limited to replicating training data, our framework balances natural movement with physical feasibility while maintaining real-time performance. This integration enables safer and more natural human-machine collaboration across applications from entertainment to assistive robotics. Achieving real-time synchronization between interacting partners faces major technical challenges. Conventional methods rely exclusively on past actor motions to generate reactions, leading to temporal mismatches. Furthermore, ensuring physical realism is more challenging, as existing solutions focus mainly on basic collision avoidance and neglect dynamic interaction forces." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 543, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 649 + ], + "type": "text", + "content": "To overcome these limitations, we present Human-X, an Auto-Regressive Action-Reaction Diffusion Planner for real-time interaction. Our framework jointly models both partners' motion histories to predict subsequent actions and reactions. We introduce multiple contact-aware losses to train realistic interactions, and combine the diffusion model with an actor-aware tracking policy to ensure physical feasibility in humanoid/robot scenarios. In summary, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 700 + ], + "type": "text", + "content": "- The first auto-regressive diffusion model that jointly predicts actions and reactions using both partners' motion history with comprehensive contract losses to ensure real-time, immersive and realistic motion synthesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 702, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 295, + 712 + ], + "type": "text", + "content": "- An actor-aware reaction policy trained with reinforce" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 190, + 553, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 190, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 321, + 190, + 553, + 225 + ], + "type": "text", + "content": "ment learning that dynamically adapts to interaction partners' motions, preventing common artifacts like foot sliding while maintaining physical feasibility." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 226, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 553, + 285 + ], + "type": "text", + "content": "- A complete real-time VR system demonstrating practical effectiveness. Experiments on Inter-X and InterHuman datasets show significant improvements in motion quality, interaction continuity, and physical accuracy compared to existing methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 294, + 406, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 406, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 406, + 306 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 314, + 465, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 465, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 465, + 326 + ], + "type": "text", + "content": "2.1. Human Motion Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 331, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 545 + ], + "type": "text", + "content": "Human-centric motion generation has seen remarkable advancements in recent years, driven by breakthroughs in large language models [1, 25] and generative frameworks [13, 20, 28, 67, 75]. Notably, the introduction of motion diffusion models [58, 71] has significantly elevated the performance of single-person motion generation. Building on these developments, researchers have increasingly focused on enriching interactions with the environment by incorporating diverse modalities, such as text [25, 62, 63, 70], audio [2, 61], and scene information [12, 22, 29, 57]. For instance, recent works [9, 60, 62, 63, 72] have leveraged text-based guidance to facilitate Human-Object Interactions (HOI), while MotionGPT [25] treats motion as a form of language, enabling seamless bidirectional translation between text and motion. Despite these advancements, existing approaches face a critical limitation: their model architectures are inherently ill-suited for real-time reaction synthesis." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 466, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 466, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 466, + 564 + ], + "type": "text", + "content": "2.2. Human-Human Interaction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Research on human-human interaction synthesis has evolved around three primary directions: (1) Interactive Behavior Modeling leverages actor-reactor role decomposition to capture interaction dynamics. For example, [56] identifies role-specific patterns, while [32, 64] use transformer architectures to model multi-agent dependencies. (2) Multimodal Interaction integrates text, audio, or scene context to enhance realism [31, 52], but often overlook dynamic physical constraints critical for real-world deployment. (3) Reactive Motion Generation employs sequence models [35] and diffusion models [27, 66] to synthesize temporally coherent reactions. These methods, however, struggle with physical" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10174" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 74, + 556, + 277 + ], + "blocks": [ + { + "bbox": [ + 62, + 74, + 556, + 277 + ], + "lines": [ + { + "bbox": [ + 62, + 74, + 556, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 74, + 556, + 277 + ], + "type": "image", + "image_path": "9d3fce72a20fd83d348ef6c4fd7a584078f361a8fc1f08ae40900361a5995903.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 290, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 290, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 555, + 357 + ], + "type": "text", + "content": "Figure 2. Overview of our immersive real-time interaction synthesis pipeline: (a) Actor Motion Capture: A human actor's movements are recorded at 30 fps by an RGB-D camera and translated into 3D poses, which are then retargeted to a humanoid character. (b) Realistic Reactor Motion Generation: An auto-regressive diffusion model, guided by optional text prompts (e.g., \"Dancing is what to do\"), generates plausible reaction motions. These motions are tracked by an actor-aware controller, which uses proprioception signals to ensure realistic, synchronized interactions. (c) Real-time VR Interface: The generated and tracked motions are rendered in simulator, providing both a third-person view and a binocular VR view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 377, + 297, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 297, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 297, + 521 + ], + "type": "text", + "content": "plausibility and temporal consistency in long-horizon interactions. Existing Reactive Generation methods face a fundamental tension: Human interaction models achieve reactivity but lack physical grounding, while physics-based animations enforce plausibility but are agnostic to interaction dynamics. While PhysReaction [36] introduced the first physics-based approach for reaction synthesis using imitation learning, it was limited to mimicking known distributions, lacking generalization ability, diversity, controllability, and the capacity to handle unseen tasks. Our work bridges this gap by unifying interaction-aware reactivity with physically constrained control." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 529, + 228, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 228, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 228, + 541 + ], + "type": "text", + "content": "2.3. Physics-based Motion Synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 714 + ], + "type": "text", + "content": "In contrast to kinematics-based methods, physics-based motion synthesis has emerged as a widely adopted approach for generating realistic and physically plausible character movements by integrating physical principles into the animation process. Early work, such as DeepMimic [45], leveraged deep reinforcement learning to enable characters to imitate diverse motion clips within a physics-based framework. Subsequent advancements [18, 46] utilized generative adversarial networks to enhance stylized motion realism. Progress continued with PHC [38] and PULSE [39], focusing on improving the diversity of generated motions. A significant milestone was achieved with PhysDiff [69], which introduced physics-guided diffusion models to mitigate artifacts such as floating and foot sliding. Despite" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 377, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 402 + ], + "type": "text", + "content": "these advancements, the physical plausibility of human-human interactions remains largely unexplored." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 412, + 370, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 370, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 370, + 424 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 431, + 418, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 418, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 418, + 444 + ], + "type": "text", + "content": "3.1. System Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 449, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 665 + ], + "type": "text", + "content": "In this section, we present a detailed overview of our immersive interaction synthesis system. To facilitate realistic and responsive interactions between two agents in immersive environments, our system takes online actor motion capture data as input, processes it through an advanced motion generation pipeline, and outputs continuous, reactive motion in real-time, as shown in Fig.2. The first key component of our system, detailed in Sec.3.2, is the real-time capture of the actor's pose and the reconstruction of its state in physics simulation. Subsequently, the Auto-regressive Interaction Diffusion Model, as described in Sec.3.3, is employed to generate the reaction of one agent in response to the actor context. Next, we introduce the Reaction Motion Tracker in Sec.3.4, which ensures that the motion of the reactor adheres to physical constraints for realism and natural behavior. Finally, we explain how VR Interface works in Sec.4.1, which enables immersive, real-time interaction with the generated motions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 672, + 545, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 545, + 685 + ], + "type": "text", + "content": "3.2. Human Pose Estimation and Reconstruction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "We firstly utilize a real-time RGB-D camera to capture images. Then, HybrIK [30] is employed for latency-free re" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10175" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "construction of the human pose. However, due to the inaccuracies in the depth information provided by HybrIK, the depth map is directly obtained from the RGB-D camera instead. Following [38, 40], we bridge SMPL's kinematic skeleton to physics simulation via applying a simple retargeting method. We utilize Isaac Gym [41] as our physics simulation platform to create a third-person perspective for the interaction between the actor and reactor. Besides, the image reading process, motion generation of the reactor, and the physical simulation within Isaac Gym are handled by separate processes. ROS2 is used as the communication framework to facilitate real-time data transmission." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 224, + 274, + 237 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 274, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 274, + 237 + ], + "type": "text", + "content": "3.3. Online Human Action-Reaction Synthesis" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "text", + "content": "The goal of online action-reaction synthesis is to generate realistic and continuous future " + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "text", + "content": " frames of reactor motions " + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^i \\coloneqq \\{x^{n:n + k - 1}\\}" + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "text", + "content": " history frames of interaction contexts " + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}^{i - 1},\\mathbf{y}^{i - 1}\\} \\coloneqq \\{x^{n - h:n - 1},y^{n - h:n - 1}\\}" + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "text", + "content": " at current frame number " + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "text", + "content": " and current window index " + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "text", + "content": ". During each window prediction, an optional text prompt " + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 243, + 296, + 327 + ], + "type": "text", + "content": " can be added to guide the reaction generation process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 334, + 244, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 244, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 244, + 346 + ], + "type": "text", + "content": "3.3.1. Reactor-centric Interaction Modeling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 506 + ], + "type": "text", + "content": "Reactor-centric Canonicalization. For individual motion synthesis, the initial frame of each sequence is typically aligned at the root to improve the generalizability of the model [14, 59]. However, using the same normalization for human-human interaction data can eliminate key details about the mutual spatial relationship. To retain relative positioning without losing generalization ability, we leverage a reactor-centric canonicalization that the pose of reactor is realigned at the root and oriented toward the forward direction, while the pose of actor is adjusted by the same transformation. This method maintains the essential spatial context between the two, enabling the model to accurately learn their interdependent movement dynamics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": "Reactor-centric Interaction Representation. We model each frame of interaction as a hybrid representation of both actor and reactor motion. For reaction representation, we follow previous works [14, 58] and use the HumanML3D representation. The reactor motion " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " is defined by " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "x^{n} = (r^{x},\\dot{r}^{x},p^{x},\\dot{p}^{x},\\theta^{x},f^{x})" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "r^x\\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " denotes the height of the roots, " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\dot{r}^x\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " denotes the angular velocity of the roots along the Y axis and the linear velocity of the roots in the XZ plane, " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "p^x\\in \\mathbb{R}^{(J - 1)\\times 3}" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " are the local joint positions, " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\dot{p}^x\\in \\mathbb{R}^{J\\times 3}" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " denotes the temporal difference of the local joint positions, " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\theta^x\\in \\mathbb{R}^{(J - 1)\\times 6}" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " is the 6D representation [74] of the joint rotations and " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "f^x\\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " are binary labels for the detection of contact. For actor motion captured from pose estimation, we define the actor " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "y^n = (r^{x\\to y},\\dot{r}^{x\\to y},p^{x\\to y},\\dot{p}^{x\\to y},f^y)" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " where all the features are in the reactor's reference frame, making it a pure" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": "relative representation consisting of relative root transition, relative root rotation, relative root linear velocity, relative joint positions, relative joint velocity and feet contact labels. To emphasize actor-reactor contacts, we introduce the binary interaction field " + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^n\\in \\mathbb{R}^{\\bar{J}\\times \\bar{J}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": " obtained by thresholding the relative distance between the selected " + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\bar{J}" + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": " joints of each agent in a contact map. For body model selection, we follow the skeleton of SMPL [37] with " + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "inline_equation", + "content": "J = 22" + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": " joints. The total reactor-centric interaction " + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": " at frame " + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": " representation can be formed as," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 394, + 200, + 555, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 200, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 394, + 200, + 555, + 214 + ], + "type": "interline_equation", + "content": "z ^ {n} = \\left(x ^ {n}, y ^ {n}, \\mathcal {I} ^ {n}\\right), \\tag {1}", + "image_path": "e0317eaa91cad84d556488df04e2c31015220cff3a3e32ecb4c079a0e91c9bf7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "inline_equation", + "content": "x^n" + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "text", + "content": " denotes the reactor motion, " + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "inline_equation", + "content": "y^n" + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "text", + "content": " denotes the relative actor motion and " + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^n" + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "text", + "content": " is the joint contact features. Thus, we can model the interaction sequence as " + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^i = \\{z^{n:n + k - 1}\\}" + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "text", + "content": " given " + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "text", + "content": " frames and window index " + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 221, + 555, + 269 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 275, + 537, + 286 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 537, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 537, + 286 + ], + "type": "text", + "content": "3.3.2. Auto-regressive Interaction Diffusion Planner" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "text", + "content": "Based on the interaction motion representation provided in Sec.3.3.1, we transfer raw motion data into a canonical space suitable for generative learning with denoising networks. For probabilistic generator, we follow the DDPM [20] framework by incrementally adding Gaussian noise to the clean data sample " + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 290, + 554, + 361 + ], + "type": "text", + "content": " steps:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 370, + 555, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 370, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 353, + 370, + 555, + 384 + ], + "type": "interline_equation", + "content": "q \\left(\\mathbf {z} _ {t} \\mid \\mathbf {z} _ {t - 1}\\right) = \\mathcal {N} \\left(\\sqrt {\\alpha_ {t}} \\mathbf {z} _ {t - 1}, (1 - \\alpha_ {t}) \\mathbf {I}\\right), \\tag {2}", + "image_path": "554008b883814f826515aa2a829daeccf1050f3b33cf2efe59acdd6c207bbca0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\alpha_{t} \\in (0,1)" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": " are constant parameters for variance scheduling. Given the noised data " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^i" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": "-th window clip, the diffusion step " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": ", the prefix interaction window " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{i-1} = z^{n-h:n-1}" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": ", and the optional text prompt " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": ", we train a transformer-based denoiser " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": " to predict the cleaned interaction primitive itself instead of predicting the noise [58]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 388, + 474, + 553, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 474, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 388, + 474, + 553, + 488 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {z}} _ {0} ^ {i} = \\mathcal {G} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, c\\right), \\tag {3}", + "image_path": "366050ea7ca9bc0b2e97956de28f3f4cb357ecc2264b04dd53a124f58016ae43.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 492, + 554, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 554, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 554, + 564 + ], + "type": "text", + "content": "where the text prompt " + }, + { + "bbox": [ + 313, + 492, + 554, + 564 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 492, + 554, + 564 + ], + "type": "text", + "content": " is randomly masked by a rate of 0.15 during training to enable unconditioned generation and classifier-free guidance [19]. The predicted motion " + }, + { + "bbox": [ + 313, + 492, + 554, + 564 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0^i" + }, + { + "bbox": [ + 313, + 492, + 554, + 564 + ], + "type": "text", + "content": " is then inserted to the interaction history for next interaction window extraction after the diffusion sampling loop [20, 53]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 578, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 554, + 603 + ], + "type": "text", + "content": "Losses The denoised motion " + }, + { + "bbox": [ + 313, + 578, + 554, + 603 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0^i" + }, + { + "bbox": [ + 313, + 578, + 554, + 603 + ], + "type": "text", + "content": " is supervised by a simple objective in DDPM [20]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 610, + 553, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 610, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 349, + 610, + 553, + 626 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s i m p l e}} = \\mathbb {E} _ {\\mathbf {z} _ {0} ^ {i} \\sim q \\left(\\mathbf {z} _ {0} ^ {i}\\right), t \\sim [ 1, T ]} \\left[ \\| \\mathbf {z} _ {0} ^ {i} - \\hat {\\mathbf {z}} _ {0} ^ {i} \\| _ {2} ^ {2} \\right]. \\tag {4}", + "image_path": "078b90489277f963e223b62ccfb0216d2a1f242a0b43fc6195a8c14a611cc9bc.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 632, + 554, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 554, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 554, + 681 + ], + "type": "text", + "content": "We extend the training objective with several auxiliary losses to enhance interaction quality and realism. To optimize foot contact quality, we add a foot contact loss on both agents:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 343, + 689, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 689, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 343, + 689, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f o o t}} = \\sum_ {j \\in \\{x ^ {\\text {f e e t}}, y ^ {\\text {f e e t}} \\}} \\| \\dot {p} ^ {j} \\left(\\hat {\\mathbf {z}} _ {0} ^ {i}\\right) \\odot f ^ {j} \\left(\\hat {\\mathbf {z}} _ {0} ^ {i}\\right) \\| _ {2} ^ {2}, \\tag {5}", + "image_path": "4f81328a50710e036bf9de784b103c82e4b7b06a454a40af8667207d3c9c9584.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10176" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\dot{p}^j (\\hat{\\mathbf{z}}_0^i)" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " is the corresponding foot joints velocity, " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "f^{j}(\\hat{\\mathbf{z}}_{0}^{i})" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " are foot contact label described in Sec 3.3.1 and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " denotes for Hadamard product. Note that " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\dot{p}^j (\\hat{\\mathbf{z}}_0^i)" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " is defined by the relative local velocity in the reactor frame, and we transform it to the global velocity in loss calculation. For interaction refinement, we design an interaction loss on contact joints distance:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 165, + 296, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 165, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 165, + 296, + 194 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i n t e r} = \\sum_ {x \\in \\bar {J} ^ {x}, y \\in \\bar {J} ^ {y}} \\| (p ^ {x} (\\hat {\\mathbf {z}} _ {0} ^ {i}) - p ^ {y} (\\hat {\\mathbf {z}} _ {0} ^ {i})) \\odot \\mathcal {I} ^ {x, y} (\\hat {\\mathbf {z}} _ {0} ^ {i}) \\| _ {2} ^ {2}, (6)", + "image_path": "82db56892cef6597209280995df7d035e68df93bf17ea3bae482a184a13e8bf4.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 201, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 286 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 201, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^{x,y}(\\hat{\\mathbf{z}}_0^i)" + }, + { + "bbox": [ + 55, + 201, + 296, + 286 + ], + "type": "text", + "content": " are binary body interaction fields and " + }, + { + "bbox": [ + 55, + 201, + 296, + 286 + ], + "type": "inline_equation", + "content": "p^x (\\hat{\\mathbf{z}}_0^i), p^y (\\hat{\\mathbf{z}}_0^i)" + }, + { + "bbox": [ + 55, + 201, + 296, + 286 + ], + "type": "text", + "content": " are contact joints positions in the same root frame to enhance strong interaction. To improve the continuity of auto-regressive generation, we finally add a prefix loss " + }, + { + "bbox": [ + 55, + 201, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{prefix}" + }, + { + "bbox": [ + 55, + 201, + 296, + 286 + ], + "type": "text", + "content": " to evaluate the difference between the end of prefix motion and the beginning of the generated motion. The full loss applied to the model is then:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 295, + 296, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 295, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 88, + 295, + 296, + 324 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\text {s i m p l e}} + \\lambda_ {\\text {f o o t}} \\mathcal {L} _ {\\text {f o o t}} + \\lambda_ {\\text {i n t e r}} \\mathcal {L} _ {\\text {i n t e r}} \\tag {7} \\\\ + \\lambda_ {p r e f i x} \\mathcal {L} _ {p r e f i x}, \\\\ \\end{array}", + "image_path": "cab1b487f3bee31a043c76deec7fa234b991a54f69c0c5d2eab7452c3ae0fcbb.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 331, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 331, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\lambda_{foot}" + }, + { + "bbox": [ + 55, + 331, + 296, + 357 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 331, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\lambda_{inter}" + }, + { + "bbox": [ + 55, + 331, + 296, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 331, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\lambda_{prefix}" + }, + { + "bbox": [ + 55, + 331, + 296, + 357 + ], + "type": "text", + "content": " are hyper-parameters to balance the loss terms from different scaling aspects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "content": "Training During training, we utilize the scheduled training strategy [3, 48, 68, 73] to enhance auto-regressive generation stability and text prompt controllability. The motion is cropped to sequences of " + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "content": " consecutive interaction windows with a random offset from the dataset. During training, " + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0^i" + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "content": " is noised to " + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^i" + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "content": " at diffusion timestep " + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "content": ", and interaction history " + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{i-1}" + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "content": " is divide into 3 stages with the sample probability " + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "content": ": ground truth for supervised learning " + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "inline_equation", + "content": "(p = 1)" + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "content": ", random sampling for scheduled training (decaying " + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "content": "), and auto-regressive prediction " + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "inline_equation", + "content": "(p = 0)" + }, + { + "bbox": [ + 55, + 369, + 296, + 501 + ], + "type": "text", + "content": ". Details in Appendix B.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 516, + 296, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 296, + 576 + ], + "type": "text", + "content": "Inference After training a denoise network " + }, + { + "bbox": [ + 55, + 516, + 296, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 516, + 296, + 576 + ], + "type": "text", + "content": ", we can auto-regressively generate reaction sequences given interaction history " + }, + { + "bbox": [ + 55, + 516, + 296, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^{i-1}" + }, + { + "bbox": [ + 55, + 516, + 296, + 576 + ], + "type": "text", + "content": ", a diffusion sampler " + }, + { + "bbox": [ + 55, + 516, + 296, + 576 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 55, + 516, + 296, + 576 + ], + "type": "text", + "content": " like [20, 53] and online text guidance " + }, + { + "bbox": [ + 55, + 516, + 296, + 576 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 516, + 296, + 576 + ], + "type": "text", + "content": " by sampling with classifier-free guidance [19]:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 583, + 295, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 583, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 69, + 583, + 295, + 620 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {G} _ {w} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, c\\right) = \\mathcal {G} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, \\varnothing\\right) \\\\ + \\quad w \\cdot \\left(\\mathcal {G} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, c\\right) - \\mathcal {G} \\left(\\mathbf {z} _ {t} ^ {i}, \\mathbf {z} ^ {i - 1}, t, \\varnothing\\right)\\right), \\tag {8} \\\\ \\end{array}", + "image_path": "6b606e5dc19e1611e6414dc37b542ba890d25a9d40a08d11e907f81636a675c5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 628, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 628, + 296, + 653 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 628, + 296, + 653 + ], + "type": "text", + "content": " is the guidance hyper-parameter, as shown in Alg.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 660, + 151, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 151, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 151, + 673 + ], + "type": "text", + "content": "3.4. Reaction Policy" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Our actor-motion data captured from virtual reality or existing datasets roughly approximates the dynamic plausibility. In contrast, the generated reactor motions is not always" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 555, + 96 + ], + "type": "text", + "content": "Algorithm 1 Online reaction sampling using auto-regressive motion diffusion model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "spans": [ + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "text", + "content": "Input: denoiser " + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "text", + "content": ", online actor loader " + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}" + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "text", + "content": ", online text loader " + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "text", + "content": ", total diffusion steps " + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "text", + "content": ", classifier-free guidance scale " + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "text", + "content": ", auto-regressive reaction sampler " + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 323, + 99, + 555, + 135 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 324, + 135, + 474, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 135, + 474, + 147 + ], + "spans": [ + { + "bbox": [ + 324, + 135, + 474, + 147 + ], + "type": "text", + "content": "Output: reactor motion sequence " + }, + { + "bbox": [ + 324, + 135, + 474, + 147 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 148, + 430, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 148, + 430, + 172 + ], + "spans": [ + { + "bbox": [ + 325, + 148, + 430, + 172 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {z} _ {T} ^ {0} \\leftarrow \\mathcal {N} (0, 1) \\\\ \\mathbf {X} \\leftarrow \\mathcal {S} (\\mathcal {G}, \\mathbf {z} _ {T} ^ {0}, T, \\varnothing , \\varnothing , w) \\\\ \\end{array}", + "image_path": "6a04ea379afff49030e3ee7a03c4b83d149f8a021dba66a143dc4fec75255290.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 411, + 172, + 554, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 172, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 411, + 172, + 554, + 183 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 411, + 172, + 554, + 183 + ], + "type": "text", + "content": " init with unconditioned sampling" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 183, + 400, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 183, + 400, + 194 + ], + "spans": [ + { + "bbox": [ + 325, + 183, + 400, + 194 + ], + "type": "text", + "content": "for " + }, + { + "bbox": [ + 325, + 183, + 400, + 194 + ], + "type": "inline_equation", + "content": "i\\gets 1" + }, + { + "bbox": [ + 325, + 183, + 400, + 194 + ], + "type": "text", + "content": " to N do" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 455, + 184, + 554, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 455, + 184, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 455, + 184, + 554, + 194 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 455, + 184, + 554, + 194 + ], + "type": "text", + "content": " num of actor windows" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 340, + 194, + 531, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 194, + 531, + 207 + ], + "spans": [ + { + "bbox": [ + 340, + 194, + 531, + 207 + ], + "type": "interline_equation", + "content": "\\mathbf {y} ^ {i - 1} \\leftarrow \\operatorname {C o l l e c t} (\\mathbf {Y}), \\mathbf {x} ^ {i - 1} \\leftarrow \\operatorname {C o l l e c t} (\\mathbf {X})", + "image_path": "8ee72da690ba03b97e066ed73d37daca5f5aa7dcd74869868870c8f063a92cb4.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 340, + 207, + 494, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 207, + 494, + 220 + ], + "spans": [ + { + "bbox": [ + 340, + 207, + 494, + 220 + ], + "type": "interline_equation", + "content": "\\mathbf {z} ^ {i - 1} \\leftarrow \\operatorname {C A N O N I C A L I Z E} \\left(\\mathbf {x} ^ {i - 1}, \\mathbf {y} ^ {i - 1}\\right)", + "image_path": "59c7205901fd295f58f08324a1ff47e1c393201362a78d875438f1189502c599.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 456, + 220, + 553, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 456, + 220, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 456, + 220, + 553, + 231 + ], + "type": "text", + "content": "> get interaction history" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 340, + 231, + 400, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 231, + 400, + 243 + ], + "spans": [ + { + "bbox": [ + 340, + 231, + 400, + 243 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {T} ^ {i} \\leftarrow \\mathcal {N} (0, 1)", + "image_path": "0a5dbfd6b263e31af8fe0818952dc818b0995703715ffb24345a86a4ffb403d3.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 340, + 243, + 422, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 243, + 422, + 255 + ], + "spans": [ + { + "bbox": [ + 340, + 243, + 422, + 255 + ], + "type": "interline_equation", + "content": "c ^ {i} \\leftarrow \\operatorname {C O L L E C T} (\\mathbf {C})", + "image_path": "ecf9e11e36e62941d86d5da22f221c6d4959ca579d1b1234897f0bc4faf09590.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 340, + 255, + 553, + 267 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 255, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 340, + 255, + 553, + 267 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {z}} _ {0} ^ {i} \\leftarrow S (\\mathcal {G}, \\mathbf {z} _ {T} ^ {i}, T, \\mathbf {z} ^ {i - 1}, c ^ {i}, w) \\triangleright C F G \\text {s a m p l i n g l o o p}", + "image_path": "37cc7d894f4e5608c489fadf5a731033f70f540d5a39d59b2479c63366611e3b.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 340, + 267, + 425, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 267, + 425, + 279 + ], + "spans": [ + { + "bbox": [ + 340, + 267, + 425, + 279 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} ^ {i} \\leftarrow \\operatorname {R e c o v e r} \\left(\\hat {\\mathbf {z}} _ {0} ^ {i}\\right)", + "image_path": "c1984adf26bd573020e5e856c1a8b05853b6118aefe8587459463f23e5bfc3ff.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 340, + 279, + 395, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 279, + 395, + 289 + ], + "spans": [ + { + "bbox": [ + 340, + 279, + 395, + 289 + ], + "type": "interline_equation", + "content": "\\mathbf {X} \\leftarrow \\mathbf {X} \\cup \\hat {\\mathbf {x}} ^ {i}", + "image_path": "ff143610f324c049d4fb836d978ebbdcf1156a87a74d522f70047feec931751f.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 362, + 292, + 553, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 292, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 362, + 292, + 553, + 302 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 362, + 292, + 553, + 302 + ], + "type": "text", + "content": " insert generated reaction to reactor sequences" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 325, + 303, + 358, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 303, + 358, + 313 + ], + "spans": [ + { + "bbox": [ + 325, + 303, + 358, + 313 + ], + "type": "text", + "content": "end for" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 325, + 315, + 362, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 315, + 362, + 325 + ], + "spans": [ + { + "bbox": [ + 325, + 315, + 362, + 325 + ], + "type": "text", + "content": "return X" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 349, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 408 + ], + "type": "text", + "content": "lying in a dynamically consistent manifold. Under the purpose of building the action-reaction controller in a physical world, we leverage the perpetual humanoid controller (PHC) [38], a universal physics tracker to track the generated reaction." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 409, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 566 + ], + "type": "text", + "content": "PHC follows the conventional pipeline of generative adversarial and goal-conditioned reinforcement learning [18, 34], it receives target joint positions or rigid body's keypoints as the policy goals and outputs physically retargeted motions. The policy consists of multiple actor-critic model in a multiplicative manner. Given arbitrary action-reaction generation, we formulate the auto-regressive interaction as a standard Markov Decision Process (MDP), denoted as tuple " + }, + { + "bbox": [ + 313, + 409, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\langle S, A, \\mathcal{P}, r, \\gamma, \\rho_0 \\rangle" + }, + { + "bbox": [ + 313, + 409, + 555, + 566 + ], + "type": "text", + "content": ", representing state space, action space, transition distribution, reward, discount factor and initial states' distribution. In our settings, the MDP is also conditioned on action-reaction tuple " + }, + { + "bbox": [ + 313, + 409, + 555, + 566 + ], + "type": "inline_equation", + "content": "\\hat{z}" + }, + { + "bbox": [ + 313, + 409, + 555, + 566 + ], + "type": "text", + "content": ", the standard bellman equation of state-action value can be reformulated as:" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 346, + 574, + 554, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 574, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 346, + 574, + 554, + 605 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Q ^ {\\pi} (s, a, \\hat {z}) = \\mathbb {E} _ {s ^ {\\prime} \\sim \\mathcal {P} (\\cdot | s, a, \\hat {z})} [ r (s, a, s ^ {\\prime}, \\hat {z}) + \\tag {9} \\\\ \\left. \\gamma \\mathbb {E} _ {a ^ {\\prime} \\sim \\pi (\\cdot | s ^ {\\prime}, \\hat {z})} [ Q ^ {\\pi} (s ^ {\\prime}, a ^ {\\prime}, \\hat {z} ^ {\\prime} ] \\right], \\\\ \\end{array}", + "image_path": "73615958a82bc4c675c0dfedc5588b867c8bf29e85b522db957229d992824764.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 613, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 555, + 662 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 613, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\hat{z} = \\langle \\hat{x},\\hat{y},y_{real}\\rangle" + }, + { + "bbox": [ + 313, + 613, + 555, + 662 + ], + "type": "text", + "content": " is the augmented interaction goal, " + }, + { + "bbox": [ + 313, + 613, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\hat{x}" + }, + { + "bbox": [ + 313, + 613, + 555, + 662 + ], + "type": "text", + "content": " means a single predicted reactor frame regarded as the target goal, " + }, + { + "bbox": [ + 313, + 613, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 313, + 613, + 555, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 613, + 555, + 662 + ], + "type": "inline_equation", + "content": "y_{real}" + }, + { + "bbox": [ + 313, + 613, + 555, + 662 + ], + "type": "text", + "content": " are the actor motions coming from the diffusion model generation and real world inputs." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "Actor-aware Reaction Policy To conserve computational and communication resources while prioritizing interaction safety, we integrate a low-frequency (0.5 Hz)" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10177" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "diffusion planner with a high-frequency (30 Hz) physics tracker and real-world actor capture interface. This frequency mismatch can cause the reaction policy to pursue stale targets, elevating collision risk when the actor's motion deviates unexpectedly. To safeguard against interpenetration, we incorporate both generated and captured actor motion histories into the policy observation and heavily penalize the trajectory inconsistency in the reward function, thereby enforcing collision-free, safety-critical responses. More details of implementations in Appendix B.2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 135, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 135, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 135, + 215 + ], + "type": "text", + "content": "4. Applications" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 220, + 192, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 192, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 192, + 232 + ], + "type": "text", + "content": "4.1. Immersive VR Interface" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 238, + 296, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 369 + ], + "type": "text", + "content": "Our system is built upon Unity, enabling real-time interaction and visualization in a VR environment. Leveraging VisionPro [42], we develop a VR platform that allows the user to wear a headset and engage in immersive, egocentric interactions with the character. The camera in Unity is positioned on the actor's head, ensuring that it moves and rotates with the actor's motions. It continuously streams real-time simulations from Unity to the VR headset, providing an immersive experience. The user can perform any actions to guide the character's movement, such as completing a duet dance with the character." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 375, + 246, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 246, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 246, + 387 + ], + "type": "text", + "content": "4.2. Human-Robot Interaction Interface" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 392, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 571 + ], + "type": "text", + "content": "Benefiting from our robust Reaction Motion Tracker, the generated motions adhere to physical constraints and exhibit high adaptability. As a result, we can seamlessly transfer physically plausible humanoid interaction to human-robot interaction once we retarget the imitation goal from SMPL[37] joints to specific robot joints. An optimization-based method is leveraged following [15, 16, 24]. First we fit the SMPL shape parameters " + }, + { + "bbox": [ + 55, + 392, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 392, + 296, + 571 + ], + "type": "text", + "content": " and use " + }, + { + "bbox": [ + 55, + 392, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 392, + 296, + 571 + ], + "type": "text", + "content": " to solve for optimal robot joint angles " + }, + { + "bbox": [ + 55, + 392, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 392, + 296, + 571 + ], + "type": "text", + "content": ". Then the corresponding joints' goal positions are calculated by forward kinematics. In the VR environment, the character can be replaced with the Unitree H1[49], enabling an immersive and realistic human-robot interaction experience. Furthermore, our approach holds potential for future deployment on physical robots." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 581, + 135, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 135, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 135, + 594 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 600, + 230, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 230, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 230, + 612 + ], + "type": "text", + "content": "5.1. Datasets and Evaluation Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 294, + 712 + ], + "type": "text", + "content": "Datasets. We evaluate our model on two datasets: InterX [65], which contains 11,388 interaction SMPL-X [43] motion sequences on 40 interaction categories; InterHuman [32], which consists of 6022 interaction SMPL [37] motion sequences with around 1.7M frames. For data processing, we use the actor-reactor annotations provided by [66]. We canonicalize all the motions into the representation stated in Sec.3.3.1 and set frame rate to 30 FPS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "Evaluation Metrics. We assess the generated reactor's sequence using three key aspects: 1) the quality of the reactor's motion independent of the actor, following kinematic motion generation works [14, 58]; 2) the physical plausibility of the reactor's motion, based on physics-based methods [59, 69]; and 3) the interaction quality between the actor and reactor, as discussed in [36, 52]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "type": "text", + "content": "For 1), we evaluate using Frechet Inception Distance (FID) [17] to measure the similarity between the generated and ground truth data distributions, Diversity to assess motion variability, and Multi-modal Distance (MMDist) to measure the distance between the generated reaction and the text prompt. For 2), we assess Penetration, the average distance between the ground and the lowest body joint position below the ground, Floating, the average distance between the ground and the lowest body joint above the ground, and Skating, the average horizontal displacement during frames exhibiting skating behavior. For 3), we measure maximum Interpenetration Volume (IV) [36] of reactor vertices intersecting the actor mesh, and cross-distance metrics " + }, + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "type": "inline_equation", + "content": "FID_{cd}" + }, + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "type": "inline_equation", + "content": "Div_{cd}" + }, + { + "bbox": [ + 313, + 157, + 555, + 348 + ], + "type": "text", + "content": " in a 100-dimensional cross-distance space [52] to evaluate interaction realism. Details of metrics calculation in Appendix C.2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 359, + 447, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 447, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 447, + 371 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "text", + "content": "In our auto-regressive denoiser, we adopt the DIT architecture [44] consisting of 8 Transformer layers with a 512-dimensional hidden size. We employ a frozen DistilBERT encoder [10, 51] as the text encoder, which is widely used in motion generation tasks [14, 59]. For training and inference, we set the binary interaction field " + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\in \\mathbb{R}^{6 \\times 6}" + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "text", + "content": ", the diffusion timestep " + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "inline_equation", + "content": "T = 8" + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "text", + "content": ", the classifier-free guidance weight " + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "inline_equation", + "content": "w = 5" + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "text", + "content": ", and the foot-contact, interaction, and prefix loss weights to " + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{foot}} = 0.2" + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{inter}} = 0.5" + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{prefix}} = 0.1" + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "text", + "content": ", respectively. In the auto-regressive setting, we use an interaction history length of " + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "inline_equation", + "content": "h = 20" + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "text", + "content": " frames and a prediction window length of " + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "inline_equation", + "content": "k = 40" + }, + { + "bbox": [ + 313, + 377, + 554, + 580 + ], + "type": "text", + "content": " frames. We train for 300K denoising steps on both the InterHuman and Inter-X datasets, and observe optimal checkpoints at 250K steps for InterHuman and 220K steps for Inter-X. All models are trained on a single NVIDIA GeForce RTX 4090d GPU for approximately 30 hours." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "For physics policy training, we use the multiplicative version of PHC network with three primitives, and construct task goal with only target body keypoints. The policy is also trained on the same device and datasets as the former stage for 20 hours. The tracking policy is executed at " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "30\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": " while physics simulation is processed in NVIDIA's Isaac Gym[41] at " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "60\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": ". In the starting phase of inference, we capture the first 1 second motion from the actor, and store it in the buffer along with masked reactor motion prefix for the next planning horizon's generation, the downstream reactor tracker receives default joint positions as tracking goals." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10178" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 71, + 298, + 178 + ], + "blocks": [ + { + "bbox": [ + 66, + 71, + 298, + 178 + ], + "lines": [ + { + "bbox": [ + 66, + 71, + 298, + 178 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 298, + 178 + ], + "type": "image", + "image_path": "1d714e11b8df8ecffa32406a56a9227abf0c06487dc8b8bf4741f2bfc9914ae5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 198, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 54, + 198, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 555, + 222 + ], + "type": "text", + "content": "Figure 3. Compared to CAMDM (top row), Human-X (bottom row) achieves more complete hand contact in tasks such as face-hitting and handshaking. Additionally, its foot movement appears more natural, as highlighted in the red and green circles." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 317, + 76, + 548, + 178 + ], + "blocks": [ + { + "bbox": [ + 317, + 76, + 548, + 178 + ], + "lines": [ + { + "bbox": [ + 317, + 76, + 548, + 178 + ], + "spans": [ + { + "bbox": [ + 317, + 76, + 548, + 178 + ], + "type": "image", + "image_path": "411b3599f11e91dabc4ffa4f42c86d50eda391247161d130da4f0a9f47982e93.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 230, + 545, + 315 + ], + "blocks": [ + { + "bbox": [ + 57, + 230, + 545, + 315 + ], + "lines": [ + { + "bbox": [ + 57, + 230, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 57, + 230, + 545, + 315 + ], + "type": "image", + "image_path": "33d889fc7db82c7f5e2a32e2ebb7fbe0b27642134625d49d4992158ba51299c0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 323, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 54, + 323, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 555, + 357 + ], + "type": "text", + "content": "Figure 4. Visualization results on Human-Robot Interaction. The robot (black skeleton) and human (orange mesh) perform a handshake on a flat plane, from arm extension and palm contact through to shake completion, illustrating our method's spatial coordination and motion coherence." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 377, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 296, + 426 + ], + "type": "text", + "content": "After warming-up, the diffusion planner generates reasonable action-reaction motion sequences, the reactor is able to continuously follow the planning and receives both real and generated actor sequences for policy correction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 436, + 216, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 216, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 216, + 449 + ], + "type": "text", + "content": "5.3. Actor-To-Reactor Evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 454, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 610 + ], + "type": "text", + "content": "Baselines. To evaluate the performance of our method, we adopt and reproduce several human reaction generation models as baselines: 1) InterFormer [8]: a transformer-based model with temporal and spatial attention for human reaction generation, 2) InterGen [32]: a diffusion-based model for generating human interaction motions conditioned on text prompts., 3) ReGenNet [66]: a diffusion-based model for generating human reactions conditioned on given human actions. and 4) CAMDM [7]: a transformer-based Conditional Auto-regressive Motion Diffusion Model. However, since it is designed for single-person motion generation, we utilize the action as a condition input to generate the corresponding reaction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Comparisons. As demonstrated in Tab. 2, in the online unconstrained reaction setting, our method exhibits significant improvements over baselines in Reaction Quality, Physical Plausibility, and Interaction Quality on both the Inter-X datasets. Unlike InterFormer's transformer architecture, our diffusion-based approach effectively captures the nuanced structure of data distributions, resulting in re" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 377, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 521 + ], + "type": "text", + "content": "actor motions that closely align with real-world motion patterns and thereby enhancing reaction quality. In comparison to InterGen and ReGenNet, our implementation of Auto-regressive Motion Diffusion Models, augmented with a Physics Tracker, enables real-time interaction with the reactor by predicting future frames. This ensures that the generated motions adhere to fundamental physical laws, producing more natural and believable animations. Furthermore, relative to CAMDM, we introduce an Interaction Loss, leading to substantial improvements in IV and " + }, + { + "bbox": [ + 313, + 377, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{cd}" + }, + { + "bbox": [ + 313, + 377, + 555, + 521 + ], + "type": "text", + "content": ", thereby facilitating more natural interactions. The experiments on the InterHuman dataset can be found in C.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 535, + 408, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 408, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 408, + 548 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 555, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 592 + ], + "type": "text", + "content": "We primarily conduct ablation studies on three aspects: Losses, Timesteps, and Horizon. For more experiments, please refer to Section C.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Losses. To verify the effectiveness of our designed loss function, we ablate " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{inter}" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{foot}" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{prefix}" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " separately. As shown in Tab. 3, removing any of them degrades performance. Notably, excluding " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{inter}" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " significantly increases " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "FID" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "FID_{cd}" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": ", underscoring its importance for generating realistic interactions. Removing " + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{foot}" + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": " raises Skating and Floating scores, showing its role in reducing foot sliding and enhancing physical plausibility." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10179" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 68, + 552, + 213 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 552, + 213 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 552, + 213 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 552, + 213 + ], + "type": "table", + "html": "
MethodsReactionPhysicsInteraction
FID↓Div.→MMDist. ↓Pene. ↓Skate. ↓Float. ↓IV ↓FIDcd ↓Divcd →
Ground Truth0.0026.0283.5240.0000.0237.9560.0240.23511.471
InterFormer [8]6.1504.9228.6480.5240.17524.1090.3326.8258.214
InterGen [32]5.4854.4176.3890.2700.14311.9150.2693.4378.730
ReGenNet [66]2.2125.0475.2350.1410.1189.8100.2181.9509.248
CAMDM [7]1.4305.6575.3180.1530.1279.8640.1452.0869.017
Human-X0.9756.0634.1150.1180.0928.6500.0761.6929.735
Human-X*1.2205.1405.0760.0260.0074.3420.0541.8338.565
", + "image_path": "4efe20ad26c42fcf971737ce37e4f66e8bdf246125673cc0e5ce03df1d3184a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 252, + 553, + 439 + ], + "blocks": [ + { + "bbox": [ + 56, + 220, + 555, + 245 + ], + "lines": [ + { + "bbox": [ + 56, + 220, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 555, + 245 + ], + "type": "text", + "content": "Table 2. Action-to-Reaction with online unconstrained reaction setting on Inter-X [65] dataset. A higher or lower value is better for ↑ or ↓, and → means the value closer to ground truth is better. * denotes the method with reaction policy designed for humanoid reaction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 252, + 553, + 439 + ], + "lines": [ + { + "bbox": [ + 58, + 252, + 553, + 439 + ], + "spans": [ + { + "bbox": [ + 58, + 252, + 553, + 439 + ], + "type": "table", + "html": "
ClassSettingsReactionPhysicsInteractionLatency (ms)
FID↓Div.→MMDist.→Penetration ↓Skating ↓Floating ↓IV ↓FIDcd ↓Divcd→
Ground Truth0.0026.0283.5240.0000.0237.9560.0240.23511.471-
Lossesw.o. Linter1.4576.1544.2640.1340.1078.7150.1192.2089.91713.6
w.o. Lfoot1.0096.1144.2050.1270.2399.1200.0941.6879.84913.6
w.o. Lprefix1.0236.1284.2250.1330.1138.8570.1001.6979.86513.6
Timesteps21.8095.5175.0070.1440.1139.0080.1532.0087.5133.8
51.1076.2634.5140.1280.1088.9080.1151.8078.0038.2
100.9886.0084.2080.1180.0968.7090.0971.6849.85316.9
1000.9085.9073.9630.1130.0938.6080.0871.6109.708170
Horizonh=10,k=301.0556.1064.1070.1230.1038.8050.1031.6879.8079.6
h=15,k=300.9876.0054.0060.1180.0978.7070.0981.6679.78710.1
h=20,k=200.9575.9563.9580.1130.0948.6580.0931.6269.7569.5
h=15,k=451.0066.0544.1050.1230.1048.8060.1021.6889.80913.6
Human-Xtext0.9265.9513.9090.1120.0878.2180.0921.6079.82213.6
unconstrained0.9756.0634.1150.1180.0928.6500.0761.6929.73513.6
", + "image_path": "082391c46ba6e6105c7e7d3851d8332b6f2394aaa80003e0f34fabd882c69a9b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 159, + 445, + 449, + 457 + ], + "lines": [ + { + "bbox": [ + 159, + 445, + 449, + 457 + ], + "spans": [ + { + "bbox": [ + 159, + 445, + 449, + 457 + ], + "type": "text", + "content": "Table 3. Ablation studies of online reaction setting on the Inter-X [65] dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 479, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 586 + ], + "type": "text", + "content": "Timesteps. During the denoising process, we experiment with different sampling timesteps: 2, 5, 10, and 100. As shown in the Tab. 3, increasing the number of timesteps improves the model's performance; however, it also leads to higher latency. When timesteps = 10, the model has already achieved satisfactory results. Considering both generation quality and latency, we ultimately select timesteps = 8 as a trade-off, ensuring high-quality generation while meeting real-time requirements." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "text", + "content": "Horizon. During inference process, the model takes a continuous sequence of " + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "text", + "content": " history frames as input to predict " + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "text", + "content": " future frames. We conduct experiments with different values of " + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "text", + "content": ", and the results in Tab. 3 show that the best performance is achieved when " + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "inline_equation", + "content": "h = 20" + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "inline_equation", + "content": "k = 40" + }, + { + "bbox": [ + 55, + 600, + 295, + 660 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 670, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 128, + 681 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "We present Human-X Interaction, a novel framework for real-time human interaction synthesis that seamlessly in-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 479, + 555, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 706 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 706 + ], + "type": "text", + "content": "tegrates autoregressive diffusion modeling with an actor-aware reaction policy. Our approach addresses the critical challenges of real-time synchronization, physical realism, and safety in human-ayar, human-humanoid and humanrobot interactions. Specifically, Human-X jointly predicts actions and reactions using both partner's motion histories, ensuring immersive and realistic motion synthesis. By incorporating comprehensive interaction-aware rewards and training the tracking policy with reinforcement learning, we achieve physically feasible and artifact-free interactions even in dynamic and safety-critical scenarios. Extensive experiments on the Inter-X and InterHuman datasets demonstrate significant improvements in motion quality, interaction continuity, and physical accuracy over existing methods. The current formulation focuses on synthesizing reactive motions and does not account for collaborative task completion between humans and machines. Extending the framework to support joint task execution with shared goals remains an open challenge." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10180" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 296, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 296, + 213 + ], + "type": "text", + "content": "This work was supported by Shanghai Local College Capacity Building Program(23010503100), NSFC (No.62406195), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (Shanghai), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (ShanghaiTech University), HPC Platform and Core Facility Platform of Computer Science and Communication of ShanghaiTech University and Shanghai Engineering Research Center of Intelligent Vision and Imaging." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 226, + 115, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 226, + 115, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 226, + 115, + 239 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 247, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 247, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 247, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 61, + 247, + 296, + 300 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 303, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 303, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 303, + 296, + 346 + ], + "type": "text", + "content": "[2] Simon Alexanderson, Rajmund Nagy, Jonas Beskow, and Gustav Eje Henter. Listen, denoise, action! audio-driven motion synthesis with diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-20, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 350, + 295, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 350, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 61, + 350, + 295, + 393 + ], + "type": "text", + "content": "[3] Samy Bengio, Oriol Vinyals, Navdeep Jaitly, and Noam Shazeer. Scheduled sampling for sequence prediction with recurrent neural networks. Advances in neural information processing systems, 28, 2015. 5, 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 396, + 296, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 296, + 460 + ], + "type": "text", + "content": "[4] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, et al. Digital life project: Autonomous 3d characters with social intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 582-592, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 463, + 296, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 463, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 463, + 296, + 518 + ], + "type": "text", + "content": "[5] Haoran Chen, Yiteng Xu, Yiming Ren, Yaoqin Ye, Xinran Li, Ning Ding, Peishan Cong, Ziyi Wang, Bushi Liu, Yuhan Chen, et al. Symbiosim: Human-in-the-loop simulation platform for bidirectional continuing learning in human-robot interaction. arXiv preprint arXiv:2502.07358, 2025. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 520, + 296, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 520, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 296, + 574 + ], + "type": "text", + "content": "[6] Jingxiao Chen, Xinyao Li, Jiahang Cao, Zhengbang Zhu, Wentao Dong, Minghuan Liu, Ying Wen, Yong Yu, Liqing Zhang, and Weinan Zhang. Rhino: Learning real-time humanoid-human-object interaction from human demonstrations. arXiv preprint arXiv:2502.13134, 2025. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 577, + 296, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 577, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 577, + 296, + 621 + ], + "type": "text", + "content": "[7] Rui Chen, Mingyi Shi, Shaoli Huang, Ping Tan, Taku Komura, and Xuelin Chen. Taming diffusion probabilistic models for character control. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 7, 8, 3, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 624, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 296, + 666 + ], + "type": "text", + "content": "[8] Baptiste Chopin, Hao Tang, Naima Otberdout, Mohamed Daoudi, and Nicu Sebe. Interaction transformer for human reaction generation. IEEE Transactions on Multimedia, 25: 8842-8854, 2023. 7, 8, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "text", + "content": "[9] Zekai Deng, Ye Shi, Kaiyang Ji, Lan Xu, Shaoli Huang, and Jingya Wang. Human-object interaction via automatically designed vlm-guided motion policy. arXiv preprint arXiv:2503.18349, 2025. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 148 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 151, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 554, + 217 + ], + "type": "text", + "content": "[11] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Thompson, Quan Vuong, Tianhe Yu, et al. Palm- e: an embodied multimodal language model. In Proceedings of the 40th International Conference on Machine Learning, pages 8469-8488, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 219, + 554, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 284 + ], + "type": "text", + "content": "[12] Yuming Du, Robin Kips, Albert Pumarola, Sebastian Starke, Ali Thabet, and Artsiom Sanakoyeu. Avatars grow legs: Generating smooth human motion from sparse tracking inputs with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 481-490, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 285, + 554, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 554, + 329 + ], + "type": "text", + "content": "[13] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 331, + 554, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 385 + ], + "type": "text", + "content": "[14] Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng. Generating diverse and natural 3d human motions from text. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5152-5161, 2022. 4, 6, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 387, + 554, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 554, + 441 + ], + "type": "text", + "content": "[15] Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. arXiv preprint arXiv:2406.08858, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 443, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 554, + 475 + ], + "type": "text", + "content": "[16] Tairan He, Zhengyi Luo, Wenli Xiao, Chong Zhang, Kris Kitani, Changliu Liu, and Guanya Shi. Learning human-to-humanoid real-time whole-body teleoperation, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 476, + 554, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 554, + 531 + ], + "type": "text", + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 533, + 554, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 554, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 554, + 565 + ], + "type": "text", + "content": "[18] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 567, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 554, + 589 + ], + "type": "text", + "content": "[19] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 590, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 623 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "type": "text", + "content": "[21] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 712 + ], + "type": "text", + "content": "[22] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10181" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Vision and Pattern Recognition, pages 16750-16761, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[23] Muhammad Gohar Javed, Chuan Guo, Li Cheng, and Xingyu Li. Intermask: 3d human interaction generation via collaborative masked modelling. arXiv preprint arXiv:2410.10010, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 186 + ], + "type": "text", + "content": "[24] Mazeyu Ji, Xuanbin Peng, Fangchen Liu, Jialong Li, Ge Yang, Xuxin Cheng, and Xiaolong Wang. Exbody2: Advanced expressive humanoid whole-body control. arXiv preprint arXiv:2412.13196, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 188, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 294, + 232 + ], + "type": "text", + "content": "[25] Biao Jiang, Xin Chen, Wen Liu, Jingyi Yu, Gang Yu, and Tao Chen. Motiongpt: Human motion as a foreign language. Advances in Neural Information Processing Systems, 36:20067-20079, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 294, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 299 + ], + "type": "text", + "content": "[26] Jianping Jiang, Weiye Xiao, Zhengyu Lin, Huaizhong Zhang, Tianxiang Ren, Yang Gao, Zhiqian Lin, Zhongang Cai, Lei Yang, and Ziwei Liu. Solami: Social vision-language-action modeling for immersive interaction with 3d autonomous characters. arXiv preprint arXiv:2412.00174, 2024. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 301, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 294, + 345 + ], + "type": "text", + "content": "[27] Wentao Jiang, Jingya Wang, Kaiyang Ji, Baoxiong Jia, Siyuan Huang, and Ye Shi. Arflow: Human action-reaction flow matching with physical guidance. arXiv preprint arXiv:2503.16973, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 347, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 347, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 347, + 294, + 369 + ], + "type": "text", + "content": "[28] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes, 2013. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 415 + ], + "type": "text", + "content": "[29] Jiaman Li, Karen Liu, and Jiajun Wu. Ego-body pose estimation via ego-head pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17142-17151, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 417, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 294, + 460 + ], + "type": "text", + "content": "[30] Jiefeng Li, Siyuan Bian, Chao Xu, Zhicun Chen, Lixin Yang, and Cewu Lu. Hybrid analytical-neural inverse kinematics for whole-body mesh recovery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 462, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 294, + 517 + ], + "type": "text", + "content": "[31] Ronghui Li, Youliang Zhang, Yachao Zhang, Yuxiang Zhang, Mingyang Su, Jie Guo, Ziwei Liu, Yebin Liu, and Xiu Li. Interdance: Reactive 3d dance generation with realistic duet interactions. arXiv preprint arXiv:2412.16982, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 519, + 295, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 519, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 519, + 295, + 563 + ], + "type": "text", + "content": "[32] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, pages 1-21, 2024. 2, 6, 7, 8, 3, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 565, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 565, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 565, + 295, + 609 + ], + "type": "text", + "content": "[33] Jing Lin, Yao Feng, Weiyang Liu, and Michael J Black. Chathuman: Language-driven 3d human understanding with retrieval-augmented tool reasoning. arXiv preprint arXiv:2405.04533, 2024. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 610, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 295, + 643 + ], + "type": "text", + "content": "[34] Minghuan Liu, Menghui Zhu, and Weinan Zhang. Goal-conditioned reinforcement learning: Problems and solutions, 2022. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 644, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 644, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 644, + 295, + 689 + ], + "type": "text", + "content": "[35] Yunze Liu, Changxi Chen, and Li Yi. Interactive humanoid: Online full-body motion reaction synthesis with social affordance canonicalization and forecasting. arXiv preprint arXiv:2312.08983, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 690, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 295, + 713 + ], + "type": "text", + "content": "[36] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-reaction: Physically plausible real-time humanoid reaction" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "synthesis via forward dynamics guided 4d imitation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3771-3780, 2024. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 149 + ], + "type": "text", + "content": "[37] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM Transactions on Graphics, 34(6), 2015. 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[38] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10895-10904, 2023. 3, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "text", + "content": "[39] Zhengyi Luo, Jinkun Cao, Josh Merel, Alexander Winkler, Jing Huang, Kris Kitani, and Weipeng Xu. Universal humanoid motion representations for physics-based control. arXiv preprint arXiv:2310.04582, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "type": "text", + "content": "[40] Zhengyi Luo, Jinkun Cao, Rawal Khirodkar, Alexander Winkler, Kris Kitani, and Weipeng Xu. Real-time simulated avatar from head-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 571-581, 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 294, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 349 + ], + "type": "text", + "content": "[41] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 4, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 370 + ], + "type": "text", + "content": "[42] Younghyo Park and Pulkit Agrawal. Using apple vision pro to train and control robots, 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 372, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 437 + ], + "type": "text", + "content": "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 480 + ], + "type": "text", + "content": "[44] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195-4205, 2023. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 525 + ], + "type": "text", + "content": "[45] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel Van de Panne. Deepmimic: Example-guided deep reinforcement learning of physics-based character skills. ACM Transactions On Graphics (TOG), 37(4):1-14, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "text", + "content": "[46] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics (ToG), 40(4):1-20, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 571, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 635 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PmLR, 2021. 4, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 636, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 691 + ], + "type": "text", + "content": "[48] Davis Rempe, Tolga Birdal, Aaron Hertzmann, Jimei Yang, Srinath Sridhar, and Leonidas J Guibas. Humor: 3d human motion model for robust pose estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11488-11499, 2021. 5, 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 551, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 551, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 551, + 713 + ], + "type": "text", + "content": "[49] Unitree Robotics. Unitree h1 robot. https://www.unitree.com, 2021. Accessed: 2023-10-10. 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10182" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[50] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), 2017. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[51] Victor Sanh, Lysandre Debut, Julien Chaumont, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019. 6, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[52] Li Siyao, Tianpei Gu, Zhitao Yang, Zhengyu Lin, Ziwei Liu, Henghui Ding, Lei Yang, and Chen Change Loy. Duolando: Follower gpt with off-policy reinforcement learning for dance accompaniment. arXiv preprint arXiv:2403.18811, 2024. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "text", + "content": "[53] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "text", + "content": "[54] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Transactions on Graphics (TOG), 39 (4):54-1, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "type": "text", + "content": "[55] Wenhui Tan, Boyuan Li, Chuhao Jin, Wenbing Huang, Xiting Wang, and Ruihua Song. Think then react: Towards unconstrained action-to-reaction motion generation. In The Thirteenth International Conference on Learning Representations, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "type": "text", + "content": "[56] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15999-16009, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[57] Jiangnan Tang, Jingya Wang, Kaiyang Ji, Lan Xu, Jingyi Yu, and Ye Shi. A unified diffusion framework for scene-aware human motion estimation from sparse signals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21251-21262, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "text", + "content": "[58] Guy Tevet, Sigal Raab, Brian Gordon, Yoni Shafir, Daniel Cohen-or, and Amit Haim Bermano. Human motion diffusion model. In The Eleventh International Conference on Learning Representations, 2023. 2, 4, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 554 + ], + "type": "text", + "content": "[59] Guy Tevet, Sigal Raab, Setareh Cohan, Daniele Reda, Zhengyi Luo, Xue Bin Peng, Amit H Bermano, and Michiel van de Panne. Closd: Closing the loop between simulation and diffusion for multi-task character control. arXiv preprint arXiv:2410.03441, 2024. 4, 6, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 600 + ], + "type": "text", + "content": "[60] Jie Tian, Ran Ji, Lingxiao Yang, Suting Ni, Yuexin Ma, Lan Xu, Jingyi Yu, Ye Shi, and Jingya Wang. Gaze-guided hand-object interaction synthesis: Dataset and method. arXiv preprint arXiv:2403.16169, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "text", + "content": "[61] Jonathan Tseng, Rodrigo Castellon, and Karen Liu. Edge: Editable dance generation from music. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 448-458, 2023. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[62] Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. Thor: Text to human-object interaction diffusion via relation intervention. arXiv preprint arXiv:2403.11208, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "text", + "content": "[63] Zhen Wu, Jiaman Li, and C Karen Liu. Human-object interaction from human-level instructions. CoRR, 2024. 2, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "text", + "content": "[64] Liang Xu, Ziyang Song, Dongliang Wang, Jing Su, Zhicheng Fang, Chenjing Ding, Weihao Gan, Yichao Yan, Xin Jin, Xiaokang Yang, et al. Actformer: A gan-based transformer towards general action-conditioned 3d human motion generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2228-2238, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 205 + ], + "type": "text", + "content": "[65] Liang Xu, Xintao Lv, Yichao Yan, Xin Jin, Shuwen Wu, Congsheng Xu, Yifan Liu, Yizhou Zhou, Fengyun Rao, Xingdong Sheng, et al. Inter-x: Towards versatile human-human interaction analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22260-22271, 2024. 2, 6, 8, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 555, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 555, + 259 + ], + "type": "text", + "content": "[66] Liang Xu, Yizhou Zhou, Yichao Yan, Xin Jin, Wenhan Zhu, Fengyun Rao, Xiaokang Yang, and Wenjun Zeng. Regennet: Towards human action-reaction synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1759-1769, 2024. 2, 6, 7, 8, 3, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "type": "text", + "content": "[67] Lingxiao Yang, Shutong Ding, Yifan Cai, Jingyi Yu, Jingya Wang, and Ye Shi. Guidance with spherical gaussian constraint for conditional diffusion. In Proceedings of the 41st International Conference on Machine Learning, pages 56071-56095, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "text", + "content": "[68] Ye Yuan, Viktor Makoviychuk, Y Guo, S Fidler, XB Peng, and K Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Trans. Graph, 42(4), 2023. 5, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 360, + 553, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 413 + ], + "type": "text", + "content": "[69] Ye Yuan, Jiaming Song, Umar Iqbal, Arash Vahdat, and Jan Kautz. Physdiff: Physics-guided human motion diffusion model. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16010-16021, 2023. 3, 6, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "text", + "content": "[70] Mingyuan Zhang, Xinying Guo, Liang Pan, Zhongang Cai, Fangzhou Hong, Huirong Li, Lei Yang, and Ziwei Liu. Remodiffuse: Retrieval-augmented motion diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 364-373, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "text", + "content": "[71] Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text-driven human motion generation with diffusion model. IEEE transactions on pattern analysis and machine intelligence, 46(6):4115-4128, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "text", + "content": "[72] Zhenhao Zhang, Ye Shi, Lingxiao Yang, Suting Ni, Qi Ye, and Jingya Wang. Openhoi: Open-world hand-object interaction synthesis with multimodal large language model. arXiv preprint arXiv:2505.18947, 2025. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 571, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 612 + ], + "type": "text", + "content": "[73] Kaifeng Zhao, Gen Li, and Siyu Tang. Dart: A diffusion-based autoregressive motion model for real-time text-driven motion control. arXiv preprint arXiv:2410.05260, 2024. 5, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 667 + ], + "type": "text", + "content": "[74] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745-5753, 2019. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[75] Kaizhen Zhu, Mokai Pan, Yuexin Ma, Yanwei Fu, Jingyi Yu, Jingya Wang, and Ye Shi. Unidb: A unified diffusion bridge framework via stochastic optimal control. arXiv preprint arXiv:2502.05749, 2025. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10183" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/d546fd0e-5080-4957-803b-5e64e0ab9d88_content_list.json b/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/d546fd0e-5080-4957-803b-5e64e0ab9d88_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..63f84b58ab3e4d047a62eb04cc7f1f947812d975 --- /dev/null +++ b/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/d546fd0e-5080-4957-803b-5e64e0ab9d88_content_list.json @@ -0,0 +1,1750 @@ +[ + { + "type": "text", + "text": "Towards Long-Horizon Vision-Language-Action System: Reasoning, Acting and Memory", + "text_level": 1, + "bbox": [ + 210, + 130, + 785, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daixun Li $^{1,\\ast}$ Yusi Zhang $^{1,\\ast}$ Mingxiang Cao $^{1,\\ast}$ Donglai Liu $^{1,\\ast}$ Weiying Xie $^{1,\\dagger}$ Tianlin Hui $^{1}$ Lunkai Lin $^{2}$ Zhiqiang Xie $^{2}$ Yunsong Li $^{1}$ Xidian University AgileX Robotics https://xdu-imagelab.github.io/MindExplore/", + "bbox": [ + 99, + 202, + 901, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2b8e8cd616ca0d87633ab0c999cf764ddc1b3d81606435c0d34f3aacc901680.jpg", + "image_caption": [ + "MindExplore System" + ], + "image_footnote": [], + "bbox": [ + 94, + 309, + 359, + 445 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/580b8f6b236f964b949835490d7766a6f350b35aa591d569a1f12538b10ffa05.jpg", + "image_caption": [ + "Reasoning Layer Validation" + ], + "image_footnote": [], + "bbox": [ + 357, + 321, + 545, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-Modal: RGB Depth LiDAR", + "bbox": [ + 568, + 327, + 756, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a45ccc13860191c2543a251c797a81ba6a21cd343588982b7ad9d45d7c350daf.jpg", + "image_caption": [ + "Dataset", + "SandGo-1k" + ], + "image_footnote": [], + "bbox": [ + 560, + 344, + 761, + 443 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chain of Thought", + "bbox": [ + 784, + 328, + 883, + 340 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6cf4b1f4d3bff592f7bcc408a0177435794eedba6912bd8410b0d7dc87ec80f.jpg", + "image_caption": [ + "SandThink-21k" + ], + "image_footnote": [], + "bbox": [ + 785, + 344, + 874, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 794, + 401, + 870, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 792, + 409, + 870, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 810, + 416, + 857, + 422 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 810, + 422, + 857, + 428 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "...", + "bbox": [ + 797, + 430, + 867, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d5b012abac6000a0afacd869a2b11b55313b02f553109f23032baa101ba044eb.jpg", + "image_caption": [ + "Zero-Shot Tasks", + "Perceive seen & unseen object", + "Figure 1. Overview of MindExplore System: A hierarchical embodied intelligence system including reasoning, acting, and memory, offering state-of-the-art generalizability in highly dynamic scenes. Unlike conventional planar embodied intelligence models that consolidate all capabilities within a single layer, MindExplore introduces a novel expert-level hierarchical embodied system architecture." + ], + "image_footnote": [], + "bbox": [ + 104, + 465, + 250, + 563 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/340f338f307043fc14744f17129e7ccb0033a656d1266646192701cbcbd217d3.jpg", + "image_caption": [ + "Long-Horizon Tasks", + "Complete multiple meta-actions" + ], + "image_footnote": [], + "bbox": [ + 285, + 465, + 431, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2474028458195b74b26fc52465a47bf8f03e9869ededa24315db589e28c430ab.jpg", + "image_caption": [ + "Cross-Expert Tasks", + "Adapt to the dynamic sand scene" + ], + "image_footnote": [], + "bbox": [ + 459, + 465, + 645, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7045b82717b103a041edd1bbcb38f8d52e653e54cc867b4bd151873aecac3ef.jpg", + "image_caption": [ + "Instruction Following Tasks", + "Complete long-horizon tasks as instructed" + ], + "image_footnote": [], + "bbox": [ + 674, + 465, + 893, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 635, + 326, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-Language-Action (VLA) is crucial for autonomous decision-making in embodied systems. While current methods have advanced single-skill abilities, their short-horizon capability limits applicability in real-world scenarios. To address this challenge, we innovatively propose MindExplore, a general hierarchical VLA system with cross-skill for long-horizon tasks in highly dynamic sand. The key insight is to iteratively align the knowledge domain of task planning and action execution. Thus, this task-oriented action enables outstanding generalization across a wide range of real-world scenarios. In the reasoning layer, task-specific chains of thought (CoT) are designed for planning long-horizon task sequences and providing meta-action signals. In the acting layer, a simple but powerful Mixture of Pol", + "bbox": [ + 88, + 666, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "icy Experts strategy is built inspired by signals and multimodal inputs for adaptively selecting skill experts and generating closed-loop action sequences. Also, it integrates a lightweight Multimodal Diffusion Policy (MMDP) to enhance spatial perception by fusing multi-visual modality features. Besides, the pioneering memory mechanism establishes feedback between the reasoning and acting layers, facilitating adaptive execution of long-horizon tasks and real-time replanning. Notably, we create SandGo-1k and SandThink-21k, the first expert-level multimodal embodied dataset and CoT dataset tailored for sandy environments. At a high execution frequency of 30 FPS, MindExplore is $3.01 \\times$ more successful than existing methods in unstructured and dynamic environments.", + "bbox": [ + 511, + 636, + 908, + 848 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 114, + 886, + 228, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors.", + "bbox": [ + 232, + 887, + 379, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6839", + "bbox": [ + 482, + 943, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current Visual-Language-Action (VLA) data and methods mainly focus on short sequence tasks of single ability [6, 20], that is, single-skill sequences limited to either arm operations or base movements, making them effective for simple tasks requiring only \"muscle memory\". [13, 20, 24]. However, these approaches fail to generalize for long-horizon tasks in real-world scenarios, particularly in complex terrain action [14, 33]. This requires necessitates the agent to handle long-horizon and multi-skill instructions in highly dynamic environments [37]. Thus, successful execution in such settings needs deep consideration—Reasoning, precise control—Acting, and adaptive feedback—Memory [5].", + "bbox": [ + 89, + 114, + 480, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve this, three key challenges must be addressed: (1) Task-oriented Chain of Thought (CoT) reasoning requires the hierarchical decomposition of coarse-grained human instructions into executable sequences of metaactions, ensuring long-horizon alignment in dynamic environments [10, 20, 23]; (2) High-Precision Multimodal Control through a lightweight spatial perception framework fuses RGB, depth, and LiDAR inputs for robust motion calibration and manipulation accuracy [17, 35]; (3) Adaptive State Tracking that dynamically memorizes system states and task progress, enabling real-time closed-loop feedback between planning and acting to mitigate error propagation from perceptual drift or control misalignment [1, 43]. Consequently, this remains an open and worthwhile challenge so far.", + "bbox": [ + 89, + 311, + 482, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose MindExplore for long-horizon tasks in complex and dynamic environments with a novel expert-level hierarchical embodied system architecture (see Figure 1). Our MindExplore is inspired by the discovery that continuous iterative alignment of the knowledge domain between task planning and action execution enables strong generalization across diverse real-world scenarios. On one hand we pursue task-oriented CoT in the Reasoning layer for effectively capturing meta-action sequences, while on the other hand we design a straightforward Mixture of Policy Experts (MoPE) strategy in the Acting Layer for flexibly generating closed-loop action sequences. Moreover, Memory Mechanism is designed to establish feedback between the reasoning and acting layers. Thus, the proposed MindExplore formulates long-horizon task execution into trainable skill expert policies, providing innovative guidance for hierarchical embodied intelligence systems without failures in highly dynamic environments. We also introduce a lightweight Multimodal Diffusion Policy (MMDP) that is dedicated to integrating RGB, depth and LiDAR data into a feature embedding space, thus enhancing spatial perception and motion control capabilities. To accurately evaluate complex tasks, we construct SandGo-1k and SandThink-21k as the first expert-level multimodal", + "bbox": [ + 89, + 537, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "embodied dataset and CoT dataset, respectively. We also carry out experiments to show that, the system maintains high accuracy and efficient execution even in highly dynamic environments such as mobile and unstructured sand, which validates our architecture. Our key contributions are summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose MindExplore, a novel expert-level hierarchical embodied system to adapt long-horizon tasks in unstructured and dynamic environments. To our best knowledge, MindExplore is the first work to formulate Reasoning, Acting and Memory into a unified architecture.", + "- We create SandGo-1k, the first embodied intelligence dataset for mobile operations in complex environments. It includes text-vision (RGB, depth, and LiDAR) data covering complex movement episodes. Building on this, we further construct a multimodal CoT dataset for highly dynamic terrains, SandThink-21k.", + "- We empirically evaluate the accuracy on real-world experiments across diverse complex terrains with 24 distinct tasks, such as steep slopes, pits, and dunes. These experiments highlight MindExplore exceptional performance for long-horizon tasks in real-world scenarios." + ], + "bbox": [ + 511, + 185, + 903, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 453, + 653, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Visual Chain-of-Thought", + "text_level": 1, + "bbox": [ + 511, + 481, + 741, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CoT reasoning [32] has become fundamental in MLLMs [40, 41], with applications in science [21], mathematics [22], and advanced question answering [36]. Recently, visual CoT has been applied to robotics [19, 29, 37] to tackle task-specific challenges. However, existing methods focus on single-instruction, single-step tasks, failing in complex scenarios that demand long-horizon task decomposition. We argue that orchestrating concise instructions into structured long-horizon tasks is a critical step toward autonomous robot decision-making.", + "bbox": [ + 511, + 507, + 906, + 659 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Embodied System", + "text_level": 1, + "bbox": [ + 511, + 681, + 689, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The long-horizon tasks in open-world environments has been a major challenge in robotics research [9]. Numerous studies leverage MLLMs with extensive world knowledge to decompose instructions into subtasks, which are executed by lightweight VLA models [4, 16, 24, 28]. The key challenges can be summarized as follows:", + "bbox": [ + 511, + 707, + 903, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Challenge 1: How can instructions be expanded into long-horizon action sequences and executed flawlessly?", + "bbox": [ + 511, + 801, + 903, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Challenge 2: How to effectively accomplish movement and manipulation in complex and dynamic terrain?", + "bbox": [ + 511, + 835, + 903, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Challenge 3: How to overcome the data missing caused by the difficulty of simulation of highly dynamic scenes?", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6840", + "bbox": [ + 482, + 943, + 514, + 954 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. The System Dataset", + "text_level": 1, + "bbox": [ + 89, + 89, + 281, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address Challenge 3, we collect a multi-task embodied operation dataset SandGo-1k, using a tracked mobile robot with a single-arm manipulator (as shown in Figure 2) in a real sand. Based on SandGo-1k, we generate a reasoning-acting embodied instruction dataset SandThink-21k.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Embodied Dataset: SandGo-1k", + "text_level": 1, + "bbox": [ + 89, + 203, + 366, + 219 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c97873e0f253848a25ba07e6c35fc10895d2cc9f7223f7317b3a5e313fabdf8d.jpg", + "image_caption": [ + "Figure 2. Left: AgileX Robotic arm degrees of freedom and environmental layout (rugged terrain, obstacles, pits, dunes). Middle: Base movement states. Right: Collected data content." + ], + "image_footnote": [], + "bbox": [ + 94, + 237, + 480, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Environmental Layout. In real-world tasks [26, 30, 31, 42], uneven sandy terrains dominate outdoor applications, making them essential for studying mobile manipulation systems. The high dynamic fluidity of sand demands adaptive adjustments, while its visual noise complicates perception and calibration. Enhancing generalization in unstructured environments remains a key challenge, and our experiments show that a system robust in sandy conditions can more easily generalize to simpler terrains like factory floors, asphalt roads, or regular soil. Therefore, to investigate the capabilities of mobile manipulation systems in highly dynamic and complex environments, we select sand as the primary experimental setting, including potholes, steep slopes, mounds, and rough roads, and construct a $6 \\times 8$ meters real-world sandbox for data collection and evaluation, as depicted in Figure 2.", + "bbox": [ + 88, + 491, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Task Setup and Collection Process. We design three long-horizon meta-tasks: (1) Cruising and Obstacle Avoidance, training the model for environmental perception and motion planning in complex terrains; (2) Sampling, enabling the model to perform instruction-following, target localization, and coordinated base-arm movement; (3) Writing, enhancing the model's fine manipulation capabilities. Based on these, we can decompose and reassemble them into five meta-actions: (1) Moving to a Target Location, (2) Crossing Obstacles, (3) Grasping a Specified Object, (4) Placing the Object into a designated Container, and (5)", + "bbox": [ + 89, + 734, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Writing. During data collection, task instructions are manually written and decomposed into sequences of meta-actions with individual instructions. Each meta-action is saved as an episode, serving as the smallest unit of data collection.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data Content. The collected data includes: (1) Sensor information, consisting of RGB-Depth images from the two cameras and LiDAR point clouds; (2) Robot state and manipulation commands, including the arm's pose and velocity, as well as the base's angular and linear velocities; (3) Text instructions for long-horizon tasks and meta-actions.", + "bbox": [ + 511, + 154, + 906, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. CoT Dataset: SandThink-21k", + "text_level": 1, + "bbox": [ + 511, + 263, + 777, + 279 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Effective robotic task execution requires goal comprehension, subgoal decomposition, and precise action sequencing. However, existing methods often lack structured reasoning mechanisms for long-horizon task decisions. To address this, we introduce a CoT reasoning dataset that enables deep reasoning before action generation, transforming concise instructions into structured long-horizon tasks for optimal execution. Our dataset follows a four-stage structure: (1) Goal Explanation; (2) Subgoal Decomposition; (3) Action Sequence Generation; (4) Action Sequence Modeling. By explicitly structuring these stages, our dataset mitigates fragmented planning and enhances decision-making. To improve model comprehension, each stage is marked with dedicated tags (e.g., , ..., ). Except for the final action sequence, intermediate responses are generated using GPT-4o, with task instructions rewritten for dataset expansion. Additionally, we incorporate 5,700 public complex surface images to leverage their similarity to sand terrain. The final dataset, SandThink-21k, consists of 21k CoT samples for model training.", + "bbox": [ + 511, + 287, + 906, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. The MindExplore System", + "text_level": 1, + "bbox": [ + 511, + 627, + 751, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Reasoning Layer", + "text_level": 1, + "bbox": [ + 511, + 656, + 679, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance reasoning abilities for long-horizon tasks, we employ a two-stage training strategy consisting of cold start followed by iterative direct preference optimization [7].", + "bbox": [ + 511, + 680, + 906, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1.1. Cold Start with Curriculum Learning", + "text_level": 1, + "bbox": [ + 511, + 742, + 823, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To promote structured reasoning in complex robotic environments, we employ SigLIP to match text instructions with collected images of Sandgo-21k in feature space, filtering out blurred image data caused by jitter, and finally cold-start training all linear layers of the Qwen2-VL-7B-Instruct model with LoRA. The training goal at this stage is to maximize the probability of generating an reasoning result given the image and instruction input. Let $X_{v}$ denotes the image input and $X_{q}$ denote the", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6841", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/97d586055f1c938e422aa255ad3073578292ff488ef0eb5f0256692580c9dd69.jpg", + "image_caption": [ + "Figure 3. MindExplore System Overview. MindExplore comprises a reasoning layer for task decomposition and an acting layer for execution. The Reasoning Layer, built on Qwen2-VL, applying Chain-of-Thought (CoT) reasoning to generate structured meta-action instructions. The acting Layer, a Mixture of Policy Experts (MoPE), includes two Multimodal Diffusion Policies (MMDPs): a base expert using RGB-Depth and LiDAR for movement, and an arm expert using RGB-Depth for manipulation. The Memory Mechanism ensures real-time feedback, enabling the reasoning layer to dynamically adjust the actions of the acting layer for optimal execution." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 385, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1e95ffaecbe3b6a19331cac635640c79f1571db1ea2022174303f0d3f8d1a2bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 88, + 897, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "instruction input. The objective formula is as follows:", + "bbox": [ + 89, + 438, + 447, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {c o l d - s t a r t}} = \\sum_ {i = 1} ^ {L} \\log P \\left(X _ {a, i} \\mid X _ {v}, X _ {q}, X _ {a, < i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 460, + 480, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $X_{r}$ denotes the reasoning output and $L$ is its length. To effectively guide training, we employ a curriculum learning approach [3], where data is arranged in order from easy to difficult. Initially, the model is trained on structured complex surface data, followed by real-world sand scenes. This progression ensures that the model first learns basic reasoning skills before transitioning to more complex tasks. By gradually increasing the task complexity, the model acquires structured reasoning capabilities and improves its ability to handle real-world uncertainty.", + "bbox": [ + 89, + 507, + 483, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.2. Enhancing Reasoning with RL", + "text_level": 1, + "bbox": [ + 89, + 664, + 349, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While cold start establishes structured reasoning, reinforcement learning is used to further refine the model's reasoning quality. We employ Iterative DPO to optimize the model's decisions in multimodal scenarios. Given an image $I$ and a task instruction $Q$ , the model generates an inference chain $R$ . DPO is based on the Bradley-Terry (BT) model:", + "bbox": [ + 89, + 680, + 483, + 758 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(R ^ {+} > R ^ {-} \\mid I, Q\\right) = \\frac {\\exp \\left(r _ {\\varphi} (I , Q , R ^ {+})\\right)}{\\exp \\left(r _ {\\varphi} (I , Q , R ^ {+})\\right) + \\exp \\left(r _ {\\varphi} (I , Q , R ^ {-})\\right)}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 765, + 482, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$r_{\\varphi}$ is a parameterized reward model, $R^{+}$ and $R^{-}$ denote preferred and non-preferred reasoning sequences.", + "bbox": [ + 89, + 797, + 482, + 827 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The policy model $\\pi_{\\theta}$ is trained to maximize the preferred alignment inference, while reference model $\\pi_{\\mathrm{ref}}$ is used to stabilize training process. The initial objective function is:", + "bbox": [ + 89, + 825, + 482, + 866 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\pi_ {\\theta}} \\mathbb {E} _ {(I, Q, R ^ {+}, R ^ {-})} [ r _ {\\varphi} (I, Q, R) ] - \\beta D _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} (R | I, Q) \\| \\pi_ {\\text {r e f}} (R | I, Q)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 871, + 482, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To stabilize training, we also incorporate a standard supervised fine-tuning loss, ensuring that the model retains the structured reasoning of supervised fine-tuning while benefiting from preference-based optimization. The loss function is given by the following formula:", + "bbox": [ + 511, + 438, + 906, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {\\mathrm {D P O}} (\\theta) = - \\mathbb {E} _ {(I, Q, R ^ {+}, R ^ {-})} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} (R ^ {+} | I , Q)}{\\pi_ {r e f} (R ^ {+} | I , Q)} \\right. \\right. \\\\ \\left. - \\beta \\log \\frac {\\pi_ {\\theta} \\left(R ^ {-} | I , Q\\right)}{\\pi_ {r e f} \\left(R ^ {-} | I , Q\\right)}\\right) + \\alpha \\frac {\\log \\pi_ {\\theta} \\left(R ^ {+} | I , Q\\right)}{| R ^ {+} |} ]. \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 545, + 903, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ denotes the trainable parameters of the model, $|R^{+}|$ denotes preferred reasoning length, the hyperparameter $\\alpha$ balances the two loss terms and $\\beta$ is a scaling factor.", + "bbox": [ + 511, + 627, + 905, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To build high-quality preference data, we refine the selection process in 3 rounds of training. In each iteration, we sample 8 inference outputs for each task and rank them using GPT-4o with preference scores ranging from 0 to 100. The selection follows a progressive rejection strategy:", + "bbox": [ + 511, + 683, + 905, + 758 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Iteration 1: The highest scoring response is selected as $R^{+}$ , while the lowest scoring response is taken as $R^{-}$ .", + "- Iteration 2: $R^{-}$ is selected as the highest ranked response below 60, increasing the task complexity.", + "- Iteration 3: $R^{-}$ is the lowest ranked response above 60, making model learn fine-grained reasoning distinctions." + ], + "bbox": [ + 511, + 768, + 903, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This approach enables models to excel in structured multimodal reasoning and long-horizon robotic planning.", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6842", + "bbox": [ + 482, + 943, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Acting Layer", + "text_level": 1, + "bbox": [ + 89, + 90, + 228, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Acting Layer generates a set of manipulation parameters and focuses on optimizing the joint probability distribution of all outputs. Therefore, generative diffusion models excelling in expressive capability and sampling quality [12] can be applied. Since the data dimensions of manipulation parameters is much smaller than that of image, the acting layer could get fast generation speed to provide real-time feedback to dynamic changes [18, 38, 39].", + "bbox": [ + 89, + 112, + 483, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Multimodal Diffusion Policy", + "text_level": 1, + "bbox": [ + 89, + 242, + 333, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Most VLAs use only single-modal sensor data, which struggles to provide precise manipulation for instruction following in complex environments. In challenging visual settings, RGB images often contain significant noise, such as the similarity in color between the cube in Figure 2 and the sandy terrain. On the other hand, models based solely on depth or point cloud data lack the ability to perceive scene details and cannot align with the semantic features of instructions. To address these, we propose a multimodal diffusion Policy model (MMDP) that jointly exploits information of multi-sensor data and textual instructions.", + "bbox": [ + 89, + 261, + 483, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ the robot manipulation visual encoder R3M, proposed in [25], to encode the RGB images. Instead of concatenating the depth and RGB images along the channel dimension, we transform them into point cloud data using a perspective projection model:", + "bbox": [ + 89, + 428, + 483, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX (u, v) = \\frac {\\left(u - c _ {x}\\right) \\cdot D (u , v)}{f _ {x}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 515, + 482, + 549 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY (u, v) = \\frac {(v - c _ {y}) \\cdot D (u , v)}{f _ {y}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 561, + 482, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ (u, v) = D (u, v), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 604, + 480, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $u$ and $v$ are the pixel coordinates in the depth image, $D(u,v)$ is the depth value of the pixel, $f_{x}$ and $f_{y}$ are the camera's focal lengths, $c_{x}$ and $c_{y}$ are the image's principal point coordinates, and $(X,Y,Z)$ represents the 3D coordinates of the point cloud data. We use the pyramid convolutional encoder from iDP3 [38] to encode point cloud data, while DistilBERT [27] is employed to encode textual instructions into feature vectors. Additionally, we propose a fusion adapters for both the sensor data and text, consisting of two linear layers, to map encoder outputs into an unified feature space.", + "bbox": [ + 89, + 628, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The manipulation parameter generation process of the MMDP is based on the conditional denoising diffusion generation of a lightweight 1D convolutional diffusion model. Specifically, we sample a random Gaussian noise $a^K$ and perform $k$ iterations of denoising using a conditional denoising network $\\epsilon_{\\theta}$ , conditioned on RGB data $r$ , depth information $d$ , LiDAR data $p$ , instruction data $i$ , and the robot's", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "state $q$ , to recover $a^K$ into the manipulation parameters $a^0$ :", + "bbox": [ + 511, + 90, + 901, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\na ^ {k - 1} = \\alpha_ {k} \\left(a ^ {k} - \\gamma_ {k} \\epsilon_ {\\theta} \\left(a ^ {k}, k, r, d, p, i, q\\right)\\right) + \\sigma_ {k} \\mathcal {N} (0, I), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 117, + 903, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{N}(0,I)$ is Gaussian noise, $\\alpha_{k},\\gamma_{k}$ , and $\\sigma_{k}$ are functions of $k$ and depend on the noise scheduler [12]. During training, we iteratively add random noise $\\epsilon^k$ for $k$ steps to the manipulation parameters, resulting in the noisy action $\\bar{a}_k a^0 +\\bar{\\beta}_k\\epsilon^k$ , where $\\bar{a}_{k}$ and $\\bar{\\beta}_k$ are the noise schedule parameters that performs one step noise adding. The denoising network $\\epsilon_{\\theta}$ is then used to predict the random noise $\\epsilon^k$ , and the training loss function is as follows:", + "bbox": [ + 511, + 148, + 906, + 270 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\operatorname {M S E} \\left(\\epsilon^ {k}, \\epsilon_ {\\theta} \\left(\\bar {\\alpha_ {k}} a ^ {0} + \\bar {\\beta_ {k}} \\epsilon^ {k}, k, r, d, p, i, q\\right)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 281, + 903, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2. Mixture of Policy Experts", + "text_level": 1, + "bbox": [ + 511, + 311, + 740, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The acting layer needs to provide precise feedback to dynamic environmental changes with relatively low model complexity, we argue that a unified manipulation model and unfiltered sensor inputs may interfere with the acting layer's accurate manipulation of specific meta-actions. Therefore, we propose a Mixture of Policy Experts Model that decouples manipulation of the robot's base and arm. The base expert is primarily responsible for executing meta-actions related to movement, outputting only the linear and angular velocities of base. Multimodal inputs are filtered to include RGB images from the base camera, base point cloud data with depth values less than $1.5\\mathrm{m}$ , and LiDAR point clouds within a horizontal radius of $1.2\\mathrm{m}$ . The arm expert performs meta-actions for grasping and placing, outputting only the pose and velocity parameters for arm manipulation. Its sensor inputs are RGB images from both the base and arm cameras, base point cloud data with depth values less than $0.8\\mathrm{m}$ and arm point clouds with depth values less than $0.5\\mathrm{m}$ .", + "bbox": [ + 511, + 330, + 906, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Memory Mechanism", + "text_level": 1, + "bbox": [ + 511, + 613, + 709, + 630 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af49e470c804a8b47dc5c1ea6b13bbb4a8b48026d6dd3ed7dee99414b28295d8.jpg", + "image_caption": [ + "Figure 4. The details of Memory Mechanism" + ], + "image_footnote": [], + "bbox": [ + 516, + 641, + 908, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance the coordination between the Reasoning Layer and the Acting Layer and achieve global control over both software and hardware, we design a memory mechanism with real-time feedback and gated routing. As shown in Figure 4, upon receiving a human instruction, the memory mechanism first processes multi-view RGB images and task instructions, which are fed into reasoning layer to generate a sequence of meta-action instructions. The gated routing network then sequentially assigns each meta-action to the corresponding VLA action expert, while real-time", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6843", + "bbox": [ + 482, + 943, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/925fcd9044d8c83ae195edf68b289b8b54c21fe3ba67bd52175fc00c8598af2b.jpg", + "table_caption": [ + "Table 1. The success rates of different VLA models across 24 tasks in four long-horizon tasks." + ], + "table_footnote": [], + "table_body": "
Coarse-Grained Long-Horizon TasksIntensive Instruction-Following Tasks
MethodTask1Task2Task3Task4Task5Task6Avg. (%)Task7Task8Task9Task10Task11Task12Avg. (%)
OpenVLA0/100/103/100/102/101/1010.00%1/103/100/100/101/102/1011.67
RDT5/103/100/103/106/105/1036.67%5/103/102/100/105/105/1033.33
MindExplore10/108/108/107/1010/1010/1088.3310/109/1010/108/109/1010/1093.33
Cross-Expert TasksZero-Shot Generalization Tasks
MethodTask13Task14Task15Task16Task17Task18Avg. (%)Task19Task20Task21Task22Task23Task24Avg. (%)
OpenVLA0/100/101/100/100/100/101.67%0/100/100/102/100/100/103.33
RDT4/103/105/101/102/106/1035.00%2/101/103/100/100/103/1015.00
MindExplore9/1010/1010/109/109/109/1093.3310/108/108/109/107/1010/1086.67
", + "bbox": [ + 94, + 107, + 906, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "multimodal sensor data is provided to the corresponding acting layer's action expert to generate control parameters, which are subsequently used to execute actions via the hardware. The gated routing network can be formatted as:", + "bbox": [ + 89, + 284, + 482, + 347 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {A E}, w _ {B E} = \\operatorname {s o f t m a x} (\\operatorname {F F N} (\\operatorname {D i s t i l l B e r t} (m i))), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 354, + 482, + 371 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM o P E (m i) = \\left\\{ \\begin{array}{l l} f _ {A E} (m i, I _ {A E}), & \\text {i f} w _ {A E} > w _ {B E} \\\\ f _ {B E} (m i, I _ {B E}), & \\text {i f} w _ {A E} \\leq w _ {B E} \\end{array} , \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 380, + 480, + 433 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $mi$ represents a single meta-action instruction, while $I_{AE}$ and $I_{BE}$ denote the multimodal sensor inputs required by the arm expert and base expert, respectively. During this process, the memory mechanism stores the initial multiview RGB images received by reasoning layer and continuously feeds the current state of the robot back to Acting Layer while it generates control parameters, enabling real-time adjustments. Additionally, the memory mechanism supplies reasoning layer with both current and historical RGB sensor images, equipping reasoning layer with multi-temporal environmental perception to determine the completion status of each meta-action. If a meta-action is successfully executed, the memory mechanism forwards the next action to acting layer. If an error occurs during execution, it triggers reasoning layer to replan the entire task, generating a new sequence of meta actions, thereby improving execution stability and robustness.", + "bbox": [ + 89, + 434, + 483, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 702, + 223, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 727, + 307, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train the Reasoning Layer with SandThink-21k and the acting layer using individual meta-actions from the SandGo-1k, with each meta-action representing an episode. Movement and crossing obstacle are used to train the base expert, while grasping, placing, and writing are used to train the arm expert. During training, we randomly apply GPT40 for instruction augmentation to simulate task instructions with varying levels of granularity. The SAM model [15] is used for semantic segmentation of RGB images, replacing the black background with different colors for data aug", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mentation. We evaluate the MindExplore system directly in the same sandpit scenario, using success rates as the metric, which is the ratio of successful trials to the total number of experiments. The tests are conducted with varying scene configurations and ensures that all comparison models are tested under the same conditions.", + "bbox": [ + 511, + 284, + 906, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. System Performance", + "text_level": 1, + "bbox": [ + 511, + 386, + 709, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2.1. Performance on Real World Tasks", + "text_level": 1, + "bbox": [ + 511, + 409, + 797, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate 24 real-world robotic tasks categorized into four types, each containing six distinct tasks. All these evaluated tasks can be divided into four categories according to their ability to complete long-horizon tasks, and we conducted 615 trials on real robots, ensuring its accuracy on meta-action and long-horizon tasks.", + "bbox": [ + 511, + 429, + 906, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Coarse-Grained Long-Horizon Tasks. These tasks involve high-level, ambiguous instructions (e.g., \"grab the branch\"), requiring the system to first interpret and structure precise task sequences before execution. The robot must break down vague commands into actionable steps, Such as object grasping after obstacle avoidance (Task 1, 2), going to position writing (Task 3, 4), and navigating challenging terrain to complete sampling (Task 5, 6).", + "bbox": [ + 511, + 520, + 905, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intensive Instruction Following Tasks. These tasks consist of sequential, interdependent sub-tasks, testing the system's compositional reasoning and execution fluency. The robot must understand and follow multi-step instructions with precision, such as continuous pick-and-place operations are performed in a dynamically changing environment (Task 7-9), and sampling is performed after multiple obstacle avoidance (Task 10-12).", + "bbox": [ + 511, + 642, + 905, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cross-Expert Tasks. These tasks demand seamless coordination between base movement and arm manipulation, requiring the integration of multiple operational skills (e.g., moving, turning, grasping, and placing) through cross-modal reasoning. The robot must reason about motion and manipulation tasks in combination, such as grasping a specified object to place in a container (Task 13-15) and writing at a specified location (Task 16-18).", + "bbox": [ + 511, + 763, + 905, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-Shot Generalization Tasks. These tasks evaluate", + "bbox": [ + 531, + 885, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6844", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b05042eba977c5b487cab9531da15317679cbe14b624aa229f7ae4d66372d95e.jpg", + "image_caption": [ + "Figure 5. During long-horizon tasks, the reasoning layer decomposes the instruction into a structured sequence of executable steps. The acting layer then orchestrates the hierarchical execution by leveraging the Base Expert and Arm Expert." + ], + "image_footnote": [], + "bbox": [ + 93, + 109, + 903, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4daa2c8da557e7fe45e9e4d9ef6500f01349722cdf001fccd28496ff6f4fbdfa.jpg", + "table_caption": [ + "Table 2. Comparison results with the most powerful closed-source and open-source MLLMs on three tasks." + ], + "table_footnote": [], + "table_body": "
ModelCoarse-Grained \nLong-Horizon TasksIntensive Instruction \nFollowing TasksCross-Expert \nMulti-TasksThe Zero-Shot \nGeneralization Tasks
ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)
GPT-4o-latest33.0247.0037.5450.0041.5961.0029.1721.00
Gemini-2.0-Flash29.5243.0033.6949.0043.5355.0024.3618.00
InternVL2.5-8B14.7133.0017.4540.0019.3743.0019.8815.00
Qwen2.5-VL-7B17.3235.0018.9344.0025.3249.0023.7319.00
DeepSeek-VL211.8427.0014.6140.0019.4943.007.185.00
Reasoning Layer69.3587.0070.2381.0075.3589.0063.1773.00
", + "bbox": [ + 94, + 306, + 906, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the ability of system to adapt previously unseen environmental terrain (Tasks 19 to 21) and object types (Tasks 22 to 24). The robot must generalize over different objects (branches, bottles, and stones), recognize and manipulate objects despite variations in texture, shape, and weight, and adjust to new ground conditions (rugged, steep slopes).", + "bbox": [ + 89, + 429, + 483, + 521 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5e88a75b01d38ed03fe82cfa621096143a2a7f9c850ca06d9fc94fe951245272.jpg", + "table_caption": [ + "Table 3. Success Rates of the Acting Layer across 5 meta-actions: Moving to a Target Location (MTL), Crossing Obstacles (CO), Grasping a Specified Object (GSO), Placing the Object into a Designated Container (POC), Writing (Wr). Avg. - Average Success Rate, BE - Base Expert, AE - Arm Expert." + ], + "table_footnote": [], + "table_body": "
MethodMTLCOGSOPOCWrAvg. (%)
ACT13/1003/6537/10055/6019/5036.86
iDP347/10027/6549/10043/6021/5050.24
BE85/10058/65---87.12
AE--79/10060/6048/5091.67
", + "bbox": [ + 91, + 608, + 480, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We select the OpenVLA [14] and the diffusion model RDT [18] as comparison methods. As shown in Table 1, in instruction-following, MindExplore's performance improves as instruction granularity increases, whereas OpenVLA and RDT perform better with coarse-grained instructions than with intensive instructions. We attribute this to MindExplore's strong reasoning capability, where more informative instructions guide the model to execute more precise actions, whereas OpenVLA and RDT tend to favor simple instruction descriptions and lack deep thoughts for long-horizon tasks. In multi-skill coordination, MindExplore outperforms single-expert models with a higher success rate due to the MoPE. In terms of generalization, compared to end-to-end MLLM and VLAs without world knowledge,", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MindExplore demonstrates strong generalization capabilities in unseen scenarios. Notably, the average sequence length for the cruising and obstacle avoidance task is 3.12, for the sampling task is 4.3, and for the writing task is 2. We find that the success rates of OpenVLA and RDT decrease progressively with the increasing meta-action length across three tasks. In Figure 5, MindExplore maintains a consistent success rates across sequences of varying lengths, benefiting from its superior task planning, decomposition capabilities, and the memory mechanism.", + "bbox": [ + 511, + 429, + 906, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.2. Performance of Different Components", + "text_level": 1, + "bbox": [ + 511, + 590, + 826, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also assess the performance of the two key components, reasoning and acting layer. For reasoning layer, we collect 480 RGB-instruction pairs across four test tasks as the evaluation dataset, and select the most powerful closed-source and open-source MLLMs for comparative experiments, including GPT-4o-latest, Gemini-2.0-Flash, InternVL2.5-8B [8], Qwen2.5-VL-7B [2], and DeepSeek-VL2 [34]. We use prompt to force all models to output a reasoning process similar to reasoning layer, and only focus on the output content of action sequence modeling, using ROUGE_L and meta-action planning accuracy to evaluate model performance. The experimental results are shown in Table 2. Reasoning layer performs best on all indicators and tasks.", + "bbox": [ + 511, + 612, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For acting layer, we compare it with two typical manipulation models, the RGB-based ACT [11] and the point-cloud-based iDP3 [38]. Since both methods are not instruction-following models, a separate manipulation model must be trained for each meta-action. In contrast, the acting layer only requires training two expert models to", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "header", + "text": "Instructions Follow Long-horizon Tasks", + "bbox": [ + 349, + 88, + 611, + 101 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6845", + "bbox": [ + 482, + 943, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/88b990b40b46678bc1ad64a2ab809fd59a03a4ee2ffc8a04203f54efbd6dec50.jpg", + "table_caption": [ + "Table 4. Ablation results with training strategy on three tasks. “+” stands for accumulation." + ], + "table_footnote": [], + "table_body": "
ModelCoarse-grained \nLong-horizon tasksIntensive instruction \nfollowing tasksCross-Expert \nmulti-tasksThe zero-shot \nGeneralization task
ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)
Qwen2-VL-7B13.5429.0015.2235.0023.1737.007.3119.00
+ Cold Start47.3562.0051.4268.0064.5775.0033.1643.00
+ iteration 154.1869.0054.5470.0067.1579.0035.3455.00
+ iteration 261.5875.0063.2979.0070.2381.0046.5764.00
+ iteration 369.3587.0070.2381.0075.3589.0063.1773.00
", + "bbox": [ + 96, + 107, + 906, + 212 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1e0d761a2d9f04a0f8ce97854d16fda22aa82b13f4e7236a4d51b7be4e7fcbec.jpg", + "table_caption": [ + "Table 5. Ablation study of the MindExplore system on the cruising and obstacle avoidance task." + ], + "table_footnote": [], + "table_body": "
MethodTask1Task2Task7Task8Task13Task14Task19Task20Avg. (%)
w/o Reasoning & Memory5/105/108/106/104/106/109/105/1060.00
w/o Memory8/107/1010/107/105/108/109/107/1076.25
MindExplore10/108/1010/109/109/1010/1010/108/1092.50
", + "bbox": [ + 96, + 239, + 906, + 308 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/38d9db2d213bd357744f36e9a8d6df1a851f2b48f97dbd405250149830c595b3.jpg", + "table_caption": [ + "Table 6. Ablation Study on Input Modalities for Meta-Actions. \"PC\" - Point Cloud, \"2D\" - 2D Images." + ], + "table_footnote": [], + "table_body": "
MethodInput ModalityDepth TypeMTLGSO
AERGB--23/50
AEDepth+RGBPC-40/50
BERGB-5/50-
BELiDAR+DepthPC37/50-
BELiDAR+Depth+RGB2D32/50-
BELiDAR+Depth+RGBPC42/50-
", + "bbox": [ + 96, + 357, + 485, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "control different hardware structures, which is parameter efficient. As shown in Table 3, the ACT performs worst on actions requiring long-distance movement, such as MTL and CO, indicating that point cloud with spatial information is crucial for movement manipulation. By combining spatial information with color and texture information from RGB images, the BE and AE achieve better alignment between object semantics and text instructions in real-world scenes, achieving the highest success rates across all meta-actions.", + "bbox": [ + 89, + 468, + 482, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 619, + 243, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation results of Reasoning Layer. Compared to DeepSeek-VL2 and InternVL2.5, although they are on par with Qwen2-VL in some benchmark indicators, they lack scalability for real-world tasks in embodied systems, such as spatial calibration, semantic perception, and fine-grained object recognition in noisy backgrounds (e.g., inability to distinguish between sand and yellow branches). Therefore, we chose Qwen2-VL as the base model of the reasoning layer. To prove the effectiveness of the Reasoning Layer training method, we performed detailed ablation experiments on three tasks, which involved adding our training methods one by one to the basic model. As shown in Table 4, cold start enables the model to obtain structured reasoning capabilities. The three rounds of iterative DPO training further improve the model's reasoning quality and achieve self-improvement by continuously building high-quality preference datasets.", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation of Acting Layer. We explore the impact of different input modalities on acting layer. As shown in Table 6, the performance of the AE using pure vision declined due to the lack of spatial information, and the RGB-based BE achieves just a $10\\%$ success rate in MTL. Similarly, the point-cloud-based BE also shows lower success rates, as point cloud data failed to align effectively with the semantic content of instructions. Additionally, we test concatenating Depth images with RGB images along the channel dimension. The results show that point cloud data in acting layer better leverages the spatial information from Depth images.", + "bbox": [ + 511, + 320, + 906, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation of MindExplore System. We conduct an ablation study on the core components of the MindExplore system. In Table 5, in the reasoning layer and acting layer system without the memory mechanism, Reasoning layer performs a single task decomposition and sequentially sends meta-action instructions to acting layer. However, without timely state feedback and error correction from acting layer, the system is highly susceptible to error accumulation, resulting in a decrease in success rate. Further, when the Reasoning layer component is removed and only the acting layer is used, the success rate on the cruising and obstacle avoidance task significantly drops due to the lack of deep reasoning and acting of the task instructions. Nonetheless, it still outperforms other VLAs, demonstrating that the MoPE structure can ensure accurate completion of base tasks.", + "bbox": [ + 511, + 487, + 908, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 724, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose MindExplore, a hierarchical embodied system designed for long-horizon tasks, including Reasoning, Acting, and Memory Mechanism. By aligning planning with action, MindExplore achieves strong generalization in dynamic environments. Besides, we introduce SandGo-1k and SandThink-21k, the first expert-level multimodal embodied and CoT datasets, bridging the data gap in challenging terrains. Real-world evaluations demonstrate MindExplore achieves better performance, ensuring high reliability in complex environments.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6846", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China (Grant 62322117, 62371365, and U24B20136), the Fundamental Research Funds for the Central Universities, and the Innovation Fund of Xidian University (Grant No. YJSJ25007).", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 207, + 187, + 222 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Muneeb Ahmed, Brejesh Lall, Rajesh Kumar, and Arzad A Kherani. Towards estimation of human intent in assistive robotic teleoperation using kinaesthetic and visual feedback. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1928-1934, 2023. 2", + "[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7", + "[3] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning (ICML), pages 41-48, 2009. 4", + "[4] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. $\\pi 0$ : A vision-language-action flow model for general robot control, 2024. arXiv preprint arXiv:2410.24164. 2", + "[5] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2", + "[6] Shizhe Chen, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Sugar: Pre-training 3d visual representations for robotics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18049-18060, 2024. 2", + "[7] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 3", + "[8] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 7", + "[9] David B D'Ambrosio, Saminda Abeyruwan, Laura Graesser, Atil Iscen, Henri Ben Amor, Alex Bewley, Barney J Reed, Krista Reymann, Leila Takayama, Yuval Tassa, et al. Achieving human level competitive robot table tennis. arXiv preprint arXiv:2408.03906, 2024. 2", + "[10] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, et al. Robothor: An open simulation-to-real embodied ai platform. In Proceedings of the IEEE/CVF Conference on Computer" + ], + "bbox": [ + 93, + 232, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition (CVPR), pages 3164-3174, 2020. 2", + "[11] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024. 7", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33:6840-6851, 2020. 5", + "[13] Stephen James, Kentaro Wada, Tristan Laidlow, and Andrew J Davison. Coarse-to-fine q-attention: Efficient learning for visual robotic manipulation via discretisation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13739–13748, 2022. 2", + "[14] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2, 7", + "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 6", + "[16] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Manipllm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18061-18070, 2024. 2", + "[17] Haitao Lin, Yanwei Fu, and Xiangyang Xue. Pourit!: Weakly-supervised liquid perception from a single image for visual closed-loop robotic pouring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 241-251, 2023. 2", + "[18] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 5, 7", + "[19] Yuecheng Liu, Dafeng Chi, Shiguang Wu, Zhanguang Zhang, Yaochen Hu, Lingfeng Zhang, Yingxue Zhang, Shuang Wu, Tongtong Cao, Guowei Huang, et al. Spatialcot: Advancing spatial reasoning through coordinate alignment and chain-of-thought for embodied task planning. arXiv preprint arXiv:2501.10074, 2025. 2", + "[20] Jingpei Lu, Florian Richter, and Michael C Yip. Markerless camera-to-robot pose estimation via self-supervised sim-to-real transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21296-21306, 2023. 2", + "[21] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems (NeurIPS), 35: 2507-2521, 2022. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6847", + "bbox": [ + 482, + 943, + 514, + 954 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 2", + "[23] Xiao Ma, Sumit Patidar, Iain Haughton, and Stephen James. Hierarchical diffusion policy for kinematics-aware multitask robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18081-18090, 2024. 2", + "[24] Weixin Mao, Weiheng Zhong, Zhou Jiang, Dong Fang, Zhongyue Zhang, Zihan Lan, Fan Jia, Tiancai Wang, Haoqiang Fan, and Osamu Yoshie. Robomatrix: A skill-centric hierarchical framework for scalable robot task planning and execution in open-world. arXiv preprint arXiv:2412.00171, 2024. 2", + "[25] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 5", + "[26] Max Polzin, Qinghua Guan, and Josie Hughes. Robotic locomotion through active and passive morphological adaptation in extreme outdoor environments. Science Robotics, 10(99): eadp6419, 2025. 3", + "[27] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.5", + "[28] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2998-3009, 2023. 2", + "[29] Chan Hee Song, Valts Blukis, Jonathan Tremblay, Stephen Tyree, Yu Su, and Stan Birchfield. Robospatial: Teaching spatial understanding to 2d and 3d vision-language models for robotics. arXiv preprint arXiv:2411.16537, 2024. 2", + "[30] Enrica Tricomi, Francesco Missiroli, Michele Xiloyannis, Nicola Lotti, Xiaohui Zhang, Marios Stefanakis, Maximilian Theisen, Jürgen Bauer, Clemens Becker, and Lorenzo Massia. Soft robotic shorts improve outdoor walking efficiency in older adults. Nature Machine Intelligence, 6(10):1145-1155, 2024. 3", + "[31] Kasun Weerakoon, Adarsh Jagan Sathyamoorthy, Mohamed Elnoor, and Dinesh Manocha. Vapor: Legged robot navigation in unstructured outdoor environments using offline reinforcement learning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 10344–10350. IEEE, 2024. 3", + "[32] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems (NeurIPS), 35:24824-24837, 2022. 2", + "[33] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Peng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. 2", + "[34] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 7", + "[35] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4737-4746, 2023. 2", + "[36] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567, 2024. 2", + "[37] Michal Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 2", + "[38] Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xialin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with improved 3d diffusion policies. arXiv preprint arXiv:2410.10803, 2024. 5, 7", + "[39] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. arXiv preprint arXiv:2403.03954, 2024. 5", + "[40] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 2", + "[41] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2", + "[42] Da Zhao, Haobo Luo, Yuxiao Tu, Chongxi Meng, and Tin Lun Lam. Snail-inspired robotic swarms: a hybrid connector drives collective adaptation in unstructured outdoor environments. Nature Communications, 15(1):3647, 2024. 3", + "[43] Allan Zhou, Moo Jin Kim, Lirui Wang, Pete Florence, and Chelsea Finn. Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17907-17917, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 849 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6848", + "bbox": [ + 482, + 944, + 514, + 954 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/d546fd0e-5080-4957-803b-5e64e0ab9d88_model.json b/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/d546fd0e-5080-4957-803b-5e64e0ab9d88_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b4c71be00118f06cd6275038d634037a16ba0f1c --- /dev/null +++ b/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/d546fd0e-5080-4957-803b-5e64e0ab9d88_model.json @@ -0,0 +1,2409 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.212, + 0.131, + 0.787, + 0.177 + ], + "angle": 0, + "content": "Towards Long-Horizon Vision-Language-Action System: Reasoning, Acting and Memory" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.203, + 0.903, + 0.277 + ], + "angle": 0, + "content": "Daixun Li\\(^{1,\\ast}\\) Yusi Zhang\\(^{1,\\ast}\\) Mingxiang Cao\\(^{1,\\ast}\\) Donglai Liu\\(^{1,\\ast}\\) Weiying Xie\\(^{1,\\dagger}\\) Tianlin Hui\\(^{1}\\) Lunkai Lin\\(^{2}\\) Zhiqiang Xie\\(^{2}\\) Yunsong Li\\(^{1}\\) Xidian University AgileX Robotics https://xdu-imagelab.github.io/MindExplore/" + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.293, + 0.404, + 0.308 + ], + "angle": 0, + "content": "MindExplore System" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.31, + 0.36, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.313, + 0.506, + 0.322 + ], + "angle": 0, + "content": "Reasoning Layer Validation" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.323, + 0.546, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.294, + 0.756, + 0.308 + ], + "angle": 0, + "content": "Dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.31, + 0.696, + 0.322 + ], + "angle": 0, + "content": "SandGo-1k" + }, + { + "type": "text", + "bbox": [ + 0.569, + 0.328, + 0.758, + 0.341 + ], + "angle": 0, + "content": "Multi-Modal: RGB Depth LiDAR" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.345, + 0.763, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.785, + 0.309, + 0.882, + 0.322 + ], + "angle": 0, + "content": "SandThink-21k" + }, + { + "type": "text", + "bbox": [ + 0.785, + 0.329, + 0.884, + 0.342 + ], + "angle": 0, + "content": "Chain of Thought" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.345, + 0.875, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.795, + 0.402, + 0.872, + 0.409 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.794, + 0.41, + 0.872, + 0.416 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.812, + 0.417, + 0.858, + 0.423 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.812, + 0.424, + 0.858, + 0.429 + ], + "angle": 0, + "content": "" + }, + { + "type": "text", + "bbox": [ + 0.799, + 0.431, + 0.868, + 0.438 + ], + "angle": 0, + "content": "..." + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.451, + 0.234, + 0.464 + ], + "angle": 0, + "content": "Zero-Shot Tasks" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.467, + 0.251, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.565, + 0.264, + 0.578 + ], + "angle": 0, + "content": "Perceive seen & unseen object" + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.451, + 0.428, + 0.465 + ], + "angle": 0, + "content": "Long-Horizon Tasks" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.467, + 0.432, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.565, + 0.448, + 0.578 + ], + "angle": 0, + "content": "Complete multiple meta-actions" + }, + { + "type": "image_caption", + "bbox": [ + 0.484, + 0.451, + 0.614, + 0.465 + ], + "angle": 0, + "content": "Cross-Expert Tasks" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.467, + 0.647, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.565, + 0.643, + 0.578 + ], + "angle": 0, + "content": "Adapt to the dynamic sand scene" + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.451, + 0.872, + 0.465 + ], + "angle": 0, + "content": "Instruction Following Tasks" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.467, + 0.895, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.565, + 0.901, + 0.578 + ], + "angle": 0, + "content": "Complete long-horizon tasks as instructed" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.584, + 0.907, + 0.627 + ], + "angle": 0, + "content": "Figure 1. Overview of MindExplore System: A hierarchical embodied intelligence system including reasoning, acting, and memory, offering state-of-the-art generalizability in highly dynamic scenes. Unlike conventional planar embodied intelligence models that consolidate all capabilities within a single layer, MindExplore introduces a novel expert-level hierarchical embodied system architecture." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.636, + 0.327, + 0.651 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.667, + 0.486, + 0.879 + ], + "angle": 0, + "content": "Vision-Language-Action (VLA) is crucial for autonomous decision-making in embodied systems. While current methods have advanced single-skill abilities, their short-horizon capability limits applicability in real-world scenarios. To address this challenge, we innovatively propose MindExplore, a general hierarchical VLA system with cross-skill for long-horizon tasks in highly dynamic sand. The key insight is to iteratively align the knowledge domain of task planning and action execution. Thus, this task-oriented action enables outstanding generalization across a wide range of real-world scenarios. In the reasoning layer, task-specific chains of thought (CoT) are designed for planning long-horizon task sequences and providing meta-action signals. In the acting layer, a simple but powerful Mixture of Pol" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.909, + 0.849 + ], + "angle": 0, + "content": "icy Experts strategy is built inspired by signals and multimodal inputs for adaptively selecting skill experts and generating closed-loop action sequences. Also, it integrates a lightweight Multimodal Diffusion Policy (MMDP) to enhance spatial perception by fusing multi-visual modality features. Besides, the pioneering memory mechanism establishes feedback between the reasoning and acting layers, facilitating adaptive execution of long-horizon tasks and real-time replanning. Notably, we create SandGo-1k and SandThink-21k, the first expert-level multimodal embodied dataset and CoT dataset tailored for sandy environments. At a high execution frequency of 30 FPS, MindExplore is \\(3.01 \\times\\) more successful than existing methods in unstructured and dynamic environments." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.23, + 0.899 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.233, + 0.888, + 0.38, + 0.9 + ], + "angle": 0, + "content": "† Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6839" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.312 + ], + "angle": 0, + "content": "Current Visual-Language-Action (VLA) data and methods mainly focus on short sequence tasks of single ability [6, 20], that is, single-skill sequences limited to either arm operations or base movements, making them effective for simple tasks requiring only \"muscle memory\". [13, 20, 24]. However, these approaches fail to generalize for long-horizon tasks in real-world scenarios, particularly in complex terrain action [14, 33]. This requires necessitates the agent to handle long-horizon and multi-skill instructions in highly dynamic environments [37]. Thus, successful execution in such settings needs deep consideration—Reasoning, precise control—Acting, and adaptive feedback—Memory [5]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.483, + 0.538 + ], + "angle": 0, + "content": "To achieve this, three key challenges must be addressed: (1) Task-oriented Chain of Thought (CoT) reasoning requires the hierarchical decomposition of coarse-grained human instructions into executable sequences of metaactions, ensuring long-horizon alignment in dynamic environments [10, 20, 23]; (2) High-Precision Multimodal Control through a lightweight spatial perception framework fuses RGB, depth, and LiDAR inputs for robust motion calibration and manipulation accuracy [17, 35]; (3) Adaptive State Tracking that dynamically memorizes system states and task progress, enabling real-time closed-loop feedback between planning and acting to mitigate error propagation from perceptual drift or control misalignment [1, 43]. Consequently, this remains an open and worthwhile challenge so far." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose MindExplore for long-horizon tasks in complex and dynamic environments with a novel expert-level hierarchical embodied system architecture (see Figure 1). Our MindExplore is inspired by the discovery that continuous iterative alignment of the knowledge domain between task planning and action execution enables strong generalization across diverse real-world scenarios. On one hand we pursue task-oriented CoT in the Reasoning layer for effectively capturing meta-action sequences, while on the other hand we design a straightforward Mixture of Policy Experts (MoPE) strategy in the Acting Layer for flexibly generating closed-loop action sequences. Moreover, Memory Mechanism is designed to establish feedback between the reasoning and acting layers. Thus, the proposed MindExplore formulates long-horizon task execution into trainable skill expert policies, providing innovative guidance for hierarchical embodied intelligence systems without failures in highly dynamic environments. We also introduce a lightweight Multimodal Diffusion Policy (MMDP) that is dedicated to integrating RGB, depth and LiDAR data into a feature embedding space, thus enhancing spatial perception and motion control capabilities. To accurately evaluate complex tasks, we construct SandGo-1k and SandThink-21k as the first expert-level multimodal" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "embodied dataset and CoT dataset, respectively. We also carry out experiments to show that, the system maintains high accuracy and efficient execution even in highly dynamic environments such as mobile and unstructured sand, which validates our architecture. Our key contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.186, + 0.905, + 0.261 + ], + "angle": 0, + "content": "- We propose MindExplore, a novel expert-level hierarchical embodied system to adapt long-horizon tasks in unstructured and dynamic environments. To our best knowledge, MindExplore is the first work to formulate Reasoning, Acting and Memory into a unified architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.262, + 0.905, + 0.351 + ], + "angle": 0, + "content": "- We create SandGo-1k, the first embodied intelligence dataset for mobile operations in complex environments. It includes text-vision (RGB, depth, and LiDAR) data covering complex movement episodes. Building on this, we further construct a multimodal CoT dataset for highly dynamic terrains, SandThink-21k." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.905, + 0.428 + ], + "angle": 0, + "content": "- We empirically evaluate the accuracy on real-world experiments across diverse complex terrains with 24 distinct tasks, such as steep slopes, pits, and dunes. These experiments highlight MindExplore exceptional performance for long-horizon tasks in real-world scenarios." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.186, + 0.905, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.454, + 0.655, + 0.469 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.743, + 0.499 + ], + "angle": 0, + "content": "2.1. Visual Chain-of-Thought" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.66 + ], + "angle": 0, + "content": "CoT reasoning [32] has become fundamental in MLLMs [40, 41], with applications in science [21], mathematics [22], and advanced question answering [36]. Recently, visual CoT has been applied to robotics [19, 29, 37] to tackle task-specific challenges. However, existing methods focus on single-instruction, single-step tasks, failing in complex scenarios that demand long-horizon task decomposition. We argue that orchestrating concise instructions into structured long-horizon tasks is a critical step toward autonomous robot decision-making." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.69, + 0.698 + ], + "angle": 0, + "content": "2.2. Embodied System" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.905, + 0.799 + ], + "angle": 0, + "content": "The long-horizon tasks in open-world environments has been a major challenge in robotics research [9]. Numerous studies leverage MLLMs with extensive world knowledge to decompose instructions into subtasks, which are executed by lightweight VLA models [4, 16, 24, 28]. The key challenges can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.802, + 0.905, + 0.833 + ], + "angle": 0, + "content": "Challenge 1: How can instructions be expanded into long-horizon action sequences and executed flawlessly?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.836, + 0.905, + 0.868 + ], + "angle": 0, + "content": "Challenge 2: How to effectively accomplish movement and manipulation in complex and dynamic terrain?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Challenge 3: How to overcome the data missing caused by the difficulty of simulation of highly dynamic scenes?" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.955 + ], + "angle": 0, + "content": "6840" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.282, + 0.108 + ], + "angle": 0, + "content": "3. The System Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.485, + 0.193 + ], + "angle": 0, + "content": "To address Challenge 3, we collect a multi-task embodied operation dataset SandGo-1k, using a tracked mobile robot with a single-arm manipulator (as shown in Figure 2) in a real sand. Based on SandGo-1k, we generate a reasoning-acting embodied instruction dataset SandThink-21k." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.204, + 0.367, + 0.22 + ], + "angle": 0, + "content": "3.1. Embodied Dataset: SandGo-1k" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.238, + 0.482, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.44, + 0.483, + 0.482 + ], + "angle": 0, + "content": "Figure 2. Left: AgileX Robotic arm degrees of freedom and environmental layout (rugged terrain, obstacles, pits, dunes). Middle: Base movement states. Right: Collected data content." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.492, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Environmental Layout. In real-world tasks [26, 30, 31, 42], uneven sandy terrains dominate outdoor applications, making them essential for studying mobile manipulation systems. The high dynamic fluidity of sand demands adaptive adjustments, while its visual noise complicates perception and calibration. Enhancing generalization in unstructured environments remains a key challenge, and our experiments show that a system robust in sandy conditions can more easily generalize to simpler terrains like factory floors, asphalt roads, or regular soil. Therefore, to investigate the capabilities of mobile manipulation systems in highly dynamic and complex environments, we select sand as the primary experimental setting, including potholes, steep slopes, mounds, and rough roads, and construct a \\(6 \\times 8\\) meters real-world sandbox for data collection and evaluation, as depicted in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Task Setup and Collection Process. We design three long-horizon meta-tasks: (1) Cruising and Obstacle Avoidance, training the model for environmental perception and motion planning in complex terrains; (2) Sampling, enabling the model to perform instruction-following, target localization, and coordinated base-arm movement; (3) Writing, enhancing the model's fine manipulation capabilities. Based on these, we can decompose and reassemble them into five meta-actions: (1) Moving to a Target Location, (2) Crossing Obstacles, (3) Grasping a Specified Object, (4) Placing the Object into a designated Container, and (5)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "Writing. During data collection, task instructions are manually written and decomposed into sequences of meta-actions with individual instructions. Each meta-action is saved as an episode, serving as the smallest unit of data collection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.155, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Data Content. The collected data includes: (1) Sensor information, consisting of RGB-Depth images from the two cameras and LiDAR point clouds; (2) Robot state and manipulation commands, including the arm's pose and velocity, as well as the base's angular and linear velocities; (3) Text instructions for long-horizon tasks and meta-actions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.264, + 0.779, + 0.28 + ], + "angle": 0, + "content": "3.2. CoT Dataset: SandThink-21k" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.908, + 0.608 + ], + "angle": 0, + "content": "Effective robotic task execution requires goal comprehension, subgoal decomposition, and precise action sequencing. However, existing methods often lack structured reasoning mechanisms for long-horizon task decisions. To address this, we introduce a CoT reasoning dataset that enables deep reasoning before action generation, transforming concise instructions into structured long-horizon tasks for optimal execution. Our dataset follows a four-stage structure: (1) Goal Explanation; (2) Subgoal Decomposition; (3) Action Sequence Generation; (4) Action Sequence Modeling. By explicitly structuring these stages, our dataset mitigates fragmented planning and enhances decision-making. To improve model comprehension, each stage is marked with dedicated tags (e.g., , ..., ). Except for the final action sequence, intermediate responses are generated using GPT-4o, with task instructions rewritten for dataset expansion. Additionally, we incorporate 5,700 public complex surface images to leverage their similarity to sand terrain. The final dataset, SandThink-21k, consists of 21k CoT samples for model training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.628, + 0.753, + 0.646 + ], + "angle": 0, + "content": "4. The MindExplore System" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.657, + 0.681, + 0.673 + ], + "angle": 0, + "content": "4.1. Reasoning Layer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.907, + 0.728 + ], + "angle": 0, + "content": "To enhance reasoning abilities for long-horizon tasks, we employ a two-stage training strategy consisting of cold start followed by iterative direct preference optimization [7]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.825, + 0.759 + ], + "angle": 0, + "content": "4.1.1. Cold Start with Curriculum Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To promote structured reasoning in complex robotic environments, we employ SigLIP to match text instructions with collected images of Sandgo-21k in feature space, filtering out blurred image data caused by jitter, and finally cold-start training all linear layers of the Qwen2-VL-7B-Instruct model with LoRA. The training goal at this stage is to maximize the probability of generating an reasoning result given the image and instruction input. Let \\( X_{v} \\) denotes the image input and \\( X_{q} \\) denote the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "6841" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.387, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.089, + 0.898, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.364, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Figure 3. MindExplore System Overview. MindExplore comprises a reasoning layer for task decomposition and an acting layer for execution. The Reasoning Layer, built on Qwen2-VL, applying Chain-of-Thought (CoT) reasoning to generate structured meta-action instructions. The acting Layer, a Mixture of Policy Experts (MoPE), includes two Multimodal Diffusion Policies (MMDPs): a base expert using RGB-Depth and LiDAR for movement, and an arm expert using RGB-Depth for manipulation. The Memory Mechanism ensures real-time feedback, enabling the reasoning layer to dynamically adjust the actions of the acting layer for optimal execution." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.439, + 0.449, + 0.455 + ], + "angle": 0, + "content": "instruction input. The objective formula is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.462, + 0.482, + 0.502 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {c o l d - s t a r t}} = \\sum_ {i = 1} ^ {L} \\log P \\left(X _ {a, i} \\mid X _ {v}, X _ {q}, X _ {a, < i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.66 + ], + "angle": 0, + "content": "where \\( X_{r} \\) denotes the reasoning output and \\( L \\) is its length. To effectively guide training, we employ a curriculum learning approach [3], where data is arranged in order from easy to difficult. Initially, the model is trained on structured complex surface data, followed by real-world sand scenes. This progression ensures that the model first learns basic reasoning skills before transitioning to more complex tasks. By gradually increasing the task complexity, the model acquires structured reasoning capabilities and improves its ability to handle real-world uncertainty." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.351, + 0.68 + ], + "angle": 0, + "content": "4.1.2. Enhancing Reasoning with RL" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.681, + 0.484, + 0.759 + ], + "angle": 0, + "content": "While cold start establishes structured reasoning, reinforcement learning is used to further refine the model's reasoning quality. We employ Iterative DPO to optimize the model's decisions in multimodal scenarios. Given an image \\(I\\) and a task instruction \\(Q\\), the model generates an inference chain \\(R\\). DPO is based on the Bradley-Terry (BT) model:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.766, + 0.483, + 0.793 + ], + "angle": 0, + "content": "\\[\nP \\left(R ^ {+} > R ^ {-} \\mid I, Q\\right) = \\frac {\\exp \\left(r _ {\\varphi} (I , Q , R ^ {+})\\right)}{\\exp \\left(r _ {\\varphi} (I , Q , R ^ {+})\\right) + \\exp \\left(r _ {\\varphi} (I , Q , R ^ {-})\\right)}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.799, + 0.483, + 0.828 + ], + "angle": 0, + "content": "\\(r_{\\varphi}\\) is a parameterized reward model, \\(R^{+}\\) and \\(R^{-}\\) denote preferred and non-preferred reasoning sequences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.827, + 0.483, + 0.867 + ], + "angle": 0, + "content": "The policy model \\(\\pi_{\\theta}\\) is trained to maximize the preferred alignment inference, while reference model \\(\\pi_{\\mathrm{ref}}\\) is used to stabilize training process. The initial objective function is:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.872, + 0.483, + 0.9 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\pi_ {\\theta}} \\mathbb {E} _ {(I, Q, R ^ {+}, R ^ {-})} [ r _ {\\varphi} (I, Q, R) ] - \\beta D _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} (R | I, Q) \\| \\pi_ {\\text {r e f}} (R | I, Q)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.515 + ], + "angle": 0, + "content": "To stabilize training, we also incorporate a standard supervised fine-tuning loss, ensuring that the model retains the structured reasoning of supervised fine-tuning while benefiting from preference-based optimization. The loss function is given by the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.546, + 0.905, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {\\mathrm {D P O}} (\\theta) = - \\mathbb {E} _ {(I, Q, R ^ {+}, R ^ {-})} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} (R ^ {+} | I , Q)}{\\pi_ {r e f} (R ^ {+} | I , Q)} \\right. \\right. \\\\ \\left. - \\beta \\log \\frac {\\pi_ {\\theta} \\left(R ^ {-} | I , Q\\right)}{\\pi_ {r e f} \\left(R ^ {-} | I , Q\\right)}\\right) + \\alpha \\frac {\\log \\pi_ {\\theta} \\left(R ^ {+} | I , Q\\right)}{| R ^ {+} |} ]. \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.675 + ], + "angle": 0, + "content": "where \\(\\theta\\) denotes the trainable parameters of the model, \\(|R^{+}|\\) denotes preferred reasoning length, the hyperparameter \\(\\alpha\\) balances the two loss terms and \\(\\beta\\) is a scaling factor." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.906, + 0.76 + ], + "angle": 0, + "content": "To build high-quality preference data, we refine the selection process in 3 rounds of training. In each iteration, we sample 8 inference outputs for each task and rank them using GPT-4o with preference scores ranging from 0 to 100. The selection follows a progressive rejection strategy:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.77, + 0.905, + 0.8 + ], + "angle": 0, + "content": "- Iteration 1: The highest scoring response is selected as \\( R^{+} \\), while the lowest scoring response is taken as \\( R^{-} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.8, + 0.905, + 0.83 + ], + "angle": 0, + "content": "- Iteration 2: \\( R^{-} \\) is selected as the highest ranked response below 60, increasing the task complexity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.831, + 0.905, + 0.86 + ], + "angle": 0, + "content": "- Iteration 3: \\( R^{-} \\) is the lowest ranked response above 60, making model learn fine-grained reasoning distinctions." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.77, + 0.905, + 0.86 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This approach enables models to excel in structured multimodal reasoning and long-horizon robotic planning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6842" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.23, + 0.108 + ], + "angle": 0, + "content": "4.2. Acting Layer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.234 + ], + "angle": 0, + "content": "The Acting Layer generates a set of manipulation parameters and focuses on optimizing the joint probability distribution of all outputs. Therefore, generative diffusion models excelling in expressive capability and sampling quality [12] can be applied. Since the data dimensions of manipulation parameters is much smaller than that of image, the acting layer could get fast generation speed to provide real-time feedback to dynamic changes [18, 38, 39]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.243, + 0.334, + 0.259 + ], + "angle": 0, + "content": "4.2.1. Multimodal Diffusion Policy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.484, + 0.428 + ], + "angle": 0, + "content": "Most VLAs use only single-modal sensor data, which struggles to provide precise manipulation for instruction following in complex environments. In challenging visual settings, RGB images often contain significant noise, such as the similarity in color between the cube in Figure 2 and the sandy terrain. On the other hand, models based solely on depth or point cloud data lack the ability to perceive scene details and cannot align with the semantic features of instructions. To address these, we propose a multimodal diffusion Policy model (MMDP) that jointly exploits information of multi-sensor data and textual instructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.505 + ], + "angle": 0, + "content": "We employ the robot manipulation visual encoder R3M, proposed in [25], to encode the RGB images. Instead of concatenating the depth and RGB images along the channel dimension, we transform them into point cloud data using a perspective projection model:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.516, + 0.483, + 0.55 + ], + "angle": 0, + "content": "\\[\nX (u, v) = \\frac {\\left(u - c _ {x}\\right) \\cdot D (u , v)}{f _ {x}}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.562, + 0.483, + 0.597 + ], + "angle": 0, + "content": "\\[\nY (u, v) = \\frac {(v - c _ {y}) \\cdot D (u , v)}{f _ {y}}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.605, + 0.482, + 0.621 + ], + "angle": 0, + "content": "\\[\nZ (u, v) = D (u, v), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.794 + ], + "angle": 0, + "content": "where \\( u \\) and \\( v \\) are the pixel coordinates in the depth image, \\( D(u,v) \\) is the depth value of the pixel, \\( f_{x} \\) and \\( f_{y} \\) are the camera's focal lengths, \\( c_{x} \\) and \\( c_{y} \\) are the image's principal point coordinates, and \\( (X,Y,Z) \\) represents the 3D coordinates of the point cloud data. We use the pyramid convolutional encoder from iDP3 [38] to encode point cloud data, while DistilBERT [27] is employed to encode textual instructions into feature vectors. Additionally, we propose a fusion adapters for both the sensor data and text, consisting of two linear layers, to map encoder outputs into an unified feature space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The manipulation parameter generation process of the MMDP is based on the conditional denoising diffusion generation of a lightweight 1D convolutional diffusion model. Specifically, we sample a random Gaussian noise \\(a^K\\) and perform \\(k\\) iterations of denoising using a conditional denoising network \\(\\epsilon_{\\theta}\\), conditioned on RGB data \\(r\\), depth information \\(d\\), LiDAR data \\(p\\), instruction data \\(i\\), and the robot's" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.091, + 0.903, + 0.107 + ], + "angle": 0, + "content": "state \\(q\\), to recover \\(a^K\\) into the manipulation parameters \\(a^0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.118, + 0.905, + 0.15 + ], + "angle": 0, + "content": "\\[\na ^ {k - 1} = \\alpha_ {k} \\left(a ^ {k} - \\gamma_ {k} \\epsilon_ {\\theta} \\left(a ^ {k}, k, r, d, p, i, q\\right)\\right) + \\sigma_ {k} \\mathcal {N} (0, I), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.15, + 0.907, + 0.271 + ], + "angle": 0, + "content": "where \\(\\mathcal{N}(0,I)\\) is Gaussian noise, \\(\\alpha_{k},\\gamma_{k}\\), and \\(\\sigma_{k}\\) are functions of \\(k\\) and depend on the noise scheduler [12]. During training, we iteratively add random noise \\(\\epsilon^k\\) for \\(k\\) steps to the manipulation parameters, resulting in the noisy action \\(\\bar{a}_k a^0 +\\bar{\\beta}_k\\epsilon^k\\), where \\(\\bar{a}_{k}\\) and \\(\\bar{\\beta}_k\\) are the noise schedule parameters that performs one step noise adding. The denoising network \\(\\epsilon_{\\theta}\\) is then used to predict the random noise \\(\\epsilon^k\\), and the training loss function is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.282, + 0.905, + 0.301 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\operatorname {M S E} \\left(\\epsilon^ {k}, \\epsilon_ {\\theta} \\left(\\bar {\\alpha_ {k}} a ^ {0} + \\bar {\\beta_ {k}} \\epsilon^ {k}, k, r, d, p, i, q\\right)\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.313, + 0.741, + 0.328 + ], + "angle": 0, + "content": "4.2.2. Mixture of Policy Experts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.332, + 0.907, + 0.604 + ], + "angle": 0, + "content": "The acting layer needs to provide precise feedback to dynamic environmental changes with relatively low model complexity, we argue that a unified manipulation model and unfiltered sensor inputs may interfere with the acting layer's accurate manipulation of specific meta-actions. Therefore, we propose a Mixture of Policy Experts Model that decouples manipulation of the robot's base and arm. The base expert is primarily responsible for executing meta-actions related to movement, outputting only the linear and angular velocities of base. Multimodal inputs are filtered to include RGB images from the base camera, base point cloud data with depth values less than \\(1.5\\mathrm{m}\\), and LiDAR point clouds within a horizontal radius of \\(1.2\\mathrm{m}\\). The arm expert performs meta-actions for grasping and placing, outputting only the pose and velocity parameters for arm manipulation. Its sensor inputs are RGB images from both the base and arm cameras, base point cloud data with depth values less than \\(0.8\\mathrm{m}\\) and arm point clouds with depth values less than \\(0.5\\mathrm{m}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.614, + 0.71, + 0.631 + ], + "angle": 0, + "content": "4.3. Memory Mechanism" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.642, + 0.909, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.73, + 0.845, + 0.744 + ], + "angle": 0, + "content": "Figure 4. The details of Memory Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To enhance the coordination between the Reasoning Layer and the Acting Layer and achieve global control over both software and hardware, we design a memory mechanism with real-time feedback and gated routing. As shown in Figure 4, upon receiving a human instruction, the memory mechanism first processes multi-view RGB images and task instructions, which are fed into reasoning layer to generate a sequence of meta-action instructions. The gated routing network then sequentially assigns each meta-action to the corresponding VLA action expert, while real-time" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6843" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.219, + 0.09, + 0.777, + 0.104 + ], + "angle": 0, + "content": "Table 1. The success rates of different VLA models across 24 tasks in four long-horizon tasks." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.108, + 0.907, + 0.273 + ], + "angle": 0, + "content": "
Coarse-Grained Long-Horizon TasksIntensive Instruction-Following Tasks
MethodTask1Task2Task3Task4Task5Task6Avg. (%)Task7Task8Task9Task10Task11Task12Avg. (%)
OpenVLA0/100/103/100/102/101/1010.00%1/103/100/100/101/102/1011.67
RDT5/103/100/103/106/105/1036.67%5/103/102/100/105/105/1033.33
MindExplore10/108/108/107/1010/1010/1088.3310/109/1010/108/109/1010/1093.33
Cross-Expert TasksZero-Shot Generalization Tasks
MethodTask13Task14Task15Task16Task17Task18Avg. (%)Task19Task20Task21Task22Task23Task24Avg. (%)
OpenVLA0/100/101/100/100/100/101.67%0/100/100/102/100/100/103.33
RDT4/103/105/101/102/106/1035.00%2/101/103/100/100/103/1015.00
MindExplore9/1010/1010/109/109/109/1093.3310/108/108/109/107/1010/1086.67
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.285, + 0.483, + 0.348 + ], + "angle": 0, + "content": "multimodal sensor data is provided to the corresponding acting layer's action expert to generate control parameters, which are subsequently used to execute actions via the hardware. The gated routing network can be formatted as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.356, + 0.483, + 0.372 + ], + "angle": 0, + "content": "\\[\nw _ {A E}, w _ {B E} = \\operatorname {s o f t m a x} (\\operatorname {F F N} (\\operatorname {D i s t i l l B e r t} (m i))), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.381, + 0.482, + 0.434 + ], + "angle": 0, + "content": "\\[\nM o P E (m i) = \\left\\{ \\begin{array}{l l} f _ {A E} (m i, I _ {A E}), & \\text {i f} w _ {A E} > w _ {B E} \\\\ f _ {B E} (m i, I _ {B E}), & \\text {i f} w _ {A E} \\leq w _ {B E} \\end{array} , \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.691 + ], + "angle": 0, + "content": "where \\(mi\\) represents a single meta-action instruction, while \\(I_{AE}\\) and \\(I_{BE}\\) denote the multimodal sensor inputs required by the arm expert and base expert, respectively. During this process, the memory mechanism stores the initial multiview RGB images received by reasoning layer and continuously feeds the current state of the robot back to Acting Layer while it generates control parameters, enabling real-time adjustments. Additionally, the memory mechanism supplies reasoning layer with both current and historical RGB sensor images, equipping reasoning layer with multi-temporal environmental perception to determine the completion status of each meta-action. If a meta-action is successfully executed, the memory mechanism forwards the next action to acting layer. If an error occurs during execution, it triggers reasoning layer to replan the entire task, generating a new sequence of meta actions, thereby improving execution stability and robustness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.703, + 0.224, + 0.72 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.308, + 0.744 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We train the Reasoning Layer with SandThink-21k and the acting layer using individual meta-actions from the SandGo-1k, with each meta-action representing an episode. Movement and crossing obstacle are used to train the base expert, while grasping, placing, and writing are used to train the arm expert. During training, we randomly apply GPT40 for instruction augmentation to simulate task instructions with varying levels of granularity. The SAM model [15] is used for semantic segmentation of RGB images, replacing the black background with different colors for data aug" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.907, + 0.376 + ], + "angle": 0, + "content": "mentation. We evaluate the MindExplore system directly in the same sandpit scenario, using success rates as the metric, which is the ratio of successful trials to the total number of experiments. The tests are conducted with varying scene configurations and ensures that all comparison models are tested under the same conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.387, + 0.71, + 0.403 + ], + "angle": 0, + "content": "5.2. System Performance" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.41, + 0.799, + 0.424 + ], + "angle": 0, + "content": "5.2.1. Performance on Real World Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.52 + ], + "angle": 0, + "content": "We evaluate 24 real-world robotic tasks categorized into four types, each containing six distinct tasks. All these evaluated tasks can be divided into four categories according to their ability to complete long-horizon tasks, and we conducted 615 trials on real robots, ensuring its accuracy on meta-action and long-horizon tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.906, + 0.642 + ], + "angle": 0, + "content": "Coarse-Grained Long-Horizon Tasks. These tasks involve high-level, ambiguous instructions (e.g., \"grab the branch\"), requiring the system to first interpret and structure precise task sequences before execution. The robot must break down vague commands into actionable steps, Such as object grasping after obstacle avoidance (Task 1, 2), going to position writing (Task 3, 4), and navigating challenging terrain to complete sampling (Task 5, 6)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.763 + ], + "angle": 0, + "content": "Intensive Instruction Following Tasks. These tasks consist of sequential, interdependent sub-tasks, testing the system's compositional reasoning and execution fluency. The robot must understand and follow multi-step instructions with precision, such as continuous pick-and-place operations are performed in a dynamically changing environment (Task 7-9), and sampling is performed after multiple obstacle avoidance (Task 10-12)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.906, + 0.884 + ], + "angle": 0, + "content": "Cross-Expert Tasks. These tasks demand seamless coordination between base movement and arm manipulation, requiring the integration of multiple operational skills (e.g., moving, turning, grasping, and placing) through cross-modal reasoning. The robot must reason about motion and manipulation tasks in combination, such as grasping a specified object to place in a container (Task 13-15) and writing at a specified location (Task 16-18)." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Zero-Shot Generalization Tasks. These tasks evaluate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6844" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.35, + 0.089, + 0.612, + 0.102 + ], + "angle": 0, + "content": "Instructions Follow Long-horizon Tasks" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.11, + 0.905, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.255, + 0.905, + 0.283 + ], + "angle": 0, + "content": "Figure 5. During long-horizon tasks, the reasoning layer decomposes the instruction into a structured sequence of executable steps. The acting layer then orchestrates the hierarchical execution by leveraging the Base Expert and Arm Expert." + }, + { + "type": "table_caption", + "bbox": [ + 0.18, + 0.287, + 0.816, + 0.3 + ], + "angle": 0, + "content": "Table 2. Comparison results with the most powerful closed-source and open-source MLLMs on three tasks." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.307, + 0.907, + 0.421 + ], + "angle": 0, + "content": "
ModelCoarse-Grained \nLong-Horizon TasksIntensive Instruction \nFollowing TasksCross-Expert \nMulti-TasksThe Zero-Shot \nGeneralization Tasks
ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)
GPT-4o-latest33.0247.0037.5450.0041.5961.0029.1721.00
Gemini-2.0-Flash29.5243.0033.6949.0043.5355.0024.3618.00
InternVL2.5-8B14.7133.0017.4540.0019.3743.0019.8815.00
Qwen2.5-VL-7B17.3235.0018.9344.0025.3249.0023.7319.00
DeepSeek-VL211.8427.0014.6140.0019.4943.007.185.00
Reasoning Layer69.3587.0070.2381.0075.3589.0063.1773.00
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.522 + ], + "angle": 0, + "content": "the ability of system to adapt previously unseen environmental terrain (Tasks 19 to 21) and object types (Tasks 22 to 24). The robot must generalize over different objects (branches, bottles, and stones), recognize and manipulate objects despite variations in texture, shape, and weight, and adjust to new ground conditions (rugged, steep slopes)." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Table 3. Success Rates of the Acting Layer across 5 meta-actions: Moving to a Target Location (MTL), Crossing Obstacles (CO), Grasping a Specified Object (GSO), Placing the Object into a Designated Container (POC), Writing (Wr). Avg. - Average Success Rate, BE - Base Expert, AE - Arm Expert." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.609, + 0.481, + 0.678 + ], + "angle": 0, + "content": "
MethodMTLCOGSOPOCWrAvg. (%)
ACT13/1003/6537/10055/6019/5036.86
iDP347/10027/6549/10043/6021/5050.24
BE85/10058/65---87.12
AE--79/10060/6048/5091.67
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We select the OpenVLA [14] and the diffusion model RDT [18] as comparison methods. As shown in Table 1, in instruction-following, MindExplore's performance improves as instruction granularity increases, whereas OpenVLA and RDT perform better with coarse-grained instructions than with intensive instructions. We attribute this to MindExplore's strong reasoning capability, where more informative instructions guide the model to execute more precise actions, whereas OpenVLA and RDT tend to favor simple instruction descriptions and lack deep thoughts for long-horizon tasks. In multi-skill coordination, MindExplore outperforms single-expert models with a higher success rate due to the MoPE. In terms of generalization, compared to end-to-end MLLM and VLAs without world knowledge," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.581 + ], + "angle": 0, + "content": "MindExplore demonstrates strong generalization capabilities in unseen scenarios. Notably, the average sequence length for the cruising and obstacle avoidance task is 3.12, for the sampling task is 4.3, and for the writing task is 2. We find that the success rates of OpenVLA and RDT decrease progressively with the increasing meta-action length across three tasks. In Figure 5, MindExplore maintains a consistent success rates across sequences of varying lengths, benefiting from its superior task planning, decomposition capabilities, and the memory mechanism." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.827, + 0.608 + ], + "angle": 0, + "content": "5.2.2. Performance of Different Components" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.809 + ], + "angle": 0, + "content": "We also assess the performance of the two key components, reasoning and acting layer. For reasoning layer, we collect 480 RGB-instruction pairs across four test tasks as the evaluation dataset, and select the most powerful closed-source and open-source MLLMs for comparative experiments, including GPT-4o-latest, Gemini-2.0-Flash, InternVL2.5-8B [8], Qwen2.5-VL-7B [2], and DeepSeek-VL2 [34]. We use prompt to force all models to output a reasoning process similar to reasoning layer, and only focus on the output content of action sequence modeling, using ROUGE_L and meta-action planning accuracy to evaluate model performance. The experimental results are shown in Table 2. Reasoning layer performs best on all indicators and tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "For acting layer, we compare it with two typical manipulation models, the RGB-based ACT [11] and the point-cloud-based iDP3 [38]. Since both methods are not instruction-following models, a separate manipulation model must be trained for each meta-action. In contrast, the acting layer only requires training two expert models to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6845" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.226, + 0.09, + 0.768, + 0.104 + ], + "angle": 0, + "content": "Table 4. Ablation results with training strategy on three tasks. “+” stands for accumulation." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.108, + 0.907, + 0.213 + ], + "angle": 0, + "content": "
ModelCoarse-grained \nLong-horizon tasksIntensive instruction \nfollowing tasksCross-Expert \nmulti-tasksThe zero-shot \nGeneralization task
ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)
Qwen2-VL-7B13.5429.0015.2235.0023.1737.007.3119.00
+ Cold Start47.3562.0051.4268.0064.5775.0033.1643.00
+ iteration 154.1869.0054.5470.0067.1579.0035.3455.00
+ iteration 261.5875.0063.2979.0070.2381.0046.5764.00
+ iteration 369.3587.0070.2381.0075.3589.0063.1773.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.214, + 0.22, + 0.782, + 0.234 + ], + "angle": 0, + "content": "Table 5. Ablation study of the MindExplore system on the cruising and obstacle avoidance task." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.24, + 0.907, + 0.309 + ], + "angle": 0, + "content": "
MethodTask1Task2Task7Task8Task13Task14Task19Task20Avg. (%)
w/o Reasoning & Memory5/105/108/106/104/106/109/105/1060.00
w/o Memory8/107/1010/107/105/108/109/107/1076.25
MindExplore10/108/1010/109/109/1010/1010/108/1092.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.319, + 0.482, + 0.348 + ], + "angle": 0, + "content": "Table 6. Ablation Study on Input Modalities for Meta-Actions. \"PC\" - Point Cloud, \"2D\" - 2D Images." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.358, + 0.486, + 0.453 + ], + "angle": 0, + "content": "
MethodInput ModalityDepth TypeMTLGSO
AERGB--23/50
AEDepth+RGBPC-40/50
BERGB-5/50-
BELiDAR+DepthPC37/50-
BELiDAR+Depth+RGB2D32/50-
BELiDAR+Depth+RGBPC42/50-
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.606 + ], + "angle": 0, + "content": "control different hardware structures, which is parameter efficient. As shown in Table 3, the ACT performs worst on actions requiring long-distance movement, such as MTL and CO, indicating that point cloud with spatial information is crucial for movement manipulation. By combining spatial information with color and texture information from RGB images, the BE and AE achieve better alignment between object semantics and text instructions in real-world scenes, achieving the highest success rates across all meta-actions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.62, + 0.245, + 0.637 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Ablation results of Reasoning Layer. Compared to DeepSeek-VL2 and InternVL2.5, although they are on par with Qwen2-VL in some benchmark indicators, they lack scalability for real-world tasks in embodied systems, such as spatial calibration, semantic perception, and fine-grained object recognition in noisy backgrounds (e.g., inability to distinguish between sand and yellow branches). Therefore, we chose Qwen2-VL as the base model of the reasoning layer. To prove the effectiveness of the Reasoning Layer training method, we performed detailed ablation experiments on three tasks, which involved adding our training methods one by one to the basic model. As shown in Table 4, cold start enables the model to obtain structured reasoning capabilities. The three rounds of iterative DPO training further improve the model's reasoning quality and achieve self-improvement by continuously building high-quality preference datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.907, + 0.487 + ], + "angle": 0, + "content": "Ablation of Acting Layer. We explore the impact of different input modalities on acting layer. As shown in Table 6, the performance of the AE using pure vision declined due to the lack of spatial information, and the RGB-based BE achieves just a \\(10\\%\\) success rate in MTL. Similarly, the point-cloud-based BE also shows lower success rates, as point cloud data failed to align effectively with the semantic content of instructions. Additionally, we test concatenating Depth images with RGB images along the channel dimension. The results show that point cloud data in acting layer better leverages the spatial information from Depth images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.909, + 0.713 + ], + "angle": 0, + "content": "Ablation of MindExplore System. We conduct an ablation study on the core components of the MindExplore system. In Table 5, in the reasoning layer and acting layer system without the memory mechanism, Reasoning layer performs a single task decomposition and sequentially sends meta-action instructions to acting layer. However, without timely state feedback and error correction from acting layer, the system is highly susceptible to error accumulation, resulting in a decrease in success rate. Further, when the Reasoning layer component is removed and only the acting layer is used, the success rate on the cruising and obstacle avoidance task significantly drops due to the lack of deep reasoning and acting of the task instructions. Nonetheless, it still outperforms other VLAs, demonstrating that the MoPE structure can ensure accurate completion of base tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.634, + 0.74 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We propose MindExplore, a hierarchical embodied system designed for long-horizon tasks, including Reasoning, Acting, and Memory Mechanism. By aligning planning with action, MindExplore achieves strong generalization in dynamic environments. Besides, we introduce SandGo-1k and SandThink-21k, the first expert-level multimodal embodied and CoT datasets, bridging the data gap in challenging terrains. Real-world evaluations demonstrate MindExplore achieves better performance, ensuring high reliability in complex environments." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6846" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.193 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China (Grant 62322117, 62371365, and U24B20136), the Fundamental Research Funds for the Central Universities, and the Innovation Fund of Xidian University (Grant No. YJSJ25007)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.208, + 0.188, + 0.223 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Muneeb Ahmed, Brejesh Lall, Rajesh Kumar, and Arzad A Kherani. Towards estimation of human intent in assistive robotic teleoperation using kinaesthetic and visual feedback. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1928-1934, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.304, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[3] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning (ICML), pages 41-48, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.419, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[4] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. \\(\\pi 0\\): A vision-language-action flow model for general robot control, 2024. arXiv preprint arXiv:2410.24164. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[5] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[6] Shizhe Chen, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Sugar: Pre-training 3d visual representations for robotics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18049-18060, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[7] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[8] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[9] David B D'Ambrosio, Saminda Abeyruwan, Laura Graesser, Atil Iscen, Henri Ben Amor, Alex Bewley, Barney J Reed, Krista Reymann, Leila Takayama, Yuval Tassa, et al. Achieving human level competitive robot table tennis. arXiv preprint arXiv:2408.03906, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, et al. Robothor: An open simulation-to-real embodied ai platform. In Proceedings of the IEEE/CVF Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.233, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Vision and Pattern Recognition (CVPR), pages 3164-3174, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[11] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33:6840-6851, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.22, + 0.905, + 0.299 + ], + "angle": 0, + "content": "[13] Stephen James, Kentaro Wada, Tristan Laidlow, and Andrew J Davison. Coarse-to-fine q-attention: Efficient learning for visual robotic manipulation via discretisation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13739–13748, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[14] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.456, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[16] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Manipllm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18061-18070, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.54, + 0.907, + 0.607 + ], + "angle": 0, + "content": "[17] Haitao Lin, Yanwei Fu, and Xiangyang Xue. Pourit!: Weakly-supervised liquid perception from a single image for visual closed-loop robotic pouring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 241-251, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[18] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[19] Yuecheng Liu, Dafeng Chi, Shiguang Wu, Zhanguang Zhang, Yaochen Hu, Lingfeng Zhang, Yingxue Zhang, Shuang Wu, Tongtong Cao, Guowei Huang, et al. Spatialcot: Advancing spatial reasoning through coordinate alignment and chain-of-thought for embodied task planning. arXiv preprint arXiv:2501.10074, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[20] Jingpei Lu, Florian Richter, and Michael C Yip. Markerless camera-to-robot pose estimation via self-supervised sim-to-real transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21296-21306, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[21] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems (NeurIPS), 35: 2507-2521, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.944, + 0.516, + 0.955 + ], + "angle": 0, + "content": "6847" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[22] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[23] Xiao Ma, Sumit Patidar, Iain Haughton, and Stephen James. Hierarchical diffusion policy for kinematics-aware multitask robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18081-18090, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[24] Weixin Mao, Weiheng Zhong, Zhou Jiang, Dong Fang, Zhongyue Zhang, Zihan Lan, Fan Jia, Tiancai Wang, Haoqiang Fan, and Osamu Yoshie. Robomatrix: A skill-centric hierarchical framework for scalable robot task planning and execution in open-world. arXiv preprint arXiv:2412.00171, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.319, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[25] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[26] Max Polzin, Qinghua Guan, and Josie Hughes. Robotic locomotion through active and passive morphological adaptation in extreme outdoor environments. Science Robotics, 10(99): eadp6419, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[27] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[28] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2998-3009, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[29] Chan Hee Song, Valts Blukis, Jonathan Tremblay, Stephen Tyree, Yu Su, and Stan Birchfield. Robospatial: Teaching spatial understanding to 2d and 3d vision-language models for robotics. arXiv preprint arXiv:2411.16537, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[30] Enrica Tricomi, Francesco Missiroli, Michele Xiloyannis, Nicola Lotti, Xiaohui Zhang, Marios Stefanakis, Maximilian Theisen, Jürgen Bauer, Clemens Becker, and Lorenzo Massia. Soft robotic shorts improve outdoor walking efficiency in older adults. Nature Machine Intelligence, 6(10):1145-1155, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.799 + ], + "angle": 0, + "content": "[31] Kasun Weerakoon, Adarsh Jagan Sathyamoorthy, Mohamed Elnoor, and Dinesh Manocha. Vapor: Legged robot navigation in unstructured outdoor environments using offline reinforcement learning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 10344–10350. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[32] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems (NeurIPS), 35:24824-24837, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[33] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Peng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[34] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.302 + ], + "angle": 0, + "content": "[35] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4737-4746, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.4 + ], + "angle": 0, + "content": "[36] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.402, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[37] Michal Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.905, + 0.514 + ], + "angle": 0, + "content": "[38] Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xialin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with improved 3d diffusion policies. arXiv preprint arXiv:2410.10803, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[39] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. arXiv preprint arXiv:2403.03954, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[40] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[41] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.699, + 0.905, + 0.766 + ], + "angle": 0, + "content": "[42] Da Zhao, Haobo Luo, Yuxiao Tu, Chongxi Meng, and Tin Lun Lam. Snail-inspired robotic swarms: a hybrid connector drives collective adaptation in unstructured outdoor environments. Nature Communications, 15(1):3647, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.769, + 0.905, + 0.851 + ], + "angle": 0, + "content": "[43] Allan Zhou, Moo Jin Kim, Lirui Wang, Pete Florence, and Chelsea Finn. Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17907-17917, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.851 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.955 + ], + "angle": 0, + "content": "6848" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/d546fd0e-5080-4957-803b-5e64e0ab9d88_origin.pdf b/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/d546fd0e-5080-4957-803b-5e64e0ab9d88_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..69d7f73087a9e1dcce7a9f38c44120eea3b03f99 --- /dev/null +++ b/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/d546fd0e-5080-4957-803b-5e64e0ab9d88_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9affe7516e49550462122cb37d20f8c99644a920145c88323d70698fdecfdde6 +size 1582594 diff --git a/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/full.md b/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/full.md new file mode 100644 index 0000000000000000000000000000000000000000..67cb1b1d70b845a60c1c9da751435ab7aa36fc7f --- /dev/null +++ b/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/full.md @@ -0,0 +1,358 @@ +# Towards Long-Horizon Vision-Language-Action System: Reasoning, Acting and Memory + +Daixun Li $^{1,\ast}$ Yusi Zhang $^{1,\ast}$ Mingxiang Cao $^{1,\ast}$ Donglai Liu $^{1,\ast}$ Weiying Xie $^{1,\dagger}$ Tianlin Hui $^{1}$ Lunkai Lin $^{2}$ Zhiqiang Xie $^{2}$ Yunsong Li $^{1}$ Xidian University AgileX Robotics https://xdu-imagelab.github.io/MindExplore/ + +![](images/a2b8e8cd616ca0d87633ab0c999cf764ddc1b3d81606435c0d34f3aacc901680.jpg) +MindExplore System + +![](images/580b8f6b236f964b949835490d7766a6f350b35aa591d569a1f12538b10ffa05.jpg) +Reasoning Layer Validation + +Multi-Modal: RGB Depth LiDAR + +![](images/a45ccc13860191c2543a251c797a81ba6a21cd343588982b7ad9d45d7c350daf.jpg) +Dataset +SandGo-1k + +Chain of Thought + +![](images/b6cf4b1f4d3bff592f7bcc408a0177435794eedba6912bd8410b0d7dc87ec80f.jpg) +SandThink-21k + + + + + + + + + +... + +![](images/d5b012abac6000a0afacd869a2b11b55313b02f553109f23032baa101ba044eb.jpg) +Zero-Shot Tasks +Perceive seen & unseen object +Figure 1. Overview of MindExplore System: A hierarchical embodied intelligence system including reasoning, acting, and memory, offering state-of-the-art generalizability in highly dynamic scenes. Unlike conventional planar embodied intelligence models that consolidate all capabilities within a single layer, MindExplore introduces a novel expert-level hierarchical embodied system architecture. + +![](images/340f338f307043fc14744f17129e7ccb0033a656d1266646192701cbcbd217d3.jpg) +Long-Horizon Tasks +Complete multiple meta-actions + +![](images/2474028458195b74b26fc52465a47bf8f03e9869ededa24315db589e28c430ab.jpg) +Cross-Expert Tasks +Adapt to the dynamic sand scene + +![](images/a7045b82717b103a041edd1bbcb38f8d52e653e54cc867b4bd151873aecac3ef.jpg) +Instruction Following Tasks +Complete long-horizon tasks as instructed + +# Abstract + +Vision-Language-Action (VLA) is crucial for autonomous decision-making in embodied systems. While current methods have advanced single-skill abilities, their short-horizon capability limits applicability in real-world scenarios. To address this challenge, we innovatively propose MindExplore, a general hierarchical VLA system with cross-skill for long-horizon tasks in highly dynamic sand. The key insight is to iteratively align the knowledge domain of task planning and action execution. Thus, this task-oriented action enables outstanding generalization across a wide range of real-world scenarios. In the reasoning layer, task-specific chains of thought (CoT) are designed for planning long-horizon task sequences and providing meta-action signals. In the acting layer, a simple but powerful Mixture of Pol + +icy Experts strategy is built inspired by signals and multimodal inputs for adaptively selecting skill experts and generating closed-loop action sequences. Also, it integrates a lightweight Multimodal Diffusion Policy (MMDP) to enhance spatial perception by fusing multi-visual modality features. Besides, the pioneering memory mechanism establishes feedback between the reasoning and acting layers, facilitating adaptive execution of long-horizon tasks and real-time replanning. Notably, we create SandGo-1k and SandThink-21k, the first expert-level multimodal embodied dataset and CoT dataset tailored for sandy environments. At a high execution frequency of 30 FPS, MindExplore is $3.01 \times$ more successful than existing methods in unstructured and dynamic environments. + +# 1. Introduction + +Current Visual-Language-Action (VLA) data and methods mainly focus on short sequence tasks of single ability [6, 20], that is, single-skill sequences limited to either arm operations or base movements, making them effective for simple tasks requiring only "muscle memory". [13, 20, 24]. However, these approaches fail to generalize for long-horizon tasks in real-world scenarios, particularly in complex terrain action [14, 33]. This requires necessitates the agent to handle long-horizon and multi-skill instructions in highly dynamic environments [37]. Thus, successful execution in such settings needs deep consideration—Reasoning, precise control—Acting, and adaptive feedback—Memory [5]. + +To achieve this, three key challenges must be addressed: (1) Task-oriented Chain of Thought (CoT) reasoning requires the hierarchical decomposition of coarse-grained human instructions into executable sequences of metaactions, ensuring long-horizon alignment in dynamic environments [10, 20, 23]; (2) High-Precision Multimodal Control through a lightweight spatial perception framework fuses RGB, depth, and LiDAR inputs for robust motion calibration and manipulation accuracy [17, 35]; (3) Adaptive State Tracking that dynamically memorizes system states and task progress, enabling real-time closed-loop feedback between planning and acting to mitigate error propagation from perceptual drift or control misalignment [1, 43]. Consequently, this remains an open and worthwhile challenge so far. + +In this paper, we propose MindExplore for long-horizon tasks in complex and dynamic environments with a novel expert-level hierarchical embodied system architecture (see Figure 1). Our MindExplore is inspired by the discovery that continuous iterative alignment of the knowledge domain between task planning and action execution enables strong generalization across diverse real-world scenarios. On one hand we pursue task-oriented CoT in the Reasoning layer for effectively capturing meta-action sequences, while on the other hand we design a straightforward Mixture of Policy Experts (MoPE) strategy in the Acting Layer for flexibly generating closed-loop action sequences. Moreover, Memory Mechanism is designed to establish feedback between the reasoning and acting layers. Thus, the proposed MindExplore formulates long-horizon task execution into trainable skill expert policies, providing innovative guidance for hierarchical embodied intelligence systems without failures in highly dynamic environments. We also introduce a lightweight Multimodal Diffusion Policy (MMDP) that is dedicated to integrating RGB, depth and LiDAR data into a feature embedding space, thus enhancing spatial perception and motion control capabilities. To accurately evaluate complex tasks, we construct SandGo-1k and SandThink-21k as the first expert-level multimodal + +embodied dataset and CoT dataset, respectively. We also carry out experiments to show that, the system maintains high accuracy and efficient execution even in highly dynamic environments such as mobile and unstructured sand, which validates our architecture. Our key contributions are summarized as follows: + +- We propose MindExplore, a novel expert-level hierarchical embodied system to adapt long-horizon tasks in unstructured and dynamic environments. To our best knowledge, MindExplore is the first work to formulate Reasoning, Acting and Memory into a unified architecture. +- We create SandGo-1k, the first embodied intelligence dataset for mobile operations in complex environments. It includes text-vision (RGB, depth, and LiDAR) data covering complex movement episodes. Building on this, we further construct a multimodal CoT dataset for highly dynamic terrains, SandThink-21k. +- We empirically evaluate the accuracy on real-world experiments across diverse complex terrains with 24 distinct tasks, such as steep slopes, pits, and dunes. These experiments highlight MindExplore exceptional performance for long-horizon tasks in real-world scenarios. + +# 2. Related Work + +# 2.1. Visual Chain-of-Thought + +CoT reasoning [32] has become fundamental in MLLMs [40, 41], with applications in science [21], mathematics [22], and advanced question answering [36]. Recently, visual CoT has been applied to robotics [19, 29, 37] to tackle task-specific challenges. However, existing methods focus on single-instruction, single-step tasks, failing in complex scenarios that demand long-horizon task decomposition. We argue that orchestrating concise instructions into structured long-horizon tasks is a critical step toward autonomous robot decision-making. + +# 2.2. Embodied System + +The long-horizon tasks in open-world environments has been a major challenge in robotics research [9]. Numerous studies leverage MLLMs with extensive world knowledge to decompose instructions into subtasks, which are executed by lightweight VLA models [4, 16, 24, 28]. The key challenges can be summarized as follows: + +Challenge 1: How can instructions be expanded into long-horizon action sequences and executed flawlessly? + +Challenge 2: How to effectively accomplish movement and manipulation in complex and dynamic terrain? + +Challenge 3: How to overcome the data missing caused by the difficulty of simulation of highly dynamic scenes? + +# 3. The System Dataset + +To address Challenge 3, we collect a multi-task embodied operation dataset SandGo-1k, using a tracked mobile robot with a single-arm manipulator (as shown in Figure 2) in a real sand. Based on SandGo-1k, we generate a reasoning-acting embodied instruction dataset SandThink-21k. + +# 3.1. Embodied Dataset: SandGo-1k + +![](images/c97873e0f253848a25ba07e6c35fc10895d2cc9f7223f7317b3a5e313fabdf8d.jpg) +Figure 2. Left: AgileX Robotic arm degrees of freedom and environmental layout (rugged terrain, obstacles, pits, dunes). Middle: Base movement states. Right: Collected data content. + +Environmental Layout. In real-world tasks [26, 30, 31, 42], uneven sandy terrains dominate outdoor applications, making them essential for studying mobile manipulation systems. The high dynamic fluidity of sand demands adaptive adjustments, while its visual noise complicates perception and calibration. Enhancing generalization in unstructured environments remains a key challenge, and our experiments show that a system robust in sandy conditions can more easily generalize to simpler terrains like factory floors, asphalt roads, or regular soil. Therefore, to investigate the capabilities of mobile manipulation systems in highly dynamic and complex environments, we select sand as the primary experimental setting, including potholes, steep slopes, mounds, and rough roads, and construct a $6 \times 8$ meters real-world sandbox for data collection and evaluation, as depicted in Figure 2. + +Task Setup and Collection Process. We design three long-horizon meta-tasks: (1) Cruising and Obstacle Avoidance, training the model for environmental perception and motion planning in complex terrains; (2) Sampling, enabling the model to perform instruction-following, target localization, and coordinated base-arm movement; (3) Writing, enhancing the model's fine manipulation capabilities. Based on these, we can decompose and reassemble them into five meta-actions: (1) Moving to a Target Location, (2) Crossing Obstacles, (3) Grasping a Specified Object, (4) Placing the Object into a designated Container, and (5) + +Writing. During data collection, task instructions are manually written and decomposed into sequences of meta-actions with individual instructions. Each meta-action is saved as an episode, serving as the smallest unit of data collection. + +Data Content. The collected data includes: (1) Sensor information, consisting of RGB-Depth images from the two cameras and LiDAR point clouds; (2) Robot state and manipulation commands, including the arm's pose and velocity, as well as the base's angular and linear velocities; (3) Text instructions for long-horizon tasks and meta-actions. + +# 3.2. CoT Dataset: SandThink-21k + +Effective robotic task execution requires goal comprehension, subgoal decomposition, and precise action sequencing. However, existing methods often lack structured reasoning mechanisms for long-horizon task decisions. To address this, we introduce a CoT reasoning dataset that enables deep reasoning before action generation, transforming concise instructions into structured long-horizon tasks for optimal execution. Our dataset follows a four-stage structure: (1) Goal Explanation; (2) Subgoal Decomposition; (3) Action Sequence Generation; (4) Action Sequence Modeling. By explicitly structuring these stages, our dataset mitigates fragmented planning and enhances decision-making. To improve model comprehension, each stage is marked with dedicated tags (e.g., , ..., ). Except for the final action sequence, intermediate responses are generated using GPT-4o, with task instructions rewritten for dataset expansion. Additionally, we incorporate 5,700 public complex surface images to leverage their similarity to sand terrain. The final dataset, SandThink-21k, consists of 21k CoT samples for model training. + +# 4. The MindExplore System + +# 4.1. Reasoning Layer + +To enhance reasoning abilities for long-horizon tasks, we employ a two-stage training strategy consisting of cold start followed by iterative direct preference optimization [7]. + +# 4.1.1. Cold Start with Curriculum Learning + +To promote structured reasoning in complex robotic environments, we employ SigLIP to match text instructions with collected images of Sandgo-21k in feature space, filtering out blurred image data caused by jitter, and finally cold-start training all linear layers of the Qwen2-VL-7B-Instruct model with LoRA. The training goal at this stage is to maximize the probability of generating an reasoning result given the image and instruction input. Let $X_{v}$ denotes the image input and $X_{q}$ denote the + +![](images/97d586055f1c938e422aa255ad3073578292ff488ef0eb5f0256692580c9dd69.jpg) +Figure 3. MindExplore System Overview. MindExplore comprises a reasoning layer for task decomposition and an acting layer for execution. The Reasoning Layer, built on Qwen2-VL, applying Chain-of-Thought (CoT) reasoning to generate structured meta-action instructions. The acting Layer, a Mixture of Policy Experts (MoPE), includes two Multimodal Diffusion Policies (MMDPs): a base expert using RGB-Depth and LiDAR for movement, and an arm expert using RGB-Depth for manipulation. The Memory Mechanism ensures real-time feedback, enabling the reasoning layer to dynamically adjust the actions of the acting layer for optimal execution. + +![](images/1e95ffaecbe3b6a19331cac635640c79f1571db1ea2022174303f0d3f8d1a2bc.jpg) + +instruction input. The objective formula is as follows: + +$$ +L _ {\text {c o l d - s t a r t}} = \sum_ {i = 1} ^ {L} \log P \left(X _ {a, i} \mid X _ {v}, X _ {q}, X _ {a, < i}\right), \tag {1} +$$ + +where $X_{r}$ denotes the reasoning output and $L$ is its length. To effectively guide training, we employ a curriculum learning approach [3], where data is arranged in order from easy to difficult. Initially, the model is trained on structured complex surface data, followed by real-world sand scenes. This progression ensures that the model first learns basic reasoning skills before transitioning to more complex tasks. By gradually increasing the task complexity, the model acquires structured reasoning capabilities and improves its ability to handle real-world uncertainty. + +# 4.1.2. Enhancing Reasoning with RL + +While cold start establishes structured reasoning, reinforcement learning is used to further refine the model's reasoning quality. We employ Iterative DPO to optimize the model's decisions in multimodal scenarios. Given an image $I$ and a task instruction $Q$ , the model generates an inference chain $R$ . DPO is based on the Bradley-Terry (BT) model: + +$$ +P \left(R ^ {+} > R ^ {-} \mid I, Q\right) = \frac {\exp \left(r _ {\varphi} (I , Q , R ^ {+})\right)}{\exp \left(r _ {\varphi} (I , Q , R ^ {+})\right) + \exp \left(r _ {\varphi} (I , Q , R ^ {-})\right)}. \tag {2} +$$ + +$r_{\varphi}$ is a parameterized reward model, $R^{+}$ and $R^{-}$ denote preferred and non-preferred reasoning sequences. + +The policy model $\pi_{\theta}$ is trained to maximize the preferred alignment inference, while reference model $\pi_{\mathrm{ref}}$ is used to stabilize training process. The initial objective function is: + +$$ +\max _ {\pi_ {\theta}} \mathbb {E} _ {(I, Q, R ^ {+}, R ^ {-})} [ r _ {\varphi} (I, Q, R) ] - \beta D _ {\mathrm {K L}} \left(\pi_ {\theta} (R | I, Q) \| \pi_ {\text {r e f}} (R | I, Q)\right). \tag {3} +$$ + +To stabilize training, we also incorporate a standard supervised fine-tuning loss, ensuring that the model retains the structured reasoning of supervised fine-tuning while benefiting from preference-based optimization. The loss function is given by the following formula: + +$$ +\begin{array}{l} L _ {\mathrm {D P O}} (\theta) = - \mathbb {E} _ {(I, Q, R ^ {+}, R ^ {-})} \left[ \log \sigma \left(\beta \log \frac {\pi_ {\theta} (R ^ {+} | I , Q)}{\pi_ {r e f} (R ^ {+} | I , Q)} \right. \right. \\ \left. - \beta \log \frac {\pi_ {\theta} \left(R ^ {-} | I , Q\right)}{\pi_ {r e f} \left(R ^ {-} | I , Q\right)}\right) + \alpha \frac {\log \pi_ {\theta} \left(R ^ {+} | I , Q\right)}{| R ^ {+} |} ]. \tag {4} \\ \end{array} +$$ + +where $\theta$ denotes the trainable parameters of the model, $|R^{+}|$ denotes preferred reasoning length, the hyperparameter $\alpha$ balances the two loss terms and $\beta$ is a scaling factor. + +To build high-quality preference data, we refine the selection process in 3 rounds of training. In each iteration, we sample 8 inference outputs for each task and rank them using GPT-4o with preference scores ranging from 0 to 100. The selection follows a progressive rejection strategy: + +- Iteration 1: The highest scoring response is selected as $R^{+}$ , while the lowest scoring response is taken as $R^{-}$ . +- Iteration 2: $R^{-}$ is selected as the highest ranked response below 60, increasing the task complexity. +- Iteration 3: $R^{-}$ is the lowest ranked response above 60, making model learn fine-grained reasoning distinctions. + +This approach enables models to excel in structured multimodal reasoning and long-horizon robotic planning. + +# 4.2. Acting Layer + +The Acting Layer generates a set of manipulation parameters and focuses on optimizing the joint probability distribution of all outputs. Therefore, generative diffusion models excelling in expressive capability and sampling quality [12] can be applied. Since the data dimensions of manipulation parameters is much smaller than that of image, the acting layer could get fast generation speed to provide real-time feedback to dynamic changes [18, 38, 39]. + +# 4.2.1. Multimodal Diffusion Policy + +Most VLAs use only single-modal sensor data, which struggles to provide precise manipulation for instruction following in complex environments. In challenging visual settings, RGB images often contain significant noise, such as the similarity in color between the cube in Figure 2 and the sandy terrain. On the other hand, models based solely on depth or point cloud data lack the ability to perceive scene details and cannot align with the semantic features of instructions. To address these, we propose a multimodal diffusion Policy model (MMDP) that jointly exploits information of multi-sensor data and textual instructions. + +We employ the robot manipulation visual encoder R3M, proposed in [25], to encode the RGB images. Instead of concatenating the depth and RGB images along the channel dimension, we transform them into point cloud data using a perspective projection model: + +$$ +X (u, v) = \frac {\left(u - c _ {x}\right) \cdot D (u , v)}{f _ {x}}, \tag {5} +$$ + +$$ +Y (u, v) = \frac {(v - c _ {y}) \cdot D (u , v)}{f _ {y}}, \tag {6} +$$ + +$$ +Z (u, v) = D (u, v), \tag {7} +$$ + +where $u$ and $v$ are the pixel coordinates in the depth image, $D(u,v)$ is the depth value of the pixel, $f_{x}$ and $f_{y}$ are the camera's focal lengths, $c_{x}$ and $c_{y}$ are the image's principal point coordinates, and $(X,Y,Z)$ represents the 3D coordinates of the point cloud data. We use the pyramid convolutional encoder from iDP3 [38] to encode point cloud data, while DistilBERT [27] is employed to encode textual instructions into feature vectors. Additionally, we propose a fusion adapters for both the sensor data and text, consisting of two linear layers, to map encoder outputs into an unified feature space. + +The manipulation parameter generation process of the MMDP is based on the conditional denoising diffusion generation of a lightweight 1D convolutional diffusion model. Specifically, we sample a random Gaussian noise $a^K$ and perform $k$ iterations of denoising using a conditional denoising network $\epsilon_{\theta}$ , conditioned on RGB data $r$ , depth information $d$ , LiDAR data $p$ , instruction data $i$ , and the robot's + +state $q$ , to recover $a^K$ into the manipulation parameters $a^0$ : + +$$ +a ^ {k - 1} = \alpha_ {k} \left(a ^ {k} - \gamma_ {k} \epsilon_ {\theta} \left(a ^ {k}, k, r, d, p, i, q\right)\right) + \sigma_ {k} \mathcal {N} (0, I), \tag {8} +$$ + +where $\mathcal{N}(0,I)$ is Gaussian noise, $\alpha_{k},\gamma_{k}$ , and $\sigma_{k}$ are functions of $k$ and depend on the noise scheduler [12]. During training, we iteratively add random noise $\epsilon^k$ for $k$ steps to the manipulation parameters, resulting in the noisy action $\bar{a}_k a^0 +\bar{\beta}_k\epsilon^k$ , where $\bar{a}_{k}$ and $\bar{\beta}_k$ are the noise schedule parameters that performs one step noise adding. The denoising network $\epsilon_{\theta}$ is then used to predict the random noise $\epsilon^k$ , and the training loss function is as follows: + +$$ +\mathcal {L} = \operatorname {M S E} \left(\epsilon^ {k}, \epsilon_ {\theta} \left(\bar {\alpha_ {k}} a ^ {0} + \bar {\beta_ {k}} \epsilon^ {k}, k, r, d, p, i, q\right)\right). \tag {9} +$$ + +# 4.2.2. Mixture of Policy Experts + +The acting layer needs to provide precise feedback to dynamic environmental changes with relatively low model complexity, we argue that a unified manipulation model and unfiltered sensor inputs may interfere with the acting layer's accurate manipulation of specific meta-actions. Therefore, we propose a Mixture of Policy Experts Model that decouples manipulation of the robot's base and arm. The base expert is primarily responsible for executing meta-actions related to movement, outputting only the linear and angular velocities of base. Multimodal inputs are filtered to include RGB images from the base camera, base point cloud data with depth values less than $1.5\mathrm{m}$ , and LiDAR point clouds within a horizontal radius of $1.2\mathrm{m}$ . The arm expert performs meta-actions for grasping and placing, outputting only the pose and velocity parameters for arm manipulation. Its sensor inputs are RGB images from both the base and arm cameras, base point cloud data with depth values less than $0.8\mathrm{m}$ and arm point clouds with depth values less than $0.5\mathrm{m}$ . + +# 4.3. Memory Mechanism + +![](images/af49e470c804a8b47dc5c1ea6b13bbb4a8b48026d6dd3ed7dee99414b28295d8.jpg) +Figure 4. The details of Memory Mechanism + +To enhance the coordination between the Reasoning Layer and the Acting Layer and achieve global control over both software and hardware, we design a memory mechanism with real-time feedback and gated routing. As shown in Figure 4, upon receiving a human instruction, the memory mechanism first processes multi-view RGB images and task instructions, which are fed into reasoning layer to generate a sequence of meta-action instructions. The gated routing network then sequentially assigns each meta-action to the corresponding VLA action expert, while real-time + +Table 1. The success rates of different VLA models across 24 tasks in four long-horizon tasks. + +
Coarse-Grained Long-Horizon TasksIntensive Instruction-Following Tasks
MethodTask1Task2Task3Task4Task5Task6Avg. (%)Task7Task8Task9Task10Task11Task12Avg. (%)
OpenVLA0/100/103/100/102/101/1010.00%1/103/100/100/101/102/1011.67
RDT5/103/100/103/106/105/1036.67%5/103/102/100/105/105/1033.33
MindExplore10/108/108/107/1010/1010/1088.3310/109/1010/108/109/1010/1093.33
Cross-Expert TasksZero-Shot Generalization Tasks
MethodTask13Task14Task15Task16Task17Task18Avg. (%)Task19Task20Task21Task22Task23Task24Avg. (%)
OpenVLA0/100/101/100/100/100/101.67%0/100/100/102/100/100/103.33
RDT4/103/105/101/102/106/1035.00%2/101/103/100/100/103/1015.00
MindExplore9/1010/1010/109/109/109/1093.3310/108/108/109/107/1010/1086.67
+ +multimodal sensor data is provided to the corresponding acting layer's action expert to generate control parameters, which are subsequently used to execute actions via the hardware. The gated routing network can be formatted as: + +$$ +w _ {A E}, w _ {B E} = \operatorname {s o f t m a x} (\operatorname {F F N} (\operatorname {D i s t i l l B e r t} (m i))), \tag {10} +$$ + +$$ +M o P E (m i) = \left\{ \begin{array}{l l} f _ {A E} (m i, I _ {A E}), & \text {i f} w _ {A E} > w _ {B E} \\ f _ {B E} (m i, I _ {B E}), & \text {i f} w _ {A E} \leq w _ {B E} \end{array} , \right. \tag {11} +$$ + +where $mi$ represents a single meta-action instruction, while $I_{AE}$ and $I_{BE}$ denote the multimodal sensor inputs required by the arm expert and base expert, respectively. During this process, the memory mechanism stores the initial multiview RGB images received by reasoning layer and continuously feeds the current state of the robot back to Acting Layer while it generates control parameters, enabling real-time adjustments. Additionally, the memory mechanism supplies reasoning layer with both current and historical RGB sensor images, equipping reasoning layer with multi-temporal environmental perception to determine the completion status of each meta-action. If a meta-action is successfully executed, the memory mechanism forwards the next action to acting layer. If an error occurs during execution, it triggers reasoning layer to replan the entire task, generating a new sequence of meta actions, thereby improving execution stability and robustness. + +# 5. Experiments + +# 5.1. Implementation Details + +We train the Reasoning Layer with SandThink-21k and the acting layer using individual meta-actions from the SandGo-1k, with each meta-action representing an episode. Movement and crossing obstacle are used to train the base expert, while grasping, placing, and writing are used to train the arm expert. During training, we randomly apply GPT40 for instruction augmentation to simulate task instructions with varying levels of granularity. The SAM model [15] is used for semantic segmentation of RGB images, replacing the black background with different colors for data aug + +mentation. We evaluate the MindExplore system directly in the same sandpit scenario, using success rates as the metric, which is the ratio of successful trials to the total number of experiments. The tests are conducted with varying scene configurations and ensures that all comparison models are tested under the same conditions. + +# 5.2. System Performance + +# 5.2.1. Performance on Real World Tasks + +We evaluate 24 real-world robotic tasks categorized into four types, each containing six distinct tasks. All these evaluated tasks can be divided into four categories according to their ability to complete long-horizon tasks, and we conducted 615 trials on real robots, ensuring its accuracy on meta-action and long-horizon tasks. + +Coarse-Grained Long-Horizon Tasks. These tasks involve high-level, ambiguous instructions (e.g., "grab the branch"), requiring the system to first interpret and structure precise task sequences before execution. The robot must break down vague commands into actionable steps, Such as object grasping after obstacle avoidance (Task 1, 2), going to position writing (Task 3, 4), and navigating challenging terrain to complete sampling (Task 5, 6). + +Intensive Instruction Following Tasks. These tasks consist of sequential, interdependent sub-tasks, testing the system's compositional reasoning and execution fluency. The robot must understand and follow multi-step instructions with precision, such as continuous pick-and-place operations are performed in a dynamically changing environment (Task 7-9), and sampling is performed after multiple obstacle avoidance (Task 10-12). + +Cross-Expert Tasks. These tasks demand seamless coordination between base movement and arm manipulation, requiring the integration of multiple operational skills (e.g., moving, turning, grasping, and placing) through cross-modal reasoning. The robot must reason about motion and manipulation tasks in combination, such as grasping a specified object to place in a container (Task 13-15) and writing at a specified location (Task 16-18). + +Zero-Shot Generalization Tasks. These tasks evaluate + +![](images/b05042eba977c5b487cab9531da15317679cbe14b624aa229f7ae4d66372d95e.jpg) +Figure 5. During long-horizon tasks, the reasoning layer decomposes the instruction into a structured sequence of executable steps. The acting layer then orchestrates the hierarchical execution by leveraging the Base Expert and Arm Expert. + +Table 2. Comparison results with the most powerful closed-source and open-source MLLMs on three tasks. + +
ModelCoarse-Grained +Long-Horizon TasksIntensive Instruction +Following TasksCross-Expert +Multi-TasksThe Zero-Shot +Generalization Tasks
ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)
GPT-4o-latest33.0247.0037.5450.0041.5961.0029.1721.00
Gemini-2.0-Flash29.5243.0033.6949.0043.5355.0024.3618.00
InternVL2.5-8B14.7133.0017.4540.0019.3743.0019.8815.00
Qwen2.5-VL-7B17.3235.0018.9344.0025.3249.0023.7319.00
DeepSeek-VL211.8427.0014.6140.0019.4943.007.185.00
Reasoning Layer69.3587.0070.2381.0075.3589.0063.1773.00
+ +the ability of system to adapt previously unseen environmental terrain (Tasks 19 to 21) and object types (Tasks 22 to 24). The robot must generalize over different objects (branches, bottles, and stones), recognize and manipulate objects despite variations in texture, shape, and weight, and adjust to new ground conditions (rugged, steep slopes). + +Table 3. Success Rates of the Acting Layer across 5 meta-actions: Moving to a Target Location (MTL), Crossing Obstacles (CO), Grasping a Specified Object (GSO), Placing the Object into a Designated Container (POC), Writing (Wr). Avg. - Average Success Rate, BE - Base Expert, AE - Arm Expert. + +
MethodMTLCOGSOPOCWrAvg. (%)
ACT13/1003/6537/10055/6019/5036.86
iDP347/10027/6549/10043/6021/5050.24
BE85/10058/65---87.12
AE--79/10060/6048/5091.67
+ +We select the OpenVLA [14] and the diffusion model RDT [18] as comparison methods. As shown in Table 1, in instruction-following, MindExplore's performance improves as instruction granularity increases, whereas OpenVLA and RDT perform better with coarse-grained instructions than with intensive instructions. We attribute this to MindExplore's strong reasoning capability, where more informative instructions guide the model to execute more precise actions, whereas OpenVLA and RDT tend to favor simple instruction descriptions and lack deep thoughts for long-horizon tasks. In multi-skill coordination, MindExplore outperforms single-expert models with a higher success rate due to the MoPE. In terms of generalization, compared to end-to-end MLLM and VLAs without world knowledge, + +MindExplore demonstrates strong generalization capabilities in unseen scenarios. Notably, the average sequence length for the cruising and obstacle avoidance task is 3.12, for the sampling task is 4.3, and for the writing task is 2. We find that the success rates of OpenVLA and RDT decrease progressively with the increasing meta-action length across three tasks. In Figure 5, MindExplore maintains a consistent success rates across sequences of varying lengths, benefiting from its superior task planning, decomposition capabilities, and the memory mechanism. + +# 5.2.2. Performance of Different Components + +We also assess the performance of the two key components, reasoning and acting layer. For reasoning layer, we collect 480 RGB-instruction pairs across four test tasks as the evaluation dataset, and select the most powerful closed-source and open-source MLLMs for comparative experiments, including GPT-4o-latest, Gemini-2.0-Flash, InternVL2.5-8B [8], Qwen2.5-VL-7B [2], and DeepSeek-VL2 [34]. We use prompt to force all models to output a reasoning process similar to reasoning layer, and only focus on the output content of action sequence modeling, using ROUGE_L and meta-action planning accuracy to evaluate model performance. The experimental results are shown in Table 2. Reasoning layer performs best on all indicators and tasks. + +For acting layer, we compare it with two typical manipulation models, the RGB-based ACT [11] and the point-cloud-based iDP3 [38]. Since both methods are not instruction-following models, a separate manipulation model must be trained for each meta-action. In contrast, the acting layer only requires training two expert models to + +Table 4. Ablation results with training strategy on three tasks. “+” stands for accumulation. + +
ModelCoarse-grained +Long-horizon tasksIntensive instruction +following tasksCross-Expert +multi-tasksThe zero-shot +Generalization task
ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)
Qwen2-VL-7B13.5429.0015.2235.0023.1737.007.3119.00
+ Cold Start47.3562.0051.4268.0064.5775.0033.1643.00
+ iteration 154.1869.0054.5470.0067.1579.0035.3455.00
+ iteration 261.5875.0063.2979.0070.2381.0046.5764.00
+ iteration 369.3587.0070.2381.0075.3589.0063.1773.00
+ +Table 5. Ablation study of the MindExplore system on the cruising and obstacle avoidance task. + +
MethodTask1Task2Task7Task8Task13Task14Task19Task20Avg. (%)
w/o Reasoning & Memory5/105/108/106/104/106/109/105/1060.00
w/o Memory8/107/1010/107/105/108/109/107/1076.25
MindExplore10/108/1010/109/109/1010/1010/108/1092.50
+ +Table 6. Ablation Study on Input Modalities for Meta-Actions. "PC" - Point Cloud, "2D" - 2D Images. + +
MethodInput ModalityDepth TypeMTLGSO
AERGB--23/50
AEDepth+RGBPC-40/50
BERGB-5/50-
BELiDAR+DepthPC37/50-
BELiDAR+Depth+RGB2D32/50-
BELiDAR+Depth+RGBPC42/50-
+ +control different hardware structures, which is parameter efficient. As shown in Table 3, the ACT performs worst on actions requiring long-distance movement, such as MTL and CO, indicating that point cloud with spatial information is crucial for movement manipulation. By combining spatial information with color and texture information from RGB images, the BE and AE achieve better alignment between object semantics and text instructions in real-world scenes, achieving the highest success rates across all meta-actions. + +# 5.3. Ablation Study + +Ablation results of Reasoning Layer. Compared to DeepSeek-VL2 and InternVL2.5, although they are on par with Qwen2-VL in some benchmark indicators, they lack scalability for real-world tasks in embodied systems, such as spatial calibration, semantic perception, and fine-grained object recognition in noisy backgrounds (e.g., inability to distinguish between sand and yellow branches). Therefore, we chose Qwen2-VL as the base model of the reasoning layer. To prove the effectiveness of the Reasoning Layer training method, we performed detailed ablation experiments on three tasks, which involved adding our training methods one by one to the basic model. As shown in Table 4, cold start enables the model to obtain structured reasoning capabilities. The three rounds of iterative DPO training further improve the model's reasoning quality and achieve self-improvement by continuously building high-quality preference datasets. + +Ablation of Acting Layer. We explore the impact of different input modalities on acting layer. As shown in Table 6, the performance of the AE using pure vision declined due to the lack of spatial information, and the RGB-based BE achieves just a $10\%$ success rate in MTL. Similarly, the point-cloud-based BE also shows lower success rates, as point cloud data failed to align effectively with the semantic content of instructions. Additionally, we test concatenating Depth images with RGB images along the channel dimension. The results show that point cloud data in acting layer better leverages the spatial information from Depth images. + +Ablation of MindExplore System. We conduct an ablation study on the core components of the MindExplore system. In Table 5, in the reasoning layer and acting layer system without the memory mechanism, Reasoning layer performs a single task decomposition and sequentially sends meta-action instructions to acting layer. However, without timely state feedback and error correction from acting layer, the system is highly susceptible to error accumulation, resulting in a decrease in success rate. Further, when the Reasoning layer component is removed and only the acting layer is used, the success rate on the cruising and obstacle avoidance task significantly drops due to the lack of deep reasoning and acting of the task instructions. Nonetheless, it still outperforms other VLAs, demonstrating that the MoPE structure can ensure accurate completion of base tasks. + +# 6. Conclusion + +We propose MindExplore, a hierarchical embodied system designed for long-horizon tasks, including Reasoning, Acting, and Memory Mechanism. By aligning planning with action, MindExplore achieves strong generalization in dynamic environments. Besides, we introduce SandGo-1k and SandThink-21k, the first expert-level multimodal embodied and CoT datasets, bridging the data gap in challenging terrains. Real-world evaluations demonstrate MindExplore achieves better performance, ensuring high reliability in complex environments. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China (Grant 62322117, 62371365, and U24B20136), the Fundamental Research Funds for the Central Universities, and the Innovation Fund of Xidian University (Grant No. YJSJ25007). + +# References + +[1] Muneeb Ahmed, Brejesh Lall, Rajesh Kumar, and Arzad A Kherani. Towards estimation of human intent in assistive robotic teleoperation using kinaesthetic and visual feedback. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1928-1934, 2023. 2 +[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7 +[3] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning (ICML), pages 41-48, 2009. 4 +[4] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. $\pi 0$ : A vision-language-action flow model for general robot control, 2024. arXiv preprint arXiv:2410.24164. 2 +[5] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2 +[6] Shizhe Chen, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Sugar: Pre-training 3d visual representations for robotics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18049-18060, 2024. 2 +[7] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 3 +[8] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 7 +[9] David B D'Ambrosio, Saminda Abeyruwan, Laura Graesser, Atil Iscen, Henri Ben Amor, Alex Bewley, Barney J Reed, Krista Reymann, Leila Takayama, Yuval Tassa, et al. Achieving human level competitive robot table tennis. arXiv preprint arXiv:2408.03906, 2024. 2 +[10] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, et al. Robothor: An open simulation-to-real embodied ai platform. In Proceedings of the IEEE/CVF Conference on Computer + +Vision and Pattern Recognition (CVPR), pages 3164-3174, 2020. 2 +[11] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024. 7 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33:6840-6851, 2020. 5 +[13] Stephen James, Kentaro Wada, Tristan Laidlow, and Andrew J Davison. Coarse-to-fine q-attention: Efficient learning for visual robotic manipulation via discretisation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13739–13748, 2022. 2 +[14] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2, 7 +[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 6 +[16] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Manipllm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18061-18070, 2024. 2 +[17] Haitao Lin, Yanwei Fu, and Xiangyang Xue. Pourit!: Weakly-supervised liquid perception from a single image for visual closed-loop robotic pouring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 241-251, 2023. 2 +[18] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 5, 7 +[19] Yuecheng Liu, Dafeng Chi, Shiguang Wu, Zhanguang Zhang, Yaochen Hu, Lingfeng Zhang, Yingxue Zhang, Shuang Wu, Tongtong Cao, Guowei Huang, et al. Spatialcot: Advancing spatial reasoning through coordinate alignment and chain-of-thought for embodied task planning. arXiv preprint arXiv:2501.10074, 2025. 2 +[20] Jingpei Lu, Florian Richter, and Michael C Yip. Markerless camera-to-robot pose estimation via self-supervised sim-to-real transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21296-21306, 2023. 2 +[21] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems (NeurIPS), 35: 2507-2521, 2022. 2 + +[22] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 2 +[23] Xiao Ma, Sumit Patidar, Iain Haughton, and Stephen James. Hierarchical diffusion policy for kinematics-aware multitask robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18081-18090, 2024. 2 +[24] Weixin Mao, Weiheng Zhong, Zhou Jiang, Dong Fang, Zhongyue Zhang, Zihan Lan, Fan Jia, Tiancai Wang, Haoqiang Fan, and Osamu Yoshie. Robomatrix: A skill-centric hierarchical framework for scalable robot task planning and execution in open-world. arXiv preprint arXiv:2412.00171, 2024. 2 +[25] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 5 +[26] Max Polzin, Qinghua Guan, and Josie Hughes. Robotic locomotion through active and passive morphological adaptation in extreme outdoor environments. Science Robotics, 10(99): eadp6419, 2025. 3 +[27] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.5 +[28] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2998-3009, 2023. 2 +[29] Chan Hee Song, Valts Blukis, Jonathan Tremblay, Stephen Tyree, Yu Su, and Stan Birchfield. Robospatial: Teaching spatial understanding to 2d and 3d vision-language models for robotics. arXiv preprint arXiv:2411.16537, 2024. 2 +[30] Enrica Tricomi, Francesco Missiroli, Michele Xiloyannis, Nicola Lotti, Xiaohui Zhang, Marios Stefanakis, Maximilian Theisen, Jürgen Bauer, Clemens Becker, and Lorenzo Massia. Soft robotic shorts improve outdoor walking efficiency in older adults. Nature Machine Intelligence, 6(10):1145-1155, 2024. 3 +[31] Kasun Weerakoon, Adarsh Jagan Sathyamoorthy, Mohamed Elnoor, and Dinesh Manocha. Vapor: Legged robot navigation in unstructured outdoor environments using offline reinforcement learning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 10344–10350. IEEE, 2024. 3 +[32] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems (NeurIPS), 35:24824-24837, 2022. 2 +[33] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin + +Peng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. 2 +[34] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 7 +[35] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4737-4746, 2023. 2 +[36] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567, 2024. 2 +[37] Michal Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 2 +[38] Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xialin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with improved 3d diffusion policies. arXiv preprint arXiv:2410.10803, 2024. 5, 7 +[39] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. arXiv preprint arXiv:2403.03954, 2024. 5 +[40] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 2 +[41] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2 +[42] Da Zhao, Haobo Luo, Yuxiao Tu, Chongxi Meng, and Tin Lun Lam. Snail-inspired robotic swarms: a hybrid connector drives collective adaptation in unstructured outdoor environments. Nature Communications, 15(1):3647, 2024. 3 +[43] Allan Zhou, Moo Jin Kim, Lirui Wang, Pete Florence, and Chelsea Finn. Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17907-17917, 2023. 2 \ No newline at end of file diff --git a/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/images.zip b/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ffcc36644548b1251a9882072ff00a55a152638e --- /dev/null +++ b/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f447cb8f2e473a7d6daaa6746597b60f7e75bff389b229f11734aa06ebdd145 +size 732186 diff --git a/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/layout.json b/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..85e67025a8015767ef6067e9ef273efa13663c8a --- /dev/null +++ b/2025/Towards Long-Horizon Vision-Language-Action System_ Reasoning, Acting and Memory/layout.json @@ -0,0 +1,8778 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 481, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 481, + 140 + ], + "type": "text", + "content": "Towards Long-Horizon Vision-Language-Action System: Reasoning, Acting and Memory" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "text", + "content": "Daixun Li" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "text", + "content": " Yusi Zhang" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "text", + "content": " Mingxiang Cao" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "text", + "content": " Donglai Liu" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "text", + "content": " Weiying Xie" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "text", + "content": " Tianlin Hui" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "text", + "content": " Lunkai Lin" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "text", + "content": " Zhiqiang Xie" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "text", + "content": " Yunsong Li" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 552, + 219 + ], + "type": "text", + "content": " Xidian University AgileX Robotics https://xdu-imagelab.github.io/MindExplore/" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 245, + 220, + 353 + ], + "blocks": [ + { + "bbox": [ + 146, + 232, + 247, + 243 + ], + "lines": [ + { + "bbox": [ + 146, + 232, + 247, + 243 + ], + "spans": [ + { + "bbox": [ + 146, + 232, + 247, + 243 + ], + "type": "text", + "content": "MindExplore System" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 245, + 220, + 353 + ], + "lines": [ + { + "bbox": [ + 58, + 245, + 220, + 353 + ], + "spans": [ + { + "bbox": [ + 58, + 245, + 220, + 353 + ], + "type": "image", + "image_path": "a2b8e8cd616ca0d87633ab0c999cf764ddc1b3d81606435c0d34f3aacc901680.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 219, + 255, + 334, + 352 + ], + "blocks": [ + { + "bbox": [ + 249, + 247, + 309, + 255 + ], + "lines": [ + { + "bbox": [ + 249, + 247, + 309, + 255 + ], + "spans": [ + { + "bbox": [ + 249, + 247, + 309, + 255 + ], + "type": "text", + "content": "Reasoning Layer Validation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 255, + 334, + 352 + ], + "lines": [ + { + "bbox": [ + 219, + 255, + 334, + 352 + ], + "spans": [ + { + "bbox": [ + 219, + 255, + 334, + 352 + ], + "type": "image", + "image_path": "580b8f6b236f964b949835490d7766a6f350b35aa591d569a1f12538b10ffa05.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 348, + 259, + 463, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 259, + 463, + 270 + ], + "spans": [ + { + "bbox": [ + 348, + 259, + 463, + 270 + ], + "type": "text", + "content": "Multi-Modal: RGB Depth LiDAR" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 343, + 273, + 466, + 351 + ], + "blocks": [ + { + "bbox": [ + 425, + 232, + 462, + 243 + ], + "lines": [ + { + "bbox": [ + 425, + 232, + 462, + 243 + ], + "spans": [ + { + "bbox": [ + 425, + 232, + 462, + 243 + ], + "type": "text", + "content": "Dataset" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 245, + 425, + 255 + ], + "lines": [ + { + "bbox": [ + 383, + 245, + 425, + 255 + ], + "spans": [ + { + "bbox": [ + 383, + 245, + 425, + 255 + ], + "type": "text", + "content": "SandGo-1k" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 273, + 466, + 351 + ], + "lines": [ + { + "bbox": [ + 343, + 273, + 466, + 351 + ], + "spans": [ + { + "bbox": [ + 343, + 273, + 466, + 351 + ], + "type": "image", + "image_path": "a45ccc13860191c2543a251c797a81ba6a21cd343588982b7ad9d45d7c350daf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 480, + 260, + 541, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 480, + 260, + 541, + 270 + ], + "spans": [ + { + "bbox": [ + 480, + 260, + 541, + 270 + ], + "type": "text", + "content": "Chain of Thought" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 481, + 273, + 535, + 312 + ], + "blocks": [ + { + "bbox": [ + 480, + 244, + 539, + 255 + ], + "lines": [ + { + "bbox": [ + 480, + 244, + 539, + 255 + ], + "spans": [ + { + "bbox": [ + 480, + 244, + 539, + 255 + ], + "type": "text", + "content": "SandThink-21k" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 273, + 535, + 312 + ], + "lines": [ + { + "bbox": [ + 481, + 273, + 535, + 312 + ], + "spans": [ + { + "bbox": [ + 481, + 273, + 535, + 312 + ], + "type": "image", + "image_path": "b6cf4b1f4d3bff592f7bcc408a0177435794eedba6912bd8410b0d7dc87ec80f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 486, + 318, + 533, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 486, + 318, + 533, + 323 + ], + "spans": [ + { + "bbox": [ + 486, + 318, + 533, + 323 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 485, + 324, + 533, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 485, + 324, + 533, + 329 + ], + "spans": [ + { + "bbox": [ + 485, + 324, + 533, + 329 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 496, + 330, + 525, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 496, + 330, + 525, + 335 + ], + "spans": [ + { + "bbox": [ + 496, + 330, + 525, + 335 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 496, + 335, + 525, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 496, + 335, + 525, + 339 + ], + "spans": [ + { + "bbox": [ + 496, + 335, + 525, + 339 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 488, + 341, + 531, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 488, + 341, + 531, + 346 + ], + "spans": [ + { + "bbox": [ + 488, + 341, + 531, + 346 + ], + "type": "text", + "content": "..." + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 64, + 369, + 153, + 446 + ], + "blocks": [ + { + "bbox": [ + 74, + 357, + 143, + 367 + ], + "lines": [ + { + "bbox": [ + 74, + 357, + 143, + 367 + ], + "spans": [ + { + "bbox": [ + 74, + 357, + 143, + 367 + ], + "type": "text", + "content": "Zero-Shot Tasks" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 369, + 153, + 446 + ], + "lines": [ + { + "bbox": [ + 64, + 369, + 153, + 446 + ], + "spans": [ + { + "bbox": [ + 64, + 369, + 153, + 446 + ], + "type": "image", + "image_path": "d5b012abac6000a0afacd869a2b11b55313b02f553109f23032baa101ba044eb.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 447, + 161, + 457 + ], + "lines": [ + { + "bbox": [ + 57, + 447, + 161, + 457 + ], + "spans": [ + { + "bbox": [ + 57, + 447, + 161, + 457 + ], + "type": "text", + "content": "Perceive seen & unseen object" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 462, + 555, + 496 + ], + "lines": [ + { + "bbox": [ + 55, + 462, + 555, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 555, + 496 + ], + "type": "text", + "content": "Figure 1. Overview of MindExplore System: A hierarchical embodied intelligence system including reasoning, acting, and memory, offering state-of-the-art generalizability in highly dynamic scenes. Unlike conventional planar embodied intelligence models that consolidate all capabilities within a single layer, MindExplore introduces a novel expert-level hierarchical embodied system architecture." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 175, + 369, + 264, + 445 + ], + "blocks": [ + { + "bbox": [ + 179, + 357, + 261, + 368 + ], + "lines": [ + { + "bbox": [ + 179, + 357, + 261, + 368 + ], + "spans": [ + { + "bbox": [ + 179, + 357, + 261, + 368 + ], + "type": "text", + "content": "Long-Horizon Tasks" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 369, + 264, + 445 + ], + "lines": [ + { + "bbox": [ + 175, + 369, + 264, + 445 + ], + "spans": [ + { + "bbox": [ + 175, + 369, + 264, + 445 + ], + "type": "image", + "image_path": "340f338f307043fc14744f17129e7ccb0033a656d1266646192701cbcbd217d3.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 447, + 274, + 457 + ], + "lines": [ + { + "bbox": [ + 166, + 447, + 274, + 457 + ], + "spans": [ + { + "bbox": [ + 166, + 447, + 274, + 457 + ], + "type": "text", + "content": "Complete multiple meta-actions" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 281, + 369, + 395, + 445 + ], + "blocks": [ + { + "bbox": [ + 296, + 357, + 375, + 368 + ], + "lines": [ + { + "bbox": [ + 296, + 357, + 375, + 368 + ], + "spans": [ + { + "bbox": [ + 296, + 357, + 375, + 368 + ], + "type": "text", + "content": "Cross-Expert Tasks" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 281, + 369, + 395, + 445 + ], + "lines": [ + { + "bbox": [ + 281, + 369, + 395, + 445 + ], + "spans": [ + { + "bbox": [ + 281, + 369, + 395, + 445 + ], + "type": "image", + "image_path": "2474028458195b74b26fc52465a47bf8f03e9869ededa24315db589e28c430ab.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 447, + 393, + 457 + ], + "lines": [ + { + "bbox": [ + 282, + 447, + 393, + 457 + ], + "spans": [ + { + "bbox": [ + 282, + 447, + 393, + 457 + ], + "type": "text", + "content": "Adapt to the dynamic sand scene" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 413, + 369, + 547, + 445 + ], + "blocks": [ + { + "bbox": [ + 421, + 357, + 533, + 368 + ], + "lines": [ + { + "bbox": [ + 421, + 357, + 533, + 368 + ], + "spans": [ + { + "bbox": [ + 421, + 357, + 533, + 368 + ], + "type": "text", + "content": "Instruction Following Tasks" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 369, + 547, + 445 + ], + "lines": [ + { + "bbox": [ + 413, + 369, + 547, + 445 + ], + "spans": [ + { + "bbox": [ + 413, + 369, + 547, + 445 + ], + "type": "image", + "image_path": "a7045b82717b103a041edd1bbcb38f8d52e653e54cc867b4bd151873aecac3ef.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 447, + 551, + 457 + ], + "lines": [ + { + "bbox": [ + 409, + 447, + 551, + 457 + ], + "spans": [ + { + "bbox": [ + 409, + 447, + 551, + 457 + ], + "type": "text", + "content": "Complete long-horizon tasks as instructed" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 152, + 503, + 200, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 503, + 200, + 515 + ], + "spans": [ + { + "bbox": [ + 152, + 503, + 200, + 515 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 54, + 528, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 528, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 528, + 297, + 696 + ], + "type": "text", + "content": "Vision-Language-Action (VLA) is crucial for autonomous decision-making in embodied systems. While current methods have advanced single-skill abilities, their short-horizon capability limits applicability in real-world scenarios. To address this challenge, we innovatively propose MindExplore, a general hierarchical VLA system with cross-skill for long-horizon tasks in highly dynamic sand. The key insight is to iteratively align the knowledge domain of task planning and action execution. Thus, this task-oriented action enables outstanding generalization across a wide range of real-world scenarios. In the reasoning layer, task-specific chains of thought (CoT) are designed for planning long-horizon task sequences and providing meta-action signals. In the acting layer, a simple but powerful Mixture of Pol" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 504, + 556, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 556, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 556, + 672 + ], + "type": "text", + "content": "icy Experts strategy is built inspired by signals and multimodal inputs for adaptively selecting skill experts and generating closed-loop action sequences. Also, it integrates a lightweight Multimodal Diffusion Policy (MMDP) to enhance spatial perception by fusing multi-visual modality features. Besides, the pioneering memory mechanism establishes feedback between the reasoning and acting layers, facilitating adaptive execution of long-horizon tasks and real-time replanning. Notably, we create SandGo-1k and SandThink-21k, the first expert-level multimodal embodied dataset and CoT dataset tailored for sandy environments. At a high execution frequency of 30 FPS, MindExplore is " + }, + { + "bbox": [ + 313, + 504, + 556, + 672 + ], + "type": "inline_equation", + "content": "3.01 \\times" + }, + { + "bbox": [ + 313, + 504, + 556, + 672 + ], + "type": "text", + "content": " more successful than existing methods in unstructured and dynamic environments." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 140, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 140, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 140, + 712 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 142, + 703, + 232, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 703, + 232, + 712 + ], + "spans": [ + { + "bbox": [ + 142, + 703, + 232, + 712 + ], + "type": "text", + "content": "† Corresponding authors." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "text", + "content": "6839" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 247 + ], + "type": "text", + "content": "Current Visual-Language-Action (VLA) data and methods mainly focus on short sequence tasks of single ability [6, 20], that is, single-skill sequences limited to either arm operations or base movements, making them effective for simple tasks requiring only \"muscle memory\". [13, 20, 24]. However, these approaches fail to generalize for long-horizon tasks in real-world scenarios, particularly in complex terrain action [14, 33]. This requires necessitates the agent to handle long-horizon and multi-skill instructions in highly dynamic environments [37]. Thus, successful execution in such settings needs deep consideration—Reasoning, precise control—Acting, and adaptive feedback—Memory [5]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 295, + 426 + ], + "type": "text", + "content": "To achieve this, three key challenges must be addressed: (1) Task-oriented Chain of Thought (CoT) reasoning requires the hierarchical decomposition of coarse-grained human instructions into executable sequences of metaactions, ensuring long-horizon alignment in dynamic environments [10, 20, 23]; (2) High-Precision Multimodal Control through a lightweight spatial perception framework fuses RGB, depth, and LiDAR inputs for robust motion calibration and manipulation accuracy [17, 35]; (3) Adaptive State Tracking that dynamically memorizes system states and task progress, enabling real-time closed-loop feedback between planning and acting to mitigate error propagation from perceptual drift or control misalignment [1, 43]. Consequently, this remains an open and worthwhile challenge so far." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we propose MindExplore for long-horizon tasks in complex and dynamic environments with a novel expert-level hierarchical embodied system architecture (see Figure 1). Our MindExplore is inspired by the discovery that continuous iterative alignment of the knowledge domain between task planning and action execution enables strong generalization across diverse real-world scenarios. On one hand we pursue task-oriented CoT in the Reasoning layer for effectively capturing meta-action sequences, while on the other hand we design a straightforward Mixture of Policy Experts (MoPE) strategy in the Acting Layer for flexibly generating closed-loop action sequences. Moreover, Memory Mechanism is designed to establish feedback between the reasoning and acting layers. Thus, the proposed MindExplore formulates long-horizon task execution into trainable skill expert policies, providing innovative guidance for hierarchical embodied intelligence systems without failures in highly dynamic environments. We also introduce a lightweight Multimodal Diffusion Policy (MMDP) that is dedicated to integrating RGB, depth and LiDAR data into a feature embedding space, thus enhancing spatial perception and motion control capabilities. To accurately evaluate complex tasks, we construct SandGo-1k and SandThink-21k as the first expert-level multimodal" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "embodied dataset and CoT dataset, respectively. We also carry out experiments to show that, the system maintains high accuracy and efficient execution even in highly dynamic environments such as mobile and unstructured sand, which validates our architecture. Our key contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 147, + 553, + 338 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 313, + 147, + 553, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 553, + 206 + ], + "type": "text", + "content": "- We propose MindExplore, a novel expert-level hierarchical embodied system to adapt long-horizon tasks in unstructured and dynamic environments. To our best knowledge, MindExplore is the first work to formulate Reasoning, Acting and Memory into a unified architecture." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 207, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 553, + 277 + ], + "type": "text", + "content": "- We create SandGo-1k, the first embodied intelligence dataset for mobile operations in complex environments. It includes text-vision (RGB, depth, and LiDAR) data covering complex movement episodes. Building on this, we further construct a multimodal CoT dataset for highly dynamic terrains, SandThink-21k." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 279, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 553, + 338 + ], + "type": "text", + "content": "- We empirically evaluate the accuracy on real-world experiments across diverse complex terrains with 24 distinct tasks, such as steep slopes, pits, and dunes. These experiments highlight MindExplore exceptional performance for long-horizon tasks in real-world scenarios." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 359, + 400, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 400, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 400, + 371 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 381, + 454, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 454, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 454, + 395 + ], + "type": "text", + "content": "2.1. Visual Chain-of-Thought" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 402, + 555, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 522 + ], + "type": "text", + "content": "CoT reasoning [32] has become fundamental in MLLMs [40, 41], with applications in science [21], mathematics [22], and advanced question answering [36]. Recently, visual CoT has been applied to robotics [19, 29, 37] to tackle task-specific challenges. However, existing methods focus on single-instruction, single-step tasks, failing in complex scenarios that demand long-horizon task decomposition. We argue that orchestrating concise instructions into structured long-horizon tasks is a critical step toward autonomous robot decision-making." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 540, + 422, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 422, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 422, + 552 + ], + "type": "text", + "content": "2.2. Embodied System" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 560, + 553, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 553, + 632 + ], + "type": "text", + "content": "The long-horizon tasks in open-world environments has been a major challenge in robotics research [9]. Numerous studies leverage MLLMs with extensive world knowledge to decompose instructions into subtasks, which are executed by lightweight VLA models [4, 16, 24, 28]. The key challenges can be summarized as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 635, + 553, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 553, + 659 + ], + "type": "text", + "content": "Challenge 1: How can instructions be expanded into long-horizon action sequences and executed flawlessly?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 662, + 553, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 553, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 553, + 687 + ], + "type": "text", + "content": "Challenge 2: How to effectively accomplish movement and manipulation in complex and dynamic terrain?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Challenge 3: How to overcome the data missing caused by the difficulty of simulation of highly dynamic scenes?" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "text", + "content": "6840" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 172, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 172, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 172, + 85 + ], + "type": "text", + "content": "3. The System Dataset" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "To address Challenge 3, we collect a multi-task embodied operation dataset SandGo-1k, using a tracked mobile robot with a single-arm manipulator (as shown in Figure 2) in a real sand. Based on SandGo-1k, we generate a reasoning-acting embodied instruction dataset SandThink-21k." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 161, + 224, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 224, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 224, + 174 + ], + "type": "text", + "content": "3.1. Embodied Dataset: SandGo-1k" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 188, + 294, + 343 + ], + "blocks": [ + { + "bbox": [ + 58, + 188, + 294, + 343 + ], + "lines": [ + { + "bbox": [ + 58, + 188, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 58, + 188, + 294, + 343 + ], + "type": "image", + "image_path": "c97873e0f253848a25ba07e6c35fc10895d2cc9f7223f7317b3a5e313fabdf8d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 348, + 295, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 348, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 295, + 381 + ], + "type": "text", + "content": "Figure 2. Left: AgileX Robotic arm degrees of freedom and environmental layout (rugged terrain, obstacles, pits, dunes). Middle: Base movement states. Right: Collected data content." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": "Environmental Layout. In real-world tasks [26, 30, 31, 42], uneven sandy terrains dominate outdoor applications, making them essential for studying mobile manipulation systems. The high dynamic fluidity of sand demands adaptive adjustments, while its visual noise complicates perception and calibration. Enhancing generalization in unstructured environments remains a key challenge, and our experiments show that a system robust in sandy conditions can more easily generalize to simpler terrains like factory floors, asphalt roads, or regular soil. Therefore, to investigate the capabilities of mobile manipulation systems in highly dynamic and complex environments, we select sand as the primary experimental setting, including potholes, steep slopes, mounds, and rough roads, and construct a " + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "inline_equation", + "content": "6 \\times 8" + }, + { + "bbox": [ + 54, + 389, + 295, + 581 + ], + "type": "text", + "content": " meters real-world sandbox for data collection and evaluation, as depicted in Figure 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 715 + ], + "type": "text", + "content": "Task Setup and Collection Process. We design three long-horizon meta-tasks: (1) Cruising and Obstacle Avoidance, training the model for environmental perception and motion planning in complex terrains; (2) Sampling, enabling the model to perform instruction-following, target localization, and coordinated base-arm movement; (3) Writing, enhancing the model's fine manipulation capabilities. Based on these, we can decompose and reassemble them into five meta-actions: (1) Moving to a Target Location, (2) Crossing Obstacles, (3) Grasping a Specified Object, (4) Placing the Object into a designated Container, and (5)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "Writing. During data collection, task instructions are manually written and decomposed into sequences of meta-actions with individual instructions. Each meta-action is saved as an episode, serving as the smallest unit of data collection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 122, + 555, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 555, + 196 + ], + "type": "text", + "content": "Data Content. The collected data includes: (1) Sensor information, consisting of RGB-Depth images from the two cameras and LiDAR point clouds; (2) Robot state and manipulation commands, including the arm's pose and velocity, as well as the base's angular and linear velocities; (3) Text instructions for long-horizon tasks and meta-actions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 209, + 476, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 476, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 476, + 221 + ], + "type": "text", + "content": "3.2. CoT Dataset: SandThink-21k" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 228, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 481 + ], + "type": "text", + "content": "Effective robotic task execution requires goal comprehension, subgoal decomposition, and precise action sequencing. However, existing methods often lack structured reasoning mechanisms for long-horizon task decisions. To address this, we introduce a CoT reasoning dataset that enables deep reasoning before action generation, transforming concise instructions into structured long-horizon tasks for optimal execution. Our dataset follows a four-stage structure: (1) Goal Explanation; (2) Subgoal Decomposition; (3) Action Sequence Generation; (4) Action Sequence Modeling. By explicitly structuring these stages, our dataset mitigates fragmented planning and enhances decision-making. To improve model comprehension, each stage is marked with dedicated tags (e.g., , ..., ). Except for the final action sequence, intermediate responses are generated using GPT-4o, with task instructions rewritten for dataset expansion. Additionally, we incorporate 5,700 public complex surface images to leverage their similarity to sand terrain. The final dataset, SandThink-21k, consists of 21k CoT samples for model training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 497, + 460, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 460, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 460, + 511 + ], + "type": "text", + "content": "4. The MindExplore System" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 520, + 416, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 416, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 416, + 533 + ], + "type": "text", + "content": "4.1. Reasoning Layer" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 539, + 555, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 576 + ], + "type": "text", + "content": "To enhance reasoning abilities for long-horizon tasks, we employ a two-stage training strategy consisting of cold start followed by iterative direct preference optimization [7]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 588, + 504, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 504, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 504, + 601 + ], + "type": "text", + "content": "4.1.1. Cold Start with Curriculum Learning" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "To promote structured reasoning in complex robotic environments, we employ SigLIP to match text instructions with collected images of Sandgo-21k in feature space, filtering out blurred image data caused by jitter, and finally cold-start training all linear layers of the Qwen2-VL-7B-Instruct model with LoRA. The training goal at this stage is to maximize the probability of generating an reasoning result given the image and instruction input. Let " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "X_{v}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " denotes the image input and " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "X_{q}" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " denote the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6841" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 236, + 281 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 236, + 281 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 236, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 236, + 281 + ], + "type": "image", + "image_path": "97d586055f1c938e422aa255ad3073578292ff488ef0eb5f0256692580c9dd69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 288, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 54, + 288, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 555, + 344 + ], + "type": "text", + "content": "Figure 3. MindExplore System Overview. MindExplore comprises a reasoning layer for task decomposition and an acting layer for execution. The Reasoning Layer, built on Qwen2-VL, applying Chain-of-Thought (CoT) reasoning to generate structured meta-action instructions. The acting Layer, a Mixture of Policy Experts (MoPE), includes two Multimodal Diffusion Policies (MMDPs): a base expert using RGB-Depth and LiDAR for movement, and an arm expert using RGB-Depth for manipulation. The Memory Mechanism ensures real-time feedback, enabling the reasoning layer to dynamically adjust the actions of the acting layer for optimal execution." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 236, + 70, + 549, + 281 + ], + "blocks": [ + { + "bbox": [ + 236, + 70, + 549, + 281 + ], + "lines": [ + { + "bbox": [ + 236, + 70, + 549, + 281 + ], + "spans": [ + { + "bbox": [ + 236, + 70, + 549, + 281 + ], + "type": "image", + "image_path": "1e95ffaecbe3b6a19331cac635640c79f1571db1ea2022174303f0d3f8d1a2bc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 347, + 274, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 274, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 274, + 360 + ], + "type": "text", + "content": "instruction input. The objective formula is as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 365, + 294, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 365, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 80, + 365, + 294, + 397 + ], + "type": "interline_equation", + "content": "L _ {\\text {c o l d - s t a r t}} = \\sum_ {i = 1} ^ {L} \\log P \\left(X _ {a, i} \\mid X _ {v}, X _ {q}, X _ {a, < i}\\right), \\tag {1}", + "image_path": "6b51ad6aa9c47a2f2a2beea2a04e9273a3d6be822bada9b1b1c66c8f94a9966e.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 402, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 402, + 296, + 522 + ], + "type": "inline_equation", + "content": "X_{r}" + }, + { + "bbox": [ + 55, + 402, + 296, + 522 + ], + "type": "text", + "content": " denotes the reasoning output and " + }, + { + "bbox": [ + 55, + 402, + 296, + 522 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 402, + 296, + 522 + ], + "type": "text", + "content": " is its length. To effectively guide training, we employ a curriculum learning approach [3], where data is arranged in order from easy to difficult. Initially, the model is trained on structured complex surface data, followed by real-world sand scenes. This progression ensures that the model first learns basic reasoning skills before transitioning to more complex tasks. By gradually increasing the task complexity, the model acquires structured reasoning capabilities and improves its ability to handle real-world uncertainty." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 526, + 214, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 214, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 214, + 538 + ], + "type": "text", + "content": "4.1.2. Enhancing Reasoning with RL" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 539, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 296, + 601 + ], + "type": "text", + "content": "While cold start establishes structured reasoning, reinforcement learning is used to further refine the model's reasoning quality. We employ Iterative DPO to optimize the model's decisions in multimodal scenarios. Given an image " + }, + { + "bbox": [ + 55, + 539, + 296, + 601 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 539, + 296, + 601 + ], + "type": "text", + "content": " and a task instruction " + }, + { + "bbox": [ + 55, + 539, + 296, + 601 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 539, + 296, + 601 + ], + "type": "text", + "content": ", the model generates an inference chain " + }, + { + "bbox": [ + 55, + 539, + 296, + 601 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 539, + 296, + 601 + ], + "type": "text", + "content": ". DPO is based on the Bradley-Terry (BT) model:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 60, + 606, + 295, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 606, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 60, + 606, + 295, + 628 + ], + "type": "interline_equation", + "content": "P \\left(R ^ {+} > R ^ {-} \\mid I, Q\\right) = \\frac {\\exp \\left(r _ {\\varphi} (I , Q , R ^ {+})\\right)}{\\exp \\left(r _ {\\varphi} (I , Q , R ^ {+})\\right) + \\exp \\left(r _ {\\varphi} (I , Q , R ^ {-})\\right)}. \\tag {2}", + "image_path": "b77a70ed49ebf60f6eef7a946f91636306a10c21ebd21fafe2c503224dec350b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 632, + 295, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 295, + 655 + ], + "type": "inline_equation", + "content": "r_{\\varphi}" + }, + { + "bbox": [ + 55, + 632, + 295, + 655 + ], + "type": "text", + "content": " is a parameterized reward model, " + }, + { + "bbox": [ + 55, + 632, + 295, + 655 + ], + "type": "inline_equation", + "content": "R^{+}" + }, + { + "bbox": [ + 55, + 632, + 295, + 655 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 632, + 295, + 655 + ], + "type": "inline_equation", + "content": "R^{-}" + }, + { + "bbox": [ + 55, + 632, + 295, + 655 + ], + "type": "text", + "content": " denote preferred and non-preferred reasoning sequences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 295, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 686 + ], + "type": "text", + "content": "The policy model " + }, + { + "bbox": [ + 55, + 654, + 295, + 686 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 55, + 654, + 295, + 686 + ], + "type": "text", + "content": " is trained to maximize the preferred alignment inference, while reference model " + }, + { + "bbox": [ + 55, + 654, + 295, + 686 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 654, + 295, + 686 + ], + "type": "text", + "content": " is used to stabilize training process. The initial objective function is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 690, + 295, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 690, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 690, + 295, + 712 + ], + "type": "interline_equation", + "content": "\\max _ {\\pi_ {\\theta}} \\mathbb {E} _ {(I, Q, R ^ {+}, R ^ {-})} [ r _ {\\varphi} (I, Q, R) ] - \\beta D _ {\\mathrm {K L}} \\left(\\pi_ {\\theta} (R | I, Q) \\| \\pi_ {\\text {r e f}} (R | I, Q)\\right). \\tag {3}", + "image_path": "995a29cc351305f36b67778335326fdbd232ab1120c32f8f04d98bd57943642e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 347, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 407 + ], + "type": "text", + "content": "To stabilize training, we also incorporate a standard supervised fine-tuning loss, ensuring that the model retains the structured reasoning of supervised fine-tuning while benefiting from preference-based optimization. The loss function is given by the following formula:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 432, + 553, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 432, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 318, + 432, + 553, + 496 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {\\mathrm {D P O}} (\\theta) = - \\mathbb {E} _ {(I, Q, R ^ {+}, R ^ {-})} \\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} (R ^ {+} | I , Q)}{\\pi_ {r e f} (R ^ {+} | I , Q)} \\right. \\right. \\\\ \\left. - \\beta \\log \\frac {\\pi_ {\\theta} \\left(R ^ {-} | I , Q\\right)}{\\pi_ {r e f} \\left(R ^ {-} | I , Q\\right)}\\right) + \\alpha \\frac {\\log \\pi_ {\\theta} \\left(R ^ {+} | I , Q\\right)}{| R ^ {+} |} ]. \\tag {4} \\\\ \\end{array}", + "image_path": "d74ca7d0ccc1ecec4c127dd72f93adfb8b0625cbd70ede38f22a1aea3ee07dbf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 497, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 534 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 497, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 497, + 554, + 534 + ], + "type": "text", + "content": " denotes the trainable parameters of the model, " + }, + { + "bbox": [ + 313, + 497, + 554, + 534 + ], + "type": "inline_equation", + "content": "|R^{+}|" + }, + { + "bbox": [ + 313, + 497, + 554, + 534 + ], + "type": "text", + "content": " denotes preferred reasoning length, the hyperparameter " + }, + { + "bbox": [ + 313, + 497, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 497, + 554, + 534 + ], + "type": "text", + "content": " balances the two loss terms and " + }, + { + "bbox": [ + 313, + 497, + 554, + 534 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 497, + 554, + 534 + ], + "type": "text", + "content": " is a scaling factor." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 541, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 554, + 601 + ], + "type": "text", + "content": "To build high-quality preference data, we refine the selection process in 3 rounds of training. In each iteration, we sample 8 inference outputs for each task and rank them using GPT-4o with preference scores ranging from 0 to 100. The selection follows a progressive rejection strategy:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 609, + 553, + 681 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 313, + 609, + 553, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 553, + 633 + ], + "type": "text", + "content": "- Iteration 1: The highest scoring response is selected as " + }, + { + "bbox": [ + 313, + 609, + 553, + 633 + ], + "type": "inline_equation", + "content": "R^{+}" + }, + { + "bbox": [ + 313, + 609, + 553, + 633 + ], + "type": "text", + "content": ", while the lowest scoring response is taken as " + }, + { + "bbox": [ + 313, + 609, + 553, + 633 + ], + "type": "inline_equation", + "content": "R^{-}" + }, + { + "bbox": [ + 313, + 609, + 553, + 633 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 633, + 553, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 553, + 657 + ], + "type": "text", + "content": "- Iteration 2: " + }, + { + "bbox": [ + 313, + 633, + 553, + 657 + ], + "type": "inline_equation", + "content": "R^{-}" + }, + { + "bbox": [ + 313, + 633, + 553, + 657 + ], + "type": "text", + "content": " is selected as the highest ranked response below 60, increasing the task complexity." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 658, + 553, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 553, + 681 + ], + "type": "text", + "content": "- Iteration 3: " + }, + { + "bbox": [ + 313, + 658, + 553, + 681 + ], + "type": "inline_equation", + "content": "R^{-}" + }, + { + "bbox": [ + 313, + 658, + 553, + 681 + ], + "type": "text", + "content": " is the lowest ranked response above 60, making model learn fine-grained reasoning distinctions." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "This approach enables models to excel in structured multimodal reasoning and long-horizon robotic planning." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "text", + "content": "6842" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 140, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 140, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 140, + 85 + ], + "type": "text", + "content": "4.2. Acting Layer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 185 + ], + "type": "text", + "content": "The Acting Layer generates a set of manipulation parameters and focuses on optimizing the joint probability distribution of all outputs. Therefore, generative diffusion models excelling in expressive capability and sampling quality [12] can be applied. Since the data dimensions of manipulation parameters is much smaller than that of image, the acting layer could get fast generation speed to provide real-time feedback to dynamic changes [18, 38, 39]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 192, + 204, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 204, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 204, + 205 + ], + "type": "text", + "content": "4.2.1. Multimodal Diffusion Policy" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 207, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 338 + ], + "type": "text", + "content": "Most VLAs use only single-modal sensor data, which struggles to provide precise manipulation for instruction following in complex environments. In challenging visual settings, RGB images often contain significant noise, such as the similarity in color between the cube in Figure 2 and the sandy terrain. On the other hand, models based solely on depth or point cloud data lack the ability to perceive scene details and cannot align with the semantic features of instructions. To address these, we propose a multimodal diffusion Policy model (MMDP) that jointly exploits information of multi-sensor data and textual instructions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 399 + ], + "type": "text", + "content": "We employ the robot manipulation visual encoder R3M, proposed in [25], to encode the RGB images. Instead of concatenating the depth and RGB images along the channel dimension, we transform them into point cloud data using a perspective projection model:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 106, + 408, + 295, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 408, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 106, + 408, + 295, + 435 + ], + "type": "interline_equation", + "content": "X (u, v) = \\frac {\\left(u - c _ {x}\\right) \\cdot D (u , v)}{f _ {x}}, \\tag {5}", + "image_path": "d75732441ba20904f1e39d9bbe4861f84341d91247d13d20440f8f51d6e1a8e1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 112, + 445, + 295, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 445, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 112, + 445, + 295, + 472 + ], + "type": "interline_equation", + "content": "Y (u, v) = \\frac {(v - c _ {y}) \\cdot D (u , v)}{f _ {y}}, \\tag {6}", + "image_path": "adeb6d028e8520e554a36131a2fad97aae29fd608c2428c7114ebbac2e751557.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 479, + 294, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 479, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 129, + 479, + 294, + 491 + ], + "type": "interline_equation", + "content": "Z (u, v) = D (u, v), \\tag {7}", + "image_path": "a1b27c7458a984af05f732e9b39f97799fe871c1de32595f1c5f36175168ffcf.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "text", + "content": " are the pixel coordinates in the depth image, " + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "inline_equation", + "content": "D(u,v)" + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "text", + "content": " is the depth value of the pixel, " + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "inline_equation", + "content": "f_{y}" + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "text", + "content": " are the camera's focal lengths, " + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "inline_equation", + "content": "c_{x}" + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "inline_equation", + "content": "c_{y}" + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "text", + "content": " are the image's principal point coordinates, and " + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "inline_equation", + "content": "(X,Y,Z)" + }, + { + "bbox": [ + 55, + 498, + 295, + 628 + ], + "type": "text", + "content": " represents the 3D coordinates of the point cloud data. We use the pyramid convolutional encoder from iDP3 [38] to encode point cloud data, while DistilBERT [27] is employed to encode textual instructions into feature vectors. Additionally, we propose a fusion adapters for both the sensor data and text, consisting of two linear layers, to map encoder outputs into an unified feature space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "The manipulation parameter generation process of the MMDP is based on the conditional denoising diffusion generation of a lightweight 1D convolutional diffusion model. Specifically, we sample a random Gaussian noise " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "a^K" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " and perform " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " iterations of denoising using a conditional denoising network " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", conditioned on RGB data " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", depth information " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", LiDAR data " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", instruction data " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", and the robot's" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 552, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 552, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 552, + 84 + ], + "type": "text", + "content": "state " + }, + { + "bbox": [ + 313, + 72, + 552, + 84 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 72, + 552, + 84 + ], + "type": "text", + "content": ", to recover " + }, + { + "bbox": [ + 313, + 72, + 552, + 84 + ], + "type": "inline_equation", + "content": "a^K" + }, + { + "bbox": [ + 313, + 72, + 552, + 84 + ], + "type": "text", + "content": " into the manipulation parameters " + }, + { + "bbox": [ + 313, + 72, + 552, + 84 + ], + "type": "inline_equation", + "content": "a^0" + }, + { + "bbox": [ + 313, + 72, + 552, + 84 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 93, + 553, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 93, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 313, + 93, + 553, + 118 + ], + "type": "interline_equation", + "content": "a ^ {k - 1} = \\alpha_ {k} \\left(a ^ {k} - \\gamma_ {k} \\epsilon_ {\\theta} \\left(a ^ {k}, k, r, d, p, i, q\\right)\\right) + \\sigma_ {k} \\mathcal {N} (0, I), \\tag {8}", + "image_path": "ca1b6c37184fff3c46bfc2fbd64f530a36489425b47d024056a40d0f6cdd7d45.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "content": " is Gaussian noise, " + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\alpha_{k},\\gamma_{k}" + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "content": " are functions of " + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "content": " and depend on the noise scheduler [12]. During training, we iteratively add random noise " + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\epsilon^k" + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "content": " steps to the manipulation parameters, resulting in the noisy action " + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\bar{a}_k a^0 +\\bar{\\beta}_k\\epsilon^k" + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\bar{a}_{k}" + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\bar{\\beta}_k" + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "content": " are the noise schedule parameters that performs one step noise adding. The denoising network " + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "content": " is then used to predict the random noise " + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "inline_equation", + "content": "\\epsilon^k" + }, + { + "bbox": [ + 313, + 118, + 555, + 214 + ], + "type": "text", + "content": ", and the training loss function is as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 329, + 223, + 553, + 238 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 223, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 329, + 223, + 553, + 238 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\operatorname {M S E} \\left(\\epsilon^ {k}, \\epsilon_ {\\theta} \\left(\\bar {\\alpha_ {k}} a ^ {0} + \\bar {\\beta_ {k}} \\epsilon^ {k}, k, r, d, p, i, q\\right)\\right). \\tag {9}", + "image_path": "65e773606ecc3cf9308c2c735b155f97ac90630fea0692807c5df16e4f0b1cb4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 247, + 453, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 453, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 453, + 259 + ], + "type": "text", + "content": "4.2.2. Mixture of Policy Experts" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 262, + 555, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 478 + ], + "type": "text", + "content": "The acting layer needs to provide precise feedback to dynamic environmental changes with relatively low model complexity, we argue that a unified manipulation model and unfiltered sensor inputs may interfere with the acting layer's accurate manipulation of specific meta-actions. Therefore, we propose a Mixture of Policy Experts Model that decouples manipulation of the robot's base and arm. The base expert is primarily responsible for executing meta-actions related to movement, outputting only the linear and angular velocities of base. Multimodal inputs are filtered to include RGB images from the base camera, base point cloud data with depth values less than " + }, + { + "bbox": [ + 313, + 262, + 555, + 478 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 313, + 262, + 555, + 478 + ], + "type": "text", + "content": ", and LiDAR point clouds within a horizontal radius of " + }, + { + "bbox": [ + 313, + 262, + 555, + 478 + ], + "type": "inline_equation", + "content": "1.2\\mathrm{m}" + }, + { + "bbox": [ + 313, + 262, + 555, + 478 + ], + "type": "text", + "content": ". The arm expert performs meta-actions for grasping and placing, outputting only the pose and velocity parameters for arm manipulation. Its sensor inputs are RGB images from both the base and arm cameras, base point cloud data with depth values less than " + }, + { + "bbox": [ + 313, + 262, + 555, + 478 + ], + "type": "inline_equation", + "content": "0.8\\mathrm{m}" + }, + { + "bbox": [ + 313, + 262, + 555, + 478 + ], + "type": "text", + "content": " and arm point clouds with depth values less than " + }, + { + "bbox": [ + 313, + 262, + 555, + 478 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{m}" + }, + { + "bbox": [ + 313, + 262, + 555, + 478 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 486, + 434, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 434, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 434, + 499 + ], + "type": "text", + "content": "4.3. Memory Mechanism" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 508, + 556, + 570 + ], + "blocks": [ + { + "bbox": [ + 316, + 508, + 556, + 570 + ], + "lines": [ + { + "bbox": [ + 316, + 508, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 316, + 508, + 556, + 570 + ], + "type": "image", + "image_path": "af49e470c804a8b47dc5c1ea6b13bbb4a8b48026d6dd3ed7dee99414b28295d8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 578, + 517, + 589 + ], + "lines": [ + { + "bbox": [ + 351, + 578, + 517, + 589 + ], + "spans": [ + { + "bbox": [ + 351, + 578, + 517, + 589 + ], + "type": "text", + "content": "Figure 4. The details of Memory Mechanism" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "To enhance the coordination between the Reasoning Layer and the Acting Layer and achieve global control over both software and hardware, we design a memory mechanism with real-time feedback and gated routing. As shown in Figure 4, upon receiving a human instruction, the memory mechanism first processes multi-view RGB images and task instructions, which are fed into reasoning layer to generate a sequence of meta-action instructions. The gated routing network then sequentially assigns each meta-action to the corresponding VLA action expert, while real-time" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "text", + "content": "6843" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 85, + 555, + 216 + ], + "blocks": [ + { + "bbox": [ + 134, + 71, + 475, + 82 + ], + "lines": [ + { + "bbox": [ + 134, + 71, + 475, + 82 + ], + "spans": [ + { + "bbox": [ + 134, + 71, + 475, + 82 + ], + "type": "text", + "content": "Table 1. The success rates of different VLA models across 24 tasks in four long-horizon tasks." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 85, + 555, + 216 + ], + "lines": [ + { + "bbox": [ + 58, + 85, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 58, + 85, + 555, + 216 + ], + "type": "table", + "html": "
Coarse-Grained Long-Horizon TasksIntensive Instruction-Following Tasks
MethodTask1Task2Task3Task4Task5Task6Avg. (%)Task7Task8Task9Task10Task11Task12Avg. (%)
OpenVLA0/100/103/100/102/101/1010.00%1/103/100/100/101/102/1011.67
RDT5/103/100/103/106/105/1036.67%5/103/102/100/105/105/1033.33
MindExplore10/108/108/107/1010/1010/1088.3310/109/1010/108/109/1010/1093.33
Cross-Expert TasksZero-Shot Generalization Tasks
MethodTask13Task14Task15Task16Task17Task18Avg. (%)Task19Task20Task21Task22Task23Task24Avg. (%)
OpenVLA0/100/101/100/100/100/101.67%0/100/100/102/100/100/103.33
RDT4/103/105/101/102/106/1035.00%2/101/103/100/100/103/1015.00
MindExplore9/1010/1010/109/109/109/1093.3310/108/108/109/107/1010/1086.67
", + "image_path": "925fcd9044d8c83ae195edf68b289b8b54c21fe3ba67bd52175fc00c8598af2b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 225, + 295, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 295, + 275 + ], + "type": "text", + "content": "multimodal sensor data is provided to the corresponding acting layer's action expert to generate control parameters, which are subsequently used to execute actions via the hardware. The gated routing network can be formatted as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 281, + 295, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 281, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 67, + 281, + 295, + 294 + ], + "type": "interline_equation", + "content": "w _ {A E}, w _ {B E} = \\operatorname {s o f t m a x} (\\operatorname {F F N} (\\operatorname {D i s t i l l B e r t} (m i))), \\tag {10}", + "image_path": "fbcea5e0adef1093bc37455fde58f746e3cea961bb946c676a301ee36e8ce554.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 301, + 294, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 301, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 67, + 301, + 294, + 343 + ], + "type": "interline_equation", + "content": "M o P E (m i) = \\left\\{ \\begin{array}{l l} f _ {A E} (m i, I _ {A E}), & \\text {i f} w _ {A E} > w _ {B E} \\\\ f _ {B E} (m i, I _ {B E}), & \\text {i f} w _ {A E} \\leq w _ {B E} \\end{array} , \\right. \\tag {11}", + "image_path": "075f6a1217374682c544c9a185739137c3cd8d36d9b455612818659564493309.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 344, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 547 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 344, + 296, + 547 + ], + "type": "inline_equation", + "content": "mi" + }, + { + "bbox": [ + 55, + 344, + 296, + 547 + ], + "type": "text", + "content": " represents a single meta-action instruction, while " + }, + { + "bbox": [ + 55, + 344, + 296, + 547 + ], + "type": "inline_equation", + "content": "I_{AE}" + }, + { + "bbox": [ + 55, + 344, + 296, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 344, + 296, + 547 + ], + "type": "inline_equation", + "content": "I_{BE}" + }, + { + "bbox": [ + 55, + 344, + 296, + 547 + ], + "type": "text", + "content": " denote the multimodal sensor inputs required by the arm expert and base expert, respectively. During this process, the memory mechanism stores the initial multiview RGB images received by reasoning layer and continuously feeds the current state of the robot back to Acting Layer while it generates control parameters, enabling real-time adjustments. Additionally, the memory mechanism supplies reasoning layer with both current and historical RGB sensor images, equipping reasoning layer with multi-temporal environmental perception to determine the completion status of each meta-action. If a meta-action is successfully executed, the memory mechanism forwards the next action to acting layer. If an error occurs during execution, it triggers reasoning layer to replan the entire task, generating a new sequence of meta actions, thereby improving execution stability and robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 556, + 137, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 137, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 137, + 570 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 188, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 188, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 188, + 589 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "We train the Reasoning Layer with SandThink-21k and the acting layer using individual meta-actions from the SandGo-1k, with each meta-action representing an episode. Movement and crossing obstacle are used to train the base expert, while grasping, placing, and writing are used to train the arm expert. During training, we randomly apply GPT40 for instruction augmentation to simulate task instructions with varying levels of granularity. The SAM model [15] is used for semantic segmentation of RGB images, replacing the black background with different colors for data aug" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 225, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 297 + ], + "type": "text", + "content": "mentation. We evaluate the MindExplore system directly in the same sandpit scenario, using success rates as the metric, which is the ratio of successful trials to the total number of experiments. The tests are conducted with varying scene configurations and ensures that all comparison models are tested under the same conditions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 306, + 434, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 434, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 434, + 319 + ], + "type": "text", + "content": "5.2. System Performance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 324, + 488, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 488, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 488, + 335 + ], + "type": "text", + "content": "5.2.1. Performance on Real World Tasks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 340, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 411 + ], + "type": "text", + "content": "We evaluate 24 real-world robotic tasks categorized into four types, each containing six distinct tasks. All these evaluated tasks can be divided into four categories according to their ability to complete long-horizon tasks, and we conducted 615 trials on real robots, ensuring its accuracy on meta-action and long-horizon tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "type": "text", + "content": "Coarse-Grained Long-Horizon Tasks. These tasks involve high-level, ambiguous instructions (e.g., \"grab the branch\"), requiring the system to first interpret and structure precise task sequences before execution. The robot must break down vague commands into actionable steps, Such as object grasping after obstacle avoidance (Task 1, 2), going to position writing (Task 3, 4), and navigating challenging terrain to complete sampling (Task 5, 6)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 509, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 604 + ], + "type": "text", + "content": "Intensive Instruction Following Tasks. These tasks consist of sequential, interdependent sub-tasks, testing the system's compositional reasoning and execution fluency. The robot must understand and follow multi-step instructions with precision, such as continuous pick-and-place operations are performed in a dynamically changing environment (Task 7-9), and sampling is performed after multiple obstacle avoidance (Task 10-12)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 700 + ], + "type": "text", + "content": "Cross-Expert Tasks. These tasks demand seamless coordination between base movement and arm manipulation, requiring the integration of multiple operational skills (e.g., moving, turning, grasping, and placing) through cross-modal reasoning. The robot must reason about motion and manipulation tasks in combination, such as grasping a specified object to place in a container (Task 13-15) and writing at a specified location (Task 16-18)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 701, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 554, + 713 + ], + "type": "text", + "content": "Zero-Shot Generalization Tasks. These tasks evaluate" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6844" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 87, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 57, + 87, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 57, + 87, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 57, + 87, + 553, + 198 + ], + "type": "image", + "image_path": "b05042eba977c5b487cab9531da15317679cbe14b624aa229f7ae4d66372d95e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 553, + 224 + ], + "type": "text", + "content": "Figure 5. During long-horizon tasks, the reasoning layer decomposes the instruction into a structured sequence of executable steps. The acting layer then orchestrates the hierarchical execution by leveraging the Base Expert and Arm Expert." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 243, + 555, + 333 + ], + "blocks": [ + { + "bbox": [ + 110, + 227, + 499, + 237 + ], + "lines": [ + { + "bbox": [ + 110, + 227, + 499, + 237 + ], + "spans": [ + { + "bbox": [ + 110, + 227, + 499, + 237 + ], + "type": "text", + "content": "Table 2. Comparison results with the most powerful closed-source and open-source MLLMs on three tasks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 243, + 555, + 333 + ], + "lines": [ + { + "bbox": [ + 58, + 243, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 58, + 243, + 555, + 333 + ], + "type": "table", + "html": "
ModelCoarse-Grained \nLong-Horizon TasksIntensive Instruction \nFollowing TasksCross-Expert \nMulti-TasksThe Zero-Shot \nGeneralization Tasks
ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)
GPT-4o-latest33.0247.0037.5450.0041.5961.0029.1721.00
Gemini-2.0-Flash29.5243.0033.6949.0043.5355.0024.3618.00
InternVL2.5-8B14.7133.0017.4540.0019.3743.0019.8815.00
Qwen2.5-VL-7B17.3235.0018.9344.0025.3249.0023.7319.00
DeepSeek-VL211.8427.0014.6140.0019.4943.007.185.00
Reasoning Layer69.3587.0070.2381.0075.3589.0063.1773.00
", + "image_path": "4daa2c8da557e7fe45e9e4d9ef6500f01349722cdf001fccd28496ff6f4fbdfa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 340, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 413 + ], + "type": "text", + "content": "the ability of system to adapt previously unseen environmental terrain (Tasks 19 to 21) and object types (Tasks 22 to 24). The robot must generalize over different objects (branches, bottles, and stones), recognize and manipulate objects despite variations in texture, shape, and weight, and adjust to new ground conditions (rugged, steep slopes)." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 482, + 294, + 536 + ], + "blocks": [ + { + "bbox": [ + 55, + 424, + 296, + 479 + ], + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 479 + ], + "type": "text", + "content": "Table 3. Success Rates of the Acting Layer across 5 meta-actions: Moving to a Target Location (MTL), Crossing Obstacles (CO), Grasping a Specified Object (GSO), Placing the Object into a Designated Container (POC), Writing (Wr). Avg. - Average Success Rate, BE - Base Expert, AE - Arm Expert." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "lines": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 294, + 536 + ], + "type": "table", + "html": "
MethodMTLCOGSOPOCWrAvg. (%)
ACT13/1003/6537/10055/6019/5036.86
iDP347/10027/6549/10043/6021/5050.24
BE85/10058/65---87.12
AE--79/10060/6048/5091.67
", + "image_path": "5e88a75b01d38ed03fe82cfa621096143a2a7f9c850ca06d9fc94fe951245272.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "We select the OpenVLA [14] and the diffusion model RDT [18] as comparison methods. As shown in Table 1, in instruction-following, MindExplore's performance improves as instruction granularity increases, whereas OpenVLA and RDT perform better with coarse-grained instructions than with intensive instructions. We attribute this to MindExplore's strong reasoning capability, where more informative instructions guide the model to execute more precise actions, whereas OpenVLA and RDT tend to favor simple instruction descriptions and lack deep thoughts for long-horizon tasks. In multi-skill coordination, MindExplore outperforms single-expert models with a higher success rate due to the MoPE. In terms of generalization, compared to end-to-end MLLM and VLAs without world knowledge," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 340, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 460 + ], + "type": "text", + "content": "MindExplore demonstrates strong generalization capabilities in unseen scenarios. Notably, the average sequence length for the cruising and obstacle avoidance task is 3.12, for the sampling task is 4.3, and for the writing task is 2. We find that the success rates of OpenVLA and RDT decrease progressively with the increasing meta-action length across three tasks. In Figure 5, MindExplore maintains a consistent success rates across sequences of varying lengths, benefiting from its superior task planning, decomposition capabilities, and the memory mechanism." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 468, + 506, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 506, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 506, + 481 + ], + "type": "text", + "content": "5.2.2. Performance of Different Components" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 485, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 640 + ], + "type": "text", + "content": "We also assess the performance of the two key components, reasoning and acting layer. For reasoning layer, we collect 480 RGB-instruction pairs across four test tasks as the evaluation dataset, and select the most powerful closed-source and open-source MLLMs for comparative experiments, including GPT-4o-latest, Gemini-2.0-Flash, InternVL2.5-8B [8], Qwen2.5-VL-7B [2], and DeepSeek-VL2 [34]. We use prompt to force all models to output a reasoning process similar to reasoning layer, and only focus on the output content of action sequence modeling, using ROUGE_L and meta-action planning accuracy to evaluate model performance. The experimental results are shown in Table 2. Reasoning layer performs best on all indicators and tasks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "For acting layer, we compare it with two typical manipulation models, the RGB-based ACT [11] and the point-cloud-based iDP3 [38]. Since both methods are not instruction-following models, a separate manipulation model must be trained for each meta-action. In contrast, the acting layer only requires training two expert models to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 214, + 70, + 374, + 80 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 70, + 374, + 80 + ], + "spans": [ + { + "bbox": [ + 214, + 70, + 374, + 80 + ], + "type": "text", + "content": "Instructions Follow Long-horizon Tasks" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 757 + ], + "type": "text", + "content": "6845" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 85, + 555, + 168 + ], + "blocks": [ + { + "bbox": [ + 138, + 71, + 470, + 82 + ], + "lines": [ + { + "bbox": [ + 138, + 71, + 470, + 82 + ], + "spans": [ + { + "bbox": [ + 138, + 71, + 470, + 82 + ], + "type": "text", + "content": "Table 4. Ablation results with training strategy on three tasks. “+” stands for accumulation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 85, + 555, + 168 + ], + "lines": [ + { + "bbox": [ + 59, + 85, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 59, + 85, + 555, + 168 + ], + "type": "table", + "html": "
ModelCoarse-grained \nLong-horizon tasksIntensive instruction \nfollowing tasksCross-Expert \nmulti-tasksThe zero-shot \nGeneralization task
ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)ROUGE_L (%)Acc. (%)
Qwen2-VL-7B13.5429.0015.2235.0023.1737.007.3119.00
+ Cold Start47.3562.0051.4268.0064.5775.0033.1643.00
+ iteration 154.1869.0054.5470.0067.1579.0035.3455.00
+ iteration 261.5875.0063.2979.0070.2381.0046.5764.00
+ iteration 369.3587.0070.2381.0075.3589.0063.1773.00
", + "image_path": "88b990b40b46678bc1ad64a2ab809fd59a03a4ee2ffc8a04203f54efbd6dec50.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 190, + 555, + 244 + ], + "blocks": [ + { + "bbox": [ + 130, + 174, + 478, + 185 + ], + "lines": [ + { + "bbox": [ + 130, + 174, + 478, + 185 + ], + "spans": [ + { + "bbox": [ + 130, + 174, + 478, + 185 + ], + "type": "text", + "content": "Table 5. Ablation study of the MindExplore system on the cruising and obstacle avoidance task." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 190, + 555, + 244 + ], + "lines": [ + { + "bbox": [ + 59, + 190, + 555, + 244 + ], + "spans": [ + { + "bbox": [ + 59, + 190, + 555, + 244 + ], + "type": "table", + "html": "
MethodTask1Task2Task7Task8Task13Task14Task19Task20Avg. (%)
w/o Reasoning & Memory5/105/108/106/104/106/109/105/1060.00
w/o Memory8/107/1010/107/105/108/109/107/1076.25
MindExplore10/108/1010/109/109/1010/1010/108/1092.50
", + "image_path": "1e0d761a2d9f04a0f8ce97854d16fda22aa82b13f4e7236a4d51b7be4e7fcbec.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 59, + 283, + 297, + 358 + ], + "blocks": [ + { + "bbox": [ + 55, + 252, + 294, + 275 + ], + "lines": [ + { + "bbox": [ + 55, + 252, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 294, + 275 + ], + "type": "text", + "content": "Table 6. Ablation Study on Input Modalities for Meta-Actions. \"PC\" - Point Cloud, \"2D\" - 2D Images." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 283, + 297, + 358 + ], + "lines": [ + { + "bbox": [ + 59, + 283, + 297, + 358 + ], + "spans": [ + { + "bbox": [ + 59, + 283, + 297, + 358 + ], + "type": "table", + "html": "
MethodInput ModalityDepth TypeMTLGSO
AERGB--23/50
AEDepth+RGBPC-40/50
BERGB-5/50-
BELiDAR+DepthPC37/50-
BELiDAR+Depth+RGB2D32/50-
BELiDAR+Depth+RGBPC42/50-
", + "image_path": "38d9db2d213bd357744f36e9a8d6df1a851f2b48f97dbd405250149830c595b3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 371, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 479 + ], + "type": "text", + "content": "control different hardware structures, which is parameter efficient. As shown in Table 3, the ACT performs worst on actions requiring long-distance movement, such as MTL and CO, indicating that point cloud with spatial information is crucial for movement manipulation. By combining spatial information with color and texture information from RGB images, the BE and AE achieve better alignment between object semantics and text instructions in real-world scenes, achieving the highest success rates across all meta-actions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 491, + 149, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 149, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 149, + 504 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "Ablation results of Reasoning Layer. Compared to DeepSeek-VL2 and InternVL2.5, although they are on par with Qwen2-VL in some benchmark indicators, they lack scalability for real-world tasks in embodied systems, such as spatial calibration, semantic perception, and fine-grained object recognition in noisy backgrounds (e.g., inability to distinguish between sand and yellow branches). Therefore, we chose Qwen2-VL as the base model of the reasoning layer. To prove the effectiveness of the Reasoning Layer training method, we performed detailed ablation experiments on three tasks, which involved adding our training methods one by one to the basic model. As shown in Table 4, cold start enables the model to obtain structured reasoning capabilities. The three rounds of iterative DPO training further improve the model's reasoning quality and achieve self-improvement by continuously building high-quality preference datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 254, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 385 + ], + "type": "text", + "content": "Ablation of Acting Layer. We explore the impact of different input modalities on acting layer. As shown in Table 6, the performance of the AE using pure vision declined due to the lack of spatial information, and the RGB-based BE achieves just a " + }, + { + "bbox": [ + 313, + 254, + 555, + 385 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 254, + 555, + 385 + ], + "type": "text", + "content": " success rate in MTL. Similarly, the point-cloud-based BE also shows lower success rates, as point cloud data failed to align effectively with the semantic content of instructions. Additionally, we test concatenating Depth images with RGB images along the channel dimension. The results show that point cloud data in acting layer better leverages the spatial information from Depth images." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 386, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 556, + 564 + ], + "type": "text", + "content": "Ablation of MindExplore System. We conduct an ablation study on the core components of the MindExplore system. In Table 5, in the reasoning layer and acting layer system without the memory mechanism, Reasoning layer performs a single task decomposition and sequentially sends meta-action instructions to acting layer. However, without timely state feedback and error correction from acting layer, the system is highly susceptible to error accumulation, resulting in a decrease in success rate. Further, when the Reasoning layer component is removed and only the acting layer is used, the success rate on the cruising and obstacle avoidance task significantly drops due to the lack of deep reasoning and acting of the task instructions. Nonetheless, it still outperforms other VLAs, demonstrating that the MoPE structure can ensure accurate completion of base tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "We propose MindExplore, a hierarchical embodied system designed for long-horizon tasks, including Reasoning, Acting, and Memory Mechanism. By aligning planning with action, MindExplore achieves strong generalization in dynamic environments. Besides, we introduce SandGo-1k and SandThink-21k, the first expert-level multimodal embodied and CoT datasets, bridging the data gap in challenging terrains. Real-world evaluations demonstrate MindExplore achieves better performance, ensuring high reliability in complex environments." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6846" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China (Grant 62322117, 62371365, and U24B20136), the Fundamental Research Funds for the Central Universities, and the Innovation Fund of Xidian University (Grant No. YJSJ25007)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 184, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "type": "text", + "content": "[1] Muneeb Ahmed, Brejesh Lall, Rajesh Kumar, and Arzad A Kherani. Towards estimation of human intent in assistive robotic teleoperation using kinaesthetic and visual feedback. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1928-1934, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 283 + ], + "type": "text", + "content": "[2] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2.5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 330 + ], + "type": "text", + "content": "[3] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning (ICML), pages 41-48, 2009. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 331, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 295, + 386 + ], + "type": "text", + "content": "[4] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. " + }, + { + "bbox": [ + 62, + 331, + 295, + 386 + ], + "type": "inline_equation", + "content": "\\pi 0" + }, + { + "bbox": [ + 62, + 331, + 295, + 386 + ], + "type": "text", + "content": ": A vision-language-action flow model for general robot control, 2024. arXiv preprint arXiv:2410.24164. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "type": "text", + "content": "[5] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. pi0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 296, + 498 + ], + "type": "text", + "content": "[6] Shizhe Chen, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Sugar: Pre-training 3d visual representations for robotics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18049-18060, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 500, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 296, + 544 + ], + "type": "text", + "content": "[7] Zixiang Chen, Yihe Deng, Huizhuo Yuan, Kaixuan Ji, and Quanquan Gu. Self-play fine-tuning converts weak language models to strong language models. arXiv preprint arXiv:2401.01335, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 601 + ], + "type": "text", + "content": "[8] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271, 2024. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 296, + 656 + ], + "type": "text", + "content": "[9] David B D'Ambrosio, Saminda Abeyruwan, Laura Graesser, Atil Iscen, Henri Ben Amor, Alex Bewley, Barney J Reed, Krista Reymann, Leila Takayama, Yuval Tassa, et al. Achieving human level competitive robot table tennis. arXiv preprint arXiv:2408.03906, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[10] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, et al. Robothor: An open simulation-to-real embodied ai platform. In Proceedings of the IEEE/CVF Conference on Computer" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "Vision and Pattern Recognition (CVPR), pages 3164-3174, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[11] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 172 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NeurIPS), 33:6840-6851, 2020. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 174, + 553, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 553, + 236 + ], + "type": "text", + "content": "[13] Stephen James, Kentaro Wada, Tristan Laidlow, and Andrew J Davison. Coarse-to-fine q-attention: Efficient learning for visual robotic manipulation via discretisation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13739–13748, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[14] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 294, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 358 + ], + "type": "text", + "content": "[15] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4015-4026, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 361, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 426 + ], + "type": "text", + "content": "[16] Xiaoqi Li, Mingxu Zhang, Yiran Geng, Haoran Geng, Yuxing Long, Yan Shen, Renrui Zhang, Jiaming Liu, and Hao Dong. Manipllm: Embodied multimodal large language model for object-centric robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18061-18070, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 427, + 555, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 555, + 480 + ], + "type": "text", + "content": "[17] Haitao Lin, Yanwei Fu, and Xiangyang Xue. Pourit!: Weakly-supervised liquid perception from a single image for visual closed-loop robotic pouring. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 241-251, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 525 + ], + "type": "text", + "content": "[18] Songming Liu, Lingxuan Wu, Bangguo Li, Hengkai Tan, Huayu Chen, Zhengyi Wang, Ke Xu, Hang Su, and Jun Zhu. Rdt-1b: a diffusion foundation model for bimanual manipulation. arXiv preprint arXiv:2410.07864, 2024. 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 526, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 592 + ], + "type": "text", + "content": "[19] Yuecheng Liu, Dafeng Chi, Shiguang Wu, Zhanguang Zhang, Yaochen Hu, Lingfeng Zhang, Yingxue Zhang, Shuang Wu, Tongtong Cao, Guowei Huang, et al. Spatialcot: Advancing spatial reasoning through coordinate alignment and chain-of-thought for embodied task planning. arXiv preprint arXiv:2501.10074, 2025. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 593, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 646 + ], + "type": "text", + "content": "[20] Jingpei Lu, Florian Richter, and Michael C Yip. Markerless camera-to-robot pose estimation via self-supervised sim-to-real transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21296-21306, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "text", + "content": "[21] Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems (NeurIPS), 35: 2507-2521, 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 747, + 315, + 756 + ], + "type": "text", + "content": "6847" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[22] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[23] Xiao Ma, Sumit Patidar, Iain Haughton, and Stephen James. Hierarchical diffusion policy for kinematics-aware multitask robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18081-18090, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 250 + ], + "type": "text", + "content": "[24] Weixin Mao, Weiheng Zhong, Zhou Jiang, Dong Fang, Zhongyue Zhang, Zihan Lan, Fan Jia, Tiancai Wang, Haoqiang Fan, and Osamu Yoshie. Robomatrix: A skill-centric hierarchical framework for scalable robot task planning and execution in open-world. arXiv preprint arXiv:2412.00171, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 252, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 294, + 296 + ], + "type": "text", + "content": "[25] Suraj Nair, Aravind Rajeswaran, Vikash Kumar, Chelsea Finn, and Abhinav Gupta. R3m: A universal visual representation for robot manipulation. arXiv preprint arXiv:2203.12601, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[26] Max Polzin, Qinghua Guan, and Josie Hughes. Robotic locomotion through active and passive morphological adaptation in extreme outdoor environments. Science Robotics, 10(99): eadp6419, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 386 + ], + "type": "text", + "content": "[27] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108, 2019.5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "text", + "content": "[28] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su. Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2998-3009, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 499 + ], + "type": "text", + "content": "[29] Chan Hee Song, Valts Blukis, Jonathan Tremblay, Stephen Tyree, Yu Su, and Stan Birchfield. Robospatial: Teaching spatial understanding to 2d and 3d vision-language models for robotics. arXiv preprint arXiv:2411.16537, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 566 + ], + "type": "text", + "content": "[30] Enrica Tricomi, Francesco Missiroli, Michele Xiloyannis, Nicola Lotti, Xiaohui Zhang, Marios Stefanakis, Maximilian Theisen, Jürgen Bauer, Clemens Becker, and Lorenzo Massia. Soft robotic shorts improve outdoor walking efficiency in older adults. Nature Machine Intelligence, 6(10):1145-1155, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 568, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 632 + ], + "type": "text", + "content": "[31] Kasun Weerakoon, Adarsh Jagan Sathyamoorthy, Mohamed Elnoor, and Dinesh Manocha. Vapor: Legged robot navigation in unstructured outdoor environments using offline reinforcement learning. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 10344–10350. IEEE, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[32] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems (NeurIPS), 35:24824-24837, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[33] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 673 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Peng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[34] Zhiyu Wu, Xiaokang Chen, Zizheng Pan, Xingchao Liu, Wen Liu, Damai Dai, Huazuo Gao, Yiyang Ma, Chengyue Wu, Bingxuan Wang, et al. Deepseek-vl2: Mixture-of-experts vision-language models for advanced multimodal understanding. arXiv preprint arXiv:2412.10302, 2024. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 239 + ], + "type": "text", + "content": "[35] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4737-4746, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 240, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 316 + ], + "type": "text", + "content": "[36] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9556-9567, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 318, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 318, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 318, + 553, + 361 + ], + "type": "text", + "content": "[37] Michal Zawalski, William Chen, Karl Pertsch, Oier Mees, Chelsea Finn, and Sergey Levine. Robotic control via embodied chain-of-thought reasoning. arXiv preprint arXiv:2407.08693, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 363, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 553, + 407 + ], + "type": "text", + "content": "[38] Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xialin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with improved 3d diffusion policies. arXiv preprint arXiv:2410.10803, 2024. 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 408, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 553, + 452 + ], + "type": "text", + "content": "[39] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations. arXiv preprint arXiv:2403.03954, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 507 + ], + "type": "text", + "content": "[40] Andy Zeng, Maria Attarian, Brian Ichter, Krzysztof Choromanski, Adrian Wong, Stefan Welker, Federico Tombari, Aveek Purohit, Michael Ryoo, Vikas Sindhwani, et al. Socratic models: Composing zero-shot multimodal reasoning with language. arXiv preprint arXiv:2204.00598, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "text", + "content": "[41] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 553, + 553, + 606 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 553, + 553, + 606 + ], + "spans": [ + { + "bbox": [ + 316, + 553, + 553, + 606 + ], + "type": "text", + "content": "[42] Da Zhao, Haobo Luo, Yuxiao Tu, Chongxi Meng, and Tin Lun Lam. Snail-inspired robotic swarms: a hybrid connector drives collective adaptation in unstructured outdoor environments. Nature Communications, 15(1):3647, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 609, + 553, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 609, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 316, + 609, + 553, + 673 + ], + "type": "text", + "content": "[43] Allan Zhou, Moo Jin Kim, Lirui Wang, Pete Florence, and Chelsea Finn. Nerf in the palm of your hand: Corrective augmentation for robotics via novel-view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17907-17917, 2023. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 756 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 756 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 756 + ], + "type": "text", + "content": "6848" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/3caf6ed2-ef92-4e47-ab15-82d341d3a49f_content_list.json b/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/3caf6ed2-ef92-4e47-ab15-82d341d3a49f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e23fc9e3ba2d344997499fde4b3a85af57429d3a --- /dev/null +++ b/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/3caf6ed2-ef92-4e47-ab15-82d341d3a49f_content_list.json @@ -0,0 +1,1670 @@ +[ + { + "type": "text", + "text": "Towards More Diverse and Challenging Pre-training for Point Cloud Learning: Self-Supervised Cross Reconstruction with Decoupled Views", + "text_level": 1, + "bbox": [ + 99, + 128, + 898, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiangdong Zhang*, Shaofeng Zhang*, Junchi Yan‡", + "bbox": [ + 295, + 202, + 699, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "School of AI, Shanghai Jiao Tong University", + "bbox": [ + 320, + 222, + 674, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangxiangdong, sherrylone, yanjunchi}@sjtu.edu.cn", + "bbox": [ + 197, + 239, + 795, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point cloud learning, especially in a self-supervised way without manual labels, has gained growing attention in both vision and learning communities due to its potential utility in a wide range of applications. Most existing generative approaches for point cloud self-supervised learning focus on recovering masked points from visible ones within a single view. Recognizing that a two-view pre-training paradigm inherently introduces greater diversity and variance, it may thus enable more challenging and informative pre-training. Inspired by this, we explore the potential of two-view learning in this domain. In this paper, we propose Point-PQAE, a cross-reconstruction generative paradigm that first generates two decoupled point clouds/views and then reconstructs one from the other. To achieve this goal, we develop a crop mechanism for point cloud view generation for the first time and further propose a novel positional encoding to represent the 3D relative position between the two decoupled views. The cross-reconstruction significantly increases the difficulty of pre-training compared to self-reconstruction, which enables our method to surpass previous single-modal self-reconstruction methods in 3D self-supervised learning. Specifically, it outperforms the self-reconstruction baseline (Point-MAE) by $6.5\\%$ , $7.0\\%$ , and $6.7\\%$ in three variants of ScanObjectNN with the MLP-LINEAR evaluation protocol. The code is available at https://github.com/aHapBean/Point-PQAE.", + "bbox": [ + 89, + 323, + 485, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 750, + 220, + 766 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D vision is gaining increasing attention for its wide applications such as autonomous driving [38] and robotics [11, 40], owing to its ability to understand the human world. The point cloud is the most popular representation form of data in 3D vision and many analyses based on point cloud are", + "bbox": [ + 89, + 775, + 483, + 853 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71fc8b35306dce4462853f98b5b0cc64850f75583a1d84e87d9db6913c22d847.jpg", + "image_caption": [ + "Self-Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 535, + 292, + 687, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/12e4a292f90bab83a13a9f8aa0c2fe8a949e9c3cfd6e2f3a13bbaceb4845174a.jpg", + "image_caption": [ + "Cross-Reconstruction (Ours)", + "Figure 1. Comparison between self-reconstruction (Point-MAE [33] and other methods based on it) and cross-reconstruction (Ours) paradigms. In the self-reconstruction paradigm, part of the input data is masked, and the autoencoder is trained to recover the missing patches. In contrast, the cross-reconstruction paradigm generates decoupled views first, with one view leveraging cross-view information to reconstruct the other, using an autoencoder backbone." + ], + "image_footnote": [], + "bbox": [ + 717, + 292, + 880, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "explored to solve various tasks such as object classification [15, 24, 34, 35, 54, 62], object detection [30, 36, 76] and segmentation [15, 24, 34, 54, 58], they typically necessitate fully-supervised training from scratch. Compared to 2D data, point cloud generally requires more expensive and labor-intensive efforts to collect and annotate, which hinders the development of the fully supervised 3D representation learning method. Self-supervised learning is one of the predominant approaches to address this issue and has been proven to be effective in 2D vision [5, 8, 14, 18, 19, 68-71, 75]. Inspired by this, self-supervised learning has been widely studied in the 3D field in recent years.", + "bbox": [ + 509, + 580, + 906, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervised learning in 3D can be mainly divided into two categories, i.e., generative methods [1, 7, 23, 25, 33, 52, 66, 73] and contrastive methods [2, 9, 10, 20, 37, 39, 60]. Similar to 2D, contrastive methods in 3D aim to learn global-discriminative information by maximizing the mutual information across views [2, 9, 60] with intra-/inter-modal information [43-49, 56, 57]. PointContrast [60] is the first to learn transformation invariance across views through the extension of InfoNCE objective [31, 60]. Dif", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. ‡Corresponding author. This paper is in part supported by Shanghai Municipal Science and Technology Major Project, China, under grant No. 2021SHZDZX0102.", + "bbox": [ + 89, + 862, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28696", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ferent from contrastive methods, the core idea of the generative methods [25, 33, 66] in 3D is inspired by MAE [19], which masks the input point cloud and uses the visible part to reconstruct the masked part. Point-MAE [33] is one of the conventional methods, which learns point-wise information through the self-reconstruction mechanism. Specifically, Point-MAE masks a high ratio of the point cloud and reconstructs the masked points through the visible part. Point-M2AE [66] proposes a hierarchical Transformer [27] with a corresponding masking strategy based on it, capturing both fine-grained and high-level semantics of 3D shapes.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing generative methods mostly focus on the masked reconstruction in a single view, where single-view learning has been proven less difficult and informative than two views [8, 14] since two views would bring more variance than a single view. To bring the benefits of the two-view learning paradigm to the generative learning field, one straightforward assumption is adopting the two-view learning paradigm in the 3D generative pertaining tasks to increase the difficulty of point-wise reconstruction. However, unlike 2D vision data, constructing two views in the 3D point cloud is more challenging, as point cloud data is less structured. Besides, cross-reconstruction (reconstructing one view from the other view) is much more difficult than self-reconstruction (reconstructing the masked single view), which involves learning both the inter-positional relation between two views and the inner spatial information within the view. The differences between these two paradigms can be seen in Fig. 1. Aimed at solving the above issues, we propose Point-PQAE (Position Query based AutoEncoder for Point Cloud), a novel framework aimed at addressing two primary challenges: i) how to construct two views in point cloud data, and ii) how to perform cross-reconstruction between two views. As shown in Fig. 2, for each point cloud, we first apply a custom-designed point cloud crop mechanism, which randomly selects two points as the central points of two views and records their positions. Following this, for each central point, the nearest points according to a specified ratio are incorporated into the view. Subsequently, we normalize the cropped point cloud and apply random augmentation (e.g., rotation) to each point cloud to achieve view decoupling. Then we calculate the view-relative positional embedding (VRPE) for these two decoupled views. Finally, we take the VRPE and one view of the point cloud as input to predict the point-wise input of the other view.", + "bbox": [ + 91, + 275, + 483, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The highlights of this paper are:", + "text_level": 1, + "bbox": [ + 89, + 791, + 318, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "i) New generative framework: We propose PointPQAE, the first framework that successfully brings cross-reconstruction to 3D generative field for point cloud self-supervised learning, with three modules: 1) decoupled views generation, 2) VRPE generation, and 3) positional query block. To our knowledge, we are the first to design", + "bbox": [ + 89, + 810, + 480, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and apply crop mechanism to point cloud self-supervised learning. Our framework, Point-PQAE, enables more informative and challenging self-supervised pre-training compared to existing self-reconstruction methods.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ii) The first relative position-aware query module for point cloud: We introduce a positional query block after the encoder to capture relative position information between decoupled views. The module applies cross-attention between the hidden representations of one view and the VRPE, enabling it to predict the other view. Our module can be easily plugged into various related tasks, e.g., distillation as stated in Appendix Sec. 4.", + "bbox": [ + 511, + 152, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "iii) Strong performance: Point-PQAE achieves new state-of-the-art performance on several benchmarks, e.g., outperforming previous methods on few-shot learning and achieving average improvements of $1.8\\%$ , $6.7\\%$ , and $4.4\\%$ over the baseline (Point-MAE) on ScanObjectNN classification across three evaluation protocols (FULL, MLP-LINEAR, MLP-3), respectively.", + "bbox": [ + 511, + 272, + 903, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 511, + 392, + 658, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SSL in 2D vision. Self-supervised learning (SSL) in 2D vision is broadly categorized into contrastive-based and masked-image-modeling (MIM) approaches. Contrastive methods learn discriminative features by creating two augmented views of an image and enforcing invariance between them. For instance, SimCLR [8] treats augmented views from the same image as positive pairs and those from different images as negatives. MoCo [18] utilizes a memory bank to store these negative samples and further proposes a momentum-updated mechanism to avoid model collapse. MIM-based methods mask portions of the input and task models with reconstructing the masked regions, encouraging fine-grained understanding. MAE [19] employs an encoder-decoder architecture, discarding masked patches before reconstruction, while SimMIM [61] replaces masked regions with a learnable parameter. Subsequent work explores variations such as masking strategies [22, 72] and feature-level reconstruction [75].", + "bbox": [ + 511, + 416, + 903, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Representation learning for point clouds. It is a challenging task due to its irregular and sparse nature when compared to other modalities such as images which are well-structured and uniform. PointContrast [60] firstly borrows the idea of contrastive learning in 2D and performs point-level invariant mapping on two transformed views of the given point cloud. On the basis of PointContrast, Cross-Point [2] further proposes inter and intra-modal contrastive objectives to enhance the global-discriminative capability. To better model multi-modal data, ACT [9] employs cross-modal auto-encoders as teacher models to acquire knowledge from other modalities. In contrast to contrastive objectives, Point-MAE [33] proposes to randomly mask points, and reconstruct the masked part through the visible part.", + "bbox": [ + 511, + 688, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28697", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/204b1936a68fac0d3f61ec2f1b08677b5bd4ca3b8b7f2f670b75507f633993b7.jpg", + "image_caption": [ + "Figure 2. Pipeline of Point-PQAE. The input point cloud is randomly cropped followed by the rotation to generate views. Then, we feed the views to the patch embedding layer and the transformer encoder, followed by the proposed positional query block. The View-Relative Positional Embedding (VRPE) is obtained through the VRPE module by extracting relative geometric relations, which is taken as \"Query\" in the cross-attention mechanism. The queried hidden embeddings are then fed to the decoder to predict the inputs of the other view." + ], + "image_footnote": [], + "bbox": [ + 112, + 89, + 883, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following Point-MAE and CrossPoint, ReCon [37] harmoniously incorporates the generative method Point-MAE and the contrastive method. Almost all the existing generative methods in 3D rely on self-reconstruction. To this end, we propose a new cross-reconstruction generative framework, named Point-PQAE, which increases pre-training difficulty by introducing cross-view reconstruction, bringing more variance to the training samples compared to single-view reconstruction. This requires the model to effectively learn both intra-view and inter-view knowledge, thereby enabling the model to learn more semantic representations.", + "bbox": [ + 89, + 453, + 483, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The proposed Point-PQAE", + "text_level": 1, + "bbox": [ + 89, + 632, + 341, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Views and embeddings generation", + "text_level": 1, + "bbox": [ + 89, + 657, + 390, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Decoupled views generation. Random Image Crop has been widely used in both the supervised domain [17, 21, 42] and the self-supervised domain [4, 8, 18], and it has been proved to be effective in practice. This method typically involves randomly selecting a rectangular region from an image and cropping it. However, to the best of our knowledge, no corresponding crop mechanism has been proposed for 3D self-supervised representation learning. Instead of directly applying the 2D crop algorithm to 3D (by randomly selecting a cube in space), which may lead to inconsistency, as the points within cubes of the same size can vary significantly due to the diverse distribution of points in 3D, we design in a smarter way: For a randomly selected crop ratio, we first select a center point randomly and then choose the corresponding number of points in the point cloud that", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "are closest to the center point.", + "bbox": [ + 511, + 454, + 712, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Formally, given an input point cloud with $p$ points $\\mathbf{X} \\in \\mathbb{R}^{p \\times 3}$ and a pre-defined minimum crop ratio $r_m$ where $0 < r_m < 1$ , we first randomly select the crop ratios $r_1$ and $r_2$ uniformly from $[r_m, 1]$ . Then, we randomly select two center points $\\mathbf{C}_1, \\mathbf{C}_2 \\in \\mathbb{R}^{1 \\times 3}$ and get $\\mathbf{X}_1$ and $\\mathbf{X}_2$ by expanding nearest $r_1 \\times p$ and $r_2 \\times p$ points to the center point $\\mathbf{C}_1$ and $\\mathbf{C}_2$ respectively in $\\mathbf{X}$ . Meanwhile, we record the absolute coordinates of cropped point clouds' geometric centers, which are denoted as $\\mathbf{L}_1 \\in \\mathbb{R}^{1 \\times 3}$ and $\\mathbf{L}_2 \\in \\mathbb{R}^{1 \\times 3}$ , respectively. The process can be formulated as:", + "bbox": [ + 511, + 469, + 906, + 621 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nr _ {1}, r _ {2} = \\text {U n i f o r m S a m p l e} ([ r _ {m}, 1 ]) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 633, + 906, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {1}, \\mathbf {C} _ {2} = \\operatorname {R a n d o m S e l e c t} (\\mathbf {X}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 652, + 903, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {1} = \\mathrm {N} - \\text {N e a r e s t} (\\mathbf {C} _ {\\mathbf {1}}, r _ {1}, \\mathbf {X}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 671, + 903, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {2} = \\mathrm {N} - \\operatorname {N e a r e s t} \\left(\\mathbf {C} _ {2}, r _ {2}, \\mathbf {X}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 690, + 903, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_1 \\in \\mathbb{R}^{r_1p \\times 3}$ and $\\mathbf{X}_2 \\in \\mathbb{R}^{r_2p \\times 3}$ mean the two cropped views. The $\\mathbf{X}_1$ and $\\mathbf{X}_2$ are normalized using the min-max normalization, centered on $\\mathbf{L}_1$ and $\\mathbf{L}_2$ , respectively. Subsequently, we apply additional augmentation to further enhance the variance between two views. We apply random rotations to the cropped views. The distinct normalization centers, along with random rotation, effectively isolate the coordinate systems of the two views and alter their fixed relative relationships. This process decouples the two point clouds, resulting in two independent views.", + "bbox": [ + 511, + 718, + 906, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point patches generation. We divide the input point cloud into irregular point patches (may overlap) via the Far-", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28698", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "thest Point Sampling(FPS) and K-Nearest Neighborhood (KNN) algorithm following Point-BERT [65]. Formally, given two point clouds (views) $\\mathbf{X}_1$ and $\\mathbf{X}_2$ , we first use the FPS algorithm to sample pre-defined $n$ points as group centers, which we called $\\mathbf{G}_1 \\in \\mathbb{R}^{n \\times 3}$ and $\\mathbf{G}_2 \\in \\mathbb{R}^{n \\times 3}$ , respectively. Then, based on the center points, we use KNN to choose the pre-defined $k$ -nearest points for each group center $\\mathbf{G}^{(i)}$ ( $0 \\leq i \\leq n$ ) and finally get $n$ groups (each group contains $k$ points). We apply the FPS and KNN on $\\mathbf{X}_1$ and $\\mathbf{X}_2$ , respectively:", + "bbox": [ + 89, + 90, + 483, + 243 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {1} = \\operatorname {F P S} \\left(\\mathbf {X} _ {1}\\right), \\quad \\mathbf {G} _ {2} = \\operatorname {F P S} \\left(\\mathbf {X} _ {2}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 253, + 480, + 270 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {1} = \\operatorname {K N N} \\left(\\mathbf {X} _ {1}, \\mathbf {G} _ {1}\\right), \\quad \\mathbf {P} _ {2} = \\operatorname {K N N} \\left(\\mathbf {X} _ {2}, \\mathbf {G} _ {2}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 273, + 482, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{G}_1, \\mathbf{G}_2 \\in \\mathbb{R}^{n \\times 3}$ and $\\mathbf{P}_1, \\mathbf{P}_2 \\in \\mathbb{R}^{n \\times (k \\cdot 3)}$ . Then, the $\\mathbf{P}_1$ and $\\mathbf{P}_2$ both are treated as a sequence of length $n$ and dimension $k \\cdot 3$ as the input.", + "bbox": [ + 89, + 303, + 483, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Backbone with positional query", + "text_level": 1, + "bbox": [ + 89, + 354, + 370, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following [33], we adopt a lightweight PointNet [34], followed by an asymmetric encoder-decoder structure consisting of Transformer blocks in the Point-PQAE. After the encoder projects the input embedding into the latent space, we use the latent representation of one view and view-relative positional embedding (VRPE) between two views to implement the position query through cross-attention [51]. We directly feed the output of the position query block to the decoder to reconstruct another view without using extra positional embedding. The last layer of the model adopts a simple prediction head to achieve the cross-reconstruction.", + "bbox": [ + 89, + 373, + 483, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Encoding. The encoder consists of the lightweight PointNet and standard Transformer blocks [51], which projects $\\mathbf{P}_1$ and $\\mathbf{P}_2$ to latent space, respectively:", + "bbox": [ + 89, + 541, + 483, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {1} = \\operatorname {E n c o d e r} \\left(\\mathbf {P} _ {1}, \\mathbf {G} _ {1}\\right), \\quad \\mathbf {H} _ {2} = \\operatorname {E n c o d e r} \\left(\\mathbf {P} _ {2}, \\mathbf {G} _ {2}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 598, + 482, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{H}_1\\in \\mathbb{R}^{n\\times D}$ and $\\mathbf{H}_2\\in \\mathbb{R}^{n\\times D}$ mean the latent representations of the view 1 and view 2. Note that $D$ is the hidden dimension of the networks.", + "bbox": [ + 89, + 627, + 482, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Positional query block. After obtaining latent representations $\\mathbf{H}_1$ and $\\mathbf{H}_2$ , we feed them into the proposed positional query block to extract cross-information. The module can be mainly divided into 3 sub-stages: attain relative information, VRPE generation, and cross-attention.", + "bbox": [ + 89, + 674, + 483, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Attain relative information. Take reconstructing view 2 from view 1 as an example. To reconstruct view 2, we have to know the latent representations of view 1 and the relative positions of view 2 centered on view 1, where the latent representations of view 1 can be obtained through the encoder network. Therefore, the key to reconstructing view 2 is to obtain the relative positions of view 2 centered on view 1. Back to the crop operation, we have recorded the geometric centers of views, denoted as $\\mathbf{L}_1$ and $\\mathbf{L}_2$ , respectively. We define the relative position of view 2 centered on view 1", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as $\\mathbf{RL}_{1\\rightarrow 2} = \\mathbf{L}_1 - \\mathbf{L}_2$ where $\\mathbf{RL}_{1\\rightarrow 2}\\in \\mathbb{R}^{1\\times 3}$ . To better align the dimension in the latter operation, we expand the $\\mathbf{RL}_{1\\rightarrow 2}$ to $\\mathbb{R}^{n\\times 3}$ by repeating the content for $n$ times. By obtaining the relative positions between the two views, we further add the group center location to help the PQ modules reconstruct each group of the inputs. Specifically, we define the group-/patch- wise relative positions as:", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R P} _ {1 \\rightarrow 2} = \\operatorname {C o n C a t} \\left(\\mathbf {G} _ {2}, \\mathbf {R L} _ {1 \\rightarrow 2}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 205, + 903, + 220 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R P} _ {2 \\rightarrow 1} = \\operatorname {C o n C a t} \\left(\\mathbf {G} _ {1}, \\mathbf {R L} _ {2 \\rightarrow 1}\\right) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 224, + 903, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{RP} \\in \\mathbb{R}^{n \\times 6}$ (as $\\mathbf{G} \\in \\mathbb{R}^{n \\times 3}$ and $\\mathbf{RL} \\in \\mathbb{R}^{n \\times 3}$ ) and $\\operatorname{ConCat}(\\cdot, \\cdot)$ is the concatenate operation.", + "bbox": [ + 511, + 246, + 903, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "View-relative positional embedding (VRPE). Instead of simply using learnable positional encoding, which could hurt the expression of the relative positions between the two views (see ablation in Sec. 4.3), we use the fixed positional encoding of relative information to cut the uncertainty. Specifically, for elements in the second dimension of $\\mathbf{RP}_{1\\rightarrow 2}$ , we use a sinusoid form [19] in Transformers to generate view-relative positional embedding $\\mathbf{VRPE}_{1\\rightarrow 2}$ :", + "bbox": [ + 511, + 277, + 905, + 398 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {V R P E} _ {1 \\rightarrow 2} ^ {i} = \\left[ \\sin \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e ^ {2 \\times \\frac {1}{D / 1 2}}}\\right), \\cos \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e ^ {2 \\times \\frac {1}{D / 1 2}}}\\right), \\right. \\\\ \\sin \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e ^ {2 \\times \\frac {2}{D / 1 2}}}\\right), \\cos \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e ^ {2 \\times \\frac {2}{D / 1 2}}}\\right), \\ldots , \\\\ \\left. \\sin \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e}\\right), \\cos \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e}\\right)\\right] \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 406, + 903, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{RP}_{1\\rightarrow 2}^i\\in \\mathbb{R}^{n\\times 1}$ , $\\mathbf{VRPE}_{1\\rightarrow 2}^i\\in \\mathbb{R}^{n\\times (D / 6)}$ ( $1\\leq i\\leq 6$ ). $e = 10000$ is pre-defined, which is also used in MAE [19]. Finally, we concatenate $\\mathbf{VRPE}_{1\\rightarrow 2}^i$ for each dimension to obtain $\\mathbf{VRPE}_{1\\rightarrow 2}\\in \\mathbb{R}^{n\\times D}$ .", + "bbox": [ + 511, + 521, + 905, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-attention. Empirically, assume the latent representations of view 1, denoted as $\\mathbf{H}_1$ , contains intrinsic features and the global characteristics of its source point cloud, then with the combination of the view-relative positional embeddings $\\mathbf{VRPE}_{1\\rightarrow 2}$ , the view 2 can be reconstructed. Specifically, we employ $\\mathbf{VRPE}_{1\\rightarrow 2}$ (as $\\mathbf{Q}$ ) and $\\mathbf{H}_1$ (as $\\mathbf{K},\\mathbf{V}$ ) in cross-attention mechanisms, as can be formalized:", + "bbox": [ + 511, + 584, + 905, + 688 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {T} _ {2} = \\operatorname {A t t n} \\left(\\mathbf {Q} _ {\\mathbf {V R P E} _ {1 \\rightarrow 2}}, \\mathbf {K} _ {\\mathbf {H} _ {1}}, \\mathbf {V} _ {\\mathbf {H} _ {1}}\\right) \\\\ = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {\\mathbf {V R P E} _ {1 \\rightarrow 2}} \\mathbf {K} _ {\\mathbf {H} _ {1}}}{\\sqrt {D}}\\right) \\mathbf {V} _ {\\mathbf {H} _ {1}} \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 696, + 903, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}_{\\mathbf{VRPE}_{1\\rightarrow 2}} = \\mathbf{VRPE}_{1\\rightarrow 2}\\mathbf{W}_{\\mathbf{Q}},\\mathbf{K}_{\\mathbf{H}_1} = \\mathbf{H}_1\\mathbf{W}_{\\mathbf{K}}$ and $\\mathbf{V_{H_1}} = \\mathbf{H_1W_V}$ , where $\\mathbf{W}_{\\mathbf{Q}},\\mathbf{W}_{\\mathbf{K}}$ , and $\\mathbf{W_V}$ are learnable parameters. Reconstructing view 1 from view 2 is a siamese condition of this, through the same way as stated above, we can obtain $\\mathbf{RL}_{2\\to 1},\\mathbf{RP}_{2\\to 1},\\mathbf{VRPE}_{2\\to 1}$ and yield $\\mathbf{T}_1$", + "bbox": [ + 511, + 755, + 905, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoding. The obtained $\\mathbf{T}_1$ and $\\mathbf{T}_2$ are fed into a decoder composed of a few transformer blocks:", + "bbox": [ + 511, + 847, + 903, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {1} = \\operatorname {D e c o d e r} \\left(\\mathbf {T} _ {1}\\right), \\quad \\mathbf {Z} _ {2} = \\operatorname {D e c o d e r} \\left(\\mathbf {T} _ {2}\\right) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 883, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28699", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, similar to [33], we adopt a projection head (composed of a fully connected layer) to predict the input point cloud as follows, where $\\mathbf{P}_{pred}^{1},\\mathbf{P}_{pred}^{2}\\in \\mathbb{R}^{n\\times k\\times 3}$ :", + "bbox": [ + 89, + 90, + 483, + 140 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {p r e d} ^ {1} = \\operatorname {R e s h a p e} \\left(\\operatorname {L i n e a r} \\left(\\mathbf {Z} _ {1}\\right)\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 150, + 482, + 169 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {p r e d} ^ {2} = \\operatorname {R e s h a p e} \\left(\\operatorname {L i n e a r} \\left(\\mathbf {Z} _ {2}\\right)\\right) \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 171, + 482, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Objective function", + "text_level": 1, + "bbox": [ + 89, + 203, + 269, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall objective is to perform cross-reconstruction, which involves predicting view 1 from view 2 and predicting view 2 from view 1. Given the predicted point patches $\\mathbf{P}_{pred}^{1}$ and $\\mathbf{P}_{pred}^{2}$ and ground truth $\\mathbf{P}_1$ and $\\mathbf{P}_2$ , we compute the cross-reconstruction loss using the $l_{2}$ Chamfer Distance loss function [12], written as $\\mathcal{L}_{cross} = \\mathcal{L}_{2\\rightarrow 1} + \\mathcal{L}_{1\\rightarrow 2}$ , where:", + "bbox": [ + 89, + 220, + 483, + 325 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {2 \\rightarrow 1} = \\frac {1}{\\left| \\mathbf {P} _ {p r e d} ^ {1} \\right|} \\sum_ {a \\in \\mathbf {P} _ {p r e d} ^ {1}} \\min _ {b \\in \\mathbf {P} _ {1}} \\| a - b \\| _ {2} ^ {2} \\tag {15} \\\\ + \\frac {1}{| \\mathbf {P} _ {1} |} \\sum_ {b \\in \\mathbf {P} _ {1}} \\min _ {a \\in \\mathbf {P} _ {p r e d} ^ {1}} \\| a - b \\| _ {2} ^ {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 334, + 482, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|\\mathbf{P}|$ is the cardinality of the set $\\mathbf{P}$ and $\\| \\cdot \\| _2$ is the $l_{2}$ distance. The $\\mathcal{L}_{1\\to 2}$ follows similarly.", + "bbox": [ + 89, + 426, + 483, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 472, + 223, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section is organized as follows: First, we present the model architecture and pre-training details. Second, we validate the effectiveness of our pre-trained model on a wide range of downstream tasks, including object classification, few-shot learning, and part segmentation. Finally, ablation studies are conducted to demonstrate the properties and robustness of our proposed Point-PQAE.", + "bbox": [ + 89, + 497, + 483, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Pre-training setups", + "text_level": 1, + "bbox": [ + 89, + 608, + 274, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-trained dataset. We use the dataset ShapeNet [6] for pre-training of Point-PQAE following previous studies [9, 33, 37]. ShapeNet [6] consists of about 51,300 clean 3D models, covering 55 common object categories.", + "bbox": [ + 89, + 627, + 483, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model structure. We adopt standard Transformer blocks [51] in the autoencoder's backbone where the encoder has 12 Transformer blocks and the decoder has 4 blocks. MLP ratio in Transformer blocks is set to 4. Each Transformer block has 384 hidden dimensions and 6 heads. The positional query module between encoder and decoder only performs cross-attention mechanism once as stated in Sec. 3.2, and the number of heads is set to 6 and remains 384 hidden dimensions.", + "bbox": [ + 89, + 689, + 483, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experiment detail. For each input point cloud, we only apply random rotation for pre-training data augmentation. After sampling 1024 points via FPS from the input point cloud, we generate two different partial point clouds by randomly cropping and rotating as illustrated in Sec. 3.1 with", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$r_m = 0.6$ . Both of them are divided into 64 patches with 32 points via FPS and KNN. The Point-PQAE model undergoes pre-training for 300 epochs using an AdamW [29] optimizer with a batch size of 128. The initial learning rate is set to 0.0005, with a weight decay of 0.05, with cosine learning rate decay [28] utilized. We visualize the pretraining results in Fig. 3. It shows Point-PQAE learns cross-knowledge well and is able to generalize excellently to other crop ratios, even though it was pre-trained with $r_m = 0.6$ .", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Transfer learning on downstream tasks", + "text_level": 1, + "bbox": [ + 511, + 232, + 849, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Transfer protocol. We use three transfer learning protocols for classification tasks following [9]: FULL, MLP-LINEAR, MLP-3, which are detailed in Appendix Sec. 3.", + "bbox": [ + 511, + 251, + 903, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D real-world object classification. The transferability of models can be well demonstrated through testing on a 3D real-world object dataset. Therefore, we transfer our pre-trained model to ScanObjectNN [50] for classification, which is one of the most challenging 3D datasets, covering approximately 15,000 real-world objects across 15 categories. We conduct experiments on three variants: OBJBG, OBJ-ONLY, and PB-T50-RS. The results are reported in Tab. 1. Our Point-PQAE surpasses previous single-modal self-supervised methods by a margin under all protocols. In contrast to the most relevant baseline, Point-MAE, PointPQAE significantly outperforms it, showing improvements of $2.4\\%$ , $1.7\\%$ , and $1.2\\%$ . Especially for the MLP-LINEAR and MLP-3 protocols, our Point-PQAE outperforms Point-MAE by $6.7\\%$ and $4.4\\%$ on average, respectively. The excellent frozen representation evaluation results indicate the superior quality of the representation learned by our Point-PQAE during pre-training. This shows that the cross-reconstruction paradigm learns more robust representations and generalizes better on real-world datasets.", + "bbox": [ + 511, + 296, + 905, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D clean object classification. We evaluate the performance of our pre-trained model for object classification using the ModelNet40 dataset [59]. ModelNet40 comprises 12,311 meticulously crafted 3D CAD models, representing 40 distinct object categories. During testing, we adhere to the standard voting method [26] to ensure fair comparisons with previous work such as [7, 33]. Standard random scaling and random translation are applied for data augmentation during training, and the input point clouds exclusively contain coordinate information, without additional normal information provided. The results are presented in Tab. 1. Our Point-PQAE yields better or comparable results compared to previous approaches.", + "bbox": [ + 511, + 598, + 905, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Few-shot learning. We conduct the few-shot learning classification on the ModelNet40 dataset to show the generalization ability of our method following the protocols in previous studies [33, 41, 53, 65]. The few-shot learning experiments are conducted in the form of \"w-way, s-shot\" including four parts where $w \\in \\{5, 10\\}$ represents the number of randomly selected classes and $s \\in \\{10, 20\\}$ means", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28700", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a3026c65407e601c8dec37467fd5b25a1e53e55810f8d88097109312c5e0de7b.jpg", + "image_caption": [ + "Cropped Rebuilt" + ], + "image_footnote": [], + "bbox": [ + 174, + 127, + 282, + 209 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b25778fcd134e75327060f3ee0fd455e866fc2a7f49f0db55068625858d11c11.jpg", + "image_caption": [ + "Input point cloud" + ], + "image_footnote": [], + "bbox": [ + 187, + 237, + 261, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1bb13258c46e6dfc24d340715ba35b1aae8dd1c3bb9c63fea0947121aa56e438.jpg", + "image_caption": [ + "Cropped Rebuilt" + ], + "image_footnote": [], + "bbox": [ + 303, + 126, + 421, + 205 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d82c0c2b96671c70322a59c7b27114f86c1f4a94eb7eae12cdf22d26cf35b299.jpg", + "image_caption": [ + "Crop ratio $20\\%$" + ], + "image_footnote": [], + "bbox": [ + 302, + 222, + 419, + 301 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a4e01672b06f5245dac48c0c7739a732b62c245b571f607a57703c76ff8d3769.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 125, + 555, + 202 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85c35790e2ae81bb1eb84686abd7d9d42a107580d4bef8973f3d8248dbbf8352.jpg", + "image_caption": [ + "Crop ratio $40\\%$" + ], + "image_footnote": [], + "bbox": [ + 434, + 220, + 549, + 308 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/784a93b547adc88ec643569db3f3db5486a1caf47ca964c47af5eb353c504ad8.jpg", + "image_caption": [ + "Cropped Rebuilt" + ], + "image_footnote": [], + "bbox": [ + 566, + 119, + 679, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b006035c5d1b4896f6c459d0193bb6dd1978934f7ef42d8bbce7da017cf7b932.jpg", + "image_caption": [ + "Crop ratio $60\\%$" + ], + "image_footnote": [], + "bbox": [ + 568, + 224, + 684, + 306 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/15e48a1586f403deefc6e9bfa332e7318339ab3c963de707e25df079dd0a5f7a.jpg", + "image_caption": [ + "Cropped Rebuilt" + ], + "image_footnote": [], + "bbox": [ + 696, + 119, + 810, + 203 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ccbff0f4bfb44b246bc01219a203df5c5d0be55cb500dc6a2ed895b8a9f959c.jpg", + "image_caption": [ + "Crop ratio $80\\%$", + "Figure 3. Cross-reconstruction results on ShapeNet. The arrow points from source point clouds to cross-reconstruction results. PointPQAE generalizes well to other crop ratios though with minimum crop ratio $r_m = 0.6$ when pre-training." + ], + "image_footnote": [], + "bbox": [ + 694, + 223, + 808, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the number of randomly selected samples for each $w$ . Each part is conducted with 10 independent trials. The mean accuracy and standard deviation are reported in Tab. 2. Our Point-PQAE demonstrates superior performance compared to previous methods in few-shot learning, even surpassing cross-modal methods that utilize strong pre-trained teachers such as ACT and ReCon under FULL protocol. Specifically, Point-PQAE achieves $2.7\\%$ , $1.1\\%$ , $4.5\\%$ , $1.9\\%$ improvement under MLP-LINEAR protocol and shows enhanced performance over the self-construction method Point-MAE.", + "bbox": [ + 89, + 388, + 483, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Part segmentation. We conduct part segmentation experiments on the ShapeNetPart [64] to validate the effectiveness of our Point-PQAE. The ShapeNetPart contains 16,881 objects covering 16 categories. We sample 2,048 points from each input point cloud, following previous work [33, 65] and divide each cloud into 128 point patches. We use the same segmentation head and utilize learned features from the 4th, 8th, and 12th layers of the Transformer block as in Point-MAE [33]. We concatenate three levels of features. Then average pooling, max pooling, and upsampling are utilized to generate features for each point and an MLP is applied for label prediction. The results are reported in Tab. 3a. The results are reported in Tab. 3a, which show that Point-PQAE achieves a comparable Inst.mIoU to the previous method, achieving $84.6\\%$ in Cls.mIoU and improving the from-scratch baseline by $1.4\\%$ .", + "bbox": [ + 89, + 547, + 483, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3D scene segmentation. Semantic segmentation on large-scale 3D scenes is challenging, requiring models possessing a deep comprehension of semantics and intricate local geometric relationships. We report semantic segmentation results on the S3DIS dataset [3] in Tab. 3b. Our Point-PQAE improves the from-scratch baseline by $2.0\\%$ and $1.4\\%$ , and outperforms the self-reconstruction method", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Point-MAE by $0.7\\%$ and $0.6\\%$ in mAcc and mIoU.", + "bbox": [ + 511, + 388, + 852, + 404 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 412, + 661, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments are conducted to show the properties of our Point-PQAE. We report the classification results on three variants of ScanObjectNN. The experiment settings are aligned with Sec. 4.2.", + "bbox": [ + 511, + 430, + 905, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "View-relative positional embedding. To analyze the effect of view-relative positional embedding in the positional query block, we conduct a series of experiments using different types of positional embeddings. We primarily consider two types of embeddings: Sinusoid (as discussed in Sec. 3.2) and Learnable, along with a None group. Learnable embedding refers to utilizing the relative position information $\\mathbf{RP}_{1\\rightarrow 2}$ and $\\mathbf{RP}_{2\\rightarrow 1}$ as input and employing an MLP composed of two linear layers with an activation function to map the 6-dimensional input to a D-dimensional output. The None group involves using randomly assigned embeddings to assess the effectiveness of VRPE design. Tab. 4a displays the results of different positional embeddings. In the None group, the model cannot learn the relative positional information across two views, leading to a significant drop in accuracy. In addition to the type of VRPE, we also experiment with Absolute Positional Embedding (APE) to compare its performance with our proposed VRPE. The APE is unsuitable for our framework since it requires a shared coordinate system for both views, making it incompatible with the normalization and rotation used in our method. We incorporate APE into our Point-PQAE by removing these operations. The results in Tab. 4a show APE performs worse and support our claim.", + "bbox": [ + 511, + 491, + 905, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data augmentation. After applying the random crop mechanism to the point cloud, we further perform data augmentation on the cropped views to generate new views.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28701", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a08e614a5cdef566bcdffd36016810f1b1387158e3af069af3a27ce27f652e3c.jpg", + "table_caption": [ + "Table 1. Classification accuracy (%) on ScanObjectNN and ModelNet40. The inference model parameters #P (M) are reported. Three variants are evaluated on ScanObjectNN and the accuracy obtained on ModelNet40 is reported for both 1K and 8K points. We compare methods using the $\\bullet$ plain Transformer architectures, e.g. Point-MAE [33], $\\circ$ hierarchical Transformer architectures and $\\circ$ dedicated architectures for 3D. The dagger(†) denotes the baseline results reported from ReCon [37] which aligns augmentation with us and ReCon." + ], + "table_footnote": [], + "table_body": "
Methods#PScanObjectNNModelNet40
OBJ_BGOBJ_ONLYPB_T50_RS1K P8K P
Supervised Learning Only
○PointNet [34]3.573.379.268.089.290.8
○PointNet++ [35]1.582.384.377.990.791.9
○DGCNN [54]1.882.886.278.192.9-
○PointCNN [24]0.686.185.578.592.2-
○SimpleView [13]---80.5±0.393.9-
○PCT [15]2.88---93.2-
with Self-Supervised Representation Learning (FULL)
•Transformer [51]22.183.084.079.191.491.8
•Point-BERT [65]22.187.488.183.093.293.8
•MaskPoint [25]-89.388.184.393.8-
•Point-MAE [33]22.190.088.385.293.894.0
○Point-M2AE [66]15.391.288.986.494.0-
•PointDif [74]22.193.391.987.6--
•Point-MAE† [33]22.192.691.988.493.894.0
•Point-PQAE22.195.093.689.694.094.3
Methods using cross-modal information and teacher models
•Joint-MAE [16]-90.988.986.194.0-
•TAP [55]22.190.489.585.794.0-
•ACT [9]22.193.391.988.293.794.0
○PointMLP+ULIP [63]---89.494.594.7
○I2P-MAE [67]15.394.291.690.194.1
•ReCon [37]43.695.293.690.694.594.7
with Self-Supervised Representation Learning (MLP-LINEAR)
•Point-MAE† [33]22.182.8±0.383.2±0.274.1±0.290.2±0.190.7±0.1
•Point-PQAE22.189.3±0.390.2±0.480.8±0.192.0±0.292.2±0.1
Methods using cross-modal information and teacher models
•ACT [9]22.185.2±0.885.8±0.276.3±0.391.4±0.291.8±0.2
•ReCon [37]43.689.5±0.289.7±0.281.4±0.192.5±0.292.7±0.1
with Self-Supervised Representation Learning (MLP-3)
•Point-MAE† [33]22.185.8±0.385.5±0.280.4±0.291.3±0.291.7±0.2
•Point-PQAE22.190.7±0.290.9±0.283.3±0.192.8±0.192.9±0.1
Methods using cross-modal information and teacher models
•ACT [9]22.187.1±0.287.9±0.481.5±0.292.7±0.293.0±0.1
•ReCon [37]43.690.6±0.290.7±0.383.8±0.493.0±0.193.4±0.1
", + "bbox": [ + 212, + 148, + 777, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data augmentation is crucial for generating diverse and decoupled views, so we conducted ablation studies on different augmentation methods. The results in Tab. 4b indicate that rotation performs best.", + "bbox": [ + 89, + 698, + 482, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the crop mechanism. Random crop has been shown to be highly effective in 2D self-supervised learning (SSL), particularly for contrastive learning [4, 8, 14], and is also vital for generative learning [19]. However, its potential in 3D SSL remains largely unexplored. Even contrastive learning methods for point cloud understanding have not explored the power of it [2, 32]. To address this gap, we first introduce a custom-designed random crop mechanism for point cloud data, using it as the", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "foundation for generating decoupled views. Without the crop augmentation, the two views will contain exactly the same shape though with normalization and rotation applied. The model would only need to infer the augmentation rather than the inter-view relationship, which means the views are not decoupled. Conversely, given two isolated cropped point clouds, followed by min-max normalization and rotation for further decoupling, there remain overlapping and non-overlapping parts, requiring the model to effectively encode the intra-view parts in order to infer the unknown inter-view points in our Point-PQAE. However, directly applying random crop as an additional augmentation to the self-reconstruction method (Point-MAE) does", + "bbox": [ + 511, + 698, + 906, + 895 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28702", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d7ff8aa1387991c682c961031975d3a1549eb7a9ecd1e1d4911e6bff7f13da8b.jpg", + "table_caption": [ + "Table 2. Few-shot classification results on ModelNet40. 10 independent trials are conducted in each experimental setting. The mean accuracy (\\%) and standard deviation are reported for each setting. The dagger( $\\dagger$ ) denotes the baseline results reported from ReCon [37] which aligns augmentation with us and ReCon." + ], + "table_footnote": [], + "table_body": "
Methods5-way10-way
10-shot20-shot10-shot20-shot
○DGCNN [54]31.6±2.840.8±4.619.9±2.116.9±1.5
○OcCo [52]90.6±2.892.5±1.982.9±1.386.5±2.2
with Self-Supervised Representation Learning (FULL)
•Transformer [51]87.8±5.293.3±4.384.6±5.589.4±6.3
•Point-BERT [65]94.6±3.196.3±2.791.0±5.492.7±5.1
•MaskPoint [25]95.0±3.797.2±1.791.4±4.093.4±3.5
•Point-MAE [33]96.3±2.597.8±1.892.6±4.195.0±3.0
○Point-M2AE [66]96.8±1.898.3±1.492.3±4.595.0±3.0
•Point-MAE† [66]96.4±2.897.8±2.092.5±4.495.2±3.9
•Point-PQAE96.9±3.298.9±1.094.1±4.296.3±2.7
Methods using cross-modal information and teacher models
•Joint-MAE [16]96.7±2.297.9±1.992.6±3.795.1±2.6
•TAP [55]97.3±1.897.8±1.993.1±2.695.8±1.0
•ACT [9]96.8±2.398.0±1.493.3±4.095.6±2.8
○I2P-MAE [67]97.0±1.898.3±1.392.6±5.095.5±3.0
•ReCon [37]97.3±1.998.9±1.293.3±3.995.8±3.0
with Self-Supervised Representation Learning (MLP-LINEAR)
•Point-MAE† [66]91.1±5.691.7±4.083.5±6.189.7±4.1
•Point-PQAE93.0±4.696.8±1.989.0±5.293.5±3.8
with Self-Supervised Representation Learning (MLP-3)
•Point-MAE† [66]95.0±2.896.7±2.490.6±4.793.8±5.0
•Point-PQAE95.3±3.498.2±1.892.0±3.894.7±3.5
", + "bbox": [ + 109, + 164, + 459, + 470 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/60f7cbfa772efafa25b00249761674892cfde2aaf64ec05b4cdd657872a49d66.jpg", + "table_caption": [ + "Table 3. Segmentation results. Cls.mIoU (\\%) and Inst.mIoU (\\%) refer to Mean intersection over union for all classes and all instances, respectively. mAcc (\\%) refers to mean accuracy.", + "(a) Part segmentation on ShapeNetPart." + ], + "table_footnote": [], + "table_body": "
MethodCls.mIoU (%)Inst.mIoU (%)
Supervised
PointNet [34]80.483.7
PointNet++ [35]81.985.1
DGCNN [54]82.385.2
Self-supervised
Transformer [51]83.484.7
CrossPoint [2]-85.5
Point-BERT [65]84.185.6
Point-MAE [33]84.286.1
Point-PQAE (Ours)84.686.1
Cross-modal
ACT [9]84.786.1
ReCon [37]84.886.4
", + "bbox": [ + 93, + 554, + 282, + 659 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e50debff01ce93a64b42ea1d9f295bded65e3673722e2c0e2cbdb5d0e38b3cfe.jpg", + "table_caption": [ + "(b) Semantic segmentation results on S3DIS Area 5." + ], + "table_footnote": [], + "table_body": "
MethodmAcc (%)mIoU (%)
Supervised
PointNet [34]49.041.1
PointNet++ [35]67.153.5
Self-supervised
Transformer [51]68.660.0
Point-MAE [33]69.960.8
Point-PQAE (Ours)70.661.4
Cross-modal
ACT [9]71.161.2
", + "bbox": [ + 295, + 555, + 483, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "not improve the performance much, as shown in Tab. 4c. In other words, it means our Point-PQAE fits crop mechanism better by harmoniously incorporating it into the decoupled views generation process. The experiment results in Tab. 4c demonstrate the critical role of random crop in our method. We set minimum crop ratio $r_m = 1.0$ when removing the crop in our Point-PQAE.", + "bbox": [ + 89, + 671, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on decoupled views generation. While crop plays an important role in our method, relying solely on the random crop mechanism to generate new views would degenerate cross-reconstruction to simply reconstructing one part from another. This is similar to the block mask self-reconstruction used in Point-MAE [33], where visible blocks are used to reconstruct masked blocks, which is also trivial for pre-training. The only difference between them", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d77a8f6fa3b9ebf8f18edf63ed6ce000e553dbd800041b96b05348d03d16bca3.jpg", + "table_caption": [ + "Table 4. Ablation study with Point-PQAE pre-training on ShapeNet. The classification results by accuracy (\\%) on three variants of ScanObjectNN are reported. Default settings are marked in gray.", + "(a) View-Relative positional embedding" + ], + "table_footnote": [], + "table_body": "
Positional EmbeddingOBJ_BGOBJ_ONLYLYPB_T50_RS
None84.585.979.3
APE (sinusoid)92.391.087.7
VRPE (learnable)93.493.189.1
VRPE (sinusoid Eq. (10))95.093.689.6
", + "bbox": [ + 568, + 161, + 854, + 217 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cff9b9ba6bdfb25d5f145124cccce63ab8884ac377f662b03cdd006b8ec7a6f1.jpg", + "table_caption": [ + "(b) Data augmentation" + ], + "table_footnote": [], + "table_body": "
Data augmentationOBJ_BGOBJ_ONLYPB_T50_RS
jitter93.391.487.3
scale93.391.788.0
rotation95.093.689.6
scale&translate92.891.788.1
rotation+scale&translate93.892.989.1
", + "bbox": [ + 568, + 232, + 852, + 299 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b151369a2f039289a3b5542ac2346c6ab97036424b8cccba172278d497fd0180.jpg", + "table_caption": [ + "(c) 3D random crop mechanism" + ], + "table_footnote": [], + "table_body": "
MethodcropOBJ_BGOBJ_ONLYPB_T50_RS
Point-MAE✘92.691.988.4
Point-MAE✓92.992.188.8
Point-PQAE✘92.992.387.7
Point-PQAE✓95.093.689.6
", + "bbox": [ + 568, + 316, + 852, + 375 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/67a6975751f5524eb4b5a34e025ad64b415be1098eeb53c67761b4e1c22d3958.jpg", + "table_caption": [ + "(d) Effects of augmentations after random crop" + ], + "table_footnote": [], + "table_body": "
norm.rotationOBJ_BGOBJ_ONLYPB_T50_RS
XX92.892.188.2
✓X93.392.687.9
X✓93.592.688.8
✓✓95.093.689.6
", + "bbox": [ + 568, + 392, + 848, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is that the cropped parts can overlap somewhat, whereas the block mask does not. What makes the cross-reconstruction different from block mask self-reconstruction are the subsequent augmentations including normalization and rotation that decouple the cropped parts. By normalizing cropped point clouds centered on the geometric centers, the coordinate systems of the two point clouds become isolated and independent from each other, with rotation further amplifying the variance between the two views, thus enabling the generation of two decoupled views. The importance of the augmentations applied after the random crop can be validated through experiments, as shown in Tab. 4d.", + "bbox": [ + 511, + 464, + 906, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 652, + 633, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel cross-reconstruction generative framework for self-supervised learning on 3D point clouds, called Point-PQAE. In contrast to well-studied self-reconstruction schemes, Point-PQAE reconstructs one cropped point cloud from another decoupled point cloud. To supply sufficient information for cross-reconstruction, we further propose a 3D view-relative positional embedding and a corresponding position-aware query module. Compared to self-reconstruction, the cross-reconstruction task, carefully designed by us, is much more challenging for pre-training, promoting the learning of richer semantic representations during pre-training. This enables our proposed Point-PQAE to outperform previous single-modal self-reconstruction methods by a margin and to perform on par with, or better than cross-modal methods.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28703", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 1", + "[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 1, 2, 7, 8", + "[3] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 6", + "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3, 7", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1", + "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5", + "[7] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. Advances in Neural Information Processing Systems, 36, 2024. 1, 5", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 2, 3, 7", + "[9] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022. 1, 2, 5, 7, 8", + "[10] Bi'an Du, Xiang Gao, Wei Hu, and Xin Li. Self-contrastive learning with hard negative sampling for self-supervised point cloud learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3133–3142, 2021. 1", + "[11] Ikenna Enebuse, Mathias Foo, Babul Salam Ksm Kader Ibrahim, Hafiz Ahmed, Fhon Supmak, and Odongo Steven Eyobu. A comparative review of hand-eye calibration techniques for vision guided robots. IEEE Access, 9:113143-113155, 2021. 1", + "[12] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 5" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In International Conference on Machine Learning, pages 3809-3820. PMLR, 2021. 7", + "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 2, 7", + "[15] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 1, 7", + "[16] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzhi Li, and Pheng-Ann Heng. Joint-mae: 2d-3d joint masked autoencoders for 3d point cloud pre-training. arXiv preprint arXiv:2302.14007, 2023. 7, 8", + "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 3", + "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 2, 4, 7", + "[20] Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6535-6545, 2021. 1", + "[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 3", + "[22] Gang Li, Heliang Zheng, Daqing Liu, Chaoyue Wang, Bing Su, and Changwen Zheng. Semmae: Semantic-guided masking for learning masked autoencoders. NeurIPS, 2022. 2", + "[23] Jiaxin Li, Ben M Chen, and Gim Hee Lee. So-net: Self-organizing network for point cloud analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9397–9406, 2018. 1", + "[24] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 1, 7", + "[25] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 1, 2, 7, 8", + "[26] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28704", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "cloud analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8895-8904, 2019. 5", + "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2", + "[28] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5", + "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[30] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 1", + "[31] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1", + "[32] Bo Pang, Hongchi Xia, and Cewu Lu. Unsupervised 3d point cloud representation learning by triangle constrained contrast for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5229-5239, 2023. 7", + "[33] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 1, 2, 4, 5, 6, 7, 8", + "[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 4, 7, 8", + "[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 7, 8", + "[36] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1", + "[37] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. arXiv preprint arXiv:2302.02318, 2023. 1, 3, 5, 7, 8", + "[38] Rui Qian, Xin Lai, and Xirong Li. 3d object detection for autonomous driving: A survey. Pattern Recognition, 130: 108796, 2022. 1", + "[39] Aditya Sanghi. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX 16, pages 626-642. Springer, 2020. 1", + "[40] Oleg Yu Sergiyenko and Vera V Tyrsa. 3d optical machine vision sensors with intelligent data management for robotic" + ], + "bbox": [ + 91, + 90, + 483, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "swarm navigation improvement. IEEE Sensors Journal, 21 (10):11262-11274, 2020. 1", + "[41] Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. Advances in Neural Information Processing Systems, 33:7212-7221, 2020. 5", + "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3", + "[43] Haoru Tan, Chuang Wang, Sitong Wu, Tieqiang Wang, Xu-Yao Zhang, and Cheng-Lin Liu. Proxy graph matching with proximal matching networks. In The Annual AAAI Conference on Artificial Intelligence (AAAI), 2021. 1", + "[44] Haoru Tan, Sitong Wu, Fei Du, Yukang Chen, Zhibin Wang, Fan Wang, and Xiaojuan Qi. Data pruning via moving-one-sample-out. In Neural Information Processing Systems (NeurIPS), 2023.", + "[45] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2023.", + "[46] Haoru Tan, Chuang Wang, Sitong Wu, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Ensemble quadratic assignment network for graph matching. International Journal of Computer Vision (IJCV), 2024.", + "[47] Haoru Tan, Sitong Wu, Zhuotao Tian, Yukang Chen, Xiaojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity consistency for vision-language pre-training. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.", + "[48] Haoru Tan, Sitong Wu, Wei Huang, Shizhen Zhao, and Xiaojuan Qi. Data pruning by information maximization. In International Conference on Learning Representations (ICLR), 2025.", + "[49] Haoru Tan, Sitong Wu, Bo Zhao, Zeke Xie, and XIAOJUAN QI. Diff-in: Data influence estimation with differential approximation, 2025. 1", + "[50] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 5", + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5, 7, 8", + "[52] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 1, 8", + "[53] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 5", + "[54] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic" + ], + "bbox": [ + 516, + 92, + 903, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28705", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 1, 7, 8", + "[55] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. Take-a-photo: 3d-to-2d generative pre-training of point cloud models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5640–5650, 2023. 7, 8", + "[56] Hai Wu, Ruifei He, Haoru Tan, Xiaojuan Qi, and Kaibin Huang. Vertical layering of quantized neural networks for heterogeneous inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[57] Sitong Wu, Haoru Tan, Yukang Chen, Shaofeng Zhang, Jingyao Li, Bei Yu, Xiaojuan Qi, and Jiaya Jia. Mixture-of-scores: Robust image-text data quality score via three lines of code. In International Conference on Computer Vision (ICCV), 2025. 1", + "[58] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. arXiv preprint arXiv:2312.10035, 2023. 1", + "[59] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 5", + "[60] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 1, 2", + "[61] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2", + "[62] Mingye Xu, Zhipeng Zhou, Yali Wang, and Yu Qiao. Towards robustness and generalization of point cloud representation: A geometry coding method and a large-scale object-level dataset. Computational Visual Media, 10(1):27-43, 2024. 1", + "[63] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1179-1189, 2023. 7", + "[64] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6", + "[65] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 4, 5, 6, 7, 8", + "[66] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae:" + ], + "bbox": [ + 91, + 92, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 1, 2, 7, 8", + "[67] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 7, 8", + "[68] Shaofeng Zhang, Qiang Zhou, Sitong Wu, Haoru Tan, Zhibin Wang, Jina Huang, and Junchi Yan. Cr2pq: Continuous relative rotary positional query for dense visual representation learning. In The Thirteenth International Conference on Learning Representations. 1", + "[69] Shaofeng Zhang, Meng Liu, and Junchi Yan. The diversified ensemble neural network. Advances in Neural Information Processing Systems, 33:16001-16011, 2020.", + "[70] Shaofeng Zhang, Feng Zhu, Junchi Yan, Rui Zhao, and Xiaokang Yang. Zero-cl: Instance and feature decorrelation for negative-free symmetric contrastive learning. In International Conference on Learning Representations, 2021.", + "[71] Shaofeng Zhang, Meng Liu, Junchi Yan, Hengrui Zhang, Lingxiao Huang, Xiaokang Yang, and Pinyin Lu. M-mix: Generating hard negatives via multi-sample mixing for contrastive learning. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, pages 2461-2470, 2022. 1", + "[72] Shaofeng Zhang, Feng Zhu, Rui Zhao, and Junchi Yan. Contextual image masking modeling via synergized contrasting without view augmentation for faster and better visual pretraining. In ICLR, 2023. 2", + "[73] Xiangdong Zhang, Shaofeng Zhang, and Junchi Yan. Pcmae: Learning to predict centers for point masked autoencoders. Advances in Neural Information Processing Systems, 37:80303-80327, 2024. 1", + "[74] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22935-22945, 2024. 7", + "[75] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2", + "[76] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1" + ], + "bbox": [ + 516, + 92, + 903, + 755 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28706", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/3caf6ed2-ef92-4e47-ab15-82d341d3a49f_model.json b/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/3caf6ed2-ef92-4e47-ab15-82d341d3a49f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..12b19c1c5af204ca41fe5d5bdea7412cc9dd305b --- /dev/null +++ b/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/3caf6ed2-ef92-4e47-ab15-82d341d3a49f_model.json @@ -0,0 +1,2620 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.13, + 0.9, + 0.177 + ], + "angle": 0, + "content": "Towards More Diverse and Challenging Pre-training for Point Cloud Learning: Self-Supervised Cross Reconstruction with Decoupled Views" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.203, + 0.7, + 0.222 + ], + "angle": 0, + "content": "Xiangdong Zhang*, Shaofeng Zhang*, Junchi Yan‡" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.223, + 0.676, + 0.24 + ], + "angle": 0, + "content": "School of AI, Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.24, + 0.796, + 0.258 + ], + "angle": 0, + "content": "{zhangxiangdong, sherrylone, yanjunchi}@sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.718 + ], + "angle": 0, + "content": "Point cloud learning, especially in a self-supervised way without manual labels, has gained growing attention in both vision and learning communities due to its potential utility in a wide range of applications. Most existing generative approaches for point cloud self-supervised learning focus on recovering masked points from visible ones within a single view. Recognizing that a two-view pre-training paradigm inherently introduces greater diversity and variance, it may thus enable more challenging and informative pre-training. Inspired by this, we explore the potential of two-view learning in this domain. In this paper, we propose Point-PQAE, a cross-reconstruction generative paradigm that first generates two decoupled point clouds/views and then reconstructs one from the other. To achieve this goal, we develop a crop mechanism for point cloud view generation for the first time and further propose a novel positional encoding to represent the 3D relative position between the two decoupled views. The cross-reconstruction significantly increases the difficulty of pre-training compared to self-reconstruction, which enables our method to surpass previous single-modal self-reconstruction methods in 3D self-supervised learning. Specifically, it outperforms the self-reconstruction baseline (Point-MAE) by \\(6.5\\%\\), \\(7.0\\%\\), and \\(6.7\\%\\) in three variants of ScanObjectNN with the MLP-LINEAR evaluation protocol. The code is available at https://github.com/aHapBean/Point-PQAE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.75, + 0.222, + 0.767 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.776, + 0.484, + 0.854 + ], + "angle": 0, + "content": "3D vision is gaining increasing attention for its wide applications such as autonomous driving [38] and robotics [11, 40], owing to its ability to understand the human world. The point cloud is the most popular representation form of data in 3D vision and many analyses based on point cloud are" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.294, + 0.689, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.445, + 0.669, + 0.455 + ], + "angle": 0, + "content": "Self-Reconstruction" + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.293, + 0.882, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.445, + 0.868, + 0.456 + ], + "angle": 0, + "content": "Cross-Reconstruction (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.46, + 0.908, + 0.571 + ], + "angle": 0, + "content": "Figure 1. Comparison between self-reconstruction (Point-MAE [33] and other methods based on it) and cross-reconstruction (Ours) paradigms. In the self-reconstruction paradigm, part of the input data is masked, and the autoencoder is trained to recover the missing patches. In contrast, the cross-reconstruction paradigm generates decoupled views first, with one view leveraging cross-view information to reconstruct the other, using an autoencoder backbone." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.581, + 0.907, + 0.763 + ], + "angle": 0, + "content": "explored to solve various tasks such as object classification [15, 24, 34, 35, 54, 62], object detection [30, 36, 76] and segmentation [15, 24, 34, 54, 58], they typically necessitate fully-supervised training from scratch. Compared to 2D data, point cloud generally requires more expensive and labor-intensive efforts to collect and annotate, which hinders the development of the fully supervised 3D representation learning method. Self-supervised learning is one of the predominant approaches to address this issue and has been proven to be effective in 2D vision [5, 8, 14, 18, 19, 68-71, 75]. Inspired by this, self-supervised learning has been widely studied in the 3D field in recent years." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Self-supervised learning in 3D can be mainly divided into two categories, i.e., generative methods [1, 7, 23, 25, 33, 52, 66, 73] and contrastive methods [2, 9, 10, 20, 37, 39, 60]. Similar to 2D, contrastive methods in 3D aim to learn global-discriminative information by maximizing the mutual information across views [2, 9, 60] with intra-/inter-modal information [43-49, 56, 57]. PointContrast [60] is the first to learn transformation invariance across views through the extension of InfoNCE objective [31, 60]. Dif" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.484, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. ‡Corresponding author. This paper is in part supported by Shanghai Municipal Science and Technology Major Project, China, under grant No. 2021SHZDZX0102." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28696" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "ferent from contrastive methods, the core idea of the generative methods [25, 33, 66] in 3D is inspired by MAE [19], which masks the input point cloud and uses the visible part to reconstruct the masked part. Point-MAE [33] is one of the conventional methods, which learns point-wise information through the self-reconstruction mechanism. Specifically, Point-MAE masks a high ratio of the point cloud and reconstructs the masked points through the visible part. Point-M2AE [66] proposes a hierarchical Transformer [27] with a corresponding masking strategy based on it, capturing both fine-grained and high-level semantics of 3D shapes." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.276, + 0.485, + 0.79 + ], + "angle": 0, + "content": "Existing generative methods mostly focus on the masked reconstruction in a single view, where single-view learning has been proven less difficult and informative than two views [8, 14] since two views would bring more variance than a single view. To bring the benefits of the two-view learning paradigm to the generative learning field, one straightforward assumption is adopting the two-view learning paradigm in the 3D generative pertaining tasks to increase the difficulty of point-wise reconstruction. However, unlike 2D vision data, constructing two views in the 3D point cloud is more challenging, as point cloud data is less structured. Besides, cross-reconstruction (reconstructing one view from the other view) is much more difficult than self-reconstruction (reconstructing the masked single view), which involves learning both the inter-positional relation between two views and the inner spatial information within the view. The differences between these two paradigms can be seen in Fig. 1. Aimed at solving the above issues, we propose Point-PQAE (Position Query based AutoEncoder for Point Cloud), a novel framework aimed at addressing two primary challenges: i) how to construct two views in point cloud data, and ii) how to perform cross-reconstruction between two views. As shown in Fig. 2, for each point cloud, we first apply a custom-designed point cloud crop mechanism, which randomly selects two points as the central points of two views and records their positions. Following this, for each central point, the nearest points according to a specified ratio are incorporated into the view. Subsequently, we normalize the cropped point cloud and apply random augmentation (e.g., rotation) to each point cloud to achieve view decoupling. Then we calculate the view-relative positional embedding (VRPE) for these two decoupled views. Finally, we take the VRPE and one view of the point cloud as input to predict the point-wise input of the other view." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.32, + 0.807 + ], + "angle": 0, + "content": "The highlights of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.482, + 0.902 + ], + "angle": 0, + "content": "i) New generative framework: We propose PointPQAE, the first framework that successfully brings cross-reconstruction to 3D generative field for point cloud self-supervised learning, with three modules: 1) decoupled views generation, 2) VRPE generation, and 3) positional query block. To our knowledge, we are the first to design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "and apply crop mechanism to point cloud self-supervised learning. Our framework, Point-PQAE, enables more informative and challenging self-supervised pre-training compared to existing self-reconstruction methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.272 + ], + "angle": 0, + "content": "ii) The first relative position-aware query module for point cloud: We introduce a positional query block after the encoder to capture relative position information between decoupled views. The module applies cross-attention between the hidden representations of one view and the VRPE, enabling it to predict the other view. Our module can be easily plugged into various related tasks, e.g., distillation as stated in Appendix Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.379 + ], + "angle": 0, + "content": "iii) Strong performance: Point-PQAE achieves new state-of-the-art performance on several benchmarks, e.g., outperforming previous methods on few-shot learning and achieving average improvements of \\(1.8\\%\\), \\(6.7\\%\\), and \\(4.4\\%\\) over the baseline (Point-MAE) on ScanObjectNN classification across three evaluation protocols (FULL, MLP-LINEAR, MLP-3), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.393, + 0.659, + 0.408 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.905, + 0.689 + ], + "angle": 0, + "content": "SSL in 2D vision. Self-supervised learning (SSL) in 2D vision is broadly categorized into contrastive-based and masked-image-modeling (MIM) approaches. Contrastive methods learn discriminative features by creating two augmented views of an image and enforcing invariance between them. For instance, SimCLR [8] treats augmented views from the same image as positive pairs and those from different images as negatives. MoCo [18] utilizes a memory bank to store these negative samples and further proposes a momentum-updated mechanism to avoid model collapse. MIM-based methods mask portions of the input and task models with reconstructing the masked regions, encouraging fine-grained understanding. MAE [19] employs an encoder-decoder architecture, discarding masked patches before reconstruction, while SimMIM [61] replaces masked regions with a learnable parameter. Subsequent work explores variations such as masking strategies [22, 72] and feature-level reconstruction [75]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Representation learning for point clouds. It is a challenging task due to its irregular and sparse nature when compared to other modalities such as images which are well-structured and uniform. PointContrast [60] firstly borrows the idea of contrastive learning in 2D and performs point-level invariant mapping on two transformed views of the given point cloud. On the basis of PointContrast, Cross-Point [2] further proposes inter and intra-modal contrastive objectives to enhance the global-discriminative capability. To better model multi-modal data, ACT [9] employs cross-modal auto-encoders as teacher models to acquire knowledge from other modalities. In contrast to contrastive objectives, Point-MAE [33] proposes to randomly mask points, and reconstruct the masked part through the visible part." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28697" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.09, + 0.885, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.391, + 0.908, + 0.448 + ], + "angle": 0, + "content": "Figure 2. Pipeline of Point-PQAE. The input point cloud is randomly cropped followed by the rotation to generate views. Then, we feed the views to the patch embedding layer and the transformer encoder, followed by the proposed positional query block. The View-Relative Positional Embedding (VRPE) is obtained through the VRPE module by extracting relative geometric relations, which is taken as \"Query\" in the cross-attention mechanism. The queried hidden embeddings are then fed to the decoder to predict the inputs of the other view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.484, + 0.622 + ], + "angle": 0, + "content": "Following Point-MAE and CrossPoint, ReCon [37] harmoniously incorporates the generative method Point-MAE and the contrastive method. Almost all the existing generative methods in 3D rely on self-reconstruction. To this end, we propose a new cross-reconstruction generative framework, named Point-PQAE, which increases pre-training difficulty by introducing cross-view reconstruction, bringing more variance to the training samples compared to single-view reconstruction. This requires the model to effectively learn both intra-view and inter-view knowledge, thereby enabling the model to learn more semantic representations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.342, + 0.65 + ], + "angle": 0, + "content": "3. The proposed Point-PQAE" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.391, + 0.673 + ], + "angle": 0, + "content": "3.1. Views and embeddings generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Decoupled views generation. Random Image Crop has been widely used in both the supervised domain [17, 21, 42] and the self-supervised domain [4, 8, 18], and it has been proved to be effective in practice. This method typically involves randomly selecting a rectangular region from an image and cropping it. However, to the best of our knowledge, no corresponding crop mechanism has been proposed for 3D self-supervised representation learning. Instead of directly applying the 2D crop algorithm to 3D (by randomly selecting a cube in space), which may lead to inconsistency, as the points within cubes of the same size can vary significantly due to the diverse distribution of points in 3D, we design in a smarter way: For a randomly selected crop ratio, we first select a center point randomly and then choose the corresponding number of points in the point cloud that" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.455, + 0.714, + 0.469 + ], + "angle": 0, + "content": "are closest to the center point." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.907, + 0.622 + ], + "angle": 0, + "content": "Formally, given an input point cloud with \\( p \\) points \\( \\mathbf{X} \\in \\mathbb{R}^{p \\times 3} \\) and a pre-defined minimum crop ratio \\( r_m \\) where \\( 0 < r_m < 1 \\), we first randomly select the crop ratios \\( r_1 \\) and \\( r_2 \\) uniformly from \\( [r_m, 1] \\). Then, we randomly select two center points \\( \\mathbf{C}_1, \\mathbf{C}_2 \\in \\mathbb{R}^{1 \\times 3} \\) and get \\( \\mathbf{X}_1 \\) and \\( \\mathbf{X}_2 \\) by expanding nearest \\( r_1 \\times p \\) and \\( r_2 \\times p \\) points to the center point \\( \\mathbf{C}_1 \\) and \\( \\mathbf{C}_2 \\) respectively in \\( \\mathbf{X} \\). Meanwhile, we record the absolute coordinates of cropped point clouds' geometric centers, which are denoted as \\( \\mathbf{L}_1 \\in \\mathbb{R}^{1 \\times 3} \\) and \\( \\mathbf{L}_2 \\in \\mathbb{R}^{1 \\times 3} \\), respectively. The process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.634, + 0.907, + 0.651 + ], + "angle": 0, + "content": "\\[\nr _ {1}, r _ {2} = \\text {U n i f o r m S a m p l e} ([ r _ {m}, 1 ]) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.653, + 0.905, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {1}, \\mathbf {C} _ {2} = \\operatorname {R a n d o m S e l e c t} (\\mathbf {X}) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.672, + 0.905, + 0.688 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {1} = \\mathrm {N} - \\text {N e a r e s t} (\\mathbf {C} _ {\\mathbf {1}}, r _ {1}, \\mathbf {X}) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.691, + 0.905, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {2} = \\mathrm {N} - \\operatorname {N e a r e s t} \\left(\\mathbf {C} _ {2}, r _ {2}, \\mathbf {X}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.907, + 0.87 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_1 \\in \\mathbb{R}^{r_1p \\times 3}\\) and \\(\\mathbf{X}_2 \\in \\mathbb{R}^{r_2p \\times 3}\\) mean the two cropped views. The \\(\\mathbf{X}_1\\) and \\(\\mathbf{X}_2\\) are normalized using the min-max normalization, centered on \\(\\mathbf{L}_1\\) and \\(\\mathbf{L}_2\\), respectively. Subsequently, we apply additional augmentation to further enhance the variance between two views. We apply random rotations to the cropped views. The distinct normalization centers, along with random rotation, effectively isolate the coordinate systems of the two views and alter their fixed relative relationships. This process decouples the two point clouds, resulting in two independent views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Point patches generation. We divide the input point cloud into irregular point patches (may overlap) via the Far-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "28698" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.244 + ], + "angle": 0, + "content": "thest Point Sampling(FPS) and K-Nearest Neighborhood (KNN) algorithm following Point-BERT [65]. Formally, given two point clouds (views) \\(\\mathbf{X}_1\\) and \\(\\mathbf{X}_2\\), we first use the FPS algorithm to sample pre-defined \\(n\\) points as group centers, which we called \\(\\mathbf{G}_1 \\in \\mathbb{R}^{n \\times 3}\\) and \\(\\mathbf{G}_2 \\in \\mathbb{R}^{n \\times 3}\\), respectively. Then, based on the center points, we use KNN to choose the pre-defined \\(k\\)-nearest points for each group center \\(\\mathbf{G}^{(i)}\\) (\\(0 \\leq i \\leq n\\)) and finally get \\(n\\) groups (each group contains \\(k\\) points). We apply the FPS and KNN on \\(\\mathbf{X}_1\\) and \\(\\mathbf{X}_2\\), respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.255, + 0.482, + 0.271 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {1} = \\operatorname {F P S} \\left(\\mathbf {X} _ {1}\\right), \\quad \\mathbf {G} _ {2} = \\operatorname {F P S} \\left(\\mathbf {X} _ {2}\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.274, + 0.483, + 0.291 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {1} = \\operatorname {K N N} \\left(\\mathbf {X} _ {1}, \\mathbf {G} _ {1}\\right), \\quad \\mathbf {P} _ {2} = \\operatorname {K N N} \\left(\\mathbf {X} _ {2}, \\mathbf {G} _ {2}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.35 + ], + "angle": 0, + "content": "where \\(\\mathbf{G}_1, \\mathbf{G}_2 \\in \\mathbb{R}^{n \\times 3}\\) and \\(\\mathbf{P}_1, \\mathbf{P}_2 \\in \\mathbb{R}^{n \\times (k \\cdot 3)}\\). Then, the \\(\\mathbf{P}_1\\) and \\(\\mathbf{P}_2\\) both are treated as a sequence of length \\(n\\) and dimension \\(k \\cdot 3\\) as the input." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.355, + 0.371, + 0.372 + ], + "angle": 0, + "content": "3.2. Backbone with positional query" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.484, + 0.541 + ], + "angle": 0, + "content": "Following [33], we adopt a lightweight PointNet [34], followed by an asymmetric encoder-decoder structure consisting of Transformer blocks in the Point-PQAE. After the encoder projects the input embedding into the latent space, we use the latent representation of one view and view-relative positional embedding (VRPE) between two views to implement the position query through cross-attention [51]. We directly feed the output of the position query block to the decoder to reconstruct another view without using extra positional embedding. The last layer of the model adopts a simple prediction head to achieve the cross-reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.587 + ], + "angle": 0, + "content": "Encoding. The encoder consists of the lightweight PointNet and standard Transformer blocks [51], which projects \\(\\mathbf{P}_1\\) and \\(\\mathbf{P}_2\\) to latent space, respectively:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.599, + 0.483, + 0.616 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {1} = \\operatorname {E n c o d e r} \\left(\\mathbf {P} _ {1}, \\mathbf {G} _ {1}\\right), \\quad \\mathbf {H} _ {2} = \\operatorname {E n c o d e r} \\left(\\mathbf {P} _ {2}, \\mathbf {G} _ {2}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.673 + ], + "angle": 0, + "content": "where \\(\\mathbf{H}_1\\in \\mathbb{R}^{n\\times D}\\) and \\(\\mathbf{H}_2\\in \\mathbb{R}^{n\\times D}\\) mean the latent representations of the view 1 and view 2. Note that \\(D\\) is the hidden dimension of the networks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.749 + ], + "angle": 0, + "content": "Positional query block. After obtaining latent representations \\(\\mathbf{H}_1\\) and \\(\\mathbf{H}_2\\), we feed them into the proposed positional query block to extract cross-information. The module can be mainly divided into 3 sub-stages: attain relative information, VRPE generation, and cross-attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Attain relative information. Take reconstructing view 2 from view 1 as an example. To reconstruct view 2, we have to know the latent representations of view 1 and the relative positions of view 2 centered on view 1, where the latent representations of view 1 can be obtained through the encoder network. Therefore, the key to reconstructing view 2 is to obtain the relative positions of view 2 centered on view 1. Back to the crop operation, we have recorded the geometric centers of views, denoted as \\(\\mathbf{L}_1\\) and \\(\\mathbf{L}_2\\), respectively. We define the relative position of view 2 centered on view 1" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "as \\(\\mathbf{RL}_{1\\rightarrow 2} = \\mathbf{L}_1 - \\mathbf{L}_2\\) where \\(\\mathbf{RL}_{1\\rightarrow 2}\\in \\mathbb{R}^{1\\times 3}\\). To better align the dimension in the latter operation, we expand the \\(\\mathbf{RL}_{1\\rightarrow 2}\\) to \\(\\mathbb{R}^{n\\times 3}\\) by repeating the content for \\(n\\) times. By obtaining the relative positions between the two views, we further add the group center location to help the PQ modules reconstruct each group of the inputs. Specifically, we define the group-/patch- wise relative positions as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.206, + 0.905, + 0.221 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R P} _ {1 \\rightarrow 2} = \\operatorname {C o n C a t} \\left(\\mathbf {G} _ {2}, \\mathbf {R L} _ {1 \\rightarrow 2}\\right), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.225, + 0.905, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R P} _ {2 \\rightarrow 1} = \\operatorname {C o n C a t} \\left(\\mathbf {G} _ {1}, \\mathbf {R L} _ {2 \\rightarrow 1}\\right) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.247, + 0.905, + 0.278 + ], + "angle": 0, + "content": "where \\(\\mathbf{RP} \\in \\mathbb{R}^{n \\times 6}\\) (as \\(\\mathbf{G} \\in \\mathbb{R}^{n \\times 3}\\) and \\(\\mathbf{RL} \\in \\mathbb{R}^{n \\times 3}\\)) and \\(\\operatorname{ConCat}(\\cdot, \\cdot)\\) is the concatenate operation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.906, + 0.4 + ], + "angle": 0, + "content": "View-relative positional embedding (VRPE). Instead of simply using learnable positional encoding, which could hurt the expression of the relative positions between the two views (see ablation in Sec. 4.3), we use the fixed positional encoding of relative information to cut the uncertainty. Specifically, for elements in the second dimension of \\(\\mathbf{RP}_{1\\rightarrow 2}\\), we use a sinusoid form [19] in Transformers to generate view-relative positional embedding \\(\\mathbf{VRPE}_{1\\rightarrow 2}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.407, + 0.905, + 0.516 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {V R P E} _ {1 \\rightarrow 2} ^ {i} = \\left[ \\sin \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e ^ {2 \\times \\frac {1}{D / 1 2}}}\\right), \\cos \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e ^ {2 \\times \\frac {1}{D / 1 2}}}\\right), \\right. \\\\ \\sin \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e ^ {2 \\times \\frac {2}{D / 1 2}}}\\right), \\cos \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e ^ {2 \\times \\frac {2}{D / 1 2}}}\\right), \\ldots , \\\\ \\left. \\sin \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e}\\right), \\cos \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e}\\right)\\right] \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.522, + 0.906, + 0.584 + ], + "angle": 0, + "content": "where \\(\\mathbf{RP}_{1\\rightarrow 2}^i\\in \\mathbb{R}^{n\\times 1}\\), \\(\\mathbf{VRPE}_{1\\rightarrow 2}^i\\in \\mathbb{R}^{n\\times (D / 6)}\\) (\\(1\\leq i\\leq 6\\)). \\(e = 10000\\) is pre-defined, which is also used in MAE [19]. Finally, we concatenate \\(\\mathbf{VRPE}_{1\\rightarrow 2}^i\\) for each dimension to obtain \\(\\mathbf{VRPE}_{1\\rightarrow 2}\\in \\mathbb{R}^{n\\times D}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Cross-attention. Empirically, assume the latent representations of view 1, denoted as \\(\\mathbf{H}_1\\), contains intrinsic features and the global characteristics of its source point cloud, then with the combination of the view-relative positional embeddings \\(\\mathbf{VRPE}_{1\\rightarrow 2}\\), the view 2 can be reconstructed. Specifically, we employ \\(\\mathbf{VRPE}_{1\\rightarrow 2}\\) (as \\(\\mathbf{Q}\\)) and \\(\\mathbf{H}_1\\) (as \\(\\mathbf{K},\\mathbf{V}\\)) in cross-attention mechanisms, as can be formalized:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.697, + 0.905, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {T} _ {2} = \\operatorname {A t t n} \\left(\\mathbf {Q} _ {\\mathbf {V R P E} _ {1 \\rightarrow 2}}, \\mathbf {K} _ {\\mathbf {H} _ {1}}, \\mathbf {V} _ {\\mathbf {H} _ {1}}\\right) \\\\ = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {\\mathbf {V R P E} _ {1 \\rightarrow 2}} \\mathbf {K} _ {\\mathbf {H} _ {1}}}{\\sqrt {D}}\\right) \\mathbf {V} _ {\\mathbf {H} _ {1}} \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.756, + 0.906, + 0.847 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}_{\\mathbf{VRPE}_{1\\rightarrow 2}} = \\mathbf{VRPE}_{1\\rightarrow 2}\\mathbf{W}_{\\mathbf{Q}},\\mathbf{K}_{\\mathbf{H}_1} = \\mathbf{H}_1\\mathbf{W}_{\\mathbf{K}}\\) and \\(\\mathbf{V_{H_1}} = \\mathbf{H_1W_V}\\) , where \\(\\mathbf{W}_{\\mathbf{Q}},\\mathbf{W}_{\\mathbf{K}}\\) , and \\(\\mathbf{W_V}\\) are learnable parameters. Reconstructing view 1 from view 2 is a siamese condition of this, through the same way as stated above, we can obtain \\(\\mathbf{RL}_{2\\to 1},\\mathbf{RP}_{2\\to 1},\\mathbf{VRPE}_{2\\to 1}\\) and yield \\(\\mathbf{T}_1\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.848, + 0.905, + 0.877 + ], + "angle": 0, + "content": "Decoding. The obtained \\(\\mathbf{T}_1\\) and \\(\\mathbf{T}_2\\) are fed into a decoder composed of a few transformer blocks:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {1} = \\operatorname {D e c o d e r} \\left(\\mathbf {T} _ {1}\\right), \\quad \\mathbf {Z} _ {2} = \\operatorname {D e c o d e r} \\left(\\mathbf {T} _ {2}\\right) \\tag {12}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28699" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.141 + ], + "angle": 0, + "content": "Finally, similar to [33], we adopt a projection head (composed of a fully connected layer) to predict the input point cloud as follows, where \\(\\mathbf{P}_{pred}^{1},\\mathbf{P}_{pred}^{2}\\in \\mathbb{R}^{n\\times k\\times 3}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.151, + 0.483, + 0.17 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {p r e d} ^ {1} = \\operatorname {R e s h a p e} \\left(\\operatorname {L i n e a r} \\left(\\mathbf {Z} _ {1}\\right)\\right), \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.172, + 0.483, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {p r e d} ^ {2} = \\operatorname {R e s h a p e} \\left(\\operatorname {L i n e a r} \\left(\\mathbf {Z} _ {2}\\right)\\right) \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.204, + 0.271, + 0.219 + ], + "angle": 0, + "content": "3.3. Objective function" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.484, + 0.327 + ], + "angle": 0, + "content": "The overall objective is to perform cross-reconstruction, which involves predicting view 1 from view 2 and predicting view 2 from view 1. Given the predicted point patches \\(\\mathbf{P}_{pred}^{1}\\) and \\(\\mathbf{P}_{pred}^{2}\\) and ground truth \\(\\mathbf{P}_1\\) and \\(\\mathbf{P}_2\\), we compute the cross-reconstruction loss using the \\(l_{2}\\) Chamfer Distance loss function [12], written as \\(\\mathcal{L}_{cross} = \\mathcal{L}_{2\\rightarrow 1} + \\mathcal{L}_{1\\rightarrow 2}\\), where:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.335, + 0.483, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {2 \\rightarrow 1} = \\frac {1}{\\left| \\mathbf {P} _ {p r e d} ^ {1} \\right|} \\sum_ {a \\in \\mathbf {P} _ {p r e d} ^ {1}} \\min _ {b \\in \\mathbf {P} _ {1}} \\| a - b \\| _ {2} ^ {2} \\tag {15} \\\\ + \\frac {1}{| \\mathbf {P} _ {1} |} \\sum_ {b \\in \\mathbf {P} _ {1}} \\min _ {a \\in \\mathbf {P} _ {p r e d} ^ {1}} \\| a - b \\| _ {2} ^ {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.459 + ], + "angle": 0, + "content": "where \\(|\\mathbf{P}|\\) is the cardinality of the set \\(\\mathbf{P}\\) and \\(\\| \\cdot \\| _2\\) is the \\(l_{2}\\) distance. The \\(\\mathcal{L}_{1\\to 2}\\) follows similarly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.473, + 0.224, + 0.489 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.484, + 0.604 + ], + "angle": 0, + "content": "This section is organized as follows: First, we present the model architecture and pre-training details. Second, we validate the effectiveness of our pre-trained model on a wide range of downstream tasks, including object classification, few-shot learning, and part segmentation. Finally, ablation studies are conducted to demonstrate the properties and robustness of our proposed Point-PQAE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.609, + 0.276, + 0.625 + ], + "angle": 0, + "content": "4.1. Pre-training setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.689 + ], + "angle": 0, + "content": "Pre-trained dataset. We use the dataset ShapeNet [6] for pre-training of Point-PQAE following previous studies [9, 33, 37]. ShapeNet [6] consists of about 51,300 clean 3D models, covering 55 common object categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.824 + ], + "angle": 0, + "content": "Model structure. We adopt standard Transformer blocks [51] in the autoencoder's backbone where the encoder has 12 Transformer blocks and the decoder has 4 blocks. MLP ratio in Transformer blocks is set to 4. Each Transformer block has 384 hidden dimensions and 6 heads. The positional query module between encoder and decoder only performs cross-attention mechanism once as stated in Sec. 3.2, and the number of heads is set to 6 and remains 384 hidden dimensions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Experiment detail. For each input point cloud, we only apply random rotation for pre-training data augmentation. After sampling 1024 points via FPS from the input point cloud, we generate two different partial point clouds by randomly cropping and rotating as illustrated in Sec. 3.1 with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "\\(r_m = 0.6\\). Both of them are divided into 64 patches with 32 points via FPS and KNN. The Point-PQAE model undergoes pre-training for 300 epochs using an AdamW [29] optimizer with a batch size of 128. The initial learning rate is set to 0.0005, with a weight decay of 0.05, with cosine learning rate decay [28] utilized. We visualize the pretraining results in Fig. 3. It shows Point-PQAE learns cross-knowledge well and is able to generalize excellently to other crop ratios, even though it was pre-trained with \\(r_m = 0.6\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.233, + 0.851, + 0.248 + ], + "angle": 0, + "content": "4.2. Transfer learning on downstream tasks" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.252, + 0.905, + 0.297 + ], + "angle": 0, + "content": "Transfer protocol. We use three transfer learning protocols for classification tasks following [9]: FULL, MLP-LINEAR, MLP-3, which are detailed in Appendix Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.906, + 0.599 + ], + "angle": 0, + "content": "3D real-world object classification. The transferability of models can be well demonstrated through testing on a 3D real-world object dataset. Therefore, we transfer our pre-trained model to ScanObjectNN [50] for classification, which is one of the most challenging 3D datasets, covering approximately 15,000 real-world objects across 15 categories. We conduct experiments on three variants: OBJBG, OBJ-ONLY, and PB-T50-RS. The results are reported in Tab. 1. Our Point-PQAE surpasses previous single-modal self-supervised methods by a margin under all protocols. In contrast to the most relevant baseline, Point-MAE, PointPQAE significantly outperforms it, showing improvements of \\(2.4\\%\\), \\(1.7\\%\\), and \\(1.2\\%\\). Especially for the MLP-LINEAR and MLP-3 protocols, our Point-PQAE outperforms Point-MAE by \\(6.7\\%\\) and \\(4.4\\%\\) on average, respectively. The excellent frozen representation evaluation results indicate the superior quality of the representation learned by our Point-PQAE during pre-training. This shows that the cross-reconstruction paradigm learns more robust representations and generalizes better on real-world datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.795 + ], + "angle": 0, + "content": "3D clean object classification. We evaluate the performance of our pre-trained model for object classification using the ModelNet40 dataset [59]. ModelNet40 comprises 12,311 meticulously crafted 3D CAD models, representing 40 distinct object categories. During testing, we adhere to the standard voting method [26] to ensure fair comparisons with previous work such as [7, 33]. Standard random scaling and random translation are applied for data augmentation during training, and the input point clouds exclusively contain coordinate information, without additional normal information provided. The results are presented in Tab. 1. Our Point-PQAE yields better or comparable results compared to previous approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Few-shot learning. We conduct the few-shot learning classification on the ModelNet40 dataset to show the generalization ability of our method following the protocols in previous studies [33, 41, 53, 65]. The few-shot learning experiments are conducted in the form of \"w-way, s-shot\" including four parts where \\( w \\in \\{5, 10\\} \\) represents the number of randomly selected classes and \\( s \\in \\{10, 20\\} \\) means" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28700" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.128, + 0.283, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.188, + 0.238, + 0.262, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.328, + 0.28, + 0.34 + ], + "angle": 0, + "content": "Input point cloud" + }, + { + "type": "image_caption", + "bbox": [ + 0.31, + 0.104, + 0.413, + 0.116 + ], + "angle": 0, + "content": "Cropped Rebuilt" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.127, + 0.422, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.223, + 0.421, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.328, + 0.404, + 0.339 + ], + "angle": 0, + "content": "Crop ratio \\(20\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.104, + 0.544, + 0.116 + ], + "angle": 0, + "content": "Cropped Rebuilt" + }, + { + "type": "image", + "bbox": [ + 0.435, + 0.126, + 0.557, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.221, + 0.55, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.328, + 0.544, + 0.339 + ], + "angle": 0, + "content": "Crop ratio \\(40\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.573, + 0.104, + 0.674, + 0.116 + ], + "angle": 0, + "content": "Cropped Rebuilt" + }, + { + "type": "image", + "bbox": [ + 0.567, + 0.121, + 0.681, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.57, + 0.226, + 0.685, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.328, + 0.67, + 0.339 + ], + "angle": 0, + "content": "Crop ratio \\(60\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.703, + 0.104, + 0.805, + 0.116 + ], + "angle": 0, + "content": "Cropped Rebuilt" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.121, + 0.812, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.224, + 0.81, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.327, + 0.802, + 0.339 + ], + "angle": 0, + "content": "Crop ratio \\(80\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.354, + 0.905, + 0.383 + ], + "angle": 0, + "content": "Figure 3. Cross-reconstruction results on ShapeNet. The arrow points from source point clouds to cross-reconstruction results. PointPQAE generalizes well to other crop ratios though with minimum crop ratio \\( r_m = 0.6 \\) when pre-training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.39, + 0.484, + 0.542 + ], + "angle": 0, + "content": "the number of randomly selected samples for each \\( w \\). Each part is conducted with 10 independent trials. The mean accuracy and standard deviation are reported in Tab. 2. Our Point-PQAE demonstrates superior performance compared to previous methods in few-shot learning, even surpassing cross-modal methods that utilize strong pre-trained teachers such as ACT and ReCon under FULL protocol. Specifically, Point-PQAE achieves \\( 2.7\\% \\), \\( 1.1\\% \\), \\( 4.5\\% \\), \\( 1.9\\% \\) improvement under MLP-LINEAR protocol and shows enhanced performance over the self-construction method Point-MAE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.79 + ], + "angle": 0, + "content": "Part segmentation. We conduct part segmentation experiments on the ShapeNetPart [64] to validate the effectiveness of our Point-PQAE. The ShapeNetPart contains 16,881 objects covering 16 categories. We sample 2,048 points from each input point cloud, following previous work [33, 65] and divide each cloud into 128 point patches. We use the same segmentation head and utilize learned features from the 4th, 8th, and 12th layers of the Transformer block as in Point-MAE [33]. We concatenate three levels of features. Then average pooling, max pooling, and upsampling are utilized to generate features for each point and an MLP is applied for label prediction. The results are reported in Tab. 3a. The results are reported in Tab. 3a, which show that Point-PQAE achieves a comparable Inst.mIoU to the previous method, achieving \\(84.6\\%\\) in Cls.mIoU and improving the from-scratch baseline by \\(1.4\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "3D scene segmentation. Semantic segmentation on large-scale 3D scenes is challenging, requiring models possessing a deep comprehension of semantics and intricate local geometric relationships. We report semantic segmentation results on the S3DIS dataset [3] in Tab. 3b. Our Point-PQAE improves the from-scratch baseline by \\(2.0\\%\\) and \\(1.4\\%\\), and outperforms the self-reconstruction method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.39, + 0.853, + 0.405 + ], + "angle": 0, + "content": "Point-MAE by \\(0.7\\%\\) and \\(0.6\\%\\) in mAcc and mIoU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.413, + 0.663, + 0.429 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.491 + ], + "angle": 0, + "content": "Experiments are conducted to show the properties of our Point-PQAE. We report the classification results on three variants of ScanObjectNN. The experiment settings are aligned with Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.906, + 0.854 + ], + "angle": 0, + "content": "View-relative positional embedding. To analyze the effect of view-relative positional embedding in the positional query block, we conduct a series of experiments using different types of positional embeddings. We primarily consider two types of embeddings: Sinusoid (as discussed in Sec. 3.2) and Learnable, along with a None group. Learnable embedding refers to utilizing the relative position information \\(\\mathbf{RP}_{1\\rightarrow 2}\\) and \\(\\mathbf{RP}_{2\\rightarrow 1}\\) as input and employing an MLP composed of two linear layers with an activation function to map the 6-dimensional input to a D-dimensional output. The None group involves using randomly assigned embeddings to assess the effectiveness of VRPE design. Tab. 4a displays the results of different positional embeddings. In the None group, the model cannot learn the relative positional information across two views, leading to a significant drop in accuracy. In addition to the type of VRPE, we also experiment with Absolute Positional Embedding (APE) to compare its performance with our proposed VRPE. The APE is unsuitable for our framework since it requires a shared coordinate system for both views, making it incompatible with the normalization and rotation used in our method. We incorporate APE into our Point-PQAE by removing these operations. The results in Tab. 4a show APE performs worse and support our claim." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Data augmentation. After applying the random crop mechanism to the point cloud, we further perform data augmentation on the cropped views to generate new views." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28701" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. Classification accuracy (%) on ScanObjectNN and ModelNet40. The inference model parameters #P (M) are reported. Three variants are evaluated on ScanObjectNN and the accuracy obtained on ModelNet40 is reported for both 1K and 8K points. We compare methods using the \\(\\bullet\\) plain Transformer architectures, e.g. Point-MAE [33], \\(\\circ\\) hierarchical Transformer architectures and \\(\\circ\\) dedicated architectures for 3D. The dagger(†) denotes the baseline results reported from ReCon [37] which aligns augmentation with us and ReCon." + }, + { + "type": "table", + "bbox": [ + 0.214, + 0.15, + 0.778, + 0.687 + ], + "angle": 0, + "content": "
Methods#PScanObjectNNModelNet40
OBJ_BGOBJ_ONLYPB_T50_RS1K P8K P
Supervised Learning Only
○PointNet [34]3.573.379.268.089.290.8
○PointNet++ [35]1.582.384.377.990.791.9
○DGCNN [54]1.882.886.278.192.9-
○PointCNN [24]0.686.185.578.592.2-
○SimpleView [13]---80.5±0.393.9-
○PCT [15]2.88---93.2-
with Self-Supervised Representation Learning (FULL)
•Transformer [51]22.183.084.079.191.491.8
•Point-BERT [65]22.187.488.183.093.293.8
•MaskPoint [25]-89.388.184.393.8-
•Point-MAE [33]22.190.088.385.293.894.0
○Point-M2AE [66]15.391.288.986.494.0-
•PointDif [74]22.193.391.987.6--
•Point-MAE† [33]22.192.691.988.493.894.0
•Point-PQAE22.195.093.689.694.094.3
Methods using cross-modal information and teacher models
•Joint-MAE [16]-90.988.986.194.0-
•TAP [55]22.190.489.585.794.0-
•ACT [9]22.193.391.988.293.794.0
○PointMLP+ULIP [63]---89.494.594.7
○I2P-MAE [67]15.394.291.690.194.1
•ReCon [37]43.695.293.690.694.594.7
with Self-Supervised Representation Learning (MLP-LINEAR)
•Point-MAE† [33]22.182.8±0.383.2±0.274.1±0.290.2±0.190.7±0.1
•Point-PQAE22.189.3±0.390.2±0.480.8±0.192.0±0.292.2±0.1
Methods using cross-modal information and teacher models
•ACT [9]22.185.2±0.885.8±0.276.3±0.391.4±0.291.8±0.2
•ReCon [37]43.689.5±0.289.7±0.281.4±0.192.5±0.292.7±0.1
with Self-Supervised Representation Learning (MLP-3)
•Point-MAE† [33]22.185.8±0.385.5±0.280.4±0.291.3±0.291.7±0.2
•Point-PQAE22.190.7±0.290.9±0.283.3±0.192.8±0.192.9±0.1
Methods using cross-modal information and teacher models
•ACT [9]22.187.1±0.287.9±0.481.5±0.292.7±0.293.0±0.1
•ReCon [37]43.690.6±0.290.7±0.383.8±0.493.0±0.193.4±0.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.699, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Data augmentation is crucial for generating diverse and decoupled views, so we conducted ablation studies on different augmentation methods. The results in Tab. 4b indicate that rotation performs best." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of the crop mechanism. Random crop has been shown to be highly effective in 2D self-supervised learning (SSL), particularly for contrastive learning [4, 8, 14], and is also vital for generative learning [19]. However, its potential in 3D SSL remains largely unexplored. Even contrastive learning methods for point cloud understanding have not explored the power of it [2, 32]. To address this gap, we first introduce a custom-designed random crop mechanism for point cloud data, using it as the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.907, + 0.896 + ], + "angle": 0, + "content": "foundation for generating decoupled views. Without the crop augmentation, the two views will contain exactly the same shape though with normalization and rotation applied. The model would only need to infer the augmentation rather than the inter-view relationship, which means the views are not decoupled. Conversely, given two isolated cropped point clouds, followed by min-max normalization and rotation for further decoupling, there remain overlapping and non-overlapping parts, requiring the model to effectively encode the intra-view parts in order to infer the unknown inter-view points in our Point-PQAE. However, directly applying random crop as an additional augmentation to the self-reconstruction method (Point-MAE) does" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28702" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.16 + ], + "angle": 0, + "content": "Table 2. Few-shot classification results on ModelNet40. 10 independent trials are conducted in each experimental setting. The mean accuracy (\\%) and standard deviation are reported for each setting. The dagger(\\(\\dagger\\)) denotes the baseline results reported from ReCon [37] which aligns augmentation with us and ReCon." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.165, + 0.46, + 0.472 + ], + "angle": 0, + "content": "
Methods5-way10-way
10-shot20-shot10-shot20-shot
○DGCNN [54]31.6±2.840.8±4.619.9±2.116.9±1.5
○OcCo [52]90.6±2.892.5±1.982.9±1.386.5±2.2
with Self-Supervised Representation Learning (FULL)
•Transformer [51]87.8±5.293.3±4.384.6±5.589.4±6.3
•Point-BERT [65]94.6±3.196.3±2.791.0±5.492.7±5.1
•MaskPoint [25]95.0±3.797.2±1.791.4±4.093.4±3.5
•Point-MAE [33]96.3±2.597.8±1.892.6±4.195.0±3.0
○Point-M2AE [66]96.8±1.898.3±1.492.3±4.595.0±3.0
•Point-MAE† [66]96.4±2.897.8±2.092.5±4.495.2±3.9
•Point-PQAE96.9±3.298.9±1.094.1±4.296.3±2.7
Methods using cross-modal information and teacher models
•Joint-MAE [16]96.7±2.297.9±1.992.6±3.795.1±2.6
•TAP [55]97.3±1.897.8±1.993.1±2.695.8±1.0
•ACT [9]96.8±2.398.0±1.493.3±4.095.6±2.8
○I2P-MAE [67]97.0±1.898.3±1.392.6±5.095.5±3.0
•ReCon [37]97.3±1.998.9±1.293.3±3.995.8±3.0
with Self-Supervised Representation Learning (MLP-LINEAR)
•Point-MAE† [66]91.1±5.691.7±4.083.5±6.189.7±4.1
•Point-PQAE93.0±4.696.8±1.989.0±5.293.5±3.8
with Self-Supervised Representation Learning (MLP-3)
•Point-MAE† [66]95.0±2.896.7±2.490.6±4.793.8±5.0
•Point-PQAE95.3±3.498.2±1.892.0±3.894.7±3.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.518 + ], + "angle": 0, + "content": "Table 3. Segmentation results. Cls.mIoU (\\%) and Inst.mIoU (\\%) refer to Mean intersection over union for all classes and all instances, respectively. mAcc (\\%) refers to mean accuracy." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.523, + 0.283, + 0.547 + ], + "angle": 0, + "content": "(a) Part segmentation on ShapeNetPart." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.555, + 0.283, + 0.66 + ], + "angle": 0, + "content": "
MethodCls.mIoU (%)Inst.mIoU (%)
Supervised
PointNet [34]80.483.7
PointNet++ [35]81.985.1
DGCNN [54]82.385.2
Self-supervised
Transformer [51]83.484.7
CrossPoint [2]-85.5
Point-BERT [65]84.185.6
Point-MAE [33]84.286.1
Point-PQAE (Ours)84.686.1
Cross-modal
ACT [9]84.786.1
ReCon [37]84.886.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.293, + 0.523, + 0.483, + 0.547 + ], + "angle": 0, + "content": "(b) Semantic segmentation results on S3DIS Area 5." + }, + { + "type": "table", + "bbox": [ + 0.297, + 0.556, + 0.484, + 0.646 + ], + "angle": 0, + "content": "
MethodmAcc (%)mIoU (%)
Supervised
PointNet [34]49.041.1
PointNet++ [35]67.153.5
Self-supervised
Transformer [51]68.660.0
Point-MAE [33]69.960.8
Point-PQAE (Ours)70.661.4
Cross-modal
ACT [9]71.161.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.779 + ], + "angle": 0, + "content": "not improve the performance much, as shown in Tab. 4c. In other words, it means our Point-PQAE fits crop mechanism better by harmoniously incorporating it into the decoupled views generation process. The experiment results in Tab. 4c demonstrate the critical role of random crop in our method. We set minimum crop ratio \\( r_m = 1.0 \\) when removing the crop in our Point-PQAE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Analysis on decoupled views generation. While crop plays an important role in our method, relying solely on the random crop mechanism to generate new views would degenerate cross-reconstruction to simply reconstructing one part from another. This is similar to the block mask self-reconstruction used in Point-MAE [33], where visible blocks are used to reconstruct masked blocks, which is also trivial for pre-training. The only difference between them" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.148 + ], + "angle": 0, + "content": "Table 4. Ablation study with Point-PQAE pre-training on ShapeNet. The classification results by accuracy (\\%) on three variants of ScanObjectNN are reported. Default settings are marked in gray." + }, + { + "type": "table_caption", + "bbox": [ + 0.604, + 0.147, + 0.816, + 0.16 + ], + "angle": 0, + "content": "(a) View-Relative positional embedding" + }, + { + "type": "table", + "bbox": [ + 0.57, + 0.162, + 0.855, + 0.218 + ], + "angle": 0, + "content": "
Positional EmbeddingOBJ_BGOBJ_ONLYLYPB_T50_RS
None84.585.979.3
APE (sinusoid)92.391.087.7
VRPE (learnable)93.493.189.1
VRPE (sinusoid Eq. (10))95.093.689.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.651, + 0.22, + 0.771, + 0.232 + ], + "angle": 0, + "content": "(b) Data augmentation" + }, + { + "type": "table", + "bbox": [ + 0.57, + 0.233, + 0.853, + 0.3 + ], + "angle": 0, + "content": "
Data augmentationOBJ_BGOBJ_ONLYPB_T50_RS
jitter93.391.487.3
scale93.391.788.0
rotation95.093.689.6
scale&translate92.891.788.1
rotation+scale&translate93.892.989.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.626, + 0.301, + 0.796, + 0.314 + ], + "angle": 0, + "content": "(c) 3D random crop mechanism" + }, + { + "type": "table", + "bbox": [ + 0.57, + 0.317, + 0.853, + 0.376 + ], + "angle": 0, + "content": "
MethodcropOBJ_BGOBJ_ONLYPB_T50_RS
Point-MAE✘92.691.988.4
Point-MAE✓92.992.188.8
Point-PQAE✘92.992.387.7
Point-PQAE✓95.093.689.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.585, + 0.378, + 0.833, + 0.39 + ], + "angle": 0, + "content": "(d) Effects of augmentations after random crop" + }, + { + "type": "table", + "bbox": [ + 0.569, + 0.393, + 0.849, + 0.456 + ], + "angle": 0, + "content": "
norm.rotationOBJ_BGOBJ_ONLYPB_T50_RS
XX92.892.188.2
✓X93.392.687.9
X✓93.592.688.8
✓✓95.093.689.6
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.907, + 0.646 + ], + "angle": 0, + "content": "is that the cropped parts can overlap somewhat, whereas the block mask does not. What makes the cross-reconstruction different from block mask self-reconstruction are the subsequent augmentations including normalization and rotation that decouple the cropped parts. By normalizing cropped point clouds centered on the geometric centers, the coordinate systems of the two point clouds become isolated and independent from each other, with rotation further amplifying the variance between the two views, thus enabling the generation of two decoupled views. The importance of the augmentations applied after the random crop can be validated through experiments, as shown in Tab. 4d." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.654, + 0.634, + 0.669 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel cross-reconstruction generative framework for self-supervised learning on 3D point clouds, called Point-PQAE. In contrast to well-studied self-reconstruction schemes, Point-PQAE reconstructs one cropped point cloud from another decoupled point cloud. To supply sufficient information for cross-reconstruction, we further propose a 3D view-relative positional embedding and a corresponding position-aware query module. Compared to self-reconstruction, the cross-reconstruction task, carefully designed by us, is much more challenging for pre-training, promoting the learning of richer semantic representations during pre-training. This enables our proposed Point-PQAE to outperform previous single-modal self-reconstruction methods by a margin and to perform on par with, or better than cross-modal methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28703" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.256 + ], + "angle": 0, + "content": "[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[3] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.371 + ], + "angle": 0, + "content": "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.482, + 0.513 + ], + "angle": 0, + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[7] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. Advances in Neural Information Processing Systems, 36, 2024. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[9] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[10] Bi'an Du, Xiang Gao, Wei Hu, and Xin Li. Self-contrastive learning with hard negative sampling for self-supervised point cloud learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3133–3142, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[11] Ikenna Enebuse, Mathias Foo, Babul Salam Ksm Kader Ibrahim, Hafiz Ahmed, Fhon Supmak, and Odongo Steven Eyobu. A comparative review of hand-eye calibration techniques for vision guided robots. IEEE Access, 9:113143-113155, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In International Conference on Machine Learning, pages 3809-3820. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[15] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.288, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[16] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzhi Li, and Pheng-Ann Heng. Joint-mae: 2d-3d joint masked autoencoders for 3d point cloud pre-training. arXiv preprint arXiv:2302.14007, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.469, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.905, + 0.606 + ], + "angle": 0, + "content": "[20] Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6535-6545, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.608, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[22] Gang Li, Heliang Zheng, Daqing Liu, Chaoyue Wang, Bing Su, and Changwen Zheng. Semmae: Semantic-guided masking for learning masked autoencoders. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[23] Jiaxin Li, Ben M Chen, and Gim Hee Lee. So-net: Self-organizing network for point cloud analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9397–9406, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.762, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[24] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[25] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28704" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "cloud analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8895-8904, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.484, + 0.205 + ], + "angle": 0, + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.275, + 0.483, + 0.33 + ], + "angle": 0, + "content": "[30] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[31] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[32] Bo Pang, Hongchi Xia, and Cewu Lu. Unsupervised 3d point cloud representation learning by triangle constrained contrast for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5229-5239, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.51 + ], + "angle": 0, + "content": "[33] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.483, + 0.637 + ], + "angle": 0, + "content": "[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.483, + 0.693 + ], + "angle": 0, + "content": "[36] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.483, + 0.761 + ], + "angle": 0, + "content": "[37] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. arXiv preprint arXiv:2302.02318, 2023. 1, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[38] Rui Qian, Xin Lai, and Xirong Li. 3d object detection for autonomous driving: A survey. Pattern Recognition, 130: 108796, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.874 + ], + "angle": 0, + "content": "[39] Aditya Sanghi. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX 16, pages 626-642. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[40] Oleg Yu Sergiyenko and Vera V Tyrsa. 3d optical machine vision sensors with intelligent data management for robotic" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "swarm navigation improvement. IEEE Sensors Journal, 21 (10):11262-11274, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[41] Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. Advances in Neural Information Processing Systems, 33:7212-7221, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[43] Haoru Tan, Chuang Wang, Sitong Wu, Tieqiang Wang, Xu-Yao Zhang, and Cheng-Lin Liu. Proxy graph matching with proximal matching networks. In The Annual AAAI Conference on Artificial Intelligence (AAAI), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[44] Haoru Tan, Sitong Wu, Fei Du, Yukang Chen, Zhibin Wang, Fan Wang, and Xiaojuan Qi. Data pruning via moving-one-sample-out. In Neural Information Processing Systems (NeurIPS), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.363 + ], + "angle": 0, + "content": "[45] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.42 + ], + "angle": 0, + "content": "[46] Haoru Tan, Chuang Wang, Sitong Wu, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Ensemble quadratic assignment network for graph matching. International Journal of Computer Vision (IJCV), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[47] Haoru Tan, Sitong Wu, Zhuotao Tian, Yukang Chen, Xiaojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity consistency for vision-language pre-training. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[48] Haoru Tan, Sitong Wu, Wei Huang, Shizhen Zhao, and Xiaojuan Qi. Data pruning by information maximization. In International Conference on Learning Representations (ICLR), 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[49] Haoru Tan, Sitong Wu, Bo Zhao, Zeke Xie, and XIAOJUAN QI. Diff-in: Data influence estimation with differential approximation, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[50] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.675, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[52] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[53] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.903 + ], + "angle": 0, + "content": "[54] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28705" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[55] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. Take-a-photo: 3d-to-2d generative pre-training of point cloud models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5640–5650, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[56] Hai Wu, Ruifei He, Haoru Tan, Xiaojuan Qi, and Kaibin Huang. Vertical layering of quantized neural networks for heterogeneous inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[57] Sitong Wu, Haoru Tan, Yukang Chen, Shaofeng Zhang, Jingyao Li, Bei Yu, Xiaojuan Qi, and Jiaya Jia. Mixture-of-scores: Robust image-text data quality score via three lines of code. In International Conference on Computer Vision (ICCV), 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[58] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. arXiv preprint arXiv:2312.10035, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[59] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[60] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.575 + ], + "angle": 0, + "content": "[61] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[62] Mingye Xu, Zhipeng Zhou, Yali Wang, and Yu Qiao. Towards robustness and generalization of point cloud representation: A geometry coding method and a large-scale object-level dataset. Computational Visual Media, 10(1):27-43, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[63] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1179-1189, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[64] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[65] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[66] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[67] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[68] Shaofeng Zhang, Qiang Zhou, Sitong Wu, Haoru Tan, Zhibin Wang, Jina Huang, and Junchi Yan. Cr2pq: Continuous relative rotary positional query for dense visual representation learning. In The Thirteenth International Conference on Learning Representations. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[69] Shaofeng Zhang, Meng Liu, and Junchi Yan. The diversified ensemble neural network. Advances in Neural Information Processing Systems, 33:16001-16011, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[70] Shaofeng Zhang, Feng Zhu, Junchi Yan, Rui Zhao, and Xiaokang Yang. Zero-cl: Instance and feature decorrelation for negative-free symmetric contrastive learning. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[71] Shaofeng Zhang, Meng Liu, Junchi Yan, Hengrui Zhang, Lingxiao Huang, Xiaokang Yang, and Pinyin Lu. M-mix: Generating hard negatives via multi-sample mixing for contrastive learning. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, pages 2461-2470, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[72] Shaofeng Zhang, Feng Zhu, Rui Zhao, and Junchi Yan. Contextual image masking modeling via synergized contrasting without view augmentation for faster and better visual pretraining. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[73] Xiangdong Zhang, Shaofeng Zhang, and Junchi Yan. Pcmae: Learning to predict centers for point masked autoencoders. Advances in Neural Information Processing Systems, 37:80303-80327, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[74] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22935-22945, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[75] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.756 + ], + "angle": 0, + "content": "[76] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28706" + } + ] +] \ No newline at end of file diff --git a/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/3caf6ed2-ef92-4e47-ab15-82d341d3a49f_origin.pdf b/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/3caf6ed2-ef92-4e47-ab15-82d341d3a49f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3d8a9f73005b2f449f0256ad16bb25458a19b4d1 --- /dev/null +++ b/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/3caf6ed2-ef92-4e47-ab15-82d341d3a49f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6a2f588e006a99628c72d195815d35f418d01d3e835aaf94f10acd8d0c7beff +size 1498860 diff --git a/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/full.md b/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1ae2f65a01671760201f1ef34ffc6cd5a40aad7b --- /dev/null +++ b/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/full.md @@ -0,0 +1,371 @@ +# Towards More Diverse and Challenging Pre-training for Point Cloud Learning: Self-Supervised Cross Reconstruction with Decoupled Views + +Xiangdong Zhang*, Shaofeng Zhang*, Junchi Yan‡ + +School of AI, Shanghai Jiao Tong University + +{zhangxiangdong, sherrylone, yanjunchi}@sjtu.edu.cn + +# Abstract + +Point cloud learning, especially in a self-supervised way without manual labels, has gained growing attention in both vision and learning communities due to its potential utility in a wide range of applications. Most existing generative approaches for point cloud self-supervised learning focus on recovering masked points from visible ones within a single view. Recognizing that a two-view pre-training paradigm inherently introduces greater diversity and variance, it may thus enable more challenging and informative pre-training. Inspired by this, we explore the potential of two-view learning in this domain. In this paper, we propose Point-PQAE, a cross-reconstruction generative paradigm that first generates two decoupled point clouds/views and then reconstructs one from the other. To achieve this goal, we develop a crop mechanism for point cloud view generation for the first time and further propose a novel positional encoding to represent the 3D relative position between the two decoupled views. The cross-reconstruction significantly increases the difficulty of pre-training compared to self-reconstruction, which enables our method to surpass previous single-modal self-reconstruction methods in 3D self-supervised learning. Specifically, it outperforms the self-reconstruction baseline (Point-MAE) by $6.5\%$ , $7.0\%$ , and $6.7\%$ in three variants of ScanObjectNN with the MLP-LINEAR evaluation protocol. The code is available at https://github.com/aHapBean/Point-PQAE. + +# 1. Introduction + +3D vision is gaining increasing attention for its wide applications such as autonomous driving [38] and robotics [11, 40], owing to its ability to understand the human world. The point cloud is the most popular representation form of data in 3D vision and many analyses based on point cloud are + +![](images/71fc8b35306dce4462853f98b5b0cc64850f75583a1d84e87d9db6913c22d847.jpg) +Self-Reconstruction + +![](images/12e4a292f90bab83a13a9f8aa0c2fe8a949e9c3cfd6e2f3a13bbaceb4845174a.jpg) +Cross-Reconstruction (Ours) +Figure 1. Comparison between self-reconstruction (Point-MAE [33] and other methods based on it) and cross-reconstruction (Ours) paradigms. In the self-reconstruction paradigm, part of the input data is masked, and the autoencoder is trained to recover the missing patches. In contrast, the cross-reconstruction paradigm generates decoupled views first, with one view leveraging cross-view information to reconstruct the other, using an autoencoder backbone. + +explored to solve various tasks such as object classification [15, 24, 34, 35, 54, 62], object detection [30, 36, 76] and segmentation [15, 24, 34, 54, 58], they typically necessitate fully-supervised training from scratch. Compared to 2D data, point cloud generally requires more expensive and labor-intensive efforts to collect and annotate, which hinders the development of the fully supervised 3D representation learning method. Self-supervised learning is one of the predominant approaches to address this issue and has been proven to be effective in 2D vision [5, 8, 14, 18, 19, 68-71, 75]. Inspired by this, self-supervised learning has been widely studied in the 3D field in recent years. + +Self-supervised learning in 3D can be mainly divided into two categories, i.e., generative methods [1, 7, 23, 25, 33, 52, 66, 73] and contrastive methods [2, 9, 10, 20, 37, 39, 60]. Similar to 2D, contrastive methods in 3D aim to learn global-discriminative information by maximizing the mutual information across views [2, 9, 60] with intra-/inter-modal information [43-49, 56, 57]. PointContrast [60] is the first to learn transformation invariance across views through the extension of InfoNCE objective [31, 60]. Dif + +ferent from contrastive methods, the core idea of the generative methods [25, 33, 66] in 3D is inspired by MAE [19], which masks the input point cloud and uses the visible part to reconstruct the masked part. Point-MAE [33] is one of the conventional methods, which learns point-wise information through the self-reconstruction mechanism. Specifically, Point-MAE masks a high ratio of the point cloud and reconstructs the masked points through the visible part. Point-M2AE [66] proposes a hierarchical Transformer [27] with a corresponding masking strategy based on it, capturing both fine-grained and high-level semantics of 3D shapes. + +Existing generative methods mostly focus on the masked reconstruction in a single view, where single-view learning has been proven less difficult and informative than two views [8, 14] since two views would bring more variance than a single view. To bring the benefits of the two-view learning paradigm to the generative learning field, one straightforward assumption is adopting the two-view learning paradigm in the 3D generative pertaining tasks to increase the difficulty of point-wise reconstruction. However, unlike 2D vision data, constructing two views in the 3D point cloud is more challenging, as point cloud data is less structured. Besides, cross-reconstruction (reconstructing one view from the other view) is much more difficult than self-reconstruction (reconstructing the masked single view), which involves learning both the inter-positional relation between two views and the inner spatial information within the view. The differences between these two paradigms can be seen in Fig. 1. Aimed at solving the above issues, we propose Point-PQAE (Position Query based AutoEncoder for Point Cloud), a novel framework aimed at addressing two primary challenges: i) how to construct two views in point cloud data, and ii) how to perform cross-reconstruction between two views. As shown in Fig. 2, for each point cloud, we first apply a custom-designed point cloud crop mechanism, which randomly selects two points as the central points of two views and records their positions. Following this, for each central point, the nearest points according to a specified ratio are incorporated into the view. Subsequently, we normalize the cropped point cloud and apply random augmentation (e.g., rotation) to each point cloud to achieve view decoupling. Then we calculate the view-relative positional embedding (VRPE) for these two decoupled views. Finally, we take the VRPE and one view of the point cloud as input to predict the point-wise input of the other view. + +# The highlights of this paper are: + +i) New generative framework: We propose PointPQAE, the first framework that successfully brings cross-reconstruction to 3D generative field for point cloud self-supervised learning, with three modules: 1) decoupled views generation, 2) VRPE generation, and 3) positional query block. To our knowledge, we are the first to design + +and apply crop mechanism to point cloud self-supervised learning. Our framework, Point-PQAE, enables more informative and challenging self-supervised pre-training compared to existing self-reconstruction methods. + +ii) The first relative position-aware query module for point cloud: We introduce a positional query block after the encoder to capture relative position information between decoupled views. The module applies cross-attention between the hidden representations of one view and the VRPE, enabling it to predict the other view. Our module can be easily plugged into various related tasks, e.g., distillation as stated in Appendix Sec. 4. + +iii) Strong performance: Point-PQAE achieves new state-of-the-art performance on several benchmarks, e.g., outperforming previous methods on few-shot learning and achieving average improvements of $1.8\%$ , $6.7\%$ , and $4.4\%$ over the baseline (Point-MAE) on ScanObjectNN classification across three evaluation protocols (FULL, MLP-LINEAR, MLP-3), respectively. + +# 2. Related works + +SSL in 2D vision. Self-supervised learning (SSL) in 2D vision is broadly categorized into contrastive-based and masked-image-modeling (MIM) approaches. Contrastive methods learn discriminative features by creating two augmented views of an image and enforcing invariance between them. For instance, SimCLR [8] treats augmented views from the same image as positive pairs and those from different images as negatives. MoCo [18] utilizes a memory bank to store these negative samples and further proposes a momentum-updated mechanism to avoid model collapse. MIM-based methods mask portions of the input and task models with reconstructing the masked regions, encouraging fine-grained understanding. MAE [19] employs an encoder-decoder architecture, discarding masked patches before reconstruction, while SimMIM [61] replaces masked regions with a learnable parameter. Subsequent work explores variations such as masking strategies [22, 72] and feature-level reconstruction [75]. + +Representation learning for point clouds. It is a challenging task due to its irregular and sparse nature when compared to other modalities such as images which are well-structured and uniform. PointContrast [60] firstly borrows the idea of contrastive learning in 2D and performs point-level invariant mapping on two transformed views of the given point cloud. On the basis of PointContrast, Cross-Point [2] further proposes inter and intra-modal contrastive objectives to enhance the global-discriminative capability. To better model multi-modal data, ACT [9] employs cross-modal auto-encoders as teacher models to acquire knowledge from other modalities. In contrast to contrastive objectives, Point-MAE [33] proposes to randomly mask points, and reconstruct the masked part through the visible part. + +![](images/204b1936a68fac0d3f61ec2f1b08677b5bd4ca3b8b7f2f670b75507f633993b7.jpg) +Figure 2. Pipeline of Point-PQAE. The input point cloud is randomly cropped followed by the rotation to generate views. Then, we feed the views to the patch embedding layer and the transformer encoder, followed by the proposed positional query block. The View-Relative Positional Embedding (VRPE) is obtained through the VRPE module by extracting relative geometric relations, which is taken as "Query" in the cross-attention mechanism. The queried hidden embeddings are then fed to the decoder to predict the inputs of the other view. + +Following Point-MAE and CrossPoint, ReCon [37] harmoniously incorporates the generative method Point-MAE and the contrastive method. Almost all the existing generative methods in 3D rely on self-reconstruction. To this end, we propose a new cross-reconstruction generative framework, named Point-PQAE, which increases pre-training difficulty by introducing cross-view reconstruction, bringing more variance to the training samples compared to single-view reconstruction. This requires the model to effectively learn both intra-view and inter-view knowledge, thereby enabling the model to learn more semantic representations. + +# 3. The proposed Point-PQAE + +# 3.1. Views and embeddings generation + +Decoupled views generation. Random Image Crop has been widely used in both the supervised domain [17, 21, 42] and the self-supervised domain [4, 8, 18], and it has been proved to be effective in practice. This method typically involves randomly selecting a rectangular region from an image and cropping it. However, to the best of our knowledge, no corresponding crop mechanism has been proposed for 3D self-supervised representation learning. Instead of directly applying the 2D crop algorithm to 3D (by randomly selecting a cube in space), which may lead to inconsistency, as the points within cubes of the same size can vary significantly due to the diverse distribution of points in 3D, we design in a smarter way: For a randomly selected crop ratio, we first select a center point randomly and then choose the corresponding number of points in the point cloud that + +are closest to the center point. + +Formally, given an input point cloud with $p$ points $\mathbf{X} \in \mathbb{R}^{p \times 3}$ and a pre-defined minimum crop ratio $r_m$ where $0 < r_m < 1$ , we first randomly select the crop ratios $r_1$ and $r_2$ uniformly from $[r_m, 1]$ . Then, we randomly select two center points $\mathbf{C}_1, \mathbf{C}_2 \in \mathbb{R}^{1 \times 3}$ and get $\mathbf{X}_1$ and $\mathbf{X}_2$ by expanding nearest $r_1 \times p$ and $r_2 \times p$ points to the center point $\mathbf{C}_1$ and $\mathbf{C}_2$ respectively in $\mathbf{X}$ . Meanwhile, we record the absolute coordinates of cropped point clouds' geometric centers, which are denoted as $\mathbf{L}_1 \in \mathbb{R}^{1 \times 3}$ and $\mathbf{L}_2 \in \mathbb{R}^{1 \times 3}$ , respectively. The process can be formulated as: + +$$ +r _ {1}, r _ {2} = \text {U n i f o r m S a m p l e} ([ r _ {m}, 1 ]) \tag {1} +$$ + +$$ +\mathbf {C} _ {1}, \mathbf {C} _ {2} = \operatorname {R a n d o m S e l e c t} (\mathbf {X}) \tag {2} +$$ + +$$ +\mathbf {X} _ {1} = \mathrm {N} - \text {N e a r e s t} (\mathbf {C} _ {\mathbf {1}}, r _ {1}, \mathbf {X}) \tag {3} +$$ + +$$ +\mathbf {X} _ {2} = \mathrm {N} - \operatorname {N e a r e s t} \left(\mathbf {C} _ {2}, r _ {2}, \mathbf {X}\right) \tag {4} +$$ + +where $\mathbf{X}_1 \in \mathbb{R}^{r_1p \times 3}$ and $\mathbf{X}_2 \in \mathbb{R}^{r_2p \times 3}$ mean the two cropped views. The $\mathbf{X}_1$ and $\mathbf{X}_2$ are normalized using the min-max normalization, centered on $\mathbf{L}_1$ and $\mathbf{L}_2$ , respectively. Subsequently, we apply additional augmentation to further enhance the variance between two views. We apply random rotations to the cropped views. The distinct normalization centers, along with random rotation, effectively isolate the coordinate systems of the two views and alter their fixed relative relationships. This process decouples the two point clouds, resulting in two independent views. + +Point patches generation. We divide the input point cloud into irregular point patches (may overlap) via the Far- + +thest Point Sampling(FPS) and K-Nearest Neighborhood (KNN) algorithm following Point-BERT [65]. Formally, given two point clouds (views) $\mathbf{X}_1$ and $\mathbf{X}_2$ , we first use the FPS algorithm to sample pre-defined $n$ points as group centers, which we called $\mathbf{G}_1 \in \mathbb{R}^{n \times 3}$ and $\mathbf{G}_2 \in \mathbb{R}^{n \times 3}$ , respectively. Then, based on the center points, we use KNN to choose the pre-defined $k$ -nearest points for each group center $\mathbf{G}^{(i)}$ ( $0 \leq i \leq n$ ) and finally get $n$ groups (each group contains $k$ points). We apply the FPS and KNN on $\mathbf{X}_1$ and $\mathbf{X}_2$ , respectively: + +$$ +\mathbf {G} _ {1} = \operatorname {F P S} \left(\mathbf {X} _ {1}\right), \quad \mathbf {G} _ {2} = \operatorname {F P S} \left(\mathbf {X} _ {2}\right) \tag {5} +$$ + +$$ +\mathbf {P} _ {1} = \operatorname {K N N} \left(\mathbf {X} _ {1}, \mathbf {G} _ {1}\right), \quad \mathbf {P} _ {2} = \operatorname {K N N} \left(\mathbf {X} _ {2}, \mathbf {G} _ {2}\right) \tag {6} +$$ + +where $\mathbf{G}_1, \mathbf{G}_2 \in \mathbb{R}^{n \times 3}$ and $\mathbf{P}_1, \mathbf{P}_2 \in \mathbb{R}^{n \times (k \cdot 3)}$ . Then, the $\mathbf{P}_1$ and $\mathbf{P}_2$ both are treated as a sequence of length $n$ and dimension $k \cdot 3$ as the input. + +# 3.2. Backbone with positional query + +Following [33], we adopt a lightweight PointNet [34], followed by an asymmetric encoder-decoder structure consisting of Transformer blocks in the Point-PQAE. After the encoder projects the input embedding into the latent space, we use the latent representation of one view and view-relative positional embedding (VRPE) between two views to implement the position query through cross-attention [51]. We directly feed the output of the position query block to the decoder to reconstruct another view without using extra positional embedding. The last layer of the model adopts a simple prediction head to achieve the cross-reconstruction. + +Encoding. The encoder consists of the lightweight PointNet and standard Transformer blocks [51], which projects $\mathbf{P}_1$ and $\mathbf{P}_2$ to latent space, respectively: + +$$ +\mathbf {H} _ {1} = \operatorname {E n c o d e r} \left(\mathbf {P} _ {1}, \mathbf {G} _ {1}\right), \quad \mathbf {H} _ {2} = \operatorname {E n c o d e r} \left(\mathbf {P} _ {2}, \mathbf {G} _ {2}\right), \tag {7} +$$ + +where $\mathbf{H}_1\in \mathbb{R}^{n\times D}$ and $\mathbf{H}_2\in \mathbb{R}^{n\times D}$ mean the latent representations of the view 1 and view 2. Note that $D$ is the hidden dimension of the networks. + +Positional query block. After obtaining latent representations $\mathbf{H}_1$ and $\mathbf{H}_2$ , we feed them into the proposed positional query block to extract cross-information. The module can be mainly divided into 3 sub-stages: attain relative information, VRPE generation, and cross-attention. + +Attain relative information. Take reconstructing view 2 from view 1 as an example. To reconstruct view 2, we have to know the latent representations of view 1 and the relative positions of view 2 centered on view 1, where the latent representations of view 1 can be obtained through the encoder network. Therefore, the key to reconstructing view 2 is to obtain the relative positions of view 2 centered on view 1. Back to the crop operation, we have recorded the geometric centers of views, denoted as $\mathbf{L}_1$ and $\mathbf{L}_2$ , respectively. We define the relative position of view 2 centered on view 1 + +as $\mathbf{RL}_{1\rightarrow 2} = \mathbf{L}_1 - \mathbf{L}_2$ where $\mathbf{RL}_{1\rightarrow 2}\in \mathbb{R}^{1\times 3}$ . To better align the dimension in the latter operation, we expand the $\mathbf{RL}_{1\rightarrow 2}$ to $\mathbb{R}^{n\times 3}$ by repeating the content for $n$ times. By obtaining the relative positions between the two views, we further add the group center location to help the PQ modules reconstruct each group of the inputs. Specifically, we define the group-/patch- wise relative positions as: + +$$ +\mathbf {R P} _ {1 \rightarrow 2} = \operatorname {C o n C a t} \left(\mathbf {G} _ {2}, \mathbf {R L} _ {1 \rightarrow 2}\right), \tag {8} +$$ + +$$ +\mathbf {R P} _ {2 \rightarrow 1} = \operatorname {C o n C a t} \left(\mathbf {G} _ {1}, \mathbf {R L} _ {2 \rightarrow 1}\right) \tag {9} +$$ + +where $\mathbf{RP} \in \mathbb{R}^{n \times 6}$ (as $\mathbf{G} \in \mathbb{R}^{n \times 3}$ and $\mathbf{RL} \in \mathbb{R}^{n \times 3}$ ) and $\operatorname{ConCat}(\cdot, \cdot)$ is the concatenate operation. + +View-relative positional embedding (VRPE). Instead of simply using learnable positional encoding, which could hurt the expression of the relative positions between the two views (see ablation in Sec. 4.3), we use the fixed positional encoding of relative information to cut the uncertainty. Specifically, for elements in the second dimension of $\mathbf{RP}_{1\rightarrow 2}$ , we use a sinusoid form [19] in Transformers to generate view-relative positional embedding $\mathbf{VRPE}_{1\rightarrow 2}$ : + +$$ +\begin{array}{l} \mathbf {V R P E} _ {1 \rightarrow 2} ^ {i} = \left[ \sin \left(\frac {\mathbf {R P} _ {1 \rightarrow 2} ^ {i}}{e ^ {2 \times \frac {1}{D / 1 2}}}\right), \cos \left(\frac {\mathbf {R P} _ {1 \rightarrow 2} ^ {i}}{e ^ {2 \times \frac {1}{D / 1 2}}}\right), \right. \\ \sin \left(\frac {\mathbf {R P} _ {1 \rightarrow 2} ^ {i}}{e ^ {2 \times \frac {2}{D / 1 2}}}\right), \cos \left(\frac {\mathbf {R P} _ {1 \rightarrow 2} ^ {i}}{e ^ {2 \times \frac {2}{D / 1 2}}}\right), \ldots , \\ \left. \sin \left(\frac {\mathbf {R P} _ {1 \rightarrow 2} ^ {i}}{e}\right), \cos \left(\frac {\mathbf {R P} _ {1 \rightarrow 2} ^ {i}}{e}\right)\right] \tag {10} \\ \end{array} +$$ + +where $\mathbf{RP}_{1\rightarrow 2}^i\in \mathbb{R}^{n\times 1}$ , $\mathbf{VRPE}_{1\rightarrow 2}^i\in \mathbb{R}^{n\times (D / 6)}$ ( $1\leq i\leq 6$ ). $e = 10000$ is pre-defined, which is also used in MAE [19]. Finally, we concatenate $\mathbf{VRPE}_{1\rightarrow 2}^i$ for each dimension to obtain $\mathbf{VRPE}_{1\rightarrow 2}\in \mathbb{R}^{n\times D}$ . + +Cross-attention. Empirically, assume the latent representations of view 1, denoted as $\mathbf{H}_1$ , contains intrinsic features and the global characteristics of its source point cloud, then with the combination of the view-relative positional embeddings $\mathbf{VRPE}_{1\rightarrow 2}$ , the view 2 can be reconstructed. Specifically, we employ $\mathbf{VRPE}_{1\rightarrow 2}$ (as $\mathbf{Q}$ ) and $\mathbf{H}_1$ (as $\mathbf{K},\mathbf{V}$ ) in cross-attention mechanisms, as can be formalized: + +$$ +\begin{array}{l} \mathbf {T} _ {2} = \operatorname {A t t n} \left(\mathbf {Q} _ {\mathbf {V R P E} _ {1 \rightarrow 2}}, \mathbf {K} _ {\mathbf {H} _ {1}}, \mathbf {V} _ {\mathbf {H} _ {1}}\right) \\ = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q} _ {\mathbf {V R P E} _ {1 \rightarrow 2}} \mathbf {K} _ {\mathbf {H} _ {1}}}{\sqrt {D}}\right) \mathbf {V} _ {\mathbf {H} _ {1}} \tag {11} \\ \end{array} +$$ + +where $\mathbf{Q}_{\mathbf{VRPE}_{1\rightarrow 2}} = \mathbf{VRPE}_{1\rightarrow 2}\mathbf{W}_{\mathbf{Q}},\mathbf{K}_{\mathbf{H}_1} = \mathbf{H}_1\mathbf{W}_{\mathbf{K}}$ and $\mathbf{V_{H_1}} = \mathbf{H_1W_V}$ , where $\mathbf{W}_{\mathbf{Q}},\mathbf{W}_{\mathbf{K}}$ , and $\mathbf{W_V}$ are learnable parameters. Reconstructing view 1 from view 2 is a siamese condition of this, through the same way as stated above, we can obtain $\mathbf{RL}_{2\to 1},\mathbf{RP}_{2\to 1},\mathbf{VRPE}_{2\to 1}$ and yield $\mathbf{T}_1$ + +Decoding. The obtained $\mathbf{T}_1$ and $\mathbf{T}_2$ are fed into a decoder composed of a few transformer blocks: + +$$ +\mathbf {Z} _ {1} = \operatorname {D e c o d e r} \left(\mathbf {T} _ {1}\right), \quad \mathbf {Z} _ {2} = \operatorname {D e c o d e r} \left(\mathbf {T} _ {2}\right) \tag {12} +$$ + +Finally, similar to [33], we adopt a projection head (composed of a fully connected layer) to predict the input point cloud as follows, where $\mathbf{P}_{pred}^{1},\mathbf{P}_{pred}^{2}\in \mathbb{R}^{n\times k\times 3}$ : + +$$ +\mathbf {P} _ {p r e d} ^ {1} = \operatorname {R e s h a p e} \left(\operatorname {L i n e a r} \left(\mathbf {Z} _ {1}\right)\right), \tag {13} +$$ + +$$ +\mathbf {P} _ {p r e d} ^ {2} = \operatorname {R e s h a p e} \left(\operatorname {L i n e a r} \left(\mathbf {Z} _ {2}\right)\right) \tag {14} +$$ + +# 3.3. Objective function + +The overall objective is to perform cross-reconstruction, which involves predicting view 1 from view 2 and predicting view 2 from view 1. Given the predicted point patches $\mathbf{P}_{pred}^{1}$ and $\mathbf{P}_{pred}^{2}$ and ground truth $\mathbf{P}_1$ and $\mathbf{P}_2$ , we compute the cross-reconstruction loss using the $l_{2}$ Chamfer Distance loss function [12], written as $\mathcal{L}_{cross} = \mathcal{L}_{2\rightarrow 1} + \mathcal{L}_{1\rightarrow 2}$ , where: + +$$ +\begin{array}{l} \mathcal {L} _ {2 \rightarrow 1} = \frac {1}{\left| \mathbf {P} _ {p r e d} ^ {1} \right|} \sum_ {a \in \mathbf {P} _ {p r e d} ^ {1}} \min _ {b \in \mathbf {P} _ {1}} \| a - b \| _ {2} ^ {2} \tag {15} \\ + \frac {1}{| \mathbf {P} _ {1} |} \sum_ {b \in \mathbf {P} _ {1}} \min _ {a \in \mathbf {P} _ {p r e d} ^ {1}} \| a - b \| _ {2} ^ {2} \\ \end{array} +$$ + +where $|\mathbf{P}|$ is the cardinality of the set $\mathbf{P}$ and $\| \cdot \| _2$ is the $l_{2}$ distance. The $\mathcal{L}_{1\to 2}$ follows similarly. + +# 4. Experiments + +This section is organized as follows: First, we present the model architecture and pre-training details. Second, we validate the effectiveness of our pre-trained model on a wide range of downstream tasks, including object classification, few-shot learning, and part segmentation. Finally, ablation studies are conducted to demonstrate the properties and robustness of our proposed Point-PQAE. + +# 4.1. Pre-training setups + +Pre-trained dataset. We use the dataset ShapeNet [6] for pre-training of Point-PQAE following previous studies [9, 33, 37]. ShapeNet [6] consists of about 51,300 clean 3D models, covering 55 common object categories. + +Model structure. We adopt standard Transformer blocks [51] in the autoencoder's backbone where the encoder has 12 Transformer blocks and the decoder has 4 blocks. MLP ratio in Transformer blocks is set to 4. Each Transformer block has 384 hidden dimensions and 6 heads. The positional query module between encoder and decoder only performs cross-attention mechanism once as stated in Sec. 3.2, and the number of heads is set to 6 and remains 384 hidden dimensions. + +Experiment detail. For each input point cloud, we only apply random rotation for pre-training data augmentation. After sampling 1024 points via FPS from the input point cloud, we generate two different partial point clouds by randomly cropping and rotating as illustrated in Sec. 3.1 with + +$r_m = 0.6$ . Both of them are divided into 64 patches with 32 points via FPS and KNN. The Point-PQAE model undergoes pre-training for 300 epochs using an AdamW [29] optimizer with a batch size of 128. The initial learning rate is set to 0.0005, with a weight decay of 0.05, with cosine learning rate decay [28] utilized. We visualize the pretraining results in Fig. 3. It shows Point-PQAE learns cross-knowledge well and is able to generalize excellently to other crop ratios, even though it was pre-trained with $r_m = 0.6$ . + +# 4.2. Transfer learning on downstream tasks + +Transfer protocol. We use three transfer learning protocols for classification tasks following [9]: FULL, MLP-LINEAR, MLP-3, which are detailed in Appendix Sec. 3. + +3D real-world object classification. The transferability of models can be well demonstrated through testing on a 3D real-world object dataset. Therefore, we transfer our pre-trained model to ScanObjectNN [50] for classification, which is one of the most challenging 3D datasets, covering approximately 15,000 real-world objects across 15 categories. We conduct experiments on three variants: OBJBG, OBJ-ONLY, and PB-T50-RS. The results are reported in Tab. 1. Our Point-PQAE surpasses previous single-modal self-supervised methods by a margin under all protocols. In contrast to the most relevant baseline, Point-MAE, PointPQAE significantly outperforms it, showing improvements of $2.4\%$ , $1.7\%$ , and $1.2\%$ . Especially for the MLP-LINEAR and MLP-3 protocols, our Point-PQAE outperforms Point-MAE by $6.7\%$ and $4.4\%$ on average, respectively. The excellent frozen representation evaluation results indicate the superior quality of the representation learned by our Point-PQAE during pre-training. This shows that the cross-reconstruction paradigm learns more robust representations and generalizes better on real-world datasets. + +3D clean object classification. We evaluate the performance of our pre-trained model for object classification using the ModelNet40 dataset [59]. ModelNet40 comprises 12,311 meticulously crafted 3D CAD models, representing 40 distinct object categories. During testing, we adhere to the standard voting method [26] to ensure fair comparisons with previous work such as [7, 33]. Standard random scaling and random translation are applied for data augmentation during training, and the input point clouds exclusively contain coordinate information, without additional normal information provided. The results are presented in Tab. 1. Our Point-PQAE yields better or comparable results compared to previous approaches. + +Few-shot learning. We conduct the few-shot learning classification on the ModelNet40 dataset to show the generalization ability of our method following the protocols in previous studies [33, 41, 53, 65]. The few-shot learning experiments are conducted in the form of "w-way, s-shot" including four parts where $w \in \{5, 10\}$ represents the number of randomly selected classes and $s \in \{10, 20\}$ means + +![](images/a3026c65407e601c8dec37467fd5b25a1e53e55810f8d88097109312c5e0de7b.jpg) +Cropped Rebuilt + +![](images/b25778fcd134e75327060f3ee0fd455e866fc2a7f49f0db55068625858d11c11.jpg) +Input point cloud + +![](images/1bb13258c46e6dfc24d340715ba35b1aae8dd1c3bb9c63fea0947121aa56e438.jpg) +Cropped Rebuilt + +![](images/d82c0c2b96671c70322a59c7b27114f86c1f4a94eb7eae12cdf22d26cf35b299.jpg) +Crop ratio $20\%$ + +![](images/a4e01672b06f5245dac48c0c7739a732b62c245b571f607a57703c76ff8d3769.jpg) + +![](images/85c35790e2ae81bb1eb84686abd7d9d42a107580d4bef8973f3d8248dbbf8352.jpg) +Crop ratio $40\%$ + +![](images/784a93b547adc88ec643569db3f3db5486a1caf47ca964c47af5eb353c504ad8.jpg) +Cropped Rebuilt + +![](images/b006035c5d1b4896f6c459d0193bb6dd1978934f7ef42d8bbce7da017cf7b932.jpg) +Crop ratio $60\%$ + +![](images/15e48a1586f403deefc6e9bfa332e7318339ab3c963de707e25df079dd0a5f7a.jpg) +Cropped Rebuilt + +![](images/1ccbff0f4bfb44b246bc01219a203df5c5d0be55cb500dc6a2ed895b8a9f959c.jpg) +Crop ratio $80\%$ +Figure 3. Cross-reconstruction results on ShapeNet. The arrow points from source point clouds to cross-reconstruction results. PointPQAE generalizes well to other crop ratios though with minimum crop ratio $r_m = 0.6$ when pre-training. + +the number of randomly selected samples for each $w$ . Each part is conducted with 10 independent trials. The mean accuracy and standard deviation are reported in Tab. 2. Our Point-PQAE demonstrates superior performance compared to previous methods in few-shot learning, even surpassing cross-modal methods that utilize strong pre-trained teachers such as ACT and ReCon under FULL protocol. Specifically, Point-PQAE achieves $2.7\%$ , $1.1\%$ , $4.5\%$ , $1.9\%$ improvement under MLP-LINEAR protocol and shows enhanced performance over the self-construction method Point-MAE. + +Part segmentation. We conduct part segmentation experiments on the ShapeNetPart [64] to validate the effectiveness of our Point-PQAE. The ShapeNetPart contains 16,881 objects covering 16 categories. We sample 2,048 points from each input point cloud, following previous work [33, 65] and divide each cloud into 128 point patches. We use the same segmentation head and utilize learned features from the 4th, 8th, and 12th layers of the Transformer block as in Point-MAE [33]. We concatenate three levels of features. Then average pooling, max pooling, and upsampling are utilized to generate features for each point and an MLP is applied for label prediction. The results are reported in Tab. 3a. The results are reported in Tab. 3a, which show that Point-PQAE achieves a comparable Inst.mIoU to the previous method, achieving $84.6\%$ in Cls.mIoU and improving the from-scratch baseline by $1.4\%$ . + +3D scene segmentation. Semantic segmentation on large-scale 3D scenes is challenging, requiring models possessing a deep comprehension of semantics and intricate local geometric relationships. We report semantic segmentation results on the S3DIS dataset [3] in Tab. 3b. Our Point-PQAE improves the from-scratch baseline by $2.0\%$ and $1.4\%$ , and outperforms the self-reconstruction method + +Point-MAE by $0.7\%$ and $0.6\%$ in mAcc and mIoU. + +# 4.3. Ablation study + +Experiments are conducted to show the properties of our Point-PQAE. We report the classification results on three variants of ScanObjectNN. The experiment settings are aligned with Sec. 4.2. + +View-relative positional embedding. To analyze the effect of view-relative positional embedding in the positional query block, we conduct a series of experiments using different types of positional embeddings. We primarily consider two types of embeddings: Sinusoid (as discussed in Sec. 3.2) and Learnable, along with a None group. Learnable embedding refers to utilizing the relative position information $\mathbf{RP}_{1\rightarrow 2}$ and $\mathbf{RP}_{2\rightarrow 1}$ as input and employing an MLP composed of two linear layers with an activation function to map the 6-dimensional input to a D-dimensional output. The None group involves using randomly assigned embeddings to assess the effectiveness of VRPE design. Tab. 4a displays the results of different positional embeddings. In the None group, the model cannot learn the relative positional information across two views, leading to a significant drop in accuracy. In addition to the type of VRPE, we also experiment with Absolute Positional Embedding (APE) to compare its performance with our proposed VRPE. The APE is unsuitable for our framework since it requires a shared coordinate system for both views, making it incompatible with the normalization and rotation used in our method. We incorporate APE into our Point-PQAE by removing these operations. The results in Tab. 4a show APE performs worse and support our claim. + +Data augmentation. After applying the random crop mechanism to the point cloud, we further perform data augmentation on the cropped views to generate new views. + +Table 1. Classification accuracy (%) on ScanObjectNN and ModelNet40. The inference model parameters #P (M) are reported. Three variants are evaluated on ScanObjectNN and the accuracy obtained on ModelNet40 is reported for both 1K and 8K points. We compare methods using the $\bullet$ plain Transformer architectures, e.g. Point-MAE [33], $\circ$ hierarchical Transformer architectures and $\circ$ dedicated architectures for 3D. The dagger(†) denotes the baseline results reported from ReCon [37] which aligns augmentation with us and ReCon. + +
Methods#PScanObjectNNModelNet40
OBJ_BGOBJ_ONLYPB_T50_RS1K P8K P
Supervised Learning Only
○PointNet [34]3.573.379.268.089.290.8
○PointNet++ [35]1.582.384.377.990.791.9
○DGCNN [54]1.882.886.278.192.9-
○PointCNN [24]0.686.185.578.592.2-
○SimpleView [13]---80.5±0.393.9-
○PCT [15]2.88---93.2-
with Self-Supervised Representation Learning (FULL)
•Transformer [51]22.183.084.079.191.491.8
•Point-BERT [65]22.187.488.183.093.293.8
•MaskPoint [25]-89.388.184.393.8-
•Point-MAE [33]22.190.088.385.293.894.0
○Point-M2AE [66]15.391.288.986.494.0-
•PointDif [74]22.193.391.987.6--
•Point-MAE† [33]22.192.691.988.493.894.0
•Point-PQAE22.195.093.689.694.094.3
Methods using cross-modal information and teacher models
•Joint-MAE [16]-90.988.986.194.0-
•TAP [55]22.190.489.585.794.0-
•ACT [9]22.193.391.988.293.794.0
○PointMLP+ULIP [63]---89.494.594.7
○I2P-MAE [67]15.394.291.690.194.1
•ReCon [37]43.695.293.690.694.594.7
with Self-Supervised Representation Learning (MLP-LINEAR)
•Point-MAE† [33]22.182.8±0.383.2±0.274.1±0.290.2±0.190.7±0.1
•Point-PQAE22.189.3±0.390.2±0.480.8±0.192.0±0.292.2±0.1
Methods using cross-modal information and teacher models
•ACT [9]22.185.2±0.885.8±0.276.3±0.391.4±0.291.8±0.2
•ReCon [37]43.689.5±0.289.7±0.281.4±0.192.5±0.292.7±0.1
with Self-Supervised Representation Learning (MLP-3)
•Point-MAE† [33]22.185.8±0.385.5±0.280.4±0.291.3±0.291.7±0.2
•Point-PQAE22.190.7±0.290.9±0.283.3±0.192.8±0.192.9±0.1
Methods using cross-modal information and teacher models
•ACT [9]22.187.1±0.287.9±0.481.5±0.292.7±0.293.0±0.1
•ReCon [37]43.690.6±0.290.7±0.383.8±0.493.0±0.193.4±0.1
+ +Data augmentation is crucial for generating diverse and decoupled views, so we conducted ablation studies on different augmentation methods. The results in Tab. 4b indicate that rotation performs best. + +Effectiveness of the crop mechanism. Random crop has been shown to be highly effective in 2D self-supervised learning (SSL), particularly for contrastive learning [4, 8, 14], and is also vital for generative learning [19]. However, its potential in 3D SSL remains largely unexplored. Even contrastive learning methods for point cloud understanding have not explored the power of it [2, 32]. To address this gap, we first introduce a custom-designed random crop mechanism for point cloud data, using it as the + +foundation for generating decoupled views. Without the crop augmentation, the two views will contain exactly the same shape though with normalization and rotation applied. The model would only need to infer the augmentation rather than the inter-view relationship, which means the views are not decoupled. Conversely, given two isolated cropped point clouds, followed by min-max normalization and rotation for further decoupling, there remain overlapping and non-overlapping parts, requiring the model to effectively encode the intra-view parts in order to infer the unknown inter-view points in our Point-PQAE. However, directly applying random crop as an additional augmentation to the self-reconstruction method (Point-MAE) does + +Table 2. Few-shot classification results on ModelNet40. 10 independent trials are conducted in each experimental setting. The mean accuracy (\%) and standard deviation are reported for each setting. The dagger( $\dagger$ ) denotes the baseline results reported from ReCon [37] which aligns augmentation with us and ReCon. + +
Methods5-way10-way
10-shot20-shot10-shot20-shot
○DGCNN [54]31.6±2.840.8±4.619.9±2.116.9±1.5
○OcCo [52]90.6±2.892.5±1.982.9±1.386.5±2.2
with Self-Supervised Representation Learning (FULL)
•Transformer [51]87.8±5.293.3±4.384.6±5.589.4±6.3
•Point-BERT [65]94.6±3.196.3±2.791.0±5.492.7±5.1
•MaskPoint [25]95.0±3.797.2±1.791.4±4.093.4±3.5
•Point-MAE [33]96.3±2.597.8±1.892.6±4.195.0±3.0
○Point-M2AE [66]96.8±1.898.3±1.492.3±4.595.0±3.0
•Point-MAE† [66]96.4±2.897.8±2.092.5±4.495.2±3.9
•Point-PQAE96.9±3.298.9±1.094.1±4.296.3±2.7
Methods using cross-modal information and teacher models
•Joint-MAE [16]96.7±2.297.9±1.992.6±3.795.1±2.6
•TAP [55]97.3±1.897.8±1.993.1±2.695.8±1.0
•ACT [9]96.8±2.398.0±1.493.3±4.095.6±2.8
○I2P-MAE [67]97.0±1.898.3±1.392.6±5.095.5±3.0
•ReCon [37]97.3±1.998.9±1.293.3±3.995.8±3.0
with Self-Supervised Representation Learning (MLP-LINEAR)
•Point-MAE† [66]91.1±5.691.7±4.083.5±6.189.7±4.1
•Point-PQAE93.0±4.696.8±1.989.0±5.293.5±3.8
with Self-Supervised Representation Learning (MLP-3)
•Point-MAE† [66]95.0±2.896.7±2.490.6±4.793.8±5.0
•Point-PQAE95.3±3.498.2±1.892.0±3.894.7±3.5
+ +Table 3. Segmentation results. Cls.mIoU (\%) and Inst.mIoU (\%) refer to Mean intersection over union for all classes and all instances, respectively. mAcc (\%) refers to mean accuracy. +(a) Part segmentation on ShapeNetPart. + +
MethodCls.mIoU (%)Inst.mIoU (%)
Supervised
PointNet [34]80.483.7
PointNet++ [35]81.985.1
DGCNN [54]82.385.2
Self-supervised
Transformer [51]83.484.7
CrossPoint [2]-85.5
Point-BERT [65]84.185.6
Point-MAE [33]84.286.1
Point-PQAE (Ours)84.686.1
Cross-modal
ACT [9]84.786.1
ReCon [37]84.886.4
+ +(b) Semantic segmentation results on S3DIS Area 5. + +
MethodmAcc (%)mIoU (%)
Supervised
PointNet [34]49.041.1
PointNet++ [35]67.153.5
Self-supervised
Transformer [51]68.660.0
Point-MAE [33]69.960.8
Point-PQAE (Ours)70.661.4
Cross-modal
ACT [9]71.161.2
+ +not improve the performance much, as shown in Tab. 4c. In other words, it means our Point-PQAE fits crop mechanism better by harmoniously incorporating it into the decoupled views generation process. The experiment results in Tab. 4c demonstrate the critical role of random crop in our method. We set minimum crop ratio $r_m = 1.0$ when removing the crop in our Point-PQAE. + +Analysis on decoupled views generation. While crop plays an important role in our method, relying solely on the random crop mechanism to generate new views would degenerate cross-reconstruction to simply reconstructing one part from another. This is similar to the block mask self-reconstruction used in Point-MAE [33], where visible blocks are used to reconstruct masked blocks, which is also trivial for pre-training. The only difference between them + +Table 4. Ablation study with Point-PQAE pre-training on ShapeNet. The classification results by accuracy (\%) on three variants of ScanObjectNN are reported. Default settings are marked in gray. +(a) View-Relative positional embedding + +
Positional EmbeddingOBJ_BGOBJ_ONLYLYPB_T50_RS
None84.585.979.3
APE (sinusoid)92.391.087.7
VRPE (learnable)93.493.189.1
VRPE (sinusoid Eq. (10))95.093.689.6
+ +(b) Data augmentation + +
Data augmentationOBJ_BGOBJ_ONLYPB_T50_RS
jitter93.391.487.3
scale93.391.788.0
rotation95.093.689.6
scale&translate92.891.788.1
rotation+scale&translate93.892.989.1
+ +(c) 3D random crop mechanism + +
MethodcropOBJ_BGOBJ_ONLYPB_T50_RS
Point-MAE✘92.691.988.4
Point-MAE✓92.992.188.8
Point-PQAE✘92.992.387.7
Point-PQAE✓95.093.689.6
+ +(d) Effects of augmentations after random crop + +
norm.rotationOBJ_BGOBJ_ONLYPB_T50_RS
XX92.892.188.2
✓X93.392.687.9
X✓93.592.688.8
✓✓95.093.689.6
+ +is that the cropped parts can overlap somewhat, whereas the block mask does not. What makes the cross-reconstruction different from block mask self-reconstruction are the subsequent augmentations including normalization and rotation that decouple the cropped parts. By normalizing cropped point clouds centered on the geometric centers, the coordinate systems of the two point clouds become isolated and independent from each other, with rotation further amplifying the variance between the two views, thus enabling the generation of two decoupled views. The importance of the augmentations applied after the random crop can be validated through experiments, as shown in Tab. 4d. + +# 5. Conclusion + +In this paper, we propose a novel cross-reconstruction generative framework for self-supervised learning on 3D point clouds, called Point-PQAE. In contrast to well-studied self-reconstruction schemes, Point-PQAE reconstructs one cropped point cloud from another decoupled point cloud. To supply sufficient information for cross-reconstruction, we further propose a 3D view-relative positional embedding and a corresponding position-aware query module. Compared to self-reconstruction, the cross-reconstruction task, carefully designed by us, is much more challenging for pre-training, promoting the learning of richer semantic representations during pre-training. This enables our proposed Point-PQAE to outperform previous single-modal self-reconstruction methods by a margin and to perform on par with, or better than cross-modal methods. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 1 +[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 1, 2, 7, 8 +[3] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 6 +[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3, 7 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1 +[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5 +[7] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. Advances in Neural Information Processing Systems, 36, 2024. 1, 5 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 2, 3, 7 +[9] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022. 1, 2, 5, 7, 8 +[10] Bi'an Du, Xiang Gao, Wei Hu, and Xin Li. Self-contrastive learning with hard negative sampling for self-supervised point cloud learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3133–3142, 2021. 1 +[11] Ikenna Enebuse, Mathias Foo, Babul Salam Ksm Kader Ibrahim, Hafiz Ahmed, Fhon Supmak, and Odongo Steven Eyobu. A comparative review of hand-eye calibration techniques for vision guided robots. IEEE Access, 9:113143-113155, 2021. 1 +[12] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 5 + +[13] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In International Conference on Machine Learning, pages 3809-3820. PMLR, 2021. 7 +[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 2, 7 +[15] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 1, 7 +[16] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzhi Li, and Pheng-Ann Heng. Joint-mae: 2d-3d joint masked autoencoders for 3d point cloud pre-training. arXiv preprint arXiv:2302.14007, 2023. 7, 8 +[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 3 +[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 2, 4, 7 +[20] Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6535-6545, 2021. 1 +[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 3 +[22] Gang Li, Heliang Zheng, Daqing Liu, Chaoyue Wang, Bing Su, and Changwen Zheng. Semmae: Semantic-guided masking for learning masked autoencoders. NeurIPS, 2022. 2 +[23] Jiaxin Li, Ben M Chen, and Gim Hee Lee. So-net: Self-organizing network for point cloud analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9397–9406, 2018. 1 +[24] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 1, 7 +[25] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 1, 2, 7, 8 +[26] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point + +cloud analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8895-8904, 2019. 5 +[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2 +[28] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5 +[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[30] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 1 +[31] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1 +[32] Bo Pang, Hongchi Xia, and Cewu Lu. Unsupervised 3d point cloud representation learning by triangle constrained contrast for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5229-5239, 2023. 7 +[33] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 1, 2, 4, 5, 6, 7, 8 +[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 4, 7, 8 +[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 7, 8 +[36] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1 +[37] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. arXiv preprint arXiv:2302.02318, 2023. 1, 3, 5, 7, 8 +[38] Rui Qian, Xin Lai, and Xirong Li. 3d object detection for autonomous driving: A survey. Pattern Recognition, 130: 108796, 2022. 1 +[39] Aditya Sanghi. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX 16, pages 626-642. Springer, 2020. 1 +[40] Oleg Yu Sergiyenko and Vera V Tyrsa. 3d optical machine vision sensors with intelligent data management for robotic + +swarm navigation improvement. IEEE Sensors Journal, 21 (10):11262-11274, 2020. 1 +[41] Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. Advances in Neural Information Processing Systems, 33:7212-7221, 2020. 5 +[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3 +[43] Haoru Tan, Chuang Wang, Sitong Wu, Tieqiang Wang, Xu-Yao Zhang, and Cheng-Lin Liu. Proxy graph matching with proximal matching networks. In The Annual AAAI Conference on Artificial Intelligence (AAAI), 2021. 1 +[44] Haoru Tan, Sitong Wu, Fei Du, Yukang Chen, Zhibin Wang, Fan Wang, and Xiaojuan Qi. Data pruning via moving-one-sample-out. In Neural Information Processing Systems (NeurIPS), 2023. +[45] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2023. +[46] Haoru Tan, Chuang Wang, Sitong Wu, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Ensemble quadratic assignment network for graph matching. International Journal of Computer Vision (IJCV), 2024. +[47] Haoru Tan, Sitong Wu, Zhuotao Tian, Yukang Chen, Xiaojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity consistency for vision-language pre-training. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. +[48] Haoru Tan, Sitong Wu, Wei Huang, Shizhen Zhao, and Xiaojuan Qi. Data pruning by information maximization. In International Conference on Learning Representations (ICLR), 2025. +[49] Haoru Tan, Sitong Wu, Bo Zhao, Zeke Xie, and XIAOJUAN QI. Diff-in: Data influence estimation with differential approximation, 2025. 1 +[50] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 5 +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5, 7, 8 +[52] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 1, 8 +[53] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 5 +[54] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic + +graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 1, 7, 8 +[55] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. Take-a-photo: 3d-to-2d generative pre-training of point cloud models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5640–5650, 2023. 7, 8 +[56] Hai Wu, Ruifei He, Haoru Tan, Xiaojuan Qi, and Kaibin Huang. Vertical layering of quantized neural networks for heterogeneous inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[57] Sitong Wu, Haoru Tan, Yukang Chen, Shaofeng Zhang, Jingyao Li, Bei Yu, Xiaojuan Qi, and Jiaya Jia. Mixture-of-scores: Robust image-text data quality score via three lines of code. In International Conference on Computer Vision (ICCV), 2025. 1 +[58] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. arXiv preprint arXiv:2312.10035, 2023. 1 +[59] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 5 +[60] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 1, 2 +[61] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2 +[62] Mingye Xu, Zhipeng Zhou, Yali Wang, and Yu Qiao. Towards robustness and generalization of point cloud representation: A geometry coding method and a large-scale object-level dataset. Computational Visual Media, 10(1):27-43, 2024. 1 +[63] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1179-1189, 2023. 7 +[64] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6 +[65] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 4, 5, 6, 7, 8 +[66] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: + +multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 1, 2, 7, 8 +[67] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 7, 8 +[68] Shaofeng Zhang, Qiang Zhou, Sitong Wu, Haoru Tan, Zhibin Wang, Jina Huang, and Junchi Yan. Cr2pq: Continuous relative rotary positional query for dense visual representation learning. In The Thirteenth International Conference on Learning Representations. 1 +[69] Shaofeng Zhang, Meng Liu, and Junchi Yan. The diversified ensemble neural network. Advances in Neural Information Processing Systems, 33:16001-16011, 2020. +[70] Shaofeng Zhang, Feng Zhu, Junchi Yan, Rui Zhao, and Xiaokang Yang. Zero-cl: Instance and feature decorrelation for negative-free symmetric contrastive learning. In International Conference on Learning Representations, 2021. +[71] Shaofeng Zhang, Meng Liu, Junchi Yan, Hengrui Zhang, Lingxiao Huang, Xiaokang Yang, and Pinyin Lu. M-mix: Generating hard negatives via multi-sample mixing for contrastive learning. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, pages 2461-2470, 2022. 1 +[72] Shaofeng Zhang, Feng Zhu, Rui Zhao, and Junchi Yan. Contextual image masking modeling via synergized contrasting without view augmentation for faster and better visual pretraining. In ICLR, 2023. 2 +[73] Xiangdong Zhang, Shaofeng Zhang, and Junchi Yan. Pcmae: Learning to predict centers for point masked autoencoders. Advances in Neural Information Processing Systems, 37:80303-80327, 2024. 1 +[74] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22935-22945, 2024. 7 +[75] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2 +[76] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1 \ No newline at end of file diff --git a/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/images.zip b/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d898ee4bdb42f5bbc89aa70b51c76350851a7df5 --- /dev/null +++ b/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afb15afbda54a96598fd8111d968ff633c6f36e171e269f91b41ec5bc9e50aad +size 634070 diff --git a/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/layout.json b/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..24f9367cfe765db14e3a905a099f18fa89c10c43 --- /dev/null +++ b/2025/Towards More Diverse and Challenging Pre-training for Point Cloud Learning_ Self-Supervised Cross Reconstruction with Decoupled Views/layout.json @@ -0,0 +1,10636 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 550, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 550, + 140 + ], + "type": "text", + "content": "Towards More Diverse and Challenging Pre-training for Point Cloud Learning: Self-Supervised Cross Reconstruction with Decoupled Views" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 160, + 428, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 160, + 428, + 175 + ], + "spans": [ + { + "bbox": [ + 181, + 160, + 428, + 175 + ], + "type": "text", + "content": "Xiangdong Zhang*, Shaofeng Zhang*, Junchi Yan‡" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 176, + 413, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 176, + 413, + 190 + ], + "spans": [ + { + "bbox": [ + 196, + 176, + 413, + 190 + ], + "type": "text", + "content": "School of AI, Shanghai Jiao Tong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 121, + 190, + 487, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 190, + 487, + 204 + ], + "spans": [ + { + "bbox": [ + 121, + 190, + 487, + 204 + ], + "type": "text", + "content": "{zhangxiangdong, sherrylone, yanjunchi}@sjtu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 256, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 568 + ], + "type": "text", + "content": "Point cloud learning, especially in a self-supervised way without manual labels, has gained growing attention in both vision and learning communities due to its potential utility in a wide range of applications. Most existing generative approaches for point cloud self-supervised learning focus on recovering masked points from visible ones within a single view. Recognizing that a two-view pre-training paradigm inherently introduces greater diversity and variance, it may thus enable more challenging and informative pre-training. Inspired by this, we explore the potential of two-view learning in this domain. In this paper, we propose Point-PQAE, a cross-reconstruction generative paradigm that first generates two decoupled point clouds/views and then reconstructs one from the other. To achieve this goal, we develop a crop mechanism for point cloud view generation for the first time and further propose a novel positional encoding to represent the 3D relative position between the two decoupled views. The cross-reconstruction significantly increases the difficulty of pre-training compared to self-reconstruction, which enables our method to surpass previous single-modal self-reconstruction methods in 3D self-supervised learning. Specifically, it outperforms the self-reconstruction baseline (Point-MAE) by " + }, + { + "bbox": [ + 55, + 256, + 297, + 568 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 55, + 256, + 297, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 256, + 297, + 568 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 55, + 256, + 297, + 568 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 256, + 297, + 568 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 55, + 256, + 297, + 568 + ], + "type": "text", + "content": " in three variants of ScanObjectNN with the MLP-LINEAR evaluation protocol. The code is available at https://github.com/aHapBean/Point-PQAE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 135, + 607 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 135, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 135, + 607 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 614, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 296, + 676 + ], + "type": "text", + "content": "3D vision is gaining increasing attention for its wide applications such as autonomous driving [38] and robotics [11, 40], owing to its ability to understand the human world. The point cloud is the most popular representation form of data in 3D vision and many analyses based on point cloud are" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 232, + 421, + 340 + ], + "blocks": [ + { + "bbox": [ + 328, + 232, + 421, + 340 + ], + "lines": [ + { + "bbox": [ + 328, + 232, + 421, + 340 + ], + "spans": [ + { + "bbox": [ + 328, + 232, + 421, + 340 + ], + "type": "image", + "image_path": "71fc8b35306dce4462853f98b5b0cc64850f75583a1d84e87d9db6913c22d847.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 352, + 409, + 360 + ], + "lines": [ + { + "bbox": [ + 354, + 352, + 409, + 360 + ], + "spans": [ + { + "bbox": [ + 354, + 352, + 409, + 360 + ], + "type": "text", + "content": "Self-Reconstruction" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 439, + 232, + 539, + 350 + ], + "blocks": [ + { + "bbox": [ + 439, + 232, + 539, + 350 + ], + "lines": [ + { + "bbox": [ + 439, + 232, + 539, + 350 + ], + "spans": [ + { + "bbox": [ + 439, + 232, + 539, + 350 + ], + "type": "image", + "image_path": "12e4a292f90bab83a13a9f8aa0c2fe8a949e9c3cfd6e2f3a13bbaceb4845174a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 352, + 531, + 361 + ], + "lines": [ + { + "bbox": [ + 454, + 352, + 531, + 361 + ], + "spans": [ + { + "bbox": [ + 454, + 352, + 531, + 361 + ], + "type": "text", + "content": "Cross-Reconstruction (Ours)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 364, + 555, + 452 + ], + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 452 + ], + "type": "text", + "content": "Figure 1. Comparison between self-reconstruction (Point-MAE [33] and other methods based on it) and cross-reconstruction (Ours) paradigms. In the self-reconstruction paradigm, part of the input data is masked, and the autoencoder is trained to recover the missing patches. In contrast, the cross-reconstruction paradigm generates decoupled views first, with one view leveraging cross-view information to reconstruct the other, using an autoencoder backbone." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 460, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 555, + 604 + ], + "type": "text", + "content": "explored to solve various tasks such as object classification [15, 24, 34, 35, 54, 62], object detection [30, 36, 76] and segmentation [15, 24, 34, 54, 58], they typically necessitate fully-supervised training from scratch. Compared to 2D data, point cloud generally requires more expensive and labor-intensive efforts to collect and annotate, which hinders the development of the fully supervised 3D representation learning method. Self-supervised learning is one of the predominant approaches to address this issue and has been proven to be effective in 2D vision [5, 8, 14, 18, 19, 68-71, 75]. Inspired by this, self-supervised learning has been widely studied in the 3D field in recent years." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Self-supervised learning in 3D can be mainly divided into two categories, i.e., generative methods [1, 7, 23, 25, 33, 52, 66, 73] and contrastive methods [2, 9, 10, 20, 37, 39, 60]. Similar to 2D, contrastive methods in 3D aim to learn global-discriminative information by maximizing the mutual information across views [2, 9, 60] with intra-/inter-modal information [43-49, 56, 57]. PointContrast [60] is the first to learn transformation invariance across views through the extension of InfoNCE objective [31, 60]. Dif" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 296, + 713 + ], + "type": "text", + "content": "*Equal contribution. ‡Corresponding author. This paper is in part supported by Shanghai Municipal Science and Technology Major Project, China, under grant No. 2021SHZDZX0102." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28696" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "ferent from contrastive methods, the core idea of the generative methods [25, 33, 66] in 3D is inspired by MAE [19], which masks the input point cloud and uses the visible part to reconstruct the masked part. Point-MAE [33] is one of the conventional methods, which learns point-wise information through the self-reconstruction mechanism. Specifically, Point-MAE masks a high ratio of the point cloud and reconstructs the masked points through the visible part. Point-M2AE [66] proposes a hierarchical Transformer [27] with a corresponding masking strategy based on it, capturing both fine-grained and high-level semantics of 3D shapes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 218, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 296, + 625 + ], + "type": "text", + "content": "Existing generative methods mostly focus on the masked reconstruction in a single view, where single-view learning has been proven less difficult and informative than two views [8, 14] since two views would bring more variance than a single view. To bring the benefits of the two-view learning paradigm to the generative learning field, one straightforward assumption is adopting the two-view learning paradigm in the 3D generative pertaining tasks to increase the difficulty of point-wise reconstruction. However, unlike 2D vision data, constructing two views in the 3D point cloud is more challenging, as point cloud data is less structured. Besides, cross-reconstruction (reconstructing one view from the other view) is much more difficult than self-reconstruction (reconstructing the masked single view), which involves learning both the inter-positional relation between two views and the inner spatial information within the view. The differences between these two paradigms can be seen in Fig. 1. Aimed at solving the above issues, we propose Point-PQAE (Position Query based AutoEncoder for Point Cloud), a novel framework aimed at addressing two primary challenges: i) how to construct two views in point cloud data, and ii) how to perform cross-reconstruction between two views. As shown in Fig. 2, for each point cloud, we first apply a custom-designed point cloud crop mechanism, which randomly selects two points as the central points of two views and records their positions. Following this, for each central point, the nearest points according to a specified ratio are incorporated into the view. Subsequently, we normalize the cropped point cloud and apply random augmentation (e.g., rotation) to each point cloud to achieve view decoupling. Then we calculate the view-relative positional embedding (VRPE) for these two decoupled views. Finally, we take the VRPE and one view of the point cloud as input to predict the point-wise input of the other view." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 627, + 195, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 195, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 195, + 639 + ], + "type": "text", + "content": "The highlights of this paper are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 714 + ], + "type": "text", + "content": "i) New generative framework: We propose PointPQAE, the first framework that successfully brings cross-reconstruction to 3D generative field for point cloud self-supervised learning, with three modules: 1) decoupled views generation, 2) VRPE generation, and 3) positional query block. To our knowledge, we are the first to design" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "and apply crop mechanism to point cloud self-supervised learning. Our framework, Point-PQAE, enables more informative and challenging self-supervised pre-training compared to existing self-reconstruction methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 121, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 215 + ], + "type": "text", + "content": "ii) The first relative position-aware query module for point cloud: We introduce a positional query block after the encoder to capture relative position information between decoupled views. The module applies cross-attention between the hidden representations of one view and the VRPE, enabling it to predict the other view. Our module can be easily plugged into various related tasks, e.g., distillation as stated in Appendix Sec. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 216, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 300 + ], + "type": "text", + "content": "iii) Strong performance: Point-PQAE achieves new state-of-the-art performance on several benchmarks, e.g., outperforming previous methods on few-shot learning and achieving average improvements of " + }, + { + "bbox": [ + 313, + 216, + 553, + 300 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 313, + 216, + 553, + 300 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 216, + 553, + 300 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 313, + 216, + 553, + 300 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 216, + 553, + 300 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 313, + 216, + 553, + 300 + ], + "type": "text", + "content": " over the baseline (Point-MAE) on ScanObjectNN classification across three evaluation protocols (FULL, MLP-LINEAR, MLP-3), respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 311, + 403, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 403, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 403, + 323 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 330, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 553, + 545 + ], + "type": "text", + "content": "SSL in 2D vision. Self-supervised learning (SSL) in 2D vision is broadly categorized into contrastive-based and masked-image-modeling (MIM) approaches. Contrastive methods learn discriminative features by creating two augmented views of an image and enforcing invariance between them. For instance, SimCLR [8] treats augmented views from the same image as positive pairs and those from different images as negatives. MoCo [18] utilizes a memory bank to store these negative samples and further proposes a momentum-updated mechanism to avoid model collapse. MIM-based methods mask portions of the input and task models with reconstructing the masked regions, encouraging fine-grained understanding. MAE [19] employs an encoder-decoder architecture, discarding masked patches before reconstruction, while SimMIM [61] replaces masked regions with a learnable parameter. Subsequent work explores variations such as masking strategies [22, 72] and feature-level reconstruction [75]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "content": "Representation learning for point clouds. It is a challenging task due to its irregular and sparse nature when compared to other modalities such as images which are well-structured and uniform. PointContrast [60] firstly borrows the idea of contrastive learning in 2D and performs point-level invariant mapping on two transformed views of the given point cloud. On the basis of PointContrast, Cross-Point [2] further proposes inter and intra-modal contrastive objectives to enhance the global-discriminative capability. To better model multi-modal data, ACT [9] employs cross-modal auto-encoders as teacher models to acquire knowledge from other modalities. In contrast to contrastive objectives, Point-MAE [33] proposes to randomly mask points, and reconstruct the masked part through the visible part." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28697" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 71, + 541, + 300 + ], + "blocks": [ + { + "bbox": [ + 69, + 71, + 541, + 300 + ], + "lines": [ + { + "bbox": [ + 69, + 71, + 541, + 300 + ], + "spans": [ + { + "bbox": [ + 69, + 71, + 541, + 300 + ], + "type": "image", + "image_path": "204b1936a68fac0d3f61ec2f1b08677b5bd4ca3b8b7f2f670b75507f633993b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 309, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 54, + 309, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 309, + 555, + 354 + ], + "type": "text", + "content": "Figure 2. Pipeline of Point-PQAE. The input point cloud is randomly cropped followed by the rotation to generate views. Then, we feed the views to the patch embedding layer and the transformer encoder, followed by the proposed positional query block. The View-Relative Positional Embedding (VRPE) is obtained through the VRPE module by extracting relative geometric relations, which is taken as \"Query\" in the cross-attention mechanism. The queried hidden embeddings are then fed to the decoder to predict the inputs of the other view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 359, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 296, + 492 + ], + "type": "text", + "content": "Following Point-MAE and CrossPoint, ReCon [37] harmoniously incorporates the generative method Point-MAE and the contrastive method. Almost all the existing generative methods in 3D rely on self-reconstruction. To this end, we propose a new cross-reconstruction generative framework, named Point-PQAE, which increases pre-training difficulty by introducing cross-view reconstruction, bringing more variance to the training samples compared to single-view reconstruction. This requires the model to effectively learn both intra-view and inter-view knowledge, thereby enabling the model to learn more semantic representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 501, + 209, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 209, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 209, + 514 + ], + "type": "text", + "content": "3. The proposed Point-PQAE" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 239, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 239, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 239, + 533 + ], + "type": "text", + "content": "3.1. Views and embeddings generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "Decoupled views generation. Random Image Crop has been widely used in both the supervised domain [17, 21, 42] and the self-supervised domain [4, 8, 18], and it has been proved to be effective in practice. This method typically involves randomly selecting a rectangular region from an image and cropping it. However, to the best of our knowledge, no corresponding crop mechanism has been proposed for 3D self-supervised representation learning. Instead of directly applying the 2D crop algorithm to 3D (by randomly selecting a cube in space), which may lead to inconsistency, as the points within cubes of the same size can vary significantly due to the diverse distribution of points in 3D, we design in a smarter way: For a randomly selected crop ratio, we first select a center point randomly and then choose the corresponding number of points in the point cloud that" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 360, + 436, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 436, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 436, + 371 + ], + "type": "text", + "content": "are closest to the center point." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": "Formally, given an input point cloud with " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " points " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{p \\times 3}" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " and a pre-defined minimum crop ratio " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "r_m" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "0 < r_m < 1" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": ", we first randomly select the crop ratios " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " uniformly from " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "[r_m, 1]" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": ". Then, we randomly select two center points " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1, \\mathbf{C}_2 \\in \\mathbb{R}^{1 \\times 3}" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " and get " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_2" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " by expanding nearest " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "r_1 \\times p" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "r_2 \\times p" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " points to the center point " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_1" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_2" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " respectively in " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": ". Meanwhile, we record the absolute coordinates of cropped point clouds' geometric centers, which are denoted as " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_1 \\in \\mathbb{R}^{1 \\times 3}" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_2 \\in \\mathbb{R}^{1 \\times 3}" + }, + { + "bbox": [ + 313, + 372, + 555, + 492 + ], + "type": "text", + "content": ", respectively. The process can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 367, + 502, + 555, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 502, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 367, + 502, + 555, + 515 + ], + "type": "interline_equation", + "content": "r _ {1}, r _ {2} = \\text {U n i f o r m S a m p l e} ([ r _ {m}, 1 ]) \\tag {1}", + "image_path": "9e511843bb0719a4d2a83b543b394b81d39af644bd4e12b6a3c9f8a747e91d6c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 517, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 517, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 359, + 517, + 553, + 529 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {1}, \\mathbf {C} _ {2} = \\operatorname {R a n d o m S e l e c t} (\\mathbf {X}) \\tag {2}", + "image_path": "47d87f1b1e5a6abbeab48aed47875b8883a154652ee2dbcf2678a715678bd456.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 377, + 532, + 553, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 532, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 377, + 532, + 553, + 544 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {1} = \\mathrm {N} - \\text {N e a r e s t} (\\mathbf {C} _ {\\mathbf {1}}, r _ {1}, \\mathbf {X}) \\tag {3}", + "image_path": "a343c4a88a88d00386cffb37a9e2b8837f1a9233536fe82c46f49ca734fee102.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 377, + 547, + 553, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 547, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 377, + 547, + 553, + 559 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {2} = \\mathrm {N} - \\operatorname {N e a r e s t} \\left(\\mathbf {C} _ {2}, r _ {2}, \\mathbf {X}\\right) \\tag {4}", + "image_path": "b17dc0d1cc6873ae27e2caffe3cf6bc458a89f1ac4dad939a87f0fe3a5572030.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1 \\in \\mathbb{R}^{r_1p \\times 3}" + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_2 \\in \\mathbb{R}^{r_2p \\times 3}" + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "text", + "content": " mean the two cropped views. The " + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1" + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_2" + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "text", + "content": " are normalized using the min-max normalization, centered on " + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_1" + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_2" + }, + { + "bbox": [ + 313, + 569, + 555, + 689 + ], + "type": "text", + "content": ", respectively. Subsequently, we apply additional augmentation to further enhance the variance between two views. We apply random rotations to the cropped views. The distinct normalization centers, along with random rotation, effectively isolate the coordinate systems of the two views and alter their fixed relative relationships. This process decouples the two point clouds, resulting in two independent views." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Point patches generation. We divide the input point cloud into irregular point patches (may overlap) via the Far-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28698" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "thest Point Sampling(FPS) and K-Nearest Neighborhood (KNN) algorithm following Point-BERT [65]. Formally, given two point clouds (views) " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_2" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ", we first use the FPS algorithm to sample pre-defined " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " points as group centers, which we called " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_1 \\in \\mathbb{R}^{n \\times 3}" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_2 \\in \\mathbb{R}^{n \\times 3}" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ", respectively. Then, based on the center points, we use KNN to choose the pre-defined " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": "-nearest points for each group center " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^{(i)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "0 \\leq i \\leq n" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ") and finally get " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " groups (each group contains " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " points). We apply the FPS and KNN on " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_1" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_2" + }, + { + "bbox": [ + 55, + 72, + 296, + 193 + ], + "type": "text", + "content": ", respectively:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 201, + 294, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 201, + 294, + 214 + ], + "spans": [ + { + "bbox": [ + 102, + 201, + 294, + 214 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {1} = \\operatorname {F P S} \\left(\\mathbf {X} _ {1}\\right), \\quad \\mathbf {G} _ {2} = \\operatorname {F P S} \\left(\\mathbf {X} _ {2}\\right) \\tag {5}", + "image_path": "58c5aadf825729d22c36658f3c8a20f5a0229029efa17c8da34037b50dfb85b9.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 82, + 217, + 295, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 217, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 82, + 217, + 295, + 230 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {1} = \\operatorname {K N N} \\left(\\mathbf {X} _ {1}, \\mathbf {G} _ {1}\\right), \\quad \\mathbf {P} _ {2} = \\operatorname {K N N} \\left(\\mathbf {X} _ {2}, \\mathbf {G} _ {2}\\right) \\tag {6}", + "image_path": "293f7da23937c7b179ea14016e2812e6da1b99141ec294aac5cf13f3b1baa697.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{G}_1, \\mathbf{G}_2 \\in \\mathbb{R}^{n \\times 3}" + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_1, \\mathbf{P}_2 \\in \\mathbb{R}^{n \\times (k \\cdot 3)}" + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "text", + "content": ". Then, the " + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_1" + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_2" + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "text", + "content": " both are treated as a sequence of length " + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "text", + "content": " and dimension " + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "inline_equation", + "content": "k \\cdot 3" + }, + { + "bbox": [ + 55, + 240, + 296, + 277 + ], + "type": "text", + "content": " as the input." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 281, + 227, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 227, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 227, + 294 + ], + "type": "text", + "content": "3.2. Backbone with positional query" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 296, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 296, + 428 + ], + "type": "text", + "content": "Following [33], we adopt a lightweight PointNet [34], followed by an asymmetric encoder-decoder structure consisting of Transformer blocks in the Point-PQAE. After the encoder projects the input embedding into the latent space, we use the latent representation of one view and view-relative positional embedding (VRPE) between two views to implement the position query through cross-attention [51]. We directly feed the output of the position query block to the decoder to reconstruct another view without using extra positional embedding. The last layer of the model adopts a simple prediction head to achieve the cross-reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 429, + 296, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 464 + ], + "type": "text", + "content": "Encoding. The encoder consists of the lightweight PointNet and standard Transformer blocks [51], which projects " + }, + { + "bbox": [ + 55, + 429, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_1" + }, + { + "bbox": [ + 55, + 429, + 296, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 429, + 296, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_2" + }, + { + "bbox": [ + 55, + 429, + 296, + 464 + ], + "type": "text", + "content": " to latent space, respectively:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 474, + 295, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 474, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 61, + 474, + 295, + 487 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {1} = \\operatorname {E n c o d e r} \\left(\\mathbf {P} _ {1}, \\mathbf {G} _ {1}\\right), \\quad \\mathbf {H} _ {2} = \\operatorname {E n c o d e r} \\left(\\mathbf {P} _ {2}, \\mathbf {G} _ {2}\\right), \\tag {7}", + "image_path": "08f3fab362c1b9695773d00e7c51aa538d0d90d368d87f8abe06c0440671718b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_1\\in \\mathbb{R}^{n\\times D}" + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_2\\in \\mathbb{R}^{n\\times D}" + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "content": " mean the latent representations of the view 1 and view 2. Note that " + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 497, + 295, + 533 + ], + "type": "text", + "content": " is the hidden dimension of the networks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 534, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 593 + ], + "type": "text", + "content": "Positional query block. After obtaining latent representations " + }, + { + "bbox": [ + 55, + 534, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_1" + }, + { + "bbox": [ + 55, + 534, + 296, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 296, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_2" + }, + { + "bbox": [ + 55, + 534, + 296, + 593 + ], + "type": "text", + "content": ", we feed them into the proposed positional query block to extract cross-information. The module can be mainly divided into 3 sub-stages: attain relative information, VRPE generation, and cross-attention." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Attain relative information. Take reconstructing view 2 from view 1 as an example. To reconstruct view 2, we have to know the latent representations of view 1 and the relative positions of view 2 centered on view 1, where the latent representations of view 1 can be obtained through the encoder network. Therefore, the key to reconstructing view 2 is to obtain the relative positions of view 2 centered on view 1. Back to the crop operation, we have recorded the geometric centers of views, denoted as " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_1" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_2" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ", respectively. We define the relative position of view 2 centered on view 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "as " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{RL}_{1\\rightarrow 2} = \\mathbf{L}_1 - \\mathbf{L}_2" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{RL}_{1\\rightarrow 2}\\in \\mathbb{R}^{1\\times 3}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": ". To better align the dimension in the latter operation, we expand the " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{RL}_{1\\rightarrow 2}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n\\times 3}" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " by repeating the content for " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " times. By obtaining the relative positions between the two views, we further add the group center location to help the PQ modules reconstruct each group of the inputs. Specifically, we define the group-/patch- wise relative positions as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 359, + 163, + 553, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 163, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 359, + 163, + 553, + 175 + ], + "type": "interline_equation", + "content": "\\mathbf {R P} _ {1 \\rightarrow 2} = \\operatorname {C o n C a t} \\left(\\mathbf {G} _ {2}, \\mathbf {R L} _ {1 \\rightarrow 2}\\right), \\tag {8}", + "image_path": "df2532c80129ff99b7658991c27dbc902c4ee226c10048e6b9f69c860337cf9f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 359, + 178, + 553, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 178, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 359, + 178, + 553, + 190 + ], + "type": "interline_equation", + "content": "\\mathbf {R P} _ {2 \\rightarrow 1} = \\operatorname {C o n C a t} \\left(\\mathbf {G} _ {1}, \\mathbf {R L} _ {2 \\rightarrow 1}\\right) \\tag {9}", + "image_path": "4e857ad2bd046805820f0b22c9d4a2f6b69a27e589acc22b098a487025dbdc5c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 195, + 553, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 553, + 220 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 195, + 553, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{RP} \\in \\mathbb{R}^{n \\times 6}" + }, + { + "bbox": [ + 313, + 195, + 553, + 220 + ], + "type": "text", + "content": " (as " + }, + { + "bbox": [ + 313, + 195, + 553, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{G} \\in \\mathbb{R}^{n \\times 3}" + }, + { + "bbox": [ + 313, + 195, + 553, + 220 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 195, + 553, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{RL} \\in \\mathbb{R}^{n \\times 3}" + }, + { + "bbox": [ + 313, + 195, + 553, + 220 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 195, + 553, + 220 + ], + "type": "inline_equation", + "content": "\\operatorname{ConCat}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 195, + 553, + 220 + ], + "type": "text", + "content": " is the concatenate operation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 220, + 554, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 554, + 316 + ], + "type": "text", + "content": "View-relative positional embedding (VRPE). Instead of simply using learnable positional encoding, which could hurt the expression of the relative positions between the two views (see ablation in Sec. 4.3), we use the fixed positional encoding of relative information to cut the uncertainty. Specifically, for elements in the second dimension of " + }, + { + "bbox": [ + 313, + 220, + 554, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{RP}_{1\\rightarrow 2}" + }, + { + "bbox": [ + 313, + 220, + 554, + 316 + ], + "type": "text", + "content": ", we use a sinusoid form [19] in Transformers to generate view-relative positional embedding " + }, + { + "bbox": [ + 313, + 220, + 554, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{VRPE}_{1\\rightarrow 2}" + }, + { + "bbox": [ + 313, + 220, + 554, + 316 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 322, + 322, + 553, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 322, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 322, + 322, + 553, + 408 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {V R P E} _ {1 \\rightarrow 2} ^ {i} = \\left[ \\sin \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e ^ {2 \\times \\frac {1}{D / 1 2}}}\\right), \\cos \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e ^ {2 \\times \\frac {1}{D / 1 2}}}\\right), \\right. \\\\ \\sin \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e ^ {2 \\times \\frac {2}{D / 1 2}}}\\right), \\cos \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e ^ {2 \\times \\frac {2}{D / 1 2}}}\\right), \\ldots , \\\\ \\left. \\sin \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e}\\right), \\cos \\left(\\frac {\\mathbf {R P} _ {1 \\rightarrow 2} ^ {i}}{e}\\right)\\right] \\tag {10} \\\\ \\end{array}", + "image_path": "4785d4282d2c2d2c0eb1ae5e71a44a57e99ab89ca7aac1ca3775e090f123151b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{RP}_{1\\rightarrow 2}^i\\in \\mathbb{R}^{n\\times 1}" + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{VRPE}_{1\\rightarrow 2}^i\\in \\mathbb{R}^{n\\times (D / 6)}" + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "inline_equation", + "content": "1\\leq i\\leq 6" + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "inline_equation", + "content": "e = 10000" + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "text", + "content": " is pre-defined, which is also used in MAE [19]. Finally, we concatenate " + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{VRPE}_{1\\rightarrow 2}^i" + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "text", + "content": " for each dimension to obtain " + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{VRPE}_{1\\rightarrow 2}\\in \\mathbb{R}^{n\\times D}" + }, + { + "bbox": [ + 313, + 413, + 554, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "text", + "content": "Cross-attention. Empirically, assume the latent representations of view 1, denoted as " + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_1" + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "text", + "content": ", contains intrinsic features and the global characteristics of its source point cloud, then with the combination of the view-relative positional embeddings " + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{VRPE}_{1\\rightarrow 2}" + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "text", + "content": ", the view 2 can be reconstructed. Specifically, we employ " + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{VRPE}_{1\\rightarrow 2}" + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "text", + "content": " (as " + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_1" + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "text", + "content": " (as " + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{K},\\mathbf{V}" + }, + { + "bbox": [ + 313, + 463, + 554, + 545 + ], + "type": "text", + "content": ") in cross-attention mechanisms, as can be formalized:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 338, + 552, + 553, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 552, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 338, + 552, + 553, + 594 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {T} _ {2} = \\operatorname {A t t n} \\left(\\mathbf {Q} _ {\\mathbf {V R P E} _ {1 \\rightarrow 2}}, \\mathbf {K} _ {\\mathbf {H} _ {1}}, \\mathbf {V} _ {\\mathbf {H} _ {1}}\\right) \\\\ = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q} _ {\\mathbf {V R P E} _ {1 \\rightarrow 2}} \\mathbf {K} _ {\\mathbf {H} _ {1}}}{\\sqrt {D}}\\right) \\mathbf {V} _ {\\mathbf {H} _ {1}} \\tag {11} \\\\ \\end{array}", + "image_path": "568d5eb77c9098b245744ceb6c7f8aa45d23d1d486f30f4cd0f910731a2919d3.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{\\mathbf{VRPE}_{1\\rightarrow 2}} = \\mathbf{VRPE}_{1\\rightarrow 2}\\mathbf{W}_{\\mathbf{Q}},\\mathbf{K}_{\\mathbf{H}_1} = \\mathbf{H}_1\\mathbf{W}_{\\mathbf{K}}" + }, + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{V_{H_1}} = \\mathbf{H_1W_V}" + }, + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{Q}},\\mathbf{W}_{\\mathbf{K}}" + }, + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{W_V}" + }, + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "text", + "content": " are learnable parameters. Reconstructing view 1 from view 2 is a siamese condition of this, through the same way as stated above, we can obtain " + }, + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{RL}_{2\\to 1},\\mathbf{RP}_{2\\to 1},\\mathbf{VRPE}_{2\\to 1}" + }, + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "text", + "content": " and yield " + }, + { + "bbox": [ + 313, + 598, + 554, + 670 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "text", + "content": "Decoding. The obtained " + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_1" + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_2" + }, + { + "bbox": [ + 313, + 671, + 553, + 694 + ], + "type": "text", + "content": " are fed into a decoder composed of a few transformer blocks:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 337, + 700, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 337, + 700, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {1} = \\operatorname {D e c o d e r} \\left(\\mathbf {T} _ {1}\\right), \\quad \\mathbf {Z} _ {2} = \\operatorname {D e c o d e r} \\left(\\mathbf {T} _ {2}\\right) \\tag {12}", + "image_path": "ea5e2c8247e76db0c5b142201199bf4a5af860d7168b484c80081f2f8af5b4e0.jpg" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28699" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 111 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 111 + ], + "type": "text", + "content": "Finally, similar to [33], we adopt a projection head (composed of a fully connected layer) to predict the input point cloud as follows, where " + }, + { + "bbox": [ + 55, + 72, + 296, + 111 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{pred}^{1},\\mathbf{P}_{pred}^{2}\\in \\mathbb{R}^{n\\times k\\times 3}" + }, + { + "bbox": [ + 55, + 72, + 296, + 111 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 107, + 119, + 295, + 134 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 119, + 295, + 134 + ], + "spans": [ + { + "bbox": [ + 107, + 119, + 295, + 134 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {p r e d} ^ {1} = \\operatorname {R e s h a p e} \\left(\\operatorname {L i n e a r} \\left(\\mathbf {Z} _ {1}\\right)\\right), \\tag {13}", + "image_path": "0408d55cc98e9810ebf331309b325c82e05ba64f85fade5bd828c5a92308872e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 108, + 136, + 295, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 136, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 108, + 136, + 295, + 152 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {p r e d} ^ {2} = \\operatorname {R e s h a p e} \\left(\\operatorname {L i n e a r} \\left(\\mathbf {Z} _ {2}\\right)\\right) \\tag {14}", + "image_path": "be9b5b9e853489993398560abc0dd7ed09196988c5fa6df7fd25393739197bcf.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 161, + 165, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 165, + 173 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 165, + 173 + ], + "type": "text", + "content": "3.3. Objective function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "text", + "content": "The overall objective is to perform cross-reconstruction, which involves predicting view 1 from view 2 and predicting view 2 from view 1. Given the predicted point patches " + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{pred}^{1}" + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{pred}^{2}" + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "text", + "content": " and ground truth " + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_1" + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_2" + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "text", + "content": ", we compute the cross-reconstruction loss using the " + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "text", + "content": " Chamfer Distance loss function [12], written as " + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cross} = \\mathcal{L}_{2\\rightarrow 1} + \\mathcal{L}_{1\\rightarrow 2}" + }, + { + "bbox": [ + 55, + 175, + 296, + 258 + ], + "type": "text", + "content": ", where:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 265, + 295, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 265, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 94, + 265, + 295, + 329 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {2 \\rightarrow 1} = \\frac {1}{\\left| \\mathbf {P} _ {p r e d} ^ {1} \\right|} \\sum_ {a \\in \\mathbf {P} _ {p r e d} ^ {1}} \\min _ {b \\in \\mathbf {P} _ {1}} \\| a - b \\| _ {2} ^ {2} \\tag {15} \\\\ + \\frac {1}{| \\mathbf {P} _ {1} |} \\sum_ {b \\in \\mathbf {P} _ {1}} \\min _ {a \\in \\mathbf {P} _ {p r e d} ^ {1}} \\| a - b \\| _ {2} ^ {2} \\\\ \\end{array}", + "image_path": "8977db2184751a1139fb41fa69642b53859ae8b46dfa9179f3cf4256f9a6cb65.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "type": "inline_equation", + "content": "|\\mathbf{P}|" + }, + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "type": "text", + "content": " is the cardinality of the set " + }, + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "type": "text", + "content": " distance. The " + }, + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{1\\to 2}" + }, + { + "bbox": [ + 55, + 338, + 296, + 363 + ], + "type": "text", + "content": " follows similarly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 374, + 137, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 137, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 137, + 387 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 394, + 296, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 296, + 478 + ], + "type": "text", + "content": "This section is organized as follows: First, we present the model architecture and pre-training details. Second, we validate the effectiveness of our pre-trained model on a wide range of downstream tasks, including object classification, few-shot learning, and part segmentation. Finally, ablation studies are conducted to demonstrate the properties and robustness of our proposed Point-PQAE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 482, + 168, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 168, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 168, + 495 + ], + "type": "text", + "content": "4.1. Pre-training setups" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 497, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 545 + ], + "type": "text", + "content": "Pre-trained dataset. We use the dataset ShapeNet [6] for pre-training of Point-PQAE following previous studies [9, 33, 37]. ShapeNet [6] consists of about 51,300 clean 3D models, covering 55 common object categories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 546, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 652 + ], + "type": "text", + "content": "Model structure. We adopt standard Transformer blocks [51] in the autoencoder's backbone where the encoder has 12 Transformer blocks and the decoder has 4 blocks. MLP ratio in Transformer blocks is set to 4. Each Transformer block has 384 hidden dimensions and 6 heads. The positional query module between encoder and decoder only performs cross-attention mechanism once as stated in Sec. 3.2, and the number of heads is set to 6 and remains 384 hidden dimensions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Experiment detail. For each input point cloud, we only apply random rotation for pre-training data augmentation. After sampling 1024 points via FPS from the input point cloud, we generate two different partial point clouds by randomly cropping and rotating as illustrated in Sec. 3.1 with" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "r_m = 0.6" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ". Both of them are divided into 64 patches with 32 points via FPS and KNN. The Point-PQAE model undergoes pre-training for 300 epochs using an AdamW [29] optimizer with a batch size of 128. The initial learning rate is set to 0.0005, with a weight decay of 0.05, with cosine learning rate decay [28] utilized. We visualize the pretraining results in Fig. 3. It shows Point-PQAE learns cross-knowledge well and is able to generalize excellently to other crop ratios, even though it was pre-trained with " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "r_m = 0.6" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 184, + 520, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 520, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 520, + 196 + ], + "type": "text", + "content": "4.2. Transfer learning on downstream tasks" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 199, + 553, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 553, + 235 + ], + "type": "text", + "content": "Transfer protocol. We use three transfer learning protocols for classification tasks following [9]: FULL, MLP-LINEAR, MLP-3, which are detailed in Appendix Sec. 3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "type": "text", + "content": "3D real-world object classification. The transferability of models can be well demonstrated through testing on a 3D real-world object dataset. Therefore, we transfer our pre-trained model to ScanObjectNN [50] for classification, which is one of the most challenging 3D datasets, covering approximately 15,000 real-world objects across 15 categories. We conduct experiments on three variants: OBJBG, OBJ-ONLY, and PB-T50-RS. The results are reported in Tab. 1. Our Point-PQAE surpasses previous single-modal self-supervised methods by a margin under all protocols. In contrast to the most relevant baseline, Point-MAE, PointPQAE significantly outperforms it, showing improvements of " + }, + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "type": "text", + "content": ". Especially for the MLP-LINEAR and MLP-3 protocols, our Point-PQAE outperforms Point-MAE by " + }, + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 313, + 235, + 554, + 474 + ], + "type": "text", + "content": " on average, respectively. The excellent frozen representation evaluation results indicate the superior quality of the representation learned by our Point-PQAE during pre-training. This shows that the cross-reconstruction paradigm learns more robust representations and generalizes better on real-world datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 474, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 629 + ], + "type": "text", + "content": "3D clean object classification. We evaluate the performance of our pre-trained model for object classification using the ModelNet40 dataset [59]. ModelNet40 comprises 12,311 meticulously crafted 3D CAD models, representing 40 distinct object categories. During testing, we adhere to the standard voting method [26] to ensure fair comparisons with previous work such as [7, 33]. Standard random scaling and random translation are applied for data augmentation during training, and the input point clouds exclusively contain coordinate information, without additional normal information provided. The results are presented in Tab. 1. Our Point-PQAE yields better or comparable results compared to previous approaches." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Few-shot learning. We conduct the few-shot learning classification on the ModelNet40 dataset to show the generalization ability of our method following the protocols in previous studies [33, 41, 53, 65]. The few-shot learning experiments are conducted in the form of \"w-way, s-shot\" including four parts where " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "w \\in \\{5, 10\\}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " represents the number of randomly selected classes and " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "s \\in \\{10, 20\\}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " means" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28700" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 101, + 173, + 166 + ], + "blocks": [ + { + "bbox": [ + 107, + 101, + 173, + 166 + ], + "lines": [ + { + "bbox": [ + 107, + 101, + 173, + 166 + ], + "spans": [ + { + "bbox": [ + 107, + 101, + 173, + 166 + ], + "type": "image", + "image_path": "a3026c65407e601c8dec37467fd5b25a1e53e55810f8d88097109312c5e0de7b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 189, + 82, + 252, + 91 + ], + "lines": [ + { + "bbox": [ + 189, + 82, + 252, + 91 + ], + "spans": [ + { + "bbox": [ + 189, + 82, + 252, + 91 + ], + "type": "text", + "content": "Cropped Rebuilt" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 115, + 188, + 160, + 232 + ], + "blocks": [ + { + "bbox": [ + 115, + 188, + 160, + 232 + ], + "lines": [ + { + "bbox": [ + 115, + 188, + 160, + 232 + ], + "spans": [ + { + "bbox": [ + 115, + 188, + 160, + 232 + ], + "type": "image", + "image_path": "b25778fcd134e75327060f3ee0fd455e866fc2a7f49f0db55068625858d11c11.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 259, + 171, + 269 + ], + "lines": [ + { + "bbox": [ + 113, + 259, + 171, + 269 + ], + "spans": [ + { + "bbox": [ + 113, + 259, + 171, + 269 + ], + "type": "text", + "content": "Input point cloud" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 186, + 100, + 258, + 163 + ], + "blocks": [ + { + "bbox": [ + 186, + 100, + 258, + 163 + ], + "lines": [ + { + "bbox": [ + 186, + 100, + 258, + 163 + ], + "spans": [ + { + "bbox": [ + 186, + 100, + 258, + 163 + ], + "type": "image", + "image_path": "1bb13258c46e6dfc24d340715ba35b1aae8dd1c3bb9c63fea0947121aa56e438.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 82, + 332, + 91 + ], + "lines": [ + { + "bbox": [ + 270, + 82, + 332, + 91 + ], + "spans": [ + { + "bbox": [ + 270, + 82, + 332, + 91 + ], + "type": "text", + "content": "Cropped Rebuilt" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 185, + 176, + 257, + 239 + ], + "blocks": [ + { + "bbox": [ + 185, + 176, + 257, + 239 + ], + "lines": [ + { + "bbox": [ + 185, + 176, + 257, + 239 + ], + "spans": [ + { + "bbox": [ + 185, + 176, + 257, + 239 + ], + "type": "image", + "image_path": "d82c0c2b96671c70322a59c7b27114f86c1f4a94eb7eae12cdf22d26cf35b299.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 259, + 247, + 268 + ], + "lines": [ + { + "bbox": [ + 195, + 259, + 247, + 268 + ], + "spans": [ + { + "bbox": [ + 195, + 259, + 247, + 268 + ], + "type": "text", + "content": "Crop ratio " + }, + { + "bbox": [ + 195, + 259, + 247, + 268 + ], + "type": "inline_equation", + "content": "20\\%" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 266, + 99, + 340, + 160 + ], + "blocks": [ + { + "bbox": [ + 266, + 99, + 340, + 160 + ], + "lines": [ + { + "bbox": [ + 266, + 99, + 340, + 160 + ], + "spans": [ + { + "bbox": [ + 266, + 99, + 340, + 160 + ], + "type": "image", + "image_path": "a4e01672b06f5245dac48c0c7739a732b62c245b571f607a57703c76ff8d3769.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 266, + 175, + 336, + 244 + ], + "blocks": [ + { + "bbox": [ + 266, + 175, + 336, + 244 + ], + "lines": [ + { + "bbox": [ + 266, + 175, + 336, + 244 + ], + "spans": [ + { + "bbox": [ + 266, + 175, + 336, + 244 + ], + "type": "image", + "image_path": "85c35790e2ae81bb1eb84686abd7d9d42a107580d4bef8973f3d8248dbbf8352.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 259, + 332, + 268 + ], + "lines": [ + { + "bbox": [ + 281, + 259, + 332, + 268 + ], + "spans": [ + { + "bbox": [ + 281, + 259, + 332, + 268 + ], + "type": "text", + "content": "Crop ratio " + }, + { + "bbox": [ + 281, + 259, + 332, + 268 + ], + "type": "inline_equation", + "content": "40\\%" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 347, + 95, + 416, + 167 + ], + "blocks": [ + { + "bbox": [ + 350, + 82, + 412, + 91 + ], + "lines": [ + { + "bbox": [ + 350, + 82, + 412, + 91 + ], + "spans": [ + { + "bbox": [ + 350, + 82, + 412, + 91 + ], + "type": "text", + "content": "Cropped Rebuilt" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 95, + 416, + 167 + ], + "lines": [ + { + "bbox": [ + 347, + 95, + 416, + 167 + ], + "spans": [ + { + "bbox": [ + 347, + 95, + 416, + 167 + ], + "type": "image", + "image_path": "784a93b547adc88ec643569db3f3db5486a1caf47ca964c47af5eb353c504ad8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 348, + 178, + 419, + 243 + ], + "blocks": [ + { + "bbox": [ + 348, + 178, + 419, + 243 + ], + "lines": [ + { + "bbox": [ + 348, + 178, + 419, + 243 + ], + "spans": [ + { + "bbox": [ + 348, + 178, + 419, + 243 + ], + "type": "image", + "image_path": "b006035c5d1b4896f6c459d0193bb6dd1978934f7ef42d8bbce7da017cf7b932.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 259, + 410, + 268 + ], + "lines": [ + { + "bbox": [ + 358, + 259, + 410, + 268 + ], + "spans": [ + { + "bbox": [ + 358, + 259, + 410, + 268 + ], + "type": "text", + "content": "Crop ratio " + }, + { + "bbox": [ + 358, + 259, + 410, + 268 + ], + "type": "inline_equation", + "content": "60\\%" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 426, + 95, + 496, + 161 + ], + "blocks": [ + { + "bbox": [ + 430, + 82, + 492, + 91 + ], + "lines": [ + { + "bbox": [ + 430, + 82, + 492, + 91 + ], + "spans": [ + { + "bbox": [ + 430, + 82, + 492, + 91 + ], + "type": "text", + "content": "Cropped Rebuilt" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 95, + 496, + 161 + ], + "lines": [ + { + "bbox": [ + 426, + 95, + 496, + 161 + ], + "spans": [ + { + "bbox": [ + 426, + 95, + 496, + 161 + ], + "type": "image", + "image_path": "15e48a1586f403deefc6e9bfa332e7318339ab3c963de707e25df079dd0a5f7a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 425, + 177, + 495, + 245 + ], + "blocks": [ + { + "bbox": [ + 425, + 177, + 495, + 245 + ], + "lines": [ + { + "bbox": [ + 425, + 177, + 495, + 245 + ], + "spans": [ + { + "bbox": [ + 425, + 177, + 495, + 245 + ], + "type": "image", + "image_path": "1ccbff0f4bfb44b246bc01219a203df5c5d0be55cb500dc6a2ed895b8a9f959c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 439, + 258, + 490, + 268 + ], + "lines": [ + { + "bbox": [ + 439, + 258, + 490, + 268 + ], + "spans": [ + { + "bbox": [ + 439, + 258, + 490, + 268 + ], + "type": "text", + "content": "Crop ratio " + }, + { + "bbox": [ + 439, + 258, + 490, + 268 + ], + "type": "inline_equation", + "content": "80\\%" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 280, + 553, + 303 + ], + "lines": [ + { + "bbox": [ + 55, + 280, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 553, + 303 + ], + "type": "text", + "content": "Figure 3. Cross-reconstruction results on ShapeNet. The arrow points from source point clouds to cross-reconstruction results. PointPQAE generalizes well to other crop ratios though with minimum crop ratio " + }, + { + "bbox": [ + 55, + 280, + 553, + 303 + ], + "type": "inline_equation", + "content": "r_m = 0.6" + }, + { + "bbox": [ + 55, + 280, + 553, + 303 + ], + "type": "text", + "content": " when pre-training." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "type": "text", + "content": "the number of randomly selected samples for each " + }, + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "type": "text", + "content": ". Each part is conducted with 10 independent trials. The mean accuracy and standard deviation are reported in Tab. 2. Our Point-PQAE demonstrates superior performance compared to previous methods in few-shot learning, even surpassing cross-modal methods that utilize strong pre-trained teachers such as ACT and ReCon under FULL protocol. Specifically, Point-PQAE achieves " + }, + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 55, + 308, + 296, + 429 + ], + "type": "text", + "content": " improvement under MLP-LINEAR protocol and shows enhanced performance over the self-construction method Point-MAE." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 434, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 625 + ], + "type": "text", + "content": "Part segmentation. We conduct part segmentation experiments on the ShapeNetPart [64] to validate the effectiveness of our Point-PQAE. The ShapeNetPart contains 16,881 objects covering 16 categories. We sample 2,048 points from each input point cloud, following previous work [33, 65] and divide each cloud into 128 point patches. We use the same segmentation head and utilize learned features from the 4th, 8th, and 12th layers of the Transformer block as in Point-MAE [33]. We concatenate three levels of features. Then average pooling, max pooling, and upsampling are utilized to generate features for each point and an MLP is applied for label prediction. The results are reported in Tab. 3a. The results are reported in Tab. 3a, which show that Point-PQAE achieves a comparable Inst.mIoU to the previous method, achieving " + }, + { + "bbox": [ + 55, + 434, + 296, + 625 + ], + "type": "inline_equation", + "content": "84.6\\%" + }, + { + "bbox": [ + 55, + 434, + 296, + 625 + ], + "type": "text", + "content": " in Cls.mIoU and improving the from-scratch baseline by " + }, + { + "bbox": [ + 55, + 434, + 296, + 625 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 55, + 434, + 296, + 625 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "3D scene segmentation. Semantic segmentation on large-scale 3D scenes is challenging, requiring models possessing a deep comprehension of semantics and intricate local geometric relationships. We report semantic segmentation results on the S3DIS dataset [3] in Tab. 3b. Our Point-PQAE improves the from-scratch baseline by " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", and outperforms the self-reconstruction method" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 308, + 522, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 522, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 522, + 320 + ], + "type": "text", + "content": "Point-MAE by " + }, + { + "bbox": [ + 313, + 308, + 522, + 320 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 313, + 308, + 522, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 308, + 522, + 320 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 313, + 308, + 522, + 320 + ], + "type": "text", + "content": " in mAcc and mIoU." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 327, + 405, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 405, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 405, + 339 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 388 + ], + "type": "text", + "content": "Experiments are conducted to show the properties of our Point-PQAE. We report the classification results on three variants of ScanObjectNN. The experiment settings are aligned with Sec. 4.2." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 389, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 554, + 676 + ], + "type": "text", + "content": "View-relative positional embedding. To analyze the effect of view-relative positional embedding in the positional query block, we conduct a series of experiments using different types of positional embeddings. We primarily consider two types of embeddings: Sinusoid (as discussed in Sec. 3.2) and Learnable, along with a None group. Learnable embedding refers to utilizing the relative position information " + }, + { + "bbox": [ + 313, + 389, + 554, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{RP}_{1\\rightarrow 2}" + }, + { + "bbox": [ + 313, + 389, + 554, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 389, + 554, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{RP}_{2\\rightarrow 1}" + }, + { + "bbox": [ + 313, + 389, + 554, + 676 + ], + "type": "text", + "content": " as input and employing an MLP composed of two linear layers with an activation function to map the 6-dimensional input to a D-dimensional output. The None group involves using randomly assigned embeddings to assess the effectiveness of VRPE design. Tab. 4a displays the results of different positional embeddings. In the None group, the model cannot learn the relative positional information across two views, leading to a significant drop in accuracy. In addition to the type of VRPE, we also experiment with Absolute Positional Embedding (APE) to compare its performance with our proposed VRPE. The APE is unsuitable for our framework since it requires a shared coordinate system for both views, making it incompatible with the normalization and rotation used in our method. We incorporate APE into our Point-PQAE by removing these operations. The results in Tab. 4a show APE performs worse and support our claim." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Data augmentation. After applying the random crop mechanism to the point cloud, we further perform data augmentation on the cropped views to generate new views." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28701" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 130, + 118, + 476, + 544 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. Classification accuracy (%) on ScanObjectNN and ModelNet40. The inference model parameters #P (M) are reported. Three variants are evaluated on ScanObjectNN and the accuracy obtained on ModelNet40 is reported for both 1K and 8K points. We compare methods using the " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " plain Transformer architectures, e.g. Point-MAE [33], " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " hierarchical Transformer architectures and " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " dedicated architectures for 3D. The dagger(†) denotes the baseline results reported from ReCon [37] which aligns augmentation with us and ReCon." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 130, + 118, + 476, + 544 + ], + "lines": [ + { + "bbox": [ + 130, + 118, + 476, + 544 + ], + "spans": [ + { + "bbox": [ + 130, + 118, + 476, + 544 + ], + "type": "table", + "html": "
Methods#PScanObjectNNModelNet40
OBJ_BGOBJ_ONLYPB_T50_RS1K P8K P
Supervised Learning Only
○PointNet [34]3.573.379.268.089.290.8
○PointNet++ [35]1.582.384.377.990.791.9
○DGCNN [54]1.882.886.278.192.9-
○PointCNN [24]0.686.185.578.592.2-
○SimpleView [13]---80.5±0.393.9-
○PCT [15]2.88---93.2-
with Self-Supervised Representation Learning (FULL)
•Transformer [51]22.183.084.079.191.491.8
•Point-BERT [65]22.187.488.183.093.293.8
•MaskPoint [25]-89.388.184.393.8-
•Point-MAE [33]22.190.088.385.293.894.0
○Point-M2AE [66]15.391.288.986.494.0-
•PointDif [74]22.193.391.987.6--
•Point-MAE† [33]22.192.691.988.493.894.0
•Point-PQAE22.195.093.689.694.094.3
Methods using cross-modal information and teacher models
•Joint-MAE [16]-90.988.986.194.0-
•TAP [55]22.190.489.585.794.0-
•ACT [9]22.193.391.988.293.794.0
○PointMLP+ULIP [63]---89.494.594.7
○I2P-MAE [67]15.394.291.690.194.1
•ReCon [37]43.695.293.690.694.594.7
with Self-Supervised Representation Learning (MLP-LINEAR)
•Point-MAE† [33]22.182.8±0.383.2±0.274.1±0.290.2±0.190.7±0.1
•Point-PQAE22.189.3±0.390.2±0.480.8±0.192.0±0.292.2±0.1
Methods using cross-modal information and teacher models
•ACT [9]22.185.2±0.885.8±0.276.3±0.391.4±0.291.8±0.2
•ReCon [37]43.689.5±0.289.7±0.281.4±0.192.5±0.292.7±0.1
with Self-Supervised Representation Learning (MLP-3)
•Point-MAE† [33]22.185.8±0.385.5±0.280.4±0.291.3±0.291.7±0.2
•Point-PQAE22.190.7±0.290.9±0.283.3±0.192.8±0.192.9±0.1
Methods using cross-modal information and teacher models
•ACT [9]22.187.1±0.287.9±0.481.5±0.292.7±0.293.0±0.1
•ReCon [37]43.690.6±0.290.7±0.383.8±0.493.0±0.193.4±0.1
", + "image_path": "a08e614a5cdef566bcdffd36016810f1b1387158e3af069af3a27ce27f652e3c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 295, + 601 + ], + "type": "text", + "content": "Data augmentation is crucial for generating diverse and decoupled views, so we conducted ablation studies on different augmentation methods. The results in Tab. 4b indicate that rotation performs best." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Effectiveness of the crop mechanism. Random crop has been shown to be highly effective in 2D self-supervised learning (SSL), particularly for contrastive learning [4, 8, 14], and is also vital for generative learning [19]. However, its potential in 3D SSL remains largely unexplored. Even contrastive learning methods for point cloud understanding have not explored the power of it [2, 32]. To address this gap, we first introduce a custom-designed random crop mechanism for point cloud data, using it as the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 553, + 555, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 555, + 709 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 555, + 709 + ], + "type": "text", + "content": "foundation for generating decoupled views. Without the crop augmentation, the two views will contain exactly the same shape though with normalization and rotation applied. The model would only need to infer the augmentation rather than the inter-view relationship, which means the views are not decoupled. Conversely, given two isolated cropped point clouds, followed by min-max normalization and rotation for further decoupling, there remain overlapping and non-overlapping parts, requiring the model to effectively encode the intra-view parts in order to infer the unknown inter-view points in our Point-PQAE. However, directly applying random crop as an additional augmentation to the self-reconstruction method (Point-MAE) does" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28702" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 130, + 281, + 373 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": "Table 2. Few-shot classification results on ModelNet40. 10 independent trials are conducted in each experimental setting. The mean accuracy (\\%) and standard deviation are reported for each setting. The dagger(" + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": ") denotes the baseline results reported from ReCon [37] which aligns augmentation with us and ReCon." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 130, + 281, + 373 + ], + "lines": [ + { + "bbox": [ + 67, + 130, + 281, + 373 + ], + "spans": [ + { + "bbox": [ + 67, + 130, + 281, + 373 + ], + "type": "table", + "html": "
Methods5-way10-way
10-shot20-shot10-shot20-shot
○DGCNN [54]31.6±2.840.8±4.619.9±2.116.9±1.5
○OcCo [52]90.6±2.892.5±1.982.9±1.386.5±2.2
with Self-Supervised Representation Learning (FULL)
•Transformer [51]87.8±5.293.3±4.384.6±5.589.4±6.3
•Point-BERT [65]94.6±3.196.3±2.791.0±5.492.7±5.1
•MaskPoint [25]95.0±3.797.2±1.791.4±4.093.4±3.5
•Point-MAE [33]96.3±2.597.8±1.892.6±4.195.0±3.0
○Point-M2AE [66]96.8±1.898.3±1.492.3±4.595.0±3.0
•Point-MAE† [66]96.4±2.897.8±2.092.5±4.495.2±3.9
•Point-PQAE96.9±3.298.9±1.094.1±4.296.3±2.7
Methods using cross-modal information and teacher models
•Joint-MAE [16]96.7±2.297.9±1.992.6±3.795.1±2.6
•TAP [55]97.3±1.897.8±1.993.1±2.695.8±1.0
•ACT [9]96.8±2.398.0±1.493.3±4.095.6±2.8
○I2P-MAE [67]97.0±1.898.3±1.392.6±5.095.5±3.0
•ReCon [37]97.3±1.998.9±1.293.3±3.995.8±3.0
with Self-Supervised Representation Learning (MLP-LINEAR)
•Point-MAE† [66]91.1±5.691.7±4.083.5±6.189.7±4.1
•Point-PQAE93.0±4.696.8±1.989.0±5.293.5±3.8
with Self-Supervised Representation Learning (MLP-3)
•Point-MAE† [66]95.0±2.896.7±2.490.6±4.793.8±5.0
•Point-PQAE95.3±3.498.2±1.892.0±3.894.7±3.5
", + "image_path": "d7ff8aa1387991c682c961031975d3a1549eb7a9ecd1e1d4911e6bff7f13da8b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 439, + 173, + 522 + ], + "blocks": [ + { + "bbox": [ + 55, + 376, + 295, + 410 + ], + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 410 + ], + "type": "text", + "content": "Table 3. Segmentation results. Cls.mIoU (\\%) and Inst.mIoU (\\%) refer to Mean intersection over union for all classes and all instances, respectively. mAcc (\\%) refers to mean accuracy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 414, + 173, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 414, + 173, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 173, + 433 + ], + "type": "text", + "content": "(a) Part segmentation on ShapeNetPart." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 439, + 173, + 522 + ], + "lines": [ + { + "bbox": [ + 57, + 439, + 173, + 522 + ], + "spans": [ + { + "bbox": [ + 57, + 439, + 173, + 522 + ], + "type": "table", + "html": "
MethodCls.mIoU (%)Inst.mIoU (%)
Supervised
PointNet [34]80.483.7
PointNet++ [35]81.985.1
DGCNN [54]82.385.2
Self-supervised
Transformer [51]83.484.7
CrossPoint [2]-85.5
Point-BERT [65]84.185.6
Point-MAE [33]84.286.1
Point-PQAE (Ours)84.686.1
Cross-modal
ACT [9]84.786.1
ReCon [37]84.886.4
", + "image_path": "60f7cbfa772efafa25b00249761674892cfde2aaf64ec05b4cdd657872a49d66.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 181, + 440, + 296, + 511 + ], + "blocks": [ + { + "bbox": [ + 179, + 414, + 295, + 433 + ], + "lines": [ + { + "bbox": [ + 179, + 414, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 179, + 414, + 295, + 433 + ], + "type": "text", + "content": "(b) Semantic segmentation results on S3DIS Area 5." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 181, + 440, + 296, + 511 + ], + "lines": [ + { + "bbox": [ + 181, + 440, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 181, + 440, + 296, + 511 + ], + "type": "table", + "html": "
MethodmAcc (%)mIoU (%)
Supervised
PointNet [34]49.041.1
PointNet++ [35]67.153.5
Self-supervised
Transformer [51]68.660.0
Point-MAE [33]69.960.8
Point-PQAE (Ours)70.661.4
Cross-modal
ACT [9]71.161.2
", + "image_path": "e50debff01ce93a64b42ea1d9f295bded65e3673722e2c0e2cbdb5d0e38b3cfe.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 532, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 616 + ], + "type": "text", + "content": "not improve the performance much, as shown in Tab. 4c. In other words, it means our Point-PQAE fits crop mechanism better by harmoniously incorporating it into the decoupled views generation process. The experiment results in Tab. 4c demonstrate the critical role of random crop in our method. We set minimum crop ratio " + }, + { + "bbox": [ + 55, + 532, + 295, + 616 + ], + "type": "inline_equation", + "content": "r_m = 1.0" + }, + { + "bbox": [ + 55, + 532, + 295, + 616 + ], + "type": "text", + "content": " when removing the crop in our Point-PQAE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "Analysis on decoupled views generation. While crop plays an important role in our method, relying solely on the random crop mechanism to generate new views would degenerate cross-reconstruction to simply reconstructing one part from another. This is similar to the block mask self-reconstruction used in Point-MAE [33], where visible blocks are used to reconstruct masked blocks, which is also trivial for pre-training. The only difference between them" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 348, + 128, + 523, + 172 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 117 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 117 + ], + "type": "text", + "content": "Table 4. Ablation study with Point-PQAE pre-training on ShapeNet. The classification results by accuracy (\\%) on three variants of ScanObjectNN are reported. Default settings are marked in gray." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 369, + 116, + 499, + 126 + ], + "lines": [ + { + "bbox": [ + 369, + 116, + 499, + 126 + ], + "spans": [ + { + "bbox": [ + 369, + 116, + 499, + 126 + ], + "type": "text", + "content": "(a) View-Relative positional embedding" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 348, + 128, + 523, + 172 + ], + "lines": [ + { + "bbox": [ + 348, + 128, + 523, + 172 + ], + "spans": [ + { + "bbox": [ + 348, + 128, + 523, + 172 + ], + "type": "table", + "html": "
Positional EmbeddingOBJ_BGOBJ_ONLYLYPB_T50_RS
None84.585.979.3
APE (sinusoid)92.391.087.7
VRPE (learnable)93.493.189.1
VRPE (sinusoid Eq. (10))95.093.689.6
", + "image_path": "d77a8f6fa3b9ebf8f18edf63ed6ce000e553dbd800041b96b05348d03d16bca3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 348, + 184, + 522, + 237 + ], + "blocks": [ + { + "bbox": [ + 398, + 174, + 471, + 183 + ], + "lines": [ + { + "bbox": [ + 398, + 174, + 471, + 183 + ], + "spans": [ + { + "bbox": [ + 398, + 174, + 471, + 183 + ], + "type": "text", + "content": "(b) Data augmentation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 348, + 184, + 522, + 237 + ], + "lines": [ + { + "bbox": [ + 348, + 184, + 522, + 237 + ], + "spans": [ + { + "bbox": [ + 348, + 184, + 522, + 237 + ], + "type": "table", + "html": "
Data augmentationOBJ_BGOBJ_ONLYPB_T50_RS
jitter93.391.487.3
scale93.391.788.0
rotation95.093.689.6
scale&translate92.891.788.1
rotation+scale&translate93.892.989.1
", + "image_path": "cff9b9ba6bdfb25d5f145124cccce63ab8884ac377f662b03cdd006b8ec7a6f1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 348, + 251, + 522, + 297 + ], + "blocks": [ + { + "bbox": [ + 383, + 238, + 487, + 248 + ], + "lines": [ + { + "bbox": [ + 383, + 238, + 487, + 248 + ], + "spans": [ + { + "bbox": [ + 383, + 238, + 487, + 248 + ], + "type": "text", + "content": "(c) 3D random crop mechanism" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 348, + 251, + 522, + 297 + ], + "lines": [ + { + "bbox": [ + 348, + 251, + 522, + 297 + ], + "spans": [ + { + "bbox": [ + 348, + 251, + 522, + 297 + ], + "type": "table", + "html": "
MethodcropOBJ_BGOBJ_ONLYPB_T50_RS
Point-MAE✘92.691.988.4
Point-MAE✓92.992.188.8
Point-PQAE✘92.992.387.7
Point-PQAE✓95.093.689.6
", + "image_path": "b151369a2f039289a3b5542ac2346c6ab97036424b8cccba172278d497fd0180.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 348, + 311, + 519, + 361 + ], + "blocks": [ + { + "bbox": [ + 358, + 299, + 509, + 308 + ], + "lines": [ + { + "bbox": [ + 358, + 299, + 509, + 308 + ], + "spans": [ + { + "bbox": [ + 358, + 299, + 509, + 308 + ], + "type": "text", + "content": "(d) Effects of augmentations after random crop" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 348, + 311, + 519, + 361 + ], + "lines": [ + { + "bbox": [ + 348, + 311, + 519, + 361 + ], + "spans": [ + { + "bbox": [ + 348, + 311, + 519, + 361 + ], + "type": "table", + "html": "
norm.rotationOBJ_BGOBJ_ONLYPB_T50_RS
XX92.892.188.2
✓X93.392.687.9
X✓93.592.688.8
✓✓95.093.689.6
", + "image_path": "67a6975751f5524eb4b5a34e025ad64b415be1098eeb53c67761b4e1c22d3958.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 368, + 555, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 555, + 511 + ], + "type": "text", + "content": "is that the cropped parts can overlap somewhat, whereas the block mask does not. What makes the cross-reconstruction different from block mask self-reconstruction are the subsequent augmentations including normalization and rotation that decouple the cropped parts. By normalizing cropped point clouds centered on the geometric centers, the coordinate systems of the two point clouds become isolated and independent from each other, with rotation further amplifying the variance between the two views, thus enabling the generation of two decoupled views. The importance of the augmentations applied after the random crop can be validated through experiments, as shown in Tab. 4d." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 517, + 388, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 388, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 388, + 529 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel cross-reconstruction generative framework for self-supervised learning on 3D point clouds, called Point-PQAE. In contrast to well-studied self-reconstruction schemes, Point-PQAE reconstructs one cropped point cloud from another decoupled point cloud. To supply sufficient information for cross-reconstruction, we further propose a 3D view-relative positional embedding and a corresponding position-aware query module. Compared to self-reconstruction, the cross-reconstruction task, carefully designed by us, is much more challenging for pre-training, promoting the learning of richer semantic representations during pre-training. This enables our proposed Point-PQAE to outperform previous single-modal self-reconstruction methods by a margin and to perform on par with, or better than cross-modal methods." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28703" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In International conference on machine learning, pages 40-49. PMLR, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 202 + ], + "type": "text", + "content": "[2] Mohamed Afham, Isuru Dissanayake, Dinithi Dissanayake, Amaya Dharmasiri, Kanchana Thilakarathna, and Ranga Rodrigo. Crosspoint: Self-supervised cross-modal contrastive learning for 3d point cloud understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9902-9912, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 296, + 258 + ], + "type": "text", + "content": "[3] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1534-1543, 2016. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "text", + "content": "[4] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 296, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 350 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 352, + 294, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 294, + 406 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 294, + 406 + ], + "type": "text", + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 294, + 452 + ], + "type": "text", + "content": "[7] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. Pointgpt: Auto-regressively generative pretraining from point clouds. Advances in Neural Information Processing Systems, 36, 2024. 1, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 454, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 294, + 498 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 2, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 500, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 294, + 553 + ], + "type": "text", + "content": "[9] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2d image transformers help 3d representation learning? arXiv preprint arXiv:2212.08320, 2022. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 556, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 294, + 609 + ], + "type": "text", + "content": "[10] Bi'an Du, Xiang Gao, Wei Hu, and Xin Li. Self-contrastive learning with hard negative sampling for self-supervised point cloud learning. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3133–3142, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 666 + ], + "type": "text", + "content": "[11] Ikenna Enebuse, Mathias Foo, Babul Salam Ksm Kader Ibrahim, Hafiz Ahmed, Fhon Supmak, and Odongo Steven Eyobu. A comparative review of hand-eye calibration techniques for vision guided robots. IEEE Access, 9:113143-113155, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[12] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In International Conference on Machine Learning, pages 3809-3820. PMLR, 2021. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "text", + "content": "[14] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 184, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 226 + ], + "type": "text", + "content": "[15] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7:187-199, 2021. 1, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "type": "text", + "content": "[16] Ziyu Guo, Renrui Zhang, Longtian Qiu, Xianzhi Li, and Pheng-Ann Heng. Joint-mae: 2d-3d joint masked autoencoders for 3d point cloud pre-training. arXiv preprint arXiv:2302.14007, 2023. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 316 + ], + "type": "text", + "content": "[17] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 316, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 369 + ], + "type": "text", + "content": "[18] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 371, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 371, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 371, + 553, + 425 + ], + "type": "text", + "content": "[19] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 2, 4, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 426, + 553, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 479 + ], + "type": "text", + "content": "[20] Siyuan Huang, Yichen Xie, Song-Chun Zhu, and Yixin Zhu. Spatio-temporal self-supervised representation learning for 3d point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6535-6545, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 481, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 481, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 317, + 481, + 553, + 524 + ], + "type": "text", + "content": "[21] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25, 2012. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 526, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 558 + ], + "type": "text", + "content": "[22] Gang Li, Heliang Zheng, Daqing Liu, Chaoyue Wang, Bing Su, and Changwen Zheng. Semmae: Semantic-guided masking for learning masked autoencoders. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 602 + ], + "type": "text", + "content": "[23] Jiaxin Li, Ben M Chen, and Gim Hee Lee. So-net: Self-organizing network for point cloud analysis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9397–9406, 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 603, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 603, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 603, + 553, + 645 + ], + "type": "text", + "content": "[24] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 1, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 690 + ], + "type": "text", + "content": "[25] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "text", + "content": "[26] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28704" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "cloud analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8895-8904, 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 162 + ], + "type": "text", + "content": "[27] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 195 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 295, + 217 + ], + "type": "text", + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 295, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 295, + 261 + ], + "type": "text", + "content": "[30] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2906-2917, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 294 + ], + "type": "text", + "content": "[31] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 350 + ], + "type": "text", + "content": "[32] Bo Pang, Hongchi Xia, and Cewu Lu. Unsupervised 3d point cloud representation learning by triangle constrained contrast for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5229-5239, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 403 + ], + "type": "text", + "content": "[33] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 459 + ], + "type": "text", + "content": "[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1, 4, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 459, + 295, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 295, + 504 + ], + "type": "text", + "content": "[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 504, + 295, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 295, + 548 + ], + "type": "text", + "content": "[36] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9277-9286, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 295, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 602 + ], + "type": "text", + "content": "[37] Zekun Qi, Runpei Dong, Guofan Fan, Zheng Ge, Xiangyu Zhang, Kaisheng Ma, and Li Yi. Contrast with reconstruct: Contrastive 3d representation learning guided by generative pretraining. arXiv preprint arXiv:2302.02318, 2023. 1, 3, 5, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 295, + 635 + ], + "type": "text", + "content": "[38] Rui Qian, Xin Lai, and Xirong Li. 3d object detection for autonomous driving: A survey. Pattern Recognition, 130: 108796, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 295, + 692 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 692 + ], + "type": "text", + "content": "[39] Aditya Sanghi. Info3d: Representation learning on 3d objects using mutual information maximization and contrastive learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIX 16, pages 626-642. Springer, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 692, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 715 + ], + "type": "text", + "content": "[40] Oleg Yu Sergiyenko and Vera V Tyrsa. 3d optical machine vision sensors with intelligent data management for robotic" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 715 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "swarm navigation improvement. IEEE Sensors Journal, 21 (10):11262-11274, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[41] Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. Advances in Neural Information Processing Systems, 33:7212-7221, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "text", + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 164, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 208 + ], + "type": "text", + "content": "[43] Haoru Tan, Chuang Wang, Sitong Wu, Tieqiang Wang, Xu-Yao Zhang, and Cheng-Lin Liu. Proxy graph matching with proximal matching networks. In The Annual AAAI Conference on Artificial Intelligence (AAAI), 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 253 + ], + "type": "text", + "content": "[44] Haoru Tan, Sitong Wu, Fei Du, Yukang Chen, Zhibin Wang, Fan Wang, and Xiaojuan Qi. Data pruning via moving-one-sample-out. In Neural Information Processing Systems (NeurIPS), 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 254, + 553, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 287 + ], + "type": "text", + "content": "[45] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 288, + 553, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 332 + ], + "type": "text", + "content": "[46] Haoru Tan, Chuang Wang, Sitong Wu, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Ensemble quadratic assignment network for graph matching. International Journal of Computer Vision (IJCV), 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 333, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 388 + ], + "type": "text", + "content": "[47] Haoru Tan, Sitong Wu, Zhuotao Tian, Yukang Chen, Xiaojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity consistency for vision-language pre-training. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 432 + ], + "type": "text", + "content": "[48] Haoru Tan, Sitong Wu, Wei Huang, Shizhen Zhao, and Xiaojuan Qi. Data pruning by information maximization. In International Conference on Learning Representations (ICLR), 2025." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "text", + "content": "[49] Haoru Tan, Sitong Wu, Bo Zhao, Zeke Xie, and XIAOJUAN QI. Diff-in: Data influence estimation with differential approximation, 2025. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 533 + ], + "type": "text", + "content": "[50] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 534, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 534, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 534, + 553, + 579 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4, 5, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 579, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 634 + ], + "type": "text", + "content": "[52] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 1, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[53] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 715 + ], + "type": "text", + "content": "[54] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28705" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 1, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 150 + ], + "type": "text", + "content": "[55] Ziyi Wang, Xumin Yu, Yongming Rao, Jie Zhou, and Jiwen Lu. Take-a-photo: 3d-to-2d generative pre-training of point cloud models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5640–5650, 2023. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[56] Hai Wu, Ruifei He, Haoru Tan, Xiaojuan Qi, and Kaibin Huang. Vertical layering of quantized neural networks for heterogeneous inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "text", + "content": "[57] Sitong Wu, Haoru Tan, Yukang Chen, Shaofeng Zhang, Jingyao Li, Bei Yu, Xiaojuan Qi, and Jiaya Jia. Mixture-of-scores: Robust image-text data quality score via three lines of code. In International Conference on Computer Vision (ICCV), 2025. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[58] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. arXiv preprint arXiv:2312.10035, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 354 + ], + "type": "text", + "content": "[59] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 421 + ], + "type": "text", + "content": "[60] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part III 16, pages 574-591. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 455 + ], + "type": "text", + "content": "[61] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "text", + "content": "[62] Mingye Xu, Zhipeng Zhou, Yali Wang, and Yu Qiao. Towards robustness and generalization of point cloud representation: A geometry coding method and a large-scale object-level dataset. Computational Visual Media, 10(1):27-43, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 578 + ], + "type": "text", + "content": "[63] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1179-1189, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 635 + ], + "type": "text", + "content": "[64] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 690 + ], + "type": "text", + "content": "[65] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[66] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 598 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 1, 2, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 162 + ], + "type": "text", + "content": "[67] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[68] Shaofeng Zhang, Qiang Zhou, Sitong Wu, Haoru Tan, Zhibin Wang, Jina Huang, and Junchi Yan. Cr2pq: Continuous relative rotary positional query for dense visual representation learning. In The Thirteenth International Conference on Learning Representations. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "text", + "content": "[69] Shaofeng Zhang, Meng Liu, and Junchi Yan. The diversified ensemble neural network. Advances in Neural Information Processing Systems, 33:16001-16011, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[70] Shaofeng Zhang, Feng Zhu, Junchi Yan, Rui Zhao, and Xiaokang Yang. Zero-cl: Instance and feature decorrelation for negative-free symmetric contrastive learning. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 363 + ], + "type": "text", + "content": "[71] Shaofeng Zhang, Meng Liu, Junchi Yan, Hengrui Zhang, Lingxiao Huang, Xiaokang Yang, and Pinyin Lu. M-mix: Generating hard negatives via multi-sample mixing for contrastive learning. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, pages 2461-2470, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 364, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 553, + 408 + ], + "type": "text", + "content": "[72] Shaofeng Zhang, Feng Zhu, Rui Zhao, and Junchi Yan. Contextual image masking modeling via synergized contrasting without view augmentation for faster and better visual pretraining. In ICLR, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "text", + "content": "[73] Xiangdong Zhang, Shaofeng Zhang, and Junchi Yan. Pcmae: Learning to predict centers for point masked autoencoders. Advances in Neural Information Processing Systems, 37:80303-80327, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 454, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 509 + ], + "type": "text", + "content": "[74] Xiao Zheng, Xiaoshui Huang, Guofeng Mei, Yuenan Hou, Zhaoyang Lyu, Bo Dai, Wanli Ouyang, and Yongshun Gong. Point cloud pre-training with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22935-22945, 2024. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "text", + "content": "[75] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 554, + 553, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 598 + ], + "type": "text", + "content": "[76] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4490-4499, 2018. 1" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28706" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/7d9586d8-b2bc-4332-a23e-c79a83d7ffa9_content_list.json b/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/7d9586d8-b2bc-4332-a23e-c79a83d7ffa9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..df0732575b4c67053088bd0273e42dbf0877a19d --- /dev/null +++ b/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/7d9586d8-b2bc-4332-a23e-c79a83d7ffa9_content_list.json @@ -0,0 +1,1698 @@ +[ + { + "type": "text", + "text": "Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation", + "text_level": 1, + "bbox": [ + 238, + 130, + 759, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kaining Ying Henghui Ding Guangquan Jie Yu-Gang Jiang Fudan University, China https://henghuiding.com/OmniAVS/", + "bbox": [ + 223, + 202, + 769, + 262 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf108340ec71e323cdd2905750ff4d55cf6c7e98c9bc0d707710454ced831582.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 297, + 256, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/27e014ace92dec0868d33332164bd0b5ca078cab7e818199ca277e445e822583.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 367, + 122, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfb71e38f2cbc00fd73757223fdfd699c396e13804e47c9b1abf69e7d35c3bd8.jpg", + "image_caption": [ + "Which object mimics another?" + ], + "image_footnote": [], + "bbox": [ + 122, + 367, + 150, + 385 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b356832cc4bf4c27a087eb56f18f4365966197e56d6741c434dc6b69e71b6a32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 297, + 416, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0879cbf93a27bc442d9c092b46b0489cc91cc2e5a760f2e92b1b471680a32342.jpg", + "image_caption": [ + "The one speak the same language as ." + ], + "image_footnote": [], + "bbox": [ + 256, + 367, + 416, + 458 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "e same language as :", + "bbox": [ + 256, + 458, + 416, + 472 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/03f6a59ebc9dabfa5515685f8a514aa134fcd4551b29c4cbf2895ba6f94d1dbc.jpg", + "image_caption": [ + "Output: It is [SEG]. The girl is imitating the barking of the dog on the right." + ], + "image_footnote": [], + "bbox": [ + 416, + 297, + 575, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e878ece4aeb966d1f0cc4ecd81796671223db454a4de2f58d4a07c4a472ec82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 387, + 575, + 457 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b762d089ef619b203f215708cfa5893416a1923a7348c53ad12db9bfb2ff6cd7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 387, + 256, + 457 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62ac3b8f240ea5a7581d5a64c24a2eb2b11edd6a3983327325440160bf7a6aaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 457, + 122, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/66bb6ebcd1680f5aa6d187ddadc08d253463c4a2ac20836fd5ad71d1efb2fac2.jpg", + "image_caption": [ + " + serves as lead" + ], + "image_footnote": [], + "bbox": [ + 256, + 476, + 416, + 545 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4f73d2949afd8505f03af5d8f0bc2992bd90fd64d18cfa10bc0a005c8feb7ee2.jpg", + "image_caption": [ + "Output: It is [SEG]. They both speak Korean.", + "Output: It is [SEG]. The guitar on the right serves as lead." + ], + "image_footnote": [], + "bbox": [ + 416, + 476, + 575, + 546 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3d5a6c3d5072f2374a6d5bf5e583f0665389304b8245edc7b63ec8958d1106d4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 476, + 735, + 546 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa058d5b82664b4ace528c5f7fba79f937b7d6870b7f432d76655ce14b97efe4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 476, + 895, + 546 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73c001faf6336adfe9d22d883bfc1fe4d065210907d1b67c56f569bfd60fb733.jpg", + "image_caption": [ + "The one similar to serves as lead." + ], + "image_footnote": [], + "bbox": [ + 96, + 549, + 122, + 565 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7b31e2f421d44310fcc98aea365eaa32654150cb3aa149a96c534b3cc963cbe9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 565, + 256, + 635 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d4547cf5b1f4a3925042b29398a989815a5ac329f3f3972505d6ea0f037bfa5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 636, + 122, + 655 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce853e2e1e27764830e2b625207a9d158f641edff7feb8bf1407a57049099e35.jpg", + "image_caption": [ + "The instrument is played by someone of the same gender and makes a similar sound in ." + ], + "image_footnote": [], + "bbox": [ + 122, + 637, + 150, + 655 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b8da93d3bd86037194882cea478beb42a2a7de146bd07435d2f129d4846d0efb.jpg", + "image_caption": [ + "is played by someone of the similar sound in $<$ sound: $\\mathbb{N}^{\\prime}$ Sound" + ], + "image_footnote": [], + "bbox": [ + 256, + 565, + 416, + 636 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/222a8844b023e73f8aa6758d6ca968d1ca70b97aea11cd0049cf57df6ec8b2ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 416, + 565, + 575, + 636 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb0853a0b31c4fbf72cac2be3e722997d23506b17fa5f87f7fd03497dcb6a8c3.jpg", + "image_caption": [ + "Output: It is [SEG]. The pipa held by the man on the right." + ], + "image_footnote": [], + "bbox": [ + 575, + 565, + 735, + 636 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/361881400489dc22938d8cf7f88a75558e1ae1f6e40130886cdbf7f1cd293614.jpg", + "image_caption": [ + "er in >.", + "Figure 1. Examples of the proposed benchmark Omnimodal Referring Audio-Visual Segmentation (OmniAVS) to show its nature and flexibility. OmniAVS introduces 3 key features: 1) It supports diverse multimodal referring expressions that flexibly combine text, speech, sound, and image for referring audio-visual segmentation; 2) It emphasizes understanding the content of audio rather than merely hearing them; 3) It incorporates complex reasoning and world knowledge in expressions, prompting models to provide explanations for their segmentation decisions. These characteristics make OmniAVS practical for real-world use and well-suited for developing omnimodal models with fine-grained perception." + ], + "image_footnote": [], + "bbox": [ + 575, + 637, + 614, + 665 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3744975a652ab5a6f46cfb8b84c786a760821bffeb0376e3218b51008d24dc8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 565, + 895, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 768, + 326, + 786 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Referring audio-visual segmentation (RAVS) has recently seen significant advancements, yet challenges remain in integrating multimodal information and deeply understand", + "bbox": [ + 89, + 804, + 482, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing and reasoning about audiovisual content. To extend the boundaries of RAVS and facilitate future research in this field, we propose Omnimodal Referring AudioVisual Segmentation (OmniAVS), a new dataset containing 2,104 videos and 61,095 multimodal referring expressions. OmniAVS stands out with three key innovations: (1) 8 types of multimodal expressions that flexibly combine text, speech, sound, and visual cues; (2) an emphasis", + "bbox": [ + 511, + 771, + 906, + 892 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Henghui Ding (henghui.ding@gmail.com) is the corresponding author with the Institute of Big Data, College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22575", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on understanding audio content beyond just detecting their presence; and (3) the inclusion of complex reasoning and world knowledge in expressions. Furthermore, we introduce Omnimodal Instructed Segmentation Assistant (OISA), to address the challenges of multimodal reasoning and fine-grained understanding of audiovisual content in OmniAVS. OISA uses MLLM to comprehend complex cues and perform reasoning-based segmentation. Extensive experiments show that OISA outperforms existing methods on OmniAVS and achieves competitive results on other related tasks.", + "bbox": [ + 88, + 90, + 482, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 258, + 222, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Referring Audio-Visual Segmentation (RAVS) [19, 68] is an emerging field focused on segmenting target objects indicated by referring expressions (e.g., text) within audio-visual scenes using audio-visual cues. The combination of referring and audio distinguishes RAVS from Referring Video Object Segmentation (RVOS) [17, 26, 31, 40, 42, 57], which uses text descriptions to refer to objects in silent videos, and Audio-Visual Segmentation (AVS) [24, 86], which segments any sound-producing object without referring expressions. RAVS enables practical applications in video conferencing and robotics, etc.", + "bbox": [ + 89, + 281, + 482, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Understanding Rather than Just Hearing. However, the association between expressions and audio cues in the previous RAVS dataset Ref-AVS [68] remains superficial, primarily limited to basic acoustic properties such as sound occurrence, intensity, and temporal sequence. For example, expressions like \"Who is making the loudest sound?\" only address the surface-level characteristics without delving into the content of the sounds. Recent works [3, 33, 53, 70, 84] in image and video segmentation have evolved from simple semantic perception to advanced reasoning understanding. However, such progress is notably absent in audio-visual segmentation. This gap naturally leads us to explore the potential for complex reasoning in audio-visual scenes. Besides, drawing insights from human auditory perception studies [6], understanding sound content is essential for sophisticated audio-visual scene reasoning. To achieve these goals, we present OmniAVS dataset to advance reasoning-based segmentation in audio-visual scenes. As shown in Figure 2(b), our dataset contains expressions that demand reasoning about audio content, such as \"Who is most likely to be sick?\". These expressions establish complex cognitive chains that go beyond basic acoustic features, e.g., sound $\\rightarrow$ coughing $\\rightarrow$ sickness. Furthermore, we enhance the dataset's interpretability by providing detailed explanations for reasoning-based expressions.", + "bbox": [ + 89, + 448, + 482, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards Omnimodal Referring. The emergence of flagship model ChatGPT-4o [51] sets a new standard for modern AI systems, emphasizing the need to process arbitrary combinations of omnimodal inputs, e.g., audio, vision, and text. In the context of RAVS, enabling omnimodal", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2f8ac4f9f80a1661769e23ef3e4c9f8eb4e7ba3b937f9f721149a1410f736e8a.jpg", + "image_caption": [ + "Figure 2. Comparison of (a) Ref-AVS [68] and (b) our proposed OmniAVS. In Ref-AVS, expressions mainly focus on surface-level sound properties, while OmniAVS demands deeper understanding and reasoning about sound content, enabling complex reasoning like identifying potential illness from coughing sounds." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 903, + 186 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "inputs can greatly enhance flexible referring and human-machine interactions. However, existing datasets [17, 68] are limited in scope, supporting only single or isolated modalities in their referring expressions. To address this issue, we include diverse expressions in OmniAVS. Beyond the most common text-only expressions, OmniAVS supports text with sound, text with image, and text with both sound and image. Additionally, we allow text to be replaced by speech*, resulting in 8 different kinds of expression types in OmniAVS, as shown in Figure 1.", + "bbox": [ + 511, + 284, + 906, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Omnimodal Instructed Segmentation Assisant. Given the unique and complex requirements of OmniAVS, which necessitate both flexible interfaces and sophisticated multimodal perception and reasoning capabilities, no existing models can be directly applied to this emerging field. To address this issue, we introduce a baseline model named Omnimodal Instructed Segmentation Assisant, based on Multimodal Large Language Model (MLLM) [13]. This model seamlessly integrates text, speech, sound, and image inputs to perform referring object segmentation while providing explanations for prediction. We identify audio-visual synchronization as a critical factor and implement Audio-Visual Interleaving to achieve temporal alignment without introducing additional parameters. Furthermore, unlike concurrent works [3] that use a single token to segment all frames, we adopt a more lightweight query propagation approach to achieve efficient and accurate segmentation. Experimental results demonstrate that the proposed method OISA possesses strong multimodal reasoning capabilities and supports multiple expression patterns, outperforming other methods on OmniAVS dataset. To further validate the effectiveness of our method OISA, we conduct evaluations on other datasets, including referring video/image segmentation and Ref-AVS, achieving competitive results compared to existing methods across all tasks.", + "bbox": [ + 511, + 436, + 908, + 813 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work makes four main contributions: i) we explore omnimodal referring audio-visual segmentation, which is more aligned with real-world scenarios and provides a solid foundation for developing future omnimodal AI systems; ii)", + "bbox": [ + 511, + 814, + 906, + 875 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "*To clarify, speech is human-produced spoken sound.", + "bbox": [ + 531, + 887, + 815, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22576", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we introduce a large-scale referring audio-visual segmentation dataset OmniAVS that enables advanced multimodal reasoning in audio-visual scenes, with a particular emphasis on deep understanding and reasoning from audio content; iii) we enhance interaction flexibility by incorporating multiple referring modalities in the proposed OmniAVS dataset, broadening the scope of referring AVS; iv) we develop OISA, a multimodal large language model designed to uniformly handle omnimodal expressions and reasoning in RAVS, with extensive experiments across different tasks.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 253, + 232, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio-Visual Scene Perception. Audio-visual scene perception [35-37, 47, 60, 66-68, 72] aims to simulate human multimodal sensory capabilities by integrating auditory and visual information. The rapid development of deep learning has enabled joint audio-visual learning, achieving significant progress in tasks such as event detection [62], robotic manipulation [64], object localization [30], action recognition [8], and object segmentation [68, 86]. The integration of multiple modalities enables models to achieve superior performance compared to single-modality approaches, particularly in handling incomplete or ambiguous information. However, datasets that comprehensively incorporate audio, visual, and language components remain scarce. Among the few available datasets, Music-AVQA [35], AVQA [72], and Ref-AVS [68] stand out as notable examples. Music-AVQA [35] and AVQA [72] primarily address question-answering tasks. Ref-AVS [68] focuses on referring audio-visual segmentation. Nevertheless, Ref-AVS's utilization of audio information is limited, as it employs relatively simple expressions without complex reasoning, thus constraining its effectiveness in real-world scenarios. To overcome these limitations, we introduce OmniAVS, which is specifically designed to maximize the utilization of audio information and seamlessly integrate it with textual descriptions.", + "bbox": [ + 91, + 280, + 483, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reasoning-based Perception. With the rapid development of MLLMs [2, 13, 16, 43-45, 58, 61, 63, 76, 81], the research community has shifted from perception based on simple semantic [17, 71] to perception based on world knowledge or complex reasoning [3, 25, 33, 53, 70, 82]. For example, GPT4RoI [82] prompts MLLM with bounding boxes to solve region-level complex tasks that require reasoning. DetGPT [53] combines a pretrained MLLM with an open-vocabulary object detector to enable reasoning-based object detection from natural language instructions. LISA [33] uses a MLLM [43] to output a [SEG] token, which is then processed by a mask head [32] to generate the masks. Leveraging the world knowledge and reasoning capabilities inherent in LLMs, LISA demonstrates superior performance in complex segmentation scenarios that require complex reasoning. Building upon this foundation, VideoLISA [3] and VISA [70] extend LISA's capabilities to", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "reasoning video object segmentation by incorporating comprehensive temporal modeling. Drawing inspiration from these works, we propose OISA, which extends reasoning capabilities to audio-visual contexts. Unlike previous works that primarily focus on vision-language reasoning, OISA leverages MLLM to perform reasoning about audio content in videos. While AnyRef [25] pioneers the use of audio and images as referring inputs for segmentation, it is constrained to static images and lacks the ability to reason about audio in dynamic video contexts. Moreover, some works [21, 39, 51] extend LLMs to multiple modalities, e.g., vison, audio, and text, but they cannot be used for segmentation tasks.", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Benchmark: OmniAVS", + "text_level": 1, + "bbox": [ + 511, + 286, + 733, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Audio-Visual Video Collection", + "text_level": 1, + "bbox": [ + 511, + 311, + 781, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The video sources consist of three main parts: i) Real-world web videos under the Creative Commons License, where we use techniques from [10, 68, 86] to align audio and visual snippets with intended semantics; ii) The open-source dataset TVQA [34], featuring extensive dialogue from TV shows; iii) Self-recorded videos, with consent from all participants. We select videos from above that meet the following two standards:", + "bbox": [ + 511, + 333, + 903, + 455 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We encourage to select videos that contain audio content with informative or reasoning-based meanings and corresponding identifiable objects in the scene.", + "- We strive to select videos with complex scenes and multiple objects, enabling more diverse referring expressions, e.g., referring to multiple targets, and makes the dataset more representative of real-world complex scenarios." + ], + "bbox": [ + 511, + 457, + 903, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After reviewing about 10,871 potential candidates, we carefully selected the most satisfactory and appropriate videos, which meet the above standards. Finally, we chose 2,104 videos to create the OmniAVS dataset, which is diverse and representative of a wide range of real-world audio-visual scenarios. After collecting the videos, our research team annotates potential objects with expressions and bounding boxes at their first appearance. This information is then used by the annotation team to achieve video mask annotation on the referred objects. Next, we will introduce the corresponding annotation process.", + "bbox": [ + 511, + 566, + 905, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Omnimodal Expression and Mask Annotation", + "text_level": 1, + "bbox": [ + 511, + 742, + 901, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The expression annotation process integrates 4 different modalities: text, speech, sound, and image. We categorize referring expressions into 8 forms: I) Text; II) Speech; III) Text with Sound; IV) Speech with Sound; V) Text with Image; VI) Speech with Image; VII) Text with Sound and Image; and VIII) Speech with Sound and Image. Each expression includes either text or speech, as they provide essential instructions for the model to understand the task and other modalities. All speech expressions are generated", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22577", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/020dc339105349dc4462c93ba5c9f092cf1fd86e99e9bbe74278c5c358697793.jpg", + "table_caption": [ + "Table 1. Statistical comparison between the newly proposed OmniAVS and other datasets of related referring video segmentation tasks. Avail., Expl., and Expr. are abbreviations for Availability of reasoning, Explanations, and Expressions, respectively." + ], + "table_footnote": [], + "table_body": "
DatasetVenueContentReferringReasoningStatistics
VideoAudioTextAudioImageAvail.Expl.VideoFrameObjectMaskExpr.Expl.
Refer-YouTubeVOS [57][ECCV'20]✓X✓XXXX3,978116k7,451131k15,009X
MeViS [17][ICCV'23]✓X✓XXXX2,00644k8,175443k28,570X
ReVOS [70][ECCV'24]✓X✓XX✓X1,042116k5,535469k35,074X
Ref-AVS [68][ECCV'24]✓✓✓XXXX4,00240k6,88878k20,261X
OmniAVS (ours)[ICCV'25]✓✓✓✓✓✓✓2,104103k4,277206k61,09534,841
", + "bbox": [ + 91, + 119, + 901, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by converting the corresponding text expressions into audio. Following previous works [21, 52], we use a combination of several TTS models [23, 50] and human readings for this conversion. In light of this, we will focus on introducing the annotation process for text-related expressions. Figure 3 shows the distribution of different expression types. The inner areas (e.g., text) are composed of their corresponding outer areas (e.g., 4 components of text).", + "bbox": [ + 88, + 256, + 480, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1) Text. This is the main component of OmniAVS. The annotation process follows these rules: 1) Expressions should relate to the sound in the video, not just visual cues. 2) Expressions should emphasize the sound's content, not just the act of making it. For example, instead of \"The dog barking\", use \"The dog warning\" as it requires understanding the sound. 3) Encourage reasoning in expressions, with necessary explanations. 4) Expressions can refer to any number of objects, from none to many (more than one).", + "bbox": [ + 88, + 378, + 482, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "III) Text with Sound. This form combines text and sound to provide a more comprehensive expression. The annotation process follows the same rules as in the Text modality, with the following additional requirements when adding audio: 1) Use natural sounds, instrumental music, or environmental sounds, e.g., dog barking, bird chirping. 2) Text should complement the sound, adding context or instructions for identifying the object(s). An example is shown in the 2nd row of Figure 1.", + "bbox": [ + 88, + 517, + 482, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "V)) Text with Image. This type combines text and image. The annotation follows the rules of Text modality with additional considerations for images: 1) Include relevant images that depict the target object(s) or provide context for the audio-visual content. 2) Ensure the text complements the image, offering instructions for identifying the object(s). An example is shown in the 3rd row of Figure 1.", + "bbox": [ + 88, + 655, + 482, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VII) Text with Sound and Image. This type combines text, sound, and image in the expressions. The text should guide identification using both sound and image. An example is shown in the 4th row of Figure 1.", + "bbox": [ + 88, + 763, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mask Annotation. After collecting expressions, the annotation team proceeds to segment the objects referred to by each expression in every video. Using the bounding box of each object's first appearance (not necessarily in the first frame) annotated in the previous step as a reference,", + "bbox": [ + 88, + 825, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/86812771fe330221065ff4e18cb7dde69c4826e805ec43c81ac0e535b25a054e.jpg", + "table_caption": [ + "Table 2. More dataset statistics." + ], + "table_footnote": [], + "table_body": "
StatisticNumber
Words per Expr.9.3
FPS (mean/min/max)4.8/3/15
Expression Types8
Frames (min/max)11/251
Obj. per Expr. (mean)1.12
Obj. per Expr. (max)7
Duration6.2 h
Videos (train/test)1,864/240
Expr. (train/test)54,304/6,791
", + "bbox": [ + 516, + 268, + 715, + 395 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a7e5b9ec7e2cad5200ca3e187196213bca0260da6ee9c3104d5dc1a79dbc8d16.jpg", + "image_caption": [ + "Figure 3. Distribution of expression types. SI: sound+image." + ], + "image_footnote": [], + "bbox": [ + 750, + 299, + 874, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the annotation team is required to carefully confirm the corresponding object in the video, then track and annotate the object's mask in all frames $\\dagger$ . We designed an interactive annotation tool to automatically load the video and corresponding object bounding boxes. Annotators use the tool to load and preview the corresponding video, then sequentially track and segment target objects in the video. The tool integrates state-of-the-art interactive video object segmentation model SAM2 [56] to assist with mask annotation, enhancing efficiency.", + "bbox": [ + 511, + 411, + 906, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dataset Statistics and Analysis", + "text_level": 1, + "bbox": [ + 511, + 571, + 784, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 shows a statistical comparison between OmniAVS and related datasets. Refer-YouTubeVOS [57], MeViS [17], and ReVOS [70] focus on silent videos without audio and provide only text expressions. While ReVOS [70] further supports reasoning expression, it lacks explanations. Ref-AVS Bench [68] includes audio-visual videos but supports only single-modality text expressions without reasoning or explanations. Compared to previous datasets, the proposed OmniAVS dataset offers multimodal content (audio-visual videos) and diverse expression modalities (text, speech, sound, image), supports reasoning with explanations, and allows for referring to arbitrary number of target objects. We compare more datasets in supplementary.", + "bbox": [ + 511, + 593, + 906, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Main Differences with Ref-AVS. As Ref-AVS [68] is the most related dataset to our proposed OmniAVS, we summarize the differences here to facilitate reference:", + "bbox": [ + 511, + 789, + 905, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Omnimodal Expressions: While Ref-AVS uses only textual referring expressions, OmniAVS supports expres", + "bbox": [ + 511, + 835, + 903, + 866 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ We follow MeViS [17] to perform full-temporal segmentation, even when objects only partially match expressions in specific time segments.", + "bbox": [ + 511, + 875, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22578", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sions that can combine text, speech, sound, image, providing a richer multimodal interactive interface.", + "bbox": [ + 102, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Enhanced Audio-Visual Integration: Unlike Ref-AVS, which primarily focuses on \"the sounding object\", Omni-AVS offers a more comprehensive audio-visual analysis. Our expressions delve deeper into the content of sounds and their contextual impact on the scene.", + "- Reasoning and Explanation: Our dataset includes numerous expressions requiring reasoning, such as \"Who is most likely to be sick?\" (see Figure 2), which necessitates inferring from audio-visual cues like sneezing. We also provide explanations for some reasoning expressions.", + "- Number of Referred Target Objects: Unlike Ref-AVS that only refers to 0 or 1 object per expression, OmniAVS allows each referring expression to indicate an arbitrary number of target objects, i.e., from 0 to many as [41].", + "- Annotation FPS: While Ref-AVS uses fixed video settings (10s, 1 FPS), OmniAVS includes videos of varying durations and annotation frame rates (3-15 FPS) to capture temporal dynamics. Table 2 provides more statistics." + ], + "bbox": [ + 91, + 121, + 482, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Method: OISA", + "text_level": 1, + "bbox": [ + 89, + 411, + 241, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Overview Architecture", + "text_level": 1, + "bbox": [ + 89, + 436, + 302, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Because of the unique nature of this task and the proposed OmniAVS dataset, there is no off-the-shelf model available. We propose an Omnimodal Instructed Segmentation Assistant (OISA) as a baseline method to support omnimodal referring and reasoning audio-visual segmentation. As shown in Figure 4, OISA consists of two main components: a MLLM for multimodal understanding and generation, and a mask head for segmentation and tracking. The MLLM includes audio encoder, vision encoder and LLM, while the mask head consists of ViT-Adapter [12], pixel decoder and mask decoder [14]. It's worth noting that the mask head is flexible and can be replaced with models like SAM [32].", + "bbox": [ + 88, + 460, + 482, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an audio-visual video, the multimodal encoder processes video and audio content to obtain corresponding vision and audio tokens. Different from VideoLLaMA [80] that concatenates these two modalities sequentially, we adopt an Audio-Visual Interleaving strategy, as described in Section 4.2, where we divide audio tokens into clips and interleave them with vision tokens. The interleaved sequence forms our Audio-Visual Content Tokens. This approach effectively synchronizes audio and frames without introducing additional parameters, which is particularly useful for scenarios requiring audio-visual alignment. See examples at Figure 2, where without proper audiovisual alignment it would be difficult to determine which object sneezed when multiple objects are making sounds simultaneously. For omnimodal expression inputs, we employ the audio encoder to process speech and sound, and the vision encoder to handle images. The resulting tokens are then", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/240914d0902495ac7ff33a954f38f4805b028935856e34480c4e60d5aae8db65.jpg", + "image_caption": [ + "Figure 4. Overview of the proposed method OISA. Vision.audio encoders and text embedding are omitted for clarity." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "integrated into the corresponding text tokens to produce the final Omnimodal Expression Tokens. All types of tokens are subsequently fed into the MLLM. The MLLM generates text responses and produces a [SEG] token representing the target object, which is then used by mask decoder for segmentation prediction. Our mask decoder uses query propagation to segment each frame, as described in Section 4.3. The query is refined online during the segmentation process, alleviating dynamic motion. Following previous work [33], we employed cross entropy loss for text generation, DICE loss [49], and binary cross entropy loss for segmentation.", + "bbox": [ + 511, + 333, + 906, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Audio-Visual Interleaving", + "text_level": 1, + "bbox": [ + 511, + 513, + 750, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For videos of arbitrary length, we uniformly sample $N$ frames from the beginning to the end. Each frame is individually processed by the vision encoder to obtain $L_{v}$ vision tokens, denoted as $\\mathbf{V} = \\{v_{1}, v_{2}, \\ldots, v_{N}\\}$ , where $v_{i} \\in \\mathbb{R}^{L_{v} \\times d}$ represents the $L_{v}$ vision tokens of the $i$ -th frame with dimension $d$ . For the audio content in the video, we use an audio encoder to process it and obtain audio tokens $\\mathbf{A} \\in \\mathbb{R}^{L_{A} \\times d}$ , where $L_{A}$ denotes the total number of audio tokens. To achieve audio-visual alignment, we segment the audio into clips corresponding to the frame rate, obtaining $\\{a_{1}, a_{2}, \\ldots, a_{N}\\}$ , where $a_{i} \\in \\mathbb{R}^{L_{a} \\times d}$ and $L_{a} = L_{A} / N$ . Next, we interleave audio token clips and vision tokens to form the audio-visual interleaving token sequence $\\{v_{1}, a_{1}, v_{2}, a_{2}, \\ldots, v_{N}, a_{N}\\}$ . In comparison, VideoLLaMA [15, 80] directly concatenates audio and visual tokens sequentially, $\\{v_{1}, \\ldots, v_{N}, a_{1}, \\ldots, a_{N}\\}$ . However, this approach makes it difficult to align audio frames, especially in segmentation tasks where the video's FPS is not fixed, thereby exacerbating the challenge of audio-visual synchronization. Besides, video-SALMONN [58] fuses audio and visual tokens by concatenating them along the feature dimension and then applying a weight matrix for mapping. This way introduces additional training parameters and causes the fused audio-visual features becoming", + "bbox": [ + 511, + 537, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22579", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4f6e8a66ef50314aa6157eb548a0b8d2d30ffb86c37dbef4db6a93a1ce8b393e.jpg", + "image_caption": [ + "Figure 5. Comparison of different mask decoder forms." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 478, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "misaligned with the original pretrained MLLM model.", + "bbox": [ + 89, + 258, + 450, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Query Propagation", + "text_level": 1, + "bbox": [ + 89, + 282, + 274, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Figure 4, once the Audio-Visual Content Tokens and Omnimodal Expression Tokens are obtained, the MLLM generates the corresponding responses. We employ the [SEG] token, following LISA [33], to represent the object embedding, which is subsequently passed to the mask decoder. While MLLMs excel at multimodal understanding and reasoning, they are not good at segmentation tasks. Previous works [3, 33] attach an additional vision encoder for plain feature extraction and a mask head for segmentation after MLLM, resulting in a redundant and suboptimal design. To address these issues, when extracting vision tokens from video, we simultaneously extract corresponding multi-scale features through ViT Adapter [12], which are then enhanced via pixel decoder [14]. These enhanced features, along with the [SEG] generated by the MLLM, are input into the mask decoder. When using mask decoder for segmentation, VideoLISA [3] employs the same [SEG] for the independent segmentation of each frame, aka One-Token-Seg-ALL (OTSA), as shown in Figure 5 (a). However, this approach has limitations. Previous studies [21, 27, 74, 75, 83] have shown that a single query often fails to adequately represent a target object, especially when rapid motion is present within the video. A single query carries a positional prior, making it difficult to encapsulate dynamic motion processes, such as an object moving from right to left, as shown in Figure 5 (a). This limitation results in ID switches, consistently tracking the target on the right part of the video. To address this, we employ a query propagation mechanism, as shown in Figure 5 (b), which allows the query to be updated online on a frame-by-frame basis [27, 79]. This approach enables the query to smoothly capture the temporal motion of the object while also effectively modeling contextual temporal information.", + "bbox": [ + 89, + 303, + 483, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 815, + 215, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We adopt InternVL2-1B [13] as our MLLM, equipped with InternViT-300M-448px [13] as vision encoder and Qwen2-0.5B-Instruct [1] as LLM. Whisper-large-v3 [54] is used as our audio encoder, with an", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "extra audio MLP to project audio features into LLM feature space. For ViT-Adapter [12], we remove all Injectors to reduce memory consumption and ensure vision tokens remain unaffected. For pixel decoder and mask decoder, we follow the default design of Mask2Former [14] but remove the self-attention module in each transformer block since there is only one query input to the mask decoder. We denote our model with the above settings as OISA-1B.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset Split. We split OmniAVS into training and testing sets, containing 1,864 videos with 54,304 expressions and 240 videos with 6,791 expressions, respectively.", + "bbox": [ + 511, + 212, + 903, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following previous work [17, 18, 20, 68, 77], we adopt $\\mathcal{J}\\& \\mathcal{F}$ as the segmentation evaluation metric, where $\\mathcal{J}$ measures region similarity (IoU) and $\\mathcal{F}$ evaluates contour accuracy. Following [41], for no-target expressions that referring to nothing in the video, $\\mathcal{J}\\& \\mathcal{F}$ is set to 1 if the prediction is empty; otherwise, it is 0. We employ METEOR [4] as evaluation metric for text explanation generation.", + "bbox": [ + 511, + 258, + 906, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Training Details", + "text_level": 1, + "bbox": [ + 511, + 390, + 671, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Audio-Text Alignment. Since InternVL2 does not support audio input natively, we introduce an audio encoder MLP to align audio features with the LLM space. This stage mainly aims to align audio features with LLM space. To achieve this goal, we use Automatic Speech Recognition datasets [9] and Audio Caption datasets [59] for audio-text alignment. During this process, we only train the audio encoder MLP while keeping all other parameters frozen.", + "bbox": [ + 511, + 412, + 905, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Omnimodal Instruct Segmentation Tuning. Following previous works [3, 33, 70, 84], we use multiple datasets for training, including 1) semantic segmentation datasets: ADE20K[85], COCO-Stuff[7], PASCAL-Part [11], PACOLVIS [55]; 2) referring segmentation datasets: RefCOCO, RefCOCO+ [78], RefCOCOg [48], ReasonSeg [33]; 4) referring video segmentation datasets: Refer-YouTube-VOS [57], Refer-DAVIS-17 [31], MeViS [17], ReVOS [70]; 5) audio-visual segmentation datasets: Ref-AVS Bench [68], OmniAVS. We use LoRA [29] to fine-tune LLM, train all parameters of the mask head, and freeze all remaining parameters. For video samples, during training, we uniformly sample 10 frames from the entire video and follow [3] to use 4 frames as dense frames and the remaining frames as sparse frames. During inference, we sample 32 frames with 4 frames as dense frames.", + "bbox": [ + 511, + 534, + 906, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 787, + 666, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablation Study on Audio-Visual Fusion. In Table 3 we evaluate different audio-visual fusion strategies. For attention baseline, we adopt a simple cross-attention block where image tokens serve as key and audio tokens serve as query/value. For weighted sum, we first pad the audio tokens to match the length of vision tokens, then perform", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22580", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/486be35dea0a1729f39117ebe5a1d24f3edd9da04155f39278e94ed7a24f0ee1.jpg", + "table_caption": [ + "Table 3. Ablation study on audio-visual fusion. AVI means Audio-Visual Interleaving. TVQA Split represents the performance evaluated on samples collected from TVQA [34]. These samples demand precise audio-visual alignment and feature a substantial amount of dialogue. See an example in Figure 6 (a)." + ], + "table_footnote": [], + "table_body": "
Fusion TypeSketchTVQA SplitOverall
Attention-37.435.8
Concat36.935.3
Weighted Sum34.831.1
AVI40.839.2
AVI + Concat42.040.5
", + "bbox": [ + 96, + 160, + 475, + 247 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/52a9a8d6c1c2cf50153d4888fe7c156aef481bcf1943fadb25cbbfe03c79bfc0.jpg", + "table_caption": [ + "Table 4. Ablation study of mask head. OTSA: One-Token-Seg-All [3]; QP: query propagation (Ours); M2F: Mask2Former. FPS is tested on one A6000 GPU (batch size=1)." + ], + "table_footnote": [], + "table_body": "
IDQuery TypeHead TypeOmniAVS Metrics
OTSAQPM2FSAMJ&FJFFPS
I✓XX✓38.136.439.84.3
II✓X✓X35.233.836.615.7
IIIX✓X✓41.239.243.24.1
IVX✓✓X40.538.742.312.3
", + "bbox": [ + 96, + 305, + 480, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "weighted sum on them. Attention-based and concatenation [15] methods achieve similar performance around $35 - 37\\%$ $\\mathcal{J}\\& \\mathcal{F}$ . Our proposed Audio-Visual Interleaving (AVI) strategy significantly improves the overall performance to $39.2\\%$ and get $40.2\\%$ in TVQA split. Previous works [13, 44] add a thumbnail after tiling to capture global multimodal features to capture global features more effectively. Similarly, we append original audio tokens A at the end of the audio-visual interleaving sequence. This supplements complete and untruncated audio tokens based on our AVI audio-visual alignment, enabling the MLLM to handle more comprehensive audio tokens and thereby improving the model performance to $40.5\\%$ $\\mathcal{J}\\& \\mathcal{F}$ .", + "bbox": [ + 88, + 410, + 482, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study on Mask Head. In Table 4 we study different mask head designs. While SAM mask head achieves competitive performance compared to M2F, it suffers from slower inference speed due to its huge ViT backbone. When using M2F as the mask head (Setting II and IV), we achieve at least $3 \\times$ speedup with a slight accuracy trade-off. Regarding query types, our QP approach (Setting III and IV) significantly outperforms OTSA in segmentation quality, improving $\\mathcal{J} \\& \\mathcal{F}$ from $38.1\\%$ to $41.2\\%$ when using SAM mask head, and from $35.2\\%$ to $40.5\\%$ when using M2F mask head. Considering the balance between performance and efficiency, we choose Setting IV (QP with M2F head) for our subsequent experiments.", + "bbox": [ + 89, + 607, + 482, + 803 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. OmniAVS Benchmark Results", + "text_level": 1, + "bbox": [ + 89, + 816, + 352, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 5, we report the OmniAVS benchmark results. For traditional referring segmentation methods [17, 68, 71], we process each modality independently and concatenate their features to create a unified query representation.", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f95e3fc287638a93422a9f7307730208dbb1c33e74710a6c286f18d483acd328.jpg", + "table_caption": [ + "Table 5. Testing on OmniAVS. We use $\\mathcal{I}\\&\\mathcal{F}$ as the default metric. All is the average result across 8 splits. MET: METEOR." + ], + "table_footnote": [], + "table_body": "
MethodAllIIIIIIIVVVIVIIVIIIMET.
LMPM [17]25.831.228.720.022.721.320.930.031.4-
EEMC [68]29.634.432.619.626.028.024.735.636.0-
MUTR [71]32.335.433.328.429.826.522.841.640.5-
LISA-7B [33]33.633.331.229.232.728.627.343.443.111.6
LISA-13B [33]36.136.432.130.435.731.630.246.745.716.5
OISA-1B (ours)41.140.138.534.938.535.935.252.653.021.7
", + "bbox": [ + 517, + 119, + 903, + 209 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regarding LISA [33], we enhance its audio processing capability by applying the same audio-text alignment method as our OISA, while maintaining its frame-by-frame inference approach. Our OISA-1B achieves state-of-the-art performance on the OmniAVS benchmark, significantly surpassing previous methods. Specifically, it attains an average score of $41.1\\%$ across all modality combinations, outperforming LISA-13B by $5.0\\%$ . These improvements are consistent across all splits, with the most notable gain of $7.3\\%$ on split VIII (speech+sound+image), showcasing our model's superior ability in multimodal input processing and reasoning. In terms of explanation generation, our model achieves a METEOR score of $21.7\\%$ , significantly surpassing LISA-13B's score of $16.5\\%$ . This indicates that our model excels not only in segmentation tasks but also in understanding and explaining its decisions. Analysis across different modality combinations reveals that splits VII (text+speech+image) and VIII (text+sound+image) achieve the highest scores of $52.6\\%$ and $53.0\\%$ , respectively. This suggests that incorporating more modalities generally enhances performance, as the model effectively leverages complementary information from various sources. Comparison with the previous dataset Ref-AVS Bench shows that our OISA-1B achieved a score of $58.0\\%$ (shown in Table 6), whereas it only attained $41.1\\%$ on our OmniAVS. This significant difference underscores the increased difficulty and complexity of our dataset.", + "bbox": [ + 511, + 244, + 906, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results and Failure Cases. In Figure 6, we present several success and failure cases of the proposed baseline method OISA. Case (a) demonstrates that OISA effectively aligns audio with visual content, identifies key dialogues in the fourth frame, and subsequently performs reasoning and accurately locates the target object, providing a clear explanation. Besides, Case (a) shows that OISA achieves robust tracking in scenarios involving viewpoint changes, occlusions, and reappearance. Case (b) illustrates the model's ability to handle multimodal expressions by understanding both the fire in the image and the siren sound to locate the fire truck. However, as shown in Case (c), OISA struggles in scenarios with complex sounds, indicating that understanding intricate sounds still requires further exploration, such as decoupling sound events [28, 46] before perception and reasoning.", + "bbox": [ + 511, + 657, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22581", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/80dbe08a52221f8e0efaaacb7d611cb6bccaf5daf49b21813c1135dfda3e411c.jpg", + "image_caption": [ + "Figure 6. Success and failure cases of OISA-1B." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 477, + 247 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/08c11c3eee3dd3426f5487a177f14420b66ebe929fb6808a7f5996ec5257a0b0.jpg", + "table_caption": [ + "Table 6. Performance on testing set of Ref-AVS [68]." + ], + "table_footnote": [], + "table_body": "
MethodSeenUnseenMix (S+U)Null
JFJFJFS (↓)
AVSBench [86]23.251.132.454.727.852.920.8
AVGSegFormer [22]33.547.036.150.134.848.617.1
GAVS [65]28.949.829.849.729.449.819.0
ReferFormer [69]31.350.130.448.830.949.517.6
R2VOS [38]25.041.027.949.826.545.418.3
EEMC [68]34.251.349.564.841.958.10.7
OISA-1B (ours)51.758.758.365.154.561.49.8
", + "bbox": [ + 96, + 301, + 477, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Results on Related Tasks", + "text_level": 1, + "bbox": [ + 89, + 449, + 316, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on Ref-AVS. As shown in Table 6, our OISA-1B achieves SOTA performance on Ref-AVS [68], greatly surpassing previous methods. Specifically, OISA-1B improves the $\\mathcal{I}$ by $17.5\\%$ and $8.8\\%$ on Seen and Unseen splits respectively compared to EEMC [68], demonstrating good generalization ability. While OISA-1B gets suboptimal score on Null split, which may inherit from hallucination in MLLM, EEMC achieves $0.7\\%$ $\\mathcal{S}$ . We argue that videos in Null split contain empty or fixed audio with unrelated expressions, so EEMC is likely overfitting to identify audio patterns without understanding. Therefore, our model's performance $(9.8\\%)$ is reasonable, and OISA-1B's advantages on other metrics also demonstrate its robustness.", + "bbox": [ + 89, + 472, + 482, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on Referring Image Segmentation. As shown in Table 7, our OISA-1B outperforms previous referring image segmentation methods on refCOCO series [48, 78], achieving state-of-the-art performance. The performance on ReasonSeg [33] falls short of LISA [33] and VideoLISA [3], likely due to the limited reasoning abilities of our smaller 0.5B LLM, which are crucial for this dataset.", + "bbox": [ + 89, + 671, + 482, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on Referring Video Segmentation. As shown in Table 8, the proposed OISA-1B achieves competitive results on MeViS [17] (43.2%), R-YTVOS [57] (62.1%), and R-DAVIS $_{17}$ [31] (65.2%). On ReVOS [70], it achieves new state-of-the-art results with 47.3% $\\mathcal{I} \\& \\mathcal{F}$ and 19.3% robustness score, surpassing VISA [70] by 0.2% and 4.0% respectively, demonstrating strong perception and reasoning ability in complex video scenes.", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a84841f22072d92c2297bab1679fa10dd84a6dedd300fb40d7498c9eab1863ee.jpg", + "table_caption": [ + "Table 7. cIoU Results on refCOCO [48, 78] and ReasonSeg [33]." + ], + "table_footnote": [], + "table_body": "
MethodrefCOCOrefCOCO+refCOCogReasonSeg
valtestAtestBvaltestAtestBvalUtestUvaltest
LAVT [73]72.775.868.862.168.455.161.262.1--
ReLA [41]73.876.570.266.071.057.765.066.0--
LISA-7B [33]74.176.571.162.467.456.566.468.554.048.4
LISA-13B [33]--------62.951.1
VISA-7B [70]72.475.568.159.864.853.165.566.457.8-
VideoLISA-3.8B [3]73.876.668.863.468.856.268.368.867.154.4
OISA-1B (ours)74.278.372.068.871.762.070.672.361.250.8
", + "bbox": [ + 519, + 104, + 901, + 205 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e4f002580f03febccb1047ae90258fbc370d5d7947fc3af60c3ef7bd361f19ae.jpg", + "table_caption": [ + "Table 8. Results on referring/reasoning video segmentation [17, 31, 57, 70]. $\\mathcal{I}\\&\\mathcal{F}$ is the default metric, $\\mathcal{R}$ is the robustness score." + ], + "table_footnote": [], + "table_body": "
MethodMeViSR-YTVOSR-DAVIS17ReVOSReVOS(R)
ReferFormer [69]31.062.961.114.94.9
MTTR [5]30.055.3-25.55.6
LMPM [17]37.2--26.43.2
LISA-7B [33]-50.258.440.99.3
TrackGPT-7B [87]40.156.463.243.611.6
VideoLISA-3.8B [3]42.361.767.7--
VISA-7B [70]44.563.070.447.115.3
OISA-1B (ours)43.262.165.247.319.3
", + "bbox": [ + 517, + 250, + 903, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 381, + 759, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce OmniAVS, a large-scale dataset supporting advanced multimodal referring and reasoning in audiovisual scenes with 8 types of expressions across text, speech, sound, and image. Compared to previous referring datasets focused on salient video or sound occurrence, OmniAVS enables deeper understanding and reasoning of audiovisual content. To address the challenges posed by OmniAVS, we propose a simple yet effective baseline OISA, a segmentation model based on multimodal large language models. Our OISA leverages audio-visual interleaving for temporal alignment and query propagation for efficient segmentation, enabling effective handling of multimodal inputs without additional parameters. Extensive experiments on 10 datasets demonstrate the effectiveness of our OISA.", + "bbox": [ + 509, + 406, + 906, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future Directions. Although our baseline OISA has shown promising results on OmniAVS, there is still much room for improvement: i) Exploring more efficient audiovisual fusion methods, such as learning more robust audiovisual representations. ii) Disentangling sounds when multiple objects make sounds simultaneously, which is crucial for subsequent reasoning. iii) How to effectively combine multiple modalities in expressions by using joint representations or cross-modal fusion. iv) Exploring more advanced segmentation models to handle complex scenarios, e.g., occlusion, disappearance and reappearance. v) Pre-training OISA on larger audio-visual datasets to further improve its generalization to unseen tasks. vi) Developing more interactive and conversational capabilities to enable multi-round interactions with users, allowing for clarification and refinement of referring expressions.", + "bbox": [ + 511, + 619, + 908, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This project was supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62472104.", + "bbox": [ + 511, + 875, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22582", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen Technical Report. arXiv, 2023. 6", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A Frontier Large Vision-Language Model with Versatile Abilities. arXiv preprint arXiv:2308.12966, 2023. 3", + "[3] Zechen Bai, Tong He, Haiyang Mei, Pichao WANG, Ziteng Gao, Joya Chen, liulei, Zheng Zhang, and Mike Zheng Shou. One Token to Seg Them All: Language Instructed Reasoning Segmentation in Videos. In Adv. Neural Inform. Process. Syst., 2024. 2, 3, 6, 7, 8", + "[4] Satanjeev Banerjee and Alon Lavie. METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments. In Assoc. Comput. Linguist. Worksh., 2005. 6", + "[5] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-End Referring Video Object Segmentation with Multimodal Transformers. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 8", + "[6] Albert S Bregman. Auditory Scene Analysis: The Perceptual Organization of Sound. MIT press, 1994. 2", + "[7] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and Stuff Classes in Context. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 6", + "[8] Jacob Chalk, Jaesung Huh, Evangelos Kazakos, Andrew Zisserman, and Dima Damen. TIM: A Time Interval Machine for Audio-Visual Action Recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3", + "[9] Guoguo Chen, Shuzhou Chai, et al. GigaSpeech: An Evolving, Multi-domain ASR Corpus with 10,000 Hours of Transcribed Audio. In Proc. Interspeech 2021, 2021. 6", + "[10] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. VGGSound: A Large-scale Audio-Visual Dataset. In IEEE Int. Conf. Acoust. Speech Signal Process., 2020. 3", + "[11] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect What You Can: Detecting and Representing Objects using Holistic Models and Body Parts. In IEEE Conf. Comput. Vis. Pattern Recog., 2014. 6", + "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision Transformer Adapter for Dense Predictions. In Int. Conf. Learn. Represent., 2023. 5, 6", + "[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites. arXiv preprint arXiv:2404.16821, 2024. 2, 3, 6, 7", + "[14] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 5, 6" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs. arXiv preprint arXiv:2406.07476, 2024. 5, 7", + "[16] Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, et al. Qwen2-Audio Technical Report. arXiv preprint arXiv:2407.10759, 2024. 3", + "[17] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A Large-scale Benchmark for Video Segmentation with Motion Expressions. In Int. Conf. Comput. Vis., 2023. 2, 3, 4, 6, 7, 8", + "[18] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A New Dataset for Video Object Segmentation in Complex Scenes. In Int. Conf. Comput. Vis., 2023. 6", + "[19] Henghui Ding, Song Tang, Shuting He, Chang Liu, Zuxuan Wu, and Yu-Gang Jiang. Multimodal referring segmentation: A survey. arXiv, 2025. 2", + "[20] Henghui Ding, Kaining Ying, Chang Liu, Shuting He, YuGang Jiang, Philip HS Torr, and Song Bai. MOSEv2: A more challenging dataset for video object segmentation in complex scenes. arXiv, 2025. 6", + "[21] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, et al. VITA: Towards Open-Source Interactive Omni Multimodal LLM. arXiv, 2024. 3, 4, 6", + "[22] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. AVSegFormer: Audio-Visual Segmentation with Transformer. In AAAI, 2024. 8", + "[23] GPT-SoVITS. https://github.com/RVC-Boss/GPT-SoVITS, 2024.4", + "[24] Ruohao Guo, Liao Qu, Dantong Niu, Yanyu Qi, Wenzhen Yue, Ji Shi, Bowei Xing, and Xianghua Ying. Open-Vocabulary Audio-Visual Semantic Segmentation. In ACM Int. Conf. Multimedia, 2024. 2", + "[25] Junwen He, Yifan Wang, Lijun Wang, Huchuan Lu, JunYan He, Jin-Peng Lan, Bin Luo, and Xuansong Xie. Multimodal Instruction Tuned LLMs with Fine-grained Visual Perception. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3", + "[26] Shuting He and Henghui Ding. Decoupling Static and Hierarchical Motion Perception for Referring Video Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2", + "[27] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A Generalized Framework for Video Instance Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 6", + "[28] John R Hershey, Zhuo Chen, Jonathan Le Roux, and Shinji Watanabe. Deep clustering: Discriminative embeddings for segmentation and separation. In IEEE Int. Conf. Acoust. Speech Signal Process., 2016. 7", + "[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In Int. Conf. Learn. Represent., 2022. 6" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22583", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Egocentric Audio-Visual Object Localization. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3", + "[31] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video Object Segmentation with Language Referring Expressions. In ACCV, 2019. 2, 6, 8", + "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment Anything. In Int. Conf. Comput. Vis., 2023. 3, 5", + "[33] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: Reasoning Segmentation via Large Language Model. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 5, 6, 7, 8", + "[34] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. TVQA: Localized, Compositional Video Question Answering. In Proc. of the Conf. on Empirical Methods in Nat. Lang. Process., 2018. 3, 7", + "[35] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to Answer Questions in Dynamic Audio-Visual Scenarios. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3", + "[36] Guangyao Li, Henghui Du, and Di Hu. Boosting Audio Visual Question Answering via Key Semantic-Aware Cues. In ACM Int. Conf. Multimedia, pages 5997-6005, 2024.", + "[37] Jia Li, Wenjie Zhao, Ziru Huang, Yunhui Guo, and Yapeng Tian. Do Audio-Visual Segmentation Models Truly Segment Sounding Objects? arXiv, 2025. 3", + "[38] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust Referring Video Object Segmentation with Cyclic Structural Consensus. In Int. Conf. Comput. Vis., 2023. 8", + "[39] Yadong Li, Haoze Sun, Mingan Lin, Tianpeng Li, Guosheng Dong, Tao Zhang, Bowen Ding, Wei Song, Zhenglin Cheng, Yuqi Huo, et al. Baichuan-Omni Technical Report. arXiv preprint arXiv:2410.08565, 2024. 3", + "[40] Linfeng Yuan and Miaojing Shi and Zijie Yue and Qijun Chen. Losh: Long-short text joint prediction network for referring video object segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2", + "[41] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized Referring Expression Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 5, 6, 8", + "[42] Chang Liu, Xudong Jiang, and Henghui Ding. Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence, 2024. 2", + "[43] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. In Adv. Neural Inform. Process. Syst., 2023. 3", + "[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 7", + "[45] Shuo Liu, Kaining Ying, Hao Zhang, Yue Yang, Yuqi Lin, Tianle Zhang, Chuanhao Li, Yu Qiao, Ping Luo, Wenqi Shao, et al. ConvBench: A Multi-Turn Conversation Evaluation Benchmark with Hierarchical Ablation Capability for Large Vision-Language Models. In Adv. Neural Inform. Process. Syst. D&B, 2024. 3" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Yi Luo and Nima Mesgarani. Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation. IEEE/ACM Trans. Audio Speech Lang. Process., 27 (8), 2019. 7", + "[47] Juncheng Ma, Peiwen Sun, Yaoting Wang, and Di Hu. Stepping Stones: A Progressive Training Strategy for Audio-Visual Semantic Segmentation. In Eur. Conf. Comput. Vis., 2024. 3", + "[48] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and Comprehension of Unambiguous Object Descriptions. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 6, 8", + "[49] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In IEEE Int. Conf. 3D Vis., 2016. 5", + "[50] OpenAI. https://platform.openai.com/docs/guides/text-to-speech, 2023.4", + "[51] OpenAI. https://openai.com/index/hello-gpt-4o, 2024.2,3", + "[52] Wenwen Pan, Haonan Shi, Zhou Zhao, Jieming Zhu, Xi-uqiang He, Zhigeng Pan, Lianli Gao, Jun Yu, Fei Wu, and Qi Tian. Wnet: Audio-Guided Video Object Segmentation via Wavelet-Based Cross-Modal Denoising Networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 4", + "[53] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, et al. DetGPT: Detect What You Need via Reasoning. In Proc. of the Conf. on Empirical Methods in Nat. Lang. Process., 2023. 2, 3", + "[54] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust Speech Recognition via Large-Scale Weak Supervision. In Int. Conf. Mach. Learn., 2023. 6", + "[55] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. PACO: Parts and Attributes of Common Objects. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 6", + "[56] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. SAM 2: Segment Anything in Images and Videos. arXiv preprint arXiv:2408.00714, 2024. 4", + "[57] Seonguk Seo, Joon-Young Lee, and Bohyung Han. URVOS: Unified Referring Video Object Segmentation Network with a Large-Scale Benchmark. In Eur. Conf. Comput. Vis., 2020. 2, 4, 6, 8", + "[58] Guangzhi Sun, Wenyi Yu, Changli Tang, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, Yuxuan Wang, and Chao Zhang. video-SALMONN: Speech-Enhanced Audio-Visual Large Language Models. In Int. Conf. Mach. Learn., 2024. 3, 5", + "[59] Luoyi Sun, Xuanan Xu, Mengyue Wu, and Weidi Xie. Auto-ACD: A Large-scale Dataset for Audio-Language Representation Learning. In ACM Int. Conf. Multimedia, 2024. 6" + ], + "bbox": [ + 514, + 92, + 905, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22584", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Peiwen Sun, Honggang Zhang, and Di Hu. Unveiling and Mitigating Bias in Audio Visual Segmentation. In ACM Int. Conf. Multimedia, 2024. 3", + "[61] Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. SALMONN: Towards Generic Hearing Abilities for Large Language Models. In Int. Conf. Learn. Represent., 2023. 3", + "[62] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-Visual Event Localization in Unconstrained Videos. In Eur. Conf. Comput. Vis., 2018. 3", + "[63] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191, 2024. 3", + "[64] Yefei Wang, Kaili Wang, Yi Wang, Di Guo, Huaping Liu, and Fuchun Sun. Audio-Visual Grounding Referring Expression for Robotic Manipulation. In IEEE Int. Conf. Robot. Autom., 2022. 3", + "[65] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting Segmentation with Sound Is Generalizable Audio-Visual Source Localization. In AAAI, 2024. 8", + "[66] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting Segmentation with Sound Is Generalizable Audio-Visual Source Localization. In AAAI, 2024. 3", + "[67] Yaoting Wang, Peiwen Sun, Yuanchao Li, Honggang Zhang, and Di Hu. Can Textual Semantics Mitigate Sounding Object Segmentation Preference? In Eur. Conf. Comput. Vis., 2024.", + "[68] Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. Ref-AVS: Refer and Segment Objects in Audio-Visual Scenes. In Eur. Conf. Comput. Vis., 2024. 2, 3, 4, 6, 7, 8", + "[69] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as Queries for Referring Video Object Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 8", + "[70] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. VISA: Reasoning Video Object Segmentation via Large Language Models. In Eur. Conf. Comput. Vis., 2024. 2, 3, 4, 6, 8", + "[71] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by Multi-Modality: A Unified Temporal Transformer for Video Object Segmentation. In AAAI, 2024. 3, 7", + "[72] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. AVQA: A Dataset for Audio-Visual Question Answering on Videos. In ACM Int. Conf. Multimedia, 2022. 3", + "[73] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. LAVT: Language-Aware Vision Transformer for Referring Image Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 8", + "[74] Kaining Ying, Zhenhua Wang, Cong Bai, and Pengfei Zhou. Isda: Position-aware instance segmentation with deformable" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "attention. In IEEE Int. Conf. Acoust. Speech Signal Process., 2022. 6", + "[75] Kaining Ying, Qing Zhong, Weian Mao, Zhenhua Wang, Hao Chen, Lin Yuanbo Wu, Yifan Liu, Chengxiang Fan, Yunzhi Zhuge, and Chunhua Shen. CTVIS: Consistent Training for Online Video Instance Segmentation. In Int. Conf. Comput. Vis., 2023. 6", + "[76] Kaining Ying, Fanqing Meng, Jin Wang, Zhiqian Li, Han Lin, Yue Yang, Hao Zhang, Wenbo Zhang, Yuqi Lin, Shuo Liu, Jiayi Lei, Quanfeng Lu, Runjian Chen, Peng Xu, Renrui Zhang, Haozhe Zhang, Peng Gao, Yali Wang, Yu Qiao, Ping Luo, Kaipeng Zhang, and Wenqi Shao. MMT-Bench: A Comprehensive Multimodal Benchmark for Evaluating Large Vision-Language Models Towards Multitask AGI. In Int. Conf. Mach. Learn., 2024. 3", + "[77] Kaining Ying, Hengrui Hu, and Henghui Ding. MOVE: Motion-guided few-shot video object segmentation. In Int. Conf. Comput. Vis., 2025. 6", + "[78] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling Context in Referring Expressions. In Eur. Conf. Comput. Vis., 2016. 6, 8", + "[79] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. MOTR: End-to-End Multiple-Object Tracking with Transformer. In Eur. Conf. Comput. Vis., 2022. 6", + "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding. In Proc. of the Conf. on Empirical Methods in Nat. Lang. Process., 2023. 5", + "[81] Renrui Zhang, Jiaming Han, Chris Liu, Aojun Zhou, Pan Lu, Yu Qiao, Hongsheng Li, and Peng Gao. LLaMA-Adapter: Efficient Fine-tuning of Large Language Models with Zero-initialized Attention. In Int. Conf. Learn. Represent., 2024. 3", + "[82] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Yu Liu, Kai Chen, and Ping Luo. GPT4RoI: Instruction Tuning Large Language Model on Region-of-Interest. arXiv preprint arXiv:2307.03601, 2023. 3", + "[83] Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. DVIS: Decoupled Video Instance Segmentation Framework. In Int. Conf. Comput. Vis., 2023. 6", + "[84] Rongkun Zheng, Lu Qi, Xi Chen, Yi Wang, Kun Wang, Yu Qiao, and Hengshuang Zhao. ViLLa: Video Reasoning Segmentation with Large Language Model. arXiv preprint arXiv:2407.14500, 2024. 2, 6", + "[85] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene Parsing through ADE20K Dataset. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 6", + "[86] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-Visual Segmentation. In Eur. Conf. Comput. Vis., 2022. 2, 3, 8", + "[87] Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with Human-Intent Reasoning. arXiv, 2023. 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22585", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/7d9586d8-b2bc-4332-a23e-c79a83d7ffa9_model.json b/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/7d9586d8-b2bc-4332-a23e-c79a83d7ffa9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..99bf091674900eac6babc00376ae333435c5cbfd --- /dev/null +++ b/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/7d9586d8-b2bc-4332-a23e-c79a83d7ffa9_model.json @@ -0,0 +1,2774 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.239, + 0.131, + 0.761, + 0.177 + ], + "angle": 0, + "content": "Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.203, + 0.77, + 0.263 + ], + "angle": 0, + "content": "Kaining Ying Henghui Ding Guangquan Jie Yu-Gang Jiang Fudan University, China https://henghuiding.com/OmniAVS/" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.298, + 0.257, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.368, + 0.123, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.368, + 0.151, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.371, + 0.331, + 0.384 + ], + "angle": 0, + "content": "Which object mimics another?" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.298, + 0.417, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.368, + 0.417, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.459, + 0.417, + 0.473 + ], + "angle": 0, + "content": "e same language as :" + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.298, + 0.576, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.371, + 0.888, + 0.384 + ], + "angle": 0, + "content": "Output: It is [SEG]. The girl is imitating the barking of the dog on the right." + }, + { + "type": "image", + "bbox": [ + 0.417, + 0.388, + 0.576, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.461, + 0.892, + 0.474 + ], + "angle": 0, + "content": "Output: It is [SEG]. They both speak Korean." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.388, + 0.257, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.458, + 0.123, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.459, + 0.473, + 0.475 + ], + "angle": 0, + "content": "The one speak the same language as ." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.477, + 0.417, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.546, + 0.417, + 0.563 + ], + "angle": 0, + "content": " + serves as lead" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.477, + 0.576, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.477, + 0.736, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.477, + 0.897, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.55, + 0.123, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.55, + 0.421, + 0.564 + ], + "angle": 0, + "content": "The one similar to serves as lead." + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.551, + 0.892, + 0.563 + ], + "angle": 0, + "content": "Output: It is [SEG]. The guitar on the right serves as lead." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.566, + 0.257, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.637, + 0.123, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.638, + 0.151, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.643, + 0.485, + 0.666 + ], + "angle": 0, + "content": "The instrument is played by someone of the same gender and makes a similar sound in ." + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.566, + 0.417, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.637, + 0.417, + 0.666 + ], + "angle": 0, + "content": "is played by someone of the similar sound in \\(<\\)sound: \\(\\mathbb{N}^{\\prime}\\)Sound" + }, + { + "type": "image", + "bbox": [ + 0.418, + 0.566, + 0.576, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.473, + 0.638, + 0.587, + 0.666 + ], + "angle": 0, + "content": "er in >." + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.566, + 0.736, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.638, + 0.615, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.566, + 0.897, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.642, + 0.89, + 0.666 + ], + "angle": 0, + "content": "Output: It is [SEG]. The pipa held by the man on the right." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.669, + 0.907, + 0.756 + ], + "angle": 0, + "content": "Figure 1. Examples of the proposed benchmark Omnimodal Referring Audio-Visual Segmentation (OmniAVS) to show its nature and flexibility. OmniAVS introduces 3 key features: 1) It supports diverse multimodal referring expressions that flexibly combine text, speech, sound, and image for referring audio-visual segmentation; 2) It emphasizes understanding the content of audio rather than merely hearing them; 3) It incorporates complex reasoning and world knowledge in expressions, prompting models to provide explanations for their segmentation decisions. These characteristics make OmniAVS practical for real-world use and well-suited for developing omnimodal models with fine-grained perception." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.77, + 0.327, + 0.787 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.805, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Referring audio-visual segmentation (RAVS) has recently seen significant advancements, yet challenges remain in integrating multimodal information and deeply understand" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Henghui Ding (henghui.ding@gmail.com) is the corresponding author with the Institute of Big Data, College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.772, + 0.907, + 0.893 + ], + "angle": 0, + "content": "ing and reasoning about audiovisual content. To extend the boundaries of RAVS and facilitate future research in this field, we propose Omnimodal Referring AudioVisual Segmentation (OmniAVS), a new dataset containing 2,104 videos and 61,095 multimodal referring expressions. OmniAVS stands out with three key innovations: (1) 8 types of multimodal expressions that flexibly combine text, speech, sound, and visual cues; (2) an emphasis" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22575" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.483, + 0.244 + ], + "angle": 0, + "content": "on understanding audio content beyond just detecting their presence; and (3) the inclusion of complex reasoning and world knowledge in expressions. Furthermore, we introduce Omnimodal Instructed Segmentation Assistant (OISA), to address the challenges of multimodal reasoning and fine-grained understanding of audiovisual content in OmniAVS. OISA uses MLLM to comprehend complex cues and perform reasoning-based segmentation. Extensive experiments show that OISA outperforms existing methods on OmniAVS and achieves competitive results on other related tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.26, + 0.223, + 0.275 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.483, + 0.447 + ], + "angle": 0, + "content": "Referring Audio-Visual Segmentation (RAVS) [19, 68] is an emerging field focused on segmenting target objects indicated by referring expressions (e.g., text) within audio-visual scenes using audio-visual cues. The combination of referring and audio distinguishes RAVS from Referring Video Object Segmentation (RVOS) [17, 26, 31, 40, 42, 57], which uses text descriptions to refer to objects in silent videos, and Audio-Visual Segmentation (AVS) [24, 86], which segments any sound-producing object without referring expressions. RAVS enables practical applications in video conferencing and robotics, etc." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Understanding Rather than Just Hearing. However, the association between expressions and audio cues in the previous RAVS dataset Ref-AVS [68] remains superficial, primarily limited to basic acoustic properties such as sound occurrence, intensity, and temporal sequence. For example, expressions like \"Who is making the loudest sound?\" only address the surface-level characteristics without delving into the content of the sounds. Recent works [3, 33, 53, 70, 84] in image and video segmentation have evolved from simple semantic perception to advanced reasoning understanding. However, such progress is notably absent in audio-visual segmentation. This gap naturally leads us to explore the potential for complex reasoning in audio-visual scenes. Besides, drawing insights from human auditory perception studies [6], understanding sound content is essential for sophisticated audio-visual scene reasoning. To achieve these goals, we present OmniAVS dataset to advance reasoning-based segmentation in audio-visual scenes. As shown in Figure 2(b), our dataset contains expressions that demand reasoning about audio content, such as \"Who is most likely to be sick?\". These expressions establish complex cognitive chains that go beyond basic acoustic features, e.g., sound \\(\\rightarrow\\) coughing \\(\\rightarrow\\) sickness. Furthermore, we enhance the dataset's interpretability by providing detailed explanations for reasoning-based expressions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Towards Omnimodal Referring. The emergence of flagship model ChatGPT-4o [51] sets a new standard for modern AI systems, emphasizing the need to process arbitrary combinations of omnimodal inputs, e.g., audio, vision, and text. In the context of RAVS, enabling omnimodal" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.193, + 0.907, + 0.263 + ], + "angle": 0, + "content": "Figure 2. Comparison of (a) Ref-AVS [68] and (b) our proposed OmniAVS. In Ref-AVS, expressions mainly focus on surface-level sound properties, while OmniAVS demands deeper understanding and reasoning about sound content, enabling complex reasoning like identifying potential illness from coughing sounds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.907, + 0.436 + ], + "angle": 0, + "content": "inputs can greatly enhance flexible referring and human-machine interactions. However, existing datasets [17, 68] are limited in scope, supporting only single or isolated modalities in their referring expressions. To address this issue, we include diverse expressions in OmniAVS. Beyond the most common text-only expressions, OmniAVS supports text with sound, text with image, and text with both sound and image. Additionally, we allow text to be replaced by speech*, resulting in 8 different kinds of expression types in OmniAVS, as shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.438, + 0.909, + 0.814 + ], + "angle": 0, + "content": "Omnimodal Instructed Segmentation Assisant. Given the unique and complex requirements of OmniAVS, which necessitate both flexible interfaces and sophisticated multimodal perception and reasoning capabilities, no existing models can be directly applied to this emerging field. To address this issue, we introduce a baseline model named Omnimodal Instructed Segmentation Assisant, based on Multimodal Large Language Model (MLLM) [13]. This model seamlessly integrates text, speech, sound, and image inputs to perform referring object segmentation while providing explanations for prediction. We identify audio-visual synchronization as a critical factor and implement Audio-Visual Interleaving to achieve temporal alignment without introducing additional parameters. Furthermore, unlike concurrent works [3] that use a single token to segment all frames, we adopt a more lightweight query propagation approach to achieve efficient and accurate segmentation. Experimental results demonstrate that the proposed method OISA possesses strong multimodal reasoning capabilities and supports multiple expression patterns, outperforming other methods on OmniAVS dataset. To further validate the effectiveness of our method OISA, we conduct evaluations on other datasets, including referring video/image segmentation and Ref-AVS, achieving competitive results compared to existing methods across all tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.815, + 0.908, + 0.876 + ], + "angle": 0, + "content": "This work makes four main contributions: i) we explore omnimodal referring audio-visual segmentation, which is more aligned with real-world scenarios and provides a solid foundation for developing future omnimodal AI systems; ii)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.888, + 0.816, + 0.901 + ], + "angle": 0, + "content": "*To clarify, speech is human-produced spoken sound." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22576" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "we introduce a large-scale referring audio-visual segmentation dataset OmniAVS that enables advanced multimodal reasoning in audio-visual scenes, with a particular emphasis on deep understanding and reasoning from audio content; iii) we enhance interaction flexibility by incorporating multiple referring modalities in the proposed OmniAVS dataset, broadening the scope of referring AVS; iv) we develop OISA, a multimodal large language model designed to uniformly handle omnimodal expressions and reasoning in RAVS, with extensive experiments across different tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.254, + 0.233, + 0.27 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.281, + 0.485, + 0.643 + ], + "angle": 0, + "content": "Audio-Visual Scene Perception. Audio-visual scene perception [35-37, 47, 60, 66-68, 72] aims to simulate human multimodal sensory capabilities by integrating auditory and visual information. The rapid development of deep learning has enabled joint audio-visual learning, achieving significant progress in tasks such as event detection [62], robotic manipulation [64], object localization [30], action recognition [8], and object segmentation [68, 86]. The integration of multiple modalities enables models to achieve superior performance compared to single-modality approaches, particularly in handling incomplete or ambiguous information. However, datasets that comprehensively incorporate audio, visual, and language components remain scarce. Among the few available datasets, Music-AVQA [35], AVQA [72], and Ref-AVS [68] stand out as notable examples. Music-AVQA [35] and AVQA [72] primarily address question-answering tasks. Ref-AVS [68] focuses on referring audio-visual segmentation. Nevertheless, Ref-AVS's utilization of audio information is limited, as it employs relatively simple expressions without complex reasoning, thus constraining its effectiveness in real-world scenarios. To overcome these limitations, we introduce OmniAVS, which is specifically designed to maximize the utilization of audio information and seamlessly integrate it with textual descriptions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Reasoning-based Perception. With the rapid development of MLLMs [2, 13, 16, 43-45, 58, 61, 63, 76, 81], the research community has shifted from perception based on simple semantic [17, 71] to perception based on world knowledge or complex reasoning [3, 25, 33, 53, 70, 82]. For example, GPT4RoI [82] prompts MLLM with bounding boxes to solve region-level complex tasks that require reasoning. DetGPT [53] combines a pretrained MLLM with an open-vocabulary object detector to enable reasoning-based object detection from natural language instructions. LISA [33] uses a MLLM [43] to output a [SEG] token, which is then processed by a mask head [32] to generate the masks. Leveraging the world knowledge and reasoning capabilities inherent in LLMs, LISA demonstrates superior performance in complex segmentation scenarios that require complex reasoning. Building upon this foundation, VideoLISA [3] and VISA [70] extend LISA's capabilities to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "reasoning video object segmentation by incorporating comprehensive temporal modeling. Drawing inspiration from these works, we propose OISA, which extends reasoning capabilities to audio-visual contexts. Unlike previous works that primarily focus on vision-language reasoning, OISA leverages MLLM to perform reasoning about audio content in videos. While AnyRef [25] pioneers the use of audio and images as referring inputs for segmentation, it is constrained to static images and lacks the ability to reason about audio in dynamic video contexts. Moreover, some works [21, 39, 51] extend LLMs to multiple modalities, e.g., vison, audio, and text, but they cannot be used for segmentation tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.287, + 0.734, + 0.303 + ], + "angle": 0, + "content": "3. Benchmark: OmniAVS" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.312, + 0.782, + 0.327 + ], + "angle": 0, + "content": "3.1. Audio-Visual Video Collection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.905, + 0.456 + ], + "angle": 0, + "content": "The video sources consist of three main parts: i) Real-world web videos under the Creative Commons License, where we use techniques from [10, 68, 86] to align audio and visual snippets with intended semantics; ii) The open-source dataset TVQA [34], featuring extensive dialogue from TV shows; iii) Self-recorded videos, with consent from all participants. We select videos from above that meet the following two standards:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.458, + 0.905, + 0.503 + ], + "angle": 0, + "content": "- We encourage to select videos that contain audio content with informative or reasoning-based meanings and corresponding identifiable objects in the scene." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.504, + 0.905, + 0.564 + ], + "angle": 0, + "content": "- We strive to select videos with complex scenes and multiple objects, enabling more diverse referring expressions, e.g., referring to multiple targets, and makes the dataset more representative of real-world complex scenarios." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.458, + 0.905, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.906, + 0.734 + ], + "angle": 0, + "content": "After reviewing about 10,871 potential candidates, we carefully selected the most satisfactory and appropriate videos, which meet the above standards. Finally, we chose 2,104 videos to create the OmniAVS dataset, which is diverse and representative of a wide range of real-world audio-visual scenarios. After collecting the videos, our research team annotates potential objects with expressions and bounding boxes at their first appearance. This information is then used by the annotation team to achieve video mask annotation on the referred objects. Next, we will introduce the corresponding annotation process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.902, + 0.759 + ], + "angle": 0, + "content": "3.2. Omnimodal Expression and Mask Annotation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The expression annotation process integrates 4 different modalities: text, speech, sound, and image. We categorize referring expressions into 8 forms: I) Text; II) Speech; III) Text with Sound; IV) Speech with Sound; V) Text with Image; VI) Speech with Image; VII) Text with Sound and Image; and VIII) Speech with Sound and Image. Each expression includes either text or speech, as they provide essential instructions for the model to understand the task and other modalities. All speech expressions are generated" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22577" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 1. Statistical comparison between the newly proposed OmniAVS and other datasets of related referring video segmentation tasks. Avail., Expl., and Expr. are abbreviations for Availability of reasoning, Explanations, and Expressions, respectively." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.12, + 0.902, + 0.231 + ], + "angle": 0, + "content": "
DatasetVenueContentReferringReasoningStatistics
VideoAudioTextAudioImageAvail.Expl.VideoFrameObjectMaskExpr.Expl.
Refer-YouTubeVOS [57][ECCV'20]✓X✓XXXX3,978116k7,451131k15,009X
MeViS [17][ICCV'23]✓X✓XXXX2,00644k8,175443k28,570X
ReVOS [70][ECCV'24]✓X✓XX✓X1,042116k5,535469k35,074X
Ref-AVS [68][ECCV'24]✓✓✓XXXX4,00240k6,88878k20,261X
OmniAVS (ours)[ICCV'25]✓✓✓✓✓✓✓2,104103k4,277206k61,09534,841
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.257, + 0.482, + 0.378 + ], + "angle": 0, + "content": "by converting the corresponding text expressions into audio. Following previous works [21, 52], we use a combination of several TTS models [23, 50] and human readings for this conversion. In light of this, we will focus on introducing the annotation process for text-related expressions. Figure 3 shows the distribution of different expression types. The inner areas (e.g., text) are composed of their corresponding outer areas (e.g., 4 components of text)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.38, + 0.483, + 0.516 + ], + "angle": 0, + "content": "1) Text. This is the main component of OmniAVS. The annotation process follows these rules: 1) Expressions should relate to the sound in the video, not just visual cues. 2) Expressions should emphasize the sound's content, not just the act of making it. For example, instead of \"The dog barking\", use \"The dog warning\" as it requires understanding the sound. 3) Encourage reasoning in expressions, with necessary explanations. 4) Expressions can refer to any number of objects, from none to many (more than one)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.518, + 0.483, + 0.653 + ], + "angle": 0, + "content": "III) Text with Sound. This form combines text and sound to provide a more comprehensive expression. The annotation process follows the same rules as in the Text modality, with the following additional requirements when adding audio: 1) Use natural sounds, instrumental music, or environmental sounds, e.g., dog barking, bird chirping. 2) Text should complement the sound, adding context or instructions for identifying the object(s). An example is shown in the 2nd row of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.656, + 0.483, + 0.762 + ], + "angle": 0, + "content": "V)) Text with Image. This type combines text and image. The annotation follows the rules of Text modality with additional considerations for images: 1) Include relevant images that depict the target object(s) or provide context for the audio-visual content. 2) Ensure the text complements the image, offering instructions for identifying the object(s). An example is shown in the 3rd row of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.764, + 0.483, + 0.825 + ], + "angle": 0, + "content": "VII) Text with Sound and Image. This type combines text, sound, and image in the expressions. The text should guide identification using both sound and image. An example is shown in the 4th row of Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Mask Annotation. After collecting expressions, the annotation team proceeds to segment the objects referred to by each expression in every video. Using the bounding box of each object's first appearance (not necessarily in the first frame) annotated in the previous step as a reference," + }, + { + "type": "table_caption", + "bbox": [ + 0.52, + 0.255, + 0.71, + 0.267 + ], + "angle": 0, + "content": "Table 2. More dataset statistics." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.27, + 0.716, + 0.396 + ], + "angle": 0, + "content": "
StatisticNumber
Words per Expr.9.3
FPS (mean/min/max)4.8/3/15
Expression Types8
Frames (min/max)11/251
Obj. per Expr. (mean)1.12
Obj. per Expr. (max)7
Duration6.2 h
Videos (train/test)1,864/240
Expr. (train/test)54,304/6,791
" + }, + { + "type": "image_caption", + "bbox": [ + 0.738, + 0.256, + 0.887, + 0.298 + ], + "angle": 0, + "content": "Figure 3. Distribution of expression types. SI: sound+image." + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.3, + 0.875, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.412, + 0.907, + 0.562 + ], + "angle": 0, + "content": "the annotation team is required to carefully confirm the corresponding object in the video, then track and annotate the object's mask in all frames \\(\\dagger\\). We designed an interactive annotation tool to automatically load the video and corresponding object bounding boxes. Annotators use the tool to load and preview the corresponding video, then sequentially track and segment target objects in the video. The tool integrates state-of-the-art interactive video object segmentation model SAM2 [56] to assist with mask annotation, enhancing efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.785, + 0.587 + ], + "angle": 0, + "content": "3.3. Dataset Statistics and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.789 + ], + "angle": 0, + "content": "Table 1 shows a statistical comparison between OmniAVS and related datasets. Refer-YouTubeVOS [57], MeViS [17], and ReVOS [70] focus on silent videos without audio and provide only text expressions. While ReVOS [70] further supports reasoning expression, it lacks explanations. Ref-AVS Bench [68] includes audio-visual videos but supports only single-modality text expressions without reasoning or explanations. Compared to previous datasets, the proposed OmniAVS dataset offers multimodal content (audio-visual videos) and diverse expression modalities (text, speech, sound, image), supports reasoning with explanations, and allows for referring to arbitrary number of target objects. We compare more datasets in supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.79, + 0.906, + 0.835 + ], + "angle": 0, + "content": "Main Differences with Ref-AVS. As Ref-AVS [68] is the most related dataset to our proposed OmniAVS, we summarize the differences here to facilitate reference:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.836, + 0.905, + 0.867 + ], + "angle": 0, + "content": "- Omnimodal Expressions: While Ref-AVS uses only textual referring expressions, OmniAVS supports expres" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) We follow MeViS [17] to perform full-temporal segmentation, even when objects only partially match expressions in specific time segments." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22578" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "sions that can combine text, speech, sound, image, providing a richer multimodal interactive interface." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.122, + 0.483, + 0.196 + ], + "angle": 0, + "content": "- Enhanced Audio-Visual Integration: Unlike Ref-AVS, which primarily focuses on \"the sounding object\", Omni-AVS offers a more comprehensive audio-visual analysis. Our expressions delve deeper into the content of sounds and their contextual impact on the scene." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.198, + 0.483, + 0.273 + ], + "angle": 0, + "content": "- Reasoning and Explanation: Our dataset includes numerous expressions requiring reasoning, such as \"Who is most likely to be sick?\" (see Figure 2), which necessitates inferring from audio-visual cues like sneezing. We also provide explanations for some reasoning expressions." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.273, + 0.483, + 0.332 + ], + "angle": 0, + "content": "- Number of Referred Target Objects: Unlike Ref-AVS that only refers to 0 or 1 object per expression, OmniAVS allows each referring expression to indicate an arbitrary number of target objects, i.e., from 0 to many as [41]." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.334, + 0.483, + 0.395 + ], + "angle": 0, + "content": "- Annotation FPS: While Ref-AVS uses fixed video settings (10s, 1 FPS), OmniAVS includes videos of varying durations and annotation frame rates (3-15 FPS) to capture temporal dynamics. Table 2 provides more statistics." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.122, + 0.483, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.412, + 0.242, + 0.427 + ], + "angle": 0, + "content": "4. Method: OISA" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.438, + 0.303, + 0.453 + ], + "angle": 0, + "content": "4.1. Overview Architecture" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.483, + 0.643 + ], + "angle": 0, + "content": "Because of the unique nature of this task and the proposed OmniAVS dataset, there is no off-the-shelf model available. We propose an Omnimodal Instructed Segmentation Assistant (OISA) as a baseline method to support omnimodal referring and reasoning audio-visual segmentation. As shown in Figure 4, OISA consists of two main components: a MLLM for multimodal understanding and generation, and a mask head for segmentation and tracking. The MLLM includes audio encoder, vision encoder and LLM, while the mask head consists of ViT-Adapter [12], pixel decoder and mask decoder [14]. It's worth noting that the mask head is flexible and can be replaced with models like SAM [32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Given an audio-visual video, the multimodal encoder processes video and audio content to obtain corresponding vision and audio tokens. Different from VideoLLaMA [80] that concatenates these two modalities sequentially, we adopt an Audio-Visual Interleaving strategy, as described in Section 4.2, where we divide audio tokens into clips and interleave them with vision tokens. The interleaved sequence forms our Audio-Visual Content Tokens. This approach effectively synchronizes audio and frames without introducing additional parameters, which is particularly useful for scenarios requiring audio-visual alignment. See examples at Figure 2, where without proper audiovisual alignment it would be difficult to determine which object sneezed when multiple objects are making sounds simultaneously. For omnimodal expression inputs, we employ the audio encoder to process speech and sound, and the vision encoder to handle images. The resulting tokens are then" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.282, + 0.907, + 0.31 + ], + "angle": 0, + "content": "Figure 4. Overview of the proposed method OISA. Vision.audio encoders and text embedding are omitted for clarity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.334, + 0.907, + 0.502 + ], + "angle": 0, + "content": "integrated into the corresponding text tokens to produce the final Omnimodal Expression Tokens. All types of tokens are subsequently fed into the MLLM. The MLLM generates text responses and produces a [SEG] token representing the target object, which is then used by mask decoder for segmentation prediction. Our mask decoder uses query propagation to segment each frame, as described in Section 4.3. The query is refined online during the segmentation process, alleviating dynamic motion. Following previous work [33], we employed cross entropy loss for text generation, DICE loss [49], and binary cross entropy loss for segmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.75, + 0.531 + ], + "angle": 0, + "content": "4.2. Audio-Visual Interleaving" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For videos of arbitrary length, we uniformly sample \\(N\\) frames from the beginning to the end. Each frame is individually processed by the vision encoder to obtain \\(L_{v}\\) vision tokens, denoted as \\(\\mathbf{V} = \\{v_{1}, v_{2}, \\ldots, v_{N}\\}\\), where \\(v_{i} \\in \\mathbb{R}^{L_{v} \\times d}\\) represents the \\(L_{v}\\) vision tokens of the \\(i\\)-th frame with dimension \\(d\\). For the audio content in the video, we use an audio encoder to process it and obtain audio tokens \\(\\mathbf{A} \\in \\mathbb{R}^{L_{A} \\times d}\\), where \\(L_{A}\\) denotes the total number of audio tokens. To achieve audio-visual alignment, we segment the audio into clips corresponding to the frame rate, obtaining \\(\\{a_{1}, a_{2}, \\ldots, a_{N}\\}\\), where \\(a_{i} \\in \\mathbb{R}^{L_{a} \\times d}\\) and \\(L_{a} = L_{A} / N\\). Next, we interleave audio token clips and vision tokens to form the audio-visual interleaving token sequence \\(\\{v_{1}, a_{1}, v_{2}, a_{2}, \\ldots, v_{N}, a_{N}\\}\\). In comparison, VideoLLaMA [15, 80] directly concatenates audio and visual tokens sequentially, \\(\\{v_{1}, \\ldots, v_{N}, a_{1}, \\ldots, a_{N}\\}\\). However, this approach makes it difficult to align audio frames, especially in segmentation tasks where the video's FPS is not fixed, thereby exacerbating the challenge of audio-visual synchronization. Besides, video-SALMONN [58] fuses audio and visual tokens by concatenating them along the feature dimension and then applying a weight matrix for mapping. This way introduces additional training parameters and causes the fused audio-visual features becoming" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22579" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.48, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.223, + 0.454, + 0.238 + ], + "angle": 0, + "content": "Figure 5. Comparison of different mask decoder forms." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.259, + 0.452, + 0.275 + ], + "angle": 0, + "content": "misaligned with the original pretrained MLLM model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.283, + 0.276, + 0.299 + ], + "angle": 0, + "content": "4.3. Query Propagation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.485, + 0.804 + ], + "angle": 0, + "content": "As shown in Figure 4, once the Audio-Visual Content Tokens and Omnimodal Expression Tokens are obtained, the MLLM generates the corresponding responses. We employ the [SEG] token, following LISA [33], to represent the object embedding, which is subsequently passed to the mask decoder. While MLLMs excel at multimodal understanding and reasoning, they are not good at segmentation tasks. Previous works [3, 33] attach an additional vision encoder for plain feature extraction and a mask head for segmentation after MLLM, resulting in a redundant and suboptimal design. To address these issues, when extracting vision tokens from video, we simultaneously extract corresponding multi-scale features through ViT Adapter [12], which are then enhanced via pixel decoder [14]. These enhanced features, along with the [SEG] generated by the MLLM, are input into the mask decoder. When using mask decoder for segmentation, VideoLISA [3] employs the same [SEG] for the independent segmentation of each frame, aka One-Token-Seg-ALL (OTSA), as shown in Figure 5 (a). However, this approach has limitations. Previous studies [21, 27, 74, 75, 83] have shown that a single query often fails to adequately represent a target object, especially when rapid motion is present within the video. A single query carries a positional prior, making it difficult to encapsulate dynamic motion processes, such as an object moving from right to left, as shown in Figure 5 (a). This limitation results in ID switches, consistently tracking the target on the right part of the video. To address this, we employ a query propagation mechanism, as shown in Figure 5 (b), which allows the query to be updated online on a frame-by-frame basis [27, 79]. This approach enables the query to smoothly capture the temporal motion of the object while also effectively modeling contextual temporal information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.816, + 0.216, + 0.833 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We adopt InternVL2-1B [13] as our MLLM, equipped with InternViT-300M-448px [13] as vision encoder and Qwen2-0.5B-Instruct [1] as LLM. Whisper-large-v3 [54] is used as our audio encoder, with an" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "extra audio MLP to project audio features into LLM feature space. For ViT-Adapter [12], we remove all Injectors to reduce memory consumption and ensure vision tokens remain unaffected. For pixel decoder and mask decoder, we follow the default design of Mask2Former [14] but remove the self-attention module in each transformer block since there is only one query input to the mask decoder. We denote our model with the above settings as OISA-1B." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.258 + ], + "angle": 0, + "content": "Dataset Split. We split OmniAVS into training and testing sets, containing 1,864 videos with 54,304 expressions and 240 videos with 6,791 expressions, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.907, + 0.38 + ], + "angle": 0, + "content": "Evaluation Metrics. Following previous work [17, 18, 20, 68, 77], we adopt \\(\\mathcal{J}\\& \\mathcal{F}\\) as the segmentation evaluation metric, where \\(\\mathcal{J}\\) measures region similarity (IoU) and \\(\\mathcal{F}\\) evaluates contour accuracy. Following [41], for no-target expressions that referring to nothing in the video, \\(\\mathcal{J}\\& \\mathcal{F}\\) is set to 1 if the prediction is empty; otherwise, it is 0. We employ METEOR [4] as evaluation metric for text explanation generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.391, + 0.673, + 0.407 + ], + "angle": 0, + "content": "5.1. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.906, + 0.534 + ], + "angle": 0, + "content": "Audio-Text Alignment. Since InternVL2 does not support audio input natively, we introduce an audio encoder MLP to align audio features with the LLM space. This stage mainly aims to align audio features with LLM space. To achieve this goal, we use Automatic Speech Recognition datasets [9] and Audio Caption datasets [59] for audio-text alignment. During this process, we only train the audio encoder MLP while keeping all other parameters frozen." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.776 + ], + "angle": 0, + "content": "Omnimodal Instruct Segmentation Tuning. Following previous works [3, 33, 70, 84], we use multiple datasets for training, including 1) semantic segmentation datasets: ADE20K[85], COCO-Stuff[7], PASCAL-Part [11], PACOLVIS [55]; 2) referring segmentation datasets: RefCOCO, RefCOCO+ [78], RefCOCOg [48], ReasonSeg [33]; 4) referring video segmentation datasets: Refer-YouTube-VOS [57], Refer-DAVIS-17 [31], MeViS [17], ReVOS [70]; 5) audio-visual segmentation datasets: Ref-AVS Bench [68], OmniAVS. We use LoRA [29] to fine-tune LLM, train all parameters of the mask head, and freeze all remaining parameters. For video samples, during training, we uniformly sample 10 frames from the entire video and follow [3] to use 4 frames as dense frames and the remaining frames as sparse frames. During inference, we sample 32 frames with 4 frames as dense frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.667, + 0.804 + ], + "angle": 0, + "content": "5.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Ablation Study on Audio-Visual Fusion. In Table 3 we evaluate different audio-visual fusion strategies. For attention baseline, we adopt a simple cross-attention block where image tokens serve as key and audio tokens serve as query/value. For weighted sum, we first pad the audio tokens to match the length of vision tokens, then perform" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22580" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.482, + 0.159 + ], + "angle": 0, + "content": "Table 3. Ablation study on audio-visual fusion. AVI means Audio-Visual Interleaving. TVQA Split represents the performance evaluated on samples collected from TVQA [34]. These samples demand precise audio-visual alignment and feature a substantial amount of dialogue. See an example in Figure 6 (a)." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.161, + 0.476, + 0.248 + ], + "angle": 0, + "content": "
Fusion TypeSketchTVQA SplitOverall
Attention-37.435.8
Concat36.935.3
Weighted Sum34.831.1
AVI40.839.2
AVI + Concat42.040.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Table 4. Ablation study of mask head. OTSA: One-Token-Seg-All [3]; QP: query propagation (Ours); M2F: Mask2Former. FPS is tested on one A6000 GPU (batch size=1)." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.306, + 0.481, + 0.392 + ], + "angle": 0, + "content": "
IDQuery TypeHead TypeOmniAVS Metrics
OTSAQPM2FSAMJ&FJFFPS
I✓XX✓38.136.439.84.3
II✓X✓X35.233.836.615.7
IIIX✓X✓41.239.243.24.1
IVX✓✓X40.538.742.312.3
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.411, + 0.483, + 0.606 + ], + "angle": 0, + "content": "weighted sum on them. Attention-based and concatenation [15] methods achieve similar performance around \\(35 - 37\\%\\) \\(\\mathcal{J}\\& \\mathcal{F}\\). Our proposed Audio-Visual Interleaving (AVI) strategy significantly improves the overall performance to \\(39.2\\%\\) and get \\(40.2\\%\\) in TVQA split. Previous works [13, 44] add a thumbnail after tiling to capture global multimodal features to capture global features more effectively. Similarly, we append original audio tokens A at the end of the audio-visual interleaving sequence. This supplements complete and untruncated audio tokens based on our AVI audio-visual alignment, enabling the MLLM to handle more comprehensive audio tokens and thereby improving the model performance to \\(40.5\\%\\) \\(\\mathcal{J}\\& \\mathcal{F}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.483, + 0.804 + ], + "angle": 0, + "content": "Ablation Study on Mask Head. In Table 4 we study different mask head designs. While SAM mask head achieves competitive performance compared to M2F, it suffers from slower inference speed due to its huge ViT backbone. When using M2F as the mask head (Setting II and IV), we achieve at least \\(3 \\times\\) speedup with a slight accuracy trade-off. Regarding query types, our QP approach (Setting III and IV) significantly outperforms OTSA in segmentation quality, improving \\(\\mathcal{J} \\& \\mathcal{F}\\) from \\(38.1\\%\\) to \\(41.2\\%\\) when using SAM mask head, and from \\(35.2\\%\\) to \\(40.5\\%\\) when using M2F mask head. Considering the balance between performance and efficiency, we choose Setting IV (QP with M2F head) for our subsequent experiments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.353, + 0.832 + ], + "angle": 0, + "content": "5.3. OmniAVS Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In Table 5, we report the OmniAVS benchmark results. For traditional referring segmentation methods [17, 68, 71], we process each modality independently and concatenate their features to create a unified query representation." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 5. Testing on OmniAVS. We use \\(\\mathcal{I}\\&\\mathcal{F}\\) as the default metric. All is the average result across 8 splits. MET: METEOR." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.12, + 0.905, + 0.21 + ], + "angle": 0, + "content": "
MethodAllIIIIIIIVVVIVIIVIIIMET.
LMPM [17]25.831.228.720.022.721.320.930.031.4-
EEMC [68]29.634.432.619.626.028.024.735.636.0-
MUTR [71]32.335.433.328.429.826.522.841.640.5-
LISA-7B [33]33.633.331.229.232.728.627.343.443.111.6
LISA-13B [33]36.136.432.130.435.731.630.246.745.716.5
OISA-1B (ours)41.140.138.534.938.535.935.252.653.021.7
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.907, + 0.653 + ], + "angle": 0, + "content": "Regarding LISA [33], we enhance its audio processing capability by applying the same audio-text alignment method as our OISA, while maintaining its frame-by-frame inference approach. Our OISA-1B achieves state-of-the-art performance on the OmniAVS benchmark, significantly surpassing previous methods. Specifically, it attains an average score of \\(41.1\\%\\) across all modality combinations, outperforming LISA-13B by \\(5.0\\%\\). These improvements are consistent across all splits, with the most notable gain of \\(7.3\\%\\) on split VIII (speech+sound+image), showcasing our model's superior ability in multimodal input processing and reasoning. In terms of explanation generation, our model achieves a METEOR score of \\(21.7\\%\\), significantly surpassing LISA-13B's score of \\(16.5\\%\\). This indicates that our model excels not only in segmentation tasks but also in understanding and explaining its decisions. Analysis across different modality combinations reveals that splits VII (text+speech+image) and VIII (text+sound+image) achieve the highest scores of \\(52.6\\%\\) and \\(53.0\\%\\), respectively. This suggests that incorporating more modalities generally enhances performance, as the model effectively leverages complementary information from various sources. Comparison with the previous dataset Ref-AVS Bench shows that our OISA-1B achieved a score of \\(58.0\\%\\) (shown in Table 6), whereas it only attained \\(41.1\\%\\) on our OmniAVS. This significant difference underscores the increased difficulty and complexity of our dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Qualitative Results and Failure Cases. In Figure 6, we present several success and failure cases of the proposed baseline method OISA. Case (a) demonstrates that OISA effectively aligns audio with visual content, identifies key dialogues in the fourth frame, and subsequently performs reasoning and accurately locates the target object, providing a clear explanation. Besides, Case (a) shows that OISA achieves robust tracking in scenarios involving viewpoint changes, occlusions, and reappearance. Case (b) illustrates the model's ability to handle multimodal expressions by understanding both the fire in the image and the siren sound to locate the fire truck. However, as shown in Case (c), OISA struggles in scenarios with complex sounds, indicating that understanding intricate sounds still requires further exploration, such as decoupling sound events [28, 46] before perception and reasoning." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22581" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.478, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.253, + 0.431, + 0.266 + ], + "angle": 0, + "content": "Figure 6. Success and failure cases of OISA-1B." + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.287, + 0.444, + 0.3 + ], + "angle": 0, + "content": "Table 6. Performance on testing set of Ref-AVS [68]." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.302, + 0.478, + 0.419 + ], + "angle": 0, + "content": "
MethodSeenUnseenMix (S+U)Null
JFJFJFS (↓)
AVSBench [86]23.251.132.454.727.852.920.8
AVGSegFormer [22]33.547.036.150.134.848.617.1
GAVS [65]28.949.829.849.729.449.819.0
ReferFormer [69]31.350.130.448.830.949.517.6
R2VOS [38]25.041.027.949.826.545.418.3
EEMC [68]34.251.349.564.841.958.10.7
OISA-1B (ours)51.758.758.365.154.561.49.8
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.45, + 0.318, + 0.463 + ], + "angle": 0, + "content": "5.4. Results on Related Tasks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.669 + ], + "angle": 0, + "content": "Results on Ref-AVS. As shown in Table 6, our OISA-1B achieves SOTA performance on Ref-AVS [68], greatly surpassing previous methods. Specifically, OISA-1B improves the \\(\\mathcal{I}\\) by \\(17.5\\%\\) and \\(8.8\\%\\) on Seen and Unseen splits respectively compared to EEMC [68], demonstrating good generalization ability. While OISA-1B gets suboptimal score on Null split, which may inherit from hallucination in MLLM, EEMC achieves \\(0.7\\%\\) \\(\\mathcal{S}\\). We argue that videos in Null split contain empty or fixed audio with unrelated expressions, so EEMC is likely overfitting to identify audio patterns without understanding. Therefore, our model's performance \\((9.8\\%)\\) is reasonable, and OISA-1B's advantages on other metrics also demonstrate its robustness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.777 + ], + "angle": 0, + "content": "Results on Referring Image Segmentation. As shown in Table 7, our OISA-1B outperforms previous referring image segmentation methods on refCOCO series [48, 78], achieving state-of-the-art performance. The performance on ReasonSeg [33] falls short of LISA [33] and VideoLISA [3], likely due to the limited reasoning abilities of our smaller 0.5B LLM, which are crucial for this dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results on Referring Video Segmentation. As shown in Table 8, the proposed OISA-1B achieves competitive results on MeViS [17] (43.2%), R-YTVOS [57] (62.1%), and R-DAVIS\\(_{17}\\) [31] (65.2%). On ReVOS [70], it achieves new state-of-the-art results with 47.3% \\(\\mathcal{I} \\& \\mathcal{F}\\) and 19.3% robustness score, surpassing VISA [70] by 0.2% and 4.0% respectively, demonstrating strong perception and reasoning ability in complex video scenes." + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.09, + 0.904, + 0.104 + ], + "angle": 0, + "content": "Table 7. cIoU Results on refCOCO [48, 78] and ReasonSeg [33]." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.106, + 0.902, + 0.207 + ], + "angle": 0, + "content": "
MethodrefCOCOrefCOCO+refCOCogReasonSeg
valtestAtestBvaltestAtestBvalUtestUvaltest
LAVT [73]72.775.868.862.168.455.161.262.1--
ReLA [41]73.876.570.266.071.057.765.066.0--
LISA-7B [33]74.176.571.162.467.456.566.468.554.048.4
LISA-13B [33]--------62.951.1
VISA-7B [70]72.475.568.159.864.853.165.566.457.8-
VideoLISA-3.8B [3]73.876.668.863.468.856.268.368.867.154.4
OISA-1B (ours)74.278.372.068.871.762.070.672.361.250.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.222, + 0.904, + 0.248 + ], + "angle": 0, + "content": "Table 8. Results on referring/reasoning video segmentation [17, 31, 57, 70]. \\(\\mathcal{I}\\&\\mathcal{F}\\) is the default metric, \\(\\mathcal{R}\\) is the robustness score." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.251, + 0.905, + 0.362 + ], + "angle": 0, + "content": "
MethodMeViSR-YTVOSR-DAVIS17ReVOSReVOS(R)
ReferFormer [69]31.062.961.114.94.9
MTTR [5]30.055.3-25.55.6
LMPM [17]37.2--26.43.2
LISA-7B [33]-50.258.440.99.3
TrackGPT-7B [87]40.156.463.243.611.6
VideoLISA-3.8B [3]42.361.767.7--
VISA-7B [70]44.563.070.447.115.3
OISA-1B (ours)43.262.165.247.319.3
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.382, + 0.761, + 0.397 + ], + "angle": 0, + "content": "6. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.407, + 0.907, + 0.619 + ], + "angle": 0, + "content": "We introduce OmniAVS, a large-scale dataset supporting advanced multimodal referring and reasoning in audiovisual scenes with 8 types of expressions across text, speech, sound, and image. Compared to previous referring datasets focused on salient video or sound occurrence, OmniAVS enables deeper understanding and reasoning of audiovisual content. To address the challenges posed by OmniAVS, we propose a simple yet effective baseline OISA, a segmentation model based on multimodal large language models. Our OISA leverages audio-visual interleaving for temporal alignment and query propagation for efficient segmentation, enabling effective handling of multimodal inputs without additional parameters. Extensive experiments on 10 datasets demonstrate the effectiveness of our OISA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.909, + 0.862 + ], + "angle": 0, + "content": "Future Directions. Although our baseline OISA has shown promising results on OmniAVS, there is still much room for improvement: i) Exploring more efficient audiovisual fusion methods, such as learning more robust audiovisual representations. ii) Disentangling sounds when multiple objects make sounds simultaneously, which is crucial for subsequent reasoning. iii) How to effectively combine multiple modalities in expressions by using joint representations or cross-modal fusion. iv) Exploring more advanced segmentation models to handle complex scenarios, e.g., occlusion, disappearance and reappearance. v) Pre-training OISA on larger audio-visual datasets to further improve its generalization to unseen tasks. vi) Developing more interactive and conversational capabilities to enable multi-round interactions with users, allowing for clarification and refinement of referring expressions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.875, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This project was supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62472104." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22582" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen Technical Report. arXiv, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.227 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A Frontier Large Vision-Language Model with Versatile Abilities. arXiv preprint arXiv:2308.12966, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.483, + 0.299 + ], + "angle": 0, + "content": "[3] Zechen Bai, Tong He, Haiyang Mei, Pichao WANG, Ziteng Gao, Joya Chen, liulei, Zheng Zhang, and Mike Zheng Shou. One Token to Seg Them All: Language Instructed Reasoning Segmentation in Videos. In Adv. Neural Inform. Process. Syst., 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.301, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[4] Satanjeev Banerjee and Alon Lavie. METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments. In Assoc. Comput. Linguist. Worksh., 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[5] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-End Referring Video Object Segmentation with Multimodal Transformers. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.483, + 0.443 + ], + "angle": 0, + "content": "[6] Albert S Bregman. Auditory Scene Analysis: The Perceptual Organization of Sound. MIT press, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[7] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and Stuff Classes in Context. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[8] Jacob Chalk, Jaesung Huh, Evangelos Kazakos, Andrew Zisserman, and Dima Damen. TIM: A Time Interval Machine for Audio-Visual Action Recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[9] Guoguo Chen, Shuzhou Chai, et al. GigaSpeech: An Evolving, Multi-domain ASR Corpus with 10,000 Hours of Transcribed Audio. In Proc. Interspeech 2021, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.589, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[10] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. VGGSound: A Large-scale Audio-Visual Dataset. In IEEE Int. Conf. Acoust. Speech Signal Process., 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.633, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[11] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect What You Can: Detecting and Representing Objects using Holistic Models and Body Parts. In IEEE Conf. Comput. Vis. Pattern Recog., 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision Transformer Adapter for Dense Predictions. In Int. Conf. Learn. Represent., 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites. arXiv preprint arXiv:2404.16821, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[15] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs. arXiv preprint arXiv:2406.07476, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[16] Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, et al. Qwen2-Audio Technical Report. arXiv preprint arXiv:2407.10759, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[17] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A Large-scale Benchmark for Video Segmentation with Motion Expressions. In Int. Conf. Comput. Vis., 2023. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[18] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A New Dataset for Video Object Segmentation in Complex Scenes. In Int. Conf. Comput. Vis., 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[19] Henghui Ding, Song Tang, Shuting He, Chang Liu, Zuxuan Wu, and Yu-Gang Jiang. Multimodal referring segmentation: A survey. arXiv, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[20] Henghui Ding, Kaining Ying, Chang Liu, Shuting He, YuGang Jiang, Philip HS Torr, and Song Bai. MOSEv2: A more challenging dataset for video object segmentation in complex scenes. arXiv, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[21] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, et al. VITA: Towards Open-Source Interactive Omni Multimodal LLM. arXiv, 2024. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[22] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. AVSegFormer: Audio-Visual Segmentation with Transformer. In AAAI, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[23] GPT-SoVITS. https://github.com/RVC-Boss/GPT-SoVITS, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[24] Ruohao Guo, Liao Qu, Dantong Niu, Yanyu Qi, Wenzhen Yue, Ji Shi, Bowei Xing, and Xianghua Ying. Open-Vocabulary Audio-Visual Semantic Segmentation. In ACM Int. Conf. Multimedia, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[25] Junwen He, Yifan Wang, Lijun Wang, Huchuan Lu, JunYan He, Jin-Peng Lan, Bin Luo, and Xuansong Xie. Multimodal Instruction Tuned LLMs with Fine-grained Visual Perception. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[26] Shuting He and Henghui Ding. Decoupling Static and Hierarchical Motion Perception for Referring Video Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[27] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A Generalized Framework for Video Instance Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[28] John R Hershey, Zhuo Chen, Jonathan Le Roux, and Shinji Watanabe. Deep clustering: Discriminative embeddings for segmentation and separation. In IEEE Int. Conf. Acoust. Speech Signal Process., 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In Int. Conf. Learn. Represent., 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22583" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[30] Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Egocentric Audio-Visual Object Localization. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[31] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video Object Segmentation with Language Referring Expressions. In ACCV, 2019. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment Anything. In Int. Conf. Comput. Vis., 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[33] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: Reasoning Segmentation via Large Language Model. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[34] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. TVQA: Localized, Compositional Video Question Answering. In Proc. of the Conf. on Empirical Methods in Nat. Lang. Process., 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[35] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to Answer Questions in Dynamic Audio-Visual Scenarios. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.441 + ], + "angle": 0, + "content": "[36] Guangyao Li, Henghui Du, and Di Hu. Boosting Audio Visual Question Answering via Key Semantic-Aware Cues. In ACM Int. Conf. Multimedia, pages 5997-6005, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.483 + ], + "angle": 0, + "content": "[37] Jia Li, Wenjie Zhao, Ziru Huang, Yunhui Guo, and Yapeng Tian. Do Audio-Visual Segmentation Models Truly Segment Sounding Objects? arXiv, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.484, + 0.537 + ], + "angle": 0, + "content": "[38] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust Referring Video Object Segmentation with Cyclic Structural Consensus. In Int. Conf. Comput. Vis., 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.538, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[39] Yadong Li, Haoze Sun, Mingan Lin, Tianpeng Li, Guosheng Dong, Tao Zhang, Bowen Ding, Wei Song, Zhenglin Cheng, Yuqi Huo, et al. Baichuan-Omni Technical Report. arXiv preprint arXiv:2410.08565, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.65 + ], + "angle": 0, + "content": "[40] Linfeng Yuan and Miaojing Shi and Zijie Yue and Qijun Chen. Losh: Long-short text joint prediction network for referring video object segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.484, + 0.692 + ], + "angle": 0, + "content": "[41] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized Referring Expression Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[42] Chang Liu, Xudong Jiang, and Henghui Ding. Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[43] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. In Adv. Neural Inform. Process. Syst., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[45] Shuo Liu, Kaining Ying, Hao Zhang, Yue Yang, Yuqi Lin, Tianle Zhang, Chuanhao Li, Yu Qiao, Ping Luo, Wenqi Shao, et al. ConvBench: A Multi-Turn Conversation Evaluation Benchmark with Hierarchical Ablation Capability for Large Vision-Language Models. In Adv. Neural Inform. Process. Syst. D&B, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[46] Yi Luo and Nima Mesgarani. Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation. IEEE/ACM Trans. Audio Speech Lang. Process., 27 (8), 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.15, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[47] Juncheng Ma, Peiwen Sun, Yaoting Wang, and Di Hu. Stepping Stones: A Progressive Training Strategy for Audio-Visual Semantic Segmentation. In Eur. Conf. Comput. Vis., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.208, + 0.906, + 0.264 + ], + "angle": 0, + "content": "[48] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and Comprehension of Unambiguous Object Descriptions. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.266, + 0.906, + 0.32 + ], + "angle": 0, + "content": "[49] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In IEEE Int. Conf. 3D Vis., 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.324, + 0.906, + 0.352 + ], + "angle": 0, + "content": "[50] OpenAI. https://platform.openai.com/docs/guides/text-to-speech, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.354, + 0.906, + 0.382 + ], + "angle": 0, + "content": "[51] OpenAI. https://openai.com/index/hello-gpt-4o, 2024.2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.384, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[52] Wenwen Pan, Haonan Shi, Zhou Zhao, Jieming Zhu, Xi-uqiang He, Zhigeng Pan, Lianli Gao, Jun Yu, Fei Wu, and Qi Tian. Wnet: Audio-Guided Video Object Segmentation via Wavelet-Based Cross-Modal Denoising Networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.457, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[53] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, et al. DetGPT: Detect What You Need via Reasoning. In Proc. of the Conf. on Empirical Methods in Nat. Lang. Process., 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.528, + 0.906, + 0.582 + ], + "angle": 0, + "content": "[54] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust Speech Recognition via Large-Scale Weak Supervision. In Int. Conf. Mach. Learn., 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.586, + 0.906, + 0.655 + ], + "angle": 0, + "content": "[55] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. PACO: Parts and Attributes of Common Objects. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.657, + 0.906, + 0.727 + ], + "angle": 0, + "content": "[56] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. SAM 2: Segment Anything in Images and Videos. arXiv preprint arXiv:2408.00714, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.729, + 0.906, + 0.784 + ], + "angle": 0, + "content": "[57] Seonguk Seo, Joon-Young Lee, and Bohyung Han. URVOS: Unified Referring Video Object Segmentation Network with a Large-Scale Benchmark. In Eur. Conf. Comput. Vis., 2020. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.787, + 0.906, + 0.856 + ], + "angle": 0, + "content": "[58] Guangzhi Sun, Wenyi Yu, Changli Tang, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, Yuxuan Wang, and Chao Zhang. video-SALMONN: Speech-Enhanced Audio-Visual Large Language Models. In Int. Conf. Mach. Learn., 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[59] Luoyi Sun, Xuanan Xu, Mengyue Wu, and Weidi Xie. Auto-ACD: A Large-scale Dataset for Audio-Language Representation Learning. In ACM Int. Conf. Multimedia, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22584" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[60] Peiwen Sun, Honggang Zhang, and Di Hu. Unveiling and Mitigating Bias in Audio Visual Segmentation. In ACM Int. Conf. Multimedia, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[61] Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. SALMONN: Towards Generic Hearing Abilities for Large Language Models. In Int. Conf. Learn. Represent., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[62] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-Visual Event Localization in Unconstrained Videos. In Eur. Conf. Comput. Vis., 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[63] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[64] Yefei Wang, Kaili Wang, Yi Wang, Di Guo, Huaping Liu, and Fuchun Sun. Audio-Visual Grounding Referring Expression for Robotic Manipulation. In IEEE Int. Conf. Robot. Autom., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[65] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting Segmentation with Sound Is Generalizable Audio-Visual Source Localization. In AAAI, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[66] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting Segmentation with Sound Is Generalizable Audio-Visual Source Localization. In AAAI, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[67] Yaoting Wang, Peiwen Sun, Yuanchao Li, Honggang Zhang, and Di Hu. Can Textual Semantics Mitigate Sounding Object Segmentation Preference? In Eur. Conf. Comput. Vis., 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[68] Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. Ref-AVS: Refer and Segment Objects in Audio-Visual Scenes. In Eur. Conf. Comput. Vis., 2024. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[69] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as Queries for Referring Video Object Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[70] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. VISA: Reasoning Video Object Segmentation via Large Language Models. In Eur. Conf. Comput. Vis., 2024. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[71] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by Multi-Modality: A Unified Temporal Transformer for Video Object Segmentation. In AAAI, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[72] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. AVQA: A Dataset for Audio-Visual Question Answering on Videos. In ACM Int. Conf. Multimedia, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[73] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. LAVT: Language-Aware Vision Transformer for Referring Image Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[74] Kaining Ying, Zhenhua Wang, Cong Bai, and Pengfei Zhou. Isda: Position-aware instance segmentation with deformable" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "attention. In IEEE Int. Conf. Acoust. Speech Signal Process., 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[75] Kaining Ying, Qing Zhong, Weian Mao, Zhenhua Wang, Hao Chen, Lin Yuanbo Wu, Yifan Liu, Chengxiang Fan, Yunzhi Zhuge, and Chunhua Shen. CTVIS: Consistent Training for Online Video Instance Segmentation. In Int. Conf. Comput. Vis., 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[76] Kaining Ying, Fanqing Meng, Jin Wang, Zhiqian Li, Han Lin, Yue Yang, Hao Zhang, Wenbo Zhang, Yuqi Lin, Shuo Liu, Jiayi Lei, Quanfeng Lu, Runjian Chen, Peng Xu, Renrui Zhang, Haozhe Zhang, Peng Gao, Yali Wang, Yu Qiao, Ping Luo, Kaipeng Zhang, and Wenqi Shao. MMT-Bench: A Comprehensive Multimodal Benchmark for Evaluating Large Vision-Language Models Towards Multitask AGI. In Int. Conf. Mach. Learn., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.301, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[77] Kaining Ying, Hengrui Hu, and Henghui Ding. MOVE: Motion-guided few-shot video object segmentation. In Int. Conf. Comput. Vis., 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[78] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling Context in Referring Expressions. In Eur. Conf. Comput. Vis., 2016. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[79] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. MOTR: End-to-End Multiple-Object Tracking with Transformer. In Eur. Conf. Comput. Vis., 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding. In Proc. of the Conf. on Empirical Methods in Nat. Lang. Process., 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.564 + ], + "angle": 0, + "content": "[81] Renrui Zhang, Jiaming Han, Chris Liu, Aojun Zhou, Pan Lu, Yu Qiao, Hongsheng Li, and Peng Gao. LLaMA-Adapter: Efficient Fine-tuning of Large Language Models with Zero-initialized Attention. In Int. Conf. Learn. Represent., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.566, + 0.905, + 0.634 + ], + "angle": 0, + "content": "[82] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Yu Liu, Kai Chen, and Ping Luo. GPT4RoI: Instruction Tuning Large Language Model on Region-of-Interest. arXiv preprint arXiv:2307.03601, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.635, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[83] Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. DVIS: Decoupled Video Instance Segmentation Framework. In Int. Conf. Comput. Vis., 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.691, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[84] Rongkun Zheng, Lu Qi, Xi Chen, Yi Wang, Kun Wang, Yu Qiao, and Hengshuang Zhao. ViLLa: Video Reasoning Segmentation with Large Language Model. arXiv preprint arXiv:2407.14500, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[85] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene Parsing through ADE20K Dataset. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[86] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-Visual Segmentation. In Eur. Conf. Comput. Vis., 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[87] Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with Human-Intent Reasoning. arXiv, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22585" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/7d9586d8-b2bc-4332-a23e-c79a83d7ffa9_origin.pdf b/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/7d9586d8-b2bc-4332-a23e-c79a83d7ffa9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4839c394ec34ce8e5fe065e2ddb76589d52f7deb --- /dev/null +++ b/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/7d9586d8-b2bc-4332-a23e-c79a83d7ffa9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dae92822a070241ab79a064660a60cd8882a8357d8ab4c47612b1c2feb88eb2c +size 3138812 diff --git a/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/full.md b/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e68da5bcf62f22e366f077912178e3a9bbf750cf --- /dev/null +++ b/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/full.md @@ -0,0 +1,348 @@ +# Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation + +Kaining Ying Henghui Ding Guangquan Jie Yu-Gang Jiang Fudan University, China https://henghuiding.com/OmniAVS/ + +![](images/bf108340ec71e323cdd2905750ff4d55cf6c7e98c9bc0d707710454ced831582.jpg) + +![](images/27e014ace92dec0868d33332164bd0b5ca078cab7e818199ca277e445e822583.jpg) + +![](images/cfb71e38f2cbc00fd73757223fdfd699c396e13804e47c9b1abf69e7d35c3bd8.jpg) +Which object mimics another? + +![](images/b356832cc4bf4c27a087eb56f18f4365966197e56d6741c434dc6b69e71b6a32.jpg) + +![](images/0879cbf93a27bc442d9c092b46b0489cc91cc2e5a760f2e92b1b471680a32342.jpg) +The one speak the same language as . + +e same language as : + +![](images/03f6a59ebc9dabfa5515685f8a514aa134fcd4551b29c4cbf2895ba6f94d1dbc.jpg) +Output: It is [SEG]. The girl is imitating the barking of the dog on the right. + +![](images/5e878ece4aeb966d1f0cc4ecd81796671223db454a4de2f58d4a07c4a472ec82.jpg) + +![](images/b762d089ef619b203f215708cfa5893416a1923a7348c53ad12db9bfb2ff6cd7.jpg) + +![](images/62ac3b8f240ea5a7581d5a64c24a2eb2b11edd6a3983327325440160bf7a6aaf.jpg) + +![](images/66bb6ebcd1680f5aa6d187ddadc08d253463c4a2ac20836fd5ad71d1efb2fac2.jpg) + + serves as lead + +![](images/4f73d2949afd8505f03af5d8f0bc2992bd90fd64d18cfa10bc0a005c8feb7ee2.jpg) +Output: It is [SEG]. They both speak Korean. +Output: It is [SEG]. The guitar on the right serves as lead. + +![](images/3d5a6c3d5072f2374a6d5bf5e583f0665389304b8245edc7b63ec8958d1106d4.jpg) + +![](images/fa058d5b82664b4ace528c5f7fba79f937b7d6870b7f432d76655ce14b97efe4.jpg) + +![](images/73c001faf6336adfe9d22d883bfc1fe4d065210907d1b67c56f569bfd60fb733.jpg) +The one similar to serves as lead. + +![](images/7b31e2f421d44310fcc98aea365eaa32654150cb3aa149a96c534b3cc963cbe9.jpg) + +![](images/7d4547cf5b1f4a3925042b29398a989815a5ac329f3f3972505d6ea0f037bfa5.jpg) + +![](images/ce853e2e1e27764830e2b625207a9d158f641edff7feb8bf1407a57049099e35.jpg) +The instrument is played by someone of the same gender and makes a similar sound in . + +![](images/b8da93d3bd86037194882cea478beb42a2a7de146bd07435d2f129d4846d0efb.jpg) +is played by someone of the similar sound in $<$ sound: $\mathbb{N}^{\prime}$ Sound + +![](images/222a8844b023e73f8aa6758d6ca968d1ca70b97aea11cd0049cf57df6ec8b2ae.jpg) + +![](images/fb0853a0b31c4fbf72cac2be3e722997d23506b17fa5f87f7fd03497dcb6a8c3.jpg) +Output: It is [SEG]. The pipa held by the man on the right. + +![](images/361881400489dc22938d8cf7f88a75558e1ae1f6e40130886cdbf7f1cd293614.jpg) +er in >. +Figure 1. Examples of the proposed benchmark Omnimodal Referring Audio-Visual Segmentation (OmniAVS) to show its nature and flexibility. OmniAVS introduces 3 key features: 1) It supports diverse multimodal referring expressions that flexibly combine text, speech, sound, and image for referring audio-visual segmentation; 2) It emphasizes understanding the content of audio rather than merely hearing them; 3) It incorporates complex reasoning and world knowledge in expressions, prompting models to provide explanations for their segmentation decisions. These characteristics make OmniAVS practical for real-world use and well-suited for developing omnimodal models with fine-grained perception. + +![](images/f3744975a652ab5a6f46cfb8b84c786a760821bffeb0376e3218b51008d24dc8.jpg) + +# Abstract + +Referring audio-visual segmentation (RAVS) has recently seen significant advancements, yet challenges remain in integrating multimodal information and deeply understand + +ing and reasoning about audiovisual content. To extend the boundaries of RAVS and facilitate future research in this field, we propose Omnimodal Referring AudioVisual Segmentation (OmniAVS), a new dataset containing 2,104 videos and 61,095 multimodal referring expressions. OmniAVS stands out with three key innovations: (1) 8 types of multimodal expressions that flexibly combine text, speech, sound, and visual cues; (2) an emphasis + +on understanding audio content beyond just detecting their presence; and (3) the inclusion of complex reasoning and world knowledge in expressions. Furthermore, we introduce Omnimodal Instructed Segmentation Assistant (OISA), to address the challenges of multimodal reasoning and fine-grained understanding of audiovisual content in OmniAVS. OISA uses MLLM to comprehend complex cues and perform reasoning-based segmentation. Extensive experiments show that OISA outperforms existing methods on OmniAVS and achieves competitive results on other related tasks. + +# 1. Introduction + +Referring Audio-Visual Segmentation (RAVS) [19, 68] is an emerging field focused on segmenting target objects indicated by referring expressions (e.g., text) within audio-visual scenes using audio-visual cues. The combination of referring and audio distinguishes RAVS from Referring Video Object Segmentation (RVOS) [17, 26, 31, 40, 42, 57], which uses text descriptions to refer to objects in silent videos, and Audio-Visual Segmentation (AVS) [24, 86], which segments any sound-producing object without referring expressions. RAVS enables practical applications in video conferencing and robotics, etc. + +Understanding Rather than Just Hearing. However, the association between expressions and audio cues in the previous RAVS dataset Ref-AVS [68] remains superficial, primarily limited to basic acoustic properties such as sound occurrence, intensity, and temporal sequence. For example, expressions like "Who is making the loudest sound?" only address the surface-level characteristics without delving into the content of the sounds. Recent works [3, 33, 53, 70, 84] in image and video segmentation have evolved from simple semantic perception to advanced reasoning understanding. However, such progress is notably absent in audio-visual segmentation. This gap naturally leads us to explore the potential for complex reasoning in audio-visual scenes. Besides, drawing insights from human auditory perception studies [6], understanding sound content is essential for sophisticated audio-visual scene reasoning. To achieve these goals, we present OmniAVS dataset to advance reasoning-based segmentation in audio-visual scenes. As shown in Figure 2(b), our dataset contains expressions that demand reasoning about audio content, such as "Who is most likely to be sick?". These expressions establish complex cognitive chains that go beyond basic acoustic features, e.g., sound $\rightarrow$ coughing $\rightarrow$ sickness. Furthermore, we enhance the dataset's interpretability by providing detailed explanations for reasoning-based expressions. + +Towards Omnimodal Referring. The emergence of flagship model ChatGPT-4o [51] sets a new standard for modern AI systems, emphasizing the need to process arbitrary combinations of omnimodal inputs, e.g., audio, vision, and text. In the context of RAVS, enabling omnimodal + +![](images/2f8ac4f9f80a1661769e23ef3e4c9f8eb4e7ba3b937f9f721149a1410f736e8a.jpg) +Figure 2. Comparison of (a) Ref-AVS [68] and (b) our proposed OmniAVS. In Ref-AVS, expressions mainly focus on surface-level sound properties, while OmniAVS demands deeper understanding and reasoning about sound content, enabling complex reasoning like identifying potential illness from coughing sounds. + +inputs can greatly enhance flexible referring and human-machine interactions. However, existing datasets [17, 68] are limited in scope, supporting only single or isolated modalities in their referring expressions. To address this issue, we include diverse expressions in OmniAVS. Beyond the most common text-only expressions, OmniAVS supports text with sound, text with image, and text with both sound and image. Additionally, we allow text to be replaced by speech*, resulting in 8 different kinds of expression types in OmniAVS, as shown in Figure 1. + +Omnimodal Instructed Segmentation Assisant. Given the unique and complex requirements of OmniAVS, which necessitate both flexible interfaces and sophisticated multimodal perception and reasoning capabilities, no existing models can be directly applied to this emerging field. To address this issue, we introduce a baseline model named Omnimodal Instructed Segmentation Assisant, based on Multimodal Large Language Model (MLLM) [13]. This model seamlessly integrates text, speech, sound, and image inputs to perform referring object segmentation while providing explanations for prediction. We identify audio-visual synchronization as a critical factor and implement Audio-Visual Interleaving to achieve temporal alignment without introducing additional parameters. Furthermore, unlike concurrent works [3] that use a single token to segment all frames, we adopt a more lightweight query propagation approach to achieve efficient and accurate segmentation. Experimental results demonstrate that the proposed method OISA possesses strong multimodal reasoning capabilities and supports multiple expression patterns, outperforming other methods on OmniAVS dataset. To further validate the effectiveness of our method OISA, we conduct evaluations on other datasets, including referring video/image segmentation and Ref-AVS, achieving competitive results compared to existing methods across all tasks. + +This work makes four main contributions: i) we explore omnimodal referring audio-visual segmentation, which is more aligned with real-world scenarios and provides a solid foundation for developing future omnimodal AI systems; ii) + +we introduce a large-scale referring audio-visual segmentation dataset OmniAVS that enables advanced multimodal reasoning in audio-visual scenes, with a particular emphasis on deep understanding and reasoning from audio content; iii) we enhance interaction flexibility by incorporating multiple referring modalities in the proposed OmniAVS dataset, broadening the scope of referring AVS; iv) we develop OISA, a multimodal large language model designed to uniformly handle omnimodal expressions and reasoning in RAVS, with extensive experiments across different tasks. + +# 2. Related Work + +Audio-Visual Scene Perception. Audio-visual scene perception [35-37, 47, 60, 66-68, 72] aims to simulate human multimodal sensory capabilities by integrating auditory and visual information. The rapid development of deep learning has enabled joint audio-visual learning, achieving significant progress in tasks such as event detection [62], robotic manipulation [64], object localization [30], action recognition [8], and object segmentation [68, 86]. The integration of multiple modalities enables models to achieve superior performance compared to single-modality approaches, particularly in handling incomplete or ambiguous information. However, datasets that comprehensively incorporate audio, visual, and language components remain scarce. Among the few available datasets, Music-AVQA [35], AVQA [72], and Ref-AVS [68] stand out as notable examples. Music-AVQA [35] and AVQA [72] primarily address question-answering tasks. Ref-AVS [68] focuses on referring audio-visual segmentation. Nevertheless, Ref-AVS's utilization of audio information is limited, as it employs relatively simple expressions without complex reasoning, thus constraining its effectiveness in real-world scenarios. To overcome these limitations, we introduce OmniAVS, which is specifically designed to maximize the utilization of audio information and seamlessly integrate it with textual descriptions. + +Reasoning-based Perception. With the rapid development of MLLMs [2, 13, 16, 43-45, 58, 61, 63, 76, 81], the research community has shifted from perception based on simple semantic [17, 71] to perception based on world knowledge or complex reasoning [3, 25, 33, 53, 70, 82]. For example, GPT4RoI [82] prompts MLLM with bounding boxes to solve region-level complex tasks that require reasoning. DetGPT [53] combines a pretrained MLLM with an open-vocabulary object detector to enable reasoning-based object detection from natural language instructions. LISA [33] uses a MLLM [43] to output a [SEG] token, which is then processed by a mask head [32] to generate the masks. Leveraging the world knowledge and reasoning capabilities inherent in LLMs, LISA demonstrates superior performance in complex segmentation scenarios that require complex reasoning. Building upon this foundation, VideoLISA [3] and VISA [70] extend LISA's capabilities to + +reasoning video object segmentation by incorporating comprehensive temporal modeling. Drawing inspiration from these works, we propose OISA, which extends reasoning capabilities to audio-visual contexts. Unlike previous works that primarily focus on vision-language reasoning, OISA leverages MLLM to perform reasoning about audio content in videos. While AnyRef [25] pioneers the use of audio and images as referring inputs for segmentation, it is constrained to static images and lacks the ability to reason about audio in dynamic video contexts. Moreover, some works [21, 39, 51] extend LLMs to multiple modalities, e.g., vison, audio, and text, but they cannot be used for segmentation tasks. + +# 3. Benchmark: OmniAVS + +# 3.1. Audio-Visual Video Collection + +The video sources consist of three main parts: i) Real-world web videos under the Creative Commons License, where we use techniques from [10, 68, 86] to align audio and visual snippets with intended semantics; ii) The open-source dataset TVQA [34], featuring extensive dialogue from TV shows; iii) Self-recorded videos, with consent from all participants. We select videos from above that meet the following two standards: + +- We encourage to select videos that contain audio content with informative or reasoning-based meanings and corresponding identifiable objects in the scene. +- We strive to select videos with complex scenes and multiple objects, enabling more diverse referring expressions, e.g., referring to multiple targets, and makes the dataset more representative of real-world complex scenarios. + +After reviewing about 10,871 potential candidates, we carefully selected the most satisfactory and appropriate videos, which meet the above standards. Finally, we chose 2,104 videos to create the OmniAVS dataset, which is diverse and representative of a wide range of real-world audio-visual scenarios. After collecting the videos, our research team annotates potential objects with expressions and bounding boxes at their first appearance. This information is then used by the annotation team to achieve video mask annotation on the referred objects. Next, we will introduce the corresponding annotation process. + +# 3.2. Omnimodal Expression and Mask Annotation + +The expression annotation process integrates 4 different modalities: text, speech, sound, and image. We categorize referring expressions into 8 forms: I) Text; II) Speech; III) Text with Sound; IV) Speech with Sound; V) Text with Image; VI) Speech with Image; VII) Text with Sound and Image; and VIII) Speech with Sound and Image. Each expression includes either text or speech, as they provide essential instructions for the model to understand the task and other modalities. All speech expressions are generated + +Table 1. Statistical comparison between the newly proposed OmniAVS and other datasets of related referring video segmentation tasks. Avail., Expl., and Expr. are abbreviations for Availability of reasoning, Explanations, and Expressions, respectively. + +
DatasetVenueContentReferringReasoningStatistics
VideoAudioTextAudioImageAvail.Expl.VideoFrameObjectMaskExpr.Expl.
Refer-YouTubeVOS [57][ECCV'20]✓X✓XXXX3,978116k7,451131k15,009X
MeViS [17][ICCV'23]✓X✓XXXX2,00644k8,175443k28,570X
ReVOS [70][ECCV'24]✓X✓XX✓X1,042116k5,535469k35,074X
Ref-AVS [68][ECCV'24]✓✓✓XXXX4,00240k6,88878k20,261X
OmniAVS (ours)[ICCV'25]✓✓✓✓✓✓✓2,104103k4,277206k61,09534,841
+ +by converting the corresponding text expressions into audio. Following previous works [21, 52], we use a combination of several TTS models [23, 50] and human readings for this conversion. In light of this, we will focus on introducing the annotation process for text-related expressions. Figure 3 shows the distribution of different expression types. The inner areas (e.g., text) are composed of their corresponding outer areas (e.g., 4 components of text). + +1) Text. This is the main component of OmniAVS. The annotation process follows these rules: 1) Expressions should relate to the sound in the video, not just visual cues. 2) Expressions should emphasize the sound's content, not just the act of making it. For example, instead of "The dog barking", use "The dog warning" as it requires understanding the sound. 3) Encourage reasoning in expressions, with necessary explanations. 4) Expressions can refer to any number of objects, from none to many (more than one). + +III) Text with Sound. This form combines text and sound to provide a more comprehensive expression. The annotation process follows the same rules as in the Text modality, with the following additional requirements when adding audio: 1) Use natural sounds, instrumental music, or environmental sounds, e.g., dog barking, bird chirping. 2) Text should complement the sound, adding context or instructions for identifying the object(s). An example is shown in the 2nd row of Figure 1. + +V)) Text with Image. This type combines text and image. The annotation follows the rules of Text modality with additional considerations for images: 1) Include relevant images that depict the target object(s) or provide context for the audio-visual content. 2) Ensure the text complements the image, offering instructions for identifying the object(s). An example is shown in the 3rd row of Figure 1. + +VII) Text with Sound and Image. This type combines text, sound, and image in the expressions. The text should guide identification using both sound and image. An example is shown in the 4th row of Figure 1. + +Mask Annotation. After collecting expressions, the annotation team proceeds to segment the objects referred to by each expression in every video. Using the bounding box of each object's first appearance (not necessarily in the first frame) annotated in the previous step as a reference, + +Table 2. More dataset statistics. + +
StatisticNumber
Words per Expr.9.3
FPS (mean/min/max)4.8/3/15
Expression Types8
Frames (min/max)11/251
Obj. per Expr. (mean)1.12
Obj. per Expr. (max)7
Duration6.2 h
Videos (train/test)1,864/240
Expr. (train/test)54,304/6,791
+ +![](images/a7e5b9ec7e2cad5200ca3e187196213bca0260da6ee9c3104d5dc1a79dbc8d16.jpg) +Figure 3. Distribution of expression types. SI: sound+image. + +the annotation team is required to carefully confirm the corresponding object in the video, then track and annotate the object's mask in all frames $\dagger$ . We designed an interactive annotation tool to automatically load the video and corresponding object bounding boxes. Annotators use the tool to load and preview the corresponding video, then sequentially track and segment target objects in the video. The tool integrates state-of-the-art interactive video object segmentation model SAM2 [56] to assist with mask annotation, enhancing efficiency. + +# 3.3. Dataset Statistics and Analysis + +Table 1 shows a statistical comparison between OmniAVS and related datasets. Refer-YouTubeVOS [57], MeViS [17], and ReVOS [70] focus on silent videos without audio and provide only text expressions. While ReVOS [70] further supports reasoning expression, it lacks explanations. Ref-AVS Bench [68] includes audio-visual videos but supports only single-modality text expressions without reasoning or explanations. Compared to previous datasets, the proposed OmniAVS dataset offers multimodal content (audio-visual videos) and diverse expression modalities (text, speech, sound, image), supports reasoning with explanations, and allows for referring to arbitrary number of target objects. We compare more datasets in supplementary. + +Main Differences with Ref-AVS. As Ref-AVS [68] is the most related dataset to our proposed OmniAVS, we summarize the differences here to facilitate reference: + +- Omnimodal Expressions: While Ref-AVS uses only textual referring expressions, OmniAVS supports expres + +sions that can combine text, speech, sound, image, providing a richer multimodal interactive interface. + +- Enhanced Audio-Visual Integration: Unlike Ref-AVS, which primarily focuses on "the sounding object", Omni-AVS offers a more comprehensive audio-visual analysis. Our expressions delve deeper into the content of sounds and their contextual impact on the scene. +- Reasoning and Explanation: Our dataset includes numerous expressions requiring reasoning, such as "Who is most likely to be sick?" (see Figure 2), which necessitates inferring from audio-visual cues like sneezing. We also provide explanations for some reasoning expressions. +- Number of Referred Target Objects: Unlike Ref-AVS that only refers to 0 or 1 object per expression, OmniAVS allows each referring expression to indicate an arbitrary number of target objects, i.e., from 0 to many as [41]. +- Annotation FPS: While Ref-AVS uses fixed video settings (10s, 1 FPS), OmniAVS includes videos of varying durations and annotation frame rates (3-15 FPS) to capture temporal dynamics. Table 2 provides more statistics. + +# 4. Method: OISA + +# 4.1. Overview Architecture + +Because of the unique nature of this task and the proposed OmniAVS dataset, there is no off-the-shelf model available. We propose an Omnimodal Instructed Segmentation Assistant (OISA) as a baseline method to support omnimodal referring and reasoning audio-visual segmentation. As shown in Figure 4, OISA consists of two main components: a MLLM for multimodal understanding and generation, and a mask head for segmentation and tracking. The MLLM includes audio encoder, vision encoder and LLM, while the mask head consists of ViT-Adapter [12], pixel decoder and mask decoder [14]. It's worth noting that the mask head is flexible and can be replaced with models like SAM [32]. + +Given an audio-visual video, the multimodal encoder processes video and audio content to obtain corresponding vision and audio tokens. Different from VideoLLaMA [80] that concatenates these two modalities sequentially, we adopt an Audio-Visual Interleaving strategy, as described in Section 4.2, where we divide audio tokens into clips and interleave them with vision tokens. The interleaved sequence forms our Audio-Visual Content Tokens. This approach effectively synchronizes audio and frames without introducing additional parameters, which is particularly useful for scenarios requiring audio-visual alignment. See examples at Figure 2, where without proper audiovisual alignment it would be difficult to determine which object sneezed when multiple objects are making sounds simultaneously. For omnimodal expression inputs, we employ the audio encoder to process speech and sound, and the vision encoder to handle images. The resulting tokens are then + +![](images/240914d0902495ac7ff33a954f38f4805b028935856e34480c4e60d5aae8db65.jpg) +Figure 4. Overview of the proposed method OISA. Vision.audio encoders and text embedding are omitted for clarity. + +integrated into the corresponding text tokens to produce the final Omnimodal Expression Tokens. All types of tokens are subsequently fed into the MLLM. The MLLM generates text responses and produces a [SEG] token representing the target object, which is then used by mask decoder for segmentation prediction. Our mask decoder uses query propagation to segment each frame, as described in Section 4.3. The query is refined online during the segmentation process, alleviating dynamic motion. Following previous work [33], we employed cross entropy loss for text generation, DICE loss [49], and binary cross entropy loss for segmentation. + +# 4.2. Audio-Visual Interleaving + +For videos of arbitrary length, we uniformly sample $N$ frames from the beginning to the end. Each frame is individually processed by the vision encoder to obtain $L_{v}$ vision tokens, denoted as $\mathbf{V} = \{v_{1}, v_{2}, \ldots, v_{N}\}$ , where $v_{i} \in \mathbb{R}^{L_{v} \times d}$ represents the $L_{v}$ vision tokens of the $i$ -th frame with dimension $d$ . For the audio content in the video, we use an audio encoder to process it and obtain audio tokens $\mathbf{A} \in \mathbb{R}^{L_{A} \times d}$ , where $L_{A}$ denotes the total number of audio tokens. To achieve audio-visual alignment, we segment the audio into clips corresponding to the frame rate, obtaining $\{a_{1}, a_{2}, \ldots, a_{N}\}$ , where $a_{i} \in \mathbb{R}^{L_{a} \times d}$ and $L_{a} = L_{A} / N$ . Next, we interleave audio token clips and vision tokens to form the audio-visual interleaving token sequence $\{v_{1}, a_{1}, v_{2}, a_{2}, \ldots, v_{N}, a_{N}\}$ . In comparison, VideoLLaMA [15, 80] directly concatenates audio and visual tokens sequentially, $\{v_{1}, \ldots, v_{N}, a_{1}, \ldots, a_{N}\}$ . However, this approach makes it difficult to align audio frames, especially in segmentation tasks where the video's FPS is not fixed, thereby exacerbating the challenge of audio-visual synchronization. Besides, video-SALMONN [58] fuses audio and visual tokens by concatenating them along the feature dimension and then applying a weight matrix for mapping. This way introduces additional training parameters and causes the fused audio-visual features becoming + +![](images/4f6e8a66ef50314aa6157eb548a0b8d2d30ffb86c37dbef4db6a93a1ce8b393e.jpg) +Figure 5. Comparison of different mask decoder forms. + +misaligned with the original pretrained MLLM model. + +# 4.3. Query Propagation + +As shown in Figure 4, once the Audio-Visual Content Tokens and Omnimodal Expression Tokens are obtained, the MLLM generates the corresponding responses. We employ the [SEG] token, following LISA [33], to represent the object embedding, which is subsequently passed to the mask decoder. While MLLMs excel at multimodal understanding and reasoning, they are not good at segmentation tasks. Previous works [3, 33] attach an additional vision encoder for plain feature extraction and a mask head for segmentation after MLLM, resulting in a redundant and suboptimal design. To address these issues, when extracting vision tokens from video, we simultaneously extract corresponding multi-scale features through ViT Adapter [12], which are then enhanced via pixel decoder [14]. These enhanced features, along with the [SEG] generated by the MLLM, are input into the mask decoder. When using mask decoder for segmentation, VideoLISA [3] employs the same [SEG] for the independent segmentation of each frame, aka One-Token-Seg-ALL (OTSA), as shown in Figure 5 (a). However, this approach has limitations. Previous studies [21, 27, 74, 75, 83] have shown that a single query often fails to adequately represent a target object, especially when rapid motion is present within the video. A single query carries a positional prior, making it difficult to encapsulate dynamic motion processes, such as an object moving from right to left, as shown in Figure 5 (a). This limitation results in ID switches, consistently tracking the target on the right part of the video. To address this, we employ a query propagation mechanism, as shown in Figure 5 (b), which allows the query to be updated online on a frame-by-frame basis [27, 79]. This approach enables the query to smoothly capture the temporal motion of the object while also effectively modeling contextual temporal information. + +# 5. Experiment + +Implementation Details. We adopt InternVL2-1B [13] as our MLLM, equipped with InternViT-300M-448px [13] as vision encoder and Qwen2-0.5B-Instruct [1] as LLM. Whisper-large-v3 [54] is used as our audio encoder, with an + +extra audio MLP to project audio features into LLM feature space. For ViT-Adapter [12], we remove all Injectors to reduce memory consumption and ensure vision tokens remain unaffected. For pixel decoder and mask decoder, we follow the default design of Mask2Former [14] but remove the self-attention module in each transformer block since there is only one query input to the mask decoder. We denote our model with the above settings as OISA-1B. + +Dataset Split. We split OmniAVS into training and testing sets, containing 1,864 videos with 54,304 expressions and 240 videos with 6,791 expressions, respectively. + +Evaluation Metrics. Following previous work [17, 18, 20, 68, 77], we adopt $\mathcal{J}\& \mathcal{F}$ as the segmentation evaluation metric, where $\mathcal{J}$ measures region similarity (IoU) and $\mathcal{F}$ evaluates contour accuracy. Following [41], for no-target expressions that referring to nothing in the video, $\mathcal{J}\& \mathcal{F}$ is set to 1 if the prediction is empty; otherwise, it is 0. We employ METEOR [4] as evaluation metric for text explanation generation. + +# 5.1. Training Details + +Audio-Text Alignment. Since InternVL2 does not support audio input natively, we introduce an audio encoder MLP to align audio features with the LLM space. This stage mainly aims to align audio features with LLM space. To achieve this goal, we use Automatic Speech Recognition datasets [9] and Audio Caption datasets [59] for audio-text alignment. During this process, we only train the audio encoder MLP while keeping all other parameters frozen. + +Omnimodal Instruct Segmentation Tuning. Following previous works [3, 33, 70, 84], we use multiple datasets for training, including 1) semantic segmentation datasets: ADE20K[85], COCO-Stuff[7], PASCAL-Part [11], PACOLVIS [55]; 2) referring segmentation datasets: RefCOCO, RefCOCO+ [78], RefCOCOg [48], ReasonSeg [33]; 4) referring video segmentation datasets: Refer-YouTube-VOS [57], Refer-DAVIS-17 [31], MeViS [17], ReVOS [70]; 5) audio-visual segmentation datasets: Ref-AVS Bench [68], OmniAVS. We use LoRA [29] to fine-tune LLM, train all parameters of the mask head, and freeze all remaining parameters. For video samples, during training, we uniformly sample 10 frames from the entire video and follow [3] to use 4 frames as dense frames and the remaining frames as sparse frames. During inference, we sample 32 frames with 4 frames as dense frames. + +# 5.2. Ablation Study + +Ablation Study on Audio-Visual Fusion. In Table 3 we evaluate different audio-visual fusion strategies. For attention baseline, we adopt a simple cross-attention block where image tokens serve as key and audio tokens serve as query/value. For weighted sum, we first pad the audio tokens to match the length of vision tokens, then perform + +Table 3. Ablation study on audio-visual fusion. AVI means Audio-Visual Interleaving. TVQA Split represents the performance evaluated on samples collected from TVQA [34]. These samples demand precise audio-visual alignment and feature a substantial amount of dialogue. See an example in Figure 6 (a). + +
Fusion TypeSketchTVQA SplitOverall
Attention-37.435.8
Concat36.935.3
Weighted Sum34.831.1
AVI40.839.2
AVI + Concat42.040.5
+ +Table 4. Ablation study of mask head. OTSA: One-Token-Seg-All [3]; QP: query propagation (Ours); M2F: Mask2Former. FPS is tested on one A6000 GPU (batch size=1). + +
IDQuery TypeHead TypeOmniAVS Metrics
OTSAQPM2FSAMJ&FJFFPS
I✓XX✓38.136.439.84.3
II✓X✓X35.233.836.615.7
IIIX✓X✓41.239.243.24.1
IVX✓✓X40.538.742.312.3
+ +weighted sum on them. Attention-based and concatenation [15] methods achieve similar performance around $35 - 37\%$ $\mathcal{J}\& \mathcal{F}$ . Our proposed Audio-Visual Interleaving (AVI) strategy significantly improves the overall performance to $39.2\%$ and get $40.2\%$ in TVQA split. Previous works [13, 44] add a thumbnail after tiling to capture global multimodal features to capture global features more effectively. Similarly, we append original audio tokens A at the end of the audio-visual interleaving sequence. This supplements complete and untruncated audio tokens based on our AVI audio-visual alignment, enabling the MLLM to handle more comprehensive audio tokens and thereby improving the model performance to $40.5\%$ $\mathcal{J}\& \mathcal{F}$ . + +Ablation Study on Mask Head. In Table 4 we study different mask head designs. While SAM mask head achieves competitive performance compared to M2F, it suffers from slower inference speed due to its huge ViT backbone. When using M2F as the mask head (Setting II and IV), we achieve at least $3 \times$ speedup with a slight accuracy trade-off. Regarding query types, our QP approach (Setting III and IV) significantly outperforms OTSA in segmentation quality, improving $\mathcal{J} \& \mathcal{F}$ from $38.1\%$ to $41.2\%$ when using SAM mask head, and from $35.2\%$ to $40.5\%$ when using M2F mask head. Considering the balance between performance and efficiency, we choose Setting IV (QP with M2F head) for our subsequent experiments. + +# 5.3. OmniAVS Benchmark Results + +In Table 5, we report the OmniAVS benchmark results. For traditional referring segmentation methods [17, 68, 71], we process each modality independently and concatenate their features to create a unified query representation. + +Table 5. Testing on OmniAVS. We use $\mathcal{I}\&\mathcal{F}$ as the default metric. All is the average result across 8 splits. MET: METEOR. + +
MethodAllIIIIIIIVVVIVIIVIIIMET.
LMPM [17]25.831.228.720.022.721.320.930.031.4-
EEMC [68]29.634.432.619.626.028.024.735.636.0-
MUTR [71]32.335.433.328.429.826.522.841.640.5-
LISA-7B [33]33.633.331.229.232.728.627.343.443.111.6
LISA-13B [33]36.136.432.130.435.731.630.246.745.716.5
OISA-1B (ours)41.140.138.534.938.535.935.252.653.021.7
+ +Regarding LISA [33], we enhance its audio processing capability by applying the same audio-text alignment method as our OISA, while maintaining its frame-by-frame inference approach. Our OISA-1B achieves state-of-the-art performance on the OmniAVS benchmark, significantly surpassing previous methods. Specifically, it attains an average score of $41.1\%$ across all modality combinations, outperforming LISA-13B by $5.0\%$ . These improvements are consistent across all splits, with the most notable gain of $7.3\%$ on split VIII (speech+sound+image), showcasing our model's superior ability in multimodal input processing and reasoning. In terms of explanation generation, our model achieves a METEOR score of $21.7\%$ , significantly surpassing LISA-13B's score of $16.5\%$ . This indicates that our model excels not only in segmentation tasks but also in understanding and explaining its decisions. Analysis across different modality combinations reveals that splits VII (text+speech+image) and VIII (text+sound+image) achieve the highest scores of $52.6\%$ and $53.0\%$ , respectively. This suggests that incorporating more modalities generally enhances performance, as the model effectively leverages complementary information from various sources. Comparison with the previous dataset Ref-AVS Bench shows that our OISA-1B achieved a score of $58.0\%$ (shown in Table 6), whereas it only attained $41.1\%$ on our OmniAVS. This significant difference underscores the increased difficulty and complexity of our dataset. + +Qualitative Results and Failure Cases. In Figure 6, we present several success and failure cases of the proposed baseline method OISA. Case (a) demonstrates that OISA effectively aligns audio with visual content, identifies key dialogues in the fourth frame, and subsequently performs reasoning and accurately locates the target object, providing a clear explanation. Besides, Case (a) shows that OISA achieves robust tracking in scenarios involving viewpoint changes, occlusions, and reappearance. Case (b) illustrates the model's ability to handle multimodal expressions by understanding both the fire in the image and the siren sound to locate the fire truck. However, as shown in Case (c), OISA struggles in scenarios with complex sounds, indicating that understanding intricate sounds still requires further exploration, such as decoupling sound events [28, 46] before perception and reasoning. + +![](images/80dbe08a52221f8e0efaaacb7d611cb6bccaf5daf49b21813c1135dfda3e411c.jpg) +Figure 6. Success and failure cases of OISA-1B. + +Table 6. Performance on testing set of Ref-AVS [68]. + +
MethodSeenUnseenMix (S+U)Null
JFJFJFS (↓)
AVSBench [86]23.251.132.454.727.852.920.8
AVGSegFormer [22]33.547.036.150.134.848.617.1
GAVS [65]28.949.829.849.729.449.819.0
ReferFormer [69]31.350.130.448.830.949.517.6
R2VOS [38]25.041.027.949.826.545.418.3
EEMC [68]34.251.349.564.841.958.10.7
OISA-1B (ours)51.758.758.365.154.561.49.8
+ +# 5.4. Results on Related Tasks + +Results on Ref-AVS. As shown in Table 6, our OISA-1B achieves SOTA performance on Ref-AVS [68], greatly surpassing previous methods. Specifically, OISA-1B improves the $\mathcal{I}$ by $17.5\%$ and $8.8\%$ on Seen and Unseen splits respectively compared to EEMC [68], demonstrating good generalization ability. While OISA-1B gets suboptimal score on Null split, which may inherit from hallucination in MLLM, EEMC achieves $0.7\%$ $\mathcal{S}$ . We argue that videos in Null split contain empty or fixed audio with unrelated expressions, so EEMC is likely overfitting to identify audio patterns without understanding. Therefore, our model's performance $(9.8\%)$ is reasonable, and OISA-1B's advantages on other metrics also demonstrate its robustness. + +Results on Referring Image Segmentation. As shown in Table 7, our OISA-1B outperforms previous referring image segmentation methods on refCOCO series [48, 78], achieving state-of-the-art performance. The performance on ReasonSeg [33] falls short of LISA [33] and VideoLISA [3], likely due to the limited reasoning abilities of our smaller 0.5B LLM, which are crucial for this dataset. + +Results on Referring Video Segmentation. As shown in Table 8, the proposed OISA-1B achieves competitive results on MeViS [17] (43.2%), R-YTVOS [57] (62.1%), and R-DAVIS $_{17}$ [31] (65.2%). On ReVOS [70], it achieves new state-of-the-art results with 47.3% $\mathcal{I} \& \mathcal{F}$ and 19.3% robustness score, surpassing VISA [70] by 0.2% and 4.0% respectively, demonstrating strong perception and reasoning ability in complex video scenes. + +Table 7. cIoU Results on refCOCO [48, 78] and ReasonSeg [33]. + +
MethodrefCOCOrefCOCO+refCOCogReasonSeg
valtestAtestBvaltestAtestBvalUtestUvaltest
LAVT [73]72.775.868.862.168.455.161.262.1--
ReLA [41]73.876.570.266.071.057.765.066.0--
LISA-7B [33]74.176.571.162.467.456.566.468.554.048.4
LISA-13B [33]--------62.951.1
VISA-7B [70]72.475.568.159.864.853.165.566.457.8-
VideoLISA-3.8B [3]73.876.668.863.468.856.268.368.867.154.4
OISA-1B (ours)74.278.372.068.871.762.070.672.361.250.8
+ +Table 8. Results on referring/reasoning video segmentation [17, 31, 57, 70]. $\mathcal{I}\&\mathcal{F}$ is the default metric, $\mathcal{R}$ is the robustness score. + +
MethodMeViSR-YTVOSR-DAVIS17ReVOSReVOS(R)
ReferFormer [69]31.062.961.114.94.9
MTTR [5]30.055.3-25.55.6
LMPM [17]37.2--26.43.2
LISA-7B [33]-50.258.440.99.3
TrackGPT-7B [87]40.156.463.243.611.6
VideoLISA-3.8B [3]42.361.767.7--
VISA-7B [70]44.563.070.447.115.3
OISA-1B (ours)43.262.165.247.319.3
+ +# 6. Conclusion and Discussion + +We introduce OmniAVS, a large-scale dataset supporting advanced multimodal referring and reasoning in audiovisual scenes with 8 types of expressions across text, speech, sound, and image. Compared to previous referring datasets focused on salient video or sound occurrence, OmniAVS enables deeper understanding and reasoning of audiovisual content. To address the challenges posed by OmniAVS, we propose a simple yet effective baseline OISA, a segmentation model based on multimodal large language models. Our OISA leverages audio-visual interleaving for temporal alignment and query propagation for efficient segmentation, enabling effective handling of multimodal inputs without additional parameters. Extensive experiments on 10 datasets demonstrate the effectiveness of our OISA. + +Future Directions. Although our baseline OISA has shown promising results on OmniAVS, there is still much room for improvement: i) Exploring more efficient audiovisual fusion methods, such as learning more robust audiovisual representations. ii) Disentangling sounds when multiple objects make sounds simultaneously, which is crucial for subsequent reasoning. iii) How to effectively combine multiple modalities in expressions by using joint representations or cross-modal fusion. iv) Exploring more advanced segmentation models to handle complex scenarios, e.g., occlusion, disappearance and reappearance. v) Pre-training OISA on larger audio-visual datasets to further improve its generalization to unseen tasks. vi) Developing more interactive and conversational capabilities to enable multi-round interactions with users, allowing for clarification and refinement of referring expressions. + +Acknowledgement. This project was supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62472104. + +# References + +[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen Technical Report. arXiv, 2023. 6 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A Frontier Large Vision-Language Model with Versatile Abilities. arXiv preprint arXiv:2308.12966, 2023. 3 +[3] Zechen Bai, Tong He, Haiyang Mei, Pichao WANG, Ziteng Gao, Joya Chen, liulei, Zheng Zhang, and Mike Zheng Shou. One Token to Seg Them All: Language Instructed Reasoning Segmentation in Videos. In Adv. Neural Inform. Process. Syst., 2024. 2, 3, 6, 7, 8 +[4] Satanjeev Banerjee and Alon Lavie. METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments. In Assoc. Comput. Linguist. Worksh., 2005. 6 +[5] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-End Referring Video Object Segmentation with Multimodal Transformers. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 8 +[6] Albert S Bregman. Auditory Scene Analysis: The Perceptual Organization of Sound. MIT press, 1994. 2 +[7] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and Stuff Classes in Context. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 6 +[8] Jacob Chalk, Jaesung Huh, Evangelos Kazakos, Andrew Zisserman, and Dima Damen. TIM: A Time Interval Machine for Audio-Visual Action Recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3 +[9] Guoguo Chen, Shuzhou Chai, et al. GigaSpeech: An Evolving, Multi-domain ASR Corpus with 10,000 Hours of Transcribed Audio. In Proc. Interspeech 2021, 2021. 6 +[10] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. VGGSound: A Large-scale Audio-Visual Dataset. In IEEE Int. Conf. Acoust. Speech Signal Process., 2020. 3 +[11] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect What You Can: Detecting and Representing Objects using Holistic Models and Body Parts. In IEEE Conf. Comput. Vis. Pattern Recog., 2014. 6 +[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision Transformer Adapter for Dense Predictions. In Int. Conf. Learn. Represent., 2023. 5, 6 +[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites. arXiv preprint arXiv:2404.16821, 2024. 2, 3, 6, 7 +[14] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 5, 6 + +[15] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs. arXiv preprint arXiv:2406.07476, 2024. 5, 7 +[16] Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, et al. Qwen2-Audio Technical Report. arXiv preprint arXiv:2407.10759, 2024. 3 +[17] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A Large-scale Benchmark for Video Segmentation with Motion Expressions. In Int. Conf. Comput. Vis., 2023. 2, 3, 4, 6, 7, 8 +[18] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A New Dataset for Video Object Segmentation in Complex Scenes. In Int. Conf. Comput. Vis., 2023. 6 +[19] Henghui Ding, Song Tang, Shuting He, Chang Liu, Zuxuan Wu, and Yu-Gang Jiang. Multimodal referring segmentation: A survey. arXiv, 2025. 2 +[20] Henghui Ding, Kaining Ying, Chang Liu, Shuting He, YuGang Jiang, Philip HS Torr, and Song Bai. MOSEv2: A more challenging dataset for video object segmentation in complex scenes. arXiv, 2025. 6 +[21] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, et al. VITA: Towards Open-Source Interactive Omni Multimodal LLM. arXiv, 2024. 3, 4, 6 +[22] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. AVSegFormer: Audio-Visual Segmentation with Transformer. In AAAI, 2024. 8 +[23] GPT-SoVITS. https://github.com/RVC-Boss/GPT-SoVITS, 2024.4 +[24] Ruohao Guo, Liao Qu, Dantong Niu, Yanyu Qi, Wenzhen Yue, Ji Shi, Bowei Xing, and Xianghua Ying. Open-Vocabulary Audio-Visual Semantic Segmentation. In ACM Int. Conf. Multimedia, 2024. 2 +[25] Junwen He, Yifan Wang, Lijun Wang, Huchuan Lu, JunYan He, Jin-Peng Lan, Bin Luo, and Xuansong Xie. Multimodal Instruction Tuned LLMs with Fine-grained Visual Perception. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3 +[26] Shuting He and Henghui Ding. Decoupling Static and Hierarchical Motion Perception for Referring Video Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2 +[27] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A Generalized Framework for Video Instance Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 6 +[28] John R Hershey, Zhuo Chen, Jonathan Le Roux, and Shinji Watanabe. Deep clustering: Discriminative embeddings for segmentation and separation. In IEEE Int. Conf. Acoust. Speech Signal Process., 2016. 7 +[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In Int. Conf. Learn. Represent., 2022. 6 + +[30] Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Egocentric Audio-Visual Object Localization. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3 +[31] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video Object Segmentation with Language Referring Expressions. In ACCV, 2019. 2, 6, 8 +[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment Anything. In Int. Conf. Comput. Vis., 2023. 3, 5 +[33] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: Reasoning Segmentation via Large Language Model. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 5, 6, 7, 8 +[34] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. TVQA: Localized, Compositional Video Question Answering. In Proc. of the Conf. on Empirical Methods in Nat. Lang. Process., 2018. 3, 7 +[35] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to Answer Questions in Dynamic Audio-Visual Scenarios. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3 +[36] Guangyao Li, Henghui Du, and Di Hu. Boosting Audio Visual Question Answering via Key Semantic-Aware Cues. In ACM Int. Conf. Multimedia, pages 5997-6005, 2024. +[37] Jia Li, Wenjie Zhao, Ziru Huang, Yunhui Guo, and Yapeng Tian. Do Audio-Visual Segmentation Models Truly Segment Sounding Objects? arXiv, 2025. 3 +[38] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust Referring Video Object Segmentation with Cyclic Structural Consensus. In Int. Conf. Comput. Vis., 2023. 8 +[39] Yadong Li, Haoze Sun, Mingan Lin, Tianpeng Li, Guosheng Dong, Tao Zhang, Bowen Ding, Wei Song, Zhenglin Cheng, Yuqi Huo, et al. Baichuan-Omni Technical Report. arXiv preprint arXiv:2410.08565, 2024. 3 +[40] Linfeng Yuan and Miaojing Shi and Zijie Yue and Qijun Chen. Losh: Long-short text joint prediction network for referring video object segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2 +[41] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized Referring Expression Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 5, 6, 8 +[42] Chang Liu, Xudong Jiang, and Henghui Ding. Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence, 2024. 2 +[43] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. In Adv. Neural Inform. Process. Syst., 2023. 3 +[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 7 +[45] Shuo Liu, Kaining Ying, Hao Zhang, Yue Yang, Yuqi Lin, Tianle Zhang, Chuanhao Li, Yu Qiao, Ping Luo, Wenqi Shao, et al. ConvBench: A Multi-Turn Conversation Evaluation Benchmark with Hierarchical Ablation Capability for Large Vision-Language Models. In Adv. Neural Inform. Process. Syst. D&B, 2024. 3 + +[46] Yi Luo and Nima Mesgarani. Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation. IEEE/ACM Trans. Audio Speech Lang. Process., 27 (8), 2019. 7 +[47] Juncheng Ma, Peiwen Sun, Yaoting Wang, and Di Hu. Stepping Stones: A Progressive Training Strategy for Audio-Visual Semantic Segmentation. In Eur. Conf. Comput. Vis., 2024. 3 +[48] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and Comprehension of Unambiguous Object Descriptions. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 6, 8 +[49] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In IEEE Int. Conf. 3D Vis., 2016. 5 +[50] OpenAI. https://platform.openai.com/docs/guides/text-to-speech, 2023.4 +[51] OpenAI. https://openai.com/index/hello-gpt-4o, 2024.2,3 +[52] Wenwen Pan, Haonan Shi, Zhou Zhao, Jieming Zhu, Xi-uqiang He, Zhigeng Pan, Lianli Gao, Jun Yu, Fei Wu, and Qi Tian. Wnet: Audio-Guided Video Object Segmentation via Wavelet-Based Cross-Modal Denoising Networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 4 +[53] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, et al. DetGPT: Detect What You Need via Reasoning. In Proc. of the Conf. on Empirical Methods in Nat. Lang. Process., 2023. 2, 3 +[54] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust Speech Recognition via Large-Scale Weak Supervision. In Int. Conf. Mach. Learn., 2023. 6 +[55] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. PACO: Parts and Attributes of Common Objects. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 6 +[56] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. SAM 2: Segment Anything in Images and Videos. arXiv preprint arXiv:2408.00714, 2024. 4 +[57] Seonguk Seo, Joon-Young Lee, and Bohyung Han. URVOS: Unified Referring Video Object Segmentation Network with a Large-Scale Benchmark. In Eur. Conf. Comput. Vis., 2020. 2, 4, 6, 8 +[58] Guangzhi Sun, Wenyi Yu, Changli Tang, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, Yuxuan Wang, and Chao Zhang. video-SALMONN: Speech-Enhanced Audio-Visual Large Language Models. In Int. Conf. Mach. Learn., 2024. 3, 5 +[59] Luoyi Sun, Xuanan Xu, Mengyue Wu, and Weidi Xie. Auto-ACD: A Large-scale Dataset for Audio-Language Representation Learning. In ACM Int. Conf. Multimedia, 2024. 6 + +[60] Peiwen Sun, Honggang Zhang, and Di Hu. Unveiling and Mitigating Bias in Audio Visual Segmentation. In ACM Int. Conf. Multimedia, 2024. 3 +[61] Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. SALMONN: Towards Generic Hearing Abilities for Large Language Models. In Int. Conf. Learn. Represent., 2023. 3 +[62] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-Visual Event Localization in Unconstrained Videos. In Eur. Conf. Comput. Vis., 2018. 3 +[63] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191, 2024. 3 +[64] Yefei Wang, Kaili Wang, Yi Wang, Di Guo, Huaping Liu, and Fuchun Sun. Audio-Visual Grounding Referring Expression for Robotic Manipulation. In IEEE Int. Conf. Robot. Autom., 2022. 3 +[65] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting Segmentation with Sound Is Generalizable Audio-Visual Source Localization. In AAAI, 2024. 8 +[66] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting Segmentation with Sound Is Generalizable Audio-Visual Source Localization. In AAAI, 2024. 3 +[67] Yaoting Wang, Peiwen Sun, Yuanchao Li, Honggang Zhang, and Di Hu. Can Textual Semantics Mitigate Sounding Object Segmentation Preference? In Eur. Conf. Comput. Vis., 2024. +[68] Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. Ref-AVS: Refer and Segment Objects in Audio-Visual Scenes. In Eur. Conf. Comput. Vis., 2024. 2, 3, 4, 6, 7, 8 +[69] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as Queries for Referring Video Object Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 8 +[70] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. VISA: Reasoning Video Object Segmentation via Large Language Models. In Eur. Conf. Comput. Vis., 2024. 2, 3, 4, 6, 8 +[71] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by Multi-Modality: A Unified Temporal Transformer for Video Object Segmentation. In AAAI, 2024. 3, 7 +[72] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. AVQA: A Dataset for Audio-Visual Question Answering on Videos. In ACM Int. Conf. Multimedia, 2022. 3 +[73] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. LAVT: Language-Aware Vision Transformer for Referring Image Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 8 +[74] Kaining Ying, Zhenhua Wang, Cong Bai, and Pengfei Zhou. Isda: Position-aware instance segmentation with deformable + +attention. In IEEE Int. Conf. Acoust. Speech Signal Process., 2022. 6 +[75] Kaining Ying, Qing Zhong, Weian Mao, Zhenhua Wang, Hao Chen, Lin Yuanbo Wu, Yifan Liu, Chengxiang Fan, Yunzhi Zhuge, and Chunhua Shen. CTVIS: Consistent Training for Online Video Instance Segmentation. In Int. Conf. Comput. Vis., 2023. 6 +[76] Kaining Ying, Fanqing Meng, Jin Wang, Zhiqian Li, Han Lin, Yue Yang, Hao Zhang, Wenbo Zhang, Yuqi Lin, Shuo Liu, Jiayi Lei, Quanfeng Lu, Runjian Chen, Peng Xu, Renrui Zhang, Haozhe Zhang, Peng Gao, Yali Wang, Yu Qiao, Ping Luo, Kaipeng Zhang, and Wenqi Shao. MMT-Bench: A Comprehensive Multimodal Benchmark for Evaluating Large Vision-Language Models Towards Multitask AGI. In Int. Conf. Mach. Learn., 2024. 3 +[77] Kaining Ying, Hengrui Hu, and Henghui Ding. MOVE: Motion-guided few-shot video object segmentation. In Int. Conf. Comput. Vis., 2025. 6 +[78] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling Context in Referring Expressions. In Eur. Conf. Comput. Vis., 2016. 6, 8 +[79] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. MOTR: End-to-End Multiple-Object Tracking with Transformer. In Eur. Conf. Comput. Vis., 2022. 6 +[80] Hang Zhang, Xin Li, and Lidong Bing. Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding. In Proc. of the Conf. on Empirical Methods in Nat. Lang. Process., 2023. 5 +[81] Renrui Zhang, Jiaming Han, Chris Liu, Aojun Zhou, Pan Lu, Yu Qiao, Hongsheng Li, and Peng Gao. LLaMA-Adapter: Efficient Fine-tuning of Large Language Models with Zero-initialized Attention. In Int. Conf. Learn. Represent., 2024. 3 +[82] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Yu Liu, Kai Chen, and Ping Luo. GPT4RoI: Instruction Tuning Large Language Model on Region-of-Interest. arXiv preprint arXiv:2307.03601, 2023. 3 +[83] Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. DVIS: Decoupled Video Instance Segmentation Framework. In Int. Conf. Comput. Vis., 2023. 6 +[84] Rongkun Zheng, Lu Qi, Xi Chen, Yi Wang, Kun Wang, Yu Qiao, and Hengshuang Zhao. ViLLa: Video Reasoning Segmentation with Large Language Model. arXiv preprint arXiv:2407.14500, 2024. 2, 6 +[85] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene Parsing through ADE20K Dataset. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 6 +[86] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-Visual Segmentation. In Eur. Conf. Comput. Vis., 2022. 2, 3, 8 +[87] Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with Human-Intent Reasoning. arXiv, 2023. 8 \ No newline at end of file diff --git a/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/images.zip b/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..422e3b1aa5a11cdd6eeb819361003887dea108d0 --- /dev/null +++ b/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e36e083e7319c1cac22a70a89f932e5970c8d7fc2347fad7f82d4848790c16ef +size 550798 diff --git a/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/layout.json b/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..078e9c6df07902c1be3096e07ef8275cff932d03 --- /dev/null +++ b/2025/Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation/layout.json @@ -0,0 +1,9733 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 103, + 465, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 103, + 465, + 140 + ], + "spans": [ + { + "bbox": [ + 146, + 103, + 465, + 140 + ], + "type": "text", + "content": "Towards Omnimodal Expressions and Reasoning in Referring Audio-Visual Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 160, + 471, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 160, + 471, + 208 + ], + "spans": [ + { + "bbox": [ + 137, + 160, + 471, + 208 + ], + "type": "text", + "content": "Kaining Ying Henghui Ding Guangquan Jie Yu-Gang Jiang Fudan University, China https://henghuiding.com/OmniAVS/" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 236, + 157, + 291 + ], + "blocks": [ + { + "bbox": [ + 59, + 236, + 157, + 291 + ], + "lines": [ + { + "bbox": [ + 59, + 236, + 157, + 291 + ], + "spans": [ + { + "bbox": [ + 59, + 236, + 157, + 291 + ], + "type": "image", + "image_path": "bf108340ec71e323cdd2905750ff4d55cf6c7e98c9bc0d707710454ced831582.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 291, + 75, + 305 + ], + "blocks": [ + { + "bbox": [ + 59, + 291, + 75, + 305 + ], + "lines": [ + { + "bbox": [ + 59, + 291, + 75, + 305 + ], + "spans": [ + { + "bbox": [ + 59, + 291, + 75, + 305 + ], + "type": "image", + "image_path": "27e014ace92dec0868d33332164bd0b5ca078cab7e818199ca277e445e822583.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 75, + 291, + 92, + 305 + ], + "blocks": [ + { + "bbox": [ + 75, + 291, + 92, + 305 + ], + "lines": [ + { + "bbox": [ + 75, + 291, + 92, + 305 + ], + "spans": [ + { + "bbox": [ + 75, + 291, + 92, + 305 + ], + "type": "image", + "image_path": "cfb71e38f2cbc00fd73757223fdfd699c396e13804e47c9b1abf69e7d35c3bd8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 293, + 202, + 304 + ], + "lines": [ + { + "bbox": [ + 92, + 293, + 202, + 304 + ], + "spans": [ + { + "bbox": [ + 92, + 293, + 202, + 304 + ], + "type": "text", + "content": "Which object mimics another?" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 157, + 236, + 255, + 291 + ], + "blocks": [ + { + "bbox": [ + 157, + 236, + 255, + 291 + ], + "lines": [ + { + "bbox": [ + 157, + 236, + 255, + 291 + ], + "spans": [ + { + "bbox": [ + 157, + 236, + 255, + 291 + ], + "type": "image", + "image_path": "b356832cc4bf4c27a087eb56f18f4365966197e56d6741c434dc6b69e71b6a32.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 157, + 291, + 255, + 363 + ], + "blocks": [ + { + "bbox": [ + 157, + 291, + 255, + 363 + ], + "lines": [ + { + "bbox": [ + 157, + 291, + 255, + 363 + ], + "spans": [ + { + "bbox": [ + 157, + 291, + 255, + 363 + ], + "type": "image", + "image_path": "0879cbf93a27bc442d9c092b46b0489cc91cc2e5a760f2e92b1b471680a32342.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 363, + 289, + 376 + ], + "lines": [ + { + "bbox": [ + 75, + 363, + 289, + 376 + ], + "spans": [ + { + "bbox": [ + 75, + 363, + 289, + 376 + ], + "type": "text", + "content": "The one speak the same language as ." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 157, + 363, + 255, + 374 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 363, + 255, + 374 + ], + "spans": [ + { + "bbox": [ + 157, + 363, + 255, + 374 + ], + "type": "text", + "content": "e same language as :" + } + ] + } + ], + "index": 10, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 255, + 236, + 352, + 291 + ], + "blocks": [ + { + "bbox": [ + 255, + 236, + 352, + 291 + ], + "lines": [ + { + "bbox": [ + 255, + 236, + 352, + 291 + ], + "spans": [ + { + "bbox": [ + 255, + 236, + 352, + 291 + ], + "type": "image", + "image_path": "03f6a59ebc9dabfa5515685f8a514aa134fcd4551b29c4cbf2895ba6f94d1dbc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 289, + 293, + 543, + 304 + ], + "lines": [ + { + "bbox": [ + 289, + 293, + 543, + 304 + ], + "spans": [ + { + "bbox": [ + 289, + 293, + 543, + 304 + ], + "type": "text", + "content": "Output: It is [SEG]. The girl is imitating the barking of the dog on the right." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 255, + 307, + 352, + 362 + ], + "blocks": [ + { + "bbox": [ + 255, + 307, + 352, + 362 + ], + "lines": [ + { + "bbox": [ + 255, + 307, + 352, + 362 + ], + "spans": [ + { + "bbox": [ + 255, + 307, + 352, + 362 + ], + "type": "image", + "image_path": "5e878ece4aeb966d1f0cc4ecd81796671223db454a4de2f58d4a07c4a472ec82.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 59, + 307, + 157, + 362 + ], + "blocks": [ + { + "bbox": [ + 59, + 307, + 157, + 362 + ], + "lines": [ + { + "bbox": [ + 59, + 307, + 157, + 362 + ], + "spans": [ + { + "bbox": [ + 59, + 307, + 157, + 362 + ], + "type": "image", + "image_path": "b762d089ef619b203f215708cfa5893416a1923a7348c53ad12db9bfb2ff6cd7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 59, + 362, + 75, + 376 + ], + "blocks": [ + { + "bbox": [ + 59, + 362, + 75, + 376 + ], + "lines": [ + { + "bbox": [ + 59, + 362, + 75, + 376 + ], + "spans": [ + { + "bbox": [ + 59, + 362, + 75, + 376 + ], + "type": "image", + "image_path": "62ac3b8f240ea5a7581d5a64c24a2eb2b11edd6a3983327325440160bf7a6aaf.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 157, + 377, + 255, + 432 + ], + "blocks": [ + { + "bbox": [ + 157, + 377, + 255, + 432 + ], + "lines": [ + { + "bbox": [ + 157, + 377, + 255, + 432 + ], + "spans": [ + { + "bbox": [ + 157, + 377, + 255, + 432 + ], + "type": "image", + "image_path": "66bb6ebcd1680f5aa6d187ddadc08d253463c4a2ac20836fd5ad71d1efb2fac2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 432, + 255, + 445 + ], + "lines": [ + { + "bbox": [ + 157, + 432, + 255, + 445 + ], + "spans": [ + { + "bbox": [ + 157, + 432, + 255, + 445 + ], + "type": "text", + "content": " + serves as lead" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 255, + 377, + 352, + 433 + ], + "blocks": [ + { + "bbox": [ + 289, + 365, + 545, + 375 + ], + "lines": [ + { + "bbox": [ + 289, + 365, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 289, + 365, + 545, + 375 + ], + "type": "text", + "content": "Output: It is [SEG]. They both speak Korean." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 377, + 352, + 433 + ], + "lines": [ + { + "bbox": [ + 255, + 377, + 352, + 433 + ], + "spans": [ + { + "bbox": [ + 255, + 377, + 352, + 433 + ], + "type": "image", + "image_path": "4f73d2949afd8505f03af5d8f0bc2992bd90fd64d18cfa10bc0a005c8feb7ee2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 436, + 545, + 445 + ], + "lines": [ + { + "bbox": [ + 345, + 436, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 345, + 436, + 545, + 445 + ], + "type": "text", + "content": "Output: It is [SEG]. The guitar on the right serves as lead." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 352, + 377, + 450, + 433 + ], + "blocks": [ + { + "bbox": [ + 352, + 377, + 450, + 433 + ], + "lines": [ + { + "bbox": [ + 352, + 377, + 450, + 433 + ], + "spans": [ + { + "bbox": [ + 352, + 377, + 450, + 433 + ], + "type": "image", + "image_path": "3d5a6c3d5072f2374a6d5bf5e583f0665389304b8245edc7b63ec8958d1106d4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 450, + 377, + 548, + 433 + ], + "blocks": [ + { + "bbox": [ + 450, + 377, + 548, + 433 + ], + "lines": [ + { + "bbox": [ + 450, + 377, + 548, + 433 + ], + "spans": [ + { + "bbox": [ + 450, + 377, + 548, + 433 + ], + "type": "image", + "image_path": "fa058d5b82664b4ace528c5f7fba79f937b7d6870b7f432d76655ce14b97efe4.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 59, + 435, + 75, + 448 + ], + "blocks": [ + { + "bbox": [ + 59, + 435, + 75, + 448 + ], + "lines": [ + { + "bbox": [ + 59, + 435, + 75, + 448 + ], + "spans": [ + { + "bbox": [ + 59, + 435, + 75, + 448 + ], + "type": "image", + "image_path": "73c001faf6336adfe9d22d883bfc1fe4d065210907d1b67c56f569bfd60fb733.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 435, + 257, + 446 + ], + "lines": [ + { + "bbox": [ + 75, + 435, + 257, + 446 + ], + "spans": [ + { + "bbox": [ + 75, + 435, + 257, + 446 + ], + "type": "text", + "content": "The one similar to serves as lead." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 59, + 448, + 157, + 503 + ], + "blocks": [ + { + "bbox": [ + 59, + 448, + 157, + 503 + ], + "lines": [ + { + "bbox": [ + 59, + 448, + 157, + 503 + ], + "spans": [ + { + "bbox": [ + 59, + 448, + 157, + 503 + ], + "type": "image", + "image_path": "7b31e2f421d44310fcc98aea365eaa32654150cb3aa149a96c534b3cc963cbe9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 59, + 504, + 75, + 519 + ], + "blocks": [ + { + "bbox": [ + 59, + 504, + 75, + 519 + ], + "lines": [ + { + "bbox": [ + 59, + 504, + 75, + 519 + ], + "spans": [ + { + "bbox": [ + 59, + 504, + 75, + 519 + ], + "type": "image", + "image_path": "7d4547cf5b1f4a3925042b29398a989815a5ac329f3f3972505d6ea0f037bfa5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 75, + 505, + 92, + 519 + ], + "blocks": [ + { + "bbox": [ + 75, + 505, + 92, + 519 + ], + "lines": [ + { + "bbox": [ + 75, + 505, + 92, + 519 + ], + "spans": [ + { + "bbox": [ + 75, + 505, + 92, + 519 + ], + "type": "image", + "image_path": "ce853e2e1e27764830e2b625207a9d158f641edff7feb8bf1407a57049099e35.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 509, + 296, + 527 + ], + "lines": [ + { + "bbox": [ + 92, + 509, + 296, + 527 + ], + "spans": [ + { + "bbox": [ + 92, + 509, + 296, + 527 + ], + "type": "text", + "content": "The instrument is played by someone of the same gender and makes a similar sound in ." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 157, + 448, + 255, + 504 + ], + "blocks": [ + { + "bbox": [ + 157, + 448, + 255, + 504 + ], + "lines": [ + { + "bbox": [ + 157, + 448, + 255, + 504 + ], + "spans": [ + { + "bbox": [ + 157, + 448, + 255, + 504 + ], + "type": "image", + "image_path": "b8da93d3bd86037194882cea478beb42a2a7de146bd07435d2f129d4846d0efb.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 504, + 255, + 527 + ], + "lines": [ + { + "bbox": [ + 157, + 504, + 255, + 527 + ], + "spans": [ + { + "bbox": [ + 157, + 504, + 255, + 527 + ], + "type": "text", + "content": "is played by someone of the similar sound in " + }, + { + "bbox": [ + 157, + 504, + 255, + 527 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 157, + 504, + 255, + 527 + ], + "type": "text", + "content": "sound: " + }, + { + "bbox": [ + 157, + 504, + 255, + 527 + ], + "type": "inline_equation", + "content": "\\mathbb{N}^{\\prime}" + }, + { + "bbox": [ + 157, + 504, + 255, + 527 + ], + "type": "text", + "content": "Sound" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 255, + 448, + 352, + 504 + ], + "blocks": [ + { + "bbox": [ + 255, + 448, + 352, + 504 + ], + "lines": [ + { + "bbox": [ + 255, + 448, + 352, + 504 + ], + "spans": [ + { + "bbox": [ + 255, + 448, + 352, + 504 + ], + "type": "image", + "image_path": "222a8844b023e73f8aa6758d6ca968d1ca70b97aea11cd0049cf57df6ec8b2ae.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 352, + 448, + 450, + 504 + ], + "blocks": [ + { + "bbox": [ + 352, + 448, + 450, + 504 + ], + "lines": [ + { + "bbox": [ + 352, + 448, + 450, + 504 + ], + "spans": [ + { + "bbox": [ + 352, + 448, + 450, + 504 + ], + "type": "image", + "image_path": "fb0853a0b31c4fbf72cac2be3e722997d23506b17fa5f87f7fd03497dcb6a8c3.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 508, + 544, + 527 + ], + "lines": [ + { + "bbox": [ + 436, + 508, + 544, + 527 + ], + "spans": [ + { + "bbox": [ + 436, + 508, + 544, + 527 + ], + "type": "text", + "content": "Output: It is [SEG]. The pipa held by the man on the right." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 352, + 505, + 376, + 527 + ], + "blocks": [ + { + "bbox": [ + 289, + 505, + 359, + 527 + ], + "lines": [ + { + "bbox": [ + 289, + 505, + 359, + 527 + ], + "spans": [ + { + "bbox": [ + 289, + 505, + 359, + 527 + ], + "type": "text", + "content": "er in >." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 352, + 505, + 376, + 527 + ], + "lines": [ + { + "bbox": [ + 352, + 505, + 376, + 527 + ], + "spans": [ + { + "bbox": [ + 352, + 505, + 376, + 527 + ], + "type": "image", + "image_path": "361881400489dc22938d8cf7f88a75558e1ae1f6e40130886cdbf7f1cd293614.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 529, + 555, + 598 + ], + "lines": [ + { + "bbox": [ + 55, + 529, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 555, + 598 + ], + "type": "text", + "content": "Figure 1. Examples of the proposed benchmark Omnimodal Referring Audio-Visual Segmentation (OmniAVS) to show its nature and flexibility. OmniAVS introduces 3 key features: 1) It supports diverse multimodal referring expressions that flexibly combine text, speech, sound, and image for referring audio-visual segmentation; 2) It emphasizes understanding the content of audio rather than merely hearing them; 3) It incorporates complex reasoning and world knowledge in expressions, prompting models to provide explanations for their segmentation decisions. These characteristics make OmniAVS practical for real-world use and well-suited for developing omnimodal models with fine-grained perception." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 450, + 448, + 548, + 504 + ], + "blocks": [ + { + "bbox": [ + 450, + 448, + 548, + 504 + ], + "lines": [ + { + "bbox": [ + 450, + 448, + 548, + 504 + ], + "spans": [ + { + "bbox": [ + 450, + 448, + 548, + 504 + ], + "type": "image", + "image_path": "f3744975a652ab5a6f46cfb8b84c786a760821bffeb0376e3218b51008d24dc8.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 152, + 609, + 200, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 609, + 200, + 623 + ], + "spans": [ + { + "bbox": [ + 152, + 609, + 200, + 623 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 55, + 637, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 295, + 673 + ], + "type": "text", + "content": "Referring audio-visual segmentation (RAVS) has recently seen significant advancements, yet challenges remain in integrating multimodal information and deeply understand" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 611, + 555, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 555, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 555, + 707 + ], + "type": "text", + "content": "ing and reasoning about audiovisual content. To extend the boundaries of RAVS and facilitate future research in this field, we propose Omnimodal Referring AudioVisual Segmentation (OmniAVS), a new dataset containing 2,104 videos and 61,095 multimodal referring expressions. OmniAVS stands out with three key innovations: (1) 8 types of multimodal expressions that flexibly combine text, speech, sound, and visual cues; (2) an emphasis" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "Henghui Ding (henghui.ding@gmail.com) is the corresponding author with the Institute of Big Data, College of Computer Science and Artificial Intelligence, Fudan University, Shanghai, China." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22575" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 295, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 295, + 193 + ], + "type": "text", + "content": "on understanding audio content beyond just detecting their presence; and (3) the inclusion of complex reasoning and world knowledge in expressions. Furthermore, we introduce Omnimodal Instructed Segmentation Assistant (OISA), to address the challenges of multimodal reasoning and fine-grained understanding of audiovisual content in OmniAVS. OISA uses MLLM to comprehend complex cues and perform reasoning-based segmentation. Extensive experiments show that OISA outperforms existing methods on OmniAVS and achieves competitive results on other related tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 136, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 136, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 136, + 217 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 223, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 295, + 354 + ], + "type": "text", + "content": "Referring Audio-Visual Segmentation (RAVS) [19, 68] is an emerging field focused on segmenting target objects indicated by referring expressions (e.g., text) within audio-visual scenes using audio-visual cues. The combination of referring and audio distinguishes RAVS from Referring Video Object Segmentation (RVOS) [17, 26, 31, 40, 42, 57], which uses text descriptions to refer to objects in silent videos, and Audio-Visual Segmentation (AVS) [24, 86], which segments any sound-producing object without referring expressions. RAVS enables practical applications in video conferencing and robotics, etc." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 355, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 295, + 653 + ], + "type": "text", + "content": "Understanding Rather than Just Hearing. However, the association between expressions and audio cues in the previous RAVS dataset Ref-AVS [68] remains superficial, primarily limited to basic acoustic properties such as sound occurrence, intensity, and temporal sequence. For example, expressions like \"Who is making the loudest sound?\" only address the surface-level characteristics without delving into the content of the sounds. Recent works [3, 33, 53, 70, 84] in image and video segmentation have evolved from simple semantic perception to advanced reasoning understanding. However, such progress is notably absent in audio-visual segmentation. This gap naturally leads us to explore the potential for complex reasoning in audio-visual scenes. Besides, drawing insights from human auditory perception studies [6], understanding sound content is essential for sophisticated audio-visual scene reasoning. To achieve these goals, we present OmniAVS dataset to advance reasoning-based segmentation in audio-visual scenes. As shown in Figure 2(b), our dataset contains expressions that demand reasoning about audio content, such as \"Who is most likely to be sick?\". These expressions establish complex cognitive chains that go beyond basic acoustic features, e.g., sound " + }, + { + "bbox": [ + 55, + 355, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 355, + 295, + 653 + ], + "type": "text", + "content": " coughing " + }, + { + "bbox": [ + 55, + 355, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 355, + 295, + 653 + ], + "type": "text", + "content": " sickness. Furthermore, we enhance the dataset's interpretability by providing detailed explanations for reasoning-based expressions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "Towards Omnimodal Referring. The emergence of flagship model ChatGPT-4o [51] sets a new standard for modern AI systems, emphasizing the need to process arbitrary combinations of omnimodal inputs, e.g., audio, vision, and text. In the context of RAVS, enabling omnimodal" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 148 + ], + "type": "image", + "image_path": "2f8ac4f9f80a1661769e23ef3e4c9f8eb4e7ba3b937f9f721149a1410f736e8a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 152, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 208 + ], + "type": "text", + "content": "Figure 2. Comparison of (a) Ref-AVS [68] and (b) our proposed OmniAVS. In Ref-AVS, expressions mainly focus on surface-level sound properties, while OmniAVS demands deeper understanding and reasoning about sound content, enabling complex reasoning like identifying potential illness from coughing sounds." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 225, + 555, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 345 + ], + "type": "text", + "content": "inputs can greatly enhance flexible referring and human-machine interactions. However, existing datasets [17, 68] are limited in scope, supporting only single or isolated modalities in their referring expressions. To address this issue, we include diverse expressions in OmniAVS. Beyond the most common text-only expressions, OmniAVS supports text with sound, text with image, and text with both sound and image. Additionally, we allow text to be replaced by speech*, resulting in 8 different kinds of expression types in OmniAVS, as shown in Figure 1." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 346, + 556, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 556, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 556, + 644 + ], + "type": "text", + "content": "Omnimodal Instructed Segmentation Assisant. Given the unique and complex requirements of OmniAVS, which necessitate both flexible interfaces and sophisticated multimodal perception and reasoning capabilities, no existing models can be directly applied to this emerging field. To address this issue, we introduce a baseline model named Omnimodal Instructed Segmentation Assisant, based on Multimodal Large Language Model (MLLM) [13]. This model seamlessly integrates text, speech, sound, and image inputs to perform referring object segmentation while providing explanations for prediction. We identify audio-visual synchronization as a critical factor and implement Audio-Visual Interleaving to achieve temporal alignment without introducing additional parameters. Furthermore, unlike concurrent works [3] that use a single token to segment all frames, we adopt a more lightweight query propagation approach to achieve efficient and accurate segmentation. Experimental results demonstrate that the proposed method OISA possesses strong multimodal reasoning capabilities and supports multiple expression patterns, outperforming other methods on OmniAVS dataset. To further validate the effectiveness of our method OISA, we conduct evaluations on other datasets, including referring video/image segmentation and Ref-AVS, achieving competitive results compared to existing methods across all tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 645, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 555, + 693 + ], + "type": "text", + "content": "This work makes four main contributions: i) we explore omnimodal referring audio-visual segmentation, which is more aligned with real-world scenarios and provides a solid foundation for developing future omnimodal AI systems; ii)" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 703, + 499, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 703, + 499, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 703, + 499, + 713 + ], + "type": "text", + "content": "*To clarify, speech is human-produced spoken sound." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22576" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "we introduce a large-scale referring audio-visual segmentation dataset OmniAVS that enables advanced multimodal reasoning in audio-visual scenes, with a particular emphasis on deep understanding and reasoning from audio content; iii) we enhance interaction flexibility by incorporating multiple referring modalities in the proposed OmniAVS dataset, broadening the scope of referring AVS; iv) we develop OISA, a multimodal large language model designed to uniformly handle omnimodal expressions and reasoning in RAVS, with extensive experiments across different tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 201, + 142, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 142, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 142, + 213 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 222, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 296, + 509 + ], + "type": "text", + "content": "Audio-Visual Scene Perception. Audio-visual scene perception [35-37, 47, 60, 66-68, 72] aims to simulate human multimodal sensory capabilities by integrating auditory and visual information. The rapid development of deep learning has enabled joint audio-visual learning, achieving significant progress in tasks such as event detection [62], robotic manipulation [64], object localization [30], action recognition [8], and object segmentation [68, 86]. The integration of multiple modalities enables models to achieve superior performance compared to single-modality approaches, particularly in handling incomplete or ambiguous information. However, datasets that comprehensively incorporate audio, visual, and language components remain scarce. Among the few available datasets, Music-AVQA [35], AVQA [72], and Ref-AVS [68] stand out as notable examples. Music-AVQA [35] and AVQA [72] primarily address question-answering tasks. Ref-AVS [68] focuses on referring audio-visual segmentation. Nevertheless, Ref-AVS's utilization of audio information is limited, as it employs relatively simple expressions without complex reasoning, thus constraining its effectiveness in real-world scenarios. To overcome these limitations, we introduce OmniAVS, which is specifically designed to maximize the utilization of audio information and seamlessly integrate it with textual descriptions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "Reasoning-based Perception. With the rapid development of MLLMs [2, 13, 16, 43-45, 58, 61, 63, 76, 81], the research community has shifted from perception based on simple semantic [17, 71] to perception based on world knowledge or complex reasoning [3, 25, 33, 53, 70, 82]. For example, GPT4RoI [82] prompts MLLM with bounding boxes to solve region-level complex tasks that require reasoning. DetGPT [53] combines a pretrained MLLM with an open-vocabulary object detector to enable reasoning-based object detection from natural language instructions. LISA [33] uses a MLLM [43] to output a [SEG] token, which is then processed by a mask head [32] to generate the masks. Leveraging the world knowledge and reasoning capabilities inherent in LLMs, LISA demonstrates superior performance in complex segmentation scenarios that require complex reasoning. Building upon this foundation, VideoLISA [3] and VISA [70] extend LISA's capabilities to" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "reasoning video object segmentation by incorporating comprehensive temporal modeling. Drawing inspiration from these works, we propose OISA, which extends reasoning capabilities to audio-visual contexts. Unlike previous works that primarily focus on vision-language reasoning, OISA leverages MLLM to perform reasoning about audio content in videos. While AnyRef [25] pioneers the use of audio and images as referring inputs for segmentation, it is constrained to static images and lacks the ability to reason about audio in dynamic video contexts. Moreover, some works [21, 39, 51] extend LLMs to multiple modalities, e.g., vison, audio, and text, but they cannot be used for segmentation tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 227, + 449, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 449, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 449, + 239 + ], + "type": "text", + "content": "3. Benchmark: OmniAVS" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 247, + 478, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 478, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 478, + 258 + ], + "type": "text", + "content": "3.1. Audio-Visual Video Collection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 264, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 361 + ], + "type": "text", + "content": "The video sources consist of three main parts: i) Real-world web videos under the Creative Commons License, where we use techniques from [10, 68, 86] to align audio and visual snippets with intended semantics; ii) The open-source dataset TVQA [34], featuring extensive dialogue from TV shows; iii) Self-recorded videos, with consent from all participants. We select videos from above that meet the following two standards:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 362, + 553, + 446 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 362, + 553, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 553, + 398 + ], + "type": "text", + "content": "- We encourage to select videos that contain audio content with informative or reasoning-based meanings and corresponding identifiable objects in the scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 399, + 553, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 553, + 446 + ], + "type": "text", + "content": "- We strive to select videos with complex scenes and multiple objects, enabling more diverse referring expressions, e.g., referring to multiple targets, and makes the dataset more representative of real-world complex scenarios." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 449, + 554, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 581 + ], + "type": "text", + "content": "After reviewing about 10,871 potential candidates, we carefully selected the most satisfactory and appropriate videos, which meet the above standards. Finally, we chose 2,104 videos to create the OmniAVS dataset, which is diverse and representative of a wide range of real-world audio-visual scenarios. After collecting the videos, our research team annotates potential objects with expressions and bounding boxes at their first appearance. This information is then used by the annotation team to achieve video mask annotation on the referred objects. Next, we will introduce the corresponding annotation process." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 588, + 552, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 552, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 552, + 601 + ], + "type": "text", + "content": "3.2. Omnimodal Expression and Mask Annotation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "The expression annotation process integrates 4 different modalities: text, speech, sound, and image. We categorize referring expressions into 8 forms: I) Text; II) Speech; III) Text with Sound; IV) Speech with Sound; V) Text with Image; VI) Speech with Image; VII) Text with Sound and Image; and VIII) Speech with Sound and Image. Each expression includes either text or speech, as they provide essential instructions for the model to understand the task and other modalities. All speech expressions are generated" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22577" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 95, + 552, + 182 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 1. Statistical comparison between the newly proposed OmniAVS and other datasets of related referring video segmentation tasks. Avail., Expl., and Expr. are abbreviations for Availability of reasoning, Explanations, and Expressions, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 95, + 552, + 182 + ], + "lines": [ + { + "bbox": [ + 56, + 95, + 552, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 552, + 182 + ], + "type": "table", + "html": "
DatasetVenueContentReferringReasoningStatistics
VideoAudioTextAudioImageAvail.Expl.VideoFrameObjectMaskExpr.Expl.
Refer-YouTubeVOS [57][ECCV'20]✓X✓XXXX3,978116k7,451131k15,009X
MeViS [17][ICCV'23]✓X✓XXXX2,00644k8,175443k28,570X
ReVOS [70][ECCV'24]✓X✓XX✓X1,042116k5,535469k35,074X
Ref-AVS [68][ECCV'24]✓✓✓XXXX4,00240k6,88878k20,261X
OmniAVS (ours)[ICCV'25]✓✓✓✓✓✓✓2,104103k4,277206k61,09534,841
", + "image_path": "020dc339105349dc4462c93ba5c9f092cf1fd86e99e9bbe74278c5c358697793.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 203, + 294, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 203, + 294, + 299 + ], + "spans": [ + { + "bbox": [ + 54, + 203, + 294, + 299 + ], + "type": "text", + "content": "by converting the corresponding text expressions into audio. Following previous works [21, 52], we use a combination of several TTS models [23, 50] and human readings for this conversion. In light of this, we will focus on introducing the annotation process for text-related expressions. Figure 3 shows the distribution of different expression types. The inner areas (e.g., text) are composed of their corresponding outer areas (e.g., 4 components of text)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 300, + 295, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 300, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 295, + 408 + ], + "type": "text", + "content": "1) Text. This is the main component of OmniAVS. The annotation process follows these rules: 1) Expressions should relate to the sound in the video, not just visual cues. 2) Expressions should emphasize the sound's content, not just the act of making it. For example, instead of \"The dog barking\", use \"The dog warning\" as it requires understanding the sound. 3) Encourage reasoning in expressions, with necessary explanations. 4) Expressions can refer to any number of objects, from none to many (more than one)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 410, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 410, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 410, + 295, + 517 + ], + "type": "text", + "content": "III) Text with Sound. This form combines text and sound to provide a more comprehensive expression. The annotation process follows the same rules as in the Text modality, with the following additional requirements when adding audio: 1) Use natural sounds, instrumental music, or environmental sounds, e.g., dog barking, bird chirping. 2) Text should complement the sound, adding context or instructions for identifying the object(s). An example is shown in the 2nd row of Figure 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 519, + 295, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 519, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 54, + 519, + 295, + 603 + ], + "type": "text", + "content": "V)) Text with Image. This type combines text and image. The annotation follows the rules of Text modality with additional considerations for images: 1) Include relevant images that depict the target object(s) or provide context for the audio-visual content. 2) Ensure the text complements the image, offering instructions for identifying the object(s). An example is shown in the 3rd row of Figure 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 605, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 653 + ], + "type": "text", + "content": "VII) Text with Sound and Image. This type combines text, sound, and image in the expressions. The text should guide identification using both sound and image. An example is shown in the 4th row of Figure 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 654, + 295, + 713 + ], + "type": "text", + "content": "Mask Annotation. After collecting expressions, the annotation team proceeds to segment the objects referred to by each expression in every video. Using the bounding box of each object's first appearance (not necessarily in the first frame) annotated in the previous step as a reference," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 213, + 438, + 313 + ], + "blocks": [ + { + "bbox": [ + 318, + 201, + 434, + 211 + ], + "lines": [ + { + "bbox": [ + 318, + 201, + 434, + 211 + ], + "spans": [ + { + "bbox": [ + 318, + 201, + 434, + 211 + ], + "type": "text", + "content": "Table 2. More dataset statistics." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 213, + 438, + 313 + ], + "lines": [ + { + "bbox": [ + 316, + 213, + 438, + 313 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 438, + 313 + ], + "type": "table", + "html": "
StatisticNumber
Words per Expr.9.3
FPS (mean/min/max)4.8/3/15
Expression Types8
Frames (min/max)11/251
Obj. per Expr. (mean)1.12
Obj. per Expr. (max)7
Duration6.2 h
Videos (train/test)1,864/240
Expr. (train/test)54,304/6,791
", + "image_path": "86812771fe330221065ff4e18cb7dde69c4826e805ec43c81ac0e535b25a054e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 459, + 237, + 535, + 312 + ], + "blocks": [ + { + "bbox": [ + 451, + 202, + 542, + 236 + ], + "lines": [ + { + "bbox": [ + 451, + 202, + 542, + 236 + ], + "spans": [ + { + "bbox": [ + 451, + 202, + 542, + 236 + ], + "type": "text", + "content": "Figure 3. Distribution of expression types. SI: sound+image." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 459, + 237, + 535, + 312 + ], + "lines": [ + { + "bbox": [ + 459, + 237, + 535, + 312 + ], + "spans": [ + { + "bbox": [ + 459, + 237, + 535, + 312 + ], + "type": "image", + "image_path": "a7e5b9ec7e2cad5200ca3e187196213bca0260da6ee9c3104d5dc1a79dbc8d16.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 326, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 445 + ], + "type": "text", + "content": "the annotation team is required to carefully confirm the corresponding object in the video, then track and annotate the object's mask in all frames " + }, + { + "bbox": [ + 313, + 326, + 555, + 445 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 326, + 555, + 445 + ], + "type": "text", + "content": ". We designed an interactive annotation tool to automatically load the video and corresponding object bounding boxes. Annotators use the tool to load and preview the corresponding video, then sequentially track and segment target objects in the video. The tool integrates state-of-the-art interactive video object segmentation model SAM2 [56] to assist with mask annotation, enhancing efficiency." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 453, + 480, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 480, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 480, + 464 + ], + "type": "text", + "content": "3.3. Dataset Statistics and Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 470, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 624 + ], + "type": "text", + "content": "Table 1 shows a statistical comparison between OmniAVS and related datasets. Refer-YouTubeVOS [57], MeViS [17], and ReVOS [70] focus on silent videos without audio and provide only text expressions. While ReVOS [70] further supports reasoning expression, it lacks explanations. Ref-AVS Bench [68] includes audio-visual videos but supports only single-modality text expressions without reasoning or explanations. Compared to previous datasets, the proposed OmniAVS dataset offers multimodal content (audio-visual videos) and diverse expression modalities (text, speech, sound, image), supports reasoning with explanations, and allows for referring to arbitrary number of target objects. We compare more datasets in supplementary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 625, + 554, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 554, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 554, + 661 + ], + "type": "text", + "content": "Main Differences with Ref-AVS. As Ref-AVS [68] is the most related dataset to our proposed OmniAVS, we summarize the differences here to facilitate reference:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 662, + 553, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 553, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 553, + 686 + ], + "type": "text", + "content": "- Omnimodal Expressions: While Ref-AVS uses only textual referring expressions, OmniAVS supports expres" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 313, + 693, + 553, + 713 + ], + "type": "text", + "content": " We follow MeViS [17] to perform full-temporal segmentation, even when objects only partially match expressions in specific time segments." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22578" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 95 + ], + "type": "text", + "content": "sions that can combine text, speech, sound, image, providing a richer multimodal interactive interface." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 312 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 96, + 295, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 155 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 155 + ], + "type": "text", + "content": "- Enhanced Audio-Visual Integration: Unlike Ref-AVS, which primarily focuses on \"the sounding object\", Omni-AVS offers a more comprehensive audio-visual analysis. Our expressions delve deeper into the content of sounds and their contextual impact on the scene." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 156, + 295, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 295, + 216 + ], + "type": "text", + "content": "- Reasoning and Explanation: Our dataset includes numerous expressions requiring reasoning, such as \"Who is most likely to be sick?\" (see Figure 2), which necessitates inferring from audio-visual cues like sneezing. We also provide explanations for some reasoning expressions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 216, + 295, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 216, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 216, + 295, + 262 + ], + "type": "text", + "content": "- Number of Referred Target Objects: Unlike Ref-AVS that only refers to 0 or 1 object per expression, OmniAVS allows each referring expression to indicate an arbitrary number of target objects, i.e., from 0 to many as [41]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 312 + ], + "type": "text", + "content": "- Annotation FPS: While Ref-AVS uses fixed video settings (10s, 1 FPS), OmniAVS includes videos of varying durations and annotation frame rates (3-15 FPS) to capture temporal dynamics. Table 2 provides more statistics." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 326, + 148, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 148, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 148, + 338 + ], + "type": "text", + "content": "4. Method: OISA" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 346, + 185, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 185, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 185, + 358 + ], + "type": "text", + "content": "4.1. Overview Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 365, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 509 + ], + "type": "text", + "content": "Because of the unique nature of this task and the proposed OmniAVS dataset, there is no off-the-shelf model available. We propose an Omnimodal Instructed Segmentation Assistant (OISA) as a baseline method to support omnimodal referring and reasoning audio-visual segmentation. As shown in Figure 4, OISA consists of two main components: a MLLM for multimodal understanding and generation, and a mask head for segmentation and tracking. The MLLM includes audio encoder, vision encoder and LLM, while the mask head consists of ViT-Adapter [12], pixel decoder and mask decoder [14]. It's worth noting that the mask head is flexible and can be replaced with models like SAM [32]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "Given an audio-visual video, the multimodal encoder processes video and audio content to obtain corresponding vision and audio tokens. Different from VideoLLaMA [80] that concatenates these two modalities sequentially, we adopt an Audio-Visual Interleaving strategy, as described in Section 4.2, where we divide audio tokens into clips and interleave them with vision tokens. The interleaved sequence forms our Audio-Visual Content Tokens. This approach effectively synchronizes audio and frames without introducing additional parameters, which is particularly useful for scenarios requiring audio-visual alignment. See examples at Figure 2, where without proper audiovisual alignment it would be difficult to determine which object sneezed when multiple objects are making sounds simultaneously. For omnimodal expression inputs, we employ the audio encoder to process speech and sound, and the vision encoder to handle images. The resulting tokens are then" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 220 + ], + "type": "image", + "image_path": "240914d0902495ac7ff33a954f38f4805b028935856e34480c4e60d5aae8db65.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 223, + 555, + 245 + ], + "lines": [ + { + "bbox": [ + 313, + 223, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 555, + 245 + ], + "type": "text", + "content": "Figure 4. Overview of the proposed method OISA. Vision.audio encoders and text embedding are omitted for clarity." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 264, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 397 + ], + "type": "text", + "content": "integrated into the corresponding text tokens to produce the final Omnimodal Expression Tokens. All types of tokens are subsequently fed into the MLLM. The MLLM generates text responses and produces a [SEG] token representing the target object, which is then used by mask decoder for segmentation prediction. Our mask decoder uses query propagation to segment each frame, as described in Section 4.3. The query is refined online during the segmentation process, alleviating dynamic motion. Following previous work [33], we employed cross entropy loss for text generation, DICE loss [49], and binary cross entropy loss for segmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 407, + 459, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 459, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 459, + 420 + ], + "type": "text", + "content": "4.2. Audio-Visual Interleaving" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "For videos of arbitrary length, we uniformly sample " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " frames from the beginning to the end. Each frame is individually processed by the vision encoder to obtain " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "L_{v}" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " vision tokens, denoted as " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{V} = \\{v_{1}, v_{2}, \\ldots, v_{N}\\}" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "v_{i} \\in \\mathbb{R}^{L_{v} \\times d}" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "L_{v}" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " vision tokens of the " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "-th frame with dimension " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": ". For the audio content in the video, we use an audio encoder to process it and obtain audio tokens " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{L_{A} \\times d}" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "L_{A}" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " denotes the total number of audio tokens. To achieve audio-visual alignment, we segment the audio into clips corresponding to the frame rate, obtaining " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{a_{1}, a_{2}, \\ldots, a_{N}\\}" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "a_{i} \\in \\mathbb{R}^{L_{a} \\times d}" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "L_{a} = L_{A} / N" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": ". Next, we interleave audio token clips and vision tokens to form the audio-visual interleaving token sequence " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{v_{1}, a_{1}, v_{2}, a_{2}, \\ldots, v_{N}, a_{N}\\}" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": ". In comparison, VideoLLaMA [15, 80] directly concatenates audio and visual tokens sequentially, " + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{v_{1}, \\ldots, v_{N}, a_{1}, \\ldots, a_{N}\\}" + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": ". However, this approach makes it difficult to align audio frames, especially in segmentation tasks where the video's FPS is not fixed, thereby exacerbating the challenge of audio-visual synchronization. Besides, video-SALMONN [58] fuses audio and visual tokens by concatenating them along the feature dimension and then applying a weight matrix for mapping. This way introduces additional training parameters and causes the fused audio-visual features becoming" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22579" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 293, + 173 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 293, + 173 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 293, + 173 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 293, + 173 + ], + "type": "image", + "image_path": "4f6e8a66ef50314aa6157eb548a0b8d2d30ffb86c37dbef4db6a93a1ce8b393e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 176, + 277, + 188 + ], + "lines": [ + { + "bbox": [ + 72, + 176, + 277, + 188 + ], + "spans": [ + { + "bbox": [ + 72, + 176, + 277, + 188 + ], + "type": "text", + "content": "Figure 5. Comparison of different mask decoder forms." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 276, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 276, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 276, + 217 + ], + "type": "text", + "content": "misaligned with the original pretrained MLLM model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 224, + 168, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 168, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 168, + 236 + ], + "type": "text", + "content": "4.3. Query Propagation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 240, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 636 + ], + "type": "text", + "content": "As shown in Figure 4, once the Audio-Visual Content Tokens and Omnimodal Expression Tokens are obtained, the MLLM generates the corresponding responses. We employ the [SEG] token, following LISA [33], to represent the object embedding, which is subsequently passed to the mask decoder. While MLLMs excel at multimodal understanding and reasoning, they are not good at segmentation tasks. Previous works [3, 33] attach an additional vision encoder for plain feature extraction and a mask head for segmentation after MLLM, resulting in a redundant and suboptimal design. To address these issues, when extracting vision tokens from video, we simultaneously extract corresponding multi-scale features through ViT Adapter [12], which are then enhanced via pixel decoder [14]. These enhanced features, along with the [SEG] generated by the MLLM, are input into the mask decoder. When using mask decoder for segmentation, VideoLISA [3] employs the same [SEG] for the independent segmentation of each frame, aka One-Token-Seg-ALL (OTSA), as shown in Figure 5 (a). However, this approach has limitations. Previous studies [21, 27, 74, 75, 83] have shown that a single query often fails to adequately represent a target object, especially when rapid motion is present within the video. A single query carries a positional prior, making it difficult to encapsulate dynamic motion processes, such as an object moving from right to left, as shown in Figure 5 (a). This limitation results in ID switches, consistently tracking the target on the right part of the video. To address this, we employ a query propagation mechanism, as shown in Figure 5 (b), which allows the query to be updated online on a frame-by-frame basis [27, 79]. This approach enables the query to smoothly capture the temporal motion of the object while also effectively modeling contextual temporal information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 646, + 132, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 132, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 132, + 659 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Implementation Details. We adopt InternVL2-1B [13] as our MLLM, equipped with InternViT-300M-448px [13] as vision encoder and Qwen2-0.5B-Instruct [1] as LLM. Whisper-large-v3 [54] is used as our audio encoder, with an" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "extra audio MLP to project audio features into LLM feature space. For ViT-Adapter [12], we remove all Injectors to reduce memory consumption and ensure vision tokens remain unaffected. For pixel decoder and mask decoder, we follow the default design of Mask2Former [14] but remove the self-attention module in each transformer block since there is only one query input to the mask decoder. We denote our model with the above settings as OISA-1B." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 168, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 204 + ], + "type": "text", + "content": "Dataset Split. We split OmniAVS into training and testing sets, containing 1,864 videos with 54,304 expressions and 240 videos with 6,791 expressions, respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 205, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 300 + ], + "type": "text", + "content": "Evaluation Metrics. Following previous work [17, 18, 20, 68, 77], we adopt " + }, + { + "bbox": [ + 313, + 205, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 313, + 205, + 555, + 300 + ], + "type": "text", + "content": " as the segmentation evaluation metric, where " + }, + { + "bbox": [ + 313, + 205, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 313, + 205, + 555, + 300 + ], + "type": "text", + "content": " measures region similarity (IoU) and " + }, + { + "bbox": [ + 313, + 205, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 205, + 555, + 300 + ], + "type": "text", + "content": " evaluates contour accuracy. Following [41], for no-target expressions that referring to nothing in the video, " + }, + { + "bbox": [ + 313, + 205, + 555, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 313, + 205, + 555, + 300 + ], + "type": "text", + "content": " is set to 1 if the prediction is empty; otherwise, it is 0. We employ METEOR [4] as evaluation metric for text explanation generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 309, + 411, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 411, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 411, + 322 + ], + "type": "text", + "content": "5.1. Training Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 327, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 422 + ], + "type": "text", + "content": "Audio-Text Alignment. Since InternVL2 does not support audio input natively, we introduce an audio encoder MLP to align audio features with the LLM space. This stage mainly aims to align audio features with LLM space. To achieve this goal, we use Automatic Speech Recognition datasets [9] and Audio Caption datasets [59] for audio-text alignment. During this process, we only train the audio encoder MLP while keeping all other parameters frozen." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 423, + 555, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 614 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 614 + ], + "type": "text", + "content": "Omnimodal Instruct Segmentation Tuning. Following previous works [3, 33, 70, 84], we use multiple datasets for training, including 1) semantic segmentation datasets: ADE20K[85], COCO-Stuff[7], PASCAL-Part [11], PACOLVIS [55]; 2) referring segmentation datasets: RefCOCO, RefCOCO+ [78], RefCOCOg [48], ReasonSeg [33]; 4) referring video segmentation datasets: Refer-YouTube-VOS [57], Refer-DAVIS-17 [31], MeViS [17], ReVOS [70]; 5) audio-visual segmentation datasets: Ref-AVS Bench [68], OmniAVS. We use LoRA [29] to fine-tune LLM, train all parameters of the mask head, and freeze all remaining parameters. For video samples, during training, we uniformly sample 10 frames from the entire video and follow [3] to use 4 frames as dense frames and the remaining frames as sparse frames. During inference, we sample 32 frames with 4 frames as dense frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 624, + 408, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 408, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 408, + 636 + ], + "type": "text", + "content": "5.2. Ablation Study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "Ablation Study on Audio-Visual Fusion. In Table 3 we evaluate different audio-visual fusion strategies. For attention baseline, we adopt a simple cross-attention block where image tokens serve as key and audio tokens serve as query/value. For weighted sum, we first pad the audio tokens to match the length of vision tokens, then perform" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22580" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 127, + 291, + 196 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 125 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 125 + ], + "type": "text", + "content": "Table 3. Ablation study on audio-visual fusion. AVI means Audio-Visual Interleaving. TVQA Split represents the performance evaluated on samples collected from TVQA [34]. These samples demand precise audio-visual alignment and feature a substantial amount of dialogue. See an example in Figure 6 (a)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 127, + 291, + 196 + ], + "lines": [ + { + "bbox": [ + 59, + 127, + 291, + 196 + ], + "spans": [ + { + "bbox": [ + 59, + 127, + 291, + 196 + ], + "type": "table", + "html": "
Fusion TypeSketchTVQA SplitOverall
Attention-37.435.8
Concat36.935.3
Weighted Sum34.831.1
AVI40.839.2
AVI + Concat42.040.5
", + "image_path": "486be35dea0a1729f39117ebe5a1d24f3edd9da04155f39278e94ed7a24f0ee1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 242, + 294, + 310 + ], + "blocks": [ + { + "bbox": [ + 55, + 208, + 295, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 240 + ], + "type": "text", + "content": "Table 4. Ablation study of mask head. OTSA: One-Token-Seg-All [3]; QP: query propagation (Ours); M2F: Mask2Former. FPS is tested on one A6000 GPU (batch size=1)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 242, + 294, + 310 + ], + "lines": [ + { + "bbox": [ + 59, + 242, + 294, + 310 + ], + "spans": [ + { + "bbox": [ + 59, + 242, + 294, + 310 + ], + "type": "table", + "html": "
IDQuery TypeHead TypeOmniAVS Metrics
OTSAQPM2FSAMJ&FJFFPS
I✓XX✓38.136.439.84.3
II✓X✓X35.233.836.615.7
IIIX✓X✓41.239.243.24.1
IVX✓✓X40.538.742.312.3
", + "image_path": "52a9a8d6c1c2cf50153d4888fe7c156aef481bcf1943fadb25cbbfe03c79bfc0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "type": "text", + "content": "weighted sum on them. Attention-based and concatenation [15] methods achieve similar performance around " + }, + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "type": "inline_equation", + "content": "35 - 37\\%" + }, + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "type": "text", + "content": ". Our proposed Audio-Visual Interleaving (AVI) strategy significantly improves the overall performance to " + }, + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "type": "inline_equation", + "content": "39.2\\%" + }, + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "type": "text", + "content": " and get " + }, + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "type": "inline_equation", + "content": "40.2\\%" + }, + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "type": "text", + "content": " in TVQA split. Previous works [13, 44] add a thumbnail after tiling to capture global multimodal features to capture global features more effectively. Similarly, we append original audio tokens A at the end of the audio-visual interleaving sequence. This supplements complete and untruncated audio tokens based on our AVI audio-visual alignment, enabling the MLLM to handle more comprehensive audio tokens and thereby improving the model performance to " + }, + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "type": "inline_equation", + "content": "40.5\\%" + }, + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{J}\\& \\mathcal{F}" + }, + { + "bbox": [ + 54, + 325, + 295, + 479 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "text", + "content": "Ablation Study on Mask Head. In Table 4 we study different mask head designs. While SAM mask head achieves competitive performance compared to M2F, it suffers from slower inference speed due to its huge ViT backbone. When using M2F as the mask head (Setting II and IV), we achieve at least " + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "text", + "content": " speedup with a slight accuracy trade-off. Regarding query types, our QP approach (Setting III and IV) significantly outperforms OTSA in segmentation quality, improving " + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{J} \\& \\mathcal{F}" + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "inline_equation", + "content": "38.1\\%" + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "inline_equation", + "content": "41.2\\%" + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "text", + "content": " when using SAM mask head, and from " + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "inline_equation", + "content": "35.2\\%" + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "inline_equation", + "content": "40.5\\%" + }, + { + "bbox": [ + 55, + 481, + 295, + 636 + ], + "type": "text", + "content": " when using M2F mask head. Considering the balance between performance and efficiency, we choose Setting IV (QP with M2F head) for our subsequent experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 647, + 216, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 216, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 216, + 658 + ], + "type": "text", + "content": "5.3. OmniAVS Benchmark Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "In Table 5, we report the OmniAVS benchmark results. For traditional referring segmentation methods [17, 68, 71], we process each modality independently and concatenate their features to create a unified query representation." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 95, + 553, + 166 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 5. Testing on OmniAVS. We use " + }, + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "type": "inline_equation", + "content": "\\mathcal{I}\\&\\mathcal{F}" + }, + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "type": "text", + "content": " as the default metric. All is the average result across 8 splits. MET: METEOR." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 95, + 553, + 166 + ], + "lines": [ + { + "bbox": [ + 317, + 95, + 553, + 166 + ], + "spans": [ + { + "bbox": [ + 317, + 95, + 553, + 166 + ], + "type": "table", + "html": "
MethodAllIIIIIIIVVVIVIIVIIIMET.
LMPM [17]25.831.228.720.022.721.320.930.031.4-
EEMC [68]29.634.432.619.626.028.024.735.636.0-
MUTR [71]32.335.433.328.429.826.522.841.640.5-
LISA-7B [33]33.633.331.229.232.728.627.343.443.111.6
LISA-13B [33]36.136.432.130.435.731.630.246.745.716.5
OISA-1B (ours)41.140.138.534.938.535.935.252.653.021.7
", + "image_path": "f95e3fc287638a93422a9f7307730208dbb1c33e74710a6c286f18d483acd328.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "text", + "content": "Regarding LISA [33], we enhance its audio processing capability by applying the same audio-text alignment method as our OISA, while maintaining its frame-by-frame inference approach. Our OISA-1B achieves state-of-the-art performance on the OmniAVS benchmark, significantly surpassing previous methods. Specifically, it attains an average score of " + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "inline_equation", + "content": "41.1\\%" + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "text", + "content": " across all modality combinations, outperforming LISA-13B by " + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "inline_equation", + "content": "5.0\\%" + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "text", + "content": ". These improvements are consistent across all splits, with the most notable gain of " + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "inline_equation", + "content": "7.3\\%" + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "text", + "content": " on split VIII (speech+sound+image), showcasing our model's superior ability in multimodal input processing and reasoning. In terms of explanation generation, our model achieves a METEOR score of " + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "inline_equation", + "content": "21.7\\%" + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "text", + "content": ", significantly surpassing LISA-13B's score of " + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "inline_equation", + "content": "16.5\\%" + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "text", + "content": ". This indicates that our model excels not only in segmentation tasks but also in understanding and explaining its decisions. Analysis across different modality combinations reveals that splits VII (text+speech+image) and VIII (text+sound+image) achieve the highest scores of " + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "inline_equation", + "content": "52.6\\%" + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "inline_equation", + "content": "53.0\\%" + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "text", + "content": ", respectively. This suggests that incorporating more modalities generally enhances performance, as the model effectively leverages complementary information from various sources. Comparison with the previous dataset Ref-AVS Bench shows that our OISA-1B achieved a score of " + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "inline_equation", + "content": "58.0\\%" + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "text", + "content": " (shown in Table 6), whereas it only attained " + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "inline_equation", + "content": "41.1\\%" + }, + { + "bbox": [ + 313, + 194, + 555, + 517 + ], + "type": "text", + "content": " on our OmniAVS. This significant difference underscores the increased difficulty and complexity of our dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 713 + ], + "type": "text", + "content": "Qualitative Results and Failure Cases. In Figure 6, we present several success and failure cases of the proposed baseline method OISA. Case (a) demonstrates that OISA effectively aligns audio with visual content, identifies key dialogues in the fourth frame, and subsequently performs reasoning and accurately locates the target object, providing a clear explanation. Besides, Case (a) shows that OISA achieves robust tracking in scenarios involving viewpoint changes, occlusions, and reappearance. Case (b) illustrates the model's ability to handle multimodal expressions by understanding both the fire in the image and the siren sound to locate the fire truck. However, as shown in Case (c), OISA struggles in scenarios with complex sounds, indicating that understanding intricate sounds still requires further exploration, such as decoupling sound events [28, 46] before perception and reasoning." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22581" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 292, + 196 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 292, + 196 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 292, + 196 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 292, + 196 + ], + "type": "image", + "image_path": "80dbe08a52221f8e0efaaacb7d611cb6bccaf5daf49b21813c1135dfda3e411c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 200, + 263, + 210 + ], + "lines": [ + { + "bbox": [ + 87, + 200, + 263, + 210 + ], + "spans": [ + { + "bbox": [ + 87, + 200, + 263, + 210 + ], + "type": "text", + "content": "Figure 6. Success and failure cases of OISA-1B." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 239, + 292, + 331 + ], + "blocks": [ + { + "bbox": [ + 78, + 227, + 271, + 237 + ], + "lines": [ + { + "bbox": [ + 78, + 227, + 271, + 237 + ], + "spans": [ + { + "bbox": [ + 78, + 227, + 271, + 237 + ], + "type": "text", + "content": "Table 6. Performance on testing set of Ref-AVS [68]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 239, + 292, + 331 + ], + "lines": [ + { + "bbox": [ + 59, + 239, + 292, + 331 + ], + "spans": [ + { + "bbox": [ + 59, + 239, + 292, + 331 + ], + "type": "table", + "html": "
MethodSeenUnseenMix (S+U)Null
JFJFJFS (↓)
AVSBench [86]23.251.132.454.727.852.920.8
AVGSegFormer [22]33.547.036.150.134.848.617.1
GAVS [65]28.949.829.849.729.449.819.0
ReferFormer [69]31.350.130.448.830.949.517.6
R2VOS [38]25.041.027.949.826.545.418.3
EEMC [68]34.251.349.564.841.958.10.7
OISA-1B (ours)51.758.758.365.154.561.49.8
", + "image_path": "08c11c3eee3dd3426f5487a177f14420b66ebe929fb6808a7f5996ec5257a0b0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 356, + 194, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 194, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 194, + 366 + ], + "type": "text", + "content": "5.4. Results on Related Tasks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "text", + "content": "Results on Ref-AVS. As shown in Table 6, our OISA-1B achieves SOTA performance on Ref-AVS [68], greatly surpassing previous methods. Specifically, OISA-1B improves the " + }, + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "inline_equation", + "content": "17.5\\%" + }, + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "inline_equation", + "content": "8.8\\%" + }, + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "text", + "content": " on Seen and Unseen splits respectively compared to EEMC [68], demonstrating good generalization ability. While OISA-1B gets suboptimal score on Null split, which may inherit from hallucination in MLLM, EEMC achieves " + }, + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "text", + "content": ". We argue that videos in Null split contain empty or fixed audio with unrelated expressions, so EEMC is likely overfitting to identify audio patterns without understanding. Therefore, our model's performance " + }, + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "inline_equation", + "content": "(9.8\\%)" + }, + { + "bbox": [ + 55, + 374, + 295, + 529 + ], + "type": "text", + "content": " is reasonable, and OISA-1B's advantages on other metrics also demonstrate its robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 532, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 615 + ], + "type": "text", + "content": "Results on Referring Image Segmentation. As shown in Table 7, our OISA-1B outperforms previous referring image segmentation methods on refCOCO series [48, 78], achieving state-of-the-art performance. The performance on ReasonSeg [33] falls short of LISA [33] and VideoLISA [3], likely due to the limited reasoning abilities of our smaller 0.5B LLM, which are crucial for this dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Results on Referring Video Segmentation. As shown in Table 8, the proposed OISA-1B achieves competitive results on MeViS [17] (43.2%), R-YTVOS [57] (62.1%), and R-DAVIS" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "_{17}" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " [31] (65.2%). On ReVOS [70], it achieves new state-of-the-art results with 47.3% " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I} \\& \\mathcal{F}" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " and 19.3% robustness score, surpassing VISA [70] by 0.2% and 4.0% respectively, demonstrating strong perception and reasoning ability in complex video scenes." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 83, + 552, + 163 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 553, + 82 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 553, + 82 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 553, + 82 + ], + "type": "text", + "content": "Table 7. cIoU Results on refCOCO [48, 78] and ReasonSeg [33]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 83, + 552, + 163 + ], + "lines": [ + { + "bbox": [ + 318, + 83, + 552, + 163 + ], + "spans": [ + { + "bbox": [ + 318, + 83, + 552, + 163 + ], + "type": "table", + "html": "
MethodrefCOCOrefCOCO+refCOCogReasonSeg
valtestAtestBvaltestAtestBvalUtestUvaltest
LAVT [73]72.775.868.862.168.455.161.262.1--
ReLA [41]73.876.570.266.071.057.765.066.0--
LISA-7B [33]74.176.571.162.467.456.566.468.554.048.4
LISA-13B [33]--------62.951.1
VISA-7B [70]72.475.568.159.864.853.165.566.457.8-
VideoLISA-3.8B [3]73.876.668.863.468.856.268.368.867.154.4
OISA-1B (ours)74.278.372.068.871.762.070.672.361.250.8
", + "image_path": "a84841f22072d92c2297bab1679fa10dd84a6dedd300fb40d7498c9eab1863ee.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 198, + 553, + 286 + ], + "blocks": [ + { + "bbox": [ + 314, + 175, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 314, + 175, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 314, + 175, + 553, + 196 + ], + "type": "text", + "content": "Table 8. Results on referring/reasoning video segmentation [17, 31, 57, 70]. " + }, + { + "bbox": [ + 314, + 175, + 553, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{I}\\&\\mathcal{F}" + }, + { + "bbox": [ + 314, + 175, + 553, + 196 + ], + "type": "text", + "content": " is the default metric, " + }, + { + "bbox": [ + 314, + 175, + 553, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 314, + 175, + 553, + 196 + ], + "type": "text", + "content": " is the robustness score." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 198, + 553, + 286 + ], + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 286 + ], + "type": "table", + "html": "
MethodMeViSR-YTVOSR-DAVIS17ReVOSReVOS(R)
ReferFormer [69]31.062.961.114.94.9
MTTR [5]30.055.3-25.55.6
LMPM [17]37.2--26.43.2
LISA-7B [33]-50.258.440.99.3
TrackGPT-7B [87]40.156.463.243.611.6
VideoLISA-3.8B [3]42.361.767.7--
VISA-7B [70]44.563.070.447.115.3
OISA-1B (ours)43.262.165.247.319.3
", + "image_path": "e4f002580f03febccb1047ae90258fbc370d5d7947fc3af60c3ef7bd361f19ae.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 302, + 465, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 465, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 465, + 314 + ], + "type": "text", + "content": "6. Conclusion and Discussion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 322, + 555, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 322, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 312, + 322, + 555, + 490 + ], + "type": "text", + "content": "We introduce OmniAVS, a large-scale dataset supporting advanced multimodal referring and reasoning in audiovisual scenes with 8 types of expressions across text, speech, sound, and image. Compared to previous referring datasets focused on salient video or sound occurrence, OmniAVS enables deeper understanding and reasoning of audiovisual content. To address the challenges posed by OmniAVS, we propose a simple yet effective baseline OISA, a segmentation model based on multimodal large language models. Our OISA leverages audio-visual interleaving for temporal alignment and query propagation for efficient segmentation, enabling effective handling of multimodal inputs without additional parameters. Extensive experiments on 10 datasets demonstrate the effectiveness of our OISA." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 491, + 556, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 556, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 556, + 682 + ], + "type": "text", + "content": "Future Directions. Although our baseline OISA has shown promising results on OmniAVS, there is still much room for improvement: i) Exploring more efficient audiovisual fusion methods, such as learning more robust audiovisual representations. ii) Disentangling sounds when multiple objects make sounds simultaneously, which is crucial for subsequent reasoning. iii) How to effectively combine multiple modalities in expressions by using joint representations or cross-modal fusion. iv) Exploring more advanced segmentation models to handle complex scenarios, e.g., occlusion, disappearance and reappearance. v) Pre-training OISA on larger audio-visual datasets to further improve its generalization to unseen tasks. vi) Developing more interactive and conversational capabilities to enable multi-round interactions with users, allowing for clarification and refinement of referring expressions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 555, + 712 + ], + "type": "text", + "content": "Acknowledgement. This project was supported by the National Natural Science Foundation of China (NSFC) under Grant No. 62472104." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22582" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen Technical Report. arXiv, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 179 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-VL: A Frontier Large Vision-Language Model with Versatile Abilities. arXiv preprint arXiv:2308.12966, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 236 + ], + "type": "text", + "content": "[3] Zechen Bai, Tong He, Haiyang Mei, Pichao WANG, Ziteng Gao, Joya Chen, liulei, Zheng Zhang, and Mike Zheng Shou. One Token to Seg Them All: Language Instructed Reasoning Segmentation in Videos. In Adv. Neural Inform. Process. Syst., 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 238, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 238, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 294, + 281 + ], + "type": "text", + "content": "[4] Satanjeev Banerjee and Alon Lavie. METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments. In Assoc. Comput. Linguist. Worksh., 2005. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 327 + ], + "type": "text", + "content": "[5] Adam Botach, Evgenii Zheltonozhskii, and Chaim Baskin. End-to-End Referring Video Object Segmentation with Multimodal Transformers. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 328, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 295, + 350 + ], + "type": "text", + "content": "[6] Albert S Bregman. Auditory Scene Analysis: The Perceptual Organization of Sound. MIT press, 1994. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 294, + 385 + ], + "type": "text", + "content": "[7] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. COCO-Stuff: Thing and Stuff Classes in Context. In IEEE Conf. Comput. Vis. Pattern Recog., 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 387, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 294, + 430 + ], + "type": "text", + "content": "[8] Jacob Chalk, Jaesung Huh, Evangelos Kazakos, Andrew Zisserman, and Dima Damen. TIM: A Time Interval Machine for Audio-Visual Action Recognition. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 432, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 295, + 464 + ], + "type": "text", + "content": "[9] Guoguo Chen, Shuzhou Chai, et al. GigaSpeech: An Evolving, Multi-domain ASR Corpus with 10,000 Hours of Transcribed Audio. In Proc. Interspeech 2021, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 466, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 466, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 57, + 466, + 294, + 499 + ], + "type": "text", + "content": "[10] Honglie Chen, Weidi Xie, Andrea Vedaldi, and Andrew Zisserman. VGGSound: A Large-scale Audio-Visual Dataset. In IEEE Int. Conf. Acoust. Speech Signal Process., 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 501, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 501, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 501, + 294, + 554 + ], + "type": "text", + "content": "[11] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect What You Can: Detecting and Representing Objects using Holistic Models and Body Parts. In IEEE Conf. Comput. Vis. Pattern Recog., 2014. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 556, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 294, + 600 + ], + "type": "text", + "content": "[12] Zhe Chen, Yuchen Duan, Wenhai Wang, Junjun He, Tong Lu, Jifeng Dai, and Yu Qiao. Vision Transformer Adapter for Dense Predictions. In Int. Conf. Learn. Represent., 2023. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 602, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 667 + ], + "type": "text", + "content": "[13] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How Far Are We to GPT-4V? Closing the Gap to Commercial Multimodal Models with Open-Source Suites. arXiv preprint arXiv:2404.16821, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[14] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 5, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[15] Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, et al. VideoLLaMA 2: Advancing Spatial-Temporal Modeling and Audio Understanding in Video-LLMs. arXiv preprint arXiv:2406.07476, 2024. 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "type": "text", + "content": "[16] Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, et al. Qwen2-Audio Technical Report. arXiv preprint arXiv:2407.10759, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "text", + "content": "[17] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, and Chen Change Loy. MeViS: A Large-scale Benchmark for Video Segmentation with Motion Expressions. In Int. Conf. Comput. Vis., 2023. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 259 + ], + "type": "text", + "content": "[18] Henghui Ding, Chang Liu, Shuting He, Xudong Jiang, Philip HS Torr, and Song Bai. MOSE: A New Dataset for Video Object Segmentation in Complex Scenes. In Int. Conf. Comput. Vis., 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 293 + ], + "type": "text", + "content": "[19] Henghui Ding, Song Tang, Shuting He, Chang Liu, Zuxuan Wu, and Yu-Gang Jiang. Multimodal referring segmentation: A survey. arXiv, 2025. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 294, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 337 + ], + "type": "text", + "content": "[20] Henghui Ding, Kaining Ying, Chang Liu, Shuting He, YuGang Jiang, Philip HS Torr, and Song Bai. MOSEv2: A more challenging dataset for video object segmentation in complex scenes. arXiv, 2025. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[21] Chaoyou Fu, Haojia Lin, Zuwei Long, Yunhang Shen, Meng Zhao, Yifan Zhang, Xiong Wang, Di Yin, Long Ma, Xiawu Zheng, et al. VITA: Towards Open-Source Interactive Omni Multimodal LLM. arXiv, 2024. 3, 4, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "text", + "content": "[22] Shengyi Gao, Zhe Chen, Guo Chen, Wenhai Wang, and Tong Lu. AVSegFormer: Audio-Visual Segmentation with Transformer. In AAAI, 2024. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 415, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 437 + ], + "type": "text", + "content": "[23] GPT-SoVITS. https://github.com/RVC-Boss/GPT-SoVITS, 2024.4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "text", + "content": "[24] Ruohao Guo, Liao Qu, Dantong Niu, Yanyu Qi, Wenzhen Yue, Ji Shi, Bowei Xing, and Xianghua Ying. Open-Vocabulary Audio-Visual Semantic Segmentation. In ACM Int. Conf. Multimedia, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 482, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 535 + ], + "type": "text", + "content": "[25] Junwen He, Yifan Wang, Lijun Wang, Huchuan Lu, JunYan He, Jin-Peng Lan, Bin Luo, and Xuansong Xie. Multimodal Instruction Tuned LLMs with Fine-grained Visual Perception. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 537, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 579 + ], + "type": "text", + "content": "[26] Shuting He and Henghui Ding. Decoupling Static and Hierarchical Motion Perception for Referring Video Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "type": "text", + "content": "[27] Miran Heo, Sukjun Hwang, Jeongseok Hyun, Hanjung Kim, Seoung Wug Oh, Joon-Young Lee, and Seon Joo Kim. A Generalized Framework for Video Instance Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[28] John R Hershey, Zhuo Chen, Jonathan Le Roux, and Shinji Watanabe. Deep clustering: Discriminative embeddings for segmentation and separation. In IEEE Int. Conf. Acoust. Speech Signal Process., 2016. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In Int. Conf. Learn. Represent., 2022. 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22583" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[30] Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Egocentric Audio-Visual Object Localization. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 139 + ], + "type": "text", + "content": "[31] Anna Khoreva, Anna Rohrbach, and Bernt Schiele. Video Object Segmentation with Language Referring Expressions. In ACCV, 2019. 2, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 183 + ], + "type": "text", + "content": "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment Anything. In Int. Conf. Comput. Vis., 2023. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 228 + ], + "type": "text", + "content": "[33] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. LISA: Reasoning Segmentation via Large Language Model. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 270 + ], + "type": "text", + "content": "[34] Jie Lei, Licheng Yu, Mohit Bansal, and Tamara L Berg. TVQA: Localized, Compositional Video Question Answering. In Proc. of the Conf. on Empirical Methods in Nat. Lang. Process., 2018. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 271, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 296, + 316 + ], + "type": "text", + "content": "[35] Guangyao Li, Yake Wei, Yapeng Tian, Chenliang Xu, JiRong Wen, and Di Hu. Learning to Answer Questions in Dynamic Audio-Visual Scenarios. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 296, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 349 + ], + "type": "text", + "content": "[36] Guangyao Li, Henghui Du, and Di Hu. Boosting Audio Visual Question Answering via Key Semantic-Aware Cues. In ACM Int. Conf. Multimedia, pages 5997-6005, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 296, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 382 + ], + "type": "text", + "content": "[37] Jia Li, Wenjie Zhao, Ziru Huang, Yunhui Guo, and Yapeng Tian. Do Audio-Visual Segmentation Models Truly Segment Sounding Objects? arXiv, 2025. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 296, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 296, + 425 + ], + "type": "text", + "content": "[38] Xiang Li, Jinglu Wang, Xiaohao Xu, Xiao Li, Bhiksha Raj, and Yan Lu. Robust Referring Video Object Segmentation with Cyclic Structural Consensus. In Int. Conf. Comput. Vis., 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "type": "text", + "content": "[39] Yadong Li, Haoze Sun, Mingan Lin, Tianpeng Li, Guosheng Dong, Tao Zhang, Bowen Ding, Wei Song, Zhenglin Cheng, Yuqi Huo, et al. Baichuan-Omni Technical Report. arXiv preprint arXiv:2410.08565, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "text", + "content": "[40] Linfeng Yuan and Miaojing Shi and Zijie Yue and Qijun Chen. Losh: Long-short text joint prediction network for referring video object segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 514, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 296, + 548 + ], + "type": "text", + "content": "[41] Chang Liu, Henghui Ding, and Xudong Jiang. GRES: Generalized Referring Expression Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 5, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 581 + ], + "type": "text", + "content": "[42] Chang Liu, Xudong Jiang, and Henghui Ding. Primitivenet: decomposing the global constraints for referring segmentation. Visual Intelligence, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 582, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 613 + ], + "type": "text", + "content": "[43] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual Instruction Tuning. In Adv. Neural Inform. Process. Syst., 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 614, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 647 + ], + "type": "text", + "content": "[44] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved Baselines with Visual Instruction Tuning. In IEEE Conf. Comput. Vis. Pattern Recog., 2024. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "type": "text", + "content": "[45] Shuo Liu, Kaining Ying, Hao Zhang, Yue Yang, Yuqi Lin, Tianle Zhang, Chuanhao Li, Yu Qiao, Ping Luo, Wenqi Shao, et al. ConvBench: A Multi-Turn Conversation Evaluation Benchmark with Hierarchical Ablation Capability for Large Vision-Language Models. In Adv. Neural Inform. Process. Syst. D&B, 2024. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 315, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 554, + 116 + ], + "type": "text", + "content": "[46] Yi Luo and Nima Mesgarani. Conv-TasNet: Surpassing Ideal Time-Frequency Magnitude Masking for Speech Separation. IEEE/ACM Trans. Audio Speech Lang. Process., 27 (8), 2019. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 554, + 162 + ], + "type": "text", + "content": "[47] Juncheng Ma, Peiwen Sun, Yaoting Wang, and Di Hu. Stepping Stones: A Progressive Training Strategy for Audio-Visual Semantic Segmentation. In Eur. Conf. Comput. Vis., 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 164, + 554, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 164, + 554, + 209 + ], + "spans": [ + { + "bbox": [ + 315, + 164, + 554, + 209 + ], + "type": "text", + "content": "[48] Junhua Mao, Jonathan Huang, Alexander Toshev, Oana Camburu, Alan L Yuille, and Kevin Murphy. Generation and Comprehension of Unambiguous Object Descriptions. In IEEE Conf. Comput. Vis. Pattern Recog., 2016. 6, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 210, + 554, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 210, + 554, + 253 + ], + "spans": [ + { + "bbox": [ + 315, + 210, + 554, + 253 + ], + "type": "text", + "content": "[49] Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation. In IEEE Int. Conf. 3D Vis., 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 256, + 554, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 256, + 554, + 278 + ], + "spans": [ + { + "bbox": [ + 315, + 256, + 554, + 278 + ], + "type": "text", + "content": "[50] OpenAI. https://platform.openai.com/docs/guides/text-to-speech, 2023.4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 280, + 554, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 280, + 554, + 302 + ], + "spans": [ + { + "bbox": [ + 315, + 280, + 554, + 302 + ], + "type": "text", + "content": "[51] OpenAI. https://openai.com/index/hello-gpt-4o, 2024.2,3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 304, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 304, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 315, + 304, + 554, + 360 + ], + "type": "text", + "content": "[52] Wenwen Pan, Haonan Shi, Zhou Zhao, Jieming Zhu, Xi-uqiang He, Zhigeng Pan, Lianli Gao, Jun Yu, Fei Wu, and Qi Tian. Wnet: Audio-Guided Video Object Segmentation via Wavelet-Based Cross-Modal Denoising Networks. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 361, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 361, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 315, + 361, + 554, + 415 + ], + "type": "text", + "content": "[53] Renjie Pi, Jiahui Gao, Shizhe Diao, Rui Pan, Hanze Dong, Jipeng Zhang, Lewei Yao, Jianhua Han, Hang Xu, Lingpeng Kong, et al. DetGPT: Detect What You Need via Reasoning. In Proc. of the Conf. on Empirical Methods in Nat. Lang. Process., 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 418, + 554, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 418, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 315, + 418, + 554, + 460 + ], + "type": "text", + "content": "[54] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust Speech Recognition via Large-Scale Weak Supervision. In Int. Conf. Mach. Learn., 2023. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 464, + 554, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 464, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 315, + 464, + 554, + 518 + ], + "type": "text", + "content": "[55] Vignesh Ramanathan, Anmol Kalia, Vladan Petrovic, Yi Wen, Baixue Zheng, Baishan Guo, Rui Wang, Aaron Marquez, Rama Kovvuri, Abhishek Kadian, et al. PACO: Parts and Attributes of Common Objects. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 520, + 554, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 520, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 315, + 520, + 554, + 575 + ], + "type": "text", + "content": "[56] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, et al. SAM 2: Segment Anything in Images and Videos. arXiv preprint arXiv:2408.00714, 2024. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 577, + 554, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 577, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 315, + 577, + 554, + 620 + ], + "type": "text", + "content": "[57] Seonguk Seo, Joon-Young Lee, and Bohyung Han. URVOS: Unified Referring Video Object Segmentation Network with a Large-Scale Benchmark. In Eur. Conf. Comput. Vis., 2020. 2, 4, 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 623, + 554, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 623, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 315, + 623, + 554, + 677 + ], + "type": "text", + "content": "[58] Guangzhi Sun, Wenyi Yu, Changli Tang, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, Yuxuan Wang, and Chao Zhang. video-SALMONN: Speech-Enhanced Audio-Visual Large Language Models. In Int. Conf. Mach. Learn., 2024. 3, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 680, + 554, + 713 + ], + "type": "text", + "content": "[59] Luoyi Sun, Xuanan Xu, Mengyue Wu, and Weidi Xie. Auto-ACD: A Large-scale Dataset for Audio-Language Representation Learning. In ACM Int. Conf. Multimedia, 2024. 6" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22584" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[60] Peiwen Sun, Honggang Zhang, and Di Hu. Unveiling and Mitigating Bias in Audio Visual Segmentation. In ACM Int. Conf. Multimedia, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 150 + ], + "type": "text", + "content": "[61] Changli Tang, Wenyi Yu, Guangzhi Sun, Xianzhao Chen, Tian Tan, Wei Li, Lu Lu, Zejun Ma, and Chao Zhang. SALMONN: Towards Generic Hearing Abilities for Large Language Models. In Int. Conf. Learn. Represent., 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "text", + "content": "[62] Yapeng Tian, Jing Shi, Bochen Li, Zhiyao Duan, and Chen-liang Xu. Audio-Visual Event Localization in Unconstrained Videos. In Eur. Conf. Comput. Vis., 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "type": "text", + "content": "[63] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution. arXiv preprint arXiv:2409.12191, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "type": "text", + "content": "[64] Yefei Wang, Kaili Wang, Yi Wang, Di Guo, Huaping Liu, and Fuchun Sun. Audio-Visual Grounding Referring Expression for Robotic Manipulation. In IEEE Int. Conf. Robot. Autom., 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "text", + "content": "[65] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting Segmentation with Sound Is Generalizable Audio-Visual Source Localization. In AAAI, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "text", + "content": "[66] Yaoting Wang, Weisong Liu, Guangyao Li, Jian Ding, Di Hu, and Xi Li. Prompting Segmentation with Sound Is Generalizable Audio-Visual Source Localization. In AAAI, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "type": "text", + "content": "[67] Yaoting Wang, Peiwen Sun, Yuanchao Li, Honggang Zhang, and Di Hu. Can Textual Semantics Mitigate Sounding Object Segmentation Preference? In Eur. Conf. Comput. Vis., 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "text", + "content": "[68] Yaoting Wang, Peiwen Sun, Dongzhan Zhou, Guangyao Li, Honggang Zhang, and Di Hu. Ref-AVS: Refer and Segment Objects in Audio-Visual Scenes. In Eur. Conf. Comput. Vis., 2024. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "text", + "content": "[69] Jiannan Wu, Yi Jiang, Peize Sun, Zehuan Yuan, and Ping Luo. Language as Queries for Referring Video Object Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 492, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 547 + ], + "type": "text", + "content": "[70] Cilin Yan, Haochen Wang, Shilin Yan, Xiaolong Jiang, Yao Hu, Guoliang Kang, Weidi Xie, and Efstratios Gavves. VISA: Reasoning Video Object Segmentation via Large Language Models. In Eur. Conf. Comput. Vis., 2024. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 602 + ], + "type": "text", + "content": "[71] Shilin Yan, Renrui Zhang, Ziyu Guo, Wenchao Chen, Wei Zhang, Hongyang Li, Yu Qiao, Hao Dong, Zhongjiang He, and Peng Gao. Referred by Multi-Modality: A Unified Temporal Transformer for Video Object Segmentation. In AAAI, 2024. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 647 + ], + "type": "text", + "content": "[72] Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. AVQA: A Dataset for Audio-Visual Question Answering on Videos. In ACM Int. Conf. Multimedia, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 691 + ], + "type": "text", + "content": "[73] Zhao Yang, Jiaqi Wang, Yansong Tang, Kai Chen, Hengshuang Zhao, and Philip HS Torr. LAVT: Language-Aware Vision Transformer for Referring Image Segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., 2022. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[74] Kaining Ying, Zhenhua Wang, Cong Bai, and Pengfei Zhou. Isda: Position-aware instance segmentation with deformable" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "attention. In IEEE Int. Conf. Acoust. Speech Signal Process., 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "type": "text", + "content": "[75] Kaining Ying, Qing Zhong, Weian Mao, Zhenhua Wang, Hao Chen, Lin Yuanbo Wu, Yifan Liu, Chengxiang Fan, Yunzhi Zhuge, and Chunhua Shen. CTVIS: Consistent Training for Online Video Instance Segmentation. In Int. Conf. Comput. Vis., 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 150, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 237 + ], + "type": "text", + "content": "[76] Kaining Ying, Fanqing Meng, Jin Wang, Zhiqian Li, Han Lin, Yue Yang, Hao Zhang, Wenbo Zhang, Yuqi Lin, Shuo Liu, Jiayi Lei, Quanfeng Lu, Runjian Chen, Peng Xu, Renrui Zhang, Haozhe Zhang, Peng Gao, Yali Wang, Yu Qiao, Ping Luo, Kaipeng Zhang, and Wenqi Shao. MMT-Bench: A Comprehensive Multimodal Benchmark for Evaluating Large Vision-Language Models Towards Multitask AGI. In Int. Conf. Mach. Learn., 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 238, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 238, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 238, + 553, + 271 + ], + "type": "text", + "content": "[77] Kaining Ying, Hengrui Hu, and Henghui Ding. MOVE: Motion-guided few-shot video object segmentation. In Int. Conf. Comput. Vis., 2025. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 271, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 553, + 304 + ], + "type": "text", + "content": "[78] Licheng Yu, Patrick Poirson, Shan Yang, Alexander C Berg, and Tamara L Berg. Modeling Context in Referring Expressions. In Eur. Conf. Comput. Vis., 2016. 6, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 304, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 349 + ], + "type": "text", + "content": "[79] Fangao Zeng, Bin Dong, Yuang Zhang, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. MOTR: End-to-End Multiple-Object Tracking with Transformer. In Eur. Conf. Comput. Vis., 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 349, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 392 + ], + "type": "text", + "content": "[80] Hang Zhang, Xin Li, and Lidong Bing. Video-LLaMA: An Instruction-tuned Audio-Visual Language Model for Video Understanding. In Proc. of the Conf. on Empirical Methods in Nat. Lang. Process., 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 393, + 553, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 446 + ], + "type": "text", + "content": "[81] Renrui Zhang, Jiaming Han, Chris Liu, Aojun Zhou, Pan Lu, Yu Qiao, Hongsheng Li, and Peng Gao. LLaMA-Adapter: Efficient Fine-tuning of Large Language Models with Zero-initialized Attention. In Int. Conf. Learn. Represent., 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 448, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 448, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 448, + 553, + 502 + ], + "type": "text", + "content": "[82] Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Yu Liu, Kai Chen, and Ping Luo. GPT4RoI: Instruction Tuning Large Language Model on Region-of-Interest. arXiv preprint arXiv:2307.03601, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 502, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 553, + 547 + ], + "type": "text", + "content": "[83] Tao Zhang, Xingye Tian, Yu Wu, Shunping Ji, Xuebo Wang, Yuan Zhang, and Pengfei Wan. DVIS: Decoupled Video Instance Segmentation Framework. In Int. Conf. Comput. Vis., 2023. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 547, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 547, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 547, + 553, + 591 + ], + "type": "text", + "content": "[84] Rongkun Zheng, Lu Qi, Xi Chen, Yi Wang, Kun Wang, Yu Qiao, and Hengshuang Zhao. ViLLa: Video Reasoning Segmentation with Large Language Model. arXiv preprint arXiv:2407.14500, 2024. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 591, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 635 + ], + "type": "text", + "content": "[85] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene Parsing through ADE20K Dataset. In IEEE Conf. Comput. Vis. Pattern Recog., 2017. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 680 + ], + "type": "text", + "content": "[86] Jinxing Zhou, Jianyuan Wang, Jiayi Zhang, Weixuan Sun, Jing Zhang, Stan Birchfield, Dan Guo, Lingpeng Kong, Meng Wang, and Yiran Zhong. Audio-Visual Segmentation. In Eur. Conf. Comput. Vis., 2022. 2, 3, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[87] Jiawen Zhu, Zhi-Qi Cheng, Jun-Yan He, Chenyang Li, Bin Luo, Huchuan Lu, Yifeng Geng, and Xuansong Xie. Tracking with Human-Intent Reasoning. arXiv, 2023. 8" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22585" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/2355f2bc-16e1-4dc5-88a7-53284ff42cd2_content_list.json b/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/2355f2bc-16e1-4dc5-88a7-53284ff42cd2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a000f06d1665e82f6fdcff23279172f76ffade90 --- /dev/null +++ b/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/2355f2bc-16e1-4dc5-88a7-53284ff42cd2_content_list.json @@ -0,0 +1,1908 @@ +[ + { + "type": "text", + "text": "Towards Open-World Generation of Stereo Images and Unsupervised Matching", + "text_level": 1, + "bbox": [ + 96, + 130, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feng Qiao Zhexiao Xiong Eric Xing Nathan Jacobs Washington University in St. Louis", + "bbox": [ + 263, + 180, + 733, + 217 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e85e509e88f5dec64869b1deea24ad1c82ada4f7de0182661ef05b5422c41462.jpg", + "image_caption": [ + "Figure 1. Comparison with diffusion-based methods for stereo image generation on the COCO dataset." + ], + "image_footnote": [], + "bbox": [ + 91, + 229, + 903, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 430, + 326, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stereo images are fundamental to numerous applications, including extended reality (XR) devices, autonomous driving, and robotics. Unfortunately, acquiring high-quality stereo images remains challenging due to the precise calibration requirements of dual-camera setups and the complexity of obtaining accurate, dense disparity maps. Existing stereo image generation methods typically focus on either visual quality for viewing or geometric accuracy for matching, but not both. We introduce GenStereo, a diffusion-based approach, to bridge this gap. The method includes two primary innovations (1) conditioning the diffusion process on a disparity-aware coordinate embedding and a warped input image, allowing for more precise stereo alignment than previous methods, and (2) an adaptive fusion mechanism that intelligently combines the diffusion-generated image with a warped image, improving both realism and disparity consistency. Through extensive training on 11 diverse stereo datasets, GenStereo demonstrates strong generalization ability. GenStereo achieves state-of-the-art performance in both stereo image generation and unsupervised stereo matching tasks. Project page is available at https://qjizhi.github.io/genstereo.", + "bbox": [ + 86, + 464, + 483, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 828, + 220, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The demand for stereo images continues growing with XR devices, autonomous driving, and robotics. However, acquiring high-quality stereo images remains chal", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lenging due to complex camera calibration and environmental constraints, limiting the development of robust stereo matching models that generalize well across diverse scenarios. Real-world datasets often provide only sparse disparity annotations [28] or lack accurate ground truth altogether [9]. Although some datasets offer dense, accurate disparity maps [2, 33], they are typically limited to specific scenarios like indoor scenes. Moreover, capturing real-world stereo data requires complex sensor setups with precise calibration, often constraining baseline distances and scene diversity. Synthetic datasets, while providing precise disparity maps, suffer from domain gaps compared to real-world scenarios [5, 56, 57].", + "bbox": [ + 511, + 431, + 906, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in monocular depth estimation (MDE) models [3, 34, 64, 65] have provided increasingly accurate dense disparity maps, playing a crucial role in image generation from single views. Meanwhile, Text-to-Image (T2I) diffusion models, such as Stable Diffusion [35], have made remarkable strides in generating diverse and high-quality images from user-provided text prompts. However, these models often struggle to maintain spatial consistency when altering the viewpoint of the generated images. While there has been some research on 3D novel view generation [38, 53], there remains a gap in utilizing image generation models to directly produce stereo images. StereoDiffusion [55] is the first approach to leverage diffusion-based generation for stereo image generation, but it struggles with pixel-level accuracy, as it applies disparity shifts in the latent space and fills occluded regions with blurry pixels that lack meaningful semantic content. On the other hand, Stable Diffusion Inpainting (SD", + "bbox": [ + 511, + 628, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26579", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inpainting) [35], though not designed for stereo image generation, faces similar challenges by incorrectly filling occluded areas with semantically inappropriate pixels. Fig. 1 shows that previous diffusion-based methods and our methods tested on the COCO dataset [23]. The scale factor $\\gamma$ is set to 0.15, meaning that the maximum disparity is $15\\%$ of the image width.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present GenStereo, a novel framework that addresses both visual quality and geometric accuracy in stereo image generation. Existing approaches rely on either warping-based methods [60], which provide geometric accuracy but struggle with semantic consistency in occluded regions, or diffusion-based approaches [55], which maintain better semantic coherence but lack precise geometric control. GenStereo bridges this gap through a carefully designed two-stream architecture inspired by recent advances in human animation [15] and view synthesis [38], achieving both geometric precision and semantic consistency.", + "bbox": [ + 89, + 198, + 480, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework employs a multi-level constraint system for geometric precision and semantic coherence, starting with disparity-aware coordinate embedding that provides implicit geometric guidance. This is followed by a cross-view attention mechanism that enables semantic feature alignment between views, ensuring consistency in challenging areas like occlusions and complex textures. The framework leverages a dual-space supervision strategy that operates in both latent and pixel spaces, with an adaptive fusion mechanism that further optimizes pixel-level accuracy. To ensure robust generalization across diverse real-world scenarios, we train our model on a diverse dataset combining 11 stereo datasets with varying scenes and baselines.", + "bbox": [ + 89, + 366, + 480, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we demonstrate how our high-quality stereo generation framework significantly improves unsupervised stereo matching learning. Previous approaches to unsupervised learning have been limited by either simplified warping and random background filling [60] or constraints to small-scale static scenes [50]. In contrast, our method enables large-scale training with diverse, photorealistic stereo images that maintain both geometric accuracy and semantic consistency. This advancement represents a significant step toward bridging the gap between supervised and unsupervised stereo matching approaches. Our contributions can be summarized as follows:", + "bbox": [ + 89, + 563, + 480, + 743 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose GenStereo, the first unified framework for open-world stereo image generation that addresses visual quality and geometric accuracy, enabling both practical applications and unsupervised stereo matching.", + "- We introduce a comprehensive multi-level constraint system that combines: (1) disparity-aware coordinate embedding with warped image conditioning for geometric guidance, (2) cross-view attention mechanism for semantic feature alignment, and (3) dual-space supervision with adaptive fusion for pixel-accuracy generation." + ], + "bbox": [ + 89, + 750, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Extensive experimental validation demonstrating state-of-the-art performance in both stereo image generation and unsupervised stereo matching.", + "bbox": [ + 513, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 151, + 653, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Conditional Diffusion Models", + "text_level": 1, + "bbox": [ + 513, + 176, + 774, + 193 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models began with DDPM [14], advancing from slow, probabilistic generation to faster sampling with DDIM [41]. Stable Diffusion [35], a pioneering T2I model, further improved efficiency by performing diffusion in a compact latent space, combining it with CLIP-based [32] text conditioning for versatile T2I generation. However, Stable Diffusion lacks fine-grained control over specific visual attributes, relying mainly on text prompts without structured guidance. To address these limitations, conditional diffusion models have emerged. ControlNet [66] is a key advancement, introducing conditioning mechanisms for structural inputs (such as edges, depth maps, or poses) that maintain flexibility while enabling precise, user-driven synthesis. However, ControlNet still faces challenges in achieving pixel-level accuracy in complex images, such as stereo image generation.", + "bbox": [ + 511, + 200, + 903, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent research [15, 19, 38, 39, 63] has explored the self-attention properties within T2I models. Text2VideoZero [19] and MVDream [39] generate consistent visuals across video frames or 3D multi-views by sharing self-attention, whileAnimate-Anyone [15] and MagicAnimate [63] apply a similar approach to produce human dance videos through fine-tuned T2I models. GenWarp [38] employs augmenting self-attention with cross-view attention between the reference and target views to generate novel images. This two-stream architecture augments the features of the denoising net. Inspired by the adaptability and control of these self-attention-based architectures, our method builds on their strengths. However, their direct application to stereo image generation faces fundamental challenges. First, they typically don't incorporate disparity information, which is crucial for stereoscopic consistency. Second, achieving the pixel-level accuracy required for comfortable stereo viewing demands specialized architectural considerations beyond existing frameworks.", + "bbox": [ + 511, + 441, + 903, + 729 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Stereo Image Generation and Inpainting", + "text_level": 1, + "bbox": [ + 513, + 741, + 861, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional view synthesis approaches, including geometry-based reconstruction [12, 13, 20] and recent NeRF-based methods [50], excel at rendering novel views but are constrained by their requirement for multiple input images of static scenes. While 3D Photography techniques [12, 40] attempt to overcome this limitation through depth-based mesh projection, they often struggle with complex occlusion handling and require sophisticated post-processing pipelines.", + "bbox": [ + 511, + 763, + 903, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The evolution of stereo generation methods has pro", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26580", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gressed from simple warping-based approaches to more sophisticated diffusion-based solutions. MfS [60] introduced a basic framework that attempted to handle occlusions by sampling random background patches from the dataset, but this led to semantic inconsistencies and visible artifacts at occlusion areas. While SD-Inpainting [35] introduced more coherent semantic content through its prior learning, it failed to maintain local consistency between inpainted regions and their surrounding context, often producing visible discontinuities in texture and structure. Mono2Stereo [59] made progress by fine-tuning SD-Inpainting specifically for stereo generation, enabling unsupervised stereo matching, though visible artifacts persist in in-painted regions. Stereodiffusion [55] marked a significant shift by introducing end-to-end diffusion-based stereo image generation, but its latent-space warping approach without explicit geometric constraints led to compromised pixel-level accuracy. Despite these advances, the challenge of generating photorealistic stereo images that maintain both visual quality and geometric fidelity remains largely unsolved, particularly for diverse real-world scenarios.", + "bbox": [ + 89, + 90, + 483, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Unsupervised Stereo Matching", + "text_level": 1, + "bbox": [ + 89, + 417, + 362, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional unsupervised stereo matching approaches primarily rely on photometric consistency. Methods like [11, 47, 48, 67] leverage photometric losses across stereo images, while others [8, 21, 58] extend this to temporal sequences. A parallel line of research explores proxy supervision strategies, either through carefully designed algorithmic supervisors [31, 45, 46] or knowledge distillation from pre-trained networks [1]. Recent domain adaptation approaches [25, 42, 61] attempt to bridge the gap between synthetic and real-world domains or facilitate cross-dataset adaptation. However, these methods often exhibit limited generalization capability beyond their target domains [1], particularly struggling with diverse real-world scenarios.", + "bbox": [ + 89, + 439, + 483, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More recent approaches have explored novel directions for stereo generation and matching. MfS [60] pioneered unsupervised stereo matching using MDE as guidance, but it relies on random background sampling for occlusion filling, leading to semantic inconsistencies. While NeRF-Stereo [50] achieves high-quality results through neural radiance fields, it requires multiple views of static scenes, limiting its practical applications. Mono2Stereo [59] leverages Stable Diffusion for inpainting occluded regions, representing a significant advance in generative stereo synthesis, though it still exhibits artifacts in occluded areas where geometric and semantic consistency is crucial.", + "bbox": [ + 89, + 636, + 483, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 829, + 191, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a left view image $I_{l}$ and its corresponding disparity map $\\hat{D}_{l}$ predicted by an MDE model, our goal is to generate a high-quality right view image $\\hat{I}_r$ that maintains both vi", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sual quality and geometric consistency. As shown in Fig. 2, our GenStereo framework processes the left view as the reference image and synthesizes the right view as the target image. The generated triplet $\\langle I_l,\\hat{D}_l,\\hat{I}_r\\rangle$ not only serves as a stereo pair for visualization but also provides training data for unsupervised stereo matching.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Disparity-Aware Coordinate Embedding", + "text_level": 1, + "bbox": [ + 511, + 189, + 862, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional inpainting-based stereo image generation methods often suffer from visible boundaries between warped and inpainted regions. To address this limitation, we propose a disparity-aware coordinate embedding scheme inspired by recent advances in coordinate-based generation [29, 38]. Our approach utilizes dual coordinate embeddings: a canonical embedding for the left view and its warped counterpart for the right view.", + "bbox": [ + 511, + 210, + 905, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we first construct a canonical 2D coordinate map $X \\in \\mathbb{R}^{h \\times w \\times 2}$ with values normalized to $[-1,1]$ . This map is transformed into Fourier features [43] through a positional encoding function $\\phi$ :", + "bbox": [ + 511, + 332, + 905, + 392 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC _ {l} = \\phi (X) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 660, + 401, + 903, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The resulting Fourier feature map $C_l$ serves as the coordinate embedding for the left view $I_l$ . We then generate the right view embedding by warping $C_l$ according to the disparity map:", + "bbox": [ + 511, + 425, + 905, + 484 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC _ {r} = \\operatorname {w a r p} \\left(C _ {l}, D _ {l}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 486, + 903, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $D_{l}$ represents ground-truth disparity during training and predicted disparity $\\hat{D}_l$ during inference. These coordinate embeddings $(C_l$ and $C_r)$ are integrated into their respective view features $(F_{l}$ and $F_{r})$ through convolutional layers, establishing strong geometric correspondence while maintaining visual consistency between views.", + "bbox": [ + 511, + 506, + 905, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach differs from GenWarp [38] by utilizing disparity maps instead of camera matrices for warping, enabling more precise pixel-level control. We further enhance geometric consistency by incorporating the warped image $I_{\\mathrm{warp}}$ as additional conditioning for the denoising U-Net.", + "bbox": [ + 511, + 597, + 905, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Cross-View Feature Enhancement", + "text_level": 1, + "bbox": [ + 511, + 680, + 812, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To facilitate effective information exchange between views, we adopt a two-parallel U-Nets framework pretrained from Stable Diffusion, inspired by GenWarp [38]. However, our work differs in its novel conditioning strategy: the reference U-Net takes concatenated $(I_l, C_l)$ as conditions to process the left view, while the denoising U-Net takes concatenated $(I_{\\mathrm{warp}}, C_r)$ as conditions when synthesizing the right view $\\hat{I}_r$ . This design ensures that each U-Net receives both image content and corresponding coordinate information for better feature extraction. As our model is conditioned on a reference image and its disparity map, we replace text condition of SD with image embedding of the left image through a CLIP image encoder in both U-Nets.", + "bbox": [ + 511, + 702, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26581", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5d1ad429e007e8b74d4ad5dc5ef20bfa88ab62ebbe8e293b87f95f9403d1bf40.jpg", + "image_caption": [ + "Figure 2. Overview of the GenStereo framework." + ], + "image_footnote": [], + "bbox": [ + 98, + 89, + 911, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To incorporate left-view information during right-view generation, we concatenate the reference features with the target features in the attention mechanism. Specifically, we compute cross-view attention as:", + "bbox": [ + 89, + 344, + 483, + 404 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq = F _ {r}, \\quad k = \\left[ F _ {l}, F _ {r} \\right], \\quad v = \\left[ F _ {l}, F _ {r} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 411, + 480, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{l}$ is derived from the reference U-Net conditioned on $(I_l,C_l)$ , while $F_{r}$ comes from the denoising U-Net conditioned on $(I_{\\mathrm{warp}},C_r)$ . This dual-stream attention mechanism allows the model to adaptively balance between semantic consistency from the reference view and geometric accuracy from the warped view, with each stream guided by both image content and coordinate information.", + "bbox": [ + 89, + 433, + 483, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training Strategy", + "text_level": 1, + "bbox": [ + 89, + 547, + 264, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mixed training [22, 34, 64] has been proven to be an effective training strategy for domain generalization in both monocular and stereo matching depth estimation. We follow the datasets summarized in [51] and utilize the most widely available public synthetic datasets. Based on our observations, real-world datasets can negatively impact performance, even when there are only minor calibration errors or slight differences in imaging between the two cameras. Tab. 1 shows the comprehensive list of datasets utilized. Datasets containing abstract images are not included, because our model aims to have a good generalization capability in the real world.", + "bbox": [ + 89, + 569, + 483, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address dataset imbalance, we employ a resampling strategy. Specifically, we sample the smaller datasets multiple times until the number reaches $10\\%$ of the largest dataset. As the Stable Diffusion model only accepts images of size $512 \\times 512$ for SD v1.5 and $768 \\times 768$ for SD v2.1, we apply a random square crop and resize.", + "bbox": [ + 89, + 750, + 483, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Pixel Space Alignment", + "text_level": 1, + "bbox": [ + 89, + 848, + 302, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an RGB image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ , Latent Diffusion Models (LDMs) first encode it into a latent representation", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/8e07f9a70d5f917486b75dabb1c974c99b00581d5a0cd628d8e20cb1bda4beb1.jpg", + "table_caption": [ + "Table 1. Our training data sources. All datasets are synthetic with the size of 684K." + ], + "table_footnote": [], + "table_body": "
DatasetIndoorOutdoor#ImagesYear
TartanAir [57]✓✓306K2020
Dynamic Replica [18]✓145K2023
IRS [56]✓103K2021
Falling Things [52]✓✓61K2018
VKITTI2 [5]✓21K2020
InfinigenSV [16]✓17K2024
SimStereo [17]✓14K2022
UnrealStereo4K [49]✓✓8K2021
Spring [27]✓5K2023
PLT-D3 [44]✓3K2024
Sintel [4]✓1K2012
", + "bbox": [ + 535, + 381, + 880, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$z = \\mathcal{E}(I)$ , where $z\\in \\mathbb{R}^{h\\times w\\times c}$ . While this latent space operation significantly reduces computational costs, it may compromise pixel-level accuracy during image generation.", + "bbox": [ + 511, + 541, + 905, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The standard LDM objective operates in latent space:", + "bbox": [ + 532, + 587, + 885, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {l a t e n t}} := \\mathbb {E} _ {z, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 609, + 903, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{t}$ is the noisy latent at timestep $t$ , $\\epsilon$ is the noise added to create $z_{t}$ , and $\\epsilon_{\\theta}(z_{t}, t)$ is the network's prediction of this noise. The model learns to denoise by predicting the noise that was added.", + "bbox": [ + 511, + 636, + 905, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To maintain pixel-level accuracy, we introduce an additional pixel-space loss by decoding both the predicted and target latent variables:", + "bbox": [ + 511, + 696, + 905, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {p i x e l}} = \\mathbb {E} _ {z, \\epsilon , t} \\left[ \\| \\mathcal {D} \\left(z _ {\\text {p r e d}}\\right) - \\mathcal {D} \\left(z _ {\\text {t a r g e t}}\\right) \\| _ {2} ^ {2} \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 750, + 903, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final objective combines both spaces:", + "bbox": [ + 511, + 776, + 792, + 791 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\text {l a t e n t}} + \\alpha L _ {\\text {p i x e l}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 800, + 903, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha = 1$ balances latent and pixel-space supervision.", + "bbox": [ + 511, + 825, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Adaptive Fusion Module", + "text_level": 1, + "bbox": [ + 511, + 848, + 741, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve seamless integration between generated and warped content, we propose an Adaptive Fusion Module", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26582", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that learns to combine $I_{gen}$ and $I_{warp}$ based on local context and confidence. The module predicts spatially-varying fusion weights through a lightweight network:", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW = \\sigma \\left(f _ {\\theta} \\left(\\operatorname {c o n c a t} \\left(I _ {g e n}, I _ {w a r p}, M\\right)\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 148, + 482, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{\\theta}$ is a $3 \\times 3$ convolutional layer and $\\sigma$ is the sigmoid activation, ensuring weights in [0, 1]. The final right view is computed as:", + "bbox": [ + 89, + 176, + 483, + 222 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {r} = M \\odot W \\odot I _ {\\text {w a r p}} + (1 - M \\odot W) \\odot I _ {\\text {g e n}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 231, + 482, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise multiplication. This formulation adaptively favors warped content in high-confidence regions ( $M \\approx 1$ ) while relying on generated content in occluded or uncertain areas. The learned weights $W$ enable smooth transitions between warped and generated regions, ensuring both geometric accuracy and visual consistency.", + "bbox": [ + 89, + 261, + 483, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Random Disparity Dropout", + "text_level": 1, + "bbox": [ + 89, + 361, + 339, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To simulate the sparsity of real-world disparity maps like KITTI's LiDAR-derived ground truth, we randomly apply disparity dropout to $10\\%$ of training samples. For each selected sample, we first generate a dropout ratio:", + "bbox": [ + 89, + 383, + 483, + 444 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr \\sim \\operatorname {U n i f o r m} (0, 1) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 455, + 482, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using this ratio, we create a binary mask where each pixel has a probability $r$ of being dropped:", + "bbox": [ + 89, + 483, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM _ {\\text {r a n d}} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {o t h e r w i s e} \\\\ 0 & \\text {w i t h p r o b a b i l i t y} r \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 523, + 482, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final mask combines this random dropout with the warping mask:", + "bbox": [ + 89, + 574, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM = M _ {\\text {w a r p}} \\vee M _ {\\text {r a n d}} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 616, + 482, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\vee$ denotes element-wise logical OR. This strategy encourages the model to handle sparse disparity inputs, improving its robustness and generalization to real-world scenarios where dense disparity maps may not be available.", + "bbox": [ + 89, + 643, + 483, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 718, + 223, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Stereo Image Generation", + "text_level": 1, + "bbox": [ + 89, + 742, + 318, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Setup. We fine-tune the pretrained Stable Diffusion (SD) UNet for 3 epochs across all experiments. For evaluation, we utilize two widely adopted stereo vision benchmarks: Middlebury 2014 [36] and KITTI 2015 [28], neither of which are included in our training set. Middlebury 2014 comprises 23 high-resolution indoor stereo pairs captured with wide-baseline stereo cameras under varying illumination conditions, providing challenging scenarios for stereo generation. KITTI 2015 contains 200 outdoor stereo", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pairs of street scenes with LiDAR-based sparse disparity ground truth, offering diverse real-world testing scenarios. Following StereoDiffusion [55], we preprocess Middlebury images to $512 \\times 512$ resolution, while KITTI images are center-cropped before resizing to $512 \\times 512$ to maintain the most informative regions.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct comprehensive comparisons against both traditional and learning-based baselines. Traditional approaches include naive solutions such as leave blank and image stretching. For learning-based comparisons, we evaluate against state-of-the-art diffusion models adapted for stereo generation: Repaint [26] and SD-Inpainting [35], as well as StereoDiffusion, which specifically targets stereo image generation. To quantitatively assess the quality of generated right-view images $(\\hat{I_r})$ against ground truth $(I_{r})$ we employ three complementary metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) with SqueezeNet backbone.", + "bbox": [ + 511, + 183, + 906, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative Results. Tab. 2 presents the generation performance on Middlebury 2014 and KITTI 2015. The sparsity of ground truth disparities in KITTI limits evaluation accuracy, making pseudo-label disparities more effective for assessing stereo consistency. This sparsity particularly affects inpainting-based methods, as they struggle to infer accurate right-view images in areas with missing depth information, especially for dynamic objects and distant regions. In contrast, Middlebury 2014, which provides dense ground truth disparities, demonstrates a smaller performance gap between using ground truth and using pseudo disparities, underscoring the advantage of high-quality disparity annotations for generation. We obtain pseudo disparities from CREStereo [22] pretrained model, where evaluation datasets are not used for training.", + "bbox": [ + 511, + 380, + 905, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Results. We present qualitative comparisons between our method and several baselines: leave blank, StereoDiffusion, and SD-Inpainting, with ground truth as reference. As shown in Fig. 3, our method demonstrates superior performance on the Middlebury 2014 dataset. While SD-Inpainting struggles with occlusion handling, often generating semantically inconsistent content in occluded regions, and StereoDiffusion exhibits loss of fine details due to latent-space warping, our approach successfully maintains both geometric accuracy and visual fidelity. Particularly noteworthy is our method's ability to generate coherent right-view images while preserving pixel-level correspondence with the left images. Additional qualitative results on the KITTI 2015 dataset and more visualizations on other datasets can be found in the supplementary materials.", + "bbox": [ + 511, + 608, + 905, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Unsupervised Learning", + "text_level": 1, + "bbox": [ + 511, + 847, + 730, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Experimental Setup. We conduct two comprehensive experiments to validate the effectiveness of our proposed gen", + "bbox": [ + 511, + 869, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26583", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/85a69e565832226dede9b9a4a70040b626e295054226f15bdade49cfbe8a2099.jpg", + "table_caption": [ + "Table 2. Quantitative results of image generation for Middlebury 2014 and KITTI 2015 datasets. The top three results for each metric are highlighted with a first, second, and third background, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsSD versionMiddlebury 2014KITTI 2015
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Leave blank-11.328+2.4830.315+0.2300.450+0.09712.980+4.9190.374+0.2860.313+0.222
-3.489-0.154-0.089-3.831-0.251-0.109
Stretch-14.842+2.7140.432+0.1900.285+0.11214.757+5.3750.429+0.2870.212+0.145
-2.753+0.265-0.089-4.694-0.271-0.100
3D Photography [40]-14.190+2.4640.427+0.2380.275+0.07314.540+7.2560.398+0.2700.210+0.099
-2.798-0.175-0.065-4.023-0.323-0.073
RePaint [26]1.515.102+2.9090.462+0.2680.311+0.08815.056+5.3660.462+0.2680.251+0.128
-2.802-0.253-0.079-4.897-0.285-0.095
SD-Inpainting [35]1.515.740+3.5110.412+0.2360.311+0.0989.792+2.8090.230+0.2360.652+0.144
-3.351-0.188-0.164-4.070-0.229-0.139
StereoDiffusion [55]1.515.456+2.6690.468+0.2520.231+0.09615.679+5.8880.481+0.2450.205+0.135
-3.313-0.205-0.088-5.487-0.310-0.099
Ours + GT1.523.699+5.4560.866+0.0840.064+0.05320.439+7.5180.746+0.1550.108+0.117
-6.033-0.141-0.034-8.541-0.274-0.048
Ours + Pseudo1.523.835+6.8420.868+0.0960.062+0.06322.749+6.7220.813+0.1100.096+0.068
-6.741-0.157-0.038-8.102-0.152-0.104
Ours + GT2.124.819+6.5190.906+0.0570.061+0.03219.836+8.4830.765+0.2450.135+0.059
-4.252-0.132-0.032-7.090-0.092-0.068
Ours + Pseudo2.125.142+6.9700.911+0.0660.060+0.06123.488+6.4440.849+0.1300.109+0.050
", + "bbox": [ + 138, + 128, + 856, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4626db6c11940104f518bea321646586a4836eb60a01189874c9d0b4cde068f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 305, + 225, + 409 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e35f46a7566228bd105a1072af012120eedf8ae32516f84ee8af4b56749ad612.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 411, + 225, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4c99701b8f14edda3eac7936d93cdd5d18465709f8f6813a3adb971b07b3eaef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 515, + 225, + 619 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6673940f90317b9467a02dd793cbc4e6f71131540b557b91668350a0610e7722.jpg", + "image_caption": [ + "Reference Image" + ], + "image_footnote": [], + "bbox": [ + 91, + 621, + 225, + 723 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d9779e095e6a380a0cebc24fa84dc627f9ae4d45cdcddccaa5ff484a174fa8d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 306, + 359, + 409 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/39c2eb229569dae5fdcf288a45c6e07c5bccc5c0f5aef1883979f7d94284c45e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 411, + 359, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08769934ae2ec1826398d26c679943933846da98f088a22d287267d2bc587082.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 515, + 359, + 619 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf51d8a69ae0c53540d400439a7fb044a4596a9f502b8310ffb55b35d1a3ab4c.jpg", + "image_caption": [ + "Leave Blank" + ], + "image_footnote": [], + "bbox": [ + 228, + 621, + 361, + 723 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/31ea98a2d156060801743b2a9e76e9d10b2f0d83f6f0dd334a2005b83caaf770.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 306, + 496, + 409 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d932d1ec3188b94be7d44086205892028959feacc77f07b6d2bda12bb03dcb48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 411, + 496, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5967345400804e8acbaf2021f80606f50817fe297f19e9d66c26e7d56e8573f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 515, + 496, + 619 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aebdb82329d735a030550ca5d57340b6d33e7d453ea266b2cfa4594b5ac12cbb.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 364, + 621, + 496, + 723 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4e2a4b4ad75882de53bf77bf3fd7cd3981195d19deafe8814bd70d992169b8d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 306, + 633, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ed3776bc4acc972c4b81d76d8746827ceb475d123e2c230b65d34e515d7fd72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 411, + 633, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a85ece81b50054519d62a78672352903f300d95e4b039a212add1e0006fc325.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 515, + 633, + 619 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c536a43c7b962480e27d95a57817f3cc142c6395346474e1405d1bbdee667c5e.jpg", + "image_caption": [ + "SD-Inpainting" + ], + "image_footnote": [], + "bbox": [ + 500, + 621, + 633, + 723 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7d1be42928f302dd7cedc9810ad721d98b3089a8fafb70ca14ff97a465b2b82e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 306, + 769, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/307bfa7e1b8efd2b97b33703d60b265416ebe370137f38ed87e6934dcd3f4b0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 411, + 769, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/713a917839ca3b21711c64fc409793faa7a8a0c1fb3b06b9210c5fb6a0e658eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 515, + 769, + 619 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9b41ab7322bd3624ae38bf3563c21dc56a6f7ea10d01e24da83af9b4f6825f87.jpg", + "image_caption": [ + "StereoDiffusion", + "Figure 3. Qualitative comparison on Middlebury 2014 with ground truth disparity maps." + ], + "image_footnote": [], + "bbox": [ + 635, + 621, + 769, + 723 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/61135b89c4a0ec761b3380772bff5b4f650c3d033c8242ee86d07e3280a000b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 306, + 903, + 409 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/852dcbbd60b718c68912436a0633ff3d9620027380d535ce9e0ac5fb1c603cb2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 411, + 903, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f4592ed5f91857f3dce4d8e70f71b8f0fa616601fa0e6ede490b273eddc97e25.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 515, + 903, + 619 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76ce018505aaa8e4052d8d7480232010c24be52ebe3da5db2fe54d1510214c21.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 771, + 621, + 903, + 723 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "eration method for unsupervised stereo matching.", + "bbox": [ + 89, + 794, + 418, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the first experiment, we fine-tune a stereo matching model pretrained on the SceneFlow dataset using the KITTI 2012 and KITTI 2015 benchmarks. We split the training and evaluation sets into 160/34 images for KITTI 2012 and 160/40 images for KITTI 2015, following [60].", + "bbox": [ + 89, + 810, + 482, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The second experiment evaluates various unsupervised", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "stereo matching approaches by generating stereo images. Specifically, KITTI 2012 [10] has 194 stereo images, KITTI 2015 has 200 stereo images, ETH3D [37] has 27 stereo pairs, and the Middlebury v3 training set [36] has 15 stereo pairs at Quarter and Half resolutions (Midd-Q, Midd-H). This diverse dataset selection allows us to comprehensively assess the generalization capability of our method across", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26584", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2dbff5b2205090804c971477d9004d877f3cfca3d5ddb5782d727cd4eb80c9fc.jpg", + "table_caption": [ + "Table 3. Comparison with other diffusion-based stereo image generation methods for unsupervised stereo matching. The first group of methods is trained on PSMNet while the second group is trained on RAFT-Stereo. SD v1.5 is used for the experiments." + ], + "table_footnote": [], + "table_body": "
MethodsKITTI 2012KITTI 2015
D1-all ↓EPE ↓D1-all ↓EPE ↓
Leave blank25.2193.54224.5682.937
SD-Inpainting [35]3.9070.8944.4901.059
StereoDiffusion [55]15.2132.2205.6511.154
Ours3.8020.8153.9330.991
Leave blank4.1840.9524.0740.957
SD-Inpainting [35]4.0160.9454.1280.972
StereoDiffusion [55]12.7561.8116.3601.174
Ours3.5370.9303.6530.943
", + "bbox": [ + 114, + 155, + 457, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "different domains and resolutions.", + "bbox": [ + 89, + 316, + 318, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the performance of various unsupervised stereo matching approaches using the D1-all metric, which measures the percentage of incorrect disparity predictions (error $>3\\mathrm{px}$ or $5\\%$ of the true disparity), the End-PointError (EPE), and the $>n\\mathrm{px}$ metric, which denotes the percentage of pixels with an EPE exceeding $n$ pixels. To demonstrate the architecture-agnostic nature of our approach, we employ three distinct stereo models: PSM-Net [6], a cost-volume-based method, RAFT-Stereo [24], an iterative refinement model, and IGEV++ [62], one of the most recent methods. For disparity estimation, we utilize the Depth Anything Model v2 (DAMv2) [65], from which we obtain the disparity map. The disparity values are normalized to the range $[0, 1]$ and subsequently scaled by a factor $\\gamma$ to enable flexibility generation.", + "bbox": [ + 88, + 334, + 485, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fine-tuning on KITTI. To ensure fair and consistent comparisons across all generation methods, we construct the training datasets using a standardized procedure. Specifically, we sample the scale factor $\\gamma$ from the set $\\{0.05, 0.1, 0.15, 0.2, 0.25\\}$ . For each selected scale, we employ various generation methods to create the right images based on 160 left images and their corresponding scaled disparity maps for both KITTI 2012 and KITTI 2015. This procedure results in a total of 800 stereo pairs.", + "bbox": [ + 89, + 563, + 483, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 3, higher generation quality leads to improved stereo matching performance. Among the compared methods, SD-Inpainting achieves the closest performance to ours due to its pixel-level warping operation. However, it falls short in providing semantic consistency in occluded areas, which limits its overall effectiveness.", + "bbox": [ + 89, + 700, + 483, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unsupervised Generalization. Following the methodology of constructing the Mono for Stereo (MfS) dataset in [60], we construct our training dataset, MfS-GenStereo, from the same datasets, including ADE20K [68], Mapillary Vistas [30], DIODE [54], Depth in the Wild [7], and COCO 2017 [23]. During the generation process, we randomly select a scale factor in the range $[0, 0.35]$ and set the maxi", + "bbox": [ + 89, + 794, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mum disparity to 256 for PSMNet. As shown in Tab. 4, our method enables unsupervised stereo matching using only left images, achieving superior performance.", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our approach with the original MfS [60] and its enhanced variant using the Depth Anything Model v2 (DAMv2), which generates the right view images by filling the occluded areas with random background patches from the dataset. The experiments indicate that, despite the better depth prior, the synthesis of right-view images remains a critical bottleneck in this methodology. Additionally, we compare with Mono2Stereo, which fine-tunes SD-Inpainting for right-view image generation, and NeRF-Stereo, which generates stereo images by reconstructing the static scenes using Neural Radiance Fields (NeRF), limiting the diversity of its datasets. Our experimental results indicate that our proposed method consistently surpasses these existing approaches in generating high-quality stereo pairs with pixel-level accuracy, demonstrating robust generalization capabilities and significantly enhancing stereo matching performance across diverse datasets.", + "bbox": [ + 511, + 137, + 908, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study.", + "text_level": 1, + "bbox": [ + 511, + 407, + 669, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Random disparity drop. Since the ground-truth disparity maps in real-world datasets (e.g., KITTI) are often derived from LiDAR, resulting in sparse and incomplete point clouds, our random disparity dropout strategy mitigates the model's reliance on dense disparity supervision. This enhancement allows the model to generate plausible right images even when only sparse disparity maps are available. This strategy significantly improves performance on datasets with sparse ground-truth annotations, demonstrating better generalization to real-world scenarios. As shown in Tab. 5, comparing rows (1) and (2) highlights the effectiveness of this strategy.", + "bbox": [ + 511, + 430, + 908, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mixed datasets. We train our model on Virtual KITTI 2 (VKITTI2) [5] and a combination of mixed datasets, and subsequently evaluate both models. When trained exclusively on VKITTI2, the model achieves comparable performance to the mixed-dataset model on KITTI 2015, owing to the relatively small domain gap between VKITTI2 and KITTI 2015. However, this single-dataset training results in notable performance degradation on other datasets, indicating limited generalization. As shown in Tab. 5, comparing rows (3) and (8) demonstrates that our mixed dataset training strategy significantly improves the model's generalization across diverse domains.", + "bbox": [ + 511, + 612, + 908, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Coordinate embedding and the warped image. Incorporating the coordinate embedding process effectively encodes spatial information and reinforces the structural relationship between stereo views. Additionally, using the warped image as a condition enhances performance by providing pixel-level guidance. The warped image serves as a coarse prediction of the right view, helping the model focus", + "bbox": [ + 511, + 795, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26585", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f48a6f4eb7b3ac59ac97cd6873681a753205dfe17d8902181b6489700fe97fb8.jpg", + "table_caption": [ + "Table 4. Zero-Shot Generalization Benchmark. The first group shows methods trained with PSMNet, the second group shows methods trained with RAFT-Stereo, and the third group represents IGEV++." + ], + "table_footnote": [], + "table_body": "
MethodsKITTI 2012KITTI 2015ETH3DMiddlebury-QMiddlebury-H
D1-allEPE>3pxD1-allEPE>3pxD1-allEPE>1pxD1-allEPE>2pxD1-allEPE>2px
PSMNet [6]26.693.87727.3528.244.00328.534.2581.98915.7415.004.94120.2021.349.89933.26
MFS [60]4.3221.0094.7034.9631.6365.1812.2040.5358.1709.2171.64812.0712.343.24217.56
MfS+DAMv2 [65]4.9311.0635.3135.5041.4015.7371.9900.4938.1459.4471.82511.9012.343.67817.21
Mono2Stereo [59]4.0770.859-4.5071.052-1.5310.397-------
NeRFStereo [50]3.6830.8514.5074.8131.4485.0411.6200.55711.687.6191.45111.048.8922.59012.88
Ours + SD v1.53.6760.8464.0633.8790.9964.1121.1140.3785.5566.8871.13910.4311.332.26919.39
Ours + SD v2.13.8070.8204.1223.8980.9784.0960.8940.3244.1015.0370.8607.4638.0461.70913.73
RAFT-Stereo [24]3.9440.8304.2395.2541.1325.4680.8830.2672.6075.1571.21710.255.7741.43611.20
NeRFStereo [50]3.7870.8554.1405.3561.4695.5610.8300.3013.3705.8191.0038.3707.4261.82010.95
Ours + SD v1.53.2240.7343.4554.6831.0424.8970.9590.2942.6673.4501.0037.9524.4461.5149.94
Ours + SD v2.13.2190.7163.0193.9500.9714.2220.8350.2772.6384.0671.1279.3824.6181.52711.11
IGEV++ [62]5.6811.0996.1966.0281.2656.2391.7010.3564.6664.2380.8286.2745.2922.3847.752
Ours + SD v2.12.8020.6863.0153.7460.9524.0041.2260.3263.4293.6820.6845.5122.7970.8425.098
", + "bbox": [ + 91, + 128, + 936, + 316 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9d4b6f94da68350eef54b95ad089221d4da91351522c7f13fe17c72ff6b0e952.jpg", + "table_caption": [ + "Table 5. Ablation studies. SD v1.5 is used for the experiments." + ], + "table_footnote": [], + "table_body": "
Disparity Source#Mixed DatasetsRandom DropCoord EmbeddingWarped ImagePixel LossFusionMiddlebury 2014KITTI 2015
PSNRSSIMLPIPSPSNRSSIMLPIPS
Ground Truth(1)✓X✓✓✓✓23.5520.8620.064116.1640.6640.1740
(2)✓✓✓✓✓✓23.6990.8660.063520.4390.7460.1079
Pseudo Disparities(3)X✓✓✓✓✓20.0370.7450.121421.0850.7770.1113
(4)✓✓✓XXX22.0330.7960.073121.8660.7790.1002
(5)✓✓X✓XX22.6970.8320.071522.3130.7980.1011
(6)✓✓✓✓XX23.4070.8550.064422.5200.8050.0974
(7)✓✓✓✓✓X23.8110.8670.062122.7160.8110.0961
(8)✓✓✓✓✓✓23.8350.8680.062022.7490.8130.0958
", + "bbox": [ + 109, + 356, + 883, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on correcting local inconsistencies rather than synthesizing the entire image from scratch. As shown in rows (4), (5), and (6) in Tab. 5, the combination of coordinate embedding and warped image significantly improves generation performance, preserving geometric consistency and reducing artifacts like stitching boundaries. While warped image embedding contributes most to performance, its effectiveness decreases with additional fine-tuning iterations, revealing instability when using warped image alone as a conditioning signal. This underscores the importance of combining coordinate embedding and warped image to achieve stable, high-quality stereo image generation.", + "bbox": [ + 88, + 513, + 480, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pixel-level loss. By decoding the latent representations back to the image space, the model learns to minimize discrepancies directly in the observed image space, enforcing pixel-level alignment. As illustrated in rows (6) and (7), the introduction of the pixel-level loss significantly improves the generation performance, highlighting its role in maintaining high-fidelity stereo image generation.", + "bbox": [ + 88, + 698, + 480, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adaptive Fusion. The proposed adaptive fusion module significantly enhances visual quality by dynamically weighting the contributions of both the generated and warped images. A comparison of rows (7) and (8) demonstrates that this fusion strategy further refines pixel alignment, resulting in sharper, more structurally coherent, and", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "perceptually accurate reconstructions. The improvement in PSNR is more pronounced than that in SSIM and LPIPS, indicating that the adaptive fusion is particularly effective for pixel-level alignment.", + "bbox": [ + 511, + 513, + 906, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 587, + 633, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented GenStereo, a novel diffusion-based framework for open-world stereo image generation with applications in unsupervised stereo matching. Our method introduces several key innovations, including disparity-aware coordinate embeddings along with warped image embeddings, pixel-level loss, and an adaptive fusion module, to ensure high-quality stereo image generation with strong geometric and semantic consistency. Through extensive experiments across diverse datasets, we show that GenStereo significantly outperforms existing methods in both stereo generation and unsupervised matching. Our ablation studies demonstrate the efficacy of each proposed component.", + "bbox": [ + 511, + 612, + 903, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Despite promising results, diffusion-based models face inherent challenges with large disparities when generating right-view images due to large unconditioned regions. While our data augmentation such as random cropping and resizing mitigates these issues and performs well in typical stereo setups, future work could explore extending these models to accommodate larger disparities.", + "bbox": [ + 511, + 792, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26586", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We gratefully acknowledge the advanced computational resources provided by Engineering IT and Research Infrastructure Services at Washington University in St. Louis.", + "bbox": [ + 89, + 114, + 483, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Filippo Aleotti, Fabio Tosi, Li Zhang, Matteo Poggi, and Stefano Mattoccia. Reversing the cycle: self-supervised deep stereo through enhanced monocular distillation. In European Conference on Computer Vision, pages 614-632. Springer, 2020. 3", + "[2] Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences, 63:1-11, 2020. 1", + "[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1", + "[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In European Conf. on Computer Vision, pages 611-625. Springer-Verlag, 2012. 4", + "[5] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv preprint arXiv:2001.10773, 2020. 1, 4, 7", + "[6] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5410-5418, 2018. 7, 8", + "[7] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 7", + "[8] Cheng Chi, Qingjie Wang, Tianyu Hao, Peng Guo, and Xin Yang. Feature-level collaboration: Joint unsupervised learning of optical flow, stereo depth and camera motion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2463-2473, 2021. 3", + "[9] Jaehoon Cho, Dongbo Min, Youngjung Kim, and Kwanghoon Sohn. Deep monocular depth estimation leveraging a large-scale outdoor stereo dataset. Expert Systems with Applications, 178:114877, 2021. 1", + "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 6", + "[11] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 3", + "[12] Peter Hedman, Suhib Alsisan, Richard Szeliski, and Johannes Kopf. Casual 3d photography. ACM Transactions on Graphics (TOG), 36(6):1-15, 2017. 2", + "[13] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 2" + ], + "bbox": [ + 93, + 200, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[15] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2", + "[16] Junpeng Jing, Ye Mao, and Krystian Mikolajczyk. Match-stereo-videos: Bidirectional alignment for consistent dynamic stereo matching. In European Conference on Computer Vision, pages 415–432. Springer, 2024. 4", + "[17] Laurent Valentin Jospin, Hamid Laga, Farid Boussaid, and Mohammed Bennamoun. Active-passive simstereo, 2022. 4", + "[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 4", + "[19] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2", + "[20] Johannes Kopf, Fabian Langguth, Daniel Scharstein, Richard Szeliski, and Michael Goesele. Image-based rendering in the gradient domain. ACM Transactions on Graphics (TOG), 32(6):1-9, 2013. 2", + "[21] Hsueh-Ying Lai, Yi-Hsuan Tsai, and Wei-Chen Chiu. Bridging stereo matching and optical flow via spatiotemporal correspondence. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1890-1899, 2019. 3", + "[22] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16263-16272, 2022. 4, 5", + "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740-755. Springer, 2014. 2, 7", + "[24] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In IEEE International Conference on 3D Vision, pages 218-227, 2021. 7, 8", + "[25] Rui Liu, Chengxi Yang, Wenxiu Sun, Xiaogang Wang, and Hongsheng Li. Stereogan: Bridging synthetic-to-real domain gap by joint optimization of domain translation and stereo matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12757-12766, 2020. 3", + "[26] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 5, 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26587", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 4", + "[28] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3061-3070, 2015. 1, 5", + "[29] Jiteng Mu, Shalini De Mello, Zhiding Yu, Nuno Vasconcelos, Xiaolong Wang, Jan Kautz, and Sifei Liu. Coordgan: Self-supervised dense correspondences emerge from gans. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10011-10020, 2022. 3", + "[30] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In IEEE international conference on computer vision, pages 4990-4999, 2017. 7", + "[31] Matteo Poggi, Alessio Tonioni, Fabio Tosi, Stefano Mattoccia, and Luigi Di Stefano. Continual adaptation for deep stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4713-4729, 2021. 3", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[33] Pierluigi Zama Ramirez, Alex Costanzino, Fabio Tosi, Matteo Poggi, Samuele Salti, Stefano Mattoccia, and Luigi Di Stefano. Booster: a benchmark for depth from images of specular and transparent surfaces. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 1, 4", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2, 3, 5, 6, 7", + "[36] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nesic, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 31-42. Springer, 2014. 5, 6", + "[37] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3260-3269, 2017. 6", + "[38] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "to novel views with semantic-preserving generative warping. Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3", + "[39] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2", + "[40] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020. 2, 6", + "[41] J. Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. International Conference on Learning Representations, 2021. 2", + "[42] Xiao Song, Guorun Yang, Xinge Zhu, Hui Zhou, Zhe Wang, and Jianping Shi. Adastereo: A simple and efficient approach for adaptive stereo matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10328-10337, 2021. 3", + "[43] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 3", + "[44] Joshua Tokarsky, Ibrahim Abdulhafiz, Satya Ayyalasomayajula, Mostafa Mohsen, Navya G. Rao, and Adam Forbes. PLT-D3: A High-fidelity Dynamic Driving Simulation Dataset for Stereo Depth and Scene Flow, 2024. 4", + "[45] Alessio Tonioni, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Unsupervised adaptation for deep stereo. In IEEE International Conference on Computer Vision, pages 1605-1613, 2017. 3", + "[46] Alessio Tonioni, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Unsupervised domain adaptation for depth prediction from images. IEEE transactions on pattern analysis and machine intelligence, 42(10):2396-2409, 2019. 3", + "[47] Alessio Tonioni, Oscar Rahnama, Thomas Joy, Luigi Di Stefano, Thalaiyasingam Ajanthan, and Philip HS Torr. Learning to adapt for stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9661-9670, 2019. 3", + "[48] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 195-204, 2019. 3", + "[49] Fabio Tosi, Yiyi Liao, Carolin Schmitt, and Andreas Geiger. Smd-nets: Stereo mixture density networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8942-8952, 2021. 4", + "[50] Fabio Tosi, Alessio Tonioni, Daniele De Gregorio, and Matteo Poggi. Nerf-supervised deep stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 855-866, 2023. 2, 3, 8", + "[51] Fabio Tosi, Luca Bartolomei, and Matteo Poggi. A survey on deep stereo matching in the twenties. International Journal of Computer Vision, 2025. 4", + "[52] Jonathan Tremblay, Thang To, and Stan Birchfield. Falling things: A synthetic dataset for 3d object detection and pose" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26588", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "estimation. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2038-2041, 2018. 4", + "[53] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, JiaBin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 1", + "[54] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z. Dai, Andrea F. Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R. Walter, and Gregory Shakhnarovich. DIODE: A Dense Indoor and Outdoor DEpth Dataset. CoRR, abs/1908.00463, 2019. 7", + "[55] Lezhong Wang, Jeppe Revall Frisvad, Mark Bo Jensen, and Siavash Arjomand Bigdeli. Stereodiffusion: Training-free stereo image generation using latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7416-7425, 2024. 1, 2, 3, 5, 6, 7", + "[56] Qiang Wang, Shizhen Zheng, Qingsong Yan, Fei Deng, Kaiyong Zhao, and Xiaowen Chu. IRS: A large naturalistic indoor robotics stereo dataset to train deep models for disparity and surface normal estimation. In IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2021. 1, 4", + "[57] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4909-4916. IEEE, 2020. 1, 4", + "[58] Yang Wang, Peng Wang, Zhenheng Yang, Chenxu Luo, Yi Yang, and Wei Xu. Unos: Unified unsupervised optical-flow and stereo-depth estimation by watching videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8071-8081, 2019. 3", + "[59] Yuran Wang, Yingping Liang, Hesong Li, and Ying Fu. Mono2stereo: Monocular knowledge transfer for enhanced stereo matching. arXiv preprint arXiv:2411.09151, 2024. 3, 8", + "[60] Jamie Watson, Oisin Mac Aodha, Daniyar Turmukhambetov, Gabriel J Brostow, and Michael Firman. Learning stereo from single images. In European Conference on Computer Vision, pages 722-740. Springer, 2020. 2, 3, 6, 7, 8", + "[61] Zhexiao Xiong, Feng Qiao, Yu Zhang, and Nathan Jacobs. Stereoflowgan: Co-training for stereo and flow with unsupervised domain adaptation. In 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023. BMVA, 2023. 3", + "[62] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: Iterative multirange geometry encoding volumes for stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 7, 8", + "[63] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 2" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 4", + "[65] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875-21911, 2024. 1, 7, 8", + "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2", + "[67] Yiran Zhong, Yuchao Dai, and Hongdong Li. Self-supervised learning for stereo matching with self-improving ability. arXiv preprint arXiv:1709.00930, 2017. 3", + "[68] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In IEEE Conference on Computer Vision and Pattern Recognition, pages 633-641, 2017. 7" + ], + "bbox": [ + 516, + 92, + 903, + 373 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26589", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/2355f2bc-16e1-4dc5-88a7-53284ff42cd2_model.json b/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/2355f2bc-16e1-4dc5-88a7-53284ff42cd2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0905d868f32736050dec53378754e46113e9d5be --- /dev/null +++ b/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/2355f2bc-16e1-4dc5-88a7-53284ff42cd2_model.json @@ -0,0 +1,2664 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.9, + 0.154 + ], + "angle": 0, + "content": "Towards Open-World Generation of Stereo Images and Unsupervised Matching" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.181, + 0.734, + 0.218 + ], + "angle": 0, + "content": "Feng Qiao Zhexiao Xiong Eric Xing Nathan Jacobs Washington University in St. Louis" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.23, + 0.905, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.193, + 0.405, + 0.802, + 0.419 + ], + "angle": 0, + "content": "Figure 1. Comparison with diffusion-based methods for stereo image generation on the COCO dataset." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.431, + 0.327, + 0.447 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.465, + 0.485, + 0.799 + ], + "angle": 0, + "content": "Stereo images are fundamental to numerous applications, including extended reality (XR) devices, autonomous driving, and robotics. Unfortunately, acquiring high-quality stereo images remains challenging due to the precise calibration requirements of dual-camera setups and the complexity of obtaining accurate, dense disparity maps. Existing stereo image generation methods typically focus on either visual quality for viewing or geometric accuracy for matching, but not both. We introduce GenStereo, a diffusion-based approach, to bridge this gap. The method includes two primary innovations (1) conditioning the diffusion process on a disparity-aware coordinate embedding and a warped input image, allowing for more precise stereo alignment than previous methods, and (2) an adaptive fusion mechanism that intelligently combines the diffusion-generated image with a warped image, improving both realism and disparity consistency. Through extensive training on 11 diverse stereo datasets, GenStereo demonstrates strong generalization ability. GenStereo achieves state-of-the-art performance in both stereo image generation and unsupervised stereo matching tasks. Project page is available at https://qjizhi.github.io/genstereo." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.829, + 0.222, + 0.845 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The demand for stereo images continues growing with XR devices, autonomous driving, and robotics. However, acquiring high-quality stereo images remains chal" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.629 + ], + "angle": 0, + "content": "lenging due to complex camera calibration and environmental constraints, limiting the development of robust stereo matching models that generalize well across diverse scenarios. Real-world datasets often provide only sparse disparity annotations [28] or lack accurate ground truth altogether [9]. Although some datasets offer dense, accurate disparity maps [2, 33], they are typically limited to specific scenarios like indoor scenes. Moreover, capturing real-world stereo data requires complex sensor setups with precise calibration, often constraining baseline distances and scene diversity. Synthetic datasets, while providing precise disparity maps, suffer from domain gaps compared to real-world scenarios [5, 56, 57]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recent advancements in monocular depth estimation (MDE) models [3, 34, 64, 65] have provided increasingly accurate dense disparity maps, playing a crucial role in image generation from single views. Meanwhile, Text-to-Image (T2I) diffusion models, such as Stable Diffusion [35], have made remarkable strides in generating diverse and high-quality images from user-provided text prompts. However, these models often struggle to maintain spatial consistency when altering the viewpoint of the generated images. While there has been some research on 3D novel view generation [38, 53], there remains a gap in utilizing image generation models to directly produce stereo images. StereoDiffusion [55] is the first approach to leverage diffusion-based generation for stereo image generation, but it struggles with pixel-level accuracy, as it applies disparity shifts in the latent space and fills occluded regions with blurry pixels that lack meaningful semantic content. On the other hand, Stable Diffusion Inpainting (SD" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26579" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "Inpainting) [35], though not designed for stereo image generation, faces similar challenges by incorrectly filling occluded areas with semantically inappropriate pixels. Fig. 1 shows that previous diffusion-based methods and our methods tested on the COCO dataset [23]. The scale factor \\(\\gamma\\) is set to 0.15, meaning that the maximum disparity is \\(15\\%\\) of the image width." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.482, + 0.365 + ], + "angle": 0, + "content": "We present GenStereo, a novel framework that addresses both visual quality and geometric accuracy in stereo image generation. Existing approaches rely on either warping-based methods [60], which provide geometric accuracy but struggle with semantic consistency in occluded regions, or diffusion-based approaches [55], which maintain better semantic coherence but lack precise geometric control. GenStereo bridges this gap through a carefully designed two-stream architecture inspired by recent advances in human animation [15] and view synthesis [38], achieving both geometric precision and semantic consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.482, + 0.562 + ], + "angle": 0, + "content": "Our framework employs a multi-level constraint system for geometric precision and semantic coherence, starting with disparity-aware coordinate embedding that provides implicit geometric guidance. This is followed by a cross-view attention mechanism that enables semantic feature alignment between views, ensuring consistency in challenging areas like occlusions and complex textures. The framework leverages a dual-space supervision strategy that operates in both latent and pixel spaces, with an adaptive fusion mechanism that further optimizes pixel-level accuracy. To ensure robust generalization across diverse real-world scenarios, we train our model on a diverse dataset combining 11 stereo datasets with varying scenes and baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.482, + 0.744 + ], + "angle": 0, + "content": "Furthermore, we demonstrate how our high-quality stereo generation framework significantly improves unsupervised stereo matching learning. Previous approaches to unsupervised learning have been limited by either simplified warping and random background filling [60] or constraints to small-scale static scenes [50]. In contrast, our method enables large-scale training with diverse, photorealistic stereo images that maintain both geometric accuracy and semantic consistency. This advancement represents a significant step toward bridging the gap between supervised and unsupervised stereo matching approaches. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.75, + 0.481, + 0.81 + ], + "angle": 0, + "content": "- We propose GenStereo, the first unified framework for open-world stereo image generation that addresses visual quality and geometric accuracy, enabling both practical applications and unsupervised stereo matching." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.811, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- We introduce a comprehensive multi-level constraint system that combines: (1) disparity-aware coordinate embedding with warped image conditioning for geometric guidance, (2) cross-view attention mechanism for semantic feature alignment, and (3) dual-space supervision with adaptive fusion for pixel-accuracy generation." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.75, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "- Extensive experimental validation demonstrating state-of-the-art performance in both stereo image generation and unsupervised stereo matching." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.152, + 0.655, + 0.168 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.178, + 0.776, + 0.194 + ], + "angle": 0, + "content": "2.1. Conditional Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.201, + 0.905, + 0.442 + ], + "angle": 0, + "content": "Diffusion models began with DDPM [14], advancing from slow, probabilistic generation to faster sampling with DDIM [41]. Stable Diffusion [35], a pioneering T2I model, further improved efficiency by performing diffusion in a compact latent space, combining it with CLIP-based [32] text conditioning for versatile T2I generation. However, Stable Diffusion lacks fine-grained control over specific visual attributes, relying mainly on text prompts without structured guidance. To address these limitations, conditional diffusion models have emerged. ControlNet [66] is a key advancement, introducing conditioning mechanisms for structural inputs (such as edges, depth maps, or poses) that maintain flexibility while enabling precise, user-driven synthesis. However, ControlNet still faces challenges in achieving pixel-level accuracy in complex images, such as stereo image generation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.443, + 0.905, + 0.73 + ], + "angle": 0, + "content": "Recent research [15, 19, 38, 39, 63] has explored the self-attention properties within T2I models. Text2VideoZero [19] and MVDream [39] generate consistent visuals across video frames or 3D multi-views by sharing self-attention, whileAnimate-Anyone [15] and MagicAnimate [63] apply a similar approach to produce human dance videos through fine-tuned T2I models. GenWarp [38] employs augmenting self-attention with cross-view attention between the reference and target views to generate novel images. This two-stream architecture augments the features of the denoising net. Inspired by the adaptability and control of these self-attention-based architectures, our method builds on their strengths. However, their direct application to stereo image generation faces fundamental challenges. First, they typically don't incorporate disparity information, which is crucial for stereoscopic consistency. Second, achieving the pixel-level accuracy required for comfortable stereo viewing demands specialized architectural considerations beyond existing frameworks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.742, + 0.862, + 0.759 + ], + "angle": 0, + "content": "2.2. Stereo Image Generation and Inpainting" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.905, + 0.885 + ], + "angle": 0, + "content": "Traditional view synthesis approaches, including geometry-based reconstruction [12, 13, 20] and recent NeRF-based methods [50], excel at rendering novel views but are constrained by their requirement for multiple input images of static scenes. While 3D Photography techniques [12, 40] attempt to overcome this limitation through depth-based mesh projection, they often struggle with complex occlusion handling and require sophisticated post-processing pipelines." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The evolution of stereo generation methods has pro" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26580" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.409 + ], + "angle": 0, + "content": "gressed from simple warping-based approaches to more sophisticated diffusion-based solutions. MfS [60] introduced a basic framework that attempted to handle occlusions by sampling random background patches from the dataset, but this led to semantic inconsistencies and visible artifacts at occlusion areas. While SD-Inpainting [35] introduced more coherent semantic content through its prior learning, it failed to maintain local consistency between inpainted regions and their surrounding context, often producing visible discontinuities in texture and structure. Mono2Stereo [59] made progress by fine-tuning SD-Inpainting specifically for stereo generation, enabling unsupervised stereo matching, though visible artifacts persist in in-painted regions. Stereodiffusion [55] marked a significant shift by introducing end-to-end diffusion-based stereo image generation, but its latent-space warping approach without explicit geometric constraints led to compromised pixel-level accuracy. Despite these advances, the challenge of generating photorealistic stereo images that maintain both visual quality and geometric fidelity remains largely unsolved, particularly for diverse real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.418, + 0.364, + 0.433 + ], + "angle": 0, + "content": "2.3. Unsupervised Stereo Matching" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.636 + ], + "angle": 0, + "content": "Traditional unsupervised stereo matching approaches primarily rely on photometric consistency. Methods like [11, 47, 48, 67] leverage photometric losses across stereo images, while others [8, 21, 58] extend this to temporal sequences. A parallel line of research explores proxy supervision strategies, either through carefully designed algorithmic supervisors [31, 45, 46] or knowledge distillation from pre-trained networks [1]. Recent domain adaptation approaches [25, 42, 61] attempt to bridge the gap between synthetic and real-world domains or facilitate cross-dataset adaptation. However, these methods often exhibit limited generalization capability beyond their target domains [1], particularly struggling with diverse real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.484, + 0.817 + ], + "angle": 0, + "content": "More recent approaches have explored novel directions for stereo generation and matching. MfS [60] pioneered unsupervised stereo matching using MDE as guidance, but it relies on random background sampling for occlusion filling, leading to semantic inconsistencies. While NeRF-Stereo [50] achieves high-quality results through neural radiance fields, it requires multiple views of static scenes, limiting its practical applications. Mono2Stereo [59] leverages Stable Diffusion for inpainting occluded regions, representing a significant advance in generative stereo synthesis, though it still exhibits artifacts in occluded areas where geometric and semantic consistency is crucial." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.192, + 0.847 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Given a left view image \\( I_{l} \\) and its corresponding disparity map \\( \\hat{D}_{l} \\) predicted by an MDE model, our goal is to generate a high-quality right view image \\( \\hat{I}_r \\) that maintains both vi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "sual quality and geometric consistency. As shown in Fig. 2, our GenStereo framework processes the left view as the reference image and synthesizes the right view as the target image. The generated triplet \\(\\langle I_l,\\hat{D}_l,\\hat{I}_r\\rangle\\) not only serves as a stereo pair for visualization but also provides training data for unsupervised stereo matching." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.19, + 0.863, + 0.207 + ], + "angle": 0, + "content": "3.1. Disparity-Aware Coordinate Embedding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.212, + 0.906, + 0.332 + ], + "angle": 0, + "content": "Traditional inpainting-based stereo image generation methods often suffer from visible boundaries between warped and inpainted regions. To address this limitation, we propose a disparity-aware coordinate embedding scheme inspired by recent advances in coordinate-based generation [29, 38]. Our approach utilizes dual coordinate embeddings: a canonical embedding for the left view and its warped counterpart for the right view." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.906, + 0.393 + ], + "angle": 0, + "content": "Specifically, we first construct a canonical 2D coordinate map \\( X \\in \\mathbb{R}^{h \\times w \\times 2} \\) with values normalized to \\([-1,1]\\). This map is transformed into Fourier features [43] through a positional encoding function \\( \\phi \\):" + }, + { + "type": "equation", + "bbox": [ + 0.661, + 0.402, + 0.905, + 0.418 + ], + "angle": 0, + "content": "\\[\nC _ {l} = \\phi (X) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.906, + 0.486 + ], + "angle": 0, + "content": "The resulting Fourier feature map \\( C_l \\) serves as the coordinate embedding for the left view \\( I_l \\). We then generate the right view embedding by warping \\( C_l \\) according to the disparity map:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.487, + 0.905, + 0.502 + ], + "angle": 0, + "content": "\\[\nC _ {r} = \\operatorname {w a r p} \\left(C _ {l}, D _ {l}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.597 + ], + "angle": 0, + "content": "where \\( D_{l} \\) represents ground-truth disparity during training and predicted disparity \\( \\hat{D}_l \\) during inference. These coordinate embeddings \\( (C_l \\) and \\( C_r) \\) are integrated into their respective view features \\( (F_{l} \\) and \\( F_{r}) \\) through convolutional layers, establishing strong geometric correspondence while maintaining visual consistency between views." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.675 + ], + "angle": 0, + "content": "Our approach differs from GenWarp [38] by utilizing disparity maps instead of camera matrices for warping, enabling more precise pixel-level control. We further enhance geometric consistency by incorporating the warped image \\(I_{\\mathrm{warp}}\\) as additional conditioning for the denoising U-Net." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.813, + 0.696 + ], + "angle": 0, + "content": "3.2. Cross-View Feature Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To facilitate effective information exchange between views, we adopt a two-parallel U-Nets framework pretrained from Stable Diffusion, inspired by GenWarp [38]. However, our work differs in its novel conditioning strategy: the reference U-Net takes concatenated \\((I_l, C_l)\\) as conditions to process the left view, while the denoising U-Net takes concatenated \\((I_{\\mathrm{warp}}, C_r)\\) as conditions when synthesizing the right view \\(\\hat{I}_r\\). This design ensures that each U-Net receives both image content and corresponding coordinate information for better feature extraction. As our model is conditioned on a reference image and its disparity map, we replace text condition of SD with image embedding of the left image through a CLIP image encoder in both U-Nets." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26581" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.912, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.35, + 0.304, + 0.645, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Overview of the GenStereo framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.484, + 0.405 + ], + "angle": 0, + "content": "To incorporate left-view information during right-view generation, we concatenate the reference features with the target features in the attention mechanism. Specifically, we compute cross-view attention as:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.412, + 0.482, + 0.429 + ], + "angle": 0, + "content": "\\[\nq = F _ {r}, \\quad k = \\left[ F _ {l}, F _ {r} \\right], \\quad v = \\left[ F _ {l}, F _ {r} \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.541 + ], + "angle": 0, + "content": "where \\(F_{l}\\) is derived from the reference U-Net conditioned on \\((I_l,C_l)\\), while \\(F_{r}\\) comes from the denoising U-Net conditioned on \\((I_{\\mathrm{warp}},C_r)\\). This dual-stream attention mechanism allows the model to adaptively balance between semantic consistency from the reference view and geometric accuracy from the warped view, with each stream guided by both image content and coordinate information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.548, + 0.265, + 0.563 + ], + "angle": 0, + "content": "3.3. Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.75 + ], + "angle": 0, + "content": "Mixed training [22, 34, 64] has been proven to be an effective training strategy for domain generalization in both monocular and stereo matching depth estimation. We follow the datasets summarized in [51] and utilize the most widely available public synthetic datasets. Based on our observations, real-world datasets can negatively impact performance, even when there are only minor calibration errors or slight differences in imaging between the two cameras. Tab. 1 shows the comprehensive list of datasets utilized. Datasets containing abstract images are not included, because our model aims to have a good generalization capability in the real world." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.484, + 0.843 + ], + "angle": 0, + "content": "To address dataset imbalance, we employ a resampling strategy. Specifically, we sample the smaller datasets multiple times until the number reaches \\(10\\%\\) of the largest dataset. As the Stable Diffusion model only accepts images of size \\(512 \\times 512\\) for SD v1.5 and \\(768 \\times 768\\) for SD v2.1, we apply a random square crop and resize." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.303, + 0.866 + ], + "angle": 0, + "content": "3.4. Pixel Space Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Given an RGB image \\(I \\in \\mathbb{R}^{H \\times W \\times 3}\\), Latent Diffusion Models (LDMs) first encode it into a latent representation" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.343, + 0.906, + 0.369 + ], + "angle": 0, + "content": "Table 1. Our training data sources. All datasets are synthetic with the size of 684K." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.382, + 0.882, + 0.519 + ], + "angle": 0, + "content": "
DatasetIndoorOutdoor#ImagesYear
TartanAir [57]✓✓306K2020
Dynamic Replica [18]✓145K2023
IRS [56]✓103K2021
Falling Things [52]✓✓61K2018
VKITTI2 [5]✓21K2020
InfinigenSV [16]✓17K2024
SimStereo [17]✓14K2022
UnrealStereo4K [49]✓✓8K2021
Spring [27]✓5K2023
PLT-D3 [44]✓3K2024
Sintel [4]✓1K2012
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.542, + 0.906, + 0.588 + ], + "angle": 0, + "content": "\\(z = \\mathcal{E}(I)\\), where \\(z\\in \\mathbb{R}^{h\\times w\\times c}\\). While this latent space operation significantly reduces computational costs, it may compromise pixel-level accuracy during image generation." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.588, + 0.886, + 0.603 + ], + "angle": 0, + "content": "The standard LDM objective operates in latent space:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.61, + 0.905, + 0.63 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {l a t e n t}} := \\mathbb {E} _ {z, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.906, + 0.696 + ], + "angle": 0, + "content": "where \\( z_{t} \\) is the noisy latent at timestep \\( t \\), \\( \\epsilon \\) is the noise added to create \\( z_{t} \\), and \\( \\epsilon_{\\theta}(z_{t}, t) \\) is the network's prediction of this noise. The model learns to denoise by predicting the noise that was added." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.698, + 0.906, + 0.743 + ], + "angle": 0, + "content": "To maintain pixel-level accuracy, we introduce an additional pixel-space loss by decoding both the predicted and target latent variables:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.75, + 0.905, + 0.77 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {p i x e l}} = \\mathbb {E} _ {z, \\epsilon , t} \\left[ \\| \\mathcal {D} \\left(z _ {\\text {p r e d}}\\right) - \\mathcal {D} \\left(z _ {\\text {t a r g e t}}\\right) \\| _ {2} ^ {2} \\right] \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.777, + 0.794, + 0.792 + ], + "angle": 0, + "content": "The final objective combines both spaces:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.801, + 0.905, + 0.817 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\text {l a t e n t}} + \\alpha L _ {\\text {p i x e l}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.892, + 0.841 + ], + "angle": 0, + "content": "where \\(\\alpha = 1\\) balances latent and pixel-space supervision." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.742, + 0.865 + ], + "angle": 0, + "content": "3.5. Adaptive Fusion Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To achieve seamless integration between generated and warped content, we propose an Adaptive Fusion Module" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26582" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "that learns to combine \\( I_{gen} \\) and \\( I_{warp} \\) based on local context and confidence. The module predicts spatially-varying fusion weights through a lightweight network:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.149, + 0.483, + 0.168 + ], + "angle": 0, + "content": "\\[\nW = \\sigma \\left(f _ {\\theta} \\left(\\operatorname {c o n c a t} \\left(I _ {g e n}, I _ {w a r p}, M\\right)\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.484, + 0.223 + ], + "angle": 0, + "content": "where \\( f_{\\theta} \\) is a \\( 3 \\times 3 \\) convolutional layer and \\( \\sigma \\) is the sigmoid activation, ensuring weights in [0, 1]. The final right view is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.232, + 0.483, + 0.252 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {r} = M \\odot W \\odot I _ {\\text {w a r p}} + (1 - M \\odot W) \\odot I _ {\\text {g e n}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.484, + 0.352 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise multiplication. This formulation adaptively favors warped content in high-confidence regions (\\(M \\approx 1\\)) while relying on generated content in occluded or uncertain areas. The learned weights \\(W\\) enable smooth transitions between warped and generated regions, ensuring both geometric accuracy and visual consistency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.362, + 0.341, + 0.379 + ], + "angle": 0, + "content": "3.6. Random Disparity Dropout" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.484, + 0.445 + ], + "angle": 0, + "content": "To simulate the sparsity of real-world disparity maps like KITTI's LiDAR-derived ground truth, we randomly apply disparity dropout to \\(10\\%\\) of training samples. For each selected sample, we first generate a dropout ratio:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.456, + 0.483, + 0.473 + ], + "angle": 0, + "content": "\\[\nr \\sim \\operatorname {U n i f o r m} (0, 1) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.515 + ], + "angle": 0, + "content": "Using this ratio, we create a binary mask where each pixel has a probability \\( r \\) of being dropped:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.524, + 0.483, + 0.567 + ], + "angle": 0, + "content": "\\[\nM _ {\\text {r a n d}} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {o t h e r w i s e} \\\\ 0 & \\text {w i t h p r o b a b i l i t y} r \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.606 + ], + "angle": 0, + "content": "The final mask combines this random dropout with the warping mask:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.617, + 0.483, + 0.635 + ], + "angle": 0, + "content": "\\[\nM = M _ {\\text {w a r p}} \\vee M _ {\\text {r a n d}} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.707 + ], + "angle": 0, + "content": "where \\(\\vee\\) denotes element-wise logical OR. This strategy encourages the model to handle sparse disparity inputs, improving its robustness and generalization to real-world scenarios where dense disparity maps may not be available." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.719, + 0.224, + 0.736 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.32, + 0.759 + ], + "angle": 0, + "content": "4.1. Stereo Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Experimental Setup. We fine-tune the pretrained Stable Diffusion (SD) UNet for 3 epochs across all experiments. For evaluation, we utilize two widely adopted stereo vision benchmarks: Middlebury 2014 [36] and KITTI 2015 [28], neither of which are included in our training set. Middlebury 2014 comprises 23 high-resolution indoor stereo pairs captured with wide-baseline stereo cameras under varying illumination conditions, providing challenging scenarios for stereo generation. KITTI 2015 contains 200 outdoor stereo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "pairs of street scenes with LiDAR-based sparse disparity ground truth, offering diverse real-world testing scenarios. Following StereoDiffusion [55], we preprocess Middlebury images to \\(512 \\times 512\\) resolution, while KITTI images are center-cropped before resizing to \\(512 \\times 512\\) to maintain the most informative regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.184, + 0.907, + 0.379 + ], + "angle": 0, + "content": "We conduct comprehensive comparisons against both traditional and learning-based baselines. Traditional approaches include naive solutions such as leave blank and image stretching. For learning-based comparisons, we evaluate against state-of-the-art diffusion models adapted for stereo generation: Repaint [26] and SD-Inpainting [35], as well as StereoDiffusion, which specifically targets stereo image generation. To quantitatively assess the quality of generated right-view images \\((\\hat{I_r})\\) against ground truth \\((I_{r})\\) we employ three complementary metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) with SqueezeNet backbone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.906, + 0.607 + ], + "angle": 0, + "content": "Quantitative Results. Tab. 2 presents the generation performance on Middlebury 2014 and KITTI 2015. The sparsity of ground truth disparities in KITTI limits evaluation accuracy, making pseudo-label disparities more effective for assessing stereo consistency. This sparsity particularly affects inpainting-based methods, as they struggle to infer accurate right-view images in areas with missing depth information, especially for dynamic objects and distant regions. In contrast, Middlebury 2014, which provides dense ground truth disparities, demonstrates a smaller performance gap between using ground truth and using pseudo disparities, underscoring the advantage of high-quality disparity annotations for generation. We obtain pseudo disparities from CREStereo [22] pretrained model, where evaluation datasets are not used for training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.906, + 0.835 + ], + "angle": 0, + "content": "Qualitative Results. We present qualitative comparisons between our method and several baselines: leave blank, StereoDiffusion, and SD-Inpainting, with ground truth as reference. As shown in Fig. 3, our method demonstrates superior performance on the Middlebury 2014 dataset. While SD-Inpainting struggles with occlusion handling, often generating semantically inconsistent content in occluded regions, and StereoDiffusion exhibits loss of fine details due to latent-space warping, our approach successfully maintains both geometric accuracy and visual fidelity. Particularly noteworthy is our method's ability to generate coherent right-view images while preserving pixel-level correspondence with the left images. Additional qualitative results on the KITTI 2015 dataset and more visualizations on other datasets can be found in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.731, + 0.864 + ], + "angle": 0, + "content": "4.2. Unsupervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Experimental Setup. We conduct two comprehensive experiments to validate the effectiveness of our proposed gen" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26583" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 2. Quantitative results of image generation for Middlebury 2014 and KITTI 2015 datasets. The top three results for each metric are highlighted with a first, second, and third background, respectively." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.13, + 0.857, + 0.294 + ], + "angle": 0, + "content": "
MethodsSD versionMiddlebury 2014KITTI 2015
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Leave blank-11.328+2.4830.315+0.2300.450+0.09712.980+4.9190.374+0.2860.313+0.222
-3.489-0.154-0.089-3.831-0.251-0.109
Stretch-14.842+2.7140.432+0.1900.285+0.11214.757+5.3750.429+0.2870.212+0.145
-2.753+0.265-0.089-4.694-0.271-0.100
3D Photography [40]-14.190+2.4640.427+0.2380.275+0.07314.540+7.2560.398+0.2700.210+0.099
-2.798-0.175-0.065-4.023-0.323-0.073
RePaint [26]1.515.102+2.9090.462+0.2680.311+0.08815.056+5.3660.462+0.2680.251+0.128
-2.802-0.253-0.079-4.897-0.285-0.095
SD-Inpainting [35]1.515.740+3.5110.412+0.2360.311+0.0989.792+2.8090.230+0.2360.652+0.144
-3.351-0.188-0.164-4.070-0.229-0.139
StereoDiffusion [55]1.515.456+2.6690.468+0.2520.231+0.09615.679+5.8880.481+0.2450.205+0.135
-3.313-0.205-0.088-5.487-0.310-0.099
Ours + GT1.523.699+5.4560.866+0.0840.064+0.05320.439+7.5180.746+0.1550.108+0.117
-6.033-0.141-0.034-8.541-0.274-0.048
Ours + Pseudo1.523.835+6.8420.868+0.0960.062+0.06322.749+6.7220.813+0.1100.096+0.068
-6.741-0.157-0.038-8.102-0.152-0.104
Ours + GT2.124.819+6.5190.906+0.0570.061+0.03219.836+8.4830.765+0.2450.135+0.059
-4.252-0.132-0.032-7.090-0.092-0.068
Ours + Pseudo2.125.142+6.9700.911+0.0660.060+0.06123.488+6.4440.849+0.1300.109+0.050
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.306, + 0.227, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.412, + 0.226, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.516, + 0.226, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.622, + 0.226, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.727, + 0.213, + 0.74 + ], + "angle": 0, + "content": "Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.307, + 0.361, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.412, + 0.361, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.516, + 0.361, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.622, + 0.362, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.257, + 0.727, + 0.335, + 0.739 + ], + "angle": 0, + "content": "Leave Blank" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.307, + 0.498, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.412, + 0.498, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.516, + 0.498, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.622, + 0.498, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.727, + 0.474, + 0.739 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.307, + 0.634, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.412, + 0.634, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.516, + 0.634, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.622, + 0.634, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.727, + 0.611, + 0.74 + ], + "angle": 0, + "content": "SD-Inpainting" + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.307, + 0.77, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.412, + 0.77, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.516, + 0.77, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.622, + 0.77, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.727, + 0.752, + 0.739 + ], + "angle": 0, + "content": "StereoDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.307, + 0.905, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.412, + 0.905, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.516, + 0.905, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.622, + 0.905, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.727, + 0.855, + 0.739 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.236, + 0.753, + 0.76, + 0.768 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison on Middlebury 2014 with ground truth disparity maps." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.795, + 0.419, + 0.809 + ], + "angle": 0, + "content": "eration method for unsupervised stereo matching." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.885 + ], + "angle": 0, + "content": "In the first experiment, we fine-tune a stereo matching model pretrained on the SceneFlow dataset using the KITTI 2012 and KITTI 2015 benchmarks. We split the training and evaluation sets into 160/34 images for KITTI 2012 and 160/40 images for KITTI 2015, following [60]." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The second experiment evaluates various unsupervised" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "stereo matching approaches by generating stereo images. Specifically, KITTI 2012 [10] has 194 stereo images, KITTI 2015 has 200 stereo images, ETH3D [37] has 27 stereo pairs, and the Middlebury v3 training set [36] has 15 stereo pairs at Quarter and Half resolutions (Midd-Q, Midd-H). This diverse dataset selection allows us to comprehensively assess the generalization capability of our method across" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26584" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.148 + ], + "angle": 0, + "content": "Table 3. Comparison with other diffusion-based stereo image generation methods for unsupervised stereo matching. The first group of methods is trained on PSMNet while the second group is trained on RAFT-Stereo. SD v1.5 is used for the experiments." + }, + { + "type": "table", + "bbox": [ + 0.116, + 0.156, + 0.458, + 0.289 + ], + "angle": 0, + "content": "
MethodsKITTI 2012KITTI 2015
D1-all ↓EPE ↓D1-all ↓EPE ↓
Leave blank25.2193.54224.5682.937
SD-Inpainting [35]3.9070.8944.4901.059
StereoDiffusion [55]15.2132.2205.6511.154
Ours3.8020.8153.9330.991
Leave blank4.1840.9524.0740.957
SD-Inpainting [35]4.0160.9454.1280.972
StereoDiffusion [55]12.7561.8116.3601.174
Ours3.5370.9303.6530.943
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.32, + 0.332 + ], + "angle": 0, + "content": "different domains and resolutions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.335, + 0.486, + 0.562 + ], + "angle": 0, + "content": "We evaluate the performance of various unsupervised stereo matching approaches using the D1-all metric, which measures the percentage of incorrect disparity predictions (error \\(>3\\mathrm{px}\\) or \\(5\\%\\) of the true disparity), the End-PointError (EPE), and the \\(>n\\mathrm{px}\\) metric, which denotes the percentage of pixels with an EPE exceeding \\(n\\) pixels. To demonstrate the architecture-agnostic nature of our approach, we employ three distinct stereo models: PSM-Net [6], a cost-volume-based method, RAFT-Stereo [24], an iterative refinement model, and IGEV++ [62], one of the most recent methods. For disparity estimation, we utilize the Depth Anything Model v2 (DAMv2) [65], from which we obtain the disparity map. The disparity values are normalized to the range \\([0, 1]\\) and subsequently scaled by a factor \\(\\gamma\\) to enable flexibility generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.485, + 0.7 + ], + "angle": 0, + "content": "Fine-tuning on KITTI. To ensure fair and consistent comparisons across all generation methods, we construct the training datasets using a standardized procedure. Specifically, we sample the scale factor \\(\\gamma\\) from the set \\(\\{0.05, 0.1, 0.15, 0.2, 0.25\\}\\). For each selected scale, we employ various generation methods to create the right images based on 160 left images and their corresponding scaled disparity maps for both KITTI 2012 and KITTI 2015. This procedure results in a total of 800 stereo pairs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.485, + 0.793 + ], + "angle": 0, + "content": "As shown in Tab. 3, higher generation quality leads to improved stereo matching performance. Among the compared methods, SD-Inpainting achieves the closest performance to ours due to its pixel-level warping operation. However, it falls short in providing semantic consistency in occluded areas, which limits its overall effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.904 + ], + "angle": 0, + "content": "Unsupervised Generalization. Following the methodology of constructing the Mono for Stereo (MfS) dataset in [60], we construct our training dataset, MfS-GenStereo, from the same datasets, including ADE20K [68], Mapillary Vistas [30], DIODE [54], Depth in the Wild [7], and COCO 2017 [23]. During the generation process, we randomly select a scale factor in the range \\([0, 0.35]\\) and set the maxi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "mum disparity to 256 for PSMNet. As shown in Tab. 4, our method enables unsupervised stereo matching using only left images, achieving superior performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.909, + 0.396 + ], + "angle": 0, + "content": "We compare our approach with the original MfS [60] and its enhanced variant using the Depth Anything Model v2 (DAMv2), which generates the right view images by filling the occluded areas with random background patches from the dataset. The experiments indicate that, despite the better depth prior, the synthesis of right-view images remains a critical bottleneck in this methodology. Additionally, we compare with Mono2Stereo, which fine-tunes SD-Inpainting for right-view image generation, and NeRF-Stereo, which generates stereo images by reconstructing the static scenes using Neural Radiance Fields (NeRF), limiting the diversity of its datasets. Our experimental results indicate that our proposed method consistently surpasses these existing approaches in generating high-quality stereo pairs with pixel-level accuracy, demonstrating robust generalization capabilities and significantly enhancing stereo matching performance across diverse datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.408, + 0.671, + 0.423 + ], + "angle": 0, + "content": "4.3. Ablation Study." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.909, + 0.612 + ], + "angle": 0, + "content": "Random disparity drop. Since the ground-truth disparity maps in real-world datasets (e.g., KITTI) are often derived from LiDAR, resulting in sparse and incomplete point clouds, our random disparity dropout strategy mitigates the model's reliance on dense disparity supervision. This enhancement allows the model to generate plausible right images even when only sparse disparity maps are available. This strategy significantly improves performance on datasets with sparse ground-truth annotations, demonstrating better generalization to real-world scenarios. As shown in Tab. 5, comparing rows (1) and (2) highlights the effectiveness of this strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Mixed datasets. We train our model on Virtual KITTI 2 (VKITTI2) [5] and a combination of mixed datasets, and subsequently evaluate both models. When trained exclusively on VKITTI2, the model achieves comparable performance to the mixed-dataset model on KITTI 2015, owing to the relatively small domain gap between VKITTI2 and KITTI 2015. However, this single-dataset training results in notable performance degradation on other datasets, indicating limited generalization. As shown in Tab. 5, comparing rows (3) and (8) demonstrates that our mixed dataset training strategy significantly improves the model's generalization across diverse domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Coordinate embedding and the warped image. Incorporating the coordinate embedding process effectively encodes spatial information and reinforces the structural relationship between stereo views. Additionally, using the warped image as a condition enhances performance by providing pixel-level guidance. The warped image serves as a coarse prediction of the right view, helping the model focus" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26585" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 4. Zero-Shot Generalization Benchmark. The first group shows methods trained with PSMNet, the second group shows methods trained with RAFT-Stereo, and the third group represents IGEV++." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.13, + 0.937, + 0.317 + ], + "angle": 0, + "content": "
MethodsKITTI 2012KITTI 2015ETH3DMiddlebury-QMiddlebury-H
D1-allEPE>3pxD1-allEPE>3pxD1-allEPE>1pxD1-allEPE>2pxD1-allEPE>2px
PSMNet [6]26.693.87727.3528.244.00328.534.2581.98915.7415.004.94120.2021.349.89933.26
MFS [60]4.3221.0094.7034.9631.6365.1812.2040.5358.1709.2171.64812.0712.343.24217.56
MfS+DAMv2 [65]4.9311.0635.3135.5041.4015.7371.9900.4938.1459.4471.82511.9012.343.67817.21
Mono2Stereo [59]4.0770.859-4.5071.052-1.5310.397-------
NeRFStereo [50]3.6830.8514.5074.8131.4485.0411.6200.55711.687.6191.45111.048.8922.59012.88
Ours + SD v1.53.6760.8464.0633.8790.9964.1121.1140.3785.5566.8871.13910.4311.332.26919.39
Ours + SD v2.13.8070.8204.1223.8980.9784.0960.8940.3244.1015.0370.8607.4638.0461.70913.73
RAFT-Stereo [24]3.9440.8304.2395.2541.1325.4680.8830.2672.6075.1571.21710.255.7741.43611.20
NeRFStereo [50]3.7870.8554.1405.3561.4695.5610.8300.3013.3705.8191.0038.3707.4261.82010.95
Ours + SD v1.53.2240.7343.4554.6831.0424.8970.9590.2942.6673.4501.0037.9524.4461.5149.94
Ours + SD v2.13.2190.7163.0193.9500.9714.2220.8350.2772.6384.0671.1279.3824.6181.52711.11
IGEV++ [62]5.6811.0996.1966.0281.2656.2391.7010.3564.6664.2380.8286.2745.2922.3847.752
Ours + SD v2.12.8020.6863.0153.7460.9524.0041.2260.3263.4293.6820.6845.5122.7970.8425.098
" + }, + { + "type": "table_caption", + "bbox": [ + 0.312, + 0.332, + 0.684, + 0.346 + ], + "angle": 0, + "content": "Table 5. Ablation studies. SD v1.5 is used for the experiments." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.357, + 0.885, + 0.489 + ], + "angle": 0, + "content": "
Disparity Source#Mixed DatasetsRandom DropCoord EmbeddingWarped ImagePixel LossFusionMiddlebury 2014KITTI 2015
PSNRSSIMLPIPSPSNRSSIMLPIPS
Ground Truth(1)✓X✓✓✓✓23.5520.8620.064116.1640.6640.1740
(2)✓✓✓✓✓✓23.6990.8660.063520.4390.7460.1079
Pseudo Disparities(3)X✓✓✓✓✓20.0370.7450.121421.0850.7770.1113
(4)✓✓✓XXX22.0330.7960.073121.8660.7790.1002
(5)✓✓X✓XX22.6970.8320.071522.3130.7980.1011
(6)✓✓✓✓XX23.4070.8550.064422.5200.8050.0974
(7)✓✓✓✓✓X23.8110.8670.062122.7160.8110.0961
(8)✓✓✓✓✓✓23.8350.8680.062022.7490.8130.0958
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.514, + 0.482, + 0.695 + ], + "angle": 0, + "content": "on correcting local inconsistencies rather than synthesizing the entire image from scratch. As shown in rows (4), (5), and (6) in Tab. 5, the combination of coordinate embedding and warped image significantly improves generation performance, preserving geometric consistency and reducing artifacts like stitching boundaries. While warped image embedding contributes most to performance, its effectiveness decreases with additional fine-tuning iterations, revealing instability when using warped image alone as a conditioning signal. This underscores the importance of combining coordinate embedding and warped image to achieve stable, high-quality stereo image generation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.699, + 0.482, + 0.806 + ], + "angle": 0, + "content": "Pixel-level loss. By decoding the latent representations back to the image space, the model learns to minimize discrepancies directly in the observed image space, enforcing pixel-level alignment. As illustrated in rows (6) and (7), the introduction of the pixel-level loss significantly improves the generation performance, highlighting its role in maintaining high-fidelity stereo image generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Adaptive Fusion. The proposed adaptive fusion module significantly enhances visual quality by dynamically weighting the contributions of both the generated and warped images. A comparison of rows (7) and (8) demonstrates that this fusion strategy further refines pixel alignment, resulting in sharper, more structurally coherent, and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.907, + 0.575 + ], + "angle": 0, + "content": "perceptually accurate reconstructions. The improvement in PSNR is more pronounced than that in SSIM and LPIPS, indicating that the adaptive fusion is particularly effective for pixel-level alignment." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.634, + 0.604 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.905, + 0.793 + ], + "angle": 0, + "content": "We presented GenStereo, a novel diffusion-based framework for open-world stereo image generation with applications in unsupervised stereo matching. Our method introduces several key innovations, including disparity-aware coordinate embeddings along with warped image embeddings, pixel-level loss, and an adaptive fusion module, to ensure high-quality stereo image generation with strong geometric and semantic consistency. Through extensive experiments across diverse datasets, we show that GenStereo significantly outperforms existing methods in both stereo generation and unsupervised matching. Our ablation studies demonstrate the efficacy of each proposed component." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Limitations. Despite promising results, diffusion-based models face inherent challenges with large disparities when generating right-view images due to large unconditioned regions. While our data augmentation such as random cropping and resizing mitigates these issues and performs well in typical stereo setups, future work could explore extending these models to accommodate larger disparities." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26586" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.162 + ], + "angle": 0, + "content": "We gratefully acknowledge the advanced computational resources provided by Engineering IT and Research Infrastructure Services at Washington University in St. Louis." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.177, + 0.188, + 0.192 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.484, + 0.27 + ], + "angle": 0, + "content": "[1] Filippo Aleotti, Fabio Tosi, Li Zhang, Matteo Poggi, and Stefano Mattoccia. Reversing the cycle: self-supervised deep stereo through enhanced monocular distillation. In European Conference on Computer Vision, pages 614-632. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.272, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[2] Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences, 63:1-11, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.329, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.387, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In European Conf. on Computer Vision, pages 611-625. Springer-Verlag, 2012. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.484, + 0.471 + ], + "angle": 0, + "content": "[5] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv preprint arXiv:2001.10773, 2020. 1, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[6] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5410-5418, 2018. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[7] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.484, + 0.629 + ], + "angle": 0, + "content": "[8] Cheng Chi, Qingjie Wang, Tianyu Hao, Peng Guo, and Xin Yang. Feature-level collaboration: Joint unsupervised learning of optical flow, stereo depth and camera motion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2463-2473, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.484, + 0.686 + ], + "angle": 0, + "content": "[9] Jaehoon Cho, Dongbo Min, Youngjung Kim, and Kwanghoon Sohn. Deep monocular depth estimation leveraging a large-scale outdoor stereo dataset. Expert Systems with Applications, 178:114877, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[11] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[12] Peter Hedman, Suhib Alsisan, Richard Szeliski, and Johannes Kopf. Casual 3d photography. ACM Transactions on Graphics (TOG), 36(6):1-15, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[13] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.201, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[15] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[16] Junpeng Jing, Ye Mao, and Krystian Mikolajczyk. Match-stereo-videos: Bidirectional alignment for consistent dynamic stereo matching. In European Conference on Computer Vision, pages 415–432. Springer, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[17] Laurent Valentin Jospin, Hamid Laga, Farid Boussaid, and Mohammed Bennamoun. Active-passive simstereo, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.28, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[19] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[20] Johannes Kopf, Fabian Langguth, Daniel Scharstein, Richard Szeliski, and Michael Goesele. Image-based rendering in the gradient domain. ACM Transactions on Graphics (TOG), 32(6):1-9, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[21] Hsueh-Ying Lai, Yi-Hsuan Tsai, and Wei-Chen Chiu. Bridging stereo matching and optical flow via spatiotemporal correspondence. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1890-1899, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.549, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[22] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16263-16272, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740-755. Springer, 2014. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[24] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In IEEE International Conference on 3D Vision, pages 218-227, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[25] Rui Liu, Chengxi Yang, Wenxiu Sun, Xiaogang Wang, and Hongsheng Li. Stereogan: Bridging synthetic-to-real domain gap by joint optimization of domain translation and stereo matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12757-12766, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[26] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26587" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[27] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[28] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3061-3070, 2015. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[29] Jiteng Mu, Shalini De Mello, Zhiding Yu, Nuno Vasconcelos, Xiaolong Wang, Jan Kautz, and Sifei Liu. Coordgan: Self-supervised dense correspondences emerge from gans. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10011-10020, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[30] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In IEEE international conference on computer vision, pages 4990-4999, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[31] Matteo Poggi, Alessio Tonioni, Fabio Tosi, Stefano Mattoccia, and Luigi Di Stefano. Continual adaptation for deep stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4713-4729, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[33] Pierluigi Zama Ramirez, Alex Costanzino, Fabio Tosi, Matteo Poggi, Samuele Salti, Stefano Mattoccia, and Luigi Di Stefano. Booster: a benchmark for depth from images of specular and transparent surfaces. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[36] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nesic, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 31-42. Springer, 2014. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[37] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3260-3269, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[38] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "to novel views with semantic-preserving generative warping. Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[39] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[40] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[41] J. Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[42] Xiao Song, Guorun Yang, Xinge Zhu, Hui Zhou, Zhe Wang, and Jianping Shi. Adastereo: A simple and efficient approach for adaptive stereo matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10328-10337, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[43] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[44] Joshua Tokarsky, Ibrahim Abdulhafiz, Satya Ayyalasomayajula, Mostafa Mohsen, Navya G. Rao, and Adam Forbes. PLT-D3: A High-fidelity Dynamic Driving Simulation Dataset for Stereo Depth and Scene Flow, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[45] Alessio Tonioni, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Unsupervised adaptation for deep stereo. In IEEE International Conference on Computer Vision, pages 1605-1613, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[46] Alessio Tonioni, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Unsupervised domain adaptation for depth prediction from images. IEEE transactions on pattern analysis and machine intelligence, 42(10):2396-2409, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[47] Alessio Tonioni, Oscar Rahnama, Thomas Joy, Luigi Di Stefano, Thalaiyasingam Ajanthan, and Philip HS Torr. Learning to adapt for stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9661-9670, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.72 + ], + "angle": 0, + "content": "[48] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 195-204, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[49] Fabio Tosi, Yiyi Liao, Carolin Schmitt, and Andreas Geiger. Smd-nets: Stereo mixture density networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8942-8952, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[50] Fabio Tosi, Alessio Tonioni, Daniele De Gregorio, and Matteo Poggi. Nerf-supervised deep stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 855-866, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[51] Fabio Tosi, Luca Bartolomei, and Matteo Poggi. A survey on deep stereo matching in the twenties. International Journal of Computer Vision, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[52] Jonathan Tremblay, Thang To, and Stan Birchfield. Falling things: A synthetic dataset for 3d object detection and pose" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26588" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "estimation. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2038-2041, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[53] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, JiaBin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[54] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z. Dai, Andrea F. Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R. Walter, and Gregory Shakhnarovich. DIODE: A Dense Indoor and Outdoor DEpth Dataset. CoRR, abs/1908.00463, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.329 + ], + "angle": 0, + "content": "[55] Lezhong Wang, Jeppe Revall Frisvad, Mark Bo Jensen, and Siavash Arjomand Bigdeli. Stereodiffusion: Training-free stereo image generation using latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7416-7425, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.331, + 0.482, + 0.411 + ], + "angle": 0, + "content": "[56] Qiang Wang, Shizhen Zheng, Qingsong Yan, Fei Deng, Kaiyong Zhao, and Xiaowen Chu. IRS: A large naturalistic indoor robotics stereo dataset to train deep models for disparity and surface normal estimation. In IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2021. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[57] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4909-4916. IEEE, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[58] Yang Wang, Peng Wang, Zhenheng Yang, Chenxu Luo, Yi Yang, and Wei Xu. Unos: Unified unsupervised optical-flow and stereo-depth estimation by watching videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8071-8081, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[59] Yuran Wang, Yingping Liang, Hesong Li, and Ying Fu. Mono2stereo: Monocular knowledge transfer for enhanced stereo matching. arXiv preprint arXiv:2411.09151, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[60] Jamie Watson, Oisin Mac Aodha, Daniyar Turmukhambetov, Gabriel J Brostow, and Michael Firman. Learning stereo from single images. In European Conference on Computer Vision, pages 722-740. Springer, 2020. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[61] Zhexiao Xiong, Feng Qiao, Yu Zhang, and Nathan Jacobs. Stereoflowgan: Co-training for stereo and flow with unsupervised domain adaptation. In 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023. BMVA, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[62] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: Iterative multirange geometry encoding volumes for stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[63] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[64] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[65] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875-21911, 2024. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[67] Yiran Zhong, Yuchao Dai, and Hongdong Li. Self-supervised learning for stereo matching with self-improving ability. arXiv preprint arXiv:1709.00930, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[68] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In IEEE Conference on Computer Vision and Pattern Recognition, pages 633-641, 2017. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26589" + } + ] +] \ No newline at end of file diff --git a/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/2355f2bc-16e1-4dc5-88a7-53284ff42cd2_origin.pdf b/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/2355f2bc-16e1-4dc5-88a7-53284ff42cd2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..73ea8a724d7d64635232406dae19ebe09f366140 --- /dev/null +++ b/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/2355f2bc-16e1-4dc5-88a7-53284ff42cd2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a8fbac4c65a0da2efb7e6fb0c47d66a08eacc6d19229050b5308d5c0d7e20455 +size 3859656 diff --git a/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/full.md b/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..930eb599f3d4edcdabd28f32b8108b88ee942783 --- /dev/null +++ b/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/full.md @@ -0,0 +1,387 @@ +# Towards Open-World Generation of Stereo Images and Unsupervised Matching + +Feng Qiao Zhexiao Xiong Eric Xing Nathan Jacobs Washington University in St. Louis + +![](images/e85e509e88f5dec64869b1deea24ad1c82ada4f7de0182661ef05b5422c41462.jpg) +Figure 1. Comparison with diffusion-based methods for stereo image generation on the COCO dataset. + +# Abstract + +Stereo images are fundamental to numerous applications, including extended reality (XR) devices, autonomous driving, and robotics. Unfortunately, acquiring high-quality stereo images remains challenging due to the precise calibration requirements of dual-camera setups and the complexity of obtaining accurate, dense disparity maps. Existing stereo image generation methods typically focus on either visual quality for viewing or geometric accuracy for matching, but not both. We introduce GenStereo, a diffusion-based approach, to bridge this gap. The method includes two primary innovations (1) conditioning the diffusion process on a disparity-aware coordinate embedding and a warped input image, allowing for more precise stereo alignment than previous methods, and (2) an adaptive fusion mechanism that intelligently combines the diffusion-generated image with a warped image, improving both realism and disparity consistency. Through extensive training on 11 diverse stereo datasets, GenStereo demonstrates strong generalization ability. GenStereo achieves state-of-the-art performance in both stereo image generation and unsupervised stereo matching tasks. Project page is available at https://qjizhi.github.io/genstereo. + +# 1. Introduction + +The demand for stereo images continues growing with XR devices, autonomous driving, and robotics. However, acquiring high-quality stereo images remains chal + +lenging due to complex camera calibration and environmental constraints, limiting the development of robust stereo matching models that generalize well across diverse scenarios. Real-world datasets often provide only sparse disparity annotations [28] or lack accurate ground truth altogether [9]. Although some datasets offer dense, accurate disparity maps [2, 33], they are typically limited to specific scenarios like indoor scenes. Moreover, capturing real-world stereo data requires complex sensor setups with precise calibration, often constraining baseline distances and scene diversity. Synthetic datasets, while providing precise disparity maps, suffer from domain gaps compared to real-world scenarios [5, 56, 57]. + +Recent advancements in monocular depth estimation (MDE) models [3, 34, 64, 65] have provided increasingly accurate dense disparity maps, playing a crucial role in image generation from single views. Meanwhile, Text-to-Image (T2I) diffusion models, such as Stable Diffusion [35], have made remarkable strides in generating diverse and high-quality images from user-provided text prompts. However, these models often struggle to maintain spatial consistency when altering the viewpoint of the generated images. While there has been some research on 3D novel view generation [38, 53], there remains a gap in utilizing image generation models to directly produce stereo images. StereoDiffusion [55] is the first approach to leverage diffusion-based generation for stereo image generation, but it struggles with pixel-level accuracy, as it applies disparity shifts in the latent space and fills occluded regions with blurry pixels that lack meaningful semantic content. On the other hand, Stable Diffusion Inpainting (SD + +Inpainting) [35], though not designed for stereo image generation, faces similar challenges by incorrectly filling occluded areas with semantically inappropriate pixels. Fig. 1 shows that previous diffusion-based methods and our methods tested on the COCO dataset [23]. The scale factor $\gamma$ is set to 0.15, meaning that the maximum disparity is $15\%$ of the image width. + +We present GenStereo, a novel framework that addresses both visual quality and geometric accuracy in stereo image generation. Existing approaches rely on either warping-based methods [60], which provide geometric accuracy but struggle with semantic consistency in occluded regions, or diffusion-based approaches [55], which maintain better semantic coherence but lack precise geometric control. GenStereo bridges this gap through a carefully designed two-stream architecture inspired by recent advances in human animation [15] and view synthesis [38], achieving both geometric precision and semantic consistency. + +Our framework employs a multi-level constraint system for geometric precision and semantic coherence, starting with disparity-aware coordinate embedding that provides implicit geometric guidance. This is followed by a cross-view attention mechanism that enables semantic feature alignment between views, ensuring consistency in challenging areas like occlusions and complex textures. The framework leverages a dual-space supervision strategy that operates in both latent and pixel spaces, with an adaptive fusion mechanism that further optimizes pixel-level accuracy. To ensure robust generalization across diverse real-world scenarios, we train our model on a diverse dataset combining 11 stereo datasets with varying scenes and baselines. + +Furthermore, we demonstrate how our high-quality stereo generation framework significantly improves unsupervised stereo matching learning. Previous approaches to unsupervised learning have been limited by either simplified warping and random background filling [60] or constraints to small-scale static scenes [50]. In contrast, our method enables large-scale training with diverse, photorealistic stereo images that maintain both geometric accuracy and semantic consistency. This advancement represents a significant step toward bridging the gap between supervised and unsupervised stereo matching approaches. Our contributions can be summarized as follows: + +- We propose GenStereo, the first unified framework for open-world stereo image generation that addresses visual quality and geometric accuracy, enabling both practical applications and unsupervised stereo matching. +- We introduce a comprehensive multi-level constraint system that combines: (1) disparity-aware coordinate embedding with warped image conditioning for geometric guidance, (2) cross-view attention mechanism for semantic feature alignment, and (3) dual-space supervision with adaptive fusion for pixel-accuracy generation. + +- Extensive experimental validation demonstrating state-of-the-art performance in both stereo image generation and unsupervised stereo matching. + +# 2. Related Work + +# 2.1. Conditional Diffusion Models + +Diffusion models began with DDPM [14], advancing from slow, probabilistic generation to faster sampling with DDIM [41]. Stable Diffusion [35], a pioneering T2I model, further improved efficiency by performing diffusion in a compact latent space, combining it with CLIP-based [32] text conditioning for versatile T2I generation. However, Stable Diffusion lacks fine-grained control over specific visual attributes, relying mainly on text prompts without structured guidance. To address these limitations, conditional diffusion models have emerged. ControlNet [66] is a key advancement, introducing conditioning mechanisms for structural inputs (such as edges, depth maps, or poses) that maintain flexibility while enabling precise, user-driven synthesis. However, ControlNet still faces challenges in achieving pixel-level accuracy in complex images, such as stereo image generation. + +Recent research [15, 19, 38, 39, 63] has explored the self-attention properties within T2I models. Text2VideoZero [19] and MVDream [39] generate consistent visuals across video frames or 3D multi-views by sharing self-attention, whileAnimate-Anyone [15] and MagicAnimate [63] apply a similar approach to produce human dance videos through fine-tuned T2I models. GenWarp [38] employs augmenting self-attention with cross-view attention between the reference and target views to generate novel images. This two-stream architecture augments the features of the denoising net. Inspired by the adaptability and control of these self-attention-based architectures, our method builds on their strengths. However, their direct application to stereo image generation faces fundamental challenges. First, they typically don't incorporate disparity information, which is crucial for stereoscopic consistency. Second, achieving the pixel-level accuracy required for comfortable stereo viewing demands specialized architectural considerations beyond existing frameworks. + +# 2.2. Stereo Image Generation and Inpainting + +Traditional view synthesis approaches, including geometry-based reconstruction [12, 13, 20] and recent NeRF-based methods [50], excel at rendering novel views but are constrained by their requirement for multiple input images of static scenes. While 3D Photography techniques [12, 40] attempt to overcome this limitation through depth-based mesh projection, they often struggle with complex occlusion handling and require sophisticated post-processing pipelines. + +The evolution of stereo generation methods has pro + +gressed from simple warping-based approaches to more sophisticated diffusion-based solutions. MfS [60] introduced a basic framework that attempted to handle occlusions by sampling random background patches from the dataset, but this led to semantic inconsistencies and visible artifacts at occlusion areas. While SD-Inpainting [35] introduced more coherent semantic content through its prior learning, it failed to maintain local consistency between inpainted regions and their surrounding context, often producing visible discontinuities in texture and structure. Mono2Stereo [59] made progress by fine-tuning SD-Inpainting specifically for stereo generation, enabling unsupervised stereo matching, though visible artifacts persist in in-painted regions. Stereodiffusion [55] marked a significant shift by introducing end-to-end diffusion-based stereo image generation, but its latent-space warping approach without explicit geometric constraints led to compromised pixel-level accuracy. Despite these advances, the challenge of generating photorealistic stereo images that maintain both visual quality and geometric fidelity remains largely unsolved, particularly for diverse real-world scenarios. + +# 2.3. Unsupervised Stereo Matching + +Traditional unsupervised stereo matching approaches primarily rely on photometric consistency. Methods like [11, 47, 48, 67] leverage photometric losses across stereo images, while others [8, 21, 58] extend this to temporal sequences. A parallel line of research explores proxy supervision strategies, either through carefully designed algorithmic supervisors [31, 45, 46] or knowledge distillation from pre-trained networks [1]. Recent domain adaptation approaches [25, 42, 61] attempt to bridge the gap between synthetic and real-world domains or facilitate cross-dataset adaptation. However, these methods often exhibit limited generalization capability beyond their target domains [1], particularly struggling with diverse real-world scenarios. + +More recent approaches have explored novel directions for stereo generation and matching. MfS [60] pioneered unsupervised stereo matching using MDE as guidance, but it relies on random background sampling for occlusion filling, leading to semantic inconsistencies. While NeRF-Stereo [50] achieves high-quality results through neural radiance fields, it requires multiple views of static scenes, limiting its practical applications. Mono2Stereo [59] leverages Stable Diffusion for inpainting occluded regions, representing a significant advance in generative stereo synthesis, though it still exhibits artifacts in occluded areas where geometric and semantic consistency is crucial. + +# 3. Methods + +Given a left view image $I_{l}$ and its corresponding disparity map $\hat{D}_{l}$ predicted by an MDE model, our goal is to generate a high-quality right view image $\hat{I}_r$ that maintains both vi + +sual quality and geometric consistency. As shown in Fig. 2, our GenStereo framework processes the left view as the reference image and synthesizes the right view as the target image. The generated triplet $\langle I_l,\hat{D}_l,\hat{I}_r\rangle$ not only serves as a stereo pair for visualization but also provides training data for unsupervised stereo matching. + +# 3.1. Disparity-Aware Coordinate Embedding + +Traditional inpainting-based stereo image generation methods often suffer from visible boundaries between warped and inpainted regions. To address this limitation, we propose a disparity-aware coordinate embedding scheme inspired by recent advances in coordinate-based generation [29, 38]. Our approach utilizes dual coordinate embeddings: a canonical embedding for the left view and its warped counterpart for the right view. + +Specifically, we first construct a canonical 2D coordinate map $X \in \mathbb{R}^{h \times w \times 2}$ with values normalized to $[-1,1]$ . This map is transformed into Fourier features [43] through a positional encoding function $\phi$ : + +$$ +C _ {l} = \phi (X) \tag {1} +$$ + +The resulting Fourier feature map $C_l$ serves as the coordinate embedding for the left view $I_l$ . We then generate the right view embedding by warping $C_l$ according to the disparity map: + +$$ +C _ {r} = \operatorname {w a r p} \left(C _ {l}, D _ {l}\right), \tag {2} +$$ + +where $D_{l}$ represents ground-truth disparity during training and predicted disparity $\hat{D}_l$ during inference. These coordinate embeddings $(C_l$ and $C_r)$ are integrated into their respective view features $(F_{l}$ and $F_{r})$ through convolutional layers, establishing strong geometric correspondence while maintaining visual consistency between views. + +Our approach differs from GenWarp [38] by utilizing disparity maps instead of camera matrices for warping, enabling more precise pixel-level control. We further enhance geometric consistency by incorporating the warped image $I_{\mathrm{warp}}$ as additional conditioning for the denoising U-Net. + +# 3.2. Cross-View Feature Enhancement + +To facilitate effective information exchange between views, we adopt a two-parallel U-Nets framework pretrained from Stable Diffusion, inspired by GenWarp [38]. However, our work differs in its novel conditioning strategy: the reference U-Net takes concatenated $(I_l, C_l)$ as conditions to process the left view, while the denoising U-Net takes concatenated $(I_{\mathrm{warp}}, C_r)$ as conditions when synthesizing the right view $\hat{I}_r$ . This design ensures that each U-Net receives both image content and corresponding coordinate information for better feature extraction. As our model is conditioned on a reference image and its disparity map, we replace text condition of SD with image embedding of the left image through a CLIP image encoder in both U-Nets. + +![](images/5d1ad429e007e8b74d4ad5dc5ef20bfa88ab62ebbe8e293b87f95f9403d1bf40.jpg) +Figure 2. Overview of the GenStereo framework. + +To incorporate left-view information during right-view generation, we concatenate the reference features with the target features in the attention mechanism. Specifically, we compute cross-view attention as: + +$$ +q = F _ {r}, \quad k = \left[ F _ {l}, F _ {r} \right], \quad v = \left[ F _ {l}, F _ {r} \right]. \tag {3} +$$ + +where $F_{l}$ is derived from the reference U-Net conditioned on $(I_l,C_l)$ , while $F_{r}$ comes from the denoising U-Net conditioned on $(I_{\mathrm{warp}},C_r)$ . This dual-stream attention mechanism allows the model to adaptively balance between semantic consistency from the reference view and geometric accuracy from the warped view, with each stream guided by both image content and coordinate information. + +# 3.3. Training Strategy + +Mixed training [22, 34, 64] has been proven to be an effective training strategy for domain generalization in both monocular and stereo matching depth estimation. We follow the datasets summarized in [51] and utilize the most widely available public synthetic datasets. Based on our observations, real-world datasets can negatively impact performance, even when there are only minor calibration errors or slight differences in imaging between the two cameras. Tab. 1 shows the comprehensive list of datasets utilized. Datasets containing abstract images are not included, because our model aims to have a good generalization capability in the real world. + +To address dataset imbalance, we employ a resampling strategy. Specifically, we sample the smaller datasets multiple times until the number reaches $10\%$ of the largest dataset. As the Stable Diffusion model only accepts images of size $512 \times 512$ for SD v1.5 and $768 \times 768$ for SD v2.1, we apply a random square crop and resize. + +# 3.4. Pixel Space Alignment + +Given an RGB image $I \in \mathbb{R}^{H \times W \times 3}$ , Latent Diffusion Models (LDMs) first encode it into a latent representation + +Table 1. Our training data sources. All datasets are synthetic with the size of 684K. + +
DatasetIndoorOutdoor#ImagesYear
TartanAir [57]✓✓306K2020
Dynamic Replica [18]✓145K2023
IRS [56]✓103K2021
Falling Things [52]✓✓61K2018
VKITTI2 [5]✓21K2020
InfinigenSV [16]✓17K2024
SimStereo [17]✓14K2022
UnrealStereo4K [49]✓✓8K2021
Spring [27]✓5K2023
PLT-D3 [44]✓3K2024
Sintel [4]✓1K2012
+ +$z = \mathcal{E}(I)$ , where $z\in \mathbb{R}^{h\times w\times c}$ . While this latent space operation significantly reduces computational costs, it may compromise pixel-level accuracy during image generation. + +The standard LDM objective operates in latent space: + +$$ +L _ {\text {l a t e n t}} := \mathbb {E} _ {z, \epsilon , t} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t\right) \| _ {2} ^ {2} \right] \tag {4} +$$ + +where $z_{t}$ is the noisy latent at timestep $t$ , $\epsilon$ is the noise added to create $z_{t}$ , and $\epsilon_{\theta}(z_{t}, t)$ is the network's prediction of this noise. The model learns to denoise by predicting the noise that was added. + +To maintain pixel-level accuracy, we introduce an additional pixel-space loss by decoding both the predicted and target latent variables: + +$$ +L _ {\text {p i x e l}} = \mathbb {E} _ {z, \epsilon , t} \left[ \| \mathcal {D} \left(z _ {\text {p r e d}}\right) - \mathcal {D} \left(z _ {\text {t a r g e t}}\right) \| _ {2} ^ {2} \right] \tag {5} +$$ + +The final objective combines both spaces: + +$$ +L = L _ {\text {l a t e n t}} + \alpha L _ {\text {p i x e l}} \tag {6} +$$ + +where $\alpha = 1$ balances latent and pixel-space supervision. + +# 3.5. Adaptive Fusion Module + +To achieve seamless integration between generated and warped content, we propose an Adaptive Fusion Module + +that learns to combine $I_{gen}$ and $I_{warp}$ based on local context and confidence. The module predicts spatially-varying fusion weights through a lightweight network: + +$$ +W = \sigma \left(f _ {\theta} \left(\operatorname {c o n c a t} \left(I _ {g e n}, I _ {w a r p}, M\right)\right)\right), \tag {7} +$$ + +where $f_{\theta}$ is a $3 \times 3$ convolutional layer and $\sigma$ is the sigmoid activation, ensuring weights in [0, 1]. The final right view is computed as: + +$$ +\hat {I} _ {r} = M \odot W \odot I _ {\text {w a r p}} + (1 - M \odot W) \odot I _ {\text {g e n}}, \tag {8} +$$ + +where $\odot$ denotes element-wise multiplication. This formulation adaptively favors warped content in high-confidence regions ( $M \approx 1$ ) while relying on generated content in occluded or uncertain areas. The learned weights $W$ enable smooth transitions between warped and generated regions, ensuring both geometric accuracy and visual consistency. + +# 3.6. Random Disparity Dropout + +To simulate the sparsity of real-world disparity maps like KITTI's LiDAR-derived ground truth, we randomly apply disparity dropout to $10\%$ of training samples. For each selected sample, we first generate a dropout ratio: + +$$ +r \sim \operatorname {U n i f o r m} (0, 1) \tag {9} +$$ + +Using this ratio, we create a binary mask where each pixel has a probability $r$ of being dropped: + +$$ +M _ {\text {r a n d}} (i, j) = \left\{ \begin{array}{l l} 1 & \text {o t h e r w i s e} \\ 0 & \text {w i t h p r o b a b i l i t y} r \end{array} \right. \tag {10} +$$ + +The final mask combines this random dropout with the warping mask: + +$$ +M = M _ {\text {w a r p}} \vee M _ {\text {r a n d}} \tag {11} +$$ + +where $\vee$ denotes element-wise logical OR. This strategy encourages the model to handle sparse disparity inputs, improving its robustness and generalization to real-world scenarios where dense disparity maps may not be available. + +# 4. Experiments + +# 4.1. Stereo Image Generation + +Experimental Setup. We fine-tune the pretrained Stable Diffusion (SD) UNet for 3 epochs across all experiments. For evaluation, we utilize two widely adopted stereo vision benchmarks: Middlebury 2014 [36] and KITTI 2015 [28], neither of which are included in our training set. Middlebury 2014 comprises 23 high-resolution indoor stereo pairs captured with wide-baseline stereo cameras under varying illumination conditions, providing challenging scenarios for stereo generation. KITTI 2015 contains 200 outdoor stereo + +pairs of street scenes with LiDAR-based sparse disparity ground truth, offering diverse real-world testing scenarios. Following StereoDiffusion [55], we preprocess Middlebury images to $512 \times 512$ resolution, while KITTI images are center-cropped before resizing to $512 \times 512$ to maintain the most informative regions. + +We conduct comprehensive comparisons against both traditional and learning-based baselines. Traditional approaches include naive solutions such as leave blank and image stretching. For learning-based comparisons, we evaluate against state-of-the-art diffusion models adapted for stereo generation: Repaint [26] and SD-Inpainting [35], as well as StereoDiffusion, which specifically targets stereo image generation. To quantitatively assess the quality of generated right-view images $(\hat{I_r})$ against ground truth $(I_{r})$ we employ three complementary metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) with SqueezeNet backbone. + +Quantitative Results. Tab. 2 presents the generation performance on Middlebury 2014 and KITTI 2015. The sparsity of ground truth disparities in KITTI limits evaluation accuracy, making pseudo-label disparities more effective for assessing stereo consistency. This sparsity particularly affects inpainting-based methods, as they struggle to infer accurate right-view images in areas with missing depth information, especially for dynamic objects and distant regions. In contrast, Middlebury 2014, which provides dense ground truth disparities, demonstrates a smaller performance gap between using ground truth and using pseudo disparities, underscoring the advantage of high-quality disparity annotations for generation. We obtain pseudo disparities from CREStereo [22] pretrained model, where evaluation datasets are not used for training. + +Qualitative Results. We present qualitative comparisons between our method and several baselines: leave blank, StereoDiffusion, and SD-Inpainting, with ground truth as reference. As shown in Fig. 3, our method demonstrates superior performance on the Middlebury 2014 dataset. While SD-Inpainting struggles with occlusion handling, often generating semantically inconsistent content in occluded regions, and StereoDiffusion exhibits loss of fine details due to latent-space warping, our approach successfully maintains both geometric accuracy and visual fidelity. Particularly noteworthy is our method's ability to generate coherent right-view images while preserving pixel-level correspondence with the left images. Additional qualitative results on the KITTI 2015 dataset and more visualizations on other datasets can be found in the supplementary materials. + +# 4.2. Unsupervised Learning + +Experimental Setup. We conduct two comprehensive experiments to validate the effectiveness of our proposed gen + +Table 2. Quantitative results of image generation for Middlebury 2014 and KITTI 2015 datasets. The top three results for each metric are highlighted with a first, second, and third background, respectively. + +
MethodsSD versionMiddlebury 2014KITTI 2015
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Leave blank-11.328+2.4830.315+0.2300.450+0.09712.980+4.9190.374+0.2860.313+0.222
-3.489-0.154-0.089-3.831-0.251-0.109
Stretch-14.842+2.7140.432+0.1900.285+0.11214.757+5.3750.429+0.2870.212+0.145
-2.753+0.265-0.089-4.694-0.271-0.100
3D Photography [40]-14.190+2.4640.427+0.2380.275+0.07314.540+7.2560.398+0.2700.210+0.099
-2.798-0.175-0.065-4.023-0.323-0.073
RePaint [26]1.515.102+2.9090.462+0.2680.311+0.08815.056+5.3660.462+0.2680.251+0.128
-2.802-0.253-0.079-4.897-0.285-0.095
SD-Inpainting [35]1.515.740+3.5110.412+0.2360.311+0.0989.792+2.8090.230+0.2360.652+0.144
-3.351-0.188-0.164-4.070-0.229-0.139
StereoDiffusion [55]1.515.456+2.6690.468+0.2520.231+0.09615.679+5.8880.481+0.2450.205+0.135
-3.313-0.205-0.088-5.487-0.310-0.099
Ours + GT1.523.699+5.4560.866+0.0840.064+0.05320.439+7.5180.746+0.1550.108+0.117
-6.033-0.141-0.034-8.541-0.274-0.048
Ours + Pseudo1.523.835+6.8420.868+0.0960.062+0.06322.749+6.7220.813+0.1100.096+0.068
-6.741-0.157-0.038-8.102-0.152-0.104
Ours + GT2.124.819+6.5190.906+0.0570.061+0.03219.836+8.4830.765+0.2450.135+0.059
-4.252-0.132-0.032-7.090-0.092-0.068
Ours + Pseudo2.125.142+6.9700.911+0.0660.060+0.06123.488+6.4440.849+0.1300.109+0.050
+ +![](images/4626db6c11940104f518bea321646586a4836eb60a01189874c9d0b4cde068f8.jpg) + +![](images/e35f46a7566228bd105a1072af012120eedf8ae32516f84ee8af4b56749ad612.jpg) + +![](images/4c99701b8f14edda3eac7936d93cdd5d18465709f8f6813a3adb971b07b3eaef.jpg) + +![](images/6673940f90317b9467a02dd793cbc4e6f71131540b557b91668350a0610e7722.jpg) +Reference Image + +![](images/d9779e095e6a380a0cebc24fa84dc627f9ae4d45cdcddccaa5ff484a174fa8d8.jpg) + +![](images/39c2eb229569dae5fdcf288a45c6e07c5bccc5c0f5aef1883979f7d94284c45e.jpg) + +![](images/08769934ae2ec1826398d26c679943933846da98f088a22d287267d2bc587082.jpg) + +![](images/cf51d8a69ae0c53540d400439a7fb044a4596a9f502b8310ffb55b35d1a3ab4c.jpg) +Leave Blank + +![](images/31ea98a2d156060801743b2a9e76e9d10b2f0d83f6f0dd334a2005b83caaf770.jpg) + +![](images/d932d1ec3188b94be7d44086205892028959feacc77f07b6d2bda12bb03dcb48.jpg) + +![](images/5967345400804e8acbaf2021f80606f50817fe297f19e9d66c26e7d56e8573f8.jpg) + +![](images/aebdb82329d735a030550ca5d57340b6d33e7d453ea266b2cfa4594b5ac12cbb.jpg) +Ground Truth + +![](images/4e2a4b4ad75882de53bf77bf3fd7cd3981195d19deafe8814bd70d992169b8d3.jpg) + +![](images/1ed3776bc4acc972c4b81d76d8746827ceb475d123e2c230b65d34e515d7fd72.jpg) + +![](images/3a85ece81b50054519d62a78672352903f300d95e4b039a212add1e0006fc325.jpg) + +![](images/c536a43c7b962480e27d95a57817f3cc142c6395346474e1405d1bbdee667c5e.jpg) +SD-Inpainting + +![](images/7d1be42928f302dd7cedc9810ad721d98b3089a8fafb70ca14ff97a465b2b82e.jpg) + +![](images/307bfa7e1b8efd2b97b33703d60b265416ebe370137f38ed87e6934dcd3f4b0f.jpg) + +![](images/713a917839ca3b21711c64fc409793faa7a8a0c1fb3b06b9210c5fb6a0e658eb.jpg) + +![](images/9b41ab7322bd3624ae38bf3563c21dc56a6f7ea10d01e24da83af9b4f6825f87.jpg) +StereoDiffusion +Figure 3. Qualitative comparison on Middlebury 2014 with ground truth disparity maps. + +![](images/61135b89c4a0ec761b3380772bff5b4f650c3d033c8242ee86d07e3280a000b7.jpg) + +![](images/852dcbbd60b718c68912436a0633ff3d9620027380d535ce9e0ac5fb1c603cb2.jpg) + +![](images/f4592ed5f91857f3dce4d8e70f71b8f0fa616601fa0e6ede490b273eddc97e25.jpg) + +![](images/76ce018505aaa8e4052d8d7480232010c24be52ebe3da5db2fe54d1510214c21.jpg) +Ours + +eration method for unsupervised stereo matching. + +In the first experiment, we fine-tune a stereo matching model pretrained on the SceneFlow dataset using the KITTI 2012 and KITTI 2015 benchmarks. We split the training and evaluation sets into 160/34 images for KITTI 2012 and 160/40 images for KITTI 2015, following [60]. + +The second experiment evaluates various unsupervised + +stereo matching approaches by generating stereo images. Specifically, KITTI 2012 [10] has 194 stereo images, KITTI 2015 has 200 stereo images, ETH3D [37] has 27 stereo pairs, and the Middlebury v3 training set [36] has 15 stereo pairs at Quarter and Half resolutions (Midd-Q, Midd-H). This diverse dataset selection allows us to comprehensively assess the generalization capability of our method across + +Table 3. Comparison with other diffusion-based stereo image generation methods for unsupervised stereo matching. The first group of methods is trained on PSMNet while the second group is trained on RAFT-Stereo. SD v1.5 is used for the experiments. + +
MethodsKITTI 2012KITTI 2015
D1-all ↓EPE ↓D1-all ↓EPE ↓
Leave blank25.2193.54224.5682.937
SD-Inpainting [35]3.9070.8944.4901.059
StereoDiffusion [55]15.2132.2205.6511.154
Ours3.8020.8153.9330.991
Leave blank4.1840.9524.0740.957
SD-Inpainting [35]4.0160.9454.1280.972
StereoDiffusion [55]12.7561.8116.3601.174
Ours3.5370.9303.6530.943
+ +different domains and resolutions. + +We evaluate the performance of various unsupervised stereo matching approaches using the D1-all metric, which measures the percentage of incorrect disparity predictions (error $>3\mathrm{px}$ or $5\%$ of the true disparity), the End-PointError (EPE), and the $>n\mathrm{px}$ metric, which denotes the percentage of pixels with an EPE exceeding $n$ pixels. To demonstrate the architecture-agnostic nature of our approach, we employ three distinct stereo models: PSM-Net [6], a cost-volume-based method, RAFT-Stereo [24], an iterative refinement model, and IGEV++ [62], one of the most recent methods. For disparity estimation, we utilize the Depth Anything Model v2 (DAMv2) [65], from which we obtain the disparity map. The disparity values are normalized to the range $[0, 1]$ and subsequently scaled by a factor $\gamma$ to enable flexibility generation. + +Fine-tuning on KITTI. To ensure fair and consistent comparisons across all generation methods, we construct the training datasets using a standardized procedure. Specifically, we sample the scale factor $\gamma$ from the set $\{0.05, 0.1, 0.15, 0.2, 0.25\}$ . For each selected scale, we employ various generation methods to create the right images based on 160 left images and their corresponding scaled disparity maps for both KITTI 2012 and KITTI 2015. This procedure results in a total of 800 stereo pairs. + +As shown in Tab. 3, higher generation quality leads to improved stereo matching performance. Among the compared methods, SD-Inpainting achieves the closest performance to ours due to its pixel-level warping operation. However, it falls short in providing semantic consistency in occluded areas, which limits its overall effectiveness. + +Unsupervised Generalization. Following the methodology of constructing the Mono for Stereo (MfS) dataset in [60], we construct our training dataset, MfS-GenStereo, from the same datasets, including ADE20K [68], Mapillary Vistas [30], DIODE [54], Depth in the Wild [7], and COCO 2017 [23]. During the generation process, we randomly select a scale factor in the range $[0, 0.35]$ and set the maxi + +mum disparity to 256 for PSMNet. As shown in Tab. 4, our method enables unsupervised stereo matching using only left images, achieving superior performance. + +We compare our approach with the original MfS [60] and its enhanced variant using the Depth Anything Model v2 (DAMv2), which generates the right view images by filling the occluded areas with random background patches from the dataset. The experiments indicate that, despite the better depth prior, the synthesis of right-view images remains a critical bottleneck in this methodology. Additionally, we compare with Mono2Stereo, which fine-tunes SD-Inpainting for right-view image generation, and NeRF-Stereo, which generates stereo images by reconstructing the static scenes using Neural Radiance Fields (NeRF), limiting the diversity of its datasets. Our experimental results indicate that our proposed method consistently surpasses these existing approaches in generating high-quality stereo pairs with pixel-level accuracy, demonstrating robust generalization capabilities and significantly enhancing stereo matching performance across diverse datasets. + +# 4.3. Ablation Study. + +Random disparity drop. Since the ground-truth disparity maps in real-world datasets (e.g., KITTI) are often derived from LiDAR, resulting in sparse and incomplete point clouds, our random disparity dropout strategy mitigates the model's reliance on dense disparity supervision. This enhancement allows the model to generate plausible right images even when only sparse disparity maps are available. This strategy significantly improves performance on datasets with sparse ground-truth annotations, demonstrating better generalization to real-world scenarios. As shown in Tab. 5, comparing rows (1) and (2) highlights the effectiveness of this strategy. + +Mixed datasets. We train our model on Virtual KITTI 2 (VKITTI2) [5] and a combination of mixed datasets, and subsequently evaluate both models. When trained exclusively on VKITTI2, the model achieves comparable performance to the mixed-dataset model on KITTI 2015, owing to the relatively small domain gap between VKITTI2 and KITTI 2015. However, this single-dataset training results in notable performance degradation on other datasets, indicating limited generalization. As shown in Tab. 5, comparing rows (3) and (8) demonstrates that our mixed dataset training strategy significantly improves the model's generalization across diverse domains. + +Coordinate embedding and the warped image. Incorporating the coordinate embedding process effectively encodes spatial information and reinforces the structural relationship between stereo views. Additionally, using the warped image as a condition enhances performance by providing pixel-level guidance. The warped image serves as a coarse prediction of the right view, helping the model focus + +Table 4. Zero-Shot Generalization Benchmark. The first group shows methods trained with PSMNet, the second group shows methods trained with RAFT-Stereo, and the third group represents IGEV++. + +
MethodsKITTI 2012KITTI 2015ETH3DMiddlebury-QMiddlebury-H
D1-allEPE>3pxD1-allEPE>3pxD1-allEPE>1pxD1-allEPE>2pxD1-allEPE>2px
PSMNet [6]26.693.87727.3528.244.00328.534.2581.98915.7415.004.94120.2021.349.89933.26
MFS [60]4.3221.0094.7034.9631.6365.1812.2040.5358.1709.2171.64812.0712.343.24217.56
MfS+DAMv2 [65]4.9311.0635.3135.5041.4015.7371.9900.4938.1459.4471.82511.9012.343.67817.21
Mono2Stereo [59]4.0770.859-4.5071.052-1.5310.397-------
NeRFStereo [50]3.6830.8514.5074.8131.4485.0411.6200.55711.687.6191.45111.048.8922.59012.88
Ours + SD v1.53.6760.8464.0633.8790.9964.1121.1140.3785.5566.8871.13910.4311.332.26919.39
Ours + SD v2.13.8070.8204.1223.8980.9784.0960.8940.3244.1015.0370.8607.4638.0461.70913.73
RAFT-Stereo [24]3.9440.8304.2395.2541.1325.4680.8830.2672.6075.1571.21710.255.7741.43611.20
NeRFStereo [50]3.7870.8554.1405.3561.4695.5610.8300.3013.3705.8191.0038.3707.4261.82010.95
Ours + SD v1.53.2240.7343.4554.6831.0424.8970.9590.2942.6673.4501.0037.9524.4461.5149.94
Ours + SD v2.13.2190.7163.0193.9500.9714.2220.8350.2772.6384.0671.1279.3824.6181.52711.11
IGEV++ [62]5.6811.0996.1966.0281.2656.2391.7010.3564.6664.2380.8286.2745.2922.3847.752
Ours + SD v2.12.8020.6863.0153.7460.9524.0041.2260.3263.4293.6820.6845.5122.7970.8425.098
+ +Table 5. Ablation studies. SD v1.5 is used for the experiments. + +
Disparity Source#Mixed DatasetsRandom DropCoord EmbeddingWarped ImagePixel LossFusionMiddlebury 2014KITTI 2015
PSNRSSIMLPIPSPSNRSSIMLPIPS
Ground Truth(1)✓X✓✓✓✓23.5520.8620.064116.1640.6640.1740
(2)✓✓✓✓✓✓23.6990.8660.063520.4390.7460.1079
Pseudo Disparities(3)X✓✓✓✓✓20.0370.7450.121421.0850.7770.1113
(4)✓✓✓XXX22.0330.7960.073121.8660.7790.1002
(5)✓✓X✓XX22.6970.8320.071522.3130.7980.1011
(6)✓✓✓✓XX23.4070.8550.064422.5200.8050.0974
(7)✓✓✓✓✓X23.8110.8670.062122.7160.8110.0961
(8)✓✓✓✓✓✓23.8350.8680.062022.7490.8130.0958
+ +on correcting local inconsistencies rather than synthesizing the entire image from scratch. As shown in rows (4), (5), and (6) in Tab. 5, the combination of coordinate embedding and warped image significantly improves generation performance, preserving geometric consistency and reducing artifacts like stitching boundaries. While warped image embedding contributes most to performance, its effectiveness decreases with additional fine-tuning iterations, revealing instability when using warped image alone as a conditioning signal. This underscores the importance of combining coordinate embedding and warped image to achieve stable, high-quality stereo image generation. + +Pixel-level loss. By decoding the latent representations back to the image space, the model learns to minimize discrepancies directly in the observed image space, enforcing pixel-level alignment. As illustrated in rows (6) and (7), the introduction of the pixel-level loss significantly improves the generation performance, highlighting its role in maintaining high-fidelity stereo image generation. + +Adaptive Fusion. The proposed adaptive fusion module significantly enhances visual quality by dynamically weighting the contributions of both the generated and warped images. A comparison of rows (7) and (8) demonstrates that this fusion strategy further refines pixel alignment, resulting in sharper, more structurally coherent, and + +perceptually accurate reconstructions. The improvement in PSNR is more pronounced than that in SSIM and LPIPS, indicating that the adaptive fusion is particularly effective for pixel-level alignment. + +# 5. Conclusion + +We presented GenStereo, a novel diffusion-based framework for open-world stereo image generation with applications in unsupervised stereo matching. Our method introduces several key innovations, including disparity-aware coordinate embeddings along with warped image embeddings, pixel-level loss, and an adaptive fusion module, to ensure high-quality stereo image generation with strong geometric and semantic consistency. Through extensive experiments across diverse datasets, we show that GenStereo significantly outperforms existing methods in both stereo generation and unsupervised matching. Our ablation studies demonstrate the efficacy of each proposed component. + +Limitations. Despite promising results, diffusion-based models face inherent challenges with large disparities when generating right-view images due to large unconditioned regions. While our data augmentation such as random cropping and resizing mitigates these issues and performs well in typical stereo setups, future work could explore extending these models to accommodate larger disparities. + +# Acknowledgments + +We gratefully acknowledge the advanced computational resources provided by Engineering IT and Research Infrastructure Services at Washington University in St. Louis. + +# References + +[1] Filippo Aleotti, Fabio Tosi, Li Zhang, Matteo Poggi, and Stefano Mattoccia. Reversing the cycle: self-supervised deep stereo through enhanced monocular distillation. In European Conference on Computer Vision, pages 614-632. Springer, 2020. 3 +[2] Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences, 63:1-11, 2020. 1 +[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1 +[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In European Conf. on Computer Vision, pages 611-625. Springer-Verlag, 2012. 4 +[5] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv preprint arXiv:2001.10773, 2020. 1, 4, 7 +[6] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5410-5418, 2018. 7, 8 +[7] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 7 +[8] Cheng Chi, Qingjie Wang, Tianyu Hao, Peng Guo, and Xin Yang. Feature-level collaboration: Joint unsupervised learning of optical flow, stereo depth and camera motion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2463-2473, 2021. 3 +[9] Jaehoon Cho, Dongbo Min, Youngjung Kim, and Kwanghoon Sohn. Deep monocular depth estimation leveraging a large-scale outdoor stereo dataset. Expert Systems with Applications, 178:114877, 2021. 1 +[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 6 +[11] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 3 +[12] Peter Hedman, Suhib Alsisan, Richard Szeliski, and Johannes Kopf. Casual 3d photography. ACM Transactions on Graphics (TOG), 36(6):1-15, 2017. 2 +[13] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 2 + +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[15] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2 +[16] Junpeng Jing, Ye Mao, and Krystian Mikolajczyk. Match-stereo-videos: Bidirectional alignment for consistent dynamic stereo matching. In European Conference on Computer Vision, pages 415–432. Springer, 2024. 4 +[17] Laurent Valentin Jospin, Hamid Laga, Farid Boussaid, and Mohammed Bennamoun. Active-passive simstereo, 2022. 4 +[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 4 +[19] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2 +[20] Johannes Kopf, Fabian Langguth, Daniel Scharstein, Richard Szeliski, and Michael Goesele. Image-based rendering in the gradient domain. ACM Transactions on Graphics (TOG), 32(6):1-9, 2013. 2 +[21] Hsueh-Ying Lai, Yi-Hsuan Tsai, and Wei-Chen Chiu. Bridging stereo matching and optical flow via spatiotemporal correspondence. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1890-1899, 2019. 3 +[22] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16263-16272, 2022. 4, 5 +[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740-755. Springer, 2014. 2, 7 +[24] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In IEEE International Conference on 3D Vision, pages 218-227, 2021. 7, 8 +[25] Rui Liu, Chengxi Yang, Wenxiu Sun, Xiaogang Wang, and Hongsheng Li. Stereogan: Bridging synthetic-to-real domain gap by joint optimization of domain translation and stereo matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12757-12766, 2020. 3 +[26] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 5, 6 + +[27] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 4 +[28] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3061-3070, 2015. 1, 5 +[29] Jiteng Mu, Shalini De Mello, Zhiding Yu, Nuno Vasconcelos, Xiaolong Wang, Jan Kautz, and Sifei Liu. Coordgan: Self-supervised dense correspondences emerge from gans. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10011-10020, 2022. 3 +[30] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In IEEE international conference on computer vision, pages 4990-4999, 2017. 7 +[31] Matteo Poggi, Alessio Tonioni, Fabio Tosi, Stefano Mattoccia, and Luigi Di Stefano. Continual adaptation for deep stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4713-4729, 2021. 3 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[33] Pierluigi Zama Ramirez, Alex Costanzino, Fabio Tosi, Matteo Poggi, Samuele Salti, Stefano Mattoccia, and Luigi Di Stefano. Booster: a benchmark for depth from images of specular and transparent surfaces. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 1, 4 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2, 3, 5, 6, 7 +[36] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nesic, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 31-42. Springer, 2014. 5, 6 +[37] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3260-3269, 2017. 6 +[38] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image + +to novel views with semantic-preserving generative warping. Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3 +[39] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2 +[40] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020. 2, 6 +[41] J. Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. International Conference on Learning Representations, 2021. 2 +[42] Xiao Song, Guorun Yang, Xinge Zhu, Hui Zhou, Zhe Wang, and Jianping Shi. Adastereo: A simple and efficient approach for adaptive stereo matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10328-10337, 2021. 3 +[43] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 3 +[44] Joshua Tokarsky, Ibrahim Abdulhafiz, Satya Ayyalasomayajula, Mostafa Mohsen, Navya G. Rao, and Adam Forbes. PLT-D3: A High-fidelity Dynamic Driving Simulation Dataset for Stereo Depth and Scene Flow, 2024. 4 +[45] Alessio Tonioni, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Unsupervised adaptation for deep stereo. In IEEE International Conference on Computer Vision, pages 1605-1613, 2017. 3 +[46] Alessio Tonioni, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Unsupervised domain adaptation for depth prediction from images. IEEE transactions on pattern analysis and machine intelligence, 42(10):2396-2409, 2019. 3 +[47] Alessio Tonioni, Oscar Rahnama, Thomas Joy, Luigi Di Stefano, Thalaiyasingam Ajanthan, and Philip HS Torr. Learning to adapt for stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9661-9670, 2019. 3 +[48] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 195-204, 2019. 3 +[49] Fabio Tosi, Yiyi Liao, Carolin Schmitt, and Andreas Geiger. Smd-nets: Stereo mixture density networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8942-8952, 2021. 4 +[50] Fabio Tosi, Alessio Tonioni, Daniele De Gregorio, and Matteo Poggi. Nerf-supervised deep stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 855-866, 2023. 2, 3, 8 +[51] Fabio Tosi, Luca Bartolomei, and Matteo Poggi. A survey on deep stereo matching in the twenties. International Journal of Computer Vision, 2025. 4 +[52] Jonathan Tremblay, Thang To, and Stan Birchfield. Falling things: A synthetic dataset for 3d object detection and pose + +estimation. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2038-2041, 2018. 4 +[53] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, JiaBin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 1 +[54] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z. Dai, Andrea F. Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R. Walter, and Gregory Shakhnarovich. DIODE: A Dense Indoor and Outdoor DEpth Dataset. CoRR, abs/1908.00463, 2019. 7 +[55] Lezhong Wang, Jeppe Revall Frisvad, Mark Bo Jensen, and Siavash Arjomand Bigdeli. Stereodiffusion: Training-free stereo image generation using latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7416-7425, 2024. 1, 2, 3, 5, 6, 7 +[56] Qiang Wang, Shizhen Zheng, Qingsong Yan, Fei Deng, Kaiyong Zhao, and Xiaowen Chu. IRS: A large naturalistic indoor robotics stereo dataset to train deep models for disparity and surface normal estimation. In IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2021. 1, 4 +[57] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4909-4916. IEEE, 2020. 1, 4 +[58] Yang Wang, Peng Wang, Zhenheng Yang, Chenxu Luo, Yi Yang, and Wei Xu. Unos: Unified unsupervised optical-flow and stereo-depth estimation by watching videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8071-8081, 2019. 3 +[59] Yuran Wang, Yingping Liang, Hesong Li, and Ying Fu. Mono2stereo: Monocular knowledge transfer for enhanced stereo matching. arXiv preprint arXiv:2411.09151, 2024. 3, 8 +[60] Jamie Watson, Oisin Mac Aodha, Daniyar Turmukhambetov, Gabriel J Brostow, and Michael Firman. Learning stereo from single images. In European Conference on Computer Vision, pages 722-740. Springer, 2020. 2, 3, 6, 7, 8 +[61] Zhexiao Xiong, Feng Qiao, Yu Zhang, and Nathan Jacobs. Stereoflowgan: Co-training for stereo and flow with unsupervised domain adaptation. In 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023. BMVA, 2023. 3 +[62] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: Iterative multirange geometry encoding volumes for stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 7, 8 +[63] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 2 + +[64] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 4 +[65] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875-21911, 2024. 1, 7, 8 +[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2 +[67] Yiran Zhong, Yuchao Dai, and Hongdong Li. Self-supervised learning for stereo matching with self-improving ability. arXiv preprint arXiv:1709.00930, 2017. 3 +[68] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In IEEE Conference on Computer Vision and Pattern Recognition, pages 633-641, 2017. 7 \ No newline at end of file diff --git a/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/images.zip b/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8c3f92fac0874d8c33f0bdbd5ce3b03ab4760c7f --- /dev/null +++ b/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:73566e9a52f2df1b48b70c52b88a9e36e60c74c29fab4da11083760172576c4a +size 845540 diff --git a/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/layout.json b/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ad81f7f5585342454656f6608226768e90a23cf1 --- /dev/null +++ b/2025/Towards Open-World Generation of Stereo Images and Unsupervised Matching/layout.json @@ -0,0 +1,9532 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "text", + "content": "Towards Open-World Generation of Stereo Images and Unsupervised Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 161, + 143, + 449, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 143, + 449, + 172 + ], + "spans": [ + { + "bbox": [ + 161, + 143, + 449, + 172 + ], + "type": "text", + "content": "Feng Qiao Zhexiao Xiong Eric Xing Nathan Jacobs Washington University in St. Louis" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 182, + 553, + 319 + ], + "blocks": [ + { + "bbox": [ + 56, + 182, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 56, + 182, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 553, + 319 + ], + "type": "image", + "image_path": "e85e509e88f5dec64869b1deea24ad1c82ada4f7de0182661ef05b5422c41462.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 320, + 490, + 331 + ], + "lines": [ + { + "bbox": [ + 118, + 320, + 490, + 331 + ], + "spans": [ + { + "bbox": [ + 118, + 320, + 490, + 331 + ], + "type": "text", + "content": "Figure 1. Comparison with diffusion-based methods for stereo image generation on the COCO dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 341, + 200, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 341, + 200, + 354 + ], + "spans": [ + { + "bbox": [ + 151, + 341, + 200, + 354 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 368, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 368, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 53, + 368, + 296, + 632 + ], + "type": "text", + "content": "Stereo images are fundamental to numerous applications, including extended reality (XR) devices, autonomous driving, and robotics. Unfortunately, acquiring high-quality stereo images remains challenging due to the precise calibration requirements of dual-camera setups and the complexity of obtaining accurate, dense disparity maps. Existing stereo image generation methods typically focus on either visual quality for viewing or geometric accuracy for matching, but not both. We introduce GenStereo, a diffusion-based approach, to bridge this gap. The method includes two primary innovations (1) conditioning the diffusion process on a disparity-aware coordinate embedding and a warped input image, allowing for more precise stereo alignment than previous methods, and (2) an adaptive fusion mechanism that intelligently combines the diffusion-generated image with a warped image, improving both realism and disparity consistency. Through extensive training on 11 diverse stereo datasets, GenStereo demonstrates strong generalization ability. GenStereo achieves state-of-the-art performance in both stereo image generation and unsupervised stereo matching tasks. Project page is available at https://qjizhi.github.io/genstereo." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 656, + 135, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 135, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 135, + 669 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "The demand for stereo images continues growing with XR devices, autonomous driving, and robotics. However, acquiring high-quality stereo images remains chal" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 498 + ], + "type": "text", + "content": "lenging due to complex camera calibration and environmental constraints, limiting the development of robust stereo matching models that generalize well across diverse scenarios. Real-world datasets often provide only sparse disparity annotations [28] or lack accurate ground truth altogether [9]. Although some datasets offer dense, accurate disparity maps [2, 33], they are typically limited to specific scenarios like indoor scenes. Moreover, capturing real-world stereo data requires complex sensor setups with precise calibration, often constraining baseline distances and scene diversity. Synthetic datasets, while providing precise disparity maps, suffer from domain gaps compared to real-world scenarios [5, 56, 57]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 714 + ], + "type": "text", + "content": "Recent advancements in monocular depth estimation (MDE) models [3, 34, 64, 65] have provided increasingly accurate dense disparity maps, playing a crucial role in image generation from single views. Meanwhile, Text-to-Image (T2I) diffusion models, such as Stable Diffusion [35], have made remarkable strides in generating diverse and high-quality images from user-provided text prompts. However, these models often struggle to maintain spatial consistency when altering the viewpoint of the generated images. While there has been some research on 3D novel view generation [38, 53], there remains a gap in utilizing image generation models to directly produce stereo images. StereoDiffusion [55] is the first approach to leverage diffusion-based generation for stereo image generation, but it struggles with pixel-level accuracy, as it applies disparity shifts in the latent space and fills occluded regions with blurry pixels that lack meaningful semantic content. On the other hand, Stable Diffusion Inpainting (SD" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26579" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "Inpainting) [35], though not designed for stereo image generation, faces similar challenges by incorrectly filling occluded areas with semantically inappropriate pixels. Fig. 1 shows that previous diffusion-based methods and our methods tested on the COCO dataset [23]. The scale factor " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " is set to 0.15, meaning that the maximum disparity is " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " of the image width." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 294, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 294, + 289 + ], + "type": "text", + "content": "We present GenStereo, a novel framework that addresses both visual quality and geometric accuracy in stereo image generation. Existing approaches rely on either warping-based methods [60], which provide geometric accuracy but struggle with semantic consistency in occluded regions, or diffusion-based approaches [55], which maintain better semantic coherence but lack precise geometric control. GenStereo bridges this gap through a carefully designed two-stream architecture inspired by recent advances in human animation [15] and view synthesis [38], achieving both geometric precision and semantic consistency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 290, + 294, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 294, + 445 + ], + "type": "text", + "content": "Our framework employs a multi-level constraint system for geometric precision and semantic coherence, starting with disparity-aware coordinate embedding that provides implicit geometric guidance. This is followed by a cross-view attention mechanism that enables semantic feature alignment between views, ensuring consistency in challenging areas like occlusions and complex textures. The framework leverages a dual-space supervision strategy that operates in both latent and pixel spaces, with an adaptive fusion mechanism that further optimizes pixel-level accuracy. To ensure robust generalization across diverse real-world scenarios, we train our model on a diverse dataset combining 11 stereo datasets with varying scenes and baselines." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 446, + 294, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 294, + 589 + ], + "type": "text", + "content": "Furthermore, we demonstrate how our high-quality stereo generation framework significantly improves unsupervised stereo matching learning. Previous approaches to unsupervised learning have been limited by either simplified warping and random background filling [60] or constraints to small-scale static scenes [50]. In contrast, our method enables large-scale training with diverse, photorealistic stereo images that maintain both geometric accuracy and semantic consistency. This advancement represents a significant step toward bridging the gap between supervised and unsupervised stereo matching approaches. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 594, + 294, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 294, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 294, + 641 + ], + "type": "text", + "content": "- We propose GenStereo, the first unified framework for open-world stereo image generation that addresses visual quality and geometric accuracy, enabling both practical applications and unsupervised stereo matching." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 294, + 713 + ], + "type": "text", + "content": "- We introduce a comprehensive multi-level constraint system that combines: (1) disparity-aware coordinate embedding with warped image conditioning for geometric guidance, (2) cross-view attention mechanism for semantic feature alignment, and (3) dual-space supervision with adaptive fusion for pixel-accuracy generation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 108 + ], + "type": "text", + "content": "- Extensive experimental validation demonstrating state-of-the-art performance in both stereo image generation and unsupervised stereo matching." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 120, + 400, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 120, + 400, + 133 + ], + "spans": [ + { + "bbox": [ + 314, + 120, + 400, + 133 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 140, + 474, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 140, + 474, + 153 + ], + "spans": [ + { + "bbox": [ + 314, + 140, + 474, + 153 + ], + "type": "text", + "content": "2.1. Conditional Diffusion Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 159, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 553, + 350 + ], + "type": "text", + "content": "Diffusion models began with DDPM [14], advancing from slow, probabilistic generation to faster sampling with DDIM [41]. Stable Diffusion [35], a pioneering T2I model, further improved efficiency by performing diffusion in a compact latent space, combining it with CLIP-based [32] text conditioning for versatile T2I generation. However, Stable Diffusion lacks fine-grained control over specific visual attributes, relying mainly on text prompts without structured guidance. To address these limitations, conditional diffusion models have emerged. ControlNet [66] is a key advancement, introducing conditioning mechanisms for structural inputs (such as edges, depth maps, or poses) that maintain flexibility while enabling precise, user-driven synthesis. However, ControlNet still faces challenges in achieving pixel-level accuracy in complex images, such as stereo image generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 350, + 553, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 553, + 578 + ], + "type": "text", + "content": "Recent research [15, 19, 38, 39, 63] has explored the self-attention properties within T2I models. Text2VideoZero [19] and MVDream [39] generate consistent visuals across video frames or 3D multi-views by sharing self-attention, whileAnimate-Anyone [15] and MagicAnimate [63] apply a similar approach to produce human dance videos through fine-tuned T2I models. GenWarp [38] employs augmenting self-attention with cross-view attention between the reference and target views to generate novel images. This two-stream architecture augments the features of the denoising net. Inspired by the adaptability and control of these self-attention-based architectures, our method builds on their strengths. However, their direct application to stereo image generation faces fundamental challenges. First, they typically don't incorporate disparity information, which is crucial for stereoscopic consistency. Second, achieving the pixel-level accuracy required for comfortable stereo viewing demands specialized architectural considerations beyond existing frameworks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 587, + 527, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 587, + 527, + 601 + ], + "spans": [ + { + "bbox": [ + 314, + 587, + 527, + 601 + ], + "type": "text", + "content": "2.2. Stereo Image Generation and Inpainting" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 700 + ], + "type": "text", + "content": "Traditional view synthesis approaches, including geometry-based reconstruction [12, 13, 20] and recent NeRF-based methods [50], excel at rendering novel views but are constrained by their requirement for multiple input images of static scenes. While 3D Photography techniques [12, 40] attempt to overcome this limitation through depth-based mesh projection, they often struggle with complex occlusion handling and require sophisticated post-processing pipelines." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "The evolution of stereo generation methods has pro" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26580" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 323 + ], + "type": "text", + "content": "gressed from simple warping-based approaches to more sophisticated diffusion-based solutions. MfS [60] introduced a basic framework that attempted to handle occlusions by sampling random background patches from the dataset, but this led to semantic inconsistencies and visible artifacts at occlusion areas. While SD-Inpainting [35] introduced more coherent semantic content through its prior learning, it failed to maintain local consistency between inpainted regions and their surrounding context, often producing visible discontinuities in texture and structure. Mono2Stereo [59] made progress by fine-tuning SD-Inpainting specifically for stereo generation, enabling unsupervised stereo matching, though visible artifacts persist in in-painted regions. Stereodiffusion [55] marked a significant shift by introducing end-to-end diffusion-based stereo image generation, but its latent-space warping approach without explicit geometric constraints led to compromised pixel-level accuracy. Despite these advances, the challenge of generating photorealistic stereo images that maintain both visual quality and geometric fidelity remains largely unsolved, particularly for diverse real-world scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 331, + 222, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 222, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 222, + 342 + ], + "type": "text", + "content": "2.3. Unsupervised Stereo Matching" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 348, + 296, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 503 + ], + "type": "text", + "content": "Traditional unsupervised stereo matching approaches primarily rely on photometric consistency. Methods like [11, 47, 48, 67] leverage photometric losses across stereo images, while others [8, 21, 58] extend this to temporal sequences. A parallel line of research explores proxy supervision strategies, either through carefully designed algorithmic supervisors [31, 45, 46] or knowledge distillation from pre-trained networks [1]. Recent domain adaptation approaches [25, 42, 61] attempt to bridge the gap between synthetic and real-world domains or facilitate cross-dataset adaptation. However, these methods often exhibit limited generalization capability beyond their target domains [1], particularly struggling with diverse real-world scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 504, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 296, + 647 + ], + "type": "text", + "content": "More recent approaches have explored novel directions for stereo generation and matching. MfS [60] pioneered unsupervised stereo matching using MDE as guidance, but it relies on random background sampling for occlusion filling, leading to semantic inconsistencies. While NeRF-Stereo [50] achieves high-quality results through neural radiance fields, it requires multiple views of static scenes, limiting its practical applications. Mono2Stereo [59] leverages Stable Diffusion for inpainting occluded regions, representing a significant advance in generative stereo synthesis, though it still exhibits artifacts in occluded areas where geometric and semantic consistency is crucial." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 657, + 117, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 117, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 117, + 670 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Given a left view image " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{l}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " and its corresponding disparity map " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{D}_{l}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " predicted by an MDE model, our goal is to generate a high-quality right view image " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{I}_r" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " that maintains both vi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "sual quality and geometric consistency. As shown in Fig. 2, our GenStereo framework processes the left view as the reference image and synthesizes the right view as the target image. The generated triplet " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\langle I_l,\\hat{D}_l,\\hat{I}_r\\rangle" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " not only serves as a stereo pair for visualization but also provides training data for unsupervised stereo matching." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 150, + 528, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 528, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 528, + 163 + ], + "type": "text", + "content": "3.1. Disparity-Aware Coordinate Embedding" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 167, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 554, + 262 + ], + "type": "text", + "content": "Traditional inpainting-based stereo image generation methods often suffer from visible boundaries between warped and inpainted regions. To address this limitation, we propose a disparity-aware coordinate embedding scheme inspired by recent advances in coordinate-based generation [29, 38]. Our approach utilizes dual coordinate embeddings: a canonical embedding for the left view and its warped counterpart for the right view." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": "Specifically, we first construct a canonical 2D coordinate map " + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{h \\times w \\times 2}" + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": " with values normalized to " + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": ". This map is transformed into Fourier features [43] through a positional encoding function " + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 263, + 554, + 311 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 404, + 318, + 553, + 331 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 318, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 404, + 318, + 553, + 331 + ], + "type": "interline_equation", + "content": "C _ {l} = \\phi (X) \\tag {1}", + "image_path": "ec0d8ee35941ec6d0cc4b597daecf399f7f00e93fd9135f73f352bbf32bc9a06.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 337, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 554, + 384 + ], + "type": "text", + "content": "The resulting Fourier feature map " + }, + { + "bbox": [ + 313, + 337, + 554, + 384 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 313, + 337, + 554, + 384 + ], + "type": "text", + "content": " serves as the coordinate embedding for the left view " + }, + { + "bbox": [ + 313, + 337, + 554, + 384 + ], + "type": "inline_equation", + "content": "I_l" + }, + { + "bbox": [ + 313, + 337, + 554, + 384 + ], + "type": "text", + "content": ". We then generate the right view embedding by warping " + }, + { + "bbox": [ + 313, + 337, + 554, + 384 + ], + "type": "inline_equation", + "content": "C_l" + }, + { + "bbox": [ + 313, + 337, + 554, + 384 + ], + "type": "text", + "content": " according to the disparity map:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 392, + 385, + 553, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 385, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 392, + 385, + 553, + 397 + ], + "type": "interline_equation", + "content": "C _ {r} = \\operatorname {w a r p} \\left(C _ {l}, D _ {l}\\right), \\tag {2}", + "image_path": "56041b9c0e316a2a3128e3a0d17ab1311dc3aa38c0ec0b64ba3ceea21a156320.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "inline_equation", + "content": "D_{l}" + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "text", + "content": " represents ground-truth disparity during training and predicted disparity " + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "inline_equation", + "content": "\\hat{D}_l" + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "text", + "content": " during inference. These coordinate embeddings " + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "inline_equation", + "content": "(C_l" + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "inline_equation", + "content": "C_r)" + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "text", + "content": " are integrated into their respective view features " + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "inline_equation", + "content": "(F_{l}" + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "inline_equation", + "content": "F_{r})" + }, + { + "bbox": [ + 313, + 401, + 554, + 472 + ], + "type": "text", + "content": " through convolutional layers, establishing strong geometric correspondence while maintaining visual consistency between views." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 473, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 534 + ], + "type": "text", + "content": "Our approach differs from GenWarp [38] by utilizing disparity maps instead of camera matrices for warping, enabling more precise pixel-level control. We further enhance geometric consistency by incorporating the warped image " + }, + { + "bbox": [ + 313, + 473, + 554, + 534 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{warp}}" + }, + { + "bbox": [ + 313, + 473, + 554, + 534 + ], + "type": "text", + "content": " as additional conditioning for the denoising U-Net." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 539, + 497, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 497, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 497, + 551 + ], + "type": "text", + "content": "3.2. Cross-View Feature Enhancement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 556, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 713 + ], + "type": "text", + "content": "To facilitate effective information exchange between views, we adopt a two-parallel U-Nets framework pretrained from Stable Diffusion, inspired by GenWarp [38]. However, our work differs in its novel conditioning strategy: the reference U-Net takes concatenated " + }, + { + "bbox": [ + 313, + 556, + 555, + 713 + ], + "type": "inline_equation", + "content": "(I_l, C_l)" + }, + { + "bbox": [ + 313, + 556, + 555, + 713 + ], + "type": "text", + "content": " as conditions to process the left view, while the denoising U-Net takes concatenated " + }, + { + "bbox": [ + 313, + 556, + 555, + 713 + ], + "type": "inline_equation", + "content": "(I_{\\mathrm{warp}}, C_r)" + }, + { + "bbox": [ + 313, + 556, + 555, + 713 + ], + "type": "text", + "content": " as conditions when synthesizing the right view " + }, + { + "bbox": [ + 313, + 556, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{I}_r" + }, + { + "bbox": [ + 313, + 556, + 555, + 713 + ], + "type": "text", + "content": ". This design ensures that each U-Net receives both image content and corresponding coordinate information for better feature extraction. As our model is conditioned on a reference image and its disparity map, we replace text condition of SD with image embedding of the left image through a CLIP image encoder in both U-Nets." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26581" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 558, + 233 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 558, + 233 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 558, + 233 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 558, + 233 + ], + "type": "image", + "image_path": "5d1ad429e007e8b74d4ad5dc5ef20bfa88ab62ebbe8e293b87f95f9403d1bf40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 240, + 394, + 251 + ], + "lines": [ + { + "bbox": [ + 214, + 240, + 394, + 251 + ], + "spans": [ + { + "bbox": [ + 214, + 240, + 394, + 251 + ], + "type": "text", + "content": "Figure 2. Overview of the GenStereo framework." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 273, + 296, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 320 + ], + "type": "text", + "content": "To incorporate left-view information during right-view generation, we concatenate the reference features with the target features in the attention mechanism. Specifically, we compute cross-view attention as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 326, + 294, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 326, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 94, + 326, + 294, + 339 + ], + "type": "interline_equation", + "content": "q = F _ {r}, \\quad k = \\left[ F _ {l}, F _ {r} \\right], \\quad v = \\left[ F _ {l}, F _ {r} \\right]. \\tag {3}", + "image_path": "84d93b2015f2c8b43d01bd67991020e4918ec582303808ed30e3e0c538c472b7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 343, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 428 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 343, + 296, + 428 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 55, + 343, + 296, + 428 + ], + "type": "text", + "content": " is derived from the reference U-Net conditioned on " + }, + { + "bbox": [ + 55, + 343, + 296, + 428 + ], + "type": "inline_equation", + "content": "(I_l,C_l)" + }, + { + "bbox": [ + 55, + 343, + 296, + 428 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 55, + 343, + 296, + 428 + ], + "type": "inline_equation", + "content": "F_{r}" + }, + { + "bbox": [ + 55, + 343, + 296, + 428 + ], + "type": "text", + "content": " comes from the denoising U-Net conditioned on " + }, + { + "bbox": [ + 55, + 343, + 296, + 428 + ], + "type": "inline_equation", + "content": "(I_{\\mathrm{warp}},C_r)" + }, + { + "bbox": [ + 55, + 343, + 296, + 428 + ], + "type": "text", + "content": ". This dual-stream attention mechanism allows the model to adaptively balance between semantic consistency from the reference view and geometric accuracy from the warped view, with each stream guided by both image content and coordinate information." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 162, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 162, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 162, + 445 + ], + "type": "text", + "content": "3.3. Training Strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 594 + ], + "type": "text", + "content": "Mixed training [22, 34, 64] has been proven to be an effective training strategy for domain generalization in both monocular and stereo matching depth estimation. We follow the datasets summarized in [51] and utilize the most widely available public synthetic datasets. Based on our observations, real-world datasets can negatively impact performance, even when there are only minor calibration errors or slight differences in imaging between the two cameras. Tab. 1 shows the comprehensive list of datasets utilized. Datasets containing abstract images are not included, because our model aims to have a good generalization capability in the real world." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 667 + ], + "type": "text", + "content": "To address dataset imbalance, we employ a resampling strategy. Specifically, we sample the smaller datasets multiple times until the number reaches " + }, + { + "bbox": [ + 55, + 594, + 296, + 667 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 667 + ], + "type": "text", + "content": " of the largest dataset. As the Stable Diffusion model only accepts images of size " + }, + { + "bbox": [ + 55, + 594, + 296, + 667 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 594, + 296, + 667 + ], + "type": "text", + "content": " for SD v1.5 and " + }, + { + "bbox": [ + 55, + 594, + 296, + 667 + ], + "type": "inline_equation", + "content": "768 \\times 768" + }, + { + "bbox": [ + 55, + 594, + 296, + 667 + ], + "type": "text", + "content": " for SD v2.1, we apply a random square crop and resize." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 185, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 185, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 185, + 685 + ], + "type": "text", + "content": "3.4. Pixel Space Alignment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Given an RGB image " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ", Latent Diffusion Models (LDMs) first encode it into a latent representation" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 328, + 302, + 539, + 411 + ], + "blocks": [ + { + "bbox": [ + 313, + 271, + 554, + 292 + ], + "lines": [ + { + "bbox": [ + 313, + 271, + 554, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 554, + 292 + ], + "type": "text", + "content": "Table 1. Our training data sources. All datasets are synthetic with the size of 684K." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 302, + 539, + 411 + ], + "lines": [ + { + "bbox": [ + 328, + 302, + 539, + 411 + ], + "spans": [ + { + "bbox": [ + 328, + 302, + 539, + 411 + ], + "type": "table", + "html": "
DatasetIndoorOutdoor#ImagesYear
TartanAir [57]✓✓306K2020
Dynamic Replica [18]✓145K2023
IRS [56]✓103K2021
Falling Things [52]✓✓61K2018
VKITTI2 [5]✓21K2020
InfinigenSV [16]✓17K2024
SimStereo [17]✓14K2022
UnrealStereo4K [49]✓✓8K2021
Spring [27]✓5K2023
PLT-D3 [44]✓3K2024
Sintel [4]✓1K2012
", + "image_path": "8e07f9a70d5f917486b75dabb1c974c99b00581d5a0cd628d8e20cb1bda4beb1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 429, + 554, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 554, + 465 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 554, + 465 + ], + "type": "inline_equation", + "content": "z = \\mathcal{E}(I)" + }, + { + "bbox": [ + 313, + 429, + 554, + 465 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 429, + 554, + 465 + ], + "type": "inline_equation", + "content": "z\\in \\mathbb{R}^{h\\times w\\times c}" + }, + { + "bbox": [ + 313, + 429, + 554, + 465 + ], + "type": "text", + "content": ". While this latent space operation significantly reduces computational costs, it may compromise pixel-level accuracy during image generation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 465, + 542, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 465, + 542, + 477 + ], + "spans": [ + { + "bbox": [ + 326, + 465, + 542, + 477 + ], + "type": "text", + "content": "The standard LDM objective operates in latent space:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 362, + 483, + 553, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 483, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 362, + 483, + 553, + 498 + ], + "type": "interline_equation", + "content": "L _ {\\text {l a t e n t}} := \\mathbb {E} _ {z, \\epsilon , t} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {4}", + "image_path": "0fe8403f8a4c849b2b8e0f7f4c6993635c3bdf4d06bfd711f68471b0cbded8e5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "type": "text", + "content": " is the noisy latent at timestep " + }, + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "type": "text", + "content": " is the noise added to create " + }, + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_{t}, t)" + }, + { + "bbox": [ + 313, + 504, + 554, + 551 + ], + "type": "text", + "content": " is the network's prediction of this noise. The model learns to denoise by predicting the noise that was added." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 588 + ], + "type": "text", + "content": "To maintain pixel-level accuracy, we introduce an additional pixel-space loss by decoding both the predicted and target latent variables:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 348, + 594, + 553, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 594, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 348, + 594, + 553, + 609 + ], + "type": "interline_equation", + "content": "L _ {\\text {p i x e l}} = \\mathbb {E} _ {z, \\epsilon , t} \\left[ \\| \\mathcal {D} \\left(z _ {\\text {p r e d}}\\right) - \\mathcal {D} \\left(z _ {\\text {t a r g e t}}\\right) \\| _ {2} ^ {2} \\right] \\tag {5}", + "image_path": "6fe488a52b3533f19fdea539c6e34f3283782368c229eb45f86cb6c38fe4f53d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 615, + 485, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 485, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 485, + 627 + ], + "type": "text", + "content": "The final objective combines both spaces:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 386, + 634, + 553, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 634, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 386, + 634, + 553, + 647 + ], + "type": "interline_equation", + "content": "L = L _ {\\text {l a t e n t}} + \\alpha L _ {\\text {p i x e l}} \\tag {6}", + "image_path": "d95c439945587af4d58260152d9fb8cd0366623f9457ba1f23a83cee88a888de.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 545, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 313, + 654, + 545, + 666 + ], + "type": "text", + "content": " balances latent and pixel-space supervision." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 672, + 454, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 454, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 454, + 685 + ], + "type": "text", + "content": "3.5. Adaptive Fusion Module" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "To achieve seamless integration between generated and warped content, we propose an Adaptive Fusion Module" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26582" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "that learns to combine " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "I_{gen}" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "I_{warp}" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": " based on local context and confidence. The module predicts spatially-varying fusion weights through a lightweight network:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 95, + 118, + 295, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 118, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 95, + 118, + 295, + 133 + ], + "type": "interline_equation", + "content": "W = \\sigma \\left(f _ {\\theta} \\left(\\operatorname {c o n c a t} \\left(I _ {g e n}, I _ {w a r p}, M\\right)\\right)\\right), \\tag {7}", + "image_path": "92b78a5733965abe79863f70c040a96557cb0a31c93692a9afbc89f2c948f4d5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 296, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 296, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 296, + 176 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 140, + 296, + 176 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 140, + 296, + 176 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 55, + 140, + 296, + 176 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 140, + 296, + 176 + ], + "type": "text", + "content": " convolutional layer and " + }, + { + "bbox": [ + 55, + 140, + 296, + 176 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 140, + 296, + 176 + ], + "type": "text", + "content": " is the sigmoid activation, ensuring weights in [0, 1]. The final right view is computed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 183, + 295, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 183, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 70, + 183, + 295, + 199 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {r} = M \\odot W \\odot I _ {\\text {w a r p}} + (1 - M \\odot W) \\odot I _ {\\text {g e n}}, \\tag {8}", + "image_path": "48ef6ba40f79d9acdf1ecb8bcdb958a8b8b2e7b17412701021becdb84664fb10.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 207, + 296, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 278 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 207, + 296, + 278 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 207, + 296, + 278 + ], + "type": "text", + "content": " denotes element-wise multiplication. This formulation adaptively favors warped content in high-confidence regions (" + }, + { + "bbox": [ + 55, + 207, + 296, + 278 + ], + "type": "inline_equation", + "content": "M \\approx 1" + }, + { + "bbox": [ + 55, + 207, + 296, + 278 + ], + "type": "text", + "content": ") while relying on generated content in occluded or uncertain areas. The learned weights " + }, + { + "bbox": [ + 55, + 207, + 296, + 278 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 207, + 296, + 278 + ], + "type": "text", + "content": " enable smooth transitions between warped and generated regions, ensuring both geometric accuracy and visual consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 286, + 208, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 208, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 208, + 300 + ], + "type": "text", + "content": "3.6. Random Disparity Dropout" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 304, + 296, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 352 + ], + "type": "text", + "content": "To simulate the sparsity of real-world disparity maps like KITTI's LiDAR-derived ground truth, we randomly apply disparity dropout to " + }, + { + "bbox": [ + 55, + 304, + 296, + 352 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 304, + 296, + 352 + ], + "type": "text", + "content": " of training samples. For each selected sample, we first generate a dropout ratio:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 136, + 361, + 295, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 361, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 136, + 361, + 295, + 374 + ], + "type": "interline_equation", + "content": "r \\sim \\operatorname {U n i f o r m} (0, 1) \\tag {9}", + "image_path": "c173a426c5c6f5ce4cf3bcfee613a91cf3e4eac1471f1b23591818886894b0db.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "text", + "content": "Using this ratio, we create a binary mask where each pixel has a probability " + }, + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "text", + "content": " of being dropped:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 97, + 415, + 295, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 415, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 97, + 415, + 295, + 449 + ], + "type": "interline_equation", + "content": "M _ {\\text {r a n d}} (i, j) = \\left\\{ \\begin{array}{l l} 1 & \\text {o t h e r w i s e} \\\\ 0 & \\text {w i t h p r o b a b i l i t y} r \\end{array} \\right. \\tag {10}", + "image_path": "8bf5f77577425d50d01deb1815aaed528d661c4e273e088178adbcae06598eab.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 455, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 479 + ], + "type": "text", + "content": "The final mask combines this random dropout with the warping mask:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 133, + 488, + 295, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 488, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 133, + 488, + 295, + 502 + ], + "type": "interline_equation", + "content": "M = M _ {\\text {w a r p}} \\vee M _ {\\text {r a n d}} \\tag {11}", + "image_path": "9b159fc1240c6f6c3158932c450331bebc0baa487416831b1aa00323f29ed8b8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 510, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 559 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 296, + 559 + ], + "type": "inline_equation", + "content": "\\vee" + }, + { + "bbox": [ + 55, + 510, + 296, + 559 + ], + "type": "text", + "content": " denotes element-wise logical OR. This strategy encourages the model to handle sparse disparity inputs, improving its robustness and generalization to real-world scenarios where dense disparity maps may not be available." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 569, + 137, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 137, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 137, + 582 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 588, + 195, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 195, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 195, + 601 + ], + "type": "text", + "content": "4.1. Stereo Image Generation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Experimental Setup. We fine-tune the pretrained Stable Diffusion (SD) UNet for 3 epochs across all experiments. For evaluation, we utilize two widely adopted stereo vision benchmarks: Middlebury 2014 [36] and KITTI 2015 [28], neither of which are included in our training set. Middlebury 2014 comprises 23 high-resolution indoor stereo pairs captured with wide-baseline stereo cameras under varying illumination conditions, providing challenging scenarios for stereo generation. KITTI 2015 contains 200 outdoor stereo" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "pairs of street scenes with LiDAR-based sparse disparity ground truth, offering diverse real-world testing scenarios. Following StereoDiffusion [55], we preprocess Middlebury images to " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " resolution, while KITTI images are center-cropped before resizing to " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " to maintain the most informative regions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 145, + 555, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 145, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 145, + 555, + 300 + ], + "type": "text", + "content": "We conduct comprehensive comparisons against both traditional and learning-based baselines. Traditional approaches include naive solutions such as leave blank and image stretching. For learning-based comparisons, we evaluate against state-of-the-art diffusion models adapted for stereo generation: Repaint [26] and SD-Inpainting [35], as well as StereoDiffusion, which specifically targets stereo image generation. To quantitatively assess the quality of generated right-view images " + }, + { + "bbox": [ + 313, + 145, + 555, + 300 + ], + "type": "inline_equation", + "content": "(\\hat{I_r})" + }, + { + "bbox": [ + 313, + 145, + 555, + 300 + ], + "type": "text", + "content": " against ground truth " + }, + { + "bbox": [ + 313, + 145, + 555, + 300 + ], + "type": "inline_equation", + "content": "(I_{r})" + }, + { + "bbox": [ + 313, + 145, + 555, + 300 + ], + "type": "text", + "content": " we employ three complementary metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) with SqueezeNet backbone." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 301, + 554, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 554, + 480 + ], + "type": "text", + "content": "Quantitative Results. Tab. 2 presents the generation performance on Middlebury 2014 and KITTI 2015. The sparsity of ground truth disparities in KITTI limits evaluation accuracy, making pseudo-label disparities more effective for assessing stereo consistency. This sparsity particularly affects inpainting-based methods, as they struggle to infer accurate right-view images in areas with missing depth information, especially for dynamic objects and distant regions. In contrast, Middlebury 2014, which provides dense ground truth disparities, demonstrates a smaller performance gap between using ground truth and using pseudo disparities, underscoring the advantage of high-quality disparity annotations for generation. We obtain pseudo disparities from CREStereo [22] pretrained model, where evaluation datasets are not used for training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 482, + 554, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 554, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 554, + 661 + ], + "type": "text", + "content": "Qualitative Results. We present qualitative comparisons between our method and several baselines: leave blank, StereoDiffusion, and SD-Inpainting, with ground truth as reference. As shown in Fig. 3, our method demonstrates superior performance on the Middlebury 2014 dataset. While SD-Inpainting struggles with occlusion handling, often generating semantically inconsistent content in occluded regions, and StereoDiffusion exhibits loss of fine details due to latent-space warping, our approach successfully maintains both geometric accuracy and visual fidelity. Particularly noteworthy is our method's ability to generate coherent right-view images while preserving pixel-level correspondence with the left images. Additional qualitative results on the KITTI 2015 dataset and more visualizations on other datasets can be found in the supplementary materials." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 671, + 447, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 447, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 447, + 684 + ], + "type": "text", + "content": "4.2. Unsupervised Learning" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": "Experimental Setup. We conduct two comprehensive experiments to validate the effectiveness of our proposed gen" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26583" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 102, + 524, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 2. Quantitative results of image generation for Middlebury 2014 and KITTI 2015 datasets. The top three results for each metric are highlighted with a first, second, and third background, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 102, + 524, + 232 + ], + "lines": [ + { + "bbox": [ + 85, + 102, + 524, + 232 + ], + "spans": [ + { + "bbox": [ + 85, + 102, + 524, + 232 + ], + "type": "table", + "html": "
MethodsSD versionMiddlebury 2014KITTI 2015
PSNR ↑SSIM ↑LPIPS ↓PSNR ↑SSIM ↑LPIPS ↓
Leave blank-11.328+2.4830.315+0.2300.450+0.09712.980+4.9190.374+0.2860.313+0.222
-3.489-0.154-0.089-3.831-0.251-0.109
Stretch-14.842+2.7140.432+0.1900.285+0.11214.757+5.3750.429+0.2870.212+0.145
-2.753+0.265-0.089-4.694-0.271-0.100
3D Photography [40]-14.190+2.4640.427+0.2380.275+0.07314.540+7.2560.398+0.2700.210+0.099
-2.798-0.175-0.065-4.023-0.323-0.073
RePaint [26]1.515.102+2.9090.462+0.2680.311+0.08815.056+5.3660.462+0.2680.251+0.128
-2.802-0.253-0.079-4.897-0.285-0.095
SD-Inpainting [35]1.515.740+3.5110.412+0.2360.311+0.0989.792+2.8090.230+0.2360.652+0.144
-3.351-0.188-0.164-4.070-0.229-0.139
StereoDiffusion [55]1.515.456+2.6690.468+0.2520.231+0.09615.679+5.8880.481+0.2450.205+0.135
-3.313-0.205-0.088-5.487-0.310-0.099
Ours + GT1.523.699+5.4560.866+0.0840.064+0.05320.439+7.5180.746+0.1550.108+0.117
-6.033-0.141-0.034-8.541-0.274-0.048
Ours + Pseudo1.523.835+6.8420.868+0.0960.062+0.06322.749+6.7220.813+0.1100.096+0.068
-6.741-0.157-0.038-8.102-0.152-0.104
Ours + GT2.124.819+6.5190.906+0.0570.061+0.03219.836+8.4830.765+0.2450.135+0.059
-4.252-0.132-0.032-7.090-0.092-0.068
Ours + Pseudo2.125.142+6.9700.911+0.0660.060+0.06123.488+6.4440.849+0.1300.109+0.050
", + "image_path": "85a69e565832226dede9b9a4a70040b626e295054226f15bdade49cfbe8a2099.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 242, + 138, + 324 + ], + "blocks": [ + { + "bbox": [ + 56, + 242, + 138, + 324 + ], + "lines": [ + { + "bbox": [ + 56, + 242, + 138, + 324 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 138, + 324 + ], + "type": "image", + "image_path": "4626db6c11940104f518bea321646586a4836eb60a01189874c9d0b4cde068f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 326, + 138, + 407 + ], + "blocks": [ + { + "bbox": [ + 56, + 326, + 138, + 407 + ], + "lines": [ + { + "bbox": [ + 56, + 326, + 138, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 326, + 138, + 407 + ], + "type": "image", + "image_path": "e35f46a7566228bd105a1072af012120eedf8ae32516f84ee8af4b56749ad612.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 408, + 138, + 491 + ], + "blocks": [ + { + "bbox": [ + 56, + 408, + 138, + 491 + ], + "lines": [ + { + "bbox": [ + 56, + 408, + 138, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 138, + 491 + ], + "type": "image", + "image_path": "4c99701b8f14edda3eac7936d93cdd5d18465709f8f6813a3adb971b07b3eaef.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 492, + 138, + 573 + ], + "blocks": [ + { + "bbox": [ + 56, + 492, + 138, + 573 + ], + "lines": [ + { + "bbox": [ + 56, + 492, + 138, + 573 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 138, + 573 + ], + "type": "image", + "image_path": "6673940f90317b9467a02dd793cbc4e6f71131540b557b91668350a0610e7722.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 575, + 130, + 586 + ], + "lines": [ + { + "bbox": [ + 65, + 575, + 130, + 586 + ], + "spans": [ + { + "bbox": [ + 65, + 575, + 130, + 586 + ], + "type": "text", + "content": "Reference Image" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 140, + 243, + 220, + 324 + ], + "blocks": [ + { + "bbox": [ + 140, + 243, + 220, + 324 + ], + "lines": [ + { + "bbox": [ + 140, + 243, + 220, + 324 + ], + "spans": [ + { + "bbox": [ + 140, + 243, + 220, + 324 + ], + "type": "image", + "image_path": "d9779e095e6a380a0cebc24fa84dc627f9ae4d45cdcddccaa5ff484a174fa8d8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 140, + 326, + 220, + 407 + ], + "blocks": [ + { + "bbox": [ + 140, + 326, + 220, + 407 + ], + "lines": [ + { + "bbox": [ + 140, + 326, + 220, + 407 + ], + "spans": [ + { + "bbox": [ + 140, + 326, + 220, + 407 + ], + "type": "image", + "image_path": "39c2eb229569dae5fdcf288a45c6e07c5bccc5c0f5aef1883979f7d94284c45e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 140, + 408, + 220, + 491 + ], + "blocks": [ + { + "bbox": [ + 140, + 408, + 220, + 491 + ], + "lines": [ + { + "bbox": [ + 140, + 408, + 220, + 491 + ], + "spans": [ + { + "bbox": [ + 140, + 408, + 220, + 491 + ], + "type": "image", + "image_path": "08769934ae2ec1826398d26c679943933846da98f088a22d287267d2bc587082.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 140, + 492, + 221, + 573 + ], + "blocks": [ + { + "bbox": [ + 140, + 492, + 221, + 573 + ], + "lines": [ + { + "bbox": [ + 140, + 492, + 221, + 573 + ], + "spans": [ + { + "bbox": [ + 140, + 492, + 221, + 573 + ], + "type": "image", + "image_path": "cf51d8a69ae0c53540d400439a7fb044a4596a9f502b8310ffb55b35d1a3ab4c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 575, + 205, + 585 + ], + "lines": [ + { + "bbox": [ + 157, + 575, + 205, + 585 + ], + "spans": [ + { + "bbox": [ + 157, + 575, + 205, + 585 + ], + "type": "text", + "content": "Leave Blank" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 223, + 243, + 304, + 324 + ], + "blocks": [ + { + "bbox": [ + 223, + 243, + 304, + 324 + ], + "lines": [ + { + "bbox": [ + 223, + 243, + 304, + 324 + ], + "spans": [ + { + "bbox": [ + 223, + 243, + 304, + 324 + ], + "type": "image", + "image_path": "31ea98a2d156060801743b2a9e76e9d10b2f0d83f6f0dd334a2005b83caaf770.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 223, + 326, + 304, + 407 + ], + "blocks": [ + { + "bbox": [ + 223, + 326, + 304, + 407 + ], + "lines": [ + { + "bbox": [ + 223, + 326, + 304, + 407 + ], + "spans": [ + { + "bbox": [ + 223, + 326, + 304, + 407 + ], + "type": "image", + "image_path": "d932d1ec3188b94be7d44086205892028959feacc77f07b6d2bda12bb03dcb48.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 223, + 408, + 304, + 491 + ], + "blocks": [ + { + "bbox": [ + 223, + 408, + 304, + 491 + ], + "lines": [ + { + "bbox": [ + 223, + 408, + 304, + 491 + ], + "spans": [ + { + "bbox": [ + 223, + 408, + 304, + 491 + ], + "type": "image", + "image_path": "5967345400804e8acbaf2021f80606f50817fe297f19e9d66c26e7d56e8573f8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 223, + 492, + 304, + 573 + ], + "blocks": [ + { + "bbox": [ + 223, + 492, + 304, + 573 + ], + "lines": [ + { + "bbox": [ + 223, + 492, + 304, + 573 + ], + "spans": [ + { + "bbox": [ + 223, + 492, + 304, + 573 + ], + "type": "image", + "image_path": "aebdb82329d735a030550ca5d57340b6d33e7d453ea266b2cfa4594b5ac12cbb.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 575, + 290, + 585 + ], + "lines": [ + { + "bbox": [ + 238, + 575, + 290, + 585 + ], + "spans": [ + { + "bbox": [ + 238, + 575, + 290, + 585 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 306, + 243, + 388, + 323 + ], + "blocks": [ + { + "bbox": [ + 306, + 243, + 388, + 323 + ], + "lines": [ + { + "bbox": [ + 306, + 243, + 388, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 388, + 323 + ], + "type": "image", + "image_path": "4e2a4b4ad75882de53bf77bf3fd7cd3981195d19deafe8814bd70d992169b8d3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 306, + 326, + 388, + 407 + ], + "blocks": [ + { + "bbox": [ + 306, + 326, + 388, + 407 + ], + "lines": [ + { + "bbox": [ + 306, + 326, + 388, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 326, + 388, + 407 + ], + "type": "image", + "image_path": "1ed3776bc4acc972c4b81d76d8746827ceb475d123e2c230b65d34e515d7fd72.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 306, + 408, + 388, + 491 + ], + "blocks": [ + { + "bbox": [ + 306, + 408, + 388, + 491 + ], + "lines": [ + { + "bbox": [ + 306, + 408, + 388, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 408, + 388, + 491 + ], + "type": "image", + "image_path": "3a85ece81b50054519d62a78672352903f300d95e4b039a212add1e0006fc325.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 306, + 492, + 388, + 573 + ], + "blocks": [ + { + "bbox": [ + 306, + 492, + 388, + 573 + ], + "lines": [ + { + "bbox": [ + 306, + 492, + 388, + 573 + ], + "spans": [ + { + "bbox": [ + 306, + 492, + 388, + 573 + ], + "type": "image", + "image_path": "c536a43c7b962480e27d95a57817f3cc142c6395346474e1405d1bbdee667c5e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 575, + 373, + 586 + ], + "lines": [ + { + "bbox": [ + 320, + 575, + 373, + 586 + ], + "spans": [ + { + "bbox": [ + 320, + 575, + 373, + 586 + ], + "type": "text", + "content": "SD-Inpainting" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 389, + 243, + 471, + 323 + ], + "blocks": [ + { + "bbox": [ + 389, + 243, + 471, + 323 + ], + "lines": [ + { + "bbox": [ + 389, + 243, + 471, + 323 + ], + "spans": [ + { + "bbox": [ + 389, + 243, + 471, + 323 + ], + "type": "image", + "image_path": "7d1be42928f302dd7cedc9810ad721d98b3089a8fafb70ca14ff97a465b2b82e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 389, + 326, + 471, + 407 + ], + "blocks": [ + { + "bbox": [ + 389, + 326, + 471, + 407 + ], + "lines": [ + { + "bbox": [ + 389, + 326, + 471, + 407 + ], + "spans": [ + { + "bbox": [ + 389, + 326, + 471, + 407 + ], + "type": "image", + "image_path": "307bfa7e1b8efd2b97b33703d60b265416ebe370137f38ed87e6934dcd3f4b0f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 389, + 408, + 471, + 491 + ], + "blocks": [ + { + "bbox": [ + 389, + 408, + 471, + 491 + ], + "lines": [ + { + "bbox": [ + 389, + 408, + 471, + 491 + ], + "spans": [ + { + "bbox": [ + 389, + 408, + 471, + 491 + ], + "type": "image", + "image_path": "713a917839ca3b21711c64fc409793faa7a8a0c1fb3b06b9210c5fb6a0e658eb.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 389, + 492, + 471, + 573 + ], + "blocks": [ + { + "bbox": [ + 389, + 492, + 471, + 573 + ], + "lines": [ + { + "bbox": [ + 389, + 492, + 471, + 573 + ], + "spans": [ + { + "bbox": [ + 389, + 492, + 471, + 573 + ], + "type": "image", + "image_path": "9b41ab7322bd3624ae38bf3563c21dc56a6f7ea10d01e24da83af9b4f6825f87.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 575, + 460, + 585 + ], + "lines": [ + { + "bbox": [ + 400, + 575, + 460, + 585 + ], + "spans": [ + { + "bbox": [ + 400, + 575, + 460, + 585 + ], + "type": "text", + "content": "StereoDiffusion" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 144, + 596, + 465, + 608 + ], + "lines": [ + { + "bbox": [ + 144, + 596, + 465, + 608 + ], + "spans": [ + { + "bbox": [ + 144, + 596, + 465, + 608 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison on Middlebury 2014 with ground truth disparity maps." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 472, + 243, + 553, + 324 + ], + "blocks": [ + { + "bbox": [ + 472, + 243, + 553, + 324 + ], + "lines": [ + { + "bbox": [ + 472, + 243, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 472, + 243, + 553, + 324 + ], + "type": "image", + "image_path": "61135b89c4a0ec761b3380772bff5b4f650c3d033c8242ee86d07e3280a000b7.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 472, + 326, + 553, + 407 + ], + "blocks": [ + { + "bbox": [ + 472, + 326, + 553, + 407 + ], + "lines": [ + { + "bbox": [ + 472, + 326, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 472, + 326, + 553, + 407 + ], + "type": "image", + "image_path": "852dcbbd60b718c68912436a0633ff3d9620027380d535ce9e0ac5fb1c603cb2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 472, + 408, + 553, + 491 + ], + "blocks": [ + { + "bbox": [ + 472, + 408, + 553, + 491 + ], + "lines": [ + { + "bbox": [ + 472, + 408, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 472, + 408, + 553, + 491 + ], + "type": "image", + "image_path": "f4592ed5f91857f3dce4d8e70f71b8f0fa616601fa0e6ede490b273eddc97e25.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 472, + 492, + 553, + 573 + ], + "blocks": [ + { + "bbox": [ + 472, + 492, + 553, + 573 + ], + "lines": [ + { + "bbox": [ + 472, + 492, + 553, + 573 + ], + "spans": [ + { + "bbox": [ + 472, + 492, + 553, + 573 + ], + "type": "image", + "image_path": "76ce018505aaa8e4052d8d7480232010c24be52ebe3da5db2fe54d1510214c21.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 575, + 523, + 585 + ], + "lines": [ + { + "bbox": [ + 503, + 575, + 523, + 585 + ], + "spans": [ + { + "bbox": [ + 503, + 575, + 523, + 585 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 629, + 256, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 256, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 256, + 640 + ], + "type": "text", + "content": "eration method for unsupervised stereo matching." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 700 + ], + "type": "text", + "content": "In the first experiment, we fine-tune a stereo matching model pretrained on the SceneFlow dataset using the KITTI 2012 and KITTI 2015 benchmarks. We split the training and evaluation sets into 160/34 images for KITTI 2012 and 160/40 images for KITTI 2015, following [60]." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "The second experiment evaluates various unsupervised" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "stereo matching approaches by generating stereo images. Specifically, KITTI 2012 [10] has 194 stereo images, KITTI 2015 has 200 stereo images, ETH3D [37] has 27 stereo pairs, and the Middlebury v3 training set [36] has 15 stereo pairs at Quarter and Half resolutions (Midd-Q, Midd-H). This diverse dataset selection allows us to comprehensively assess the generalization capability of our method across" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26584" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 123, + 280, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 117 + ], + "type": "text", + "content": "Table 3. Comparison with other diffusion-based stereo image generation methods for unsupervised stereo matching. The first group of methods is trained on PSMNet while the second group is trained on RAFT-Stereo. SD v1.5 is used for the experiments." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 123, + 280, + 228 + ], + "lines": [ + { + "bbox": [ + 70, + 123, + 280, + 228 + ], + "spans": [ + { + "bbox": [ + 70, + 123, + 280, + 228 + ], + "type": "table", + "html": "
MethodsKITTI 2012KITTI 2015
D1-all ↓EPE ↓D1-all ↓EPE ↓
Leave blank25.2193.54224.5682.937
SD-Inpainting [35]3.9070.8944.4901.059
StereoDiffusion [55]15.2132.2205.6511.154
Ours3.8020.8153.9330.991
Leave blank4.1840.9524.0740.957
SD-Inpainting [35]4.0160.9454.1280.972
StereoDiffusion [55]12.7561.8116.3601.174
Ours3.5370.9303.6530.943
", + "image_path": "2dbff5b2205090804c971477d9004d877f3cfca3d5ddb5782d727cd4eb80c9fc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 195, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 195, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 195, + 262 + ], + "type": "text", + "content": "different domains and resolutions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": "We evaluate the performance of various unsupervised stereo matching approaches using the D1-all metric, which measures the percentage of incorrect disparity predictions (error " + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "inline_equation", + "content": ">3\\mathrm{px}" + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": " of the true disparity), the End-PointError (EPE), and the " + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "inline_equation", + "content": ">n\\mathrm{px}" + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": " metric, which denotes the percentage of pixels with an EPE exceeding " + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": " pixels. To demonstrate the architecture-agnostic nature of our approach, we employ three distinct stereo models: PSM-Net [6], a cost-volume-based method, RAFT-Stereo [24], an iterative refinement model, and IGEV++ [62], one of the most recent methods. For disparity estimation, we utilize the Depth Anything Model v2 (DAMv2) [65], from which we obtain the disparity map. The disparity values are normalized to the range " + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "inline_equation", + "content": "[0, 1]" + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": " and subsequently scaled by a factor " + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 265, + 297, + 445 + ], + "type": "text", + "content": " to enable flexibility generation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 446, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 554 + ], + "type": "text", + "content": "Fine-tuning on KITTI. To ensure fair and consistent comparisons across all generation methods, we construct the training datasets using a standardized procedure. Specifically, we sample the scale factor " + }, + { + "bbox": [ + 55, + 446, + 296, + 554 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 446, + 296, + 554 + ], + "type": "text", + "content": " from the set " + }, + { + "bbox": [ + 55, + 446, + 296, + 554 + ], + "type": "inline_equation", + "content": "\\{0.05, 0.1, 0.15, 0.2, 0.25\\}" + }, + { + "bbox": [ + 55, + 446, + 296, + 554 + ], + "type": "text", + "content": ". For each selected scale, we employ various generation methods to create the right images based on 160 left images and their corresponding scaled disparity maps for both KITTI 2012 and KITTI 2015. This procedure results in a total of 800 stereo pairs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 555, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 628 + ], + "type": "text", + "content": "As shown in Tab. 3, higher generation quality leads to improved stereo matching performance. Among the compared methods, SD-Inpainting achieves the closest performance to ours due to its pixel-level warping operation. However, it falls short in providing semantic consistency in occluded areas, which limits its overall effectiveness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "content": "Unsupervised Generalization. Following the methodology of constructing the Mono for Stereo (MfS) dataset in [60], we construct our training dataset, MfS-GenStereo, from the same datasets, including ADE20K [68], Mapillary Vistas [30], DIODE [54], Depth in the Wild [7], and COCO 2017 [23]. During the generation process, we randomly select a scale factor in the range " + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "inline_equation", + "content": "[0, 0.35]" + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "content": " and set the maxi" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "mum disparity to 256 for PSMNet. As shown in Tab. 4, our method enables unsupervised stereo matching using only left images, achieving superior performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 109, + 556, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 556, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 556, + 313 + ], + "type": "text", + "content": "We compare our approach with the original MfS [60] and its enhanced variant using the Depth Anything Model v2 (DAMv2), which generates the right view images by filling the occluded areas with random background patches from the dataset. The experiments indicate that, despite the better depth prior, the synthesis of right-view images remains a critical bottleneck in this methodology. Additionally, we compare with Mono2Stereo, which fine-tunes SD-Inpainting for right-view image generation, and NeRF-Stereo, which generates stereo images by reconstructing the static scenes using Neural Radiance Fields (NeRF), limiting the diversity of its datasets. Our experimental results indicate that our proposed method consistently surpasses these existing approaches in generating high-quality stereo pairs with pixel-level accuracy, demonstrating robust generalization capabilities and significantly enhancing stereo matching performance across diverse datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 323, + 410, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 410, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 410, + 335 + ], + "type": "text", + "content": "4.3. Ablation Study." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 341, + 556, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 556, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 556, + 484 + ], + "type": "text", + "content": "Random disparity drop. Since the ground-truth disparity maps in real-world datasets (e.g., KITTI) are often derived from LiDAR, resulting in sparse and incomplete point clouds, our random disparity dropout strategy mitigates the model's reliance on dense disparity supervision. This enhancement allows the model to generate plausible right images even when only sparse disparity maps are available. This strategy significantly improves performance on datasets with sparse ground-truth annotations, demonstrating better generalization to real-world scenarios. As shown in Tab. 5, comparing rows (1) and (2) highlights the effectiveness of this strategy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 485, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 556, + 628 + ], + "type": "text", + "content": "Mixed datasets. We train our model on Virtual KITTI 2 (VKITTI2) [5] and a combination of mixed datasets, and subsequently evaluate both models. When trained exclusively on VKITTI2, the model achieves comparable performance to the mixed-dataset model on KITTI 2015, owing to the relatively small domain gap between VKITTI2 and KITTI 2015. However, this single-dataset training results in notable performance degradation on other datasets, indicating limited generalization. As shown in Tab. 5, comparing rows (3) and (8) demonstrates that our mixed dataset training strategy significantly improves the model's generalization across diverse domains." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 630, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 557, + 715 + ], + "type": "text", + "content": "Coordinate embedding and the warped image. Incorporating the coordinate embedding process effectively encodes spatial information and reinforces the structural relationship between stereo views. Additionally, using the warped image as a condition enhances performance by providing pixel-level guidance. The warped image serves as a coarse prediction of the right view, helping the model focus" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26585" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 573, + 251 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 4. Zero-Shot Generalization Benchmark. The first group shows methods trained with PSMNet, the second group shows methods trained with RAFT-Stereo, and the third group represents IGEV++." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 573, + 251 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 573, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 573, + 251 + ], + "type": "table", + "html": "
MethodsKITTI 2012KITTI 2015ETH3DMiddlebury-QMiddlebury-H
D1-allEPE>3pxD1-allEPE>3pxD1-allEPE>1pxD1-allEPE>2pxD1-allEPE>2px
PSMNet [6]26.693.87727.3528.244.00328.534.2581.98915.7415.004.94120.2021.349.89933.26
MFS [60]4.3221.0094.7034.9631.6365.1812.2040.5358.1709.2171.64812.0712.343.24217.56
MfS+DAMv2 [65]4.9311.0635.3135.5041.4015.7371.9900.4938.1459.4471.82511.9012.343.67817.21
Mono2Stereo [59]4.0770.859-4.5071.052-1.5310.397-------
NeRFStereo [50]3.6830.8514.5074.8131.4485.0411.6200.55711.687.6191.45111.048.8922.59012.88
Ours + SD v1.53.6760.8464.0633.8790.9964.1121.1140.3785.5566.8871.13910.4311.332.26919.39
Ours + SD v2.13.8070.8204.1223.8980.9784.0960.8940.3244.1015.0370.8607.4638.0461.70913.73
RAFT-Stereo [24]3.9440.8304.2395.2541.1325.4680.8830.2672.6075.1571.21710.255.7741.43611.20
NeRFStereo [50]3.7870.8554.1405.3561.4695.5610.8300.3013.3705.8191.0038.3707.4261.82010.95
Ours + SD v1.53.2240.7343.4554.6831.0424.8970.9590.2942.6673.4501.0037.9524.4461.5149.94
Ours + SD v2.13.2190.7163.0193.9500.9714.2220.8350.2772.6384.0671.1279.3824.6181.52711.11
IGEV++ [62]5.6811.0996.1966.0281.2656.2391.7010.3564.6664.2380.8286.2745.2922.3847.752
Ours + SD v2.12.8020.6863.0153.7460.9524.0041.2260.3263.4293.6820.6845.5122.7970.8425.098
", + "image_path": "f48a6f4eb7b3ac59ac97cd6873681a753205dfe17d8902181b6489700fe97fb8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 67, + 282, + 541, + 387 + ], + "blocks": [ + { + "bbox": [ + 190, + 262, + 418, + 274 + ], + "lines": [ + { + "bbox": [ + 190, + 262, + 418, + 274 + ], + "spans": [ + { + "bbox": [ + 190, + 262, + 418, + 274 + ], + "type": "text", + "content": "Table 5. Ablation studies. SD v1.5 is used for the experiments." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 282, + 541, + 387 + ], + "lines": [ + { + "bbox": [ + 67, + 282, + 541, + 387 + ], + "spans": [ + { + "bbox": [ + 67, + 282, + 541, + 387 + ], + "type": "table", + "html": "
Disparity Source#Mixed DatasetsRandom DropCoord EmbeddingWarped ImagePixel LossFusionMiddlebury 2014KITTI 2015
PSNRSSIMLPIPSPSNRSSIMLPIPS
Ground Truth(1)✓X✓✓✓✓23.5520.8620.064116.1640.6640.1740
(2)✓✓✓✓✓✓23.6990.8660.063520.4390.7460.1079
Pseudo Disparities(3)X✓✓✓✓✓20.0370.7450.121421.0850.7770.1113
(4)✓✓✓XXX22.0330.7960.073121.8660.7790.1002
(5)✓✓X✓XX22.6970.8320.071522.3130.7980.1011
(6)✓✓✓✓XX23.4070.8550.064422.5200.8050.0974
(7)✓✓✓✓✓X23.8110.8670.062122.7160.8110.0961
(8)✓✓✓✓✓✓23.8350.8680.062022.7490.8130.0958
", + "image_path": "9d4b6f94da68350eef54b95ad089221d4da91351522c7f13fe17c72ff6b0e952.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 407, + 294, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 407, + 294, + 550 + ], + "spans": [ + { + "bbox": [ + 54, + 407, + 294, + 550 + ], + "type": "text", + "content": "on correcting local inconsistencies rather than synthesizing the entire image from scratch. As shown in rows (4), (5), and (6) in Tab. 5, the combination of coordinate embedding and warped image significantly improves generation performance, preserving geometric consistency and reducing artifacts like stitching boundaries. While warped image embedding contributes most to performance, its effectiveness decreases with additional fine-tuning iterations, revealing instability when using warped image alone as a conditioning signal. This underscores the importance of combining coordinate embedding and warped image to achieve stable, high-quality stereo image generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 553, + 294, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 553, + 294, + 638 + ], + "spans": [ + { + "bbox": [ + 54, + 553, + 294, + 638 + ], + "type": "text", + "content": "Pixel-level loss. By decoding the latent representations back to the image space, the model learns to minimize discrepancies directly in the observed image space, enforcing pixel-level alignment. As illustrated in rows (6) and (7), the introduction of the pixel-level loss significantly improves the generation performance, highlighting its role in maintaining high-fidelity stereo image generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "Adaptive Fusion. The proposed adaptive fusion module significantly enhances visual quality by dynamically weighting the contributions of both the generated and warped images. A comparison of rows (7) and (8) demonstrates that this fusion strategy further refines pixel alignment, resulting in sharper, more structurally coherent, and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 407, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 455 + ], + "type": "text", + "content": "perceptually accurate reconstructions. The improvement in PSNR is more pronounced than that in SSIM and LPIPS, indicating that the adaptive fusion is particularly effective for pixel-level alignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 465, + 388, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 388, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 388, + 478 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 485, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 553, + 628 + ], + "type": "text", + "content": "We presented GenStereo, a novel diffusion-based framework for open-world stereo image generation with applications in unsupervised stereo matching. Our method introduces several key innovations, including disparity-aware coordinate embeddings along with warped image embeddings, pixel-level loss, and an adaptive fusion module, to ensure high-quality stereo image generation with strong geometric and semantic consistency. Through extensive experiments across diverse datasets, we show that GenStereo significantly outperforms existing methods in both stereo generation and unsupervised matching. Our ablation studies demonstrate the efficacy of each proposed component." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 628, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 553, + 712 + ], + "type": "text", + "content": "Limitations. Despite promising results, diffusion-based models face inherent challenges with large disparities when generating right-view images due to large unconditioned regions. While our data augmentation such as random cropping and resizing mitigates these issues and performs well in typical stereo setups, future work could explore extending these models to accommodate larger disparities." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26586" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "type": "text", + "content": "We gratefully acknowledge the advanced computational resources provided by Engineering IT and Research Infrastructure Services at Washington University in St. Louis." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 159, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 213 + ], + "type": "text", + "content": "[1] Filippo Aleotti, Fabio Tosi, Li Zhang, Matteo Poggi, and Stefano Mattoccia. Reversing the cycle: self-supervised deep stereo through enhanced monocular distillation. In European Conference on Computer Vision, pages 614-632. Springer, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 215, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 215, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 215, + 296, + 258 + ], + "type": "text", + "content": "[2] Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences, 63:1-11, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 260, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 260, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 61, + 260, + 296, + 304 + ], + "type": "text", + "content": "[3] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 306, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 306, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 306, + 296, + 350 + ], + "type": "text", + "content": "[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In European Conf. on Computer Vision, pages 611-625. Springer-Verlag, 2012. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 351, + 296, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 296, + 373 + ], + "type": "text", + "content": "[5] Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual katii 2. arXiv preprint arXiv:2001.10773, 2020. 1, 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 374, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 296, + 407 + ], + "type": "text", + "content": "[6] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In IEEE Conference on Computer Vision and Pattern Recognition, pages 5410-5418, 2018. 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 409, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 296, + 441 + ], + "type": "text", + "content": "[7] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 443, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 296, + 498 + ], + "type": "text", + "content": "[8] Cheng Chi, Qingjie Wang, Tianyu Hao, Peng Guo, and Xin Yang. Feature-level collaboration: Joint unsupervised learning of optical flow, stereo depth and camera motion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2463-2473, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 499, + 296, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 296, + 543 + ], + "type": "text", + "content": "[9] Jaehoon Cho, Dongbo Min, Youngjung Kim, and Kwanghoon Sohn. Deep monocular depth estimation leveraging a large-scale outdoor stereo dataset. Expert Systems with Applications, 178:114877, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 544, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 296, + 588 + ], + "type": "text", + "content": "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361. IEEE, 2012. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 296, + 633 + ], + "type": "text", + "content": "[11] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "type": "text", + "content": "[12] Peter Hedman, Suhib Alsisan, Richard Szeliski, and Johannes Kopf. Casual 3d photography. ACM Transactions on Graphics (TOG), 36(6):1-15, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 669, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 296, + 712 + ], + "type": "text", + "content": "[13] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics (ToG), 37(6):1-15, 2018. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 150 + ], + "type": "text", + "content": "[15] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[16] Junpeng Jing, Ye Mao, and Krystian Mikolajczyk. Match-stereo-videos: Bidirectional alignment for consistent dynamic stereo matching. In European Conference on Computer Vision, pages 415–432. Springer, 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 219 + ], + "type": "text", + "content": "[17] Laurent Valentin Jospin, Hamid Laga, Farid Boussaid, and Mohammed Bennamoun. Active-passive simstereo, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 221, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 221, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 221, + 553, + 275 + ], + "type": "text", + "content": "[18] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Dynamic stereo: Consistent dynamic depth from stereo videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13229-13239, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 342 + ], + "type": "text", + "content": "[19] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 344, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 387 + ], + "type": "text", + "content": "[20] Johannes Kopf, Fabian Langguth, Daniel Scharstein, Richard Szeliski, and Michael Goesele. Image-based rendering in the gradient domain. ACM Transactions on Graphics (TOG), 32(6):1-9, 2013. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 388, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 433 + ], + "type": "text", + "content": "[21] Hsueh-Ying Lai, Yi-Hsuan Tsai, and Wei-Chen Chiu. Bridging stereo matching and optical flow via spatiotemporal correspondence. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1890-1899, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 434, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 498 + ], + "type": "text", + "content": "[22] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16263-16272, 2022. 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision, pages 740-755. Springer, 2014. 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "text", + "content": "[24] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In IEEE International Conference on 3D Vision, pages 218-227, 2021. 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "type": "text", + "content": "[25] Rui Liu, Chengxi Yang, Wenxiu Sun, Xiaogang Wang, and Hongsheng Li. Stereogan: Bridging synthetic-to-real domain gap by joint optimization of domain translation and stereo matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12757-12766, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[26] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461-11471, 2022. 5, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26587" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[27] Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nalivayko, and Andres Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "text", + "content": "[28] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3061-3070, 2015. 1, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 218 + ], + "type": "text", + "content": "[29] Jiteng Mu, Shalini De Mello, Zhiding Yu, Nuno Vasconcelos, Xiaolong Wang, Jan Kautz, and Sifei Liu. Coordgan: Self-supervised dense correspondences emerge from gans. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10011-10020, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 263 + ], + "type": "text", + "content": "[30] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In IEEE international conference on computer vision, pages 4990-4999, 2017. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 308 + ], + "type": "text", + "content": "[31] Matteo Poggi, Alessio Tonioni, Fabio Tosi, Stefano Mattoccia, and Luigi Di Stefano. Continual adaptation for deep stereo. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4713-4729, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 376 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 432 + ], + "type": "text", + "content": "[33] Pierluigi Zama Ramirez, Alex Costanzino, Fabio Tosi, Matteo Poggi, Samuele Salti, Stefano Mattoccia, and Luigi Di Stefano. Booster: a benchmark for depth from images of specular and transparent surfaces. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "type": "text", + "content": "[34] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 44(3):1623-1637, 2020. 1, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 612 + ], + "type": "text", + "content": "[36] Daniel Scharstein, Heiko Hirschmüller, York Kitajima, Greg Krathwohl, Nera Nesic, Xi Wang, and Porter Westling. High-resolution stereo datasets with subpixel-accurate ground truth. In Pattern Recognition: 36th German Conference, GCPR 2014, Münster, Germany, September 2-5, 2014, Proceedings 36, pages 31-42. Springer, 2014. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 678 + ], + "type": "text", + "content": "[37] Thomas Schops, Johannes L Schonberger, Silvano Galliani, Torsten Sattler, Konrad Schindler, Marc Pollefeys, and Andreas Geiger. A multi-view stereo benchmark with high-resolution images and multi-camera videos. In IEEE Conference on Computer Vision and Pattern Recognition, pages 3260-3269, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 714 + ], + "type": "text", + "content": "[38] Junyoung Seo, Kazumi Fukuda, Takashi Shibuya, Takuya Narihira, Naoki Murata, Shoukang Hu, Chieh-Hsin Lai, Seungryong Kim, and Yuki Mitsufuji. Genwarp: Single image" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "to novel views with semantic-preserving generative warping. Advances in Neural Information Processing Systems (NeurIPS), 2024. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 139 + ], + "type": "text", + "content": "[39] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "text", + "content": "[40] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 217 + ], + "type": "text", + "content": "[41] J. Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 270 + ], + "type": "text", + "content": "[42] Xiao Song, Guorun Yang, Xinge Zhu, Hui Zhou, Zhe Wang, and Jianping Shi. Adastereo: A simple and efficient approach for adaptive stereo matching. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10328-10337, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 337 + ], + "type": "text", + "content": "[43] Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 33:7537-7547, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[44] Joshua Tokarsky, Ibrahim Abdulhafiz, Satya Ayyalasomayajula, Mostafa Mohsen, Navya G. Rao, and Adam Forbes. PLT-D3: A High-fidelity Dynamic Driving Simulation Dataset for Stereo Depth and Scene Flow, 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[45] Alessio Tonioni, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Unsupervised adaptation for deep stereo. In IEEE International Conference on Computer Vision, pages 1605-1613, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "text", + "content": "[46] Alessio Tonioni, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Unsupervised domain adaptation for depth prediction from images. IEEE transactions on pattern analysis and machine intelligence, 42(10):2396-2409, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 471, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 524 + ], + "type": "text", + "content": "[47] Alessio Tonioni, Oscar Rahnama, Thomas Joy, Luigi Di Stefano, Thalaiyasingam Ajanthan, and Philip HS Torr. Learning to adapt for stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9661-9670, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 525, + 553, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 570 + ], + "type": "text", + "content": "[48] Alessio Tonioni, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, and Luigi Di Stefano. Real-time self-adaptive deep stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 195-204, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "text", + "content": "[49] Fabio Tosi, Yiyi Liao, Carolin Schmitt, and Andreas Geiger. Smd-nets: Stereo mixture density networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8942-8952, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "type": "text", + "content": "[50] Fabio Tosi, Alessio Tonioni, Daniele De Gregorio, and Matteo Poggi. Nerf-supervised deep stereo. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 855-866, 2023. 2, 3, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "text", + "content": "[51] Fabio Tosi, Luca Bartolomei, and Matteo Poggi. A survey on deep stereo matching in the twenties. International Journal of Computer Vision, 2025. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 714 + ], + "type": "text", + "content": "[52] Jonathan Tremblay, Thang To, and Stan Birchfield. Falling things: A synthetic dataset for 3d object detection and pose" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26588" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "text", + "content": "estimation. In IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 2038-2041, 2018. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 148 + ], + "type": "text", + "content": "[53] Hung-Yu Tseng, Qinbo Li, Changil Kim, Suhib Alsisan, JiaBin Huang, and Johannes Kopf. Consistent view synthesis with pose-guided diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16773-16783, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 205 + ], + "type": "text", + "content": "[54] Igor Vasiljevic, Nick Kolkin, Shanyi Zhang, Ruotian Luo, Haochen Wang, Falcon Z. Dai, Andrea F. Daniele, Mohammadreza Mostajabi, Steven Basart, Matthew R. Walter, and Gregory Shakhnarovich. DIODE: A Dense Indoor and Outdoor DEpth Dataset. CoRR, abs/1908.00463, 2019. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 260 + ], + "type": "text", + "content": "[55] Lezhong Wang, Jeppe Revall Frisvad, Mark Bo Jensen, and Siavash Arjomand Bigdeli. Stereodiffusion: Training-free stereo image generation using latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7416-7425, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 262, + 294, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 294, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 294, + 325 + ], + "type": "text", + "content": "[56] Qiang Wang, Shizhen Zheng, Qingsong Yan, Fei Deng, Kaiyong Zhao, and Xiaowen Chu. IRS: A large naturalistic indoor robotics stereo dataset to train deep models for disparity and surface normal estimation. In IEEE International Conference on Multimedia and Expo, pages 1-6. IEEE, 2021. 1, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "text", + "content": "[57] Wenshan Wang, Delong Zhu, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. Tartanair: A dataset to push the limits of visual slam. In IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4909-4916. IEEE, 2020. 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 448 + ], + "type": "text", + "content": "[58] Yang Wang, Peng Wang, Zhenheng Yang, Chenxu Luo, Yi Yang, and Wei Xu. Unos: Unified unsupervised optical-flow and stereo-depth estimation by watching videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8071-8081, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "text", + "content": "[59] Yuran Wang, Yingping Liang, Hesong Li, and Ying Fu. Mono2stereo: Monocular knowledge transfer for enhanced stereo matching. arXiv preprint arXiv:2411.09151, 2024. 3, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[60] Jamie Watson, Oisin Mac Aodha, Daniyar Turmukhambetov, Gabriel J Brostow, and Michael Firman. Learning stereo from single images. In European Conference on Computer Vision, pages 722-740. Springer, 2020. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "text", + "content": "[61] Zhexiao Xiong, Feng Qiao, Yu Zhang, and Nathan Jacobs. Stereoflowgan: Co-training for stereo and flow with unsupervised domain adaptation. In 34th British Machine Vision Conference 2023, BMVC 2023, Aberdeen, UK, November 20-24, 2023. BMVA, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 645 + ], + "type": "text", + "content": "[62] Gangwei Xu, Xianqi Wang, Zhaoxing Zhang, Junda Cheng, Chunyuan Liao, and Xin Yang. Igev++: Iterative multirange geometry encoding volumes for stereo matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[63] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 296 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "text", + "content": "[64] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 171 + ], + "type": "text", + "content": "[65] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875-21911, 2024. 1, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[66] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "text", + "content": "[67] Yiran Zhong, Yuchao Dai, and Hongdong Li. Self-supervised learning for stereo matching with self-improving ability. arXiv preprint arXiv:1709.00930, 2017. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 296 + ], + "type": "text", + "content": "[68] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In IEEE Conference on Computer Vision and Pattern Recognition, pages 633-641, 2017. 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26589" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file